diff --git a/.gitattributes b/.gitattributes index 65c0522b169b4fb5176928a3a391cee01357d4fb..8cb6ae57079c3abd1f015939d8b47a3f9cea312c 100644 --- a/.gitattributes +++ b/.gitattributes @@ -2343,3 +2343,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2025/Towards[[:space:]]Transformer-Based[[:space:]]Aligned[[:space:]]Generation[[:space:]]with[[:space:]]Self-Coherence[[:space:]]Guidance/0430b3b2-c194-4129-b2be-acaa49b5720f_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/Towards[[:space:]]Unbiased[[:space:]]and[[:space:]]Robust[[:space:]]Spatio-Temporal[[:space:]]Scene[[:space:]]Graph[[:space:]]Generation[[:space:]]and[[:space:]]Anticipation/61a4ff36-4fd8-43f1-a4cb-46547e183676_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/Towards[[:space:]]Understanding[[:space:]]How[[:space:]]Knowledge[[:space:]]Evolves[[:space:]]in[[:space:]]Large[[:space:]]Vision-Language[[:space:]]Models/feb88a82-d0b8-411b-8ce3-184786042b8f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Towards[[:space:]]Understanding[[:space:]]and[[:space:]]Quantifying[[:space:]]Uncertainty[[:space:]]for[[:space:]]Text-to-Image[[:space:]]Generation/c3ad3744-8ee2-4117-a4a1-7d23c58e3fe4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Towards[[:space:]]Universal[[:space:]]AI-Generated[[:space:]]Image[[:space:]]Detection[[:space:]]by[[:space:]]Variational[[:space:]]Information[[:space:]]Bottleneck[[:space:]]Network/c23bce64-f04c-4e24-b700-3d6473e34043_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Towards[[:space:]]Universal[[:space:]]Dataset[[:space:]]Distillation[[:space:]]via[[:space:]]Task-Driven[[:space:]]Diffusion/cd8b4c9e-1814-423f-bdad-2b631dc6a75a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Towards[[:space:]]Universal[[:space:]]Soccer[[:space:]]Video[[:space:]]Understanding/bd5f526c-bc30-45cc-bbcd-e151dfc21f1a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Towards[[:space:]]Visual[[:space:]]Discrimination[[:space:]]and[[:space:]]Reasoning[[:space:]]of[[:space:]]Real-World[[:space:]]Physical[[:space:]]Dynamics_[[:space:]]Physics-Grounded[[:space:]]Anomaly[[:space:]]Detection/e496d86c-0d1a-42d1-b098-7f552a70c4ff_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Towards[[:space:]]Zero-Shot[[:space:]]Anomaly[[:space:]]Detection[[:space:]]and[[:space:]]Reasoning[[:space:]]with[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Models/9698d3f1-26f6-4c2e-aef0-60a71f074ba9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Towards[[:space:]]a[[:space:]]Universal[[:space:]]Synthetic[[:space:]]Video[[:space:]]Detector_[[:space:]]From[[:space:]]Face[[:space:]]or[[:space:]]Background[[:space:]]Manipulations[[:space:]]to[[:space:]]Fully[[:space:]]AI-Generated[[:space:]]Content/49360ca2-47e3-4c7b-81b1-6a87621996c4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Tra-MoE_[[:space:]]Learning[[:space:]]Trajectory[[:space:]]Prediction[[:space:]]Model[[:space:]]from[[:space:]]Multiple[[:space:]]Domains[[:space:]]for[[:space:]]Adaptive[[:space:]]Policy[[:space:]]Conditioning/d41e3854-0a29-4aee-85fc-0fcf3b667c89_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TraF-Align_[[:space:]]Trajectory-aware[[:space:]]Feature[[:space:]]Alignment[[:space:]]for[[:space:]]Asynchronous[[:space:]]Multi-agent[[:space:]]Perception/0f12efc0-153e-4cda-95dd-b16491feb48e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Track[[:space:]]Any[[:space:]]Anomalous[[:space:]]Object_A[[:space:]]Granular[[:space:]]Video[[:space:]]Anomaly[[:space:]]Detection[[:space:]]Pipeline/1ddd1714-9e44-4427-b96d-74620c052e99_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Track4Gen_[[:space:]]Teaching[[:space:]]Video[[:space:]]Diffusion[[:space:]]Models[[:space:]]to[[:space:]]Track[[:space:]]Points[[:space:]]Improves[[:space:]]Video[[:space:]]Generation/52b14d4e-8da8-4b5e-a32c-278d64dd36e6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Tracktention_[[:space:]]Leveraging[[:space:]]Point[[:space:]]Tracking[[:space:]]to[[:space:]]Attend[[:space:]]Videos[[:space:]]Faster[[:space:]]and[[:space:]]Better/55c83ffe-0183-42a7-a82c-e5f5408552e5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Training[[:space:]]Data[[:space:]]Provenance[[:space:]]Verification_[[:space:]]Did[[:space:]]Your[[:space:]]Model[[:space:]]Use[[:space:]]Synthetic[[:space:]]Data[[:space:]]from[[:space:]]My[[:space:]]Generative[[:space:]]Model[[:space:]]for[[:space:]]Training_/95a43e45-ca4b-40d2-9bb6-79d7e388dedb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Training-free[[:space:]]Dense-Aligned[[:space:]]Diffusion[[:space:]]Guidance[[:space:]]for[[:space:]]Modular[[:space:]]Conditional[[:space:]]Image[[:space:]]Synthesis/237b334f-3be2-4e40-8ac1-466a51400496_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Training-free[[:space:]]Neural[[:space:]]Architecture[[:space:]]Search[[:space:]]through[[:space:]]Variance[[:space:]]of[[:space:]]Knowledge[[:space:]]of[[:space:]]Deep[[:space:]]Network[[:space:]]Weights/740054ef-86e4-4bac-b7b3-53d57b1f3a66_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Trajectory[[:space:]]Mamba_[[:space:]]Efficient[[:space:]]Attention-Mamba[[:space:]]Forecasting[[:space:]]Model[[:space:]]Based[[:space:]]on[[:space:]]Selective[[:space:]]SSM/14500b01-a80f-4244-a6f9-1c844f2302da_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TransPixeler_[[:space:]]Advancing[[:space:]]Text-to-Video[[:space:]]Generation[[:space:]]with[[:space:]]Transparency/a57aba62-4ac6-4f75-a668-a4f223b5c05f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Transfer[[:space:]]Your[[:space:]]Perspective_[[:space:]]Controllable[[:space:]]3D[[:space:]]Generation[[:space:]]from[[:space:]]Any[[:space:]]Viewpoint[[:space:]]in[[:space:]]a[[:space:]]Driving[[:space:]]Scene/0e711492-3ab2-4748-8147-c897ce7d76b8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Transformers[[:space:]]without[[:space:]]Normalization/05ce47c0-7126-423f-aeff-039d314c6fa4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Traversing[[:space:]]Distortion-Perception[[:space:]]Tradeoff[[:space:]]using[[:space:]]a[[:space:]]Single[[:space:]]Score-Based[[:space:]]Generative[[:space:]]Model/dcdd686b-eeb1-4343-a7e6-20b5cd329198_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TreeMeshGPT_[[:space:]]Artistic[[:space:]]Mesh[[:space:]]Generation[[:space:]]with[[:space:]]Autoregressive[[:space:]]Tree[[:space:]]Sequencing/cc43c3b2-b6bf-472a-9dd6-c74d4f03cee9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TriTex_[[:space:]]Learning[[:space:]]Texture[[:space:]]from[[:space:]]a[[:space:]]Single[[:space:]]Mesh[[:space:]]via[[:space:]]Triplane[[:space:]]Semantic[[:space:]]Features/f0b8cf58-1685-4de9-9df0-451922c97e91_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Tripartite[[:space:]]Weight-Space[[:space:]]Ensemble[[:space:]]for[[:space:]]Few-Shot[[:space:]]Class-Incremental[[:space:]]Learning/6894bed0-33d2-4729-8056-07cf7489b89f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Tuning[[:space:]]the[[:space:]]Frequencies_[[:space:]]Robust[[:space:]]Training[[:space:]]for[[:space:]]Sinusoidal[[:space:]]Neural[[:space:]]Networks/79e8b6d6-50a7-4215-bd2a-fb5524e17539_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Turbo3D_[[:space:]]Ultra-fast[[:space:]]Text-to-3D[[:space:]]Generation/bd4523a6-3f35-4dc6-9b48-90da3f36ebea_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TurboFill_[[:space:]]Adapting[[:space:]]Few-step[[:space:]]Text-to-image[[:space:]]Model[[:space:]]for[[:space:]]Fast[[:space:]]Image[[:space:]]Inpainting/6302d669-b0e3-4a35-adf2-a42f09f692e9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Twinner_[[:space:]]Shining[[:space:]]Light[[:space:]]on[[:space:]]Digital[[:space:]]Twins[[:space:]]in[[:space:]]a[[:space:]]Few[[:space:]]Snaps/0293a039-5500-41cc-90fa-fe4f8710c6e9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Two[[:space:]]by[[:space:]]Two_[[:space:]]Learning[[:space:]]Multi-Task[[:space:]]Pairwise[[:space:]]Objects[[:space:]]Assembly[[:space:]]for[[:space:]]Generalizable[[:space:]]Robot[[:space:]]Manipulation/57fa886e-a784-460f-94a7-de5f04b8ab96_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Two[[:space:]]is[[:space:]]Better[[:space:]]than[[:space:]]One_[[:space:]]Efficient[[:space:]]Ensemble[[:space:]]Defense[[:space:]]for[[:space:]]Robust[[:space:]]and[[:space:]]Compact[[:space:]]Models/f8d31d9f-8dff-45eb-9d02-43167ab5a77a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Type-R_[[:space:]]Automatically[[:space:]]Retouching[[:space:]]Typos[[:space:]]for[[:space:]]Text-to-Image[[:space:]]Generation/6f176b19-a787-4b85-9ebe-081a3f4759be_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/U-Know-DiffPAN_[[:space:]]An[[:space:]]Uncertainty-aware[[:space:]]Knowledge[[:space:]]Distillation[[:space:]]Diffusion[[:space:]]Framework[[:space:]]with[[:space:]]Details[[:space:]]Enhancement[[:space:]]for[[:space:]]PAN-Sharpening/ca20f1f2-4ee9-4b83-80ca-6e2c71410bad_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/UA-Pose_[[:space:]]Uncertainty-Aware[[:space:]]6D[[:space:]]Object[[:space:]]Pose[[:space:]]Estimation[[:space:]]and[[:space:]]Online[[:space:]]Object[[:space:]]Completion[[:space:]]with[[:space:]]Partial[[:space:]]References/0d3743ff-3af5-445c-ae4c-ba5a33b9d7b4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/UCM-VeID[[:space:]]V2_[[:space:]]A[[:space:]]Richer[[:space:]]Dataset[[:space:]]and[[:space:]]A[[:space:]]Pre-training[[:space:]]Method[[:space:]]for[[:space:]]UAV[[:space:]]Cross-Modality[[:space:]]Vehicle[[:space:]]Re-Identification/9b7d21db-8ee1-4718-8ddf-98653daa1c9e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/UCOD-DPL_[[:space:]]Unsupervised[[:space:]]Camouflaged[[:space:]]Object[[:space:]]Detection[[:space:]]via[[:space:]]Dynamic[[:space:]]Pseudo-label[[:space:]]Learning/dec1b26d-4dd9-4c06-a461-c03d0e31f356_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/UHD-processer_[[:space:]]Unified[[:space:]]UHD[[:space:]]Image[[:space:]]Restoration[[:space:]]with[[:space:]]Progressive[[:space:]]Frequency[[:space:]]Learning[[:space:]]and[[:space:]]Degradation-aware[[:space:]]Prompts/2e2768b6-2887-4bd5-935c-d0b880ebe319_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/UIBDiffusion_[[:space:]]Universal[[:space:]]Imperceptible[[:space:]]Backdoor[[:space:]]Attack[[:space:]]for[[:space:]]Diffusion[[:space:]]Models/7f0574b2-e282-4c3e-a072-3d64d026983a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/UMFN_[[:space:]]Unified[[:space:]]Multi-Domain[[:space:]]Face[[:space:]]Normalization[[:space:]]for[[:space:]]Joint[[:space:]]Cross-domain[[:space:]]Prototype[[:space:]]Learning[[:space:]]and[[:space:]]Heterogeneous[[:space:]]Face[[:space:]]Recognition/3e3f9780-5e1d-4d68-aee1-68dcb88c5368_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/UMotion_[[:space:]]Uncertainty-driven[[:space:]]Human[[:space:]]Motion[[:space:]]Estimation[[:space:]]from[[:space:]]Inertial[[:space:]]and[[:space:]]Ultra-wideband[[:space:]]Units/20c112a8-f19a-4ffc-8e5b-a5d02b15caa1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/UNEM_[[:space:]]UNrolled[[:space:]]Generalized[[:space:]]EM[[:space:]]for[[:space:]]Transductive[[:space:]]Few-Shot[[:space:]]Learning/592279b8-3f9e-4b39-91e5-638d8677cb8a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/UNIALIGN_[[:space:]]Scaling[[:space:]]Multimodal[[:space:]]Alignment[[:space:]]within[[:space:]]One[[:space:]]Unified[[:space:]]Model/7e109366-fc0f-438c-9b91-2ba7f07e59e9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/UNIC-Adapter_[[:space:]]Unified[[:space:]]Image-instruction[[:space:]]Adapter[[:space:]]with[[:space:]]Multi-modal[[:space:]]Transformer[[:space:]]for[[:space:]]Image[[:space:]]Generation/409eb6c5-d571-4968-a329-2c1ebf51e321_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/UNICL-SAM_[[:space:]]Uncertainty-Driven[[:space:]]In-Context[[:space:]]Segmentation[[:space:]]with[[:space:]]Part[[:space:]]Prototype[[:space:]]Discovery/fc453b95-4cba-4bb9-95fb-6c8d34e4ccaa_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/UNOPose_[[:space:]]Unseen[[:space:]]Object[[:space:]]Pose[[:space:]]Estimation[[:space:]]with[[:space:]]an[[:space:]]Unposed[[:space:]]RGB-D[[:space:]]Reference[[:space:]]Image/94d3840c-f3fe-4b2b-9388-48ccee10c9d7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/UPME_[[:space:]]An[[:space:]]Unsupervised[[:space:]]Peer[[:space:]]Review[[:space:]]Framework[[:space:]]for[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Model[[:space:]]Evaluation/42fa34eb-b512-4300-b1c0-40e45c00118f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/URWKV_[[:space:]]Unified[[:space:]]RWKV[[:space:]]Model[[:space:]]with[[:space:]]Multi-state[[:space:]]Perspective[[:space:]]for[[:space:]]Low-light[[:space:]]Image[[:space:]]Restoration/e077664a-8838-4206-9246-0d60f57e032f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/USP-Gaussian_[[:space:]]Unifying[[:space:]]Spike-based[[:space:]]Image[[:space:]]Reconstruction,[[:space:]]Pose[[:space:]]Correction[[:space:]]and[[:space:]]Gaussian[[:space:]]Splatting/62c2f0f5-8ea0-4959-8e97-f59984bffdfe_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/UVGS_[[:space:]]Reimagining[[:space:]]Unstructured[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]using[[:space:]]UV[[:space:]]Mapping/77087a0b-d1cb-4230-a846-7760f7510d17_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/UWAV_[[:space:]]Uncertainty-weighted[[:space:]]Weakly-supervised[[:space:]]Audio-Visual[[:space:]]Video[[:space:]]Parsing/82f7769f-b767-4619-96ea-539823053afb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/UltraFusion_[[:space:]]Ultra[[:space:]]High[[:space:]]Dynamic[[:space:]]Imaging[[:space:]]using[[:space:]]Exposure[[:space:]]Fusion/c2c26ec1-c8af-49f8-87ba-b3c749531067_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/UnCommon[[:space:]]Objects[[:space:]]in[[:space:]]3D/4d2bd976-41fd-45e5-bfc6-8a4578f73725_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Unbiased[[:space:]]Video[[:space:]]Scene[[:space:]]Graph[[:space:]]Generation[[:space:]]via[[:space:]]Visual[[:space:]]and[[:space:]]Semantic[[:space:]]Dual[[:space:]]Debiasing/168f707b-a7fa-4e7c-93c1-aac7eef770eb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Unbiasing[[:space:]]through[[:space:]]Textual[[:space:]]Descriptions_[[:space:]]Mitigating[[:space:]]Representation[[:space:]]Bias[[:space:]]in[[:space:]]Video[[:space:]]Benchmarks/5edfcfce-ad23-41c6-9a43-d3cb9f8bcb90_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Unboxed_[[:space:]]Geometrically[[:space:]]and[[:space:]]Temporally[[:space:]]Consistent[[:space:]]Video[[:space:]]Outpainting/3a973275-71b1-47b0-9c32-b162722059fe_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Uncertain[[:space:]]Multimodal[[:space:]]Intention[[:space:]]and[[:space:]]Emotion[[:space:]]Understanding[[:space:]]in[[:space:]]the[[:space:]]Wild/50195041-da23-4b5c-8721-3858d2b20dc9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Uncertainty[[:space:]]Meets[[:space:]]Diversity_[[:space:]]A[[:space:]]Comprehensive[[:space:]]Active[[:space:]]Learning[[:space:]]Framework[[:space:]]for[[:space:]]Indoor[[:space:]]3D[[:space:]]Object[[:space:]]Detection/d2e21437-1e7e-4dc1-839a-f40a2bf9489e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Uncertainty[[:space:]]Weighted[[:space:]]Gradients[[:space:]]for[[:space:]]Model[[:space:]]Calibration/792a362b-85b5-4092-b750-5040ac391b64_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Uncertainty-Instructed[[:space:]]Structure[[:space:]]Injection[[:space:]]for[[:space:]]Generalizable[[:space:]]HD[[:space:]]Map[[:space:]]Construction/308073a9-0b23-46de-a5fd-816ea2a49dc7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Uncertainty-guided[[:space:]]Perturbation[[:space:]]for[[:space:]]Image[[:space:]]Super-Resolution[[:space:]]Diffusion[[:space:]]Model/10430da4-a80e-4106-a1b5-f1210ade542e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Understanding[[:space:]]Fine-tuning[[:space:]]CLIP[[:space:]]for[[:space:]]Open-vocabulary[[:space:]]Semantic[[:space:]]Segmentation[[:space:]]in[[:space:]]Hyperbolic[[:space:]]Space/4933d0fc-2175-4f82-8204-a2ef1c424606_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Understanding[[:space:]]Multi-Task[[:space:]]Activities[[:space:]]from[[:space:]]Single-Task[[:space:]]Videos/b633800a-1331-4862-80d6-5ccbeba54c12_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Understanding[[:space:]]Multi-layered[[:space:]]Transmission[[:space:]]Matrices/0c5698ec-c1a0-4efc-b1d6-15a1169745e5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Uni-Renderer_[[:space:]]Unifying[[:space:]]Rendering[[:space:]]and[[:space:]]Inverse[[:space:]]Rendering[[:space:]]Via[[:space:]]Dual[[:space:]]Stream[[:space:]]Diffusion/ef5cd413-b783-4606-bcb0-152603271134_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Uni4D_[[:space:]]Unifying[[:space:]]Visual[[:space:]]Foundation[[:space:]]Models[[:space:]]for[[:space:]]4D[[:space:]]Modeling[[:space:]]from[[:space:]]a[[:space:]]Single[[:space:]]Video/b0bfb21a-5579-4e48-8be6-e0a25d4d029f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/UniAP_[[:space:]]Unifying[[:space:]]Inter-[[:space:]]and[[:space:]]Intra-Layer[[:space:]]Automatic[[:space:]]Parallelism[[:space:]]by[[:space:]]Mixed[[:space:]]Integer[[:space:]]Quadratic[[:space:]]Programming/b1ddd4d6-5b56-44ba-82ed-6323538fa782_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2025/Towards Understanding and Quantifying Uncertainty for Text-to-Image Generation/c3ad3744-8ee2-4117-a4a1-7d23c58e3fe4_content_list.json b/2025/Towards Understanding and Quantifying Uncertainty for Text-to-Image Generation/c3ad3744-8ee2-4117-a4a1-7d23c58e3fe4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8abae98597941a385d4a3f687314528736747e03 --- /dev/null +++ b/2025/Towards Understanding and Quantifying Uncertainty for Text-to-Image Generation/c3ad3744-8ee2-4117-a4a1-7d23c58e3fe4_content_list.json @@ -0,0 +1,1544 @@ +[ + { + "type": "text", + "text": "Towards Understanding and Quantifying Uncertainty for Text-to-Image Generation", + "text_level": 1, + "bbox": [ + 228, + 130, + 772, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gianni Franchi, $^{1}$ * Nacim Belkhir, $^{2}$ * Dat Nguyen Trong, $^{1}$ Guoxuan Xia, $^{3}$ Andrea Pilzer $^{4}$ \nU2IS, ENSTA Paris, Institut Polytechnique de Paris, $^{1}$ Mirai, $^{2}$ Imperial College London, $^{3}$ NVIDIA $^{4}$", + "bbox": [ + 258, + 202, + 738, + 276 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0df14e7c0fe8aff5150a47ddde96930d88820f306268a239773e70e5602c43fa.jpg", + "image_caption": [ + "Figure 1. Examples of Applications for Uncertainty Quantification in Text-to-Image Generation. Text-to-image generation models may exhibit uncertainty, and that need to be quantified since it can provide insights into the model's training dataset, aiding in deepfake prevention, detecting model biases, and protecting copyrighted content from unauthorized generation." + ], + "image_footnote": [], + "bbox": [ + 114, + 308, + 377, + 597 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e4cb42dfba865827dd7ac22e034ddfa0a402644ec557b371ebb37dce40429cc8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 379, + 308, + 879, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 660, + 328, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Uncertainty quantification in text-to-image (T2I) generative models is crucial for understanding model behavior and improving output reliability. In this paper, we are the first to quantify and evaluate the uncertainty of T2I models with respect to the prompt. Alongside adapting existing approaches designed to measure uncertainty in the image space, we also introduce Prompt-based Uncertainty Estimation for T2I models (PUNC), a novel method leveraging Large Vision-Language Models (LVLMs) to better address uncertainties arising from the semantics of the prompt and generated images. PUNC utilizes a LVLM to caption a gen", + "bbox": [ + 88, + 694, + 485, + 861 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "erated image, and then compares the caption with the original prompt in the more semantically meaningful text space. PUNC also enables the disentanglement of both aleatoric and epistemic uncertainties via precision and recall, which image-space approaches are unable to do. Extensive experiments demonstrate that PUNC outperforms state-of-the-art uncertainty estimation techniques across various settings. Uncertainty quantification in text-to-image generation models can be used on various applications including bias detection, copyright protection, and OOD detection. We also introduce a comprehensive dataset of text prompts and generation pairs to foster further research in uncertainty quantification for generative models. Our findings illustrate that PUNC not only achieves competitive performance but also enables novel applications in evaluating", + "bbox": [ + 511, + 662, + 908, + 888 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* equal contribution\nresponding author - gianni.franchi@ensta-paris.fr", + "bbox": [ + 109, + 875, + 370, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "+", + "bbox": [ + 473, + 875, + 482, + 883 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8062", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and improving the trustworthiness of text-to-image models. The code is available at https://github.com/ENSTA-U2IS-AI/Uncertainty_diffusion", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 165, + 222, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generative models have revolutionized artificial intelligence, making significant strides in fields such as robotics [2], healthcare [28], and physics [33]. These models have elevated the performance of AI systems by enabling the generation of complex data representations. However, while generative models offer immense potential, they do not address all challenges, especially those related to robustness and uncertainty in Deep Neural Networks(DNNs)—issues that have become even more pronounced in generative settings.", + "bbox": [ + 89, + 190, + 482, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we focus on the uncertainty quantification in deep generative models, in particular, we are the first to investigate modern text-to-image (T2I) models where the input is a natural language prompt [8, 18, 48-50]. These models have surged in popularity in recent years due to their capacity to generate high-quality images conveniently in a user-guided fashion, finding usage in a wide array of real-world applications. We are the first to investigate and propose a dedicated approach for quantifying uncertainty in text-to-image generation, aiming to fill a critical gap in generative AI research. The task of quantifying uncertainty in these models is not only complex but also crucial. Understanding the uncertainty in generative models can be invaluable for applications such as Out-of-Distribution (OOD) detection, trustworthy AI, and decision-making in high-stakes domains. To the best of our knowledge, no existing work has comprehensively addressed the challenge of uncertainty quantification in text-to-image generative models. However, given the semantic nature of the uncertainty we aim to measure, our work is closely related to recent studies on text-to-image alignment [37, 52, 67].", + "bbox": [ + 89, + 342, + 482, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing research explores uncertainty estimation and OOD detection in the image/output space [4, 5, 23, 39]. For example, given a test image, can we determine whether or not it is OOD? On the other hand, the uncertainty of a generation with respect to conditioning (such as text) is underexplored. We argue that this is an important area to be addressed, as it represents a widespread real-world application of deep generative models. In this case, it remains an open question as to how should we define uncertainty, and given a definition, how should we quantify it? As illustrated in Fig. 1, focusing on uncertainty quantification in text-to-image generation could enable novel applications, allowing us to extract and interpret the knowledge embedded within models using language.", + "bbox": [ + 89, + 659, + 482, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To answer these questions, we first argue that uncertainties should be rooted in the semantics of the image", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and the prompt. We then adapt and evaluate a number of existing image-space approaches, as well as proposing a novel method for uncertainty quantification in T2I generative models, which we refer to as \"Prompt-based Uncertainty Estimation for T2I Generation\" (PUNC). Our approach leverages the growing capabilities of Large Vision-Language Models (LVLMs) to extract the semantics from generated images. LVLMs, which have been trained on vast amounts of text and visual data, can serve as powerful tools for interpreting the underlying meaning of text/image inputs and assessing how variations in phrasing or semantics impact the uncertainty of the generated outputs. Our hypothesis is that as LVLMs become increasingly adept at understanding complex relationships between text and images, they will allow us to better analyze and quantify uncertainty. Through this paper, we will demonstrate Prompt-based Uncertainty Estimation for T2I (PUNC) effectiveness in quantifying uncertainties in generative models. Showcasing its ability to decompose and interpret different types of uncertainties arising from the multimodal nature of the input, offering a new perspective on uncertainty quantification in generative AI.", + "bbox": [ + 511, + 90, + 903, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our contributions are as follows: (1) We introduce a new task aimed at quantifying uncertainty (with respect to the prompt) in text-to-image generation models, and we share a dataset of prompts to support future research on uncertainty in these models. (2) We both adapt existing image-space approaches to our tasks, as well as propose a novel straightforward method, PUNC, designed to better quantify semantic uncertainty in text-to-image generation. (3) Extensive experiments demonstrate that PUNC, despite its simplicity and computational efficiency, achieves competitive performance. (4) We highlight the practical utility of uncertainty quantification in text-to-image generation by showcasing additional applications where this approach could be beneficial.", + "bbox": [ + 511, + 422, + 903, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 648, + 653, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep learning models typically encounter two primary types of uncertainty: aleatoric and epistemic uncertainty [26]. Over the years, a variety of methods have been developed to quantify these uncertainties. Ensemble methods [16, 32, 34, 58, 61], for instance, have often achieved state-of-the-art performance by leveraging multiple models to assess prediction confidence. However, these methods tend to be computationally expensive due to the need to train and maintain several models. Another approach involves Bayesian methods [21], especially Bayesian Neural Networks (BNNs) [19, 27, 42, 57], which offer a theoretically grounded way to quantify uncertainty. Additionally, there are techniques aimed at directly predicting uncertainty within DNNs [10, 20, 30]. Despite their effectiveness, most of these methods are challenging to apply to text-to-image", + "bbox": [ + 511, + 674, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8063", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f22022d38cc577c40f1d07107d04862ce1c9139eb16d6a4dce45345b8c7c4a8d.jpg", + "image_caption": [ + "Generation/Image-space Uncertainty (existing work)", + "Figure 2. Diagram illustrating generation/image-space uncertainty which is considered in previous work and condition/prompt-space uncertainty which is investigated in our work." + ], + "image_footnote": [], + "bbox": [ + 93, + 102, + 480, + 205 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/61e51754cec18673e6e57272b130161e06703729facb467446902ddcc7dc5863.jpg", + "image_caption": [ + "Figure 3. Three generations from PixArt- $\\Sigma$ [7] illustrating uncertainty with regards to prompt semantics. For the Normal image, we used an ImageNet-inspired prompt; for the corrupted image, additional corruption was applied to the prompt (e.g. fish was perturbed to fis) increasing the aleatoric uncertainty; and for the out-of-distribution (OOD) case, the model was prompted to generate an image of the Prime Minister of Japan Kishida Fumio under similar conditions, where epistemic uncertainty was injected in the form of a semantic concept the model was not familiar with." + ], + "image_footnote": [], + "bbox": [ + 114, + 290, + 460, + 386 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "generation, as they rely on ensembling predictions or accessing model uncertainty in ways that are difficult to implement for generative models without explicit access to intermediate predictions.", + "bbox": [ + 89, + 559, + 483, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generative models, as a subset of deep learning, also exhibit uncertainty. In the past, traditional generative models were employed for uncertainty quantification and used to help estimate the uncertainty within standard DNNs [6, 38, 56, 64]. With the recent popularity of diffusion models, research has begun exploring whether image-based diffusion models can detect out-of-distribution (OOD) samples [4, 23, 39] or quantify their own uncertainty [5, 6, 14]. Whilst these approaches align with our goals, they are limited to image-space tasks, where the input is an image rather than a textual prompt. This input difference fundamentally alters the nature of uncertainty in the model. Our approach, in contrast, specifically addresses text-to-image generation, a setting where prompt ambiguity and multimodal complexity introduce new dimensions of uncertainty. We demonstrate that quantifying uncertainty in this context can lead to valuable insights and applications across various downstream tasks.", + "bbox": [ + 89, + 628, + 483, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Uncertainty in Image Generation Models", + "text_level": 1, + "bbox": [ + 511, + 89, + 880, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider a deep generative model $p_{\\theta}(\\boldsymbol{x})$ with parameters $\\theta$ that models the data distribution $p_{\\mathrm{data}}(\\boldsymbol{x})$ of $\\boldsymbol{x} \\in \\mathbb{R}^D$ . We may be interested in measuring the uncertainty of our model $p_{\\theta}$ with respect to some test datum $\\boldsymbol{x}^*$ . For example, consider the case where we want to determine whether or not $\\boldsymbol{x}^*$ was drawn from $p_{\\mathrm{data}}(\\boldsymbol{x})$ , i.e. out-of-distribution (OOD) detection. This can be achieved by quantifying the uncertainty of $p_{\\theta}$ that arises from a lack of knowledge of $\\boldsymbol{x}^*$ , i.e. epistemic uncertainty [26]. A range of existing work that leverages diffusion-based generative models for OOD image detection [23, 39] fall into the above paradigm. From hereon we will refer to this as generation/image-space uncertainty. This is illustrated in Fig. 2 (top).", + "bbox": [ + 511, + 114, + 906, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, however, we are primarily interested in a different scenario: text-to-image generation. Consider instead a conditional model $p_{\\theta}(\\pmb{x}|\\pmb{c})$ that models the conditional data distribution $p_{\\mathrm{data}}(\\pmb{x}|\\pmb{c})$ . This describes real-world text-to-image generation, where image $\\pmb{x}$ is drawn from $p_{\\theta}(\\pmb{x}|\\pmb{c}^*)$ conditioned on some test text prompt $\\pmb{c}^*$ . In this case, we wish to quantify the uncertainty of $p_{\\theta}$ not with respect to some test $\\pmb{x}^*$ (e.g. image) but instead with respect to a test condition $\\pmb{c}^*$ . We will hereon refer to this as condition/prompt-space uncertainty. This is illustrated in Fig. 2 (bottom).", + "bbox": [ + 511, + 311, + 905, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We argue that in this case, uncertainties should be concerned with semantics, as intuitively, these are what inherently link the user's intentions, represented by the text prompt, and the generated output image. This is similar to recent discourse on uncertainty estimation for large language models [31]. If we consider the semantic concepts in the prompt and generated image, then we can define useful concepts of uncertainty (illustrated in Fig. 3):", + "bbox": [ + 511, + 478, + 905, + 598 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Aleatoric uncertainty is irreducible uncertainty in the data distribution $p_{\\mathrm{data}}(\\boldsymbol{x}|\\boldsymbol{c})$ . From the perspective of semantics, high aleatoric uncertainty would be where a variation of generated concepts may arise from a single prompt. For example, a spelling mistake where fish is mistyped as fis in the prompt may result in either a \"fish\" or a \"fist\" appearing in the image.", + "- Epistemic uncertainty should correspond to a model's lack of knowledge of semantic concepts found in a prompt. For example, a model trained on ImageNet will not know what the prime minister of Japan Kishida Fumio looks like, and thus will have high epistemic uncertainty for this semantic concept." + ], + "bbox": [ + 511, + 598, + 905, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We note that broadly speaking uncertainty estimation for deep conditional models is well explored [20-22, 24, 26, 32, 34, 41, 59-61, 63, 66]. However, research in this area is generally concerned with problem settings where 1) the", + "bbox": [ + 511, + 795, + 905, + 854 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1We note that although there is a discussion over how to define and quantify epistemic uncertainty [21, 26, 41, 45, 60], in this work, we stick to an intuitive albeit informal view based on a deep model's knowledge.", + "bbox": [ + 511, + 862, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8064", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "models explicitly evaluate $p_{\\theta}(\\pmb{x}|c)$ and 2) downstream applications of uncertainty relate directly to the numerical values of $\\pmb{x}$ and $p_{\\theta}(\\pmb{x}|c)$ . For example, a cross-entropy-trained image classifier explicitly predicts the probability mass $P_{\\theta}(\\pmb{x}|c^{*})$ of label $\\pmb{x}$ given test image $c^{*}$ . $P_{\\theta}(\\pmb{x}|c^{*})$ directly relates to the likelihood of classification error – which is an event that we may want to detect using an uncertainty measure, i.e. downstream misclassification detection [22, 24]. Thus, we are able to easily extract useful measures of uncertainty from the model $P_{\\theta}$ , e.g. the maximum of the softmax.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "On the other hand, text-to-image generation models are not as clear-cut. They may not explicitly evaluate $p_{\\theta}(\\pmb{x}|c^{*})$ (e.g. diffusion-based models [25, 54]), and even when they do (e.g. vector-quantised autoregressive models [17, 55]) it is hard to directly relate (latent) pixel values directly to our above definitions of uncertainty. For example, simple variations in contrast and brightness that may leave semantics completely unchanged could result in a very high variance in $p_{\\theta}(\\pmb{x}|c^{*})$ . To give a simple example, inverting the black and white stripes on a zebra would lead to a large Euclidean distance in the image space whilst preserving the semantics of the image. In the next section, we present and discuss a range of potential methods to quantify the uncertainty of text-to-image models with respect to a given prompt.", + "bbox": [ + 89, + 243, + 482, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Techniques to Quantify Uncertainty", + "text_level": 1, + "bbox": [ + 89, + 469, + 413, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We quantify text-to-image uncertainty by adapting existing methods and proposing a novel approach named Prompt-based UNCertainty estimation (PUNC)(Fig. 4).", + "bbox": [ + 89, + 496, + 482, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Adapting Existing Approaches", + "text_level": 1, + "bbox": [ + 89, + 551, + 362, + 569 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There are a number of existing methods for quantifying the image-space uncertainty of diffusion models. As diffusion-based models represent a majority of current text-to-image foundation models [8, 18, 48, 50], we choose to adapt these in order to quantify prompt-space uncertainty (Fig. 4). We first present a brief primer on diffusion models to the reader.", + "bbox": [ + 89, + 575, + 482, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1.1. Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 676, + 256, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion models [25] generate data by iteratively denoising pure noise. The forward process over time $t \\in [0,1]$ for $\\boldsymbol{x}_t \\in \\mathbb{R}^D$ , conditioned on a text prompt $c$ , follows:", + "bbox": [ + 89, + 695, + 482, + 742 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}, \\boldsymbol {c}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t}; a _ {t} \\boldsymbol {x} _ {0}, \\sigma_ {t} ^ {2} \\boldsymbol {I}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 753, + 482, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{x}_0$ is sampled from an unknown data distribution $p(\\pmb{x}_0)$ , and $a_{t},\\sigma_{t}$ define the noise schedule. This formulation allows diffusion models to progressively refine noisy inputs, enabling high-quality data generation through a learned denoising process. By leveraging stochastic differential equations (SDEs) or their deterministic counterparts,", + "bbox": [ + 89, + 784, + 482, + 875 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "these models sample from complex distributions in a structured manner. Various solvers [29, 40, 53, 54, 62, 65, 69] can then be employed to efficiently generate realistic samples with a limited number of denoising steps.", + "bbox": [ + 511, + 90, + 903, + 152 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Latent diffusion models (LDMs) [39] perform this process within a compressed latent space rather than directly on pixel space. By first encoding images into a more compact representation, LDMs reduce computational demands, enabling high-quality high-resolution generations. We will not consider the latent representation to keep the notation simple, even though experiments will use LDMs.", + "bbox": [ + 511, + 160, + 906, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1.2. Time-step based approaches", + "text_level": 1, + "bbox": [ + 511, + 299, + 756, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "DDPM-OOD [23]. This approach, developed for image-space OOD detection, involves noising a test sample $\\boldsymbol{x}_0^*$ to different points in the forward process to get $\\{\\boldsymbol{x}_{t_1}^*,\\boldsymbol{x}_{t_2}^*,\\dots \\}$ and denoising them, generating multiple reconstructions $\\{\\hat{x}_{0_1},\\hat{x}_{0_2},\\ldots \\}$ . Similarity is then measured between each reconstruction and the original image. The idea is that test samples $\\boldsymbol{x}_0^*$ for which the model has higher epistemic uncertainty will lead to reconstructions that differ from the test sample. However, in our case, we do not have a test image $\\boldsymbol{x}_0^*$ . In order to adapt DDPM-OOD we propose measuring similarity between reconstructions at different $t$ and a baseline generation, $S(\\hat{x}_{0_i},\\hat{x}_0)$ , where $\\hat{x}_0$ represents an image directly generated using the original test prompt $c^*$ alone. Both aleatoric uncertainty (e.g. ambiguous prompt) and epistemic uncertainty (e.g. unfamiliar concepts in prompt) could lead to more divergent reconstructions.", + "bbox": [ + 511, + 325, + 906, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LMD [39]. Another similar approach for image-space OOD detection, known as Learned Manifold Denoising (LMD), instead explores alternative transforms on $\\boldsymbol{x}_0^*$ to noising. The idea is to lift $\\boldsymbol{x}_0^*$ from its original data manifold and then map it back using the trained diffusion model. If the sample is in-distribution, the mapping process should return it close to its original position on the data manifold. Conversely, if it is OOD, the mapped position will likely diverge significantly from the original. The corruption process typically involves masking portions of the image, with various mask patterns applied. Similar to DDPM-OOD, similarity is measured between the original image and the reconstructed version after applying the diffusion process. Similar to before, since we lack a testimage, we measure the similarity between reconstructed images and $\\hat{\\boldsymbol{x}}_{0_i}$ , the image generated directly from the prompt without perturbations.", + "bbox": [ + 511, + 569, + 908, + 810 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarity Metric. Various methods are used to evaluate similarity between images. Mean-squared error (MSE) is commonly applied between input and reconstruction, while some approaches favor the Learned Perceptual Image Patch Similarity (LPIPS) [70] metric. The choice of similarity metric is a crucial component of these techniques.", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2We stick with this non-standard notation for the sake of consistency.", + "bbox": [ + 107, + 886, + 473, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8065", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/028fd6b0c4c53b8460bc355b984d6d2584ea5339d9111611314b3422a07d66e0.jpg", + "image_caption": [ + "Figure 4. Illustration showing the different baselines and PUNC. PUNC leverages a LVLM to describe generated images and assess similarity with the original prompt, providing a refined uncertainty score. In contrast, baseline methods employ traditional techniques such as noise injection, ensembling, or masking to quantify uncertainty, followed by image-based similarity scoring." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 906, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.3. Test-Time Ensembling Approaches", + "text_level": 1, + "bbox": [ + 89, + 314, + 382, + 329 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Test-time ensembling techniques take inspiration from approaches used in the large language model (LLM) community [31], where multiple generations are produced to assess consistency. Given a prompt $c$ , we generate two images $\\{\\hat{x}_1, \\hat{x}_2\\}$ and measure the similarity between the two outputs. We denote this technique 2XDM. The underlying rationale is that if the model has high confidence, it will generate similar images. However, if uncertainty is high, it will generate substantially different images. We note that this is similar to the approach proposed in [5], however, that approach uses an ensemble of diffusion models, which is not practical in the setting where we want to quantify the uncertainty of a single pretrained T2I model.", + "bbox": [ + 88, + 332, + 483, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A key question remains regarding how to find the best measure similarity, and we propose using the same techniques applied in the time-step-based methods. A significant advantage of test-time ensembling approaches is that they can be applied generically to all T2I models. In contrast, timestep-based approaches require a multi-step generation process, limiting the broadness of their applicability. We note that all above approaches require multiple generations, incurring a notable computational overhead at inference time. For example, DDPM-OOD requires $\\sim 50\\times$ the number of network inferences as a single generation (if the hyperparameters of the original paper are used) [23].", + "bbox": [ + 88, + 530, + 483, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Prompt-based Uncertainty Estimation", + "text_level": 1, + "bbox": [ + 89, + 720, + 419, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we propose a new method, Prompt-based Uncertainty estimation for text-to-image generation (PUNC), which leverages Large Vision-Language Models (LVLM) in conjunction with T2I models. We aim to assess the uncertainty in generated images by analyzing the alignment between initial text prompts and reconstructed textual descriptions of the generated images.", + "bbox": [ + 89, + 744, + 483, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Background on Large Vision-Language Model (LVLM)", + "text_level": 1, + "bbox": [ + 89, + 869, + 482, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A LVLM integrates a large language model (LLM) with", + "bbox": [ + 89, + 885, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "an image encoder, allowing for robust, multimodal understanding. Given a text prompt $c$ and an image $x$ , a LVLM model, with parameters $\\omega$ , includes:", + "bbox": [ + 511, + 314, + 903, + 358 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Image Encoder: $f_{\\omega}^{\\mathrm{img}}(\\cdot)$ processes the input image $x$ and produces an embedding: $z^{\\mathrm{img}} = f_{\\omega}^{\\mathrm{img}}(x)$", + "- Text Processor with LLM: $f_{\\omega}^{\\mathrm{txt}}(\\cdot, \\cdot)$ takes the prompt $c$ and the image embedding $z^{\\mathrm{img}}$ as inputs, generating a descriptive or answer-based response: $\\hat{c} = f_{\\omega}^{\\mathrm{txt}}(c, z^{\\mathrm{img}})$" + ], + "bbox": [ + 511, + 359, + 903, + 438 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thus, a LVLM system generates an interpretation $\\hat{c}$ that reflects the content of the image given an initial prompt.", + "bbox": [ + 511, + 440, + 903, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prompt-Based Uncertainty Estimation for T2I Generation (PUNC) The previously introduced approaches to uncertainty estimation rely on measuring similarity between images. For a given text prompt $c$ , a T2I model generates an image $x \\sim p_{\\theta}(x|c)$ . We reason that, even under low uncertainty, the model can produce images that vary in aspects such as color, texture, or object positioning, while still faithfully representing the semantics of prompt $c$ . As previously discussed in Sec. 3, this may result in image-based similarity metrics failing to capture uncertainty at a semantic level, since large differences in the image space may not necessarily correspond to meaningful semantic differences.", + "bbox": [ + 511, + 488, + 905, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose that advanced LVLMs, such as Molmo [11], LLAMA 3 [15], and GPT-4 [1], are a strong candidate for extracting the semantics from an image, through describing an image via natural language. Building on this insight, our proposed method, PUNC, introduces a novel approach that bypasses direct image comparison and instead evaluates the alignment between generated images and the original prompts by calculating an uncertainty score in the more semantically meaningful text-space. The steps for implementing PUNC are as follows:", + "bbox": [ + 511, + 669, + 905, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Step 1: Initial Prompt and Image Generation Given a test prompt $c^*$ , we use the T2I model to sample an image $x$ :", + "bbox": [ + 511, + 820, + 903, + 849 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} \\sim p _ {\\boldsymbol {\\theta}} (\\boldsymbol {x} | \\boldsymbol {c} ^ {*})\n$$\n", + "text_format": "latex", + "bbox": [ + 658, + 859, + 758, + 876 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Step 2: Image Interpretation via LVLM With the", + "bbox": [ + 511, + 885, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8066", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bfd49245a54fb630faa3175b2d199995fca0356c85229b1aa37a4ac1798f6ca0.jpg", + "image_caption": [ + "Figure 5. Disentangling Aleatoric and Epistemic Uncertainty." + ], + "image_footnote": [], + "bbox": [ + 127, + 87, + 446, + 208 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "generated image $\\pmb{x}$ and the initial prompt $c^*$ , the LVLM produces a new descriptive caption $\\hat{c}$ :", + "bbox": [ + 89, + 258, + 482, + 291 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {c}} = f _ {\\omega} ^ {\\mathrm {t x t}} \\left(\\boldsymbol {c} ^ {*}, f _ {\\omega} ^ {\\mathrm {i m g}} (\\boldsymbol {x})\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 301, + 359, + 319 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Step 3: Uncertainty Score Calculation The core of PUNC lies in measuring the alignment between $\\mathbf{c}$ and $\\hat{\\mathbf{c}}$ to assess the diffusion model's confidence. Specifically, the similarity score $S(\\mathbf{c},\\hat{\\mathbf{c}})$ is computed as:", + "bbox": [ + 89, + 330, + 483, + 391 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nS (\\boldsymbol {c} ^ {*}, \\hat {\\boldsymbol {c}}) = \\operatorname {s i m} (\\boldsymbol {c} ^ {*}, \\hat {\\boldsymbol {c}})\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 402, + 362, + 420 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "High similarity suggests low uncertainty, indicating the generated image closely matches the initial prompt. Conversely, a lower score reveals higher uncertainty, signifying potential divergence from the prompt's intended meaning.", + "bbox": [ + 89, + 431, + 482, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Thus, PUNC is able to predict semantic uncertainties by utilising the powerful image understanding ability of LVLMs and computing similarities in the text space.", + "bbox": [ + 89, + 492, + 483, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prompt Similarity: Aleatoric and Epistemic via Precision and Recall. In Sec. 3 we discussed the different types of semantic uncertainty for T2I generation. Here we will present a method for disentangling these uncertainties using the precision and recall of the caption given the prompt. Consider an abstract scenario where the prompt and the generated image each contain a discrete set of semantic concepts. We can then calculate the precision and recall of the image concepts with respect to the prompt:", + "bbox": [ + 89, + 537, + 482, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Intuitively, a lack of knowledge about concepts in the prompt (epistemic uncertainty) will result in fewer concepts being preserved in the image, reducing recall. On the other hand, a lack of specificity in the prompt (aleatoric uncertainty) will result in additional concepts being generated, reducing precision. Thus we can use this abstract form of semantic-concept-based recall to measure epistemic uncertainty, and use precision to measure aleatoric uncertainty.", + "bbox": [ + 89, + 674, + 482, + 795 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Although we cannot directly measure concepts, we can approximate them using natural language. Moreover, existing similarity scores for text already have the above notions of precision and recall integrated within [36, 43, 72], allowing PUNC to disentangle uncertainty by comparing the caption of the generated image with the original prompt. We note that the image-based approaches in", + "bbox": [ + 89, + 795, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sec. 4.1 are unable to disentangle aleatoric from epistemic uncertainty. We focus on two different existing text-based similarity metrics, ROUGE [36], and BERTScore [72]. The former calculates similarity (precision or recall) by directly matching (sequences of) words between two text sequences. The latter on the other hand measures similarity in the deep feature space of a text encoder.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As illustrated in Fig. 5, we define Prompt Similarity as the semantic overlap between the input prompt and the generated descriptive caption. More formally, let $P_{\\mathrm{GT}}$ denote the set of semantic concepts in the ground-truth prompt used to generate an image, and $P_{\\mathrm{VQA}}$ the set of concepts extracted from the LVLM-generated caption. The resulting uncertainties can be characterized as follows:", + "bbox": [ + 511, + 196, + 906, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Epistemic Uncertainty, which captures missing concepts in $P_{\\mathrm{VQA}}$ that are present in $P_{\\mathrm{GT}}$ , reflects an incomplete understanding of the intended semantics. It is defined as:", + "bbox": [ + 511, + 301, + 903, + 348 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nU _ {\\text {e p i s t e m i c}} = 1 - \\frac {\\left| P _ {\\mathrm {G T}} \\cap P _ {\\mathrm {V Q A}} \\right|}{\\left| P _ {\\mathrm {G T}} \\right|} = 1 - \\text {R e c a l l} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 358, + 903, + 391 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Aleatoric Uncertainty, on the other hand, arises from spurious or extra concepts in $P_{\\mathrm{VQA}}$ that do not appear in $P_{\\mathrm{GT}}$ , indicating ambiguity in the model's response. It is defined as:", + "bbox": [ + 511, + 402, + 903, + 460 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nU _ {\\text {a l e a t o r i c}} = 1 - \\frac {\\left| P _ {\\mathrm {G T}} \\cap P _ {\\mathrm {V Q A}} \\right|}{\\left| P _ {\\mathrm {V Q A}} \\right|} = 1 - \\text {P r e c i s i o n} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 470, + 903, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Further details on the computation of precision and recall can be found in Appendix A.", + "bbox": [ + 511, + 513, + 906, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments & Discussions", + "text_level": 1, + "bbox": [ + 511, + 558, + 766, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We test PUNC on Various task and various dataset and compared PUNC to DDPM-OOD [23], LMD [39] and 2XDM. We first introduce on Sec. 5 how we have constitute the dataset of prompts. Then on Sec. 5.1 we discuss of different results on these dataset. Then Appendix F explains how we can use the uncertainty of text-to-image model to detect if a model has been trained with some concept related to some fictional character of politicians. Appendix F.3 shows how uncertainty can help to detect bias of T2I models. To conduct our experiments and evaluate the various techniques for quantifying uncertainty, we utilized four T2I models: Stable Diffusion 1.5 (SDv1.5) [50], SDXS [68], SDXL [48], and PixArt- $\\Sigma$ [7]. These diverse T2I models allow us to test a range of architectures, each with unique characteristics and potential weaknesses. SDv1.5, being one of the earliest models, likely has a smaller and less diverse training dataset compared to the others. In contrast, SDXS generates in a single-step, different to the multi-step processes in other models.", + "bbox": [ + 511, + 583, + 906, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset of prompts To conduct our experiments, we require both an in-distribution dataset of prompts and", + "bbox": [ + 511, + 871, + 903, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8067", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "datasets with uncertainty. For the in-distribution prompts, we use the dataset proposed by [13], where high-quality descriptions of ImageNet [12] images are generated using GPT-4[13], referred to as Normal. To simulate epistemic uncertainty, we create out-of-distribution (OOD) datasets using images from underrepresented domains like remote sensing, texture, and microscopic datasets, captioned with LLAVA Next[35], labeled as Remote Sensing, Texture, and Microscopic. For aleatoric uncertainty, we generate two datasets: Vague, containing 2,000 minimally descriptive prompts (e.g., \"An image of ***\"), and Corrupted, derived from Normal by adding grammatical errors (Level 1) and removing $50\\%$ of words (Level 2). A detailed summary of these datasets is presented in Tab. 1 and Appendix E. In Appendix C, we provide the details of the hyperparameters used in our experiments.", + "bbox": [ + 89, + 90, + 480, + 333 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/171fe859994b014f2125487c611fc0c28b4bf4d771eef1699d8169708a968c43.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset TypeDataset NameDescription
In-Distribution (ID)Normal20k High-quality detailed prompts from ImageNet [12] images using GPT-4[13].
Out-of-Distribution (OOD)Remote Sensing Texture Microscopic20k Prompts from remote sensing images [47]5k Prompts from texture datasets10k Prompts from microscopic images [44]
Aleatoric UncertaintyVagueAdversarialCorrupted2k prompts with minimal context, e.g., "An image of ***," where "****" is an ImageNet class name, creating ambiguous input.1k altered prompts from Normal using Unlearn-DiffAtk [73]20k Altered prompts from Normal:• Level 1: Grammatical mistakes and spelling errors.• Level 2: Random removal of 50% of words in Level 1 prompts.
", + "bbox": [ + 130, + 345, + 444, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Summary of prompt datasets for unc. quantification.", + "bbox": [ + 104, + 486, + 467, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Results for Uncertainty Estimation", + "text_level": 1, + "bbox": [ + 89, + 527, + 393, + 544 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 3 presents the results of PUNC when using Molmo as the LVLM to assess epistemic uncertainty in the T2I models. Our approach, PUNC, outperforms most state-of-the-art techniques on average. However, we observe more mixed results on the Texture dataset, likely due to uncertainty about the presence of texture images in the training data of the T2I models.", + "bbox": [ + 89, + 550, + 482, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 4, we show the results of PUNC under aleatoric uncertainty, again using Molmo as the LVLM. Our approach outperforms most state-of-the-art methods in this setting as well. Notably, while the ROUGE score tends to perform better for epistemic uncertainty, the optimal metric is less clear for aleatoric uncertainty. Here, the choice between ROUGE and BERTScore may depend on the specific application: because corruption introduces spelling and grammatical errors, BERTScore—which captures semantic similarity—may be more suitable. For additional results using different LVLMs, please refer to Appendix D and Appendix B. In Appendix G, we present qualitative results.", + "bbox": [ + 89, + 656, + 482, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Applications of uncertainty", + "text_level": 1, + "bbox": [ + 89, + 848, + 336, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our study explores multiple applications of uncertainty quantification in T2I models, with detailed results provided", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in Appendix F. First, we investigate deepfake detection by examining the model's ability to generate recognizable images of prominent politicians (see Appendix F.1). We examined whether diffusion models could generate convincing images of G7 politicians. Using 100 prompts containing their names, we generated images with four diffusion models and queried a LVLM to determine if the intended politician was present in each image (yes/no), thus measuring concept recognition uncertainty. Results are presented in Table 2.", + "bbox": [ + 511, + 90, + 903, + 241 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we address copyright concerns by evaluating whether T2I models can accurately replicate iconic, copyrighted characters such as Mickey Mouse and Darth Vader (see Appendix F.2). Finally, we analyze bias in T2I models by testing their capacity to generate individuals across different genders and races for specific job roles (see Appendix F.3).", + "bbox": [ + 511, + 242, + 903, + 347 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Using a LVLM, we evaluate the presence of specific individuals, copyrighted characters, or demographic concepts in the generated images, demonstrating how uncertainty estimation can aid in detecting potential misuses of generative models. These experiments illustrate the value of text-based uncertainty quantification in analyzing T2I model performance across ethically and legally sensitive applications, enabling assessments that are challenging to achieve with image-based generation alone.", + "bbox": [ + 511, + 348, + 906, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 497, + 633, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this work, we set out to be first to investigate uncertainty estimation T2I generation, where the uncertainty is with respect to the text prompt. We both adapt existing image-space uncertainty approaches, as well as present Prompt-based UNCertainty Estimation for T2I models (PUNC), the first approach designed specifically to quantify uncertainty in T2I generation. By leveraging the interpretive power of Large Vision-Language Models to extract semantics from a generated image, PUNC addresses the challenges associated with measuring multimodal uncertainty, enabling a deeper understanding of model behavior with respect to prompt-based conditioning. Our results validate that PUNC provides reliable, fine-grained assessments of aleatoric and epistemic uncertainties, outperforming existing methods that work for image generation.", + "bbox": [ + 511, + 523, + 903, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "PUNC opens avenues for new applications, such as improving model robustness in biased or OOD scenarios and safeguarding against unauthorized content generation, including deepfakes and copyrighted characters. Our experiments show that diffusion models, while powerful, still exhibit significant variances in their ability to adhere to prompt details, particularly with ambiguous, corrupt prompts or OOD prompts. By identifying such uncertainties, PUNC not only supports the development of more trustworthy generative AI but also aids in aligning these", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8068", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b80375c0c9bb1ca15cccb13e9d02894a56d3ece7d1a15a401578a1fdb70a4e62.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Justin TrudeauKing CharlesFumio KishidaOlaf ScholzEmmanuel MacronJoe BidenGiorgia MeloniAVERAGE
SDXL9494888898962195.5
SDv1.57860628894801878
SDXS8040708294762872.5
PixArt3676627894901574
AVERAGE7267.570.5849585.520.570.71
", + "bbox": [ + 215, + 88, + 781, + 151 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b660b9d8dd8679298c7986b79d8b083a83775bebe3e10881d71011565a1153b1.jpg", + "table_caption": [ + "Table 2. Accuracy (%) of diffusion models in generating recognizable images of G7 politicians. Higher accuracy indicates lower uncertainty in the model's ability to represent specific individuals." + ], + "table_footnote": [], + "table_body": "
auroc ↑SDXS aupr ↑fpr95↓auroc ↑PixArt aupr ↑fpr95↓auroc ↑SDv1.5 aupr ↑fpr95↓auroc ↑SDXL aupr ↑fpr95↓auroc ↑Average aupr ↑fpr95↓
microscopicDDPM-OOD mse_and_lpips39.61%56.38%100.00%71.64%71.52%93.62%87.25%83.67%51.43%66.17%70.53%81.68%
LMD lpips alex45.82%45.24%99.10%16.81%31.61%99.88%5.75%28.12%100.00%22.79%34.99%99.66%
2XDM lpips alex24.49%32.60%99.91%24.32%33.21%99.99%16.19%32.42%99.94%30.58%36.99%98.25%23.90%33.81%99.52%
PUNC ROUGE (1_recall)84.00%85.10%57.07%95.47%97.06%30.84%87.24%90.66%48.90%83.56%84.31%52.78%87.57%89.28%47.40%
PUNC ROUGE (L_recall)87.64%89.56%55.28%96.96%98.02%16.85%90.20%93.40%43.56%88.19%89.84%47.44%90.75%92.71%40.78%
PUNC BERT (recall)88.45%89.71%48.68%97.83%98.41%10.32%87.75%91.70%51.03%83.10%85.46%58.14%89.28%91.32%42.04%
remote sensingDDPM-OOD mse_and_lpips80.00%85.37%99.28%74.36%71.89%71.00%65.47%58.25%71.97%73.28%71.84%80.75%
LMD lpips alex90.65%93.11%74.49%78.93%79.31%81.05%86.60%85.39%49.19%85.39%85.94%68.24%
2XDM lpips alex53.88%59.21%98.23%54.43%57.99%98.78%30.66%38.80%98.76%23.54%35.35%96.23%40.63%47.83%98.00%
PUNC ROUGE (1_recall)85.59%86.59%63.87%94.14%95.97%49.29%73.24%72.93%79.56%73.07%74.54%83.48%81.51%82.51%69.05%
PUNC ROUGE (L_recall)89.85%90.63%51.96%97.01%97.94%16.08%83.63%84.65%65.35%85.68%87.09%64.00%89.04%90.08%49.34%
PUNC BERT (recall)80.50%82.56%75.19%91.93%93.08%43.30%61.27%64.74%90.08%60.71%65.32%90.94%73.60%76.42%74.88%
TextureDDPM-OOD mse_and_lpips73.22%62.02%89.24%70.21%55.97%99.80%74.24%51.97%84.71%72.56%56.65%91.25%
LMD lpips alex63.40%50.47%97.52%26.74%16.06%99.89%35.66%23.22%99.95%41.93%29.92%99.12%
2XDM lpips alex32.80%20.84%99.98%37.21%23.36%99.37%29.58%24.27%99.93%28.87%16.49%98.88%32.11%21.24%99.54%
PUNC ROUGE (1_recall)40.77%72.88%95.54%83.26%94.80%73.48%46.61%75.64%90.87%49.16%78.04%93.57%54.95%80.34%88.36%
PUNC ROUGE (L_recall)45.64%76.90%95.99%84.10%95.13%74.21%47.60%77.70%93.14%50.71%80.05%94.35%57.01%82.44%89.42%
PUNC BERT (recall)49.11%79.11%94.65%84.40%95.14%69.99%47.60%78.80%93.85%46.72%79.20%96.87%56.96%83.06%88.84%
", + "bbox": [ + 135, + 205, + 864, + 393 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4433cec5bf99037f5c9ec0bab56b699020f6329088988d04ac39c4039ab20c8f.jpg", + "table_caption": [ + "Table 3. Performance comparison of PUNC applied to different Out of Distribution concepts and different text-to-image models" + ], + "table_footnote": [], + "table_body": "
auroc ↑SDXS aupr ↑fpr95↓auroc ↑PixArt aupr ↑fpr95↓auroc ↑SDv1.5 aupr ↑fpr95↓auroc ↑SDXL aupr ↑fpr95↓auroc ↑Average aupr ↑fpr95↓
VagueDDPM-OOD mse.and.Lpips31.41%6.06%96.45%50.06%9.30%96.52%40.59%5.36%97.37%40.69%6.91%96.78%
LMD lipps alex39.26%6.91%96.35%47.62%9.54%97.70%40.77%5.70%97.95%42.55%7.38%97.33%
2XDM lipps alex51.28%10.10%94.35%43.82%8.57%96.89%50.97%10.09%95.96%57.83%10.01%94.37%50.98%9.69%95.39%
PUNC ROUGE (1_precision)99.99%100.00%0.00%100.00%100.00%0.00%99.99%100.00%0.00%99.99%100.00%0.00%99.99%100.00%0.00%
PUNC ROUGE (L_precision)99.99%100.00%0.00%100.00%100.00%0.00%99.99%100.00%0.00%100.00%100.00%0.00%100.00%100.00%0.00%
PUNC BERT (precision)60.63%92.24%82.85%67.02%94.54%83.10%37.35%86.51%92.70%70.85%95.13%78.00%58.96%92.11%84.16%
AdversarialDDPM-OOD mse.and.Lpips39.54%1.78%95.72%50.89%2.54%95.91%38.97%1.10%98.38%43.13%1.81%96.67%
LMD lipps alex43.71%1.95%96.24%53.29%2.89%94.43%41.64%1.17%97.48%46.21%2.00%96.05%
2XDM lipps alex36.94%1.94%98.92%47.95%2.54%96.03%53.04%2.83%95.69%55.19%1.96%96.20%48.28%2.32%96.71%
PUNC ROUGE (1_precision)64.14%97.77%48.80%76.32%98.90%48.00%63.59%97.73%49.50%62.97%97.64%49.20%66.76%98.01%48.87%
PUNC ROUGE (L_precision)63.20%97.75%49.00%76.26%98.90%48.40%64.09%97.90%49.70%63.76%97.87%49.60%66.83%98.11%49.17%
PUNC BERT (precision)73.98%98.72%78.31%72.95%98.93%86.40%45.09%96.93%94.18%72.69%98.80%81.40%66.18%98.34%85.07%
Corrupt l11DDPM-OOD mse.and.Lpips48.87%44.82%87.43%53.20%54.43%95.48%41.02%29.54%96.32%47.70%42.93%93.08%
LMD lipps alex51.23%47.48%89.99%57.53%56.49%90.92%40.94%30.30%96.32%49.90%44.76%92.41%
2XDM lipps alex25.48%36.16%99.29%50.43%49.86%90.10%54.97%54.29%95.69%54.73%38.27%93.29%46.40%44.65%94.59%
PUNC ROUGE (1_precision)28.75%38.34%98.77%54.58%55.29%96.02%27.69%37.57%98.78%27.45%37.27%98.54%34.62%42.12%98.03%
PUNC ROUGE (L_precision)27.29%38.11%99.15%54.32%55.59%96.16%27.30%38.07%99.45%27.35%37.89%99.37%34.07%42.42%98.53%
PUNC BERT (precision)85.85%84.45%73.94%79.43%78.83%87.21%54.97%53.73%94.51%75.76%72.91%84.65%74.00%72.48%85.08%
Corrupt l12DDPM-OOD lipps alex43.23%43.94%91.23%59.46%60.14%93.71%48.79%34.44%95.74%50.49%46.17%93.56%
LMD lipps alex46.90%45.65%92.04%59.94%58.90%90.54%40.88%30.42%96.35%49.24%44.99%92.98%
2XDM lipps alex58.15%58.50%93.98%52.08%53.36%90.19%53.57%54.17%96.34%60.24%44.23%91.78%56.01%52.57%93.07%
PUNC ROUGE (1_precision)28.76%38.10%98.59%54.77%55.49%96.02%27.66%37.42%98.75%27.39%37.23%98.53%34.64%42.06%97.97%
PUNC ROUGE (L_precision)27.33%37.87%99.16%54.51%55.87%96.11%27.31%37.92%99.35%27.33%37.83%99.36%34.12%42.37%98.50%
PUNC BERT (precision)85.96%84.23%73.54%79.47%78.80%87.05%54.78%53.51%94.79%76.04%73.44%84.58%74.06%72.49%84.99%
", + "bbox": [ + 133, + 433, + 864, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Performance comparison all applications (Vague, Adversarial, Corrupt Lvl1, Corrup Lvl2)", + "bbox": [ + 189, + 679, + 799, + 693 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "models with ethical and regulatory standards.", + "bbox": [ + 89, + 720, + 390, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our contributions also include a dataset specifically tailored to uncertainty quantification in text-to-image models, which we hope will serve as a benchmark for future research. Ultimately, this work highlights the potential of prompt-based uncertainty estimation as a critical tool for enhancing the reliability and safety of generative models in real-world applications.", + "bbox": [ + 89, + 742, + 482, + 848 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. While PUNC demonstrates promising results in quantifying uncertainty for text-to-image generation, several limitations remain. First, our approach", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "relies heavily on the interpretive abilities of LVLMs, which, despite their effectiveness, may introduce biases or inaccuracies inherited from their training data. Second, our analysis focuses primarily on English-language prompts, limiting our method's applicability to multilingual or culturally diverse datasets.", + "bbox": [ + 511, + 719, + 906, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was performed using HPC resources from GENCI-IDRIS (Grant 2023-[AD011011970R3]), and EuroHPC Development access to LEONARDO, hosted by CINECA.", + "bbox": [ + 511, + 829, + 906, + 891 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8069", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 5", + "[2] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2", + "[3] Nouar AlDahoul, Talal Rahwan, and Yasir Zaki. Aigenerated faces free from racial and gender stereotypes. arXiv preprint arXiv:2402.01002, 2024. 6", + "[4] Georges Le Bellier and Nicolas Audebert. Detecting Out-Of-Distribution Earth Observation Images with Diffusion Models. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 481–491, Los Alamitos, CA, USA, 2024. IEEE Computer Society. 2, 3", + "[5] Lucas Berry, Axel Brando, and David Meger. Shedding light on large generative networks: Estimating epistemic uncertainty in diffusion models. In The 40th Conference on Uncertainty in Artificial Intelligence, 2024. 2, 3, 5", + "[6] Matthew A Chan, Maria J Molina, and Christopher A Metzler. Hyper-diffusion: Estimating epistemic and aleatoric uncertainty with a single model. arXiv preprint arXiv:2402.03478, 2024. 3", + "[7] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart $\\backslash$ sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 3, 6, 2", + "[8] Junsong Chen, Jincheng YU, Chongjian GE, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- $\\mathbb{S}\\backslash \\mathbb{alpha}\\mathbb{S}$ : Fast training of diffusion transformer for photorealistic text-to-image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 2, 4", + "[9] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In Proceedings of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014. 2, 3", + "[10] Charles Corbière, Nicolas Thome, Avner Bar-Hen, Matthieu Cord, and Patrick Pérez. Addressing failure prediction by learning model confidence. Advances in Neural Information Processing Systems, 32, 2019. 2", + "[11] Matt Deitke, Christopher Clark, Sangho Lee, Rohun Tripathi, Yue Yang, Jae Sung Park, Mohammadreza Salehi, Niklas Muennighoff, Kyle Lo, Luca Soldaini, et al. Molmo and pixmo: Open weights and open data for state-of-the-art multimodal models. arXiv preprint arXiv:2409.17146, 2024. 5, 2", + "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7, 2" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Ning Ding, Yehui Tang, Zhongqian Fu, Chao Xu, Kai Han, and Yunhe Wang. Gpt4image: Can large pre-trained models help vision models on perception tasks? arXiv e-prints, pages arXiv-2306, 2023. 7, 2", + "[14] Zhekai Du and Jingjing Li. Diffusion-based probabilistic uncertainty estimation for active domain adaptation. Advances in Neural Information Processing Systems, 36: 17129-17155, 2023. 3", + "[15] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 5, 2", + "[16] Nikita Durasov, Timur Bagautdinov, Pierre Baque, and Pascal Fua. Masksembles for uncertainty estimation. In CVPR, 2021. 2", + "[17] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 4", + "[18] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In Proceedings of the 41st International Conference on Machine Learning, pages 12606-12633. PMLR, 2024. 2, 4", + "[19] Gianni Franchi, Andrei Bursuc, Emanuel Aldea, Séverine Dubuisson, and Isabelle Bloch. TRADI: Tracking deep neural network weight distributions. In ECCV, 2020. 2", + "[20] Gianni Franchi, Xuanlong Yu, Andrei Bursuc, Emanuel Aldea, Severine Dubuisson, and David Filliat. Latent discriminant deterministic uncertainty. In Computer Vision – ECCV 2022, pages 243–260, Cham, 2022. Springer Nature Switzerland. 2, 3", + "[21] Yarin Gal. Uncertainty in Deep Learning. PhD thesis, University of Cambridge, 2016. 2, 3", + "[22] Yonatan Geifman and Ran El-Yaniv. Selective classification for deep neural networks. In Proceedings of the 31st International Conference on Neural Information Processing Systems, page 4885-4894, Red Hook, NY, USA, 2017. Curran Associates Inc. 3, 4", + "[23] Mark S Graham, Walter HL Pinaya, Petru-Daniel Tudosiu, Parashkev Nachev, Sebastien Ourselin, and Jorge Cardoso. Denoising diffusion models for out-of-distribution detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2948-2957, 2023. 2, 3, 4, 5, 6", + "[24] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In International Conference on Learning Representations, 2017. 3, 4", + "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4" + ], + "bbox": [ + 93, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8070", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Eyke Hüllermeier and Willem Waegeman. *Aleatoric and epistemic uncertainty in machine learning: an introduction to concepts and methods*. *Machine Learning*, 110(3), 2021. 2, 3", + "[27] Pavel Izmailov, Sharad Vikram, Matthew D Hoffman, and Andrew Gordon Gordon Wilson. What are bayesian neural network posteriors really like? In ICML, 2021. 2", + "[28] John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, et al. Highly accurate protein structure prediction with alphafold. nature, 596(7873):583-589, 2021. 2", + "[29] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems, pages 26565-26577. Curran Associates, Inc., 2022. 4", + "[30] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 2", + "[31] Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. In The Eleventh International Conference on Learning Representations, 2023. 3, 5", + "[32] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. In NeurIPS, 2017. 2, 3", + "[33] Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, et al. Learning skillful medium-range global weather forecasting. Science, 382(6677):1416-1421, 2023. 2", + "[34] Olivier Laurent, Adrien Lafage, Enzo Tartaglione, Geoffrey Daniel, Jean-Marc Martinez, Andrei Bursuc, and Gianni Franchi. Packed-ensembles for efficient uncertainty estimation. In ICLR, 2023. 2, 3", + "[35] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 7, 2, 3", + "[36] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6, 1", + "[37] Zhiqiu Lin, Xinyue Chen, Deepak Pathak, Pengchuan Zhang, and Deva Ramanan. Revisiting the role of language priors in vision-language models. In *Forty-first International Conference on Machine Learning*. 2", + "[38] Krzysztof Lis, Krishna Nakka, Pascal Fua, and Mathieu Salzmann. Detecting the unexpected via image resynthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2152-2161, 2019. 3", + "[39] Zhenzhen Liu, Jin Peng Zhou, Yufan Wang, and Kilian Q Weinberger. Unsupervised out-of-distribution detection with diffusion inpainting. In International Conference on Machine Learning, pages 22528-22538. PMLR, 2023. 2, 3, 4, 6", + "[40] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan LI, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. In Advances in Neural Information Processing Systems, pages 5775-5787. Curran Associates, Inc., 2022. 4" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Andrey Malinin and Mark Gales. Predictive uncertainty estimation via prior networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 3", + "[42] Radford M Neal et al. Mcmc using hamiltonian dynamics. Handbook of markov chain monte carlo, 2011. 2", + "[43] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6", + "[44] Gangadhar Payyavula. Nuclei segmentation in microscope cell images. https://www.kaggle.com/datasets/gangadhar/nuclei-segmentation-in-microscope-cell-images, 2018. Accessed: 2024-09-01. 7, 2, 3", + "[45] Tim Pearce, Alexandra Brintrup, and Jun Zhu. Understanding softmax confidence and uncertainty, 2021. 3", + "[46] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2", + "[47] Jeremy Pinto. Bigearthnet. https://github.com/ jerpint/bigearthnet, 2022. Accessed: 2024-09-01. 7, 2, 3", + "[48] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2023. 2, 4, 6", + "[49] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Proceedings of the 38th International Conference on Machine Learning, pages 8821-8831. PMLR, 2021.", + "[50] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4, 6", + "[51] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2, 3", + "[52] Jaskirat Singh and Liang Zheng. Divide, evaluate, and refine: Evaluating and improving text-to-image alignment with iterative vqa feedback. Advances in Neural Information Processing Systems, 36:70799-70811, 2023. 2", + "[53] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 4", + "[54] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 4" + ], + "bbox": [ + 514, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8071", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. arXiv preprint arXiv:2409.17146, 2024. 4", + "[56] Tomas Vojir, Tomáš Šipka, Rahaf Aljundi, Nikolay Chumerin, Daniel Olmeda Reino, and Jiri Matas. Road anomaly detection by partial image reconstruction with segmentation coupling. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15651-15660, 2021. 3", + "[57] Max Welling and Yee W Teh. Bayesian learning via stochastic gradient Langevin dynamics. In ICML, 2011. 2", + "[58] Yeming Wen, Dustin Tran, and Jimmy Ba. BatchEnsemble: an alternative approach to efficient ensemble and lifelong learning. In ICLR, 2019. 2", + "[59] Guoxuan Xia and Christos-Savvas Bouganis. Augmenting softmax information for selective classification with out-of-distribution data. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 1995-2012, 2022. 3", + "[60] Guoxuan Xia and Christos-Savvas Bouganis. On the usefulness of deep ensemble diversity for out-of-distribution detection, 2022. 3", + "[61] Guoxuan Xia and Christos-Savvas Bouganis. Window-based early-exit cascades for uncertainty estimation: When deep ensembles are more efficient than single models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17368-17380, 2023. 2, 3", + "[62] Guoxuan Xia, Duolikun Danier, Ayan Das, Stathi Fotiadis, Farhang Nabiei, Ushnish Sengupta, and Alberto Bernacchia. Score Normalization for a Faster Diffusion Exponential Integrator Sampler. In NeurIPS 2023 Workshop on Diffusion Models, 2023. 4", + "[63] Guoxuan Xia, Olivier Laurent, Gianni Franchi, and Christos-Savvas Bouganis. Towards understanding why label smoothing degrades selective classification and how to fix it, 2024. 3", + "[64] Yingda Xia, Yi Zhang, Fengze Liu, Wei Shen, and Alan L Yuille. Synthesize then compare: Detecting failures and anomalies for semantic segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 145-161. Springer, 2020. 3", + "[65] Yilun Xu, Mingyang Deng, Xiang Cheng, Yonglong Tian, Ziming Liu, and Tommi S. Jaakkola. Restart sampling for improving generative processes. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 4", + "[66] Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. Generalized out-of-distribution detection: A survey. International Journal of Computer Vision, 2024. 3", + "[67] Michal Yarom, Yonatan Bitton, Soravit Changpinyo, Roeah Aharoni, Jonathan Herzig, Oran Lang, Eran Ofek, and Idan Szpektor. What you see is what you read? improving text-image alignment evaluation. Advances in Neural Information Processing Systems, 36:1601-1619, 2023. 2", + "[68] Xuanwu Yin Yuda Song, Zehao Sun. Sdxs: Real-time one-step latent diffusion models with image conditions. arxiv, 2024. 6, 2", + "[69] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. In The Eleventh" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "International Conference on Learning Representations, 2023.4", + "[70] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 4", + "[71] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 1", + "[72] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675, 2019. 6, 1", + "[73] Yimeng Zhang, Jinghan Jia, Xin Chen, Aochuan Chen, Yihua Zhang, Jiancheng Liu, Ke Ding, and Sijia Liu. To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images ... for now, 2024. 7, 3" + ], + "bbox": [ + 516, + 92, + 903, + 359 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8072", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Towards Understanding and Quantifying Uncertainty for Text-to-Image Generation/c3ad3744-8ee2-4117-a4a1-7d23c58e3fe4_model.json b/2025/Towards Understanding and Quantifying Uncertainty for Text-to-Image Generation/c3ad3744-8ee2-4117-a4a1-7d23c58e3fe4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..91c3a22604d3a3ffad447be29751e6d5e36d88ae --- /dev/null +++ b/2025/Towards Understanding and Quantifying Uncertainty for Text-to-Image Generation/c3ad3744-8ee2-4117-a4a1-7d23c58e3fe4_model.json @@ -0,0 +1,2356 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.229, + 0.131, + 0.773, + 0.177 + ], + "angle": 0, + "content": "Towards Understanding and Quantifying Uncertainty for Text-to-Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.259, + 0.203, + 0.739, + 0.277 + ], + "angle": 0, + "content": "Gianni Franchi,\\(^{1}\\)* Nacim Belkhir,\\(^{2}\\)* Dat Nguyen Trong,\\(^{1}\\) Guoxuan Xia,\\(^{3}\\) Andrea Pilzer\\(^{4}\\) \nU2IS, ENSTA Paris, Institut Polytechnique de Paris, \\(^{1}\\)Mirai,\\(^{2}\\) Imperial College London,\\(^{3}\\) NVIDIA\\(^{4}\\)" + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.309, + 0.379, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.309, + 0.88, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.606, + 0.908, + 0.649 + ], + "angle": 0, + "content": "Figure 1. Examples of Applications for Uncertainty Quantification in Text-to-Image Generation. Text-to-image generation models may exhibit uncertainty, and that need to be quantified since it can provide insights into the model's training dataset, aiding in deepfake prevention, detecting model biases, and protecting copyrighted content from unauthorized generation." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.661, + 0.329, + 0.677 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.695, + 0.486, + 0.862 + ], + "angle": 0, + "content": "Uncertainty quantification in text-to-image (T2I) generative models is crucial for understanding model behavior and improving output reliability. In this paper, we are the first to quantify and evaluate the uncertainty of T2I models with respect to the prompt. Alongside adapting existing approaches designed to measure uncertainty in the image space, we also introduce Prompt-based Uncertainty Estimation for T2I models (PUNC), a novel method leveraging Large Vision-Language Models (LVLMs) to better address uncertainties arising from the semantics of the prompt and generated images. PUNC utilizes a LVLM to caption a gen" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.663, + 0.909, + 0.89 + ], + "angle": 0, + "content": "erated image, and then compares the caption with the original prompt in the more semantically meaningful text space. PUNC also enables the disentanglement of both aleatoric and epistemic uncertainties via precision and recall, which image-space approaches are unable to do. Extensive experiments demonstrate that PUNC outperforms state-of-the-art uncertainty estimation techniques across various settings. Uncertainty quantification in text-to-image generation models can be used on various applications including bias detection, copyright protection, and OOD detection. We also introduce a comprehensive dataset of text prompts and generation pairs to foster further research in uncertainty quantification for generative models. Our findings illustrate that PUNC not only achieves competitive performance but also enables novel applications in evaluating" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.875, + 0.372, + 0.901 + ], + "angle": 0, + "content": "* equal contribution\nresponding author - gianni.franchi@ensta-paris.fr" + }, + { + "type": "page_number", + "bbox": [ + 0.474, + 0.875, + 0.483, + 0.884 + ], + "angle": 0, + "content": "+" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8062" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "and improving the trustworthiness of text-to-image models. The code is available at https://github.com/ENSTA-U2IS-AI/Uncertainty_diffusion" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.166, + 0.223, + 0.182 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.191, + 0.483, + 0.342 + ], + "angle": 0, + "content": "Generative models have revolutionized artificial intelligence, making significant strides in fields such as robotics [2], healthcare [28], and physics [33]. These models have elevated the performance of AI systems by enabling the generation of complex data representations. However, while generative models offer immense potential, they do not address all challenges, especially those related to robustness and uncertainty in Deep Neural Networks(DNNs)—issues that have become even more pronounced in generative settings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.343, + 0.483, + 0.659 + ], + "angle": 0, + "content": "In this paper, we focus on the uncertainty quantification in deep generative models, in particular, we are the first to investigate modern text-to-image (T2I) models where the input is a natural language prompt [8, 18, 48-50]. These models have surged in popularity in recent years due to their capacity to generate high-quality images conveniently in a user-guided fashion, finding usage in a wide array of real-world applications. We are the first to investigate and propose a dedicated approach for quantifying uncertainty in text-to-image generation, aiming to fill a critical gap in generative AI research. The task of quantifying uncertainty in these models is not only complex but also crucial. Understanding the uncertainty in generative models can be invaluable for applications such as Out-of-Distribution (OOD) detection, trustworthy AI, and decision-making in high-stakes domains. To the best of our knowledge, no existing work has comprehensively addressed the challenge of uncertainty quantification in text-to-image generative models. However, given the semantic nature of the uncertainty we aim to measure, our work is closely related to recent studies on text-to-image alignment [37, 52, 67]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Existing research explores uncertainty estimation and OOD detection in the image/output space [4, 5, 23, 39]. For example, given a test image, can we determine whether or not it is OOD? On the other hand, the uncertainty of a generation with respect to conditioning (such as text) is underexplored. We argue that this is an important area to be addressed, as it represents a widespread real-world application of deep generative models. In this case, it remains an open question as to how should we define uncertainty, and given a definition, how should we quantify it? As illustrated in Fig. 1, focusing on uncertainty quantification in text-to-image generation could enable novel applications, allowing us to extract and interpret the knowledge embedded within models using language." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To answer these questions, we first argue that uncertainties should be rooted in the semantics of the image" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.424 + ], + "angle": 0, + "content": "and the prompt. We then adapt and evaluate a number of existing image-space approaches, as well as proposing a novel method for uncertainty quantification in T2I generative models, which we refer to as \"Prompt-based Uncertainty Estimation for T2I Generation\" (PUNC). Our approach leverages the growing capabilities of Large Vision-Language Models (LVLMs) to extract the semantics from generated images. LVLMs, which have been trained on vast amounts of text and visual data, can serve as powerful tools for interpreting the underlying meaning of text/image inputs and assessing how variations in phrasing or semantics impact the uncertainty of the generated outputs. Our hypothesis is that as LVLMs become increasingly adept at understanding complex relationships between text and images, they will allow us to better analyze and quantify uncertainty. Through this paper, we will demonstrate Prompt-based Uncertainty Estimation for T2I (PUNC) effectiveness in quantifying uncertainties in generative models. Showcasing its ability to decompose and interpret different types of uncertainties arising from the multimodal nature of the input, offering a new perspective on uncertainty quantification in generative AI." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.905, + 0.635 + ], + "angle": 0, + "content": "To summarize, our contributions are as follows: (1) We introduce a new task aimed at quantifying uncertainty (with respect to the prompt) in text-to-image generation models, and we share a dataset of prompts to support future research on uncertainty in these models. (2) We both adapt existing image-space approaches to our tasks, as well as propose a novel straightforward method, PUNC, designed to better quantify semantic uncertainty in text-to-image generation. (3) Extensive experiments demonstrate that PUNC, despite its simplicity and computational efficiency, achieves competitive performance. (4) We highlight the practical utility of uncertainty quantification in text-to-image generation by showcasing additional applications where this approach could be beneficial." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.649, + 0.655, + 0.665 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Deep learning models typically encounter two primary types of uncertainty: aleatoric and epistemic uncertainty [26]. Over the years, a variety of methods have been developed to quantify these uncertainties. Ensemble methods [16, 32, 34, 58, 61], for instance, have often achieved state-of-the-art performance by leveraging multiple models to assess prediction confidence. However, these methods tend to be computationally expensive due to the need to train and maintain several models. Another approach involves Bayesian methods [21], especially Bayesian Neural Networks (BNNs) [19, 27, 42, 57], which offer a theoretically grounded way to quantify uncertainty. Additionally, there are techniques aimed at directly predicting uncertainty within DNNs [10, 20, 30]. Despite their effectiveness, most of these methods are challenging to apply to text-to-image" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8063" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.144, + 0.09, + 0.43, + 0.102 + ], + "angle": 0, + "content": "Generation/Image-space Uncertainty (existing work)" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.103, + 0.482, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.216, + 0.483, + 0.258 + ], + "angle": 0, + "content": "Figure 2. Diagram illustrating generation/image-space uncertainty which is considered in previous work and condition/prompt-space uncertainty which is investigated in our work." + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.291, + 0.462, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.401, + 0.484, + 0.526 + ], + "angle": 0, + "content": "Figure 3. Three generations from PixArt-\\(\\Sigma\\) [7] illustrating uncertainty with regards to prompt semantics. For the Normal image, we used an ImageNet-inspired prompt; for the corrupted image, additional corruption was applied to the prompt (e.g. fish was perturbed to fis) increasing the aleatoric uncertainty; and for the out-of-distribution (OOD) case, the model was prompted to generate an image of the Prime Minister of Japan Kishida Fumio under similar conditions, where epistemic uncertainty was injected in the form of a semantic concept the model was not familiar with." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.56, + 0.484, + 0.62 + ], + "angle": 0, + "content": "generation, as they rely on ensembling predictions or accessing model uncertainty in ways that are difficult to implement for generative models without explicit access to intermediate predictions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.9 + ], + "angle": 0, + "content": "Generative models, as a subset of deep learning, also exhibit uncertainty. In the past, traditional generative models were employed for uncertainty quantification and used to help estimate the uncertainty within standard DNNs [6, 38, 56, 64]. With the recent popularity of diffusion models, research has begun exploring whether image-based diffusion models can detect out-of-distribution (OOD) samples [4, 23, 39] or quantify their own uncertainty [5, 6, 14]. Whilst these approaches align with our goals, they are limited to image-space tasks, where the input is an image rather than a textual prompt. This input difference fundamentally alters the nature of uncertainty in the model. Our approach, in contrast, specifically addresses text-to-image generation, a setting where prompt ambiguity and multimodal complexity introduce new dimensions of uncertainty. We demonstrate that quantifying uncertainty in this context can lead to valuable insights and applications across various downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.09, + 0.882, + 0.108 + ], + "angle": 0, + "content": "3. Uncertainty in Image Generation Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.907, + 0.312 + ], + "angle": 0, + "content": "Consider a deep generative model \\(p_{\\theta}(\\boldsymbol{x})\\) with parameters \\(\\theta\\) that models the data distribution \\(p_{\\mathrm{data}}(\\boldsymbol{x})\\) of \\(\\boldsymbol{x} \\in \\mathbb{R}^D\\). We may be interested in measuring the uncertainty of our model \\(p_{\\theta}\\) with respect to some test datum \\(\\boldsymbol{x}^*\\). For example, consider the case where we want to determine whether or not \\(\\boldsymbol{x}^*\\) was drawn from \\(p_{\\mathrm{data}}(\\boldsymbol{x})\\), i.e. out-of-distribution (OOD) detection. This can be achieved by quantifying the uncertainty of \\(p_{\\theta}\\) that arises from a lack of knowledge of \\(\\boldsymbol{x}^*\\), i.e. epistemic uncertainty [26]. A range of existing work that leverages diffusion-based generative models for OOD image detection [23, 39] fall into the above paradigm. From hereon we will refer to this as generation/image-space uncertainty. This is illustrated in Fig. 2 (top)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.312, + 0.906, + 0.478 + ], + "angle": 0, + "content": "In this work, however, we are primarily interested in a different scenario: text-to-image generation. Consider instead a conditional model \\( p_{\\theta}(\\pmb{x}|\\pmb{c}) \\) that models the conditional data distribution \\( p_{\\mathrm{data}}(\\pmb{x}|\\pmb{c}) \\). This describes real-world text-to-image generation, where image \\( \\pmb{x} \\) is drawn from \\( p_{\\theta}(\\pmb{x}|\\pmb{c}^*) \\) conditioned on some test text prompt \\( \\pmb{c}^* \\). In this case, we wish to quantify the uncertainty of \\( p_{\\theta} \\) not with respect to some test \\( \\pmb{x}^* \\) (e.g. image) but instead with respect to a test condition \\( \\pmb{c}^* \\). We will hereon refer to this as condition/prompt-space uncertainty. This is illustrated in Fig. 2 (bottom)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.479, + 0.906, + 0.599 + ], + "angle": 0, + "content": "We argue that in this case, uncertainties should be concerned with semantics, as intuitively, these are what inherently link the user's intentions, represented by the text prompt, and the generated output image. This is similar to recent discourse on uncertainty estimation for large language models [31]. If we consider the semantic concepts in the prompt and generated image, then we can define useful concepts of uncertainty (illustrated in Fig. 3):" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.599, + 0.906, + 0.704 + ], + "angle": 0, + "content": "- Aleatoric uncertainty is irreducible uncertainty in the data distribution \\( p_{\\mathrm{data}}(\\boldsymbol{x}|\\boldsymbol{c}) \\). From the perspective of semantics, high aleatoric uncertainty would be where a variation of generated concepts may arise from a single prompt. For example, a spelling mistake where fish is mistyped as fis in the prompt may result in either a \"fish\" or a \"fist\" appearing in the image." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.705, + 0.906, + 0.794 + ], + "angle": 0, + "content": "- Epistemic uncertainty should correspond to a model's lack of knowledge of semantic concepts found in a prompt. For example, a model trained on ImageNet will not know what the prime minister of Japan Kishida Fumio looks like, and thus will have high epistemic uncertainty for this semantic concept." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.599, + 0.906, + 0.794 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.856 + ], + "angle": 0, + "content": "We note that broadly speaking uncertainty estimation for deep conditional models is well explored [20-22, 24, 26, 32, 34, 41, 59-61, 63, 66]. However, research in this area is generally concerned with problem settings where 1) the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.512, + 0.863, + 0.907, + 0.901 + ], + "angle": 0, + "content": "1We note that although there is a discussion over how to define and quantify epistemic uncertainty [21, 26, 41, 45, 60], in this work, we stick to an intuitive albeit informal view based on a deep model's knowledge." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8064" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "models explicitly evaluate \\( p_{\\theta}(\\pmb{x}|c) \\) and 2) downstream applications of uncertainty relate directly to the numerical values of \\( \\pmb{x} \\) and \\( p_{\\theta}(\\pmb{x}|c) \\). For example, a cross-entropy-trained image classifier explicitly predicts the probability mass \\( P_{\\theta}(\\pmb{x}|c^{*}) \\) of label \\( \\pmb{x} \\) given test image \\( c^{*} \\). \\( P_{\\theta}(\\pmb{x}|c^{*}) \\) directly relates to the likelihood of classification error – which is an event that we may want to detect using an uncertainty measure, i.e. downstream misclassification detection [22, 24]. Thus, we are able to easily extract useful measures of uncertainty from the model \\( P_{\\theta} \\), e.g. the maximum of the softmax." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.244, + 0.483, + 0.455 + ], + "angle": 0, + "content": "On the other hand, text-to-image generation models are not as clear-cut. They may not explicitly evaluate \\( p_{\\theta}(\\pmb{x}|c^{*}) \\) (e.g. diffusion-based models [25, 54]), and even when they do (e.g. vector-quantised autoregressive models [17, 55]) it is hard to directly relate (latent) pixel values directly to our above definitions of uncertainty. For example, simple variations in contrast and brightness that may leave semantics completely unchanged could result in a very high variance in \\( p_{\\theta}(\\pmb{x}|c^{*}) \\). To give a simple example, inverting the black and white stripes on a zebra would lead to a large Euclidean distance in the image space whilst preserving the semantics of the image. In the next section, we present and discuss a range of potential methods to quantify the uncertainty of text-to-image models with respect to a given prompt." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.47, + 0.415, + 0.489 + ], + "angle": 0, + "content": "4. Techniques to Quantify Uncertainty" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.497, + 0.483, + 0.543 + ], + "angle": 0, + "content": "We quantify text-to-image uncertainty by adapting existing methods and proposing a novel approach named Prompt-based UNCertainty estimation (PUNC)(Fig. 4)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.553, + 0.363, + 0.57 + ], + "angle": 0, + "content": "4.1. Adapting Existing Approaches" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.576, + 0.483, + 0.668 + ], + "angle": 0, + "content": "There are a number of existing methods for quantifying the image-space uncertainty of diffusion models. As diffusion-based models represent a majority of current text-to-image foundation models [8, 18, 48, 50], we choose to adapt these in order to quantify prompt-space uncertainty (Fig. 4). We first present a brief primer on diffusion models to the reader." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.677, + 0.258, + 0.691 + ], + "angle": 0, + "content": "4.1.1. Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.696, + 0.483, + 0.743 + ], + "angle": 0, + "content": "Diffusion models [25] generate data by iteratively denoising pure noise. The forward process over time \\( t \\in [0,1] \\) for \\( \\boldsymbol{x}_t \\in \\mathbb{R}^D \\), conditioned on a text prompt \\( c \\), follows:" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.754, + 0.483, + 0.772 + ], + "angle": 0, + "content": "\\[\np \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}, \\boldsymbol {c}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t}; a _ {t} \\boldsymbol {x} _ {0}, \\sigma_ {t} ^ {2} \\boldsymbol {I}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.785, + 0.483, + 0.876 + ], + "angle": 0, + "content": "where \\( \\pmb{x}_0 \\) is sampled from an unknown data distribution \\( p(\\pmb{x}_0) \\), and \\( a_{t},\\sigma_{t} \\) define the noise schedule. This formulation allows diffusion models to progressively refine noisy inputs, enabling high-quality data generation through a learned denoising process. By leveraging stochastic differential equations (SDEs) or their deterministic counterparts," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.153 + ], + "angle": 0, + "content": "these models sample from complex distributions in a structured manner. Various solvers [29, 40, 53, 54, 62, 65, 69] can then be employed to efficiently generate realistic samples with a limited number of denoising steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.161, + 0.907, + 0.268 + ], + "angle": 0, + "content": "Latent diffusion models (LDMs) [39] perform this process within a compressed latent space rather than directly on pixel space. By first encoding images into a more compact representation, LDMs reduce computational demands, enabling high-quality high-resolution generations. We will not consider the latent representation to keep the notation simple, even though experiments will use LDMs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.3, + 0.757, + 0.315 + ], + "angle": 0, + "content": "4.1.2. Time-step based approaches" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.327, + 0.907, + 0.569 + ], + "angle": 0, + "content": "DDPM-OOD [23]. This approach, developed for image-space OOD detection, involves noising a test sample \\( \\boldsymbol{x}_0^* \\) to different points in the forward process to get \\( \\{\\boldsymbol{x}_{t_1}^*,\\boldsymbol{x}_{t_2}^*,\\dots \\} \\) and denoising them, generating multiple reconstructions \\( \\{\\hat{x}_{0_1},\\hat{x}_{0_2},\\ldots \\} \\). Similarity is then measured between each reconstruction and the original image. The idea is that test samples \\( \\boldsymbol{x}_0^* \\) for which the model has higher epistemic uncertainty will lead to reconstructions that differ from the test sample. However, in our case, we do not have a test image \\( \\boldsymbol{x}_0^* \\). In order to adapt DDPM-OOD we propose measuring similarity between reconstructions at different \\( t \\) and a baseline generation, \\( S(\\hat{x}_{0_i},\\hat{x}_0) \\), where \\( \\hat{x}_0 \\) represents an image directly generated using the original test prompt \\( c^* \\) alone. Both aleatoric uncertainty (e.g. ambiguous prompt) and epistemic uncertainty (e.g. unfamiliar concepts in prompt) could lead to more divergent reconstructions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.909, + 0.811 + ], + "angle": 0, + "content": "LMD [39]. Another similar approach for image-space OOD detection, known as Learned Manifold Denoising (LMD), instead explores alternative transforms on \\( \\boldsymbol{x}_0^* \\) to noising. The idea is to lift \\( \\boldsymbol{x}_0^* \\) from its original data manifold and then map it back using the trained diffusion model. If the sample is in-distribution, the mapping process should return it close to its original position on the data manifold. Conversely, if it is OOD, the mapped position will likely diverge significantly from the original. The corruption process typically involves masking portions of the image, with various mask patterns applied. Similar to DDPM-OOD, similarity is measured between the original image and the reconstructed version after applying the diffusion process. Similar to before, since we lack a testimage, we measure the similarity between reconstructed images and \\( \\hat{\\boldsymbol{x}}_{0_i} \\), the image generated directly from the prompt without perturbations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Similarity Metric. Various methods are used to evaluate similarity between images. Mean-squared error (MSE) is commonly applied between input and reconstruction, while some approaches favor the Learned Perceptual Image Patch Similarity (LPIPS) [70] metric. The choice of similarity metric is a crucial component of these techniques." + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.887, + 0.474, + 0.901 + ], + "angle": 0, + "content": "2We stick with this non-standard notation for the sake of consistency." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8065" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.907, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.259, + 0.908, + 0.302 + ], + "angle": 0, + "content": "Figure 4. Illustration showing the different baselines and PUNC. PUNC leverages a LVLM to describe generated images and assess similarity with the original prompt, providing a refined uncertainty score. In contrast, baseline methods employ traditional techniques such as noise injection, ensembling, or masking to quantify uncertainty, followed by image-based similarity scoring." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.315, + 0.383, + 0.33 + ], + "angle": 0, + "content": "4.1.3. Test-Time Ensembling Approaches" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.333, + 0.484, + 0.529 + ], + "angle": 0, + "content": "Test-time ensembling techniques take inspiration from approaches used in the large language model (LLM) community [31], where multiple generations are produced to assess consistency. Given a prompt \\( c \\), we generate two images \\( \\{\\hat{x}_1, \\hat{x}_2\\} \\) and measure the similarity between the two outputs. We denote this technique 2XDM. The underlying rationale is that if the model has high confidence, it will generate similar images. However, if uncertainty is high, it will generate substantially different images. We note that this is similar to the approach proposed in [5], however, that approach uses an ensemble of diffusion models, which is not practical in the setting where we want to quantify the uncertainty of a single pretrained T2I model." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.531, + 0.484, + 0.713 + ], + "angle": 0, + "content": "A key question remains regarding how to find the best measure similarity, and we propose using the same techniques applied in the time-step-based methods. A significant advantage of test-time ensembling approaches is that they can be applied generically to all T2I models. In contrast, timestep-based approaches require a multi-step generation process, limiting the broadness of their applicability. We note that all above approaches require multiple generations, incurring a notable computational overhead at inference time. For example, DDPM-OOD requires \\(\\sim 50\\times\\) the number of network inferences as a single generation (if the hyperparameters of the original paper are used) [23]." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.722, + 0.421, + 0.737 + ], + "angle": 0, + "content": "4.2. Prompt-based Uncertainty Estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.745, + 0.484, + 0.851 + ], + "angle": 0, + "content": "In this section, we propose a new method, Prompt-based Uncertainty estimation for text-to-image generation (PUNC), which leverages Large Vision-Language Models (LVLM) in conjunction with T2I models. We aim to assess the uncertainty in generated images by analyzing the alignment between initial text prompts and reconstructed textual descriptions of the generated images." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.885 + ], + "angle": 0, + "content": "Background on Large Vision-Language Model (LVLM)" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "A LVLM integrates a large language model (LLM) with" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.315, + 0.905, + 0.359 + ], + "angle": 0, + "content": "an image encoder, allowing for robust, multimodal understanding. Given a text prompt \\( c \\) and an image \\( x \\), a LVLM model, with parameters \\( \\omega \\), includes:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.361, + 0.905, + 0.394 + ], + "angle": 0, + "content": "- Image Encoder: \\( f_{\\omega}^{\\mathrm{img}}(\\cdot) \\) processes the input image \\( x \\) and produces an embedding: \\( z^{\\mathrm{img}} = f_{\\omega}^{\\mathrm{img}}(x) \\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.394, + 0.905, + 0.439 + ], + "angle": 0, + "content": "- Text Processor with LLM: \\( f_{\\omega}^{\\mathrm{txt}}(\\cdot, \\cdot) \\) takes the prompt \\( c \\) and the image embedding \\( z^{\\mathrm{img}} \\) as inputs, generating a descriptive or answer-based response: \\( \\hat{c} = f_{\\omega}^{\\mathrm{txt}}(c, z^{\\mathrm{img}}) \\)" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.361, + 0.905, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.905, + 0.472 + ], + "angle": 0, + "content": "Thus, a LVLM system generates an interpretation \\(\\hat{c}\\) that reflects the content of the image given an initial prompt." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.906, + 0.67 + ], + "angle": 0, + "content": "Prompt-Based Uncertainty Estimation for T2I Generation (PUNC) The previously introduced approaches to uncertainty estimation rely on measuring similarity between images. For a given text prompt \\( c \\), a T2I model generates an image \\( x \\sim p_{\\theta}(x|c) \\). We reason that, even under low uncertainty, the model can produce images that vary in aspects such as color, texture, or object positioning, while still faithfully representing the semantics of prompt \\( c \\). As previously discussed in Sec. 3, this may result in image-based similarity metrics failing to capture uncertainty at a semantic level, since large differences in the image space may not necessarily correspond to meaningful semantic differences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.67, + 0.906, + 0.82 + ], + "angle": 0, + "content": "We propose that advanced LVLMs, such as Molmo [11], LLAMA 3 [15], and GPT-4 [1], are a strong candidate for extracting the semantics from an image, through describing an image via natural language. Building on this insight, our proposed method, PUNC, introduces a novel approach that bypasses direct image comparison and instead evaluates the alignment between generated images and the original prompts by calculating an uncertainty score in the more semantically meaningful text-space. The steps for implementing PUNC are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.821, + 0.905, + 0.851 + ], + "angle": 0, + "content": "Step 1: Initial Prompt and Image Generation Given a test prompt \\( c^* \\), we use the T2I model to sample an image \\( x \\):" + }, + { + "type": "equation", + "bbox": [ + 0.66, + 0.861, + 0.759, + 0.877 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} \\sim p _ {\\boldsymbol {\\theta}} (\\boldsymbol {x} | \\boldsymbol {c} ^ {*})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Step 2: Image Interpretation via LVLM With the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8066" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.088, + 0.447, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.219, + 0.47, + 0.233 + ], + "angle": 0, + "content": "Figure 5. Disentangling Aleatoric and Epistemic Uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.26, + 0.483, + 0.292 + ], + "angle": 0, + "content": "generated image \\( \\pmb{x} \\) and the initial prompt \\( c^* \\), the LVLM produces a new descriptive caption \\( \\hat{c} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.302, + 0.361, + 0.32 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {c}} = f _ {\\omega} ^ {\\mathrm {t x t}} \\left(\\boldsymbol {c} ^ {*}, f _ {\\omega} ^ {\\mathrm {i m g}} (\\boldsymbol {x})\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.331, + 0.485, + 0.392 + ], + "angle": 0, + "content": "Step 3: Uncertainty Score Calculation The core of PUNC lies in measuring the alignment between \\( \\mathbf{c} \\) and \\( \\hat{\\mathbf{c}} \\) to assess the diffusion model's confidence. Specifically, the similarity score \\( S(\\mathbf{c},\\hat{\\mathbf{c}}) \\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.404, + 0.363, + 0.421 + ], + "angle": 0, + "content": "\\[\nS (\\boldsymbol {c} ^ {*}, \\hat {\\boldsymbol {c}}) = \\operatorname {s i m} (\\boldsymbol {c} ^ {*}, \\hat {\\boldsymbol {c}})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.483, + 0.493 + ], + "angle": 0, + "content": "High similarity suggests low uncertainty, indicating the generated image closely matches the initial prompt. Conversely, a lower score reveals higher uncertainty, signifying potential divergence from the prompt's intended meaning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.484, + 0.539 + ], + "angle": 0, + "content": "Thus, PUNC is able to predict semantic uncertainties by utilising the powerful image understanding ability of LVLMs and computing similarities in the text space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.675 + ], + "angle": 0, + "content": "Prompt Similarity: Aleatoric and Epistemic via Precision and Recall. In Sec. 3 we discussed the different types of semantic uncertainty for T2I generation. Here we will present a method for disentangling these uncertainties using the precision and recall of the caption given the prompt. Consider an abstract scenario where the prompt and the generated image each contain a discrete set of semantic concepts. We can then calculate the precision and recall of the image concepts with respect to the prompt:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.796 + ], + "angle": 0, + "content": "Intuitively, a lack of knowledge about concepts in the prompt (epistemic uncertainty) will result in fewer concepts being preserved in the image, reducing recall. On the other hand, a lack of specificity in the prompt (aleatoric uncertainty) will result in additional concepts being generated, reducing precision. Thus we can use this abstract form of semantic-concept-based recall to measure epistemic uncertainty, and use precision to measure aleatoric uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Although we cannot directly measure concepts, we can approximate them using natural language. Moreover, existing similarity scores for text already have the above notions of precision and recall integrated within [36, 43, 72], allowing PUNC to disentangle uncertainty by comparing the caption of the generated image with the original prompt. We note that the image-based approaches in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "Sec. 4.1 are unable to disentangle aleatoric from epistemic uncertainty. We focus on two different existing text-based similarity metrics, ROUGE [36], and BERTScore [72]. The former calculates similarity (precision or recall) by directly matching (sequences of) words between two text sequences. The latter on the other hand measures similarity in the deep feature space of a text encoder." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.198, + 0.907, + 0.303 + ], + "angle": 0, + "content": "As illustrated in Fig. 5, we define Prompt Similarity as the semantic overlap between the input prompt and the generated descriptive caption. More formally, let \\( P_{\\mathrm{GT}} \\) denote the set of semantic concepts in the ground-truth prompt used to generate an image, and \\( P_{\\mathrm{VQA}} \\) the set of concepts extracted from the LVLM-generated caption. The resulting uncertainties can be characterized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.303, + 0.905, + 0.349 + ], + "angle": 0, + "content": "Epistemic Uncertainty, which captures missing concepts in \\( P_{\\mathrm{VQA}} \\) that are present in \\( P_{\\mathrm{GT}} \\), reflects an incomplete understanding of the intended semantics. It is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.359, + 0.905, + 0.392 + ], + "angle": 0, + "content": "\\[\nU _ {\\text {e p i s t e m i c}} = 1 - \\frac {\\left| P _ {\\mathrm {G T}} \\cap P _ {\\mathrm {V Q A}} \\right|}{\\left| P _ {\\mathrm {G T}} \\right|} = 1 - \\text {R e c a l l} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.403, + 0.905, + 0.462 + ], + "angle": 0, + "content": "Aleatoric Uncertainty, on the other hand, arises from spurious or extra concepts in \\( P_{\\mathrm{VQA}} \\) that do not appear in \\( P_{\\mathrm{GT}} \\), indicating ambiguity in the model's response. It is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.471, + 0.905, + 0.505 + ], + "angle": 0, + "content": "\\[\nU _ {\\text {a l e a t o r i c}} = 1 - \\frac {\\left| P _ {\\mathrm {G T}} \\cap P _ {\\mathrm {V Q A}} \\right|}{\\left| P _ {\\mathrm {V Q A}} \\right|} = 1 - \\text {P r e c i s i o n} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.515, + 0.907, + 0.546 + ], + "angle": 0, + "content": "Further details on the computation of precision and recall can be found in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.559, + 0.767, + 0.575 + ], + "angle": 0, + "content": "5. Experiments & Discussions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.871 + ], + "angle": 0, + "content": "We test PUNC on Various task and various dataset and compared PUNC to DDPM-OOD [23], LMD [39] and 2XDM. We first introduce on Sec. 5 how we have constitute the dataset of prompts. Then on Sec. 5.1 we discuss of different results on these dataset. Then Appendix F explains how we can use the uncertainty of text-to-image model to detect if a model has been trained with some concept related to some fictional character of politicians. Appendix F.3 shows how uncertainty can help to detect bias of T2I models. To conduct our experiments and evaluate the various techniques for quantifying uncertainty, we utilized four T2I models: Stable Diffusion 1.5 (SDv1.5) [50], SDXS [68], SDXL [48], and PixArt- \\(\\Sigma\\) [7]. These diverse T2I models allow us to test a range of architectures, each with unique characteristics and potential weaknesses. SDv1.5, being one of the earliest models, likely has a smaller and less diverse training dataset compared to the others. In contrast, SDXS generates in a single-step, different to the multi-step processes in other models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.872, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Dataset of prompts To conduct our experiments, we require both an in-distribution dataset of prompts and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8067" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.334 + ], + "angle": 0, + "content": "datasets with uncertainty. For the in-distribution prompts, we use the dataset proposed by [13], where high-quality descriptions of ImageNet [12] images are generated using GPT-4[13], referred to as Normal. To simulate epistemic uncertainty, we create out-of-distribution (OOD) datasets using images from underrepresented domains like remote sensing, texture, and microscopic datasets, captioned with LLAVA Next[35], labeled as Remote Sensing, Texture, and Microscopic. For aleatoric uncertainty, we generate two datasets: Vague, containing 2,000 minimally descriptive prompts (e.g., \"An image of ***\"), and Corrupted, derived from Normal by adding grammatical errors (Level 1) and removing \\(50\\%\\) of words (Level 2). A detailed summary of these datasets is presented in Tab. 1 and Appendix E. In Appendix C, we provide the details of the hyperparameters used in our experiments." + }, + { + "type": "table", + "bbox": [ + 0.132, + 0.346, + 0.445, + 0.476 + ], + "angle": 0, + "content": "
Dataset TypeDataset NameDescription
In-Distribution (ID)Normal20k High-quality detailed prompts from ImageNet [12] images using GPT-4[13].
Out-of-Distribution (OOD)Remote Sensing Texture Microscopic20k Prompts from remote sensing images [47]5k Prompts from texture datasets10k Prompts from microscopic images [44]
Aleatoric UncertaintyVagueAdversarialCorrupted2k prompts with minimal context, e.g., "An image of ***," where "****" is an ImageNet class name, creating ambiguous input.1k altered prompts from Normal using Unlearn-DiffAtk [73]20k Altered prompts from Normal:• Level 1: Grammatical mistakes and spelling errors.• Level 2: Random removal of 50% of words in Level 1 prompts.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.105, + 0.487, + 0.468, + 0.501 + ], + "angle": 0, + "content": "Table 1. Summary of prompt datasets for unc. quantification." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.528, + 0.395, + 0.545 + ], + "angle": 0, + "content": "5.1. Results for Uncertainty Estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.656 + ], + "angle": 0, + "content": "Tab. 3 presents the results of PUNC when using Molmo as the LVLM to assess epistemic uncertainty in the T2I models. Our approach, PUNC, outperforms most state-of-the-art techniques on average. However, we observe more mixed results on the Texture dataset, likely due to uncertainty about the presence of texture images in the training data of the T2I models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.483, + 0.839 + ], + "angle": 0, + "content": "In Tab. 4, we show the results of PUNC under aleatoric uncertainty, again using Molmo as the LVLM. Our approach outperforms most state-of-the-art methods in this setting as well. Notably, while the ROUGE score tends to perform better for epistemic uncertainty, the optimal metric is less clear for aleatoric uncertainty. Here, the choice between ROUGE and BERTScore may depend on the specific application: because corruption introduces spelling and grammatical errors, BERTScore—which captures semantic similarity—may be more suitable. For additional results using different LVLMs, please refer to Appendix D and Appendix B. In Appendix G, we present qualitative results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.337, + 0.866 + ], + "angle": 0, + "content": "5.2. Applications of uncertainty" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Our study explores multiple applications of uncertainty quantification in T2I models, with detailed results provided" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.242 + ], + "angle": 0, + "content": "in Appendix F. First, we investigate deepfake detection by examining the model's ability to generate recognizable images of prominent politicians (see Appendix F.1). We examined whether diffusion models could generate convincing images of G7 politicians. Using 100 prompts containing their names, we generated images with four diffusion models and queried a LVLM to determine if the intended politician was present in each image (yes/no), thus measuring concept recognition uncertainty. Results are presented in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.243, + 0.905, + 0.348 + ], + "angle": 0, + "content": "Next, we address copyright concerns by evaluating whether T2I models can accurately replicate iconic, copyrighted characters such as Mickey Mouse and Darth Vader (see Appendix F.2). Finally, we analyze bias in T2I models by testing their capacity to generate individuals across different genders and races for specific job roles (see Appendix F.3)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.349, + 0.907, + 0.485 + ], + "angle": 0, + "content": "Using a LVLM, we evaluate the presence of specific individuals, copyrighted characters, or demographic concepts in the generated images, demonstrating how uncertainty estimation can aid in detecting potential misuses of generative models. These experiments illustrate the value of text-based uncertainty quantification in analyzing T2I model performance across ethically and legally sensitive applications, enabling assessments that are challenging to achieve with image-based generation alone." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.498, + 0.634, + 0.515 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.905, + 0.75 + ], + "angle": 0, + "content": "In this work, we set out to be first to investigate uncertainty estimation T2I generation, where the uncertainty is with respect to the text prompt. We both adapt existing image-space uncertainty approaches, as well as present Prompt-based UNCertainty Estimation for T2I models (PUNC), the first approach designed specifically to quantify uncertainty in T2I generation. By leveraging the interpretive power of Large Vision-Language Models to extract semantics from a generated image, PUNC addresses the challenges associated with measuring multimodal uncertainty, enabling a deeper understanding of model behavior with respect to prompt-based conditioning. Our results validate that PUNC provides reliable, fine-grained assessments of aleatoric and epistemic uncertainties, outperforming existing methods that work for image generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.751, + 0.907, + 0.902 + ], + "angle": 0, + "content": "PUNC opens avenues for new applications, such as improving model robustness in biased or OOD scenarios and safeguarding against unauthorized content generation, including deepfakes and copyrighted characters. Our experiments show that diffusion models, while powerful, still exhibit significant variances in their ability to adhere to prompt details, particularly with ambiguous, corrupt prompts or OOD prompts. By identifying such uncertainties, PUNC not only supports the development of more trustworthy generative AI but also aids in aligning these" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8068" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.217, + 0.089, + 0.782, + 0.152 + ], + "angle": 0, + "content": "
Justin TrudeauKing CharlesFumio KishidaOlaf ScholzEmmanuel MacronJoe BidenGiorgia MeloniAVERAGE
SDXL9494888898962195.5
SDv1.57860628894801878
SDXS8040708294762872.5
PixArt3676627894901574
AVERAGE7267.570.5849585.520.570.71
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.163, + 0.907, + 0.192 + ], + "angle": 0, + "content": "Table 2. Accuracy (%) of diffusion models in generating recognizable images of G7 politicians. Higher accuracy indicates lower uncertainty in the model's ability to represent specific individuals." + }, + { + "type": "table", + "bbox": [ + 0.136, + 0.206, + 0.865, + 0.395 + ], + "angle": 0, + "content": "
auroc ↑SDXS aupr ↑fpr95↓auroc ↑PixArt aupr ↑fpr95↓auroc ↑SDv1.5 aupr ↑fpr95↓auroc ↑SDXL aupr ↑fpr95↓auroc ↑Average aupr ↑fpr95↓
microscopicDDPM-OOD mse_and_lpips39.61%56.38%100.00%71.64%71.52%93.62%87.25%83.67%51.43%66.17%70.53%81.68%
LMD lpips alex45.82%45.24%99.10%16.81%31.61%99.88%5.75%28.12%100.00%22.79%34.99%99.66%
2XDM lpips alex24.49%32.60%99.91%24.32%33.21%99.99%16.19%32.42%99.94%30.58%36.99%98.25%23.90%33.81%99.52%
PUNC ROUGE (1_recall)84.00%85.10%57.07%95.47%97.06%30.84%87.24%90.66%48.90%83.56%84.31%52.78%87.57%89.28%47.40%
PUNC ROUGE (L_recall)87.64%89.56%55.28%96.96%98.02%16.85%90.20%93.40%43.56%88.19%89.84%47.44%90.75%92.71%40.78%
PUNC BERT (recall)88.45%89.71%48.68%97.83%98.41%10.32%87.75%91.70%51.03%83.10%85.46%58.14%89.28%91.32%42.04%
remote sensingDDPM-OOD mse_and_lpips80.00%85.37%99.28%74.36%71.89%71.00%65.47%58.25%71.97%73.28%71.84%80.75%
LMD lpips alex90.65%93.11%74.49%78.93%79.31%81.05%86.60%85.39%49.19%85.39%85.94%68.24%
2XDM lpips alex53.88%59.21%98.23%54.43%57.99%98.78%30.66%38.80%98.76%23.54%35.35%96.23%40.63%47.83%98.00%
PUNC ROUGE (1_recall)85.59%86.59%63.87%94.14%95.97%49.29%73.24%72.93%79.56%73.07%74.54%83.48%81.51%82.51%69.05%
PUNC ROUGE (L_recall)89.85%90.63%51.96%97.01%97.94%16.08%83.63%84.65%65.35%85.68%87.09%64.00%89.04%90.08%49.34%
PUNC BERT (recall)80.50%82.56%75.19%91.93%93.08%43.30%61.27%64.74%90.08%60.71%65.32%90.94%73.60%76.42%74.88%
TextureDDPM-OOD mse_and_lpips73.22%62.02%89.24%70.21%55.97%99.80%74.24%51.97%84.71%72.56%56.65%91.25%
LMD lpips alex63.40%50.47%97.52%26.74%16.06%99.89%35.66%23.22%99.95%41.93%29.92%99.12%
2XDM lpips alex32.80%20.84%99.98%37.21%23.36%99.37%29.58%24.27%99.93%28.87%16.49%98.88%32.11%21.24%99.54%
PUNC ROUGE (1_recall)40.77%72.88%95.54%83.26%94.80%73.48%46.61%75.64%90.87%49.16%78.04%93.57%54.95%80.34%88.36%
PUNC ROUGE (L_recall)45.64%76.90%95.99%84.10%95.13%74.21%47.60%77.70%93.14%50.71%80.05%94.35%57.01%82.44%89.42%
PUNC BERT (recall)49.11%79.11%94.65%84.40%95.14%69.99%47.60%78.80%93.85%46.72%79.20%96.87%56.96%83.06%88.84%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.097, + 0.406, + 0.9, + 0.42 + ], + "angle": 0, + "content": "Table 3. Performance comparison of PUNC applied to different Out of Distribution concepts and different text-to-image models" + }, + { + "type": "table", + "bbox": [ + 0.135, + 0.434, + 0.865, + 0.669 + ], + "angle": 0, + "content": "
auroc ↑SDXS aupr ↑fpr95↓auroc ↑PixArt aupr ↑fpr95↓auroc ↑SDv1.5 aupr ↑fpr95↓auroc ↑SDXL aupr ↑fpr95↓auroc ↑Average aupr ↑fpr95↓
VagueDDPM-OOD mse.and.Lpips31.41%6.06%96.45%50.06%9.30%96.52%40.59%5.36%97.37%40.69%6.91%96.78%
LMD lipps alex39.26%6.91%96.35%47.62%9.54%97.70%40.77%5.70%97.95%42.55%7.38%97.33%
2XDM lipps alex51.28%10.10%94.35%43.82%8.57%96.89%50.97%10.09%95.96%57.83%10.01%94.37%50.98%9.69%95.39%
PUNC ROUGE (1_precision)99.99%100.00%0.00%100.00%100.00%0.00%99.99%100.00%0.00%99.99%100.00%0.00%99.99%100.00%0.00%
PUNC ROUGE (L_precision)99.99%100.00%0.00%100.00%100.00%0.00%99.99%100.00%0.00%100.00%100.00%0.00%100.00%100.00%0.00%
PUNC BERT (precision)60.63%92.24%82.85%67.02%94.54%83.10%37.35%86.51%92.70%70.85%95.13%78.00%58.96%92.11%84.16%
AdversarialDDPM-OOD mse.and.Lpips39.54%1.78%95.72%50.89%2.54%95.91%38.97%1.10%98.38%43.13%1.81%96.67%
LMD lipps alex43.71%1.95%96.24%53.29%2.89%94.43%41.64%1.17%97.48%46.21%2.00%96.05%
2XDM lipps alex36.94%1.94%98.92%47.95%2.54%96.03%53.04%2.83%95.69%55.19%1.96%96.20%48.28%2.32%96.71%
PUNC ROUGE (1_precision)64.14%97.77%48.80%76.32%98.90%48.00%63.59%97.73%49.50%62.97%97.64%49.20%66.76%98.01%48.87%
PUNC ROUGE (L_precision)63.20%97.75%49.00%76.26%98.90%48.40%64.09%97.90%49.70%63.76%97.87%49.60%66.83%98.11%49.17%
PUNC BERT (precision)73.98%98.72%78.31%72.95%98.93%86.40%45.09%96.93%94.18%72.69%98.80%81.40%66.18%98.34%85.07%
Corrupt l11DDPM-OOD mse.and.Lpips48.87%44.82%87.43%53.20%54.43%95.48%41.02%29.54%96.32%47.70%42.93%93.08%
LMD lipps alex51.23%47.48%89.99%57.53%56.49%90.92%40.94%30.30%96.32%49.90%44.76%92.41%
2XDM lipps alex25.48%36.16%99.29%50.43%49.86%90.10%54.97%54.29%95.69%54.73%38.27%93.29%46.40%44.65%94.59%
PUNC ROUGE (1_precision)28.75%38.34%98.77%54.58%55.29%96.02%27.69%37.57%98.78%27.45%37.27%98.54%34.62%42.12%98.03%
PUNC ROUGE (L_precision)27.29%38.11%99.15%54.32%55.59%96.16%27.30%38.07%99.45%27.35%37.89%99.37%34.07%42.42%98.53%
PUNC BERT (precision)85.85%84.45%73.94%79.43%78.83%87.21%54.97%53.73%94.51%75.76%72.91%84.65%74.00%72.48%85.08%
Corrupt l12DDPM-OOD lipps alex43.23%43.94%91.23%59.46%60.14%93.71%48.79%34.44%95.74%50.49%46.17%93.56%
LMD lipps alex46.90%45.65%92.04%59.94%58.90%90.54%40.88%30.42%96.35%49.24%44.99%92.98%
2XDM lipps alex58.15%58.50%93.98%52.08%53.36%90.19%53.57%54.17%96.34%60.24%44.23%91.78%56.01%52.57%93.07%
PUNC ROUGE (1_precision)28.76%38.10%98.59%54.77%55.49%96.02%27.66%37.42%98.75%27.39%37.23%98.53%34.64%42.06%97.97%
PUNC ROUGE (L_precision)27.33%37.87%99.16%54.51%55.87%96.11%27.31%37.92%99.35%27.33%37.83%99.36%34.12%42.37%98.50%
PUNC BERT (precision)85.96%84.23%73.54%79.47%78.80%87.05%54.78%53.51%94.79%76.04%73.44%84.58%74.06%72.49%84.99%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.191, + 0.68, + 0.8, + 0.694 + ], + "angle": 0, + "content": "Table 4. Performance comparison all applications (Vague, Adversarial, Corrupt Lvl1, Corrup Lvl2)" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.721, + 0.391, + 0.736 + ], + "angle": 0, + "content": "models with ethical and regulatory standards." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.743, + 0.483, + 0.849 + ], + "angle": 0, + "content": "Our contributions also include a dataset specifically tailored to uncertainty quantification in text-to-image models, which we hope will serve as a benchmark for future research. Ultimately, this work highlights the potential of prompt-based uncertainty estimation as a critical tool for enhancing the reliability and safety of generative models in real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Limitations. While PUNC demonstrates promising results in quantifying uncertainty for text-to-image generation, several limitations remain. First, our approach" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.812 + ], + "angle": 0, + "content": "relies heavily on the interpretive abilities of LVLMs, which, despite their effectiveness, may introduce biases or inaccuracies inherited from their training data. Second, our analysis focuses primarily on English-language prompts, limiting our method's applicability to multilingual or culturally diverse datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.83, + 0.907, + 0.892 + ], + "angle": 0, + "content": "Acknowledgments. This work was performed using HPC resources from GENCI-IDRIS (Grant 2023-[AD011011970R3]), and EuroHPC Development access to LEONARDO, hosted by CINECA." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8069" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.483, + 0.297 + ], + "angle": 0, + "content": "[3] Nouar AlDahoul, Talal Rahwan, and Yasir Zaki. Aigenerated faces free from racial and gender stereotypes. arXiv preprint arXiv:2402.01002, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.484, + 0.382 + ], + "angle": 0, + "content": "[4] Georges Le Bellier and Nicolas Audebert. Detecting Out-Of-Distribution Earth Observation Images with Diffusion Models. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 481–491, Los Alamitos, CA, USA, 2024. IEEE Computer Society. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.483, + 0.437 + ], + "angle": 0, + "content": "[5] Lucas Berry, Axel Brando, and David Meger. Shedding light on large generative networks: Estimating epistemic uncertainty in diffusion models. In The 40th Conference on Uncertainty in Artificial Intelligence, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.483, + 0.493 + ], + "angle": 0, + "content": "[6] Matthew A Chan, Maria J Molina, and Christopher A Metzler. Hyper-diffusion: Estimating epistemic and aleatoric uncertainty with a single model. arXiv preprint arXiv:2402.03478, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.495, + 0.484, + 0.564 + ], + "angle": 0, + "content": "[7] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart\\(\\backslash\\)sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 3, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.484, + 0.648 + ], + "angle": 0, + "content": "[8] Junsong Chen, Jincheng YU, Chongjian GE, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-\\(\\mathbb{S}\\backslash \\mathbb{alpha}\\mathbb{S}\\): Fast training of diffusion transformer for photorealistic text-to-image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.484, + 0.704 + ], + "angle": 0, + "content": "[9] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In Proceedings of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.483, + 0.76 + ], + "angle": 0, + "content": "[10] Charles Corbière, Nicolas Thome, Avner Bar-Hen, Matthieu Cord, and Patrick Pérez. Addressing failure prediction by learning model confidence. Advances in Neural Information Processing Systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[11] Matt Deitke, Christopher Clark, Sangho Lee, Rohun Tripathi, Yue Yang, Jae Sung Park, Mohammadreza Salehi, Niklas Muennighoff, Kyle Lo, Luca Soldaini, et al. Molmo and pixmo: Open weights and open data for state-of-the-art multimodal models. arXiv preprint arXiv:2409.17146, 2024. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.845, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7, 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[13] Ning Ding, Yehui Tang, Zhongqian Fu, Chao Xu, Kai Han, and Yunhe Wang. Gpt4image: Can large pre-trained models help vision models on perception tasks? arXiv e-prints, pages arXiv-2306, 2023. 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.206 + ], + "angle": 0, + "content": "[14] Zhekai Du and Jingjing Li. Diffusion-based probabilistic uncertainty estimation for active domain adaptation. Advances in Neural Information Processing Systems, 36: 17129-17155, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[15] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.281, + 0.905, + 0.322 + ], + "angle": 0, + "content": "[16] Nikita Durasov, Timur Bagautdinov, Pierre Baque, and Pascal Fua. Masksembles for uncertainty estimation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.324, + 0.905, + 0.393 + ], + "angle": 0, + "content": "[17] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.397, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[18] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In Proceedings of the 41st International Conference on Machine Learning, pages 12606-12633. PMLR, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[19] Gianni Franchi, Andrei Bursuc, Emanuel Aldea, Séverine Dubuisson, and Isabelle Bloch. TRADI: Tracking deep neural network weight distributions. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.541, + 0.905, + 0.61 + ], + "angle": 0, + "content": "[20] Gianni Franchi, Xuanlong Yu, Andrei Bursuc, Emanuel Aldea, Severine Dubuisson, and David Filliat. Latent discriminant deterministic uncertainty. In Computer Vision – ECCV 2022, pages 243–260, Cham, 2022. Springer Nature Switzerland. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.612, + 0.905, + 0.642 + ], + "angle": 0, + "content": "[21] Yarin Gal. Uncertainty in Deep Learning. PhD thesis, University of Cambridge, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.643, + 0.905, + 0.712 + ], + "angle": 0, + "content": "[22] Yonatan Geifman and Ran El-Yaniv. Selective classification for deep neural networks. In Proceedings of the 31st International Conference on Neural Information Processing Systems, page 4885-4894, Red Hook, NY, USA, 2017. Curran Associates Inc. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.714, + 0.905, + 0.798 + ], + "angle": 0, + "content": "[23] Mark S Graham, Walter HL Pinaya, Petru-Daniel Tudosiu, Parashkev Nachev, Sebastien Ourselin, and Jorge Cardoso. Denoising diffusion models for out-of-distribution detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2948-2957, 2023. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.801, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[24] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In International Conference on Learning Representations, 2017. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.858, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "8070" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[26] Eyke Hüllermeier and Willem Waegeman. *Aleatoric and epistemic uncertainty in machine learning: an introduction to concepts and methods*. *Machine Learning*, 110(3), 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[27] Pavel Izmailov, Sharad Vikram, Matthew D Hoffman, and Andrew Gordon Gordon Wilson. What are bayesian neural network posteriors really like? In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[28] John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, et al. Highly accurate protein structure prediction with alphafold. nature, 596(7873):583-589, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[29] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems, pages 26565-26577. Curran Associates, Inc., 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[30] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[31] Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. In The Eleventh International Conference on Learning Representations, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.441 + ], + "angle": 0, + "content": "[32] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. In NeurIPS, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[33] Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, et al. Learning skillful medium-range global weather forecasting. Science, 382(6677):1416-1421, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.566 + ], + "angle": 0, + "content": "[34] Olivier Laurent, Adrien Lafage, Enzo Tartaglione, Geoffrey Daniel, Jean-Marc Martinez, Andrei Bursuc, and Gianni Franchi. Packed-ensembles for efficient uncertainty estimation. In ICLR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.622 + ], + "angle": 0, + "content": "[35] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 7, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[36] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.482, + 0.72 + ], + "angle": 0, + "content": "[37] Zhiqiu Lin, Xinyue Chen, Deepak Pathak, Pengchuan Zhang, and Deva Ramanan. Revisiting the role of language priors in vision-language models. In *Forty-first International Conference on Machine Learning*. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.721, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[38] Krzysztof Lis, Krishna Nakka, Pascal Fua, and Mathieu Salzmann. Detecting the unexpected via image resynthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2152-2161, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[39] Zhenzhen Liu, Jin Peng Zhou, Yufan Wang, and Kilian Q Weinberger. Unsupervised out-of-distribution detection with diffusion inpainting. In International Conference on Machine Learning, pages 22528-22538. PMLR, 2023. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[40] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan LI, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. In Advances in Neural Information Processing Systems, pages 5775-5787. Curran Associates, Inc., 2022. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[41] Andrey Malinin and Mark Gales. Predictive uncertainty estimation via prior networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.136, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[42] Radford M Neal et al. Mcmc using hamiltonian dynamics. Handbook of markov chain monte carlo, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[43] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[44] Gangadhar Payyavula. Nuclei segmentation in microscope cell images. https://www.kaggle.com/datasets/gangadhar/nuclei-segmentation-in-microscope-cell-images, 2018. Accessed: 2024-09-01. 7, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.334 + ], + "angle": 0, + "content": "[45] Tim Pearce, Alexandra Brintrup, and Jun Zhu. Understanding softmax confidence and uncertainty, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[46] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[47] Jeremy Pinto. Bigearthnet. https://github.com/ jerpint/bigearthnet, 2022. Accessed: 2024-09-01. 7, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.905, + 0.505 + ], + "angle": 0, + "content": "[48] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2023. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.506, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[49] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Proceedings of the 38th International Conference on Machine Learning, pages 8821-8831. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.646 + ], + "angle": 0, + "content": "[50] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[51] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[52] Jaskirat Singh and Liang Zheng. Divide, evaluate, and refine: Evaluating and improving text-to-image alignment with iterative vqa feedback. Advances in Neural Information Processing Systems, 36:70799-70811, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[53] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[54] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 4" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "8071" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[55] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. arXiv preprint arXiv:2409.17146, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.148, + 0.482, + 0.231 + ], + "angle": 0, + "content": "[56] Tomas Vojir, Tomáš Šipka, Rahaf Aljundi, Nikolay Chumerin, Daniel Olmeda Reino, and Jiri Matas. Road anomaly detection by partial image reconstruction with segmentation coupling. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15651-15660, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.232, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[57] Max Welling and Yee W Teh. Bayesian learning via stochastic gradient Langevin dynamics. In ICML, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[58] Yeming Wen, Dustin Tran, and Jimmy Ba. BatchEnsemble: an alternative approach to efficient ensemble and lifelong learning. In ICLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[59] Guoxuan Xia and Christos-Savvas Bouganis. Augmenting softmax information for selective classification with out-of-distribution data. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 1995-2012, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.398 + ], + "angle": 0, + "content": "[60] Guoxuan Xia and Christos-Savvas Bouganis. On the usefulness of deep ensemble diversity for out-of-distribution detection, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[61] Guoxuan Xia and Christos-Savvas Bouganis. Window-based early-exit cascades for uncertainty estimation: When deep ensembles are more efficient than single models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17368-17380, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.537 + ], + "angle": 0, + "content": "[62] Guoxuan Xia, Duolikun Danier, Ayan Das, Stathi Fotiadis, Farhang Nabiei, Ushnish Sengupta, and Alberto Bernacchia. Score Normalization for a Faster Diffusion Exponential Integrator Sampler. In NeurIPS 2023 Workshop on Diffusion Models, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[63] Guoxuan Xia, Olivier Laurent, Gianni Franchi, and Christos-Savvas Bouganis. Towards understanding why label smoothing degrades selective classification and how to fix it, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[64] Yingda Xia, Yi Zhang, Fengze Liu, Wei Shen, and Alan L Yuille. Synthesize then compare: Detecting failures and anomalies for semantic segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 145-161. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[65] Yilun Xu, Mingyang Deng, Xiang Cheng, Yonglong Tian, Ziming Liu, and Tommi S. Jaakkola. Restart sampling for improving generative processes. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[66] Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. Generalized out-of-distribution detection: A survey. International Journal of Computer Vision, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[67] Michal Yarom, Yonatan Bitton, Soravit Changpinyo, Roeah Aharoni, Jonathan Herzig, Oran Lang, Eran Ofek, and Idan Szpektor. What you see is what you read? improving text-image alignment evaluation. Advances in Neural Information Processing Systems, 36:1601-1619, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[68] Xuanwu Yin Yuda Song, Zehao Sun. Sdxs: Real-time one-step latent diffusion models with image conditions. arxiv, 2024. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[69] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. In The Eleventh" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "International Conference on Learning Representations, 2023.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[70] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[71] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[72] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675, 2019. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[73] Yimeng Zhang, Jinghan Jia, Xin Chen, Aochuan Chen, Yihua Zhang, Jiancheng Liu, Ke Ding, and Sijia Liu. To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images ... for now, 2024. 7, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8072" + } + ] +] \ No newline at end of file diff --git a/2025/Towards Understanding and Quantifying Uncertainty for Text-to-Image Generation/c3ad3744-8ee2-4117-a4a1-7d23c58e3fe4_origin.pdf b/2025/Towards Understanding and Quantifying Uncertainty for Text-to-Image Generation/c3ad3744-8ee2-4117-a4a1-7d23c58e3fe4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2855aaa2233c637494cd4a0c0b8c1e52a19b0376 --- /dev/null +++ b/2025/Towards Understanding and Quantifying Uncertainty for Text-to-Image Generation/c3ad3744-8ee2-4117-a4a1-7d23c58e3fe4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7948aacacc50c00553efe146feebba3aa782fd6102d0378a179d6e9822c9079a +size 4573219 diff --git a/2025/Towards Understanding and Quantifying Uncertainty for Text-to-Image Generation/full.md b/2025/Towards Understanding and Quantifying Uncertainty for Text-to-Image Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..877c3c69f03ec41634de79d0582ba3b532b3744d --- /dev/null +++ b/2025/Towards Understanding and Quantifying Uncertainty for Text-to-Image Generation/full.md @@ -0,0 +1,310 @@ +# Towards Understanding and Quantifying Uncertainty for Text-to-Image Generation + +Gianni Franchi, $^{1}$ * Nacim Belkhir, $^{2}$ * Dat Nguyen Trong, $^{1}$ Guoxuan Xia, $^{3}$ Andrea Pilzer $^{4}$ +U2IS, ENSTA Paris, Institut Polytechnique de Paris, $^{1}$ Mirai, $^{2}$ Imperial College London, $^{3}$ NVIDIA $^{4}$ + +![](images/0df14e7c0fe8aff5150a47ddde96930d88820f306268a239773e70e5602c43fa.jpg) +Figure 1. Examples of Applications for Uncertainty Quantification in Text-to-Image Generation. Text-to-image generation models may exhibit uncertainty, and that need to be quantified since it can provide insights into the model's training dataset, aiding in deepfake prevention, detecting model biases, and protecting copyrighted content from unauthorized generation. + +![](images/e4cb42dfba865827dd7ac22e034ddfa0a402644ec557b371ebb37dce40429cc8.jpg) + +# Abstract + +Uncertainty quantification in text-to-image (T2I) generative models is crucial for understanding model behavior and improving output reliability. In this paper, we are the first to quantify and evaluate the uncertainty of T2I models with respect to the prompt. Alongside adapting existing approaches designed to measure uncertainty in the image space, we also introduce Prompt-based Uncertainty Estimation for T2I models (PUNC), a novel method leveraging Large Vision-Language Models (LVLMs) to better address uncertainties arising from the semantics of the prompt and generated images. PUNC utilizes a LVLM to caption a gen + +erated image, and then compares the caption with the original prompt in the more semantically meaningful text space. PUNC also enables the disentanglement of both aleatoric and epistemic uncertainties via precision and recall, which image-space approaches are unable to do. Extensive experiments demonstrate that PUNC outperforms state-of-the-art uncertainty estimation techniques across various settings. Uncertainty quantification in text-to-image generation models can be used on various applications including bias detection, copyright protection, and OOD detection. We also introduce a comprehensive dataset of text prompts and generation pairs to foster further research in uncertainty quantification for generative models. Our findings illustrate that PUNC not only achieves competitive performance but also enables novel applications in evaluating + +and improving the trustworthiness of text-to-image models. The code is available at https://github.com/ENSTA-U2IS-AI/Uncertainty_diffusion + +# 1. Introduction + +Generative models have revolutionized artificial intelligence, making significant strides in fields such as robotics [2], healthcare [28], and physics [33]. These models have elevated the performance of AI systems by enabling the generation of complex data representations. However, while generative models offer immense potential, they do not address all challenges, especially those related to robustness and uncertainty in Deep Neural Networks(DNNs)—issues that have become even more pronounced in generative settings. + +In this paper, we focus on the uncertainty quantification in deep generative models, in particular, we are the first to investigate modern text-to-image (T2I) models where the input is a natural language prompt [8, 18, 48-50]. These models have surged in popularity in recent years due to their capacity to generate high-quality images conveniently in a user-guided fashion, finding usage in a wide array of real-world applications. We are the first to investigate and propose a dedicated approach for quantifying uncertainty in text-to-image generation, aiming to fill a critical gap in generative AI research. The task of quantifying uncertainty in these models is not only complex but also crucial. Understanding the uncertainty in generative models can be invaluable for applications such as Out-of-Distribution (OOD) detection, trustworthy AI, and decision-making in high-stakes domains. To the best of our knowledge, no existing work has comprehensively addressed the challenge of uncertainty quantification in text-to-image generative models. However, given the semantic nature of the uncertainty we aim to measure, our work is closely related to recent studies on text-to-image alignment [37, 52, 67]. + +Existing research explores uncertainty estimation and OOD detection in the image/output space [4, 5, 23, 39]. For example, given a test image, can we determine whether or not it is OOD? On the other hand, the uncertainty of a generation with respect to conditioning (such as text) is underexplored. We argue that this is an important area to be addressed, as it represents a widespread real-world application of deep generative models. In this case, it remains an open question as to how should we define uncertainty, and given a definition, how should we quantify it? As illustrated in Fig. 1, focusing on uncertainty quantification in text-to-image generation could enable novel applications, allowing us to extract and interpret the knowledge embedded within models using language. + +To answer these questions, we first argue that uncertainties should be rooted in the semantics of the image + +and the prompt. We then adapt and evaluate a number of existing image-space approaches, as well as proposing a novel method for uncertainty quantification in T2I generative models, which we refer to as "Prompt-based Uncertainty Estimation for T2I Generation" (PUNC). Our approach leverages the growing capabilities of Large Vision-Language Models (LVLMs) to extract the semantics from generated images. LVLMs, which have been trained on vast amounts of text and visual data, can serve as powerful tools for interpreting the underlying meaning of text/image inputs and assessing how variations in phrasing or semantics impact the uncertainty of the generated outputs. Our hypothesis is that as LVLMs become increasingly adept at understanding complex relationships between text and images, they will allow us to better analyze and quantify uncertainty. Through this paper, we will demonstrate Prompt-based Uncertainty Estimation for T2I (PUNC) effectiveness in quantifying uncertainties in generative models. Showcasing its ability to decompose and interpret different types of uncertainties arising from the multimodal nature of the input, offering a new perspective on uncertainty quantification in generative AI. + +To summarize, our contributions are as follows: (1) We introduce a new task aimed at quantifying uncertainty (with respect to the prompt) in text-to-image generation models, and we share a dataset of prompts to support future research on uncertainty in these models. (2) We both adapt existing image-space approaches to our tasks, as well as propose a novel straightforward method, PUNC, designed to better quantify semantic uncertainty in text-to-image generation. (3) Extensive experiments demonstrate that PUNC, despite its simplicity and computational efficiency, achieves competitive performance. (4) We highlight the practical utility of uncertainty quantification in text-to-image generation by showcasing additional applications where this approach could be beneficial. + +# 2. Related Work + +Deep learning models typically encounter two primary types of uncertainty: aleatoric and epistemic uncertainty [26]. Over the years, a variety of methods have been developed to quantify these uncertainties. Ensemble methods [16, 32, 34, 58, 61], for instance, have often achieved state-of-the-art performance by leveraging multiple models to assess prediction confidence. However, these methods tend to be computationally expensive due to the need to train and maintain several models. Another approach involves Bayesian methods [21], especially Bayesian Neural Networks (BNNs) [19, 27, 42, 57], which offer a theoretically grounded way to quantify uncertainty. Additionally, there are techniques aimed at directly predicting uncertainty within DNNs [10, 20, 30]. Despite their effectiveness, most of these methods are challenging to apply to text-to-image + +![](images/f22022d38cc577c40f1d07107d04862ce1c9139eb16d6a4dce45345b8c7c4a8d.jpg) +Generation/Image-space Uncertainty (existing work) +Figure 2. Diagram illustrating generation/image-space uncertainty which is considered in previous work and condition/prompt-space uncertainty which is investigated in our work. + +![](images/61e51754cec18673e6e57272b130161e06703729facb467446902ddcc7dc5863.jpg) +Figure 3. Three generations from PixArt- $\Sigma$ [7] illustrating uncertainty with regards to prompt semantics. For the Normal image, we used an ImageNet-inspired prompt; for the corrupted image, additional corruption was applied to the prompt (e.g. fish was perturbed to fis) increasing the aleatoric uncertainty; and for the out-of-distribution (OOD) case, the model was prompted to generate an image of the Prime Minister of Japan Kishida Fumio under similar conditions, where epistemic uncertainty was injected in the form of a semantic concept the model was not familiar with. + +generation, as they rely on ensembling predictions or accessing model uncertainty in ways that are difficult to implement for generative models without explicit access to intermediate predictions. + +Generative models, as a subset of deep learning, also exhibit uncertainty. In the past, traditional generative models were employed for uncertainty quantification and used to help estimate the uncertainty within standard DNNs [6, 38, 56, 64]. With the recent popularity of diffusion models, research has begun exploring whether image-based diffusion models can detect out-of-distribution (OOD) samples [4, 23, 39] or quantify their own uncertainty [5, 6, 14]. Whilst these approaches align with our goals, they are limited to image-space tasks, where the input is an image rather than a textual prompt. This input difference fundamentally alters the nature of uncertainty in the model. Our approach, in contrast, specifically addresses text-to-image generation, a setting where prompt ambiguity and multimodal complexity introduce new dimensions of uncertainty. We demonstrate that quantifying uncertainty in this context can lead to valuable insights and applications across various downstream tasks. + +# 3. Uncertainty in Image Generation Models + +Consider a deep generative model $p_{\theta}(\boldsymbol{x})$ with parameters $\theta$ that models the data distribution $p_{\mathrm{data}}(\boldsymbol{x})$ of $\boldsymbol{x} \in \mathbb{R}^D$ . We may be interested in measuring the uncertainty of our model $p_{\theta}$ with respect to some test datum $\boldsymbol{x}^*$ . For example, consider the case where we want to determine whether or not $\boldsymbol{x}^*$ was drawn from $p_{\mathrm{data}}(\boldsymbol{x})$ , i.e. out-of-distribution (OOD) detection. This can be achieved by quantifying the uncertainty of $p_{\theta}$ that arises from a lack of knowledge of $\boldsymbol{x}^*$ , i.e. epistemic uncertainty [26]. A range of existing work that leverages diffusion-based generative models for OOD image detection [23, 39] fall into the above paradigm. From hereon we will refer to this as generation/image-space uncertainty. This is illustrated in Fig. 2 (top). + +In this work, however, we are primarily interested in a different scenario: text-to-image generation. Consider instead a conditional model $p_{\theta}(\pmb{x}|\pmb{c})$ that models the conditional data distribution $p_{\mathrm{data}}(\pmb{x}|\pmb{c})$ . This describes real-world text-to-image generation, where image $\pmb{x}$ is drawn from $p_{\theta}(\pmb{x}|\pmb{c}^*)$ conditioned on some test text prompt $\pmb{c}^*$ . In this case, we wish to quantify the uncertainty of $p_{\theta}$ not with respect to some test $\pmb{x}^*$ (e.g. image) but instead with respect to a test condition $\pmb{c}^*$ . We will hereon refer to this as condition/prompt-space uncertainty. This is illustrated in Fig. 2 (bottom). + +We argue that in this case, uncertainties should be concerned with semantics, as intuitively, these are what inherently link the user's intentions, represented by the text prompt, and the generated output image. This is similar to recent discourse on uncertainty estimation for large language models [31]. If we consider the semantic concepts in the prompt and generated image, then we can define useful concepts of uncertainty (illustrated in Fig. 3): + +- Aleatoric uncertainty is irreducible uncertainty in the data distribution $p_{\mathrm{data}}(\boldsymbol{x}|\boldsymbol{c})$ . From the perspective of semantics, high aleatoric uncertainty would be where a variation of generated concepts may arise from a single prompt. For example, a spelling mistake where fish is mistyped as fis in the prompt may result in either a "fish" or a "fist" appearing in the image. +- Epistemic uncertainty should correspond to a model's lack of knowledge of semantic concepts found in a prompt. For example, a model trained on ImageNet will not know what the prime minister of Japan Kishida Fumio looks like, and thus will have high epistemic uncertainty for this semantic concept. + +We note that broadly speaking uncertainty estimation for deep conditional models is well explored [20-22, 24, 26, 32, 34, 41, 59-61, 63, 66]. However, research in this area is generally concerned with problem settings where 1) the + +models explicitly evaluate $p_{\theta}(\pmb{x}|c)$ and 2) downstream applications of uncertainty relate directly to the numerical values of $\pmb{x}$ and $p_{\theta}(\pmb{x}|c)$ . For example, a cross-entropy-trained image classifier explicitly predicts the probability mass $P_{\theta}(\pmb{x}|c^{*})$ of label $\pmb{x}$ given test image $c^{*}$ . $P_{\theta}(\pmb{x}|c^{*})$ directly relates to the likelihood of classification error – which is an event that we may want to detect using an uncertainty measure, i.e. downstream misclassification detection [22, 24]. Thus, we are able to easily extract useful measures of uncertainty from the model $P_{\theta}$ , e.g. the maximum of the softmax. + +On the other hand, text-to-image generation models are not as clear-cut. They may not explicitly evaluate $p_{\theta}(\pmb{x}|c^{*})$ (e.g. diffusion-based models [25, 54]), and even when they do (e.g. vector-quantised autoregressive models [17, 55]) it is hard to directly relate (latent) pixel values directly to our above definitions of uncertainty. For example, simple variations in contrast and brightness that may leave semantics completely unchanged could result in a very high variance in $p_{\theta}(\pmb{x}|c^{*})$ . To give a simple example, inverting the black and white stripes on a zebra would lead to a large Euclidean distance in the image space whilst preserving the semantics of the image. In the next section, we present and discuss a range of potential methods to quantify the uncertainty of text-to-image models with respect to a given prompt. + +# 4. Techniques to Quantify Uncertainty + +We quantify text-to-image uncertainty by adapting existing methods and proposing a novel approach named Prompt-based UNCertainty estimation (PUNC)(Fig. 4). + +# 4.1. Adapting Existing Approaches + +There are a number of existing methods for quantifying the image-space uncertainty of diffusion models. As diffusion-based models represent a majority of current text-to-image foundation models [8, 18, 48, 50], we choose to adapt these in order to quantify prompt-space uncertainty (Fig. 4). We first present a brief primer on diffusion models to the reader. + +# 4.1.1. Diffusion Models + +Diffusion models [25] generate data by iteratively denoising pure noise. The forward process over time $t \in [0,1]$ for $\boldsymbol{x}_t \in \mathbb{R}^D$ , conditioned on a text prompt $c$ , follows: + +$$ +p \left(\boldsymbol {x} _ {t} \mid \boldsymbol {x} _ {0}, \boldsymbol {c}\right) = \mathcal {N} \left(\boldsymbol {x} _ {t}; a _ {t} \boldsymbol {x} _ {0}, \sigma_ {t} ^ {2} \boldsymbol {I}\right), \tag {1} +$$ + +where $\pmb{x}_0$ is sampled from an unknown data distribution $p(\pmb{x}_0)$ , and $a_{t},\sigma_{t}$ define the noise schedule. This formulation allows diffusion models to progressively refine noisy inputs, enabling high-quality data generation through a learned denoising process. By leveraging stochastic differential equations (SDEs) or their deterministic counterparts, + +these models sample from complex distributions in a structured manner. Various solvers [29, 40, 53, 54, 62, 65, 69] can then be employed to efficiently generate realistic samples with a limited number of denoising steps. + +Latent diffusion models (LDMs) [39] perform this process within a compressed latent space rather than directly on pixel space. By first encoding images into a more compact representation, LDMs reduce computational demands, enabling high-quality high-resolution generations. We will not consider the latent representation to keep the notation simple, even though experiments will use LDMs. + +# 4.1.2. Time-step based approaches + +DDPM-OOD [23]. This approach, developed for image-space OOD detection, involves noising a test sample $\boldsymbol{x}_0^*$ to different points in the forward process to get $\{\boldsymbol{x}_{t_1}^*,\boldsymbol{x}_{t_2}^*,\dots \}$ and denoising them, generating multiple reconstructions $\{\hat{x}_{0_1},\hat{x}_{0_2},\ldots \}$ . Similarity is then measured between each reconstruction and the original image. The idea is that test samples $\boldsymbol{x}_0^*$ for which the model has higher epistemic uncertainty will lead to reconstructions that differ from the test sample. However, in our case, we do not have a test image $\boldsymbol{x}_0^*$ . In order to adapt DDPM-OOD we propose measuring similarity between reconstructions at different $t$ and a baseline generation, $S(\hat{x}_{0_i},\hat{x}_0)$ , where $\hat{x}_0$ represents an image directly generated using the original test prompt $c^*$ alone. Both aleatoric uncertainty (e.g. ambiguous prompt) and epistemic uncertainty (e.g. unfamiliar concepts in prompt) could lead to more divergent reconstructions. + +LMD [39]. Another similar approach for image-space OOD detection, known as Learned Manifold Denoising (LMD), instead explores alternative transforms on $\boldsymbol{x}_0^*$ to noising. The idea is to lift $\boldsymbol{x}_0^*$ from its original data manifold and then map it back using the trained diffusion model. If the sample is in-distribution, the mapping process should return it close to its original position on the data manifold. Conversely, if it is OOD, the mapped position will likely diverge significantly from the original. The corruption process typically involves masking portions of the image, with various mask patterns applied. Similar to DDPM-OOD, similarity is measured between the original image and the reconstructed version after applying the diffusion process. Similar to before, since we lack a testimage, we measure the similarity between reconstructed images and $\hat{\boldsymbol{x}}_{0_i}$ , the image generated directly from the prompt without perturbations. + +Similarity Metric. Various methods are used to evaluate similarity between images. Mean-squared error (MSE) is commonly applied between input and reconstruction, while some approaches favor the Learned Perceptual Image Patch Similarity (LPIPS) [70] metric. The choice of similarity metric is a crucial component of these techniques. + +![](images/028fd6b0c4c53b8460bc355b984d6d2584ea5339d9111611314b3422a07d66e0.jpg) +Figure 4. Illustration showing the different baselines and PUNC. PUNC leverages a LVLM to describe generated images and assess similarity with the original prompt, providing a refined uncertainty score. In contrast, baseline methods employ traditional techniques such as noise injection, ensembling, or masking to quantify uncertainty, followed by image-based similarity scoring. + +# 4.1.3. Test-Time Ensembling Approaches + +Test-time ensembling techniques take inspiration from approaches used in the large language model (LLM) community [31], where multiple generations are produced to assess consistency. Given a prompt $c$ , we generate two images $\{\hat{x}_1, \hat{x}_2\}$ and measure the similarity between the two outputs. We denote this technique 2XDM. The underlying rationale is that if the model has high confidence, it will generate similar images. However, if uncertainty is high, it will generate substantially different images. We note that this is similar to the approach proposed in [5], however, that approach uses an ensemble of diffusion models, which is not practical in the setting where we want to quantify the uncertainty of a single pretrained T2I model. + +A key question remains regarding how to find the best measure similarity, and we propose using the same techniques applied in the time-step-based methods. A significant advantage of test-time ensembling approaches is that they can be applied generically to all T2I models. In contrast, timestep-based approaches require a multi-step generation process, limiting the broadness of their applicability. We note that all above approaches require multiple generations, incurring a notable computational overhead at inference time. For example, DDPM-OOD requires $\sim 50\times$ the number of network inferences as a single generation (if the hyperparameters of the original paper are used) [23]. + +# 4.2. Prompt-based Uncertainty Estimation + +In this section, we propose a new method, Prompt-based Uncertainty estimation for text-to-image generation (PUNC), which leverages Large Vision-Language Models (LVLM) in conjunction with T2I models. We aim to assess the uncertainty in generated images by analyzing the alignment between initial text prompts and reconstructed textual descriptions of the generated images. + +# Background on Large Vision-Language Model (LVLM) + +A LVLM integrates a large language model (LLM) with + +an image encoder, allowing for robust, multimodal understanding. Given a text prompt $c$ and an image $x$ , a LVLM model, with parameters $\omega$ , includes: + +- Image Encoder: $f_{\omega}^{\mathrm{img}}(\cdot)$ processes the input image $x$ and produces an embedding: $z^{\mathrm{img}} = f_{\omega}^{\mathrm{img}}(x)$ +- Text Processor with LLM: $f_{\omega}^{\mathrm{txt}}(\cdot, \cdot)$ takes the prompt $c$ and the image embedding $z^{\mathrm{img}}$ as inputs, generating a descriptive or answer-based response: $\hat{c} = f_{\omega}^{\mathrm{txt}}(c, z^{\mathrm{img}})$ + +Thus, a LVLM system generates an interpretation $\hat{c}$ that reflects the content of the image given an initial prompt. + +Prompt-Based Uncertainty Estimation for T2I Generation (PUNC) The previously introduced approaches to uncertainty estimation rely on measuring similarity between images. For a given text prompt $c$ , a T2I model generates an image $x \sim p_{\theta}(x|c)$ . We reason that, even under low uncertainty, the model can produce images that vary in aspects such as color, texture, or object positioning, while still faithfully representing the semantics of prompt $c$ . As previously discussed in Sec. 3, this may result in image-based similarity metrics failing to capture uncertainty at a semantic level, since large differences in the image space may not necessarily correspond to meaningful semantic differences. + +We propose that advanced LVLMs, such as Molmo [11], LLAMA 3 [15], and GPT-4 [1], are a strong candidate for extracting the semantics from an image, through describing an image via natural language. Building on this insight, our proposed method, PUNC, introduces a novel approach that bypasses direct image comparison and instead evaluates the alignment between generated images and the original prompts by calculating an uncertainty score in the more semantically meaningful text-space. The steps for implementing PUNC are as follows: + +Step 1: Initial Prompt and Image Generation Given a test prompt $c^*$ , we use the T2I model to sample an image $x$ : + +$$ +\boldsymbol {x} \sim p _ {\boldsymbol {\theta}} (\boldsymbol {x} | \boldsymbol {c} ^ {*}) +$$ + +Step 2: Image Interpretation via LVLM With the + +![](images/bfd49245a54fb630faa3175b2d199995fca0356c85229b1aa37a4ac1798f6ca0.jpg) +Figure 5. Disentangling Aleatoric and Epistemic Uncertainty. + +generated image $\pmb{x}$ and the initial prompt $c^*$ , the LVLM produces a new descriptive caption $\hat{c}$ : + +$$ +\hat {\boldsymbol {c}} = f _ {\omega} ^ {\mathrm {t x t}} \left(\boldsymbol {c} ^ {*}, f _ {\omega} ^ {\mathrm {i m g}} (\boldsymbol {x})\right) +$$ + +Step 3: Uncertainty Score Calculation The core of PUNC lies in measuring the alignment between $\mathbf{c}$ and $\hat{\mathbf{c}}$ to assess the diffusion model's confidence. Specifically, the similarity score $S(\mathbf{c},\hat{\mathbf{c}})$ is computed as: + +$$ +S (\boldsymbol {c} ^ {*}, \hat {\boldsymbol {c}}) = \operatorname {s i m} (\boldsymbol {c} ^ {*}, \hat {\boldsymbol {c}}) +$$ + +High similarity suggests low uncertainty, indicating the generated image closely matches the initial prompt. Conversely, a lower score reveals higher uncertainty, signifying potential divergence from the prompt's intended meaning. + +Thus, PUNC is able to predict semantic uncertainties by utilising the powerful image understanding ability of LVLMs and computing similarities in the text space. + +Prompt Similarity: Aleatoric and Epistemic via Precision and Recall. In Sec. 3 we discussed the different types of semantic uncertainty for T2I generation. Here we will present a method for disentangling these uncertainties using the precision and recall of the caption given the prompt. Consider an abstract scenario where the prompt and the generated image each contain a discrete set of semantic concepts. We can then calculate the precision and recall of the image concepts with respect to the prompt: + +Intuitively, a lack of knowledge about concepts in the prompt (epistemic uncertainty) will result in fewer concepts being preserved in the image, reducing recall. On the other hand, a lack of specificity in the prompt (aleatoric uncertainty) will result in additional concepts being generated, reducing precision. Thus we can use this abstract form of semantic-concept-based recall to measure epistemic uncertainty, and use precision to measure aleatoric uncertainty. + +Although we cannot directly measure concepts, we can approximate them using natural language. Moreover, existing similarity scores for text already have the above notions of precision and recall integrated within [36, 43, 72], allowing PUNC to disentangle uncertainty by comparing the caption of the generated image with the original prompt. We note that the image-based approaches in + +Sec. 4.1 are unable to disentangle aleatoric from epistemic uncertainty. We focus on two different existing text-based similarity metrics, ROUGE [36], and BERTScore [72]. The former calculates similarity (precision or recall) by directly matching (sequences of) words between two text sequences. The latter on the other hand measures similarity in the deep feature space of a text encoder. + +As illustrated in Fig. 5, we define Prompt Similarity as the semantic overlap between the input prompt and the generated descriptive caption. More formally, let $P_{\mathrm{GT}}$ denote the set of semantic concepts in the ground-truth prompt used to generate an image, and $P_{\mathrm{VQA}}$ the set of concepts extracted from the LVLM-generated caption. The resulting uncertainties can be characterized as follows: + +Epistemic Uncertainty, which captures missing concepts in $P_{\mathrm{VQA}}$ that are present in $P_{\mathrm{GT}}$ , reflects an incomplete understanding of the intended semantics. It is defined as: + +$$ +U _ {\text {e p i s t e m i c}} = 1 - \frac {\left| P _ {\mathrm {G T}} \cap P _ {\mathrm {V Q A}} \right|}{\left| P _ {\mathrm {G T}} \right|} = 1 - \text {R e c a l l} \tag {2} +$$ + +Aleatoric Uncertainty, on the other hand, arises from spurious or extra concepts in $P_{\mathrm{VQA}}$ that do not appear in $P_{\mathrm{GT}}$ , indicating ambiguity in the model's response. It is defined as: + +$$ +U _ {\text {a l e a t o r i c}} = 1 - \frac {\left| P _ {\mathrm {G T}} \cap P _ {\mathrm {V Q A}} \right|}{\left| P _ {\mathrm {V Q A}} \right|} = 1 - \text {P r e c i s i o n} \tag {3} +$$ + +Further details on the computation of precision and recall can be found in Appendix A. + +# 5. Experiments & Discussions + +We test PUNC on Various task and various dataset and compared PUNC to DDPM-OOD [23], LMD [39] and 2XDM. We first introduce on Sec. 5 how we have constitute the dataset of prompts. Then on Sec. 5.1 we discuss of different results on these dataset. Then Appendix F explains how we can use the uncertainty of text-to-image model to detect if a model has been trained with some concept related to some fictional character of politicians. Appendix F.3 shows how uncertainty can help to detect bias of T2I models. To conduct our experiments and evaluate the various techniques for quantifying uncertainty, we utilized four T2I models: Stable Diffusion 1.5 (SDv1.5) [50], SDXS [68], SDXL [48], and PixArt- $\Sigma$ [7]. These diverse T2I models allow us to test a range of architectures, each with unique characteristics and potential weaknesses. SDv1.5, being one of the earliest models, likely has a smaller and less diverse training dataset compared to the others. In contrast, SDXS generates in a single-step, different to the multi-step processes in other models. + +Dataset of prompts To conduct our experiments, we require both an in-distribution dataset of prompts and + +datasets with uncertainty. For the in-distribution prompts, we use the dataset proposed by [13], where high-quality descriptions of ImageNet [12] images are generated using GPT-4[13], referred to as Normal. To simulate epistemic uncertainty, we create out-of-distribution (OOD) datasets using images from underrepresented domains like remote sensing, texture, and microscopic datasets, captioned with LLAVA Next[35], labeled as Remote Sensing, Texture, and Microscopic. For aleatoric uncertainty, we generate two datasets: Vague, containing 2,000 minimally descriptive prompts (e.g., "An image of ***"), and Corrupted, derived from Normal by adding grammatical errors (Level 1) and removing $50\%$ of words (Level 2). A detailed summary of these datasets is presented in Tab. 1 and Appendix E. In Appendix C, we provide the details of the hyperparameters used in our experiments. + +
Dataset TypeDataset NameDescription
In-Distribution (ID)Normal20k High-quality detailed prompts from ImageNet [12] images using GPT-4[13].
Out-of-Distribution (OOD)Remote Sensing Texture Microscopic20k Prompts from remote sensing images [47]5k Prompts from texture datasets10k Prompts from microscopic images [44]
Aleatoric UncertaintyVagueAdversarialCorrupted2k prompts with minimal context, e.g., "An image of ***," where "****" is an ImageNet class name, creating ambiguous input.1k altered prompts from Normal using Unlearn-DiffAtk [73]20k Altered prompts from Normal:• Level 1: Grammatical mistakes and spelling errors.• Level 2: Random removal of 50% of words in Level 1 prompts.
+ +Table 1. Summary of prompt datasets for unc. quantification. + +# 5.1. Results for Uncertainty Estimation + +Tab. 3 presents the results of PUNC when using Molmo as the LVLM to assess epistemic uncertainty in the T2I models. Our approach, PUNC, outperforms most state-of-the-art techniques on average. However, we observe more mixed results on the Texture dataset, likely due to uncertainty about the presence of texture images in the training data of the T2I models. + +In Tab. 4, we show the results of PUNC under aleatoric uncertainty, again using Molmo as the LVLM. Our approach outperforms most state-of-the-art methods in this setting as well. Notably, while the ROUGE score tends to perform better for epistemic uncertainty, the optimal metric is less clear for aleatoric uncertainty. Here, the choice between ROUGE and BERTScore may depend on the specific application: because corruption introduces spelling and grammatical errors, BERTScore—which captures semantic similarity—may be more suitable. For additional results using different LVLMs, please refer to Appendix D and Appendix B. In Appendix G, we present qualitative results. + +# 5.2. Applications of uncertainty + +Our study explores multiple applications of uncertainty quantification in T2I models, with detailed results provided + +in Appendix F. First, we investigate deepfake detection by examining the model's ability to generate recognizable images of prominent politicians (see Appendix F.1). We examined whether diffusion models could generate convincing images of G7 politicians. Using 100 prompts containing their names, we generated images with four diffusion models and queried a LVLM to determine if the intended politician was present in each image (yes/no), thus measuring concept recognition uncertainty. Results are presented in Table 2. + +Next, we address copyright concerns by evaluating whether T2I models can accurately replicate iconic, copyrighted characters such as Mickey Mouse and Darth Vader (see Appendix F.2). Finally, we analyze bias in T2I models by testing their capacity to generate individuals across different genders and races for specific job roles (see Appendix F.3). + +Using a LVLM, we evaluate the presence of specific individuals, copyrighted characters, or demographic concepts in the generated images, demonstrating how uncertainty estimation can aid in detecting potential misuses of generative models. These experiments illustrate the value of text-based uncertainty quantification in analyzing T2I model performance across ethically and legally sensitive applications, enabling assessments that are challenging to achieve with image-based generation alone. + +# 6. Conclusion + +In this work, we set out to be first to investigate uncertainty estimation T2I generation, where the uncertainty is with respect to the text prompt. We both adapt existing image-space uncertainty approaches, as well as present Prompt-based UNCertainty Estimation for T2I models (PUNC), the first approach designed specifically to quantify uncertainty in T2I generation. By leveraging the interpretive power of Large Vision-Language Models to extract semantics from a generated image, PUNC addresses the challenges associated with measuring multimodal uncertainty, enabling a deeper understanding of model behavior with respect to prompt-based conditioning. Our results validate that PUNC provides reliable, fine-grained assessments of aleatoric and epistemic uncertainties, outperforming existing methods that work for image generation. + +PUNC opens avenues for new applications, such as improving model robustness in biased or OOD scenarios and safeguarding against unauthorized content generation, including deepfakes and copyrighted characters. Our experiments show that diffusion models, while powerful, still exhibit significant variances in their ability to adhere to prompt details, particularly with ambiguous, corrupt prompts or OOD prompts. By identifying such uncertainties, PUNC not only supports the development of more trustworthy generative AI but also aids in aligning these + +
Justin TrudeauKing CharlesFumio KishidaOlaf ScholzEmmanuel MacronJoe BidenGiorgia MeloniAVERAGE
SDXL9494888898962195.5
SDv1.57860628894801878
SDXS8040708294762872.5
PixArt3676627894901574
AVERAGE7267.570.5849585.520.570.71
+ +Table 2. Accuracy (%) of diffusion models in generating recognizable images of G7 politicians. Higher accuracy indicates lower uncertainty in the model's ability to represent specific individuals. + +
auroc ↑SDXS aupr ↑fpr95↓auroc ↑PixArt aupr ↑fpr95↓auroc ↑SDv1.5 aupr ↑fpr95↓auroc ↑SDXL aupr ↑fpr95↓auroc ↑Average aupr ↑fpr95↓
microscopicDDPM-OOD mse_and_lpips39.61%56.38%100.00%71.64%71.52%93.62%87.25%83.67%51.43%66.17%70.53%81.68%
LMD lpips alex45.82%45.24%99.10%16.81%31.61%99.88%5.75%28.12%100.00%22.79%34.99%99.66%
2XDM lpips alex24.49%32.60%99.91%24.32%33.21%99.99%16.19%32.42%99.94%30.58%36.99%98.25%23.90%33.81%99.52%
PUNC ROUGE (1_recall)84.00%85.10%57.07%95.47%97.06%30.84%87.24%90.66%48.90%83.56%84.31%52.78%87.57%89.28%47.40%
PUNC ROUGE (L_recall)87.64%89.56%55.28%96.96%98.02%16.85%90.20%93.40%43.56%88.19%89.84%47.44%90.75%92.71%40.78%
PUNC BERT (recall)88.45%89.71%48.68%97.83%98.41%10.32%87.75%91.70%51.03%83.10%85.46%58.14%89.28%91.32%42.04%
remote sensingDDPM-OOD mse_and_lpips80.00%85.37%99.28%74.36%71.89%71.00%65.47%58.25%71.97%73.28%71.84%80.75%
LMD lpips alex90.65%93.11%74.49%78.93%79.31%81.05%86.60%85.39%49.19%85.39%85.94%68.24%
2XDM lpips alex53.88%59.21%98.23%54.43%57.99%98.78%30.66%38.80%98.76%23.54%35.35%96.23%40.63%47.83%98.00%
PUNC ROUGE (1_recall)85.59%86.59%63.87%94.14%95.97%49.29%73.24%72.93%79.56%73.07%74.54%83.48%81.51%82.51%69.05%
PUNC ROUGE (L_recall)89.85%90.63%51.96%97.01%97.94%16.08%83.63%84.65%65.35%85.68%87.09%64.00%89.04%90.08%49.34%
PUNC BERT (recall)80.50%82.56%75.19%91.93%93.08%43.30%61.27%64.74%90.08%60.71%65.32%90.94%73.60%76.42%74.88%
TextureDDPM-OOD mse_and_lpips73.22%62.02%89.24%70.21%55.97%99.80%74.24%51.97%84.71%72.56%56.65%91.25%
LMD lpips alex63.40%50.47%97.52%26.74%16.06%99.89%35.66%23.22%99.95%41.93%29.92%99.12%
2XDM lpips alex32.80%20.84%99.98%37.21%23.36%99.37%29.58%24.27%99.93%28.87%16.49%98.88%32.11%21.24%99.54%
PUNC ROUGE (1_recall)40.77%72.88%95.54%83.26%94.80%73.48%46.61%75.64%90.87%49.16%78.04%93.57%54.95%80.34%88.36%
PUNC ROUGE (L_recall)45.64%76.90%95.99%84.10%95.13%74.21%47.60%77.70%93.14%50.71%80.05%94.35%57.01%82.44%89.42%
PUNC BERT (recall)49.11%79.11%94.65%84.40%95.14%69.99%47.60%78.80%93.85%46.72%79.20%96.87%56.96%83.06%88.84%
+ +Table 3. Performance comparison of PUNC applied to different Out of Distribution concepts and different text-to-image models + +
auroc ↑SDXS aupr ↑fpr95↓auroc ↑PixArt aupr ↑fpr95↓auroc ↑SDv1.5 aupr ↑fpr95↓auroc ↑SDXL aupr ↑fpr95↓auroc ↑Average aupr ↑fpr95↓
VagueDDPM-OOD mse.and.Lpips31.41%6.06%96.45%50.06%9.30%96.52%40.59%5.36%97.37%40.69%6.91%96.78%
LMD lipps alex39.26%6.91%96.35%47.62%9.54%97.70%40.77%5.70%97.95%42.55%7.38%97.33%
2XDM lipps alex51.28%10.10%94.35%43.82%8.57%96.89%50.97%10.09%95.96%57.83%10.01%94.37%50.98%9.69%95.39%
PUNC ROUGE (1_precision)99.99%100.00%0.00%100.00%100.00%0.00%99.99%100.00%0.00%99.99%100.00%0.00%99.99%100.00%0.00%
PUNC ROUGE (L_precision)99.99%100.00%0.00%100.00%100.00%0.00%99.99%100.00%0.00%100.00%100.00%0.00%100.00%100.00%0.00%
PUNC BERT (precision)60.63%92.24%82.85%67.02%94.54%83.10%37.35%86.51%92.70%70.85%95.13%78.00%58.96%92.11%84.16%
AdversarialDDPM-OOD mse.and.Lpips39.54%1.78%95.72%50.89%2.54%95.91%38.97%1.10%98.38%43.13%1.81%96.67%
LMD lipps alex43.71%1.95%96.24%53.29%2.89%94.43%41.64%1.17%97.48%46.21%2.00%96.05%
2XDM lipps alex36.94%1.94%98.92%47.95%2.54%96.03%53.04%2.83%95.69%55.19%1.96%96.20%48.28%2.32%96.71%
PUNC ROUGE (1_precision)64.14%97.77%48.80%76.32%98.90%48.00%63.59%97.73%49.50%62.97%97.64%49.20%66.76%98.01%48.87%
PUNC ROUGE (L_precision)63.20%97.75%49.00%76.26%98.90%48.40%64.09%97.90%49.70%63.76%97.87%49.60%66.83%98.11%49.17%
PUNC BERT (precision)73.98%98.72%78.31%72.95%98.93%86.40%45.09%96.93%94.18%72.69%98.80%81.40%66.18%98.34%85.07%
Corrupt l11DDPM-OOD mse.and.Lpips48.87%44.82%87.43%53.20%54.43%95.48%41.02%29.54%96.32%47.70%42.93%93.08%
LMD lipps alex51.23%47.48%89.99%57.53%56.49%90.92%40.94%30.30%96.32%49.90%44.76%92.41%
2XDM lipps alex25.48%36.16%99.29%50.43%49.86%90.10%54.97%54.29%95.69%54.73%38.27%93.29%46.40%44.65%94.59%
PUNC ROUGE (1_precision)28.75%38.34%98.77%54.58%55.29%96.02%27.69%37.57%98.78%27.45%37.27%98.54%34.62%42.12%98.03%
PUNC ROUGE (L_precision)27.29%38.11%99.15%54.32%55.59%96.16%27.30%38.07%99.45%27.35%37.89%99.37%34.07%42.42%98.53%
PUNC BERT (precision)85.85%84.45%73.94%79.43%78.83%87.21%54.97%53.73%94.51%75.76%72.91%84.65%74.00%72.48%85.08%
Corrupt l12DDPM-OOD lipps alex43.23%43.94%91.23%59.46%60.14%93.71%48.79%34.44%95.74%50.49%46.17%93.56%
LMD lipps alex46.90%45.65%92.04%59.94%58.90%90.54%40.88%30.42%96.35%49.24%44.99%92.98%
2XDM lipps alex58.15%58.50%93.98%52.08%53.36%90.19%53.57%54.17%96.34%60.24%44.23%91.78%56.01%52.57%93.07%
PUNC ROUGE (1_precision)28.76%38.10%98.59%54.77%55.49%96.02%27.66%37.42%98.75%27.39%37.23%98.53%34.64%42.06%97.97%
PUNC ROUGE (L_precision)27.33%37.87%99.16%54.51%55.87%96.11%27.31%37.92%99.35%27.33%37.83%99.36%34.12%42.37%98.50%
PUNC BERT (precision)85.96%84.23%73.54%79.47%78.80%87.05%54.78%53.51%94.79%76.04%73.44%84.58%74.06%72.49%84.99%
+ +Table 4. Performance comparison all applications (Vague, Adversarial, Corrupt Lvl1, Corrup Lvl2) + +models with ethical and regulatory standards. + +Our contributions also include a dataset specifically tailored to uncertainty quantification in text-to-image models, which we hope will serve as a benchmark for future research. Ultimately, this work highlights the potential of prompt-based uncertainty estimation as a critical tool for enhancing the reliability and safety of generative models in real-world applications. + +Limitations. While PUNC demonstrates promising results in quantifying uncertainty for text-to-image generation, several limitations remain. First, our approach + +relies heavily on the interpretive abilities of LVLMs, which, despite their effectiveness, may introduce biases or inaccuracies inherited from their training data. Second, our analysis focuses primarily on English-language prompts, limiting our method's applicability to multilingual or culturally diverse datasets. + +Acknowledgments. This work was performed using HPC resources from GENCI-IDRIS (Grant 2023-[AD011011970R3]), and EuroHPC Development access to LEONARDO, hosted by CINECA. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 5 +[2] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2 +[3] Nouar AlDahoul, Talal Rahwan, and Yasir Zaki. Aigenerated faces free from racial and gender stereotypes. arXiv preprint arXiv:2402.01002, 2024. 6 +[4] Georges Le Bellier and Nicolas Audebert. Detecting Out-Of-Distribution Earth Observation Images with Diffusion Models. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 481–491, Los Alamitos, CA, USA, 2024. IEEE Computer Society. 2, 3 +[5] Lucas Berry, Axel Brando, and David Meger. Shedding light on large generative networks: Estimating epistemic uncertainty in diffusion models. In The 40th Conference on Uncertainty in Artificial Intelligence, 2024. 2, 3, 5 +[6] Matthew A Chan, Maria J Molina, and Christopher A Metzler. Hyper-diffusion: Estimating epistemic and aleatoric uncertainty with a single model. arXiv preprint arXiv:2402.03478, 2024. 3 +[7] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart $\backslash$ sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 3, 6, 2 +[8] Junsong Chen, Jincheng YU, Chongjian GE, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- $\mathbb{S}\backslash \mathbb{alpha}\mathbb{S}$ : Fast training of diffusion transformer for photorealistic text-to-image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 2, 4 +[9] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In Proceedings of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014. 2, 3 +[10] Charles Corbière, Nicolas Thome, Avner Bar-Hen, Matthieu Cord, and Patrick Pérez. Addressing failure prediction by learning model confidence. Advances in Neural Information Processing Systems, 32, 2019. 2 +[11] Matt Deitke, Christopher Clark, Sangho Lee, Rohun Tripathi, Yue Yang, Jae Sung Park, Mohammadreza Salehi, Niklas Muennighoff, Kyle Lo, Luca Soldaini, et al. Molmo and pixmo: Open weights and open data for state-of-the-art multimodal models. arXiv preprint arXiv:2409.17146, 2024. 5, 2 +[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7, 2 + +[13] Ning Ding, Yehui Tang, Zhongqian Fu, Chao Xu, Kai Han, and Yunhe Wang. Gpt4image: Can large pre-trained models help vision models on perception tasks? arXiv e-prints, pages arXiv-2306, 2023. 7, 2 +[14] Zhekai Du and Jingjing Li. Diffusion-based probabilistic uncertainty estimation for active domain adaptation. Advances in Neural Information Processing Systems, 36: 17129-17155, 2023. 3 +[15] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 5, 2 +[16] Nikita Durasov, Timur Bagautdinov, Pierre Baque, and Pascal Fua. Masksembles for uncertainty estimation. In CVPR, 2021. 2 +[17] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 4 +[18] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In Proceedings of the 41st International Conference on Machine Learning, pages 12606-12633. PMLR, 2024. 2, 4 +[19] Gianni Franchi, Andrei Bursuc, Emanuel Aldea, Séverine Dubuisson, and Isabelle Bloch. TRADI: Tracking deep neural network weight distributions. In ECCV, 2020. 2 +[20] Gianni Franchi, Xuanlong Yu, Andrei Bursuc, Emanuel Aldea, Severine Dubuisson, and David Filliat. Latent discriminant deterministic uncertainty. In Computer Vision – ECCV 2022, pages 243–260, Cham, 2022. Springer Nature Switzerland. 2, 3 +[21] Yarin Gal. Uncertainty in Deep Learning. PhD thesis, University of Cambridge, 2016. 2, 3 +[22] Yonatan Geifman and Ran El-Yaniv. Selective classification for deep neural networks. In Proceedings of the 31st International Conference on Neural Information Processing Systems, page 4885-4894, Red Hook, NY, USA, 2017. Curran Associates Inc. 3, 4 +[23] Mark S Graham, Walter HL Pinaya, Petru-Daniel Tudosiu, Parashkev Nachev, Sebastien Ourselin, and Jorge Cardoso. Denoising diffusion models for out-of-distribution detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2948-2957, 2023. 2, 3, 4, 5, 6 +[24] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In International Conference on Learning Representations, 2017. 3, 4 +[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4 + +[26] Eyke Hüllermeier and Willem Waegeman. *Aleatoric and epistemic uncertainty in machine learning: an introduction to concepts and methods*. *Machine Learning*, 110(3), 2021. 2, 3 +[27] Pavel Izmailov, Sharad Vikram, Matthew D Hoffman, and Andrew Gordon Gordon Wilson. What are bayesian neural network posteriors really like? In ICML, 2021. 2 +[28] John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, et al. Highly accurate protein structure prediction with alphafold. nature, 596(7873):583-589, 2021. 2 +[29] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems, pages 26565-26577. Curran Associates, Inc., 2022. 4 +[30] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 2 +[31] Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. In The Eleventh International Conference on Learning Representations, 2023. 3, 5 +[32] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. In NeurIPS, 2017. 2, 3 +[33] Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, et al. Learning skillful medium-range global weather forecasting. Science, 382(6677):1416-1421, 2023. 2 +[34] Olivier Laurent, Adrien Lafage, Enzo Tartaglione, Geoffrey Daniel, Jean-Marc Martinez, Andrei Bursuc, and Gianni Franchi. Packed-ensembles for efficient uncertainty estimation. In ICLR, 2023. 2, 3 +[35] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 7, 2, 3 +[36] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6, 1 +[37] Zhiqiu Lin, Xinyue Chen, Deepak Pathak, Pengchuan Zhang, and Deva Ramanan. Revisiting the role of language priors in vision-language models. In *Forty-first International Conference on Machine Learning*. 2 +[38] Krzysztof Lis, Krishna Nakka, Pascal Fua, and Mathieu Salzmann. Detecting the unexpected via image resynthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2152-2161, 2019. 3 +[39] Zhenzhen Liu, Jin Peng Zhou, Yufan Wang, and Kilian Q Weinberger. Unsupervised out-of-distribution detection with diffusion inpainting. In International Conference on Machine Learning, pages 22528-22538. PMLR, 2023. 2, 3, 4, 6 +[40] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan LI, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. In Advances in Neural Information Processing Systems, pages 5775-5787. Curran Associates, Inc., 2022. 4 + +[41] Andrey Malinin and Mark Gales. Predictive uncertainty estimation via prior networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 3 +[42] Radford M Neal et al. Mcmc using hamiltonian dynamics. Handbook of markov chain monte carlo, 2011. 2 +[43] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6 +[44] Gangadhar Payyavula. Nuclei segmentation in microscope cell images. https://www.kaggle.com/datasets/gangadhar/nuclei-segmentation-in-microscope-cell-images, 2018. Accessed: 2024-09-01. 7, 2, 3 +[45] Tim Pearce, Alexandra Brintrup, and Jun Zhu. Understanding softmax confidence and uncertainty, 2021. 3 +[46] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2 +[47] Jeremy Pinto. Bigearthnet. https://github.com/ jerpint/bigearthnet, 2022. Accessed: 2024-09-01. 7, 2, 3 +[48] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2023. 2, 4, 6 +[49] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Proceedings of the 38th International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. +[50] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4, 6 +[51] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2, 3 +[52] Jaskirat Singh and Liang Zheng. Divide, evaluate, and refine: Evaluating and improving text-to-image alignment with iterative vqa feedback. Advances in Neural Information Processing Systems, 36:70799-70811, 2023. 2 +[53] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 4 +[54] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 4 + +[55] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. arXiv preprint arXiv:2409.17146, 2024. 4 +[56] Tomas Vojir, Tomáš Šipka, Rahaf Aljundi, Nikolay Chumerin, Daniel Olmeda Reino, and Jiri Matas. Road anomaly detection by partial image reconstruction with segmentation coupling. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15651-15660, 2021. 3 +[57] Max Welling and Yee W Teh. Bayesian learning via stochastic gradient Langevin dynamics. In ICML, 2011. 2 +[58] Yeming Wen, Dustin Tran, and Jimmy Ba. BatchEnsemble: an alternative approach to efficient ensemble and lifelong learning. In ICLR, 2019. 2 +[59] Guoxuan Xia and Christos-Savvas Bouganis. Augmenting softmax information for selective classification with out-of-distribution data. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 1995-2012, 2022. 3 +[60] Guoxuan Xia and Christos-Savvas Bouganis. On the usefulness of deep ensemble diversity for out-of-distribution detection, 2022. 3 +[61] Guoxuan Xia and Christos-Savvas Bouganis. Window-based early-exit cascades for uncertainty estimation: When deep ensembles are more efficient than single models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17368-17380, 2023. 2, 3 +[62] Guoxuan Xia, Duolikun Danier, Ayan Das, Stathi Fotiadis, Farhang Nabiei, Ushnish Sengupta, and Alberto Bernacchia. Score Normalization for a Faster Diffusion Exponential Integrator Sampler. In NeurIPS 2023 Workshop on Diffusion Models, 2023. 4 +[63] Guoxuan Xia, Olivier Laurent, Gianni Franchi, and Christos-Savvas Bouganis. Towards understanding why label smoothing degrades selective classification and how to fix it, 2024. 3 +[64] Yingda Xia, Yi Zhang, Fengze Liu, Wei Shen, and Alan L Yuille. Synthesize then compare: Detecting failures and anomalies for semantic segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 145-161. Springer, 2020. 3 +[65] Yilun Xu, Mingyang Deng, Xiang Cheng, Yonglong Tian, Ziming Liu, and Tommi S. Jaakkola. Restart sampling for improving generative processes. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 4 +[66] Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. Generalized out-of-distribution detection: A survey. International Journal of Computer Vision, 2024. 3 +[67] Michal Yarom, Yonatan Bitton, Soravit Changpinyo, Roeah Aharoni, Jonathan Herzig, Oran Lang, Eran Ofek, and Idan Szpektor. What you see is what you read? improving text-image alignment evaluation. Advances in Neural Information Processing Systems, 36:1601-1619, 2023. 2 +[68] Xuanwu Yin Yuda Song, Zehao Sun. Sdxs: Real-time one-step latent diffusion models with image conditions. arxiv, 2024. 6, 2 +[69] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. In The Eleventh + +International Conference on Learning Representations, 2023.4 +[70] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 4 +[71] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 1 +[72] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675, 2019. 6, 1 +[73] Yimeng Zhang, Jinghan Jia, Xin Chen, Aochuan Chen, Yihua Zhang, Jiancheng Liu, Ke Ding, and Sijia Liu. To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images ... for now, 2024. 7, 3 \ No newline at end of file diff --git a/2025/Towards Understanding and Quantifying Uncertainty for Text-to-Image Generation/images.zip b/2025/Towards Understanding and Quantifying Uncertainty for Text-to-Image Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..47b248b3d831de31e44d3d8d3c6c69d5cec873fe --- /dev/null +++ b/2025/Towards Understanding and Quantifying Uncertainty for Text-to-Image Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:737a2eaa794beb17489c38a478c2f7944d0e2a984f8c7ffcead45e528a0b16d6 +size 679094 diff --git a/2025/Towards Understanding and Quantifying Uncertainty for Text-to-Image Generation/layout.json b/2025/Towards Understanding and Quantifying Uncertainty for Text-to-Image Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c6a6200028aec767caa8bb445acfd5da2c759253 --- /dev/null +++ b/2025/Towards Understanding and Quantifying Uncertainty for Text-to-Image Generation/layout.json @@ -0,0 +1,8911 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 103, + 473, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 103, + 473, + 140 + ], + "spans": [ + { + "bbox": [ + 140, + 103, + 473, + 140 + ], + "type": "text", + "content": "Towards Understanding and Quantifying Uncertainty for Text-to-Image Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 158, + 160, + 452, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 160, + 452, + 219 + ], + "spans": [ + { + "bbox": [ + 158, + 160, + 452, + 219 + ], + "type": "text", + "content": "Gianni Franchi," + }, + { + "bbox": [ + 158, + 160, + 452, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 158, + 160, + 452, + 219 + ], + "type": "text", + "content": "* Nacim Belkhir," + }, + { + "bbox": [ + 158, + 160, + 452, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 158, + 160, + 452, + 219 + ], + "type": "text", + "content": "* Dat Nguyen Trong," + }, + { + "bbox": [ + 158, + 160, + 452, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 158, + 160, + 452, + 219 + ], + "type": "text", + "content": " Guoxuan Xia," + }, + { + "bbox": [ + 158, + 160, + 452, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 158, + 160, + 452, + 219 + ], + "type": "text", + "content": " Andrea Pilzer" + }, + { + "bbox": [ + 158, + 160, + 452, + 219 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 158, + 160, + 452, + 219 + ], + "type": "text", + "content": " \nU2IS, ENSTA Paris, Institut Polytechnique de Paris, " + }, + { + "bbox": [ + 158, + 160, + 452, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 158, + 160, + 452, + 219 + ], + "type": "text", + "content": "Mirai," + }, + { + "bbox": [ + 158, + 160, + 452, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 158, + 160, + 452, + 219 + ], + "type": "text", + "content": " Imperial College London," + }, + { + "bbox": [ + 158, + 160, + 452, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 158, + 160, + 452, + 219 + ], + "type": "text", + "content": " NVIDIA" + }, + { + "bbox": [ + 158, + 160, + 452, + 219 + ], + "type": "inline_equation", + "content": "^{4}" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 70, + 244, + 231, + 473 + ], + "blocks": [ + { + "bbox": [ + 70, + 244, + 231, + 473 + ], + "lines": [ + { + "bbox": [ + 70, + 244, + 231, + 473 + ], + "spans": [ + { + "bbox": [ + 70, + 244, + 231, + 473 + ], + "type": "image", + "image_path": "0df14e7c0fe8aff5150a47ddde96930d88820f306268a239773e70e5602c43fa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 479, + 555, + 514 + ], + "lines": [ + { + "bbox": [ + 54, + 479, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 54, + 479, + 555, + 514 + ], + "type": "text", + "content": "Figure 1. Examples of Applications for Uncertainty Quantification in Text-to-Image Generation. Text-to-image generation models may exhibit uncertainty, and that need to be quantified since it can provide insights into the model's training dataset, aiding in deepfake prevention, detecting model biases, and protecting copyrighted content from unauthorized generation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 232, + 244, + 538, + 473 + ], + "blocks": [ + { + "bbox": [ + 232, + 244, + 538, + 473 + ], + "lines": [ + { + "bbox": [ + 232, + 244, + 538, + 473 + ], + "spans": [ + { + "bbox": [ + 232, + 244, + 538, + 473 + ], + "type": "image", + "image_path": "e4cb42dfba865827dd7ac22e034ddfa0a402644ec557b371ebb37dce40429cc8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 523, + 201, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 523, + 201, + 536 + ], + "spans": [ + { + "bbox": [ + 151, + 523, + 201, + 536 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 550, + 297, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 550, + 297, + 682 + ], + "spans": [ + { + "bbox": [ + 54, + 550, + 297, + 682 + ], + "type": "text", + "content": "Uncertainty quantification in text-to-image (T2I) generative models is crucial for understanding model behavior and improving output reliability. In this paper, we are the first to quantify and evaluate the uncertainty of T2I models with respect to the prompt. Alongside adapting existing approaches designed to measure uncertainty in the image space, we also introduce Prompt-based Uncertainty Estimation for T2I models (PUNC), a novel method leveraging Large Vision-Language Models (LVLMs) to better address uncertainties arising from the semantics of the prompt and generated images. PUNC utilizes a LVLM to caption a gen" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 525, + 556, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 556, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 556, + 704 + ], + "type": "text", + "content": "erated image, and then compares the caption with the original prompt in the more semantically meaningful text space. PUNC also enables the disentanglement of both aleatoric and epistemic uncertainties via precision and recall, which image-space approaches are unable to do. Extensive experiments demonstrate that PUNC outperforms state-of-the-art uncertainty estimation techniques across various settings. Uncertainty quantification in text-to-image generation models can be used on various applications including bias detection, copyright protection, and OOD detection. We also introduce a comprehensive dataset of text prompts and generation pairs to foster further research in uncertainty quantification for generative models. Our findings illustrate that PUNC not only achieves competitive performance but also enables novel applications in evaluating" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 693, + 227, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 693, + 227, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 693, + 227, + 713 + ], + "type": "text", + "content": "* equal contribution\nresponding author - gianni.franchi@ensta-paris.fr" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 290, + 693, + 295, + 700 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 693, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 290, + 693, + 295, + 700 + ], + "type": "text", + "content": "+" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8062" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "and improving the trustworthiness of text-to-image models. The code is available at https://github.com/ENSTA-U2IS-AI/Uncertainty_diffusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 131, + 136, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 131, + 136, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 131, + 136, + 144 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 151, + 295, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 151, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 151, + 295, + 270 + ], + "type": "text", + "content": "Generative models have revolutionized artificial intelligence, making significant strides in fields such as robotics [2], healthcare [28], and physics [33]. These models have elevated the performance of AI systems by enabling the generation of complex data representations. However, while generative models offer immense potential, they do not address all challenges, especially those related to robustness and uncertainty in Deep Neural Networks(DNNs)—issues that have become even more pronounced in generative settings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 271, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 295, + 521 + ], + "type": "text", + "content": "In this paper, we focus on the uncertainty quantification in deep generative models, in particular, we are the first to investigate modern text-to-image (T2I) models where the input is a natural language prompt [8, 18, 48-50]. These models have surged in popularity in recent years due to their capacity to generate high-quality images conveniently in a user-guided fashion, finding usage in a wide array of real-world applications. We are the first to investigate and propose a dedicated approach for quantifying uncertainty in text-to-image generation, aiming to fill a critical gap in generative AI research. The task of quantifying uncertainty in these models is not only complex but also crucial. Understanding the uncertainty in generative models can be invaluable for applications such as Out-of-Distribution (OOD) detection, trustworthy AI, and decision-making in high-stakes domains. To the best of our knowledge, no existing work has comprehensively addressed the challenge of uncertainty quantification in text-to-image generative models. However, given the semantic nature of the uncertainty we aim to measure, our work is closely related to recent studies on text-to-image alignment [37, 52, 67]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 522, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 689 + ], + "type": "text", + "content": "Existing research explores uncertainty estimation and OOD detection in the image/output space [4, 5, 23, 39]. For example, given a test image, can we determine whether or not it is OOD? On the other hand, the uncertainty of a generation with respect to conditioning (such as text) is underexplored. We argue that this is an important area to be addressed, as it represents a widespread real-world application of deep generative models. In this case, it remains an open question as to how should we define uncertainty, and given a definition, how should we quantify it? As illustrated in Fig. 1, focusing on uncertainty quantification in text-to-image generation could enable novel applications, allowing us to extract and interpret the knowledge embedded within models using language." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "To answer these questions, we first argue that uncertainties should be rooted in the semantics of the image" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 335 + ], + "type": "text", + "content": "and the prompt. We then adapt and evaluate a number of existing image-space approaches, as well as proposing a novel method for uncertainty quantification in T2I generative models, which we refer to as \"Prompt-based Uncertainty Estimation for T2I Generation\" (PUNC). Our approach leverages the growing capabilities of Large Vision-Language Models (LVLMs) to extract the semantics from generated images. LVLMs, which have been trained on vast amounts of text and visual data, can serve as powerful tools for interpreting the underlying meaning of text/image inputs and assessing how variations in phrasing or semantics impact the uncertainty of the generated outputs. Our hypothesis is that as LVLMs become increasingly adept at understanding complex relationships between text and images, they will allow us to better analyze and quantify uncertainty. Through this paper, we will demonstrate Prompt-based Uncertainty Estimation for T2I (PUNC) effectiveness in quantifying uncertainties in generative models. Showcasing its ability to decompose and interpret different types of uncertainties arising from the multimodal nature of the input, offering a new perspective on uncertainty quantification in generative AI." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 335, + 553, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 553, + 502 + ], + "type": "text", + "content": "To summarize, our contributions are as follows: (1) We introduce a new task aimed at quantifying uncertainty (with respect to the prompt) in text-to-image generation models, and we share a dataset of prompts to support future research on uncertainty in these models. (2) We both adapt existing image-space approaches to our tasks, as well as propose a novel straightforward method, PUNC, designed to better quantify semantic uncertainty in text-to-image generation. (3) Extensive experiments demonstrate that PUNC, despite its simplicity and computational efficiency, achieves competitive performance. (4) We highlight the practical utility of uncertainty quantification in text-to-image generation by showcasing additional applications where this approach could be beneficial." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 514, + 400, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 400, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 400, + 526 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "text", + "content": "Deep learning models typically encounter two primary types of uncertainty: aleatoric and epistemic uncertainty [26]. Over the years, a variety of methods have been developed to quantify these uncertainties. Ensemble methods [16, 32, 34, 58, 61], for instance, have often achieved state-of-the-art performance by leveraging multiple models to assess prediction confidence. However, these methods tend to be computationally expensive due to the need to train and maintain several models. Another approach involves Bayesian methods [21], especially Bayesian Neural Networks (BNNs) [19, 27, 42, 57], which offer a theoretically grounded way to quantify uncertainty. Additionally, there are techniques aimed at directly predicting uncertainty within DNNs [10, 20, 30]. Despite their effectiveness, most of these methods are challenging to apply to text-to-image" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8063" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 81, + 294, + 163 + ], + "blocks": [ + { + "bbox": [ + 88, + 71, + 263, + 80 + ], + "lines": [ + { + "bbox": [ + 88, + 71, + 263, + 80 + ], + "spans": [ + { + "bbox": [ + 88, + 71, + 263, + 80 + ], + "type": "text", + "content": "Generation/Image-space Uncertainty (existing work)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 81, + 294, + 163 + ], + "lines": [ + { + "bbox": [ + 57, + 81, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 57, + 81, + 294, + 163 + ], + "type": "image", + "image_path": "f22022d38cc577c40f1d07107d04862ce1c9139eb16d6a4dce45345b8c7c4a8d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 171, + 295, + 204 + ], + "lines": [ + { + "bbox": [ + 55, + 171, + 295, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 295, + 204 + ], + "type": "text", + "content": "Figure 2. Diagram illustrating generation/image-space uncertainty which is considered in previous work and condition/prompt-space uncertainty which is investigated in our work." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 70, + 230, + 282, + 306 + ], + "blocks": [ + { + "bbox": [ + 70, + 230, + 282, + 306 + ], + "lines": [ + { + "bbox": [ + 70, + 230, + 282, + 306 + ], + "spans": [ + { + "bbox": [ + 70, + 230, + 282, + 306 + ], + "type": "image", + "image_path": "61e51754cec18673e6e57272b130161e06703729facb467446902ddcc7dc5863.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 317, + 296, + 416 + ], + "lines": [ + { + "bbox": [ + 55, + 317, + 296, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 296, + 416 + ], + "type": "text", + "content": "Figure 3. Three generations from PixArt-" + }, + { + "bbox": [ + 55, + 317, + 296, + 416 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 55, + 317, + 296, + 416 + ], + "type": "text", + "content": " [7] illustrating uncertainty with regards to prompt semantics. For the Normal image, we used an ImageNet-inspired prompt; for the corrupted image, additional corruption was applied to the prompt (e.g. fish was perturbed to fis) increasing the aleatoric uncertainty; and for the out-of-distribution (OOD) case, the model was prompted to generate an image of the Prime Minister of Japan Kishida Fumio under similar conditions, where epistemic uncertainty was injected in the form of a semantic concept the model was not familiar with." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 443, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 296, + 491 + ], + "type": "text", + "content": "generation, as they rely on ensembling predictions or accessing model uncertainty in ways that are difficult to implement for generative models without explicit access to intermediate predictions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 498, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 712 + ], + "type": "text", + "content": "Generative models, as a subset of deep learning, also exhibit uncertainty. In the past, traditional generative models were employed for uncertainty quantification and used to help estimate the uncertainty within standard DNNs [6, 38, 56, 64]. With the recent popularity of diffusion models, research has begun exploring whether image-based diffusion models can detect out-of-distribution (OOD) samples [4, 23, 39] or quantify their own uncertainty [5, 6, 14]. Whilst these approaches align with our goals, they are limited to image-space tasks, where the input is an image rather than a textual prompt. This input difference fundamentally alters the nature of uncertainty in the model. Our approach, in contrast, specifically addresses text-to-image generation, a setting where prompt ambiguity and multimodal complexity introduce new dimensions of uncertainty. We demonstrate that quantifying uncertainty in this context can lead to valuable insights and applications across various downstream tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 71, + 539, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 71, + 539, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 539, + 85 + ], + "type": "text", + "content": "3. Uncertainty in Image Generation Models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "content": "Consider a deep generative model " + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\boldsymbol{x})" + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "content": " that models the data distribution " + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{data}}(\\boldsymbol{x})" + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x} \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "content": ". We may be interested in measuring the uncertainty of our model " + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "content": " with respect to some test datum " + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}^*" + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "content": ". For example, consider the case where we want to determine whether or not " + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}^*" + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "content": " was drawn from " + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{data}}(\\boldsymbol{x})" + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "content": ", i.e. out-of-distribution (OOD) detection. This can be achieved by quantifying the uncertainty of " + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "content": " that arises from a lack of knowledge of " + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}^*" + }, + { + "bbox": [ + 313, + 91, + 555, + 247 + ], + "type": "text", + "content": ", i.e. epistemic uncertainty [26]. A range of existing work that leverages diffusion-based generative models for OOD image detection [23, 39] fall into the above paradigm. From hereon we will refer to this as generation/image-space uncertainty. This is illustrated in Fig. 2 (top)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 247, + 554, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 554, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 554, + 378 + ], + "type": "text", + "content": "In this work, however, we are primarily interested in a different scenario: text-to-image generation. Consider instead a conditional model " + }, + { + "bbox": [ + 313, + 247, + 554, + 378 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\pmb{x}|\\pmb{c})" + }, + { + "bbox": [ + 313, + 247, + 554, + 378 + ], + "type": "text", + "content": " that models the conditional data distribution " + }, + { + "bbox": [ + 313, + 247, + 554, + 378 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{data}}(\\pmb{x}|\\pmb{c})" + }, + { + "bbox": [ + 313, + 247, + 554, + 378 + ], + "type": "text", + "content": ". This describes real-world text-to-image generation, where image " + }, + { + "bbox": [ + 313, + 247, + 554, + 378 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 247, + 554, + 378 + ], + "type": "text", + "content": " is drawn from " + }, + { + "bbox": [ + 313, + 247, + 554, + 378 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\pmb{x}|\\pmb{c}^*)" + }, + { + "bbox": [ + 313, + 247, + 554, + 378 + ], + "type": "text", + "content": " conditioned on some test text prompt " + }, + { + "bbox": [ + 313, + 247, + 554, + 378 + ], + "type": "inline_equation", + "content": "\\pmb{c}^*" + }, + { + "bbox": [ + 313, + 247, + 554, + 378 + ], + "type": "text", + "content": ". In this case, we wish to quantify the uncertainty of " + }, + { + "bbox": [ + 313, + 247, + 554, + 378 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 313, + 247, + 554, + 378 + ], + "type": "text", + "content": " not with respect to some test " + }, + { + "bbox": [ + 313, + 247, + 554, + 378 + ], + "type": "inline_equation", + "content": "\\pmb{x}^*" + }, + { + "bbox": [ + 313, + 247, + 554, + 378 + ], + "type": "text", + "content": " (e.g. image) but instead with respect to a test condition " + }, + { + "bbox": [ + 313, + 247, + 554, + 378 + ], + "type": "inline_equation", + "content": "\\pmb{c}^*" + }, + { + "bbox": [ + 313, + 247, + 554, + 378 + ], + "type": "text", + "content": ". We will hereon refer to this as condition/prompt-space uncertainty. This is illustrated in Fig. 2 (bottom)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 379, + 554, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 554, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 554, + 474 + ], + "type": "text", + "content": "We argue that in this case, uncertainties should be concerned with semantics, as intuitively, these are what inherently link the user's intentions, represented by the text prompt, and the generated output image. This is similar to recent discourse on uncertainty estimation for large language models [31]. If we consider the semantic concepts in the prompt and generated image, then we can define useful concepts of uncertainty (illustrated in Fig. 3):" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 474, + 554, + 628 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 314, + 474, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 474, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 314, + 474, + 554, + 557 + ], + "type": "text", + "content": "- Aleatoric uncertainty is irreducible uncertainty in the data distribution " + }, + { + "bbox": [ + 314, + 474, + 554, + 557 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{data}}(\\boldsymbol{x}|\\boldsymbol{c})" + }, + { + "bbox": [ + 314, + 474, + 554, + 557 + ], + "type": "text", + "content": ". From the perspective of semantics, high aleatoric uncertainty would be where a variation of generated concepts may arise from a single prompt. For example, a spelling mistake where fish is mistyped as fis in the prompt may result in either a \"fish\" or a \"fist\" appearing in the image." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 558, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 628 + ], + "type": "text", + "content": "- Epistemic uncertainty should correspond to a model's lack of knowledge of semantic concepts found in a prompt. For example, a model trained on ImageNet will not know what the prime minister of Japan Kishida Fumio looks like, and thus will have high epistemic uncertainty for this semantic concept." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 630, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 677 + ], + "type": "text", + "content": "We note that broadly speaking uncertainty estimation for deep conditional models is well explored [20-22, 24, 26, 32, 34, 41, 59-61, 63, 66]. However, research in this area is generally concerned with problem settings where 1) the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "text", + "content": "1We note that although there is a discussion over how to define and quantify epistemic uncertainty [21, 26, 41, 45, 60], in this work, we stick to an intuitive albeit informal view based on a deep model's knowledge." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8064" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "models explicitly evaluate " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\pmb{x}|c)" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": " and 2) downstream applications of uncertainty relate directly to the numerical values of " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\pmb{x}|c)" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": ". For example, a cross-entropy-trained image classifier explicitly predicts the probability mass " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "P_{\\theta}(\\pmb{x}|c^{*})" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": " of label " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": " given test image " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "c^{*}" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "P_{\\theta}(\\pmb{x}|c^{*})" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": " directly relates to the likelihood of classification error – which is an event that we may want to detect using an uncertainty measure, i.e. downstream misclassification detection [22, 24]. Thus, we are able to easily extract useful measures of uncertainty from the model " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "P_{\\theta}" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": ", e.g. the maximum of the softmax." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 193, + 295, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 193, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 295, + 360 + ], + "type": "text", + "content": "On the other hand, text-to-image generation models are not as clear-cut. They may not explicitly evaluate " + }, + { + "bbox": [ + 55, + 193, + 295, + 360 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\pmb{x}|c^{*})" + }, + { + "bbox": [ + 55, + 193, + 295, + 360 + ], + "type": "text", + "content": " (e.g. diffusion-based models [25, 54]), and even when they do (e.g. vector-quantised autoregressive models [17, 55]) it is hard to directly relate (latent) pixel values directly to our above definitions of uncertainty. For example, simple variations in contrast and brightness that may leave semantics completely unchanged could result in a very high variance in " + }, + { + "bbox": [ + 55, + 193, + 295, + 360 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\pmb{x}|c^{*})" + }, + { + "bbox": [ + 55, + 193, + 295, + 360 + ], + "type": "text", + "content": ". To give a simple example, inverting the black and white stripes on a zebra would lead to a large Euclidean distance in the image space whilst preserving the semantics of the image. In the next section, we present and discuss a range of potential methods to quantify the uncertainty of text-to-image models with respect to a given prompt." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 372, + 253, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 253, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 253, + 387 + ], + "type": "text", + "content": "4. Techniques to Quantify Uncertainty" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 393, + 295, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 393, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 295, + 430 + ], + "type": "text", + "content": "We quantify text-to-image uncertainty by adapting existing methods and proposing a novel approach named Prompt-based UNCertainty estimation (PUNC)(Fig. 4)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 437, + 222, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 222, + 451 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 222, + 451 + ], + "type": "text", + "content": "4.1. Adapting Existing Approaches" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 456, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 529 + ], + "type": "text", + "content": "There are a number of existing methods for quantifying the image-space uncertainty of diffusion models. As diffusion-based models represent a majority of current text-to-image foundation models [8, 18, 48, 50], we choose to adapt these in order to quantify prompt-space uncertainty (Fig. 4). We first present a brief primer on diffusion models to the reader." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 536, + 157, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 157, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 157, + 547 + ], + "type": "text", + "content": "4.1.1. Diffusion Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 551, + 295, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 295, + 588 + ], + "type": "text", + "content": "Diffusion models [25] generate data by iteratively denoising pure noise. The forward process over time " + }, + { + "bbox": [ + 55, + 551, + 295, + 588 + ], + "type": "inline_equation", + "content": "t \\in [0,1]" + }, + { + "bbox": [ + 55, + 551, + 295, + 588 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 551, + 295, + 588 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_t \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 55, + 551, + 295, + 588 + ], + "type": "text", + "content": ", conditioned on a text prompt " + }, + { + "bbox": [ + 55, + 551, + 295, + 588 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 551, + 295, + 588 + ], + "type": "text", + "content": ", follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 597, + 295, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 597, + 295, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 295, + 611 + ], + "type": "interline_equation", + "content": "p \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}, \\boldsymbol {c}\\right) = \\mathcal {N} \\left(\\boldsymbol {x} _ {t}; a _ {t} \\boldsymbol {x} _ {0}, \\sigma_ {t} ^ {2} \\boldsymbol {I}\\right), \\tag {1}", + "image_path": "50be9fdaca19391d2fa346d4e8edee05dfbffd44efb736b75e938f13cf48f489.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 621, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 295, + 693 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 621, + 295, + 693 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 55, + 621, + 295, + 693 + ], + "type": "text", + "content": " is sampled from an unknown data distribution " + }, + { + "bbox": [ + 55, + 621, + 295, + 693 + ], + "type": "inline_equation", + "content": "p(\\pmb{x}_0)" + }, + { + "bbox": [ + 55, + 621, + 295, + 693 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 621, + 295, + 693 + ], + "type": "inline_equation", + "content": "a_{t},\\sigma_{t}" + }, + { + "bbox": [ + 55, + 621, + 295, + 693 + ], + "type": "text", + "content": " define the noise schedule. This formulation allows diffusion models to progressively refine noisy inputs, enabling high-quality data generation through a learned denoising process. By leveraging stochastic differential equations (SDEs) or their deterministic counterparts," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 121 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 121 + ], + "type": "text", + "content": "these models sample from complex distributions in a structured manner. Various solvers [29, 40, 53, 54, 62, 65, 69] can then be employed to efficiently generate realistic samples with a limited number of denoising steps." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 127, + 555, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 127, + 555, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 127, + 555, + 212 + ], + "type": "text", + "content": "Latent diffusion models (LDMs) [39] perform this process within a compressed latent space rather than directly on pixel space. By first encoding images into a more compact representation, LDMs reduce computational demands, enabling high-quality high-resolution generations. We will not consider the latent representation to keep the notation simple, even though experiments will use LDMs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 237, + 463, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 237, + 463, + 249 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 463, + 249 + ], + "type": "text", + "content": "4.1.2. Time-step based approaches" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 258, + 555, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 450 + ], + "type": "text", + "content": "DDPM-OOD [23]. This approach, developed for image-space OOD detection, involves noising a test sample " + }, + { + "bbox": [ + 313, + 258, + 555, + 450 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_0^*" + }, + { + "bbox": [ + 313, + 258, + 555, + 450 + ], + "type": "text", + "content": " to different points in the forward process to get " + }, + { + "bbox": [ + 313, + 258, + 555, + 450 + ], + "type": "inline_equation", + "content": "\\{\\boldsymbol{x}_{t_1}^*,\\boldsymbol{x}_{t_2}^*,\\dots \\}" + }, + { + "bbox": [ + 313, + 258, + 555, + 450 + ], + "type": "text", + "content": " and denoising them, generating multiple reconstructions " + }, + { + "bbox": [ + 313, + 258, + 555, + 450 + ], + "type": "inline_equation", + "content": "\\{\\hat{x}_{0_1},\\hat{x}_{0_2},\\ldots \\}" + }, + { + "bbox": [ + 313, + 258, + 555, + 450 + ], + "type": "text", + "content": ". Similarity is then measured between each reconstruction and the original image. The idea is that test samples " + }, + { + "bbox": [ + 313, + 258, + 555, + 450 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_0^*" + }, + { + "bbox": [ + 313, + 258, + 555, + 450 + ], + "type": "text", + "content": " for which the model has higher epistemic uncertainty will lead to reconstructions that differ from the test sample. However, in our case, we do not have a test image " + }, + { + "bbox": [ + 313, + 258, + 555, + 450 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_0^*" + }, + { + "bbox": [ + 313, + 258, + 555, + 450 + ], + "type": "text", + "content": ". In order to adapt DDPM-OOD we propose measuring similarity between reconstructions at different " + }, + { + "bbox": [ + 313, + 258, + 555, + 450 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 258, + 555, + 450 + ], + "type": "text", + "content": " and a baseline generation, " + }, + { + "bbox": [ + 313, + 258, + 555, + 450 + ], + "type": "inline_equation", + "content": "S(\\hat{x}_{0_i},\\hat{x}_0)" + }, + { + "bbox": [ + 313, + 258, + 555, + 450 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 258, + 555, + 450 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 313, + 258, + 555, + 450 + ], + "type": "text", + "content": " represents an image directly generated using the original test prompt " + }, + { + "bbox": [ + 313, + 258, + 555, + 450 + ], + "type": "inline_equation", + "content": "c^*" + }, + { + "bbox": [ + 313, + 258, + 555, + 450 + ], + "type": "text", + "content": " alone. Both aleatoric uncertainty (e.g. ambiguous prompt) and epistemic uncertainty (e.g. unfamiliar concepts in prompt) could lead to more divergent reconstructions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 451, + 556, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 556, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 556, + 642 + ], + "type": "text", + "content": "LMD [39]. Another similar approach for image-space OOD detection, known as Learned Manifold Denoising (LMD), instead explores alternative transforms on " + }, + { + "bbox": [ + 313, + 451, + 556, + 642 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_0^*" + }, + { + "bbox": [ + 313, + 451, + 556, + 642 + ], + "type": "text", + "content": " to noising. The idea is to lift " + }, + { + "bbox": [ + 313, + 451, + 556, + 642 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_0^*" + }, + { + "bbox": [ + 313, + 451, + 556, + 642 + ], + "type": "text", + "content": " from its original data manifold and then map it back using the trained diffusion model. If the sample is in-distribution, the mapping process should return it close to its original position on the data manifold. Conversely, if it is OOD, the mapped position will likely diverge significantly from the original. The corruption process typically involves masking portions of the image, with various mask patterns applied. Similar to DDPM-OOD, similarity is measured between the original image and the reconstructed version after applying the diffusion process. Similar to before, since we lack a testimage, we measure the similarity between reconstructed images and " + }, + { + "bbox": [ + 313, + 451, + 556, + 642 + ], + "type": "inline_equation", + "content": "\\hat{\\boldsymbol{x}}_{0_i}" + }, + { + "bbox": [ + 313, + 451, + 556, + 642 + ], + "type": "text", + "content": ", the image generated directly from the prompt without perturbations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Similarity Metric. Various methods are used to evaluate similarity between images. Mean-squared error (MSE) is commonly applied between input and reconstruction, while some approaches favor the Learned Perceptual Image Patch Similarity (LPIPS) [70] metric. The choice of similarity metric is a crucial component of these techniques." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 290, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 290, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 290, + 713 + ], + "type": "text", + "content": "2We stick with this non-standard notation for the sake of consistency." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8065" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 555, + 196 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 555, + 196 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 555, + 196 + ], + "type": "image", + "image_path": "028fd6b0c4c53b8460bc355b984d6d2584ea5339d9111611314b3422a07d66e0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 205, + 555, + 239 + ], + "lines": [ + { + "bbox": [ + 54, + 205, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 54, + 205, + 555, + 239 + ], + "type": "text", + "content": "Figure 4. Illustration showing the different baselines and PUNC. PUNC leverages a LVLM to describe generated images and assess similarity with the original prompt, providing a refined uncertainty score. In contrast, baseline methods employ traditional techniques such as noise injection, ensembling, or masking to quantify uncertainty, followed by image-based similarity scoring." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 249, + 234, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 249, + 234, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 234, + 261 + ], + "type": "text", + "content": "4.1.3. Test-Time Ensembling Approaches" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 263, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 263, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 54, + 263, + 296, + 418 + ], + "type": "text", + "content": "Test-time ensembling techniques take inspiration from approaches used in the large language model (LLM) community [31], where multiple generations are produced to assess consistency. Given a prompt " + }, + { + "bbox": [ + 54, + 263, + 296, + 418 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 54, + 263, + 296, + 418 + ], + "type": "text", + "content": ", we generate two images " + }, + { + "bbox": [ + 54, + 263, + 296, + 418 + ], + "type": "inline_equation", + "content": "\\{\\hat{x}_1, \\hat{x}_2\\}" + }, + { + "bbox": [ + 54, + 263, + 296, + 418 + ], + "type": "text", + "content": " and measure the similarity between the two outputs. We denote this technique 2XDM. The underlying rationale is that if the model has high confidence, it will generate similar images. However, if uncertainty is high, it will generate substantially different images. We note that this is similar to the approach proposed in [5], however, that approach uses an ensemble of diffusion models, which is not practical in the setting where we want to quantify the uncertainty of a single pretrained T2I model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 420, + 296, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 420, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 54, + 420, + 296, + 564 + ], + "type": "text", + "content": "A key question remains regarding how to find the best measure similarity, and we propose using the same techniques applied in the time-step-based methods. A significant advantage of test-time ensembling approaches is that they can be applied generically to all T2I models. In contrast, timestep-based approaches require a multi-step generation process, limiting the broadness of their applicability. We note that all above approaches require multiple generations, incurring a notable computational overhead at inference time. For example, DDPM-OOD requires " + }, + { + "bbox": [ + 54, + 420, + 296, + 564 + ], + "type": "inline_equation", + "content": "\\sim 50\\times" + }, + { + "bbox": [ + 54, + 420, + 296, + 564 + ], + "type": "text", + "content": " the number of network inferences as a single generation (if the hyperparameters of the original paper are used) [23]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 571, + 257, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 257, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 257, + 583 + ], + "type": "text", + "content": "4.2. Prompt-based Uncertainty Estimation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 590, + 296, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 296, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 296, + 673 + ], + "type": "text", + "content": "In this section, we propose a new method, Prompt-based Uncertainty estimation for text-to-image generation (PUNC), which leverages Large Vision-Language Models (LVLM) in conjunction with T2I models. We aim to assess the uncertainty in generated images by analyzing the alignment between initial text prompts and reconstructed textual descriptions of the generated images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 700 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 700 + ], + "type": "text", + "content": "Background on Large Vision-Language Model (LVLM)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "A LVLM integrates a large language model (LLM) with" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 249, + 553, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 553, + 284 + ], + "type": "text", + "content": "an image encoder, allowing for robust, multimodal understanding. Given a text prompt " + }, + { + "bbox": [ + 313, + 249, + 553, + 284 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 249, + 553, + 284 + ], + "type": "text", + "content": " and an image " + }, + { + "bbox": [ + 313, + 249, + 553, + 284 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 249, + 553, + 284 + ], + "type": "text", + "content": ", a LVLM model, with parameters " + }, + { + "bbox": [ + 313, + 249, + 553, + 284 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 249, + 553, + 284 + ], + "type": "text", + "content": ", includes:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 285, + 553, + 347 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 313, + 285, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 553, + 312 + ], + "type": "text", + "content": "- Image Encoder: " + }, + { + "bbox": [ + 313, + 285, + 553, + 312 + ], + "type": "inline_equation", + "content": "f_{\\omega}^{\\mathrm{img}}(\\cdot)" + }, + { + "bbox": [ + 313, + 285, + 553, + 312 + ], + "type": "text", + "content": " processes the input image " + }, + { + "bbox": [ + 313, + 285, + 553, + 312 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 285, + 553, + 312 + ], + "type": "text", + "content": " and produces an embedding: " + }, + { + "bbox": [ + 313, + 285, + 553, + 312 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{img}} = f_{\\omega}^{\\mathrm{img}}(x)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 312, + 553, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 347 + ], + "type": "text", + "content": "- Text Processor with LLM: " + }, + { + "bbox": [ + 313, + 312, + 553, + 347 + ], + "type": "inline_equation", + "content": "f_{\\omega}^{\\mathrm{txt}}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 312, + 553, + 347 + ], + "type": "text", + "content": " takes the prompt " + }, + { + "bbox": [ + 313, + 312, + 553, + 347 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 312, + 553, + 347 + ], + "type": "text", + "content": " and the image embedding " + }, + { + "bbox": [ + 313, + 312, + 553, + 347 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{img}}" + }, + { + "bbox": [ + 313, + 312, + 553, + 347 + ], + "type": "text", + "content": " as inputs, generating a descriptive or answer-based response: " + }, + { + "bbox": [ + 313, + 312, + 553, + 347 + ], + "type": "inline_equation", + "content": "\\hat{c} = f_{\\omega}^{\\mathrm{txt}}(c, z^{\\mathrm{img}})" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 349, + 553, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 553, + 373 + ], + "type": "text", + "content": "Thus, a LVLM system generates an interpretation " + }, + { + "bbox": [ + 313, + 349, + 553, + 373 + ], + "type": "inline_equation", + "content": "\\hat{c}" + }, + { + "bbox": [ + 313, + 349, + 553, + 373 + ], + "type": "text", + "content": " that reflects the content of the image given an initial prompt." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 387, + 554, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 554, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 554, + 530 + ], + "type": "text", + "content": "Prompt-Based Uncertainty Estimation for T2I Generation (PUNC) The previously introduced approaches to uncertainty estimation rely on measuring similarity between images. For a given text prompt " + }, + { + "bbox": [ + 313, + 387, + 554, + 530 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 387, + 554, + 530 + ], + "type": "text", + "content": ", a T2I model generates an image " + }, + { + "bbox": [ + 313, + 387, + 554, + 530 + ], + "type": "inline_equation", + "content": "x \\sim p_{\\theta}(x|c)" + }, + { + "bbox": [ + 313, + 387, + 554, + 530 + ], + "type": "text", + "content": ". We reason that, even under low uncertainty, the model can produce images that vary in aspects such as color, texture, or object positioning, while still faithfully representing the semantics of prompt " + }, + { + "bbox": [ + 313, + 387, + 554, + 530 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 387, + 554, + 530 + ], + "type": "text", + "content": ". As previously discussed in Sec. 3, this may result in image-based similarity metrics failing to capture uncertainty at a semantic level, since large differences in the image space may not necessarily correspond to meaningful semantic differences." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 530, + 554, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 554, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 554, + 649 + ], + "type": "text", + "content": "We propose that advanced LVLMs, such as Molmo [11], LLAMA 3 [15], and GPT-4 [1], are a strong candidate for extracting the semantics from an image, through describing an image via natural language. Building on this insight, our proposed method, PUNC, introduces a novel approach that bypasses direct image comparison and instead evaluates the alignment between generated images and the original prompts by calculating an uncertainty score in the more semantically meaningful text-space. The steps for implementing PUNC are as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 650, + 553, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 650, + 553, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 650, + 553, + 673 + ], + "type": "text", + "content": "Step 1: Initial Prompt and Image Generation Given a test prompt " + }, + { + "bbox": [ + 313, + 650, + 553, + 673 + ], + "type": "inline_equation", + "content": "c^*" + }, + { + "bbox": [ + 313, + 650, + 553, + 673 + ], + "type": "text", + "content": ", we use the T2I model to sample an image " + }, + { + "bbox": [ + 313, + 650, + 553, + 673 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 650, + 553, + 673 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 403, + 681, + 464, + 694 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 681, + 464, + 694 + ], + "spans": [ + { + "bbox": [ + 403, + 681, + 464, + 694 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} \\sim p _ {\\boldsymbol {\\theta}} (\\boldsymbol {x} | \\boldsymbol {c} ^ {*})", + "image_path": "a17ba5d1a6bfa8cd030c2aada8e442cfd810895a7a3e45983adbeddf4d3a56f5.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "content": "Step 2: Image Interpretation via LVLM With the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8066" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 69, + 273, + 165 + ], + "blocks": [ + { + "bbox": [ + 78, + 69, + 273, + 165 + ], + "lines": [ + { + "bbox": [ + 78, + 69, + 273, + 165 + ], + "spans": [ + { + "bbox": [ + 78, + 69, + 273, + 165 + ], + "type": "image", + "image_path": "bfd49245a54fb630faa3175b2d199995fca0356c85229b1aa37a4ac1798f6ca0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 173, + 287, + 184 + ], + "lines": [ + { + "bbox": [ + 63, + 173, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 63, + 173, + 287, + 184 + ], + "type": "text", + "content": "Figure 5. Disentangling Aleatoric and Epistemic Uncertainty." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 205, + 295, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 295, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 295, + 231 + ], + "type": "text", + "content": "generated image " + }, + { + "bbox": [ + 55, + 205, + 295, + 231 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 205, + 295, + 231 + ], + "type": "text", + "content": " and the initial prompt " + }, + { + "bbox": [ + 55, + 205, + 295, + 231 + ], + "type": "inline_equation", + "content": "c^*" + }, + { + "bbox": [ + 55, + 205, + 295, + 231 + ], + "type": "text", + "content": ", the LVLM produces a new descriptive caption " + }, + { + "bbox": [ + 55, + 205, + 295, + 231 + ], + "type": "inline_equation", + "content": "\\hat{c}" + }, + { + "bbox": [ + 55, + 205, + 295, + 231 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 239, + 220, + 253 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 239, + 220, + 253 + ], + "spans": [ + { + "bbox": [ + 130, + 239, + 220, + 253 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {c}} = f _ {\\omega} ^ {\\mathrm {t x t}} \\left(\\boldsymbol {c} ^ {*}, f _ {\\omega} ^ {\\mathrm {i m g}} (\\boldsymbol {x})\\right)", + "image_path": "e1b9f254f07c8ab2cb01a8a5c99e1cb7773d0de1015e4129c904cb60da1017a8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 262, + 296, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 296, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 296, + 310 + ], + "type": "text", + "content": "Step 3: Uncertainty Score Calculation The core of PUNC lies in measuring the alignment between " + }, + { + "bbox": [ + 55, + 262, + 296, + 310 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 262, + 296, + 310 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 262, + 296, + 310 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{c}}" + }, + { + "bbox": [ + 55, + 262, + 296, + 310 + ], + "type": "text", + "content": " to assess the diffusion model's confidence. Specifically, the similarity score " + }, + { + "bbox": [ + 55, + 262, + 296, + 310 + ], + "type": "inline_equation", + "content": "S(\\mathbf{c},\\hat{\\mathbf{c}})" + }, + { + "bbox": [ + 55, + 262, + 296, + 310 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 129, + 319, + 222, + 333 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 319, + 222, + 333 + ], + "spans": [ + { + "bbox": [ + 129, + 319, + 222, + 333 + ], + "type": "interline_equation", + "content": "S (\\boldsymbol {c} ^ {*}, \\hat {\\boldsymbol {c}}) = \\operatorname {s i m} (\\boldsymbol {c} ^ {*}, \\hat {\\boldsymbol {c}})", + "image_path": "f9c348a164cbb4e59b4e08be6bc04ecdcaabfd2955aec9d72188534a0e18c289.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 342, + 295, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 390 + ], + "type": "text", + "content": "High similarity suggests low uncertainty, indicating the generated image closely matches the initial prompt. Conversely, a lower score reveals higher uncertainty, signifying potential divergence from the prompt's intended meaning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 390, + 296, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 296, + 426 + ], + "type": "text", + "content": "Thus, PUNC is able to predict semantic uncertainties by utilising the powerful image understanding ability of LVLMs and computing similarities in the text space." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 426, + 295, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 534 + ], + "type": "text", + "content": "Prompt Similarity: Aleatoric and Epistemic via Precision and Recall. In Sec. 3 we discussed the different types of semantic uncertainty for T2I generation. Here we will present a method for disentangling these uncertainties using the precision and recall of the caption given the prompt. Consider an abstract scenario where the prompt and the generated image each contain a discrete set of semantic concepts. We can then calculate the precision and recall of the image concepts with respect to the prompt:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 534, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 630 + ], + "type": "text", + "content": "Intuitively, a lack of knowledge about concepts in the prompt (epistemic uncertainty) will result in fewer concepts being preserved in the image, reducing recall. On the other hand, a lack of specificity in the prompt (aleatoric uncertainty) will result in additional concepts being generated, reducing precision. Thus we can use this abstract form of semantic-concept-based recall to measure epistemic uncertainty, and use precision to measure aleatoric uncertainty." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "Although we cannot directly measure concepts, we can approximate them using natural language. Moreover, existing similarity scores for text already have the above notions of precision and recall integrated within [36, 43, 72], allowing PUNC to disentangle uncertainty by comparing the caption of the generated image with the original prompt. We note that the image-based approaches in" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "Sec. 4.1 are unable to disentangle aleatoric from epistemic uncertainty. We focus on two different existing text-based similarity metrics, ROUGE [36], and BERTScore [72]. The former calculates similarity (precision or recall) by directly matching (sequences of) words between two text sequences. The latter on the other hand measures similarity in the deep feature space of a text encoder." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 156, + 555, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 239 + ], + "type": "text", + "content": "As illustrated in Fig. 5, we define Prompt Similarity as the semantic overlap between the input prompt and the generated descriptive caption. More formally, let " + }, + { + "bbox": [ + 313, + 156, + 555, + 239 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{GT}}" + }, + { + "bbox": [ + 313, + 156, + 555, + 239 + ], + "type": "text", + "content": " denote the set of semantic concepts in the ground-truth prompt used to generate an image, and " + }, + { + "bbox": [ + 313, + 156, + 555, + 239 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{VQA}}" + }, + { + "bbox": [ + 313, + 156, + 555, + 239 + ], + "type": "text", + "content": " the set of concepts extracted from the LVLM-generated caption. The resulting uncertainties can be characterized as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 239, + 553, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 553, + 276 + ], + "type": "text", + "content": "Epistemic Uncertainty, which captures missing concepts in " + }, + { + "bbox": [ + 313, + 239, + 553, + 276 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{VQA}}" + }, + { + "bbox": [ + 313, + 239, + 553, + 276 + ], + "type": "text", + "content": " that are present in " + }, + { + "bbox": [ + 313, + 239, + 553, + 276 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{GT}}" + }, + { + "bbox": [ + 313, + 239, + 553, + 276 + ], + "type": "text", + "content": ", reflects an incomplete understanding of the intended semantics. It is defined as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 343, + 284, + 553, + 310 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 284, + 553, + 310 + ], + "spans": [ + { + "bbox": [ + 343, + 284, + 553, + 310 + ], + "type": "interline_equation", + "content": "U _ {\\text {e p i s t e m i c}} = 1 - \\frac {\\left| P _ {\\mathrm {G T}} \\cap P _ {\\mathrm {V Q A}} \\right|}{\\left| P _ {\\mathrm {G T}} \\right|} = 1 - \\text {R e c a l l} \\tag {2}", + "image_path": "825d4c8bc11c831f8934002418d7848680d9b80927464a8c38d0828e3fd1c73c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 319, + 553, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 553, + 365 + ], + "type": "text", + "content": "Aleatoric Uncertainty, on the other hand, arises from spurious or extra concepts in " + }, + { + "bbox": [ + 313, + 319, + 553, + 365 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{VQA}}" + }, + { + "bbox": [ + 313, + 319, + 553, + 365 + ], + "type": "text", + "content": " that do not appear in " + }, + { + "bbox": [ + 313, + 319, + 553, + 365 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{GT}}" + }, + { + "bbox": [ + 313, + 319, + 553, + 365 + ], + "type": "text", + "content": ", indicating ambiguity in the model's response. It is defined as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 339, + 373, + 553, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 373, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 339, + 373, + 553, + 399 + ], + "type": "interline_equation", + "content": "U _ {\\text {a l e a t o r i c}} = 1 - \\frac {\\left| P _ {\\mathrm {G T}} \\cap P _ {\\mathrm {V Q A}} \\right|}{\\left| P _ {\\mathrm {V Q A}} \\right|} = 1 - \\text {P r e c i s i o n} \\tag {3}", + "image_path": "7345adf0b2305a3797c75ec488b77316d89914546e0cccfb5fdf24ff463a713e.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 407, + 555, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 555, + 432 + ], + "type": "text", + "content": "Further details on the computation of precision and recall can be found in Appendix A." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 442, + 469, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 469, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 469, + 455 + ], + "type": "text", + "content": "5. Experiments & Discussions" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 462, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 689 + ], + "type": "text", + "content": "We test PUNC on Various task and various dataset and compared PUNC to DDPM-OOD [23], LMD [39] and 2XDM. We first introduce on Sec. 5 how we have constitute the dataset of prompts. Then on Sec. 5.1 we discuss of different results on these dataset. Then Appendix F explains how we can use the uncertainty of text-to-image model to detect if a model has been trained with some concept related to some fictional character of politicians. Appendix F.3 shows how uncertainty can help to detect bias of T2I models. To conduct our experiments and evaluate the various techniques for quantifying uncertainty, we utilized four T2I models: Stable Diffusion 1.5 (SDv1.5) [50], SDXS [68], SDXL [48], and PixArt- " + }, + { + "bbox": [ + 313, + 462, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 462, + 555, + 689 + ], + "type": "text", + "content": " [7]. These diverse T2I models allow us to test a range of architectures, each with unique characteristics and potential weaknesses. SDv1.5, being one of the earliest models, likely has a smaller and less diverse training dataset compared to the others. In contrast, SDXS generates in a single-step, different to the multi-step processes in other models." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 690, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 690, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 690, + 553, + 714 + ], + "type": "text", + "content": "Dataset of prompts To conduct our experiments, we require both an in-distribution dataset of prompts and" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8067" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 264 + ], + "type": "text", + "content": "datasets with uncertainty. For the in-distribution prompts, we use the dataset proposed by [13], where high-quality descriptions of ImageNet [12] images are generated using GPT-4[13], referred to as Normal. To simulate epistemic uncertainty, we create out-of-distribution (OOD) datasets using images from underrepresented domains like remote sensing, texture, and microscopic datasets, captioned with LLAVA Next[35], labeled as Remote Sensing, Texture, and Microscopic. For aleatoric uncertainty, we generate two datasets: Vague, containing 2,000 minimally descriptive prompts (e.g., \"An image of ***\"), and Corrupted, derived from Normal by adding grammatical errors (Level 1) and removing " + }, + { + "bbox": [ + 55, + 72, + 294, + 264 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 72, + 294, + 264 + ], + "type": "text", + "content": " of words (Level 2). A detailed summary of these datasets is presented in Tab. 1 and Appendix E. In Appendix C, we provide the details of the hyperparameters used in our experiments." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 80, + 274, + 272, + 376 + ], + "blocks": [ + { + "bbox": [ + 80, + 274, + 272, + 376 + ], + "lines": [ + { + "bbox": [ + 80, + 274, + 272, + 376 + ], + "spans": [ + { + "bbox": [ + 80, + 274, + 272, + 376 + ], + "type": "table", + "html": "
Dataset TypeDataset NameDescription
In-Distribution (ID)Normal20k High-quality detailed prompts from ImageNet [12] images using GPT-4[13].
Out-of-Distribution (OOD)Remote Sensing Texture Microscopic20k Prompts from remote sensing images [47]5k Prompts from texture datasets10k Prompts from microscopic images [44]
Aleatoric UncertaintyVagueAdversarialCorrupted2k prompts with minimal context, e.g., "An image of ***," where "****" is an ImageNet class name, creating ambiguous input.1k altered prompts from Normal using Unlearn-DiffAtk [73]20k Altered prompts from Normal:• Level 1: Grammatical mistakes and spelling errors.• Level 2: Random removal of 50% of words in Level 1 prompts.
", + "image_path": "171fe859994b014f2125487c611fc0c28b4bf4d771eef1699d8169708a968c43.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 64, + 385, + 286, + 396 + ], + "lines": [ + { + "bbox": [ + 64, + 385, + 286, + 396 + ], + "spans": [ + { + "bbox": [ + 64, + 385, + 286, + 396 + ], + "type": "text", + "content": "Table 1. Summary of prompt datasets for unc. quantification." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 418, + 241, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 241, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 241, + 431 + ], + "type": "text", + "content": "5.1. Results for Uncertainty Estimation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 436, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 519 + ], + "type": "text", + "content": "Tab. 3 presents the results of PUNC when using Molmo as the LVLM to assess epistemic uncertainty in the T2I models. Our approach, PUNC, outperforms most state-of-the-art techniques on average. However, we observe more mixed results on the Texture dataset, likely due to uncertainty about the presence of texture images in the training data of the T2I models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 520, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 664 + ], + "type": "text", + "content": "In Tab. 4, we show the results of PUNC under aleatoric uncertainty, again using Molmo as the LVLM. Our approach outperforms most state-of-the-art methods in this setting as well. Notably, while the ROUGE score tends to perform better for epistemic uncertainty, the optimal metric is less clear for aleatoric uncertainty. Here, the choice between ROUGE and BERTScore may depend on the specific application: because corruption introduces spelling and grammatical errors, BERTScore—which captures semantic similarity—may be more suitable. For additional results using different LVLMs, please refer to Appendix D and Appendix B. In Appendix G, we present qualitative results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 672, + 206, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 206, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 206, + 685 + ], + "type": "text", + "content": "5.2. Applications of uncertainty" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "Our study explores multiple applications of uncertainty quantification in T2I models, with detailed results provided" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "content": "in Appendix F. First, we investigate deepfake detection by examining the model's ability to generate recognizable images of prominent politicians (see Appendix F.1). We examined whether diffusion models could generate convincing images of G7 politicians. Using 100 prompts containing their names, we generated images with four diffusion models and queried a LVLM to determine if the intended politician was present in each image (yes/no), thus measuring concept recognition uncertainty. Results are presented in Table 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 192, + 553, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 553, + 275 + ], + "type": "text", + "content": "Next, we address copyright concerns by evaluating whether T2I models can accurately replicate iconic, copyrighted characters such as Mickey Mouse and Darth Vader (see Appendix F.2). Finally, we analyze bias in T2I models by testing their capacity to generate individuals across different genders and races for specific job roles (see Appendix F.3)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 276, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 555, + 384 + ], + "type": "text", + "content": "Using a LVLM, we evaluate the presence of specific individuals, copyrighted characters, or demographic concepts in the generated images, demonstrating how uncertainty estimation can aid in detecting potential misuses of generative models. These experiments illustrate the value of text-based uncertainty quantification in analyzing T2I model performance across ethically and legally sensitive applications, enabling assessments that are challenging to achieve with image-based generation alone." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 394, + 388, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 394, + 388, + 407 + ], + "spans": [ + { + "bbox": [ + 314, + 394, + 388, + 407 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 415, + 553, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 553, + 594 + ], + "type": "text", + "content": "In this work, we set out to be first to investigate uncertainty estimation T2I generation, where the uncertainty is with respect to the text prompt. We both adapt existing image-space uncertainty approaches, as well as present Prompt-based UNCertainty Estimation for T2I models (PUNC), the first approach designed specifically to quantify uncertainty in T2I generation. By leveraging the interpretive power of Large Vision-Language Models to extract semantics from a generated image, PUNC addresses the challenges associated with measuring multimodal uncertainty, enabling a deeper understanding of model behavior with respect to prompt-based conditioning. Our results validate that PUNC provides reliable, fine-grained assessments of aleatoric and epistemic uncertainties, outperforming existing methods that work for image generation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "PUNC opens avenues for new applications, such as improving model robustness in biased or OOD scenarios and safeguarding against unauthorized content generation, including deepfakes and copyrighted characters. Our experiments show that diffusion models, while powerful, still exhibit significant variances in their ability to adhere to prompt details, particularly with ambiguous, corrupt prompts or OOD prompts. By identifying such uncertainties, PUNC not only supports the development of more trustworthy generative AI but also aids in aligning these" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8068" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 70, + 478, + 120 + ], + "blocks": [ + { + "bbox": [ + 132, + 70, + 478, + 120 + ], + "lines": [ + { + "bbox": [ + 132, + 70, + 478, + 120 + ], + "spans": [ + { + "bbox": [ + 132, + 70, + 478, + 120 + ], + "type": "table", + "html": "
Justin TrudeauKing CharlesFumio KishidaOlaf ScholzEmmanuel MacronJoe BidenGiorgia MeloniAVERAGE
SDXL9494888898962195.5
SDv1.57860628894801878
SDXS8040708294762872.5
PixArt3676627894901574
AVERAGE7267.570.5849585.520.570.71
", + "image_path": "b80375c0c9bb1ca15cccb13e9d02894a56d3ece7d1a15a401578a1fdb70a4e62.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 83, + 163, + 529, + 312 + ], + "blocks": [ + { + "bbox": [ + 55, + 129, + 555, + 152 + ], + "lines": [ + { + "bbox": [ + 55, + 129, + 555, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 555, + 152 + ], + "type": "text", + "content": "Table 2. Accuracy (%) of diffusion models in generating recognizable images of G7 politicians. Higher accuracy indicates lower uncertainty in the model's ability to represent specific individuals." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 163, + 529, + 312 + ], + "lines": [ + { + "bbox": [ + 83, + 163, + 529, + 312 + ], + "spans": [ + { + "bbox": [ + 83, + 163, + 529, + 312 + ], + "type": "table", + "html": "
auroc ↑SDXS aupr ↑fpr95↓auroc ↑PixArt aupr ↑fpr95↓auroc ↑SDv1.5 aupr ↑fpr95↓auroc ↑SDXL aupr ↑fpr95↓auroc ↑Average aupr ↑fpr95↓
microscopicDDPM-OOD mse_and_lpips39.61%56.38%100.00%71.64%71.52%93.62%87.25%83.67%51.43%66.17%70.53%81.68%
LMD lpips alex45.82%45.24%99.10%16.81%31.61%99.88%5.75%28.12%100.00%22.79%34.99%99.66%
2XDM lpips alex24.49%32.60%99.91%24.32%33.21%99.99%16.19%32.42%99.94%30.58%36.99%98.25%23.90%33.81%99.52%
PUNC ROUGE (1_recall)84.00%85.10%57.07%95.47%97.06%30.84%87.24%90.66%48.90%83.56%84.31%52.78%87.57%89.28%47.40%
PUNC ROUGE (L_recall)87.64%89.56%55.28%96.96%98.02%16.85%90.20%93.40%43.56%88.19%89.84%47.44%90.75%92.71%40.78%
PUNC BERT (recall)88.45%89.71%48.68%97.83%98.41%10.32%87.75%91.70%51.03%83.10%85.46%58.14%89.28%91.32%42.04%
remote sensingDDPM-OOD mse_and_lpips80.00%85.37%99.28%74.36%71.89%71.00%65.47%58.25%71.97%73.28%71.84%80.75%
LMD lpips alex90.65%93.11%74.49%78.93%79.31%81.05%86.60%85.39%49.19%85.39%85.94%68.24%
2XDM lpips alex53.88%59.21%98.23%54.43%57.99%98.78%30.66%38.80%98.76%23.54%35.35%96.23%40.63%47.83%98.00%
PUNC ROUGE (1_recall)85.59%86.59%63.87%94.14%95.97%49.29%73.24%72.93%79.56%73.07%74.54%83.48%81.51%82.51%69.05%
PUNC ROUGE (L_recall)89.85%90.63%51.96%97.01%97.94%16.08%83.63%84.65%65.35%85.68%87.09%64.00%89.04%90.08%49.34%
PUNC BERT (recall)80.50%82.56%75.19%91.93%93.08%43.30%61.27%64.74%90.08%60.71%65.32%90.94%73.60%76.42%74.88%
TextureDDPM-OOD mse_and_lpips73.22%62.02%89.24%70.21%55.97%99.80%74.24%51.97%84.71%72.56%56.65%91.25%
LMD lpips alex63.40%50.47%97.52%26.74%16.06%99.89%35.66%23.22%99.95%41.93%29.92%99.12%
2XDM lpips alex32.80%20.84%99.98%37.21%23.36%99.37%29.58%24.27%99.93%28.87%16.49%98.88%32.11%21.24%99.54%
PUNC ROUGE (1_recall)40.77%72.88%95.54%83.26%94.80%73.48%46.61%75.64%90.87%49.16%78.04%93.57%54.95%80.34%88.36%
PUNC ROUGE (L_recall)45.64%76.90%95.99%84.10%95.13%74.21%47.60%77.70%93.14%50.71%80.05%94.35%57.01%82.44%89.42%
PUNC BERT (recall)49.11%79.11%94.65%84.40%95.14%69.99%47.60%78.80%93.85%46.72%79.20%96.87%56.96%83.06%88.84%
", + "image_path": "b660b9d8dd8679298c7986b79d8b083a83775bebe3e10881d71011565a1153b1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 82, + 343, + 529, + 529 + ], + "blocks": [ + { + "bbox": [ + 59, + 321, + 550, + 332 + ], + "lines": [ + { + "bbox": [ + 59, + 321, + 550, + 332 + ], + "spans": [ + { + "bbox": [ + 59, + 321, + 550, + 332 + ], + "type": "text", + "content": "Table 3. Performance comparison of PUNC applied to different Out of Distribution concepts and different text-to-image models" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 343, + 529, + 529 + ], + "lines": [ + { + "bbox": [ + 82, + 343, + 529, + 529 + ], + "spans": [ + { + "bbox": [ + 82, + 343, + 529, + 529 + ], + "type": "table", + "html": "
auroc ↑SDXS aupr ↑fpr95↓auroc ↑PixArt aupr ↑fpr95↓auroc ↑SDv1.5 aupr ↑fpr95↓auroc ↑SDXL aupr ↑fpr95↓auroc ↑Average aupr ↑fpr95↓
VagueDDPM-OOD mse.and.Lpips31.41%6.06%96.45%50.06%9.30%96.52%40.59%5.36%97.37%40.69%6.91%96.78%
LMD lipps alex39.26%6.91%96.35%47.62%9.54%97.70%40.77%5.70%97.95%42.55%7.38%97.33%
2XDM lipps alex51.28%10.10%94.35%43.82%8.57%96.89%50.97%10.09%95.96%57.83%10.01%94.37%50.98%9.69%95.39%
PUNC ROUGE (1_precision)99.99%100.00%0.00%100.00%100.00%0.00%99.99%100.00%0.00%99.99%100.00%0.00%99.99%100.00%0.00%
PUNC ROUGE (L_precision)99.99%100.00%0.00%100.00%100.00%0.00%99.99%100.00%0.00%100.00%100.00%0.00%100.00%100.00%0.00%
PUNC BERT (precision)60.63%92.24%82.85%67.02%94.54%83.10%37.35%86.51%92.70%70.85%95.13%78.00%58.96%92.11%84.16%
AdversarialDDPM-OOD mse.and.Lpips39.54%1.78%95.72%50.89%2.54%95.91%38.97%1.10%98.38%43.13%1.81%96.67%
LMD lipps alex43.71%1.95%96.24%53.29%2.89%94.43%41.64%1.17%97.48%46.21%2.00%96.05%
2XDM lipps alex36.94%1.94%98.92%47.95%2.54%96.03%53.04%2.83%95.69%55.19%1.96%96.20%48.28%2.32%96.71%
PUNC ROUGE (1_precision)64.14%97.77%48.80%76.32%98.90%48.00%63.59%97.73%49.50%62.97%97.64%49.20%66.76%98.01%48.87%
PUNC ROUGE (L_precision)63.20%97.75%49.00%76.26%98.90%48.40%64.09%97.90%49.70%63.76%97.87%49.60%66.83%98.11%49.17%
PUNC BERT (precision)73.98%98.72%78.31%72.95%98.93%86.40%45.09%96.93%94.18%72.69%98.80%81.40%66.18%98.34%85.07%
Corrupt l11DDPM-OOD mse.and.Lpips48.87%44.82%87.43%53.20%54.43%95.48%41.02%29.54%96.32%47.70%42.93%93.08%
LMD lipps alex51.23%47.48%89.99%57.53%56.49%90.92%40.94%30.30%96.32%49.90%44.76%92.41%
2XDM lipps alex25.48%36.16%99.29%50.43%49.86%90.10%54.97%54.29%95.69%54.73%38.27%93.29%46.40%44.65%94.59%
PUNC ROUGE (1_precision)28.75%38.34%98.77%54.58%55.29%96.02%27.69%37.57%98.78%27.45%37.27%98.54%34.62%42.12%98.03%
PUNC ROUGE (L_precision)27.29%38.11%99.15%54.32%55.59%96.16%27.30%38.07%99.45%27.35%37.89%99.37%34.07%42.42%98.53%
PUNC BERT (precision)85.85%84.45%73.94%79.43%78.83%87.21%54.97%53.73%94.51%75.76%72.91%84.65%74.00%72.48%85.08%
Corrupt l12DDPM-OOD lipps alex43.23%43.94%91.23%59.46%60.14%93.71%48.79%34.44%95.74%50.49%46.17%93.56%
LMD lipps alex46.90%45.65%92.04%59.94%58.90%90.54%40.88%30.42%96.35%49.24%44.99%92.98%
2XDM lipps alex58.15%58.50%93.98%52.08%53.36%90.19%53.57%54.17%96.34%60.24%44.23%91.78%56.01%52.57%93.07%
PUNC ROUGE (1_precision)28.76%38.10%98.59%54.77%55.49%96.02%27.66%37.42%98.75%27.39%37.23%98.53%34.64%42.06%97.97%
PUNC ROUGE (L_precision)27.33%37.87%99.16%54.51%55.87%96.11%27.31%37.92%99.35%27.33%37.83%99.36%34.12%42.37%98.50%
PUNC BERT (precision)85.96%84.23%73.54%79.47%78.80%87.05%54.78%53.51%94.79%76.04%73.44%84.58%74.06%72.49%84.99%
", + "image_path": "4433cec5bf99037f5c9ec0bab56b699020f6329088988d04ac39c4039ab20c8f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 538, + 489, + 549 + ], + "lines": [ + { + "bbox": [ + 116, + 538, + 489, + 549 + ], + "spans": [ + { + "bbox": [ + 116, + 538, + 489, + 549 + ], + "type": "text", + "content": "Table 4. Performance comparison all applications (Vague, Adversarial, Corrupt Lvl1, Corrup Lvl2)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 571, + 239, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 239, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 239, + 582 + ], + "type": "text", + "content": "models with ethical and regulatory standards." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 588, + 295, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 295, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 295, + 672 + ], + "type": "text", + "content": "Our contributions also include a dataset specifically tailored to uncertainty quantification in text-to-image models, which we hope will serve as a benchmark for future research. Ultimately, this work highlights the potential of prompt-based uncertainty estimation as a critical tool for enhancing the reliability and safety of generative models in real-world applications." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Limitations. While PUNC demonstrates promising results in quantifying uncertainty for text-to-image generation, several limitations remain. First, our approach" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 570, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 643 + ], + "type": "text", + "content": "relies heavily on the interpretive abilities of LVLMs, which, despite their effectiveness, may introduce biases or inaccuracies inherited from their training data. Second, our analysis focuses primarily on English-language prompts, limiting our method's applicability to multilingual or culturally diverse datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 657, + 555, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 657, + 555, + 706 + ], + "spans": [ + { + "bbox": [ + 313, + 657, + 555, + 706 + ], + "type": "text", + "content": "Acknowledgments. This work was performed using HPC resources from GENCI-IDRIS (Grant 2023-[AD011011970R3]), and EuroHPC Development access to LEONARDO, hosted by CINECA." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8069" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "type": "text", + "content": "[2] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 202, + 295, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 295, + 235 + ], + "type": "text", + "content": "[3] Nouar AlDahoul, Talal Rahwan, and Yasir Zaki. Aigenerated faces free from racial and gender stereotypes. arXiv preprint arXiv:2402.01002, 2024. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 296, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 296, + 302 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 296, + 302 + ], + "type": "text", + "content": "[4] Georges Le Bellier and Nicolas Audebert. Detecting Out-Of-Distribution Earth Observation Images with Diffusion Models. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 481–491, Los Alamitos, CA, USA, 2024. IEEE Computer Society. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 302, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 295, + 346 + ], + "type": "text", + "content": "[5] Lucas Berry, Axel Brando, and David Meger. Shedding light on large generative networks: Estimating epistemic uncertainty in diffusion models. In The 40th Conference on Uncertainty in Artificial Intelligence, 2024. 2, 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 346, + 295, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 295, + 390 + ], + "type": "text", + "content": "[6] Matthew A Chan, Maria J Molina, and Christopher A Metzler. Hyper-diffusion: Estimating epistemic and aleatoric uncertainty with a single model. arXiv preprint arXiv:2402.03478, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 392, + 296, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 296, + 446 + ], + "type": "text", + "content": "[7] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart" + }, + { + "bbox": [ + 62, + 392, + 296, + 446 + ], + "type": "inline_equation", + "content": "\\backslash" + }, + { + "bbox": [ + 62, + 392, + 296, + 446 + ], + "type": "text", + "content": "sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 3, 6, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 447, + 296, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 296, + 513 + ], + "type": "text", + "content": "[8] Junsong Chen, Jincheng YU, Chongjian GE, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-" + }, + { + "bbox": [ + 62, + 447, + 296, + 513 + ], + "type": "inline_equation", + "content": "\\mathbb{S}\\backslash \\mathbb{alpha}\\mathbb{S}" + }, + { + "bbox": [ + 62, + 447, + 296, + 513 + ], + "type": "text", + "content": ": Fast training of diffusion transformer for photorealistic text-to-image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 514, + 296, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 296, + 557 + ], + "type": "text", + "content": "[9] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In Proceedings of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 558, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 295, + 601 + ], + "type": "text", + "content": "[10] Charles Corbière, Nicolas Thome, Avner Bar-Hen, Matthieu Cord, and Patrick Pérez. Addressing failure prediction by learning model confidence. Advances in Neural Information Processing Systems, 32, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 602, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 668 + ], + "type": "text", + "content": "[11] Matt Deitke, Christopher Clark, Sangho Lee, Rohun Tripathi, Yue Yang, Jae Sung Park, Mohammadreza Salehi, Niklas Muennighoff, Kyle Lo, Luca Soldaini, et al. Molmo and pixmo: Open weights and open data for state-of-the-art multimodal models. arXiv preprint arXiv:2409.17146, 2024. 5, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 669, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 669, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 669, + 295, + 713 + ], + "type": "text", + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7, 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[13] Ning Ding, Yehui Tang, Zhongqian Fu, Chao Xu, Kai Han, and Yunhe Wang. Gpt4image: Can large pre-trained models help vision models on perception tasks? arXiv e-prints, pages arXiv-2306, 2023. 7, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 119, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 163 + ], + "type": "text", + "content": "[14] Zhekai Du and Jingjing Li. Diffusion-based probabilistic uncertainty estimation for active domain adaptation. Advances in Neural Information Processing Systems, 36: 17129-17155, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 164, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 219 + ], + "type": "text", + "content": "[15] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 5, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 222, + 553, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 222, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 316, + 222, + 553, + 255 + ], + "type": "text", + "content": "[16] Nikita Durasov, Timur Bagautdinov, Pierre Baque, and Pascal Fua. Masksembles for uncertainty estimation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 256, + 553, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 256, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 316, + 256, + 553, + 311 + ], + "type": "text", + "content": "[17] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 314, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 314, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 314, + 553, + 392 + ], + "type": "text", + "content": "[18] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In Proceedings of the 41st International Conference on Machine Learning, pages 12606-12633. PMLR, 2024. 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 393, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 426 + ], + "type": "text", + "content": "[19] Gianni Franchi, Andrei Bursuc, Emanuel Aldea, Séverine Dubuisson, and Isabelle Bloch. TRADI: Tracking deep neural network weight distributions. In ECCV, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 428, + 553, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 428, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 316, + 428, + 553, + 483 + ], + "type": "text", + "content": "[20] Gianni Franchi, Xuanlong Yu, Andrei Bursuc, Emanuel Aldea, Severine Dubuisson, and David Filliat. Latent discriminant deterministic uncertainty. In Computer Vision – ECCV 2022, pages 243–260, Cham, 2022. Springer Nature Switzerland. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 484, + 553, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 484, + 553, + 508 + ], + "spans": [ + { + "bbox": [ + 316, + 484, + 553, + 508 + ], + "type": "text", + "content": "[21] Yarin Gal. Uncertainty in Deep Learning. PhD thesis, University of Cambridge, 2016. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 509, + 553, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 509, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 317, + 509, + 553, + 563 + ], + "type": "text", + "content": "[22] Yonatan Geifman and Ran El-Yaniv. Selective classification for deep neural networks. In Proceedings of the 31st International Conference on Neural Information Processing Systems, page 4885-4894, Red Hook, NY, USA, 2017. Curran Associates Inc. 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 565, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 565, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 317, + 565, + 553, + 632 + ], + "type": "text", + "content": "[23] Mark S Graham, Walter HL Pinaya, Petru-Daniel Tudosiu, Parashkev Nachev, Sebastien Ourselin, and Jorge Cardoso. Denoising diffusion models for out-of-distribution detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2948-2957, 2023. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 634, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 634, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 634, + 553, + 678 + ], + "type": "text", + "content": "[24] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In International Conference on Learning Representations, 2017. 3, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 679, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 679, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 679, + 553, + 713 + ], + "type": "text", + "content": "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8070" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[26] Eyke Hüllermeier and Willem Waegeman. *Aleatoric and epistemic uncertainty in machine learning: an introduction to concepts and methods*. *Machine Learning*, 110(3), 2021. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 139 + ], + "type": "text", + "content": "[27] Pavel Izmailov, Sharad Vikram, Matthew D Hoffman, and Andrew Gordon Gordon Wilson. What are bayesian neural network posteriors really like? In ICML, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "type": "text", + "content": "[28] John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, et al. Highly accurate protein structure prediction with alphafold. nature, 596(7873):583-589, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 194, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 239 + ], + "type": "text", + "content": "[29] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems, pages 26565-26577. Curran Associates, Inc., 2022. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 272 + ], + "type": "text", + "content": "[30] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "type": "text", + "content": "[31] Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. In The Eleventh International Conference on Learning Representations, 2023. 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 316, + 294, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 349 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 349 + ], + "type": "text", + "content": "[32] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. In NeurIPS, 2017. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "type": "text", + "content": "[33] Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, et al. Learning skillful medium-range global weather forecasting. Science, 382(6677):1416-1421, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 404, + 294, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 448 + ], + "type": "text", + "content": "[34] Olivier Laurent, Adrien Lafage, Enzo Tartaglione, Geoffrey Daniel, Jean-Marc Martinez, Andrei Bursuc, and Gianni Franchi. Packed-ensembles for efficient uncertainty estimation. In ICLR, 2023. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 449, + 294, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 492 + ], + "type": "text", + "content": "[35] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 7, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 492, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 294, + 525 + ], + "type": "text", + "content": "[36] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6, 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 525, + 294, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 294, + 570 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 294, + 570 + ], + "type": "text", + "content": "[37] Zhiqiu Lin, Xinyue Chen, Deepak Pathak, Pengchuan Zhang, and Deva Ramanan. Revisiting the role of language priors in vision-language models. In *Forty-first International Conference on Machine Learning*. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "type": "text", + "content": "[38] Krzysztof Lis, Krishna Nakka, Pascal Fua, and Mathieu Salzmann. Detecting the unexpected via image resynthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2152-2161, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "type": "text", + "content": "[39] Zhenzhen Liu, Jin Peng Zhou, Yufan Wang, and Kilian Q Weinberger. Unsupervised out-of-distribution detection with diffusion inpainting. In International Conference on Machine Learning, pages 22528-22538. PMLR, 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[40] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan LI, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. In Advances in Neural Information Processing Systems, pages 5775-5787. Curran Associates, Inc., 2022. 4" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "type": "text", + "content": "[41] Andrey Malinin and Mark Gales. Predictive uncertainty estimation via prior networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2018. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 107, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 107, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 315, + 107, + 553, + 129 + ], + "type": "text", + "content": "[42] Radford M Neal et al. Mcmc using hamiltonian dynamics. Handbook of markov chain monte carlo, 2011. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "type": "text", + "content": "[43] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 186, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 240 + ], + "type": "text", + "content": "[44] Gangadhar Payyavula. Nuclei segmentation in microscope cell images. https://www.kaggle.com/datasets/gangadhar/nuclei-segmentation-in-microscope-cell-images, 2018. Accessed: 2024-09-01. 7, 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 241, + 553, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 264 + ], + "type": "text", + "content": "[45] Tim Pearce, Alexandra Brintrup, and Jun Zhu. Understanding softmax confidence and uncertainty, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "text", + "content": "[46] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 310, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 342 + ], + "type": "text", + "content": "[47] Jeremy Pinto. Bigearthnet. https://github.com/ jerpint/bigearthnet, 2022. Accessed: 2024-09-01. 7, 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 344, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 553, + 399 + ], + "type": "text", + "content": "[48] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2023. 2, 4, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 400, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 400, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 400, + 553, + 454 + ], + "type": "text", + "content": "[49] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Proceedings of the 38th International Conference on Machine Learning, pages 8821-8831. PMLR, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 456, + 553, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 511 + ], + "type": "text", + "content": "[50] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 512, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 578 + ], + "type": "text", + "content": "[51] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "type": "text", + "content": "[52] Jaskirat Singh and Liang Zheng. Divide, evaluate, and refine: Evaluating and improving text-to-image alignment with iterative vqa feedback. Advances in Neural Information Processing Systems, 36:70799-70811, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 624, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 657 + ], + "type": "text", + "content": "[53] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[54] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 4" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8071" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[55] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. arXiv preprint arXiv:2409.17146, 2024. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 117, + 294, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 117, + 294, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 117, + 294, + 182 + ], + "type": "text", + "content": "[56] Tomas Vojir, Tomáš Šipka, Rahaf Aljundi, Nikolay Chumerin, Daniel Olmeda Reino, and Jiri Matas. Road anomaly detection by partial image reconstruction with segmentation coupling. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15651-15660, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 183, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 183, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 183, + 294, + 205 + ], + "type": "text", + "content": "[57] Max Welling and Yee W Teh. Bayesian learning via stochastic gradient Langevin dynamics. In ICML, 2011. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 237 + ], + "type": "text", + "content": "[58] Yeming Wen, Dustin Tran, and Jimmy Ba. BatchEnsemble: an alternative approach to efficient ensemble and lifelong learning. In ICLR, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "text", + "content": "[59] Guoxuan Xia and Christos-Savvas Bouganis. Augmenting softmax information for selective classification with out-of-distribution data. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 1995-2012, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 283, + 294, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 315 + ], + "type": "text", + "content": "[60] Guoxuan Xia and Christos-Savvas Bouganis. On the usefulness of deep ensemble diversity for out-of-distribution detection, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "type": "text", + "content": "[61] Guoxuan Xia and Christos-Savvas Bouganis. Window-based early-exit cascades for uncertainty estimation: When deep ensembles are more efficient than single models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17368-17380, 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 372, + 294, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 425 + ], + "type": "text", + "content": "[62] Guoxuan Xia, Duolikun Danier, Ayan Das, Stathi Fotiadis, Farhang Nabiei, Ushnish Sengupta, and Alberto Bernacchia. Score Normalization for a Faster Diffusion Exponential Integrator Sampler. In NeurIPS 2023 Workshop on Diffusion Models, 2023. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 426, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 459 + ], + "type": "text", + "content": "[63] Guoxuan Xia, Olivier Laurent, Gianni Franchi, and Christos-Savvas Bouganis. Towards understanding why label smoothing degrades selective classification and how to fix it, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 460, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 525 + ], + "type": "text", + "content": "[64] Yingda Xia, Yi Zhang, Fengze Liu, Wei Shen, and Alan L Yuille. Synthesize then compare: Detecting failures and anomalies for semantic segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 145-161. Springer, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 526, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 569 + ], + "type": "text", + "content": "[65] Yilun Xu, Mingyang Deng, Xiang Cheng, Yonglong Tian, Ziming Liu, and Tommi S. Jaakkola. Restart sampling for improving generative processes. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "type": "text", + "content": "[66] Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. Generalized out-of-distribution detection: A survey. International Journal of Computer Vision, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 603, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 658 + ], + "type": "text", + "content": "[67] Michal Yarom, Yonatan Bitton, Soravit Changpinyo, Roeah Aharoni, Jonathan Herzig, Oran Lang, Eran Ofek, and Idan Szpektor. What you see is what you read? improving text-image alignment evaluation. Advances in Neural Information Processing Systems, 36:1601-1619, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 690 + ], + "type": "text", + "content": "[68] Xuanwu Yin Yuda Song, Zehao Sun. Sdxs: Real-time one-step latent diffusion models with image conditions. arxiv, 2024. 6, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[69] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. In The Eleventh" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 285 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "International Conference on Learning Representations, 2023.4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[70] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "text", + "content": "[71] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 208, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 240 + ], + "type": "text", + "content": "[72] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675, 2019. 6, 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "text", + "content": "[73] Yimeng Zhang, Jinghan Jia, Xin Chen, Aochuan Chen, Yihua Zhang, Jiancheng Liu, Ke Ding, and Sijia Liu. To generate or not? safety-driven unlearned diffusion models are still easy to generate unsafe images ... for now, 2024. 7, 3" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8072" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network/c23bce64-f04c-4e24-b700-3d6473e34043_content_list.json b/2025/Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network/c23bce64-f04c-4e24-b700-3d6473e34043_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b1a4e33f01db080a72e49672deaf1888ba10bdb7 --- /dev/null +++ b/2025/Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network/c23bce64-f04c-4e24-b700-3d6473e34043_content_list.json @@ -0,0 +1,1359 @@ +[ + { + "type": "text", + "text": "Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network", + "text_level": 1, + "bbox": [ + 107, + 128, + 890, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haifeng Zhang, Qinghui He, Xiuli Bi, Weisheng Li, Bo Liu*, Bin Xiao Chongqing University of Posts and Telecommunications", + "bbox": [ + 209, + 203, + 785, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "s220331143@stu.cqupt.edu.cn, s220232014@stu.cqupt.edu.cn", + "bbox": [ + 250, + 241, + 746, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{bix1, liws, boliu, xiaobin}@cqupt.edu.cn", + "bbox": [ + 313, + 258, + 676, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The rapid advancement of generative models has significantly improved the quality of generated images. Meanwhile, it challenges information authenticity and credibility. Current generated image detection methods based on large-scale pre-trained multimodal models have achieved impressive results. Although these models provide abundant features, the authentication task-related features are often submerged. Consequently, those authentication task-irrelated features cause models to learn superficial biases, thereby harming their generalization performance across different model genera (e.g., GANs and Diffusion Models). To this end, we proposed VIB-Net, which uses Variational Information Bottlenecks to enforce authentication task-related feature learning. We tested and analyzed the proposed method and existing methods on samples generated by 17 different generative models. Compared to SOTA methods, VIB-Net achieved a $5.55\\%$ improvement in mAP and a $9.33\\%$ increase in accuracy. Notably, in generalization tests on unseen generative models from different series, VIB-Net improved mAP by $12.48\\%$ and accuracy by $23.59\\%$ over SOTA methods. The code is available at https://github.com/oceanzhf/VIBAIGCDetect.", + "bbox": [ + 89, + 340, + 485, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 700, + 220, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The continuous advancements in deep learning have led to the emergence of numerous well-known image generation models, such as GANs[9][15] and Diffusion[12][31]. Images generated by these models have reached a level where observers can hardly distinguish them from real images, and the flexibility of tools based on these models has also greatly improved [6]. With a few simple steps, people can easily obtain high-quality images as needed. While this presents a tremendous opportunity for fields like visual arts, it also serves as a paradise for individuals looking to fab", + "bbox": [ + 89, + 727, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/432848555de54ec22b2ec2160073a539e39993a3a52b1f214925a0ea15c4b742.jpg", + "image_caption": [ + "Figure 1. Each method's overall performance('P') on test data and generalization performance ('G') on unseen data are linked by dashed lines. Our detector demonstrates a clear advantage over all SOTA methods in both metrics." + ], + "image_footnote": [], + "bbox": [ + 516, + 308, + 905, + 486 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ricate misinformation. Given the ease of creating manipulated images today and the potential risk of using them for malicious purposes, there is an urgent need to develop detectors that help distinguish between real and generated images.", + "bbox": [ + 511, + 580, + 906, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing detection methods [4][39][25][43] focus on identifying specific traces left by generators in either the spatial or frequency domain of generated images, thus establishing a decision boundary between real and generated images. These detection methods are highly effective in detecting images generated by GANs and earlier diffusion models. However, their performance drops significantly when facing images generated by recent, popular models that visually resemble real images. Although some error reconstruction-based methods [40][23] have been developed specifically for detecting images generated by diffusion models, they similarly experience performance declines when applied to GAN-generated images. Consequently, current research increasingly focuses on improving the cross-model genera detection capabilities of detectors.", + "bbox": [ + 511, + 657, + 908, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generalization across different domains is a primary re", + "bbox": [ + 531, + 885, + 903, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding Author", + "bbox": [ + 107, + 887, + 241, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23828", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "search challenge in deep learning. In this context, the emergence of large-scale pre-trained multimodal models has introduced new approaches and promising results. For instance, by using the image encoder from the CLIP [30] model to extract image features and training a classifier with linear probing, there is a notable improvement in generalization performance over models like ResNet50[28][6]. However, compared to traditional classification tasks, real and generated images share subtle, harder-to-capture distinguishing features. This implies that the features extracted by large-scale pre-trained models for the task of distinguishing real from fake images may contain a considerable amount of task-irrelevant features. Task-irrelevant features can either have no impact on the realizness of the image or, though effective in isolation, become redundant when combined with other features. These features not only interfere with the classifier's ability to accurately learn decision boundaries but also hinder the model's ability to capture and distinguish task-relevant characteristics. More critically, task-irrelevant features may lead the model to learn superficial biases. For example, if a fake image contains a tree while the real image does not, the model might erroneously associate the presence of the tree with being a fake image. Consequently, when detecting other images, the model may mistakenly classify a real image as fake if it contains a tree. Therefore, when dealing with high-quality generated images with subtle distinguishing features, it is challenging to fine-tune the model to separate essential from unnecessary information. This can lead to an overfitting of the incidental statistical correlations between unnecessary information and target images, ultimately undermining the model's generalization and discrimination capabilities.", + "bbox": [ + 93, + 90, + 480, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "How can we harness the powerful feature extraction capability of large models while minimizing the interference of unnecessary information? By approaching this from a novel perspective, we propose addressing the issue using Information Bottleneck Theory. This theory aims to create latent encodings that not only effectively represent the target but also maximally compress the original input, thereby avoiding the impact of task-irrelevant features on the model. Specifically, we introduce a method called VIB-Net, which couples a large-scale pre-trained multimodal model with Variational Information Bottleneck (VIB) [1]. Considering that feature representations extracted by the model's image encoder are more task-specific, we decouple and filter these features using latent variables. This method maximizes the mutual information between the latent feature and the classification output, while minimizing the mutual information between the latent feature and the input, resulting in a concise feature that should only contain the effective information needed for the task. This concise latent feature helps the model avoid overfitting to irrelevant features and the incidental correlations with the task, preventing the model", + "bbox": [ + 93, + 584, + 480, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "from learning superficial biases, which is beneficial to the model's generalization performance.", + "bbox": [ + 516, + 92, + 903, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows: (1) To the best of our knowledge, we are the first to apply Information Bottleneck theory and techniques to the detection of generated images, demonstrating the potential of this theory and approach in this field. (2) We analyze the limitations of methods based on large-scale pre-trained models in detecting images generated by unseen generative models, and present our theory of what could be wrong with the existing paradigm. (3) We propose VIB-Net for the detection of generated images, which uses VIB to decouple and filter the general features extracted by large-scale pre-trained multimodal models, removing task-irrelevant information to enhance detection performance. (4) A large number of experiments validate the advantages of our method (see Fig. 1), especially in the improvement of generalization performance $(+23.59\\%$ accuracy).", + "bbox": [ + 516, + 122, + 903, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 516, + 377, + 648, + 392 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generalization Performance. The continuous advancement of generated image models presents ongoing challenges for generated image detection, making the generalization performance of detection models on unseen generators an increasing focus for researchers[14][21]. [39] explored the effects of data preprocessing and augmentation methods on the generalization performance of a ResNet50[11]-based detector. [44] argues that the differences between real and fake images are subtle and localized, framing forgery detection as a fine-grained task that could positively impact generalization performance. [3] takes a frequency-domain approach, mapping real images to a dense subspace and synthetic images outside this subspace to enhance generalization. While these methods have achieved certain successes on different models within the same generation series, they show limitations across different series. This is due to the fact that, with the development of generative algorithms, subtle traces in generated images are becoming increasingly difficult to capture. Additionally, artifacts in the frequency domain differ between model series and are gradually decreasing in visibility.", + "bbox": [ + 516, + 402, + 903, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The impressive results of large-scale pre-trained models in terms of transferability and generalization have sparked significant interest among researchers in the field of generated image detection. [28] argues that existing detectors perform asymmetric learning, where they use CLIP's image encoder to extract image features, exploring the potential of large models in detecting different model families. [6] suggests that detectors based on large models rely on higher-level semantic features, which enhances the generalization ability of the detectors. [19] uses both image and text features extracted by the CLIP model, and then employs a multilayer perceptron to detect images generated by", + "bbox": [ + 516, + 720, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23829", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "diffusion models, improving generalization performance. Information Bottleneck. The Information Bottleneck (IB) [38] is an innovative information-theoretic method used for pattern analysis and representation learning. It provides an elegant balance between data compression and information retention[13]. VIB [1] is a variational approximation of IB, aimed at parameterizing IB using neural networks. [2] demonstrated VIB's capability in improving classification calibration and out-of-distribution data. [24] explored VIB as a regularization technique to address overfitting issues in NLP models. [41] proposed a dual-stream information bottleneck method to utilize in-distribution data to enhance out-of-distribution target detection tasks.", + "bbox": [ + 89, + 90, + 485, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on the above analysis, current methods for detecting generated images using large-scale pre-trained models simply utilize the general representations extracted from images, without thoroughly exploring the relationship between these general features and downstream tasks. Additionally, Information Bottleneck is an effective technique for improving performance in cross-domain and cross-distribution tasks. Therefore, unlike existing methods, our approach leverages Information Bottleneck techniques to explore the potential of large-scale pre-trained models in generated image detection from a completely new perspective, achieving outstanding performance.", + "bbox": [ + 89, + 289, + 483, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 484, + 230, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem background", + "text_level": 1, + "bbox": [ + 89, + 511, + 290, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A common paradigm for using large-scale pretrained models is to leverage a large-scale pretrained multimodal model (e.g., CLIP) to learn the general representation of real and generated images, and then apply this feature representation to downstream tasks[28][6]. Building on their findings, we further consider the following points: (1) It is well-known that the features extracted by large-scale pre-trained visual encoders are general-purpose, meaning they contain many general features, such as semantic information. (2) Another widely accepted notion is that some critical features for distinguishing real from generated images are subtle and local.", + "bbox": [ + 89, + 532, + 483, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Further combining these two points, we propose the following hypotheses: (i) The general features extracted by large-scale pre-trained multimodal models are more beneficial for real-fake image classification tasks than those extracted by pre-trained visual models. (ii) These general features contain task-irrelevant features for real-fake classification. (iii) These task-irrelevant features affect the classification of real and fake images. Next, we will delve into these three hypotheses.", + "bbox": [ + 89, + 700, + 483, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Analysis of general feature representation", + "text_level": 1, + "bbox": [ + 89, + 847, + 447, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the first hypothesis, We use image pairs of real and generated images to aid our analysis[6], generating textual de", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e1410341985ff647e745d20a946c655965b19f0cbbf303dfb911f10c27cde35e.jpg", + "image_caption": [ + "Figure 2. UMAP visualization of features extracted from Dalle2 and Dalle3 generated images using large-scale pretrained multimodal(CLIP) and unimodal models(ResNet)." + ], + "image_footnote": [], + "bbox": [ + 521, + 89, + 905, + 308 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "scriptions for 1,000 real images and using these descriptions to produce 1,000 corresponding images with DALL-E 2 and DALL-E 3. We then compare features extracted by large-scale pre-trained multimodal models(CLIP[30]) and pre-trained visual models(ResNet[11]) through UMAP [26] visualization of 3,000 images. The results show that features extracted by multimodal models (e.g., CLIP) exhibit greater distribution differences for real images and DALL-E 2-generated images, suggesting multimodal models may be better at distinguishing real from generated images. However, for DALL-E 3-generated images, both model types show similar feature distributions, possibly due to advancements in image generation tools, which complicate the classification task and require further fine-tuning.", + "bbox": [ + 511, + 397, + 906, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To further explore hypotheses (ii) and (iii), we visualized the extracted features using both category and authenticity as labels to observe their distribution differences. Specifically, we prepared images of two categories (car and cat), each containing 500 real images and 500 generated images. We then used CLIP to extract their features and performed UMAP dimensionality reduction for visualization, with the results shown in Fig. 3a. We can see that the features extracted by CLIP clearly form two clusters: the left cluster mainly consists of light blue and light red points, while the right cluster is mainly composed of light pink and light purple points. These two clusters correspond to category 1 and category 2, respectively. This indicates that CLIP can significantly distinguish category information. However, this clear separation of category information may affect the decision-making process for classifying real vs. generated images. When examining the distinction between real and generated images, we notice that the right cluster shows a noticeable separation of real and generated features, with", + "bbox": [ + 511, + 613, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23830", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fe8f885df46bbd6b886ae2693b1dc8ce8b2ffa83bcc61143cf31da8d0c17917d.jpg", + "image_caption": [ + "(a) Visualization of features extracted by CLIP.", + "Figure 3. UMAP visualization of features extracted from real and generated images of two category using CLIP and VIB-Net." + ], + "image_footnote": [], + "bbox": [ + 91, + 89, + 284, + 212 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1ab822a33b6db285e7ab7f352b8aedae2588e0ed8278793c8c97bcbb3a4451d0.jpg", + "image_caption": [ + "(b) Visualization of features extracted by VIB-Net." + ], + "image_footnote": [], + "bbox": [ + 285, + 90, + 480, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "two different colors observed within the cluster. This suggests that CLIP is capable of distinguishing real and generated features within category 2. However, in the left cluster, CLIP does not show significant ability to differentiate between real and generated features. We observe that in the left cluster (category 1), the two colors are mixed together, showing poor distinguishability. From these observations, it is easy to infer that the internal separation of real or generated images caused by semantic category information makes it more difficult for the classifier to learn the decision boundary. More importantly, the strong sensitivity of large-scale pre-trained multimodal models like CLIP to semantic information leads the model to tend to learn superficial biases.", + "bbox": [ + 88, + 313, + 482, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on this viewpoint, we consider whether it is possible to minimize the irrelevant features within these multimodal features while retaining the ones that are relevant to the classification task. This would help the model avoid learning superficial biases, allow the classifier to more easily learn decision boundaries, and thereby enhance the model's discrimination ability and generalization capability. This is the motivation of this paper. As shown in Fig. 3b, the method(VIB-Net) proposed in this paper achieves significant results.", + "bbox": [ + 88, + 527, + 482, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Approach", + "text_level": 1, + "bbox": [ + 89, + 705, + 199, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the issues analyzed in Sec. 3, an effective approach is to filter the multimodal features in order to obtain the representations most beneficial for the real-fake classification task, which serves as the starting point of this paper. Overall, our method is illustrated in Fig. 4. VIB attempts to filter out irrelevant features and to find the most concise representation, making it easier for the classifier to learn the features needed for distinguishing between real and fake tasks, thereby significantly improving the accuracy of the classification task. Below, we will provide a detailed description of our method.", + "bbox": [ + 88, + 734, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Information bottleneck constraint", + "text_level": 1, + "bbox": [ + 511, + 90, + 808, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To specifically optimize the filtering of irrelevant information in the input features while retaining task-relevant information, we apply the information bottleneck (IB) principle as a constraint on the features to be optimized. Let $X$ , $Z$ , and $Y$ represent the input, the latent compressed representation, and the output target, respectively. The goal of IB is to find the maximum compressed representation $Z$ of the input $X$ , while preserving as much information about the output $Y$ as possible. The maximization objective function is defined as follows:", + "bbox": [ + 511, + 112, + 906, + 263 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {I B} = I (Z, Y) - \\beta I (Z, X) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 272, + 906, + 289 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where $I(.,.)$ denotes mutual information and $\\beta$ is a trade-off parameter. The first term constrains the latent compressed representation to retain sufficient information for downstream tasks, while the second term ensures that irrelevant and redundant information is filtered out.", + "bbox": [ + 511, + 296, + 906, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Variational information bottleneck", + "text_level": 1, + "bbox": [ + 511, + 380, + 815, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To parameterize the information bottleneck (IB) using deep neural networks, we perform variational inference on Equation (1). First, based on the definition of mutual information, Equation (1) can be rewritten as:", + "bbox": [ + 511, + 402, + 905, + 463 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {I B} = \\int d y d z p (y, z) \\log {\\frac {p (y | z)}{p (y)}} \\\\ - \\beta \\int d x d z p (x, z) \\log \\frac {p (z \\mid x)}{p (z)} \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 482, + 906, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where $p(y|z)$ is defined by the encoder and the Markov chain as follows:", + "bbox": [ + 511, + 556, + 905, + 585 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np (y \\mid z) = \\int d x \\frac {p (y \\mid x) p (z \\mid x) p (x)}{p (z)} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 592, + 905, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since it is not computable here, we consider setting $q(y|z)$ as a parameterized approximation of $p(y|z)$ , serving as a \"decoder,\" which can be approximated by a linear layer. We introduce the Kullback-Leibler (KL) divergence to obtain an estimizable representation. As the KL divergence satisfies $\\mathrm{KL}(p(y|z),q(y|z))\\geq 0$ , and by applying the Markov assumption, $p(y,z)$ can be rewritten as $p(y,z) = \\int dx p(x)p(y|x)p(z|x)$ , thereby providing a lower bound for the first term of the objective function:", + "bbox": [ + 511, + 632, + 906, + 768 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI (Z, Y) \\geq \\int d x d y d z p (x) p (y | x) p (z | x) \\log q (y | z) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 775, + 903, + 818 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At the same time, let $r(z)$ be an estimate of the prior probability $p(z)$ , used to approximate the marginal distribution, thereby obtaining an upper bound for the second term:", + "bbox": [ + 511, + 819, + 903, + 864 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI (Z, X) \\leq \\int d x d z p (x) p (z | x) \\log \\frac {p (z | x)}{r (z)} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 871, + 903, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23831", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f2654d25fbfe695a4e4ecf0ae41380f6fa32b64f9234168d03dc86580330d688.jpg", + "image_caption": [ + "Figure 4. Overview of VIB-Net. We utilize Variational Information Bottleneck (VIB) techniques to map the image features extracted by CLIP's visual encoder into latent features $z$ , where the information in $z$ is filtered based on the Information Bottleneck (IB). Subsequently, the latent features $z$ are used as input to an MLP to learn the decision boundary." + ], + "image_footnote": [], + "bbox": [ + 89, + 88, + 903, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The encoder we use is defined as $p(z|x) = p_{\\theta}(z|x) = \\mathcal{N}\\left(z|f^{\\mu}(x), f^{\\Sigma}(x)\\right)$ , where $f(.)$ represents a shallow MLP, $\\mu$ denotes the K-dimensional mean vector of the output $z$ , and $\\Sigma$ represents the $K \\times K$ covariance matrix. We select a parameterized Gaussian distribution for both $p_{\\theta}(z|x)$ and $r(z)$ to allow for the analytical computation of KL divergence. Additionally, we employ the reparameterization trick to estimate gradients [18]. Combining all the components above yields the following objective function:", + "bbox": [ + 88, + 378, + 483, + 517 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {V I B} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} z \\sim p _ {\\theta} (z | x) [ - \\log q (y _ {\\mathrm {n}} | z _ {n}) ] \\\\ + \\beta \\mathrm {K L} [ p _ {\\theta} (z | x _ {n}), r (z) ] \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 537, + 482, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Workflow of the overall method", + "text_level": 1, + "bbox": [ + 89, + 607, + 372, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 4, the parts in VIB-Net work in the following way: First, the real and generated images $\\{x_{1}, x_{2}, x_{3}, \\dots, x_{n}\\}$ are input into the CLIP image encoder, yielding the corresponding image features $\\{t_{1}, t_{2}, t_{3}, \\dots, t_{n}\\}$ . Next, these feature vectors are passed into the Variational Information Bottleneck (VIB). Specifically, the image features are processed through a shallow MLP, then a linear layer $f(.)$ calculates $\\mu$ and $\\Sigma$ , which are transformed using a softmax activation. This layer has $K$ dimensions, representing the bottleneck size. At this point, we obtain the latent compressed feature $z$ , which is then fed into a linear layer $q(y|z)$ to output the target task label $y$ .", + "bbox": [ + 89, + 628, + 483, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 823, + 223, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Settings", + "text_level": 1, + "bbox": [ + 89, + 848, + 189, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To better evaluate the model's performance, we utilize two mainstream public datasets: ForenSynths[39] and", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Genimage[46]. The ForenSynths dataset primarily consists of images generated by GANs, while the Genimage dataset mainly includes images generated by diffusion models. Consequently, we set up two training and testing schemes: (1) Training on ForenSynths and then testing on both ForenSynths and Genimage, and (2) similarly, training on Genimage and testing on both Genimage and ForenSynths. This setup allows us to more effectively validate the generalization performance of the detector on cross model genera scenarios. The details are as follows:", + "bbox": [ + 511, + 378, + 906, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training set:", + "text_level": 1, + "bbox": [ + 513, + 532, + 607, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The ForenSynths training set consists of 20 different categories, each containing 18,000 real images from the LSUN dataset [42] and an equal number of images generated using ProGAN[15]. Following the setup in Genimage, we use images generated by Stable Diffusion v1.4 [32] in the Genimage dataset, along with their corresponding real images, as the training set. In total, these include 32,000 real and generated images.", + "bbox": [ + 511, + 547, + 905, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Testing set:", + "text_level": 1, + "bbox": [ + 513, + 670, + 596, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "11 models from ForenSynths The test set includes images generated by 11 different models, primarily from the GAN family[15][45][29][5][16]. The real images are sourced from six datasets, including ImageNet[34], LSUN[42], FaceForensics++[33], CelebA-HQ[15], CelebA[22], and COCO[20].", + "bbox": [ + 511, + 686, + 905, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "7 models from Genimage To validate the method across different generation models, we use images generated by seven advanced generators along with their corresponding real images as part of the test set. These seven generation models include Stable Diffusion v1.4/1.5[32], VQDM[10], ADM[7], GLIDE[27], Wukong, and Midjourney.", + "bbox": [ + 511, + 777, + 906, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Real-vs-Fake classification baselines: We compare VIB-Net with existing state-of-the-art detection methods, in", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23832", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ec2b24faa11d5f7a5f1a5dcca0ab1010d0fb1a26e14f0229abea264ef53b80fa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Detection methodDiffusion ModelsGenerative Adversarial NetworksOthersAP
SDV1.4SDV1.5ADMGLIDEMidjourneyWukongVQDMPro-GANCycle-Big-GANStyle-GANStar-GANGau-GANDeep-fakeAverage
CNNSpot[39]99.9899.8351.1058.8067.9399.8049.9253.1550.2349.7955.9847.0756.0854.8654.0363.24
Fusing[14]99.9097.9869.3094.2081.2099.9084.6067.6387.7969.3767.9091.2043.0872.5689.4281.07
Lgrad[36]99.9499.9258.5284.0091.0699.7256.3483.5990.2447.5182.7499.1949.2566.4965.0978.24
Univfd[28]96.0496.2666.3493.7392.0890.9874.5351.7763.4275.8154.1254.9365.9970.2483.3475.31
NPR[37]100.0099.9794.7095.8095.50100.0086.3083.3094.9072.0082.7097.3066.0085.3095.9089.98
CLIPping[17]93.9793.1068.0087.4477.3486.5277.1788.5488.4485.3377.2389.8281.5661.1957.3780.87
Ours100.0099.9795.4997.1397.8199.9397.0096.5998.4497.1784.3197.6096.9481.3293.2795.53
", + "bbox": [ + 94, + 88, + 906, + 258 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2c33959f89ed5c07d0089de82cedaf079f6dd783e06c4f9afa7a02d5978a4abf.jpg", + "table_caption": [ + "Table 1. The $AP$ values of different methods based on diffusion sources. Models outside the Diffusion Models (DMs) column are considered as the generalization domain across generative model series." + ], + "table_footnote": [], + "table_body": "
Detection methodDiffusion ModelsGenerative Adversarial NetworksOthersACC
SDV1.4SDV1.5ADMGLIDEMidjourneyWukongVQDMPro-GANCycle-Big-GANStyle-GANStar-GANGau-GANDeep-SANAverage
CNNSpot[39]99.4899.3550.1050.9056.4297.9050.0450.2749.8150.1050.9849.7750.3851.9850.22
Fusing[14]99.9099.9151.3057.5052.3099.9064.2051.2052.4053.5050.2058.2049.3251.0264.84
Lgrad[36]99.1299.0553.0064.2476.3497.5350.9361.6160.7448.8261.4350.1749.7050.1756.49
Univfd[28]83.5584.8053.3575.3071.6073.5555.1058.6559.3061.4556.8061.4555.3058.4072.00
NPR[37]100.0099.9073.0089.7082.30100.0068.3060.3067.2059.2058.0073.2052.0074.8089.60
CLIPping[17]96.0795.4870.1485.0077.6688.8679.3588.7888.4889.5780.6990.8285.9466.9161.64
Ours99.5599.2073.8574.2588.0598.2589.3589.7088.6091.2074.1080.7087.1572.0081.50
", + "bbox": [ + 94, + 309, + 906, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Similar to Tab. 1, we compare these methods using classification accuracy, which is influenced by manually set thresholds. For a fair comparison, the classification threshold for all models is set to 0.5.", + "bbox": [ + 89, + 492, + 906, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cluding CNNDetection (CVPR 2020) [39], FreDect (ICML 2020) [8], Fusing (ICIP 2022) [14], LGrad (CVPR 2023) [36], DIRE (ICCV 2023) [40], Ojha (CVPR 2023) [28], NPR (CVPR 2024) [37], CLIPping(ICMR 2024) [17]. To ensure a fair comparison, we re-train and predict using the official code on ForenSynths and Genimage.", + "bbox": [ + 88, + 544, + 482, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics We follow existing research and consider two metrics: Average Precision (AP) and Accuracy (ACC). Average Precision (AP) is threshold-independent and evaluates the overall performance of the model. Accuracy is defined as the number of correct predictions divided by the number of test images, and this metric is influenced by the decision threshold. Given that there is no prior information in real-world scenarios, we standardize the decision threshold to 0.5.", + "bbox": [ + 88, + 654, + 482, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details Our method is implemented using the PyTorch framework on the NVIDIA Tesla V100. The model is trained with the Adam optimizer, a learning rate of $2 \\times 10^{-4}$ , and a batch size of 256. Following the same data augmentation strategy as in [39][28], which includes Gaussian blur and JPEG compression, experiments have shown that this is critical for the model's generalization ability.", + "bbox": [ + 88, + 794, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Performance evaluation based on Diffusion sources", + "text_level": 1, + "bbox": [ + 511, + 544, + 906, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It is well-known that images generated by the current state-of-the-art diffusion models are more realistic than those generated by other types of generative models. To validate the effectiveness of the proposed method, we designed a challenging experimental scenario. Specifically, we trained various detector using Stable Diffusion v1.4 and then tested them on images generated by 15 different models, which include the latest diffusion-based generative models as well as classic GANs. The results are presented in Tabs. 1 and 2. Bold indicates the highest value in the column, and underline indicates the second highest value in the column. This applies to the other tables as well.", + "bbox": [ + 509, + 582, + 906, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first evaluated the performance of each model on the latest generative models. As shown in Tabs. 1 and 2, advanced diffusion models significantly challenged most detection models, causing a sharp drop in performance. For example, Univfd and Lgrad experienced a $40\\%$ decrease in accuracy and precision when tested on ADM, Midjourney, and VQDM. This decline is likely due to the high-quality images produced by these models, which make it difficult for detection methods to extract distinguishing fea", + "bbox": [ + 509, + 763, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23833", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/96b08620198dd1fda7f1a026e64f21038d153e055719354ff20bd9af4f1aad29.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Detection methodGenerative Adversarial NetworksPerceptual lossOthersDiffusion ModelsAP
Pro-GANCycle-GANBig-GANStyle-GANStar-GANGau-GANCRNIMLEDeep-fakeSANSDV1.4SDV1.5ADMGLIDEMidjourneyWukongVQDMAverage
CNNSpot[39]99.9996.4087.5096.9494.2498.2896.6198.7664.4255.8952.8653.2565.1468.1056.6051.1569.4976.80
FreDect[8]99.9984.7793.6288.9799.4882.8582.2769.7770.7749.5038.5038.4163.7254.7347.2540.4486.0170.06
Fusing[14]100.0095.5090.7699.4899.8288.3290.3296.0271.1277.3365.3065.6274.8577.4469.9164.5375.4282.46
Lgrad[36]99.9094.0190.7599.8099.9879.2961.6467.0571.7145.0970.9071.7271.8375.9671.4266.5170.2376.93
Univfd[28]100.0099.2198.3197.9899.3599.8096.7299.0082.0482.1885.4882.3084.3484.0469.1090.1394.9690.88
NPR[37]100.0098.5087.8099.8099.9085.5050.0050.0082.4071.6084.0084.6074.6085.7085.4080.5081.2082.44
CLIPping[17]99.8594.0391.2795.2499.0589.5279.8780.3373.8956.6258.2857.8577.2577.9352.4360.6580.0077.89
Ours100.0099.8099.2998.7999.7299.9990.8996.8392.6491.6287.2486.9887.8888.5375.6890.9296.5193.14
", + "bbox": [ + 94, + 88, + 906, + 252 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2b78035a5b0191d0a345b4c6ace643fa0f58d9a5fd5c3da9ffc538edebe59679.jpg", + "table_caption": [ + "Table 3. The $AP$ values of different methods trained on GAN source images. Generative models other than those in the GANs column can be considered as the generalization domain." + ], + "table_footnote": [], + "table_body": "
Detection methodGenerative Adversarial NetworksPerceptual lossOthersDiffusion ModelsACC
Pro-GANCycle-GANBig-GANStyle-GANStar-GANGau-GANCRNIMLEDeep-fakeSANSDV1.4SDV1.5ADMGLIDEMidjourneyWukongVQDMAverage
CNNSpot[39]99.9987.5971.1889.9594.6081.4485.8785.8751.6950.0050.8250.8860.2057.8550.7751.1356.2069.18
FreDect[8]99.3678.7681.9778.0194.6280.5760.0460.1363.2950.0040.0240.3864.6655.4346.8941.5478.9565.57
Fusing[14]99.9087.0177.3285.2197.0476.9585.1788.5453.7654.5651.0451.3456.5057.1552.1251.6755.0969.43
DIRE[40]100.0076.6572.1285.03100.0067.79----49.7449.8375.7871.7558.0154.4653.68-
Lgrad[36]99.8086.9485.6391.0899.2772.4950.4950.5056.4244.4763.0363.6767.1066.1056.2063.6067.0269.64
Univfd[28]99.9098.5094.5084.4095.8599.5059.5072.0067.4056.5063.1063.5766.9061.7057.8571.0685.0076.31
NPR[37]99.8096.1084.4097.7099.3082.5050.0050.0080.2069.2076.6077.9069.7077.3077.8076.1078.1078.98
CLIPping[17]99.8896.7494.7794.8799.4794.1585.2387.7376.4860.2761.2660.7880.7780.6953.9563.9484.7680.93
Ours99.9999.0095.7591.2598.9599.7071.2586.7583.2070.5071.5570.0071.4569.4061.2575.9086.6582.50
", + "bbox": [ + 94, + 304, + 906, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Accuracy of different methods trained on GAN source images. The data from [40] is obtained from [37], and ”-” indicates that the method did not show data for the corresponding model.", + "bbox": [ + 89, + 491, + 906, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tures. We also observed that these models performed better on generative models with distributions similar to the training data (e.g., SDV1.5, Wukong). This suggests that they may have learned superficial biases from irrelevant features, which are superficially correlated with labels and do not generalize well to images from different distributions. In contrast, NPR, which detects authenticity by analyzing pixel relationships, partially mitigates these biases, resulting in less performance decline. When testing on GANs, we observed significant performance drops, particularly on GauGAN, where the accuracy and precision were around $50 - 60\\%$ . This further highlights how even subtle relevant features and surface biases can severely impact the generalization performance of these models.", + "bbox": [ + 88, + 539, + 482, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In contrast, Our method outperforms the baseline model, Univfd, with improvements of $20.22\\%$ in average precision and $20.46\\%$ in accuracy. This gap is even more pronounced across generative model families (GANs), where our method surpasses Univfd by $34.17\\%$ in precision and $26.41\\%$ in accuracy. Compared to state-of-the-art methods, our approach improves average precision by $5.55\\%$ and accuracy by $9.33\\%$ . For other generative models (GANs), VIB-Net achieves gains of $12.48\\%$ in average precision and", + "bbox": [ + 89, + 763, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "23.59% in accuracy. This is due to VIB-Net's ability to retain compact deep features while eliminating irrelevant surface features, enhancing generalization. Overall, these results highlight the advantages of our method.", + "bbox": [ + 511, + 539, + 906, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Performance evaluation based on GAN sources", + "text_level": 1, + "bbox": [ + 511, + 606, + 903, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "So far, we have explored the generalization performance of using images generated by specific diffusion models as training data. Next, we trained using images generated by ProGAN and then tested on 18 different generative models to explore the performance of the detector on unseen generative models. The results are shown in Tabs. 3 and 4. From the tables, we can see that almost all models perform well when detecting images generated by other GAN variants. This is because, compared to diffusion-based generation methods, images generated by these GAN variants are visually more distinguishable, making it easier for the model to learn the distinguishing features of real and fake. However, when detecting images generated by models other than GAN variants, the performance of most models sharply declines. This suggests that even when the training data contains rich features related to real/fake classification, it does not necessarily ensure better generalization performance. This is because the model may overfit the incidental correla", + "bbox": [ + 509, + 628, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23834", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/03c6eff0437e82330a33e46b3b5553ef28669794ee5a81cf45ccb08c1d454389.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 88, + 903, + 188 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/446f4f542fe26f5297875262a6a6efa70423d50c08a5ff47202b7b01c2126714.jpg", + "image_caption": [ + "Figure 5. Ablation study on the network architecture. The $AP$ and $ACC$ values of different VIB-Net variants on various models. CLIP:ViT-L/14 shows better performance." + ], + "image_footnote": [], + "bbox": [ + 91, + 195, + 903, + 296 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tions between irrelevant features and the task, which harms generalization performance. Univfd performs better across model families, but since it extracts general features, many of these features are unrelated to the real/fake classification task, which naturally affects the decision boundary.", + "bbox": [ + 89, + 354, + 482, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In contrast, our method shows better performance in detecting real and fake images. Compared to the baseline Univfd, our method achieved improvements of $2.26\\%$ and $6.19\\%$ in average precision and accuracy. In OOD scenarios (excluding GANs), our method outperforms the baseline Univfd by $3.22\\%$ and $8.48\\%$ . Compared with advanced methods, our method achieved improvements of $10.7\\%$ to $23.08\\%$ in average precision. These results further validate the effectiveness of our proposed method.", + "bbox": [ + 89, + 430, + 483, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Effect of VIB-Net across different backbone", + "text_level": 1, + "bbox": [ + 89, + 575, + 464, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "So far, we have seen the impressive performance of VIB-Net with the CLIP:ViT-L/14 backbone. In this section, we will explore the effect of transferring the VIB component to other multimodal models. We consider the following backbone variants to form VIB-Net: (1) FLAVA-FULL[35] (2) CLIP:ResNet50 (3) CLIP:ViT-B/16 (4) CLIP:ViT-L/14[30]. For each VIB-Net variant, we train using images generated by Stable Diffusion v1.4 and corresponding real images. Fig. 5 show the AP and ACC results of these variants on the same generative model. We observe that among these multimodal models, ViT-L/14 achieves the best performance, which can be attributed to its larger architecture and attention-based global receptive field. More importantly, all these variants exhibit strong performance, which can be attributed to the unique capabilities of our method.", + "bbox": [ + 89, + 597, + 482, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Analysis of latent features", + "text_level": 1, + "bbox": [ + 89, + 833, + 326, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, we apply our method to address the issues analyzed in Sec. 3.2. In Sec. 3.2, we identified that some irrelevant features in CLIP's general features could affect the classi-", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "fier's decision boundary, causing the model to learn superficial biases, which in turn impacts its generalization performance. As a solution, we proposed VIB-Net. Following the same experimental setup as in Sec. 3.2, we visualized the latent features extracted by VIB-Net using UMAP, and obtained encouraging results, as shown in Fig. 3b. We found that the latent features no longer exhibit obvious irrelevant features related to the real/fake image classification task, as the two separated clusters now belong to the same category (real or generated), unlike in Fig. 3a, where features of the same category were significantly separated. This demonstrates the advantage of our method.", + "bbox": [ + 511, + 354, + 906, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 546, + 633, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a universal real-fake image detection method, VIB-Net. First, we found that the general features extracted by current methods based on large-scale pre-trained models contain irrelevant features that are unrelated to the task of distinguishing real from fake images. Experiments have shown that these features not only affect classification decisions, but may also cause the model to learn superficial biases, thereby damaging its generalization performance. As an alternative, we use VIB to remove task-irrelevant features, aiming to obtain the most concise image representations. These simplified features are maximally correlated with the real-fake classification task.", + "bbox": [ + 511, + 571, + 908, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 511, + 763, + 681, + 781 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China under Grant 62376046 and Grant 62172067, and in part by the Natural Science Foundation of Chongqing under Grant CSTB2023NSCQ-MSX0341, the Science and Technology Research Program of Chongqing Municipal Education Commission (Grant No. KJQN202200635).", + "bbox": [ + 511, + 787, + 908, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23835", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck. arXiv preprint arXiv:1612.00410, 2016. 2, 3", + "[2] Alexander A Alemi, Ian Fischer, and Joshua V Dillon. Uncertainty in the variational information bottleneck. arXiv preprint arXiv:1807.00906, 2018. 3", + "[3] Xiuli Bi, Bo Liu, Fan Yang, Bin Xiao, Weisheng Li, Gao Huang, and Pamela C Cosman. Detecting generated images by real images only. arXiv preprint arXiv:2311.00962, 2023. 2", + "[4] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 103-120. Springer, 2020.", + "[5] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8789-8797, 2018. 5", + "[6] Davide Cozzolino, Giovanni Poggi, Riccardo Corvi, Matthias Nießner, and Luisa Verdoliva. Raising the bar of ai-generated image detection with clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4356-4366, 2024. 1, 2, 3", + "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 5", + "[8] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In International conference on machine learning, pages 3247-3258. PMLR, 2020. 6, 7", + "[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2014. 1", + "[10] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10696-10706, 2022. 5", + "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2, 3", + "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1", + "[13] Shizhe Hu, Zhengzheng Lou, Xiaoqiang Yan, and Yangdong Ye. A survey on information bottleneck. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Yan Ju, Shan Jia, Lipeng Ke, Hongfei Xue, Koki Nagano, and Siwei Lyu. Fusing global and local features for generalized ai-synthesized image detection. In 2022 IEEE International Conference on Image Processing (ICIP), pages 3465-3469. IEEE, 2022. 2, 6, 7", + "[15] Tero Karras. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 1, 5", + "[16] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 5", + "[17] Sohail Ahmed Khan and Duc-Tien Dang-Nguyen. Clipping the deception: Adapting vision-language models for universal deepfake detection. In Proceedings of the 2024 International Conference on Multimedia Retrieval, pages 1006-1015, 2024. 6, 7", + "[18] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 5", + "[19] Li Lin, Irene Amerini, Xin Wang, Shu Hu, et al. Robust clip-based detector for exposing diffusion model-generated images. arXiv preprint arXiv:2404.12908, 2024. 2", + "[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5", + "[21] Bo Liu, Fan Yang, Xiuli Bi, Bin Xiao, Weisheng Li, and Xinbo Gao. Detecting generated images by real images. In European Conference on Computer Vision, pages 95-110. Springer, 2022. 2", + "[22] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 5", + "[23] Yunpeng Luo, Junlong Du, Ke Yan, and Shouhong Ding. Lare^ 2: Latent reconstruction error based method for diffusion-generated image detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17006-17015, 2024. 1", + "[24] Rabeeh Karimi Mahabadi, Yonatan Belinkov, and James Henderson. Variational information bottleneck for effective low-resource fine-tuning. arXiv preprint arXiv:2106.05469, 2021.3", + "[25] Francesco Marra, Diego Gragnaniello, Luisa Verdoliva, and Giovanni Poggi. Do gans leave artificial fingerprints? In 2019 IEEE conference on multimedia information processing and retrieval (MIPR), pages 506-511. IEEE, 2019. 1", + "[26] Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. 3", + "[27] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 5" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23836", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24480-24489, 2023. 2, 3, 6, 7", + "[29] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 5", + "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 8", + "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1", + "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 5", + "[33] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics++: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1-11, 2019. 5", + "[34] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 5", + "[35] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15638-15650, 2022. 8", + "[36] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, and Yunchao Wei. Learning on gradients: Generalized artifacts representation for gan-generated images detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12105-12114, 2023. 6, 7", + "[37] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28130-28139, 2024. 6, 7", + "[38] Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. 3" + ], + "bbox": [ + 91, + 90, + 480, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8695-8704, 2020. 1, 2, 5, 6, 7", + "[40] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22445-22455, 2023. 1, 6, 7", + "[41] Among Wu and Cheng Deng. Tib: Detecting unknown objects via two-stream information bottleneck. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3", + "[42] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 5", + "[43] Ning Yu, Larry S Davis, and Mario Fritz. Attributing fake images to gans: Learning and analyzing gan fingerprints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7556-7566, 2019. 1", + "[44] Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deepfake detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2185–2194, 2021. 2", + "[45] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 5", + "[46] Mingjian Zhu, Hanting Chen, Qiangyu Yan, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang. Genimage: A million-scale benchmark for detecting ai-generated image. Advances in Neural Information Processing Systems, 36, 2024. 5" + ], + "bbox": [ + 516, + 90, + 906, + 599 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23837", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network/c23bce64-f04c-4e24-b700-3d6473e34043_model.json b/2025/Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network/c23bce64-f04c-4e24-b700-3d6473e34043_model.json new file mode 100644 index 0000000000000000000000000000000000000000..65d256a67042eeeaee17c44f9033851e6a279362 --- /dev/null +++ b/2025/Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network/c23bce64-f04c-4e24-b700-3d6473e34043_model.json @@ -0,0 +1,1859 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.13, + 0.892, + 0.175 + ], + "angle": 0, + "content": "Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network" + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.204, + 0.787, + 0.24 + ], + "angle": 0, + "content": "Haifeng Zhang, Qinghui He, Xiuli Bi, Weisheng Li, Bo Liu*, Bin Xiao Chongqing University of Posts and Telecommunications" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.242, + 0.748, + 0.257 + ], + "angle": 0, + "content": "s220331143@stu.cqupt.edu.cn, s220232014@stu.cqupt.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.259, + 0.678, + 0.275 + ], + "angle": 0, + "content": "{bix1, liws, boliu, xiaobin}@cqupt.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.341, + 0.486, + 0.675 + ], + "angle": 0, + "content": "The rapid advancement of generative models has significantly improved the quality of generated images. Meanwhile, it challenges information authenticity and credibility. Current generated image detection methods based on large-scale pre-trained multimodal models have achieved impressive results. Although these models provide abundant features, the authentication task-related features are often submerged. Consequently, those authentication task-irrelated features cause models to learn superficial biases, thereby harming their generalization performance across different model genera (e.g., GANs and Diffusion Models). To this end, we proposed VIB-Net, which uses Variational Information Bottlenecks to enforce authentication task-related feature learning. We tested and analyzed the proposed method and existing methods on samples generated by 17 different generative models. Compared to SOTA methods, VIB-Net achieved a \\(5.55\\%\\) improvement in mAP and a \\(9.33\\%\\) increase in accuracy. Notably, in generalization tests on unseen generative models from different series, VIB-Net improved mAP by \\(12.48\\%\\) and accuracy by \\(23.59\\%\\) over SOTA methods. The code is available at https://github.com/oceanzhf/VIBAIGCDetect." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.702, + 0.222, + 0.718 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.728, + 0.485, + 0.879 + ], + "angle": 0, + "content": "The continuous advancements in deep learning have led to the emergence of numerous well-known image generation models, such as GANs[9][15] and Diffusion[12][31]. Images generated by these models have reached a level where observers can hardly distinguish them from real images, and the flexibility of tools based on these models has also greatly improved [6]. With a few simple steps, people can easily obtain high-quality images as needed. While this presents a tremendous opportunity for fields like visual arts, it also serves as a paradise for individuals looking to fab" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.309, + 0.906, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.498, + 0.909, + 0.554 + ], + "angle": 0, + "content": "Figure 1. Each method's overall performance('P') on test data and generalization performance ('G') on unseen data are linked by dashed lines. Our detector demonstrates a clear advantage over all SOTA methods in both metrics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.907, + 0.657 + ], + "angle": 0, + "content": "ricate misinformation. Given the ease of creating manipulated images today and the potential risk of using them for malicious purposes, there is an urgent need to develop detectors that help distinguish between real and generated images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.909, + 0.885 + ], + "angle": 0, + "content": "Existing detection methods [4][39][25][43] focus on identifying specific traces left by generators in either the spatial or frequency domain of generated images, thus establishing a decision boundary between real and generated images. These detection methods are highly effective in detecting images generated by GANs and earlier diffusion models. However, their performance drops significantly when facing images generated by recent, popular models that visually resemble real images. Although some error reconstruction-based methods [40][23] have been developed specifically for detecting images generated by diffusion models, they similarly experience performance declines when applied to GAN-generated images. Consequently, current research increasingly focuses on improving the cross-model genera detection capabilities of detectors." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Generalization across different domains is a primary re" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.888, + 0.242, + 0.901 + ], + "angle": 0, + "content": "* Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23828" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.482, + 0.573 + ], + "angle": 0, + "content": "search challenge in deep learning. In this context, the emergence of large-scale pre-trained multimodal models has introduced new approaches and promising results. For instance, by using the image encoder from the CLIP [30] model to extract image features and training a classifier with linear probing, there is a notable improvement in generalization performance over models like ResNet50[28][6]. However, compared to traditional classification tasks, real and generated images share subtle, harder-to-capture distinguishing features. This implies that the features extracted by large-scale pre-trained models for the task of distinguishing real from fake images may contain a considerable amount of task-irrelevant features. Task-irrelevant features can either have no impact on the realizness of the image or, though effective in isolation, become redundant when combined with other features. These features not only interfere with the classifier's ability to accurately learn decision boundaries but also hinder the model's ability to capture and distinguish task-relevant characteristics. More critically, task-irrelevant features may lead the model to learn superficial biases. For example, if a fake image contains a tree while the real image does not, the model might erroneously associate the presence of the tree with being a fake image. Consequently, when detecting other images, the model may mistakenly classify a real image as fake if it contains a tree. Therefore, when dealing with high-quality generated images with subtle distinguishing features, it is challenging to fine-tune the model to separate essential from unnecessary information. This can lead to an overfitting of the incidental statistical correlations between unnecessary information and target images, ultimately undermining the model's generalization and discrimination capabilities." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.585, + 0.482, + 0.9 + ], + "angle": 0, + "content": "How can we harness the powerful feature extraction capability of large models while minimizing the interference of unnecessary information? By approaching this from a novel perspective, we propose addressing the issue using Information Bottleneck Theory. This theory aims to create latent encodings that not only effectively represent the target but also maximally compress the original input, thereby avoiding the impact of task-irrelevant features on the model. Specifically, we introduce a method called VIB-Net, which couples a large-scale pre-trained multimodal model with Variational Information Bottleneck (VIB) [1]. Considering that feature representations extracted by the model's image encoder are more task-specific, we decouple and filter these features using latent variables. This method maximizes the mutual information between the latent feature and the classification output, while minimizing the mutual information between the latent feature and the input, resulting in a concise feature that should only contain the effective information needed for the task. This concise latent feature helps the model avoid overfitting to irrelevant features and the incidental correlations with the task, preventing the model" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.093, + 0.904, + 0.121 + ], + "angle": 0, + "content": "from learning superficial biases, which is beneficial to the model's generalization performance." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.123, + 0.904, + 0.362 + ], + "angle": 0, + "content": "In summary, our contributions are as follows: (1) To the best of our knowledge, we are the first to apply Information Bottleneck theory and techniques to the detection of generated images, demonstrating the potential of this theory and approach in this field. (2) We analyze the limitations of methods based on large-scale pre-trained models in detecting images generated by unseen generative models, and present our theory of what could be wrong with the existing paradigm. (3) We propose VIB-Net for the detection of generated images, which uses VIB to decouple and filter the general features extracted by large-scale pre-trained multimodal models, removing task-irrelevant information to enhance detection performance. (4) A large number of experiments validate the advantages of our method (see Fig. 1), especially in the improvement of generalization performance \\((+23.59\\%\\) accuracy)." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.378, + 0.649, + 0.393 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.403, + 0.904, + 0.719 + ], + "angle": 0, + "content": "Generalization Performance. The continuous advancement of generated image models presents ongoing challenges for generated image detection, making the generalization performance of detection models on unseen generators an increasing focus for researchers[14][21]. [39] explored the effects of data preprocessing and augmentation methods on the generalization performance of a ResNet50[11]-based detector. [44] argues that the differences between real and fake images are subtle and localized, framing forgery detection as a fine-grained task that could positively impact generalization performance. [3] takes a frequency-domain approach, mapping real images to a dense subspace and synthetic images outside this subspace to enhance generalization. While these methods have achieved certain successes on different models within the same generation series, they show limitations across different series. This is due to the fact that, with the development of generative algorithms, subtle traces in generated images are becoming increasingly difficult to capture. Additionally, artifacts in the frequency domain differ between model series and are gradually decreasing in visibility." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.721, + 0.904, + 0.9 + ], + "angle": 0, + "content": "The impressive results of large-scale pre-trained models in terms of transferability and generalization have sparked significant interest among researchers in the field of generated image detection. [28] argues that existing detectors perform asymmetric learning, where they use CLIP's image encoder to extract image features, exploring the potential of large models in detecting different model families. [6] suggests that detectors based on large models rely on higher-level semantic features, which enhances the generalization ability of the detectors. [19] uses both image and text features extracted by the CLIP model, and then employs a multilayer perceptron to detect images generated by" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23829" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.288 + ], + "angle": 0, + "content": "diffusion models, improving generalization performance. Information Bottleneck. The Information Bottleneck (IB) [38] is an innovative information-theoretic method used for pattern analysis and representation learning. It provides an elegant balance between data compression and information retention[13]. VIB [1] is a variational approximation of IB, aimed at parameterizing IB using neural networks. [2] demonstrated VIB's capability in improving classification calibration and out-of-distribution data. [24] explored VIB as a regularization technique to address overfitting issues in NLP models. [41] proposed a dual-stream information bottleneck method to utilize in-distribution data to enhance out-of-distribution target detection tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.29, + 0.485, + 0.471 + ], + "angle": 0, + "content": "Based on the above analysis, current methods for detecting generated images using large-scale pre-trained models simply utilize the general representations extracted from images, without thoroughly exploring the relationship between these general features and downstream tasks. Additionally, Information Bottleneck is an effective technique for improving performance in cross-domain and cross-distribution tasks. Therefore, unlike existing methods, our approach leverages Information Bottleneck techniques to explore the potential of large-scale pre-trained models in generated image detection from a completely new perspective, achieving outstanding performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.486, + 0.231, + 0.501 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.512, + 0.291, + 0.527 + ], + "angle": 0, + "content": "3.1. Problem background" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.484, + 0.7 + ], + "angle": 0, + "content": "A common paradigm for using large-scale pretrained models is to leverage a large-scale pretrained multimodal model (e.g., CLIP) to learn the general representation of real and generated images, and then apply this feature representation to downstream tasks[28][6]. Building on their findings, we further consider the following points: (1) It is well-known that the features extracted by large-scale pre-trained visual encoders are general-purpose, meaning they contain many general features, such as semantic information. (2) Another widely accepted notion is that some critical features for distinguishing real from generated images are subtle and local." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.701, + 0.484, + 0.837 + ], + "angle": 0, + "content": "Further combining these two points, we propose the following hypotheses: (i) The general features extracted by large-scale pre-trained multimodal models are more beneficial for real-fake image classification tasks than those extracted by pre-trained visual models. (ii) These general features contain task-irrelevant features for real-fake classification. (iii) These task-irrelevant features affect the classification of real and fake images. Next, we will delve into these three hypotheses." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.449, + 0.865 + ], + "angle": 0, + "content": "3.2. Analysis of general feature representation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "For the first hypothesis, We use image pairs of real and generated images to aid our analysis[6], generating textual de" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.09, + 0.906, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.32, + 0.907, + 0.362 + ], + "angle": 0, + "content": "Figure 2. UMAP visualization of features extracted from Dalle2 and Dalle3 generated images using large-scale pretrained multimodal(CLIP) and unimodal models(ResNet)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.907, + 0.61 + ], + "angle": 0, + "content": "scriptions for 1,000 real images and using these descriptions to produce 1,000 corresponding images with DALL-E 2 and DALL-E 3. We then compare features extracted by large-scale pre-trained multimodal models(CLIP[30]) and pre-trained visual models(ResNet[11]) through UMAP [26] visualization of 3,000 images. The results show that features extracted by multimodal models (e.g., CLIP) exhibit greater distribution differences for real images and DALL-E 2-generated images, suggesting multimodal models may be better at distinguishing real from generated images. However, for DALL-E 3-generated images, both model types show similar feature distributions, possibly due to advancements in image generation tools, which complicate the classification task and require further fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To further explore hypotheses (ii) and (iii), we visualized the extracted features using both category and authenticity as labels to observe their distribution differences. Specifically, we prepared images of two categories (car and cat), each containing 500 real images and 500 generated images. We then used CLIP to extract their features and performed UMAP dimensionality reduction for visualization, with the results shown in Fig. 3a. We can see that the features extracted by CLIP clearly form two clusters: the left cluster mainly consists of light blue and light red points, while the right cluster is mainly composed of light pink and light purple points. These two clusters correspond to category 1 and category 2, respectively. This indicates that CLIP can significantly distinguish category information. However, this clear separation of category information may affect the decision-making process for classifying real vs. generated images. When examining the distinction between real and generated images, we notice that the right cluster shows a noticeable separation of real and generated features, with" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23830" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.09, + 0.285, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.216, + 0.285, + 0.24 + ], + "angle": 0, + "content": "(a) Visualization of features extracted by CLIP." + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.091, + 0.482, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.216, + 0.483, + 0.24 + ], + "angle": 0, + "content": "(b) Visualization of features extracted by VIB-Net." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.252, + 0.483, + 0.281 + ], + "angle": 0, + "content": "Figure 3. UMAP visualization of features extracted from real and generated images of two category using CLIP and VIB-Net." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.314, + 0.483, + 0.525 + ], + "angle": 0, + "content": "two different colors observed within the cluster. This suggests that CLIP is capable of distinguishing real and generated features within category 2. However, in the left cluster, CLIP does not show significant ability to differentiate between real and generated features. We observe that in the left cluster (category 1), the two colors are mixed together, showing poor distinguishability. From these observations, it is easy to infer that the internal separation of real or generated images caused by semantic category information makes it more difficult for the classifier to learn the decision boundary. More importantly, the strong sensitivity of large-scale pre-trained multimodal models like CLIP to semantic information leads the model to tend to learn superficial biases." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.529, + 0.483, + 0.68 + ], + "angle": 0, + "content": "Based on this viewpoint, we consider whether it is possible to minimize the irrelevant features within these multimodal features while retaining the ones that are relevant to the classification task. This would help the model avoid learning superficial biases, allow the classifier to more easily learn decision boundaries, and thereby enhance the model's discrimination ability and generalization capability. This is the motivation of this paper. As shown in Fig. 3b, the method(VIB-Net) proposed in this paper achieves significant results." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.706, + 0.2, + 0.723 + ], + "angle": 0, + "content": "4. Approach" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Based on the issues analyzed in Sec. 3, an effective approach is to filter the multimodal features in order to obtain the representations most beneficial for the real-fake classification task, which serves as the starting point of this paper. Overall, our method is illustrated in Fig. 4. VIB attempts to filter out irrelevant features and to find the most concise representation, making it easier for the classifier to learn the features needed for distinguishing between real and fake tasks, thereby significantly improving the accuracy of the classification task. Below, we will provide a detailed description of our method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.81, + 0.106 + ], + "angle": 0, + "content": "4.1. Information bottleneck constraint" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.907, + 0.264 + ], + "angle": 0, + "content": "To specifically optimize the filtering of irrelevant information in the input features while retaining task-relevant information, we apply the information bottleneck (IB) principle as a constraint on the features to be optimized. Let \\( X \\), \\( Z \\), and \\( Y \\) represent the input, the latent compressed representation, and the output target, respectively. The goal of IB is to find the maximum compressed representation \\( Z \\) of the input \\( X \\), while preserving as much information about the output \\( Y \\) as possible. The maximization objective function is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.273, + 0.907, + 0.29 + ], + "angle": 0, + "content": "\\[\nL _ {I B} = I (Z, Y) - \\beta I (Z, X) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.297, + 0.907, + 0.373 + ], + "angle": 0, + "content": "Where \\( I(.,.) \\) denotes mutual information and \\( \\beta \\) is a trade-off parameter. The first term constrains the latent compressed representation to retain sufficient information for downstream tasks, while the second term ensures that irrelevant and redundant information is filtered out." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.381, + 0.816, + 0.395 + ], + "angle": 0, + "content": "4.2. Variational information bottleneck" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.906, + 0.464 + ], + "angle": 0, + "content": "To parameterize the information bottleneck (IB) using deep neural networks, we perform variational inference on Equation (1). First, based on the definition of mutual information, Equation (1) can be rewritten as:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.483, + 0.907, + 0.551 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {I B} = \\int d y d z p (y, z) \\log {\\frac {p (y | z)}{p (y)}} \\\\ - \\beta \\int d x d z p (x, z) \\log \\frac {p (z \\mid x)}{p (z)} \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.558, + 0.906, + 0.587 + ], + "angle": 0, + "content": "Where \\( p(y|z) \\) is defined by the encoder and the Markov chain as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.593, + 0.906, + 0.626 + ], + "angle": 0, + "content": "\\[\np (y \\mid z) = \\int d x \\frac {p (y \\mid x) p (z \\mid x) p (x)}{p (z)} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.633, + 0.907, + 0.769 + ], + "angle": 0, + "content": "Since it is not computable here, we consider setting \\( q(y|z) \\) as a parameterized approximation of \\( p(y|z) \\), serving as a \"decoder,\" which can be approximated by a linear layer. We introduce the Kullback-Leibler (KL) divergence to obtain an estimizable representation. As the KL divergence satisfies \\( \\mathrm{KL}(p(y|z),q(y|z))\\geq 0 \\), and by applying the Markov assumption, \\( p(y,z) \\) can be rewritten as \\( p(y,z) = \\int dx p(x)p(y|x)p(z|x) \\), thereby providing a lower bound for the first term of the objective function:" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.776, + 0.905, + 0.819 + ], + "angle": 0, + "content": "\\[\nI (Z, Y) \\geq \\int d x d y d z p (x) p (y | x) p (z | x) \\log q (y | z) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.82, + 0.905, + 0.865 + ], + "angle": 0, + "content": "At the same time, let \\( r(z) \\) be an estimate of the prior probability \\( p(z) \\), used to approximate the marginal distribution, thereby obtaining an upper bound for the second term:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.872, + 0.905, + 0.906 + ], + "angle": 0, + "content": "\\[\nI (Z, X) \\leq \\int d x d z p (x) p (z | x) \\log \\frac {p (z | x)}{r (z)} \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23831" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.089, + 0.905, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.311, + 0.908, + 0.355 + ], + "angle": 0, + "content": "Figure 4. Overview of VIB-Net. We utilize Variational Information Bottleneck (VIB) techniques to map the image features extracted by CLIP's visual encoder into latent features \\( z \\), where the information in \\( z \\) is filtered based on the Information Bottleneck (IB). Subsequently, the latent features \\( z \\) are used as input to an MLP to learn the decision boundary." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.38, + 0.484, + 0.518 + ], + "angle": 0, + "content": "The encoder we use is defined as \\( p(z|x) = p_{\\theta}(z|x) = \\mathcal{N}\\left(z|f^{\\mu}(x), f^{\\Sigma}(x)\\right) \\), where \\( f(.) \\) represents a shallow MLP, \\( \\mu \\) denotes the K-dimensional mean vector of the output \\( z \\), and \\( \\Sigma \\) represents the \\( K \\times K \\) covariance matrix. We select a parameterized Gaussian distribution for both \\( p_{\\theta}(z|x) \\) and \\( r(z) \\) to allow for the analytical computation of KL divergence. Additionally, we employ the reparameterization trick to estimate gradients [18]. Combining all the components above yields the following objective function:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.538, + 0.483, + 0.6 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {V I B} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} z \\sim p _ {\\theta} (z | x) [ - \\log q (y _ {\\mathrm {n}} | z _ {n}) ] \\\\ + \\beta \\mathrm {K L} [ p _ {\\theta} (z | x _ {n}), r (z) ] \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.608, + 0.373, + 0.623 + ], + "angle": 0, + "content": "4.3. Workflow of the overall method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.484, + 0.812 + ], + "angle": 0, + "content": "As shown in Fig. 4, the parts in VIB-Net work in the following way: First, the real and generated images \\(\\{x_{1}, x_{2}, x_{3}, \\dots, x_{n}\\}\\) are input into the CLIP image encoder, yielding the corresponding image features \\(\\{t_{1}, t_{2}, t_{3}, \\dots, t_{n}\\}\\). Next, these feature vectors are passed into the Variational Information Bottleneck (VIB). Specifically, the image features are processed through a shallow MLP, then a linear layer \\(f(.)\\) calculates \\(\\mu\\) and \\(\\Sigma\\), which are transformed using a softmax activation. This layer has \\(K\\) dimensions, representing the bottleneck size. At this point, we obtain the latent compressed feature \\(z\\), which is then fed into a linear layer \\(q(y|z)\\) to output the target task label \\(y\\)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.824, + 0.224, + 0.841 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.191, + 0.866 + ], + "angle": 0, + "content": "5.1. Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To better evaluate the model's performance, we utilize two mainstream public datasets: ForenSynths[39] and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.907, + 0.53 + ], + "angle": 0, + "content": "Genimage[46]. The ForenSynths dataset primarily consists of images generated by GANs, while the Genimage dataset mainly includes images generated by diffusion models. Consequently, we set up two training and testing schemes: (1) Training on ForenSynths and then testing on both ForenSynths and Genimage, and (2) similarly, training on Genimage and testing on both Genimage and ForenSynths. This setup allows us to more effectively validate the generalization performance of the detector on cross model genera scenarios. The details are as follows:" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.533, + 0.609, + 0.547 + ], + "angle": 0, + "content": "Training set:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.906, + 0.669 + ], + "angle": 0, + "content": "The ForenSynths training set consists of 20 different categories, each containing 18,000 real images from the LSUN dataset [42] and an equal number of images generated using ProGAN[15]. Following the setup in Genimage, we use images generated by Stable Diffusion v1.4 [32] in the Genimage dataset, along with their corresponding real images, as the training set. In total, these include 32,000 real and generated images." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.671, + 0.598, + 0.685 + ], + "angle": 0, + "content": "Testing set:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.687, + 0.906, + 0.778 + ], + "angle": 0, + "content": "11 models from ForenSynths The test set includes images generated by 11 different models, primarily from the GAN family[15][45][29][5][16]. The real images are sourced from six datasets, including ImageNet[34], LSUN[42], FaceForensics++[33], CelebA-HQ[15], CelebA[22], and COCO[20]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.779, + 0.907, + 0.87 + ], + "angle": 0, + "content": "7 models from Genimage To validate the method across different generation models, we use images generated by seven advanced generators along with their corresponding real images as part of the test set. These seven generation models include Stable Diffusion v1.4/1.5[32], VQDM[10], ADM[7], GLIDE[27], Wukong, and Midjourney." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Real-vs-Fake classification baselines: We compare VIB-Net with existing state-of-the-art detection methods, in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23832" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.907, + 0.259 + ], + "angle": 0, + "content": "
Detection methodDiffusion ModelsGenerative Adversarial NetworksOthersAP
SDV1.4SDV1.5ADMGLIDEMidjourneyWukongVQDMPro-GANCycle-Big-GANStyle-GANStar-GANGau-GANDeep-fakeAverage
CNNSpot[39]99.9899.8351.1058.8067.9399.8049.9253.1550.2349.7955.9847.0756.0854.8654.0363.24
Fusing[14]99.9097.9869.3094.2081.2099.9084.6067.6387.7969.3767.9091.2043.0872.5689.4281.07
Lgrad[36]99.9499.9258.5284.0091.0699.7256.3483.5990.2447.5182.7499.1949.2566.4965.0978.24
Univfd[28]96.0496.2666.3493.7392.0890.9874.5351.7763.4275.8154.1254.9365.9970.2483.3475.31
NPR[37]100.0099.9794.7095.8095.50100.0086.3083.3094.9072.0082.7097.3066.0085.3095.9089.98
CLIPping[17]93.9793.1068.0087.4477.3486.5277.1788.5488.4485.3377.2389.8281.5661.1957.3780.87
Ours100.0099.9795.4997.1397.8199.9397.0096.5998.4497.1784.3197.6096.9481.3293.2795.53
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.268, + 0.907, + 0.298 + ], + "angle": 0, + "content": "Table 1. The \\(AP\\) values of different methods based on diffusion sources. Models outside the Diffusion Models (DMs) column are considered as the generalization domain across generative model series." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.31, + 0.907, + 0.482 + ], + "angle": 0, + "content": "
Detection methodDiffusion ModelsGenerative Adversarial NetworksOthersACC
SDV1.4SDV1.5ADMGLIDEMidjourneyWukongVQDMPro-GANCycle-Big-GANStyle-GANStar-GANGau-GANDeep-SANAverage
CNNSpot[39]99.4899.3550.1050.9056.4297.9050.0450.2749.8150.1050.9849.7750.3851.9850.22
Fusing[14]99.9099.9151.3057.5052.3099.9064.2051.2052.4053.5050.2058.2049.3251.0264.84
Lgrad[36]99.1299.0553.0064.2476.3497.5350.9361.6160.7448.8261.4350.1749.7050.1756.49
Univfd[28]83.5584.8053.3575.3071.6073.5555.1058.6559.3061.4556.8061.4555.3058.4072.00
NPR[37]100.0099.9073.0089.7082.30100.0068.3060.3067.2059.2058.0073.2052.0074.8089.60
CLIPping[17]96.0795.4870.1485.0077.6688.8679.3588.7888.4889.5780.6990.8285.9466.9161.64
Ours99.5599.2073.8574.2588.0598.2589.3589.7088.6091.2074.1080.7087.1572.0081.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.493, + 0.907, + 0.522 + ], + "angle": 0, + "content": "Table 2. Similar to Tab. 1, we compare these methods using classification accuracy, which is influenced by manually set thresholds. For a fair comparison, the classification threshold for all models is set to 0.5." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.545, + 0.483, + 0.652 + ], + "angle": 0, + "content": "cluding CNNDetection (CVPR 2020) [39], FreDect (ICML 2020) [8], Fusing (ICIP 2022) [14], LGrad (CVPR 2023) [36], DIRE (ICCV 2023) [40], Ojha (CVPR 2023) [28], NPR (CVPR 2024) [37], CLIPping(ICMR 2024) [17]. To ensure a fair comparison, we re-train and predict using the official code on ForenSynths and Genimage." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.655, + 0.483, + 0.79 + ], + "angle": 0, + "content": "Evaluation metrics We follow existing research and consider two metrics: Average Precision (AP) and Accuracy (ACC). Average Precision (AP) is threshold-independent and evaluates the overall performance of the model. Accuracy is defined as the number of correct predictions divided by the number of test images, and this metric is influenced by the decision threshold. Given that there is no prior information in real-world scenarios, we standardize the decision threshold to 0.5." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.795, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Implementation Details Our method is implemented using the PyTorch framework on the NVIDIA Tesla V100. The model is trained with the Adam optimizer, a learning rate of \\(2 \\times 10^{-4}\\), and a batch size of 256. Following the same data augmentation strategy as in [39][28], which includes Gaussian blur and JPEG compression, experiments have shown that this is critical for the model's generalization ability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.545, + 0.907, + 0.575 + ], + "angle": 0, + "content": "5.2. Performance evaluation based on Diffusion sources" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.583, + 0.907, + 0.764 + ], + "angle": 0, + "content": "It is well-known that images generated by the current state-of-the-art diffusion models are more realistic than those generated by other types of generative models. To validate the effectiveness of the proposed method, we designed a challenging experimental scenario. Specifically, we trained various detector using Stable Diffusion v1.4 and then tested them on images generated by 15 different models, which include the latest diffusion-based generative models as well as classic GANs. The results are presented in Tabs. 1 and 2. Bold indicates the highest value in the column, and underline indicates the second highest value in the column. This applies to the other tables as well." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We first evaluated the performance of each model on the latest generative models. As shown in Tabs. 1 and 2, advanced diffusion models significantly challenged most detection models, causing a sharp drop in performance. For example, Univfd and Lgrad experienced a \\(40\\%\\) decrease in accuracy and precision when tested on ADM, Midjourney, and VQDM. This decline is likely due to the high-quality images produced by these models, which make it difficult for detection methods to extract distinguishing fea" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23833" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.908, + 0.253 + ], + "angle": 0, + "content": "
Detection methodGenerative Adversarial NetworksPerceptual lossOthersDiffusion ModelsAP
Pro-GANCycle-GANBig-GANStyle-GANStar-GANGau-GANCRNIMLEDeep-fakeSANSDV1.4SDV1.5ADMGLIDEMidjourneyWukongVQDMAverage
CNNSpot[39]99.9996.4087.5096.9494.2498.2896.6198.7664.4255.8952.8653.2565.1468.1056.6051.1569.4976.80
FreDect[8]99.9984.7793.6288.9799.4882.8582.2769.7770.7749.5038.5038.4163.7254.7347.2540.4486.0170.06
Fusing[14]100.0095.5090.7699.4899.8288.3290.3296.0271.1277.3365.3065.6274.8577.4469.9164.5375.4282.46
Lgrad[36]99.9094.0190.7599.8099.9879.2961.6467.0571.7145.0970.9071.7271.8375.9671.4266.5170.2376.93
Univfd[28]100.0099.2198.3197.9899.3599.8096.7299.0082.0482.1885.4882.3084.3484.0469.1090.1394.9690.88
NPR[37]100.0098.5087.8099.8099.9085.5050.0050.0082.4071.6084.0084.6074.6085.7085.4080.5081.2082.44
CLIPping[17]99.8594.0391.2795.2499.0589.5279.8780.3373.8956.6258.2857.8577.2577.9352.4360.6580.0077.89
Ours100.0099.8099.2998.7999.7299.9990.8996.8392.6491.6287.2486.9887.8888.5375.6890.9296.5193.14
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.263, + 0.907, + 0.293 + ], + "angle": 0, + "content": "Table 3. The \\(AP\\) values of different methods trained on GAN source images. Generative models other than those in the GANs column can be considered as the generalization domain." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.305, + 0.907, + 0.481 + ], + "angle": 0, + "content": "
Detection methodGenerative Adversarial NetworksPerceptual lossOthersDiffusion ModelsACC
Pro-GANCycle-GANBig-GANStyle-GANStar-GANGau-GANCRNIMLEDeep-fakeSANSDV1.4SDV1.5ADMGLIDEMidjourneyWukongVQDMAverage
CNNSpot[39]99.9987.5971.1889.9594.6081.4485.8785.8751.6950.0050.8250.8860.2057.8550.7751.1356.2069.18
FreDect[8]99.3678.7681.9778.0194.6280.5760.0460.1363.2950.0040.0240.3864.6655.4346.8941.5478.9565.57
Fusing[14]99.9087.0177.3285.2197.0476.9585.1788.5453.7654.5651.0451.3456.5057.1552.1251.6755.0969.43
DIRE[40]100.0076.6572.1285.03100.0067.79----49.7449.8375.7871.7558.0154.4653.68-
Lgrad[36]99.8086.9485.6391.0899.2772.4950.4950.5056.4244.4763.0363.6767.1066.1056.2063.6067.0269.64
Univfd[28]99.9098.5094.5084.4095.8599.5059.5072.0067.4056.5063.1063.5766.9061.7057.8571.0685.0076.31
NPR[37]99.8096.1084.4097.7099.3082.5050.0050.0080.2069.2076.6077.9069.7077.3077.8076.1078.1078.98
CLIPping[17]99.8896.7494.7794.8799.4794.1585.2387.7376.4860.2761.2660.7880.7780.6953.9563.9484.7680.93
Ours99.9999.0095.7591.2598.9599.7071.2586.7583.2070.5071.5570.0071.4569.4061.2575.9086.6582.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.492, + 0.907, + 0.521 + ], + "angle": 0, + "content": "Table 4. Accuracy of different methods trained on GAN source images. The data from [40] is obtained from [37], and ”-” indicates that the method did not show data for the corresponding model." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.54, + 0.483, + 0.751 + ], + "angle": 0, + "content": "tures. We also observed that these models performed better on generative models with distributions similar to the training data (e.g., SDV1.5, Wukong). This suggests that they may have learned superficial biases from irrelevant features, which are superficially correlated with labels and do not generalize well to images from different distributions. In contrast, NPR, which detects authenticity by analyzing pixel relationships, partially mitigates these biases, resulting in less performance decline. When testing on GANs, we observed significant performance drops, particularly on GauGAN, where the accuracy and precision were around \\(50 - 60\\%\\). This further highlights how even subtle relevant features and surface biases can severely impact the generalization performance of these models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.903 + ], + "angle": 0, + "content": "In contrast, Our method outperforms the baseline model, Univfd, with improvements of \\(20.22\\%\\) in average precision and \\(20.46\\%\\) in accuracy. This gap is even more pronounced across generative model families (GANs), where our method surpasses Univfd by \\(34.17\\%\\) in precision and \\(26.41\\%\\) in accuracy. Compared to state-of-the-art methods, our approach improves average precision by \\(5.55\\%\\) and accuracy by \\(9.33\\%\\). For other generative models (GANs), VIB-Net achieves gains of \\(12.48\\%\\) in average precision and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.54, + 0.907, + 0.601 + ], + "angle": 0, + "content": "23.59% in accuracy. This is due to VIB-Net's ability to retain compact deep features while eliminating irrelevant surface features, enhancing generalization. Overall, these results highlight the advantages of our method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.905, + 0.623 + ], + "angle": 0, + "content": "5.3. Performance evaluation based on GAN sources" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.629, + 0.907, + 0.901 + ], + "angle": 0, + "content": "So far, we have explored the generalization performance of using images generated by specific diffusion models as training data. Next, we trained using images generated by ProGAN and then tested on 18 different generative models to explore the performance of the detector on unseen generative models. The results are shown in Tabs. 3 and 4. From the tables, we can see that almost all models perform well when detecting images generated by other GAN variants. This is because, compared to diffusion-based generation methods, images generated by these GAN variants are visually more distinguishable, making it easier for the model to learn the distinguishing features of real and fake. However, when detecting images generated by models other than GAN variants, the performance of most models sharply declines. This suggests that even when the training data contains rich features related to real/fake classification, it does not necessarily ensure better generalization performance. This is because the model may overfit the incidental correla" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23834" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.905, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.196, + 0.905, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.308, + 0.908, + 0.337 + ], + "angle": 0, + "content": "Figure 5. Ablation study on the network architecture. The \\(AP\\) and \\(ACC\\) values of different VIB-Net variants on various models. CLIP:ViT-L/14 shows better performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.483, + 0.431 + ], + "angle": 0, + "content": "tions between irrelevant features and the task, which harms generalization performance. Univfd performs better across model families, but since it extracts general features, many of these features are unrelated to the real/fake classification task, which naturally affects the decision boundary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.431, + 0.484, + 0.567 + ], + "angle": 0, + "content": "In contrast, our method shows better performance in detecting real and fake images. Compared to the baseline Univfd, our method achieved improvements of \\(2.26\\%\\) and \\(6.19\\%\\) in average precision and accuracy. In OOD scenarios (excluding GANs), our method outperforms the baseline Univfd by \\(3.22\\%\\) and \\(8.48\\%\\). Compared with advanced methods, our method achieved improvements of \\(10.7\\%\\) to \\(23.08\\%\\) in average precision. These results further validate the effectiveness of our proposed method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.576, + 0.465, + 0.591 + ], + "angle": 0, + "content": "5.4. Effect of VIB-Net across different backbone" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.825 + ], + "angle": 0, + "content": "So far, we have seen the impressive performance of VIB-Net with the CLIP:ViT-L/14 backbone. In this section, we will explore the effect of transferring the VIB component to other multimodal models. We consider the following backbone variants to form VIB-Net: (1) FLAVA-FULL[35] (2) CLIP:ResNet50 (3) CLIP:ViT-B/16 (4) CLIP:ViT-L/14[30]. For each VIB-Net variant, we train using images generated by Stable Diffusion v1.4 and corresponding real images. Fig. 5 show the AP and ACC results of these variants on the same generative model. We observe that among these multimodal models, ViT-L/14 achieves the best performance, which can be attributed to its larger architecture and attention-based global receptive field. More importantly, all these variants exhibit strong performance, which can be attributed to the unique capabilities of our method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.327, + 0.85 + ], + "angle": 0, + "content": "5.5. Analysis of latent features" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Finally, we apply our method to address the issues analyzed in Sec. 3.2. In Sec. 3.2, we identified that some irrelevant features in CLIP's general features could affect the classi-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.907, + 0.537 + ], + "angle": 0, + "content": "fier's decision boundary, causing the model to learn superficial biases, which in turn impacts its generalization performance. As a solution, we proposed VIB-Net. Following the same experimental setup as in Sec. 3.2, we visualized the latent features extracted by VIB-Net using UMAP, and obtained encouraging results, as shown in Fig. 3b. We found that the latent features no longer exhibit obvious irrelevant features related to the real/fake image classification task, as the two separated clusters now belong to the same category (real or generated), unlike in Fig. 3a, where features of the same category were significantly separated. This demonstrates the advantage of our method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.547, + 0.634, + 0.563 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.573, + 0.909, + 0.753 + ], + "angle": 0, + "content": "We propose a universal real-fake image detection method, VIB-Net. First, we found that the general features extracted by current methods based on large-scale pre-trained models contain irrelevant features that are unrelated to the task of distinguishing real from fake images. Experiments have shown that these features not only affect classification decisions, but may also cause the model to learn superficial biases, thereby damaging its generalization performance. As an alternative, we use VIB to remove task-irrelevant features, aiming to obtain the most concise image representations. These simplified features are maximally correlated with the real-fake classification task." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.765, + 0.682, + 0.782 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.789, + 0.909, + 0.888 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62376046 and Grant 62172067, and in part by the Natural Science Foundation of Chongqing under Grant CSTB2023NSCQ-MSX0341, the Science and Technology Research Program of Chongqing Municipal Education Commission (Grant No. KJQN202200635)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23835" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.157 + ], + "angle": 0, + "content": "[1] Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck. arXiv preprint arXiv:1612.00410, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.484, + 0.2 + ], + "angle": 0, + "content": "[2] Alexander A Alemi, Ian Fischer, and Joshua V Dillon. Uncertainty in the variational information bottleneck. arXiv preprint arXiv:1807.00906, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.203, + 0.483, + 0.257 + ], + "angle": 0, + "content": "[3] Xiuli Bi, Bo Liu, Fan Yang, Bin Xiao, Weisheng Li, Gao Huang, and Pamela C Cosman. Detecting generated images by real images only. arXiv preprint arXiv:2311.00962, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.26, + 0.483, + 0.342 + ], + "angle": 0, + "content": "[4] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 103-120. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.345, + 0.483, + 0.426 + ], + "angle": 0, + "content": "[5] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8789-8797, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.431, + 0.484, + 0.499 + ], + "angle": 0, + "content": "[6] Davide Cozzolino, Giovanni Poggi, Riccardo Corvi, Matthias Nießner, and Luisa Verdoliva. Raising the bar of ai-generated image detection with clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4356-4366, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.502, + 0.483, + 0.543 + ], + "angle": 0, + "content": "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.545, + 0.483, + 0.613 + ], + "angle": 0, + "content": "[8] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In International conference on machine learning, pages 3247-3258. PMLR, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.483, + 0.684 + ], + "angle": 0, + "content": "[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.688, + 0.483, + 0.757 + ], + "angle": 0, + "content": "[10] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10696-10706, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.759, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Shizhe Hu, Zhengzheng Lou, Xiaoqiang Yan, and Yangdong Ye. A survey on information bottleneck. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[14] Yan Ju, Shan Jia, Lipeng Ke, Hongfei Xue, Koki Nagano, and Siwei Lyu. Fusing global and local features for generalized ai-synthesized image detection. In 2022 IEEE International Conference on Image Processing (ICIP), pages 3465-3469. IEEE, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[15] Tero Karras. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[16] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[17] Sohail Ahmed Khan and Duc-Tien Dang-Nguyen. Clipping the deception: Adapting vision-language models for universal deepfake detection. In Proceedings of the 2024 International Conference on Multimedia Retrieval, pages 1006-1015, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.362 + ], + "angle": 0, + "content": "[18] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[19] Li Lin, Irene Amerini, Xin Wang, Shu Hu, et al. Robust clip-based detector for exposing diffusion model-generated images. arXiv preprint arXiv:2404.12908, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.407, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.492, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[21] Bo Liu, Fan Yang, Xiuli Bi, Bin Xiao, Weisheng Li, and Xinbo Gao. Detecting generated images by real images. In European Conference on Computer Vision, pages 95-110. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.548, + 0.907, + 0.602 + ], + "angle": 0, + "content": "[22] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.605, + 0.907, + 0.673 + ], + "angle": 0, + "content": "[23] Yunpeng Luo, Junlong Du, Ke Yan, and Shouhong Ding. Lare^ 2: Latent reconstruction error based method for diffusion-generated image detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17006-17015, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[24] Rabeeh Karimi Mahabadi, Yonatan Belinkov, and James Henderson. Variational information bottleneck for effective low-resource fine-tuning. arXiv preprint arXiv:2106.05469, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[25] Francesco Marra, Diego Gragnaniello, Luisa Verdoliva, and Giovanni Poggi. Do gans leave artificial fingerprints? In 2019 IEEE conference on multimedia information processing and retrieval (MIPR), pages 506-511. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[26] Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[27] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23836" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[28] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24480-24489, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[29] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[33] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics++: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1-11, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[34] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[35] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15638-15650, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.772 + ], + "angle": 0, + "content": "[36] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, and Yunchao Wei. Learning on gradients: Generalized artifacts representation for gan-generated images detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12105-12114, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[37] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28130-28139, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[38] Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.163 + ], + "angle": 0, + "content": "[39] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8695-8704, 2020. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.906, + 0.233 + ], + "angle": 0, + "content": "[40] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22445-22455, 2023. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.906, + 0.275 + ], + "angle": 0, + "content": "[41] Among Wu and Cheng Deng. Tib: Detecting unknown objects via two-stream information bottleneck. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.906, + 0.332 + ], + "angle": 0, + "content": "[42] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.906, + 0.389 + ], + "angle": 0, + "content": "[43] Ning Yu, Larry S Davis, and Mario Fritz. Attributing fake images to gans: Learning and analyzing gan fingerprints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7556-7566, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.906, + 0.459 + ], + "angle": 0, + "content": "[44] Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deepfake detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2185–2194, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.46, + 0.906, + 0.529 + ], + "angle": 0, + "content": "[45] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.531, + 0.906, + 0.6 + ], + "angle": 0, + "content": "[46] Mingjian Zhu, Hanting Chen, Qiangyu Yan, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang. Genimage: A million-scale benchmark for detecting ai-generated image. Advances in Neural Information Processing Systems, 36, 2024. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23837" + } + ] +] \ No newline at end of file diff --git a/2025/Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network/c23bce64-f04c-4e24-b700-3d6473e34043_origin.pdf b/2025/Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network/c23bce64-f04c-4e24-b700-3d6473e34043_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f01efa13a6c9e6b65a461c7c59b2b19a19bcb373 --- /dev/null +++ b/2025/Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network/c23bce64-f04c-4e24-b700-3d6473e34043_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b9b4059a94666b77757e3602a8008a712e66b7716ebb6bc8373dac6cc32eb7ab +size 3264067 diff --git a/2025/Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network/full.md b/2025/Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9ae2e58b3320e4e18ec4a8ef38d0d9d6671a7f20 --- /dev/null +++ b/2025/Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network/full.md @@ -0,0 +1,262 @@ +# Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network + +Haifeng Zhang, Qinghui He, Xiuli Bi, Weisheng Li, Bo Liu*, Bin Xiao Chongqing University of Posts and Telecommunications + +s220331143@stu.cqupt.edu.cn, s220232014@stu.cqupt.edu.cn + +{bix1, liws, boliu, xiaobin}@cqupt.edu.cn + +# Abstract + +The rapid advancement of generative models has significantly improved the quality of generated images. Meanwhile, it challenges information authenticity and credibility. Current generated image detection methods based on large-scale pre-trained multimodal models have achieved impressive results. Although these models provide abundant features, the authentication task-related features are often submerged. Consequently, those authentication task-irrelated features cause models to learn superficial biases, thereby harming their generalization performance across different model genera (e.g., GANs and Diffusion Models). To this end, we proposed VIB-Net, which uses Variational Information Bottlenecks to enforce authentication task-related feature learning. We tested and analyzed the proposed method and existing methods on samples generated by 17 different generative models. Compared to SOTA methods, VIB-Net achieved a $5.55\%$ improvement in mAP and a $9.33\%$ increase in accuracy. Notably, in generalization tests on unseen generative models from different series, VIB-Net improved mAP by $12.48\%$ and accuracy by $23.59\%$ over SOTA methods. The code is available at https://github.com/oceanzhf/VIBAIGCDetect. + +# 1. Introduction + +The continuous advancements in deep learning have led to the emergence of numerous well-known image generation models, such as GANs[9][15] and Diffusion[12][31]. Images generated by these models have reached a level where observers can hardly distinguish them from real images, and the flexibility of tools based on these models has also greatly improved [6]. With a few simple steps, people can easily obtain high-quality images as needed. While this presents a tremendous opportunity for fields like visual arts, it also serves as a paradise for individuals looking to fab + +![](images/432848555de54ec22b2ec2160073a539e39993a3a52b1f214925a0ea15c4b742.jpg) +Figure 1. Each method's overall performance('P') on test data and generalization performance ('G') on unseen data are linked by dashed lines. Our detector demonstrates a clear advantage over all SOTA methods in both metrics. + +ricate misinformation. Given the ease of creating manipulated images today and the potential risk of using them for malicious purposes, there is an urgent need to develop detectors that help distinguish between real and generated images. + +Existing detection methods [4][39][25][43] focus on identifying specific traces left by generators in either the spatial or frequency domain of generated images, thus establishing a decision boundary between real and generated images. These detection methods are highly effective in detecting images generated by GANs and earlier diffusion models. However, their performance drops significantly when facing images generated by recent, popular models that visually resemble real images. Although some error reconstruction-based methods [40][23] have been developed specifically for detecting images generated by diffusion models, they similarly experience performance declines when applied to GAN-generated images. Consequently, current research increasingly focuses on improving the cross-model genera detection capabilities of detectors. + +Generalization across different domains is a primary re + +search challenge in deep learning. In this context, the emergence of large-scale pre-trained multimodal models has introduced new approaches and promising results. For instance, by using the image encoder from the CLIP [30] model to extract image features and training a classifier with linear probing, there is a notable improvement in generalization performance over models like ResNet50[28][6]. However, compared to traditional classification tasks, real and generated images share subtle, harder-to-capture distinguishing features. This implies that the features extracted by large-scale pre-trained models for the task of distinguishing real from fake images may contain a considerable amount of task-irrelevant features. Task-irrelevant features can either have no impact on the realizness of the image or, though effective in isolation, become redundant when combined with other features. These features not only interfere with the classifier's ability to accurately learn decision boundaries but also hinder the model's ability to capture and distinguish task-relevant characteristics. More critically, task-irrelevant features may lead the model to learn superficial biases. For example, if a fake image contains a tree while the real image does not, the model might erroneously associate the presence of the tree with being a fake image. Consequently, when detecting other images, the model may mistakenly classify a real image as fake if it contains a tree. Therefore, when dealing with high-quality generated images with subtle distinguishing features, it is challenging to fine-tune the model to separate essential from unnecessary information. This can lead to an overfitting of the incidental statistical correlations between unnecessary information and target images, ultimately undermining the model's generalization and discrimination capabilities. + +How can we harness the powerful feature extraction capability of large models while minimizing the interference of unnecessary information? By approaching this from a novel perspective, we propose addressing the issue using Information Bottleneck Theory. This theory aims to create latent encodings that not only effectively represent the target but also maximally compress the original input, thereby avoiding the impact of task-irrelevant features on the model. Specifically, we introduce a method called VIB-Net, which couples a large-scale pre-trained multimodal model with Variational Information Bottleneck (VIB) [1]. Considering that feature representations extracted by the model's image encoder are more task-specific, we decouple and filter these features using latent variables. This method maximizes the mutual information between the latent feature and the classification output, while minimizing the mutual information between the latent feature and the input, resulting in a concise feature that should only contain the effective information needed for the task. This concise latent feature helps the model avoid overfitting to irrelevant features and the incidental correlations with the task, preventing the model + +from learning superficial biases, which is beneficial to the model's generalization performance. + +In summary, our contributions are as follows: (1) To the best of our knowledge, we are the first to apply Information Bottleneck theory and techniques to the detection of generated images, demonstrating the potential of this theory and approach in this field. (2) We analyze the limitations of methods based on large-scale pre-trained models in detecting images generated by unseen generative models, and present our theory of what could be wrong with the existing paradigm. (3) We propose VIB-Net for the detection of generated images, which uses VIB to decouple and filter the general features extracted by large-scale pre-trained multimodal models, removing task-irrelevant information to enhance detection performance. (4) A large number of experiments validate the advantages of our method (see Fig. 1), especially in the improvement of generalization performance $(+23.59\%$ accuracy). + +# 2. Related work + +Generalization Performance. The continuous advancement of generated image models presents ongoing challenges for generated image detection, making the generalization performance of detection models on unseen generators an increasing focus for researchers[14][21]. [39] explored the effects of data preprocessing and augmentation methods on the generalization performance of a ResNet50[11]-based detector. [44] argues that the differences between real and fake images are subtle and localized, framing forgery detection as a fine-grained task that could positively impact generalization performance. [3] takes a frequency-domain approach, mapping real images to a dense subspace and synthetic images outside this subspace to enhance generalization. While these methods have achieved certain successes on different models within the same generation series, they show limitations across different series. This is due to the fact that, with the development of generative algorithms, subtle traces in generated images are becoming increasingly difficult to capture. Additionally, artifacts in the frequency domain differ between model series and are gradually decreasing in visibility. + +The impressive results of large-scale pre-trained models in terms of transferability and generalization have sparked significant interest among researchers in the field of generated image detection. [28] argues that existing detectors perform asymmetric learning, where they use CLIP's image encoder to extract image features, exploring the potential of large models in detecting different model families. [6] suggests that detectors based on large models rely on higher-level semantic features, which enhances the generalization ability of the detectors. [19] uses both image and text features extracted by the CLIP model, and then employs a multilayer perceptron to detect images generated by + +diffusion models, improving generalization performance. Information Bottleneck. The Information Bottleneck (IB) [38] is an innovative information-theoretic method used for pattern analysis and representation learning. It provides an elegant balance between data compression and information retention[13]. VIB [1] is a variational approximation of IB, aimed at parameterizing IB using neural networks. [2] demonstrated VIB's capability in improving classification calibration and out-of-distribution data. [24] explored VIB as a regularization technique to address overfitting issues in NLP models. [41] proposed a dual-stream information bottleneck method to utilize in-distribution data to enhance out-of-distribution target detection tasks. + +Based on the above analysis, current methods for detecting generated images using large-scale pre-trained models simply utilize the general representations extracted from images, without thoroughly exploring the relationship between these general features and downstream tasks. Additionally, Information Bottleneck is an effective technique for improving performance in cross-domain and cross-distribution tasks. Therefore, unlike existing methods, our approach leverages Information Bottleneck techniques to explore the potential of large-scale pre-trained models in generated image detection from a completely new perspective, achieving outstanding performance. + +# 3. Preliminaries + +# 3.1. Problem background + +A common paradigm for using large-scale pretrained models is to leverage a large-scale pretrained multimodal model (e.g., CLIP) to learn the general representation of real and generated images, and then apply this feature representation to downstream tasks[28][6]. Building on their findings, we further consider the following points: (1) It is well-known that the features extracted by large-scale pre-trained visual encoders are general-purpose, meaning they contain many general features, such as semantic information. (2) Another widely accepted notion is that some critical features for distinguishing real from generated images are subtle and local. + +Further combining these two points, we propose the following hypotheses: (i) The general features extracted by large-scale pre-trained multimodal models are more beneficial for real-fake image classification tasks than those extracted by pre-trained visual models. (ii) These general features contain task-irrelevant features for real-fake classification. (iii) These task-irrelevant features affect the classification of real and fake images. Next, we will delve into these three hypotheses. + +# 3.2. Analysis of general feature representation + +For the first hypothesis, We use image pairs of real and generated images to aid our analysis[6], generating textual de + +![](images/e1410341985ff647e745d20a946c655965b19f0cbbf303dfb911f10c27cde35e.jpg) +Figure 2. UMAP visualization of features extracted from Dalle2 and Dalle3 generated images using large-scale pretrained multimodal(CLIP) and unimodal models(ResNet). + +scriptions for 1,000 real images and using these descriptions to produce 1,000 corresponding images with DALL-E 2 and DALL-E 3. We then compare features extracted by large-scale pre-trained multimodal models(CLIP[30]) and pre-trained visual models(ResNet[11]) through UMAP [26] visualization of 3,000 images. The results show that features extracted by multimodal models (e.g., CLIP) exhibit greater distribution differences for real images and DALL-E 2-generated images, suggesting multimodal models may be better at distinguishing real from generated images. However, for DALL-E 3-generated images, both model types show similar feature distributions, possibly due to advancements in image generation tools, which complicate the classification task and require further fine-tuning. + +To further explore hypotheses (ii) and (iii), we visualized the extracted features using both category and authenticity as labels to observe their distribution differences. Specifically, we prepared images of two categories (car and cat), each containing 500 real images and 500 generated images. We then used CLIP to extract their features and performed UMAP dimensionality reduction for visualization, with the results shown in Fig. 3a. We can see that the features extracted by CLIP clearly form two clusters: the left cluster mainly consists of light blue and light red points, while the right cluster is mainly composed of light pink and light purple points. These two clusters correspond to category 1 and category 2, respectively. This indicates that CLIP can significantly distinguish category information. However, this clear separation of category information may affect the decision-making process for classifying real vs. generated images. When examining the distinction between real and generated images, we notice that the right cluster shows a noticeable separation of real and generated features, with + +![](images/fe8f885df46bbd6b886ae2693b1dc8ce8b2ffa83bcc61143cf31da8d0c17917d.jpg) +(a) Visualization of features extracted by CLIP. +Figure 3. UMAP visualization of features extracted from real and generated images of two category using CLIP and VIB-Net. + +![](images/1ab822a33b6db285e7ab7f352b8aedae2588e0ed8278793c8c97bcbb3a4451d0.jpg) +(b) Visualization of features extracted by VIB-Net. + +two different colors observed within the cluster. This suggests that CLIP is capable of distinguishing real and generated features within category 2. However, in the left cluster, CLIP does not show significant ability to differentiate between real and generated features. We observe that in the left cluster (category 1), the two colors are mixed together, showing poor distinguishability. From these observations, it is easy to infer that the internal separation of real or generated images caused by semantic category information makes it more difficult for the classifier to learn the decision boundary. More importantly, the strong sensitivity of large-scale pre-trained multimodal models like CLIP to semantic information leads the model to tend to learn superficial biases. + +Based on this viewpoint, we consider whether it is possible to minimize the irrelevant features within these multimodal features while retaining the ones that are relevant to the classification task. This would help the model avoid learning superficial biases, allow the classifier to more easily learn decision boundaries, and thereby enhance the model's discrimination ability and generalization capability. This is the motivation of this paper. As shown in Fig. 3b, the method(VIB-Net) proposed in this paper achieves significant results. + +# 4. Approach + +Based on the issues analyzed in Sec. 3, an effective approach is to filter the multimodal features in order to obtain the representations most beneficial for the real-fake classification task, which serves as the starting point of this paper. Overall, our method is illustrated in Fig. 4. VIB attempts to filter out irrelevant features and to find the most concise representation, making it easier for the classifier to learn the features needed for distinguishing between real and fake tasks, thereby significantly improving the accuracy of the classification task. Below, we will provide a detailed description of our method. + +# 4.1. Information bottleneck constraint + +To specifically optimize the filtering of irrelevant information in the input features while retaining task-relevant information, we apply the information bottleneck (IB) principle as a constraint on the features to be optimized. Let $X$ , $Z$ , and $Y$ represent the input, the latent compressed representation, and the output target, respectively. The goal of IB is to find the maximum compressed representation $Z$ of the input $X$ , while preserving as much information about the output $Y$ as possible. The maximization objective function is defined as follows: + +$$ +L _ {I B} = I (Z, Y) - \beta I (Z, X) \tag {1} +$$ + +Where $I(.,.)$ denotes mutual information and $\beta$ is a trade-off parameter. The first term constrains the latent compressed representation to retain sufficient information for downstream tasks, while the second term ensures that irrelevant and redundant information is filtered out. + +# 4.2. Variational information bottleneck + +To parameterize the information bottleneck (IB) using deep neural networks, we perform variational inference on Equation (1). First, based on the definition of mutual information, Equation (1) can be rewritten as: + +$$ +\begin{array}{l} L _ {I B} = \int d y d z p (y, z) \log {\frac {p (y | z)}{p (y)}} \\ - \beta \int d x d z p (x, z) \log \frac {p (z \mid x)}{p (z)} \tag {2} \\ \end{array} +$$ + +Where $p(y|z)$ is defined by the encoder and the Markov chain as follows: + +$$ +p (y \mid z) = \int d x \frac {p (y \mid x) p (z \mid x) p (x)}{p (z)} \tag {3} +$$ + +Since it is not computable here, we consider setting $q(y|z)$ as a parameterized approximation of $p(y|z)$ , serving as a "decoder," which can be approximated by a linear layer. We introduce the Kullback-Leibler (KL) divergence to obtain an estimizable representation. As the KL divergence satisfies $\mathrm{KL}(p(y|z),q(y|z))\geq 0$ , and by applying the Markov assumption, $p(y,z)$ can be rewritten as $p(y,z) = \int dx p(x)p(y|x)p(z|x)$ , thereby providing a lower bound for the first term of the objective function: + +$$ +I (Z, Y) \geq \int d x d y d z p (x) p (y | x) p (z | x) \log q (y | z) \tag {4} +$$ + +At the same time, let $r(z)$ be an estimate of the prior probability $p(z)$ , used to approximate the marginal distribution, thereby obtaining an upper bound for the second term: + +$$ +I (Z, X) \leq \int d x d z p (x) p (z | x) \log \frac {p (z | x)}{r (z)} \tag {5} +$$ + +![](images/f2654d25fbfe695a4e4ecf0ae41380f6fa32b64f9234168d03dc86580330d688.jpg) +Figure 4. Overview of VIB-Net. We utilize Variational Information Bottleneck (VIB) techniques to map the image features extracted by CLIP's visual encoder into latent features $z$ , where the information in $z$ is filtered based on the Information Bottleneck (IB). Subsequently, the latent features $z$ are used as input to an MLP to learn the decision boundary. + +The encoder we use is defined as $p(z|x) = p_{\theta}(z|x) = \mathcal{N}\left(z|f^{\mu}(x), f^{\Sigma}(x)\right)$ , where $f(.)$ represents a shallow MLP, $\mu$ denotes the K-dimensional mean vector of the output $z$ , and $\Sigma$ represents the $K \times K$ covariance matrix. We select a parameterized Gaussian distribution for both $p_{\theta}(z|x)$ and $r(z)$ to allow for the analytical computation of KL divergence. Additionally, we employ the reparameterization trick to estimate gradients [18]. Combining all the components above yields the following objective function: + +$$ +\begin{array}{l} L _ {V I B} = \frac {1}{N} \sum_ {n = 1} ^ {N} z \sim p _ {\theta} (z | x) [ - \log q (y _ {\mathrm {n}} | z _ {n}) ] \\ + \beta \mathrm {K L} [ p _ {\theta} (z | x _ {n}), r (z) ] \tag {6} \\ \end{array} +$$ + +# 4.3. Workflow of the overall method + +As shown in Fig. 4, the parts in VIB-Net work in the following way: First, the real and generated images $\{x_{1}, x_{2}, x_{3}, \dots, x_{n}\}$ are input into the CLIP image encoder, yielding the corresponding image features $\{t_{1}, t_{2}, t_{3}, \dots, t_{n}\}$ . Next, these feature vectors are passed into the Variational Information Bottleneck (VIB). Specifically, the image features are processed through a shallow MLP, then a linear layer $f(.)$ calculates $\mu$ and $\Sigma$ , which are transformed using a softmax activation. This layer has $K$ dimensions, representing the bottleneck size. At this point, we obtain the latent compressed feature $z$ , which is then fed into a linear layer $q(y|z)$ to output the target task label $y$ . + +# 5. Experiments + +# 5.1. Settings + +To better evaluate the model's performance, we utilize two mainstream public datasets: ForenSynths[39] and + +Genimage[46]. The ForenSynths dataset primarily consists of images generated by GANs, while the Genimage dataset mainly includes images generated by diffusion models. Consequently, we set up two training and testing schemes: (1) Training on ForenSynths and then testing on both ForenSynths and Genimage, and (2) similarly, training on Genimage and testing on both Genimage and ForenSynths. This setup allows us to more effectively validate the generalization performance of the detector on cross model genera scenarios. The details are as follows: + +# Training set: + +The ForenSynths training set consists of 20 different categories, each containing 18,000 real images from the LSUN dataset [42] and an equal number of images generated using ProGAN[15]. Following the setup in Genimage, we use images generated by Stable Diffusion v1.4 [32] in the Genimage dataset, along with their corresponding real images, as the training set. In total, these include 32,000 real and generated images. + +# Testing set: + +11 models from ForenSynths The test set includes images generated by 11 different models, primarily from the GAN family[15][45][29][5][16]. The real images are sourced from six datasets, including ImageNet[34], LSUN[42], FaceForensics++[33], CelebA-HQ[15], CelebA[22], and COCO[20]. + +7 models from Genimage To validate the method across different generation models, we use images generated by seven advanced generators along with their corresponding real images as part of the test set. These seven generation models include Stable Diffusion v1.4/1.5[32], VQDM[10], ADM[7], GLIDE[27], Wukong, and Midjourney. + +Real-vs-Fake classification baselines: We compare VIB-Net with existing state-of-the-art detection methods, in + +
Detection methodDiffusion ModelsGenerative Adversarial NetworksOthersAP
SDV1.4SDV1.5ADMGLIDEMidjourneyWukongVQDMPro-GANCycle-Big-GANStyle-GANStar-GANGau-GANDeep-fakeAverage
CNNSpot[39]99.9899.8351.1058.8067.9399.8049.9253.1550.2349.7955.9847.0756.0854.8654.0363.24
Fusing[14]99.9097.9869.3094.2081.2099.9084.6067.6387.7969.3767.9091.2043.0872.5689.4281.07
Lgrad[36]99.9499.9258.5284.0091.0699.7256.3483.5990.2447.5182.7499.1949.2566.4965.0978.24
Univfd[28]96.0496.2666.3493.7392.0890.9874.5351.7763.4275.8154.1254.9365.9970.2483.3475.31
NPR[37]100.0099.9794.7095.8095.50100.0086.3083.3094.9072.0082.7097.3066.0085.3095.9089.98
CLIPping[17]93.9793.1068.0087.4477.3486.5277.1788.5488.4485.3377.2389.8281.5661.1957.3780.87
Ours100.0099.9795.4997.1397.8199.9397.0096.5998.4497.1784.3197.6096.9481.3293.2795.53
+ +Table 1. The $AP$ values of different methods based on diffusion sources. Models outside the Diffusion Models (DMs) column are considered as the generalization domain across generative model series. + +
Detection methodDiffusion ModelsGenerative Adversarial NetworksOthersACC
SDV1.4SDV1.5ADMGLIDEMidjourneyWukongVQDMPro-GANCycle-Big-GANStyle-GANStar-GANGau-GANDeep-SANAverage
CNNSpot[39]99.4899.3550.1050.9056.4297.9050.0450.2749.8150.1050.9849.7750.3851.9850.22
Fusing[14]99.9099.9151.3057.5052.3099.9064.2051.2052.4053.5050.2058.2049.3251.0264.84
Lgrad[36]99.1299.0553.0064.2476.3497.5350.9361.6160.7448.8261.4350.1749.7050.1756.49
Univfd[28]83.5584.8053.3575.3071.6073.5555.1058.6559.3061.4556.8061.4555.3058.4072.00
NPR[37]100.0099.9073.0089.7082.30100.0068.3060.3067.2059.2058.0073.2052.0074.8089.60
CLIPping[17]96.0795.4870.1485.0077.6688.8679.3588.7888.4889.5780.6990.8285.9466.9161.64
Ours99.5599.2073.8574.2588.0598.2589.3589.7088.6091.2074.1080.7087.1572.0081.50
+ +Table 2. Similar to Tab. 1, we compare these methods using classification accuracy, which is influenced by manually set thresholds. For a fair comparison, the classification threshold for all models is set to 0.5. + +cluding CNNDetection (CVPR 2020) [39], FreDect (ICML 2020) [8], Fusing (ICIP 2022) [14], LGrad (CVPR 2023) [36], DIRE (ICCV 2023) [40], Ojha (CVPR 2023) [28], NPR (CVPR 2024) [37], CLIPping(ICMR 2024) [17]. To ensure a fair comparison, we re-train and predict using the official code on ForenSynths and Genimage. + +Evaluation metrics We follow existing research and consider two metrics: Average Precision (AP) and Accuracy (ACC). Average Precision (AP) is threshold-independent and evaluates the overall performance of the model. Accuracy is defined as the number of correct predictions divided by the number of test images, and this metric is influenced by the decision threshold. Given that there is no prior information in real-world scenarios, we standardize the decision threshold to 0.5. + +Implementation Details Our method is implemented using the PyTorch framework on the NVIDIA Tesla V100. The model is trained with the Adam optimizer, a learning rate of $2 \times 10^{-4}$ , and a batch size of 256. Following the same data augmentation strategy as in [39][28], which includes Gaussian blur and JPEG compression, experiments have shown that this is critical for the model's generalization ability. + +# 5.2. Performance evaluation based on Diffusion sources + +It is well-known that images generated by the current state-of-the-art diffusion models are more realistic than those generated by other types of generative models. To validate the effectiveness of the proposed method, we designed a challenging experimental scenario. Specifically, we trained various detector using Stable Diffusion v1.4 and then tested them on images generated by 15 different models, which include the latest diffusion-based generative models as well as classic GANs. The results are presented in Tabs. 1 and 2. Bold indicates the highest value in the column, and underline indicates the second highest value in the column. This applies to the other tables as well. + +We first evaluated the performance of each model on the latest generative models. As shown in Tabs. 1 and 2, advanced diffusion models significantly challenged most detection models, causing a sharp drop in performance. For example, Univfd and Lgrad experienced a $40\%$ decrease in accuracy and precision when tested on ADM, Midjourney, and VQDM. This decline is likely due to the high-quality images produced by these models, which make it difficult for detection methods to extract distinguishing fea + +
Detection methodGenerative Adversarial NetworksPerceptual lossOthersDiffusion ModelsAP
Pro-GANCycle-GANBig-GANStyle-GANStar-GANGau-GANCRNIMLEDeep-fakeSANSDV1.4SDV1.5ADMGLIDEMidjourneyWukongVQDMAverage
CNNSpot[39]99.9996.4087.5096.9494.2498.2896.6198.7664.4255.8952.8653.2565.1468.1056.6051.1569.4976.80
FreDect[8]99.9984.7793.6288.9799.4882.8582.2769.7770.7749.5038.5038.4163.7254.7347.2540.4486.0170.06
Fusing[14]100.0095.5090.7699.4899.8288.3290.3296.0271.1277.3365.3065.6274.8577.4469.9164.5375.4282.46
Lgrad[36]99.9094.0190.7599.8099.9879.2961.6467.0571.7145.0970.9071.7271.8375.9671.4266.5170.2376.93
Univfd[28]100.0099.2198.3197.9899.3599.8096.7299.0082.0482.1885.4882.3084.3484.0469.1090.1394.9690.88
NPR[37]100.0098.5087.8099.8099.9085.5050.0050.0082.4071.6084.0084.6074.6085.7085.4080.5081.2082.44
CLIPping[17]99.8594.0391.2795.2499.0589.5279.8780.3373.8956.6258.2857.8577.2577.9352.4360.6580.0077.89
Ours100.0099.8099.2998.7999.7299.9990.8996.8392.6491.6287.2486.9887.8888.5375.6890.9296.5193.14
+ +Table 3. The $AP$ values of different methods trained on GAN source images. Generative models other than those in the GANs column can be considered as the generalization domain. + +
Detection methodGenerative Adversarial NetworksPerceptual lossOthersDiffusion ModelsACC
Pro-GANCycle-GANBig-GANStyle-GANStar-GANGau-GANCRNIMLEDeep-fakeSANSDV1.4SDV1.5ADMGLIDEMidjourneyWukongVQDMAverage
CNNSpot[39]99.9987.5971.1889.9594.6081.4485.8785.8751.6950.0050.8250.8860.2057.8550.7751.1356.2069.18
FreDect[8]99.3678.7681.9778.0194.6280.5760.0460.1363.2950.0040.0240.3864.6655.4346.8941.5478.9565.57
Fusing[14]99.9087.0177.3285.2197.0476.9585.1788.5453.7654.5651.0451.3456.5057.1552.1251.6755.0969.43
DIRE[40]100.0076.6572.1285.03100.0067.79----49.7449.8375.7871.7558.0154.4653.68-
Lgrad[36]99.8086.9485.6391.0899.2772.4950.4950.5056.4244.4763.0363.6767.1066.1056.2063.6067.0269.64
Univfd[28]99.9098.5094.5084.4095.8599.5059.5072.0067.4056.5063.1063.5766.9061.7057.8571.0685.0076.31
NPR[37]99.8096.1084.4097.7099.3082.5050.0050.0080.2069.2076.6077.9069.7077.3077.8076.1078.1078.98
CLIPping[17]99.8896.7494.7794.8799.4794.1585.2387.7376.4860.2761.2660.7880.7780.6953.9563.9484.7680.93
Ours99.9999.0095.7591.2598.9599.7071.2586.7583.2070.5071.5570.0071.4569.4061.2575.9086.6582.50
+ +Table 4. Accuracy of different methods trained on GAN source images. The data from [40] is obtained from [37], and ”-” indicates that the method did not show data for the corresponding model. + +tures. We also observed that these models performed better on generative models with distributions similar to the training data (e.g., SDV1.5, Wukong). This suggests that they may have learned superficial biases from irrelevant features, which are superficially correlated with labels and do not generalize well to images from different distributions. In contrast, NPR, which detects authenticity by analyzing pixel relationships, partially mitigates these biases, resulting in less performance decline. When testing on GANs, we observed significant performance drops, particularly on GauGAN, where the accuracy and precision were around $50 - 60\%$ . This further highlights how even subtle relevant features and surface biases can severely impact the generalization performance of these models. + +In contrast, Our method outperforms the baseline model, Univfd, with improvements of $20.22\%$ in average precision and $20.46\%$ in accuracy. This gap is even more pronounced across generative model families (GANs), where our method surpasses Univfd by $34.17\%$ in precision and $26.41\%$ in accuracy. Compared to state-of-the-art methods, our approach improves average precision by $5.55\%$ and accuracy by $9.33\%$ . For other generative models (GANs), VIB-Net achieves gains of $12.48\%$ in average precision and + +23.59% in accuracy. This is due to VIB-Net's ability to retain compact deep features while eliminating irrelevant surface features, enhancing generalization. Overall, these results highlight the advantages of our method. + +# 5.3. Performance evaluation based on GAN sources + +So far, we have explored the generalization performance of using images generated by specific diffusion models as training data. Next, we trained using images generated by ProGAN and then tested on 18 different generative models to explore the performance of the detector on unseen generative models. The results are shown in Tabs. 3 and 4. From the tables, we can see that almost all models perform well when detecting images generated by other GAN variants. This is because, compared to diffusion-based generation methods, images generated by these GAN variants are visually more distinguishable, making it easier for the model to learn the distinguishing features of real and fake. However, when detecting images generated by models other than GAN variants, the performance of most models sharply declines. This suggests that even when the training data contains rich features related to real/fake classification, it does not necessarily ensure better generalization performance. This is because the model may overfit the incidental correla + +![](images/03c6eff0437e82330a33e46b3b5553ef28669794ee5a81cf45ccb08c1d454389.jpg) + +![](images/446f4f542fe26f5297875262a6a6efa70423d50c08a5ff47202b7b01c2126714.jpg) +Figure 5. Ablation study on the network architecture. The $AP$ and $ACC$ values of different VIB-Net variants on various models. CLIP:ViT-L/14 shows better performance. + +tions between irrelevant features and the task, which harms generalization performance. Univfd performs better across model families, but since it extracts general features, many of these features are unrelated to the real/fake classification task, which naturally affects the decision boundary. + +In contrast, our method shows better performance in detecting real and fake images. Compared to the baseline Univfd, our method achieved improvements of $2.26\%$ and $6.19\%$ in average precision and accuracy. In OOD scenarios (excluding GANs), our method outperforms the baseline Univfd by $3.22\%$ and $8.48\%$ . Compared with advanced methods, our method achieved improvements of $10.7\%$ to $23.08\%$ in average precision. These results further validate the effectiveness of our proposed method. + +# 5.4. Effect of VIB-Net across different backbone + +So far, we have seen the impressive performance of VIB-Net with the CLIP:ViT-L/14 backbone. In this section, we will explore the effect of transferring the VIB component to other multimodal models. We consider the following backbone variants to form VIB-Net: (1) FLAVA-FULL[35] (2) CLIP:ResNet50 (3) CLIP:ViT-B/16 (4) CLIP:ViT-L/14[30]. For each VIB-Net variant, we train using images generated by Stable Diffusion v1.4 and corresponding real images. Fig. 5 show the AP and ACC results of these variants on the same generative model. We observe that among these multimodal models, ViT-L/14 achieves the best performance, which can be attributed to its larger architecture and attention-based global receptive field. More importantly, all these variants exhibit strong performance, which can be attributed to the unique capabilities of our method. + +# 5.5. Analysis of latent features + +Finally, we apply our method to address the issues analyzed in Sec. 3.2. In Sec. 3.2, we identified that some irrelevant features in CLIP's general features could affect the classi- + +fier's decision boundary, causing the model to learn superficial biases, which in turn impacts its generalization performance. As a solution, we proposed VIB-Net. Following the same experimental setup as in Sec. 3.2, we visualized the latent features extracted by VIB-Net using UMAP, and obtained encouraging results, as shown in Fig. 3b. We found that the latent features no longer exhibit obvious irrelevant features related to the real/fake image classification task, as the two separated clusters now belong to the same category (real or generated), unlike in Fig. 3a, where features of the same category were significantly separated. This demonstrates the advantage of our method. + +# 6. Conclusion + +We propose a universal real-fake image detection method, VIB-Net. First, we found that the general features extracted by current methods based on large-scale pre-trained models contain irrelevant features that are unrelated to the task of distinguishing real from fake images. Experiments have shown that these features not only affect classification decisions, but may also cause the model to learn superficial biases, thereby damaging its generalization performance. As an alternative, we use VIB to remove task-irrelevant features, aiming to obtain the most concise image representations. These simplified features are maximally correlated with the real-fake classification task. + +# Acknowledgements + +This work was supported in part by the National Natural Science Foundation of China under Grant 62376046 and Grant 62172067, and in part by the Natural Science Foundation of Chongqing under Grant CSTB2023NSCQ-MSX0341, the Science and Technology Research Program of Chongqing Municipal Education Commission (Grant No. KJQN202200635). + +# References + +[1] Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck. arXiv preprint arXiv:1612.00410, 2016. 2, 3 +[2] Alexander A Alemi, Ian Fischer, and Joshua V Dillon. Uncertainty in the variational information bottleneck. arXiv preprint arXiv:1807.00906, 2018. 3 +[3] Xiuli Bi, Bo Liu, Fan Yang, Bin Xiao, Weisheng Li, Gao Huang, and Pamela C Cosman. Detecting generated images by real images only. arXiv preprint arXiv:2311.00962, 2023. 2 +[4] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 103-120. Springer, 2020. +[5] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8789-8797, 2018. 5 +[6] Davide Cozzolino, Giovanni Poggi, Riccardo Corvi, Matthias Nießner, and Luisa Verdoliva. Raising the bar of ai-generated image detection with clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4356-4366, 2024. 1, 2, 3 +[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 5 +[8] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In International conference on machine learning, pages 3247-3258. PMLR, 2020. 6, 7 +[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2014. 1 +[10] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10696-10706, 2022. 5 +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2, 3 +[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1 +[13] Shizhe Hu, Zhengzheng Lou, Xiaoqiang Yan, and Yangdong Ye. A survey on information bottleneck. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3 + +[14] Yan Ju, Shan Jia, Lipeng Ke, Hongfei Xue, Koki Nagano, and Siwei Lyu. Fusing global and local features for generalized ai-synthesized image detection. In 2022 IEEE International Conference on Image Processing (ICIP), pages 3465-3469. IEEE, 2022. 2, 6, 7 +[15] Tero Karras. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 1, 5 +[16] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 5 +[17] Sohail Ahmed Khan and Duc-Tien Dang-Nguyen. Clipping the deception: Adapting vision-language models for universal deepfake detection. In Proceedings of the 2024 International Conference on Multimedia Retrieval, pages 1006-1015, 2024. 6, 7 +[18] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 5 +[19] Li Lin, Irene Amerini, Xin Wang, Shu Hu, et al. Robust clip-based detector for exposing diffusion model-generated images. arXiv preprint arXiv:2404.12908, 2024. 2 +[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5 +[21] Bo Liu, Fan Yang, Xiuli Bi, Bin Xiao, Weisheng Li, and Xinbo Gao. Detecting generated images by real images. In European Conference on Computer Vision, pages 95-110. Springer, 2022. 2 +[22] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 5 +[23] Yunpeng Luo, Junlong Du, Ke Yan, and Shouhong Ding. Lare^ 2: Latent reconstruction error based method for diffusion-generated image detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17006-17015, 2024. 1 +[24] Rabeeh Karimi Mahabadi, Yonatan Belinkov, and James Henderson. Variational information bottleneck for effective low-resource fine-tuning. arXiv preprint arXiv:2106.05469, 2021.3 +[25] Francesco Marra, Diego Gragnaniello, Luisa Verdoliva, and Giovanni Poggi. Do gans leave artificial fingerprints? In 2019 IEEE conference on multimedia information processing and retrieval (MIPR), pages 506-511. IEEE, 2019. 1 +[26] Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. 3 +[27] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 5 + +[28] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24480-24489, 2023. 2, 3, 6, 7 +[29] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 5 +[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 8 +[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1 +[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 5 +[33] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics++: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1-11, 2019. 5 +[34] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 5 +[35] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15638-15650, 2022. 8 +[36] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, and Yunchao Wei. Learning on gradients: Generalized artifacts representation for gan-generated images detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12105-12114, 2023. 6, 7 +[37] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28130-28139, 2024. 6, 7 +[38] Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. 3 + +[39] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8695-8704, 2020. 1, 2, 5, 6, 7 +[40] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22445-22455, 2023. 1, 6, 7 +[41] Among Wu and Cheng Deng. Tib: Detecting unknown objects via two-stream information bottleneck. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3 +[42] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 5 +[43] Ning Yu, Larry S Davis, and Mario Fritz. Attributing fake images to gans: Learning and analyzing gan fingerprints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7556-7566, 2019. 1 +[44] Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deepfake detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2185–2194, 2021. 2 +[45] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 5 +[46] Mingjian Zhu, Hanting Chen, Qiangyu Yan, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang. Genimage: A million-scale benchmark for detecting ai-generated image. Advances in Neural Information Processing Systems, 36, 2024. 5 \ No newline at end of file diff --git a/2025/Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network/images.zip b/2025/Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b9c3d68b73c45d8a7e1a2506b77fe6c60cffd97e --- /dev/null +++ b/2025/Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:612e7c18bec565f4ea639888abf9975e5f914852f40adb3b0a504ea9c8157ceb +size 714411 diff --git a/2025/Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network/layout.json b/2025/Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6d7ad3a2bd7c4b4503fdd96744ebb45d5ea23fc6 --- /dev/null +++ b/2025/Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network/layout.json @@ -0,0 +1,6913 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 102, + 545, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 102, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 66, + 102, + 545, + 138 + ], + "type": "text", + "content": "Towards Universal AI-Generated Image Detection by Variational Information Bottleneck Network" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 128, + 161, + 481, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 161, + 481, + 190 + ], + "spans": [ + { + "bbox": [ + 128, + 161, + 481, + 190 + ], + "type": "text", + "content": "Haifeng Zhang, Qinghui He, Xiuli Bi, Weisheng Li, Bo Liu*, Bin Xiao Chongqing University of Posts and Telecommunications" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 153, + 191, + 457, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 191, + 457, + 203 + ], + "spans": [ + { + "bbox": [ + 153, + 191, + 457, + 203 + ], + "type": "text", + "content": "s220331143@stu.cqupt.edu.cn, s220232014@stu.cqupt.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 192, + 205, + 414, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 205, + 414, + 217 + ], + "spans": [ + { + "bbox": [ + 192, + 205, + 414, + 217 + ], + "type": "text", + "content": "{bix1, liws, boliu, xiaobin}@cqupt.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 270, + 297, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 297, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 297, + 534 + ], + "type": "text", + "content": "The rapid advancement of generative models has significantly improved the quality of generated images. Meanwhile, it challenges information authenticity and credibility. Current generated image detection methods based on large-scale pre-trained multimodal models have achieved impressive results. Although these models provide abundant features, the authentication task-related features are often submerged. Consequently, those authentication task-irrelated features cause models to learn superficial biases, thereby harming their generalization performance across different model genera (e.g., GANs and Diffusion Models). To this end, we proposed VIB-Net, which uses Variational Information Bottlenecks to enforce authentication task-related feature learning. We tested and analyzed the proposed method and existing methods on samples generated by 17 different generative models. Compared to SOTA methods, VIB-Net achieved a " + }, + { + "bbox": [ + 55, + 270, + 297, + 534 + ], + "type": "inline_equation", + "content": "5.55\\%" + }, + { + "bbox": [ + 55, + 270, + 297, + 534 + ], + "type": "text", + "content": " improvement in mAP and a " + }, + { + "bbox": [ + 55, + 270, + 297, + 534 + ], + "type": "inline_equation", + "content": "9.33\\%" + }, + { + "bbox": [ + 55, + 270, + 297, + 534 + ], + "type": "text", + "content": " increase in accuracy. Notably, in generalization tests on unseen generative models from different series, VIB-Net improved mAP by " + }, + { + "bbox": [ + 55, + 270, + 297, + 534 + ], + "type": "inline_equation", + "content": "12.48\\%" + }, + { + "bbox": [ + 55, + 270, + 297, + 534 + ], + "type": "text", + "content": " and accuracy by " + }, + { + "bbox": [ + 55, + 270, + 297, + 534 + ], + "type": "inline_equation", + "content": "23.59\\%" + }, + { + "bbox": [ + 55, + 270, + 297, + 534 + ], + "type": "text", + "content": " over SOTA methods. The code is available at https://github.com/oceanzhf/VIBAIGCDetect." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 555, + 135, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 135, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 135, + 568 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 576, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 296, + 696 + ], + "type": "text", + "content": "The continuous advancements in deep learning have led to the emergence of numerous well-known image generation models, such as GANs[9][15] and Diffusion[12][31]. Images generated by these models have reached a level where observers can hardly distinguish them from real images, and the flexibility of tools based on these models has also greatly improved [6]. With a few simple steps, people can easily obtain high-quality images as needed. While this presents a tremendous opportunity for fields like visual arts, it also serves as a paradise for individuals looking to fab" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 244, + 554, + 385 + ], + "blocks": [ + { + "bbox": [ + 316, + 244, + 554, + 385 + ], + "lines": [ + { + "bbox": [ + 316, + 244, + 554, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 244, + 554, + 385 + ], + "type": "image", + "image_path": "432848555de54ec22b2ec2160073a539e39993a3a52b1f214925a0ea15c4b742.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 394, + 556, + 438 + ], + "lines": [ + { + "bbox": [ + 313, + 394, + 556, + 438 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 556, + 438 + ], + "type": "text", + "content": "Figure 1. Each method's overall performance('P') on test data and generalization performance ('G') on unseen data are linked by dashed lines. Our detector demonstrates a clear advantage over all SOTA methods in both metrics." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 460, + 555, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 520 + ], + "type": "text", + "content": "ricate misinformation. Given the ease of creating manipulated images today and the potential risk of using them for malicious purposes, there is an urgent need to develop detectors that help distinguish between real and generated images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 521, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 700 + ], + "type": "text", + "content": "Existing detection methods [4][39][25][43] focus on identifying specific traces left by generators in either the spatial or frequency domain of generated images, thus establishing a decision boundary between real and generated images. These detection methods are highly effective in detecting images generated by GANs and earlier diffusion models. However, their performance drops significantly when facing images generated by recent, popular models that visually resemble real images. Although some error reconstruction-based methods [40][23] have been developed specifically for detecting images generated by diffusion models, they similarly experience performance declines when applied to GAN-generated images. Consequently, current research increasingly focuses on improving the cross-model genera detection capabilities of detectors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "content": "Generalization across different domains is a primary re" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 703, + 148, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 703, + 148, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 703, + 148, + 713 + ], + "type": "text", + "content": "* Corresponding Author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23828" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 453 + ], + "type": "text", + "content": "search challenge in deep learning. In this context, the emergence of large-scale pre-trained multimodal models has introduced new approaches and promising results. For instance, by using the image encoder from the CLIP [30] model to extract image features and training a classifier with linear probing, there is a notable improvement in generalization performance over models like ResNet50[28][6]. However, compared to traditional classification tasks, real and generated images share subtle, harder-to-capture distinguishing features. This implies that the features extracted by large-scale pre-trained models for the task of distinguishing real from fake images may contain a considerable amount of task-irrelevant features. Task-irrelevant features can either have no impact on the realizness of the image or, though effective in isolation, become redundant when combined with other features. These features not only interfere with the classifier's ability to accurately learn decision boundaries but also hinder the model's ability to capture and distinguish task-relevant characteristics. More critically, task-irrelevant features may lead the model to learn superficial biases. For example, if a fake image contains a tree while the real image does not, the model might erroneously associate the presence of the tree with being a fake image. Consequently, when detecting other images, the model may mistakenly classify a real image as fake if it contains a tree. Therefore, when dealing with high-quality generated images with subtle distinguishing features, it is challenging to fine-tune the model to separate essential from unnecessary information. This can lead to an overfitting of the incidental statistical correlations between unnecessary information and target images, ultimately undermining the model's generalization and discrimination capabilities." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 463, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 463, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 463, + 294, + 712 + ], + "type": "text", + "content": "How can we harness the powerful feature extraction capability of large models while minimizing the interference of unnecessary information? By approaching this from a novel perspective, we propose addressing the issue using Information Bottleneck Theory. This theory aims to create latent encodings that not only effectively represent the target but also maximally compress the original input, thereby avoiding the impact of task-irrelevant features on the model. Specifically, we introduce a method called VIB-Net, which couples a large-scale pre-trained multimodal model with Variational Information Bottleneck (VIB) [1]. Considering that feature representations extracted by the model's image encoder are more task-specific, we decouple and filter these features using latent variables. This method maximizes the mutual information between the latent feature and the classification output, while minimizing the mutual information between the latent feature and the input, resulting in a concise feature that should only contain the effective information needed for the task. This concise latent feature helps the model avoid overfitting to irrelevant features and the incidental correlations with the task, preventing the model" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "text", + "content": "from learning superficial biases, which is beneficial to the model's generalization performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 316, + 97, + 553, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 286 + ], + "type": "text", + "content": "In summary, our contributions are as follows: (1) To the best of our knowledge, we are the first to apply Information Bottleneck theory and techniques to the detection of generated images, demonstrating the potential of this theory and approach in this field. (2) We analyze the limitations of methods based on large-scale pre-trained models in detecting images generated by unseen generative models, and present our theory of what could be wrong with the existing paradigm. (3) We propose VIB-Net for the detection of generated images, which uses VIB to decouple and filter the general features extracted by large-scale pre-trained multimodal models, removing task-irrelevant information to enhance detection performance. (4) A large number of experiments validate the advantages of our method (see Fig. 1), especially in the improvement of generalization performance " + }, + { + "bbox": [ + 316, + 97, + 553, + 286 + ], + "type": "inline_equation", + "content": "(+23.59\\%" + }, + { + "bbox": [ + 316, + 97, + 553, + 286 + ], + "type": "text", + "content": " accuracy)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 316, + 299, + 397, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 397, + 311 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 397, + 311 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 316, + 319, + 553, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 569 + ], + "type": "text", + "content": "Generalization Performance. The continuous advancement of generated image models presents ongoing challenges for generated image detection, making the generalization performance of detection models on unseen generators an increasing focus for researchers[14][21]. [39] explored the effects of data preprocessing and augmentation methods on the generalization performance of a ResNet50[11]-based detector. [44] argues that the differences between real and fake images are subtle and localized, framing forgery detection as a fine-grained task that could positively impact generalization performance. [3] takes a frequency-domain approach, mapping real images to a dense subspace and synthetic images outside this subspace to enhance generalization. While these methods have achieved certain successes on different models within the same generation series, they show limitations across different series. This is due to the fact that, with the development of generative algorithms, subtle traces in generated images are becoming increasingly difficult to capture. Additionally, artifacts in the frequency domain differ between model series and are gradually decreasing in visibility." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 316, + 571, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 712 + ], + "type": "text", + "content": "The impressive results of large-scale pre-trained models in terms of transferability and generalization have sparked significant interest among researchers in the field of generated image detection. [28] argues that existing detectors perform asymmetric learning, where they use CLIP's image encoder to extract image features, exploring the potential of large models in detecting different model families. [6] suggests that detectors based on large models rely on higher-level semantic features, which enhances the generalization ability of the detectors. [19] uses both image and text features extracted by the CLIP model, and then employs a multilayer perceptron to detect images generated by" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23829" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 228 + ], + "type": "text", + "content": "diffusion models, improving generalization performance. Information Bottleneck. The Information Bottleneck (IB) [38] is an innovative information-theoretic method used for pattern analysis and representation learning. It provides an elegant balance between data compression and information retention[13]. VIB [1] is a variational approximation of IB, aimed at parameterizing IB using neural networks. [2] demonstrated VIB's capability in improving classification calibration and out-of-distribution data. [24] explored VIB as a regularization technique to address overfitting issues in NLP models. [41] proposed a dual-stream information bottleneck method to utilize in-distribution data to enhance out-of-distribution target detection tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 229, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 229, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 296, + 373 + ], + "type": "text", + "content": "Based on the above analysis, current methods for detecting generated images using large-scale pre-trained models simply utilize the general representations extracted from images, without thoroughly exploring the relationship between these general features and downstream tasks. Additionally, Information Bottleneck is an effective technique for improving performance in cross-domain and cross-distribution tasks. Therefore, unlike existing methods, our approach leverages Information Bottleneck techniques to explore the potential of large-scale pre-trained models in generated image detection from a completely new perspective, achieving outstanding performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 384, + 141, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 141, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 141, + 396 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 405, + 178, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 178, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 178, + 417 + ], + "type": "text", + "content": "3.1. Problem background" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 422, + 296, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 554 + ], + "type": "text", + "content": "A common paradigm for using large-scale pretrained models is to leverage a large-scale pretrained multimodal model (e.g., CLIP) to learn the general representation of real and generated images, and then apply this feature representation to downstream tasks[28][6]. Building on their findings, we further consider the following points: (1) It is well-known that the features extracted by large-scale pre-trained visual encoders are general-purpose, meaning they contain many general features, such as semantic information. (2) Another widely accepted notion is that some critical features for distinguishing real from generated images are subtle and local." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 555, + 296, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 296, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 296, + 662 + ], + "type": "text", + "content": "Further combining these two points, we propose the following hypotheses: (i) The general features extracted by large-scale pre-trained multimodal models are more beneficial for real-fake image classification tasks than those extracted by pre-trained visual models. (ii) These general features contain task-irrelevant features for real-fake classification. (iii) These task-irrelevant features affect the classification of real and fake images. Next, we will delve into these three hypotheses." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 671, + 274, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 274, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 274, + 685 + ], + "type": "text", + "content": "3.2. Analysis of general feature representation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "For the first hypothesis, We use image pairs of real and generated images to aid our analysis[6], generating textual de" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 319, + 71, + 554, + 244 + ], + "blocks": [ + { + "bbox": [ + 319, + 71, + 554, + 244 + ], + "lines": [ + { + "bbox": [ + 319, + 71, + 554, + 244 + ], + "spans": [ + { + "bbox": [ + 319, + 71, + 554, + 244 + ], + "type": "image", + "image_path": "e1410341985ff647e745d20a946c655965b19f0cbbf303dfb911f10c27cde35e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 253, + 555, + 286 + ], + "lines": [ + { + "bbox": [ + 313, + 253, + 555, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 555, + 286 + ], + "type": "text", + "content": "Figure 2. UMAP visualization of features extracted from Dalle2 and Dalle3 generated images using large-scale pretrained multimodal(CLIP) and unimodal models(ResNet)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 315, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 555, + 483 + ], + "type": "text", + "content": "scriptions for 1,000 real images and using these descriptions to produce 1,000 corresponding images with DALL-E 2 and DALL-E 3. We then compare features extracted by large-scale pre-trained multimodal models(CLIP[30]) and pre-trained visual models(ResNet[11]) through UMAP [26] visualization of 3,000 images. The results show that features extracted by multimodal models (e.g., CLIP) exhibit greater distribution differences for real images and DALL-E 2-generated images, suggesting multimodal models may be better at distinguishing real from generated images. However, for DALL-E 3-generated images, both model types show similar feature distributions, possibly due to advancements in image generation tools, which complicate the classification task and require further fine-tuning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 486, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 713 + ], + "type": "text", + "content": "To further explore hypotheses (ii) and (iii), we visualized the extracted features using both category and authenticity as labels to observe their distribution differences. Specifically, we prepared images of two categories (car and cat), each containing 500 real images and 500 generated images. We then used CLIP to extract their features and performed UMAP dimensionality reduction for visualization, with the results shown in Fig. 3a. We can see that the features extracted by CLIP clearly form two clusters: the left cluster mainly consists of light blue and light red points, while the right cluster is mainly composed of light pink and light purple points. These two clusters correspond to category 1 and category 2, respectively. This indicates that CLIP can significantly distinguish category information. However, this clear separation of category information may affect the decision-making process for classifying real vs. generated images. When examining the distinction between real and generated images, we notice that the right cluster shows a noticeable separation of real and generated features, with" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23830" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 71, + 174, + 168 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 174, + 168 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 174, + 168 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 174, + 168 + ], + "type": "image", + "image_path": "fe8f885df46bbd6b886ae2693b1dc8ce8b2ffa83bcc61143cf31da8d0c17917d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 171, + 174, + 190 + ], + "lines": [ + { + "bbox": [ + 56, + 171, + 174, + 190 + ], + "spans": [ + { + "bbox": [ + 56, + 171, + 174, + 190 + ], + "type": "text", + "content": "(a) Visualization of features extracted by CLIP." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 199, + 295, + 222 + ], + "lines": [ + { + "bbox": [ + 55, + 199, + 295, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 199, + 295, + 222 + ], + "type": "text", + "content": "Figure 3. UMAP visualization of features extracted from real and generated images of two category using CLIP and VIB-Net." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 175, + 72, + 294, + 168 + ], + "blocks": [ + { + "bbox": [ + 175, + 72, + 294, + 168 + ], + "lines": [ + { + "bbox": [ + 175, + 72, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 175, + 72, + 294, + 168 + ], + "type": "image", + "image_path": "1ab822a33b6db285e7ab7f352b8aedae2588e0ed8278793c8c97bcbb3a4451d0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 171, + 295, + 190 + ], + "lines": [ + { + "bbox": [ + 176, + 171, + 295, + 190 + ], + "spans": [ + { + "bbox": [ + 176, + 171, + 295, + 190 + ], + "type": "text", + "content": "(b) Visualization of features extracted by VIB-Net." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 248, + 295, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 248, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 54, + 248, + 295, + 415 + ], + "type": "text", + "content": "two different colors observed within the cluster. This suggests that CLIP is capable of distinguishing real and generated features within category 2. However, in the left cluster, CLIP does not show significant ability to differentiate between real and generated features. We observe that in the left cluster (category 1), the two colors are mixed together, showing poor distinguishability. From these observations, it is easy to infer that the internal separation of real or generated images caused by semantic category information makes it more difficult for the classifier to learn the decision boundary. More importantly, the strong sensitivity of large-scale pre-trained multimodal models like CLIP to semantic information leads the model to tend to learn superficial biases." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 418, + 295, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 418, + 295, + 538 + ], + "spans": [ + { + "bbox": [ + 54, + 418, + 295, + 538 + ], + "type": "text", + "content": "Based on this viewpoint, we consider whether it is possible to minimize the irrelevant features within these multimodal features while retaining the ones that are relevant to the classification task. This would help the model avoid learning superficial biases, allow the classifier to more easily learn decision boundaries, and thereby enhance the model's discrimination ability and generalization capability. This is the motivation of this paper. As shown in Fig. 3b, the method(VIB-Net) proposed in this paper achieves significant results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 559, + 122, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 122, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 122, + 572 + ], + "type": "text", + "content": "4. Approach" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "text", + "content": "Based on the issues analyzed in Sec. 3, an effective approach is to filter the multimodal features in order to obtain the representations most beneficial for the real-fake classification task, which serves as the starting point of this paper. Overall, our method is illustrated in Fig. 4. VIB attempts to filter out irrelevant features and to find the most concise representation, making it easier for the classifier to learn the features needed for distinguishing between real and fake tasks, thereby significantly improving the accuracy of the classification task. Below, we will provide a detailed description of our method." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 495, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 495, + 83 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 495, + 83 + ], + "type": "text", + "content": "4.1. Information bottleneck constraint" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 89, + 555, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 209 + ], + "type": "text", + "content": "To specifically optimize the filtering of irrelevant information in the input features while retaining task-relevant information, we apply the information bottleneck (IB) principle as a constraint on the features to be optimized. Let " + }, + { + "bbox": [ + 313, + 89, + 555, + 209 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 89, + 555, + 209 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 89, + 555, + 209 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 89, + 555, + 209 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 89, + 555, + 209 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 313, + 89, + 555, + 209 + ], + "type": "text", + "content": " represent the input, the latent compressed representation, and the output target, respectively. The goal of IB is to find the maximum compressed representation " + }, + { + "bbox": [ + 313, + 89, + 555, + 209 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 89, + 555, + 209 + ], + "type": "text", + "content": " of the input " + }, + { + "bbox": [ + 313, + 89, + 555, + 209 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 89, + 555, + 209 + ], + "type": "text", + "content": ", while preserving as much information about the output " + }, + { + "bbox": [ + 313, + 89, + 555, + 209 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 313, + 89, + 555, + 209 + ], + "type": "text", + "content": " as possible. The maximization objective function is defined as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 373, + 216, + 555, + 229 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 216, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 373, + 216, + 555, + 229 + ], + "type": "interline_equation", + "content": "L _ {I B} = I (Z, Y) - \\beta I (Z, X) \\tag {1}", + "image_path": "52778efe8a584ff44be42b04f3a5139dee421d95f6a6f11d87634f10de132f6a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 235, + 555, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 555, + 295 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 313, + 235, + 555, + 295 + ], + "type": "inline_equation", + "content": "I(.,.)" + }, + { + "bbox": [ + 313, + 235, + 555, + 295 + ], + "type": "text", + "content": " denotes mutual information and " + }, + { + "bbox": [ + 313, + 235, + 555, + 295 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 235, + 555, + 295 + ], + "type": "text", + "content": " is a trade-off parameter. The first term constrains the latent compressed representation to retain sufficient information for downstream tasks, while the second term ensures that irrelevant and redundant information is filtered out." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 301, + 499, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 499, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 499, + 312 + ], + "type": "text", + "content": "4.2. Variational information bottleneck" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 319, + 554, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 554, + 367 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 554, + 367 + ], + "type": "text", + "content": "To parameterize the information bottleneck (IB) using deep neural networks, we perform variational inference on Equation (1). First, based on the definition of mutual information, Equation (1) can be rewritten as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 361, + 382, + 555, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 382, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 361, + 382, + 555, + 436 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {I B} = \\int d y d z p (y, z) \\log {\\frac {p (y | z)}{p (y)}} \\\\ - \\beta \\int d x d z p (x, z) \\log \\frac {p (z \\mid x)}{p (z)} \\tag {2} \\\\ \\end{array}", + "image_path": "1730bfccb3d6d7aac6a7513a915aa6b960d0a961c1f7765d56d821d2fe0d4f3d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 441, + 554, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 554, + 464 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 313, + 441, + 554, + 464 + ], + "type": "inline_equation", + "content": "p(y|z)" + }, + { + "bbox": [ + 313, + 441, + 554, + 464 + ], + "type": "text", + "content": " is defined by the encoder and the Markov chain as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 354, + 469, + 554, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 469, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 354, + 469, + 554, + 495 + ], + "type": "interline_equation", + "content": "p (y \\mid z) = \\int d x \\frac {p (y \\mid x) p (z \\mid x) p (x)}{p (z)} \\tag {3}", + "image_path": "19fe718d470e69348f8fb75ff93daf30ff1756c94da507623a634a5bf9c7ecbc.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "type": "text", + "content": "Since it is not computable here, we consider setting " + }, + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "type": "inline_equation", + "content": "q(y|z)" + }, + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "type": "text", + "content": " as a parameterized approximation of " + }, + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "type": "inline_equation", + "content": "p(y|z)" + }, + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "type": "text", + "content": ", serving as a \"decoder,\" which can be approximated by a linear layer. We introduce the Kullback-Leibler (KL) divergence to obtain an estimizable representation. As the KL divergence satisfies " + }, + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "type": "inline_equation", + "content": "\\mathrm{KL}(p(y|z),q(y|z))\\geq 0" + }, + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "type": "text", + "content": ", and by applying the Markov assumption, " + }, + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "type": "inline_equation", + "content": "p(y,z)" + }, + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "type": "text", + "content": " can be rewritten as " + }, + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "type": "inline_equation", + "content": "p(y,z) = \\int dx p(x)p(y|x)p(z|x)" + }, + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "type": "text", + "content": ", thereby providing a lower bound for the first term of the objective function:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 614, + 553, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 614, + 553, + 648 + ], + "spans": [ + { + "bbox": [ + 321, + 614, + 553, + 648 + ], + "type": "interline_equation", + "content": "I (Z, Y) \\geq \\int d x d y d z p (x) p (y | x) p (z | x) \\log q (y | z) \\tag {4}", + "image_path": "bc8c47d9de2ae28babe009113d8fa841eb5602d58ca5325d1061723559b0c190.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 649, + 553, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 649, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 649, + 553, + 685 + ], + "type": "text", + "content": "At the same time, let " + }, + { + "bbox": [ + 313, + 649, + 553, + 685 + ], + "type": "inline_equation", + "content": "r(z)" + }, + { + "bbox": [ + 313, + 649, + 553, + 685 + ], + "type": "text", + "content": " be an estimate of the prior probability " + }, + { + "bbox": [ + 313, + 649, + 553, + 685 + ], + "type": "inline_equation", + "content": "p(z)" + }, + { + "bbox": [ + 313, + 649, + 553, + 685 + ], + "type": "text", + "content": ", used to approximate the marginal distribution, thereby obtaining an upper bound for the second term:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 340, + 690, + 553, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 690, + 553, + 717 + ], + "spans": [ + { + "bbox": [ + 340, + 690, + 553, + 717 + ], + "type": "interline_equation", + "content": "I (Z, X) \\leq \\int d x d z p (x) p (z | x) \\log \\frac {p (z | x)}{r (z)} \\tag {5}", + "image_path": "942473551b43352e6dbf17da5b3046f4024e92e4471ab5a89008f6a71795be4c.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23831" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 553, + 239 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 553, + 239 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 553, + 239 + ], + "type": "image", + "image_path": "f2654d25fbfe695a4e4ecf0ae41380f6fa32b64f9234168d03dc86580330d688.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 246, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 54, + 246, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 54, + 246, + 555, + 281 + ], + "type": "text", + "content": "Figure 4. Overview of VIB-Net. We utilize Variational Information Bottleneck (VIB) techniques to map the image features extracted by CLIP's visual encoder into latent features " + }, + { + "bbox": [ + 54, + 246, + 555, + 281 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 54, + 246, + 555, + 281 + ], + "type": "text", + "content": ", where the information in " + }, + { + "bbox": [ + 54, + 246, + 555, + 281 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 54, + 246, + 555, + 281 + ], + "type": "text", + "content": " is filtered based on the Information Bottleneck (IB). Subsequently, the latent features " + }, + { + "bbox": [ + 54, + 246, + 555, + 281 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 54, + 246, + 555, + 281 + ], + "type": "text", + "content": " are used as input to an MLP to learn the decision boundary." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 300, + 296, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 300, + 296, + 410 + ], + "spans": [ + { + "bbox": [ + 54, + 300, + 296, + 410 + ], + "type": "text", + "content": "The encoder we use is defined as " + }, + { + "bbox": [ + 54, + 300, + 296, + 410 + ], + "type": "inline_equation", + "content": "p(z|x) = p_{\\theta}(z|x) = \\mathcal{N}\\left(z|f^{\\mu}(x), f^{\\Sigma}(x)\\right)" + }, + { + "bbox": [ + 54, + 300, + 296, + 410 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 300, + 296, + 410 + ], + "type": "inline_equation", + "content": "f(.)" + }, + { + "bbox": [ + 54, + 300, + 296, + 410 + ], + "type": "text", + "content": " represents a shallow MLP, " + }, + { + "bbox": [ + 54, + 300, + 296, + 410 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 54, + 300, + 296, + 410 + ], + "type": "text", + "content": " denotes the K-dimensional mean vector of the output " + }, + { + "bbox": [ + 54, + 300, + 296, + 410 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 54, + 300, + 296, + 410 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 300, + 296, + 410 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 54, + 300, + 296, + 410 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 54, + 300, + 296, + 410 + ], + "type": "inline_equation", + "content": "K \\times K" + }, + { + "bbox": [ + 54, + 300, + 296, + 410 + ], + "type": "text", + "content": " covariance matrix. We select a parameterized Gaussian distribution for both " + }, + { + "bbox": [ + 54, + 300, + 296, + 410 + ], + "type": "inline_equation", + "content": "p_{\\theta}(z|x)" + }, + { + "bbox": [ + 54, + 300, + 296, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 300, + 296, + 410 + ], + "type": "inline_equation", + "content": "r(z)" + }, + { + "bbox": [ + 54, + 300, + 296, + 410 + ], + "type": "text", + "content": " to allow for the analytical computation of KL divergence. Additionally, we employ the reparameterization trick to estimate gradients [18]. Combining all the components above yields the following objective function:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 426, + 295, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 426, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 85, + 426, + 295, + 475 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {V I B} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} z \\sim p _ {\\theta} (z | x) [ - \\log q (y _ {\\mathrm {n}} | z _ {n}) ] \\\\ + \\beta \\mathrm {K L} [ p _ {\\theta} (z | x _ {n}), r (z) ] \\tag {6} \\\\ \\end{array}", + "image_path": "5f8a38f5cc927016b77e4d61bd4c8e0180c43190289885e14d98ba4faf5b3533.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 481, + 228, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 228, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 228, + 493 + ], + "type": "text", + "content": "4.3. Workflow of the overall method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 498, + 296, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 643 + ], + "type": "text", + "content": "As shown in Fig. 4, the parts in VIB-Net work in the following way: First, the real and generated images " + }, + { + "bbox": [ + 55, + 498, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\{x_{1}, x_{2}, x_{3}, \\dots, x_{n}\\}" + }, + { + "bbox": [ + 55, + 498, + 296, + 643 + ], + "type": "text", + "content": " are input into the CLIP image encoder, yielding the corresponding image features " + }, + { + "bbox": [ + 55, + 498, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\{t_{1}, t_{2}, t_{3}, \\dots, t_{n}\\}" + }, + { + "bbox": [ + 55, + 498, + 296, + 643 + ], + "type": "text", + "content": ". Next, these feature vectors are passed into the Variational Information Bottleneck (VIB). Specifically, the image features are processed through a shallow MLP, then a linear layer " + }, + { + "bbox": [ + 55, + 498, + 296, + 643 + ], + "type": "inline_equation", + "content": "f(.)" + }, + { + "bbox": [ + 55, + 498, + 296, + 643 + ], + "type": "text", + "content": " calculates " + }, + { + "bbox": [ + 55, + 498, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 498, + 296, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 498, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 55, + 498, + 296, + 643 + ], + "type": "text", + "content": ", which are transformed using a softmax activation. This layer has " + }, + { + "bbox": [ + 55, + 498, + 296, + 643 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 498, + 296, + 643 + ], + "type": "text", + "content": " dimensions, representing the bottleneck size. At this point, we obtain the latent compressed feature " + }, + { + "bbox": [ + 55, + 498, + 296, + 643 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 498, + 296, + 643 + ], + "type": "text", + "content": ", which is then fed into a linear layer " + }, + { + "bbox": [ + 55, + 498, + 296, + 643 + ], + "type": "inline_equation", + "content": "q(y|z)" + }, + { + "bbox": [ + 55, + 498, + 296, + 643 + ], + "type": "text", + "content": " to output the target task label " + }, + { + "bbox": [ + 55, + 498, + 296, + 643 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 498, + 296, + 643 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 652, + 137, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 137, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 137, + 666 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 672, + 116, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 116, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 116, + 685 + ], + "type": "text", + "content": "5.1. Settings" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "To better evaluate the model's performance, we utilize two mainstream public datasets: ForenSynths[39] and" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 300, + 555, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 419 + ], + "type": "text", + "content": "Genimage[46]. The ForenSynths dataset primarily consists of images generated by GANs, while the Genimage dataset mainly includes images generated by diffusion models. Consequently, we set up two training and testing schemes: (1) Training on ForenSynths and then testing on both ForenSynths and Genimage, and (2) similarly, training on Genimage and testing on both Genimage and ForenSynths. This setup allows us to more effectively validate the generalization performance of the detector on cross model genera scenarios. The details are as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 422, + 372, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 422, + 372, + 433 + ], + "spans": [ + { + "bbox": [ + 314, + 422, + 372, + 433 + ], + "type": "text", + "content": "Training set:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 434, + 554, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 554, + 529 + ], + "type": "text", + "content": "The ForenSynths training set consists of 20 different categories, each containing 18,000 real images from the LSUN dataset [42] and an equal number of images generated using ProGAN[15]. Following the setup in Genimage, we use images generated by Stable Diffusion v1.4 [32] in the Genimage dataset, along with their corresponding real images, as the training set. In total, these include 32,000 real and generated images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 531, + 365, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 531, + 365, + 542 + ], + "spans": [ + { + "bbox": [ + 314, + 531, + 365, + 542 + ], + "type": "text", + "content": "Testing set:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 544, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 616 + ], + "type": "text", + "content": "11 models from ForenSynths The test set includes images generated by 11 different models, primarily from the GAN family[15][45][29][5][16]. The real images are sourced from six datasets, including ImageNet[34], LSUN[42], FaceForensics++[33], CelebA-HQ[15], CelebA[22], and COCO[20]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 616, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 555, + 689 + ], + "type": "text", + "content": "7 models from Genimage To validate the method across different generation models, we use images generated by seven advanced generators along with their corresponding real images as part of the test set. These seven generation models include Stable Diffusion v1.4/1.5[32], VQDM[10], ADM[7], GLIDE[27], Wukong, and Midjourney." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Real-vs-Fake classification baselines: We compare VIB-Net with existing state-of-the-art detection methods, in" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23832" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 555, + 205 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 205 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 205 + ], + "type": "table", + "html": "
Detection methodDiffusion ModelsGenerative Adversarial NetworksOthersAP
SDV1.4SDV1.5ADMGLIDEMidjourneyWukongVQDMPro-GANCycle-Big-GANStyle-GANStar-GANGau-GANDeep-fakeAverage
CNNSpot[39]99.9899.8351.1058.8067.9399.8049.9253.1550.2349.7955.9847.0756.0854.8654.0363.24
Fusing[14]99.9097.9869.3094.2081.2099.9084.6067.6387.7969.3767.9091.2043.0872.5689.4281.07
Lgrad[36]99.9499.9258.5284.0091.0699.7256.3483.5990.2447.5182.7499.1949.2566.4965.0978.24
Univfd[28]96.0496.2666.3493.7392.0890.9874.5351.7763.4275.8154.1254.9365.9970.2483.3475.31
NPR[37]100.0099.9794.7095.8095.50100.0086.3083.3094.9072.0082.7097.3066.0085.3095.9089.98
CLIPping[17]93.9793.1068.0087.4477.3486.5277.1788.5488.4485.3377.2389.8281.5661.1957.3780.87
Ours100.0099.9795.4997.1397.8199.9397.0096.5998.4497.1784.3197.6096.9481.3293.2795.53
", + "image_path": "ec2b24faa11d5f7a5f1a5dcca0ab1010d0fb1a26e14f0229abea264ef53b80fa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 245, + 555, + 381 + ], + "blocks": [ + { + "bbox": [ + 55, + 212, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 212, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 555, + 236 + ], + "type": "text", + "content": "Table 1. The " + }, + { + "bbox": [ + 55, + 212, + 555, + 236 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 55, + 212, + 555, + 236 + ], + "type": "text", + "content": " values of different methods based on diffusion sources. Models outside the Diffusion Models (DMs) column are considered as the generalization domain across generative model series." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 245, + 555, + 381 + ], + "lines": [ + { + "bbox": [ + 58, + 245, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 58, + 245, + 555, + 381 + ], + "type": "table", + "html": "
Detection methodDiffusion ModelsGenerative Adversarial NetworksOthersACC
SDV1.4SDV1.5ADMGLIDEMidjourneyWukongVQDMPro-GANCycle-Big-GANStyle-GANStar-GANGau-GANDeep-SANAverage
CNNSpot[39]99.4899.3550.1050.9056.4297.9050.0450.2749.8150.1050.9849.7750.3851.9850.22
Fusing[14]99.9099.9151.3057.5052.3099.9064.2051.2052.4053.5050.2058.2049.3251.0264.84
Lgrad[36]99.1299.0553.0064.2476.3497.5350.9361.6160.7448.8261.4350.1749.7050.1756.49
Univfd[28]83.5584.8053.3575.3071.6073.5555.1058.6559.3061.4556.8061.4555.3058.4072.00
NPR[37]100.0099.9073.0089.7082.30100.0068.3060.3067.2059.2058.0073.2052.0074.8089.60
CLIPping[17]96.0795.4870.1485.0077.6688.8679.3588.7888.4889.5780.6990.8285.9466.9161.64
Ours99.5599.2073.8574.2588.0598.2589.3589.7088.6091.2074.1080.7087.1572.0081.50
", + "image_path": "2c33959f89ed5c07d0089de82cedaf079f6dd783e06c4f9afa7a02d5978a4abf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 390, + 555, + 413 + ], + "lines": [ + { + "bbox": [ + 55, + 390, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 555, + 413 + ], + "type": "text", + "content": "Table 2. Similar to Tab. 1, we compare these methods using classification accuracy, which is influenced by manually set thresholds. For a fair comparison, the classification threshold for all models is set to 0.5." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 431, + 295, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 431, + 295, + 516 + ], + "spans": [ + { + "bbox": [ + 54, + 431, + 295, + 516 + ], + "type": "text", + "content": "cluding CNNDetection (CVPR 2020) [39], FreDect (ICML 2020) [8], Fusing (ICIP 2022) [14], LGrad (CVPR 2023) [36], DIRE (ICCV 2023) [40], Ojha (CVPR 2023) [28], NPR (CVPR 2024) [37], CLIPping(ICMR 2024) [17]. To ensure a fair comparison, we re-train and predict using the official code on ForenSynths and Genimage." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 518, + 295, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 518, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 54, + 518, + 295, + 625 + ], + "type": "text", + "content": "Evaluation metrics We follow existing research and consider two metrics: Average Precision (AP) and Accuracy (ACC). Average Precision (AP) is threshold-independent and evaluates the overall performance of the model. Accuracy is defined as the number of correct predictions divided by the number of test images, and this metric is influenced by the decision threshold. Given that there is no prior information in real-world scenarios, we standardize the decision threshold to 0.5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 629, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 629, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 629, + 295, + 715 + ], + "type": "text", + "content": "Implementation Details Our method is implemented using the PyTorch framework on the NVIDIA Tesla V100. The model is trained with the Adam optimizer, a learning rate of " + }, + { + "bbox": [ + 54, + 629, + 295, + 715 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 54, + 629, + 295, + 715 + ], + "type": "text", + "content": ", and a batch size of 256. Following the same data augmentation strategy as in [39][28], which includes Gaussian blur and JPEG compression, experiments have shown that this is critical for the model's generalization ability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 431, + 555, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 455 + ], + "type": "text", + "content": "5.2. Performance evaluation based on Diffusion sources" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 461, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 461, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 312, + 461, + 555, + 605 + ], + "type": "text", + "content": "It is well-known that images generated by the current state-of-the-art diffusion models are more realistic than those generated by other types of generative models. To validate the effectiveness of the proposed method, we designed a challenging experimental scenario. Specifically, we trained various detector using Stable Diffusion v1.4 and then tested them on images generated by 15 different models, which include the latest diffusion-based generative models as well as classic GANs. The results are presented in Tabs. 1 and 2. Bold indicates the highest value in the column, and underline indicates the second highest value in the column. This applies to the other tables as well." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "type": "text", + "content": "We first evaluated the performance of each model on the latest generative models. As shown in Tabs. 1 and 2, advanced diffusion models significantly challenged most detection models, causing a sharp drop in performance. For example, Univfd and Lgrad experienced a " + }, + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "type": "text", + "content": " decrease in accuracy and precision when tested on ADM, Midjourney, and VQDM. This decline is likely due to the high-quality images produced by these models, which make it difficult for detection methods to extract distinguishing fea" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23833" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 555, + 200 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 200 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 200 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 200 + ], + "type": "table", + "html": "
Detection methodGenerative Adversarial NetworksPerceptual lossOthersDiffusion ModelsAP
Pro-GANCycle-GANBig-GANStyle-GANStar-GANGau-GANCRNIMLEDeep-fakeSANSDV1.4SDV1.5ADMGLIDEMidjourneyWukongVQDMAverage
CNNSpot[39]99.9996.4087.5096.9494.2498.2896.6198.7664.4255.8952.8653.2565.1468.1056.6051.1569.4976.80
FreDect[8]99.9984.7793.6288.9799.4882.8582.2769.7770.7749.5038.5038.4163.7254.7347.2540.4486.0170.06
Fusing[14]100.0095.5090.7699.4899.8288.3290.3296.0271.1277.3365.3065.6274.8577.4469.9164.5375.4282.46
Lgrad[36]99.9094.0190.7599.8099.9879.2961.6467.0571.7145.0970.9071.7271.8375.9671.4266.5170.2376.93
Univfd[28]100.0099.2198.3197.9899.3599.8096.7299.0082.0482.1885.4882.3084.3484.0469.1090.1394.9690.88
NPR[37]100.0098.5087.8099.8099.9085.5050.0050.0082.4071.6084.0084.6074.6085.7085.4080.5081.2082.44
CLIPping[17]99.8594.0391.2795.2499.0589.5279.8780.3373.8956.6258.2857.8577.2577.9352.4360.6580.0077.89
Ours100.0099.8099.2998.7999.7299.9990.8996.8392.6491.6287.2486.9887.8888.5375.6890.9296.5193.14
", + "image_path": "96b08620198dd1fda7f1a026e64f21038d153e055719354ff20bd9af4f1aad29.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 241, + 555, + 380 + ], + "blocks": [ + { + "bbox": [ + 55, + 208, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 208, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 555, + 232 + ], + "type": "text", + "content": "Table 3. The " + }, + { + "bbox": [ + 55, + 208, + 555, + 232 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 55, + 208, + 555, + 232 + ], + "type": "text", + "content": " values of different methods trained on GAN source images. Generative models other than those in the GANs column can be considered as the generalization domain." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 241, + 555, + 380 + ], + "lines": [ + { + "bbox": [ + 58, + 241, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 58, + 241, + 555, + 380 + ], + "type": "table", + "html": "
Detection methodGenerative Adversarial NetworksPerceptual lossOthersDiffusion ModelsACC
Pro-GANCycle-GANBig-GANStyle-GANStar-GANGau-GANCRNIMLEDeep-fakeSANSDV1.4SDV1.5ADMGLIDEMidjourneyWukongVQDMAverage
CNNSpot[39]99.9987.5971.1889.9594.6081.4485.8785.8751.6950.0050.8250.8860.2057.8550.7751.1356.2069.18
FreDect[8]99.3678.7681.9778.0194.6280.5760.0460.1363.2950.0040.0240.3864.6655.4346.8941.5478.9565.57
Fusing[14]99.9087.0177.3285.2197.0476.9585.1788.5453.7654.5651.0451.3456.5057.1552.1251.6755.0969.43
DIRE[40]100.0076.6572.1285.03100.0067.79----49.7449.8375.7871.7558.0154.4653.68-
Lgrad[36]99.8086.9485.6391.0899.2772.4950.4950.5056.4244.4763.0363.6767.1066.1056.2063.6067.0269.64
Univfd[28]99.9098.5094.5084.4095.8599.5059.5072.0067.4056.5063.1063.5766.9061.7057.8571.0685.0076.31
NPR[37]99.8096.1084.4097.7099.3082.5050.0050.0080.2069.2076.6077.9069.7077.3077.8076.1078.1078.98
CLIPping[17]99.8896.7494.7794.8799.4794.1585.2387.7376.4860.2761.2660.7880.7780.6953.9563.9484.7680.93
Ours99.9999.0095.7591.2598.9599.7071.2586.7583.2070.5071.5570.0071.4569.4061.2575.9086.6582.50
", + "image_path": "2b78035a5b0191d0a345b4c6ace643fa0f58d9a5fd5c3da9ffc538edebe59679.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 389, + 555, + 412 + ], + "lines": [ + { + "bbox": [ + 55, + 389, + 555, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 555, + 412 + ], + "type": "text", + "content": "Table 4. Accuracy of different methods trained on GAN source images. The data from [40] is obtained from [37], and ”-” indicates that the method did not show data for the corresponding model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 427, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 427, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 54, + 427, + 295, + 594 + ], + "type": "text", + "content": "tures. We also observed that these models performed better on generative models with distributions similar to the training data (e.g., SDV1.5, Wukong). This suggests that they may have learned superficial biases from irrelevant features, which are superficially correlated with labels and do not generalize well to images from different distributions. In contrast, NPR, which detects authenticity by analyzing pixel relationships, partially mitigates these biases, resulting in less performance decline. When testing on GANs, we observed significant performance drops, particularly on GauGAN, where the accuracy and precision were around " + }, + { + "bbox": [ + 54, + 427, + 295, + 594 + ], + "type": "inline_equation", + "content": "50 - 60\\%" + }, + { + "bbox": [ + 54, + 427, + 295, + 594 + ], + "type": "text", + "content": ". This further highlights how even subtle relevant features and surface biases can severely impact the generalization performance of these models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": "In contrast, Our method outperforms the baseline model, Univfd, with improvements of " + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "inline_equation", + "content": "20.22\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": " in average precision and " + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "inline_equation", + "content": "20.46\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": " in accuracy. This gap is even more pronounced across generative model families (GANs), where our method surpasses Univfd by " + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "inline_equation", + "content": "34.17\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": " in precision and " + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "inline_equation", + "content": "26.41\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": " in accuracy. Compared to state-of-the-art methods, our approach improves average precision by " + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "inline_equation", + "content": "5.55\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": " and accuracy by " + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "inline_equation", + "content": "9.33\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": ". For other generative models (GANs), VIB-Net achieves gains of " + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "inline_equation", + "content": "12.48\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": " in average precision and" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 427, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 555, + 475 + ], + "type": "text", + "content": "23.59% in accuracy. This is due to VIB-Net's ability to retain compact deep features while eliminating irrelevant surface features, enhancing generalization. Overall, these results highlight the advantages of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 480, + 553, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 553, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 553, + 493 + ], + "type": "text", + "content": "5.3. Performance evaluation based on GAN sources" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 498, + 555, + 713 + ], + "type": "text", + "content": "So far, we have explored the generalization performance of using images generated by specific diffusion models as training data. Next, we trained using images generated by ProGAN and then tested on 18 different generative models to explore the performance of the detector on unseen generative models. The results are shown in Tabs. 3 and 4. From the tables, we can see that almost all models perform well when detecting images generated by other GAN variants. This is because, compared to diffusion-based generation methods, images generated by these GAN variants are visually more distinguishable, making it easier for the model to learn the distinguishing features of real and fake. However, when detecting images generated by models other than GAN variants, the performance of most models sharply declines. This suggests that even when the training data contains rich features related to real/fake classification, it does not necessarily ensure better generalization performance. This is because the model may overfit the incidental correla" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23834" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 553, + 149 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 149 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 149 + ], + "type": "image", + "image_path": "03c6eff0437e82330a33e46b3b5553ef28669794ee5a81cf45ccb08c1d454389.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 155, + 553, + 235 + ], + "blocks": [ + { + "bbox": [ + 56, + 155, + 553, + 235 + ], + "lines": [ + { + "bbox": [ + 56, + 155, + 553, + 235 + ], + "spans": [ + { + "bbox": [ + 56, + 155, + 553, + 235 + ], + "type": "image", + "image_path": "446f4f542fe26f5297875262a6a6efa70423d50c08a5ff47202b7b01c2126714.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 243, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 243, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 555, + 266 + ], + "type": "text", + "content": "Figure 5. Ablation study on the network architecture. The " + }, + { + "bbox": [ + 55, + 243, + 555, + 266 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 55, + 243, + 555, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 243, + 555, + 266 + ], + "type": "inline_equation", + "content": "ACC" + }, + { + "bbox": [ + 55, + 243, + 555, + 266 + ], + "type": "text", + "content": " values of different VIB-Net variants on various models. CLIP:ViT-L/14 shows better performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 281, + 295, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 295, + 341 + ], + "type": "text", + "content": "tions between irrelevant features and the task, which harms generalization performance. Univfd performs better across model families, but since it extracts general features, many of these features are unrelated to the real/fake classification task, which naturally affects the decision boundary." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 341, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 296, + 449 + ], + "type": "text", + "content": "In contrast, our method shows better performance in detecting real and fake images. Compared to the baseline Univfd, our method achieved improvements of " + }, + { + "bbox": [ + 55, + 341, + 296, + 449 + ], + "type": "inline_equation", + "content": "2.26\\%" + }, + { + "bbox": [ + 55, + 341, + 296, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 341, + 296, + 449 + ], + "type": "inline_equation", + "content": "6.19\\%" + }, + { + "bbox": [ + 55, + 341, + 296, + 449 + ], + "type": "text", + "content": " in average precision and accuracy. In OOD scenarios (excluding GANs), our method outperforms the baseline Univfd by " + }, + { + "bbox": [ + 55, + 341, + 296, + 449 + ], + "type": "inline_equation", + "content": "3.22\\%" + }, + { + "bbox": [ + 55, + 341, + 296, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 341, + 296, + 449 + ], + "type": "inline_equation", + "content": "8.48\\%" + }, + { + "bbox": [ + 55, + 341, + 296, + 449 + ], + "type": "text", + "content": ". Compared with advanced methods, our method achieved improvements of " + }, + { + "bbox": [ + 55, + 341, + 296, + 449 + ], + "type": "inline_equation", + "content": "10.7\\%" + }, + { + "bbox": [ + 55, + 341, + 296, + 449 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 341, + 296, + 449 + ], + "type": "inline_equation", + "content": "23.08\\%" + }, + { + "bbox": [ + 55, + 341, + 296, + 449 + ], + "type": "text", + "content": " in average precision. These results further validate the effectiveness of our proposed method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 456, + 284, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 284, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 284, + 468 + ], + "type": "text", + "content": "5.4. Effect of VIB-Net across different backbone" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 473, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 653 + ], + "type": "text", + "content": "So far, we have seen the impressive performance of VIB-Net with the CLIP:ViT-L/14 backbone. In this section, we will explore the effect of transferring the VIB component to other multimodal models. We consider the following backbone variants to form VIB-Net: (1) FLAVA-FULL[35] (2) CLIP:ResNet50 (3) CLIP:ViT-B/16 (4) CLIP:ViT-L/14[30]. For each VIB-Net variant, we train using images generated by Stable Diffusion v1.4 and corresponding real images. Fig. 5 show the AP and ACC results of these variants on the same generative model. We observe that among these multimodal models, ViT-L/14 achieves the best performance, which can be attributed to its larger architecture and attention-based global receptive field. More importantly, all these variants exhibit strong performance, which can be attributed to the unique capabilities of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 660, + 200, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 200, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 200, + 673 + ], + "type": "text", + "content": "5.5. Analysis of latent features" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Finally, we apply our method to address the issues analyzed in Sec. 3.2. In Sec. 3.2, we identified that some irrelevant features in CLIP's general features could affect the classi-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 281, + 555, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 425 + ], + "type": "text", + "content": "fier's decision boundary, causing the model to learn superficial biases, which in turn impacts its generalization performance. As a solution, we proposed VIB-Net. Following the same experimental setup as in Sec. 3.2, we visualized the latent features extracted by VIB-Net using UMAP, and obtained encouraging results, as shown in Fig. 3b. We found that the latent features no longer exhibit obvious irrelevant features related to the real/fake image classification task, as the two separated clusters now belong to the same category (real or generated), unlike in Fig. 3a, where features of the same category were significantly separated. This demonstrates the advantage of our method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 433, + 388, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 388, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 388, + 445 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 453, + 556, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 556, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 556, + 596 + ], + "type": "text", + "content": "We propose a universal real-fake image detection method, VIB-Net. First, we found that the general features extracted by current methods based on large-scale pre-trained models contain irrelevant features that are unrelated to the task of distinguishing real from fake images. Experiments have shown that these features not only affect classification decisions, but may also cause the model to learn superficial biases, thereby damaging its generalization performance. As an alternative, we use VIB to remove task-irrelevant features, aiming to obtain the most concise image representations. These simplified features are maximally correlated with the real-fake classification task." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 605, + 417, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 417, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 417, + 619 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 624, + 556, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 556, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 556, + 703 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62376046 and Grant 62172067, and in part by the Natural Science Foundation of Chongqing under Grant CSTB2023NSCQ-MSX0341, the Science and Technology Research Program of Chongqing Municipal Education Commission (Grant No. KJQN202200635)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23835" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "type": "text", + "content": "[1] Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck. arXiv preprint arXiv:1612.00410, 2016. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 158 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 158 + ], + "type": "text", + "content": "[2] Alexander A Alemi, Ian Fischer, and Joshua V Dillon. Uncertainty in the variational information bottleneck. arXiv preprint arXiv:1807.00906, 2018. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 160, + 295, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 160, + 295, + 203 + ], + "spans": [ + { + "bbox": [ + 62, + 160, + 295, + 203 + ], + "type": "text", + "content": "[3] Xiuli Bi, Bo Liu, Fan Yang, Bin Xiao, Weisheng Li, Gao Huang, and Pamela C Cosman. Detecting generated images by real images only. arXiv preprint arXiv:2311.00962, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 205, + 295, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 295, + 270 + ], + "type": "text", + "content": "[4] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 103-120. Springer, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 273, + 295, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 273, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 62, + 273, + 295, + 337 + ], + "type": "text", + "content": "[5] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8789-8797, 2018. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 341, + 296, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 341, + 296, + 395 + ], + "spans": [ + { + "bbox": [ + 62, + 341, + 296, + 395 + ], + "type": "text", + "content": "[6] Davide Cozzolino, Giovanni Poggi, Riccardo Corvi, Matthias Nießner, and Luisa Verdoliva. Raising the bar of ai-generated image detection with clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4356-4366, 2024. 1, 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 397, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 397, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 397, + 295, + 430 + ], + "type": "text", + "content": "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 431, + 295, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 431, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 431, + 295, + 485 + ], + "type": "text", + "content": "[8] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In International conference on machine learning, pages 3247-3258. PMLR, 2020. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 488, + 295, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 295, + 541 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 295, + 541 + ], + "type": "text", + "content": "[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2014. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 544, + 295, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 295, + 599 + ], + "type": "text", + "content": "[10] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10696-10706, 2022. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 601, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 295, + 644 + ], + "type": "text", + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 678 + ], + "type": "text", + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[13] Shizhe Hu, Zhengzheng Lou, Xiaoqiang Yan, and Yangdong Ye. A survey on information bottleneck. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[14] Yan Ju, Shan Jia, Lipeng Ke, Hongfei Xue, Koki Nagano, and Siwei Lyu. Fusing global and local features for generalized ai-synthesized image detection. In 2022 IEEE International Conference on Image Processing (ICIP), pages 3465-3469. IEEE, 2022. 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 129, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 161 + ], + "type": "text", + "content": "[15] Tero Karras. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 1, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 163, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 207 + ], + "type": "text", + "content": "[16] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "type": "text", + "content": "[17] Sohail Ahmed Khan and Duc-Tien Dang-Nguyen. Clipping the deception: Adapting vision-language models for universal deepfake detection. In Proceedings of the 2024 International Conference on Multimedia Retrieval, pages 1006-1015, 2024. 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 264, + 553, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 286 + ], + "type": "text", + "content": "[18] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 288, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 553, + 319 + ], + "type": "text", + "content": "[19] Li Lin, Irene Amerini, Xin Wang, Shu Hu, et al. Robust clip-based detector for exposing diffusion model-generated images. arXiv preprint arXiv:2404.12908, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 322, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 322, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 317, + 322, + 553, + 388 + ], + "type": "text", + "content": "[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 389, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 389, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 317, + 389, + 553, + 431 + ], + "type": "text", + "content": "[21] Bo Liu, Fan Yang, Xiuli Bi, Bin Xiao, Weisheng Li, and Xinbo Gao. Detecting generated images by real images. In European Conference on Computer Vision, pages 95-110. Springer, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 434, + 555, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 555, + 476 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 555, + 476 + ], + "type": "text", + "content": "[22] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 479, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 479, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 479, + 555, + 533 + ], + "type": "text", + "content": "[23] Yunpeng Luo, Junlong Du, Ke Yan, and Shouhong Ding. Lare^ 2: Latent reconstruction error based method for diffusion-generated image detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17006-17015, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 535, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 577 + ], + "type": "text", + "content": "[24] Rabeeh Karimi Mahabadi, Yonatan Belinkov, and James Henderson. Variational information bottleneck for effective low-resource fine-tuning. arXiv preprint arXiv:2106.05469, 2021.3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 580, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 553, + 624 + ], + "type": "text", + "content": "[25] Francesco Marra, Diego Gragnaniello, Luisa Verdoliva, and Giovanni Poggi. Do gans leave artificial fingerprints? In 2019 IEEE conference on multimedia information processing and retrieval (MIPR), pages 506-511. IEEE, 2019. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 625, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 657 + ], + "type": "text", + "content": "[26] Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "text", + "content": "[27] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 5" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23836" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[28] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24480-24489, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "text", + "content": "[29] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2337-2346, 2019. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 186, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 251 + ], + "type": "text", + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "type": "text", + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 309, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 365 + ], + "type": "text", + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 365, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 421 + ], + "type": "text", + "content": "[33] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics++: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1-11, 2019. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 422, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 477 + ], + "type": "text", + "content": "[34] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 478, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 544 + ], + "type": "text", + "content": "[35] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15638-15650, 2022. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 545, + 294, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 611 + ], + "type": "text", + "content": "[36] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, and Yunchao Wei. Learning on gradients: Generalized artifacts representation for gan-generated images detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12105-12114, 2023. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 613, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 678 + ], + "type": "text", + "content": "[37] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28130-28139, 2024. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "type": "text", + "content": "[38] Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. 3" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 475 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 129 + ], + "type": "text", + "content": "[39] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8695-8704, 2020. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 129, + 554, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 184 + ], + "type": "text", + "content": "[40] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22445-22455, 2023. 1, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 185, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 554, + 217 + ], + "type": "text", + "content": "[41] Among Wu and Cheng Deng. Tib: Detecting unknown objects via two-stream information bottleneck. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 219, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 554, + 262 + ], + "type": "text", + "content": "[42] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "type": "text", + "content": "[43] Ning Yu, Larry S Davis, and Mario Fritz. Attributing fake images to gans: Learning and analyzing gan fingerprints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7556-7566, 2019. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 308, + 554, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 554, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 554, + 363 + ], + "type": "text", + "content": "[44] Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deepfake detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2185–2194, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 364, + 554, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 364, + 554, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 364, + 554, + 418 + ], + "type": "text", + "content": "[45] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 420, + 554, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 420, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 420, + 554, + 475 + ], + "type": "text", + "content": "[46] Mingjian Zhu, Hanting Chen, Qiangyu Yan, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang. Genimage: A million-scale benchmark for detecting ai-generated image. Advances in Neural Information Processing Systems, 36, 2024. 5" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23837" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Towards Universal Dataset Distillation via Task-Driven Diffusion/cd8b4c9e-1814-423f-bdad-2b631dc6a75a_content_list.json b/2025/Towards Universal Dataset Distillation via Task-Driven Diffusion/cd8b4c9e-1814-423f-bdad-2b631dc6a75a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c22c5f15f40a9842dd171e7e5149caeba65164e9 --- /dev/null +++ b/2025/Towards Universal Dataset Distillation via Task-Driven Diffusion/cd8b4c9e-1814-423f-bdad-2b631dc6a75a_content_list.json @@ -0,0 +1,1622 @@ +[ + { + "type": "text", + "text": "Towards Universal Dataset Distillation via Task-Driven Diffusion", + "text_level": 1, + "bbox": [ + 169, + 130, + 826, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ding Qi $^{1*}$ Jian Li $^{2\\dagger}$ Junyao Gao $^{1}$ Shuguang Dou $^{1}$ Ying Tai $^{5}$ \nJianlong Hu $^{2}$ Bo Zhao $^{4}$ Yabiao Wang $^{3,2\\dagger}$ Chengjie Wang $^{4,2}$ Cairong Zhao $^{1\\dagger}$ $^{1}$ Tongji University $^{2}$ Tencent YouTu Lab $^{3}$ Zhejiang University \n $^{4}$ Shanghai Jiao Tong University $^{5}$ Nanjing University", + "bbox": [ + 161, + 178, + 833, + 253 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/273080066be2f11e90e6d1d4791effb0af3c1dd2d9a9a35dc7bff1253d4caa6d.jpg", + "image_caption": [ + "(a) One Universal Dataset Distillation Framework for Mutli-Tasks" + ], + "image_footnote": [], + "bbox": [ + 124, + 268, + 598, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/29377090049304128148076e6e350a1f9d66a781866c0624fbfc3fec438bfb75.jpg", + "image_caption": [ + "(b) Improved Performance by UniDD", + "Figure 1. (a) We introduce the Universal Dataset Distillation Framework (UniDD), a novel dataset distillation framework supporting a range of vision tasks, including classification, object detection, and segmentation. UniDD incorporates diverse label types, such as class labels, bounding box coordinates, and pixel-level masks, facilitating the synthesis of task-specific datasets. (b) UniDD improves the performance of dataset distillation across diverse tasks on ImageNet [8], Pascal VOC [11, 12], and MS COCO [28]." + ], + "image_footnote": [], + "bbox": [ + 629, + 279, + 867, + 438 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 563, + 326, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dataset distillation (DD) condenses key information from large-scale datasets into smaller synthetic datasets, reducing storage and computational costs for training networks. However, most recent research has primarily focused on image classification tasks, with limited exploration in detection and segmentation. Two key challenges remain: (i) Task Optimization Heterogeneity, where existing methods focus on class-level information but fail to address the diverse needs of detection and segmentation, and (ii) Inflexible Image Generation, where current generation methods rely on global updates for single-class targets and lack localized optimization for specific object regions. To address these challenges, we propose UniDD, a universal dataset distillation framework built on a task-driven diffusion model for diverse DD tasks, as shown in Fig. 1. Our approach operates in two stages: Universal Task Knowledge Mining, which captures task-relevant information through task", + "bbox": [ + 88, + 595, + 485, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "specific proxy model training, and Universal Task-Driven Diffusion, where these proxies guide the diffusion process to generate task-specific synthetic images. Extensive experiments across ImageNet-1K, Pascal VOC, and MS COCO demonstrate that UniDD consistently outperforms state-of-the-art methods. In particular, on ImageNet-1K with IPC-10, UniDD surpasses previous diffusion-based methods by $6.1\\%$ , while also reducing deployment costs.", + "bbox": [ + 511, + 564, + 908, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 720, + 645, + 736 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dataset distillation (DD) [44], as an innovative data optimization technique in machine learning [13, 14, 20, 22, 45], condenses essential information from large datasets into smaller proxy sets. Training on these highly ensemble proxy sets enables the model to use much fewer resources while providing comparable performance. Prior DD methods excelled in image classification using various techniques like Gradient Matching [26, 48, 50], Distribution Matching [49, 51], Training Trajectory Matching [3, 7], and Decoupled Methods [42, 46].", + "bbox": [ + 509, + 750, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done when Ding Qi is an intern in Tencent YouTu Lab, and the advisor is Jian Li.", + "bbox": [ + 89, + 862, + 482, + 886 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 107, + 887, + 246, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10557", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite advancements in DD, research on key visual tasks—detection and segmentation—remains limited. We identify the following key challenges: (i) Task Optimization Heterogeneity. Existing DD methods are primarily focused on image classification, optimizing distillation by compressing class-level information, but they fall short in addressing the diverse needs of tasks like detection [27] and segmentation, such as object localization and pixel-level information. (ii) Inflexible Image Generation. Current image generation methods rely on global updates for single-class targets and lack localized optimization for specific object regions. These limitations become especially evident in images with multiple objects or complex structures, hindering the ability of existing DD to generate distilled data that meets the requirements of diverse tasks. Consequently, the applicability of DD in the visual domain remains constrained.", + "bbox": [ + 89, + 90, + 480, + 345 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these limitations, we propose a Universal Dataset Distillation (UniDD) framework that optimizes for multi-task learning, including classification, detection, and segmentation, overcoming the shortcomings in task adaptation and image generation. Our approach consists of two stages: Universal Task Knowledge Mining and Universal Task-Driven Diffusion. In the first stage, we introduce Task-Specific Proxy $(\\mathcal{T}_{\\mathcal{SP}})$ models, trained on real datasets to extract task-specific information. This allows the $\\mathcal{T}_{\\mathcal{SP}}$ models to capture and store category information for classification, bounding box positions for detection, and mask data for segmentation within their parameters. Additionally, previous research has often overlooked the importance of contextual information, particularly in high-resolution datasets like ImageNet [8], Pascal VOC [11, 12] and MS COCO [28], where rich object-context relationships exist. To address this, we introduce Task-Aligned Contextual Prompting, which uses a Vision-Language model [33] to generate contextual prompts for real images. In the second stage, we introduce a Task-Driven Diffusion for image synthesis. Specifically, the $\\mathcal{T}_{\\mathcal{SP}}$ models trained in the first stage supervise the denoising process of the diffusion model by calculating the task loss on the predicted clean images. This guides the diffusion model in generating images that match the given labels. Notably, since the $\\mathcal{T}_{\\mathcal{SP}}$ models already align with the target domain, there is no need to retrain the diffusion model on the real set, significantly reducing the cost in practical applications.", + "bbox": [ + 91, + 351, + 483, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are threefold:", + "bbox": [ + 109, + 777, + 403, + 791 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We identify key challenges in extending dataset distillation beyond classification and propose the first universal framework for DD across classification, detection, and segmentation tasks.", + "- We develop a two-stage approach, Universal Task Knowledge Mining and Universal Task-Driven Diffusion for Data Synthesis, which flexibly synthesizes images for" + ], + "bbox": [ + 91, + 794, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "multi-tasks. Additionally, we address the previously overlooked need to align potential contextual information within datasets.", + "bbox": [ + 526, + 90, + 903, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Our method achieves state-of-the-art performance in classification tasks, on the ImageNet-1K dataset, our approach surpasses the previous state-of-the-art method by over $6.1\\%$ . Additionally, we establish new benchmarks for detection and segmentation on VOC and COCO, showing significant improvements over coreset methods.", + "bbox": [ + 513, + 136, + 903, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 513, + 242, + 660, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Dataset Distillation", + "text_level": 1, + "bbox": [ + 513, + 268, + 697, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dataset distillation [44], a dataset reduction method, addresses the challenge of handling substantial data by synthesizing a smaller, representative dataset. Existing dataset distillation methods can be categorized into meta-learning and data matching frameworks based on whether they explicitly mimic the performance of target data. In the meta-learning framework, the distilled data are treated as hyperparameters and optimized in a nested loop fashion according to the trained model's risk. For example, KIP [32] performs kernel ridge regression using the Neural Tangent Kernel (NTK) [25]. The data matching framework includes gradient matching, distribution matching, and trajectory matching. These methods update distilled data by imitating the influence of target data. For example, DC [50] and DM [49] enforce synthetic and real images to have similar gradients or distribution at every training step. MTT [3] proposes to imitate long-range trajectories of optimization steps for real images, thereby better mimicking the training dynamics of real images for synthetic images. However, the aforementioned gradient matching methods require a large number of iterations. To address this, DREAM [31] and IDC [26] focus on more efficient dataset distillation training. Due to the unsatisfactory performance of DM, CAFE [43] and IDM [51] propose improved distribution matching between the real and synthetic data. For the trajectory matching method, DATM [19] scales trajectory matching-based methods to larger synthetic datasets, achieving lossless dataset distillation.", + "bbox": [ + 511, + 291, + 906, + 714 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite significant progress in previous work, there have been few breakthroughs in dataset distillation for detection and segmentation tasks. In this work, we introduce the first universal framework that simultaneously handles classification, object detection, and semantic segmentation, significantly expanding the generality and scope of dataset distillation research.", + "bbox": [ + 511, + 715, + 905, + 819 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Diffusion Models", + "text_level": 1, + "bbox": [ + 513, + 832, + 681, + 847 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models [15, 16, 37, 47] are generative models that show superior fidelity and diversity compared to GANs [17]. It aims to learn the mapping from a Gaussian", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10558", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8c94ce46423189613163e117cc42deebceb0535c1c09b9fdc51f7a33944a2e6a.jpg", + "image_caption": [ + "(a) Synthetic images for classification." + ], + "image_footnote": [], + "bbox": [ + 94, + 90, + 354, + 287 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6a727fa95fd69b053327cd236578d7e039dc29ffce88b311fa0bedf52568970d.jpg", + "image_caption": [ + "(b) Synthetic images for detection." + ], + "image_footnote": [], + "bbox": [ + 367, + 90, + 629, + 287 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b749891040b76d77d0cb4ecba985295e6e5ca13b40b01681cb87fd578d07b433.jpg", + "image_caption": [ + "(c) Synthetic images for segmentation.", + "Figure 2. Visualization of synthetic images generated by UniDD. (a) For classification, images are synthesized on the ImageWoof [8] dataset. (b) For detection, images are synthesized on the Pascal VOC dataset with specified bounding box coordinates. (c) For segmentation, images are synthesized on the Pascal VOC [11, 12] dataset with pixel-level masks. Our UniDD is the first approach to enable high-realism data distillation across classification, detection, and segmentation tasks." + ], + "image_footnote": [], + "bbox": [ + 642, + 90, + 903, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "distribution to the real dataset distribution by adding noise in the diffusion process and reversing the noise to the raw image in the denoising process. In the forward process, diffusion models progressively add noise to a clean sample $z_0$ , resulting in a noisy sample $z_t$ :", + "bbox": [ + 89, + 378, + 483, + 455 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t} = \\sqrt {\\alpha_ {t}} z _ {0} + \\sqrt {1 - \\alpha_ {t}} \\epsilon , \\quad \\epsilon \\sim \\mathcal {N} (0, I). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 460, + 482, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the reverse process, a denoising network $\\epsilon_{\\theta}$ estimates the noise and iteratively recovers the clean sample $z_0$ :", + "bbox": [ + 89, + 484, + 483, + 513 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {\\theta} (z _ {t}, t) \\approx \\frac {z _ {t} - \\sqrt {\\alpha_ {t}} z _ {0}}{\\sqrt {1 - \\alpha_ {t}}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 518, + 482, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DDIM [40] provides the following formula for predicting a clean data sample:", + "bbox": [ + 89, + 556, + 483, + 587 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {0} = \\frac {z _ {t} - \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {\\theta} (z _ {t} , t)}{\\sqrt {\\alpha_ {t}}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 590, + 482, + 627 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This formula predicts the clean data point $\\hat{z}_0$ from the noisy sample $z_{t}$ , enabling deterministic sampling for efficient and high-quality image synthesis.", + "bbox": [ + 89, + 631, + 483, + 675 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To introduce additional conditions into the diffusion model, classifier guidance [9] has been proposed to train an additional classifier on the noise to guide the model to generate images that are closer to the specified condition. Additionally, [23] introduces classifier-free guidance that trains both an unconditioned and conditioned model simultaneously, achieving more efficient guidance. Following these, recent methods [1, 10, 30] have been proposed to generate high-quality images from diverse conditions.", + "bbox": [ + 89, + 676, + 483, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Universal Dataset Distillation Framework", + "text_level": 1, + "bbox": [ + 89, + 823, + 465, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Definition", + "text_level": 1, + "bbox": [ + 89, + 848, + 272, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Universal Dataset Distillation. Current dataset distillation research primarily targets classification tasks and lacks", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a generalizable approach for detection and segmentation. To address this, we propose a universal dataset distillation standard that constructs a compact synthetic dataset $\\mathcal{S} = (\\hat{x}_1, \\hat{y}_1), \\dots, (\\hat{x}_{|\\mathcal{S}|}, \\hat{y}_{|\\mathcal{S}|})$ from a larger real dataset $\\mathcal{T} = (x_1, y_1), \\dots, (x_{|\\mathcal{T}|}, y_{|\\mathcal{T}|})$ , with $|S| \\ll |\\mathcal{T}|$ . The primary difference between tasks lies in label structure: classification uses single-class labels, detection includes bounding boxes with class information, and segmentation requires pixel-level masks. Our goal is for models trained on the distilled dataset $\\mathcal{S}$ to perform comparably to those trained on the full dataset $\\mathcal{T}$ across classification, detection, and segmentation tasks, as formalized by:", + "bbox": [ + 511, + 378, + 906, + 561 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {(x, y) \\sim P _ {D}} \\left| \\ell \\left(\\theta_ {\\mathcal {T}} ^ {t a s k} (x), y\\right) - \\ell \\left(\\theta_ {\\mathcal {S}} ^ {t a s k} (x), y\\right) \\right| \\leq \\epsilon , \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 570, + 906, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $P_{D}$ denotes the distribution of test data, $\\ell$ represents the performance metric specific to each task, $task\\in$ {class,detect,segment}, and $\\epsilon$ is a small tolerance value. For each task, the parameter set $\\theta_{\\mathcal{T}}^{task}$ on the real dataset $\\mathcal{T}$ is optimized as follows:", + "bbox": [ + 511, + 598, + 906, + 674 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {\\mathcal {T}} ^ {t a s k} = \\arg \\min _ {\\theta} \\mathbb {E} _ {(x, y) \\in \\mathcal {T}} \\left[ \\ell \\left(\\phi_ {\\theta} ^ {t a s k} (x), y\\right) \\right], \\qquad (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 683, + 906, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similarly, $\\theta_S^{task}$ is optimized using the synthetic dataset $S$ .", + "bbox": [ + 511, + 717, + 898, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Universal Task Knowledge Mining", + "text_level": 1, + "bbox": [ + 511, + 742, + 815, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Task-Specific Proxy Training. Building a universal DD framework is challenging due to the diverse optimization objectives in classification, detection, and segmentation tasks. Recent decoupled DD studies have reached a consensus [34, 42, 46]: a fully trained network can capture and retain essential dataset information. Inspired by this, we extend to general tasks, training a classifier, detector, and segmenter on task-specific datasets to extract critical information—class labels for classification, bounding boxes for", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10559", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/697bfc9651c9aab0d9e97c88dab77dad9859a6862ec3e7f7180dc844b1d8405e.jpg", + "image_caption": [ + "Figure 3. Universal Dataset Distillation Framework. This framework consists of two stages: in the Universal Task Knowledge Mining stage, we introduce an easily deployable task-specific proxy $(\\mathcal{T}_{\\mathcal{SP}})$ model trained on the target dataset and a vision-language model to extract task-aligned context. In the Universal Task-Driven Diffusion stage, we apply the gradient of the supervision loss, derived from the $\\mathcal{T}_{\\mathcal{SP}}$ model's prediction of the clean image $\\hat{z}_0$ , at each sampling step." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 906, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "detection, and pixel-level masks for segmentation. We call this Task-Specific Proxy $(\\mathcal{T}_{\\mathcal{SP}})$ Model, defined as:", + "bbox": [ + 89, + 446, + 482, + 477 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta = \\arg \\min _ {\\theta} \\mathcal {L} _ {\\text {t a s k}} (\\mathcal {F} (x), y), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 489, + 482, + 510 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{F}$ represents the $\\mathcal{T}_{\\mathcal{SP}}$ model trained on the target data, $\\theta$ denotes the model's parameters, $\\mathcal{L}_{task}$ represents the loss function specific to the task.", + "bbox": [ + 89, + 522, + 482, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The $\\mathcal{T}_{\\mathcal{SP}}$ model choice is flexible, allowing adaptation for optimal performance on target datasets. For example, ResNet-18 [21] can serve as a lightweight, fast-deploy option for simple classification, while Faster R-CNN [36] may be used for complex detection tasks to capture key detection information. Unlike previous diffusion-based dataset distillation methods, our approach relies on task-driven learning to align synthetic and real datasets, avoiding costly diffusion model fine-tuning on the target dataset and significantly reducing deployment overhead.", + "bbox": [ + 89, + 568, + 482, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Task-Aligned Contextual Prompting. Previous diffusion-based dataset distillation has largely focused on classification tasks and overlooked the role of natural language in guiding image generation. For diverse tasks, textual prompts must go beyond describing the primary objects in images. As inputs to Stable Diffusion models, prompts can also guide the generation process. To address this, we use a vision-language model to extract natural language descriptions and design task-specific questions to generate task-aligned prompts, a method we call Task-Aligned Contextual Prompting. Following GPT-4o [33], we create tailored prompts for different tasks. For classification, prompts de", + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "scribe object classes, while for detection and segmentation, they also include object relationships and the interaction between objects and the background, providing richer contextual guidance.", + "bbox": [ + 511, + 446, + 906, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Universal Task-Driven Diffusion.", + "text_level": 1, + "bbox": [ + 511, + 516, + 802, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by recent advancements in guidance diffusion techniques [1, 9], we propose a Task-Driven Diffusion approach for Dataset Distillation. As illustrated in Fig. 3, during each sampling step $S(z_{t}, t)$ of the diffusion process, the trained $\\mathcal{T}_{\\mathcal{SP}}$ model $\\mathcal{F}$ provides task-specific guidance. Given an label $\\tilde{a}$ , the original noise prediction $\\epsilon_{\\theta}(z_{t}, t, \\tilde{c})$ is adjusted using the output of $\\mathcal{F}$ . The adjusted noise prediction $\\hat{\\epsilon}$ is formulated as follows:", + "bbox": [ + 511, + 537, + 906, + 657 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\epsilon} _ {\\theta} \\left(z _ {t}, t, \\tilde {c}\\right) = \\epsilon_ {\\theta} \\left(z _ {t}, t, \\tilde {c}\\right) + \\sqrt {1 - \\alpha_ {t}} \\nabla_ {z _ {t}} \\mathcal {L} _ {\\text {t a s k}} \\left(\\tilde {a}, \\mathcal {F} \\left(z _ {t}\\right)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 667, + 906, + 685 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{task}}$ represents the task-specific loss function and $\\tilde{a} = \\{a_{cl}, a_{bbox}, a_{mask}\\}$ may denote a class label $a_{cl}$ , a bounding box $a_{bbox}$ , or a pixel-level segmentation mask $a_{mask}$ . This adjustment mechanism effectively guides the diffusion process to produce data closely aligned with the intended distribution for each task. $\\tilde{c}$ is a textual prompt.", + "bbox": [ + 511, + 694, + 905, + 784 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Because $\\mathcal{F}$ is trained on clean images, it may not perform well directly on noisy inputs. Therefore, we first use Equation 3 to estimate the clean image $\\hat{z}_0$ . Once $\\hat{z}_0$ is obtained, we apply $\\mathcal{F}$ to compute the task-driven loss as follows:", + "bbox": [ + 511, + 785, + 905, + 845 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\epsilon} _ {\\theta} \\left(z _ {t}, t, \\tilde {c}\\right)) = \\epsilon_ {\\theta} \\left(z _ {t}, t, \\tilde {c}\\right) + s (t) \\cdot \\nabla_ {z _ {t}} \\mathcal {L} _ {\\text {t a s k}} (\\tilde {a}, \\mathcal {F} (\\hat {z} _ {0})), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 857, + 906, + 875 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $s(t)$ is the task-driven strength at each sampling step,", + "bbox": [ + 511, + 885, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10560", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "defined as:", + "bbox": [ + 89, + 90, + 166, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns (t) = \\lambda_ {s} \\cdot \\sqrt {1 - \\alpha_ {t}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 104, + 480, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $\\lambda_{s}$ being a constant coefficient controlling the overall intensity of the task-driven. The term $\\nabla_{z_t}\\mathcal{L}_{task}(\\tilde{a},\\mathcal{F}(\\hat{z}_0))$ represents the gradient of the loss function computed based on the $\\mathcal{T}_{\\mathcal{SP}}$ model's prediction of the clean image $\\hat{z}_0$ . Specifically, this gradient is computed as:", + "bbox": [ + 89, + 128, + 483, + 204 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {z _ {t}} \\mathcal {L} _ {\\text {t a s k}} (\\tilde {a}, \\mathcal {F} (\\hat {z} _ {0})) = \\nabla_ {z _ {t}} \\mathcal {L} _ {\\text {t a s k}} \\left(\\tilde {a}, \\mathcal {F} \\left(\\frac {z _ {t} - \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {\\theta}}{\\sqrt {\\alpha_ {t}}}\\right)\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 213, + 483, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike previous approaches [18, 41] that required finetuning the diffusion model, such as those involving SD [37], to align with new data domains, our method leverages a $\\mathcal{T}_{\\mathcal{SP}}$ model trained specifically on the target dataset. This model encapsulates critical information about the target task, effectively driving the diffusion process to generate data that is well-aligned with the target task without the need for additional finetuning. This not only conserves time and resources but also mitigates the risk of overfitting the diffusion model to the specific task.", + "bbox": [ + 89, + 267, + 483, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proxy-Driven High-Realism Refinement. To ensure the high realism and accuracy of the generated images, we employ a trained $\\mathcal{T}_{\\mathcal{SP}}$ model, to score each image. By setting a well-defined threshold, we filter out low-quality images based on their scores. The remaining images are then relabeled using the same $\\mathcal{T}_{\\mathcal{SP}}$ model. This combined process of filtering and relabeling can be expressed as:", + "bbox": [ + 89, + 417, + 483, + 525 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\{x, y \\} = \\left\\{x _ {i}, \\mathcal {F} _ {\\text {l a b e l}} \\left(x _ {i}\\right) \\mid \\operatorname {s c o r e} \\left(\\mathcal {F} \\left(x _ {i}\\right)\\right) \\geq \\tau \\right\\} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 535, + 482, + 554 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{F}(x_i)$ represents the model's output score for image $x_{i}$ , $\\tau$ is the quality threshold, and $\\mathcal{F}_{\\mathrm{label}}$ generates the accurate label $y_{i}$ for the filtered image. This ensures that both the generated images and their labels are of high quality and closely match the desired data distribution.", + "bbox": [ + 89, + 564, + 483, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 642, + 223, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setups", + "text_level": 1, + "bbox": [ + 89, + 667, + 287, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and Evaluation Metric. To enhance practicality, we prioritize high-resolution and large-scale datasets. (1) Classification task: our experiments are primarily conducted on the full ImageNet 1K dataset [8], with IPC settings commonly set to 10 and 50, following the configurations in [41]. Additionally, based on the dataset partitioning settings from previous work [3], we also use the ImageNet 10-class subsets, such as ImageWoof, ImageFruit, ImageNette, and ImageMeow. These 10-class subsets contain very similar categories, making them challenging for dataset distillation. Top-1 accuracy $(\\%)$ is reported in the experimental tables. (2) Object detection task: we conduct experiments on the widely-used Pascal VOC [11, 12] and MS COCO [28] datasets, using mAP (Mean Average", + "bbox": [ + 89, + 688, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 Universal Dataset Distillation Approach", + "text_level": 1, + "bbox": [ + 514, + 90, + 870, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: Real data $\\mathcal{T}$ , Diffusion model $\\epsilon_{\\theta}$ , $\\mathcal{T}_{\\mathcal{SP}}$ model $\\mathcal{F}$ , Sampling steps $T$ , Task-Driven strength $s(t)$", + "bbox": [ + 513, + 107, + 905, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: Synthetic dataset $S$ .", + "bbox": [ + 514, + 138, + 710, + 152 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Step I: Universal Task Knowledge Mining", + "2: Train $\\mathcal{T}_{\\mathcal{SP}}$ model $\\mathcal{F}$ on $\\mathcal{T}$ using Equ. 6", + "3: Use GPT-4o to generate task-aligned context for classification, detection, and segmentation datasets" + ], + "bbox": [ + 522, + 155, + 903, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4: Step II: Universal Task-Driven Diffusion for Data Synthetic", + "bbox": [ + 522, + 215, + 903, + 244 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "5: for each sampling step $t = 1$ to $T$ do", + "6: Generate noisy sample $z_{t}$ .", + "7: Predict clean image $\\hat{z}_0$ using Equ. 3", + "8: Calculate guidance loss $\\nabla_{z_t}\\mathcal{L}_{task}(c,\\mathcal{F}(\\hat{z}_0))$", + "9: Adjust noise prediction using Equ. 8", + "10: Update $z_{t-1}$ using adjusted noise prediction.", + "11: end for", + "12: Score each generated image $x_{i}$ using $\\mathcal{F}(x_i)$ .", + "13: Filter and relabel synthetic set based on threshold $\\tau$ using Equ. 11." + ], + "bbox": [ + 519, + 246, + 903, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Precision) and $AP_{50}$ (Average Precision at IoU threshold of 0.50) as evaluation metric. (3) Semantic segmentation task: we utilize the Pascal VOC dataset, with mIoU (Mean Intersection over Union) as evaluation metrics.", + "bbox": [ + 511, + 428, + 905, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Unless otherwise specified, the Task-Specific Proxy $\\mathcal{T}_{\\mathcal{SP}}$ model used for the classification task is ResNet-18 [21], the detector is Faster R-CNN [36], and the segmentation model is LR-ASPP [24], all utilizing the official pre-trained weights from PyTorch. For the diffusion model, we use SD v1.5 [37] as the baseline. For quantitative comparisons, we generate $256 \\times 256$ images for classification tasks and $512 \\times 512$ images for detection and segmentation tasks. Synthetic images are created using 250 denoising steps based on a predefined number of random noise samples, aligned with the IPC settings. The threshold $\\tau$ for Proxy-Driven High-Realism Refinement is set to 0.9 for classification, 0.85 for detection, and 0.75 for segmentation. Generally, this operation results in a data discard rate of approximately $5 - 10\\%$ . All experiments are conducted on NVIDIA V100 GPUs, with further details provided in the special experiments section. We generate the distilled dataset 3 times, train the network from scratch 5 times per dataset (with pre-trained backbone weights for the detector and segmenter), and evaluate on the original test set, reporting results as $\\bar{x} \\pm std$ .", + "bbox": [ + 511, + 489, + 906, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Results on Classification", + "text_level": 1, + "bbox": [ + 511, + 816, + 736, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We perform comprehensive comparisons across datasets of various scales. For large-scale datasets ImageNet-1k, we compare TESLA [7], SRe $^2$ L [46], D $^4$ M [41], RDED [42], and MiniMax Diffusion [18] (abbreviated as MiMxDiff in", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10561", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/af2b02f0dee62e5c55deefb8ab940f3c5d2775ef44dd83859d3426b594a6e0c0.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Performance comparison with state-of-the-art methods on large-scale dataset ImageNet-1K. All methods employ ResNet18 as test model. All standard deviations in this table are less than one. The results of Full dataset come from the official PyTorch. The best results are marked in bold, while the second-best results are underlined." + ], + "table_body": "
DatasetIPCMethodAccuracy(%)
ImageNet-1K1000Full dataset69.8
10TESLA7.7
SRe2L21.3
D4M27.9
RDED42.0
MiMxDiff44.3
UniDD (Ours)50.4 (↑6.1)
50SRe2L46.8
D4M55.2
RDED56.5
MiMxDiff58.6
UniDD (Ours)62.8 (↑4.2)
", + "bbox": [ + 93, + 88, + 480, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the table). For small-scale dataset ImageNet-Subset, we compare Random selection, DM [49], MTT [3], HaBa [29], DataDAM [38], and MiniMax Diffusion.", + "bbox": [ + 89, + 431, + 482, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative comparison on ImageNet-1K and ImageNet-Subset. As shown in the Tab 1, we utilize ResNet-18 as the $\\mathcal{T}_{\\mathcal{SP}}$ model, trained on ImageNet-1K. We report results for IPC-10 and IPC-50, evaluating all outcomes with the ResNet-18 model. To ensure a fair comparison, our training and testing pipeline is aligned with that of RDED and $\\mathrm{D^4M}$ . Our method, UniDD, achieves state-of-the-art performance on this large-scale dataset. Specifically, at IPC-10, UniDD surpasses optimization-based methods, delivering substantial gains over TESLA and $\\mathrm{SRe}^2\\mathrm{L}$ . Additionally, it demonstrates significant advantages over diffusion-based methods, outperforming $\\mathrm{D^4M}$ by $22.5\\%$ and MiniMax Diffusion by $6.1\\%$ . These results underscore UniDD's effectiveness by using a lightweight $\\mathcal{T}_{\\mathcal{SP}}$ model to guide SD denoising, aligning generative and target domains without requiring diffusion model finetuning, thus advancing performance in dataset distillation.", + "bbox": [ + 89, + 478, + 482, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in the Tab 2, we conduct experiments on the ImageNet-subset under the IPC-10 setting. For consistency, the test model is standardized to a 5-layer ConvNet across all methods, except for MiniMax Diffusion, which uses a 6-layer ConvNet. Our method stands out among the comparison methods, outperforming the second-best by $8.2\\%$ on ImageWoof, $9.1\\%$ on ImageFruit, and $6.6\\%$ on ImageMeow. Additionally, our method surpasses the diffusion-based MiniMax Diffusion by $9.8\\%$ on ImageWoof, further demonstrating the effectiveness of UniDD approach in ag", + "bbox": [ + 89, + 750, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "gregating information from the original dataset.", + "bbox": [ + 511, + 90, + 828, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Result. In Fig. 2 (a), we present the synthetic images generated by UniDD on ImageWoof. Benefiting from the powerful generative capability of SD and the use of a $\\mathcal{T}_{\\mathcal{SP}}$ model to align the real and synthetic data domains, our synthetic images for the classification task exhibit significantly higher realism compared to optimization-based methods [3, 46, 50]. However, unlike other diffusion-based dataset distillation methods [18, 41], we do not finetune SD, greatly reducing the deployment time of UniDD.", + "bbox": [ + 511, + 107, + 903, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Results on Detection and Segmentation", + "text_level": 1, + "bbox": [ + 511, + 256, + 848, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We deploy our UniDD framework for detection and segmentation tasks and conduct experiments on the Pascal VOC and MS COCO datasets. Since dataset distillation is rarely applied to these tasks, we reference earlier dataset distillation works (DD [44], DC [50]) and compare UniDD with core-set selection methods. Specifically, we construct the following baseline methods: Random [35], K-center [39], and Herding [2, 5]. Additionally, we develop a method that selects a subset with balanced class representation, which we refer to as Uniform.", + "bbox": [ + 511, + 279, + 903, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Result. In Table 3, we compare coreset selection methods with UniDD across different compression ratios on the object detection task. According to the official MMDetection implementation [4], the detector Faster R-CNN [36] achieves $51.4\\%$ mAP and $80.3\\%$ $\\mathrm{AP}_{50}$ on Pascal VOC with the full dataset, and $32.6\\%$ mAP and $51.4\\%$ $\\mathrm{AP}_{50}$ on MS COCO, providing an approximate performance upper bound. On Pascal VOC, UniDD demonstrates significant performance gains, achieving $7.7\\%$ mAP and $19.2\\%$ $\\mathrm{AP}_{50}$ higher than the Random method at an extremely low compression ratio of $0.5\\%$ . Additionally, UniDD surpasses Uniform by $11.1\\%$ mAP and $21.2\\%$ $\\mathrm{AP}_{50}$ at a $1\\%$ compression ratio. On the more challenging MS COCO dataset, UniDD also shows a clear advantage, surpassing the best coreset selection method by $3.7\\%$ , $3.4\\%$ , and $3.4\\%$ mAP at compression ratios of $0.25\\%$ , $0.5\\%$ , and $1\\%$ , respectively. This indicates UniDD's robustness, achieving strong results even on more complex datasets, which can be attributed to the Task Knowledge Mining in the first stage that extracts sufficient task-relevant information.", + "bbox": [ + 511, + 431, + 903, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 4, we present experimental results on the VOC dataset for the semantic segmentation task. Based on the MMSegmentation [6] implementation, LR-ASPP with a MobileNetV3 backbone achieves $32.9\\%$ mIoU on the full dataset. UniDD outperforms coreset selection methods at all compression ratios, demonstrating its effectiveness in dataset distillation for segmentation tasks. This result further underscores UniDD's versatility across different tasks. By leveraging task-driven diffusion synthesis alongside task knowledge mining, UniDD generates synthetic data well-suited to complex, pixel-level predictions, illustrating its", + "bbox": [ + 511, + 734, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10562", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9ac17681a769dc58ea4f70fb98a6be3e550ad19b2c698138fcc146c167facb21.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetRandomDMMTTHaBaDataDAMMiMxDiffUniDD (Ours)Full
ImageNette47.7±2.458.1±0.363.0 ±1.364.7 ±1.659.4±0.462.0±0.266.0±0.787.4 ±1.0
ImageWoof27.0±1.931.4±0.535.8 ±1.838.6 ±1.234.4±0.437.0±1.046.8±0.967.0 ±1.3
ImageFruit21.4±1.2-40.3 ±1.342.5 ±1.6--51.6±0.563.9 ±2.0
ImageMeow29.0±1.1-40.4 ±2.242.9 ±0.9--49.5±0.766.7 ±1.1
", + "bbox": [ + 99, + 88, + 895, + 189 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8ffe4f26155fc76f3ae3557195cd552a06d0dcbf656e869a985da23a576b5bcc.jpg", + "table_caption": [ + "Table 2. Performance comparison with other state-of-the-art methods on ImageNet-Subset. IPC is set to 10. The best results are marked in bold, while the second-best results are underlined. “-” were not officially implemented in the original papers." + ], + "table_footnote": [], + "table_body": "
MethodsObject Detection
Pascal VOCMS COCO
mAPAP50mAPAP50
Ratio0.5%0.25%
Random0.8±0.23.1±0.40.5±0.11.7±0.3
Uniform0.9±0.13.4±0.30.8±0.22.4±0.5
K-Center0.5±0.12.1±0.30.4±0.11.5±0.2
Herding0.6±0.22.4±0.20.5±0.11.8±0.4
UniDD (Ours)8.5±0.422.3±0.64.5±0.310.3±0.4
Ratio1%0.5%
Random4.2±0.513.7±0.63.7±0.210.1±0.3
Uniform5.7±0.217.7±0.43.4±0.49.5±0.6
K-Center3.6±0.612.3±0.33.2±0.59.3±0.5
Herding3.5±0.511.9±0.53.5±0.39.7±0.3
UniDD (Ours)16.8±0.538.9±0.77.1±0.416.9±0.3
Ratio2%1%
Random12.4±0.434.3±0.57.2±0.817.3±0.9
Uniform13.8±0.336.2±0.47.4±0.517.6±0.5
K-Center10.9±0.629.3±0.66.1±0.315.4±0.6
Herding10.4±0.428.7±0.76.7±0.416.3±0.7
UniDD (Ours)23.9±0.548.5±0.610.8±0.422.5±0.5
Full51.4±0.880.3±0.432.6±0.751.4±0.8
", + "bbox": [ + 94, + 253, + 480, + 563 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/18f5cb838b2dda3a4d1303afa850090e13a56caf7efc736833c2f0c0ee600c6c.jpg", + "table_caption": [ + "Table 3. Performance comparison with coreset selection methods on Pascal VOC and MS COCO for the object detection task, using Faster R-CNN as the detector. \"Ratio\" denotes the compression ratio." + ], + "table_footnote": [], + "table_body": "
MethodsSemantic Segmentation (mIoU)
Ratio: 3.5%Ratio: 7%Ratio: 14%Full
Random6.7±0.49.6±0.511.3±0.4
Uniform6.5±0.69.3±0.412.1±0.5
K-Center6.1±0.48.5±0.610.4±0.732.9±0.6
Herding6.2±0.78.7±0.410.6±0.3
UniDD (Ours)10.3±0.512.3±0.314.9±0.7
", + "bbox": [ + 94, + 646, + 480, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "versatility across a range of vision tasks.", + "bbox": [ + 89, + 838, + 359, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Result. As shown in Fig. 2 (b) and 2 (c), we present visualizations of detection and segmentation data synthesized using the UniDD framework. These images", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "are generated with $\\mathcal{T}_{\\mathcal{SP}}$ models, Faster R-CNN for detection and LR-ASPP for segmentation, which effectively steer the diffusion process toward the specified labels, bounding boxes, and masks. Previous research in dataset distillation has predominantly focused on single-category label synthesis, where each image aligns with a single class. In contrast, our UniDD framework overcomes this limitation, enabling dataset distillation for more complex, multi-class tasks. This advancement not only broadens the applicability of dataset distillation but also offers valuable insights for future research in this domain.", + "bbox": [ + 511, + 256, + 906, + 421 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 433, + 666, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Component Analysis. We conduct a comprehensive ablation study on the proposed UniDD, analyzing its components: (1) Baseline (a standard SD v1.5 checkpoint), (2) Task-driven Diffusion with the $\\mathcal{T}_{\\mathcal{SP}}$ model, (3) Task-Aligned Contextual Prompting (abbreviated as Context), and (4) Proxy-Driven High-Realism Refinement (abbreviated as Refine). As shown in Table 5, the proposed task-driven approach significantly improves the performance of the baseline, demonstrating that this guidance effectively bridges the gap between real and synthetic datasets. The contextual information is more effective for more complex detection and segmentation datasets, as it reveals underlying target relationships. The refinement module also proves to be effective, as the randomness of the SD model can lead to the generation of images that deviate from the annotations or result in failed generations.", + "bbox": [ + 511, + 455, + 908, + 699 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7f22ca9639b56d5f84fd761931c9b64315b1de352de9424b41fd7c3c733dfecb.jpg", + "table_caption": [ + "Table 4. Performance comparison with coreset selection methods on Pascal VOC for the semantic segmentation task, using LR-ASPP as the segmenter. The metric is mIoU, and \"Ratio\" denotes the compression ratio." + ], + "table_footnote": [], + "table_body": "
MethodsDataset
Cls. / Accuracy (%)Det. / mAP (%)Seg. / mIoU (%)
ImageNetteImageWoofVOCVOC
Baseline60.4±0.939.1±1.111.5±0.27.2±0.3
Baseline+TSP64.2±0.845.3±0.614.7±0.48.9±0.5
Baseline+TSRP+Context64.7±0.446.1±0.715.9±0.39.6±0.4
Baseline+TSP+Context+Refine66.0±0.746.8±0.916.8±0.510.3±0.5
", + "bbox": [ + 514, + 712, + 903, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Ablation study on UniDD. For classification datasets, IPC is set to 10 with ResNet-18 as the test model. For detection on VOC, the compression ratio is set to $1\\%$ with Faster R-CNN as the test model. For segmentation on VOC, the ratio is set to $3.5\\%$ with LR-ASPP as the test model.", + "bbox": [ + 511, + 797, + 906, + 867 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross-Architecture Analysis. Cross-Architecture experi", + "bbox": [ + 511, + 885, + 903, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10563", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8167e876d94a483cab78fcb5c4737e0d923e8288bd77a481dbabd652b184c829.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetTest Model\\( T_{SP} \\)Model
ResNet-18ResNet-50ResNet-101
ImageNetteResNet-1866.0±0.768.8±0.769.5±0.7
ResNet-5067.2±0.669.5±1.170.7±0.9
ResNet-10169.3±0.970.9±1.272.1±0.6
ImageWoofResNet-1846.8±0.949.2±0.550.1±0.8
ResNet-5047.3±0.750.0±0.850.9±0.6
ResNet-10148.9±1.151.5±0.652.4±0.7
", + "bbox": [ + 91, + 88, + 480, + 207 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. $\\mathcal{T}_{\\mathcal{SP}}$ Model Ablation and Cross-Architecture Testing: We use ResNet-18, ResNet-50, and ResNet-101 as $\\mathcal{T}_{\\mathcal{SP}}$ models to generate synthetic images. During testing, the same three models are used. IPC is set to 10.", + "bbox": [ + 89, + 217, + 482, + 272 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/710da0b5dfd0a15b8f27cf0cd423dfc27eb92886c2f2843ff03c914356ef0257.jpg", + "image_caption": [ + "(a) Accuracy vs. Guidance strength" + ], + "image_footnote": [], + "bbox": [ + 102, + 297, + 272, + 428 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/28c224eb7f4d1179d114e2528c56f2684e8881d26b53783ef767363b2beb0521.jpg", + "image_caption": [ + "(b) Accuracy vs. Sample Step" + ], + "image_footnote": [], + "bbox": [ + 290, + 297, + 460, + 426 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/56268c4bf81de374ccba9e17d01f26bc4e2c049a1a2d33703cbd143c4e5f24bf.jpg", + "image_caption": [ + "Figure 4. Hyper-parameter analysis on (a) guidance strength $\\lambda_{s}$ and (b) sampling step $t$ . The star symbol indicates the parameters chosen for trade-off.", + "(a) Training Time overhead", + "Figure 5. Comparison of distillation cost on ImageNet-1k using our UniDD with comparison methods (MTT [3], $\\mathrm{SRe}^2\\mathrm{L}$ [46], and $\\mathrm{D^4M}$ [41]), tested on V100 GPUs." + ], + "image_footnote": [], + "bbox": [ + 99, + 518, + 277, + 652 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3980bfe19f23ab16a4f8b59e65acfff76214a08a92e6d61e9526d8bf476c7cb1.jpg", + "image_caption": [ + "(b) Training GPU overhead" + ], + "image_footnote": [], + "bbox": [ + 287, + 518, + 464, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ments are crucial for validating the generalizability of synthetic data. We compare different combinations of $\\mathcal{T}_{\\mathcal{SP}}$ models and testing models. In this study, we set up three $\\mathcal{T}_{\\mathcal{SP}}$ models: ResNet-18, ResNet-50, and ResNet-101. These models are also used during testing to evaluate the cross-architecture performance of the synthetic datasets. As shown in Table. 6, using more complex $\\mathcal{T}_{\\mathcal{SP}}$ models improves the guidance effectiveness. Moreover, when testing across different architectures, the performance does not degrade as seen in previous optimization-based meth", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ods [3, 49, 50], demonstrating the generalizability of our synthetic data.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Guidance Strength Analysis. As shown in Fig. 4 (a), we analyze the performance variations of ImageWoof and ImageNette at IPC-10 under different guidance strength settings $\\lambda_{s}$ , with Conv5 as the test model. The best results are achieved at $\\lambda_{s} = 100$ , but further increasing the guidance strength interferes with the quality of diffusion-generated images, leading to a decline in test performance.", + "bbox": [ + 511, + 121, + 903, + 226 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sample Step Analysis. As shown in Fig. 4 (b), we analyze the impact of different SD sample steps on the performance of synthetic datasets under the IPC-10 setting, with Image-Woof and ImageNette resolutions set to $256 \\times 256$ . Performance improves significantly with sample steps below 250, but gains are minimal beyond this point while time consumption increases. Thus, we select 250 steps as the optimal trade-off.", + "bbox": [ + 511, + 227, + 903, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Distillation Cost Analysis. As shown in Fig. 5, we evaluate the time and GPU cost of various DD methods. Previous methods primarily focus on the time required to synthesize images, often neglecting the time needed to train or deploy the DD model. For MTT [3], the deployment time corresponds to the expert trajectory training time, while for $\\mathrm{D^4M}$ [41], it is the time required to finetune the SD model. The slight increase in UniDD's distillation time and GPU cost is mainly due to the added supervision process of the $\\mathcal{T}_{\\mathcal{SP}}$ model and the refinement stage, where approximately $5\\% -10\\%$ of low-quality generated data is discarded. Combining the deployment and distillation costs for comparison, UniDD demonstrates the best advantage, indicating its superior suitability for practical deployment.", + "bbox": [ + 511, + 348, + 906, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 573, + 633, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we introduce the first Universal Dataset Distillation (UniDD) framework, capable of handling classification, detection, and segmentation tasks within a unified approach. UniDD follows a two-stage process: in the first stage, Task-Specific Proxy Training and Task-Aligned Contextual Prompting extract critical information from real datasets. In the second stage, Universal Taskdriven Diffusion synthesizes images with specific categories, bounding boxes, and segmentation masks tailored to each task. We conduct extensive experiments to validate UniDD's effectiveness across all three tasks. For classification, UniDD achieves state-of-the-art results on ImageNet1k and ImageNet-Subset. In detection and segmentation, we establish the first dataset distillation benchmark on Pascal VOC and MS COCO, comparing UniDD to coreset selection methods. Encouraging results demonstrate that our UniDD consistently outperforms these methods, especially under extremely low compression ratios.", + "bbox": [ + 511, + 598, + 906, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported by National Natural Science Fund of China (62473286).", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10564", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 843-852, 2023. 3, 4", + "[2] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European conference on computer vision (ECCV), pages 233-248, 2018. 6", + "[3] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A Efros, and Jun-Yan Zhu. Dataset distillation by matching training trajectories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4750-4759, 2022. 1, 2, 5, 6, 8", + "[4] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 6", + "[5] Yutian Chen, Max Welling, and Alex Smola. Super-samples from kernel herding. arXiv preprint arXiv:1203.3472, 2012. 6", + "[6] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.6", + "[7] Justin Cui, Ruochen Wang, Si Si, and Cho-Jui Hsieh. Scaling up dataset distillation toImagenet-1k with constant memory. In International Conference on Machine Learning, pages 6565–6590. PMLR, 2023. 1, 5", + "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1, 2, 3, 5", + "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3, 4", + "[10] Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. Advances in Neural Information Processing Systems, 36:16222-16239, 2023. 3", + "[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 1, 2, 3, 5", + "[12] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. International journal of computer vision, 111:98-136, 2015. 1, 2, 3, 5", + "[13] Junyao Gao, Xinyang Jiang, Huishuai Zhang, Yifan Yang, Shuguang Dou, Dongsheng Li, Duoqian Miao, Cheng Deng," + ], + "bbox": [ + 93, + 114, + 483, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Cairong Zhao. Similarity distribution based membership inference attack on person re-identification. In Proceedings of the AAAI conference on artificial intelligence, pages 14820-14828, 2023. 1", + "[14] Junyao Gao, Xinyang Jiang, Shuguang Dou, Dongsheng Li, Duoqian Miao, and Cairong Zhao. Re-id-leak: Membership inference attacks against person re-identification. International Journal of Computer Vision, 132(10):4673-4687, 2024. 1", + "[15] Junyao Gao, Yanchen Liu, Yanan Sun, Yinhao Tang, Yanhong Zeng, Kai Chen, and Cairong Zhao. Styleshot: A snapshot on any style. arXiv preprint arXiv:2407.01414, 2024. 2", + "[16] Junyao Gao, Yanan Sun, Fei Shen, Xin Jiang, Zhening Xing, Kai Chen, and Cairong Zhao. Faceshot: Bring any character into life. arXiv preprint arXiv:2503.00740, 2025. 2", + "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2", + "[18] Jianyang Gu, Saeed Vahidian, Vyacheslav Kungurtsev, Haonan Wang, Wei Jiang, Yang You, and Yiran Chen. Efficient dataset distillation via minimax diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15793-15803, 2024. 5, 6", + "[19] Ziyao Guo, Kai Wang, George Cazenavette, Hui Li, Kaipeng Zhang, and Yang You. Towards lossless dataset distillation via difficulty-aligned trajectory matching. arXiv preprint arXiv:2310.05773, 2023. 2", + "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1", + "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4, 5", + "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1", + "[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3", + "[24] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1314-1324, 2019. 5", + "[25] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018. 2", + "[26] Jang-Hyun Kim, Jinuk Kim, Seong Joon Oh, Sangwoo Yun, Hwanjun Song, Joonhyun Jeong, Jung-Woo Ha, and Hyun Oh Song. Dataset condensation via efficient synthetic-data parameterization. In International Conference on Machine Learning, pages 11102-11118. PMLR, 2022. 1, 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10565", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Jian Li, Yabiao Wang, Changan Wang, Ying Tai, Jianjun Qian, Jian Yang, Chengjie Wang, Jilin Li, and Feiyue Huang. Dsfd: dual shot face detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5060-5069, 2019. 2", + "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1, 2, 5", + "[29] Songhua Liu, Kai Wang, Xingyi Yang, Jingwen Ye, and Xin chao Wang. Dataset distillation via factorization. Advances in Neural Information Processing Systems, 35:1100-1113, 2022. 6", + "[30] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 289-299, 2023. 3", + "[31] Yanqing Liu, Jianyang Gu, Kai Wang, Zheng Zhu, Wei Jiang, and Yang You. Dream: Efficient dataset distillation by representative matching. arXiv preprint arXiv:2302.14416, 2023. 2", + "[32] Timothy Nguyen, Zhourong Chen, and Jaehoon Lee. Dataset meta-learning from kernel ridge-regression. In International Conference on Learning Representations, 2020. 2", + "[33] OpenAI. Gpt-4 technical report. https://openai.com/research/gpt-4, 2023. Accessed: 2024-11-13. 2, 4", + "[34] Ding Qi, Jian Li, Jinlong Peng, Bo Zhao, Shuguang Dou, Jialin Li, Jiangning Zhang, Yabiao Wang, Chengjie Wang, and Cairong Zhao. Fetch and forge: Efficient dataset condensation for object detection. Advances in Neural Information Processing Systems, 37:119283-119300, 2024. 3", + "[35] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 6", + "[36] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 4, 5, 6", + "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 5", + "[38] Ahmad Sajedi, Samir Khaki, Ehsan Amjadian, Lucy Z Liu, Yuri A Lawryshyn, and Konstantinos N Plataniotis. Datadam: Efficient dataset distillation with attention matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17097-17107, 2023. 6", + "[39] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. arXiv preprint arXiv:1708.00489, 2017. 6" + ], + "bbox": [ + 91, + 92, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3", + "[41] Duo Su, Junjie Hou, Weizhi Gao, Yingjie Tian, and Bowen Tang. D^4: Dataset distillation via disentangled diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5809-5818, 2024. 5, 6, 8", + "[42] Peng Sun, Bei Shi, Daiwei Yu, and Tao Lin. On the diversity and realism of distilled dataset: An efficient dataset distillation paradigm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9390-9399, 2024. 1, 3, 5", + "[43] Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Shuo Yang, Shuo Wang, Guan Huang, Hakan Bilen, Xinchao Wang, and Yang You. Cafe: Learning to condense dataset by aligning features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12196-12205, 2022. 2", + "[44] Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A Efros. Dataset distillation. arXiv preprint arXiv:1811.10959, 2018. 1, 2, 6", + "[45] Yichao Yan, Zanwei Zhou, Zi Wang, Jingnan Gao, and Xiaokang Yang. Dialoguenerf: Towards realistic avatar face-to-face conversation video generation. Visual Intelligence, 2 (1):24, 2024. 1", + "[46] Zeyuan Yin, Eric Xing, and Zhiqiang Shen. Squeeze, recover and relabel: Dataset condensation at imagenet scale from a new perspective. arXiv preprint arXiv:2306.13092, 2023. 1, 3, 5, 6, 8", + "[47] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2", + "[48] Bo Zhao and Hakan Bilen. Dataset condensation with differentiable siamese augmentation. In International Conference on Machine Learning, pages 12674-12685. PMLR, 2021. 1", + "[49] Bo Zhao and Hakan Bilen. Dataset condensation with distribution matching. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6514-6523, 2023. 1, 2, 6, 8", + "[50] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. Dataset condensation with gradient matching. In International Conference on Learning Representations, 2020. 1, 2, 6, 8", + "[51] Ganlong Zhao, Guanbin Li, Yipeng Qin, and Yizhou Yu. Improved distribution matching for dataset condensation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7856-7865, 2023. 1, 2" + ], + "bbox": [ + 516, + 92, + 903, + 781 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10566", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Towards Universal Dataset Distillation via Task-Driven Diffusion/cd8b4c9e-1814-423f-bdad-2b631dc6a75a_model.json b/2025/Towards Universal Dataset Distillation via Task-Driven Diffusion/cd8b4c9e-1814-423f-bdad-2b631dc6a75a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3467099bbb2f4ce97044cac0932bc4abdbbe366a --- /dev/null +++ b/2025/Towards Universal Dataset Distillation via Task-Driven Diffusion/cd8b4c9e-1814-423f-bdad-2b631dc6a75a_model.json @@ -0,0 +1,2398 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.171, + 0.131, + 0.828, + 0.152 + ], + "angle": 0, + "content": "Towards Universal Dataset Distillation via Task-Driven Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.179, + 0.834, + 0.254 + ], + "angle": 0, + "content": "Ding Qi\\(^{1*}\\) Jian Li\\(^{2\\dagger}\\) Junyao Gao\\(^{1}\\) Shuguang Dou\\(^{1}\\) Ying Tai\\(^{5}\\) \nJianlong Hu\\(^{2}\\) Bo Zhao\\(^{4}\\) Yabiao Wang\\(^{3,2\\dagger}\\) Chengjie Wang\\(^{4,2}\\) Cairong Zhao\\(^{1\\dagger}\\) \n\\(^{1}\\)Tongji University \\(^{2}\\)Tencent YouTu Lab \\(^{3}\\)Zhejiang University \n\\(^{4}\\)Shanghai Jiao Tong University \\(^{5}\\)Nanjing University" + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.269, + 0.599, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.151, + 0.463, + 0.551, + 0.476 + ], + "angle": 0, + "content": "(a) One Universal Dataset Distillation Framework for Mutli-Tasks" + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.28, + 0.868, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.462, + 0.874, + 0.477 + ], + "angle": 0, + "content": "(b) Improved Performance by UniDD" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.489, + 0.908, + 0.545 + ], + "angle": 0, + "content": "Figure 1. (a) We introduce the Universal Dataset Distillation Framework (UniDD), a novel dataset distillation framework supporting a range of vision tasks, including classification, object detection, and segmentation. UniDD incorporates diverse label types, such as class labels, bounding box coordinates, and pixel-level masks, facilitating the synthesis of task-specific datasets. (b) UniDD improves the performance of dataset distillation across diverse tasks on ImageNet [8], Pascal VOC [11, 12], and MS COCO [28]." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.564, + 0.327, + 0.579 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.596, + 0.486, + 0.854 + ], + "angle": 0, + "content": "Dataset distillation (DD) condenses key information from large-scale datasets into smaller synthetic datasets, reducing storage and computational costs for training networks. However, most recent research has primarily focused on image classification tasks, with limited exploration in detection and segmentation. Two key challenges remain: (i) Task Optimization Heterogeneity, where existing methods focus on class-level information but fail to address the diverse needs of detection and segmentation, and (ii) Inflexible Image Generation, where current generation methods rely on global updates for single-class targets and lack localized optimization for specific object regions. To address these challenges, we propose UniDD, a universal dataset distillation framework built on a task-driven diffusion model for diverse DD tasks, as shown in Fig. 1. Our approach operates in two stages: Universal Task Knowledge Mining, which captures task-relevant information through task" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.909, + 0.687 + ], + "angle": 0, + "content": "specific proxy model training, and Universal Task-Driven Diffusion, where these proxies guide the diffusion process to generate task-specific synthetic images. Extensive experiments across ImageNet-1K, Pascal VOC, and MS COCO demonstrate that UniDD consistently outperforms state-of-the-art methods. In particular, on ImageNet-1K with IPC-10, UniDD surpasses previous diffusion-based methods by \\(6.1\\%\\), while also reducing deployment costs." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.722, + 0.646, + 0.737 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Dataset distillation (DD) [44], as an innovative data optimization technique in machine learning [13, 14, 20, 22, 45], condenses essential information from large datasets into smaller proxy sets. Training on these highly ensemble proxy sets enables the model to use much fewer resources while providing comparable performance. Prior DD methods excelled in image classification using various techniques like Gradient Matching [26, 48, 50], Distribution Matching [49, 51], Training Trajectory Matching [3, 7], and Decoupled Methods [42, 46]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.483, + 0.887 + ], + "angle": 0, + "content": "*Work done when Ding Qi is an intern in Tencent YouTu Lab, and the advisor is Jian Li." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.888, + 0.248, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.863, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10557" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.346 + ], + "angle": 0, + "content": "Despite advancements in DD, research on key visual tasks—detection and segmentation—remains limited. We identify the following key challenges: (i) Task Optimization Heterogeneity. Existing DD methods are primarily focused on image classification, optimizing distillation by compressing class-level information, but they fall short in addressing the diverse needs of tasks like detection [27] and segmentation, such as object localization and pixel-level information. (ii) Inflexible Image Generation. Current image generation methods rely on global updates for single-class targets and lack localized optimization for specific object regions. These limitations become especially evident in images with multiple objects or complex structures, hindering the ability of existing DD to generate distilled data that meets the requirements of diverse tasks. Consequently, the applicability of DD in the visual domain remains constrained." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.352, + 0.485, + 0.775 + ], + "angle": 0, + "content": "To address these limitations, we propose a Universal Dataset Distillation (UniDD) framework that optimizes for multi-task learning, including classification, detection, and segmentation, overcoming the shortcomings in task adaptation and image generation. Our approach consists of two stages: Universal Task Knowledge Mining and Universal Task-Driven Diffusion. In the first stage, we introduce Task-Specific Proxy \\((\\mathcal{T}_{\\mathcal{SP}})\\) models, trained on real datasets to extract task-specific information. This allows the \\(\\mathcal{T}_{\\mathcal{SP}}\\) models to capture and store category information for classification, bounding box positions for detection, and mask data for segmentation within their parameters. Additionally, previous research has often overlooked the importance of contextual information, particularly in high-resolution datasets like ImageNet [8], Pascal VOC [11, 12] and MS COCO [28], where rich object-context relationships exist. To address this, we introduce Task-Aligned Contextual Prompting, which uses a Vision-Language model [33] to generate contextual prompts for real images. In the second stage, we introduce a Task-Driven Diffusion for image synthesis. Specifically, the \\(\\mathcal{T}_{\\mathcal{SP}}\\) models trained in the first stage supervise the denoising process of the diffusion model by calculating the task loss on the predicted clean images. This guides the diffusion model in generating images that match the given labels. Notably, since the \\(\\mathcal{T}_{\\mathcal{SP}}\\) models already align with the target domain, there is no need to retrain the diffusion model on the real set, significantly reducing the cost in practical applications." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.778, + 0.405, + 0.792 + ], + "angle": 0, + "content": "In summary, our contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.795, + 0.483, + 0.855 + ], + "angle": 0, + "content": "- We identify key challenges in extending dataset distillation beyond classification and propose the first universal framework for DD across classification, detection, and segmentation tasks." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "- We develop a two-stage approach, Universal Task Knowledge Mining and Universal Task-Driven Diffusion for Data Synthesis, which flexibly synthesizes images for" + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.795, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.136 + ], + "angle": 0, + "content": "multi-tasks. Additionally, we address the previously overlooked need to align potential contextual information within datasets." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.137, + 0.905, + 0.228 + ], + "angle": 0, + "content": "- Our method achieves state-of-the-art performance in classification tasks, on the ImageNet-1K dataset, our approach surpasses the previous state-of-the-art method by over \\(6.1\\%\\). Additionally, we establish new benchmarks for detection and segmentation on VOC and COCO, showing significant improvements over coreset methods." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.243, + 0.661, + 0.259 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.269, + 0.698, + 0.284 + ], + "angle": 0, + "content": "2.1. Dataset Distillation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.292, + 0.907, + 0.715 + ], + "angle": 0, + "content": "Dataset distillation [44], a dataset reduction method, addresses the challenge of handling substantial data by synthesizing a smaller, representative dataset. Existing dataset distillation methods can be categorized into meta-learning and data matching frameworks based on whether they explicitly mimic the performance of target data. In the meta-learning framework, the distilled data are treated as hyperparameters and optimized in a nested loop fashion according to the trained model's risk. For example, KIP [32] performs kernel ridge regression using the Neural Tangent Kernel (NTK) [25]. The data matching framework includes gradient matching, distribution matching, and trajectory matching. These methods update distilled data by imitating the influence of target data. For example, DC [50] and DM [49] enforce synthetic and real images to have similar gradients or distribution at every training step. MTT [3] proposes to imitate long-range trajectories of optimization steps for real images, thereby better mimicking the training dynamics of real images for synthetic images. However, the aforementioned gradient matching methods require a large number of iterations. To address this, DREAM [31] and IDC [26] focus on more efficient dataset distillation training. Due to the unsatisfactory performance of DM, CAFE [43] and IDM [51] propose improved distribution matching between the real and synthetic data. For the trajectory matching method, DATM [19] scales trajectory matching-based methods to larger synthetic datasets, achieving lossless dataset distillation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.716, + 0.906, + 0.82 + ], + "angle": 0, + "content": "Despite significant progress in previous work, there have been few breakthroughs in dataset distillation for detection and segmentation tasks. In this work, we introduce the first universal framework that simultaneously handles classification, object detection, and semantic segmentation, significantly expanding the generality and scope of dataset distillation research." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.833, + 0.682, + 0.848 + ], + "angle": 0, + "content": "2.2. Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Diffusion models [15, 16, 37, 47] are generative models that show superior fidelity and diversity compared to GANs [17]. It aims to learn the mapping from a Gaussian" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10558" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.092, + 0.356, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.291, + 0.326, + 0.303 + ], + "angle": 0, + "content": "(a) Synthetic images for classification." + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.092, + 0.63, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.407, + 0.291, + 0.59, + 0.304 + ], + "angle": 0, + "content": "(b) Synthetic images for detection." + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.091, + 0.905, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.291, + 0.875, + 0.304 + ], + "angle": 0, + "content": "(c) Synthetic images for segmentation." + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.315, + 0.907, + 0.372 + ], + "angle": 0, + "content": "Figure 2. Visualization of synthetic images generated by UniDD. (a) For classification, images are synthesized on the ImageWoof [8] dataset. (b) For detection, images are synthesized on the Pascal VOC dataset with specified bounding box coordinates. (c) For segmentation, images are synthesized on the Pascal VOC [11, 12] dataset with pixel-level masks. Our UniDD is the first approach to enable high-realism data distillation across classification, detection, and segmentation tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.484, + 0.456 + ], + "angle": 0, + "content": "distribution to the real dataset distribution by adding noise in the diffusion process and reversing the noise to the raw image in the denoising process. In the forward process, diffusion models progressively add noise to a clean sample \\( z_0 \\), resulting in a noisy sample \\( z_t \\):" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.462, + 0.483, + 0.48 + ], + "angle": 0, + "content": "\\[\nz _ {t} = \\sqrt {\\alpha_ {t}} z _ {0} + \\sqrt {1 - \\alpha_ {t}} \\epsilon , \\quad \\epsilon \\sim \\mathcal {N} (0, I). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.485, + 0.484, + 0.515 + ], + "angle": 0, + "content": "In the reverse process, a denoising network \\(\\epsilon_{\\theta}\\) estimates the noise and iteratively recovers the clean sample \\(z_0\\):" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.52, + 0.483, + 0.555 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {\\theta} (z _ {t}, t) \\approx \\frac {z _ {t} - \\sqrt {\\alpha_ {t}} z _ {0}}{\\sqrt {1 - \\alpha_ {t}}}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.558, + 0.484, + 0.588 + ], + "angle": 0, + "content": "DDIM [40] provides the following formula for predicting a clean data sample:" + }, + { + "type": "equation", + "bbox": [ + 0.189, + 0.592, + 0.483, + 0.628 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {0} = \\frac {z _ {t} - \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {\\theta} (z _ {t} , t)}{\\sqrt {\\alpha_ {t}}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.632, + 0.484, + 0.676 + ], + "angle": 0, + "content": "This formula predicts the clean data point \\(\\hat{z}_0\\) from the noisy sample \\(z_{t}\\), enabling deterministic sampling for efficient and high-quality image synthesis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.677, + 0.484, + 0.813 + ], + "angle": 0, + "content": "To introduce additional conditions into the diffusion model, classifier guidance [9] has been proposed to train an additional classifier on the noise to guide the model to generate images that are closer to the specified condition. Additionally, [23] introduces classifier-free guidance that trains both an unconditioned and conditioned model simultaneously, achieving more efficient guidance. Following these, recent methods [1, 10, 30] have been proposed to generate high-quality images from diverse conditions." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.824, + 0.466, + 0.84 + ], + "angle": 0, + "content": "3. Universal Dataset Distillation Framework" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.274, + 0.864 + ], + "angle": 0, + "content": "3.1. Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Universal Dataset Distillation. Current dataset distillation research primarily targets classification tasks and lacks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.907, + 0.562 + ], + "angle": 0, + "content": "a generalizable approach for detection and segmentation. To address this, we propose a universal dataset distillation standard that constructs a compact synthetic dataset \\( \\mathcal{S} = (\\hat{x}_1, \\hat{y}_1), \\dots, (\\hat{x}_{|\\mathcal{S}|}, \\hat{y}_{|\\mathcal{S}|}) \\) from a larger real dataset \\( \\mathcal{T} = (x_1, y_1), \\dots, (x_{|\\mathcal{T}|}, y_{|\\mathcal{T}|}) \\), with \\( |S| \\ll |\\mathcal{T}| \\). The primary difference between tasks lies in label structure: classification uses single-class labels, detection includes bounding boxes with class information, and segmentation requires pixel-level masks. Our goal is for models trained on the distilled dataset \\( \\mathcal{S} \\) to perform comparably to those trained on the full dataset \\( \\mathcal{T} \\) across classification, detection, and segmentation tasks, as formalized by:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.571, + 0.907, + 0.591 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {(x, y) \\sim P _ {D}} \\left| \\ell \\left(\\theta_ {\\mathcal {T}} ^ {t a s k} (x), y\\right) - \\ell \\left(\\theta_ {\\mathcal {S}} ^ {t a s k} (x), y\\right) \\right| \\leq \\epsilon , \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.599, + 0.907, + 0.675 + ], + "angle": 0, + "content": "where \\(P_{D}\\) denotes the distribution of test data, \\(\\ell\\) represents the performance metric specific to each task, \\(task\\in\\) {class,detect,segment}, and \\(\\epsilon\\) is a small tolerance value. For each task, the parameter set \\(\\theta_{\\mathcal{T}}^{task}\\) on the real dataset \\(\\mathcal{T}\\) is optimized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.684, + 0.907, + 0.709 + ], + "angle": 0, + "content": "\\[\n\\theta_ {\\mathcal {T}} ^ {t a s k} = \\arg \\min _ {\\theta} \\mathbb {E} _ {(x, y) \\in \\mathcal {T}} \\left[ \\ell \\left(\\phi_ {\\theta} ^ {t a s k} (x), y\\right) \\right], \\qquad (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.718, + 0.899, + 0.736 + ], + "angle": 0, + "content": "Similarly, \\(\\theta_S^{task}\\) is optimized using the synthetic dataset \\(S\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.816, + 0.76 + ], + "angle": 0, + "content": "3.2. Universal Task Knowledge Mining" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Task-Specific Proxy Training. Building a universal DD framework is challenging due to the diverse optimization objectives in classification, detection, and segmentation tasks. Recent decoupled DD studies have reached a consensus [34, 42, 46]: a fully trained network can capture and retain essential dataset information. Inspired by this, we extend to general tasks, training a classifier, detector, and segmenter on task-specific datasets to extract critical information—class labels for classification, bounding boxes for" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10559" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.907, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.383, + 0.907, + 0.44 + ], + "angle": 0, + "content": "Figure 3. Universal Dataset Distillation Framework. This framework consists of two stages: in the Universal Task Knowledge Mining stage, we introduce an easily deployable task-specific proxy \\((\\mathcal{T}_{\\mathcal{SP}})\\) model trained on the target dataset and a vision-language model to extract task-aligned context. In the Universal Task-Driven Diffusion stage, we apply the gradient of the supervision loss, derived from the \\(\\mathcal{T}_{\\mathcal{SP}}\\) model's prediction of the clean image \\(\\hat{z}_0\\), at each sampling step." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.483, + 0.478 + ], + "angle": 0, + "content": "detection, and pixel-level masks for segmentation. We call this Task-Specific Proxy \\((\\mathcal{T}_{\\mathcal{SP}})\\) Model, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.49, + 0.483, + 0.511 + ], + "angle": 0, + "content": "\\[\n\\theta = \\arg \\min _ {\\theta} \\mathcal {L} _ {\\text {t a s k}} (\\mathcal {F} (x), y), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.523, + 0.483, + 0.568 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}\\) represents the \\(\\mathcal{T}_{\\mathcal{SP}}\\) model trained on the target data, \\(\\theta\\) denotes the model's parameters, \\(\\mathcal{L}_{task}\\) represents the loss function specific to the task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.719 + ], + "angle": 0, + "content": "The \\(\\mathcal{T}_{\\mathcal{SP}}\\) model choice is flexible, allowing adaptation for optimal performance on target datasets. For example, ResNet-18 [21] can serve as a lightweight, fast-deploy option for simple classification, while Faster R-CNN [36] may be used for complex detection tasks to capture key detection information. Unlike previous diffusion-based dataset distillation methods, our approach relies on task-driven learning to align synthetic and real datasets, avoiding costly diffusion model fine-tuning on the target dataset and significantly reducing deployment overhead." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Task-Aligned Contextual Prompting. Previous diffusion-based dataset distillation has largely focused on classification tasks and overlooked the role of natural language in guiding image generation. For diverse tasks, textual prompts must go beyond describing the primary objects in images. As inputs to Stable Diffusion models, prompts can also guide the generation process. To address this, we use a vision-language model to extract natural language descriptions and design task-specific questions to generate task-aligned prompts, a method we call Task-Aligned Contextual Prompting. Following GPT-4o [33], we create tailored prompts for different tasks. For classification, prompts de" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.907, + 0.508 + ], + "angle": 0, + "content": "scribe object classes, while for detection and segmentation, they also include object relationships and the interaction between objects and the background, providing richer contextual guidance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.517, + 0.803, + 0.532 + ], + "angle": 0, + "content": "3.3. Universal Task-Driven Diffusion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.659 + ], + "angle": 0, + "content": "Inspired by recent advancements in guidance diffusion techniques [1, 9], we propose a Task-Driven Diffusion approach for Dataset Distillation. As illustrated in Fig. 3, during each sampling step \\( S(z_{t}, t) \\) of the diffusion process, the trained \\( \\mathcal{T}_{\\mathcal{SP}} \\) model \\( \\mathcal{F} \\) provides task-specific guidance. Given an label \\( \\tilde{a} \\), the original noise prediction \\( \\epsilon_{\\theta}(z_{t}, t, \\tilde{c}) \\) is adjusted using the output of \\( \\mathcal{F} \\). The adjusted noise prediction \\( \\hat{\\epsilon} \\) is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.669, + 0.907, + 0.686 + ], + "angle": 0, + "content": "\\[\n\\hat {\\epsilon} _ {\\theta} \\left(z _ {t}, t, \\tilde {c}\\right) = \\epsilon_ {\\theta} \\left(z _ {t}, t, \\tilde {c}\\right) + \\sqrt {1 - \\alpha_ {t}} \\nabla_ {z _ {t}} \\mathcal {L} _ {\\text {t a s k}} \\left(\\tilde {a}, \\mathcal {F} \\left(z _ {t}\\right)\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.695, + 0.906, + 0.785 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{task}}\\) represents the task-specific loss function and \\(\\tilde{a} = \\{a_{cl}, a_{bbox}, a_{mask}\\}\\) may denote a class label \\(a_{cl}\\), a bounding box \\(a_{bbox}\\), or a pixel-level segmentation mask \\(a_{mask}\\). This adjustment mechanism effectively guides the diffusion process to produce data closely aligned with the intended distribution for each task. \\(\\tilde{c}\\) is a textual prompt." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.786, + 0.906, + 0.847 + ], + "angle": 0, + "content": "Because \\(\\mathcal{F}\\) is trained on clean images, it may not perform well directly on noisy inputs. Therefore, we first use Equation 3 to estimate the clean image \\(\\hat{z}_0\\). Once \\(\\hat{z}_0\\) is obtained, we apply \\(\\mathcal{F}\\) to compute the task-driven loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.858, + 0.907, + 0.875 + ], + "angle": 0, + "content": "\\[\n\\hat {\\epsilon} _ {\\theta} \\left(z _ {t}, t, \\tilde {c}\\right)) = \\epsilon_ {\\theta} \\left(z _ {t}, t, \\tilde {c}\\right) + s (t) \\cdot \\nabla_ {z _ {t}} \\mathcal {L} _ {\\text {t a s k}} (\\tilde {a}, \\mathcal {F} (\\hat {z} _ {0})), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\( s(t) \\) is the task-driven strength at each sampling step," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10560" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.167, + 0.105 + ], + "angle": 0, + "content": "defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.106, + 0.482, + 0.123 + ], + "angle": 0, + "content": "\\[\ns (t) = \\lambda_ {s} \\cdot \\sqrt {1 - \\alpha_ {t}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.129, + 0.484, + 0.205 + ], + "angle": 0, + "content": "with \\(\\lambda_{s}\\) being a constant coefficient controlling the overall intensity of the task-driven. The term \\(\\nabla_{z_t}\\mathcal{L}_{task}(\\tilde{a},\\mathcal{F}(\\hat{z}_0))\\) represents the gradient of the loss function computed based on the \\(\\mathcal{T}_{\\mathcal{SP}}\\) model's prediction of the clean image \\(\\hat{z}_0\\). Specifically, this gradient is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.214, + 0.484, + 0.259 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {z _ {t}} \\mathcal {L} _ {\\text {t a s k}} (\\tilde {a}, \\mathcal {F} (\\hat {z} _ {0})) = \\nabla_ {z _ {t}} \\mathcal {L} _ {\\text {t a s k}} \\left(\\tilde {a}, \\mathcal {F} \\left(\\frac {z _ {t} - \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {\\theta}}{\\sqrt {\\alpha_ {t}}}\\right)\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.268, + 0.484, + 0.418 + ], + "angle": 0, + "content": "Unlike previous approaches [18, 41] that required finetuning the diffusion model, such as those involving SD [37], to align with new data domains, our method leverages a \\(\\mathcal{T}_{\\mathcal{SP}}\\) model trained specifically on the target dataset. This model encapsulates critical information about the target task, effectively driving the diffusion process to generate data that is well-aligned with the target task without the need for additional finetuning. This not only conserves time and resources but also mitigates the risk of overfitting the diffusion model to the specific task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.419, + 0.484, + 0.526 + ], + "angle": 0, + "content": "Proxy-Driven High-Realism Refinement. To ensure the high realism and accuracy of the generated images, we employ a trained \\(\\mathcal{T}_{\\mathcal{SP}}\\) model, to score each image. By setting a well-defined threshold, we filter out low-quality images based on their scores. The remaining images are then relabeled using the same \\(\\mathcal{T}_{\\mathcal{SP}}\\) model. This combined process of filtering and relabeling can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.536, + 0.483, + 0.555 + ], + "angle": 0, + "content": "\\[\n\\{x, y \\} = \\left\\{x _ {i}, \\mathcal {F} _ {\\text {l a b e l}} \\left(x _ {i}\\right) \\mid \\operatorname {s c o r e} \\left(\\mathcal {F} \\left(x _ {i}\\right)\\right) \\geq \\tau \\right\\} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.484, + 0.64 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}(x_i)\\) represents the model's output score for image \\(x_{i}\\), \\(\\tau\\) is the quality threshold, and \\(\\mathcal{F}_{\\mathrm{label}}\\) generates the accurate label \\(y_{i}\\) for the filtered image. This ensures that both the generated images and their labels are of high quality and closely match the desired data distribution." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.643, + 0.224, + 0.66 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.668, + 0.289, + 0.684 + ], + "angle": 0, + "content": "4.1. Experimental Setups" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Datasets and Evaluation Metric. To enhance practicality, we prioritize high-resolution and large-scale datasets. (1) Classification task: our experiments are primarily conducted on the full ImageNet 1K dataset [8], with IPC settings commonly set to 10 and 50, following the configurations in [41]. Additionally, based on the dataset partitioning settings from previous work [3], we also use the ImageNet 10-class subsets, such as ImageWoof, ImageFruit, ImageNette, and ImageMeow. These 10-class subsets contain very similar categories, making them challenging for dataset distillation. Top-1 accuracy \\((\\%)\\) is reported in the experimental tables. (2) Object detection task: we conduct experiments on the widely-used Pascal VOC [11, 12] and MS COCO [28] datasets, using mAP (Mean Average" + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.091, + 0.871, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Universal Dataset Distillation Approach" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.108, + 0.906, + 0.138 + ], + "angle": 0, + "content": "Input: Real data \\(\\mathcal{T}\\), Diffusion model \\(\\epsilon_{\\theta}\\), \\(\\mathcal{T}_{\\mathcal{SP}}\\) model \\(\\mathcal{F}\\), Sampling steps \\(T\\), Task-Driven strength \\(s(t)\\)" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.139, + 0.712, + 0.153 + ], + "angle": 0, + "content": "Output: Synthetic dataset \\(S\\)." + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.156, + 0.838, + 0.171 + ], + "angle": 0, + "content": "1: Step I: Universal Task Knowledge Mining" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.172, + 0.8, + 0.186 + ], + "angle": 0, + "content": "2: Train \\(\\mathcal{T}_{\\mathcal{SP}}\\) model \\(\\mathcal{F}\\) on \\(\\mathcal{T}\\) using Equ. 6" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.186, + 0.905, + 0.216 + ], + "angle": 0, + "content": "3: Use GPT-4o to generate task-aligned context for classification, detection, and segmentation datasets" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.156, + 0.905, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.217, + 0.905, + 0.246 + ], + "angle": 0, + "content": "4: Step II: Universal Task-Driven Diffusion for Data Synthetic" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.247, + 0.789, + 0.262 + ], + "angle": 0, + "content": "5: for each sampling step \\( t = 1 \\) to \\( T \\) do" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.263, + 0.74, + 0.277 + ], + "angle": 0, + "content": "6: Generate noisy sample \\( z_{t} \\)." + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.278, + 0.805, + 0.291 + ], + "angle": 0, + "content": "7: Predict clean image \\(\\hat{z}_0\\) using Equ. 3" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.292, + 0.863, + 0.308 + ], + "angle": 0, + "content": "8: Calculate guidance loss \\(\\nabla_{z_t}\\mathcal{L}_{task}(c,\\mathcal{F}(\\hat{z}_0))\\)" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.308, + 0.809, + 0.322 + ], + "angle": 0, + "content": "9: Adjust noise prediction using Equ. 8" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.323, + 0.862, + 0.337 + ], + "angle": 0, + "content": "10: Update \\( z_{t-1} \\) using adjusted noise prediction." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.338, + 0.597, + 0.351 + ], + "angle": 0, + "content": "11: end for" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.353, + 0.835, + 0.367 + ], + "angle": 0, + "content": "12: Score each generated image \\( x_{i} \\) using \\( \\mathcal{F}(x_i) \\)." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.368, + 0.904, + 0.397 + ], + "angle": 0, + "content": "13: Filter and relabel synthetic set based on threshold \\(\\tau\\) using Equ. 11." + }, + { + "type": "list", + "bbox": [ + 0.52, + 0.247, + 0.904, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.906, + 0.489 + ], + "angle": 0, + "content": "Precision) and \\(AP_{50}\\) (Average Precision at IoU threshold of 0.50) as evaluation metric. (3) Semantic segmentation task: we utilize the Pascal VOC dataset, with mIoU (Mean Intersection over Union) as evaluation metrics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.907, + 0.807 + ], + "angle": 0, + "content": "Implementation Details. Unless otherwise specified, the Task-Specific Proxy \\(\\mathcal{T}_{\\mathcal{SP}}\\) model used for the classification task is ResNet-18 [21], the detector is Faster R-CNN [36], and the segmentation model is LR-ASPP [24], all utilizing the official pre-trained weights from PyTorch. For the diffusion model, we use SD v1.5 [37] as the baseline. For quantitative comparisons, we generate \\(256 \\times 256\\) images for classification tasks and \\(512 \\times 512\\) images for detection and segmentation tasks. Synthetic images are created using 250 denoising steps based on a predefined number of random noise samples, aligned with the IPC settings. The threshold \\(\\tau\\) for Proxy-Driven High-Realism Refinement is set to 0.9 for classification, 0.85 for detection, and 0.75 for segmentation. Generally, this operation results in a data discard rate of approximately \\(5 - 10\\%\\). All experiments are conducted on NVIDIA V100 GPUs, with further details provided in the special experiments section. We generate the distilled dataset 3 times, train the network from scratch 5 times per dataset (with pre-trained backbone weights for the detector and segmenter), and evaluate on the original test set, reporting results as \\(\\bar{x} \\pm std\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.738, + 0.833 + ], + "angle": 0, + "content": "4.2. Results on Classification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.901 + ], + "angle": 0, + "content": "We perform comprehensive comparisons across datasets of various scales. For large-scale datasets ImageNet-1k, we compare TESLA [7], SRe\\(^2\\)L [46], D\\(^4\\)M [41], RDED [42], and MiniMax Diffusion [18] (abbreviated as MiMxDiff in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10561" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.312 + ], + "angle": 0, + "content": "
DatasetIPCMethodAccuracy(%)
ImageNet-1K1000Full dataset69.8
10TESLA7.7
SRe2L21.3
D4M27.9
RDED42.0
MiMxDiff44.3
UniDD (Ours)50.4 (↑6.1)
50SRe2L46.8
D4M55.2
RDED56.5
MiMxDiff58.6
UniDD (Ours)62.8 (↑4.2)
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.322, + 0.483, + 0.406 + ], + "angle": 0, + "content": "Table 1. Performance comparison with state-of-the-art methods on large-scale dataset ImageNet-1K. All methods employ ResNet18 as test model. All standard deviations in this table are less than one. The results of Full dataset come from the official PyTorch. The best results are marked in bold, while the second-best results are underlined." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.483, + 0.478 + ], + "angle": 0, + "content": "the table). For small-scale dataset ImageNet-Subset, we compare Random selection, DM [49], MTT [3], HaBa [29], DataDAM [38], and MiniMax Diffusion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Quantitative comparison on ImageNet-1K and ImageNet-Subset. As shown in the Tab 1, we utilize ResNet-18 as the \\(\\mathcal{T}_{\\mathcal{SP}}\\) model, trained on ImageNet-1K. We report results for IPC-10 and IPC-50, evaluating all outcomes with the ResNet-18 model. To ensure a fair comparison, our training and testing pipeline is aligned with that of RDED and \\(\\mathrm{D^4M}\\). Our method, UniDD, achieves state-of-the-art performance on this large-scale dataset. Specifically, at IPC-10, UniDD surpasses optimization-based methods, delivering substantial gains over TESLA and \\(\\mathrm{SRe}^2\\mathrm{L}\\). Additionally, it demonstrates significant advantages over diffusion-based methods, outperforming \\(\\mathrm{D^4M}\\) by \\(22.5\\%\\) and MiniMax Diffusion by \\(6.1\\%\\). These results underscore UniDD's effectiveness by using a lightweight \\(\\mathcal{T}_{\\mathcal{SP}}\\) model to guide SD denoising, aligning generative and target domains without requiring diffusion model finetuning, thus advancing performance in dataset distillation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.903 + ], + "angle": 0, + "content": "As shown in the Tab 2, we conduct experiments on the ImageNet-subset under the IPC-10 setting. For consistency, the test model is standardized to a 5-layer ConvNet across all methods, except for MiniMax Diffusion, which uses a 6-layer ConvNet. Our method stands out among the comparison methods, outperforming the second-best by \\(8.2\\%\\) on ImageWoof, \\(9.1\\%\\) on ImageFruit, and \\(6.6\\%\\) on ImageMeow. Additionally, our method surpasses the diffusion-based MiniMax Diffusion by \\(9.8\\%\\) on ImageWoof, further demonstrating the effectiveness of UniDD approach in ag" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.829, + 0.107 + ], + "angle": 0, + "content": "gregating information from the original dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.108, + 0.905, + 0.244 + ], + "angle": 0, + "content": "Qualitative Result. In Fig. 2 (a), we present the synthetic images generated by UniDD on ImageWoof. Benefiting from the powerful generative capability of SD and the use of a \\(\\mathcal{T}_{\\mathcal{SP}}\\) model to align the real and synthetic data domains, our synthetic images for the classification task exhibit significantly higher realism compared to optimization-based methods [3, 46, 50]. However, unlike other diffusion-based dataset distillation methods [18, 41], we do not finetune SD, greatly reducing the deployment time of UniDD." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.257, + 0.849, + 0.273 + ], + "angle": 0, + "content": "4.3. Results on Detection and Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.28, + 0.905, + 0.43 + ], + "angle": 0, + "content": "We deploy our UniDD framework for detection and segmentation tasks and conduct experiments on the Pascal VOC and MS COCO datasets. Since dataset distillation is rarely applied to these tasks, we reference earlier dataset distillation works (DD [44], DC [50]) and compare UniDD with core-set selection methods. Specifically, we construct the following baseline methods: Random [35], K-center [39], and Herding [2, 5]. Additionally, we develop a method that selects a subset with balanced class representation, which we refer to as Uniform." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.905, + 0.734 + ], + "angle": 0, + "content": "Quantitative Result. In Table 3, we compare coreset selection methods with UniDD across different compression ratios on the object detection task. According to the official MMDetection implementation [4], the detector Faster R-CNN [36] achieves \\(51.4\\%\\) mAP and \\(80.3\\%\\) \\(\\mathrm{AP}_{50}\\) on Pascal VOC with the full dataset, and \\(32.6\\%\\) mAP and \\(51.4\\%\\) \\(\\mathrm{AP}_{50}\\) on MS COCO, providing an approximate performance upper bound. On Pascal VOC, UniDD demonstrates significant performance gains, achieving \\(7.7\\%\\) mAP and \\(19.2\\%\\) \\(\\mathrm{AP}_{50}\\) higher than the Random method at an extremely low compression ratio of \\(0.5\\%\\). Additionally, UniDD surpasses Uniform by \\(11.1\\%\\) mAP and \\(21.2\\%\\) \\(\\mathrm{AP}_{50}\\) at a \\(1\\%\\) compression ratio. On the more challenging MS COCO dataset, UniDD also shows a clear advantage, surpassing the best coreset selection method by \\(3.7\\%\\), \\(3.4\\%\\), and \\(3.4\\%\\) mAP at compression ratios of \\(0.25\\%\\), \\(0.5\\%\\), and \\(1\\%\\), respectively. This indicates UniDD's robustness, achieving strong results even on more complex datasets, which can be attributed to the Task Knowledge Mining in the first stage that extracts sufficient task-relevant information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In Table 4, we present experimental results on the VOC dataset for the semantic segmentation task. Based on the MMSegmentation [6] implementation, LR-ASPP with a MobileNetV3 backbone achieves \\(32.9\\%\\) mIoU on the full dataset. UniDD outperforms coreset selection methods at all compression ratios, demonstrating its effectiveness in dataset distillation for segmentation tasks. This result further underscores UniDD's versatility across different tasks. By leveraging task-driven diffusion synthesis alongside task knowledge mining, UniDD generates synthetic data well-suited to complex, pixel-level predictions, illustrating its" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10562" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.101, + 0.089, + 0.897, + 0.19 + ], + "angle": 0, + "content": "
DatasetRandomDMMTTHaBaDataDAMMiMxDiffUniDD (Ours)Full
ImageNette47.7±2.458.1±0.363.0 ±1.364.7 ±1.659.4±0.462.0±0.266.0±0.787.4 ±1.0
ImageWoof27.0±1.931.4±0.535.8 ±1.838.6 ±1.234.4±0.437.0±1.046.8±0.967.0 ±1.3
ImageFruit21.4±1.2-40.3 ±1.342.5 ±1.6--51.6±0.563.9 ±2.0
ImageMeow29.0±1.1-40.4 ±2.242.9 ±0.9--49.5±0.766.7 ±1.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.202, + 0.908, + 0.232 + ], + "angle": 0, + "content": "Table 2. Performance comparison with other state-of-the-art methods on ImageNet-Subset. IPC is set to 10. The best results are marked in bold, while the second-best results are underlined. “-” were not officially implemented in the original papers." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.254, + 0.481, + 0.564 + ], + "angle": 0, + "content": "
MethodsObject Detection
Pascal VOCMS COCO
mAPAP50mAPAP50
Ratio0.5%0.25%
Random0.8±0.23.1±0.40.5±0.11.7±0.3
Uniform0.9±0.13.4±0.30.8±0.22.4±0.5
K-Center0.5±0.12.1±0.30.4±0.11.5±0.2
Herding0.6±0.22.4±0.20.5±0.11.8±0.4
UniDD (Ours)8.5±0.422.3±0.64.5±0.310.3±0.4
Ratio1%0.5%
Random4.2±0.513.7±0.63.7±0.210.1±0.3
Uniform5.7±0.217.7±0.43.4±0.49.5±0.6
K-Center3.6±0.612.3±0.33.2±0.59.3±0.5
Herding3.5±0.511.9±0.53.5±0.39.7±0.3
UniDD (Ours)16.8±0.538.9±0.77.1±0.416.9±0.3
Ratio2%1%
Random12.4±0.434.3±0.57.2±0.817.3±0.9
Uniform13.8±0.336.2±0.47.4±0.517.6±0.5
K-Center10.9±0.629.3±0.66.1±0.315.4±0.6
Herding10.4±0.428.7±0.76.7±0.416.3±0.7
UniDD (Ours)23.9±0.548.5±0.610.8±0.422.5±0.5
Full51.4±0.880.3±0.432.6±0.751.4±0.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.574, + 0.483, + 0.63 + ], + "angle": 0, + "content": "Table 3. Performance comparison with coreset selection methods on Pascal VOC and MS COCO for the object detection task, using Faster R-CNN as the detector. \"Ratio\" denotes the compression ratio." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.647, + 0.481, + 0.743 + ], + "angle": 0, + "content": "
MethodsSemantic Segmentation (mIoU)
Ratio: 3.5%Ratio: 7%Ratio: 14%Full
Random6.7±0.49.6±0.511.3±0.4
Uniform6.5±0.69.3±0.412.1±0.5
K-Center6.1±0.48.5±0.610.4±0.732.9±0.6
Herding6.2±0.78.7±0.410.6±0.3
UniDD (Ours)10.3±0.512.3±0.314.9±0.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.753, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Table 4. Performance comparison with coreset selection methods on Pascal VOC for the semantic segmentation task, using LR-ASPP as the segmenter. The metric is mIoU, and \"Ratio\" denotes the compression ratio." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.839, + 0.361, + 0.854 + ], + "angle": 0, + "content": "versatility across a range of vision tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Qualitative Result. As shown in Fig. 2 (b) and 2 (c), we present visualizations of detection and segmentation data synthesized using the UniDD framework. These images" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.257, + 0.907, + 0.422 + ], + "angle": 0, + "content": "are generated with \\(\\mathcal{T}_{\\mathcal{SP}}\\) models, Faster R-CNN for detection and LR-ASPP for segmentation, which effectively steer the diffusion process toward the specified labels, bounding boxes, and masks. Previous research in dataset distillation has predominantly focused on single-category label synthesis, where each image aligns with a single class. In contrast, our UniDD framework overcomes this limitation, enabling dataset distillation for more complex, multi-class tasks. This advancement not only broadens the applicability of dataset distillation but also offers valuable insights for future research in this domain." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.434, + 0.667, + 0.451 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.909, + 0.7 + ], + "angle": 0, + "content": "Component Analysis. We conduct a comprehensive ablation study on the proposed UniDD, analyzing its components: (1) Baseline (a standard SD v1.5 checkpoint), (2) Task-driven Diffusion with the \\(\\mathcal{T}_{\\mathcal{SP}}\\) model, (3) Task-Aligned Contextual Prompting (abbreviated as Context), and (4) Proxy-Driven High-Realism Refinement (abbreviated as Refine). As shown in Table 5, the proposed task-driven approach significantly improves the performance of the baseline, demonstrating that this guidance effectively bridges the gap between real and synthetic datasets. The contextual information is more effective for more complex detection and segmentation datasets, as it reveals underlying target relationships. The refinement module also proves to be effective, as the randomness of the SD model can lead to the generation of images that deviate from the annotations or result in failed generations." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.713, + 0.905, + 0.788 + ], + "angle": 0, + "content": "
MethodsDataset
Cls. / Accuracy (%)Det. / mAP (%)Seg. / mIoU (%)
ImageNetteImageWoofVOCVOC
Baseline60.4±0.939.1±1.111.5±0.27.2±0.3
Baseline+TSP64.2±0.845.3±0.614.7±0.48.9±0.5
Baseline+TSRP+Context64.7±0.446.1±0.715.9±0.39.6±0.4
Baseline+TSP+Context+Refine66.0±0.746.8±0.916.8±0.510.3±0.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.798, + 0.907, + 0.868 + ], + "angle": 0, + "content": "Table 5. Ablation study on UniDD. For classification datasets, IPC is set to 10 with ResNet-18 as the test model. For detection on VOC, the compression ratio is set to \\(1\\%\\) with Faster R-CNN as the test model. For segmentation on VOC, the ratio is set to \\(3.5\\%\\) with LR-ASPP as the test model." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Cross-Architecture Analysis. Cross-Architecture experi" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10563" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.482, + 0.208 + ], + "angle": 0, + "content": "
DatasetTest Model\\( T_{SP} \\)Model
ResNet-18ResNet-50ResNet-101
ImageNetteResNet-1866.0±0.768.8±0.769.5±0.7
ResNet-5067.2±0.669.5±1.170.7±0.9
ResNet-10169.3±0.970.9±1.272.1±0.6
ImageWoofResNet-1846.8±0.949.2±0.550.1±0.8
ResNet-5047.3±0.750.0±0.850.9±0.6
ResNet-10148.9±1.151.5±0.652.4±0.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.218, + 0.483, + 0.273 + ], + "angle": 0, + "content": "Table 6. \\( \\mathcal{T}_{\\mathcal{SP}} \\) Model Ablation and Cross-Architecture Testing: We use ResNet-18, ResNet-50, and ResNet-101 as \\( \\mathcal{T}_{\\mathcal{SP}} \\) models to generate synthetic images. During testing, the same three models are used. IPC is set to 10." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.299, + 0.273, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.431, + 0.276, + 0.442 + ], + "angle": 0, + "content": "(a) Accuracy vs. Guidance strength" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.299, + 0.462, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.431, + 0.452, + 0.442 + ], + "angle": 0, + "content": "(b) Accuracy vs. Sample Step" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.455, + 0.483, + 0.496 + ], + "angle": 0, + "content": "Figure 4. Hyper-parameter analysis on (a) guidance strength \\(\\lambda_{s}\\) and (b) sampling step \\(t\\). The star symbol indicates the parameters chosen for trade-off." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.519, + 0.278, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.657, + 0.26, + 0.668 + ], + "angle": 0, + "content": "(a) Training Time overhead" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.519, + 0.465, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.32, + 0.657, + 0.446, + 0.668 + ], + "angle": 0, + "content": "(b) Training GPU overhead" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.68, + 0.483, + 0.723 + ], + "angle": 0, + "content": "Figure 5. Comparison of distillation cost on ImageNet-1k using our UniDD with comparison methods (MTT [3], \\(\\mathrm{SRe}^2\\mathrm{L}\\) [46], and \\(\\mathrm{D^4M}\\) [41]), tested on V100 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "ments are crucial for validating the generalizability of synthetic data. We compare different combinations of \\(\\mathcal{T}_{\\mathcal{SP}}\\) models and testing models. In this study, we set up three \\(\\mathcal{T}_{\\mathcal{SP}}\\) models: ResNet-18, ResNet-50, and ResNet-101. These models are also used during testing to evaluate the cross-architecture performance of the synthetic datasets. As shown in Table. 6, using more complex \\(\\mathcal{T}_{\\mathcal{SP}}\\) models improves the guidance effectiveness. Moreover, when testing across different architectures, the performance does not degrade as seen in previous optimization-based meth" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "ods [3, 49, 50], demonstrating the generalizability of our synthetic data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.905, + 0.227 + ], + "angle": 0, + "content": "Guidance Strength Analysis. As shown in Fig. 4 (a), we analyze the performance variations of ImageWoof and ImageNette at IPC-10 under different guidance strength settings \\(\\lambda_{s}\\), with Conv5 as the test model. The best results are achieved at \\(\\lambda_{s} = 100\\), but further increasing the guidance strength interferes with the quality of diffusion-generated images, leading to a decline in test performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.905, + 0.346 + ], + "angle": 0, + "content": "Sample Step Analysis. As shown in Fig. 4 (b), we analyze the impact of different SD sample steps on the performance of synthetic datasets under the IPC-10 setting, with Image-Woof and ImageNette resolutions set to \\(256 \\times 256\\). Performance improves significantly with sample steps below 250, but gains are minimal beyond this point while time consumption increases. Thus, we select 250 steps as the optimal trade-off." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.349, + 0.907, + 0.561 + ], + "angle": 0, + "content": "Distillation Cost Analysis. As shown in Fig. 5, we evaluate the time and GPU cost of various DD methods. Previous methods primarily focus on the time required to synthesize images, often neglecting the time needed to train or deploy the DD model. For MTT [3], the deployment time corresponds to the expert trajectory training time, while for \\(\\mathrm{D^4M}\\) [41], it is the time required to finetune the SD model. The slight increase in UniDD's distillation time and GPU cost is mainly due to the added supervision process of the \\(\\mathcal{T}_{\\mathcal{SP}}\\) model and the refinement stage, where approximately \\(5\\% -10\\%\\) of low-quality generated data is discarded. Combining the deployment and distillation costs for comparison, UniDD demonstrates the best advantage, indicating its superior suitability for practical deployment." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.574, + 0.634, + 0.589 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.87 + ], + "angle": 0, + "content": "In this study, we introduce the first Universal Dataset Distillation (UniDD) framework, capable of handling classification, detection, and segmentation tasks within a unified approach. UniDD follows a two-stage process: in the first stage, Task-Specific Proxy Training and Task-Aligned Contextual Prompting extract critical information from real datasets. In the second stage, Universal Taskdriven Diffusion synthesizes images with specific categories, bounding boxes, and segmentation masks tailored to each task. We conduct extensive experiments to validate UniDD's effectiveness across all three tasks. For classification, UniDD achieves state-of-the-art results on ImageNet1k and ImageNet-Subset. In detection and segmentation, we establish the first dataset distillation benchmark on Pascal VOC and MS COCO, comparing UniDD to coreset selection methods. Encouraging results demonstrate that our UniDD consistently outperforms these methods, especially under extremely low compression ratios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported by National Natural Science Fund of China (62473286)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10564" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.197 + ], + "angle": 0, + "content": "[1] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 843-852, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.2, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European conference on computer vision (ECCV), pages 233-248, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.483, + 0.325 + ], + "angle": 0, + "content": "[3] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A Efros, and Jun-Yan Zhu. Dataset distillation by matching training trajectories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4750-4759, 2022. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.326, + 0.484, + 0.435 + ], + "angle": 0, + "content": "[4] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.437, + 0.483, + 0.478 + ], + "angle": 0, + "content": "[5] Yutian Chen, Max Welling, and Alex Smola. Super-samples from kernel herding. arXiv preprint arXiv:1203.3472, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.48, + 0.483, + 0.535 + ], + "angle": 0, + "content": "[6] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.536, + 0.483, + 0.59 + ], + "angle": 0, + "content": "[7] Justin Cui, Ruochen Wang, Si Si, and Cho-Jui Hsieh. Scaling up dataset distillation toImagenet-1k with constant memory. In International Conference on Machine Learning, pages 6565–6590. PMLR, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.592, + 0.483, + 0.647 + ], + "angle": 0, + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.483, + 0.69 + ], + "angle": 0, + "content": "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[10] Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. Advances in Neural Information Processing Systems, 36:16222-16239, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[12] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. International journal of computer vision, 111:98-136, 2015. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.903 + ], + "angle": 0, + "content": "[13] Junyao Gao, Xinyang Jiang, Huishuai Zhang, Yifan Yang, Shuguang Dou, Dongsheng Li, Duoqian Miao, Cheng Deng," + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "and Cairong Zhao. Similarity distribution based membership inference attack on person re-identification. In Proceedings of the AAAI conference on artificial intelligence, pages 14820-14828, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.216 + ], + "angle": 0, + "content": "[14] Junyao Gao, Xinyang Jiang, Shuguang Dou, Dongsheng Li, Duoqian Miao, and Cairong Zhao. Re-id-leak: Membership inference attacks against person re-identification. International Journal of Computer Vision, 132(10):4673-4687, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[15] Junyao Gao, Yanchen Liu, Yanan Sun, Yinhao Tang, Yanhong Zeng, Kai Chen, and Cairong Zhao. Styleshot: A snapshot on any style. arXiv preprint arXiv:2407.01414, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[16] Junyao Gao, Yanan Sun, Fei Shen, Xin Jiang, Zhening Xing, Kai Chen, and Cairong Zhao. Faceshot: Bring any character into life. arXiv preprint arXiv:2503.00740, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.372 + ], + "angle": 0, + "content": "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.372, + 0.907, + 0.44 + ], + "angle": 0, + "content": "[18] Jianyang Gu, Saeed Vahidian, Vyacheslav Kungurtsev, Haonan Wang, Wei Jiang, Yang You, and Yiran Chen. Efficient dataset distillation via minimax diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15793-15803, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[19] Ziyao Guo, Kai Wang, George Cazenavette, Hui Li, Kaipeng Zhang, and Yang You. Towards lossless dataset distillation via difficulty-aligned trajectory matching. arXiv preprint arXiv:2310.05773, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.608 + ], + "angle": 0, + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.676 + ], + "angle": 0, + "content": "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.706 + ], + "angle": 0, + "content": "[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[24] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1314-1324, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[25] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Jang-Hyun Kim, Jinuk Kim, Seong Joon Oh, Sangwoo Yun, Hwanjun Song, Joonhyun Jeong, Jung-Woo Ha, and Hyun Oh Song. Dataset condensation via efficient synthetic-data parameterization. In International Conference on Machine Learning, pages 11102-11118. PMLR, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10565" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[27] Jian Li, Yabiao Wang, Changan Wang, Ying Tai, Jianjun Qian, Jian Yang, Chengjie Wang, Jilin Li, and Feiyue Huang. Dsfd: dual shot face detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5060-5069, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[29] Songhua Liu, Kai Wang, Xingyi Yang, Jingwen Ye, and Xin chao Wang. Dataset distillation via factorization. Advances in Neural Information Processing Systems, 35:1100-1113, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.301, + 0.483, + 0.384 + ], + "angle": 0, + "content": "[30] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 289-299, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.438 + ], + "angle": 0, + "content": "[31] Yanqing Liu, Jianyang Gu, Kai Wang, Zheng Zhu, Wei Jiang, and Yang You. Dream: Efficient dataset distillation by representative matching. arXiv preprint arXiv:2302.14416, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.441, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[32] Timothy Nguyen, Zhourong Chen, and Jaehoon Lee. Dataset meta-learning from kernel ridge-regression. In International Conference on Learning Representations, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[33] OpenAI. Gpt-4 technical report. https://openai.com/research/gpt-4, 2023. Accessed: 2024-11-13. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[34] Ding Qi, Jian Li, Jinlong Peng, Bo Zhao, Shuguang Dou, Jialin Li, Jiangning Zhang, Yabiao Wang, Chengjie Wang, and Cairong Zhao. Fetch and forge: Efficient dataset condensation for object detection. Advances in Neural Information Processing Systems, 37:119283-119300, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[35] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[36] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[38] Ahmad Sajedi, Samir Khaki, Ehsan Amjadian, Lucy Z Liu, Yuri A Lawryshyn, and Konstantinos N Plataniotis. Datadam: Efficient dataset distillation with attention matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17097-17107, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[39] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. arXiv preprint arXiv:1708.00489, 2017. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[41] Duo Su, Junjie Hou, Weizhi Gao, Yingjie Tian, and Bowen Tang. D^4: Dataset distillation via disentangled diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5809-5818, 2024. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[42] Peng Sun, Bei Shi, Daiwei Yu, and Tao Lin. On the diversity and realism of distilled dataset: An efficient dataset distillation paradigm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9390-9399, 2024. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[43] Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Shuo Yang, Shuo Wang, Guan Huang, Hakan Bilen, Xinchao Wang, and Yang You. Cafe: Learning to condense dataset by aligning features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12196-12205, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[44] Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A Efros. Dataset distillation. arXiv preprint arXiv:1811.10959, 2018. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.403, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[45] Yichao Yan, Zanwei Zhou, Zi Wang, Jingnan Gao, and Xiaokang Yang. Dialoguenerf: Towards realistic avatar face-to-face conversation video generation. Visual Intelligence, 2 (1):24, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.514 + ], + "angle": 0, + "content": "[46] Zeyuan Yin, Eric Xing, and Zhiqiang Shen. Squeeze, recover and relabel: Dataset condensation at imagenet scale from a new perspective. arXiv preprint arXiv:2306.13092, 2023. 1, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.517, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[47] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.574, + 0.905, + 0.614 + ], + "angle": 0, + "content": "[48] Bo Zhao and Hakan Bilen. Dataset condensation with differentiable siamese augmentation. In International Conference on Machine Learning, pages 12674-12685. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.616, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[49] Bo Zhao and Hakan Bilen. Dataset condensation with distribution matching. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6514-6523, 2023. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.673, + 0.905, + 0.714 + ], + "angle": 0, + "content": "[50] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. Dataset condensation with gradient matching. In International Conference on Learning Representations, 2020. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.716, + 0.905, + 0.782 + ], + "angle": 0, + "content": "[51] Ganlong Zhao, Guanbin Li, Yipeng Qin, and Yizhou Yu. Improved distribution matching for dataset condensation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7856-7865, 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.782 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10566" + } + ] +] \ No newline at end of file diff --git a/2025/Towards Universal Dataset Distillation via Task-Driven Diffusion/cd8b4c9e-1814-423f-bdad-2b631dc6a75a_origin.pdf b/2025/Towards Universal Dataset Distillation via Task-Driven Diffusion/cd8b4c9e-1814-423f-bdad-2b631dc6a75a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e547dffde54f8680697f06d59757ae3fa5fcb01c --- /dev/null +++ b/2025/Towards Universal Dataset Distillation via Task-Driven Diffusion/cd8b4c9e-1814-423f-bdad-2b631dc6a75a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:84e0686f3b43bafc7fb6a1a9330f093437069b096dd172aebee39099a1f7099d +size 4585278 diff --git a/2025/Towards Universal Dataset Distillation via Task-Driven Diffusion/full.md b/2025/Towards Universal Dataset Distillation via Task-Driven Diffusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8f64db62b5b045084f4320ac8b97e69c0149cd3b --- /dev/null +++ b/2025/Towards Universal Dataset Distillation via Task-Driven Diffusion/full.md @@ -0,0 +1,338 @@ +# Towards Universal Dataset Distillation via Task-Driven Diffusion + +Ding Qi $^{1*}$ Jian Li $^{2\dagger}$ Junyao Gao $^{1}$ Shuguang Dou $^{1}$ Ying Tai $^{5}$ +Jianlong Hu $^{2}$ Bo Zhao $^{4}$ Yabiao Wang $^{3,2\dagger}$ Chengjie Wang $^{4,2}$ Cairong Zhao $^{1\dagger}$ $^{1}$ Tongji University $^{2}$ Tencent YouTu Lab $^{3}$ Zhejiang University + $^{4}$ Shanghai Jiao Tong University $^{5}$ Nanjing University + +![](images/273080066be2f11e90e6d1d4791effb0af3c1dd2d9a9a35dc7bff1253d4caa6d.jpg) +(a) One Universal Dataset Distillation Framework for Mutli-Tasks + +![](images/29377090049304128148076e6e350a1f9d66a781866c0624fbfc3fec438bfb75.jpg) +(b) Improved Performance by UniDD +Figure 1. (a) We introduce the Universal Dataset Distillation Framework (UniDD), a novel dataset distillation framework supporting a range of vision tasks, including classification, object detection, and segmentation. UniDD incorporates diverse label types, such as class labels, bounding box coordinates, and pixel-level masks, facilitating the synthesis of task-specific datasets. (b) UniDD improves the performance of dataset distillation across diverse tasks on ImageNet [8], Pascal VOC [11, 12], and MS COCO [28]. + +# Abstract + +Dataset distillation (DD) condenses key information from large-scale datasets into smaller synthetic datasets, reducing storage and computational costs for training networks. However, most recent research has primarily focused on image classification tasks, with limited exploration in detection and segmentation. Two key challenges remain: (i) Task Optimization Heterogeneity, where existing methods focus on class-level information but fail to address the diverse needs of detection and segmentation, and (ii) Inflexible Image Generation, where current generation methods rely on global updates for single-class targets and lack localized optimization for specific object regions. To address these challenges, we propose UniDD, a universal dataset distillation framework built on a task-driven diffusion model for diverse DD tasks, as shown in Fig. 1. Our approach operates in two stages: Universal Task Knowledge Mining, which captures task-relevant information through task + +specific proxy model training, and Universal Task-Driven Diffusion, where these proxies guide the diffusion process to generate task-specific synthetic images. Extensive experiments across ImageNet-1K, Pascal VOC, and MS COCO demonstrate that UniDD consistently outperforms state-of-the-art methods. In particular, on ImageNet-1K with IPC-10, UniDD surpasses previous diffusion-based methods by $6.1\%$ , while also reducing deployment costs. + +# 1. Introduction + +Dataset distillation (DD) [44], as an innovative data optimization technique in machine learning [13, 14, 20, 22, 45], condenses essential information from large datasets into smaller proxy sets. Training on these highly ensemble proxy sets enables the model to use much fewer resources while providing comparable performance. Prior DD methods excelled in image classification using various techniques like Gradient Matching [26, 48, 50], Distribution Matching [49, 51], Training Trajectory Matching [3, 7], and Decoupled Methods [42, 46]. + +Despite advancements in DD, research on key visual tasks—detection and segmentation—remains limited. We identify the following key challenges: (i) Task Optimization Heterogeneity. Existing DD methods are primarily focused on image classification, optimizing distillation by compressing class-level information, but they fall short in addressing the diverse needs of tasks like detection [27] and segmentation, such as object localization and pixel-level information. (ii) Inflexible Image Generation. Current image generation methods rely on global updates for single-class targets and lack localized optimization for specific object regions. These limitations become especially evident in images with multiple objects or complex structures, hindering the ability of existing DD to generate distilled data that meets the requirements of diverse tasks. Consequently, the applicability of DD in the visual domain remains constrained. + +To address these limitations, we propose a Universal Dataset Distillation (UniDD) framework that optimizes for multi-task learning, including classification, detection, and segmentation, overcoming the shortcomings in task adaptation and image generation. Our approach consists of two stages: Universal Task Knowledge Mining and Universal Task-Driven Diffusion. In the first stage, we introduce Task-Specific Proxy $(\mathcal{T}_{\mathcal{SP}})$ models, trained on real datasets to extract task-specific information. This allows the $\mathcal{T}_{\mathcal{SP}}$ models to capture and store category information for classification, bounding box positions for detection, and mask data for segmentation within their parameters. Additionally, previous research has often overlooked the importance of contextual information, particularly in high-resolution datasets like ImageNet [8], Pascal VOC [11, 12] and MS COCO [28], where rich object-context relationships exist. To address this, we introduce Task-Aligned Contextual Prompting, which uses a Vision-Language model [33] to generate contextual prompts for real images. In the second stage, we introduce a Task-Driven Diffusion for image synthesis. Specifically, the $\mathcal{T}_{\mathcal{SP}}$ models trained in the first stage supervise the denoising process of the diffusion model by calculating the task loss on the predicted clean images. This guides the diffusion model in generating images that match the given labels. Notably, since the $\mathcal{T}_{\mathcal{SP}}$ models already align with the target domain, there is no need to retrain the diffusion model on the real set, significantly reducing the cost in practical applications. + +In summary, our contributions are threefold: + +- We identify key challenges in extending dataset distillation beyond classification and propose the first universal framework for DD across classification, detection, and segmentation tasks. +- We develop a two-stage approach, Universal Task Knowledge Mining and Universal Task-Driven Diffusion for Data Synthesis, which flexibly synthesizes images for + +multi-tasks. Additionally, we address the previously overlooked need to align potential contextual information within datasets. + +- Our method achieves state-of-the-art performance in classification tasks, on the ImageNet-1K dataset, our approach surpasses the previous state-of-the-art method by over $6.1\%$ . Additionally, we establish new benchmarks for detection and segmentation on VOC and COCO, showing significant improvements over coreset methods. + +# 2. Related works + +# 2.1. Dataset Distillation + +Dataset distillation [44], a dataset reduction method, addresses the challenge of handling substantial data by synthesizing a smaller, representative dataset. Existing dataset distillation methods can be categorized into meta-learning and data matching frameworks based on whether they explicitly mimic the performance of target data. In the meta-learning framework, the distilled data are treated as hyperparameters and optimized in a nested loop fashion according to the trained model's risk. For example, KIP [32] performs kernel ridge regression using the Neural Tangent Kernel (NTK) [25]. The data matching framework includes gradient matching, distribution matching, and trajectory matching. These methods update distilled data by imitating the influence of target data. For example, DC [50] and DM [49] enforce synthetic and real images to have similar gradients or distribution at every training step. MTT [3] proposes to imitate long-range trajectories of optimization steps for real images, thereby better mimicking the training dynamics of real images for synthetic images. However, the aforementioned gradient matching methods require a large number of iterations. To address this, DREAM [31] and IDC [26] focus on more efficient dataset distillation training. Due to the unsatisfactory performance of DM, CAFE [43] and IDM [51] propose improved distribution matching between the real and synthetic data. For the trajectory matching method, DATM [19] scales trajectory matching-based methods to larger synthetic datasets, achieving lossless dataset distillation. + +Despite significant progress in previous work, there have been few breakthroughs in dataset distillation for detection and segmentation tasks. In this work, we introduce the first universal framework that simultaneously handles classification, object detection, and semantic segmentation, significantly expanding the generality and scope of dataset distillation research. + +# 2.2. Diffusion Models + +Diffusion models [15, 16, 37, 47] are generative models that show superior fidelity and diversity compared to GANs [17]. It aims to learn the mapping from a Gaussian + +![](images/8c94ce46423189613163e117cc42deebceb0535c1c09b9fdc51f7a33944a2e6a.jpg) +(a) Synthetic images for classification. + +![](images/6a727fa95fd69b053327cd236578d7e039dc29ffce88b311fa0bedf52568970d.jpg) +(b) Synthetic images for detection. + +![](images/b749891040b76d77d0cb4ecba985295e6e5ca13b40b01681cb87fd578d07b433.jpg) +(c) Synthetic images for segmentation. +Figure 2. Visualization of synthetic images generated by UniDD. (a) For classification, images are synthesized on the ImageWoof [8] dataset. (b) For detection, images are synthesized on the Pascal VOC dataset with specified bounding box coordinates. (c) For segmentation, images are synthesized on the Pascal VOC [11, 12] dataset with pixel-level masks. Our UniDD is the first approach to enable high-realism data distillation across classification, detection, and segmentation tasks. + +distribution to the real dataset distribution by adding noise in the diffusion process and reversing the noise to the raw image in the denoising process. In the forward process, diffusion models progressively add noise to a clean sample $z_0$ , resulting in a noisy sample $z_t$ : + +$$ +z _ {t} = \sqrt {\alpha_ {t}} z _ {0} + \sqrt {1 - \alpha_ {t}} \epsilon , \quad \epsilon \sim \mathcal {N} (0, I). \tag {1} +$$ + +In the reverse process, a denoising network $\epsilon_{\theta}$ estimates the noise and iteratively recovers the clean sample $z_0$ : + +$$ +\epsilon_ {\theta} (z _ {t}, t) \approx \frac {z _ {t} - \sqrt {\alpha_ {t}} z _ {0}}{\sqrt {1 - \alpha_ {t}}}. \tag {2} +$$ + +DDIM [40] provides the following formula for predicting a clean data sample: + +$$ +\hat {z} _ {0} = \frac {z _ {t} - \sqrt {1 - \alpha_ {t}} \epsilon_ {\theta} (z _ {t} , t)}{\sqrt {\alpha_ {t}}}. \tag {3} +$$ + +This formula predicts the clean data point $\hat{z}_0$ from the noisy sample $z_{t}$ , enabling deterministic sampling for efficient and high-quality image synthesis. + +To introduce additional conditions into the diffusion model, classifier guidance [9] has been proposed to train an additional classifier on the noise to guide the model to generate images that are closer to the specified condition. Additionally, [23] introduces classifier-free guidance that trains both an unconditioned and conditioned model simultaneously, achieving more efficient guidance. Following these, recent methods [1, 10, 30] have been proposed to generate high-quality images from diverse conditions. + +# 3. Universal Dataset Distillation Framework + +# 3.1. Problem Definition + +Universal Dataset Distillation. Current dataset distillation research primarily targets classification tasks and lacks + +a generalizable approach for detection and segmentation. To address this, we propose a universal dataset distillation standard that constructs a compact synthetic dataset $\mathcal{S} = (\hat{x}_1, \hat{y}_1), \dots, (\hat{x}_{|\mathcal{S}|}, \hat{y}_{|\mathcal{S}|})$ from a larger real dataset $\mathcal{T} = (x_1, y_1), \dots, (x_{|\mathcal{T}|}, y_{|\mathcal{T}|})$ , with $|S| \ll |\mathcal{T}|$ . The primary difference between tasks lies in label structure: classification uses single-class labels, detection includes bounding boxes with class information, and segmentation requires pixel-level masks. Our goal is for models trained on the distilled dataset $\mathcal{S}$ to perform comparably to those trained on the full dataset $\mathcal{T}$ across classification, detection, and segmentation tasks, as formalized by: + +$$ +\mathbb {E} _ {(x, y) \sim P _ {D}} \left| \ell \left(\theta_ {\mathcal {T}} ^ {t a s k} (x), y\right) - \ell \left(\theta_ {\mathcal {S}} ^ {t a s k} (x), y\right) \right| \leq \epsilon , \tag {4} +$$ + +where $P_{D}$ denotes the distribution of test data, $\ell$ represents the performance metric specific to each task, $task\in$ {class,detect,segment}, and $\epsilon$ is a small tolerance value. For each task, the parameter set $\theta_{\mathcal{T}}^{task}$ on the real dataset $\mathcal{T}$ is optimized as follows: + +$$ +\theta_ {\mathcal {T}} ^ {t a s k} = \arg \min _ {\theta} \mathbb {E} _ {(x, y) \in \mathcal {T}} \left[ \ell \left(\phi_ {\theta} ^ {t a s k} (x), y\right) \right], \qquad (5) +$$ + +Similarly, $\theta_S^{task}$ is optimized using the synthetic dataset $S$ . + +# 3.2. Universal Task Knowledge Mining + +Task-Specific Proxy Training. Building a universal DD framework is challenging due to the diverse optimization objectives in classification, detection, and segmentation tasks. Recent decoupled DD studies have reached a consensus [34, 42, 46]: a fully trained network can capture and retain essential dataset information. Inspired by this, we extend to general tasks, training a classifier, detector, and segmenter on task-specific datasets to extract critical information—class labels for classification, bounding boxes for + +![](images/697bfc9651c9aab0d9e97c88dab77dad9859a6862ec3e7f7180dc844b1d8405e.jpg) +Figure 3. Universal Dataset Distillation Framework. This framework consists of two stages: in the Universal Task Knowledge Mining stage, we introduce an easily deployable task-specific proxy $(\mathcal{T}_{\mathcal{SP}})$ model trained on the target dataset and a vision-language model to extract task-aligned context. In the Universal Task-Driven Diffusion stage, we apply the gradient of the supervision loss, derived from the $\mathcal{T}_{\mathcal{SP}}$ model's prediction of the clean image $\hat{z}_0$ , at each sampling step. + +detection, and pixel-level masks for segmentation. We call this Task-Specific Proxy $(\mathcal{T}_{\mathcal{SP}})$ Model, defined as: + +$$ +\theta = \arg \min _ {\theta} \mathcal {L} _ {\text {t a s k}} (\mathcal {F} (x), y), \tag {6} +$$ + +where $\mathcal{F}$ represents the $\mathcal{T}_{\mathcal{SP}}$ model trained on the target data, $\theta$ denotes the model's parameters, $\mathcal{L}_{task}$ represents the loss function specific to the task. + +The $\mathcal{T}_{\mathcal{SP}}$ model choice is flexible, allowing adaptation for optimal performance on target datasets. For example, ResNet-18 [21] can serve as a lightweight, fast-deploy option for simple classification, while Faster R-CNN [36] may be used for complex detection tasks to capture key detection information. Unlike previous diffusion-based dataset distillation methods, our approach relies on task-driven learning to align synthetic and real datasets, avoiding costly diffusion model fine-tuning on the target dataset and significantly reducing deployment overhead. + +Task-Aligned Contextual Prompting. Previous diffusion-based dataset distillation has largely focused on classification tasks and overlooked the role of natural language in guiding image generation. For diverse tasks, textual prompts must go beyond describing the primary objects in images. As inputs to Stable Diffusion models, prompts can also guide the generation process. To address this, we use a vision-language model to extract natural language descriptions and design task-specific questions to generate task-aligned prompts, a method we call Task-Aligned Contextual Prompting. Following GPT-4o [33], we create tailored prompts for different tasks. For classification, prompts de + +scribe object classes, while for detection and segmentation, they also include object relationships and the interaction between objects and the background, providing richer contextual guidance. + +# 3.3. Universal Task-Driven Diffusion. + +Inspired by recent advancements in guidance diffusion techniques [1, 9], we propose a Task-Driven Diffusion approach for Dataset Distillation. As illustrated in Fig. 3, during each sampling step $S(z_{t}, t)$ of the diffusion process, the trained $\mathcal{T}_{\mathcal{SP}}$ model $\mathcal{F}$ provides task-specific guidance. Given an label $\tilde{a}$ , the original noise prediction $\epsilon_{\theta}(z_{t}, t, \tilde{c})$ is adjusted using the output of $\mathcal{F}$ . The adjusted noise prediction $\hat{\epsilon}$ is formulated as follows: + +$$ +\hat {\epsilon} _ {\theta} \left(z _ {t}, t, \tilde {c}\right) = \epsilon_ {\theta} \left(z _ {t}, t, \tilde {c}\right) + \sqrt {1 - \alpha_ {t}} \nabla_ {z _ {t}} \mathcal {L} _ {\text {t a s k}} \left(\tilde {a}, \mathcal {F} \left(z _ {t}\right)\right), \tag {7} +$$ + +where $\mathcal{L}_{\mathrm{task}}$ represents the task-specific loss function and $\tilde{a} = \{a_{cl}, a_{bbox}, a_{mask}\}$ may denote a class label $a_{cl}$ , a bounding box $a_{bbox}$ , or a pixel-level segmentation mask $a_{mask}$ . This adjustment mechanism effectively guides the diffusion process to produce data closely aligned with the intended distribution for each task. $\tilde{c}$ is a textual prompt. + +Because $\mathcal{F}$ is trained on clean images, it may not perform well directly on noisy inputs. Therefore, we first use Equation 3 to estimate the clean image $\hat{z}_0$ . Once $\hat{z}_0$ is obtained, we apply $\mathcal{F}$ to compute the task-driven loss as follows: + +$$ +\hat {\epsilon} _ {\theta} \left(z _ {t}, t, \tilde {c}\right)) = \epsilon_ {\theta} \left(z _ {t}, t, \tilde {c}\right) + s (t) \cdot \nabla_ {z _ {t}} \mathcal {L} _ {\text {t a s k}} (\tilde {a}, \mathcal {F} (\hat {z} _ {0})), \tag {8} +$$ + +where $s(t)$ is the task-driven strength at each sampling step, + +defined as: + +$$ +s (t) = \lambda_ {s} \cdot \sqrt {1 - \alpha_ {t}}, \tag {9} +$$ + +with $\lambda_{s}$ being a constant coefficient controlling the overall intensity of the task-driven. The term $\nabla_{z_t}\mathcal{L}_{task}(\tilde{a},\mathcal{F}(\hat{z}_0))$ represents the gradient of the loss function computed based on the $\mathcal{T}_{\mathcal{SP}}$ model's prediction of the clean image $\hat{z}_0$ . Specifically, this gradient is computed as: + +$$ +\nabla_ {z _ {t}} \mathcal {L} _ {\text {t a s k}} (\tilde {a}, \mathcal {F} (\hat {z} _ {0})) = \nabla_ {z _ {t}} \mathcal {L} _ {\text {t a s k}} \left(\tilde {a}, \mathcal {F} \left(\frac {z _ {t} - \sqrt {1 - \alpha_ {t}} \epsilon_ {\theta}}{\sqrt {\alpha_ {t}}}\right)\right). \tag {10} +$$ + +Unlike previous approaches [18, 41] that required finetuning the diffusion model, such as those involving SD [37], to align with new data domains, our method leverages a $\mathcal{T}_{\mathcal{SP}}$ model trained specifically on the target dataset. This model encapsulates critical information about the target task, effectively driving the diffusion process to generate data that is well-aligned with the target task without the need for additional finetuning. This not only conserves time and resources but also mitigates the risk of overfitting the diffusion model to the specific task. + +Proxy-Driven High-Realism Refinement. To ensure the high realism and accuracy of the generated images, we employ a trained $\mathcal{T}_{\mathcal{SP}}$ model, to score each image. By setting a well-defined threshold, we filter out low-quality images based on their scores. The remaining images are then relabeled using the same $\mathcal{T}_{\mathcal{SP}}$ model. This combined process of filtering and relabeling can be expressed as: + +$$ +\{x, y \} = \left\{x _ {i}, \mathcal {F} _ {\text {l a b e l}} \left(x _ {i}\right) \mid \operatorname {s c o r e} \left(\mathcal {F} \left(x _ {i}\right)\right) \geq \tau \right\} \tag {11} +$$ + +where $\mathcal{F}(x_i)$ represents the model's output score for image $x_{i}$ , $\tau$ is the quality threshold, and $\mathcal{F}_{\mathrm{label}}$ generates the accurate label $y_{i}$ for the filtered image. This ensures that both the generated images and their labels are of high quality and closely match the desired data distribution. + +# 4. Experiments + +# 4.1. Experimental Setups + +Datasets and Evaluation Metric. To enhance practicality, we prioritize high-resolution and large-scale datasets. (1) Classification task: our experiments are primarily conducted on the full ImageNet 1K dataset [8], with IPC settings commonly set to 10 and 50, following the configurations in [41]. Additionally, based on the dataset partitioning settings from previous work [3], we also use the ImageNet 10-class subsets, such as ImageWoof, ImageFruit, ImageNette, and ImageMeow. These 10-class subsets contain very similar categories, making them challenging for dataset distillation. Top-1 accuracy $(\%)$ is reported in the experimental tables. (2) Object detection task: we conduct experiments on the widely-used Pascal VOC [11, 12] and MS COCO [28] datasets, using mAP (Mean Average + +# Algorithm 1 Universal Dataset Distillation Approach + +Input: Real data $\mathcal{T}$ , Diffusion model $\epsilon_{\theta}$ , $\mathcal{T}_{\mathcal{SP}}$ model $\mathcal{F}$ , Sampling steps $T$ , Task-Driven strength $s(t)$ + +Output: Synthetic dataset $S$ . + +1: Step I: Universal Task Knowledge Mining +2: Train $\mathcal{T}_{\mathcal{SP}}$ model $\mathcal{F}$ on $\mathcal{T}$ using Equ. 6 +3: Use GPT-4o to generate task-aligned context for classification, detection, and segmentation datasets + +4: Step II: Universal Task-Driven Diffusion for Data Synthetic + +5: for each sampling step $t = 1$ to $T$ do +6: Generate noisy sample $z_{t}$ . +7: Predict clean image $\hat{z}_0$ using Equ. 3 +8: Calculate guidance loss $\nabla_{z_t}\mathcal{L}_{task}(c,\mathcal{F}(\hat{z}_0))$ +9: Adjust noise prediction using Equ. 8 +10: Update $z_{t-1}$ using adjusted noise prediction. +11: end for +12: Score each generated image $x_{i}$ using $\mathcal{F}(x_i)$ . +13: Filter and relabel synthetic set based on threshold $\tau$ using Equ. 11. + +Precision) and $AP_{50}$ (Average Precision at IoU threshold of 0.50) as evaluation metric. (3) Semantic segmentation task: we utilize the Pascal VOC dataset, with mIoU (Mean Intersection over Union) as evaluation metrics. + +Implementation Details. Unless otherwise specified, the Task-Specific Proxy $\mathcal{T}_{\mathcal{SP}}$ model used for the classification task is ResNet-18 [21], the detector is Faster R-CNN [36], and the segmentation model is LR-ASPP [24], all utilizing the official pre-trained weights from PyTorch. For the diffusion model, we use SD v1.5 [37] as the baseline. For quantitative comparisons, we generate $256 \times 256$ images for classification tasks and $512 \times 512$ images for detection and segmentation tasks. Synthetic images are created using 250 denoising steps based on a predefined number of random noise samples, aligned with the IPC settings. The threshold $\tau$ for Proxy-Driven High-Realism Refinement is set to 0.9 for classification, 0.85 for detection, and 0.75 for segmentation. Generally, this operation results in a data discard rate of approximately $5 - 10\%$ . All experiments are conducted on NVIDIA V100 GPUs, with further details provided in the special experiments section. We generate the distilled dataset 3 times, train the network from scratch 5 times per dataset (with pre-trained backbone weights for the detector and segmenter), and evaluate on the original test set, reporting results as $\bar{x} \pm std$ . + +# 4.2. Results on Classification + +We perform comprehensive comparisons across datasets of various scales. For large-scale datasets ImageNet-1k, we compare TESLA [7], SRe $^2$ L [46], D $^4$ M [41], RDED [42], and MiniMax Diffusion [18] (abbreviated as MiMxDiff in + +
DatasetIPCMethodAccuracy(%)
ImageNet-1K1000Full dataset69.8
10TESLA7.7
SRe2L21.3
D4M27.9
RDED42.0
MiMxDiff44.3
UniDD (Ours)50.4 (↑6.1)
50SRe2L46.8
D4M55.2
RDED56.5
MiMxDiff58.6
UniDD (Ours)62.8 (↑4.2)
+ +Table 1. Performance comparison with state-of-the-art methods on large-scale dataset ImageNet-1K. All methods employ ResNet18 as test model. All standard deviations in this table are less than one. The results of Full dataset come from the official PyTorch. The best results are marked in bold, while the second-best results are underlined. + +the table). For small-scale dataset ImageNet-Subset, we compare Random selection, DM [49], MTT [3], HaBa [29], DataDAM [38], and MiniMax Diffusion. + +Quantitative comparison on ImageNet-1K and ImageNet-Subset. As shown in the Tab 1, we utilize ResNet-18 as the $\mathcal{T}_{\mathcal{SP}}$ model, trained on ImageNet-1K. We report results for IPC-10 and IPC-50, evaluating all outcomes with the ResNet-18 model. To ensure a fair comparison, our training and testing pipeline is aligned with that of RDED and $\mathrm{D^4M}$ . Our method, UniDD, achieves state-of-the-art performance on this large-scale dataset. Specifically, at IPC-10, UniDD surpasses optimization-based methods, delivering substantial gains over TESLA and $\mathrm{SRe}^2\mathrm{L}$ . Additionally, it demonstrates significant advantages over diffusion-based methods, outperforming $\mathrm{D^4M}$ by $22.5\%$ and MiniMax Diffusion by $6.1\%$ . These results underscore UniDD's effectiveness by using a lightweight $\mathcal{T}_{\mathcal{SP}}$ model to guide SD denoising, aligning generative and target domains without requiring diffusion model finetuning, thus advancing performance in dataset distillation. + +As shown in the Tab 2, we conduct experiments on the ImageNet-subset under the IPC-10 setting. For consistency, the test model is standardized to a 5-layer ConvNet across all methods, except for MiniMax Diffusion, which uses a 6-layer ConvNet. Our method stands out among the comparison methods, outperforming the second-best by $8.2\%$ on ImageWoof, $9.1\%$ on ImageFruit, and $6.6\%$ on ImageMeow. Additionally, our method surpasses the diffusion-based MiniMax Diffusion by $9.8\%$ on ImageWoof, further demonstrating the effectiveness of UniDD approach in ag + +gregating information from the original dataset. + +Qualitative Result. In Fig. 2 (a), we present the synthetic images generated by UniDD on ImageWoof. Benefiting from the powerful generative capability of SD and the use of a $\mathcal{T}_{\mathcal{SP}}$ model to align the real and synthetic data domains, our synthetic images for the classification task exhibit significantly higher realism compared to optimization-based methods [3, 46, 50]. However, unlike other diffusion-based dataset distillation methods [18, 41], we do not finetune SD, greatly reducing the deployment time of UniDD. + +# 4.3. Results on Detection and Segmentation + +We deploy our UniDD framework for detection and segmentation tasks and conduct experiments on the Pascal VOC and MS COCO datasets. Since dataset distillation is rarely applied to these tasks, we reference earlier dataset distillation works (DD [44], DC [50]) and compare UniDD with core-set selection methods. Specifically, we construct the following baseline methods: Random [35], K-center [39], and Herding [2, 5]. Additionally, we develop a method that selects a subset with balanced class representation, which we refer to as Uniform. + +Quantitative Result. In Table 3, we compare coreset selection methods with UniDD across different compression ratios on the object detection task. According to the official MMDetection implementation [4], the detector Faster R-CNN [36] achieves $51.4\%$ mAP and $80.3\%$ $\mathrm{AP}_{50}$ on Pascal VOC with the full dataset, and $32.6\%$ mAP and $51.4\%$ $\mathrm{AP}_{50}$ on MS COCO, providing an approximate performance upper bound. On Pascal VOC, UniDD demonstrates significant performance gains, achieving $7.7\%$ mAP and $19.2\%$ $\mathrm{AP}_{50}$ higher than the Random method at an extremely low compression ratio of $0.5\%$ . Additionally, UniDD surpasses Uniform by $11.1\%$ mAP and $21.2\%$ $\mathrm{AP}_{50}$ at a $1\%$ compression ratio. On the more challenging MS COCO dataset, UniDD also shows a clear advantage, surpassing the best coreset selection method by $3.7\%$ , $3.4\%$ , and $3.4\%$ mAP at compression ratios of $0.25\%$ , $0.5\%$ , and $1\%$ , respectively. This indicates UniDD's robustness, achieving strong results even on more complex datasets, which can be attributed to the Task Knowledge Mining in the first stage that extracts sufficient task-relevant information. + +In Table 4, we present experimental results on the VOC dataset for the semantic segmentation task. Based on the MMSegmentation [6] implementation, LR-ASPP with a MobileNetV3 backbone achieves $32.9\%$ mIoU on the full dataset. UniDD outperforms coreset selection methods at all compression ratios, demonstrating its effectiveness in dataset distillation for segmentation tasks. This result further underscores UniDD's versatility across different tasks. By leveraging task-driven diffusion synthesis alongside task knowledge mining, UniDD generates synthetic data well-suited to complex, pixel-level predictions, illustrating its + +
DatasetRandomDMMTTHaBaDataDAMMiMxDiffUniDD (Ours)Full
ImageNette47.7±2.458.1±0.363.0 ±1.364.7 ±1.659.4±0.462.0±0.266.0±0.787.4 ±1.0
ImageWoof27.0±1.931.4±0.535.8 ±1.838.6 ±1.234.4±0.437.0±1.046.8±0.967.0 ±1.3
ImageFruit21.4±1.2-40.3 ±1.342.5 ±1.6--51.6±0.563.9 ±2.0
ImageMeow29.0±1.1-40.4 ±2.242.9 ±0.9--49.5±0.766.7 ±1.1
+ +Table 2. Performance comparison with other state-of-the-art methods on ImageNet-Subset. IPC is set to 10. The best results are marked in bold, while the second-best results are underlined. “-” were not officially implemented in the original papers. + +
MethodsObject Detection
Pascal VOCMS COCO
mAPAP50mAPAP50
Ratio0.5%0.25%
Random0.8±0.23.1±0.40.5±0.11.7±0.3
Uniform0.9±0.13.4±0.30.8±0.22.4±0.5
K-Center0.5±0.12.1±0.30.4±0.11.5±0.2
Herding0.6±0.22.4±0.20.5±0.11.8±0.4
UniDD (Ours)8.5±0.422.3±0.64.5±0.310.3±0.4
Ratio1%0.5%
Random4.2±0.513.7±0.63.7±0.210.1±0.3
Uniform5.7±0.217.7±0.43.4±0.49.5±0.6
K-Center3.6±0.612.3±0.33.2±0.59.3±0.5
Herding3.5±0.511.9±0.53.5±0.39.7±0.3
UniDD (Ours)16.8±0.538.9±0.77.1±0.416.9±0.3
Ratio2%1%
Random12.4±0.434.3±0.57.2±0.817.3±0.9
Uniform13.8±0.336.2±0.47.4±0.517.6±0.5
K-Center10.9±0.629.3±0.66.1±0.315.4±0.6
Herding10.4±0.428.7±0.76.7±0.416.3±0.7
UniDD (Ours)23.9±0.548.5±0.610.8±0.422.5±0.5
Full51.4±0.880.3±0.432.6±0.751.4±0.8
+ +Table 3. Performance comparison with coreset selection methods on Pascal VOC and MS COCO for the object detection task, using Faster R-CNN as the detector. "Ratio" denotes the compression ratio. + +
MethodsSemantic Segmentation (mIoU)
Ratio: 3.5%Ratio: 7%Ratio: 14%Full
Random6.7±0.49.6±0.511.3±0.4
Uniform6.5±0.69.3±0.412.1±0.5
K-Center6.1±0.48.5±0.610.4±0.732.9±0.6
Herding6.2±0.78.7±0.410.6±0.3
UniDD (Ours)10.3±0.512.3±0.314.9±0.7
+ +versatility across a range of vision tasks. + +Qualitative Result. As shown in Fig. 2 (b) and 2 (c), we present visualizations of detection and segmentation data synthesized using the UniDD framework. These images + +are generated with $\mathcal{T}_{\mathcal{SP}}$ models, Faster R-CNN for detection and LR-ASPP for segmentation, which effectively steer the diffusion process toward the specified labels, bounding boxes, and masks. Previous research in dataset distillation has predominantly focused on single-category label synthesis, where each image aligns with a single class. In contrast, our UniDD framework overcomes this limitation, enabling dataset distillation for more complex, multi-class tasks. This advancement not only broadens the applicability of dataset distillation but also offers valuable insights for future research in this domain. + +# 4.4. Ablation Study + +Component Analysis. We conduct a comprehensive ablation study on the proposed UniDD, analyzing its components: (1) Baseline (a standard SD v1.5 checkpoint), (2) Task-driven Diffusion with the $\mathcal{T}_{\mathcal{SP}}$ model, (3) Task-Aligned Contextual Prompting (abbreviated as Context), and (4) Proxy-Driven High-Realism Refinement (abbreviated as Refine). As shown in Table 5, the proposed task-driven approach significantly improves the performance of the baseline, demonstrating that this guidance effectively bridges the gap between real and synthetic datasets. The contextual information is more effective for more complex detection and segmentation datasets, as it reveals underlying target relationships. The refinement module also proves to be effective, as the randomness of the SD model can lead to the generation of images that deviate from the annotations or result in failed generations. + +Table 4. Performance comparison with coreset selection methods on Pascal VOC for the semantic segmentation task, using LR-ASPP as the segmenter. The metric is mIoU, and "Ratio" denotes the compression ratio. + +
MethodsDataset
Cls. / Accuracy (%)Det. / mAP (%)Seg. / mIoU (%)
ImageNetteImageWoofVOCVOC
Baseline60.4±0.939.1±1.111.5±0.27.2±0.3
Baseline+TSP64.2±0.845.3±0.614.7±0.48.9±0.5
Baseline+TSRP+Context64.7±0.446.1±0.715.9±0.39.6±0.4
Baseline+TSP+Context+Refine66.0±0.746.8±0.916.8±0.510.3±0.5
+ +Table 5. Ablation study on UniDD. For classification datasets, IPC is set to 10 with ResNet-18 as the test model. For detection on VOC, the compression ratio is set to $1\%$ with Faster R-CNN as the test model. For segmentation on VOC, the ratio is set to $3.5\%$ with LR-ASPP as the test model. + +Cross-Architecture Analysis. Cross-Architecture experi + +
DatasetTest Model\( T_{SP} \)Model
ResNet-18ResNet-50ResNet-101
ImageNetteResNet-1866.0±0.768.8±0.769.5±0.7
ResNet-5067.2±0.669.5±1.170.7±0.9
ResNet-10169.3±0.970.9±1.272.1±0.6
ImageWoofResNet-1846.8±0.949.2±0.550.1±0.8
ResNet-5047.3±0.750.0±0.850.9±0.6
ResNet-10148.9±1.151.5±0.652.4±0.7
+ +Table 6. $\mathcal{T}_{\mathcal{SP}}$ Model Ablation and Cross-Architecture Testing: We use ResNet-18, ResNet-50, and ResNet-101 as $\mathcal{T}_{\mathcal{SP}}$ models to generate synthetic images. During testing, the same three models are used. IPC is set to 10. + +![](images/710da0b5dfd0a15b8f27cf0cd423dfc27eb92886c2f2843ff03c914356ef0257.jpg) +(a) Accuracy vs. Guidance strength + +![](images/28c224eb7f4d1179d114e2528c56f2684e8881d26b53783ef767363b2beb0521.jpg) +(b) Accuracy vs. Sample Step + +![](images/56268c4bf81de374ccba9e17d01f26bc4e2c049a1a2d33703cbd143c4e5f24bf.jpg) +Figure 4. Hyper-parameter analysis on (a) guidance strength $\lambda_{s}$ and (b) sampling step $t$ . The star symbol indicates the parameters chosen for trade-off. +(a) Training Time overhead +Figure 5. Comparison of distillation cost on ImageNet-1k using our UniDD with comparison methods (MTT [3], $\mathrm{SRe}^2\mathrm{L}$ [46], and $\mathrm{D^4M}$ [41]), tested on V100 GPUs. + +![](images/3980bfe19f23ab16a4f8b59e65acfff76214a08a92e6d61e9526d8bf476c7cb1.jpg) +(b) Training GPU overhead + +ments are crucial for validating the generalizability of synthetic data. We compare different combinations of $\mathcal{T}_{\mathcal{SP}}$ models and testing models. In this study, we set up three $\mathcal{T}_{\mathcal{SP}}$ models: ResNet-18, ResNet-50, and ResNet-101. These models are also used during testing to evaluate the cross-architecture performance of the synthetic datasets. As shown in Table. 6, using more complex $\mathcal{T}_{\mathcal{SP}}$ models improves the guidance effectiveness. Moreover, when testing across different architectures, the performance does not degrade as seen in previous optimization-based meth + +ods [3, 49, 50], demonstrating the generalizability of our synthetic data. + +Guidance Strength Analysis. As shown in Fig. 4 (a), we analyze the performance variations of ImageWoof and ImageNette at IPC-10 under different guidance strength settings $\lambda_{s}$ , with Conv5 as the test model. The best results are achieved at $\lambda_{s} = 100$ , but further increasing the guidance strength interferes with the quality of diffusion-generated images, leading to a decline in test performance. + +Sample Step Analysis. As shown in Fig. 4 (b), we analyze the impact of different SD sample steps on the performance of synthetic datasets under the IPC-10 setting, with Image-Woof and ImageNette resolutions set to $256 \times 256$ . Performance improves significantly with sample steps below 250, but gains are minimal beyond this point while time consumption increases. Thus, we select 250 steps as the optimal trade-off. + +Distillation Cost Analysis. As shown in Fig. 5, we evaluate the time and GPU cost of various DD methods. Previous methods primarily focus on the time required to synthesize images, often neglecting the time needed to train or deploy the DD model. For MTT [3], the deployment time corresponds to the expert trajectory training time, while for $\mathrm{D^4M}$ [41], it is the time required to finetune the SD model. The slight increase in UniDD's distillation time and GPU cost is mainly due to the added supervision process of the $\mathcal{T}_{\mathcal{SP}}$ model and the refinement stage, where approximately $5\% -10\%$ of low-quality generated data is discarded. Combining the deployment and distillation costs for comparison, UniDD demonstrates the best advantage, indicating its superior suitability for practical deployment. + +# 5. Conclusion + +In this study, we introduce the first Universal Dataset Distillation (UniDD) framework, capable of handling classification, detection, and segmentation tasks within a unified approach. UniDD follows a two-stage process: in the first stage, Task-Specific Proxy Training and Task-Aligned Contextual Prompting extract critical information from real datasets. In the second stage, Universal Taskdriven Diffusion synthesizes images with specific categories, bounding boxes, and segmentation masks tailored to each task. We conduct extensive experiments to validate UniDD's effectiveness across all three tasks. For classification, UniDD achieves state-of-the-art results on ImageNet1k and ImageNet-Subset. In detection and segmentation, we establish the first dataset distillation benchmark on Pascal VOC and MS COCO, comparing UniDD to coreset selection methods. Encouraging results demonstrate that our UniDD consistently outperforms these methods, especially under extremely low compression ratios. + +Acknowledgments. This work was supported by National Natural Science Fund of China (62473286). + +# References + +[1] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 843-852, 2023. 3, 4 +[2] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European conference on computer vision (ECCV), pages 233-248, 2018. 6 +[3] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A Efros, and Jun-Yan Zhu. Dataset distillation by matching training trajectories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4750-4759, 2022. 1, 2, 5, 6, 8 +[4] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 6 +[5] Yutian Chen, Max Welling, and Alex Smola. Super-samples from kernel herding. arXiv preprint arXiv:1203.3472, 2012. 6 +[6] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.6 +[7] Justin Cui, Ruochen Wang, Si Si, and Cho-Jui Hsieh. Scaling up dataset distillation toImagenet-1k with constant memory. In International Conference on Machine Learning, pages 6565–6590. PMLR, 2023. 1, 5 +[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1, 2, 3, 5 +[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3, 4 +[10] Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. Advances in Neural Information Processing Systems, 36:16222-16239, 2023. 3 +[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 1, 2, 3, 5 +[12] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. International journal of computer vision, 111:98-136, 2015. 1, 2, 3, 5 +[13] Junyao Gao, Xinyang Jiang, Huishuai Zhang, Yifan Yang, Shuguang Dou, Dongsheng Li, Duoqian Miao, Cheng Deng, + +and Cairong Zhao. Similarity distribution based membership inference attack on person re-identification. In Proceedings of the AAAI conference on artificial intelligence, pages 14820-14828, 2023. 1 +[14] Junyao Gao, Xinyang Jiang, Shuguang Dou, Dongsheng Li, Duoqian Miao, and Cairong Zhao. Re-id-leak: Membership inference attacks against person re-identification. International Journal of Computer Vision, 132(10):4673-4687, 2024. 1 +[15] Junyao Gao, Yanchen Liu, Yanan Sun, Yinhao Tang, Yanhong Zeng, Kai Chen, and Cairong Zhao. Styleshot: A snapshot on any style. arXiv preprint arXiv:2407.01414, 2024. 2 +[16] Junyao Gao, Yanan Sun, Fei Shen, Xin Jiang, Zhening Xing, Kai Chen, and Cairong Zhao. Faceshot: Bring any character into life. arXiv preprint arXiv:2503.00740, 2025. 2 +[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2 +[18] Jianyang Gu, Saeed Vahidian, Vyacheslav Kungurtsev, Haonan Wang, Wei Jiang, Yang You, and Yiran Chen. Efficient dataset distillation via minimax diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15793-15803, 2024. 5, 6 +[19] Ziyao Guo, Kai Wang, George Cazenavette, Hui Li, Kaipeng Zhang, and Yang You. Towards lossless dataset distillation via difficulty-aligned trajectory matching. arXiv preprint arXiv:2310.05773, 2023. 2 +[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1 +[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4, 5 +[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1 +[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3 +[24] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1314-1324, 2019. 5 +[25] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018. 2 +[26] Jang-Hyun Kim, Jinuk Kim, Seong Joon Oh, Sangwoo Yun, Hwanjun Song, Joonhyun Jeong, Jung-Woo Ha, and Hyun Oh Song. Dataset condensation via efficient synthetic-data parameterization. In International Conference on Machine Learning, pages 11102-11118. PMLR, 2022. 1, 2 + +[27] Jian Li, Yabiao Wang, Changan Wang, Ying Tai, Jianjun Qian, Jian Yang, Chengjie Wang, Jilin Li, and Feiyue Huang. Dsfd: dual shot face detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5060-5069, 2019. 2 +[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1, 2, 5 +[29] Songhua Liu, Kai Wang, Xingyi Yang, Jingwen Ye, and Xin chao Wang. Dataset distillation via factorization. Advances in Neural Information Processing Systems, 35:1100-1113, 2022. 6 +[30] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 289-299, 2023. 3 +[31] Yanqing Liu, Jianyang Gu, Kai Wang, Zheng Zhu, Wei Jiang, and Yang You. Dream: Efficient dataset distillation by representative matching. arXiv preprint arXiv:2302.14416, 2023. 2 +[32] Timothy Nguyen, Zhourong Chen, and Jaehoon Lee. Dataset meta-learning from kernel ridge-regression. In International Conference on Learning Representations, 2020. 2 +[33] OpenAI. Gpt-4 technical report. https://openai.com/research/gpt-4, 2023. Accessed: 2024-11-13. 2, 4 +[34] Ding Qi, Jian Li, Jinlong Peng, Bo Zhao, Shuguang Dou, Jialin Li, Jiangning Zhang, Yabiao Wang, Chengjie Wang, and Cairong Zhao. Fetch and forge: Efficient dataset condensation for object detection. Advances in Neural Information Processing Systems, 37:119283-119300, 2024. 3 +[35] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 6 +[36] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 4, 5, 6 +[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 5 +[38] Ahmad Sajedi, Samir Khaki, Ehsan Amjadian, Lucy Z Liu, Yuri A Lawryshyn, and Konstantinos N Plataniotis. Datadam: Efficient dataset distillation with attention matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17097-17107, 2023. 6 +[39] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. arXiv preprint arXiv:1708.00489, 2017. 6 + +[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3 +[41] Duo Su, Junjie Hou, Weizhi Gao, Yingjie Tian, and Bowen Tang. D^4: Dataset distillation via disentangled diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5809-5818, 2024. 5, 6, 8 +[42] Peng Sun, Bei Shi, Daiwei Yu, and Tao Lin. On the diversity and realism of distilled dataset: An efficient dataset distillation paradigm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9390-9399, 2024. 1, 3, 5 +[43] Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Shuo Yang, Shuo Wang, Guan Huang, Hakan Bilen, Xinchao Wang, and Yang You. Cafe: Learning to condense dataset by aligning features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12196-12205, 2022. 2 +[44] Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A Efros. Dataset distillation. arXiv preprint arXiv:1811.10959, 2018. 1, 2, 6 +[45] Yichao Yan, Zanwei Zhou, Zi Wang, Jingnan Gao, and Xiaokang Yang. Dialoguenerf: Towards realistic avatar face-to-face conversation video generation. Visual Intelligence, 2 (1):24, 2024. 1 +[46] Zeyuan Yin, Eric Xing, and Zhiqiang Shen. Squeeze, recover and relabel: Dataset condensation at imagenet scale from a new perspective. arXiv preprint arXiv:2306.13092, 2023. 1, 3, 5, 6, 8 +[47] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2 +[48] Bo Zhao and Hakan Bilen. Dataset condensation with differentiable siamese augmentation. In International Conference on Machine Learning, pages 12674-12685. PMLR, 2021. 1 +[49] Bo Zhao and Hakan Bilen. Dataset condensation with distribution matching. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6514-6523, 2023. 1, 2, 6, 8 +[50] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. Dataset condensation with gradient matching. In International Conference on Learning Representations, 2020. 1, 2, 6, 8 +[51] Ganlong Zhao, Guanbin Li, Yipeng Qin, and Yizhou Yu. Improved distribution matching for dataset condensation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7856-7865, 2023. 1, 2 \ No newline at end of file diff --git a/2025/Towards Universal Dataset Distillation via Task-Driven Diffusion/images.zip b/2025/Towards Universal Dataset Distillation via Task-Driven Diffusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cdb18e5e201effc5f84040c50928f05c07e7b43f --- /dev/null +++ b/2025/Towards Universal Dataset Distillation via Task-Driven Diffusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4b0686c7d02b839dd5d304a75e8e60ecdb9296f74202278a867ad64aaa6926e1 +size 647860 diff --git a/2025/Towards Universal Dataset Distillation via Task-Driven Diffusion/layout.json b/2025/Towards Universal Dataset Distillation via Task-Driven Diffusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3814a098d0234b018705fdd41d8b3cc4b1a8f20a --- /dev/null +++ b/2025/Towards Universal Dataset Distillation via Task-Driven Diffusion/layout.json @@ -0,0 +1,10417 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 104, + 103, + 506, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 103, + 506, + 120 + ], + "spans": [ + { + "bbox": [ + 104, + 103, + 506, + 120 + ], + "type": "text", + "content": "Towards Universal Dataset Distillation via Task-Driven Diffusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "spans": [ + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "text", + "content": "Ding Qi" + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "text", + "content": " Jian Li" + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "text", + "content": " Junyao Gao" + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "text", + "content": " Shuguang Dou" + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "text", + "content": " Ying Tai" + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "text", + "content": " \nJianlong Hu" + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "text", + "content": " Bo Zhao" + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "text", + "content": " Yabiao Wang" + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "inline_equation", + "content": "^{3,2\\dagger}" + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "text", + "content": " Chengjie Wang" + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "inline_equation", + "content": "^{4,2}" + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "text", + "content": " Cairong Zhao" + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "text", + "content": "Tongji University " + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "text", + "content": "Tencent YouTu Lab " + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "text", + "content": "Zhejiang University \n" + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "text", + "content": "Shanghai Jiao Tong University " + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 99, + 141, + 510, + 201 + ], + "type": "text", + "content": "Nanjing University" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 76, + 213, + 366, + 361 + ], + "blocks": [ + { + "bbox": [ + 76, + 213, + 366, + 361 + ], + "lines": [ + { + "bbox": [ + 76, + 213, + 366, + 361 + ], + "spans": [ + { + "bbox": [ + 76, + 213, + 366, + 361 + ], + "type": "image", + "image_path": "273080066be2f11e90e6d1d4791effb0af3c1dd2d9a9a35dc7bff1253d4caa6d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 92, + 366, + 337, + 376 + ], + "lines": [ + { + "bbox": [ + 92, + 366, + 337, + 376 + ], + "spans": [ + { + "bbox": [ + 92, + 366, + 337, + 376 + ], + "type": "text", + "content": "(a) One Universal Dataset Distillation Framework for Mutli-Tasks" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 385, + 221, + 531, + 347 + ], + "blocks": [ + { + "bbox": [ + 385, + 221, + 531, + 347 + ], + "lines": [ + { + "bbox": [ + 385, + 221, + 531, + 347 + ], + "spans": [ + { + "bbox": [ + 385, + 221, + 531, + 347 + ], + "type": "image", + "image_path": "29377090049304128148076e6e350a1f9d66a781866c0624fbfc3fec438bfb75.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 365, + 534, + 377 + ], + "lines": [ + { + "bbox": [ + 394, + 365, + 534, + 377 + ], + "spans": [ + { + "bbox": [ + 394, + 365, + 534, + 377 + ], + "type": "text", + "content": "(b) Improved Performance by UniDD" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 387, + 555, + 431 + ], + "lines": [ + { + "bbox": [ + 54, + 387, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 54, + 387, + 555, + 431 + ], + "type": "text", + "content": "Figure 1. (a) We introduce the Universal Dataset Distillation Framework (UniDD), a novel dataset distillation framework supporting a range of vision tasks, including classification, object detection, and segmentation. UniDD incorporates diverse label types, such as class labels, bounding box coordinates, and pixel-level masks, facilitating the synthesis of task-specific datasets. (b) UniDD improves the performance of dataset distillation across diverse tasks on ImageNet [8], Pascal VOC [11, 12], and MS COCO [28]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 446, + 200, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 446, + 200, + 458 + ], + "spans": [ + { + "bbox": [ + 151, + 446, + 200, + 458 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 472, + 297, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 472, + 297, + 676 + ], + "spans": [ + { + "bbox": [ + 54, + 472, + 297, + 676 + ], + "type": "text", + "content": "Dataset distillation (DD) condenses key information from large-scale datasets into smaller synthetic datasets, reducing storage and computational costs for training networks. However, most recent research has primarily focused on image classification tasks, with limited exploration in detection and segmentation. Two key challenges remain: (i) Task Optimization Heterogeneity, where existing methods focus on class-level information but fail to address the diverse needs of detection and segmentation, and (ii) Inflexible Image Generation, where current generation methods rely on global updates for single-class targets and lack localized optimization for specific object regions. To address these challenges, we propose UniDD, a universal dataset distillation framework built on a task-driven diffusion model for diverse DD tasks, as shown in Fig. 1. Our approach operates in two stages: Universal Task Knowledge Mining, which captures task-relevant information through task" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 447, + 556, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 556, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 556, + 544 + ], + "type": "text", + "content": "specific proxy model training, and Universal Task-Driven Diffusion, where these proxies guide the diffusion process to generate task-specific synthetic images. Extensive experiments across ImageNet-1K, Pascal VOC, and MS COCO demonstrate that UniDD consistently outperforms state-of-the-art methods. In particular, on ImageNet-1K with IPC-10, UniDD surpasses previous diffusion-based methods by " + }, + { + "bbox": [ + 313, + 447, + 556, + 544 + ], + "type": "inline_equation", + "content": "6.1\\%" + }, + { + "bbox": [ + 313, + 447, + 556, + 544 + ], + "type": "text", + "content": ", while also reducing deployment costs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 571, + 395, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 571, + 395, + 583 + ], + "spans": [ + { + "bbox": [ + 314, + 571, + 395, + 583 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 594, + 556, + 715 + ], + "type": "text", + "content": "Dataset distillation (DD) [44], as an innovative data optimization technique in machine learning [13, 14, 20, 22, 45], condenses essential information from large datasets into smaller proxy sets. Training on these highly ensemble proxy sets enables the model to use much fewer resources while providing comparable performance. Prior DD methods excelled in image classification using various techniques like Gradient Matching [26, 48, 50], Distribution Matching [49, 51], Training Trajectory Matching [3, 7], and Decoupled Methods [42, 46]." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 683, + 295, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 702 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 702 + ], + "type": "text", + "content": "*Work done when Ding Qi is an intern in Tencent YouTu Lab, and the advisor is Jian Li." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 66, + 703, + 151, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 703, + 151, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 703, + 151, + 712 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10557" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 274 + ], + "type": "text", + "content": "Despite advancements in DD, research on key visual tasks—detection and segmentation—remains limited. We identify the following key challenges: (i) Task Optimization Heterogeneity. Existing DD methods are primarily focused on image classification, optimizing distillation by compressing class-level information, but they fall short in addressing the diverse needs of tasks like detection [27] and segmentation, such as object localization and pixel-level information. (ii) Inflexible Image Generation. Current image generation methods rely on global updates for single-class targets and lack localized optimization for specific object regions. These limitations become especially evident in images with multiple objects or complex structures, hindering the ability of existing DD to generate distilled data that meets the requirements of diverse tasks. Consequently, the applicability of DD in the visual domain remains constrained." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 278, + 296, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 278, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 278, + 296, + 613 + ], + "type": "text", + "content": "To address these limitations, we propose a Universal Dataset Distillation (UniDD) framework that optimizes for multi-task learning, including classification, detection, and segmentation, overcoming the shortcomings in task adaptation and image generation. Our approach consists of two stages: Universal Task Knowledge Mining and Universal Task-Driven Diffusion. In the first stage, we introduce Task-Specific Proxy " + }, + { + "bbox": [ + 56, + 278, + 296, + 613 + ], + "type": "inline_equation", + "content": "(\\mathcal{T}_{\\mathcal{SP}})" + }, + { + "bbox": [ + 56, + 278, + 296, + 613 + ], + "type": "text", + "content": " models, trained on real datasets to extract task-specific information. This allows the " + }, + { + "bbox": [ + 56, + 278, + 296, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathcal{SP}}" + }, + { + "bbox": [ + 56, + 278, + 296, + 613 + ], + "type": "text", + "content": " models to capture and store category information for classification, bounding box positions for detection, and mask data for segmentation within their parameters. Additionally, previous research has often overlooked the importance of contextual information, particularly in high-resolution datasets like ImageNet [8], Pascal VOC [11, 12] and MS COCO [28], where rich object-context relationships exist. To address this, we introduce Task-Aligned Contextual Prompting, which uses a Vision-Language model [33] to generate contextual prompts for real images. In the second stage, we introduce a Task-Driven Diffusion for image synthesis. Specifically, the " + }, + { + "bbox": [ + 56, + 278, + 296, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathcal{SP}}" + }, + { + "bbox": [ + 56, + 278, + 296, + 613 + ], + "type": "text", + "content": " models trained in the first stage supervise the denoising process of the diffusion model by calculating the task loss on the predicted clean images. This guides the diffusion model in generating images that match the given labels. Notably, since the " + }, + { + "bbox": [ + 56, + 278, + 296, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathcal{SP}}" + }, + { + "bbox": [ + 56, + 278, + 296, + 613 + ], + "type": "text", + "content": " models already align with the target domain, there is no need to retrain the diffusion model on the real set, significantly reducing the cost in practical applications." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 616, + 247, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 616, + 247, + 627 + ], + "spans": [ + { + "bbox": [ + 67, + 616, + 247, + 627 + ], + "type": "text", + "content": "In summary, our contributions are threefold:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 629, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 629, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 629, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 629, + 295, + 677 + ], + "type": "text", + "content": "- We identify key challenges in extending dataset distillation beyond classification and propose the first universal framework for DD across classification, detection, and segmentation tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 677, + 296, + 714 + ], + "type": "text", + "content": "- We develop a two-stage approach, Universal Task Knowledge Mining and Universal Task-Driven Diffusion for Data Synthesis, which flexibly synthesizes images for" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 553, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 107 + ], + "type": "text", + "content": "multi-tasks. Additionally, we address the previously overlooked need to align potential contextual information within datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 108, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 108, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 314, + 108, + 553, + 180 + ], + "type": "text", + "content": "- Our method achieves state-of-the-art performance in classification tasks, on the ImageNet-1K dataset, our approach surpasses the previous state-of-the-art method by over " + }, + { + "bbox": [ + 314, + 108, + 553, + 180 + ], + "type": "inline_equation", + "content": "6.1\\%" + }, + { + "bbox": [ + 314, + 108, + 553, + 180 + ], + "type": "text", + "content": ". Additionally, we establish new benchmarks for detection and segmentation on VOC and COCO, showing significant improvements over coreset methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 192, + 404, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 192, + 404, + 205 + ], + "spans": [ + { + "bbox": [ + 314, + 192, + 404, + 205 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 213, + 427, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 213, + 427, + 224 + ], + "spans": [ + { + "bbox": [ + 314, + 213, + 427, + 224 + ], + "type": "text", + "content": "2.1. Dataset Distillation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 231, + 555, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 555, + 566 + ], + "type": "text", + "content": "Dataset distillation [44], a dataset reduction method, addresses the challenge of handling substantial data by synthesizing a smaller, representative dataset. Existing dataset distillation methods can be categorized into meta-learning and data matching frameworks based on whether they explicitly mimic the performance of target data. In the meta-learning framework, the distilled data are treated as hyperparameters and optimized in a nested loop fashion according to the trained model's risk. For example, KIP [32] performs kernel ridge regression using the Neural Tangent Kernel (NTK) [25]. The data matching framework includes gradient matching, distribution matching, and trajectory matching. These methods update distilled data by imitating the influence of target data. For example, DC [50] and DM [49] enforce synthetic and real images to have similar gradients or distribution at every training step. MTT [3] proposes to imitate long-range trajectories of optimization steps for real images, thereby better mimicking the training dynamics of real images for synthetic images. However, the aforementioned gradient matching methods require a large number of iterations. To address this, DREAM [31] and IDC [26] focus on more efficient dataset distillation training. Due to the unsatisfactory performance of DM, CAFE [43] and IDM [51] propose improved distribution matching between the real and synthetic data. For the trajectory matching method, DATM [19] scales trajectory matching-based methods to larger synthetic datasets, achieving lossless dataset distillation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 567, + 554, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 554, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 554, + 649 + ], + "type": "text", + "content": "Despite significant progress in previous work, there have been few breakthroughs in dataset distillation for detection and segmentation tasks. In this work, we introduce the first universal framework that simultaneously handles classification, object detection, and semantic segmentation, significantly expanding the generality and scope of dataset distillation research." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 659, + 417, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 659, + 417, + 671 + ], + "spans": [ + { + "bbox": [ + 314, + 659, + 417, + 671 + ], + "type": "text", + "content": "2.2. Diffusion Models" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Diffusion models [15, 16, 37, 47] are generative models that show superior fidelity and diversity compared to GANs [17]. It aims to learn the mapping from a Gaussian" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10558" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 217, + 228 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 217, + 228 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 217, + 228 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 217, + 228 + ], + "type": "image", + "image_path": "8c94ce46423189613163e117cc42deebceb0535c1c09b9fdc51f7a33944a2e6a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 230, + 199, + 239 + ], + "lines": [ + { + "bbox": [ + 75, + 230, + 199, + 239 + ], + "spans": [ + { + "bbox": [ + 75, + 230, + 199, + 239 + ], + "type": "text", + "content": "(a) Synthetic images for classification." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 225, + 72, + 385, + 228 + ], + "blocks": [ + { + "bbox": [ + 225, + 72, + 385, + 228 + ], + "lines": [ + { + "bbox": [ + 225, + 72, + 385, + 228 + ], + "spans": [ + { + "bbox": [ + 225, + 72, + 385, + 228 + ], + "type": "image", + "image_path": "6a727fa95fd69b053327cd236578d7e039dc29ffce88b311fa0bedf52568970d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 230, + 361, + 240 + ], + "lines": [ + { + "bbox": [ + 249, + 230, + 361, + 240 + ], + "spans": [ + { + "bbox": [ + 249, + 230, + 361, + 240 + ], + "type": "text", + "content": "(b) Synthetic images for detection." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 393, + 72, + 553, + 228 + ], + "blocks": [ + { + "bbox": [ + 393, + 72, + 553, + 228 + ], + "lines": [ + { + "bbox": [ + 393, + 72, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 393, + 72, + 553, + 228 + ], + "type": "image", + "image_path": "b749891040b76d77d0cb4ecba985295e6e5ca13b40b01681cb87fd578d07b433.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 230, + 535, + 240 + ], + "lines": [ + { + "bbox": [ + 410, + 230, + 535, + 240 + ], + "spans": [ + { + "bbox": [ + 410, + 230, + 535, + 240 + ], + "type": "text", + "content": "(c) Synthetic images for segmentation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 249, + 555, + 294 + ], + "lines": [ + { + "bbox": [ + 54, + 249, + 555, + 294 + ], + "spans": [ + { + "bbox": [ + 54, + 249, + 555, + 294 + ], + "type": "text", + "content": "Figure 2. Visualization of synthetic images generated by UniDD. (a) For classification, images are synthesized on the ImageWoof [8] dataset. (b) For detection, images are synthesized on the Pascal VOC dataset with specified bounding box coordinates. (c) For segmentation, images are synthesized on the Pascal VOC [11, 12] dataset with pixel-level masks. Our UniDD is the first approach to enable high-realism data distillation across classification, detection, and segmentation tasks." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 300, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 361 + ], + "type": "text", + "content": "distribution to the real dataset distribution by adding noise in the diffusion process and reversing the noise to the raw image in the denoising process. In the forward process, diffusion models progressively add noise to a clean sample " + }, + { + "bbox": [ + 55, + 300, + 296, + 361 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 55, + 300, + 296, + 361 + ], + "type": "text", + "content": ", resulting in a noisy sample " + }, + { + "bbox": [ + 55, + 300, + 296, + 361 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 55, + 300, + 296, + 361 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 90, + 365, + 295, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 365, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 90, + 365, + 295, + 380 + ], + "type": "interline_equation", + "content": "z _ {t} = \\sqrt {\\alpha_ {t}} z _ {0} + \\sqrt {1 - \\alpha_ {t}} \\epsilon , \\quad \\epsilon \\sim \\mathcal {N} (0, I). \\tag {1}", + "image_path": "dea91aaa4df85a6a69a16ea47866274895e57f698a2957c96f07babaa5c32fa6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 384, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 407 + ], + "type": "text", + "content": "In the reverse process, a denoising network " + }, + { + "bbox": [ + 55, + 384, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 384, + 296, + 407 + ], + "type": "text", + "content": " estimates the noise and iteratively recovers the clean sample " + }, + { + "bbox": [ + 55, + 384, + 296, + 407 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 55, + 384, + 296, + 407 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 123, + 411, + 295, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 411, + 295, + 439 + ], + "spans": [ + { + "bbox": [ + 123, + 411, + 295, + 439 + ], + "type": "interline_equation", + "content": "\\epsilon_ {\\theta} (z _ {t}, t) \\approx \\frac {z _ {t} - \\sqrt {\\alpha_ {t}} z _ {0}}{\\sqrt {1 - \\alpha_ {t}}}. \\tag {2}", + "image_path": "7a979b183c6b5e85a9e77258c33715558e6dae6f2f655a9a69ef9d2a2722eb41.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 441, + 296, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 296, + 465 + ], + "type": "text", + "content": "DDIM [40] provides the following formula for predicting a clean data sample:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 468, + 295, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 468, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 115, + 468, + 295, + 497 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {0} = \\frac {z _ {t} - \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {\\theta} (z _ {t} , t)}{\\sqrt {\\alpha_ {t}}}. \\tag {3}", + "image_path": "2bf9bcbffe3ae014e4a75f59d2bf97f30190dd5f5743a57115c245b5d0009fe8.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 500, + 296, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 500, + 296, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 296, + 535 + ], + "type": "text", + "content": "This formula predicts the clean data point " + }, + { + "bbox": [ + 55, + 500, + 296, + 535 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 55, + 500, + 296, + 535 + ], + "type": "text", + "content": " from the noisy sample " + }, + { + "bbox": [ + 55, + 500, + 296, + 535 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 55, + 500, + 296, + 535 + ], + "type": "text", + "content": ", enabling deterministic sampling for efficient and high-quality image synthesis." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 536, + 296, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 296, + 643 + ], + "type": "text", + "content": "To introduce additional conditions into the diffusion model, classifier guidance [9] has been proposed to train an additional classifier on the noise to guide the model to generate images that are closer to the specified condition. Additionally, [23] introduces classifier-free guidance that trains both an unconditioned and conditioned model simultaneously, achieving more efficient guidance. Following these, recent methods [1, 10, 30] have been proposed to generate high-quality images from diverse conditions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 652, + 285, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 285, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 285, + 665 + ], + "type": "text", + "content": "3. Universal Dataset Distillation Framework" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 672, + 167, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 167, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 167, + 684 + ], + "type": "text", + "content": "3.1. Problem Definition" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Universal Dataset Distillation. Current dataset distillation research primarily targets classification tasks and lacks" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 300, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 445 + ], + "type": "text", + "content": "a generalizable approach for detection and segmentation. To address this, we propose a universal dataset distillation standard that constructs a compact synthetic dataset " + }, + { + "bbox": [ + 313, + 300, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{S} = (\\hat{x}_1, \\hat{y}_1), \\dots, (\\hat{x}_{|\\mathcal{S}|}, \\hat{y}_{|\\mathcal{S}|})" + }, + { + "bbox": [ + 313, + 300, + 555, + 445 + ], + "type": "text", + "content": " from a larger real dataset " + }, + { + "bbox": [ + 313, + 300, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = (x_1, y_1), \\dots, (x_{|\\mathcal{T}|}, y_{|\\mathcal{T}|})" + }, + { + "bbox": [ + 313, + 300, + 555, + 445 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 300, + 555, + 445 + ], + "type": "inline_equation", + "content": "|S| \\ll |\\mathcal{T}|" + }, + { + "bbox": [ + 313, + 300, + 555, + 445 + ], + "type": "text", + "content": ". The primary difference between tasks lies in label structure: classification uses single-class labels, detection includes bounding boxes with class information, and segmentation requires pixel-level masks. Our goal is for models trained on the distilled dataset " + }, + { + "bbox": [ + 313, + 300, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 313, + 300, + 555, + 445 + ], + "type": "text", + "content": " to perform comparably to those trained on the full dataset " + }, + { + "bbox": [ + 313, + 300, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 300, + 555, + 445 + ], + "type": "text", + "content": " across classification, detection, and segmentation tasks, as formalized by:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 329, + 452, + 555, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 452, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 329, + 452, + 555, + 468 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {(x, y) \\sim P _ {D}} \\left| \\ell \\left(\\theta_ {\\mathcal {T}} ^ {t a s k} (x), y\\right) - \\ell \\left(\\theta_ {\\mathcal {S}} ^ {t a s k} (x), y\\right) \\right| \\leq \\epsilon , \\tag {4}", + "image_path": "e3f2d99e2f5e88dfd93ed2feed3735cef621dd2ddba5566bb554466dcb43d6a4.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 474, + 555, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 534 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 474, + 555, + 534 + ], + "type": "inline_equation", + "content": "P_{D}" + }, + { + "bbox": [ + 313, + 474, + 555, + 534 + ], + "type": "text", + "content": " denotes the distribution of test data, " + }, + { + "bbox": [ + 313, + 474, + 555, + 534 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 313, + 474, + 555, + 534 + ], + "type": "text", + "content": " represents the performance metric specific to each task, " + }, + { + "bbox": [ + 313, + 474, + 555, + 534 + ], + "type": "inline_equation", + "content": "task\\in" + }, + { + "bbox": [ + 313, + 474, + 555, + 534 + ], + "type": "text", + "content": " {class,detect,segment}, and " + }, + { + "bbox": [ + 313, + 474, + 555, + 534 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 474, + 555, + 534 + ], + "type": "text", + "content": " is a small tolerance value. For each task, the parameter set " + }, + { + "bbox": [ + 313, + 474, + 555, + 534 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{T}}^{task}" + }, + { + "bbox": [ + 313, + 474, + 555, + 534 + ], + "type": "text", + "content": " on the real dataset " + }, + { + "bbox": [ + 313, + 474, + 555, + 534 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 474, + 555, + 534 + ], + "type": "text", + "content": " is optimized as follows:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 343, + 541, + 555, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 541, + 555, + 561 + ], + "spans": [ + { + "bbox": [ + 343, + 541, + 555, + 561 + ], + "type": "interline_equation", + "content": "\\theta_ {\\mathcal {T}} ^ {t a s k} = \\arg \\min _ {\\theta} \\mathbb {E} _ {(x, y) \\in \\mathcal {T}} \\left[ \\ell \\left(\\phi_ {\\theta} ^ {t a s k} (x), y\\right) \\right], \\qquad (5)", + "image_path": "ce4306fc5134891bc6810f9cdfb27ab6255d219e9f609faa3a215579815580d3.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 568, + 550, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 550, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 550, + 582 + ], + "type": "text", + "content": "Similarly, " + }, + { + "bbox": [ + 313, + 568, + 550, + 582 + ], + "type": "inline_equation", + "content": "\\theta_S^{task}" + }, + { + "bbox": [ + 313, + 568, + 550, + 582 + ], + "type": "text", + "content": " is optimized using the synthetic dataset " + }, + { + "bbox": [ + 313, + 568, + 550, + 582 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 568, + 550, + 582 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 588, + 499, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 499, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 499, + 601 + ], + "type": "text", + "content": "3.2. Universal Task Knowledge Mining" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "Task-Specific Proxy Training. Building a universal DD framework is challenging due to the diverse optimization objectives in classification, detection, and segmentation tasks. Recent decoupled DD studies have reached a consensus [34, 42, 46]: a fully trained network can capture and retain essential dataset information. Inspired by this, we extend to general tasks, training a classifier, detector, and segmenter on task-specific datasets to extract critical information—class labels for classification, bounding boxes for" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10559" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 555, + 297 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 555, + 297 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 555, + 297 + ], + "type": "image", + "image_path": "697bfc9651c9aab0d9e97c88dab77dad9859a6862ec3e7f7180dc844b1d8405e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 303, + 555, + 348 + ], + "lines": [ + { + "bbox": [ + 55, + 303, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 555, + 348 + ], + "type": "text", + "content": "Figure 3. Universal Dataset Distillation Framework. This framework consists of two stages: in the Universal Task Knowledge Mining stage, we introduce an easily deployable task-specific proxy " + }, + { + "bbox": [ + 55, + 303, + 555, + 348 + ], + "type": "inline_equation", + "content": "(\\mathcal{T}_{\\mathcal{SP}})" + }, + { + "bbox": [ + 55, + 303, + 555, + 348 + ], + "type": "text", + "content": " model trained on the target dataset and a vision-language model to extract task-aligned context. In the Universal Task-Driven Diffusion stage, we apply the gradient of the supervision loss, derived from the " + }, + { + "bbox": [ + 55, + 303, + 555, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathcal{SP}}" + }, + { + "bbox": [ + 55, + 303, + 555, + 348 + ], + "type": "text", + "content": " model's prediction of the clean image " + }, + { + "bbox": [ + 55, + 303, + 555, + 348 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 55, + 303, + 555, + 348 + ], + "type": "text", + "content": ", at each sampling step." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 354, + 295, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 295, + 378 + ], + "type": "text", + "content": "detection, and pixel-level masks for segmentation. We call this Task-Specific Proxy " + }, + { + "bbox": [ + 55, + 354, + 295, + 378 + ], + "type": "inline_equation", + "content": "(\\mathcal{T}_{\\mathcal{SP}})" + }, + { + "bbox": [ + 55, + 354, + 295, + 378 + ], + "type": "text", + "content": " Model, defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 388, + 295, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 388, + 295, + 404 + ], + "spans": [ + { + "bbox": [ + 115, + 388, + 295, + 404 + ], + "type": "interline_equation", + "content": "\\theta = \\arg \\min _ {\\theta} \\mathcal {L} _ {\\text {t a s k}} (\\mathcal {F} (x), y), \\tag {6}", + "image_path": "e371a7e7949be6a6d33a8c5502a6ca07073587d7dfd20a4eba754bc29c50ba2f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 414, + 295, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 295, + 449 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 414, + 295, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 414, + 295, + 449 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 55, + 414, + 295, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathcal{SP}}" + }, + { + "bbox": [ + 55, + 414, + 295, + 449 + ], + "type": "text", + "content": " model trained on the target data, " + }, + { + "bbox": [ + 55, + 414, + 295, + 449 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 414, + 295, + 449 + ], + "type": "text", + "content": " denotes the model's parameters, " + }, + { + "bbox": [ + 55, + 414, + 295, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{task}" + }, + { + "bbox": [ + 55, + 414, + 295, + 449 + ], + "type": "text", + "content": " represents the loss function specific to the task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 450, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 569 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 55, + 450, + 295, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathcal{SP}}" + }, + { + "bbox": [ + 55, + 450, + 295, + 569 + ], + "type": "text", + "content": " model choice is flexible, allowing adaptation for optimal performance on target datasets. For example, ResNet-18 [21] can serve as a lightweight, fast-deploy option for simple classification, while Faster R-CNN [36] may be used for complex detection tasks to capture key detection information. Unlike previous diffusion-based dataset distillation methods, our approach relies on task-driven learning to align synthetic and real datasets, avoiding costly diffusion model fine-tuning on the target dataset and significantly reducing deployment overhead." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": "Task-Aligned Contextual Prompting. Previous diffusion-based dataset distillation has largely focused on classification tasks and overlooked the role of natural language in guiding image generation. For diverse tasks, textual prompts must go beyond describing the primary objects in images. As inputs to Stable Diffusion models, prompts can also guide the generation process. To address this, we use a vision-language model to extract natural language descriptions and design task-specific questions to generate task-aligned prompts, a method we call Task-Aligned Contextual Prompting. Following GPT-4o [33], we create tailored prompts for different tasks. For classification, prompts de" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 354, + 555, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 402 + ], + "type": "text", + "content": "scribe object classes, while for detection and segmentation, they also include object relationships and the interaction between objects and the background, providing richer contextual guidance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 409, + 491, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 491, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 491, + 421 + ], + "type": "text", + "content": "3.3. Universal Task-Driven Diffusion." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "content": "Inspired by recent advancements in guidance diffusion techniques [1, 9], we propose a Task-Driven Diffusion approach for Dataset Distillation. As illustrated in Fig. 3, during each sampling step " + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "inline_equation", + "content": "S(z_{t}, t)" + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "content": " of the diffusion process, the trained " + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathcal{SP}}" + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "content": " model " + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "content": " provides task-specific guidance. Given an label " + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\tilde{a}" + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "content": ", the original noise prediction " + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_{t}, t, \\tilde{c})" + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "content": " is adjusted using the output of " + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "content": ". The adjusted noise prediction " + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}" + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "content": " is formulated as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 321, + 529, + 555, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 529, + 555, + 543 + ], + "spans": [ + { + "bbox": [ + 321, + 529, + 555, + 543 + ], + "type": "interline_equation", + "content": "\\hat {\\epsilon} _ {\\theta} \\left(z _ {t}, t, \\tilde {c}\\right) = \\epsilon_ {\\theta} \\left(z _ {t}, t, \\tilde {c}\\right) + \\sqrt {1 - \\alpha_ {t}} \\nabla_ {z _ {t}} \\mathcal {L} _ {\\text {t a s k}} \\left(\\tilde {a}, \\mathcal {F} \\left(z _ {t}\\right)\\right), \\tag {7}", + "image_path": "02247e67dce830579c1991ca643b396e1d79776587714b6636b2ffa60e39f3f5.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 550, + 554, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 554, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 554, + 621 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 550, + 554, + 621 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{task}}" + }, + { + "bbox": [ + 313, + 550, + 554, + 621 + ], + "type": "text", + "content": " represents the task-specific loss function and " + }, + { + "bbox": [ + 313, + 550, + 554, + 621 + ], + "type": "inline_equation", + "content": "\\tilde{a} = \\{a_{cl}, a_{bbox}, a_{mask}\\}" + }, + { + "bbox": [ + 313, + 550, + 554, + 621 + ], + "type": "text", + "content": " may denote a class label " + }, + { + "bbox": [ + 313, + 550, + 554, + 621 + ], + "type": "inline_equation", + "content": "a_{cl}" + }, + { + "bbox": [ + 313, + 550, + 554, + 621 + ], + "type": "text", + "content": ", a bounding box " + }, + { + "bbox": [ + 313, + 550, + 554, + 621 + ], + "type": "inline_equation", + "content": "a_{bbox}" + }, + { + "bbox": [ + 313, + 550, + 554, + 621 + ], + "type": "text", + "content": ", or a pixel-level segmentation mask " + }, + { + "bbox": [ + 313, + 550, + 554, + 621 + ], + "type": "inline_equation", + "content": "a_{mask}" + }, + { + "bbox": [ + 313, + 550, + 554, + 621 + ], + "type": "text", + "content": ". This adjustment mechanism effectively guides the diffusion process to produce data closely aligned with the intended distribution for each task. " + }, + { + "bbox": [ + 313, + 550, + 554, + 621 + ], + "type": "inline_equation", + "content": "\\tilde{c}" + }, + { + "bbox": [ + 313, + 550, + 554, + 621 + ], + "type": "text", + "content": " is a textual prompt." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 622, + 554, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 622, + 554, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 622, + 554, + 670 + ], + "type": "text", + "content": "Because " + }, + { + "bbox": [ + 313, + 622, + 554, + 670 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 622, + 554, + 670 + ], + "type": "text", + "content": " is trained on clean images, it may not perform well directly on noisy inputs. Therefore, we first use Equation 3 to estimate the clean image " + }, + { + "bbox": [ + 313, + 622, + 554, + 670 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 313, + 622, + 554, + 670 + ], + "type": "text", + "content": ". Once " + }, + { + "bbox": [ + 313, + 622, + 554, + 670 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 313, + 622, + 554, + 670 + ], + "type": "text", + "content": " is obtained, we apply " + }, + { + "bbox": [ + 313, + 622, + 554, + 670 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 622, + 554, + 670 + ], + "type": "text", + "content": " to compute the task-driven loss as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 679, + 555, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 679, + 555, + 693 + ], + "spans": [ + { + "bbox": [ + 318, + 679, + 555, + 693 + ], + "type": "interline_equation", + "content": "\\hat {\\epsilon} _ {\\theta} \\left(z _ {t}, t, \\tilde {c}\\right)) = \\epsilon_ {\\theta} \\left(z _ {t}, t, \\tilde {c}\\right) + s (t) \\cdot \\nabla_ {z _ {t}} \\mathcal {L} _ {\\text {t a s k}} (\\tilde {a}, \\mathcal {F} (\\hat {z} _ {0})), \\tag {8}", + "image_path": "14d85e9ea7085400acb08486bc7c7c5effd8b92add27d0c1fa35fe937b8151a3.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 701, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 701, + 554, + 714 + ], + "type": "inline_equation", + "content": "s(t)" + }, + { + "bbox": [ + 313, + 701, + 554, + 714 + ], + "type": "text", + "content": " is the task-driven strength at each sampling step," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10560" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 102, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 102, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 102, + 83 + ], + "type": "text", + "content": "defined as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 132, + 83, + 294, + 97 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 83, + 294, + 97 + ], + "spans": [ + { + "bbox": [ + 132, + 83, + 294, + 97 + ], + "type": "interline_equation", + "content": "s (t) = \\lambda_ {s} \\cdot \\sqrt {1 - \\alpha_ {t}}, \\tag {9}", + "image_path": "d499c62f1bad40be844a7596f831045a4b6a4a13a251603d2823d3eee682288a.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 102, + 296, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 102, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 55, + 102, + 296, + 162 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 55, + 102, + 296, + 162 + ], + "type": "inline_equation", + "content": "\\lambda_{s}" + }, + { + "bbox": [ + 55, + 102, + 296, + 162 + ], + "type": "text", + "content": " being a constant coefficient controlling the overall intensity of the task-driven. The term " + }, + { + "bbox": [ + 55, + 102, + 296, + 162 + ], + "type": "inline_equation", + "content": "\\nabla_{z_t}\\mathcal{L}_{task}(\\tilde{a},\\mathcal{F}(\\hat{z}_0))" + }, + { + "bbox": [ + 55, + 102, + 296, + 162 + ], + "type": "text", + "content": " represents the gradient of the loss function computed based on the " + }, + { + "bbox": [ + 55, + 102, + 296, + 162 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathcal{SP}}" + }, + { + "bbox": [ + 55, + 102, + 296, + 162 + ], + "type": "text", + "content": " model's prediction of the clean image " + }, + { + "bbox": [ + 55, + 102, + 296, + 162 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 55, + 102, + 296, + 162 + ], + "type": "text", + "content": ". Specifically, this gradient is computed as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 169, + 296, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 169, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 296, + 205 + ], + "type": "interline_equation", + "content": "\\nabla_ {z _ {t}} \\mathcal {L} _ {\\text {t a s k}} (\\tilde {a}, \\mathcal {F} (\\hat {z} _ {0})) = \\nabla_ {z _ {t}} \\mathcal {L} _ {\\text {t a s k}} \\left(\\tilde {a}, \\mathcal {F} \\left(\\frac {z _ {t} - \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {\\theta}}{\\sqrt {\\alpha_ {t}}}\\right)\\right). \\tag {10}", + "image_path": "94091e7ff6e698ad3ff06a07cf4df74499b04d2314313234bf6e2fe90a4ddff4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 212, + 296, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 296, + 331 + ], + "type": "text", + "content": "Unlike previous approaches [18, 41] that required finetuning the diffusion model, such as those involving SD [37], to align with new data domains, our method leverages a " + }, + { + "bbox": [ + 55, + 212, + 296, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathcal{SP}}" + }, + { + "bbox": [ + 55, + 212, + 296, + 331 + ], + "type": "text", + "content": " model trained specifically on the target dataset. This model encapsulates critical information about the target task, effectively driving the diffusion process to generate data that is well-aligned with the target task without the need for additional finetuning. This not only conserves time and resources but also mitigates the risk of overfitting the diffusion model to the specific task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 331, + 296, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 416 + ], + "type": "text", + "content": "Proxy-Driven High-Realism Refinement. To ensure the high realism and accuracy of the generated images, we employ a trained " + }, + { + "bbox": [ + 55, + 331, + 296, + 416 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathcal{SP}}" + }, + { + "bbox": [ + 55, + 331, + 296, + 416 + ], + "type": "text", + "content": " model, to score each image. By setting a well-defined threshold, we filter out low-quality images based on their scores. The remaining images are then relabeled using the same " + }, + { + "bbox": [ + 55, + 331, + 296, + 416 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathcal{SP}}" + }, + { + "bbox": [ + 55, + 331, + 296, + 416 + ], + "type": "text", + "content": " model. This combined process of filtering and relabeling can be expressed as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 75, + 424, + 295, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 424, + 295, + 439 + ], + "spans": [ + { + "bbox": [ + 75, + 424, + 295, + 439 + ], + "type": "interline_equation", + "content": "\\{x, y \\} = \\left\\{x _ {i}, \\mathcal {F} _ {\\text {l a b e l}} \\left(x _ {i}\\right) \\mid \\operatorname {s c o r e} \\left(\\mathcal {F} \\left(x _ {i}\\right)\\right) \\geq \\tau \\right\\} \\tag {11}", + "image_path": "5a75e3f5ea59b01eeebdfd1d03ee5d7e32f1cab3521cc478071f1bd83c5b033c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 447, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 296, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 447, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(x_i)" + }, + { + "bbox": [ + 55, + 447, + 296, + 506 + ], + "type": "text", + "content": " represents the model's output score for image " + }, + { + "bbox": [ + 55, + 447, + 296, + 506 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 55, + 447, + 296, + 506 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 447, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 447, + 296, + 506 + ], + "type": "text", + "content": " is the quality threshold, and " + }, + { + "bbox": [ + 55, + 447, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{label}}" + }, + { + "bbox": [ + 55, + 447, + 296, + 506 + ], + "type": "text", + "content": " generates the accurate label " + }, + { + "bbox": [ + 55, + 447, + 296, + 506 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 55, + 447, + 296, + 506 + ], + "type": "text", + "content": " for the filtered image. This ensures that both the generated images and their labels are of high quality and closely match the desired data distribution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 509, + 137, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 137, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 137, + 522 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 529, + 176, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 176, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 176, + 541 + ], + "type": "text", + "content": "4.1. Experimental Setups" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": "Datasets and Evaluation Metric. To enhance practicality, we prioritize high-resolution and large-scale datasets. (1) Classification task: our experiments are primarily conducted on the full ImageNet 1K dataset [8], with IPC settings commonly set to 10 and 50, following the configurations in [41]. Additionally, based on the dataset partitioning settings from previous work [3], we also use the ImageNet 10-class subsets, such as ImageWoof, ImageFruit, ImageNette, and ImageMeow. These 10-class subsets contain very similar categories, making them challenging for dataset distillation. Top-1 accuracy " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": " is reported in the experimental tables. (2) Object detection task: we conduct experiments on the widely-used Pascal VOC [11, 12] and MS COCO [28] datasets, using mAP (Mean Average" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 72, + 533, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 533, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 533, + 84 + ], + "type": "text", + "content": "Algorithm 1 Universal Dataset Distillation Approach" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 85, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 85, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 314, + 85, + 554, + 109 + ], + "type": "text", + "content": "Input: Real data " + }, + { + "bbox": [ + 314, + 85, + 554, + 109 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 314, + 85, + 554, + 109 + ], + "type": "text", + "content": ", Diffusion model " + }, + { + "bbox": [ + 314, + 85, + 554, + 109 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 314, + 85, + 554, + 109 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 314, + 85, + 554, + 109 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathcal{SP}}" + }, + { + "bbox": [ + 314, + 85, + 554, + 109 + ], + "type": "text", + "content": " model " + }, + { + "bbox": [ + 314, + 85, + 554, + 109 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 314, + 85, + 554, + 109 + ], + "type": "text", + "content": ", Sampling steps " + }, + { + "bbox": [ + 314, + 85, + 554, + 109 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 314, + 85, + 554, + 109 + ], + "type": "text", + "content": ", Task-Driven strength " + }, + { + "bbox": [ + 314, + 85, + 554, + 109 + ], + "type": "inline_equation", + "content": "s(t)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 110, + 435, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 110, + 435, + 121 + ], + "spans": [ + { + "bbox": [ + 315, + 110, + 435, + 121 + ], + "type": "text", + "content": "Output: Synthetic dataset " + }, + { + "bbox": [ + 315, + 110, + 435, + 121 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 315, + 110, + 435, + 121 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 123, + 553, + 171 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 320, + 123, + 512, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 123, + 512, + 135 + ], + "spans": [ + { + "bbox": [ + 320, + 123, + 512, + 135 + ], + "type": "text", + "content": "1: Step I: Universal Task Knowledge Mining" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 136, + 489, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 136, + 489, + 147 + ], + "spans": [ + { + "bbox": [ + 320, + 136, + 489, + 147 + ], + "type": "text", + "content": "2: Train " + }, + { + "bbox": [ + 320, + 136, + 489, + 147 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathcal{SP}}" + }, + { + "bbox": [ + 320, + 136, + 489, + 147 + ], + "type": "text", + "content": " model " + }, + { + "bbox": [ + 320, + 136, + 489, + 147 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 320, + 136, + 489, + 147 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 320, + 136, + 489, + 147 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 320, + 136, + 489, + 147 + ], + "type": "text", + "content": " using Equ. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 147, + 553, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 147, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 320, + 147, + 553, + 171 + ], + "type": "text", + "content": "3: Use GPT-4o to generate task-aligned context for classification, detection, and segmentation datasets" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 320, + 171, + 553, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 171, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 320, + 171, + 553, + 194 + ], + "type": "text", + "content": "4: Step II: Universal Task-Driven Diffusion for Data Synthetic" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 318, + 195, + 553, + 314 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 320, + 195, + 482, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 195, + 482, + 207 + ], + "spans": [ + { + "bbox": [ + 320, + 195, + 482, + 207 + ], + "type": "text", + "content": "5: for each sampling step " + }, + { + "bbox": [ + 320, + 195, + 482, + 207 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 320, + 195, + 482, + 207 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 320, + 195, + 482, + 207 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 320, + 195, + 482, + 207 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 208, + 452, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 208, + 452, + 219 + ], + "spans": [ + { + "bbox": [ + 320, + 208, + 452, + 219 + ], + "type": "text", + "content": "6: Generate noisy sample " + }, + { + "bbox": [ + 320, + 208, + 452, + 219 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 320, + 208, + 452, + 219 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 220, + 492, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 220, + 492, + 230 + ], + "spans": [ + { + "bbox": [ + 320, + 220, + 492, + 230 + ], + "type": "text", + "content": "7: Predict clean image " + }, + { + "bbox": [ + 320, + 220, + 492, + 230 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 320, + 220, + 492, + 230 + ], + "type": "text", + "content": " using Equ. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 231, + 528, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 231, + 528, + 243 + ], + "spans": [ + { + "bbox": [ + 320, + 231, + 528, + 243 + ], + "type": "text", + "content": "8: Calculate guidance loss " + }, + { + "bbox": [ + 320, + 231, + 528, + 243 + ], + "type": "inline_equation", + "content": "\\nabla_{z_t}\\mathcal{L}_{task}(c,\\mathcal{F}(\\hat{z}_0))" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 243, + 495, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 243, + 495, + 255 + ], + "spans": [ + { + "bbox": [ + 320, + 243, + 495, + 255 + ], + "type": "text", + "content": "9: Adjust noise prediction using Equ. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 318, + 255, + 527, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 255, + 527, + 266 + ], + "spans": [ + { + "bbox": [ + 318, + 255, + 527, + 266 + ], + "type": "text", + "content": "10: Update " + }, + { + "bbox": [ + 318, + 255, + 527, + 266 + ], + "type": "inline_equation", + "content": "z_{t-1}" + }, + { + "bbox": [ + 318, + 255, + 527, + 266 + ], + "type": "text", + "content": " using adjusted noise prediction." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 318, + 267, + 365, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 267, + 365, + 277 + ], + "spans": [ + { + "bbox": [ + 318, + 267, + 365, + 277 + ], + "type": "text", + "content": "11: end for" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 318, + 279, + 511, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 279, + 511, + 290 + ], + "spans": [ + { + "bbox": [ + 318, + 279, + 511, + 290 + ], + "type": "text", + "content": "12: Score each generated image " + }, + { + "bbox": [ + 318, + 279, + 511, + 290 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 318, + 279, + 511, + 290 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 318, + 279, + 511, + 290 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(x_i)" + }, + { + "bbox": [ + 318, + 279, + 511, + 290 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 318, + 291, + 553, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 291, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 318, + 291, + 553, + 314 + ], + "type": "text", + "content": "13: Filter and relabel synthetic set based on threshold " + }, + { + "bbox": [ + 318, + 291, + 553, + 314 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 318, + 291, + 553, + 314 + ], + "type": "text", + "content": " using Equ. 11." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 339, + 554, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 554, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 554, + 387 + ], + "type": "text", + "content": "Precision) and " + }, + { + "bbox": [ + 313, + 339, + 554, + 387 + ], + "type": "inline_equation", + "content": "AP_{50}" + }, + { + "bbox": [ + 313, + 339, + 554, + 387 + ], + "type": "text", + "content": " (Average Precision at IoU threshold of 0.50) as evaluation metric. (3) Semantic segmentation task: we utilize the Pascal VOC dataset, with mIoU (Mean Intersection over Union) as evaluation metrics." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 388, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 639 + ], + "type": "text", + "content": "Implementation Details. Unless otherwise specified, the Task-Specific Proxy " + }, + { + "bbox": [ + 313, + 388, + 555, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathcal{SP}}" + }, + { + "bbox": [ + 313, + 388, + 555, + 639 + ], + "type": "text", + "content": " model used for the classification task is ResNet-18 [21], the detector is Faster R-CNN [36], and the segmentation model is LR-ASPP [24], all utilizing the official pre-trained weights from PyTorch. For the diffusion model, we use SD v1.5 [37] as the baseline. For quantitative comparisons, we generate " + }, + { + "bbox": [ + 313, + 388, + 555, + 639 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 388, + 555, + 639 + ], + "type": "text", + "content": " images for classification tasks and " + }, + { + "bbox": [ + 313, + 388, + 555, + 639 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 388, + 555, + 639 + ], + "type": "text", + "content": " images for detection and segmentation tasks. Synthetic images are created using 250 denoising steps based on a predefined number of random noise samples, aligned with the IPC settings. The threshold " + }, + { + "bbox": [ + 313, + 388, + 555, + 639 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 388, + 555, + 639 + ], + "type": "text", + "content": " for Proxy-Driven High-Realism Refinement is set to 0.9 for classification, 0.85 for detection, and 0.75 for segmentation. Generally, this operation results in a data discard rate of approximately " + }, + { + "bbox": [ + 313, + 388, + 555, + 639 + ], + "type": "inline_equation", + "content": "5 - 10\\%" + }, + { + "bbox": [ + 313, + 388, + 555, + 639 + ], + "type": "text", + "content": ". All experiments are conducted on NVIDIA V100 GPUs, with further details provided in the special experiments section. We generate the distilled dataset 3 times, train the network from scratch 5 times per dataset (with pre-trained backbone weights for the detector and segmenter), and evaluate on the original test set, reporting results as " + }, + { + "bbox": [ + 313, + 388, + 555, + 639 + ], + "type": "inline_equation", + "content": "\\bar{x} \\pm std" + }, + { + "bbox": [ + 313, + 388, + 555, + 639 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 647, + 451, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 451, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 451, + 659 + ], + "type": "text", + "content": "4.2. Results on Classification" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "We perform comprehensive comparisons across datasets of various scales. For large-scale datasets ImageNet-1k, we compare TESLA [7], SRe" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "L [46], D" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "inline_equation", + "content": "^4" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "M [41], RDED [42], and MiniMax Diffusion [18] (abbreviated as MiMxDiff in" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10561" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 294, + 247 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 247 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 247 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 247 + ], + "type": "table", + "html": "
DatasetIPCMethodAccuracy(%)
ImageNet-1K1000Full dataset69.8
10TESLA7.7
SRe2L21.3
D4M27.9
RDED42.0
MiMxDiff44.3
UniDD (Ours)50.4 (↑6.1)
50SRe2L46.8
D4M55.2
RDED56.5
MiMxDiff58.6
UniDD (Ours)62.8 (↑4.2)
", + "image_path": "af2b02f0dee62e5c55deefb8ab940f3c5d2775ef44dd83859d3426b594a6e0c0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 255, + 295, + 321 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 295, + 321 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 295, + 321 + ], + "type": "text", + "content": "Table 1. Performance comparison with state-of-the-art methods on large-scale dataset ImageNet-1K. All methods employ ResNet18 as test model. All standard deviations in this table are less than one. The results of Full dataset come from the official PyTorch. The best results are marked in bold, while the second-best results are underlined." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 342, + 295, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 378 + ], + "type": "text", + "content": "the table). For small-scale dataset ImageNet-Subset, we compare Random selection, DM [49], MTT [3], HaBa [29], DataDAM [38], and MiniMax Diffusion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 379, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 295, + 593 + ], + "type": "text", + "content": "Quantitative comparison on ImageNet-1K and ImageNet-Subset. As shown in the Tab 1, we utilize ResNet-18 as the " + }, + { + "bbox": [ + 55, + 379, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathcal{SP}}" + }, + { + "bbox": [ + 55, + 379, + 295, + 593 + ], + "type": "text", + "content": " model, trained on ImageNet-1K. We report results for IPC-10 and IPC-50, evaluating all outcomes with the ResNet-18 model. To ensure a fair comparison, our training and testing pipeline is aligned with that of RDED and " + }, + { + "bbox": [ + 55, + 379, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\mathrm{D^4M}" + }, + { + "bbox": [ + 55, + 379, + 295, + 593 + ], + "type": "text", + "content": ". Our method, UniDD, achieves state-of-the-art performance on this large-scale dataset. Specifically, at IPC-10, UniDD surpasses optimization-based methods, delivering substantial gains over TESLA and " + }, + { + "bbox": [ + 55, + 379, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\mathrm{SRe}^2\\mathrm{L}" + }, + { + "bbox": [ + 55, + 379, + 295, + 593 + ], + "type": "text", + "content": ". Additionally, it demonstrates significant advantages over diffusion-based methods, outperforming " + }, + { + "bbox": [ + 55, + 379, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\mathrm{D^4M}" + }, + { + "bbox": [ + 55, + 379, + 295, + 593 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 379, + 295, + 593 + ], + "type": "inline_equation", + "content": "22.5\\%" + }, + { + "bbox": [ + 55, + 379, + 295, + 593 + ], + "type": "text", + "content": " and MiniMax Diffusion by " + }, + { + "bbox": [ + 55, + 379, + 295, + 593 + ], + "type": "inline_equation", + "content": "6.1\\%" + }, + { + "bbox": [ + 55, + 379, + 295, + 593 + ], + "type": "text", + "content": ". These results underscore UniDD's effectiveness by using a lightweight " + }, + { + "bbox": [ + 55, + 379, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathcal{SP}}" + }, + { + "bbox": [ + 55, + 379, + 295, + 593 + ], + "type": "text", + "content": " model to guide SD denoising, aligning generative and target domains without requiring diffusion model finetuning, thus advancing performance in dataset distillation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 594, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 715 + ], + "type": "text", + "content": "As shown in the Tab 2, we conduct experiments on the ImageNet-subset under the IPC-10 setting. For consistency, the test model is standardized to a 5-layer ConvNet across all methods, except for MiniMax Diffusion, which uses a 6-layer ConvNet. Our method stands out among the comparison methods, outperforming the second-best by " + }, + { + "bbox": [ + 55, + 594, + 295, + 715 + ], + "type": "inline_equation", + "content": "8.2\\%" + }, + { + "bbox": [ + 55, + 594, + 295, + 715 + ], + "type": "text", + "content": " on ImageWoof, " + }, + { + "bbox": [ + 55, + 594, + 295, + 715 + ], + "type": "inline_equation", + "content": "9.1\\%" + }, + { + "bbox": [ + 55, + 594, + 295, + 715 + ], + "type": "text", + "content": " on ImageFruit, and " + }, + { + "bbox": [ + 55, + 594, + 295, + 715 + ], + "type": "inline_equation", + "content": "6.6\\%" + }, + { + "bbox": [ + 55, + 594, + 295, + 715 + ], + "type": "text", + "content": " on ImageMeow. Additionally, our method surpasses the diffusion-based MiniMax Diffusion by " + }, + { + "bbox": [ + 55, + 594, + 295, + 715 + ], + "type": "inline_equation", + "content": "9.8\\%" + }, + { + "bbox": [ + 55, + 594, + 295, + 715 + ], + "type": "text", + "content": " on ImageWoof, further demonstrating the effectiveness of UniDD approach in ag" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 507, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 507, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 507, + 84 + ], + "type": "text", + "content": "gregating information from the original dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 85, + 553, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 85, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 85, + 553, + 193 + ], + "type": "text", + "content": "Qualitative Result. In Fig. 2 (a), we present the synthetic images generated by UniDD on ImageWoof. Benefiting from the powerful generative capability of SD and the use of a " + }, + { + "bbox": [ + 313, + 85, + 553, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathcal{SP}}" + }, + { + "bbox": [ + 313, + 85, + 553, + 193 + ], + "type": "text", + "content": " model to align the real and synthetic data domains, our synthetic images for the classification task exhibit significantly higher realism compared to optimization-based methods [3, 46, 50]. However, unlike other diffusion-based dataset distillation methods [18, 41], we do not finetune SD, greatly reducing the deployment time of UniDD." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 203, + 519, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 203, + 519, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 519, + 216 + ], + "type": "text", + "content": "4.3. Results on Detection and Segmentation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 221, + 553, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 553, + 340 + ], + "type": "text", + "content": "We deploy our UniDD framework for detection and segmentation tasks and conduct experiments on the Pascal VOC and MS COCO datasets. Since dataset distillation is rarely applied to these tasks, we reference earlier dataset distillation works (DD [44], DC [50]) and compare UniDD with core-set selection methods. Specifically, we construct the following baseline methods: Random [35], K-center [39], and Herding [2, 5]. Additionally, we develop a method that selects a subset with balanced class representation, which we refer to as Uniform." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "text", + "content": "Quantitative Result. In Table 3, we compare coreset selection methods with UniDD across different compression ratios on the object detection task. According to the official MMDetection implementation [4], the detector Faster R-CNN [36] achieves " + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "inline_equation", + "content": "51.4\\%" + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "inline_equation", + "content": "80.3\\%" + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "text", + "content": " on Pascal VOC with the full dataset, and " + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "inline_equation", + "content": "32.6\\%" + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "inline_equation", + "content": "51.4\\%" + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "text", + "content": " on MS COCO, providing an approximate performance upper bound. On Pascal VOC, UniDD demonstrates significant performance gains, achieving " + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "inline_equation", + "content": "7.7\\%" + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "inline_equation", + "content": "19.2\\%" + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "text", + "content": " higher than the Random method at an extremely low compression ratio of " + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "text", + "content": ". Additionally, UniDD surpasses Uniform by " + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "inline_equation", + "content": "11.1\\%" + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "inline_equation", + "content": "21.2\\%" + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "text", + "content": " at a " + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "text", + "content": " compression ratio. On the more challenging MS COCO dataset, UniDD also shows a clear advantage, surpassing the best coreset selection method by " + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "inline_equation", + "content": "3.7\\%" + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "text", + "content": " mAP at compression ratios of " + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "inline_equation", + "content": "0.25\\%" + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 342, + 553, + 581 + ], + "type": "text", + "content": ", respectively. This indicates UniDD's robustness, achieving strong results even on more complex datasets, which can be attributed to the Task Knowledge Mining in the first stage that extracts sufficient task-relevant information." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "In Table 4, we present experimental results on the VOC dataset for the semantic segmentation task. Based on the MMSegmentation [6] implementation, LR-ASPP with a MobileNetV3 backbone achieves " + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "32.9\\%" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": " mIoU on the full dataset. UniDD outperforms coreset selection methods at all compression ratios, demonstrating its effectiveness in dataset distillation for segmentation tasks. This result further underscores UniDD's versatility across different tasks. By leveraging task-driven diffusion synthesis alongside task knowledge mining, UniDD generates synthetic data well-suited to complex, pixel-level predictions, illustrating its" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10562" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 548, + 150 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 548, + 150 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 548, + 150 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 548, + 150 + ], + "type": "table", + "html": "
DatasetRandomDMMTTHaBaDataDAMMiMxDiffUniDD (Ours)Full
ImageNette47.7±2.458.1±0.363.0 ±1.364.7 ±1.659.4±0.462.0±0.266.0±0.787.4 ±1.0
ImageWoof27.0±1.931.4±0.535.8 ±1.838.6 ±1.234.4±0.437.0±1.046.8±0.967.0 ±1.3
ImageFruit21.4±1.2-40.3 ±1.342.5 ±1.6--51.6±0.563.9 ±2.0
ImageMeow29.0±1.1-40.4 ±2.242.9 ±0.9--49.5±0.766.7 ±1.1
", + "image_path": "9ac17681a769dc58ea4f70fb98a6be3e550ad19b2c698138fcc146c167facb21.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 201, + 294, + 446 + ], + "blocks": [ + { + "bbox": [ + 55, + 159, + 555, + 183 + ], + "lines": [ + { + "bbox": [ + 55, + 159, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 555, + 183 + ], + "type": "text", + "content": "Table 2. Performance comparison with other state-of-the-art methods on ImageNet-Subset. IPC is set to 10. The best results are marked in bold, while the second-best results are underlined. “-” were not officially implemented in the original papers." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 201, + 294, + 446 + ], + "lines": [ + { + "bbox": [ + 58, + 201, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 58, + 201, + 294, + 446 + ], + "type": "table", + "html": "
MethodsObject Detection
Pascal VOCMS COCO
mAPAP50mAPAP50
Ratio0.5%0.25%
Random0.8±0.23.1±0.40.5±0.11.7±0.3
Uniform0.9±0.13.4±0.30.8±0.22.4±0.5
K-Center0.5±0.12.1±0.30.4±0.11.5±0.2
Herding0.6±0.22.4±0.20.5±0.11.8±0.4
UniDD (Ours)8.5±0.422.3±0.64.5±0.310.3±0.4
Ratio1%0.5%
Random4.2±0.513.7±0.63.7±0.210.1±0.3
Uniform5.7±0.217.7±0.43.4±0.49.5±0.6
K-Center3.6±0.612.3±0.33.2±0.59.3±0.5
Herding3.5±0.511.9±0.53.5±0.39.7±0.3
UniDD (Ours)16.8±0.538.9±0.77.1±0.416.9±0.3
Ratio2%1%
Random12.4±0.434.3±0.57.2±0.817.3±0.9
Uniform13.8±0.336.2±0.47.4±0.517.6±0.5
K-Center10.9±0.629.3±0.66.1±0.315.4±0.6
Herding10.4±0.428.7±0.76.7±0.416.3±0.7
UniDD (Ours)23.9±0.548.5±0.610.8±0.422.5±0.5
Full51.4±0.880.3±0.432.6±0.751.4±0.8
", + "image_path": "8ffe4f26155fc76f3ae3557195cd552a06d0dcbf656e869a985da23a576b5bcc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 512, + 294, + 588 + ], + "blocks": [ + { + "bbox": [ + 55, + 454, + 295, + 498 + ], + "lines": [ + { + "bbox": [ + 55, + 454, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 295, + 498 + ], + "type": "text", + "content": "Table 3. Performance comparison with coreset selection methods on Pascal VOC and MS COCO for the object detection task, using Faster R-CNN as the detector. \"Ratio\" denotes the compression ratio." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 512, + 294, + 588 + ], + "lines": [ + { + "bbox": [ + 58, + 512, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 58, + 512, + 294, + 588 + ], + "type": "table", + "html": "
MethodsSemantic Segmentation (mIoU)
Ratio: 3.5%Ratio: 7%Ratio: 14%Full
Random6.7±0.49.6±0.511.3±0.4
Uniform6.5±0.69.3±0.412.1±0.5
K-Center6.1±0.48.5±0.610.4±0.732.9±0.6
Herding6.2±0.78.7±0.410.6±0.3
UniDD (Ours)10.3±0.512.3±0.314.9±0.7
", + "image_path": "18f5cb838b2dda3a4d1303afa850090e13a56caf7efc736833c2f0c0ee600c6c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 664, + 220, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 664, + 220, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 664, + 220, + 676 + ], + "type": "text", + "content": "versatility across a range of vision tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Qualitative Result. As shown in Fig. 2 (b) and 2 (c), we present visualizations of detection and segmentation data synthesized using the UniDD framework. These images" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 203, + 555, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 203, + 555, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 555, + 334 + ], + "type": "text", + "content": "are generated with " + }, + { + "bbox": [ + 313, + 203, + 555, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathcal{SP}}" + }, + { + "bbox": [ + 313, + 203, + 555, + 334 + ], + "type": "text", + "content": " models, Faster R-CNN for detection and LR-ASPP for segmentation, which effectively steer the diffusion process toward the specified labels, bounding boxes, and masks. Previous research in dataset distillation has predominantly focused on single-category label synthesis, where each image aligns with a single class. In contrast, our UniDD framework overcomes this limitation, enabling dataset distillation for more complex, multi-class tasks. This advancement not only broadens the applicability of dataset distillation but also offers valuable insights for future research in this domain." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 343, + 408, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 408, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 408, + 357 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 361, + 556, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 556, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 556, + 554 + ], + "type": "text", + "content": "Component Analysis. We conduct a comprehensive ablation study on the proposed UniDD, analyzing its components: (1) Baseline (a standard SD v1.5 checkpoint), (2) Task-driven Diffusion with the " + }, + { + "bbox": [ + 313, + 361, + 556, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathcal{SP}}" + }, + { + "bbox": [ + 313, + 361, + 556, + 554 + ], + "type": "text", + "content": " model, (3) Task-Aligned Contextual Prompting (abbreviated as Context), and (4) Proxy-Driven High-Realism Refinement (abbreviated as Refine). As shown in Table 5, the proposed task-driven approach significantly improves the performance of the baseline, demonstrating that this guidance effectively bridges the gap between real and synthetic datasets. The contextual information is more effective for more complex detection and segmentation datasets, as it reveals underlying target relationships. The refinement module also proves to be effective, as the randomness of the SD model can lead to the generation of images that deviate from the annotations or result in failed generations." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 315, + 564, + 553, + 624 + ], + "blocks": [ + { + "bbox": [ + 55, + 596, + 295, + 640 + ], + "lines": [ + { + "bbox": [ + 55, + 596, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 295, + 640 + ], + "type": "text", + "content": "Table 4. Performance comparison with coreset selection methods on Pascal VOC for the semantic segmentation task, using LR-ASPP as the segmenter. The metric is mIoU, and \"Ratio\" denotes the compression ratio." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 564, + 553, + 624 + ], + "lines": [ + { + "bbox": [ + 315, + 564, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 315, + 564, + 553, + 624 + ], + "type": "table", + "html": "
MethodsDataset
Cls. / Accuracy (%)Det. / mAP (%)Seg. / mIoU (%)
ImageNetteImageWoofVOCVOC
Baseline60.4±0.939.1±1.111.5±0.27.2±0.3
Baseline+TSP64.2±0.845.3±0.614.7±0.48.9±0.5
Baseline+TSRP+Context64.7±0.446.1±0.715.9±0.39.6±0.4
Baseline+TSP+Context+Refine66.0±0.746.8±0.916.8±0.510.3±0.5
", + "image_path": "7f22ca9639b56d5f84fd761931c9b64315b1de352de9424b41fd7c3c733dfecb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 632, + 555, + 687 + ], + "lines": [ + { + "bbox": [ + 313, + 632, + 555, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 555, + 687 + ], + "type": "text", + "content": "Table 5. Ablation study on UniDD. For classification datasets, IPC is set to 10 with ResNet-18 as the test model. For detection on VOC, the compression ratio is set to " + }, + { + "bbox": [ + 313, + 632, + 555, + 687 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 632, + 555, + 687 + ], + "type": "text", + "content": " with Faster R-CNN as the test model. For segmentation on VOC, the ratio is set to " + }, + { + "bbox": [ + 313, + 632, + 555, + 687 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 313, + 632, + 555, + 687 + ], + "type": "text", + "content": " with LR-ASPP as the test model." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "content": "Cross-Architecture Analysis. Cross-Architecture experi" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10563" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 294, + 164 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 164 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 164 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 164 + ], + "type": "table", + "html": "
DatasetTest Model\\( T_{SP} \\)Model
ResNet-18ResNet-50ResNet-101
ImageNetteResNet-1866.0±0.768.8±0.769.5±0.7
ResNet-5067.2±0.669.5±1.170.7±0.9
ResNet-10169.3±0.970.9±1.272.1±0.6
ImageWoofResNet-1846.8±0.949.2±0.550.1±0.8
ResNet-5047.3±0.750.0±0.850.9±0.6
ResNet-10148.9±1.151.5±0.652.4±0.7
", + "image_path": "8167e876d94a483cab78fcb5c4737e0d923e8288bd77a481dbabd652b184c829.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 172, + 295, + 216 + ], + "lines": [ + { + "bbox": [ + 55, + 172, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 172, + 295, + 216 + ], + "type": "text", + "content": "Table 6. " + }, + { + "bbox": [ + 55, + 172, + 295, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathcal{SP}}" + }, + { + "bbox": [ + 55, + 172, + 295, + 216 + ], + "type": "text", + "content": " Model Ablation and Cross-Architecture Testing: We use ResNet-18, ResNet-50, and ResNet-101 as " + }, + { + "bbox": [ + 55, + 172, + 295, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathcal{SP}}" + }, + { + "bbox": [ + 55, + 172, + 295, + 216 + ], + "type": "text", + "content": " models to generate synthetic images. During testing, the same three models are used. IPC is set to 10." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 63, + 236, + 167, + 339 + ], + "blocks": [ + { + "bbox": [ + 63, + 236, + 167, + 339 + ], + "lines": [ + { + "bbox": [ + 63, + 236, + 167, + 339 + ], + "spans": [ + { + "bbox": [ + 63, + 236, + 167, + 339 + ], + "type": "image", + "image_path": "710da0b5dfd0a15b8f27cf0cd423dfc27eb92886c2f2843ff03c914356ef0257.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 73, + 341, + 168, + 350 + ], + "lines": [ + { + "bbox": [ + 73, + 341, + 168, + 350 + ], + "spans": [ + { + "bbox": [ + 73, + 341, + 168, + 350 + ], + "type": "text", + "content": "(a) Accuracy vs. Guidance strength" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 178, + 236, + 282, + 338 + ], + "blocks": [ + { + "bbox": [ + 178, + 236, + 282, + 338 + ], + "lines": [ + { + "bbox": [ + 178, + 236, + 282, + 338 + ], + "spans": [ + { + "bbox": [ + 178, + 236, + 282, + 338 + ], + "type": "image", + "image_path": "28c224eb7f4d1179d114e2528c56f2684e8881d26b53783ef767363b2beb0521.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 196, + 341, + 276, + 350 + ], + "lines": [ + { + "bbox": [ + 196, + 341, + 276, + 350 + ], + "spans": [ + { + "bbox": [ + 196, + 341, + 276, + 350 + ], + "type": "text", + "content": "(b) Accuracy vs. Sample Step" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 61, + 411, + 170, + 517 + ], + "blocks": [ + { + "bbox": [ + 55, + 360, + 295, + 392 + ], + "lines": [ + { + "bbox": [ + 55, + 360, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 295, + 392 + ], + "type": "text", + "content": "Figure 4. Hyper-parameter analysis on (a) guidance strength " + }, + { + "bbox": [ + 55, + 360, + 295, + 392 + ], + "type": "inline_equation", + "content": "\\lambda_{s}" + }, + { + "bbox": [ + 55, + 360, + 295, + 392 + ], + "type": "text", + "content": " and (b) sampling step " + }, + { + "bbox": [ + 55, + 360, + 295, + 392 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 360, + 295, + 392 + ], + "type": "text", + "content": ". The star symbol indicates the parameters chosen for trade-off." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 411, + 170, + 517 + ], + "lines": [ + { + "bbox": [ + 61, + 411, + 170, + 517 + ], + "spans": [ + { + "bbox": [ + 61, + 411, + 170, + 517 + ], + "type": "image", + "image_path": "56268c4bf81de374ccba9e17d01f26bc4e2c049a1a2d33703cbd143c4e5f24bf.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 520, + 159, + 529 + ], + "lines": [ + { + "bbox": [ + 82, + 520, + 159, + 529 + ], + "spans": [ + { + "bbox": [ + 82, + 520, + 159, + 529 + ], + "type": "text", + "content": "(a) Training Time overhead" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 538, + 295, + 572 + ], + "lines": [ + { + "bbox": [ + 55, + 538, + 295, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 295, + 572 + ], + "type": "text", + "content": "Figure 5. Comparison of distillation cost on ImageNet-1k using our UniDD with comparison methods (MTT [3], " + }, + { + "bbox": [ + 55, + 538, + 295, + 572 + ], + "type": "inline_equation", + "content": "\\mathrm{SRe}^2\\mathrm{L}" + }, + { + "bbox": [ + 55, + 538, + 295, + 572 + ], + "type": "text", + "content": " [46], and " + }, + { + "bbox": [ + 55, + 538, + 295, + 572 + ], + "type": "inline_equation", + "content": "\\mathrm{D^4M}" + }, + { + "bbox": [ + 55, + 538, + 295, + 572 + ], + "type": "text", + "content": " [41]), tested on V100 GPUs." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 176, + 411, + 284, + 517 + ], + "blocks": [ + { + "bbox": [ + 176, + 411, + 284, + 517 + ], + "lines": [ + { + "bbox": [ + 176, + 411, + 284, + 517 + ], + "spans": [ + { + "bbox": [ + 176, + 411, + 284, + 517 + ], + "type": "image", + "image_path": "3980bfe19f23ab16a4f8b59e65acfff76214a08a92e6d61e9526d8bf476c7cb1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 520, + 272, + 529 + ], + "lines": [ + { + "bbox": [ + 195, + 520, + 272, + 529 + ], + "spans": [ + { + "bbox": [ + 195, + 520, + 272, + 529 + ], + "type": "text", + "content": "(b) Training GPU overhead" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "ments are crucial for validating the generalizability of synthetic data. We compare different combinations of " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathcal{SP}}" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": " models and testing models. In this study, we set up three " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathcal{SP}}" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": " models: ResNet-18, ResNet-50, and ResNet-101. These models are also used during testing to evaluate the cross-architecture performance of the synthetic datasets. As shown in Table. 6, using more complex " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathcal{SP}}" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": " models improves the guidance effectiveness. Moreover, when testing across different architectures, the performance does not degrade as seen in previous optimization-based meth" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "ods [3, 49, 50], demonstrating the generalizability of our synthetic data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 96, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 553, + 179 + ], + "type": "text", + "content": "Guidance Strength Analysis. As shown in Fig. 4 (a), we analyze the performance variations of ImageWoof and ImageNette at IPC-10 under different guidance strength settings " + }, + { + "bbox": [ + 313, + 96, + 553, + 179 + ], + "type": "inline_equation", + "content": "\\lambda_{s}" + }, + { + "bbox": [ + 313, + 96, + 553, + 179 + ], + "type": "text", + "content": ", with Conv5 as the test model. The best results are achieved at " + }, + { + "bbox": [ + 313, + 96, + 553, + 179 + ], + "type": "inline_equation", + "content": "\\lambda_{s} = 100" + }, + { + "bbox": [ + 313, + 96, + 553, + 179 + ], + "type": "text", + "content": ", but further increasing the guidance strength interferes with the quality of diffusion-generated images, leading to a decline in test performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 180, + 553, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 274 + ], + "type": "text", + "content": "Sample Step Analysis. As shown in Fig. 4 (b), we analyze the impact of different SD sample steps on the performance of synthetic datasets under the IPC-10 setting, with Image-Woof and ImageNette resolutions set to " + }, + { + "bbox": [ + 313, + 180, + 553, + 274 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 180, + 553, + 274 + ], + "type": "text", + "content": ". Performance improves significantly with sample steps below 250, but gains are minimal beyond this point while time consumption increases. Thus, we select 250 steps as the optimal trade-off." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 276, + 555, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 555, + 444 + ], + "type": "text", + "content": "Distillation Cost Analysis. As shown in Fig. 5, we evaluate the time and GPU cost of various DD methods. Previous methods primarily focus on the time required to synthesize images, often neglecting the time needed to train or deploy the DD model. For MTT [3], the deployment time corresponds to the expert trajectory training time, while for " + }, + { + "bbox": [ + 313, + 276, + 555, + 444 + ], + "type": "inline_equation", + "content": "\\mathrm{D^4M}" + }, + { + "bbox": [ + 313, + 276, + 555, + 444 + ], + "type": "text", + "content": " [41], it is the time required to finetune the SD model. The slight increase in UniDD's distillation time and GPU cost is mainly due to the added supervision process of the " + }, + { + "bbox": [ + 313, + 276, + 555, + 444 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathcal{SP}}" + }, + { + "bbox": [ + 313, + 276, + 555, + 444 + ], + "type": "text", + "content": " model and the refinement stage, where approximately " + }, + { + "bbox": [ + 313, + 276, + 555, + 444 + ], + "type": "inline_equation", + "content": "5\\% -10\\%" + }, + { + "bbox": [ + 313, + 276, + 555, + 444 + ], + "type": "text", + "content": " of low-quality generated data is discarded. Combining the deployment and distillation costs for comparison, UniDD demonstrates the best advantage, indicating its superior suitability for practical deployment." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 454, + 388, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 388, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 388, + 466 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 474, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 689 + ], + "type": "text", + "content": "In this study, we introduce the first Universal Dataset Distillation (UniDD) framework, capable of handling classification, detection, and segmentation tasks within a unified approach. UniDD follows a two-stage process: in the first stage, Task-Specific Proxy Training and Task-Aligned Contextual Prompting extract critical information from real datasets. In the second stage, Universal Taskdriven Diffusion synthesizes images with specific categories, bounding boxes, and segmentation masks tailored to each task. We conduct extensive experiments to validate UniDD's effectiveness across all three tasks. For classification, UniDD achieves state-of-the-art results on ImageNet1k and ImageNet-Subset. In detection and segmentation, we establish the first dataset distillation benchmark on Pascal VOC and MS COCO, comparing UniDD to coreset selection methods. Encouraging results demonstrate that our UniDD consistently outperforms these methods, especially under extremely low compression ratios." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Acknowledgments. This work was supported by National Natural Science Fund of China (62473286)." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10564" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "text", + "content": "[1] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 843-852, 2023. 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 158, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 158, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 296, + 201 + ], + "type": "text", + "content": "[2] Francisco M Castro, Manuel J Marín-Jiménez, Nicolas Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European conference on computer vision (ECCV), pages 233-248, 2018. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 202, + 295, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 295, + 257 + ], + "type": "text", + "content": "[3] George Cazenavette, Tongzhou Wang, Antonio Torralba, Alexei A Efros, and Jun-Yan Zhu. Dataset distillation by matching training trajectories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4750-4759, 2022. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 258, + 296, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 296, + 344 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 296, + 344 + ], + "type": "text", + "content": "[4] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 346, + 295, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 295, + 378 + ], + "type": "text", + "content": "[5] Yutian Chen, Max Welling, and Alex Smola. Super-samples from kernel herding. arXiv preprint arXiv:1203.3472, 2012. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "type": "text", + "content": "[6] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 424, + 295, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 424, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 62, + 424, + 295, + 467 + ], + "type": "text", + "content": "[7] Justin Cui, Ruochen Wang, Si Si, and Cho-Jui Hsieh. Scaling up dataset distillation toImagenet-1k with constant memory. In International Conference on Machine Learning, pages 6565–6590. PMLR, 2023. 1, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 468, + 295, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 295, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 295, + 512 + ], + "type": "text", + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1, 2, 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 514, + 295, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 295, + 546 + ], + "type": "text", + "content": "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "type": "text", + "content": "[10] Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. Advances in Neural Information Processing Systems, 36:16222-16239, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "text", + "content": "[11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 1, 2, 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "text", + "content": "[12] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. International journal of computer vision, 111:98-136, 2015. 1, 2, 3, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 295, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 715 + ], + "type": "text", + "content": "[13] Junyao Gao, Xinyang Jiang, Huishuai Zhang, Yifan Yang, Shuguang Dou, Dongsheng Li, Duoqian Miao, Cheng Deng," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 116 + ], + "type": "text", + "content": "and Cairong Zhao. Similarity distribution based membership inference attack on person re-identification. In Proceedings of the AAAI conference on artificial intelligence, pages 14820-14828, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "text", + "content": "[14] Junyao Gao, Xinyang Jiang, Shuguang Dou, Dongsheng Li, Duoqian Miao, and Cairong Zhao. Re-id-leak: Membership inference attacks against person re-identification. International Journal of Computer Vision, 132(10):4673-4687, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 173, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 215 + ], + "type": "text", + "content": "[15] Junyao Gao, Yanchen Liu, Yanan Sun, Yinhao Tang, Yanhong Zeng, Kai Chen, and Cairong Zhao. Styleshot: A snapshot on any style. arXiv preprint arXiv:2407.01414, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 249 + ], + "type": "text", + "content": "[16] Junyao Gao, Yanan Sun, Fei Shen, Xin Jiang, Zhening Xing, Kai Chen, and Cairong Zhao. Faceshot: Bring any character into life. arXiv preprint arXiv:2503.00740, 2025. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 250, + 553, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 294 + ], + "type": "text", + "content": "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 294, + 555, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 555, + 348 + ], + "type": "text", + "content": "[18] Jianyang Gu, Saeed Vahidian, Vyacheslav Kungurtsev, Haonan Wang, Wei Jiang, Yang You, and Yiran Chen. Efficient dataset distillation via minimax diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15793-15803, 2024. 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "text", + "content": "[19] Ziyao Guo, Kai Wang, George Cazenavette, Hui Li, Kaipeng Zhang, and Yang You. Towards lossless dataset distillation via difficulty-aligned trajectory matching. arXiv preprint arXiv:2310.05773, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 393, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 437 + ], + "type": "text", + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 438, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 481 + ], + "type": "text", + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 482, + 553, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 535 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 535 + ], + "type": "text", + "content": "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 537, + 553, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 559 + ], + "type": "text", + "content": "[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 559, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 613 + ], + "type": "text", + "content": "[24] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1314-1324, 2019. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 614, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 657 + ], + "type": "text", + "content": "[25] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[26] Jang-Hyun Kim, Jinuk Kim, Seong Joon Oh, Sangwoo Yun, Hwanjun Song, Joonhyun Jeong, Jung-Woo Ha, and Hyun Oh Song. Dataset condensation via efficient synthetic-data parameterization. In International Conference on Machine Learning, pages 11102-11118. PMLR, 2022. 1, 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10565" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "text", + "content": "[27] Jian Li, Yabiao Wang, Changan Wang, Ying Tai, Jianjun Qian, Jian Yang, Chengjie Wang, Jilin Li, and Feiyue Huang. Dsfd: dual shot face detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5060-5069, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 194 + ], + "type": "text", + "content": "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1, 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 195, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 237 + ], + "type": "text", + "content": "[29] Songhua Liu, Kai Wang, Xingyi Yang, Jingwen Ye, and Xin chao Wang. Dataset distillation via factorization. Advances in Neural Information Processing Systems, 35:1100-1113, 2022. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 238, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 238, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 238, + 295, + 304 + ], + "type": "text", + "content": "[30] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 289-299, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 304, + 294, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 346 + ], + "type": "text", + "content": "[31] Yanqing Liu, Jianyang Gu, Kai Wang, Zheng Zhu, Wei Jiang, and Yang You. Dream: Efficient dataset distillation by representative matching. arXiv preprint arXiv:2302.14416, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 349, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 349, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 349, + 294, + 381 + ], + "type": "text", + "content": "[32] Timothy Nguyen, Zhourong Chen, and Jaehoon Lee. Dataset meta-learning from kernel ridge-regression. In International Conference on Learning Representations, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 382, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 294, + 414 + ], + "type": "text", + "content": "[33] OpenAI. Gpt-4 technical report. https://openai.com/research/gpt-4, 2023. Accessed: 2024-11-13. 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 470 + ], + "type": "text", + "content": "[34] Ding Qi, Jian Li, Jinlong Peng, Bo Zhao, Shuguang Dou, Jialin Li, Jiangning Zhang, Yabiao Wang, Chengjie Wang, and Cairong Zhao. Fetch and forge: Efficient dataset condensation for object detection. Advances in Neural Information Processing Systems, 37:119283-119300, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "text", + "content": "[35] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2001-2010, 2017. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 525, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 294, + 569 + ], + "type": "text", + "content": "[36] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 4, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 570, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 295, + 624 + ], + "type": "text", + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 295, + 680 + ], + "type": "text", + "content": "[38] Ahmad Sajedi, Samir Khaki, Ehsan Amjadian, Lucy Z Liu, Yuri A Lawryshyn, and Konstantinos N Plataniotis. Datadam: Efficient dataset distillation with attention matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17097-17107, 2023. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 681, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 712 + ], + "type": "text", + "content": "[39] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. arXiv preprint arXiv:1708.00489, 2017. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 619 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "text", + "content": "[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "text", + "content": "[41] Duo Su, Junjie Hou, Weizhi Gao, Yingjie Tian, and Bowen Tang. D^4: Dataset distillation via disentangled diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5809-5818, 2024. 5, 6, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "text", + "content": "[42] Peng Sun, Bei Shi, Daiwei Yu, and Tao Lin. On the diversity and realism of distilled dataset: An efficient dataset distillation paradigm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9390-9399, 2024. 1, 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 219, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 282 + ], + "type": "text", + "content": "[43] Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Shuo Yang, Shuo Wang, Guan Huang, Hakan Bilen, Xinchao Wang, and Yang You. Cafe: Learning to condense dataset by aligning features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12196-12205, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 285, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 317 + ], + "type": "text", + "content": "[44] Tongzhou Wang, Jun-Yan Zhu, Antonio Torralba, and Alexei A Efros. Dataset distillation. arXiv preprint arXiv:1811.10959, 2018. 1, 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "type": "text", + "content": "[45] Yichao Yan, Zanwei Zhou, Zi Wang, Jingnan Gao, and Xiaokang Yang. Dialoguenerf: Towards realistic avatar face-to-face conversation video generation. Visual Intelligence, 2 (1):24, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 365, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 407 + ], + "type": "text", + "content": "[46] Zeyuan Yin, Eric Xing, and Zhiqiang Shen. Squeeze, recover and relabel: Dataset condensation at imagenet scale from a new perspective. arXiv preprint arXiv:2306.13092, 2023. 1, 3, 5, 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 409, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 409, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 409, + 553, + 453 + ], + "type": "text", + "content": "[47] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 454, + 553, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 454, + 553, + 486 + ], + "spans": [ + { + "bbox": [ + 316, + 454, + 553, + 486 + ], + "type": "text", + "content": "[48] Bo Zhao and Hakan Bilen. Dataset condensation with differentiable siamese augmentation. In International Conference on Machine Learning, pages 12674-12685. PMLR, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 487, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 553, + 531 + ], + "type": "text", + "content": "[49] Bo Zhao and Hakan Bilen. Dataset condensation with distribution matching. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6514-6523, 2023. 1, 2, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 533, + 553, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 565 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 565 + ], + "type": "text", + "content": "[50] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. Dataset condensation with gradient matching. In International Conference on Learning Representations, 2020. 1, 2, 6, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 567, + 553, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 619 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 619 + ], + "type": "text", + "content": "[51] Ganlong Zhao, Guanbin Li, Yipeng Qin, and Yizhou Yu. Improved distribution matching for dataset condensation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7856-7865, 2023. 1, 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10566" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Towards Universal Soccer Video Understanding/bd5f526c-bc30-45cc-bbcd-e151dfc21f1a_content_list.json b/2025/Towards Universal Soccer Video Understanding/bd5f526c-bc30-45cc-bbcd-e151dfc21f1a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b6897eacf3aca0981e9318205957953804b1ea23 --- /dev/null +++ b/2025/Towards Universal Soccer Video Understanding/bd5f526c-bc30-45cc-bbcd-e151dfc21f1a_content_list.json @@ -0,0 +1,1840 @@ +[ + { + "type": "text", + "text": "Towards Universal Soccer Video Understanding", + "text_level": 1, + "bbox": [ + 256, + 130, + 741, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiayuan Rao $^{1,2*}$ , Haoning Wu $^{1,2*}$ , Hao Jiang $^{3}$ , Ya Zhang $^{1}$ , Yanfeng Wang $^{1\\dagger}$ , Weidi Xie $^{1\\dagger}$ $^{1}$ School of Artificial Intelligence, Shanghai Jiao Tong University, China \n $^{2}$ CMIC, Shanghai Jiao Tong University, China \nhttps://jyrao.github.io/UniSoccer/", + "bbox": [ + 151, + 179, + 843, + 262 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 311, + 326, + 328 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As a globally celebrated sport, soccer has attracted widespread interest from fans all over the world. This paper aims to develop a comprehensive multi-modal framework for soccer video understanding. Specifically, we make the following contributions in this paper: (i) we introduce SoccerReplay-1988, the largest multi-modal soccer dataset to date, featuring videos and detailed annotations from 1,988 complete matches, with an automated annotation pipeline; (ii) we present an advanced soccer-specific visual encoder, MatchVision, which leverages spatiotemporal information across soccer videos and excels in various downstream tasks; (iii) we conduct extensive experiments and ablation studies on event classification, commentary generation, and multi-view foul recognition. MatchVision demonstrates state-of-the-art performance on all of them, substantially outperforming existing models, which highlights the superiority of our proposed data and model. We believe that this work will offer a standard paradigm for sports understanding research.", + "bbox": [ + 88, + 342, + 483, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"Football is one of the world's best means of communication. It is impartial, apolitical, and universal.\"", + "bbox": [ + 89, + 637, + 482, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "— Franz Beckenbauer (1945 - 2024)", + "bbox": [ + 223, + 667, + 482, + 683 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 704, + 220, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Soccer, celebrated worldwide for its significant commercial value, has recently seen great research interest in integrating artificial intelligence (AI) for soccer video understanding. This is primarily motivated by the sport's complexity and the growing demand for enhanced analytics and improved viewing experiences. AI systems facilitate tactical analysis [49], allowing coaches to devise better strategies by uncovering patterns not apparent to the naked eye. In addition, it also supports automated content generation and enriches", + "bbox": [ + 89, + 729, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b768a20c91a1e5521d31cbbdecbc86cf5ce93b5b2297350c47f6ff106847ef5b.jpg", + "image_caption": [ + "Videos in SoccerReplay-1988" + ], + "image_footnote": [], + "bbox": [ + 519, + 297, + 547, + 315 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d8fceea74ff9cd5dcca5ddf474b1fa93d9e505c697fe8845e096779cc9f8715b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 300, + 777, + 315 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ed384a51e59726eec53f918b94be181ccdfedb6766313c36bbcbed6f6b542de5.jpg", + "image_caption": [ + "Figure 1. Overview. We present SoccerReplay-1988, the largest soccer dataset to date, and a powerful soccer-specific visual encoder, MatchVision, capable of excelling in various tasks such as event classification and commentary generation." + ], + "image_footnote": [], + "bbox": [ + 516, + 319, + 903, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "fan engagement through interactive and personalized content [34, 36, 40]. These capabilities promote a deeper understanding of soccer, simplify content creation, and foster a more engaging experience for fans and professionals.", + "bbox": [ + 511, + 626, + 906, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing research in soccer video analysis primarily revolves around the SoccerNet series datasets [7, 10, 15], which comprise 500 full-match videos for benchmarking various tasks, such as action spotting [10, 15] and commentary generation [34, 36, 40]. Despite this extensive coverage, the focus has predominantly been on designing specialized models for task-specific applications, leading to fragmented and incompatible solutions. Such fragmentation underscores the need for a unified framework capable of integrating diverse demands, enabling more holistic and scalable advancements in soccer video understanding.", + "bbox": [ + 511, + 686, + 908, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we introduce SoccerReplay-1988, the largest and most comprehensive multi-modal soccer video dataset to date, featuring 1,988 complete match videos with", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*: These authors contribute equally to this work.", + "bbox": [ + 89, + 875, + 349, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†: Corresponding author.", + "bbox": [ + 91, + 888, + 225, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8384", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "rich annotations, such as event labels and textual commentaries. This dataset offers a solid foundation for developing advanced soccer understanding models and establishes a challenging new benchmark for the field. Additionally, we have harmonized existing datasets to be compatible with ours, further expanding the available data resources.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Leveraging this dataset, we develop MatchVision, an advanced soccer-specific visual encoder tailored for diverse soccer understanding tasks. It employs the cutting-edge visual-language foundation model as the backbone, e.g., SigLIP [58]. We further extend framewise visual features into spatiotemporal representations with temporal attentions [3], by training on diverse visual-language tasks on SoccerReplay-1988, as depicted in Figure 1. As a result, MatchVision exhibits strong adaptability across various tasks, such as event classification and commentary generation, serving as a universal and unified framework for comprehensive soccer video understanding.", + "bbox": [ + 89, + 181, + 482, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, we make the following contributions in this paper: (i) we construct SoccerReplay-1988, the largest and most diverse soccer video dataset to date, featuring videos of 1,988 soccer matches with rich annotations, supported by an automated curation pipeline. This provides a solid foundation for developing robust and comprehensive soccer understanding models; (ii) we present a powerful soccer-specific visual encoder, termed MatchVision, which effectively leverages spatiotemporal information in soccer videos, and can adapt to various tasks such as event classification and commentary generation, serving as a unified framework for soccer understanding; (iii) we establish more comprehensive and challenging benchmarks based on our dataset, enabling more professional evaluation of soccer understanding models; (iv) extensive experiments and ablation studies demonstrate the superiority of our data and model across various downstream tasks, achieving state-of-the-art performance on both existing benchmarks and our newly established ones. We believe this work offers a viable paradigm for future sports video understanding.", + "bbox": [ + 91, + 363, + 482, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 676, + 238, + 691 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sports Understanding [44] is an evolving field that encompasses multiple research topics and integrates diverse data modalities, covering various tasks such as action spotting [10, 15, 16], commentary generation [34, 37, 40, 51, 52, 57], athlete analysis [41, 55], tactical planning [49], sports health [39], and intelligent refereeing [22, 23]. Furthermore, with the rapid development of multimodal large language models (MLLMs), recent efforts [26, 50, 53, 54] have attempted to build more generalized frameworks to uniformly handle a variety of sports understanding tasks.", + "bbox": [ + 89, + 700, + 482, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual-Language Models [1, 27, 28, 38, 58] have exhibited remarkable performance across extensive applications like classification, segmentation, image-text retrieval, and image", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "captioning. Recent efforts have ventured into more challenging video understanding [29, 30, 42, 59, 60] tasks, such as temporal alignment [17, 31], dense captioning [5, 56, 62], and audio description [18-20]. However, these efforts typically focus on general scenarios, limiting their adaptability to specific professional fields. Thus, this paper aims to bridge this gap by advancing visual-language models tailored for comprehensive soccer understanding.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Soccer Game Analysis [9] has primarily focused on tasks such as action spotting [15], replay grounding [10, 61], commentary generation [34, 36, 40], player tracking [8], state reconstruction [43], camera calibration [6, 8, 10] and foul recognition [22, 23], as facilitated by the SoccerNet [7, 10, 14, 15] series datasets, with 500 full-match videos from 2015 to 2017. Unlike existing methods that target designing specific models for distinct tasks, this paper aims to design a unified multi-modal framework that leverages spatiotemporal information within videos, serving as a specialized visual encoder for soccer video understanding.", + "bbox": [ + 511, + 214, + 906, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. SoccerReplay-1988 Dataset", + "text_level": 1, + "bbox": [ + 511, + 397, + 764, + 414 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To establish a solid foundation for soccer understanding, we construct SoccerReplay-1988, the largest soccer dataset to date. Here, we first outline our data collection details and an overview of the dataset in Sec. 3.1; followed by elaborating on our automated data curation pipeline in Sec. 3.2; lastly, in Sec. 3.3, we present the data statistics and discussion.", + "bbox": [ + 511, + 422, + 903, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Dataset Collection", + "text_level": 1, + "bbox": [ + 511, + 525, + 689, + 539 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To construct the SoccerReplay-1988 dataset, we have collected untrimmed, full-match videos from the Internet, encompassing a total of 1,988 matches from six European major soccer leagues and championships1, spanning the 2014-15 to 2023-24 seasons. For each match, we acquire textual commentaries with second-level timestamps from a sports text live website2, with part of them annotated with specific event types such as corner and goal. Additionally, we also incorporate extensive metadata, including detailed background information about the games, players, coaches, referees, and teams, providing a solid foundation for future soccer understanding research.", + "bbox": [ + 511, + 547, + 903, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We partition the SoccerReplay-1988 dataset into train, validation, and test sets, containing 1,488, 250, and 250 full-match videos with diverse and comprehensive annotations, respectively. These sets provide rich training data for downstream tasks, such as event classification and commentary generation, while establishing comprehensive and challenging benchmarks for soccer understanding, as further discussed in subsequent sections.", + "bbox": [ + 511, + 729, + 903, + 849 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1Premier (England), Laliga (Spain), Bundesliga (Germany), Serie-a (Italy), League-1 (France) and UEFA Champions League.", + "bbox": [ + 511, + 862, + 903, + 887 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2flashscore.com", + "bbox": [ + 514, + 887, + 602, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8385", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ca6edb824bec182c0601875550b600151eadd1bf8ae1a66dbf292087b9c65f7b.jpg", + "image_caption": [ + "Figure 2. Automated Data Curation Pipeline. The collected soccer video data are automatically processed for temporal alignment, event summarization, and anonymization by our curation pipeline." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 477, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Automated Data Curation", + "text_level": 1, + "bbox": [ + 89, + 369, + 328, + 386 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the potential noise in raw data, such as irrelevant video content, inaccurate timestamps, and incomplete event annotations, we design an automated data curation pipeline, comprising (i) temporal alignment, (ii) event summarization, and (iii) anonymization, as illustrated in Figure 2.", + "bbox": [ + 89, + 392, + 483, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Temporal Alignment. Here, we divide match videos into two halves, each starting at kick-off, and adopt the temporal alignment model from MatchTime [40], to synchronize textual commentary timestamps with those of video frames.", + "bbox": [ + 89, + 470, + 483, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Event Summarization. For samples without event annotations, we leverage LLaMA-3-70B [12] to summarize the events based on textual commentaries. Concretely, we have expanded the event categories from 17 in SoccerNet [10] to 24 types, for finer-grained soccer understanding, for example, categorizing penalties into scored and missed, and integrating modern soccer regulations like VAR. The resulting 24 event labels include: 'corner', 'goal', 'injury', 'own goal', 'penalty', 'penalty missed', 'red card', 'second yellow card', 'substitution', 'start of game (half)', 'end of game (half)', 'yellow card', 'throw in', 'free kick', 'saved by goal-keeper', 'shot off target', 'clearance', 'lead to corner', 'off-side', 'var', 'foul (no card)', 'statistics and summary', 'ball possession', and 'ball out of play'. More details on the used prompts are provided in the Appendix.", + "bbox": [ + 89, + 534, + 483, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Anonymization. Similar to [34], we extract all person and team entity names from the metadata of SoccerReplay-1988, and replace them in textual commentaries with placeholders, such as \"[PLAYER]\", \"[TEAM]\", \"[COACH]\", and \"[REFEREE]\", ensuring consistency across tasks.", + "bbox": [ + 89, + 763, + 482, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, our data curation pipeline can seamlessly extend to existing datasets, converting the SoccerNet series [10, 34] into our unified data format, termed SoccerNetpro. This expansion further enlarges the standardized", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/a4ff8d585d0999336f8d7af8e33491bd175228e27dd5d83ccbdf23ffa33fcb7b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Existing Datasets
# GameDuration(h)# Event# Anno.# Com.
SoccerNet-v1 [15]50076476.7k-
SoccerNet-v2 [10]50076417110k-
MatchTime [40]4717161414k37k
GOAL [36]2025.5--8.9k
Our Curated Datasets
# GameDuration(h)# Event# Anno.# Com.
SoccerNet-pro50076424102k37k
SoccerReplay-19881,9883,32324150k150k
Integrated2,4884,08724252k187k
", + "bbox": [ + 517, + 88, + 905, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Statistics of Soccer Datasets. Our SoccerReplay-1988 significantly surpasses existing datasets in both scale and diversity. Here, # Anno. and # Com. refer to the number of event annotations and textual commentaries, respectively.", + "bbox": [ + 511, + 270, + 906, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "datasets available for soccer understanding tasks.", + "bbox": [ + 511, + 340, + 836, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Statistics & Discussion", + "text_level": 1, + "bbox": [ + 511, + 369, + 723, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dataset Statistics. As shown in Table 1, our dataset encompasses 3,323 hours of footage from 1,988 soccer matches, with an average duration of 100.3 minutes per match. The videos range in resolution from 360p to 720p and frame rates between 25 and 30 FPS.", + "bbox": [ + 511, + 393, + 905, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For textual annotations, this dataset features approximately 150K commentaries, averaging 76 per match, precisely temporal-aligned by the robust alignment model from MatchTime [40]. These commentaries cover 4,467 unique words, significantly surpassing the 2,873 words in existing datasets [34, 40], greatly enriching textual diversity. Automated event summarization based on these commentaries has yielded about 150K event annotations. Notably, a random sampling of $2\\%$ of the data yields $98\\%$ manual verification accuracy, ensuring high-quality automated labeling.", + "bbox": [ + 511, + 469, + 905, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SoccerReplay-test Benchmark. To facilitate a more comprehensive evaluation of soccer understanding models, we integrate 250 matches from SoccerReplay-1988 with 50 matches from the curated SoccerNet-pro, establishing SoccerReplay-test, a more challenging benchmark for event classification and commentary generation. This benchmark features nearly four times larger than existing datasets and comprises finer-grained event labels, richer textual commentaries, and up-to-date soccer regulations.", + "bbox": [ + 511, + 625, + 905, + 761 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Discussion. To summarize, SoccerReplay-1988 exhibits advancements in three aspects: (i) it is the largest soccer video dataset to date, with nearly four times more videos than existing datasets; (ii) it features more professional and diverse annotations, more suitable for fine-grained and comprehensive soccer understanding tasks; (iii) It employs an automated curation pipeline for annotations and is thus scalable to provide a solid data foundation for future research. All data from SoccerReplay-1988, including videos", + "bbox": [ + 511, + 763, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8386", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e4c6e688d3460a7c195de0a3737ceca94017c7c0480ff3b70fd2a1036e793496.jpg", + "image_caption": [ + "(a) Structure of MatchVision" + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 450, + 353 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/977119086969bdcf69088f68e5878a2f6416a7dae1682c8d3d5c485bda592645.jpg", + "image_caption": [ + "(b) MatchVision Pretraining", + "(c) Downstream Tasks", + "Figure 3. Overview of MatchVision. (a) The model architecture and its spatiotemporal feature extraction process; (b) Details of visual encoder pretraining, including supervised classification and video-language contrastive learning; (c) Implementation details of specific heads for various downstream tasks, including commentary generation, foul recognition, and event classification." + ], + "image_footnote": [], + "bbox": [ + 464, + 89, + 903, + 352 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and annotations, are open-source for non-commercial use.", + "bbox": [ + 89, + 435, + 478, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 89, + 462, + 181, + 478 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this paper, we aim to develop a soccer-specific visual encoder, MatchVision, tailored for diverse soccer video analysis tasks. We start by outlining our problem formulation in Sec. 4.1. Next, in Sec. 4.2, we detail the architecture of MatchVision. The training procedures are thoroughly discussed in Sec. 4.3. Finally, we describe the configurations for downstream tasks in Sec. 4.4, demonstrating the practical applications and effectiveness of our model.", + "bbox": [ + 89, + 488, + 483, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 617, + 294, + 631 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we tackle the challenge of analyzing soccer video segments, denoted as $\\mathcal{V} \\in \\mathbb{R}^{T \\times 3 \\times H \\times W}$ . Our goal is to utilize the visual encoder $(\\Phi_{\\mathrm{MatchVision}})$ to extract spatiotemporal features from these segments, which are then processed by multiple task-specific heads, formulated as:", + "bbox": [ + 89, + 638, + 483, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E}, \\mathbf {C}, \\mathbf {F} = \\Psi (\\Phi_ {\\text {M a t c h V i s i o n}} (\\mathcal {V}))\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 724, + 392, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\Psi = \\{\\Psi_{\\mathrm{cls}}, \\Psi_{\\mathrm{Cmt}}, \\Psi_{\\mathrm{Foul}}\\}$ represents the task-specific heads, with $\\mathbf{E}$ , $\\mathbf{C}$ , and $\\mathbf{F}$ denoting the output event types, textual commentaries, and foul types, respectively. This unified framework effectively learns relevant spatiotemporal features, and enables seamless integration across various downstream tasks for comprehensive soccer understanding.", + "bbox": [ + 89, + 750, + 482, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Architecture", + "text_level": 1, + "bbox": [ + 89, + 848, + 225, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MatchVision comprises three key components: (i) Token Embedding, (ii) Spatiotemporal Attention Block, and (iii)", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Aggregation Layer, as depicted in Figure 3.", + "bbox": [ + 511, + 435, + 802, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Token Embedding. In accordance with the convention in Vision Transformer [11], each frame $(\\mathcal{I}_i)$ from the video segment $(\\mathcal{V} = \\{\\mathcal{I}_1, \\mathcal{I}_2, \\dots, \\mathcal{I}_T\\})$ is divided into $M$ non-overlapping patches of size $P \\times P$ that span the entire frame.", + "bbox": [ + 511, + 452, + 905, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These patches are flattened into vectors $(\\mathbf{x}_i^p)$ , where $p$ and $i$ denote the spatial and temporal positions, respectively. Each vector is transformed via an embedding layer $(\\Phi_{\\mathrm{Emb}})$ into a token vector of size $\\mathbb{R}^{1\\times D}$ , and then added with a spatial position embedding $(\\mathbf{e}_{\\mathrm{s}}^{\\mathrm{pos}}\\in \\mathbb{R}^{M\\times D})$ . Subsequently, we concatenate a [c1s] token along with each frame. Finally, a temporal positional embedding $(\\mathbf{e}_{\\mathrm{t}}^{\\mathrm{pos}}\\in \\mathbb{R}^{T\\times D})$ is added across features of all frames, as formulated below:", + "bbox": [ + 511, + 513, + 906, + 633 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {y} _ {i} = \\left[ \\mathbf {x} _ {i} ^ {\\mathrm {c l s}}, \\Phi_ {\\mathrm {E m b}} \\left(\\left[ \\mathbf {x} _ {i} ^ {1}, \\dots , \\mathbf {x} _ {i} ^ {M} \\right]\\right) + \\mathbf {e} _ {\\mathrm {s}} ^ {\\mathrm {p o s}} \\right] \\\\ \\mathbf {z} = \\left[ \\mathbf {y} _ {1}, \\dots , \\mathbf {y} _ {T} \\right] + \\mathbf {e} _ {\\mathrm {t}} ^ {\\text {p o s}} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 640, + 846, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $[\\cdot ,\\cdot ]$ denotes concatenation, and $\\mathbf{y}_i\\in \\mathbb{R}^{(M + 1)\\times D}$ represents the frame-wise features. The embedded features (z) will then serve as input for spatiotemporal attention blocks. Spatiotemporal Attention Block. Similar to TimeSformer [3], we utilize interleaved temporal and spatial attention to integrate spatiotemporal information in soccer videos. Concretely, each spatiotemporal attention block comprises a temporal self-attention layer and a spatial self-attention layer, i.e., $\\phi_t(\\cdot)$ and $\\phi_s(\\cdot)$ , respectively.", + "bbox": [ + 511, + 685, + 905, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a video feature $(\\mathbf{z} \\in \\mathbb{R}^{T \\times (M + 1) \\times D})$ , we alternate temporal and spatial attention: temporal attention facilitates interactions among tokens at the same spatial positions across distinct frames, while spatial attention enables interactions among tokens within the same frame. Residual con", + "bbox": [ + 511, + 824, + 905, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8387", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "nections are employed in each layer. After passing through a total of $\\mathbf{K}$ spatiotemporal attention blocks, the resulting feature $(\\mathcal{F})$ captures both intra-frame and inter-frame relationships, i.e., $\\mathcal{F} = [\\phi_s(\\phi_t(\\mathbf{z}))]^{\\mathbf{K}} \\in \\mathbb{R}^{T \\times (M + 1) \\times D}$ .", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Aggregation Layer. To obtain video-level features, we employ an aggregation layer on the frame-wise spatiotemporal features. Specifically, for the $i$ -th frame, we utilize spatial self-attention to aggregate information into its [cls] token, denoted as $\\hat{\\mathcal{F}}_i^{\\mathrm{cls}} = \\Phi_{\\mathrm{Agg}}(\\mathcal{F}_i)$ . Concatenating the [cls] tokens of all frames yields the final video feature $(\\mathcal{F}_{\\mathcal{V}})$ , that effectively encapsulates spatiotemporal characteristics of soccer video segments, thus enabling it to be applicable for various downstream soccer understanding tasks. This process can be formulated as:", + "bbox": [ + 89, + 154, + 482, + 305 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\mathcal {V}} = \\Phi_ {\\mathrm {M a t c h V i s i o n}} (\\mathcal {V}) = [ \\hat {\\mathcal {F}} _ {1} ^ {\\mathrm {c l s}}, \\dots , \\hat {\\mathcal {F}} _ {T} ^ {\\mathrm {c l s}} ] \\in \\mathbb {R} ^ {T \\times D}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 318, + 460, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. MatchVision Pretraining", + "text_level": 1, + "bbox": [ + 89, + 349, + 318, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this part, we aim to pretrain the visual encoder with triplet samples $(\\{\\mathcal{V},\\mathbf{E},\\mathbf{C}\\})$ , comprising videos, event labels, and textual commentaries. Concretely, we investigate two distinct pretraining strategies: supervised classification and video-language contrastive learning.", + "bbox": [ + 89, + 373, + 482, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Supervised Classification. One way to pretrain the visual encoder is supervised learning on event classification. To be specific, the extracted visual features $(\\mathcal{F}_{\\mathcal{V}})$ are aggregated by a temporal self-attention layer into a learnable [cls] token, denoted as $\\mathcal{F}_{\\mathcal{V}}^{\\mathrm{cls}}$ , similar to the spatial-wise aggregation mentioned above. This token is then fed into a linear classifier, and trained with a cross-entropy loss for event classification. The objective is denoted as $\\mathcal{L}_{\\mathrm{sup}}$ .", + "bbox": [ + 89, + 452, + 482, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Video-Language Contrastive Learning. As an alternative, we can also pretrain our visual encoder with video-text contrastive learning. Specifically, we adopt simple average pooling on the video feature to get the aggregated visual feature $(\\mathcal{F}_{\\mathcal{V}}^{\\mathrm{Avg}})$ , and encode the textual commentary (C) with a text encoder $(\\Phi_{\\mathrm{Text}})$ . We train the model with sigmoid loss ( $\\mathcal{L}_{\\mathrm{sigmoid}}$ ), as used in SigLIP [58]. Note that, some video clips may have highly similar commentaries, for example, 'start of the game', we treat the commentaries with high similarity in the same batch as positive samples when calculating loss functions. This can be expressed as follows:", + "bbox": [ + 89, + 575, + 482, + 743 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n t r a}} = \\mathcal {L} _ {\\text {s i g m o i d}} \\left(\\mathcal {F} _ {\\mathcal {V}} ^ {\\text {A v g}}, \\Phi_ {\\text {T e x t}} (\\mathbf {C})\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 753, + 413, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Downstream Tasks", + "text_level": 1, + "bbox": [ + 89, + 787, + 272, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After the pretraining mentioned above, MatchVision can now serve as a versatile visual encoder, to map the soccer video segments into visual features $(\\mathcal{F}_{\\mathcal{V}})$ , for training task-specific heads $\\Psi = \\{\\Psi_{\\mathrm{cls}}, \\Psi_{\\mathrm{Cmt}}, \\Psi_{\\mathrm{Foul}}\\}$ across different downstream tasks, including: (i) event classification, (ii) commentary generation, and (iii) foul recognition.", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Event Classification. Similar to supervised classification above, we concatenate a learnable [cls] token to aggregate frame-wise visual features via temporal self-attention. This token is then fed into a linear classifier for event classification. The event classification head $(\\Psi_{\\mathrm{cls}})$ is trained with a cross-entropy loss while freezing the visual encoder.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Commentary Generation. We follow the paradigm in MatchTime [40] to generate anonymized textual commentary for soccer video clips. Concretely, the commentary generation head $(\\Psi_{\\mathrm{Cmt}})$ employs a Perceiver [25] aggregator to consolidate visual features, which are then projected by a trainable MLP, serving as prefix embeddings for a large language model (LLM). Subsequently, an off-the-shelf LLM decodes these embeddings into textual commentary. We adopt the negative log-likelihood loss, commonly used for auto-regressive next-token prediction.", + "bbox": [ + 511, + 185, + 906, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Foul Recognition. As outlined in [22], the foul recognition task takes multi-view videos from the same scene as inputs, with each sample annotated with a foul class (8 types) and severity (4 levels). We encode these multi-view videos with MatchVision, and aggregate the extracted features into a single feature vector, via either max or average pooling, following the common practice. Subsequently, the foul recognition head $(\\Psi_{\\mathrm{Foul}})$ employs a shared MLP and two task-specific linear classifiers, to predict foul type and severity, respectively. Similar to event classification, we use the combination of cross-entropy losses on the foul type and severity classification to jointly train $\\Psi_{\\mathrm{Foul}}$ .", + "bbox": [ + 511, + 339, + 905, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discussion. Pretraining MatchVision on large-scale soccer data equips it with substantial domain-specific knowledge, enabling it to serve as a universal visual encoder adaptable to various downstream soccer understanding tasks.", + "bbox": [ + 511, + 523, + 905, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 599, + 645, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section begins with implementation details in Sec. 5.1; followed by quantitative evaluations across downstream tasks in Sec. 5.2; then, we conduct ablation studies on our SoccerReplay-test benchmark to analyze the effectiveness of the proposed dataset and model in Sec. 5.3; finally, we provide qualitative results for comparison in Sec. 5.4.", + "bbox": [ + 511, + 625, + 905, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 727, + 730, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our experiments, video segments are sampled at 1FPS around annotated timestamps, capturing a 30-second window for each sample. Frames are resized to $224 \\times 224$ pixels as inputs. We initialize the embedding layer, spatial attention layers, aggregation layer, and text encoder of MatchVision with pretrained weights from SigLIP Base-16 [58] and adopt LLaMA-3 (8B) [12] as the off-the-shelf LLM decoder for commentary generation. All experiments are conducted on $4 \\times$ Nvidia H800 GPUs with the AdamW [33] optimizer. Next, we elaborate on the training and evaluation details", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8388", + "bbox": [ + 480, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d502cee976c4dea31f3fc7e5a902e3e873bd8fdba7e42283767eafe88893f1b7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Visual EncoderDatasetClassification (%)Commentary
SNMTSRAcc.@1Acc.@3Acc.@5B@1B@4MR-LC
Off-the-shelf Models
I3D [4]XXX45.482.593.226.775.5724.1723.1218.73
C3D [45]XXX47.885.195.028.136.6424.5224.2327.88
ResNet [21]XXX47.284.694.427.346.5724.7224.4327.29
CLIP [38]XXX48.585.595.226.256.5124.2724.7528.17
InternVideo [48]XXX49.987.095.927.126.5425.0224.8229.90
SigLIP [58]XXX50.286.795.627.856.9825.1625.0331.38
Pretrain with Supervised Classification
Baidu [61]✓XX56.491.997.331.208.8826.5626.6138.93
SigLIP✓XX55.989.694.928.517.3925.9625.9435.71
SigLIP✓✓✓57.991.797.530.958.5625.7926.1738.24
MatchVision✓XX82.596.698.829.457.9226.0126.2136.15
MatchVision✓✓✓84.097.399.231.059.0626.9427.9342.20
Pretrain with Visual-Language Contrastive Learning
SigLIPX✓X55.488.897.028.727.7225.9126.1732.27
SigLIPX✓✓66.893.798.630.358.1226.0526.3839.41
MatchVisionX✓X58.989.097.130.337.9725.4826.3333.87
MatchVisionX✓✓67.993.998.631.949.1226.2427.5640.76
Pretrain with Hybrid Supervised-Contrastive Training
SigLIP✓✓X71.294.598.728.637.8225.7425.3534.09
SigLIP✓✓✓67.193.298.130.718.7826.2626.7441.82
MatchVision✓✓X76.496.099.030.658.3325.2826.3137.23
MatchVision✓✓✓80.197.199.133.589.1426.8228.2144.18
", + "bbox": [ + 142, + 88, + 852, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative Results on Event Classification and Commentary Generation. Here, SN, MT, and SR represent finetuning with curated SoccerNet-v2 [10], MatchTime [40], and SoccerReplay-1988, respectively. B, M, R-L, and C refer to BLEU, METEOR, ROUGE-L, and CIDEr metrics, respectively. Within each unit, we denote the best performance in RED and the second-best performance in BLUE.", + "bbox": [ + 89, + 508, + 906, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "about visual encoder pretraining and downstream tasks.", + "bbox": [ + 89, + 561, + 457, + 577 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visual Encoder Pretraining. For both pretraining strategies, we use a batch size of 40 for 15 epochs. The learning rate for all randomly initialized modules, including the temporal attention layers, aggregator layer, and linear classifier, is set to $1 \\times 10^{-4}$ . Meanwhile, the learning rate for modules initialized with pretrained parameters (including the text encoder) is set to $5 \\times 10^{-5}$ . In contrastive training, we adopt a multi-positive strategy where each textual commentary, based on its event label, considers closely related categories (e.g. \"start of game\" and \"offside\") as positive samples.", + "bbox": [ + 89, + 580, + 482, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Downstream Tasks. In all downstream tasks, unless otherwise specified, we use the frozen visual encoder for feature extraction and only train the task-specific heads with a learning rate of $1 \\times 10^{-4}$ for 30 epochs. The batch sizes for event classification, commentary generation, and foul recognition are set to 40, 32, and 8, respectively. We adopt specific evaluation metrics for these three tasks: (i) For event classification, we use the top-1/3/5 classification accuracy; (ii) For commentary generation, we employ several commonly-used language evaluation metrics, including BLEU [35], METEOR [2], ROUGE-L [32], and", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CIDEr [47]; (iii) For foul recognition, we follow the common practice, and report top-1/2 and top-1 accuracy for the foul type and severity classification, respectively.", + "bbox": [ + 511, + 561, + 906, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benchmarks & Baselines. To ensure fair and reliable comparisons with existing work, we evaluate event classification (24 types) on 100 matches from curated SoccerNetv2 [10] test set; commentary generation on 49 matches from SN-Caption-test-align benchmark manually aligned in [40]; and foul recognition on MVFoul [22]. We consider various baselines: for the first two tasks, this includes off-the-shelf general visual encoders such as ResNet [21], C3D [45], I3D [4], CLIP [38], SigLIP [58], and InternVideo [48], along with Baidu [61] and SigLIP finetuned with soccer-specific data. For foul recognition, we follow previous work [22, 23] and adopt ResNet [21], R(2+1)D [46], and MViT [13] jointly finetuned with classifiers, as baselines.", + "bbox": [ + 511, + 609, + 908, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2.Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 511, + 816, + 735, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As depicted in Table 2, we draw two observations on event classification and commentary generation: (i) visual encoders trained on soccer data substantially outperform off-the-shelf general encoders (ResNet, C3D, I3D, CLIP, and", + "bbox": [ + 511, + 840, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8389", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1e7396307c4fef1618267f34b5fe5bb12fa48b57efeace8acc9710a23730fc40.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Visual EncoderFoul ClassSeverity
BackboneTrainAgg.Acc.@1Acc.@2Acc.@1
ResNet [21]✓Mean0.310.560.34
Max0.320.600.32
R(2+1)D [46]✓Mean0.310.550.34
Max0.320.560.39
MViT [13]✓Mean0.400.650.38
Max0.470.690.43
MatchVision×Mean0.440.530.58
Max0.350.700.46
", + "bbox": [ + 96, + 88, + 475, + 267 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "InternVideo), underscoring the necessity of building specialized models for soccer understanding; (ii) almost all visual encoders, across all training settings, benefit from SoccerReplay-1988, emphasizing the value of constructing large-scale, high-quality data for soccer understanding. Next, we will delve into each task to discuss the results.", + "bbox": [ + 89, + 333, + 482, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Event Classification. With identical training strategies and data, MatchVision considerably outperforms other methods in classification accuracy, demonstrating the superiority of its architecture, which effectively leverages spatiotemporal features within soccer videos. For example, MatchVision achieves a Top-1 accuracy of $82.5\\%$ , significantly surpassing SigLIP's $55.9\\%$ under the same training conditions. Moreover, models trained via supervised classification excel others, primarily because the pre-training task shares the same objectives as the downstream event classification task.", + "bbox": [ + 89, + 426, + 482, + 578 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Commentary Generation. Visual encoders trained with visual-language contrastive learning exhibit better commentary generation performance than those trained with supervised classification, as this strategy better captures correlations between visual and textual features. Additionally, while MatchVision trained solely on SoccerNet slightly underperforms Baidu [61], incorporating SoccerReplay-1988 enables it to outperform on most metrics. This demonstrates that MatchVision can take advantage of large-scale datasets. Finally, a hybrid training approach, starting with supervised classification followed by visual-language contrastive learning, enables MatchVision to achieve optimal performance. This indicates that learning coarse-grained tasks such as classification provides a foundation for fine-grained tasks like commentary generation, and fully leveraging data unlocks the potential of soccer understanding.", + "bbox": [ + 89, + 580, + 482, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Foul Recognition. As demonstrated in Table 3, MatchVision achieves performance comparable to jointly finetuned state-of-the-art methods in foul recognition, even with a frozen visual encoder. This highlights that MatchVision effectively learns substantial knowledge from large-scale soc", + "bbox": [ + 89, + 824, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/21608bf50b0ecb1bd5348751dc6d315743bf72ac03842ae185193d4c36abf0d4.jpg", + "table_caption": [ + "Table 3. Quantitative Results on Multi-view Foul Recognition. Our frozen MatchVision encoder can achieve comparable performance with other jointly finetuned visual encoders." + ], + "table_footnote": [], + "table_body": "
PretrainClassification(%)
Sup.Contra.SRAcc.@1Acc.@3Acc.@5
✓XX62.6783.0089.81
✓X✓68.0386.9092.38
X✓X46.9775.5385.85
X✓✓57.4183.1391.00
✓✓X56.8680.3088.09
✓✓✓63.5985.2191.63
", + "bbox": [ + 547, + 88, + 870, + 234 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablations on Event Classification. We explore the impact of various training settings of our MatchVision encoder on the SoccerReplay-test benchmark. Here, Sup., Contra., and SR refer to supervised classification, visual-language contrastive learning, and the SoccerReplay-1988 dataset, respectively.", + "bbox": [ + 511, + 239, + 906, + 311 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cer data and adapts seamlessly to downstream tasks. Comparisons with additional baselines from the SoccerNet foul recognition challenges [9] are provided in the Appendix.", + "bbox": [ + 511, + 328, + 906, + 375 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 386, + 679, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation experiments on event classification and commentary generation using our SoccerReplay-test benchmark. These experiments validate the effectiveness of our proposed dataset and model, while establishing a baseline for future evaluations on this benchmark.", + "bbox": [ + 511, + 409, + 906, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Event Classification. We evaluate event classification on 300 matches from our SoccerReplay-test benchmark using the MatchVision visual encoder pretrained with various strategies. Features are extracted by MatchVision and processed with a learnable aggregation layer and a linear classifier. The default training set is our curated SoccerNet-pro. As shown in Table 4, integrating SoccerReplay-1988 for training results in significant performance improvements across all pretraining strategies, yielding the significance of our dataset. Additionally, supervised classification outperforms visual-language contrastive learning and hybrid pretraining. This is due to its closer alignment with downstream event classification task, and the scale of event annotations is far larger than that of textual commentaries, further confirming the substantial benefits of data scaling for boosting soccer understanding.", + "bbox": [ + 511, + 488, + 908, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Commentary Generation. With the pretrained MatchVision encoder, we train the commentary generation head on the MatchTime [40] and SoccerReplay-1988 datasets using various training strategies. By default, only the Perceiver [25] aggregation layer and projection layer within the head are trained. For joint training with the LLM decoder, considering computational costs, we incorporate LoRA [24] layers while freezing the original LLM layers. As shown in Table 5, incorporating SoccerReplay-1988 significantly improves performance on all metrics, confirming substantial advantages of our proposed dataset. This performance", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8390", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9d80294ea391f74578532a7a932bfd89ac9a197638371fc49f6dce60d8272ed9.jpg", + "image_caption": [ + "Soccer Videos" + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 119, + 104 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9da4b340fa5e701f874da4eab9b7b914104a894fefb77e3bfec73dbc28c6aaee.jpg", + "image_caption": [ + "Events" + ], + "image_footnote": [], + "bbox": [ + 374, + 92, + 392, + 104 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dfe11a125f3ab53dc3d4ff9c58672ea87bd1c11d4c1cb608159d5405b50da324.jpg", + "image_caption": [ + "Commentaries" + ], + "image_footnote": [], + "bbox": [ + 521, + 90, + 537, + 104 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8cdd4dbbecb8a9ec6e3fdaac06188a5bdb8d9593f88138abc2c3f127dc7a9875.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 107, + 367, + 150 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/68732df897e97929bec2c06a6cd4a0e00c3c827ef4c30cf5c5ca0c4f68c81307.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 118, + 459, + 143 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fc9c2bb97514f72b61de963098cb78a266efe3146a378467055940157f707b37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 526, + 114, + 888, + 147 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/52ed70c15dc4709676df0e4b456ec14925d160a95ba8aee8054e3fadab030edd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 154, + 367, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8276801c2dc956cf715cbb6e35392feb9ec6c5c1e35246581e3815e139c44262.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 162, + 514, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/516894ea4acf31fed702636a2327972b70c6c0686e548ec9008c7942bb866031.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 155, + 890, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/270fd650638147ed4738f17a034c436bce1a5e856c8100fe1265415bc8ef1e7e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 198, + 367, + 241 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b0de7aef3c8941434bfbb3eea98b327a676dacfc6f7f83d3791ac06e9b00f0f7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 209, + 488, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a380b47f26580073f501fca6a80af2d27b0dff7e9f36d75a2471d72d0c93d535.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 200, + 890, + 241 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3eefcf41d5f78e47d7d847d40c2c7378f9003d67dcc7f9abc8a12c41c3846a2e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 243, + 367, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3b4d60d20742533f20aab26bf153e9159875fb7e3fc21dc2d1ab6a2e9a7e5be3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 253, + 477, + 281 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6bdbf1b0e4e1e262d7f85f385ddf1be5a1f0c0e018b967e3eade70b46797fd21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 247, + 888, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c4b3b98a4be399f96a73b58101fd17cc5b827266907bce570f39aa35c3e2d7ff.jpg", + "image_caption": [ + "Figure 4. Qualitative Results for Event Classification and Commentary Generation. Here, \"w/o SR\" and \"w/ SR\" indicate models trained without and with the SoccerReplay-1988 dataset, respectively. Incorporating SoccerReplay-1988 improves event classification accuracy. Moreover, this enriched training data enables the model to demonstrate several advantages in commentary generation: (a) more detailed descriptions, (b) greater linguistic variety, (c) higher event depiction accuracy, (d) better adherence to updated rules, and (e) improved specificity in scenario response." + ], + "image_footnote": [], + "bbox": [ + 96, + 290, + 367, + 333 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9b8addab99c97babf3f33b9d8f1d4cb75828a7be35583514c63ea2e65e97bcee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 300, + 460, + 327 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/39aeeb1cf36f84a1d0fe1d9adf45176e87fd7b4d6118c5b90ca344994a7a1b2b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 297, + 879, + 330 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e1760a007d98c191a6ad49503d990e87dd10e26631844807f40b023b88dfe480.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TrainableCommentary Metrics
VLB@1B@4MR-LC
Trained on MatchTime
XX21.653.2721.0217.7912.90
✓X27.627.0224.0323.5130.77
X✓27.046.4124.1523.8831.91
✓✓27.496.9624.5023.3330.81
Trained on MatchTime & SoccerReplay-1988
XX24.174.0920.5120.7015.70
✓X28.988.3924.4525.3545.85
X✓27.547.7624.5024.7042.79
✓✓29.218.2225.2525.5443.18
", + "bbox": [ + 112, + 435, + 457, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablations on Commentary Generation. We investigate the impact of different training strategies and datasets on MatchVision using the SoccerReplay-test benchmark. 'V' and 'L' denote the visual encoder and the LLM decoder, respectively.", + "bbox": [ + 89, + 650, + 482, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "gap also reflects the challenges of our established benchmark, which features diverse vocabulary, richer semantics, and updated soccer rules. Additionally, jointly finetuning the visual encoder and the LLM decoder provides a feasible approach for further improvements.", + "bbox": [ + 89, + 717, + 483, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Qualitative Comparisons", + "text_level": 1, + "bbox": [ + 89, + 801, + 318, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As depicted in Figure 4, we present qualitative results of MatchVision on the SoccerReplay-test benchmark, comparing models pretrained with and without SoccerReplay-1988. For event classification, incorporating our data improves accuracy, and even in misclassified cases, the results", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "remain contextually relevant. For commentary generation, hybrid training on SoccerReplay-1988 enables MatchVision to produce richer, more detailed textual commentary, reflecting a deeper understanding of soccer dynamics. More qualitative results are available in the Appendix.", + "bbox": [ + 511, + 438, + 906, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 541, + 633, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we establish a unified, scalable multi-modal framework for soccer understanding. Specifically, we introduce SoccerReplay-1988, the largest and most comprehensive soccer video dataset to date, annotated by an automated curation pipeline. This provides a solid foundation for developing soccer understanding models and serves as a more challenging benchmark. Built upon this, we develop MatchVision, an advanced soccer-specific visual encoder, which effectively leverages spatiotemporal information within soccer videos and can be applied to various tasks such as event classification and commentary generation. Extensive experiments demonstrate the superiority of our model, with MatchVision achieving state-of-the-art performance on both existing benchmarks and our newly established one. We believe this work will set a viable, universal paradigm for future research in sports understanding.", + "bbox": [ + 511, + 571, + 906, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 513, + 840, + 671, + 857 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is funded by National Key R&D Program of China (No.2022ZD0161400).", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8391", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, pages 23716-23736, 2022. 2", + "[2] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, pages 65-72, 2005. 6", + "[3] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In Proceedings of the International Conference on Machine Learning, page 4, 2021. 2, 4", + "[4] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 6", + "[5] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, and Sergey Tulyakov. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2", + "[6] Anthony Cioppa, Adrien Deliege, Floriane Magera, Silvio Giancola, Olivier Barnich, Bernard Ghanem, and Marc Van Droogenbroeck. Camera calibration and player localization in soccernet-v2 and investigation of their representations for action spotting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 4537-4546, 2021. 2", + "[7] Anthony Cioppa, Adrien Deliege, Silvio Giancola, Bernard Ghanem, and Marc Van Droogenbroeck. Scaling up soccer-net with multi-view spatial localization and re-identification. Scientific data, 9(1):355, 2022. 1, 2", + "[8] Anthony Cioppa, Silvio Giancola, Adrien Deliege, Le Kang, Xin Zhou, Zhiyu Cheng, Bernard Ghanem, and Marc Van Droogenbroeck. Soccernet-tracking: Multiple object tracking dataset and benchmark in soccer videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 3491-3502, 2022. 2", + "[9] Anthony Cioppa, Silvio Giancola, Vladimir Somers, Victor Joos, Floriane Magera, Jan Held, Seyed Abolfazl Ghasemzadeh, Xin Zhou, Karolina Seweryn, Mateusz Kowalczyk, Zuzanna Mroz, Szymon Lukasik, Michal Halon, Hassan Mkhallati, Adrien Deliege, Carlos Hinojosa, Karen Sanchez, Amir M. Mansourian, Pierre Miralles, Olivier Barnich, Christophe De Vleeschouwer, Alexandre Alahi, Bernard Ghanem, Marc Van Droogenbroeck, Adam Gorski, Albert Clapes, Andrei Boiarov, Anton Afanasiev, Artur Xarles, Atom Scott, ByoungKwon Lim, Calvin Yeung, Cristian Gonzalez, Dominic Rufenacht, Enzo Pacilio, Fabian Deuser, Faisal Sami Altawijri, Francisco Cachon, HanKyul Kim, Haobo Wang, Hyeonmin Choe, Hyunwoo J Kim, Il-Min" + ], + "bbox": [ + 99, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Kim, Jae-Mo Kang, Jamshid Tursunboev, Jian Yang, Jihwan Hong, Jimin Lee, Jing Zhang, Junseok Lee, Kexin Zhang, Konrad Habel, Licheng Jiao, Linyi Li, Marc Gutierrez-Perez, Marcelo Ortega, Menglong Li, Milosz Lopatto, Nikita Kasatkin, Nikolay Nemtsev, Norbert Oswald, Oleg Udin, Pavel Kononov, Pei Geng, Saad Ghazai Alotaibi, Sehyung Kim, Sergei Ulasen, Sergio Escalera, Shanshan Zhang, Shuyuan Yang, Sunghwan Moon, Thomas B. Moeslund, Vasyl Shandyba, Vladimir Golovkin, Wei Dai, WonTaek Chung, Xinyu Liu, Yongqiang Zhu, Youngseo Kim, Yuan Li, Yuting Yang, Yuxuan Xiao, Zehua Cheng, and Zhihao Li. Soccernet 2024 challenges results. arXiv preprint arXiv:2409.10587, 2024. 2, 7", + "[10] Adrien Deliege, Anthony Cioppa, Silvio Giancola, Meisam J Seikavandi, Jacob V Dueholm, Kamal Nasrollahi, Bernard Ghanem, Thomas B Moeslund, and Marc Van Droogenbroeck. Soccernet-v2: A dataset and benchmarks for holistic understanding of broadcast soccer videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 4508-4519, 2021. 1, 2, 3, 6", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations, 2021. 4", + "[12] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 3, 5", + "[13] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6824-6835, 2021. 6, 7", + "[14] Sushant Gautam, Mehdi Houshmand Sarkhoosh, Jan Held, Cise Midoglu, Anthony Cioppa, Silvio Giancola, Vajira Thambawita, Michael A Riegler, Pål Halvorsen, and Mubarak Shah. Soccernet-echoes: A soccer game audio commentary dataset. arXiv preprint arXiv:2405.07354, 2024. 2", + "[15] Silvio Giancola, Mohieddine Amine, Tarek Dghaily, and Bernard Ghanem. Soccernet: A scalable dataset for action spotting in soccer videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 1711-1721, 2018. 1, 2, 3", + "[16] Xiaofan Gu, Xinwei Xue, and Feng Wang. Fine-grained action recognition on a novel basketball dataset. In International Conference on Acoustics, Speech, and Signal Processing, pages 2563-2567, 2020. 2", + "[17] Tengda Han, Weidi Xie, and Andrew Zisserman. Temporal alignment networks for long-term video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2906-2916, 2022. 2", + "[18] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. Autoad: Movie description in" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8392", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2", + "[19] Tengda Han, Max Bain, Arsha Nagrani, Gul Varol, Weidi Xie, and Andrew Zisserman. Autoad ii: The sequel-who, when, and what in movie audio description. In Proceedings of the International Conference on Computer Vision, 2023.", + "[20] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. Autoad iii: The prequel - back to the pixels. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2", + "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6, 7", + "[22] Jan Held, Anthony Cioppa, Silvio Giancola, Abdullah Hamdi, Bernard Ghanem, and Marc Van Droogenbroeck. Vars: Video assistant referee system for automated soccer decision making from multiple views. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 5085-5096, 2023. 2, 5, 6", + "[23] Jan Held, Hani Itani, Anthony Cioppa, Silvio Giancola, Bernard Ghanem, and Marc Van Droogenbroeck. X-vars: Introducing explainability in football refereeing with multimodal large language models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 3267-3279, 2024. 2, 6", + "[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In Proceedings of the International Conference on Learning Representations, 2022. 7", + "[25] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In Proceedings of the International Conference on Machine Learning, pages 4651-4664, 2021. 5, 7", + "[26] Haopeng Li, Andong Deng, Qiuhong Ke, Jun Liu, Hossein Rahmani, Yulan Guo, Bernt Schiele, and Chen Chen. Sports-qa: A large-scale video question answering benchmark for complex and professional sports. arXiv preprint arXiv:2401.01505, 2024. 2", + "[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the International Conference on Machine Learning, pages 12888-12900, 2022. 2", + "[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the International Conference on Machine Learning, pages 19730–19742, 2023. 2", + "[29] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, Limin Wang, and Yu Qiao. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 2" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In Proceedings of the European Conference on Computer Vision, 2024. 2", + "[31] Zeqian Li, Qirui Chen, Tengda Han, Ya Zhang, Yanfeng Wang, and Weidi Xie. Multi-sentence grounding for long-term instructional video. In Proceedings of the European Conference on Computer Vision, 2024. 2", + "[32] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, 2004. 6", + "[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Proceedings of the International Conference on Learning Representations, 2019. 5", + "[34] Hassan Mkhallati, Anthony Cioppa, Silvio Giancola, Bernard Ghanem, and Marc Van Droogenbroeck. Socernetc caption: Dense video captioning for soccer broadcasts commentaries. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 5074-5085, 2023. 1, 2, 3", + "[35] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Association for Computational Linguistics, pages 311-318, 2002. 6", + "[36] Ji Qi, Jifan Yu, Teng Tu, Kunyu Gao, Yifan Xu, Xinyu Guan, Xiaozhi Wang, Bin Xu, Lei Hou, Juanzi Li, et al. Goal: A challenging knowledge-grounded video captioning benchmark for real-time soccer commentary generation. In Proceedings of the ACM International Conference on Information and Knowledge Management, pages 5391-5395, 2023. 1, 2, 3", + "[37] Mengshi Qi, Yunhong Wang, Annan Li, and Jiebo Luo. Sports video captioning via attentive motion representation and group relationship modeling. IEEE Transactions on Circuits and Systems for Video Technology, 2019. 2", + "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, 2021. 2, 6", + "[39] Prem N Ramkumar, Bryan C Luu, Heather S Haeberle, Jaret M Karnuta, Benedict U Nwachukwu, and Riley J Williams. Sports medicine and artificial intelligence: a primer. The American Journal of Sports Medicine, 50(4): 1166-1174, 2022. 2", + "[40] Jiayuan Rao, Haoning Wu, Chang Liu, Yanfeng Wang, and Weidi Xie. Matchtime: Towards automatic soccer game commentary generation. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2024. 1, 2, 3, 5, 6, 7", + "[41] Dian Shao, Yue Zhao, Bo Dai, and Dahua Lin. Finegym: A hierarchical video dataset for fine-grained action understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2616-2625, 2020. 2", + "[42] Yan Shu, Zheng Liu, Peitian Zhang, Minghao Qin, Junjie Zhou, Zhengyang Liang, Tiejun Huang, and Bo Zhao." + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8393", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Video-xl: Extra-long vision language model for hour-scale video understanding. arXiv preprint arXiv:2409.14485, 2024. 2", + "[43] Vladimir Somers, Victor Joos, Anthony Cioppa, Silvio Giancola, Seyed Abolfazl Ghasemzadeh, Floriane Magera, Baptiste Standaert, Amir M Mansourian, Xin Zhou, Shohreh Kasaei, et al. Soccernet game state reconstruction: End-to-end athlete tracking and identification on a minimap. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 3293-3305, 2024. 2", + "[44] Graham Thomas, Rikke Gade, Thomas B Moeslund, Peter Carr, and Adrian Hilton. Computer vision for sports: Current applications and research topics. Computer Vision and Image Understanding, 159:3-18, 2017. 2", + "[45] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the International Conference on Computer Vision, pages 4489-4497, 2015. 6", + "[46] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 6, 7", + "[47] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4566-4575, 2015. 6", + "[48] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 6", + "[49] Zhe Wang, Petar Velicković, Daniel Hennes, Nenad Tomasev, Laurel Prince, Michael Kaisers, Yoram Bachrach, Romuald Elie, Li Kevin Wenliang, Federico Piccinini, et al. Tacticai: an ai assistant for football tactics. Nature Communications, 15(1):1-13, 2024. 1, 2", + "[50] Dekun Wu, He Zhao, Xingce Bao, and Richard P Wildes. Sports video analysis on large-scale data. In Proceedings of the European Conference on Computer Vision, 2022. 2", + "[51] Zeyu Xi, Ge Shi, Xuefen Li, Junchi Yan, Zun Li, Lifang Wu, Zilin Liu, and Liang Wang. A simple yet effective knowledge guided method for entity-aware video captioning on a basketball benchmark. Neurocomputing, 2025. 2", + "[52] Zeyu Xi, Ge Shi, Haoying Sun, Bowen Zhang, Shuyi Li, and Lifang Wu. Eika: Explicit & implicit knowledge-augmented network for entity-aware sports video captioning. Expert Systems with Applications, 2025. 2", + "[53] Haotian Xia, Zhengbang Yang, Yuqing Wang, Rhys Tracy, Yun Zhao, Dongdong Huang, Zezhi Chen, Yan Zhu, Yuanfang Wang, and Weining Shen. Sportqa: A benchmark for sports understanding in large language models. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics, 2024. 2", + "[54] Haotian Xia, Zhengbang Yang, Junbo Zou, Rhys Tracy, Yuqing Wang, Chi Lu, Christopher Lai, Yanjun He, Xun" + ], + "bbox": [ + 91, + 90, + 482, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Shao, Zhuoqing Xie, et al. Sportu: A comprehensive sports understanding benchmark for multimodal large language models. In Proceedings of the International Conference on Learning Representations, 2025. 2", + "[55] Jinglin Xu, Yongming Rao, Xumin Yu, Guangyi Chen, Jie Zhou, and Jiwen Lu. Finediving: A fine-grained dataset for procedure-aware action quality assessment. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2949-2958, 2022. 2", + "[56] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 2", + "[57] Huanyu Yu, Shuo Cheng, Bingbing Ni, Minsi Wang, Jian Zhang, and Xiaokang Yang. Fine-grained video captioning for sports narrative. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2", + "[58] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the International Conference on Computer Vision, pages 11975-11986, 2023. 2, 5, 6", + "[59] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2023. 2", + "[60] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024. 2", + "[61] Xin Zhou, Le Kang, Zhiyu Cheng, Bo He, and Jingyu Xin. Feature combination meets attention: Baidu soccer embeddings and transformer based temporal detection. arXiv preprint arXiv:2106.14447, 2021. 2, 6, 7", + "[62] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 669 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8394", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Towards Universal Soccer Video Understanding/bd5f526c-bc30-45cc-bbcd-e151dfc21f1a_model.json b/2025/Towards Universal Soccer Video Understanding/bd5f526c-bc30-45cc-bbcd-e151dfc21f1a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d49c3327950fb2b3d0ef25fca357152c6c2c0252 --- /dev/null +++ b/2025/Towards Universal Soccer Video Understanding/bd5f526c-bc30-45cc-bbcd-e151dfc21f1a_model.json @@ -0,0 +1,2554 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.258, + 0.131, + 0.743, + 0.154 + ], + "angle": 0, + "content": "Towards Universal Soccer Video Understanding" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.18, + 0.844, + 0.263 + ], + "angle": 0, + "content": "Jiayuan Rao\\(^{1,2*}\\), Haoning Wu\\(^{1,2*}\\), Hao Jiang\\(^{3}\\), Ya Zhang\\(^{1}\\), Yanfeng Wang\\(^{1\\dagger}\\), Weidi Xie\\(^{1\\dagger}\\) \n\\(^{1}\\)School of Artificial Intelligence, Shanghai Jiao Tong University, China \n\\(^{2}\\)CMIC, Shanghai Jiao Tong University, China \nhttps://jyrao.github.io/UniSoccer/" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.312, + 0.327, + 0.329 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.343, + 0.485, + 0.631 + ], + "angle": 0, + "content": "As a globally celebrated sport, soccer has attracted widespread interest from fans all over the world. This paper aims to develop a comprehensive multi-modal framework for soccer video understanding. Specifically, we make the following contributions in this paper: (i) we introduce SoccerReplay-1988, the largest multi-modal soccer dataset to date, featuring videos and detailed annotations from 1,988 complete matches, with an automated annotation pipeline; (ii) we present an advanced soccer-specific visual encoder, MatchVision, which leverages spatiotemporal information across soccer videos and excels in various downstream tasks; (iii) we conduct extensive experiments and ablation studies on event classification, commentary generation, and multi-view foul recognition. MatchVision demonstrates state-of-the-art performance on all of them, substantially outperforming existing models, which highlights the superiority of our proposed data and model. We believe that this work will offer a standard paradigm for sports understanding research." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.638, + 0.483, + 0.668 + ], + "angle": 0, + "content": "\"Football is one of the world's best means of communication. It is impartial, apolitical, and universal.\"" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.669, + 0.483, + 0.684 + ], + "angle": 0, + "content": "— Franz Beckenbauer (1945 - 2024)" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.705, + 0.222, + 0.721 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.73, + 0.484, + 0.866 + ], + "angle": 0, + "content": "Soccer, celebrated worldwide for its significant commercial value, has recently seen great research interest in integrating artificial intelligence (AI) for soccer video understanding. This is primarily motivated by the sport's complexity and the growing demand for enhanced analytics and improved viewing experiences. AI systems facilitate tactical analysis [49], allowing coaches to devise better strategies by uncovering patterns not apparent to the naked eye. In addition, it also supports automated content generation and enriches" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.299, + 0.548, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.552, + 0.302, + 0.748, + 0.317 + ], + "angle": 0, + "content": "Videos in SoccerReplay-1988" + }, + { + "type": "image", + "bbox": [ + 0.753, + 0.301, + 0.778, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.32, + 0.904, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.556, + 0.907, + 0.613 + ], + "angle": 0, + "content": "Figure 1. Overview. We present SoccerReplay-1988, the largest soccer dataset to date, and a powerful soccer-specific visual encoder, MatchVision, capable of excelling in various tasks such as event classification and commentary generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.907, + 0.687 + ], + "angle": 0, + "content": "fan engagement through interactive and personalized content [34, 36, 40]. These capabilities promote a deeper understanding of soccer, simplify content creation, and foster a more engaging experience for fans and professionals." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.909, + 0.854 + ], + "angle": 0, + "content": "Existing research in soccer video analysis primarily revolves around the SoccerNet series datasets [7, 10, 15], which comprise 500 full-match videos for benchmarking various tasks, such as action spotting [10, 15] and commentary generation [34, 36, 40]. Despite this extensive coverage, the focus has predominantly been on designing specialized models for task-specific applications, leading to fragmented and incompatible solutions. Such fragmentation underscores the need for a unified framework capable of integrating diverse demands, enabling more holistic and scalable advancements in soccer video understanding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce SoccerReplay-1988, the largest and most comprehensive multi-modal soccer video dataset to date, featuring 1,988 complete match videos with" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.351, + 0.888 + ], + "angle": 0, + "content": "*: These authors contribute equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.889, + 0.227, + 0.9 + ], + "angle": 0, + "content": "†: Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.875, + 0.351, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8384" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "rich annotations, such as event labels and textual commentaries. This dataset offers a solid foundation for developing advanced soccer understanding models and establishes a challenging new benchmark for the field. Additionally, we have harmonized existing datasets to be compatible with ours, further expanding the available data resources." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.483, + 0.363 + ], + "angle": 0, + "content": "Leveraging this dataset, we develop MatchVision, an advanced soccer-specific visual encoder tailored for diverse soccer understanding tasks. It employs the cutting-edge visual-language foundation model as the backbone, e.g., SigLIP [58]. We further extend framewise visual features into spatiotemporal representations with temporal attentions [3], by training on diverse visual-language tasks on SoccerReplay-1988, as depicted in Figure 1. As a result, MatchVision exhibits strong adaptability across various tasks, such as event classification and commentary generation, serving as a universal and unified framework for comprehensive soccer video understanding." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.364, + 0.483, + 0.665 + ], + "angle": 0, + "content": "To summarize, we make the following contributions in this paper: (i) we construct SoccerReplay-1988, the largest and most diverse soccer video dataset to date, featuring videos of 1,988 soccer matches with rich annotations, supported by an automated curation pipeline. This provides a solid foundation for developing robust and comprehensive soccer understanding models; (ii) we present a powerful soccer-specific visual encoder, termed MatchVision, which effectively leverages spatiotemporal information in soccer videos, and can adapt to various tasks such as event classification and commentary generation, serving as a unified framework for soccer understanding; (iii) we establish more comprehensive and challenging benchmarks based on our dataset, enabling more professional evaluation of soccer understanding models; (iv) extensive experiments and ablation studies demonstrate the superiority of our data and model across various downstream tasks, achieving state-of-the-art performance on both existing benchmarks and our newly established ones. We believe this work offers a viable paradigm for future sports video understanding." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.677, + 0.24, + 0.693 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.702, + 0.483, + 0.853 + ], + "angle": 0, + "content": "Sports Understanding [44] is an evolving field that encompasses multiple research topics and integrates diverse data modalities, covering various tasks such as action spotting [10, 15, 16], commentary generation [34, 37, 40, 51, 52, 57], athlete analysis [41, 55], tactical planning [49], sports health [39], and intelligent refereeing [22, 23]. Furthermore, with the rapid development of multimodal large language models (MLLMs), recent efforts [26, 50, 53, 54] have attempted to build more generalized frameworks to uniformly handle a variety of sports understanding tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Visual-Language Models [1, 27, 28, 38, 58] have exhibited remarkable performance across extensive applications like classification, segmentation, image-text retrieval, and image" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "captioning. Recent efforts have ventured into more challenging video understanding [29, 30, 42, 59, 60] tasks, such as temporal alignment [17, 31], dense captioning [5, 56, 62], and audio description [18-20]. However, these efforts typically focus on general scenarios, limiting their adaptability to specific professional fields. Thus, this paper aims to bridge this gap by advancing visual-language models tailored for comprehensive soccer understanding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.215, + 0.907, + 0.382 + ], + "angle": 0, + "content": "Soccer Game Analysis [9] has primarily focused on tasks such as action spotting [15], replay grounding [10, 61], commentary generation [34, 36, 40], player tracking [8], state reconstruction [43], camera calibration [6, 8, 10] and foul recognition [22, 23], as facilitated by the SoccerNet [7, 10, 14, 15] series datasets, with 500 full-match videos from 2015 to 2017. Unlike existing methods that target designing specific models for distinct tasks, this paper aims to design a unified multi-modal framework that leverages spatiotemporal information within videos, serving as a specialized visual encoder for soccer video understanding." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.398, + 0.766, + 0.415 + ], + "angle": 0, + "content": "3. SoccerReplay-1988 Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.905, + 0.514 + ], + "angle": 0, + "content": "To establish a solid foundation for soccer understanding, we construct SoccerReplay-1988, the largest soccer dataset to date. Here, we first outline our data collection details and an overview of the dataset in Sec. 3.1; followed by elaborating on our automated data curation pipeline in Sec. 3.2; lastly, in Sec. 3.3, we present the data statistics and discussion." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.526, + 0.691, + 0.54 + ], + "angle": 0, + "content": "3.1. Dataset Collection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.905, + 0.73 + ], + "angle": 0, + "content": "To construct the SoccerReplay-1988 dataset, we have collected untrimmed, full-match videos from the Internet, encompassing a total of 1,988 matches from six European major soccer leagues and championships1, spanning the 2014-15 to 2023-24 seasons. For each match, we acquire textual commentaries with second-level timestamps from a sports text live website2, with part of them annotated with specific event types such as corner and goal. Additionally, we also incorporate extensive metadata, including detailed background information about the games, players, coaches, referees, and teams, providing a solid foundation for future soccer understanding research." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.731, + 0.905, + 0.851 + ], + "angle": 0, + "content": "We partition the SoccerReplay-1988 dataset into train, validation, and test sets, containing 1,488, 250, and 250 full-match videos with diverse and comprehensive annotations, respectively. These sets provide rich training data for downstream tasks, such as event classification and commentary generation, while establishing comprehensive and challenging benchmarks for soccer understanding, as further discussed in subsequent sections." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.863, + 0.905, + 0.888 + ], + "angle": 0, + "content": "1Premier (England), Laliga (Spain), Bundesliga (Germany), Serie-a (Italy), League-1 (France) and UEFA Champions League." + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.888, + 0.603, + 0.9 + ], + "angle": 0, + "content": "2flashscore.com" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.863, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8385" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.478, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.317, + 0.483, + 0.361 + ], + "angle": 0, + "content": "Figure 2. Automated Data Curation Pipeline. The collected soccer video data are automatically processed for temporal alignment, event summarization, and anonymization by our curation pipeline." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.371, + 0.33, + 0.387 + ], + "angle": 0, + "content": "3.2. Automated Data Curation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.393, + 0.484, + 0.47 + ], + "angle": 0, + "content": "Given the potential noise in raw data, such as irrelevant video content, inaccurate timestamps, and incomplete event annotations, we design an automated data curation pipeline, comprising (i) temporal alignment, (ii) event summarization, and (iii) anonymization, as illustrated in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.472, + 0.484, + 0.532 + ], + "angle": 0, + "content": "Temporal Alignment. Here, we divide match videos into two halves, each starting at kick-off, and adopt the temporal alignment model from MatchTime [40], to synchronize textual commentary timestamps with those of video frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.484, + 0.763 + ], + "angle": 0, + "content": "Event Summarization. For samples without event annotations, we leverage LLaMA-3-70B [12] to summarize the events based on textual commentaries. Concretely, we have expanded the event categories from 17 in SoccerNet [10] to 24 types, for finer-grained soccer understanding, for example, categorizing penalties into scored and missed, and integrating modern soccer regulations like VAR. The resulting 24 event labels include: 'corner', 'goal', 'injury', 'own goal', 'penalty', 'penalty missed', 'red card', 'second yellow card', 'substitution', 'start of game (half)', 'end of game (half)', 'yellow card', 'throw in', 'free kick', 'saved by goal-keeper', 'shot off target', 'clearance', 'lead to corner', 'off-side', 'var', 'foul (no card)', 'statistics and summary', 'ball possession', and 'ball out of play'. More details on the used prompts are provided in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.841 + ], + "angle": 0, + "content": "Anonymization. Similar to [34], we extract all person and team entity names from the metadata of SoccerReplay-1988, and replace them in textual commentaries with placeholders, such as \"[PLAYER]\", \"[TEAM]\", \"[COACH]\", and \"[REFEREE]\", ensuring consistency across tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Moreover, our data curation pipeline can seamlessly extend to existing datasets, converting the SoccerNet series [10, 34] into our unified data format, termed SoccerNetpro. This expansion further enlarges the standardized" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.906, + 0.264 + ], + "angle": 0, + "content": "
Existing Datasets
# GameDuration(h)# Event# Anno.# Com.
SoccerNet-v1 [15]50076476.7k-
SoccerNet-v2 [10]50076417110k-
MatchTime [40]4717161414k37k
GOAL [36]2025.5--8.9k
Our Curated Datasets
# GameDuration(h)# Event# Anno.# Com.
SoccerNet-pro50076424102k37k
SoccerReplay-19881,9883,32324150k150k
Integrated2,4884,08724252k187k
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.271, + 0.907, + 0.328 + ], + "angle": 0, + "content": "Table 1. Statistics of Soccer Datasets. Our SoccerReplay-1988 significantly surpasses existing datasets in both scale and diversity. Here, # Anno. and # Com. refer to the number of event annotations and textual commentaries, respectively." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.342, + 0.838, + 0.358 + ], + "angle": 0, + "content": "datasets available for soccer understanding tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.371, + 0.724, + 0.385 + ], + "angle": 0, + "content": "3.3. Statistics & Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.906, + 0.468 + ], + "angle": 0, + "content": "Dataset Statistics. As shown in Table 1, our dataset encompasses 3,323 hours of footage from 1,988 soccer matches, with an average duration of 100.3 minutes per match. The videos range in resolution from 360p to 720p and frame rates between 25 and 30 FPS." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.906, + 0.622 + ], + "angle": 0, + "content": "For textual annotations, this dataset features approximately 150K commentaries, averaging 76 per match, precisely temporal-aligned by the robust alignment model from MatchTime [40]. These commentaries cover 4,467 unique words, significantly surpassing the 2,873 words in existing datasets [34, 40], greatly enriching textual diversity. Automated event summarization based on these commentaries has yielded about 150K event annotations. Notably, a random sampling of \\(2\\%\\) of the data yields \\(98\\%\\) manual verification accuracy, ensuring high-quality automated labeling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.625, + 0.906, + 0.762 + ], + "angle": 0, + "content": "SoccerReplay-test Benchmark. To facilitate a more comprehensive evaluation of soccer understanding models, we integrate 250 matches from SoccerReplay-1988 with 50 matches from the curated SoccerNet-pro, establishing SoccerReplay-test, a more challenging benchmark for event classification and commentary generation. This benchmark features nearly four times larger than existing datasets and comprises finer-grained event labels, richer textual commentaries, and up-to-date soccer regulations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Discussion. To summarize, SoccerReplay-1988 exhibits advancements in three aspects: (i) it is the largest soccer video dataset to date, with nearly four times more videos than existing datasets; (ii) it features more professional and diverse annotations, more suitable for fine-grained and comprehensive soccer understanding tasks; (iii) It employs an automated curation pipeline for annotations and is thus scalable to provide a solid data foundation for future research. All data from SoccerReplay-1988, including videos" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8386" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.452, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.21, + 0.356, + 0.369, + 0.368 + ], + "angle": 0, + "content": "(a) Structure of MatchVision" + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.09, + 0.904, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.49, + 0.356, + 0.648, + 0.369 + ], + "angle": 0, + "content": "(b) MatchVision Pretraining" + }, + { + "type": "image_caption", + "bbox": [ + 0.737, + 0.356, + 0.862, + 0.368 + ], + "angle": 0, + "content": "(c) Downstream Tasks" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.378, + 0.907, + 0.421 + ], + "angle": 0, + "content": "Figure 3. Overview of MatchVision. (a) The model architecture and its spatiotemporal feature extraction process; (b) Details of visual encoder pretraining, including supervised classification and video-language contrastive learning; (c) Implementation details of specific heads for various downstream tasks, including commentary generation, foul recognition, and event classification." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.436, + 0.479, + 0.451 + ], + "angle": 0, + "content": "and annotations, are open-source for non-commercial use." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.463, + 0.182, + 0.479 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.484, + 0.609 + ], + "angle": 0, + "content": "In this paper, we aim to develop a soccer-specific visual encoder, MatchVision, tailored for diverse soccer video analysis tasks. We start by outlining our problem formulation in Sec. 4.1. Next, in Sec. 4.2, we detail the architecture of MatchVision. The training procedures are thoroughly discussed in Sec. 4.3. Finally, we describe the configurations for downstream tasks in Sec. 4.4, demonstrating the practical applications and effectiveness of our model." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.618, + 0.295, + 0.632 + ], + "angle": 0, + "content": "4.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.64, + 0.484, + 0.716 + ], + "angle": 0, + "content": "In this work, we tackle the challenge of analyzing soccer video segments, denoted as \\(\\mathcal{V} \\in \\mathbb{R}^{T \\times 3 \\times H \\times W}\\). Our goal is to utilize the visual encoder \\((\\Phi_{\\mathrm{MatchVision}})\\) to extract spatiotemporal features from these segments, which are then processed by multiple task-specific heads, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.725, + 0.393, + 0.741 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E}, \\mathbf {C}, \\mathbf {F} = \\Psi (\\Phi_ {\\text {M a t c h V i s i o n}} (\\mathcal {V}))\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.841 + ], + "angle": 0, + "content": "Here, \\(\\Psi = \\{\\Psi_{\\mathrm{cls}}, \\Psi_{\\mathrm{Cmt}}, \\Psi_{\\mathrm{Foul}}\\}\\) represents the task-specific heads, with \\(\\mathbf{E}\\), \\(\\mathbf{C}\\), and \\(\\mathbf{F}\\) denoting the output event types, textual commentaries, and foul types, respectively. This unified framework effectively learns relevant spatiotemporal features, and enables seamless integration across various downstream tasks for comprehensive soccer understanding." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.226, + 0.863 + ], + "angle": 0, + "content": "4.2. Architecture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "MatchVision comprises three key components: (i) Token Embedding, (ii) Spatiotemporal Attention Block, and (iii)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.436, + 0.803, + 0.451 + ], + "angle": 0, + "content": "Aggregation Layer, as depicted in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.906, + 0.514 + ], + "angle": 0, + "content": "Token Embedding. In accordance with the convention in Vision Transformer [11], each frame \\((\\mathcal{I}_i)\\) from the video segment \\((\\mathcal{V} = \\{\\mathcal{I}_1, \\mathcal{I}_2, \\dots, \\mathcal{I}_T\\})\\) is divided into \\(M\\) non-overlapping patches of size \\(P \\times P\\) that span the entire frame." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.514, + 0.907, + 0.634 + ], + "angle": 0, + "content": "These patches are flattened into vectors \\((\\mathbf{x}_i^p)\\), where \\(p\\) and \\(i\\) denote the spatial and temporal positions, respectively. Each vector is transformed via an embedding layer \\((\\Phi_{\\mathrm{Emb}})\\) into a token vector of size \\(\\mathbb{R}^{1\\times D}\\), and then added with a spatial position embedding \\((\\mathbf{e}_{\\mathrm{s}}^{\\mathrm{pos}}\\in \\mathbb{R}^{M\\times D})\\). Subsequently, we concatenate a [c1s] token along with each frame. Finally, a temporal positional embedding \\((\\mathbf{e}_{\\mathrm{t}}^{\\mathrm{pos}}\\in \\mathbb{R}^{T\\times D})\\) is added across features of all frames, as formulated below:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.641, + 0.848, + 0.678 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {y} _ {i} = \\left[ \\mathbf {x} _ {i} ^ {\\mathrm {c l s}}, \\Phi_ {\\mathrm {E m b}} \\left(\\left[ \\mathbf {x} _ {i} ^ {1}, \\dots , \\mathbf {x} _ {i} ^ {M} \\right]\\right) + \\mathbf {e} _ {\\mathrm {s}} ^ {\\mathrm {p o s}} \\right] \\\\ \\mathbf {z} = \\left[ \\mathbf {y} _ {1}, \\dots , \\mathbf {y} _ {T} \\right] + \\mathbf {e} _ {\\mathrm {t}} ^ {\\text {p o s}} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.906, + 0.825 + ], + "angle": 0, + "content": "Here, \\([\\cdot ,\\cdot ]\\) denotes concatenation, and \\(\\mathbf{y}_i\\in \\mathbb{R}^{(M + 1)\\times D}\\) represents the frame-wise features. The embedded features (z) will then serve as input for spatiotemporal attention blocks. Spatiotemporal Attention Block. Similar to TimeSformer [3], we utilize interleaved temporal and spatial attention to integrate spatiotemporal information in soccer videos. Concretely, each spatiotemporal attention block comprises a temporal self-attention layer and a spatial self-attention layer, i.e., \\(\\phi_t(\\cdot)\\) and \\(\\phi_s(\\cdot)\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.9 + ], + "angle": 0, + "content": "Given a video feature \\((\\mathbf{z} \\in \\mathbb{R}^{T \\times (M + 1) \\times D})\\), we alternate temporal and spatial attention: temporal attention facilitates interactions among tokens at the same spatial positions across distinct frames, while spatial attention enables interactions among tokens within the same frame. Residual con" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8387" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "nections are employed in each layer. After passing through a total of \\(\\mathbf{K}\\) spatiotemporal attention blocks, the resulting feature \\((\\mathcal{F})\\) captures both intra-frame and inter-frame relationships, i.e., \\(\\mathcal{F} = [\\phi_s(\\phi_t(\\mathbf{z}))]^{\\mathbf{K}} \\in \\mathbb{R}^{T \\times (M + 1) \\times D}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.155, + 0.483, + 0.306 + ], + "angle": 0, + "content": "Aggregation Layer. To obtain video-level features, we employ an aggregation layer on the frame-wise spatiotemporal features. Specifically, for the \\(i\\)-th frame, we utilize spatial self-attention to aggregate information into its [cls] token, denoted as \\(\\hat{\\mathcal{F}}_i^{\\mathrm{cls}} = \\Phi_{\\mathrm{Agg}}(\\mathcal{F}_i)\\). Concatenating the [cls] tokens of all frames yields the final video feature \\((\\mathcal{F}_{\\mathcal{V}})\\), that effectively encapsulates spatiotemporal characteristics of soccer video segments, thus enabling it to be applicable for various downstream soccer understanding tasks. This process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.319, + 0.462, + 0.339 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\mathcal {V}} = \\Phi_ {\\mathrm {M a t c h V i s i o n}} (\\mathcal {V}) = [ \\hat {\\mathcal {F}} _ {1} ^ {\\mathrm {c l s}}, \\dots , \\hat {\\mathcal {F}} _ {T} ^ {\\mathrm {c l s}} ] \\in \\mathbb {R} ^ {T \\times D}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.351, + 0.32, + 0.367 + ], + "angle": 0, + "content": "4.3. MatchVision Pretraining" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.374, + 0.483, + 0.45 + ], + "angle": 0, + "content": "In this part, we aim to pretrain the visual encoder with triplet samples \\((\\{\\mathcal{V},\\mathbf{E},\\mathbf{C}\\})\\), comprising videos, event labels, and textual commentaries. Concretely, we investigate two distinct pretraining strategies: supervised classification and video-language contrastive learning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.453, + 0.483, + 0.573 + ], + "angle": 0, + "content": "Supervised Classification. One way to pretrain the visual encoder is supervised learning on event classification. To be specific, the extracted visual features \\((\\mathcal{F}_{\\mathcal{V}})\\) are aggregated by a temporal self-attention layer into a learnable [cls] token, denoted as \\(\\mathcal{F}_{\\mathcal{V}}^{\\mathrm{cls}}\\), similar to the spatial-wise aggregation mentioned above. This token is then fed into a linear classifier, and trained with a cross-entropy loss for event classification. The objective is denoted as \\(\\mathcal{L}_{\\mathrm{sup}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.577, + 0.483, + 0.744 + ], + "angle": 0, + "content": "Video-Language Contrastive Learning. As an alternative, we can also pretrain our visual encoder with video-text contrastive learning. Specifically, we adopt simple average pooling on the video feature to get the aggregated visual feature \\((\\mathcal{F}_{\\mathcal{V}}^{\\mathrm{Avg}})\\), and encode the textual commentary (C) with a text encoder \\((\\Phi_{\\mathrm{Text}})\\). We train the model with sigmoid loss (\\(\\mathcal{L}_{\\mathrm{sigmoid}}\\)), as used in SigLIP [58]. Note that, some video clips may have highly similar commentaries, for example, 'start of the game', we treat the commentaries with high similarity in the same batch as positive samples when calculating loss functions. This can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.755, + 0.414, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n t r a}} = \\mathcal {L} _ {\\text {s i g m o i d}} \\left(\\mathcal {F} _ {\\mathcal {V}} ^ {\\text {A v g}}, \\Phi_ {\\text {T e x t}} (\\mathbf {C})\\right)\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.274, + 0.803 + ], + "angle": 0, + "content": "4.4. Downstream Tasks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "After the pretraining mentioned above, MatchVision can now serve as a versatile visual encoder, to map the soccer video segments into visual features \\((\\mathcal{F}_{\\mathcal{V}})\\), for training task-specific heads \\(\\Psi = \\{\\Psi_{\\mathrm{cls}}, \\Psi_{\\mathrm{Cmt}}, \\Psi_{\\mathrm{Foul}}\\}\\) across different downstream tasks, including: (i) event classification, (ii) commentary generation, and (iii) foul recognition." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "Event Classification. Similar to supervised classification above, we concatenate a learnable [cls] token to aggregate frame-wise visual features via temporal self-attention. This token is then fed into a linear classifier for event classification. The event classification head \\((\\Psi_{\\mathrm{cls}})\\) is trained with a cross-entropy loss while freezing the visual encoder." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.186, + 0.907, + 0.337 + ], + "angle": 0, + "content": "Commentary Generation. We follow the paradigm in MatchTime [40] to generate anonymized textual commentary for soccer video clips. Concretely, the commentary generation head \\((\\Psi_{\\mathrm{Cmt}})\\) employs a Perceiver [25] aggregator to consolidate visual features, which are then projected by a trainable MLP, serving as prefix embeddings for a large language model (LLM). Subsequently, an off-the-shelf LLM decodes these embeddings into textual commentary. We adopt the negative log-likelihood loss, commonly used for auto-regressive next-token prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.906, + 0.521 + ], + "angle": 0, + "content": "Foul Recognition. As outlined in [22], the foul recognition task takes multi-view videos from the same scene as inputs, with each sample annotated with a foul class (8 types) and severity (4 levels). We encode these multi-view videos with MatchVision, and aggregate the extracted features into a single feature vector, via either max or average pooling, following the common practice. Subsequently, the foul recognition head \\((\\Psi_{\\mathrm{Foul}})\\) employs a shared MLP and two task-specific linear classifiers, to predict foul type and severity, respectively. Similar to event classification, we use the combination of cross-entropy losses on the foul type and severity classification to jointly train \\(\\Psi_{\\mathrm{Foul}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.906, + 0.584 + ], + "angle": 0, + "content": "Discussion. Pretraining MatchVision on large-scale soccer data equips it with substantial domain-specific knowledge, enabling it to serve as a universal visual encoder adaptable to various downstream soccer understanding tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.6, + 0.646, + 0.617 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.625, + 0.906, + 0.716 + ], + "angle": 0, + "content": "This section begins with implementation details in Sec. 5.1; followed by quantitative evaluations across downstream tasks in Sec. 5.2; then, we conduct ablation studies on our SoccerReplay-test benchmark to analyze the effectiveness of the proposed dataset and model in Sec. 5.3; finally, we provide qualitative results for comparison in Sec. 5.4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.728, + 0.731, + 0.743 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In our experiments, video segments are sampled at 1FPS around annotated timestamps, capturing a 30-second window for each sample. Frames are resized to \\(224 \\times 224\\) pixels as inputs. We initialize the embedding layer, spatial attention layers, aggregation layer, and text encoder of MatchVision with pretrained weights from SigLIP Base-16 [58] and adopt LLaMA-3 (8B) [12] as the off-the-shelf LLM decoder for commentary generation. All experiments are conducted on \\(4 \\times\\) Nvidia H800 GPUs with the AdamW [33] optimizer. Next, we elaborate on the training and evaluation details" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8388" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.143, + 0.089, + 0.854, + 0.502 + ], + "angle": 0, + "content": "
Visual EncoderDatasetClassification (%)Commentary
SNMTSRAcc.@1Acc.@3Acc.@5B@1B@4MR-LC
Off-the-shelf Models
I3D [4]XXX45.482.593.226.775.5724.1723.1218.73
C3D [45]XXX47.885.195.028.136.6424.5224.2327.88
ResNet [21]XXX47.284.694.427.346.5724.7224.4327.29
CLIP [38]XXX48.585.595.226.256.5124.2724.7528.17
InternVideo [48]XXX49.987.095.927.126.5425.0224.8229.90
SigLIP [58]XXX50.286.795.627.856.9825.1625.0331.38
Pretrain with Supervised Classification
Baidu [61]✓XX56.491.997.331.208.8826.5626.6138.93
SigLIP✓XX55.989.694.928.517.3925.9625.9435.71
SigLIP✓✓✓57.991.797.530.958.5625.7926.1738.24
MatchVision✓XX82.596.698.829.457.9226.0126.2136.15
MatchVision✓✓✓84.097.399.231.059.0626.9427.9342.20
Pretrain with Visual-Language Contrastive Learning
SigLIPX✓X55.488.897.028.727.7225.9126.1732.27
SigLIPX✓✓66.893.798.630.358.1226.0526.3839.41
MatchVisionX✓X58.989.097.130.337.9725.4826.3333.87
MatchVisionX✓✓67.993.998.631.949.1226.2427.5640.76
Pretrain with Hybrid Supervised-Contrastive Training
SigLIP✓✓X71.294.598.728.637.8225.7425.3534.09
SigLIP✓✓✓67.193.298.130.718.7826.2626.7441.82
MatchVision✓✓X76.496.099.030.658.3325.2826.3137.23
MatchVision✓✓✓80.197.199.133.589.1426.8228.2144.18
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.509, + 0.907, + 0.552 + ], + "angle": 0, + "content": "Table 2. Quantitative Results on Event Classification and Commentary Generation. Here, SN, MT, and SR represent finetuning with curated SoccerNet-v2 [10], MatchTime [40], and SoccerReplay-1988, respectively. B, M, R-L, and C refer to BLEU, METEOR, ROUGE-L, and CIDEr metrics, respectively. Within each unit, we denote the best performance in RED and the second-best performance in BLUE." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.563, + 0.459, + 0.578 + ], + "angle": 0, + "content": "about visual encoder pretraining and downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.732 + ], + "angle": 0, + "content": "Visual Encoder Pretraining. For both pretraining strategies, we use a batch size of 40 for 15 epochs. The learning rate for all randomly initialized modules, including the temporal attention layers, aggregator layer, and linear classifier, is set to \\(1 \\times 10^{-4}\\). Meanwhile, the learning rate for modules initialized with pretrained parameters (including the text encoder) is set to \\(5 \\times 10^{-5}\\). In contrastive training, we adopt a multi-positive strategy where each textual commentary, based on its event label, considers closely related categories (e.g. \"start of game\" and \"offside\") as positive samples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Downstream Tasks. In all downstream tasks, unless otherwise specified, we use the frozen visual encoder for feature extraction and only train the task-specific heads with a learning rate of \\(1 \\times 10^{-4}\\) for 30 epochs. The batch sizes for event classification, commentary generation, and foul recognition are set to 40, 32, and 8, respectively. We adopt specific evaluation metrics for these three tasks: (i) For event classification, we use the top-1/3/5 classification accuracy; (ii) For commentary generation, we employ several commonly-used language evaluation metrics, including BLEU [35], METEOR [2], ROUGE-L [32], and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.562, + 0.907, + 0.608 + ], + "angle": 0, + "content": "CIDEr [47]; (iii) For foul recognition, we follow the common practice, and report top-1/2 and top-1 accuracy for the foul type and severity classification, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.909, + 0.807 + ], + "angle": 0, + "content": "Benchmarks & Baselines. To ensure fair and reliable comparisons with existing work, we evaluate event classification (24 types) on 100 matches from curated SoccerNetv2 [10] test set; commentary generation on 49 matches from SN-Caption-test-align benchmark manually aligned in [40]; and foul recognition on MVFoul [22]. We consider various baselines: for the first two tasks, this includes off-the-shelf general visual encoders such as ResNet [21], C3D [45], I3D [4], CLIP [38], SigLIP [58], and InternVideo [48], along with Baidu [61] and SigLIP finetuned with soccer-specific data. For foul recognition, we follow previous work [22, 23] and adopt ResNet [21], R(2+1)D [46], and MViT [13] jointly finetuned with classifiers, as baselines." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.736, + 0.835 + ], + "angle": 0, + "content": "5.2.Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.908, + 0.903 + ], + "angle": 0, + "content": "As depicted in Table 2, we draw two observations on event classification and commentary generation: (i) visual encoders trained on soccer data substantially outperform off-the-shelf general encoders (ResNet, C3D, I3D, CLIP, and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8389" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.476, + 0.268 + ], + "angle": 0, + "content": "
Visual EncoderFoul ClassSeverity
BackboneTrainAgg.Acc.@1Acc.@2Acc.@1
ResNet [21]✓Mean0.310.560.34
Max0.320.600.32
R(2+1)D [46]✓Mean0.310.550.34
Max0.320.560.39
MViT [13]✓Mean0.400.650.38
Max0.470.690.43
MatchVision×Mean0.440.530.58
Max0.350.700.46
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.276, + 0.483, + 0.318 + ], + "angle": 0, + "content": "Table 3. Quantitative Results on Multi-view Foul Recognition. Our frozen MatchVision encoder can achieve comparable performance with other jointly finetuned visual encoders." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.334, + 0.483, + 0.424 + ], + "angle": 0, + "content": "InternVideo), underscoring the necessity of building specialized models for soccer understanding; (ii) almost all visual encoders, across all training settings, benefit from SoccerReplay-1988, emphasizing the value of constructing large-scale, high-quality data for soccer understanding. Next, we will delve into each task to discuss the results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.483, + 0.579 + ], + "angle": 0, + "content": "Event Classification. With identical training strategies and data, MatchVision considerably outperforms other methods in classification accuracy, demonstrating the superiority of its architecture, which effectively leverages spatiotemporal features within soccer videos. For example, MatchVision achieves a Top-1 accuracy of \\(82.5\\%\\), significantly surpassing SigLIP's \\(55.9\\%\\) under the same training conditions. Moreover, models trained via supervised classification excel others, primarily because the pre-training task shares the same objectives as the downstream event classification task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Commentary Generation. Visual encoders trained with visual-language contrastive learning exhibit better commentary generation performance than those trained with supervised classification, as this strategy better captures correlations between visual and textual features. Additionally, while MatchVision trained solely on SoccerNet slightly underperforms Baidu [61], incorporating SoccerReplay-1988 enables it to outperform on most metrics. This demonstrates that MatchVision can take advantage of large-scale datasets. Finally, a hybrid training approach, starting with supervised classification followed by visual-language contrastive learning, enables MatchVision to achieve optimal performance. This indicates that learning coarse-grained tasks such as classification provides a foundation for fine-grained tasks like commentary generation, and fully leveraging data unlocks the potential of soccer understanding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Foul Recognition. As demonstrated in Table 3, MatchVision achieves performance comparable to jointly finetuned state-of-the-art methods in foul recognition, even with a frozen visual encoder. This highlights that MatchVision effectively learns substantial knowledge from large-scale soc" + }, + { + "type": "table", + "bbox": [ + 0.548, + 0.089, + 0.872, + 0.235 + ], + "angle": 0, + "content": "
PretrainClassification(%)
Sup.Contra.SRAcc.@1Acc.@3Acc.@5
✓XX62.6783.0089.81
✓X✓68.0386.9092.38
X✓X46.9775.5385.85
X✓✓57.4183.1391.00
✓✓X56.8680.3088.09
✓✓✓63.5985.2191.63
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.241, + 0.907, + 0.312 + ], + "angle": 0, + "content": "Table 4. Ablations on Event Classification. We explore the impact of various training settings of our MatchVision encoder on the SoccerReplay-test benchmark. Here, Sup., Contra., and SR refer to supervised classification, visual-language contrastive learning, and the SoccerReplay-1988 dataset, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.329, + 0.907, + 0.376 + ], + "angle": 0, + "content": "cer data and adapts seamlessly to downstream tasks. Comparisons with additional baselines from the SoccerNet foul recognition challenges [9] are provided in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.387, + 0.68, + 0.403 + ], + "angle": 0, + "content": "5.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.485 + ], + "angle": 0, + "content": "We conduct ablation experiments on event classification and commentary generation using our SoccerReplay-test benchmark. These experiments validate the effectiveness of our proposed dataset and model, while establishing a baseline for future evaluations on this benchmark." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.909, + 0.732 + ], + "angle": 0, + "content": "Event Classification. We evaluate event classification on 300 matches from our SoccerReplay-test benchmark using the MatchVision visual encoder pretrained with various strategies. Features are extracted by MatchVision and processed with a learnable aggregation layer and a linear classifier. The default training set is our curated SoccerNet-pro. As shown in Table 4, integrating SoccerReplay-1988 for training results in significant performance improvements across all pretraining strategies, yielding the significance of our dataset. Additionally, supervised classification outperforms visual-language contrastive learning and hybrid pretraining. This is due to its closer alignment with downstream event classification task, and the scale of event annotations is far larger than that of textual commentaries, further confirming the substantial benefits of data scaling for boosting soccer understanding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Commentary Generation. With the pretrained MatchVision encoder, we train the commentary generation head on the MatchTime [40] and SoccerReplay-1988 datasets using various training strategies. By default, only the Perceiver [25] aggregation layer and projection layer within the head are trained. For joint training with the LLM decoder, considering computational costs, we incorporate LoRA [24] layers while freezing the original LLM layers. As shown in Table 5, incorporating SoccerReplay-1988 significantly improves performance on all metrics, confirming substantial advantages of our proposed dataset. This performance" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "8390" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.092, + 0.12, + 0.105 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.095, + 0.192, + 0.104 + ], + "angle": 0, + "content": "Soccer Videos" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.093, + 0.393, + 0.105 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.095, + 0.43, + 0.103 + ], + "angle": 0, + "content": "Events" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.092, + 0.539, + 0.105 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.094, + 0.617, + 0.103 + ], + "angle": 0, + "content": "Commentaries" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.108, + 0.369, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.119, + 0.46, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.115, + 0.89, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.155, + 0.368, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.163, + 0.515, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.156, + 0.892, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.199, + 0.368, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.21, + 0.489, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.202, + 0.892, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.244, + 0.368, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.255, + 0.478, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.248, + 0.89, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.291, + 0.368, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.301, + 0.461, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.298, + 0.88, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.35, + 0.907, + 0.42 + ], + "angle": 0, + "content": "Figure 4. Qualitative Results for Event Classification and Commentary Generation. Here, \"w/o SR\" and \"w/ SR\" indicate models trained without and with the SoccerReplay-1988 dataset, respectively. Incorporating SoccerReplay-1988 improves event classification accuracy. Moreover, this enriched training data enables the model to demonstrate several advantages in commentary generation: (a) more detailed descriptions, (b) greater linguistic variety, (c) higher event depiction accuracy, (d) better adherence to updated rules, and (e) improved specificity in scenario response." + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.436, + 0.458, + 0.643 + ], + "angle": 0, + "content": "
TrainableCommentary Metrics
VLB@1B@4MR-LC
Trained on MatchTime
XX21.653.2721.0217.7912.90
✓X27.627.0224.0323.5130.77
X✓27.046.4124.1523.8831.91
✓✓27.496.9624.5023.3330.81
Trained on MatchTime & SoccerReplay-1988
XX24.174.0920.5120.7015.70
✓X28.988.3924.4525.3545.85
X✓27.547.7624.5024.7042.79
✓✓29.218.2225.2525.5443.18
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.651, + 0.483, + 0.707 + ], + "angle": 0, + "content": "Table 5. Ablations on Commentary Generation. We investigate the impact of different training strategies and datasets on MatchVision using the SoccerReplay-test benchmark. 'V' and 'L' denote the visual encoder and the LLM decoder, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.484, + 0.794 + ], + "angle": 0, + "content": "gap also reflects the challenges of our established benchmark, which features diverse vocabulary, richer semantics, and updated soccer rules. Additionally, jointly finetuning the visual encoder and the LLM decoder provides a feasible approach for further improvements." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.32, + 0.82 + ], + "angle": 0, + "content": "5.4. Qualitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "As depicted in Figure 4, we present qualitative results of MatchVision on the SoccerReplay-test benchmark, comparing models pretrained with and without SoccerReplay-1988. For event classification, incorporating our data improves accuracy, and even in misclassified cases, the results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.907, + 0.515 + ], + "angle": 0, + "content": "remain contextually relevant. For commentary generation, hybrid training on SoccerReplay-1988 enables MatchVision to produce richer, more detailed textual commentary, reflecting a deeper understanding of soccer dynamics. More qualitative results are available in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.542, + 0.634, + 0.558 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.572, + 0.907, + 0.815 + ], + "angle": 0, + "content": "In this paper, we establish a unified, scalable multi-modal framework for soccer understanding. Specifically, we introduce SoccerReplay-1988, the largest and most comprehensive soccer video dataset to date, annotated by an automated curation pipeline. This provides a solid foundation for developing soccer understanding models and serves as a more challenging benchmark. Built upon this, we develop MatchVision, an advanced soccer-specific visual encoder, which effectively leverages spatiotemporal information within soccer videos and can be applied to various tasks such as event classification and commentary generation. Extensive experiments demonstrate the superiority of our model, with MatchVision achieving state-of-the-art performance on both existing benchmarks and our newly established one. We believe this work will set a viable, universal paradigm for future research in sports understanding." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.841, + 0.673, + 0.858 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This work is funded by National Key R&D Program of China (No.2022ZD0161400)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8391" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.197 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, pages 23716-23736, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.2, + 0.484, + 0.269 + ], + "angle": 0, + "content": "[2] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, pages 65-72, 2005. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.483, + 0.325 + ], + "angle": 0, + "content": "[3] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In Proceedings of the International Conference on Machine Learning, page 4, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.328, + 0.484, + 0.382 + ], + "angle": 0, + "content": "[4] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.384, + 0.482, + 0.479 + ], + "angle": 0, + "content": "[5] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, and Sergey Tulyakov. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.482, + 0.482, + 0.578 + ], + "angle": 0, + "content": "[6] Anthony Cioppa, Adrien Deliege, Floriane Magera, Silvio Giancola, Olivier Barnich, Bernard Ghanem, and Marc Van Droogenbroeck. Camera calibration and player localization in soccernet-v2 and investigation of their representations for action spotting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 4537-4546, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.581, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[7] Anthony Cioppa, Adrien Deliege, Silvio Giancola, Bernard Ghanem, and Marc Van Droogenbroeck. Scaling up soccer-net with multi-view spatial localization and re-identification. Scientific data, 9(1):355, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.637, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[8] Anthony Cioppa, Silvio Giancola, Adrien Deliege, Le Kang, Xin Zhou, Zhiyu Cheng, Bernard Ghanem, and Marc Van Droogenbroeck. Soccernet-tracking: Multiple object tracking dataset and benchmark in soccer videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 3491-3502, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.722, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[9] Anthony Cioppa, Silvio Giancola, Vladimir Somers, Victor Joos, Floriane Magera, Jan Held, Seyed Abolfazl Ghasemzadeh, Xin Zhou, Karolina Seweryn, Mateusz Kowalczyk, Zuzanna Mroz, Szymon Lukasik, Michal Halon, Hassan Mkhallati, Adrien Deliege, Carlos Hinojosa, Karen Sanchez, Amir M. Mansourian, Pierre Miralles, Olivier Barnich, Christophe De Vleeschouwer, Alexandre Alahi, Bernard Ghanem, Marc Van Droogenbroeck, Adam Gorski, Albert Clapes, Andrei Boiarov, Anton Afanasiev, Artur Xarles, Atom Scott, ByoungKwon Lim, Calvin Yeung, Cristian Gonzalez, Dominic Rufenacht, Enzo Pacilio, Fabian Deuser, Faisal Sami Altawijri, Francisco Cachon, HanKyul Kim, Haobo Wang, Hyeonmin Choe, Hyunwoo J Kim, Il-Min" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.272 + ], + "angle": 0, + "content": "Kim, Jae-Mo Kang, Jamshid Tursunboev, Jian Yang, Jihwan Hong, Jimin Lee, Jing Zhang, Junseok Lee, Kexin Zhang, Konrad Habel, Licheng Jiao, Linyi Li, Marc Gutierrez-Perez, Marcelo Ortega, Menglong Li, Milosz Lopatto, Nikita Kasatkin, Nikolay Nemtsev, Norbert Oswald, Oleg Udin, Pavel Kononov, Pei Geng, Saad Ghazai Alotaibi, Sehyung Kim, Sergei Ulasen, Sergio Escalera, Shanshan Zhang, Shuyuan Yang, Sunghwan Moon, Thomas B. Moeslund, Vasyl Shandyba, Vladimir Golovkin, Wei Dai, WonTaek Chung, Xinyu Liu, Yongqiang Zhu, Youngseo Kim, Yuan Li, Yuting Yang, Yuxuan Xiao, Zehua Cheng, and Zhihao Li. Soccernet 2024 challenges results. arXiv preprint arXiv:2409.10587, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[10] Adrien Deliege, Anthony Cioppa, Silvio Giancola, Meisam J Seikavandi, Jacob V Dueholm, Kamal Nasrollahi, Bernard Ghanem, Thomas B Moeslund, and Marc Van Droogenbroeck. Soccernet-v2: A dataset and benchmarks for holistic understanding of broadcast soccer videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 4508-4519, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.371, + 0.905, + 0.467 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.469, + 0.905, + 0.535 + ], + "angle": 0, + "content": "[12] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.538, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[13] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6824-6835, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.689 + ], + "angle": 0, + "content": "[14] Sushant Gautam, Mehdi Houshmand Sarkhoosh, Jan Held, Cise Midoglu, Anthony Cioppa, Silvio Giancola, Vajira Thambawita, Michael A Riegler, Pål Halvorsen, and Mubarak Shah. Soccernet-echoes: A soccer game audio commentary dataset. arXiv preprint arXiv:2405.07354, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.692, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[15] Silvio Giancola, Mohieddine Amine, Tarek Dghaily, and Bernard Ghanem. Soccernet: A scalable dataset for action spotting in soccer videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 1711-1721, 2018. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.762, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[16] Xiaofan Gu, Xinwei Xue, and Feng Wang. Fine-grained action recognition on a novel basketball dataset. In International Conference on Acoustics, Speech, and Signal Processing, pages 2563-2567, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[17] Tengda Han, Weidi Xie, and Andrew Zisserman. Temporal alignment networks for long-term video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2906-2916, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[18] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. Autoad: Movie description in" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8392" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[19] Tengda Han, Max Bain, Arsha Nagrani, Gul Varol, Weidi Xie, and Andrew Zisserman. Autoad ii: The sequel-who, when, and what in movie audio description. In Proceedings of the International Conference on Computer Vision, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.18, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[20] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. Autoad iii: The prequel - back to the pixels. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.482, + 0.292 + ], + "angle": 0, + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.294, + 0.482, + 0.377 + ], + "angle": 0, + "content": "[22] Jan Held, Anthony Cioppa, Silvio Giancola, Abdullah Hamdi, Bernard Ghanem, and Marc Van Droogenbroeck. Vars: Video assistant referee system for automated soccer decision making from multiple views. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 5085-5096, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.379, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[23] Jan Held, Hani Itani, Anthony Cioppa, Silvio Giancola, Bernard Ghanem, and Marc Van Droogenbroeck. X-vars: Introducing explainability in football refereeing with multimodal large language models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 3267-3279, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In Proceedings of the International Conference on Learning Representations, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[25] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In Proceedings of the International Conference on Machine Learning, pages 4651-4664, 2021. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.605, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[26] Haopeng Li, Andong Deng, Qiuhong Ke, Jun Liu, Hossein Rahmani, Yulan Guo, Bernt Schiele, and Chen Chen. Sports-qa: A large-scale video question answering benchmark for complex and professional sports. arXiv preprint arXiv:2401.01505, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.745 + ], + "angle": 0, + "content": "[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the International Conference on Machine Learning, pages 12888-12900, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the International Conference on Machine Learning, pages 19730–19742, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[29] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, Limin Wang, and Yu Qiao. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[30] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In Proceedings of the European Conference on Computer Vision, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[31] Zeqian Li, Qirui Chen, Tengda Han, Ya Zhang, Yanfeng Wang, and Weidi Xie. Multi-sentence grounding for long-term instructional video. In Proceedings of the European Conference on Computer Vision, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[32] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Proceedings of the International Conference on Learning Representations, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[34] Hassan Mkhallati, Anthony Cioppa, Silvio Giancola, Bernard Ghanem, and Marc Van Droogenbroeck. Socernetc caption: Dense video captioning for soccer broadcasts commentaries. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 5074-5085, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[35] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Association for Computational Linguistics, pages 311-318, 2002. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[36] Ji Qi, Jifan Yu, Teng Tu, Kunyu Gao, Yifan Xu, Xinyu Guan, Xiaozhi Wang, Bin Xu, Lei Hou, Juanzi Li, et al. Goal: A challenging knowledge-grounded video captioning benchmark for real-time soccer commentary generation. In Proceedings of the ACM International Conference on Information and Knowledge Management, pages 5391-5395, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[37] Mengshi Qi, Yunhong Wang, Annan Li, and Jiebo Luo. Sports video captioning via attentive motion representation and group relationship modeling. IEEE Transactions on Circuits and Systems for Video Technology, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[39] Prem N Ramkumar, Bryan C Luu, Heather S Haeberle, Jaret M Karnuta, Benedict U Nwachukwu, and Riley J Williams. Sports medicine and artificial intelligence: a primer. The American Journal of Sports Medicine, 50(4): 1166-1174, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[40] Jiayuan Rao, Haoning Wu, Chang Liu, Yanfeng Wang, and Weidi Xie. Matchtime: Towards automatic soccer game commentary generation. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2024. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[41] Dian Shao, Yue Zhao, Bo Dai, and Dahua Lin. Finegym: A hierarchical video dataset for fine-grained action understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2616-2625, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[42] Yan Shu, Zheng Liu, Peitian Zhang, Minghao Qin, Junjie Zhou, Zhengyang Liang, Tiejun Huang, and Bo Zhao." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "8393" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "Video-xl: Extra-long vision language model for hour-scale video understanding. arXiv preprint arXiv:2409.14485, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.234 + ], + "angle": 0, + "content": "[43] Vladimir Somers, Victor Joos, Anthony Cioppa, Silvio Giancola, Seyed Abolfazl Ghasemzadeh, Floriane Magera, Baptiste Standaert, Amir M Mansourian, Xin Zhou, Shohreh Kasaei, et al. Soccernet game state reconstruction: End-to-end athlete tracking and identification on a minimap. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 3293-3305, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[44] Graham Thomas, Rikke Gade, Thomas B Moeslund, Peter Carr, and Adrian Hilton. Computer vision for sports: Current applications and research topics. Computer Vision and Image Understanding, 159:3-18, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.359 + ], + "angle": 0, + "content": "[45] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the International Conference on Computer Vision, pages 4489-4497, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[46] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.483, + 0.489 + ], + "angle": 0, + "content": "[47] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4566-4575, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.483, + 0.559 + ], + "angle": 0, + "content": "[48] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.483, + 0.63 + ], + "angle": 0, + "content": "[49] Zhe Wang, Petar Velicković, Daniel Hennes, Nenad Tomasev, Laurel Prince, Michael Kaisers, Yoram Bachrach, Romuald Elie, Li Kevin Wenliang, Federico Piccinini, et al. Tacticai: an ai assistant for football tactics. Nature Communications, 15(1):1-13, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.483, + 0.673 + ], + "angle": 0, + "content": "[50] Dekun Wu, He Zhao, Xingce Bao, and Richard P Wildes. Sports video analysis on large-scale data. In Proceedings of the European Conference on Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[51] Zeyu Xi, Ge Shi, Xuefen Li, Junchi Yan, Zun Li, Lifang Wu, Zilin Liu, and Liang Wang. A simple yet effective knowledge guided method for entity-aware video captioning on a basketball benchmark. Neurocomputing, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[52] Zeyu Xi, Ge Shi, Haoying Sun, Bowen Zhang, Shuyi Li, and Lifang Wu. Eika: Explicit & implicit knowledge-augmented network for entity-aware sports video captioning. Expert Systems with Applications, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[53] Haotian Xia, Zhengbang Yang, Yuqing Wang, Rhys Tracy, Yun Zhao, Dongdong Huang, Zezhi Chen, Yan Zhu, Yuanfang Wang, and Weining Shen. Sportqa: A benchmark for sports understanding in large language models. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[54] Haotian Xia, Zhengbang Yang, Junbo Zou, Rhys Tracy, Yuqing Wang, Chi Lu, Christopher Lai, Yanjun He, Xun" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.149 + ], + "angle": 0, + "content": "Shao, Zhuoqing Xie, et al. Sportu: A comprehensive sports understanding benchmark for multimodal large language models. In Proceedings of the International Conference on Learning Representations, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[55] Jinglin Xu, Yongming Rao, Xumin Yu, Guangyi Chen, Jie Zhou, and Jiwen Lu. Finediving: A fine-grained dataset for procedure-aware action quality assessment. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2949-2958, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[56] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[57] Huanyu Yu, Shuo Cheng, Bingbing Ni, Minsi Wang, Jian Zhang, and Xiaokang Yang. Fine-grained video captioning for sports narrative. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[58] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the International Conference on Computer Vision, pages 11975-11986, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[59] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.475, + 0.905, + 0.543 + ], + "angle": 0, + "content": "[60] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.545, + 0.905, + 0.6 + ], + "angle": 0, + "content": "[61] Xin Zhou, Le Kang, Zhiyu Cheng, Bo He, and Jingyu Xin. Feature combination meets attention: Baidu soccer embeddings and transformer based temporal detection. arXiv preprint arXiv:2106.14447, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.601, + 0.905, + 0.67 + ], + "angle": 0, + "content": "[62] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8394" + } + ] +] \ No newline at end of file diff --git a/2025/Towards Universal Soccer Video Understanding/bd5f526c-bc30-45cc-bbcd-e151dfc21f1a_origin.pdf b/2025/Towards Universal Soccer Video Understanding/bd5f526c-bc30-45cc-bbcd-e151dfc21f1a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0e098d479f042bb3271b65cc8f482c0cb882727f --- /dev/null +++ b/2025/Towards Universal Soccer Video Understanding/bd5f526c-bc30-45cc-bbcd-e151dfc21f1a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0e040d999cee2b6f12893a0062c59b25c6b4f72c8cdd5de5d684b7b8360bcefe +size 7755598 diff --git a/2025/Towards Universal Soccer Video Understanding/full.md b/2025/Towards Universal Soccer Video Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cd0c844b859923da9887442a391d1a570b64090e --- /dev/null +++ b/2025/Towards Universal Soccer Video Understanding/full.md @@ -0,0 +1,348 @@ +# Towards Universal Soccer Video Understanding + +Jiayuan Rao $^{1,2*}$ , Haoning Wu $^{1,2*}$ , Hao Jiang $^{3}$ , Ya Zhang $^{1}$ , Yanfeng Wang $^{1\dagger}$ , Weidi Xie $^{1\dagger}$ $^{1}$ School of Artificial Intelligence, Shanghai Jiao Tong University, China + $^{2}$ CMIC, Shanghai Jiao Tong University, China +https://jyrao.github.io/UniSoccer/ + +# Abstract + +As a globally celebrated sport, soccer has attracted widespread interest from fans all over the world. This paper aims to develop a comprehensive multi-modal framework for soccer video understanding. Specifically, we make the following contributions in this paper: (i) we introduce SoccerReplay-1988, the largest multi-modal soccer dataset to date, featuring videos and detailed annotations from 1,988 complete matches, with an automated annotation pipeline; (ii) we present an advanced soccer-specific visual encoder, MatchVision, which leverages spatiotemporal information across soccer videos and excels in various downstream tasks; (iii) we conduct extensive experiments and ablation studies on event classification, commentary generation, and multi-view foul recognition. MatchVision demonstrates state-of-the-art performance on all of them, substantially outperforming existing models, which highlights the superiority of our proposed data and model. We believe that this work will offer a standard paradigm for sports understanding research. + +"Football is one of the world's best means of communication. It is impartial, apolitical, and universal." + +— Franz Beckenbauer (1945 - 2024) + +# 1. Introduction + +Soccer, celebrated worldwide for its significant commercial value, has recently seen great research interest in integrating artificial intelligence (AI) for soccer video understanding. This is primarily motivated by the sport's complexity and the growing demand for enhanced analytics and improved viewing experiences. AI systems facilitate tactical analysis [49], allowing coaches to devise better strategies by uncovering patterns not apparent to the naked eye. In addition, it also supports automated content generation and enriches + +![](images/b768a20c91a1e5521d31cbbdecbc86cf5ce93b5b2297350c47f6ff106847ef5b.jpg) +Videos in SoccerReplay-1988 + +![](images/d8fceea74ff9cd5dcca5ddf474b1fa93d9e505c697fe8845e096779cc9f8715b.jpg) + +![](images/ed384a51e59726eec53f918b94be181ccdfedb6766313c36bbcbed6f6b542de5.jpg) +Figure 1. Overview. We present SoccerReplay-1988, the largest soccer dataset to date, and a powerful soccer-specific visual encoder, MatchVision, capable of excelling in various tasks such as event classification and commentary generation. + +fan engagement through interactive and personalized content [34, 36, 40]. These capabilities promote a deeper understanding of soccer, simplify content creation, and foster a more engaging experience for fans and professionals. + +Existing research in soccer video analysis primarily revolves around the SoccerNet series datasets [7, 10, 15], which comprise 500 full-match videos for benchmarking various tasks, such as action spotting [10, 15] and commentary generation [34, 36, 40]. Despite this extensive coverage, the focus has predominantly been on designing specialized models for task-specific applications, leading to fragmented and incompatible solutions. Such fragmentation underscores the need for a unified framework capable of integrating diverse demands, enabling more holistic and scalable advancements in soccer video understanding. + +In this paper, we introduce SoccerReplay-1988, the largest and most comprehensive multi-modal soccer video dataset to date, featuring 1,988 complete match videos with + +rich annotations, such as event labels and textual commentaries. This dataset offers a solid foundation for developing advanced soccer understanding models and establishes a challenging new benchmark for the field. Additionally, we have harmonized existing datasets to be compatible with ours, further expanding the available data resources. + +Leveraging this dataset, we develop MatchVision, an advanced soccer-specific visual encoder tailored for diverse soccer understanding tasks. It employs the cutting-edge visual-language foundation model as the backbone, e.g., SigLIP [58]. We further extend framewise visual features into spatiotemporal representations with temporal attentions [3], by training on diverse visual-language tasks on SoccerReplay-1988, as depicted in Figure 1. As a result, MatchVision exhibits strong adaptability across various tasks, such as event classification and commentary generation, serving as a universal and unified framework for comprehensive soccer video understanding. + +To summarize, we make the following contributions in this paper: (i) we construct SoccerReplay-1988, the largest and most diverse soccer video dataset to date, featuring videos of 1,988 soccer matches with rich annotations, supported by an automated curation pipeline. This provides a solid foundation for developing robust and comprehensive soccer understanding models; (ii) we present a powerful soccer-specific visual encoder, termed MatchVision, which effectively leverages spatiotemporal information in soccer videos, and can adapt to various tasks such as event classification and commentary generation, serving as a unified framework for soccer understanding; (iii) we establish more comprehensive and challenging benchmarks based on our dataset, enabling more professional evaluation of soccer understanding models; (iv) extensive experiments and ablation studies demonstrate the superiority of our data and model across various downstream tasks, achieving state-of-the-art performance on both existing benchmarks and our newly established ones. We believe this work offers a viable paradigm for future sports video understanding. + +# 2. Related Works + +Sports Understanding [44] is an evolving field that encompasses multiple research topics and integrates diverse data modalities, covering various tasks such as action spotting [10, 15, 16], commentary generation [34, 37, 40, 51, 52, 57], athlete analysis [41, 55], tactical planning [49], sports health [39], and intelligent refereeing [22, 23]. Furthermore, with the rapid development of multimodal large language models (MLLMs), recent efforts [26, 50, 53, 54] have attempted to build more generalized frameworks to uniformly handle a variety of sports understanding tasks. + +Visual-Language Models [1, 27, 28, 38, 58] have exhibited remarkable performance across extensive applications like classification, segmentation, image-text retrieval, and image + +captioning. Recent efforts have ventured into more challenging video understanding [29, 30, 42, 59, 60] tasks, such as temporal alignment [17, 31], dense captioning [5, 56, 62], and audio description [18-20]. However, these efforts typically focus on general scenarios, limiting their adaptability to specific professional fields. Thus, this paper aims to bridge this gap by advancing visual-language models tailored for comprehensive soccer understanding. + +Soccer Game Analysis [9] has primarily focused on tasks such as action spotting [15], replay grounding [10, 61], commentary generation [34, 36, 40], player tracking [8], state reconstruction [43], camera calibration [6, 8, 10] and foul recognition [22, 23], as facilitated by the SoccerNet [7, 10, 14, 15] series datasets, with 500 full-match videos from 2015 to 2017. Unlike existing methods that target designing specific models for distinct tasks, this paper aims to design a unified multi-modal framework that leverages spatiotemporal information within videos, serving as a specialized visual encoder for soccer video understanding. + +# 3. SoccerReplay-1988 Dataset + +To establish a solid foundation for soccer understanding, we construct SoccerReplay-1988, the largest soccer dataset to date. Here, we first outline our data collection details and an overview of the dataset in Sec. 3.1; followed by elaborating on our automated data curation pipeline in Sec. 3.2; lastly, in Sec. 3.3, we present the data statistics and discussion. + +# 3.1. Dataset Collection + +To construct the SoccerReplay-1988 dataset, we have collected untrimmed, full-match videos from the Internet, encompassing a total of 1,988 matches from six European major soccer leagues and championships1, spanning the 2014-15 to 2023-24 seasons. For each match, we acquire textual commentaries with second-level timestamps from a sports text live website2, with part of them annotated with specific event types such as corner and goal. Additionally, we also incorporate extensive metadata, including detailed background information about the games, players, coaches, referees, and teams, providing a solid foundation for future soccer understanding research. + +We partition the SoccerReplay-1988 dataset into train, validation, and test sets, containing 1,488, 250, and 250 full-match videos with diverse and comprehensive annotations, respectively. These sets provide rich training data for downstream tasks, such as event classification and commentary generation, while establishing comprehensive and challenging benchmarks for soccer understanding, as further discussed in subsequent sections. + +![](images/ca6edb824bec182c0601875550b600151eadd1bf8ae1a66dbf292087b9c65f7b.jpg) +Figure 2. Automated Data Curation Pipeline. The collected soccer video data are automatically processed for temporal alignment, event summarization, and anonymization by our curation pipeline. + +# 3.2. Automated Data Curation + +Given the potential noise in raw data, such as irrelevant video content, inaccurate timestamps, and incomplete event annotations, we design an automated data curation pipeline, comprising (i) temporal alignment, (ii) event summarization, and (iii) anonymization, as illustrated in Figure 2. + +Temporal Alignment. Here, we divide match videos into two halves, each starting at kick-off, and adopt the temporal alignment model from MatchTime [40], to synchronize textual commentary timestamps with those of video frames. + +Event Summarization. For samples without event annotations, we leverage LLaMA-3-70B [12] to summarize the events based on textual commentaries. Concretely, we have expanded the event categories from 17 in SoccerNet [10] to 24 types, for finer-grained soccer understanding, for example, categorizing penalties into scored and missed, and integrating modern soccer regulations like VAR. The resulting 24 event labels include: 'corner', 'goal', 'injury', 'own goal', 'penalty', 'penalty missed', 'red card', 'second yellow card', 'substitution', 'start of game (half)', 'end of game (half)', 'yellow card', 'throw in', 'free kick', 'saved by goal-keeper', 'shot off target', 'clearance', 'lead to corner', 'off-side', 'var', 'foul (no card)', 'statistics and summary', 'ball possession', and 'ball out of play'. More details on the used prompts are provided in the Appendix. + +Anonymization. Similar to [34], we extract all person and team entity names from the metadata of SoccerReplay-1988, and replace them in textual commentaries with placeholders, such as "[PLAYER]", "[TEAM]", "[COACH]", and "[REFEREE]", ensuring consistency across tasks. + +Moreover, our data curation pipeline can seamlessly extend to existing datasets, converting the SoccerNet series [10, 34] into our unified data format, termed SoccerNetpro. This expansion further enlarges the standardized + +
Existing Datasets
# GameDuration(h)# Event# Anno.# Com.
SoccerNet-v1 [15]50076476.7k-
SoccerNet-v2 [10]50076417110k-
MatchTime [40]4717161414k37k
GOAL [36]2025.5--8.9k
Our Curated Datasets
# GameDuration(h)# Event# Anno.# Com.
SoccerNet-pro50076424102k37k
SoccerReplay-19881,9883,32324150k150k
Integrated2,4884,08724252k187k
+ +Table 1. Statistics of Soccer Datasets. Our SoccerReplay-1988 significantly surpasses existing datasets in both scale and diversity. Here, # Anno. and # Com. refer to the number of event annotations and textual commentaries, respectively. + +datasets available for soccer understanding tasks. + +# 3.3. Statistics & Discussion + +Dataset Statistics. As shown in Table 1, our dataset encompasses 3,323 hours of footage from 1,988 soccer matches, with an average duration of 100.3 minutes per match. The videos range in resolution from 360p to 720p and frame rates between 25 and 30 FPS. + +For textual annotations, this dataset features approximately 150K commentaries, averaging 76 per match, precisely temporal-aligned by the robust alignment model from MatchTime [40]. These commentaries cover 4,467 unique words, significantly surpassing the 2,873 words in existing datasets [34, 40], greatly enriching textual diversity. Automated event summarization based on these commentaries has yielded about 150K event annotations. Notably, a random sampling of $2\%$ of the data yields $98\%$ manual verification accuracy, ensuring high-quality automated labeling. + +SoccerReplay-test Benchmark. To facilitate a more comprehensive evaluation of soccer understanding models, we integrate 250 matches from SoccerReplay-1988 with 50 matches from the curated SoccerNet-pro, establishing SoccerReplay-test, a more challenging benchmark for event classification and commentary generation. This benchmark features nearly four times larger than existing datasets and comprises finer-grained event labels, richer textual commentaries, and up-to-date soccer regulations. + +Discussion. To summarize, SoccerReplay-1988 exhibits advancements in three aspects: (i) it is the largest soccer video dataset to date, with nearly four times more videos than existing datasets; (ii) it features more professional and diverse annotations, more suitable for fine-grained and comprehensive soccer understanding tasks; (iii) It employs an automated curation pipeline for annotations and is thus scalable to provide a solid data foundation for future research. All data from SoccerReplay-1988, including videos + +![](images/e4c6e688d3460a7c195de0a3737ceca94017c7c0480ff3b70fd2a1036e793496.jpg) +(a) Structure of MatchVision + +![](images/977119086969bdcf69088f68e5878a2f6416a7dae1682c8d3d5c485bda592645.jpg) +(b) MatchVision Pretraining +(c) Downstream Tasks +Figure 3. Overview of MatchVision. (a) The model architecture and its spatiotemporal feature extraction process; (b) Details of visual encoder pretraining, including supervised classification and video-language contrastive learning; (c) Implementation details of specific heads for various downstream tasks, including commentary generation, foul recognition, and event classification. + +and annotations, are open-source for non-commercial use. + +# 4. Method + +In this paper, we aim to develop a soccer-specific visual encoder, MatchVision, tailored for diverse soccer video analysis tasks. We start by outlining our problem formulation in Sec. 4.1. Next, in Sec. 4.2, we detail the architecture of MatchVision. The training procedures are thoroughly discussed in Sec. 4.3. Finally, we describe the configurations for downstream tasks in Sec. 4.4, demonstrating the practical applications and effectiveness of our model. + +# 4.1. Problem Formulation + +In this work, we tackle the challenge of analyzing soccer video segments, denoted as $\mathcal{V} \in \mathbb{R}^{T \times 3 \times H \times W}$ . Our goal is to utilize the visual encoder $(\Phi_{\mathrm{MatchVision}})$ to extract spatiotemporal features from these segments, which are then processed by multiple task-specific heads, formulated as: + +$$ +\mathbf {E}, \mathbf {C}, \mathbf {F} = \Psi (\Phi_ {\text {M a t c h V i s i o n}} (\mathcal {V})) +$$ + +Here, $\Psi = \{\Psi_{\mathrm{cls}}, \Psi_{\mathrm{Cmt}}, \Psi_{\mathrm{Foul}}\}$ represents the task-specific heads, with $\mathbf{E}$ , $\mathbf{C}$ , and $\mathbf{F}$ denoting the output event types, textual commentaries, and foul types, respectively. This unified framework effectively learns relevant spatiotemporal features, and enables seamless integration across various downstream tasks for comprehensive soccer understanding. + +# 4.2. Architecture + +MatchVision comprises three key components: (i) Token Embedding, (ii) Spatiotemporal Attention Block, and (iii) + +Aggregation Layer, as depicted in Figure 3. + +Token Embedding. In accordance with the convention in Vision Transformer [11], each frame $(\mathcal{I}_i)$ from the video segment $(\mathcal{V} = \{\mathcal{I}_1, \mathcal{I}_2, \dots, \mathcal{I}_T\})$ is divided into $M$ non-overlapping patches of size $P \times P$ that span the entire frame. + +These patches are flattened into vectors $(\mathbf{x}_i^p)$ , where $p$ and $i$ denote the spatial and temporal positions, respectively. Each vector is transformed via an embedding layer $(\Phi_{\mathrm{Emb}})$ into a token vector of size $\mathbb{R}^{1\times D}$ , and then added with a spatial position embedding $(\mathbf{e}_{\mathrm{s}}^{\mathrm{pos}}\in \mathbb{R}^{M\times D})$ . Subsequently, we concatenate a [c1s] token along with each frame. Finally, a temporal positional embedding $(\mathbf{e}_{\mathrm{t}}^{\mathrm{pos}}\in \mathbb{R}^{T\times D})$ is added across features of all frames, as formulated below: + +$$ +\begin{array}{l} \mathbf {y} _ {i} = \left[ \mathbf {x} _ {i} ^ {\mathrm {c l s}}, \Phi_ {\mathrm {E m b}} \left(\left[ \mathbf {x} _ {i} ^ {1}, \dots , \mathbf {x} _ {i} ^ {M} \right]\right) + \mathbf {e} _ {\mathrm {s}} ^ {\mathrm {p o s}} \right] \\ \mathbf {z} = \left[ \mathbf {y} _ {1}, \dots , \mathbf {y} _ {T} \right] + \mathbf {e} _ {\mathrm {t}} ^ {\text {p o s}} \\ \end{array} +$$ + +Here, $[\cdot ,\cdot ]$ denotes concatenation, and $\mathbf{y}_i\in \mathbb{R}^{(M + 1)\times D}$ represents the frame-wise features. The embedded features (z) will then serve as input for spatiotemporal attention blocks. Spatiotemporal Attention Block. Similar to TimeSformer [3], we utilize interleaved temporal and spatial attention to integrate spatiotemporal information in soccer videos. Concretely, each spatiotemporal attention block comprises a temporal self-attention layer and a spatial self-attention layer, i.e., $\phi_t(\cdot)$ and $\phi_s(\cdot)$ , respectively. + +Given a video feature $(\mathbf{z} \in \mathbb{R}^{T \times (M + 1) \times D})$ , we alternate temporal and spatial attention: temporal attention facilitates interactions among tokens at the same spatial positions across distinct frames, while spatial attention enables interactions among tokens within the same frame. Residual con + +nections are employed in each layer. After passing through a total of $\mathbf{K}$ spatiotemporal attention blocks, the resulting feature $(\mathcal{F})$ captures both intra-frame and inter-frame relationships, i.e., $\mathcal{F} = [\phi_s(\phi_t(\mathbf{z}))]^{\mathbf{K}} \in \mathbb{R}^{T \times (M + 1) \times D}$ . + +Aggregation Layer. To obtain video-level features, we employ an aggregation layer on the frame-wise spatiotemporal features. Specifically, for the $i$ -th frame, we utilize spatial self-attention to aggregate information into its [cls] token, denoted as $\hat{\mathcal{F}}_i^{\mathrm{cls}} = \Phi_{\mathrm{Agg}}(\mathcal{F}_i)$ . Concatenating the [cls] tokens of all frames yields the final video feature $(\mathcal{F}_{\mathcal{V}})$ , that effectively encapsulates spatiotemporal characteristics of soccer video segments, thus enabling it to be applicable for various downstream soccer understanding tasks. This process can be formulated as: + +$$ +\mathcal {F} _ {\mathcal {V}} = \Phi_ {\mathrm {M a t c h V i s i o n}} (\mathcal {V}) = [ \hat {\mathcal {F}} _ {1} ^ {\mathrm {c l s}}, \dots , \hat {\mathcal {F}} _ {T} ^ {\mathrm {c l s}} ] \in \mathbb {R} ^ {T \times D} +$$ + +# 4.3. MatchVision Pretraining + +In this part, we aim to pretrain the visual encoder with triplet samples $(\{\mathcal{V},\mathbf{E},\mathbf{C}\})$ , comprising videos, event labels, and textual commentaries. Concretely, we investigate two distinct pretraining strategies: supervised classification and video-language contrastive learning. + +Supervised Classification. One way to pretrain the visual encoder is supervised learning on event classification. To be specific, the extracted visual features $(\mathcal{F}_{\mathcal{V}})$ are aggregated by a temporal self-attention layer into a learnable [cls] token, denoted as $\mathcal{F}_{\mathcal{V}}^{\mathrm{cls}}$ , similar to the spatial-wise aggregation mentioned above. This token is then fed into a linear classifier, and trained with a cross-entropy loss for event classification. The objective is denoted as $\mathcal{L}_{\mathrm{sup}}$ . + +Video-Language Contrastive Learning. As an alternative, we can also pretrain our visual encoder with video-text contrastive learning. Specifically, we adopt simple average pooling on the video feature to get the aggregated visual feature $(\mathcal{F}_{\mathcal{V}}^{\mathrm{Avg}})$ , and encode the textual commentary (C) with a text encoder $(\Phi_{\mathrm{Text}})$ . We train the model with sigmoid loss ( $\mathcal{L}_{\mathrm{sigmoid}}$ ), as used in SigLIP [58]. Note that, some video clips may have highly similar commentaries, for example, 'start of the game', we treat the commentaries with high similarity in the same batch as positive samples when calculating loss functions. This can be expressed as follows: + +$$ +\mathcal {L} _ {\text {c o n t r a}} = \mathcal {L} _ {\text {s i g m o i d}} \left(\mathcal {F} _ {\mathcal {V}} ^ {\text {A v g}}, \Phi_ {\text {T e x t}} (\mathbf {C})\right) +$$ + +# 4.4. Downstream Tasks + +After the pretraining mentioned above, MatchVision can now serve as a versatile visual encoder, to map the soccer video segments into visual features $(\mathcal{F}_{\mathcal{V}})$ , for training task-specific heads $\Psi = \{\Psi_{\mathrm{cls}}, \Psi_{\mathrm{Cmt}}, \Psi_{\mathrm{Foul}}\}$ across different downstream tasks, including: (i) event classification, (ii) commentary generation, and (iii) foul recognition. + +Event Classification. Similar to supervised classification above, we concatenate a learnable [cls] token to aggregate frame-wise visual features via temporal self-attention. This token is then fed into a linear classifier for event classification. The event classification head $(\Psi_{\mathrm{cls}})$ is trained with a cross-entropy loss while freezing the visual encoder. + +Commentary Generation. We follow the paradigm in MatchTime [40] to generate anonymized textual commentary for soccer video clips. Concretely, the commentary generation head $(\Psi_{\mathrm{Cmt}})$ employs a Perceiver [25] aggregator to consolidate visual features, which are then projected by a trainable MLP, serving as prefix embeddings for a large language model (LLM). Subsequently, an off-the-shelf LLM decodes these embeddings into textual commentary. We adopt the negative log-likelihood loss, commonly used for auto-regressive next-token prediction. + +Foul Recognition. As outlined in [22], the foul recognition task takes multi-view videos from the same scene as inputs, with each sample annotated with a foul class (8 types) and severity (4 levels). We encode these multi-view videos with MatchVision, and aggregate the extracted features into a single feature vector, via either max or average pooling, following the common practice. Subsequently, the foul recognition head $(\Psi_{\mathrm{Foul}})$ employs a shared MLP and two task-specific linear classifiers, to predict foul type and severity, respectively. Similar to event classification, we use the combination of cross-entropy losses on the foul type and severity classification to jointly train $\Psi_{\mathrm{Foul}}$ . + +Discussion. Pretraining MatchVision on large-scale soccer data equips it with substantial domain-specific knowledge, enabling it to serve as a universal visual encoder adaptable to various downstream soccer understanding tasks. + +# 5. Experiments + +This section begins with implementation details in Sec. 5.1; followed by quantitative evaluations across downstream tasks in Sec. 5.2; then, we conduct ablation studies on our SoccerReplay-test benchmark to analyze the effectiveness of the proposed dataset and model in Sec. 5.3; finally, we provide qualitative results for comparison in Sec. 5.4. + +# 5.1. Implementation Details + +In our experiments, video segments are sampled at 1FPS around annotated timestamps, capturing a 30-second window for each sample. Frames are resized to $224 \times 224$ pixels as inputs. We initialize the embedding layer, spatial attention layers, aggregation layer, and text encoder of MatchVision with pretrained weights from SigLIP Base-16 [58] and adopt LLaMA-3 (8B) [12] as the off-the-shelf LLM decoder for commentary generation. All experiments are conducted on $4 \times$ Nvidia H800 GPUs with the AdamW [33] optimizer. Next, we elaborate on the training and evaluation details + +
Visual EncoderDatasetClassification (%)Commentary
SNMTSRAcc.@1Acc.@3Acc.@5B@1B@4MR-LC
Off-the-shelf Models
I3D [4]XXX45.482.593.226.775.5724.1723.1218.73
C3D [45]XXX47.885.195.028.136.6424.5224.2327.88
ResNet [21]XXX47.284.694.427.346.5724.7224.4327.29
CLIP [38]XXX48.585.595.226.256.5124.2724.7528.17
InternVideo [48]XXX49.987.095.927.126.5425.0224.8229.90
SigLIP [58]XXX50.286.795.627.856.9825.1625.0331.38
Pretrain with Supervised Classification
Baidu [61]✓XX56.491.997.331.208.8826.5626.6138.93
SigLIP✓XX55.989.694.928.517.3925.9625.9435.71
SigLIP✓✓✓57.991.797.530.958.5625.7926.1738.24
MatchVision✓XX82.596.698.829.457.9226.0126.2136.15
MatchVision✓✓✓84.097.399.231.059.0626.9427.9342.20
Pretrain with Visual-Language Contrastive Learning
SigLIPX✓X55.488.897.028.727.7225.9126.1732.27
SigLIPX✓✓66.893.798.630.358.1226.0526.3839.41
MatchVisionX✓X58.989.097.130.337.9725.4826.3333.87
MatchVisionX✓✓67.993.998.631.949.1226.2427.5640.76
Pretrain with Hybrid Supervised-Contrastive Training
SigLIP✓✓X71.294.598.728.637.8225.7425.3534.09
SigLIP✓✓✓67.193.298.130.718.7826.2626.7441.82
MatchVision✓✓X76.496.099.030.658.3325.2826.3137.23
MatchVision✓✓✓80.197.199.133.589.1426.8228.2144.18
+ +Table 2. Quantitative Results on Event Classification and Commentary Generation. Here, SN, MT, and SR represent finetuning with curated SoccerNet-v2 [10], MatchTime [40], and SoccerReplay-1988, respectively. B, M, R-L, and C refer to BLEU, METEOR, ROUGE-L, and CIDEr metrics, respectively. Within each unit, we denote the best performance in RED and the second-best performance in BLUE. + +about visual encoder pretraining and downstream tasks. + +Visual Encoder Pretraining. For both pretraining strategies, we use a batch size of 40 for 15 epochs. The learning rate for all randomly initialized modules, including the temporal attention layers, aggregator layer, and linear classifier, is set to $1 \times 10^{-4}$ . Meanwhile, the learning rate for modules initialized with pretrained parameters (including the text encoder) is set to $5 \times 10^{-5}$ . In contrastive training, we adopt a multi-positive strategy where each textual commentary, based on its event label, considers closely related categories (e.g. "start of game" and "offside") as positive samples. + +Downstream Tasks. In all downstream tasks, unless otherwise specified, we use the frozen visual encoder for feature extraction and only train the task-specific heads with a learning rate of $1 \times 10^{-4}$ for 30 epochs. The batch sizes for event classification, commentary generation, and foul recognition are set to 40, 32, and 8, respectively. We adopt specific evaluation metrics for these three tasks: (i) For event classification, we use the top-1/3/5 classification accuracy; (ii) For commentary generation, we employ several commonly-used language evaluation metrics, including BLEU [35], METEOR [2], ROUGE-L [32], and + +CIDEr [47]; (iii) For foul recognition, we follow the common practice, and report top-1/2 and top-1 accuracy for the foul type and severity classification, respectively. + +Benchmarks & Baselines. To ensure fair and reliable comparisons with existing work, we evaluate event classification (24 types) on 100 matches from curated SoccerNetv2 [10] test set; commentary generation on 49 matches from SN-Caption-test-align benchmark manually aligned in [40]; and foul recognition on MVFoul [22]. We consider various baselines: for the first two tasks, this includes off-the-shelf general visual encoders such as ResNet [21], C3D [45], I3D [4], CLIP [38], SigLIP [58], and InternVideo [48], along with Baidu [61] and SigLIP finetuned with soccer-specific data. For foul recognition, we follow previous work [22, 23] and adopt ResNet [21], R(2+1)D [46], and MViT [13] jointly finetuned with classifiers, as baselines. + +# 5.2.Quantitative Evaluation + +As depicted in Table 2, we draw two observations on event classification and commentary generation: (i) visual encoders trained on soccer data substantially outperform off-the-shelf general encoders (ResNet, C3D, I3D, CLIP, and + +
Visual EncoderFoul ClassSeverity
BackboneTrainAgg.Acc.@1Acc.@2Acc.@1
ResNet [21]✓Mean0.310.560.34
Max0.320.600.32
R(2+1)D [46]✓Mean0.310.550.34
Max0.320.560.39
MViT [13]✓Mean0.400.650.38
Max0.470.690.43
MatchVision×Mean0.440.530.58
Max0.350.700.46
+ +InternVideo), underscoring the necessity of building specialized models for soccer understanding; (ii) almost all visual encoders, across all training settings, benefit from SoccerReplay-1988, emphasizing the value of constructing large-scale, high-quality data for soccer understanding. Next, we will delve into each task to discuss the results. + +Event Classification. With identical training strategies and data, MatchVision considerably outperforms other methods in classification accuracy, demonstrating the superiority of its architecture, which effectively leverages spatiotemporal features within soccer videos. For example, MatchVision achieves a Top-1 accuracy of $82.5\%$ , significantly surpassing SigLIP's $55.9\%$ under the same training conditions. Moreover, models trained via supervised classification excel others, primarily because the pre-training task shares the same objectives as the downstream event classification task. + +Commentary Generation. Visual encoders trained with visual-language contrastive learning exhibit better commentary generation performance than those trained with supervised classification, as this strategy better captures correlations between visual and textual features. Additionally, while MatchVision trained solely on SoccerNet slightly underperforms Baidu [61], incorporating SoccerReplay-1988 enables it to outperform on most metrics. This demonstrates that MatchVision can take advantage of large-scale datasets. Finally, a hybrid training approach, starting with supervised classification followed by visual-language contrastive learning, enables MatchVision to achieve optimal performance. This indicates that learning coarse-grained tasks such as classification provides a foundation for fine-grained tasks like commentary generation, and fully leveraging data unlocks the potential of soccer understanding. + +Foul Recognition. As demonstrated in Table 3, MatchVision achieves performance comparable to jointly finetuned state-of-the-art methods in foul recognition, even with a frozen visual encoder. This highlights that MatchVision effectively learns substantial knowledge from large-scale soc + +Table 3. Quantitative Results on Multi-view Foul Recognition. Our frozen MatchVision encoder can achieve comparable performance with other jointly finetuned visual encoders. + +
PretrainClassification(%)
Sup.Contra.SRAcc.@1Acc.@3Acc.@5
✓XX62.6783.0089.81
✓X✓68.0386.9092.38
X✓X46.9775.5385.85
X✓✓57.4183.1391.00
✓✓X56.8680.3088.09
✓✓✓63.5985.2191.63
+ +Table 4. Ablations on Event Classification. We explore the impact of various training settings of our MatchVision encoder on the SoccerReplay-test benchmark. Here, Sup., Contra., and SR refer to supervised classification, visual-language contrastive learning, and the SoccerReplay-1988 dataset, respectively. + +cer data and adapts seamlessly to downstream tasks. Comparisons with additional baselines from the SoccerNet foul recognition challenges [9] are provided in the Appendix. + +# 5.3. Ablation Studies + +We conduct ablation experiments on event classification and commentary generation using our SoccerReplay-test benchmark. These experiments validate the effectiveness of our proposed dataset and model, while establishing a baseline for future evaluations on this benchmark. + +Event Classification. We evaluate event classification on 300 matches from our SoccerReplay-test benchmark using the MatchVision visual encoder pretrained with various strategies. Features are extracted by MatchVision and processed with a learnable aggregation layer and a linear classifier. The default training set is our curated SoccerNet-pro. As shown in Table 4, integrating SoccerReplay-1988 for training results in significant performance improvements across all pretraining strategies, yielding the significance of our dataset. Additionally, supervised classification outperforms visual-language contrastive learning and hybrid pretraining. This is due to its closer alignment with downstream event classification task, and the scale of event annotations is far larger than that of textual commentaries, further confirming the substantial benefits of data scaling for boosting soccer understanding. + +Commentary Generation. With the pretrained MatchVision encoder, we train the commentary generation head on the MatchTime [40] and SoccerReplay-1988 datasets using various training strategies. By default, only the Perceiver [25] aggregation layer and projection layer within the head are trained. For joint training with the LLM decoder, considering computational costs, we incorporate LoRA [24] layers while freezing the original LLM layers. As shown in Table 5, incorporating SoccerReplay-1988 significantly improves performance on all metrics, confirming substantial advantages of our proposed dataset. This performance + +![](images/9d80294ea391f74578532a7a932bfd89ac9a197638371fc49f6dce60d8272ed9.jpg) +Soccer Videos + +![](images/9da4b340fa5e701f874da4eab9b7b914104a894fefb77e3bfec73dbc28c6aaee.jpg) +Events + +![](images/dfe11a125f3ab53dc3d4ff9c58672ea87bd1c11d4c1cb608159d5405b50da324.jpg) +Commentaries + +![](images/8cdd4dbbecb8a9ec6e3fdaac06188a5bdb8d9593f88138abc2c3f127dc7a9875.jpg) + +![](images/68732df897e97929bec2c06a6cd4a0e00c3c827ef4c30cf5c5ca0c4f68c81307.jpg) + +![](images/fc9c2bb97514f72b61de963098cb78a266efe3146a378467055940157f707b37.jpg) + +![](images/52ed70c15dc4709676df0e4b456ec14925d160a95ba8aee8054e3fadab030edd.jpg) + +![](images/8276801c2dc956cf715cbb6e35392feb9ec6c5c1e35246581e3815e139c44262.jpg) + +![](images/516894ea4acf31fed702636a2327972b70c6c0686e548ec9008c7942bb866031.jpg) + +![](images/270fd650638147ed4738f17a034c436bce1a5e856c8100fe1265415bc8ef1e7e.jpg) + +![](images/b0de7aef3c8941434bfbb3eea98b327a676dacfc6f7f83d3791ac06e9b00f0f7.jpg) + +![](images/a380b47f26580073f501fca6a80af2d27b0dff7e9f36d75a2471d72d0c93d535.jpg) + +![](images/3eefcf41d5f78e47d7d847d40c2c7378f9003d67dcc7f9abc8a12c41c3846a2e.jpg) + +![](images/3b4d60d20742533f20aab26bf153e9159875fb7e3fc21dc2d1ab6a2e9a7e5be3.jpg) + +![](images/6bdbf1b0e4e1e262d7f85f385ddf1be5a1f0c0e018b967e3eade70b46797fd21.jpg) + +![](images/c4b3b98a4be399f96a73b58101fd17cc5b827266907bce570f39aa35c3e2d7ff.jpg) +Figure 4. Qualitative Results for Event Classification and Commentary Generation. Here, "w/o SR" and "w/ SR" indicate models trained without and with the SoccerReplay-1988 dataset, respectively. Incorporating SoccerReplay-1988 improves event classification accuracy. Moreover, this enriched training data enables the model to demonstrate several advantages in commentary generation: (a) more detailed descriptions, (b) greater linguistic variety, (c) higher event depiction accuracy, (d) better adherence to updated rules, and (e) improved specificity in scenario response. + +![](images/9b8addab99c97babf3f33b9d8f1d4cb75828a7be35583514c63ea2e65e97bcee.jpg) + +![](images/39aeeb1cf36f84a1d0fe1d9adf45176e87fd7b4d6118c5b90ca344994a7a1b2b.jpg) + +
TrainableCommentary Metrics
VLB@1B@4MR-LC
Trained on MatchTime
XX21.653.2721.0217.7912.90
✓X27.627.0224.0323.5130.77
X✓27.046.4124.1523.8831.91
✓✓27.496.9624.5023.3330.81
Trained on MatchTime & SoccerReplay-1988
XX24.174.0920.5120.7015.70
✓X28.988.3924.4525.3545.85
X✓27.547.7624.5024.7042.79
✓✓29.218.2225.2525.5443.18
+ +Table 5. Ablations on Commentary Generation. We investigate the impact of different training strategies and datasets on MatchVision using the SoccerReplay-test benchmark. 'V' and 'L' denote the visual encoder and the LLM decoder, respectively. + +gap also reflects the challenges of our established benchmark, which features diverse vocabulary, richer semantics, and updated soccer rules. Additionally, jointly finetuning the visual encoder and the LLM decoder provides a feasible approach for further improvements. + +# 5.4. Qualitative Comparisons + +As depicted in Figure 4, we present qualitative results of MatchVision on the SoccerReplay-test benchmark, comparing models pretrained with and without SoccerReplay-1988. For event classification, incorporating our data improves accuracy, and even in misclassified cases, the results + +remain contextually relevant. For commentary generation, hybrid training on SoccerReplay-1988 enables MatchVision to produce richer, more detailed textual commentary, reflecting a deeper understanding of soccer dynamics. More qualitative results are available in the Appendix. + +# 6. Conclusion + +In this paper, we establish a unified, scalable multi-modal framework for soccer understanding. Specifically, we introduce SoccerReplay-1988, the largest and most comprehensive soccer video dataset to date, annotated by an automated curation pipeline. This provides a solid foundation for developing soccer understanding models and serves as a more challenging benchmark. Built upon this, we develop MatchVision, an advanced soccer-specific visual encoder, which effectively leverages spatiotemporal information within soccer videos and can be applied to various tasks such as event classification and commentary generation. Extensive experiments demonstrate the superiority of our model, with MatchVision achieving state-of-the-art performance on both existing benchmarks and our newly established one. We believe this work will set a viable, universal paradigm for future research in sports understanding. + +# Acknowledgments + +This work is funded by National Key R&D Program of China (No.2022ZD0161400). + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, pages 23716-23736, 2022. 2 +[2] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, pages 65-72, 2005. 6 +[3] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In Proceedings of the International Conference on Machine Learning, page 4, 2021. 2, 4 +[4] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 6 +[5] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, and Sergey Tulyakov. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2 +[6] Anthony Cioppa, Adrien Deliege, Floriane Magera, Silvio Giancola, Olivier Barnich, Bernard Ghanem, and Marc Van Droogenbroeck. Camera calibration and player localization in soccernet-v2 and investigation of their representations for action spotting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 4537-4546, 2021. 2 +[7] Anthony Cioppa, Adrien Deliege, Silvio Giancola, Bernard Ghanem, and Marc Van Droogenbroeck. Scaling up soccer-net with multi-view spatial localization and re-identification. Scientific data, 9(1):355, 2022. 1, 2 +[8] Anthony Cioppa, Silvio Giancola, Adrien Deliege, Le Kang, Xin Zhou, Zhiyu Cheng, Bernard Ghanem, and Marc Van Droogenbroeck. Soccernet-tracking: Multiple object tracking dataset and benchmark in soccer videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 3491-3502, 2022. 2 +[9] Anthony Cioppa, Silvio Giancola, Vladimir Somers, Victor Joos, Floriane Magera, Jan Held, Seyed Abolfazl Ghasemzadeh, Xin Zhou, Karolina Seweryn, Mateusz Kowalczyk, Zuzanna Mroz, Szymon Lukasik, Michal Halon, Hassan Mkhallati, Adrien Deliege, Carlos Hinojosa, Karen Sanchez, Amir M. Mansourian, Pierre Miralles, Olivier Barnich, Christophe De Vleeschouwer, Alexandre Alahi, Bernard Ghanem, Marc Van Droogenbroeck, Adam Gorski, Albert Clapes, Andrei Boiarov, Anton Afanasiev, Artur Xarles, Atom Scott, ByoungKwon Lim, Calvin Yeung, Cristian Gonzalez, Dominic Rufenacht, Enzo Pacilio, Fabian Deuser, Faisal Sami Altawijri, Francisco Cachon, HanKyul Kim, Haobo Wang, Hyeonmin Choe, Hyunwoo J Kim, Il-Min + +Kim, Jae-Mo Kang, Jamshid Tursunboev, Jian Yang, Jihwan Hong, Jimin Lee, Jing Zhang, Junseok Lee, Kexin Zhang, Konrad Habel, Licheng Jiao, Linyi Li, Marc Gutierrez-Perez, Marcelo Ortega, Menglong Li, Milosz Lopatto, Nikita Kasatkin, Nikolay Nemtsev, Norbert Oswald, Oleg Udin, Pavel Kononov, Pei Geng, Saad Ghazai Alotaibi, Sehyung Kim, Sergei Ulasen, Sergio Escalera, Shanshan Zhang, Shuyuan Yang, Sunghwan Moon, Thomas B. Moeslund, Vasyl Shandyba, Vladimir Golovkin, Wei Dai, WonTaek Chung, Xinyu Liu, Yongqiang Zhu, Youngseo Kim, Yuan Li, Yuting Yang, Yuxuan Xiao, Zehua Cheng, and Zhihao Li. Soccernet 2024 challenges results. arXiv preprint arXiv:2409.10587, 2024. 2, 7 +[10] Adrien Deliege, Anthony Cioppa, Silvio Giancola, Meisam J Seikavandi, Jacob V Dueholm, Kamal Nasrollahi, Bernard Ghanem, Thomas B Moeslund, and Marc Van Droogenbroeck. Soccernet-v2: A dataset and benchmarks for holistic understanding of broadcast soccer videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 4508-4519, 2021. 1, 2, 3, 6 +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations, 2021. 4 +[12] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 3, 5 +[13] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6824-6835, 2021. 6, 7 +[14] Sushant Gautam, Mehdi Houshmand Sarkhoosh, Jan Held, Cise Midoglu, Anthony Cioppa, Silvio Giancola, Vajira Thambawita, Michael A Riegler, Pål Halvorsen, and Mubarak Shah. Soccernet-echoes: A soccer game audio commentary dataset. arXiv preprint arXiv:2405.07354, 2024. 2 +[15] Silvio Giancola, Mohieddine Amine, Tarek Dghaily, and Bernard Ghanem. Soccernet: A scalable dataset for action spotting in soccer videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 1711-1721, 2018. 1, 2, 3 +[16] Xiaofan Gu, Xinwei Xue, and Feng Wang. Fine-grained action recognition on a novel basketball dataset. In International Conference on Acoustics, Speech, and Signal Processing, pages 2563-2567, 2020. 2 +[17] Tengda Han, Weidi Xie, and Andrew Zisserman. Temporal alignment networks for long-term video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2906-2916, 2022. 2 +[18] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. Autoad: Movie description in + +context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2 +[19] Tengda Han, Max Bain, Arsha Nagrani, Gul Varol, Weidi Xie, and Andrew Zisserman. Autoad ii: The sequel-who, when, and what in movie audio description. In Proceedings of the International Conference on Computer Vision, 2023. +[20] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. Autoad iii: The prequel - back to the pixels. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2 +[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6, 7 +[22] Jan Held, Anthony Cioppa, Silvio Giancola, Abdullah Hamdi, Bernard Ghanem, and Marc Van Droogenbroeck. Vars: Video assistant referee system for automated soccer decision making from multiple views. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 5085-5096, 2023. 2, 5, 6 +[23] Jan Held, Hani Itani, Anthony Cioppa, Silvio Giancola, Bernard Ghanem, and Marc Van Droogenbroeck. X-vars: Introducing explainability in football refereeing with multimodal large language models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 3267-3279, 2024. 2, 6 +[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In Proceedings of the International Conference on Learning Representations, 2022. 7 +[25] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In Proceedings of the International Conference on Machine Learning, pages 4651-4664, 2021. 5, 7 +[26] Haopeng Li, Andong Deng, Qiuhong Ke, Jun Liu, Hossein Rahmani, Yulan Guo, Bernt Schiele, and Chen Chen. Sports-qa: A large-scale video question answering benchmark for complex and professional sports. arXiv preprint arXiv:2401.01505, 2024. 2 +[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the International Conference on Machine Learning, pages 12888-12900, 2022. 2 +[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the International Conference on Machine Learning, pages 19730–19742, 2023. 2 +[29] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, Limin Wang, and Yu Qiao. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 2 + +[30] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In Proceedings of the European Conference on Computer Vision, 2024. 2 +[31] Zeqian Li, Qirui Chen, Tengda Han, Ya Zhang, Yanfeng Wang, and Weidi Xie. Multi-sentence grounding for long-term instructional video. In Proceedings of the European Conference on Computer Vision, 2024. 2 +[32] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, 2004. 6 +[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Proceedings of the International Conference on Learning Representations, 2019. 5 +[34] Hassan Mkhallati, Anthony Cioppa, Silvio Giancola, Bernard Ghanem, and Marc Van Droogenbroeck. Socernetc caption: Dense video captioning for soccer broadcasts commentaries. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 5074-5085, 2023. 1, 2, 3 +[35] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Association for Computational Linguistics, pages 311-318, 2002. 6 +[36] Ji Qi, Jifan Yu, Teng Tu, Kunyu Gao, Yifan Xu, Xinyu Guan, Xiaozhi Wang, Bin Xu, Lei Hou, Juanzi Li, et al. Goal: A challenging knowledge-grounded video captioning benchmark for real-time soccer commentary generation. In Proceedings of the ACM International Conference on Information and Knowledge Management, pages 5391-5395, 2023. 1, 2, 3 +[37] Mengshi Qi, Yunhong Wang, Annan Li, and Jiebo Luo. Sports video captioning via attentive motion representation and group relationship modeling. IEEE Transactions on Circuits and Systems for Video Technology, 2019. 2 +[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, 2021. 2, 6 +[39] Prem N Ramkumar, Bryan C Luu, Heather S Haeberle, Jaret M Karnuta, Benedict U Nwachukwu, and Riley J Williams. Sports medicine and artificial intelligence: a primer. The American Journal of Sports Medicine, 50(4): 1166-1174, 2022. 2 +[40] Jiayuan Rao, Haoning Wu, Chang Liu, Yanfeng Wang, and Weidi Xie. Matchtime: Towards automatic soccer game commentary generation. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2024. 1, 2, 3, 5, 6, 7 +[41] Dian Shao, Yue Zhao, Bo Dai, and Dahua Lin. Finegym: A hierarchical video dataset for fine-grained action understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2616-2625, 2020. 2 +[42] Yan Shu, Zheng Liu, Peitian Zhang, Minghao Qin, Junjie Zhou, Zhengyang Liang, Tiejun Huang, and Bo Zhao. + +Video-xl: Extra-long vision language model for hour-scale video understanding. arXiv preprint arXiv:2409.14485, 2024. 2 +[43] Vladimir Somers, Victor Joos, Anthony Cioppa, Silvio Giancola, Seyed Abolfazl Ghasemzadeh, Floriane Magera, Baptiste Standaert, Amir M Mansourian, Xin Zhou, Shohreh Kasaei, et al. Soccernet game state reconstruction: End-to-end athlete tracking and identification on a minimap. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 3293-3305, 2024. 2 +[44] Graham Thomas, Rikke Gade, Thomas B Moeslund, Peter Carr, and Adrian Hilton. Computer vision for sports: Current applications and research topics. Computer Vision and Image Understanding, 159:3-18, 2017. 2 +[45] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the International Conference on Computer Vision, pages 4489-4497, 2015. 6 +[46] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 6, 7 +[47] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4566-4575, 2015. 6 +[48] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 6 +[49] Zhe Wang, Petar Velicković, Daniel Hennes, Nenad Tomasev, Laurel Prince, Michael Kaisers, Yoram Bachrach, Romuald Elie, Li Kevin Wenliang, Federico Piccinini, et al. Tacticai: an ai assistant for football tactics. Nature Communications, 15(1):1-13, 2024. 1, 2 +[50] Dekun Wu, He Zhao, Xingce Bao, and Richard P Wildes. Sports video analysis on large-scale data. In Proceedings of the European Conference on Computer Vision, 2022. 2 +[51] Zeyu Xi, Ge Shi, Xuefen Li, Junchi Yan, Zun Li, Lifang Wu, Zilin Liu, and Liang Wang. A simple yet effective knowledge guided method for entity-aware video captioning on a basketball benchmark. Neurocomputing, 2025. 2 +[52] Zeyu Xi, Ge Shi, Haoying Sun, Bowen Zhang, Shuyi Li, and Lifang Wu. Eika: Explicit & implicit knowledge-augmented network for entity-aware sports video captioning. Expert Systems with Applications, 2025. 2 +[53] Haotian Xia, Zhengbang Yang, Yuqing Wang, Rhys Tracy, Yun Zhao, Dongdong Huang, Zezhi Chen, Yan Zhu, Yuanfang Wang, and Weining Shen. Sportqa: A benchmark for sports understanding in large language models. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics, 2024. 2 +[54] Haotian Xia, Zhengbang Yang, Junbo Zou, Rhys Tracy, Yuqing Wang, Chi Lu, Christopher Lai, Yanjun He, Xun + +Shao, Zhuoqing Xie, et al. Sportu: A comprehensive sports understanding benchmark for multimodal large language models. In Proceedings of the International Conference on Learning Representations, 2025. 2 +[55] Jinglin Xu, Yongming Rao, Xumin Yu, Guangyi Chen, Jie Zhou, and Jiwen Lu. Finediving: A fine-grained dataset for procedure-aware action quality assessment. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2949-2958, 2022. 2 +[56] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 2 +[57] Huanyu Yu, Shuo Cheng, Bingbing Ni, Minsi Wang, Jian Zhang, and Xiaokang Yang. Fine-grained video captioning for sports narrative. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2 +[58] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the International Conference on Computer Vision, pages 11975-11986, 2023. 2, 5, 6 +[59] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2023. 2 +[60] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024. 2 +[61] Xin Zhou, Le Kang, Zhiyu Cheng, Bo He, and Jingyu Xin. Feature combination meets attention: Baidu soccer embeddings and transformer based temporal detection. arXiv preprint arXiv:2106.14447, 2021. 2, 6, 7 +[62] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2 \ No newline at end of file diff --git a/2025/Towards Universal Soccer Video Understanding/images.zip b/2025/Towards Universal Soccer Video Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dd279266a95cb081e8c7158a2e381f7b75cc8734 --- /dev/null +++ b/2025/Towards Universal Soccer Video Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:662e11fa11125dd224724691dcbdc952fd4f3369d2148b293cd50500324c8547 +size 721991 diff --git a/2025/Towards Universal Soccer Video Understanding/layout.json b/2025/Towards Universal Soccer Video Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cc8fde35eb613c4fd06a2019bd71a857808db6c3 --- /dev/null +++ b/2025/Towards Universal Soccer Video Understanding/layout.json @@ -0,0 +1,9044 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 157, + 103, + 454, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 103, + 454, + 121 + ], + "spans": [ + { + "bbox": [ + 157, + 103, + 454, + 121 + ], + "type": "text", + "content": "Towards Universal Soccer Video Understanding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 142, + 516, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 142, + 516, + 208 + ], + "spans": [ + { + "bbox": [ + 93, + 142, + 516, + 208 + ], + "type": "text", + "content": "Jiayuan Rao" + }, + { + "bbox": [ + 93, + 142, + 516, + 208 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 93, + 142, + 516, + 208 + ], + "type": "text", + "content": ", Haoning Wu" + }, + { + "bbox": [ + 93, + 142, + 516, + 208 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 93, + 142, + 516, + 208 + ], + "type": "text", + "content": ", Hao Jiang" + }, + { + "bbox": [ + 93, + 142, + 516, + 208 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 93, + 142, + 516, + 208 + ], + "type": "text", + "content": ", Ya Zhang" + }, + { + "bbox": [ + 93, + 142, + 516, + 208 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 93, + 142, + 516, + 208 + ], + "type": "text", + "content": ", Yanfeng Wang" + }, + { + "bbox": [ + 93, + 142, + 516, + 208 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 93, + 142, + 516, + 208 + ], + "type": "text", + "content": ", Weidi Xie" + }, + { + "bbox": [ + 93, + 142, + 516, + 208 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 93, + 142, + 516, + 208 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 93, + 142, + 516, + 208 + ], + "type": "text", + "content": "School of Artificial Intelligence, Shanghai Jiao Tong University, China \n" + }, + { + "bbox": [ + 93, + 142, + 516, + 208 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 93, + 142, + 516, + 208 + ], + "type": "text", + "content": "CMIC, Shanghai Jiao Tong University, China \nhttps://jyrao.github.io/UniSoccer/" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 247, + 200, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 247, + 200, + 260 + ], + "spans": [ + { + "bbox": [ + 152, + 247, + 200, + 260 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 271, + 296, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 271, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 54, + 271, + 296, + 499 + ], + "type": "text", + "content": "As a globally celebrated sport, soccer has attracted widespread interest from fans all over the world. This paper aims to develop a comprehensive multi-modal framework for soccer video understanding. Specifically, we make the following contributions in this paper: (i) we introduce SoccerReplay-1988, the largest multi-modal soccer dataset to date, featuring videos and detailed annotations from 1,988 complete matches, with an automated annotation pipeline; (ii) we present an advanced soccer-specific visual encoder, MatchVision, which leverages spatiotemporal information across soccer videos and excels in various downstream tasks; (iii) we conduct extensive experiments and ablation studies on event classification, commentary generation, and multi-view foul recognition. MatchVision demonstrates state-of-the-art performance on all of them, substantially outperforming existing models, which highlights the superiority of our proposed data and model. We believe that this work will offer a standard paradigm for sports understanding research." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 505, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 295, + 529 + ], + "type": "text", + "content": "\"Football is one of the world's best means of communication. It is impartial, apolitical, and universal.\"" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 529, + 295, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 529, + 295, + 541 + ], + "spans": [ + { + "bbox": [ + 137, + 529, + 295, + 541 + ], + "type": "text", + "content": "— Franz Beckenbauer (1945 - 2024)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 558, + 135, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 558, + 135, + 571 + ], + "spans": [ + { + "bbox": [ + 56, + 558, + 135, + 571 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 578, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 296, + 685 + ], + "type": "text", + "content": "Soccer, celebrated worldwide for its significant commercial value, has recently seen great research interest in integrating artificial intelligence (AI) for soccer video understanding. This is primarily motivated by the sport's complexity and the growing demand for enhanced analytics and improved viewing experiences. AI systems facilitate tactical analysis [49], allowing coaches to devise better strategies by uncovering patterns not apparent to the naked eye. In addition, it also supports automated content generation and enriches" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 236, + 335, + 250 + ], + "blocks": [ + { + "bbox": [ + 318, + 236, + 335, + 250 + ], + "lines": [ + { + "bbox": [ + 318, + 236, + 335, + 250 + ], + "spans": [ + { + "bbox": [ + 318, + 236, + 335, + 250 + ], + "type": "image", + "image_path": "b768a20c91a1e5521d31cbbdecbc86cf5ce93b5b2297350c47f6ff106847ef5b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 239, + 457, + 251 + ], + "lines": [ + { + "bbox": [ + 337, + 239, + 457, + 251 + ], + "spans": [ + { + "bbox": [ + 337, + 239, + 457, + 251 + ], + "type": "text", + "content": "Videos in SoccerReplay-1988" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 460, + 238, + 476, + 250 + ], + "blocks": [ + { + "bbox": [ + 460, + 238, + 476, + 250 + ], + "lines": [ + { + "bbox": [ + 460, + 238, + 476, + 250 + ], + "spans": [ + { + "bbox": [ + 460, + 238, + 476, + 250 + ], + "type": "image", + "image_path": "d8fceea74ff9cd5dcca5ddf474b1fa93d9e505c697fe8845e096779cc9f8715b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 316, + 253, + 553, + 434 + ], + "blocks": [ + { + "bbox": [ + 316, + 253, + 553, + 434 + ], + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 434 + ], + "type": "image", + "image_path": "ed384a51e59726eec53f918b94be181ccdfedb6766313c36bbcbed6f6b542de5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 440, + 555, + 485 + ], + "lines": [ + { + "bbox": [ + 313, + 440, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 555, + 485 + ], + "type": "text", + "content": "Figure 1. Overview. We present SoccerReplay-1988, the largest soccer dataset to date, and a powerful soccer-specific visual encoder, MatchVision, capable of excelling in various tasks such as event classification and commentary generation." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 496, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 555, + 544 + ], + "type": "text", + "content": "fan engagement through interactive and personalized content [34, 36, 40]. These capabilities promote a deeper understanding of soccer, simplify content creation, and foster a more engaging experience for fans and professionals." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 544, + 556, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 556, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 556, + 676 + ], + "type": "text", + "content": "Existing research in soccer video analysis primarily revolves around the SoccerNet series datasets [7, 10, 15], which comprise 500 full-match videos for benchmarking various tasks, such as action spotting [10, 15] and commentary generation [34, 36, 40]. Despite this extensive coverage, the focus has predominantly been on designing specialized models for task-specific applications, leading to fragmented and incompatible solutions. Such fragmentation underscores the need for a unified framework capable of integrating diverse demands, enabling more holistic and scalable advancements in soccer video understanding." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "In this paper, we introduce SoccerReplay-1988, the largest and most comprehensive multi-modal soccer video dataset to date, featuring 1,988 complete match videos with" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 693, + 214, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 214, + 703 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 214, + 703 + ], + "type": "text", + "content": "*: These authors contribute equally to this work." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 56, + 704, + 138, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 704, + 138, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 704, + 138, + 712 + ], + "type": "text", + "content": "†: Corresponding author." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8384" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "rich annotations, such as event labels and textual commentaries. This dataset offers a solid foundation for developing advanced soccer understanding models and establishes a challenging new benchmark for the field. Additionally, we have harmonized existing datasets to be compatible with ours, further expanding the available data resources." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 295, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 287 + ], + "type": "text", + "content": "Leveraging this dataset, we develop MatchVision, an advanced soccer-specific visual encoder tailored for diverse soccer understanding tasks. It employs the cutting-edge visual-language foundation model as the backbone, e.g., SigLIP [58]. We further extend framewise visual features into spatiotemporal representations with temporal attentions [3], by training on diverse visual-language tasks on SoccerReplay-1988, as depicted in Figure 1. As a result, MatchVision exhibits strong adaptability across various tasks, such as event classification and commentary generation, serving as a universal and unified framework for comprehensive soccer video understanding." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 288, + 295, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 288, + 295, + 526 + ], + "spans": [ + { + "bbox": [ + 56, + 288, + 295, + 526 + ], + "type": "text", + "content": "To summarize, we make the following contributions in this paper: (i) we construct SoccerReplay-1988, the largest and most diverse soccer video dataset to date, featuring videos of 1,988 soccer matches with rich annotations, supported by an automated curation pipeline. This provides a solid foundation for developing robust and comprehensive soccer understanding models; (ii) we present a powerful soccer-specific visual encoder, termed MatchVision, which effectively leverages spatiotemporal information in soccer videos, and can adapt to various tasks such as event classification and commentary generation, serving as a unified framework for soccer understanding; (iii) we establish more comprehensive and challenging benchmarks based on our dataset, enabling more professional evaluation of soccer understanding models; (iv) extensive experiments and ablation studies demonstrate the superiority of our data and model across various downstream tasks, achieving state-of-the-art performance on both existing benchmarks and our newly established ones. We believe this work offers a viable paradigm for future sports video understanding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 536, + 146, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 146, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 146, + 548 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 555, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 675 + ], + "type": "text", + "content": "Sports Understanding [44] is an evolving field that encompasses multiple research topics and integrates diverse data modalities, covering various tasks such as action spotting [10, 15, 16], commentary generation [34, 37, 40, 51, 52, 57], athlete analysis [41, 55], tactical planning [49], sports health [39], and intelligent refereeing [22, 23]. Furthermore, with the rapid development of multimodal large language models (MLLMs), recent efforts [26, 50, 53, 54] have attempted to build more generalized frameworks to uniformly handle a variety of sports understanding tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Visual-Language Models [1, 27, 28, 38, 58] have exhibited remarkable performance across extensive applications like classification, segmentation, image-text retrieval, and image" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "captioning. Recent efforts have ventured into more challenging video understanding [29, 30, 42, 59, 60] tasks, such as temporal alignment [17, 31], dense captioning [5, 56, 62], and audio description [18-20]. However, these efforts typically focus on general scenarios, limiting their adaptability to specific professional fields. Thus, this paper aims to bridge this gap by advancing visual-language models tailored for comprehensive soccer understanding." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 170, + 555, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 170, + 555, + 302 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 555, + 302 + ], + "type": "text", + "content": "Soccer Game Analysis [9] has primarily focused on tasks such as action spotting [15], replay grounding [10, 61], commentary generation [34, 36, 40], player tracking [8], state reconstruction [43], camera calibration [6, 8, 10] and foul recognition [22, 23], as facilitated by the SoccerNet [7, 10, 14, 15] series datasets, with 500 full-match videos from 2015 to 2017. Unlike existing methods that target designing specific models for distinct tasks, this paper aims to design a unified multi-modal framework that leverages spatiotemporal information within videos, serving as a specialized visual encoder for soccer video understanding." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 315, + 468, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 468, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 468, + 328 + ], + "type": "text", + "content": "3. SoccerReplay-1988 Dataset" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 335, + 553, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 553, + 407 + ], + "type": "text", + "content": "To establish a solid foundation for soccer understanding, we construct SoccerReplay-1988, the largest soccer dataset to date. Here, we first outline our data collection details and an overview of the dataset in Sec. 3.1; followed by elaborating on our automated data curation pipeline in Sec. 3.2; lastly, in Sec. 3.3, we present the data statistics and discussion." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 416, + 422, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 422, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 422, + 427 + ], + "type": "text", + "content": "3.1. Dataset Collection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 434, + 553, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 553, + 578 + ], + "type": "text", + "content": "To construct the SoccerReplay-1988 dataset, we have collected untrimmed, full-match videos from the Internet, encompassing a total of 1,988 matches from six European major soccer leagues and championships1, spanning the 2014-15 to 2023-24 seasons. For each match, we acquire textual commentaries with second-level timestamps from a sports text live website2, with part of them annotated with specific event types such as corner and goal. Additionally, we also incorporate extensive metadata, including detailed background information about the games, players, coaches, referees, and teams, providing a solid foundation for future soccer understanding research." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 578, + 553, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 553, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 553, + 673 + ], + "type": "text", + "content": "We partition the SoccerReplay-1988 dataset into train, validation, and test sets, containing 1,488, 250, and 250 full-match videos with diverse and comprehensive annotations, respectively. These sets provide rich training data for downstream tasks, such as event classification and commentary generation, while establishing comprehensive and challenging benchmarks for soccer understanding, as further discussed in subsequent sections." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 683, + 553, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 553, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 553, + 703 + ], + "type": "text", + "content": "1Premier (England), Laliga (Spain), Bundesliga (Germany), Serie-a (Italy), League-1 (France) and UEFA Champions League." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 703, + 369, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 703, + 369, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 703, + 369, + 712 + ], + "type": "text", + "content": "2flashscore.com" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8385" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 292, + 243 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 292, + 243 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 292, + 243 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 292, + 243 + ], + "type": "image", + "image_path": "ca6edb824bec182c0601875550b600151eadd1bf8ae1a66dbf292087b9c65f7b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 251, + 295, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 295, + 285 + ], + "type": "text", + "content": "Figure 2. Automated Data Curation Pipeline. The collected soccer video data are automatically processed for temporal alignment, event summarization, and anonymization by our curation pipeline." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 293, + 201, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 201, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 201, + 306 + ], + "type": "text", + "content": "3.2. Automated Data Curation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 311, + 296, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 311, + 296, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 296, + 372 + ], + "type": "text", + "content": "Given the potential noise in raw data, such as irrelevant video content, inaccurate timestamps, and incomplete event annotations, we design an automated data curation pipeline, comprising (i) temporal alignment, (ii) event summarization, and (iii) anonymization, as illustrated in Figure 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 373, + 296, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 421 + ], + "type": "text", + "content": "Temporal Alignment. Here, we divide match videos into two halves, each starting at kick-off, and adopt the temporal alignment model from MatchTime [40], to synchronize textual commentary timestamps with those of video frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 423, + 296, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 296, + 604 + ], + "type": "text", + "content": "Event Summarization. For samples without event annotations, we leverage LLaMA-3-70B [12] to summarize the events based on textual commentaries. Concretely, we have expanded the event categories from 17 in SoccerNet [10] to 24 types, for finer-grained soccer understanding, for example, categorizing penalties into scored and missed, and integrating modern soccer regulations like VAR. The resulting 24 event labels include: 'corner', 'goal', 'injury', 'own goal', 'penalty', 'penalty missed', 'red card', 'second yellow card', 'substitution', 'start of game (half)', 'end of game (half)', 'yellow card', 'throw in', 'free kick', 'saved by goal-keeper', 'shot off target', 'clearance', 'lead to corner', 'off-side', 'var', 'foul (no card)', 'statistics and summary', 'ball possession', and 'ball out of play'. More details on the used prompts are provided in the Appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "text", + "content": "Anonymization. Similar to [34], we extract all person and team entity names from the metadata of SoccerReplay-1988, and replace them in textual commentaries with placeholders, such as \"[PLAYER]\", \"[TEAM]\", \"[COACH]\", and \"[REFEREE]\", ensuring consistency across tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "Moreover, our data curation pipeline can seamlessly extend to existing datasets, converting the SoccerNet series [10, 34] into our unified data format, termed SoccerNetpro. This expansion further enlarges the standardized" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 554, + 209 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 554, + 209 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 554, + 209 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 554, + 209 + ], + "type": "table", + "html": "
Existing Datasets
# GameDuration(h)# Event# Anno.# Com.
SoccerNet-v1 [15]50076476.7k-
SoccerNet-v2 [10]50076417110k-
MatchTime [40]4717161414k37k
GOAL [36]2025.5--8.9k
Our Curated Datasets
# GameDuration(h)# Event# Anno.# Com.
SoccerNet-pro50076424102k37k
SoccerReplay-19881,9883,32324150k150k
Integrated2,4884,08724252k187k
", + "image_path": "a4ff8d585d0999336f8d7af8e33491bd175228e27dd5d83ccbdf23ffa33fcb7b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 214, + 555, + 259 + ], + "lines": [ + { + "bbox": [ + 313, + 214, + 555, + 259 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 555, + 259 + ], + "type": "text", + "content": "Table 1. Statistics of Soccer Datasets. Our SoccerReplay-1988 significantly surpasses existing datasets in both scale and diversity. Here, # Anno. and # Com. refer to the number of event annotations and textual commentaries, respectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 270, + 512, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 512, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 512, + 283 + ], + "type": "text", + "content": "datasets available for soccer understanding tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 293, + 443, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 443, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 443, + 304 + ], + "type": "text", + "content": "3.3. Statistics & Discussion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 312, + 554, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 554, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 554, + 370 + ], + "type": "text", + "content": "Dataset Statistics. As shown in Table 1, our dataset encompasses 3,323 hours of footage from 1,988 soccer matches, with an average duration of 100.3 minutes per match. The videos range in resolution from 360p to 720p and frame rates between 25 and 30 FPS." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 372, + 554, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 554, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 554, + 492 + ], + "type": "text", + "content": "For textual annotations, this dataset features approximately 150K commentaries, averaging 76 per match, precisely temporal-aligned by the robust alignment model from MatchTime [40]. These commentaries cover 4,467 unique words, significantly surpassing the 2,873 words in existing datasets [34, 40], greatly enriching textual diversity. Automated event summarization based on these commentaries has yielded about 150K event annotations. Notably, a random sampling of " + }, + { + "bbox": [ + 313, + 372, + 554, + 492 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 313, + 372, + 554, + 492 + ], + "type": "text", + "content": " of the data yields " + }, + { + "bbox": [ + 313, + 372, + 554, + 492 + ], + "type": "inline_equation", + "content": "98\\%" + }, + { + "bbox": [ + 313, + 372, + 554, + 492 + ], + "type": "text", + "content": " manual verification accuracy, ensuring high-quality automated labeling." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 495, + 554, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 554, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 554, + 603 + ], + "type": "text", + "content": "SoccerReplay-test Benchmark. To facilitate a more comprehensive evaluation of soccer understanding models, we integrate 250 matches from SoccerReplay-1988 with 50 matches from the curated SoccerNet-pro, establishing SoccerReplay-test, a more challenging benchmark for event classification and commentary generation. This benchmark features nearly four times larger than existing datasets and comprises finer-grained event labels, richer textual commentaries, and up-to-date soccer regulations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": "Discussion. To summarize, SoccerReplay-1988 exhibits advancements in three aspects: (i) it is the largest soccer video dataset to date, with nearly four times more videos than existing datasets; (ii) it features more professional and diverse annotations, more suitable for fine-grained and comprehensive soccer understanding tasks; (iii) It employs an automated curation pipeline for annotations and is thus scalable to provide a solid data foundation for future research. All data from SoccerReplay-1988, including videos" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8386" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 276, + 280 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 276, + 280 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 276, + 280 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 276, + 280 + ], + "type": "image", + "image_path": "e4c6e688d3460a7c195de0a3737ceca94017c7c0480ff3b70fd2a1036e793496.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 128, + 281, + 225, + 291 + ], + "lines": [ + { + "bbox": [ + 128, + 281, + 225, + 291 + ], + "spans": [ + { + "bbox": [ + 128, + 281, + 225, + 291 + ], + "type": "text", + "content": "(a) Structure of MatchVision" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 284, + 71, + 553, + 279 + ], + "blocks": [ + { + "bbox": [ + 284, + 71, + 553, + 279 + ], + "lines": [ + { + "bbox": [ + 284, + 71, + 553, + 279 + ], + "spans": [ + { + "bbox": [ + 284, + 71, + 553, + 279 + ], + "type": "image", + "image_path": "977119086969bdcf69088f68e5878a2f6416a7dae1682c8d3d5c485bda592645.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 281, + 396, + 292 + ], + "lines": [ + { + "bbox": [ + 299, + 281, + 396, + 292 + ], + "spans": [ + { + "bbox": [ + 299, + 281, + 396, + 292 + ], + "type": "text", + "content": "(b) MatchVision Pretraining" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 451, + 281, + 527, + 291 + ], + "lines": [ + { + "bbox": [ + 451, + 281, + 527, + 291 + ], + "spans": [ + { + "bbox": [ + 451, + 281, + 527, + 291 + ], + "type": "text", + "content": "(c) Downstream Tasks" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 299, + 555, + 333 + ], + "lines": [ + { + "bbox": [ + 54, + 299, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 54, + 299, + 555, + 333 + ], + "type": "text", + "content": "Figure 3. Overview of MatchVision. (a) The model architecture and its spatiotemporal feature extraction process; (b) Details of visual encoder pretraining, including supervised classification and video-language contrastive learning; (c) Implementation details of specific heads for various downstream tasks, including commentary generation, foul recognition, and event classification." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 345, + 293, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 293, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 293, + 357 + ], + "type": "text", + "content": "and annotations, are open-source for non-commercial use." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 366, + 111, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 111, + 379 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 111, + 379 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 387, + 296, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 296, + 482 + ], + "type": "text", + "content": "In this paper, we aim to develop a soccer-specific visual encoder, MatchVision, tailored for diverse soccer video analysis tasks. We start by outlining our problem formulation in Sec. 4.1. Next, in Sec. 4.2, we detail the architecture of MatchVision. The training procedures are thoroughly discussed in Sec. 4.3. Finally, we describe the configurations for downstream tasks in Sec. 4.4, demonstrating the practical applications and effectiveness of our model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 489, + 180, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 489, + 180, + 500 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 180, + 500 + ], + "type": "text", + "content": "4.1. Problem Formulation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 506, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 296, + 567 + ], + "type": "text", + "content": "In this work, we tackle the challenge of analyzing soccer video segments, denoted as " + }, + { + "bbox": [ + 55, + 506, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{V} \\in \\mathbb{R}^{T \\times 3 \\times H \\times W}" + }, + { + "bbox": [ + 55, + 506, + 296, + 567 + ], + "type": "text", + "content": ". Our goal is to utilize the visual encoder " + }, + { + "bbox": [ + 55, + 506, + 296, + 567 + ], + "type": "inline_equation", + "content": "(\\Phi_{\\mathrm{MatchVision}})" + }, + { + "bbox": [ + 55, + 506, + 296, + 567 + ], + "type": "text", + "content": " to extract spatiotemporal features from these segments, which are then processed by multiple task-specific heads, formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 110, + 574, + 240, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 574, + 240, + 586 + ], + "spans": [ + { + "bbox": [ + 110, + 574, + 240, + 586 + ], + "type": "interline_equation", + "content": "\\mathbf {E}, \\mathbf {C}, \\mathbf {F} = \\Psi (\\Phi_ {\\text {M a t c h V i s i o n}} (\\mathcal {V}))", + "image_path": "2db9ef356842691b74a34d6010ebcb6dbfc13f7379058adc2f8834b5ccd66104.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 594, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 666 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 594, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\Psi = \\{\\Psi_{\\mathrm{cls}}, \\Psi_{\\mathrm{Cmt}}, \\Psi_{\\mathrm{Foul}}\\}" + }, + { + "bbox": [ + 55, + 594, + 295, + 666 + ], + "type": "text", + "content": " represents the task-specific heads, with " + }, + { + "bbox": [ + 55, + 594, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 55, + 594, + 295, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 594, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 55, + 594, + 295, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 594, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 55, + 594, + 295, + 666 + ], + "type": "text", + "content": " denoting the output event types, textual commentaries, and foul types, respectively. This unified framework effectively learns relevant spatiotemporal features, and enables seamless integration across various downstream tasks for comprehensive soccer understanding." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 672, + 138, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 138, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 138, + 683 + ], + "type": "text", + "content": "4.2. Architecture" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "MatchVision comprises three key components: (i) Token Embedding, (ii) Spatiotemporal Attention Block, and (iii)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 345, + 491, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 491, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 491, + 357 + ], + "type": "text", + "content": "Aggregation Layer, as depicted in Figure 3." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "text", + "content": "Token Embedding. In accordance with the convention in Vision Transformer [11], each frame " + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "inline_equation", + "content": "(\\mathcal{I}_i)" + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "text", + "content": " from the video segment " + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "inline_equation", + "content": "(\\mathcal{V} = \\{\\mathcal{I}_1, \\mathcal{I}_2, \\dots, \\mathcal{I}_T\\})" + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "text", + "content": " is divided into " + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "text", + "content": " non-overlapping patches of size " + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "inline_equation", + "content": "P \\times P" + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "text", + "content": " that span the entire frame." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 407, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 555, + 502 + ], + "type": "text", + "content": "These patches are flattened into vectors " + }, + { + "bbox": [ + 313, + 407, + 555, + 502 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_i^p)" + }, + { + "bbox": [ + 313, + 407, + 555, + 502 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 407, + 555, + 502 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 407, + 555, + 502 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 407, + 555, + 502 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 407, + 555, + 502 + ], + "type": "text", + "content": " denote the spatial and temporal positions, respectively. Each vector is transformed via an embedding layer " + }, + { + "bbox": [ + 313, + 407, + 555, + 502 + ], + "type": "inline_equation", + "content": "(\\Phi_{\\mathrm{Emb}})" + }, + { + "bbox": [ + 313, + 407, + 555, + 502 + ], + "type": "text", + "content": " into a token vector of size " + }, + { + "bbox": [ + 313, + 407, + 555, + 502 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{1\\times D}" + }, + { + "bbox": [ + 313, + 407, + 555, + 502 + ], + "type": "text", + "content": ", and then added with a spatial position embedding " + }, + { + "bbox": [ + 313, + 407, + 555, + 502 + ], + "type": "inline_equation", + "content": "(\\mathbf{e}_{\\mathrm{s}}^{\\mathrm{pos}}\\in \\mathbb{R}^{M\\times D})" + }, + { + "bbox": [ + 313, + 407, + 555, + 502 + ], + "type": "text", + "content": ". Subsequently, we concatenate a [c1s] token along with each frame. Finally, a temporal positional embedding " + }, + { + "bbox": [ + 313, + 407, + 555, + 502 + ], + "type": "inline_equation", + "content": "(\\mathbf{e}_{\\mathrm{t}}^{\\mathrm{pos}}\\in \\mathbb{R}^{T\\times D})" + }, + { + "bbox": [ + 313, + 407, + 555, + 502 + ], + "type": "text", + "content": " is added across features of all frames, as formulated below:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 348, + 507, + 518, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 507, + 518, + 536 + ], + "spans": [ + { + "bbox": [ + 348, + 507, + 518, + 536 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {y} _ {i} = \\left[ \\mathbf {x} _ {i} ^ {\\mathrm {c l s}}, \\Phi_ {\\mathrm {E m b}} \\left(\\left[ \\mathbf {x} _ {i} ^ {1}, \\dots , \\mathbf {x} _ {i} ^ {M} \\right]\\right) + \\mathbf {e} _ {\\mathrm {s}} ^ {\\mathrm {p o s}} \\right] \\\\ \\mathbf {z} = \\left[ \\mathbf {y} _ {1}, \\dots , \\mathbf {y} _ {T} \\right] + \\mathbf {e} _ {\\mathrm {t}} ^ {\\text {p o s}} \\\\ \\end{array}", + "image_path": "6f6058956393aa8cbd37ec15851830ac141b58f56d6c68b8b85bfcc87afa84a6.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 543, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 554, + 653 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 543, + 554, + 653 + ], + "type": "inline_equation", + "content": "[\\cdot ,\\cdot ]" + }, + { + "bbox": [ + 313, + 543, + 554, + 653 + ], + "type": "text", + "content": " denotes concatenation, and " + }, + { + "bbox": [ + 313, + 543, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i\\in \\mathbb{R}^{(M + 1)\\times D}" + }, + { + "bbox": [ + 313, + 543, + 554, + 653 + ], + "type": "text", + "content": " represents the frame-wise features. The embedded features (z) will then serve as input for spatiotemporal attention blocks. Spatiotemporal Attention Block. Similar to TimeSformer [3], we utilize interleaved temporal and spatial attention to integrate spatiotemporal information in soccer videos. Concretely, each spatiotemporal attention block comprises a temporal self-attention layer and a spatial self-attention layer, i.e., " + }, + { + "bbox": [ + 313, + 543, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\phi_t(\\cdot)" + }, + { + "bbox": [ + 313, + 543, + 554, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 543, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\phi_s(\\cdot)" + }, + { + "bbox": [ + 313, + 543, + 554, + 653 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 653, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 712 + ], + "type": "text", + "content": "Given a video feature " + }, + { + "bbox": [ + 313, + 653, + 554, + 712 + ], + "type": "inline_equation", + "content": "(\\mathbf{z} \\in \\mathbb{R}^{T \\times (M + 1) \\times D})" + }, + { + "bbox": [ + 313, + 653, + 554, + 712 + ], + "type": "text", + "content": ", we alternate temporal and spatial attention: temporal attention facilitates interactions among tokens at the same spatial positions across distinct frames, while spatial attention enables interactions among tokens within the same frame. Residual con" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8387" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "nections are employed in each layer. After passing through a total of " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " spatiotemporal attention blocks, the resulting feature " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "(\\mathcal{F})" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " captures both intra-frame and inter-frame relationships, i.e., " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{F} = [\\phi_s(\\phi_t(\\mathbf{z}))]^{\\mathbf{K}} \\in \\mathbb{R}^{T \\times (M + 1) \\times D}" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 122, + 295, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 295, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 295, + 242 + ], + "type": "text", + "content": "Aggregation Layer. To obtain video-level features, we employ an aggregation layer on the frame-wise spatiotemporal features. Specifically, for the " + }, + { + "bbox": [ + 55, + 122, + 295, + 242 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 122, + 295, + 242 + ], + "type": "text", + "content": "-th frame, we utilize spatial self-attention to aggregate information into its [cls] token, denoted as " + }, + { + "bbox": [ + 55, + 122, + 295, + 242 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_i^{\\mathrm{cls}} = \\Phi_{\\mathrm{Agg}}(\\mathcal{F}_i)" + }, + { + "bbox": [ + 55, + 122, + 295, + 242 + ], + "type": "text", + "content": ". Concatenating the [cls] tokens of all frames yields the final video feature " + }, + { + "bbox": [ + 55, + 122, + 295, + 242 + ], + "type": "inline_equation", + "content": "(\\mathcal{F}_{\\mathcal{V}})" + }, + { + "bbox": [ + 55, + 122, + 295, + 242 + ], + "type": "text", + "content": ", that effectively encapsulates spatiotemporal characteristics of soccer video segments, thus enabling it to be applicable for various downstream soccer understanding tasks. This process can be formulated as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 252, + 282, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 252, + 282, + 268 + ], + "spans": [ + { + "bbox": [ + 69, + 252, + 282, + 268 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\mathcal {V}} = \\Phi_ {\\mathrm {M a t c h V i s i o n}} (\\mathcal {V}) = [ \\hat {\\mathcal {F}} _ {1} ^ {\\mathrm {c l s}}, \\dots , \\hat {\\mathcal {F}} _ {T} ^ {\\mathrm {c l s}} ] \\in \\mathbb {R} ^ {T \\times D}", + "image_path": "e709d753755367491d1d835565965e5a8e6d6f18b16949d3829ea1edb86893af.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 277, + 195, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 195, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 195, + 290 + ], + "type": "text", + "content": "4.3. MatchVision Pretraining" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 296, + 295, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 296, + 295, + 356 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 295, + 356 + ], + "type": "text", + "content": "In this part, we aim to pretrain the visual encoder with triplet samples " + }, + { + "bbox": [ + 55, + 296, + 295, + 356 + ], + "type": "inline_equation", + "content": "(\\{\\mathcal{V},\\mathbf{E},\\mathbf{C}\\})" + }, + { + "bbox": [ + 55, + 296, + 295, + 356 + ], + "type": "text", + "content": ", comprising videos, event labels, and textual commentaries. Concretely, we investigate two distinct pretraining strategies: supervised classification and video-language contrastive learning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 358, + 295, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 295, + 453 + ], + "type": "text", + "content": "Supervised Classification. One way to pretrain the visual encoder is supervised learning on event classification. To be specific, the extracted visual features " + }, + { + "bbox": [ + 55, + 358, + 295, + 453 + ], + "type": "inline_equation", + "content": "(\\mathcal{F}_{\\mathcal{V}})" + }, + { + "bbox": [ + 55, + 358, + 295, + 453 + ], + "type": "text", + "content": " are aggregated by a temporal self-attention layer into a learnable [cls] token, denoted as " + }, + { + "bbox": [ + 55, + 358, + 295, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathcal{V}}^{\\mathrm{cls}}" + }, + { + "bbox": [ + 55, + 358, + 295, + 453 + ], + "type": "text", + "content": ", similar to the spatial-wise aggregation mentioned above. This token is then fed into a linear classifier, and trained with a cross-entropy loss for event classification. The objective is denoted as " + }, + { + "bbox": [ + 55, + 358, + 295, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sup}}" + }, + { + "bbox": [ + 55, + 358, + 295, + 453 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 456, + 295, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 589 + ], + "type": "text", + "content": "Video-Language Contrastive Learning. As an alternative, we can also pretrain our visual encoder with video-text contrastive learning. Specifically, we adopt simple average pooling on the video feature to get the aggregated visual feature " + }, + { + "bbox": [ + 55, + 456, + 295, + 589 + ], + "type": "inline_equation", + "content": "(\\mathcal{F}_{\\mathcal{V}}^{\\mathrm{Avg}})" + }, + { + "bbox": [ + 55, + 456, + 295, + 589 + ], + "type": "text", + "content": ", and encode the textual commentary (C) with a text encoder " + }, + { + "bbox": [ + 55, + 456, + 295, + 589 + ], + "type": "inline_equation", + "content": "(\\Phi_{\\mathrm{Text}})" + }, + { + "bbox": [ + 55, + 456, + 295, + 589 + ], + "type": "text", + "content": ". We train the model with sigmoid loss (" + }, + { + "bbox": [ + 55, + 456, + 295, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sigmoid}}" + }, + { + "bbox": [ + 55, + 456, + 295, + 589 + ], + "type": "text", + "content": "), as used in SigLIP [58]. Note that, some video clips may have highly similar commentaries, for example, 'start of the game', we treat the commentaries with high similarity in the same batch as positive samples when calculating loss functions. This can be expressed as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 98, + 597, + 253, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 597, + 253, + 613 + ], + "spans": [ + { + "bbox": [ + 98, + 597, + 253, + 613 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n t r a}} = \\mathcal {L} _ {\\text {s i g m o i d}} \\left(\\mathcal {F} _ {\\mathcal {V}} ^ {\\text {A v g}}, \\Phi_ {\\text {T e x t}} (\\mathbf {C})\\right)", + "image_path": "43bbba471229b2ea8e2833b1df8839af35fc55bf9c99198f3b2f4b09bf931e3e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 624, + 167, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 167, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 167, + 635 + ], + "type": "text", + "content": "4.4. Downstream Tasks" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "After the pretraining mentioned above, MatchVision can now serve as a versatile visual encoder, to map the soccer video segments into visual features " + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "inline_equation", + "content": "(\\mathcal{F}_{\\mathcal{V}})" + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": ", for training task-specific heads " + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\Psi = \\{\\Psi_{\\mathrm{cls}}, \\Psi_{\\mathrm{Cmt}}, \\Psi_{\\mathrm{Foul}}\\}" + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": " across different downstream tasks, including: (i) event classification, (ii) commentary generation, and (iii) foul recognition." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "Event Classification. Similar to supervised classification above, we concatenate a learnable [cls] token to aggregate frame-wise visual features via temporal self-attention. This token is then fed into a linear classifier for event classification. The event classification head " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "(\\Psi_{\\mathrm{cls}})" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " is trained with a cross-entropy loss while freezing the visual encoder." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 147, + 555, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 147, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 147, + 555, + 266 + ], + "type": "text", + "content": "Commentary Generation. We follow the paradigm in MatchTime [40] to generate anonymized textual commentary for soccer video clips. Concretely, the commentary generation head " + }, + { + "bbox": [ + 313, + 147, + 555, + 266 + ], + "type": "inline_equation", + "content": "(\\Psi_{\\mathrm{Cmt}})" + }, + { + "bbox": [ + 313, + 147, + 555, + 266 + ], + "type": "text", + "content": " employs a Perceiver [25] aggregator to consolidate visual features, which are then projected by a trainable MLP, serving as prefix embeddings for a large language model (LLM). Subsequently, an off-the-shelf LLM decodes these embeddings into textual commentary. We adopt the negative log-likelihood loss, commonly used for auto-regressive next-token prediction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 269, + 554, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 554, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 554, + 412 + ], + "type": "text", + "content": "Foul Recognition. As outlined in [22], the foul recognition task takes multi-view videos from the same scene as inputs, with each sample annotated with a foul class (8 types) and severity (4 levels). We encode these multi-view videos with MatchVision, and aggregate the extracted features into a single feature vector, via either max or average pooling, following the common practice. Subsequently, the foul recognition head " + }, + { + "bbox": [ + 313, + 269, + 554, + 412 + ], + "type": "inline_equation", + "content": "(\\Psi_{\\mathrm{Foul}})" + }, + { + "bbox": [ + 313, + 269, + 554, + 412 + ], + "type": "text", + "content": " employs a shared MLP and two task-specific linear classifiers, to predict foul type and severity, respectively. Similar to event classification, we use the combination of cross-entropy losses on the foul type and severity classification to jointly train " + }, + { + "bbox": [ + 313, + 269, + 554, + 412 + ], + "type": "inline_equation", + "content": "\\Psi_{\\mathrm{Foul}}" + }, + { + "bbox": [ + 313, + 269, + 554, + 412 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 415, + 554, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 462 + ], + "type": "text", + "content": "Discussion. Pretraining MatchVision on large-scale soccer data equips it with substantial domain-specific knowledge, enabling it to serve as a universal visual encoder adaptable to various downstream soccer understanding tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 475, + 395, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 395, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 395, + 488 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 495, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 554, + 567 + ], + "type": "text", + "content": "This section begins with implementation details in Sec. 5.1; followed by quantitative evaluations across downstream tasks in Sec. 5.2; then, we conduct ablation studies on our SoccerReplay-test benchmark to analyze the effectiveness of the proposed dataset and model in Sec. 5.3; finally, we provide qualitative results for comparison in Sec. 5.4." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 576, + 447, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 447, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 447, + 588 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "In our experiments, video segments are sampled at 1FPS around annotated timestamps, capturing a 30-second window for each sample. Frames are resized to " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " pixels as inputs. We initialize the embedding layer, spatial attention layers, aggregation layer, and text encoder of MatchVision with pretrained weights from SigLIP Base-16 [58] and adopt LLaMA-3 (8B) [12] as the off-the-shelf LLM decoder for commentary generation. All experiments are conducted on " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " Nvidia H800 GPUs with the AdamW [33] optimizer. Next, we elaborate on the training and evaluation details" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "text", + "content": "8388" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 87, + 70, + 522, + 397 + ], + "blocks": [ + { + "bbox": [ + 87, + 70, + 522, + 397 + ], + "lines": [ + { + "bbox": [ + 87, + 70, + 522, + 397 + ], + "spans": [ + { + "bbox": [ + 87, + 70, + 522, + 397 + ], + "type": "table", + "html": "
Visual EncoderDatasetClassification (%)Commentary
SNMTSRAcc.@1Acc.@3Acc.@5B@1B@4MR-LC
Off-the-shelf Models
I3D [4]XXX45.482.593.226.775.5724.1723.1218.73
C3D [45]XXX47.885.195.028.136.6424.5224.2327.88
ResNet [21]XXX47.284.694.427.346.5724.7224.4327.29
CLIP [38]XXX48.585.595.226.256.5124.2724.7528.17
InternVideo [48]XXX49.987.095.927.126.5425.0224.8229.90
SigLIP [58]XXX50.286.795.627.856.9825.1625.0331.38
Pretrain with Supervised Classification
Baidu [61]✓XX56.491.997.331.208.8826.5626.6138.93
SigLIP✓XX55.989.694.928.517.3925.9625.9435.71
SigLIP✓✓✓57.991.797.530.958.5625.7926.1738.24
MatchVision✓XX82.596.698.829.457.9226.0126.2136.15
MatchVision✓✓✓84.097.399.231.059.0626.9427.9342.20
Pretrain with Visual-Language Contrastive Learning
SigLIPX✓X55.488.897.028.727.7225.9126.1732.27
SigLIPX✓✓66.893.798.630.358.1226.0526.3839.41
MatchVisionX✓X58.989.097.130.337.9725.4826.3333.87
MatchVisionX✓✓67.993.998.631.949.1226.2427.5640.76
Pretrain with Hybrid Supervised-Contrastive Training
SigLIP✓✓X71.294.598.728.637.8225.7425.3534.09
SigLIP✓✓✓67.193.298.130.718.7826.2626.7441.82
MatchVision✓✓X76.496.099.030.658.3325.2826.3137.23
MatchVision✓✓✓80.197.199.133.589.1426.8228.2144.18
", + "image_path": "d502cee976c4dea31f3fc7e5a902e3e873bd8fdba7e42283767eafe88893f1b7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 403, + 555, + 437 + ], + "lines": [ + { + "bbox": [ + 55, + 403, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 555, + 437 + ], + "type": "text", + "content": "Table 2. Quantitative Results on Event Classification and Commentary Generation. Here, SN, MT, and SR represent finetuning with curated SoccerNet-v2 [10], MatchTime [40], and SoccerReplay-1988, respectively. B, M, R-L, and C refer to BLEU, METEOR, ROUGE-L, and CIDEr metrics, respectively. Within each unit, we denote the best performance in RED and the second-best performance in BLUE." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 445, + 280, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 280, + 457 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 280, + 457 + ], + "type": "text", + "content": "about visual encoder pretraining and downstream tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 460, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 579 + ], + "type": "text", + "content": "Visual Encoder Pretraining. For both pretraining strategies, we use a batch size of 40 for 15 epochs. The learning rate for all randomly initialized modules, including the temporal attention layers, aggregator layer, and linear classifier, is set to " + }, + { + "bbox": [ + 55, + 460, + 295, + 579 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 55, + 460, + 295, + 579 + ], + "type": "text", + "content": ". Meanwhile, the learning rate for modules initialized with pretrained parameters (including the text encoder) is set to " + }, + { + "bbox": [ + 55, + 460, + 295, + 579 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 55, + 460, + 295, + 579 + ], + "type": "text", + "content": ". In contrastive training, we adopt a multi-positive strategy where each textual commentary, based on its event label, considers closely related categories (e.g. \"start of game\" and \"offside\") as positive samples." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "Downstream Tasks. In all downstream tasks, unless otherwise specified, we use the frozen visual encoder for feature extraction and only train the task-specific heads with a learning rate of " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " for 30 epochs. The batch sizes for event classification, commentary generation, and foul recognition are set to 40, 32, and 8, respectively. We adopt specific evaluation metrics for these three tasks: (i) For event classification, we use the top-1/3/5 classification accuracy; (ii) For commentary generation, we employ several commonly-used language evaluation metrics, including BLEU [35], METEOR [2], ROUGE-L [32], and" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 445, + 555, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 481 + ], + "type": "text", + "content": "CIDEr [47]; (iii) For foul recognition, we follow the common practice, and report top-1/2 and top-1 accuracy for the foul type and severity classification, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 483, + 556, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 556, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 556, + 639 + ], + "type": "text", + "content": "Benchmarks & Baselines. To ensure fair and reliable comparisons with existing work, we evaluate event classification (24 types) on 100 matches from curated SoccerNetv2 [10] test set; commentary generation on 49 matches from SN-Caption-test-align benchmark manually aligned in [40]; and foul recognition on MVFoul [22]. We consider various baselines: for the first two tasks, this includes off-the-shelf general visual encoders such as ResNet [21], C3D [45], I3D [4], CLIP [38], SigLIP [58], and InternVideo [48], along with Baidu [61] and SigLIP finetuned with soccer-specific data. For foul recognition, we follow previous work [22, 23] and adopt ResNet [21], R(2+1)D [46], and MViT [13] jointly finetuned with classifiers, as baselines." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 647, + 450, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 450, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 450, + 661 + ], + "type": "text", + "content": "5.2.Quantitative Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "text", + "content": "As depicted in Table 2, we draw two observations on event classification and commentary generation: (i) visual encoders trained on soccer data substantially outperform off-the-shelf general encoders (ResNet, C3D, I3D, CLIP, and" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8389" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 291, + 212 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 291, + 212 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 291, + 212 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 291, + 212 + ], + "type": "table", + "html": "
Visual EncoderFoul ClassSeverity
BackboneTrainAgg.Acc.@1Acc.@2Acc.@1
ResNet [21]✓Mean0.310.560.34
Max0.320.600.32
R(2+1)D [46]✓Mean0.310.550.34
Max0.320.560.39
MViT [13]✓Mean0.400.650.38
Max0.470.690.43
MatchVision×Mean0.440.530.58
Max0.350.700.46
", + "image_path": "1e7396307c4fef1618267f34b5fe5bb12fa48b57efeace8acc9710a23730fc40.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 264, + 295, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 295, + 335 + ], + "type": "text", + "content": "InternVideo), underscoring the necessity of building specialized models for soccer understanding; (ii) almost all visual encoders, across all training settings, benefit from SoccerReplay-1988, emphasizing the value of constructing large-scale, high-quality data for soccer understanding. Next, we will delve into each task to discuss the results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 338, + 295, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 295, + 458 + ], + "type": "text", + "content": "Event Classification. With identical training strategies and data, MatchVision considerably outperforms other methods in classification accuracy, demonstrating the superiority of its architecture, which effectively leverages spatiotemporal features within soccer videos. For example, MatchVision achieves a Top-1 accuracy of " + }, + { + "bbox": [ + 55, + 338, + 295, + 458 + ], + "type": "inline_equation", + "content": "82.5\\%" + }, + { + "bbox": [ + 55, + 338, + 295, + 458 + ], + "type": "text", + "content": ", significantly surpassing SigLIP's " + }, + { + "bbox": [ + 55, + 338, + 295, + 458 + ], + "type": "inline_equation", + "content": "55.9\\%" + }, + { + "bbox": [ + 55, + 338, + 295, + 458 + ], + "type": "text", + "content": " under the same training conditions. Moreover, models trained via supervised classification excel others, primarily because the pre-training task shares the same objectives as the downstream event classification task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 460, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 652 + ], + "type": "text", + "content": "Commentary Generation. Visual encoders trained with visual-language contrastive learning exhibit better commentary generation performance than those trained with supervised classification, as this strategy better captures correlations between visual and textual features. Additionally, while MatchVision trained solely on SoccerNet slightly underperforms Baidu [61], incorporating SoccerReplay-1988 enables it to outperform on most metrics. This demonstrates that MatchVision can take advantage of large-scale datasets. Finally, a hybrid training approach, starting with supervised classification followed by visual-language contrastive learning, enables MatchVision to achieve optimal performance. This indicates that learning coarse-grained tasks such as classification provides a foundation for fine-grained tasks like commentary generation, and fully leveraging data unlocks the potential of soccer understanding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "text", + "content": "Foul Recognition. As demonstrated in Table 3, MatchVision achieves performance comparable to jointly finetuned state-of-the-art methods in foul recognition, even with a frozen visual encoder. This highlights that MatchVision effectively learns substantial knowledge from large-scale soc" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 335, + 70, + 533, + 186 + ], + "blocks": [ + { + "bbox": [ + 55, + 218, + 295, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 218, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 295, + 251 + ], + "type": "text", + "content": "Table 3. Quantitative Results on Multi-view Foul Recognition. Our frozen MatchVision encoder can achieve comparable performance with other jointly finetuned visual encoders." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 335, + 70, + 533, + 186 + ], + "lines": [ + { + "bbox": [ + 335, + 70, + 533, + 186 + ], + "spans": [ + { + "bbox": [ + 335, + 70, + 533, + 186 + ], + "type": "table", + "html": "
PretrainClassification(%)
Sup.Contra.SRAcc.@1Acc.@3Acc.@5
✓XX62.6783.0089.81
✓X✓68.0386.9092.38
X✓X46.9775.5385.85
X✓✓57.4183.1391.00
✓✓X56.8680.3088.09
✓✓✓63.5985.2191.63
", + "image_path": "21608bf50b0ecb1bd5348751dc6d315743bf72ac03842ae185193d4c36abf0d4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 190, + 555, + 247 + ], + "lines": [ + { + "bbox": [ + 313, + 190, + 555, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 555, + 247 + ], + "type": "text", + "content": "Table 4. Ablations on Event Classification. We explore the impact of various training settings of our MatchVision encoder on the SoccerReplay-test benchmark. Here, Sup., Contra., and SR refer to supervised classification, visual-language contrastive learning, and the SoccerReplay-1988 dataset, respectively." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 260, + 555, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 555, + 297 + ], + "type": "text", + "content": "cer data and adapts seamlessly to downstream tasks. Comparisons with additional baselines from the SoccerNet foul recognition challenges [9] are provided in the Appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 306, + 416, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 416, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 416, + 319 + ], + "type": "text", + "content": "5.3. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 324, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 384 + ], + "type": "text", + "content": "We conduct ablation experiments on event classification and commentary generation using our SoccerReplay-test benchmark. These experiments validate the effectiveness of our proposed dataset and model, while establishing a baseline for future evaluations on this benchmark." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 387, + 556, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 556, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 556, + 579 + ], + "type": "text", + "content": "Event Classification. We evaluate event classification on 300 matches from our SoccerReplay-test benchmark using the MatchVision visual encoder pretrained with various strategies. Features are extracted by MatchVision and processed with a learnable aggregation layer and a linear classifier. The default training set is our curated SoccerNet-pro. As shown in Table 4, integrating SoccerReplay-1988 for training results in significant performance improvements across all pretraining strategies, yielding the significance of our dataset. Additionally, supervised classification outperforms visual-language contrastive learning and hybrid pretraining. This is due to its closer alignment with downstream event classification task, and the scale of event annotations is far larger than that of textual commentaries, further confirming the substantial benefits of data scaling for boosting soccer understanding." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Commentary Generation. With the pretrained MatchVision encoder, we train the commentary generation head on the MatchTime [40] and SoccerReplay-1988 datasets using various training strategies. By default, only the Perceiver [25] aggregation layer and projection layer within the head are trained. For joint training with the LLM decoder, considering computational costs, we incorporate LoRA [24] layers while freezing the original LLM layers. As shown in Table 5, incorporating SoccerReplay-1988 significantly improves performance on all metrics, confirming substantial advantages of our proposed dataset. This performance" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8390" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 73, + 83 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 73, + 83 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 73, + 83 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 73, + 83 + ], + "type": "image", + "image_path": "9d80294ea391f74578532a7a932bfd89ac9a197638371fc49f6dce60d8272ed9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 75, + 117, + 82 + ], + "lines": [ + { + "bbox": [ + 74, + 75, + 117, + 82 + ], + "spans": [ + { + "bbox": [ + 74, + 75, + 117, + 82 + ], + "type": "text", + "content": "Soccer Videos" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 229, + 73, + 240, + 83 + ], + "blocks": [ + { + "bbox": [ + 229, + 73, + 240, + 83 + ], + "lines": [ + { + "bbox": [ + 229, + 73, + 240, + 83 + ], + "spans": [ + { + "bbox": [ + 229, + 73, + 240, + 83 + ], + "type": "image", + "image_path": "9da4b340fa5e701f874da4eab9b7b914104a894fefb77e3bfec73dbc28c6aaee.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 75, + 263, + 81 + ], + "lines": [ + { + "bbox": [ + 241, + 75, + 263, + 81 + ], + "spans": [ + { + "bbox": [ + 241, + 75, + 263, + 81 + ], + "type": "text", + "content": "Events" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 319, + 72, + 329, + 83 + ], + "blocks": [ + { + "bbox": [ + 319, + 72, + 329, + 83 + ], + "lines": [ + { + "bbox": [ + 319, + 72, + 329, + 83 + ], + "spans": [ + { + "bbox": [ + 319, + 72, + 329, + 83 + ], + "type": "image", + "image_path": "dfe11a125f3ab53dc3d4ff9c58672ea87bd1c11d4c1cb608159d5405b50da324.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 74, + 377, + 81 + ], + "lines": [ + { + "bbox": [ + 331, + 74, + 377, + 81 + ], + "spans": [ + { + "bbox": [ + 331, + 74, + 377, + 81 + ], + "type": "text", + "content": "Commentaries" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 59, + 85, + 225, + 119 + ], + "blocks": [ + { + "bbox": [ + 59, + 85, + 225, + 119 + ], + "lines": [ + { + "bbox": [ + 59, + 85, + 225, + 119 + ], + "spans": [ + { + "bbox": [ + 59, + 85, + 225, + 119 + ], + "type": "image", + "image_path": "8cdd4dbbecb8a9ec6e3fdaac06188a5bdb8d9593f88138abc2c3f127dc7a9875.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 228, + 94, + 281, + 114 + ], + "blocks": [ + { + "bbox": [ + 228, + 94, + 281, + 114 + ], + "lines": [ + { + "bbox": [ + 228, + 94, + 281, + 114 + ], + "spans": [ + { + "bbox": [ + 228, + 94, + 281, + 114 + ], + "type": "image", + "image_path": "68732df897e97929bec2c06a6cd4a0e00c3c827ef4c30cf5c5ca0c4f68c81307.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 322, + 91, + 544, + 117 + ], + "blocks": [ + { + "bbox": [ + 322, + 91, + 544, + 117 + ], + "lines": [ + { + "bbox": [ + 322, + 91, + 544, + 117 + ], + "spans": [ + { + "bbox": [ + 322, + 91, + 544, + 117 + ], + "type": "image", + "image_path": "fc9c2bb97514f72b61de963098cb78a266efe3146a378467055940157f707b37.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 59, + 122, + 225, + 156 + ], + "blocks": [ + { + "bbox": [ + 59, + 122, + 225, + 156 + ], + "lines": [ + { + "bbox": [ + 59, + 122, + 225, + 156 + ], + "spans": [ + { + "bbox": [ + 59, + 122, + 225, + 156 + ], + "type": "image", + "image_path": "52ed70c15dc4709676df0e4b456ec14925d160a95ba8aee8054e3fadab030edd.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 228, + 129, + 315, + 150 + ], + "blocks": [ + { + "bbox": [ + 228, + 129, + 315, + 150 + ], + "lines": [ + { + "bbox": [ + 228, + 129, + 315, + 150 + ], + "spans": [ + { + "bbox": [ + 228, + 129, + 315, + 150 + ], + "type": "image", + "image_path": "8276801c2dc956cf715cbb6e35392feb9ec6c5c1e35246581e3815e139c44262.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 323, + 123, + 545, + 156 + ], + "blocks": [ + { + "bbox": [ + 323, + 123, + 545, + 156 + ], + "lines": [ + { + "bbox": [ + 323, + 123, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 323, + 123, + 545, + 156 + ], + "type": "image", + "image_path": "516894ea4acf31fed702636a2327972b70c6c0686e548ec9008c7942bb866031.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 59, + 157, + 225, + 191 + ], + "blocks": [ + { + "bbox": [ + 59, + 157, + 225, + 191 + ], + "lines": [ + { + "bbox": [ + 59, + 157, + 225, + 191 + ], + "spans": [ + { + "bbox": [ + 59, + 157, + 225, + 191 + ], + "type": "image", + "image_path": "270fd650638147ed4738f17a034c436bce1a5e856c8100fe1265415bc8ef1e7e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 228, + 166, + 299, + 186 + ], + "blocks": [ + { + "bbox": [ + 228, + 166, + 299, + 186 + ], + "lines": [ + { + "bbox": [ + 228, + 166, + 299, + 186 + ], + "spans": [ + { + "bbox": [ + 228, + 166, + 299, + 186 + ], + "type": "image", + "image_path": "b0de7aef3c8941434bfbb3eea98b327a676dacfc6f7f83d3791ac06e9b00f0f7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 323, + 159, + 545, + 191 + ], + "blocks": [ + { + "bbox": [ + 323, + 159, + 545, + 191 + ], + "lines": [ + { + "bbox": [ + 323, + 159, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 323, + 159, + 545, + 191 + ], + "type": "image", + "image_path": "a380b47f26580073f501fca6a80af2d27b0dff7e9f36d75a2471d72d0c93d535.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 59, + 193, + 225, + 228 + ], + "blocks": [ + { + "bbox": [ + 59, + 193, + 225, + 228 + ], + "lines": [ + { + "bbox": [ + 59, + 193, + 225, + 228 + ], + "spans": [ + { + "bbox": [ + 59, + 193, + 225, + 228 + ], + "type": "image", + "image_path": "3eefcf41d5f78e47d7d847d40c2c7378f9003d67dcc7f9abc8a12c41c3846a2e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 228, + 201, + 292, + 223 + ], + "blocks": [ + { + "bbox": [ + 228, + 201, + 292, + 223 + ], + "lines": [ + { + "bbox": [ + 228, + 201, + 292, + 223 + ], + "spans": [ + { + "bbox": [ + 228, + 201, + 292, + 223 + ], + "type": "image", + "image_path": "3b4d60d20742533f20aab26bf153e9159875fb7e3fc21dc2d1ab6a2e9a7e5be3.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 323, + 196, + 544, + 228 + ], + "blocks": [ + { + "bbox": [ + 323, + 196, + 544, + 228 + ], + "lines": [ + { + "bbox": [ + 323, + 196, + 544, + 228 + ], + "spans": [ + { + "bbox": [ + 323, + 196, + 544, + 228 + ], + "type": "image", + "image_path": "6bdbf1b0e4e1e262d7f85f385ddf1be5a1f0c0e018b967e3eade70b46797fd21.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 59, + 230, + 225, + 264 + ], + "blocks": [ + { + "bbox": [ + 59, + 230, + 225, + 264 + ], + "lines": [ + { + "bbox": [ + 59, + 230, + 225, + 264 + ], + "spans": [ + { + "bbox": [ + 59, + 230, + 225, + 264 + ], + "type": "image", + "image_path": "c4b3b98a4be399f96a73b58101fd17cc5b827266907bce570f39aa35c3e2d7ff.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 277, + 555, + 332 + ], + "lines": [ + { + "bbox": [ + 55, + 277, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 555, + 332 + ], + "type": "text", + "content": "Figure 4. Qualitative Results for Event Classification and Commentary Generation. Here, \"w/o SR\" and \"w/ SR\" indicate models trained without and with the SoccerReplay-1988 dataset, respectively. Incorporating SoccerReplay-1988 improves event classification accuracy. Moreover, this enriched training data enables the model to demonstrate several advantages in commentary generation: (a) more detailed descriptions, (b) greater linguistic variety, (c) higher event depiction accuracy, (d) better adherence to updated rules, and (e) improved specificity in scenario response." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 228, + 238, + 282, + 259 + ], + "blocks": [ + { + "bbox": [ + 228, + 238, + 282, + 259 + ], + "lines": [ + { + "bbox": [ + 228, + 238, + 282, + 259 + ], + "spans": [ + { + "bbox": [ + 228, + 238, + 282, + 259 + ], + "type": "image", + "image_path": "9b8addab99c97babf3f33b9d8f1d4cb75828a7be35583514c63ea2e65e97bcee.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 323, + 236, + 538, + 262 + ], + "blocks": [ + { + "bbox": [ + 323, + 236, + 538, + 262 + ], + "lines": [ + { + "bbox": [ + 323, + 236, + 538, + 262 + ], + "spans": [ + { + "bbox": [ + 323, + 236, + 538, + 262 + ], + "type": "image", + "image_path": "39aeeb1cf36f84a1d0fe1d9adf45176e87fd7b4d6118c5b90ca344994a7a1b2b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "table", + "bbox": [ + 69, + 345, + 280, + 509 + ], + "blocks": [ + { + "bbox": [ + 69, + 345, + 280, + 509 + ], + "lines": [ + { + "bbox": [ + 69, + 345, + 280, + 509 + ], + "spans": [ + { + "bbox": [ + 69, + 345, + 280, + 509 + ], + "type": "table", + "html": "
TrainableCommentary Metrics
VLB@1B@4MR-LC
Trained on MatchTime
XX21.653.2721.0217.7912.90
✓X27.627.0224.0323.5130.77
X✓27.046.4124.1523.8831.91
✓✓27.496.9624.5023.3330.81
Trained on MatchTime & SoccerReplay-1988
XX24.174.0920.5120.7015.70
✓X28.988.3924.4525.3545.85
X✓27.547.7624.5024.7042.79
✓✓29.218.2225.2525.5443.18
", + "image_path": "e1760a007d98c191a6ad49503d990e87dd10e26631844807f40b023b88dfe480.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 515, + 295, + 559 + ], + "lines": [ + { + "bbox": [ + 55, + 515, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 295, + 559 + ], + "type": "text", + "content": "Table 5. Ablations on Commentary Generation. We investigate the impact of different training strategies and datasets on MatchVision using the SoccerReplay-test benchmark. 'V' and 'L' denote the visual encoder and the LLM decoder, respectively." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 568, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 296, + 628 + ], + "type": "text", + "content": "gap also reflects the challenges of our established benchmark, which features diverse vocabulary, richer semantics, and updated soccer rules. Additionally, jointly finetuning the visual encoder and the LLM decoder provides a feasible approach for further improvements." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 635, + 195, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 195, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 195, + 649 + ], + "type": "text", + "content": "5.4. Qualitative Comparisons" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "As depicted in Figure 4, we present qualitative results of MatchVision on the SoccerReplay-test benchmark, comparing models pretrained with and without SoccerReplay-1988. For event classification, incorporating our data improves accuracy, and even in misclassified cases, the results" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 347, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 407 + ], + "type": "text", + "content": "remain contextually relevant. For commentary generation, hybrid training on SoccerReplay-1988 enables MatchVision to produce richer, more detailed textual commentary, reflecting a deeper understanding of soccer dynamics. More qualitative results are available in the Appendix." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 314, + 429, + 388, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 429, + 388, + 441 + ], + "spans": [ + { + "bbox": [ + 314, + 429, + 388, + 441 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 453, + 555, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 555, + 645 + ], + "type": "text", + "content": "In this paper, we establish a unified, scalable multi-modal framework for soccer understanding. Specifically, we introduce SoccerReplay-1988, the largest and most comprehensive soccer video dataset to date, annotated by an automated curation pipeline. This provides a solid foundation for developing soccer understanding models and serves as a more challenging benchmark. Built upon this, we develop MatchVision, an advanced soccer-specific visual encoder, which effectively leverages spatiotemporal information within soccer videos and can be applied to various tasks such as event classification and commentary generation. Extensive experiments demonstrate the superiority of our model, with MatchVision achieving state-of-the-art performance on both existing benchmarks and our newly established one. We believe this work will set a viable, universal paradigm for future research in sports understanding." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 314, + 666, + 411, + 679 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 666, + 411, + 679 + ], + "spans": [ + { + "bbox": [ + 314, + 666, + 411, + 679 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "This work is funded by National Key R&D Program of China (No.2022ZD0161400)." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8391" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, pages 23716-23736, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 158, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 158, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 296, + 213 + ], + "type": "text", + "content": "[2] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, pages 65-72, 2005. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 214, + 295, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 295, + 257 + ], + "type": "text", + "content": "[3] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In Proceedings of the International Conference on Machine Learning, page 4, 2021. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 259, + 296, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 259, + 296, + 302 + ], + "spans": [ + { + "bbox": [ + 62, + 259, + 296, + 302 + ], + "type": "text", + "content": "[4] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 304, + 294, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 304, + 294, + 379 + ], + "spans": [ + { + "bbox": [ + 62, + 304, + 294, + 379 + ], + "type": "text", + "content": "[5] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, and Sergey Tulyakov. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 381, + 294, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 381, + 294, + 457 + ], + "spans": [ + { + "bbox": [ + 62, + 381, + 294, + 457 + ], + "type": "text", + "content": "[6] Anthony Cioppa, Adrien Deliege, Floriane Magera, Silvio Giancola, Olivier Barnich, Bernard Ghanem, and Marc Van Droogenbroeck. Camera calibration and player localization in soccernet-v2 and investigation of their representations for action spotting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 4537-4546, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 460, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 294, + 502 + ], + "type": "text", + "content": "[7] Anthony Cioppa, Adrien Deliege, Silvio Giancola, Bernard Ghanem, and Marc Van Droogenbroeck. Scaling up soccer-net with multi-view spatial localization and re-identification. Scientific data, 9(1):355, 2022. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 504, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 504, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 62, + 504, + 294, + 569 + ], + "type": "text", + "content": "[8] Anthony Cioppa, Silvio Giancola, Adrien Deliege, Le Kang, Xin Zhou, Zhiyu Cheng, Bernard Ghanem, and Marc Van Droogenbroeck. Soccernet-tracking: Multiple object tracking dataset and benchmark in soccer videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 3491-3502, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 571, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 571, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 571, + 294, + 713 + ], + "type": "text", + "content": "[9] Anthony Cioppa, Silvio Giancola, Vladimir Somers, Victor Joos, Floriane Magera, Jan Held, Seyed Abolfazl Ghasemzadeh, Xin Zhou, Karolina Seweryn, Mateusz Kowalczyk, Zuzanna Mroz, Szymon Lukasik, Michal Halon, Hassan Mkhallati, Adrien Deliege, Carlos Hinojosa, Karen Sanchez, Amir M. Mansourian, Pierre Miralles, Olivier Barnich, Christophe De Vleeschouwer, Alexandre Alahi, Bernard Ghanem, Marc Van Droogenbroeck, Adam Gorski, Albert Clapes, Andrei Boiarov, Anton Afanasiev, Artur Xarles, Atom Scott, ByoungKwon Lim, Calvin Yeung, Cristian Gonzalez, Dominic Rufenacht, Enzo Pacilio, Fabian Deuser, Faisal Sami Altawijri, Francisco Cachon, HanKyul Kim, Haobo Wang, Hyeonmin Choe, Hyunwoo J Kim, Il-Min" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 215 + ], + "type": "text", + "content": "Kim, Jae-Mo Kang, Jamshid Tursunboev, Jian Yang, Jihwan Hong, Jimin Lee, Jing Zhang, Junseok Lee, Kexin Zhang, Konrad Habel, Licheng Jiao, Linyi Li, Marc Gutierrez-Perez, Marcelo Ortega, Menglong Li, Milosz Lopatto, Nikita Kasatkin, Nikolay Nemtsev, Norbert Oswald, Oleg Udin, Pavel Kononov, Pei Geng, Saad Ghazai Alotaibi, Sehyung Kim, Sergei Ulasen, Sergio Escalera, Shanshan Zhang, Shuyuan Yang, Sunghwan Moon, Thomas B. Moeslund, Vasyl Shandyba, Vladimir Golovkin, Wei Dai, WonTaek Chung, Xinyu Liu, Yongqiang Zhu, Youngseo Kim, Yuan Li, Yuting Yang, Yuxuan Xiao, Zehua Cheng, and Zhihao Li. Soccernet 2024 challenges results. arXiv preprint arXiv:2409.10587, 2024. 2, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 217, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 293 + ], + "type": "text", + "content": "[10] Adrien Deliege, Anthony Cioppa, Silvio Giancola, Meisam J Seikavandi, Jacob V Dueholm, Kamal Nasrollahi, Bernard Ghanem, Thomas B Moeslund, and Marc Van Droogenbroeck. Soccernet-v2: A dataset and benchmarks for holistic understanding of broadcast soccer videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 4508-4519, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 293, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 293, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 317, + 293, + 553, + 369 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations, 2021. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 371, + 553, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 371, + 553, + 423 + ], + "spans": [ + { + "bbox": [ + 317, + 371, + 553, + 423 + ], + "type": "text", + "content": "[12] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 3, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 426, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 426, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 317, + 426, + 553, + 480 + ], + "type": "text", + "content": "[13] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6824-6835, 2021. 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 482, + 553, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 545 + ], + "type": "text", + "content": "[14] Sushant Gautam, Mehdi Houshmand Sarkhoosh, Jan Held, Cise Midoglu, Anthony Cioppa, Silvio Giancola, Vajira Thambawita, Michael A Riegler, Pål Halvorsen, and Mubarak Shah. Soccernet-echoes: A soccer game audio commentary dataset. arXiv preprint arXiv:2405.07354, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 548, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 602 + ], + "type": "text", + "content": "[15] Silvio Giancola, Mohieddine Amine, Tarek Dghaily, and Bernard Ghanem. Soccernet: A scalable dataset for action spotting in soccer videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 1711-1721, 2018. 1, 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 603, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 603, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 317, + 603, + 553, + 646 + ], + "type": "text", + "content": "[16] Xiaofan Gu, Xinwei Xue, and Feng Wang. Fine-grained action recognition on a novel basketball dataset. In International Conference on Acoustics, Speech, and Signal Processing, pages 2563-2567, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 690 + ], + "type": "text", + "content": "[17] Tengda Han, Weidi Xie, and Andrew Zisserman. Temporal alignment networks for long-term video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2906-2916, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[18] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. Autoad: Movie description in" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8392" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[19] Tengda Han, Max Bain, Arsha Nagrani, Gul Varol, Weidi Xie, and Andrew Zisserman. Autoad ii: The sequel-who, when, and what in movie audio description. In Proceedings of the International Conference on Computer Vision, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 142, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 142, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 142, + 294, + 185 + ], + "type": "text", + "content": "[20] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. Autoad iii: The prequel - back to the pixels. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 231 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 231 + ], + "type": "text", + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 294, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 294, + 298 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 294, + 298 + ], + "type": "text", + "content": "[22] Jan Held, Anthony Cioppa, Silvio Giancola, Abdullah Hamdi, Bernard Ghanem, and Marc Van Droogenbroeck. Vars: Video assistant referee system for automated soccer decision making from multiple views. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 5085-5096, 2023. 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 300, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 300, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 300, + 294, + 365 + ], + "type": "text", + "content": "[23] Jan Held, Hani Itani, Anthony Cioppa, Silvio Giancola, Bernard Ghanem, and Marc Van Droogenbroeck. X-vars: Introducing explainability in football refereeing with multimodal large language models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 3267-3279, 2024. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 366, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 294, + 421 + ], + "type": "text", + "content": "[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In Proceedings of the International Conference on Learning Representations, 2022. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 477 + ], + "type": "text", + "content": "[25] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In Proceedings of the International Conference on Machine Learning, pages 4651-4664, 2021. 5, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "type": "text", + "content": "[26] Haopeng Li, Andong Deng, Qiuhong Ke, Jun Liu, Hossein Rahmani, Yulan Guo, Bernt Schiele, and Chen Chen. Sports-qa: A large-scale video question answering benchmark for complex and professional sports. arXiv preprint arXiv:2401.01505, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 535, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 590 + ], + "type": "text", + "content": "[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the International Conference on Machine Learning, pages 12888-12900, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 591, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 294, + 646 + ], + "type": "text", + "content": "[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the International Conference on Machine Learning, pages 19730–19742, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "text", + "content": "[29] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, Limin Wang, and Yu Qiao. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[30] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In Proceedings of the European Conference on Computer Vision, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "text", + "content": "[31] Zeqian Li, Qirui Chen, Tengda Han, Ya Zhang, Yanfeng Wang, and Weidi Xie. Multi-sentence grounding for long-term instructional video. In Proceedings of the European Conference on Computer Vision, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "text", + "content": "[32] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, 2004. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 198, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 230 + ], + "type": "text", + "content": "[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Proceedings of the International Conference on Learning Representations, 2019. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 232, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 297 + ], + "type": "text", + "content": "[34] Hassan Mkhallati, Anthony Cioppa, Silvio Giancola, Bernard Ghanem, and Marc Van Droogenbroeck. Socernetc caption: Dense video captioning for soccer broadcasts commentaries. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 5074-5085, 2023. 1, 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 299, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 553, + 342 + ], + "type": "text", + "content": "[35] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Association for Computational Linguistics, pages 311-318, 2002. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 344, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 553, + 420 + ], + "type": "text", + "content": "[36] Ji Qi, Jifan Yu, Teng Tu, Kunyu Gao, Yifan Xu, Xinyu Guan, Xiaozhi Wang, Bin Xu, Lei Hou, Juanzi Li, et al. Goal: A challenging knowledge-grounded video captioning benchmark for real-time soccer commentary generation. In Proceedings of the ACM International Conference on Information and Knowledge Management, pages 5391-5395, 2023. 1, 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 422, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 465 + ], + "type": "text", + "content": "[37] Mengshi Qi, Yunhong Wang, Annan Li, and Jiebo Luo. Sports video captioning via attentive motion representation and group relationship modeling. IEEE Transactions on Circuits and Systems for Video Technology, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 468, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 553, + 533 + ], + "type": "text", + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, 2021. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 534, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 553, + 588 + ], + "type": "text", + "content": "[39] Prem N Ramkumar, Bryan C Luu, Heather S Haeberle, Jaret M Karnuta, Benedict U Nwachukwu, and Riley J Williams. Sports medicine and artificial intelligence: a primer. The American Journal of Sports Medicine, 50(4): 1166-1174, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 590, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 644 + ], + "type": "text", + "content": "[40] Jiayuan Rao, Haoning Wu, Chang Liu, Yanfeng Wang, and Weidi Xie. Matchtime: Towards automatic soccer game commentary generation. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2024. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "text", + "content": "[41] Dian Shao, Yue Zhao, Bo Dai, and Dahua Lin. Finegym: A hierarchical video dataset for fine-grained action understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2616-2625, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "type": "text", + "content": "[42] Yan Shu, Zheng Liu, Peitian Zhang, Minghao Qin, Junjie Zhou, Zhengyang Liang, Tiejun Huang, and Bo Zhao." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8393" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "text", + "content": "Video-xl: Extra-long vision language model for hour-scale video understanding. arXiv preprint arXiv:2409.14485, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 185 + ], + "type": "text", + "content": "[43] Vladimir Somers, Victor Joos, Anthony Cioppa, Silvio Giancola, Seyed Abolfazl Ghasemzadeh, Floriane Magera, Baptiste Standaert, Amir M Mansourian, Xin Zhou, Shohreh Kasaei, et al. Soccernet game state reconstruction: End-to-end athlete tracking and identification on a minimap. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 3293-3305, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 186, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 229 + ], + "type": "text", + "content": "[44] Graham Thomas, Rikke Gade, Thomas B Moeslund, Peter Carr, and Adrian Hilton. Computer vision for sports: Current applications and research topics. Computer Vision and Image Understanding, 159:3-18, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 231, + 294, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 284 + ], + "type": "text", + "content": "[45] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the International Conference on Computer Vision, pages 4489-4497, 2015. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 342 + ], + "type": "text", + "content": "[46] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 343, + 295, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 295, + 387 + ], + "type": "text", + "content": "[47] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4566-4575, 2015. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 295, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 295, + 442 + ], + "type": "text", + "content": "[48] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 444, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 295, + 498 + ], + "type": "text", + "content": "[49] Zhe Wang, Petar Velicković, Daniel Hennes, Nenad Tomasev, Laurel Prince, Michael Kaisers, Yoram Bachrach, Romuald Elie, Li Kevin Wenliang, Federico Piccinini, et al. Tacticai: an ai assistant for football tactics. Nature Communications, 15(1):1-13, 2024. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 500, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 295, + 533 + ], + "type": "text", + "content": "[50] Dekun Wu, He Zhao, Xingce Bao, and Richard P Wildes. Sports video analysis on large-scale data. In Proceedings of the European Conference on Computer Vision, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 295, + 578 + ], + "type": "text", + "content": "[51] Zeyu Xi, Ge Shi, Xuefen Li, Junchi Yan, Zun Li, Lifang Wu, Zilin Liu, and Liang Wang. A simple yet effective knowledge guided method for entity-aware video captioning on a basketball benchmark. Neurocomputing, 2025. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 295, + 624 + ], + "type": "text", + "content": "[52] Zeyu Xi, Ge Shi, Haoying Sun, Bowen Zhang, Shuyi Li, and Lifang Wu. Eika: Explicit & implicit knowledge-augmented network for entity-aware sports video captioning. Expert Systems with Applications, 2025. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 690 + ], + "type": "text", + "content": "[53] Haotian Xia, Zhengbang Yang, Yuqing Wang, Rhys Tracy, Yun Zhao, Dongdong Huang, Zezhi Chen, Yan Zhu, Yuanfang Wang, and Weining Shen. Sportqa: A benchmark for sports understanding in large language models. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 295, + 714 + ], + "type": "text", + "content": "[54] Haotian Xia, Zhengbang Yang, Junbo Zou, Rhys Tracy, Yuqing Wang, Chi Lu, Christopher Lai, Yanjun He, Xun" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 530 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 118 + ], + "type": "text", + "content": "Shao, Zhuoqing Xie, et al. Sportu: A comprehensive sports understanding benchmark for multimodal large language models. In Proceedings of the International Conference on Learning Representations, 2025. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 173 + ], + "type": "text", + "content": "[55] Jinglin Xu, Yongming Rao, Xumin Yu, Guangyi Chen, Jie Zhou, and Jiwen Lu. Finediving: A fine-grained dataset for procedure-aware action quality assessment. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2949-2958, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 239 + ], + "type": "text", + "content": "[56] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 241, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 284 + ], + "type": "text", + "content": "[57] Huanyu Yu, Shuo Cheng, Bingbing Ni, Minsi Wang, Jian Zhang, and Xiaokang Yang. Fine-grained video captioning for sports narrative. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 285, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 329 + ], + "type": "text", + "content": "[58] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the International Conference on Computer Vision, pages 11975-11986, 2023. 2, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "text", + "content": "[59] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 376, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 430 + ], + "type": "text", + "content": "[60] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 431, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 431, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 431, + 553, + 475 + ], + "type": "text", + "content": "[61] Xin Zhou, Le Kang, Zhiyu Cheng, Bo He, and Jingyu Xin. Feature combination meets attention: Baidu soccer embeddings and transformer based temporal detection. arXiv preprint arXiv:2106.14447, 2021. 2, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 475, + 553, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 553, + 530 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 553, + 530 + ], + "type": "text", + "content": "[62] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8394" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Towards Visual Discrimination and Reasoning of Real-World Physical Dynamics_ Physics-Grounded Anomaly Detection/e496d86c-0d1a-42d1-b098-7f552a70c4ff_content_list.json b/2025/Towards Visual Discrimination and Reasoning of Real-World Physical Dynamics_ Physics-Grounded Anomaly Detection/e496d86c-0d1a-42d1-b098-7f552a70c4ff_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..31d4a5771beb8e6d3a37f05e51ccb22952ee5083 --- /dev/null +++ b/2025/Towards Visual Discrimination and Reasoning of Real-World Physical Dynamics_ Physics-Grounded Anomaly Detection/e496d86c-0d1a-42d1-b098-7f552a70c4ff_content_list.json @@ -0,0 +1,1689 @@ +[ + { + "type": "text", + "text": "Towards Visual Discrimination and Reasoning of Real-World Physical Dynamics: Physics-Grounded Anomaly Detection", + "text_level": 1, + "bbox": [ + 94, + 128, + 908, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenqiao Li $^{1*}$ Yao Gu $^{1*}$ Xintao Chen $^{1*}$ Xiaohao Xu $^{2\\dagger}$ Ming Hu $^{3}$ Xiaonan Huang $^{2}$ Yingna Wu $^{1\\dagger}$ $^{1}$ ShanghaiTech University $^{2}$ University of Michigan, Ann Arbor $^{3}$ Monash University", + "bbox": [ + 99, + 200, + 944, + 241 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/75d6ea3cc4000b98b604cf49e2eedc38b5b4f3a012591964ab21612bc71ce942.jpg", + "image_caption": [ + "Figure 1. Towards visual discrimination of physical dynamics in real-world industrial object anomaly detection. We illustrate objects, interactions, and time-sequenced videos from the Physics-Grounded Anomaly Detection dataset: (a) Object; (b) Interaction: Applied actions shown with directional arrows; (c) Video with Physical Dynamics: Temporal sequences showing normal and abnormal states, highlighting anomalies like leaks, misalignments, and cracks. By focusing on the dynamic behaviors of complex objects, we enhance understanding of interactions and failure modes in real-world settings, where both structure and motion contribute to anomaly detection." + ], + "image_footnote": [], + "bbox": [ + 96, + 253, + 535, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/406ce781989bb97be996195c92fcaa5cde7a373227f124df11e22d77c6ab559c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 250, + 906, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 640, + 326, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans detect real-world object anomalies by perceiving, interacting, and reasoning based on object-conditioned physical knowledge. The long-term goal of Industrial Anomaly Detection (IAD) is to enable machines to autonomously replicate this skill. However, current IAD algorithms are largely developed and tested on static, semantically simple datasets, which diverge from real-world scenarios where physical understanding and reasoning are essential. To bridge this gap, we introduce the Physics Anomaly Detection (Phys-AD) dataset, the first large-scale, real-world, physics-grounded video dataset for industrial anomaly detection. Collected using a real robot arm and motor, Phys-AD provides a diverse", + "bbox": [ + 88, + 674, + 485, + 856 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "set of dynamic, semantically rich scenarios. The dataset includes more than 6400 videos across 22 real-world object categories, interacting with robot arms and motors, and exhibits 47 types of anomalies. Anomaly detection in Phys-AD requires visual reasoning, combining both physical knowledge and video content to determine object abnormality. We benchmark state-of-the-art anomaly detection methods under three settings: unsupervised AD, weakly-supervised AD, and video-understanding AD, highlighting their limitations in handling physics-grounded anomalies. Additionally, we introduce the Physics Anomaly Explanation (PAEval) metric, designed to assess the ability of visual-language foundation models to not only detect anomalies but also provide accurate explanations for their underlying physical causes. Our project is available at https://guyao2023.github.io/Phys-AD/.", + "bbox": [ + 511, + 641, + 910, + 869 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equally contribute to this work.", + "bbox": [ + 107, + 875, + 284, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors", + "bbox": [ + 109, + 887, + 238, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "30409", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fc4dc8c941e0683bab05b70a96d80230028a00681e4d6d0074810515380bb537.jpg", + "image_caption": [ + "Figure 2. Human-like decision-making process for physics-grounded object anomaly detection. We illustrate the sequential approach of a human-like agent for evaluating an object's normality. First, the agent perceives relevant physical attributes (e.g., plastic and elastic), then interacts by performing a physical action (e.g., squeezing), and finally reasons based on the vision feedback and attributes changes (e.g., surface shape change) to determine whether the object is normal or anomalous. This mirrors a human's natural process of reasoning over physics in objects." + ], + "image_footnote": [], + "bbox": [ + 101, + 90, + 478, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 467, + 222, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Industrial anomaly detection (IAD) is a critical subfield in computer vision and industrial automation, aiming to identify defects or irregularities in products during manufacturing. As shown in Fig. 2, the ultimate vision is to create autonomous systems that not only perceive but also interact with and reason about objects to discriminate between anomalies and normal states, integrating complex physical principles to detect anomalies in dynamic, real-world scenarios. For example, a human inspecting a water bottle for anomalies wouldn't rely solely on visual observation; they might rotate or invert the bottle, using physical interactions and cues, such as noticing a loose cap or an irregular internal flow, to detect issues.", + "bbox": [ + 89, + 492, + 483, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Central to advancing IAD [9] is the availability of high-quality datasets that bridge the gap between academic research and industrial needs. Datasets like MVTec-AD [4], MPDD [21], and VisA [65] have played foundational roles, enabling algorithm development for image-based anomaly detection and bringing IAD to the forefront of computer vision research. While these datasets have significantly advanced single-image anomaly detection, recent datasets, e.g., MVTec-3D [7], Real3D [29], and Anomaly-ShapeNet [27], have extended IAD to 3D, aligning research more closely with the needs of complex real-world industrial settings.", + "bbox": [ + 89, + 688, + 483, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Yet, as factories increasingly rely on robotic arms and automated systems to perform sophisticated inspections, the limitations of current IAD datasets become apparent.", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing benchmarks focus on static, semantically simple environments, overlooking the physical priors and interactive reasoning required in real-world industrial contexts. This gap highlights a growing need for datasets that not only reflect real-world physical constraints but also challenge models to reason dynamically about anomalies.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To bridge this gap, as shown in Fig. 1, we introduce the Physics Anomaly Detection (Phys-AD) dataset, the first large-scale, physics-grounded video dataset for industrial anomaly detection. Phys-AD features over 6,400 videos of 22 categories and 49 objects interacting with robotic arms and motors, capturing 47 types of anomalies that require visual reasoning informed by physical knowledge. The short video clips in the dataset range from 60 to 240 frames in length and are filmed in real-time industrial environment, fully capturing the interaction process between robotic arm or motor and industrial objects. Additionally, to ensure our dataset meets industrial demands and matches the complexity and diversity of the real physical world, we selected industrial objects of different physical qualities, various interaction methods, and anomalies that reflect diverse physical principles and require different reasoning process. Specifically, we selected 22 object categories spanning across metals, plastics, fluid, amorphous substances and articulated objects with diverse appearances. For interaction, we use mechanical grippers, robotic arms, and motors, incorporating various interaction modes such as pressing, rotating, squeezing, and driving to handle different types of objects.", + "bbox": [ + 511, + 184, + 906, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We benchmark state-of-the-art anomaly detection methods in three key configurations: unsupervised anomaly detection, weakly-supervised anomaly detection, and video-understanding based anomaly detection. Our findings reveal critical gaps in their ability to handle the complexities of physics-grounded scenarios, where anomalies often arise from dynamic, interdependent interactions. To advance the field, we also introduce the Physical Anomaly Explanation (PAEval) metric, designed to assess both detection performance and a model's ability to explain anomalies by identifying underlying physical causes. Furthermore, our benchmark reveals the fragility of existing methods in tackling these challenging conditions, underscoring the need for approaches that better understand object dynamics and temporal coherence in anomaly detection.", + "bbox": [ + 511, + 518, + 908, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as followings:", + "text_level": 1, + "bbox": [ + 531, + 747, + 901, + 762 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel task of detecting physical-based industrial anomalies in real-world that involves perception, physical and visual reasoning.", + "- We present Phys-AD, the first large-scale, physics-grounded video dataset specifically designed for industrial anomaly detection in real world, containing objects of different physical qualities, multiple interaction methods and various physical reasoning process.", + "- We benchmarking the anomaly detection and reasoning" + ], + "bbox": [ + 511, + 763, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "30410", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/2affb83bc3d5314d7dc6b2b25346b44074ea92e0f46ef8701d79e44a343e635b.jpg", + "table_caption": [ + "Table 1. Comparison of Phys-AD with existing industrial anomaly detection datasets. Our Phys-AD dataset is the first to consider complex objects with physical dynamics. 'Syn', 'IR', 'D', and 'PC' denote Synthetic, Infrared, Depth, and Point Cloud, respectively. #Anomaly indicates the number of anomaly types." + ], + "table_footnote": [], + "table_body": "
DatasetYearTypeModalitySample Statistics
Class #AnomalyPhysics
MVTec-AD [3]2019RealRGB15-X
BTAD [38]2021RealRGB33X
MPDD [21]2021RealRGB68X
VisA [66]2021RealRGB12-X
MVTec LOCO-AD [5]2022RealRGB5-X
MAD [64]2023Syn+RealRGB203X
LOCO-Annotations [62]2024RealRGB55X
Real-IAD [46]2024RealRGB308X
GDXray [37]2015RealX-ray515X
PVEL-AD [43]2023RealIR110X
MVTec3D-AD [6]2021RealRGB-D103-5X
Eyecandies [8]2022SynRGB-D103X
Real3D-AD [30]2023RealPC122X
Anomaly-ShapeNet [27]2024SynPC406X
Phys-AD (Ours)2024RealRGB4947✓
", + "bbox": [ + 91, + 161, + 483, + 392 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "performance of popular video AD methods and Visual Language Foundation Models on the Phys-AD dataset in several settings, establishing a practical and challenging benchmark to promote the development of the physics-related anomaly detection field.", + "bbox": [ + 102, + 419, + 483, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 511, + 227, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Industrial anomaly detection datasets. Existing datasets primarily focus on static, semantically simple scenarios, deviate significantly from real world where physical understanding and reasoning are essential. Datasets like MVTec-AD [3], BTAD [38], MPDD [21], and VisA [66] focus on surface level image anomaly detection with one single-view RGB image, limiting their effectiveness in capturing holistic object structures. MVTec-LOCO-AD [5], focusing on structure and local information in industries, is limited by its relatively simple and constrained data content. While MVTec3D-AD [6] and Eyecandies [8] incorporate depth data, they remain static and single-view image anomaly detection, neglecting object level information. To explore object level anomaly detection, multi-view IAD datasets like MAD [64] and Real-IAD [46], point cloud IAD datasets like Real3D-AD [30] and Anomaly-ShapeNet [27], offering richer visual and geometric cues, but are still limited to static objects without dynamic interaction or reasoning. In summary, current industrial anomaly detection datasets focus on relatively simple and static anomaly detection scenarios, lacking complex physical rules, dynamic interactions, and visual reasoning requirements. Therefore, existing IAD datasets generally applied to limited industrial scenarios and there is no IAD dataset could meet the demands of detecting", + "bbox": [ + 89, + 537, + 485, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "complex anomalies in real world which need physical priors and reasoning.", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video anomaly detection. Deep learning methods [18, 35, 54, 61] now dominate Video Anomaly Detection (VAD), categorized into unsupervised, weakly-supervised, and fully-supervised approaches. Unsupervised methods learn normal patterns via reconstruction [16, 18, 52], prediction [32], or hybrids [33], while some methods [45, 55] train with both unlabeled normal and abnormal data. Weakly-supervised methods[14, 49, 60, 63] use video-level or glance-based annotations, and fully-supervised methods[25, 31] remain rare due to costly frame-level labeling. Visual-language models like CLIP [42] have recently been applied to enhance anomaly detection [24, 41, 51], focusing on semantic anomalies. Open-vocabulary VAD [50] and prompt-based anomaly scoring [56] leverage LLMs [53, 61], but performance relies heavily on the base models, often lacking domain-specific tuning. However, existing video anomaly detection algorithms lack the capability to handle complex industrial anomaly detection scenarios and understand physical rules. This gap highlights the need for models that can capture dynamic behaviors and physical laws in industrial environments, as addressed by Phys-AD, which targets industrial anomalies in objects of various physical properties.", + "bbox": [ + 511, + 125, + 908, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual reasoning. Visual reasoning is a critical task in computer vision, aiming to enable machines to interpret perceptual information like humans. Several visual reasoning tasks have been proposed to evaluate reasoning capabilities, including Visual Question Answering (VQA), 2D puzzles, and physical dynamics prediction. In VQA, agents are required to combine natural language and visual cues to answer questions [1, 17, 19, 23]. For 2D puzzles, tasks involve discovering relationships among visual elements and making inferences [22, 26, 39, 57, 58]. Physical dynamics prediction tasks require machines to perceive and reason about physical interactions [2, 12, 13, 20]. In contrast to these work, we introduce the first benchmark featuring real-world industrial objects with dynamic physical properties, focusing on distinguishing diverse dynamics through vision.", + "bbox": [ + 511, + 460, + 910, + 690 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Dataset: Phys-AD", + "text_level": 1, + "bbox": [ + 511, + 702, + 689, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Object Preparation and Interaction Selection", + "text_level": 1, + "bbox": [ + 511, + 727, + 897, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We selected 22 categories spanning across various materials, including metals, plastics, amorphous substances and articulated objects, with diverse shapes, sizes, and physical properties. For the physical properties of different objects, we correspondingly select different interaction methods like push, rotate, pull with robotics arms and motors. For instance, we use robotics arms to grab and extruding the deformable objects like basketball to determine whether there are any elastic anomalies or surface defects by the morphological change of the objects. To make our dataset more practical", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "30411", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d1fb0c57315c70053b2eee90ea39523b6bc62eca99f6e996b4571a08712f0d89.jpg", + "image_caption": [ + "U Disk" + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 225, + 145 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/04c3a0c007d031f41abd5c7c98604aad9ab1aaf4129e18d28427e92deee1eb37.jpg", + "image_caption": [ + "Zipper" + ], + "image_footnote": [], + "bbox": [ + 227, + 88, + 359, + 145 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/45bcc2f974a31f7bc5fa856043b21b280954cb11438e7f549aef3542277029ad.jpg", + "image_caption": [ + "Rolling Bearing" + ], + "image_footnote": [], + "bbox": [ + 364, + 89, + 496, + 146 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b7d51d7075f154f89366e928da49136a71a85cee7477c389e2eb7f9afe4c02c5.jpg", + "image_caption": [ + "Sticky Roller" + ], + "image_footnote": [], + "bbox": [ + 501, + 90, + 632, + 147 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/138a01ee9acd66f4fd6acc7aa80c8d225c1ae73de2587940ad0378ef6c70897b.jpg", + "image_caption": [ + "Slider" + ], + "image_footnote": [], + "bbox": [ + 637, + 90, + 767, + 146 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6b60ccaaf439eb96ea39c5a3b9fa413b510be46701455468c04e51538753bdc3.jpg", + "image_caption": [ + "Toothpaste" + ], + "image_footnote": [], + "bbox": [ + 772, + 90, + 903, + 146 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0b3e191f36b3ed06901894a1493699ae31532b6433e29d9994adcf9e77502b7d.jpg", + "image_caption": [ + "Door Lock" + ], + "image_footnote": [], + "bbox": [ + 93, + 164, + 225, + 222 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/de166171b1fd2c1d1d14cfe9f71d84aeee36e3408be644fb0740baed93afadf8.jpg", + "image_caption": [ + "Toy Car" + ], + "image_footnote": [], + "bbox": [ + 228, + 164, + 359, + 222 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/90f777c39de74347453e303735bb358ab7dd2c4abc98c699069be3d12fd72216.jpg", + "image_caption": [ + "Clip" + ], + "image_footnote": [], + "bbox": [ + 364, + 164, + 496, + 222 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1324e3603fe14c0ee8596a894f5b4b71efea4e5df255eef133e2c529b00767b4.jpg", + "image_caption": [ + "Fan" + ], + "image_footnote": [], + "bbox": [ + 501, + 165, + 632, + 222 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/79a5bcee9d78b800f3c96612ba744acd04d2d65ab5444af43ed720b70edc3bcb.jpg", + "image_caption": [ + "Rubber Band" + ], + "image_footnote": [], + "bbox": [ + 637, + 165, + 767, + 222 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a93b776c01182baf642c844fd4003427bef1d20f803d1921c7b718ef3b4de6e3.jpg", + "image_caption": [ + "Magnet" + ], + "image_footnote": [], + "bbox": [ + 772, + 165, + 903, + 223 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ef6d8ff60375a009ba23164350c9f70fa0b56f37d63ffd49a93da1a6e504579a.jpg", + "image_caption": [ + "Servo" + ], + "image_footnote": [], + "bbox": [ + 91, + 244, + 223, + 301 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4d6dc56067eb39e38a2c6d664695e2645359b85f86b75c0ba77d4771ba20a87f.jpg", + "image_caption": [ + "Gear Assembly" + ], + "image_footnote": [], + "bbox": [ + 228, + 244, + 359, + 301 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/143b400b63356474d7a583349339b8c3fd2ca6429bf0a91836a0f9968ad6acbb.jpg", + "image_caption": [ + "Bottled Water" + ], + "image_footnote": [], + "bbox": [ + 364, + 244, + 496, + 301 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/db4e4ea50cb523a34663c3de3d8555feba8fc3e602d608de9919f4a00022692f.jpg", + "image_caption": [ + "Screw" + ], + "image_footnote": [], + "bbox": [ + 501, + 244, + 632, + 301 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5d07bbf049c4ab67cb7cefb294f93b1adecd62b7b746b16a2b25c80a1f2688ff.jpg", + "image_caption": [ + "Spherical Bearing" + ], + "image_footnote": [], + "bbox": [ + 637, + 244, + 767, + 301 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ae3b8ab1d44dca0c371de14f0f6e04fbffda0755c5a46cb2291463b29b4ca39e.jpg", + "image_caption": [ + "Motion" + ], + "image_footnote": [], + "bbox": [ + 792, + 253, + 888, + 276 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/30b520f00ef352421bb17433f5d67fdc49435e0a6a68430c5acb0b5e28c6e3da.jpg", + "image_caption": [ + "Button" + ], + "image_footnote": [], + "bbox": [ + 93, + 325, + 223, + 388 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d477cbe0731ea4b43372c2b74c2807e6242f0ecfdb1ee561599aa52988b7e88b.jpg", + "image_caption": [ + "Clock" + ], + "image_footnote": [], + "bbox": [ + 228, + 325, + 359, + 388 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/32405e8d76a58bd525059e3dbd12ba4e95ee2af92a48352e8327c8bf0cb5a0a5.jpg", + "image_caption": [ + "Caster Wheel" + ], + "image_footnote": [], + "bbox": [ + 364, + 325, + 493, + 388 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d020a461c07f5b611e5475dca37bd6451a9ee9fd00ee94a5502173746bfcce9d.jpg", + "image_caption": [ + "Hinge" + ], + "image_footnote": [], + "bbox": [ + 501, + 325, + 632, + 388 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5d39dd0d1961330a708b001bbd0af0330e7e2274175ee736aa65e240d83b341e.jpg", + "image_caption": [ + "Ball" + ], + "image_footnote": [], + "bbox": [ + 637, + 325, + 769, + 388 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ee9640bd724f2776c510eb025a226563c9518fd3ee2fc98b17e081ca1a1b534d.jpg", + "image_caption": [ + "Direction For Interaction", + "Figure 3. Interactions for understanding implicit physical laws in the Phys-AD dataset. We showcase various object interactions from the Phys-AD dataset, where different actions (indicated by motion directions) are used to explore and reason about the underlying physical properties and behaviors of each object. The colored arrows indicate the interaction directions and axes, highlighting how physical interactions reveal the implicit physics governing each object." + ], + "image_footnote": [], + "bbox": [ + 772, + 325, + 903, + 398 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b3f99f89cbed8591fccb3110e6a970871e8521d23d9bbcbaaea6ed729034642a.jpg", + "image_caption": [ + "Normal: The U Disk rotates freely over time." + ], + "image_footnote": [], + "bbox": [ + 94, + 491, + 493, + 558 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6768b2f107bdf1fc48650b414447219c6c19cc384203b8c07f0ee7c7436e2cb8.jpg", + "image_caption": [ + "Abnormal: The U Disk is stuck over time.", + "(a) Persistent Anomalies" + ], + "image_footnote": [], + "bbox": [ + 94, + 597, + 493, + 662 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fc43f6a365f2a63ec36586f1a0cf725753bce008123e8aed6dd463dcf945a622.jpg", + "image_caption": [ + "Normal: The sticky roller rotates smoothly on the surface." + ], + "image_footnote": [], + "bbox": [ + 504, + 491, + 901, + 558 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0fd51d148225da32c20aed5e679cb20d87bc532ca66ea2e17693bcb16094c8c2.jpg", + "image_caption": [ + "Normal: The sticky roller functions properly at the beginning.", + "Abnormal: The sticky roller breaks at a certain point.", + "(b) Intermittent Anomalies", + "Figure 4. Categorization of anomalies based on persistence in the Phys-AD dataset. We show examples of normal and abnormal functioning in common objects, divided into two anomaly types: persistent and intermittent. (a) Persistent anomalies, such as continuous obstruction in the U Disk or permanent malfunction of the Sticky Roller, are visible throughout the operation. (b) In contrast, intermittent anomalies, like occasional jamming of the U Disk or breakage in the Sticky Roller after initial operation, only appear at specific points in time. This classification provides insight into both constant and sporadic failures in object interactions." + ], + "image_footnote": [], + "bbox": [ + 504, + 599, + 901, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and challenging, we introduce 47 distinct defect types, some of which just rely on single frame content and physical rules for anomaly reasoning, while the other part needs to combine the content of the whole video and physical priors together to judge whether the object is abnormal or not. Different objects information with their corresponding interaction methods and anomalies are listed in Table 2. Fig 3 shows the interac", + "bbox": [ + 89, + 782, + 485, + 888 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tions for understanding implicit physical laws in the Phys-AD dataset. Fig 4 provides two representative examples: one where a U Disk requires analyzing the entire video content to determine if there is an anomaly, and another where a sticky roller only requires a few frames from the video to judge whether the object is abnormal.", + "bbox": [ + 511, + 782, + 906, + 873 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "30412", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c3ffd1eb93313cb723ec32e3219be946461890aafac571d6445aa6c70c56801c.jpg", + "image_caption": [ + "Figure 5. Data collection pipeline for the Phys-AD dataset. (a) Manipulation of a toothpaste tube using a UR5 robotic arm. (b) Manipulation of a U Disk and fan via servo and motor." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 287, + 208 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e4add3a5da0de42ca6b80027e5c1fbdd5fd5651b741bdb4da4c922c84fbf789b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 289, + 90, + 480, + 207 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/cbcf4e193c05768860b9b372ece0ae1bbe405dad03bd5e2829372390fe2dbb53.jpg", + "table_caption": [ + "Table 2. Interaction methods and anomalies of the object categories in our Phys-AD dataset. Motor. means one object driven by motor. Multi. means one object with multiple anomalies." + ], + "table_footnote": [], + "table_body": "
CategoryInteractionAnomaly Types
CarDrag, SlideDifferent wheels stuck
FanMotor., RotateStuck, Uneven Rotation, Vibration
Rolling BearingMotor., RotateLack of friction
Spherical BearingGrab, RotateInternal block
ServoGrab, Motor., RotateAngle restricted, Vibration, No calibration
ClipPressUnable to press down, Unable to rebound
U DiskGrab, Motor., RotateCover jam
HingeGrab, RotateAngle restricted, Shaft off
Sticky RollerGrab, PullDetach, Unable to rotate
Caster WheelSlideAxle axis stuck, Swivel axis stuck
ScrewPress, RotateLoosen, Unable to insert
LockMotor., RotateLatch jam, Loose
GearMotor., RotateStuck, Not meshed, Multi.
ClockMotor., RotatePointer stops, Vibration
SlideGrab, SlideDetach, Shedding, Jam
ZipperGrab, CloseStuck, Unable to close
ButtonPressNo light, Unable to press down, Unable to rebound, Multi.
LiquidGrab, ShakeWater out, Foreign body
Rubber BandStretchCrack
BallPinch, PressInsufficient gas, Leakage
MagnetGrab, Press, MoveDegaussing, Shell detached
ToothpastePinch, PressLeakage
", + "bbox": [ + 94, + 319, + 480, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Data Collection and Processing", + "text_level": 1, + "bbox": [ + 89, + 694, + 366, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data collection pipeline. Most of the data collection for the Phys-AD dataset is driven by manipulation-guided video sequences, captured using a UR5 robotic arm equipped with an RGB camera (see Fig. 5a). In order to reproduce real industrial scenes, we also include adjustable light sources in our data capture process. The RGB Camera, with a 1080p resolution and a frame rate of 60 FPS, providing high-quality video sequences. Some kinds of the objects like U Disk that are not suitable for the manipulation of robotics arms, are driven by the motor or servo (Fig. 5b). After the data collection, we used video editing software to crop out irrelevant frames and retain the complete interaction process.", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5c88e1548cab93edd3fe5a31ee082684208e011607747650969dc1fda726e079.jpg", + "table_caption": [ + "Table 3. Phys-AD dataset statistics. We denote the total number of frames and videos for each category as #Images and #Videos. Note that the Train split does not contain anomaly data." + ], + "table_footnote": [], + "table_body": "
CategoryTrain #FramesTrain #VideosTest #FramesTest #VideosTest #Videos Nor.Test #Videos Ab.Obj. TypesDef. Types
Car18,00030036,00060015045053
Fan32,40018064,8003609027033
Rolling Bearing3,600603,60060303011
Spherical Bearing3,600603,60060303011
Servo14,40012028,8002406018013
Clip28,80024043,20036012024042
U Disk14,40024014,40024012012041
Hinge3,600307,20060154512
Sticky Roller5,400308,10045153012
Caster Wheel5,4003010,80060154512
Screw5,400308,10045153012
Lock7,20012010,8001806012012
Gear21,60018054,0004509036034
Clock27,00015040,3202247315152
Slide7,2006018,0001503012013
Zipper14,40012021,6001806012022
Button21,60012054,0003006024044
Liquid5,400308,10045153012
Rubber Band10,8006010,80060303011
Ball21,6009032,400135453032
Magnet10,8006016,20090306022
Toothpaste16,2009016,20090454531
Total298,8002400511,0204,0341,1982,8364947
", + "bbox": [ + 516, + 133, + 903, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Data Statistics", + "text_level": 1, + "bbox": [ + 513, + 376, + 661, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset sample distribution. Table 3 provides a detailed breakdown of the Phys-AD dataset, which includes information on category, the number of training videos and frames, the number of testing videos and frames, the distribution of normal and abnormal samples in the test set, and the number of object and defect types. The length of the video clips in the dataset ranges from 60 to 240 frames, ensuring to fully capture the interaction process in a short time. The dataset contains 2400 training videos and 4034 test videos spreading across 22 categories, 49 object types, and 47 defect types. The frames are extracted from the videos at a rate of 60 FPS. The test set includes 1,198 normal samples and 2,836 abnormal samples. Overall, the Phys-AD dataset covers a wide range of categories and contains a large scale of data, which helps to train more robust anomaly detection models and provides a reasonable evaluation setting.", + "bbox": [ + 509, + 398, + 906, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Interaction and defect types across categories. Table 2 details the defect types and corresponding interaction methods for each category in the Phys-AD dataset. Covering 22 categories and including 47 defect types, the dataset features various interactions (e.g., Rotate, Grab). Importantly, interaction methods or defect types with identical names can differ across categories—for example, \"rotation\" for a fan describes the motor-driven movement of its blades, while for a hinge it refers to the turning of its pages around the shaft under robotic control. Furthermore, to detect anomalies with complex physical properties in a human-like manner, we sometimes combine multiple interactions. For instance, when inspecting a screw, we first press it into the hole and then use a robotic arm to rotate it further. Overall, Phys-AD provides a diverse, physics-grounded anomaly detection framework that fosters the development of advanced methods for tackling real-world challenges.", + "bbox": [ + 511, + 643, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "30413", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/37b08d3a354aa271391e904fcbd765b6bb939dde01b0e2dacc73c68dd212e63c.jpg", + "image_caption": [ + "Figure 6. PhysAD-Agent: A Large Language Model (LLM)-powered system for physics anomaly detection label augmentation and automatic evaluation. This agent framework consists of two main components: (a) Rich Annotation Generation, where an LLM-based data agent generates detailed anomaly annotations based on video, anomaly state, prompt, and human-annotated physical priors, and (b) Automatic Evaluation, where an LLM-based evaluation agent assesses model predictions to calculate the Physics Anomaly Explanation (PAEval) metric." + ], + "image_footnote": [], + "bbox": [ + 94, + 90, + 480, + 351 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Labels. For the test set videos, we provide two types of labels. First, there are the common video-level labels: all anomalous videos in the test set are labeled as 1, while normal videos are labeled as 0. For evaluating visual-language models, we also provide text labels. Specifically, for each type of anomaly in the test set, we manually design a textual description label. This label includes both a description of the video content and a physical explanation of the reason for the anomaly. To ensure diversity in our text labels, we also use ChatGPT-4o for text augmentation. For further details on the labeling process, please refer to Fig 6a.", + "bbox": [ + 89, + 508, + 483, + 675 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Phys-AD Benchmark", + "text_level": 1, + "bbox": [ + 89, + 686, + 290, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Problem Definition and Challenges", + "text_level": 1, + "bbox": [ + 89, + 712, + 395, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We establish unsupervised and weakly supervised AD settings for Phys-AD, using unsupervised as the default in our problem definition.", + "bbox": [ + 89, + 734, + 482, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We formulate our Phys-AD setting as two steps.", + "bbox": [ + 89, + 780, + 406, + 795 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Step 1: Rules deduction Given a set of training objects $\\mathcal{T} = \\{t_i\\}_{i=1}^N$ from category $c_i$ , we first use robotics arms and motors to interact with $\\mathcal{T}$ in corresponding methods $I_i$ , and we get the interaction process in video format as $V_i$ . Then, we feed the video sequence $V_i$ and the corresponding category's physical prior $P_i$ together to the deduction function $f$ to obtain the normal rules $r_i$ . After rules deduction for all the", + "bbox": [ + 89, + 795, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "categories, we get the rules bank $\\mathcal{M} = \\{r_i\\}_{i=1}^N$ . Step 1 can be formulated as below:", + "bbox": [ + 511, + 89, + 906, + 119 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} \\Theta I _ {i} = V _ {i} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 669, + 133, + 906, + 148 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nf \\left(V _ {i} \\mid P _ {i}\\right) = r _ {i} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 663, + 159, + 906, + 175 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} = \\left\\{r _ {i} \\right\\} _ {i = 1} ^ {N} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 665, + 180, + 906, + 199 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\Theta$ represents interaction between objects and robotics arms. $N$ is the total number of categories.", + "bbox": [ + 511, + 205, + 906, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Step 2: Anomaly reasoning During test time, we first use robotics arms and motors to interact with $\\mathcal{T}$ from the test set in corresponding methods $I_{i}$ , and we get the interaction process in video format as $V_{i}$ . Then, we feed the video sequence $V_{i}$ and rules $r_i$ from step 1 to the reasoner $R$ to predict the object's anomaly score $S$ . Step 2 can be formulated as:", + "bbox": [ + 511, + 234, + 906, + 325 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} \\Theta I _ {i} = V _ {i} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 671, + 337, + 906, + 353 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nR \\left(V _ {i} \\mid r _ {i}\\right) = S, r _ {i} \\in \\mathcal {M} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 364, + 906, + 380 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Challenges. Key challenges include combining video information with objects' physical priors to deduct normal rules, capturing long-term temporal dependencies and fine-grained frame-level information, ensuring robust reasoning for anomaly prediction, and generalizing across diverse anomaly patterns. Addressing these challenges is essential for advancing anomaly detection in real-world environment.", + "bbox": [ + 511, + 386, + 908, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation Settings", + "text_level": 1, + "bbox": [ + 511, + 501, + 699, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Unsupervised AD. Unsupervised AD is the most common IAD setting for existing IAD datasets and algorithms. Under unsupervised AD setting, the training set contains only normal video data, and the algorithm needs to capture the normal distribution of the data from the training set. In the test set, both normal and abnormal data are included, and the algorithm must distinguish between normal and abnormal data based on the distribution learned during training.", + "bbox": [ + 511, + 523, + 906, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Weakly-supervised AD. Our dataset is in the form of videos for complex anomaly detection in industrial scenarios. In the video context, it is inevitable that we need to discuss weakly supervised anomaly detection. Under weakly-supervised Phys-AD setting, 2~4 video-level labeled anomaly samples are sampled from all possible anomaly classes in the test set in our Phys-AD dataset. These sampled anomalies are then removed from the test set. It's worth noting that we only provide video level label in our test set. This is because most anomalies in our test set require reasoning based on the entire video content and physical knowledge of the objects.", + "bbox": [ + 511, + 643, + 908, + 810 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Video-Understanding AD. Video-Understanding models are another potential solution to our Phys-AD setting. Similar to unsupervised setting, we only provide normal videos during training. The visual language models need to provide explicit normal rules in language format during training. In the test phase, these video-understanding AD", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "30414", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ce97f53bb8944f82e5793cc25a917db16081023a8abd500324d9f0518cf27c17.jpg", + "table_caption": [ + "Table 4. Video-level AUROC (↑) result of 22 categories on Phys-AD dataset. We include Unsupervised, Weakly-supervised and Video-understanding methods. The best per-category result for each class of methods is highlighted in **bold**. 'ZS ImgB', 'V-ChatGPT', 'V-LLaMA', 'V-LLaVA' denote ZS ImageBind, Video-ChatGPT, Video-LLaMA and Video-LLaVA." + ], + "table_footnote": [], + "table_body": "
Category.UnsupervisedWeakly-supervisedVideo-understanding
MPN [34]MemAE [16]MNAD.p [40]MNAD.r [40]SVM [44]VADClip [51]S3R [48]MGFN [10]LAVAD [56]ZS Clip [42]ZS ImgB [15]V-ChatGPT [36]V-LLaMA [59]V-LLaVA [28]
Car0.2290.5230.4920.9440.5870.5810.6060.5710.5570.5000.5000.5000.6780.522
Fan0.8110.3710.8100.5420.5000.6240.6400.5420.5100.5000.5000.5490.5920.611
Rolling Bearing0.3530.0440.3520.8000.9330.5890.6010.6800.5320.5000.5000.3000.5000.500
Spherical Bearing0.1130.0920.9620.8130.6500.5000.6820.5830.4350.5000.5000.4500.5500.500
Servo0.3640.4450.9750.8780.5000.5180.5920.5560.5020.5000.5000.5060.6830.464
Clip0.5350.4430.6300.3330.5000.4120.5610.5630.5160.5000.5000.6690.5560.458
U Disk0.2400.6170.6090.9400.5000.5300.5490.5700.5130.5000.5000.5650.5750.500
Hinge0.7690.8700.7050.8950.5000.7370.5610.5500.5640.5000.5000.5000.6920.500
Sticky Roller0.9670.9670.4510.9360.5000.5420.8350.6690.2660.5000.5000.4500.5440.467
Caster Wheel0.3640.5230.2710.5080.5000.5870.6760.6760.6150.5000.5000.4440.6420.500
Screw0.5220.5670.6800.5470.5000.5000.6570.5410.6880.5000.5000.4720.2560.550
Lock0.5630.5970.4300.6410.7330.4520.5230.6260.3410.5000.5000.2790.4940.500
Gear0.5290.5100.6520.6940.5000.5190.5800.5440.6030.5000.5000.5440.6030.517
Clock0.3400.5420.3950.5870.5000.5000.5490.5090.5000.5000.5000.5010.3600.500
Slide0.5170.9620.9170.7840.5000.5310.6110.5680.4250.5000.5000.5620.5670.179
Zipper0.8150.5920.8290.4210.5000.5040.6360.6330.5350.5000.5000.5470.4890.500
Button0.8530.3650.6600.5680.5000.6270.5150.5660.4390.5000.5000.5150.5170.360
Liquid0.1840.7000.6710.8310.5000.5420.5950.7930.5040.5000.5000.4100.2780.217
Rubber Band0.3740.3680.3660.3070.5670.4820.6230.6040.5110.5000.5000.5170.4500.450
Ball0.5430.3830.7280.6870.5000.5000.6670.5670.6030.5000.5000.5620.6360.533
Magnet0.6710.4640.6910.4380.5000.5000.5480.7190.5020.5000.5000.6830.3000.400
Toothpaste0.5870.8890.5200.6310.5000.5000.6860.7110.5620.5000.5000.3760.5500.467
Average0.5110.5380.6270.6690.5440.5350.6130.6060.5100.5000.5000.4960.5230.463
", + "bbox": [ + 93, + 136, + 903, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "methods have to truly understand the videos and predict the anomalies based on the language rules withdrawn from the training phase.", + "bbox": [ + 89, + 484, + 480, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 89, + 556, + 274, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Physics anomaly accuracy metrics. We use the Area Under the Receiver Operating Characteristic Curve (AUROC) to evaluate video-level anomaly detection performance. We also report the average precision (AP), i.e. the area under the video-level precision-recall curve and the acc (ACCURACY), following previous works [47, 49].", + "bbox": [ + 89, + 579, + 482, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Physics Anomaly Explanation (PAEval) metric. Directly utilizing existing VLMs [59][36] to understand the videos in our dataset and detect the anomalies is a potential solution to our challenge. The key point is: Can existing VLMs truly understand physical rules and reason in a right way? Specifically designed for video-understanding methods, we introduce a new metric named Physics Anomaly Explanation (PAEval) metric. To be more specific, PAEval evaluate the anomaly detection performance of algorithms based on VLMs from three different perspectives: classification, description, and explanation. Classification refers to the traditional anomaly detection metrics like AUROC, etc. Inspired by the works [11, 61], PAEval also introduces two additional evaluation metrics: description and physical explanation. Description refers to the model's ability to", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "correctly describe the content of the video, used to assess whether the model has the capability to describe physical phenomena. Explanation refers to the model's ability to correctly explain the physical causes of anomalies in the video, used to evaluate the VLM's reasoning ability. To provide labels for description and explanation metrics, we manually labeled each type of defects from all categories and performed data augmentation by ChatGPT to ensure the robustness of the detection. The whole pipeline of PAEval metric is depicted in Fig 6.", + "bbox": [ + 511, + 484, + 906, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Benchmarking Results", + "text_level": 1, + "bbox": [ + 511, + 654, + 725, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Benchmark Methods Selection", + "text_level": 1, + "bbox": [ + 511, + 680, + 784, + 695 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the Phys-AD setting, we select popular and reproducible video anomaly detection methods across unsupervised, weakly-supervised, and video-understanding setting. In the unsupervised setting, we focus on reconstruction, prediction, and embedding-based models like MemAE [16] and MNAD [40], which use memory modules to enhance anomaly discrimination. For weakly-supervised anomaly detection, we adopt methods such as VadCLIP [51] and MGFN [10], which use feature magnitudes and vision-language associations. In video-understanding, we evaluate video-language models like Video-ChatGPT [36] and image-language models like CLIP [42] to predict detailed anomaly descriptions and scores through cross-modal embeddings.", + "bbox": [ + 511, + 703, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "30415", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d8144d5b51de9e11ae4893f04065b9163b95c6453985312f9855c3e67382fa9d.jpg", + "table_caption": [ + "Table 5. Physics Anomaly Explanation (PAEval) metric results on our Phys-AD dataset." + ], + "table_footnote": [], + "table_body": "
MethodsClassification (↑)Description (↑)Explanation (↑)
LAVAD[56]0.5100.1570.000
Video-ChatGPT[36]0.4960.1310.160
Video-LLaMA[59]0.5230.1370.303
Video-LLaVA[28]0.4630.2190.282
", + "bbox": [ + 93, + 119, + 480, + 189 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Code and Experiment details. We provide code and toolkit for our dataset and benchmark. More experiment details are listed in Appendix.", + "bbox": [ + 89, + 209, + 482, + 255 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Main Findings", + "text_level": 1, + "bbox": [ + 89, + 263, + 238, + 280 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Overall anomaly detection benchmarking results. Table 4 shows that existing video anomaly detection and videowhildunderstanding methods achieve limited performance on the Phys-AD dataset, with the highest AUROC only reaching $66.9\\%$ for MNAD.r [40]. This result underscores the heightened complexity of Phys-AD compared to existing datasets, highlighting a domain shift that challenges current industrial anomaly detection algorithms, which are often tuned to visually distinct anomaly patterns in single frames rather than complex temporal or physical cues.", + "bbox": [ + 89, + 285, + 483, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Anomaly detection via unsupervised and weakly supervised methods. Our experiment includes unsupervised methods (e.g., MemAE [16], MPN [34], MNAD [40]) and weakly supervised methods (e.g., S3R [48], MGFN [10], VADClip [51]). Unsupervised methods like MNAD.p [40] performed better on temporal anomalies (81.0% for fan, 68.0% for screw) by leveraging prediction-based approaches, which excel in scenarios requiring temporal understanding. Weakly supervised methods improve baseline scores across complex classes, preventing extreme low scores in challenging categories such as spherical bearing. However, weakly supervised methods show marginally lower performance in simpler anomaly classes, indicating trade-offs introduced by anomaly samples in training.", + "bbox": [ + 89, + 440, + 483, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unsupervised anomaly detection method performance by category. Among unsupervised methods, MemAE [16] achieves high AUROC on objects with spatial anomalies (e.g., sticky roller, $96.7\\%$ ) but underperforms in temporal anomaly classes like rolling and spherical bearings, where it achieves only $4.4\\%$ and $9.2\\%$ , respectively. MNAD [40] improves temporal sensitivity by incorporating temporal prototypes, achieving better scores for bearings but still struggling with complex anomaly types, revealing limitations in purely prototype-based approaches.", + "bbox": [ + 89, + 655, + 482, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Anomaly detection via Video-Language Models. VLM-based methods (e.g., Video-ChatGPT [36], VideoLLaMA [59]) struggle, with top-performing VideoLLaMA [59] reaching only $52.3\\%$ average AUROC. Performance is impacted by reliance on pre-trained weights that are not optimized for physics-grounded video content, evident", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "from low AUROC scores in categories with nuanced physical dynamics, such as hinges and screws. Further, PAEval results suggest that these models lack effective reasoning about object physical dynamics and behaviors influenced by physical forces, underscoring a gap between VLM capabilities and the demands of IAD tasks.", + "bbox": [ + 511, + 90, + 906, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Anomaly explanation via Video-Language Models. Table 5 reports PAEval metric results, with the best-performing VLM achieving only $21.9\\%$ in description and $30.3\\%$ in explanation. These findings emphasize that current VLMs lack the depth in physical reasoning and temporal coherence required for understanding real-world physics-based scenarios, which our Phys-AD dataset demands.", + "bbox": [ + 511, + 185, + 908, + 291 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Summary. Overall, the results reveal Phys-AD's unique challenge in requiring both high spatial detail and temporal comprehension, areas where existing methods and models underperform. This analysis points to the need for future research in models that integrate temporal reasoning with physics-based anomaly detection.", + "bbox": [ + 511, + 294, + 908, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 398, + 633, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce the first industrial anomaly detection task focusing on real-world scenarios where physical understanding and reasoning are essential for anomaly detection. We present the Physics Anomaly Detection (Phys-AD) dataset, a large-scale, physics-grounded video dataset with over 6400 videos across 22 categories and 49 object types interacting with robotic systems, capturing 47 anomaly types that necessitate visual and physical understanding. We assess Phys-AD, highlighting the lack of baseline approaches for high-level reasoning in anomaly detection. Additionally, we propose the Physics Anomaly Explanation (PAEval) metric to evaluate visual language models (VLMs) on physics-based reasoning. Experiments show that current VLMs fall short of human-level understanding in physics-based anomaly scenarios. This work marks a milestone for industrial anomaly detection, promoting physics-grounded reasoning in complex industrial settings.", + "bbox": [ + 511, + 422, + 908, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation and future work. Although our Phys-AD dataset provides a large variety of objects with diverse physical properties and various types of interaction methods, we plan to add even more diverse interaction methods and objects in the future to better meet the demands of complex real-world industrial scenarios. Due to the significant differences between our Phys-AD dataset and current industrial anomaly detection and video anomaly detection datasets, most existing anomaly detection algorithms cannot be directly applied to our dataset. In the future, we will test more algorithms on our Phys-AD dataset and provide experimental results across more settings like zero-shot, few-shot, semi-supervised settings, etc.", + "bbox": [ + 511, + 684, + 908, + 866 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "30416", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In ICCV, 2015. 3", + "[2] Fabien Baradel, Natalia Neverova, Julien Mille, Greg Mori, and Christian Wolf. Cophy: Counterfactual learning of physical dynamics. arXiv preprint arXiv:1909.12000, 2019.3", + "[3] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad — a comprehensive real-world dataset for unsupervised anomaly detection. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9584-9592, 2019. 3", + "[4] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. MVtec AD – A comprehensive real-world dataset for unsupervised anomaly detection. In CVPR, 2019. 2", + "[5] Paul Bergmann, Kilian Batzner, Michael Fauser, David Sattlegger, and Carsten Steger. Beyond dents and scratches: Logical constraints in unsupervised anomaly detection and localization. International Journal of Computer Vision, 130 (4):947-969, 2022. 3", + "[6] Paul Bergmann, Xin Jin, David Sattlegger, and Carsten Steger. The mvtec 3d-ad dataset for unsupervised 3d anomaly detection and localization. In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications. SCITEPRESS - Science and Technology Publications, 2022. 3", + "[7] Paul Bergmann, Xin Jin, David Sattlegger, and Carsten Steger. The MVTec 3d-AD dataset for unsupervised 3d anomaly detection and localization. In VISAPP, 2022. 2", + "[8] Luca Bonfiglioli, Marco Toschi, Davide Silvestri, Nicola Fioraio, and Daniele De Gregorio. The eyecandies dataset for unsupervised multimodal anomaly detection and localization. In Proceedings of the Asian Conference on Computer Vision, pages 3586-3602, 2022. 3", + "[9] Yunkang Cao, Xiaohao Xu, Jiangning Zhang, Yuqi Cheng, Xiaonan Huang, Guansong Pang, and Weiming Shen. A survey on visual anomaly detection: Challenge, approach, and prospect. arXiv preprint arXiv:2401.16402, 2024. 2", + "[10] Yingxian Chen, Zhengzhe Liu, Baoheng Zhang, Wilton Fok, Xiaojuan Qi, and Yik-Chung Wu. Mgfn: Magnitude-contrastive glance-and-focus network for weakly-supervised video anomaly detection. In Proceedings of the AAAI conference on artificial intelligence, pages 387-395, 2023. 7, 8", + "[11] Hang Du, Sicheng Zhang, Binzhu Xie, Guoshun Nan, Jiayang Zhang, Junrui Xu, Hangyu Liu, Sicong Leng, Jiangming Liu, Hehe Fan, et al. Uncovering what why and how: A comprehensive benchmark for causation understanding of video anomaly. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18793-18803, 2024. 7", + "[12] Jiafei Duan, Samson Yu, and Cheston Tan. Space: A simulator for physical interactions and causal learning in 3d environments. In ICCV, pages 2058-2063, 2021. 3", + "[13] Jiafei Duan, Samson Yu, Soujanya Poria, Bihan Wen, and Cheston Tan. Pip: Physical interaction prediction via mental" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "simulation with span selection. In ECCV, pages 405-421. Springer, 2022. 3", + "[14] Jia-Chang Feng, Fa-Ting Hong, and Wei-Shi Zheng. Mist: Multiple instance self-training framework for video anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14009-14018, 2021. 3", + "[15] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 7", + "[16] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1705-1714, 2019. 3, 7, 8", + "[17] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 3", + "[18] Mahmudul Hasan, Jonghyun Choi, Jan Neumann, Amit K Roy-Chowdhury, and Larry S Davis. Learning temporal regularity in video sequences. In CVPR, 2016. 3", + "[19] Drew A. Hudson and Christopher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, 2019. 3", + "[20] Steeven Janny, Fabien Baradel, Natalia Neverova, Madiha Nadri, Greg Mori, and Christian Wolf. Filtered-cophy: Unsupervised learning of counterfactual physics in pixel space. In ICLR, 2022. 3", + "[21] Stepan Jezek, Martin Jonak, Radim Burget, Pavel Dvorak, and Milos Skotak. Deep learning-based defect detection of metal parts: evaluating current methods in complex conditions. In 2021 13th International Congress on Ultra Modern Telecommunications and Control Systems and Workshops (ICUMT), pages 66-71, 2021. 2, 3", + "[22] Huaizu Jiang, Xiaojian Ma, Weili Nie, Zhiding Yu, Yuke Zhu, and Anima Anandkumar. Bongard-hoi: Benchmarking few-shot visual reasoning for human-object interactions. In CVPR, pages 19056-19065, 2022. 3", + "[23] Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C. Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In CVPR, 2017. 3", + "[24] Hyekang Kevin Joo, Khoa Vo, Kashu Yamazaki, and Ngan Le. Clip-tsa: Clip-assisted temporal self-attention for weakly-supervised video anomaly detection. In 2023 IEEE International Conference on Image Processing (ICIP), pages 3230-3234. IEEE, 2023. 3", + "[25] Federico Landi, Cees GM Snok, and Rita Cucchiara. Anomaly locality in video surveillance. arXiv preprint arXiv:1901.10364, 2019. 3", + "[26] Ting Li, Charles Dubout, Emma K Wampler, Steven Yantis, Donald Geman, et al. Comparing machines and humans on a visual categorization test. 2011. 3" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "30417", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Wenqiao Li, Xiaohao Xu, Yao Gu, Bozhong Zheng, Shenghua Gao, and Yingna Wu. Towards scalable 3d anomaly detection and localization: A benchmark via 3d anomaly synthesis and a self-supervised learning network. arXiv preprint, 2023. 2, 3", + "[28] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 7, 8", + "[29] Jiaqi Liu, Guoyang Xie, Rui Chen, et al. Real3d-ad: A dataset of point cloud anomaly detection. NeurIPS, 2023. 2", + "[30] Jiaqi Liu, Guoyang Xie, Ruitao Chen, Xinpeng Li, Jinbao Wang, Yong Liu, Chengjie Wang, and Feng Zheng. Real3dad: A dataset of point cloud anomaly detection. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[31] Kun Liu and Huadong Ma. Exploring background-bias for anomaly detection in surveillance videos. In Proceedings of the 27th ACM International Conference on Multimedia, pages 1490-1499, 2019. 3", + "[32] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection-a new baseline. In CVPR, 2018. 3", + "[33] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13588-13597, 2021. 3", + "[34] Hui Lv, Chen Chen, Zhen Cui, Chunyan Xu, Yong Li, and Jian Yang. Learning normal dynamics in videos with meta prototype network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15425-15434, 2021. 7, 8", + "[35] Hui Lv, Zhongqi Yue, Qianru Sun, Bin Luo, Zhen Cui, and Hanwang Zhang. Unbiased multiple instance learning for weakly supervised video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8022-8031, 2023. 3", + "[36] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 7, 8", + "[37] Domingo Mery, Vladimir Riffo, Uwe Zscherpel, German Mondragon, Ivan Lillo, Irene Zuccar, Hans Lobel, and Miguel Carrasco. Gdxray: The database of x-ray images for nondestructive testing. Journal of Nondestructive Evaluation, 34 (4):42, 2015. 3", + "[38] Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. Vt-adj: A vision transformer network for image anomaly detection and localization. In 2021 IEEE 30th International Symposium on Industrial Electronics (ISIE). IEEE, 2021. 3", + "[39] Weili Nie, Zhiding Yu, Lei Mao, Ankit B Patel, Yuke Zhu, and Animashree Anandkumar. Bongard-logo: A new benchmark for human-level concept learning and reasoning. In NeurIPS, 2020. 3", + "[40] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In Proceed-" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14372-14381, 2020. 7, 8", + "[41] Yujiang Pu, Xiaoyu Wu, and Shengjin Wang. Learning prompt-enhanced context features for weakly-supervised video anomaly detection. arXiv preprint arXiv:2306.14451, 2023. 3", + "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 7", + "[43] Binyi Su, Zhong Zhou, and Haiyong Chen. Pvel-ad: A large-scale open-world dataset for photovoltaic cell anomaly detection. IEEE Transactions on Industrial Informatics, 19 (1):404-413, 2023. 3", + "[44] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In CVPR, 2018. 7", + "[45] Kamalakar Vijay Thakare, Yash Raghuwanshi, Debi Prosad Dogra, Heeseung Choi, and Ig-Jae Kim. Dyannet: A scene dynamicity guided self-trained video anomaly detection network. In Proceedings of the IEEE/CVF Winter conference on applications of computer vision, pages 5541-5550, 2023. 3", + "[46] Chengjie Wang, Wenbing Zhu, Bin-Bin Gao, Zhenye Gan, Jianning Zhang, Zhihao Gu, Shuguang Qian, Mingang Chen, and Lizhuang Ma. Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detection, 2024. 3", + "[47] Yue Wang, Jinlong Peng, Jiangning Zhang, Ran Yi, Yabiao Wang, and Chengjie Wang. Multimodal industrial anomaly detection via hybrid fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8032-8041, 2023. 7", + "[48] Jhih-Ciang Wu, He-Yen Hsieh, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng-Luh Liu. Self-supervised sparse representation for video anomaly detection. In European Conference on Computer Vision, pages 729-745. Springer, 2022. 7, 8", + "[49] Peng Wu, Jing Liu, Yujia Shi, Yujia Sun, Fangtao Shao, Zhaoyang Wu, and Zhiwei Yang. Not only look, but also listen: Learning multimodal violence detection under weak supervision. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 322-339. Springer, 2020. 3, 7", + "[50] Peng Wu, Xuerong Zhou, Guansong Pang, Yujia Sun, Jing Liu, Peng Wang, and Yanning Zhang. Open-vocabulary video anomaly detection. arXiv preprint arXiv:2311.07042, 2023. 3", + "[51] Peng Wu, Xuerong Zhou, Guansong Pang, Lingru Zhou, Qingsen Yan, Peng Wang, and Yanning Zhang. Vadclip: Adapting vision-language models for weakly supervised video anomaly detection. arXiv preprint arXiv:2308.11681, 2023.3,7,8", + "[52] Dan Xu, Yan Yan, Elisa Ricci, and Nicu Sebe. Detecting anomalous events in videos by learning deep representations of appearance and motion. Computer Vision and Image Understanding, 156:117-127, 2017. 3", + "[53] Xiaohao Xu, Yunkang Cao, Yongqi Chen, Weiming Shen, and Xiaonan Huang. Customizing visual-language foundation models for multi-modal anomaly detection and reasoning. arXiv preprint arXiv:2403.11083, 2024. 3" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "30418", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Zhiwei Yang, Jing Liu, Zhaoyang Wu, Peng Wu, and Xiaotao Liu. Video event restoration based on keyframes for video anomaly detection. In CVPR, 2023. 3", + "[55] M Zaigham Zaheer, Arif Mahmood, M Haris Khan, Mattia Segu, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14744-14754, 2022. 3", + "[56] Luca Zanella, Willi Menapace, Massimiliano Mancini, Yiming Wang, and Elisa Ricci. Harnessing large language models for training-free video anomaly detection. arXiv preprint arXiv:2404.01014, 2024. 3, 7, 8", + "[57] Aimen Zerroug, Mohit Vaishnav, Julien Colin, Sebastian Musslick, and Thomas Serre. A benchmark for compositional visual reasoning. arXiv preprint arXiv:2206.05379, 2022. 3", + "[58] Chi Zhang, Feng Gao, Baoxiong Jia, Yixin Zhu, and Song-Chun Zhu. Raven: A dataset for relational and analogical visual reasoning. In CVPR, 2019. 3", + "[59] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 7, 8", + "[60] Huaxin Zhang, Xiang Wang, Xiaohao Xu, Xiaonan Huang, Chuchu Han, Yuehuan Wang, Changxin Gao, Shanjun Zhang, and Nong Sang. Glancevad: Exploring glance supervision for label-efficient video anomaly detection. arXiv preprint arXiv:2403.06154, 2024. 3", + "[61] Huaxin Zhang, Xiaohao Xu, Xiang Wang, Jialong Zuo, Chuchu Han, Xiaonan Huang, Changxin Gao, Yuehuan Wang, and Nong Sang. Holmes-vad: Towards unbiased and explainable video anomaly detection via multi-modal llm. arXiv preprint arXiv:2406.12235, 2024. 3, 7", + "[62] Yiheng Zhang, Yunkang Cao, Xiaohao Xu, and Weiming Shen. Logic: An llm-driven framework for logical anomaly detection. IEEE Transactions on Automation Science and Engineering, pages 1-0, 2024. 3", + "[63] Jia-Xing Zhong, Nannan Li, Weijie Kong, Shan Liu, Thomas H Li, and Ge Li. Graph convolutional label noise cleaner: Train a plug-and-play action classifier for anomaly detection. In CVPR, 2019. 3", + "[64] Qiang Zhou, Weize Li, Lihan Jiang, Guoliang Wang, Guyue Zhou, Shanghang Zhang, and Hao Zhao. Pad: A dataset and benchmark for pose-agnostic anomaly detection, 2023. 3", + "[65] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In ECCV, 2022. 2", + "[66] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation, 2022. 3" + ], + "bbox": [ + 91, + 92, + 482, + 785 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "30419", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Towards Visual Discrimination and Reasoning of Real-World Physical Dynamics_ Physics-Grounded Anomaly Detection/e496d86c-0d1a-42d1-b098-7f552a70c4ff_model.json b/2025/Towards Visual Discrimination and Reasoning of Real-World Physical Dynamics_ Physics-Grounded Anomaly Detection/e496d86c-0d1a-42d1-b098-7f552a70c4ff_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3e96d54784bdc29e59426b1336e5c7c82349b154 --- /dev/null +++ b/2025/Towards Visual Discrimination and Reasoning of Real-World Physical Dynamics_ Physics-Grounded Anomaly Detection/e496d86c-0d1a-42d1-b098-7f552a70c4ff_model.json @@ -0,0 +1,2697 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.13, + 0.91, + 0.177 + ], + "angle": 0, + "content": "Towards Visual Discrimination and Reasoning of Real-World Physical Dynamics: Physics-Grounded Anomaly Detection" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.202, + 0.946, + 0.242 + ], + "angle": 0, + "content": "Wenqiao Li\\(^{1*}\\) Yao Gu\\(^{1*}\\) Xintao Chen\\(^{1*}\\) Xiaohao Xu\\(^{2\\dagger}\\) Ming Hu\\(^{3}\\) Xiaonan Huang\\(^{2}\\) Yingna Wu\\(^{1\\dagger}\\) \n\\(^{1}\\)ShanghaiTech University \\(^{2}\\)University of Michigan, Ann Arbor \\(^{3}\\)Monash University" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.254, + 0.536, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.25, + 0.907, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.547, + 0.91, + 0.618 + ], + "angle": 0, + "content": "Figure 1. Towards visual discrimination of physical dynamics in real-world industrial object anomaly detection. We illustrate objects, interactions, and time-sequenced videos from the Physics-Grounded Anomaly Detection dataset: (a) Object; (b) Interaction: Applied actions shown with directional arrows; (c) Video with Physical Dynamics: Temporal sequences showing normal and abnormal states, highlighting anomalies like leaks, misalignments, and cracks. By focusing on the dynamic behaviors of complex objects, we enhance understanding of interactions and failure modes in real-world settings, where both structure and motion contribute to anomaly detection." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.641, + 0.327, + 0.656 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.675, + 0.486, + 0.857 + ], + "angle": 0, + "content": "Humans detect real-world object anomalies by perceiving, interacting, and reasoning based on object-conditioned physical knowledge. The long-term goal of Industrial Anomaly Detection (IAD) is to enable machines to autonomously replicate this skill. However, current IAD algorithms are largely developed and tested on static, semantically simple datasets, which diverge from real-world scenarios where physical understanding and reasoning are essential. To bridge this gap, we introduce the Physics Anomaly Detection (Phys-AD) dataset, the first large-scale, real-world, physics-grounded video dataset for industrial anomaly detection. Collected using a real robot arm and motor, Phys-AD provides a diverse" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.911, + 0.87 + ], + "angle": 0, + "content": "set of dynamic, semantically rich scenarios. The dataset includes more than 6400 videos across 22 real-world object categories, interacting with robot arms and motors, and exhibits 47 types of anomalies. Anomaly detection in Phys-AD requires visual reasoning, combining both physical knowledge and video content to determine object abnormality. We benchmark state-of-the-art anomaly detection methods under three settings: unsupervised AD, weakly-supervised AD, and video-understanding AD, highlighting their limitations in handling physics-grounded anomalies. Additionally, we introduce the Physics Anomaly Explanation (PAEval) metric, designed to assess the ability of visual-language foundation models to not only detect anomalies but also provide accurate explanations for their underlying physical causes. Our project is available at https://guyao2023.github.io/Phys-AD/." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.285, + 0.888 + ], + "angle": 0, + "content": "*Equally contribute to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.24, + 0.9 + ], + "angle": 0, + "content": "† Corresponding authors" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.285, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30409" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.092, + 0.48, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.318, + 0.488, + 0.442 + ], + "angle": 0, + "content": "Figure 2. Human-like decision-making process for physics-grounded object anomaly detection. We illustrate the sequential approach of a human-like agent for evaluating an object's normality. First, the agent perceives relevant physical attributes (e.g., plastic and elastic), then interacts by performing a physical action (e.g., squeezing), and finally reasons based on the vision feedback and attributes changes (e.g., surface shape change) to determine whether the object is normal or anomalous. This mirrors a human's natural process of reasoning over physics in objects." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.468, + 0.223, + 0.484 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.484, + 0.688 + ], + "angle": 0, + "content": "Industrial anomaly detection (IAD) is a critical subfield in computer vision and industrial automation, aiming to identify defects or irregularities in products during manufacturing. As shown in Fig. 2, the ultimate vision is to create autonomous systems that not only perceive but also interact with and reason about objects to discriminate between anomalies and normal states, integrating complex physical principles to detect anomalies in dynamic, real-world scenarios. For example, a human inspecting a water bottle for anomalies wouldn't rely solely on visual observation; they might rotate or invert the bottle, using physical interactions and cues, such as noticing a loose cap or an irregular internal flow, to detect issues." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.856 + ], + "angle": 0, + "content": "Central to advancing IAD [9] is the availability of high-quality datasets that bridge the gap between academic research and industrial needs. Datasets like MVTec-AD [4], MPDD [21], and VisA [65] have played foundational roles, enabling algorithm development for image-based anomaly detection and bringing IAD to the forefront of computer vision research. While these datasets have significantly advanced single-image anomaly detection, recent datasets, e.g., MVTec-3D [7], Real3D [29], and Anomaly-ShapeNet [27], have extended IAD to 3D, aligning research more closely with the needs of complex real-world industrial settings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Yet, as factories increasingly rely on robotic arms and automated systems to perform sophisticated inspections, the limitations of current IAD datasets become apparent." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "Existing benchmarks focus on static, semantically simple environments, overlooking the physical priors and interactive reasoning required in real-world industrial contexts. This gap highlights a growing need for datasets that not only reflect real-world physical constraints but also challenge models to reason dynamically about anomalies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.185, + 0.907, + 0.517 + ], + "angle": 0, + "content": "To bridge this gap, as shown in Fig. 1, we introduce the Physics Anomaly Detection (Phys-AD) dataset, the first large-scale, physics-grounded video dataset for industrial anomaly detection. Phys-AD features over 6,400 videos of 22 categories and 49 objects interacting with robotic arms and motors, capturing 47 types of anomalies that require visual reasoning informed by physical knowledge. The short video clips in the dataset range from 60 to 240 frames in length and are filmed in real-time industrial environment, fully capturing the interaction process between robotic arm or motor and industrial objects. Additionally, to ensure our dataset meets industrial demands and matches the complexity and diversity of the real physical world, we selected industrial objects of different physical qualities, various interaction methods, and anomalies that reflect diverse physical principles and require different reasoning process. Specifically, we selected 22 object categories spanning across metals, plastics, fluid, amorphous substances and articulated objects with diverse appearances. For interaction, we use mechanical grippers, robotic arms, and motors, incorporating various interaction modes such as pressing, rotating, squeezing, and driving to handle different types of objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.909, + 0.746 + ], + "angle": 0, + "content": "We benchmark state-of-the-art anomaly detection methods in three key configurations: unsupervised anomaly detection, weakly-supervised anomaly detection, and video-understanding based anomaly detection. Our findings reveal critical gaps in their ability to handle the complexities of physics-grounded scenarios, where anomalies often arise from dynamic, interdependent interactions. To advance the field, we also introduce the Physical Anomaly Explanation (PAEval) metric, designed to assess both detection performance and a model's ability to explain anomalies by identifying underlying physical causes. Furthermore, our benchmark reveals the fragility of existing methods in tackling these challenging conditions, underscoring the need for approaches that better understand object dynamics and temporal coherence in anomaly detection." + }, + { + "type": "title", + "bbox": [ + 0.532, + 0.748, + 0.902, + 0.763 + ], + "angle": 0, + "content": "Our contributions can be summarized as followings:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.765, + 0.907, + 0.81 + ], + "angle": 0, + "content": "- We introduce a novel task of detecting physical-based industrial anomalies in real-world that involves perception, physical and visual reasoning." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.811, + 0.908, + 0.886 + ], + "angle": 0, + "content": "- We present Phys-AD, the first large-scale, physics-grounded video dataset specifically designed for industrial anomaly detection in real world, containing objects of different physical qualities, multiple interaction methods and various physical reasoning process." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "- We benchmarking the anomaly detection and reasoning" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.765, + 0.908, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30410" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.16 + ], + "angle": 0, + "content": "Table 1. Comparison of Phys-AD with existing industrial anomaly detection datasets. Our Phys-AD dataset is the first to consider complex objects with physical dynamics. 'Syn', 'IR', 'D', and 'PC' denote Synthetic, Infrared, Depth, and Point Cloud, respectively. #Anomaly indicates the number of anomaly types." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.162, + 0.485, + 0.393 + ], + "angle": 0, + "content": "
DatasetYearTypeModalitySample Statistics
Class #AnomalyPhysics
MVTec-AD [3]2019RealRGB15-X
BTAD [38]2021RealRGB33X
MPDD [21]2021RealRGB68X
VisA [66]2021RealRGB12-X
MVTec LOCO-AD [5]2022RealRGB5-X
MAD [64]2023Syn+RealRGB203X
LOCO-Annotations [62]2024RealRGB55X
Real-IAD [46]2024RealRGB308X
GDXray [37]2015RealX-ray515X
PVEL-AD [43]2023RealIR110X
MVTec3D-AD [6]2021RealRGB-D103-5X
Eyecandies [8]2022SynRGB-D103X
Real3D-AD [30]2023RealPC122X
Anomaly-ShapeNet [27]2024SynPC406X
Phys-AD (Ours)2024RealRGB4947✓
" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.42, + 0.484, + 0.496 + ], + "angle": 0, + "content": "performance of popular video AD methods and Visual Language Foundation Models on the Phys-AD dataset in several settings, establishing a practical and challenging benchmark to promote the development of the physics-related anomaly detection field." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.512, + 0.228, + 0.528 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Industrial anomaly detection datasets. Existing datasets primarily focus on static, semantically simple scenarios, deviate significantly from real world where physical understanding and reasoning are essential. Datasets like MVTec-AD [3], BTAD [38], MPDD [21], and VisA [66] focus on surface level image anomaly detection with one single-view RGB image, limiting their effectiveness in capturing holistic object structures. MVTec-LOCO-AD [5], focusing on structure and local information in industries, is limited by its relatively simple and constrained data content. While MVTec3D-AD [6] and Eyecandies [8] incorporate depth data, they remain static and single-view image anomaly detection, neglecting object level information. To explore object level anomaly detection, multi-view IAD datasets like MAD [64] and Real-IAD [46], point cloud IAD datasets like Real3D-AD [30] and Anomaly-ShapeNet [27], offering richer visual and geometric cues, but are still limited to static objects without dynamic interaction or reasoning. In summary, current industrial anomaly detection datasets focus on relatively simple and static anomaly detection scenarios, lacking complex physical rules, dynamic interactions, and visual reasoning requirements. Therefore, existing IAD datasets generally applied to limited industrial scenarios and there is no IAD dataset could meet the demands of detecting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "complex anomalies in real world which need physical priors and reasoning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.126, + 0.909, + 0.459 + ], + "angle": 0, + "content": "Video anomaly detection. Deep learning methods [18, 35, 54, 61] now dominate Video Anomaly Detection (VAD), categorized into unsupervised, weakly-supervised, and fully-supervised approaches. Unsupervised methods learn normal patterns via reconstruction [16, 18, 52], prediction [32], or hybrids [33], while some methods [45, 55] train with both unlabeled normal and abnormal data. Weakly-supervised methods[14, 49, 60, 63] use video-level or glance-based annotations, and fully-supervised methods[25, 31] remain rare due to costly frame-level labeling. Visual-language models like CLIP [42] have recently been applied to enhance anomaly detection [24, 41, 51], focusing on semantic anomalies. Open-vocabulary VAD [50] and prompt-based anomaly scoring [56] leverage LLMs [53, 61], but performance relies heavily on the base models, often lacking domain-specific tuning. However, existing video anomaly detection algorithms lack the capability to handle complex industrial anomaly detection scenarios and understand physical rules. This gap highlights the need for models that can capture dynamic behaviors and physical laws in industrial environments, as addressed by Phys-AD, which targets industrial anomalies in objects of various physical properties." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.911, + 0.691 + ], + "angle": 0, + "content": "Visual reasoning. Visual reasoning is a critical task in computer vision, aiming to enable machines to interpret perceptual information like humans. Several visual reasoning tasks have been proposed to evaluate reasoning capabilities, including Visual Question Answering (VQA), 2D puzzles, and physical dynamics prediction. In VQA, agents are required to combine natural language and visual cues to answer questions [1, 17, 19, 23]. For 2D puzzles, tasks involve discovering relationships among visual elements and making inferences [22, 26, 39, 57, 58]. Physical dynamics prediction tasks require machines to perceive and reason about physical interactions [2, 12, 13, 20]. In contrast to these work, we introduce the first benchmark featuring real-world industrial objects with dynamic physical properties, focusing on distinguishing diverse dynamics through vision." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.703, + 0.69, + 0.719 + ], + "angle": 0, + "content": "3. Dataset: Phys-AD" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.728, + 0.898, + 0.744 + ], + "angle": 0, + "content": "3.1. Object Preparation and Interaction Selection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We selected 22 categories spanning across various materials, including metals, plastics, amorphous substances and articulated objects, with diverse shapes, sizes, and physical properties. For the physical properties of different objects, we correspondingly select different interaction methods like push, rotate, pull with robotics arms and motors. For instance, we use robotics arms to grab and extruding the deformable objects like basketball to determine whether there are any elastic anomalies or surface defects by the morphological change of the objects. To make our dataset more practical" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30411" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.226, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.148, + 0.18, + 0.159 + ], + "angle": 0, + "content": "U Disk" + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.089, + 0.361, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.276, + 0.147, + 0.317, + 0.16 + ], + "angle": 0, + "content": "Zipper" + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.09, + 0.498, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.148, + 0.483, + 0.16 + ], + "angle": 0, + "content": "Rolling Bearing" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.091, + 0.633, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.149, + 0.606, + 0.16 + ], + "angle": 0, + "content": "Sticky Roller" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.091, + 0.769, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.148, + 0.719, + 0.159 + ], + "angle": 0, + "content": "Slider" + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.091, + 0.904, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.808, + 0.147, + 0.871, + 0.159 + ], + "angle": 0, + "content": "Toothpaste" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.165, + 0.226, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.225, + 0.192, + 0.235 + ], + "angle": 0, + "content": "Door Lock" + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.165, + 0.361, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.272, + 0.224, + 0.319, + 0.237 + ], + "angle": 0, + "content": "Toy Car" + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.165, + 0.498, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.417, + 0.224, + 0.444, + 0.237 + ], + "angle": 0, + "content": "Clip" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.166, + 0.633, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.225, + 0.582, + 0.236 + ], + "angle": 0, + "content": "Fan" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.166, + 0.769, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.225, + 0.741, + 0.237 + ], + "angle": 0, + "content": "Rubber Band" + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.166, + 0.905, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.819, + 0.226, + 0.865, + 0.238 + ], + "angle": 0, + "content": "Magnet" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.245, + 0.225, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.306, + 0.169, + 0.316 + ], + "angle": 0, + "content": "Servo" + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.245, + 0.361, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.305, + 0.34, + 0.318 + ], + "angle": 0, + "content": "Gear Assembly" + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.245, + 0.498, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.305, + 0.469, + 0.316 + ], + "angle": 0, + "content": "Bottled Water" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.245, + 0.633, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.306, + 0.582, + 0.316 + ], + "angle": 0, + "content": "Screw" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.245, + 0.769, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.305, + 0.758, + 0.318 + ], + "angle": 0, + "content": "Spherical Bearing" + }, + { + "type": "image", + "bbox": [ + 0.793, + 0.254, + 0.889, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.814, + 0.29, + 0.861, + 0.301 + ], + "angle": 0, + "content": "Motion" + }, + { + "type": "image_caption", + "bbox": [ + 0.796, + 0.304, + 0.882, + 0.327 + ], + "angle": 0, + "content": "Direction For Interaction" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.326, + 0.225, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.392, + 0.171, + 0.402 + ], + "angle": 0, + "content": "Button" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.327, + 0.361, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.275, + 0.392, + 0.31, + 0.402 + ], + "angle": 0, + "content": "Clock" + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.326, + 0.495, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.392, + 0.467, + 0.403 + ], + "angle": 0, + "content": "Caster Wheel" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.327, + 0.633, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.392, + 0.585, + 0.404 + ], + "angle": 0, + "content": "Hinge" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.327, + 0.77, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.693, + 0.392, + 0.718, + 0.404 + ], + "angle": 0, + "content": "Ball" + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.327, + 0.904, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.408, + 0.907, + 0.465 + ], + "angle": 0, + "content": "Figure 3. Interactions for understanding implicit physical laws in the Phys-AD dataset. We showcase various object interactions from the Phys-AD dataset, where different actions (indicated by motion directions) are used to explore and reason about the underlying physical properties and behaviors of each object. The colored arrows indicate the interaction directions and axes, highlighting how physical interactions reveal the implicit physics governing each object." + }, + { + "type": "image_caption", + "bbox": [ + 0.169, + 0.477, + 0.411, + 0.489 + ], + "angle": 0, + "content": "Normal: The U Disk rotates freely over time." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.492, + 0.495, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.178, + 0.582, + 0.404, + 0.594 + ], + "angle": 0, + "content": "Abnormal: The U Disk is stuck over time." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.598, + 0.495, + 0.664 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.669, + 0.358, + 0.68 + ], + "angle": 0, + "content": "(a) Persistent Anomalies" + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.477, + 0.856, + 0.489 + ], + "angle": 0, + "content": "Normal: The sticky roller rotates smoothly on the surface." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.492, + 0.902, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.573, + 0.72, + 0.598 + ], + "angle": 0, + "content": "Normal: The sticky roller functions properly at the beginning." + }, + { + "type": "image_caption", + "bbox": [ + 0.749, + 0.573, + 0.902, + 0.597 + ], + "angle": 0, + "content": "Abnormal: The sticky roller breaks at a certain point." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.601, + 0.902, + 0.665 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.67, + 0.786, + 0.681 + ], + "angle": 0, + "content": "(b) Intermittent Anomalies" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.687, + 0.907, + 0.757 + ], + "angle": 0, + "content": "Figure 4. Categorization of anomalies based on persistence in the Phys-AD dataset. We show examples of normal and abnormal functioning in common objects, divided into two anomaly types: persistent and intermittent. (a) Persistent anomalies, such as continuous obstruction in the U Disk or permanent malfunction of the Sticky Roller, are visible throughout the operation. (b) In contrast, intermittent anomalies, like occasional jamming of the U Disk or breakage in the Sticky Roller after initial operation, only appear at specific points in time. This classification provides insight into both constant and sporadic failures in object interactions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.784, + 0.486, + 0.889 + ], + "angle": 0, + "content": "and challenging, we introduce 47 distinct defect types, some of which just rely on single frame content and physical rules for anomaly reasoning, while the other part needs to combine the content of the whole video and physical priors together to judge whether the object is abnormal or not. Different objects information with their corresponding interaction methods and anomalies are listed in Table 2. Fig 3 shows the interac" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.784, + 0.907, + 0.874 + ], + "angle": 0, + "content": "tions for understanding implicit physical laws in the Phys-AD dataset. Fig 4 provides two representative examples: one where a U Disk requires analyzing the entire video content to determine if there is an anomaly, and another where a sticky roller only requires a few frames from the video to judge whether the object is abnormal." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30412" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.288, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.091, + 0.482, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.213, + 0.483, + 0.255 + ], + "angle": 0, + "content": "Figure 5. Data collection pipeline for the Phys-AD dataset. (a) Manipulation of a toothpaste tube using a UR5 robotic arm. (b) Manipulation of a U Disk and fan via servo and motor." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.275, + 0.483, + 0.317 + ], + "angle": 0, + "content": "Table 2. Interaction methods and anomalies of the object categories in our Phys-AD dataset. Motor. means one object driven by motor. Multi. means one object with multiple anomalies." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.32, + 0.481, + 0.671 + ], + "angle": 0, + "content": "
CategoryInteractionAnomaly Types
CarDrag, SlideDifferent wheels stuck
FanMotor., RotateStuck, Uneven Rotation, Vibration
Rolling BearingMotor., RotateLack of friction
Spherical BearingGrab, RotateInternal block
ServoGrab, Motor., RotateAngle restricted, Vibration, No calibration
ClipPressUnable to press down, Unable to rebound
U DiskGrab, Motor., RotateCover jam
HingeGrab, RotateAngle restricted, Shaft off
Sticky RollerGrab, PullDetach, Unable to rotate
Caster WheelSlideAxle axis stuck, Swivel axis stuck
ScrewPress, RotateLoosen, Unable to insert
LockMotor., RotateLatch jam, Loose
GearMotor., RotateStuck, Not meshed, Multi.
ClockMotor., RotatePointer stops, Vibration
SlideGrab, SlideDetach, Shedding, Jam
ZipperGrab, CloseStuck, Unable to close
ButtonPressNo light, Unable to press down, Unable to rebound, Multi.
LiquidGrab, ShakeWater out, Foreign body
Rubber BandStretchCrack
BallPinch, PressInsufficient gas, Leakage
MagnetGrab, Press, MoveDegaussing, Shell detached
ToothpastePinch, PressLeakage
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.695, + 0.367, + 0.712 + ], + "angle": 0, + "content": "3.2. Data Collection and Processing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Data collection pipeline. Most of the data collection for the Phys-AD dataset is driven by manipulation-guided video sequences, captured using a UR5 robotic arm equipped with an RGB camera (see Fig. 5a). In order to reproduce real industrial scenes, we also include adjustable light sources in our data capture process. The RGB Camera, with a 1080p resolution and a frame rate of 60 FPS, providing high-quality video sequences. Some kinds of the objects like U Disk that are not suitable for the manipulation of robotics arms, are driven by the motor or servo (Fig. 5b). After the data collection, we used video editing software to crop out irrelevant frames and retain the complete interaction process." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.132 + ], + "angle": 0, + "content": "Table 3. Phys-AD dataset statistics. We denote the total number of frames and videos for each category as #Images and #Videos. Note that the Train split does not contain anomaly data." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.135, + 0.905, + 0.358 + ], + "angle": 0, + "content": "
CategoryTrain #FramesTrain #VideosTest #FramesTest #VideosTest #Videos Nor.Test #Videos Ab.Obj. TypesDef. Types
Car18,00030036,00060015045053
Fan32,40018064,8003609027033
Rolling Bearing3,600603,60060303011
Spherical Bearing3,600603,60060303011
Servo14,40012028,8002406018013
Clip28,80024043,20036012024042
U Disk14,40024014,40024012012041
Hinge3,600307,20060154512
Sticky Roller5,400308,10045153012
Caster Wheel5,4003010,80060154512
Screw5,400308,10045153012
Lock7,20012010,8001806012012
Gear21,60018054,0004509036034
Clock27,00015040,3202247315152
Slide7,2006018,0001503012013
Zipper14,40012021,6001806012022
Button21,60012054,0003006024044
Liquid5,400308,10045153012
Rubber Band10,8006010,80060303011
Ball21,6009032,400135453032
Magnet10,8006016,20090306022
Toothpaste16,2009016,20090454531
Total298,8002400511,0204,0341,1982,8364947
" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.377, + 0.662, + 0.391 + ], + "angle": 0, + "content": "3.3. Data Statistics" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.399, + 0.908, + 0.642 + ], + "angle": 0, + "content": "Dataset sample distribution. Table 3 provides a detailed breakdown of the Phys-AD dataset, which includes information on category, the number of training videos and frames, the number of testing videos and frames, the distribution of normal and abnormal samples in the test set, and the number of object and defect types. The length of the video clips in the dataset ranges from 60 to 240 frames, ensuring to fully capture the interaction process in a short time. The dataset contains 2400 training videos and 4034 test videos spreading across 22 categories, 49 object types, and 47 defect types. The frames are extracted from the videos at a rate of 60 FPS. The test set includes 1,198 normal samples and 2,836 abnormal samples. Overall, the Phys-AD dataset covers a wide range of categories and contains a large scale of data, which helps to train more robust anomaly detection models and provides a reasonable evaluation setting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Interaction and defect types across categories. Table 2 details the defect types and corresponding interaction methods for each category in the Phys-AD dataset. Covering 22 categories and including 47 defect types, the dataset features various interactions (e.g., Rotate, Grab). Importantly, interaction methods or defect types with identical names can differ across categories—for example, \"rotation\" for a fan describes the motor-driven movement of its blades, while for a hinge it refers to the turning of its pages around the shaft under robotic control. Furthermore, to detect anomalies with complex physical properties in a human-like manner, we sometimes combine multiple interactions. For instance, when inspecting a screw, we first press it into the hole and then use a robotic arm to rotate it further. Overall, Phys-AD provides a diverse, physics-grounded anomaly detection framework that fosters the development of advanced methods for tackling real-world challenges." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30413" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.091, + 0.482, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.358, + 0.485, + 0.484 + ], + "angle": 0, + "content": "Figure 6. PhysAD-Agent: A Large Language Model (LLM)-powered system for physics anomaly detection label augmentation and automatic evaluation. This agent framework consists of two main components: (a) Rich Annotation Generation, where an LLM-based data agent generates detailed anomaly annotations based on video, anomaly state, prompt, and human-annotated physical priors, and (b) Automatic Evaluation, where an LLM-based evaluation agent assesses model predictions to calculate the Physics Anomaly Explanation (PAEval) metric." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.484, + 0.676 + ], + "angle": 0, + "content": "Labels. For the test set videos, we provide two types of labels. First, there are the common video-level labels: all anomalous videos in the test set are labeled as 1, while normal videos are labeled as 0. For evaluating visual-language models, we also provide text labels. Specifically, for each type of anomaly in the test set, we manually design a textual description label. This label includes both a description of the video content and a physical explanation of the reason for the anomaly. To ensure diversity in our text labels, we also use ChatGPT-4o for text augmentation. For further details on the labeling process, please refer to Fig 6a." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.688, + 0.292, + 0.705 + ], + "angle": 0, + "content": "4. Phys-AD Benchmark" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.396, + 0.73 + ], + "angle": 0, + "content": "4.1. Problem Definition and Challenges" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.779 + ], + "angle": 0, + "content": "We establish unsupervised and weakly supervised AD settings for Phys-AD, using unsupervised as the default in our problem definition." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.781, + 0.407, + 0.796 + ], + "angle": 0, + "content": "We formulate our Phys-AD setting as two steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Step 1: Rules deduction Given a set of training objects \\(\\mathcal{T} = \\{t_i\\}_{i=1}^N\\) from category \\(c_i\\), we first use robotics arms and motors to interact with \\(\\mathcal{T}\\) in corresponding methods \\(I_i\\), and we get the interaction process in video format as \\(V_i\\). Then, we feed the video sequence \\(V_i\\) and the corresponding category's physical prior \\(P_i\\) together to the deduction function \\(f\\) to obtain the normal rules \\(r_i\\). After rules deduction for all the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.09, + 0.907, + 0.121 + ], + "angle": 0, + "content": "categories, we get the rules bank \\(\\mathcal{M} = \\{r_i\\}_{i=1}^N\\). Step 1 can be formulated as below:" + }, + { + "type": "equation", + "bbox": [ + 0.671, + 0.134, + 0.907, + 0.15 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} \\Theta I _ {i} = V _ {i} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.664, + 0.16, + 0.907, + 0.176 + ], + "angle": 0, + "content": "\\[\nf \\left(V _ {i} \\mid P _ {i}\\right) = r _ {i} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.666, + 0.181, + 0.907, + 0.2 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} = \\left\\{r _ {i} \\right\\} _ {i = 1} ^ {N} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.206, + 0.908, + 0.235 + ], + "angle": 0, + "content": "\\(\\Theta\\) represents interaction between objects and robotics arms. \\(N\\) is the total number of categories." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.236, + 0.907, + 0.327 + ], + "angle": 0, + "content": "Step 2: Anomaly reasoning During test time, we first use robotics arms and motors to interact with \\(\\mathcal{T}\\) from the test set in corresponding methods \\(I_{i}\\), and we get the interaction process in video format as \\(V_{i}\\). Then, we feed the video sequence \\(V_{i}\\) and rules \\(r_i\\) from step 1 to the reasoner \\(R\\) to predict the object's anomaly score \\(S\\). Step 2 can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.672, + 0.338, + 0.907, + 0.354 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} \\Theta I _ {i} = V _ {i} \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.639, + 0.365, + 0.907, + 0.381 + ], + "angle": 0, + "content": "\\[\nR \\left(V _ {i} \\mid r _ {i}\\right) = S, r _ {i} \\in \\mathcal {M} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.909, + 0.493 + ], + "angle": 0, + "content": "Challenges. Key challenges include combining video information with objects' physical priors to deduct normal rules, capturing long-term temporal dependencies and fine-grained frame-level information, ensuring robust reasoning for anomaly prediction, and generalizing across diverse anomaly patterns. Addressing these challenges is essential for advancing anomaly detection in real-world environment." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.502, + 0.7, + 0.518 + ], + "angle": 0, + "content": "4.2. Evaluation Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.644 + ], + "angle": 0, + "content": "Unsupervised AD. Unsupervised AD is the most common IAD setting for existing IAD datasets and algorithms. Under unsupervised AD setting, the training set contains only normal video data, and the algorithm needs to capture the normal distribution of the data from the training set. In the test set, both normal and abnormal data are included, and the algorithm must distinguish between normal and abnormal data based on the distribution learned during training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.909, + 0.811 + ], + "angle": 0, + "content": "Weakly-supervised AD. Our dataset is in the form of videos for complex anomaly detection in industrial scenarios. In the video context, it is inevitable that we need to discuss weakly supervised anomaly detection. Under weakly-supervised Phys-AD setting, 2~4 video-level labeled anomaly samples are sampled from all possible anomaly classes in the test set in our Phys-AD dataset. These sampled anomalies are then removed from the test set. It's worth noting that we only provide video level label in our test set. This is because most anomalies in our test set require reasoning based on the entire video content and physical knowledge of the objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Video-Understanding AD. Video-Understanding models are another potential solution to our Phys-AD setting. Similar to unsupervised setting, we only provide normal videos during training. The visual language models need to provide explicit normal rules in language format during training. In the test phase, these video-understanding AD" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30414" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.131 + ], + "angle": 0, + "content": "Table 4. Video-level AUROC (↑) result of 22 categories on Phys-AD dataset. We include Unsupervised, Weakly-supervised and Video-understanding methods. The best per-category result for each class of methods is highlighted in **bold**. 'ZS ImgB', 'V-ChatGPT', 'V-LLaMA', 'V-LLaVA' denote ZS ImageBind, Video-ChatGPT, Video-LLaMA and Video-LLaVA." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.137, + 0.905, + 0.466 + ], + "angle": 0, + "content": "
Category.UnsupervisedWeakly-supervisedVideo-understanding
MPN [34]MemAE [16]MNAD.p [40]MNAD.r [40]SVM [44]VADClip [51]S3R [48]MGFN [10]LAVAD [56]ZS Clip [42]ZS ImgB [15]V-ChatGPT [36]V-LLaMA [59]V-LLaVA [28]
Car0.2290.5230.4920.9440.5870.5810.6060.5710.5570.5000.5000.5000.6780.522
Fan0.8110.3710.8100.5420.5000.6240.6400.5420.5100.5000.5000.5490.5920.611
Rolling Bearing0.3530.0440.3520.8000.9330.5890.6010.6800.5320.5000.5000.3000.5000.500
Spherical Bearing0.1130.0920.9620.8130.6500.5000.6820.5830.4350.5000.5000.4500.5500.500
Servo0.3640.4450.9750.8780.5000.5180.5920.5560.5020.5000.5000.5060.6830.464
Clip0.5350.4430.6300.3330.5000.4120.5610.5630.5160.5000.5000.6690.5560.458
U Disk0.2400.6170.6090.9400.5000.5300.5490.5700.5130.5000.5000.5650.5750.500
Hinge0.7690.8700.7050.8950.5000.7370.5610.5500.5640.5000.5000.5000.6920.500
Sticky Roller0.9670.9670.4510.9360.5000.5420.8350.6690.2660.5000.5000.4500.5440.467
Caster Wheel0.3640.5230.2710.5080.5000.5870.6760.6760.6150.5000.5000.4440.6420.500
Screw0.5220.5670.6800.5470.5000.5000.6570.5410.6880.5000.5000.4720.2560.550
Lock0.5630.5970.4300.6410.7330.4520.5230.6260.3410.5000.5000.2790.4940.500
Gear0.5290.5100.6520.6940.5000.5190.5800.5440.6030.5000.5000.5440.6030.517
Clock0.3400.5420.3950.5870.5000.5000.5490.5090.5000.5000.5000.5010.3600.500
Slide0.5170.9620.9170.7840.5000.5310.6110.5680.4250.5000.5000.5620.5670.179
Zipper0.8150.5920.8290.4210.5000.5040.6360.6330.5350.5000.5000.5470.4890.500
Button0.8530.3650.6600.5680.5000.6270.5150.5660.4390.5000.5000.5150.5170.360
Liquid0.1840.7000.6710.8310.5000.5420.5950.7930.5040.5000.5000.4100.2780.217
Rubber Band0.3740.3680.3660.3070.5670.4820.6230.6040.5110.5000.5000.5170.4500.450
Ball0.5430.3830.7280.6870.5000.5000.6670.5670.6030.5000.5000.5620.6360.533
Magnet0.6710.4640.6910.4380.5000.5000.5480.7190.5020.5000.5000.6830.3000.400
Toothpaste0.5870.8890.5200.6310.5000.5000.6860.7110.5620.5000.5000.3760.5500.467
Average0.5110.5380.6270.6690.5440.5350.6130.6060.5100.5000.5000.4960.5230.463
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.485, + 0.482, + 0.531 + ], + "angle": 0, + "content": "methods have to truly understand the videos and predict the anomalies based on the language rules withdrawn from the training phase." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.557, + 0.275, + 0.572 + ], + "angle": 0, + "content": "4.3. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.483, + 0.671 + ], + "angle": 0, + "content": "Physics anomaly accuracy metrics. We use the Area Under the Receiver Operating Characteristic Curve (AUROC) to evaluate video-level anomaly detection performance. We also report the average precision (AP), i.e. the area under the video-level precision-recall curve and the acc (ACCURACY), following previous works [47, 49]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Physics Anomaly Explanation (PAEval) metric. Directly utilizing existing VLMs [59][36] to understand the videos in our dataset and detect the anomalies is a potential solution to our challenge. The key point is: Can existing VLMs truly understand physical rules and reason in a right way? Specifically designed for video-understanding methods, we introduce a new metric named Physics Anomaly Explanation (PAEval) metric. To be more specific, PAEval evaluate the anomaly detection performance of algorithms based on VLMs from three different perspectives: classification, description, and explanation. Classification refers to the traditional anomaly detection metrics like AUROC, etc. Inspired by the works [11, 61], PAEval also introduces two additional evaluation metrics: description and physical explanation. Description refers to the model's ability to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.907, + 0.637 + ], + "angle": 0, + "content": "correctly describe the content of the video, used to assess whether the model has the capability to describe physical phenomena. Explanation refers to the model's ability to correctly explain the physical causes of anomalies in the video, used to evaluate the VLM's reasoning ability. To provide labels for description and explanation metrics, we manually labeled each type of defects from all categories and performed data augmentation by ChatGPT to ensure the robustness of the detection. The whole pipeline of PAEval metric is depicted in Fig 6." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.655, + 0.727, + 0.672 + ], + "angle": 0, + "content": "5. Benchmarking Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.681, + 0.785, + 0.696 + ], + "angle": 0, + "content": "5.1. Benchmark Methods Selection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.901 + ], + "angle": 0, + "content": "For the Phys-AD setting, we select popular and reproducible video anomaly detection methods across unsupervised, weakly-supervised, and video-understanding setting. In the unsupervised setting, we focus on reconstruction, prediction, and embedding-based models like MemAE [16] and MNAD [40], which use memory modules to enhance anomaly discrimination. For weakly-supervised anomaly detection, we adopt methods such as VadCLIP [51] and MGFN [10], which use feature magnitudes and vision-language associations. In video-understanding, we evaluate video-language models like Video-ChatGPT [36] and image-language models like CLIP [42] to predict detailed anomaly descriptions and scores through cross-modal embeddings." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30415" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.09, + 0.483, + 0.117 + ], + "angle": 0, + "content": "Table 5. Physics Anomaly Explanation (PAEval) metric results on our Phys-AD dataset." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.12, + 0.481, + 0.19 + ], + "angle": 0, + "content": "
MethodsClassification (↑)Description (↑)Explanation (↑)
LAVAD[56]0.5100.1570.000
Video-ChatGPT[36]0.4960.1310.160
Video-LLaMA[59]0.5230.1370.303
Video-LLaVA[28]0.4630.2190.282
" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.21, + 0.483, + 0.256 + ], + "angle": 0, + "content": "Code and Experiment details. We provide code and toolkit for our dataset and benchmark. More experiment details are listed in Appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.264, + 0.24, + 0.281 + ], + "angle": 0, + "content": "5.2. Main Findings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.286, + 0.484, + 0.438 + ], + "angle": 0, + "content": "Overall anomaly detection benchmarking results. Table 4 shows that existing video anomaly detection and videowhildunderstanding methods achieve limited performance on the Phys-AD dataset, with the highest AUROC only reaching \\(66.9\\%\\) for MNAD.r [40]. This result underscores the heightened complexity of Phys-AD compared to existing datasets, highlighting a domain shift that challenges current industrial anomaly detection algorithms, which are often tuned to visually distinct anomaly patterns in single frames rather than complex temporal or physical cues." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.441, + 0.484, + 0.652 + ], + "angle": 0, + "content": "Anomaly detection via unsupervised and weakly supervised methods. Our experiment includes unsupervised methods (e.g., MemAE [16], MPN [34], MNAD [40]) and weakly supervised methods (e.g., S3R [48], MGFN [10], VADClip [51]). Unsupervised methods like MNAD.p [40] performed better on temporal anomalies (81.0% for fan, 68.0% for screw) by leveraging prediction-based approaches, which excel in scenarios requiring temporal understanding. Weakly supervised methods improve baseline scores across complex classes, preventing extreme low scores in challenging categories such as spherical bearing. However, weakly supervised methods show marginally lower performance in simpler anomaly classes, indicating trade-offs introduced by anomaly samples in training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.483, + 0.807 + ], + "angle": 0, + "content": "Unsupervised anomaly detection method performance by category. Among unsupervised methods, MemAE [16] achieves high AUROC on objects with spatial anomalies (e.g., sticky roller, \\(96.7\\%\\)) but underperforms in temporal anomaly classes like rolling and spherical bearings, where it achieves only \\(4.4\\%\\) and \\(9.2\\%\\), respectively. MNAD [40] improves temporal sensitivity by incorporating temporal prototypes, achieving better scores for bearings but still struggling with complex anomaly types, revealing limitations in purely prototype-based approaches." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Anomaly detection via Video-Language Models. VLM-based methods (e.g., Video-ChatGPT [36], VideoLLaMA [59]) struggle, with top-performing VideoLLaMA [59] reaching only \\(52.3\\%\\) average AUROC. Performance is impacted by reliance on pre-trained weights that are not optimized for physics-grounded video content, evident" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.182 + ], + "angle": 0, + "content": "from low AUROC scores in categories with nuanced physical dynamics, such as hinges and screws. Further, PAEval results suggest that these models lack effective reasoning about object physical dynamics and behaviors influenced by physical forces, underscoring a gap between VLM capabilities and the demands of IAD tasks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.186, + 0.909, + 0.292 + ], + "angle": 0, + "content": "Anomaly explanation via Video-Language Models. Table 5 reports PAEval metric results, with the best-performing VLM achieving only \\(21.9\\%\\) in description and \\(30.3\\%\\) in explanation. These findings emphasize that current VLMs lack the depth in physical reasoning and temporal coherence required for understanding real-world physics-based scenarios, which our Phys-AD dataset demands." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.295, + 0.909, + 0.387 + ], + "angle": 0, + "content": "Summary. Overall, the results reveal Phys-AD's unique challenge in requiring both high spatial detail and temporal comprehension, areas where existing methods and models underperform. This analysis points to the need for future research in models that integrate temporal reasoning with physics-based anomaly detection." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.4, + 0.634, + 0.415 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.909, + 0.681 + ], + "angle": 0, + "content": "In this paper, we introduce the first industrial anomaly detection task focusing on real-world scenarios where physical understanding and reasoning are essential for anomaly detection. We present the Physics Anomaly Detection (Phys-AD) dataset, a large-scale, physics-grounded video dataset with over 6400 videos across 22 categories and 49 object types interacting with robotic systems, capturing 47 anomaly types that necessitate visual and physical understanding. We assess Phys-AD, highlighting the lack of baseline approaches for high-level reasoning in anomaly detection. Additionally, we propose the Physics Anomaly Explanation (PAEval) metric to evaluate visual language models (VLMs) on physics-based reasoning. Experiments show that current VLMs fall short of human-level understanding in physics-based anomaly scenarios. This work marks a milestone for industrial anomaly detection, promoting physics-grounded reasoning in complex industrial settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.685, + 0.909, + 0.867 + ], + "angle": 0, + "content": "Limitation and future work. Although our Phys-AD dataset provides a large variety of objects with diverse physical properties and various types of interaction methods, we plan to add even more diverse interaction methods and objects in the future to better meet the demands of complex real-world industrial scenarios. Due to the significant differences between our Phys-AD dataset and current industrial anomaly detection and video anomaly detection datasets, most existing anomaly detection algorithms cannot be directly applied to our dataset. In the future, we will test more algorithms on our Phys-AD dataset and provide experimental results across more settings like zero-shot, few-shot, semi-supervised settings, etc." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "30416" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.157 + ], + "angle": 0, + "content": "[1] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In ICCV, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.484, + 0.201 + ], + "angle": 0, + "content": "[2] Fabien Baradel, Natalia Neverova, Julien Mille, Greg Mori, and Christian Wolf. Cophy: Counterfactual learning of physical dynamics. arXiv preprint arXiv:1909.12000, 2019.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.203, + 0.483, + 0.272 + ], + "angle": 0, + "content": "[3] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad — a comprehensive real-world dataset for unsupervised anomaly detection. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9584-9592, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.274, + 0.483, + 0.315 + ], + "angle": 0, + "content": "[4] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. MVtec AD – A comprehensive real-world dataset for unsupervised anomaly detection. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.318, + 0.484, + 0.386 + ], + "angle": 0, + "content": "[5] Paul Bergmann, Kilian Batzner, Michael Fauser, David Sattlegger, and Carsten Steger. Beyond dents and scratches: Logical constraints in unsupervised anomaly detection and localization. International Journal of Computer Vision, 130 (4):947-969, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.389, + 0.484, + 0.471 + ], + "angle": 0, + "content": "[6] Paul Bergmann, Xin Jin, David Sattlegger, and Carsten Steger. The mvtec 3d-ad dataset for unsupervised 3d anomaly detection and localization. In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications. SCITEPRESS - Science and Technology Publications, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.474, + 0.483, + 0.514 + ], + "angle": 0, + "content": "[7] Paul Bergmann, Xin Jin, David Sattlegger, and Carsten Steger. The MVTec 3d-AD dataset for unsupervised 3d anomaly detection and localization. In VISAPP, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.484, + 0.586 + ], + "angle": 0, + "content": "[8] Luca Bonfiglioli, Marco Toschi, Davide Silvestri, Nicola Fioraio, and Daniele De Gregorio. The eyecandies dataset for unsupervised multimodal anomaly detection and localization. In Proceedings of the Asian Conference on Computer Vision, pages 3586-3602, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.589, + 0.484, + 0.643 + ], + "angle": 0, + "content": "[9] Yunkang Cao, Xiaohao Xu, Jiangning Zhang, Yuqi Cheng, Xiaonan Huang, Guansong Pang, and Weiming Shen. A survey on visual anomaly detection: Challenge, approach, and prospect. arXiv preprint arXiv:2401.16402, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.646, + 0.484, + 0.728 + ], + "angle": 0, + "content": "[10] Yingxian Chen, Zhengzhe Liu, Baoheng Zhang, Wilton Fok, Xiaojuan Qi, and Yik-Chung Wu. Mgfn: Magnitude-contrastive glance-and-focus network for weakly-supervised video anomaly detection. In Proceedings of the AAAI conference on artificial intelligence, pages 387-395, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.731, + 0.484, + 0.827 + ], + "angle": 0, + "content": "[11] Hang Du, Sicheng Zhang, Binzhu Xie, Guoshun Nan, Jiayang Zhang, Junrui Xu, Hangyu Liu, Sicong Leng, Jiangming Liu, Hehe Fan, et al. Uncovering what why and how: A comprehensive benchmark for causation understanding of video anomaly. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18793-18803, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.83, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[12] Jiafei Duan, Samson Yu, and Cheston Tan. Space: A simulator for physical interactions and causal learning in 3d environments. In ICCV, pages 2058-2063, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Jiafei Duan, Samson Yu, Soujanya Poria, Bihan Wen, and Cheston Tan. Pip: Physical interaction prediction via mental" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.907, + 0.12 + ], + "angle": 0, + "content": "simulation with span selection. In ECCV, pages 405-421. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.907, + 0.188 + ], + "angle": 0, + "content": "[14] Jia-Chang Feng, Fa-Ting Hong, and Wei-Shi Zheng. Mist: Multiple instance self-training framework for video anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14009-14018, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.907, + 0.259 + ], + "angle": 0, + "content": "[15] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.261, + 0.907, + 0.354 + ], + "angle": 0, + "content": "[16] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1705-1714, 2019. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.357, + 0.906, + 0.412 + ], + "angle": 0, + "content": "[17] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.906, + 0.455 + ], + "angle": 0, + "content": "[18] Mahmudul Hasan, Jonghyun Choi, Jan Neumann, Amit K Roy-Chowdhury, and Larry S Davis. Learning temporal regularity in video sequences. In CVPR, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.906, + 0.496 + ], + "angle": 0, + "content": "[19] Drew A. Hudson and Christopher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.907, + 0.551 + ], + "angle": 0, + "content": "[20] Steeven Janny, Fabien Baradel, Natalia Neverova, Madiha Nadri, Greg Mori, and Christian Wolf. Filtered-cophy: Unsupervised learning of counterfactual physics in pixel space. In ICLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.554, + 0.907, + 0.635 + ], + "angle": 0, + "content": "[21] Stepan Jezek, Martin Jonak, Radim Burget, Pavel Dvorak, and Milos Skotak. Deep learning-based defect detection of metal parts: evaluating current methods in complex conditions. In 2021 13th International Congress on Ultra Modern Telecommunications and Control Systems and Workshops (ICUMT), pages 66-71, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.906, + 0.691 + ], + "angle": 0, + "content": "[22] Huaizu Jiang, Xiaojian Ma, Weili Nie, Zhiding Yu, Yuke Zhu, and Anima Anandkumar. Bongard-hoi: Benchmarking few-shot visual reasoning for human-object interactions. In CVPR, pages 19056-19065, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.693, + 0.907, + 0.747 + ], + "angle": 0, + "content": "[23] Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C. Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In CVPR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.907, + 0.815 + ], + "angle": 0, + "content": "[24] Hyekang Kevin Joo, Khoa Vo, Kashu Yamazaki, and Ngan Le. Clip-tsa: Clip-assisted temporal self-attention for weakly-supervised video anomaly detection. In 2023 IEEE International Conference on Image Processing (ICIP), pages 3230-3234. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.907, + 0.858 + ], + "angle": 0, + "content": "[25] Federico Landi, Cees GM Snok, and Rita Cucchiara. Anomaly locality in video surveillance. arXiv preprint arXiv:1901.10364, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[26] Ting Li, Charles Dubout, Emma K Wampler, Steven Yantis, Donald Geman, et al. Comparing machines and humans on a visual categorization test. 2011. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30417" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.148 + ], + "angle": 0, + "content": "[27] Wenqiao Li, Xiaohao Xu, Yao Gu, Bozhong Zheng, Shenghua Gao, and Yingna Wu. Towards scalable 3d anomaly detection and localization: A benchmark via 3d anomaly synthesis and a self-supervised learning network. arXiv preprint, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.483, + 0.204 + ], + "angle": 0, + "content": "[28] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.207, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[29] Jiaqi Liu, Guoyang Xie, Rui Chen, et al. Real3d-ad: A dataset of point cloud anomaly detection. NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[30] Jiaqi Liu, Guoyang Xie, Ruitao Chen, Xinpeng Li, Jinbao Wang, Yong Liu, Chengjie Wang, and Feng Zheng. Real3dad: A dataset of point cloud anomaly detection. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.483, + 0.347 + ], + "angle": 0, + "content": "[31] Kun Liu and Huadong Ma. Exploring background-bias for anomaly detection in surveillance videos. In Proceedings of the 27th ACM International Conference on Multimedia, pages 1490-1499, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.483, + 0.39 + ], + "angle": 0, + "content": "[32] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection-a new baseline. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.393, + 0.483, + 0.474 + ], + "angle": 0, + "content": "[33] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13588-13597, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.483, + 0.545 + ], + "angle": 0, + "content": "[34] Hui Lv, Chen Chen, Zhen Cui, Chunyan Xu, Yong Li, and Jian Yang. Learning normal dynamics in videos with meta prototype network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15425-15434, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.483, + 0.616 + ], + "angle": 0, + "content": "[35] Hui Lv, Zhongqi Yue, Qianru Sun, Bin Luo, Zhen Cui, and Hanwang Zhang. Unbiased multiple instance learning for weakly supervised video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8022-8031, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.483, + 0.673 + ], + "angle": 0, + "content": "[36] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.483, + 0.743 + ], + "angle": 0, + "content": "[37] Domingo Mery, Vladimir Riffo, Uwe Zscherpel, German Mondragon, Ivan Lillo, Irene Zuccar, Hans Lobel, and Miguel Carrasco. Gdxray: The database of x-ray images for nondestructive testing. Journal of Nondestructive Evaluation, 34 (4):42, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[38] Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. Vt-adj: A vision transformer network for image anomaly detection and localization. In 2021 IEEE 30th International Symposium on Industrial Electronics (ISIE). IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[39] Weili Nie, Zhiding Yu, Lei Mao, Ankit B Patel, Yuke Zhu, and Animashree Anandkumar. Bongard-logo: A new benchmark for human-level concept learning and reasoning. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[40] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14372-14381, 2020. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[41] Yujiang Pu, Xiaoyu Wu, and Shengjin Wang. Learning prompt-enhanced context features for weakly-supervised video anomaly detection. arXiv preprint arXiv:2306.14451, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.907, + 0.244 + ], + "angle": 0, + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[43] Binyi Su, Zhong Zhou, and Haiyong Chen. Pvel-ad: A large-scale open-world dataset for photovoltaic cell anomaly detection. IEEE Transactions on Industrial Informatics, 19 (1):404-413, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.329 + ], + "angle": 0, + "content": "[44] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In CVPR, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[45] Kamalakar Vijay Thakare, Yash Raghuwanshi, Debi Prosad Dogra, Heeseung Choi, and Ig-Jae Kim. Dyannet: A scene dynamicity guided self-trained video anomaly detection network. In Proceedings of the IEEE/CVF Winter conference on applications of computer vision, pages 5541-5550, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.401, + 0.907, + 0.466 + ], + "angle": 0, + "content": "[46] Chengjie Wang, Wenbing Zhu, Bin-Bin Gao, Zhenye Gan, Jianning Zhang, Zhihao Gu, Shuguang Qian, Mingang Chen, and Lizhuang Ma. Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detection, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[47] Yue Wang, Jinlong Peng, Jiangning Zhang, Ran Yi, Yabiao Wang, and Chengjie Wang. Multimodal industrial anomaly detection via hybrid fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8032-8041, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[48] Jhih-Ciang Wu, He-Yen Hsieh, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng-Luh Liu. Self-supervised sparse representation for video anomaly detection. In European Conference on Computer Vision, pages 729-745. Springer, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.597, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[49] Peng Wu, Jing Liu, Yujia Shi, Yujia Sun, Fangtao Shao, Zhaoyang Wu, and Zhiwei Yang. Not only look, but also listen: Learning multimodal violence detection under weak supervision. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 322-339. Springer, 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[50] Peng Wu, Xuerong Zhou, Guansong Pang, Yujia Sun, Jing Liu, Peng Wang, and Yanning Zhang. Open-vocabulary video anomaly detection. arXiv preprint arXiv:2311.07042, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[51] Peng Wu, Xuerong Zhou, Guansong Pang, Lingru Zhou, Qingsen Yan, Peng Wang, and Yanning Zhang. Vadclip: Adapting vision-language models for weakly supervised video anomaly detection. arXiv preprint arXiv:2308.11681, 2023.3,7,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[52] Dan Xu, Yan Yan, Elisa Ricci, and Nicu Sebe. Detecting anomalous events in videos by learning deep representations of appearance and motion. Computer Vision and Image Understanding, 156:117-127, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[53] Xiaohao Xu, Yunkang Cao, Yongqi Chen, Weiming Shen, and Xiaonan Huang. Customizing visual-language foundation models for multi-modal anomaly detection and reasoning. arXiv preprint arXiv:2403.11083, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "30418" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[54] Zhiwei Yang, Jing Liu, Zhaoyang Wu, Peng Wu, and Xiaotao Liu. Video event restoration based on keyframes for video anomaly detection. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[55] M Zaigham Zaheer, Arif Mahmood, M Haris Khan, Mattia Segu, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14744-14754, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.261 + ], + "angle": 0, + "content": "[56] Luca Zanella, Willi Menapace, Massimiliano Mancini, Yiming Wang, and Elisa Ricci. Harnessing large language models for training-free video anomaly detection. arXiv preprint arXiv:2404.01014, 2024. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.304 + ], + "angle": 0, + "content": "[57] Aimen Zerroug, Mohit Vaishnav, Julien Colin, Sebastian Musslick, and Thomas Serre. A benchmark for compositional visual reasoning. arXiv preprint arXiv:2206.05379, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.483, + 0.347 + ], + "angle": 0, + "content": "[58] Chi Zhang, Feng Gao, Baoxiong Jia, Yixin Zhu, and Song-Chun Zhu. Raven: A dataset for relational and analogical visual reasoning. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.483, + 0.389 + ], + "angle": 0, + "content": "[59] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.483, + 0.459 + ], + "angle": 0, + "content": "[60] Huaxin Zhang, Xiang Wang, Xiaohao Xu, Xiaonan Huang, Chuchu Han, Yuehuan Wang, Changxin Gao, Shanjun Zhang, and Nong Sang. Glancevad: Exploring glance supervision for label-efficient video anomaly detection. arXiv preprint arXiv:2403.06154, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.483, + 0.53 + ], + "angle": 0, + "content": "[61] Huaxin Zhang, Xiaohao Xu, Xiang Wang, Jialong Zuo, Chuchu Han, Xiaonan Huang, Changxin Gao, Yuehuan Wang, and Nong Sang. Holmes-vad: Towards unbiased and explainable video anomaly detection via multi-modal llm. arXiv preprint arXiv:2406.12235, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.483, + 0.587 + ], + "angle": 0, + "content": "[62] Yiheng Zhang, Yunkang Cao, Xiaohao Xu, and Weiming Shen. Logic: An llm-driven framework for logical anomaly detection. IEEE Transactions on Automation Science and Engineering, pages 1-0, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.483, + 0.642 + ], + "angle": 0, + "content": "[63] Jia-Xing Zhong, Nannan Li, Weijie Kong, Shan Liu, Thomas H Li, and Ge Li. Graph convolutional label noise cleaner: Train a plug-and-play action classifier for anomaly detection. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.645, + 0.483, + 0.686 + ], + "angle": 0, + "content": "[64] Qiang Zhou, Weize Li, Lihan Jiang, Guoliang Wang, Guyue Zhou, Shanghang Zhang, and Hao Zhao. Pad: A dataset and benchmark for pose-agnostic anomaly detection, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.483, + 0.741 + ], + "angle": 0, + "content": "[65] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.744, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[66] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.786 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30419" + } + ] +] \ No newline at end of file diff --git a/2025/Towards Visual Discrimination and Reasoning of Real-World Physical Dynamics_ Physics-Grounded Anomaly Detection/e496d86c-0d1a-42d1-b098-7f552a70c4ff_origin.pdf b/2025/Towards Visual Discrimination and Reasoning of Real-World Physical Dynamics_ Physics-Grounded Anomaly Detection/e496d86c-0d1a-42d1-b098-7f552a70c4ff_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0512d2eb358c3cfdd1722e89de55dbbc206a5c7a --- /dev/null +++ b/2025/Towards Visual Discrimination and Reasoning of Real-World Physical Dynamics_ Physics-Grounded Anomaly Detection/e496d86c-0d1a-42d1-b098-7f552a70c4ff_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8bfb8a7a032072416846dba924bf24ba41acba814630d98a722422664d82b18 +size 1307578 diff --git a/2025/Towards Visual Discrimination and Reasoning of Real-World Physical Dynamics_ Physics-Grounded Anomaly Detection/full.md b/2025/Towards Visual Discrimination and Reasoning of Real-World Physical Dynamics_ Physics-Grounded Anomaly Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..927bbf7bf5abea59ad3e5f875443c82ab391e08a --- /dev/null +++ b/2025/Towards Visual Discrimination and Reasoning of Real-World Physical Dynamics_ Physics-Grounded Anomaly Detection/full.md @@ -0,0 +1,349 @@ +# Towards Visual Discrimination and Reasoning of Real-World Physical Dynamics: Physics-Grounded Anomaly Detection + +Wenqiao Li $^{1*}$ Yao Gu $^{1*}$ Xintao Chen $^{1*}$ Xiaohao Xu $^{2\dagger}$ Ming Hu $^{3}$ Xiaonan Huang $^{2}$ Yingna Wu $^{1\dagger}$ $^{1}$ ShanghaiTech University $^{2}$ University of Michigan, Ann Arbor $^{3}$ Monash University + +![](images/75d6ea3cc4000b98b604cf49e2eedc38b5b4f3a012591964ab21612bc71ce942.jpg) +Figure 1. Towards visual discrimination of physical dynamics in real-world industrial object anomaly detection. We illustrate objects, interactions, and time-sequenced videos from the Physics-Grounded Anomaly Detection dataset: (a) Object; (b) Interaction: Applied actions shown with directional arrows; (c) Video with Physical Dynamics: Temporal sequences showing normal and abnormal states, highlighting anomalies like leaks, misalignments, and cracks. By focusing on the dynamic behaviors of complex objects, we enhance understanding of interactions and failure modes in real-world settings, where both structure and motion contribute to anomaly detection. + +![](images/406ce781989bb97be996195c92fcaa5cde7a373227f124df11e22d77c6ab559c.jpg) + +# Abstract + +Humans detect real-world object anomalies by perceiving, interacting, and reasoning based on object-conditioned physical knowledge. The long-term goal of Industrial Anomaly Detection (IAD) is to enable machines to autonomously replicate this skill. However, current IAD algorithms are largely developed and tested on static, semantically simple datasets, which diverge from real-world scenarios where physical understanding and reasoning are essential. To bridge this gap, we introduce the Physics Anomaly Detection (Phys-AD) dataset, the first large-scale, real-world, physics-grounded video dataset for industrial anomaly detection. Collected using a real robot arm and motor, Phys-AD provides a diverse + +set of dynamic, semantically rich scenarios. The dataset includes more than 6400 videos across 22 real-world object categories, interacting with robot arms and motors, and exhibits 47 types of anomalies. Anomaly detection in Phys-AD requires visual reasoning, combining both physical knowledge and video content to determine object abnormality. We benchmark state-of-the-art anomaly detection methods under three settings: unsupervised AD, weakly-supervised AD, and video-understanding AD, highlighting their limitations in handling physics-grounded anomalies. Additionally, we introduce the Physics Anomaly Explanation (PAEval) metric, designed to assess the ability of visual-language foundation models to not only detect anomalies but also provide accurate explanations for their underlying physical causes. Our project is available at https://guyao2023.github.io/Phys-AD/. + +![](images/fc4dc8c941e0683bab05b70a96d80230028a00681e4d6d0074810515380bb537.jpg) +Figure 2. Human-like decision-making process for physics-grounded object anomaly detection. We illustrate the sequential approach of a human-like agent for evaluating an object's normality. First, the agent perceives relevant physical attributes (e.g., plastic and elastic), then interacts by performing a physical action (e.g., squeezing), and finally reasons based on the vision feedback and attributes changes (e.g., surface shape change) to determine whether the object is normal or anomalous. This mirrors a human's natural process of reasoning over physics in objects. + +# 1. Introduction + +Industrial anomaly detection (IAD) is a critical subfield in computer vision and industrial automation, aiming to identify defects or irregularities in products during manufacturing. As shown in Fig. 2, the ultimate vision is to create autonomous systems that not only perceive but also interact with and reason about objects to discriminate between anomalies and normal states, integrating complex physical principles to detect anomalies in dynamic, real-world scenarios. For example, a human inspecting a water bottle for anomalies wouldn't rely solely on visual observation; they might rotate or invert the bottle, using physical interactions and cues, such as noticing a loose cap or an irregular internal flow, to detect issues. + +Central to advancing IAD [9] is the availability of high-quality datasets that bridge the gap between academic research and industrial needs. Datasets like MVTec-AD [4], MPDD [21], and VisA [65] have played foundational roles, enabling algorithm development for image-based anomaly detection and bringing IAD to the forefront of computer vision research. While these datasets have significantly advanced single-image anomaly detection, recent datasets, e.g., MVTec-3D [7], Real3D [29], and Anomaly-ShapeNet [27], have extended IAD to 3D, aligning research more closely with the needs of complex real-world industrial settings. + +Yet, as factories increasingly rely on robotic arms and automated systems to perform sophisticated inspections, the limitations of current IAD datasets become apparent. + +Existing benchmarks focus on static, semantically simple environments, overlooking the physical priors and interactive reasoning required in real-world industrial contexts. This gap highlights a growing need for datasets that not only reflect real-world physical constraints but also challenge models to reason dynamically about anomalies. + +To bridge this gap, as shown in Fig. 1, we introduce the Physics Anomaly Detection (Phys-AD) dataset, the first large-scale, physics-grounded video dataset for industrial anomaly detection. Phys-AD features over 6,400 videos of 22 categories and 49 objects interacting with robotic arms and motors, capturing 47 types of anomalies that require visual reasoning informed by physical knowledge. The short video clips in the dataset range from 60 to 240 frames in length and are filmed in real-time industrial environment, fully capturing the interaction process between robotic arm or motor and industrial objects. Additionally, to ensure our dataset meets industrial demands and matches the complexity and diversity of the real physical world, we selected industrial objects of different physical qualities, various interaction methods, and anomalies that reflect diverse physical principles and require different reasoning process. Specifically, we selected 22 object categories spanning across metals, plastics, fluid, amorphous substances and articulated objects with diverse appearances. For interaction, we use mechanical grippers, robotic arms, and motors, incorporating various interaction modes such as pressing, rotating, squeezing, and driving to handle different types of objects. + +We benchmark state-of-the-art anomaly detection methods in three key configurations: unsupervised anomaly detection, weakly-supervised anomaly detection, and video-understanding based anomaly detection. Our findings reveal critical gaps in their ability to handle the complexities of physics-grounded scenarios, where anomalies often arise from dynamic, interdependent interactions. To advance the field, we also introduce the Physical Anomaly Explanation (PAEval) metric, designed to assess both detection performance and a model's ability to explain anomalies by identifying underlying physical causes. Furthermore, our benchmark reveals the fragility of existing methods in tackling these challenging conditions, underscoring the need for approaches that better understand object dynamics and temporal coherence in anomaly detection. + +# Our contributions can be summarized as followings: + +- We introduce a novel task of detecting physical-based industrial anomalies in real-world that involves perception, physical and visual reasoning. +- We present Phys-AD, the first large-scale, physics-grounded video dataset specifically designed for industrial anomaly detection in real world, containing objects of different physical qualities, multiple interaction methods and various physical reasoning process. +- We benchmarking the anomaly detection and reasoning + +Table 1. Comparison of Phys-AD with existing industrial anomaly detection datasets. Our Phys-AD dataset is the first to consider complex objects with physical dynamics. 'Syn', 'IR', 'D', and 'PC' denote Synthetic, Infrared, Depth, and Point Cloud, respectively. #Anomaly indicates the number of anomaly types. + +
DatasetYearTypeModalitySample Statistics
Class #AnomalyPhysics
MVTec-AD [3]2019RealRGB15-X
BTAD [38]2021RealRGB33X
MPDD [21]2021RealRGB68X
VisA [66]2021RealRGB12-X
MVTec LOCO-AD [5]2022RealRGB5-X
MAD [64]2023Syn+RealRGB203X
LOCO-Annotations [62]2024RealRGB55X
Real-IAD [46]2024RealRGB308X
GDXray [37]2015RealX-ray515X
PVEL-AD [43]2023RealIR110X
MVTec3D-AD [6]2021RealRGB-D103-5X
Eyecandies [8]2022SynRGB-D103X
Real3D-AD [30]2023RealPC122X
Anomaly-ShapeNet [27]2024SynPC406X
Phys-AD (Ours)2024RealRGB4947✓
+ +performance of popular video AD methods and Visual Language Foundation Models on the Phys-AD dataset in several settings, establishing a practical and challenging benchmark to promote the development of the physics-related anomaly detection field. + +# 2. Related work + +Industrial anomaly detection datasets. Existing datasets primarily focus on static, semantically simple scenarios, deviate significantly from real world where physical understanding and reasoning are essential. Datasets like MVTec-AD [3], BTAD [38], MPDD [21], and VisA [66] focus on surface level image anomaly detection with one single-view RGB image, limiting their effectiveness in capturing holistic object structures. MVTec-LOCO-AD [5], focusing on structure and local information in industries, is limited by its relatively simple and constrained data content. While MVTec3D-AD [6] and Eyecandies [8] incorporate depth data, they remain static and single-view image anomaly detection, neglecting object level information. To explore object level anomaly detection, multi-view IAD datasets like MAD [64] and Real-IAD [46], point cloud IAD datasets like Real3D-AD [30] and Anomaly-ShapeNet [27], offering richer visual and geometric cues, but are still limited to static objects without dynamic interaction or reasoning. In summary, current industrial anomaly detection datasets focus on relatively simple and static anomaly detection scenarios, lacking complex physical rules, dynamic interactions, and visual reasoning requirements. Therefore, existing IAD datasets generally applied to limited industrial scenarios and there is no IAD dataset could meet the demands of detecting + +complex anomalies in real world which need physical priors and reasoning. + +Video anomaly detection. Deep learning methods [18, 35, 54, 61] now dominate Video Anomaly Detection (VAD), categorized into unsupervised, weakly-supervised, and fully-supervised approaches. Unsupervised methods learn normal patterns via reconstruction [16, 18, 52], prediction [32], or hybrids [33], while some methods [45, 55] train with both unlabeled normal and abnormal data. Weakly-supervised methods[14, 49, 60, 63] use video-level or glance-based annotations, and fully-supervised methods[25, 31] remain rare due to costly frame-level labeling. Visual-language models like CLIP [42] have recently been applied to enhance anomaly detection [24, 41, 51], focusing on semantic anomalies. Open-vocabulary VAD [50] and prompt-based anomaly scoring [56] leverage LLMs [53, 61], but performance relies heavily on the base models, often lacking domain-specific tuning. However, existing video anomaly detection algorithms lack the capability to handle complex industrial anomaly detection scenarios and understand physical rules. This gap highlights the need for models that can capture dynamic behaviors and physical laws in industrial environments, as addressed by Phys-AD, which targets industrial anomalies in objects of various physical properties. + +Visual reasoning. Visual reasoning is a critical task in computer vision, aiming to enable machines to interpret perceptual information like humans. Several visual reasoning tasks have been proposed to evaluate reasoning capabilities, including Visual Question Answering (VQA), 2D puzzles, and physical dynamics prediction. In VQA, agents are required to combine natural language and visual cues to answer questions [1, 17, 19, 23]. For 2D puzzles, tasks involve discovering relationships among visual elements and making inferences [22, 26, 39, 57, 58]. Physical dynamics prediction tasks require machines to perceive and reason about physical interactions [2, 12, 13, 20]. In contrast to these work, we introduce the first benchmark featuring real-world industrial objects with dynamic physical properties, focusing on distinguishing diverse dynamics through vision. + +# 3. Dataset: Phys-AD + +# 3.1. Object Preparation and Interaction Selection + +We selected 22 categories spanning across various materials, including metals, plastics, amorphous substances and articulated objects, with diverse shapes, sizes, and physical properties. For the physical properties of different objects, we correspondingly select different interaction methods like push, rotate, pull with robotics arms and motors. For instance, we use robotics arms to grab and extruding the deformable objects like basketball to determine whether there are any elastic anomalies or surface defects by the morphological change of the objects. To make our dataset more practical + +![](images/d1fb0c57315c70053b2eee90ea39523b6bc62eca99f6e996b4571a08712f0d89.jpg) +U Disk + +![](images/04c3a0c007d031f41abd5c7c98604aad9ab1aaf4129e18d28427e92deee1eb37.jpg) +Zipper + +![](images/45bcc2f974a31f7bc5fa856043b21b280954cb11438e7f549aef3542277029ad.jpg) +Rolling Bearing + +![](images/b7d51d7075f154f89366e928da49136a71a85cee7477c389e2eb7f9afe4c02c5.jpg) +Sticky Roller + +![](images/138a01ee9acd66f4fd6acc7aa80c8d225c1ae73de2587940ad0378ef6c70897b.jpg) +Slider + +![](images/6b60ccaaf439eb96ea39c5a3b9fa413b510be46701455468c04e51538753bdc3.jpg) +Toothpaste + +![](images/0b3e191f36b3ed06901894a1493699ae31532b6433e29d9994adcf9e77502b7d.jpg) +Door Lock + +![](images/de166171b1fd2c1d1d14cfe9f71d84aeee36e3408be644fb0740baed93afadf8.jpg) +Toy Car + +![](images/90f777c39de74347453e303735bb358ab7dd2c4abc98c699069be3d12fd72216.jpg) +Clip + +![](images/1324e3603fe14c0ee8596a894f5b4b71efea4e5df255eef133e2c529b00767b4.jpg) +Fan + +![](images/79a5bcee9d78b800f3c96612ba744acd04d2d65ab5444af43ed720b70edc3bcb.jpg) +Rubber Band + +![](images/a93b776c01182baf642c844fd4003427bef1d20f803d1921c7b718ef3b4de6e3.jpg) +Magnet + +![](images/ef6d8ff60375a009ba23164350c9f70fa0b56f37d63ffd49a93da1a6e504579a.jpg) +Servo + +![](images/4d6dc56067eb39e38a2c6d664695e2645359b85f86b75c0ba77d4771ba20a87f.jpg) +Gear Assembly + +![](images/143b400b63356474d7a583349339b8c3fd2ca6429bf0a91836a0f9968ad6acbb.jpg) +Bottled Water + +![](images/db4e4ea50cb523a34663c3de3d8555feba8fc3e602d608de9919f4a00022692f.jpg) +Screw + +![](images/5d07bbf049c4ab67cb7cefb294f93b1adecd62b7b746b16a2b25c80a1f2688ff.jpg) +Spherical Bearing + +![](images/ae3b8ab1d44dca0c371de14f0f6e04fbffda0755c5a46cb2291463b29b4ca39e.jpg) +Motion + +![](images/30b520f00ef352421bb17433f5d67fdc49435e0a6a68430c5acb0b5e28c6e3da.jpg) +Button + +![](images/d477cbe0731ea4b43372c2b74c2807e6242f0ecfdb1ee561599aa52988b7e88b.jpg) +Clock + +![](images/32405e8d76a58bd525059e3dbd12ba4e95ee2af92a48352e8327c8bf0cb5a0a5.jpg) +Caster Wheel + +![](images/d020a461c07f5b611e5475dca37bd6451a9ee9fd00ee94a5502173746bfcce9d.jpg) +Hinge + +![](images/5d39dd0d1961330a708b001bbd0af0330e7e2274175ee736aa65e240d83b341e.jpg) +Ball + +![](images/ee9640bd724f2776c510eb025a226563c9518fd3ee2fc98b17e081ca1a1b534d.jpg) +Direction For Interaction +Figure 3. Interactions for understanding implicit physical laws in the Phys-AD dataset. We showcase various object interactions from the Phys-AD dataset, where different actions (indicated by motion directions) are used to explore and reason about the underlying physical properties and behaviors of each object. The colored arrows indicate the interaction directions and axes, highlighting how physical interactions reveal the implicit physics governing each object. + +![](images/b3f99f89cbed8591fccb3110e6a970871e8521d23d9bbcbaaea6ed729034642a.jpg) +Normal: The U Disk rotates freely over time. + +![](images/6768b2f107bdf1fc48650b414447219c6c19cc384203b8c07f0ee7c7436e2cb8.jpg) +Abnormal: The U Disk is stuck over time. +(a) Persistent Anomalies + +![](images/fc43f6a365f2a63ec36586f1a0cf725753bce008123e8aed6dd463dcf945a622.jpg) +Normal: The sticky roller rotates smoothly on the surface. + +![](images/0fd51d148225da32c20aed5e679cb20d87bc532ca66ea2e17693bcb16094c8c2.jpg) +Normal: The sticky roller functions properly at the beginning. +Abnormal: The sticky roller breaks at a certain point. +(b) Intermittent Anomalies +Figure 4. Categorization of anomalies based on persistence in the Phys-AD dataset. We show examples of normal and abnormal functioning in common objects, divided into two anomaly types: persistent and intermittent. (a) Persistent anomalies, such as continuous obstruction in the U Disk or permanent malfunction of the Sticky Roller, are visible throughout the operation. (b) In contrast, intermittent anomalies, like occasional jamming of the U Disk or breakage in the Sticky Roller after initial operation, only appear at specific points in time. This classification provides insight into both constant and sporadic failures in object interactions. + +and challenging, we introduce 47 distinct defect types, some of which just rely on single frame content and physical rules for anomaly reasoning, while the other part needs to combine the content of the whole video and physical priors together to judge whether the object is abnormal or not. Different objects information with their corresponding interaction methods and anomalies are listed in Table 2. Fig 3 shows the interac + +tions for understanding implicit physical laws in the Phys-AD dataset. Fig 4 provides two representative examples: one where a U Disk requires analyzing the entire video content to determine if there is an anomaly, and another where a sticky roller only requires a few frames from the video to judge whether the object is abnormal. + +![](images/c3ffd1eb93313cb723ec32e3219be946461890aafac571d6445aa6c70c56801c.jpg) +Figure 5. Data collection pipeline for the Phys-AD dataset. (a) Manipulation of a toothpaste tube using a UR5 robotic arm. (b) Manipulation of a U Disk and fan via servo and motor. + +![](images/e4add3a5da0de42ca6b80027e5c1fbdd5fd5651b741bdb4da4c922c84fbf789b.jpg) + +Table 2. Interaction methods and anomalies of the object categories in our Phys-AD dataset. Motor. means one object driven by motor. Multi. means one object with multiple anomalies. + +
CategoryInteractionAnomaly Types
CarDrag, SlideDifferent wheels stuck
FanMotor., RotateStuck, Uneven Rotation, Vibration
Rolling BearingMotor., RotateLack of friction
Spherical BearingGrab, RotateInternal block
ServoGrab, Motor., RotateAngle restricted, Vibration, No calibration
ClipPressUnable to press down, Unable to rebound
U DiskGrab, Motor., RotateCover jam
HingeGrab, RotateAngle restricted, Shaft off
Sticky RollerGrab, PullDetach, Unable to rotate
Caster WheelSlideAxle axis stuck, Swivel axis stuck
ScrewPress, RotateLoosen, Unable to insert
LockMotor., RotateLatch jam, Loose
GearMotor., RotateStuck, Not meshed, Multi.
ClockMotor., RotatePointer stops, Vibration
SlideGrab, SlideDetach, Shedding, Jam
ZipperGrab, CloseStuck, Unable to close
ButtonPressNo light, Unable to press down, Unable to rebound, Multi.
LiquidGrab, ShakeWater out, Foreign body
Rubber BandStretchCrack
BallPinch, PressInsufficient gas, Leakage
MagnetGrab, Press, MoveDegaussing, Shell detached
ToothpastePinch, PressLeakage
+ +# 3.2. Data Collection and Processing + +Data collection pipeline. Most of the data collection for the Phys-AD dataset is driven by manipulation-guided video sequences, captured using a UR5 robotic arm equipped with an RGB camera (see Fig. 5a). In order to reproduce real industrial scenes, we also include adjustable light sources in our data capture process. The RGB Camera, with a 1080p resolution and a frame rate of 60 FPS, providing high-quality video sequences. Some kinds of the objects like U Disk that are not suitable for the manipulation of robotics arms, are driven by the motor or servo (Fig. 5b). After the data collection, we used video editing software to crop out irrelevant frames and retain the complete interaction process. + +Table 3. Phys-AD dataset statistics. We denote the total number of frames and videos for each category as #Images and #Videos. Note that the Train split does not contain anomaly data. + +
CategoryTrain #FramesTrain #VideosTest #FramesTest #VideosTest #Videos Nor.Test #Videos Ab.Obj. TypesDef. Types
Car18,00030036,00060015045053
Fan32,40018064,8003609027033
Rolling Bearing3,600603,60060303011
Spherical Bearing3,600603,60060303011
Servo14,40012028,8002406018013
Clip28,80024043,20036012024042
U Disk14,40024014,40024012012041
Hinge3,600307,20060154512
Sticky Roller5,400308,10045153012
Caster Wheel5,4003010,80060154512
Screw5,400308,10045153012
Lock7,20012010,8001806012012
Gear21,60018054,0004509036034
Clock27,00015040,3202247315152
Slide7,2006018,0001503012013
Zipper14,40012021,6001806012022
Button21,60012054,0003006024044
Liquid5,400308,10045153012
Rubber Band10,8006010,80060303011
Ball21,6009032,400135453032
Magnet10,8006016,20090306022
Toothpaste16,2009016,20090454531
Total298,8002400511,0204,0341,1982,8364947
+ +# 3.3. Data Statistics + +Dataset sample distribution. Table 3 provides a detailed breakdown of the Phys-AD dataset, which includes information on category, the number of training videos and frames, the number of testing videos and frames, the distribution of normal and abnormal samples in the test set, and the number of object and defect types. The length of the video clips in the dataset ranges from 60 to 240 frames, ensuring to fully capture the interaction process in a short time. The dataset contains 2400 training videos and 4034 test videos spreading across 22 categories, 49 object types, and 47 defect types. The frames are extracted from the videos at a rate of 60 FPS. The test set includes 1,198 normal samples and 2,836 abnormal samples. Overall, the Phys-AD dataset covers a wide range of categories and contains a large scale of data, which helps to train more robust anomaly detection models and provides a reasonable evaluation setting. + +Interaction and defect types across categories. Table 2 details the defect types and corresponding interaction methods for each category in the Phys-AD dataset. Covering 22 categories and including 47 defect types, the dataset features various interactions (e.g., Rotate, Grab). Importantly, interaction methods or defect types with identical names can differ across categories—for example, "rotation" for a fan describes the motor-driven movement of its blades, while for a hinge it refers to the turning of its pages around the shaft under robotic control. Furthermore, to detect anomalies with complex physical properties in a human-like manner, we sometimes combine multiple interactions. For instance, when inspecting a screw, we first press it into the hole and then use a robotic arm to rotate it further. Overall, Phys-AD provides a diverse, physics-grounded anomaly detection framework that fosters the development of advanced methods for tackling real-world challenges. + +![](images/37b08d3a354aa271391e904fcbd765b6bb939dde01b0e2dacc73c68dd212e63c.jpg) +Figure 6. PhysAD-Agent: A Large Language Model (LLM)-powered system for physics anomaly detection label augmentation and automatic evaluation. This agent framework consists of two main components: (a) Rich Annotation Generation, where an LLM-based data agent generates detailed anomaly annotations based on video, anomaly state, prompt, and human-annotated physical priors, and (b) Automatic Evaluation, where an LLM-based evaluation agent assesses model predictions to calculate the Physics Anomaly Explanation (PAEval) metric. + +Labels. For the test set videos, we provide two types of labels. First, there are the common video-level labels: all anomalous videos in the test set are labeled as 1, while normal videos are labeled as 0. For evaluating visual-language models, we also provide text labels. Specifically, for each type of anomaly in the test set, we manually design a textual description label. This label includes both a description of the video content and a physical explanation of the reason for the anomaly. To ensure diversity in our text labels, we also use ChatGPT-4o for text augmentation. For further details on the labeling process, please refer to Fig 6a. + +# 4. Phys-AD Benchmark + +# 4.1. Problem Definition and Challenges + +We establish unsupervised and weakly supervised AD settings for Phys-AD, using unsupervised as the default in our problem definition. + +We formulate our Phys-AD setting as two steps. + +Step 1: Rules deduction Given a set of training objects $\mathcal{T} = \{t_i\}_{i=1}^N$ from category $c_i$ , we first use robotics arms and motors to interact with $\mathcal{T}$ in corresponding methods $I_i$ , and we get the interaction process in video format as $V_i$ . Then, we feed the video sequence $V_i$ and the corresponding category's physical prior $P_i$ together to the deduction function $f$ to obtain the normal rules $r_i$ . After rules deduction for all the + +categories, we get the rules bank $\mathcal{M} = \{r_i\}_{i=1}^N$ . Step 1 can be formulated as below: + +$$ +\mathcal {T} \Theta I _ {i} = V _ {i} \tag {1} +$$ + +$$ +f \left(V _ {i} \mid P _ {i}\right) = r _ {i} \tag {2} +$$ + +$$ +\mathcal {M} = \left\{r _ {i} \right\} _ {i = 1} ^ {N} \tag {3} +$$ + +$\Theta$ represents interaction between objects and robotics arms. $N$ is the total number of categories. + +Step 2: Anomaly reasoning During test time, we first use robotics arms and motors to interact with $\mathcal{T}$ from the test set in corresponding methods $I_{i}$ , and we get the interaction process in video format as $V_{i}$ . Then, we feed the video sequence $V_{i}$ and rules $r_i$ from step 1 to the reasoner $R$ to predict the object's anomaly score $S$ . Step 2 can be formulated as: + +$$ +\mathcal {T} \Theta I _ {i} = V _ {i} \tag {4} +$$ + +$$ +R \left(V _ {i} \mid r _ {i}\right) = S, r _ {i} \in \mathcal {M} \tag {5} +$$ + +Challenges. Key challenges include combining video information with objects' physical priors to deduct normal rules, capturing long-term temporal dependencies and fine-grained frame-level information, ensuring robust reasoning for anomaly prediction, and generalizing across diverse anomaly patterns. Addressing these challenges is essential for advancing anomaly detection in real-world environment. + +# 4.2. Evaluation Settings + +Unsupervised AD. Unsupervised AD is the most common IAD setting for existing IAD datasets and algorithms. Under unsupervised AD setting, the training set contains only normal video data, and the algorithm needs to capture the normal distribution of the data from the training set. In the test set, both normal and abnormal data are included, and the algorithm must distinguish between normal and abnormal data based on the distribution learned during training. + +Weakly-supervised AD. Our dataset is in the form of videos for complex anomaly detection in industrial scenarios. In the video context, it is inevitable that we need to discuss weakly supervised anomaly detection. Under weakly-supervised Phys-AD setting, 2~4 video-level labeled anomaly samples are sampled from all possible anomaly classes in the test set in our Phys-AD dataset. These sampled anomalies are then removed from the test set. It's worth noting that we only provide video level label in our test set. This is because most anomalies in our test set require reasoning based on the entire video content and physical knowledge of the objects. + +Video-Understanding AD. Video-Understanding models are another potential solution to our Phys-AD setting. Similar to unsupervised setting, we only provide normal videos during training. The visual language models need to provide explicit normal rules in language format during training. In the test phase, these video-understanding AD + +Table 4. Video-level AUROC (↑) result of 22 categories on Phys-AD dataset. We include Unsupervised, Weakly-supervised and Video-understanding methods. The best per-category result for each class of methods is highlighted in **bold**. 'ZS ImgB', 'V-ChatGPT', 'V-LLaMA', 'V-LLaVA' denote ZS ImageBind, Video-ChatGPT, Video-LLaMA and Video-LLaVA. + +
Category.UnsupervisedWeakly-supervisedVideo-understanding
MPN [34]MemAE [16]MNAD.p [40]MNAD.r [40]SVM [44]VADClip [51]S3R [48]MGFN [10]LAVAD [56]ZS Clip [42]ZS ImgB [15]V-ChatGPT [36]V-LLaMA [59]V-LLaVA [28]
Car0.2290.5230.4920.9440.5870.5810.6060.5710.5570.5000.5000.5000.6780.522
Fan0.8110.3710.8100.5420.5000.6240.6400.5420.5100.5000.5000.5490.5920.611
Rolling Bearing0.3530.0440.3520.8000.9330.5890.6010.6800.5320.5000.5000.3000.5000.500
Spherical Bearing0.1130.0920.9620.8130.6500.5000.6820.5830.4350.5000.5000.4500.5500.500
Servo0.3640.4450.9750.8780.5000.5180.5920.5560.5020.5000.5000.5060.6830.464
Clip0.5350.4430.6300.3330.5000.4120.5610.5630.5160.5000.5000.6690.5560.458
U Disk0.2400.6170.6090.9400.5000.5300.5490.5700.5130.5000.5000.5650.5750.500
Hinge0.7690.8700.7050.8950.5000.7370.5610.5500.5640.5000.5000.5000.6920.500
Sticky Roller0.9670.9670.4510.9360.5000.5420.8350.6690.2660.5000.5000.4500.5440.467
Caster Wheel0.3640.5230.2710.5080.5000.5870.6760.6760.6150.5000.5000.4440.6420.500
Screw0.5220.5670.6800.5470.5000.5000.6570.5410.6880.5000.5000.4720.2560.550
Lock0.5630.5970.4300.6410.7330.4520.5230.6260.3410.5000.5000.2790.4940.500
Gear0.5290.5100.6520.6940.5000.5190.5800.5440.6030.5000.5000.5440.6030.517
Clock0.3400.5420.3950.5870.5000.5000.5490.5090.5000.5000.5000.5010.3600.500
Slide0.5170.9620.9170.7840.5000.5310.6110.5680.4250.5000.5000.5620.5670.179
Zipper0.8150.5920.8290.4210.5000.5040.6360.6330.5350.5000.5000.5470.4890.500
Button0.8530.3650.6600.5680.5000.6270.5150.5660.4390.5000.5000.5150.5170.360
Liquid0.1840.7000.6710.8310.5000.5420.5950.7930.5040.5000.5000.4100.2780.217
Rubber Band0.3740.3680.3660.3070.5670.4820.6230.6040.5110.5000.5000.5170.4500.450
Ball0.5430.3830.7280.6870.5000.5000.6670.5670.6030.5000.5000.5620.6360.533
Magnet0.6710.4640.6910.4380.5000.5000.5480.7190.5020.5000.5000.6830.3000.400
Toothpaste0.5870.8890.5200.6310.5000.5000.6860.7110.5620.5000.5000.3760.5500.467
Average0.5110.5380.6270.6690.5440.5350.6130.6060.5100.5000.5000.4960.5230.463
+ +methods have to truly understand the videos and predict the anomalies based on the language rules withdrawn from the training phase. + +# 4.3. Evaluation Metrics + +Physics anomaly accuracy metrics. We use the Area Under the Receiver Operating Characteristic Curve (AUROC) to evaluate video-level anomaly detection performance. We also report the average precision (AP), i.e. the area under the video-level precision-recall curve and the acc (ACCURACY), following previous works [47, 49]. + +Physics Anomaly Explanation (PAEval) metric. Directly utilizing existing VLMs [59][36] to understand the videos in our dataset and detect the anomalies is a potential solution to our challenge. The key point is: Can existing VLMs truly understand physical rules and reason in a right way? Specifically designed for video-understanding methods, we introduce a new metric named Physics Anomaly Explanation (PAEval) metric. To be more specific, PAEval evaluate the anomaly detection performance of algorithms based on VLMs from three different perspectives: classification, description, and explanation. Classification refers to the traditional anomaly detection metrics like AUROC, etc. Inspired by the works [11, 61], PAEval also introduces two additional evaluation metrics: description and physical explanation. Description refers to the model's ability to + +correctly describe the content of the video, used to assess whether the model has the capability to describe physical phenomena. Explanation refers to the model's ability to correctly explain the physical causes of anomalies in the video, used to evaluate the VLM's reasoning ability. To provide labels for description and explanation metrics, we manually labeled each type of defects from all categories and performed data augmentation by ChatGPT to ensure the robustness of the detection. The whole pipeline of PAEval metric is depicted in Fig 6. + +# 5. Benchmarking Results + +# 5.1. Benchmark Methods Selection + +For the Phys-AD setting, we select popular and reproducible video anomaly detection methods across unsupervised, weakly-supervised, and video-understanding setting. In the unsupervised setting, we focus on reconstruction, prediction, and embedding-based models like MemAE [16] and MNAD [40], which use memory modules to enhance anomaly discrimination. For weakly-supervised anomaly detection, we adopt methods such as VadCLIP [51] and MGFN [10], which use feature magnitudes and vision-language associations. In video-understanding, we evaluate video-language models like Video-ChatGPT [36] and image-language models like CLIP [42] to predict detailed anomaly descriptions and scores through cross-modal embeddings. + +Table 5. Physics Anomaly Explanation (PAEval) metric results on our Phys-AD dataset. + +
MethodsClassification (↑)Description (↑)Explanation (↑)
LAVAD[56]0.5100.1570.000
Video-ChatGPT[36]0.4960.1310.160
Video-LLaMA[59]0.5230.1370.303
Video-LLaVA[28]0.4630.2190.282
+ +Code and Experiment details. We provide code and toolkit for our dataset and benchmark. More experiment details are listed in Appendix. + +# 5.2. Main Findings + +Overall anomaly detection benchmarking results. Table 4 shows that existing video anomaly detection and videowhildunderstanding methods achieve limited performance on the Phys-AD dataset, with the highest AUROC only reaching $66.9\%$ for MNAD.r [40]. This result underscores the heightened complexity of Phys-AD compared to existing datasets, highlighting a domain shift that challenges current industrial anomaly detection algorithms, which are often tuned to visually distinct anomaly patterns in single frames rather than complex temporal or physical cues. + +Anomaly detection via unsupervised and weakly supervised methods. Our experiment includes unsupervised methods (e.g., MemAE [16], MPN [34], MNAD [40]) and weakly supervised methods (e.g., S3R [48], MGFN [10], VADClip [51]). Unsupervised methods like MNAD.p [40] performed better on temporal anomalies (81.0% for fan, 68.0% for screw) by leveraging prediction-based approaches, which excel in scenarios requiring temporal understanding. Weakly supervised methods improve baseline scores across complex classes, preventing extreme low scores in challenging categories such as spherical bearing. However, weakly supervised methods show marginally lower performance in simpler anomaly classes, indicating trade-offs introduced by anomaly samples in training. + +Unsupervised anomaly detection method performance by category. Among unsupervised methods, MemAE [16] achieves high AUROC on objects with spatial anomalies (e.g., sticky roller, $96.7\%$ ) but underperforms in temporal anomaly classes like rolling and spherical bearings, where it achieves only $4.4\%$ and $9.2\%$ , respectively. MNAD [40] improves temporal sensitivity by incorporating temporal prototypes, achieving better scores for bearings but still struggling with complex anomaly types, revealing limitations in purely prototype-based approaches. + +Anomaly detection via Video-Language Models. VLM-based methods (e.g., Video-ChatGPT [36], VideoLLaMA [59]) struggle, with top-performing VideoLLaMA [59] reaching only $52.3\%$ average AUROC. Performance is impacted by reliance on pre-trained weights that are not optimized for physics-grounded video content, evident + +from low AUROC scores in categories with nuanced physical dynamics, such as hinges and screws. Further, PAEval results suggest that these models lack effective reasoning about object physical dynamics and behaviors influenced by physical forces, underscoring a gap between VLM capabilities and the demands of IAD tasks. + +Anomaly explanation via Video-Language Models. Table 5 reports PAEval metric results, with the best-performing VLM achieving only $21.9\%$ in description and $30.3\%$ in explanation. These findings emphasize that current VLMs lack the depth in physical reasoning and temporal coherence required for understanding real-world physics-based scenarios, which our Phys-AD dataset demands. + +Summary. Overall, the results reveal Phys-AD's unique challenge in requiring both high spatial detail and temporal comprehension, areas where existing methods and models underperform. This analysis points to the need for future research in models that integrate temporal reasoning with physics-based anomaly detection. + +# 6. Conclusion + +In this paper, we introduce the first industrial anomaly detection task focusing on real-world scenarios where physical understanding and reasoning are essential for anomaly detection. We present the Physics Anomaly Detection (Phys-AD) dataset, a large-scale, physics-grounded video dataset with over 6400 videos across 22 categories and 49 object types interacting with robotic systems, capturing 47 anomaly types that necessitate visual and physical understanding. We assess Phys-AD, highlighting the lack of baseline approaches for high-level reasoning in anomaly detection. Additionally, we propose the Physics Anomaly Explanation (PAEval) metric to evaluate visual language models (VLMs) on physics-based reasoning. Experiments show that current VLMs fall short of human-level understanding in physics-based anomaly scenarios. This work marks a milestone for industrial anomaly detection, promoting physics-grounded reasoning in complex industrial settings. + +Limitation and future work. Although our Phys-AD dataset provides a large variety of objects with diverse physical properties and various types of interaction methods, we plan to add even more diverse interaction methods and objects in the future to better meet the demands of complex real-world industrial scenarios. Due to the significant differences between our Phys-AD dataset and current industrial anomaly detection and video anomaly detection datasets, most existing anomaly detection algorithms cannot be directly applied to our dataset. In the future, we will test more algorithms on our Phys-AD dataset and provide experimental results across more settings like zero-shot, few-shot, semi-supervised settings, etc. + +# References + +[1] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In ICCV, 2015. 3 +[2] Fabien Baradel, Natalia Neverova, Julien Mille, Greg Mori, and Christian Wolf. Cophy: Counterfactual learning of physical dynamics. arXiv preprint arXiv:1909.12000, 2019.3 +[3] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad — a comprehensive real-world dataset for unsupervised anomaly detection. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9584-9592, 2019. 3 +[4] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. MVtec AD – A comprehensive real-world dataset for unsupervised anomaly detection. In CVPR, 2019. 2 +[5] Paul Bergmann, Kilian Batzner, Michael Fauser, David Sattlegger, and Carsten Steger. Beyond dents and scratches: Logical constraints in unsupervised anomaly detection and localization. International Journal of Computer Vision, 130 (4):947-969, 2022. 3 +[6] Paul Bergmann, Xin Jin, David Sattlegger, and Carsten Steger. The mvtec 3d-ad dataset for unsupervised 3d anomaly detection and localization. In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications. SCITEPRESS - Science and Technology Publications, 2022. 3 +[7] Paul Bergmann, Xin Jin, David Sattlegger, and Carsten Steger. The MVTec 3d-AD dataset for unsupervised 3d anomaly detection and localization. In VISAPP, 2022. 2 +[8] Luca Bonfiglioli, Marco Toschi, Davide Silvestri, Nicola Fioraio, and Daniele De Gregorio. The eyecandies dataset for unsupervised multimodal anomaly detection and localization. In Proceedings of the Asian Conference on Computer Vision, pages 3586-3602, 2022. 3 +[9] Yunkang Cao, Xiaohao Xu, Jiangning Zhang, Yuqi Cheng, Xiaonan Huang, Guansong Pang, and Weiming Shen. A survey on visual anomaly detection: Challenge, approach, and prospect. arXiv preprint arXiv:2401.16402, 2024. 2 +[10] Yingxian Chen, Zhengzhe Liu, Baoheng Zhang, Wilton Fok, Xiaojuan Qi, and Yik-Chung Wu. Mgfn: Magnitude-contrastive glance-and-focus network for weakly-supervised video anomaly detection. In Proceedings of the AAAI conference on artificial intelligence, pages 387-395, 2023. 7, 8 +[11] Hang Du, Sicheng Zhang, Binzhu Xie, Guoshun Nan, Jiayang Zhang, Junrui Xu, Hangyu Liu, Sicong Leng, Jiangming Liu, Hehe Fan, et al. Uncovering what why and how: A comprehensive benchmark for causation understanding of video anomaly. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18793-18803, 2024. 7 +[12] Jiafei Duan, Samson Yu, and Cheston Tan. Space: A simulator for physical interactions and causal learning in 3d environments. In ICCV, pages 2058-2063, 2021. 3 +[13] Jiafei Duan, Samson Yu, Soujanya Poria, Bihan Wen, and Cheston Tan. Pip: Physical interaction prediction via mental + +simulation with span selection. In ECCV, pages 405-421. Springer, 2022. 3 +[14] Jia-Chang Feng, Fa-Ting Hong, and Wei-Shi Zheng. Mist: Multiple instance self-training framework for video anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14009-14018, 2021. 3 +[15] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 7 +[16] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1705-1714, 2019. 3, 7, 8 +[17] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 3 +[18] Mahmudul Hasan, Jonghyun Choi, Jan Neumann, Amit K Roy-Chowdhury, and Larry S Davis. Learning temporal regularity in video sequences. In CVPR, 2016. 3 +[19] Drew A. Hudson and Christopher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, 2019. 3 +[20] Steeven Janny, Fabien Baradel, Natalia Neverova, Madiha Nadri, Greg Mori, and Christian Wolf. Filtered-cophy: Unsupervised learning of counterfactual physics in pixel space. In ICLR, 2022. 3 +[21] Stepan Jezek, Martin Jonak, Radim Burget, Pavel Dvorak, and Milos Skotak. Deep learning-based defect detection of metal parts: evaluating current methods in complex conditions. In 2021 13th International Congress on Ultra Modern Telecommunications and Control Systems and Workshops (ICUMT), pages 66-71, 2021. 2, 3 +[22] Huaizu Jiang, Xiaojian Ma, Weili Nie, Zhiding Yu, Yuke Zhu, and Anima Anandkumar. Bongard-hoi: Benchmarking few-shot visual reasoning for human-object interactions. In CVPR, pages 19056-19065, 2022. 3 +[23] Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C. Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In CVPR, 2017. 3 +[24] Hyekang Kevin Joo, Khoa Vo, Kashu Yamazaki, and Ngan Le. Clip-tsa: Clip-assisted temporal self-attention for weakly-supervised video anomaly detection. In 2023 IEEE International Conference on Image Processing (ICIP), pages 3230-3234. IEEE, 2023. 3 +[25] Federico Landi, Cees GM Snok, and Rita Cucchiara. Anomaly locality in video surveillance. arXiv preprint arXiv:1901.10364, 2019. 3 +[26] Ting Li, Charles Dubout, Emma K Wampler, Steven Yantis, Donald Geman, et al. Comparing machines and humans on a visual categorization test. 2011. 3 + +[27] Wenqiao Li, Xiaohao Xu, Yao Gu, Bozhong Zheng, Shenghua Gao, and Yingna Wu. Towards scalable 3d anomaly detection and localization: A benchmark via 3d anomaly synthesis and a self-supervised learning network. arXiv preprint, 2023. 2, 3 +[28] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 7, 8 +[29] Jiaqi Liu, Guoyang Xie, Rui Chen, et al. Real3d-ad: A dataset of point cloud anomaly detection. NeurIPS, 2023. 2 +[30] Jiaqi Liu, Guoyang Xie, Ruitao Chen, Xinpeng Li, Jinbao Wang, Yong Liu, Chengjie Wang, and Feng Zheng. Real3dad: A dataset of point cloud anomaly detection. Advances in Neural Information Processing Systems, 36, 2024. 3 +[31] Kun Liu and Huadong Ma. Exploring background-bias for anomaly detection in surveillance videos. In Proceedings of the 27th ACM International Conference on Multimedia, pages 1490-1499, 2019. 3 +[32] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection-a new baseline. In CVPR, 2018. 3 +[33] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13588-13597, 2021. 3 +[34] Hui Lv, Chen Chen, Zhen Cui, Chunyan Xu, Yong Li, and Jian Yang. Learning normal dynamics in videos with meta prototype network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15425-15434, 2021. 7, 8 +[35] Hui Lv, Zhongqi Yue, Qianru Sun, Bin Luo, Zhen Cui, and Hanwang Zhang. Unbiased multiple instance learning for weakly supervised video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8022-8031, 2023. 3 +[36] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 7, 8 +[37] Domingo Mery, Vladimir Riffo, Uwe Zscherpel, German Mondragon, Ivan Lillo, Irene Zuccar, Hans Lobel, and Miguel Carrasco. Gdxray: The database of x-ray images for nondestructive testing. Journal of Nondestructive Evaluation, 34 (4):42, 2015. 3 +[38] Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. Vt-adj: A vision transformer network for image anomaly detection and localization. In 2021 IEEE 30th International Symposium on Industrial Electronics (ISIE). IEEE, 2021. 3 +[39] Weili Nie, Zhiding Yu, Lei Mao, Ankit B Patel, Yuke Zhu, and Animashree Anandkumar. Bongard-logo: A new benchmark for human-level concept learning and reasoning. In NeurIPS, 2020. 3 +[40] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In Proceed- + +ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14372-14381, 2020. 7, 8 +[41] Yujiang Pu, Xiaoyu Wu, and Shengjin Wang. Learning prompt-enhanced context features for weakly-supervised video anomaly detection. arXiv preprint arXiv:2306.14451, 2023. 3 +[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 7 +[43] Binyi Su, Zhong Zhou, and Haiyong Chen. Pvel-ad: A large-scale open-world dataset for photovoltaic cell anomaly detection. IEEE Transactions on Industrial Informatics, 19 (1):404-413, 2023. 3 +[44] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In CVPR, 2018. 7 +[45] Kamalakar Vijay Thakare, Yash Raghuwanshi, Debi Prosad Dogra, Heeseung Choi, and Ig-Jae Kim. Dyannet: A scene dynamicity guided self-trained video anomaly detection network. In Proceedings of the IEEE/CVF Winter conference on applications of computer vision, pages 5541-5550, 2023. 3 +[46] Chengjie Wang, Wenbing Zhu, Bin-Bin Gao, Zhenye Gan, Jianning Zhang, Zhihao Gu, Shuguang Qian, Mingang Chen, and Lizhuang Ma. Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detection, 2024. 3 +[47] Yue Wang, Jinlong Peng, Jiangning Zhang, Ran Yi, Yabiao Wang, and Chengjie Wang. Multimodal industrial anomaly detection via hybrid fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8032-8041, 2023. 7 +[48] Jhih-Ciang Wu, He-Yen Hsieh, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng-Luh Liu. Self-supervised sparse representation for video anomaly detection. In European Conference on Computer Vision, pages 729-745. Springer, 2022. 7, 8 +[49] Peng Wu, Jing Liu, Yujia Shi, Yujia Sun, Fangtao Shao, Zhaoyang Wu, and Zhiwei Yang. Not only look, but also listen: Learning multimodal violence detection under weak supervision. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 322-339. Springer, 2020. 3, 7 +[50] Peng Wu, Xuerong Zhou, Guansong Pang, Yujia Sun, Jing Liu, Peng Wang, and Yanning Zhang. Open-vocabulary video anomaly detection. arXiv preprint arXiv:2311.07042, 2023. 3 +[51] Peng Wu, Xuerong Zhou, Guansong Pang, Lingru Zhou, Qingsen Yan, Peng Wang, and Yanning Zhang. Vadclip: Adapting vision-language models for weakly supervised video anomaly detection. arXiv preprint arXiv:2308.11681, 2023.3,7,8 +[52] Dan Xu, Yan Yan, Elisa Ricci, and Nicu Sebe. Detecting anomalous events in videos by learning deep representations of appearance and motion. Computer Vision and Image Understanding, 156:117-127, 2017. 3 +[53] Xiaohao Xu, Yunkang Cao, Yongqi Chen, Weiming Shen, and Xiaonan Huang. Customizing visual-language foundation models for multi-modal anomaly detection and reasoning. arXiv preprint arXiv:2403.11083, 2024. 3 + +[54] Zhiwei Yang, Jing Liu, Zhaoyang Wu, Peng Wu, and Xiaotao Liu. Video event restoration based on keyframes for video anomaly detection. In CVPR, 2023. 3 +[55] M Zaigham Zaheer, Arif Mahmood, M Haris Khan, Mattia Segu, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14744-14754, 2022. 3 +[56] Luca Zanella, Willi Menapace, Massimiliano Mancini, Yiming Wang, and Elisa Ricci. Harnessing large language models for training-free video anomaly detection. arXiv preprint arXiv:2404.01014, 2024. 3, 7, 8 +[57] Aimen Zerroug, Mohit Vaishnav, Julien Colin, Sebastian Musslick, and Thomas Serre. A benchmark for compositional visual reasoning. arXiv preprint arXiv:2206.05379, 2022. 3 +[58] Chi Zhang, Feng Gao, Baoxiong Jia, Yixin Zhu, and Song-Chun Zhu. Raven: A dataset for relational and analogical visual reasoning. In CVPR, 2019. 3 +[59] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 7, 8 +[60] Huaxin Zhang, Xiang Wang, Xiaohao Xu, Xiaonan Huang, Chuchu Han, Yuehuan Wang, Changxin Gao, Shanjun Zhang, and Nong Sang. Glancevad: Exploring glance supervision for label-efficient video anomaly detection. arXiv preprint arXiv:2403.06154, 2024. 3 +[61] Huaxin Zhang, Xiaohao Xu, Xiang Wang, Jialong Zuo, Chuchu Han, Xiaonan Huang, Changxin Gao, Yuehuan Wang, and Nong Sang. Holmes-vad: Towards unbiased and explainable video anomaly detection via multi-modal llm. arXiv preprint arXiv:2406.12235, 2024. 3, 7 +[62] Yiheng Zhang, Yunkang Cao, Xiaohao Xu, and Weiming Shen. Logic: An llm-driven framework for logical anomaly detection. IEEE Transactions on Automation Science and Engineering, pages 1-0, 2024. 3 +[63] Jia-Xing Zhong, Nannan Li, Weijie Kong, Shan Liu, Thomas H Li, and Ge Li. Graph convolutional label noise cleaner: Train a plug-and-play action classifier for anomaly detection. In CVPR, 2019. 3 +[64] Qiang Zhou, Weize Li, Lihan Jiang, Guoliang Wang, Guyue Zhou, Shanghang Zhang, and Hao Zhao. Pad: A dataset and benchmark for pose-agnostic anomaly detection, 2023. 3 +[65] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In ECCV, 2022. 2 +[66] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation, 2022. 3 \ No newline at end of file diff --git a/2025/Towards Visual Discrimination and Reasoning of Real-World Physical Dynamics_ Physics-Grounded Anomaly Detection/images.zip b/2025/Towards Visual Discrimination and Reasoning of Real-World Physical Dynamics_ Physics-Grounded Anomaly Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b5407d3de2cb7e69c65e1bef5099e39dc0bb636f --- /dev/null +++ b/2025/Towards Visual Discrimination and Reasoning of Real-World Physical Dynamics_ Physics-Grounded Anomaly Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:18674aaffabf880257995cbcd8ef3216381832b284b7184dc4fa30f9572bfb4e +size 800142 diff --git a/2025/Towards Visual Discrimination and Reasoning of Real-World Physical Dynamics_ Physics-Grounded Anomaly Detection/layout.json b/2025/Towards Visual Discrimination and Reasoning of Real-World Physical Dynamics_ Physics-Grounded Anomaly Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f1208371812095402f751c6ef4189755173d90a3 --- /dev/null +++ b/2025/Towards Visual Discrimination and Reasoning of Real-World Physical Dynamics_ Physics-Grounded Anomaly Detection/layout.json @@ -0,0 +1,9177 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 102, + 556, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 102, + 556, + 140 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 556, + 140 + ], + "type": "text", + "content": "Towards Visual Discrimination and Reasoning of Real-World Physical Dynamics: Physics-Grounded Anomaly Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 159, + 578, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 578, + 191 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 578, + 191 + ], + "type": "text", + "content": "Wenqiao Li" + }, + { + "bbox": [ + 61, + 159, + 578, + 191 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 61, + 159, + 578, + 191 + ], + "type": "text", + "content": " Yao Gu" + }, + { + "bbox": [ + 61, + 159, + 578, + 191 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 61, + 159, + 578, + 191 + ], + "type": "text", + "content": " Xintao Chen" + }, + { + "bbox": [ + 61, + 159, + 578, + 191 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 61, + 159, + 578, + 191 + ], + "type": "text", + "content": " Xiaohao Xu" + }, + { + "bbox": [ + 61, + 159, + 578, + 191 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 61, + 159, + 578, + 191 + ], + "type": "text", + "content": " Ming Hu" + }, + { + "bbox": [ + 61, + 159, + 578, + 191 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 159, + 578, + 191 + ], + "type": "text", + "content": " Xiaonan Huang" + }, + { + "bbox": [ + 61, + 159, + 578, + 191 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 159, + 578, + 191 + ], + "type": "text", + "content": " Yingna Wu" + }, + { + "bbox": [ + 61, + 159, + 578, + 191 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 61, + 159, + 578, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 159, + 578, + 191 + ], + "type": "text", + "content": "ShanghaiTech University " + }, + { + "bbox": [ + 61, + 159, + 578, + 191 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 159, + 578, + 191 + ], + "type": "text", + "content": "University of Michigan, Ann Arbor " + }, + { + "bbox": [ + 61, + 159, + 578, + 191 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 159, + 578, + 191 + ], + "type": "text", + "content": "Monash University" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 59, + 201, + 328, + 430 + ], + "blocks": [ + { + "bbox": [ + 59, + 201, + 328, + 430 + ], + "lines": [ + { + "bbox": [ + 59, + 201, + 328, + 430 + ], + "spans": [ + { + "bbox": [ + 59, + 201, + 328, + 430 + ], + "type": "image", + "image_path": "75d6ea3cc4000b98b604cf49e2eedc38b5b4f3a012591964ab21612bc71ce942.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 433, + 556, + 489 + ], + "lines": [ + { + "bbox": [ + 54, + 433, + 556, + 489 + ], + "spans": [ + { + "bbox": [ + 54, + 433, + 556, + 489 + ], + "type": "text", + "content": "Figure 1. Towards visual discrimination of physical dynamics in real-world industrial object anomaly detection. We illustrate objects, interactions, and time-sequenced videos from the Physics-Grounded Anomaly Detection dataset: (a) Object; (b) Interaction: Applied actions shown with directional arrows; (c) Video with Physical Dynamics: Temporal sequences showing normal and abnormal states, highlighting anomalies like leaks, misalignments, and cracks. By focusing on the dynamic behaviors of complex objects, we enhance understanding of interactions and failure modes in real-world settings, where both structure and motion contribute to anomaly detection." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 331, + 198, + 555, + 430 + ], + "blocks": [ + { + "bbox": [ + 331, + 198, + 555, + 430 + ], + "lines": [ + { + "bbox": [ + 331, + 198, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 331, + 198, + 555, + 430 + ], + "type": "image", + "image_path": "406ce781989bb97be996195c92fcaa5cde7a373227f124df11e22d77c6ab559c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 507, + 200, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 507, + 200, + 519 + ], + "spans": [ + { + "bbox": [ + 152, + 507, + 200, + 519 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 534, + 297, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 534, + 297, + 678 + ], + "spans": [ + { + "bbox": [ + 54, + 534, + 297, + 678 + ], + "type": "text", + "content": "Humans detect real-world object anomalies by perceiving, interacting, and reasoning based on object-conditioned physical knowledge. The long-term goal of Industrial Anomaly Detection (IAD) is to enable machines to autonomously replicate this skill. However, current IAD algorithms are largely developed and tested on static, semantically simple datasets, which diverge from real-world scenarios where physical understanding and reasoning are essential. To bridge this gap, we introduce the Physics Anomaly Detection (Phys-AD) dataset, the first large-scale, real-world, physics-grounded video dataset for industrial anomaly detection. Collected using a real robot arm and motor, Phys-AD provides a diverse" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 508, + 557, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 557, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 557, + 689 + ], + "type": "text", + "content": "set of dynamic, semantically rich scenarios. The dataset includes more than 6400 videos across 22 real-world object categories, interacting with robot arms and motors, and exhibits 47 types of anomalies. Anomaly detection in Phys-AD requires visual reasoning, combining both physical knowledge and video content to determine object abnormality. We benchmark state-of-the-art anomaly detection methods under three settings: unsupervised AD, weakly-supervised AD, and video-understanding AD, highlighting their limitations in handling physics-grounded anomalies. Additionally, we introduce the Physics Anomaly Explanation (PAEval) metric, designed to assess the ability of visual-language foundation models to not only detect anomalies but also provide accurate explanations for their underlying physical causes. Our project is available at https://guyao2023.github.io/Phys-AD/." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 174, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 174, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 174, + 703 + ], + "type": "text", + "content": "*Equally contribute to this work." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "text", + "content": "† Corresponding authors" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30409" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 72, + 293, + 247 + ], + "blocks": [ + { + "bbox": [ + 62, + 72, + 293, + 247 + ], + "lines": [ + { + "bbox": [ + 62, + 72, + 293, + 247 + ], + "spans": [ + { + "bbox": [ + 62, + 72, + 293, + 247 + ], + "type": "image", + "image_path": "fc4dc8c941e0683bab05b70a96d80230028a00681e4d6d0074810515380bb537.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 251, + 298, + 350 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 298, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 298, + 350 + ], + "type": "text", + "content": "Figure 2. Human-like decision-making process for physics-grounded object anomaly detection. We illustrate the sequential approach of a human-like agent for evaluating an object's normality. First, the agent perceives relevant physical attributes (e.g., plastic and elastic), then interacts by performing a physical action (e.g., squeezing), and finally reasons based on the vision feedback and attributes changes (e.g., surface shape change) to determine whether the object is normal or anomalous. This mirrors a human's natural process of reasoning over physics in objects." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 370, + 136, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 370, + 136, + 383 + ], + "spans": [ + { + "bbox": [ + 56, + 370, + 136, + 383 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 390, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 296, + 544 + ], + "type": "text", + "content": "Industrial anomaly detection (IAD) is a critical subfield in computer vision and industrial automation, aiming to identify defects or irregularities in products during manufacturing. As shown in Fig. 2, the ultimate vision is to create autonomous systems that not only perceive but also interact with and reason about objects to discriminate between anomalies and normal states, integrating complex physical principles to detect anomalies in dynamic, real-world scenarios. For example, a human inspecting a water bottle for anomalies wouldn't rely solely on visual observation; they might rotate or invert the bottle, using physical interactions and cues, such as noticing a loose cap or an irregular internal flow, to detect issues." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 545, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 677 + ], + "type": "text", + "content": "Central to advancing IAD [9] is the availability of high-quality datasets that bridge the gap between academic research and industrial needs. Datasets like MVTec-AD [4], MPDD [21], and VisA [65] have played foundational roles, enabling algorithm development for image-based anomaly detection and bringing IAD to the forefront of computer vision research. While these datasets have significantly advanced single-image anomaly detection, recent datasets, e.g., MVTec-3D [7], Real3D [29], and Anomaly-ShapeNet [27], have extended IAD to 3D, aligning research more closely with the needs of complex real-world industrial settings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Yet, as factories increasingly rely on robotic arms and automated systems to perform sophisticated inspections, the limitations of current IAD datasets become apparent." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "Existing benchmarks focus on static, semantically simple environments, overlooking the physical priors and interactive reasoning required in real-world industrial contexts. This gap highlights a growing need for datasets that not only reflect real-world physical constraints but also challenge models to reason dynamically about anomalies." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 146, + 555, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 146, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 146, + 555, + 409 + ], + "type": "text", + "content": "To bridge this gap, as shown in Fig. 1, we introduce the Physics Anomaly Detection (Phys-AD) dataset, the first large-scale, physics-grounded video dataset for industrial anomaly detection. Phys-AD features over 6,400 videos of 22 categories and 49 objects interacting with robotic arms and motors, capturing 47 types of anomalies that require visual reasoning informed by physical knowledge. The short video clips in the dataset range from 60 to 240 frames in length and are filmed in real-time industrial environment, fully capturing the interaction process between robotic arm or motor and industrial objects. Additionally, to ensure our dataset meets industrial demands and matches the complexity and diversity of the real physical world, we selected industrial objects of different physical qualities, various interaction methods, and anomalies that reflect diverse physical principles and require different reasoning process. Specifically, we selected 22 object categories spanning across metals, plastics, fluid, amorphous substances and articulated objects with diverse appearances. For interaction, we use mechanical grippers, robotic arms, and motors, incorporating various interaction modes such as pressing, rotating, squeezing, and driving to handle different types of objects." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 411, + 556, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 556, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 556, + 590 + ], + "type": "text", + "content": "We benchmark state-of-the-art anomaly detection methods in three key configurations: unsupervised anomaly detection, weakly-supervised anomaly detection, and video-understanding based anomaly detection. Our findings reveal critical gaps in their ability to handle the complexities of physics-grounded scenarios, where anomalies often arise from dynamic, interdependent interactions. To advance the field, we also introduce the Physical Anomaly Explanation (PAEval) metric, designed to assess both detection performance and a model's ability to explain anomalies by identifying underlying physical causes. Furthermore, our benchmark reveals the fragility of existing methods in tackling these challenging conditions, underscoring the need for approaches that better understand object dynamics and temporal coherence in anomaly detection." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 325, + 592, + 552, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 592, + 552, + 604 + ], + "spans": [ + { + "bbox": [ + 325, + 592, + 552, + 604 + ], + "type": "text", + "content": "Our contributions can be summarized as followings:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 314, + 605, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 605, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 314, + 605, + 555, + 641 + ], + "type": "text", + "content": "- We introduce a novel task of detecting physical-based industrial anomalies in real-world that involves perception, physical and visual reasoning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 642, + 555, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 701 + ], + "type": "text", + "content": "- We present Phys-AD, the first large-scale, physics-grounded video dataset specifically designed for industrial anomaly detection in real world, containing objects of different physical qualities, multiple interaction methods and various physical reasoning process." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "content": "- We benchmarking the anomaly detection and reasoning" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30410" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 128, + 296, + 311 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "type": "text", + "content": "Table 1. Comparison of Phys-AD with existing industrial anomaly detection datasets. Our Phys-AD dataset is the first to consider complex objects with physical dynamics. 'Syn', 'IR', 'D', and 'PC' denote Synthetic, Infrared, Depth, and Point Cloud, respectively. #Anomaly indicates the number of anomaly types." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 128, + 296, + 311 + ], + "lines": [ + { + "bbox": [ + 56, + 128, + 296, + 311 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 296, + 311 + ], + "type": "table", + "html": "
DatasetYearTypeModalitySample Statistics
Class #AnomalyPhysics
MVTec-AD [3]2019RealRGB15-X
BTAD [38]2021RealRGB33X
MPDD [21]2021RealRGB68X
VisA [66]2021RealRGB12-X
MVTec LOCO-AD [5]2022RealRGB5-X
MAD [64]2023Syn+RealRGB203X
LOCO-Annotations [62]2024RealRGB55X
Real-IAD [46]2024RealRGB308X
GDXray [37]2015RealX-ray515X
PVEL-AD [43]2023RealIR110X
MVTec3D-AD [6]2021RealRGB-D103-5X
Eyecandies [8]2022SynRGB-D103X
Real3D-AD [30]2023RealPC122X
Anomaly-ShapeNet [27]2024SynPC406X
Phys-AD (Ours)2024RealRGB4947✓
", + "image_path": "2affb83bc3d5314d7dc6b2b25346b44074ea92e0f46ef8701d79e44a343e635b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 63, + 332, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 332, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 63, + 332, + 296, + 392 + ], + "type": "text", + "content": "performance of popular video AD methods and Visual Language Foundation Models on the Phys-AD dataset in several settings, establishing a practical and challenging benchmark to promote the development of the physics-related anomaly detection field." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 405, + 139, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 139, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 139, + 418 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "type": "text", + "content": "Industrial anomaly detection datasets. Existing datasets primarily focus on static, semantically simple scenarios, deviate significantly from real world where physical understanding and reasoning are essential. Datasets like MVTec-AD [3], BTAD [38], MPDD [21], and VisA [66] focus on surface level image anomaly detection with one single-view RGB image, limiting their effectiveness in capturing holistic object structures. MVTec-LOCO-AD [5], focusing on structure and local information in industries, is limited by its relatively simple and constrained data content. While MVTec3D-AD [6] and Eyecandies [8] incorporate depth data, they remain static and single-view image anomaly detection, neglecting object level information. To explore object level anomaly detection, multi-view IAD datasets like MAD [64] and Real-IAD [46], point cloud IAD datasets like Real3D-AD [30] and Anomaly-ShapeNet [27], offering richer visual and geometric cues, but are still limited to static objects without dynamic interaction or reasoning. In summary, current industrial anomaly detection datasets focus on relatively simple and static anomaly detection scenarios, lacking complex physical rules, dynamic interactions, and visual reasoning requirements. Therefore, existing IAD datasets generally applied to limited industrial scenarios and there is no IAD dataset could meet the demands of detecting" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "complex anomalies in real world which need physical priors and reasoning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 99, + 556, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 99, + 556, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 99, + 556, + 363 + ], + "type": "text", + "content": "Video anomaly detection. Deep learning methods [18, 35, 54, 61] now dominate Video Anomaly Detection (VAD), categorized into unsupervised, weakly-supervised, and fully-supervised approaches. Unsupervised methods learn normal patterns via reconstruction [16, 18, 52], prediction [32], or hybrids [33], while some methods [45, 55] train with both unlabeled normal and abnormal data. Weakly-supervised methods[14, 49, 60, 63] use video-level or glance-based annotations, and fully-supervised methods[25, 31] remain rare due to costly frame-level labeling. Visual-language models like CLIP [42] have recently been applied to enhance anomaly detection [24, 41, 51], focusing on semantic anomalies. Open-vocabulary VAD [50] and prompt-based anomaly scoring [56] leverage LLMs [53, 61], but performance relies heavily on the base models, often lacking domain-specific tuning. However, existing video anomaly detection algorithms lack the capability to handle complex industrial anomaly detection scenarios and understand physical rules. This gap highlights the need for models that can capture dynamic behaviors and physical laws in industrial environments, as addressed by Phys-AD, which targets industrial anomalies in objects of various physical properties." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 365, + 557, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 557, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 557, + 547 + ], + "type": "text", + "content": "Visual reasoning. Visual reasoning is a critical task in computer vision, aiming to enable machines to interpret perceptual information like humans. Several visual reasoning tasks have been proposed to evaluate reasoning capabilities, including Visual Question Answering (VQA), 2D puzzles, and physical dynamics prediction. In VQA, agents are required to combine natural language and visual cues to answer questions [1, 17, 19, 23]. For 2D puzzles, tasks involve discovering relationships among visual elements and making inferences [22, 26, 39, 57, 58]. Physical dynamics prediction tasks require machines to perceive and reason about physical interactions [2, 12, 13, 20]. In contrast to these work, we introduce the first benchmark featuring real-world industrial objects with dynamic physical properties, focusing on distinguishing diverse dynamics through vision." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 556, + 422, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 422, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 422, + 569 + ], + "type": "text", + "content": "3. Dataset: Phys-AD" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 576, + 549, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 549, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 549, + 589 + ], + "type": "text", + "content": "3.1. Object Preparation and Interaction Selection" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "We selected 22 categories spanning across various materials, including metals, plastics, amorphous substances and articulated objects, with diverse shapes, sizes, and physical properties. For the physical properties of different objects, we correspondingly select different interaction methods like push, rotate, pull with robotics arms and motors. For instance, we use robotics arms to grab and extruding the deformable objects like basketball to determine whether there are any elastic anomalies or surface defects by the morphological change of the objects. To make our dataset more practical" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30411" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 138, + 115 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 138, + 115 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 138, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 138, + 115 + ], + "type": "image", + "image_path": "d1fb0c57315c70053b2eee90ea39523b6bc62eca99f6e996b4571a08712f0d89.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 117, + 110, + 125 + ], + "lines": [ + { + "bbox": [ + 83, + 117, + 110, + 125 + ], + "spans": [ + { + "bbox": [ + 83, + 117, + 110, + 125 + ], + "type": "text", + "content": "U Disk" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 139, + 70, + 220, + 115 + ], + "blocks": [ + { + "bbox": [ + 139, + 70, + 220, + 115 + ], + "lines": [ + { + "bbox": [ + 139, + 70, + 220, + 115 + ], + "spans": [ + { + "bbox": [ + 139, + 70, + 220, + 115 + ], + "type": "image", + "image_path": "04c3a0c007d031f41abd5c7c98604aad9ab1aaf4129e18d28427e92deee1eb37.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 116, + 194, + 126 + ], + "lines": [ + { + "bbox": [ + 168, + 116, + 194, + 126 + ], + "spans": [ + { + "bbox": [ + 168, + 116, + 194, + 126 + ], + "type": "text", + "content": "Zipper" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 223, + 71, + 304, + 116 + ], + "blocks": [ + { + "bbox": [ + 223, + 71, + 304, + 116 + ], + "lines": [ + { + "bbox": [ + 223, + 71, + 304, + 116 + ], + "spans": [ + { + "bbox": [ + 223, + 71, + 304, + 116 + ], + "type": "image", + "image_path": "45bcc2f974a31f7bc5fa856043b21b280954cb11438e7f549aef3542277029ad.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 117, + 295, + 126 + ], + "lines": [ + { + "bbox": [ + 239, + 117, + 295, + 126 + ], + "spans": [ + { + "bbox": [ + 239, + 117, + 295, + 126 + ], + "type": "text", + "content": "Rolling Bearing" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 387, + 117 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 387, + 117 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 387, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 387, + 117 + ], + "type": "image", + "image_path": "b7d51d7075f154f89366e928da49136a71a85cee7477c389e2eb7f9afe4c02c5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 118, + 370, + 126 + ], + "lines": [ + { + "bbox": [ + 323, + 118, + 370, + 126 + ], + "spans": [ + { + "bbox": [ + 323, + 118, + 370, + 126 + ], + "type": "text", + "content": "Sticky Roller" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 390, + 72, + 470, + 116 + ], + "blocks": [ + { + "bbox": [ + 390, + 72, + 470, + 116 + ], + "lines": [ + { + "bbox": [ + 390, + 72, + 470, + 116 + ], + "spans": [ + { + "bbox": [ + 390, + 72, + 470, + 116 + ], + "type": "image", + "image_path": "138a01ee9acd66f4fd6acc7aa80c8d225c1ae73de2587940ad0378ef6c70897b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 117, + 440, + 125 + ], + "lines": [ + { + "bbox": [ + 417, + 117, + 440, + 125 + ], + "spans": [ + { + "bbox": [ + 417, + 117, + 440, + 125 + ], + "type": "text", + "content": "Slider" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 473, + 72, + 553, + 116 + ], + "blocks": [ + { + "bbox": [ + 473, + 72, + 553, + 116 + ], + "lines": [ + { + "bbox": [ + 473, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 473, + 72, + 553, + 116 + ], + "type": "image", + "image_path": "6b60ccaaf439eb96ea39c5a3b9fa413b510be46701455468c04e51538753bdc3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 494, + 116, + 533, + 125 + ], + "lines": [ + { + "bbox": [ + 494, + 116, + 533, + 125 + ], + "spans": [ + { + "bbox": [ + 494, + 116, + 533, + 125 + ], + "type": "text", + "content": "Toothpaste" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 57, + 130, + 138, + 176 + ], + "blocks": [ + { + "bbox": [ + 57, + 130, + 138, + 176 + ], + "lines": [ + { + "bbox": [ + 57, + 130, + 138, + 176 + ], + "spans": [ + { + "bbox": [ + 57, + 130, + 138, + 176 + ], + "type": "image", + "image_path": "0b3e191f36b3ed06901894a1493699ae31532b6433e29d9994adcf9e77502b7d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 178, + 117, + 186 + ], + "lines": [ + { + "bbox": [ + 78, + 178, + 117, + 186 + ], + "spans": [ + { + "bbox": [ + 78, + 178, + 117, + 186 + ], + "type": "text", + "content": "Door Lock" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 140, + 130, + 220, + 176 + ], + "blocks": [ + { + "bbox": [ + 140, + 130, + 220, + 176 + ], + "lines": [ + { + "bbox": [ + 140, + 130, + 220, + 176 + ], + "spans": [ + { + "bbox": [ + 140, + 130, + 220, + 176 + ], + "type": "image", + "image_path": "de166171b1fd2c1d1d14cfe9f71d84aeee36e3408be644fb0740baed93afadf8.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 166, + 177, + 195, + 187 + ], + "lines": [ + { + "bbox": [ + 166, + 177, + 195, + 187 + ], + "spans": [ + { + "bbox": [ + 166, + 177, + 195, + 187 + ], + "type": "text", + "content": "Toy Car" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 223, + 130, + 304, + 176 + ], + "blocks": [ + { + "bbox": [ + 223, + 130, + 304, + 176 + ], + "lines": [ + { + "bbox": [ + 223, + 130, + 304, + 176 + ], + "spans": [ + { + "bbox": [ + 223, + 130, + 304, + 176 + ], + "type": "image", + "image_path": "90f777c39de74347453e303735bb358ab7dd2c4abc98c699069be3d12fd72216.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 177, + 271, + 187 + ], + "lines": [ + { + "bbox": [ + 255, + 177, + 271, + 187 + ], + "spans": [ + { + "bbox": [ + 255, + 177, + 271, + 187 + ], + "type": "text", + "content": "Clip" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 307, + 131, + 387, + 176 + ], + "blocks": [ + { + "bbox": [ + 307, + 131, + 387, + 176 + ], + "lines": [ + { + "bbox": [ + 307, + 131, + 387, + 176 + ], + "spans": [ + { + "bbox": [ + 307, + 131, + 387, + 176 + ], + "type": "image", + "image_path": "1324e3603fe14c0ee8596a894f5b4b71efea4e5df255eef133e2c529b00767b4.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 178, + 356, + 186 + ], + "lines": [ + { + "bbox": [ + 342, + 178, + 356, + 186 + ], + "spans": [ + { + "bbox": [ + 342, + 178, + 356, + 186 + ], + "type": "text", + "content": "Fan" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 390, + 131, + 470, + 176 + ], + "blocks": [ + { + "bbox": [ + 390, + 131, + 470, + 176 + ], + "lines": [ + { + "bbox": [ + 390, + 131, + 470, + 176 + ], + "spans": [ + { + "bbox": [ + 390, + 131, + 470, + 176 + ], + "type": "image", + "image_path": "79a5bcee9d78b800f3c96612ba744acd04d2d65ab5444af43ed720b70edc3bcb.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 178, + 453, + 187 + ], + "lines": [ + { + "bbox": [ + 406, + 178, + 453, + 187 + ], + "spans": [ + { + "bbox": [ + 406, + 178, + 453, + 187 + ], + "type": "text", + "content": "Rubber Band" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 473, + 131, + 553, + 177 + ], + "blocks": [ + { + "bbox": [ + 473, + 131, + 553, + 177 + ], + "lines": [ + { + "bbox": [ + 473, + 131, + 553, + 177 + ], + "spans": [ + { + "bbox": [ + 473, + 131, + 553, + 177 + ], + "type": "image", + "image_path": "a93b776c01182baf642c844fd4003427bef1d20f803d1921c7b718ef3b4de6e3.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 501, + 178, + 529, + 188 + ], + "lines": [ + { + "bbox": [ + 501, + 178, + 529, + 188 + ], + "spans": [ + { + "bbox": [ + 501, + 178, + 529, + 188 + ], + "type": "text", + "content": "Magnet" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 56, + 194, + 137, + 239 + ], + "blocks": [ + { + "bbox": [ + 56, + 194, + 137, + 239 + ], + "lines": [ + { + "bbox": [ + 56, + 194, + 137, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 137, + 239 + ], + "type": "image", + "image_path": "ef6d8ff60375a009ba23164350c9f70fa0b56f37d63ffd49a93da1a6e504579a.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 81, + 242, + 103, + 250 + ], + "lines": [ + { + "bbox": [ + 81, + 242, + 103, + 250 + ], + "spans": [ + { + "bbox": [ + 81, + 242, + 103, + 250 + ], + "type": "text", + "content": "Servo" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 140, + 194, + 220, + 239 + ], + "blocks": [ + { + "bbox": [ + 140, + 194, + 220, + 239 + ], + "lines": [ + { + "bbox": [ + 140, + 194, + 220, + 239 + ], + "spans": [ + { + "bbox": [ + 140, + 194, + 220, + 239 + ], + "type": "image", + "image_path": "4d6dc56067eb39e38a2c6d664695e2645359b85f86b75c0ba77d4771ba20a87f.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 241, + 208, + 251 + ], + "lines": [ + { + "bbox": [ + 154, + 241, + 208, + 251 + ], + "spans": [ + { + "bbox": [ + 154, + 241, + 208, + 251 + ], + "type": "text", + "content": "Gear Assembly" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 223, + 194, + 304, + 239 + ], + "blocks": [ + { + "bbox": [ + 223, + 194, + 304, + 239 + ], + "lines": [ + { + "bbox": [ + 223, + 194, + 304, + 239 + ], + "spans": [ + { + "bbox": [ + 223, + 194, + 304, + 239 + ], + "type": "image", + "image_path": "143b400b63356474d7a583349339b8c3fd2ca6429bf0a91836a0f9968ad6acbb.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 241, + 287, + 250 + ], + "lines": [ + { + "bbox": [ + 238, + 241, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 238, + 241, + 287, + 250 + ], + "type": "text", + "content": "Bottled Water" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 307, + 194, + 387, + 239 + ], + "blocks": [ + { + "bbox": [ + 307, + 194, + 387, + 239 + ], + "lines": [ + { + "bbox": [ + 307, + 194, + 387, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 387, + 239 + ], + "type": "image", + "image_path": "db4e4ea50cb523a34663c3de3d8555feba8fc3e602d608de9919f4a00022692f.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 242, + 356, + 250 + ], + "lines": [ + { + "bbox": [ + 333, + 242, + 356, + 250 + ], + "spans": [ + { + "bbox": [ + 333, + 242, + 356, + 250 + ], + "type": "text", + "content": "Screw" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 390, + 194, + 470, + 239 + ], + "blocks": [ + { + "bbox": [ + 390, + 194, + 470, + 239 + ], + "lines": [ + { + "bbox": [ + 390, + 194, + 470, + 239 + ], + "spans": [ + { + "bbox": [ + 390, + 194, + 470, + 239 + ], + "type": "image", + "image_path": "5d07bbf049c4ab67cb7cefb294f93b1adecd62b7b746b16a2b25c80a1f2688ff.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 241, + 463, + 251 + ], + "lines": [ + { + "bbox": [ + 400, + 241, + 463, + 251 + ], + "spans": [ + { + "bbox": [ + 400, + 241, + 463, + 251 + ], + "type": "text", + "content": "Spherical Bearing" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 485, + 201, + 544, + 219 + ], + "blocks": [ + { + "bbox": [ + 485, + 201, + 544, + 219 + ], + "lines": [ + { + "bbox": [ + 485, + 201, + 544, + 219 + ], + "spans": [ + { + "bbox": [ + 485, + 201, + 544, + 219 + ], + "type": "image", + "image_path": "ae3b8ab1d44dca0c371de14f0f6e04fbffda0755c5a46cb2291463b29b4ca39e.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 498, + 229, + 526, + 238 + ], + "lines": [ + { + "bbox": [ + 498, + 229, + 526, + 238 + ], + "spans": [ + { + "bbox": [ + 498, + 229, + 526, + 238 + ], + "type": "text", + "content": "Motion" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 57, + 258, + 137, + 308 + ], + "blocks": [ + { + "bbox": [ + 57, + 258, + 137, + 308 + ], + "lines": [ + { + "bbox": [ + 57, + 258, + 137, + 308 + ], + "spans": [ + { + "bbox": [ + 57, + 258, + 137, + 308 + ], + "type": "image", + "image_path": "30b520f00ef352421bb17433f5d67fdc49435e0a6a68430c5acb0b5e28c6e3da.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 310, + 104, + 318 + ], + "lines": [ + { + "bbox": [ + 80, + 310, + 104, + 318 + ], + "spans": [ + { + "bbox": [ + 80, + 310, + 104, + 318 + ], + "type": "text", + "content": "Button" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 140, + 258, + 220, + 308 + ], + "blocks": [ + { + "bbox": [ + 140, + 258, + 220, + 308 + ], + "lines": [ + { + "bbox": [ + 140, + 258, + 220, + 308 + ], + "spans": [ + { + "bbox": [ + 140, + 258, + 220, + 308 + ], + "type": "image", + "image_path": "d477cbe0731ea4b43372c2b74c2807e6242f0ecfdb1ee561599aa52988b7e88b.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 310, + 189, + 318 + ], + "lines": [ + { + "bbox": [ + 168, + 310, + 189, + 318 + ], + "spans": [ + { + "bbox": [ + 168, + 310, + 189, + 318 + ], + "type": "text", + "content": "Clock" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 223, + 258, + 302, + 308 + ], + "blocks": [ + { + "bbox": [ + 223, + 258, + 302, + 308 + ], + "lines": [ + { + "bbox": [ + 223, + 258, + 302, + 308 + ], + "spans": [ + { + "bbox": [ + 223, + 258, + 302, + 308 + ], + "type": "image", + "image_path": "32405e8d76a58bd525059e3dbd12ba4e95ee2af92a48352e8327c8bf0cb5a0a5.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 310, + 285, + 319 + ], + "lines": [ + { + "bbox": [ + 239, + 310, + 285, + 319 + ], + "spans": [ + { + "bbox": [ + 239, + 310, + 285, + 319 + ], + "type": "text", + "content": "Caster Wheel" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 307, + 258, + 387, + 308 + ], + "blocks": [ + { + "bbox": [ + 307, + 258, + 387, + 308 + ], + "lines": [ + { + "bbox": [ + 307, + 258, + 387, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 258, + 387, + 308 + ], + "type": "image", + "image_path": "d020a461c07f5b611e5475dca37bd6451a9ee9fd00ee94a5502173746bfcce9d.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 310, + 358, + 319 + ], + "lines": [ + { + "bbox": [ + 335, + 310, + 358, + 319 + ], + "spans": [ + { + "bbox": [ + 335, + 310, + 358, + 319 + ], + "type": "text", + "content": "Hinge" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 390, + 258, + 471, + 308 + ], + "blocks": [ + { + "bbox": [ + 390, + 258, + 471, + 308 + ], + "lines": [ + { + "bbox": [ + 390, + 258, + 471, + 308 + ], + "spans": [ + { + "bbox": [ + 390, + 258, + 471, + 308 + ], + "type": "image", + "image_path": "5d39dd0d1961330a708b001bbd0af0330e7e2274175ee736aa65e240d83b341e.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 424, + 310, + 439, + 319 + ], + "lines": [ + { + "bbox": [ + 424, + 310, + 439, + 319 + ], + "spans": [ + { + "bbox": [ + 424, + 310, + 439, + 319 + ], + "type": "text", + "content": "Ball" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 473, + 258, + 553, + 316 + ], + "blocks": [ + { + "bbox": [ + 487, + 240, + 539, + 258 + ], + "lines": [ + { + "bbox": [ + 487, + 240, + 539, + 258 + ], + "spans": [ + { + "bbox": [ + 487, + 240, + 539, + 258 + ], + "type": "text", + "content": "Direction For Interaction" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 473, + 258, + 553, + 316 + ], + "lines": [ + { + "bbox": [ + 473, + 258, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 473, + 258, + 553, + 316 + ], + "type": "image", + "image_path": "ee9640bd724f2776c510eb025a226563c9518fd3ee2fc98b17e081ca1a1b534d.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 323, + 555, + 368 + ], + "lines": [ + { + "bbox": [ + 55, + 323, + 555, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 555, + 368 + ], + "type": "text", + "content": "Figure 3. Interactions for understanding implicit physical laws in the Phys-AD dataset. We showcase various object interactions from the Phys-AD dataset, where different actions (indicated by motion directions) are used to explore and reason about the underlying physical properties and behaviors of each object. The colored arrows indicate the interaction directions and axes, highlighting how physical interactions reveal the implicit physics governing each object." + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 58, + 389, + 302, + 442 + ], + "blocks": [ + { + "bbox": [ + 103, + 377, + 251, + 387 + ], + "lines": [ + { + "bbox": [ + 103, + 377, + 251, + 387 + ], + "spans": [ + { + "bbox": [ + 103, + 377, + 251, + 387 + ], + "type": "text", + "content": "Normal: The U Disk rotates freely over time." + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 389, + 302, + 442 + ], + "lines": [ + { + "bbox": [ + 58, + 389, + 302, + 442 + ], + "spans": [ + { + "bbox": [ + 58, + 389, + 302, + 442 + ], + "type": "image", + "image_path": "b3f99f89cbed8591fccb3110e6a970871e8521d23d9bbcbaaea6ed729034642a.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 58, + 473, + 302, + 525 + ], + "blocks": [ + { + "bbox": [ + 108, + 460, + 247, + 470 + ], + "lines": [ + { + "bbox": [ + 108, + 460, + 247, + 470 + ], + "spans": [ + { + "bbox": [ + 108, + 460, + 247, + 470 + ], + "type": "text", + "content": "Abnormal: The U Disk is stuck over time." + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 473, + 302, + 525 + ], + "lines": [ + { + "bbox": [ + 58, + 473, + 302, + 525 + ], + "spans": [ + { + "bbox": [ + 58, + 473, + 302, + 525 + ], + "type": "image", + "image_path": "6768b2f107bdf1fc48650b414447219c6c19cc384203b8c07f0ee7c7436e2cb8.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 529, + 219, + 538 + ], + "lines": [ + { + "bbox": [ + 137, + 529, + 219, + 538 + ], + "spans": [ + { + "bbox": [ + 137, + 529, + 219, + 538 + ], + "type": "text", + "content": "(a) Persistent Anomalies" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 309, + 389, + 552, + 442 + ], + "blocks": [ + { + "bbox": [ + 331, + 377, + 523, + 387 + ], + "lines": [ + { + "bbox": [ + 331, + 377, + 523, + 387 + ], + "spans": [ + { + "bbox": [ + 331, + 377, + 523, + 387 + ], + "type": "text", + "content": "Normal: The sticky roller rotates smoothly on the surface." + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 389, + 552, + 442 + ], + "lines": [ + { + "bbox": [ + 309, + 389, + 552, + 442 + ], + "spans": [ + { + "bbox": [ + 309, + 389, + 552, + 442 + ], + "type": "image", + "image_path": "fc43f6a365f2a63ec36586f1a0cf725753bce008123e8aed6dd463dcf945a622.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 309, + 475, + 552, + 526 + ], + "blocks": [ + { + "bbox": [ + 321, + 453, + 440, + 473 + ], + "lines": [ + { + "bbox": [ + 321, + 453, + 440, + 473 + ], + "spans": [ + { + "bbox": [ + 321, + 453, + 440, + 473 + ], + "type": "text", + "content": "Normal: The sticky roller functions properly at the beginning." + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 458, + 453, + 552, + 472 + ], + "lines": [ + { + "bbox": [ + 458, + 453, + 552, + 472 + ], + "spans": [ + { + "bbox": [ + 458, + 453, + 552, + 472 + ], + "type": "text", + "content": "Abnormal: The sticky roller breaks at a certain point." + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 475, + 552, + 526 + ], + "lines": [ + { + "bbox": [ + 309, + 475, + 552, + 526 + ], + "spans": [ + { + "bbox": [ + 309, + 475, + 552, + 526 + ], + "type": "image", + "image_path": "0fd51d148225da32c20aed5e679cb20d87bc532ca66ea2e17693bcb16094c8c2.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 530, + 481, + 539 + ], + "lines": [ + { + "bbox": [ + 391, + 530, + 481, + 539 + ], + "spans": [ + { + "bbox": [ + 391, + 530, + 481, + 539 + ], + "type": "text", + "content": "(b) Intermittent Anomalies" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 544, + 555, + 599 + ], + "lines": [ + { + "bbox": [ + 55, + 544, + 555, + 599 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 555, + 599 + ], + "type": "text", + "content": "Figure 4. Categorization of anomalies based on persistence in the Phys-AD dataset. We show examples of normal and abnormal functioning in common objects, divided into two anomaly types: persistent and intermittent. (a) Persistent anomalies, such as continuous obstruction in the U Disk or permanent malfunction of the Sticky Roller, are visible throughout the operation. (b) In contrast, intermittent anomalies, like occasional jamming of the U Disk or breakage in the Sticky Roller after initial operation, only appear at specific points in time. This classification provides insight into both constant and sporadic failures in object interactions." + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_caption" + } + ], + "index": 58 + }, + { + "bbox": [ + 55, + 620, + 297, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 297, + 704 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 297, + 704 + ], + "type": "text", + "content": "and challenging, we introduce 47 distinct defect types, some of which just rely on single frame content and physical rules for anomaly reasoning, while the other part needs to combine the content of the whole video and physical priors together to judge whether the object is abnormal or not. Different objects information with their corresponding interaction methods and anomalies are listed in Table 2. Fig 3 shows the interac" + } + ] + } + ], + "index": 61 + }, + { + "bbox": [ + 313, + 620, + 555, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 620, + 555, + 692 + ], + "spans": [ + { + "bbox": [ + 313, + 620, + 555, + 692 + ], + "type": "text", + "content": "tions for understanding implicit physical laws in the Phys-AD dataset. Fig 4 provides two representative examples: one where a U Disk requires analyzing the entire video content to determine if there is an anomaly, and another where a sticky roller only requires a few frames from the video to judge whether the object is abnormal." + } + ] + } + ], + "index": 62 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30412" + } + ] + } + ], + "index": 63 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 176, + 165 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 176, + 165 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 176, + 165 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 176, + 165 + ], + "type": "image", + "image_path": "c3ffd1eb93313cb723ec32e3219be946461890aafac571d6445aa6c70c56801c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 168, + 295, + 201 + ], + "lines": [ + { + "bbox": [ + 55, + 168, + 295, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 295, + 201 + ], + "type": "text", + "content": "Figure 5. Data collection pipeline for the Phys-AD dataset. (a) Manipulation of a toothpaste tube using a UR5 robotic arm. (b) Manipulation of a U Disk and fan via servo and motor." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 177, + 72, + 294, + 164 + ], + "blocks": [ + { + "bbox": [ + 177, + 72, + 294, + 164 + ], + "lines": [ + { + "bbox": [ + 177, + 72, + 294, + 164 + ], + "spans": [ + { + "bbox": [ + 177, + 72, + 294, + 164 + ], + "type": "image", + "image_path": "e4add3a5da0de42ca6b80027e5c1fbdd5fd5651b741bdb4da4c922c84fbf789b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 253, + 294, + 531 + ], + "blocks": [ + { + "bbox": [ + 55, + 217, + 295, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 217, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 295, + 251 + ], + "type": "text", + "content": "Table 2. Interaction methods and anomalies of the object categories in our Phys-AD dataset. Motor. means one object driven by motor. Multi. means one object with multiple anomalies." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 253, + 294, + 531 + ], + "lines": [ + { + "bbox": [ + 58, + 253, + 294, + 531 + ], + "spans": [ + { + "bbox": [ + 58, + 253, + 294, + 531 + ], + "type": "table", + "html": "
CategoryInteractionAnomaly Types
CarDrag, SlideDifferent wheels stuck
FanMotor., RotateStuck, Uneven Rotation, Vibration
Rolling BearingMotor., RotateLack of friction
Spherical BearingGrab, RotateInternal block
ServoGrab, Motor., RotateAngle restricted, Vibration, No calibration
ClipPressUnable to press down, Unable to rebound
U DiskGrab, Motor., RotateCover jam
HingeGrab, RotateAngle restricted, Shaft off
Sticky RollerGrab, PullDetach, Unable to rotate
Caster WheelSlideAxle axis stuck, Swivel axis stuck
ScrewPress, RotateLoosen, Unable to insert
LockMotor., RotateLatch jam, Loose
GearMotor., RotateStuck, Not meshed, Multi.
ClockMotor., RotatePointer stops, Vibration
SlideGrab, SlideDetach, Shedding, Jam
ZipperGrab, CloseStuck, Unable to close
ButtonPressNo light, Unable to press down, Unable to rebound, Multi.
LiquidGrab, ShakeWater out, Foreign body
Rubber BandStretchCrack
BallPinch, PressInsufficient gas, Leakage
MagnetGrab, Press, MoveDegaussing, Shell detached
ToothpastePinch, PressLeakage
", + "image_path": "cbcf4e193c05768860b9b372ece0ae1bbe405dad03bd5e2829372390fe2dbb53.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 550, + 224, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 224, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 224, + 563 + ], + "type": "text", + "content": "3.2. Data Collection and Processing" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "Data collection pipeline. Most of the data collection for the Phys-AD dataset is driven by manipulation-guided video sequences, captured using a UR5 robotic arm equipped with an RGB camera (see Fig. 5a). In order to reproduce real industrial scenes, we also include adjustable light sources in our data capture process. The RGB Camera, with a 1080p resolution and a frame rate of 60 FPS, providing high-quality video sequences. Some kinds of the objects like U Disk that are not suitable for the manipulation of robotics arms, are driven by the motor or servo (Fig. 5b). After the data collection, we used video editing software to crop out irrelevant frames and retain the complete interaction process." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 316, + 106, + 553, + 283 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 104 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 104 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 104 + ], + "type": "text", + "content": "Table 3. Phys-AD dataset statistics. We denote the total number of frames and videos for each category as #Images and #Videos. Note that the Train split does not contain anomaly data." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 106, + 553, + 283 + ], + "lines": [ + { + "bbox": [ + 316, + 106, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 553, + 283 + ], + "type": "table", + "html": "
CategoryTrain #FramesTrain #VideosTest #FramesTest #VideosTest #Videos Nor.Test #Videos Ab.Obj. TypesDef. Types
Car18,00030036,00060015045053
Fan32,40018064,8003609027033
Rolling Bearing3,600603,60060303011
Spherical Bearing3,600603,60060303011
Servo14,40012028,8002406018013
Clip28,80024043,20036012024042
U Disk14,40024014,40024012012041
Hinge3,600307,20060154512
Sticky Roller5,400308,10045153012
Caster Wheel5,4003010,80060154512
Screw5,400308,10045153012
Lock7,20012010,8001806012012
Gear21,60018054,0004509036034
Clock27,00015040,3202247315152
Slide7,2006018,0001503012013
Zipper14,40012021,6001806012022
Button21,60012054,0003006024044
Liquid5,400308,10045153012
Rubber Band10,8006010,80060303011
Ball21,6009032,400135453032
Magnet10,8006016,20090306022
Toothpaste16,2009016,20090454531
Total298,8002400511,0204,0341,1982,8364947
", + "image_path": "5c88e1548cab93edd3fe5a31ee082684208e011607747650969dc1fda726e079.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 298, + 405, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 298, + 405, + 309 + ], + "spans": [ + { + "bbox": [ + 314, + 298, + 405, + 309 + ], + "type": "text", + "content": "3.3. Data Statistics" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 316, + 555, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 316, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 312, + 316, + 555, + 508 + ], + "type": "text", + "content": "Dataset sample distribution. Table 3 provides a detailed breakdown of the Phys-AD dataset, which includes information on category, the number of training videos and frames, the number of testing videos and frames, the distribution of normal and abnormal samples in the test set, and the number of object and defect types. The length of the video clips in the dataset ranges from 60 to 240 frames, ensuring to fully capture the interaction process in a short time. The dataset contains 2400 training videos and 4034 test videos spreading across 22 categories, 49 object types, and 47 defect types. The frames are extracted from the videos at a rate of 60 FPS. The test set includes 1,198 normal samples and 2,836 abnormal samples. Overall, the Phys-AD dataset covers a wide range of categories and contains a large scale of data, which helps to train more robust anomaly detection models and provides a reasonable evaluation setting." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": "Interaction and defect types across categories. Table 2 details the defect types and corresponding interaction methods for each category in the Phys-AD dataset. Covering 22 categories and including 47 defect types, the dataset features various interactions (e.g., Rotate, Grab). Importantly, interaction methods or defect types with identical names can differ across categories—for example, \"rotation\" for a fan describes the motor-driven movement of its blades, while for a hinge it refers to the turning of its pages around the shaft under robotic control. Furthermore, to detect anomalies with complex physical properties in a human-like manner, we sometimes combine multiple interactions. For instance, when inspecting a screw, we first press it into the hole and then use a robotic arm to rotate it further. Overall, Phys-AD provides a diverse, physics-grounded anomaly detection framework that fosters the development of advanced methods for tackling real-world challenges." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30413" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 294, + 278 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 294, + 278 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 294, + 278 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 294, + 278 + ], + "type": "image", + "image_path": "37b08d3a354aa271391e904fcbd765b6bb939dde01b0e2dacc73c68dd212e63c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 283, + 296, + 383 + ], + "lines": [ + { + "bbox": [ + 55, + 283, + 296, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 296, + 383 + ], + "type": "text", + "content": "Figure 6. PhysAD-Agent: A Large Language Model (LLM)-powered system for physics anomaly detection label augmentation and automatic evaluation. This agent framework consists of two main components: (a) Rich Annotation Generation, where an LLM-based data agent generates detailed anomaly annotations based on video, anomaly state, prompt, and human-annotated physical priors, and (b) Automatic Evaluation, where an LLM-based evaluation agent assesses model predictions to calculate the Physics Anomaly Explanation (PAEval) metric." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 403, + 296, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 535 + ], + "type": "text", + "content": "Labels. For the test set videos, we provide two types of labels. First, there are the common video-level labels: all anomalous videos in the test set are labeled as 1, while normal videos are labeled as 0. For evaluating visual-language models, we also provide text labels. Specifically, for each type of anomaly in the test set, we manually design a textual description label. This label includes both a description of the video content and a physical explanation of the reason for the anomaly. To ensure diversity in our text labels, we also use ChatGPT-4o for text augmentation. For further details on the labeling process, please refer to Fig 6a." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 544, + 178, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 178, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 178, + 558 + ], + "type": "text", + "content": "4. Phys-AD Benchmark" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 564, + 242, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 242, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 242, + 578 + ], + "type": "text", + "content": "4.1. Problem Definition and Challenges" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 616 + ], + "type": "text", + "content": "We establish unsupervised and weakly supervised AD settings for Phys-AD, using unsupervised as the default in our problem definition." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 249, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 249, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 249, + 630 + ], + "type": "text", + "content": "We formulate our Phys-AD setting as two steps." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": "Step 1: Rules deduction Given a set of training objects " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{t_i\\}_{i=1}^N" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " from category " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": ", we first use robotics arms and motors to interact with " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " in corresponding methods " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": ", and we get the interaction process in video format as " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "V_i" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": ". Then, we feed the video sequence " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "V_i" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " and the corresponding category's physical prior " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "P_i" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " together to the deduction function " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " to obtain the normal rules " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": ". After rules deduction for all the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 71, + 555, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 95 + ], + "type": "text", + "content": "categories, we get the rules bank " + }, + { + "bbox": [ + 313, + 71, + 555, + 95 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\{r_i\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 71, + 555, + 95 + ], + "type": "text", + "content": ". Step 1 can be formulated as below:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 410, + 106, + 555, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 106, + 555, + 118 + ], + "spans": [ + { + "bbox": [ + 410, + 106, + 555, + 118 + ], + "type": "interline_equation", + "content": "\\mathcal {T} \\Theta I _ {i} = V _ {i} \\tag {1}", + "image_path": "7be1f3991f8a9486f7afc8fc298fa4388f8cec4189b565b5ba9e88818a05bbbd.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 406, + 126, + 555, + 139 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 126, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 406, + 126, + 555, + 139 + ], + "type": "interline_equation", + "content": "f \\left(V _ {i} \\mid P _ {i}\\right) = r _ {i} \\tag {2}", + "image_path": "7060e9c6873cff59148f58129900c8e540899b5f7561b49db65a611fc5023387.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 407, + 143, + 555, + 158 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 407, + 143, + 555, + 158 + ], + "spans": [ + { + "bbox": [ + 407, + 143, + 555, + 158 + ], + "type": "interline_equation", + "content": "\\mathcal {M} = \\left\\{r _ {i} \\right\\} _ {i = 1} ^ {N} \\tag {3}", + "image_path": "a47d2c475d2620d51670bce990db4dab993018c6dfdcc766ee94a83613e1bbee.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 163, + 555, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 555, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 555, + 186 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 163, + 555, + 186 + ], + "type": "text", + "content": " represents interaction between objects and robotics arms. " + }, + { + "bbox": [ + 313, + 163, + 555, + 186 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 163, + 555, + 186 + ], + "type": "text", + "content": " is the total number of categories." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 186, + 555, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 555, + 258 + ], + "type": "text", + "content": "Step 2: Anomaly reasoning During test time, we first use robotics arms and motors to interact with " + }, + { + "bbox": [ + 313, + 186, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 186, + 555, + 258 + ], + "type": "text", + "content": " from the test set in corresponding methods " + }, + { + "bbox": [ + 313, + 186, + 555, + 258 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 313, + 186, + 555, + 258 + ], + "type": "text", + "content": ", and we get the interaction process in video format as " + }, + { + "bbox": [ + 313, + 186, + 555, + 258 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 313, + 186, + 555, + 258 + ], + "type": "text", + "content": ". Then, we feed the video sequence " + }, + { + "bbox": [ + 313, + 186, + 555, + 258 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 313, + 186, + 555, + 258 + ], + "type": "text", + "content": " and rules " + }, + { + "bbox": [ + 313, + 186, + 555, + 258 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 313, + 186, + 555, + 258 + ], + "type": "text", + "content": " from step 1 to the reasoner " + }, + { + "bbox": [ + 313, + 186, + 555, + 258 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 186, + 555, + 258 + ], + "type": "text", + "content": " to predict the object's anomaly score " + }, + { + "bbox": [ + 313, + 186, + 555, + 258 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 186, + 555, + 258 + ], + "type": "text", + "content": ". Step 2 can be formulated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 411, + 267, + 555, + 280 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 267, + 555, + 280 + ], + "spans": [ + { + "bbox": [ + 411, + 267, + 555, + 280 + ], + "type": "interline_equation", + "content": "\\mathcal {T} \\Theta I _ {i} = V _ {i} \\tag {4}", + "image_path": "94d1e2eb962fd654d9dec0add711606b7f44fbe12cbcc18dda13d74827cd3b5e.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 391, + 289, + 555, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 289, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 391, + 289, + 555, + 301 + ], + "type": "interline_equation", + "content": "R \\left(V _ {i} \\mid r _ {i}\\right) = S, r _ {i} \\in \\mathcal {M} \\tag {5}", + "image_path": "9cc6ff18e67702cbe565f0a5527cf99f70410004cc93bc6f30970f238396e9ca.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 306, + 556, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 556, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 556, + 390 + ], + "type": "text", + "content": "Challenges. Key challenges include combining video information with objects' physical priors to deduct normal rules, capturing long-term temporal dependencies and fine-grained frame-level information, ensuring robust reasoning for anomaly prediction, and generalizing across diverse anomaly patterns. Addressing these challenges is essential for advancing anomaly detection in real-world environment." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 397, + 428, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 428, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 428, + 410 + ], + "type": "text", + "content": "4.2. Evaluation Settings" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 415, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 510 + ], + "type": "text", + "content": "Unsupervised AD. Unsupervised AD is the most common IAD setting for existing IAD datasets and algorithms. Under unsupervised AD setting, the training set contains only normal video data, and the algorithm needs to capture the normal distribution of the data from the training set. In the test set, both normal and abnormal data are included, and the algorithm must distinguish between normal and abnormal data based on the distribution learned during training." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 510, + 556, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 642 + ], + "type": "text", + "content": "Weakly-supervised AD. Our dataset is in the form of videos for complex anomaly detection in industrial scenarios. In the video context, it is inevitable that we need to discuss weakly supervised anomaly detection. Under weakly-supervised Phys-AD setting, 2~4 video-level labeled anomaly samples are sampled from all possible anomaly classes in the test set in our Phys-AD dataset. These sampled anomalies are then removed from the test set. It's worth noting that we only provide video level label in our test set. This is because most anomalies in our test set require reasoning based on the entire video content and physical knowledge of the objects." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Video-Understanding AD. Video-Understanding models are another potential solution to our Phys-AD setting. Similar to unsupervised setting, we only provide normal videos during training. The visual language models need to provide explicit normal rules in language format during training. In the test phase, these video-understanding AD" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30414" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 108, + 553, + 369 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 4. Video-level AUROC (↑) result of 22 categories on Phys-AD dataset. We include Unsupervised, Weakly-supervised and Video-understanding methods. The best per-category result for each class of methods is highlighted in **bold**. 'ZS ImgB', 'V-ChatGPT', 'V-LLaMA', 'V-LLaVA' denote ZS ImageBind, Video-ChatGPT, Video-LLaMA and Video-LLaVA." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 108, + 553, + 369 + ], + "lines": [ + { + "bbox": [ + 57, + 108, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 57, + 108, + 553, + 369 + ], + "type": "table", + "html": "
Category.UnsupervisedWeakly-supervisedVideo-understanding
MPN [34]MemAE [16]MNAD.p [40]MNAD.r [40]SVM [44]VADClip [51]S3R [48]MGFN [10]LAVAD [56]ZS Clip [42]ZS ImgB [15]V-ChatGPT [36]V-LLaMA [59]V-LLaVA [28]
Car0.2290.5230.4920.9440.5870.5810.6060.5710.5570.5000.5000.5000.6780.522
Fan0.8110.3710.8100.5420.5000.6240.6400.5420.5100.5000.5000.5490.5920.611
Rolling Bearing0.3530.0440.3520.8000.9330.5890.6010.6800.5320.5000.5000.3000.5000.500
Spherical Bearing0.1130.0920.9620.8130.6500.5000.6820.5830.4350.5000.5000.4500.5500.500
Servo0.3640.4450.9750.8780.5000.5180.5920.5560.5020.5000.5000.5060.6830.464
Clip0.5350.4430.6300.3330.5000.4120.5610.5630.5160.5000.5000.6690.5560.458
U Disk0.2400.6170.6090.9400.5000.5300.5490.5700.5130.5000.5000.5650.5750.500
Hinge0.7690.8700.7050.8950.5000.7370.5610.5500.5640.5000.5000.5000.6920.500
Sticky Roller0.9670.9670.4510.9360.5000.5420.8350.6690.2660.5000.5000.4500.5440.467
Caster Wheel0.3640.5230.2710.5080.5000.5870.6760.6760.6150.5000.5000.4440.6420.500
Screw0.5220.5670.6800.5470.5000.5000.6570.5410.6880.5000.5000.4720.2560.550
Lock0.5630.5970.4300.6410.7330.4520.5230.6260.3410.5000.5000.2790.4940.500
Gear0.5290.5100.6520.6940.5000.5190.5800.5440.6030.5000.5000.5440.6030.517
Clock0.3400.5420.3950.5870.5000.5000.5490.5090.5000.5000.5000.5010.3600.500
Slide0.5170.9620.9170.7840.5000.5310.6110.5680.4250.5000.5000.5620.5670.179
Zipper0.8150.5920.8290.4210.5000.5040.6360.6330.5350.5000.5000.5470.4890.500
Button0.8530.3650.6600.5680.5000.6270.5150.5660.4390.5000.5000.5150.5170.360
Liquid0.1840.7000.6710.8310.5000.5420.5950.7930.5040.5000.5000.4100.2780.217
Rubber Band0.3740.3680.3660.3070.5670.4820.6230.6040.5110.5000.5000.5170.4500.450
Ball0.5430.3830.7280.6870.5000.5000.6670.5670.6030.5000.5000.5620.6360.533
Magnet0.6710.4640.6910.4380.5000.5000.5480.7190.5020.5000.5000.6830.3000.400
Toothpaste0.5870.8890.5200.6310.5000.5000.6860.7110.5620.5000.5000.3760.5500.467
Average0.5110.5380.6270.6690.5440.5350.6130.6060.5100.5000.5000.4960.5230.463
", + "image_path": "ce97f53bb8944f82e5793cc25a917db16081023a8abd500324d9f0518cf27c17.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 384, + 294, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 294, + 420 + ], + "type": "text", + "content": "methods have to truly understand the videos and predict the anomalies based on the language rules withdrawn from the training phase." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 441, + 168, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 168, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 168, + 453 + ], + "type": "text", + "content": "4.3. Evaluation Metrics" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 459, + 295, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 295, + 531 + ], + "type": "text", + "content": "Physics anomaly accuracy metrics. We use the Area Under the Receiver Operating Characteristic Curve (AUROC) to evaluate video-level anomaly detection performance. We also report the average precision (AP), i.e. the area under the video-level precision-recall curve and the acc (ACCURACY), following previous works [47, 49]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "Physics Anomaly Explanation (PAEval) metric. Directly utilizing existing VLMs [59][36] to understand the videos in our dataset and detect the anomalies is a potential solution to our challenge. The key point is: Can existing VLMs truly understand physical rules and reason in a right way? Specifically designed for video-understanding methods, we introduce a new metric named Physics Anomaly Explanation (PAEval) metric. To be more specific, PAEval evaluate the anomaly detection performance of algorithms based on VLMs from three different perspectives: classification, description, and explanation. Classification refers to the traditional anomaly detection metrics like AUROC, etc. Inspired by the works [11, 61], PAEval also introduces two additional evaluation metrics: description and physical explanation. Description refers to the model's ability to" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 384, + 555, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 504 + ], + "type": "text", + "content": "correctly describe the content of the video, used to assess whether the model has the capability to describe physical phenomena. Explanation refers to the model's ability to correctly explain the physical causes of anomalies in the video, used to evaluate the VLM's reasoning ability. To provide labels for description and explanation metrics, we manually labeled each type of defects from all categories and performed data augmentation by ChatGPT to ensure the robustness of the detection. The whole pipeline of PAEval metric is depicted in Fig 6." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 518, + 444, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 444, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 444, + 532 + ], + "type": "text", + "content": "5. Benchmarking Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 539, + 480, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 480, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 480, + 551 + ], + "type": "text", + "content": "5.1. Benchmark Methods Selection" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "content": "For the Phys-AD setting, we select popular and reproducible video anomaly detection methods across unsupervised, weakly-supervised, and video-understanding setting. In the unsupervised setting, we focus on reconstruction, prediction, and embedding-based models like MemAE [16] and MNAD [40], which use memory modules to enhance anomaly discrimination. For weakly-supervised anomaly detection, we adopt methods such as VadCLIP [51] and MGFN [10], which use feature magnitudes and vision-language associations. In video-understanding, we evaluate video-language models like Video-ChatGPT [36] and image-language models like CLIP [42] to predict detailed anomaly descriptions and scores through cross-modal embeddings." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30415" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 95, + 294, + 150 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 295, + 92 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 295, + 92 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 295, + 92 + ], + "type": "text", + "content": "Table 5. Physics Anomaly Explanation (PAEval) metric results on our Phys-AD dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 95, + 294, + 150 + ], + "lines": [ + { + "bbox": [ + 57, + 95, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 57, + 95, + 294, + 150 + ], + "type": "table", + "html": "
MethodsClassification (↑)Description (↑)Explanation (↑)
LAVAD[56]0.5100.1570.000
Video-ChatGPT[36]0.4960.1310.160
Video-LLaMA[59]0.5230.1370.303
Video-LLaVA[28]0.4630.2190.282
", + "image_path": "d8144d5b51de9e11ae4893f04065b9163b95c6453985312f9855c3e67382fa9d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 166, + 295, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 166, + 295, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 166, + 295, + 202 + ], + "type": "text", + "content": "Code and Experiment details. We provide code and toolkit for our dataset and benchmark. More experiment details are listed in Appendix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 209, + 146, + 222 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 146, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 146, + 222 + ], + "type": "text", + "content": "5.2. Main Findings" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 226, + 296, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 226, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 296, + 346 + ], + "type": "text", + "content": "Overall anomaly detection benchmarking results. Table 4 shows that existing video anomaly detection and videowhildunderstanding methods achieve limited performance on the Phys-AD dataset, with the highest AUROC only reaching " + }, + { + "bbox": [ + 55, + 226, + 296, + 346 + ], + "type": "inline_equation", + "content": "66.9\\%" + }, + { + "bbox": [ + 55, + 226, + 296, + 346 + ], + "type": "text", + "content": " for MNAD.r [40]. This result underscores the heightened complexity of Phys-AD compared to existing datasets, highlighting a domain shift that challenges current industrial anomaly detection algorithms, which are often tuned to visually distinct anomaly patterns in single frames rather than complex temporal or physical cues." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 349, + 296, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 296, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 296, + 516 + ], + "type": "text", + "content": "Anomaly detection via unsupervised and weakly supervised methods. Our experiment includes unsupervised methods (e.g., MemAE [16], MPN [34], MNAD [40]) and weakly supervised methods (e.g., S3R [48], MGFN [10], VADClip [51]). Unsupervised methods like MNAD.p [40] performed better on temporal anomalies (81.0% for fan, 68.0% for screw) by leveraging prediction-based approaches, which excel in scenarios requiring temporal understanding. Weakly supervised methods improve baseline scores across complex classes, preventing extreme low scores in challenging categories such as spherical bearing. However, weakly supervised methods show marginally lower performance in simpler anomaly classes, indicating trade-offs introduced by anomaly samples in training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 519, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 295, + 639 + ], + "type": "text", + "content": "Unsupervised anomaly detection method performance by category. Among unsupervised methods, MemAE [16] achieves high AUROC on objects with spatial anomalies (e.g., sticky roller, " + }, + { + "bbox": [ + 55, + 519, + 295, + 639 + ], + "type": "inline_equation", + "content": "96.7\\%" + }, + { + "bbox": [ + 55, + 519, + 295, + 639 + ], + "type": "text", + "content": ") but underperforms in temporal anomaly classes like rolling and spherical bearings, where it achieves only " + }, + { + "bbox": [ + 55, + 519, + 295, + 639 + ], + "type": "inline_equation", + "content": "4.4\\%" + }, + { + "bbox": [ + 55, + 519, + 295, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 519, + 295, + 639 + ], + "type": "inline_equation", + "content": "9.2\\%" + }, + { + "bbox": [ + 55, + 519, + 295, + 639 + ], + "type": "text", + "content": ", respectively. MNAD [40] improves temporal sensitivity by incorporating temporal prototypes, achieving better scores for bearings but still struggling with complex anomaly types, revealing limitations in purely prototype-based approaches." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "Anomaly detection via Video-Language Models. VLM-based methods (e.g., Video-ChatGPT [36], VideoLLaMA [59]) struggle, with top-performing VideoLLaMA [59] reaching only " + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "inline_equation", + "content": "52.3\\%" + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": " average AUROC. Performance is impacted by reliance on pre-trained weights that are not optimized for physics-grounded video content, evident" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": "from low AUROC scores in categories with nuanced physical dynamics, such as hinges and screws. Further, PAEval results suggest that these models lack effective reasoning about object physical dynamics and behaviors influenced by physical forces, underscoring a gap between VLM capabilities and the demands of IAD tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 147, + 556, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 147, + 556, + 231 + ], + "spans": [ + { + "bbox": [ + 313, + 147, + 556, + 231 + ], + "type": "text", + "content": "Anomaly explanation via Video-Language Models. Table 5 reports PAEval metric results, with the best-performing VLM achieving only " + }, + { + "bbox": [ + 313, + 147, + 556, + 231 + ], + "type": "inline_equation", + "content": "21.9\\%" + }, + { + "bbox": [ + 313, + 147, + 556, + 231 + ], + "type": "text", + "content": " in description and " + }, + { + "bbox": [ + 313, + 147, + 556, + 231 + ], + "type": "inline_equation", + "content": "30.3\\%" + }, + { + "bbox": [ + 313, + 147, + 556, + 231 + ], + "type": "text", + "content": " in explanation. These findings emphasize that current VLMs lack the depth in physical reasoning and temporal coherence required for understanding real-world physics-based scenarios, which our Phys-AD dataset demands." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 233, + 556, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 233, + 556, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 556, + 306 + ], + "type": "text", + "content": "Summary. Overall, the results reveal Phys-AD's unique challenge in requiring both high spatial detail and temporal comprehension, areas where existing methods and models underperform. This analysis points to the need for future research in models that integrate temporal reasoning with physics-based anomaly detection." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 316, + 388, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 316, + 388, + 328 + ], + "spans": [ + { + "bbox": [ + 314, + 316, + 388, + 328 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 335, + 556, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 556, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 556, + 539 + ], + "type": "text", + "content": "In this paper, we introduce the first industrial anomaly detection task focusing on real-world scenarios where physical understanding and reasoning are essential for anomaly detection. We present the Physics Anomaly Detection (Phys-AD) dataset, a large-scale, physics-grounded video dataset with over 6400 videos across 22 categories and 49 object types interacting with robotic systems, capturing 47 anomaly types that necessitate visual and physical understanding. We assess Phys-AD, highlighting the lack of baseline approaches for high-level reasoning in anomaly detection. Additionally, we propose the Physics Anomaly Explanation (PAEval) metric to evaluate visual language models (VLMs) on physics-based reasoning. Experiments show that current VLMs fall short of human-level understanding in physics-based anomaly scenarios. This work marks a milestone for industrial anomaly detection, promoting physics-grounded reasoning in complex industrial settings." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 542, + 556, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 556, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 556, + 686 + ], + "type": "text", + "content": "Limitation and future work. Although our Phys-AD dataset provides a large variety of objects with diverse physical properties and various types of interaction methods, we plan to add even more diverse interaction methods and objects in the future to better meet the demands of complex real-world industrial scenarios. Due to the significant differences between our Phys-AD dataset and current industrial anomaly detection and video anomaly detection datasets, most existing anomaly detection algorithms cannot be directly applied to our dataset. In the future, we will test more algorithms on our Phys-AD dataset and provide experimental results across more settings like zero-shot, few-shot, semi-supervised settings, etc." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30416" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "type": "text", + "content": "[1] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In ICCV, 2015. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 159 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 159 + ], + "type": "text", + "content": "[2] Fabien Baradel, Natalia Neverova, Julien Mille, Greg Mori, and Christian Wolf. Cophy: Counterfactual learning of physical dynamics. arXiv preprint arXiv:1909.12000, 2019.3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 160, + 295, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 160, + 295, + 215 + ], + "spans": [ + { + "bbox": [ + 62, + 160, + 295, + 215 + ], + "type": "text", + "content": "[3] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad — a comprehensive real-world dataset for unsupervised anomaly detection. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9584-9592, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 217, + 295, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 217, + 295, + 249 + ], + "spans": [ + { + "bbox": [ + 62, + 217, + 295, + 249 + ], + "type": "text", + "content": "[4] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. MVtec AD – A comprehensive real-world dataset for unsupervised anomaly detection. In CVPR, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 251, + 296, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 296, + 305 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 296, + 305 + ], + "type": "text", + "content": "[5] Paul Bergmann, Kilian Batzner, Michael Fauser, David Sattlegger, and Carsten Steger. Beyond dents and scratches: Logical constraints in unsupervised anomaly detection and localization. International Journal of Computer Vision, 130 (4):947-969, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 308, + 296, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 296, + 373 + ], + "type": "text", + "content": "[6] Paul Bergmann, Xin Jin, David Sattlegger, and Carsten Steger. The mvtec 3d-ad dataset for unsupervised 3d anomaly detection and localization. In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications. SCITEPRESS - Science and Technology Publications, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 375, + 295, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 375, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 375, + 295, + 407 + ], + "type": "text", + "content": "[7] Paul Bergmann, Xin Jin, David Sattlegger, and Carsten Steger. The MVTec 3d-AD dataset for unsupervised 3d anomaly detection and localization. In VISAPP, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 409, + 296, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 296, + 464 + ], + "type": "text", + "content": "[8] Luca Bonfiglioli, Marco Toschi, Davide Silvestri, Nicola Fioraio, and Daniele De Gregorio. The eyecandies dataset for unsupervised multimodal anomaly detection and localization. In Proceedings of the Asian Conference on Computer Vision, pages 3586-3602, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 466, + 296, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 466, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 466, + 296, + 509 + ], + "type": "text", + "content": "[9] Yunkang Cao, Xiaohao Xu, Jiangning Zhang, Yuqi Cheng, Xiaonan Huang, Guansong Pang, and Weiming Shen. A survey on visual anomaly detection: Challenge, approach, and prospect. arXiv preprint arXiv:2401.16402, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 511, + 296, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 511, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 57, + 511, + 296, + 576 + ], + "type": "text", + "content": "[10] Yingxian Chen, Zhengzhe Liu, Baoheng Zhang, Wilton Fok, Xiaojuan Qi, and Yik-Chung Wu. Mgfn: Magnitude-contrastive glance-and-focus network for weakly-supervised video anomaly detection. In Proceedings of the AAAI conference on artificial intelligence, pages 387-395, 2023. 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 578, + 296, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 296, + 654 + ], + "type": "text", + "content": "[11] Hang Du, Sicheng Zhang, Binzhu Xie, Guoshun Nan, Jiayang Zhang, Junrui Xu, Hangyu Liu, Sicong Leng, Jiangming Liu, Hehe Fan, et al. Uncovering what why and how: A comprehensive benchmark for causation understanding of video anomaly. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18793-18803, 2024. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 657, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 657, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 657, + 295, + 689 + ], + "type": "text", + "content": "[12] Jiafei Duan, Samson Yu, and Cheston Tan. Space: A simulator for physical interactions and causal learning in 3d environments. In ICCV, pages 2058-2063, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[13] Jiafei Duan, Samson Yu, Soujanya Poria, Bihan Wen, and Cheston Tan. Pip: Physical interaction prediction via mental" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 555, + 95 + ], + "type": "text", + "content": "simulation with span selection. In ECCV, pages 405-421. Springer, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 96, + 555, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 555, + 148 + ], + "type": "text", + "content": "[14] Jia-Chang Feng, Fa-Ting Hong, and Wei-Shi Zheng. Mist: Multiple instance self-training framework for video anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14009-14018, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 151, + 555, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 555, + 205 + ], + "type": "text", + "content": "[15] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 206, + 555, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 206, + 555, + 280 + ], + "spans": [ + { + "bbox": [ + 317, + 206, + 555, + 280 + ], + "type": "text", + "content": "[16] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1705-1714, 2019. 3, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 282, + 554, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 282, + 554, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 282, + 554, + 326 + ], + "type": "text", + "content": "[17] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 327, + 554, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 554, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 554, + 360 + ], + "type": "text", + "content": "[18] Mahmudul Hasan, Jonghyun Choi, Jan Neumann, Amit K Roy-Chowdhury, and Larry S Davis. Learning temporal regularity in video sequences. In CVPR, 2016. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 361, + 554, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 554, + 392 + ], + "type": "text", + "content": "[19] Drew A. Hudson and Christopher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 394, + 555, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 555, + 436 + ], + "type": "text", + "content": "[20] Steeven Janny, Fabien Baradel, Natalia Neverova, Madiha Nadri, Greg Mori, and Christian Wolf. Filtered-cophy: Unsupervised learning of counterfactual physics in pixel space. In ICLR, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 438, + 555, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 438, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 438, + 555, + 502 + ], + "type": "text", + "content": "[21] Stepan Jezek, Martin Jonak, Radim Burget, Pavel Dvorak, and Milos Skotak. Deep learning-based defect detection of metal parts: evaluating current methods in complex conditions. In 2021 13th International Congress on Ultra Modern Telecommunications and Control Systems and Workshops (ICUMT), pages 66-71, 2021. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 504, + 554, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 554, + 547 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 554, + 547 + ], + "type": "text", + "content": "[22] Huaizu Jiang, Xiaojian Ma, Weili Nie, Zhiding Yu, Yuke Zhu, and Anima Anandkumar. Bongard-hoi: Benchmarking few-shot visual reasoning for human-object interactions. In CVPR, pages 19056-19065, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 548, + 555, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 555, + 591 + ], + "type": "text", + "content": "[23] Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C. Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In CVPR, 2017. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 593, + 555, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 555, + 645 + ], + "type": "text", + "content": "[24] Hyekang Kevin Joo, Khoa Vo, Kashu Yamazaki, and Ngan Le. Clip-tsa: Clip-assisted temporal self-attention for weakly-supervised video anomaly detection. In 2023 IEEE International Conference on Image Processing (ICIP), pages 3230-3234. IEEE, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 555, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 679 + ], + "type": "text", + "content": "[25] Federico Landi, Cees GM Snok, and Rita Cucchiara. Anomaly locality in video surveillance. arXiv preprint arXiv:1901.10364, 2019. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 681, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 712 + ], + "type": "text", + "content": "[26] Ting Li, Charles Dubout, Emma K Wampler, Steven Yantis, Donald Geman, et al. Comparing machines and humans on a visual categorization test. 2011. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30417" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 117 + ], + "type": "text", + "content": "[27] Wenqiao Li, Xiaohao Xu, Yao Gu, Bozhong Zheng, Shenghua Gao, and Yingna Wu. Towards scalable 3d anomaly detection and localization: A benchmark via 3d anomaly synthesis and a self-supervised learning network. arXiv preprint, 2023. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 161 + ], + "type": "text", + "content": "[28] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 163, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 163, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 57, + 163, + 294, + 185 + ], + "type": "text", + "content": "[29] Jiaqi Liu, Guoyang Xie, Rui Chen, et al. Real3d-ad: A dataset of point cloud anomaly detection. NeurIPS, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 229 + ], + "type": "text", + "content": "[30] Jiaqi Liu, Guoyang Xie, Ruitao Chen, Xinpeng Li, Jinbao Wang, Yong Liu, Chengjie Wang, and Feng Zheng. Real3dad: A dataset of point cloud anomaly detection. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 295, + 274 + ], + "type": "text", + "content": "[31] Kun Liu and Huadong Ma. Exploring background-bias for anomaly detection in surveillance videos. In Proceedings of the 27th ACM International Conference on Multimedia, pages 1490-1499, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 277, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 295, + 308 + ], + "type": "text", + "content": "[32] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection-a new baseline. In CVPR, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 311, + 295, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 311, + 295, + 375 + ], + "spans": [ + { + "bbox": [ + 56, + 311, + 295, + 375 + ], + "type": "text", + "content": "[33] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13588-13597, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 377, + 295, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 295, + 431 + ], + "type": "text", + "content": "[34] Hui Lv, Chen Chen, Zhen Cui, Chunyan Xu, Yong Li, and Jian Yang. Learning normal dynamics in videos with meta prototype network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15425-15434, 2021. 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 295, + 487 + ], + "type": "text", + "content": "[35] Hui Lv, Zhongqi Yue, Qianru Sun, Bin Luo, Zhen Cui, and Hanwang Zhang. Unbiased multiple instance learning for weakly supervised video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8022-8031, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 490, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 295, + 533 + ], + "type": "text", + "content": "[36] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 295, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 295, + 588 + ], + "type": "text", + "content": "[37] Domingo Mery, Vladimir Riffo, Uwe Zscherpel, German Mondragon, Ivan Lillo, Irene Zuccar, Hans Lobel, and Miguel Carrasco. Gdxray: The database of x-ray images for nondestructive testing. Journal of Nondestructive Evaluation, 34 (4):42, 2015. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 295, + 644 + ], + "type": "text", + "content": "[38] Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. Vt-adj: A vision transformer network for image anomaly detection and localization. In 2021 IEEE 30th International Symposium on Industrial Electronics (ISIE). IEEE, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "type": "text", + "content": "[39] Weili Nie, Zhiding Yu, Lei Mao, Ankit B Patel, Yuke Zhu, and Animashree Anandkumar. Bongard-logo: A new benchmark for human-level concept learning and reasoning. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "text", + "content": "[40] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In Proceed-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14372-14381, 2020. 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "text", + "content": "[41] Yujiang Pu, Xiaoyu Wu, and Shengjin Wang. Learning prompt-enhanced context features for weakly-supervised video anomaly detection. arXiv preprint arXiv:2306.14451, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 129, + 555, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 193 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 193 + ], + "type": "text", + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "type": "text", + "content": "[43] Binyi Su, Zhong Zhou, and Haiyong Chen. Pvel-ad: A large-scale open-world dataset for photovoltaic cell anomaly detection. IEEE Transactions on Industrial Informatics, 19 (1):404-413, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 239, + 553, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 260 + ], + "type": "text", + "content": "[44] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In CVPR, 2018. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 262, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 316 + ], + "type": "text", + "content": "[45] Kamalakar Vijay Thakare, Yash Raghuwanshi, Debi Prosad Dogra, Heeseung Choi, and Ig-Jae Kim. Dyannet: A scene dynamicity guided self-trained video anomaly detection network. In Proceedings of the IEEE/CVF Winter conference on applications of computer vision, pages 5541-5550, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 317, + 555, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 317, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 317, + 555, + 369 + ], + "type": "text", + "content": "[46] Chengjie Wang, Wenbing Zhu, Bin-Bin Gao, Zhenye Gan, Jianning Zhang, Zhihao Gu, Shuguang Qian, Mingang Chen, and Lizhuang Ma. Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detection, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 372, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 426 + ], + "type": "text", + "content": "[47] Yue Wang, Jinlong Peng, Jiangning Zhang, Ran Yi, Yabiao Wang, and Chengjie Wang. Multimodal industrial anomaly detection via hybrid fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8032-8041, 2023. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 427, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 553, + 470 + ], + "type": "text", + "content": "[48] Jhih-Ciang Wu, He-Yen Hsieh, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng-Luh Liu. Self-supervised sparse representation for video anomaly detection. In European Conference on Computer Vision, pages 729-745. Springer, 2022. 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 472, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 553, + 536 + ], + "type": "text", + "content": "[49] Peng Wu, Jing Liu, Yujia Shi, Yujia Sun, Fangtao Shao, Zhaoyang Wu, and Zhiwei Yang. Not only look, but also listen: Learning multimodal violence detection under weak supervision. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 322-339. Springer, 2020. 3, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 537, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 569 + ], + "type": "text", + "content": "[50] Peng Wu, Xuerong Zhou, Guansong Pang, Yujia Sun, Jing Liu, Peng Wang, and Yanning Zhang. Open-vocabulary video anomaly detection. arXiv preprint arXiv:2311.07042, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 571, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 624 + ], + "type": "text", + "content": "[51] Peng Wu, Xuerong Zhou, Guansong Pang, Lingru Zhou, Qingsen Yan, Peng Wang, and Yanning Zhang. Vadclip: Adapting vision-language models for weakly supervised video anomaly detection. arXiv preprint arXiv:2308.11681, 2023.3,7,8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "text", + "content": "[52] Dan Xu, Yan Yan, Elisa Ricci, and Nicu Sebe. Detecting anomalous events in videos by learning deep representations of appearance and motion. Computer Vision and Image Understanding, 156:117-127, 2017. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[53] Xiaohao Xu, Yunkang Cao, Yongqi Chen, Weiming Shen, and Xiaonan Huang. Customizing visual-language foundation models for multi-modal anomaly detection and reasoning. arXiv preprint arXiv:2403.11083, 2024. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30418" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 622 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[54] Zhiwei Yang, Jing Liu, Zhaoyang Wu, Peng Wu, and Xiaotao Liu. Video event restoration based on keyframes for video anomaly detection. In CVPR, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "text", + "content": "[55] M Zaigham Zaheer, Arif Mahmood, M Haris Khan, Mattia Segu, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14744-14754, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 206 + ], + "type": "text", + "content": "[56] Luca Zanella, Willi Menapace, Massimiliano Mancini, Yiming Wang, and Elisa Ricci. Harnessing large language models for training-free video anomaly detection. arXiv preprint arXiv:2404.01014, 2024. 3, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 240 + ], + "type": "text", + "content": "[57] Aimen Zerroug, Mohit Vaishnav, Julien Colin, Sebastian Musslick, and Thomas Serre. A benchmark for compositional visual reasoning. arXiv preprint arXiv:2206.05379, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 295, + 274 + ], + "type": "text", + "content": "[58] Chi Zhang, Feng Gao, Baoxiong Jia, Yixin Zhu, and Song-Chun Zhu. Raven: A dataset for relational and analogical visual reasoning. In CVPR, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 295, + 308 + ], + "type": "text", + "content": "[59] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 309, + 295, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 295, + 363 + ], + "type": "text", + "content": "[60] Huaxin Zhang, Xiang Wang, Xiaohao Xu, Xiaonan Huang, Chuchu Han, Yuehuan Wang, Changxin Gao, Shanjun Zhang, and Nong Sang. Glancevad: Exploring glance supervision for label-efficient video anomaly detection. arXiv preprint arXiv:2403.06154, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 295, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 295, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 295, + 419 + ], + "type": "text", + "content": "[61] Huaxin Zhang, Xiaohao Xu, Xiang Wang, Jialong Zuo, Chuchu Han, Xiaonan Huang, Changxin Gao, Yuehuan Wang, and Nong Sang. Holmes-vad: Towards unbiased and explainable video anomaly detection via multi-modal llm. arXiv preprint arXiv:2406.12235, 2024. 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 421, + 295, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 295, + 464 + ], + "type": "text", + "content": "[62] Yiheng Zhang, Yunkang Cao, Xiaohao Xu, and Weiming Shen. Logic: An llm-driven framework for logical anomaly detection. IEEE Transactions on Automation Science and Engineering, pages 1-0, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 466, + 295, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 295, + 508 + ], + "type": "text", + "content": "[63] Jia-Xing Zhong, Nannan Li, Weijie Kong, Shan Liu, Thomas H Li, and Ge Li. Graph convolutional label noise cleaner: Train a plug-and-play action classifier for anomaly detection. In CVPR, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 510, + 295, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 295, + 543 + ], + "type": "text", + "content": "[64] Qiang Zhou, Weize Li, Lihan Jiang, Guoliang Wang, Guyue Zhou, Shanghang Zhang, and Hao Zhao. Pad: A dataset and benchmark for pose-agnostic anomaly detection, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 544, + 295, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 295, + 586 + ], + "type": "text", + "content": "[65] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In ECCV, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 589, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 589, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 589, + 295, + 622 + ], + "type": "text", + "content": "[66] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation, 2022. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30419" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models/9698d3f1-26f6-4c2e-aef0-60a71f074ba9_content_list.json b/2025/Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models/9698d3f1-26f6-4c2e-aef0-60a71f074ba9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ccb95cdc46195b0f659a905cd37018fb20e4fbcc --- /dev/null +++ b/2025/Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models/9698d3f1-26f6-4c2e-aef0-60a71f074ba9_content_list.json @@ -0,0 +1,2192 @@ +[ + { + "type": "text", + "text": "Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models", + "text_level": 1, + "bbox": [ + 223, + 130, + 776, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiacong Xu $^{1*}$ Shao-Yuan Lo $^{2}$ Bardia Safaei $^{1}$ Vishal M. Patel $^{1}$ Isht Dwivedi $^{2}$ $^{1}$ Johns Hopkins University $^{2}$ Honda Research Institute USA", + "bbox": [ + 181, + 202, + 815, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jxu155, bsafaei1, vpatel36}@jhu.edu {shao-yuan.lo, idwivedi}@honda-ri.com", + "bbox": [ + 165, + 241, + 826, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zero-Shot Anomaly Detection (ZSAD) is an emerging AD paradigm. Unlike the traditional unsupervised AD setting that requires a large number of normal samples to train a model, ZSAD is more practical for handling data-restricted real-world scenarios. Recently, Multimodal Large Language Models (MLLMs) have shown revolutionary reasoning capabilities in various vision tasks. However, the reasoning of image abnormalities remains underexplored due to the lack of corresponding datasets and benchmarks. To facilitate research in AD & reasoning, we establish the first visual instruction tuning dataset, Anomaly-Instruct-125k, and the evaluation benchmark, VisA-D&R. Through investigation with our benchmark, we reveal that current MLLMs like GPT-4o cannot accurately detect and describe fine-grained anomalous details in images. To address this, we propose Anomaly-OneVision (Anomaly-OV), the first specialist visual assistant for ZSAD and reasoning. Inspired by human behavior in visual inspection, Anomaly-OV leverages a Look-Twice Feature Matching (LTFM) mechanism to adaptively select and emphasize abnormal visual tokens. Extensive experiments demonstrate that Anomaly-OV achieves significant improvements over advanced generalist models in both detection and reasoning. Extensions to medical and 3D AD are provided for future study. The link to our project page: https://xujiacong.github.io/Anomaly-OV/", + "bbox": [ + 89, + 321, + 485, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 729, + 220, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual Anomaly Detection (AD) is a well-established task in computer vision, extensively applied in scenarios such as industrial defect inspection [2, 3, 5, 12, 35, 69, 76, 77, 92, 98] and medical image diagnosis [1, 24, 31, 36, 89, 90, 103, 106]. In the traditional unsupervised AD (a.k.a. one-class AD) setting, models learn the distribution of normal visual features from normal samples and are required to identify anomaly samples during inference. While recent advance", + "bbox": [ + 89, + 755, + 485, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/87ca03e000aaeacd11d7caf00c0f3e4662be350fb651c38b989f04967522b4d5.jpg", + "image_caption": [ + "Figure 1. Visualization of the image-level AUROC comparison between our Anomaly-OV and current state-of-the-art ZSAD methods (WinCLIP [38], AnoVL [19], AnomalyCLIP [110], AdaCLIP [6]). Notably, our zero-shot performance on VisA even surpasses most recent advances in the few-shot setting [28, 51, 112]." + ], + "image_footnote": [], + "bbox": [ + 516, + 290, + 906, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ments [9, 25, 32–34, 37, 42, 82, 84, 95, 97, 104] have significantly improved the detection performance, these approaches assume the availability of a substantial number of normal samples. However, this assumption becomes impractical in certain scenarios due to strict data privacy policies and the significant human effort required for data classification, sometimes involving experts or specialists. Therefore, Zero-Shot Anomaly Detection (ZSAD) is emerging as a popular research direction, leading to the development of many innovative methods [6, 17, 27, 38, 43, 52, 78, 79, 110, 113].", + "bbox": [ + 511, + 685, + 908, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in Multimodal Large Language Models (MLLMs) [7, 15, 44, 45, 47, 48, 57, 58, 111] have shown revolutionary reasoning capabilities in various vision tasks [14, 29, 67, 70, 80, 91, 94, 107, 109]. However, the rea", + "bbox": [ + 511, + 839, + 910, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Most of this work was done when J. Xu was an intern at HRI-USA.", + "bbox": [ + 107, + 886, + 472, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20370", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "soning of image abnormalities has not been explored due to the challenges of collecting large-scale datasets and establishing benchmarks. Existing methods simply predict the likelihood of an anomaly without providing rationales [6, 11, 19, 38, 110]. In contrast, for better interpretability, robustness, and trustworthiness, people would expect models to explain why an image is considered anomalous and provide visual evidence. Interestingly, we find that recent advanced MLLMs, such as GPT-4o [72], fall short in AD & reasoning. As shown in Figure 2, while the detection is correct, the explanation from GPT-4o lacks accuracy, indicating a gap in a comprehensive understanding of the anomaly.", + "bbox": [ + 89, + 90, + 483, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To expedite research in AD & reasoning, we establish the first visual instruction tuning dataset, Anomaly-Instruct-125k, and the evaluation benchmark, VisA-D&R, through intensive human efforts. After evaluating current generalist MLLMs, we observe that these models fail to accurately detect and describe fine-grained anomalous details in images. To address this, we propose Anomaly-OneVision (Anomaly-OV), the first specialist visual assistant for ZSAD and reasoning. Unlike existing ZSAD methods [6, 11, 19, 38, 110], Anomaly-OV directly learns object-awareness abnormality embeddings in feature space using only the visual encoder. Inspired by human behavior in visual inspection, Anomaly-OV employs a Look-Twice Feature Matching (LTFM) mechanism to assist its LLM in adaptively selecting and emphasizing the most suspicious abnormal visual tokens.", + "bbox": [ + 89, + 272, + 485, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments demonstrate that Anomaly-OV achieves significant improvements over advanced generalist models in both detection and reasoning. Extended results of Anomaly-OV, from applications in industrial defect detection to 3D inspection and medical image diagnosis, are provided for future study. With precise descriptions and rationales of visual anomalies, our model can infer potential causes (see Figure 2), assess current impacts, and offer improvement suggestions, positioning itself as a reliable assistant for visual inspection. Our contributions are in two folds:", + "bbox": [ + 89, + 500, + 483, + 648 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We establish the first visual instruction tuning dataset and benchmark for anomaly detection and reasoning.", + "- We propose the first specialist visual assistant with state-of-the-art performance for this new impactful domain." + ], + "bbox": [ + 91, + 650, + 483, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 723, + 232, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal Large Language Models. Vision-Language Models (VLMs), such as CLIP [73], exhibit robust zero-shot classification capabilities and have been applied to a range of downstream vision tasks [10, 53, 56, 61, 86]. Combining a VLM's vision encoder and an LLM [20, 62, 74], MLLMs [15, 48, 49, 57, 111] enable text-based interactions related to visual content. MLLMs have shown remarkable reasoning capability, particularly when incorporated with prompting strategies such as Chain-of-Thought [66, 88, 105]. Recent studies have harnessed MLLMs to provide reasoning for", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/13bec8769ce795b7c1bf630673506d45bbe937ad7c185058bb030de265045c11.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 526, + 95, + 617, + 161 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/25d9ca1d82563a8f93cbf3b8f88b477c4c9f7175c783c82e46e24750d5b1726d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 95, + 723, + 161 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "GT: The top-left candle shows a missing chunk of wax and an irregular indentation near the wick.", + "bbox": [ + 728, + 99, + 890, + 157 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Do you see any anomalies in the image?", + "bbox": [ + 581, + 181, + 859, + 194 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fa1ec3fecb2f2f7a21fcf15c025d1cc5e0202f2e27fadac7851d6ee400a6f176.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 867, + 172, + 895, + 196 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1028b0d63a08272fee646e1912e0f6c44eeb5805de96a4139539c94247c752f7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 526, + 198, + 550, + 217 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The candle in the top left appears to have a damaged or missing wick, which is unusual compared to the other candles.", + "bbox": [ + 555, + 200, + 846, + 234 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/01126e20ac9b600e067b4454d12d690d857d9b915c2df3b957d02dc35024e36f.jpg", + "image_caption": [ + "Figure 2. Industrial image anomaly reasoning results from GPT-4o [72] and our Anomaly-OV. The responses for fine-grained anomaly reasoning are highlighted, with the ground truth given for reference." + ], + "image_footnote": [], + "bbox": [ + 524, + 243, + 552, + 266 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Yes, there is an anomaly in the top left candle. It has a small hole or crack near the wick, which is not present in the other candles.", + "bbox": [ + 558, + 244, + 857, + 290 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "What is the cause for the anomaly?", + "bbox": [ + 611, + 299, + 859, + 311 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0853c89fa1b7fdb473555b96d7d47ac50e03bb39bccc3193e3b271babe46203d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 867, + 290, + 895, + 315 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ec51fc591b957718275465597f05c2ab9414421124c387006f6117f901285a47.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 313, + 552, + 337 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The anomaly could be due to manufacturing defects, such as improper molding or cooling, or it might result from external damage like a puncture.", + "bbox": [ + 557, + 316, + 857, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "downstream tasks, e.g., video anomaly detection [67, 94], affective computing [14, 29], and visual commonsense reasoning [109], revealing more interpretability.", + "bbox": [ + 511, + 439, + 906, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unsupervised Anomaly Detection. Due to the scarcity and difficulty of collecting anomalous data, researchers often focus on the unsupervised AD setting, which exclusively uses normal data to train an AD model. Earlier studies, such as reconstruction-based [63, 69, 99], student-teacher [18, 85, 102], and augmentation-based [46] approaches, assume a large amount of normal data is available. These traditional approaches are less practical when data are restricted or expensive, such as in the medical domain.", + "bbox": [ + 511, + 489, + 906, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Zero-Shot Anomaly Detection. Unlike unsupervised AD [35, 77] and few-shot AD [22, 28, 36, 51, 112], ZSAD models directly access the likelihood of abnormality for a given image without requiring data specific to the target object. Existing works [6, 19, 38, 110] accomplish ZSAD by comparing visual and textual features encoded by visual and text encoders of CLIP and constructing their positive (anomaly) and negative (normal) prompts in the format of:", + "bbox": [ + 511, + 633, + 908, + 753 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {P} ^ {+} = \\left[ V _ {1} \\right] \\left[ V _ {2} \\right] \\dots \\left[ V _ {n} \\right] [ o b j e c t ] \\\\ \\mathcal {P} ^ {-} = \\left[ W _ {1} \\right] \\left[ W _ {2} \\right] \\dots \\left[ W _ {n} \\right] [ o b j e c t ] \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 762, + 815, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $V_{i}$ and $W_{i}$ are handcrafted or learnable tokens, and object refers to the word object or the class name of the object. However, simply utilizing object to represent all kinds of objects cannot capture the class-awareness abnormality types. Also, for an intelligent visual assistant, the images should be totally blind to the user (object-agnostic).", + "bbox": [ + 511, + 809, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20371", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c9b0b96175b01f00b0c7a0ecc318ecfbc5b060ab17aea4571e60fc25539fed13.jpg", + "image_caption": [ + "Figure 3. Overview of the Anomaly-0V architecture. It consists of two training stages: (1) professional training for the anomaly expert, and (2) visual instruction tuning for anomaly detection and reasoning. Text and visual tokens are distinguished by different colors." + ], + "image_footnote": [], + "bbox": [ + 89, + 87, + 906, + 440 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 503, + 183, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 529, + 220, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training an MLLM from scratch demands extensive data and computational resources to align the visual and textual embedding spaces and develop robust instruction-following capabilities. Recent studies [23, 83, 93] reveal that pretrained MLLMs function as generalists, possessing a broad knowledge base but underperforming in specialized domains. Therefore, our goal is to introduce an auxiliary specialist or expert model designed to guide the generalist in selecting and utilizing critical visual tokens. This approach circumvents the need for large-scale pre-training while preserving the generalization capacity of the original model.", + "bbox": [ + 88, + 551, + 482, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We choose $LLaVA-OneVision$ [44] as our base MLLM because it is open-sourced and performs similarly to other commercial models. $LLaVA-OneVision$ follows the model architectures for LLaVA family [47, 57-59] and other generic MLLMs, which typically consist of three major components: Visual Encoder, Projector, and LLM. The visual encoder [73, 100] extracts the visual information from the raw images, the projector aligns the spaces of visual features with the word embedding, and the LLM is responsible for textual instruction processing and complex reasoning. Since the image resolution for CLIP pre-training is fixed, $LLaVA-OneVision$ leverages AnyRes with pooling strategy to scale", + "bbox": [ + 89, + 719, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "up the input raw image resolution. Specifically, the high-resolution images are divided into a prototyped number of crops, and the visual encoder independently processes the image crops before final spatial pooling.", + "bbox": [ + 511, + 505, + 906, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Architecture Overview", + "text_level": 1, + "bbox": [ + 511, + 575, + 725, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the same image-splitting strategy AnyRes as LLaVA-OneVision, the input high-resolution image is split into several crops, and the new image set can be written as:", + "bbox": [ + 511, + 597, + 906, + 643 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} = \\left\\{I _ {0}, I _ {1}, I _ {2}, \\dots , I _ {n - 1} \\right\\} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 652, + 906, + 670 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $I_0$ is the resized original image and $I_{j\\neq 0}$ refers to the image crops. As shown in Figure 3, the image set $\\mathcal{I}$ will be processed by the visual encoder $\\mathcal{F}_{\\theta}$ to generate the final visual features $\\{\\mathbf{v}_j^o\\}$ . Similar to AnomalyCLIP [110], we store the outputs for four selected layers in the ViT [21] to capture the image representations from different levels and apply four adapters to compress the feature dimension. Then, the extracted visual features can be written as:", + "bbox": [ + 511, + 680, + 908, + 800 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {j} ^ {i} = \\mathcal {F} _ {\\theta} ^ {i} (I _ {j}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 663, + 811, + 906, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $i$ denotes the $i$ -th level and $j$ refers to the index of corresponding image in $\\mathcal{I}$ . These multi-level features have been demonstrated to be effective in capturing fine-grained local semantics by recent works [6, 28, 110].", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20372", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f22e345a2ef3107a4d55a4c39a16623ba30cea442cceb34a735f0e07cee87875.jpg", + "image_caption": [ + "Figure 4. Simulation of visual anomaly inspection by humans." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 480, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The large-scale pre-trained CLIP models align the projection spaces of the textual and visual encoder. Therefore, the encoded image features already contain the class information required by ZSAD. To avoid human involvement in object classification and reduce the model complexity, we remove the heavy text encoder commonly utilized in existing works and let the visual model itself parse the information for suspicious classes or objects. Specifically, the output visual features for the original image $\\mathbf{v}_0^o$ are leveraged to provide the global description of the target object or regions in the look-back path. With the multi-level features and the global embeddings, the LTFM module is responsible for the recognition and localization of suspicious tokens.", + "bbox": [ + 88, + 203, + 483, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Drawing inspiration from human visual inspection, where suspicious objects or regions are identified and then inspected closely (see Figure 4), we design the VT selector module for aggregating (zooming in) the crucial visual tokens and explicitly assisting the LLM in distinguishing these tokens from many irrelevant ones when dealing with instructions regarding anomaly detection and reasoning. Additionally, the original visual features are preserved to maintain the generalization capability of the base model on regular instructions, such as Can you describe the content of the image?", + "bbox": [ + 89, + 400, + 483, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Look-Twice Feature Matching", + "text_level": 1, + "bbox": [ + 89, + 575, + 359, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the global object information $\\mathbf{v}_0^o$ provided by the lookback path, we generate the class-awareness abnormality description by merging $\\mathbf{v}_0^o$ with two learnable embeddings: $\\mathbf{e}^{+}\\in \\mathbb{R}^{D}$ and $\\mathbf{e}^{-}\\in \\mathbb{R}^{D}$ , where $+$ and $-$ indicate positive (anomalous) and negative (normal) patterns and $D$ is the embedding dimension. Specifically, a linear layer $\\mathcal{T}_i^o$ is applied along the token dimension to select and fuse useful tokens from $\\mathbf{v}_0^o$ , and then the fused vector will be concatenated with $\\mathbf{e}^{+}$ and $\\mathbf{e}^{-}$ independently and pass through two MLPs $\\{\\mathcal{G}_i^+, \\mathcal{G}_i^-\\}$ to generate the abnormality and normality descriptions $\\{\\mathbf{d}_i^+, \\mathbf{d}_i^-\\}$ , which can be represented by:", + "bbox": [ + 89, + 597, + 483, + 763 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\mathbf {d} _ {i} ^ {+}, \\mathbf {d} _ {i} ^ {-} \\right\\} = \\left\\{ \\begin{array}{l} \\mathcal {G} _ {i} ^ {+} \\left(\\mathbf {e} ^ {+} \\circ \\mathcal {T} _ {i} ^ {o} \\left(\\mathbf {v} _ {0} ^ {o}\\right)\\right) \\\\ \\mathcal {G} _ {i} ^ {-} \\left(\\mathbf {e} ^ {-} \\circ \\mathcal {T} _ {i} ^ {o} \\left(\\mathbf {v} _ {0} ^ {o}\\right)\\right) \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 775, + 483, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The visual features extracted from different levels of the ViT focus on different scales of semantics. Thus, the parameters of $\\mathcal{T}_i^o$ and $\\{\\mathcal{G}_i^+, \\mathcal{G}_i^-\\}$ should be independent for different levels, where $i$ indicate the level number.", + "bbox": [ + 89, + 824, + 483, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similar to the zero-shot classification mechanism of CLIP", + "bbox": [ + 109, + 885, + 482, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "models, we calculate the possibilities of each patch token in $\\mathbf{v}_j^i$ belonging to the anomalous patterns by combining cosine similarity and softmax operations:", + "bbox": [ + 511, + 90, + 905, + 136 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {m} _ {j} ^ {i} = \\frac {\\exp \\left(< \\mathbf {d} _ {i} ^ {+} , \\mathbf {v} _ {j} ^ {i} > / \\tau\\right)}{\\exp \\left(< \\mathbf {d} _ {i} ^ {+} , \\mathbf {v} _ {j} ^ {i} > / \\tau\\right) + \\exp \\left(< \\mathbf {d} _ {i} ^ {-} , \\mathbf {v} _ {j} ^ {i} > / \\tau\\right)} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 145, + 906, + 184 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{m}_j^i$ represents the significance map for visual tokens, $\\tau$ is the temperature hyperparameter, and $<,>$ refers to the cosine similarity operator. The patch weight in $\\mathbf{m}_j^i$ indicates the closeness of the corresponding visual token to the anomalous pattern. Then, all the maps are averaged to capture the token significances from low to high levels:", + "bbox": [ + 511, + 193, + 906, + 284 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {m} _ {j} = \\sum_ {i = 1} ^ {4} \\mathbf {m} _ {j} ^ {i} / 4 \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 292, + 906, + 334 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The visual features are leveraged twice in the forward and look-back paths, so this module is named by Look-Twice Feature Matching (LTFM), following the nature of two-step human visual inspection shown in Figure 4.", + "bbox": [ + 511, + 342, + 905, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Visual Token Selector", + "text_level": 1, + "bbox": [ + 511, + 411, + 715, + 425 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Under the image cropping strategy widely applied in recent MLLMs, there will be a large number of visual tokens for a high-resolution image, e.g., 7290 tokens for an image with $1152 \\times 1152$ resolution in LLaVA-OneVision. While these tokens provide rich visual details, the LLM is required to pick the most useful information when adapting to a specific task. When the LLM lacks enough knowledge in this domain, the token-picking process will become complicated. Thus, our solution is to introduce a specialist or expert who knows which token is crucial or not and assist the LLM in selecting and emphasizing (zooming in) the crucial tokens.", + "bbox": [ + 511, + 433, + 906, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the encoded visual tokens $\\{\\mathbf{v}_j^o\\}$ for each image crop in $\\mathcal{I}$ and the corresponding significance map $\\mathbf{m}_j$ , the suspicious tokens are emphasized by direct multiplication of the two tensors. Then, the normal tokens will be scaled to zero while the anomalous tokens will be maintained. After that, spatial average pooling $\\mathcal{P}$ is applied to reduce the number of tokens. This process can be written as:", + "bbox": [ + 511, + 599, + 905, + 705 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {q} _ {j} = \\mathcal {P} \\left(\\mathbf {v} _ {j} ^ {o} \\odot \\mathbf {m} _ {j}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 715, + 906, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{q}_j\\in \\mathbb{R}^{h\\times w\\times D}$ refers to the pooled query tokens. Empirically, setting $h = w = 2$ provides a better trade-off than other options. Then, a Q-Former $\\mathcal{Q}$ [49] is leveraged to aggregate the correlated tokens in the original output by forwarding $\\mathbf{q}_j$ as the query and $\\mathbf{v}_j^o$ as the key and value:", + "bbox": [ + 511, + 742, + 905, + 821 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {j} ^ {s} = \\mathcal {Q} \\left(\\mathbf {q} _ {j}, \\mathbf {v} _ {j} ^ {o}, \\mathbf {v} _ {j} ^ {o}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 829, + 906, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Visual Token Selector (VT Selector) serves as a tool for the anomaly expert to hand-pick visual tokens that contain the most suspicious semantics for a given image.", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20373", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/27870aa2e157a6e267299eafa2aaa3056f439cbc70d1a89ab877c999c695d5e7.jpg", + "image_caption": [ + "Figure 5. Composition of the instruction data in Anomaly-Instruct-125k. There are four main types of image samples: in-the-wild, industrial, medical, and $3D$ (in the format of multi-view images), covering most image anomaly detection tasks and enabling the possibility of a unified assistant for visual inspection. The reasoning words are highlighted in blue. For more information about dataset establishment, statistics, and the data collection pipeline, please refer to Section A1 in the supplementary." + ], + "image_footnote": [ + "Q: Do you notice any anomalies in the image? A: Yes, the anomaly in the image is a fine, dark crack visible on the ceramic cup, starting from the rim and extending downward.", + "Q: How can I prevent similar anomalies in the future? A: Avoid rapid temperature changes, handle the cup carefully to minimize impacts, and consider using cups made with thicker, more durable materials." + ], + "bbox": [ + 99, + 97, + 155, + 152 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1cae4e97724f842ce372d6f3700a1026cebd396a15bf77bfa7c68dac0e191126.jpg", + "image_caption": [], + "image_footnote": [ + "Q: Are there any anomalies in this cable image? A: Yes, the green wire is missing its internal copper conductors, unlike the blue and brown wires, which contain complete sets of bundled copper strands.", + "Q: What could be the reason behind this anomaly? A: The absence of copper in the green wire could be due to manufacturing defects such as failure to insert the strands or a mechanical issue where the strands were removed or misplaced during production or handling." + ], + "bbox": [ + 300, + 104, + 352, + 146 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/236f9ae2d43bd439000f023802369ecec417947d1e99b3582af1ac9171fb9000.jpg", + "image_caption": [], + "image_footnote": [ + "Q: What does the image show? A: The image shows a densely cellular tissue sample with features indicating potential metastatic breast cancer.", + "Q: What suggests that there might be malignant activity in the tissue sample? A: Malignant activity is suggested by irregularly shaped cells, hyperchromatic nuclei, and an increased nuclear-to-cytoplasmic ratio, indicative of abnormal and increased DNA content typical in cancerous cells." + ], + "bbox": [ + 506, + 104, + 560, + 146 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2e43f15b428805c298907b2430f159ad648f7a757bfd215f12ca344cd3fa05b1.jpg", + "image_caption": [], + "image_footnote": [ + "Q: What kind of object are you examining in the images? A: I'm examining a round cookie with a golden-brown surface and visible chocolate chips embedded across its surface. Q: Are there any significant anomalies detected on this cookie? A: Yes, there are two significant anomalies on the cookie: a central defect featuring a noticeable hole and an edge defect showing a gap or crack." + ], + "bbox": [ + 709, + 97, + 754, + 128 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Inference and Loss", + "text_level": 1, + "bbox": [ + 89, + 297, + 272, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Anomaly Prediction. In the traditional anomaly detection task, the model predicts the possibility of the image being abnormal. To achieve anomaly score prediction, we aggregate the anomaly information from all the image crops by an average operation weighted on the significance maps:", + "bbox": [ + 89, + 320, + 483, + 396 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {r} (\\mathcal {I}) = \\frac {\\sum_ {j , k} \\mathbf {v} _ {j} ^ {\\mathrm {s}} [ k ] \\cdot \\mathcal {P} (\\mathbf {m} _ {j}) [ k ]}{\\sum_ {j , k} \\mathcal {P} (\\mathbf {m} _ {j}) [ k ]} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 406, + 483, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{P}$ is the same spatial pooling in VT Selector and $\\mathbf{r}(\\mathcal{I})$ is a vector containing the global anomaly information for the entire image. Then, the anomaly expert can calculate the image-level abnormal possibility by parsing $\\mathbf{r}(\\mathcal{I})$ :", + "bbox": [ + 89, + 455, + 483, + 516 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {s c o r e} (\\mathcal {I}) = \\operatorname {S i g m o i d} \\left(\\mathcal {G} ^ {o} (\\mathbf {r} (\\mathcal {I}))\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 527, + 482, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{G}^o$ is an MLP for distinguishing normal and abnormal semantics. To handle the unbalanced sample distribution, we employ the balanced BCE loss as the professional training objective for the anomaly expert components.", + "bbox": [ + 89, + 556, + 483, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Text Generation. Instead of directly forwarding the concatenation of the original $\\{\\mathbf{v}_j^o\\}$ and the selected $\\{\\mathbf{r}(\\mathcal{I}),\\mathbf{v}_j^s\\}$ visual tokens into the LLM, we apply an indication prompt with suspicious feature: in the middle of the two series of tokens, which will highlight the selected tokens for LLM when handling anomaly-related instructions. This approach can be considered a form of prompt engineering in MLLMs. Besides, the is chosen from {highly, moderately, slightly} and is determined by score $(\\mathcal{I})$ and predefined thresholds $\\{\\mathbf{s}_{low},\\mathbf{s}_{high}\\}$ . When the input image $\\mathcal{I}$ has a high likelihood of anomaly, the LLM will place greater emphasis on the selected tokens; otherwise, these tokens will have less significance. The text generation is implemented by the original auto-regressive token prediction mechanism of LLM:", + "bbox": [ + 89, + 625, + 483, + 849 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np \\left(X _ {a} \\mid \\mathcal {I}, X _ {q}\\right) = \\prod_ {t = 1} ^ {L} p _ {\\theta} \\left(x _ {t} \\mid \\mathcal {I}, X _ {q, < t}, X _ {a, < t}\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 863, + 483, + 904 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $X_{a, < t}$ and $X_{q, < t}$ are the answer and instruction tokens from all prior turns before the current prediction token $x_{t}$ for a sequence of length $L$ . The entire model is parameterized by $\\theta$ and trained by the original language model cross-entropy loss for each predicted answer token $x_{t}$ .", + "bbox": [ + 511, + 299, + 906, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Dataset and Benchmark", + "text_level": 1, + "bbox": [ + 511, + 388, + 740, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The lack of multimodal instruction-following data for image anomaly detection and reasoning hinders the development of special intelligent assistants in this domain. Even though AnomalyGPT [28] introduces a prompt tuning dataset by simulating the anomalies, the scale of their dataset and the diversity of their instructions and answers are limited, only focusing on anomaly localization. To resolve the data scarcity issue, we establish the first large-scale instruction tuning dataset: Anomaly-Instruct-125k and the corresponding anomaly detection and reasoning benchmark: VisA-D&R.", + "bbox": [ + 511, + 414, + 906, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Anomaly-Instruct-125k", + "text_level": 1, + "bbox": [ + 511, + 575, + 730, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LLaVA [57] builds its instruction tuning dataset by leveraging the image caption and bounding boxes available in the COCO dataset[55] to prompt the text-only GPT-4. ShareGPT4V [8] provides a higher-quality dataset by directly prompting GPT-4V [71]. However, there is no image caption provided in existing anomaly detection datasets [1, 2], and no matter GPT-4V [71] or most recent GPT-4o [72] cannot accurately locate and describe the anomalies in the image without explicit human involvement.", + "bbox": [ + 511, + 598, + 906, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To resolve these issues, we design a new prompt pipeline for accurate anomaly description generation. Since most of the datasets contain annotations for anomaly types, we manually combine the class name and anomaly type, such as a [capsule] with [poke] on surface. If the anomaly masks are provided, we draw bounding boxes on the images to highlight the anomalous area. The short description and the image with (or w/o) bounding boxes are used to prompt GPT-4o to generate the detailed image and anomaly descriptions. Then, we employ an in-context learning strategy similar to LLaVA to create the instructions.", + "bbox": [ + 511, + 734, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20374", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Detection:", + "text_level": 1, + "bbox": [ + 104, + 95, + 174, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Q: Are there any defects for the object in the image? Please reply with 'Yes' or 'No'.", + "bbox": [ + 104, + 106, + 457, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Reasoning:", + "text_level": 1, + "bbox": [ + 106, + 136, + 174, + 147 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Q1: Do you observe any anomalies in the image?", + "Q2: Can you describe the anomalies you observed?", + "Q3: What is the potential cause for the anomalies?", + "Q4: How can such anomalies be prevented in the future?" + ], + "bbox": [ + 106, + 147, + 464, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 6. Prompt examples in VisA-D&R for detection and reasoning. The complex reasoning instructions are highlighted.", + "bbox": [ + 89, + 205, + 485, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For a unified visual inspection dataset, precise instruction data is collected from MVtec AD [2], the training set of BMAD [1], Anomaly-ShapeNet [50], Real3D-AD [60], and MVtec-3D AD [4], covering both 2D to 3D data across industry to medical domains. The 3D point cloud data are converted into 9 multi-view images, and the corresponding masks are rendered using predefined camera positions. However, the diversities and scales of these datasets are relatively limited, probably due to the difficulty of collecting anomaly images. To scale up the instruction data, we introduce an automatic anomaly data collection pipeline combining GPT4o [72] and Google Image Search [26] for image collection, data cleaning, and instruction generation. Finally, 72k in-the-wild images (named as WebAD) targeting anomaly detection are collected, significantly enriching our instruction dataset. Several samples from Anomaly-Instruct-125k are shown in Figure 5. The instructions are mainly in the format of multi-round conversations, covering anomaly detection and description in low-level reasoning and potential cause and future suggestions for complex understanding.", + "bbox": [ + 91, + 247, + 483, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.VisA-D&R", + "text_level": 1, + "bbox": [ + 89, + 560, + 212, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VisA [115] is a classic but challenging industrial anomaly detection dataset, providing fine-grained anomaly type and segmentation for each image. For evaluation of the anomaly detection and reasoning performance on existing and future methods, we select 10 classes from VisA and follow a similar data generation pipeline of Anomaly-Instruct-125k to create the benchmark. Differently, significant human effort has been invested in meticulously reviewing all generated images and anomaly descriptions. Wrong descriptions are picked out and re-annotated by humans before utilizing them for Q&A generation. Totally, the benchmark consists of 761 normal samples and 1000 anomalous ones.", + "bbox": [ + 89, + 582, + 482, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For evaluating detection performance, questions designed to elicit a one-word answer are used to prompt the MLLMs (Figure 6), with results quantified using Accuracy, Precision, Recall, and F1-score. We divide the reasoning performance into two parts: low-level reasoning that focuses on the description of visual defects or anomalies and complex reasoning requiring the MLLMs to provide the potential cause and future improvement strategies for the detected anomalies, where ROUGE-L [54], Sentence-BERT (SBERT) [75], and", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GPT-Score (GPT-4 as the judge [57]) are utilized to quantify the similarity between generated text and ground truth. Note that low-level reasoning is highly correlated to detection performance, while anomaly-type descriptions of low-level reasoning determine the output of complex reasoning.", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 179, + 645, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Training & Evaluation", + "text_level": 1, + "bbox": [ + 511, + 203, + 723, + 219 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "There are two independent training stages for Anomaly-OV. In Stage 1, the components of the anomaly expert are trained to obtain the token selection capability, targeting traditional ZSAD. This stage utilizes all of the data with anomaly labels in Anomaly-Instruct-125k. Similar to previous works [6, 110], when evaluating the model on the datasets contained in the training set, the corresponding datasets are replaced by VisA [115]. In Stage 2, the anomaly expert and visual encoder are frozen, while the projector and LLM are trainable. In addition to our instruction dataset, we sample around $350\\mathrm{k}$ data from the original training recipe of $LLaVA-OneVision$ to maintain the generalization ability. For more details on training, please refer to the supplementary.", + "bbox": [ + 511, + 224, + 906, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The ZSAD performance for the anomaly expert is evaluated on nine benchmarks, including MVTec AD [2], VisA [115], AITEX [81], ELPV [16], BTAD [68], and MPDD [39] for industrial inspection, and BrainMRI [40], HeadCT [41], and Br35H [30] for medical diagnosis. AUROC (Area Under the Receiver Operating Characteristic) is leveraged to quantify the image-level AD performance. For text-based anomaly detection, both normal and anomaly data are employed to assess the accuracy by examining if the generated text contains the word Yes. Differently, only anomaly data are utilized to prompt the MLLMs to determine their anomaly reasoning capabilities since the justifications of normality are similar for different models.", + "bbox": [ + 511, + 422, + 908, + 619 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5e5d27628c50ccc09f575834cbf0dd33c03975aa672e3fb46e9a21e1699bbc39.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMVTecVisAHeadCTBrainMRI
Full Model94.091.197.693.9
w/o. Look-back92.890.596.693.5
w/o. e+ & e-92.190.194.792.9
w/o. Q-Former91.789.992.895.1
w/o. WebAD88.588.991.293.4
", + "bbox": [ + 519, + 628, + 900, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Ablation study for the anomaly expert of Anomaly-OV. w/o. Look-back refers to the removal of $\\mathbf{v}_0^o$ in LTFM.", + "bbox": [ + 511, + 734, + 906, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Zero-Shot Anomaly Detection", + "text_level": 1, + "bbox": [ + 511, + 772, + 779, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 2, compared with existing methods, the anomaly expert of Anomaly-OV achieves significant image-level AUROC improvements on most of the ZSAD benchmarks, which demonstrates that the text encoder widely applied in existing models is not necessary. The success of our model mainly originates from the extra data of WebAD (Table 1), which enables the model to learn more generic", + "bbox": [ + 511, + 794, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20375", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6785ac6a944ad439782076c5c22e88f54c165fec2fc8715737663589f75f81fa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelIndustrial DefectsMedical AnomaliesAverage
MVTec ADVisAAITEXELPVBTADMPDDBrainMRIHeadCTBr35H
CLIP [73]74.166.471.059.234.554.373.956.578.463.1
CoOp [108]88.862.866.273.066.855.161.378.486.070.9
WinCLIP [38]91.878.873.074.068.263.692.690.080.579.2
APRIL-GAN [11]86.278.057.665.573.673.089.389.193.178.4
AnoVL [19]92.579.272.570.680.368.988.781.688.480.3
AnomalyCLIP [110]91.582.162.281.588.377.090.393.494.684.5
AdaCLIP [6]89.285.864.579.788.676.094.891.497.785.3
Ours94.091.172.083.089.081.793.997.695.588.6
", + "bbox": [ + 99, + 88, + 898, + 244 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "semantics for normality and abnormality from the data distribution in the absence of the text encoder. This observation also reveals that large-scale in-the-wild online data can benefit zero-shot performance in anomaly detection.", + "bbox": [ + 89, + 297, + 485, + 358 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While the Q-Former reduces the model performance on BrainMRI, it shows effectiveness on most benchmarks, indicating the importance of token aggregation. Similarly, the look-back information and two learnable embeddings are required for describing class-awareness abnormality and distinguishing positive and negative features, respectively. As previously discussed, the anomaly expert is responsible for selecting suspicious visual tokens for the LLM, and the significance maps shown in Figure 7 demonstrate the interpretable token selection mechanism. The high intensities are automatically distributed around the anomalous areas even without any supervision of the anomaly masks.", + "bbox": [ + 89, + 359, + 483, + 541 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/64f50cfa285d9f428edf3462c2cab1e9bdb560c5289d050d2012e58827b12991.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 104, + 551, + 222, + 621 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0981baf11f7c8786c1a1a03b1ed604d3cfd7543e706fbc85a5b2ca1336e324fb.jpg", + "image_caption": [ + "Figure 7. Visualization of the significance map on VisA samples." + ], + "image_footnote": [], + "bbox": [ + 106, + 623, + 220, + 691 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/67319f985b80c7763a079f5228d21f9acfc0ba797304cbc6bb0961b1e5991746.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 554, + 346, + 621 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fc0d1a80b23d087fe076abffd6415aab48435be08723deaadd71ed472a66b3e1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 623, + 346, + 691 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7c5d82ad91c401f62339a88ed2869ef5c6f00df1516237dd8d3ed6a69835e079.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 554, + 470, + 621 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/447de3945edc43d51dfcbac8faf9e4d3338c3b8041cb0ddd06aba3772e56e6b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 623, + 468, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Anomaly Detection & Reasoning", + "text_level": 1, + "bbox": [ + 89, + 727, + 377, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With the strong capabilities of the anomaly expert for zero-shot detection and suspicious token selection, Anomaly-OV accomplishes significant improvement in text-based anomaly detection and reasoning over other open-sourced generalist MLLMs, as shown in Table 4. Here are a few observations: i) While a larger language model cannot guarantee better detection performance, it always provides greater reasoning ability; ii) Most of the existing MLLMs present much lower recall than precision, indicating their insensitivity to visual anomalies; iii) GPT-4o shows stronger reasoning ability", + "bbox": [ + 89, + 750, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7d15dd20b305220bf22bf59e6d6ae8e9f20eaa8b777608b66d7da0991a4f4c48.jpg", + "table_caption": [ + "Table 2. Quantitative comparison of Image-level AUROC on different ZSAD methods (some of the results are borrowed from [6, 110, 114]). The best and the second-best results are bolded and underlined, respectively. Please refer to the supplementary for more detailed results." + ], + "table_footnote": [], + "table_body": "
PCB Example: Excessive Solder
UserAre there any anomalies visible in the image?
GPT-4oThe image shows an HC-SR04 ultrasonic sensor. There don’t appear to be any visible anomalies with the sensor itself. The components and pins seem intact and properly aligned.
LLaVA-OVThe image shows an ultrasonic sensor module, specifically the HC-SR04 model. There are no visible anomalies in the image; it appears to be a standard representation of this type of sensor module.
Ours-7BYes, there is a noticeable anomaly in the image: a short circuit on the PCB, indicated by a white, crusty residue near the ‘Echo’ pin.
", + "bbox": [ + 526, + 297, + 895, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Anomaly-OV presents more accurate anomaly detection.", + "bbox": [ + 513, + 566, + 901, + 582 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "compared to other open-sourced models. Table 3 and Table 5 provide the qualitative comparison of our Anomaly-OV with its base model LLaVA-OV-7B [44] and GPT-4o [72]. Both GPT-4o and LLaVA-OV show insensitivity to anomalous features and cannot accurately detect the anomaly in the image. Sometimes, GPT-4o knows the image is anomalous but fails to describe the anomalies precisely.", + "bbox": [ + 511, + 597, + 906, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We provide the fine-tuned version of the base model LLaVA-OV-0.5B on Anomaly-Instruct-125k, which presents much higher accuracy and more balanced precision and recall than its original version. This demonstrates the effectiveness of our instruction-tuning dataset. By integrating the anomaly expert with the base model, our Anomaly-OV-0.5B achieves 0.08 accuracy and 0.06 F1-score improvements in text-based anomaly detection and better reasoning capability in low-level and complex settings. Equipped with a larger language model, Anomaly-OV-7B provides the best detection performance among all the existing MLLMs and shows comparable reasoning ability with GPT-4o. Notably, we observe that the anomaly expert restricts the detection perfor", + "bbox": [ + 511, + 704, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20376", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7149edef65c33bd2790ceca3d01d4a3790cbbe64059fd625bcbe5d54447f6f39.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelAnomaly DetectionLow-level ReasoningComplex Reasoning
AccuracyPrecisionRecallF1-scoreROUGE-LSBERTGPT-ScoreSBERTGPT-Score
GPT-4V [71]0.680.900.490.550.160.653.310.775.64
GPT-4o [72]0.700.830.710.680.240.714.840.816.89
Qwen2-VL-2B [87]0.650.870.550.590.220.551.940.744.26
Qwen2-VL-7B [87]0.760.910.690.750.250.613.090.684.62
InternVL-2-8B [13]0.740.780.810.760.230.733.690.805.08
InternVL-2-26B [13]0.730.860.660.680.210.744.130.805.49
IXC-2.5-7B [101]0.720.880.630.670.210.582.450.775.14
LLaVA-OV-0.5B [44]0.540.700.190.280.200.632.540.814.34
LLaVA-OV-7B [44]0.710.950.560.630.240.663.570.795.44
LLaVA-OV-0.5B*0.710.770.840.760.310.703.690.825.31
Anomaly-OV-0.5B0.790.860.830.820.330.723.870.835.67
Anomaly-OV-7B0.790.830.860.830.340.734.260.846.34
", + "bbox": [ + 94, + 88, + 903, + 300 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Quantitative comparison of text-based anomaly detection and reasoning for MLLMs. Notably, the Accuracy and F1-score for the anomaly expert of Anomaly-OV can be calculated as $\\{0.78, 0.77\\}$ with threshold 0.5. * indicates the model is fine-tuned on our dataset.", + "bbox": [ + 89, + 310, + 906, + 339 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mance of Anomaly-OV. Therefore, the design of a stronger anomaly expert is suggested for future works.", + "bbox": [ + 89, + 354, + 483, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Macaroni Example: Yellowish Spot", + "bbox": [ + 106, + 398, + 307, + 412 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/88fd3c7885aa08856f47465b158091edfd8ec145c9f1f282665e988598f21199.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 417, + 416, + 489 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8fc53e512b4661142e28b5d2641742ab4d2b10a5f35a271445453bd0cc1f31d8.jpg", + "table_caption": [ + "User", + "Do you see any anomalies in the image?" + ], + "table_footnote": [], + "table_body": "
GPT-4oThe image shows four pieces of elbow macaroni on a green background. The anomaly is that the macaroni pieces are not whole; they are cut in half.
LLaVA-OVThe image shows four pieces of pasta, specifically macaroni shells, arranged on a green textured surface. The pasta appears to be uniformly colored and shaped, with no visible defects or anomalies.
Ours-7BYes, there is an anomaly in the image. The bottom right pasta piece has a noticeable yellowish discoloration on its surface.
", + "bbox": [ + 99, + 507, + 470, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Anomaly-OV presents more precise anomaly reasoning.", + "bbox": [ + 91, + 655, + 475, + 670 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Extension", + "text_level": 1, + "bbox": [ + 89, + 674, + 204, + 689 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With the generalization and multi-image processing capabilities of MLLMs, it is possible to build a unified assistant for visual inspection. Table 6 demonstrates the comprehensive knowledge of Anomaly-OV (without using Anomaly-ShapeNet [50] for training) on 3D and medical (testing set of BMAD [1]) AD & reasoning. More data, benchmarks, and investigation on a unified model are meaningful.", + "bbox": [ + 89, + 696, + 485, + 803 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 815, + 209, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we establish the first large-scale visual instruction tuning dataset, Anomaly-Instruct-125k, and the corresponding benchmark, VisA-D&R, to address the data scarcity issue for visual anomaly detection and reasoning.", + "bbox": [ + 89, + 839, + 485, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3D Example: Bulge", + "bbox": [ + 535, + 354, + 648, + 368 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/04a92329325abad74a691dc739a11fb8b70be3374afe8a3918381a73b2a06cd4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 372, + 710, + 445 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/56fbef7823f2c266934b4fbe1b91fbc04875b8ebd7da10dcf49856622d26e33d.jpg", + "image_caption": [ + "···" + ], + "image_footnote": [], + "bbox": [ + 722, + 372, + 833, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User", + "bbox": [ + 537, + 448, + 565, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Is there any anomaly present in the bowl?", + "bbox": [ + 598, + 448, + 818, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ours-7B", + "bbox": [ + 535, + 465, + 584, + 476 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Yes, there is a bulge on the side of the bowl, which disrupts its otherwise smooth and symmetrical surface.", + "bbox": [ + 596, + 465, + 879, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Medical Example: Pneumonia", + "bbox": [ + 537, + 494, + 709, + 507 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/02534aa62c713e51f9815a7cd628b8ba19ef8c12742b7262babaea7aca80ebc9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 511, + 821, + 584 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User", + "bbox": [ + 537, + 587, + 565, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "What is the most likely diagnosis for this patient?", + "bbox": [ + 596, + 587, + 859, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ours-7B", + "bbox": [ + 535, + 603, + 584, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Based on the chest X-ray, the most likely diagnosis for this patient is pneumonia. The presence of bilateral patchy opacities in the lower lung fields is a common finding in pneumonia, which is an infection that causes inflammation in the air sacs of the lungs.", + "bbox": [ + 596, + 603, + 880, + 665 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User", + "bbox": [ + 537, + 669, + 586, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "What should the patient do next?", + "bbox": [ + 596, + 669, + 771, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ours-7B", + "bbox": [ + 535, + 686, + 584, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The patient should consult a healthcare professional for further evaluation and appropriate treatment. This may include antibiotics to treat the infection, supportive care such as oxygen therapy, and monitoring of the patient's condition.", + "bbox": [ + 596, + 686, + 879, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Extension to 3D and medical AD & reasoning.", + "bbox": [ + 540, + 762, + 872, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Then, a specialist MLLM, Anomaly-OV, targeting visual inspection is introduced to serve as the baseline in this domain. Anomaly-OV leverages an anomaly expert to assist the LLM with suspicious visual token selection and presents significant improvements on both traditional ZSAD and text-based anomaly detection and reasoning tasks over existing methods. Extension to 3D and medical domains is demonstrated.", + "bbox": [ + 511, + 794, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20377", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jinan Bao, Hanshi Sun, Hanqiu Deng, Yinsheng He, Zhaoxiang Zhang, and Xingyu Li. Bmad: Benchmarks for medical anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4042-4053, 2024. 1, 5, 6, 8, 14", + "[2] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592-9600, 2019. 1, 5, 6, 14, 15", + "[3] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4183–4192, 2020. 1", + "[4] Paul Bergmann, Xin Jin, David Sattlegger, and Carsten Steger. The mvtec 3d-ad dataset for unsupervised 3d anomaly detection and localization. In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications. SCITEPRESS - Science and Technology Publications, 2022. 6", + "[5] Tri Cao, Jiawen Zhu, and Guansong Pang. Anomaly detection under distribution shift. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6511-6523, 2023. 1", + "[6] Yunkang Cao, Jiangning Zhang, Luca Frittoli, Yuqi Cheng, Weiming Shen, and Giacomo Boracchi. Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly detection. In European Conference on Computer Vision, pages 55-72. Springer, 2025. 1, 2, 3, 6, 7", + "[7] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. 1", + "[8] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions, 2023. 5, 14", + "[9] Qiyu Chen, Huiyuan Luo, Chengkan Lv, and Zhengtao Zhang. A unified anomaly synthesis strategy with gradient ascent for industrial anomaly detection and localization. arXiv preprint arXiv:2407.09359, 2024. 1", + "[10] Runnan Chen, Youquan Liu, Lingdong Kong, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, Yu Qiao, and Wenping Wang. Clip2scene: Towards label-efficient 3d scene understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7030, 2023. 2", + "[11] Xuhai Chen, Yue Han, and Jiangning Zhang. April-gan: A zero-/few-shot anomaly classification and segmentation method for cvpr 2023 vand workshop challenge tracks 1&2: 1st place on zero-shot ad and 4th place on few-shot ad, 2023. 2, 7" + ], + "bbox": [ + 99, + 114, + 485, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Yuanhong Chen, Yu Tian, Guansong Pang, and Gustavo Carneiro. Deep one-class classification via interpolated gaussian descriptor. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 383-392, 2022. 1", + "[13] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, Ji Ma, Jiaqi Wang, Xiaoyi Dong, Hang Yan, Hewei Guo, Conghui He, Botian Shi, Zhenjiang Jin, Chao Xu, Bin Wang, Xingjian Wei, Wei Li, Wenjian Zhang, Bo Zhang, Pinlong Cai, Licheng Wen, Xiangchao Yan, Min Dou, Lewei Lu, Xizhou Zhu, Tong Lu, Dahua Lin, Yu Qiao, Jifeng Dai, and Wenhai Wang. How far are we to gpt-4v? closing the gap to commercial multimodal models with opensource suites, 2024. 8", + "[14] Zebang Cheng, Zhi-Qi Cheng, Jun-Yan He, Jingdong Sun, Kai Wang, Yuxiang Lin, Zheng Lian, Xiaojiang Peng, and Alexander Hauptmann. Emotion-llama: Multimodal emotion recognition and reasoning with instruction tuning. In Conference on Neural Information Processing Systems, 2024. 1, 2", + "[15] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2", + "[16] Sergiu Deitsch, Vincent Christlein, Stephan Berger, Claudia Buerhop-Lutz, Andreas Maier, Florian Gallwitz, and Christian Riess. Automatic classification of defective photovoltaic module cells in electroluminescence images. Solar Energy, 185:455-468, 2019. 6", + "[17] Chenghao Deng, Haote Xu, Xiaolu Chen, Haodi Xu, Xiaotong Tu, Xinghao Ding, and Yue Huang. Simclip: Refining image-text alignment with simple prompts for zero-/few-shot anomaly detection. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 1761-1770, 2024. 1", + "[18] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In IEEE/CVF conference on computer vision and pattern recognition, 2022. 2", + "[19] Hanqiu Deng, Zhaoxiang Zhang, Jinan Bao, and Xingyu Li. Anovl: Adapting vision-language models for unified zero-shot anomaly localization. arXiv preprint arXiv:2308.15939, 2023. 1, 2, 7", + "[20] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. 2", + "[21] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 3", + "[22] Zheng Fang, Xiaoyang Wang, Haocheng Li, Jiejie Liu, Qiugui Hu, and Jimin Xiao. Fastrecon: Few-shot industrial" + ], + "bbox": [ + 522, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20378", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "anomaly detection via fast feature reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17481-17490, 2023. 2", + "[23] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J Black. Chatpose: Chatting about 3d human pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2093-2103, 2024. 3", + "[24] Tharindu Fernando, Harshala Gammulle, Simon Denman, Sridha Sridharan, and Clinton Fookes. Deep learning for medical anomaly detection-a survey. ACM Computing Surveys (CSUR), 54(7):1-37, 2021. 1", + "[25] Matic Fučka, Vitjan Zavrtanik, and Danijel Skočaj. Transfusion-a transparency-based diffusion model for anomaly detection. In European conference on computer vision, pages 91-108. Springer, 2025. 1", + "[26] Google. Google-images-search 1.4.7, 2024. https://pypi.org/project/Google-Images-Search.6, 14", + "[27] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Hao Li, Ming Tang, and Jinqiao Wang. Filo: Zero-shot anomaly detection by fine-grained description and high-quality localization. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2041-2049, 2024. 1", + "[28] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. Anomalygpt: Detecting industrial anomalies using large vision-language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1932-1940, 2024. 1, 2, 3, 5", + "[29] Yuxiang Guo, Faizan Siddiqui, Yang Zhao, Rama Chellappa, and Shao-Yuan Lo. Stimuvar: Spatiotemporal stimuli-aware video affective reasoning with multimodal large language models. arXiv preprint arXiv:2409.00304, 2024. 1, 2", + "[30] Ahmed Hamada. Br35h: Brain tumor detection 2020, 2020. 6", + "[31] Changhee Han, Leonardo Rundo, Kohei Murao, Tomoyuki Noguchi, Yuki Shimahara, Zoltan Adam Milacski, Saori Koshino, Evis Sala, Hideki Nakayama, and Shin'ichi Satoh. Madgan: Unsupervised medical anomaly detection gan using multiple adjacent brain mri slice reconstruction. BMC bioinformatics, 22:1-20, 2021. 1", + "[32] Liren He, Zhengkai Jiang, Jinlong Peng, Liang Liu, Qianggang Du, Xiaobin Hu, Wenbing Zhu, Mingmin Chi, Yabiao Wang, and Chengjie Wang. Learning unified reference representation for unsupervised multi-class anomaly detection. arXiv preprint arXiv:2403.11561, 2024. 1", + "[33] Chih-Hui Ho, Kuan-Chuan Peng, and Nuno Vasconcelos. Long-tailed anomaly detection with learnable class names. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12435-12446, 2024.", + "[34] Jinlei Hou, Yingying Zhang, Qiaoyong Zhong, Di Xie, Shiliang Pu, and Hong Zhou. Divide-and-assemble: Learning block-wise memory for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8791-8800, 2021. 1", + "[35] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based" + ], + "bbox": [ + 99, + 92, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "few-shot anomaly detection. In European Conference on Computer Vision, pages 303-319. Springer, 2022. 1, 2", + "[36] Chaoqin Huang, Aofan Jiang, Jinghao Feng, Ya Zhang, Xin chao Wang, and Yanfeng Wang. Adapting visual-language models for generalizable anomaly detection in medical images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11375-11385, 2024. 1, 2", + "[37] Brian KS Isaac-Medina, Yona Falinie A Gaus, Neelanjan Bhowmik, and Toby P Breckon. Towards open-world object-based anomaly detection via self-supervised outlier synthesis. In European Conference on Computer Vision (ECCV), 2024. 1", + "[38] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19606-19616, 2023. 1, 2, 7", + "[39] Stepan Jezek, Martin Jonak, Radim Burget, Pavel Dvorak, and Milos Skotak. Deep learning-based defect detection of metal parts: evaluating current methods in complex conditions. In 2021 13th International congress on ultra modern telecommunications and control systems and workshops (ICUMT), pages 66-71. IEEE, 2021. 6", + "[40] Pranita Balaji Kanade and PP Gumaste. Brain tumor detection using mri images. *Brain*, 3(2):146-150, 2015. 6", + "[41] Felipe Campos Kitamura. Head ct - hemorrhage, 2018. 6", + "[42] Mingyu Lee and Jongwon Choi. Text-guided variational image generation for industrial anomaly detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26519-26528, 2024. 1", + "[43] Aodong Li, Chen Qiu, Marius Kloft, Padhraic Smyth, Maja Rudolph, and Stephan Mandt. Zero-shot anomaly detection via batch normalization. Advances in Neural Information Processing Systems, 36, 2024. 1", + "[44] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer, 2024. 1, 3, 7, 8, 15", + "[45] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. LLaVA-med: Training a large language-and-vision assistant for biomedicine in one day. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 1, 15", + "[46] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In IEEE/CVF conference on computer vision and pattern recognition, 2021. 2", + "[47] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava-last-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 1, 3", + "[48] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Interna-" + ], + "bbox": [ + 522, + 92, + 908, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20379", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tional conference on machine learning, pages 12888-12900. \nPMLR, 2022. 1, 2", + "[49] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2, 4", + "[50] Wenqiao Li, Xiaohao Xu, Yao Gu, Bozhong Zheng, Shenghua Gao, and Yingna Wu. Towards scalable 3d anomaly detection and localization: A benchmark via 3d anomaly synthesis and a self-supervised learning network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22207-22216, 2024. 6, 8", + "[51] Xiaofan Li, Zhizhong Zhang, Xin Tan, Chengwei Chen, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Prompt: Learning prompts with only normal samples for few-shot anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16838-16848, 2024. 1, 2", + "[52] Yiting Li, Adam Goodge, Fayao Liu, and Chuan-Sheng Foo. Promptad: Zero-shot anomaly detection using text prompts. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1093-1102, 2024. 1", + "[53] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 2", + "[54] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6", + "[55] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5", + "[56] Yuqi Lin, Minghao Chen, Wenxiao Wang, Boxi Wu, Ke Li, Binbin Lin, Haifeng Liu, and Xiaofei He. Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15305-15314, 2023. 2", + "[57] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2, 3, 5, 6, 16", + "[58] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1", + "[59] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024.3", + "[60] Jiaqi Liu, Guoyang Xie, ruitao chen, Xinpeng Li, Jinbao Wang, Yong Liu, Chengjie Wang, and Feng Zheng. Real3d" + ], + "bbox": [ + 99, + 92, + 485, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "AD: A dataset of point cloud anomaly detection. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 6", + "[61] Jie Liu, Yixiao Zhang, Jie-Neng Chen, Junfei Xiao, Yongyi Lu, Bennett A Landman, Yixuan Yuan, Alan Yuille, Yucheng Tang, and Zongwei Zhou. Clip-driven universal model for organ segmentation and tumor detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21152-21164, 2023. 2", + "[62] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach, 2019. 2", + "[63] Shao-Yuan Lo, Poojan Oza, and Vishal M Patel. Adversarially robust one-class novelty detection. In IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2", + "[64] Ilya Loshchilov and Frank Hutter. SGDR: Stochastic gradient descent with warm restarts. In International Conference on Learning Representations, 2017. 15", + "[65] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 15", + "[66] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 2", + "[67] Hui Lv and Qianru Sun. Video anomaly detection and explanation via large language models. arXiv preprint arXiv:2401.05702, 2024. 1, 2", + "[68] Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. Vt-adt: A vision transformer network for image anomaly detection and localization. In 2021 IEEE 30th International Symposium on Industrial Electronics (ISIE), pages 01-06. IEEE, 2021. 6", + "[69] Shancong Mou, Xiaoyi Gu, Meng Cao, Haoping Bai, Ping Huang, Jiulong Shan, and Jianjun Shi. RGI: robust GAN-inversion for mask-free image inpainting and unsupervised pixel-wise anomaly detection. In The Eleventh International Conference on Learning Representations, 2023. 1, 2", + "[70] Ming Nie, Renyuan Peng, Chunwei Wang, Xinyue Cai, Jianhua Han, Hang Xu, and Li Zhang. Reason2drive: Towards interpretable and chain-based reasoning for autonomous driving. In European Conference on Computer Vision, pages 292-308. Springer, 2025. 1", + "[71] OpenAI. Gpt-4v(isdiction) system card, 2023. https://openai.com/index/gpt-4v-system-card.5,8", + "[72] OpenAI. Gpt-4o system card, 2024. https://openai.com/index/gpt-4o-system-card. 2, 5, 6, 7, 8, 14, 15", + "[73] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 7, 14" + ], + "bbox": [ + 524, + 92, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20380", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[74] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer, 2023. 2", + "[75] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks, 2019. 6", + "[76] Tal Reiss and Yedid Hoshen. Mean-shifted contrastive loss for anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2155-2162, 2023. 1", + "[77] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318-14328, 2022. 1, 2", + "[78] Fumiaki Sato, Ryo Hachiuma, and Taiki Sekii. Prompt-guided zero-shot anomaly action recognition using pretrained deep skeleton features. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6471-6480, 2023. 1", + "[79] Eli Schwartz, Assaf Arbelle, Leonid Karlinsky, Sivan Harary, Florian Scheidegger, Sivan Doveh, and Raja Giryes. Mae-day: Mae for few-and zero-shot anomaly-detection. Computer Vision and Image Understanding, 241:103958, 2024. 1", + "[80] Pierre Sermanet, Tianli Ding, Jeffrey Zhao, Fei Xia, Debidatta Dwibedi, Keerthana Gopalakrishnan, Christine Chan, Gabriel Dulac-Arnold, Sharath Maddineni, Nikhil J Joshi, et al. Robovqa: Multimodal long-horizon reasoning for robotics. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 645-652. IEEE, 2024. 1", + "[81] Javier Silvestre-Blanes, Teresa Albero-Albero, Ignacio Miralles, Rubén Pérez-Llorens, and Jorge Moreno. A public fabric database for defect detection methods and results. Autex Research Journal, 19(4):363–374, 2019. 6", + "[82] Luc PJ Sträter, Mohammadreza Salehi, Efstratios Gavves, Cees GM Snoek, and Yuki M Asano. Generalad: Anomaly detection across domains by attending to distorted features. arXiv preprint arXiv:2407.12427, 2024. 1", + "[83] Haomiao Sun, Mingjie He, Tianheng Lian, Hu Han, and Shiguang Shan. Face-mllm: A large face perception model, 2024. 3", + "[84] Jiaqi Tang, Hao Lu, Xiaogang Xu, Ruizheng Wu, Sixing Hu, Tong Zhang, Tsz Wa Cheng, Ming Ge, Ying-Cong Chen, and Fugee Tsung. An incremental unified framework for small defect inspection. In European Conference on Computer Vision, pages 307-324. Springer, 2025. 1", + "[85] Tran Dinh Tien, Anh Tuan Nguyen, Nguyen Hoang Tran, Ta Duc Huy, Soan Duong, Chanh D Tr Nguyen, and Steven QH Truong. Revisiting reverse distillation for anomaly detection. In IEEE/CVF conference on computer vision and pattern recognition, 2023. 2", + "[86] Hualiang Wang, Yi Li, Huifeng Yao, and Xiaomeng Li. Clipn for zero-shot ood detection: Teaching clip to say no. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1802-1812, 2023. 2" + ], + "bbox": [ + 99, + 92, + 483, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[87] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution, 2024. 8", + "[88] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. In Conference on Neural Information Processing Systems, 2022. 2", + "[89] Qi Wei, Yinhao Ren, Rui Hou, Bibo Shi, Joseph Y Lo, and Lawrence Carin. Anomaly detection for medical images based on a one-class classification. In Medical Imaging 2018: Computer-Aided Diagnosis, pages 375-380. SPIE, 2018. 1", + "[90] Julia Wolleb, Florentin Bieder, Robin Sandkühler, and Philippe C Cattin. Diffusion models for medical anomaly detection. In International Conference on Medical image computing and computer-assisted intervention, pages 35-45. Springer, 2022. 1", + "[91] Binzhu Xie, Sicheng Zhang, Zitang Zhou, Bo Li, Yuanhan Zhang, Jack Hessel, Jingkang Yang, and Ziwei Liu. Funqa: Towards surprising video comprehension. In European Conference on Computer Vision, pages 39-57. Springer, 2025. 1", + "[92] Guoyang Xie, Jinbao Wang, Jiaqi Liu, Yaochu Jin, and Feng Zheng. Pushing the limits of fewshot anomaly detection in industry vision: Graphcore. In The Eleventh International Conference on Learning Representations, 2023. 1", + "[93] Hongxia Xie, Chu-Jun Peng, Yu-Wen Tseng, Hung-Jen Chen, Chan-Feng Hsu, Hong-Han Shuai, and Wen-Huang Cheng. Emovit: Revolutionizing emotion insights with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26596-26605, 2024. 3", + "[94] Yuchen Yang, Kwonjoon Lee, Behzad Dariush, Yinzhi Cao, and Shao-Yuan Lo. Follow the rules: reasoning for video anomaly detection with large language models. arXiv preprint arXiv:2407.10299, 2024. 1, 2", + "[95] Hang Yao, Ming Liu, Haolin Wang, Zhicun Yin, Zifei Yan, Xiaopeng Hong, and Wangmeng Zuo. Glad: Towards better reconstruction with global and local adaptive diffusion models for unsupervised anomaly detection. arXiv preprint arXiv:2406.07487, 2024. 1", + "[96] Huanjin Yao, Wenhao Wu, Taojiannan Yang, YuXin Song, Mengxi Zhang, Haocheng Feng, Yifan Sun, Zhiheng Li, Wanli Ouyang, and Jingdong Wang. Dense connector for mllms, 2024. 17", + "[97] Xincheng Yao, Ruoqi Li, Zefeng Qian, Lu Wang, and Chongyang Zhang. Hierarchical gaussian mixture normalizing flow modeling for unified anomaly detection. arXiv preprint arXiv:2403.13349, 2024. 1", + "[98] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. In Advances in Neural Information Processing Systems, 2022. 1" + ], + "bbox": [ + 522, + 92, + 906, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "20381", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[99] Vitjan Zavrtanik, Matej Kristan, and Danijel Skoçaj. Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In IEEE/CVF international conference on computer vision, 2021. 2", + "[100] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3", + "[101] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, Songyang Zhang, Wenwei Zhang, Yining Li, Yang Gao, Peng Sun, Xinyue Zhang, Wei Li, Jingwen Li, Wenhai Wang, Hang Yan, Conghui He, Xingcheng Zhang, Kai Chen, Jifeng Dai, Yu Qiao, Dahua Lin, and Jiaqi Wang. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output, 2024. 8", + "[102] Xuan Zhang, Shiyu Li, Xi Li, Ping Huang, Jiulong Shan, and Ting Chen. Destseg: Segmentation guided denoising studentteacher for anomaly detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2", + "[103] Ximiao Zhang, Min Xu, Dehui Qiu, Ruixin Yan, Ning Lang, and Xiuzhuang Zhou. Mediclip: Adapting clip for few-shot medical image anomaly detection. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 458-468. Springer, 2024. 1", + "[104] Ximiao Zhang, Min Xu, and Xiuzhuang Zhou. Realnet: A feature selection network with realistic synthetic anomaly for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16699-16708, 2024. 1", + "[105] Zhuosheng Zhang, Aston Zhang, Mu Li, hai zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. Transactions on Machine Learning Research, 2024. 2", + "[106] He Zhao, Yuexiang Li, Nanjun He, Kai Ma, Leyuan Fang, Huiqi Li, and Yefeng Zheng. Anomaly detection for medical images using self-supervised and translation-consistent features. IEEE Transactions on Medical Imaging, 40(12): 3641-3651, 2021. 1", + "[107] Gengze Zhou, Yicong Hong, and Qi Wu. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7641-7649, 2024. 1", + "[108] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 7", + "[109] Kaiwen Zhou, Kwonjoon Lee, Teruhisa Mitsu, and Xin Eric Wang. Vicor: Bridging visual understanding and commonsense reasoning with large language models. In Findings of the Association for Computational Linguistics, 202'. 1, 2", + "[110] Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. AnomalyCLIP: Object-agnostic prompt learning for zero-shot anomaly detection. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 6, 7", + "[111] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. MiniGPT-4: Enhancing vision-language" + ], + "bbox": [ + 91, + 92, + 485, + 902 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "understanding with advanced large language models. In The Twelfth International Conference on Learning Representations, 2024. 1, 2", + "[112] Jiawen Zhu and Guansong Pang. Toward generalist anomaly detection via in-context residual learning with few-shot sample prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17826-17836, 2024. 1, 2", + "[113] Jiaqi Zhu, Shaofeng Cai, Fang Deng, Beng Chin Ooi, and Junran Wu. Do llms understand visual anomalies? uncovering llm's capabilities in zero-shot anomaly detection. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 48-57, 2024. 1", + "[114] Jiawen Zhu, Yew-Soon Ong, Chunhua Shen, and Guansong Pang. Fine-grained abnormality prompt learning for zero-shot anomaly detection, 2024. 7", + "[115] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In European Conference on Computer Vision, pages 392-408. Springer, 2022. 6, 15" + ], + "bbox": [ + 516, + 92, + 906, + 388 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "20382", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2025/Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models/9698d3f1-26f6-4c2e-aef0-60a71f074ba9_model.json b/2025/Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models/9698d3f1-26f6-4c2e-aef0-60a71f074ba9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fb1fb0f75854dddc69f1ac7138f529d7aab889e4 --- /dev/null +++ b/2025/Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models/9698d3f1-26f6-4c2e-aef0-60a71f074ba9_model.json @@ -0,0 +1,3548 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.225, + 0.131, + 0.777, + 0.177 + ], + "angle": 0, + "content": "Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.203, + 0.816, + 0.24 + ], + "angle": 0, + "content": "Jiacong Xu\\(^{1*}\\) Shao-Yuan Lo\\(^{2}\\) Bardia Safaei\\(^{1}\\) Vishal M. Patel\\(^{1}\\) Isht Dwivedi\\(^{2}\\) \n\\(^{1}\\)Johns Hopkins University \\(^{2}\\)Honda Research Institute USA" + }, + { + "type": "text", + "bbox": [ + 0.166, + 0.242, + 0.828, + 0.259 + ], + "angle": 0, + "content": "{jxu155, bsafaei1, vpatel36}@jhu.edu {shao-yuan.lo, idwivedi}@honda-ri.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.323, + 0.486, + 0.701 + ], + "angle": 0, + "content": "Zero-Shot Anomaly Detection (ZSAD) is an emerging AD paradigm. Unlike the traditional unsupervised AD setting that requires a large number of normal samples to train a model, ZSAD is more practical for handling data-restricted real-world scenarios. Recently, Multimodal Large Language Models (MLLMs) have shown revolutionary reasoning capabilities in various vision tasks. However, the reasoning of image abnormalities remains underexplored due to the lack of corresponding datasets and benchmarks. To facilitate research in AD & reasoning, we establish the first visual instruction tuning dataset, Anomaly-Instruct-125k, and the evaluation benchmark, VisA-D&R. Through investigation with our benchmark, we reveal that current MLLMs like GPT-4o cannot accurately detect and describe fine-grained anomalous details in images. To address this, we propose Anomaly-OneVision (Anomaly-OV), the first specialist visual assistant for ZSAD and reasoning. Inspired by human behavior in visual inspection, Anomaly-OV leverages a Look-Twice Feature Matching (LTFM) mechanism to adaptively select and emphasize abnormal visual tokens. Extensive experiments demonstrate that Anomaly-OV achieves significant improvements over advanced generalist models in both detection and reasoning. Extensions to medical and 3D AD are provided for future study. The link to our project page: https://xujiacong.github.io/Anomaly-OV/" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.731, + 0.222, + 0.746 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.756, + 0.486, + 0.879 + ], + "angle": 0, + "content": "Visual Anomaly Detection (AD) is a well-established task in computer vision, extensively applied in scenarios such as industrial defect inspection [2, 3, 5, 12, 35, 69, 76, 77, 92, 98] and medical image diagnosis [1, 24, 31, 36, 89, 90, 103, 106]. In the traditional unsupervised AD (a.k.a. one-class AD) setting, models learn the distribution of normal visual features from normal samples and are required to identify anomaly samples during inference. While recent advance" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.291, + 0.907, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.587, + 0.909, + 0.658 + ], + "angle": 0, + "content": "Figure 1. Visualization of the image-level AUROC comparison between our Anomaly-OV and current state-of-the-art ZSAD methods (WinCLIP [38], AnoVL [19], AnomalyCLIP [110], AdaCLIP [6]). Notably, our zero-shot performance on VisA even surpasses most recent advances in the few-shot setting [28, 51, 112]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.909, + 0.838 + ], + "angle": 0, + "content": "ments [9, 25, 32–34, 37, 42, 82, 84, 95, 97, 104] have significantly improved the detection performance, these approaches assume the availability of a substantial number of normal samples. However, this assumption becomes impractical in certain scenarios due to strict data privacy policies and the significant human effort required for data classification, sometimes involving experts or specialists. Therefore, Zero-Shot Anomaly Detection (ZSAD) is emerging as a popular research direction, leading to the development of many innovative methods [6, 17, 27, 38, 43, 52, 78, 79, 110, 113]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Recent advances in Multimodal Large Language Models (MLLMs) [7, 15, 44, 45, 47, 48, 57, 58, 111] have shown revolutionary reasoning capabilities in various vision tasks [14, 29, 67, 70, 80, 91, 94, 107, 109]. However, the rea" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.473, + 0.9 + ], + "angle": 0, + "content": "*Most of this work was done when J. Xu was an intern at HRI-USA." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20370" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.272 + ], + "angle": 0, + "content": "soning of image abnormalities has not been explored due to the challenges of collecting large-scale datasets and establishing benchmarks. Existing methods simply predict the likelihood of an anomaly without providing rationales [6, 11, 19, 38, 110]. In contrast, for better interpretability, robustness, and trustworthiness, people would expect models to explain why an image is considered anomalous and provide visual evidence. Interestingly, we find that recent advanced MLLMs, such as GPT-4o [72], fall short in AD & reasoning. As shown in Figure 2, while the detection is correct, the explanation from GPT-4o lacks accuracy, indicating a gap in a comprehensive understanding of the anomaly." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.486, + 0.499 + ], + "angle": 0, + "content": "To expedite research in AD & reasoning, we establish the first visual instruction tuning dataset, Anomaly-Instruct-125k, and the evaluation benchmark, VisA-D&R, through intensive human efforts. After evaluating current generalist MLLMs, we observe that these models fail to accurately detect and describe fine-grained anomalous details in images. To address this, we propose Anomaly-OneVision (Anomaly-OV), the first specialist visual assistant for ZSAD and reasoning. Unlike existing ZSAD methods [6, 11, 19, 38, 110], Anomaly-OV directly learns object-awareness abnormality embeddings in feature space using only the visual encoder. Inspired by human behavior in visual inspection, Anomaly-OV employs a Look-Twice Feature Matching (LTFM) mechanism to assist its LLM in adaptively selecting and emphasizing the most suspicious abnormal visual tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.5, + 0.485, + 0.65 + ], + "angle": 0, + "content": "Extensive experiments demonstrate that Anomaly-OV achieves significant improvements over advanced generalist models in both detection and reasoning. Extended results of Anomaly-OV, from applications in industrial defect detection to 3D inspection and medical image diagnosis, are provided for future study. With precise descriptions and rationales of visual anomalies, our model can infer potential causes (see Figure 2), assess current impacts, and offer improvement suggestions, positioning itself as a reliable assistant for visual inspection. Our contributions are in two folds:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.651, + 0.483, + 0.68 + ], + "angle": 0, + "content": "- We establish the first visual instruction tuning dataset and benchmark for anomaly detection and reasoning." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.681, + 0.484, + 0.71 + ], + "angle": 0, + "content": "- We propose the first specialist visual assistant with state-of-the-art performance for this new impactful domain." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.651, + 0.484, + 0.71 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.724, + 0.233, + 0.74 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Multimodal Large Language Models. Vision-Language Models (VLMs), such as CLIP [73], exhibit robust zero-shot classification capabilities and have been applied to a range of downstream vision tasks [10, 53, 56, 61, 86]. Combining a VLM's vision encoder and an LLM [20, 62, 74], MLLMs [15, 48, 49, 57, 111] enable text-based interactions related to visual content. MLLMs have shown remarkable reasoning capability, particularly when incorporated with prompting strategies such as Chain-of-Thought [66, 88, 105]. Recent studies have harnessed MLLMs to provide reasoning for" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.097, + 0.619, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.097, + 0.724, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.73, + 0.101, + 0.892, + 0.158 + ], + "angle": 0, + "content": "GT: The top-left candle shows a missing chunk of wax and an irregular indentation near the wick." + }, + { + "type": "text", + "bbox": [ + 0.583, + 0.182, + 0.86, + 0.195 + ], + "angle": 0, + "content": "Do you see any anomalies in the image?" + }, + { + "type": "image", + "bbox": [ + 0.868, + 0.173, + 0.897, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.199, + 0.551, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.556, + 0.201, + 0.848, + 0.235 + ], + "angle": 0, + "content": "The candle in the top left appears to have a damaged or missing wick, which is unusual compared to the other candles." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.244, + 0.553, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.559, + 0.246, + 0.858, + 0.291 + ], + "angle": 0, + "content": "Yes, there is an anomaly in the top left candle. It has a small hole or crack near the wick, which is not present in the other candles." + }, + { + "type": "text", + "bbox": [ + 0.612, + 0.3, + 0.86, + 0.312 + ], + "angle": 0, + "content": "What is the cause for the anomaly?" + }, + { + "type": "image", + "bbox": [ + 0.868, + 0.291, + 0.897, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.314, + 0.553, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.558, + 0.318, + 0.858, + 0.362 + ], + "angle": 0, + "content": "The anomaly could be due to manufacturing defects, such as improper molding or cooling, or it might result from external damage like a puncture." + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.38, + 0.907, + 0.423 + ], + "angle": 0, + "content": "Figure 2. Industrial image anomaly reasoning results from GPT-4o [72] and our Anomaly-OV. The responses for fine-grained anomaly reasoning are highlighted, with the ground truth given for reference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.908, + 0.485 + ], + "angle": 0, + "content": "downstream tasks, e.g., video anomaly detection [67, 94], affective computing [14, 29], and visual commonsense reasoning [109], revealing more interpretability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.907, + 0.627 + ], + "angle": 0, + "content": "Unsupervised Anomaly Detection. Due to the scarcity and difficulty of collecting anomalous data, researchers often focus on the unsupervised AD setting, which exclusively uses normal data to train an AD model. Earlier studies, such as reconstruction-based [63, 69, 99], student-teacher [18, 85, 102], and augmentation-based [46] approaches, assume a large amount of normal data is available. These traditional approaches are less practical when data are restricted or expensive, such as in the medical domain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.909, + 0.755 + ], + "angle": 0, + "content": "Zero-Shot Anomaly Detection. Unlike unsupervised AD [35, 77] and few-shot AD [22, 28, 36, 51, 112], ZSAD models directly access the likelihood of abnormality for a given image without requiring data specific to the target object. Existing works [6, 19, 38, 110] accomplish ZSAD by comparing visual and textual features encoded by visual and text encoders of CLIP and constructing their positive (anomaly) and negative (normal) prompts in the format of:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.763, + 0.816, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {P} ^ {+} = \\left[ V _ {1} \\right] \\left[ V _ {2} \\right] \\dots \\left[ V _ {n} \\right] [ o b j e c t ] \\\\ \\mathcal {P} ^ {-} = \\left[ W _ {1} \\right] \\left[ W _ {2} \\right] \\dots \\left[ W _ {n} \\right] [ o b j e c t ] \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\( V_{i} \\) and \\( W_{i} \\) are handcrafted or learnable tokens, and object refers to the word object or the class name of the object. However, simply utilizing object to represent all kinds of objects cannot capture the class-awareness abnormality types. Also, for an intelligent visual assistant, the images should be totally blind to the user (object-agnostic)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20371" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.088, + 0.907, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.451, + 0.908, + 0.481 + ], + "angle": 0, + "content": "Figure 3. Overview of the Anomaly-0V architecture. It consists of two training stages: (1) professional training for the anomaly expert, and (2) visual instruction tuning for anomaly detection and reasoning. Text and visual tokens are distinguished by different colors." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.505, + 0.184, + 0.52 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.53, + 0.221, + 0.547 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.553, + 0.483, + 0.719 + ], + "angle": 0, + "content": "Training an MLLM from scratch demands extensive data and computational resources to align the visual and textual embedding spaces and develop robust instruction-following capabilities. Recent studies [23, 83, 93] reveal that pretrained MLLMs function as generalists, possessing a broad knowledge base but underperforming in specialized domains. Therefore, our goal is to introduce an auxiliary specialist or expert model designed to guide the generalist in selecting and utilizing critical visual tokens. This approach circumvents the need for large-scale pre-training while preserving the generalization capacity of the original model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.486, + 0.902 + ], + "angle": 0, + "content": "We choose \\(LLaVA-OneVision\\) [44] as our base MLLM because it is open-sourced and performs similarly to other commercial models. \\(LLaVA-OneVision\\) follows the model architectures for LLaVA family [47, 57-59] and other generic MLLMs, which typically consist of three major components: Visual Encoder, Projector, and LLM. The visual encoder [73, 100] extracts the visual information from the raw images, the projector aligns the spaces of visual features with the word embedding, and the LLM is responsible for textual instruction processing and complex reasoning. Since the image resolution for CLIP pre-training is fixed, \\(LLaVA-OneVision\\) leverages AnyRes with pooling strategy to scale" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.908, + 0.568 + ], + "angle": 0, + "content": "up the input raw image resolution. Specifically, the high-resolution images are divided into a prototyped number of crops, and the visual encoder independently processes the image crops before final spatial pooling." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.576, + 0.727, + 0.59 + ], + "angle": 0, + "content": "3.2. Architecture Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.908, + 0.644 + ], + "angle": 0, + "content": "With the same image-splitting strategy AnyRes as LLaVA-OneVision, the input high-resolution image is split into several crops, and the new image set can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.654, + 0.907, + 0.671 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} = \\left\\{I _ {0}, I _ {1}, I _ {2}, \\dots , I _ {n - 1} \\right\\} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.681, + 0.909, + 0.801 + ], + "angle": 0, + "content": "where \\( I_0 \\) is the resized original image and \\( I_{j\\neq 0} \\) refers to the image crops. As shown in Figure 3, the image set \\( \\mathcal{I} \\) will be processed by the visual encoder \\( \\mathcal{F}_{\\theta} \\) to generate the final visual features \\( \\{\\mathbf{v}_j^o\\} \\). Similar to AnomalyCLIP [110], we store the outputs for four selected layers in the ViT [21] to capture the image representations from different levels and apply four adapters to compress the feature dimension. Then, the extracted visual features can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.665, + 0.812, + 0.907, + 0.831 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {j} ^ {i} = \\mathcal {F} _ {\\theta} ^ {i} (I _ {j}) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.902 + ], + "angle": 0, + "content": "where \\(i\\) denotes the \\(i\\)-th level and \\(j\\) refers to the index of corresponding image in \\(\\mathcal{I}\\). These multi-level features have been demonstrated to be effective in capturing fine-grained local semantics by recent works [6, 28, 110]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20372" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.482, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.178, + 0.471, + 0.193 + ], + "angle": 0, + "content": "Figure 4. Simulation of visual anomaly inspection by humans." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.204, + 0.484, + 0.4 + ], + "angle": 0, + "content": "The large-scale pre-trained CLIP models align the projection spaces of the textual and visual encoder. Therefore, the encoded image features already contain the class information required by ZSAD. To avoid human involvement in object classification and reduce the model complexity, we remove the heavy text encoder commonly utilized in existing works and let the visual model itself parse the information for suspicious classes or objects. Specifically, the output visual features for the original image \\(\\mathbf{v}_0^o\\) are leveraged to provide the global description of the target object or regions in the look-back path. With the multi-level features and the global embeddings, the LTFM module is responsible for the recognition and localization of suspicious tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.401, + 0.484, + 0.566 + ], + "angle": 0, + "content": "Drawing inspiration from human visual inspection, where suspicious objects or regions are identified and then inspected closely (see Figure 4), we design the VT selector module for aggregating (zooming in) the crucial visual tokens and explicitly assisting the LLM in distinguishing these tokens from many irrelevant ones when dealing with instructions regarding anomaly detection and reasoning. Additionally, the original visual features are preserved to maintain the generalization capability of the base model on regular instructions, such as Can you describe the content of the image?" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.576, + 0.36, + 0.592 + ], + "angle": 0, + "content": "3.3. Look-Twice Feature Matching" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.484, + 0.765 + ], + "angle": 0, + "content": "Given the global object information \\(\\mathbf{v}_0^o\\) provided by the lookback path, we generate the class-awareness abnormality description by merging \\(\\mathbf{v}_0^o\\) with two learnable embeddings: \\(\\mathbf{e}^{+}\\in \\mathbb{R}^{D}\\) and \\(\\mathbf{e}^{-}\\in \\mathbb{R}^{D}\\), where \\(+\\) and \\(-\\) indicate positive (anomalous) and negative (normal) patterns and \\(D\\) is the embedding dimension. Specifically, a linear layer \\(\\mathcal{T}_i^o\\) is applied along the token dimension to select and fuse useful tokens from \\(\\mathbf{v}_0^o\\), and then the fused vector will be concatenated with \\(\\mathbf{e}^{+}\\) and \\(\\mathbf{e}^{-}\\) independently and pass through two MLPs \\(\\{\\mathcal{G}_i^+, \\mathcal{G}_i^-\\}\\) to generate the abnormality and normality descriptions \\(\\{\\mathbf{d}_i^+, \\mathbf{d}_i^-\\}\\), which can be represented by:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.776, + 0.484, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\mathbf {d} _ {i} ^ {+}, \\mathbf {d} _ {i} ^ {-} \\right\\} = \\left\\{ \\begin{array}{l} \\mathcal {G} _ {i} ^ {+} \\left(\\mathbf {e} ^ {+} \\circ \\mathcal {T} _ {i} ^ {o} \\left(\\mathbf {v} _ {0} ^ {o}\\right)\\right) \\\\ \\mathcal {G} _ {i} ^ {-} \\left(\\mathbf {e} ^ {-} \\circ \\mathcal {T} _ {i} ^ {o} \\left(\\mathbf {v} _ {0} ^ {o}\\right)\\right) \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.884 + ], + "angle": 0, + "content": "The visual features extracted from different levels of the ViT focus on different scales of semantics. Thus, the parameters of \\(\\mathcal{T}_i^o\\) and \\(\\{\\mathcal{G}_i^+, \\mathcal{G}_i^-\\}\\) should be independent for different levels, where \\(i\\) indicate the level number." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.9 + ], + "angle": 0, + "content": "Similar to the zero-shot classification mechanism of CLIP" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.137 + ], + "angle": 0, + "content": "models, we calculate the possibilities of each patch token in \\(\\mathbf{v}_j^i\\) belonging to the anomalous patterns by combining cosine similarity and softmax operations:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.146, + 0.907, + 0.185 + ], + "angle": 0, + "content": "\\[\n\\mathbf {m} _ {j} ^ {i} = \\frac {\\exp \\left(< \\mathbf {d} _ {i} ^ {+} , \\mathbf {v} _ {j} ^ {i} > / \\tau\\right)}{\\exp \\left(< \\mathbf {d} _ {i} ^ {+} , \\mathbf {v} _ {j} ^ {i} > / \\tau\\right) + \\exp \\left(< \\mathbf {d} _ {i} ^ {-} , \\mathbf {v} _ {j} ^ {i} > / \\tau\\right)} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.194, + 0.908, + 0.285 + ], + "angle": 0, + "content": "where \\(\\mathbf{m}_j^i\\) represents the significance map for visual tokens, \\(\\tau\\) is the temperature hyperparameter, and \\(<,>\\) refers to the cosine similarity operator. The patch weight in \\(\\mathbf{m}_j^i\\) indicates the closeness of the corresponding visual token to the anomalous pattern. Then, all the maps are averaged to capture the token significances from low to high levels:" + }, + { + "type": "equation", + "bbox": [ + 0.653, + 0.294, + 0.907, + 0.335 + ], + "angle": 0, + "content": "\\[\n\\mathbf {m} _ {j} = \\sum_ {i = 1} ^ {4} \\mathbf {m} _ {j} ^ {i} / 4 \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.343, + 0.906, + 0.404 + ], + "angle": 0, + "content": "The visual features are leveraged twice in the forward and look-back paths, so this module is named by Look-Twice Feature Matching (LTFM), following the nature of two-step human visual inspection shown in Figure 4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.412, + 0.716, + 0.426 + ], + "angle": 0, + "content": "3.4. Visual Token Selector" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.434, + 0.907, + 0.6 + ], + "angle": 0, + "content": "Under the image cropping strategy widely applied in recent MLLMs, there will be a large number of visual tokens for a high-resolution image, e.g., 7290 tokens for an image with \\(1152 \\times 1152\\) resolution in LLaVA-OneVision. While these tokens provide rich visual details, the LLM is required to pick the most useful information when adapting to a specific task. When the LLM lacks enough knowledge in this domain, the token-picking process will become complicated. Thus, our solution is to introduce a specialist or expert who knows which token is crucial or not and assist the LLM in selecting and emphasizing (zooming in) the crucial tokens." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.601, + 0.906, + 0.706 + ], + "angle": 0, + "content": "Given the encoded visual tokens \\(\\{\\mathbf{v}_j^o\\}\\) for each image crop in \\(\\mathcal{I}\\) and the corresponding significance map \\(\\mathbf{m}_j\\), the suspicious tokens are emphasized by direct multiplication of the two tensors. Then, the normal tokens will be scaled to zero while the anomalous tokens will be maintained. After that, spatial average pooling \\(\\mathcal{P}\\) is applied to reduce the number of tokens. This process can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.716, + 0.907, + 0.734 + ], + "angle": 0, + "content": "\\[\n\\mathbf {q} _ {j} = \\mathcal {P} \\left(\\mathbf {v} _ {j} ^ {o} \\odot \\mathbf {m} _ {j}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.743, + 0.906, + 0.822 + ], + "angle": 0, + "content": "where \\(\\mathbf{q}_j\\in \\mathbb{R}^{h\\times w\\times D}\\) refers to the pooled query tokens. Empirically, setting \\(h = w = 2\\) provides a better trade-off than other options. Then, a Q-Former \\(\\mathcal{Q}\\) [49] is leveraged to aggregate the correlated tokens in the original output by forwarding \\(\\mathbf{q}_j\\) as the query and \\(\\mathbf{v}_j^o\\) as the key and value:" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.83, + 0.907, + 0.849 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {j} ^ {s} = \\mathcal {Q} \\left(\\mathbf {q} _ {j}, \\mathbf {v} _ {j} ^ {o}, \\mathbf {v} _ {j} ^ {o}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "The Visual Token Selector (VT Selector) serves as a tool for the anomaly expert to hand-pick visual tokens that contain the most suspicious semantics for a given image." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20373" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.098, + 0.156, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.159, + 0.097, + 0.278, + 0.143 + ], + "angle": 0, + "content": "Q: Do you notice any anomalies in the image? A: Yes, the anomaly in the image is a fine, dark crack visible on the ceramic cup, starting from the rim and extending downward." + }, + { + "type": "image_footnote", + "bbox": [ + 0.159, + 0.148, + 0.28, + 0.201 + ], + "angle": 0, + "content": "Q: How can I prevent similar anomalies in the future? A: Avoid rapid temperature changes, handle the cup carefully to minimize impacts, and consider using cups made with thicker, more durable materials." + }, + { + "type": "list", + "bbox": [ + 0.159, + 0.097, + 0.28, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.105, + 0.354, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.356, + 0.097, + 0.498, + 0.143 + ], + "angle": 0, + "content": "Q: Are there any anomalies in this cable image? A: Yes, the green wire is missing its internal copper conductors, unlike the blue and brown wires, which contain complete sets of bundled copper strands." + }, + { + "type": "image_footnote", + "bbox": [ + 0.356, + 0.148, + 0.498, + 0.207 + ], + "angle": 0, + "content": "Q: What could be the reason behind this anomaly? A: The absence of copper in the green wire could be due to manufacturing defects such as failure to insert the strands or a mechanical issue where the strands were removed or misplaced during production or handling." + }, + { + "type": "list", + "bbox": [ + 0.356, + 0.097, + 0.498, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.105, + 0.561, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.562, + 0.097, + 0.688, + 0.13 + ], + "angle": 0, + "content": "Q: What does the image show? A: The image shows a densely cellular tissue sample with features indicating potential metastatic breast cancer." + }, + { + "type": "image_footnote", + "bbox": [ + 0.562, + 0.136, + 0.692, + 0.206 + ], + "angle": 0, + "content": "Q: What suggests that there might be malignant activity in the tissue sample? A: Malignant activity is suggested by irregularly shaped cells, hyperchromatic nuclei, and an increased nuclear-to-cytoplasmic ratio, indicative of abnormal and increased DNA content typical in cancerous cells." + }, + { + "type": "list", + "bbox": [ + 0.562, + 0.097, + 0.692, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.098, + 0.755, + 0.13 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.758, + 0.097, + 0.892, + 0.2 + ], + "angle": 0, + "content": "Q: What kind of object are you examining in the images? A: I'm examining a round cookie with a golden-brown surface and visible chocolate chips embedded across its surface. Q: Are there any significant anomalies detected on this cookie? A: Yes, there are two significant anomalies on the cookie: a central defect featuring a noticeable hole and an edge defect showing a gap or crack." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.224, + 0.908, + 0.281 + ], + "angle": 0, + "content": "Figure 5. Composition of the instruction data in Anomaly-Instruct-125k. There are four main types of image samples: in-the-wild, industrial, medical, and \\(3D\\) (in the format of multi-view images), covering most image anomaly detection tasks and enabling the possibility of a unified assistant for visual inspection. The reasoning words are highlighted in blue. For more information about dataset establishment, statistics, and the data collection pipeline, please refer to Section A1 in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.299, + 0.273, + 0.314 + ], + "angle": 0, + "content": "3.5. Inference and Loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.321, + 0.484, + 0.397 + ], + "angle": 0, + "content": "Anomaly Prediction. In the traditional anomaly detection task, the model predicts the possibility of the image being abnormal. To achieve anomaly score prediction, we aggregate the anomaly information from all the image crops by an average operation weighted on the significance maps:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.407, + 0.484, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\mathbf {r} (\\mathcal {I}) = \\frac {\\sum_ {j , k} \\mathbf {v} _ {j} ^ {\\mathrm {s}} [ k ] \\cdot \\mathcal {P} (\\mathbf {m} _ {j}) [ k ]}{\\sum_ {j , k} \\mathcal {P} (\\mathbf {m} _ {j}) [ k ]} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.456, + 0.484, + 0.517 + ], + "angle": 0, + "content": "where \\(\\mathcal{P}\\) is the same spatial pooling in VT Selector and \\(\\mathbf{r}(\\mathcal{I})\\) is a vector containing the global anomaly information for the entire image. Then, the anomaly expert can calculate the image-level abnormal possibility by parsing \\(\\mathbf{r}(\\mathcal{I})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.529, + 0.483, + 0.546 + ], + "angle": 0, + "content": "\\[\n\\operatorname {s c o r e} (\\mathcal {I}) = \\operatorname {S i g m o i d} \\left(\\mathcal {G} ^ {o} (\\mathbf {r} (\\mathcal {I}))\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.557, + 0.484, + 0.618 + ], + "angle": 0, + "content": "where \\(\\mathcal{G}^o\\) is an MLP for distinguishing normal and abnormal semantics. To handle the unbalanced sample distribution, we employ the balanced BCE loss as the professional training objective for the anomaly expert components." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.484, + 0.851 + ], + "angle": 0, + "content": "Text Generation. Instead of directly forwarding the concatenation of the original \\(\\{\\mathbf{v}_j^o\\}\\) and the selected \\(\\{\\mathbf{r}(\\mathcal{I}),\\mathbf{v}_j^s\\}\\) visual tokens into the LLM, we apply an indication prompt with suspicious feature: in the middle of the two series of tokens, which will highlight the selected tokens for LLM when handling anomaly-related instructions. This approach can be considered a form of prompt engineering in MLLMs. Besides, the is chosen from {highly, moderately, slightly} and is determined by score \\((\\mathcal{I})\\) and predefined thresholds \\(\\{\\mathbf{s}_{low},\\mathbf{s}_{high}\\}\\). When the input image \\(\\mathcal{I}\\) has a high likelihood of anomaly, the LLM will place greater emphasis on the selected tokens; otherwise, these tokens will have less significance. The text generation is implemented by the original auto-regressive token prediction mechanism of LLM:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.864, + 0.484, + 0.905 + ], + "angle": 0, + "content": "\\[\np \\left(X _ {a} \\mid \\mathcal {I}, X _ {q}\\right) = \\prod_ {t = 1} ^ {L} p _ {\\theta} \\left(x _ {t} \\mid \\mathcal {I}, X _ {q, < t}, X _ {a, < t}\\right) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.3, + 0.908, + 0.376 + ], + "angle": 0, + "content": "where \\( X_{a, < t} \\) and \\( X_{q, < t} \\) are the answer and instruction tokens from all prior turns before the current prediction token \\( x_{t} \\) for a sequence of length \\( L \\). The entire model is parameterized by \\( \\theta \\) and trained by the original language model cross-entropy loss for each predicted answer token \\( x_{t} \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.39, + 0.741, + 0.405 + ], + "angle": 0, + "content": "4. Dataset and Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.415, + 0.907, + 0.567 + ], + "angle": 0, + "content": "The lack of multimodal instruction-following data for image anomaly detection and reasoning hinders the development of special intelligent assistants in this domain. Even though AnomalyGPT [28] introduces a prompt tuning dataset by simulating the anomalies, the scale of their dataset and the diversity of their instructions and answers are limited, only focusing on anomaly localization. To resolve the data scarcity issue, we establish the first large-scale instruction tuning dataset: Anomaly-Instruct-125k and the corresponding anomaly detection and reasoning benchmark: VisA-D&R." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.577, + 0.731, + 0.592 + ], + "angle": 0, + "content": "4.1. Anomaly-Instruct-125k" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.734 + ], + "angle": 0, + "content": "LLaVA [57] builds its instruction tuning dataset by leveraging the image caption and bounding boxes available in the COCO dataset[55] to prompt the text-only GPT-4. ShareGPT4V [8] provides a higher-quality dataset by directly prompting GPT-4V [71]. However, there is no image caption provided in existing anomaly detection datasets [1, 2], and no matter GPT-4V [71] or most recent GPT-4o [72] cannot accurately locate and describe the anomalies in the image without explicit human involvement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To resolve these issues, we design a new prompt pipeline for accurate anomaly description generation. Since most of the datasets contain annotations for anomaly types, we manually combine the class name and anomaly type, such as a [capsule] with [poke] on surface. If the anomaly masks are provided, we draw bounding boxes on the images to highlight the anomalous area. The short description and the image with (or w/o) bounding boxes are used to prompt GPT-4o to generate the detailed image and anomaly descriptions. Then, we employ an in-context learning strategy similar to LLaVA to create the instructions." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20374" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.106, + 0.097, + 0.175, + 0.107 + ], + "angle": 0, + "content": "Detection:" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.107, + 0.458, + 0.128 + ], + "angle": 0, + "content": "Q: Are there any defects for the object in the image? Please reply with 'Yes' or 'No'." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.137, + 0.175, + 0.148 + ], + "angle": 0, + "content": "Reasoning:" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.148, + 0.414, + 0.157 + ], + "angle": 0, + "content": "Q1: Do you observe any anomalies in the image?" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.157, + 0.425, + 0.167 + ], + "angle": 0, + "content": "Q2: Can you describe the anomalies you observed?" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.167, + 0.439, + 0.177 + ], + "angle": 0, + "content": "Q3: What is the potential cause for the anomalies?" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.177, + 0.465, + 0.187 + ], + "angle": 0, + "content": "Q4: How can such anomalies be prevented in the future?" + }, + { + "type": "list", + "bbox": [ + 0.107, + 0.148, + 0.465, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.207, + 0.486, + 0.236 + ], + "angle": 0, + "content": "Figure 6. Prompt examples in VisA-D&R for detection and reasoning. The complex reasoning instructions are highlighted." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.248, + 0.485, + 0.552 + ], + "angle": 0, + "content": "For a unified visual inspection dataset, precise instruction data is collected from MVtec AD [2], the training set of BMAD [1], Anomaly-ShapeNet [50], Real3D-AD [60], and MVtec-3D AD [4], covering both 2D to 3D data across industry to medical domains. The 3D point cloud data are converted into 9 multi-view images, and the corresponding masks are rendered using predefined camera positions. However, the diversities and scales of these datasets are relatively limited, probably due to the difficulty of collecting anomaly images. To scale up the instruction data, we introduce an automatic anomaly data collection pipeline combining GPT4o [72] and Google Image Search [26] for image collection, data cleaning, and instruction generation. Finally, 72k in-the-wild images (named as WebAD) targeting anomaly detection are collected, significantly enriching our instruction dataset. Several samples from Anomaly-Instruct-125k are shown in Figure 5. The instructions are mainly in the format of multi-round conversations, covering anomaly detection and description in low-level reasoning and potential cause and future suggestions for complex understanding." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.561, + 0.214, + 0.575 + ], + "angle": 0, + "content": "4.2.VisA-D&R" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.483, + 0.764 + ], + "angle": 0, + "content": "VisA [115] is a classic but challenging industrial anomaly detection dataset, providing fine-grained anomaly type and segmentation for each image. For evaluation of the anomaly detection and reasoning performance on existing and future methods, we select 10 classes from VisA and follow a similar data generation pipeline of Anomaly-Instruct-125k to create the benchmark. Differently, significant human effort has been invested in meticulously reviewing all generated images and anomaly descriptions. Wrong descriptions are picked out and re-annotated by humans before utilizing them for Q&A generation. Totally, the benchmark consists of 761 normal samples and 1000 anomalous ones." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "For evaluating detection performance, questions designed to elicit a one-word answer are used to prompt the MLLMs (Figure 6), with results quantified using Accuracy, Precision, Recall, and F1-score. We divide the reasoning performance into two parts: low-level reasoning that focuses on the description of visual defects or anomalies and complex reasoning requiring the MLLMs to provide the potential cause and future improvement strategies for the detected anomalies, where ROUGE-L [54], Sentence-BERT (SBERT) [75], and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "GPT-Score (GPT-4 as the judge [57]) are utilized to quantify the similarity between generated text and ground truth. Note that low-level reasoning is highly correlated to detection performance, while anomaly-type descriptions of low-level reasoning determine the output of complex reasoning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.18, + 0.646, + 0.197 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.204, + 0.725, + 0.22 + ], + "angle": 0, + "content": "5.1. Training & Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.226, + 0.908, + 0.422 + ], + "angle": 0, + "content": "There are two independent training stages for Anomaly-OV. In Stage 1, the components of the anomaly expert are trained to obtain the token selection capability, targeting traditional ZSAD. This stage utilizes all of the data with anomaly labels in Anomaly-Instruct-125k. Similar to previous works [6, 110], when evaluating the model on the datasets contained in the training set, the corresponding datasets are replaced by VisA [115]. In Stage 2, the anomaly expert and visual encoder are frozen, while the projector and LLM are trainable. In addition to our instruction dataset, we sample around \\(350\\mathrm{k}\\) data from the original training recipe of \\(LLaVA-OneVision\\) to maintain the generalization ability. For more details on training, please refer to the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.423, + 0.909, + 0.62 + ], + "angle": 0, + "content": "The ZSAD performance for the anomaly expert is evaluated on nine benchmarks, including MVTec AD [2], VisA [115], AITEX [81], ELPV [16], BTAD [68], and MPDD [39] for industrial inspection, and BrainMRI [40], HeadCT [41], and Br35H [30] for medical diagnosis. AUROC (Area Under the Receiver Operating Characteristic) is leveraged to quantify the image-level AD performance. For text-based anomaly detection, both normal and anomaly data are employed to assess the accuracy by examining if the generated text contains the word Yes. Differently, only anomaly data are utilized to prompt the MLLMs to determine their anomaly reasoning capabilities since the justifications of normality are similar for different models." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.63, + 0.901, + 0.726 + ], + "angle": 0, + "content": "
MethodMVTecVisAHeadCTBrainMRI
Full Model94.091.197.693.9
w/o. Look-back92.890.596.693.5
w/o. e+ & e-92.190.194.792.9
w/o. Q-Former91.789.992.895.1
w/o. WebAD88.588.991.293.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.735, + 0.908, + 0.763 + ], + "angle": 0, + "content": "Table 1. Ablation study for the anomaly expert of Anomaly-OV. w/o. Look-back refers to the removal of \\(\\mathbf{v}_0^o\\) in LTFM." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.78, + 0.79 + ], + "angle": 0, + "content": "5.2. Zero-Shot Anomaly Detection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.901 + ], + "angle": 0, + "content": "As shown in Table 2, compared with existing methods, the anomaly expert of Anomaly-OV achieves significant image-level AUROC improvements on most of the ZSAD benchmarks, which demonstrates that the text encoder widely applied in existing models is not necessary. The success of our model mainly originates from the extra data of WebAD (Table 1), which enables the model to learn more generic" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20375" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.9, + 0.245 + ], + "angle": 0, + "content": "
ModelIndustrial DefectsMedical AnomaliesAverage
MVTec ADVisAAITEXELPVBTADMPDDBrainMRIHeadCTBr35H
CLIP [73]74.166.471.059.234.554.373.956.578.463.1
CoOp [108]88.862.866.273.066.855.161.378.486.070.9
WinCLIP [38]91.878.873.074.068.263.692.690.080.579.2
APRIL-GAN [11]86.278.057.665.573.673.089.389.193.178.4
AnoVL [19]92.579.272.570.680.368.988.781.688.480.3
AnomalyCLIP [110]91.582.162.281.588.377.090.393.494.684.5
AdaCLIP [6]89.285.864.579.788.676.094.891.497.785.3
Ours94.091.172.083.089.081.793.997.695.588.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.255, + 0.908, + 0.283 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of Image-level AUROC on different ZSAD methods (some of the results are borrowed from [6, 110, 114]). The best and the second-best results are bolded and underlined, respectively. Please refer to the supplementary for more detailed results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.299, + 0.486, + 0.359 + ], + "angle": 0, + "content": "semantics for normality and abnormality from the data distribution in the absence of the text encoder. This observation also reveals that large-scale in-the-wild online data can benefit zero-shot performance in anomaly detection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.36, + 0.485, + 0.542 + ], + "angle": 0, + "content": "While the Q-Former reduces the model performance on BrainMRI, it shows effectiveness on most benchmarks, indicating the importance of token aggregation. Similarly, the look-back information and two learnable embeddings are required for describing class-awareness abnormality and distinguishing positive and negative features, respectively. As previously discussed, the anomaly expert is responsible for selecting suspicious visual tokens for the LLM, and the significance maps shown in Figure 7 demonstrate the interpretable token selection mechanism. The high intensities are automatically distributed around the anomalous areas even without any supervision of the anomaly masks." + }, + { + "type": "image", + "bbox": [ + 0.105, + 0.553, + 0.223, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.624, + 0.222, + 0.692 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.555, + 0.348, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.624, + 0.347, + 0.692 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.555, + 0.471, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.624, + 0.47, + 0.692 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.704, + 0.48, + 0.72 + ], + "angle": 0, + "content": "Figure 7. Visualization of the significance map on VisA samples." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.379, + 0.745 + ], + "angle": 0, + "content": "5.3. Anomaly Detection & Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.486, + 0.902 + ], + "angle": 0, + "content": "With the strong capabilities of the anomaly expert for zero-shot detection and suspicious token selection, Anomaly-OV accomplishes significant improvement in text-based anomaly detection and reasoning over other open-sourced generalist MLLMs, as shown in Table 4. Here are a few observations: i) While a larger language model cannot guarantee better detection performance, it always provides greater reasoning ability; ii) Most of the existing MLLMs present much lower recall than precision, indicating their insensitivity to visual anomalies; iii) GPT-4o shows stronger reasoning ability" + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.299, + 0.897, + 0.558 + ], + "angle": 0, + "content": "
PCB Example: Excessive Solder
UserAre there any anomalies visible in the image?
GPT-4oThe image shows an HC-SR04 ultrasonic sensor. There don’t appear to be any visible anomalies with the sensor itself. The components and pins seem intact and properly aligned.
LLaVA-OVThe image shows an ultrasonic sensor module, specifically the HC-SR04 model. There are no visible anomalies in the image; it appears to be a standard representation of this type of sensor module.
Ours-7BYes, there is a noticeable anomaly in the image: a short circuit on the PCB, indicated by a white, crusty residue near the ‘Echo’ pin.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.568, + 0.902, + 0.583 + ], + "angle": 0, + "content": "Table 3. Anomaly-OV presents more accurate anomaly detection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.907, + 0.703 + ], + "angle": 0, + "content": "compared to other open-sourced models. Table 3 and Table 5 provide the qualitative comparison of our Anomaly-OV with its base model LLaVA-OV-7B [44] and GPT-4o [72]. Both GPT-4o and LLaVA-OV show insensitivity to anomalous features and cannot accurately detect the anomaly in the image. Sometimes, GPT-4o knows the image is anomalous but fails to describe the anomalies precisely." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We provide the fine-tuned version of the base model LLaVA-OV-0.5B on Anomaly-Instruct-125k, which presents much higher accuracy and more balanced precision and recall than its original version. This demonstrates the effectiveness of our instruction-tuning dataset. By integrating the anomaly expert with the base model, our Anomaly-OV-0.5B achieves 0.08 accuracy and 0.06 F1-score improvements in text-based anomaly detection and better reasoning capability in low-level and complex settings. Equipped with a larger language model, Anomaly-OV-7B provides the best detection performance among all the existing MLLMs and shows comparable reasoning ability with GPT-4o. Notably, we observe that the anomaly expert restricts the detection perfor" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20376" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.905, + 0.301 + ], + "angle": 0, + "content": "
ModelAnomaly DetectionLow-level ReasoningComplex Reasoning
AccuracyPrecisionRecallF1-scoreROUGE-LSBERTGPT-ScoreSBERTGPT-Score
GPT-4V [71]0.680.900.490.550.160.653.310.775.64
GPT-4o [72]0.700.830.710.680.240.714.840.816.89
Qwen2-VL-2B [87]0.650.870.550.590.220.551.940.744.26
Qwen2-VL-7B [87]0.760.910.690.750.250.613.090.684.62
InternVL-2-8B [13]0.740.780.810.760.230.733.690.805.08
InternVL-2-26B [13]0.730.860.660.680.210.744.130.805.49
IXC-2.5-7B [101]0.720.880.630.670.210.582.450.775.14
LLaVA-OV-0.5B [44]0.540.700.190.280.200.632.540.814.34
LLaVA-OV-7B [44]0.710.950.560.630.240.663.570.795.44
LLaVA-OV-0.5B*0.710.770.840.760.310.703.690.825.31
Anomaly-OV-0.5B0.790.860.830.820.330.723.870.835.67
Anomaly-OV-7B0.790.830.860.830.340.734.260.846.34
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.311, + 0.908, + 0.34 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison of text-based anomaly detection and reasoning for MLLMs. Notably, the Accuracy and F1-score for the anomaly expert of Anomaly-OV can be calculated as \\(\\{0.78, 0.77\\}\\) with threshold 0.5. * indicates the model is fine-tuned on our dataset." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.355, + 0.484, + 0.387 + ], + "angle": 0, + "content": "mance of Anomaly-OV. Therefore, the design of a stronger anomaly expert is suggested for future works." + }, + { + "type": "table_caption", + "bbox": [ + 0.107, + 0.4, + 0.308, + 0.413 + ], + "angle": 0, + "content": "Macaroni Example: Yellowish Spot" + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.418, + 0.417, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.107, + 0.493, + 0.136, + 0.504 + ], + "angle": 0, + "content": "User" + }, + { + "type": "table_caption", + "bbox": [ + 0.184, + 0.493, + 0.397, + 0.506 + ], + "angle": 0, + "content": "Do you see any anomalies in the image?" + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.508, + 0.472, + 0.644 + ], + "angle": 0, + "content": "
GPT-4oThe image shows four pieces of elbow macaroni on a green background. The anomaly is that the macaroni pieces are not whole; they are cut in half.
LLaVA-OVThe image shows four pieces of pasta, specifically macaroni shells, arranged on a green textured surface. The pasta appears to be uniformly colored and shaped, with no visible defects or anomalies.
Ours-7BYes, there is an anomaly in the image. The bottom right pasta piece has a noticeable yellowish discoloration on its surface.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.656, + 0.477, + 0.671 + ], + "angle": 0, + "content": "Table 5. Anomaly-OV presents more precise anomaly reasoning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.675, + 0.205, + 0.69 + ], + "angle": 0, + "content": "5.4. Extension" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.697, + 0.486, + 0.804 + ], + "angle": 0, + "content": "With the generalization and multi-image processing capabilities of MLLMs, it is possible to build a unified assistant for visual inspection. Table 6 demonstrates the comprehensive knowledge of Anomaly-OV (without using Anomaly-ShapeNet [50] for training) on 3D and medical (testing set of BMAD [1]) AD & reasoning. More data, benchmarks, and investigation on a unified model are meaningful." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.816, + 0.21, + 0.831 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.486, + 0.903 + ], + "angle": 0, + "content": "In this paper, we establish the first large-scale visual instruction tuning dataset, Anomaly-Instruct-125k, and the corresponding benchmark, VisA-D&R, to address the data scarcity issue for visual anomaly detection and reasoning." + }, + { + "type": "table_caption", + "bbox": [ + 0.536, + 0.356, + 0.65, + 0.369 + ], + "angle": 0, + "content": "3D Example: Bulge" + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.373, + 0.712, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.373, + 0.834, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.84, + 0.404, + 0.862, + 0.412 + ], + "angle": 0, + "content": "···" + }, + { + "type": "table_caption", + "bbox": [ + 0.538, + 0.449, + 0.566, + 0.461 + ], + "angle": 0, + "content": "User" + }, + { + "type": "table_caption", + "bbox": [ + 0.599, + 0.449, + 0.819, + 0.462 + ], + "angle": 0, + "content": "Is there any anomaly present in the bowl?" + }, + { + "type": "table_caption", + "bbox": [ + 0.537, + 0.466, + 0.586, + 0.477 + ], + "angle": 0, + "content": "Ours-7B" + }, + { + "type": "table_caption", + "bbox": [ + 0.597, + 0.466, + 0.88, + 0.491 + ], + "angle": 0, + "content": "Yes, there is a bulge on the side of the bowl, which disrupts its otherwise smooth and symmetrical surface." + }, + { + "type": "table_caption", + "bbox": [ + 0.538, + 0.495, + 0.71, + 0.508 + ], + "angle": 0, + "content": "Medical Example: Pneumonia" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.512, + 0.822, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.538, + 0.588, + 0.566, + 0.599 + ], + "angle": 0, + "content": "User" + }, + { + "type": "table_caption", + "bbox": [ + 0.598, + 0.588, + 0.86, + 0.601 + ], + "angle": 0, + "content": "What is the most likely diagnosis for this patient?" + }, + { + "type": "table_caption", + "bbox": [ + 0.537, + 0.604, + 0.586, + 0.615 + ], + "angle": 0, + "content": "Ours-7B" + }, + { + "type": "table_caption", + "bbox": [ + 0.598, + 0.604, + 0.881, + 0.666 + ], + "angle": 0, + "content": "Based on the chest X-ray, the most likely diagnosis for this patient is pneumonia. The presence of bilateral patchy opacities in the lower lung fields is a common finding in pneumonia, which is an infection that causes inflammation in the air sacs of the lungs." + }, + { + "type": "table_caption", + "bbox": [ + 0.538, + 0.67, + 0.587, + 0.682 + ], + "angle": 0, + "content": "User" + }, + { + "type": "table_caption", + "bbox": [ + 0.598, + 0.67, + 0.772, + 0.683 + ], + "angle": 0, + "content": "What should the patient do next?" + }, + { + "type": "table_caption", + "bbox": [ + 0.537, + 0.687, + 0.586, + 0.698 + ], + "angle": 0, + "content": "Ours-7B" + }, + { + "type": "table_caption", + "bbox": [ + 0.598, + 0.687, + 0.88, + 0.748 + ], + "angle": 0, + "content": "The patient should consult a healthcare professional for further evaluation and appropriate treatment. This may include antibiotics to treat the infection, supportive care such as oxygen therapy, and monitoring of the patient's condition." + }, + { + "type": "table_caption", + "bbox": [ + 0.542, + 0.763, + 0.873, + 0.777 + ], + "angle": 0, + "content": "Table 6. Extension to 3D and medical AD & reasoning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Then, a specialist MLLM, Anomaly-OV, targeting visual inspection is introduced to serve as the baseline in this domain. Anomaly-OV leverages an anomaly expert to assist the LLM with suspicious visual token selection and presents significant improvements on both traditional ZSAD and text-based anomaly detection and reasoning tasks over existing methods. Extension to 3D and medical domains is demonstrated." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20377" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.115, + 0.486, + 0.184 + ], + "angle": 0, + "content": "[1] Jinan Bao, Hanshi Sun, Hanqiu Deng, Yinsheng He, Zhaoxiang Zhang, and Xingyu Li. Bmad: Benchmarks for medical anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4042-4053, 2024. 1, 5, 6, 8, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.187, + 0.485, + 0.257 + ], + "angle": 0, + "content": "[2] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592-9600, 2019. 1, 5, 6, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.259, + 0.483, + 0.328 + ], + "angle": 0, + "content": "[3] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4183–4192, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.33, + 0.485, + 0.413 + ], + "angle": 0, + "content": "[4] Paul Bergmann, Xin Jin, David Sattlegger, and Carsten Steger. The mvtec 3d-ad dataset for unsupervised 3d anomaly detection and localization. In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications. SCITEPRESS - Science and Technology Publications, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.415, + 0.485, + 0.471 + ], + "angle": 0, + "content": "[5] Tri Cao, Jiawen Zhu, and Guansong Pang. Anomaly detection under distribution shift. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6511-6523, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.473, + 0.485, + 0.543 + ], + "angle": 0, + "content": "[6] Yunkang Cao, Jiangning Zhang, Luca Frittoli, Yuqi Cheng, Weiming Shen, and Giacomo Boracchi. Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly detection. In European Conference on Computer Vision, pages 55-72. Springer, 2025. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.545, + 0.485, + 0.627 + ], + "angle": 0, + "content": "[7] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.63, + 0.485, + 0.685 + ], + "angle": 0, + "content": "[8] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions, 2023. 5, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.688, + 0.485, + 0.743 + ], + "angle": 0, + "content": "[9] Qiyu Chen, Huiyuan Luo, Chengkan Lv, and Zhengtao Zhang. A unified anomaly synthesis strategy with gradient ascent for industrial anomaly detection and localization. arXiv preprint arXiv:2407.09359, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.746, + 0.485, + 0.829 + ], + "angle": 0, + "content": "[10] Runnan Chen, Youquan Liu, Lingdong Kong, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, Yu Qiao, and Wenping Wang. Clip2scene: Towards label-efficient 3d scene understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7030, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.831, + 0.485, + 0.9 + ], + "angle": 0, + "content": "[11] Xuhai Chen, Yue Han, and Jiangning Zhang. April-gan: A zero-/few-shot anomaly classification and segmentation method for cvpr 2023 vand workshop challenge tracks 1&2: 1st place on zero-shot ad and 4th place on few-shot ad, 2023. 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.907, + 0.148 + ], + "angle": 0, + "content": "[12] Yuanhong Chen, Yu Tian, Guansong Pang, and Gustavo Carneiro. Deep one-class classification via interpolated gaussian descriptor. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 383-392, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.15, + 0.908, + 0.288 + ], + "angle": 0, + "content": "[13] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, Ji Ma, Jiaqi Wang, Xiaoyi Dong, Hang Yan, Hewei Guo, Conghui He, Botian Shi, Zhenjiang Jin, Chao Xu, Bin Wang, Xingjian Wei, Wei Li, Wenjian Zhang, Bo Zhang, Pinlong Cai, Licheng Wen, Xiangchao Yan, Min Dou, Lewei Lu, Xizhou Zhu, Tong Lu, Dahua Lin, Yu Qiao, Jifeng Dai, and Wenhai Wang. How far are we to gpt-4v? closing the gap to commercial multimodal models with opensource suites, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.29, + 0.908, + 0.373 + ], + "angle": 0, + "content": "[14] Zebang Cheng, Zhi-Qi Cheng, Jun-Yan He, Jingdong Sun, Kai Wang, Yuxiang Lin, Zheng Lian, Xiaojiang Peng, and Alexander Hauptmann. Emotion-llama: Multimodal emotion recognition and reasoning with instruction tuning. In Conference on Neural Information Processing Systems, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.375, + 0.908, + 0.457 + ], + "angle": 0, + "content": "[15] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.459, + 0.908, + 0.529 + ], + "angle": 0, + "content": "[16] Sergiu Deitsch, Vincent Christlein, Stephan Berger, Claudia Buerhop-Lutz, Andreas Maier, Florian Gallwitz, and Christian Riess. Automatic classification of defective photovoltaic module cells in electroluminescence images. Solar Energy, 185:455-468, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.531, + 0.908, + 0.613 + ], + "angle": 0, + "content": "[17] Chenghao Deng, Haote Xu, Xiaolu Chen, Haodi Xu, Xiaotong Tu, Xinghao Ding, and Yue Huang. Simclip: Refining image-text alignment with simple prompts for zero-/few-shot anomaly detection. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 1761-1770, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.616, + 0.908, + 0.671 + ], + "angle": 0, + "content": "[18] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In IEEE/CVF conference on computer vision and pattern recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.673, + 0.908, + 0.729 + ], + "angle": 0, + "content": "[19] Hanqiu Deng, Zhaoxiang Zhang, Jinan Bao, and Xingyu Li. Anovl: Adapting vision-language models for unified zero-shot anomaly localization. arXiv preprint arXiv:2308.15939, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.731, + 0.908, + 0.773 + ], + "angle": 0, + "content": "[20] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.774, + 0.908, + 0.87 + ], + "angle": 0, + "content": "[21] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.873, + 0.908, + 0.902 + ], + "angle": 0, + "content": "[22] Zheng Fang, Xiaoyang Wang, Haocheng Li, Jiejie Liu, Qiugui Hu, and Jimin Xiao. Fastrecon: Few-shot industrial" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.908, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20378" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.131, + 0.093, + 0.486, + 0.135 + ], + "angle": 0, + "content": "anomaly detection via fast feature reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17481-17490, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.136, + 0.486, + 0.205 + ], + "angle": 0, + "content": "[23] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J Black. Chatpose: Chatting about 3d human pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2093-2103, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.207, + 0.485, + 0.262 + ], + "angle": 0, + "content": "[24] Tharindu Fernando, Harshala Gammulle, Simon Denman, Sridha Sridharan, and Clinton Fookes. Deep learning for medical anomaly detection-a survey. ACM Computing Surveys (CSUR), 54(7):1-37, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.264, + 0.484, + 0.319 + ], + "angle": 0, + "content": "[25] Matic Fučka, Vitjan Zavrtanik, and Danijel Skočaj. Transfusion-a transparency-based diffusion model for anomaly detection. In European conference on computer vision, pages 91-108. Springer, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.321, + 0.485, + 0.36 + ], + "angle": 0, + "content": "[26] Google. Google-images-search 1.4.7, 2024. https://pypi.org/project/Google-Images-Search.6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.363, + 0.484, + 0.433 + ], + "angle": 0, + "content": "[27] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Hao Li, Ming Tang, and Jinqiao Wang. Filo: Zero-shot anomaly detection by fine-grained description and high-quality localization. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2041-2049, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.434, + 0.484, + 0.503 + ], + "angle": 0, + "content": "[28] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. Anomalygpt: Detecting industrial anomalies using large vision-language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1932-1940, 2024. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.505, + 0.484, + 0.56 + ], + "angle": 0, + "content": "[29] Yuxiang Guo, Faizan Siddiqui, Yang Zhao, Rama Chellappa, and Shao-Yuan Lo. Stimuvar: Spatiotemporal stimuli-aware video affective reasoning with multimodal large language models. arXiv preprint arXiv:2409.00304, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.562, + 0.484, + 0.587 + ], + "angle": 0, + "content": "[30] Ahmed Hamada. Br35h: Brain tumor detection 2020, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.591, + 0.484, + 0.673 + ], + "angle": 0, + "content": "[31] Changhee Han, Leonardo Rundo, Kohei Murao, Tomoyuki Noguchi, Yuki Shimahara, Zoltan Adam Milacski, Saori Koshino, Evis Sala, Hideki Nakayama, and Shin'ichi Satoh. Madgan: Unsupervised medical anomaly detection gan using multiple adjacent brain mri slice reconstruction. BMC bioinformatics, 22:1-20, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.676, + 0.484, + 0.744 + ], + "angle": 0, + "content": "[32] Liren He, Zhengkai Jiang, Jinlong Peng, Liang Liu, Qianggang Du, Xiaobin Hu, Wenbing Zhu, Mingmin Chi, Yabiao Wang, and Chengjie Wang. Learning unified reference representation for unsupervised multi-class anomaly detection. arXiv preprint arXiv:2403.11561, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.746, + 0.484, + 0.801 + ], + "angle": 0, + "content": "[33] Chih-Hui Ho, Kuan-Chuan Peng, and Nuno Vasconcelos. Long-tailed anomaly detection with learnable class names. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12435-12446, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.803, + 0.484, + 0.872 + ], + "angle": 0, + "content": "[34] Jinlei Hou, Yingying Zhang, Qiaoyong Zhong, Di Xie, Shiliang Pu, and Hong Zhou. Divide-and-assemble: Learning block-wise memory for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8791-8800, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.874, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[35] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.907, + 0.121 + ], + "angle": 0, + "content": "few-shot anomaly detection. In European Conference on Computer Vision, pages 303-319. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.123, + 0.909, + 0.205 + ], + "angle": 0, + "content": "[36] Chaoqin Huang, Aofan Jiang, Jinghao Feng, Ya Zhang, Xin chao Wang, and Yanfeng Wang. Adapting visual-language models for generalizable anomaly detection in medical images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11375-11385, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.207, + 0.908, + 0.275 + ], + "angle": 0, + "content": "[37] Brian KS Isaac-Medina, Yona Falinie A Gaus, Neelanjan Bhowmik, and Toby P Breckon. Towards open-world object-based anomaly detection via self-supervised outlier synthesis. In European Conference on Computer Vision (ECCV), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.278, + 0.908, + 0.348 + ], + "angle": 0, + "content": "[38] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19606-19616, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.349, + 0.908, + 0.431 + ], + "angle": 0, + "content": "[39] Stepan Jezek, Martin Jonak, Radim Burget, Pavel Dvorak, and Milos Skotak. Deep learning-based defect detection of metal parts: evaluating current methods in complex conditions. In 2021 13th International congress on ultra modern telecommunications and control systems and workshops (ICUMT), pages 66-71. IEEE, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.433, + 0.907, + 0.46 + ], + "angle": 0, + "content": "[40] Pranita Balaji Kanade and PP Gumaste. Brain tumor detection using mri images. *Brain*, 3(2):146-150, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.462, + 0.897, + 0.476 + ], + "angle": 0, + "content": "[41] Felipe Campos Kitamura. Head ct - hemorrhage, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.478, + 0.908, + 0.545 + ], + "angle": 0, + "content": "[42] Mingyu Lee and Jongwon Choi. Text-guided variational image generation for industrial anomaly detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26519-26528, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.548, + 0.906, + 0.603 + ], + "angle": 0, + "content": "[43] Aodong Li, Chen Qiu, Marius Kloft, Padhraic Smyth, Maja Rudolph, and Stephan Mandt. Zero-shot anomaly detection via batch normalization. Advances in Neural Information Processing Systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.605, + 0.907, + 0.659 + ], + "angle": 0, + "content": "[44] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer, 2024. 1, 3, 7, 8, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.661, + 0.908, + 0.744 + ], + "angle": 0, + "content": "[45] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. LLaVA-med: Training a large language-and-vision assistant for biomedicine in one day. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 1, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.746, + 0.906, + 0.801 + ], + "angle": 0, + "content": "[46] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In IEEE/CVF conference on computer vision and pattern recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.803, + 0.908, + 0.858 + ], + "angle": 0, + "content": "[47] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava-last-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.86, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[48] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Interna-" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.909, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20379" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.485, + 0.119 + ], + "angle": 0, + "content": "tional conference on machine learning, pages 12888-12900. \nPMLR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.122, + 0.486, + 0.189 + ], + "angle": 0, + "content": "[49] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.191, + 0.485, + 0.285 + ], + "angle": 0, + "content": "[50] Wenqiao Li, Xiaohao Xu, Yao Gu, Bozhong Zheng, Shenghua Gao, and Yingna Wu. Towards scalable 3d anomaly detection and localization: A benchmark via 3d anomaly synthesis and a self-supervised learning network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22207-22216, 2024. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.288, + 0.484, + 0.37 + ], + "angle": 0, + "content": "[51] Xiaofan Li, Zhizhong Zhang, Xin Tan, Chengwei Chen, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Prompt: Learning prompts with only normal samples for few-shot anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16838-16848, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.372, + 0.484, + 0.427 + ], + "angle": 0, + "content": "[52] Yiting Li, Adam Goodge, Fayao Liu, and Chuan-Sheng Foo. Promptad: Zero-shot anomaly detection using text prompts. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1093-1102, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.428, + 0.484, + 0.509 + ], + "angle": 0, + "content": "[53] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.511, + 0.482, + 0.551 + ], + "angle": 0, + "content": "[54] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.553, + 0.484, + 0.636 + ], + "angle": 0, + "content": "[55] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.637, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[56] Yuqi Lin, Minghao Chen, Wenxiao Wang, Boxi Wu, Ke Li, Binbin Lin, Haifeng Liu, and Xiaofei He. Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15305-15314, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.721, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[57] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2, 3, 5, 6, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.484, + 0.831 + ], + "angle": 0, + "content": "[58] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.484, + 0.873 + ], + "angle": 0, + "content": "[59] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.874, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[60] Jiaqi Liu, Guoyang Xie, ruitao chen, Xinpeng Li, Jinbao Wang, Yong Liu, Chengjie Wang, and Feng Zheng. Real3d" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.908, + 0.134 + ], + "angle": 0, + "content": "AD: A dataset of point cloud anomaly detection. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.135, + 0.907, + 0.218 + ], + "angle": 0, + "content": "[61] Jie Liu, Yixiao Zhang, Jie-Neng Chen, Junfei Xiao, Yongyi Lu, Bennett A Landman, Yixuan Yuan, Alan Yuille, Yucheng Tang, and Zongwei Zhou. Clip-driven universal model for organ segmentation and tumor detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21152-21164, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.219, + 0.907, + 0.273 + ], + "angle": 0, + "content": "[62] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.274, + 0.907, + 0.327 + ], + "angle": 0, + "content": "[63] Shao-Yuan Lo, Poojan Oza, and Vishal M Patel. Adversarially robust one-class novelty detection. In IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.33, + 0.907, + 0.372 + ], + "angle": 0, + "content": "[64] Ilya Loshchilov and Frank Hutter. SGDR: Stochastic gradient descent with warm restarts. In International Conference on Learning Representations, 2017. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.372, + 0.906, + 0.413 + ], + "angle": 0, + "content": "[65] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.414, + 0.907, + 0.495 + ], + "angle": 0, + "content": "[66] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.497, + 0.906, + 0.537 + ], + "angle": 0, + "content": "[67] Hui Lv and Qianru Sun. Video anomaly detection and explanation via large language models. arXiv preprint arXiv:2401.05702, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.539, + 0.908, + 0.609 + ], + "angle": 0, + "content": "[68] Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. Vt-adt: A vision transformer network for image anomaly detection and localization. In 2021 IEEE 30th International Symposium on Industrial Electronics (ISIE), pages 01-06. IEEE, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.609, + 0.908, + 0.678 + ], + "angle": 0, + "content": "[69] Shancong Mou, Xiaoyi Gu, Meng Cao, Haoping Bai, Ping Huang, Jiulong Shan, and Jianjun Shi. RGI: robust GAN-inversion for mask-free image inpainting and unsupervised pixel-wise anomaly detection. In The Eleventh International Conference on Learning Representations, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.679, + 0.907, + 0.748 + ], + "angle": 0, + "content": "[70] Ming Nie, Renyuan Peng, Chunwei Wang, Xinyue Cai, Jianhua Han, Hang Xu, and Li Zhang. Reason2drive: Towards interpretable and chain-based reasoning for autonomous driving. In European Conference on Computer Vision, pages 292-308. Springer, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.749, + 0.906, + 0.775 + ], + "angle": 0, + "content": "[71] OpenAI. Gpt-4v(isdiction) system card, 2023. https://openai.com/index/gpt-4v-system-card.5,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.776, + 0.908, + 0.815 + ], + "angle": 0, + "content": "[72] OpenAI. Gpt-4o system card, 2024. https://openai.com/index/gpt-4o-system-card. 2, 5, 6, 7, 8, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.818, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[73] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 7, 14" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20380" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.484, + 0.147 + ], + "angle": 0, + "content": "[74] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.15, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[75] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.179, + 0.484, + 0.233 + ], + "angle": 0, + "content": "[76] Tal Reiss and Yedid Hoshen. Mean-shifted contrastive loss for anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2155-2162, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.236, + 0.484, + 0.305 + ], + "angle": 0, + "content": "[77] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318-14328, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.307, + 0.484, + 0.375 + ], + "angle": 0, + "content": "[78] Fumiaki Sato, Ryo Hachiuma, and Taiki Sekii. Prompt-guided zero-shot anomaly action recognition using pretrained deep skeleton features. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6471-6480, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.378, + 0.484, + 0.445 + ], + "angle": 0, + "content": "[79] Eli Schwartz, Assaf Arbelle, Leonid Karlinsky, Sivan Harary, Florian Scheidegger, Sivan Doveh, and Raja Giryes. Mae-day: Mae for few-and zero-shot anomaly-detection. Computer Vision and Image Understanding, 241:103958, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.448, + 0.484, + 0.544 + ], + "angle": 0, + "content": "[80] Pierre Sermanet, Tianli Ding, Jeffrey Zhao, Fei Xia, Debidatta Dwibedi, Keerthana Gopalakrishnan, Christine Chan, Gabriel Dulac-Arnold, Sharath Maddineni, Nikhil J Joshi, et al. Robovqa: Multimodal long-horizon reasoning for robotics. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 645-652. IEEE, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.547, + 0.484, + 0.602 + ], + "angle": 0, + "content": "[81] Javier Silvestre-Blanes, Teresa Albero-Albero, Ignacio Miralles, Rubén Pérez-Llorens, and Jorge Moreno. A public fabric database for defect detection methods and results. Autex Research Journal, 19(4):363–374, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.604, + 0.484, + 0.659 + ], + "angle": 0, + "content": "[82] Luc PJ Sträter, Mohammadreza Salehi, Efstratios Gavves, Cees GM Snoek, and Yuki M Asano. Generalad: Anomaly detection across domains by attending to distorted features. arXiv preprint arXiv:2407.12427, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.661, + 0.484, + 0.701 + ], + "angle": 0, + "content": "[83] Haomiao Sun, Mingjie He, Tianheng Lian, Hu Han, and Shiguang Shan. Face-mllm: A large face perception model, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[84] Jiaqi Tang, Hao Lu, Xiaogang Xu, Ruizheng Wu, Sixing Hu, Tong Zhang, Tsz Wa Cheng, Ming Ge, Ying-Cong Chen, and Fugee Tsung. An incremental unified framework for small defect inspection. In European Conference on Computer Vision, pages 307-324. Springer, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[85] Tran Dinh Tien, Anh Tuan Nguyen, Nguyen Hoang Tran, Ta Duc Huy, Soan Duong, Chanh D Tr Nguyen, and Steven QH Truong. Revisiting reverse distillation for anomaly detection. In IEEE/CVF conference on computer vision and pattern recognition, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[86] Hualiang Wang, Yi Li, Huifeng Yao, and Xiaomeng Li. Clipn for zero-shot ood detection: Teaching clip to say no. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1802-1812, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.093, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[87] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.179, + 0.907, + 0.246 + ], + "angle": 0, + "content": "[88] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. In Conference on Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.249, + 0.907, + 0.316 + ], + "angle": 0, + "content": "[89] Qi Wei, Yinhao Ren, Rui Hou, Bibo Shi, Joseph Y Lo, and Lawrence Carin. Anomaly detection for medical images based on a one-class classification. In Medical Imaging 2018: Computer-Aided Diagnosis, pages 375-380. SPIE, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.32, + 0.907, + 0.388 + ], + "angle": 0, + "content": "[90] Julia Wolleb, Florentin Bieder, Robin Sandkühler, and Philippe C Cattin. Diffusion models for medical anomaly detection. In International Conference on Medical image computing and computer-assisted intervention, pages 35-45. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.391, + 0.907, + 0.459 + ], + "angle": 0, + "content": "[91] Binzhu Xie, Sicheng Zhang, Zitang Zhou, Bo Li, Yuanhan Zhang, Jack Hessel, Jingkang Yang, and Ziwei Liu. Funqa: Towards surprising video comprehension. In European Conference on Computer Vision, pages 39-57. Springer, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.462, + 0.906, + 0.517 + ], + "angle": 0, + "content": "[92] Guoyang Xie, Jinbao Wang, Jiaqi Liu, Yaochu Jin, and Feng Zheng. Pushing the limits of fewshot anomaly detection in industry vision: Graphcore. In The Eleventh International Conference on Learning Representations, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.519, + 0.907, + 0.601 + ], + "angle": 0, + "content": "[93] Hongxia Xie, Chu-Jun Peng, Yu-Wen Tseng, Hung-Jen Chen, Chan-Feng Hsu, Hong-Han Shuai, and Wen-Huang Cheng. Emovit: Revolutionizing emotion insights with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26596-26605, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.604, + 0.906, + 0.659 + ], + "angle": 0, + "content": "[94] Yuchen Yang, Kwonjoon Lee, Behzad Dariush, Yinzhi Cao, and Shao-Yuan Lo. Follow the rules: reasoning for video anomaly detection with large language models. arXiv preprint arXiv:2407.10299, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.661, + 0.907, + 0.729 + ], + "angle": 0, + "content": "[95] Hang Yao, Ming Liu, Haolin Wang, Zhicun Yin, Zifei Yan, Xiaopeng Hong, and Wangmeng Zuo. Glad: Towards better reconstruction with global and local adaptive diffusion models for unsupervised anomaly detection. arXiv preprint arXiv:2406.07487, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.732, + 0.907, + 0.786 + ], + "angle": 0, + "content": "[96] Huanjin Yao, Wenhao Wu, Taojiannan Yang, YuXin Song, Mengxi Zhang, Haocheng Feng, Yifan Sun, Zhiheng Li, Wanli Ouyang, and Jingdong Wang. Dense connector for mllms, 2024. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.789, + 0.907, + 0.844 + ], + "angle": 0, + "content": "[97] Xincheng Yao, Ruoqi Li, Zefeng Qian, Lu Wang, and Chongyang Zhang. Hierarchical gaussian mixture normalizing flow modeling for unified anomaly detection. arXiv preprint arXiv:2403.13349, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.846, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[98] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. In Advances in Neural Information Processing Systems, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20381" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[99] Vitjan Zavrtanik, Matej Kristan, and Danijel Skoçaj. Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In IEEE/CVF international conference on computer vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.204 + ], + "angle": 0, + "content": "[100] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.484, + 0.315 + ], + "angle": 0, + "content": "[101] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, Songyang Zhang, Wenwei Zhang, Yining Li, Yang Gao, Peng Sun, Xinyue Zhang, Wei Li, Jingwen Li, Wenhai Wang, Hang Yan, Conghui He, Xingcheng Zhang, Kai Chen, Jifeng Dai, Yu Qiao, Dahua Lin, and Jiaqi Wang. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.484, + 0.371 + ], + "angle": 0, + "content": "[102] Xuan Zhang, Shiyu Li, Xi Li, Ping Huang, Jiulong Shan, and Ting Chen. Destseg: Segmentation guided denoising studentteacher for anomaly detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.484, + 0.44 + ], + "angle": 0, + "content": "[103] Ximiao Zhang, Min Xu, Dehui Qiu, Ruixin Yan, Ning Lang, and Xiuzhuang Zhou. Mediclip: Adapting clip for few-shot medical image anomaly detection. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 458-468. Springer, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.484, + 0.51 + ], + "angle": 0, + "content": "[104] Ximiao Zhang, Min Xu, and Xiuzhuang Zhou. Realnet: A feature selection network with realistic synthetic anomaly for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16699-16708, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.484, + 0.566 + ], + "angle": 0, + "content": "[105] Zhuosheng Zhang, Aston Zhang, Mu Li, hai zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. Transactions on Machine Learning Research, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.484, + 0.635 + ], + "angle": 0, + "content": "[106] He Zhao, Yuexiang Li, Nanjun He, Kai Ma, Leyuan Fang, Huiqi Li, and Yefeng Zheng. Anomaly detection for medical images using self-supervised and translation-consistent features. IEEE Transactions on Medical Imaging, 40(12): 3641-3651, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.484, + 0.691 + ], + "angle": 0, + "content": "[107] Gengze Zhou, Yicong Hong, and Qi Wu. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7641-7649, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.484, + 0.746 + ], + "angle": 0, + "content": "[108] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.484, + 0.803 + ], + "angle": 0, + "content": "[109] Kaiwen Zhou, Kwonjoon Lee, Teruhisa Mitsu, and Xin Eric Wang. Vicor: Bridging visual understanding and commonsense reasoning with large language models. In Findings of the Association for Computational Linguistics, 202'. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[110] Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. AnomalyCLIP: Object-agnostic prompt learning for zero-shot anomaly detection. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.903 + ], + "angle": 0, + "content": "[111] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. MiniGPT-4: Enhancing vision-language" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "understanding with advanced large language models. In The Twelfth International Conference on Learning Representations, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.204 + ], + "angle": 0, + "content": "[112] Jiawen Zhu and Guansong Pang. Toward generalist anomaly detection via in-context residual learning with few-shot sample prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17826-17836, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.907, + 0.275 + ], + "angle": 0, + "content": "[113] Jiaqi Zhu, Shaofeng Cai, Fang Deng, Beng Chin Ooi, and Junran Wu. Do llms understand visual anomalies? uncovering llm's capabilities in zero-shot anomaly detection. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 48-57, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.907, + 0.318 + ], + "angle": 0, + "content": "[114] Jiawen Zhu, Yew-Soon Ong, Chunhua Shen, and Guansong Pang. Fine-grained abnormality prompt learning for zero-shot anomaly detection, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.907, + 0.389 + ], + "angle": 0, + "content": "[115] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In European Conference on Computer Vision, pages 392-408. Springer, 2022. 6, 15" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20382" + } + ] +] \ No newline at end of file diff --git a/2025/Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models/9698d3f1-26f6-4c2e-aef0-60a71f074ba9_origin.pdf b/2025/Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models/9698d3f1-26f6-4c2e-aef0-60a71f074ba9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8eae72ebf4ff4b66a779a9fa715b2132640270bc --- /dev/null +++ b/2025/Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models/9698d3f1-26f6-4c2e-aef0-60a71f074ba9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6450f755868e2711934435f2e2b2e3d1d32933181dd0fe57f294b21fd65944f3 +size 5366530 diff --git a/2025/Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models/full.md b/2025/Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f03722a9a63a789a3a2a2d3eece799af5a46ac9e --- /dev/null +++ b/2025/Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models/full.md @@ -0,0 +1,471 @@ +# Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models + +Jiacong Xu $^{1*}$ Shao-Yuan Lo $^{2}$ Bardia Safaei $^{1}$ Vishal M. Patel $^{1}$ Isht Dwivedi $^{2}$ $^{1}$ Johns Hopkins University $^{2}$ Honda Research Institute USA + +{jxu155, bsafaei1, vpatel36}@jhu.edu {shao-yuan.lo, idwivedi}@honda-ri.com + +# Abstract + +Zero-Shot Anomaly Detection (ZSAD) is an emerging AD paradigm. Unlike the traditional unsupervised AD setting that requires a large number of normal samples to train a model, ZSAD is more practical for handling data-restricted real-world scenarios. Recently, Multimodal Large Language Models (MLLMs) have shown revolutionary reasoning capabilities in various vision tasks. However, the reasoning of image abnormalities remains underexplored due to the lack of corresponding datasets and benchmarks. To facilitate research in AD & reasoning, we establish the first visual instruction tuning dataset, Anomaly-Instruct-125k, and the evaluation benchmark, VisA-D&R. Through investigation with our benchmark, we reveal that current MLLMs like GPT-4o cannot accurately detect and describe fine-grained anomalous details in images. To address this, we propose Anomaly-OneVision (Anomaly-OV), the first specialist visual assistant for ZSAD and reasoning. Inspired by human behavior in visual inspection, Anomaly-OV leverages a Look-Twice Feature Matching (LTFM) mechanism to adaptively select and emphasize abnormal visual tokens. Extensive experiments demonstrate that Anomaly-OV achieves significant improvements over advanced generalist models in both detection and reasoning. Extensions to medical and 3D AD are provided for future study. The link to our project page: https://xujiacong.github.io/Anomaly-OV/ + +# 1. Introduction + +Visual Anomaly Detection (AD) is a well-established task in computer vision, extensively applied in scenarios such as industrial defect inspection [2, 3, 5, 12, 35, 69, 76, 77, 92, 98] and medical image diagnosis [1, 24, 31, 36, 89, 90, 103, 106]. In the traditional unsupervised AD (a.k.a. one-class AD) setting, models learn the distribution of normal visual features from normal samples and are required to identify anomaly samples during inference. While recent advance + +![](images/87ca03e000aaeacd11d7caf00c0f3e4662be350fb651c38b989f04967522b4d5.jpg) +Figure 1. Visualization of the image-level AUROC comparison between our Anomaly-OV and current state-of-the-art ZSAD methods (WinCLIP [38], AnoVL [19], AnomalyCLIP [110], AdaCLIP [6]). Notably, our zero-shot performance on VisA even surpasses most recent advances in the few-shot setting [28, 51, 112]. + +ments [9, 25, 32–34, 37, 42, 82, 84, 95, 97, 104] have significantly improved the detection performance, these approaches assume the availability of a substantial number of normal samples. However, this assumption becomes impractical in certain scenarios due to strict data privacy policies and the significant human effort required for data classification, sometimes involving experts or specialists. Therefore, Zero-Shot Anomaly Detection (ZSAD) is emerging as a popular research direction, leading to the development of many innovative methods [6, 17, 27, 38, 43, 52, 78, 79, 110, 113]. + +Recent advances in Multimodal Large Language Models (MLLMs) [7, 15, 44, 45, 47, 48, 57, 58, 111] have shown revolutionary reasoning capabilities in various vision tasks [14, 29, 67, 70, 80, 91, 94, 107, 109]. However, the rea + +soning of image abnormalities has not been explored due to the challenges of collecting large-scale datasets and establishing benchmarks. Existing methods simply predict the likelihood of an anomaly without providing rationales [6, 11, 19, 38, 110]. In contrast, for better interpretability, robustness, and trustworthiness, people would expect models to explain why an image is considered anomalous and provide visual evidence. Interestingly, we find that recent advanced MLLMs, such as GPT-4o [72], fall short in AD & reasoning. As shown in Figure 2, while the detection is correct, the explanation from GPT-4o lacks accuracy, indicating a gap in a comprehensive understanding of the anomaly. + +To expedite research in AD & reasoning, we establish the first visual instruction tuning dataset, Anomaly-Instruct-125k, and the evaluation benchmark, VisA-D&R, through intensive human efforts. After evaluating current generalist MLLMs, we observe that these models fail to accurately detect and describe fine-grained anomalous details in images. To address this, we propose Anomaly-OneVision (Anomaly-OV), the first specialist visual assistant for ZSAD and reasoning. Unlike existing ZSAD methods [6, 11, 19, 38, 110], Anomaly-OV directly learns object-awareness abnormality embeddings in feature space using only the visual encoder. Inspired by human behavior in visual inspection, Anomaly-OV employs a Look-Twice Feature Matching (LTFM) mechanism to assist its LLM in adaptively selecting and emphasizing the most suspicious abnormal visual tokens. + +Extensive experiments demonstrate that Anomaly-OV achieves significant improvements over advanced generalist models in both detection and reasoning. Extended results of Anomaly-OV, from applications in industrial defect detection to 3D inspection and medical image diagnosis, are provided for future study. With precise descriptions and rationales of visual anomalies, our model can infer potential causes (see Figure 2), assess current impacts, and offer improvement suggestions, positioning itself as a reliable assistant for visual inspection. Our contributions are in two folds: + +- We establish the first visual instruction tuning dataset and benchmark for anomaly detection and reasoning. +- We propose the first specialist visual assistant with state-of-the-art performance for this new impactful domain. + +# 2. Related Work + +Multimodal Large Language Models. Vision-Language Models (VLMs), such as CLIP [73], exhibit robust zero-shot classification capabilities and have been applied to a range of downstream vision tasks [10, 53, 56, 61, 86]. Combining a VLM's vision encoder and an LLM [20, 62, 74], MLLMs [15, 48, 49, 57, 111] enable text-based interactions related to visual content. MLLMs have shown remarkable reasoning capability, particularly when incorporated with prompting strategies such as Chain-of-Thought [66, 88, 105]. Recent studies have harnessed MLLMs to provide reasoning for + +![](images/13bec8769ce795b7c1bf630673506d45bbe937ad7c185058bb030de265045c11.jpg) + +![](images/25d9ca1d82563a8f93cbf3b8f88b477c4c9f7175c783c82e46e24750d5b1726d.jpg) + +GT: The top-left candle shows a missing chunk of wax and an irregular indentation near the wick. + +Do you see any anomalies in the image? + +![](images/fa1ec3fecb2f2f7a21fcf15c025d1cc5e0202f2e27fadac7851d6ee400a6f176.jpg) + +![](images/1028b0d63a08272fee646e1912e0f6c44eeb5805de96a4139539c94247c752f7.jpg) + +The candle in the top left appears to have a damaged or missing wick, which is unusual compared to the other candles. + +![](images/01126e20ac9b600e067b4454d12d690d857d9b915c2df3b957d02dc35024e36f.jpg) +Figure 2. Industrial image anomaly reasoning results from GPT-4o [72] and our Anomaly-OV. The responses for fine-grained anomaly reasoning are highlighted, with the ground truth given for reference. + +Yes, there is an anomaly in the top left candle. It has a small hole or crack near the wick, which is not present in the other candles. + +What is the cause for the anomaly? + +![](images/0853c89fa1b7fdb473555b96d7d47ac50e03bb39bccc3193e3b271babe46203d.jpg) + +![](images/ec51fc591b957718275465597f05c2ab9414421124c387006f6117f901285a47.jpg) + +The anomaly could be due to manufacturing defects, such as improper molding or cooling, or it might result from external damage like a puncture. + +downstream tasks, e.g., video anomaly detection [67, 94], affective computing [14, 29], and visual commonsense reasoning [109], revealing more interpretability. + +Unsupervised Anomaly Detection. Due to the scarcity and difficulty of collecting anomalous data, researchers often focus on the unsupervised AD setting, which exclusively uses normal data to train an AD model. Earlier studies, such as reconstruction-based [63, 69, 99], student-teacher [18, 85, 102], and augmentation-based [46] approaches, assume a large amount of normal data is available. These traditional approaches are less practical when data are restricted or expensive, such as in the medical domain. + +Zero-Shot Anomaly Detection. Unlike unsupervised AD [35, 77] and few-shot AD [22, 28, 36, 51, 112], ZSAD models directly access the likelihood of abnormality for a given image without requiring data specific to the target object. Existing works [6, 19, 38, 110] accomplish ZSAD by comparing visual and textual features encoded by visual and text encoders of CLIP and constructing their positive (anomaly) and negative (normal) prompts in the format of: + +$$ +\begin{array}{l} \mathcal {P} ^ {+} = \left[ V _ {1} \right] \left[ V _ {2} \right] \dots \left[ V _ {n} \right] [ o b j e c t ] \\ \mathcal {P} ^ {-} = \left[ W _ {1} \right] \left[ W _ {2} \right] \dots \left[ W _ {n} \right] [ o b j e c t ] \\ \end{array} +$$ + +where $V_{i}$ and $W_{i}$ are handcrafted or learnable tokens, and object refers to the word object or the class name of the object. However, simply utilizing object to represent all kinds of objects cannot capture the class-awareness abnormality types. Also, for an intelligent visual assistant, the images should be totally blind to the user (object-agnostic). + +![](images/c9b0b96175b01f00b0c7a0ecc318ecfbc5b060ab17aea4571e60fc25539fed13.jpg) +Figure 3. Overview of the Anomaly-0V architecture. It consists of two training stages: (1) professional training for the anomaly expert, and (2) visual instruction tuning for anomaly detection and reasoning. Text and visual tokens are distinguished by different colors. + +# 3. Method + +# 3.1. Preliminary + +Training an MLLM from scratch demands extensive data and computational resources to align the visual and textual embedding spaces and develop robust instruction-following capabilities. Recent studies [23, 83, 93] reveal that pretrained MLLMs function as generalists, possessing a broad knowledge base but underperforming in specialized domains. Therefore, our goal is to introduce an auxiliary specialist or expert model designed to guide the generalist in selecting and utilizing critical visual tokens. This approach circumvents the need for large-scale pre-training while preserving the generalization capacity of the original model. + +We choose $LLaVA-OneVision$ [44] as our base MLLM because it is open-sourced and performs similarly to other commercial models. $LLaVA-OneVision$ follows the model architectures for LLaVA family [47, 57-59] and other generic MLLMs, which typically consist of three major components: Visual Encoder, Projector, and LLM. The visual encoder [73, 100] extracts the visual information from the raw images, the projector aligns the spaces of visual features with the word embedding, and the LLM is responsible for textual instruction processing and complex reasoning. Since the image resolution for CLIP pre-training is fixed, $LLaVA-OneVision$ leverages AnyRes with pooling strategy to scale + +up the input raw image resolution. Specifically, the high-resolution images are divided into a prototyped number of crops, and the visual encoder independently processes the image crops before final spatial pooling. + +# 3.2. Architecture Overview + +With the same image-splitting strategy AnyRes as LLaVA-OneVision, the input high-resolution image is split into several crops, and the new image set can be written as: + +$$ +\mathcal {I} = \left\{I _ {0}, I _ {1}, I _ {2}, \dots , I _ {n - 1} \right\} \tag {1} +$$ + +where $I_0$ is the resized original image and $I_{j\neq 0}$ refers to the image crops. As shown in Figure 3, the image set $\mathcal{I}$ will be processed by the visual encoder $\mathcal{F}_{\theta}$ to generate the final visual features $\{\mathbf{v}_j^o\}$ . Similar to AnomalyCLIP [110], we store the outputs for four selected layers in the ViT [21] to capture the image representations from different levels and apply four adapters to compress the feature dimension. Then, the extracted visual features can be written as: + +$$ +\mathbf {v} _ {j} ^ {i} = \mathcal {F} _ {\theta} ^ {i} (I _ {j}) \tag {2} +$$ + +where $i$ denotes the $i$ -th level and $j$ refers to the index of corresponding image in $\mathcal{I}$ . These multi-level features have been demonstrated to be effective in capturing fine-grained local semantics by recent works [6, 28, 110]. + +![](images/f22e345a2ef3107a4d55a4c39a16623ba30cea442cceb34a735f0e07cee87875.jpg) +Figure 4. Simulation of visual anomaly inspection by humans. + +The large-scale pre-trained CLIP models align the projection spaces of the textual and visual encoder. Therefore, the encoded image features already contain the class information required by ZSAD. To avoid human involvement in object classification and reduce the model complexity, we remove the heavy text encoder commonly utilized in existing works and let the visual model itself parse the information for suspicious classes or objects. Specifically, the output visual features for the original image $\mathbf{v}_0^o$ are leveraged to provide the global description of the target object or regions in the look-back path. With the multi-level features and the global embeddings, the LTFM module is responsible for the recognition and localization of suspicious tokens. + +Drawing inspiration from human visual inspection, where suspicious objects or regions are identified and then inspected closely (see Figure 4), we design the VT selector module for aggregating (zooming in) the crucial visual tokens and explicitly assisting the LLM in distinguishing these tokens from many irrelevant ones when dealing with instructions regarding anomaly detection and reasoning. Additionally, the original visual features are preserved to maintain the generalization capability of the base model on regular instructions, such as Can you describe the content of the image? + +# 3.3. Look-Twice Feature Matching + +Given the global object information $\mathbf{v}_0^o$ provided by the lookback path, we generate the class-awareness abnormality description by merging $\mathbf{v}_0^o$ with two learnable embeddings: $\mathbf{e}^{+}\in \mathbb{R}^{D}$ and $\mathbf{e}^{-}\in \mathbb{R}^{D}$ , where $+$ and $-$ indicate positive (anomalous) and negative (normal) patterns and $D$ is the embedding dimension. Specifically, a linear layer $\mathcal{T}_i^o$ is applied along the token dimension to select and fuse useful tokens from $\mathbf{v}_0^o$ , and then the fused vector will be concatenated with $\mathbf{e}^{+}$ and $\mathbf{e}^{-}$ independently and pass through two MLPs $\{\mathcal{G}_i^+, \mathcal{G}_i^-\}$ to generate the abnormality and normality descriptions $\{\mathbf{d}_i^+, \mathbf{d}_i^-\}$ , which can be represented by: + +$$ +\left\{\mathbf {d} _ {i} ^ {+}, \mathbf {d} _ {i} ^ {-} \right\} = \left\{ \begin{array}{l} \mathcal {G} _ {i} ^ {+} \left(\mathbf {e} ^ {+} \circ \mathcal {T} _ {i} ^ {o} \left(\mathbf {v} _ {0} ^ {o}\right)\right) \\ \mathcal {G} _ {i} ^ {-} \left(\mathbf {e} ^ {-} \circ \mathcal {T} _ {i} ^ {o} \left(\mathbf {v} _ {0} ^ {o}\right)\right) \end{array} \right. \tag {3} +$$ + +The visual features extracted from different levels of the ViT focus on different scales of semantics. Thus, the parameters of $\mathcal{T}_i^o$ and $\{\mathcal{G}_i^+, \mathcal{G}_i^-\}$ should be independent for different levels, where $i$ indicate the level number. + +Similar to the zero-shot classification mechanism of CLIP + +models, we calculate the possibilities of each patch token in $\mathbf{v}_j^i$ belonging to the anomalous patterns by combining cosine similarity and softmax operations: + +$$ +\mathbf {m} _ {j} ^ {i} = \frac {\exp \left(< \mathbf {d} _ {i} ^ {+} , \mathbf {v} _ {j} ^ {i} > / \tau\right)}{\exp \left(< \mathbf {d} _ {i} ^ {+} , \mathbf {v} _ {j} ^ {i} > / \tau\right) + \exp \left(< \mathbf {d} _ {i} ^ {-} , \mathbf {v} _ {j} ^ {i} > / \tau\right)} \tag {4} +$$ + +where $\mathbf{m}_j^i$ represents the significance map for visual tokens, $\tau$ is the temperature hyperparameter, and $<,>$ refers to the cosine similarity operator. The patch weight in $\mathbf{m}_j^i$ indicates the closeness of the corresponding visual token to the anomalous pattern. Then, all the maps are averaged to capture the token significances from low to high levels: + +$$ +\mathbf {m} _ {j} = \sum_ {i = 1} ^ {4} \mathbf {m} _ {j} ^ {i} / 4 \tag {5} +$$ + +The visual features are leveraged twice in the forward and look-back paths, so this module is named by Look-Twice Feature Matching (LTFM), following the nature of two-step human visual inspection shown in Figure 4. + +# 3.4. Visual Token Selector + +Under the image cropping strategy widely applied in recent MLLMs, there will be a large number of visual tokens for a high-resolution image, e.g., 7290 tokens for an image with $1152 \times 1152$ resolution in LLaVA-OneVision. While these tokens provide rich visual details, the LLM is required to pick the most useful information when adapting to a specific task. When the LLM lacks enough knowledge in this domain, the token-picking process will become complicated. Thus, our solution is to introduce a specialist or expert who knows which token is crucial or not and assist the LLM in selecting and emphasizing (zooming in) the crucial tokens. + +Given the encoded visual tokens $\{\mathbf{v}_j^o\}$ for each image crop in $\mathcal{I}$ and the corresponding significance map $\mathbf{m}_j$ , the suspicious tokens are emphasized by direct multiplication of the two tensors. Then, the normal tokens will be scaled to zero while the anomalous tokens will be maintained. After that, spatial average pooling $\mathcal{P}$ is applied to reduce the number of tokens. This process can be written as: + +$$ +\mathbf {q} _ {j} = \mathcal {P} \left(\mathbf {v} _ {j} ^ {o} \odot \mathbf {m} _ {j}\right) \tag {6} +$$ + +where $\mathbf{q}_j\in \mathbb{R}^{h\times w\times D}$ refers to the pooled query tokens. Empirically, setting $h = w = 2$ provides a better trade-off than other options. Then, a Q-Former $\mathcal{Q}$ [49] is leveraged to aggregate the correlated tokens in the original output by forwarding $\mathbf{q}_j$ as the query and $\mathbf{v}_j^o$ as the key and value: + +$$ +\mathbf {v} _ {j} ^ {s} = \mathcal {Q} \left(\mathbf {q} _ {j}, \mathbf {v} _ {j} ^ {o}, \mathbf {v} _ {j} ^ {o}\right) \tag {7} +$$ + +The Visual Token Selector (VT Selector) serves as a tool for the anomaly expert to hand-pick visual tokens that contain the most suspicious semantics for a given image. + +Figure 5. Composition of the instruction data in Anomaly-Instruct-125k. There are four main types of image samples: in-the-wild, industrial, medical, and $3D$ (in the format of multi-view images), covering most image anomaly detection tasks and enabling the possibility of a unified assistant for visual inspection. The reasoning words are highlighted in blue. For more information about dataset establishment, statistics, and the data collection pipeline, please refer to Section A1 in the supplementary. +![](images/27870aa2e157a6e267299eafa2aaa3056f439cbc70d1a89ab877c999c695d5e7.jpg) +Q: Do you notice any anomalies in the image? A: Yes, the anomaly in the image is a fine, dark crack visible on the ceramic cup, starting from the rim and extending downward. +Q: How can I prevent similar anomalies in the future? A: Avoid rapid temperature changes, handle the cup carefully to minimize impacts, and consider using cups made with thicker, more durable materials. + +![](images/1cae4e97724f842ce372d6f3700a1026cebd396a15bf77bfa7c68dac0e191126.jpg) +Q: Are there any anomalies in this cable image? A: Yes, the green wire is missing its internal copper conductors, unlike the blue and brown wires, which contain complete sets of bundled copper strands. +Q: What could be the reason behind this anomaly? A: The absence of copper in the green wire could be due to manufacturing defects such as failure to insert the strands or a mechanical issue where the strands were removed or misplaced during production or handling. + +![](images/236f9ae2d43bd439000f023802369ecec417947d1e99b3582af1ac9171fb9000.jpg) +Q: What does the image show? A: The image shows a densely cellular tissue sample with features indicating potential metastatic breast cancer. +Q: What suggests that there might be malignant activity in the tissue sample? A: Malignant activity is suggested by irregularly shaped cells, hyperchromatic nuclei, and an increased nuclear-to-cytoplasmic ratio, indicative of abnormal and increased DNA content typical in cancerous cells. + +![](images/2e43f15b428805c298907b2430f159ad648f7a757bfd215f12ca344cd3fa05b1.jpg) +Q: What kind of object are you examining in the images? A: I'm examining a round cookie with a golden-brown surface and visible chocolate chips embedded across its surface. Q: Are there any significant anomalies detected on this cookie? A: Yes, there are two significant anomalies on the cookie: a central defect featuring a noticeable hole and an edge defect showing a gap or crack. + +# 3.5. Inference and Loss + +Anomaly Prediction. In the traditional anomaly detection task, the model predicts the possibility of the image being abnormal. To achieve anomaly score prediction, we aggregate the anomaly information from all the image crops by an average operation weighted on the significance maps: + +$$ +\mathbf {r} (\mathcal {I}) = \frac {\sum_ {j , k} \mathbf {v} _ {j} ^ {\mathrm {s}} [ k ] \cdot \mathcal {P} (\mathbf {m} _ {j}) [ k ]}{\sum_ {j , k} \mathcal {P} (\mathbf {m} _ {j}) [ k ]} \tag {8} +$$ + +where $\mathcal{P}$ is the same spatial pooling in VT Selector and $\mathbf{r}(\mathcal{I})$ is a vector containing the global anomaly information for the entire image. Then, the anomaly expert can calculate the image-level abnormal possibility by parsing $\mathbf{r}(\mathcal{I})$ : + +$$ +\operatorname {s c o r e} (\mathcal {I}) = \operatorname {S i g m o i d} \left(\mathcal {G} ^ {o} (\mathbf {r} (\mathcal {I}))\right) \tag {9} +$$ + +where $\mathcal{G}^o$ is an MLP for distinguishing normal and abnormal semantics. To handle the unbalanced sample distribution, we employ the balanced BCE loss as the professional training objective for the anomaly expert components. + +Text Generation. Instead of directly forwarding the concatenation of the original $\{\mathbf{v}_j^o\}$ and the selected $\{\mathbf{r}(\mathcal{I}),\mathbf{v}_j^s\}$ visual tokens into the LLM, we apply an indication prompt with suspicious feature: in the middle of the two series of tokens, which will highlight the selected tokens for LLM when handling anomaly-related instructions. This approach can be considered a form of prompt engineering in MLLMs. Besides, the is chosen from {highly, moderately, slightly} and is determined by score $(\mathcal{I})$ and predefined thresholds $\{\mathbf{s}_{low},\mathbf{s}_{high}\}$ . When the input image $\mathcal{I}$ has a high likelihood of anomaly, the LLM will place greater emphasis on the selected tokens; otherwise, these tokens will have less significance. The text generation is implemented by the original auto-regressive token prediction mechanism of LLM: + +$$ +p \left(X _ {a} \mid \mathcal {I}, X _ {q}\right) = \prod_ {t = 1} ^ {L} p _ {\theta} \left(x _ {t} \mid \mathcal {I}, X _ {q, < t}, X _ {a, < t}\right) \tag {10} +$$ + +where $X_{a, < t}$ and $X_{q, < t}$ are the answer and instruction tokens from all prior turns before the current prediction token $x_{t}$ for a sequence of length $L$ . The entire model is parameterized by $\theta$ and trained by the original language model cross-entropy loss for each predicted answer token $x_{t}$ . + +# 4. Dataset and Benchmark + +The lack of multimodal instruction-following data for image anomaly detection and reasoning hinders the development of special intelligent assistants in this domain. Even though AnomalyGPT [28] introduces a prompt tuning dataset by simulating the anomalies, the scale of their dataset and the diversity of their instructions and answers are limited, only focusing on anomaly localization. To resolve the data scarcity issue, we establish the first large-scale instruction tuning dataset: Anomaly-Instruct-125k and the corresponding anomaly detection and reasoning benchmark: VisA-D&R. + +# 4.1. Anomaly-Instruct-125k + +LLaVA [57] builds its instruction tuning dataset by leveraging the image caption and bounding boxes available in the COCO dataset[55] to prompt the text-only GPT-4. ShareGPT4V [8] provides a higher-quality dataset by directly prompting GPT-4V [71]. However, there is no image caption provided in existing anomaly detection datasets [1, 2], and no matter GPT-4V [71] or most recent GPT-4o [72] cannot accurately locate and describe the anomalies in the image without explicit human involvement. + +To resolve these issues, we design a new prompt pipeline for accurate anomaly description generation. Since most of the datasets contain annotations for anomaly types, we manually combine the class name and anomaly type, such as a [capsule] with [poke] on surface. If the anomaly masks are provided, we draw bounding boxes on the images to highlight the anomalous area. The short description and the image with (or w/o) bounding boxes are used to prompt GPT-4o to generate the detailed image and anomaly descriptions. Then, we employ an in-context learning strategy similar to LLaVA to create the instructions. + +# Detection: + +Q: Are there any defects for the object in the image? Please reply with 'Yes' or 'No'. + +# Reasoning: + +Q1: Do you observe any anomalies in the image? +Q2: Can you describe the anomalies you observed? +Q3: What is the potential cause for the anomalies? +Q4: How can such anomalies be prevented in the future? + +Figure 6. Prompt examples in VisA-D&R for detection and reasoning. The complex reasoning instructions are highlighted. + +For a unified visual inspection dataset, precise instruction data is collected from MVtec AD [2], the training set of BMAD [1], Anomaly-ShapeNet [50], Real3D-AD [60], and MVtec-3D AD [4], covering both 2D to 3D data across industry to medical domains. The 3D point cloud data are converted into 9 multi-view images, and the corresponding masks are rendered using predefined camera positions. However, the diversities and scales of these datasets are relatively limited, probably due to the difficulty of collecting anomaly images. To scale up the instruction data, we introduce an automatic anomaly data collection pipeline combining GPT4o [72] and Google Image Search [26] for image collection, data cleaning, and instruction generation. Finally, 72k in-the-wild images (named as WebAD) targeting anomaly detection are collected, significantly enriching our instruction dataset. Several samples from Anomaly-Instruct-125k are shown in Figure 5. The instructions are mainly in the format of multi-round conversations, covering anomaly detection and description in low-level reasoning and potential cause and future suggestions for complex understanding. + +# 4.2.VisA-D&R + +VisA [115] is a classic but challenging industrial anomaly detection dataset, providing fine-grained anomaly type and segmentation for each image. For evaluation of the anomaly detection and reasoning performance on existing and future methods, we select 10 classes from VisA and follow a similar data generation pipeline of Anomaly-Instruct-125k to create the benchmark. Differently, significant human effort has been invested in meticulously reviewing all generated images and anomaly descriptions. Wrong descriptions are picked out and re-annotated by humans before utilizing them for Q&A generation. Totally, the benchmark consists of 761 normal samples and 1000 anomalous ones. + +For evaluating detection performance, questions designed to elicit a one-word answer are used to prompt the MLLMs (Figure 6), with results quantified using Accuracy, Precision, Recall, and F1-score. We divide the reasoning performance into two parts: low-level reasoning that focuses on the description of visual defects or anomalies and complex reasoning requiring the MLLMs to provide the potential cause and future improvement strategies for the detected anomalies, where ROUGE-L [54], Sentence-BERT (SBERT) [75], and + +GPT-Score (GPT-4 as the judge [57]) are utilized to quantify the similarity between generated text and ground truth. Note that low-level reasoning is highly correlated to detection performance, while anomaly-type descriptions of low-level reasoning determine the output of complex reasoning. + +# 5. Experiments + +# 5.1. Training & Evaluation + +There are two independent training stages for Anomaly-OV. In Stage 1, the components of the anomaly expert are trained to obtain the token selection capability, targeting traditional ZSAD. This stage utilizes all of the data with anomaly labels in Anomaly-Instruct-125k. Similar to previous works [6, 110], when evaluating the model on the datasets contained in the training set, the corresponding datasets are replaced by VisA [115]. In Stage 2, the anomaly expert and visual encoder are frozen, while the projector and LLM are trainable. In addition to our instruction dataset, we sample around $350\mathrm{k}$ data from the original training recipe of $LLaVA-OneVision$ to maintain the generalization ability. For more details on training, please refer to the supplementary. + +The ZSAD performance for the anomaly expert is evaluated on nine benchmarks, including MVTec AD [2], VisA [115], AITEX [81], ELPV [16], BTAD [68], and MPDD [39] for industrial inspection, and BrainMRI [40], HeadCT [41], and Br35H [30] for medical diagnosis. AUROC (Area Under the Receiver Operating Characteristic) is leveraged to quantify the image-level AD performance. For text-based anomaly detection, both normal and anomaly data are employed to assess the accuracy by examining if the generated text contains the word Yes. Differently, only anomaly data are utilized to prompt the MLLMs to determine their anomaly reasoning capabilities since the justifications of normality are similar for different models. + +
MethodMVTecVisAHeadCTBrainMRI
Full Model94.091.197.693.9
w/o. Look-back92.890.596.693.5
w/o. e+ & e-92.190.194.792.9
w/o. Q-Former91.789.992.895.1
w/o. WebAD88.588.991.293.4
+ +Table 1. Ablation study for the anomaly expert of Anomaly-OV. w/o. Look-back refers to the removal of $\mathbf{v}_0^o$ in LTFM. + +# 5.2. Zero-Shot Anomaly Detection + +As shown in Table 2, compared with existing methods, the anomaly expert of Anomaly-OV achieves significant image-level AUROC improvements on most of the ZSAD benchmarks, which demonstrates that the text encoder widely applied in existing models is not necessary. The success of our model mainly originates from the extra data of WebAD (Table 1), which enables the model to learn more generic + +
ModelIndustrial DefectsMedical AnomaliesAverage
MVTec ADVisAAITEXELPVBTADMPDDBrainMRIHeadCTBr35H
CLIP [73]74.166.471.059.234.554.373.956.578.463.1
CoOp [108]88.862.866.273.066.855.161.378.486.070.9
WinCLIP [38]91.878.873.074.068.263.692.690.080.579.2
APRIL-GAN [11]86.278.057.665.573.673.089.389.193.178.4
AnoVL [19]92.579.272.570.680.368.988.781.688.480.3
AnomalyCLIP [110]91.582.162.281.588.377.090.393.494.684.5
AdaCLIP [6]89.285.864.579.788.676.094.891.497.785.3
Ours94.091.172.083.089.081.793.997.695.588.6
+ +semantics for normality and abnormality from the data distribution in the absence of the text encoder. This observation also reveals that large-scale in-the-wild online data can benefit zero-shot performance in anomaly detection. + +While the Q-Former reduces the model performance on BrainMRI, it shows effectiveness on most benchmarks, indicating the importance of token aggregation. Similarly, the look-back information and two learnable embeddings are required for describing class-awareness abnormality and distinguishing positive and negative features, respectively. As previously discussed, the anomaly expert is responsible for selecting suspicious visual tokens for the LLM, and the significance maps shown in Figure 7 demonstrate the interpretable token selection mechanism. The high intensities are automatically distributed around the anomalous areas even without any supervision of the anomaly masks. + +![](images/64f50cfa285d9f428edf3462c2cab1e9bdb560c5289d050d2012e58827b12991.jpg) + +![](images/0981baf11f7c8786c1a1a03b1ed604d3cfd7543e706fbc85a5b2ca1336e324fb.jpg) +Figure 7. Visualization of the significance map on VisA samples. + +![](images/67319f985b80c7763a079f5228d21f9acfc0ba797304cbc6bb0961b1e5991746.jpg) + +![](images/fc0d1a80b23d087fe076abffd6415aab48435be08723deaadd71ed472a66b3e1.jpg) + +![](images/7c5d82ad91c401f62339a88ed2869ef5c6f00df1516237dd8d3ed6a69835e079.jpg) + +![](images/447de3945edc43d51dfcbac8faf9e4d3338c3b8041cb0ddd06aba3772e56e6b4.jpg) + +# 5.3. Anomaly Detection & Reasoning + +With the strong capabilities of the anomaly expert for zero-shot detection and suspicious token selection, Anomaly-OV accomplishes significant improvement in text-based anomaly detection and reasoning over other open-sourced generalist MLLMs, as shown in Table 4. Here are a few observations: i) While a larger language model cannot guarantee better detection performance, it always provides greater reasoning ability; ii) Most of the existing MLLMs present much lower recall than precision, indicating their insensitivity to visual anomalies; iii) GPT-4o shows stronger reasoning ability + +Table 2. Quantitative comparison of Image-level AUROC on different ZSAD methods (some of the results are borrowed from [6, 110, 114]). The best and the second-best results are bolded and underlined, respectively. Please refer to the supplementary for more detailed results. + +
PCB Example: Excessive Solder
UserAre there any anomalies visible in the image?
GPT-4oThe image shows an HC-SR04 ultrasonic sensor. There don’t appear to be any visible anomalies with the sensor itself. The components and pins seem intact and properly aligned.
LLaVA-OVThe image shows an ultrasonic sensor module, specifically the HC-SR04 model. There are no visible anomalies in the image; it appears to be a standard representation of this type of sensor module.
Ours-7BYes, there is a noticeable anomaly in the image: a short circuit on the PCB, indicated by a white, crusty residue near the ‘Echo’ pin.
+ +Table 3. Anomaly-OV presents more accurate anomaly detection. + +compared to other open-sourced models. Table 3 and Table 5 provide the qualitative comparison of our Anomaly-OV with its base model LLaVA-OV-7B [44] and GPT-4o [72]. Both GPT-4o and LLaVA-OV show insensitivity to anomalous features and cannot accurately detect the anomaly in the image. Sometimes, GPT-4o knows the image is anomalous but fails to describe the anomalies precisely. + +We provide the fine-tuned version of the base model LLaVA-OV-0.5B on Anomaly-Instruct-125k, which presents much higher accuracy and more balanced precision and recall than its original version. This demonstrates the effectiveness of our instruction-tuning dataset. By integrating the anomaly expert with the base model, our Anomaly-OV-0.5B achieves 0.08 accuracy and 0.06 F1-score improvements in text-based anomaly detection and better reasoning capability in low-level and complex settings. Equipped with a larger language model, Anomaly-OV-7B provides the best detection performance among all the existing MLLMs and shows comparable reasoning ability with GPT-4o. Notably, we observe that the anomaly expert restricts the detection perfor + +
ModelAnomaly DetectionLow-level ReasoningComplex Reasoning
AccuracyPrecisionRecallF1-scoreROUGE-LSBERTGPT-ScoreSBERTGPT-Score
GPT-4V [71]0.680.900.490.550.160.653.310.775.64
GPT-4o [72]0.700.830.710.680.240.714.840.816.89
Qwen2-VL-2B [87]0.650.870.550.590.220.551.940.744.26
Qwen2-VL-7B [87]0.760.910.690.750.250.613.090.684.62
InternVL-2-8B [13]0.740.780.810.760.230.733.690.805.08
InternVL-2-26B [13]0.730.860.660.680.210.744.130.805.49
IXC-2.5-7B [101]0.720.880.630.670.210.582.450.775.14
LLaVA-OV-0.5B [44]0.540.700.190.280.200.632.540.814.34
LLaVA-OV-7B [44]0.710.950.560.630.240.663.570.795.44
LLaVA-OV-0.5B*0.710.770.840.760.310.703.690.825.31
Anomaly-OV-0.5B0.790.860.830.820.330.723.870.835.67
Anomaly-OV-7B0.790.830.860.830.340.734.260.846.34
+ +Table 4. Quantitative comparison of text-based anomaly detection and reasoning for MLLMs. Notably, the Accuracy and F1-score for the anomaly expert of Anomaly-OV can be calculated as $\{0.78, 0.77\}$ with threshold 0.5. * indicates the model is fine-tuned on our dataset. + +mance of Anomaly-OV. Therefore, the design of a stronger anomaly expert is suggested for future works. + +Macaroni Example: Yellowish Spot + +![](images/88fd3c7885aa08856f47465b158091edfd8ec145c9f1f282665e988598f21199.jpg) + +User +Do you see any anomalies in the image? + +
GPT-4oThe image shows four pieces of elbow macaroni on a green background. The anomaly is that the macaroni pieces are not whole; they are cut in half.
LLaVA-OVThe image shows four pieces of pasta, specifically macaroni shells, arranged on a green textured surface. The pasta appears to be uniformly colored and shaped, with no visible defects or anomalies.
Ours-7BYes, there is an anomaly in the image. The bottom right pasta piece has a noticeable yellowish discoloration on its surface.
+ +Table 5. Anomaly-OV presents more precise anomaly reasoning. + +# 5.4. Extension + +With the generalization and multi-image processing capabilities of MLLMs, it is possible to build a unified assistant for visual inspection. Table 6 demonstrates the comprehensive knowledge of Anomaly-OV (without using Anomaly-ShapeNet [50] for training) on 3D and medical (testing set of BMAD [1]) AD & reasoning. More data, benchmarks, and investigation on a unified model are meaningful. + +# 6. Conclusion + +In this paper, we establish the first large-scale visual instruction tuning dataset, Anomaly-Instruct-125k, and the corresponding benchmark, VisA-D&R, to address the data scarcity issue for visual anomaly detection and reasoning. + +3D Example: Bulge + +![](images/04a92329325abad74a691dc739a11fb8b70be3374afe8a3918381a73b2a06cd4.jpg) + +![](images/56fbef7823f2c266934b4fbe1b91fbc04875b8ebd7da10dcf49856622d26e33d.jpg) +··· + +User + +Is there any anomaly present in the bowl? + +Ours-7B + +Yes, there is a bulge on the side of the bowl, which disrupts its otherwise smooth and symmetrical surface. + +Medical Example: Pneumonia + +![](images/02534aa62c713e51f9815a7cd628b8ba19ef8c12742b7262babaea7aca80ebc9.jpg) + +User + +What is the most likely diagnosis for this patient? + +Ours-7B + +Based on the chest X-ray, the most likely diagnosis for this patient is pneumonia. The presence of bilateral patchy opacities in the lower lung fields is a common finding in pneumonia, which is an infection that causes inflammation in the air sacs of the lungs. + +User + +What should the patient do next? + +Ours-7B + +The patient should consult a healthcare professional for further evaluation and appropriate treatment. This may include antibiotics to treat the infection, supportive care such as oxygen therapy, and monitoring of the patient's condition. + +Table 6. Extension to 3D and medical AD & reasoning. + +Then, a specialist MLLM, Anomaly-OV, targeting visual inspection is introduced to serve as the baseline in this domain. Anomaly-OV leverages an anomaly expert to assist the LLM with suspicious visual token selection and presents significant improvements on both traditional ZSAD and text-based anomaly detection and reasoning tasks over existing methods. Extension to 3D and medical domains is demonstrated. + +# References + +[1] Jinan Bao, Hanshi Sun, Hanqiu Deng, Yinsheng He, Zhaoxiang Zhang, and Xingyu Li. Bmad: Benchmarks for medical anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4042-4053, 2024. 1, 5, 6, 8, 14 +[2] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592-9600, 2019. 1, 5, 6, 14, 15 +[3] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4183–4192, 2020. 1 +[4] Paul Bergmann, Xin Jin, David Sattlegger, and Carsten Steger. The mvtec 3d-ad dataset for unsupervised 3d anomaly detection and localization. In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications. SCITEPRESS - Science and Technology Publications, 2022. 6 +[5] Tri Cao, Jiawen Zhu, and Guansong Pang. Anomaly detection under distribution shift. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6511-6523, 2023. 1 +[6] Yunkang Cao, Jiangning Zhang, Luca Frittoli, Yuqi Cheng, Weiming Shen, and Giacomo Boracchi. Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly detection. In European Conference on Computer Vision, pages 55-72. Springer, 2025. 1, 2, 3, 6, 7 +[7] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. 1 +[8] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions, 2023. 5, 14 +[9] Qiyu Chen, Huiyuan Luo, Chengkan Lv, and Zhengtao Zhang. A unified anomaly synthesis strategy with gradient ascent for industrial anomaly detection and localization. arXiv preprint arXiv:2407.09359, 2024. 1 +[10] Runnan Chen, Youquan Liu, Lingdong Kong, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, Yu Qiao, and Wenping Wang. Clip2scene: Towards label-efficient 3d scene understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7030, 2023. 2 +[11] Xuhai Chen, Yue Han, and Jiangning Zhang. April-gan: A zero-/few-shot anomaly classification and segmentation method for cvpr 2023 vand workshop challenge tracks 1&2: 1st place on zero-shot ad and 4th place on few-shot ad, 2023. 2, 7 + +[12] Yuanhong Chen, Yu Tian, Guansong Pang, and Gustavo Carneiro. Deep one-class classification via interpolated gaussian descriptor. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 383-392, 2022. 1 +[13] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, Ji Ma, Jiaqi Wang, Xiaoyi Dong, Hang Yan, Hewei Guo, Conghui He, Botian Shi, Zhenjiang Jin, Chao Xu, Bin Wang, Xingjian Wei, Wei Li, Wenjian Zhang, Bo Zhang, Pinlong Cai, Licheng Wen, Xiangchao Yan, Min Dou, Lewei Lu, Xizhou Zhu, Tong Lu, Dahua Lin, Yu Qiao, Jifeng Dai, and Wenhai Wang. How far are we to gpt-4v? closing the gap to commercial multimodal models with opensource suites, 2024. 8 +[14] Zebang Cheng, Zhi-Qi Cheng, Jun-Yan He, Jingdong Sun, Kai Wang, Yuxiang Lin, Zheng Lian, Xiaojiang Peng, and Alexander Hauptmann. Emotion-llama: Multimodal emotion recognition and reasoning with instruction tuning. In Conference on Neural Information Processing Systems, 2024. 1, 2 +[15] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2 +[16] Sergiu Deitsch, Vincent Christlein, Stephan Berger, Claudia Buerhop-Lutz, Andreas Maier, Florian Gallwitz, and Christian Riess. Automatic classification of defective photovoltaic module cells in electroluminescence images. Solar Energy, 185:455-468, 2019. 6 +[17] Chenghao Deng, Haote Xu, Xiaolu Chen, Haodi Xu, Xiaotong Tu, Xinghao Ding, and Yue Huang. Simclip: Refining image-text alignment with simple prompts for zero-/few-shot anomaly detection. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 1761-1770, 2024. 1 +[18] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In IEEE/CVF conference on computer vision and pattern recognition, 2022. 2 +[19] Hanqiu Deng, Zhaoxiang Zhang, Jinan Bao, and Xingyu Li. Anovl: Adapting vision-language models for unified zero-shot anomaly localization. arXiv preprint arXiv:2308.15939, 2023. 1, 2, 7 +[20] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. 2 +[21] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 3 +[22] Zheng Fang, Xiaoyang Wang, Haocheng Li, Jiejie Liu, Qiugui Hu, and Jimin Xiao. Fastrecon: Few-shot industrial + +anomaly detection via fast feature reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17481-17490, 2023. 2 +[23] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J Black. Chatpose: Chatting about 3d human pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2093-2103, 2024. 3 +[24] Tharindu Fernando, Harshala Gammulle, Simon Denman, Sridha Sridharan, and Clinton Fookes. Deep learning for medical anomaly detection-a survey. ACM Computing Surveys (CSUR), 54(7):1-37, 2021. 1 +[25] Matic Fučka, Vitjan Zavrtanik, and Danijel Skočaj. Transfusion-a transparency-based diffusion model for anomaly detection. In European conference on computer vision, pages 91-108. Springer, 2025. 1 +[26] Google. Google-images-search 1.4.7, 2024. https://pypi.org/project/Google-Images-Search.6, 14 +[27] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Hao Li, Ming Tang, and Jinqiao Wang. Filo: Zero-shot anomaly detection by fine-grained description and high-quality localization. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2041-2049, 2024. 1 +[28] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. Anomalygpt: Detecting industrial anomalies using large vision-language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1932-1940, 2024. 1, 2, 3, 5 +[29] Yuxiang Guo, Faizan Siddiqui, Yang Zhao, Rama Chellappa, and Shao-Yuan Lo. Stimuvar: Spatiotemporal stimuli-aware video affective reasoning with multimodal large language models. arXiv preprint arXiv:2409.00304, 2024. 1, 2 +[30] Ahmed Hamada. Br35h: Brain tumor detection 2020, 2020. 6 +[31] Changhee Han, Leonardo Rundo, Kohei Murao, Tomoyuki Noguchi, Yuki Shimahara, Zoltan Adam Milacski, Saori Koshino, Evis Sala, Hideki Nakayama, and Shin'ichi Satoh. Madgan: Unsupervised medical anomaly detection gan using multiple adjacent brain mri slice reconstruction. BMC bioinformatics, 22:1-20, 2021. 1 +[32] Liren He, Zhengkai Jiang, Jinlong Peng, Liang Liu, Qianggang Du, Xiaobin Hu, Wenbing Zhu, Mingmin Chi, Yabiao Wang, and Chengjie Wang. Learning unified reference representation for unsupervised multi-class anomaly detection. arXiv preprint arXiv:2403.11561, 2024. 1 +[33] Chih-Hui Ho, Kuan-Chuan Peng, and Nuno Vasconcelos. Long-tailed anomaly detection with learnable class names. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12435-12446, 2024. +[34] Jinlei Hou, Yingying Zhang, Qiaoyong Zhong, Di Xie, Shiliang Pu, and Hong Zhou. Divide-and-assemble: Learning block-wise memory for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8791-8800, 2021. 1 +[35] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based + +few-shot anomaly detection. In European Conference on Computer Vision, pages 303-319. Springer, 2022. 1, 2 +[36] Chaoqin Huang, Aofan Jiang, Jinghao Feng, Ya Zhang, Xin chao Wang, and Yanfeng Wang. Adapting visual-language models for generalizable anomaly detection in medical images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11375-11385, 2024. 1, 2 +[37] Brian KS Isaac-Medina, Yona Falinie A Gaus, Neelanjan Bhowmik, and Toby P Breckon. Towards open-world object-based anomaly detection via self-supervised outlier synthesis. In European Conference on Computer Vision (ECCV), 2024. 1 +[38] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19606-19616, 2023. 1, 2, 7 +[39] Stepan Jezek, Martin Jonak, Radim Burget, Pavel Dvorak, and Milos Skotak. Deep learning-based defect detection of metal parts: evaluating current methods in complex conditions. In 2021 13th International congress on ultra modern telecommunications and control systems and workshops (ICUMT), pages 66-71. IEEE, 2021. 6 +[40] Pranita Balaji Kanade and PP Gumaste. Brain tumor detection using mri images. *Brain*, 3(2):146-150, 2015. 6 +[41] Felipe Campos Kitamura. Head ct - hemorrhage, 2018. 6 +[42] Mingyu Lee and Jongwon Choi. Text-guided variational image generation for industrial anomaly detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26519-26528, 2024. 1 +[43] Aodong Li, Chen Qiu, Marius Kloft, Padhraic Smyth, Maja Rudolph, and Stephan Mandt. Zero-shot anomaly detection via batch normalization. Advances in Neural Information Processing Systems, 36, 2024. 1 +[44] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer, 2024. 1, 3, 7, 8, 15 +[45] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. LLaVA-med: Training a large language-and-vision assistant for biomedicine in one day. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 1, 15 +[46] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In IEEE/CVF conference on computer vision and pattern recognition, 2021. 2 +[47] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava-last-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 1, 3 +[48] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Interna- + +tional conference on machine learning, pages 12888-12900. +PMLR, 2022. 1, 2 +[49] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2, 4 +[50] Wenqiao Li, Xiaohao Xu, Yao Gu, Bozhong Zheng, Shenghua Gao, and Yingna Wu. Towards scalable 3d anomaly detection and localization: A benchmark via 3d anomaly synthesis and a self-supervised learning network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22207-22216, 2024. 6, 8 +[51] Xiaofan Li, Zhizhong Zhang, Xin Tan, Chengwei Chen, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Prompt: Learning prompts with only normal samples for few-shot anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16838-16848, 2024. 1, 2 +[52] Yiting Li, Adam Goodge, Fayao Liu, and Chuan-Sheng Foo. Promptad: Zero-shot anomaly detection using text prompts. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1093-1102, 2024. 1 +[53] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 2 +[54] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6 +[55] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5 +[56] Yuqi Lin, Minghao Chen, Wenxiao Wang, Boxi Wu, Ke Li, Binbin Lin, Haifeng Liu, and Xiaofei He. Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15305-15314, 2023. 2 +[57] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2, 3, 5, 6, 16 +[58] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1 +[59] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024.3 +[60] Jiaqi Liu, Guoyang Xie, ruitao chen, Xinpeng Li, Jinbao Wang, Yong Liu, Chengjie Wang, and Feng Zheng. Real3d + +AD: A dataset of point cloud anomaly detection. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 6 +[61] Jie Liu, Yixiao Zhang, Jie-Neng Chen, Junfei Xiao, Yongyi Lu, Bennett A Landman, Yixuan Yuan, Alan Yuille, Yucheng Tang, and Zongwei Zhou. Clip-driven universal model for organ segmentation and tumor detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21152-21164, 2023. 2 +[62] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach, 2019. 2 +[63] Shao-Yuan Lo, Poojan Oza, and Vishal M Patel. Adversarially robust one-class novelty detection. In IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2 +[64] Ilya Loshchilov and Frank Hutter. SGDR: Stochastic gradient descent with warm restarts. In International Conference on Learning Representations, 2017. 15 +[65] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 15 +[66] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 2 +[67] Hui Lv and Qianru Sun. Video anomaly detection and explanation via large language models. arXiv preprint arXiv:2401.05702, 2024. 1, 2 +[68] Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. Vt-adt: A vision transformer network for image anomaly detection and localization. In 2021 IEEE 30th International Symposium on Industrial Electronics (ISIE), pages 01-06. IEEE, 2021. 6 +[69] Shancong Mou, Xiaoyi Gu, Meng Cao, Haoping Bai, Ping Huang, Jiulong Shan, and Jianjun Shi. RGI: robust GAN-inversion for mask-free image inpainting and unsupervised pixel-wise anomaly detection. In The Eleventh International Conference on Learning Representations, 2023. 1, 2 +[70] Ming Nie, Renyuan Peng, Chunwei Wang, Xinyue Cai, Jianhua Han, Hang Xu, and Li Zhang. Reason2drive: Towards interpretable and chain-based reasoning for autonomous driving. In European Conference on Computer Vision, pages 292-308. Springer, 2025. 1 +[71] OpenAI. Gpt-4v(isdiction) system card, 2023. https://openai.com/index/gpt-4v-system-card.5,8 +[72] OpenAI. Gpt-4o system card, 2024. https://openai.com/index/gpt-4o-system-card. 2, 5, 6, 7, 8, 14, 15 +[73] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 7, 14 + +[74] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer, 2023. 2 +[75] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks, 2019. 6 +[76] Tal Reiss and Yedid Hoshen. Mean-shifted contrastive loss for anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2155-2162, 2023. 1 +[77] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318-14328, 2022. 1, 2 +[78] Fumiaki Sato, Ryo Hachiuma, and Taiki Sekii. Prompt-guided zero-shot anomaly action recognition using pretrained deep skeleton features. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6471-6480, 2023. 1 +[79] Eli Schwartz, Assaf Arbelle, Leonid Karlinsky, Sivan Harary, Florian Scheidegger, Sivan Doveh, and Raja Giryes. Mae-day: Mae for few-and zero-shot anomaly-detection. Computer Vision and Image Understanding, 241:103958, 2024. 1 +[80] Pierre Sermanet, Tianli Ding, Jeffrey Zhao, Fei Xia, Debidatta Dwibedi, Keerthana Gopalakrishnan, Christine Chan, Gabriel Dulac-Arnold, Sharath Maddineni, Nikhil J Joshi, et al. Robovqa: Multimodal long-horizon reasoning for robotics. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 645-652. IEEE, 2024. 1 +[81] Javier Silvestre-Blanes, Teresa Albero-Albero, Ignacio Miralles, Rubén Pérez-Llorens, and Jorge Moreno. A public fabric database for defect detection methods and results. Autex Research Journal, 19(4):363–374, 2019. 6 +[82] Luc PJ Sträter, Mohammadreza Salehi, Efstratios Gavves, Cees GM Snoek, and Yuki M Asano. Generalad: Anomaly detection across domains by attending to distorted features. arXiv preprint arXiv:2407.12427, 2024. 1 +[83] Haomiao Sun, Mingjie He, Tianheng Lian, Hu Han, and Shiguang Shan. Face-mllm: A large face perception model, 2024. 3 +[84] Jiaqi Tang, Hao Lu, Xiaogang Xu, Ruizheng Wu, Sixing Hu, Tong Zhang, Tsz Wa Cheng, Ming Ge, Ying-Cong Chen, and Fugee Tsung. An incremental unified framework for small defect inspection. In European Conference on Computer Vision, pages 307-324. Springer, 2025. 1 +[85] Tran Dinh Tien, Anh Tuan Nguyen, Nguyen Hoang Tran, Ta Duc Huy, Soan Duong, Chanh D Tr Nguyen, and Steven QH Truong. Revisiting reverse distillation for anomaly detection. In IEEE/CVF conference on computer vision and pattern recognition, 2023. 2 +[86] Hualiang Wang, Yi Li, Huifeng Yao, and Xiaomeng Li. Clipn for zero-shot ood detection: Teaching clip to say no. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1802-1812, 2023. 2 + +[87] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution, 2024. 8 +[88] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. In Conference on Neural Information Processing Systems, 2022. 2 +[89] Qi Wei, Yinhao Ren, Rui Hou, Bibo Shi, Joseph Y Lo, and Lawrence Carin. Anomaly detection for medical images based on a one-class classification. In Medical Imaging 2018: Computer-Aided Diagnosis, pages 375-380. SPIE, 2018. 1 +[90] Julia Wolleb, Florentin Bieder, Robin Sandkühler, and Philippe C Cattin. Diffusion models for medical anomaly detection. In International Conference on Medical image computing and computer-assisted intervention, pages 35-45. Springer, 2022. 1 +[91] Binzhu Xie, Sicheng Zhang, Zitang Zhou, Bo Li, Yuanhan Zhang, Jack Hessel, Jingkang Yang, and Ziwei Liu. Funqa: Towards surprising video comprehension. In European Conference on Computer Vision, pages 39-57. Springer, 2025. 1 +[92] Guoyang Xie, Jinbao Wang, Jiaqi Liu, Yaochu Jin, and Feng Zheng. Pushing the limits of fewshot anomaly detection in industry vision: Graphcore. In The Eleventh International Conference on Learning Representations, 2023. 1 +[93] Hongxia Xie, Chu-Jun Peng, Yu-Wen Tseng, Hung-Jen Chen, Chan-Feng Hsu, Hong-Han Shuai, and Wen-Huang Cheng. Emovit: Revolutionizing emotion insights with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26596-26605, 2024. 3 +[94] Yuchen Yang, Kwonjoon Lee, Behzad Dariush, Yinzhi Cao, and Shao-Yuan Lo. Follow the rules: reasoning for video anomaly detection with large language models. arXiv preprint arXiv:2407.10299, 2024. 1, 2 +[95] Hang Yao, Ming Liu, Haolin Wang, Zhicun Yin, Zifei Yan, Xiaopeng Hong, and Wangmeng Zuo. Glad: Towards better reconstruction with global and local adaptive diffusion models for unsupervised anomaly detection. arXiv preprint arXiv:2406.07487, 2024. 1 +[96] Huanjin Yao, Wenhao Wu, Taojiannan Yang, YuXin Song, Mengxi Zhang, Haocheng Feng, Yifan Sun, Zhiheng Li, Wanli Ouyang, and Jingdong Wang. Dense connector for mllms, 2024. 17 +[97] Xincheng Yao, Ruoqi Li, Zefeng Qian, Lu Wang, and Chongyang Zhang. Hierarchical gaussian mixture normalizing flow modeling for unified anomaly detection. arXiv preprint arXiv:2403.13349, 2024. 1 +[98] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. In Advances in Neural Information Processing Systems, 2022. 1 + +[99] Vitjan Zavrtanik, Matej Kristan, and Danijel Skoçaj. Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In IEEE/CVF international conference on computer vision, 2021. 2 +[100] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3 +[101] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, Songyang Zhang, Wenwei Zhang, Yining Li, Yang Gao, Peng Sun, Xinyue Zhang, Wei Li, Jingwen Li, Wenhai Wang, Hang Yan, Conghui He, Xingcheng Zhang, Kai Chen, Jifeng Dai, Yu Qiao, Dahua Lin, and Jiaqi Wang. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output, 2024. 8 +[102] Xuan Zhang, Shiyu Li, Xi Li, Ping Huang, Jiulong Shan, and Ting Chen. Destseg: Segmentation guided denoising studentteacher for anomaly detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2 +[103] Ximiao Zhang, Min Xu, Dehui Qiu, Ruixin Yan, Ning Lang, and Xiuzhuang Zhou. Mediclip: Adapting clip for few-shot medical image anomaly detection. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 458-468. Springer, 2024. 1 +[104] Ximiao Zhang, Min Xu, and Xiuzhuang Zhou. Realnet: A feature selection network with realistic synthetic anomaly for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16699-16708, 2024. 1 +[105] Zhuosheng Zhang, Aston Zhang, Mu Li, hai zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. Transactions on Machine Learning Research, 2024. 2 +[106] He Zhao, Yuexiang Li, Nanjun He, Kai Ma, Leyuan Fang, Huiqi Li, and Yefeng Zheng. Anomaly detection for medical images using self-supervised and translation-consistent features. IEEE Transactions on Medical Imaging, 40(12): 3641-3651, 2021. 1 +[107] Gengze Zhou, Yicong Hong, and Qi Wu. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7641-7649, 2024. 1 +[108] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 7 +[109] Kaiwen Zhou, Kwonjoon Lee, Teruhisa Mitsu, and Xin Eric Wang. Vicor: Bridging visual understanding and commonsense reasoning with large language models. In Findings of the Association for Computational Linguistics, 202'. 1, 2 +[110] Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. AnomalyCLIP: Object-agnostic prompt learning for zero-shot anomaly detection. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 6, 7 +[111] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. MiniGPT-4: Enhancing vision-language + +understanding with advanced large language models. In The Twelfth International Conference on Learning Representations, 2024. 1, 2 +[112] Jiawen Zhu and Guansong Pang. Toward generalist anomaly detection via in-context residual learning with few-shot sample prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17826-17836, 2024. 1, 2 +[113] Jiaqi Zhu, Shaofeng Cai, Fang Deng, Beng Chin Ooi, and Junran Wu. Do llms understand visual anomalies? uncovering llm's capabilities in zero-shot anomaly detection. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 48-57, 2024. 1 +[114] Jiawen Zhu, Yew-Soon Ong, Chunhua Shen, and Guansong Pang. Fine-grained abnormality prompt learning for zero-shot anomaly detection, 2024. 7 +[115] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In European Conference on Computer Vision, pages 392-408. Springer, 2022. 6, 15 \ No newline at end of file diff --git a/2025/Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models/images.zip b/2025/Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ce48802181ce5faf4dd4fad0a230ed03d13c62a8 --- /dev/null +++ b/2025/Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00ac7d3028648fc096f40f073b0c3cac593afe12897bbcc63e354807ae83ebf2 +size 623487 diff --git a/2025/Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models/layout.json b/2025/Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d89d18a838902ff83d2ab34940eff30a2f332ab1 --- /dev/null +++ b/2025/Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models/layout.json @@ -0,0 +1,12196 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 137, + 103, + 475, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 103, + 475, + 140 + ], + "spans": [ + { + "bbox": [ + 137, + 103, + 475, + 140 + ], + "type": "text", + "content": "Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 160, + 499, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 160, + 499, + 190 + ], + "spans": [ + { + "bbox": [ + 111, + 160, + 499, + 190 + ], + "type": "text", + "content": "Jiacong Xu" + }, + { + "bbox": [ + 111, + 160, + 499, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 111, + 160, + 499, + 190 + ], + "type": "text", + "content": " Shao-Yuan Lo" + }, + { + "bbox": [ + 111, + 160, + 499, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 160, + 499, + 190 + ], + "type": "text", + "content": " Bardia Safaei" + }, + { + "bbox": [ + 111, + 160, + 499, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 160, + 499, + 190 + ], + "type": "text", + "content": " Vishal M. Patel" + }, + { + "bbox": [ + 111, + 160, + 499, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 160, + 499, + 190 + ], + "type": "text", + "content": " Isht Dwivedi" + }, + { + "bbox": [ + 111, + 160, + 499, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 160, + 499, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 160, + 499, + 190 + ], + "type": "text", + "content": "Johns Hopkins University " + }, + { + "bbox": [ + 111, + 160, + 499, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 160, + 499, + 190 + ], + "type": "text", + "content": "Honda Research Institute USA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 101, + 191, + 506, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 191, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 101, + 191, + 506, + 205 + ], + "type": "text", + "content": "{jxu155, bsafaei1, vpatel36}@jhu.edu {shao-yuan.lo, idwivedi}@honda-ri.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 255, + 297, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 297, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 297, + 555 + ], + "type": "text", + "content": "Zero-Shot Anomaly Detection (ZSAD) is an emerging AD paradigm. Unlike the traditional unsupervised AD setting that requires a large number of normal samples to train a model, ZSAD is more practical for handling data-restricted real-world scenarios. Recently, Multimodal Large Language Models (MLLMs) have shown revolutionary reasoning capabilities in various vision tasks. However, the reasoning of image abnormalities remains underexplored due to the lack of corresponding datasets and benchmarks. To facilitate research in AD & reasoning, we establish the first visual instruction tuning dataset, Anomaly-Instruct-125k, and the evaluation benchmark, VisA-D&R. Through investigation with our benchmark, we reveal that current MLLMs like GPT-4o cannot accurately detect and describe fine-grained anomalous details in images. To address this, we propose Anomaly-OneVision (Anomaly-OV), the first specialist visual assistant for ZSAD and reasoning. Inspired by human behavior in visual inspection, Anomaly-OV leverages a Look-Twice Feature Matching (LTFM) mechanism to adaptively select and emphasize abnormal visual tokens. Extensive experiments demonstrate that Anomaly-OV achieves significant improvements over advanced generalist models in both detection and reasoning. Extensions to medical and 3D AD are provided for future study. The link to our project page: https://xujiacong.github.io/Anomaly-OV/" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 578, + 135, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 135, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 135, + 590 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 598, + 297, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 297, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 297, + 696 + ], + "type": "text", + "content": "Visual Anomaly Detection (AD) is a well-established task in computer vision, extensively applied in scenarios such as industrial defect inspection [2, 3, 5, 12, 35, 69, 76, 77, 92, 98] and medical image diagnosis [1, 24, 31, 36, 89, 90, 103, 106]. In the traditional unsupervised AD (a.k.a. one-class AD) setting, models learn the distribution of normal visual features from normal samples and are required to identify anomaly samples during inference. While recent advance" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 230, + 555, + 459 + ], + "blocks": [ + { + "bbox": [ + 316, + 230, + 555, + 459 + ], + "lines": [ + { + "bbox": [ + 316, + 230, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 555, + 459 + ], + "type": "image", + "image_path": "87ca03e000aaeacd11d7caf00c0f3e4662be350fb651c38b989f04967522b4d5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 464, + 556, + 521 + ], + "lines": [ + { + "bbox": [ + 313, + 464, + 556, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 556, + 521 + ], + "type": "text", + "content": "Figure 1. Visualization of the image-level AUROC comparison between our Anomaly-OV and current state-of-the-art ZSAD methods (WinCLIP [38], AnoVL [19], AnomalyCLIP [110], AdaCLIP [6]). Notably, our zero-shot performance on VisA even surpasses most recent advances in the few-shot setting [28, 51, 112]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 543, + 556, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 556, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 556, + 663 + ], + "type": "text", + "content": "ments [9, 25, 32–34, 37, 42, 82, 84, 95, 97, 104] have significantly improved the detection performance, these approaches assume the availability of a substantial number of normal samples. However, this assumption becomes impractical in certain scenarios due to strict data privacy policies and the significant human effort required for data classification, sometimes involving experts or specialists. Therefore, Zero-Shot Anomaly Detection (ZSAD) is emerging as a popular research direction, leading to the development of many innovative methods [6, 17, 27, 38, 43, 52, 78, 79, 110, 113]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 557, + 714 + ], + "type": "text", + "content": "Recent advances in Multimodal Large Language Models (MLLMs) [7, 15, 44, 45, 47, 48, 57, 58, 111] have shown revolutionary reasoning capabilities in various vision tasks [14, 29, 67, 70, 80, 91, 94, 107, 109]. However, the rea" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 289, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 289, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 289, + 712 + ], + "type": "text", + "content": "*Most of this work was done when J. Xu was an intern at HRI-USA." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20370" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": "soning of image abnormalities has not been explored due to the challenges of collecting large-scale datasets and establishing benchmarks. Existing methods simply predict the likelihood of an anomaly without providing rationales [6, 11, 19, 38, 110]. In contrast, for better interpretability, robustness, and trustworthiness, people would expect models to explain why an image is considered anomalous and provide visual evidence. Interestingly, we find that recent advanced MLLMs, such as GPT-4o [72], fall short in AD & reasoning. As shown in Figure 2, while the detection is correct, the explanation from GPT-4o lacks accuracy, indicating a gap in a comprehensive understanding of the anomaly." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 216, + 297, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 297, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 297, + 395 + ], + "type": "text", + "content": "To expedite research in AD & reasoning, we establish the first visual instruction tuning dataset, Anomaly-Instruct-125k, and the evaluation benchmark, VisA-D&R, through intensive human efforts. After evaluating current generalist MLLMs, we observe that these models fail to accurately detect and describe fine-grained anomalous details in images. To address this, we propose Anomaly-OneVision (Anomaly-OV), the first specialist visual assistant for ZSAD and reasoning. Unlike existing ZSAD methods [6, 11, 19, 38, 110], Anomaly-OV directly learns object-awareness abnormality embeddings in feature space using only the visual encoder. Inspired by human behavior in visual inspection, Anomaly-OV employs a Look-Twice Feature Matching (LTFM) mechanism to assist its LLM in adaptively selecting and emphasizing the most suspicious abnormal visual tokens." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 396, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 514 + ], + "type": "text", + "content": "Extensive experiments demonstrate that Anomaly-OV achieves significant improvements over advanced generalist models in both detection and reasoning. Extended results of Anomaly-OV, from applications in industrial defect detection to 3D inspection and medical image diagnosis, are provided for future study. With precise descriptions and rationales of visual anomalies, our model can infer potential causes (see Figure 2), assess current impacts, and offer improvement suggestions, positioning itself as a reliable assistant for visual inspection. Our contributions are in two folds:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 515, + 296, + 562 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 515, + 295, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 295, + 538 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 295, + 538 + ], + "type": "text", + "content": "- We establish the first visual instruction tuning dataset and benchmark for anomaly detection and reasoning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 539, + 296, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 539, + 296, + 562 + ], + "spans": [ + { + "bbox": [ + 56, + 539, + 296, + 562 + ], + "type": "text", + "content": "- We propose the first specialist visual assistant with state-of-the-art performance for this new impactful domain." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 573, + 142, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 573, + 142, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 573, + 142, + 586 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "Multimodal Large Language Models. Vision-Language Models (VLMs), such as CLIP [73], exhibit robust zero-shot classification capabilities and have been applied to a range of downstream vision tasks [10, 53, 56, 61, 86]. Combining a VLM's vision encoder and an LLM [20, 62, 74], MLLMs [15, 48, 49, 57, 111] enable text-based interactions related to visual content. MLLMs have shown remarkable reasoning capability, particularly when incorporated with prompting strategies such as Chain-of-Thought [66, 88, 105]. Recent studies have harnessed MLLMs to provide reasoning for" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 322, + 76, + 378, + 128 + ], + "blocks": [ + { + "bbox": [ + 322, + 76, + 378, + 128 + ], + "lines": [ + { + "bbox": [ + 322, + 76, + 378, + 128 + ], + "spans": [ + { + "bbox": [ + 322, + 76, + 378, + 128 + ], + "type": "image", + "image_path": "13bec8769ce795b7c1bf630673506d45bbe937ad7c185058bb030de265045c11.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 389, + 76, + 443, + 128 + ], + "blocks": [ + { + "bbox": [ + 389, + 76, + 443, + 128 + ], + "lines": [ + { + "bbox": [ + 389, + 76, + 443, + 128 + ], + "spans": [ + { + "bbox": [ + 389, + 76, + 443, + 128 + ], + "type": "image", + "image_path": "25d9ca1d82563a8f93cbf3b8f88b477c4c9f7175c783c82e46e24750d5b1726d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 446, + 79, + 545, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 446, + 79, + 545, + 125 + ], + "spans": [ + { + "bbox": [ + 446, + 79, + 545, + 125 + ], + "type": "text", + "content": "GT: The top-left candle shows a missing chunk of wax and an irregular indentation near the wick." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 356, + 144, + 526, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 144, + 526, + 154 + ], + "spans": [ + { + "bbox": [ + 356, + 144, + 526, + 154 + ], + "type": "text", + "content": "Do you see any anomalies in the image?" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 531, + 137, + 548, + 156 + ], + "blocks": [ + { + "bbox": [ + 531, + 137, + 548, + 156 + ], + "lines": [ + { + "bbox": [ + 531, + 137, + 548, + 156 + ], + "spans": [ + { + "bbox": [ + 531, + 137, + 548, + 156 + ], + "type": "image", + "image_path": "fa1ec3fecb2f2f7a21fcf15c025d1cc5e0202f2e27fadac7851d6ee400a6f176.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 322, + 157, + 337, + 172 + ], + "blocks": [ + { + "bbox": [ + 322, + 157, + 337, + 172 + ], + "lines": [ + { + "bbox": [ + 322, + 157, + 337, + 172 + ], + "spans": [ + { + "bbox": [ + 322, + 157, + 337, + 172 + ], + "type": "image", + "image_path": "1028b0d63a08272fee646e1912e0f6c44eeb5805de96a4139539c94247c752f7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 340, + 159, + 518, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 159, + 518, + 186 + ], + "spans": [ + { + "bbox": [ + 340, + 159, + 518, + 186 + ], + "type": "text", + "content": "The candle in the top left appears to have a damaged or missing wick, which is unusual compared to the other candles." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 321, + 193, + 338, + 211 + ], + "blocks": [ + { + "bbox": [ + 321, + 193, + 338, + 211 + ], + "lines": [ + { + "bbox": [ + 321, + 193, + 338, + 211 + ], + "spans": [ + { + "bbox": [ + 321, + 193, + 338, + 211 + ], + "type": "image", + "image_path": "01126e20ac9b600e067b4454d12d690d857d9b915c2df3b957d02dc35024e36f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 300, + 555, + 335 + ], + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 335 + ], + "type": "text", + "content": "Figure 2. Industrial image anomaly reasoning results from GPT-4o [72] and our Anomaly-OV. The responses for fine-grained anomaly reasoning are highlighted, with the ground truth given for reference." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 342, + 194, + 525, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 194, + 525, + 230 + ], + "spans": [ + { + "bbox": [ + 342, + 194, + 525, + 230 + ], + "type": "text", + "content": "Yes, there is an anomaly in the top left candle. It has a small hole or crack near the wick, which is not present in the other candles." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 374, + 237, + 526, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 237, + 526, + 247 + ], + "spans": [ + { + "bbox": [ + 374, + 237, + 526, + 247 + ], + "type": "text", + "content": "What is the cause for the anomaly?" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 531, + 230, + 548, + 250 + ], + "blocks": [ + { + "bbox": [ + 531, + 230, + 548, + 250 + ], + "lines": [ + { + "bbox": [ + 531, + 230, + 548, + 250 + ], + "spans": [ + { + "bbox": [ + 531, + 230, + 548, + 250 + ], + "type": "image", + "image_path": "0853c89fa1b7fdb473555b96d7d47ac50e03bb39bccc3193e3b271babe46203d.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 321, + 248, + 338, + 267 + ], + "blocks": [ + { + "bbox": [ + 321, + 248, + 338, + 267 + ], + "lines": [ + { + "bbox": [ + 321, + 248, + 338, + 267 + ], + "spans": [ + { + "bbox": [ + 321, + 248, + 338, + 267 + ], + "type": "image", + "image_path": "ec51fc591b957718275465597f05c2ab9414421124c387006f6117f901285a47.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 341, + 251, + 525, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 251, + 525, + 286 + ], + "spans": [ + { + "bbox": [ + 341, + 251, + 525, + 286 + ], + "type": "text", + "content": "The anomaly could be due to manufacturing defects, such as improper molding or cooling, or it might result from external damage like a puncture." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 348, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 555, + 384 + ], + "type": "text", + "content": "downstream tasks, e.g., video anomaly detection [67, 94], affective computing [14, 29], and visual commonsense reasoning [109], revealing more interpretability." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 388, + 555, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 496 + ], + "type": "text", + "content": "Unsupervised Anomaly Detection. Due to the scarcity and difficulty of collecting anomalous data, researchers often focus on the unsupervised AD setting, which exclusively uses normal data to train an AD model. Earlier studies, such as reconstruction-based [63, 69, 99], student-teacher [18, 85, 102], and augmentation-based [46] approaches, assume a large amount of normal data is available. These traditional approaches are less practical when data are restricted or expensive, such as in the medical domain." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 502, + 556, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 556, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 556, + 597 + ], + "type": "text", + "content": "Zero-Shot Anomaly Detection. Unlike unsupervised AD [35, 77] and few-shot AD [22, 28, 36, 51, 112], ZSAD models directly access the likelihood of abnormality for a given image without requiring data specific to the target object. Existing works [6, 19, 38, 110] accomplish ZSAD by comparing visual and textual features encoded by visual and text encoders of CLIP and constructing their positive (anomaly) and negative (normal) prompts in the format of:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 368, + 604, + 499, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 604, + 499, + 634 + ], + "spans": [ + { + "bbox": [ + 368, + 604, + 499, + 634 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {P} ^ {+} = \\left[ V _ {1} \\right] \\left[ V _ {2} \\right] \\dots \\left[ V _ {n} \\right] [ o b j e c t ] \\\\ \\mathcal {P} ^ {-} = \\left[ W _ {1} \\right] \\left[ W _ {2} \\right] \\dots \\left[ W _ {n} \\right] [ o b j e c t ] \\\\ \\end{array}", + "image_path": "4c59b16e65bebdaccd6d0df7cc2815289e59b657283c04a446794328544aceba.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "W_{i}" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": " are handcrafted or learnable tokens, and object refers to the word object or the class name of the object. However, simply utilizing object to represent all kinds of objects cannot capture the class-awareness abnormality types. Also, for an intelligent visual assistant, the images should be totally blind to the user (object-agnostic)." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20371" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 69, + 555, + 349 + ], + "blocks": [ + { + "bbox": [ + 55, + 69, + 555, + 349 + ], + "lines": [ + { + "bbox": [ + 55, + 69, + 555, + 349 + ], + "spans": [ + { + "bbox": [ + 55, + 69, + 555, + 349 + ], + "type": "image", + "image_path": "c9b0b96175b01f00b0c7a0ecc318ecfbc5b060ab17aea4571e60fc25539fed13.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 357, + 555, + 380 + ], + "lines": [ + { + "bbox": [ + 55, + 357, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 555, + 380 + ], + "type": "text", + "content": "Figure 3. Overview of the Anomaly-0V architecture. It consists of two training stages: (1) professional training for the anomaly expert, and (2) visual instruction tuning for anomaly detection and reasoning. Text and visual tokens are distinguished by different colors." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 399, + 112, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 112, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 112, + 411 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 419, + 135, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 135, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 135, + 433 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 437, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 295, + 569 + ], + "type": "text", + "content": "Training an MLLM from scratch demands extensive data and computational resources to align the visual and textual embedding spaces and develop robust instruction-following capabilities. Recent studies [23, 83, 93] reveal that pretrained MLLMs function as generalists, possessing a broad knowledge base but underperforming in specialized domains. Therefore, our goal is to introduce an auxiliary specialist or expert model designed to guide the generalist in selecting and utilizing critical visual tokens. This approach circumvents the need for large-scale pre-training while preserving the generalization capacity of the original model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": "We choose " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "LLaVA-OneVision" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": " [44] as our base MLLM because it is open-sourced and performs similarly to other commercial models. " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "LLaVA-OneVision" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": " follows the model architectures for LLaVA family [47, 57-59] and other generic MLLMs, which typically consist of three major components: Visual Encoder, Projector, and LLM. The visual encoder [73, 100] extracts the visual information from the raw images, the projector aligns the spaces of visual features with the word embedding, and the LLM is responsible for textual instruction processing and complex reasoning. Since the image resolution for CLIP pre-training is fixed, " + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "LLaVA-OneVision" + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": " leverages AnyRes with pooling strategy to scale" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 400, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 555, + 449 + ], + "type": "text", + "content": "up the input raw image resolution. Specifically, the high-resolution images are divided into a prototyped number of crops, and the visual encoder independently processes the image crops before final spatial pooling." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 456, + 444, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 444, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 444, + 467 + ], + "type": "text", + "content": "3.2. Architecture Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 473, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 510 + ], + "type": "text", + "content": "With the same image-splitting strategy AnyRes as LLaVA-OneVision, the input high-resolution image is split into several crops, and the new image set can be written as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 381, + 517, + 555, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 517, + 555, + 531 + ], + "spans": [ + { + "bbox": [ + 381, + 517, + 555, + 531 + ], + "type": "interline_equation", + "content": "\\mathcal {I} = \\left\\{I _ {0}, I _ {1}, I _ {2}, \\dots , I _ {n - 1} \\right\\} \\tag {1}", + "image_path": "0e01d481d09a926ba19bd0e00fff8337bac370465ed47f825347dc80be97d418.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 539, + 556, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 556, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 556, + 634 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 539, + 556, + 634 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 313, + 539, + 556, + 634 + ], + "type": "text", + "content": " is the resized original image and " + }, + { + "bbox": [ + 313, + 539, + 556, + 634 + ], + "type": "inline_equation", + "content": "I_{j\\neq 0}" + }, + { + "bbox": [ + 313, + 539, + 556, + 634 + ], + "type": "text", + "content": " refers to the image crops. As shown in Figure 3, the image set " + }, + { + "bbox": [ + 313, + 539, + 556, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 539, + 556, + 634 + ], + "type": "text", + "content": " will be processed by the visual encoder " + }, + { + "bbox": [ + 313, + 539, + 556, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}" + }, + { + "bbox": [ + 313, + 539, + 556, + 634 + ], + "type": "text", + "content": " to generate the final visual features " + }, + { + "bbox": [ + 313, + 539, + 556, + 634 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{v}_j^o\\}" + }, + { + "bbox": [ + 313, + 539, + 556, + 634 + ], + "type": "text", + "content": ". Similar to AnomalyCLIP [110], we store the outputs for four selected layers in the ViT [21] to capture the image representations from different levels and apply four adapters to compress the feature dimension. Then, the extracted visual features can be written as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 406, + 643, + 555, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 643, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 406, + 643, + 555, + 658 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {j} ^ {i} = \\mathcal {F} _ {\\theta} ^ {i} (I _ {j}) \\tag {2}", + "image_path": "7a5ee3b14900e7f4c10a502f8f2a360ade0aa7440bf6100e66068efb96c57176.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "-th level and " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " refers to the index of corresponding image in " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": ". These multi-level features have been demonstrated to be effective in capturing fine-grained local semantics by recent works [6, 28, 110]." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20372" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 294, + 133 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 294, + 133 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 294, + 133 + ], + "type": "image", + "image_path": "f22e345a2ef3107a4d55a4c39a16623ba30cea442cceb34a735f0e07cee87875.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 140, + 288, + 152 + ], + "lines": [ + { + "bbox": [ + 61, + 140, + 288, + 152 + ], + "spans": [ + { + "bbox": [ + 61, + 140, + 288, + 152 + ], + "type": "text", + "content": "Figure 4. Simulation of visual anomaly inspection by humans." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 161, + 296, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 161, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 54, + 161, + 296, + 316 + ], + "type": "text", + "content": "The large-scale pre-trained CLIP models align the projection spaces of the textual and visual encoder. Therefore, the encoded image features already contain the class information required by ZSAD. To avoid human involvement in object classification and reduce the model complexity, we remove the heavy text encoder commonly utilized in existing works and let the visual model itself parse the information for suspicious classes or objects. Specifically, the output visual features for the original image " + }, + { + "bbox": [ + 54, + 161, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_0^o" + }, + { + "bbox": [ + 54, + 161, + 296, + 316 + ], + "type": "text", + "content": " are leveraged to provide the global description of the target object or regions in the look-back path. With the multi-level features and the global embeddings, the LTFM module is responsible for the recognition and localization of suspicious tokens." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 317, + 296, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 296, + 448 + ], + "type": "text", + "content": "Drawing inspiration from human visual inspection, where suspicious objects or regions are identified and then inspected closely (see Figure 4), we design the VT selector module for aggregating (zooming in) the crucial visual tokens and explicitly assisting the LLM in distinguishing these tokens from many irrelevant ones when dealing with instructions regarding anomaly detection and reasoning. Additionally, the original visual features are preserved to maintain the generalization capability of the base model on regular instructions, such as Can you describe the content of the image?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 456, + 220, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 220, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 220, + 468 + ], + "type": "text", + "content": "3.3. Look-Twice Feature Matching" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 473, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 296, + 605 + ], + "type": "text", + "content": "Given the global object information " + }, + { + "bbox": [ + 55, + 473, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_0^o" + }, + { + "bbox": [ + 55, + 473, + 296, + 605 + ], + "type": "text", + "content": " provided by the lookback path, we generate the class-awareness abnormality description by merging " + }, + { + "bbox": [ + 55, + 473, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_0^o" + }, + { + "bbox": [ + 55, + 473, + 296, + 605 + ], + "type": "text", + "content": " with two learnable embeddings: " + }, + { + "bbox": [ + 55, + 473, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{e}^{+}\\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 55, + 473, + 296, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 473, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{e}^{-}\\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 55, + 473, + 296, + 605 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 473, + 296, + 605 + ], + "type": "inline_equation", + "content": "+" + }, + { + "bbox": [ + 55, + 473, + 296, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 473, + 296, + 605 + ], + "type": "inline_equation", + "content": "-" + }, + { + "bbox": [ + 55, + 473, + 296, + 605 + ], + "type": "text", + "content": " indicate positive (anomalous) and negative (normal) patterns and " + }, + { + "bbox": [ + 55, + 473, + 296, + 605 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 473, + 296, + 605 + ], + "type": "text", + "content": " is the embedding dimension. Specifically, a linear layer " + }, + { + "bbox": [ + 55, + 473, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i^o" + }, + { + "bbox": [ + 55, + 473, + 296, + 605 + ], + "type": "text", + "content": " is applied along the token dimension to select and fuse useful tokens from " + }, + { + "bbox": [ + 55, + 473, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_0^o" + }, + { + "bbox": [ + 55, + 473, + 296, + 605 + ], + "type": "text", + "content": ", and then the fused vector will be concatenated with " + }, + { + "bbox": [ + 55, + 473, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{e}^{+}" + }, + { + "bbox": [ + 55, + 473, + 296, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 473, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{e}^{-}" + }, + { + "bbox": [ + 55, + 473, + 296, + 605 + ], + "type": "text", + "content": " independently and pass through two MLPs " + }, + { + "bbox": [ + 55, + 473, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{G}_i^+, \\mathcal{G}_i^-\\}" + }, + { + "bbox": [ + 55, + 473, + 296, + 605 + ], + "type": "text", + "content": " to generate the abnormality and normality descriptions " + }, + { + "bbox": [ + 55, + 473, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{d}_i^+, \\mathbf{d}_i^-\\}" + }, + { + "bbox": [ + 55, + 473, + 296, + 605 + ], + "type": "text", + "content": ", which can be represented by:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 101, + 614, + 296, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 614, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 101, + 614, + 296, + 647 + ], + "type": "interline_equation", + "content": "\\left\\{\\mathbf {d} _ {i} ^ {+}, \\mathbf {d} _ {i} ^ {-} \\right\\} = \\left\\{ \\begin{array}{l} \\mathcal {G} _ {i} ^ {+} \\left(\\mathbf {e} ^ {+} \\circ \\mathcal {T} _ {i} ^ {o} \\left(\\mathbf {v} _ {0} ^ {o}\\right)\\right) \\\\ \\mathcal {G} _ {i} ^ {-} \\left(\\mathbf {e} ^ {-} \\circ \\mathcal {T} _ {i} ^ {o} \\left(\\mathbf {v} _ {0} ^ {o}\\right)\\right) \\end{array} \\right. \\tag {3}", + "image_path": "fcaadde5688f715cc7d9d01bae6bcc8bb7a1290e010f04502fb44d87cf591d52.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 296, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 700 + ], + "type": "text", + "content": "The visual features extracted from different levels of the ViT focus on different scales of semantics. Thus, the parameters of " + }, + { + "bbox": [ + 55, + 653, + 296, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i^o" + }, + { + "bbox": [ + 55, + 653, + 296, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 653, + 296, + 700 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{G}_i^+, \\mathcal{G}_i^-\\}" + }, + { + "bbox": [ + 55, + 653, + 296, + 700 + ], + "type": "text", + "content": " should be independent for different levels, where " + }, + { + "bbox": [ + 55, + 653, + 296, + 700 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 653, + 296, + 700 + ], + "type": "text", + "content": " indicate the level number." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 701, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 712 + ], + "type": "text", + "content": "Similar to the zero-shot classification mechanism of CLIP" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "content": "models, we calculate the possibilities of each patch token in " + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_j^i" + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "content": " belonging to the anomalous patterns by combining cosine similarity and softmax operations:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 321, + 115, + 555, + 146 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 115, + 555, + 146 + ], + "spans": [ + { + "bbox": [ + 321, + 115, + 555, + 146 + ], + "type": "interline_equation", + "content": "\\mathbf {m} _ {j} ^ {i} = \\frac {\\exp \\left(< \\mathbf {d} _ {i} ^ {+} , \\mathbf {v} _ {j} ^ {i} > / \\tau\\right)}{\\exp \\left(< \\mathbf {d} _ {i} ^ {+} , \\mathbf {v} _ {j} ^ {i} > / \\tau\\right) + \\exp \\left(< \\mathbf {d} _ {i} ^ {-} , \\mathbf {v} _ {j} ^ {i} > / \\tau\\right)} \\tag {4}", + "image_path": "7cf92f71e013ff69575d84a4f82b43c80e52c1ecdffdc55e6bcb7ef4476e9faf.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 153, + 555, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 153, + 555, + 225 + ], + "spans": [ + { + "bbox": [ + 313, + 153, + 555, + 225 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 153, + 555, + 225 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_j^i" + }, + { + "bbox": [ + 313, + 153, + 555, + 225 + ], + "type": "text", + "content": " represents the significance map for visual tokens, " + }, + { + "bbox": [ + 313, + 153, + 555, + 225 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 153, + 555, + 225 + ], + "type": "text", + "content": " is the temperature hyperparameter, and " + }, + { + "bbox": [ + 313, + 153, + 555, + 225 + ], + "type": "inline_equation", + "content": "<,>" + }, + { + "bbox": [ + 313, + 153, + 555, + 225 + ], + "type": "text", + "content": " refers to the cosine similarity operator. The patch weight in " + }, + { + "bbox": [ + 313, + 153, + 555, + 225 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_j^i" + }, + { + "bbox": [ + 313, + 153, + 555, + 225 + ], + "type": "text", + "content": " indicates the closeness of the corresponding visual token to the anomalous pattern. Then, all the maps are averaged to capture the token significances from low to high levels:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 399, + 232, + 555, + 265 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 232, + 555, + 265 + ], + "spans": [ + { + "bbox": [ + 399, + 232, + 555, + 265 + ], + "type": "interline_equation", + "content": "\\mathbf {m} _ {j} = \\sum_ {i = 1} ^ {4} \\mathbf {m} _ {j} ^ {i} / 4 \\tag {5}", + "image_path": "878b1c28afc6473a33a64e28c8b963c57bb75da87b35707d0900c55f99b88c77.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 271, + 554, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 554, + 319 + ], + "type": "text", + "content": "The visual features are leveraged twice in the forward and look-back paths, so this module is named by Look-Twice Feature Matching (LTFM), following the nature of two-step human visual inspection shown in Figure 4." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 326, + 438, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 438, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 438, + 337 + ], + "type": "text", + "content": "3.4. Visual Token Selector" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 343, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 555, + 475 + ], + "type": "text", + "content": "Under the image cropping strategy widely applied in recent MLLMs, there will be a large number of visual tokens for a high-resolution image, e.g., 7290 tokens for an image with " + }, + { + "bbox": [ + 313, + 343, + 555, + 475 + ], + "type": "inline_equation", + "content": "1152 \\times 1152" + }, + { + "bbox": [ + 313, + 343, + 555, + 475 + ], + "type": "text", + "content": " resolution in LLaVA-OneVision. While these tokens provide rich visual details, the LLM is required to pick the most useful information when adapting to a specific task. When the LLM lacks enough knowledge in this domain, the token-picking process will become complicated. Thus, our solution is to introduce a specialist or expert who knows which token is crucial or not and assist the LLM in selecting and emphasizing (zooming in) the crucial tokens." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 475, + 554, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 554, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 554, + 559 + ], + "type": "text", + "content": "Given the encoded visual tokens " + }, + { + "bbox": [ + 313, + 475, + 554, + 559 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{v}_j^o\\}" + }, + { + "bbox": [ + 313, + 475, + 554, + 559 + ], + "type": "text", + "content": " for each image crop in " + }, + { + "bbox": [ + 313, + 475, + 554, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 475, + 554, + 559 + ], + "type": "text", + "content": " and the corresponding significance map " + }, + { + "bbox": [ + 313, + 475, + 554, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_j" + }, + { + "bbox": [ + 313, + 475, + 554, + 559 + ], + "type": "text", + "content": ", the suspicious tokens are emphasized by direct multiplication of the two tensors. Then, the normal tokens will be scaled to zero while the anomalous tokens will be maintained. After that, spatial average pooling " + }, + { + "bbox": [ + 313, + 475, + 554, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 475, + 554, + 559 + ], + "type": "text", + "content": " is applied to reduce the number of tokens. This process can be written as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 394, + 567, + 555, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 567, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 394, + 567, + 555, + 581 + ], + "type": "interline_equation", + "content": "\\mathbf {q} _ {j} = \\mathcal {P} \\left(\\mathbf {v} _ {j} ^ {o} \\odot \\mathbf {m} _ {j}\\right) \\tag {6}", + "image_path": "cd818998894b43ffb175a608c503f52bc07734b217def8343b8f622c8f85bc29.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 588, + 554, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 554, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 554, + 651 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 588, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_j\\in \\mathbb{R}^{h\\times w\\times D}" + }, + { + "bbox": [ + 313, + 588, + 554, + 651 + ], + "type": "text", + "content": " refers to the pooled query tokens. Empirically, setting " + }, + { + "bbox": [ + 313, + 588, + 554, + 651 + ], + "type": "inline_equation", + "content": "h = w = 2" + }, + { + "bbox": [ + 313, + 588, + 554, + 651 + ], + "type": "text", + "content": " provides a better trade-off than other options. Then, a Q-Former " + }, + { + "bbox": [ + 313, + 588, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 313, + 588, + 554, + 651 + ], + "type": "text", + "content": " [49] is leveraged to aggregate the correlated tokens in the original output by forwarding " + }, + { + "bbox": [ + 313, + 588, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_j" + }, + { + "bbox": [ + 313, + 588, + 554, + 651 + ], + "type": "text", + "content": " as the query and " + }, + { + "bbox": [ + 313, + 588, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_j^o" + }, + { + "bbox": [ + 313, + 588, + 554, + 651 + ], + "type": "text", + "content": " as the key and value:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 392, + 657, + 555, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 657, + 555, + 672 + ], + "spans": [ + { + "bbox": [ + 392, + 657, + 555, + 672 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {j} ^ {s} = \\mathcal {Q} \\left(\\mathbf {q} _ {j}, \\mathbf {v} _ {j} ^ {o}, \\mathbf {v} _ {j} ^ {o}\\right) \\tag {7}", + "image_path": "44ed6804bc1081187f63ac56fdf3c0df23d157d60f08dfce520d6516a41b9ff4.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "The Visual Token Selector (VT Selector) serves as a tool for the anomaly expert to hand-pick visual tokens that contain the most suspicious semantics for a given image." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20373" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 77, + 95, + 121 + ], + "blocks": [ + { + "bbox": [ + 61, + 77, + 95, + 121 + ], + "lines": [ + { + "bbox": [ + 61, + 77, + 95, + 121 + ], + "spans": [ + { + "bbox": [ + 61, + 77, + 95, + 121 + ], + "type": "image", + "image_path": "27870aa2e157a6e267299eafa2aaa3056f439cbc70d1a89ab877c999c695d5e7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 76, + 170, + 113 + ], + "lines": [ + { + "bbox": [ + 97, + 76, + 170, + 113 + ], + "spans": [ + { + "bbox": [ + 97, + 76, + 170, + 113 + ], + "type": "text", + "content": "Q: Do you notice any anomalies in the image? A: Yes, the anomaly in the image is a fine, dark crack visible on the ceramic cup, starting from the rim and extending downward." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 97, + 117, + 171, + 159 + ], + "lines": [ + { + "bbox": [ + 97, + 117, + 171, + 159 + ], + "spans": [ + { + "bbox": [ + 97, + 117, + 171, + 159 + ], + "type": "text", + "content": "Q: How can I prevent similar anomalies in the future? A: Avoid rapid temperature changes, handle the cup carefully to minimize impacts, and consider using cups made with thicker, more durable materials." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 55, + 177, + 555, + 222 + ], + "lines": [ + { + "bbox": [ + 55, + 177, + 555, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 177, + 555, + 222 + ], + "type": "text", + "content": "Figure 5. Composition of the instruction data in Anomaly-Instruct-125k. There are four main types of image samples: in-the-wild, industrial, medical, and " + }, + { + "bbox": [ + 55, + 177, + 555, + 222 + ], + "type": "inline_equation", + "content": "3D" + }, + { + "bbox": [ + 55, + 177, + 555, + 222 + ], + "type": "text", + "content": " (in the format of multi-view images), covering most image anomaly detection tasks and enabling the possibility of a unified assistant for visual inspection. The reasoning words are highlighted in blue. For more information about dataset establishment, statistics, and the data collection pipeline, please refer to Section A1 in the supplementary." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 184, + 83, + 216, + 116 + ], + "blocks": [ + { + "bbox": [ + 184, + 83, + 216, + 116 + ], + "lines": [ + { + "bbox": [ + 184, + 83, + 216, + 116 + ], + "spans": [ + { + "bbox": [ + 184, + 83, + 216, + 116 + ], + "type": "image", + "image_path": "1cae4e97724f842ce372d6f3700a1026cebd396a15bf77bfa7c68dac0e191126.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 76, + 304, + 113 + ], + "lines": [ + { + "bbox": [ + 217, + 76, + 304, + 113 + ], + "spans": [ + { + "bbox": [ + 217, + 76, + 304, + 113 + ], + "type": "text", + "content": "Q: Are there any anomalies in this cable image? A: Yes, the green wire is missing its internal copper conductors, unlike the blue and brown wires, which contain complete sets of bundled copper strands." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 217, + 117, + 304, + 163 + ], + "lines": [ + { + "bbox": [ + 217, + 117, + 304, + 163 + ], + "spans": [ + { + "bbox": [ + 217, + 117, + 304, + 163 + ], + "type": "text", + "content": "Q: What could be the reason behind this anomaly? A: The absence of copper in the green wire could be due to manufacturing defects such as failure to insert the strands or a mechanical issue where the strands were removed or misplaced during production or handling." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 83, + 343, + 116 + ], + "blocks": [ + { + "bbox": [ + 310, + 83, + 343, + 116 + ], + "lines": [ + { + "bbox": [ + 310, + 83, + 343, + 116 + ], + "spans": [ + { + "bbox": [ + 310, + 83, + 343, + 116 + ], + "type": "image", + "image_path": "236f9ae2d43bd439000f023802369ecec417947d1e99b3582af1ac9171fb9000.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 76, + 421, + 102 + ], + "lines": [ + { + "bbox": [ + 343, + 76, + 421, + 102 + ], + "spans": [ + { + "bbox": [ + 343, + 76, + 421, + 102 + ], + "type": "text", + "content": "Q: What does the image show? A: The image shows a densely cellular tissue sample with features indicating potential metastatic breast cancer." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 343, + 107, + 423, + 163 + ], + "lines": [ + { + "bbox": [ + 343, + 107, + 423, + 163 + ], + "spans": [ + { + "bbox": [ + 343, + 107, + 423, + 163 + ], + "type": "text", + "content": "Q: What suggests that there might be malignant activity in the tissue sample? A: Malignant activity is suggested by irregularly shaped cells, hyperchromatic nuclei, and an increased nuclear-to-cytoplasmic ratio, indicative of abnormal and increased DNA content typical in cancerous cells." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 434, + 77, + 462, + 102 + ], + "blocks": [ + { + "bbox": [ + 434, + 77, + 462, + 102 + ], + "lines": [ + { + "bbox": [ + 434, + 77, + 462, + 102 + ], + "spans": [ + { + "bbox": [ + 434, + 77, + 462, + 102 + ], + "type": "image", + "image_path": "2e43f15b428805c298907b2430f159ad648f7a757bfd215f12ca344cd3fa05b1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 76, + 545, + 158 + ], + "lines": [ + { + "bbox": [ + 463, + 76, + 545, + 158 + ], + "spans": [ + { + "bbox": [ + 463, + 76, + 545, + 158 + ], + "type": "text", + "content": "Q: What kind of object are you examining in the images? A: I'm examining a round cookie with a golden-brown surface and visible chocolate chips embedded across its surface. Q: Are there any significant anomalies detected on this cookie? A: Yes, there are two significant anomalies on the cookie: a central defect featuring a noticeable hole and an edge defect showing a gap or crack." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 236, + 167, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 167, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 167, + 248 + ], + "type": "text", + "content": "3.5. Inference and Loss" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 254, + 296, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 296, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 296, + 314 + ], + "type": "text", + "content": "Anomaly Prediction. In the traditional anomaly detection task, the model predicts the possibility of the image being abnormal. To achieve anomaly score prediction, we aggregate the anomaly information from all the image crops by an average operation weighted on the significance maps:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 111, + 322, + 296, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 322, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 111, + 322, + 296, + 352 + ], + "type": "interline_equation", + "content": "\\mathbf {r} (\\mathcal {I}) = \\frac {\\sum_ {j , k} \\mathbf {v} _ {j} ^ {\\mathrm {s}} [ k ] \\cdot \\mathcal {P} (\\mathbf {m} _ {j}) [ k ]}{\\sum_ {j , k} \\mathcal {P} (\\mathbf {m} _ {j}) [ k ]} \\tag {8}", + "image_path": "14f21ae3ac1be6f7b4c9582ebcd87c76d5d7f7a8347c835906eacd93f5c8afcc.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 361, + 296, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 409 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 361, + 296, + 409 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 361, + 296, + 409 + ], + "type": "text", + "content": " is the same spatial pooling in VT Selector and " + }, + { + "bbox": [ + 55, + 361, + 296, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(\\mathcal{I})" + }, + { + "bbox": [ + 55, + 361, + 296, + 409 + ], + "type": "text", + "content": " is a vector containing the global anomaly information for the entire image. Then, the anomaly expert can calculate the image-level abnormal possibility by parsing " + }, + { + "bbox": [ + 55, + 361, + 296, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(\\mathcal{I})" + }, + { + "bbox": [ + 55, + 361, + 296, + 409 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 418, + 295, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 418, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 295, + 432 + ], + "type": "interline_equation", + "content": "\\operatorname {s c o r e} (\\mathcal {I}) = \\operatorname {S i g m o i d} \\left(\\mathcal {G} ^ {o} (\\mathbf {r} (\\mathcal {I}))\\right) \\tag {9}", + "image_path": "a243b30c75adcc35a2ed5e6d457206b7fa557bf2d54e3c1cfebb501a9f82d50c.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 441, + 296, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 296, + 489 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 441, + 296, + 489 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^o" + }, + { + "bbox": [ + 55, + 441, + 296, + 489 + ], + "type": "text", + "content": " is an MLP for distinguishing normal and abnormal semantics. To handle the unbalanced sample distribution, we employ the balanced BCE loss as the professional training objective for the anomaly expert components." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 495, + 296, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 673 + ], + "type": "text", + "content": "Text Generation. Instead of directly forwarding the concatenation of the original " + }, + { + "bbox": [ + 55, + 495, + 296, + 673 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{v}_j^o\\}" + }, + { + "bbox": [ + 55, + 495, + 296, + 673 + ], + "type": "text", + "content": " and the selected " + }, + { + "bbox": [ + 55, + 495, + 296, + 673 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{r}(\\mathcal{I}),\\mathbf{v}_j^s\\}" + }, + { + "bbox": [ + 55, + 495, + 296, + 673 + ], + "type": "text", + "content": " visual tokens into the LLM, we apply an indication prompt with suspicious feature: in the middle of the two series of tokens, which will highlight the selected tokens for LLM when handling anomaly-related instructions. This approach can be considered a form of prompt engineering in MLLMs. Besides, the is chosen from {highly, moderately, slightly} and is determined by score " + }, + { + "bbox": [ + 55, + 495, + 296, + 673 + ], + "type": "inline_equation", + "content": "(\\mathcal{I})" + }, + { + "bbox": [ + 55, + 495, + 296, + 673 + ], + "type": "text", + "content": " and predefined thresholds " + }, + { + "bbox": [ + 55, + 495, + 296, + 673 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{s}_{low},\\mathbf{s}_{high}\\}" + }, + { + "bbox": [ + 55, + 495, + 296, + 673 + ], + "type": "text", + "content": ". When the input image " + }, + { + "bbox": [ + 55, + 495, + 296, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 495, + 296, + 673 + ], + "type": "text", + "content": " has a high likelihood of anomaly, the LLM will place greater emphasis on the selected tokens; otherwise, these tokens will have less significance. The text generation is implemented by the original auto-regressive token prediction mechanism of LLM:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 86, + 684, + 296, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 684, + 296, + 716 + ], + "spans": [ + { + "bbox": [ + 86, + 684, + 296, + 716 + ], + "type": "interline_equation", + "content": "p \\left(X _ {a} \\mid \\mathcal {I}, X _ {q}\\right) = \\prod_ {t = 1} ^ {L} p _ {\\theta} \\left(x _ {t} \\mid \\mathcal {I}, X _ {q, < t}, X _ {a, < t}\\right) \\tag {10}", + "image_path": "c0c730df2c37037d6bd3d2c842669eecd192349be725ff53a3748ede7f764e8f.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 237, + 555, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 237, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 555, + 297 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 237, + 555, + 297 + ], + "type": "inline_equation", + "content": "X_{a, < t}" + }, + { + "bbox": [ + 313, + 237, + 555, + 297 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 237, + 555, + 297 + ], + "type": "inline_equation", + "content": "X_{q, < t}" + }, + { + "bbox": [ + 313, + 237, + 555, + 297 + ], + "type": "text", + "content": " are the answer and instruction tokens from all prior turns before the current prediction token " + }, + { + "bbox": [ + 313, + 237, + 555, + 297 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 237, + 555, + 297 + ], + "type": "text", + "content": " for a sequence of length " + }, + { + "bbox": [ + 313, + 237, + 555, + 297 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 237, + 555, + 297 + ], + "type": "text", + "content": ". The entire model is parameterized by " + }, + { + "bbox": [ + 313, + 237, + 555, + 297 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 237, + 555, + 297 + ], + "type": "text", + "content": " and trained by the original language model cross-entropy loss for each predicted answer token " + }, + { + "bbox": [ + 313, + 237, + 555, + 297 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 237, + 555, + 297 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 308, + 453, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 453, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 453, + 320 + ], + "type": "text", + "content": "4. Dataset and Benchmark" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 328, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 555, + 449 + ], + "type": "text", + "content": "The lack of multimodal instruction-following data for image anomaly detection and reasoning hinders the development of special intelligent assistants in this domain. Even though AnomalyGPT [28] introduces a prompt tuning dataset by simulating the anomalies, the scale of their dataset and the diversity of their instructions and answers are limited, only focusing on anomaly localization. To resolve the data scarcity issue, we establish the first large-scale instruction tuning dataset: Anomaly-Instruct-125k and the corresponding anomaly detection and reasoning benchmark: VisA-D&R." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 456, + 447, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 447, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 447, + 468 + ], + "type": "text", + "content": "4.1. Anomaly-Instruct-125k" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 474, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 581 + ], + "type": "text", + "content": "LLaVA [57] builds its instruction tuning dataset by leveraging the image caption and bounding boxes available in the COCO dataset[55] to prompt the text-only GPT-4. ShareGPT4V [8] provides a higher-quality dataset by directly prompting GPT-4V [71]. However, there is no image caption provided in existing anomaly detection datasets [1, 2], and no matter GPT-4V [71] or most recent GPT-4o [72] cannot accurately locate and describe the anomalies in the image without explicit human involvement." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": "To resolve these issues, we design a new prompt pipeline for accurate anomaly description generation. Since most of the datasets contain annotations for anomaly types, we manually combine the class name and anomaly type, such as a [capsule] with [poke] on surface. If the anomaly masks are provided, we draw bounding boxes on the images to highlight the anomalous area. The short description and the image with (or w/o) bounding boxes are used to prompt GPT-4o to generate the detailed image and anomaly descriptions. Then, we employ an in-context learning strategy similar to LLaVA to create the instructions." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20374" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 64, + 76, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 76, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 64, + 76, + 107, + 84 + ], + "type": "text", + "content": "Detection:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 84, + 280, + 101 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 84, + 280, + 101 + ], + "spans": [ + { + "bbox": [ + 64, + 84, + 280, + 101 + ], + "type": "text", + "content": "Q: Are there any defects for the object in the image? Please reply with 'Yes' or 'No'." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 65, + 108, + 107, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 108, + 107, + 117 + ], + "spans": [ + { + "bbox": [ + 65, + 108, + 107, + 117 + ], + "type": "text", + "content": "Reasoning:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 117, + 284, + 148 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 65, + 117, + 253, + 124 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 117, + 253, + 124 + ], + "spans": [ + { + "bbox": [ + 65, + 117, + 253, + 124 + ], + "type": "text", + "content": "Q1: Do you observe any anomalies in the image?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 65, + 124, + 260, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 124, + 260, + 132 + ], + "spans": [ + { + "bbox": [ + 65, + 124, + 260, + 132 + ], + "type": "text", + "content": "Q2: Can you describe the anomalies you observed?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 65, + 132, + 268, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 132, + 268, + 140 + ], + "spans": [ + { + "bbox": [ + 65, + 132, + 268, + 140 + ], + "type": "text", + "content": "Q3: What is the potential cause for the anomalies?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 65, + 140, + 284, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 140, + 284, + 148 + ], + "spans": [ + { + "bbox": [ + 65, + 140, + 284, + 148 + ], + "type": "text", + "content": "Q4: How can such anomalies be prevented in the future?" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 163, + 297, + 186 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 297, + 186 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 297, + 186 + ], + "type": "text", + "content": "Figure 6. Prompt examples in VisA-D&R for detection and reasoning. The complex reasoning instructions are highlighted." + } + ] + } + ], + "index": 8, + "type": "text" + }, + { + "bbox": [ + 56, + 196, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 296, + 437 + ], + "type": "text", + "content": "For a unified visual inspection dataset, precise instruction data is collected from MVtec AD [2], the training set of BMAD [1], Anomaly-ShapeNet [50], Real3D-AD [60], and MVtec-3D AD [4], covering both 2D to 3D data across industry to medical domains. The 3D point cloud data are converted into 9 multi-view images, and the corresponding masks are rendered using predefined camera positions. However, the diversities and scales of these datasets are relatively limited, probably due to the difficulty of collecting anomaly images. To scale up the instruction data, we introduce an automatic anomaly data collection pipeline combining GPT4o [72] and Google Image Search [26] for image collection, data cleaning, and instruction generation. Finally, 72k in-the-wild images (named as WebAD) targeting anomaly detection are collected, significantly enriching our instruction dataset. Several samples from Anomaly-Instruct-125k are shown in Figure 5. The instructions are mainly in the format of multi-round conversations, covering anomaly detection and description in low-level reasoning and potential cause and future suggestions for complex understanding." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 444, + 130, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 130, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 130, + 455 + ], + "type": "text", + "content": "4.2.VisA-D&R" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 461, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 295, + 605 + ], + "type": "text", + "content": "VisA [115] is a classic but challenging industrial anomaly detection dataset, providing fine-grained anomaly type and segmentation for each image. For evaluation of the anomaly detection and reasoning performance on existing and future methods, we select 10 classes from VisA and follow a similar data generation pipeline of Anomaly-Instruct-125k to create the benchmark. Differently, significant human effort has been invested in meticulously reviewing all generated images and anomaly descriptions. Wrong descriptions are picked out and re-annotated by humans before utilizing them for Q&A generation. Totally, the benchmark consists of 761 normal samples and 1000 anomalous ones." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "For evaluating detection performance, questions designed to elicit a one-word answer are used to prompt the MLLMs (Figure 6), with results quantified using Accuracy, Precision, Recall, and F1-score. We divide the reasoning performance into two parts: low-level reasoning that focuses on the description of visual defects or anomalies and complex reasoning requiring the MLLMs to provide the potential cause and future improvement strategies for the detected anomalies, where ROUGE-L [54], Sentence-BERT (SBERT) [75], and" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "GPT-Score (GPT-4 as the judge [57]) are utilized to quantify the similarity between generated text and ground truth. Note that low-level reasoning is highly correlated to detection performance, while anomaly-type descriptions of low-level reasoning determine the output of complex reasoning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 142, + 395, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 142, + 395, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 142, + 395, + 156 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 161, + 443, + 174 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 161, + 443, + 174 + ], + "spans": [ + { + "bbox": [ + 313, + 161, + 443, + 174 + ], + "type": "text", + "content": "5.1. Training & Evaluation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 178, + 555, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 555, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 555, + 334 + ], + "type": "text", + "content": "There are two independent training stages for Anomaly-OV. In Stage 1, the components of the anomaly expert are trained to obtain the token selection capability, targeting traditional ZSAD. This stage utilizes all of the data with anomaly labels in Anomaly-Instruct-125k. Similar to previous works [6, 110], when evaluating the model on the datasets contained in the training set, the corresponding datasets are replaced by VisA [115]. In Stage 2, the anomaly expert and visual encoder are frozen, while the projector and LLM are trainable. In addition to our instruction dataset, we sample around " + }, + { + "bbox": [ + 313, + 178, + 555, + 334 + ], + "type": "inline_equation", + "content": "350\\mathrm{k}" + }, + { + "bbox": [ + 313, + 178, + 555, + 334 + ], + "type": "text", + "content": " data from the original training recipe of " + }, + { + "bbox": [ + 313, + 178, + 555, + 334 + ], + "type": "inline_equation", + "content": "LLaVA-OneVision" + }, + { + "bbox": [ + 313, + 178, + 555, + 334 + ], + "type": "text", + "content": " to maintain the generalization ability. For more details on training, please refer to the supplementary." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 335, + 556, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 556, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 556, + 491 + ], + "type": "text", + "content": "The ZSAD performance for the anomaly expert is evaluated on nine benchmarks, including MVTec AD [2], VisA [115], AITEX [81], ELPV [16], BTAD [68], and MPDD [39] for industrial inspection, and BrainMRI [40], HeadCT [41], and Br35H [30] for medical diagnosis. AUROC (Area Under the Receiver Operating Characteristic) is leveraged to quantify the image-level AD performance. For text-based anomaly detection, both normal and anomaly data are employed to assess the accuracy by examining if the generated text contains the word Yes. Differently, only anomaly data are utilized to prompt the MLLMs to determine their anomaly reasoning capabilities since the justifications of normality are similar for different models." + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 318, + 498, + 551, + 574 + ], + "blocks": [ + { + "bbox": [ + 318, + 498, + 551, + 574 + ], + "lines": [ + { + "bbox": [ + 318, + 498, + 551, + 574 + ], + "spans": [ + { + "bbox": [ + 318, + 498, + 551, + 574 + ], + "type": "table", + "html": "
MethodMVTecVisAHeadCTBrainMRI
Full Model94.091.197.693.9
w/o. Look-back92.890.596.693.5
w/o. e+ & e-92.190.194.792.9
w/o. Q-Former91.789.992.895.1
w/o. WebAD88.588.991.293.4
", + "image_path": "5e5d27628c50ccc09f575834cbf0dd33c03975aa672e3fb46e9a21e1699bbc39.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 582, + 555, + 604 + ], + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 604 + ], + "type": "text", + "content": "Table 1. Ablation study for the anomaly expert of Anomaly-OV. w/o. Look-back refers to the removal of " + }, + { + "bbox": [ + 313, + 582, + 555, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_0^o" + }, + { + "bbox": [ + 313, + 582, + 555, + 604 + ], + "type": "text", + "content": " in LTFM." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 612, + 477, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 477, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 477, + 625 + ], + "type": "text", + "content": "5.2. Zero-Shot Anomaly Detection" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": "As shown in Table 2, compared with existing methods, the anomaly expert of Anomaly-OV achieves significant image-level AUROC improvements on most of the ZSAD benchmarks, which demonstrates that the text encoder widely applied in existing models is not necessary. The success of our model mainly originates from the extra data of WebAD (Table 1), which enables the model to learn more generic" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20375" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 550, + 194 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 550, + 194 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 550, + 194 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 550, + 194 + ], + "type": "table", + "html": "
ModelIndustrial DefectsMedical AnomaliesAverage
MVTec ADVisAAITEXELPVBTADMPDDBrainMRIHeadCTBr35H
CLIP [73]74.166.471.059.234.554.373.956.578.463.1
CoOp [108]88.862.866.273.066.855.161.378.486.070.9
WinCLIP [38]91.878.873.074.068.263.692.690.080.579.2
APRIL-GAN [11]86.278.057.665.573.673.089.389.193.178.4
AnoVL [19]92.579.272.570.680.368.988.781.688.480.3
AnomalyCLIP [110]91.582.162.281.588.377.090.393.494.684.5
AdaCLIP [6]89.285.864.579.788.676.094.891.497.785.3
Ours94.091.172.083.089.081.793.997.695.588.6
", + "image_path": "6785ac6a944ad439782076c5c22e88f54c165fec2fc8715737663589f75f81fa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 236, + 297, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 297, + 284 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 297, + 284 + ], + "type": "text", + "content": "semantics for normality and abnormality from the data distribution in the absence of the text encoder. This observation also reveals that large-scale in-the-wild online data can benefit zero-shot performance in anomaly detection." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 285, + 296, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 296, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 296, + 429 + ], + "type": "text", + "content": "While the Q-Former reduces the model performance on BrainMRI, it shows effectiveness on most benchmarks, indicating the importance of token aggregation. Similarly, the look-back information and two learnable embeddings are required for describing class-awareness abnormality and distinguishing positive and negative features, respectively. As previously discussed, the anomaly expert is responsible for selecting suspicious visual tokens for the LLM, and the significance maps shown in Figure 7 demonstrate the interpretable token selection mechanism. The high intensities are automatically distributed around the anomalous areas even without any supervision of the anomaly masks." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 64, + 437, + 136, + 492 + ], + "blocks": [ + { + "bbox": [ + 64, + 437, + 136, + 492 + ], + "lines": [ + { + "bbox": [ + 64, + 437, + 136, + 492 + ], + "spans": [ + { + "bbox": [ + 64, + 437, + 136, + 492 + ], + "type": "image", + "image_path": "64f50cfa285d9f428edf3462c2cab1e9bdb560c5289d050d2012e58827b12991.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 65, + 494, + 135, + 548 + ], + "blocks": [ + { + "bbox": [ + 65, + 494, + 135, + 548 + ], + "lines": [ + { + "bbox": [ + 65, + 494, + 135, + 548 + ], + "spans": [ + { + "bbox": [ + 65, + 494, + 135, + 548 + ], + "type": "image", + "image_path": "0981baf11f7c8786c1a1a03b1ed604d3cfd7543e706fbc85a5b2ca1336e324fb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 557, + 293, + 570 + ], + "lines": [ + { + "bbox": [ + 56, + 557, + 293, + 570 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 293, + 570 + ], + "type": "text", + "content": "Figure 7. Visualization of the significance map on VisA samples." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 141, + 439, + 212, + 492 + ], + "blocks": [ + { + "bbox": [ + 141, + 439, + 212, + 492 + ], + "lines": [ + { + "bbox": [ + 141, + 439, + 212, + 492 + ], + "spans": [ + { + "bbox": [ + 141, + 439, + 212, + 492 + ], + "type": "image", + "image_path": "67319f985b80c7763a079f5228d21f9acfc0ba797304cbc6bb0961b1e5991746.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 141, + 494, + 212, + 548 + ], + "blocks": [ + { + "bbox": [ + 141, + 494, + 212, + 548 + ], + "lines": [ + { + "bbox": [ + 141, + 494, + 212, + 548 + ], + "spans": [ + { + "bbox": [ + 141, + 494, + 212, + 548 + ], + "type": "image", + "image_path": "fc0d1a80b23d087fe076abffd6415aab48435be08723deaadd71ed472a66b3e1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 217, + 439, + 288, + 492 + ], + "blocks": [ + { + "bbox": [ + 217, + 439, + 288, + 492 + ], + "lines": [ + { + "bbox": [ + 217, + 439, + 288, + 492 + ], + "spans": [ + { + "bbox": [ + 217, + 439, + 288, + 492 + ], + "type": "image", + "image_path": "7c5d82ad91c401f62339a88ed2869ef5c6f00df1516237dd8d3ed6a69835e079.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 217, + 494, + 287, + 548 + ], + "blocks": [ + { + "bbox": [ + 217, + 494, + 287, + 548 + ], + "lines": [ + { + "bbox": [ + 217, + 494, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 217, + 494, + 287, + 548 + ], + "type": "image", + "image_path": "447de3945edc43d51dfcbac8faf9e4d3338c3b8041cb0ddd06aba3772e56e6b4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 576, + 231, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 231, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 231, + 590 + ], + "type": "text", + "content": "5.3. Anomaly Detection & Reasoning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "content": "With the strong capabilities of the anomaly expert for zero-shot detection and suspicious token selection, Anomaly-OV accomplishes significant improvement in text-based anomaly detection and reasoning over other open-sourced generalist MLLMs, as shown in Table 4. Here are a few observations: i) While a larger language model cannot guarantee better detection performance, it always provides greater reasoning ability; ii) Most of the existing MLLMs present much lower recall than precision, indicating their insensitivity to visual anomalies; iii) GPT-4o shows stronger reasoning ability" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 322, + 236, + 548, + 441 + ], + "blocks": [ + { + "bbox": [ + 55, + 201, + 555, + 224 + ], + "lines": [ + { + "bbox": [ + 55, + 201, + 555, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 555, + 224 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of Image-level AUROC on different ZSAD methods (some of the results are borrowed from [6, 110, 114]). The best and the second-best results are bolded and underlined, respectively. Please refer to the supplementary for more detailed results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 236, + 548, + 441 + ], + "lines": [ + { + "bbox": [ + 322, + 236, + 548, + 441 + ], + "spans": [ + { + "bbox": [ + 322, + 236, + 548, + 441 + ], + "type": "table", + "html": "
PCB Example: Excessive Solder
UserAre there any anomalies visible in the image?
GPT-4oThe image shows an HC-SR04 ultrasonic sensor. There don’t appear to be any visible anomalies with the sensor itself. The components and pins seem intact and properly aligned.
LLaVA-OVThe image shows an ultrasonic sensor module, specifically the HC-SR04 model. There are no visible anomalies in the image; it appears to be a standard representation of this type of sensor module.
Ours-7BYes, there is a noticeable anomaly in the image: a short circuit on the PCB, indicated by a white, crusty residue near the ‘Echo’ pin.
", + "image_path": "7d15dd20b305220bf22bf59e6d6ae8e9f20eaa8b777608b66d7da0991a4f4c48.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 449, + 552, + 461 + ], + "lines": [ + { + "bbox": [ + 314, + 449, + 552, + 461 + ], + "spans": [ + { + "bbox": [ + 314, + 449, + 552, + 461 + ], + "type": "text", + "content": "Table 3. Anomaly-OV presents more accurate anomaly detection." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 473, + 555, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 556 + ], + "type": "text", + "content": "compared to other open-sourced models. Table 3 and Table 5 provide the qualitative comparison of our Anomaly-OV with its base model LLaVA-OV-7B [44] and GPT-4o [72]. Both GPT-4o and LLaVA-OV show insensitivity to anomalous features and cannot accurately detect the anomaly in the image. Sometimes, GPT-4o knows the image is anomalous but fails to describe the anomalies precisely." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": "We provide the fine-tuned version of the base model LLaVA-OV-0.5B on Anomaly-Instruct-125k, which presents much higher accuracy and more balanced precision and recall than its original version. This demonstrates the effectiveness of our instruction-tuning dataset. By integrating the anomaly expert with the base model, our Anomaly-OV-0.5B achieves 0.08 accuracy and 0.06 F1-score improvements in text-based anomaly detection and better reasoning capability in low-level and complex settings. Equipped with a larger language model, Anomaly-OV-7B provides the best detection performance among all the existing MLLMs and shows comparable reasoning ability with GPT-4o. Notably, we observe that the anomaly expert restricts the detection perfor" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20376" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 238 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 238 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 238 + ], + "type": "table", + "html": "
ModelAnomaly DetectionLow-level ReasoningComplex Reasoning
AccuracyPrecisionRecallF1-scoreROUGE-LSBERTGPT-ScoreSBERTGPT-Score
GPT-4V [71]0.680.900.490.550.160.653.310.775.64
GPT-4o [72]0.700.830.710.680.240.714.840.816.89
Qwen2-VL-2B [87]0.650.870.550.590.220.551.940.744.26
Qwen2-VL-7B [87]0.760.910.690.750.250.613.090.684.62
InternVL-2-8B [13]0.740.780.810.760.230.733.690.805.08
InternVL-2-26B [13]0.730.860.660.680.210.744.130.805.49
IXC-2.5-7B [101]0.720.880.630.670.210.582.450.775.14
LLaVA-OV-0.5B [44]0.540.700.190.280.200.632.540.814.34
LLaVA-OV-7B [44]0.710.950.560.630.240.663.570.795.44
LLaVA-OV-0.5B*0.710.770.840.760.310.703.690.825.31
Anomaly-OV-0.5B0.790.860.830.820.330.723.870.835.67
Anomaly-OV-7B0.790.830.860.830.340.734.260.846.34
", + "image_path": "7149edef65c33bd2790ceca3d01d4a3790cbbe64059fd625bcbe5d54447f6f39.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 246, + 555, + 269 + ], + "lines": [ + { + "bbox": [ + 55, + 246, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 246, + 555, + 269 + ], + "type": "text", + "content": "Table 4. Quantitative comparison of text-based anomaly detection and reasoning for MLLMs. Notably, the Accuracy and F1-score for the anomaly expert of Anomaly-OV can be calculated as " + }, + { + "bbox": [ + 55, + 246, + 555, + 269 + ], + "type": "inline_equation", + "content": "\\{0.78, 0.77\\}" + }, + { + "bbox": [ + 55, + 246, + 555, + 269 + ], + "type": "text", + "content": " with threshold 0.5. * indicates the model is fine-tuned on our dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 281, + 296, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 306 + ], + "type": "text", + "content": "mance of Anomaly-OV. Therefore, the design of a stronger anomaly expert is suggested for future works." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 316, + 188, + 327 + ], + "lines": [ + { + "bbox": [ + 65, + 316, + 188, + 327 + ], + "spans": [ + { + "bbox": [ + 65, + 316, + 188, + 327 + ], + "type": "text", + "content": "Macaroni Example: Yellowish Spot" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 112, + 331, + 255, + 388 + ], + "blocks": [ + { + "bbox": [ + 112, + 331, + 255, + 388 + ], + "lines": [ + { + "bbox": [ + 112, + 331, + 255, + 388 + ], + "spans": [ + { + "bbox": [ + 112, + 331, + 255, + 388 + ], + "type": "image", + "image_path": "88fd3c7885aa08856f47465b158091edfd8ec145c9f1f282665e988598f21199.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 61, + 402, + 288, + 510 + ], + "blocks": [ + { + "bbox": [ + 65, + 390, + 83, + 399 + ], + "lines": [ + { + "bbox": [ + 65, + 390, + 83, + 399 + ], + "spans": [ + { + "bbox": [ + 65, + 390, + 83, + 399 + ], + "type": "text", + "content": "User" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 112, + 390, + 242, + 400 + ], + "lines": [ + { + "bbox": [ + 112, + 390, + 242, + 400 + ], + "spans": [ + { + "bbox": [ + 112, + 390, + 242, + 400 + ], + "type": "text", + "content": "Do you see any anomalies in the image?" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 402, + 288, + 510 + ], + "lines": [ + { + "bbox": [ + 61, + 402, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 61, + 402, + 288, + 510 + ], + "type": "table", + "html": "
GPT-4oThe image shows four pieces of elbow macaroni on a green background. The anomaly is that the macaroni pieces are not whole; they are cut in half.
LLaVA-OVThe image shows four pieces of pasta, specifically macaroni shells, arranged on a green textured surface. The pasta appears to be uniformly colored and shaped, with no visible defects or anomalies.
Ours-7BYes, there is an anomaly in the image. The bottom right pasta piece has a noticeable yellowish discoloration on its surface.
", + "image_path": "8fc53e512b4661142e28b5d2641742ab4d2b10a5f35a271445453bd0cc1f31d8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 519, + 291, + 531 + ], + "lines": [ + { + "bbox": [ + 56, + 519, + 291, + 531 + ], + "spans": [ + { + "bbox": [ + 56, + 519, + 291, + 531 + ], + "type": "text", + "content": "Table 5. Anomaly-OV presents more precise anomaly reasoning." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 534, + 125, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 125, + 546 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 125, + 546 + ], + "type": "text", + "content": "5.4. Extension" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 552, + 297, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 297, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 297, + 636 + ], + "type": "text", + "content": "With the generalization and multi-image processing capabilities of MLLMs, it is possible to build a unified assistant for visual inspection. Table 6 demonstrates the comprehensive knowledge of Anomaly-OV (without using Anomaly-ShapeNet [50] for training) on 3D and medical (testing set of BMAD [1]) AD & reasoning. More data, benchmarks, and investigation on a unified model are meaningful." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 646, + 128, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 646, + 128, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 646, + 128, + 658 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "text", + "content": "In this paper, we establish the first large-scale visual instruction tuning dataset, Anomaly-Instruct-125k, and the corresponding benchmark, VisA-D&R, to address the data scarcity issue for visual anomaly detection and reasoning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 328, + 281, + 397, + 292 + ], + "lines": [ + { + "bbox": [ + 328, + 281, + 397, + 292 + ], + "spans": [ + { + "bbox": [ + 328, + 281, + 397, + 292 + ], + "type": "text", + "content": "3D Example: Bulge" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 367, + 295, + 435, + 353 + ], + "blocks": [ + { + "bbox": [ + 367, + 295, + 435, + 353 + ], + "lines": [ + { + "bbox": [ + 367, + 295, + 435, + 353 + ], + "spans": [ + { + "bbox": [ + 367, + 295, + 435, + 353 + ], + "type": "image", + "image_path": "04a92329325abad74a691dc739a11fb8b70be3374afe8a3918381a73b2a06cd4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 442, + 295, + 510, + 353 + ], + "blocks": [ + { + "bbox": [ + 442, + 295, + 510, + 353 + ], + "lines": [ + { + "bbox": [ + 442, + 295, + 510, + 353 + ], + "spans": [ + { + "bbox": [ + 442, + 295, + 510, + 353 + ], + "type": "image", + "image_path": "56fbef7823f2c266934b4fbe1b91fbc04875b8ebd7da10dcf49856622d26e33d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 514, + 319, + 527, + 326 + ], + "lines": [ + { + "bbox": [ + 514, + 319, + 527, + 326 + ], + "spans": [ + { + "bbox": [ + 514, + 319, + 527, + 326 + ], + "type": "text", + "content": "···" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 329, + 355, + 346, + 365 + ], + "lines": [ + { + "bbox": [ + 329, + 355, + 346, + 365 + ], + "spans": [ + { + "bbox": [ + 329, + 355, + 346, + 365 + ], + "type": "text", + "content": "User" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 366, + 355, + 501, + 365 + ], + "lines": [ + { + "bbox": [ + 366, + 355, + 501, + 365 + ], + "spans": [ + { + "bbox": [ + 366, + 355, + 501, + 365 + ], + "type": "text", + "content": "Is there any anomaly present in the bowl?" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 328, + 369, + 358, + 377 + ], + "lines": [ + { + "bbox": [ + 328, + 369, + 358, + 377 + ], + "spans": [ + { + "bbox": [ + 328, + 369, + 358, + 377 + ], + "type": "text", + "content": "Ours-7B" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 365, + 369, + 538, + 388 + ], + "lines": [ + { + "bbox": [ + 365, + 369, + 538, + 388 + ], + "spans": [ + { + "bbox": [ + 365, + 369, + 538, + 388 + ], + "type": "text", + "content": "Yes, there is a bulge on the side of the bowl, which disrupts its otherwise smooth and symmetrical surface." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 329, + 392, + 434, + 402 + ], + "lines": [ + { + "bbox": [ + 329, + 392, + 434, + 402 + ], + "spans": [ + { + "bbox": [ + 329, + 392, + 434, + 402 + ], + "type": "text", + "content": "Medical Example: Pneumonia" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 365, + 405, + 503, + 463 + ], + "blocks": [ + { + "bbox": [ + 365, + 405, + 503, + 463 + ], + "lines": [ + { + "bbox": [ + 365, + 405, + 503, + 463 + ], + "spans": [ + { + "bbox": [ + 365, + 405, + 503, + 463 + ], + "type": "image", + "image_path": "02534aa62c713e51f9815a7cd628b8ba19ef8c12742b7262babaea7aca80ebc9.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 329, + 465, + 346, + 474 + ], + "lines": [ + { + "bbox": [ + 329, + 465, + 346, + 474 + ], + "spans": [ + { + "bbox": [ + 329, + 465, + 346, + 474 + ], + "type": "text", + "content": "User" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 365, + 465, + 526, + 475 + ], + "lines": [ + { + "bbox": [ + 365, + 465, + 526, + 475 + ], + "spans": [ + { + "bbox": [ + 365, + 465, + 526, + 475 + ], + "type": "text", + "content": "What is the most likely diagnosis for this patient?" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 328, + 478, + 358, + 487 + ], + "lines": [ + { + "bbox": [ + 328, + 478, + 358, + 487 + ], + "spans": [ + { + "bbox": [ + 328, + 478, + 358, + 487 + ], + "type": "text", + "content": "Ours-7B" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 365, + 478, + 539, + 527 + ], + "lines": [ + { + "bbox": [ + 365, + 478, + 539, + 527 + ], + "spans": [ + { + "bbox": [ + 365, + 478, + 539, + 527 + ], + "type": "text", + "content": "Based on the chest X-ray, the most likely diagnosis for this patient is pneumonia. The presence of bilateral patchy opacities in the lower lung fields is a common finding in pneumonia, which is an infection that causes inflammation in the air sacs of the lungs." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 329, + 530, + 359, + 540 + ], + "lines": [ + { + "bbox": [ + 329, + 530, + 359, + 540 + ], + "spans": [ + { + "bbox": [ + 329, + 530, + 359, + 540 + ], + "type": "text", + "content": "User" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 365, + 530, + 472, + 540 + ], + "lines": [ + { + "bbox": [ + 365, + 530, + 472, + 540 + ], + "spans": [ + { + "bbox": [ + 365, + 530, + 472, + 540 + ], + "type": "text", + "content": "What should the patient do next?" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 328, + 544, + 358, + 552 + ], + "lines": [ + { + "bbox": [ + 328, + 544, + 358, + 552 + ], + "spans": [ + { + "bbox": [ + 328, + 544, + 358, + 552 + ], + "type": "text", + "content": "Ours-7B" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 365, + 544, + 538, + 592 + ], + "lines": [ + { + "bbox": [ + 365, + 544, + 538, + 592 + ], + "spans": [ + { + "bbox": [ + 365, + 544, + 538, + 592 + ], + "type": "text", + "content": "The patient should consult a healthcare professional for further evaluation and appropriate treatment. This may include antibiotics to treat the infection, supportive care such as oxygen therapy, and monitoring of the patient's condition." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 331, + 604, + 534, + 615 + ], + "lines": [ + { + "bbox": [ + 331, + 604, + 534, + 615 + ], + "spans": [ + { + "bbox": [ + 331, + 604, + 534, + 615 + ], + "type": "text", + "content": "Table 6. Extension to 3D and medical AD & reasoning." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "content": "Then, a specialist MLLM, Anomaly-OV, targeting visual inspection is introduced to serve as the baseline in this domain. Anomaly-OV leverages an anomaly expert to assist the LLM with suspicious visual token selection and presents significant improvements on both traditional ZSAD and text-based anomaly detection and reasoning tasks over existing methods. Extension to 3D and medical domains is demonstrated." + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20377" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 91, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 91, + 297, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 91, + 297, + 145 + ], + "spans": [ + { + "bbox": [ + 66, + 91, + 297, + 145 + ], + "type": "text", + "content": "[1] Jinan Bao, Hanshi Sun, Hanqiu Deng, Yinsheng He, Zhaoxiang Zhang, and Xingyu Li. Bmad: Benchmarks for medical anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4042-4053, 2024. 1, 5, 6, 8, 14" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 148, + 296, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 148, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 66, + 148, + 296, + 203 + ], + "type": "text", + "content": "[2] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592-9600, 2019. 1, 5, 6, 14, 15" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 205, + 295, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 205, + 295, + 259 + ], + "spans": [ + { + "bbox": [ + 66, + 205, + 295, + 259 + ], + "type": "text", + "content": "[3] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4183–4192, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 261, + 296, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 261, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 66, + 261, + 296, + 327 + ], + "type": "text", + "content": "[4] Paul Bergmann, Xin Jin, David Sattlegger, and Carsten Steger. The mvtec 3d-ad dataset for unsupervised 3d anomaly detection and localization. In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications. SCITEPRESS - Science and Technology Publications, 2022. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 328, + 296, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 328, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 66, + 328, + 296, + 373 + ], + "type": "text", + "content": "[5] Tri Cao, Jiawen Zhu, and Guansong Pang. Anomaly detection under distribution shift. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6511-6523, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 374, + 296, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 374, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 66, + 374, + 296, + 430 + ], + "type": "text", + "content": "[6] Yunkang Cao, Jiangning Zhang, Luca Frittoli, Yuqi Cheng, Weiming Shen, and Giacomo Boracchi. Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly detection. In European Conference on Computer Vision, pages 55-72. Springer, 2025. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 431, + 296, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 431, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 66, + 431, + 296, + 496 + ], + "type": "text", + "content": "[7] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 498, + 296, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 498, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 66, + 498, + 296, + 542 + ], + "type": "text", + "content": "[8] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions, 2023. 5, 14" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 544, + 296, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 544, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 66, + 544, + 296, + 588 + ], + "type": "text", + "content": "[9] Qiyu Chen, Huiyuan Luo, Chengkan Lv, and Zhengtao Zhang. A unified anomaly synthesis strategy with gradient ascent for industrial anomaly detection and localization. arXiv preprint arXiv:2407.09359, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 590, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 590, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 61, + 590, + 296, + 656 + ], + "type": "text", + "content": "[10] Runnan Chen, Youquan Liu, Lingdong Kong, Xinge Zhu, Yuexin Ma, Yikang Li, Yuenan Hou, Yu Qiao, and Wenping Wang. Clip2scene: Towards label-efficient 3d scene understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7030, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 658, + 296, + 712 + ], + "type": "text", + "content": "[11] Xuhai Chen, Yue Han, and Jiangning Zhang. April-gan: A zero-/few-shot anomaly classification and segmentation method for cvpr 2023 vand workshop challenge tracks 1&2: 1st place on zero-shot ad and 4th place on few-shot ad, 2023. 2, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 117 + ], + "type": "text", + "content": "[12] Yuanhong Chen, Yu Tian, Guansong Pang, and Gustavo Carneiro. Deep one-class classification via interpolated gaussian descriptor. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 383-392, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 118, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 555, + 228 + ], + "type": "text", + "content": "[13] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, Ji Ma, Jiaqi Wang, Xiaoyi Dong, Hang Yan, Hewei Guo, Conghui He, Botian Shi, Zhenjiang Jin, Chao Xu, Bin Wang, Xingjian Wei, Wei Li, Wenjian Zhang, Bo Zhang, Pinlong Cai, Licheng Wen, Xiangchao Yan, Min Dou, Lewei Lu, Xizhou Zhu, Tong Lu, Dahua Lin, Yu Qiao, Jifeng Dai, and Wenhai Wang. How far are we to gpt-4v? closing the gap to commercial multimodal models with opensource suites, 2024. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 229, + 555, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 229, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 320, + 229, + 555, + 295 + ], + "type": "text", + "content": "[14] Zebang Cheng, Zhi-Qi Cheng, Jun-Yan He, Jingdong Sun, Kai Wang, Yuxiang Lin, Zheng Lian, Xiaojiang Peng, and Alexander Hauptmann. Emotion-llama: Multimodal emotion recognition and reasoning with instruction tuning. In Conference on Neural Information Processing Systems, 2024. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 297, + 555, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 297, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 320, + 297, + 555, + 361 + ], + "type": "text", + "content": "[15] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 363, + 555, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 363, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 320, + 363, + 555, + 418 + ], + "type": "text", + "content": "[16] Sergiu Deitsch, Vincent Christlein, Stephan Berger, Claudia Buerhop-Lutz, Andreas Maier, Florian Gallwitz, and Christian Riess. Automatic classification of defective photovoltaic module cells in electroluminescence images. Solar Energy, 185:455-468, 2019. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 420, + 555, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 420, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 320, + 420, + 555, + 485 + ], + "type": "text", + "content": "[17] Chenghao Deng, Haote Xu, Xiaolu Chen, Haodi Xu, Xiaotong Tu, Xinghao Ding, and Yue Huang. Simclip: Refining image-text alignment with simple prompts for zero-/few-shot anomaly detection. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 1761-1770, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 487, + 555, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 487, + 555, + 531 + ], + "spans": [ + { + "bbox": [ + 320, + 487, + 555, + 531 + ], + "type": "text", + "content": "[18] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In IEEE/CVF conference on computer vision and pattern recognition, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 533, + 555, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 533, + 555, + 577 + ], + "spans": [ + { + "bbox": [ + 320, + 533, + 555, + 577 + ], + "type": "text", + "content": "[19] Hanqiu Deng, Zhaoxiang Zhang, Jinan Bao, and Xingyu Li. Anovl: Adapting vision-language models for unified zero-shot anomaly localization. arXiv preprint arXiv:2308.15939, 2023. 1, 2, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 578, + 555, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 578, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 320, + 578, + 555, + 612 + ], + "type": "text", + "content": "[20] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 613, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 613, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 320, + 613, + 555, + 689 + ], + "type": "text", + "content": "[21] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 691, + 555, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 691, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 320, + 691, + 555, + 714 + ], + "type": "text", + "content": "[22] Zheng Fang, Xiaoyang Wang, Haocheng Li, Jiejie Liu, Qiugui Hu, and Jimin Xiao. Fastrecon: Few-shot industrial" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20378" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 80, + 73, + 297, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 73, + 297, + 106 + ], + "spans": [ + { + "bbox": [ + 80, + 73, + 297, + 106 + ], + "type": "text", + "content": "anomaly detection via fast feature reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17481-17490, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 107, + 297, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 107, + 297, + 162 + ], + "spans": [ + { + "bbox": [ + 61, + 107, + 297, + 162 + ], + "type": "text", + "content": "[23] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J Black. Chatpose: Chatting about 3d human pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2093-2103, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 163, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 163, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 61, + 163, + 296, + 207 + ], + "type": "text", + "content": "[24] Tharindu Fernando, Harshala Gammulle, Simon Denman, Sridha Sridharan, and Clinton Fookes. Deep learning for medical anomaly detection-a survey. ACM Computing Surveys (CSUR), 54(7):1-37, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 209, + 296, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 209, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 61, + 209, + 296, + 252 + ], + "type": "text", + "content": "[25] Matic Fučka, Vitjan Zavrtanik, and Danijel Skočaj. Transfusion-a transparency-based diffusion model for anomaly detection. In European conference on computer vision, pages 91-108. Springer, 2025. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 254, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 254, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 61, + 254, + 296, + 285 + ], + "type": "text", + "content": "[26] Google. Google-images-search 1.4.7, 2024. https://pypi.org/project/Google-Images-Search.6, 14" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 287, + 296, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 287, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 61, + 287, + 296, + 342 + ], + "type": "text", + "content": "[27] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Hao Li, Ming Tang, and Jinqiao Wang. Filo: Zero-shot anomaly detection by fine-grained description and high-quality localization. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2041-2049, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 343, + 296, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 343, + 296, + 398 + ], + "spans": [ + { + "bbox": [ + 61, + 343, + 296, + 398 + ], + "type": "text", + "content": "[28] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. Anomalygpt: Detecting industrial anomalies using large vision-language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1932-1940, 2024. 1, 2, 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 399, + 296, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 399, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 61, + 399, + 296, + 443 + ], + "type": "text", + "content": "[29] Yuxiang Guo, Faizan Siddiqui, Yang Zhao, Rama Chellappa, and Shao-Yuan Lo. Stimuvar: Spatiotemporal stimuli-aware video affective reasoning with multimodal large language models. arXiv preprint arXiv:2409.00304, 2024. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 445, + 296, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 445, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 61, + 445, + 296, + 464 + ], + "type": "text", + "content": "[30] Ahmed Hamada. Br35h: Brain tumor detection 2020, 2020. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 468, + 296, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 468, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 61, + 468, + 296, + 533 + ], + "type": "text", + "content": "[31] Changhee Han, Leonardo Rundo, Kohei Murao, Tomoyuki Noguchi, Yuki Shimahara, Zoltan Adam Milacski, Saori Koshino, Evis Sala, Hideki Nakayama, and Shin'ichi Satoh. Madgan: Unsupervised medical anomaly detection gan using multiple adjacent brain mri slice reconstruction. BMC bioinformatics, 22:1-20, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 535, + 296, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 535, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 61, + 535, + 296, + 589 + ], + "type": "text", + "content": "[32] Liren He, Zhengkai Jiang, Jinlong Peng, Liang Liu, Qianggang Du, Xiaobin Hu, Wenbing Zhu, Mingmin Chi, Yabiao Wang, and Chengjie Wang. Learning unified reference representation for unsupervised multi-class anomaly detection. arXiv preprint arXiv:2403.11561, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 590, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 590, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 61, + 590, + 296, + 634 + ], + "type": "text", + "content": "[33] Chih-Hui Ho, Kuan-Chuan Peng, and Nuno Vasconcelos. Long-tailed anomaly detection with learnable class names. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12435-12446, 2024." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 635, + 296, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 635, + 296, + 690 + ], + "spans": [ + { + "bbox": [ + 61, + 635, + 296, + 690 + ], + "type": "text", + "content": "[34] Jinlei Hou, Yingying Zhang, Qiaoyong Zhong, Di Xie, Shiliang Pu, and Hong Zhou. Divide-and-assemble: Learning block-wise memory for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8791-8800, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 61, + 692, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 692, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 692, + 296, + 713 + ], + "type": "text", + "content": "[35] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 556, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 339, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 555, + 95 + ], + "type": "text", + "content": "few-shot anomaly detection. In European Conference on Computer Vision, pages 303-319. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 97, + 556, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 97, + 556, + 162 + ], + "spans": [ + { + "bbox": [ + 320, + 97, + 556, + 162 + ], + "type": "text", + "content": "[36] Chaoqin Huang, Aofan Jiang, Jinghao Feng, Ya Zhang, Xin chao Wang, and Yanfeng Wang. Adapting visual-language models for generalizable anomaly detection in medical images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11375-11385, 2024. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 163, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 163, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 320, + 163, + 555, + 217 + ], + "type": "text", + "content": "[37] Brian KS Isaac-Medina, Yona Falinie A Gaus, Neelanjan Bhowmik, and Toby P Breckon. Towards open-world object-based anomaly detection via self-supervised outlier synthesis. In European Conference on Computer Vision (ECCV), 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 220, + 555, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 220, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 320, + 220, + 555, + 275 + ], + "type": "text", + "content": "[38] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19606-19616, 2023. 1, 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 276, + 555, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 276, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 320, + 276, + 555, + 341 + ], + "type": "text", + "content": "[39] Stepan Jezek, Martin Jonak, Radim Burget, Pavel Dvorak, and Milos Skotak. Deep learning-based defect detection of metal parts: evaluating current methods in complex conditions. In 2021 13th International congress on ultra modern telecommunications and control systems and workshops (ICUMT), pages 66-71. IEEE, 2021. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 342, + 555, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 342, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 320, + 342, + 555, + 364 + ], + "type": "text", + "content": "[40] Pranita Balaji Kanade and PP Gumaste. Brain tumor detection using mri images. *Brain*, 3(2):146-150, 2015. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 365, + 548, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 365, + 548, + 376 + ], + "spans": [ + { + "bbox": [ + 320, + 365, + 548, + 376 + ], + "type": "text", + "content": "[41] Felipe Campos Kitamura. Head ct - hemorrhage, 2018. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 378, + 555, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 378, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 320, + 378, + 555, + 431 + ], + "type": "text", + "content": "[42] Mingyu Lee and Jongwon Choi. Text-guided variational image generation for industrial anomaly detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26519-26528, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 434, + 554, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 434, + 554, + 477 + ], + "spans": [ + { + "bbox": [ + 320, + 434, + 554, + 477 + ], + "type": "text", + "content": "[43] Aodong Li, Chen Qiu, Marius Kloft, Padhraic Smyth, Maja Rudolph, and Stephan Mandt. Zero-shot anomaly detection via batch normalization. Advances in Neural Information Processing Systems, 36, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 479, + 555, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 479, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 320, + 479, + 555, + 521 + ], + "type": "text", + "content": "[44] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer, 2024. 1, 3, 7, 8, 15" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 523, + 555, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 523, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 320, + 523, + 555, + 589 + ], + "type": "text", + "content": "[45] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. LLaVA-med: Training a large language-and-vision assistant for biomedicine in one day. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 1, 15" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 590, + 554, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 590, + 554, + 634 + ], + "spans": [ + { + "bbox": [ + 320, + 590, + 554, + 634 + ], + "type": "text", + "content": "[46] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In IEEE/CVF conference on computer vision and pattern recognition, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 635, + 555, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 635, + 555, + 679 + ], + "spans": [ + { + "bbox": [ + 320, + 635, + 555, + 679 + ], + "type": "text", + "content": "[47] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava-last-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 1, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 681, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 681, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 681, + 555, + 713 + ], + "type": "text", + "content": "[48] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Interna-" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20379" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 81, + 73, + 296, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 296, + 94 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 296, + 94 + ], + "type": "text", + "content": "tional conference on machine learning, pages 12888-12900. \nPMLR, 2022. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 96, + 297, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 96, + 297, + 149 + ], + "spans": [ + { + "bbox": [ + 61, + 96, + 297, + 149 + ], + "type": "text", + "content": "[49] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 151, + 296, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 151, + 296, + 225 + ], + "spans": [ + { + "bbox": [ + 61, + 151, + 296, + 225 + ], + "type": "text", + "content": "[50] Wenqiao Li, Xiaohao Xu, Yao Gu, Bozhong Zheng, Shenghua Gao, and Yingna Wu. Towards scalable 3d anomaly detection and localization: A benchmark via 3d anomaly synthesis and a self-supervised learning network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22207-22216, 2024. 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 228, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 296, + 293 + ], + "type": "text", + "content": "[51] Xiaofan Li, Zhizhong Zhang, Xin Tan, Chengwei Chen, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Prompt: Learning prompts with only normal samples for few-shot anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16838-16848, 2024. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 294, + 296, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 294, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 294, + 296, + 338 + ], + "type": "text", + "content": "[52] Yiting Li, Adam Goodge, Fayao Liu, and Chuan-Sheng Foo. Promptad: Zero-shot anomaly detection using text prompts. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1093-1102, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 338, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 338, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 62, + 338, + 296, + 403 + ], + "type": "text", + "content": "[53] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 404, + 294, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 404, + 294, + 436 + ], + "spans": [ + { + "bbox": [ + 62, + 404, + 294, + 436 + ], + "type": "text", + "content": "[54] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 437, + 296, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 437, + 296, + 503 + ], + "spans": [ + { + "bbox": [ + 62, + 437, + 296, + 503 + ], + "type": "text", + "content": "[55] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 504, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 504, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 62, + 504, + 294, + 569 + ], + "type": "text", + "content": "[56] Yuqi Lin, Minghao Chen, Wenxiao Wang, Boxi Wu, Ke Li, Binbin Lin, Haifeng Liu, and Xiaofei He. Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15305-15314, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 571, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 571, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 571, + 296, + 612 + ], + "type": "text", + "content": "[57] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1, 2, 3, 5, 6, 16" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 614, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 296, + 658 + ], + "type": "text", + "content": "[58] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 658, + 296, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 296, + 691 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 296, + 691 + ], + "type": "text", + "content": "[59] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024.3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 692, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 692, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 692, + 296, + 713 + ], + "type": "text", + "content": "[60] Jiaqi Liu, Guoyang Xie, ruitao chen, Xinpeng Li, Jinbao Wang, Yong Liu, Chengjie Wang, and Feng Zheng. Real3d" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 339, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 555, + 106 + ], + "type": "text", + "content": "AD: A dataset of point cloud anomaly detection. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 106, + 555, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 106, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 321, + 106, + 555, + 172 + ], + "type": "text", + "content": "[61] Jie Liu, Yixiao Zhang, Jie-Neng Chen, Junfei Xiao, Yongyi Lu, Bennett A Landman, Yixuan Yuan, Alan Yuille, Yucheng Tang, and Zongwei Zhou. Clip-driven universal model for organ segmentation and tumor detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21152-21164, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 173, + 555, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 173, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 321, + 173, + 555, + 216 + ], + "type": "text", + "content": "[62] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 217, + 555, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 217, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 321, + 217, + 555, + 258 + ], + "type": "text", + "content": "[63] Shao-Yuan Lo, Poojan Oza, and Vishal M Patel. Adversarially robust one-class novelty detection. In IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 261, + 555, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 261, + 555, + 294 + ], + "spans": [ + { + "bbox": [ + 321, + 261, + 555, + 294 + ], + "type": "text", + "content": "[64] Ilya Loshchilov and Frank Hutter. SGDR: Stochastic gradient descent with warm restarts. In International Conference on Learning Representations, 2017. 15" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 294, + 554, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 294, + 554, + 327 + ], + "spans": [ + { + "bbox": [ + 321, + 294, + 554, + 327 + ], + "type": "text", + "content": "[65] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 15" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 327, + 555, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 327, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 321, + 327, + 555, + 392 + ], + "type": "text", + "content": "[66] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 393, + 554, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 393, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 321, + 393, + 554, + 425 + ], + "type": "text", + "content": "[67] Hui Lv and Qianru Sun. Video anomaly detection and explanation via large language models. arXiv preprint arXiv:2401.05702, 2024. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 426, + 555, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 426, + 555, + 482 + ], + "spans": [ + { + "bbox": [ + 321, + 426, + 555, + 482 + ], + "type": "text", + "content": "[68] Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. Vt-adt: A vision transformer network for image anomaly detection and localization. In 2021 IEEE 30th International Symposium on Industrial Electronics (ISIE), pages 01-06. IEEE, 2021. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 482, + 555, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 482, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 321, + 482, + 555, + 536 + ], + "type": "text", + "content": "[69] Shancong Mou, Xiaoyi Gu, Meng Cao, Haoping Bai, Ping Huang, Jiulong Shan, and Jianjun Shi. RGI: robust GAN-inversion for mask-free image inpainting and unsupervised pixel-wise anomaly detection. In The Eleventh International Conference on Learning Representations, 2023. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 537, + 555, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 537, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 321, + 537, + 555, + 592 + ], + "type": "text", + "content": "[70] Ming Nie, Renyuan Peng, Chunwei Wang, Xinyue Cai, Jianhua Han, Hang Xu, and Li Zhang. Reason2drive: Towards interpretable and chain-based reasoning for autonomous driving. In European Conference on Computer Vision, pages 292-308. Springer, 2025. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 593, + 554, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 593, + 554, + 613 + ], + "spans": [ + { + "bbox": [ + 321, + 593, + 554, + 613 + ], + "type": "text", + "content": "[71] OpenAI. Gpt-4v(isdiction) system card, 2023. https://openai.com/index/gpt-4v-system-card.5,8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 614, + 555, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 614, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 321, + 614, + 555, + 645 + ], + "type": "text", + "content": "[72] OpenAI. Gpt-4o system card, 2024. https://openai.com/index/gpt-4o-system-card. 2, 5, 6, 7, 8, 14, 15" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 647, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 647, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 647, + 555, + 713 + ], + "type": "text", + "content": "[73] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 7, 14" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20380" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 73, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 296, + 116 + ], + "type": "text", + "content": "[74] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 295, + 139 + ], + "type": "text", + "content": "[75] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks, 2019. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 141, + 296, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 141, + 296, + 184 + ], + "spans": [ + { + "bbox": [ + 62, + 141, + 296, + 184 + ], + "type": "text", + "content": "[76] Tal Reiss and Yedid Hoshen. Mean-shifted contrastive loss for anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2155-2162, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 186, + 296, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 186, + 296, + 241 + ], + "spans": [ + { + "bbox": [ + 62, + 186, + 296, + 241 + ], + "type": "text", + "content": "[77] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318-14328, 2022. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 243, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 243, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 62, + 243, + 296, + 297 + ], + "type": "text", + "content": "[78] Fumiaki Sato, Ryo Hachiuma, and Taiki Sekii. Prompt-guided zero-shot anomaly action recognition using pretrained deep skeleton features. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6471-6480, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 299, + 296, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 299, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 62, + 299, + 296, + 352 + ], + "type": "text", + "content": "[79] Eli Schwartz, Assaf Arbelle, Leonid Karlinsky, Sivan Harary, Florian Scheidegger, Sivan Doveh, and Raja Giryes. Mae-day: Mae for few-and zero-shot anomaly-detection. Computer Vision and Image Understanding, 241:103958, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 354, + 296, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 354, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 354, + 296, + 430 + ], + "type": "text", + "content": "[80] Pierre Sermanet, Tianli Ding, Jeffrey Zhao, Fei Xia, Debidatta Dwibedi, Keerthana Gopalakrishnan, Christine Chan, Gabriel Dulac-Arnold, Sharath Maddineni, Nikhil J Joshi, et al. Robovqa: Multimodal long-horizon reasoning for robotics. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 645-652. IEEE, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 433, + 296, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 433, + 296, + 476 + ], + "spans": [ + { + "bbox": [ + 62, + 433, + 296, + 476 + ], + "type": "text", + "content": "[81] Javier Silvestre-Blanes, Teresa Albero-Albero, Ignacio Miralles, Rubén Pérez-Llorens, and Jorge Moreno. A public fabric database for defect detection methods and results. Autex Research Journal, 19(4):363–374, 2019. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 478, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 478, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 478, + 296, + 521 + ], + "type": "text", + "content": "[82] Luc PJ Sträter, Mohammadreza Salehi, Efstratios Gavves, Cees GM Snoek, and Yuki M Asano. Generalad: Anomaly detection across domains by attending to distorted features. arXiv preprint arXiv:2407.12427, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 523, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 523, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 523, + 296, + 555 + ], + "type": "text", + "content": "[83] Haomiao Sun, Mingjie He, Tianheng Lian, Hu Han, and Shiguang Shan. Face-mllm: A large face perception model, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 557, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 296, + 612 + ], + "type": "text", + "content": "[84] Jiaqi Tang, Hao Lu, Xiaogang Xu, Ruizheng Wu, Sixing Hu, Tong Zhang, Tsz Wa Cheng, Ming Ge, Ying-Cong Chen, and Fugee Tsung. An incremental unified framework for small defect inspection. In European Conference on Computer Vision, pages 307-324. Springer, 2025. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 614, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 296, + 668 + ], + "type": "text", + "content": "[85] Tran Dinh Tien, Anh Tuan Nguyen, Nguyen Hoang Tran, Ta Duc Huy, Soan Duong, Chanh D Tr Nguyen, and Steven QH Truong. Revisiting reverse distillation for anomaly detection. In IEEE/CVF conference on computer vision and pattern recognition, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 296, + 713 + ], + "type": "text", + "content": "[86] Hualiang Wang, Yi Li, Huifeng Yao, and Xiaomeng Li. Clipn for zero-shot ood detection: Teaching clip to say no. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1802-1812, 2023. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 139 + ], + "type": "text", + "content": "[87] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution, 2024. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 141, + 555, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 141, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 320, + 141, + 555, + 194 + ], + "type": "text", + "content": "[88] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. In Conference on Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 197, + 555, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 197, + 555, + 250 + ], + "spans": [ + { + "bbox": [ + 320, + 197, + 555, + 250 + ], + "type": "text", + "content": "[89] Qi Wei, Yinhao Ren, Rui Hou, Bibo Shi, Joseph Y Lo, and Lawrence Carin. Anomaly detection for medical images based on a one-class classification. In Medical Imaging 2018: Computer-Aided Diagnosis, pages 375-380. SPIE, 2018. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 253, + 555, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 253, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 320, + 253, + 555, + 307 + ], + "type": "text", + "content": "[90] Julia Wolleb, Florentin Bieder, Robin Sandkühler, and Philippe C Cattin. Diffusion models for medical anomaly detection. In International Conference on Medical image computing and computer-assisted intervention, pages 35-45. Springer, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 309, + 555, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 309, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 320, + 309, + 555, + 363 + ], + "type": "text", + "content": "[91] Binzhu Xie, Sicheng Zhang, Zitang Zhou, Bo Li, Yuanhan Zhang, Jack Hessel, Jingkang Yang, and Ziwei Liu. Funqa: Towards surprising video comprehension. In European Conference on Computer Vision, pages 39-57. Springer, 2025. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 365, + 554, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 365, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 320, + 365, + 554, + 409 + ], + "type": "text", + "content": "[92] Guoyang Xie, Jinbao Wang, Jiaqi Liu, Yaochu Jin, and Feng Zheng. Pushing the limits of fewshot anomaly detection in industry vision: Graphcore. In The Eleventh International Conference on Learning Representations, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 411, + 555, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 411, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 320, + 411, + 555, + 475 + ], + "type": "text", + "content": "[93] Hongxia Xie, Chu-Jun Peng, Yu-Wen Tseng, Hung-Jen Chen, Chan-Feng Hsu, Hong-Han Shuai, and Wen-Huang Cheng. Emovit: Revolutionizing emotion insights with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26596-26605, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 478, + 554, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 478, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 320, + 478, + 554, + 521 + ], + "type": "text", + "content": "[94] Yuchen Yang, Kwonjoon Lee, Behzad Dariush, Yinzhi Cao, and Shao-Yuan Lo. Follow the rules: reasoning for video anomaly detection with large language models. arXiv preprint arXiv:2407.10299, 2024. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 523, + 555, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 523, + 555, + 577 + ], + "spans": [ + { + "bbox": [ + 320, + 523, + 555, + 577 + ], + "type": "text", + "content": "[95] Hang Yao, Ming Liu, Haolin Wang, Zhicun Yin, Zifei Yan, Xiaopeng Hong, and Wangmeng Zuo. Glad: Towards better reconstruction with global and local adaptive diffusion models for unsupervised anomaly detection. arXiv preprint arXiv:2406.07487, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 579, + 555, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 579, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 320, + 579, + 555, + 622 + ], + "type": "text", + "content": "[96] Huanjin Yao, Wenhao Wu, Taojiannan Yang, YuXin Song, Mengxi Zhang, Haocheng Feng, Yifan Sun, Zhiheng Li, Wanli Ouyang, and Jingdong Wang. Dense connector for mllms, 2024. 17" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 624, + 555, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 624, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 320, + 624, + 555, + 668 + ], + "type": "text", + "content": "[97] Xincheng Yao, Ruoqi Li, Zefeng Qian, Lu Wang, and Chongyang Zhang. Hierarchical gaussian mixture normalizing flow modeling for unified anomaly detection. arXiv preprint arXiv:2403.13349, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 670, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 670, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 670, + 555, + 713 + ], + "type": "text", + "content": "[98] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. In Advances in Neural Information Processing Systems, 2022. 1" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20381" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 297, + 116 + ], + "type": "text", + "content": "[99] Vitjan Zavrtanik, Matej Kristan, and Danijel Skoçaj. Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In IEEE/CVF international conference on computer vision, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 161 + ], + "type": "text", + "content": "[100] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 296, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 296, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 296, + 249 + ], + "type": "text", + "content": "[101] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, Songyang Zhang, Wenwei Zhang, Yining Li, Yang Gao, Peng Sun, Xinyue Zhang, Wei Li, Jingwen Li, Wenhai Wang, Hang Yan, Conghui He, Xingcheng Zhang, Kai Chen, Jifeng Dai, Yu Qiao, Dahua Lin, and Jiaqi Wang. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output, 2024. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 250, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 296, + 293 + ], + "type": "text", + "content": "[102] Xuan Zhang, Shiyu Li, Xi Li, Ping Huang, Jiulong Shan, and Ting Chen. Destseg: Segmentation guided denoising studentteacher for anomaly detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 294, + 296, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 296, + 348 + ], + "type": "text", + "content": "[103] Ximiao Zhang, Min Xu, Dehui Qiu, Ruixin Yan, Ning Lang, and Xiuzhuang Zhou. Mediclip: Adapting clip for few-shot medical image anomaly detection. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 458-468. Springer, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 350, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 296, + 403 + ], + "type": "text", + "content": "[104] Ximiao Zhang, Min Xu, and Xiuzhuang Zhou. Realnet: A feature selection network with realistic synthetic anomaly for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16699-16708, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 404, + 296, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 296, + 448 + ], + "type": "text", + "content": "[105] Zhuosheng Zhang, Aston Zhang, Mu Li, hai zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. Transactions on Machine Learning Research, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 449, + 296, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 296, + 502 + ], + "type": "text", + "content": "[106] He Zhao, Yuexiang Li, Nanjun He, Kai Ma, Leyuan Fang, Huiqi Li, and Yefeng Zheng. Anomaly detection for medical images using self-supervised and translation-consistent features. IEEE Transactions on Medical Imaging, 40(12): 3641-3651, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 504, + 296, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 296, + 547 + ], + "type": "text", + "content": "[107] Gengze Zhou, Yicong Hong, and Qi Wu. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7641-7649, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 548, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 296, + 590 + ], + "type": "text", + "content": "[108] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 296, + 635 + ], + "type": "text", + "content": "[109] Kaiwen Zhou, Kwonjoon Lee, Teruhisa Mitsu, and Xin Eric Wang. Vicor: Bridging visual understanding and commonsense reasoning with large language models. In Findings of the Association for Computational Linguistics, 202'. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 296, + 689 + ], + "type": "text", + "content": "[110] Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. AnomalyCLIP: Object-agnostic prompt learning for zero-shot anomaly detection. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "type": "text", + "content": "[111] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. MiniGPT-4: Enhancing vision-language" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 308 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 338, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 338, + 73, + 555, + 106 + ], + "type": "text", + "content": "understanding with advanced large language models. In The Twelfth International Conference on Learning Representations, 2024. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 555, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 161 + ], + "type": "text", + "content": "[112] Jiawen Zhu and Guansong Pang. Toward generalist anomaly detection via in-context residual learning with few-shot sample prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17826-17836, 2024. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 555, + 217 + ], + "type": "text", + "content": "[113] Jiaqi Zhu, Shaofeng Cai, Fang Deng, Beng Chin Ooi, and Junran Wu. Do llms understand visual anomalies? uncovering llm's capabilities in zero-shot anomaly detection. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 48-57, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 219, + 555, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 555, + 251 + ], + "type": "text", + "content": "[114] Jiawen Zhu, Yew-Soon Ong, Chunhua Shen, and Guansong Pang. Fine-grained abnormality prompt learning for zero-shot anomaly detection, 2024. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 253, + 555, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 555, + 308 + ], + "type": "text", + "content": "[115] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In European Conference on Computer Vision, pages 392-408. Springer, 2022. 6, 15" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20382" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Towards a Universal Synthetic Video Detector_ From Face or Background Manipulations to Fully AI-Generated Content/49360ca2-47e3-4c7b-81b1-6a87621996c4_content_list.json b/2025/Towards a Universal Synthetic Video Detector_ From Face or Background Manipulations to Fully AI-Generated Content/49360ca2-47e3-4c7b-81b1-6a87621996c4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b35a6bdd8af44ac7dc20bebbb4b23950f51e166c --- /dev/null +++ b/2025/Towards a Universal Synthetic Video Detector_ From Face or Background Manipulations to Fully AI-Generated Content/49360ca2-47e3-4c7b-81b1-6a87621996c4_content_list.json @@ -0,0 +1,1635 @@ +[ + { + "type": "text", + "text": "Towards a Universal Synthetic Video Detector: From Face or Background Manipulations to Fully AI-Generated Content", + "text_level": 1, + "bbox": [ + 125, + 128, + 870, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rohit Kundu $^{1,2}$ , Hao Xiong $^{1}$ , Vishal Mohanty $^{1}$ , Athula Balachandran $^{1}$ , Amit K. Roy-Chowdhury $^{2}$ , $^{1}$ Google LLC, $^{2}$ University of California, Riverside", + "bbox": [ + 112, + 202, + 883, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{rohit.kundu@email, amertc@ece}.ucr.edu; {rohitkun, haoxg, vishal, athula}@google.com", + "bbox": [ + 120, + 241, + 870, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing DeepFake detection techniques primarily focus on facial manipulations, such as face-swapping or lip-syncing. However, advancements in text-to-video (T2V) and image-to-video (I2V) generative models now allow fully AI-generated synthetic content and seamless background alterations, challenging face-centric detection methods and demanding more versatile approaches.", + "bbox": [ + 89, + 324, + 483, + 430 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this, we introduce the Universal Network for Identifying Tampered and synthEtic videos (UNITE) model, which, unlike traditional detectors, captures full-frame manipulations. UNITE extends detection capabilities to scenarios without faces, non-human subjects, and complex background modifications. It leverages a transformer-based architecture that processes domain-agnostic features extracted from videos via the SigLIP-So400M foundation model. Given limited datasets encompassing both facial/background alterations and T2V/I2V content, we integrate task-irrelevant data alongside standard DeepFake datasets in training. We further mitigate the model's tendency to over-focus on faces by incorporating an attention-diversity (AD) loss, which promotes diverse spatial attention across video frames. Combining AD loss with cross-entropy improves detection performance across varied contexts. Comparative evaluations demonstrate that UNITE outperforms state-of-the-art detectors on datasets featuring face/background manipulations and fully synthetic T2V/I2V videos, showcasing its adaptability and generalizable detection capabilities.", + "bbox": [ + 89, + 431, + 483, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 782, + 220, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The rise of synthetic media, especially DeepFakes, has transformed content perception and interaction online. Hyper-realistic images produced by technologies like $\\mathrm{FLUX}_{1.1}$ [1] are challenging for even humans to identify as fake, and with the development of their video model underway, this underscores the critical need for effective detec", + "bbox": [ + 89, + 809, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7f4d537dbea8ce1cfa0f687526832537988d457109714d1137feb379a094bc2b.jpg", + "image_caption": [ + "Figure 1. Problem Overview: Existing DeepFake detection methods focus on identifying face-manipulated videos, most of which cannot perform inference unless there is a face detected in the video. However, with advancements like seamless background modifications (e.g., AVID [61]) and hyper-realistic content from games like GTA-V [24] and T2V/I2V models [9], a more comprehensive approach is needed. A model trained with only cross-entropy (CE) loss, using full frames, automatically focuses on the face, capturing temporal discontinuities through its transformer architecture, performing better than random $(\\approx)$ on T2V/I2V content but struggling with background manipulations. UNITE, with its attention-diversity (AD) loss, effectively detects both face/background manipulations and fully synthetic content." + ], + "image_footnote": [], + "bbox": [ + 535, + 290, + 885, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion methods for media generated by deep neural networks. Traditional DeepFake generators [28, 49] focus on manipulating human faces through face-swapping and lip-syncing. However, powerful text-to-video (T2V) and image-to-video (I2V) models [42, 52, 63] have expanded manipulation possibilities beyond faces.", + "bbox": [ + 511, + 700, + 905, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While conventional detectors [10, 45, 57] perform well on older, face-centric DeepFake datasets [17, 30, 43], they often struggle with newer manipulations involving full scenes or backgrounds. DeepFake-O-Meter [26] is an open-source tool (consisting of several state-of-the-art models) for DeepFake detection, but it cannot run inference unless a human face is visible in the image/video. The rapid spread", + "bbox": [ + 511, + 794, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28050", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of misinformation, particularly during critical periods such as elections, highlights the need for generalizable detection models capable of identifying diverse manipulations, including face, background, and fully AI-generated T2V/I2V content with/without human subjects (overview in Fig. 1).", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we present a Universal Network for Identifying Tampered and synthEtic videos or UNITE model, which detects both partially manipulated (foreground/background) and fully synthetic videos. Unlike detectors focused solely on face detection, UNITE analyzes entire frames, regardless of whether a human subject is present in the videos.", + "bbox": [ + 89, + 167, + 480, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given the inherent domain gaps in DeepFake datasets [44, 64], even when generated by similar techniques, we leverage the SigLIP-So400m [4] foundation model to extract domain-agnostic features. This serves as inputs to a learnable transformer with multi-head attention, enabling effective detection by capturing temporal inconsistencies in synthetic content. However, preliminary experiments show that training a transformer architecture solely with cross-entropy (CE) loss often leads to the attention heads converging on the face region (Fig. 3). As a result, the model struggles during inference when handling videos where a real human subject is placed in a manipulated background or content generated by T2V [8, 52] and I2V [35, 63] models. To address this, we introduce an \"attention-diversity\" (AD) loss that encourages attention heads to focus on different spatial regions, enhancing the model's ability to capture critical cues from both foreground and background.", + "bbox": [ + 91, + 275, + 482, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Due to the limited availability of open-source datasets covering face or background manipulations and fully AI-generated content, we employ innovative training strategies for UNITE. This includes integrating task-irrelevant data with standard DeepFake datasets to simulate AI-generated synthetic media. Beyond the popularly used FaceForensics++ [43] dataset, we utilize the SAIL-VOS-3D [24] dataset, originally designed for 3D video object segmentation in the game GTA-V. As this dataset is fully synthetic, it helps simulate AI-generated media, enhancing our model's ability to detect diverse forms of synthetic manipulation.", + "bbox": [ + 89, + 532, + 482, + 698 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 109, + 700, + 434, + 714 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose UNITE, a model for detecting partially manipulated (foreground/background) and fully synthetic videos, moving beyond face-centric DeepFake detection.", + "- Unlike detectors relying on face detection and cropping, our model can process full video frames and can detect fakes even if there are no human subjects.", + "- Using the SigLIP-So400m foundation model [4], we extract domain-agnostic features, enabling generalization to in-the-wild DeepFakes.", + "- We introduce an attention-diversity loss, encouraging attention heads to focus on diverse spatial regions, enhancing detection beyond face manipulations." + ], + "bbox": [ + 91, + 719, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Unlike existing methods that evaluate on a few datasets, we comprehensively assess UNITE on a broad range of face and synthetic datasets with various T2V/I2V generators, outperforming detection in foreground, background, and T2V/I2V manipulations.", + "bbox": [ + 513, + 90, + 903, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 180, + 653, + 195 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Face-centric DeepFake Detection: Methods such as [13, 19, 40] focus on modeling spatial inconsistencies for DeepFake detection. Concas et al. [13] track specific facial features (eyes, nose, mouth) in a high-frequency domain, as fake videos exhibit distinct behaviors in this space compared to real ones. However, due to their dataset-centric architectures, it is not a generalizable approach. PUDD [40] learns person-specific prototypes by modeling patterns from real videos and assessing how closely inference videos match these prototypes, making it less effective for detecting in-the-wild DeepFakes. LAA-Net [39] attempts to improve generalizability through a multi-task attention module that detects artifacts via heatmap and self-consistency regression, but it works by analyzing every still frame of a video, thus limiting its scalability. The authors in [36] tackled the challenging problem of detecting and localizing expression swaps, where the number of manipulated pixels is even fewer compared to identity swaps. Although the authors obtained good localization performance, their reliance on a CNN-based architecture limits their ability to capture temporal inconsistencies effectively.", + "bbox": [ + 511, + 205, + 903, + 522 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DPNet [50] and ID-Reveal [16] are identity-aware temporal artifact modeling methods: DPNet [50] focuses on interpretable, prototype-based detection of unnatural movements through dynamic feature representations, while IDReveal [16] leverages metric learning on real data to detect biometric inconsistencies. However these methods assume access to authentic reference videos, making them unfit for in-the-wild DeepFake detection. Shifting from identity-specific methods, TALL [57] constructs composite \"thumbnail\" images from four consecutive frames taken at random timestamps for temporal analysis. TI2Net [32] captures temporal anomalies by subtracting consecutive frame features but relies on the assumption that FaceSwap manipulations show pronounced inter-frame discrepancies.", + "bbox": [ + 511, + 523, + 903, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Choi et al. [11] approach DeepFake detection by targeting suppressed variance in style-based latent temporal features via a StyleGRU module. However, this method is limited to cropped face regions, making it ineffective for T2V/I2V or background-manipulated videos.", + "bbox": [ + 511, + 734, + 903, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Synthetic Video Detection: While there has been considerable effort in synthetic image detection [14, 15, 55], synthetic video detection is a scarcely explored area. DeMamba [9], a recently proposed synthetic video detector, analyzes small zones of video frames to capture how pixels change spatiotemporally. Its continuous scanning approach", + "bbox": [ + 511, + 810, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28051", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/293854bf54d05f4417715994a0c7ca60da622532c26dfd731035498dd8ac84a2.jpg", + "image_caption": [ + "Figure 2. UNITE architecture overview: We extract domain-agnostic features $(\\xi)$ using the SigLIP-So400m foundation model [4] to mitigate domain gaps between DeepFake datasets (Sec. 3.2). These embeddings, combined with positional encodings, are input to a transformer with multi-head attention and MLP layers (Sec. 3.3), culminating in a classifier for final predictions. AD-loss (Sec. 3.4) encourages the model attention to span diverse spatial regions." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 441, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "better tracks subtle changes and patterns, making it easier to spot manipulated or fake content. The authors also propose a million scale T2V/I2V dataset (which we call the DeMamba dataset hereforth) to evaluate their performance. However, the method was not evaluated on typical DeepFake datasets [30, 43, 53] and thus is not fit for human-centric DeepFake detection. In contrast our UNITE model is built to detect partially manipulated as well as fully AI-generated videos. A key reason for the scarcity of synthetic video detection methods in the literature is the lack of standardized datasets for training and evaluation.", + "bbox": [ + 89, + 449, + 483, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 89, + 628, + 264, + 646 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section we first describe our problem setup (Sec. 3.1), then we describe the foundation model-based video encoding (Sec. 3.2), our trainable transformer architecture (Sec. 3.3) and finally the novel loss function for training the UNITE model (Sec. 3.4). The overview of the UNITE architecture is shown in Fig. 2.", + "bbox": [ + 89, + 655, + 483, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Setup", + "text_level": 1, + "bbox": [ + 89, + 756, + 241, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For each video in the dataset, we implement a frame sampling strategy where every other frame is extracted. From the resulting collection of frames, we construct video segment samples consisting of $n_f = 64$ consecutive frames, which are defined as a single data sample, denoted as \"v\". This selection aligns with the context window established for our video transformer model, ensuring that each sample retains the temporal coherence necessary for effective train-", + "bbox": [ + 89, + 780, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing. In instances where the video duration results in fewer than $n_f$ frames, we apply padding (ablation in supplementary) to ensure that each sample remains uniform in size.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given that the original video-level classification labels are known and fake videos contain manipulations throughout their entirety (i.e., no videos have manipulations limited to specific frames), each $n_f$ -frame segment generated from a given video is assigned the same label (denoted by $l \\in 0, \\dots, n_c$ ) as the source video, with number of possible classes as $n_c$ . The aggregation of these samples across all videos forms our final dataset, $v_i, l_{i=1}^N \\in \\mathcal{V}$ , where $N$ is the total number of samples obtained. This approach serves as a data augmentation technique, increasing the number of training samples without additional data collection.", + "bbox": [ + 511, + 137, + 906, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Domain-Agnostic Feature Extraction", + "text_level": 1, + "bbox": [ + 511, + 313, + 833, + 329 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our objective is to leverage the trained UNITE model for detecting in-the-wild DeepFakes, necessitating careful consideration of the domain gap [7, 34] between the train and test datasets. To effectively mitigate the impact of domain discrepancies, we utilize the SigLIP foundation model [59] to extract domain-agnostic features. Specifically, we employ the \"shape-optimized\" ViT [4] image encoder, whose architecture is rooted in [18]. The SigLIP-So400m [4] model was contrastively pretrained on 3B diverse examples using a sigmoid loss, enabling it to capture robust, generalizable features while maintaining a compact size (400M parameters). This extensive pretraining makes SigLIP particularly adept at extracting domain-invariant representations, crucial for handling the significant variability encountered in DeepFake datasets and real-world scenarios.", + "bbox": [ + 511, + 337, + 906, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For every frame $f_{j}$ (resized to $\\mathbb{R}^{384\\times 384\\times 3}$ ), of a video sample $v_{i}\\in \\mathcal{V}$ , we obtain an image encoding from the frozen SigLIP-So400m [4] model as $e_j = SigLIP(f_j)\\in \\mathbb{R}^{t_s\\times d_s}$ , where $j\\in \\{1,2,\\dots ,n_f\\}$ , since there are $n_f$ frames per video sample (as explained in Sec. 3.1), $t_s = 729$ represents the token length and $d_{s} = 1152$ represents the feature dimension. We concatenate these frame-level features for the video sample $v_{i}$ , while maintaining the order of the frames, to obtain a video segment encoding denoted by $\\xi_{i}\\in \\mathbb{R}^{n_{f}\\times t_{s}\\times d_{s}}$ , which are used as input to our trainable transformer architecture. This encoded video dataset can be denoted as $\\{\\xi_i,l_i\\}_{i = 1}^N\\in \\mathcal{V}_\\xi$ .", + "bbox": [ + 511, + 564, + 908, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Learnable Transformer Architecture", + "text_level": 1, + "bbox": [ + 511, + 756, + 833, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The UNITE architecture is a multi-head self-attention (MHSA) based transformer model designed specifically for video classification tasks. It leverages the strengths of Transformer networks to process sequences of video frame embeddings effectively, allowing for robust classification performance. The UNITE architecture allows for adjustable depth, signifying the number of encoder blocks stacked within the model. A deeper architecture can capture more", + "bbox": [ + 511, + 780, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28052", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "intricate features and dependencies at the cost of increased computational complexity. For our architecture the depth is set to 4. Ablation experiments with different transformer depths are provided in Sec. 4.4.", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 Encoder Block", + "text_level": 1, + "bbox": [ + 89, + 172, + 246, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each encoder block consists of the following components:", + "bbox": [ + 89, + 196, + 478, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-Head Self-Attention Layer: This layer allows the model to focus on different parts of the input sequence simultaneously. With 12 attention heads in our architecture, it captures diverse interactions and dependencies among frames. Each head computes attention scores independently, learning multiple representations of the input data. This enhances the model's ability to detect subtle temporal variations, which is crucial for video classification. The use of scaled dot-product attention further helps mitigate large gradients and stabilizes training.", + "bbox": [ + 89, + 213, + 483, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Layer Normalization and Residual Connections: To aid gradient flow during training, each sub-layer incorporates residual connections that add the input back to the output, mitigating vanishing gradient issues in deep networks. Layer normalization follows to stabilize activations and enhance convergence speed. Additionally, dropout is applied post-attention to reduce overfitting by randomly deactivating a portion of neurons during training.", + "bbox": [ + 89, + 364, + 483, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Feed-Forward Network (MLP): The second sub-layer is a point-wise feed-forward network with two dense layers separated by a GELU activation [23]. This MLP enables nonlinear feature transformations, enhancing the model's expressiveness. By projecting attention outputs into a higher-dimensional space before reducing them, it captures complex interactions beyond linear transformations.", + "bbox": [ + 89, + 486, + 483, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attention maps are derived from the outputs of the encoder blocks to be used for computing the Attention-Diversity loss (Sec. 3.4) and to illustrate how the model prioritizes different spatial regions in the frames during the classification process (Fig. 3). This interpretability is valuable for understanding the model's decision-making, particularly in complex video classification scenarios.", + "bbox": [ + 89, + 592, + 483, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2 Positional Encoding", + "text_level": 1, + "bbox": [ + 89, + 719, + 284, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We utilize a sine-cosine positional encoding scheme following the original transformer [51] to provide a unique position identifier for each input token. For every frame $f_{j}$ in the video sample with encoded feature dimension $d_{s}$ (as per Sec. 3.2), we compute the positional encoding for odd and even indexed feature dimensions as,", + "bbox": [ + 89, + 743, + 483, + 834 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP E _ {(j, 2 i + 1)} = \\cos \\left(\\frac {j}{1 0 0 0 0 ^ {\\frac {2 i}{d _ {s}}}}\\right), \\quad P E _ {(j, 2 i)} = \\sin \\left(\\frac {j}{1 0 0 0 0 ^ {\\frac {2 i}{d _ {s}}}}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 857, + 488, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where, $i$ denotes the feature index in the encoded feature dimension $d_{s}$ . We embed these encodings directly into the input tokens of our video transformer to provide positional context, for better temporal modeling during the attention computation with minimal computational overhead.", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Attention-Diversity Loss", + "text_level": 1, + "bbox": [ + 511, + 174, + 738, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training a multi-head attention transformer with only cross-entropy loss makes the attention maps focus only on the face regions of the frame (as evident in Fig. 3). However, we aim to detect fake videos where the manipulation might not be in the face at all, and instead be in the background (like video inpainting with the AVID model [61]). To ensure that the attention heads focus on diverse spatial regions of the video frames, we devise the \"Attention-Diversity\" (AD) loss.", + "bbox": [ + 511, + 198, + 905, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "AD-loss is designed to minimize overlap among attention maps (obtained from the first encoder block of the trainable transformer architecture in 3.3.1) while maintaining consistency across different input video samples. The attention outputs from the video transformer model $\\mathcal{A} \\in \\mathbb{R}^{n_h \\times t_s \\times d_s}$ are used to pool the input SigLIP-So400m features $\\xi \\in \\mathbb{R}^{n_f \\times t_s \\times d_s}$ , resulting in a pooled feature tensor $\\mathcal{P} \\in \\mathbb{R}^{n_h \\times n_f}$ according to,", + "bbox": [ + 511, + 319, + 906, + 439 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} = \\sum_ {j = 1} ^ {t _ {s}} \\sum_ {k = 1} ^ {d _ {s}} \\mathcal {A} _ {h, j, k} \\cdot \\xi_ {f, j, k}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 450, + 905, + 493 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f$ indexes the frames $(n_f)$ , $h$ indexes the number of attention heads $(n_h)$ , and $j, k$ are spatial positions.", + "bbox": [ + 511, + 497, + 905, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We compute \"feature centers\" which are points in the feature space that serve as anchors for the learned representations of different classes. These centers represent the average features of samples from a particular class, allowing the model to capture essential characteristics of that class. In the context of the AD-loss, feature centers $\\mathcal{C} \\in \\mathbb{R}^{n_h \\times n_f}$ are dynamically updated in each training iteration $\\tau$ based on the pooled feature vectors $\\mathcal{P}$ , following,", + "bbox": [ + 511, + 527, + 905, + 648 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} ^ {\\tau} = \\mathcal {C} ^ {\\tau - 1} - \\eta \\left(\\mathcal {C} ^ {\\tau - 1} - \\frac {1}{B} \\sum_ {b = 1} ^ {B} \\mathcal {P} _ {b}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 659, + 905, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\eta$ represents the learning rate (set to 0.05 in our experiments) for feature center updates and $B$ is the batch dimension. The feature centers are initialized with zeros $(\\mathcal{C}^0 = [0]_{n_h\\times n_f})$ at the beginning of the model training.", + "bbox": [ + 511, + 704, + 905, + 765 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "AD-loss consists of two distinct components: within-class loss and between-class loss inspired by the Fisher Discriminant Analysis for deep networks [22]. The within-class term calculates the distance between the pooled feature vectors $\\mathcal{P}$ and their corresponding feature centers $\\mathcal{C}$ , promoting closeness among similar samples. The between-class term measures the distance between feature centers of different classes, encouraging them to be spaced apart in the feature space, encouraging separability.", + "bbox": [ + 511, + 766, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28053", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The within-class loss term is calculated as", + "bbox": [ + 109, + 90, + 392, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {w i t h i n}} = \\max \\left(\\| \\mathcal {P} - \\mathcal {C} \\| _ {2} - \\delta_ {\\text {w i t h i n}}, 0\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 119, + 480, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\delta_{\\mathrm{within}} \\in \\mathbb{R}^{n_c}$ is a hyperparameter controlling the allowable distance between feature vectors and their respective feature centers for inputs of the same class, encouraging tighter clustering and $\\| \\cdot \\|_2$ represents $L_2$ normalization.", + "bbox": [ + 89, + 140, + 482, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The between-class loss is calculated as", + "bbox": [ + 109, + 200, + 369, + 215 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {b e t w e e n}} = \\sum_ {\\substack {k \\neq l \\\\ (k, l) \\in \\left(n _ {h}, n _ {h}\\right)}} \\max \\left(\\delta_ {\\text {b e t w e e n}} - \\| \\mathcal {C} _ {k} - \\mathcal {C} _ {l} \\| _ {2}, 0\\right), \\tag{5}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 234, + 480, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\delta_{\\mathrm{between}}$ is a predefined hyperparameter that ensures a minimum distance between feature centers of different classes. Thus the AD-loss is a simple addition of these two components $\\mathcal{L}_{AD} = \\mathcal{L}_{\\mathrm{within}} + \\mathcal{L}_{\\mathrm{between}}$ , making the final objective function for training the UNITE model,", + "bbox": [ + 89, + 291, + 483, + 367 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {U N I T E}} = \\lambda_ {1} \\cdot \\mathcal {L} _ {C E} + \\lambda_ {2} \\cdot \\mathcal {L} _ {A D}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 381, + 482, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{CE}$ is the traditional cross-entropy loss, and $\\lambda_{1}$ and $\\lambda_{2}$ are loss-weighting hyperparameters.", + "bbox": [ + 89, + 400, + 483, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 441, + 223, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 89, + 465, + 192, + 481 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Most DeepFake datasets focus on face manipulations, such as face-swaps or lip-syncing, and lack fully AI-generated video content. The DeMamba dataset [9] features videos from T2V/I2V models, including powerful diffusion models like SORA [5]. It is not a human-centric dataset and primarily contains videos from natural scenes. We use its validation split to evaluate UNITE.", + "bbox": [ + 89, + 489, + 483, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To train UNITE, we employ the FaceForensics++ (FF++) [43] dataset (c23 i.e., medium-quality) and the SAIL-VOS-3D dataset [24], which includes synthetic (although not AI-generated) GTA-V game videos featuring human subjects suitable for DeepFake detection. To evaluate the UNITE model, we use:", + "bbox": [ + 89, + 595, + 483, + 686 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Face manipulated data: $\\mathrm{FF}++$ [43] (in-domain evaluation), CelebDF [30], DeeperForensics [25], Deepfake-TIMIT [27], HifiFace [53], UADFV [58].", + "- Background manipulated data: Sample videos from the AVID [61] model provided publicly by the authors.", + "- Fully synthetic data: GTA-V [24] (in-domain evaluation), DeMamba [9].", + "- In-the-wild DeepFakes: Publicly available videos from the New York Times DeepFake quiz [48]." + ], + "bbox": [ + 89, + 686, + 482, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Results and Discussion", + "text_level": 1, + "bbox": [ + 89, + 832, + 303, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics: Although most DeepFake detection methods only report accuracy, we evaluate UNITE using additional metrics: (1) AUC (Area Under the Precision-Recall", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Curve); (2) Precision@0.5 and (3) Recall@0.5 at a 0.5 confidence threshold; (4) Precs@Rec=0.8 (precision at 0.8 recall); (5) Rec@Precs=0.8 (recall at 0.8 precision).", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Details: UNITE is trained using an AdamW optimizer [33] with an initial learning rate of 0.0001 and a decay rate of 0.5 every 1000 steps. For AD-loss, $\\delta_{\\mathrm{between}} = 0.5$ and $\\delta_{\\mathrm{within}}$ is set to $[0.01, -2]$ for binary and $[0.01, -2, 1]$ for fine-grained classification in Sec. 4.3 (see supplementary for sensitivity). Loss weights are $\\lambda_1 = \\lambda_2 = 0.5$ . Training uses a batch size of 32 for 25 epochs on 8 TPUv3 chips, with the framework implemented in TensorFlow.", + "bbox": [ + 511, + 137, + 905, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative Results: The results obtained by the UNITE model when trained on $\\mathrm{FF}++$ [43] alone, and when trained with both $\\mathrm{FF}++$ [43] and GTA-V [24] data and evaluated on the various datasets mentioned in Sec. 4.1 are shown in Table 1. On the AVID [61] and DeMamba datasets [9], the model performs poorly when trained with only face manipulated data, but the performance enhances several-fold when the GTA-V [24] synthetic data, even though it is not AI-generated, is used in training.", + "bbox": [ + 511, + 261, + 906, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Interestingly, when evaluating on face-manipulated datasets such as CelebDF [30] and DeeperForensics [25], we observe a performance boost, particularly for CelebDF [30], when GTA-V [24] data is included in the training set. Intuitively, one would expect that training solely on $\\mathrm{FF}++$ [43] would yield similar results since the inference set consists of similar face-manipulated data. This deviation from the expected behavior comes from the contribution of the AD-Loss, as the ablation study (Fig. 4) using only CE-loss compared to both CE and AD-losses reveals that the performance is enhanced with GTA-V [24] data in training only in the latter (detailed discussion in Sec. 4.4).", + "bbox": [ + 511, + 398, + 905, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate UNITE on in-the-wild DeepFakes, we attempted the recent New York Times quiz [48] (denoted by NYTimes) which provides 10 videos (4 real and 6 fake) for testing DeepFake detectors. UNITE got 8 of those 10 videos correct, even though some of the face-swap videos were indistinguishable from real videos, even by humans. On all the AI-generated videos, UNITE made the correct prediction (when trained on $\\mathrm{FF} + +$ [43] and GTA-V [24]).", + "bbox": [ + 511, + 582, + 905, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Attention Heatmaps: To analyze the impact of CE and AD-losses on attention distribution, we extracted attention features from the first encoder of the UNITE model, comparing models trained with CE loss alone versus that trained with both CE and AD-losses. The resulting heatmaps, shown in Fig. 3, reveal that the model trained with only CE loss tends to concentrate primarily on the face region. In contrast, the model trained with the combined $\\mathrm{CE + AD}$ loss demonstrates a broader attention span across the frame, as evidenced by a lighter, more distributed bluish tone in the heatmaps, indicating increased spatial diversity in the model's focus. This is why on the example shown in the first row (face-manipulated video) both models give the correct", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28054", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/42739651164fe353323c6279da5dce08e66572740d45c8b707abe18c536a5005.jpg", + "table_caption": [ + "Table 1. Results from the UNITE model trained with (1) FF++ [43] only and (2) FF++ [43] combined with GTA-V [24]. All other results reflect cross-dataset evaluations except for FF++ and GTA-V (when trained). Performance gains are highlighted in green." + ], + "table_footnote": [], + "table_body": "
TrainTest
FF++GTA-VDatasetAccuracyAUCPrecision@0.5Recall@0.5Precs@Rec=0.8Rec@Precs=0.8
Face Manipulated Data
✓FF++99.53%99.77%99.94%99.49%99.94%99.94%
✓CelebDF72.61%94.05%96.45%61.22%80.45%61.22%
✓DeeperForensics91.35%100.00%100.00%91.35%100.00%91.35%
✓DeepfakeTIMIT86.90%86.46%83.61%83.97%88.90%81.33%
✓HiFiFace63.63%62.47%67.12%63.63%59.30%63.63%
✓UADFV94.12%94.38%95.68%97.11%93.79%94.38%
✓✓FF++99.96%(+0.43)99.89%(+0.12)100.00%(+0.06)99.84%(+0.35)100.00%(+0.06)99.96%(+0.02)
✓✓CelebDF95.11%(+22.50)94.36%(+0.31)96.82%(+0.37)68.75%(+7.53)96.53%(+16.08)68.75%(+7.53)
✓✓DeeperForensics99.62%(+8.27)100.00%(+0.00)100.00%(+0.00)99.62%(+8.27)100.00%(+0.00)99.63%(+8.28)
✓✓DeepfakeTIMIT91.90%(+5.00)91.33%(+4.87)90.45%(+6.84)88.39%(+4.42)100.00%(+11.10)91.95%(+10.62)
✓✓HiFiFace75.62%(+11.99)81.24%(+18.77)79.55%(+12.43)71.71%(+8.08)75.62%(+16.32)72.47%(+8.84)
✓✓UADFV97.01%(+2.89)94.95%(+0.57)96.89%(+1.21)100.00%(+2.89)94.12%(+0.33)100.00%(+5.62)
Background Manipulated Data
✓AVID41.67%33.33%33.33%41.67%41.67%33.33%
✓✓AVID100.00%(+58.33)100.00%(+66.67)100.00%(+66.67)100.00%(+58.33)100.00%(+58.33)100.00%(+66.67)
Fully Synthetic Data
✓GTA-V60.16%61.52%60.16%58.73%63.29%58.73%
✓DeMamba61.47%57.38%67.73%33.01%62.15%54.16%
✓✓GTA-V100.00%(+39.84)100.00%(+38.48)100.00%(+39.84)100.00%(+41.27)100.00%(+36.71)100.00%(+41.27)
✓✓DeMamba87.12%(+25.65)93.75%(+36.67)92.76%(+25.03)89.60%(+56.59)89.81%(+27.66)92.12%(+37.96)
In-the-wild DeepFakes
✓NYTimes [48]50.00%53.74%50.00%25.00%0.00%0.00%
✓✓NYTimes [48]80.00%(+30.00)97.42%(+43.68)83.33%(+33.33)83.33%(+58.33)60.00%(+60.00)100.00%(+100.00)
", + "bbox": [ + 171, + 117, + 821, + 358 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/026a53759fa5e9c90ff0e7f220988e58d42c475cf1f306716a2082362aedd22d.jpg", + "table_caption": [ + "Table 2. SOTA Comparison on Face-Manipulated Data: We compared the performance of UNITE with recent DeepFake detectors, in terms of detection accuracy on various face manipulated datasets. UNITE outperforms the existing methods. Bold shows the current best results and the previous best and second-best results are highlighted in red and blue respectively." + ], + "table_footnote": [], + "table_body": "
MethodFF++CelebDFDeeperForensicsUADFV
TALL [57]98.65%90.79%99.62%-
ISTVT [62]99.00%84.10%98.60%-
Concas et al. [13]99.49%---
PUDD [40]-95.10%--
TI2Net [32]99.95%68.22%76.08%-
LRNet [46]99.89%53.20%56.77%-
Choi et al. [11]-89.00%99.00%-
Lin et al. [31]98.28%74.42%--
Guo et al. [21]99.24%84.97%--
Li et al. (Res-152) [29]---93.80%
HeadPose [58]---89.00%
CViT [56]93.00%--93.75%
FakeCatcher [12]94.65%91.50%--
MesoNet [3]---82.40%
UNITE (Ours)99.96%95.11%99.62%97.01%
", + "bbox": [ + 109, + 449, + 457, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "prediction (\"fake\") with high confidence, but on the row-2 example (video generated from the Sora [6] T2V model) the model trained with only CE loss predicted \"real\" with a confidence score of $99.25\\%$ (which is wrong), but the model trained with $\\mathrm{CE + AD}$ losses predicted \"fake\" with a confidence of $100.00\\%$ , which is correct.", + "bbox": [ + 89, + 611, + 483, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison on Face Manipulated Data: UNITE is compared with recent state-of-the-art (SOTA) DeepFake detection models in Table 2. The methods we compared against were specifically designed for detecting face manipulations and often crop faces from the videos, during both training and inference of the models. However, UNITE, which is designed to detect a diverse range of fake videos, still outperforms these detectors.", + "bbox": [ + 89, + 703, + 483, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison on Synthetic Data: We evaluated UNITE against state-of-the-art synthetic video detectors on the DeMamba dataset [9] (validation split), with results in Table 3. While existing detectors were trained on the DeMamba train split and validated on the DeMamba validation split,", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c22e4507b864d24b58be75ce23f856e165498a1d17c85f7886055912690090c3.jpg", + "image_caption": [ + "(a) Original Frame" + ], + "image_footnote": [], + "bbox": [ + 526, + 359, + 647, + 411 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9a7f55b96974cf7b358b430d9ef5d44c715dd9f04ae58d9d55339da946fc69ed.jpg", + "image_caption": [ + "(b) CE Loss only" + ], + "image_footnote": [], + "bbox": [ + 650, + 359, + 769, + 411 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a73dc7545cdac92a63c283887cef990df632ac6489198467c2b382e0b2d6c15f.jpg", + "image_caption": [ + "(c) $\\mathrm{CE + AD}$ Loss" + ], + "image_footnote": [], + "bbox": [ + 772, + 359, + 893, + 411 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/66b04e8a8f21be849327484594cfd6b4ecf64172824ab28252ce8e38960c5b84.jpg", + "image_caption": [ + "(d) Original Frame" + ], + "image_footnote": [], + "bbox": [ + 527, + 424, + 647, + 477 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c6f70e908b9eb1083ccd7c19c4d8aafcce441fe1ac95d4a3093f9532bc7b1cf2.jpg", + "image_caption": [ + "(e) CE Loss only" + ], + "image_footnote": [], + "bbox": [ + 650, + 424, + 769, + 477 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f78bfdc6431dce203fc1738fdbc24287c4c17126ed02ee3507b3c365f47ad81c.jpg", + "image_caption": [ + "(f) $\\mathrm{CE + AD}$ Loss", + "Figure 3. Comparison of attention heatmaps when UNITE is trained with only CE loss (second column) versus CE and AD-losses (third column). The CE-loss heatmap predominantly focuses on the face region, whereas the $\\mathrm{CE + AD}$ loss heatmap demonstrates more distributed attention across the entire frame, as indicated by a broader bluish tone in the heatmaps. Both samples are \"fake\" videos- (a) is taken from Celeb-DF [30] and (d) is taken from DeMamba [9] and generated by OpenAI's Sora [6] (Best viewed as GIFs provided in the supplementary material). Yellow to blue shades show increasing attention in the plot." + ], + "image_footnote": [], + "bbox": [ + 772, + 424, + 893, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "UNITE was trained on FF++ [43] and GTA-V [24]. Despite not being trained on DeMamba, UNITE's average performance (including real videos) surpassed current SOTA detectors, with competitive generator-wise results.", + "bbox": [ + 511, + 638, + 906, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Thus, UNITE reliably detects a diverse range of fake videos (even in cross-domain settings), including T2V/I2V videos and background manipulations, while consistently maintaining state-of-the-art performance on traditional face-manipulated data. UNITE eliminates the need to have separate DeepFake and T2V/I2V video detector models by by handling both tasks within a single trained model.", + "bbox": [ + 511, + 700, + 906, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Finegrained Evaluation", + "text_level": 1, + "bbox": [ + 511, + 816, + 732, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In fake media detection, flagging videos partially or fully manipulated by AI also necessitates fine-grained classification into three classes: (1) real videos, (2) partially manipulated videos, and (3) fully synthetic videos. While binary", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28055", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/285d59e5e6e07288f95b3fdfee0aaabba479d0a223322ee43e05064bdcead95c.jpg", + "table_caption": [ + "Table 3. SOTA Comparison on Synthetic Data: On the DeMamba dataset (validation split), we compare the performance of UNITE, which was NOT trained on DeMamba train split, against state-of-the-art detectors which were trained on DeMamba train split (results taken from Chen et al. [9]). We report the results ( $P = \\text{Precision}$ and $R = \\text{Recall}$ ) on the individual T2V/I2V generators and the average performance across the entire validation set ( $Avg$ , which also includes real videos). Although the direct comparison is unfair against UNITE which was trained with FF++ [43] and GTA-V [24], our method still outperforms these synthetic video detectors. Bold shows the current best results and the previous best and second-best results are highlighted in red and blue respectively." + ], + "table_footnote": [], + "table_body": "
MethodMetricsSora [6]Morph Studio [2]Runway ML (Gen2) [42]HotShot [38]Lavie [54]Show-1 [60]Moon Valley [37]Crafter [8]Model Scope [52]Wild Scrape [9]Avg
TALL [57]P71.15%96.89%98.51%79.38%84.59%79.38%98.79%99.02%92.70%76.47%87.91%
R91.07%98.28%97.83%83.00%76.57%79.57%99.52%98.93%94.14%66.31%88.52%
F3Net [41]P68.27%99.89%99.67%89.35%57.00%36.57%99.52%99.71%93.80%88.41%88.73%
R83.93%99.71%98.62%77.57%85.24%63.17%99.58%99.89%89.43%76.78%81.88%
NPR [47]P91.07%99.57%99.49%24.29%89.64%57.71%97.12%99.86%94.29%87.80%82.45%
R91.07%99.57%99.49%24.29%89.64%57.71%97.12%99.86%94.29%87.80%84.08%
STIL [20]P57.21%99.08%99.32%86.19%82.24%70.43%99.25%98.96%97.18%81.32%87.12%
R67.86%96.00%98.41%96.14%77.14%80.43%97.44%96.93%96.29%68.36%82.22%
MINTIME-CLIP-B [9]P83.21%99.99%99.67%50.84%99.20%99.27%99.76%99.99%91.83%91.77%91.55%
R89.29%100.00%98.99%26.43%96.79%98.14%99.84%100.00%84.29%82.38%87.62%
FTCN-CLIP-B [9]P91.79%99.99%99.79%45.94%99.76%97.80%99.99%99.99%94.69%92.32%92.21%
R87.50%100.00%98.91%17.71%97.71%91.86%100.00%100.00%85.29%82.83%86.18%
CLIP-B-PT [9]P67.80%43.56%70.88%29.97%52.97%35.36%55.52%66.03%44.23%42.99%44.83%
R85.71%82.43%90.36%71.00%79.29%75.43%89.62%86.29%82.14%75.16%81.74%
DeMamba-CLIP-PT [9]P25.87%95.14%96.23%73.43%83.31%75.49%90.17%95.06%95.05%69.95%79.97%
R58.93%96.43%93.12%68.00%69.36%69.00%89.14%91.86%96.14%56.59%78.86%
XCLIP-B-PT [9]P16.39%72.16%87.77%39.86%65.57%54.26%75.23%84.80%61.60%55.28%61.29%
R81.34%82.15%83.35%80.98%81.82%81.55%82.14%82.98%81.93%81.10%81.93%
DeMamba-XCLIP-PT [9]P18.26%93.50%94.72%69.94%78.08%71.50%83.95%92.23%93.54%68.10%76.38%
R66.07%95.86%94.64%77.86%75.36%80.29%90.89%92.50%96.00%66.41%83.59%
XCLIP-B-FT [9]P64.42%99.73%96.78%70.98%90.35%77.28%97.34%99.84%82.01%88.97%86.77%
R82.14%99.57%93.62%61.29%79.36%69.71%97.92%99.79%77.14%83.59%84.41%
UNITE (Ours)P88.57%100.00%100.00%90.16%89.91%98.34%99.52%100.00%98.96%92.56%92.76%
R92.11%100.00%94.62%96.93%98.12%99.86%98.69%100.00%96.29%89.89%89.60%
", + "bbox": [ + 109, + 172, + 879, + 411 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/06feaeed88e23c1a4a5632e9ccdd54161f908fe2f012c75ed66e2b745f5ebb78.jpg", + "table_caption": [ + "Table 4. Results obtained by the UNITE model on finegrained classes: detecting whether a video is real, partially manipulated or fully AI-generated. Performance gains are mentioned in green." + ], + "table_footnote": [], + "table_body": "
TrainTest
FF++GTA-VDatasetAccuracy
✓FF++96.46%
✓CelebDF60.40%
✓DeeperForensics71.04%
✓DeepFakeTIMIT80.68%
✓HifiFace43.27%
✓UADFV90.37%
✓AVID50.00%
✓GTA-V0.00%
✓DeMamba38.29%
✓✓FF++97.70%(+1.24)
✓✓CelebDF70.17%(+9.77)
✓✓DeeperForensics80.59%(+9.55)
✓✓DeepFakeTIMIT81.04%(+0.36)
✓✓HifiFace64.20%(+20.93)
✓✓UADFV92.61%(+1.24)
✓✓AVID62.50%(+12.50)
✓✓GTA-V100.00%(+100.00)
✓✓DeMamba65.84%(+27.55)
", + "bbox": [ + 148, + 463, + 418, + 665 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "real vs. fake classification may suffice to curb misinformation, fine-grained classification offers greater explainability in otherwise black-box models.", + "bbox": [ + 89, + 671, + 482, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Using this defined convention of finegrained classes, all the face manipulated datasets (CelebDF [30], DeeperForensics [25], etc.) and the videos from the AVID [61] model (background manipulations) fall under class-2, and the videos from the GTA-V [24] and the DeMamba [9] datasets fall in class-3. In this setup, only the final classification layer of the UNITE transformer architecture is changed to 3 neurons with softmax activation. The results obtained are presented in Table 4. As expected, when the model is trained only with $\\mathrm{FF}++$ [43], the performance on fully synthetic data is poor (since the model has not seen any synthetic videos). When the GTA-V [24] data is also", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4e0ee692c4c5de2ef0684ec2901942f259d3d5e21d2f9aed9f0a466693c70f68.jpg", + "image_caption": [ + "(a) UNITE trained on $\\mathrm{FF} + +$ [43] only" + ], + "image_footnote": [], + "bbox": [ + 540, + 420, + 880, + 508 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/731f76544b4eb7ad0882ec6ef0bec3dce866c8aa404905fbfc7f87320a267252.jpg", + "image_caption": [ + "(b) UNITE trained on $\\mathrm{FF} + +$ [43] and GTA-V [24]", + "Figure 4. Ablation Results to show the effect of changing the loss functions used to train the UNITE model. The combination of the cross-entropy (CE) and Attention Diversity (AD) losses always performs better, and the contribution from the AD-loss component increases significantly when using fully synthetic data for training." + ], + "image_footnote": [], + "bbox": [ + 540, + 527, + 879, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "used in training, the performance is much higher.", + "bbox": [ + 511, + 710, + 836, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 739, + 666, + 756 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted an analysis to assess the contribution of individual loss functions to the training of the UNITE model. The outcomes of this ablation study are presented in Fig. 4. Specifically, we evaluated the impact of the AD-loss in two scenarios: (1) when UNITE was trained solely on the FF++ dataset [43] (Fig. 4(a)), and (2) when UNITE was trained using a combination of the FF++ [43] and the GTA-V [24] datasets (Fig. 4(b)). The combination of the two losses always performed better than the individual losses.", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28056", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e2581f9420b0e9c89632475fbb1f99e00a2b12ff8998a84ceffb81c5cc3e3920.jpg", + "image_caption": [ + "Figure 5. No. of frames vs. Performance: Performance analysis of UNITE based on the number of frames sampled per video segment. The results illustrate that as the number of frames increases from 1 to 64 (context window), the detection accuracy improves, showcasing UNITE's ability to effectively capture temporal inconsistencies in fake videos." + ], + "image_footnote": [], + "bbox": [ + 140, + 93, + 436, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect on Synthetic Data: UNITE performed significantly better on the fully synthetic datasets GTA-V [24] and DeMamba [9] when the model was training includes GTA-V [24] data. However, for the DeMamba dataset (cross-data evaluation), the performance with only CE-loss is similar regardless of the training data. It is the AD-loss component that provides the accuracy boost in Fig 4(b).", + "bbox": [ + 89, + 323, + 482, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect on Background-Manipulated Data: In Fig. 4(a), we observe that when UNITE was trained solely on $\\mathrm{FF}++$ [43] using only the CE-loss, its performance on the AVID dataset [61] was $0.00\\%$ . This outcome can be attributed to the fact that the fake data in $\\mathrm{FF}++$ [43] contains exclusively face-manipulated videos, causing the UNITE transformer's attention heads to focus predominantly on the facial regions. As AVID [61] comprises only background-manipulated videos, UNITE failed to detect any of the fakes under such conditions. However, when the AD-loss was introduced, the attention heads of UNITE were encouraged to diversify their spatial focus, significantly improving performance. Incorporating synthetic data, as shown in Fig. 4(b), further enhanced this effect, achieving $100\\%$ accuracy on AVID when using both CE-loss and AD-loss.", + "bbox": [ + 89, + 430, + 482, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect on Face-Manipulated Data: For the evaluation on face-manipulated datasets such as CelebDF [30] and UADFV [58], one might expect comparable results between training solely on $\\mathrm{FF} + +$ [43] and incorporating GTA-V [24] due to their perceived irrelevance in evaluating face-centric manipulations. However, as shown in Fig. 4(b), training with fully synthetic data leads to a substantial performance improvement. Notably, in both Fig. 4(a) and (b), using only CE loss yields similar results, but the inclusion of the AD-loss component significantly enhances overall detection accuracy. The diversity in the training data makes the UNITE model learn more discriminative attention map features through the AD-loss component, which in turn helps even in detecting face-manipulated videos.", + "bbox": [ + 89, + 657, + 482, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Frames vs. Performance: For inference with UNITE, we evaluated the impact of temporal context by vary-", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/69c7bd286988af9c27ba996e301c20fc5c1e44c982413c17694203378e6c34a8.jpg", + "image_caption": [ + "Figure 6. Transformer Depth Evaluation: Performance comparison of UNITE with varying the number of encoder blocks (depth). UNITE performs optimally in cross-domain settings when the depth is 4. Greater depths overfit to the training domains $(\\mathrm{FF} + +$ [43] and GTA-V [24]), while a depth of 2 is insufficient to capture the complexity of the data." + ], + "image_footnote": [], + "bbox": [ + 563, + 94, + 859, + 231 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing the number of frames sampled from video segments to $\\{1,8,16,32,64\\}$ , with 64 frames corresponding to our model's context window. As shown in Fig. 5, performance improves as more frames are included, demonstrating that UNITE effectively captures temporal discontinuities in fake videos. This highlights the model's ability to leverage temporal cues for enhanced detection accuracy.", + "bbox": [ + 511, + 329, + 906, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Transformer Depth Ablation: We evaluate the impact of varying the depth of the transformer architecture in UNITE by adjusting the number of encoder blocks, with results shown in Fig. 6. We tested depths of $\\{2,4,6,8\\}$ and observed that the model achieved the best performance with 4 encoder blocks, which is the depth used in our final architecture. This suggests that a moderate depth strikes the optimal balance between model complexity and performance, especially in cross-dataset settings, providing robust detection capabilities without overfitting.", + "bbox": [ + 511, + 435, + 908, + 587 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 601, + 633, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Traditional DeepFake detectors, focused solely on face-manipulated content, struggle against newer, more sophisticated forgery methods involving T2V, I2V, and background manipulations. To address this, we proposed UNITE, leveraging the fully synthetic GTA-V [24] dataset for training. While this synthetic data is not AI-generated, our results show it effectively enhances UNITE's ability to detect AI-generated content and background manipulations.", + "bbox": [ + 511, + 627, + 905, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A key innovation in our approach is the AD-loss, which encourages the model to explore diverse spatial regions, enhancing its detection of manipulated areas beyond just the face region. This allows UNITE to excel not only in detecting AI-generated and background-altered videos but also in identifying traditional face-manipulated content. Our ablation studies demonstrate the robustness of UNITE, making it a valuable tool in synthetic video detection.", + "bbox": [ + 511, + 750, + 906, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: This work was supported in part by funding from YouTube (Google).", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28057", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Flux 1.1. https://blackforestlabs.ai/, 2024.1", + "[2] Morph studio. https://www.morphstudio.com/, 2024.7", + "[3] Darius Afchar, Vincent Nozick, Junichi Yamagishi, and Isao Echizen. Mesonet: a compact facial video forgery detection network. In 2018 IEEE international workshop on information forensics and security (WIFS), pages 1-7. IEEE, 2018. 6", + "[4] Ibrahim M Alabdulmohsin, Xiaohua Zhai, Alexander Kolesnikov, and Lucas Beyer. Getting vit in shape: Scaling laws for compute-optimal model design. Advances in Neural Information Processing Systems, 36, 2024. 2, 3", + "[5] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 5", + "[6] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. 2024. URL https://openai.com/research/video-generation-models-as-world-simulators, 3, 2024. 6, 7", + "[7] Baoying Chen and Shunquan Tan. Featuretransfer: Unsupervised domain adaptation for cross-domain deepfake detection. Security and Communication Networks, 2021(1): 9942754, 2021. 3", + "[8] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 7", + "[9] Haoxing Chen, Yan Hong, Zizheng Huang, Zhuoer Xu, Zhangxuan Gu, Yaohui Li, Jun Lan, Huijia Zhu, Jianfu Zhang, Weiqiang Wang, et al. Demamba: Ai-generated video detection on million-scale genvideo benchmark. arXiv preprint arXiv:2405.19707, 2024. 1, 2, 5, 6, 7, 8", + "[10] Jikang Cheng, Zhiyuan Yan, Ying Zhang, Yuhao Luo, Zhongyuan Wang, and Chen Li. Can we leave deepfake data behind in training deepfake detector? arXiv preprint arXiv:2408.17052, 2024. 1", + "[11] Jongwook Choi, Taehoon Kim, Yonghyun Jeong, Seungryul Baek, and Jongwon Choi. Exploiting style latent flows for generalizing deepfake video detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1133-1143, 2024. 2, 6", + "[12] Umur Aybars Ciftci, Ilke Demir, and Lijun Yin. Fakecatcher: Detection of synthetic portrait videos using biological signals. IEEE transactions on pattern analysis and machine intelligence, 2020. 6", + "[13] Sara Concas, Simone Maurizio La Cava, Roberto Casula, Giulia Orrù, Giovanni Puglisi, and Gian Luca Marcialis. Quality-based artifact modeling for facial deepfake detection in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3845–3854, 2024. 2, 6" + ], + "bbox": [ + 93, + 114, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Riccardo Corvi, Davide Cozzolino, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. Intriguing properties of synthetic images: from generative adversarial networks to diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 973-982, 2023. 2", + "[15] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 2", + "[16] Davide Cozzolino, Andreas Rössler, Justus Thies, Matthias Nießner, and Luisa Verdoliva. Id-reveal: Identity-aware deepfake video detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15108-15117, 2021. 2", + "[17] Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. The deepfake detection challenge (dfdc) dataset. arXiv preprint arXiv:2006.07397, 2020. 1", + "[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3", + "[19] Ricard Durall, Margret Keuper, Franz-Josef Pfundt, and Janis Keuper. Unmasking deepfakes with simple features. arXiv preprint arXiv:1911.00686, 2019. 2", + "[20] Zhihao Gu, Yang Chen, Taiping Yao, Shouhong Ding, Jilin Li, Feiyue Huang, and Lizhuang Ma. Spatiotemporal inconsistency learning for deepfake video detection. In Proceedings of the 29th ACM international conference on multimedia, pages 3473-3481, 2021. 7", + "[21] Ying Guo, Cheng Zhen, and Pengfei Yan. Controllable guide-space for generalizable face forgery detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20818-20827, 2023. 6", + "[22] Harald Hanselmann, Shen Yan, and Hermann Ney. Deep fisher faces. In BMVC, 2017. 4", + "[23] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 4", + "[24] Yuan-Ting Hu, Jiahong Wang, Raymond A Yeh, and Alexander G Schwing. Sail-vos 3d: A synthetic dataset and baselines for object detection and 3d mesh reconstruction from video data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1418-1428, 2021. 1, 2, 5, 6, 7, 8", + "[25] Liming Jiang, Ren Li, Wayne Wu, Chen Qian, and Chen Change Loy. Deeperforensics-1.0: A large-scale dataset for real-world face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2889–2898, 2020. 5, 7", + "[26] Yan Ju, Chengzhe Sun, Shan Jia, Shuwei Hou, Zhaofeng Si, Soumyya Kanti Datta, Lipeng Ke, Riky Zhou, Anita Nikolich, and Siwei Lyu. Deepfake-o-meter v2.0: An open platform for deepfake detection. In 2024 IEEE 7th Inter-" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28058", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "national Conference on Multimedia Information Processing and Retrieval (MIPR), pages 439-445. IEEE, 2024. 1", + "[27] Pavel Korshunov and Sébastien Marcel. Deepfakes: a new threat to face recognition? assessment and detection. arXiv preprint arXiv:1812.08685, 2018. 5", + "[28] Lingzhi Li, Jianmin Bao, Hao Yang, Dong Chen, and Fang Wen. Faceshifter: Towards high fidelity and occlusion aware face swapping. arXiv preprint arXiv:1912.13457, 2019. 1", + "[29] Y Li. Exposing deepfake videos by detecting face warping artifact acts. arXiv preprint arXiv:1811.00656, 2018. 6", + "[30] Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3207-3216, 2020. 1, 3, 5, 6, 7, 8", + "[31] Li Lin, Xinan He, Yan Ju, Xin Wang, Feng Ding, and Shu Hu. Preserving fairness generalization in deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16815-16825, 2024. 6", + "[32] Baoping Liu, Bo Liu, Ming Ding, Tianqing Zhu, and Xin Yu. Ti2net: temporal identity inconsistency network for deepfake detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4691-4700, 2023. 2, 6", + "[33] I Loshchilov and F Hutter. Decoupled weight decay regularization. *ICLR*, 2019. 5", + "[34] Qingxuan Lv, Yuezun Li, Junyu Dong, Sheng Chen, Hui Yu, Huiyu Zhou, and Shu Zhang. Domainforensics: Exposing face forgery across domains via bi-directional adaptation. IEEE Transactions on Information Forensics and Security, 2024. 3", + "[35] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 2", + "[36] Ghazal Mazaheri and Amit K Roy-Chowdhury. Detection and localization of facial expression manipulations. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1035-1045, 2022. 2", + "[37] moonvalley.ai. moonvalley.ai. https://moonvalley.ai/, 2022.7", + "[38] John Mullan, Duncan Crawbuck, and Aakash Sastry. Hotshot-XL, 2023. 7", + "[39] Dat Nguyen, Nesryne Mejri, Inder Pal Singh, Polina Kuleshova, Marcella Astrid, Anis Kacem, Enjie Ghorbel, and Djamila Aouada. Laa-net: Localized artifact attention network for quality-agnostic and generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17395–17405, 2024. 2", + "[40] Alvaro Lopez Pellicer, Yi Li, and Plamen Angelov. Pudd: Towards robust multi-modal prototype-based deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3809-3817, 2024. 2, 6" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In European conference on computer vision, pages 86-103. Springer, 2020. 7", + "[42] Runway Research. Text driven video generation. https://research.runwayml.com/gen2, 2023.1,7", + "[43] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics++: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1-11, 2019. 1, 2, 3, 5, 6, 7, 8", + "[44] Md Shamim Seraj, Ankita Singh, and Shayok Chakraborty. Semi-supervised deep domain adaptation for deepfake detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1061-1071, 2024. 2", + "[45] Ke Sun, Shen Chen, Taiping Yao, Hong Liu, Xiaoshuai Sun, Shouhong Ding, and Rongrong Ji. Diffusionfake: Enhancing generalization in deepfake detection via guided stable diffusion. NeurIPS, 2024. 1", + "[46] Zekun Sun, Yujie Han, Zeyu Hua, Na Ruan, and Weijia Jia. Improving the efficiency and robustness of deepfakes detection through precise geometric features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3609-3618, 2021. 6", + "[47] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28130-28139, 2024. 7", + "[48] The New York Times. A.i. can now create lifelike videos. can you tell what's real? https://www.nytimes.com/interactive/2024/09/09/technology/ai-video-deepfake-runway-kling-quiz.html, 2024.5,6", + "[49] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG), 38(4):1-12, 2019. 1", + "[50] Loc Trinh, Michael Tsang, Sirisha Rambhatla, and Yan Liu. Interpretable and trustworthy deepfake detection via dynamic prototypes. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1973-1983, 2021. 2", + "[51] A Vaswani, N Shazeer, N Parmar, J Uszkoreit, L Jones, A Gomez, L Kaiser, and I Polosukhin. Attention is all you need. NeurIPS, 2017. 4", + "[52] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1, 2, 7", + "[53] Yuhan Wang, Xu Chen, Junwei Zhu, Wenqing Chu, Ying Tai, Chengjie Wang, Jilin Li, Yongjian Wu, Feiyue Huang, and Rongrong Ji. Hififace: 3d shape and semantic prior guided high fidelity face swapping. Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, 2021. 3, 5" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28059", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 7", + "[55] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22445-22455, 2023. 2", + "[56] Deressa Wodajo and Solomon Atnafu. Deepfake video detection using convolutional vision transformer. arXiv preprint arXiv:2102.11126, 2021. 6", + "[57] Yuting Xu, Jian Liang, Gengyun Jia, Ziming Yang, Yanhao Zhang, and Ran He. Tall: Thumbnail layout for deepfake video detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22658-22668, 2023. 1, 2, 6, 7", + "[58] Xin Yang, Yuezun Li, and Siwei Lyu. Exposing deep fakes using inconsistent head poses. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 8261-8265. IEEE, 2019. 5, 6, 8", + "[59] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3", + "[60] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. International Journal of Computer Vision, pages 1–15, 2024. 7", + "[61] Zhixing Zhang, Bichen Wu, Xiaoyan Wang, Yaqiao Luo, Luxin Zhang, Yinan Zhao, Peter Vajda, Dimitris Metaxas, and Licheng Yu. Avid: Any-length video inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7162-7172, 2024. 1, 4, 5, 7, 8", + "[62] Cairong Zhao, Chutian Wang, Guosheng Hu, Haonan Chen, Chun Liu, and Jinhui Tang. Istvt: interpretable spatial-temporal video transformer for deepfake detection. IEEE Transactions on Information Forensics and Security, 18: 1335-1348, 2023. 6", + "[63] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1, 2", + "[64] Xinye Zhou, Hu Han, Shiguang Shan, and Xin Chen. Fine-grained open-set deepfake detection via unsupervised domain adaptation. IEEE Transactions on Information Forensics and Security, 2024. 2" + ], + "bbox": [ + 91, + 90, + 480, + 796 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "28060", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Towards a Universal Synthetic Video Detector_ From Face or Background Manipulations to Fully AI-Generated Content/49360ca2-47e3-4c7b-81b1-6a87621996c4_model.json b/2025/Towards a Universal Synthetic Video Detector_ From Face or Background Manipulations to Fully AI-Generated Content/49360ca2-47e3-4c7b-81b1-6a87621996c4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..129c23755d4d03f8a45696a31d8d797b463773d9 --- /dev/null +++ b/2025/Towards a Universal Synthetic Video Detector_ From Face or Background Manipulations to Fully AI-Generated Content/49360ca2-47e3-4c7b-81b1-6a87621996c4_model.json @@ -0,0 +1,2477 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.13, + 0.872, + 0.177 + ], + "angle": 0, + "content": "Towards a Universal Synthetic Video Detector: From Face or Background Manipulations to Fully AI-Generated Content" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.203, + 0.884, + 0.24 + ], + "angle": 0, + "content": "Rohit Kundu\\(^{1,2}\\), Hao Xiong\\(^{1}\\), Vishal Mohanty\\(^{1}\\), Athula Balachandran\\(^{1}\\), Amit K. Roy-Chowdhury\\(^{2}\\), \\(^{1}\\)Google LLC, \\(^{2}\\)University of California, Riverside" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.242, + 0.872, + 0.259 + ], + "angle": 0, + "content": "{rohit.kundu@email, amertc@ece}.ucr.edu; {rohitkun, haoxg, vishal, athula}@google.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.484, + 0.431 + ], + "angle": 0, + "content": "Existing DeepFake detection techniques primarily focus on facial manipulations, such as face-swapping or lip-syncing. However, advancements in text-to-video (T2V) and image-to-video (I2V) generative models now allow fully AI-generated synthetic content and seamless background alterations, challenging face-centric detection methods and demanding more versatile approaches." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.484, + 0.751 + ], + "angle": 0, + "content": "To address this, we introduce the Universal Network for Identifying Tampered and synthEtic videos (UNITE) model, which, unlike traditional detectors, captures full-frame manipulations. UNITE extends detection capabilities to scenarios without faces, non-human subjects, and complex background modifications. It leverages a transformer-based architecture that processes domain-agnostic features extracted from videos via the SigLIP-So400M foundation model. Given limited datasets encompassing both facial/background alterations and T2V/I2V content, we integrate task-irrelevant data alongside standard DeepFake datasets in training. We further mitigate the model's tendency to over-focus on faces by incorporating an attention-diversity (AD) loss, which promotes diverse spatial attention across video frames. Combining AD loss with cross-entropy improves detection performance across varied contexts. Comparative evaluations demonstrate that UNITE outperforms state-of-the-art detectors on datasets featuring face/background manipulations and fully synthetic T2V/I2V videos, showcasing its adaptability and generalizable detection capabilities." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.784, + 0.222, + 0.799 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The rise of synthetic media, especially DeepFakes, has transformed content perception and interaction online. Hyper-realistic images produced by technologies like \\(\\mathrm{FLUX}_{1.1}\\) [1] are challenging for even humans to identify as fake, and with the development of their video model underway, this underscores the critical need for effective detec" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.291, + 0.886, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.689 + ], + "angle": 0, + "content": "Figure 1. Problem Overview: Existing DeepFake detection methods focus on identifying face-manipulated videos, most of which cannot perform inference unless there is a face detected in the video. However, with advancements like seamless background modifications (e.g., AVID [61]) and hyper-realistic content from games like GTA-V [24] and T2V/I2V models [9], a more comprehensive approach is needed. A model trained with only cross-entropy (CE) loss, using full frames, automatically focuses on the face, capturing temporal discontinuities through its transformer architecture, performing better than random \\((\\approx)\\) on T2V/I2V content but struggling with background manipulations. UNITE, with its attention-diversity (AD) loss, effectively detects both face/background manipulations and fully synthetic content." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.702, + 0.906, + 0.793 + ], + "angle": 0, + "content": "tion methods for media generated by deep neural networks. Traditional DeepFake generators [28, 49] focus on manipulating human faces through face-swapping and lip-syncing. However, powerful text-to-video (T2V) and image-to-video (I2V) models [42, 52, 63] have expanded manipulation possibilities beyond faces." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.903 + ], + "angle": 0, + "content": "While conventional detectors [10, 45, 57] perform well on older, face-centric DeepFake datasets [17, 30, 43], they often struggle with newer manipulations involving full scenes or backgrounds. DeepFake-O-Meter [26] is an open-source tool (consisting of several state-of-the-art models) for DeepFake detection, but it cannot run inference unless a human face is visible in the image/video. The rapid spread" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28050" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "of misinformation, particularly during critical periods such as elections, highlights the need for generalizable detection models capable of identifying diverse manipulations, including face, background, and fully AI-generated T2V/I2V content with/without human subjects (overview in Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.169, + 0.482, + 0.273 + ], + "angle": 0, + "content": "To address these challenges, we present a Universal Network for Identifying Tampered and synthEtic videos or UNITE model, which detects both partially manipulated (foreground/background) and fully synthetic videos. Unlike detectors focused solely on face detection, UNITE analyzes entire frames, regardless of whether a human subject is present in the videos." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.276, + 0.483, + 0.532 + ], + "angle": 0, + "content": "Given the inherent domain gaps in DeepFake datasets [44, 64], even when generated by similar techniques, we leverage the SigLIP-So400m [4] foundation model to extract domain-agnostic features. This serves as inputs to a learnable transformer with multi-head attention, enabling effective detection by capturing temporal inconsistencies in synthetic content. However, preliminary experiments show that training a transformer architecture solely with cross-entropy (CE) loss often leads to the attention heads converging on the face region (Fig. 3). As a result, the model struggles during inference when handling videos where a real human subject is placed in a manipulated background or content generated by T2V [8, 52] and I2V [35, 63] models. To address this, we introduce an \"attention-diversity\" (AD) loss that encourages attention heads to focus on different spatial regions, enhancing the model's ability to capture critical cues from both foreground and background." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.533, + 0.483, + 0.699 + ], + "angle": 0, + "content": "Due to the limited availability of open-source datasets covering face or background manipulations and fully AI-generated content, we employ innovative training strategies for UNITE. This includes integrating task-irrelevant data with standard DeepFake datasets to simulate AI-generated synthetic media. Beyond the popularly used FaceForensics++ [43] dataset, we utilize the SAIL-VOS-3D [24] dataset, originally designed for 3D video object segmentation in the game GTA-V. As this dataset is fully synthetic, it helps simulate AI-generated media, enhancing our model's ability to detect diverse forms of synthetic manipulation." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.701, + 0.435, + 0.715 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.72, + 0.483, + 0.765 + ], + "angle": 0, + "content": "- We propose UNITE, a model for detecting partially manipulated (foreground/background) and fully synthetic videos, moving beyond face-centric DeepFake detection." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.766, + 0.483, + 0.81 + ], + "angle": 0, + "content": "- Unlike detectors relying on face detection and cropping, our model can process full video frames and can detect fakes even if there are no human subjects." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.811, + 0.483, + 0.856 + ], + "angle": 0, + "content": "- Using the SigLIP-So400m foundation model [4], we extract domain-agnostic features, enabling generalization to in-the-wild DeepFakes." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- We introduce an attention-diversity loss, encouraging attention heads to focus on diverse spatial regions, enhancing detection beyond face manipulations." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "- Unlike existing methods that evaluate on a few datasets, we comprehensively assess UNITE on a broad range of face and synthetic datasets with various T2V/I2V generators, outperforming detection in foreground, background, and T2V/I2V manipulations." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.181, + 0.655, + 0.196 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.206, + 0.905, + 0.523 + ], + "angle": 0, + "content": "Face-centric DeepFake Detection: Methods such as [13, 19, 40] focus on modeling spatial inconsistencies for DeepFake detection. Concas et al. [13] track specific facial features (eyes, nose, mouth) in a high-frequency domain, as fake videos exhibit distinct behaviors in this space compared to real ones. However, due to their dataset-centric architectures, it is not a generalizable approach. PUDD [40] learns person-specific prototypes by modeling patterns from real videos and assessing how closely inference videos match these prototypes, making it less effective for detecting in-the-wild DeepFakes. LAA-Net [39] attempts to improve generalizability through a multi-task attention module that detects artifacts via heatmap and self-consistency regression, but it works by analyzing every still frame of a video, thus limiting its scalability. The authors in [36] tackled the challenging problem of detecting and localizing expression swaps, where the number of manipulated pixels is even fewer compared to identity swaps. Although the authors obtained good localization performance, their reliance on a CNN-based architecture limits their ability to capture temporal inconsistencies effectively." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.524, + 0.905, + 0.735 + ], + "angle": 0, + "content": "DPNet [50] and ID-Reveal [16] are identity-aware temporal artifact modeling methods: DPNet [50] focuses on interpretable, prototype-based detection of unnatural movements through dynamic feature representations, while IDReveal [16] leverages metric learning on real data to detect biometric inconsistencies. However these methods assume access to authentic reference videos, making them unfit for in-the-wild DeepFake detection. Shifting from identity-specific methods, TALL [57] constructs composite \"thumbnail\" images from four consecutive frames taken at random timestamps for temporal analysis. TI2Net [32] captures temporal anomalies by subtracting consecutive frame features but relies on the assumption that FaceSwap manipulations show pronounced inter-frame discrepancies." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.735, + 0.905, + 0.81 + ], + "angle": 0, + "content": "Choi et al. [11] approach DeepFake detection by targeting suppressed variance in style-based latent temporal features via a StyleGRU module. However, this method is limited to cropped face regions, making it ineffective for T2V/I2V or background-manipulated videos." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.811, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Synthetic Video Detection: While there has been considerable effort in synthetic image detection [14, 15, 55], synthetic video detection is a scarcely explored area. DeMamba [9], a recently proposed synthetic video detector, analyzes small zones of video frames to capture how pixels change spatiotemporally. Its continuous scanning approach" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28051" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.442, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.338, + 0.484, + 0.436 + ], + "angle": 0, + "content": "Figure 2. UNITE architecture overview: We extract domain-agnostic features \\((\\xi)\\) using the SigLIP-So400m foundation model [4] to mitigate domain gaps between DeepFake datasets (Sec. 3.2). These embeddings, combined with positional encodings, are input to a transformer with multi-head attention and MLP layers (Sec. 3.3), culminating in a classifier for final predictions. AD-loss (Sec. 3.4) encourages the model attention to span diverse spatial regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.45, + 0.484, + 0.616 + ], + "angle": 0, + "content": "better tracks subtle changes and patterns, making it easier to spot manipulated or fake content. The authors also propose a million scale T2V/I2V dataset (which we call the DeMamba dataset hereforth) to evaluate their performance. However, the method was not evaluated on typical DeepFake datasets [30, 43, 53] and thus is not fit for human-centric DeepFake detection. In contrast our UNITE model is built to detect partially manipulated as well as fully AI-generated videos. A key reason for the scarcity of synthetic video detection methods in the literature is the lack of standardized datasets for training and evaluation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.63, + 0.265, + 0.647 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.484, + 0.748 + ], + "angle": 0, + "content": "In this section we first describe our problem setup (Sec. 3.1), then we describe the foundation model-based video encoding (Sec. 3.2), our trainable transformer architecture (Sec. 3.3) and finally the novel loss function for training the UNITE model (Sec. 3.4). The overview of the UNITE architecture is shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.242, + 0.773 + ], + "angle": 0, + "content": "3.1. Problem Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.901 + ], + "angle": 0, + "content": "For each video in the dataset, we implement a frame sampling strategy where every other frame is extracted. From the resulting collection of frames, we construct video segment samples consisting of \\( n_f = 64 \\) consecutive frames, which are defined as a single data sample, denoted as \"v\". This selection aligns with the context window established for our video transformer model, ensuring that each sample retains the temporal coherence necessary for effective train-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "ing. In instances where the video duration results in fewer than \\( n_f \\) frames, we apply padding (ablation in supplementary) to ensure that each sample remains uniform in size." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.907, + 0.304 + ], + "angle": 0, + "content": "Given that the original video-level classification labels are known and fake videos contain manipulations throughout their entirety (i.e., no videos have manipulations limited to specific frames), each \\( n_f \\)-frame segment generated from a given video is assigned the same label (denoted by \\( l \\in 0, \\dots, n_c \\)) as the source video, with number of possible classes as \\( n_c \\). The aggregation of these samples across all videos forms our final dataset, \\( v_i, l_{i=1}^N \\in \\mathcal{V} \\), where \\( N \\) is the total number of samples obtained. This approach serves as a data augmentation technique, increasing the number of training samples without additional data collection." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.314, + 0.834, + 0.33 + ], + "angle": 0, + "content": "3.2. Domain-Agnostic Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.338, + 0.907, + 0.563 + ], + "angle": 0, + "content": "Our objective is to leverage the trained UNITE model for detecting in-the-wild DeepFakes, necessitating careful consideration of the domain gap [7, 34] between the train and test datasets. To effectively mitigate the impact of domain discrepancies, we utilize the SigLIP foundation model [59] to extract domain-agnostic features. Specifically, we employ the \"shape-optimized\" ViT [4] image encoder, whose architecture is rooted in [18]. The SigLIP-So400m [4] model was contrastively pretrained on 3B diverse examples using a sigmoid loss, enabling it to capture robust, generalizable features while maintaining a compact size (400M parameters). This extensive pretraining makes SigLIP particularly adept at extracting domain-invariant representations, crucial for handling the significant variability encountered in DeepFake datasets and real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.909, + 0.749 + ], + "angle": 0, + "content": "For every frame \\( f_{j} \\) (resized to \\( \\mathbb{R}^{384\\times 384\\times 3} \\)), of a video sample \\( v_{i}\\in \\mathcal{V} \\), we obtain an image encoding from the frozen SigLIP-So400m [4] model as \\( e_j = SigLIP(f_j)\\in \\mathbb{R}^{t_s\\times d_s} \\), where \\( j\\in \\{1,2,\\dots ,n_f\\} \\), since there are \\( n_f \\) frames per video sample (as explained in Sec. 3.1), \\( t_s = 729 \\) represents the token length and \\( d_{s} = 1152 \\) represents the feature dimension. We concatenate these frame-level features for the video sample \\( v_{i} \\), while maintaining the order of the frames, to obtain a video segment encoding denoted by \\( \\xi_{i}\\in \\mathbb{R}^{n_{f}\\times t_{s}\\times d_{s}} \\), which are used as input to our trainable transformer architecture. This encoded video dataset can be denoted as \\( \\{\\xi_i,l_i\\}_{i = 1}^N\\in \\mathcal{V}_\\xi \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.834, + 0.772 + ], + "angle": 0, + "content": "3.3. Learnable Transformer Architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The UNITE architecture is a multi-head self-attention (MHSA) based transformer model designed specifically for video classification tasks. It leverages the strengths of Transformer networks to process sequences of video frame embeddings effectively, allowing for robust classification performance. The UNITE architecture allows for adjustable depth, signifying the number of encoder blocks stacked within the model. A deeper architecture can capture more" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28052" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "intricate features and dependencies at the cost of increased computational complexity. For our architecture the depth is set to 4. Ablation experiments with different transformer depths are provided in Sec. 4.4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.173, + 0.248, + 0.187 + ], + "angle": 0, + "content": "3.3.1 Encoder Block" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.198, + 0.48, + 0.213 + ], + "angle": 0, + "content": "Each encoder block consists of the following components:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.214, + 0.484, + 0.365 + ], + "angle": 0, + "content": "Multi-Head Self-Attention Layer: This layer allows the model to focus on different parts of the input sequence simultaneously. With 12 attention heads in our architecture, it captures diverse interactions and dependencies among frames. Each head computes attention scores independently, learning multiple representations of the input data. This enhances the model's ability to detect subtle temporal variations, which is crucial for video classification. The use of scaled dot-product attention further helps mitigate large gradients and stabilizes training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.365, + 0.484, + 0.487 + ], + "angle": 0, + "content": "Layer Normalization and Residual Connections: To aid gradient flow during training, each sub-layer incorporates residual connections that add the input back to the output, mitigating vanishing gradient issues in deep networks. Layer normalization follows to stabilize activations and enhance convergence speed. Additionally, dropout is applied post-attention to reduce overfitting by randomly deactivating a portion of neurons during training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.487, + 0.484, + 0.592 + ], + "angle": 0, + "content": "Feed-Forward Network (MLP): The second sub-layer is a point-wise feed-forward network with two dense layers separated by a GELU activation [23]. This MLP enables nonlinear feature transformations, enhancing the model's expressiveness. By projecting attention outputs into a higher-dimensional space before reducing them, it captures complex interactions beyond linear transformations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.593, + 0.484, + 0.699 + ], + "angle": 0, + "content": "Attention maps are derived from the outputs of the encoder blocks to be used for computing the Attention-Diversity loss (Sec. 3.4) and to illustrate how the model prioritizes different spatial regions in the frames during the classification process (Fig. 3). This interpretability is valuable for understanding the model's decision-making, particularly in complex video classification scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.72, + 0.285, + 0.735 + ], + "angle": 0, + "content": "3.3.2 Positional Encoding" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.744, + 0.484, + 0.835 + ], + "angle": 0, + "content": "We utilize a sine-cosine positional encoding scheme following the original transformer [51] to provide a unique position identifier for each input token. For every frame \\( f_{j} \\) in the video sample with encoded feature dimension \\( d_{s} \\) (as per Sec. 3.2), we compute the positional encoding for odd and even indexed feature dimensions as," + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.858, + 0.49, + 0.901 + ], + "angle": 0, + "content": "\\[\nP E _ {(j, 2 i + 1)} = \\cos \\left(\\frac {j}{1 0 0 0 0 ^ {\\frac {2 i}{d _ {s}}}}\\right), \\quad P E _ {(j, 2 i)} = \\sin \\left(\\frac {j}{1 0 0 0 0 ^ {\\frac {2 i}{d _ {s}}}}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "where, \\( i \\) denotes the feature index in the encoded feature dimension \\( d_{s} \\). We embed these encodings directly into the input tokens of our video transformer to provide positional context, for better temporal modeling during the attention computation with minimal computational overhead." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.175, + 0.74, + 0.192 + ], + "angle": 0, + "content": "3.4. Attention-Diversity Loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.906, + 0.319 + ], + "angle": 0, + "content": "Training a multi-head attention transformer with only cross-entropy loss makes the attention maps focus only on the face regions of the frame (as evident in Fig. 3). However, we aim to detect fake videos where the manipulation might not be in the face at all, and instead be in the background (like video inpainting with the AVID model [61]). To ensure that the attention heads focus on diverse spatial regions of the video frames, we devise the \"Attention-Diversity\" (AD) loss." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.32, + 0.907, + 0.44 + ], + "angle": 0, + "content": "AD-loss is designed to minimize overlap among attention maps (obtained from the first encoder block of the trainable transformer architecture in 3.3.1) while maintaining consistency across different input video samples. The attention outputs from the video transformer model \\(\\mathcal{A} \\in \\mathbb{R}^{n_h \\times t_s \\times d_s}\\) are used to pool the input SigLIP-So400m features \\(\\xi \\in \\mathbb{R}^{n_f \\times t_s \\times d_s}\\), resulting in a pooled feature tensor \\(\\mathcal{P} \\in \\mathbb{R}^{n_h \\times n_f}\\) according to," + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.452, + 0.906, + 0.494 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} = \\sum_ {j = 1} ^ {t _ {s}} \\sum_ {k = 1} ^ {d _ {s}} \\mathcal {A} _ {h, j, k} \\cdot \\xi_ {f, j, k}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.498, + 0.906, + 0.528 + ], + "angle": 0, + "content": "where \\( f \\) indexes the frames \\( (n_f) \\), \\( h \\) indexes the number of attention heads \\( (n_h) \\), and \\( j, k \\) are spatial positions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.529, + 0.906, + 0.65 + ], + "angle": 0, + "content": "We compute \"feature centers\" which are points in the feature space that serve as anchors for the learned representations of different classes. These centers represent the average features of samples from a particular class, allowing the model to capture essential characteristics of that class. In the context of the AD-loss, feature centers \\(\\mathcal{C} \\in \\mathbb{R}^{n_h \\times n_f}\\) are dynamically updated in each training iteration \\(\\tau\\) based on the pooled feature vectors \\(\\mathcal{P}\\), following," + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.66, + 0.906, + 0.702 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} ^ {\\tau} = \\mathcal {C} ^ {\\tau - 1} - \\eta \\left(\\mathcal {C} ^ {\\tau - 1} - \\frac {1}{B} \\sum_ {b = 1} ^ {B} \\mathcal {P} _ {b}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.705, + 0.906, + 0.766 + ], + "angle": 0, + "content": "where \\(\\eta\\) represents the learning rate (set to 0.05 in our experiments) for feature center updates and \\(B\\) is the batch dimension. The feature centers are initialized with zeros \\((\\mathcal{C}^0 = [0]_{n_h\\times n_f})\\) at the beginning of the model training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.767, + 0.907, + 0.902 + ], + "angle": 0, + "content": "AD-loss consists of two distinct components: within-class loss and between-class loss inspired by the Fisher Discriminant Analysis for deep networks [22]. The within-class term calculates the distance between the pooled feature vectors \\(\\mathcal{P}\\) and their corresponding feature centers \\(\\mathcal{C}\\), promoting closeness among similar samples. The between-class term measures the distance between feature centers of different classes, encouraging them to be spaced apart in the feature space, encouraging separability." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28053" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.111, + 0.092, + 0.393, + 0.106 + ], + "angle": 0, + "content": "The within-class loss term is calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.121, + 0.482, + 0.138 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {w i t h i n}} = \\max \\left(\\| \\mathcal {P} - \\mathcal {C} \\| _ {2} - \\delta_ {\\text {w i t h i n}}, 0\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.141, + 0.483, + 0.202 + ], + "angle": 0, + "content": "where \\(\\delta_{\\mathrm{within}} \\in \\mathbb{R}^{n_c}\\) is a hyperparameter controlling the allowable distance between feature vectors and their respective feature centers for inputs of the same class, encouraging tighter clustering and \\(\\| \\cdot \\|_2\\) represents \\(L_2\\) normalization." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.202, + 0.37, + 0.216 + ], + "angle": 0, + "content": "The between-class loss is calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.235, + 0.482, + 0.291 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {b e t w e e n}} = \\sum_ {\\substack {k \\neq l \\\\ (k, l) \\in \\left(n _ {h}, n _ {h}\\right)}} \\max \\left(\\delta_ {\\text {b e t w e e n}} - \\| \\mathcal {C} _ {k} - \\mathcal {C} _ {l} \\| _ {2}, 0\\right), \\tag{5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.292, + 0.484, + 0.368 + ], + "angle": 0, + "content": "where \\(\\delta_{\\mathrm{between}}\\) is a predefined hyperparameter that ensures a minimum distance between feature centers of different classes. Thus the AD-loss is a simple addition of these two components \\(\\mathcal{L}_{AD} = \\mathcal{L}_{\\mathrm{within}} + \\mathcal{L}_{\\mathrm{between}}\\), making the final objective function for training the UNITE model," + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.382, + 0.483, + 0.398 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {U N I T E}} = \\lambda_ {1} \\cdot \\mathcal {L} _ {C E} + \\lambda_ {2} \\cdot \\mathcal {L} _ {A D}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.401, + 0.484, + 0.432 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{CE}\\) is the traditional cross-entropy loss, and \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are loss-weighting hyperparameters." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.443, + 0.224, + 0.459 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.467, + 0.194, + 0.482 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.484, + 0.596 + ], + "angle": 0, + "content": "Most DeepFake datasets focus on face manipulations, such as face-swaps or lip-syncing, and lack fully AI-generated video content. The DeMamba dataset [9] features videos from T2V/I2V models, including powerful diffusion models like SORA [5]. It is not a human-centric dataset and primarily contains videos from natural scenes. We use its validation split to evaluate UNITE." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.484, + 0.687 + ], + "angle": 0, + "content": "To train UNITE, we employ the FaceForensics++ (FF++) [43] dataset (c23 i.e., medium-quality) and the SAIL-VOS-3D dataset [24], which includes synthetic (although not AI-generated) GTA-V game videos featuring human subjects suitable for DeepFake detection. To evaluate the UNITE model, we use:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.688, + 0.483, + 0.733 + ], + "angle": 0, + "content": "- Face manipulated data: \\(\\mathrm{FF}++\\) [43] (in-domain evaluation), CelebDF [30], DeeperForensics [25], Deepfake-TIMIT [27], HifiFace [53], UADFV [58]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.734, + 0.483, + 0.763 + ], + "angle": 0, + "content": "- Background manipulated data: Sample videos from the AVID [61] model provided publicly by the authors." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.764, + 0.483, + 0.793 + ], + "angle": 0, + "content": "- Fully synthetic data: GTA-V [24] (in-domain evaluation), DeMamba [9]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.794, + 0.483, + 0.824 + ], + "angle": 0, + "content": "- In-the-wild DeepFakes: Publicly available videos from the New York Times DeepFake quiz [48]." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.688, + 0.483, + 0.824 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.304, + 0.848 + ], + "angle": 0, + "content": "4.2. Results and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metrics: Although most DeepFake detection methods only report accuracy, we evaluate UNITE using additional metrics: (1) AUC (Area Under the Precision-Recall" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "Curve); (2) Precision@0.5 and (3) Recall@0.5 at a 0.5 confidence threshold; (4) Precs@Rec=0.8 (precision at 0.8 recall); (5) Rec@Precs=0.8 (recall at 0.8 precision)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.906, + 0.26 + ], + "angle": 0, + "content": "Training Details: UNITE is trained using an AdamW optimizer [33] with an initial learning rate of 0.0001 and a decay rate of 0.5 every 1000 steps. For AD-loss, \\(\\delta_{\\mathrm{between}} = 0.5\\) and \\(\\delta_{\\mathrm{within}}\\) is set to \\([0.01, -2]\\) for binary and \\([0.01, -2, 1]\\) for fine-grained classification in Sec. 4.3 (see supplementary for sensitivity). Loss weights are \\(\\lambda_1 = \\lambda_2 = 0.5\\). Training uses a batch size of 32 for 25 epochs on 8 TPUv3 chips, with the framework implemented in TensorFlow." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.262, + 0.907, + 0.398 + ], + "angle": 0, + "content": "Quantitative Results: The results obtained by the UNITE model when trained on \\(\\mathrm{FF}++\\) [43] alone, and when trained with both \\(\\mathrm{FF}++\\) [43] and GTA-V [24] data and evaluated on the various datasets mentioned in Sec. 4.1 are shown in Table 1. On the AVID [61] and DeMamba datasets [9], the model performs poorly when trained with only face manipulated data, but the performance enhances several-fold when the GTA-V [24] synthetic data, even though it is not AI-generated, is used in training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.906, + 0.58 + ], + "angle": 0, + "content": "Interestingly, when evaluating on face-manipulated datasets such as CelebDF [30] and DeeperForensics [25], we observe a performance boost, particularly for CelebDF [30], when GTA-V [24] data is included in the training set. Intuitively, one would expect that training solely on \\(\\mathrm{FF}++\\) [43] would yield similar results since the inference set consists of similar face-manipulated data. This deviation from the expected behavior comes from the contribution of the AD-Loss, as the ablation study (Fig. 4) using only CE-loss compared to both CE and AD-losses reveals that the performance is enhanced with GTA-V [24] data in training only in the latter (detailed discussion in Sec. 4.4)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.906, + 0.703 + ], + "angle": 0, + "content": "To evaluate UNITE on in-the-wild DeepFakes, we attempted the recent New York Times quiz [48] (denoted by NYTimes) which provides 10 videos (4 real and 6 fake) for testing DeepFake detectors. UNITE got 8 of those 10 videos correct, even though some of the face-swap videos were indistinguishable from real videos, even by humans. On all the AI-generated videos, UNITE made the correct prediction (when trained on \\(\\mathrm{FF} + +\\) [43] and GTA-V [24])." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Attention Heatmaps: To analyze the impact of CE and AD-losses on attention distribution, we extracted attention features from the first encoder of the UNITE model, comparing models trained with CE loss alone versus that trained with both CE and AD-losses. The resulting heatmaps, shown in Fig. 3, reveal that the model trained with only CE loss tends to concentrate primarily on the face region. In contrast, the model trained with the combined \\(\\mathrm{CE + AD}\\) loss demonstrates a broader attention span across the frame, as evidenced by a lighter, more distributed bluish tone in the heatmaps, indicating increased spatial diversity in the model's focus. This is why on the example shown in the first row (face-manipulated video) both models give the correct" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28054" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.117 + ], + "angle": 0, + "content": "Table 1. Results from the UNITE model trained with (1) FF++ [43] only and (2) FF++ [43] combined with GTA-V [24]. All other results reflect cross-dataset evaluations except for FF++ and GTA-V (when trained). Performance gains are highlighted in green." + }, + { + "type": "table", + "bbox": [ + 0.172, + 0.118, + 0.823, + 0.359 + ], + "angle": 0, + "content": "
TrainTest
FF++GTA-VDatasetAccuracyAUCPrecision@0.5Recall@0.5Precs@Rec=0.8Rec@Precs=0.8
Face Manipulated Data
✓FF++99.53%99.77%99.94%99.49%99.94%99.94%
✓CelebDF72.61%94.05%96.45%61.22%80.45%61.22%
✓DeeperForensics91.35%100.00%100.00%91.35%100.00%91.35%
✓DeepfakeTIMIT86.90%86.46%83.61%83.97%88.90%81.33%
✓HiFiFace63.63%62.47%67.12%63.63%59.30%63.63%
✓UADFV94.12%94.38%95.68%97.11%93.79%94.38%
✓✓FF++99.96%(+0.43)99.89%(+0.12)100.00%(+0.06)99.84%(+0.35)100.00%(+0.06)99.96%(+0.02)
✓✓CelebDF95.11%(+22.50)94.36%(+0.31)96.82%(+0.37)68.75%(+7.53)96.53%(+16.08)68.75%(+7.53)
✓✓DeeperForensics99.62%(+8.27)100.00%(+0.00)100.00%(+0.00)99.62%(+8.27)100.00%(+0.00)99.63%(+8.28)
✓✓DeepfakeTIMIT91.90%(+5.00)91.33%(+4.87)90.45%(+6.84)88.39%(+4.42)100.00%(+11.10)91.95%(+10.62)
✓✓HiFiFace75.62%(+11.99)81.24%(+18.77)79.55%(+12.43)71.71%(+8.08)75.62%(+16.32)72.47%(+8.84)
✓✓UADFV97.01%(+2.89)94.95%(+0.57)96.89%(+1.21)100.00%(+2.89)94.12%(+0.33)100.00%(+5.62)
Background Manipulated Data
✓AVID41.67%33.33%33.33%41.67%41.67%33.33%
✓✓AVID100.00%(+58.33)100.00%(+66.67)100.00%(+66.67)100.00%(+58.33)100.00%(+58.33)100.00%(+66.67)
Fully Synthetic Data
✓GTA-V60.16%61.52%60.16%58.73%63.29%58.73%
✓DeMamba61.47%57.38%67.73%33.01%62.15%54.16%
✓✓GTA-V100.00%(+39.84)100.00%(+38.48)100.00%(+39.84)100.00%(+41.27)100.00%(+36.71)100.00%(+41.27)
✓✓DeMamba87.12%(+25.65)93.75%(+36.67)92.76%(+25.03)89.60%(+56.59)89.81%(+27.66)92.12%(+37.96)
In-the-wild DeepFakes
✓NYTimes [48]50.00%53.74%50.00%25.00%0.00%0.00%
✓✓NYTimes [48]80.00%(+30.00)97.42%(+43.68)83.33%(+33.33)83.33%(+58.33)60.00%(+60.00)100.00%(+100.00)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.363, + 0.484, + 0.446 + ], + "angle": 0, + "content": "Table 2. SOTA Comparison on Face-Manipulated Data: We compared the performance of UNITE with recent DeepFake detectors, in terms of detection accuracy on various face manipulated datasets. UNITE outperforms the existing methods. Bold shows the current best results and the previous best and second-best results are highlighted in red and blue respectively." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.45, + 0.458, + 0.608 + ], + "angle": 0, + "content": "
MethodFF++CelebDFDeeperForensicsUADFV
TALL [57]98.65%90.79%99.62%-
ISTVT [62]99.00%84.10%98.60%-
Concas et al. [13]99.49%---
PUDD [40]-95.10%--
TI2Net [32]99.95%68.22%76.08%-
LRNet [46]99.89%53.20%56.77%-
Choi et al. [11]-89.00%99.00%-
Lin et al. [31]98.28%74.42%--
Guo et al. [21]99.24%84.97%--
Li et al. (Res-152) [29]---93.80%
HeadPose [58]---89.00%
CViT [56]93.00%--93.75%
FakeCatcher [12]94.65%91.50%--
MesoNet [3]---82.40%
UNITE (Ours)99.96%95.11%99.62%97.01%
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.484, + 0.702 + ], + "angle": 0, + "content": "prediction (\"fake\") with high confidence, but on the row-2 example (video generated from the Sora [6] T2V model) the model trained with only CE loss predicted \"real\" with a confidence score of \\(99.25\\%\\) (which is wrong), but the model trained with \\(\\mathrm{CE + AD}\\) losses predicted \"fake\" with a confidence of \\(100.00\\%\\), which is correct." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.484, + 0.824 + ], + "angle": 0, + "content": "Comparison on Face Manipulated Data: UNITE is compared with recent state-of-the-art (SOTA) DeepFake detection models in Table 2. The methods we compared against were specifically designed for detecting face manipulations and often crop faces from the videos, during both training and inference of the models. However, UNITE, which is designed to detect a diverse range of fake videos, still outperforms these detectors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Comparison on Synthetic Data: We evaluated UNITE against state-of-the-art synthetic video detectors on the DeMamba dataset [9] (validation split), with results in Table 3. While existing detectors were trained on the DeMamba train split and validated on the DeMamba validation split," + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.361, + 0.648, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.414, + 0.639, + 0.424 + ], + "angle": 0, + "content": "(a) Original Frame" + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.361, + 0.77, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.414, + 0.756, + 0.424 + ], + "angle": 0, + "content": "(b) CE Loss only" + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.361, + 0.895, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.789, + 0.414, + 0.88, + 0.424 + ], + "angle": 0, + "content": "(c) \\(\\mathrm{CE + AD}\\) Loss" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.425, + 0.648, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.481, + 0.639, + 0.492 + ], + "angle": 0, + "content": "(d) Original Frame" + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.425, + 0.77, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.481, + 0.756, + 0.492 + ], + "angle": 0, + "content": "(e) CE Loss only" + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.425, + 0.895, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.789, + 0.481, + 0.879, + 0.492 + ], + "angle": 0, + "content": "(f) \\(\\mathrm{CE + AD}\\) Loss" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.493, + 0.907, + 0.632 + ], + "angle": 0, + "content": "Figure 3. Comparison of attention heatmaps when UNITE is trained with only CE loss (second column) versus CE and AD-losses (third column). The CE-loss heatmap predominantly focuses on the face region, whereas the \\(\\mathrm{CE + AD}\\) loss heatmap demonstrates more distributed attention across the entire frame, as indicated by a broader bluish tone in the heatmaps. Both samples are \"fake\" videos- (a) is taken from Celeb-DF [30] and (d) is taken from DeMamba [9] and generated by OpenAI's Sora [6] (Best viewed as GIFs provided in the supplementary material). Yellow to blue shades show increasing attention in the plot." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.907, + 0.7 + ], + "angle": 0, + "content": "UNITE was trained on FF++ [43] and GTA-V [24]. Despite not being trained on DeMamba, UNITE's average performance (including real videos) surpassed current SOTA detectors, with competitive generator-wise results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.701, + 0.907, + 0.807 + ], + "angle": 0, + "content": "Thus, UNITE reliably detects a diverse range of fake videos (even in cross-domain settings), including T2V/I2V videos and background manipulations, while consistently maintaining state-of-the-art performance on traditional face-manipulated data. UNITE eliminates the need to have separate DeepFake and T2V/I2V video detector models by by handling both tasks within a single trained model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.817, + 0.733, + 0.833 + ], + "angle": 0, + "content": "4.3. Finegrained Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In fake media detection, flagging videos partially or fully manipulated by AI also necessitates fine-grained classification into three classes: (1) real videos, (2) partially manipulated videos, and (3) fully synthetic videos. While binary" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28055" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.172 + ], + "angle": 0, + "content": "Table 3. SOTA Comparison on Synthetic Data: On the DeMamba dataset (validation split), we compare the performance of UNITE, which was NOT trained on DeMamba train split, against state-of-the-art detectors which were trained on DeMamba train split (results taken from Chen et al. [9]). We report the results (\\(P = \\text{Precision}\\) and \\(R = \\text{Recall}\\)) on the individual T2V/I2V generators and the average performance across the entire validation set (\\(Avg\\), which also includes real videos). Although the direct comparison is unfair against UNITE which was trained with FF++ [43] and GTA-V [24], our method still outperforms these synthetic video detectors. Bold shows the current best results and the previous best and second-best results are highlighted in red and blue respectively." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.173, + 0.88, + 0.412 + ], + "angle": 0, + "content": "
MethodMetricsSora [6]Morph Studio [2]Runway ML (Gen2) [42]HotShot [38]Lavie [54]Show-1 [60]Moon Valley [37]Crafter [8]Model Scope [52]Wild Scrape [9]Avg
TALL [57]P71.15%96.89%98.51%79.38%84.59%79.38%98.79%99.02%92.70%76.47%87.91%
R91.07%98.28%97.83%83.00%76.57%79.57%99.52%98.93%94.14%66.31%88.52%
F3Net [41]P68.27%99.89%99.67%89.35%57.00%36.57%99.52%99.71%93.80%88.41%88.73%
R83.93%99.71%98.62%77.57%85.24%63.17%99.58%99.89%89.43%76.78%81.88%
NPR [47]P91.07%99.57%99.49%24.29%89.64%57.71%97.12%99.86%94.29%87.80%82.45%
R91.07%99.57%99.49%24.29%89.64%57.71%97.12%99.86%94.29%87.80%84.08%
STIL [20]P57.21%99.08%99.32%86.19%82.24%70.43%99.25%98.96%97.18%81.32%87.12%
R67.86%96.00%98.41%96.14%77.14%80.43%97.44%96.93%96.29%68.36%82.22%
MINTIME-CLIP-B [9]P83.21%99.99%99.67%50.84%99.20%99.27%99.76%99.99%91.83%91.77%91.55%
R89.29%100.00%98.99%26.43%96.79%98.14%99.84%100.00%84.29%82.38%87.62%
FTCN-CLIP-B [9]P91.79%99.99%99.79%45.94%99.76%97.80%99.99%99.99%94.69%92.32%92.21%
R87.50%100.00%98.91%17.71%97.71%91.86%100.00%100.00%85.29%82.83%86.18%
CLIP-B-PT [9]P67.80%43.56%70.88%29.97%52.97%35.36%55.52%66.03%44.23%42.99%44.83%
R85.71%82.43%90.36%71.00%79.29%75.43%89.62%86.29%82.14%75.16%81.74%
DeMamba-CLIP-PT [9]P25.87%95.14%96.23%73.43%83.31%75.49%90.17%95.06%95.05%69.95%79.97%
R58.93%96.43%93.12%68.00%69.36%69.00%89.14%91.86%96.14%56.59%78.86%
XCLIP-B-PT [9]P16.39%72.16%87.77%39.86%65.57%54.26%75.23%84.80%61.60%55.28%61.29%
R81.34%82.15%83.35%80.98%81.82%81.55%82.14%82.98%81.93%81.10%81.93%
DeMamba-XCLIP-PT [9]P18.26%93.50%94.72%69.94%78.08%71.50%83.95%92.23%93.54%68.10%76.38%
R66.07%95.86%94.64%77.86%75.36%80.29%90.89%92.50%96.00%66.41%83.59%
XCLIP-B-FT [9]P64.42%99.73%96.78%70.98%90.35%77.28%97.34%99.84%82.01%88.97%86.77%
R82.14%99.57%93.62%61.29%79.36%69.71%97.92%99.79%77.14%83.59%84.41%
UNITE (Ours)P88.57%100.00%100.00%90.16%89.91%98.34%99.52%100.00%98.96%92.56%92.76%
R92.11%100.00%94.62%96.93%98.12%99.86%98.69%100.00%96.29%89.89%89.60%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.417, + 0.483, + 0.459 + ], + "angle": 0, + "content": "Table 4. Results obtained by the UNITE model on finegrained classes: detecting whether a video is real, partially manipulated or fully AI-generated. Performance gains are mentioned in green." + }, + { + "type": "table", + "bbox": [ + 0.149, + 0.464, + 0.419, + 0.666 + ], + "angle": 0, + "content": "
TrainTest
FF++GTA-VDatasetAccuracy
✓FF++96.46%
✓CelebDF60.40%
✓DeeperForensics71.04%
✓DeepFakeTIMIT80.68%
✓HifiFace43.27%
✓UADFV90.37%
✓AVID50.00%
✓GTA-V0.00%
✓DeMamba38.29%
✓✓FF++97.70%(+1.24)
✓✓CelebDF70.17%(+9.77)
✓✓DeeperForensics80.59%(+9.55)
✓✓DeepFakeTIMIT81.04%(+0.36)
✓✓HifiFace64.20%(+20.93)
✓✓UADFV92.61%(+1.24)
✓✓AVID62.50%(+12.50)
✓✓GTA-V100.00%(+100.00)
✓✓DeMamba65.84%(+27.55)
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.483, + 0.716 + ], + "angle": 0, + "content": "real vs. fake classification may suffice to curb misinformation, fine-grained classification offers greater explainability in otherwise black-box models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Using this defined convention of finegrained classes, all the face manipulated datasets (CelebDF [30], DeeperForensics [25], etc.) and the videos from the AVID [61] model (background manipulations) fall under class-2, and the videos from the GTA-V [24] and the DeMamba [9] datasets fall in class-3. In this setup, only the final classification layer of the UNITE transformer architecture is changed to 3 neurons with softmax activation. The results obtained are presented in Table 4. As expected, when the model is trained only with \\(\\mathrm{FF}++\\) [43], the performance on fully synthetic data is poor (since the model has not seen any synthetic videos). When the GTA-V [24] data is also" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.421, + 0.882, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.611, + 0.512, + 0.81, + 0.524 + ], + "angle": 0, + "content": "(a) UNITE trained on \\(\\mathrm{FF} + +\\) [43] only" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.529, + 0.88, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.579, + 0.62, + 0.842, + 0.631 + ], + "angle": 0, + "content": "(b) UNITE trained on \\(\\mathrm{FF} + +\\) [43] and GTA-V [24]" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.632, + 0.906, + 0.701 + ], + "angle": 0, + "content": "Figure 4. Ablation Results to show the effect of changing the loss functions used to train the UNITE model. The combination of the cross-entropy (CE) and Attention Diversity (AD) losses always performs better, and the contribution from the AD-loss component increases significantly when using fully synthetic data for training." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.711, + 0.838, + 0.727 + ], + "angle": 0, + "content": "used in training, the performance is much higher." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.667, + 0.757 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We conducted an analysis to assess the contribution of individual loss functions to the training of the UNITE model. The outcomes of this ablation study are presented in Fig. 4. Specifically, we evaluated the impact of the AD-loss in two scenarios: (1) when UNITE was trained solely on the FF++ dataset [43] (Fig. 4(a)), and (2) when UNITE was trained using a combination of the FF++ [43] and the GTA-V [24] datasets (Fig. 4(b)). The combination of the two losses always performed better than the individual losses." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28056" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.094, + 0.437, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.231, + 0.483, + 0.314 + ], + "angle": 0, + "content": "Figure 5. No. of frames vs. Performance: Performance analysis of UNITE based on the number of frames sampled per video segment. The results illustrate that as the number of frames increases from 1 to 64 (context window), the detection accuracy improves, showcasing UNITE's ability to effectively capture temporal inconsistencies in fake videos." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.483, + 0.43 + ], + "angle": 0, + "content": "Effect on Synthetic Data: UNITE performed significantly better on the fully synthetic datasets GTA-V [24] and DeMamba [9] when the model was training includes GTA-V [24] data. However, for the DeMamba dataset (cross-data evaluation), the performance with only CE-loss is similar regardless of the training data. It is the AD-loss component that provides the accuracy boost in Fig 4(b)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.431, + 0.483, + 0.657 + ], + "angle": 0, + "content": "Effect on Background-Manipulated Data: In Fig. 4(a), we observe that when UNITE was trained solely on \\(\\mathrm{FF}++\\) [43] using only the CE-loss, its performance on the AVID dataset [61] was \\(0.00\\%\\). This outcome can be attributed to the fact that the fake data in \\(\\mathrm{FF}++\\) [43] contains exclusively face-manipulated videos, causing the UNITE transformer's attention heads to focus predominantly on the facial regions. As AVID [61] comprises only background-manipulated videos, UNITE failed to detect any of the fakes under such conditions. However, when the AD-loss was introduced, the attention heads of UNITE were encouraged to diversify their spatial focus, significantly improving performance. Incorporating synthetic data, as shown in Fig. 4(b), further enhanced this effect, achieving \\(100\\%\\) accuracy on AVID when using both CE-loss and AD-loss." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.869 + ], + "angle": 0, + "content": "Effect on Face-Manipulated Data: For the evaluation on face-manipulated datasets such as CelebDF [30] and UADFV [58], one might expect comparable results between training solely on \\(\\mathrm{FF} + +\\) [43] and incorporating GTA-V [24] due to their perceived irrelevance in evaluating face-centric manipulations. However, as shown in Fig. 4(b), training with fully synthetic data leads to a substantial performance improvement. Notably, in both Fig. 4(a) and (b), using only CE loss yields similar results, but the inclusion of the AD-loss component significantly enhances overall detection accuracy. The diversity in the training data makes the UNITE model learn more discriminative attention map features through the AD-loss component, which in turn helps even in detecting face-manipulated videos." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Frames vs. Performance: For inference with UNITE, we evaluated the impact of temporal context by vary-" + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.095, + 0.861, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.237, + 0.907, + 0.321 + ], + "angle": 0, + "content": "Figure 6. Transformer Depth Evaluation: Performance comparison of UNITE with varying the number of encoder blocks (depth). UNITE performs optimally in cross-domain settings when the depth is 4. Greater depths overfit to the training domains \\((\\mathrm{FF} + +\\) [43] and GTA-V [24]), while a depth of 2 is insufficient to capture the complexity of the data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.907, + 0.435 + ], + "angle": 0, + "content": "ing the number of frames sampled from video segments to \\(\\{1,8,16,32,64\\}\\), with 64 frames corresponding to our model's context window. As shown in Fig. 5, performance improves as more frames are included, demonstrating that UNITE effectively captures temporal discontinuities in fake videos. This highlights the model's ability to leverage temporal cues for enhanced detection accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.436, + 0.909, + 0.588 + ], + "angle": 0, + "content": "Transformer Depth Ablation: We evaluate the impact of varying the depth of the transformer architecture in UNITE by adjusting the number of encoder blocks, with results shown in Fig. 6. We tested depths of \\(\\{2,4,6,8\\}\\) and observed that the model achieved the best performance with 4 encoder blocks, which is the depth used in our final architecture. This suggests that a moderate depth strikes the optimal balance between model complexity and performance, especially in cross-dataset settings, providing robust detection capabilities without overfitting." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.602, + 0.634, + 0.618 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.906, + 0.749 + ], + "angle": 0, + "content": "Traditional DeepFake detectors, focused solely on face-manipulated content, struggle against newer, more sophisticated forgery methods involving T2V, I2V, and background manipulations. To address this, we proposed UNITE, leveraging the fully synthetic GTA-V [24] dataset for training. While this synthetic data is not AI-generated, our results show it effectively enhances UNITE's ability to detect AI-generated content and background manipulations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.87 + ], + "angle": 0, + "content": "A key innovation in our approach is the AD-loss, which encourages the model to explore diverse spatial regions, enhancing its detection of manipulated areas beyond just the face region. This allows UNITE to excel not only in detecting AI-generated and background-altered videos but also in identifying traditional face-manipulated content. Our ablation studies demonstrate the robustness of UNITE, making it a valuable tool in synthetic video detection." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Acknowledgements: This work was supported in part by funding from YouTube (Google)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28057" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.476, + 0.129 + ], + "angle": 0, + "content": "[1] Flux 1.1. https://blackforestlabs.ai/, 2024.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.131, + 0.483, + 0.157 + ], + "angle": 0, + "content": "[2] Morph studio. https://www.morphstudio.com/, 2024.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.161, + 0.483, + 0.229 + ], + "angle": 0, + "content": "[3] Darius Afchar, Vincent Nozick, Junichi Yamagishi, and Isao Echizen. Mesonet: a compact facial video forgery detection network. In 2018 IEEE international workshop on information forensics and security (WIFS), pages 1-7. IEEE, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.232, + 0.483, + 0.288 + ], + "angle": 0, + "content": "[4] Ibrahim M Alabdulmohsin, Xiaohua Zhai, Alexander Kolesnikov, and Lucas Beyer. Getting vit in shape: Scaling laws for compute-optimal model design. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.29, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[5] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.361, + 0.483, + 0.43 + ], + "angle": 0, + "content": "[6] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. 2024. URL https://openai.com/research/video-generation-models-as-world-simulators, 3, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.433, + 0.483, + 0.487 + ], + "angle": 0, + "content": "[7] Baoying Chen and Shunquan Tan. Featuretransfer: Unsupervised domain adaptation for cross-domain deepfake detection. Security and Communication Networks, 2021(1): 9942754, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.49, + 0.483, + 0.559 + ], + "angle": 0, + "content": "[8] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.562, + 0.483, + 0.63 + ], + "angle": 0, + "content": "[9] Haoxing Chen, Yan Hong, Zizheng Huang, Zhuoer Xu, Zhangxuan Gu, Yaohui Li, Jun Lan, Huijia Zhu, Jianfu Zhang, Weiqiang Wang, et al. Demamba: Ai-generated video detection on million-scale genvideo benchmark. arXiv preprint arXiv:2405.19707, 2024. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.633, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[10] Jikang Cheng, Zhiyuan Yan, Ying Zhang, Yuhao Luo, Zhongyuan Wang, and Chen Li. Can we leave deepfake data behind in training deepfake detector? arXiv preprint arXiv:2408.17052, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.69, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[11] Jongwook Choi, Taehoon Kim, Yonghyun Jeong, Seungryul Baek, and Jongwon Choi. Exploiting style latent flows for generalizing deepfake video detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1133-1143, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[12] Umur Aybars Ciftci, Ilke Demir, and Lijun Yin. Fakecatcher: Detection of synthetic portrait videos using biological signals. IEEE transactions on pattern analysis and machine intelligence, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[13] Sara Concas, Simone Maurizio La Cava, Roberto Casula, Giulia Orrù, Giovanni Puglisi, and Gian Luca Marcialis. Quality-based artifact modeling for facial deepfake detection in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3845–3854, 2024. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.174 + ], + "angle": 0, + "content": "[14] Riccardo Corvi, Davide Cozzolino, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. Intriguing properties of synthetic images: from generative adversarial networks to diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 973-982, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[15] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.263, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[16] Davide Cozzolino, Andreas Rössler, Justus Thies, Matthias Nießner, and Luisa Verdoliva. Id-reveal: Identity-aware deepfake video detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15108-15117, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.334, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[17] Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. The deepfake detection challenge (dfdc) dataset. arXiv preprint arXiv:2006.07397, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.391, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.462, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[19] Ricard Durall, Margret Keuper, Franz-Josef Pfundt, and Janis Keuper. Unmasking deepfakes with simple features. arXiv preprint arXiv:1911.00686, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[20] Zhihao Gu, Yang Chen, Taiping Yao, Shouhong Ding, Jilin Li, Feiyue Huang, and Lizhuang Ma. Spatiotemporal inconsistency learning for deepfake video detection. In Proceedings of the 29th ACM international conference on multimedia, pages 3473-3481, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.576, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[21] Ying Guo, Cheng Zhen, and Pengfei Yan. Controllable guide-space for generalizable face forgery detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20818-20827, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.633, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[22] Harald Hanselmann, Shen Yan, and Hermann Ney. Deep fisher faces. In BMVC, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.662, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[23] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.691, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[24] Yuan-Ting Hu, Jiahong Wang, Raymond A Yeh, and Alexander G Schwing. Sail-vos 3d: A synthetic dataset and baselines for object detection and 3d mesh reconstruction from video data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1418-1428, 2021. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[25] Liming Jiang, Ren Li, Wayne Wu, Chen Qian, and Chen Change Loy. Deeperforensics-1.0: A large-scale dataset for real-world face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2889–2898, 2020. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.847, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[26] Yan Ju, Chengzhe Sun, Shan Jia, Shuwei Hou, Zhaofeng Si, Soumyya Kanti Datta, Lipeng Ke, Riky Zhou, Anita Nikolich, and Siwei Lyu. Deepfake-o-meter v2.0: An open platform for deepfake detection. In 2024 IEEE 7th Inter-" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28058" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "national Conference on Multimedia Information Processing and Retrieval (MIPR), pages 439-445. IEEE, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.123, + 0.483, + 0.164 + ], + "angle": 0, + "content": "[27] Pavel Korshunov and Sébastien Marcel. Deepfakes: a new threat to face recognition? assessment and detection. arXiv preprint arXiv:1812.08685, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.167, + 0.482, + 0.208 + ], + "angle": 0, + "content": "[28] Lingzhi Li, Jianmin Bao, Hao Yang, Dong Chen, and Fang Wen. Faceshifter: Towards high fidelity and occlusion aware face swapping. arXiv preprint arXiv:1912.13457, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.211, + 0.482, + 0.239 + ], + "angle": 0, + "content": "[29] Y Li. Exposing deepfake videos by detecting face warping artifact acts. arXiv preprint arXiv:1811.00656, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.241, + 0.483, + 0.309 + ], + "angle": 0, + "content": "[30] Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3207-3216, 2020. 1, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.312, + 0.483, + 0.38 + ], + "angle": 0, + "content": "[31] Li Lin, Xinan He, Yan Ju, Xin Wang, Feng Ding, and Shu Hu. Preserving fairness generalization in deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16815-16825, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.384, + 0.483, + 0.452 + ], + "angle": 0, + "content": "[32] Baoping Liu, Bo Liu, Ming Ding, Tianqing Zhu, and Xin Yu. Ti2net: temporal identity inconsistency network for deepfake detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4691-4700, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.455, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[33] I Loshchilov and F Hutter. Decoupled weight decay regularization. *ICLR*, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.485, + 0.482, + 0.553 + ], + "angle": 0, + "content": "[34] Qingxuan Lv, Yuezun Li, Junyu Dong, Sheng Chen, Hui Yu, Huiyu Zhou, and Shu Zhang. Domainforensics: Exposing face forgery across domains via bi-directional adaptation. IEEE Transactions on Information Forensics and Security, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.557, + 0.482, + 0.612 + ], + "angle": 0, + "content": "[35] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.615, + 0.482, + 0.671 + ], + "angle": 0, + "content": "[36] Ghazal Mazaheri and Amit K Roy-Chowdhury. Detection and localization of facial expression manipulations. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1035-1045, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.673, + 0.481, + 0.699 + ], + "angle": 0, + "content": "[37] moonvalley.ai. moonvalley.ai. https://moonvalley.ai/, 2022.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.703, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[38] John Mullan, Duncan Crawbuck, and Aakash Sastry. Hotshot-XL, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[39] Dat Nguyen, Nesryne Mejri, Inder Pal Singh, Polina Kuleshova, Marcella Astrid, Anis Kacem, Enjie Ghorbel, and Djamila Aouada. Laa-net: Localized artifact attention network for quality-agnostic and generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17395–17405, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[40] Alvaro Lopez Pellicer, Yi Li, and Plamen Angelov. Pudd: Towards robust multi-modal prototype-based deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3809-3817, 2024. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[41] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In European conference on computer vision, pages 86-103. Springer, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.904, + 0.177 + ], + "angle": 0, + "content": "[42] Runway Research. Text driven video generation. https://research.runwayml.com/gen2, 2023.1,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.18, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[43] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics++: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1-11, 2019. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.251, + 0.905, + 0.306 + ], + "angle": 0, + "content": "[44] Md Shamim Seraj, Ankita Singh, and Shayok Chakraborty. Semi-supervised deep domain adaptation for deepfake detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1061-1071, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.308, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[45] Ke Sun, Shen Chen, Taiping Yao, Hong Liu, Xiaoshuai Sun, Shouhong Ding, and Rongrong Ji. Diffusionfake: Enhancing generalization in deepfake detection via guided stable diffusion. NeurIPS, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.907, + 0.433 + ], + "angle": 0, + "content": "[46] Zekun Sun, Yujie Han, Zeyu Hua, Na Ruan, and Weijia Jia. Improving the efficiency and robustness of deepfakes detection through precise geometric features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3609-3618, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[47] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28130-28139, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[48] The New York Times. A.i. can now create lifelike videos. can you tell what's real? https://www.nytimes.com/interactive/2024/09/09/technology/ai-video-deepfake-runway-kling-quiz.html, 2024.5,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[49] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG), 38(4):1-12, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[50] Loc Trinh, Michael Tsang, Sirisha Rambhatla, and Yan Liu. Interpretable and trustworthy deepfake detection via dynamic prototypes. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1973-1983, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.719, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[51] A Vaswani, N Shazeer, N Parmar, J Uszkoreit, L Jones, A Gomez, L Kaiser, and I Polosukhin. Attention is all you need. NeurIPS, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[52] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[53] Yuhan Wang, Xu Chen, Junwei Zhu, Wenqing Chu, Ying Tai, Chengjie Wang, Jilin Li, Yongjian Wu, Feiyue Huang, and Rongrong Ji. Hififace: 3d shape and semantic prior guided high fidelity face swapping. Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, 2021. 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28059" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[54] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[55] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22445-22455, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[56] Deressa Wodajo and Solomon Atnafu. Deepfake video detection using convolutional vision transformer. arXiv preprint arXiv:2102.11126, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.345 + ], + "angle": 0, + "content": "[57] Yuting Xu, Jian Liang, Gengyun Jia, Ziming Yang, Yanhao Zhang, and Ran He. Tall: Thumbnail layout for deepfake video detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22658-22668, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.347, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[58] Xin Yang, Yuezun Li, and Siwei Lyu. Exposing deep fakes using inconsistent head poses. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 8261-8265. IEEE, 2019. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.404, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[59] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.46, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[60] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. International Journal of Computer Vision, pages 1–15, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.531, + 0.482, + 0.613 + ], + "angle": 0, + "content": "[61] Zhixing Zhang, Bichen Wu, Xiaoyan Wang, Yaqiao Luo, Luxin Zhang, Yinan Zhao, Peter Vajda, Dimitris Metaxas, and Licheng Yu. Avid: Any-length video inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7162-7172, 2024. 1, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.615, + 0.482, + 0.683 + ], + "angle": 0, + "content": "[62] Cairong Zhao, Chutian Wang, Guosheng Hu, Haonan Chen, Chun Liu, and Jinhui Tang. Istvt: interpretable spatial-temporal video transformer for deepfake detection. IEEE Transactions on Information Forensics and Security, 18: 1335-1348, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.685, + 0.482, + 0.74 + ], + "angle": 0, + "content": "[63] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.742, + 0.482, + 0.797 + ], + "angle": 0, + "content": "[64] Xinye Zhou, Hu Han, Shiguang Shan, and Xin Chen. Fine-grained open-set deepfake detection via unsupervised domain adaptation. IEEE Transactions on Information Forensics and Security, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.797 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "28060" + } + ] +] \ No newline at end of file diff --git a/2025/Towards a Universal Synthetic Video Detector_ From Face or Background Manipulations to Fully AI-Generated Content/49360ca2-47e3-4c7b-81b1-6a87621996c4_origin.pdf b/2025/Towards a Universal Synthetic Video Detector_ From Face or Background Manipulations to Fully AI-Generated Content/49360ca2-47e3-4c7b-81b1-6a87621996c4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3b56e13663f4e942cb7813df676547757cef6b5a --- /dev/null +++ b/2025/Towards a Universal Synthetic Video Detector_ From Face or Background Manipulations to Fully AI-Generated Content/49360ca2-47e3-4c7b-81b1-6a87621996c4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:353eb54ebe234c165fb2f8fff55dfd8e415c61627cd214944b17d4c881902ecd +size 8006347 diff --git a/2025/Towards a Universal Synthetic Video Detector_ From Face or Background Manipulations to Fully AI-Generated Content/full.md b/2025/Towards a Universal Synthetic Video Detector_ From Face or Background Manipulations to Fully AI-Generated Content/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c3345fb4a7e213c39877d0ea087f8e29757f609e --- /dev/null +++ b/2025/Towards a Universal Synthetic Video Detector_ From Face or Background Manipulations to Fully AI-Generated Content/full.md @@ -0,0 +1,337 @@ +# Towards a Universal Synthetic Video Detector: From Face or Background Manipulations to Fully AI-Generated Content + +Rohit Kundu $^{1,2}$ , Hao Xiong $^{1}$ , Vishal Mohanty $^{1}$ , Athula Balachandran $^{1}$ , Amit K. Roy-Chowdhury $^{2}$ , $^{1}$ Google LLC, $^{2}$ University of California, Riverside + +{rohit.kundu@email, amertc@ece}.ucr.edu; {rohitkun, haoxg, vishal, athula}@google.com + +# Abstract + +Existing DeepFake detection techniques primarily focus on facial manipulations, such as face-swapping or lip-syncing. However, advancements in text-to-video (T2V) and image-to-video (I2V) generative models now allow fully AI-generated synthetic content and seamless background alterations, challenging face-centric detection methods and demanding more versatile approaches. + +To address this, we introduce the Universal Network for Identifying Tampered and synthEtic videos (UNITE) model, which, unlike traditional detectors, captures full-frame manipulations. UNITE extends detection capabilities to scenarios without faces, non-human subjects, and complex background modifications. It leverages a transformer-based architecture that processes domain-agnostic features extracted from videos via the SigLIP-So400M foundation model. Given limited datasets encompassing both facial/background alterations and T2V/I2V content, we integrate task-irrelevant data alongside standard DeepFake datasets in training. We further mitigate the model's tendency to over-focus on faces by incorporating an attention-diversity (AD) loss, which promotes diverse spatial attention across video frames. Combining AD loss with cross-entropy improves detection performance across varied contexts. Comparative evaluations demonstrate that UNITE outperforms state-of-the-art detectors on datasets featuring face/background manipulations and fully synthetic T2V/I2V videos, showcasing its adaptability and generalizable detection capabilities. + +# 1. Introduction + +The rise of synthetic media, especially DeepFakes, has transformed content perception and interaction online. Hyper-realistic images produced by technologies like $\mathrm{FLUX}_{1.1}$ [1] are challenging for even humans to identify as fake, and with the development of their video model underway, this underscores the critical need for effective detec + +![](images/7f4d537dbea8ce1cfa0f687526832537988d457109714d1137feb379a094bc2b.jpg) +Figure 1. Problem Overview: Existing DeepFake detection methods focus on identifying face-manipulated videos, most of which cannot perform inference unless there is a face detected in the video. However, with advancements like seamless background modifications (e.g., AVID [61]) and hyper-realistic content from games like GTA-V [24] and T2V/I2V models [9], a more comprehensive approach is needed. A model trained with only cross-entropy (CE) loss, using full frames, automatically focuses on the face, capturing temporal discontinuities through its transformer architecture, performing better than random $(\approx)$ on T2V/I2V content but struggling with background manipulations. UNITE, with its attention-diversity (AD) loss, effectively detects both face/background manipulations and fully synthetic content. + +tion methods for media generated by deep neural networks. Traditional DeepFake generators [28, 49] focus on manipulating human faces through face-swapping and lip-syncing. However, powerful text-to-video (T2V) and image-to-video (I2V) models [42, 52, 63] have expanded manipulation possibilities beyond faces. + +While conventional detectors [10, 45, 57] perform well on older, face-centric DeepFake datasets [17, 30, 43], they often struggle with newer manipulations involving full scenes or backgrounds. DeepFake-O-Meter [26] is an open-source tool (consisting of several state-of-the-art models) for DeepFake detection, but it cannot run inference unless a human face is visible in the image/video. The rapid spread + +of misinformation, particularly during critical periods such as elections, highlights the need for generalizable detection models capable of identifying diverse manipulations, including face, background, and fully AI-generated T2V/I2V content with/without human subjects (overview in Fig. 1). + +To address these challenges, we present a Universal Network for Identifying Tampered and synthEtic videos or UNITE model, which detects both partially manipulated (foreground/background) and fully synthetic videos. Unlike detectors focused solely on face detection, UNITE analyzes entire frames, regardless of whether a human subject is present in the videos. + +Given the inherent domain gaps in DeepFake datasets [44, 64], even when generated by similar techniques, we leverage the SigLIP-So400m [4] foundation model to extract domain-agnostic features. This serves as inputs to a learnable transformer with multi-head attention, enabling effective detection by capturing temporal inconsistencies in synthetic content. However, preliminary experiments show that training a transformer architecture solely with cross-entropy (CE) loss often leads to the attention heads converging on the face region (Fig. 3). As a result, the model struggles during inference when handling videos where a real human subject is placed in a manipulated background or content generated by T2V [8, 52] and I2V [35, 63] models. To address this, we introduce an "attention-diversity" (AD) loss that encourages attention heads to focus on different spatial regions, enhancing the model's ability to capture critical cues from both foreground and background. + +Due to the limited availability of open-source datasets covering face or background manipulations and fully AI-generated content, we employ innovative training strategies for UNITE. This includes integrating task-irrelevant data with standard DeepFake datasets to simulate AI-generated synthetic media. Beyond the popularly used FaceForensics++ [43] dataset, we utilize the SAIL-VOS-3D [24] dataset, originally designed for 3D video object segmentation in the game GTA-V. As this dataset is fully synthetic, it helps simulate AI-generated media, enhancing our model's ability to detect diverse forms of synthetic manipulation. + +Our contributions can be summarized as follows: + +- We propose UNITE, a model for detecting partially manipulated (foreground/background) and fully synthetic videos, moving beyond face-centric DeepFake detection. +- Unlike detectors relying on face detection and cropping, our model can process full video frames and can detect fakes even if there are no human subjects. +- Using the SigLIP-So400m foundation model [4], we extract domain-agnostic features, enabling generalization to in-the-wild DeepFakes. +- We introduce an attention-diversity loss, encouraging attention heads to focus on diverse spatial regions, enhancing detection beyond face manipulations. + +- Unlike existing methods that evaluate on a few datasets, we comprehensively assess UNITE on a broad range of face and synthetic datasets with various T2V/I2V generators, outperforming detection in foreground, background, and T2V/I2V manipulations. + +# 2. Related Work + +Face-centric DeepFake Detection: Methods such as [13, 19, 40] focus on modeling spatial inconsistencies for DeepFake detection. Concas et al. [13] track specific facial features (eyes, nose, mouth) in a high-frequency domain, as fake videos exhibit distinct behaviors in this space compared to real ones. However, due to their dataset-centric architectures, it is not a generalizable approach. PUDD [40] learns person-specific prototypes by modeling patterns from real videos and assessing how closely inference videos match these prototypes, making it less effective for detecting in-the-wild DeepFakes. LAA-Net [39] attempts to improve generalizability through a multi-task attention module that detects artifacts via heatmap and self-consistency regression, but it works by analyzing every still frame of a video, thus limiting its scalability. The authors in [36] tackled the challenging problem of detecting and localizing expression swaps, where the number of manipulated pixels is even fewer compared to identity swaps. Although the authors obtained good localization performance, their reliance on a CNN-based architecture limits their ability to capture temporal inconsistencies effectively. + +DPNet [50] and ID-Reveal [16] are identity-aware temporal artifact modeling methods: DPNet [50] focuses on interpretable, prototype-based detection of unnatural movements through dynamic feature representations, while IDReveal [16] leverages metric learning on real data to detect biometric inconsistencies. However these methods assume access to authentic reference videos, making them unfit for in-the-wild DeepFake detection. Shifting from identity-specific methods, TALL [57] constructs composite "thumbnail" images from four consecutive frames taken at random timestamps for temporal analysis. TI2Net [32] captures temporal anomalies by subtracting consecutive frame features but relies on the assumption that FaceSwap manipulations show pronounced inter-frame discrepancies. + +Choi et al. [11] approach DeepFake detection by targeting suppressed variance in style-based latent temporal features via a StyleGRU module. However, this method is limited to cropped face regions, making it ineffective for T2V/I2V or background-manipulated videos. + +Synthetic Video Detection: While there has been considerable effort in synthetic image detection [14, 15, 55], synthetic video detection is a scarcely explored area. DeMamba [9], a recently proposed synthetic video detector, analyzes small zones of video frames to capture how pixels change spatiotemporally. Its continuous scanning approach + +![](images/293854bf54d05f4417715994a0c7ca60da622532c26dfd731035498dd8ac84a2.jpg) +Figure 2. UNITE architecture overview: We extract domain-agnostic features $(\xi)$ using the SigLIP-So400m foundation model [4] to mitigate domain gaps between DeepFake datasets (Sec. 3.2). These embeddings, combined with positional encodings, are input to a transformer with multi-head attention and MLP layers (Sec. 3.3), culminating in a classifier for final predictions. AD-loss (Sec. 3.4) encourages the model attention to span diverse spatial regions. + +better tracks subtle changes and patterns, making it easier to spot manipulated or fake content. The authors also propose a million scale T2V/I2V dataset (which we call the DeMamba dataset hereforth) to evaluate their performance. However, the method was not evaluated on typical DeepFake datasets [30, 43, 53] and thus is not fit for human-centric DeepFake detection. In contrast our UNITE model is built to detect partially manipulated as well as fully AI-generated videos. A key reason for the scarcity of synthetic video detection methods in the literature is the lack of standardized datasets for training and evaluation. + +# 3. Proposed Method + +In this section we first describe our problem setup (Sec. 3.1), then we describe the foundation model-based video encoding (Sec. 3.2), our trainable transformer architecture (Sec. 3.3) and finally the novel loss function for training the UNITE model (Sec. 3.4). The overview of the UNITE architecture is shown in Fig. 2. + +# 3.1. Problem Setup + +For each video in the dataset, we implement a frame sampling strategy where every other frame is extracted. From the resulting collection of frames, we construct video segment samples consisting of $n_f = 64$ consecutive frames, which are defined as a single data sample, denoted as "v". This selection aligns with the context window established for our video transformer model, ensuring that each sample retains the temporal coherence necessary for effective train- + +ing. In instances where the video duration results in fewer than $n_f$ frames, we apply padding (ablation in supplementary) to ensure that each sample remains uniform in size. + +Given that the original video-level classification labels are known and fake videos contain manipulations throughout their entirety (i.e., no videos have manipulations limited to specific frames), each $n_f$ -frame segment generated from a given video is assigned the same label (denoted by $l \in 0, \dots, n_c$ ) as the source video, with number of possible classes as $n_c$ . The aggregation of these samples across all videos forms our final dataset, $v_i, l_{i=1}^N \in \mathcal{V}$ , where $N$ is the total number of samples obtained. This approach serves as a data augmentation technique, increasing the number of training samples without additional data collection. + +# 3.2. Domain-Agnostic Feature Extraction + +Our objective is to leverage the trained UNITE model for detecting in-the-wild DeepFakes, necessitating careful consideration of the domain gap [7, 34] between the train and test datasets. To effectively mitigate the impact of domain discrepancies, we utilize the SigLIP foundation model [59] to extract domain-agnostic features. Specifically, we employ the "shape-optimized" ViT [4] image encoder, whose architecture is rooted in [18]. The SigLIP-So400m [4] model was contrastively pretrained on 3B diverse examples using a sigmoid loss, enabling it to capture robust, generalizable features while maintaining a compact size (400M parameters). This extensive pretraining makes SigLIP particularly adept at extracting domain-invariant representations, crucial for handling the significant variability encountered in DeepFake datasets and real-world scenarios. + +For every frame $f_{j}$ (resized to $\mathbb{R}^{384\times 384\times 3}$ ), of a video sample $v_{i}\in \mathcal{V}$ , we obtain an image encoding from the frozen SigLIP-So400m [4] model as $e_j = SigLIP(f_j)\in \mathbb{R}^{t_s\times d_s}$ , where $j\in \{1,2,\dots ,n_f\}$ , since there are $n_f$ frames per video sample (as explained in Sec. 3.1), $t_s = 729$ represents the token length and $d_{s} = 1152$ represents the feature dimension. We concatenate these frame-level features for the video sample $v_{i}$ , while maintaining the order of the frames, to obtain a video segment encoding denoted by $\xi_{i}\in \mathbb{R}^{n_{f}\times t_{s}\times d_{s}}$ , which are used as input to our trainable transformer architecture. This encoded video dataset can be denoted as $\{\xi_i,l_i\}_{i = 1}^N\in \mathcal{V}_\xi$ . + +# 3.3. Learnable Transformer Architecture + +The UNITE architecture is a multi-head self-attention (MHSA) based transformer model designed specifically for video classification tasks. It leverages the strengths of Transformer networks to process sequences of video frame embeddings effectively, allowing for robust classification performance. The UNITE architecture allows for adjustable depth, signifying the number of encoder blocks stacked within the model. A deeper architecture can capture more + +intricate features and dependencies at the cost of increased computational complexity. For our architecture the depth is set to 4. Ablation experiments with different transformer depths are provided in Sec. 4.4. + +# 3.3.1 Encoder Block + +Each encoder block consists of the following components: + +Multi-Head Self-Attention Layer: This layer allows the model to focus on different parts of the input sequence simultaneously. With 12 attention heads in our architecture, it captures diverse interactions and dependencies among frames. Each head computes attention scores independently, learning multiple representations of the input data. This enhances the model's ability to detect subtle temporal variations, which is crucial for video classification. The use of scaled dot-product attention further helps mitigate large gradients and stabilizes training. + +Layer Normalization and Residual Connections: To aid gradient flow during training, each sub-layer incorporates residual connections that add the input back to the output, mitigating vanishing gradient issues in deep networks. Layer normalization follows to stabilize activations and enhance convergence speed. Additionally, dropout is applied post-attention to reduce overfitting by randomly deactivating a portion of neurons during training. + +Feed-Forward Network (MLP): The second sub-layer is a point-wise feed-forward network with two dense layers separated by a GELU activation [23]. This MLP enables nonlinear feature transformations, enhancing the model's expressiveness. By projecting attention outputs into a higher-dimensional space before reducing them, it captures complex interactions beyond linear transformations. + +Attention maps are derived from the outputs of the encoder blocks to be used for computing the Attention-Diversity loss (Sec. 3.4) and to illustrate how the model prioritizes different spatial regions in the frames during the classification process (Fig. 3). This interpretability is valuable for understanding the model's decision-making, particularly in complex video classification scenarios. + +# 3.3.2 Positional Encoding + +We utilize a sine-cosine positional encoding scheme following the original transformer [51] to provide a unique position identifier for each input token. For every frame $f_{j}$ in the video sample with encoded feature dimension $d_{s}$ (as per Sec. 3.2), we compute the positional encoding for odd and even indexed feature dimensions as, + +$$ +P E _ {(j, 2 i + 1)} = \cos \left(\frac {j}{1 0 0 0 0 ^ {\frac {2 i}{d _ {s}}}}\right), \quad P E _ {(j, 2 i)} = \sin \left(\frac {j}{1 0 0 0 0 ^ {\frac {2 i}{d _ {s}}}}\right), \tag {1} +$$ + +where, $i$ denotes the feature index in the encoded feature dimension $d_{s}$ . We embed these encodings directly into the input tokens of our video transformer to provide positional context, for better temporal modeling during the attention computation with minimal computational overhead. + +# 3.4. Attention-Diversity Loss + +Training a multi-head attention transformer with only cross-entropy loss makes the attention maps focus only on the face regions of the frame (as evident in Fig. 3). However, we aim to detect fake videos where the manipulation might not be in the face at all, and instead be in the background (like video inpainting with the AVID model [61]). To ensure that the attention heads focus on diverse spatial regions of the video frames, we devise the "Attention-Diversity" (AD) loss. + +AD-loss is designed to minimize overlap among attention maps (obtained from the first encoder block of the trainable transformer architecture in 3.3.1) while maintaining consistency across different input video samples. The attention outputs from the video transformer model $\mathcal{A} \in \mathbb{R}^{n_h \times t_s \times d_s}$ are used to pool the input SigLIP-So400m features $\xi \in \mathbb{R}^{n_f \times t_s \times d_s}$ , resulting in a pooled feature tensor $\mathcal{P} \in \mathbb{R}^{n_h \times n_f}$ according to, + +$$ +\mathcal {P} = \sum_ {j = 1} ^ {t _ {s}} \sum_ {k = 1} ^ {d _ {s}} \mathcal {A} _ {h, j, k} \cdot \xi_ {f, j, k}, \tag {2} +$$ + +where $f$ indexes the frames $(n_f)$ , $h$ indexes the number of attention heads $(n_h)$ , and $j, k$ are spatial positions. + +We compute "feature centers" which are points in the feature space that serve as anchors for the learned representations of different classes. These centers represent the average features of samples from a particular class, allowing the model to capture essential characteristics of that class. In the context of the AD-loss, feature centers $\mathcal{C} \in \mathbb{R}^{n_h \times n_f}$ are dynamically updated in each training iteration $\tau$ based on the pooled feature vectors $\mathcal{P}$ , following, + +$$ +\mathcal {C} ^ {\tau} = \mathcal {C} ^ {\tau - 1} - \eta \left(\mathcal {C} ^ {\tau - 1} - \frac {1}{B} \sum_ {b = 1} ^ {B} \mathcal {P} _ {b}\right), \tag {3} +$$ + +where $\eta$ represents the learning rate (set to 0.05 in our experiments) for feature center updates and $B$ is the batch dimension. The feature centers are initialized with zeros $(\mathcal{C}^0 = [0]_{n_h\times n_f})$ at the beginning of the model training. + +AD-loss consists of two distinct components: within-class loss and between-class loss inspired by the Fisher Discriminant Analysis for deep networks [22]. The within-class term calculates the distance between the pooled feature vectors $\mathcal{P}$ and their corresponding feature centers $\mathcal{C}$ , promoting closeness among similar samples. The between-class term measures the distance between feature centers of different classes, encouraging them to be spaced apart in the feature space, encouraging separability. + +The within-class loss term is calculated as + +$$ +\mathcal {L} _ {\text {w i t h i n}} = \max \left(\| \mathcal {P} - \mathcal {C} \| _ {2} - \delta_ {\text {w i t h i n}}, 0\right), \tag {4} +$$ + +where $\delta_{\mathrm{within}} \in \mathbb{R}^{n_c}$ is a hyperparameter controlling the allowable distance between feature vectors and their respective feature centers for inputs of the same class, encouraging tighter clustering and $\| \cdot \|_2$ represents $L_2$ normalization. + +The between-class loss is calculated as + +$$ +\mathcal {L} _ {\text {b e t w e e n}} = \sum_ {\substack {k \neq l \\ (k, l) \in \left(n _ {h}, n _ {h}\right)}} \max \left(\delta_ {\text {b e t w e e n}} - \| \mathcal {C} _ {k} - \mathcal {C} _ {l} \| _ {2}, 0\right), \tag{5} +$$ + +where $\delta_{\mathrm{between}}$ is a predefined hyperparameter that ensures a minimum distance between feature centers of different classes. Thus the AD-loss is a simple addition of these two components $\mathcal{L}_{AD} = \mathcal{L}_{\mathrm{within}} + \mathcal{L}_{\mathrm{between}}$ , making the final objective function for training the UNITE model, + +$$ +\mathcal {L} _ {\text {U N I T E}} = \lambda_ {1} \cdot \mathcal {L} _ {C E} + \lambda_ {2} \cdot \mathcal {L} _ {A D}, \tag {6} +$$ + +where $\mathcal{L}_{CE}$ is the traditional cross-entropy loss, and $\lambda_{1}$ and $\lambda_{2}$ are loss-weighting hyperparameters. + +# 4. Experiments + +# 4.1. Datasets + +Most DeepFake datasets focus on face manipulations, such as face-swaps or lip-syncing, and lack fully AI-generated video content. The DeMamba dataset [9] features videos from T2V/I2V models, including powerful diffusion models like SORA [5]. It is not a human-centric dataset and primarily contains videos from natural scenes. We use its validation split to evaluate UNITE. + +To train UNITE, we employ the FaceForensics++ (FF++) [43] dataset (c23 i.e., medium-quality) and the SAIL-VOS-3D dataset [24], which includes synthetic (although not AI-generated) GTA-V game videos featuring human subjects suitable for DeepFake detection. To evaluate the UNITE model, we use: + +- Face manipulated data: $\mathrm{FF}++$ [43] (in-domain evaluation), CelebDF [30], DeeperForensics [25], Deepfake-TIMIT [27], HifiFace [53], UADFV [58]. +- Background manipulated data: Sample videos from the AVID [61] model provided publicly by the authors. +- Fully synthetic data: GTA-V [24] (in-domain evaluation), DeMamba [9]. +- In-the-wild DeepFakes: Publicly available videos from the New York Times DeepFake quiz [48]. + +# 4.2. Results and Discussion + +Evaluation Metrics: Although most DeepFake detection methods only report accuracy, we evaluate UNITE using additional metrics: (1) AUC (Area Under the Precision-Recall + +Curve); (2) Precision@0.5 and (3) Recall@0.5 at a 0.5 confidence threshold; (4) Precs@Rec=0.8 (precision at 0.8 recall); (5) Rec@Precs=0.8 (recall at 0.8 precision). + +Training Details: UNITE is trained using an AdamW optimizer [33] with an initial learning rate of 0.0001 and a decay rate of 0.5 every 1000 steps. For AD-loss, $\delta_{\mathrm{between}} = 0.5$ and $\delta_{\mathrm{within}}$ is set to $[0.01, -2]$ for binary and $[0.01, -2, 1]$ for fine-grained classification in Sec. 4.3 (see supplementary for sensitivity). Loss weights are $\lambda_1 = \lambda_2 = 0.5$ . Training uses a batch size of 32 for 25 epochs on 8 TPUv3 chips, with the framework implemented in TensorFlow. + +Quantitative Results: The results obtained by the UNITE model when trained on $\mathrm{FF}++$ [43] alone, and when trained with both $\mathrm{FF}++$ [43] and GTA-V [24] data and evaluated on the various datasets mentioned in Sec. 4.1 are shown in Table 1. On the AVID [61] and DeMamba datasets [9], the model performs poorly when trained with only face manipulated data, but the performance enhances several-fold when the GTA-V [24] synthetic data, even though it is not AI-generated, is used in training. + +Interestingly, when evaluating on face-manipulated datasets such as CelebDF [30] and DeeperForensics [25], we observe a performance boost, particularly for CelebDF [30], when GTA-V [24] data is included in the training set. Intuitively, one would expect that training solely on $\mathrm{FF}++$ [43] would yield similar results since the inference set consists of similar face-manipulated data. This deviation from the expected behavior comes from the contribution of the AD-Loss, as the ablation study (Fig. 4) using only CE-loss compared to both CE and AD-losses reveals that the performance is enhanced with GTA-V [24] data in training only in the latter (detailed discussion in Sec. 4.4). + +To evaluate UNITE on in-the-wild DeepFakes, we attempted the recent New York Times quiz [48] (denoted by NYTimes) which provides 10 videos (4 real and 6 fake) for testing DeepFake detectors. UNITE got 8 of those 10 videos correct, even though some of the face-swap videos were indistinguishable from real videos, even by humans. On all the AI-generated videos, UNITE made the correct prediction (when trained on $\mathrm{FF} + +$ [43] and GTA-V [24]). + +Attention Heatmaps: To analyze the impact of CE and AD-losses on attention distribution, we extracted attention features from the first encoder of the UNITE model, comparing models trained with CE loss alone versus that trained with both CE and AD-losses. The resulting heatmaps, shown in Fig. 3, reveal that the model trained with only CE loss tends to concentrate primarily on the face region. In contrast, the model trained with the combined $\mathrm{CE + AD}$ loss demonstrates a broader attention span across the frame, as evidenced by a lighter, more distributed bluish tone in the heatmaps, indicating increased spatial diversity in the model's focus. This is why on the example shown in the first row (face-manipulated video) both models give the correct + +Table 1. Results from the UNITE model trained with (1) FF++ [43] only and (2) FF++ [43] combined with GTA-V [24]. All other results reflect cross-dataset evaluations except for FF++ and GTA-V (when trained). Performance gains are highlighted in green. + +
TrainTest
FF++GTA-VDatasetAccuracyAUCPrecision@0.5Recall@0.5Precs@Rec=0.8Rec@Precs=0.8
Face Manipulated Data
✓FF++99.53%99.77%99.94%99.49%99.94%99.94%
✓CelebDF72.61%94.05%96.45%61.22%80.45%61.22%
✓DeeperForensics91.35%100.00%100.00%91.35%100.00%91.35%
✓DeepfakeTIMIT86.90%86.46%83.61%83.97%88.90%81.33%
✓HiFiFace63.63%62.47%67.12%63.63%59.30%63.63%
✓UADFV94.12%94.38%95.68%97.11%93.79%94.38%
✓✓FF++99.96%(+0.43)99.89%(+0.12)100.00%(+0.06)99.84%(+0.35)100.00%(+0.06)99.96%(+0.02)
✓✓CelebDF95.11%(+22.50)94.36%(+0.31)96.82%(+0.37)68.75%(+7.53)96.53%(+16.08)68.75%(+7.53)
✓✓DeeperForensics99.62%(+8.27)100.00%(+0.00)100.00%(+0.00)99.62%(+8.27)100.00%(+0.00)99.63%(+8.28)
✓✓DeepfakeTIMIT91.90%(+5.00)91.33%(+4.87)90.45%(+6.84)88.39%(+4.42)100.00%(+11.10)91.95%(+10.62)
✓✓HiFiFace75.62%(+11.99)81.24%(+18.77)79.55%(+12.43)71.71%(+8.08)75.62%(+16.32)72.47%(+8.84)
✓✓UADFV97.01%(+2.89)94.95%(+0.57)96.89%(+1.21)100.00%(+2.89)94.12%(+0.33)100.00%(+5.62)
Background Manipulated Data
✓AVID41.67%33.33%33.33%41.67%41.67%33.33%
✓✓AVID100.00%(+58.33)100.00%(+66.67)100.00%(+66.67)100.00%(+58.33)100.00%(+58.33)100.00%(+66.67)
Fully Synthetic Data
✓GTA-V60.16%61.52%60.16%58.73%63.29%58.73%
✓DeMamba61.47%57.38%67.73%33.01%62.15%54.16%
✓✓GTA-V100.00%(+39.84)100.00%(+38.48)100.00%(+39.84)100.00%(+41.27)100.00%(+36.71)100.00%(+41.27)
✓✓DeMamba87.12%(+25.65)93.75%(+36.67)92.76%(+25.03)89.60%(+56.59)89.81%(+27.66)92.12%(+37.96)
In-the-wild DeepFakes
✓NYTimes [48]50.00%53.74%50.00%25.00%0.00%0.00%
✓✓NYTimes [48]80.00%(+30.00)97.42%(+43.68)83.33%(+33.33)83.33%(+58.33)60.00%(+60.00)100.00%(+100.00)
+ +Table 2. SOTA Comparison on Face-Manipulated Data: We compared the performance of UNITE with recent DeepFake detectors, in terms of detection accuracy on various face manipulated datasets. UNITE outperforms the existing methods. Bold shows the current best results and the previous best and second-best results are highlighted in red and blue respectively. + +
MethodFF++CelebDFDeeperForensicsUADFV
TALL [57]98.65%90.79%99.62%-
ISTVT [62]99.00%84.10%98.60%-
Concas et al. [13]99.49%---
PUDD [40]-95.10%--
TI2Net [32]99.95%68.22%76.08%-
LRNet [46]99.89%53.20%56.77%-
Choi et al. [11]-89.00%99.00%-
Lin et al. [31]98.28%74.42%--
Guo et al. [21]99.24%84.97%--
Li et al. (Res-152) [29]---93.80%
HeadPose [58]---89.00%
CViT [56]93.00%--93.75%
FakeCatcher [12]94.65%91.50%--
MesoNet [3]---82.40%
UNITE (Ours)99.96%95.11%99.62%97.01%
+ +prediction ("fake") with high confidence, but on the row-2 example (video generated from the Sora [6] T2V model) the model trained with only CE loss predicted "real" with a confidence score of $99.25\%$ (which is wrong), but the model trained with $\mathrm{CE + AD}$ losses predicted "fake" with a confidence of $100.00\%$ , which is correct. + +Comparison on Face Manipulated Data: UNITE is compared with recent state-of-the-art (SOTA) DeepFake detection models in Table 2. The methods we compared against were specifically designed for detecting face manipulations and often crop faces from the videos, during both training and inference of the models. However, UNITE, which is designed to detect a diverse range of fake videos, still outperforms these detectors. + +Comparison on Synthetic Data: We evaluated UNITE against state-of-the-art synthetic video detectors on the DeMamba dataset [9] (validation split), with results in Table 3. While existing detectors were trained on the DeMamba train split and validated on the DeMamba validation split, + +![](images/c22e4507b864d24b58be75ce23f856e165498a1d17c85f7886055912690090c3.jpg) +(a) Original Frame + +![](images/9a7f55b96974cf7b358b430d9ef5d44c715dd9f04ae58d9d55339da946fc69ed.jpg) +(b) CE Loss only + +![](images/a73dc7545cdac92a63c283887cef990df632ac6489198467c2b382e0b2d6c15f.jpg) +(c) $\mathrm{CE + AD}$ Loss + +![](images/66b04e8a8f21be849327484594cfd6b4ecf64172824ab28252ce8e38960c5b84.jpg) +(d) Original Frame + +![](images/c6f70e908b9eb1083ccd7c19c4d8aafcce441fe1ac95d4a3093f9532bc7b1cf2.jpg) +(e) CE Loss only + +![](images/f78bfdc6431dce203fc1738fdbc24287c4c17126ed02ee3507b3c365f47ad81c.jpg) +(f) $\mathrm{CE + AD}$ Loss +Figure 3. Comparison of attention heatmaps when UNITE is trained with only CE loss (second column) versus CE and AD-losses (third column). The CE-loss heatmap predominantly focuses on the face region, whereas the $\mathrm{CE + AD}$ loss heatmap demonstrates more distributed attention across the entire frame, as indicated by a broader bluish tone in the heatmaps. Both samples are "fake" videos- (a) is taken from Celeb-DF [30] and (d) is taken from DeMamba [9] and generated by OpenAI's Sora [6] (Best viewed as GIFs provided in the supplementary material). Yellow to blue shades show increasing attention in the plot. + +UNITE was trained on FF++ [43] and GTA-V [24]. Despite not being trained on DeMamba, UNITE's average performance (including real videos) surpassed current SOTA detectors, with competitive generator-wise results. + +Thus, UNITE reliably detects a diverse range of fake videos (even in cross-domain settings), including T2V/I2V videos and background manipulations, while consistently maintaining state-of-the-art performance on traditional face-manipulated data. UNITE eliminates the need to have separate DeepFake and T2V/I2V video detector models by by handling both tasks within a single trained model. + +# 4.3. Finegrained Evaluation + +In fake media detection, flagging videos partially or fully manipulated by AI also necessitates fine-grained classification into three classes: (1) real videos, (2) partially manipulated videos, and (3) fully synthetic videos. While binary + +Table 3. SOTA Comparison on Synthetic Data: On the DeMamba dataset (validation split), we compare the performance of UNITE, which was NOT trained on DeMamba train split, against state-of-the-art detectors which were trained on DeMamba train split (results taken from Chen et al. [9]). We report the results ( $P = \text{Precision}$ and $R = \text{Recall}$ ) on the individual T2V/I2V generators and the average performance across the entire validation set ( $Avg$ , which also includes real videos). Although the direct comparison is unfair against UNITE which was trained with FF++ [43] and GTA-V [24], our method still outperforms these synthetic video detectors. Bold shows the current best results and the previous best and second-best results are highlighted in red and blue respectively. + +
MethodMetricsSora [6]Morph Studio [2]Runway ML (Gen2) [42]HotShot [38]Lavie [54]Show-1 [60]Moon Valley [37]Crafter [8]Model Scope [52]Wild Scrape [9]Avg
TALL [57]P71.15%96.89%98.51%79.38%84.59%79.38%98.79%99.02%92.70%76.47%87.91%
R91.07%98.28%97.83%83.00%76.57%79.57%99.52%98.93%94.14%66.31%88.52%
F3Net [41]P68.27%99.89%99.67%89.35%57.00%36.57%99.52%99.71%93.80%88.41%88.73%
R83.93%99.71%98.62%77.57%85.24%63.17%99.58%99.89%89.43%76.78%81.88%
NPR [47]P91.07%99.57%99.49%24.29%89.64%57.71%97.12%99.86%94.29%87.80%82.45%
R91.07%99.57%99.49%24.29%89.64%57.71%97.12%99.86%94.29%87.80%84.08%
STIL [20]P57.21%99.08%99.32%86.19%82.24%70.43%99.25%98.96%97.18%81.32%87.12%
R67.86%96.00%98.41%96.14%77.14%80.43%97.44%96.93%96.29%68.36%82.22%
MINTIME-CLIP-B [9]P83.21%99.99%99.67%50.84%99.20%99.27%99.76%99.99%91.83%91.77%91.55%
R89.29%100.00%98.99%26.43%96.79%98.14%99.84%100.00%84.29%82.38%87.62%
FTCN-CLIP-B [9]P91.79%99.99%99.79%45.94%99.76%97.80%99.99%99.99%94.69%92.32%92.21%
R87.50%100.00%98.91%17.71%97.71%91.86%100.00%100.00%85.29%82.83%86.18%
CLIP-B-PT [9]P67.80%43.56%70.88%29.97%52.97%35.36%55.52%66.03%44.23%42.99%44.83%
R85.71%82.43%90.36%71.00%79.29%75.43%89.62%86.29%82.14%75.16%81.74%
DeMamba-CLIP-PT [9]P25.87%95.14%96.23%73.43%83.31%75.49%90.17%95.06%95.05%69.95%79.97%
R58.93%96.43%93.12%68.00%69.36%69.00%89.14%91.86%96.14%56.59%78.86%
XCLIP-B-PT [9]P16.39%72.16%87.77%39.86%65.57%54.26%75.23%84.80%61.60%55.28%61.29%
R81.34%82.15%83.35%80.98%81.82%81.55%82.14%82.98%81.93%81.10%81.93%
DeMamba-XCLIP-PT [9]P18.26%93.50%94.72%69.94%78.08%71.50%83.95%92.23%93.54%68.10%76.38%
R66.07%95.86%94.64%77.86%75.36%80.29%90.89%92.50%96.00%66.41%83.59%
XCLIP-B-FT [9]P64.42%99.73%96.78%70.98%90.35%77.28%97.34%99.84%82.01%88.97%86.77%
R82.14%99.57%93.62%61.29%79.36%69.71%97.92%99.79%77.14%83.59%84.41%
UNITE (Ours)P88.57%100.00%100.00%90.16%89.91%98.34%99.52%100.00%98.96%92.56%92.76%
R92.11%100.00%94.62%96.93%98.12%99.86%98.69%100.00%96.29%89.89%89.60%
+ +Table 4. Results obtained by the UNITE model on finegrained classes: detecting whether a video is real, partially manipulated or fully AI-generated. Performance gains are mentioned in green. + +
TrainTest
FF++GTA-VDatasetAccuracy
✓FF++96.46%
✓CelebDF60.40%
✓DeeperForensics71.04%
✓DeepFakeTIMIT80.68%
✓HifiFace43.27%
✓UADFV90.37%
✓AVID50.00%
✓GTA-V0.00%
✓DeMamba38.29%
✓✓FF++97.70%(+1.24)
✓✓CelebDF70.17%(+9.77)
✓✓DeeperForensics80.59%(+9.55)
✓✓DeepFakeTIMIT81.04%(+0.36)
✓✓HifiFace64.20%(+20.93)
✓✓UADFV92.61%(+1.24)
✓✓AVID62.50%(+12.50)
✓✓GTA-V100.00%(+100.00)
✓✓DeMamba65.84%(+27.55)
+ +real vs. fake classification may suffice to curb misinformation, fine-grained classification offers greater explainability in otherwise black-box models. + +Using this defined convention of finegrained classes, all the face manipulated datasets (CelebDF [30], DeeperForensics [25], etc.) and the videos from the AVID [61] model (background manipulations) fall under class-2, and the videos from the GTA-V [24] and the DeMamba [9] datasets fall in class-3. In this setup, only the final classification layer of the UNITE transformer architecture is changed to 3 neurons with softmax activation. The results obtained are presented in Table 4. As expected, when the model is trained only with $\mathrm{FF}++$ [43], the performance on fully synthetic data is poor (since the model has not seen any synthetic videos). When the GTA-V [24] data is also + +![](images/4e0ee692c4c5de2ef0684ec2901942f259d3d5e21d2f9aed9f0a466693c70f68.jpg) +(a) UNITE trained on $\mathrm{FF} + +$ [43] only + +![](images/731f76544b4eb7ad0882ec6ef0bec3dce866c8aa404905fbfc7f87320a267252.jpg) +(b) UNITE trained on $\mathrm{FF} + +$ [43] and GTA-V [24] +Figure 4. Ablation Results to show the effect of changing the loss functions used to train the UNITE model. The combination of the cross-entropy (CE) and Attention Diversity (AD) losses always performs better, and the contribution from the AD-loss component increases significantly when using fully synthetic data for training. + +used in training, the performance is much higher. + +# 4.4. Ablation Study + +We conducted an analysis to assess the contribution of individual loss functions to the training of the UNITE model. The outcomes of this ablation study are presented in Fig. 4. Specifically, we evaluated the impact of the AD-loss in two scenarios: (1) when UNITE was trained solely on the FF++ dataset [43] (Fig. 4(a)), and (2) when UNITE was trained using a combination of the FF++ [43] and the GTA-V [24] datasets (Fig. 4(b)). The combination of the two losses always performed better than the individual losses. + +![](images/e2581f9420b0e9c89632475fbb1f99e00a2b12ff8998a84ceffb81c5cc3e3920.jpg) +Figure 5. No. of frames vs. Performance: Performance analysis of UNITE based on the number of frames sampled per video segment. The results illustrate that as the number of frames increases from 1 to 64 (context window), the detection accuracy improves, showcasing UNITE's ability to effectively capture temporal inconsistencies in fake videos. + +Effect on Synthetic Data: UNITE performed significantly better on the fully synthetic datasets GTA-V [24] and DeMamba [9] when the model was training includes GTA-V [24] data. However, for the DeMamba dataset (cross-data evaluation), the performance with only CE-loss is similar regardless of the training data. It is the AD-loss component that provides the accuracy boost in Fig 4(b). + +Effect on Background-Manipulated Data: In Fig. 4(a), we observe that when UNITE was trained solely on $\mathrm{FF}++$ [43] using only the CE-loss, its performance on the AVID dataset [61] was $0.00\%$ . This outcome can be attributed to the fact that the fake data in $\mathrm{FF}++$ [43] contains exclusively face-manipulated videos, causing the UNITE transformer's attention heads to focus predominantly on the facial regions. As AVID [61] comprises only background-manipulated videos, UNITE failed to detect any of the fakes under such conditions. However, when the AD-loss was introduced, the attention heads of UNITE were encouraged to diversify their spatial focus, significantly improving performance. Incorporating synthetic data, as shown in Fig. 4(b), further enhanced this effect, achieving $100\%$ accuracy on AVID when using both CE-loss and AD-loss. + +Effect on Face-Manipulated Data: For the evaluation on face-manipulated datasets such as CelebDF [30] and UADFV [58], one might expect comparable results between training solely on $\mathrm{FF} + +$ [43] and incorporating GTA-V [24] due to their perceived irrelevance in evaluating face-centric manipulations. However, as shown in Fig. 4(b), training with fully synthetic data leads to a substantial performance improvement. Notably, in both Fig. 4(a) and (b), using only CE loss yields similar results, but the inclusion of the AD-loss component significantly enhances overall detection accuracy. The diversity in the training data makes the UNITE model learn more discriminative attention map features through the AD-loss component, which in turn helps even in detecting face-manipulated videos. + +Frames vs. Performance: For inference with UNITE, we evaluated the impact of temporal context by vary- + +![](images/69c7bd286988af9c27ba996e301c20fc5c1e44c982413c17694203378e6c34a8.jpg) +Figure 6. Transformer Depth Evaluation: Performance comparison of UNITE with varying the number of encoder blocks (depth). UNITE performs optimally in cross-domain settings when the depth is 4. Greater depths overfit to the training domains $(\mathrm{FF} + +$ [43] and GTA-V [24]), while a depth of 2 is insufficient to capture the complexity of the data. + +ing the number of frames sampled from video segments to $\{1,8,16,32,64\}$ , with 64 frames corresponding to our model's context window. As shown in Fig. 5, performance improves as more frames are included, demonstrating that UNITE effectively captures temporal discontinuities in fake videos. This highlights the model's ability to leverage temporal cues for enhanced detection accuracy. + +Transformer Depth Ablation: We evaluate the impact of varying the depth of the transformer architecture in UNITE by adjusting the number of encoder blocks, with results shown in Fig. 6. We tested depths of $\{2,4,6,8\}$ and observed that the model achieved the best performance with 4 encoder blocks, which is the depth used in our final architecture. This suggests that a moderate depth strikes the optimal balance between model complexity and performance, especially in cross-dataset settings, providing robust detection capabilities without overfitting. + +# 5. Conclusion + +Traditional DeepFake detectors, focused solely on face-manipulated content, struggle against newer, more sophisticated forgery methods involving T2V, I2V, and background manipulations. To address this, we proposed UNITE, leveraging the fully synthetic GTA-V [24] dataset for training. While this synthetic data is not AI-generated, our results show it effectively enhances UNITE's ability to detect AI-generated content and background manipulations. + +A key innovation in our approach is the AD-loss, which encourages the model to explore diverse spatial regions, enhancing its detection of manipulated areas beyond just the face region. This allows UNITE to excel not only in detecting AI-generated and background-altered videos but also in identifying traditional face-manipulated content. Our ablation studies demonstrate the robustness of UNITE, making it a valuable tool in synthetic video detection. + +Acknowledgements: This work was supported in part by funding from YouTube (Google). + +# References + +[1] Flux 1.1. https://blackforestlabs.ai/, 2024.1 +[2] Morph studio. https://www.morphstudio.com/, 2024.7 +[3] Darius Afchar, Vincent Nozick, Junichi Yamagishi, and Isao Echizen. Mesonet: a compact facial video forgery detection network. In 2018 IEEE international workshop on information forensics and security (WIFS), pages 1-7. IEEE, 2018. 6 +[4] Ibrahim M Alabdulmohsin, Xiaohua Zhai, Alexander Kolesnikov, and Lucas Beyer. Getting vit in shape: Scaling laws for compute-optimal model design. Advances in Neural Information Processing Systems, 36, 2024. 2, 3 +[5] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 5 +[6] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. 2024. URL https://openai.com/research/video-generation-models-as-world-simulators, 3, 2024. 6, 7 +[7] Baoying Chen and Shunquan Tan. Featuretransfer: Unsupervised domain adaptation for cross-domain deepfake detection. Security and Communication Networks, 2021(1): 9942754, 2021. 3 +[8] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 7 +[9] Haoxing Chen, Yan Hong, Zizheng Huang, Zhuoer Xu, Zhangxuan Gu, Yaohui Li, Jun Lan, Huijia Zhu, Jianfu Zhang, Weiqiang Wang, et al. Demamba: Ai-generated video detection on million-scale genvideo benchmark. arXiv preprint arXiv:2405.19707, 2024. 1, 2, 5, 6, 7, 8 +[10] Jikang Cheng, Zhiyuan Yan, Ying Zhang, Yuhao Luo, Zhongyuan Wang, and Chen Li. Can we leave deepfake data behind in training deepfake detector? arXiv preprint arXiv:2408.17052, 2024. 1 +[11] Jongwook Choi, Taehoon Kim, Yonghyun Jeong, Seungryul Baek, and Jongwon Choi. Exploiting style latent flows for generalizing deepfake video detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1133-1143, 2024. 2, 6 +[12] Umur Aybars Ciftci, Ilke Demir, and Lijun Yin. Fakecatcher: Detection of synthetic portrait videos using biological signals. IEEE transactions on pattern analysis and machine intelligence, 2020. 6 +[13] Sara Concas, Simone Maurizio La Cava, Roberto Casula, Giulia Orrù, Giovanni Puglisi, and Gian Luca Marcialis. Quality-based artifact modeling for facial deepfake detection in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3845–3854, 2024. 2, 6 + +[14] Riccardo Corvi, Davide Cozzolino, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. Intriguing properties of synthetic images: from generative adversarial networks to diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 973-982, 2023. 2 +[15] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 2 +[16] Davide Cozzolino, Andreas Rössler, Justus Thies, Matthias Nießner, and Luisa Verdoliva. Id-reveal: Identity-aware deepfake video detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15108-15117, 2021. 2 +[17] Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. The deepfake detection challenge (dfdc) dataset. arXiv preprint arXiv:2006.07397, 2020. 1 +[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3 +[19] Ricard Durall, Margret Keuper, Franz-Josef Pfundt, and Janis Keuper. Unmasking deepfakes with simple features. arXiv preprint arXiv:1911.00686, 2019. 2 +[20] Zhihao Gu, Yang Chen, Taiping Yao, Shouhong Ding, Jilin Li, Feiyue Huang, and Lizhuang Ma. Spatiotemporal inconsistency learning for deepfake video detection. In Proceedings of the 29th ACM international conference on multimedia, pages 3473-3481, 2021. 7 +[21] Ying Guo, Cheng Zhen, and Pengfei Yan. Controllable guide-space for generalizable face forgery detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20818-20827, 2023. 6 +[22] Harald Hanselmann, Shen Yan, and Hermann Ney. Deep fisher faces. In BMVC, 2017. 4 +[23] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 4 +[24] Yuan-Ting Hu, Jiahong Wang, Raymond A Yeh, and Alexander G Schwing. Sail-vos 3d: A synthetic dataset and baselines for object detection and 3d mesh reconstruction from video data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1418-1428, 2021. 1, 2, 5, 6, 7, 8 +[25] Liming Jiang, Ren Li, Wayne Wu, Chen Qian, and Chen Change Loy. Deeperforensics-1.0: A large-scale dataset for real-world face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2889–2898, 2020. 5, 7 +[26] Yan Ju, Chengzhe Sun, Shan Jia, Shuwei Hou, Zhaofeng Si, Soumyya Kanti Datta, Lipeng Ke, Riky Zhou, Anita Nikolich, and Siwei Lyu. Deepfake-o-meter v2.0: An open platform for deepfake detection. In 2024 IEEE 7th Inter- + +national Conference on Multimedia Information Processing and Retrieval (MIPR), pages 439-445. IEEE, 2024. 1 +[27] Pavel Korshunov and Sébastien Marcel. Deepfakes: a new threat to face recognition? assessment and detection. arXiv preprint arXiv:1812.08685, 2018. 5 +[28] Lingzhi Li, Jianmin Bao, Hao Yang, Dong Chen, and Fang Wen. Faceshifter: Towards high fidelity and occlusion aware face swapping. arXiv preprint arXiv:1912.13457, 2019. 1 +[29] Y Li. Exposing deepfake videos by detecting face warping artifact acts. arXiv preprint arXiv:1811.00656, 2018. 6 +[30] Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3207-3216, 2020. 1, 3, 5, 6, 7, 8 +[31] Li Lin, Xinan He, Yan Ju, Xin Wang, Feng Ding, and Shu Hu. Preserving fairness generalization in deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16815-16825, 2024. 6 +[32] Baoping Liu, Bo Liu, Ming Ding, Tianqing Zhu, and Xin Yu. Ti2net: temporal identity inconsistency network for deepfake detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4691-4700, 2023. 2, 6 +[33] I Loshchilov and F Hutter. Decoupled weight decay regularization. *ICLR*, 2019. 5 +[34] Qingxuan Lv, Yuezun Li, Junyu Dong, Sheng Chen, Hui Yu, Huiyu Zhou, and Shu Zhang. Domainforensics: Exposing face forgery across domains via bi-directional adaptation. IEEE Transactions on Information Forensics and Security, 2024. 3 +[35] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 2 +[36] Ghazal Mazaheri and Amit K Roy-Chowdhury. Detection and localization of facial expression manipulations. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1035-1045, 2022. 2 +[37] moonvalley.ai. moonvalley.ai. https://moonvalley.ai/, 2022.7 +[38] John Mullan, Duncan Crawbuck, and Aakash Sastry. Hotshot-XL, 2023. 7 +[39] Dat Nguyen, Nesryne Mejri, Inder Pal Singh, Polina Kuleshova, Marcella Astrid, Anis Kacem, Enjie Ghorbel, and Djamila Aouada. Laa-net: Localized artifact attention network for quality-agnostic and generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17395–17405, 2024. 2 +[40] Alvaro Lopez Pellicer, Yi Li, and Plamen Angelov. Pudd: Towards robust multi-modal prototype-based deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3809-3817, 2024. 2, 6 + +[41] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In European conference on computer vision, pages 86-103. Springer, 2020. 7 +[42] Runway Research. Text driven video generation. https://research.runwayml.com/gen2, 2023.1,7 +[43] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics++: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1-11, 2019. 1, 2, 3, 5, 6, 7, 8 +[44] Md Shamim Seraj, Ankita Singh, and Shayok Chakraborty. Semi-supervised deep domain adaptation for deepfake detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1061-1071, 2024. 2 +[45] Ke Sun, Shen Chen, Taiping Yao, Hong Liu, Xiaoshuai Sun, Shouhong Ding, and Rongrong Ji. Diffusionfake: Enhancing generalization in deepfake detection via guided stable diffusion. NeurIPS, 2024. 1 +[46] Zekun Sun, Yujie Han, Zeyu Hua, Na Ruan, and Weijia Jia. Improving the efficiency and robustness of deepfakes detection through precise geometric features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3609-3618, 2021. 6 +[47] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28130-28139, 2024. 7 +[48] The New York Times. A.i. can now create lifelike videos. can you tell what's real? https://www.nytimes.com/interactive/2024/09/09/technology/ai-video-deepfake-runway-kling-quiz.html, 2024.5,6 +[49] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG), 38(4):1-12, 2019. 1 +[50] Loc Trinh, Michael Tsang, Sirisha Rambhatla, and Yan Liu. Interpretable and trustworthy deepfake detection via dynamic prototypes. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1973-1983, 2021. 2 +[51] A Vaswani, N Shazeer, N Parmar, J Uszkoreit, L Jones, A Gomez, L Kaiser, and I Polosukhin. Attention is all you need. NeurIPS, 2017. 4 +[52] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1, 2, 7 +[53] Yuhan Wang, Xu Chen, Junwei Zhu, Wenqing Chu, Ying Tai, Chengjie Wang, Jilin Li, Yongjian Wu, Feiyue Huang, and Rongrong Ji. Hififace: 3d shape and semantic prior guided high fidelity face swapping. Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, 2021. 3, 5 + +[54] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 7 +[55] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22445-22455, 2023. 2 +[56] Deressa Wodajo and Solomon Atnafu. Deepfake video detection using convolutional vision transformer. arXiv preprint arXiv:2102.11126, 2021. 6 +[57] Yuting Xu, Jian Liang, Gengyun Jia, Ziming Yang, Yanhao Zhang, and Ran He. Tall: Thumbnail layout for deepfake video detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22658-22668, 2023. 1, 2, 6, 7 +[58] Xin Yang, Yuezun Li, and Siwei Lyu. Exposing deep fakes using inconsistent head poses. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 8261-8265. IEEE, 2019. 5, 6, 8 +[59] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3 +[60] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. International Journal of Computer Vision, pages 1–15, 2024. 7 +[61] Zhixing Zhang, Bichen Wu, Xiaoyan Wang, Yaqiao Luo, Luxin Zhang, Yinan Zhao, Peter Vajda, Dimitris Metaxas, and Licheng Yu. Avid: Any-length video inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7162-7172, 2024. 1, 4, 5, 7, 8 +[62] Cairong Zhao, Chutian Wang, Guosheng Hu, Haonan Chen, Chun Liu, and Jinhui Tang. Istvt: interpretable spatial-temporal video transformer for deepfake detection. IEEE Transactions on Information Forensics and Security, 18: 1335-1348, 2023. 6 +[63] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1, 2 +[64] Xinye Zhou, Hu Han, Shiguang Shan, and Xin Chen. Fine-grained open-set deepfake detection via unsupervised domain adaptation. IEEE Transactions on Information Forensics and Security, 2024. 2 \ No newline at end of file diff --git a/2025/Towards a Universal Synthetic Video Detector_ From Face or Background Manipulations to Fully AI-Generated Content/images.zip b/2025/Towards a Universal Synthetic Video Detector_ From Face or Background Manipulations to Fully AI-Generated Content/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1c75cb2785c82ab9d21ab805e6891edc96b8e30a --- /dev/null +++ b/2025/Towards a Universal Synthetic Video Detector_ From Face or Background Manipulations to Fully AI-Generated Content/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:81b3f9eceaac4bfd814574ae6faa54c65568f2d06252b76ead4e3083ef9954c5 +size 586882 diff --git a/2025/Towards a Universal Synthetic Video Detector_ From Face or Background Manipulations to Fully AI-Generated Content/layout.json b/2025/Towards a Universal Synthetic Video Detector_ From Face or Background Manipulations to Fully AI-Generated Content/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..40cec94594193911cc44ae7dfb0ba78653cc93a2 --- /dev/null +++ b/2025/Towards a Universal Synthetic Video Detector_ From Face or Background Manipulations to Fully AI-Generated Content/layout.json @@ -0,0 +1,9295 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 102, + 533, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 102, + 533, + 140 + ], + "spans": [ + { + "bbox": [ + 77, + 102, + 533, + 140 + ], + "type": "text", + "content": "Towards a Universal Synthetic Video Detector: From Face or Background Manipulations to Fully AI-Generated Content" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 160, + 541, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 160, + 541, + 190 + ], + "spans": [ + { + "bbox": [ + 69, + 160, + 541, + 190 + ], + "type": "text", + "content": "Rohit Kundu" + }, + { + "bbox": [ + 69, + 160, + 541, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 69, + 160, + 541, + 190 + ], + "type": "text", + "content": ", Hao Xiong" + }, + { + "bbox": [ + 69, + 160, + 541, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 160, + 541, + 190 + ], + "type": "text", + "content": ", Vishal Mohanty" + }, + { + "bbox": [ + 69, + 160, + 541, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 160, + 541, + 190 + ], + "type": "text", + "content": ", Athula Balachandran" + }, + { + "bbox": [ + 69, + 160, + 541, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 160, + 541, + 190 + ], + "type": "text", + "content": ", Amit K. Roy-Chowdhury" + }, + { + "bbox": [ + 69, + 160, + 541, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 69, + 160, + 541, + 190 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 69, + 160, + 541, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 160, + 541, + 190 + ], + "type": "text", + "content": "Google LLC, " + }, + { + "bbox": [ + 69, + 160, + 541, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 69, + 160, + 541, + 190 + ], + "type": "text", + "content": "University of California, Riverside" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 74, + 191, + 533, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 191, + 533, + 205 + ], + "spans": [ + { + "bbox": [ + 74, + 191, + 533, + 205 + ], + "type": "text", + "content": "{rohit.kundu@email, amertc@ece}.ucr.edu; {rohitkun, haoxg, vishal, athula}@google.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 257, + 296, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 296, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 296, + 341 + ], + "type": "text", + "content": "Existing DeepFake detection techniques primarily focus on facial manipulations, such as face-swapping or lip-syncing. However, advancements in text-to-video (T2V) and image-to-video (I2V) generative models now allow fully AI-generated synthetic content and seamless background alterations, challenging face-centric detection methods and demanding more versatile approaches." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 342, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 594 + ], + "type": "text", + "content": "To address this, we introduce the Universal Network for Identifying Tampered and synthEtic videos (UNITE) model, which, unlike traditional detectors, captures full-frame manipulations. UNITE extends detection capabilities to scenarios without faces, non-human subjects, and complex background modifications. It leverages a transformer-based architecture that processes domain-agnostic features extracted from videos via the SigLIP-So400M foundation model. Given limited datasets encompassing both facial/background alterations and T2V/I2V content, we integrate task-irrelevant data alongside standard DeepFake datasets in training. We further mitigate the model's tendency to over-focus on faces by incorporating an attention-diversity (AD) loss, which promotes diverse spatial attention across video frames. Combining AD loss with cross-entropy improves detection performance across varied contexts. Comparative evaluations demonstrate that UNITE outperforms state-of-the-art detectors on datasets featuring face/background manipulations and fully synthetic T2V/I2V videos, showcasing its adaptability and generalizable detection capabilities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 620, + 135, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 620, + 135, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 620, + 135, + 632 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": "The rise of synthetic media, especially DeepFakes, has transformed content perception and interaction online. Hyper-realistic images produced by technologies like " + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{FLUX}_{1.1}" + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": " [1] are challenging for even humans to identify as fake, and with the development of their video model underway, this underscores the critical need for effective detec" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 328, + 230, + 542, + 399 + ], + "blocks": [ + { + "bbox": [ + 328, + 230, + 542, + 399 + ], + "lines": [ + { + "bbox": [ + 328, + 230, + 542, + 399 + ], + "spans": [ + { + "bbox": [ + 328, + 230, + 542, + 399 + ], + "type": "image", + "image_path": "7f4d537dbea8ce1cfa0f687526832537988d457109714d1137feb379a094bc2b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 402, + 555, + 545 + ], + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 545 + ], + "type": "text", + "content": "Figure 1. Problem Overview: Existing DeepFake detection methods focus on identifying face-manipulated videos, most of which cannot perform inference unless there is a face detected in the video. However, with advancements like seamless background modifications (e.g., AVID [61]) and hyper-realistic content from games like GTA-V [24] and T2V/I2V models [9], a more comprehensive approach is needed. A model trained with only cross-entropy (CE) loss, using full frames, automatically focuses on the face, capturing temporal discontinuities through its transformer architecture, performing better than random " + }, + { + "bbox": [ + 313, + 402, + 555, + 545 + ], + "type": "inline_equation", + "content": "(\\approx)" + }, + { + "bbox": [ + 313, + 402, + 555, + 545 + ], + "type": "text", + "content": " on T2V/I2V content but struggling with background manipulations. UNITE, with its attention-diversity (AD) loss, effectively detects both face/background manipulations and fully synthetic content." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 555, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 554, + 628 + ], + "type": "text", + "content": "tion methods for media generated by deep neural networks. Traditional DeepFake generators [28, 49] focus on manipulating human faces through face-swapping and lip-syncing. However, powerful text-to-video (T2V) and image-to-video (I2V) models [42, 52, 63] have expanded manipulation possibilities beyond faces." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "While conventional detectors [10, 45, 57] perform well on older, face-centric DeepFake datasets [17, 30, 43], they often struggle with newer manipulations involving full scenes or backgrounds. DeepFake-O-Meter [26] is an open-source tool (consisting of several state-of-the-art models) for DeepFake detection, but it cannot run inference unless a human face is visible in the image/video. The rapid spread" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28050" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "of misinformation, particularly during critical periods such as elections, highlights the need for generalizable detection models capable of identifying diverse manipulations, including face, background, and fully AI-generated T2V/I2V content with/without human subjects (overview in Fig. 1)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 294, + 216 + ], + "type": "text", + "content": "To address these challenges, we present a Universal Network for Identifying Tampered and synthEtic videos or UNITE model, which detects both partially manipulated (foreground/background) and fully synthetic videos. Unlike detectors focused solely on face detection, UNITE analyzes entire frames, regardless of whether a human subject is present in the videos." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 218, + 295, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 218, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 218, + 295, + 421 + ], + "type": "text", + "content": "Given the inherent domain gaps in DeepFake datasets [44, 64], even when generated by similar techniques, we leverage the SigLIP-So400m [4] foundation model to extract domain-agnostic features. This serves as inputs to a learnable transformer with multi-head attention, enabling effective detection by capturing temporal inconsistencies in synthetic content. However, preliminary experiments show that training a transformer architecture solely with cross-entropy (CE) loss often leads to the attention heads converging on the face region (Fig. 3). As a result, the model struggles during inference when handling videos where a real human subject is placed in a manipulated background or content generated by T2V [8, 52] and I2V [35, 63] models. To address this, we introduce an \"attention-diversity\" (AD) loss that encourages attention heads to focus on different spatial regions, enhancing the model's ability to capture critical cues from both foreground and background." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 422, + 295, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 295, + 553 + ], + "type": "text", + "content": "Due to the limited availability of open-source datasets covering face or background manipulations and fully AI-generated content, we employ innovative training strategies for UNITE. This includes integrating task-irrelevant data with standard DeepFake datasets to simulate AI-generated synthetic media. Beyond the popularly used FaceForensics++ [43] dataset, we utilize the SAIL-VOS-3D [24] dataset, originally designed for 3D video object segmentation in the game GTA-V. As this dataset is fully synthetic, it helps simulate AI-generated media, enhancing our model's ability to detect diverse forms of synthetic manipulation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 555, + 266, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 555, + 266, + 566 + ], + "spans": [ + { + "bbox": [ + 67, + 555, + 266, + 566 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 570, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 56, + 570, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 295, + 605 + ], + "type": "text", + "content": "- We propose UNITE, a model for detecting partially manipulated (foreground/background) and fully synthetic videos, moving beyond face-centric DeepFake detection." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 606, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 606, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 56, + 606, + 295, + 641 + ], + "type": "text", + "content": "- Unlike detectors relying on face detection and cropping, our model can process full video frames and can detect fakes even if there are no human subjects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 642, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 642, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 642, + 295, + 677 + ], + "type": "text", + "content": "- Using the SigLIP-So400m foundation model [4], we extract domain-agnostic features, enabling generalization to in-the-wild DeepFakes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 677, + 295, + 713 + ], + "type": "text", + "content": "- We introduce an attention-diversity loss, encouraging attention heads to focus on diverse spatial regions, enhancing detection beyond face manipulations." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 553, + 133 + ], + "type": "text", + "content": "- Unlike existing methods that evaluate on a few datasets, we comprehensively assess UNITE on a broad range of face and synthetic datasets with various T2V/I2V generators, outperforming detection in foreground, background, and T2V/I2V manipulations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 143, + 400, + 155 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 143, + 400, + 155 + ], + "spans": [ + { + "bbox": [ + 314, + 143, + 400, + 155 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 163, + 553, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 553, + 414 + ], + "type": "text", + "content": "Face-centric DeepFake Detection: Methods such as [13, 19, 40] focus on modeling spatial inconsistencies for DeepFake detection. Concas et al. [13] track specific facial features (eyes, nose, mouth) in a high-frequency domain, as fake videos exhibit distinct behaviors in this space compared to real ones. However, due to their dataset-centric architectures, it is not a generalizable approach. PUDD [40] learns person-specific prototypes by modeling patterns from real videos and assessing how closely inference videos match these prototypes, making it less effective for detecting in-the-wild DeepFakes. LAA-Net [39] attempts to improve generalizability through a multi-task attention module that detects artifacts via heatmap and self-consistency regression, but it works by analyzing every still frame of a video, thus limiting its scalability. The authors in [36] tackled the challenging problem of detecting and localizing expression swaps, where the number of manipulated pixels is even fewer compared to identity swaps. Although the authors obtained good localization performance, their reliance on a CNN-based architecture limits their ability to capture temporal inconsistencies effectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 415, + 553, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 553, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 553, + 582 + ], + "type": "text", + "content": "DPNet [50] and ID-Reveal [16] are identity-aware temporal artifact modeling methods: DPNet [50] focuses on interpretable, prototype-based detection of unnatural movements through dynamic feature representations, while IDReveal [16] leverages metric learning on real data to detect biometric inconsistencies. However these methods assume access to authentic reference videos, making them unfit for in-the-wild DeepFake detection. Shifting from identity-specific methods, TALL [57] constructs composite \"thumbnail\" images from four consecutive frames taken at random timestamps for temporal analysis. TI2Net [32] captures temporal anomalies by subtracting consecutive frame features but relies on the assumption that FaceSwap manipulations show pronounced inter-frame discrepancies." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 553, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 641 + ], + "type": "text", + "content": "Choi et al. [11] approach DeepFake detection by targeting suppressed variance in style-based latent temporal features via a StyleGRU module. However, this method is limited to cropped face regions, making it ineffective for T2V/I2V or background-manipulated videos." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "content": "Synthetic Video Detection: While there has been considerable effort in synthetic image detection [14, 15, 55], synthetic video detection is a scarcely explored area. DeMamba [9], a recently proposed synthetic video detector, analyzes small zones of video frames to capture how pixels change spatiotemporally. Its continuous scanning approach" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28051" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 270, + 266 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 270, + 266 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 270, + 266 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 270, + 266 + ], + "type": "image", + "image_path": "293854bf54d05f4417715994a0c7ca60da622532c26dfd731035498dd8ac84a2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 267, + 296, + 345 + ], + "lines": [ + { + "bbox": [ + 55, + 267, + 296, + 345 + ], + "spans": [ + { + "bbox": [ + 55, + 267, + 296, + 345 + ], + "type": "text", + "content": "Figure 2. UNITE architecture overview: We extract domain-agnostic features " + }, + { + "bbox": [ + 55, + 267, + 296, + 345 + ], + "type": "inline_equation", + "content": "(\\xi)" + }, + { + "bbox": [ + 55, + 267, + 296, + 345 + ], + "type": "text", + "content": " using the SigLIP-So400m foundation model [4] to mitigate domain gaps between DeepFake datasets (Sec. 3.2). These embeddings, combined with positional encodings, are input to a transformer with multi-head attention and MLP layers (Sec. 3.3), culminating in a classifier for final predictions. AD-loss (Sec. 3.4) encourages the model attention to span diverse spatial regions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 356, + 296, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 356, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 296, + 487 + ], + "type": "text", + "content": "better tracks subtle changes and patterns, making it easier to spot manipulated or fake content. The authors also propose a million scale T2V/I2V dataset (which we call the DeMamba dataset hereforth) to evaluate their performance. However, the method was not evaluated on typical DeepFake datasets [30, 43, 53] and thus is not fit for human-centric DeepFake detection. In contrast our UNITE model is built to detect partially manipulated as well as fully AI-generated videos. A key reason for the scarcity of synthetic video detection methods in the literature is the lack of standardized datasets for training and evaluation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 498, + 162, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 162, + 512 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 162, + 512 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 519, + 296, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 296, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 296, + 592 + ], + "type": "text", + "content": "In this section we first describe our problem setup (Sec. 3.1), then we describe the foundation model-based video encoding (Sec. 3.2), our trainable transformer architecture (Sec. 3.3) and finally the novel loss function for training the UNITE model (Sec. 3.4). The overview of the UNITE architecture is shown in Fig. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 599, + 148, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 148, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 148, + 612 + ], + "type": "text", + "content": "3.1. Problem Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": "For each video in the dataset, we implement a frame sampling strategy where every other frame is extracted. From the resulting collection of frames, we construct video segment samples consisting of " + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "inline_equation", + "content": "n_f = 64" + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": " consecutive frames, which are defined as a single data sample, denoted as \"v\". This selection aligns with the context window established for our video transformer model, ensuring that each sample retains the temporal coherence necessary for effective train-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "ing. In instances where the video duration results in fewer than " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "n_f" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": " frames, we apply padding (ablation in supplementary) to ensure that each sample remains uniform in size." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 109, + 555, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 555, + 240 + ], + "type": "text", + "content": "Given that the original video-level classification labels are known and fake videos contain manipulations throughout their entirety (i.e., no videos have manipulations limited to specific frames), each " + }, + { + "bbox": [ + 313, + 109, + 555, + 240 + ], + "type": "inline_equation", + "content": "n_f" + }, + { + "bbox": [ + 313, + 109, + 555, + 240 + ], + "type": "text", + "content": "-frame segment generated from a given video is assigned the same label (denoted by " + }, + { + "bbox": [ + 313, + 109, + 555, + 240 + ], + "type": "inline_equation", + "content": "l \\in 0, \\dots, n_c" + }, + { + "bbox": [ + 313, + 109, + 555, + 240 + ], + "type": "text", + "content": ") as the source video, with number of possible classes as " + }, + { + "bbox": [ + 313, + 109, + 555, + 240 + ], + "type": "inline_equation", + "content": "n_c" + }, + { + "bbox": [ + 313, + 109, + 555, + 240 + ], + "type": "text", + "content": ". The aggregation of these samples across all videos forms our final dataset, " + }, + { + "bbox": [ + 313, + 109, + 555, + 240 + ], + "type": "inline_equation", + "content": "v_i, l_{i=1}^N \\in \\mathcal{V}" + }, + { + "bbox": [ + 313, + 109, + 555, + 240 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 109, + 555, + 240 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 109, + 555, + 240 + ], + "type": "text", + "content": " is the total number of samples obtained. This approach serves as a data augmentation technique, increasing the number of training samples without additional data collection." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 248, + 510, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 510, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 510, + 261 + ], + "type": "text", + "content": "3.2. Domain-Agnostic Feature Extraction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 267, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 267, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 555, + 445 + ], + "type": "text", + "content": "Our objective is to leverage the trained UNITE model for detecting in-the-wild DeepFakes, necessitating careful consideration of the domain gap [7, 34] between the train and test datasets. To effectively mitigate the impact of domain discrepancies, we utilize the SigLIP foundation model [59] to extract domain-agnostic features. Specifically, we employ the \"shape-optimized\" ViT [4] image encoder, whose architecture is rooted in [18]. The SigLIP-So400m [4] model was contrastively pretrained on 3B diverse examples using a sigmoid loss, enabling it to capture robust, generalizable features while maintaining a compact size (400M parameters). This extensive pretraining makes SigLIP particularly adept at extracting domain-invariant representations, crucial for handling the significant variability encountered in DeepFake datasets and real-world scenarios." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 447, + 556, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 556, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 556, + 593 + ], + "type": "text", + "content": "For every frame " + }, + { + "bbox": [ + 313, + 447, + 556, + 593 + ], + "type": "inline_equation", + "content": "f_{j}" + }, + { + "bbox": [ + 313, + 447, + 556, + 593 + ], + "type": "text", + "content": " (resized to " + }, + { + "bbox": [ + 313, + 447, + 556, + 593 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{384\\times 384\\times 3}" + }, + { + "bbox": [ + 313, + 447, + 556, + 593 + ], + "type": "text", + "content": "), of a video sample " + }, + { + "bbox": [ + 313, + 447, + 556, + 593 + ], + "type": "inline_equation", + "content": "v_{i}\\in \\mathcal{V}" + }, + { + "bbox": [ + 313, + 447, + 556, + 593 + ], + "type": "text", + "content": ", we obtain an image encoding from the frozen SigLIP-So400m [4] model as " + }, + { + "bbox": [ + 313, + 447, + 556, + 593 + ], + "type": "inline_equation", + "content": "e_j = SigLIP(f_j)\\in \\mathbb{R}^{t_s\\times d_s}" + }, + { + "bbox": [ + 313, + 447, + 556, + 593 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 447, + 556, + 593 + ], + "type": "inline_equation", + "content": "j\\in \\{1,2,\\dots ,n_f\\}" + }, + { + "bbox": [ + 313, + 447, + 556, + 593 + ], + "type": "text", + "content": ", since there are " + }, + { + "bbox": [ + 313, + 447, + 556, + 593 + ], + "type": "inline_equation", + "content": "n_f" + }, + { + "bbox": [ + 313, + 447, + 556, + 593 + ], + "type": "text", + "content": " frames per video sample (as explained in Sec. 3.1), " + }, + { + "bbox": [ + 313, + 447, + 556, + 593 + ], + "type": "inline_equation", + "content": "t_s = 729" + }, + { + "bbox": [ + 313, + 447, + 556, + 593 + ], + "type": "text", + "content": " represents the token length and " + }, + { + "bbox": [ + 313, + 447, + 556, + 593 + ], + "type": "inline_equation", + "content": "d_{s} = 1152" + }, + { + "bbox": [ + 313, + 447, + 556, + 593 + ], + "type": "text", + "content": " represents the feature dimension. We concatenate these frame-level features for the video sample " + }, + { + "bbox": [ + 313, + 447, + 556, + 593 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 313, + 447, + 556, + 593 + ], + "type": "text", + "content": ", while maintaining the order of the frames, to obtain a video segment encoding denoted by " + }, + { + "bbox": [ + 313, + 447, + 556, + 593 + ], + "type": "inline_equation", + "content": "\\xi_{i}\\in \\mathbb{R}^{n_{f}\\times t_{s}\\times d_{s}}" + }, + { + "bbox": [ + 313, + 447, + 556, + 593 + ], + "type": "text", + "content": ", which are used as input to our trainable transformer architecture. This encoded video dataset can be denoted as " + }, + { + "bbox": [ + 313, + 447, + 556, + 593 + ], + "type": "inline_equation", + "content": "\\{\\xi_i,l_i\\}_{i = 1}^N\\in \\mathcal{V}_\\xi" + }, + { + "bbox": [ + 313, + 447, + 556, + 593 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 599, + 510, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 510, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 510, + 611 + ], + "type": "text", + "content": "3.3. Learnable Transformer Architecture" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "content": "The UNITE architecture is a multi-head self-attention (MHSA) based transformer model designed specifically for video classification tasks. It leverages the strengths of Transformer networks to process sequences of video frame embeddings effectively, allowing for robust classification performance. The UNITE architecture allows for adjustable depth, signifying the number of encoder blocks stacked within the model. A deeper architecture can capture more" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28052" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "intricate features and dependencies at the cost of increased computational complexity. For our architecture the depth is set to 4. Ablation experiments with different transformer depths are provided in Sec. 4.4." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 137, + 151, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 137, + 151, + 148 + ], + "spans": [ + { + "bbox": [ + 55, + 137, + 151, + 148 + ], + "type": "text", + "content": "3.3.1 Encoder Block" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 156, + 293, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 293, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 293, + 168 + ], + "type": "text", + "content": "Each encoder block consists of the following components:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 169, + 296, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 169, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 296, + 289 + ], + "type": "text", + "content": "Multi-Head Self-Attention Layer: This layer allows the model to focus on different parts of the input sequence simultaneously. With 12 attention heads in our architecture, it captures diverse interactions and dependencies among frames. Each head computes attention scores independently, learning multiple representations of the input data. This enhances the model's ability to detect subtle temporal variations, which is crucial for video classification. The use of scaled dot-product attention further helps mitigate large gradients and stabilizes training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 289, + 296, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 385 + ], + "type": "text", + "content": "Layer Normalization and Residual Connections: To aid gradient flow during training, each sub-layer incorporates residual connections that add the input back to the output, mitigating vanishing gradient issues in deep networks. Layer normalization follows to stabilize activations and enhance convergence speed. Additionally, dropout is applied post-attention to reduce overfitting by randomly deactivating a portion of neurons during training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 385, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 296, + 468 + ], + "type": "text", + "content": "Feed-Forward Network (MLP): The second sub-layer is a point-wise feed-forward network with two dense layers separated by a GELU activation [23]. This MLP enables nonlinear feature transformations, enhancing the model's expressiveness. By projecting attention outputs into a higher-dimensional space before reducing them, it captures complex interactions beyond linear transformations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 469, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 296, + 553 + ], + "type": "text", + "content": "Attention maps are derived from the outputs of the encoder blocks to be used for computing the Attention-Diversity loss (Sec. 3.4) and to illustrate how the model prioritizes different spatial regions in the frames during the classification process (Fig. 3). This interpretability is valuable for understanding the model's decision-making, particularly in complex video classification scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 174, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 174, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 174, + 582 + ], + "type": "text", + "content": "3.3.2 Positional Encoding" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 589, + 296, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 589, + 296, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 296, + 661 + ], + "type": "text", + "content": "We utilize a sine-cosine positional encoding scheme following the original transformer [51] to provide a unique position identifier for each input token. For every frame " + }, + { + "bbox": [ + 55, + 589, + 296, + 661 + ], + "type": "inline_equation", + "content": "f_{j}" + }, + { + "bbox": [ + 55, + 589, + 296, + 661 + ], + "type": "text", + "content": " in the video sample with encoded feature dimension " + }, + { + "bbox": [ + 55, + 589, + 296, + 661 + ], + "type": "inline_equation", + "content": "d_{s}" + }, + { + "bbox": [ + 55, + 589, + 296, + 661 + ], + "type": "text", + "content": " (as per Sec. 3.2), we compute the positional encoding for odd and even indexed feature dimensions as," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 679, + 299, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 679, + 299, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 679, + 299, + 713 + ], + "type": "interline_equation", + "content": "P E _ {(j, 2 i + 1)} = \\cos \\left(\\frac {j}{1 0 0 0 0 ^ {\\frac {2 i}{d _ {s}}}}\\right), \\quad P E _ {(j, 2 i)} = \\sin \\left(\\frac {j}{1 0 0 0 0 ^ {\\frac {2 i}{d _ {s}}}}\\right), \\tag {1}", + "image_path": "6f4c3bab341443aa74caee8b65e41af9c380aad11504adc0a400815ef02215f4.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " denotes the feature index in the encoded feature dimension " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "d_{s}" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": ". We embed these encodings directly into the input tokens of our video transformer to provide positional context, for better temporal modeling during the attention computation with minimal computational overhead." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 138, + 452, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 138, + 452, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 138, + 452, + 152 + ], + "type": "text", + "content": "3.4. Attention-Diversity Loss" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 157, + 554, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 554, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 554, + 252 + ], + "type": "text", + "content": "Training a multi-head attention transformer with only cross-entropy loss makes the attention maps focus only on the face regions of the frame (as evident in Fig. 3). However, we aim to detect fake videos where the manipulation might not be in the face at all, and instead be in the background (like video inpainting with the AVID model [61]). To ensure that the attention heads focus on diverse spatial regions of the video frames, we devise the \"Attention-Diversity\" (AD) loss." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 253, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 555, + 348 + ], + "type": "text", + "content": "AD-loss is designed to minimize overlap among attention maps (obtained from the first encoder block of the trainable transformer architecture in 3.3.1) while maintaining consistency across different input video samples. The attention outputs from the video transformer model " + }, + { + "bbox": [ + 313, + 253, + 555, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{A} \\in \\mathbb{R}^{n_h \\times t_s \\times d_s}" + }, + { + "bbox": [ + 313, + 253, + 555, + 348 + ], + "type": "text", + "content": " are used to pool the input SigLIP-So400m features " + }, + { + "bbox": [ + 313, + 253, + 555, + 348 + ], + "type": "inline_equation", + "content": "\\xi \\in \\mathbb{R}^{n_f \\times t_s \\times d_s}" + }, + { + "bbox": [ + 313, + 253, + 555, + 348 + ], + "type": "text", + "content": ", resulting in a pooled feature tensor " + }, + { + "bbox": [ + 313, + 253, + 555, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{P} \\in \\mathbb{R}^{n_h \\times n_f}" + }, + { + "bbox": [ + 313, + 253, + 555, + 348 + ], + "type": "text", + "content": " according to," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 376, + 357, + 554, + 391 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 357, + 554, + 391 + ], + "spans": [ + { + "bbox": [ + 376, + 357, + 554, + 391 + ], + "type": "interline_equation", + "content": "\\mathcal {P} = \\sum_ {j = 1} ^ {t _ {s}} \\sum_ {k = 1} ^ {d _ {s}} \\mathcal {A} _ {h, j, k} \\cdot \\xi_ {f, j, k}, \\tag {2}", + "image_path": "7b4ae65b76c5490d2a3527af4ac68f0c945fec52f4095a5b9e02cd4cdfb87adb.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 394, + 554, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 554, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 554, + 418 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 394, + 554, + 418 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 394, + 554, + 418 + ], + "type": "text", + "content": " indexes the frames " + }, + { + "bbox": [ + 313, + 394, + 554, + 418 + ], + "type": "inline_equation", + "content": "(n_f)" + }, + { + "bbox": [ + 313, + 394, + 554, + 418 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 394, + 554, + 418 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 394, + 554, + 418 + ], + "type": "text", + "content": " indexes the number of attention heads " + }, + { + "bbox": [ + 313, + 394, + 554, + 418 + ], + "type": "inline_equation", + "content": "(n_h)" + }, + { + "bbox": [ + 313, + 394, + 554, + 418 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 394, + 554, + 418 + ], + "type": "inline_equation", + "content": "j, k" + }, + { + "bbox": [ + 313, + 394, + 554, + 418 + ], + "type": "text", + "content": " are spatial positions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 418, + 554, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 554, + 514 + ], + "type": "text", + "content": "We compute \"feature centers\" which are points in the feature space that serve as anchors for the learned representations of different classes. These centers represent the average features of samples from a particular class, allowing the model to capture essential characteristics of that class. In the context of the AD-loss, feature centers " + }, + { + "bbox": [ + 313, + 418, + 554, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{C} \\in \\mathbb{R}^{n_h \\times n_f}" + }, + { + "bbox": [ + 313, + 418, + 554, + 514 + ], + "type": "text", + "content": " are dynamically updated in each training iteration " + }, + { + "bbox": [ + 313, + 418, + 554, + 514 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 418, + 554, + 514 + ], + "type": "text", + "content": " based on the pooled feature vectors " + }, + { + "bbox": [ + 313, + 418, + 554, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 418, + 554, + 514 + ], + "type": "text", + "content": ", following," + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 353, + 522, + 554, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 522, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 353, + 522, + 554, + 555 + ], + "type": "interline_equation", + "content": "\\mathcal {C} ^ {\\tau} = \\mathcal {C} ^ {\\tau - 1} - \\eta \\left(\\mathcal {C} ^ {\\tau - 1} - \\frac {1}{B} \\sum_ {b = 1} ^ {B} \\mathcal {P} _ {b}\\right), \\tag {3}", + "image_path": "83287f9aa1426c9c843da3dd8dd23aae9a52a77ed1e1ae2ff304994ea7e28e97.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 558, + 554, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 606 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 558, + 554, + 606 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 558, + 554, + 606 + ], + "type": "text", + "content": " represents the learning rate (set to 0.05 in our experiments) for feature center updates and " + }, + { + "bbox": [ + 313, + 558, + 554, + 606 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 558, + 554, + 606 + ], + "type": "text", + "content": " is the batch dimension. The feature centers are initialized with zeros " + }, + { + "bbox": [ + 313, + 558, + 554, + 606 + ], + "type": "inline_equation", + "content": "(\\mathcal{C}^0 = [0]_{n_h\\times n_f})" + }, + { + "bbox": [ + 313, + 558, + 554, + 606 + ], + "type": "text", + "content": " at the beginning of the model training." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 607, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 607, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 607, + 555, + 714 + ], + "type": "text", + "content": "AD-loss consists of two distinct components: within-class loss and between-class loss inspired by the Fisher Discriminant Analysis for deep networks [22]. The within-class term calculates the distance between the pooled feature vectors " + }, + { + "bbox": [ + 313, + 607, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 607, + 555, + 714 + ], + "type": "text", + "content": " and their corresponding feature centers " + }, + { + "bbox": [ + 313, + 607, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 607, + 555, + 714 + ], + "type": "text", + "content": ", promoting closeness among similar samples. The between-class term measures the distance between feature centers of different classes, encouraging them to be spaced apart in the feature space, encouraging separability." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28053" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 72, + 240, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 240, + 83 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 240, + 83 + ], + "type": "text", + "content": "The within-class loss term is calculated as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 96, + 95, + 294, + 109 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 95, + 294, + 109 + ], + "spans": [ + { + "bbox": [ + 96, + 95, + 294, + 109 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {w i t h i n}} = \\max \\left(\\| \\mathcal {P} - \\mathcal {C} \\| _ {2} - \\delta_ {\\text {w i t h i n}}, 0\\right), \\tag {4}", + "image_path": "14cfcfb6dc8f74c69c377b1f395ca75f2025949b80bcef9e3a2de18a77b34478.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 111, + 295, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 111, + 295, + 159 + ], + "spans": [ + { + "bbox": [ + 55, + 111, + 295, + 159 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 111, + 295, + 159 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{within}} \\in \\mathbb{R}^{n_c}" + }, + { + "bbox": [ + 55, + 111, + 295, + 159 + ], + "type": "text", + "content": " is a hyperparameter controlling the allowable distance between feature vectors and their respective feature centers for inputs of the same class, encouraging tighter clustering and " + }, + { + "bbox": [ + 55, + 111, + 295, + 159 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_2" + }, + { + "bbox": [ + 55, + 111, + 295, + 159 + ], + "type": "text", + "content": " represents " + }, + { + "bbox": [ + 55, + 111, + 295, + 159 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 55, + 111, + 295, + 159 + ], + "type": "text", + "content": " normalization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 159, + 226, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 159, + 226, + 171 + ], + "spans": [ + { + "bbox": [ + 67, + 159, + 226, + 171 + ], + "type": "text", + "content": "The between-class loss is calculated as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 186, + 294, + 230 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 186, + 294, + 230 + ], + "spans": [ + { + "bbox": [ + 61, + 186, + 294, + 230 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {b e t w e e n}} = \\sum_ {\\substack {k \\neq l \\\\ (k, l) \\in \\left(n _ {h}, n _ {h}\\right)}} \\max \\left(\\delta_ {\\text {b e t w e e n}} - \\| \\mathcal {C} _ {k} - \\mathcal {C} _ {l} \\| _ {2}, 0\\right), \\tag{5}", + "image_path": "a5cadf1f9b1474f4bb296c41b6602587a5c6a46a0520f37ed1d9dde882227346.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 231, + 296, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 231, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 296, + 291 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 231, + 296, + 291 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{between}}" + }, + { + "bbox": [ + 55, + 231, + 296, + 291 + ], + "type": "text", + "content": " is a predefined hyperparameter that ensures a minimum distance between feature centers of different classes. Thus the AD-loss is a simple addition of these two components " + }, + { + "bbox": [ + 55, + 231, + 296, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{AD} = \\mathcal{L}_{\\mathrm{within}} + \\mathcal{L}_{\\mathrm{between}}" + }, + { + "bbox": [ + 55, + 231, + 296, + 291 + ], + "type": "text", + "content": ", making the final objective function for training the UNITE model," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 108, + 302, + 295, + 315 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 302, + 295, + 315 + ], + "spans": [ + { + "bbox": [ + 108, + 302, + 295, + 315 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {U N I T E}} = \\lambda_ {1} \\cdot \\mathcal {L} _ {C E} + \\lambda_ {2} \\cdot \\mathcal {L} _ {A D}, \\tag {6}", + "image_path": "c38fc20f4d1bb2ac980597b4e6c178ddd4ae89b31b1c9b86493c62d3b81d79d7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 317, + 296, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 296, + 342 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 317, + 296, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + }, + { + "bbox": [ + 55, + 317, + 296, + 342 + ], + "type": "text", + "content": " is the traditional cross-entropy loss, and " + }, + { + "bbox": [ + 55, + 317, + 296, + 342 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 55, + 317, + 296, + 342 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 317, + 296, + 342 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 55, + 317, + 296, + 342 + ], + "type": "text", + "content": " are loss-weighting hyperparameters." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 350, + 137, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 137, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 137, + 363 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 369, + 118, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 118, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 118, + 381 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 388, + 296, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 472 + ], + "type": "text", + "content": "Most DeepFake datasets focus on face manipulations, such as face-swaps or lip-syncing, and lack fully AI-generated video content. The DeMamba dataset [9] features videos from T2V/I2V models, including powerful diffusion models like SORA [5]. It is not a human-centric dataset and primarily contains videos from natural scenes. We use its validation split to evaluate UNITE." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 472, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 544 + ], + "type": "text", + "content": "To train UNITE, we employ the FaceForensics++ (FF++) [43] dataset (c23 i.e., medium-quality) and the SAIL-VOS-3D dataset [24], which includes synthetic (although not AI-generated) GTA-V game videos featuring human subjects suitable for DeepFake detection. To evaluate the UNITE model, we use:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 544, + 295, + 652 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 55, + 544, + 295, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 580 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 580 + ], + "type": "text", + "content": "- Face manipulated data: " + }, + { + "bbox": [ + 55, + 544, + 295, + 580 + ], + "type": "inline_equation", + "content": "\\mathrm{FF}++" + }, + { + "bbox": [ + 55, + 544, + 295, + 580 + ], + "type": "text", + "content": " [43] (in-domain evaluation), CelebDF [30], DeeperForensics [25], Deepfake-TIMIT [27], HifiFace [53], UADFV [58]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 581, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 295, + 604 + ], + "type": "text", + "content": "- Background manipulated data: Sample videos from the AVID [61] model provided publicly by the authors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "type": "text", + "content": "- Fully synthetic data: GTA-V [24] (in-domain evaluation), DeMamba [9]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 628, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 295, + 652 + ], + "type": "text", + "content": "- In-the-wild DeepFakes: Publicly available videos from the New York Times DeepFake quiz [48]." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 659, + 186, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 186, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 186, + 671 + ], + "type": "text", + "content": "4.2. Results and Discussion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Evaluation Metrics: Although most DeepFake detection methods only report accuracy, we evaluate UNITE using additional metrics: (1) AUC (Area Under the Precision-Recall" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "Curve); (2) Precision@0.5 and (3) Recall@0.5 at a 0.5 confidence threshold; (4) Precs@Rec=0.8 (precision at 0.8 recall); (5) Rec@Precs=0.8 (recall at 0.8 precision)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 109, + 554, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 554, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 554, + 205 + ], + "type": "text", + "content": "Training Details: UNITE is trained using an AdamW optimizer [33] with an initial learning rate of 0.0001 and a decay rate of 0.5 every 1000 steps. For AD-loss, " + }, + { + "bbox": [ + 313, + 109, + 554, + 205 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{between}} = 0.5" + }, + { + "bbox": [ + 313, + 109, + 554, + 205 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 109, + 554, + 205 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{within}}" + }, + { + "bbox": [ + 313, + 109, + 554, + 205 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 313, + 109, + 554, + 205 + ], + "type": "inline_equation", + "content": "[0.01, -2]" + }, + { + "bbox": [ + 313, + 109, + 554, + 205 + ], + "type": "text", + "content": " for binary and " + }, + { + "bbox": [ + 313, + 109, + 554, + 205 + ], + "type": "inline_equation", + "content": "[0.01, -2, 1]" + }, + { + "bbox": [ + 313, + 109, + 554, + 205 + ], + "type": "text", + "content": " for fine-grained classification in Sec. 4.3 (see supplementary for sensitivity). Loss weights are " + }, + { + "bbox": [ + 313, + 109, + 554, + 205 + ], + "type": "inline_equation", + "content": "\\lambda_1 = \\lambda_2 = 0.5" + }, + { + "bbox": [ + 313, + 109, + 554, + 205 + ], + "type": "text", + "content": ". Training uses a batch size of 32 for 25 epochs on 8 TPUv3 chips, with the framework implemented in TensorFlow." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 207, + 555, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 555, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 555, + 315 + ], + "type": "text", + "content": "Quantitative Results: The results obtained by the UNITE model when trained on " + }, + { + "bbox": [ + 313, + 207, + 555, + 315 + ], + "type": "inline_equation", + "content": "\\mathrm{FF}++" + }, + { + "bbox": [ + 313, + 207, + 555, + 315 + ], + "type": "text", + "content": " [43] alone, and when trained with both " + }, + { + "bbox": [ + 313, + 207, + 555, + 315 + ], + "type": "inline_equation", + "content": "\\mathrm{FF}++" + }, + { + "bbox": [ + 313, + 207, + 555, + 315 + ], + "type": "text", + "content": " [43] and GTA-V [24] data and evaluated on the various datasets mentioned in Sec. 4.1 are shown in Table 1. On the AVID [61] and DeMamba datasets [9], the model performs poorly when trained with only face manipulated data, but the performance enhances several-fold when the GTA-V [24] synthetic data, even though it is not AI-generated, is used in training." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 316, + 554, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 554, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 554, + 459 + ], + "type": "text", + "content": "Interestingly, when evaluating on face-manipulated datasets such as CelebDF [30] and DeeperForensics [25], we observe a performance boost, particularly for CelebDF [30], when GTA-V [24] data is included in the training set. Intuitively, one would expect that training solely on " + }, + { + "bbox": [ + 313, + 316, + 554, + 459 + ], + "type": "inline_equation", + "content": "\\mathrm{FF}++" + }, + { + "bbox": [ + 313, + 316, + 554, + 459 + ], + "type": "text", + "content": " [43] would yield similar results since the inference set consists of similar face-manipulated data. This deviation from the expected behavior comes from the contribution of the AD-Loss, as the ablation study (Fig. 4) using only CE-loss compared to both CE and AD-losses reveals that the performance is enhanced with GTA-V [24] data in training only in the latter (detailed discussion in Sec. 4.4)." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 461, + 554, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 554, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 554, + 556 + ], + "type": "text", + "content": "To evaluate UNITE on in-the-wild DeepFakes, we attempted the recent New York Times quiz [48] (denoted by NYTimes) which provides 10 videos (4 real and 6 fake) for testing DeepFake detectors. UNITE got 8 of those 10 videos correct, even though some of the face-swap videos were indistinguishable from real videos, even by humans. On all the AI-generated videos, UNITE made the correct prediction (when trained on " + }, + { + "bbox": [ + 313, + 461, + 554, + 556 + ], + "type": "inline_equation", + "content": "\\mathrm{FF} + +" + }, + { + "bbox": [ + 313, + 461, + 554, + 556 + ], + "type": "text", + "content": " [43] and GTA-V [24])." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "Attention Heatmaps: To analyze the impact of CE and AD-losses on attention distribution, we extracted attention features from the first encoder of the UNITE model, comparing models trained with CE loss alone versus that trained with both CE and AD-losses. The resulting heatmaps, shown in Fig. 3, reveal that the model trained with only CE loss tends to concentrate primarily on the face region. In contrast, the model trained with the combined " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{CE + AD}" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": " loss demonstrates a broader attention span across the frame, as evidenced by a lighter, more distributed bluish tone in the heatmaps, indicating increased spatial diversity in the model's focus. This is why on the example shown in the first row (face-manipulated video) both models give the correct" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28054" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 105, + 93, + 503, + 284 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 92 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 92 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 92 + ], + "type": "text", + "content": "Table 1. Results from the UNITE model trained with (1) FF++ [43] only and (2) FF++ [43] combined with GTA-V [24]. All other results reflect cross-dataset evaluations except for FF++ and GTA-V (when trained). Performance gains are highlighted in green." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 105, + 93, + 503, + 284 + ], + "lines": [ + { + "bbox": [ + 105, + 93, + 503, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 503, + 284 + ], + "type": "table", + "html": "
TrainTest
FF++GTA-VDatasetAccuracyAUCPrecision@0.5Recall@0.5Precs@Rec=0.8Rec@Precs=0.8
Face Manipulated Data
✓FF++99.53%99.77%99.94%99.49%99.94%99.94%
✓CelebDF72.61%94.05%96.45%61.22%80.45%61.22%
✓DeeperForensics91.35%100.00%100.00%91.35%100.00%91.35%
✓DeepfakeTIMIT86.90%86.46%83.61%83.97%88.90%81.33%
✓HiFiFace63.63%62.47%67.12%63.63%59.30%63.63%
✓UADFV94.12%94.38%95.68%97.11%93.79%94.38%
✓✓FF++99.96%(+0.43)99.89%(+0.12)100.00%(+0.06)99.84%(+0.35)100.00%(+0.06)99.96%(+0.02)
✓✓CelebDF95.11%(+22.50)94.36%(+0.31)96.82%(+0.37)68.75%(+7.53)96.53%(+16.08)68.75%(+7.53)
✓✓DeeperForensics99.62%(+8.27)100.00%(+0.00)100.00%(+0.00)99.62%(+8.27)100.00%(+0.00)99.63%(+8.28)
✓✓DeepfakeTIMIT91.90%(+5.00)91.33%(+4.87)90.45%(+6.84)88.39%(+4.42)100.00%(+11.10)91.95%(+10.62)
✓✓HiFiFace75.62%(+11.99)81.24%(+18.77)79.55%(+12.43)71.71%(+8.08)75.62%(+16.32)72.47%(+8.84)
✓✓UADFV97.01%(+2.89)94.95%(+0.57)96.89%(+1.21)100.00%(+2.89)94.12%(+0.33)100.00%(+5.62)
Background Manipulated Data
✓AVID41.67%33.33%33.33%41.67%41.67%33.33%
✓✓AVID100.00%(+58.33)100.00%(+66.67)100.00%(+66.67)100.00%(+58.33)100.00%(+58.33)100.00%(+66.67)
Fully Synthetic Data
✓GTA-V60.16%61.52%60.16%58.73%63.29%58.73%
✓DeMamba61.47%57.38%67.73%33.01%62.15%54.16%
✓✓GTA-V100.00%(+39.84)100.00%(+38.48)100.00%(+39.84)100.00%(+41.27)100.00%(+36.71)100.00%(+41.27)
✓✓DeMamba87.12%(+25.65)93.75%(+36.67)92.76%(+25.03)89.60%(+56.59)89.81%(+27.66)92.12%(+37.96)
In-the-wild DeepFakes
✓NYTimes [48]50.00%53.74%50.00%25.00%0.00%0.00%
✓✓NYTimes [48]80.00%(+30.00)97.42%(+43.68)83.33%(+33.33)83.33%(+58.33)60.00%(+60.00)100.00%(+100.00)
", + "image_path": "42739651164fe353323c6279da5dce08e66572740d45c8b707abe18c536a5005.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 67, + 356, + 280, + 481 + ], + "blocks": [ + { + "bbox": [ + 55, + 287, + 296, + 353 + ], + "lines": [ + { + "bbox": [ + 55, + 287, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 296, + 353 + ], + "type": "text", + "content": "Table 2. SOTA Comparison on Face-Manipulated Data: We compared the performance of UNITE with recent DeepFake detectors, in terms of detection accuracy on various face manipulated datasets. UNITE outperforms the existing methods. Bold shows the current best results and the previous best and second-best results are highlighted in red and blue respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 356, + 280, + 481 + ], + "lines": [ + { + "bbox": [ + 67, + 356, + 280, + 481 + ], + "spans": [ + { + "bbox": [ + 67, + 356, + 280, + 481 + ], + "type": "table", + "html": "
MethodFF++CelebDFDeeperForensicsUADFV
TALL [57]98.65%90.79%99.62%-
ISTVT [62]99.00%84.10%98.60%-
Concas et al. [13]99.49%---
PUDD [40]-95.10%--
TI2Net [32]99.95%68.22%76.08%-
LRNet [46]99.89%53.20%56.77%-
Choi et al. [11]-89.00%99.00%-
Lin et al. [31]98.28%74.42%--
Guo et al. [21]99.24%84.97%--
Li et al. (Res-152) [29]---93.80%
HeadPose [58]---89.00%
CViT [56]93.00%--93.75%
FakeCatcher [12]94.65%91.50%--
MesoNet [3]---82.40%
UNITE (Ours)99.96%95.11%99.62%97.01%
", + "image_path": "026a53759fa5e9c90ff0e7f220988e58d42c475cf1f306716a2082362aedd22d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 484, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 296, + 555 + ], + "type": "text", + "content": "prediction (\"fake\") with high confidence, but on the row-2 example (video generated from the Sora [6] T2V model) the model trained with only CE loss predicted \"real\" with a confidence score of " + }, + { + "bbox": [ + 55, + 484, + 296, + 555 + ], + "type": "inline_equation", + "content": "99.25\\%" + }, + { + "bbox": [ + 55, + 484, + 296, + 555 + ], + "type": "text", + "content": " (which is wrong), but the model trained with " + }, + { + "bbox": [ + 55, + 484, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\mathrm{CE + AD}" + }, + { + "bbox": [ + 55, + 484, + 296, + 555 + ], + "type": "text", + "content": " losses predicted \"fake\" with a confidence of " + }, + { + "bbox": [ + 55, + 484, + 296, + 555 + ], + "type": "inline_equation", + "content": "100.00\\%" + }, + { + "bbox": [ + 55, + 484, + 296, + 555 + ], + "type": "text", + "content": ", which is correct." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 557, + 296, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 296, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 296, + 652 + ], + "type": "text", + "content": "Comparison on Face Manipulated Data: UNITE is compared with recent state-of-the-art (SOTA) DeepFake detection models in Table 2. The methods we compared against were specifically designed for detecting face manipulations and often crop faces from the videos, during both training and inference of the models. However, UNITE, which is designed to detect a diverse range of fake videos, still outperforms these detectors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Comparison on Synthetic Data: We evaluated UNITE against state-of-the-art synthetic video detectors on the DeMamba dataset [9] (validation split), with results in Table 3. While existing detectors were trained on the DeMamba train split and validated on the DeMamba validation split," + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 322, + 285, + 396, + 326 + ], + "blocks": [ + { + "bbox": [ + 322, + 285, + 396, + 326 + ], + "lines": [ + { + "bbox": [ + 322, + 285, + 396, + 326 + ], + "spans": [ + { + "bbox": [ + 322, + 285, + 396, + 326 + ], + "type": "image", + "image_path": "c22e4507b864d24b58be75ce23f856e165498a1d17c85f7886055912690090c3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 327, + 391, + 335 + ], + "lines": [ + { + "bbox": [ + 328, + 327, + 391, + 335 + ], + "spans": [ + { + "bbox": [ + 328, + 327, + 391, + 335 + ], + "type": "text", + "content": "(a) Original Frame" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 398, + 285, + 471, + 326 + ], + "blocks": [ + { + "bbox": [ + 398, + 285, + 471, + 326 + ], + "lines": [ + { + "bbox": [ + 398, + 285, + 471, + 326 + ], + "spans": [ + { + "bbox": [ + 398, + 285, + 471, + 326 + ], + "type": "image", + "image_path": "9a7f55b96974cf7b358b430d9ef5d44c715dd9f04ae58d9d55339da946fc69ed.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 327, + 462, + 335 + ], + "lines": [ + { + "bbox": [ + 406, + 327, + 462, + 335 + ], + "spans": [ + { + "bbox": [ + 406, + 327, + 462, + 335 + ], + "type": "text", + "content": "(b) CE Loss only" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 473, + 285, + 547, + 326 + ], + "blocks": [ + { + "bbox": [ + 473, + 285, + 547, + 326 + ], + "lines": [ + { + "bbox": [ + 473, + 285, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 473, + 285, + 547, + 326 + ], + "type": "image", + "image_path": "a73dc7545cdac92a63c283887cef990df632ac6489198467c2b382e0b2d6c15f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 482, + 327, + 538, + 335 + ], + "lines": [ + { + "bbox": [ + 482, + 327, + 538, + 335 + ], + "spans": [ + { + "bbox": [ + 482, + 327, + 538, + 335 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 482, + 327, + 538, + 335 + ], + "type": "inline_equation", + "content": "\\mathrm{CE + AD}" + }, + { + "bbox": [ + 482, + 327, + 538, + 335 + ], + "type": "text", + "content": " Loss" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 323, + 336, + 396, + 378 + ], + "blocks": [ + { + "bbox": [ + 323, + 336, + 396, + 378 + ], + "lines": [ + { + "bbox": [ + 323, + 336, + 396, + 378 + ], + "spans": [ + { + "bbox": [ + 323, + 336, + 396, + 378 + ], + "type": "image", + "image_path": "66b04e8a8f21be849327484594cfd6b4ecf64172824ab28252ce8e38960c5b84.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 380, + 391, + 389 + ], + "lines": [ + { + "bbox": [ + 329, + 380, + 391, + 389 + ], + "spans": [ + { + "bbox": [ + 329, + 380, + 391, + 389 + ], + "type": "text", + "content": "(d) Original Frame" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 398, + 336, + 471, + 378 + ], + "blocks": [ + { + "bbox": [ + 398, + 336, + 471, + 378 + ], + "lines": [ + { + "bbox": [ + 398, + 336, + 471, + 378 + ], + "spans": [ + { + "bbox": [ + 398, + 336, + 471, + 378 + ], + "type": "image", + "image_path": "c6f70e908b9eb1083ccd7c19c4d8aafcce441fe1ac95d4a3093f9532bc7b1cf2.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 380, + 462, + 389 + ], + "lines": [ + { + "bbox": [ + 406, + 380, + 462, + 389 + ], + "spans": [ + { + "bbox": [ + 406, + 380, + 462, + 389 + ], + "type": "text", + "content": "(e) CE Loss only" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 473, + 336, + 547, + 378 + ], + "blocks": [ + { + "bbox": [ + 473, + 336, + 547, + 378 + ], + "lines": [ + { + "bbox": [ + 473, + 336, + 547, + 378 + ], + "spans": [ + { + "bbox": [ + 473, + 336, + 547, + 378 + ], + "type": "image", + "image_path": "f78bfdc6431dce203fc1738fdbc24287c4c17126ed02ee3507b3c365f47ad81c.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 482, + 380, + 537, + 389 + ], + "lines": [ + { + "bbox": [ + 482, + 380, + 537, + 389 + ], + "spans": [ + { + "bbox": [ + 482, + 380, + 537, + 389 + ], + "type": "text", + "content": "(f) " + }, + { + "bbox": [ + 482, + 380, + 537, + 389 + ], + "type": "inline_equation", + "content": "\\mathrm{CE + AD}" + }, + { + "bbox": [ + 482, + 380, + 537, + 389 + ], + "type": "text", + "content": " Loss" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 390, + 555, + 500 + ], + "lines": [ + { + "bbox": [ + 313, + 390, + 555, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 555, + 500 + ], + "type": "text", + "content": "Figure 3. Comparison of attention heatmaps when UNITE is trained with only CE loss (second column) versus CE and AD-losses (third column). The CE-loss heatmap predominantly focuses on the face region, whereas the " + }, + { + "bbox": [ + 313, + 390, + 555, + 500 + ], + "type": "inline_equation", + "content": "\\mathrm{CE + AD}" + }, + { + "bbox": [ + 313, + 390, + 555, + 500 + ], + "type": "text", + "content": " loss heatmap demonstrates more distributed attention across the entire frame, as indicated by a broader bluish tone in the heatmaps. Both samples are \"fake\" videos- (a) is taken from Celeb-DF [30] and (d) is taken from DeMamba [9] and generated by OpenAI's Sora [6] (Best viewed as GIFs provided in the supplementary material). Yellow to blue shades show increasing attention in the plot." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 506, + 555, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 554 + ], + "type": "text", + "content": "UNITE was trained on FF++ [43] and GTA-V [24]. Despite not being trained on DeMamba, UNITE's average performance (including real videos) surpassed current SOTA detectors, with competitive generator-wise results." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 555, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 555, + 639 + ], + "type": "text", + "content": "Thus, UNITE reliably detects a diverse range of fake videos (even in cross-domain settings), including T2V/I2V videos and background manipulations, while consistently maintaining state-of-the-art performance on traditional face-manipulated data. UNITE eliminates the need to have separate DeepFake and T2V/I2V video detector models by by handling both tasks within a single trained model." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 647, + 448, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 448, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 448, + 659 + ], + "type": "text", + "content": "4.3. Finegrained Evaluation" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "In fake media detection, flagging videos partially or fully manipulated by AI also necessitates fine-grained classification into three classes: (1) real videos, (2) partially manipulated videos, and (3) fully synthetic videos. While binary" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28055" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 137, + 538, + 326 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 136 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 136 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 136 + ], + "type": "text", + "content": "Table 3. SOTA Comparison on Synthetic Data: On the DeMamba dataset (validation split), we compare the performance of UNITE, which was NOT trained on DeMamba train split, against state-of-the-art detectors which were trained on DeMamba train split (results taken from Chen et al. [9]). We report the results (" + }, + { + "bbox": [ + 55, + 70, + 555, + 136 + ], + "type": "inline_equation", + "content": "P = \\text{Precision}" + }, + { + "bbox": [ + 55, + 70, + 555, + 136 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 70, + 555, + 136 + ], + "type": "inline_equation", + "content": "R = \\text{Recall}" + }, + { + "bbox": [ + 55, + 70, + 555, + 136 + ], + "type": "text", + "content": ") on the individual T2V/I2V generators and the average performance across the entire validation set (" + }, + { + "bbox": [ + 55, + 70, + 555, + 136 + ], + "type": "inline_equation", + "content": "Avg" + }, + { + "bbox": [ + 55, + 70, + 555, + 136 + ], + "type": "text", + "content": ", which also includes real videos). Although the direct comparison is unfair against UNITE which was trained with FF++ [43] and GTA-V [24], our method still outperforms these synthetic video detectors. Bold shows the current best results and the previous best and second-best results are highlighted in red and blue respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 137, + 538, + 326 + ], + "lines": [ + { + "bbox": [ + 67, + 137, + 538, + 326 + ], + "spans": [ + { + "bbox": [ + 67, + 137, + 538, + 326 + ], + "type": "table", + "html": "
MethodMetricsSora [6]Morph Studio [2]Runway ML (Gen2) [42]HotShot [38]Lavie [54]Show-1 [60]Moon Valley [37]Crafter [8]Model Scope [52]Wild Scrape [9]Avg
TALL [57]P71.15%96.89%98.51%79.38%84.59%79.38%98.79%99.02%92.70%76.47%87.91%
R91.07%98.28%97.83%83.00%76.57%79.57%99.52%98.93%94.14%66.31%88.52%
F3Net [41]P68.27%99.89%99.67%89.35%57.00%36.57%99.52%99.71%93.80%88.41%88.73%
R83.93%99.71%98.62%77.57%85.24%63.17%99.58%99.89%89.43%76.78%81.88%
NPR [47]P91.07%99.57%99.49%24.29%89.64%57.71%97.12%99.86%94.29%87.80%82.45%
R91.07%99.57%99.49%24.29%89.64%57.71%97.12%99.86%94.29%87.80%84.08%
STIL [20]P57.21%99.08%99.32%86.19%82.24%70.43%99.25%98.96%97.18%81.32%87.12%
R67.86%96.00%98.41%96.14%77.14%80.43%97.44%96.93%96.29%68.36%82.22%
MINTIME-CLIP-B [9]P83.21%99.99%99.67%50.84%99.20%99.27%99.76%99.99%91.83%91.77%91.55%
R89.29%100.00%98.99%26.43%96.79%98.14%99.84%100.00%84.29%82.38%87.62%
FTCN-CLIP-B [9]P91.79%99.99%99.79%45.94%99.76%97.80%99.99%99.99%94.69%92.32%92.21%
R87.50%100.00%98.91%17.71%97.71%91.86%100.00%100.00%85.29%82.83%86.18%
CLIP-B-PT [9]P67.80%43.56%70.88%29.97%52.97%35.36%55.52%66.03%44.23%42.99%44.83%
R85.71%82.43%90.36%71.00%79.29%75.43%89.62%86.29%82.14%75.16%81.74%
DeMamba-CLIP-PT [9]P25.87%95.14%96.23%73.43%83.31%75.49%90.17%95.06%95.05%69.95%79.97%
R58.93%96.43%93.12%68.00%69.36%69.00%89.14%91.86%96.14%56.59%78.86%
XCLIP-B-PT [9]P16.39%72.16%87.77%39.86%65.57%54.26%75.23%84.80%61.60%55.28%61.29%
R81.34%82.15%83.35%80.98%81.82%81.55%82.14%82.98%81.93%81.10%81.93%
DeMamba-XCLIP-PT [9]P18.26%93.50%94.72%69.94%78.08%71.50%83.95%92.23%93.54%68.10%76.38%
R66.07%95.86%94.64%77.86%75.36%80.29%90.89%92.50%96.00%66.41%83.59%
XCLIP-B-FT [9]P64.42%99.73%96.78%70.98%90.35%77.28%97.34%99.84%82.01%88.97%86.77%
R82.14%99.57%93.62%61.29%79.36%69.71%97.92%99.79%77.14%83.59%84.41%
UNITE (Ours)P88.57%100.00%100.00%90.16%89.91%98.34%99.52%100.00%98.96%92.56%92.76%
R92.11%100.00%94.62%96.93%98.12%99.86%98.69%100.00%96.29%89.89%89.60%
", + "image_path": "285d59e5e6e07288f95b3fdfee0aaabba479d0a223322ee43e05064bdcead95c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 91, + 367, + 256, + 527 + ], + "blocks": [ + { + "bbox": [ + 55, + 330, + 295, + 363 + ], + "lines": [ + { + "bbox": [ + 55, + 330, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 295, + 363 + ], + "type": "text", + "content": "Table 4. Results obtained by the UNITE model on finegrained classes: detecting whether a video is real, partially manipulated or fully AI-generated. Performance gains are mentioned in green." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 91, + 367, + 256, + 527 + ], + "lines": [ + { + "bbox": [ + 91, + 367, + 256, + 527 + ], + "spans": [ + { + "bbox": [ + 91, + 367, + 256, + 527 + ], + "type": "table", + "html": "
TrainTest
FF++GTA-VDatasetAccuracy
✓FF++96.46%
✓CelebDF60.40%
✓DeeperForensics71.04%
✓DeepFakeTIMIT80.68%
✓HifiFace43.27%
✓UADFV90.37%
✓AVID50.00%
✓GTA-V0.00%
✓DeMamba38.29%
✓✓FF++97.70%(+1.24)
✓✓CelebDF70.17%(+9.77)
✓✓DeeperForensics80.59%(+9.55)
✓✓DeepFakeTIMIT81.04%(+0.36)
✓✓HifiFace64.20%(+20.93)
✓✓UADFV92.61%(+1.24)
✓✓AVID62.50%(+12.50)
✓✓GTA-V100.00%(+100.00)
✓✓DeMamba65.84%(+27.55)
", + "image_path": "06feaeed88e23c1a4a5632e9ccdd54161f908fe2f012c75ed66e2b745f5ebb78.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 532, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 567 + ], + "type": "text", + "content": "real vs. fake classification may suffice to curb misinformation, fine-grained classification offers greater explainability in otherwise black-box models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "Using this defined convention of finegrained classes, all the face manipulated datasets (CelebDF [30], DeeperForensics [25], etc.) and the videos from the AVID [61] model (background manipulations) fall under class-2, and the videos from the GTA-V [24] and the DeMamba [9] datasets fall in class-3. In this setup, only the final classification layer of the UNITE transformer architecture is changed to 3 neurons with softmax activation. The results obtained are presented in Table 4. As expected, when the model is trained only with " + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{FF}++" + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": " [43], the performance on fully synthetic data is poor (since the model has not seen any synthetic videos). When the GTA-V [24] data is also" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 331, + 333, + 539, + 403 + ], + "blocks": [ + { + "bbox": [ + 331, + 333, + 539, + 403 + ], + "lines": [ + { + "bbox": [ + 331, + 333, + 539, + 403 + ], + "spans": [ + { + "bbox": [ + 331, + 333, + 539, + 403 + ], + "type": "image", + "image_path": "4e0ee692c4c5de2ef0684ec2901942f259d3d5e21d2f9aed9f0a466693c70f68.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 405, + 495, + 415 + ], + "lines": [ + { + "bbox": [ + 373, + 405, + 495, + 415 + ], + "spans": [ + { + "bbox": [ + 373, + 405, + 495, + 415 + ], + "type": "text", + "content": "(a) UNITE trained on " + }, + { + "bbox": [ + 373, + 405, + 495, + 415 + ], + "type": "inline_equation", + "content": "\\mathrm{FF} + +" + }, + { + "bbox": [ + 373, + 405, + 495, + 415 + ], + "type": "text", + "content": " [43] only" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 331, + 418, + 538, + 488 + ], + "blocks": [ + { + "bbox": [ + 331, + 418, + 538, + 488 + ], + "lines": [ + { + "bbox": [ + 331, + 418, + 538, + 488 + ], + "spans": [ + { + "bbox": [ + 331, + 418, + 538, + 488 + ], + "type": "image", + "image_path": "731f76544b4eb7ad0882ec6ef0bec3dce866c8aa404905fbfc7f87320a267252.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 491, + 515, + 499 + ], + "lines": [ + { + "bbox": [ + 354, + 491, + 515, + 499 + ], + "spans": [ + { + "bbox": [ + 354, + 491, + 515, + 499 + ], + "type": "text", + "content": "(b) UNITE trained on " + }, + { + "bbox": [ + 354, + 491, + 515, + 499 + ], + "type": "inline_equation", + "content": "\\mathrm{FF} + +" + }, + { + "bbox": [ + 354, + 491, + 515, + 499 + ], + "type": "text", + "content": " [43] and GTA-V [24]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 500, + 554, + 555 + ], + "lines": [ + { + "bbox": [ + 313, + 500, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 500, + 554, + 555 + ], + "type": "text", + "content": "Figure 4. Ablation Results to show the effect of changing the loss functions used to train the UNITE model. The combination of the cross-entropy (CE) and Attention Diversity (AD) losses always performs better, and the contribution from the AD-loss component increases significantly when using fully synthetic data for training." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 563, + 512, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 512, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 512, + 575 + ], + "type": "text", + "content": "used in training, the performance is much higher." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 586, + 408, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 408, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 408, + 599 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "We conducted an analysis to assess the contribution of individual loss functions to the training of the UNITE model. The outcomes of this ablation study are presented in Fig. 4. Specifically, we evaluated the impact of the AD-loss in two scenarios: (1) when UNITE was trained solely on the FF++ dataset [43] (Fig. 4(a)), and (2) when UNITE was trained using a combination of the FF++ [43] and the GTA-V [24] datasets (Fig. 4(b)). The combination of the two losses always performed better than the individual losses." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28056" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 74, + 267, + 182 + ], + "blocks": [ + { + "bbox": [ + 86, + 74, + 267, + 182 + ], + "lines": [ + { + "bbox": [ + 86, + 74, + 267, + 182 + ], + "spans": [ + { + "bbox": [ + 86, + 74, + 267, + 182 + ], + "type": "image", + "image_path": "e2581f9420b0e9c89632475fbb1f99e00a2b12ff8998a84ceffb81c5cc3e3920.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 182, + 295, + 248 + ], + "lines": [ + { + "bbox": [ + 55, + 182, + 295, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 295, + 248 + ], + "type": "text", + "content": "Figure 5. No. of frames vs. Performance: Performance analysis of UNITE based on the number of frames sampled per video segment. The results illustrate that as the number of frames increases from 1 to 64 (context window), the detection accuracy improves, showcasing UNITE's ability to effectively capture temporal inconsistencies in fake videos." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 256, + 295, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 295, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 295, + 340 + ], + "type": "text", + "content": "Effect on Synthetic Data: UNITE performed significantly better on the fully synthetic datasets GTA-V [24] and DeMamba [9] when the model was training includes GTA-V [24] data. However, for the DeMamba dataset (cross-data evaluation), the performance with only CE-loss is similar regardless of the training data. It is the AD-loss component that provides the accuracy boost in Fig 4(b)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 341, + 295, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 295, + 520 + ], + "type": "text", + "content": "Effect on Background-Manipulated Data: In Fig. 4(a), we observe that when UNITE was trained solely on " + }, + { + "bbox": [ + 55, + 341, + 295, + 520 + ], + "type": "inline_equation", + "content": "\\mathrm{FF}++" + }, + { + "bbox": [ + 55, + 341, + 295, + 520 + ], + "type": "text", + "content": " [43] using only the CE-loss, its performance on the AVID dataset [61] was " + }, + { + "bbox": [ + 55, + 341, + 295, + 520 + ], + "type": "inline_equation", + "content": "0.00\\%" + }, + { + "bbox": [ + 55, + 341, + 295, + 520 + ], + "type": "text", + "content": ". This outcome can be attributed to the fact that the fake data in " + }, + { + "bbox": [ + 55, + 341, + 295, + 520 + ], + "type": "inline_equation", + "content": "\\mathrm{FF}++" + }, + { + "bbox": [ + 55, + 341, + 295, + 520 + ], + "type": "text", + "content": " [43] contains exclusively face-manipulated videos, causing the UNITE transformer's attention heads to focus predominantly on the facial regions. As AVID [61] comprises only background-manipulated videos, UNITE failed to detect any of the fakes under such conditions. However, when the AD-loss was introduced, the attention heads of UNITE were encouraged to diversify their spatial focus, significantly improving performance. Incorporating synthetic data, as shown in Fig. 4(b), further enhanced this effect, achieving " + }, + { + "bbox": [ + 55, + 341, + 295, + 520 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 55, + 341, + 295, + 520 + ], + "type": "text", + "content": " accuracy on AVID when using both CE-loss and AD-loss." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 521, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 688 + ], + "type": "text", + "content": "Effect on Face-Manipulated Data: For the evaluation on face-manipulated datasets such as CelebDF [30] and UADFV [58], one might expect comparable results between training solely on " + }, + { + "bbox": [ + 55, + 521, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\mathrm{FF} + +" + }, + { + "bbox": [ + 55, + 521, + 295, + 688 + ], + "type": "text", + "content": " [43] and incorporating GTA-V [24] due to their perceived irrelevance in evaluating face-centric manipulations. However, as shown in Fig. 4(b), training with fully synthetic data leads to a substantial performance improvement. Notably, in both Fig. 4(a) and (b), using only CE loss yields similar results, but the inclusion of the AD-loss component significantly enhances overall detection accuracy. The diversity in the training data makes the UNITE model learn more discriminative attention map features through the AD-loss component, which in turn helps even in detecting face-manipulated videos." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Frames vs. Performance: For inference with UNITE, we evaluated the impact of temporal context by vary-" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 345, + 75, + 526, + 183 + ], + "blocks": [ + { + "bbox": [ + 345, + 75, + 526, + 183 + ], + "lines": [ + { + "bbox": [ + 345, + 75, + 526, + 183 + ], + "spans": [ + { + "bbox": [ + 345, + 75, + 526, + 183 + ], + "type": "image", + "image_path": "69c7bd286988af9c27ba996e301c20fc5c1e44c982413c17694203378e6c34a8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 187, + 555, + 254 + ], + "lines": [ + { + "bbox": [ + 313, + 187, + 555, + 254 + ], + "spans": [ + { + "bbox": [ + 313, + 187, + 555, + 254 + ], + "type": "text", + "content": "Figure 6. Transformer Depth Evaluation: Performance comparison of UNITE with varying the number of encoder blocks (depth). UNITE performs optimally in cross-domain settings when the depth is 4. Greater depths overfit to the training domains " + }, + { + "bbox": [ + 313, + 187, + 555, + 254 + ], + "type": "inline_equation", + "content": "(\\mathrm{FF} + +" + }, + { + "bbox": [ + 313, + 187, + 555, + 254 + ], + "type": "text", + "content": " [43] and GTA-V [24]), while a depth of 2 is insufficient to capture the complexity of the data." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 261, + 555, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 555, + 344 + ], + "type": "text", + "content": "ing the number of frames sampled from video segments to " + }, + { + "bbox": [ + 313, + 261, + 555, + 344 + ], + "type": "inline_equation", + "content": "\\{1,8,16,32,64\\}" + }, + { + "bbox": [ + 313, + 261, + 555, + 344 + ], + "type": "text", + "content": ", with 64 frames corresponding to our model's context window. As shown in Fig. 5, performance improves as more frames are included, demonstrating that UNITE effectively captures temporal discontinuities in fake videos. This highlights the model's ability to leverage temporal cues for enhanced detection accuracy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 345, + 556, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 556, + 465 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 556, + 465 + ], + "type": "text", + "content": "Transformer Depth Ablation: We evaluate the impact of varying the depth of the transformer architecture in UNITE by adjusting the number of encoder blocks, with results shown in Fig. 6. We tested depths of " + }, + { + "bbox": [ + 313, + 345, + 556, + 465 + ], + "type": "inline_equation", + "content": "\\{2,4,6,8\\}" + }, + { + "bbox": [ + 313, + 345, + 556, + 465 + ], + "type": "text", + "content": " and observed that the model achieved the best performance with 4 encoder blocks, which is the depth used in our final architecture. This suggests that a moderate depth strikes the optimal balance between model complexity and performance, especially in cross-dataset settings, providing robust detection capabilities without overfitting." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 476, + 388, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 388, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 388, + 489 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 497, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 554, + 593 + ], + "type": "text", + "content": "Traditional DeepFake detectors, focused solely on face-manipulated content, struggle against newer, more sophisticated forgery methods involving T2V, I2V, and background manipulations. To address this, we proposed UNITE, leveraging the fully synthetic GTA-V [24] dataset for training. While this synthetic data is not AI-generated, our results show it effectively enhances UNITE's ability to detect AI-generated content and background manipulations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "type": "text", + "content": "A key innovation in our approach is the AD-loss, which encourages the model to explore diverse spatial regions, enhancing its detection of manipulated areas beyond just the face region. This allows UNITE to excel not only in detecting AI-generated and background-altered videos but also in identifying traditional face-manipulated content. Our ablation studies demonstrate the robustness of UNITE, making it a valuable tool in synthetic video detection." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "Acknowledgements: This work was supported in part by funding from YouTube (Google)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28057" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 291, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 291, + 102 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 291, + 102 + ], + "type": "text", + "content": "[1] Flux 1.1. https://blackforestlabs.ai/, 2024.1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 103, + 295, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 295, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 295, + 124 + ], + "type": "text", + "content": "[2] Morph studio. https://www.morphstudio.com/, 2024.7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 127, + 295, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 127, + 295, + 181 + ], + "spans": [ + { + "bbox": [ + 62, + 127, + 295, + 181 + ], + "type": "text", + "content": "[3] Darius Afchar, Vincent Nozick, Junichi Yamagishi, and Isao Echizen. Mesonet: a compact facial video forgery detection network. In 2018 IEEE international workshop on information forensics and security (WIFS), pages 1-7. IEEE, 2018. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 183, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 183, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 62, + 183, + 295, + 228 + ], + "type": "text", + "content": "[4] Ibrahim M Alabdulmohsin, Xiaohua Zhai, Alexander Kolesnikov, and Lucas Beyer. Getting vit in shape: Scaling laws for compute-optimal model design. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 229, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 229, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 229, + 295, + 282 + ], + "type": "text", + "content": "[5] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 285, + 295, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 295, + 340 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 295, + 340 + ], + "type": "text", + "content": "[6] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, et al. Video generation models as world simulators. 2024. URL https://openai.com/research/video-generation-models-as-world-simulators, 3, 2024. 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 342, + 295, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 295, + 385 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 295, + 385 + ], + "type": "text", + "content": "[7] Baoying Chen and Shunquan Tan. Featuretransfer: Unsupervised domain adaptation for cross-domain deepfake detection. Security and Communication Networks, 2021(1): 9942754, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 388, + 295, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 295, + 442 + ], + "type": "text", + "content": "[8] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 445, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 295, + 498 + ], + "type": "text", + "content": "[9] Haoxing Chen, Yan Hong, Zizheng Huang, Zhuoer Xu, Zhangxuan Gu, Yaohui Li, Jun Lan, Huijia Zhu, Jianfu Zhang, Weiqiang Wang, et al. Demamba: Ai-generated video detection on million-scale genvideo benchmark. arXiv preprint arXiv:2405.19707, 2024. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 501, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 501, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 57, + 501, + 295, + 544 + ], + "type": "text", + "content": "[10] Jikang Cheng, Zhiyuan Yan, Ying Zhang, Yuhao Luo, Zhongyuan Wang, and Chen Li. Can we leave deepfake data behind in training deepfake detector? arXiv preprint arXiv:2408.17052, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 546, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 546, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 546, + 295, + 601 + ], + "type": "text", + "content": "[11] Jongwook Choi, Taehoon Kim, Yonghyun Jeong, Seungryul Baek, and Jongwon Choi. Exploiting style latent flows for generalizing deepfake video detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1133-1143, 2024. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 602, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 646 + ], + "type": "text", + "content": "[12] Umur Aybars Ciftci, Ilke Demir, and Lijun Yin. Fakecatcher: Detection of synthetic portrait videos using biological signals. IEEE transactions on pattern analysis and machine intelligence, 2020. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 712 + ], + "type": "text", + "content": "[13] Sara Concas, Simone Maurizio La Cava, Roberto Casula, Giulia Orrù, Giovanni Puglisi, and Gian Luca Marcialis. Quality-based artifact modeling for facial deepfake detection in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3845–3854, 2024. 2, 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "type": "text", + "content": "[14] Riccardo Corvi, Davide Cozzolino, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. Intriguing properties of synthetic images: from generative adversarial networks to diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 973-982, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 140, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 205 + ], + "type": "text", + "content": "[15] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 208, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 208, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 317, + 208, + 553, + 261 + ], + "type": "text", + "content": "[16] Davide Cozzolino, Andreas Rössler, Justus Thies, Matthias Nießner, and Luisa Verdoliva. Id-reveal: Identity-aware deepfake video detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15108-15117, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 264, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 264, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 317, + 264, + 553, + 306 + ], + "type": "text", + "content": "[17] Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. The deepfake detection challenge (dfdc) dataset. arXiv preprint arXiv:2006.07397, 2020. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 309, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 309, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 317, + 309, + 553, + 363 + ], + "type": "text", + "content": "[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 365, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 365, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 317, + 365, + 553, + 397 + ], + "type": "text", + "content": "[19] Ricard Durall, Margret Keuper, Franz-Josef Pfundt, and Janis Keuper. Unmasking deepfakes with simple features. arXiv preprint arXiv:1911.00686, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 399, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 453 + ], + "type": "text", + "content": "[20] Zhihao Gu, Yang Chen, Taiping Yao, Shouhong Ding, Jilin Li, Feiyue Huang, and Lizhuang Ma. Spatiotemporal inconsistency learning for deepfake video detection. In Proceedings of the 29th ACM international conference on multimedia, pages 3473-3481, 2021. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 456, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 498 + ], + "type": "text", + "content": "[21] Ying Guo, Cheng Zhen, and Pengfei Yan. Controllable guide-space for generalizable face forgery detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20818-20827, 2023. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 501, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 501, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 553, + 520 + ], + "type": "text", + "content": "[22] Harald Hanselmann, Shen Yan, and Hermann Ney. Deep fisher faces. In BMVC, 2017. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 524, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 524, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 524, + 553, + 544 + ], + "type": "text", + "content": "[23] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 547, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 547, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 317, + 547, + 553, + 611 + ], + "type": "text", + "content": "[24] Yuan-Ting Hu, Jiahong Wang, Raymond A Yeh, and Alexander G Schwing. Sail-vos 3d: A synthetic dataset and baselines for object detection and 3d mesh reconstruction from video data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1418-1428, 2021. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 614, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 553, + 668 + ], + "type": "text", + "content": "[25] Liming Jiang, Ren Li, Wayne Wu, Chen Qian, and Chen Change Loy. Deeperforensics-1.0: A large-scale dataset for real-world face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2889–2898, 2020. 5, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 553, + 712 + ], + "type": "text", + "content": "[26] Yan Ju, Chengzhe Sun, Shan Jia, Shuwei Hou, Zhaofeng Si, Soumyya Kanti Datta, Lipeng Ke, Riky Zhou, Anita Nikolich, and Siwei Lyu. Deepfake-o-meter v2.0: An open platform for deepfake detection. In 2024 IEEE 7th Inter-" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28058" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "national Conference on Multimedia Information Processing and Retrieval (MIPR), pages 439-445. IEEE, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 97, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 97, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 57, + 97, + 295, + 129 + ], + "type": "text", + "content": "[27] Pavel Korshunov and Sébastien Marcel. Deepfakes: a new threat to face recognition? assessment and detection. arXiv preprint arXiv:1812.08685, 2018. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 132, + 294, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 132, + 294, + 164 + ], + "spans": [ + { + "bbox": [ + 57, + 132, + 294, + 164 + ], + "type": "text", + "content": "[28] Lingzhi Li, Jianmin Bao, Hao Yang, Dong Chen, and Fang Wen. Faceshifter: Towards high fidelity and occlusion aware face swapping. arXiv preprint arXiv:1912.13457, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 167, + 294, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 167, + 294, + 189 + ], + "spans": [ + { + "bbox": [ + 57, + 167, + 294, + 189 + ], + "type": "text", + "content": "[29] Y Li. Exposing deepfake videos by detecting face warping artifact acts. arXiv preprint arXiv:1811.00656, 2018. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 190, + 295, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 190, + 295, + 244 + ], + "spans": [ + { + "bbox": [ + 56, + 190, + 295, + 244 + ], + "type": "text", + "content": "[30] Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3207-3216, 2020. 1, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 247, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 247, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 56, + 247, + 295, + 300 + ], + "type": "text", + "content": "[31] Li Lin, Xinan He, Yan Ju, Xin Wang, Feng Ding, and Shu Hu. Preserving fairness generalization in deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16815-16825, 2024. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 304, + 295, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 295, + 357 + ], + "type": "text", + "content": "[32] Baoping Liu, Bo Liu, Ming Ding, Tianqing Zhu, and Xin Yu. Ti2net: temporal identity inconsistency network for deepfake detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4691-4700, 2023. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 360, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 360, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 57, + 360, + 294, + 381 + ], + "type": "text", + "content": "[33] I Loshchilov and F Hutter. Decoupled weight decay regularization. *ICLR*, 2019. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 384, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 384, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 57, + 384, + 294, + 437 + ], + "type": "text", + "content": "[34] Qingxuan Lv, Yuezun Li, Junyu Dong, Sheng Chen, Hui Yu, Huiyu Zhou, and Shu Zhang. Domainforensics: Exposing face forgery across domains via bi-directional adaptation. IEEE Transactions on Information Forensics and Security, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 441, + 294, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 441, + 294, + 484 + ], + "spans": [ + { + "bbox": [ + 57, + 441, + 294, + 484 + ], + "type": "text", + "content": "[35] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 487, + 294, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 487, + 294, + 531 + ], + "spans": [ + { + "bbox": [ + 57, + 487, + 294, + 531 + ], + "type": "text", + "content": "[36] Ghazal Mazaheri and Amit K Roy-Chowdhury. Detection and localization of facial expression manipulations. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1035-1045, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 533, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 294, + 553 + ], + "type": "text", + "content": "[37] moonvalley.ai. moonvalley.ai. https://moonvalley.ai/, 2022.7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 556, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 556, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 57, + 556, + 294, + 578 + ], + "type": "text", + "content": "[38] John Mullan, Duncan Crawbuck, and Aakash Sastry. Hotshot-XL, 2023. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 580, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 295, + 656 + ], + "type": "text", + "content": "[39] Dat Nguyen, Nesryne Mejri, Inder Pal Singh, Polina Kuleshova, Marcella Astrid, Anis Kacem, Enjie Ghorbel, and Djamila Aouada. Laa-net: Localized artifact attention network for quality-agnostic and generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17395–17405, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "text", + "content": "[40] Alvaro Lopez Pellicer, Yi Li, and Plamen Angelov. Pudd: Towards robust multi-modal prototype-based deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3809-3817, 2024. 2, 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[41] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In European conference on computer vision, pages 86-103. Springer, 2020. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 119, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 140 + ], + "type": "text", + "content": "[42] Runway Research. Text driven video generation. https://research.runwayml.com/gen2, 2023.1,7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 142, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 142, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 142, + 553, + 196 + ], + "type": "text", + "content": "[43] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics++: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1-11, 2019. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 198, + 553, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 242 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 242 + ], + "type": "text", + "content": "[44] Md Shamim Seraj, Ankita Singh, and Shayok Chakraborty. Semi-supervised deep domain adaptation for deepfake detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1061-1071, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "text", + "content": "[45] Ke Sun, Shen Chen, Taiping Yao, Hong Liu, Xiaoshuai Sun, Shouhong Ding, and Rongrong Ji. Diffusionfake: Enhancing generalization in deepfake detection via guided stable diffusion. NeurIPS, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 288, + 555, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 555, + 342 + ], + "type": "text", + "content": "[46] Zekun Sun, Yujie Han, Zeyu Hua, Na Ruan, and Weijia Jia. Improving the efficiency and robustness of deepfakes detection through precise geometric features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3609-3618, 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 344, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 553, + 408 + ], + "type": "text", + "content": "[47] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28130-28139, 2024. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "text", + "content": "[48] The New York Times. A.i. can now create lifelike videos. can you tell what's real? https://www.nytimes.com/interactive/2024/09/09/technology/ai-video-deepfake-runway-kling-quiz.html, 2024.5,6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 468, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 510 + ], + "type": "text", + "content": "[49] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. Acm Transactions on Graphics (TOG), 38(4):1-12, 2019. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 512, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 553, + 567 + ], + "type": "text", + "content": "[50] Loc Trinh, Michael Tsang, Sirisha Rambhatla, and Yan Liu. Interpretable and trustworthy deepfake detection via dynamic prototypes. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1973-1983, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 569, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 569, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 569, + 553, + 600 + ], + "type": "text", + "content": "[51] A Vaswani, N Shazeer, N Parmar, J Uszkoreit, L Jones, A Gomez, L Kaiser, and I Polosukhin. Attention is all you need. NeurIPS, 2017. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 602, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 645 + ], + "type": "text", + "content": "[52] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1, 2, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[53] Yuhan Wang, Xu Chen, Junwei Zhu, Wenqing Chu, Ying Tai, Chengjie Wang, Jilin Li, Yongjian Wu, Feiyue Huang, and Rongrong Ji. Hififace: 3d shape and semantic prior guided high fidelity face swapping. Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, 2021. 3, 5" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28059" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 631 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[54] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "text", + "content": "[55] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22445-22455, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 217 + ], + "type": "text", + "content": "[56] Deressa Wodajo and Solomon Atnafu. Deepfake video detection using convolutional vision transformer. arXiv preprint arXiv:2102.11126, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 273 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 273 + ], + "type": "text", + "content": "[57] Yuting Xu, Jian Liang, Gengyun Jia, Ziming Yang, Yanhao Zhang, and Ran He. Tall: Thumbnail layout for deepfake video detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22658-22668, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 319 + ], + "type": "text", + "content": "[58] Xin Yang, Yuezun Li, and Siwei Lyu. Exposing deep fakes using inconsistent head poses. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 8261-8265. IEEE, 2019. 5, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 363 + ], + "type": "text", + "content": "[59] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 364, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 364, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 294, + 419 + ], + "type": "text", + "content": "[60] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. International Journal of Computer Vision, pages 1–15, 2024. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 420, + 294, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 420, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 56, + 420, + 294, + 485 + ], + "type": "text", + "content": "[61] Zhixing Zhang, Bichen Wu, Xiaoyan Wang, Yaqiao Luo, Luxin Zhang, Yinan Zhao, Peter Vajda, Dimitris Metaxas, and Licheng Yu. Avid: Any-length video inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7162-7172, 2024. 1, 4, 5, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 487, + 294, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 294, + 540 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 294, + 540 + ], + "type": "text", + "content": "[62] Cairong Zhao, Chutian Wang, Guosheng Hu, Haonan Chen, Chun Liu, and Jinhui Tang. Istvt: interpretable spatial-temporal video transformer for deepfake detection. IEEE Transactions on Information Forensics and Security, 18: 1335-1348, 2023. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 542, + 294, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 542, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 542, + 294, + 586 + ], + "type": "text", + "content": "[63] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 587, + 294, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 587, + 294, + 631 + ], + "spans": [ + { + "bbox": [ + 56, + 587, + 294, + 631 + ], + "type": "text", + "content": "[64] Xinye Zhou, Hu Han, Shiguang Shan, and Xin Chen. Fine-grained open-set deepfake detection via unsupervised domain adaptation. IEEE Transactions on Information Forensics and Security, 2024. 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28060" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Tra-MoE_ Learning Trajectory Prediction Model from Multiple Domains for Adaptive Policy Conditioning/d41e3854-0a29-4aee-85fc-0fcf3b667c89_content_list.json b/2025/Tra-MoE_ Learning Trajectory Prediction Model from Multiple Domains for Adaptive Policy Conditioning/d41e3854-0a29-4aee-85fc-0fcf3b667c89_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e65796220cdeac19fd4dd02c70f985bc01597146 --- /dev/null +++ b/2025/Tra-MoE_ Learning Trajectory Prediction Model from Multiple Domains for Adaptive Policy Conditioning/d41e3854-0a29-4aee-85fc-0fcf3b667c89_content_list.json @@ -0,0 +1,1515 @@ +[ + { + "type": "text", + "text": "Tra-MoE: Learning Trajectory Prediction Model from Multiple Domains for Adaptive Policy Conditioning", + "text_level": 1, + "bbox": [ + 130, + 128, + 869, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiange Yang $^{1,2}$ Haoyi Zhu $^{2,3}$ Yating Wang $^{2,4}$ Gangshan Wu $^{1}$ Tong He $^{2}$ Limin Wang $^{1,2*}$", + "bbox": [ + 107, + 200, + 895, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ State Key Laboratory for Novel Software Technology, Nanjing University, China $^{2}$ Shanghai AI Laboratory, $^{3}$ University of Science and Technology of China, $^{4}$ Tongji University", + "bbox": [ + 125, + 220, + 870, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jiangyang.jgy@gmail.com, {gswu,lmwang}@nju.edu.cn", + "bbox": [ + 272, + 258, + 720, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhuhaoyi, wangyating, hetong}@pjlab.org.cn", + "bbox": [ + 308, + 277, + 687, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/MCG-NJU/Tra-MoE", + "bbox": [ + 326, + 294, + 665, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 344, + 326, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning from multiple domains is a primary factor that influences the generalization of a single unified robot system. In this paper, we aim to learn the trajectory prediction model by using broad out-of-domain data to improve its performance and generalization ability. Trajectory model is designed to predict any-point trajectories in the current frame given an instruction and can provide detailed control guidance for robotic policy learning. To handle the diverse out-of-domain data distribution, we propose a sparsely-gated MoE (Top-1 gating strategy) architecture for trajectory model, coined as Tra-MoE. The sparse activation design enables good balance between parameter cooperation and specialization, effectively benefiting from large-scale out-of-domain data while maintaining constant FLOPs per token. In addition, we further introduce an adaptive policy conditioning technique by learning 2D mask representations for predicted trajectories, which is explicitly aligned with image observations to guide action prediction more flexibly. We perform extensive experiments on both simulation and real-world scenarios to verify the effectiveness of Tra-MoE and adaptive policy conditioning technique. We also conduct a comprehensive empirical study to train Tra-MoE, demonstrating that our Tra-MoE consistently exhibits superior performance compared to the dense baseline model, even when the latter is scaled to match Tra-MoE's parameter count.", + "bbox": [ + 89, + 376, + 485, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 762, + 220, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In computer vision and natural language processing, significant progress has been made through learning models from multiple domains data. For example, many works use board data to perform self-supervised pre-training, and then adapt to diverse downstream tasks through zero-shot [1, 10, 49] or fine-tuning [13, 15, 23, 58]. Additionally, some research ef", + "bbox": [ + 89, + 785, + 485, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b20601886bd5d635928fdf1cbebeaae57f139d3d6ba8796761e8f9210cfab061.jpg", + "image_caption": [ + "Figure 1. (a) Training with small-scale, in-domain data. (b) Joint training with in-domain data and large-scale, out-of-domain data." + ], + "image_footnote": [], + "bbox": [ + 558, + 345, + 870, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "forts [43, 51, 71, 72] focus on learning a generalist model capable of unifying various tasks. In contrast, in robot learning, due to data scarcity and homogeneity, some groundbreaking methods [2, 37] resort to training only using in-domain data. Recently, some works [3, 4, 52, 61, 69] have shifted to learn from action-free human video data. A particularly scalable and promising paradigm [5, 20, 59, 63, 65] involves learning a trajectory prediction model from action-free video data, and then using small-scale action-labeled demonstrations to learn trajectory-guided policies, thereby achieving greater sample efficiency and better generalization.", + "bbox": [ + 511, + 608, + 908, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the potential benefits, it still remains underexplored how to effectively leverage broad out-of-domain video data for jointly training trajectory models. In particular, the impact of incorporating data encompassing diverse environments, objects, skills and embodiments has not fully investigated for in-domain tasks. For instance, ATM [63], a notable work in this field, mainly relies on in-domain data (same tasks and environments) for training both trajectory", + "bbox": [ + 511, + 779, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 236, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6960", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1116d206222c344b3495caf32192dc759e91b4b7ea155d6cd6e4cf158685afb2.jpg", + "image_caption": [ + "Figure 2. Training trajectory prediction model from multiple domains. (a) The visualization of Dataset $\\mathcal{D}_{\\text {ood }}$ and $\\mathcal{D}_{\\text {in }}$ . The former may contain additional environments, objects, skills and embodiments. (b) Our pipeline: first co-training trajectory prediction model and then adapting it for downstream policy learning." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 297, + 257 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dbc45401311683cf540e6205e3be6a9a5b51e719c61443d62b51f8072d7ecab6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 308, + 89, + 477, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "models and policies. We argue that learning from multiple domains is the primary factor to build a powerful trajectory model for policy guidance. Therefore, we propose to train trajectory model under the setup of jointly learning from large-scale out-of-domain and small-scale in-domain data, as shown in Fig. 2(a). However, directly using large-scale out-of-domain video data for hybrid pre-training involves two critical challenges: (i) Joint training on data from different environments, objects, skills, and embodiments will add learning difficulty and might lead to optimization conflicts. It is unclear how to effectively fuse the complementary information of multiple domains data within a single unified model. Naively expanding the out-of-domain training data does not necessarily lead to improved performance on in-domain tasks, as shown in Fig. 1, which results in a 5.6 $(57.6\\to 52.0)$ performance decrease on in-domain tasks. (ii) How to ensure both high performance and computational efficiency during trajectory model scaling up process.", + "bbox": [ + 88, + 340, + 482, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we design a new sparsely-gated Mixture-of-Expert (MoE) architecture [53] to scale up our trajectory model, coined as Tra-MoE. Specifically, we integrate several MoE blocks to replace the original transformer blocks. This design enables efficient joint training of most parameters on the broad out-of-domain data, capturing the complementary patterns for mutual cooperation and the shared knowledge for better generalization. Meanwhile, for different data and different tokens within the same data, our Tra-MoE naturally forms a specialization when activating different experts. To maintain high computational efficiency while scaling up model capacity, we implement a top-1 gating strategy for token-choice, ensuring constant FLOPs per token. In sparse MoE, some auxiliary losses are commonly used to enhance performance, including the router z-loss [74] for improving training stability and the load-balancing loss [36] for balancing expert activations. We further conduct a comprehensive empirical study on scaling up trajectory models and find that the touter z-loss improves", + "bbox": [ + 89, + 614, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "performance while the load-balancing loss tends to cause performance degradation. Furthermore, we observe that when expanding pre-training data from small-scale in-domain to large-scale in-domain and out-of-domain data, employing MoE (Tra-MoE) can lead to obvious performance improvement on in-domain tasks, whereas the dense counterpart (Tra-baseline) suffers from performance drop, as illustrated in Fig. 1. Notably, even when expanding the dense Tra-baseline's parameter to match that of Tra-MoE, performance improvements remain elusive. These findings collectively demonstrate the effectiveness of our sparsely-gated MoE architecture, attributing its capability of balancing parameter cooperation and specialization when expanding training with out-of-domain data.", + "bbox": [ + 511, + 90, + 906, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, given a pre-trained trajectory model, how to effectively condition the policy module using the 2D trajectories still remain an unresolved challenge. We further propose an adaptive policy conditioning technique for trajectory-guided policy. This enables better spatial alignment between 2D trajectories and images, while achieving more flexible trajectory representation and guidance for the policy. Finally, we also conduct extensive real-world experiments to further strengthen our conclusions. In summary, our main contributions can be summarized as follows:", + "bbox": [ + 511, + 301, + 908, + 452 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a sparsely-gated MoE architecture to learn trajectory models, called as Tra-MoE, with large-scale, out-of-domain, and action-free video data, coupled with a comprehensive empirical study on its training techniques.", + "- Extensive simulation and real-world environments demonstrate that our Tra-MoE (with Top-1 gating) can more effectively benefit from large-scale out-of-domain data (such as cross-embediments data and different physics engine simulated data) compared to the dense baseline.", + "- We demonstrate that Tra-MoE still significantly outperform expanded dense counterpart with equivalent parameters, highlighting the effectiveness of sparse MoE architecture in leveraging out-of-domain data through better parameter cooperation and specialization.", + "- We propose an adaptive policy conditioning technique for trajectory-guided policy, which ensures adaptive 2D trajectory representation and explicit spatial alignment with the observations, thereby achieving superior performance." + ], + "bbox": [ + 513, + 455, + 906, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 739, + 653, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning from multiple domains in Robot Learning. Scaling laws [30] have fueled substantial advancements in various fields. In robot learning, many recent studies [7, 8, 29, 51] investigate how compute, model size, and training data quantity affect the model performance on robotic manipulation tasks. Based on this, recent works have focused on collecting larger-scale real-world robotic datasets [11, 19, 32, 60, 64], and have successfully trained several robotics foundational models [6, 9, 16, 33, 38, 47, 57,", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6961", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[62] that demonstrate superior transfer capabilities. It is worth noting that these works mainly employ pre-training and fine-tuning paradigm rather than joint co-training paradigm, while focusing on evaluating the combination generalization and neglecting the performance impact on in-domain tasks. Additionally, these works directly combine heterogeneous robotic data with different action and observation spaces for joint training, potentially leading to sub-optimal solutions. In contrast, we leverage sparsely-gated MoE to incorporate larger-scale, out-of-domain, and action-less video data. Furthermore, our experiments demonstrate the importance of MoE for expanding out-of-domain data due to its superior capacity for parameter cooperation and specialization.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mixture-of-Experts. The Mixture-of-Experts (MoE) framework [12] is built upon a simple yet effective concept: dividing a model into specialized components, or 'experts', each focusing on distinct tasks or data aspects. This paradigm allows for selective activation of relevant experts based on the input, thereby maintaining computational efficiency while leveraging a vast reservoir of specialized knowledge through increased model capacity. Recently, in computer vision, natural language processing, and multi-modal, there has been a surge of foundation models trained based on the MoE architecture, including Swin-Moe [24], MoE-LLaVA [39], DeepSeek-V2 [40], Mixtral-8x22B [28], and so on. MoE typically encompasses dense [48] and sparse [14] variants. Dense MoE activates all experts in each iteration, while sparse MoE activates only a subset, generally resulting in lower computational overhead. However, sparse MoE can suffer from expert load imbalance and training instability issues. Techniques such as adding noise, load balancing loss [36], and router z-loss [74] are often employed to mitigate these issues and improve performance. In our work, we conduct a comprehensive empirical study of these techniques when training our sparsely-gated MoE-based trajectory model.", + "bbox": [ + 91, + 292, + 486, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Representations and Techniques for Policy Conditioning. Policy conditioning equips a multi-task robot policy with the ability to modulate its behavior by incorporating task-specific guidance information. This information, which may take various forms, guides the policy's action generation, enabling a single model to effectively learn and perform diverse tasks. Classic policy conditioning representations often encompass task identifiers [50], natural language instructions [8, 22, 45, 54], goal images [44], and demonstration videos [25, 27]. These are typically integrated with visual observations within the policy through concatenation or attention mechanisms. Recent works have explored alternative policy conditioning representations to enhance generalization and sample efficiency. These include leveraging internet-trained models to obtain bounding box [55] and object mask [42, 66, 68, 73], as well as learning plan [4], affordance [3, 46], future observations [17, 18, 34, 67], 2D or 3D trajectory [5, 61, 63, 65, 70] from human video data,", + "bbox": [ + 89, + 628, + 486, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and also hand-drawn trajectory [21] or sketch [56]. Different from these approaches that directly utilize raw or manually crafted signals for integration with visual observations, we introduce an adaptive policy conditioning technique that explicitly maps 2D trajectories onto the visual observations in space and encodes them as learnable embeddings. These designs ensure adaptive 2D trajectory representation and explicit spatial alignment with the visual observations.", + "bbox": [ + 511, + 90, + 908, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 224, + 648, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first describe our problem formulation in Sec. 3.1. Subsequently, we give an overview of our proposed framework, which is composed of a sparsely-gated MoE-based trajectory prediction model and a trajectory-guided policy model using the adaptive policy conditioning technique, detailed in Sec. 3.2. Finally, we describe our training process and loss functions in Sec. 3.3.", + "bbox": [ + 511, + 250, + 908, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 511, + 364, + 718, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following the setup of previous works [5, 63, 65], our approach encompasses a trajectory prediction model to conduct language-conditioned any-point trajectory prediction and a trajectory-guided policy model to predict executable robot actions. In particular, as depicted in Fig. 2, we consider two types of datasets for training, which are large-scale, out-of-domain, and action-free videos $\\mathcal{D}_{\\text{ood}}$ , as well as small-scale, in-domain, and action-labeled demonstrations $\\mathcal{D}_{\\text{in}}$ . For trajectory prediction model, we employ $\\mathcal{D}_{\\text{ood}}$ and $\\mathcal{D}_{\\text{in}}$ for joint pre-training, whereas for trajectory-guided policy model, we only use $\\mathcal{D}_{\\text{in}}$ for training.", + "bbox": [ + 511, + 386, + 908, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Architecture", + "text_level": 1, + "bbox": [ + 511, + 561, + 648, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ATM Revisited. In this paper, we use ATM [63] as the baseline for our method. The overall pipeline of ATM is illustrated on Fig. 2(b). It first learns a trajectory model from video data to perform language-conditioned trajectory prediction. The predicted trajectories is then used to learn a trajectory-guided policy using robot demonstrations. Specifically, as for trajectory model, its objective is to predict the future of any-point trajectories in a frame. Formally, given an image observation $o_t$ at timestep $t$ , any set of 2D query points simplified on the image frame $\\mathbf{p}_t = \\{p_{t,k}\\}_{k=1}^K$ , and a language instruction $\\ell$ , the trajectory model learns a mapping $\\mathbf{p}_{t:t+H} = \\tau_\\theta(o_t, \\mathbf{p}_t, \\ell)$ that predicts the coordinates of these query points in the camera frame for the next $H$ time steps, where $\\mathbf{p}_{t:t+H} \\in \\mathbb{R}^{H \\times K \\times 2}$ . For the trajectory-guided policy, it leverages the current observation $o_t$ and the predicted trajectory $\\mathbf{p}_{t:t+H}$ to predict the subsequent robot actions.", + "bbox": [ + 511, + 583, + 908, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sparsely-gated MoE-based trajectory model (Tra-MoE). ATM proposes a track transformer to implement the trajectory model. In track transformer, language tokens, image tokens, and track tokens to be decoded involve global interactions within the transformer block. However, directly", + "bbox": [ + 511, + 825, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6962", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1b360506ebbfb7a53d46babead0df0d6744f34a203836d82d0e49690726e4618.jpg", + "image_caption": [ + "Figure 3. (a) The pipeline of our sparsely-gated MoE-based trajectory model (Tra-MoE). (b) The pipeline of our trajectory-guided policy using the adaptive policy conditioning technique. Mapping means concatenating the trajectory mask with image observations, while learnable refers to setting each point in the trajectory mask as a learnable embedding." + ], + "image_footnote": [], + "bbox": [ + 151, + 88, + 557, + 414 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d1ccaa97ebefa6d3fb7b4012c22df2c42abd6eb7a4331b344e1bc0685f8f71ae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 89, + 836, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "processing diverse domains and multiple modalities data within a unified transformer may yield suboptimal results. Hence, we propose our Tra-MoE framework. To develop our sparsely-gated MoE-based trajectory model, as depicted in Fig. 3(a), we replace certain layers of the transformer with MoE blocks. In each MoE block, it incorporates multiple feedforward networks (FFNs), each designated as an expert, and utilizes a gating function to activate a selected subset (Top- $K$ ) of these experts. In this paper, we set $K$ to a fixed value of 1 to ensure that the FLOPs per token remain constant relative to the baseline. Formally, the gating network $\\mathcal{G}$ , parameterized by $\\Theta$ and typically consisting of a linear-softmax network, yields the output $\\mathcal{G}(\\mathbf{x_s};\\Theta)$ , where $\\mathbf{x_s}$ represents our token sequences input. We formulate the sparsely-gated token-choice mechanism as follows:", + "bbox": [ + 88, + 477, + 485, + 704 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} \\left(\\mathbf {x} _ {\\mathrm {s}}; \\boldsymbol {\\Theta}\\right) _ {i} = \\operatorname {s o f t m a x} (\\operatorname {T o p - K} (g \\left(\\mathbf {x} _ {\\mathrm {s}}; \\boldsymbol {\\Theta}\\right), k)) _ {i}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 746, + 485, + 763 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {T o p - K} (g \\left(\\mathbf {x} _ {\\mathbf {s}}; \\boldsymbol {\\Theta}\\right), k) _ {i} = \\left\\{ \\begin{array}{l l} g (\\mathbf {x}; \\boldsymbol {\\Theta}) _ {i}, & \\text {c o n d ,} \\\\ - \\infty , & \\text {o t h .} \\end{array} \\right., \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 768, + 483, + 806 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {c o n d}: \\text {i f} g (\\mathbf {x}; \\boldsymbol {\\Theta}) _ {i} \\text {i s i n t o p - k e l e m e n t s o f} g (\\mathbf {x}; \\boldsymbol {\\Theta}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 810, + 483, + 828 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More specifically, $\\operatorname{Top - K}(\\cdot ,k)$ function retains only the top- $k$ values in a vector, setting the rest to $-\\infty$ . Consequently, the output of the sparsely-gated MoE layer can be", + "bbox": [ + 89, + 854, + 485, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "formulated as:", + "bbox": [ + 513, + 477, + 612, + 491 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\text {s p a r s e}} ^ {\\mathrm {M o E}} \\left(\\mathbf {x} _ {\\mathbf {s}}; \\boldsymbol {\\Theta}, \\left\\{\\mathbf {W} _ {i} \\right\\} _ {i = 1} ^ {K}\\right) = \\sum_ {i = 1} ^ {K} \\mathcal {G} \\left(\\mathbf {x} _ {\\mathbf {s}}; \\boldsymbol {\\Theta}\\right) _ {i} f _ {i} \\left(\\mathbf {x} _ {\\mathbf {s}}; \\mathbf {W} _ {i}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 500, + 906, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{i}$ represents the i-th expert, usually a linear-GELU-linear FFN layer, is parameterized by $\\mathbf{W}_i$ .", + "bbox": [ + 511, + 568, + 906, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Trajectory-guided policy using the adaptive policy conditioning technique. After completing the trajectory model learning, we develop our trajectory-guided policy through the adaptive policy conditioning technique based on the initial ATM, as shown in Fig. 3(b). In our policy model, independently encoded track tokens and image tokens undergo a early fusion within the fusion transformer. Subsequently, the fused features are further integrated with the raw trajectory data through a late fusion mechanism along the channel dimension, ensuring the preservation of original trajectory information for more accurate action prediction.", + "bbox": [ + 511, + 598, + 908, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fundamentally, the policy model bridges the gap between the predicted 2D any-point trajectories and the executable 3D robot actions, functioning as an inverse dynamic model and eliminating the necessity of other task specification. However, how to effectively condition the policy using the predicted 2D trajectories remains an unresolved challenge. Based on our analysis, we present two key insights: (i) Explicit spatial alignment and fusion of the 2D trajectories with image observations turns out to be highly beneficial.", + "bbox": [ + 511, + 763, + 910, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6963", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It simplifies the mapping of 2D trajectories to image space and highlights motion-relevant regions, enabling the policy model to focus on important areas and reducing the learning difficulty. (ii) Trajectory positions exhibit different characteristics: starting points typically emphasize local motion, while endpoints focus more on global trend. Therefore, an adaptive trajectory representation and conditioning technique is expected to flexibly guide the policy by effectively capturing these variations.", + "bbox": [ + 89, + 90, + 483, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the aforementioned analysis, we propose an adaptive policy conditioning technique to incorporate trajectory information into 2D images. Specifically, we construct an additional mask modality, and populate it with 2D trajectories based on their specific spatial locations, while setting the value of each point in the trajectory mask as a learnable embedding. Before encoding the image, we concatenate this mask modality with the image observations on the channel dimension, forming a tensor of dimensions $H \\times W \\times 4$ , as shown in Fig. 3(b). To sum up, our adaptive policy conditioning technique yield an adaptive 2D trajectory representation, which could be explicitly aligned with the image observations to guide the policy prediction more flexibly.", + "bbox": [ + 89, + 227, + 483, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training", + "text_level": 1, + "bbox": [ + 89, + 430, + 194, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We argue that trajectory models could be shared by multiple domains data. Therefore, we first pre-train a generalizable trajectory model on $\\mathcal{D}_{\\text{ood}} \\cup \\mathcal{D}_{\\text{in}}$ . The $\\mathcal{D}_{\\text{ood}}$ typically contains additional environments, objects, skills, and embodiments. Then we train the trajectory-guided policy model on $\\mathcal{D}_{\\text{in}}$ based on the pre-trained trajectory model. In trajectory-guided policy training phase, we freeze the trajectory model. Pre-training sparsely-gated MoE-based trajectory model. Following ATM, we train the trajectory model using the ground truth tracks generated by CoTracker [31]. We use MSE to optimize the trajectory prediction loss $\\mathcal{L}_{\\text{tra}}$ . Additionally, we also consider adding an auxiliary load-balancing loss [36] to balance the activation frequency among experts. Given a batch of queries $\\mathcal{B}$ , it contains B samples and each sample contains S tokens. We define $\\mathcal{Q}_i$ as the proportion of tokens distributed to expert $i$ , and $\\mathcal{P}_i$ as the proportion of the gating probability assigned to expert $i$ . The $\\mathcal{L}_{\\text{lo-ba}}$ can be formulated as follows:", + "bbox": [ + 89, + 452, + 483, + 724 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Q} _ {i} = \\frac {1}{S \\cdot B} \\sum_ {x _ {s} ^ {m} \\in \\mathcal {B}} \\mathbb {1} \\left\\{\\operatorname {a r g m a x} \\mathcal {G} \\left(\\mathbf {x} _ {\\mathbf {s}} ^ {\\mathbf {m}}; \\boldsymbol {\\Theta}\\right) = i \\right\\}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 742, + 483, + 779 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} _ {i} = \\frac {1}{S \\cdot B} \\sum_ {x _ {\\mathrm {s}} ^ {m} \\in \\mathcal {B}} \\mathcal {G} \\left(\\mathbf {x} _ {\\mathrm {s}} ^ {\\mathrm {m}}; \\boldsymbol {\\Theta}\\right) _ {i}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 782, + 483, + 819 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {l o - b a}} = N \\sum_ {i = 1} ^ {N} \\mathcal {Q} _ {i} \\mathcal {P} _ {i}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 824, + 483, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, ST-MoE [74] proposed using the router $z$ -loss to penalize large logits entering the gating network,", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "thereby improving training stability. We similarly applies it to the training of our trajectory model, as shown below:", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {z} (x) = \\frac {1}{S} \\sum_ {k = 1} ^ {S} \\left(\\log \\sum_ {i = 1} ^ {N} e ^ {\\mathbf {g} _ {i} ^ {(k)}}\\right) ^ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 132, + 906, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $S$ is the total number of tokens, $N$ is the total number of experts, and $\\mathbf{g} \\in \\mathcal{R}^{S \\times N}$ are the logits going into the router. In summary, the trajectory model's training loss is as follow:", + "bbox": [ + 511, + 181, + 906, + 229 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {t r a}} \\cdot \\mathcal {L} _ {\\text {t r a}} + \\lambda_ {\\text {l o - b a}} \\cdot \\mathcal {L} _ {\\text {l o - b a}} + \\lambda_ {z} \\cdot \\mathcal {L} _ {z}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 256, + 906, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training trajectory-guided policy through adaptive policy conditioning technique. For trajectory-guided policy training, we employ the paradigm of behavior cloning. Specifically, we use the MSE between predicted actions and ground truths as the loss function.", + "bbox": [ + 511, + 284, + 906, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 375, + 645, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We perform evaluation experiments using the LIBERO benchmark [41] and a real low-cost dual-arm robot [35]. In experiments, we quantitatively find that the success rate of downstream manipulation tasks is generally positively correlated with the performance of the trajectory model. Detailed quantitative experiment results can be found in the appendix. Meanwhile, CoTracker [31] can provide relatively accurate trajectory annotations; therefore, we report the average success rate of downstream manipulation tasks as our metric. Our experiments aim to study the following questions:", + "bbox": [ + 511, + 398, + 906, + 550 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Q1: Does the techniques of router z-loss, load-balancing loss, and adding noise to gating logits enhance the training of sparsely-gated Tra-MoE?", + "Q2: Does sparsely-gated Tra-MoE, compared to the dense Tra-baseline, benefit more from large-scale out-of-domain data (such as cross-embediments and different physics engine data) and demonstrate superior scaling up capabilities?", + "Q3: When we scale up the dense Tra-baseline to match the parameter count of Tra-MoE, does the sparsely-gated Tra-MoE still maintain its advantage?", + "Q4: Is our proposed adaptive policy conditioning technique for trajectory-guided policy more effective than ATM?" + ], + "bbox": [ + 511, + 551, + 906, + 732 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Simulation Experiments", + "text_level": 1, + "bbox": [ + 511, + 742, + 736, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experiment setup. The LIBERO [41] is divided into five categories: LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, LIBERO-Long, and LIBERO-90. LIBERO-90 includes 90 tasks, while each of the other categories contains 10 tasks. Compared to other categories, LIBERO-90 presents a highly varied selection of object categories, layouts, environments, and skill goals. It is typically used for pre-training rather than joint co-training. Therefore, our main ablation experiments select 90 tasks from LIBERO-90 to construct", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6964", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9504d9d45d6a4b0a101bfc2bfb7656c82744da79452a9b4ebd7335a1fd79c306.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SpatialGoalObjectLongAvg.
062.073.071.021.556.9
5·10-469.573.568.022.558.4
1·10-462.581.073.528.561.4
1·10-560.065.568.527.055.3
1·10-660.079.053.031.555.9
", + "bbox": [ + 116, + 88, + 472, + 189 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0f0712d7bceb9eb4623303c792b6ec47f95513fa514dcad77622b0c57072fcfd.jpg", + "table_caption": [ + "(a) The value of $\\lambda_z$" + ], + "table_footnote": [], + "table_body": "
SpatialGoalObjectLongAvg.
w/o. noise62.073.071.022.556.9
w. noise71.567.057.527.055.8
", + "bbox": [ + 102, + 209, + 478, + 266 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/03a497c14e81d85247113fd03473c7b66e52786ef05f766549381404274a1fcc.jpg", + "table_caption": [ + "(c) Adding a noise term to the gating logits." + ], + "table_footnote": [], + "table_body": "
Num.SpatialGoalObjectLongAvg.
149.567.056.535.052.0
259.072.565.527.056.0
364.068.071.017.555.1
462.073.071.021.556.9
", + "bbox": [ + 117, + 300, + 460, + 386 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/750356bdb0dea26187e847fb2fc535848a12c8434edd50e8202d3ed706bd395d.jpg", + "table_caption": [ + "(e) The number of experts." + ], + "table_footnote": [], + "table_body": "
SpatialGoalObjectLongAvg.
062.073.071.022.556.9
1·10-358.574.060.018.052.6
1·10-561.069.556.026.053.1
1·10-766.567.073.522.057.3
", + "bbox": [ + 526, + 88, + 879, + 174 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/17bdff4583247d83ff735f2dd80fe5f356987ff135fffb08860a3e448251094e.jpg", + "table_caption": [ + "(b) The value of $\\lambda_{lo - ba}$ ." + ], + "table_footnote": [], + "table_body": "
SpatialGoalObjectLongAvg.
MoE62.581.073.528.561.4
Width66.550.561.515.048.4
Depth61.575.546.526.552.5
", + "bbox": [ + 529, + 209, + 875, + 280 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6139101233449309a3573c9cefcb724605042ca4863f91c124222887c560fdb4.jpg", + "table_caption": [ + "(d) Scaling Tra-baseline from width and depth." + ], + "table_footnote": [], + "table_body": "
MoEOODSpatialGoalObjectLongAvg.
67.568.568.026.557.6
✓63.064.060.519.551.8
✓49.567.056.535.052.0
✓✓62.581.073.528.561.4
", + "bbox": [ + 500, + 300, + 895, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "dataset $\\mathcal{D}_{\\text{ood}}$ , utilizing 20 action-free videos per task; and choose 40 tasks from LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, and LIBERO-Long to build dataset $\\mathcal{D}_{\\text{in}}$ , using 10 action-labeled demonstrations per task. In summary, for training Tra-MoE, our main ablation experiments use a total of 2, 200 action-free videos, with an out-of-domain to in-domain data ratio of $9:2$ . Additionally, on this basis, we also further select 92 tasks from the RLbench [26] (using CoppeliaSim rather than MuJoCo as physics engine) simulation environment and generate 5 videos for each task, totaling 2, 660 videos in conjunction with the LIBERO data, to train Tra-MoE. The different physics engines and dynamics properties further increase the data distribution discrepancy in $\\mathcal{D}_{\\text{ood}}$ . Finally, as for the trajectory-guided policy, we train a separate multi-task policy for each category. Each policy undergoes behavioral cloning training with 100 action-labeled demonstrations from 10 tasks.", + "bbox": [ + 88, + 431, + 485, + 689 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implement details. Building on ATM [63], we develop our sparsely-gated Tra-MoE and adaptive policy conditioning technique. Specifically, when using only LIBERO data, we replace the 1th, 5th, and 8th layers of the track transformer with MoE blocks. When further incorporating RLbench data, we additionally replace the 2th and 7th layers. Unless otherwise specified, we set the number of experts to 4 by default and use a top-1 gating strategy for token choice to ensure the FLOPs per token remain constant.", + "bbox": [ + 88, + 694, + 482, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiment results and analysis. We present empirical simulation results to address the aforementioned research questions and make the following findings:", + "bbox": [ + 89, + 835, + 482, + 881 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Result 1: In Tab. 1a, we examine the effect of the router", + "bbox": [ + 89, + 885, + 482, + 898 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/694232bc2cfc422c4abc74e342ff3c257c79792d291197b661c0015ed70541ee.jpg", + "table_caption": [ + "(f) Scaling with MoE and OOD data.", + "Table 1. The ablation experiments of the sparsely-gated Tra-MoE. All baselines are the track transformer of ATM." + ], + "table_footnote": [], + "table_body": "
SpatialGoalObjectLongAvg.
ATM51.042.055.519.041.9
Tra-MoE54.062.061.041.054.5
", + "bbox": [ + 524, + 430, + 893, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. The comparison results when training trajectory model using RLbench video data.", + "bbox": [ + 511, + 497, + 903, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "z-loss weight $\\lambda_z$ . The results indicate that a higher $\\lambda_z$ can effectively enhance performance, with an improvement of 4.5 (61.4 vs. 56.9) at $\\lambda_z = 1 \\cdot 10^{-4}$ . In Tab. 1b, we study the impact of the load-balancing loss weight $\\lambda_{lo - ba}$ . The results show a significant performance decline with larger values of $\\lambda_{lo - ba}$ , with only a slight improvement ( $56.7 \\rightarrow 57.3$ ) at $\\lambda_{lo - ba} = 1 \\cdot 10^{-7}$ . In Tab. 1c, we verify the effectiveness of adding noise to the gating logits, which results in a performance decrease of by 1.1 ( $56.9 \\rightarrow 55.8$ ).", + "bbox": [ + 511, + 534, + 906, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finding 1: Based on the above results, we find that using router z-loss to penalize large logits entering the gating network can effectively improve training stability and lead to performance gains. Meanwhile, the load-balancing loss used to balance the activation frequency among experts and the addition of noise to enhance expert allocation exploration cannot bring effective performance improvements. An intuitive explanation is that when the data distribution is uneven, the load-balancing loss might force experts to set shared parameters on data with a large domain gap, which could lead to mutual counteraction of learning gradients. This undermines the advantage of parameter specialization in sparsely-gated MoE. Similarly, adding noise is also detrimental to learning parameter specialization among experts.", + "bbox": [ + 511, + 672, + 908, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Result 2: We comprehensively investigate the impact of in", + "bbox": [ + 511, + 885, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6965", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/32262f4e4a8ee1cb12f5562878940f70d642710c03559def97db4bef228462b5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SpatialGoalObjectLongAvg.
ATM62.581.073.528.561.4
+ DM69.058.085.533.561.5
+ AM69.577.088.030.566.3
", + "bbox": [ + 114, + 88, + 460, + 160 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. The ablation experiments of the adaptive policy conditioning technique. DM (RT-Trajectory style) and AM respectively represent hand-drawn mask and adaptive mask.", + "bbox": [ + 89, + 171, + 482, + 213 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "producing sparsely-gated MoE and large-scale out-of-domain data on scaling up trajectory models in Tab. 1f. Firstly, when training only with in-domain data, Tra-baseline achieves an average success rate of 57.6, while Tra-MoE only achieves an average success rate of 51.8. Secondly, when adding out-of-domain data to jointly train Tra-baseline, the average success rate drops to $52.0$ ( $57.6 \\rightarrow 52.0$ ), with particularly significant performance declines in LIBERO-Spatial and LIBERO-Object. However, when adding out-of-domain data to jointly train Tra-MoE, the average success rate increases to $61.4$ ( $51.8 \\rightarrow 61.4$ ). Additionally, in Tab. 1e, we also explore the impact of the number of experts in sparsely-gated MoE. The results show that using just two experts can lead to an increase of 4.0 (52.0 vs. 56.0) in the average success rate. With additional experts, performance generally trends upward, albeit with some fluctuations. Finally, we report the comparison results of training trajectory model by further using RLbench data in Tab. 2. The results indicate that although the introduction of RLbench data led to a performance drop in LIBERO, our Tra-MoE still outperform the baseline by 12.6 (54.5 vs. 41.9).", + "bbox": [ + 91, + 220, + 483, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finding 2: Based on the above results, we have the following findings and analysis: (i) When the data scale is small, naively expanding the model capacity through sparsely-gated MoE does not lead to performance improvements. A explanation is that the larger model may lead to a overfitting problem when training on small-scale data. Moreover, small-scale data often leads to some experts not being sufficiently trained, thereby affecting performance. (ii) Naively introducing out-of-domain data for joint training often leads to performance degradation in the target domain. However, simultaneously incorporating out-of-domain data and expanding models with sparsely-gated MoE may improve performance or alleviate performance degradation. This phenomenon suggests that although out-of-domain data generally enhances model generalization, it could hinder task-specific performance in the target domain due to insufficient in-domain data learning. In contrast, expanding the model with out-of-domain data using sparsely-gated MoE achieves a more optimal balance between parameter cooperation and specialization across diverse data and even different tokens within the same data input. This design allows these large-scale data to jointly train most of parameters, ensuring the capture of general patterns for mutual cooperation. (iii) Even with a few experts, Tra-MoE exhibits significant scaling up", + "bbox": [ + 91, + 537, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "capabilities. Additionally, further increasing the number of experts generally leads to performance improvement. However, this improvement is not guaranteed due to factors like data-expert complexity matching and training stability.", + "bbox": [ + 511, + 90, + 906, + 151 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Result 3: In Tab. 1d, we also investigate the performance comparison between the dense Tra-baseline and the sparse Tra-MoE with the same number of parameters. Specifically, we expand Tra-baseline to match the model capacity of Tra-MoE by increasing its model width or depth separately. These expanded versions achieve average success rates of 48.4 and 52.5 respectively, which are significantly lower than Tra-MoE's 61.4.", + "bbox": [ + 511, + 154, + 908, + 273 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finding 3: Based on these results, we are surprised to find that simply expanding the dense Tra-baseline in depth and width can not effectively improve performance. This highlights the importance of the sparsely-gated MoE architecture when training with large-scale out-of-domain data, as it ensures dynamic activation of different experts based on different data input and different tokens within the same data input, significantly reducing optimization conflict problems.", + "bbox": [ + 511, + 276, + 908, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Result 4: In Tab. 3, we investigate the effectiveness of the adaptive policy conditioning. Specifically, building upon ATM, we first introduce an additional hand-drawn mask modality before image encoding. In this mask, the first half of the trajectory is set to 128, while the second half is set to 255. We call it 'ATM + hand-drawn mask' and this is similar to RT-Trajectory [21]. Next, we replace the hand-drawn mask with an adaptive mask, where each point on the trajectory is instantiated as a learnable embedding. We call it 'ATM + adaptive mask'. The results show that 'ATM + hand-drawn mask' achieves only a 0.1 improvement compared to the baseline. However, except for a significant performance decrease in LIBERO-Goal, it achieves performance improvements of 6.5, 12.0, and 5.0 in LIBERO-Spatial, LIBERO-Object, and LIBERO-Long respectively. In addition, the 'ATM + adaptive mask' achieves a performance improvement of 4.9 (66.3 vs. 61.4) compared to the baseline. It further improves performance across all suites compared to 'ATM + hand-drawn mask', especially achieving a 19.0 performance gains in LIBERO-Goal (77.0 vs. 58.0).", + "bbox": [ + 511, + 400, + 908, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finding 4: Based on these results, we find that explicitly mapping 2D trajectories to visual observations significantly improves the performance of LIBERO-Spatial, LIBERO-Object, and LIBERO-Long, but causes a notable decrease in LIBERO-Goal. By instantiating the points in the 2D trajectory as learnable embeddings, we not only further enhances the performance of LIBERO-Spatial, LIBERO-Object, and LIBERO-Long, but also significantly improves LIBERO-Goal. We analyze that the decline in LIBERO-Goal is mainly due to overlapping front and back parts of some trajectories in the 2D images, leading to mapping errors. However, our adaptive trajectory representation and conditioning technique effectively mitigates this issue. In summary, these", + "bbox": [ + 511, + 704, + 910, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6966", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bd6dc923dc11f3b6cfa2edcdc0fc56ce008272cd16430bccf28e46cbcbea5b57.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Human DataMoEAdaptive Trajectory MaskPourPushPick and PassTissueFoldAverage
40.045.050.030.025.038.0
✓45.035.050.025.035.038.0
✓✓40.050.060.035.045.046.0
✓✓✓60.070.065.035.050.056.0
", + "bbox": [ + 125, + 89, + 870, + 175 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. The ablation of ood human data, MoE architecture, and adaptive policy conditioning technique on our real-world tasks.", + "bbox": [ + 119, + 186, + 872, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/84503aacec77ac5c54356f89d97bcde412c1ed4dba170d2a1d766ebf2aa969d0.jpg", + "image_caption": [ + "Figure 4. Left: The real-world experiments hardware platform setup. Right: The real-world tasks evaluation demonstrations." + ], + "image_footnote": [], + "bbox": [ + 155, + 204, + 834, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "experiments collectively demonstrate the effectiveness of explicitly aligning 2D trajectories with images in space as well as applying the adaptive conditioning technique.", + "bbox": [ + 89, + 435, + 483, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Real-World Experiments", + "text_level": 1, + "bbox": [ + 89, + 494, + 318, + 511 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Experiment setup. As shown in Fig 4, we utilize a dual-arm robot [35] to conduct real-world experiments. We evaluate 5 tasks, which include two single-arm tasks: pouring water and pushing a vegetable to the side of the cutting board, as well as three dual-arm tasks: pulling out tissues, folding towels, and picking up and passing a holder. Specifically, for each task, we collect 50 action-labeled robot demonstrations. Concurrently, we also collect 50 videos of human (cross-embediments) performing for each task. Compared to the robot data, the human video encompasses more environment variations and distractor objects. Please see the appendix for more detailed evaluation details and real-world setup.", + "bbox": [ + 89, + 518, + 483, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implement details. Following our simulation experiments, we utilize two camera views and uniformly resize their resolution to $128 \\times 128$ for trajectory model and policy training. We train our real-world Tra-MoE on borh human video and robot video data, with a separate policy trained for each task. Therefore, our real-world Tra-MoE integrates data across different embodiments (human, single-arm robot, and dual-arm robot). They have significantly different visual appearances and dynamics properties. Finally, we also present some real-world task evaluation demonstrations in Fig. 4.", + "bbox": [ + 89, + 702, + 483, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Experiment results and analysis. To further strengthen our conclusions that sparsely-gated MoE can better learn trajectory model from out-of-domain data and our adaptive policy", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "condition technique could lead to more flexible guidance for policy learning, we conduct real-world ablation experiments, as shown in Tab. 4. The results indicate that simply introducing human video data cannot lead to performance improvement. However, when expanding the model by MoE architecture, the performance increases from $38\\%$ to $46\\%$ . We speculate this may be due to the significant differences in appearance and dynamics between human and robot, with the former typically moving much faster and more flexibly. Finally, when further incorporating an adaptive trajectory mask for policy conditioning, the performance is further increased from $46\\%$ to $56\\%$ .", + "bbox": [ + 511, + 435, + 906, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 619, + 633, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we have introduced Tra-MoE, a novel framework that utilizes a sparsely-gated MoE architecture with Top-1 gating strategy to learn trajectory model from multiple domains. Tra-MoE achieves superior parameter cooperation and specialization, enabling more effective utilization of large-scale out-of-domain data while maintaining constant FLOPs per token. We train our Tra-MoE and conduct a comprehensive empirical study, demonstrating that TraMoE consistently exhibits superior performance compared to the dense baseline, even when the latter is scaled to match Tra-MoE's parameter count. Additionally, we propose an adaptive policy conditioning technique for trajectory-guided policy, ensuring adaptive 2D trajectory representation and explicit spatial alignment with observations, thereby achieving superior performance. Extensive experiments in both simulated and real-world environments validate our findings and demonstrate the effectiveness of our proposed methods.", + "bbox": [ + 511, + 643, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6967", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 89, + 90, + 230, + 106 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the National Key R&D Program of China (No. 2022ZD0160900), Jiangsu Frontier Technology Research and Development Program (No. BF2024076), and the Collaborative Innovation Center of Novel Software Technology and Industrialization.", + "bbox": [ + 89, + 106, + 485, + 183 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 191, + 187, + 208 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1", + "[2] OpenAI: Marcin Andrychowicz, Bowen Baker, Maciek Chogiej, Rafal Jozefowicz, Bob McGrew, Jakub Pachocki, Arthur Petron, Matthias Plappert, Glenn Powell, Alex Ray, et al. Learning dexterous in-hand manipulation. *IJRR*, 39(1): 3-20, 2020. 1", + "[3] Shikhar Bahl, Russell Mendonca, Lili Chen, Unnat Jain, and Deepak Pathak. Affordances from human videos as a versatile representation for robotics. In CVPR, pages 13778-13790, 2023. 1, 3", + "[4] Homanga Bharadhwaj, Abhinav Gupta, Vikash Kumar, and Shubham Tulsiani. Towards generalizable zero-shot manipulation via translating human interaction plans. In ICRA, pages 6904-6911. IEEE, 2024. 1, 3", + "[5] Homanga Bharadhwaj, Roozbeh Mottaghi, Abhinav Gupta, and Shubham Tulsiani. Track2act: Predicting point tracks from internet videos enables diverse zero-shot robot manipulation. arXiv preprint arXiv:2405.01527, 2024. 1, 3", + "[6] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. 2", + "[7] Konstantinos Bousmalis, Giulia Vezzani, Dushyant Rao, Coline Manon Devin, Alex X Lee, Maria Bauza Villalonga, Todor Davchev, Yuxiang Zhou, Agrim Gupta, Akhil Raju, et al. Robocat: A self-improving generalist agent for robotic manipulation. TMLR, 2023. 2", + "[8] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 2, 3", + "[9] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 2", + "[10] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1", + "[11] Qingwen Bu, Jisong Cai, Li Chen, Xiuqi Cui, Yan Ding, Siyuan Feng, Shenyuan Gao, Xindong He, Xu Huang, Shu Jiang, et al. Agibot world colosseo: A large-scale manipulation platform for scalable and intelligent embodied systems. arXiv preprint arXiv:2503.06669, 2025. 2" + ], + "bbox": [ + 93, + 217, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Weilin Cai, Juyong Jiang, Fan Wang, Jing Tang, Sunghun Kim, and Jiayi Huang. A survey on mixture of experts. arXiv preprint arXiv:2407.06204, 2024. 3", + "[13] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, pages 9640-9649, 2021. 1", + "[14] Zitian Chen, Yikang Shen, Mingyu Ding, Zhenfang Chen, Hengshuang Zhao, Erik G Learned-Miller, and Chuang Gan. Mod-squad: Designing mixtures of experts as modular multitask learners. In CVPR, pages 11828-11837, 2023. 3", + "[15] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1", + "[16] Ria Doshi, Homer Rich Walke, Oier Mees, Sudeep Dasari, and Sergey Levine. Scaling cross-embodied learning: One policy for manipulation, navigation, locomotion and aviation. In CoRL, 2", + "[17] Yilun Du, Sherry Yang, Pete Florence, Fei Xia, Ayzaan Wahid, Pierre Sermanet, Tianhe Yu, Pieter Abbeel, Joshua B Tenenbaum, Leslie Pack Kaelbling, et al. Video language planning. In The Twelfth International Conference on Learning Representations. 3", + "[18] Yilun Du, Sherry Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Josh Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation. NeurIPS, 36, 2024. 3", + "[19] Hao-Shu Fang, Hongjie Fang, Zhenyu Tang, Jirong Liu, Junbo Wang, Haoyi Zhu, and Cewu Lu. Rh20t: A robotic dataset for learning diverse skills in one-shot. In RSS 2023 Workshop on Learning for Task and Motion Planning, 2023.", + "[20] Chongkai Gao, Haozhuo Zhang, Zhixuan Xu, Cai Zhehao, and Lin Shao. Flip: Flow-centric generative planning as general-purpose manipulation world model. In ICLR. 1", + "[21] Jiayuan Gu, Sean Kirmani, Paul Wohlhart, Yao Lu, Montserrat Gonzalez Arenas, Kanishka Rao, Wenhao Yu, Chuyuan Fu, Keerthana Gopalakrishnan, Zhuo Xu, et al. Rt-trajectory: Robotic task generalization via hindsight trajectory sketches. In ICLR. 3, 7", + "[22] Pierre-Louis Guhur, Shizhe Chen, Ricardo Garcia Pinel, Makarand Tapaswi, Ivan Laptev, and Cordelia Schmid. Instruction-driven history-aware policies for robotic manipulations. In CORL, pages 175-187. PMLR, 2023. 3", + "[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1", + "[24] Changho Hwang, Wei Cui, Yifan Xiong, Ziyue Yang, Ze Liu, Han Hu, Zilong Wang, Rafael Salas, Jithin Jose, Prabhat Ram, et al. Tutel: Adaptive mixture-of-experts at scale. Proceedings of Machine Learning and Systems, 5:269-287, 2023. 3", + "[25] Vidhi Jain, Maria Attarian, Nikhil J Joshi, Ayzaan Wahid, Danny Driess, Quan Vuong, Pannag R Sanketi, Pierre Sermanet, Stefan Welker, Christine Chan, et al. Vid2robot: End-to-end video-conditioned policy learning with cross-attention transformers. arXiv preprint arXiv:2403.12943, 2024. 3", + "[26] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J Davison. Rlbench: The robot learning benchmark & learning environment. RAL, 5(2):3019-3026, 2020. 6" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6968", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kappler, Frederik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. Bc-z: Zero-shot task generalization with robotic imitation learning. In CoRL, pages 991-1002. PMLR, 2022. 3", + "[28] Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. Mixtral of experts. arXiv preprint arXiv:2401.04088, 2024. 3", + "[29] Yunfan Jiang, Agrim Gupta, Zichen Zhang, Guanzhi Wang, Yongqiang Dou, Yanjun Chen, Li Fei-Fei, Anima Anandkumar, Yuke Zhu, and Linxi Fan. Vima: General robot manipulation with multimodal prompts. arXiv preprint arXiv:2210.03094, 2(3):6, 2022. 2", + "[30] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 2", + "[31] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv preprint arXiv:2307.07635, 2023. 5", + "[32] Alexander Khazatsky, Karl Pertsch, Suraj Nair, Ashwin Balakrishna, Sudeep Dasari, Siddharth Karamcheti, Soroush Nasiriany, Mohan Kumar Srirama, Lawrence Yunliang Chen, Kirsty Ellis, et al. Droid: A large-scale in-the-wild robot manipulation dataset. arXiv preprint arXiv:2403.12945, 2024. 2", + "[33] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 2", + "[34] Po-Chen Ko, Jiayuan Mao, Yilun Du, Shao-Hua Sun, and Joshua B Tenenbaum. Learning to act from actionless videos through dense correspondences. In ICLR. 3", + "[35] Alexander Koch. Low-cost robot arm. https://github.com/AlexanderKoch-Koch/low_costrobot, 2024. GitHub repository. 5, 8", + "[36] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan First, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. arXiv preprint arXiv:2006.16668, 2020. 2, 3, 5", + "[37] Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuomotor policies. JMLR, 17(39):1-40, 2016. 1", + "[38] Xinghang Li, Minghuan Liu, Hanbo Zhang, Cunjun Yu, Jie Xu, Hongtao Wu, Chilam Cheang, Ya Jing, Weinan Zhang, Huaping Liu, et al. Vision-language foundation models as effective robot imitators. In ICLR. 2", + "[39] Bin Lin, Zhenyu Tang, Yang Ye, Jiaxi Cui, Bin Zhu, Peng Jin, Junwu Zhang, Munan Ning, and Li Yuan. Moe-llava: Mixture of experts for large vision-language models. arXiv preprint arXiv:2401.15947, 2024. 3", + "[40] Aixin Liu, Bei Feng, Bin Wang, Bingxuan Wang, Bo Liu, Chenggang Zhao, Chengqi Dengr, Chong Ruan, Damai Dai," + ], + "bbox": [ + 91, + 92, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Daya Guo, et al. Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model. arXiv preprint arXiv:2405.04434, 2024. 3", + "[41] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning. NeurIPS, 36, 2024. 5", + "[42] Fangchen Liu, Kuan Fang, Pieter Abbeel, and Sergey Levine. Moka: Open-vocabulary robotic manipulation through mark-based visual prompting. In First Workshop on Vision-Language Models for Navigation and Manipulation at ICRA 2024. 3", + "[43] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. In ICLR, 2022. 1", + "[44] Corey Lynch, Mohi Khansari, Ted Xiao, Vikash Kumar, Jonathan Tompson, Sergey Levine, and Pierre Sermanet. Learning latent plans from play. In CoRL, pages 1113-1132. PMLR, 2020. 3", + "[45] Corey Lynch, Ayzaan Wahid, Jonathan Tompson, Tianli Ding, James Betker, Robert Baruch, Travis Armstrong, and Pete Florence. Interactive language: Talking to robots in real time. RAL, 2023. 3", + "[46] Soroush Nasiriany, Sean Kirmani, Tianli Ding, Laura Smith, Yuke Zhu, Danny Driess, Dorsa Sadigh, and Ted Xiao. Rt-affordance: Affordances are versatile intermediate representations for robot manipulation. arXiv preprint arXiv:2411.02704, 2024. 3", + "[47] Abhishek Padalkar, Acorn Pooley, Ajinkya Jain, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anikait Singh, Anthony Brohan, et al. Open x-embediment: Robotic learning datasets and rt-x models. arXiv preprint arXiv:2310.08864, 2023. 2", + "[48] Joan Puigcerver, Carlos Riquelme, Basil Mustafa, and Neil Houlsby. From sparse to soft mixtures of experts. arXiv preprint arXiv:2308.00951, 2023. 3", + "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICLR, pages 8748-8763. PMLR, 2021. 1", + "[50] Rouhollah Rahmatizadeh, Pooya Abolghasemi, Ladislau Böloni, and Sergey Levine. Vision-based multi-task manipulation for inexpensive robots using end-to-end learning from demonstration. In ICRA, pages 3758-3765. IEEE, 2018. 3", + "[51] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. arXiv preprint arXiv:2205.06175, 2022. 1, 2", + "[52] Younggyo Seo, Kimin Lee, Stephen L James, and Pieter Abbeel. Reinforcement learning with action-free pre-training from videos. In ICLR, pages 19561-19579. PMLR, 2022. 1", + "[53] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6969", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In CoRL, pages 894-906. PMLR, 2022. 3", + "[55] Austin Stone, Ted Xiao, Yao Lu, Keerthana Gopalakrishnan, Kuang-Huei Lee, Quan Vuong, Paul Wohlhart, Brianna Zitkovich, Fei Xia, Chelsea Finn, et al. Open-world object manipulation using pre-trained vision-language models. In CoRL, 2023. 3", + "[56] Priya Sundaresan, Quan Vuong, Jiayuan Gu, Peng Xu, Ted Xiao, Sean Kirmani, Tianhe Yu, Michael Stark, Ajinkya Jain, Karol Hausman, et al. Rt-sketch: Goal-conditioned imitation learning from hand-drawn sketches. 2024. 3", + "[57] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. 2", + "[58] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 1", + "[59] Mel Vecerik, Carl Doersch, Yi Yang, Todor Davchev, Yusuf Aytar, Guangyao Zhou, Raia Hadsell, Lourdes Agapito, and Jon Scholz. Robotap: Tracking arbitrary points for few-shot visual imitation. In ICRA, pages 5397-5403. IEEE, 2024. 1", + "[60] Homer Rich Walke, Kevin Black, Tony Z Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen-Estruch, Andre Wang He, Vivek Myers, Moo Jin Kim, Max Du, et al. Bridgedata v2: A dataset for robot learning at scale. In CoRL, pages 1723-1736. PMLR, 2023. 2", + "[61] Chen Wang, Linxi Fan, Jiankai Sun, Ruohan Zhang, Li Fei-Fei, Danfei Xu, Yuke Zhu, and Anima Anandkumar. Mimicplay: Long-horizon imitation learning by watching human play. In CoRL. 1, 3", + "[62] Lirui Wang, Xinlei Chen, Jialiang Zhao, and Kaiming He. Scaling proprioceptive-visual learning with heterogeneous pre-trained transformers. CoRR, abs/2409.20537, 2024. 3", + "[63] Chuan Wen, Xingyu Lin, John So, Kai Chen, Qi Dou, Yang Gao, and Pieter Abbeel. Any-point trajectory modeling for policy learning. arXiv preprint arXiv:2401.00025, 2023. 1, 3, 6", + "[64] Kun Wu, Chengkai Hou, Jiaming Liu, Zhengping Che, Xiaozhu Ju, Zhuqin Yang, Meng Li, Yinuo Zhao, Zhiyuan Xu, Guang Yang, et al. Robomind: Benchmark on multi-embediment intelligence normative data for robot manipulation. arXiv preprint arXiv:2412.13877, 2024. 2", + "[65] Mengda Xu, Zhenjia Xu, Yinghao Xu, Cheng Chi, Gordon Wetzstein, Manuela Veloso, and Shuran Song. Flow as the cross-domain manipulation interface. arXiv preprint arXiv:2407.15208, 2024. 1, 3", + "[66] Jiange Yang, Wenhui Tan, Chuhao Jin, Bei Liu, Jianlong Fu, Ruihua Song, and Limin Wang. Pave the way to grasp anything: Transferring foundation models for universal pickplace robots. arXiv preprint arXiv:2306.05716, 1(2), 2023. 3", + "[67] Jiange Yang, Bei Liu, Jianlong Fu, Bocheng Pan, Gangshan Wu, and Limin Wang. Spatiotemporal predictive pre-training for robotic motor control. arXiv preprint arXiv:2403.05304, 2024. 3" + ], + "bbox": [ + 91, + 90, + 485, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[68] Jiange Yang, Wenhui Tan, Chuhao Jin, Keling Yao, Bei Liu, Jianlong Fu, Ruihua Song, Gangshan Wu, and Limin Wang. Transferring foundation models for generalizable robotic manipulation. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), pages 1999-2010, 2025. 3", + "[69] Seonghyeon Ye, Joel Jang, Byeongguk Jeon, Sejune Joo, Jianwei Yang, Baolin Peng, Ajay Mandlekar, Reuben Tan, Yu-Wei Chao, Bill Yuchen Lin, et al. Latent action pretraining from videos. arXiv preprint arXiv:2410.11758, 2024. 1", + "[70] Chengbo Yuan, Chuan Wen, Tong Zhang, and Yang Gao. General flow as foundation affordance for scalable robot learning. arXiv preprint arXiv:2401.11439, 2024. 3", + "[71] Jinguo Zhu, Xizhou Zhu, Wenhai Wang, Xiaohua Wang, Hongsheng Li, Xiaogang Wang, and Jifeng Dai. Uniperceiver-moe: Learning sparse generalist models with conditional moes. NeurIPS, 35:2664-2678, 2022. 1", + "[72] Xizhou Zhu, Jinguo Zhu, Hao Li, Xiaoshi Wu, Hongsheng Li, Xiaohua Wang, and Jifeng Dai. Uni-perceiver: Pre-training unified architecture for generic perception for zero-shot and few-shot tasks. In CVPR, pages 16804-16815, 2022. 1", + "[73] Yifeng Zhu, Zhenyu Jiang, Peter Stone, and Yuke Zhu. Learning generalizable manipulation policies with object-centric 3d representations. In CoRL. 3", + "[74] Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam Shazeer, and William Fedus. St-moe: Designing stable and transferable sparse expert models. arXiv preprint arXiv:2202.08906, 2022. 2, 3, 5" + ], + "bbox": [ + 514, + 90, + 906, + 487 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6970", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Tra-MoE_ Learning Trajectory Prediction Model from Multiple Domains for Adaptive Policy Conditioning/d41e3854-0a29-4aee-85fc-0fcf3b667c89_model.json b/2025/Tra-MoE_ Learning Trajectory Prediction Model from Multiple Domains for Adaptive Policy Conditioning/d41e3854-0a29-4aee-85fc-0fcf3b667c89_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0575527b28cb8147205743ef24f3979e2fa436c7 --- /dev/null +++ b/2025/Tra-MoE_ Learning Trajectory Prediction Model from Multiple Domains for Adaptive Policy Conditioning/d41e3854-0a29-4aee-85fc-0fcf3b667c89_model.json @@ -0,0 +1,2389 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.131, + 0.13, + 0.87, + 0.177 + ], + "angle": 0, + "content": "Tra-MoE: Learning Trajectory Prediction Model from Multiple Domains for Adaptive Policy Conditioning" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.202, + 0.896, + 0.221 + ], + "angle": 0, + "content": "Jiange Yang \\(^{1,2}\\) Haoyi Zhu \\(^{2,3}\\) Yating Wang \\(^{2,4}\\) Gangshan Wu \\(^{1}\\) Tong He \\(^{2}\\) Limin Wang \\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.222, + 0.872, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\)State Key Laboratory for Novel Software Technology, Nanjing University, China \\(^{2}\\)Shanghai AI Laboratory, \\(^{3}\\)University of Science and Technology of China, \\(^{4}\\)Tongji University" + }, + { + "type": "text", + "bbox": [ + 0.274, + 0.26, + 0.721, + 0.275 + ], + "angle": 0, + "content": "jiangyang.jgy@gmail.com, {gswu,lmwang}@nju.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.31, + 0.278, + 0.689, + 0.292 + ], + "angle": 0, + "content": "{zhuhaoyi, wangyating, hetong}@pjlab.org.cn" + }, + { + "type": "text", + "bbox": [ + 0.327, + 0.295, + 0.666, + 0.309 + ], + "angle": 0, + "content": "https://github.com/MCG-NJU/Tra-MoE" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.345, + 0.327, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.377, + 0.486, + 0.755 + ], + "angle": 0, + "content": "Learning from multiple domains is a primary factor that influences the generalization of a single unified robot system. In this paper, we aim to learn the trajectory prediction model by using broad out-of-domain data to improve its performance and generalization ability. Trajectory model is designed to predict any-point trajectories in the current frame given an instruction and can provide detailed control guidance for robotic policy learning. To handle the diverse out-of-domain data distribution, we propose a sparsely-gated MoE (Top-1 gating strategy) architecture for trajectory model, coined as Tra-MoE. The sparse activation design enables good balance between parameter cooperation and specialization, effectively benefiting from large-scale out-of-domain data while maintaining constant FLOPs per token. In addition, we further introduce an adaptive policy conditioning technique by learning 2D mask representations for predicted trajectories, which is explicitly aligned with image observations to guide action prediction more flexibly. We perform extensive experiments on both simulation and real-world scenarios to verify the effectiveness of Tra-MoE and adaptive policy conditioning technique. We also conduct a comprehensive empirical study to train Tra-MoE, demonstrating that our Tra-MoE consistently exhibits superior performance compared to the dense baseline model, even when the latter is scaled to match Tra-MoE's parameter count." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.763, + 0.222, + 0.779 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.786, + 0.486, + 0.879 + ], + "angle": 0, + "content": "In computer vision and natural language processing, significant progress has been made through learning models from multiple domains data. For example, many works use board data to perform self-supervised pre-training, and then adapt to diverse downstream tasks through zero-shot [1, 10, 49] or fine-tuning [13, 15, 23, 58]. Additionally, some research ef" + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.347, + 0.871, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.568, + 0.907, + 0.597 + ], + "angle": 0, + "content": "Figure 1. (a) Training with small-scale, in-domain data. (b) Joint training with in-domain data and large-scale, out-of-domain data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.909, + 0.777 + ], + "angle": 0, + "content": "forts [43, 51, 71, 72] focus on learning a generalist model capable of unifying various tasks. In contrast, in robot learning, due to data scarcity and homogeneity, some groundbreaking methods [2, 37] resort to training only using in-domain data. Recently, some works [3, 4, 52, 61, 69] have shifted to learn from action-free human video data. A particularly scalable and promising paradigm [5, 20, 59, 63, 65] involves learning a trajectory prediction model from action-free video data, and then using small-scale action-labeled demonstrations to learn trajectory-guided policies, thereby achieving greater sample efficiency and better generalization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Despite the potential benefits, it still remains underexplored how to effectively leverage broad out-of-domain video data for jointly training trajectory models. In particular, the impact of incorporating data encompassing diverse environments, objects, skills and embodiments has not fully investigated for in-domain tasks. For instance, ATM [63], a notable work in this field, mainly relies on in-domain data (same tasks and environments) for training both trajectory" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.237, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6960" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.092, + 0.298, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.09, + 0.478, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.267, + 0.483, + 0.337 + ], + "angle": 0, + "content": "Figure 2. Training trajectory prediction model from multiple domains. (a) The visualization of Dataset \\(\\mathcal{D}_{\\text {ood }}\\) and \\(\\mathcal{D}_{\\text {in }}\\). The former may contain additional environments, objects, skills and embodiments. (b) Our pipeline: first co-training trajectory prediction model and then adapting it for downstream policy learning." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.342, + 0.483, + 0.614 + ], + "angle": 0, + "content": "models and policies. We argue that learning from multiple domains is the primary factor to build a powerful trajectory model for policy guidance. Therefore, we propose to train trajectory model under the setup of jointly learning from large-scale out-of-domain and small-scale in-domain data, as shown in Fig. 2(a). However, directly using large-scale out-of-domain video data for hybrid pre-training involves two critical challenges: (i) Joint training on data from different environments, objects, skills, and embodiments will add learning difficulty and might lead to optimization conflicts. It is unclear how to effectively fuse the complementary information of multiple domains data within a single unified model. Naively expanding the out-of-domain training data does not necessarily lead to improved performance on in-domain tasks, as shown in Fig. 1, which results in a 5.6 \\((57.6\\to 52.0)\\) performance decrease on in-domain tasks. (ii) How to ensure both high performance and computational efficiency during trajectory model scaling up process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To address these challenges, we design a new sparsely-gated Mixture-of-Expert (MoE) architecture [53] to scale up our trajectory model, coined as Tra-MoE. Specifically, we integrate several MoE blocks to replace the original transformer blocks. This design enables efficient joint training of most parameters on the broad out-of-domain data, capturing the complementary patterns for mutual cooperation and the shared knowledge for better generalization. Meanwhile, for different data and different tokens within the same data, our Tra-MoE naturally forms a specialization when activating different experts. To maintain high computational efficiency while scaling up model capacity, we implement a top-1 gating strategy for token-choice, ensuring constant FLOPs per token. In sparse MoE, some auxiliary losses are commonly used to enhance performance, including the router z-loss [74] for improving training stability and the load-balancing loss [36] for balancing expert activations. We further conduct a comprehensive empirical study on scaling up trajectory models and find that the touter z-loss improves" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.302 + ], + "angle": 0, + "content": "performance while the load-balancing loss tends to cause performance degradation. Furthermore, we observe that when expanding pre-training data from small-scale in-domain to large-scale in-domain and out-of-domain data, employing MoE (Tra-MoE) can lead to obvious performance improvement on in-domain tasks, whereas the dense counterpart (Tra-baseline) suffers from performance drop, as illustrated in Fig. 1. Notably, even when expanding the dense Tra-baseline's parameter to match that of Tra-MoE, performance improvements remain elusive. These findings collectively demonstrate the effectiveness of our sparsely-gated MoE architecture, attributing its capability of balancing parameter cooperation and specialization when expanding training with out-of-domain data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.303, + 0.909, + 0.453 + ], + "angle": 0, + "content": "Furthermore, given a pre-trained trajectory model, how to effectively condition the policy module using the 2D trajectories still remain an unresolved challenge. We further propose an adaptive policy conditioning technique for trajectory-guided policy. This enables better spatial alignment between 2D trajectories and images, while achieving more flexible trajectory representation and guidance for the policy. Finally, we also conduct extensive real-world experiments to further strengthen our conclusions. In summary, our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.456, + 0.907, + 0.516 + ], + "angle": 0, + "content": "- We propose a sparsely-gated MoE architecture to learn trajectory models, called as Tra-MoE, with large-scale, out-of-domain, and action-free video data, coupled with a comprehensive empirical study on its training techniques." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.517, + 0.907, + 0.591 + ], + "angle": 0, + "content": "- Extensive simulation and real-world environments demonstrate that our Tra-MoE (with Top-1 gating) can more effectively benefit from large-scale out-of-domain data (such as cross-embediments data and different physics engine simulated data) compared to the dense baseline." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.592, + 0.907, + 0.667 + ], + "angle": 0, + "content": "- We demonstrate that Tra-MoE still significantly outperform expanded dense counterpart with equivalent parameters, highlighting the effectiveness of sparse MoE architecture in leveraging out-of-domain data through better parameter cooperation and specialization." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.668, + 0.907, + 0.727 + ], + "angle": 0, + "content": "- We propose an adaptive policy conditioning technique for trajectory-guided policy, which ensures adaptive 2D trajectory representation and explicit spatial alignment with the observations, thereby achieving superior performance." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.456, + 0.907, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.74, + 0.655, + 0.755 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Learning from multiple domains in Robot Learning. Scaling laws [30] have fueled substantial advancements in various fields. In robot learning, many recent studies [7, 8, 29, 51] investigate how compute, model size, and training data quantity affect the model performance on robotic manipulation tasks. Based on this, recent works have focused on collecting larger-scale real-world robotic datasets [11, 19, 32, 60, 64], and have successfully trained several robotics foundational models [6, 9, 16, 33, 38, 47, 57," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6961" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[62] that demonstrate superior transfer capabilities. It is worth noting that these works mainly employ pre-training and fine-tuning paradigm rather than joint co-training paradigm, while focusing on evaluating the combination generalization and neglecting the performance impact on in-domain tasks. Additionally, these works directly combine heterogeneous robotic data with different action and observation spaces for joint training, potentially leading to sub-optimal solutions. In contrast, we leverage sparsely-gated MoE to incorporate larger-scale, out-of-domain, and action-less video data. Furthermore, our experiments demonstrate the importance of MoE for expanding out-of-domain data due to its superior capacity for parameter cooperation and specialization." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.293, + 0.487, + 0.626 + ], + "angle": 0, + "content": "Mixture-of-Experts. The Mixture-of-Experts (MoE) framework [12] is built upon a simple yet effective concept: dividing a model into specialized components, or 'experts', each focusing on distinct tasks or data aspects. This paradigm allows for selective activation of relevant experts based on the input, thereby maintaining computational efficiency while leveraging a vast reservoir of specialized knowledge through increased model capacity. Recently, in computer vision, natural language processing, and multi-modal, there has been a surge of foundation models trained based on the MoE architecture, including Swin-Moe [24], MoE-LLaVA [39], DeepSeek-V2 [40], Mixtral-8x22B [28], and so on. MoE typically encompasses dense [48] and sparse [14] variants. Dense MoE activates all experts in each iteration, while sparse MoE activates only a subset, generally resulting in lower computational overhead. However, sparse MoE can suffer from expert load imbalance and training instability issues. Techniques such as adding noise, load balancing loss [36], and router z-loss [74] are often employed to mitigate these issues and improve performance. In our work, we conduct a comprehensive empirical study of these techniques when training our sparsely-gated MoE-based trajectory model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.487, + 0.902 + ], + "angle": 0, + "content": "Representations and Techniques for Policy Conditioning. Policy conditioning equips a multi-task robot policy with the ability to modulate its behavior by incorporating task-specific guidance information. This information, which may take various forms, guides the policy's action generation, enabling a single model to effectively learn and perform diverse tasks. Classic policy conditioning representations often encompass task identifiers [50], natural language instructions [8, 22, 45, 54], goal images [44], and demonstration videos [25, 27]. These are typically integrated with visual observations within the policy through concatenation or attention mechanisms. Recent works have explored alternative policy conditioning representations to enhance generalization and sample efficiency. These include leveraging internet-trained models to obtain bounding box [55] and object mask [42, 66, 68, 73], as well as learning plan [4], affordance [3, 46], future observations [17, 18, 34, 67], 2D or 3D trajectory [5, 61, 63, 65, 70] from human video data," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.213 + ], + "angle": 0, + "content": "and also hand-drawn trajectory [21] or sketch [56]. Different from these approaches that directly utilize raw or manually crafted signals for integration with visual observations, we introduce an adaptive policy conditioning technique that explicitly maps 2D trajectories onto the visual observations in space and encodes them as learnable embeddings. These designs ensure adaptive 2D trajectory representation and explicit spatial alignment with the visual observations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.226, + 0.65, + 0.243 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.251, + 0.909, + 0.356 + ], + "angle": 0, + "content": "In this section, we first describe our problem formulation in Sec. 3.1. Subsequently, we give an overview of our proposed framework, which is composed of a sparsely-gated MoE-based trajectory prediction model and a trajectory-guided policy model using the adaptive policy conditioning technique, detailed in Sec. 3.2. Finally, we describe our training process and loss functions in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.366, + 0.719, + 0.38 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.909, + 0.553 + ], + "angle": 0, + "content": "Following the setup of previous works [5, 63, 65], our approach encompasses a trajectory prediction model to conduct language-conditioned any-point trajectory prediction and a trajectory-guided policy model to predict executable robot actions. In particular, as depicted in Fig. 2, we consider two types of datasets for training, which are large-scale, out-of-domain, and action-free videos \\( \\mathcal{D}_{\\text{ood}} \\), as well as small-scale, in-domain, and action-labeled demonstrations \\( \\mathcal{D}_{\\text{in}} \\). For trajectory prediction model, we employ \\( \\mathcal{D}_{\\text{ood}} \\) and \\( \\mathcal{D}_{\\text{in}} \\) for joint pre-training, whereas for trajectory-guided policy model, we only use \\( \\mathcal{D}_{\\text{in}} \\) for training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.562, + 0.649, + 0.576 + ], + "angle": 0, + "content": "3.2. Architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.909, + 0.827 + ], + "angle": 0, + "content": "ATM Revisited. In this paper, we use ATM [63] as the baseline for our method. The overall pipeline of ATM is illustrated on Fig. 2(b). It first learns a trajectory model from video data to perform language-conditioned trajectory prediction. The predicted trajectories is then used to learn a trajectory-guided policy using robot demonstrations. Specifically, as for trajectory model, its objective is to predict the future of any-point trajectories in a frame. Formally, given an image observation \\(o_t\\) at timestep \\(t\\), any set of 2D query points simplified on the image frame \\(\\mathbf{p}_t = \\{p_{t,k}\\}_{k=1}^K\\), and a language instruction \\(\\ell\\), the trajectory model learns a mapping \\(\\mathbf{p}_{t:t+H} = \\tau_\\theta(o_t, \\mathbf{p}_t, \\ell)\\) that predicts the coordinates of these query points in the camera frame for the next \\(H\\) time steps, where \\(\\mathbf{p}_{t:t+H} \\in \\mathbb{R}^{H \\times K \\times 2}\\). For the trajectory-guided policy, it leverages the current observation \\(o_t\\) and the predicted trajectory \\(\\mathbf{p}_{t:t+H}\\) to predict the subsequent robot actions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.827, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Sparsely-gated MoE-based trajectory model (Tra-MoE). ATM proposes a track transformer to implement the trajectory model. In track transformer, language tokens, image tokens, and track tokens to be decoded involve global interactions within the transformer block. However, directly" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6962" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.152, + 0.089, + 0.558, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.09, + 0.837, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.427, + 0.908, + 0.47 + ], + "angle": 0, + "content": "Figure 3. (a) The pipeline of our sparsely-gated MoE-based trajectory model (Tra-MoE). (b) The pipeline of our trajectory-guided policy using the adaptive policy conditioning technique. Mapping means concatenating the trajectory mask with image observations, while learnable refers to setting each point in the trajectory mask as a learnable embedding." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.478, + 0.486, + 0.705 + ], + "angle": 0, + "content": "processing diverse domains and multiple modalities data within a unified transformer may yield suboptimal results. Hence, we propose our Tra-MoE framework. To develop our sparsely-gated MoE-based trajectory model, as depicted in Fig. 3(a), we replace certain layers of the transformer with MoE blocks. In each MoE block, it incorporates multiple feedforward networks (FFNs), each designated as an expert, and utilizes a gating function to activate a selected subset (Top- \\(K\\)) of these experts. In this paper, we set \\(K\\) to a fixed value of 1 to ensure that the FLOPs per token remain constant relative to the baseline. Formally, the gating network \\(\\mathcal{G}\\), parameterized by \\(\\Theta\\) and typically consisting of a linear-softmax network, yields the output \\(\\mathcal{G}(\\mathbf{x_s};\\Theta)\\), where \\(\\mathbf{x_s}\\) represents our token sequences input. We formulate the sparsely-gated token-choice mechanism as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.747, + 0.486, + 0.765 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} \\left(\\mathbf {x} _ {\\mathrm {s}}; \\boldsymbol {\\Theta}\\right) _ {i} = \\operatorname {s o f t m a x} (\\operatorname {T o p - K} (g \\left(\\mathbf {x} _ {\\mathrm {s}}; \\boldsymbol {\\Theta}\\right), k)) _ {i}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.769, + 0.484, + 0.808 + ], + "angle": 0, + "content": "\\[\n\\operatorname {T o p - K} (g \\left(\\mathbf {x} _ {\\mathbf {s}}; \\boldsymbol {\\Theta}\\right), k) _ {i} = \\left\\{ \\begin{array}{l l} g (\\mathbf {x}; \\boldsymbol {\\Theta}) _ {i}, & \\text {c o n d ,} \\\\ - \\infty , & \\text {o t h .} \\end{array} \\right., \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.811, + 0.484, + 0.829 + ], + "angle": 0, + "content": "\\[\n\\operatorname {c o n d}: \\text {i f} g (\\mathbf {x}; \\boldsymbol {\\Theta}) _ {i} \\text {i s i n t o p - k e l e m e n t s o f} g (\\mathbf {x}; \\boldsymbol {\\Theta}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.486, + 0.903 + ], + "angle": 0, + "content": "More specifically, \\(\\operatorname{Top - K}(\\cdot ,k)\\) function retains only the top- \\(k\\) values in a vector, setting the rest to \\(-\\infty\\). Consequently, the output of the sparsely-gated MoE layer can be" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.478, + 0.613, + 0.492 + ], + "angle": 0, + "content": "formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.501, + 0.907, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\text {s p a r s e}} ^ {\\mathrm {M o E}} \\left(\\mathbf {x} _ {\\mathbf {s}}; \\boldsymbol {\\Theta}, \\left\\{\\mathbf {W} _ {i} \\right\\} _ {i = 1} ^ {K}\\right) = \\sum_ {i = 1} ^ {K} \\mathcal {G} \\left(\\mathbf {x} _ {\\mathbf {s}}; \\boldsymbol {\\Theta}\\right) _ {i} f _ {i} \\left(\\mathbf {x} _ {\\mathbf {s}}; \\mathbf {W} _ {i}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.908, + 0.599 + ], + "angle": 0, + "content": "where \\(f_{i}\\) represents the i-th expert, usually a linear-GELU-linear FFN layer, is parameterized by \\(\\mathbf{W}_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.765 + ], + "angle": 0, + "content": "Trajectory-guided policy using the adaptive policy conditioning technique. After completing the trajectory model learning, we develop our trajectory-guided policy through the adaptive policy conditioning technique based on the initial ATM, as shown in Fig. 3(b). In our policy model, independently encoded track tokens and image tokens undergo a early fusion within the fusion transformer. Subsequently, the fused features are further integrated with the raw trajectory data through a late fusion mechanism along the channel dimension, ensuring the preservation of original trajectory information for more accurate action prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Fundamentally, the policy model bridges the gap between the predicted 2D any-point trajectories and the executable 3D robot actions, functioning as an inverse dynamic model and eliminating the necessity of other task specification. However, how to effectively condition the policy using the predicted 2D trajectories remains an unresolved challenge. Based on our analysis, we present two key insights: (i) Explicit spatial alignment and fusion of the 2D trajectories with image observations turns out to be highly beneficial." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6963" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.227 + ], + "angle": 0, + "content": "It simplifies the mapping of 2D trajectories to image space and highlights motion-relevant regions, enabling the policy model to focus on important areas and reducing the learning difficulty. (ii) Trajectory positions exhibit different characteristics: starting points typically emphasize local motion, while endpoints focus more on global trend. Therefore, an adaptive trajectory representation and conditioning technique is expected to flexibly guide the policy by effectively capturing these variations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.484, + 0.424 + ], + "angle": 0, + "content": "Based on the aforementioned analysis, we propose an adaptive policy conditioning technique to incorporate trajectory information into 2D images. Specifically, we construct an additional mask modality, and populate it with 2D trajectories based on their specific spatial locations, while setting the value of each point in the trajectory mask as a learnable embedding. Before encoding the image, we concatenate this mask modality with the image observations on the channel dimension, forming a tensor of dimensions \\( H \\times W \\times 4 \\), as shown in Fig. 3(b). To sum up, our adaptive policy conditioning technique yield an adaptive 2D trajectory representation, which could be explicitly aligned with the image observations to guide the policy prediction more flexibly." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.431, + 0.196, + 0.447 + ], + "angle": 0, + "content": "3.3. Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.453, + 0.485, + 0.725 + ], + "angle": 0, + "content": "We argue that trajectory models could be shared by multiple domains data. Therefore, we first pre-train a generalizable trajectory model on \\( \\mathcal{D}_{\\text{ood}} \\cup \\mathcal{D}_{\\text{in}} \\). The \\( \\mathcal{D}_{\\text{ood}} \\) typically contains additional environments, objects, skills, and embodiments. Then we train the trajectory-guided policy model on \\( \\mathcal{D}_{\\text{in}} \\) based on the pre-trained trajectory model. In trajectory-guided policy training phase, we freeze the trajectory model. Pre-training sparsely-gated MoE-based trajectory model. Following ATM, we train the trajectory model using the ground truth tracks generated by CoTracker [31]. We use MSE to optimize the trajectory prediction loss \\( \\mathcal{L}_{\\text{tra}} \\). Additionally, we also consider adding an auxiliary load-balancing loss [36] to balance the activation frequency among experts. Given a batch of queries \\( \\mathcal{B} \\), it contains B samples and each sample contains S tokens. We define \\( \\mathcal{Q}_i \\) as the proportion of tokens distributed to expert \\( i \\), and \\( \\mathcal{P}_i \\) as the proportion of the gating probability assigned to expert \\( i \\). The \\( \\mathcal{L}_{\\text{lo-ba}} \\) can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.743, + 0.484, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Q} _ {i} = \\frac {1}{S \\cdot B} \\sum_ {x _ {s} ^ {m} \\in \\mathcal {B}} \\mathbb {1} \\left\\{\\operatorname {a r g m a x} \\mathcal {G} \\left(\\mathbf {x} _ {\\mathbf {s}} ^ {\\mathbf {m}}; \\boldsymbol {\\Theta}\\right) = i \\right\\}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.784, + 0.484, + 0.82 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} _ {i} = \\frac {1}{S \\cdot B} \\sum_ {x _ {\\mathrm {s}} ^ {m} \\in \\mathcal {B}} \\mathcal {G} \\left(\\mathbf {x} _ {\\mathrm {s}} ^ {\\mathrm {m}}; \\boldsymbol {\\Theta}\\right) _ {i}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.825, + 0.484, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {l o - b a}} = N \\sum_ {i = 1} ^ {N} \\mathcal {Q} _ {i} \\mathcal {P} _ {i}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Additionally, ST-MoE [74] proposed using the router \\( z \\)-loss to penalize large logits entering the gating network," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "thereby improving training stability. We similarly applies it to the training of our trajectory model, as shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.133, + 0.907, + 0.178 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {z} (x) = \\frac {1}{S} \\sum_ {k = 1} ^ {S} \\left(\\log \\sum_ {i = 1} ^ {N} e ^ {\\mathbf {g} _ {i} ^ {(k)}}\\right) ^ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.183, + 0.908, + 0.23 + ], + "angle": 0, + "content": "where \\(S\\) is the total number of tokens, \\(N\\) is the total number of experts, and \\(\\mathbf{g} \\in \\mathcal{R}^{S \\times N}\\) are the logits going into the router. In summary, the trajectory model's training loss is as follow:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.257, + 0.907, + 0.273 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {t r a}} \\cdot \\mathcal {L} _ {\\text {t r a}} + \\lambda_ {\\text {l o - b a}} \\cdot \\mathcal {L} _ {\\text {l o - b a}} + \\lambda_ {z} \\cdot \\mathcal {L} _ {z}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.285, + 0.908, + 0.36 + ], + "angle": 0, + "content": "Training trajectory-guided policy through adaptive policy conditioning technique. For trajectory-guided policy training, we employ the paradigm of behavior cloning. Specifically, we use the MSE between predicted actions and ground truths as the loss function." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.375, + 0.646, + 0.392 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.4, + 0.908, + 0.551 + ], + "angle": 0, + "content": "We perform evaluation experiments using the LIBERO benchmark [41] and a real low-cost dual-arm robot [35]. In experiments, we quantitatively find that the success rate of downstream manipulation tasks is generally positively correlated with the performance of the trajectory model. Detailed quantitative experiment results can be found in the appendix. Meanwhile, CoTracker [31] can provide relatively accurate trajectory annotations; therefore, we report the average success rate of downstream manipulation tasks as our metric. Our experiments aim to study the following questions:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.552, + 0.905, + 0.596 + ], + "angle": 0, + "content": "Q1: Does the techniques of router z-loss, load-balancing loss, and adding noise to gating logits enhance the training of sparsely-gated Tra-MoE?" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.597, + 0.907, + 0.657 + ], + "angle": 0, + "content": "Q2: Does sparsely-gated Tra-MoE, compared to the dense Tra-baseline, benefit more from large-scale out-of-domain data (such as cross-embediments and different physics engine data) and demonstrate superior scaling up capabilities?" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.658, + 0.905, + 0.702 + ], + "angle": 0, + "content": "Q3: When we scale up the dense Tra-baseline to match the parameter count of Tra-MoE, does the sparsely-gated Tra-MoE still maintain its advantage?" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.703, + 0.905, + 0.733 + ], + "angle": 0, + "content": "Q4: Is our proposed adaptive policy conditioning technique for trajectory-guided policy more effective than ATM?" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.552, + 0.907, + 0.733 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.738, + 0.759 + ], + "angle": 0, + "content": "4.1. Simulation Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Experiment setup. The LIBERO [41] is divided into five categories: LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, LIBERO-Long, and LIBERO-90. LIBERO-90 includes 90 tasks, while each of the other categories contains 10 tasks. Compared to other categories, LIBERO-90 presents a highly varied selection of object categories, layouts, environments, and skill goals. It is typically used for pre-training rather than joint co-training. Therefore, our main ablation experiments select 90 tasks from LIBERO-90 to construct" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6964" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.117, + 0.089, + 0.473, + 0.19 + ], + "angle": 0, + "content": "
SpatialGoalObjectLongAvg.
062.073.071.021.556.9
5·10-469.573.568.022.558.4
1·10-462.581.073.528.561.4
1·10-560.065.568.527.055.3
1·10-660.079.053.031.555.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.238, + 0.193, + 0.345, + 0.205 + ], + "angle": 0, + "content": "(a) The value of \\(\\lambda_z\\)" + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.21, + 0.479, + 0.267 + ], + "angle": 0, + "content": "
SpatialGoalObjectLongAvg.
w/o. noise62.073.071.022.556.9
w. noise71.567.057.527.055.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.176, + 0.269, + 0.408, + 0.281 + ], + "angle": 0, + "content": "(c) Adding a noise term to the gating logits." + }, + { + "type": "table", + "bbox": [ + 0.119, + 0.301, + 0.462, + 0.387 + ], + "angle": 0, + "content": "
Num.SpatialGoalObjectLongAvg.
149.567.056.535.052.0
259.072.565.527.056.0
364.068.071.017.555.1
462.073.071.021.556.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.22, + 0.389, + 0.364, + 0.402 + ], + "angle": 0, + "content": "(e) The number of experts." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.089, + 0.88, + 0.175 + ], + "angle": 0, + "content": "
SpatialGoalObjectLongAvg.
062.073.071.022.556.9
1·10-358.574.060.018.052.6
1·10-561.069.556.026.053.1
1·10-766.567.073.522.057.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.638, + 0.178, + 0.771, + 0.19 + ], + "angle": 0, + "content": "(b) The value of \\(\\lambda_{lo - ba}\\)." + }, + { + "type": "table", + "bbox": [ + 0.53, + 0.21, + 0.876, + 0.281 + ], + "angle": 0, + "content": "
SpatialGoalObjectLongAvg.
MoE62.581.073.528.561.4
Width66.550.561.515.048.4
Depth61.575.546.526.552.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.58, + 0.283, + 0.829, + 0.296 + ], + "angle": 0, + "content": "(d) Scaling Tra-baseline from width and depth." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.301, + 0.897, + 0.387 + ], + "angle": 0, + "content": "
MoEOODSpatialGoalObjectLongAvg.
67.568.568.026.557.6
✓63.064.060.519.551.8
✓49.567.056.535.052.0
✓✓62.581.073.528.561.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.602, + 0.389, + 0.798, + 0.401 + ], + "angle": 0, + "content": "(f) Scaling with MoE and OOD data." + }, + { + "type": "table_caption", + "bbox": [ + 0.16, + 0.403, + 0.835, + 0.417 + ], + "angle": 0, + "content": "Table 1. The ablation experiments of the sparsely-gated Tra-MoE. All baselines are the track transformer of ATM." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.433, + 0.486, + 0.69 + ], + "angle": 0, + "content": "dataset \\( \\mathcal{D}_{\\text{ood}} \\), utilizing 20 action-free videos per task; and choose 40 tasks from LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, and LIBERO-Long to build dataset \\( \\mathcal{D}_{\\text{in}} \\), using 10 action-labeled demonstrations per task. In summary, for training Tra-MoE, our main ablation experiments use a total of 2, 200 action-free videos, with an out-of-domain to in-domain data ratio of \\( 9:2 \\). Additionally, on this basis, we also further select 92 tasks from the RLbench [26] (using CoppeliaSim rather than MuJoCo as physics engine) simulation environment and generate 5 videos for each task, totaling 2, 660 videos in conjunction with the LIBERO data, to train Tra-MoE. The different physics engines and dynamics properties further increase the data distribution discrepancy in \\( \\mathcal{D}_{\\text{ood}} \\). Finally, as for the trajectory-guided policy, we train a separate multi-task policy for each category. Each policy undergoes behavioral cloning training with 100 action-labeled demonstrations from 10 tasks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.695, + 0.483, + 0.831 + ], + "angle": 0, + "content": "Implement details. Building on ATM [63], we develop our sparsely-gated Tra-MoE and adaptive policy conditioning technique. Specifically, when using only LIBERO data, we replace the 1th, 5th, and 8th layers of the track transformer with MoE blocks. When further incorporating RLbench data, we additionally replace the 2th and 7th layers. Unless otherwise specified, we set the number of experts to 4 by default and use a top-1 gating strategy for token choice to ensure the FLOPs per token remain constant." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.836, + 0.483, + 0.882 + ], + "angle": 0, + "content": "Experiment results and analysis. We present empirical simulation results to address the aforementioned research questions and make the following findings:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.886, + 0.483, + 0.9 + ], + "angle": 0, + "content": "Result 1: In Tab. 1a, we examine the effect of the router" + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.431, + 0.895, + 0.488 + ], + "angle": 0, + "content": "
SpatialGoalObjectLongAvg.
ATM51.042.055.519.041.9
Tra-MoE54.062.061.041.054.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.498, + 0.905, + 0.527 + ], + "angle": 0, + "content": "Table 2. The comparison results when training trajectory model using RLbench video data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.672 + ], + "angle": 0, + "content": "z-loss weight \\(\\lambda_z\\). The results indicate that a higher \\(\\lambda_z\\) can effectively enhance performance, with an improvement of 4.5 (61.4 vs. 56.9) at \\(\\lambda_z = 1 \\cdot 10^{-4}\\). In Tab. 1b, we study the impact of the load-balancing loss weight \\(\\lambda_{lo - ba}\\). The results show a significant performance decline with larger values of \\(\\lambda_{lo - ba}\\), with only a slight improvement (\\(56.7 \\rightarrow 57.3\\)) at \\(\\lambda_{lo - ba} = 1 \\cdot 10^{-7}\\). In Tab. 1c, we verify the effectiveness of adding noise to the gating logits, which results in a performance decrease of by 1.1 (\\(56.9 \\rightarrow 55.8\\))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.909, + 0.884 + ], + "angle": 0, + "content": "Finding 1: Based on the above results, we find that using router z-loss to penalize large logits entering the gating network can effectively improve training stability and lead to performance gains. Meanwhile, the load-balancing loss used to balance the activation frequency among experts and the addition of noise to enhance expert allocation exploration cannot bring effective performance improvements. An intuitive explanation is that when the data distribution is uneven, the load-balancing loss might force experts to set shared parameters on data with a large domain gap, which could lead to mutual counteraction of learning gradients. This undermines the advantage of parameter specialization in sparsely-gated MoE. Similarly, adding noise is also detrimental to learning parameter specialization among experts." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Result 2: We comprehensively investigate the impact of in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6965" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.089, + 0.462, + 0.161 + ], + "angle": 0, + "content": "
SpatialGoalObjectLongAvg.
ATM62.581.073.528.561.4
+ DM69.058.085.533.561.5
+ AM69.577.088.030.566.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.172, + 0.483, + 0.214 + ], + "angle": 0, + "content": "Table 3. The ablation experiments of the adaptive policy conditioning technique. DM (RT-Trajectory style) and AM respectively represent hand-drawn mask and adaptive mask." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.221, + 0.485, + 0.537 + ], + "angle": 0, + "content": "producing sparsely-gated MoE and large-scale out-of-domain data on scaling up trajectory models in Tab. 1f. Firstly, when training only with in-domain data, Tra-baseline achieves an average success rate of 57.6, while Tra-MoE only achieves an average success rate of 51.8. Secondly, when adding out-of-domain data to jointly train Tra-baseline, the average success rate drops to \\(52.0\\) (\\(57.6 \\rightarrow 52.0\\)), with particularly significant performance declines in LIBERO-Spatial and LIBERO-Object. However, when adding out-of-domain data to jointly train Tra-MoE, the average success rate increases to \\(61.4\\) (\\(51.8 \\rightarrow 61.4\\)). Additionally, in Tab. 1e, we also explore the impact of the number of experts in sparsely-gated MoE. The results show that using just two experts can lead to an increase of 4.0 (52.0 vs. 56.0) in the average success rate. With additional experts, performance generally trends upward, albeit with some fluctuations. Finally, we report the comparison results of training trajectory model by further using RLbench data in Tab. 2. The results indicate that although the introduction of RLbench data led to a performance drop in LIBERO, our Tra-MoE still outperform the baseline by 12.6 (54.5 vs. 41.9)." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.539, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Finding 2: Based on the above results, we have the following findings and analysis: (i) When the data scale is small, naively expanding the model capacity through sparsely-gated MoE does not lead to performance improvements. A explanation is that the larger model may lead to a overfitting problem when training on small-scale data. Moreover, small-scale data often leads to some experts not being sufficiently trained, thereby affecting performance. (ii) Naively introducing out-of-domain data for joint training often leads to performance degradation in the target domain. However, simultaneously incorporating out-of-domain data and expanding models with sparsely-gated MoE may improve performance or alleviate performance degradation. This phenomenon suggests that although out-of-domain data generally enhances model generalization, it could hinder task-specific performance in the target domain due to insufficient in-domain data learning. In contrast, expanding the model with out-of-domain data using sparsely-gated MoE achieves a more optimal balance between parameter cooperation and specialization across diverse data and even different tokens within the same data input. This design allows these large-scale data to jointly train most of parameters, ensuring the capture of general patterns for mutual cooperation. (iii) Even with a few experts, Tra-MoE exhibits significant scaling up" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.152 + ], + "angle": 0, + "content": "capabilities. Additionally, further increasing the number of experts generally leads to performance improvement. However, this improvement is not guaranteed due to factors like data-expert complexity matching and training stability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.155, + 0.909, + 0.274 + ], + "angle": 0, + "content": "Result 3: In Tab. 1d, we also investigate the performance comparison between the dense Tra-baseline and the sparse Tra-MoE with the same number of parameters. Specifically, we expand Tra-baseline to match the model capacity of Tra-MoE by increasing its model width or depth separately. These expanded versions achieve average success rates of 48.4 and 52.5 respectively, which are significantly lower than Tra-MoE's 61.4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.277, + 0.909, + 0.399 + ], + "angle": 0, + "content": "Finding 3: Based on these results, we are surprised to find that simply expanding the dense Tra-baseline in depth and width can not effectively improve performance. This highlights the importance of the sparsely-gated MoE architecture when training with large-scale out-of-domain data, as it ensures dynamic activation of different experts based on different data input and different tokens within the same data input, significantly reducing optimization conflict problems." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.401, + 0.909, + 0.702 + ], + "angle": 0, + "content": "Result 4: In Tab. 3, we investigate the effectiveness of the adaptive policy conditioning. Specifically, building upon ATM, we first introduce an additional hand-drawn mask modality before image encoding. In this mask, the first half of the trajectory is set to 128, while the second half is set to 255. We call it 'ATM + hand-drawn mask' and this is similar to RT-Trajectory [21]. Next, we replace the hand-drawn mask with an adaptive mask, where each point on the trajectory is instantiated as a learnable embedding. We call it 'ATM + adaptive mask'. The results show that 'ATM + hand-drawn mask' achieves only a 0.1 improvement compared to the baseline. However, except for a significant performance decrease in LIBERO-Goal, it achieves performance improvements of 6.5, 12.0, and 5.0 in LIBERO-Spatial, LIBERO-Object, and LIBERO-Long respectively. In addition, the 'ATM + adaptive mask' achieves a performance improvement of 4.9 (66.3 vs. 61.4) compared to the baseline. It further improves performance across all suites compared to 'ATM + hand-drawn mask', especially achieving a 19.0 performance gains in LIBERO-Goal (77.0 vs. 58.0)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.911, + 0.901 + ], + "angle": 0, + "content": "Finding 4: Based on these results, we find that explicitly mapping 2D trajectories to visual observations significantly improves the performance of LIBERO-Spatial, LIBERO-Object, and LIBERO-Long, but causes a notable decrease in LIBERO-Goal. By instantiating the points in the 2D trajectory as learnable embeddings, we not only further enhances the performance of LIBERO-Spatial, LIBERO-Object, and LIBERO-Long, but also significantly improves LIBERO-Goal. We analyze that the decline in LIBERO-Goal is mainly due to overlapping front and back parts of some trajectories in the 2D images, leading to mapping errors. However, our adaptive trajectory representation and conditioning technique effectively mitigates this issue. In summary, these" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6966" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.127, + 0.09, + 0.872, + 0.176 + ], + "angle": 0, + "content": "
Human DataMoEAdaptive Trajectory MaskPourPushPick and PassTissueFoldAverage
40.045.050.030.025.038.0
✓45.035.050.025.035.038.0
✓✓40.050.060.035.045.046.0
✓✓✓60.070.065.035.050.056.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.12, + 0.187, + 0.874, + 0.2 + ], + "angle": 0, + "content": "Table 4. The ablation of ood human data, MoE architecture, and adaptive policy conditioning technique on our real-world tasks." + }, + { + "type": "image", + "bbox": [ + 0.156, + 0.205, + 0.835, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.127, + 0.413, + 0.868, + 0.427 + ], + "angle": 0, + "content": "Figure 4. Left: The real-world experiments hardware platform setup. Right: The real-world tasks evaluation demonstrations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.436, + 0.484, + 0.482 + ], + "angle": 0, + "content": "experiments collectively demonstrate the effectiveness of explicitly aligning 2D trajectories with images in space as well as applying the adaptive conditioning technique." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.496, + 0.32, + 0.512 + ], + "angle": 0, + "content": "4.2. Real-World Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.484, + 0.701 + ], + "angle": 0, + "content": "Experiment setup. As shown in Fig 4, we utilize a dual-arm robot [35] to conduct real-world experiments. We evaluate 5 tasks, which include two single-arm tasks: pouring water and pushing a vegetable to the side of the cutting board, as well as three dual-arm tasks: pulling out tissues, folding towels, and picking up and passing a holder. Specifically, for each task, we collect 50 action-labeled robot demonstrations. Concurrently, we also collect 50 videos of human (cross-embediments) performing for each task. Compared to the robot data, the human video encompasses more environment variations and distractor objects. Please see the appendix for more detailed evaluation details and real-world setup." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.484, + 0.854 + ], + "angle": 0, + "content": "Implement details. Following our simulation experiments, we utilize two camera views and uniformly resize their resolution to \\(128 \\times 128\\) for trajectory model and policy training. We train our real-world Tra-MoE on borh human video and robot video data, with a separate policy trained for each task. Therefore, our real-world Tra-MoE integrates data across different embodiments (human, single-arm robot, and dual-arm robot). They have significantly different visual appearances and dynamics properties. Finally, we also present some real-world task evaluation demonstrations in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Experiment results and analysis. To further strengthen our conclusions that sparsely-gated MoE can better learn trajectory model from out-of-domain data and our adaptive policy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.436, + 0.908, + 0.617 + ], + "angle": 0, + "content": "condition technique could lead to more flexible guidance for policy learning, we conduct real-world ablation experiments, as shown in Tab. 4. The results indicate that simply introducing human video data cannot lead to performance improvement. However, when expanding the model by MoE architecture, the performance increases from \\(38\\%\\) to \\(46\\%\\). We speculate this may be due to the significant differences in appearance and dynamics between human and robot, with the former typically moving much faster and more flexibly. Finally, when further incorporating an adaptive trajectory mask for policy conditioning, the performance is further increased from \\(46\\%\\) to \\(56\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.62, + 0.634, + 0.635 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this work, we have introduced Tra-MoE, a novel framework that utilizes a sparsely-gated MoE architecture with Top-1 gating strategy to learn trajectory model from multiple domains. Tra-MoE achieves superior parameter cooperation and specialization, enabling more effective utilization of large-scale out-of-domain data while maintaining constant FLOPs per token. We train our Tra-MoE and conduct a comprehensive empirical study, demonstrating that TraMoE consistently exhibits superior performance compared to the dense baseline, even when the latter is scaled to match Tra-MoE's parameter count. Additionally, we propose an adaptive policy conditioning technique for trajectory-guided policy, ensuring adaptive 2D trajectory representation and explicit spatial alignment with observations, thereby achieving superior performance. Extensive experiments in both simulated and real-world environments validate our findings and demonstrate the effectiveness of our proposed methods." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6967" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.092, + 0.232, + 0.107 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.107, + 0.486, + 0.184 + ], + "angle": 0, + "content": "This work is supported by the National Key R&D Program of China (No. 2022ZD0160900), Jiangsu Frontier Technology Research and Development Program (No. BF2024076), and the Collaborative Innovation Center of Novel Software Technology and Industrialization." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.193, + 0.188, + 0.209 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.218, + 0.484, + 0.274 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.276, + 0.484, + 0.343 + ], + "angle": 0, + "content": "[2] OpenAI: Marcin Andrychowicz, Bowen Baker, Maciek Chogiej, Rafal Jozefowicz, Bob McGrew, Jakub Pachocki, Arthur Petron, Matthias Plappert, Glenn Powell, Alex Ray, et al. Learning dexterous in-hand manipulation. *IJRR*, 39(1): 3-20, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.346, + 0.484, + 0.4 + ], + "angle": 0, + "content": "[3] Shikhar Bahl, Russell Mendonca, Lili Chen, Unnat Jain, and Deepak Pathak. Affordances from human videos as a versatile representation for robotics. In CVPR, pages 13778-13790, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.403, + 0.484, + 0.458 + ], + "angle": 0, + "content": "[4] Homanga Bharadhwaj, Abhinav Gupta, Vikash Kumar, and Shubham Tulsiani. Towards generalizable zero-shot manipulation via translating human interaction plans. In ICRA, pages 6904-6911. IEEE, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.461, + 0.484, + 0.516 + ], + "angle": 0, + "content": "[5] Homanga Bharadhwaj, Roozbeh Mottaghi, Abhinav Gupta, and Shubham Tulsiani. Track2act: Predicting point tracks from internet videos enables diverse zero-shot robot manipulation. arXiv preprint arXiv:2405.01527, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.518, + 0.484, + 0.586 + ], + "angle": 0, + "content": "[6] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.589, + 0.484, + 0.657 + ], + "angle": 0, + "content": "[7] Konstantinos Bousmalis, Giulia Vezzani, Dushyant Rao, Coline Manon Devin, Alex X Lee, Maria Bauza Villalonga, Todor Davchev, Yuxiang Zhou, Agrim Gupta, Akhil Raju, et al. Robocat: A self-improving generalist agent for robotic manipulation. TMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.661, + 0.484, + 0.729 + ], + "angle": 0, + "content": "[8] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.732, + 0.484, + 0.8 + ], + "angle": 0, + "content": "[9] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.484, + 0.83 + ], + "angle": 0, + "content": "[10] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[11] Qingwen Bu, Jisong Cai, Li Chen, Xiuqi Cui, Yan Ding, Siyuan Feng, Shenyuan Gao, Xindong He, Xu Huang, Shu Jiang, et al. Agibot world colosseo: A large-scale manipulation platform for scalable and intelligent embodied systems. arXiv preprint arXiv:2503.06669, 2025. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.218, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.135 + ], + "angle": 0, + "content": "[12] Weilin Cai, Juyong Jiang, Fan Wang, Jing Tang, Sunghun Kim, and Jiayi Huang. A survey on mixture of experts. arXiv preprint arXiv:2407.06204, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[13] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, pages 9640-9649, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.907, + 0.232 + ], + "angle": 0, + "content": "[14] Zitian Chen, Yikang Shen, Mingyu Ding, Zhenfang Chen, Hengshuang Zhao, Erik G Learned-Miller, and Chuang Gan. Mod-squad: Designing mixtures of experts as modular multitask learners. In CVPR, pages 11828-11837, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.906, + 0.273 + ], + "angle": 0, + "content": "[15] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.907, + 0.327 + ], + "angle": 0, + "content": "[16] Ria Doshi, Homer Rich Walke, Oier Mees, Sudeep Dasari, and Sergey Levine. Scaling cross-embodied learning: One policy for manipulation, navigation, locomotion and aviation. In CoRL, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.907, + 0.398 + ], + "angle": 0, + "content": "[17] Yilun Du, Sherry Yang, Pete Florence, Fei Xia, Ayzaan Wahid, Pierre Sermanet, Tianhe Yu, Pieter Abbeel, Joshua B Tenenbaum, Leslie Pack Kaelbling, et al. Video language planning. In The Twelfth International Conference on Learning Representations. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.401, + 0.907, + 0.453 + ], + "angle": 0, + "content": "[18] Yilun Du, Sherry Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Josh Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation. NeurIPS, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.456, + 0.907, + 0.523 + ], + "angle": 0, + "content": "[19] Hao-Shu Fang, Hongjie Fang, Zhenyu Tang, Jirong Liu, Junbo Wang, Haoyi Zhu, and Cewu Lu. Rh20t: A robotic dataset for learning diverse skills in one-shot. In RSS 2023 Workshop on Learning for Task and Motion Planning, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.906, + 0.566 + ], + "angle": 0, + "content": "[20] Chongkai Gao, Haozhuo Zhang, Zhixuan Xu, Cai Zhehao, and Lin Shao. Flip: Flow-centric generative planning as general-purpose manipulation world model. In ICLR. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.907, + 0.635 + ], + "angle": 0, + "content": "[21] Jiayuan Gu, Sean Kirmani, Paul Wohlhart, Yao Lu, Montserrat Gonzalez Arenas, Kanishka Rao, Wenhao Yu, Chuyuan Fu, Keerthana Gopalakrishnan, Zhuo Xu, et al. Rt-trajectory: Robotic task generalization via hindsight trajectory sketches. In ICLR. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.907, + 0.691 + ], + "angle": 0, + "content": "[22] Pierre-Louis Guhur, Shizhe Chen, Ricardo Garcia Pinel, Makarand Tapaswi, Ivan Laptev, and Cordelia Schmid. Instruction-driven history-aware policies for robotic manipulations. In CORL, pages 175-187. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.693, + 0.906, + 0.734 + ], + "angle": 0, + "content": "[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.907, + 0.789 + ], + "angle": 0, + "content": "[24] Changho Hwang, Wei Cui, Yifan Xiong, Ziyue Yang, Ze Liu, Han Hu, Zilong Wang, Rafael Salas, Jithin Jose, Prabhat Ram, et al. Tutel: Adaptive mixture-of-experts at scale. Proceedings of Machine Learning and Systems, 5:269-287, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.907, + 0.858 + ], + "angle": 0, + "content": "[25] Vidhi Jain, Maria Attarian, Nikhil J Joshi, Ayzaan Wahid, Danny Driess, Quan Vuong, Pannag R Sanketi, Pierre Sermanet, Stefan Welker, Christine Chan, et al. Vid2robot: End-to-end video-conditioned policy learning with cross-attention transformers. arXiv preprint arXiv:2403.12943, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[26] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J Davison. Rlbench: The robot learning benchmark & learning environment. RAL, 5(2):3019-3026, 2020. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6968" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[27] Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kappler, Frederik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. Bc-z: Zero-shot task generalization with robotic imitation learning. In CoRL, pages 991-1002. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.217 + ], + "angle": 0, + "content": "[28] Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. Mixtral of experts. arXiv preprint arXiv:2401.04088, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.484, + 0.286 + ], + "angle": 0, + "content": "[29] Yunfan Jiang, Agrim Gupta, Zichen Zhang, Guanzhi Wang, Yongqiang Dou, Yanjun Chen, Li Fei-Fei, Anima Anandkumar, Yuke Zhu, and Linxi Fan. Vima: General robot manipulation with multimodal prompts. arXiv preprint arXiv:2210.03094, 2(3):6, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.484, + 0.355 + ], + "angle": 0, + "content": "[30] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.484, + 0.412 + ], + "angle": 0, + "content": "[31] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv preprint arXiv:2307.07635, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.484, + 0.493 + ], + "angle": 0, + "content": "[32] Alexander Khazatsky, Karl Pertsch, Suraj Nair, Ashwin Balakrishna, Sudeep Dasari, Siddharth Karamcheti, Soroush Nasiriany, Mohan Kumar Srirama, Lawrence Yunliang Chen, Kirsty Ellis, et al. Droid: A large-scale in-the-wild robot manipulation dataset. arXiv preprint arXiv:2403.12945, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.484, + 0.564 + ], + "angle": 0, + "content": "[33] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.484, + 0.607 + ], + "angle": 0, + "content": "[34] Po-Chen Ko, Jiayuan Mao, Yilun Du, Shao-Hua Sun, and Joshua B Tenenbaum. Learning to act from actionless videos through dense correspondences. In ICLR. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.484, + 0.649 + ], + "angle": 0, + "content": "[35] Alexander Koch. Low-cost robot arm. https://github.com/AlexanderKoch-Koch/low_costrobot, 2024. GitHub repository. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.484, + 0.718 + ], + "angle": 0, + "content": "[36] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan First, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. arXiv preprint arXiv:2006.16668, 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.484, + 0.761 + ], + "angle": 0, + "content": "[37] Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuomotor policies. JMLR, 17(39):1-40, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[38] Xinghang Li, Minghuan Liu, Hanbo Zhang, Cunjun Yu, Jie Xu, Hongtao Wu, Chilam Cheang, Ya Jing, Weinan Zhang, Huaping Liu, et al. Vision-language foundation models as effective robot imitators. In ICLR. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.484, + 0.872 + ], + "angle": 0, + "content": "[39] Bin Lin, Zhenyu Tang, Yang Ye, Jiaxi Cui, Bin Zhu, Peng Jin, Junwu Zhang, Munan Ning, and Li Yuan. Moe-llava: Mixture of experts for large vision-language models. arXiv preprint arXiv:2401.15947, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[40] Aixin Liu, Bei Feng, Bin Wang, Bingxuan Wang, Bo Liu, Chenggang Zhao, Chengqi Dengr, Chong Ruan, Damai Dai," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Daya Guo, et al. Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model. arXiv preprint arXiv:2405.04434, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.177 + ], + "angle": 0, + "content": "[41] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning. NeurIPS, 36, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.907, + 0.246 + ], + "angle": 0, + "content": "[42] Fangchen Liu, Kuan Fang, Pieter Abbeel, and Sergey Levine. Moka: Open-vocabulary robotic manipulation through mark-based visual prompting. In First Workshop on Vision-Language Models for Navigation and Manipulation at ICRA 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.907, + 0.303 + ], + "angle": 0, + "content": "[43] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. In ICLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.306, + 0.907, + 0.36 + ], + "angle": 0, + "content": "[44] Corey Lynch, Mohi Khansari, Ted Xiao, Vikash Kumar, Jonathan Tompson, Sergey Levine, and Pierre Sermanet. Learning latent plans from play. In CoRL, pages 1113-1132. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.907, + 0.417 + ], + "angle": 0, + "content": "[45] Corey Lynch, Ayzaan Wahid, Jonathan Tompson, Tianli Ding, James Betker, Robert Baruch, Travis Armstrong, and Pete Florence. Interactive language: Talking to robots in real time. RAL, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.907, + 0.488 + ], + "angle": 0, + "content": "[46] Soroush Nasiriany, Sean Kirmani, Tianli Ding, Laura Smith, Yuke Zhu, Danny Driess, Dorsa Sadigh, and Ted Xiao. Rt-affordance: Affordances are versatile intermediate representations for robot manipulation. arXiv preprint arXiv:2411.02704, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.907, + 0.559 + ], + "angle": 0, + "content": "[47] Abhishek Padalkar, Acorn Pooley, Ajinkya Jain, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anikait Singh, Anthony Brohan, et al. Open x-embediment: Robotic learning datasets and rt-x models. arXiv preprint arXiv:2310.08864, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.907, + 0.602 + ], + "angle": 0, + "content": "[48] Joan Puigcerver, Carlos Riquelme, Basil Mustafa, and Neil Houlsby. From sparse to soft mixtures of experts. arXiv preprint arXiv:2308.00951, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.907, + 0.673 + ], + "angle": 0, + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICLR, pages 8748-8763. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.907, + 0.731 + ], + "angle": 0, + "content": "[50] Rouhollah Rahmatizadeh, Pooya Abolghasemi, Ladislau Böloni, and Sergey Levine. Vision-based multi-task manipulation for inexpensive robots using end-to-end learning from demonstration. In ICRA, pages 3758-3765. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.907, + 0.8 + ], + "angle": 0, + "content": "[51] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. arXiv preprint arXiv:2205.06175, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.907, + 0.845 + ], + "angle": 0, + "content": "[52] Younggyo Seo, Kimin Lee, Stephen L James, and Pieter Abbeel. Reinforcement learning with action-free pre-training from videos. In ICLR, pages 19561-19579. PMLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[53] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6969" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.135 + ], + "angle": 0, + "content": "[54] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In CoRL, pages 894-906. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.485, + 0.205 + ], + "angle": 0, + "content": "[55] Austin Stone, Ted Xiao, Yao Lu, Keerthana Gopalakrishnan, Kuang-Huei Lee, Quan Vuong, Paul Wohlhart, Brianna Zitkovich, Fei Xia, Chelsea Finn, et al. Open-world object manipulation using pre-trained vision-language models. In CoRL, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.262 + ], + "angle": 0, + "content": "[56] Priya Sundaresan, Quan Vuong, Jiayuan Gu, Peng Xu, Ted Xiao, Sean Kirmani, Tianhe Yu, Michael Stark, Ajinkya Jain, Karol Hausman, et al. Rt-sketch: Goal-conditioned imitation learning from hand-drawn sketches. 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.484, + 0.319 + ], + "angle": 0, + "content": "[57] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.484, + 0.362 + ], + "angle": 0, + "content": "[58] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.484, + 0.419 + ], + "angle": 0, + "content": "[59] Mel Vecerik, Carl Doersch, Yi Yang, Todor Davchev, Yusuf Aytar, Guangyao Zhou, Raia Hadsell, Lourdes Agapito, and Jon Scholz. Robotap: Tracking arbitrary points for few-shot visual imitation. In ICRA, pages 5397-5403. IEEE, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.484, + 0.488 + ], + "angle": 0, + "content": "[60] Homer Rich Walke, Kevin Black, Tony Z Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen-Estruch, Andre Wang He, Vivek Myers, Moo Jin Kim, Max Du, et al. Bridgedata v2: A dataset for robot learning at scale. In CoRL, pages 1723-1736. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.484, + 0.546 + ], + "angle": 0, + "content": "[61] Chen Wang, Linxi Fan, Jiankai Sun, Ruohan Zhang, Li Fei-Fei, Danfei Xu, Yuke Zhu, and Anima Anandkumar. Mimicplay: Long-horizon imitation learning by watching human play. In CoRL. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.484, + 0.589 + ], + "angle": 0, + "content": "[62] Lirui Wang, Xinlei Chen, Jialiang Zhao, and Kaiming He. Scaling proprioceptive-visual learning with heterogeneous pre-trained transformers. CoRR, abs/2409.20537, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.484, + 0.645 + ], + "angle": 0, + "content": "[63] Chuan Wen, Xingyu Lin, John So, Kai Chen, Qi Dou, Yang Gao, and Pieter Abbeel. Any-point trajectory modeling for policy learning. arXiv preprint arXiv:2401.00025, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.648, + 0.484, + 0.716 + ], + "angle": 0, + "content": "[64] Kun Wu, Chengkai Hou, Jiaming Liu, Zhengping Che, Xiaozhu Ju, Zhuqin Yang, Meng Li, Yinuo Zhao, Zhiyuan Xu, Guang Yang, et al. Robomind: Benchmark on multi-embediment intelligence normative data for robot manipulation. arXiv preprint arXiv:2412.13877, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[65] Mengda Xu, Zhenjia Xu, Yinghao Xu, Cheng Chi, Gordon Wetzstein, Manuela Veloso, and Shuran Song. Flow as the cross-domain manipulation interface. arXiv preprint arXiv:2407.15208, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[66] Jiange Yang, Wenhui Tan, Chuhao Jin, Bei Liu, Jianlong Fu, Ruihua Song, and Limin Wang. Pave the way to grasp anything: Transferring foundation models for universal pickplace robots. arXiv preprint arXiv:2306.05716, 1(2), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[67] Jiange Yang, Bei Liu, Jianlong Fu, Bocheng Pan, Gangshan Wu, and Limin Wang. Spatiotemporal predictive pre-training for robotic motor control. arXiv preprint arXiv:2403.05304, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.908, + 0.174 + ], + "angle": 0, + "content": "[68] Jiange Yang, Wenhui Tan, Chuhao Jin, Keling Yao, Bei Liu, Jianlong Fu, Ruihua Song, Gangshan Wu, and Limin Wang. Transferring foundation models for generalizable robotic manipulation. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), pages 1999-2010, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.178, + 0.907, + 0.233 + ], + "angle": 0, + "content": "[69] Seonghyeon Ye, Joel Jang, Byeongguk Jeon, Sejune Joo, Jianwei Yang, Baolin Peng, Ajay Mandlekar, Reuben Tan, Yu-Wei Chao, Bill Yuchen Lin, et al. Latent action pretraining from videos. arXiv preprint arXiv:2410.11758, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.234, + 0.907, + 0.275 + ], + "angle": 0, + "content": "[70] Chengbo Yuan, Chuan Wen, Tong Zhang, and Yang Gao. General flow as foundation affordance for scalable robot learning. arXiv preprint arXiv:2401.11439, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.277, + 0.907, + 0.331 + ], + "angle": 0, + "content": "[71] Jinguo Zhu, Xizhou Zhu, Wenhai Wang, Xiaohua Wang, Hongsheng Li, Xiaogang Wang, and Jifeng Dai. Uniperceiver-moe: Learning sparse generalist models with conditional moes. NeurIPS, 35:2664-2678, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.333, + 0.907, + 0.389 + ], + "angle": 0, + "content": "[72] Xizhou Zhu, Jinguo Zhu, Hao Li, Xiaoshi Wu, Hongsheng Li, Xiaohua Wang, and Jifeng Dai. Uni-perceiver: Pre-training unified architecture for generic perception for zero-shot and few-shot tasks. In CVPR, pages 16804-16815, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.39, + 0.907, + 0.432 + ], + "angle": 0, + "content": "[73] Yifeng Zhu, Zhenyu Jiang, Peter Stone, and Yuke Zhu. Learning generalizable manipulation policies with object-centric 3d representations. In CoRL. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.433, + 0.908, + 0.488 + ], + "angle": 0, + "content": "[74] Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam Shazeer, and William Fedus. St-moe: Designing stable and transferable sparse expert models. arXiv preprint arXiv:2202.08906, 2022. 2, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.908, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "6970" + } + ] +] \ No newline at end of file diff --git a/2025/Tra-MoE_ Learning Trajectory Prediction Model from Multiple Domains for Adaptive Policy Conditioning/d41e3854-0a29-4aee-85fc-0fcf3b667c89_origin.pdf b/2025/Tra-MoE_ Learning Trajectory Prediction Model from Multiple Domains for Adaptive Policy Conditioning/d41e3854-0a29-4aee-85fc-0fcf3b667c89_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b84216de27186e29f5c4ebdbd8b9f3cddad07df5 --- /dev/null +++ b/2025/Tra-MoE_ Learning Trajectory Prediction Model from Multiple Domains for Adaptive Policy Conditioning/d41e3854-0a29-4aee-85fc-0fcf3b667c89_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:df984963d6de2c2be9fab1a77af5ad87c7c8ad512eacc9da9f441d4737f6bada +size 1784290 diff --git a/2025/Tra-MoE_ Learning Trajectory Prediction Model from Multiple Domains for Adaptive Policy Conditioning/full.md b/2025/Tra-MoE_ Learning Trajectory Prediction Model from Multiple Domains for Adaptive Policy Conditioning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4a09e56086d3639b2c1b3782a1c8a8ea3812c163 --- /dev/null +++ b/2025/Tra-MoE_ Learning Trajectory Prediction Model from Multiple Domains for Adaptive Policy Conditioning/full.md @@ -0,0 +1,323 @@ +# Tra-MoE: Learning Trajectory Prediction Model from Multiple Domains for Adaptive Policy Conditioning + +Jiange Yang $^{1,2}$ Haoyi Zhu $^{2,3}$ Yating Wang $^{2,4}$ Gangshan Wu $^{1}$ Tong He $^{2}$ Limin Wang $^{1,2*}$ + +$^{1}$ State Key Laboratory for Novel Software Technology, Nanjing University, China $^{2}$ Shanghai AI Laboratory, $^{3}$ University of Science and Technology of China, $^{4}$ Tongji University + +jiangyang.jgy@gmail.com, {gswu,lmwang}@nju.edu.cn + +{zhuhaoyi, wangyating, hetong}@pjlab.org.cn + +https://github.com/MCG-NJU/Tra-MoE + +# Abstract + +Learning from multiple domains is a primary factor that influences the generalization of a single unified robot system. In this paper, we aim to learn the trajectory prediction model by using broad out-of-domain data to improve its performance and generalization ability. Trajectory model is designed to predict any-point trajectories in the current frame given an instruction and can provide detailed control guidance for robotic policy learning. To handle the diverse out-of-domain data distribution, we propose a sparsely-gated MoE (Top-1 gating strategy) architecture for trajectory model, coined as Tra-MoE. The sparse activation design enables good balance between parameter cooperation and specialization, effectively benefiting from large-scale out-of-domain data while maintaining constant FLOPs per token. In addition, we further introduce an adaptive policy conditioning technique by learning 2D mask representations for predicted trajectories, which is explicitly aligned with image observations to guide action prediction more flexibly. We perform extensive experiments on both simulation and real-world scenarios to verify the effectiveness of Tra-MoE and adaptive policy conditioning technique. We also conduct a comprehensive empirical study to train Tra-MoE, demonstrating that our Tra-MoE consistently exhibits superior performance compared to the dense baseline model, even when the latter is scaled to match Tra-MoE's parameter count. + +# 1. Introduction + +In computer vision and natural language processing, significant progress has been made through learning models from multiple domains data. For example, many works use board data to perform self-supervised pre-training, and then adapt to diverse downstream tasks through zero-shot [1, 10, 49] or fine-tuning [13, 15, 23, 58]. Additionally, some research ef + +![](images/b20601886bd5d635928fdf1cbebeaae57f139d3d6ba8796761e8f9210cfab061.jpg) +Figure 1. (a) Training with small-scale, in-domain data. (b) Joint training with in-domain data and large-scale, out-of-domain data. + +forts [43, 51, 71, 72] focus on learning a generalist model capable of unifying various tasks. In contrast, in robot learning, due to data scarcity and homogeneity, some groundbreaking methods [2, 37] resort to training only using in-domain data. Recently, some works [3, 4, 52, 61, 69] have shifted to learn from action-free human video data. A particularly scalable and promising paradigm [5, 20, 59, 63, 65] involves learning a trajectory prediction model from action-free video data, and then using small-scale action-labeled demonstrations to learn trajectory-guided policies, thereby achieving greater sample efficiency and better generalization. + +Despite the potential benefits, it still remains underexplored how to effectively leverage broad out-of-domain video data for jointly training trajectory models. In particular, the impact of incorporating data encompassing diverse environments, objects, skills and embodiments has not fully investigated for in-domain tasks. For instance, ATM [63], a notable work in this field, mainly relies on in-domain data (same tasks and environments) for training both trajectory + +![](images/1116d206222c344b3495caf32192dc759e91b4b7ea155d6cd6e4cf158685afb2.jpg) +Figure 2. Training trajectory prediction model from multiple domains. (a) The visualization of Dataset $\mathcal{D}_{\text {ood }}$ and $\mathcal{D}_{\text {in }}$ . The former may contain additional environments, objects, skills and embodiments. (b) Our pipeline: first co-training trajectory prediction model and then adapting it for downstream policy learning. + +![](images/dbc45401311683cf540e6205e3be6a9a5b51e719c61443d62b51f8072d7ecab6.jpg) + +models and policies. We argue that learning from multiple domains is the primary factor to build a powerful trajectory model for policy guidance. Therefore, we propose to train trajectory model under the setup of jointly learning from large-scale out-of-domain and small-scale in-domain data, as shown in Fig. 2(a). However, directly using large-scale out-of-domain video data for hybrid pre-training involves two critical challenges: (i) Joint training on data from different environments, objects, skills, and embodiments will add learning difficulty and might lead to optimization conflicts. It is unclear how to effectively fuse the complementary information of multiple domains data within a single unified model. Naively expanding the out-of-domain training data does not necessarily lead to improved performance on in-domain tasks, as shown in Fig. 1, which results in a 5.6 $(57.6\to 52.0)$ performance decrease on in-domain tasks. (ii) How to ensure both high performance and computational efficiency during trajectory model scaling up process. + +To address these challenges, we design a new sparsely-gated Mixture-of-Expert (MoE) architecture [53] to scale up our trajectory model, coined as Tra-MoE. Specifically, we integrate several MoE blocks to replace the original transformer blocks. This design enables efficient joint training of most parameters on the broad out-of-domain data, capturing the complementary patterns for mutual cooperation and the shared knowledge for better generalization. Meanwhile, for different data and different tokens within the same data, our Tra-MoE naturally forms a specialization when activating different experts. To maintain high computational efficiency while scaling up model capacity, we implement a top-1 gating strategy for token-choice, ensuring constant FLOPs per token. In sparse MoE, some auxiliary losses are commonly used to enhance performance, including the router z-loss [74] for improving training stability and the load-balancing loss [36] for balancing expert activations. We further conduct a comprehensive empirical study on scaling up trajectory models and find that the touter z-loss improves + +performance while the load-balancing loss tends to cause performance degradation. Furthermore, we observe that when expanding pre-training data from small-scale in-domain to large-scale in-domain and out-of-domain data, employing MoE (Tra-MoE) can lead to obvious performance improvement on in-domain tasks, whereas the dense counterpart (Tra-baseline) suffers from performance drop, as illustrated in Fig. 1. Notably, even when expanding the dense Tra-baseline's parameter to match that of Tra-MoE, performance improvements remain elusive. These findings collectively demonstrate the effectiveness of our sparsely-gated MoE architecture, attributing its capability of balancing parameter cooperation and specialization when expanding training with out-of-domain data. + +Furthermore, given a pre-trained trajectory model, how to effectively condition the policy module using the 2D trajectories still remain an unresolved challenge. We further propose an adaptive policy conditioning technique for trajectory-guided policy. This enables better spatial alignment between 2D trajectories and images, while achieving more flexible trajectory representation and guidance for the policy. Finally, we also conduct extensive real-world experiments to further strengthen our conclusions. In summary, our main contributions can be summarized as follows: + +- We propose a sparsely-gated MoE architecture to learn trajectory models, called as Tra-MoE, with large-scale, out-of-domain, and action-free video data, coupled with a comprehensive empirical study on its training techniques. +- Extensive simulation and real-world environments demonstrate that our Tra-MoE (with Top-1 gating) can more effectively benefit from large-scale out-of-domain data (such as cross-embediments data and different physics engine simulated data) compared to the dense baseline. +- We demonstrate that Tra-MoE still significantly outperform expanded dense counterpart with equivalent parameters, highlighting the effectiveness of sparse MoE architecture in leveraging out-of-domain data through better parameter cooperation and specialization. +- We propose an adaptive policy conditioning technique for trajectory-guided policy, which ensures adaptive 2D trajectory representation and explicit spatial alignment with the observations, thereby achieving superior performance. + +# 2. Related Work + +Learning from multiple domains in Robot Learning. Scaling laws [30] have fueled substantial advancements in various fields. In robot learning, many recent studies [7, 8, 29, 51] investigate how compute, model size, and training data quantity affect the model performance on robotic manipulation tasks. Based on this, recent works have focused on collecting larger-scale real-world robotic datasets [11, 19, 32, 60, 64], and have successfully trained several robotics foundational models [6, 9, 16, 33, 38, 47, 57, + +[62] that demonstrate superior transfer capabilities. It is worth noting that these works mainly employ pre-training and fine-tuning paradigm rather than joint co-training paradigm, while focusing on evaluating the combination generalization and neglecting the performance impact on in-domain tasks. Additionally, these works directly combine heterogeneous robotic data with different action and observation spaces for joint training, potentially leading to sub-optimal solutions. In contrast, we leverage sparsely-gated MoE to incorporate larger-scale, out-of-domain, and action-less video data. Furthermore, our experiments demonstrate the importance of MoE for expanding out-of-domain data due to its superior capacity for parameter cooperation and specialization. + +Mixture-of-Experts. The Mixture-of-Experts (MoE) framework [12] is built upon a simple yet effective concept: dividing a model into specialized components, or 'experts', each focusing on distinct tasks or data aspects. This paradigm allows for selective activation of relevant experts based on the input, thereby maintaining computational efficiency while leveraging a vast reservoir of specialized knowledge through increased model capacity. Recently, in computer vision, natural language processing, and multi-modal, there has been a surge of foundation models trained based on the MoE architecture, including Swin-Moe [24], MoE-LLaVA [39], DeepSeek-V2 [40], Mixtral-8x22B [28], and so on. MoE typically encompasses dense [48] and sparse [14] variants. Dense MoE activates all experts in each iteration, while sparse MoE activates only a subset, generally resulting in lower computational overhead. However, sparse MoE can suffer from expert load imbalance and training instability issues. Techniques such as adding noise, load balancing loss [36], and router z-loss [74] are often employed to mitigate these issues and improve performance. In our work, we conduct a comprehensive empirical study of these techniques when training our sparsely-gated MoE-based trajectory model. + +Representations and Techniques for Policy Conditioning. Policy conditioning equips a multi-task robot policy with the ability to modulate its behavior by incorporating task-specific guidance information. This information, which may take various forms, guides the policy's action generation, enabling a single model to effectively learn and perform diverse tasks. Classic policy conditioning representations often encompass task identifiers [50], natural language instructions [8, 22, 45, 54], goal images [44], and demonstration videos [25, 27]. These are typically integrated with visual observations within the policy through concatenation or attention mechanisms. Recent works have explored alternative policy conditioning representations to enhance generalization and sample efficiency. These include leveraging internet-trained models to obtain bounding box [55] and object mask [42, 66, 68, 73], as well as learning plan [4], affordance [3, 46], future observations [17, 18, 34, 67], 2D or 3D trajectory [5, 61, 63, 65, 70] from human video data, + +and also hand-drawn trajectory [21] or sketch [56]. Different from these approaches that directly utilize raw or manually crafted signals for integration with visual observations, we introduce an adaptive policy conditioning technique that explicitly maps 2D trajectories onto the visual observations in space and encodes them as learnable embeddings. These designs ensure adaptive 2D trajectory representation and explicit spatial alignment with the visual observations. + +# 3. Methodology + +In this section, we first describe our problem formulation in Sec. 3.1. Subsequently, we give an overview of our proposed framework, which is composed of a sparsely-gated MoE-based trajectory prediction model and a trajectory-guided policy model using the adaptive policy conditioning technique, detailed in Sec. 3.2. Finally, we describe our training process and loss functions in Sec. 3.3. + +# 3.1. Problem Formulation + +Following the setup of previous works [5, 63, 65], our approach encompasses a trajectory prediction model to conduct language-conditioned any-point trajectory prediction and a trajectory-guided policy model to predict executable robot actions. In particular, as depicted in Fig. 2, we consider two types of datasets for training, which are large-scale, out-of-domain, and action-free videos $\mathcal{D}_{\text{ood}}$ , as well as small-scale, in-domain, and action-labeled demonstrations $\mathcal{D}_{\text{in}}$ . For trajectory prediction model, we employ $\mathcal{D}_{\text{ood}}$ and $\mathcal{D}_{\text{in}}$ for joint pre-training, whereas for trajectory-guided policy model, we only use $\mathcal{D}_{\text{in}}$ for training. + +# 3.2. Architecture + +ATM Revisited. In this paper, we use ATM [63] as the baseline for our method. The overall pipeline of ATM is illustrated on Fig. 2(b). It first learns a trajectory model from video data to perform language-conditioned trajectory prediction. The predicted trajectories is then used to learn a trajectory-guided policy using robot demonstrations. Specifically, as for trajectory model, its objective is to predict the future of any-point trajectories in a frame. Formally, given an image observation $o_t$ at timestep $t$ , any set of 2D query points simplified on the image frame $\mathbf{p}_t = \{p_{t,k}\}_{k=1}^K$ , and a language instruction $\ell$ , the trajectory model learns a mapping $\mathbf{p}_{t:t+H} = \tau_\theta(o_t, \mathbf{p}_t, \ell)$ that predicts the coordinates of these query points in the camera frame for the next $H$ time steps, where $\mathbf{p}_{t:t+H} \in \mathbb{R}^{H \times K \times 2}$ . For the trajectory-guided policy, it leverages the current observation $o_t$ and the predicted trajectory $\mathbf{p}_{t:t+H}$ to predict the subsequent robot actions. + +Sparsely-gated MoE-based trajectory model (Tra-MoE). ATM proposes a track transformer to implement the trajectory model. In track transformer, language tokens, image tokens, and track tokens to be decoded involve global interactions within the transformer block. However, directly + +![](images/1b360506ebbfb7a53d46babead0df0d6744f34a203836d82d0e49690726e4618.jpg) +Figure 3. (a) The pipeline of our sparsely-gated MoE-based trajectory model (Tra-MoE). (b) The pipeline of our trajectory-guided policy using the adaptive policy conditioning technique. Mapping means concatenating the trajectory mask with image observations, while learnable refers to setting each point in the trajectory mask as a learnable embedding. + +![](images/d1ccaa97ebefa6d3fb7b4012c22df2c42abd6eb7a4331b344e1bc0685f8f71ae.jpg) + +processing diverse domains and multiple modalities data within a unified transformer may yield suboptimal results. Hence, we propose our Tra-MoE framework. To develop our sparsely-gated MoE-based trajectory model, as depicted in Fig. 3(a), we replace certain layers of the transformer with MoE blocks. In each MoE block, it incorporates multiple feedforward networks (FFNs), each designated as an expert, and utilizes a gating function to activate a selected subset (Top- $K$ ) of these experts. In this paper, we set $K$ to a fixed value of 1 to ensure that the FLOPs per token remain constant relative to the baseline. Formally, the gating network $\mathcal{G}$ , parameterized by $\Theta$ and typically consisting of a linear-softmax network, yields the output $\mathcal{G}(\mathbf{x_s};\Theta)$ , where $\mathbf{x_s}$ represents our token sequences input. We formulate the sparsely-gated token-choice mechanism as follows: + +$$ +\mathcal {G} \left(\mathbf {x} _ {\mathrm {s}}; \boldsymbol {\Theta}\right) _ {i} = \operatorname {s o f t m a x} (\operatorname {T o p - K} (g \left(\mathbf {x} _ {\mathrm {s}}; \boldsymbol {\Theta}\right), k)) _ {i}, \tag {1} +$$ + +$$ +\operatorname {T o p - K} (g \left(\mathbf {x} _ {\mathbf {s}}; \boldsymbol {\Theta}\right), k) _ {i} = \left\{ \begin{array}{l l} g (\mathbf {x}; \boldsymbol {\Theta}) _ {i}, & \text {c o n d ,} \\ - \infty , & \text {o t h .} \end{array} \right., \tag {2} +$$ + +$$ +\operatorname {c o n d}: \text {i f} g (\mathbf {x}; \boldsymbol {\Theta}) _ {i} \text {i s i n t o p - k e l e m e n t s o f} g (\mathbf {x}; \boldsymbol {\Theta}). \tag {3} +$$ + +More specifically, $\operatorname{Top - K}(\cdot ,k)$ function retains only the top- $k$ values in a vector, setting the rest to $-\infty$ . Consequently, the output of the sparsely-gated MoE layer can be + +formulated as: + +$$ +\mathcal {F} _ {\text {s p a r s e}} ^ {\mathrm {M o E}} \left(\mathbf {x} _ {\mathbf {s}}; \boldsymbol {\Theta}, \left\{\mathbf {W} _ {i} \right\} _ {i = 1} ^ {K}\right) = \sum_ {i = 1} ^ {K} \mathcal {G} \left(\mathbf {x} _ {\mathbf {s}}; \boldsymbol {\Theta}\right) _ {i} f _ {i} \left(\mathbf {x} _ {\mathbf {s}}; \mathbf {W} _ {i}\right), \tag {4} +$$ + +where $f_{i}$ represents the i-th expert, usually a linear-GELU-linear FFN layer, is parameterized by $\mathbf{W}_i$ . + +Trajectory-guided policy using the adaptive policy conditioning technique. After completing the trajectory model learning, we develop our trajectory-guided policy through the adaptive policy conditioning technique based on the initial ATM, as shown in Fig. 3(b). In our policy model, independently encoded track tokens and image tokens undergo a early fusion within the fusion transformer. Subsequently, the fused features are further integrated with the raw trajectory data through a late fusion mechanism along the channel dimension, ensuring the preservation of original trajectory information for more accurate action prediction. + +Fundamentally, the policy model bridges the gap between the predicted 2D any-point trajectories and the executable 3D robot actions, functioning as an inverse dynamic model and eliminating the necessity of other task specification. However, how to effectively condition the policy using the predicted 2D trajectories remains an unresolved challenge. Based on our analysis, we present two key insights: (i) Explicit spatial alignment and fusion of the 2D trajectories with image observations turns out to be highly beneficial. + +It simplifies the mapping of 2D trajectories to image space and highlights motion-relevant regions, enabling the policy model to focus on important areas and reducing the learning difficulty. (ii) Trajectory positions exhibit different characteristics: starting points typically emphasize local motion, while endpoints focus more on global trend. Therefore, an adaptive trajectory representation and conditioning technique is expected to flexibly guide the policy by effectively capturing these variations. + +Based on the aforementioned analysis, we propose an adaptive policy conditioning technique to incorporate trajectory information into 2D images. Specifically, we construct an additional mask modality, and populate it with 2D trajectories based on their specific spatial locations, while setting the value of each point in the trajectory mask as a learnable embedding. Before encoding the image, we concatenate this mask modality with the image observations on the channel dimension, forming a tensor of dimensions $H \times W \times 4$ , as shown in Fig. 3(b). To sum up, our adaptive policy conditioning technique yield an adaptive 2D trajectory representation, which could be explicitly aligned with the image observations to guide the policy prediction more flexibly. + +# 3.3. Training + +We argue that trajectory models could be shared by multiple domains data. Therefore, we first pre-train a generalizable trajectory model on $\mathcal{D}_{\text{ood}} \cup \mathcal{D}_{\text{in}}$ . The $\mathcal{D}_{\text{ood}}$ typically contains additional environments, objects, skills, and embodiments. Then we train the trajectory-guided policy model on $\mathcal{D}_{\text{in}}$ based on the pre-trained trajectory model. In trajectory-guided policy training phase, we freeze the trajectory model. Pre-training sparsely-gated MoE-based trajectory model. Following ATM, we train the trajectory model using the ground truth tracks generated by CoTracker [31]. We use MSE to optimize the trajectory prediction loss $\mathcal{L}_{\text{tra}}$ . Additionally, we also consider adding an auxiliary load-balancing loss [36] to balance the activation frequency among experts. Given a batch of queries $\mathcal{B}$ , it contains B samples and each sample contains S tokens. We define $\mathcal{Q}_i$ as the proportion of tokens distributed to expert $i$ , and $\mathcal{P}_i$ as the proportion of the gating probability assigned to expert $i$ . The $\mathcal{L}_{\text{lo-ba}}$ can be formulated as follows: + +$$ +\mathcal {Q} _ {i} = \frac {1}{S \cdot B} \sum_ {x _ {s} ^ {m} \in \mathcal {B}} \mathbb {1} \left\{\operatorname {a r g m a x} \mathcal {G} \left(\mathbf {x} _ {\mathbf {s}} ^ {\mathbf {m}}; \boldsymbol {\Theta}\right) = i \right\}, \tag {5} +$$ + +$$ +\mathcal {P} _ {i} = \frac {1}{S \cdot B} \sum_ {x _ {\mathrm {s}} ^ {m} \in \mathcal {B}} \mathcal {G} \left(\mathbf {x} _ {\mathrm {s}} ^ {\mathrm {m}}; \boldsymbol {\Theta}\right) _ {i}, \tag {6} +$$ + +$$ +\mathcal {L} _ {\mathrm {l o - b a}} = N \sum_ {i = 1} ^ {N} \mathcal {Q} _ {i} \mathcal {P} _ {i}. \tag {7} +$$ + +Additionally, ST-MoE [74] proposed using the router $z$ -loss to penalize large logits entering the gating network, + +thereby improving training stability. We similarly applies it to the training of our trajectory model, as shown below: + +$$ +\mathcal {L} _ {z} (x) = \frac {1}{S} \sum_ {k = 1} ^ {S} \left(\log \sum_ {i = 1} ^ {N} e ^ {\mathbf {g} _ {i} ^ {(k)}}\right) ^ {2}, \tag {8} +$$ + +where $S$ is the total number of tokens, $N$ is the total number of experts, and $\mathbf{g} \in \mathcal{R}^{S \times N}$ are the logits going into the router. In summary, the trajectory model's training loss is as follow: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \lambda_ {\text {t r a}} \cdot \mathcal {L} _ {\text {t r a}} + \lambda_ {\text {l o - b a}} \cdot \mathcal {L} _ {\text {l o - b a}} + \lambda_ {z} \cdot \mathcal {L} _ {z}. \tag {9} +$$ + +Training trajectory-guided policy through adaptive policy conditioning technique. For trajectory-guided policy training, we employ the paradigm of behavior cloning. Specifically, we use the MSE between predicted actions and ground truths as the loss function. + +# 4. Experiments + +We perform evaluation experiments using the LIBERO benchmark [41] and a real low-cost dual-arm robot [35]. In experiments, we quantitatively find that the success rate of downstream manipulation tasks is generally positively correlated with the performance of the trajectory model. Detailed quantitative experiment results can be found in the appendix. Meanwhile, CoTracker [31] can provide relatively accurate trajectory annotations; therefore, we report the average success rate of downstream manipulation tasks as our metric. Our experiments aim to study the following questions: + +Q1: Does the techniques of router z-loss, load-balancing loss, and adding noise to gating logits enhance the training of sparsely-gated Tra-MoE? +Q2: Does sparsely-gated Tra-MoE, compared to the dense Tra-baseline, benefit more from large-scale out-of-domain data (such as cross-embediments and different physics engine data) and demonstrate superior scaling up capabilities? +Q3: When we scale up the dense Tra-baseline to match the parameter count of Tra-MoE, does the sparsely-gated Tra-MoE still maintain its advantage? +Q4: Is our proposed adaptive policy conditioning technique for trajectory-guided policy more effective than ATM? + +# 4.1. Simulation Experiments + +Experiment setup. The LIBERO [41] is divided into five categories: LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, LIBERO-Long, and LIBERO-90. LIBERO-90 includes 90 tasks, while each of the other categories contains 10 tasks. Compared to other categories, LIBERO-90 presents a highly varied selection of object categories, layouts, environments, and skill goals. It is typically used for pre-training rather than joint co-training. Therefore, our main ablation experiments select 90 tasks from LIBERO-90 to construct + +
SpatialGoalObjectLongAvg.
062.073.071.021.556.9
5·10-469.573.568.022.558.4
1·10-462.581.073.528.561.4
1·10-560.065.568.527.055.3
1·10-660.079.053.031.555.9
+ +(a) The value of $\lambda_z$ + +
SpatialGoalObjectLongAvg.
w/o. noise62.073.071.022.556.9
w. noise71.567.057.527.055.8
+ +(c) Adding a noise term to the gating logits. + +
Num.SpatialGoalObjectLongAvg.
149.567.056.535.052.0
259.072.565.527.056.0
364.068.071.017.555.1
462.073.071.021.556.9
+ +(e) The number of experts. + +
SpatialGoalObjectLongAvg.
062.073.071.022.556.9
1·10-358.574.060.018.052.6
1·10-561.069.556.026.053.1
1·10-766.567.073.522.057.3
+ +(b) The value of $\lambda_{lo - ba}$ . + +
SpatialGoalObjectLongAvg.
MoE62.581.073.528.561.4
Width66.550.561.515.048.4
Depth61.575.546.526.552.5
+ +(d) Scaling Tra-baseline from width and depth. + +
MoEOODSpatialGoalObjectLongAvg.
67.568.568.026.557.6
✓63.064.060.519.551.8
✓49.567.056.535.052.0
✓✓62.581.073.528.561.4
+ +dataset $\mathcal{D}_{\text{ood}}$ , utilizing 20 action-free videos per task; and choose 40 tasks from LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, and LIBERO-Long to build dataset $\mathcal{D}_{\text{in}}$ , using 10 action-labeled demonstrations per task. In summary, for training Tra-MoE, our main ablation experiments use a total of 2, 200 action-free videos, with an out-of-domain to in-domain data ratio of $9:2$ . Additionally, on this basis, we also further select 92 tasks from the RLbench [26] (using CoppeliaSim rather than MuJoCo as physics engine) simulation environment and generate 5 videos for each task, totaling 2, 660 videos in conjunction with the LIBERO data, to train Tra-MoE. The different physics engines and dynamics properties further increase the data distribution discrepancy in $\mathcal{D}_{\text{ood}}$ . Finally, as for the trajectory-guided policy, we train a separate multi-task policy for each category. Each policy undergoes behavioral cloning training with 100 action-labeled demonstrations from 10 tasks. + +Implement details. Building on ATM [63], we develop our sparsely-gated Tra-MoE and adaptive policy conditioning technique. Specifically, when using only LIBERO data, we replace the 1th, 5th, and 8th layers of the track transformer with MoE blocks. When further incorporating RLbench data, we additionally replace the 2th and 7th layers. Unless otherwise specified, we set the number of experts to 4 by default and use a top-1 gating strategy for token choice to ensure the FLOPs per token remain constant. + +Experiment results and analysis. We present empirical simulation results to address the aforementioned research questions and make the following findings: + +Result 1: In Tab. 1a, we examine the effect of the router + +(f) Scaling with MoE and OOD data. +Table 1. The ablation experiments of the sparsely-gated Tra-MoE. All baselines are the track transformer of ATM. + +
SpatialGoalObjectLongAvg.
ATM51.042.055.519.041.9
Tra-MoE54.062.061.041.054.5
+ +Table 2. The comparison results when training trajectory model using RLbench video data. + +z-loss weight $\lambda_z$ . The results indicate that a higher $\lambda_z$ can effectively enhance performance, with an improvement of 4.5 (61.4 vs. 56.9) at $\lambda_z = 1 \cdot 10^{-4}$ . In Tab. 1b, we study the impact of the load-balancing loss weight $\lambda_{lo - ba}$ . The results show a significant performance decline with larger values of $\lambda_{lo - ba}$ , with only a slight improvement ( $56.7 \rightarrow 57.3$ ) at $\lambda_{lo - ba} = 1 \cdot 10^{-7}$ . In Tab. 1c, we verify the effectiveness of adding noise to the gating logits, which results in a performance decrease of by 1.1 ( $56.9 \rightarrow 55.8$ ). + +Finding 1: Based on the above results, we find that using router z-loss to penalize large logits entering the gating network can effectively improve training stability and lead to performance gains. Meanwhile, the load-balancing loss used to balance the activation frequency among experts and the addition of noise to enhance expert allocation exploration cannot bring effective performance improvements. An intuitive explanation is that when the data distribution is uneven, the load-balancing loss might force experts to set shared parameters on data with a large domain gap, which could lead to mutual counteraction of learning gradients. This undermines the advantage of parameter specialization in sparsely-gated MoE. Similarly, adding noise is also detrimental to learning parameter specialization among experts. + +Result 2: We comprehensively investigate the impact of in + +
SpatialGoalObjectLongAvg.
ATM62.581.073.528.561.4
+ DM69.058.085.533.561.5
+ AM69.577.088.030.566.3
+ +Table 3. The ablation experiments of the adaptive policy conditioning technique. DM (RT-Trajectory style) and AM respectively represent hand-drawn mask and adaptive mask. + +producing sparsely-gated MoE and large-scale out-of-domain data on scaling up trajectory models in Tab. 1f. Firstly, when training only with in-domain data, Tra-baseline achieves an average success rate of 57.6, while Tra-MoE only achieves an average success rate of 51.8. Secondly, when adding out-of-domain data to jointly train Tra-baseline, the average success rate drops to $52.0$ ( $57.6 \rightarrow 52.0$ ), with particularly significant performance declines in LIBERO-Spatial and LIBERO-Object. However, when adding out-of-domain data to jointly train Tra-MoE, the average success rate increases to $61.4$ ( $51.8 \rightarrow 61.4$ ). Additionally, in Tab. 1e, we also explore the impact of the number of experts in sparsely-gated MoE. The results show that using just two experts can lead to an increase of 4.0 (52.0 vs. 56.0) in the average success rate. With additional experts, performance generally trends upward, albeit with some fluctuations. Finally, we report the comparison results of training trajectory model by further using RLbench data in Tab. 2. The results indicate that although the introduction of RLbench data led to a performance drop in LIBERO, our Tra-MoE still outperform the baseline by 12.6 (54.5 vs. 41.9). + +Finding 2: Based on the above results, we have the following findings and analysis: (i) When the data scale is small, naively expanding the model capacity through sparsely-gated MoE does not lead to performance improvements. A explanation is that the larger model may lead to a overfitting problem when training on small-scale data. Moreover, small-scale data often leads to some experts not being sufficiently trained, thereby affecting performance. (ii) Naively introducing out-of-domain data for joint training often leads to performance degradation in the target domain. However, simultaneously incorporating out-of-domain data and expanding models with sparsely-gated MoE may improve performance or alleviate performance degradation. This phenomenon suggests that although out-of-domain data generally enhances model generalization, it could hinder task-specific performance in the target domain due to insufficient in-domain data learning. In contrast, expanding the model with out-of-domain data using sparsely-gated MoE achieves a more optimal balance between parameter cooperation and specialization across diverse data and even different tokens within the same data input. This design allows these large-scale data to jointly train most of parameters, ensuring the capture of general patterns for mutual cooperation. (iii) Even with a few experts, Tra-MoE exhibits significant scaling up + +capabilities. Additionally, further increasing the number of experts generally leads to performance improvement. However, this improvement is not guaranteed due to factors like data-expert complexity matching and training stability. + +Result 3: In Tab. 1d, we also investigate the performance comparison between the dense Tra-baseline and the sparse Tra-MoE with the same number of parameters. Specifically, we expand Tra-baseline to match the model capacity of Tra-MoE by increasing its model width or depth separately. These expanded versions achieve average success rates of 48.4 and 52.5 respectively, which are significantly lower than Tra-MoE's 61.4. + +Finding 3: Based on these results, we are surprised to find that simply expanding the dense Tra-baseline in depth and width can not effectively improve performance. This highlights the importance of the sparsely-gated MoE architecture when training with large-scale out-of-domain data, as it ensures dynamic activation of different experts based on different data input and different tokens within the same data input, significantly reducing optimization conflict problems. + +Result 4: In Tab. 3, we investigate the effectiveness of the adaptive policy conditioning. Specifically, building upon ATM, we first introduce an additional hand-drawn mask modality before image encoding. In this mask, the first half of the trajectory is set to 128, while the second half is set to 255. We call it 'ATM + hand-drawn mask' and this is similar to RT-Trajectory [21]. Next, we replace the hand-drawn mask with an adaptive mask, where each point on the trajectory is instantiated as a learnable embedding. We call it 'ATM + adaptive mask'. The results show that 'ATM + hand-drawn mask' achieves only a 0.1 improvement compared to the baseline. However, except for a significant performance decrease in LIBERO-Goal, it achieves performance improvements of 6.5, 12.0, and 5.0 in LIBERO-Spatial, LIBERO-Object, and LIBERO-Long respectively. In addition, the 'ATM + adaptive mask' achieves a performance improvement of 4.9 (66.3 vs. 61.4) compared to the baseline. It further improves performance across all suites compared to 'ATM + hand-drawn mask', especially achieving a 19.0 performance gains in LIBERO-Goal (77.0 vs. 58.0). + +Finding 4: Based on these results, we find that explicitly mapping 2D trajectories to visual observations significantly improves the performance of LIBERO-Spatial, LIBERO-Object, and LIBERO-Long, but causes a notable decrease in LIBERO-Goal. By instantiating the points in the 2D trajectory as learnable embeddings, we not only further enhances the performance of LIBERO-Spatial, LIBERO-Object, and LIBERO-Long, but also significantly improves LIBERO-Goal. We analyze that the decline in LIBERO-Goal is mainly due to overlapping front and back parts of some trajectories in the 2D images, leading to mapping errors. However, our adaptive trajectory representation and conditioning technique effectively mitigates this issue. In summary, these + +
Human DataMoEAdaptive Trajectory MaskPourPushPick and PassTissueFoldAverage
40.045.050.030.025.038.0
✓45.035.050.025.035.038.0
✓✓40.050.060.035.045.046.0
✓✓✓60.070.065.035.050.056.0
+ +Table 4. The ablation of ood human data, MoE architecture, and adaptive policy conditioning technique on our real-world tasks. + +![](images/84503aacec77ac5c54356f89d97bcde412c1ed4dba170d2a1d766ebf2aa969d0.jpg) +Figure 4. Left: The real-world experiments hardware platform setup. Right: The real-world tasks evaluation demonstrations. + +experiments collectively demonstrate the effectiveness of explicitly aligning 2D trajectories with images in space as well as applying the adaptive conditioning technique. + +# 4.2. Real-World Experiments + +Experiment setup. As shown in Fig 4, we utilize a dual-arm robot [35] to conduct real-world experiments. We evaluate 5 tasks, which include two single-arm tasks: pouring water and pushing a vegetable to the side of the cutting board, as well as three dual-arm tasks: pulling out tissues, folding towels, and picking up and passing a holder. Specifically, for each task, we collect 50 action-labeled robot demonstrations. Concurrently, we also collect 50 videos of human (cross-embediments) performing for each task. Compared to the robot data, the human video encompasses more environment variations and distractor objects. Please see the appendix for more detailed evaluation details and real-world setup. + +Implement details. Following our simulation experiments, we utilize two camera views and uniformly resize their resolution to $128 \times 128$ for trajectory model and policy training. We train our real-world Tra-MoE on borh human video and robot video data, with a separate policy trained for each task. Therefore, our real-world Tra-MoE integrates data across different embodiments (human, single-arm robot, and dual-arm robot). They have significantly different visual appearances and dynamics properties. Finally, we also present some real-world task evaluation demonstrations in Fig. 4. + +Experiment results and analysis. To further strengthen our conclusions that sparsely-gated MoE can better learn trajectory model from out-of-domain data and our adaptive policy + +condition technique could lead to more flexible guidance for policy learning, we conduct real-world ablation experiments, as shown in Tab. 4. The results indicate that simply introducing human video data cannot lead to performance improvement. However, when expanding the model by MoE architecture, the performance increases from $38\%$ to $46\%$ . We speculate this may be due to the significant differences in appearance and dynamics between human and robot, with the former typically moving much faster and more flexibly. Finally, when further incorporating an adaptive trajectory mask for policy conditioning, the performance is further increased from $46\%$ to $56\%$ . + +# 5. Conclusion + +In this work, we have introduced Tra-MoE, a novel framework that utilizes a sparsely-gated MoE architecture with Top-1 gating strategy to learn trajectory model from multiple domains. Tra-MoE achieves superior parameter cooperation and specialization, enabling more effective utilization of large-scale out-of-domain data while maintaining constant FLOPs per token. We train our Tra-MoE and conduct a comprehensive empirical study, demonstrating that TraMoE consistently exhibits superior performance compared to the dense baseline, even when the latter is scaled to match Tra-MoE's parameter count. Additionally, we propose an adaptive policy conditioning technique for trajectory-guided policy, ensuring adaptive 2D trajectory representation and explicit spatial alignment with observations, thereby achieving superior performance. Extensive experiments in both simulated and real-world environments validate our findings and demonstrate the effectiveness of our proposed methods. + +# Acknowledgements + +This work is supported by the National Key R&D Program of China (No. 2022ZD0160900), Jiangsu Frontier Technology Research and Development Program (No. BF2024076), and the Collaborative Innovation Center of Novel Software Technology and Industrialization. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1 +[2] OpenAI: Marcin Andrychowicz, Bowen Baker, Maciek Chogiej, Rafal Jozefowicz, Bob McGrew, Jakub Pachocki, Arthur Petron, Matthias Plappert, Glenn Powell, Alex Ray, et al. Learning dexterous in-hand manipulation. *IJRR*, 39(1): 3-20, 2020. 1 +[3] Shikhar Bahl, Russell Mendonca, Lili Chen, Unnat Jain, and Deepak Pathak. Affordances from human videos as a versatile representation for robotics. In CVPR, pages 13778-13790, 2023. 1, 3 +[4] Homanga Bharadhwaj, Abhinav Gupta, Vikash Kumar, and Shubham Tulsiani. Towards generalizable zero-shot manipulation via translating human interaction plans. In ICRA, pages 6904-6911. IEEE, 2024. 1, 3 +[5] Homanga Bharadhwaj, Roozbeh Mottaghi, Abhinav Gupta, and Shubham Tulsiani. Track2act: Predicting point tracks from internet videos enables diverse zero-shot robot manipulation. arXiv preprint arXiv:2405.01527, 2024. 1, 3 +[6] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. 2 +[7] Konstantinos Bousmalis, Giulia Vezzani, Dushyant Rao, Coline Manon Devin, Alex X Lee, Maria Bauza Villalonga, Todor Davchev, Yuxiang Zhou, Agrim Gupta, Akhil Raju, et al. Robocat: A self-improving generalist agent for robotic manipulation. TMLR, 2023. 2 +[8] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 2, 3 +[9] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 2 +[10] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1 +[11] Qingwen Bu, Jisong Cai, Li Chen, Xiuqi Cui, Yan Ding, Siyuan Feng, Shenyuan Gao, Xindong He, Xu Huang, Shu Jiang, et al. Agibot world colosseo: A large-scale manipulation platform for scalable and intelligent embodied systems. arXiv preprint arXiv:2503.06669, 2025. 2 + +[12] Weilin Cai, Juyong Jiang, Fan Wang, Jing Tang, Sunghun Kim, and Jiayi Huang. A survey on mixture of experts. arXiv preprint arXiv:2407.06204, 2024. 3 +[13] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, pages 9640-9649, 2021. 1 +[14] Zitian Chen, Yikang Shen, Mingyu Ding, Zhenfang Chen, Hengshuang Zhao, Erik G Learned-Miller, and Chuang Gan. Mod-squad: Designing mixtures of experts as modular multitask learners. In CVPR, pages 11828-11837, 2023. 3 +[15] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1 +[16] Ria Doshi, Homer Rich Walke, Oier Mees, Sudeep Dasari, and Sergey Levine. Scaling cross-embodied learning: One policy for manipulation, navigation, locomotion and aviation. In CoRL, 2 +[17] Yilun Du, Sherry Yang, Pete Florence, Fei Xia, Ayzaan Wahid, Pierre Sermanet, Tianhe Yu, Pieter Abbeel, Joshua B Tenenbaum, Leslie Pack Kaelbling, et al. Video language planning. In The Twelfth International Conference on Learning Representations. 3 +[18] Yilun Du, Sherry Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Josh Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation. NeurIPS, 36, 2024. 3 +[19] Hao-Shu Fang, Hongjie Fang, Zhenyu Tang, Jirong Liu, Junbo Wang, Haoyi Zhu, and Cewu Lu. Rh20t: A robotic dataset for learning diverse skills in one-shot. In RSS 2023 Workshop on Learning for Task and Motion Planning, 2023. +[20] Chongkai Gao, Haozhuo Zhang, Zhixuan Xu, Cai Zhehao, and Lin Shao. Flip: Flow-centric generative planning as general-purpose manipulation world model. In ICLR. 1 +[21] Jiayuan Gu, Sean Kirmani, Paul Wohlhart, Yao Lu, Montserrat Gonzalez Arenas, Kanishka Rao, Wenhao Yu, Chuyuan Fu, Keerthana Gopalakrishnan, Zhuo Xu, et al. Rt-trajectory: Robotic task generalization via hindsight trajectory sketches. In ICLR. 3, 7 +[22] Pierre-Louis Guhur, Shizhe Chen, Ricardo Garcia Pinel, Makarand Tapaswi, Ivan Laptev, and Cordelia Schmid. Instruction-driven history-aware policies for robotic manipulations. In CORL, pages 175-187. PMLR, 2023. 3 +[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1 +[24] Changho Hwang, Wei Cui, Yifan Xiong, Ziyue Yang, Ze Liu, Han Hu, Zilong Wang, Rafael Salas, Jithin Jose, Prabhat Ram, et al. Tutel: Adaptive mixture-of-experts at scale. Proceedings of Machine Learning and Systems, 5:269-287, 2023. 3 +[25] Vidhi Jain, Maria Attarian, Nikhil J Joshi, Ayzaan Wahid, Danny Driess, Quan Vuong, Pannag R Sanketi, Pierre Sermanet, Stefan Welker, Christine Chan, et al. Vid2robot: End-to-end video-conditioned policy learning with cross-attention transformers. arXiv preprint arXiv:2403.12943, 2024. 3 +[26] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J Davison. Rlbench: The robot learning benchmark & learning environment. RAL, 5(2):3019-3026, 2020. 6 + +[27] Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kappler, Frederik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. Bc-z: Zero-shot task generalization with robotic imitation learning. In CoRL, pages 991-1002. PMLR, 2022. 3 +[28] Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. Mixtral of experts. arXiv preprint arXiv:2401.04088, 2024. 3 +[29] Yunfan Jiang, Agrim Gupta, Zichen Zhang, Guanzhi Wang, Yongqiang Dou, Yanjun Chen, Li Fei-Fei, Anima Anandkumar, Yuke Zhu, and Linxi Fan. Vima: General robot manipulation with multimodal prompts. arXiv preprint arXiv:2210.03094, 2(3):6, 2022. 2 +[30] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 2 +[31] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv preprint arXiv:2307.07635, 2023. 5 +[32] Alexander Khazatsky, Karl Pertsch, Suraj Nair, Ashwin Balakrishna, Sudeep Dasari, Siddharth Karamcheti, Soroush Nasiriany, Mohan Kumar Srirama, Lawrence Yunliang Chen, Kirsty Ellis, et al. Droid: A large-scale in-the-wild robot manipulation dataset. arXiv preprint arXiv:2403.12945, 2024. 2 +[33] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 2 +[34] Po-Chen Ko, Jiayuan Mao, Yilun Du, Shao-Hua Sun, and Joshua B Tenenbaum. Learning to act from actionless videos through dense correspondences. In ICLR. 3 +[35] Alexander Koch. Low-cost robot arm. https://github.com/AlexanderKoch-Koch/low_costrobot, 2024. GitHub repository. 5, 8 +[36] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan First, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. arXiv preprint arXiv:2006.16668, 2020. 2, 3, 5 +[37] Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuomotor policies. JMLR, 17(39):1-40, 2016. 1 +[38] Xinghang Li, Minghuan Liu, Hanbo Zhang, Cunjun Yu, Jie Xu, Hongtao Wu, Chilam Cheang, Ya Jing, Weinan Zhang, Huaping Liu, et al. Vision-language foundation models as effective robot imitators. In ICLR. 2 +[39] Bin Lin, Zhenyu Tang, Yang Ye, Jiaxi Cui, Bin Zhu, Peng Jin, Junwu Zhang, Munan Ning, and Li Yuan. Moe-llava: Mixture of experts for large vision-language models. arXiv preprint arXiv:2401.15947, 2024. 3 +[40] Aixin Liu, Bei Feng, Bin Wang, Bingxuan Wang, Bo Liu, Chenggang Zhao, Chengqi Dengr, Chong Ruan, Damai Dai, + +Daya Guo, et al. Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model. arXiv preprint arXiv:2405.04434, 2024. 3 +[41] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning. NeurIPS, 36, 2024. 5 +[42] Fangchen Liu, Kuan Fang, Pieter Abbeel, and Sergey Levine. Moka: Open-vocabulary robotic manipulation through mark-based visual prompting. In First Workshop on Vision-Language Models for Navigation and Manipulation at ICRA 2024. 3 +[43] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. In ICLR, 2022. 1 +[44] Corey Lynch, Mohi Khansari, Ted Xiao, Vikash Kumar, Jonathan Tompson, Sergey Levine, and Pierre Sermanet. Learning latent plans from play. In CoRL, pages 1113-1132. PMLR, 2020. 3 +[45] Corey Lynch, Ayzaan Wahid, Jonathan Tompson, Tianli Ding, James Betker, Robert Baruch, Travis Armstrong, and Pete Florence. Interactive language: Talking to robots in real time. RAL, 2023. 3 +[46] Soroush Nasiriany, Sean Kirmani, Tianli Ding, Laura Smith, Yuke Zhu, Danny Driess, Dorsa Sadigh, and Ted Xiao. Rt-affordance: Affordances are versatile intermediate representations for robot manipulation. arXiv preprint arXiv:2411.02704, 2024. 3 +[47] Abhishek Padalkar, Acorn Pooley, Ajinkya Jain, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anikait Singh, Anthony Brohan, et al. Open x-embediment: Robotic learning datasets and rt-x models. arXiv preprint arXiv:2310.08864, 2023. 2 +[48] Joan Puigcerver, Carlos Riquelme, Basil Mustafa, and Neil Houlsby. From sparse to soft mixtures of experts. arXiv preprint arXiv:2308.00951, 2023. 3 +[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICLR, pages 8748-8763. PMLR, 2021. 1 +[50] Rouhollah Rahmatizadeh, Pooya Abolghasemi, Ladislau Böloni, and Sergey Levine. Vision-based multi-task manipulation for inexpensive robots using end-to-end learning from demonstration. In ICRA, pages 3758-3765. IEEE, 2018. 3 +[51] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. arXiv preprint arXiv:2205.06175, 2022. 1, 2 +[52] Younggyo Seo, Kimin Lee, Stephen L James, and Pieter Abbeel. Reinforcement learning with action-free pre-training from videos. In ICLR, pages 19561-19579. PMLR, 2022. 1 +[53] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017. 2 + +[54] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In CoRL, pages 894-906. PMLR, 2022. 3 +[55] Austin Stone, Ted Xiao, Yao Lu, Keerthana Gopalakrishnan, Kuang-Huei Lee, Quan Vuong, Paul Wohlhart, Brianna Zitkovich, Fei Xia, Chelsea Finn, et al. Open-world object manipulation using pre-trained vision-language models. In CoRL, 2023. 3 +[56] Priya Sundaresan, Quan Vuong, Jiayuan Gu, Peng Xu, Ted Xiao, Sean Kirmani, Tianhe Yu, Michael Stark, Ajinkya Jain, Karol Hausman, et al. Rt-sketch: Goal-conditioned imitation learning from hand-drawn sketches. 2024. 3 +[57] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. 2 +[58] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 1 +[59] Mel Vecerik, Carl Doersch, Yi Yang, Todor Davchev, Yusuf Aytar, Guangyao Zhou, Raia Hadsell, Lourdes Agapito, and Jon Scholz. Robotap: Tracking arbitrary points for few-shot visual imitation. In ICRA, pages 5397-5403. IEEE, 2024. 1 +[60] Homer Rich Walke, Kevin Black, Tony Z Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen-Estruch, Andre Wang He, Vivek Myers, Moo Jin Kim, Max Du, et al. Bridgedata v2: A dataset for robot learning at scale. In CoRL, pages 1723-1736. PMLR, 2023. 2 +[61] Chen Wang, Linxi Fan, Jiankai Sun, Ruohan Zhang, Li Fei-Fei, Danfei Xu, Yuke Zhu, and Anima Anandkumar. Mimicplay: Long-horizon imitation learning by watching human play. In CoRL. 1, 3 +[62] Lirui Wang, Xinlei Chen, Jialiang Zhao, and Kaiming He. Scaling proprioceptive-visual learning with heterogeneous pre-trained transformers. CoRR, abs/2409.20537, 2024. 3 +[63] Chuan Wen, Xingyu Lin, John So, Kai Chen, Qi Dou, Yang Gao, and Pieter Abbeel. Any-point trajectory modeling for policy learning. arXiv preprint arXiv:2401.00025, 2023. 1, 3, 6 +[64] Kun Wu, Chengkai Hou, Jiaming Liu, Zhengping Che, Xiaozhu Ju, Zhuqin Yang, Meng Li, Yinuo Zhao, Zhiyuan Xu, Guang Yang, et al. Robomind: Benchmark on multi-embediment intelligence normative data for robot manipulation. arXiv preprint arXiv:2412.13877, 2024. 2 +[65] Mengda Xu, Zhenjia Xu, Yinghao Xu, Cheng Chi, Gordon Wetzstein, Manuela Veloso, and Shuran Song. Flow as the cross-domain manipulation interface. arXiv preprint arXiv:2407.15208, 2024. 1, 3 +[66] Jiange Yang, Wenhui Tan, Chuhao Jin, Bei Liu, Jianlong Fu, Ruihua Song, and Limin Wang. Pave the way to grasp anything: Transferring foundation models for universal pickplace robots. arXiv preprint arXiv:2306.05716, 1(2), 2023. 3 +[67] Jiange Yang, Bei Liu, Jianlong Fu, Bocheng Pan, Gangshan Wu, and Limin Wang. Spatiotemporal predictive pre-training for robotic motor control. arXiv preprint arXiv:2403.05304, 2024. 3 + +[68] Jiange Yang, Wenhui Tan, Chuhao Jin, Keling Yao, Bei Liu, Jianlong Fu, Ruihua Song, Gangshan Wu, and Limin Wang. Transferring foundation models for generalizable robotic manipulation. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), pages 1999-2010, 2025. 3 +[69] Seonghyeon Ye, Joel Jang, Byeongguk Jeon, Sejune Joo, Jianwei Yang, Baolin Peng, Ajay Mandlekar, Reuben Tan, Yu-Wei Chao, Bill Yuchen Lin, et al. Latent action pretraining from videos. arXiv preprint arXiv:2410.11758, 2024. 1 +[70] Chengbo Yuan, Chuan Wen, Tong Zhang, and Yang Gao. General flow as foundation affordance for scalable robot learning. arXiv preprint arXiv:2401.11439, 2024. 3 +[71] Jinguo Zhu, Xizhou Zhu, Wenhai Wang, Xiaohua Wang, Hongsheng Li, Xiaogang Wang, and Jifeng Dai. Uniperceiver-moe: Learning sparse generalist models with conditional moes. NeurIPS, 35:2664-2678, 2022. 1 +[72] Xizhou Zhu, Jinguo Zhu, Hao Li, Xiaoshi Wu, Hongsheng Li, Xiaohua Wang, and Jifeng Dai. Uni-perceiver: Pre-training unified architecture for generic perception for zero-shot and few-shot tasks. In CVPR, pages 16804-16815, 2022. 1 +[73] Yifeng Zhu, Zhenyu Jiang, Peter Stone, and Yuke Zhu. Learning generalizable manipulation policies with object-centric 3d representations. In CoRL. 3 +[74] Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam Shazeer, and William Fedus. St-moe: Designing stable and transferable sparse expert models. arXiv preprint arXiv:2202.08906, 2022. 2, 3, 5 \ No newline at end of file diff --git a/2025/Tra-MoE_ Learning Trajectory Prediction Model from Multiple Domains for Adaptive Policy Conditioning/images.zip b/2025/Tra-MoE_ Learning Trajectory Prediction Model from Multiple Domains for Adaptive Policy Conditioning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dada2c0d1141856bd5117571da01c410c329ea0e --- /dev/null +++ b/2025/Tra-MoE_ Learning Trajectory Prediction Model from Multiple Domains for Adaptive Policy Conditioning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ca36ea14986ad168fc5cfe0fc3fbb7316bf23bcd149f97ce0081ecf4a754d56 +size 460906 diff --git a/2025/Tra-MoE_ Learning Trajectory Prediction Model from Multiple Domains for Adaptive Policy Conditioning/layout.json b/2025/Tra-MoE_ Learning Trajectory Prediction Model from Multiple Domains for Adaptive Policy Conditioning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9f0785240f958a8cd4d153053b117a41df719b10 --- /dev/null +++ b/2025/Tra-MoE_ Learning Trajectory Prediction Model from Multiple Domains for Adaptive Policy Conditioning/layout.json @@ -0,0 +1,8758 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 80, + 102, + 532, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 102, + 532, + 140 + ], + "spans": [ + { + "bbox": [ + 80, + 102, + 532, + 140 + ], + "type": "text", + "content": "Tra-MoE: Learning Trajectory Prediction Model from Multiple Domains for Adaptive Policy Conditioning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 159, + 548, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 159, + 548, + 175 + ], + "spans": [ + { + "bbox": [ + 66, + 159, + 548, + 175 + ], + "type": "text", + "content": "Jiange Yang " + }, + { + "bbox": [ + 66, + 159, + 548, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 66, + 159, + 548, + 175 + ], + "type": "text", + "content": " Haoyi Zhu " + }, + { + "bbox": [ + 66, + 159, + 548, + 175 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 66, + 159, + 548, + 175 + ], + "type": "text", + "content": " Yating Wang " + }, + { + "bbox": [ + 66, + 159, + 548, + 175 + ], + "type": "inline_equation", + "content": "^{2,4}" + }, + { + "bbox": [ + 66, + 159, + 548, + 175 + ], + "type": "text", + "content": " Gangshan Wu " + }, + { + "bbox": [ + 66, + 159, + 548, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 159, + 548, + 175 + ], + "type": "text", + "content": " Tong He " + }, + { + "bbox": [ + 66, + 159, + 548, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 66, + 159, + 548, + 175 + ], + "type": "text", + "content": " Limin Wang " + }, + { + "bbox": [ + 66, + 159, + 548, + 175 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 175, + 533, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 175, + 533, + 204 + ], + "spans": [ + { + "bbox": [ + 77, + 175, + 533, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 175, + 533, + 204 + ], + "type": "text", + "content": "State Key Laboratory for Novel Software Technology, Nanjing University, China " + }, + { + "bbox": [ + 77, + 175, + 533, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 175, + 533, + 204 + ], + "type": "text", + "content": "Shanghai AI Laboratory, " + }, + { + "bbox": [ + 77, + 175, + 533, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 77, + 175, + 533, + 204 + ], + "type": "text", + "content": "University of Science and Technology of China, " + }, + { + "bbox": [ + 77, + 175, + 533, + 204 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 77, + 175, + 533, + 204 + ], + "type": "text", + "content": "Tongji University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 167, + 205, + 441, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 205, + 441, + 217 + ], + "spans": [ + { + "bbox": [ + 167, + 205, + 441, + 217 + ], + "type": "text", + "content": "jiangyang.jgy@gmail.com, {gswu,lmwang}@nju.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 189, + 220, + 421, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 220, + 421, + 231 + ], + "spans": [ + { + "bbox": [ + 189, + 220, + 421, + 231 + ], + "type": "text", + "content": "{zhuhaoyi, wangyating, hetong}@pjlab.org.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 200, + 233, + 407, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 233, + 407, + 244 + ], + "spans": [ + { + "bbox": [ + 200, + 233, + 407, + 244 + ], + "type": "text", + "content": "https://github.com/MCG-NJU/Tra-MoE" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 298, + 297, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 297, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 297, + 597 + ], + "type": "text", + "content": "Learning from multiple domains is a primary factor that influences the generalization of a single unified robot system. In this paper, we aim to learn the trajectory prediction model by using broad out-of-domain data to improve its performance and generalization ability. Trajectory model is designed to predict any-point trajectories in the current frame given an instruction and can provide detailed control guidance for robotic policy learning. To handle the diverse out-of-domain data distribution, we propose a sparsely-gated MoE (Top-1 gating strategy) architecture for trajectory model, coined as Tra-MoE. The sparse activation design enables good balance between parameter cooperation and specialization, effectively benefiting from large-scale out-of-domain data while maintaining constant FLOPs per token. In addition, we further introduce an adaptive policy conditioning technique by learning 2D mask representations for predicted trajectories, which is explicitly aligned with image observations to guide action prediction more flexibly. We perform extensive experiments on both simulation and real-world scenarios to verify the effectiveness of Tra-MoE and adaptive policy conditioning technique. We also conduct a comprehensive empirical study to train Tra-MoE, demonstrating that our Tra-MoE consistently exhibits superior performance compared to the dense baseline model, even when the latter is scaled to match Tra-MoE's parameter count." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 622, + 297, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 622, + 297, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 622, + 297, + 696 + ], + "type": "text", + "content": "In computer vision and natural language processing, significant progress has been made through learning models from multiple domains data. For example, many works use board data to perform self-supervised pre-training, and then adapt to diverse downstream tasks through zero-shot [1, 10, 49] or fine-tuning [13, 15, 23, 58]. Additionally, some research ef" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 342, + 274, + 533, + 441 + ], + "blocks": [ + { + "bbox": [ + 342, + 274, + 533, + 441 + ], + "lines": [ + { + "bbox": [ + 342, + 274, + 533, + 441 + ], + "spans": [ + { + "bbox": [ + 342, + 274, + 533, + 441 + ], + "type": "image", + "image_path": "b20601886bd5d635928fdf1cbebeaae57f139d3d6ba8796761e8f9210cfab061.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 449, + 555, + 472 + ], + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 472 + ], + "type": "text", + "content": "Figure 1. (a) Training with small-scale, in-domain data. (b) Joint training with in-domain data and large-scale, out-of-domain data." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 482, + 556, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 556, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 556, + 615 + ], + "type": "text", + "content": "forts [43, 51, 71, 72] focus on learning a generalist model capable of unifying various tasks. In contrast, in robot learning, due to data scarcity and homogeneity, some groundbreaking methods [2, 37] resort to training only using in-domain data. Recently, some works [3, 4, 52, 61, 69] have shifted to learn from action-free human video data. A particularly scalable and promising paradigm [5, 20, 59, 63, 65] involves learning a trajectory prediction model from action-free video data, and then using small-scale action-labeled demonstrations to learn trajectory-guided policies, thereby achieving greater sample efficiency and better generalization." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "content": "Despite the potential benefits, it still remains underexplored how to effectively leverage broad out-of-domain video data for jointly training trajectory models. In particular, the impact of incorporating data encompassing diverse environments, objects, skills and embodiments has not fully investigated for in-domain tasks. For instance, ATM [63], a notable work in this field, mainly relies on in-domain data (same tasks and environments) for training both trajectory" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6960" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 182, + 204 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 182, + 204 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 182, + 204 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 182, + 204 + ], + "type": "image", + "image_path": "1116d206222c344b3495caf32192dc759e91b4b7ea155d6cd6e4cf158685afb2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 211, + 295, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 211, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 295, + 266 + ], + "type": "text", + "content": "Figure 2. Training trajectory prediction model from multiple domains. (a) The visualization of Dataset " + }, + { + "bbox": [ + 55, + 211, + 295, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\text {ood }}" + }, + { + "bbox": [ + 55, + 211, + 295, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 211, + 295, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\text {in }}" + }, + { + "bbox": [ + 55, + 211, + 295, + 266 + ], + "type": "text", + "content": ". The former may contain additional environments, objects, skills and embodiments. (b) Our pipeline: first co-training trajectory prediction model and then adapting it for downstream policy learning." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 189, + 71, + 292, + 205 + ], + "blocks": [ + { + "bbox": [ + 189, + 71, + 292, + 205 + ], + "lines": [ + { + "bbox": [ + 189, + 71, + 292, + 205 + ], + "spans": [ + { + "bbox": [ + 189, + 71, + 292, + 205 + ], + "type": "image", + "image_path": "dbc45401311683cf540e6205e3be6a9a5b51e719c61443d62b51f8072d7ecab6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 270, + 295, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 295, + 486 + ], + "type": "text", + "content": "models and policies. We argue that learning from multiple domains is the primary factor to build a powerful trajectory model for policy guidance. Therefore, we propose to train trajectory model under the setup of jointly learning from large-scale out-of-domain and small-scale in-domain data, as shown in Fig. 2(a). However, directly using large-scale out-of-domain video data for hybrid pre-training involves two critical challenges: (i) Joint training on data from different environments, objects, skills, and embodiments will add learning difficulty and might lead to optimization conflicts. It is unclear how to effectively fuse the complementary information of multiple domains data within a single unified model. Naively expanding the out-of-domain training data does not necessarily lead to improved performance on in-domain tasks, as shown in Fig. 1, which results in a 5.6 " + }, + { + "bbox": [ + 54, + 270, + 295, + 486 + ], + "type": "inline_equation", + "content": "(57.6\\to 52.0)" + }, + { + "bbox": [ + 54, + 270, + 295, + 486 + ], + "type": "text", + "content": " performance decrease on in-domain tasks. (ii) How to ensure both high performance and computational efficiency during trajectory model scaling up process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 487, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 295, + 713 + ], + "type": "text", + "content": "To address these challenges, we design a new sparsely-gated Mixture-of-Expert (MoE) architecture [53] to scale up our trajectory model, coined as Tra-MoE. Specifically, we integrate several MoE blocks to replace the original transformer blocks. This design enables efficient joint training of most parameters on the broad out-of-domain data, capturing the complementary patterns for mutual cooperation and the shared knowledge for better generalization. Meanwhile, for different data and different tokens within the same data, our Tra-MoE naturally forms a specialization when activating different experts. To maintain high computational efficiency while scaling up model capacity, we implement a top-1 gating strategy for token-choice, ensuring constant FLOPs per token. In sparse MoE, some auxiliary losses are commonly used to enhance performance, including the router z-loss [74] for improving training stability and the load-balancing loss [36] for balancing expert activations. We further conduct a comprehensive empirical study on scaling up trajectory models and find that the touter z-loss improves" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 239 + ], + "type": "text", + "content": "performance while the load-balancing loss tends to cause performance degradation. Furthermore, we observe that when expanding pre-training data from small-scale in-domain to large-scale in-domain and out-of-domain data, employing MoE (Tra-MoE) can lead to obvious performance improvement on in-domain tasks, whereas the dense counterpart (Tra-baseline) suffers from performance drop, as illustrated in Fig. 1. Notably, even when expanding the dense Tra-baseline's parameter to match that of Tra-MoE, performance improvements remain elusive. These findings collectively demonstrate the effectiveness of our sparsely-gated MoE architecture, attributing its capability of balancing parameter cooperation and specialization when expanding training with out-of-domain data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 239, + 556, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 556, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 556, + 358 + ], + "type": "text", + "content": "Furthermore, given a pre-trained trajectory model, how to effectively condition the policy module using the 2D trajectories still remain an unresolved challenge. We further propose an adaptive policy conditioning technique for trajectory-guided policy. This enables better spatial alignment between 2D trajectories and images, while achieving more flexible trajectory representation and guidance for the policy. Finally, we also conduct extensive real-world experiments to further strengthen our conclusions. In summary, our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 361, + 555, + 575 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 361, + 555, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 361, + 555, + 408 + ], + "spans": [ + { + "bbox": [ + 314, + 361, + 555, + 408 + ], + "type": "text", + "content": "- We propose a sparsely-gated MoE architecture to learn trajectory models, called as Tra-MoE, with large-scale, out-of-domain, and action-free video data, coupled with a comprehensive empirical study on its training techniques." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 409, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 409, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 314, + 409, + 555, + 468 + ], + "type": "text", + "content": "- Extensive simulation and real-world environments demonstrate that our Tra-MoE (with Top-1 gating) can more effectively benefit from large-scale out-of-domain data (such as cross-embediments data and different physics engine simulated data) compared to the dense baseline." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 468, + 555, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 468, + 555, + 528 + ], + "spans": [ + { + "bbox": [ + 314, + 468, + 555, + 528 + ], + "type": "text", + "content": "- We demonstrate that Tra-MoE still significantly outperform expanded dense counterpart with equivalent parameters, highlighting the effectiveness of sparse MoE architecture in leveraging out-of-domain data through better parameter cooperation and specialization." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 529, + 555, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 529, + 555, + 575 + ], + "spans": [ + { + "bbox": [ + 314, + 529, + 555, + 575 + ], + "type": "text", + "content": "- We propose an adaptive policy conditioning technique for trajectory-guided policy, which ensures adaptive 2D trajectory representation and explicit spatial alignment with the observations, thereby achieving superior performance." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 586, + 400, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 586, + 400, + 597 + ], + "spans": [ + { + "bbox": [ + 314, + 586, + 400, + 597 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "Learning from multiple domains in Robot Learning. Scaling laws [30] have fueled substantial advancements in various fields. In robot learning, many recent studies [7, 8, 29, 51] investigate how compute, model size, and training data quantity affect the model performance on robotic manipulation tasks. Based on this, recent works have focused on collecting larger-scale real-world robotic datasets [11, 19, 32, 60, 64], and have successfully trained several robotics foundational models [6, 9, 16, 33, 38, 47, 57," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "6961" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "[62] that demonstrate superior transfer capabilities. It is worth noting that these works mainly employ pre-training and fine-tuning paradigm rather than joint co-training paradigm, while focusing on evaluating the combination generalization and neglecting the performance impact on in-domain tasks. Additionally, these works directly combine heterogeneous robotic data with different action and observation spaces for joint training, potentially leading to sub-optimal solutions. In contrast, we leverage sparsely-gated MoE to incorporate larger-scale, out-of-domain, and action-less video data. Furthermore, our experiments demonstrate the importance of MoE for expanding out-of-domain data due to its superior capacity for parameter cooperation and specialization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 232, + 298, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 298, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 298, + 495 + ], + "type": "text", + "content": "Mixture-of-Experts. The Mixture-of-Experts (MoE) framework [12] is built upon a simple yet effective concept: dividing a model into specialized components, or 'experts', each focusing on distinct tasks or data aspects. This paradigm allows for selective activation of relevant experts based on the input, thereby maintaining computational efficiency while leveraging a vast reservoir of specialized knowledge through increased model capacity. Recently, in computer vision, natural language processing, and multi-modal, there has been a surge of foundation models trained based on the MoE architecture, including Swin-Moe [24], MoE-LLaVA [39], DeepSeek-V2 [40], Mixtral-8x22B [28], and so on. MoE typically encompasses dense [48] and sparse [14] variants. Dense MoE activates all experts in each iteration, while sparse MoE activates only a subset, generally resulting in lower computational overhead. However, sparse MoE can suffer from expert load imbalance and training instability issues. Techniques such as adding noise, load balancing loss [36], and router z-loss [74] are often employed to mitigate these issues and improve performance. In our work, we conduct a comprehensive empirical study of these techniques when training our sparsely-gated MoE-based trajectory model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 498, + 298, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 298, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 298, + 714 + ], + "type": "text", + "content": "Representations and Techniques for Policy Conditioning. Policy conditioning equips a multi-task robot policy with the ability to modulate its behavior by incorporating task-specific guidance information. This information, which may take various forms, guides the policy's action generation, enabling a single model to effectively learn and perform diverse tasks. Classic policy conditioning representations often encompass task identifiers [50], natural language instructions [8, 22, 45, 54], goal images [44], and demonstration videos [25, 27]. These are typically integrated with visual observations within the policy through concatenation or attention mechanisms. Recent works have explored alternative policy conditioning representations to enhance generalization and sample efficiency. These include leveraging internet-trained models to obtain bounding box [55] and object mask [42, 66, 68, 73], as well as learning plan [4], affordance [3, 46], future observations [17, 18, 34, 67], 2D or 3D trajectory [5, 61, 63, 65, 70] from human video data," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 556, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 168 + ], + "type": "text", + "content": "and also hand-drawn trajectory [21] or sketch [56]. Different from these approaches that directly utilize raw or manually crafted signals for integration with visual observations, we introduce an adaptive policy conditioning technique that explicitly maps 2D trajectories onto the visual observations in space and encodes them as learnable embeddings. These designs ensure adaptive 2D trajectory representation and explicit spatial alignment with the visual observations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 178, + 397, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 397, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 397, + 192 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 198, + 556, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 556, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 556, + 281 + ], + "type": "text", + "content": "In this section, we first describe our problem formulation in Sec. 3.1. Subsequently, we give an overview of our proposed framework, which is composed of a sparsely-gated MoE-based trajectory prediction model and a trajectory-guided policy model using the adaptive policy conditioning technique, detailed in Sec. 3.2. Finally, we describe our training process and loss functions in Sec. 3.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 289, + 440, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 440, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 440, + 300 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 306, + 556, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 556, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 556, + 437 + ], + "type": "text", + "content": "Following the setup of previous works [5, 63, 65], our approach encompasses a trajectory prediction model to conduct language-conditioned any-point trajectory prediction and a trajectory-guided policy model to predict executable robot actions. In particular, as depicted in Fig. 2, we consider two types of datasets for training, which are large-scale, out-of-domain, and action-free videos " + }, + { + "bbox": [ + 313, + 306, + 556, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\text{ood}}" + }, + { + "bbox": [ + 313, + 306, + 556, + 437 + ], + "type": "text", + "content": ", as well as small-scale, in-domain, and action-labeled demonstrations " + }, + { + "bbox": [ + 313, + 306, + 556, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\text{in}}" + }, + { + "bbox": [ + 313, + 306, + 556, + 437 + ], + "type": "text", + "content": ". For trajectory prediction model, we employ " + }, + { + "bbox": [ + 313, + 306, + 556, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\text{ood}}" + }, + { + "bbox": [ + 313, + 306, + 556, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 306, + 556, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\text{in}}" + }, + { + "bbox": [ + 313, + 306, + 556, + 437 + ], + "type": "text", + "content": " for joint pre-training, whereas for trajectory-guided policy model, we only use " + }, + { + "bbox": [ + 313, + 306, + 556, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\text{in}}" + }, + { + "bbox": [ + 313, + 306, + 556, + 437 + ], + "type": "text", + "content": " for training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 445, + 397, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 397, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 397, + 456 + ], + "type": "text", + "content": "3.2. Architecture" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 462, + 556, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 556, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 556, + 654 + ], + "type": "text", + "content": "ATM Revisited. In this paper, we use ATM [63] as the baseline for our method. The overall pipeline of ATM is illustrated on Fig. 2(b). It first learns a trajectory model from video data to perform language-conditioned trajectory prediction. The predicted trajectories is then used to learn a trajectory-guided policy using robot demonstrations. Specifically, as for trajectory model, its objective is to predict the future of any-point trajectories in a frame. Formally, given an image observation " + }, + { + "bbox": [ + 313, + 462, + 556, + 654 + ], + "type": "inline_equation", + "content": "o_t" + }, + { + "bbox": [ + 313, + 462, + 556, + 654 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 313, + 462, + 556, + 654 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 462, + 556, + 654 + ], + "type": "text", + "content": ", any set of 2D query points simplified on the image frame " + }, + { + "bbox": [ + 313, + 462, + 556, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_t = \\{p_{t,k}\\}_{k=1}^K" + }, + { + "bbox": [ + 313, + 462, + 556, + 654 + ], + "type": "text", + "content": ", and a language instruction " + }, + { + "bbox": [ + 313, + 462, + 556, + 654 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 313, + 462, + 556, + 654 + ], + "type": "text", + "content": ", the trajectory model learns a mapping " + }, + { + "bbox": [ + 313, + 462, + 556, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{t:t+H} = \\tau_\\theta(o_t, \\mathbf{p}_t, \\ell)" + }, + { + "bbox": [ + 313, + 462, + 556, + 654 + ], + "type": "text", + "content": " that predicts the coordinates of these query points in the camera frame for the next " + }, + { + "bbox": [ + 313, + 462, + 556, + 654 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 462, + 556, + 654 + ], + "type": "text", + "content": " time steps, where " + }, + { + "bbox": [ + 313, + 462, + 556, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{t:t+H} \\in \\mathbb{R}^{H \\times K \\times 2}" + }, + { + "bbox": [ + 313, + 462, + 556, + 654 + ], + "type": "text", + "content": ". For the trajectory-guided policy, it leverages the current observation " + }, + { + "bbox": [ + 313, + 462, + 556, + 654 + ], + "type": "inline_equation", + "content": "o_t" + }, + { + "bbox": [ + 313, + 462, + 556, + 654 + ], + "type": "text", + "content": " and the predicted trajectory " + }, + { + "bbox": [ + 313, + 462, + 556, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{t:t+H}" + }, + { + "bbox": [ + 313, + 462, + 556, + 654 + ], + "type": "text", + "content": " to predict the subsequent robot actions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "content": "Sparsely-gated MoE-based trajectory model (Tra-MoE). ATM proposes a track transformer to implement the trajectory model. In track transformer, language tokens, image tokens, and track tokens to be decoded involve global interactions within the transformer block. However, directly" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6962" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 93, + 70, + 341, + 328 + ], + "blocks": [ + { + "bbox": [ + 93, + 70, + 341, + 328 + ], + "lines": [ + { + "bbox": [ + 93, + 70, + 341, + 328 + ], + "spans": [ + { + "bbox": [ + 93, + 70, + 341, + 328 + ], + "type": "image", + "image_path": "1b360506ebbfb7a53d46babead0df0d6744f34a203836d82d0e49690726e4618.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 338, + 555, + 372 + ], + "lines": [ + { + "bbox": [ + 54, + 338, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 54, + 338, + 555, + 372 + ], + "type": "text", + "content": "Figure 3. (a) The pipeline of our sparsely-gated MoE-based trajectory model (Tra-MoE). (b) The pipeline of our trajectory-guided policy using the adaptive policy conditioning technique. Mapping means concatenating the trajectory mask with image observations, while learnable refers to setting each point in the trajectory mask as a learnable embedding." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 340, + 71, + 512, + 327 + ], + "blocks": [ + { + "bbox": [ + 340, + 71, + 512, + 327 + ], + "lines": [ + { + "bbox": [ + 340, + 71, + 512, + 327 + ], + "spans": [ + { + "bbox": [ + 340, + 71, + 512, + 327 + ], + "type": "image", + "image_path": "d1ccaa97ebefa6d3fb7b4012c22df2c42abd6eb7a4331b344e1bc0685f8f71ae.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 378, + 297, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 378, + 297, + 558 + ], + "spans": [ + { + "bbox": [ + 54, + 378, + 297, + 558 + ], + "type": "text", + "content": "processing diverse domains and multiple modalities data within a unified transformer may yield suboptimal results. Hence, we propose our Tra-MoE framework. To develop our sparsely-gated MoE-based trajectory model, as depicted in Fig. 3(a), we replace certain layers of the transformer with MoE blocks. In each MoE block, it incorporates multiple feedforward networks (FFNs), each designated as an expert, and utilizes a gating function to activate a selected subset (Top- " + }, + { + "bbox": [ + 54, + 378, + 297, + 558 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 378, + 297, + 558 + ], + "type": "text", + "content": ") of these experts. In this paper, we set " + }, + { + "bbox": [ + 54, + 378, + 297, + 558 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 378, + 297, + 558 + ], + "type": "text", + "content": " to a fixed value of 1 to ensure that the FLOPs per token remain constant relative to the baseline. Formally, the gating network " + }, + { + "bbox": [ + 54, + 378, + 297, + 558 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 54, + 378, + 297, + 558 + ], + "type": "text", + "content": ", parameterized by " + }, + { + "bbox": [ + 54, + 378, + 297, + 558 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 54, + 378, + 297, + 558 + ], + "type": "text", + "content": " and typically consisting of a linear-softmax network, yields the output " + }, + { + "bbox": [ + 54, + 378, + 297, + 558 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\mathbf{x_s};\\Theta)" + }, + { + "bbox": [ + 54, + 378, + 297, + 558 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 378, + 297, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{x_s}" + }, + { + "bbox": [ + 54, + 378, + 297, + 558 + ], + "type": "text", + "content": " represents our token sequences input. We formulate the sparsely-gated token-choice mechanism as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 591, + 297, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 591, + 297, + 605 + ], + "spans": [ + { + "bbox": [ + 67, + 591, + 297, + 605 + ], + "type": "interline_equation", + "content": "\\mathcal {G} \\left(\\mathbf {x} _ {\\mathrm {s}}; \\boldsymbol {\\Theta}\\right) _ {i} = \\operatorname {s o f t m a x} (\\operatorname {T o p - K} (g \\left(\\mathbf {x} _ {\\mathrm {s}}; \\boldsymbol {\\Theta}\\right), k)) _ {i}, \\tag {1}", + "image_path": "89922b85d479a41a8dcbcb9d4921206d4f451fc02e3a69d4b39df42a237e9653.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 609, + 296, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 609, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 69, + 609, + 296, + 639 + ], + "type": "interline_equation", + "content": "\\operatorname {T o p - K} (g \\left(\\mathbf {x} _ {\\mathbf {s}}; \\boldsymbol {\\Theta}\\right), k) _ {i} = \\left\\{ \\begin{array}{l l} g (\\mathbf {x}; \\boldsymbol {\\Theta}) _ {i}, & \\text {c o n d ,} \\\\ - \\infty , & \\text {o t h .} \\end{array} \\right., \\tag {2}", + "image_path": "986cfe69c204ad189004f53b0d485732e8853b5714b45122cfbdfb8e94970ecf.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 642, + 296, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 642, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 67, + 642, + 296, + 656 + ], + "type": "interline_equation", + "content": "\\operatorname {c o n d}: \\text {i f} g (\\mathbf {x}; \\boldsymbol {\\Theta}) _ {i} \\text {i s i n t o p - k e l e m e n t s o f} g (\\mathbf {x}; \\boldsymbol {\\Theta}). \\tag {3}", + "image_path": "67b86fd906a4d994129cbe667affbaa11f31289a5395db62a614e8d9434af318.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 297, + 715 + ], + "type": "text", + "content": "More specifically, " + }, + { + "bbox": [ + 55, + 677, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\operatorname{Top - K}(\\cdot ,k)" + }, + { + "bbox": [ + 55, + 677, + 297, + 715 + ], + "type": "text", + "content": " function retains only the top- " + }, + { + "bbox": [ + 55, + 677, + 297, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 677, + 297, + 715 + ], + "type": "text", + "content": " values in a vector, setting the rest to " + }, + { + "bbox": [ + 55, + 677, + 297, + 715 + ], + "type": "inline_equation", + "content": "-\\infty" + }, + { + "bbox": [ + 55, + 677, + 297, + 715 + ], + "type": "text", + "content": ". Consequently, the output of the sparsely-gated MoE layer can be" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 378, + 375, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 378, + 375, + 389 + ], + "spans": [ + { + "bbox": [ + 314, + 378, + 375, + 389 + ], + "type": "text", + "content": "formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 321, + 396, + 555, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 396, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 321, + 396, + 555, + 441 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\text {s p a r s e}} ^ {\\mathrm {M o E}} \\left(\\mathbf {x} _ {\\mathbf {s}}; \\boldsymbol {\\Theta}, \\left\\{\\mathbf {W} _ {i} \\right\\} _ {i = 1} ^ {K}\\right) = \\sum_ {i = 1} ^ {K} \\mathcal {G} \\left(\\mathbf {x} _ {\\mathbf {s}}; \\boldsymbol {\\Theta}\\right) _ {i} f _ {i} \\left(\\mathbf {x} _ {\\mathbf {s}}; \\mathbf {W} _ {i}\\right), \\tag {4}", + "image_path": "5063acae8555820fa3d229af29190649644cdbac847b4fe7dd9307c58c5e0547.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 450, + 555, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 474 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 450, + 555, + 474 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 313, + 450, + 555, + 474 + ], + "type": "text", + "content": " represents the i-th expert, usually a linear-GELU-linear FFN layer, is parameterized by " + }, + { + "bbox": [ + 313, + 450, + 555, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_i" + }, + { + "bbox": [ + 313, + 450, + 555, + 474 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 474, + 556, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 605 + ], + "type": "text", + "content": "Trajectory-guided policy using the adaptive policy conditioning technique. After completing the trajectory model learning, we develop our trajectory-guided policy through the adaptive policy conditioning technique based on the initial ATM, as shown in Fig. 3(b). In our policy model, independently encoded track tokens and image tokens undergo a early fusion within the fusion transformer. Subsequently, the fused features are further integrated with the raw trajectory data through a late fusion mechanism along the channel dimension, ensuring the preservation of original trajectory information for more accurate action prediction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 605, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 557, + 714 + ], + "type": "text", + "content": "Fundamentally, the policy model bridges the gap between the predicted 2D any-point trajectories and the executable 3D robot actions, functioning as an inverse dynamic model and eliminating the necessity of other task specification. However, how to effectively condition the policy using the predicted 2D trajectories remains an unresolved challenge. Based on our analysis, we present two key insights: (i) Explicit spatial alignment and fusion of the 2D trajectories with image observations turns out to be highly beneficial." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6963" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "text", + "content": "It simplifies the mapping of 2D trajectories to image space and highlights motion-relevant regions, enabling the policy model to focus on important areas and reducing the learning difficulty. (ii) Trajectory positions exhibit different characteristics: starting points typically emphasize local motion, while endpoints focus more on global trend. Therefore, an adaptive trajectory representation and conditioning technique is expected to flexibly guide the policy by effectively capturing these variations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 180, + 296, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 296, + 335 + ], + "type": "text", + "content": "Based on the aforementioned analysis, we propose an adaptive policy conditioning technique to incorporate trajectory information into 2D images. Specifically, we construct an additional mask modality, and populate it with 2D trajectories based on their specific spatial locations, while setting the value of each point in the trajectory mask as a learnable embedding. Before encoding the image, we concatenate this mask modality with the image observations on the channel dimension, forming a tensor of dimensions " + }, + { + "bbox": [ + 55, + 180, + 296, + 335 + ], + "type": "inline_equation", + "content": "H \\times W \\times 4" + }, + { + "bbox": [ + 55, + 180, + 296, + 335 + ], + "type": "text", + "content": ", as shown in Fig. 3(b). To sum up, our adaptive policy conditioning technique yield an adaptive 2D trajectory representation, which could be explicitly aligned with the image observations to guide the policy prediction more flexibly." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 341, + 119, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 119, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 119, + 354 + ], + "type": "text", + "content": "3.3. Training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 358, + 296, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 296, + 574 + ], + "type": "text", + "content": "We argue that trajectory models could be shared by multiple domains data. Therefore, we first pre-train a generalizable trajectory model on " + }, + { + "bbox": [ + 55, + 358, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\text{ood}} \\cup \\mathcal{D}_{\\text{in}}" + }, + { + "bbox": [ + 55, + 358, + 296, + 574 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 55, + 358, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\text{ood}}" + }, + { + "bbox": [ + 55, + 358, + 296, + 574 + ], + "type": "text", + "content": " typically contains additional environments, objects, skills, and embodiments. Then we train the trajectory-guided policy model on " + }, + { + "bbox": [ + 55, + 358, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\text{in}}" + }, + { + "bbox": [ + 55, + 358, + 296, + 574 + ], + "type": "text", + "content": " based on the pre-trained trajectory model. In trajectory-guided policy training phase, we freeze the trajectory model. Pre-training sparsely-gated MoE-based trajectory model. Following ATM, we train the trajectory model using the ground truth tracks generated by CoTracker [31]. We use MSE to optimize the trajectory prediction loss " + }, + { + "bbox": [ + 55, + 358, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{tra}}" + }, + { + "bbox": [ + 55, + 358, + 296, + 574 + ], + "type": "text", + "content": ". Additionally, we also consider adding an auxiliary load-balancing loss [36] to balance the activation frequency among experts. Given a batch of queries " + }, + { + "bbox": [ + 55, + 358, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 55, + 358, + 296, + 574 + ], + "type": "text", + "content": ", it contains B samples and each sample contains S tokens. We define " + }, + { + "bbox": [ + 55, + 358, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_i" + }, + { + "bbox": [ + 55, + 358, + 296, + 574 + ], + "type": "text", + "content": " as the proportion of tokens distributed to expert " + }, + { + "bbox": [ + 55, + 358, + 296, + 574 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 358, + 296, + 574 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 358, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 55, + 358, + 296, + 574 + ], + "type": "text", + "content": " as the proportion of the gating probability assigned to expert " + }, + { + "bbox": [ + 55, + 358, + 296, + 574 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 358, + 296, + 574 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 55, + 358, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{lo-ba}}" + }, + { + "bbox": [ + 55, + 358, + 296, + 574 + ], + "type": "text", + "content": " can be formulated as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 83, + 588, + 296, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 588, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 83, + 588, + 296, + 617 + ], + "type": "interline_equation", + "content": "\\mathcal {Q} _ {i} = \\frac {1}{S \\cdot B} \\sum_ {x _ {s} ^ {m} \\in \\mathcal {B}} \\mathbb {1} \\left\\{\\operatorname {a r g m a x} \\mathcal {G} \\left(\\mathbf {x} _ {\\mathbf {s}} ^ {\\mathbf {m}}; \\boldsymbol {\\Theta}\\right) = i \\right\\}, \\tag {5}", + "image_path": "e41efa0aa5a5900120eb68ae1295e4195dbd5e092e03e3ad65dd267347a85e55.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 82, + 620, + 296, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 620, + 296, + 649 + ], + "spans": [ + { + "bbox": [ + 82, + 620, + 296, + 649 + ], + "type": "interline_equation", + "content": "\\mathcal {P} _ {i} = \\frac {1}{S \\cdot B} \\sum_ {x _ {\\mathrm {s}} ^ {m} \\in \\mathcal {B}} \\mathcal {G} \\left(\\mathbf {x} _ {\\mathrm {s}} ^ {\\mathrm {m}}; \\boldsymbol {\\Theta}\\right) _ {i}, \\tag {6}", + "image_path": "c9ccf59ae9c7ebab79ef67aa8ca1b641f8e10ba9ce46d85196058efea39d8c1a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 82, + 653, + 296, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 653, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 82, + 653, + 296, + 685 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {l o - b a}} = N \\sum_ {i = 1} ^ {N} \\mathcal {Q} _ {i} \\mathcal {P} _ {i}. \\tag {7}", + "image_path": "cb7768c0cb4fa999a29c5f95a861407b73bbe90d1a260c060bfc1f303e673ea9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Additionally, ST-MoE [74] proposed using the router " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "-loss to penalize large logits entering the gating network," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "thereby improving training stability. We similarly applies it to the training of our trajectory model, as shown below:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 363, + 105, + 555, + 140 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 105, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 363, + 105, + 555, + 140 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {z} (x) = \\frac {1}{S} \\sum_ {k = 1} ^ {S} \\left(\\log \\sum_ {i = 1} ^ {N} e ^ {\\mathbf {g} _ {i} ^ {(k)}}\\right) ^ {2}, \\tag {8}", + "image_path": "0c1dac1303f509fae0338e930ae82cec916b3faae4aef104618fd078e2d8ecef.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 144, + 555, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 182 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 182 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 144, + 555, + 182 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 144, + 555, + 182 + ], + "type": "text", + "content": " is the total number of tokens, " + }, + { + "bbox": [ + 313, + 144, + 555, + 182 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 144, + 555, + 182 + ], + "type": "text", + "content": " is the total number of experts, and " + }, + { + "bbox": [ + 313, + 144, + 555, + 182 + ], + "type": "inline_equation", + "content": "\\mathbf{g} \\in \\mathcal{R}^{S \\times N}" + }, + { + "bbox": [ + 313, + 144, + 555, + 182 + ], + "type": "text", + "content": " are the logits going into the router. In summary, the trajectory model's training loss is as follow:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 339, + 203, + 555, + 216 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 203, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 339, + 203, + 555, + 216 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {t r a}} \\cdot \\mathcal {L} _ {\\text {t r a}} + \\lambda_ {\\text {l o - b a}} \\cdot \\mathcal {L} _ {\\text {l o - b a}} + \\lambda_ {z} \\cdot \\mathcal {L} _ {z}. \\tag {9}", + "image_path": "3bdca18344fe7b4953e65454cbd096b06b67954a95b62c7423588873f1835bcd.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 225, + 555, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 225, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 555, + 285 + ], + "type": "text", + "content": "Training trajectory-guided policy through adaptive policy conditioning technique. For trajectory-guided policy training, we employ the paradigm of behavior cloning. Specifically, we use the MSE between predicted actions and ground truths as the loss function." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 297, + 395, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 297, + 395, + 310 + ], + "spans": [ + { + "bbox": [ + 314, + 297, + 395, + 310 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 316, + 555, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 436 + ], + "type": "text", + "content": "We perform evaluation experiments using the LIBERO benchmark [41] and a real low-cost dual-arm robot [35]. In experiments, we quantitatively find that the success rate of downstream manipulation tasks is generally positively correlated with the performance of the trajectory model. Detailed quantitative experiment results can be found in the appendix. Meanwhile, CoTracker [31] can provide relatively accurate trajectory annotations; therefore, we report the average success rate of downstream manipulation tasks as our metric. Our experiments aim to study the following questions:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 437, + 555, + 580 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 313, + 437, + 553, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 553, + 472 + ], + "type": "text", + "content": "Q1: Does the techniques of router z-loss, load-balancing loss, and adding noise to gating logits enhance the training of sparsely-gated Tra-MoE?" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 472, + 555, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 520 + ], + "type": "text", + "content": "Q2: Does sparsely-gated Tra-MoE, compared to the dense Tra-baseline, benefit more from large-scale out-of-domain data (such as cross-embediments and different physics engine data) and demonstrate superior scaling up capabilities?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 521, + 553, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 553, + 555 + ], + "type": "text", + "content": "Q3: When we scale up the dense Tra-baseline to match the parameter count of Tra-MoE, does the sparsely-gated Tra-MoE still maintain its advantage?" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 556, + 553, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 553, + 580 + ], + "type": "text", + "content": "Q4: Is our proposed adaptive policy conditioning technique for trajectory-guided policy more effective than ATM?" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 588, + 451, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 451, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 451, + 601 + ], + "type": "text", + "content": "4.1. Simulation Experiments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "Experiment setup. The LIBERO [41] is divided into five categories: LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, LIBERO-Long, and LIBERO-90. LIBERO-90 includes 90 tasks, while each of the other categories contains 10 tasks. Compared to other categories, LIBERO-90 presents a highly varied selection of object categories, layouts, environments, and skill goals. It is typically used for pre-training rather than joint co-training. Therefore, our main ablation experiments select 90 tasks from LIBERO-90 to construct" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6964" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 70, + 289, + 150 + ], + "blocks": [ + { + "bbox": [ + 71, + 70, + 289, + 150 + ], + "lines": [ + { + "bbox": [ + 71, + 70, + 289, + 150 + ], + "spans": [ + { + "bbox": [ + 71, + 70, + 289, + 150 + ], + "type": "table", + "html": "
SpatialGoalObjectLongAvg.
062.073.071.021.556.9
5·10-469.573.568.022.558.4
1·10-462.581.073.528.561.4
1·10-560.065.568.527.055.3
1·10-660.079.053.031.555.9
", + "image_path": "9504d9d45d6a4b0a101bfc2bfb7656c82744da79452a9b4ebd7335a1fd79c306.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 63, + 166, + 293, + 211 + ], + "blocks": [ + { + "bbox": [ + 145, + 152, + 211, + 162 + ], + "lines": [ + { + "bbox": [ + 145, + 152, + 211, + 162 + ], + "spans": [ + { + "bbox": [ + 145, + 152, + 211, + 162 + ], + "type": "text", + "content": "(a) The value of " + }, + { + "bbox": [ + 145, + 152, + 211, + 162 + ], + "type": "inline_equation", + "content": "\\lambda_z" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 166, + 293, + 211 + ], + "lines": [ + { + "bbox": [ + 63, + 166, + 293, + 211 + ], + "spans": [ + { + "bbox": [ + 63, + 166, + 293, + 211 + ], + "type": "table", + "html": "
SpatialGoalObjectLongAvg.
w/o. noise62.073.071.022.556.9
w. noise71.567.057.527.055.8
", + "image_path": "0f0712d7bceb9eb4623303c792b6ec47f95513fa514dcad77622b0c57072fcfd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 72, + 238, + 282, + 306 + ], + "blocks": [ + { + "bbox": [ + 107, + 213, + 249, + 222 + ], + "lines": [ + { + "bbox": [ + 107, + 213, + 249, + 222 + ], + "spans": [ + { + "bbox": [ + 107, + 213, + 249, + 222 + ], + "type": "text", + "content": "(c) Adding a noise term to the gating logits." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 238, + 282, + 306 + ], + "lines": [ + { + "bbox": [ + 72, + 238, + 282, + 306 + ], + "spans": [ + { + "bbox": [ + 72, + 238, + 282, + 306 + ], + "type": "table", + "html": "
Num.SpatialGoalObjectLongAvg.
149.567.056.535.052.0
259.072.565.527.056.0
364.068.071.017.555.1
462.073.071.021.556.9
", + "image_path": "03a497c14e81d85247113fd03473c7b66e52786ef05f766549381404274a1fcc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 322, + 70, + 538, + 138 + ], + "blocks": [ + { + "bbox": [ + 134, + 308, + 222, + 318 + ], + "lines": [ + { + "bbox": [ + 134, + 308, + 222, + 318 + ], + "spans": [ + { + "bbox": [ + 134, + 308, + 222, + 318 + ], + "type": "text", + "content": "(e) The number of experts." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 70, + 538, + 138 + ], + "lines": [ + { + "bbox": [ + 322, + 70, + 538, + 138 + ], + "spans": [ + { + "bbox": [ + 322, + 70, + 538, + 138 + ], + "type": "table", + "html": "
SpatialGoalObjectLongAvg.
062.073.071.022.556.9
1·10-358.574.060.018.052.6
1·10-561.069.556.026.053.1
1·10-766.567.073.522.057.3
", + "image_path": "750356bdb0dea26187e847fb2fc535848a12c8434edd50e8202d3ed706bd395d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 324, + 166, + 536, + 222 + ], + "blocks": [ + { + "bbox": [ + 390, + 140, + 471, + 150 + ], + "lines": [ + { + "bbox": [ + 390, + 140, + 471, + 150 + ], + "spans": [ + { + "bbox": [ + 390, + 140, + 471, + 150 + ], + "type": "text", + "content": "(b) The value of " + }, + { + "bbox": [ + 390, + 140, + 471, + 150 + ], + "type": "inline_equation", + "content": "\\lambda_{lo - ba}" + }, + { + "bbox": [ + 390, + 140, + 471, + 150 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 166, + 536, + 222 + ], + "lines": [ + { + "bbox": [ + 324, + 166, + 536, + 222 + ], + "spans": [ + { + "bbox": [ + 324, + 166, + 536, + 222 + ], + "type": "table", + "html": "
SpatialGoalObjectLongAvg.
MoE62.581.073.528.561.4
Width66.550.561.515.048.4
Depth61.575.546.526.552.5
", + "image_path": "17bdff4583247d83ff735f2dd80fe5f356987ff135fffb08860a3e448251094e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 306, + 238, + 548, + 306 + ], + "blocks": [ + { + "bbox": [ + 354, + 224, + 507, + 234 + ], + "lines": [ + { + "bbox": [ + 354, + 224, + 507, + 234 + ], + "spans": [ + { + "bbox": [ + 354, + 224, + 507, + 234 + ], + "type": "text", + "content": "(d) Scaling Tra-baseline from width and depth." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 238, + 548, + 306 + ], + "lines": [ + { + "bbox": [ + 306, + 238, + 548, + 306 + ], + "spans": [ + { + "bbox": [ + 306, + 238, + 548, + 306 + ], + "type": "table", + "html": "
MoEOODSpatialGoalObjectLongAvg.
67.568.568.026.557.6
✓63.064.060.519.551.8
✓49.567.056.535.052.0
✓✓62.581.073.528.561.4
", + "image_path": "6139101233449309a3573c9cefcb724605042ca4863f91c124222887c560fdb4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 342, + 297, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 342, + 297, + 546 + ], + "spans": [ + { + "bbox": [ + 54, + 342, + 297, + 546 + ], + "type": "text", + "content": "dataset " + }, + { + "bbox": [ + 54, + 342, + 297, + 546 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\text{ood}}" + }, + { + "bbox": [ + 54, + 342, + 297, + 546 + ], + "type": "text", + "content": ", utilizing 20 action-free videos per task; and choose 40 tasks from LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, and LIBERO-Long to build dataset " + }, + { + "bbox": [ + 54, + 342, + 297, + 546 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\text{in}}" + }, + { + "bbox": [ + 54, + 342, + 297, + 546 + ], + "type": "text", + "content": ", using 10 action-labeled demonstrations per task. In summary, for training Tra-MoE, our main ablation experiments use a total of 2, 200 action-free videos, with an out-of-domain to in-domain data ratio of " + }, + { + "bbox": [ + 54, + 342, + 297, + 546 + ], + "type": "inline_equation", + "content": "9:2" + }, + { + "bbox": [ + 54, + 342, + 297, + 546 + ], + "type": "text", + "content": ". Additionally, on this basis, we also further select 92 tasks from the RLbench [26] (using CoppeliaSim rather than MuJoCo as physics engine) simulation environment and generate 5 videos for each task, totaling 2, 660 videos in conjunction with the LIBERO data, to train Tra-MoE. The different physics engines and dynamics properties further increase the data distribution discrepancy in " + }, + { + "bbox": [ + 54, + 342, + 297, + 546 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\text{ood}}" + }, + { + "bbox": [ + 54, + 342, + 297, + 546 + ], + "type": "text", + "content": ". Finally, as for the trajectory-guided policy, we train a separate multi-task policy for each category. Each policy undergoes behavioral cloning training with 100 action-labeled demonstrations from 10 tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 54, + 550, + 295, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 550, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 54, + 550, + 295, + 658 + ], + "type": "text", + "content": "Implement details. Building on ATM [63], we develop our sparsely-gated Tra-MoE and adaptive policy conditioning technique. Specifically, when using only LIBERO data, we replace the 1th, 5th, and 8th layers of the track transformer with MoE blocks. When further incorporating RLbench data, we additionally replace the 2th and 7th layers. Unless otherwise specified, we set the number of experts to 4 by default and use a top-1 gating strategy for token choice to ensure the FLOPs per token remain constant." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 662, + 295, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 662, + 295, + 698 + ], + "spans": [ + { + "bbox": [ + 55, + 662, + 295, + 698 + ], + "type": "text", + "content": "Experiment results and analysis. We present empirical simulation results to address the aforementioned research questions and make the following findings:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 701, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 712 + ], + "type": "text", + "content": "Result 1: In Tab. 1a, we examine the effect of the router" + } + ] + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 321, + 341, + 547, + 386 + ], + "blocks": [ + { + "bbox": [ + 368, + 308, + 488, + 317 + ], + "lines": [ + { + "bbox": [ + 368, + 308, + 488, + 317 + ], + "spans": [ + { + "bbox": [ + 368, + 308, + 488, + 317 + ], + "type": "text", + "content": "(f) Scaling with MoE and OOD data." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 97, + 319, + 511, + 330 + ], + "lines": [ + { + "bbox": [ + 97, + 319, + 511, + 330 + ], + "spans": [ + { + "bbox": [ + 97, + 319, + 511, + 330 + ], + "type": "text", + "content": "Table 1. The ablation experiments of the sparsely-gated Tra-MoE. All baselines are the track transformer of ATM." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 341, + 547, + 386 + ], + "lines": [ + { + "bbox": [ + 321, + 341, + 547, + 386 + ], + "spans": [ + { + "bbox": [ + 321, + 341, + 547, + 386 + ], + "type": "table", + "html": "
SpatialGoalObjectLongAvg.
ATM51.042.055.519.041.9
Tra-MoE54.062.061.041.054.5
", + "image_path": "694232bc2cfc422c4abc74e342ff3c257c79792d291197b661c0015ed70541ee.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 394, + 553, + 417 + ], + "lines": [ + { + "bbox": [ + 313, + 394, + 553, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 553, + 417 + ], + "type": "text", + "content": "Table 2. The comparison results when training trajectory model using RLbench video data." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 423, + 555, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 532 + ], + "type": "text", + "content": "z-loss weight " + }, + { + "bbox": [ + 313, + 423, + 555, + 532 + ], + "type": "inline_equation", + "content": "\\lambda_z" + }, + { + "bbox": [ + 313, + 423, + 555, + 532 + ], + "type": "text", + "content": ". The results indicate that a higher " + }, + { + "bbox": [ + 313, + 423, + 555, + 532 + ], + "type": "inline_equation", + "content": "\\lambda_z" + }, + { + "bbox": [ + 313, + 423, + 555, + 532 + ], + "type": "text", + "content": " can effectively enhance performance, with an improvement of 4.5 (61.4 vs. 56.9) at " + }, + { + "bbox": [ + 313, + 423, + 555, + 532 + ], + "type": "inline_equation", + "content": "\\lambda_z = 1 \\cdot 10^{-4}" + }, + { + "bbox": [ + 313, + 423, + 555, + 532 + ], + "type": "text", + "content": ". In Tab. 1b, we study the impact of the load-balancing loss weight " + }, + { + "bbox": [ + 313, + 423, + 555, + 532 + ], + "type": "inline_equation", + "content": "\\lambda_{lo - ba}" + }, + { + "bbox": [ + 313, + 423, + 555, + 532 + ], + "type": "text", + "content": ". The results show a significant performance decline with larger values of " + }, + { + "bbox": [ + 313, + 423, + 555, + 532 + ], + "type": "inline_equation", + "content": "\\lambda_{lo - ba}" + }, + { + "bbox": [ + 313, + 423, + 555, + 532 + ], + "type": "text", + "content": ", with only a slight improvement (" + }, + { + "bbox": [ + 313, + 423, + 555, + 532 + ], + "type": "inline_equation", + "content": "56.7 \\rightarrow 57.3" + }, + { + "bbox": [ + 313, + 423, + 555, + 532 + ], + "type": "text", + "content": ") at " + }, + { + "bbox": [ + 313, + 423, + 555, + 532 + ], + "type": "inline_equation", + "content": "\\lambda_{lo - ba} = 1 \\cdot 10^{-7}" + }, + { + "bbox": [ + 313, + 423, + 555, + 532 + ], + "type": "text", + "content": ". In Tab. 1c, we verify the effectiveness of adding noise to the gating logits, which results in a performance decrease of by 1.1 (" + }, + { + "bbox": [ + 313, + 423, + 555, + 532 + ], + "type": "inline_equation", + "content": "56.9 \\rightarrow 55.8" + }, + { + "bbox": [ + 313, + 423, + 555, + 532 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 533, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 700 + ], + "type": "text", + "content": "Finding 1: Based on the above results, we find that using router z-loss to penalize large logits entering the gating network can effectively improve training stability and lead to performance gains. Meanwhile, the load-balancing loss used to balance the activation frequency among experts and the addition of noise to enhance expert allocation exploration cannot bring effective performance improvements. An intuitive explanation is that when the data distribution is uneven, the load-balancing loss might force experts to set shared parameters on data with a large domain gap, which could lead to mutual counteraction of learning gradients. This undermines the advantage of parameter specialization in sparsely-gated MoE. Similarly, adding noise is also detrimental to learning parameter specialization among experts." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "type": "text", + "content": "Result 2: We comprehensively investigate the impact of in" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6965" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 282, + 127 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 282, + 127 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 282, + 127 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 282, + 127 + ], + "type": "table", + "html": "
SpatialGoalObjectLongAvg.
ATM62.581.073.528.561.4
+ DM69.058.085.533.561.5
+ AM69.577.088.030.566.3
", + "image_path": "32262f4e4a8ee1cb12f5562878940f70d642710c03559def97db4bef228462b5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 136, + 295, + 169 + ], + "lines": [ + { + "bbox": [ + 55, + 136, + 295, + 169 + ], + "spans": [ + { + "bbox": [ + 55, + 136, + 295, + 169 + ], + "type": "text", + "content": "Table 3. The ablation experiments of the adaptive policy conditioning technique. DM (RT-Trajectory style) and AM respectively represent hand-drawn mask and adaptive mask." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 56, + 175, + 296, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 296, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 296, + 425 + ], + "type": "text", + "content": "producing sparsely-gated MoE and large-scale out-of-domain data on scaling up trajectory models in Tab. 1f. Firstly, when training only with in-domain data, Tra-baseline achieves an average success rate of 57.6, while Tra-MoE only achieves an average success rate of 51.8. Secondly, when adding out-of-domain data to jointly train Tra-baseline, the average success rate drops to " + }, + { + "bbox": [ + 56, + 175, + 296, + 425 + ], + "type": "inline_equation", + "content": "52.0" + }, + { + "bbox": [ + 56, + 175, + 296, + 425 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 56, + 175, + 296, + 425 + ], + "type": "inline_equation", + "content": "57.6 \\rightarrow 52.0" + }, + { + "bbox": [ + 56, + 175, + 296, + 425 + ], + "type": "text", + "content": "), with particularly significant performance declines in LIBERO-Spatial and LIBERO-Object. However, when adding out-of-domain data to jointly train Tra-MoE, the average success rate increases to " + }, + { + "bbox": [ + 56, + 175, + 296, + 425 + ], + "type": "inline_equation", + "content": "61.4" + }, + { + "bbox": [ + 56, + 175, + 296, + 425 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 56, + 175, + 296, + 425 + ], + "type": "inline_equation", + "content": "51.8 \\rightarrow 61.4" + }, + { + "bbox": [ + 56, + 175, + 296, + 425 + ], + "type": "text", + "content": "). Additionally, in Tab. 1e, we also explore the impact of the number of experts in sparsely-gated MoE. The results show that using just two experts can lead to an increase of 4.0 (52.0 vs. 56.0) in the average success rate. With additional experts, performance generally trends upward, albeit with some fluctuations. Finally, we report the comparison results of training trajectory model by further using RLbench data in Tab. 2. The results indicate that although the introduction of RLbench data led to a performance drop in LIBERO, our Tra-MoE still outperform the baseline by 12.6 (54.5 vs. 41.9)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 426, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 297, + 714 + ], + "type": "text", + "content": "Finding 2: Based on the above results, we have the following findings and analysis: (i) When the data scale is small, naively expanding the model capacity through sparsely-gated MoE does not lead to performance improvements. A explanation is that the larger model may lead to a overfitting problem when training on small-scale data. Moreover, small-scale data often leads to some experts not being sufficiently trained, thereby affecting performance. (ii) Naively introducing out-of-domain data for joint training often leads to performance degradation in the target domain. However, simultaneously incorporating out-of-domain data and expanding models with sparsely-gated MoE may improve performance or alleviate performance degradation. This phenomenon suggests that although out-of-domain data generally enhances model generalization, it could hinder task-specific performance in the target domain due to insufficient in-domain data learning. In contrast, expanding the model with out-of-domain data using sparsely-gated MoE achieves a more optimal balance between parameter cooperation and specialization across diverse data and even different tokens within the same data input. This design allows these large-scale data to jointly train most of parameters, ensuring the capture of general patterns for mutual cooperation. (iii) Even with a few experts, Tra-MoE exhibits significant scaling up" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "type": "text", + "content": "capabilities. Additionally, further increasing the number of experts generally leads to performance improvement. However, this improvement is not guaranteed due to factors like data-expert complexity matching and training stability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 122, + 556, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 122, + 556, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 122, + 556, + 217 + ], + "type": "text", + "content": "Result 3: In Tab. 1d, we also investigate the performance comparison between the dense Tra-baseline and the sparse Tra-MoE with the same number of parameters. Specifically, we expand Tra-baseline to match the model capacity of Tra-MoE by increasing its model width or depth separately. These expanded versions achieve average success rates of 48.4 and 52.5 respectively, which are significantly lower than Tra-MoE's 61.4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 219, + 556, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 219, + 556, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 556, + 316 + ], + "type": "text", + "content": "Finding 3: Based on these results, we are surprised to find that simply expanding the dense Tra-baseline in depth and width can not effectively improve performance. This highlights the importance of the sparsely-gated MoE architecture when training with large-scale out-of-domain data, as it ensures dynamic activation of different experts based on different data input and different tokens within the same data input, significantly reducing optimization conflict problems." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 317, + 556, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 556, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 556, + 555 + ], + "type": "text", + "content": "Result 4: In Tab. 3, we investigate the effectiveness of the adaptive policy conditioning. Specifically, building upon ATM, we first introduce an additional hand-drawn mask modality before image encoding. In this mask, the first half of the trajectory is set to 128, while the second half is set to 255. We call it 'ATM + hand-drawn mask' and this is similar to RT-Trajectory [21]. Next, we replace the hand-drawn mask with an adaptive mask, where each point on the trajectory is instantiated as a learnable embedding. We call it 'ATM + adaptive mask'. The results show that 'ATM + hand-drawn mask' achieves only a 0.1 improvement compared to the baseline. However, except for a significant performance decrease in LIBERO-Goal, it achieves performance improvements of 6.5, 12.0, and 5.0 in LIBERO-Spatial, LIBERO-Object, and LIBERO-Long respectively. In addition, the 'ATM + adaptive mask' achieves a performance improvement of 4.9 (66.3 vs. 61.4) compared to the baseline. It further improves performance across all suites compared to 'ATM + hand-drawn mask', especially achieving a 19.0 performance gains in LIBERO-Goal (77.0 vs. 58.0)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 558, + 557, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 557, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 557, + 713 + ], + "type": "text", + "content": "Finding 4: Based on these results, we find that explicitly mapping 2D trajectories to visual observations significantly improves the performance of LIBERO-Spatial, LIBERO-Object, and LIBERO-Long, but causes a notable decrease in LIBERO-Goal. By instantiating the points in the 2D trajectory as learnable embeddings, we not only further enhances the performance of LIBERO-Spatial, LIBERO-Object, and LIBERO-Long, but also significantly improves LIBERO-Goal. We analyze that the decline in LIBERO-Goal is mainly due to overlapping front and back parts of some trajectories in the 2D images, leading to mapping errors. However, our adaptive trajectory representation and conditioning technique effectively mitigates this issue. In summary, these" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6966" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 71, + 533, + 139 + ], + "blocks": [ + { + "bbox": [ + 77, + 71, + 533, + 139 + ], + "lines": [ + { + "bbox": [ + 77, + 71, + 533, + 139 + ], + "spans": [ + { + "bbox": [ + 77, + 71, + 533, + 139 + ], + "type": "table", + "html": "
Human DataMoEAdaptive Trajectory MaskPourPushPick and PassTissueFoldAverage
40.045.050.030.025.038.0
✓45.035.050.025.035.038.0
✓✓40.050.060.035.045.046.0
✓✓✓60.070.065.035.050.056.0
", + "image_path": "bd6dc923dc11f3b6cfa2edcdc0fc56ce008272cd16430bccf28e46cbcbea5b57.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 73, + 148, + 534, + 158 + ], + "lines": [ + { + "bbox": [ + 73, + 148, + 534, + 158 + ], + "spans": [ + { + "bbox": [ + 73, + 148, + 534, + 158 + ], + "type": "text", + "content": "Table 4. The ablation of ood human data, MoE architecture, and adaptive policy conditioning technique on our real-world tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 95, + 162, + 511, + 316 + ], + "blocks": [ + { + "bbox": [ + 95, + 162, + 511, + 316 + ], + "lines": [ + { + "bbox": [ + 95, + 162, + 511, + 316 + ], + "spans": [ + { + "bbox": [ + 95, + 162, + 511, + 316 + ], + "type": "image", + "image_path": "84503aacec77ac5c54356f89d97bcde412c1ed4dba170d2a1d766ebf2aa969d0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 327, + 531, + 338 + ], + "lines": [ + { + "bbox": [ + 77, + 327, + 531, + 338 + ], + "spans": [ + { + "bbox": [ + 77, + 327, + 531, + 338 + ], + "type": "text", + "content": "Figure 4. Left: The real-world experiments hardware platform setup. Right: The real-world tasks evaluation demonstrations." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 296, + 381 + ], + "type": "text", + "content": "experiments collectively demonstrate the effectiveness of explicitly aligning 2D trajectories with images in space as well as applying the adaptive conditioning technique." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 392, + 195, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 195, + 405 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 195, + 405 + ], + "type": "text", + "content": "4.2. Real-World Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 411, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 296, + 555 + ], + "type": "text", + "content": "Experiment setup. As shown in Fig 4, we utilize a dual-arm robot [35] to conduct real-world experiments. We evaluate 5 tasks, which include two single-arm tasks: pouring water and pushing a vegetable to the side of the cutting board, as well as three dual-arm tasks: pulling out tissues, folding towels, and picking up and passing a holder. Specifically, for each task, we collect 50 action-labeled robot demonstrations. Concurrently, we also collect 50 videos of human (cross-embediments) performing for each task. Compared to the robot data, the human video encompasses more environment variations and distractor objects. Please see the appendix for more detailed evaluation details and real-world setup." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 556, + 296, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 296, + 676 + ], + "type": "text", + "content": "Implement details. Following our simulation experiments, we utilize two camera views and uniformly resize their resolution to " + }, + { + "bbox": [ + 55, + 556, + 296, + 676 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 55, + 556, + 296, + 676 + ], + "type": "text", + "content": " for trajectory model and policy training. We train our real-world Tra-MoE on borh human video and robot video data, with a separate policy trained for each task. Therefore, our real-world Tra-MoE integrates data across different embodiments (human, single-arm robot, and dual-arm robot). They have significantly different visual appearances and dynamics properties. Finally, we also present some real-world task evaluation demonstrations in Fig. 4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Experiment results and analysis. To further strengthen our conclusions that sparsely-gated MoE can better learn trajectory model from out-of-domain data and our adaptive policy" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 345, + 555, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 555, + 488 + ], + "type": "text", + "content": "condition technique could lead to more flexible guidance for policy learning, we conduct real-world ablation experiments, as shown in Tab. 4. The results indicate that simply introducing human video data cannot lead to performance improvement. However, when expanding the model by MoE architecture, the performance increases from " + }, + { + "bbox": [ + 313, + 345, + 555, + 488 + ], + "type": "inline_equation", + "content": "38\\%" + }, + { + "bbox": [ + 313, + 345, + 555, + 488 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 345, + 555, + 488 + ], + "type": "inline_equation", + "content": "46\\%" + }, + { + "bbox": [ + 313, + 345, + 555, + 488 + ], + "type": "text", + "content": ". We speculate this may be due to the significant differences in appearance and dynamics between human and robot, with the former typically moving much faster and more flexibly. Finally, when further incorporating an adaptive trajectory mask for policy conditioning, the performance is further increased from " + }, + { + "bbox": [ + 313, + 345, + 555, + 488 + ], + "type": "inline_equation", + "content": "46\\%" + }, + { + "bbox": [ + 313, + 345, + 555, + 488 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 345, + 555, + 488 + ], + "type": "inline_equation", + "content": "56\\%" + }, + { + "bbox": [ + 313, + 345, + 555, + 488 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 491, + 388, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 491, + 388, + 502 + ], + "spans": [ + { + "bbox": [ + 314, + 491, + 388, + 502 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": "In this work, we have introduced Tra-MoE, a novel framework that utilizes a sparsely-gated MoE architecture with Top-1 gating strategy to learn trajectory model from multiple domains. Tra-MoE achieves superior parameter cooperation and specialization, enabling more effective utilization of large-scale out-of-domain data while maintaining constant FLOPs per token. We train our Tra-MoE and conduct a comprehensive empirical study, demonstrating that TraMoE consistently exhibits superior performance compared to the dense baseline, even when the latter is scaled to match Tra-MoE's parameter count. Additionally, we propose an adaptive policy conditioning technique for trajectory-guided policy, ensuring adaptive 2D trajectory representation and explicit spatial alignment with observations, thereby achieving superior performance. Extensive experiments in both simulated and real-world environments validate our findings and demonstrate the effectiveness of our proposed methods." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6967" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 141, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 141, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 141, + 84 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 84, + 297, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 84, + 297, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 84, + 297, + 145 + ], + "type": "text", + "content": "This work is supported by the National Key R&D Program of China (No. 2022ZD0160900), Jiangsu Frontier Technology Research and Development Program (No. BF2024076), and the Collaborative Innovation Center of Novel Software Technology and Industrialization." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 115, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 115, + 165 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 115, + 165 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 172, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 172, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 172, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 61, + 172, + 296, + 217 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 218, + 296, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 218, + 296, + 271 + ], + "spans": [ + { + "bbox": [ + 61, + 218, + 296, + 271 + ], + "type": "text", + "content": "[2] OpenAI: Marcin Andrychowicz, Bowen Baker, Maciek Chogiej, Rafal Jozefowicz, Bob McGrew, Jakub Pachocki, Arthur Petron, Matthias Plappert, Glenn Powell, Alex Ray, et al. Learning dexterous in-hand manipulation. *IJRR*, 39(1): 3-20, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 274, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 296, + 316 + ], + "type": "text", + "content": "[3] Shikhar Bahl, Russell Mendonca, Lili Chen, Unnat Jain, and Deepak Pathak. Affordances from human videos as a versatile representation for robotics. In CVPR, pages 13778-13790, 2023. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 319, + 296, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 296, + 362 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 296, + 362 + ], + "type": "text", + "content": "[4] Homanga Bharadhwaj, Abhinav Gupta, Vikash Kumar, and Shubham Tulsiani. Towards generalizable zero-shot manipulation via translating human interaction plans. In ICRA, pages 6904-6911. IEEE, 2024. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 365, + 296, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 296, + 408 + ], + "type": "text", + "content": "[5] Homanga Bharadhwaj, Roozbeh Mottaghi, Abhinav Gupta, and Shubham Tulsiani. Track2act: Predicting point tracks from internet videos enables diverse zero-shot robot manipulation. arXiv preprint arXiv:2405.01527, 2024. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 410, + 296, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 410, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 410, + 296, + 464 + ], + "type": "text", + "content": "[6] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 466, + 296, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 466, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 466, + 296, + 520 + ], + "type": "text", + "content": "[7] Konstantinos Bousmalis, Giulia Vezzani, Dushyant Rao, Coline Manon Devin, Alex X Lee, Maria Bauza Villalonga, Todor Davchev, Yuxiang Zhou, Agrim Gupta, Akhil Raju, et al. Robocat: A self-improving generalist agent for robotic manipulation. TMLR, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 523, + 296, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 523, + 296, + 577 + ], + "spans": [ + { + "bbox": [ + 62, + 523, + 296, + 577 + ], + "type": "text", + "content": "[8] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 579, + 296, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 579, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 62, + 579, + 296, + 633 + ], + "type": "text", + "content": "[9] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 296, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 296, + 657 + ], + "type": "text", + "content": "[10] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "text", + "content": "[11] Qingwen Bu, Jisong Cai, Li Chen, Xiuqi Cui, Yan Ding, Siyuan Feng, Shenyuan Gao, Xindong He, Xu Huang, Shu Jiang, et al. Agibot world colosseo: A large-scale manipulation platform for scalable and intelligent embodied systems. arXiv preprint arXiv:2503.06669, 2025. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "text", + "content": "[12] Weilin Cai, Juyong Jiang, Fan Wang, Jing Tang, Sunghun Kim, and Jiayi Huang. A survey on mixture of experts. arXiv preprint arXiv:2407.06204, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 107, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 555, + 139 + ], + "type": "text", + "content": "[13] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, pages 9640-9649, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 140, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 555, + 183 + ], + "type": "text", + "content": "[14] Zitian Chen, Yikang Shen, Mingyu Ding, Zhenfang Chen, Hengshuang Zhao, Erik G Learned-Miller, and Chuang Gan. Mod-squad: Designing mixtures of experts as modular multitask learners. In CVPR, pages 11828-11837, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 185, + 554, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 554, + 216 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 554, + 216 + ], + "type": "text", + "content": "[15] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 217, + 555, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 555, + 258 + ], + "type": "text", + "content": "[16] Ria Doshi, Homer Rich Walke, Oier Mees, Sudeep Dasari, and Sergey Levine. Scaling cross-embodied learning: One policy for manipulation, navigation, locomotion and aviation. In CoRL, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 261, + 555, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 555, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 555, + 315 + ], + "type": "text", + "content": "[17] Yilun Du, Sherry Yang, Pete Florence, Fei Xia, Ayzaan Wahid, Pierre Sermanet, Tianhe Yu, Pieter Abbeel, Joshua B Tenenbaum, Leslie Pack Kaelbling, et al. Video language planning. In The Twelfth International Conference on Learning Representations. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 317, + 555, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 317, + 555, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 317, + 555, + 358 + ], + "type": "text", + "content": "[18] Yilun Du, Sherry Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Josh Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation. NeurIPS, 36, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 361, + 555, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 555, + 414 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 555, + 414 + ], + "type": "text", + "content": "[19] Hao-Shu Fang, Hongjie Fang, Zhenyu Tang, Jirong Liu, Junbo Wang, Haoyi Zhu, and Cewu Lu. Rh20t: A robotic dataset for learning diverse skills in one-shot. In RSS 2023 Workshop on Learning for Task and Motion Planning, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 416, + 554, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 554, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 554, + 448 + ], + "type": "text", + "content": "[20] Chongkai Gao, Haozhuo Zhang, Zhixuan Xu, Cai Zhehao, and Lin Shao. Flip: Flow-centric generative planning as general-purpose manipulation world model. In ICLR. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 449, + 555, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 555, + 502 + ], + "type": "text", + "content": "[21] Jiayuan Gu, Sean Kirmani, Paul Wohlhart, Yao Lu, Montserrat Gonzalez Arenas, Kanishka Rao, Wenhao Yu, Chuyuan Fu, Keerthana Gopalakrishnan, Zhuo Xu, et al. Rt-trajectory: Robotic task generalization via hindsight trajectory sketches. In ICLR. 3, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 504, + 555, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 555, + 547 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 555, + 547 + ], + "type": "text", + "content": "[22] Pierre-Louis Guhur, Shizhe Chen, Ricardo Garcia Pinel, Makarand Tapaswi, Ivan Laptev, and Cordelia Schmid. Instruction-driven history-aware policies for robotic manipulations. In CORL, pages 175-187. PMLR, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 548, + 554, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 554, + 581 + ], + "type": "text", + "content": "[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 582, + 555, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 555, + 624 + ], + "type": "text", + "content": "[24] Changho Hwang, Wei Cui, Yifan Xiong, Ziyue Yang, Ze Liu, Han Hu, Zilong Wang, Rafael Salas, Jithin Jose, Prabhat Ram, et al. Tutel: Adaptive mixture-of-experts at scale. Proceedings of Machine Learning and Systems, 5:269-287, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 625, + 555, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 555, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 555, + 679 + ], + "type": "text", + "content": "[25] Vidhi Jain, Maria Attarian, Nikhil J Joshi, Ayzaan Wahid, Danny Driess, Quan Vuong, Pannag R Sanketi, Pierre Sermanet, Stefan Welker, Christine Chan, et al. Vid2robot: End-to-end video-conditioned policy learning with cross-attention transformers. arXiv preprint arXiv:2403.12943, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "type": "text", + "content": "[26] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J Davison. Rlbench: The robot learning benchmark & learning environment. RAL, 5(2):3019-3026, 2020. 6" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6968" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "type": "text", + "content": "[27] Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kappler, Frederik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. Bc-z: Zero-shot task generalization with robotic imitation learning. In CoRL, pages 991-1002. PMLR, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 171 + ], + "type": "text", + "content": "[28] Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. Mixtral of experts. arXiv preprint arXiv:2401.04088, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 296, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 296, + 226 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 296, + 226 + ], + "type": "text", + "content": "[29] Yunfan Jiang, Agrim Gupta, Zichen Zhang, Guanzhi Wang, Yongqiang Dou, Yanjun Chen, Li Fei-Fei, Anima Anandkumar, Yuke Zhu, and Linxi Fan. Vima: General robot manipulation with multimodal prompts. arXiv preprint arXiv:2210.03094, 2(3):6, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 296, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 296, + 281 + ], + "type": "text", + "content": "[30] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 283, + 296, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 296, + 326 + ], + "type": "text", + "content": "[31] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv preprint arXiv:2307.07635, 2023. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 296, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 296, + 390 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 296, + 390 + ], + "type": "text", + "content": "[32] Alexander Khazatsky, Karl Pertsch, Suraj Nair, Ashwin Balakrishna, Sudeep Dasari, Siddharth Karamcheti, Soroush Nasiriany, Mohan Kumar Srirama, Lawrence Yunliang Chen, Kirsty Ellis, et al. Droid: A large-scale in-the-wild robot manipulation dataset. arXiv preprint arXiv:2403.12945, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 393, + 296, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 296, + 446 + ], + "type": "text", + "content": "[33] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 449, + 296, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 296, + 480 + ], + "type": "text", + "content": "[34] Po-Chen Ko, Jiayuan Mao, Yilun Du, Shao-Hua Sun, and Joshua B Tenenbaum. Learning to act from actionless videos through dense correspondences. In ICLR. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 481, + 296, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 296, + 514 + ], + "type": "text", + "content": "[35] Alexander Koch. Low-cost robot arm. https://github.com/AlexanderKoch-Koch/low_costrobot, 2024. GitHub repository. 5, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 515, + 296, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 296, + 568 + ], + "type": "text", + "content": "[36] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan First, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. arXiv preprint arXiv:2006.16668, 2020. 2, 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 570, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 296, + 602 + ], + "type": "text", + "content": "[37] Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuomotor policies. JMLR, 17(39):1-40, 2016. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 603, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 296, + 645 + ], + "type": "text", + "content": "[38] Xinghang Li, Minghuan Liu, Hanbo Zhang, Cunjun Yu, Jie Xu, Hongtao Wu, Chilam Cheang, Ya Jing, Weinan Zhang, Huaping Liu, et al. Vision-language foundation models as effective robot imitators. In ICLR. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 296, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 690 + ], + "type": "text", + "content": "[39] Bin Lin, Zhenyu Tang, Yang Ye, Jiaxi Cui, Bin Zhu, Peng Jin, Junwu Zhang, Munan Ning, and Li Yuan. Moe-llava: Mixture of experts for large vision-language models. arXiv preprint arXiv:2401.15947, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 692, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 296, + 713 + ], + "type": "text", + "content": "[40] Aixin Liu, Bei Feng, Bin Wang, Bingxuan Wang, Bo Liu, Chenggang Zhao, Chengqi Dengr, Chong Ruan, Damai Dai," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 105 + ], + "type": "text", + "content": "Daya Guo, et al. Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model. arXiv preprint arXiv:2405.04434, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 140 + ], + "type": "text", + "content": "[41] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning. NeurIPS, 36, 2024. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 555, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 555, + 194 + ], + "type": "text", + "content": "[42] Fangchen Liu, Kuan Fang, Pieter Abbeel, and Sergey Levine. Moka: Open-vocabulary robotic manipulation through mark-based visual prompting. In First Workshop on Vision-Language Models for Navigation and Manipulation at ICRA 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 555, + 239 + ], + "type": "text", + "content": "[43] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. In ICLR, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 242, + 555, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 242, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 242, + 555, + 285 + ], + "type": "text", + "content": "[44] Corey Lynch, Mohi Khansari, Ted Xiao, Vikash Kumar, Jonathan Tompson, Sergey Levine, and Pierre Sermanet. Learning latent plans from play. In CoRL, pages 1113-1132. PMLR, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 287, + 555, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 555, + 330 + ], + "type": "text", + "content": "[45] Corey Lynch, Ayzaan Wahid, Jonathan Tompson, Tianli Ding, James Betker, Robert Baruch, Travis Armstrong, and Pete Florence. Interactive language: Talking to robots in real time. RAL, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 332, + 555, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 555, + 386 + ], + "type": "text", + "content": "[46] Soroush Nasiriany, Sean Kirmani, Tianli Ding, Laura Smith, Yuke Zhu, Danny Driess, Dorsa Sadigh, and Ted Xiao. Rt-affordance: Affordances are versatile intermediate representations for robot manipulation. arXiv preprint arXiv:2411.02704, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 388, + 555, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 555, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 555, + 442 + ], + "type": "text", + "content": "[47] Abhishek Padalkar, Acorn Pooley, Ajinkya Jain, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anikait Singh, Anthony Brohan, et al. Open x-embediment: Robotic learning datasets and rt-x models. arXiv preprint arXiv:2310.08864, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 445, + 555, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 555, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 555, + 476 + ], + "type": "text", + "content": "[48] Joan Puigcerver, Carlos Riquelme, Basil Mustafa, and Neil Houlsby. From sparse to soft mixtures of experts. arXiv preprint arXiv:2308.00951, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 478, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 555, + 533 + ], + "type": "text", + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICLR, pages 8748-8763. PMLR, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 535, + 555, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 555, + 578 + ], + "type": "text", + "content": "[50] Rouhollah Rahmatizadeh, Pooya Abolghasemi, Ladislau Böloni, and Sergey Levine. Vision-based multi-task manipulation for inexpensive robots using end-to-end learning from demonstration. In ICRA, pages 3758-3765. IEEE, 2018. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 580, + 555, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 555, + 633 + ], + "type": "text", + "content": "[51] Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. arXiv preprint arXiv:2205.06175, 2022. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 635, + 555, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 555, + 669 + ], + "type": "text", + "content": "[52] Younggyo Seo, Kimin Lee, Stephen L James, and Pieter Abbeel. Reinforcement learning with action-free pre-training from videos. In ICLR, pages 19561-19579. PMLR, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "text", + "content": "[53] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6969" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 106 + ], + "type": "text", + "content": "[54] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In CoRL, pages 894-906. PMLR, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "text", + "content": "[55] Austin Stone, Ted Xiao, Yao Lu, Keerthana Gopalakrishnan, Kuang-Huei Lee, Quan Vuong, Paul Wohlhart, Brianna Zitkovich, Fei Xia, Chelsea Finn, et al. Open-world object manipulation using pre-trained vision-language models. In CoRL, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "type": "text", + "content": "[56] Priya Sundaresan, Quan Vuong, Jiayuan Gu, Peng Xu, Ted Xiao, Sean Kirmani, Tianhe Yu, Michael Stark, Ajinkya Jain, Karol Hausman, et al. Rt-sketch: Goal-conditioned imitation learning from hand-drawn sketches. 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 296, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 296, + 252 + ], + "type": "text", + "content": "[57] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 296, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 296, + 286 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 296, + 286 + ], + "type": "text", + "content": "[58] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. In NeurIPS, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 287, + 296, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 296, + 331 + ], + "type": "text", + "content": "[59] Mel Vecerik, Carl Doersch, Yi Yang, Todor Davchev, Yusuf Aytar, Guangyao Zhou, Raia Hadsell, Lourdes Agapito, and Jon Scholz. Robotap: Tracking arbitrary points for few-shot visual imitation. In ICRA, pages 5397-5403. IEEE, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 296, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 296, + 386 + ], + "type": "text", + "content": "[60] Homer Rich Walke, Kevin Black, Tony Z Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen-Estruch, Andre Wang He, Vivek Myers, Moo Jin Kim, Max Du, et al. Bridgedata v2: A dataset for robot learning at scale. In CoRL, pages 1723-1736. PMLR, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 296, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 296, + 432 + ], + "type": "text", + "content": "[61] Chen Wang, Linxi Fan, Jiankai Sun, Ruohan Zhang, Li Fei-Fei, Danfei Xu, Yuke Zhu, and Anima Anandkumar. Mimicplay: Long-horizon imitation learning by watching human play. In CoRL. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 296, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 296, + 466 + ], + "type": "text", + "content": "[62] Lirui Wang, Xinlei Chen, Jialiang Zhao, and Kaiming He. Scaling proprioceptive-visual learning with heterogeneous pre-trained transformers. CoRR, abs/2409.20537, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 468, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 296, + 510 + ], + "type": "text", + "content": "[63] Chuan Wen, Xingyu Lin, John So, Kai Chen, Qi Dou, Yang Gao, and Pieter Abbeel. Any-point trajectory modeling for policy learning. arXiv preprint arXiv:2401.00025, 2023. 1, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 513, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 513, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 513, + 296, + 567 + ], + "type": "text", + "content": "[64] Kun Wu, Chengkai Hou, Jiaming Liu, Zhengping Che, Xiaozhu Ju, Zhuqin Yang, Meng Li, Yinuo Zhao, Zhiyuan Xu, Guang Yang, et al. Robomind: Benchmark on multi-embediment intelligence normative data for robot manipulation. arXiv preprint arXiv:2412.13877, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 568, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 296, + 612 + ], + "type": "text", + "content": "[65] Mengda Xu, Zhenjia Xu, Yinghao Xu, Cheng Chi, Gordon Wetzstein, Manuela Veloso, and Shuran Song. Flow as the cross-domain manipulation interface. arXiv preprint arXiv:2407.15208, 2024. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 613, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 296, + 667 + ], + "type": "text", + "content": "[66] Jiange Yang, Wenhui Tan, Chuhao Jin, Bei Liu, Jianlong Fu, Ruihua Song, and Limin Wang. Pave the way to grasp anything: Transferring foundation models for universal pickplace robots. arXiv preprint arXiv:2306.05716, 1(2), 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 669, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 296, + 712 + ], + "type": "text", + "content": "[67] Jiange Yang, Bei Liu, Jianlong Fu, Bocheng Pan, Gangshan Wu, and Limin Wang. Spatiotemporal predictive pre-training for robotic motor control. arXiv preprint arXiv:2403.05304, 2024. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 555, + 386 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 315, + 72, + 555, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 555, + 137 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 555, + 137 + ], + "type": "text", + "content": "[68] Jiange Yang, Wenhui Tan, Chuhao Jin, Keling Yao, Bei Liu, Jianlong Fu, Ruihua Song, Gangshan Wu, and Limin Wang. Transferring foundation models for generalizable robotic manipulation. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), pages 1999-2010, 2025. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 140, + 555, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 140, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 315, + 140, + 555, + 184 + ], + "type": "text", + "content": "[69] Seonghyeon Ye, Joel Jang, Byeongguk Jeon, Sejune Joo, Jianwei Yang, Baolin Peng, Ajay Mandlekar, Reuben Tan, Yu-Wei Chao, Bill Yuchen Lin, et al. Latent action pretraining from videos. arXiv preprint arXiv:2410.11758, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 185, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 185, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 315, + 185, + 555, + 217 + ], + "type": "text", + "content": "[70] Chengbo Yuan, Chuan Wen, Tong Zhang, and Yang Gao. General flow as foundation affordance for scalable robot learning. arXiv preprint arXiv:2401.11439, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 219, + 555, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 219, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 315, + 219, + 555, + 262 + ], + "type": "text", + "content": "[71] Jinguo Zhu, Xizhou Zhu, Wenhai Wang, Xiaohua Wang, Hongsheng Li, Xiaogang Wang, and Jifeng Dai. Uniperceiver-moe: Learning sparse generalist models with conditional moes. NeurIPS, 35:2664-2678, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 263, + 555, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 263, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 315, + 263, + 555, + 308 + ], + "type": "text", + "content": "[72] Xizhou Zhu, Jinguo Zhu, Hao Li, Xiaoshi Wu, Hongsheng Li, Xiaohua Wang, and Jifeng Dai. Uni-perceiver: Pre-training unified architecture for generic perception for zero-shot and few-shot tasks. In CVPR, pages 16804-16815, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 308, + 555, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 308, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 315, + 308, + 555, + 342 + ], + "type": "text", + "content": "[73] Yifeng Zhu, Zhenyu Jiang, Peter Stone, and Yuke Zhu. Learning generalizable manipulation policies with object-centric 3d representations. In CoRL. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 342, + 555, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 342, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 315, + 342, + 555, + 386 + ], + "type": "text", + "content": "[74] Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam Shazeer, and William Fedus. St-moe: Designing stable and transferable sparse expert models. arXiv preprint arXiv:2202.08906, 2022. 2, 3, 5" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "6970" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TraF-Align_ Trajectory-aware Feature Alignment for Asynchronous Multi-agent Perception/0f12efc0-153e-4cda-95dd-b16491feb48e_content_list.json b/2025/TraF-Align_ Trajectory-aware Feature Alignment for Asynchronous Multi-agent Perception/0f12efc0-153e-4cda-95dd-b16491feb48e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..04f2e2e246e65d1166b09b33a092d55523c8fa6d --- /dev/null +++ b/2025/TraF-Align_ Trajectory-aware Feature Alignment for Asynchronous Multi-agent Perception/0f12efc0-153e-4cda-95dd-b16491feb48e_content_list.json @@ -0,0 +1,1335 @@ +[ + { + "type": "text", + "text": "TraF-Align: Trajectory-aware Feature Alignment for Asynchronous Multi-agent Perception", + "text_level": 1, + "bbox": [ + 156, + 130, + 843, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhiying Song $^{1}$ , Lei Yang $^{1}$ , Fuxi Wen $^{1,2,\\dagger}$ , Jun Li $^{1,2}$ $^{1}$ School of Vehicle and Mobility, Tsinghua University †Corresponding author \n $^{2}$ State Key Lab of Intelligent Green Vehicle and Mobility \n{song-zy24,yanglei20}@ mails.tsinghua.edu.cn, {wenfuxi, lijun1958}@tsinghua.edu.cn", + "bbox": [ + 143, + 202, + 849, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cooperative perception presents significant potential for enhancing the sensing capabilities of individual vehicles, however, inter-agent latency remains a critical challenge. Latencies cause misalignments in both spatial and semantic features, complicating the fusion of real-time observations from the ego vehicle with delayed data from others. To address these issues, we propose TraF-Align, a novel framework that learns the flow path of features by predicting the feature-level trajectory of objects from past observations up to the ego vehicle's current time. By generating temporally ordered sampling points along these paths, TraF-Align directs attention from the current-time query to relevant historical features along each trajectory, supporting the reconstruction of current-time features and promoting semantic interaction across multiple frames. This approach corrects spatial misalignment and ensures semantic consistency across agents, effectively compensating for motion and achieving coherent feature fusion. Experiments on two real-world datasets, V2V4Real and DAIR-V2X-Seq, show that TraF-Align sets a new benchmark for asynchronous cooperative perception. The code is available at https://github.com/zhyingS/TraF-Align.", + "bbox": [ + 88, + 342, + 483, + 675 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 705, + 220, + 722 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Intelligent vehicles rely on advanced onboard sensors, such as LiDAR and cameras, to perceive and understand the surrounding world. Nevertheless, the limited field of view and detection range of onboard sensors restrict their ability to perceive objects in occluded areas and at greater distances [2, 23, 36, 47]. Vehicle-to-everything (V2X) communication-enabled cooperative perception systems bridge this gap by exchanging perception information between vehicles and other agents, thereby extending the perception capability of a single agent [3, 4, 20, 41, 45].", + "bbox": [ + 89, + 732, + 482, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In real-world scenarios, cooperative perception systems", + "bbox": [ + 109, + 885, + 482, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f3d4dfeee0aaaa6a9eee9851008ffb564a0a3cc14d5fed9de5113269dbcf1eaf.jpg", + "image_caption": [ + "Figure 1. Asynchronous observations of ego vehicle and other agents result in spatial and semantic misalignments." + ], + "image_footnote": [], + "bbox": [ + 529, + 309, + 901, + 506 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "are cursed with inter-agent time delays [33, 35]. Consider a network with an ego agent and agent $i$ , where their observations of the same object are shown in Figure 1. Latency in agent $i$ 's message, captured before ego's observation time, can result in either a complete loss of features (1-I) or feature misalignment relative to ego's observation (1-II). Corresponding to Figure 1-I and 1-II, there are two main challenges: i) Spatial misalignment, where observations from other agents fail to align with the ground truth in space, lacking any reference from the ego vehicle's observations. ii) Semantic misalignment, where both the ego vehicle and another agent observe the same object but fail to recognize it as the same entity. These misalignments lead to misplaced objects and increased false positives [29, 30].", + "bbox": [ + 511, + 579, + 906, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the first challenge, some studies have proposed potential solutions, by predicting the current-frame features using historical frames [15], or predicting the pixel-level flow of features and then extrapolating them [44] or directly warping them to their future positions [34]. For instance, FFNet [44] uses a self-supervised approach to learn the first-order derivatives of features between consecutive", + "bbox": [ + 511, + 795, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12048", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "frames, and CoBEVFlow [34] generates an instance-level region of interest (ROI) and calculates feature motion vectors by matching and tracking these regions. However, predicting the movement of high-dimensional neural features is non-trivial. Despite their efforts, these methods struggle with larger delays and lack end-to-end capability. For the second challenge, the significance of semantic misalignment has not been well investigated in previous work.", + "bbox": [ + 89, + 90, + 480, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these misalignments, our core idea is to reconstruct current-time features and achieve semantic consensus across agents on multi-frame LiDAR features within a unified framework. We guide the flow and interaction of high-dimensional features by learning the motion of low-dimensional trajectory features. A trajectory field is defined as a spatiotemporal map of each object's anticipated path, guiding the generation of temporally ordered sampling points along each trajectory. Cross-attention between each query and its aligned sampling points allows the query to access relevant past features along the trajectory, enabling effective feature reconstruction and semantic interaction for robust spatial alignment and inter-agent consensus. Consider the spatial misalignment scenario shown in Figure 1-I. By aggregating multi-frame point cloud features from agent $i$ , we learn the object's past trajectory distribution and predict its motion direction up to the ego vehicle's current time, forming a trajectory field. This field guides attention from the predicted position to historical features along the same trajectory, enabling the reconstruction of the object's current features near the ground truth and compensating for motion. For semantic misalignment cases (Figure 1-II), while delays may misalign agent $i$ 's observations with the ego vehicle's, the continuity of object motion implies overlap with point clouds from past ego frames. By learning the trajectory distribution from the ego's multi-frame point clouds and applying attention to compute feature correlations along the trajectory, the model recognizes that all features along the trajectory correspond to the same object, thus achieving cross-agent semantic alignment.", + "bbox": [ + 91, + 217, + 486, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on the abovementioned idea, we propose TraFAlign, a novel framework for asynchronous multi-agent cooperative perception (as shown in Figure 2). The proposed method includes three key parts: i) a field predictor that generates synchronized trajectory fields aligned with the ego vehicle's time, ii) an offset generator that produces sampling points for each query position and distributes them along the corresponding trajectory, and iii) attention layers where each query position is treated as a token, with attention focused only on positions specified by the offset generator. Extensive experiments are conducted on two widely used large-scale real-world datasets, V2V4Real [38] and DAIR-V2X-Seq [43], one for vehicle-to-vehicle (V2V) cooperative perception, another for vehicle-to-infrastructure (V2I) applications. Experimental results show that state-of", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the-art (SOTA) performance is achieved for TraF-Align in both object detection precision and robustness to inter-agent latency. The contributions are summarized as follows:", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose TraF-Align, a novel framework for asynchronous cooperative perception, addressing spatial and semantic misalignments caused by time delays.", + "- We leverage a field predictor and an offset generator to align trajectory fields and distribute attention points along object trajectories, enabling robust feature reconstruction and interaction along spatiotemporally aligned paths.", + "- We introduce a field loss and an offset loss to supervise trajectory alignment and the generation of attention points, providing an end-to-end trainable solution." + ], + "bbox": [ + 511, + 136, + 903, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 301, + 650, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Asynchronous V2X cooperative perception. V2X cooperative perception extends the sensing range of single vehicles by exchanging information with other agents via wireless communication [13]. Previous studies have conducted extensive research under time-synchronized conditions, such as F-Cooper [3], V2VNet [33], V2X-ViT [35, 39], ERMVP [45], etc. In the presence of latency, previous research emphasizes using historical data to estimate and compensate for object motion. For instance, at the output level, the object state filters are designed to predict and update the dynamic state of the received objects [1]. Using raw data, Zhang et al. [46] and Dao et al. [8] estimated the motion of point clouds and predicted compensatory flows for late-arriving data. However, these levels may encounter challenges, such as short prediction time horizons or high communication data volumes. At the feature level, which has a better balance between accuracy and transmission bandwidth [18], SyncNet [15] employs a pyramidal LSTM to iteratively process and learn from delayed features, enabling the prediction of current-frame features. FFNet [44] predicts inter-frame feature changes in a self-supervised feature manner to synchronize features across agents. Nevertheless, predicting high-dimensional features is challenging and prone to uncontrollable variations. Additionally, CoBEVFlow [34] extracts and warps object-specific features to anticipated locations by predicting instance-wise motion vectors. However, this approach requires generating ROIs and tracking their motion, resulting in a non-end-to-end two-stage method. This paper uses multi-frame LiDAR sequences to predict low-dimensional trajectory features and guide attention to fuse features on past trajectories. By operating directly at the feature level, our approach enables end-to-end training and avoids the complexities of high-dimensional feature prediction.", + "bbox": [ + 511, + 325, + 906, + 840 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning from LiDAR sequences across agents. Multiframe point cloud fusion is widely applied in single-vehicle autonomous systems to amalgamate semantic data across different perspectives [5, 21, 22, 24, 32, 42]. However, its", + "bbox": [ + 511, + 840, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12049", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2cdf8216f524cb86daeb1fdb5422271a921dc7b3ef7ca977ca25426faa07647e.jpg", + "image_caption": [ + "Figure 2. Proposed Architecture. With asynchronous LiDAR inputs from the ego agent (time $t$ ) and agent $i$ (time $s = t - \\tau_t^i$ ), BEV features are extracted using an onboard sparse encoder. Agent $i$ 's features are transmitted to the ego agent and stored in memory. TraFAlign predicts a trajectory field up to time $t$ , generates target attention positions, and uses features at these positions as keys/values in attention layers. Finally, the multi-agent features are fused and processed by the head to generate cooperative predictions at time $t$ ." + ], + "image_footnote": [], + "bbox": [ + 101, + 85, + 897, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "application in V2X cooperative perception remains under-explored. Although studies like SCOPE [40] and MRC-Net [11] have leveraged recurrent neural networks to fuse sequences from ego vehicles, these methods are resource-intensive and limited by sequence length. This has restricted their use to ego vehicles rather than a broader agent network, necessitating a more straightforward multi-frame fusion network. In cooperative perception scenarios, multiframe point clouds offer crucial motion cues that enable agents to synchronize their understanding of object positions despite time delays, thereby correlating trajectories across agents. This aspect has yet to be fully explored at the feature level. This paper proposes a novel approach using trajectory fields and customized attention with offsets for direct multi-frame fusion, thus establishing a consistent understanding of object motion across the network.", + "bbox": [ + 88, + 454, + 485, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 714, + 183, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem formulation", + "text_level": 1, + "bbox": [ + 89, + 741, + 290, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider a group of networked agents indexed by $\\mathcal{I} = \\{e, 1, 2, \\dots, N - 1\\}$ , each is equipped with LiDAR and communication units. Index $e$ indicates the ego agent, which integrates information from other agents to extend its sensing range and reduce the blind spots. The LiDAR observation of the $i$ th agent at its local timestamp $s$ is denoted as $L_s^i$ , where $i \\in \\mathcal{I}$ . A shared neural network $f(\\cdot)$ processes this observation into deep features. At each timestamp $s$ , agent $i$ transmits $F_s^i = f(L_s^i)$ to the ego agent,", + "bbox": [ + 89, + 763, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "who then accumulates these features in the cache, forming a historical feature set for agent $i$ at time $s$ , represented as $\\mathcal{F}_s^i = \\{F_{s - k}^i\\}_{k\\in K}$ , where $K = \\{k\\in \\mathbb{Z}|1\\leqslant k\\leqslant m_i\\}$ , $m_{i}$ is the maximal sequence length for agent $i$ . Accounting for delays in the collaborative process, messages from various agents arrive at ego at different times. Let $\\tau_t^i$ be the delay for agent $i$ 's message, and this means the information sent by agent $i$ at time $t - \\tau_t^i$ is received by the ego at time $t$ . If ignoring feature extraction delay, then $s = t - \\tau_t^i$ . The task is to fuse these multi-agent features and estimate the state of objects, i.e., the bounding boxes, at ego timestamp $t$ within the detection range. The information available for fusion includes: i) the ego's historical features $\\mathcal{F}_t^e$ , and ii) the historical features from other agents, denoted as $\\{\\mathcal{F}_{t - \\tau_t^i}^i\\}_{i\\in \\mathcal{T}\\backslash \\{e\\}}$ . The proposed method is shown in Figure 2.", + "bbox": [ + 511, + 454, + 908, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Feature extraction", + "text_level": 1, + "bbox": [ + 511, + 696, + 692, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The encoder converts single-frame raw point cloud data into BEV features. Similar to PointPillars [14], the point cloud is first discretized into a regular grid on the $x - y$ plane. Within each pillar, the point features are encapsulated as the vector $[x,y,z,x_c,y_c,z_c,x_p,y_p,z_p]\\in \\mathbb{R}^9$ . Here, $[x,y,z]$ is the point's position, while $[x_c,y_c,z_c]$ indicates the center of the pillar. The vector $[x_p,y_p,z_p]$ represents the point's offset from the pillar center. The pillar feature extraction and backbone network architectures are adopted from PillarNext [16]. Multi-Layer Perceptron (MLP) serves to encode point-wise features, while a sparse variant of ResNet [10] is utilized as the backbone to distill deeper and more", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12050", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "complex features with size $(C, H, W)$ . The sparse network is implemented using Spatially Sparse Convolution (SpConv) library [6]. The extracted single-frame features are transmitted and stored in the ego's feature bank $(\\mathcal{F}_t^e$ and $\\mathcal{F}_s^i)$ , which maintains a fixed capacity by admitting new features and discarding the oldest ones, ensuring an up-to-date repository for feature reuse. The extracted features of each agent are then processed using TraF-Align separately.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temporal Encoding. Temporal information is decoupled from the point features to facilitate the reuse of historical features across frames. This allows for the immediate retrieval of historical features at any timestamp. Inspired by the positional embedding in Transformer [31], a temporal embedding $TE \\in \\mathbb{R}^{C \\times H \\times W}$ is defined as", + "bbox": [ + 89, + 212, + 480, + 301 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT E _ {(\\tau , 2 j)} = \\sin (\\xi), T E _ {(\\tau , 2 j + 1)} = \\cos (\\xi), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 313, + 480, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\xi = \\tau / \\varepsilon^{2j/C}$ , $j$ indexes the dimension along $C$ channels, $\\tau$ represents the number of frames delayed from the current frame, and $\\varepsilon$ is set to eight in this study. Each feature within the memory cache, $\\mathcal{F}_t^e$ for the ego agent and $\\mathcal{F}_s^i$ for agent $i$ , associated with a specific temporal delay, is concatenated with a corresponding temporal embedding. These concatenated features are refined through a convolutional layer with a $1 \\times 1$ kernel and then warped to time $t$ by ego-motion compensation.", + "bbox": [ + 89, + 340, + 482, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Trajectory field", + "text_level": 1, + "bbox": [ + 89, + 486, + 246, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each agent processes its aggregated multi-frame feature map via a field predictor, which generates a trajectory field consisting of position and orientation fields. The position field creates heatmap peaks along the object's trajectory, while the orientation field captures the inverse tangent direction of the trajectory. Due to delays, agents' local observations may be outdated. To achieve temporal alignment, the field predictor predicts the trajectory features of all objects up to the ego vehicle's latest time (with temporal information encoded in the features via Equation 1). This ensures that, for the same object, trajectory fields from different viewpoints spatially overlap and are aligned temporally to the ego vehicle's current time. Our model uses a variant of UNet [25] as the field predictor, whose architecture is detailed in the supplementary material, along with a supervision signal to guide the learning process.", + "bbox": [ + 88, + 508, + 482, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 3, when generating the ground truth field, we first extract bounding box annotations of each object from the local perspective of each agent within the maximum temporal window, starting from the object's initial appearance up to the ego's current time. A trajectory is then constructed by connecting the centers of these bounding boxes. Specifically, the length of the object's trajectory is given by $\\lceil \\tau_i \\cdot \\omega_i \\rceil + m_i$ , $\\forall i \\in \\mathcal{I}$ , where $\\lceil \\cdot \\rceil$ denotes the ceiling function, $\\omega_i$ is the LiDAR sampling frequency $(\\mathrm{s}^{-1})$ of agent $i$ , and $\\tau_i$ represents its cooperative latency (s). The", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0af345f65626421913b5a14c7b1c3fb65639dfffa01dae3308eeaa52744f6a10.jpg", + "image_caption": [ + "Figure 3. Illustration of trajectory field, which includes a position field indicating the occupancy grid of the trajectory and a direction field depicting the trajectory's flow orientation." + ], + "image_footnote": [], + "bbox": [ + 522, + 93, + 898, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "trajectory field is then generated by interpolating along the trajectory and projecting it onto the BEV feature space. If two object trajectories overlap, the trajectory with the more recent timestamp will overwrite the older one.", + "bbox": [ + 511, + 422, + 903, + 482 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For cases in Figure 1-I, such a design compels the network to predict the trajectory field up to the present moment, thereby supporting the learning of motion information. Failing to do so would result in the network detecting only the delayed bounding box. Compared to the direct forecasting of high-dimensional features from the backbone network, predicting the trajectory field provides a more straightforward and more reliable method with predictable outcomes.", + "bbox": [ + 511, + 484, + 905, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Trajectory-aware attention", + "text_level": 1, + "bbox": [ + 511, + 635, + 756, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We flatten the input feature map $F^i (\\forall i \\in \\mathcal{I})$ of size $C \\times H \\times W$ into $H \\times W$ embeddings, each serving as a query $q \\in \\mathbb{R}^{1 \\times C}$ . For each query, the offset generator creates a response set $R \\in \\mathbb{R}^{n \\times C}$ , capturing features at designated attention positions that are learned based on predefined ground truth. $n$ is the number of attention positions for each query, and it's set to 18 in this paper. To generate the ground truth offsets, we select $n$ pixels on the feature map that satisfy three conditions: i) The position field has a response (indicating the presence of an object), ii) The pixels belong to the same trajectory as the query (indicating the same object), and iii) The pixels correspond to earlier timestamps than the query (indicating historical positions of the object). As shown in Figure 4b, this ground truth aligns each query with preceding trajectory positions, enabling it to draw in relevant historical information. The", + "bbox": [ + 511, + 657, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12051", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "trajectory field has effectively modeled the motion and flow of the historical features, offering valuable guidance to the offset generator. With this rich information, the generator can rely on simple convolution layers to produce good attention offsets, and we use PReLU activation [9] to provide flexibility in adjusting the offsets' directional orientation. Given the response set $R$ , the feature of query $q$ is updated or reconstructed by the attention process:", + "bbox": [ + 89, + 90, + 483, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {A t t e n t i o n} (q, R) = \\text {S o f t m a x} \\left[ \\frac {q W ^ {q} \\left(R W ^ {k}\\right) ^ {T}}{\\sqrt {d}} \\right] R W ^ {v}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 219, + 483, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $W^{q}, W^{k}, W^{v} \\in \\mathbb{R}^{C \\times d}$ are the query, key and value projection matrices. Similar to the original Transformer [31], we also incorporate feed-forward networks, addition and normalization layers, and a multi-head attention design, the details are shown in Figure 4a. Using attention along the trajectory transfers features to the current time step and enables feature interaction along the trajectory. This allows the model to recognize which features belong to the same object, aligning the features across agents for effective fusion. As a result, each agent aligns its feature map to the ego vehicle's latest time step.", + "bbox": [ + 89, + 268, + 483, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Inter-agent fusion, detection head and loss", + "text_level": 1, + "bbox": [ + 89, + 444, + 452, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While inter-agent fusion has been extensively studied [3, 36, 45], this paper adopts a simplified approach by concatenating features from all agents and utilizing a basic convolution layer for fusion. We use an anchor-based detection head and employ three losses: a detection loss, a field loss, and an offset loss:", + "bbox": [ + 89, + 465, + 483, + 555 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {d e t e c t i o n}} + \\alpha \\mathcal {L} _ {\\text {f i e l d}} + \\beta \\mathcal {L} _ {\\text {o f f s e t}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 566, + 480, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{detection}}$ is consistent with the loss used in [14], $\\mathcal{L}_{\\mathrm{field}}$ includes a focal loss [17] for the position field and an $L1$ loss for the orientation field. The constants $\\alpha$ and $\\beta$ balance the contributions of each loss term. Here, we set $\\alpha = \\beta = 0.05$ . Considering the offset loss, because the generated attention positions are unordered, we first apply the Sinkhorn algorithm [7, 26, 27] to compute the matching probabilities $P_{jk} \\in \\mathbb{R}^{n \\times n}$ between the generated attention positions and the ground truth. This is done using the L1 distance matrix $C_{jk} \\in \\mathbb{R}^{n \\times n}$ as the cost for the pixel located at $(j, k)$ on the feature map. The offset loss is then defined as:", + "bbox": [ + 89, + 590, + 483, + 771 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {o f f s e t}} = \\sum_ {j, k} P _ {j k} \\cdot C _ {j k}, \\forall j \\in [ 1, H ], k \\in [ 1, W ]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 777, + 483, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 830, + 223, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate robustness to cross-agent latency, we conduct 3D object detection experiments on two large-scale real-world datasets, V2V4Real [38] and DAIR-V2X-Seq [43].", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/55564513160de392774c28db019896855f74f3937e31077014e26ee013505a5b.jpg", + "image_caption": [ + "(a) Trajectory-aware attention layers." + ], + "image_footnote": [], + "bbox": [ + 527, + 87, + 903, + 233 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e47348146adec98c21b18cdd1b1ad9c97eedf19a6115575a1a39490bceb3a2f8.jpg", + "image_caption": [ + "(b) Query and ground truth attention positions.", + "Figure 4. Attention positions and trajectory-aware attention layers." + ], + "image_footnote": [], + "bbox": [ + 529, + 253, + 898, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "These datasets encompass both V2V and V2I scenarios, as well as sparse and dense LiDAR conditions. The detection results are evaluated using Average Precision (AP) at Intersection-over-Union (IoU) thresholds of 0.5 (AP50) and 0.7 (AP70). To enable a fair comparison, we omit feature compression operations from the methods under evaluation.", + "bbox": [ + 511, + 460, + 905, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation details", + "text_level": 1, + "bbox": [ + 511, + 561, + 728, + 577 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and Benchmarks. V2V4Real [38] is specialized for V2V cooperative detection in urban settings, and DAIR-V2X-Seq [43] captures sequential data in V2I scenarios at intersections. The benchmarks include single-agent perception relying on the ego vehicle's data for detection, alongside eight SOTA intermediate fusion models: F-Cooper [3], V2VNet [33], Where2comm [12], AttFuse [36], V2X-ViT [35], CoBEVT [37], MRCNet [11], and ERMVP [45].", + "bbox": [ + 511, + 583, + 905, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training details. All the models are trained within the same codebase for 60 epochs on nine RTX 3090 GPUs, employing the one-cycle learning rate strategy [28] with AdamW [19] as the optimizer. The point clouds are segmented at a grid size of $0.4 \\times 0.4\\mathrm{m}$ , and the backbone's feature map size output is one-fourth the original size. During training on V2V4Real, all models shuffle the vehicle list and randomly select the ego vehicle at the start of each epoch. Standard LiDAR data augmentations are applied, including random flip, scaling, rotation, and translation. Multi-frame methods are trained with random latency augmentation, following a uniform distribution from 0 to $400~\\mathrm{ms}$ . TraF-Align leverages two LiDAR frames for the ego vehicle and four", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12052", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e322f68fde4a5aa4df1725cd69255fcc4ffaf9c03f3b71f7714021cea3620972.jpg", + "table_caption": [ + "Table 1. Performance comparison of vehicle class on V2V4Real validation and DAIR-V2X-Seq validation dataset with different latencies. The results are reported in AP 50/70 (%)." + ], + "table_footnote": [], + "table_body": "
MethodV2V4RealDAIR-V2X-Seq
0ms100 ms200 ms300 ms400 ms0ms100 ms200 ms300 ms400 ms
Single Agent48.75/30.5348.75/30.5348.75/30.5348.75/30.5348.75/30.5351.31/40.4551.31/40.4551.31/40.4551.31/40.4551.31/40.45
AttFuse [36]63.84/33.3059.16/27.3655.57/26.0353.25/25.4451.69/25.0968.94/49.6865.83/45.7663.02/44.2461.61/43.2860.19/42.53
F-Cooper [3]66.86/33.3962.15/27.6158.03/26.3055.31/25.4653.92/24.9569.47/50.0165.85/44.9662.74/43.2161.12/42.2860.01/41.70
V2VNet [33]67.11/35.4761.97/29.1458.09/27.8556.26/27.4354.62/26.7571.44/51.9167.26/46.6363.84/44.5961.68/43.2459.78/42.08
Where2comm [12]67.03/37.5561.86/31.5258.22/30.2056.22/29.7154.47/29.0768.37/48.5064.81/44.3261.51/42.8059.66/41.6358.25/40.80
V2X-ViT [35]68.51/35.8863.45/29.5759.42/28.0457.35/27.6555.60/27.2070.21/50.9266.42/46.4462.58/44.2961.00/43.5659.63/42.66
MRCNet [11]67.83/37.3762.65/31.5958.98/30.3956.70/29.5355.03/28.9469.27/50.6765.90/45.6062.65/44.1560.09/42.3958.19/41.35
CoBEVT [37]70.59/38.8265.50/32.1161.55/30.4058.91/29.5457.05/29.2671.51/51.3167.60/46.5564.38/45.0362.39/44.1160.74/43.08
ERMVP [45]70.74/38.4565.53/31.9161.63/30.5359.24/30.1857.26/29.3171.19/53.8567.04/48.0963.67/46.3761.23/44.7659.55/43.93
TraF-Align (Ours)74.28/44.1473.04/39.9971.51/39.5470.07/38.9969.41/38.4976.90/58.3173.95/53.6872.69/52.8371.66/52.3371.22/52.01
", + "bbox": [ + 91, + 128, + 898, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for the cooperative vehicle on V2V4Real, while using four LiDAR frames for the vehicle and infrastructure on DAIR-V2X-Seq. The number of attention layers is set to two, each containing four attention heads.", + "bbox": [ + 88, + 321, + 480, + 382 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Latency settings. To comprehensively evaluate robustness to time delays, we introduce delays of 0, 100, 200, 300, and $400\\mathrm{ms}$ to one of the agents separately, extensively testing the model's performance under each delay condition. Unless specified otherwise, these delays are applied to the point cloud data of the non-ego agent.", + "bbox": [ + 88, + 383, + 482, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with state-of-the-art", + "text_level": 1, + "bbox": [ + 89, + 484, + 382, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Detection performance. Table 1 presents a comparison of detection performance on two real-world datasets for vehicle class, which includes cars, vans, trucks, buses, etc. It's observed that the proposed method i) achieves state-of-the-art performance across all benchmarks under various latency conditions on both datasets. ii) In synchronous scenarios, i.e., when the delay is $0\\mathrm{ms}$ , our model improves AP 50/70 by $25.53\\% / 13.61\\%$ (on V2V4Real) and $25.59\\% / 18.26\\%$ (on DAIR-V2X-Seq) compared to single-agent perception, it also surpasses ERMVP, the SOTA performance of feature-level methods, by $3.54\\% / 5.69\\%$ and $5.71\\% / 4.46\\%$ , respectively.", + "bbox": [ + 88, + 507, + 482, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Robustness to inter-agent latency. As shown in Table 1, our model demonstrates stronger robustness in the presence of time delays. Even under a $400~\\mathrm{ms}$ delay, our AP decreases by only $4.87\\% /5.65\\%$ (on V2V4Real) and $5.68\\% /6.30\\%$ (on DAIR-V2X-Seq) compared to the synchronous scenario. This performance significantly outperforms the ERMVP, which shows declines of $13.48\\% /9.14\\%$ and $11.64\\% /9.92\\%$ , respectively.", + "bbox": [ + 88, + 688, + 482, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To gain a deeper understanding of the impact of latency on collaboration, we plot the Precision-Recall curves of the models under various latency conditions, as shown in Figure 5. It is obvious that cooperation is effective in enhancing the recall limit of single-vehicle perception, helping the ego vehicle to detect a more significant number of objects.", + "bbox": [ + 88, + 810, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d7d5e17242f79c59992fe9cc8f2b2a1a2073234a6b1571b037c6d9470d18cc6c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 526, + 321, + 704, + 479 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/33902ffe0d7955c3ac3b544de806c2fa363eab9a2ef4149b0a5e5d9f1ded6ee2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 320, + 895, + 479 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1a9ad1243c0ef0fbc8106297e8e9cc52ee1392f15c426150cc3e9b800a8dee69.jpg", + "image_caption": [ + "Figure 5. Precision-Recall curves for models on the V2V4Real validation dataset under various latency conditions." + ], + "image_footnote": [], + "bbox": [ + 524, + 484, + 704, + 672 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dd396af796c57a7a09b42fec6a53f420d55d851cace7b4f4551f0fcc0ef62558.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 484, + 897, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, at a $200\\mathrm{ms}$ latency, state-of-the-art methods exhibit lower precision in the low-recall range compared to single-vehicle perception. This suggests that the delayed collaborative information interferes with the objects that the single vehicle could detect, corresponding to the semantic misalignment issues depicted in Figure 1-II. In areas not visible to the ego vehicle, specifically at high recall levels, the precision of SOTA methods significantly declines with the introduction of latency. This indicates that spatial misalignment issues (Figure 1-I) become dominant, lead", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12053", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing to decreased performance. These findings reinforce our analysis of latency's impact on cooperative perception. Furthermore, using ERMVP as an example, the differences between the PR curves diminish as latency increases, indicating that existing models progressively lose their effectiveness in managing delays. In contrast, TraF-Align retains the original performance of single-agent perception in lowrecall regions and shows no significant decline in performance in high-recall areas due to latency. This highlights our model's ability to establish strong semantic consistency between the ego vehicle's data and the received information, demonstrating robust reconstruction capability and effectively addressing the scenarios in Figures 1-I and 1-II.", + "bbox": [ + 89, + 90, + 483, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation studies", + "text_level": 1, + "bbox": [ + 89, + 300, + 251, + 315 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conducted ablation studies on the effectiveness of significant components and loss and multiframe point clouds using average precision as the evaluation metric. Other ablation experiments and analyses (using PR curves) are presented in the supplementary material.", + "bbox": [ + 89, + 323, + 483, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Major Components. To evaluate the contribution of modules, we progressively remove i) the field predictor, ii) the offset generator, and iii) the attention layers. The results are shown in Table 2. The experiments are conducted on the DAIR-V2X-Seq dataset. To ensure a comprehensive assessment of each component's contributions, we train the model for 120 epochs, double the duration of the experiments presented in Table 1. Additionally, we test the effects of adding latency to both the ego and non-ego agents to fully evaluate the impact of each module. The consistent rise in detection precision over different latency settings demonstrates the effectiveness of each introduced component. Notably, applying latency to vehicles results in lower AP than when latency is applied to infrastructure. This discrepancy arises from the different distributions of object visibility between the vehicle agent and roadside in", + "bbox": [ + 89, + 400, + 483, + 642 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0f35fae71d44b655b919c9be8faebeede34fcc70d6af33dc6bf11f5af92e8fed.jpg", + "table_caption": [ + "Table 2. Results of the ablation study of the core components on the DAIR-V2X-Seq dataset." + ], + "table_footnote": [ + "FP: Field predictor. OG: Offset generator. AL: Attention layers." + ], + "table_body": "
FP OG ALLatency of Infrastructure
0ms100ms200ms300ms400ms
75.35/54.7273.01/50.8171.88/50.1771.04/49.7770.09/49.45
✓76.31/57.0274.60/53.5373.90/52.3472.91/51.99
✓77.11/58.6375.15/54.7774.18/53.7773.24/52.99
✓✓77.83/59.7275.74/55.8574.99/54.82
FP OG ALLatency of Vehicle
0ms100ms200ms300ms400ms
75.35/54.7274.31/49.6172.17/43.3267.19/34.8256.93/24.51
✓76.31/57.0275.10/51.9372.91/46.2269.22/39.15
✓77.11/58.6375.76/53.4873.44/47.5569.92/40.52
✓✓77.83/59.7276.41/54.8874.24/49.46
", + "bbox": [ + 91, + 707, + 480, + 882 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b03905b131ae06e2b47176f978f2208e3eb522d897484cb47d1bead45e31a7bc.jpg", + "table_caption": [ + "Table 3. Results of the ablation study of the loss on the DAIR-V2X-Seq dataset." + ], + "table_footnote": [ + "FL:Field loss. OL: Offset loss." + ], + "table_body": "
FL OLLatency of Infrastructure
0ms100ms200ms300ms400ms
✓76.43/57.7474.44/53.9173.65/52.9772.71/52.1071.83/51.87
77.12/58.3674.96/54.5174.02/53.6872.98/52.9172.01/52.27
✓77.15/58.7075.19/54.5874.31/53.6673.68/53.0772.52/52.28
✓77.83/59.7275.74/55.8574.99/54.8274.10/54.0772.89/53.21
FL OLLatency of Vehicle
0ms100ms200ms300ms400ms
✓76.43/57.7475.09/52.8172.84/46.5469.14/39.1261.53/29.15
77.12/58.3675.69/53.2372.76/46.8569.25/39.7862.13/30.41
✓77.15/58.7075.80/53.9073.34/48.4970.12/41.8063.21/32.02
✓77.83/59.7276.41/54.8874.24/49.4670.75/42.6263.59/32.31
", + "bbox": [ + 516, + 128, + 903, + 303 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/730d661a4453700fb5be3e7a5cb5543a368244299c1e2f39368ab5567b3902f5.jpg", + "image_caption": [ + "(a) LiDAR points." + ], + "image_footnote": [], + "bbox": [ + 524, + 335, + 633, + 541 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4f8815cb30ce6da13f277e6a32c79dbc91b6687d022eb8dc87d423dcd62c34f6.jpg", + "image_caption": [ + "(b) Trajectory field.", + "Figure 6. Visualization of TraF-Align in real-world scenarios from the DAIR-V2X-Seq dataset with latency of $400~\\mathrm{ms}$ ." + ], + "image_footnote": [], + "bbox": [ + 638, + 335, + 746, + 545 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a3b0ddecd3e40ab65b9a8bfcd3225a97fcdd0821f18f7e2c632b84b4ac408f3b.jpg", + "image_caption": [ + "(c) Deep features." + ], + "image_footnote": [], + "bbox": [ + 764, + 335, + 880, + 544 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "frastructure agent in the dataset.", + "bbox": [ + 511, + 628, + 727, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of loss. Table 3 presents the ablation study of field loss and offset loss, both designed to guide the attention biases, under the same settings as in Table 2. The results show that introducing the field loss provides better initial values for the offset generator, enhancing the model's performance. Optimal results are achieved only when both losses are incorporated.", + "bbox": [ + 511, + 643, + 906, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of LiDAR frames. Table 4 presents the ablation study conducted on V2V4Real, investigating the impact of different historical frame numbers for both the ego and cooperative vehicle on detection performance and robustness to time delays. It's observed that increasing the ego frames primarily enhances AP70, indicating improved accuracy in object localization, while AP50 shows no significant increase. Conversely, increasing the CV frames enhances both AP50 and the model's robustness to time delays, as most latency originates from the cooperative vehicle's Li", + "bbox": [ + 511, + 750, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12054", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/695ed3c096d1b7b5b1e7404d998ec06bb86258471cf4611290aa867e0e8d4986.jpg", + "image_caption": [ + "Figure 7. Qualitative results in real-world scenarios from the DAIR-V2X-Seq dataset with latency of $400\\mathrm{ms}$ . Green and red boxes denote the ground truths and detection results, respectively." + ], + "image_footnote": [], + "bbox": [ + 106, + 85, + 903, + 281 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d5d1248b7c8bcaa2d1c427ad7851ba727862a3b3bb32a101d5528f036df64954.jpg", + "table_caption": [ + "Table 4. Results of the ablation study of the number of LiDAR frames on the V2V4Real dataset." + ], + "table_footnote": [], + "table_body": "
Ego CVLatency of Cooperative Vehicle (CV)
0ms100ms200ms300ms400ms
1173.35/40.5469.76/34.0366.51/32.5465.17/32.1364.28/32.28
4172.85/44.8169.45/38.3266.73/37.1666.00/37.2865.67/37.32
4273.02/45.0071.74/39.8869.71/38.2567.88/37.4966.98/37.39
1471.28/42.1670.28/39.1869.31/38.1567.68/37.5266.47/36.90
2474.28/44.1473.04/39.9971.51/39.5470.07/38.9969.41/38.49
4474.76/44.0373.28/39.5871.64/38.5870.10/37.7069.08/37.08
", + "bbox": [ + 96, + 382, + 478, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DAR. Furthermore, having too many point cloud frames for the ego vehicle may not yield optimal results, as it could introduce additional motion blur issues, which have been proved in the context of single-vehicle perception [24].", + "bbox": [ + 89, + 522, + 483, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative results", + "text_level": 1, + "bbox": [ + 89, + 590, + 269, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of detection results. Figure 7a-7d compare TraF-Align with three previous methods on a real-world scenario from DAIR-V2X-Seq dataset under a latency of $400~\\mathrm{ms}$ . It is observed that in region $A_{1}$ , due to the latency in roadside (Agent 2), V2VNet, COBEVT, and V2X-ViT all exhibit a false positive detection and fail to capture the correct bounding box; a similar issue occurs in region $A_{2}$ . The reasons for this phenomenon have been discussed in section 1. In contrast, TraF-Align not only identifies the object but also avoids false positives. This improvement is due to our trajectory-guided attention, which enables both agents to achieve a consensus on the same object.", + "bbox": [ + 89, + 613, + 483, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of trajectory field. Figure 6b illustrates the predicted trajectory fields (specifically, the position field) from both the vehicle (Agent 1) and roadside (Agent 2) perspectives. Despite the delays in the roadside point clouds, the predicted trajectory field is aligned to the latest frame.", + "bbox": [ + 89, + 795, + 483, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of attention. Figure 7e shows the response within the $A_{1}$ box of Figure 7d from the roadside perspec", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tive, with six different query positions. Since the ego data is undelayed, the offset distribution centers around the query points, resembling a convolution kernel. In contrast, the roadside agent extends attention along the trajectory. After fusion by TraF-Align, as shown in Figure 6c, the roadside features align toward the vehicles, demonstrating spatial alignment. The feature intensity near the trajectory is uniform, indicating semantic alignment. More results are available in the supplementary material.", + "bbox": [ + 511, + 344, + 906, + 482 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 513, + 496, + 640, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose TraF-Align, a novel cooperative perception framework robust to inter-agent latencies. Our method handles delays by predicting the flow of object features and aligning them across agents through attention within the expected trajectory paths. Extensive experiments on real-world datasets demonstrate its effectiveness.", + "bbox": [ + 511, + 522, + 905, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future work. TraF-Align is primarily trained on data with regular latencies (multiples of 100 ms). In practical applications, more realistic latency settings should be considered. Additionally, while TraF-Align is robust to latency, detection precision under delay can still be improved. The main challenges are: i) Designing more accurate trajectory field predictors and attention position generators to enhance feature reconstruction capabilities; ii) Improving object localization precision. Although multi-frame information improves motion understanding, it also introduces motion blur, making precise object localization challenging and iii) Addressing pose estimation errors. Multi-frame processing relies on accurate inter-frame and inter-agent pose alignment, errors in these estimates can degrade performance. Future work will explore methods to reduce the impact of these factors.", + "bbox": [ + 511, + 613, + 906, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This research is supported by the National Key Research and Development Program of China under Grant 2024YFB2505803.", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12055", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kunyang Cai, Ting Qu, Bingzhao Gao, and Hong Chen. Consensus-based distributed cooperative perception for connected and automated vehicles. IEEE Transactions on Intelligent Transportation Systems, 24(8):8188-8208, 2023. 2", + "[2] Antoine Caillot, Safa Ouerghi, Pascal Vasseur, Rémi Boutteau, and Yohan Dupuis. Survey on cooperative perception in an automotive context. IEEE Transactions on Intelligent Transportation Systems, 23(9):14204-14223, 2022. 1", + "[3] Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. In Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, pages 88-100, 2019. 1, 2, 5, 6", + "[4] Qi Chen, Sihai Tang, Qing Yang, and Song Fu. Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds. In 2019 IEEE 39th International Conference on Distributed Computing Systems (ICDCS), pages 514-524. IEEE, 2019. 1", + "[5] Xuesong Chen, Shaoshuai Shi, Benjamin Zhu, Ka Chun Cheung, Hang Xu, and Hongsheng Li. Mppnet: Multi-frame feature intertwining with proxy points for 3d temporal object detection. In European Conference on Computer Vision, pages 680-697. Springer, 2022. 2", + "[6] Spconv Contributors. Spconv: Spatially sparse convolution library. https://github.com/traveller59/spconv, 2022.4", + "[7] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013. 5", + "[8] Minh-Quan Dao, Julie Stephany Berrio, Vincent Frémont, Mao Shan, Elwan Héry, and Stewart Worrall. Practical collaborative perception: A framework for asynchronous and multi-agent 3d object detection. IEEE Transactions on Intelligent Transportation Systems, 2024. 2", + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015. 5", + "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3", + "[11] Shixin Hong, Yu Liu, Zhi Li, Shaohui Li, and You He. Multiagent collaborative perception via motion-aware robust communication network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15301-15310, 2024. 3, 5, 6", + "[12] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. Advances in neural information processing systems, 35:4874-4886, 2022. 5, 6", + "[13] Tao Huang, Jianan Liu, Xi Zhou, Dinh C Nguyen, Mostafa Rahimi Azghadi, Yuxuan Xia, Qing-Long Han, and Sumei Sun. V2x cooperative perception for autonomous" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "driving: Recent advances and challenges. arXiv preprint arXiv:2310.03525, 2023. 2", + "[14] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 3, 5", + "[15] Zixing Lei, Shunli Ren, Yue Hu, Wenjun Zhang, and Siheng Chen. Latency-aware collaborative perception. In Computer Vision - ECCV 2022, pages 316-332, Cham, 2022. Springer Nature Switzerland. 1, 2", + "[16] Jinyu Li, Chenxu Luo, and Xiaodong Yang. Pillarnext: Rethinking network designs for 3d object detection in lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17567-17576, 2023. 3", + "[17] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5", + "[18] Si Liu, Chen Gao, Yuan Chen, Xingyu Peng, Xianghao Kong, Kun Wang, Runsheng Xu, Wentao Jiang, Hao Xiang, Jiaqi Ma, et al. Towards vehicle-to-everything autonomous driving: A survey on collaborative perception. arXiv preprint arXiv:2308.16714, 2023. 2", + "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 5", + "[20] Yifan Lu, Yue Hu, Yiqi Zhong, Dequan Wang, Yanfeng Wang, and Siheng Chen. An extensible framework for open heterogeneous collaborative perception. In The Twelfth International Conference on Learning Representations, 2024. 1", + "[21] Tao Ma, Xuemeng Yang, Hongbin Zhou, Xin Li, Botian Shi, Junjie Liu, Yuchen Yang, Zhizheng Liu, Liang He, Yu Qiao, et al. Detzero: Rethinking offboard 3d object detection with long-term sequential point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6736-6747, 2023. 2", + "[22] Charles R Qi, Yin Zhou, Mahyar Najibi, Pei Sun, Khoa Vo, Boyang Deng, and Dragomir Anguelov. Offboard 3d object detection from point cloud sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6134-6144, 2021. 2", + "[23] Hang Qiu, Po-Han Huang, Namo Asavisanu, Xiaochen Liu, Konstantinos Psounis, and Ramesh Govindan. Autocast: scalable infrastructure-less cooperative perception for distributed collaborative driving. In Proceedings of the 20th Annual International Conference on Mobile Systems, Applications and Services, pages 128-141, 2022. 1", + "[24] Yao Rong, Xiangyu Wei, Tianwei Lin, Yueyu Wang, and Enkelejda Kasneci. Dynstatf: An efficient feature fusion strategy for lidar 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3237-3246, 2023. 2, 8", + "[25] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12056", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4", + "[26] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 5", + "[27] Richard Sinkhorn and Paul Knopp. Concerning nonnegative matrices and doubly stochastic matrices. Pacific Journal of Mathematics, 21(2):343-348, 1967. 5", + "[28] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, pages 369-386. SPIE, 2019. 5", + "[29] Zhiying Song, Fuxi Wen, Hailiang Zhang, and Jun Li. A cooperative perception system robust to localization errors. In 2023 IEEE Intelligent Vehicles Symposium (IV), pages 1-6. IEEE, 2023. 1", + "[30] Zhiying Song, Tenghui Xie, Hailiang Zhang, Jiaxin Liu, Fuxi Wen, and Jun Li. A spatial calibration method for robust cooperative perception. IEEE Robotics and Automation Letters, 9(5):4011-4018, 2024. 1", + "[31] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4, 5", + "[32] Haiyan Wang and Yingli Tian. Sequential point clouds: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2", + "[33] Tsun-Hsuan Wang, Sivaban Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 605–621. Springer, 2020. 1, 2, 5, 6", + "[34] Sizhe Wei, Yuxi Wei, Yue Hu, Yifan Lu, Yiqi Zhong, Siheng Chen, and Ya Zhang. Asynchrony-robust collaborative perception via bird's eye view flow. In Advances in Neural Information Processing Systems, pages 28462-28477. Curran Associates, Inc., 2023. 1, 2", + "[35] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In European conference on computer vision, pages 107-124. Springer, 2022. 1, 2, 5, 6", + "[36] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 1, 5, 6", + "[37] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view" + ], + "bbox": [ + 91, + 90, + 483, + 883 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "semantic segmentation with sparse transformers. In Conference on Robot Learning, pages 989-1000. PMLR, 2023. 5, 6", + "[38] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13712–13722, 2023. 2, 5", + "[39] Runsheng Xu, Chia-Ju Chen, Zhengzhong Tu, and Ming-Hsuan Yang. V2x-vitv2: Improved vision transformers for vehicle-to-everything cooperative perception. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2", + "[40] Kun Yang, Dingkang Yang, Jingyu Zhang, Mingcheng Li, Yang Liu, Jing Liu, Hanqi Wang, Peng Sun, and Liang Song. Spatio-temporal domain awareness for multi-agent collaborative perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23383-23392, 2023. 3", + "[41] Lei Yang, Kaicheng Yu, Tao Tang, Jun Li, Kun Yuan, Li Wang, Xinyu Zhang, and Peng Chen. Bevheight: A robust framework for vision-based roadside 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21611-21620, 2023. 1", + "[42] Zetong Yang, Yin Zhou, Zhifeng Chen, and Jiquan Ngiam. 3d-man: 3d multi-frame attention network for object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1863-1872, 2021. 2", + "[43] Haibao Yu, Wenxian Yang, Hongzhi Ruan, Zhenwei Yang, Yingjuan Tang, Xu Gao, Xin Hao, Yifeng Shi, Yifeng Pan, Ning Sun, et al. V2x-seq: A large-scale sequential dataset for vehicle-infrastructure cooperative perception and forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5486-5495, 2023. 2, 5", + "[44] Haibao Yu, Yingjuan Tang, Enze Xie, Jilei Mao, Ping Luo, and Zaiqing Nie. Flow-based feature fusion for vehicleinfrastructure cooperative 3d object detection. Advances in Neural Information Processing Systems, 36, 2024. 1, 2", + "[45] Jingyu Zhang, Kun Yang, Yilei Wang, Hanqi Wang, Peng Sun, and Liang Song. ERMVP: Communication-efficient and collaboration-robust multi-vehicle perception in challenging environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12575–12584, 2024. 1, 2, 5, 6", + "[46] Qingzhao Zhang, Xumiao Zhang, Ruiyang Zhu, Fan Bai, Mohammad Naserian, and Z Morley Mao. Robust real-time multi-vehicle collaboration on asynchronous sensors. In Proceedings of the 29th Annual International Conference on Mobile Computing and Networking, pages 1–15, 2023. 2", + "[47] Xumiao Zhang, Anlan Zhang, Jiachen Sun, Xiao Zhu, Y Ethan Guo, Feng Qian, and Z Morley Mao. Emp: Edge-assisted multi-vehicle perception. In Proceedings of the 27th Annual International Conference on Mobile Computing and Networking, pages 545–558, 2021. 1" + ], + "bbox": [ + 516, + 92, + 905, + 883 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12057", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/TraF-Align_ Trajectory-aware Feature Alignment for Asynchronous Multi-agent Perception/0f12efc0-153e-4cda-95dd-b16491feb48e_model.json b/2025/TraF-Align_ Trajectory-aware Feature Alignment for Asynchronous Multi-agent Perception/0f12efc0-153e-4cda-95dd-b16491feb48e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8a63ddf244838f83de43378d6126413de2390aa7 --- /dev/null +++ b/2025/TraF-Align_ Trajectory-aware Feature Alignment for Asynchronous Multi-agent Perception/0f12efc0-153e-4cda-95dd-b16491feb48e_model.json @@ -0,0 +1,1980 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.157, + 0.131, + 0.844, + 0.177 + ], + "angle": 0, + "content": "TraF-Align: Trajectory-aware Feature Alignment for Asynchronous Multi-agent Perception" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.203, + 0.85, + 0.277 + ], + "angle": 0, + "content": "Zhiying Song\\(^{1}\\), Lei Yang\\(^{1}\\), Fuxi Wen\\(^{1,2,\\dagger}\\), Jun Li\\(^{1,2}\\) \n\\(^{1}\\)School of Vehicle and Mobility, Tsinghua University †Corresponding author \n\\(^{2}\\)State Key Lab of Intelligent Green Vehicle and Mobility \n{song-zy24,yanglei20}@ mails.tsinghua.edu.cn, {wenfuxi, lijun1958}@tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.343, + 0.485, + 0.676 + ], + "angle": 0, + "content": "Cooperative perception presents significant potential for enhancing the sensing capabilities of individual vehicles, however, inter-agent latency remains a critical challenge. Latencies cause misalignments in both spatial and semantic features, complicating the fusion of real-time observations from the ego vehicle with delayed data from others. To address these issues, we propose TraF-Align, a novel framework that learns the flow path of features by predicting the feature-level trajectory of objects from past observations up to the ego vehicle's current time. By generating temporally ordered sampling points along these paths, TraF-Align directs attention from the current-time query to relevant historical features along each trajectory, supporting the reconstruction of current-time features and promoting semantic interaction across multiple frames. This approach corrects spatial misalignment and ensures semantic consistency across agents, effectively compensating for motion and achieving coherent feature fusion. Experiments on two real-world datasets, V2V4Real and DAIR-V2X-Seq, show that TraF-Align sets a new benchmark for asynchronous cooperative perception. The code is available at https://github.com/zhyingS/TraF-Align." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.707, + 0.222, + 0.723 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.483, + 0.885 + ], + "angle": 0, + "content": "Intelligent vehicles rely on advanced onboard sensors, such as LiDAR and cameras, to perceive and understand the surrounding world. Nevertheless, the limited field of view and detection range of onboard sensors restrict their ability to perceive objects in occluded areas and at greater distances [2, 23, 36, 47]. Vehicle-to-everything (V2X) communication-enabled cooperative perception systems bridge this gap by exchanging perception information between vehicles and other agents, thereby extending the perception capability of a single agent [3, 4, 20, 41, 45]." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In real-world scenarios, cooperative perception systems" + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.31, + 0.903, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.517, + 0.907, + 0.546 + ], + "angle": 0, + "content": "Figure 1. Asynchronous observations of ego vehicle and other agents result in spatial and semantic misalignments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.907, + 0.793 + ], + "angle": 0, + "content": "are cursed with inter-agent time delays [33, 35]. Consider a network with an ego agent and agent \\( i \\), where their observations of the same object are shown in Figure 1. Latency in agent \\( i \\)'s message, captured before ego's observation time, can result in either a complete loss of features (1-I) or feature misalignment relative to ego's observation (1-II). Corresponding to Figure 1-I and 1-II, there are two main challenges: i) Spatial misalignment, where observations from other agents fail to align with the ground truth in space, lacking any reference from the ego vehicle's observations. ii) Semantic misalignment, where both the ego vehicle and another agent observe the same object but fail to recognize it as the same entity. These misalignments lead to misplaced objects and increased false positives [29, 30]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To address the first challenge, some studies have proposed potential solutions, by predicting the current-frame features using historical frames [15], or predicting the pixel-level flow of features and then extrapolating them [44] or directly warping them to their future positions [34]. For instance, FFNet [44] uses a self-supervised approach to learn the first-order derivatives of features between consecutive" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12048" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.213 + ], + "angle": 0, + "content": "frames, and CoBEVFlow [34] generates an instance-level region of interest (ROI) and calculates feature motion vectors by matching and tracking these regions. However, predicting the movement of high-dimensional neural features is non-trivial. Despite their efforts, these methods struggle with larger delays and lack end-to-end capability. For the second challenge, the significance of semantic misalignment has not been well investigated in previous work." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.218, + 0.487, + 0.671 + ], + "angle": 0, + "content": "To address these misalignments, our core idea is to reconstruct current-time features and achieve semantic consensus across agents on multi-frame LiDAR features within a unified framework. We guide the flow and interaction of high-dimensional features by learning the motion of low-dimensional trajectory features. A trajectory field is defined as a spatiotemporal map of each object's anticipated path, guiding the generation of temporally ordered sampling points along each trajectory. Cross-attention between each query and its aligned sampling points allows the query to access relevant past features along the trajectory, enabling effective feature reconstruction and semantic interaction for robust spatial alignment and inter-agent consensus. Consider the spatial misalignment scenario shown in Figure 1-I. By aggregating multi-frame point cloud features from agent \\( i \\), we learn the object's past trajectory distribution and predict its motion direction up to the ego vehicle's current time, forming a trajectory field. This field guides attention from the predicted position to historical features along the same trajectory, enabling the reconstruction of the object's current features near the ground truth and compensating for motion. For semantic misalignment cases (Figure 1-II), while delays may misalign agent \\( i \\)'s observations with the ego vehicle's, the continuity of object motion implies overlap with point clouds from past ego frames. By learning the trajectory distribution from the ego's multi-frame point clouds and applying attention to compute feature correlations along the trajectory, the model recognizes that all features along the trajectory correspond to the same object, thus achieving cross-agent semantic alignment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Building on the abovementioned idea, we propose TraFAlign, a novel framework for asynchronous multi-agent cooperative perception (as shown in Figure 2). The proposed method includes three key parts: i) a field predictor that generates synchronized trajectory fields aligned with the ego vehicle's time, ii) an offset generator that produces sampling points for each query position and distributes them along the corresponding trajectory, and iii) attention layers where each query position is treated as a token, with attention focused only on positions specified by the offset generator. Extensive experiments are conducted on two widely used large-scale real-world datasets, V2V4Real [38] and DAIR-V2X-Seq [43], one for vehicle-to-vehicle (V2V) cooperative perception, another for vehicle-to-infrastructure (V2I) applications. Experimental results show that state-of" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "the-art (SOTA) performance is achieved for TraF-Align in both object detection precision and robustness to inter-agent latency. The contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.137, + 0.905, + 0.182 + ], + "angle": 0, + "content": "- We propose TraF-Align, a novel framework for asynchronous cooperative perception, addressing spatial and semantic misalignments caused by time delays." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.183, + 0.905, + 0.242 + ], + "angle": 0, + "content": "- We leverage a field predictor and an offset generator to align trajectory fields and distribute attention points along object trajectories, enabling robust feature reconstruction and interaction along spatiotemporally aligned paths." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.243, + 0.905, + 0.288 + ], + "angle": 0, + "content": "- We introduce a field loss and an offset loss to supervise trajectory alignment and the generation of attention points, providing an end-to-end trainable solution." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.137, + 0.905, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.302, + 0.651, + 0.317 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.327, + 0.908, + 0.841 + ], + "angle": 0, + "content": "Asynchronous V2X cooperative perception. V2X cooperative perception extends the sensing range of single vehicles by exchanging information with other agents via wireless communication [13]. Previous studies have conducted extensive research under time-synchronized conditions, such as F-Cooper [3], V2VNet [33], V2X-ViT [35, 39], ERMVP [45], etc. In the presence of latency, previous research emphasizes using historical data to estimate and compensate for object motion. For instance, at the output level, the object state filters are designed to predict and update the dynamic state of the received objects [1]. Using raw data, Zhang et al. [46] and Dao et al. [8] estimated the motion of point clouds and predicted compensatory flows for late-arriving data. However, these levels may encounter challenges, such as short prediction time horizons or high communication data volumes. At the feature level, which has a better balance between accuracy and transmission bandwidth [18], SyncNet [15] employs a pyramidal LSTM to iteratively process and learn from delayed features, enabling the prediction of current-frame features. FFNet [44] predicts inter-frame feature changes in a self-supervised feature manner to synchronize features across agents. Nevertheless, predicting high-dimensional features is challenging and prone to uncontrollable variations. Additionally, CoBEVFlow [34] extracts and warps object-specific features to anticipated locations by predicting instance-wise motion vectors. However, this approach requires generating ROIs and tracking their motion, resulting in a non-end-to-end two-stage method. This paper uses multi-frame LiDAR sequences to predict low-dimensional trajectory features and guide attention to fuse features on past trajectories. By operating directly at the feature level, our approach enables end-to-end training and avoids the complexities of high-dimensional feature prediction." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.841, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Learning from LiDAR sequences across agents. Multiframe point cloud fusion is widely applied in single-vehicle autonomous systems to amalgamate semantic data across different perspectives [5, 21, 22, 24, 32, 42]. However, its" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12049" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.087, + 0.898, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.373, + 0.908, + 0.43 + ], + "angle": 0, + "content": "Figure 2. Proposed Architecture. With asynchronous LiDAR inputs from the ego agent (time \\( t \\)) and agent \\( i \\) (time \\( s = t - \\tau_t^i \\)), BEV features are extracted using an onboard sparse encoder. Agent \\( i \\)'s features are transmitted to the ego agent and stored in memory. TraFAlign predicts a trajectory field up to time \\( t \\), generates target attention positions, and uses features at these positions as keys/values in attention layers. Finally, the multi-agent features are fused and processed by the head to generate cooperative predictions at time \\( t \\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.455, + 0.486, + 0.699 + ], + "angle": 0, + "content": "application in V2X cooperative perception remains under-explored. Although studies like SCOPE [40] and MRC-Net [11] have leveraged recurrent neural networks to fuse sequences from ego vehicles, these methods are resource-intensive and limited by sequence length. This has restricted their use to ego vehicles rather than a broader agent network, necessitating a more straightforward multi-frame fusion network. In cooperative perception scenarios, multiframe point clouds offer crucial motion cues that enable agents to synchronize their understanding of object positions despite time delays, thereby correlating trajectories across agents. This aspect has yet to be fully explored at the feature level. This paper proposes a novel approach using trajectory fields and customized attention with offsets for direct multi-frame fusion, thus establishing a consistent understanding of object motion across the network." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.715, + 0.184, + 0.731 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.742, + 0.291, + 0.756 + ], + "angle": 0, + "content": "3.1. Problem formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Consider a group of networked agents indexed by \\(\\mathcal{I} = \\{e, 1, 2, \\dots, N - 1\\}\\), each is equipped with LiDAR and communication units. Index \\(e\\) indicates the ego agent, which integrates information from other agents to extend its sensing range and reduce the blind spots. The LiDAR observation of the \\(i\\)th agent at its local timestamp \\(s\\) is denoted as \\(L_s^i\\), where \\(i \\in \\mathcal{I}\\). A shared neural network \\(f(\\cdot)\\) processes this observation into deep features. At each timestamp \\(s\\), agent \\(i\\) transmits \\(F_s^i = f(L_s^i)\\) to the ego agent," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.455, + 0.909, + 0.685 + ], + "angle": 0, + "content": "who then accumulates these features in the cache, forming a historical feature set for agent \\(i\\) at time \\(s\\), represented as \\(\\mathcal{F}_s^i = \\{F_{s - k}^i\\}_{k\\in K}\\), where \\(K = \\{k\\in \\mathbb{Z}|1\\leqslant k\\leqslant m_i\\}\\), \\(m_{i}\\) is the maximal sequence length for agent \\(i\\). Accounting for delays in the collaborative process, messages from various agents arrive at ego at different times. Let \\(\\tau_t^i\\) be the delay for agent \\(i\\)'s message, and this means the information sent by agent \\(i\\) at time \\(t - \\tau_t^i\\) is received by the ego at time \\(t\\). If ignoring feature extraction delay, then \\(s = t - \\tau_t^i\\). The task is to fuse these multi-agent features and estimate the state of objects, i.e., the bounding boxes, at ego timestamp \\(t\\) within the detection range. The information available for fusion includes: i) the ego's historical features \\(\\mathcal{F}_t^e\\), and ii) the historical features from other agents, denoted as \\(\\{\\mathcal{F}_{t - \\tau_t^i}^i\\}_{i\\in \\mathcal{T}\\backslash \\{e\\}}\\). The proposed method is shown in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.697, + 0.693, + 0.712 + ], + "angle": 0, + "content": "3.2. Feature extraction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The encoder converts single-frame raw point cloud data into BEV features. Similar to PointPillars [14], the point cloud is first discretized into a regular grid on the \\(x - y\\) plane. Within each pillar, the point features are encapsulated as the vector \\([x,y,z,x_c,y_c,z_c,x_p,y_p,z_p]\\in \\mathbb{R}^9\\). Here, \\([x,y,z]\\) is the point's position, while \\([x_c,y_c,z_c]\\) indicates the center of the pillar. The vector \\([x_p,y_p,z_p]\\) represents the point's offset from the pillar center. The pillar feature extraction and backbone network architectures are adopted from PillarNext [16]. Multi-Layer Perceptron (MLP) serves to encode point-wise features, while a sparse variant of ResNet [10] is utilized as the backbone to distill deeper and more" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12050" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "complex features with size \\((C, H, W)\\). The sparse network is implemented using Spatially Sparse Convolution (SpConv) library [6]. The extracted single-frame features are transmitted and stored in the ego's feature bank \\((\\mathcal{F}_t^e\\) and \\(\\mathcal{F}_s^i)\\), which maintains a fixed capacity by admitting new features and discarding the oldest ones, ensuring an up-to-date repository for feature reuse. The extracted features of each agent are then processed using TraF-Align separately." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.213, + 0.482, + 0.303 + ], + "angle": 0, + "content": "Temporal Encoding. Temporal information is decoupled from the point features to facilitate the reuse of historical features across frames. This allows for the immediate retrieval of historical features at any timestamp. Inspired by the positional embedding in Transformer [31], a temporal embedding \\( TE \\in \\mathbb{R}^{C \\times H \\times W} \\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.314, + 0.482, + 0.332 + ], + "angle": 0, + "content": "\\[\nT E _ {(\\tau , 2 j)} = \\sin (\\xi), T E _ {(\\tau , 2 j + 1)} = \\cos (\\xi), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.341, + 0.483, + 0.478 + ], + "angle": 0, + "content": "where \\(\\xi = \\tau / \\varepsilon^{2j/C}\\), \\(j\\) indexes the dimension along \\(C\\) channels, \\(\\tau\\) represents the number of frames delayed from the current frame, and \\(\\varepsilon\\) is set to eight in this study. Each feature within the memory cache, \\(\\mathcal{F}_t^e\\) for the ego agent and \\(\\mathcal{F}_s^i\\) for agent \\(i\\), associated with a specific temporal delay, is concatenated with a corresponding temporal embedding. These concatenated features are refined through a convolutional layer with a \\(1 \\times 1\\) kernel and then warped to time \\(t\\) by ego-motion compensation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.487, + 0.247, + 0.503 + ], + "angle": 0, + "content": "3.3. Trajectory field" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.509, + 0.483, + 0.75 + ], + "angle": 0, + "content": "Each agent processes its aggregated multi-frame feature map via a field predictor, which generates a trajectory field consisting of position and orientation fields. The position field creates heatmap peaks along the object's trajectory, while the orientation field captures the inverse tangent direction of the trajectory. Due to delays, agents' local observations may be outdated. To achieve temporal alignment, the field predictor predicts the trajectory features of all objects up to the ego vehicle's latest time (with temporal information encoded in the features via Equation 1). This ensures that, for the same object, trajectory fields from different viewpoints spatially overlap and are aligned temporally to the ego vehicle's current time. Our model uses a variant of UNet [25] as the field predictor, whose architecture is detailed in the supplementary material, along with a supervision signal to guide the learning process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.751, + 0.483, + 0.902 + ], + "angle": 0, + "content": "As shown in Figure 3, when generating the ground truth field, we first extract bounding box annotations of each object from the local perspective of each agent within the maximum temporal window, starting from the object's initial appearance up to the ego's current time. A trajectory is then constructed by connecting the centers of these bounding boxes. Specifically, the length of the object's trajectory is given by \\( \\lceil \\tau_i \\cdot \\omega_i \\rceil + m_i \\), \\( \\forall i \\in \\mathcal{I} \\), where \\( \\lceil \\cdot \\rceil \\) denotes the ceiling function, \\( \\omega_i \\) is the LiDAR sampling frequency \\( (\\mathrm{s}^{-1}) \\) of agent \\( i \\), and \\( \\tau_i \\) represents its cooperative latency (s). The" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.094, + 0.9, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.352, + 0.905, + 0.394 + ], + "angle": 0, + "content": "Figure 3. Illustration of trajectory field, which includes a position field indicating the occupancy grid of the trajectory and a direction field depicting the trajectory's flow orientation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.905, + 0.483 + ], + "angle": 0, + "content": "trajectory field is then generated by interpolating along the trajectory and projecting it onto the BEV feature space. If two object trajectories overlap, the trajectory with the more recent timestamp will overwrite the older one." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.906, + 0.621 + ], + "angle": 0, + "content": "For cases in Figure 1-I, such a design compels the network to predict the trajectory field up to the present moment, thereby supporting the learning of motion information. Failing to do so would result in the network detecting only the delayed bounding box. Compared to the direct forecasting of high-dimensional features from the backbone network, predicting the trajectory field provides a more straightforward and more reliable method with predictable outcomes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.636, + 0.757, + 0.651 + ], + "angle": 0, + "content": "3.4. Trajectory-aware attention" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We flatten the input feature map \\( F^i (\\forall i \\in \\mathcal{I}) \\) of size \\( C \\times H \\times W \\) into \\( H \\times W \\) embeddings, each serving as a query \\( q \\in \\mathbb{R}^{1 \\times C} \\). For each query, the offset generator creates a response set \\( R \\in \\mathbb{R}^{n \\times C} \\), capturing features at designated attention positions that are learned based on predefined ground truth. \\( n \\) is the number of attention positions for each query, and it's set to 18 in this paper. To generate the ground truth offsets, we select \\( n \\) pixels on the feature map that satisfy three conditions: i) The position field has a response (indicating the presence of an object), ii) The pixels belong to the same trajectory as the query (indicating the same object), and iii) The pixels correspond to earlier timestamps than the query (indicating historical positions of the object). As shown in Figure 4b, this ground truth aligns each query with preceding trajectory positions, enabling it to draw in relevant historical information. The" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12051" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.213 + ], + "angle": 0, + "content": "trajectory field has effectively modeled the motion and flow of the historical features, offering valuable guidance to the offset generator. With this rich information, the generator can rely on simple convolution layers to produce good attention offsets, and we use PReLU activation [9] to provide flexibility in adjusting the offsets' directional orientation. Given the response set \\( R \\), the feature of query \\( q \\) is updated or reconstructed by the attention process:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.22, + 0.484, + 0.262 + ], + "angle": 0, + "content": "\\[\n\\text {A t t e n t i o n} (q, R) = \\text {S o f t m a x} \\left[ \\frac {q W ^ {q} \\left(R W ^ {k}\\right) ^ {T}}{\\sqrt {d}} \\right] R W ^ {v}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.269, + 0.485, + 0.438 + ], + "angle": 0, + "content": "where \\( W^{q}, W^{k}, W^{v} \\in \\mathbb{R}^{C \\times d} \\) are the query, key and value projection matrices. Similar to the original Transformer [31], we also incorporate feed-forward networks, addition and normalization layers, and a multi-head attention design, the details are shown in Figure 4a. Using attention along the trajectory transfers features to the current time step and enables feature interaction along the trajectory. This allows the model to recognize which features belong to the same object, aligning the features across agents for effective fusion. As a result, each agent aligns its feature map to the ego vehicle's latest time step." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.445, + 0.453, + 0.461 + ], + "angle": 0, + "content": "3.5. Inter-agent fusion, detection head and loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.467, + 0.484, + 0.556 + ], + "angle": 0, + "content": "While inter-agent fusion has been extensively studied [3, 36, 45], this paper adopts a simplified approach by concatenating features from all agents and utilizing a basic convolution layer for fusion. We use an anchor-based detection head and employ three losses: a detection loss, a field loss, and an offset loss:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.567, + 0.482, + 0.582 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {d e t e c t i o n}} + \\alpha \\mathcal {L} _ {\\text {f i e l d}} + \\beta \\mathcal {L} _ {\\text {o f f s e t}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.591, + 0.484, + 0.772 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{detection}}\\) is consistent with the loss used in [14], \\(\\mathcal{L}_{\\mathrm{field}}\\) includes a focal loss [17] for the position field and an \\(L1\\) loss for the orientation field. The constants \\(\\alpha\\) and \\(\\beta\\) balance the contributions of each loss term. Here, we set \\(\\alpha = \\beta = 0.05\\). Considering the offset loss, because the generated attention positions are unordered, we first apply the Sinkhorn algorithm [7, 26, 27] to compute the matching probabilities \\(P_{jk} \\in \\mathbb{R}^{n \\times n}\\) between the generated attention positions and the ground truth. This is done using the L1 distance matrix \\(C_{jk} \\in \\mathbb{R}^{n \\times n}\\) as the cost for the pixel located at \\((j, k)\\) on the feature map. The offset loss is then defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.778, + 0.484, + 0.812 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {o f f s e t}} = \\sum_ {j, k} P _ {j k} \\cdot C _ {j k}, \\forall j \\in [ 1, H ], k \\in [ 1, W ]. \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.831, + 0.224, + 0.848 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To evaluate robustness to cross-agent latency, we conduct 3D object detection experiments on two large-scale real-world datasets, V2V4Real [38] and DAIR-V2X-Seq [43]." + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.088, + 0.905, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.24, + 0.808, + 0.252 + ], + "angle": 0, + "content": "(a) Trajectory-aware attention layers." + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.255, + 0.9, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.587, + 0.396, + 0.835, + 0.41 + ], + "angle": 0, + "content": "(b) Query and ground truth attention positions." + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.421, + 0.905, + 0.436 + ], + "angle": 0, + "content": "Figure 4. Attention positions and trajectory-aware attention layers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.906, + 0.553 + ], + "angle": 0, + "content": "These datasets encompass both V2V and V2I scenarios, as well as sparse and dense LiDAR conditions. The detection results are evaluated using Average Precision (AP) at Intersection-over-Union (IoU) thresholds of 0.5 (AP50) and 0.7 (AP70). To enable a fair comparison, we omit feature compression operations from the methods under evaluation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.562, + 0.729, + 0.578 + ], + "angle": 0, + "content": "4.1. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.704 + ], + "angle": 0, + "content": "Datasets and Benchmarks. V2V4Real [38] is specialized for V2V cooperative detection in urban settings, and DAIR-V2X-Seq [43] captures sequential data in V2I scenarios at intersections. The benchmarks include single-agent perception relying on the ego vehicle's data for detection, alongside eight SOTA intermediate fusion models: F-Cooper [3], V2VNet [33], Where2comm [12], AttFuse [36], V2X-ViT [35], CoBEVT [37], MRCNet [11], and ERMVP [45]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Training details. All the models are trained within the same codebase for 60 epochs on nine RTX 3090 GPUs, employing the one-cycle learning rate strategy [28] with AdamW [19] as the optimizer. The point clouds are segmented at a grid size of \\(0.4 \\times 0.4\\mathrm{m}\\), and the backbone's feature map size output is one-fourth the original size. During training on V2V4Real, all models shuffle the vehicle list and randomly select the ego vehicle at the start of each epoch. Standard LiDAR data augmentations are applied, including random flip, scaling, rotation, and translation. Multi-frame methods are trained with random latency augmentation, following a uniform distribution from 0 to \\(400~\\mathrm{ms}\\). TraF-Align leverages two LiDAR frames for the ego vehicle and four" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12052" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Performance comparison of vehicle class on V2V4Real validation and DAIR-V2X-Seq validation dataset with different latencies. The results are reported in AP 50/70 (%)." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.13, + 0.9, + 0.297 + ], + "angle": 0, + "content": "
MethodV2V4RealDAIR-V2X-Seq
0ms100 ms200 ms300 ms400 ms0ms100 ms200 ms300 ms400 ms
Single Agent48.75/30.5348.75/30.5348.75/30.5348.75/30.5348.75/30.5351.31/40.4551.31/40.4551.31/40.4551.31/40.4551.31/40.45
AttFuse [36]63.84/33.3059.16/27.3655.57/26.0353.25/25.4451.69/25.0968.94/49.6865.83/45.7663.02/44.2461.61/43.2860.19/42.53
F-Cooper [3]66.86/33.3962.15/27.6158.03/26.3055.31/25.4653.92/24.9569.47/50.0165.85/44.9662.74/43.2161.12/42.2860.01/41.70
V2VNet [33]67.11/35.4761.97/29.1458.09/27.8556.26/27.4354.62/26.7571.44/51.9167.26/46.6363.84/44.5961.68/43.2459.78/42.08
Where2comm [12]67.03/37.5561.86/31.5258.22/30.2056.22/29.7154.47/29.0768.37/48.5064.81/44.3261.51/42.8059.66/41.6358.25/40.80
V2X-ViT [35]68.51/35.8863.45/29.5759.42/28.0457.35/27.6555.60/27.2070.21/50.9266.42/46.4462.58/44.2961.00/43.5659.63/42.66
MRCNet [11]67.83/37.3762.65/31.5958.98/30.3956.70/29.5355.03/28.9469.27/50.6765.90/45.6062.65/44.1560.09/42.3958.19/41.35
CoBEVT [37]70.59/38.8265.50/32.1161.55/30.4058.91/29.5457.05/29.2671.51/51.3167.60/46.5564.38/45.0362.39/44.1160.74/43.08
ERMVP [45]70.74/38.4565.53/31.9161.63/30.5359.24/30.1857.26/29.3171.19/53.8567.04/48.0963.67/46.3761.23/44.7659.55/43.93
TraF-Align (Ours)74.28/44.1473.04/39.9971.51/39.5470.07/38.9969.41/38.4976.90/58.3173.95/53.6872.69/52.8371.66/52.3371.22/52.01
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.323, + 0.482, + 0.383 + ], + "angle": 0, + "content": "for the cooperative vehicle on V2V4Real, while using four LiDAR frames for the vehicle and infrastructure on DAIR-V2X-Seq. The number of attention layers is set to two, each containing four attention heads." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.384, + 0.483, + 0.476 + ], + "angle": 0, + "content": "Latency settings. To comprehensively evaluate robustness to time delays, we introduce delays of 0, 100, 200, 300, and \\(400\\mathrm{ms}\\) to one of the agents separately, extensively testing the model's performance under each delay condition. Unless specified otherwise, these delays are applied to the point cloud data of the non-ego agent." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.485, + 0.383, + 0.501 + ], + "angle": 0, + "content": "4.2. Comparison with state-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.508, + 0.483, + 0.688 + ], + "angle": 0, + "content": "Detection performance. Table 1 presents a comparison of detection performance on two real-world datasets for vehicle class, which includes cars, vans, trucks, buses, etc. It's observed that the proposed method i) achieves state-of-the-art performance across all benchmarks under various latency conditions on both datasets. ii) In synchronous scenarios, i.e., when the delay is \\(0\\mathrm{ms}\\), our model improves AP 50/70 by \\(25.53\\% / 13.61\\%\\) (on V2V4Real) and \\(25.59\\% / 18.26\\%\\) (on DAIR-V2X-Seq) compared to single-agent perception, it also surpasses ERMVP, the SOTA performance of feature-level methods, by \\(3.54\\% / 5.69\\%\\) and \\(5.71\\% / 4.46\\%\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.689, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Robustness to inter-agent latency. As shown in Table 1, our model demonstrates stronger robustness in the presence of time delays. Even under a \\(400~\\mathrm{ms}\\) delay, our AP decreases by only \\(4.87\\% /5.65\\%\\) (on V2V4Real) and \\(5.68\\% /6.30\\%\\) (on DAIR-V2X-Seq) compared to the synchronous scenario. This performance significantly outperforms the ERMVP, which shows declines of \\(13.48\\% /9.14\\%\\) and \\(11.64\\% /9.92\\%\\) , respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To gain a deeper understanding of the impact of latency on collaboration, we plot the Precision-Recall curves of the models under various latency conditions, as shown in Figure 5. It is obvious that cooperation is effective in enhancing the recall limit of single-vehicle perception, helping the ego vehicle to detect a more significant number of objects." + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.322, + 0.705, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.321, + 0.897, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.485, + 0.705, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.485, + 0.898, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.685, + 0.908, + 0.714 + ], + "angle": 0, + "content": "Figure 5. Precision-Recall curves for models on the V2V4Real validation dataset under various latency conditions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "However, at a \\(200\\mathrm{ms}\\) latency, state-of-the-art methods exhibit lower precision in the low-recall range compared to single-vehicle perception. This suggests that the delayed collaborative information interferes with the objects that the single vehicle could detect, corresponding to the semantic misalignment issues depicted in Figure 1-II. In areas not visible to the ego vehicle, specifically at high recall levels, the precision of SOTA methods significantly declines with the introduction of latency. This indicates that spatial misalignment issues (Figure 1-I) become dominant, lead" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12053" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.288 + ], + "angle": 0, + "content": "ing to decreased performance. These findings reinforce our analysis of latency's impact on cooperative perception. Furthermore, using ERMVP as an example, the differences between the PR curves diminish as latency increases, indicating that existing models progressively lose their effectiveness in managing delays. In contrast, TraF-Align retains the original performance of single-agent perception in lowrecall regions and shows no significant decline in performance in high-recall areas due to latency. This highlights our model's ability to establish strong semantic consistency between the ego vehicle's data and the received information, demonstrating robust reconstruction capability and effectively addressing the scenarios in Figures 1-I and 1-II." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.301, + 0.253, + 0.316 + ], + "angle": 0, + "content": "4.3. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.484, + 0.4 + ], + "angle": 0, + "content": "In this section, we conducted ablation studies on the effectiveness of significant components and loss and multiframe point clouds using average precision as the evaluation metric. Other ablation experiments and analyses (using PR curves) are presented in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.401, + 0.484, + 0.643 + ], + "angle": 0, + "content": "Effectiveness of Major Components. To evaluate the contribution of modules, we progressively remove i) the field predictor, ii) the offset generator, and iii) the attention layers. The results are shown in Table 2. The experiments are conducted on the DAIR-V2X-Seq dataset. To ensure a comprehensive assessment of each component's contributions, we train the model for 120 epochs, double the duration of the experiments presented in Table 1. Additionally, we test the effects of adding latency to both the ego and non-ego agents to fully evaluate the impact of each module. The consistent rise in detection precision over different latency settings demonstrates the effectiveness of each introduced component. Notably, applying latency to vehicles results in lower AP than when latency is applied to infrastructure. This discrepancy arises from the different distributions of object visibility between the vehicle agent and roadside in" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.668, + 0.483, + 0.696 + ], + "angle": 0, + "content": "Table 2. Results of the ablation study of the core components on the DAIR-V2X-Seq dataset." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.708, + 0.482, + 0.883 + ], + "angle": 0, + "content": "
FP OG ALLatency of Infrastructure
0ms100ms200ms300ms400ms
75.35/54.7273.01/50.8171.88/50.1771.04/49.7770.09/49.45
✓76.31/57.0274.60/53.5373.90/52.3472.91/51.99
✓77.11/58.6375.15/54.7774.18/53.7773.24/52.99
✓✓77.83/59.7275.74/55.8574.99/54.82
FP OG ALLatency of Vehicle
0ms100ms200ms300ms400ms
75.35/54.7274.31/49.6172.17/43.3267.19/34.8256.93/24.51
✓76.31/57.0275.10/51.9372.91/46.2269.22/39.15
✓77.11/58.6375.76/53.4873.44/47.5569.92/40.52
✓✓77.83/59.7276.41/54.8874.24/49.46
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.111, + 0.885, + 0.452, + 0.898 + ], + "angle": 0, + "content": "FP: Field predictor. OG: Offset generator. AL: Attention layers." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.118 + ], + "angle": 0, + "content": "Table 3. Results of the ablation study of the loss on the DAIR-V2X-Seq dataset." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.129, + 0.905, + 0.304 + ], + "angle": 0, + "content": "
FL OLLatency of Infrastructure
0ms100ms200ms300ms400ms
✓76.43/57.7474.44/53.9173.65/52.9772.71/52.1071.83/51.87
77.12/58.3674.96/54.5174.02/53.6872.98/52.9172.01/52.27
✓77.15/58.7075.19/54.5874.31/53.6673.68/53.0772.52/52.28
✓77.83/59.7275.74/55.8574.99/54.8274.10/54.0772.89/53.21
FL OLLatency of Vehicle
0ms100ms200ms300ms400ms
✓76.43/57.7475.09/52.8172.84/46.5469.14/39.1261.53/29.15
77.12/58.3675.69/53.2372.76/46.8569.25/39.7862.13/30.41
✓77.15/58.7075.80/53.9073.34/48.4970.12/41.8063.21/32.02
✓77.83/59.7276.41/54.8874.24/49.4670.75/42.6263.59/32.31
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.536, + 0.306, + 0.704, + 0.319 + ], + "angle": 0, + "content": "FL:Field loss. OL: Offset loss." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.336, + 0.634, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.547, + 0.626, + 0.56 + ], + "angle": 0, + "content": "(a) LiDAR points." + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.336, + 0.747, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.643, + 0.548, + 0.746, + 0.56 + ], + "angle": 0, + "content": "(b) Trajectory field." + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.336, + 0.882, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.779, + 0.548, + 0.874, + 0.56 + ], + "angle": 0, + "content": "(c) Deep features." + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.572, + 0.907, + 0.6 + ], + "angle": 0, + "content": "Figure 6. Visualization of TraF-Align in real-world scenarios from the DAIR-V2X-Seq dataset with latency of \\(400~\\mathrm{ms}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.629, + 0.728, + 0.643 + ], + "angle": 0, + "content": "frastructure agent in the dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.748 + ], + "angle": 0, + "content": "Effectiveness of loss. Table 3 presents the ablation study of field loss and offset loss, both designed to guide the attention biases, under the same settings as in Table 2. The results show that introducing the field loss provides better initial values for the offset generator, enhancing the model's performance. Optimal results are achieved only when both losses are incorporated." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Impact of LiDAR frames. Table 4 presents the ablation study conducted on V2V4Real, investigating the impact of different historical frame numbers for both the ego and cooperative vehicle on detection performance and robustness to time delays. It's observed that increasing the ego frames primarily enhances AP70, indicating improved accuracy in object localization, while AP50 shows no significant increase. Conversely, increasing the CV frames enhances both AP50 and the model's robustness to time delays, as most latency originates from the cooperative vehicle's Li" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12054" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.087, + 0.905, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.291, + 0.908, + 0.32 + ], + "angle": 0, + "content": "Figure 7. Qualitative results in real-world scenarios from the DAIR-V2X-Seq dataset with latency of \\(400\\mathrm{ms}\\). Green and red boxes denote the ground truths and detection results, respectively." + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.343, + 0.483, + 0.371 + ], + "angle": 0, + "content": "Table 4. Results of the ablation study of the number of LiDAR frames on the V2V4Real dataset." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.383, + 0.48, + 0.5 + ], + "angle": 0, + "content": "
Ego CVLatency of Cooperative Vehicle (CV)
0ms100ms200ms300ms400ms
1173.35/40.5469.76/34.0366.51/32.5465.17/32.1364.28/32.28
4172.85/44.8169.45/38.3266.73/37.1666.00/37.2865.67/37.32
4273.02/45.0071.74/39.8869.71/38.2567.88/37.4966.98/37.39
1471.28/42.1670.28/39.1869.31/38.1567.68/37.5266.47/36.90
2474.28/44.1473.04/39.9971.51/39.5470.07/38.9969.41/38.49
4474.76/44.0373.28/39.5871.64/38.5870.10/37.7069.08/37.08
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.523, + 0.484, + 0.584 + ], + "angle": 0, + "content": "DAR. Furthermore, having too many point cloud frames for the ego vehicle may not yield optimal results, as it could introduce additional motion blur issues, which have been proved in the context of single-vehicle perception [24]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.592, + 0.27, + 0.608 + ], + "angle": 0, + "content": "4.4. Qualitative results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.795 + ], + "angle": 0, + "content": "Visualization of detection results. Figure 7a-7d compare TraF-Align with three previous methods on a real-world scenario from DAIR-V2X-Seq dataset under a latency of \\(400~\\mathrm{ms}\\). It is observed that in region \\(A_{1}\\), due to the latency in roadside (Agent 2), V2VNet, COBEVT, and V2X-ViT all exhibit a false positive detection and fail to capture the correct bounding box; a similar issue occurs in region \\(A_{2}\\). The reasons for this phenomenon have been discussed in section 1. In contrast, TraF-Align not only identifies the object but also avoids false positives. This improvement is due to our trajectory-guided attention, which enables both agents to achieve a consensus on the same object." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.871 + ], + "angle": 0, + "content": "Visualization of trajectory field. Figure 6b illustrates the predicted trajectory fields (specifically, the position field) from both the vehicle (Agent 1) and roadside (Agent 2) perspectives. Despite the delays in the roadside point clouds, the predicted trajectory field is aligned to the latest frame." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Visualization of attention. Figure 7e shows the response within the \\(A_{1}\\) box of Figure 7d from the roadside perspec" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.345, + 0.907, + 0.483 + ], + "angle": 0, + "content": "tive, with six different query positions. Since the ego data is undelayed, the offset distribution centers around the query points, resembling a convolution kernel. In contrast, the roadside agent extends attention along the trajectory. After fusion by TraF-Align, as shown in Figure 6c, the roadside features align toward the vehicles, demonstrating spatial alignment. The feature intensity near the trajectory is uniform, indicating semantic alignment. More results are available in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.497, + 0.642, + 0.513 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.906, + 0.612 + ], + "angle": 0, + "content": "We propose TraF-Align, a novel cooperative perception framework robust to inter-agent latencies. Our method handles delays by predicting the flow of object features and aligning them across agents through attention within the expected trajectory paths. Extensive experiments on real-world datasets demonstrate its effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.854 + ], + "angle": 0, + "content": "Limitations and future work. TraF-Align is primarily trained on data with regular latencies (multiples of 100 ms). In practical applications, more realistic latency settings should be considered. Additionally, while TraF-Align is robust to latency, detection precision under delay can still be improved. The main challenges are: i) Designing more accurate trajectory field predictors and attention position generators to enhance feature reconstruction capabilities; ii) Improving object localization precision. Although multi-frame information improves motion understanding, it also introduces motion blur, making precise object localization challenging and iii) Addressing pose estimation errors. Multi-frame processing relies on accurate inter-frame and inter-agent pose alignment, errors in these estimates can degrade performance. Future work will explore methods to reduce the impact of these factors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments. This research is supported by the National Key Research and Development Program of China under Grant 2024YFB2505803." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12055" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.171 + ], + "angle": 0, + "content": "[1] Kunyang Cai, Ting Qu, Bingzhao Gao, and Hong Chen. Consensus-based distributed cooperative perception for connected and automated vehicles. IEEE Transactions on Intelligent Transportation Systems, 24(8):8188-8208, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.483, + 0.227 + ], + "angle": 0, + "content": "[2] Antoine Caillot, Safa Ouerghi, Pascal Vasseur, Rémi Boutteau, and Yohan Dupuis. Survey on cooperative perception in an automotive context. IEEE Transactions on Intelligent Transportation Systems, 23(9):14204-14223, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.483, + 0.297 + ], + "angle": 0, + "content": "[3] Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. In Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, pages 88-100, 2019. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.482, + 0.367 + ], + "angle": 0, + "content": "[4] Qi Chen, Sihai Tang, Qing Yang, and Song Fu. Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds. In 2019 IEEE 39th International Conference on Distributed Computing Systems (ICDCS), pages 514-524. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.482, + 0.437 + ], + "angle": 0, + "content": "[5] Xuesong Chen, Shaoshuai Shi, Benjamin Zhu, Ka Chun Cheung, Hang Xu, and Hongsheng Li. Mppnet: Multi-frame feature intertwining with proxy points for 3d temporal object detection. In European Conference on Computer Vision, pages 680-697. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.482, + 0.479 + ], + "angle": 0, + "content": "[6] Spconv Contributors. Spconv: Spatially sparse convolution library. https://github.com/traveller59/spconv, 2022.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.482, + 0.522 + ], + "angle": 0, + "content": "[7] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.482, + 0.592 + ], + "angle": 0, + "content": "[8] Minh-Quan Dao, Julie Stephany Berrio, Vincent Frémont, Mao Shan, Elwan Héry, and Stewart Worrall. Practical collaborative perception: A framework for asynchronous and multi-agent 3d object detection. IEEE Transactions on Intelligent Transportation Systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.482, + 0.661 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[11] Shixin Hong, Yu Liu, Zhi Li, Shaohui Li, and You He. Multiagent collaborative perception via motion-aware robust communication network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15301-15310, 2024. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[12] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. Advances in neural information processing systems, 35:4874-4886, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[13] Tao Huang, Jianan Liu, Xi Zhou, Dinh C Nguyen, Mostafa Rahimi Azghadi, Yuxuan Xia, Qing-Long Han, and Sumei Sun. V2x cooperative perception for autonomous" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "driving: Recent advances and challenges. arXiv preprint arXiv:2310.03525, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.907, + 0.191 + ], + "angle": 0, + "content": "[14] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.192, + 0.906, + 0.247 + ], + "angle": 0, + "content": "[15] Zixing Lei, Shunli Ren, Yue Hu, Wenjun Zhang, and Siheng Chen. Latency-aware collaborative perception. In Computer Vision - ECCV 2022, pages 316-332, Cham, 2022. Springer Nature Switzerland. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.906, + 0.318 + ], + "angle": 0, + "content": "[16] Jinyu Li, Chenxu Luo, and Xiaodong Yang. Pillarnext: Rethinking network designs for 3d object detection in lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17567-17576, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.906, + 0.376 + ], + "angle": 0, + "content": "[17] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.377, + 0.906, + 0.446 + ], + "angle": 0, + "content": "[18] Si Liu, Chen Gao, Yuan Chen, Xingyu Peng, Xianghao Kong, Kun Wang, Runsheng Xu, Wentao Jiang, Hao Xiang, Jiaqi Ma, et al. Towards vehicle-to-everything autonomous driving: A survey on collaborative perception. arXiv preprint arXiv:2308.16714, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.477, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[20] Yifan Lu, Yue Hu, Yiqi Zhong, Dequan Wang, Yanfeng Wang, and Siheng Chen. An extensible framework for open heterogeneous collaborative perception. In The Twelfth International Conference on Learning Representations, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.548, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[21] Tao Ma, Xuemeng Yang, Hongbin Zhou, Xin Li, Botian Shi, Junjie Liu, Yuchen Yang, Zhizheng Liu, Liang He, Yu Qiao, et al. Detzero: Rethinking offboard 3d object detection with long-term sequential point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6736-6747, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.906, + 0.702 + ], + "angle": 0, + "content": "[22] Charles R Qi, Yin Zhou, Mahyar Najibi, Pei Sun, Khoa Vo, Boyang Deng, and Dragomir Anguelov. Offboard 3d object detection from point cloud sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6134-6144, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[23] Hang Qiu, Po-Han Huang, Namo Asavisanu, Xiaochen Liu, Konstantinos Psounis, and Ramesh Govindan. Autocast: scalable infrastructure-less cooperative perception for distributed collaborative driving. In Proceedings of the 20th Annual International Conference on Mobile Systems, Applications and Services, pages 128-141, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.788, + 0.906, + 0.857 + ], + "angle": 0, + "content": "[24] Yao Rong, Xiangyu Wei, Tianwei Lin, Yueyu Wang, and Enkelejda Kasneci. Dynstatf: An efficient feature fusion strategy for lidar 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3237-3246, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[25] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12056" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.484, + 0.135 + ], + "angle": 0, + "content": "intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.207 + ], + "angle": 0, + "content": "[26] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.484, + 0.251 + ], + "angle": 0, + "content": "[27] Richard Sinkhorn and Paul Knopp. Concerning nonnegative matrices and doubly stochastic matrices. Pacific Journal of Mathematics, 21(2):343-348, 1967. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.252, + 0.483, + 0.322 + ], + "angle": 0, + "content": "[28] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, pages 369-386. SPIE, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.323, + 0.483, + 0.379 + ], + "angle": 0, + "content": "[29] Zhiying Song, Fuxi Wen, Hailiang Zhang, and Jun Li. A cooperative perception system robust to localization errors. In 2023 IEEE Intelligent Vehicles Symposium (IV), pages 1-6. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.381, + 0.483, + 0.437 + ], + "angle": 0, + "content": "[30] Zhiying Song, Tenghui Xie, Hailiang Zhang, Jiaxin Liu, Fuxi Wen, and Jun Li. A spatial calibration method for robust cooperative perception. IEEE Robotics and Automation Letters, 9(5):4011-4018, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.439, + 0.483, + 0.495 + ], + "angle": 0, + "content": "[31] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.483, + 0.54 + ], + "angle": 0, + "content": "[32] Haiyan Wang and Yingli Tian. Sequential point clouds: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.541, + 0.483, + 0.638 + ], + "angle": 0, + "content": "[33] Tsun-Hsuan Wang, Sivaban Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 605–621. Springer, 2020. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.64, + 0.483, + 0.71 + ], + "angle": 0, + "content": "[34] Sizhe Wei, Yuxi Wei, Yue Hu, Yifan Lu, Yiqi Zhong, Siheng Chen, and Ya Zhang. Asynchrony-robust collaborative perception via bird's eye view flow. In Advances in Neural Information Processing Systems, pages 28462-28477. Curran Associates, Inc., 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.712, + 0.483, + 0.782 + ], + "angle": 0, + "content": "[35] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In European conference on computer vision, pages 107-124. Springer, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.784, + 0.483, + 0.853 + ], + "angle": 0, + "content": "[36] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.855, + 0.483, + 0.884 + ], + "angle": 0, + "content": "[37] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.884 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.133 + ], + "angle": 0, + "content": "semantic segmentation with sparse transformers. In Conference on Robot Learning, pages 989-1000. PMLR, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[38] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13712–13722, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.906, + 0.286 + ], + "angle": 0, + "content": "[39] Runsheng Xu, Chia-Ju Chen, Zhengzhong Tu, and Ming-Hsuan Yang. V2x-vitv2: Improved vision transformers for vehicle-to-everything cooperative perception. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.906, + 0.37 + ], + "angle": 0, + "content": "[40] Kun Yang, Dingkang Yang, Jingyu Zhang, Mingcheng Li, Yang Liu, Jing Liu, Hanqi Wang, Peng Sun, and Liang Song. Spatio-temporal domain awareness for multi-agent collaborative perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23383-23392, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.371, + 0.906, + 0.44 + ], + "angle": 0, + "content": "[41] Lei Yang, Kaicheng Yu, Tao Tang, Jun Li, Kun Yuan, Li Wang, Xinyu Zhang, and Peng Chen. Bevheight: A robust framework for vision-based roadside 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21611-21620, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.441, + 0.906, + 0.508 + ], + "angle": 0, + "content": "[42] Zetong Yang, Yin Zhou, Zhifeng Chen, and Jiquan Ngiam. 3d-man: 3d multi-frame attention network for object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1863-1872, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.509, + 0.906, + 0.606 + ], + "angle": 0, + "content": "[43] Haibao Yu, Wenxian Yang, Hongzhi Ruan, Zhenwei Yang, Yingjuan Tang, Xu Gao, Xin Hao, Yifeng Shi, Yifeng Pan, Ning Sun, et al. V2x-seq: A large-scale sequential dataset for vehicle-infrastructure cooperative perception and forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5486-5495, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.607, + 0.906, + 0.662 + ], + "angle": 0, + "content": "[44] Haibao Yu, Yingjuan Tang, Enze Xie, Jilei Mao, Ping Luo, and Zaiqing Nie. Flow-based feature fusion for vehicleinfrastructure cooperative 3d object detection. Advances in Neural Information Processing Systems, 36, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.662, + 0.906, + 0.745 + ], + "angle": 0, + "content": "[45] Jingyu Zhang, Kun Yang, Yilei Wang, Hanqi Wang, Peng Sun, and Liang Song. ERMVP: Communication-efficient and collaboration-robust multi-vehicle perception in challenging environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12575–12584, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.745, + 0.906, + 0.814 + ], + "angle": 0, + "content": "[46] Qingzhao Zhang, Xumiao Zhang, Ruiyang Zhu, Fan Bai, Mohammad Naserian, and Z Morley Mao. Robust real-time multi-vehicle collaboration on asynchronous sensors. In Proceedings of the 29th Annual International Conference on Mobile Computing and Networking, pages 1–15, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.814, + 0.906, + 0.884 + ], + "angle": 0, + "content": "[47] Xumiao Zhang, Anlan Zhang, Jiachen Sun, Xiao Zhu, Y Ethan Guo, Feng Qian, and Z Morley Mao. Emp: Edge-assisted multi-vehicle perception. In Proceedings of the 27th Annual International Conference on Mobile Computing and Networking, pages 545–558, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.884 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12057" + } + ] +] \ No newline at end of file diff --git a/2025/TraF-Align_ Trajectory-aware Feature Alignment for Asynchronous Multi-agent Perception/0f12efc0-153e-4cda-95dd-b16491feb48e_origin.pdf b/2025/TraF-Align_ Trajectory-aware Feature Alignment for Asynchronous Multi-agent Perception/0f12efc0-153e-4cda-95dd-b16491feb48e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bc4fd8ec96c2ddb0fcedd7dbcae4e2e9cb225eab --- /dev/null +++ b/2025/TraF-Align_ Trajectory-aware Feature Alignment for Asynchronous Multi-agent Perception/0f12efc0-153e-4cda-95dd-b16491feb48e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed7c3e8901619fd05c750692e9dbe24b7c4ea5dc009d3bd51973179cbb2b090f +size 7566799 diff --git a/2025/TraF-Align_ Trajectory-aware Feature Alignment for Asynchronous Multi-agent Perception/full.md b/2025/TraF-Align_ Trajectory-aware Feature Alignment for Asynchronous Multi-agent Perception/full.md new file mode 100644 index 0000000000000000000000000000000000000000..98acfc6e23f25212192974363bf6b16b4137b1ec --- /dev/null +++ b/2025/TraF-Align_ Trajectory-aware Feature Alignment for Asynchronous Multi-agent Perception/full.md @@ -0,0 +1,271 @@ +# TraF-Align: Trajectory-aware Feature Alignment for Asynchronous Multi-agent Perception + +Zhiying Song $^{1}$ , Lei Yang $^{1}$ , Fuxi Wen $^{1,2,\dagger}$ , Jun Li $^{1,2}$ $^{1}$ School of Vehicle and Mobility, Tsinghua University †Corresponding author + $^{2}$ State Key Lab of Intelligent Green Vehicle and Mobility +{song-zy24,yanglei20}@ mails.tsinghua.edu.cn, {wenfuxi, lijun1958}@tsinghua.edu.cn + +# Abstract + +Cooperative perception presents significant potential for enhancing the sensing capabilities of individual vehicles, however, inter-agent latency remains a critical challenge. Latencies cause misalignments in both spatial and semantic features, complicating the fusion of real-time observations from the ego vehicle with delayed data from others. To address these issues, we propose TraF-Align, a novel framework that learns the flow path of features by predicting the feature-level trajectory of objects from past observations up to the ego vehicle's current time. By generating temporally ordered sampling points along these paths, TraF-Align directs attention from the current-time query to relevant historical features along each trajectory, supporting the reconstruction of current-time features and promoting semantic interaction across multiple frames. This approach corrects spatial misalignment and ensures semantic consistency across agents, effectively compensating for motion and achieving coherent feature fusion. Experiments on two real-world datasets, V2V4Real and DAIR-V2X-Seq, show that TraF-Align sets a new benchmark for asynchronous cooperative perception. The code is available at https://github.com/zhyingS/TraF-Align. + +# 1. Introduction + +Intelligent vehicles rely on advanced onboard sensors, such as LiDAR and cameras, to perceive and understand the surrounding world. Nevertheless, the limited field of view and detection range of onboard sensors restrict their ability to perceive objects in occluded areas and at greater distances [2, 23, 36, 47]. Vehicle-to-everything (V2X) communication-enabled cooperative perception systems bridge this gap by exchanging perception information between vehicles and other agents, thereby extending the perception capability of a single agent [3, 4, 20, 41, 45]. + +In real-world scenarios, cooperative perception systems + +![](images/f3d4dfeee0aaaa6a9eee9851008ffb564a0a3cc14d5fed9de5113269dbcf1eaf.jpg) +Figure 1. Asynchronous observations of ego vehicle and other agents result in spatial and semantic misalignments. + +are cursed with inter-agent time delays [33, 35]. Consider a network with an ego agent and agent $i$ , where their observations of the same object are shown in Figure 1. Latency in agent $i$ 's message, captured before ego's observation time, can result in either a complete loss of features (1-I) or feature misalignment relative to ego's observation (1-II). Corresponding to Figure 1-I and 1-II, there are two main challenges: i) Spatial misalignment, where observations from other agents fail to align with the ground truth in space, lacking any reference from the ego vehicle's observations. ii) Semantic misalignment, where both the ego vehicle and another agent observe the same object but fail to recognize it as the same entity. These misalignments lead to misplaced objects and increased false positives [29, 30]. + +To address the first challenge, some studies have proposed potential solutions, by predicting the current-frame features using historical frames [15], or predicting the pixel-level flow of features and then extrapolating them [44] or directly warping them to their future positions [34]. For instance, FFNet [44] uses a self-supervised approach to learn the first-order derivatives of features between consecutive + +frames, and CoBEVFlow [34] generates an instance-level region of interest (ROI) and calculates feature motion vectors by matching and tracking these regions. However, predicting the movement of high-dimensional neural features is non-trivial. Despite their efforts, these methods struggle with larger delays and lack end-to-end capability. For the second challenge, the significance of semantic misalignment has not been well investigated in previous work. + +To address these misalignments, our core idea is to reconstruct current-time features and achieve semantic consensus across agents on multi-frame LiDAR features within a unified framework. We guide the flow and interaction of high-dimensional features by learning the motion of low-dimensional trajectory features. A trajectory field is defined as a spatiotemporal map of each object's anticipated path, guiding the generation of temporally ordered sampling points along each trajectory. Cross-attention between each query and its aligned sampling points allows the query to access relevant past features along the trajectory, enabling effective feature reconstruction and semantic interaction for robust spatial alignment and inter-agent consensus. Consider the spatial misalignment scenario shown in Figure 1-I. By aggregating multi-frame point cloud features from agent $i$ , we learn the object's past trajectory distribution and predict its motion direction up to the ego vehicle's current time, forming a trajectory field. This field guides attention from the predicted position to historical features along the same trajectory, enabling the reconstruction of the object's current features near the ground truth and compensating for motion. For semantic misalignment cases (Figure 1-II), while delays may misalign agent $i$ 's observations with the ego vehicle's, the continuity of object motion implies overlap with point clouds from past ego frames. By learning the trajectory distribution from the ego's multi-frame point clouds and applying attention to compute feature correlations along the trajectory, the model recognizes that all features along the trajectory correspond to the same object, thus achieving cross-agent semantic alignment. + +Building on the abovementioned idea, we propose TraFAlign, a novel framework for asynchronous multi-agent cooperative perception (as shown in Figure 2). The proposed method includes three key parts: i) a field predictor that generates synchronized trajectory fields aligned with the ego vehicle's time, ii) an offset generator that produces sampling points for each query position and distributes them along the corresponding trajectory, and iii) attention layers where each query position is treated as a token, with attention focused only on positions specified by the offset generator. Extensive experiments are conducted on two widely used large-scale real-world datasets, V2V4Real [38] and DAIR-V2X-Seq [43], one for vehicle-to-vehicle (V2V) cooperative perception, another for vehicle-to-infrastructure (V2I) applications. Experimental results show that state-of + +the-art (SOTA) performance is achieved for TraF-Align in both object detection precision and robustness to inter-agent latency. The contributions are summarized as follows: + +- We propose TraF-Align, a novel framework for asynchronous cooperative perception, addressing spatial and semantic misalignments caused by time delays. +- We leverage a field predictor and an offset generator to align trajectory fields and distribute attention points along object trajectories, enabling robust feature reconstruction and interaction along spatiotemporally aligned paths. +- We introduce a field loss and an offset loss to supervise trajectory alignment and the generation of attention points, providing an end-to-end trainable solution. + +# 2. Related work + +Asynchronous V2X cooperative perception. V2X cooperative perception extends the sensing range of single vehicles by exchanging information with other agents via wireless communication [13]. Previous studies have conducted extensive research under time-synchronized conditions, such as F-Cooper [3], V2VNet [33], V2X-ViT [35, 39], ERMVP [45], etc. In the presence of latency, previous research emphasizes using historical data to estimate and compensate for object motion. For instance, at the output level, the object state filters are designed to predict and update the dynamic state of the received objects [1]. Using raw data, Zhang et al. [46] and Dao et al. [8] estimated the motion of point clouds and predicted compensatory flows for late-arriving data. However, these levels may encounter challenges, such as short prediction time horizons or high communication data volumes. At the feature level, which has a better balance between accuracy and transmission bandwidth [18], SyncNet [15] employs a pyramidal LSTM to iteratively process and learn from delayed features, enabling the prediction of current-frame features. FFNet [44] predicts inter-frame feature changes in a self-supervised feature manner to synchronize features across agents. Nevertheless, predicting high-dimensional features is challenging and prone to uncontrollable variations. Additionally, CoBEVFlow [34] extracts and warps object-specific features to anticipated locations by predicting instance-wise motion vectors. However, this approach requires generating ROIs and tracking their motion, resulting in a non-end-to-end two-stage method. This paper uses multi-frame LiDAR sequences to predict low-dimensional trajectory features and guide attention to fuse features on past trajectories. By operating directly at the feature level, our approach enables end-to-end training and avoids the complexities of high-dimensional feature prediction. + +Learning from LiDAR sequences across agents. Multiframe point cloud fusion is widely applied in single-vehicle autonomous systems to amalgamate semantic data across different perspectives [5, 21, 22, 24, 32, 42]. However, its + +![](images/2cdf8216f524cb86daeb1fdb5422271a921dc7b3ef7ca977ca25426faa07647e.jpg) +Figure 2. Proposed Architecture. With asynchronous LiDAR inputs from the ego agent (time $t$ ) and agent $i$ (time $s = t - \tau_t^i$ ), BEV features are extracted using an onboard sparse encoder. Agent $i$ 's features are transmitted to the ego agent and stored in memory. TraFAlign predicts a trajectory field up to time $t$ , generates target attention positions, and uses features at these positions as keys/values in attention layers. Finally, the multi-agent features are fused and processed by the head to generate cooperative predictions at time $t$ . + +application in V2X cooperative perception remains under-explored. Although studies like SCOPE [40] and MRC-Net [11] have leveraged recurrent neural networks to fuse sequences from ego vehicles, these methods are resource-intensive and limited by sequence length. This has restricted their use to ego vehicles rather than a broader agent network, necessitating a more straightforward multi-frame fusion network. In cooperative perception scenarios, multiframe point clouds offer crucial motion cues that enable agents to synchronize their understanding of object positions despite time delays, thereby correlating trajectories across agents. This aspect has yet to be fully explored at the feature level. This paper proposes a novel approach using trajectory fields and customized attention with offsets for direct multi-frame fusion, thus establishing a consistent understanding of object motion across the network. + +# 3. Method + +# 3.1. Problem formulation + +Consider a group of networked agents indexed by $\mathcal{I} = \{e, 1, 2, \dots, N - 1\}$ , each is equipped with LiDAR and communication units. Index $e$ indicates the ego agent, which integrates information from other agents to extend its sensing range and reduce the blind spots. The LiDAR observation of the $i$ th agent at its local timestamp $s$ is denoted as $L_s^i$ , where $i \in \mathcal{I}$ . A shared neural network $f(\cdot)$ processes this observation into deep features. At each timestamp $s$ , agent $i$ transmits $F_s^i = f(L_s^i)$ to the ego agent, + +who then accumulates these features in the cache, forming a historical feature set for agent $i$ at time $s$ , represented as $\mathcal{F}_s^i = \{F_{s - k}^i\}_{k\in K}$ , where $K = \{k\in \mathbb{Z}|1\leqslant k\leqslant m_i\}$ , $m_{i}$ is the maximal sequence length for agent $i$ . Accounting for delays in the collaborative process, messages from various agents arrive at ego at different times. Let $\tau_t^i$ be the delay for agent $i$ 's message, and this means the information sent by agent $i$ at time $t - \tau_t^i$ is received by the ego at time $t$ . If ignoring feature extraction delay, then $s = t - \tau_t^i$ . The task is to fuse these multi-agent features and estimate the state of objects, i.e., the bounding boxes, at ego timestamp $t$ within the detection range. The information available for fusion includes: i) the ego's historical features $\mathcal{F}_t^e$ , and ii) the historical features from other agents, denoted as $\{\mathcal{F}_{t - \tau_t^i}^i\}_{i\in \mathcal{T}\backslash \{e\}}$ . The proposed method is shown in Figure 2. + +# 3.2. Feature extraction + +The encoder converts single-frame raw point cloud data into BEV features. Similar to PointPillars [14], the point cloud is first discretized into a regular grid on the $x - y$ plane. Within each pillar, the point features are encapsulated as the vector $[x,y,z,x_c,y_c,z_c,x_p,y_p,z_p]\in \mathbb{R}^9$ . Here, $[x,y,z]$ is the point's position, while $[x_c,y_c,z_c]$ indicates the center of the pillar. The vector $[x_p,y_p,z_p]$ represents the point's offset from the pillar center. The pillar feature extraction and backbone network architectures are adopted from PillarNext [16]. Multi-Layer Perceptron (MLP) serves to encode point-wise features, while a sparse variant of ResNet [10] is utilized as the backbone to distill deeper and more + +complex features with size $(C, H, W)$ . The sparse network is implemented using Spatially Sparse Convolution (SpConv) library [6]. The extracted single-frame features are transmitted and stored in the ego's feature bank $(\mathcal{F}_t^e$ and $\mathcal{F}_s^i)$ , which maintains a fixed capacity by admitting new features and discarding the oldest ones, ensuring an up-to-date repository for feature reuse. The extracted features of each agent are then processed using TraF-Align separately. + +Temporal Encoding. Temporal information is decoupled from the point features to facilitate the reuse of historical features across frames. This allows for the immediate retrieval of historical features at any timestamp. Inspired by the positional embedding in Transformer [31], a temporal embedding $TE \in \mathbb{R}^{C \times H \times W}$ is defined as + +$$ +T E _ {(\tau , 2 j)} = \sin (\xi), T E _ {(\tau , 2 j + 1)} = \cos (\xi), \tag {1} +$$ + +where $\xi = \tau / \varepsilon^{2j/C}$ , $j$ indexes the dimension along $C$ channels, $\tau$ represents the number of frames delayed from the current frame, and $\varepsilon$ is set to eight in this study. Each feature within the memory cache, $\mathcal{F}_t^e$ for the ego agent and $\mathcal{F}_s^i$ for agent $i$ , associated with a specific temporal delay, is concatenated with a corresponding temporal embedding. These concatenated features are refined through a convolutional layer with a $1 \times 1$ kernel and then warped to time $t$ by ego-motion compensation. + +# 3.3. Trajectory field + +Each agent processes its aggregated multi-frame feature map via a field predictor, which generates a trajectory field consisting of position and orientation fields. The position field creates heatmap peaks along the object's trajectory, while the orientation field captures the inverse tangent direction of the trajectory. Due to delays, agents' local observations may be outdated. To achieve temporal alignment, the field predictor predicts the trajectory features of all objects up to the ego vehicle's latest time (with temporal information encoded in the features via Equation 1). This ensures that, for the same object, trajectory fields from different viewpoints spatially overlap and are aligned temporally to the ego vehicle's current time. Our model uses a variant of UNet [25] as the field predictor, whose architecture is detailed in the supplementary material, along with a supervision signal to guide the learning process. + +As shown in Figure 3, when generating the ground truth field, we first extract bounding box annotations of each object from the local perspective of each agent within the maximum temporal window, starting from the object's initial appearance up to the ego's current time. A trajectory is then constructed by connecting the centers of these bounding boxes. Specifically, the length of the object's trajectory is given by $\lceil \tau_i \cdot \omega_i \rceil + m_i$ , $\forall i \in \mathcal{I}$ , where $\lceil \cdot \rceil$ denotes the ceiling function, $\omega_i$ is the LiDAR sampling frequency $(\mathrm{s}^{-1})$ of agent $i$ , and $\tau_i$ represents its cooperative latency (s). The + +![](images/0af345f65626421913b5a14c7b1c3fb65639dfffa01dae3308eeaa52744f6a10.jpg) +Figure 3. Illustration of trajectory field, which includes a position field indicating the occupancy grid of the trajectory and a direction field depicting the trajectory's flow orientation. + +trajectory field is then generated by interpolating along the trajectory and projecting it onto the BEV feature space. If two object trajectories overlap, the trajectory with the more recent timestamp will overwrite the older one. + +For cases in Figure 1-I, such a design compels the network to predict the trajectory field up to the present moment, thereby supporting the learning of motion information. Failing to do so would result in the network detecting only the delayed bounding box. Compared to the direct forecasting of high-dimensional features from the backbone network, predicting the trajectory field provides a more straightforward and more reliable method with predictable outcomes. + +# 3.4. Trajectory-aware attention + +We flatten the input feature map $F^i (\forall i \in \mathcal{I})$ of size $C \times H \times W$ into $H \times W$ embeddings, each serving as a query $q \in \mathbb{R}^{1 \times C}$ . For each query, the offset generator creates a response set $R \in \mathbb{R}^{n \times C}$ , capturing features at designated attention positions that are learned based on predefined ground truth. $n$ is the number of attention positions for each query, and it's set to 18 in this paper. To generate the ground truth offsets, we select $n$ pixels on the feature map that satisfy three conditions: i) The position field has a response (indicating the presence of an object), ii) The pixels belong to the same trajectory as the query (indicating the same object), and iii) The pixels correspond to earlier timestamps than the query (indicating historical positions of the object). As shown in Figure 4b, this ground truth aligns each query with preceding trajectory positions, enabling it to draw in relevant historical information. The + +trajectory field has effectively modeled the motion and flow of the historical features, offering valuable guidance to the offset generator. With this rich information, the generator can rely on simple convolution layers to produce good attention offsets, and we use PReLU activation [9] to provide flexibility in adjusting the offsets' directional orientation. Given the response set $R$ , the feature of query $q$ is updated or reconstructed by the attention process: + +$$ +\text {A t t e n t i o n} (q, R) = \text {S o f t m a x} \left[ \frac {q W ^ {q} \left(R W ^ {k}\right) ^ {T}}{\sqrt {d}} \right] R W ^ {v}, \tag {2} +$$ + +where $W^{q}, W^{k}, W^{v} \in \mathbb{R}^{C \times d}$ are the query, key and value projection matrices. Similar to the original Transformer [31], we also incorporate feed-forward networks, addition and normalization layers, and a multi-head attention design, the details are shown in Figure 4a. Using attention along the trajectory transfers features to the current time step and enables feature interaction along the trajectory. This allows the model to recognize which features belong to the same object, aligning the features across agents for effective fusion. As a result, each agent aligns its feature map to the ego vehicle's latest time step. + +# 3.5. Inter-agent fusion, detection head and loss + +While inter-agent fusion has been extensively studied [3, 36, 45], this paper adopts a simplified approach by concatenating features from all agents and utilizing a basic convolution layer for fusion. We use an anchor-based detection head and employ three losses: a detection loss, a field loss, and an offset loss: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {d e t e c t i o n}} + \alpha \mathcal {L} _ {\text {f i e l d}} + \beta \mathcal {L} _ {\text {o f f s e t}}, \tag {3} +$$ + +where $\mathcal{L}_{\mathrm{detection}}$ is consistent with the loss used in [14], $\mathcal{L}_{\mathrm{field}}$ includes a focal loss [17] for the position field and an $L1$ loss for the orientation field. The constants $\alpha$ and $\beta$ balance the contributions of each loss term. Here, we set $\alpha = \beta = 0.05$ . Considering the offset loss, because the generated attention positions are unordered, we first apply the Sinkhorn algorithm [7, 26, 27] to compute the matching probabilities $P_{jk} \in \mathbb{R}^{n \times n}$ between the generated attention positions and the ground truth. This is done using the L1 distance matrix $C_{jk} \in \mathbb{R}^{n \times n}$ as the cost for the pixel located at $(j, k)$ on the feature map. The offset loss is then defined as: + +$$ +\mathcal {L} _ {\text {o f f s e t}} = \sum_ {j, k} P _ {j k} \cdot C _ {j k}, \forall j \in [ 1, H ], k \in [ 1, W ]. \tag {4} +$$ + +# 4. Experiments + +To evaluate robustness to cross-agent latency, we conduct 3D object detection experiments on two large-scale real-world datasets, V2V4Real [38] and DAIR-V2X-Seq [43]. + +![](images/55564513160de392774c28db019896855f74f3937e31077014e26ee013505a5b.jpg) +(a) Trajectory-aware attention layers. + +![](images/e47348146adec98c21b18cdd1b1ad9c97eedf19a6115575a1a39490bceb3a2f8.jpg) +(b) Query and ground truth attention positions. +Figure 4. Attention positions and trajectory-aware attention layers. + +These datasets encompass both V2V and V2I scenarios, as well as sparse and dense LiDAR conditions. The detection results are evaluated using Average Precision (AP) at Intersection-over-Union (IoU) thresholds of 0.5 (AP50) and 0.7 (AP70). To enable a fair comparison, we omit feature compression operations from the methods under evaluation. + +# 4.1. Implementation details + +Datasets and Benchmarks. V2V4Real [38] is specialized for V2V cooperative detection in urban settings, and DAIR-V2X-Seq [43] captures sequential data in V2I scenarios at intersections. The benchmarks include single-agent perception relying on the ego vehicle's data for detection, alongside eight SOTA intermediate fusion models: F-Cooper [3], V2VNet [33], Where2comm [12], AttFuse [36], V2X-ViT [35], CoBEVT [37], MRCNet [11], and ERMVP [45]. + +Training details. All the models are trained within the same codebase for 60 epochs on nine RTX 3090 GPUs, employing the one-cycle learning rate strategy [28] with AdamW [19] as the optimizer. The point clouds are segmented at a grid size of $0.4 \times 0.4\mathrm{m}$ , and the backbone's feature map size output is one-fourth the original size. During training on V2V4Real, all models shuffle the vehicle list and randomly select the ego vehicle at the start of each epoch. Standard LiDAR data augmentations are applied, including random flip, scaling, rotation, and translation. Multi-frame methods are trained with random latency augmentation, following a uniform distribution from 0 to $400~\mathrm{ms}$ . TraF-Align leverages two LiDAR frames for the ego vehicle and four + +Table 1. Performance comparison of vehicle class on V2V4Real validation and DAIR-V2X-Seq validation dataset with different latencies. The results are reported in AP 50/70 (%). + +
MethodV2V4RealDAIR-V2X-Seq
0ms100 ms200 ms300 ms400 ms0ms100 ms200 ms300 ms400 ms
Single Agent48.75/30.5348.75/30.5348.75/30.5348.75/30.5348.75/30.5351.31/40.4551.31/40.4551.31/40.4551.31/40.4551.31/40.45
AttFuse [36]63.84/33.3059.16/27.3655.57/26.0353.25/25.4451.69/25.0968.94/49.6865.83/45.7663.02/44.2461.61/43.2860.19/42.53
F-Cooper [3]66.86/33.3962.15/27.6158.03/26.3055.31/25.4653.92/24.9569.47/50.0165.85/44.9662.74/43.2161.12/42.2860.01/41.70
V2VNet [33]67.11/35.4761.97/29.1458.09/27.8556.26/27.4354.62/26.7571.44/51.9167.26/46.6363.84/44.5961.68/43.2459.78/42.08
Where2comm [12]67.03/37.5561.86/31.5258.22/30.2056.22/29.7154.47/29.0768.37/48.5064.81/44.3261.51/42.8059.66/41.6358.25/40.80
V2X-ViT [35]68.51/35.8863.45/29.5759.42/28.0457.35/27.6555.60/27.2070.21/50.9266.42/46.4462.58/44.2961.00/43.5659.63/42.66
MRCNet [11]67.83/37.3762.65/31.5958.98/30.3956.70/29.5355.03/28.9469.27/50.6765.90/45.6062.65/44.1560.09/42.3958.19/41.35
CoBEVT [37]70.59/38.8265.50/32.1161.55/30.4058.91/29.5457.05/29.2671.51/51.3167.60/46.5564.38/45.0362.39/44.1160.74/43.08
ERMVP [45]70.74/38.4565.53/31.9161.63/30.5359.24/30.1857.26/29.3171.19/53.8567.04/48.0963.67/46.3761.23/44.7659.55/43.93
TraF-Align (Ours)74.28/44.1473.04/39.9971.51/39.5470.07/38.9969.41/38.4976.90/58.3173.95/53.6872.69/52.8371.66/52.3371.22/52.01
+ +for the cooperative vehicle on V2V4Real, while using four LiDAR frames for the vehicle and infrastructure on DAIR-V2X-Seq. The number of attention layers is set to two, each containing four attention heads. + +Latency settings. To comprehensively evaluate robustness to time delays, we introduce delays of 0, 100, 200, 300, and $400\mathrm{ms}$ to one of the agents separately, extensively testing the model's performance under each delay condition. Unless specified otherwise, these delays are applied to the point cloud data of the non-ego agent. + +# 4.2. Comparison with state-of-the-art + +Detection performance. Table 1 presents a comparison of detection performance on two real-world datasets for vehicle class, which includes cars, vans, trucks, buses, etc. It's observed that the proposed method i) achieves state-of-the-art performance across all benchmarks under various latency conditions on both datasets. ii) In synchronous scenarios, i.e., when the delay is $0\mathrm{ms}$ , our model improves AP 50/70 by $25.53\% / 13.61\%$ (on V2V4Real) and $25.59\% / 18.26\%$ (on DAIR-V2X-Seq) compared to single-agent perception, it also surpasses ERMVP, the SOTA performance of feature-level methods, by $3.54\% / 5.69\%$ and $5.71\% / 4.46\%$ , respectively. + +Robustness to inter-agent latency. As shown in Table 1, our model demonstrates stronger robustness in the presence of time delays. Even under a $400~\mathrm{ms}$ delay, our AP decreases by only $4.87\% /5.65\%$ (on V2V4Real) and $5.68\% /6.30\%$ (on DAIR-V2X-Seq) compared to the synchronous scenario. This performance significantly outperforms the ERMVP, which shows declines of $13.48\% /9.14\%$ and $11.64\% /9.92\%$ , respectively. + +To gain a deeper understanding of the impact of latency on collaboration, we plot the Precision-Recall curves of the models under various latency conditions, as shown in Figure 5. It is obvious that cooperation is effective in enhancing the recall limit of single-vehicle perception, helping the ego vehicle to detect a more significant number of objects. + +![](images/d7d5e17242f79c59992fe9cc8f2b2a1a2073234a6b1571b037c6d9470d18cc6c.jpg) + +![](images/33902ffe0d7955c3ac3b544de806c2fa363eab9a2ef4149b0a5e5d9f1ded6ee2.jpg) + +![](images/1a9ad1243c0ef0fbc8106297e8e9cc52ee1392f15c426150cc3e9b800a8dee69.jpg) +Figure 5. Precision-Recall curves for models on the V2V4Real validation dataset under various latency conditions. + +![](images/dd396af796c57a7a09b42fec6a53f420d55d851cace7b4f4551f0fcc0ef62558.jpg) + +However, at a $200\mathrm{ms}$ latency, state-of-the-art methods exhibit lower precision in the low-recall range compared to single-vehicle perception. This suggests that the delayed collaborative information interferes with the objects that the single vehicle could detect, corresponding to the semantic misalignment issues depicted in Figure 1-II. In areas not visible to the ego vehicle, specifically at high recall levels, the precision of SOTA methods significantly declines with the introduction of latency. This indicates that spatial misalignment issues (Figure 1-I) become dominant, lead + +ing to decreased performance. These findings reinforce our analysis of latency's impact on cooperative perception. Furthermore, using ERMVP as an example, the differences between the PR curves diminish as latency increases, indicating that existing models progressively lose their effectiveness in managing delays. In contrast, TraF-Align retains the original performance of single-agent perception in lowrecall regions and shows no significant decline in performance in high-recall areas due to latency. This highlights our model's ability to establish strong semantic consistency between the ego vehicle's data and the received information, demonstrating robust reconstruction capability and effectively addressing the scenarios in Figures 1-I and 1-II. + +# 4.3. Ablation studies + +In this section, we conducted ablation studies on the effectiveness of significant components and loss and multiframe point clouds using average precision as the evaluation metric. Other ablation experiments and analyses (using PR curves) are presented in the supplementary material. + +Effectiveness of Major Components. To evaluate the contribution of modules, we progressively remove i) the field predictor, ii) the offset generator, and iii) the attention layers. The results are shown in Table 2. The experiments are conducted on the DAIR-V2X-Seq dataset. To ensure a comprehensive assessment of each component's contributions, we train the model for 120 epochs, double the duration of the experiments presented in Table 1. Additionally, we test the effects of adding latency to both the ego and non-ego agents to fully evaluate the impact of each module. The consistent rise in detection precision over different latency settings demonstrates the effectiveness of each introduced component. Notably, applying latency to vehicles results in lower AP than when latency is applied to infrastructure. This discrepancy arises from the different distributions of object visibility between the vehicle agent and roadside in + +Table 2. Results of the ablation study of the core components on the DAIR-V2X-Seq dataset. + +
FP OG ALLatency of Infrastructure
0ms100ms200ms300ms400ms
75.35/54.7273.01/50.8171.88/50.1771.04/49.7770.09/49.45
✓76.31/57.0274.60/53.5373.90/52.3472.91/51.99
✓77.11/58.6375.15/54.7774.18/53.7773.24/52.99
✓✓77.83/59.7275.74/55.8574.99/54.82
FP OG ALLatency of Vehicle
0ms100ms200ms300ms400ms
75.35/54.7274.31/49.6172.17/43.3267.19/34.8256.93/24.51
✓76.31/57.0275.10/51.9372.91/46.2269.22/39.15
✓77.11/58.6375.76/53.4873.44/47.5569.92/40.52
✓✓77.83/59.7276.41/54.8874.24/49.46
+ +FP: Field predictor. OG: Offset generator. AL: Attention layers. + +Table 3. Results of the ablation study of the loss on the DAIR-V2X-Seq dataset. + +
FL OLLatency of Infrastructure
0ms100ms200ms300ms400ms
✓76.43/57.7474.44/53.9173.65/52.9772.71/52.1071.83/51.87
77.12/58.3674.96/54.5174.02/53.6872.98/52.9172.01/52.27
✓77.15/58.7075.19/54.5874.31/53.6673.68/53.0772.52/52.28
✓77.83/59.7275.74/55.8574.99/54.8274.10/54.0772.89/53.21
FL OLLatency of Vehicle
0ms100ms200ms300ms400ms
✓76.43/57.7475.09/52.8172.84/46.5469.14/39.1261.53/29.15
77.12/58.3675.69/53.2372.76/46.8569.25/39.7862.13/30.41
✓77.15/58.7075.80/53.9073.34/48.4970.12/41.8063.21/32.02
✓77.83/59.7276.41/54.8874.24/49.4670.75/42.6263.59/32.31
+ +FL:Field loss. OL: Offset loss. + +![](images/730d661a4453700fb5be3e7a5cb5543a368244299c1e2f39368ab5567b3902f5.jpg) +(a) LiDAR points. + +![](images/4f8815cb30ce6da13f277e6a32c79dbc91b6687d022eb8dc87d423dcd62c34f6.jpg) +(b) Trajectory field. +Figure 6. Visualization of TraF-Align in real-world scenarios from the DAIR-V2X-Seq dataset with latency of $400~\mathrm{ms}$ . + +![](images/a3b0ddecd3e40ab65b9a8bfcd3225a97fcdd0821f18f7e2c632b84b4ac408f3b.jpg) +(c) Deep features. + +frastructure agent in the dataset. + +Effectiveness of loss. Table 3 presents the ablation study of field loss and offset loss, both designed to guide the attention biases, under the same settings as in Table 2. The results show that introducing the field loss provides better initial values for the offset generator, enhancing the model's performance. Optimal results are achieved only when both losses are incorporated. + +Impact of LiDAR frames. Table 4 presents the ablation study conducted on V2V4Real, investigating the impact of different historical frame numbers for both the ego and cooperative vehicle on detection performance and robustness to time delays. It's observed that increasing the ego frames primarily enhances AP70, indicating improved accuracy in object localization, while AP50 shows no significant increase. Conversely, increasing the CV frames enhances both AP50 and the model's robustness to time delays, as most latency originates from the cooperative vehicle's Li + +![](images/695ed3c096d1b7b5b1e7404d998ec06bb86258471cf4611290aa867e0e8d4986.jpg) +Figure 7. Qualitative results in real-world scenarios from the DAIR-V2X-Seq dataset with latency of $400\mathrm{ms}$ . Green and red boxes denote the ground truths and detection results, respectively. + +Table 4. Results of the ablation study of the number of LiDAR frames on the V2V4Real dataset. + +
Ego CVLatency of Cooperative Vehicle (CV)
0ms100ms200ms300ms400ms
1173.35/40.5469.76/34.0366.51/32.5465.17/32.1364.28/32.28
4172.85/44.8169.45/38.3266.73/37.1666.00/37.2865.67/37.32
4273.02/45.0071.74/39.8869.71/38.2567.88/37.4966.98/37.39
1471.28/42.1670.28/39.1869.31/38.1567.68/37.5266.47/36.90
2474.28/44.1473.04/39.9971.51/39.5470.07/38.9969.41/38.49
4474.76/44.0373.28/39.5871.64/38.5870.10/37.7069.08/37.08
+ +DAR. Furthermore, having too many point cloud frames for the ego vehicle may not yield optimal results, as it could introduce additional motion blur issues, which have been proved in the context of single-vehicle perception [24]. + +# 4.4. Qualitative results + +Visualization of detection results. Figure 7a-7d compare TraF-Align with three previous methods on a real-world scenario from DAIR-V2X-Seq dataset under a latency of $400~\mathrm{ms}$ . It is observed that in region $A_{1}$ , due to the latency in roadside (Agent 2), V2VNet, COBEVT, and V2X-ViT all exhibit a false positive detection and fail to capture the correct bounding box; a similar issue occurs in region $A_{2}$ . The reasons for this phenomenon have been discussed in section 1. In contrast, TraF-Align not only identifies the object but also avoids false positives. This improvement is due to our trajectory-guided attention, which enables both agents to achieve a consensus on the same object. + +Visualization of trajectory field. Figure 6b illustrates the predicted trajectory fields (specifically, the position field) from both the vehicle (Agent 1) and roadside (Agent 2) perspectives. Despite the delays in the roadside point clouds, the predicted trajectory field is aligned to the latest frame. + +Visualization of attention. Figure 7e shows the response within the $A_{1}$ box of Figure 7d from the roadside perspec + +tive, with six different query positions. Since the ego data is undelayed, the offset distribution centers around the query points, resembling a convolution kernel. In contrast, the roadside agent extends attention along the trajectory. After fusion by TraF-Align, as shown in Figure 6c, the roadside features align toward the vehicles, demonstrating spatial alignment. The feature intensity near the trajectory is uniform, indicating semantic alignment. More results are available in the supplementary material. + +# 5. Conclusions + +We propose TraF-Align, a novel cooperative perception framework robust to inter-agent latencies. Our method handles delays by predicting the flow of object features and aligning them across agents through attention within the expected trajectory paths. Extensive experiments on real-world datasets demonstrate its effectiveness. + +Limitations and future work. TraF-Align is primarily trained on data with regular latencies (multiples of 100 ms). In practical applications, more realistic latency settings should be considered. Additionally, while TraF-Align is robust to latency, detection precision under delay can still be improved. The main challenges are: i) Designing more accurate trajectory field predictors and attention position generators to enhance feature reconstruction capabilities; ii) Improving object localization precision. Although multi-frame information improves motion understanding, it also introduces motion blur, making precise object localization challenging and iii) Addressing pose estimation errors. Multi-frame processing relies on accurate inter-frame and inter-agent pose alignment, errors in these estimates can degrade performance. Future work will explore methods to reduce the impact of these factors. + +Acknowledgments. This research is supported by the National Key Research and Development Program of China under Grant 2024YFB2505803. + +# References + +[1] Kunyang Cai, Ting Qu, Bingzhao Gao, and Hong Chen. Consensus-based distributed cooperative perception for connected and automated vehicles. IEEE Transactions on Intelligent Transportation Systems, 24(8):8188-8208, 2023. 2 +[2] Antoine Caillot, Safa Ouerghi, Pascal Vasseur, Rémi Boutteau, and Yohan Dupuis. Survey on cooperative perception in an automotive context. IEEE Transactions on Intelligent Transportation Systems, 23(9):14204-14223, 2022. 1 +[3] Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. In Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, pages 88-100, 2019. 1, 2, 5, 6 +[4] Qi Chen, Sihai Tang, Qing Yang, and Song Fu. Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds. In 2019 IEEE 39th International Conference on Distributed Computing Systems (ICDCS), pages 514-524. IEEE, 2019. 1 +[5] Xuesong Chen, Shaoshuai Shi, Benjamin Zhu, Ka Chun Cheung, Hang Xu, and Hongsheng Li. Mppnet: Multi-frame feature intertwining with proxy points for 3d temporal object detection. In European Conference on Computer Vision, pages 680-697. Springer, 2022. 2 +[6] Spconv Contributors. Spconv: Spatially sparse convolution library. https://github.com/traveller59/spconv, 2022.4 +[7] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013. 5 +[8] Minh-Quan Dao, Julie Stephany Berrio, Vincent Frémont, Mao Shan, Elwan Héry, and Stewart Worrall. Practical collaborative perception: A framework for asynchronous and multi-agent 3d object detection. IEEE Transactions on Intelligent Transportation Systems, 2024. 2 +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015. 5 +[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3 +[11] Shixin Hong, Yu Liu, Zhi Li, Shaohui Li, and You He. Multiagent collaborative perception via motion-aware robust communication network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15301-15310, 2024. 3, 5, 6 +[12] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. Advances in neural information processing systems, 35:4874-4886, 2022. 5, 6 +[13] Tao Huang, Jianan Liu, Xi Zhou, Dinh C Nguyen, Mostafa Rahimi Azghadi, Yuxuan Xia, Qing-Long Han, and Sumei Sun. V2x cooperative perception for autonomous + +driving: Recent advances and challenges. arXiv preprint arXiv:2310.03525, 2023. 2 +[14] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 3, 5 +[15] Zixing Lei, Shunli Ren, Yue Hu, Wenjun Zhang, and Siheng Chen. Latency-aware collaborative perception. In Computer Vision - ECCV 2022, pages 316-332, Cham, 2022. Springer Nature Switzerland. 1, 2 +[16] Jinyu Li, Chenxu Luo, and Xiaodong Yang. Pillarnext: Rethinking network designs for 3d object detection in lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17567-17576, 2023. 3 +[17] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5 +[18] Si Liu, Chen Gao, Yuan Chen, Xingyu Peng, Xianghao Kong, Kun Wang, Runsheng Xu, Wentao Jiang, Hao Xiang, Jiaqi Ma, et al. Towards vehicle-to-everything autonomous driving: A survey on collaborative perception. arXiv preprint arXiv:2308.16714, 2023. 2 +[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 5 +[20] Yifan Lu, Yue Hu, Yiqi Zhong, Dequan Wang, Yanfeng Wang, and Siheng Chen. An extensible framework for open heterogeneous collaborative perception. In The Twelfth International Conference on Learning Representations, 2024. 1 +[21] Tao Ma, Xuemeng Yang, Hongbin Zhou, Xin Li, Botian Shi, Junjie Liu, Yuchen Yang, Zhizheng Liu, Liang He, Yu Qiao, et al. Detzero: Rethinking offboard 3d object detection with long-term sequential point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6736-6747, 2023. 2 +[22] Charles R Qi, Yin Zhou, Mahyar Najibi, Pei Sun, Khoa Vo, Boyang Deng, and Dragomir Anguelov. Offboard 3d object detection from point cloud sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6134-6144, 2021. 2 +[23] Hang Qiu, Po-Han Huang, Namo Asavisanu, Xiaochen Liu, Konstantinos Psounis, and Ramesh Govindan. Autocast: scalable infrastructure-less cooperative perception for distributed collaborative driving. In Proceedings of the 20th Annual International Conference on Mobile Systems, Applications and Services, pages 128-141, 2022. 1 +[24] Yao Rong, Xiangyu Wei, Tianwei Lin, Yueyu Wang, and Enkelejda Kasneci. Dynstatf: An efficient feature fusion strategy for lidar 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3237-3246, 2023. 2, 8 +[25] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted + +intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4 +[26] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 5 +[27] Richard Sinkhorn and Paul Knopp. Concerning nonnegative matrices and doubly stochastic matrices. Pacific Journal of Mathematics, 21(2):343-348, 1967. 5 +[28] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, pages 369-386. SPIE, 2019. 5 +[29] Zhiying Song, Fuxi Wen, Hailiang Zhang, and Jun Li. A cooperative perception system robust to localization errors. In 2023 IEEE Intelligent Vehicles Symposium (IV), pages 1-6. IEEE, 2023. 1 +[30] Zhiying Song, Tenghui Xie, Hailiang Zhang, Jiaxin Liu, Fuxi Wen, and Jun Li. A spatial calibration method for robust cooperative perception. IEEE Robotics and Automation Letters, 9(5):4011-4018, 2024. 1 +[31] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4, 5 +[32] Haiyan Wang and Yingli Tian. Sequential point clouds: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 +[33] Tsun-Hsuan Wang, Sivaban Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 605–621. Springer, 2020. 1, 2, 5, 6 +[34] Sizhe Wei, Yuxi Wei, Yue Hu, Yifan Lu, Yiqi Zhong, Siheng Chen, and Ya Zhang. Asynchrony-robust collaborative perception via bird's eye view flow. In Advances in Neural Information Processing Systems, pages 28462-28477. Curran Associates, Inc., 2023. 1, 2 +[35] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In European conference on computer vision, pages 107-124. Springer, 2022. 1, 2, 5, 6 +[36] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 1, 5, 6 +[37] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view + +semantic segmentation with sparse transformers. In Conference on Robot Learning, pages 989-1000. PMLR, 2023. 5, 6 +[38] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13712–13722, 2023. 2, 5 +[39] Runsheng Xu, Chia-Ju Chen, Zhengzhong Tu, and Ming-Hsuan Yang. V2x-vitv2: Improved vision transformers for vehicle-to-everything cooperative perception. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 +[40] Kun Yang, Dingkang Yang, Jingyu Zhang, Mingcheng Li, Yang Liu, Jing Liu, Hanqi Wang, Peng Sun, and Liang Song. Spatio-temporal domain awareness for multi-agent collaborative perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23383-23392, 2023. 3 +[41] Lei Yang, Kaicheng Yu, Tao Tang, Jun Li, Kun Yuan, Li Wang, Xinyu Zhang, and Peng Chen. Bevheight: A robust framework for vision-based roadside 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21611-21620, 2023. 1 +[42] Zetong Yang, Yin Zhou, Zhifeng Chen, and Jiquan Ngiam. 3d-man: 3d multi-frame attention network for object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1863-1872, 2021. 2 +[43] Haibao Yu, Wenxian Yang, Hongzhi Ruan, Zhenwei Yang, Yingjuan Tang, Xu Gao, Xin Hao, Yifeng Shi, Yifeng Pan, Ning Sun, et al. V2x-seq: A large-scale sequential dataset for vehicle-infrastructure cooperative perception and forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5486-5495, 2023. 2, 5 +[44] Haibao Yu, Yingjuan Tang, Enze Xie, Jilei Mao, Ping Luo, and Zaiqing Nie. Flow-based feature fusion for vehicleinfrastructure cooperative 3d object detection. Advances in Neural Information Processing Systems, 36, 2024. 1, 2 +[45] Jingyu Zhang, Kun Yang, Yilei Wang, Hanqi Wang, Peng Sun, and Liang Song. ERMVP: Communication-efficient and collaboration-robust multi-vehicle perception in challenging environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12575–12584, 2024. 1, 2, 5, 6 +[46] Qingzhao Zhang, Xumiao Zhang, Ruiyang Zhu, Fan Bai, Mohammad Naserian, and Z Morley Mao. Robust real-time multi-vehicle collaboration on asynchronous sensors. In Proceedings of the 29th Annual International Conference on Mobile Computing and Networking, pages 1–15, 2023. 2 +[47] Xumiao Zhang, Anlan Zhang, Jiachen Sun, Xiao Zhu, Y Ethan Guo, Feng Qian, and Z Morley Mao. Emp: Edge-assisted multi-vehicle perception. In Proceedings of the 27th Annual International Conference on Mobile Computing and Networking, pages 545–558, 2021. 1 \ No newline at end of file diff --git a/2025/TraF-Align_ Trajectory-aware Feature Alignment for Asynchronous Multi-agent Perception/images.zip b/2025/TraF-Align_ Trajectory-aware Feature Alignment for Asynchronous Multi-agent Perception/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7ab9f21d4a334a79ccf9f3a7dc207108628eef88 --- /dev/null +++ b/2025/TraF-Align_ Trajectory-aware Feature Alignment for Asynchronous Multi-agent Perception/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:612815d3fdea327e8e91829acd263174b1a2d2644700e43efee4615aa28c58fa +size 762898 diff --git a/2025/TraF-Align_ Trajectory-aware Feature Alignment for Asynchronous Multi-agent Perception/layout.json b/2025/TraF-Align_ Trajectory-aware Feature Alignment for Asynchronous Multi-agent Perception/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3194cc60d0720acee94d8f8c5039810ab7414ead --- /dev/null +++ b/2025/TraF-Align_ Trajectory-aware Feature Alignment for Asynchronous Multi-agent Perception/layout.json @@ -0,0 +1,8217 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 103, + 516, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 103, + 516, + 140 + ], + "spans": [ + { + "bbox": [ + 96, + 103, + 516, + 140 + ], + "type": "text", + "content": "TraF-Align: Trajectory-aware Feature Alignment for Asynchronous Multi-agent Perception" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 160, + 520, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 160, + 520, + 219 + ], + "spans": [ + { + "bbox": [ + 88, + 160, + 520, + 219 + ], + "type": "text", + "content": "Zhiying Song" + }, + { + "bbox": [ + 88, + 160, + 520, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 520, + 219 + ], + "type": "text", + "content": ", Lei Yang" + }, + { + "bbox": [ + 88, + 160, + 520, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 520, + 219 + ], + "type": "text", + "content": ", Fuxi Wen" + }, + { + "bbox": [ + 88, + 160, + 520, + 219 + ], + "type": "inline_equation", + "content": "^{1,2,\\dagger}" + }, + { + "bbox": [ + 88, + 160, + 520, + 219 + ], + "type": "text", + "content": ", Jun Li" + }, + { + "bbox": [ + 88, + 160, + 520, + 219 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 88, + 160, + 520, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 520, + 219 + ], + "type": "text", + "content": "School of Vehicle and Mobility, Tsinghua University †Corresponding author \n" + }, + { + "bbox": [ + 88, + 160, + 520, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 160, + 520, + 219 + ], + "type": "text", + "content": "State Key Lab of Intelligent Green Vehicle and Mobility \n{song-zy24,yanglei20}@ mails.tsinghua.edu.cn, {wenfuxi, lijun1958}@tsinghua.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 271, + 296, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 271, + 296, + 535 + ], + "spans": [ + { + "bbox": [ + 54, + 271, + 296, + 535 + ], + "type": "text", + "content": "Cooperative perception presents significant potential for enhancing the sensing capabilities of individual vehicles, however, inter-agent latency remains a critical challenge. Latencies cause misalignments in both spatial and semantic features, complicating the fusion of real-time observations from the ego vehicle with delayed data from others. To address these issues, we propose TraF-Align, a novel framework that learns the flow path of features by predicting the feature-level trajectory of objects from past observations up to the ego vehicle's current time. By generating temporally ordered sampling points along these paths, TraF-Align directs attention from the current-time query to relevant historical features along each trajectory, supporting the reconstruction of current-time features and promoting semantic interaction across multiple frames. This approach corrects spatial misalignment and ensures semantic consistency across agents, effectively compensating for motion and achieving coherent feature fusion. Experiments on two real-world datasets, V2V4Real and DAIR-V2X-Seq, show that TraF-Align sets a new benchmark for asynchronous cooperative perception. The code is available at https://github.com/zhyingS/TraF-Align." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 559, + 135, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 135, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 135, + 572 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 580, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 700 + ], + "type": "text", + "content": "Intelligent vehicles rely on advanced onboard sensors, such as LiDAR and cameras, to perceive and understand the surrounding world. Nevertheless, the limited field of view and detection range of onboard sensors restrict their ability to perceive objects in occluded areas and at greater distances [2, 23, 36, 47]. Vehicle-to-everything (V2X) communication-enabled cooperative perception systems bridge this gap by exchanging perception information between vehicles and other agents, thereby extending the perception capability of a single agent [3, 4, 20, 41, 45]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "content": "In real-world scenarios, cooperative perception systems" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 324, + 245, + 552, + 401 + ], + "blocks": [ + { + "bbox": [ + 324, + 245, + 552, + 401 + ], + "lines": [ + { + "bbox": [ + 324, + 245, + 552, + 401 + ], + "spans": [ + { + "bbox": [ + 324, + 245, + 552, + 401 + ], + "type": "image", + "image_path": "f3d4dfeee0aaaa6a9eee9851008ffb564a0a3cc14d5fed9de5113269dbcf1eaf.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 409, + 555, + 432 + ], + "lines": [ + { + "bbox": [ + 313, + 409, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 555, + 432 + ], + "type": "text", + "content": "Figure 1. Asynchronous observations of ego vehicle and other agents result in spatial and semantic misalignments." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 459, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 555, + 628 + ], + "type": "text", + "content": "are cursed with inter-agent time delays [33, 35]. Consider a network with an ego agent and agent " + }, + { + "bbox": [ + 313, + 459, + 555, + 628 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 459, + 555, + 628 + ], + "type": "text", + "content": ", where their observations of the same object are shown in Figure 1. Latency in agent " + }, + { + "bbox": [ + 313, + 459, + 555, + 628 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 459, + 555, + 628 + ], + "type": "text", + "content": "'s message, captured before ego's observation time, can result in either a complete loss of features (1-I) or feature misalignment relative to ego's observation (1-II). Corresponding to Figure 1-I and 1-II, there are two main challenges: i) Spatial misalignment, where observations from other agents fail to align with the ground truth in space, lacking any reference from the ego vehicle's observations. ii) Semantic misalignment, where both the ego vehicle and another agent observe the same object but fail to recognize it as the same entity. These misalignments lead to misplaced objects and increased false positives [29, 30]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "content": "To address the first challenge, some studies have proposed potential solutions, by predicting the current-frame features using historical frames [15], or predicting the pixel-level flow of features and then extrapolating them [44] or directly warping them to their future positions [34]. For instance, FFNet [44] uses a self-supervised approach to learn the first-order derivatives of features between consecutive" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12048" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": "frames, and CoBEVFlow [34] generates an instance-level region of interest (ROI) and calculates feature motion vectors by matching and tracking these regions. However, predicting the movement of high-dimensional neural features is non-trivial. Despite their efforts, these methods struggle with larger delays and lack end-to-end capability. For the second challenge, the significance of semantic misalignment has not been well investigated in previous work." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 172, + 298, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 298, + 531 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 298, + 531 + ], + "type": "text", + "content": "To address these misalignments, our core idea is to reconstruct current-time features and achieve semantic consensus across agents on multi-frame LiDAR features within a unified framework. We guide the flow and interaction of high-dimensional features by learning the motion of low-dimensional trajectory features. A trajectory field is defined as a spatiotemporal map of each object's anticipated path, guiding the generation of temporally ordered sampling points along each trajectory. Cross-attention between each query and its aligned sampling points allows the query to access relevant past features along the trajectory, enabling effective feature reconstruction and semantic interaction for robust spatial alignment and inter-agent consensus. Consider the spatial misalignment scenario shown in Figure 1-I. By aggregating multi-frame point cloud features from agent " + }, + { + "bbox": [ + 56, + 172, + 298, + 531 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 56, + 172, + 298, + 531 + ], + "type": "text", + "content": ", we learn the object's past trajectory distribution and predict its motion direction up to the ego vehicle's current time, forming a trajectory field. This field guides attention from the predicted position to historical features along the same trajectory, enabling the reconstruction of the object's current features near the ground truth and compensating for motion. For semantic misalignment cases (Figure 1-II), while delays may misalign agent " + }, + { + "bbox": [ + 56, + 172, + 298, + 531 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 56, + 172, + 298, + 531 + ], + "type": "text", + "content": "'s observations with the ego vehicle's, the continuity of object motion implies overlap with point clouds from past ego frames. By learning the trajectory distribution from the ego's multi-frame point clouds and applying attention to compute feature correlations along the trajectory, the model recognizes that all features along the trajectory correspond to the same object, thus achieving cross-agent semantic alignment." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "Building on the abovementioned idea, we propose TraFAlign, a novel framework for asynchronous multi-agent cooperative perception (as shown in Figure 2). The proposed method includes three key parts: i) a field predictor that generates synchronized trajectory fields aligned with the ego vehicle's time, ii) an offset generator that produces sampling points for each query position and distributes them along the corresponding trajectory, and iii) attention layers where each query position is treated as a token, with attention focused only on positions specified by the offset generator. Extensive experiments are conducted on two widely used large-scale real-world datasets, V2V4Real [38] and DAIR-V2X-Seq [43], one for vehicle-to-vehicle (V2V) cooperative perception, another for vehicle-to-infrastructure (V2I) applications. Experimental results show that state-of" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "the-art (SOTA) performance is achieved for TraF-Align in both object detection precision and robustness to inter-agent latency. The contributions are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 108, + 553, + 228 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 314, + 108, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 108, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 314, + 108, + 553, + 144 + ], + "type": "text", + "content": "- We propose TraF-Align, a novel framework for asynchronous cooperative perception, addressing spatial and semantic misalignments caused by time delays." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 144, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 191 + ], + "type": "text", + "content": "- We leverage a field predictor and an offset generator to align trajectory fields and distribute attention points along object trajectories, enabling robust feature reconstruction and interaction along spatiotemporally aligned paths." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 192, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 553, + 228 + ], + "type": "text", + "content": "- We introduce a field loss and an offset loss to supervise trajectory alignment and the generation of attention points, providing an end-to-end trainable solution." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 239, + 398, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 239, + 398, + 251 + ], + "spans": [ + { + "bbox": [ + 314, + 239, + 398, + 251 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 258, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 666 + ], + "type": "text", + "content": "Asynchronous V2X cooperative perception. V2X cooperative perception extends the sensing range of single vehicles by exchanging information with other agents via wireless communication [13]. Previous studies have conducted extensive research under time-synchronized conditions, such as F-Cooper [3], V2VNet [33], V2X-ViT [35, 39], ERMVP [45], etc. In the presence of latency, previous research emphasizes using historical data to estimate and compensate for object motion. For instance, at the output level, the object state filters are designed to predict and update the dynamic state of the received objects [1]. Using raw data, Zhang et al. [46] and Dao et al. [8] estimated the motion of point clouds and predicted compensatory flows for late-arriving data. However, these levels may encounter challenges, such as short prediction time horizons or high communication data volumes. At the feature level, which has a better balance between accuracy and transmission bandwidth [18], SyncNet [15] employs a pyramidal LSTM to iteratively process and learn from delayed features, enabling the prediction of current-frame features. FFNet [44] predicts inter-frame feature changes in a self-supervised feature manner to synchronize features across agents. Nevertheless, predicting high-dimensional features is challenging and prone to uncontrollable variations. Additionally, CoBEVFlow [34] extracts and warps object-specific features to anticipated locations by predicting instance-wise motion vectors. However, this approach requires generating ROIs and tracking their motion, resulting in a non-end-to-end two-stage method. This paper uses multi-frame LiDAR sequences to predict low-dimensional trajectory features and guide attention to fuse features on past trajectories. By operating directly at the feature level, our approach enables end-to-end training and avoids the complexities of high-dimensional feature prediction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "content": "Learning from LiDAR sequences across agents. Multiframe point cloud fusion is widely applied in single-vehicle autonomous systems to amalgamate semantic data across different perspectives [5, 21, 22, 24, 32, 42]. However, its" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12049" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 68, + 549, + 286 + ], + "blocks": [ + { + "bbox": [ + 62, + 68, + 549, + 286 + ], + "lines": [ + { + "bbox": [ + 62, + 68, + 549, + 286 + ], + "spans": [ + { + "bbox": [ + 62, + 68, + 549, + 286 + ], + "type": "image", + "image_path": "2cdf8216f524cb86daeb1fdb5422271a921dc7b3ef7ca977ca25426faa07647e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 295, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 54, + 295, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 54, + 295, + 555, + 340 + ], + "type": "text", + "content": "Figure 2. Proposed Architecture. With asynchronous LiDAR inputs from the ego agent (time " + }, + { + "bbox": [ + 54, + 295, + 555, + 340 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 295, + 555, + 340 + ], + "type": "text", + "content": ") and agent " + }, + { + "bbox": [ + 54, + 295, + 555, + 340 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 54, + 295, + 555, + 340 + ], + "type": "text", + "content": " (time " + }, + { + "bbox": [ + 54, + 295, + 555, + 340 + ], + "type": "inline_equation", + "content": "s = t - \\tau_t^i" + }, + { + "bbox": [ + 54, + 295, + 555, + 340 + ], + "type": "text", + "content": "), BEV features are extracted using an onboard sparse encoder. Agent " + }, + { + "bbox": [ + 54, + 295, + 555, + 340 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 54, + 295, + 555, + 340 + ], + "type": "text", + "content": "'s features are transmitted to the ego agent and stored in memory. TraFAlign predicts a trajectory field up to time " + }, + { + "bbox": [ + 54, + 295, + 555, + 340 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 295, + 555, + 340 + ], + "type": "text", + "content": ", generates target attention positions, and uses features at these positions as keys/values in attention layers. Finally, the multi-agent features are fused and processed by the head to generate cooperative predictions at time " + }, + { + "bbox": [ + 54, + 295, + 555, + 340 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 295, + 555, + 340 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 360, + 297, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 360, + 297, + 553 + ], + "spans": [ + { + "bbox": [ + 54, + 360, + 297, + 553 + ], + "type": "text", + "content": "application in V2X cooperative perception remains under-explored. Although studies like SCOPE [40] and MRC-Net [11] have leveraged recurrent neural networks to fuse sequences from ego vehicles, these methods are resource-intensive and limited by sequence length. This has restricted their use to ego vehicles rather than a broader agent network, necessitating a more straightforward multi-frame fusion network. In cooperative perception scenarios, multiframe point clouds offer crucial motion cues that enable agents to synchronize their understanding of object positions despite time delays, thereby correlating trajectories across agents. This aspect has yet to be fully explored at the feature level. This paper proposes a novel approach using trajectory fields and customized attention with offsets for direct multi-frame fusion, thus establishing a consistent understanding of object motion across the network." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 566, + 112, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 112, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 112, + 578 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 587, + 178, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 178, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 178, + 598 + ], + "type": "text", + "content": "3.1. Problem formulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "Consider a group of networked agents indexed by " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\{e, 1, 2, \\dots, N - 1\\}" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": ", each is equipped with LiDAR and communication units. Index " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " indicates the ego agent, which integrates information from other agents to extend its sensing range and reduce the blind spots. The LiDAR observation of the " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "th agent at its local timestamp " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " is denoted as " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "L_s^i" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "i \\in \\mathcal{I}" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": ". A shared neural network " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " processes this observation into deep features. At each timestamp " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": ", agent " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " transmits " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "F_s^i = f(L_s^i)" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " to the ego agent," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "text", + "content": "who then accumulates these features in the cache, forming a historical feature set for agent " + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "text", + "content": ", represented as " + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s^i = \\{F_{s - k}^i\\}_{k\\in K}" + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "inline_equation", + "content": "K = \\{k\\in \\mathbb{Z}|1\\leqslant k\\leqslant m_i\\}" + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "text", + "content": " is the maximal sequence length for agent " + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "text", + "content": ". Accounting for delays in the collaborative process, messages from various agents arrive at ego at different times. Let " + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "inline_equation", + "content": "\\tau_t^i" + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "text", + "content": " be the delay for agent " + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "text", + "content": "'s message, and this means the information sent by agent " + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "inline_equation", + "content": "t - \\tau_t^i" + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "text", + "content": " is received by the ego at time " + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "text", + "content": ". If ignoring feature extraction delay, then " + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "inline_equation", + "content": "s = t - \\tau_t^i" + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "text", + "content": ". The task is to fuse these multi-agent features and estimate the state of objects, i.e., the bounding boxes, at ego timestamp " + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "text", + "content": " within the detection range. The information available for fusion includes: i) the ego's historical features " + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_t^e" + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "text", + "content": ", and ii) the historical features from other agents, denoted as " + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}_{t - \\tau_t^i}^i\\}_{i\\in \\mathcal{T}\\backslash \\{e\\}}" + }, + { + "bbox": [ + 313, + 360, + 556, + 542 + ], + "type": "text", + "content": ". The proposed method is shown in Figure 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 552, + 424, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 424, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 424, + 563 + ], + "type": "text", + "content": "3.2. Feature extraction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "The encoder converts single-frame raw point cloud data into BEV features. Similar to PointPillars [14], the point cloud is first discretized into a regular grid on the " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "x - y" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " plane. Within each pillar, the point features are encapsulated as the vector " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "[x,y,z,x_c,y_c,z_c,x_p,y_p,z_p]\\in \\mathbb{R}^9" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "[x,y,z]" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " is the point's position, while " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "[x_c,y_c,z_c]" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " indicates the center of the pillar. The vector " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "[x_p,y_p,z_p]" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " represents the point's offset from the pillar center. The pillar feature extraction and backbone network architectures are adopted from PillarNext [16]. Multi-Layer Perceptron (MLP) serves to encode point-wise features, while a sparse variant of ResNet [10] is utilized as the backbone to distill deeper and more" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12050" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "complex features with size " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "(C, H, W)" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": ". The sparse network is implemented using Spatially Sparse Convolution (SpConv) library [6]. The extracted single-frame features are transmitted and stored in the ego's feature bank " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "(\\mathcal{F}_t^e" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s^i)" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": ", which maintains a fixed capacity by admitting new features and discarding the oldest ones, ensuring an up-to-date repository for feature reuse. The extracted features of each agent are then processed using TraF-Align separately." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 168, + 294, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 294, + 239 + ], + "type": "text", + "content": "Temporal Encoding. Temporal information is decoupled from the point features to facilitate the reuse of historical features across frames. This allows for the immediate retrieval of historical features at any timestamp. Inspired by the positional embedding in Transformer [31], a temporal embedding " + }, + { + "bbox": [ + 55, + 168, + 294, + 239 + ], + "type": "inline_equation", + "content": "TE \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 55, + 168, + 294, + 239 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 88, + 248, + 294, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 248, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 88, + 248, + 294, + 262 + ], + "type": "interline_equation", + "content": "T E _ {(\\tau , 2 j)} = \\sin (\\xi), T E _ {(\\tau , 2 j + 1)} = \\cos (\\xi), \\tag {1}", + "image_path": "504ab07e36a3f643fee27691e424f5b2df3a6e2cd56457af451e00c9b1d75815.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 270, + 295, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 295, + 378 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 270, + 295, + 378 + ], + "type": "inline_equation", + "content": "\\xi = \\tau / \\varepsilon^{2j/C}" + }, + { + "bbox": [ + 55, + 270, + 295, + 378 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 270, + 295, + 378 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 270, + 295, + 378 + ], + "type": "text", + "content": " indexes the dimension along " + }, + { + "bbox": [ + 55, + 270, + 295, + 378 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 270, + 295, + 378 + ], + "type": "text", + "content": " channels, " + }, + { + "bbox": [ + 55, + 270, + 295, + 378 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 270, + 295, + 378 + ], + "type": "text", + "content": " represents the number of frames delayed from the current frame, and " + }, + { + "bbox": [ + 55, + 270, + 295, + 378 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 55, + 270, + 295, + 378 + ], + "type": "text", + "content": " is set to eight in this study. Each feature within the memory cache, " + }, + { + "bbox": [ + 55, + 270, + 295, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_t^e" + }, + { + "bbox": [ + 55, + 270, + 295, + 378 + ], + "type": "text", + "content": " for the ego agent and " + }, + { + "bbox": [ + 55, + 270, + 295, + 378 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s^i" + }, + { + "bbox": [ + 55, + 270, + 295, + 378 + ], + "type": "text", + "content": " for agent " + }, + { + "bbox": [ + 55, + 270, + 295, + 378 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 270, + 295, + 378 + ], + "type": "text", + "content": ", associated with a specific temporal delay, is concatenated with a corresponding temporal embedding. These concatenated features are refined through a convolutional layer with a " + }, + { + "bbox": [ + 55, + 270, + 295, + 378 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 55, + 270, + 295, + 378 + ], + "type": "text", + "content": " kernel and then warped to time " + }, + { + "bbox": [ + 55, + 270, + 295, + 378 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 270, + 295, + 378 + ], + "type": "text", + "content": " by ego-motion compensation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 385, + 151, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 151, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 151, + 398 + ], + "type": "text", + "content": "3.3. Trajectory field" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 403, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 403, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 54, + 403, + 295, + 594 + ], + "type": "text", + "content": "Each agent processes its aggregated multi-frame feature map via a field predictor, which generates a trajectory field consisting of position and orientation fields. The position field creates heatmap peaks along the object's trajectory, while the orientation field captures the inverse tangent direction of the trajectory. Due to delays, agents' local observations may be outdated. To achieve temporal alignment, the field predictor predicts the trajectory features of all objects up to the ego vehicle's latest time (with temporal information encoded in the features via Equation 1). This ensures that, for the same object, trajectory fields from different viewpoints spatially overlap and are aligned temporally to the ego vehicle's current time. Our model uses a variant of UNet [25] as the field predictor, whose architecture is detailed in the supplementary material, along with a supervision signal to guide the learning process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "As shown in Figure 3, when generating the ground truth field, we first extract bounding box annotations of each object from the local perspective of each agent within the maximum temporal window, starting from the object's initial appearance up to the ego's current time. A trajectory is then constructed by connecting the centers of these bounding boxes. Specifically, the length of the object's trajectory is given by " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\lceil \\tau_i \\cdot \\omega_i \\rceil + m_i" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\forall i \\in \\mathcal{I}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\lceil \\cdot \\rceil" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " denotes the ceiling function, " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\omega_i" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " is the LiDAR sampling frequency " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "(\\mathrm{s}^{-1})" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " of agent " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\tau_i" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " represents its cooperative latency (s). The" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 320, + 74, + 550, + 270 + ], + "blocks": [ + { + "bbox": [ + 320, + 74, + 550, + 270 + ], + "lines": [ + { + "bbox": [ + 320, + 74, + 550, + 270 + ], + "spans": [ + { + "bbox": [ + 320, + 74, + 550, + 270 + ], + "type": "image", + "image_path": "0af345f65626421913b5a14c7b1c3fb65639dfffa01dae3308eeaa52744f6a10.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 278, + 553, + 312 + ], + "lines": [ + { + "bbox": [ + 313, + 278, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 553, + 312 + ], + "type": "text", + "content": "Figure 3. Illustration of trajectory field, which includes a position field indicating the occupancy grid of the trajectory and a direction field depicting the trajectory's flow orientation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 335, + 553, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 553, + 382 + ], + "type": "text", + "content": "trajectory field is then generated by interpolating along the trajectory and projecting it onto the BEV feature space. If two object trajectories overlap, the trajectory with the more recent timestamp will overwrite the older one." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 384, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 554, + 491 + ], + "type": "text", + "content": "For cases in Figure 1-I, such a design compels the network to predict the trajectory field up to the present moment, thereby supporting the learning of motion information. Failing to do so would result in the network detecting only the delayed bounding box. Compared to the direct forecasting of high-dimensional features from the backbone network, predicting the trajectory field provides a more straightforward and more reliable method with predictable outcomes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 503, + 463, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 463, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 463, + 515 + ], + "type": "text", + "content": "3.4. Trajectory-aware attention" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": "We flatten the input feature map " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "F^i (\\forall i \\in \\mathcal{I})" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "C \\times H \\times W" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " embeddings, each serving as a query " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "q \\in \\mathbb{R}^{1 \\times C}" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": ". For each query, the offset generator creates a response set " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "R \\in \\mathbb{R}^{n \\times C}" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": ", capturing features at designated attention positions that are learned based on predefined ground truth. " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " is the number of attention positions for each query, and it's set to 18 in this paper. To generate the ground truth offsets, we select " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " pixels on the feature map that satisfy three conditions: i) The position field has a response (indicating the presence of an object), ii) The pixels belong to the same trajectory as the query (indicating the same object), and iii) The pixels correspond to earlier timestamps than the query (indicating historical positions of the object). As shown in Figure 4b, this ground truth aligns each query with preceding trajectory positions, enabling it to draw in relevant historical information. The" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12051" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": "trajectory field has effectively modeled the motion and flow of the historical features, offering valuable guidance to the offset generator. With this rich information, the generator can rely on simple convolution layers to produce good attention offsets, and we use PReLU activation [9] to provide flexibility in adjusting the offsets' directional orientation. Given the response set " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": ", the feature of query " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " is updated or reconstructed by the attention process:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 174, + 296, + 207 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 174, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 61, + 174, + 296, + 207 + ], + "type": "interline_equation", + "content": "\\text {A t t e n t i o n} (q, R) = \\text {S o f t m a x} \\left[ \\frac {q W ^ {q} \\left(R W ^ {k}\\right) ^ {T}}{\\sqrt {d}} \\right] R W ^ {v}, \\tag {2}", + "image_path": "8a490c4d8906c34a7482f402efb81990d615a8e49649301e5930af62ef1635f6.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 213, + 296, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 296, + 346 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 213, + 296, + 346 + ], + "type": "inline_equation", + "content": "W^{q}, W^{k}, W^{v} \\in \\mathbb{R}^{C \\times d}" + }, + { + "bbox": [ + 55, + 213, + 296, + 346 + ], + "type": "text", + "content": " are the query, key and value projection matrices. Similar to the original Transformer [31], we also incorporate feed-forward networks, addition and normalization layers, and a multi-head attention design, the details are shown in Figure 4a. Using attention along the trajectory transfers features to the current time step and enables feature interaction along the trajectory. This allows the model to recognize which features belong to the same object, aligning the features across agents for effective fusion. As a result, each agent aligns its feature map to the ego vehicle's latest time step." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 352, + 277, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 277, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 277, + 365 + ], + "type": "text", + "content": "3.5. Inter-agent fusion, detection head and loss" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 369, + 296, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 440 + ], + "type": "text", + "content": "While inter-agent fusion has been extensively studied [3, 36, 45], this paper adopts a simplified approach by concatenating features from all agents and utilizing a basic convolution layer for fusion. We use an anchor-based detection head and employ three losses: a detection loss, a field loss, and an offset loss:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 449, + 294, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 449, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 449, + 294, + 460 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {d e t e c t i o n}} + \\alpha \\mathcal {L} _ {\\text {f i e l d}} + \\beta \\mathcal {L} _ {\\text {o f f s e t}}, \\tag {3}", + "image_path": "59e461986409cd0c48b829bb516c59d1fe0288f4d9d5bbe7923f6ff4308583cb.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 468, + 296, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 611 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 468, + 296, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{detection}}" + }, + { + "bbox": [ + 55, + 468, + 296, + 611 + ], + "type": "text", + "content": " is consistent with the loss used in [14], " + }, + { + "bbox": [ + 55, + 468, + 296, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{field}}" + }, + { + "bbox": [ + 55, + 468, + 296, + 611 + ], + "type": "text", + "content": " includes a focal loss [17] for the position field and an " + }, + { + "bbox": [ + 55, + 468, + 296, + 611 + ], + "type": "inline_equation", + "content": "L1" + }, + { + "bbox": [ + 55, + 468, + 296, + 611 + ], + "type": "text", + "content": " loss for the orientation field. The constants " + }, + { + "bbox": [ + 55, + 468, + 296, + 611 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 468, + 296, + 611 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 468, + 296, + 611 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 468, + 296, + 611 + ], + "type": "text", + "content": " balance the contributions of each loss term. Here, we set " + }, + { + "bbox": [ + 55, + 468, + 296, + 611 + ], + "type": "inline_equation", + "content": "\\alpha = \\beta = 0.05" + }, + { + "bbox": [ + 55, + 468, + 296, + 611 + ], + "type": "text", + "content": ". Considering the offset loss, because the generated attention positions are unordered, we first apply the Sinkhorn algorithm [7, 26, 27] to compute the matching probabilities " + }, + { + "bbox": [ + 55, + 468, + 296, + 611 + ], + "type": "inline_equation", + "content": "P_{jk} \\in \\mathbb{R}^{n \\times n}" + }, + { + "bbox": [ + 55, + 468, + 296, + 611 + ], + "type": "text", + "content": " between the generated attention positions and the ground truth. This is done using the L1 distance matrix " + }, + { + "bbox": [ + 55, + 468, + 296, + 611 + ], + "type": "inline_equation", + "content": "C_{jk} \\in \\mathbb{R}^{n \\times n}" + }, + { + "bbox": [ + 55, + 468, + 296, + 611 + ], + "type": "text", + "content": " as the cost for the pixel located at " + }, + { + "bbox": [ + 55, + 468, + 296, + 611 + ], + "type": "inline_equation", + "content": "(j, k)" + }, + { + "bbox": [ + 55, + 468, + 296, + 611 + ], + "type": "text", + "content": " on the feature map. The offset loss is then defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 616, + 296, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 616, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 70, + 616, + 296, + 643 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {o f f s e t}} = \\sum_ {j, k} P _ {j k} \\cdot C _ {j k}, \\forall j \\in [ 1, H ], k \\in [ 1, W ]. \\tag {4}", + "image_path": "0e13527a5a435f0405cb22da71a73c017422aed8b3369dd8ded9b133ce4f4fb2.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 658, + 137, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 137, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 137, + 671 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "To evaluate robustness to cross-agent latency, we conduct 3D object detection experiments on two large-scale real-world datasets, V2V4Real [38] and DAIR-V2X-Seq [43]." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 323, + 69, + 553, + 185 + ], + "blocks": [ + { + "bbox": [ + 323, + 69, + 553, + 185 + ], + "lines": [ + { + "bbox": [ + 323, + 69, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 323, + 69, + 553, + 185 + ], + "type": "image", + "image_path": "55564513160de392774c28db019896855f74f3937e31077014e26ee013505a5b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 190, + 494, + 199 + ], + "lines": [ + { + "bbox": [ + 374, + 190, + 494, + 199 + ], + "spans": [ + { + "bbox": [ + 374, + 190, + 494, + 199 + ], + "type": "text", + "content": "(a) Trajectory-aware attention layers." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 324, + 201, + 550, + 312 + ], + "blocks": [ + { + "bbox": [ + 324, + 201, + 550, + 312 + ], + "lines": [ + { + "bbox": [ + 324, + 201, + 550, + 312 + ], + "spans": [ + { + "bbox": [ + 324, + 201, + 550, + 312 + ], + "type": "image", + "image_path": "e47348146adec98c21b18cdd1b1ad9c97eedf19a6115575a1a39490bceb3a2f8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 313, + 511, + 324 + ], + "lines": [ + { + "bbox": [ + 359, + 313, + 511, + 324 + ], + "spans": [ + { + "bbox": [ + 359, + 313, + 511, + 324 + ], + "type": "text", + "content": "(b) Query and ground truth attention positions." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 314, + 333, + 553, + 345 + ], + "lines": [ + { + "bbox": [ + 314, + 333, + 553, + 345 + ], + "spans": [ + { + "bbox": [ + 314, + 333, + 553, + 345 + ], + "type": "text", + "content": "Figure 4. Attention positions and trajectory-aware attention layers." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 365, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 554, + 437 + ], + "type": "text", + "content": "These datasets encompass both V2V and V2I scenarios, as well as sparse and dense LiDAR conditions. The detection results are evaluated using Average Precision (AP) at Intersection-over-Union (IoU) thresholds of 0.5 (AP50) and 0.7 (AP70). To enable a fair comparison, we omit feature compression operations from the methods under evaluation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 445, + 446, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 446, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 446, + 457 + ], + "type": "text", + "content": "4.1. Implementation details" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 462, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 557 + ], + "type": "text", + "content": "Datasets and Benchmarks. V2V4Real [38] is specialized for V2V cooperative detection in urban settings, and DAIR-V2X-Seq [43] captures sequential data in V2I scenarios at intersections. The benchmarks include single-agent perception relying on the ego vehicle's data for detection, alongside eight SOTA intermediate fusion models: F-Cooper [3], V2VNet [33], Where2comm [12], AttFuse [36], V2X-ViT [35], CoBEVT [37], MRCNet [11], and ERMVP [45]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "Training details. All the models are trained within the same codebase for 60 epochs on nine RTX 3090 GPUs, employing the one-cycle learning rate strategy [28] with AdamW [19] as the optimizer. The point clouds are segmented at a grid size of " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "0.4 \\times 0.4\\mathrm{m}" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": ", and the backbone's feature map size output is one-fourth the original size. During training on V2V4Real, all models shuffle the vehicle list and randomly select the ego vehicle at the start of each epoch. Standard LiDAR data augmentations are applied, including random flip, scaling, rotation, and translation. Multi-frame methods are trained with random latency augmentation, following a uniform distribution from 0 to " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "400~\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": ". TraF-Align leverages two LiDAR frames for the ego vehicle and four" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12052" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 102, + 550, + 235 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Performance comparison of vehicle class on V2V4Real validation and DAIR-V2X-Seq validation dataset with different latencies. The results are reported in AP 50/70 (%)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 102, + 550, + 235 + ], + "lines": [ + { + "bbox": [ + 56, + 102, + 550, + 235 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 550, + 235 + ], + "type": "table", + "html": "
MethodV2V4RealDAIR-V2X-Seq
0ms100 ms200 ms300 ms400 ms0ms100 ms200 ms300 ms400 ms
Single Agent48.75/30.5348.75/30.5348.75/30.5348.75/30.5348.75/30.5351.31/40.4551.31/40.4551.31/40.4551.31/40.4551.31/40.45
AttFuse [36]63.84/33.3059.16/27.3655.57/26.0353.25/25.4451.69/25.0968.94/49.6865.83/45.7663.02/44.2461.61/43.2860.19/42.53
F-Cooper [3]66.86/33.3962.15/27.6158.03/26.3055.31/25.4653.92/24.9569.47/50.0165.85/44.9662.74/43.2161.12/42.2860.01/41.70
V2VNet [33]67.11/35.4761.97/29.1458.09/27.8556.26/27.4354.62/26.7571.44/51.9167.26/46.6363.84/44.5961.68/43.2459.78/42.08
Where2comm [12]67.03/37.5561.86/31.5258.22/30.2056.22/29.7154.47/29.0768.37/48.5064.81/44.3261.51/42.8059.66/41.6358.25/40.80
V2X-ViT [35]68.51/35.8863.45/29.5759.42/28.0457.35/27.6555.60/27.2070.21/50.9266.42/46.4462.58/44.2961.00/43.5659.63/42.66
MRCNet [11]67.83/37.3762.65/31.5958.98/30.3956.70/29.5355.03/28.9469.27/50.6765.90/45.6062.65/44.1560.09/42.3958.19/41.35
CoBEVT [37]70.59/38.8265.50/32.1161.55/30.4058.91/29.5457.05/29.2671.51/51.3167.60/46.5564.38/45.0362.39/44.1160.74/43.08
ERMVP [45]70.74/38.4565.53/31.9161.63/30.5359.24/30.1857.26/29.3171.19/53.8567.04/48.0963.67/46.3761.23/44.7659.55/43.93
TraF-Align (Ours)74.28/44.1473.04/39.9971.51/39.5470.07/38.9969.41/38.4976.90/58.3173.95/53.6872.69/52.8371.66/52.3371.22/52.01
", + "image_path": "e322f68fde4a5aa4df1725cd69255fcc4ffaf9c03f3b71f7714021cea3620972.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 255, + 294, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 255, + 294, + 303 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 294, + 303 + ], + "type": "text", + "content": "for the cooperative vehicle on V2V4Real, while using four LiDAR frames for the vehicle and infrastructure on DAIR-V2X-Seq. The number of attention layers is set to two, each containing four attention heads." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 304, + 295, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 304, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 54, + 304, + 295, + 376 + ], + "type": "text", + "content": "Latency settings. To comprehensively evaluate robustness to time delays, we introduce delays of 0, 100, 200, 300, and " + }, + { + "bbox": [ + 54, + 304, + 295, + 376 + ], + "type": "inline_equation", + "content": "400\\mathrm{ms}" + }, + { + "bbox": [ + 54, + 304, + 295, + 376 + ], + "type": "text", + "content": " to one of the agents separately, extensively testing the model's performance under each delay condition. Unless specified otherwise, these delays are applied to the point cloud data of the non-ego agent." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 384, + 234, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 234, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 234, + 396 + ], + "type": "text", + "content": "4.2. Comparison with state-of-the-art" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 402, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 402, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 54, + 402, + 295, + 544 + ], + "type": "text", + "content": "Detection performance. Table 1 presents a comparison of detection performance on two real-world datasets for vehicle class, which includes cars, vans, trucks, buses, etc. It's observed that the proposed method i) achieves state-of-the-art performance across all benchmarks under various latency conditions on both datasets. ii) In synchronous scenarios, i.e., when the delay is " + }, + { + "bbox": [ + 54, + 402, + 295, + 544 + ], + "type": "inline_equation", + "content": "0\\mathrm{ms}" + }, + { + "bbox": [ + 54, + 402, + 295, + 544 + ], + "type": "text", + "content": ", our model improves AP 50/70 by " + }, + { + "bbox": [ + 54, + 402, + 295, + 544 + ], + "type": "inline_equation", + "content": "25.53\\% / 13.61\\%" + }, + { + "bbox": [ + 54, + 402, + 295, + 544 + ], + "type": "text", + "content": " (on V2V4Real) and " + }, + { + "bbox": [ + 54, + 402, + 295, + 544 + ], + "type": "inline_equation", + "content": "25.59\\% / 18.26\\%" + }, + { + "bbox": [ + 54, + 402, + 295, + 544 + ], + "type": "text", + "content": " (on DAIR-V2X-Seq) compared to single-agent perception, it also surpasses ERMVP, the SOTA performance of feature-level methods, by " + }, + { + "bbox": [ + 54, + 402, + 295, + 544 + ], + "type": "inline_equation", + "content": "3.54\\% / 5.69\\%" + }, + { + "bbox": [ + 54, + 402, + 295, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 402, + 295, + 544 + ], + "type": "inline_equation", + "content": "5.71\\% / 4.46\\%" + }, + { + "bbox": [ + 54, + 402, + 295, + 544 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 545, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 545, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 54, + 545, + 295, + 641 + ], + "type": "text", + "content": "Robustness to inter-agent latency. As shown in Table 1, our model demonstrates stronger robustness in the presence of time delays. Even under a " + }, + { + "bbox": [ + 54, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "400~\\mathrm{ms}" + }, + { + "bbox": [ + 54, + 545, + 295, + 641 + ], + "type": "text", + "content": " delay, our AP decreases by only " + }, + { + "bbox": [ + 54, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "4.87\\% /5.65\\%" + }, + { + "bbox": [ + 54, + 545, + 295, + 641 + ], + "type": "text", + "content": " (on V2V4Real) and " + }, + { + "bbox": [ + 54, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "5.68\\% /6.30\\%" + }, + { + "bbox": [ + 54, + 545, + 295, + 641 + ], + "type": "text", + "content": " (on DAIR-V2X-Seq) compared to the synchronous scenario. This performance significantly outperforms the ERMVP, which shows declines of " + }, + { + "bbox": [ + 54, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "13.48\\% /9.14\\%" + }, + { + "bbox": [ + 54, + 545, + 295, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "11.64\\% /9.92\\%" + }, + { + "bbox": [ + 54, + 545, + 295, + 641 + ], + "type": "text", + "content": " , respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 642, + 296, + 714 + ], + "type": "text", + "content": "To gain a deeper understanding of the impact of latency on collaboration, we plot the Precision-Recall curves of the models under various latency conditions, as shown in Figure 5. It is obvious that cooperation is effective in enhancing the recall limit of single-vehicle perception, helping the ego vehicle to detect a more significant number of objects." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 322, + 255, + 431, + 380 + ], + "blocks": [ + { + "bbox": [ + 322, + 255, + 431, + 380 + ], + "lines": [ + { + "bbox": [ + 322, + 255, + 431, + 380 + ], + "spans": [ + { + "bbox": [ + 322, + 255, + 431, + 380 + ], + "type": "image", + "image_path": "d7d5e17242f79c59992fe9cc8f2b2a1a2073234a6b1571b037c6d9470d18cc6c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 437, + 254, + 548, + 380 + ], + "blocks": [ + { + "bbox": [ + 437, + 254, + 548, + 380 + ], + "lines": [ + { + "bbox": [ + 437, + 254, + 548, + 380 + ], + "spans": [ + { + "bbox": [ + 437, + 254, + 548, + 380 + ], + "type": "image", + "image_path": "33902ffe0d7955c3ac3b544de806c2fa363eab9a2ef4149b0a5e5d9f1ded6ee2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 321, + 384, + 431, + 533 + ], + "blocks": [ + { + "bbox": [ + 321, + 384, + 431, + 533 + ], + "lines": [ + { + "bbox": [ + 321, + 384, + 431, + 533 + ], + "spans": [ + { + "bbox": [ + 321, + 384, + 431, + 533 + ], + "type": "image", + "image_path": "1a9ad1243c0ef0fbc8106297e8e9cc52ee1392f15c426150cc3e9b800a8dee69.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 542, + 555, + 565 + ], + "lines": [ + { + "bbox": [ + 313, + 542, + 555, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 555, + 565 + ], + "type": "text", + "content": "Figure 5. Precision-Recall curves for models on the V2V4Real validation dataset under various latency conditions." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 438, + 384, + 549, + 533 + ], + "blocks": [ + { + "bbox": [ + 438, + 384, + 549, + 533 + ], + "lines": [ + { + "bbox": [ + 438, + 384, + 549, + 533 + ], + "spans": [ + { + "bbox": [ + 438, + 384, + 549, + 533 + ], + "type": "image", + "image_path": "dd396af796c57a7a09b42fec6a53f420d55d851cace7b4f4551f0fcc0ef62558.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "However, at a " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "200\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " latency, state-of-the-art methods exhibit lower precision in the low-recall range compared to single-vehicle perception. This suggests that the delayed collaborative information interferes with the objects that the single vehicle could detect, corresponding to the semantic misalignment issues depicted in Figure 1-II. In areas not visible to the ego vehicle, specifically at high recall levels, the precision of SOTA methods significantly declines with the introduction of latency. This indicates that spatial misalignment issues (Figure 1-I) become dominant, lead" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12053" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": "ing to decreased performance. These findings reinforce our analysis of latency's impact on cooperative perception. Furthermore, using ERMVP as an example, the differences between the PR curves diminish as latency increases, indicating that existing models progressively lose their effectiveness in managing delays. In contrast, TraF-Align retains the original performance of single-agent perception in lowrecall regions and shows no significant decline in performance in high-recall areas due to latency. This highlights our model's ability to establish strong semantic consistency between the ego vehicle's data and the received information, demonstrating robust reconstruction capability and effectively addressing the scenarios in Figures 1-I and 1-II." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 238, + 154, + 250 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 154, + 250 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 154, + 250 + ], + "type": "text", + "content": "4.3. Ablation studies" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 256, + 296, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 296, + 316 + ], + "type": "text", + "content": "In this section, we conducted ablation studies on the effectiveness of significant components and loss and multiframe point clouds using average precision as the evaluation metric. Other ablation experiments and analyses (using PR curves) are presented in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 317, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 296, + 509 + ], + "type": "text", + "content": "Effectiveness of Major Components. To evaluate the contribution of modules, we progressively remove i) the field predictor, ii) the offset generator, and iii) the attention layers. The results are shown in Table 2. The experiments are conducted on the DAIR-V2X-Seq dataset. To ensure a comprehensive assessment of each component's contributions, we train the model for 120 epochs, double the duration of the experiments presented in Table 1. Additionally, we test the effects of adding latency to both the ego and non-ego agents to fully evaluate the impact of each module. The consistent rise in detection precision over different latency settings demonstrates the effectiveness of each introduced component. Notably, applying latency to vehicles results in lower AP than when latency is applied to infrastructure. This discrepancy arises from the different distributions of object visibility between the vehicle agent and roadside in" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 56, + 560, + 294, + 699 + ], + "blocks": [ + { + "bbox": [ + 55, + 529, + 295, + 551 + ], + "lines": [ + { + "bbox": [ + 55, + 529, + 295, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 295, + 551 + ], + "type": "text", + "content": "Table 2. Results of the ablation study of the core components on the DAIR-V2X-Seq dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 560, + 294, + 699 + ], + "lines": [ + { + "bbox": [ + 56, + 560, + 294, + 699 + ], + "spans": [ + { + "bbox": [ + 56, + 560, + 294, + 699 + ], + "type": "table", + "html": "
FP OG ALLatency of Infrastructure
0ms100ms200ms300ms400ms
75.35/54.7273.01/50.8171.88/50.1771.04/49.7770.09/49.45
✓76.31/57.0274.60/53.5373.90/52.3472.91/51.99
✓77.11/58.6375.15/54.7774.18/53.7773.24/52.99
✓✓77.83/59.7275.74/55.8574.99/54.82
FP OG ALLatency of Vehicle
0ms100ms200ms300ms400ms
75.35/54.7274.31/49.6172.17/43.3267.19/34.8256.93/24.51
✓76.31/57.0275.10/51.9372.91/46.2269.22/39.15
✓77.11/58.6375.76/53.4873.44/47.5569.92/40.52
✓✓77.83/59.7276.41/54.8874.24/49.46
", + "image_path": "0f35fae71d44b655b919c9be8faebeede34fcc70d6af33dc6bf11f5af92e8fed.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 67, + 700, + 276, + 711 + ], + "lines": [ + { + "bbox": [ + 67, + 700, + 276, + 711 + ], + "spans": [ + { + "bbox": [ + 67, + 700, + 276, + 711 + ], + "type": "text", + "content": "FP: Field predictor. OG: Offset generator. AL: Attention layers." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 316, + 102, + 553, + 240 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "type": "text", + "content": "Table 3. Results of the ablation study of the loss on the DAIR-V2X-Seq dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 102, + 553, + 240 + ], + "lines": [ + { + "bbox": [ + 316, + 102, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 102, + 553, + 240 + ], + "type": "table", + "html": "
FL OLLatency of Infrastructure
0ms100ms200ms300ms400ms
✓76.43/57.7474.44/53.9173.65/52.9772.71/52.1071.83/51.87
77.12/58.3674.96/54.5174.02/53.6872.98/52.9172.01/52.27
✓77.15/58.7075.19/54.5874.31/53.6673.68/53.0772.52/52.28
✓77.83/59.7275.74/55.8574.99/54.8274.10/54.0772.89/53.21
FL OLLatency of Vehicle
0ms100ms200ms300ms400ms
✓76.43/57.7475.09/52.8172.84/46.5469.14/39.1261.53/29.15
77.12/58.3675.69/53.2372.76/46.8569.25/39.7862.13/30.41
✓77.15/58.7075.80/53.9073.34/48.4970.12/41.8063.21/32.02
✓77.83/59.7276.41/54.8874.24/49.4670.75/42.6263.59/32.31
", + "image_path": "b03905b131ae06e2b47176f978f2208e3eb522d897484cb47d1bead45e31a7bc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 328, + 242, + 430, + 252 + ], + "lines": [ + { + "bbox": [ + 328, + 242, + 430, + 252 + ], + "spans": [ + { + "bbox": [ + 328, + 242, + 430, + 252 + ], + "type": "text", + "content": "FL:Field loss. OL: Offset loss." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 321, + 266, + 388, + 429 + ], + "blocks": [ + { + "bbox": [ + 321, + 266, + 388, + 429 + ], + "lines": [ + { + "bbox": [ + 321, + 266, + 388, + 429 + ], + "spans": [ + { + "bbox": [ + 321, + 266, + 388, + 429 + ], + "type": "image", + "image_path": "730d661a4453700fb5be3e7a5cb5543a368244299c1e2f39368ab5567b3902f5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 433, + 383, + 443 + ], + "lines": [ + { + "bbox": [ + 323, + 433, + 383, + 443 + ], + "spans": [ + { + "bbox": [ + 323, + 433, + 383, + 443 + ], + "type": "text", + "content": "(a) LiDAR points." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 391, + 266, + 457, + 432 + ], + "blocks": [ + { + "bbox": [ + 391, + 266, + 457, + 432 + ], + "lines": [ + { + "bbox": [ + 391, + 266, + 457, + 432 + ], + "spans": [ + { + "bbox": [ + 391, + 266, + 457, + 432 + ], + "type": "image", + "image_path": "4f8815cb30ce6da13f277e6a32c79dbc91b6687d022eb8dc87d423dcd62c34f6.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 393, + 434, + 456, + 443 + ], + "lines": [ + { + "bbox": [ + 393, + 434, + 456, + 443 + ], + "spans": [ + { + "bbox": [ + 393, + 434, + 456, + 443 + ], + "type": "text", + "content": "(b) Trajectory field." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 453, + 555, + 475 + ], + "lines": [ + { + "bbox": [ + 313, + 453, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 555, + 475 + ], + "type": "text", + "content": "Figure 6. Visualization of TraF-Align in real-world scenarios from the DAIR-V2X-Seq dataset with latency of " + }, + { + "bbox": [ + 313, + 453, + 555, + 475 + ], + "type": "inline_equation", + "content": "400~\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 453, + 555, + 475 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 468, + 266, + 539, + 431 + ], + "blocks": [ + { + "bbox": [ + 468, + 266, + 539, + 431 + ], + "lines": [ + { + "bbox": [ + 468, + 266, + 539, + 431 + ], + "spans": [ + { + "bbox": [ + 468, + 266, + 539, + 431 + ], + "type": "image", + "image_path": "a3b0ddecd3e40ab65b9a8bfcd3225a97fcdd0821f18f7e2c632b84b4ac408f3b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 476, + 434, + 534, + 443 + ], + "lines": [ + { + "bbox": [ + 476, + 434, + 534, + 443 + ], + "spans": [ + { + "bbox": [ + 476, + 434, + 534, + 443 + ], + "type": "text", + "content": "(c) Deep features." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 498, + 445, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 445, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 445, + 509 + ], + "type": "text", + "content": "frastructure agent in the dataset." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 510, + 555, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 592 + ], + "type": "text", + "content": "Effectiveness of loss. Table 3 presents the ablation study of field loss and offset loss, both designed to guide the attention biases, under the same settings as in Table 2. The results show that introducing the field loss provides better initial values for the offset generator, enhancing the model's performance. Optimal results are achieved only when both losses are incorporated." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": "Impact of LiDAR frames. Table 4 presents the ablation study conducted on V2V4Real, investigating the impact of different historical frame numbers for both the ego and cooperative vehicle on detection performance and robustness to time delays. It's observed that increasing the ego frames primarily enhances AP70, indicating improved accuracy in object localization, while AP50 shows no significant increase. Conversely, increasing the CV frames enhances both AP50 and the model's robustness to time delays, as most latency originates from the cooperative vehicle's Li" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12054" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 68, + 553, + 223 + ], + "blocks": [ + { + "bbox": [ + 65, + 68, + 553, + 223 + ], + "lines": [ + { + "bbox": [ + 65, + 68, + 553, + 223 + ], + "spans": [ + { + "bbox": [ + 65, + 68, + 553, + 223 + ], + "type": "image", + "image_path": "695ed3c096d1b7b5b1e7404d998ec06bb86258471cf4611290aa867e0e8d4986.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 230, + 555, + 253 + ], + "lines": [ + { + "bbox": [ + 55, + 230, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 555, + 253 + ], + "type": "text", + "content": "Figure 7. Qualitative results in real-world scenarios from the DAIR-V2X-Seq dataset with latency of " + }, + { + "bbox": [ + 55, + 230, + 555, + 253 + ], + "type": "inline_equation", + "content": "400\\mathrm{ms}" + }, + { + "bbox": [ + 55, + 230, + 555, + 253 + ], + "type": "text", + "content": ". Green and red boxes denote the ground truths and detection results, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 303, + 293, + 396 + ], + "blocks": [ + { + "bbox": [ + 55, + 271, + 295, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 271, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 295, + 293 + ], + "type": "text", + "content": "Table 4. Results of the ablation study of the number of LiDAR frames on the V2V4Real dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 303, + 293, + 396 + ], + "lines": [ + { + "bbox": [ + 59, + 303, + 293, + 396 + ], + "spans": [ + { + "bbox": [ + 59, + 303, + 293, + 396 + ], + "type": "table", + "html": "
Ego CVLatency of Cooperative Vehicle (CV)
0ms100ms200ms300ms400ms
1173.35/40.5469.76/34.0366.51/32.5465.17/32.1364.28/32.28
4172.85/44.8169.45/38.3266.73/37.1666.00/37.2865.67/37.32
4273.02/45.0071.74/39.8869.71/38.2567.88/37.4966.98/37.39
1471.28/42.1670.28/39.1869.31/38.1567.68/37.5266.47/36.90
2474.28/44.1473.04/39.9971.51/39.5470.07/38.9969.41/38.49
4474.76/44.0373.28/39.5871.64/38.5870.10/37.7069.08/37.08
", + "image_path": "d5d1248b7c8bcaa2d1c427ad7851ba727862a3b3bb32a101d5528f036df64954.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 414, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 296, + 462 + ], + "type": "text", + "content": "DAR. Furthermore, having too many point cloud frames for the ego vehicle may not yield optimal results, as it could introduce additional motion blur issues, which have been proved in the context of single-vehicle perception [24]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 468, + 165, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 165, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 165, + 481 + ], + "type": "text", + "content": "4.4. Qualitative results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "type": "text", + "content": "Visualization of detection results. Figure 7a-7d compare TraF-Align with three previous methods on a real-world scenario from DAIR-V2X-Seq dataset under a latency of " + }, + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "type": "inline_equation", + "content": "400~\\mathrm{ms}" + }, + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "type": "text", + "content": ". It is observed that in region " + }, + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "type": "inline_equation", + "content": "A_{1}" + }, + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "type": "text", + "content": ", due to the latency in roadside (Agent 2), V2VNet, COBEVT, and V2X-ViT all exhibit a false positive detection and fail to capture the correct bounding box; a similar issue occurs in region " + }, + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "type": "inline_equation", + "content": "A_{2}" + }, + { + "bbox": [ + 55, + 486, + 296, + 629 + ], + "type": "text", + "content": ". The reasons for this phenomenon have been discussed in section 1. In contrast, TraF-Align not only identifies the object but also avoids false positives. This improvement is due to our trajectory-guided attention, which enables both agents to achieve a consensus on the same object." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 630, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 689 + ], + "type": "text", + "content": "Visualization of trajectory field. Figure 6b illustrates the predicted trajectory fields (specifically, the position field) from both the vehicle (Agent 1) and roadside (Agent 2) perspectives. Despite the delays in the roadside point clouds, the predicted trajectory field is aligned to the latest frame." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Visualization of attention. Figure 7e shows the response within the " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "A_{1}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " box of Figure 7d from the roadside perspec" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 273, + 555, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 555, + 382 + ], + "type": "text", + "content": "tive, with six different query positions. Since the ego data is undelayed, the offset distribution centers around the query points, resembling a convolution kernel. In contrast, the roadside agent extends attention along the trajectory. After fusion by TraF-Align, as shown in Figure 6c, the roadside features align toward the vehicles, demonstrating spatial alignment. The feature intensity near the trajectory is uniform, indicating semantic alignment. More results are available in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 393, + 392, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 393, + 392, + 406 + ], + "spans": [ + { + "bbox": [ + 314, + 393, + 392, + 406 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 414, + 554, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 554, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 554, + 484 + ], + "type": "text", + "content": "We propose TraF-Align, a novel cooperative perception framework robust to inter-agent latencies. Our method handles delays by predicting the flow of object features and aligning them across agents through attention within the expected trajectory paths. Extensive experiments on real-world datasets demonstrate its effectiveness." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 486, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 676 + ], + "type": "text", + "content": "Limitations and future work. TraF-Align is primarily trained on data with regular latencies (multiples of 100 ms). In practical applications, more realistic latency settings should be considered. Additionally, while TraF-Align is robust to latency, detection precision under delay can still be improved. The main challenges are: i) Designing more accurate trajectory field predictors and attention position generators to enhance feature reconstruction capabilities; ii) Improving object localization precision. Although multi-frame information improves motion understanding, it also introduces motion blur, making precise object localization challenging and iii) Addressing pose estimation errors. Multi-frame processing relies on accurate inter-frame and inter-agent pose alignment, errors in these estimates can degrade performance. Future work will explore methods to reduce the impact of these factors." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "Acknowledgments. This research is supported by the National Key Research and Development Program of China under Grant 2024YFB2505803." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12055" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 135 + ], + "type": "text", + "content": "[1] Kunyang Cai, Ting Qu, Bingzhao Gao, and Hong Chen. Consensus-based distributed cooperative perception for connected and automated vehicles. IEEE Transactions on Intelligent Transportation Systems, 24(8):8188-8208, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 295, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 295, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 295, + 179 + ], + "type": "text", + "content": "[2] Antoine Caillot, Safa Ouerghi, Pascal Vasseur, Rémi Boutteau, and Yohan Dupuis. Survey on cooperative perception in an automotive context. IEEE Transactions on Intelligent Transportation Systems, 23(9):14204-14223, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 295, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 295, + 235 + ], + "type": "text", + "content": "[3] Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. In Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, pages 88-100, 2019. 1, 2, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 294, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 294, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 294, + 290 + ], + "type": "text", + "content": "[4] Qi Chen, Sihai Tang, Qing Yang, and Song Fu. Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds. In 2019 IEEE 39th International Conference on Distributed Computing Systems (ICDCS), pages 514-524. IEEE, 2019. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 291, + 294, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 294, + 346 + ], + "type": "text", + "content": "[5] Xuesong Chen, Shaoshuai Shi, Benjamin Zhu, Ka Chun Cheung, Hang Xu, and Hongsheng Li. Mppnet: Multi-frame feature intertwining with proxy points for 3d temporal object detection. In European Conference on Computer Vision, pages 680-697. Springer, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 346, + 294, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 294, + 379 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 294, + 379 + ], + "type": "text", + "content": "[6] Spconv Contributors. Spconv: Spatially sparse convolution library. https://github.com/traveller59/spconv, 2022.4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 380, + 294, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 294, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 294, + 413 + ], + "type": "text", + "content": "[7] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 414, + 294, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 294, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 294, + 468 + ], + "type": "text", + "content": "[8] Minh-Quan Dao, Julie Stephany Berrio, Vincent Frémont, Mao Shan, Elwan Héry, and Stewart Worrall. Practical collaborative perception: A framework for asynchronous and multi-agent 3d object detection. IEEE Transactions on Intelligent Transportation Systems, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 469, + 294, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 294, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 294, + 523 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 525, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 294, + 568 + ], + "type": "text", + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 569, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 294, + 624 + ], + "type": "text", + "content": "[11] Shixin Hong, Yu Liu, Zhi Li, Shaohui Li, and You He. Multiagent collaborative perception via motion-aware robust communication network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15301-15310, 2024. 3, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 678 + ], + "type": "text", + "content": "[12] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. Advances in neural information processing systems, 35:4874-4886, 2022. 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "text", + "content": "[13] Tao Huang, Jianan Liu, Xi Zhou, Dinh C Nguyen, Mostafa Rahimi Azghadi, Yuxuan Xia, Qing-Long Han, and Sumei Sun. V2x cooperative perception for autonomous" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "text", + "content": "driving: Recent advances and challenges. arXiv preprint arXiv:2310.03525, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 96, + 555, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 555, + 151 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 555, + 151 + ], + "type": "text", + "content": "[14] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 152, + 554, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 554, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 554, + 195 + ], + "type": "text", + "content": "[15] Zixing Lei, Shunli Ren, Yue Hu, Wenjun Zhang, and Siheng Chen. Latency-aware collaborative perception. In Computer Vision - ECCV 2022, pages 316-332, Cham, 2022. Springer Nature Switzerland. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 554, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 554, + 251 + ], + "type": "text", + "content": "[16] Jinyu Li, Chenxu Luo, and Xiaodong Yang. Pillarnext: Rethinking network designs for 3d object detection in lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17567-17576, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 253, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 554, + 297 + ], + "type": "text", + "content": "[17] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 298, + 554, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 298, + 554, + 353 + ], + "spans": [ + { + "bbox": [ + 317, + 298, + 554, + 353 + ], + "type": "text", + "content": "[18] Si Liu, Chen Gao, Yuan Chen, Xingyu Peng, Xianghao Kong, Kun Wang, Runsheng Xu, Wentao Jiang, Hao Xiang, Jiaqi Ma, et al. Towards vehicle-to-everything autonomous driving: A survey on collaborative perception. arXiv preprint arXiv:2308.16714, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 354, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 376 + ], + "type": "text", + "content": "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 377, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 377, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 317, + 377, + 553, + 431 + ], + "type": "text", + "content": "[20] Yifan Lu, Yue Hu, Yiqi Zhong, Dequan Wang, Yanfeng Wang, and Siheng Chen. An extensible framework for open heterogeneous collaborative perception. In The Twelfth International Conference on Learning Representations, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 434, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 553, + 499 + ], + "type": "text", + "content": "[21] Tao Ma, Xuemeng Yang, Hongbin Zhou, Xin Li, Botian Shi, Junjie Liu, Yuchen Yang, Zhizheng Liu, Liang He, Yu Qiao, et al. Detzero: Rethinking offboard 3d object detection with long-term sequential point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6736-6747, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 500, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 554, + 555 + ], + "type": "text", + "content": "[22] Charles R Qi, Yin Zhou, Mahyar Najibi, Pei Sun, Khoa Vo, Boyang Deng, and Dragomir Anguelov. Offboard 3d object detection from point cloud sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6134-6144, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 557, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 553, + 623 + ], + "type": "text", + "content": "[23] Hang Qiu, Po-Han Huang, Namo Asavisanu, Xiaochen Liu, Konstantinos Psounis, and Ramesh Govindan. Autocast: scalable infrastructure-less cooperative perception for distributed collaborative driving. In Proceedings of the 20th Annual International Conference on Mobile Systems, Applications and Services, pages 128-141, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 624, + 554, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 554, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 554, + 678 + ], + "type": "text", + "content": "[24] Yao Rong, Xiangyu Wei, Tianwei Lin, Yueyu Wang, and Enkelejda Kasneci. Dynstatf: An efficient feature fusion strategy for lidar 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3237-3246, 2023. 2, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "text", + "content": "[25] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12056" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 700 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 296, + 106 + ], + "type": "text", + "content": "intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 163 + ], + "type": "text", + "content": "[26] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 296, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 296, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 296, + 198 + ], + "type": "text", + "content": "[27] Richard Sinkhorn and Paul Knopp. Concerning nonnegative matrices and doubly stochastic matrices. Pacific Journal of Mathematics, 21(2):343-348, 1967. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 199, + 295, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 199, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 199, + 295, + 255 + ], + "type": "text", + "content": "[28] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, pages 369-386. SPIE, 2019. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 255, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 295, + 300 + ], + "type": "text", + "content": "[29] Zhiying Song, Fuxi Wen, Hailiang Zhang, and Jun Li. A cooperative perception system robust to localization errors. In 2023 IEEE Intelligent Vehicles Symposium (IV), pages 1-6. IEEE, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 301, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 301, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 56, + 301, + 295, + 346 + ], + "type": "text", + "content": "[30] Zhiying Song, Tenghui Xie, Hailiang Zhang, Jiaxin Liu, Fuxi Wen, and Jun Li. A spatial calibration method for robust cooperative perception. IEEE Robotics and Automation Letters, 9(5):4011-4018, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 347, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 347, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 347, + 295, + 392 + ], + "type": "text", + "content": "[31] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 393, + 295, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 295, + 427 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 295, + 427 + ], + "type": "text", + "content": "[32] Haiyan Wang and Yingli Tian. Sequential point clouds: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 428, + 295, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 428, + 295, + 505 + ], + "spans": [ + { + "bbox": [ + 56, + 428, + 295, + 505 + ], + "type": "text", + "content": "[33] Tsun-Hsuan Wang, Sivaban Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 605–621. Springer, 2020. 1, 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 506, + 295, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 506, + 295, + 562 + ], + "spans": [ + { + "bbox": [ + 56, + 506, + 295, + 562 + ], + "type": "text", + "content": "[34] Sizhe Wei, Yuxi Wei, Yue Hu, Yifan Lu, Yiqi Zhong, Siheng Chen, and Ya Zhang. Asynchrony-robust collaborative perception via bird's eye view flow. In Advances in Neural Information Processing Systems, pages 28462-28477. Curran Associates, Inc., 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 563, + 295, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 563, + 295, + 619 + ], + "spans": [ + { + "bbox": [ + 56, + 563, + 295, + 619 + ], + "type": "text", + "content": "[35] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In European conference on computer vision, pages 107-124. Springer, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 620, + 295, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 620, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 56, + 620, + 295, + 675 + ], + "type": "text", + "content": "[36] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 1, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 677, + 295, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 677, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 56, + 677, + 295, + 700 + ], + "type": "text", + "content": "[37] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 700 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 105 + ], + "type": "text", + "content": "semantic segmentation with sparse transformers. In Conference on Robot Learning, pages 989-1000. PMLR, 2023. 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 554, + 173 + ], + "type": "text", + "content": "[38] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13712–13722, 2023. 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 174, + 554, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 554, + 226 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 554, + 226 + ], + "type": "text", + "content": "[39] Runsheng Xu, Chia-Ju Chen, Zhengzhong Tu, and Ming-Hsuan Yang. V2x-vitv2: Improved vision transformers for vehicle-to-everything cooperative perception. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 228, + 554, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 554, + 293 + ], + "type": "text", + "content": "[40] Kun Yang, Dingkang Yang, Jingyu Zhang, Mingcheng Li, Yang Liu, Jing Liu, Hanqi Wang, Peng Sun, and Liang Song. Spatio-temporal domain awareness for multi-agent collaborative perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23383-23392, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 293, + 554, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 293, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 293, + 554, + 348 + ], + "type": "text", + "content": "[41] Lei Yang, Kaicheng Yu, Tao Tang, Jun Li, Kun Yuan, Li Wang, Xinyu Zhang, and Peng Chen. Bevheight: A robust framework for vision-based roadside 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21611-21620, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 349, + 554, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 349, + 554, + 402 + ], + "spans": [ + { + "bbox": [ + 316, + 349, + 554, + 402 + ], + "type": "text", + "content": "[42] Zetong Yang, Yin Zhou, Zhifeng Chen, and Jiquan Ngiam. 3d-man: 3d multi-frame attention network for object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1863-1872, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 403, + 554, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 403, + 554, + 479 + ], + "spans": [ + { + "bbox": [ + 316, + 403, + 554, + 479 + ], + "type": "text", + "content": "[43] Haibao Yu, Wenxian Yang, Hongzhi Ruan, Zhenwei Yang, Yingjuan Tang, Xu Gao, Xin Hao, Yifeng Shi, Yifeng Pan, Ning Sun, et al. V2x-seq: A large-scale sequential dataset for vehicle-infrastructure cooperative perception and forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5486-5495, 2023. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 480, + 554, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 480, + 554, + 524 + ], + "spans": [ + { + "bbox": [ + 316, + 480, + 554, + 524 + ], + "type": "text", + "content": "[44] Haibao Yu, Yingjuan Tang, Enze Xie, Jilei Mao, Ping Luo, and Zaiqing Nie. Flow-based feature fusion for vehicleinfrastructure cooperative 3d object detection. Advances in Neural Information Processing Systems, 36, 2024. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 524, + 554, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 524, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 524, + 554, + 590 + ], + "type": "text", + "content": "[45] Jingyu Zhang, Kun Yang, Yilei Wang, Hanqi Wang, Peng Sun, and Liang Song. ERMVP: Communication-efficient and collaboration-robust multi-vehicle perception in challenging environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12575–12584, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 590, + 554, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 554, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 554, + 644 + ], + "type": "text", + "content": "[46] Qingzhao Zhang, Xumiao Zhang, Ruiyang Zhu, Fan Bai, Mohammad Naserian, and Z Morley Mao. Robust real-time multi-vehicle collaboration on asynchronous sensors. In Proceedings of the 29th Annual International Conference on Mobile Computing and Networking, pages 1–15, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 644, + 554, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 644, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 316, + 644, + 554, + 700 + ], + "type": "text", + "content": "[47] Xumiao Zhang, Anlan Zhang, Jiachen Sun, Xiao Zhu, Y Ethan Guo, Feng Qian, and Z Morley Mao. Emp: Edge-assisted multi-vehicle perception. In Proceedings of the 27th Annual International Conference on Mobile Computing and Networking, pages 545–558, 2021. 1" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12057" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Track Any Anomalous Object_A Granular Video Anomaly Detection Pipeline/1ddd1714-9e44-4427-b96d-74620c052e99_content_list.json b/2025/Track Any Anomalous Object_A Granular Video Anomaly Detection Pipeline/1ddd1714-9e44-4427-b96d-74620c052e99_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8b714e52f3efef59123cd11da6673895853754bb --- /dev/null +++ b/2025/Track Any Anomalous Object_A Granular Video Anomaly Detection Pipeline/1ddd1714-9e44-4427-b96d-74620c052e99_content_list.json @@ -0,0 +1,1465 @@ +[ + { + "type": "text", + "text": "Track Any Anomalous Object: A Granular Video Anomaly Detection Pipeline", + "text_level": 1, + "bbox": [ + 104, + 130, + 893, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuzhi Huang $^{1*}$ , Chenxin Li $^{2*†}$ , Haitao Zhang $^{1}$ , Zixu Lin $^{1}$ , Yunlong Lin $^{1}$ , Hengyu Liu $^{2}$ , Wuyang Li $^{2}$ , Xinyu Liu $^{2}$ , Jiechao Gao $^{3}$ , Yue Huang $^{1†}$ , Xinghao Ding $^{1}$ and Yixuan Yuan $^{2}$", + "bbox": [ + 99, + 186, + 911, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Key Laboratory of Multimedia Trusted Perception and Efficien Computing, Ministry of Education of China, Xiamen University School of Informatics, Xiamen University", + "bbox": [ + 200, + 224, + 810, + 279 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 The Chinese University of Hong Kong 3 The University of Virginia", + "bbox": [ + 223, + 277, + 789, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 345, + 326, + 362 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video anomaly detection (VAD) is crucial in scenarios such as surveillance and autonomous driving, where timely detection of unexpected activities is essential. Albeit existing methods have primarily focused on detecting anomalous objects in videos—either by identifying anomalous frames or objects—they often neglect finer-grained analysis, such as anomalous pixels, which limits their ability to capture a broader range of anomalies. To address this challenge, we propose an innovative VAD framework called Track Any Anomalous Object (TAO), which introduces a Granular Video Anomaly Detection Framework that, for the first time, integrates the detection of multiple fine-grained anomalous objects into a unified framework. Unlike methods that assign anomaly scores to every pixel at each moment, our approach transforms the problem into pixel-level tracking of anomalous objects. By linking anomaly scores to subsequent tasks such as image segmentation and video tracking, our method eliminates the need for threshold selection and achieves more precise anomaly localization, even in long and challenging video sequences. Experiments on extensive datasets demonstrate that TAO achieves state-of-the-art performance, setting a new progress for VAD by providing a practical, granular, and holistic solution. For more information, visit the project page at: https://tao-25.github.io/", + "bbox": [ + 88, + 378, + 485, + 741 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 750, + 220, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video Anomaly Detection (VAD) involves identifying unusual or unexpected activities in surveillance videos and has significant applications in areas such as security monitoring (e.g., detecting violent behavior) and autonomous driving (e.g., recognizing traffic accidents). Current VAD research has developed along two main directions, which", + "bbox": [ + 89, + 773, + 483, + 866 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e025d956f5427181d2ecc637d75e15e453a7ca7bd5a48ab72235bf197258fcec.jpg", + "image_caption": [ + "Figure 1. Analyzing the Limitations of Existing VAD Models. Video anomaly detection (VAD) models are predominantly frame-centric or object-centric. Frame-centric methods detect anomalies in frames without localizing them, while object-centric methods identify anomalous objects but lack pixel-level accuracy. Pixel-centric models address these gaps by providing pixel-level localization, delivering fine-grained segmentation and precise delineation of anomalies, particularly for overlapping objects where traditional methods struggle." + ], + "image_footnote": [], + "bbox": [ + 524, + 332, + 901, + 510 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "are reflected in both methodologies and evaluation benchmarks. Frame-centric methods [1, 2, 11, 32, 34, 39, 44, 57] and their corresponding frame-level benchmarks [18, 36, 49, 57] focus on analyzing entire frames to detect global anomalies (e.g., fires or smoke), typically evaluated using frame-level AUC metrics. In contrast, object-centric methods [3, 8-10, 19, 38, 41, 43, 48] and their object-level benchmarks [10, 19, 40] leverage pre-trained feature extractors to detect object-specific anomalies (e.g., human falls or vehicle accidents), evaluated using metrics such as TBDC [40] and RBDC [40], which comprehensively measure temporal consistency and spatial localization precision.", + "bbox": [ + 511, + 672, + 906, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite these advances, a critical gap remains in real-world applications of VAD where fine-grained localization of anomalous objects is essential. As illustrated in Fig. 1,", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution † Corresponding author", + "bbox": [ + 89, + 875, + 344, + 888 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8689", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "current approaches suffer from fundamental shortage as frame-centric methods can only identify the presence of anomalies without localizing specific regions, while object-centric methods, though more precise, lack pixel-level accuracy. These limitations become particularly pronounced in complex scenarios with overlapping anomalies, where accurate delineation of object boundaries and shapes is crucial [4, 12, 14, 20, 28, 46, 50]. This observation motivates us to consider a more challenging yet practical scenario: Can we achieve both object-level structural integrity and pixel-level precision in a granular anomaly detection?", + "bbox": [ + 89, + 90, + 480, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this challenge, we propose enhancing existing VAD benchmarks with pixel-level anomaly evaluation. While a straightforward approach would be to employ an anomaly segmentation task for pixel-wise classification [6, 13, 16, 22, 27, 62], this method falls short when dealing with multiple, potentially overlapping anomalous objects. Instead, inspired by instance segmentation techniques [15, 17, 24, 45, 52], we develop a novel evaluation framework that simultaneously considers both object-level and pixel-level anomaly detection accuracy. This framework ensures that detected anomalies accurately correspond to their true shapes, precise positions, and spatial distributions, while effectively minimizing false positives caused by local feature similarities or noise.", + "bbox": [ + 89, + 258, + 482, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, implementing precise pixel-level tracking in videos presents significant challenges, particularly in maintaining semantic and temporal consistency across frames. Traditional supervised approaches require extensive labeled data, which is scarce in existing datasets with pixel-level annotations. To overcome these challenges, we leverage SAM2 [42], a large-scale pre-trained segmentation model capable of processing both static images and video streams in real-time without requiring additional fine-tuning on specific anomaly datasets. Its ability to precisely delineate object contours and structures makes it particularly effective in handling complex scenarios involving overlapping anomalies or occlusions.", + "bbox": [ + 89, + 473, + 482, + 669 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building upon these foundations, we propose TAO, an integrated framework that combines object-centric anomaly detection algorithms with SAM2 to create a simple yet effective fine-grained video anomaly detection system. For videos containing anomalous frames, our object-centric detection model generates bounding boxes for detected anomalies. After robust filtering, these boxes serve as prompting boxes for SAM2, which then generates segmentation masks for the anomalous objects. Evaluated on our proposed benchmark, our method achieves state-of-the-art performance on two widely-used datasets, demonstrating its effectiveness in bridging the gap between object-level detection and pixel-precise segmentation. The contributions of our paper are as follows:", + "bbox": [ + 89, + 671, + 482, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We present a new testing standard that unifies pixel-level", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and object-level assessment in video anomaly detection, addressing the limitations of existing metrics in complex scenarios with multiple or overlapping anomalies.", + "bbox": [ + 526, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce TAO, a streamlined framework that integrates object-centric anomaly detection with the segmentation capabilities from vision foundation model, enabling precise pixel-level tracking of anomalous objects without requiring additional fine-tuning.", + "- Through extensive experiments on UCSD Ped2 and ShanghaiTech Campus datasets, we demonstrate that our approach achieves state-of-the-art performance under both traditional metrics and our proposed benchmark." + ], + "bbox": [ + 511, + 136, + 903, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 286, + 653, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Anomaly Detection. Video anomaly detection (VAD) presents significant challenges due to the rarity of anomaly data and the wide variety of abnormal events, which hinder the generalization capabilities of existing models across diverse scenarios. Traditional VAD methods can be broadly classified into frame-centric and object-centric approaches. Frame-centric methods [1, 11, 32, 34, 39, 44, 55-57] analyze entire frames or sequences, often leveraging reconstruction or prediction errors to identify anomalies. These methods are particularly effective for detecting global events, such as fires or smoke. However, they struggle with localized anomalies and are prone to interference from normal regions within a frame, limiting their precision in complex scenarios [11, 32, 34, 39, 53]. Object-centric methods [7-10, 19, 41, 43, 54, 58] focus on specific objects within frames by employing pre-trained object detection models to extract bounding boxes and assess their abnormality. This targeted approach is better suited for detecting anomalies related to individuals or objects, such as human falls or vehicle accidents. By concentrating on specific regions and reducing redundant information, object-centric methods achieve improved accuracy and robustness, especially in complex, multi-object environments. Building on these foundations, our approach is the first to integrate large-scale pre-trained models for pixel-level fine-grained detection in VAD. This advancement enables precise anomaly localization, bridging the gap between frame-level analysis and object-level detection, and significantly enhancing overall detection performance.", + "bbox": [ + 511, + 311, + 903, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision Foundation Models. In recent years, large-scale pre-trained language models (LLMs) and vision-language models (VLMs) have shown remarkable potential in advancing video anomaly detection. Models such as BLIP-2 [26], LLaVA [29], LAVAD [59], VadCLIP [51], VideoChatGPT [35], and Video-LLaMA [61] have significantly improved the understanding of complex visual tasks by integrating visual and linguistic information. Concurrently, prompting techniques have gained traction in large-scale vision models, enabling substantial progress in tasks such", + "bbox": [ + 511, + 750, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8690", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d519cad781a79038c645689825d2eb24ba22e10e45c532beeec93b559cf60ddd.jpg", + "image_caption": [ + "Figure 2. Pipeline of our proposed TAO. We first generate bounding boxes to identify objects in each frame. Next, we score these boxes using object-centric video anomaly detection algorithms to extract potential anomalous boxes. To ensure robustness, we apply filtering to eliminate redundant boxes. Finally, the filtered boxes and original frames are input into a prompt-based segmentation model to produce pixel-level anomaly segmentation masks." + ], + "image_footnote": [], + "bbox": [ + 101, + 95, + 563, + 498 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d07cfa6c08efc611527ce94c625577b26a2954cc59c421483f06cfa248da8c35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 95, + 901, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "as image segmentation. By employing semantic prompts (e.g., free-form text) and spatial prompts (e.g., points or bounding boxes), these models achieve highly accurate segmentation guided by input cues. Segmentation Foundation Models (SFMs), such as the Segment Anything Model (SAM) [21] and SEEM [64], have demonstrated exceptional performance, particularly in zero-shot generalization and multi-modal prompting, allowing for cross-task segmentation. Building on these advancements, the Segment Anything Model 2 (SAM2) [42] further improves segmentation, particularly for video-based applications. SAM2 integrates prompting techniques to not only segment static objects with precision but also capture dynamic, complex objects in video contexts. This capability enables SAM2 to deliver finer segmentation and enhanced localization accuracy in video anomaly detection, addressing the challenges posed by real-world variability and complexity. This work leverages SAM2's advanced video segmentation capabilities to tackle the challenge of fine-grained video anomaly detection, achieving precise and robust anomaly identification in", + "bbox": [ + 89, + 599, + 485, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "complex scenarios.", + "bbox": [ + 513, + 599, + 643, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 631, + 604, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preliminary. In the proposed video anomaly detection (VAD) benchmark, a video clip is represented as a sequence $\\{f_1,f_2,\\ldots ,f_{N_f}\\} \\subseteq \\mathcal{C}$ , where $\\mathcal{C}$ is the set of all possible video clips, and $N_{f}$ denotes the total number of frames. Each frame $f_{i}$ is expressed as $f_{i} = [p_{i,1},p_{i,2},\\dots ,p_{i,N_{i}}]$ , where $N_{i}$ is the total number of pixels in the frame, and $p_{i,j}\\in \\mathcal{P}$ represents the $j$ -th pixel, with $\\mathcal{P}$ as the set of possible pixel values. Each pixel $p_{i,j}$ is classified as either \"normal\" or \"anomalous\".", + "bbox": [ + 511, + 657, + 906, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. To achieve precise anomaly localization in videos, we design a streamlined pipeline for pixel-level segmentation in video anomaly detection. Our pipeline comprises four stages: bounding boxes extraction, anomalous boxes extraction, robust filtering, and segmentation inference. The process begins by generating bounding boxes to identify objects of interest. These boxes are then scored us", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8691", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6adbbf4d76a6cb8125318ce77626324346965b73396d3c8cda8c9bec6039ebd1.jpg", + "image_caption": [ + "Figure 3. Impact of Redundant Segmentation on Performance. Potential anomalous boxes are used as prompts for SAM2 segmentation (red masks: true anomalies, green masks: redundant results). Initially, SAM2 tracks true anomalies accurately, but tracking errors accumulate, leading to catastrophic forgetting and a collapse in Pixel-F1 performance." + ], + "image_footnote": [], + "bbox": [ + 102, + 90, + 691, + 189 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/daaff1e3489f33ad5e52695132fea706efa7667581fbb2a3cb06768da8caff8b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 692, + 90, + 898, + 190 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Boxes Robustness Filtering", + "text_level": 1, + "bbox": [ + 91, + 272, + 364, + 286 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Input: A set of anomalous bounding boxes $\\mathcal{B}_{\\mathrm{anomaly}} = \\{b_i \\mid s_i > \\tau\\}$ for each frame $f_i \\in \\{f_1, f_2, \\ldots, f_n\\}$ , where $s_i$ is the anomaly score and $\\tau$ is the threshold; parameters $k$ (tracking window size), $h$ (overlap threshold), $l$ (save interval), and $m$ (frame match threshold)", + "2: Output: Filtered bounding boxes $\\mathcal{B}_{\\mathrm{filtered}}$ with their corresponding object labels $\\mathcal{L}_j$", + "3: Initialize the list $\\mathcal{B}_i$ to store the filtered bounding boxes from frame $f_{i}$", + "4: for each frame $f_{i} \\in \\{f_{1}, \\ldots, f_{n}\\}$ do", + "5: Step 1: Inherit Existing Anomalous Boxes", + "6: for each bounding box $b_{j} \\in \\mathcal{B}_{\\mathrm{anomaly}}$ and $b_{p}$ from frame $f_{i - k}$ to $f_{i - 1}$ do" + ], + "bbox": [ + 99, + 290, + 483, + 482 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n7: \\quad \\text {i f} \\sum_ {p = i - k} ^ {i - 1} \\mathbb {I} (\\operatorname {I o U} (b _ {j}, b _ {p}) > h) \\geq m \\text {t h e n}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 482, + 421, + 515 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "8: Assign $\\mathcal{L}_j = \\mathcal{L}_p$", + "9: Remove $b_{j}$ from $\\mathcal{B}_{i}$" + ], + "bbox": [ + 101, + 515, + 323, + 545 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "10: end if", + "11: end for", + "12: Step 2: Assign New Anomalous Boxes", + "13: for each remaining $b_{j} \\in \\mathcal{B}_{\\mathrm{anomaly}}$ and $b_{p}$ from frame $f_{i + 1}$ to $f_{i + k}$ do" + ], + "bbox": [ + 96, + 546, + 482, + 619 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {i f} \\sum_ {p = i + 1} ^ {i + k} \\mathbb {I} \\left(\\operatorname {I o U} \\left(b _ {j}, b _ {p}\\right) > h\\right) \\geq m \\text {t h e n} \\tag {14:}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 619, + 421, + 652 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "15: Assign a new object label $\\mathcal{L}_j = \\mathcal{F}(b_j)$", + "16: Save the tuple $\\mathcal{T}_{\\mathrm{box}} = (f_i, b_j, \\mathcal{L}_j)$", + "17: end if", + "18: end for", + "19: Step 3: Save Boxes Every $l$ Frames", + "20: if $i \\bmod l == 0$ then", + "21: Save all bounding box tuples $\\mathcal{T}_{\\mathrm{box}} = (f_i, b_j, \\mathcal{L}_j)$ in the current frame" + ], + "bbox": [ + 94, + 619, + 480, + 770 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "22: end if", + "23: end for" + ], + "bbox": [ + 96, + 772, + 189, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing object-centric anomaly detection algorithms to extract potential anomalies. A tracking-based filtering step follows to eliminate redundant boxes and retain only robust ones. Finally, the filtered boxes and original frames are fed into", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a prompt-based segmentation model to generate pixel-level anomaly masks. The pipeline is illustrated in Fig. 2.", + "bbox": [ + 511, + 272, + 906, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Anomalous Boxes Extraction", + "text_level": 1, + "bbox": [ + 511, + 311, + 772, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Bounding Boxes Extraction. To achieve precise pixel-level segmentation in video anomaly detection, pre-trained object detection algorithms are used to generate bounding boxes for objects within a frame, serving as potential prompts for guiding the segmentation model. For a given input frame $f$ , the object detection model $\\mathcal{D}$ outputs $m$ bounding boxes $b_{1}, b_{2}, \\ldots, b_{m}$ , each paired with a class label $y_{1}, y_{2}, \\ldots, y_{m}$ :", + "bbox": [ + 511, + 333, + 906, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\{(b _ {1}, y _ {1}), (b _ {2}, y _ {2}), \\dots , (b _ {m}, y _ {m}) \\} = \\mathcal {D} (f) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 463, + 905, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After the extraction of object bounding boxes in each frame, object-centric VAD algorithms compute an anomaly score $s_i$ for each object within its bounding box $b_i$ . This score is derived from features such as pose, speed, and depth, and can be formally represented as:", + "bbox": [ + 511, + 486, + 905, + 561 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i} = \\mathcal {A} \\left(b _ {i}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 669, + 574, + 903, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{A}$ represents the scoring function used in object-centric video anomaly detection algorithms. To isolate anomalous objects, we filter the bounding boxes based on an anomaly score threshold. Specifically, we retain only the bounding boxes where the anomaly score exceeds a predefined threshold $\\tau$ , as shown in the following expression:", + "bbox": [ + 511, + 602, + 905, + 693 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {B} _ {\\text {a n o m a l y}} = \\left\\{b _ {i} \\mid s _ {i} > \\tau \\right\\} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 704, + 903, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tau$ denotes the anomaly score threshold, $s_i$ is the anomaly score for the $i$ -th object, and $b_i$ is the corresponding bounding box.", + "bbox": [ + 511, + 733, + 905, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Robust Filtering of Anomalous Boxes", + "text_level": 1, + "bbox": [ + 511, + 787, + 834, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The anomalous boxes detection from Sec. 3.1 contains numerous false positives due to overlapping characteristics between normal and abnormal boxes. This overlap complicates threshold setting and leads to misclassifications. These redundant boxes can significantly degrade segmentation performance by generating inaccurate results and", + "bbox": [ + 511, + 809, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8692", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "overwhelming the model's tracking capabilities, potentially causing the loss of anomalous targets in subsequent frames.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Fig. 3, using potential anomalous boxes as SAM2 prompts at fixed intervals reveals the impact of tracking errors. While red masks indicate true anomalies and green masks show redundant results, the model initially tracks anomalous objects effectively. However, tracking errors accumulate over time, leading to an increasing number of incorrectly tracked objects. This deterioration ultimately results in catastrophic forgetting of true anomalies, reflected in steadily declining Pixel-F1 scores.", + "bbox": [ + 88, + 122, + 482, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enhance robustness and address redundancy, we present the Boxes Robustness Filtering algorithm, which exploits the temporal consistency of true anomalous boxes. Unlike redundant boxes that appear sporadically, true anomalous boxes maintain consistent tracking of the same target. The algorithm, detailed in Alg. 1, operates in three key steps. First, the inheritance step computes the similarity between each bounding box $b_{j}$ in the current frame $f_{i}$ and bounding boxes $b_{p}$ from the previous $k$ frames. If the Intersection over Union (IoU) exceeds a threshold $h$ for at least $m$ frames, $b_{j}$ inherits the label $\\mathcal{L}_{p}$ from earlier frames, ensuring consistent tracking of previously identified anomalies. Second, the assignment step addresses bounding boxes that fail to inherit a label by comparing them with bounding boxes in the subsequent $k$ frames. If the IoU exceeds $h$ for at least $m$ frames, a new label $\\mathcal{L}_{j} = \\mathcal{F}(b_{j})$ is assigned, effectively labeling newly emerging anomalies. Finally, the saving step records bounding boxes and their labels every $l$ -th frame, systematically capturing all identified anomalies. By iteratively applying these steps, the Boxes Robustness Filtering algorithm achieves robust tracking and localization of anomalous objects. Each bounding box and its label are stored as a tuple $\\mathcal{T}_{\\mathrm{box}} = (f_i,b_j,\\mathcal{L}_j)$ , ensuring spatial and temporal consistency across video frames. This approach reduces redundancy and strengthens anomaly detection accuracy, providing a reliable basis for subsequent segmentation and analysis.", + "bbox": [ + 91, + 261, + 485, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Inference of the Segmentation Model", + "text_level": 1, + "bbox": [ + 89, + 680, + 410, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using the tuple $\\mathcal{T}_{\\mathrm{box}} = (f_i, b_j, \\mathcal{L}_j)$ obtained from Sec. 3.2, we extract the center point of each bounding box $b_j$ . Specifically, given the coordinates of the box $b_j = [x_1, y_1, x_2, y_2]$ , the center point is calculated: $\\mathbf{c}_j = \\left(\\frac{x_1 + x_2}{2}, \\frac{y_1 + y_2}{2}\\right)$ . Effectively, these prompts are not saved for every frame but are instead collected at regular intervals, specifically every $l$ -th frame, as outlined in Sec. 3.2. This ensures efficient storage and processing while still capturing the necessary information for accurate segmentation [23, 25]. For each selected frame, both the center point $\\mathbf{c}_j$ and the bounding box $b_j$ are stored as prompts. Rather than processing each frame individually, we aggregate prompts from all saved frames $\\{f_1, f_{1+l}, f_{1+2l}, \\ldots, f_{N_f}\\}$ , where $l$ denotes the saving in", + "bbox": [ + 89, + 703, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "terval. The aggregated prompts are:", + "bbox": [ + 511, + 90, + 751, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} = \\left\\{\\left(\\mathbf {c} _ {j}, b _ {j}, f _ {i}\\right) \\mid i = 1, 1 + l, 1 + 2 l, \\dots , N _ {f} \\right\\} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 118, + 906, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{I}$ represents the set of prompts that includes the center points, bounding box coordinates $b_{j}$ , and original frames $f_{i}$ for the selected frames.", + "bbox": [ + 511, + 147, + 903, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By inputting this prompt information $\\mathcal{I}$ saved at regular intervals into SAM2, which serves as the prompt-based segmentation model $\\mathcal{M}$ , we leverage its feature propagation capabilities to efficiently generate pixel-level segmentation results for the entire video. SAM2 propagates the segmentation results from the prompted frames to the remaining frames, producing pixel-level segmentation across the entire video:", + "bbox": [ + 511, + 193, + 906, + 311 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} = \\mathcal {M} (\\mathcal {I}) = \\left\\{\\mathcal {S} _ {1}, \\mathcal {S} _ {2}, \\dots , \\mathcal {S} _ {N _ {f}} \\right\\} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 325, + 906, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $S_{i} = [p_{1}, p_{2}, \\ldots, p_{M}]$ denotes the segmentation result for frame $f_{i}$ , where each pixel $p_{j}$ is either \"normal\" or \"anomalous\". This approach enables the generation of the complete segmentation set $S$ , efficiently utilizing contextual information from the prompts.", + "bbox": [ + 511, + 354, + 906, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 445, + 645, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. A New Comprehensive VAD Benchmark", + "text_level": 1, + "bbox": [ + 511, + 469, + 859, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benchmark Design. We present a novel benchmark framework that integrates both pixel-level and object-level evaluation metrics to advance VAD model assessment. This dual-level approach addresses the limitations of using single evaluation metrics. Pixel-level metrics excel at identifying subtle irregularities through independent pixel evaluation, but struggle to capture broader spatial patterns. Object-level metrics complement this by assessing global characteristics like shape and position, enabling coherent anomaly tracking and recognition. By combining these complementary perspectives, our framework provides a comprehensive evaluation of both fine-grained details and holistic anomaly patterns, particularly valuable for real-world applications.", + "bbox": [ + 511, + 491, + 906, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benchmark Datasets. UCSD Ped2 [47] contains 16 training and 12 test videos (240×360 pixels) captured by a fixed overhead camera. The training set consists of normal pedestrian activities, while the test set includes anomalous events such as bikers, skateboarders, and vehicles. ShanghaiTech Campus [30], one of the largest VAD datasets, comprises 330 training and 107 test videos (480×856 pixels). While the training set features normal scenarios, the test set contains various anomalies including robbery, fighting, and unauthorized cycling in pedestrian zones.", + "bbox": [ + 511, + 688, + 906, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benchmark Metrics. For pixel-level evaluation, we employ four metrics: Pixel-AUROC measures discrimination between normal and anomalous pixels across thresholds, Pixel-AP assesses detection precision by balancing false", + "bbox": [ + 511, + 839, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8693", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ca47cc552283084344bf2a84c30cf3fe841af99c1dd7705aa375cd0447b02c4e.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison with anomaly detection models adapted for video anomaly detection on the UCSD Ped2 dataset." + ], + "image_footnote": [], + "bbox": [ + 102, + 119, + 906, + 388 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9ce8179e09099c3d993f6cb6c8dc05eaef753ba741631cd15e6d6eb15d15bc16.jpg", + "table_caption": [ + "Table 1. Quantitative comparison on video anomaly detection with pixel-level and object-level metrics on the UCSD Ped2 dataset." + ], + "table_footnote": [], + "table_body": "
MethodPixel-levelObject-level
Pixel-AUROC ↑Pixel-AP ↑Pixel-AUPRO ↑Pixel-F1 ↑RBDC ↑TBDC ↑
AdaCLIP (Zero-shot) [5]51.021.3233.982.615.810.6
AdaCLIP (Fully fine-tuned) [5]53.064.9750.6611.1912.315.5
AnomalyCLIP (Zero-shot) [63]51.6321.2036.345.927.511.2
AnomalyCLIP (Fully fine-tuned) [63]54.2523.7338.597.4813.121.0
DDAD (Fully trained) [37]55.875.6115.122.6718.0113.29
SimpleNet (Fully trained) [33]52.4920.5144.0510.7151.1827.75
DRAEM (Fully trained) [60]69.5830.6335.7810.8944.2670.64
TAO (Partially fine-tuned)75.1150.7872.9764.1283.693.2
", + "bbox": [ + 116, + 426, + 885, + 580 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c19c8f93b5ffbdad881219fcc7054e37868c49e658c49616437688764a31b4b3.jpg", + "table_caption": [ + "Table 2. Quantitative comparison on video anomaly detection with RBDC and TBDC metrics on the UCSD Ped2 dataset." + ], + "table_footnote": [], + "table_body": "
MethodRBDC ↑TBDC ↑
OCAD [19]52.772.8
SS-VAD [40]62.580.5
SiVL [41]7489.3
AED-SSMTL [10]72.891.2
TAO (Ours)83.693.2
", + "bbox": [ + 135, + 642, + 439, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "positives and negatives, Pixel-AUPRO quantifies segmentation accuracy through region overlap, and Pixel-F1 provides an integrated precision-recall measure. For object-level assessment, we utilize Region-Based Detection Criterion (RBDC) and Track-Based Detection Criterion (TBDC). RBDC measures spatial accuracy using Intersection over Union (IoU) with threshold $\\alpha$ , while TBDC evaluates temporal consistency in tracking anomalous regions across", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fefe2174b3d0a8f56ca09e83f6b73f5d78410db33c69988bda16b7e8fbdd86b5.jpg", + "table_caption": [ + "Table 3. Quantitative comparison on video anomaly detection with RBDC and TBDC metrics on the ShanghaiTech Campus dataset." + ], + "table_footnote": [], + "table_body": "
MethodRBDC ↑TBDC ↑
OCAD [19]20.744.5
BAF-AT [9]41.378.8
AED-SSMTL [10]43.284.1
HF2VAD [31]45.484.5
STPT [38]51.684.6
TAO (Ours)62.185.4
", + "bbox": [ + 558, + 643, + 862, + 767 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "frames. Together, these metrics provide comprehensive evaluation of both spatial precision and temporal coherence.", + "bbox": [ + 511, + 791, + 906, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 832, + 732, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our framework builds upon SAM2 [42], utilizing an object-level anomaly detection algorithm [43] to extract anomalous bounding boxes as segmentation prompts. We em", + "bbox": [ + 511, + 854, + 905, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8694", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ploy SAM2 [42] as the prompt-based segmentation model, which integrates these prompts with original frames for precise anomaly segmentation. Our implementation adopts a partial fine-tuning strategy, where only the object-level VAD algorithm is fine-tuned on target datasets, while SAM2 retains its pre-trained weights for inference. In the bounding box threshold filtering stage, we assign anomaly scores based on pose and depth features, with thresholds set to $\\tau = 1.5$ for UCSD Ped2 and $\\tau = 1.6$ for ShanghaiTech Campus. The subsequent robustness filtering stage employs a tracking window of $k = 5$ , frame match threshold of $m = 3$ , and box overlap threshold of $h = 0.2$ , with save intervals of $l = 5$ and $l = 15$ for UCSD Ped2 and ShanghaiTech Campus respectively. For segmentation, we utilize SAM2 with hiera base-plus weights on an NVIDIA RTX 4090 GPU. To ensure robust evaluation, overlapping boxes of the same anomaly are merged, and segmentation probability maps are uniformly binarized. For object-level metrics, we derive bounding boxes from segmentation extremal coordinates. Given anomaly pixels $P = (x_{1},y_{1}),(x_{2},y_{2}),\\ldots ,(x_{N},y_{N})$ , we compute the box coordinates as: $x_{\\mathrm{min}} = \\min_{(x_i,y_i)\\in P}x_i$ , $x_{\\mathrm{max}} = \\max_{(x_i,y_i)\\in P}x_i$ , $y_{\\mathrm{min}} = \\min_{(x_i,y_i)\\in P}y_i$ , $y_{\\mathrm{max}} = \\max_{(x_i,y_i)\\in P}y_i$ , where the resulting bounding box $(x_{\\mathrm{min}},y_{\\mathrm{min}},x_{\\mathrm{max}},y_{\\mathrm{max}})$ is used for computing RBDC and TBDC metrics, providing robust spatial evaluation at the object level.", + "bbox": [ + 93, + 90, + 480, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison to Anomaly Detection Models", + "text_level": 1, + "bbox": [ + 93, + 513, + 450, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given the lack of pixel-level segmentation capabilities in current video anomaly detection models, we benchmark several state-of-the-art image anomaly detection algorithms, including AdaCLIP [5], AnomalyCLIP [63], DRAEM [60], DDAD [37], and SimpleNet [33], to evaluate both pixel-level and object-level metrics on the UCSD Ped2 dataset. AdaCLIP and AnomalyCLIP were selected for their pre-trained vision-language architectures, which closely align with SAM2's design and demonstrate strong zero-shot capabilities. To ensure a fair comparison, we assess both AdaCLIP and AnomalyCLIP under zero-shot and fully fine-tuned settings. In contrast, DRAEM, DDAD, and SimpleNet, as fully trained models, provide robust baselines for pixel-level and object-level anomaly detection, representing diverse methodological approaches. For anomaly detection, video streams are processed frame-by-frame through each algorithm, generating heatmaps that identify potential anomalies. Pixel-level segmentation is achieved by applying a threshold to these heatmaps, allowing precise localization of anomalous regions.", + "bbox": [ + 93, + 537, + 480, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Tab.1, the experimental results clearly demonstrate the advantages of our model in both pixel-level and object-level anomaly detection. While most benchmarked models struggle to balance these two aspects, of", + "bbox": [ + 93, + 840, + 480, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ten excelling in one while underperforming in the other, our model achieves robust performance across both dimensions. This balance reflects its strong and versatile design, enabling precise pixel-level segmentation and maintaining spatial-temporal consistency at the object level. Moreover, our partial fine-tuning approach effectively leverages pretrained features, minimizing the need for extensive dataset-specific training. In contrast, fully trained models such as DRAEM, DDAD, and SimpleNet, as well as vision-language models like AdaCLIP and AnomalyCLIP, fail to achieve comparable results. Fig. 4 further illustrates the strengths of our method. The segmentation results produced by our model are more complete and detailed, accurately delineating object boundaries while preserving the integrity of anomalous objects.", + "bbox": [ + 516, + 90, + 903, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Comparison to Conventional Video Anomaly Detection Models", + "text_level": 1, + "bbox": [ + 516, + 321, + 903, + 351 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To assess the performance of our proposed method, we conducted comprehensive experiments on two widely used datasets, UCSD Ped2 and ShanghaiTech Campus. Our approach was benchmarked against several state-of-the-art video anomaly detection models, including SS-VAD [40], SiVL [41], AED-SSMTL [10], HF2VAD [31], BAF-AT [9], and STPT [38]. Due to the absence of fine-grained segmentation capabilities in traditional methods, the evaluation was focused on object-centric anomaly detection and limited to object-level metrics, TBDC [40] and RBDC [40]. These metrics provide a robust and meaningful basis for comparing the effectiveness of our model with existing approaches.", + "bbox": [ + 516, + 364, + 903, + 544 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method achieves state-of-the-art performance on both the UCSD Ped2 and ShanghaiTech Campus datasets. On UCSD Ped2, it outperforms existing methods in TBDC and RBDC metrics, demonstrating precise tracking and localization of anomalous objects across video frames. On the more challenging ShanghaiTech Campus dataset, characterized by complex backgrounds and subtle anomalies, our method consistently surpasses baselines by delivering balanced performance across all evaluation metrics. Unlike existing approaches, which often excel in isolated metrics but fail to generalize, our model effectively integrates object-centric detection and fine-grained segmentation, enabling robust anomaly localization and tracking under diverse real-world conditions, as shown in Tabs. 2 and 3.", + "bbox": [ + 516, + 549, + 903, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To conclude, our method establishes a new standard in video anomaly detection by combining object-centric detection with fine-grained segmentation. This integration enables precise localization and robust tracking of anomalous objects, addressing the limitations of traditional methods that struggle with balanced performance across metrics. Our approach demonstrates versatility and effectiveness across diverse datasets, highlighting its potential for real-world applications.", + "bbox": [ + 516, + 763, + 903, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8695", + "bbox": [ + 483, + 946, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b3857c62f07e8040ec03a99059073408460f68da0319f50de03db886c411a6a8.jpg", + "table_caption": [ + "Table 4. Ablation study on proposed key components, evaluated with pixel-level and object-level metrics on the UCSD Ped2 dataset." + ], + "table_footnote": [], + "table_body": "
Video Track ModeBox Robust Filtering AlgorithmPixel-levelObject-level
Pixel-AUROC ↑Pixel-AP ↑Pixel-AUPRO ↑Pixel-F1 ↑RBDC ↑TBDC ↑
××41.7225.3040.5532.9941.071.6
×✓39.0622.8038.0930.5639.371.7
✓×68.9038.4767.2352.1770.876.9
✓✓75.1150.7872.9764.1283.693.2
", + "bbox": [ + 112, + 114, + 885, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bb0f3153cbed6e9a20805bb3acac1a71f14e06be1d248bafe9dc1aa92d06c558.jpg", + "image_caption": [ + "Figure 5. Sensitivity Analysis of Anomaly Threshold $(\\tau)$ . We evaluate performance under varying threshold on ShanghaiTech Campus dataset." + ], + "image_footnote": [], + "bbox": [ + 101, + 260, + 460, + 435 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9469c4c2b498ef09a28583c1d54f6ed4cf3d9f0c9b5d36e37c227dab2b3dccf7.jpg", + "image_caption": [ + "Figure 6. Robust Testing of SAM2 Backbones. We compare the performance across different SAM2 model variants." + ], + "image_footnote": [], + "bbox": [ + 101, + 513, + 473, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Further Empirical Study", + "text_level": 1, + "bbox": [ + 89, + 727, + 326, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study. We conduct ablation experiments on the UCSD Ped2 dataset to evaluate two components: video tracking mode and box robust filtering. Video tracking mode, a SAM2 component, enables temporal analysis, while disabling it reduces the model to SAM, performing frame-by-frame segmentation. Disabling box robust filtering uses raw anomaly-thresholded boxes as prompts. Tab. 4 shows the effects of each component on pixel- and object-level metrics. The baseline model, with both components disabled, performs the worst. Enabling box robust filter", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing alone yields minimal improvement, while activating video tracking mode significantly boosts pixel-level accuracy. The best performance is achieved with both components enabled, demonstrating their synergistic effect. These results confirm the necessity of both temporal tracking and robust filtering in our framework.", + "bbox": [ + 511, + 255, + 903, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sensitivity Analysis. We conduct a sensitivity analysis to assess how variations in the anomaly threshold $\\tau$ affect our framework's performance. Fig. 5 shows the impact of $\\tau$ on four metrics—Pixel-AUROC, Pixel-AP, TBDC, and RBDC—using the ShanghaiTech Campus dataset. As $\\tau$ ranges from 1.0 to 1.9, the framework performs stably, with optimal results at $\\tau = 1.6$ . TBDC is most sensitive, peaking at $85\\%$ at $\\tau = 1.6$ , while RBDC stays around $62\\%$ . Pixel-level metrics, Pixel-AUROC and Pixel-AP, also show steady trends, reflecting the framework's balance of fine-grained anomaly detection and object consistency. This analysis highlights the robustness and adaptability of our framework across threshold variations, ensuring reliable detection in diverse scenarios.", + "bbox": [ + 511, + 348, + 906, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Robustness Testing. We evaluate our framework's robustness through extensive testing across different SAM2 backbone architectures (Fig. 6). Our analysis covers models from the lightweight Tiny variant to the full-scale Large model, with Base+ consistently yielding the best results. This stability suggests that the framework's effectiveness is not dependent on backbone capacity. While larger models show slight improvements in pixel-level metrics, performance remains strong across all variants. These results validate the framework's reliability and robustness, demonstrating its suitability for deployment in a wide range of scenarios, irrespective of computational constraints.", + "bbox": [ + 511, + 561, + 908, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 766, + 633, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents TAO, a framework integrating coarse object detection with fine-grained anomaly localization in videos. Leveraging SAM2 and our dual-level evaluation, it enables precise segmentation and robust tracking. Experiments on UCSD Ped2 and ShanghaiTech datasets demonstrate state-of-the-art performance in pixel- and object-level metrics, advancing video anomaly detection.", + "bbox": [ + 511, + 794, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8696", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 89, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China under Grant 82172033, Grant U19B2031, Grant 61971369, Grant 52105126, Grant 82272071, and Grant 62271430; and in part by the Dreams Foundation of Jianghuai Advance Technology Center and in part by the Open Fund of the National Key Laboratory of Infrared Detection Technologies.", + "bbox": [ + 89, + 114, + 485, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 232, + 187, + 248 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sunghyun Ahn, Youngwan Jo, Kijung Lee, and Sanghyun Park. Videopatchcore: An effective method to memorize normality for video anomaly detection. arXiv preprint arXiv:2409.16225, 2024. 1, 2", + "[2] Jinbin Bai, Zhen Dong, Aosong Feng, Xiao Zhang, Tian Ye, Kaicheng Zhou, and Mike Zheng Shou. Integrating view conditions for image synthesis. arXiv preprint arXiv:2310.16002, 2023. 1", + "[3] Jinbin Bai, Wei Chow, Ling Yang, Xiangtai Li, Juncheng Li, Hanwang Zhang, and Shuicheng Yan. Humanedit: A high-quality human-rewarded dataset for instruction-based image editing. arXiv preprint arXiv:2412.04280, 2024. 1", + "[4] Jinbin Bai, Tian Ye, Wei Chow, Enxin Song, Qing-Guo Chen, Xiangtai Li, Zhen Dong, Lei Zhu, and Shuicheng Yan. Meissonic: Revitalizing Masked Generative Transformers for Efficient High-Resolution Text-to-Image Synthesis. arXiv preprint arXiv:2410.08261, 2024. 2", + "[5] Yunkang Cao, Jiangning Zhang, Luca Frittoli, Yuqi Cheng, Weiming Shen, and Giacomo Boracchi. Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly detection. In European Conference on Computer Vision, pages 55-72. Springer, 2025. 6, 7", + "[6] Zhiyuan Ding, Qi Dong, Haote Xu, Chenxin Li, Xinghao Ding, and Yue Huang. Unsupervised anomaly segmentation for brain lesions using dual semantic-manifold reconstruction. In International Conference on Neural Information Processing, pages 133-144. Springer International Publishing Cham, 2022. 2", + "[7] Keval Doshi and Yasin Yilmaz. Any-shot sequential anomaly detection in surveillance videos. 2020. 2", + "[8] Keval Doshi and Yasin Yilmaz. Continual learning for anomaly detection in surveillance videos. 2020. 1", + "[9] Mariana-Iuliana Georgescu, RaduTudor Ionescu, Fahad-Shahbaz Khan, Marius Popescu, and Mubarak Shah. A background-agnostic framework with adversarial training for abnormal event detection in video. IEEE Transactions on Pattern Analysis and Machine Intelligence, IEEE Transactions on Pattern Analysis and Machine Intelligence. 6, 7", + "[10] Mariana-Iuliana Georgescu, Antonio Barbalau, Radu Tudor Ionescu, Fahad Shahbaz Khan, Marius Popescu, and Mubarak Shah. Anomaly detection in video via self-supervised and multi-task learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12742-12752, 2021. 1, 2, 6, 7", + "[11] Mahmudul Hasan, Jonghyun Choi, Jan Neumann, Amit K. Roy-Chowdhury, and Larry S. Davis. Learning temporal" + ], + "bbox": [ + 93, + 258, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "regularity in video sequences. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 2", + "[12] Chunming He, Kai Li, Guoxia Xu, Yulun Zhang, Runze Hu, Zhenhua Guo, and Xiu Li. Degradation-resistant unfolding network for heterogeneous image fusion. In ICCV, pages 12611-12621, 2023. 2", + "[13] Chunming He, Kai Li, Yachao Zhang, Guoxia Xu, Longxiang Tang, Yulun Zhang, Zhenhua Guo, and Xiu Li. Weakly-supervised concealed object segmentation with sam-based pseudo labeling and multi-scale feature grouping. NeurIPS, 36, 2024. 2", + "[14] Chunming He, Kai Li, Yachao Zhang, Yulun Zhang, Zhenhua Guo, Xiu Li, Martin Danelljan, and Fisher Yu. Strategic preys make acute predators: Enhancing camouflaged object detectors by generating camouflaged objects. ICLR, 2024. 2", + "[15] Chunming He, Chengyu Fang, Yulun Zhang, Kai Li, Longxiang Tang, Chenyu You, Fengyang Xiao, Zhenhua Guo, and Xiu Li. Reti-diff: Illumination degradation image restoration with retinae-based latent diffusion model. *ICLR*, 2025. 2", + "[16] Chunming He, Yuqi Shen, Chengyu Fang, Fengyang Xiao, Longxiang Tang, Yulun Zhang, Wangmeng Zuo, Zhenhua Guo, and Xiu Li. Diffusion models in low-level vision: A survey. TPAMI, 2025. 2", + "[17] Chunming He, Rihan Zhang, Fengyang Xiao, Chenyu Fang, Longxiang Tang, Yulun Zhang, Linghe Kong, Deng-Ping Fan, Kai Li, and Sina Farsiu. Run: Reversible unfolding network for concealed object segmentation. arXiv preprint arXiv:2501.18783, 2025. 2", + "[18] Ryota Hinami, Tao Mei, and Shin'ichi Satoh. Joint detection and recounting of abnormal events by learning deep generic knowledge. In 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 1", + "[19] Radu Tudor Ionescu, Fahad Shahbaz Khan, Mariana-Iuliana Georgescu, and Ling Shao. Object-centric auto-encoders and dummy anomalies for abnormal event detection in video. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 6", + "[20] Hamza Karim, Keval Doshi, and Yasin Yilmaz. Real-time weakly supervised video anomaly detection. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 6848-6856, 2024. 2", + "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 3", + "[22] Chenxin Li, Yunlong Zhang, Jiongcheng Li, Yue Huang, and Xinghao Ding. Unsupervised anomaly segmentation using image-semantic cycle translation. arXiv preprint arXiv:2103.09094, 2021. 2", + "[23] Chenxin Li, Yunlong Zhang, Zhehan Liang, Wenao Ma, Yue Huang, and Xinghao Ding. Consistent posterior distributions under vessel-mixing: a regularization for cross-domain retinal artery/vein classification. In 2021 IEEE International Conference on Image Processing (ICIP), pages 61-65. IEEE, 2021. 5" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8697", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Chenxin Li, Wenao Ma, Liyan Sun, Xinghao Ding, Yue Huang, Guisheng Wang, and Yizhou Yu. Hierarchical deep network with uncertainty-aware semi-supervised learning for vessel segmentation. Neural Computing and Applications, pages 1-14, 2022. 2", + "[25] Chenxin Li, Xinyu Liu, Wuyang Li, Cheng Wang, Hengyu Liu, Yifan Liu, Zhen Chen, and Yixuan Yuan. U-kan makes strong backbone for medical image segmentation and generation. arXiv preprint arXiv:2406.02918, 2024. 5", + "[26] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2", + "[27] Huangxing Lin, Yunlong Lin, Jingyuan Xia, Linyu Fan, Feifei Li, Yingying Wang, and Xinghao Ding. Fusion2void: Unsupervised multi-focus image fusion based on image inpainting. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2", + "[28] Yunlong Lin, Tian Ye, Sixiang Chen, Zhenqi Fu, Yingying Wang, Wenhao Chai, Zhaohu Xing, Lei Zhu, and Xinghao Ding. Aglldiff: Guiding diffusion models towards unsupervised training-free real-world low-light image enhancement. arXiv preprint arXiv:2407.14900, 2024. 2", + "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2", + "[30] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection-a new baseline. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6536-6545, 2018. 5", + "[31] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13588-13597, 2021. 6, 7", + "[32] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2", + "[33] Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang. Simplenet: A simple network for image anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20402-20411, 2023. 6, 7", + "[34] Shao-Yuan Lo, Poojan Oza, and Vishal M. Patel. Adversarially robust one-class novelty detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, page 1-12, 2022. 1, 2", + "[35] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 2", + "[36] Jakub Micorek, Horst Possegger, Dominik Narnhofer, Horst Bischof, and Mateusz Kozinski. MULDE: Multiscale Log-Density Estimation via Denoising Score Matching for Video" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Anomaly Detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18868-18877, 2024. 1", + "[37] Arian Mousakhan, Thomas Brox, and Jawad Tayyub. Anomaly detection with conditioned denoising diffusion models. arXiv preprint arXiv:2305.15956, 2023. 6, 7", + "[38] Yassine Naji, Aleksandr Setkov, Angelique Loesch, Michèle Gouiffes, and Romaric Audigier. Spatio-temporal predictive tasks for abnormal event detection in videos. 2022. 1, 6, 7", + "[39] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2", + "[40] Bharathkumar Ramachandra and Michael Jones. Street scene: A new dataset and evaluation protocol for video anomaly detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2569-2578, 2020. 1, 6, 7", + "[41] Bharathkumar Ramachandra, Michael J. Jones, and Ranga Raju Vatsavai. Learning a distance function with a siamese network to localize anomalies in videos. In 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), 2020. 1, 2, 6, 7", + "[42] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2, 3, 6, 7", + "[43] Tal Reiss and Yedid Hoshen. Attribute-based representations for accurate and interpretable video anomaly detection. 2022. 1, 2, 6", + "[44] Chenrui Shi, Che Sun, Yuwei Wu, and Yunde Jia. Video anomaly detection via sequentially learning multiple pretext tasks. 1, 2", + "[45] Liyan Sun, Chenxin Li, Xinghao Ding, Yue Huang, Zhong Chen, Guisheng Wang, Yizhou Yu, and John Paisley. Few-shot medical image segmentation using a global correlation network with discriminative embedding. Computers in biology and medicine, 140:105067, 2022. 2", + "[46] Yu Tian, Guansong Pang, Yuanhong Chen, Rajvinder Singh, Johan W Verjans, and Gustavo Carneiro. Weakly-supervised video anomaly detection with robust temporal feature magnitude learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4975-4986, 2021. 2", + "[47] Shu Wang and Zhenjiang Miao. Anomaly detection in crowd scene. In IEEE 10th International Conference on Signal Processing Proceedings, pages 1220-1223. IEEE, 2010. 5", + "[48] Yingying Wang, Yunlong Lin, Ge Meng, Zhenqi Fu, Yuhang Dong, Linyu Fan, Hedeng Yu, Xinghao Ding, and Yue Huang. Learning high-frequency feature enhancement and alignment for pan-sharpening. In Proceedings of the 31st ACM International Conference on Multimedia, pages 358-367, 2023. 1", + "[49] Yingying Wang, Yunlong Lin, Xuanhua He, Hui Zheng, Keyu Yan, Linyu Fan, Yue Huang, and Xinghao Ding. Learning diffusion high-quality priors for pan-sharpening: A two-stage approach with time-aware adapter fine-tuning. IEEE Transactions on Geoscience and Remote Sensing, 2025. 1" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8698", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Jhih-Ciang Wu, He-Yen Hsieh, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng-Luh Liu. Self-supervised sparse representation for video anomaly detection. In European Conference on Computer Vision, pages 729-745. Springer, 2022. 2", + "[51] Peng Wu, Xuerong Zhou, Guansong Pang, Lingru Zhou, Qingsen Yan, Peng Wang, and Yanning Zhang. Vadclip: Adapting vision-language models for weakly supervised video anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6074-6082, 2024. 2", + "[52] Fengyang Xiao, Sujie Hu, Yuqi Shen, Chengyu Fang, Jina Huang, Chunming He, Longxiang Tang, Ziyun Yang, and Xiu Li. A survey of camouflaged object detection and beyond. CAIAIR, 2024. 2", + "[53] Zhaohu Xing, Lequan Yu, Liang Wan, Tong Han, and Lei Zhu. Nestedformer: Nested modality-aware transformer for brain tumor segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 140-150. Springer, 2022. 2", + "[54] Zhaohu Xing, Liang Wan, Huazhu Fu, Guang Yang, and Lei Zhu. Diff-unet: A diffusion embedded network for volumetric segmentation. arXiv preprint arXiv:2303.10326, 2023.2", + "[55] Zhaohu Xing, Sicheng Yang, Sixiang Chen, Tian Ye, Yijun Yang, Jing Qin, and Lei Zhu. Cross-conditioned diffusion model for medical image to image translation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 201-211. Springer, 2024. 2", + "[56] Zhaohu Xing, Tian Ye, Yijun Yang, Guang Liu, and Lei Zhu. Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 578-588. Springer, 2024.", + "[57] Yuchen Yang, Kwonjoon Lee, Behzad Dariush, Yinzhi Cao, and Shao-Yuan Lo. Follow the rules: Reasoning for video anomaly detection with large language models. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 1, 2", + "[58] Guang Yu, Siqi Wang, Zhiping Cai, En Zhu, Chuanfu Xu, Jianping Yin, and Marius Kloft. Cloze test helps: Effective video anomaly detection via learning to complete video events. In Proceedings of the 28th ACM International Conference on Multimedia, 2020. 2", + "[59] Luca Zanella, Willi Menapace, Massimiliano Mancini, Yiming Wang, and Elisa Ricci. Harnessing large language models for training-free video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18527-18536, 2024. 2", + "[60] Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8330-8339, 2021. 6, 7", + "[61] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 2", + "[62] Yunlong Zhang, Chenxin Li, Xin Lin, Liyan Sun, Yihong Zhuang, Yue Huang, Xinghao Ding, Xiaqing Liu, and" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yizhou Yu. Generator versus segmentor: Pseudo-healthy synthesis. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part VI 24, pages 150–160. Springer International Publishing, 2021. 2", + "[63] Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. Anomalyclip: Object-agnostic prompt learning for zero-shot anomaly detection. arXiv preprint arXiv:2310.18961, 2023. 6, 7", + "[64] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36, 2024. 3" + ], + "bbox": [ + 516, + 92, + 903, + 287 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8699", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Track Any Anomalous Object_A Granular Video Anomaly Detection Pipeline/1ddd1714-9e44-4427-b96d-74620c052e99_model.json b/2025/Track Any Anomalous Object_A Granular Video Anomaly Detection Pipeline/1ddd1714-9e44-4427-b96d-74620c052e99_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1610caac56d565e26c4deb8181abab19ce1055fa --- /dev/null +++ b/2025/Track Any Anomalous Object_A Granular Video Anomaly Detection Pipeline/1ddd1714-9e44-4427-b96d-74620c052e99_model.json @@ -0,0 +1,2400 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.106, + 0.131, + 0.895, + 0.154 + ], + "angle": 0, + "content": "Track Any Anomalous Object: A Granular Video Anomaly Detection Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.188, + 0.912, + 0.226 + ], + "angle": 0, + "content": "Yuzhi Huang\\(^{1*}\\), Chenxin Li\\(^{2*†}\\), Haitao Zhang\\(^{1}\\), Zixu Lin\\(^{1}\\), Yunlong Lin\\(^{1}\\), Hengyu Liu\\(^{2}\\), Wuyang Li\\(^{2}\\), Xinyu Liu\\(^{2}\\), Jiechao Gao\\(^{3}\\), Yue Huang\\(^{1†}\\), Xinghao Ding\\(^{1}\\) and Yixuan Yuan\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.226, + 0.812, + 0.28 + ], + "angle": 0, + "content": "1 Key Laboratory of Multimedia Trusted Perception and Efficien Computing, Ministry of Education of China, Xiamen University School of Informatics, Xiamen University" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.279, + 0.79, + 0.298 + ], + "angle": 0, + "content": "2 The Chinese University of Hong Kong 3 The University of Virginia" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.347, + 0.327, + 0.363 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.379, + 0.486, + 0.742 + ], + "angle": 0, + "content": "Video anomaly detection (VAD) is crucial in scenarios such as surveillance and autonomous driving, where timely detection of unexpected activities is essential. Albeit existing methods have primarily focused on detecting anomalous objects in videos—either by identifying anomalous frames or objects—they often neglect finer-grained analysis, such as anomalous pixels, which limits their ability to capture a broader range of anomalies. To address this challenge, we propose an innovative VAD framework called Track Any Anomalous Object (TAO), which introduces a Granular Video Anomaly Detection Framework that, for the first time, integrates the detection of multiple fine-grained anomalous objects into a unified framework. Unlike methods that assign anomaly scores to every pixel at each moment, our approach transforms the problem into pixel-level tracking of anomalous objects. By linking anomaly scores to subsequent tasks such as image segmentation and video tracking, our method eliminates the need for threshold selection and achieves more precise anomaly localization, even in long and challenging video sequences. Experiments on extensive datasets demonstrate that TAO achieves state-of-the-art performance, setting a new progress for VAD by providing a practical, granular, and holistic solution. For more information, visit the project page at: https://tao-25.github.io/" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.75, + 0.222, + 0.765 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.775, + 0.484, + 0.867 + ], + "angle": 0, + "content": "Video Anomaly Detection (VAD) involves identifying unusual or unexpected activities in surveillance videos and has significant applications in areas such as security monitoring (e.g., detecting violent behavior) and autonomous driving (e.g., recognizing traffic accidents). Current VAD research has developed along two main directions, which" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.333, + 0.903, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.522, + 0.907, + 0.648 + ], + "angle": 0, + "content": "Figure 1. Analyzing the Limitations of Existing VAD Models. Video anomaly detection (VAD) models are predominantly frame-centric or object-centric. Frame-centric methods detect anomalies in frames without localizing them, while object-centric methods identify anomalous objects but lack pixel-level accuracy. Pixel-centric models address these gaps by providing pixel-level localization, delivering fine-grained segmentation and precise delineation of anomalies, particularly for overlapping objects where traditional methods struggle." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.856 + ], + "angle": 0, + "content": "are reflected in both methodologies and evaluation benchmarks. Frame-centric methods [1, 2, 11, 32, 34, 39, 44, 57] and their corresponding frame-level benchmarks [18, 36, 49, 57] focus on analyzing entire frames to detect global anomalies (e.g., fires or smoke), typically evaluated using frame-level AUC metrics. In contrast, object-centric methods [3, 8-10, 19, 38, 41, 43, 48] and their object-level benchmarks [10, 19, 40] leverage pre-trained feature extractors to detect object-specific anomalies (e.g., human falls or vehicle accidents), evaluated using metrics such as TBDC [40] and RBDC [40], which comprehensively measure temporal consistency and spatial localization precision." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Despite these advances, a critical gap remains in real-world applications of VAD where fine-grained localization of anomalous objects is essential. As illustrated in Fig. 1," + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.876, + 0.345, + 0.889 + ], + "angle": 0, + "content": "* Equal contribution † Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8689" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.258 + ], + "angle": 0, + "content": "current approaches suffer from fundamental shortage as frame-centric methods can only identify the presence of anomalies without localizing specific regions, while object-centric methods, though more precise, lack pixel-level accuracy. These limitations become particularly pronounced in complex scenarios with overlapping anomalies, where accurate delineation of object boundaries and shapes is crucial [4, 12, 14, 20, 28, 46, 50]. This observation motivates us to consider a more challenging yet practical scenario: Can we achieve both object-level structural integrity and pixel-level precision in a granular anomaly detection?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.26, + 0.483, + 0.472 + ], + "angle": 0, + "content": "To address this challenge, we propose enhancing existing VAD benchmarks with pixel-level anomaly evaluation. While a straightforward approach would be to employ an anomaly segmentation task for pixel-wise classification [6, 13, 16, 22, 27, 62], this method falls short when dealing with multiple, potentially overlapping anomalous objects. Instead, inspired by instance segmentation techniques [15, 17, 24, 45, 52], we develop a novel evaluation framework that simultaneously considers both object-level and pixel-level anomaly detection accuracy. This framework ensures that detected anomalies accurately correspond to their true shapes, precise positions, and spatial distributions, while effectively minimizing false positives caused by local feature similarities or noise." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.474, + 0.483, + 0.67 + ], + "angle": 0, + "content": "However, implementing precise pixel-level tracking in videos presents significant challenges, particularly in maintaining semantic and temporal consistency across frames. Traditional supervised approaches require extensive labeled data, which is scarce in existing datasets with pixel-level annotations. To overcome these challenges, we leverage SAM2 [42], a large-scale pre-trained segmentation model capable of processing both static images and video streams in real-time without requiring additional fine-tuning on specific anomaly datasets. Its ability to precisely delineate object contours and structures makes it particularly effective in handling complex scenarios involving overlapping anomalies or occlusions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Building upon these foundations, we propose TAO, an integrated framework that combines object-centric anomaly detection algorithms with SAM2 to create a simple yet effective fine-grained video anomaly detection system. For videos containing anomalous frames, our object-centric detection model generates bounding boxes for detected anomalies. After robust filtering, these boxes serve as prompting boxes for SAM2, which then generates segmentation masks for the anomalous objects. Evaluated on our proposed benchmark, our method achieves state-of-the-art performance on two widely-used datasets, demonstrating its effectiveness in bridging the gap between object-level detection and pixel-precise segmentation. The contributions of our paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- We present a new testing standard that unifies pixel-level" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "and object-level assessment in video anomaly detection, addressing the limitations of existing metrics in complex scenarios with multiple or overlapping anomalies." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.137, + 0.905, + 0.212 + ], + "angle": 0, + "content": "- We introduce TAO, a streamlined framework that integrates object-centric anomaly detection with the segmentation capabilities from vision foundation model, enabling precise pixel-level tracking of anomalous objects without requiring additional fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.212, + 0.905, + 0.273 + ], + "angle": 0, + "content": "- Through extensive experiments on UCSD Ped2 and ShanghaiTech Campus datasets, we demonstrate that our approach achieves state-of-the-art performance under both traditional metrics and our proposed benchmark." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.137, + 0.905, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.287, + 0.655, + 0.303 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.312, + 0.905, + 0.749 + ], + "angle": 0, + "content": "Video Anomaly Detection. Video anomaly detection (VAD) presents significant challenges due to the rarity of anomaly data and the wide variety of abnormal events, which hinder the generalization capabilities of existing models across diverse scenarios. Traditional VAD methods can be broadly classified into frame-centric and object-centric approaches. Frame-centric methods [1, 11, 32, 34, 39, 44, 55-57] analyze entire frames or sequences, often leveraging reconstruction or prediction errors to identify anomalies. These methods are particularly effective for detecting global events, such as fires or smoke. However, they struggle with localized anomalies and are prone to interference from normal regions within a frame, limiting their precision in complex scenarios [11, 32, 34, 39, 53]. Object-centric methods [7-10, 19, 41, 43, 54, 58] focus on specific objects within frames by employing pre-trained object detection models to extract bounding boxes and assess their abnormality. This targeted approach is better suited for detecting anomalies related to individuals or objects, such as human falls or vehicle accidents. By concentrating on specific regions and reducing redundant information, object-centric methods achieve improved accuracy and robustness, especially in complex, multi-object environments. Building on these foundations, our approach is the first to integrate large-scale pre-trained models for pixel-level fine-grained detection in VAD. This advancement enables precise anomaly localization, bridging the gap between frame-level analysis and object-level detection, and significantly enhancing overall detection performance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.75, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Vision Foundation Models. In recent years, large-scale pre-trained language models (LLMs) and vision-language models (VLMs) have shown remarkable potential in advancing video anomaly detection. Models such as BLIP-2 [26], LLaVA [29], LAVAD [59], VadCLIP [51], VideoChatGPT [35], and Video-LLaMA [61] have significantly improved the understanding of complex visual tasks by integrating visual and linguistic information. Concurrently, prompting techniques have gained traction in large-scale vision models, enabling substantial progress in tasks such" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8690" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.096, + 0.565, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.096, + 0.902, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.516, + 0.908, + 0.573 + ], + "angle": 0, + "content": "Figure 2. Pipeline of our proposed TAO. We first generate bounding boxes to identify objects in each frame. Next, we score these boxes using object-centric video anomaly detection algorithms to extract potential anomalous boxes. To ensure robustness, we apply filtering to eliminate redundant boxes. Finally, the filtered boxes and original frames are input into a prompt-based segmentation model to produce pixel-level anomaly segmentation masks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.486, + 0.901 + ], + "angle": 0, + "content": "as image segmentation. By employing semantic prompts (e.g., free-form text) and spatial prompts (e.g., points or bounding boxes), these models achieve highly accurate segmentation guided by input cues. Segmentation Foundation Models (SFMs), such as the Segment Anything Model (SAM) [21] and SEEM [64], have demonstrated exceptional performance, particularly in zero-shot generalization and multi-modal prompting, allowing for cross-task segmentation. Building on these advancements, the Segment Anything Model 2 (SAM2) [42] further improves segmentation, particularly for video-based applications. SAM2 integrates prompting techniques to not only segment static objects with precision but also capture dynamic, complex objects in video contexts. This capability enables SAM2 to deliver finer segmentation and enhanced localization accuracy in video anomaly detection, addressing the challenges posed by real-world variability and complexity. This work leverages SAM2's advanced video segmentation capabilities to tackle the challenge of fine-grained video anomaly detection, achieving precise and robust anomaly identification in" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.6, + 0.644, + 0.614 + ], + "angle": 0, + "content": "complex scenarios." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.632, + 0.605, + 0.646 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.907, + 0.793 + ], + "angle": 0, + "content": "Preliminary. In the proposed video anomaly detection (VAD) benchmark, a video clip is represented as a sequence \\(\\{f_1,f_2,\\ldots ,f_{N_f}\\} \\subseteq \\mathcal{C}\\), where \\(\\mathcal{C}\\) is the set of all possible video clips, and \\(N_{f}\\) denotes the total number of frames. Each frame \\(f_{i}\\) is expressed as \\(f_{i} = [p_{i,1},p_{i,2},\\dots ,p_{i,N_{i}}]\\), where \\(N_{i}\\) is the total number of pixels in the frame, and \\(p_{i,j}\\in \\mathcal{P}\\) represents the \\(j\\)-th pixel, with \\(\\mathcal{P}\\) as the set of possible pixel values. Each pixel \\(p_{i,j}\\) is classified as either \"normal\" or \"anomalous\"." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Overview. To achieve precise anomaly localization in videos, we design a streamlined pipeline for pixel-level segmentation in video anomaly detection. Our pipeline comprises four stages: bounding boxes extraction, anomalous boxes extraction, robust filtering, and segmentation inference. The process begins by generating bounding boxes to identify objects of interest. These boxes are then scored us" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8691" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.103, + 0.092, + 0.692, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.092, + 0.9, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.204, + 0.908, + 0.247 + ], + "angle": 0, + "content": "Figure 3. Impact of Redundant Segmentation on Performance. Potential anomalous boxes are used as prompts for SAM2 segmentation (red masks: true anomalies, green masks: redundant results). Initially, SAM2 tracks true anomalies accurately, but tracking errors accumulate, leading to catastrophic forgetting and a collapse in Pixel-F1 performance." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.273, + 0.365, + 0.287 + ], + "angle": 0, + "content": "Algorithm 1 Boxes Robustness Filtering" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.291, + 0.484, + 0.366 + ], + "angle": 0, + "content": "1: Input: A set of anomalous bounding boxes \\(\\mathcal{B}_{\\mathrm{anomaly}} = \\{b_i \\mid s_i > \\tau\\}\\) for each frame \\(f_i \\in \\{f_1, f_2, \\ldots, f_n\\}\\), where \\(s_i\\) is the anomaly score and \\(\\tau\\) is the threshold; parameters \\(k\\) (tracking window size), \\(h\\) (overlap threshold), \\(l\\) (save interval), and \\(m\\) (frame match threshold)" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.366, + 0.483, + 0.395 + ], + "angle": 0, + "content": "2: Output: Filtered bounding boxes \\(\\mathcal{B}_{\\mathrm{filtered}}\\) with their corresponding object labels \\(\\mathcal{L}_j\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.396, + 0.482, + 0.424 + ], + "angle": 0, + "content": "3: Initialize the list \\(\\mathcal{B}_i\\) to store the filtered bounding boxes from frame \\(f_{i}\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.424, + 0.365, + 0.439 + ], + "angle": 0, + "content": "4: for each frame \\( f_{i} \\in \\{f_{1}, \\ldots, f_{n}\\} \\) do" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.439, + 0.44, + 0.452 + ], + "angle": 0, + "content": "5: Step 1: Inherit Existing Anomalous Boxes" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.453, + 0.483, + 0.483 + ], + "angle": 0, + "content": "6: for each bounding box \\( b_{j} \\in \\mathcal{B}_{\\mathrm{anomaly}} \\) and \\( b_{p} \\) from frame \\( f_{i - k} \\) to \\( f_{i - 1} \\) do" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.291, + 0.484, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.483, + 0.423, + 0.516 + ], + "angle": 0, + "content": "\\[\n7: \\quad \\text {i f} \\sum_ {p = i - k} ^ {i - 1} \\mathbb {I} (\\operatorname {I o U} (b _ {j}, b _ {p}) > h) \\geq m \\text {t h e n}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.516, + 0.305, + 0.531 + ], + "angle": 0, + "content": "8: Assign \\(\\mathcal{L}_j = \\mathcal{L}_p\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.531, + 0.324, + 0.546 + ], + "angle": 0, + "content": "9: Remove \\( b_{j} \\) from \\( \\mathcal{B}_{i} \\)" + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.516, + 0.324, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.547, + 0.214, + 0.559 + ], + "angle": 0, + "content": "10: end if" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.562, + 0.198, + 0.574 + ], + "angle": 0, + "content": "11: end for" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.575, + 0.411, + 0.589 + ], + "angle": 0, + "content": "12: Step 2: Assign New Anomalous Boxes" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.589, + 0.483, + 0.62 + ], + "angle": 0, + "content": "13: for each remaining \\( b_{j} \\in \\mathcal{B}_{\\mathrm{anomaly}} \\) and \\( b_{p} \\) from frame \\( f_{i + 1} \\) to \\( f_{i + k} \\) do" + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.547, + 0.483, + 0.62 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.62, + 0.422, + 0.653 + ], + "angle": 0, + "content": "\\[\n\\text {i f} \\sum_ {p = i + 1} ^ {i + k} \\mathbb {I} \\left(\\operatorname {I o U} \\left(b _ {j}, b _ {p}\\right) > h\\right) \\geq m \\text {t h e n} \\tag {14:}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.654, + 0.446, + 0.669 + ], + "angle": 0, + "content": "15: Assign a new object label \\(\\mathcal{L}_j = \\mathcal{F}(b_j)\\)" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.669, + 0.412, + 0.684 + ], + "angle": 0, + "content": "16: Save the tuple \\(\\mathcal{T}_{\\mathrm{box}} = (f_i, b_j, \\mathcal{L}_j)\\)" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.684, + 0.214, + 0.696 + ], + "angle": 0, + "content": "17: end if" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.699, + 0.198, + 0.711 + ], + "angle": 0, + "content": "18: end for" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.712, + 0.39, + 0.727 + ], + "angle": 0, + "content": "19: Step 3: Save Boxes Every \\(l\\) Frames" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.728, + 0.291, + 0.74 + ], + "angle": 0, + "content": "20: if \\(i \\bmod l == 0\\) then" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.742, + 0.481, + 0.771 + ], + "angle": 0, + "content": "21: Save all bounding box tuples \\(\\mathcal{T}_{\\mathrm{box}} = (f_i, b_j, \\mathcal{L}_j)\\) in the current frame" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.62, + 0.481, + 0.771 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.773, + 0.19, + 0.786 + ], + "angle": 0, + "content": "22: end if" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.787, + 0.174, + 0.8 + ], + "angle": 0, + "content": "23: end for" + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.773, + 0.19, + 0.8 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "ing object-centric anomaly detection algorithms to extract potential anomalies. A tracking-based filtering step follows to eliminate redundant boxes and retain only robust ones. Finally, the filtered boxes and original frames are fed into" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.907, + 0.304 + ], + "angle": 0, + "content": "a prompt-based segmentation model to generate pixel-level anomaly masks. The pipeline is illustrated in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.312, + 0.774, + 0.327 + ], + "angle": 0, + "content": "3.1. Anomalous Boxes Extraction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.907, + 0.457 + ], + "angle": 0, + "content": "Bounding Boxes Extraction. To achieve precise pixel-level segmentation in video anomaly detection, pre-trained object detection algorithms are used to generate bounding boxes for objects within a frame, serving as potential prompts for guiding the segmentation model. For a given input frame \\( f \\), the object detection model \\( \\mathcal{D} \\) outputs \\( m \\) bounding boxes \\( b_{1}, b_{2}, \\ldots, b_{m} \\), each paired with a class label \\( y_{1}, y_{2}, \\ldots, y_{m} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.464, + 0.906, + 0.481 + ], + "angle": 0, + "content": "\\[\n\\{(b _ {1}, y _ {1}), (b _ {2}, y _ {2}), \\dots , (b _ {m}, y _ {m}) \\} = \\mathcal {D} (f) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.906, + 0.563 + ], + "angle": 0, + "content": "After the extraction of object bounding boxes in each frame, object-centric VAD algorithms compute an anomaly score \\( s_i \\) for each object within its bounding box \\( b_i \\). This score is derived from features such as pose, speed, and depth, and can be formally represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.671, + 0.575, + 0.905, + 0.591 + ], + "angle": 0, + "content": "\\[\ns _ {i} = \\mathcal {A} \\left(b _ {i}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.603, + 0.906, + 0.694 + ], + "angle": 0, + "content": "where \\(\\mathcal{A}\\) represents the scoring function used in object-centric video anomaly detection algorithms. To isolate anomalous objects, we filter the bounding boxes based on an anomaly score threshold. Specifically, we retain only the bounding boxes where the anomaly score exceeds a predefined threshold \\(\\tau\\), as shown in the following expression:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.705, + 0.905, + 0.722 + ], + "angle": 0, + "content": "\\[\n\\mathcal {B} _ {\\text {a n o m a l y}} = \\left\\{b _ {i} \\mid s _ {i} > \\tau \\right\\} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.906, + 0.779 + ], + "angle": 0, + "content": "where \\(\\tau\\) denotes the anomaly score threshold, \\(s_i\\) is the anomaly score for the \\(i\\)-th object, and \\(b_i\\) is the corresponding bounding box." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.836, + 0.804 + ], + "angle": 0, + "content": "3.2. Robust Filtering of Anomalous Boxes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The anomalous boxes detection from Sec. 3.1 contains numerous false positives due to overlapping characteristics between normal and abnormal boxes. This overlap complicates threshold setting and leads to misclassifications. These redundant boxes can significantly degrade segmentation performance by generating inaccurate results and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8692" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "overwhelming the model's tracking capabilities, potentially causing the loss of anomalous targets in subsequent frames." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.123, + 0.483, + 0.259 + ], + "angle": 0, + "content": "As illustrated in Fig. 3, using potential anomalous boxes as SAM2 prompts at fixed intervals reveals the impact of tracking errors. While red masks indicate true anomalies and green masks show redundant results, the model initially tracks anomalous objects effectively. However, tracking errors accumulate over time, leading to an increasing number of incorrectly tracked objects. This deterioration ultimately results in catastrophic forgetting of true anomalies, reflected in steadily declining Pixel-F1 scores." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.262, + 0.486, + 0.668 + ], + "angle": 0, + "content": "To enhance robustness and address redundancy, we present the Boxes Robustness Filtering algorithm, which exploits the temporal consistency of true anomalous boxes. Unlike redundant boxes that appear sporadically, true anomalous boxes maintain consistent tracking of the same target. The algorithm, detailed in Alg. 1, operates in three key steps. First, the inheritance step computes the similarity between each bounding box \\( b_{j} \\) in the current frame \\( f_{i} \\) and bounding boxes \\( b_{p} \\) from the previous \\( k \\) frames. If the Intersection over Union (IoU) exceeds a threshold \\( h \\) for at least \\( m \\) frames, \\( b_{j} \\) inherits the label \\( \\mathcal{L}_{p} \\) from earlier frames, ensuring consistent tracking of previously identified anomalies. Second, the assignment step addresses bounding boxes that fail to inherit a label by comparing them with bounding boxes in the subsequent \\( k \\) frames. If the IoU exceeds \\( h \\) for at least \\( m \\) frames, a new label \\( \\mathcal{L}_{j} = \\mathcal{F}(b_{j}) \\) is assigned, effectively labeling newly emerging anomalies. Finally, the saving step records bounding boxes and their labels every \\( l \\)-th frame, systematically capturing all identified anomalies. By iteratively applying these steps, the Boxes Robustness Filtering algorithm achieves robust tracking and localization of anomalous objects. Each bounding box and its label are stored as a tuple \\( \\mathcal{T}_{\\mathrm{box}} = (f_i,b_j,\\mathcal{L}_j) \\), ensuring spatial and temporal consistency across video frames. This approach reduces redundancy and strengthens anomaly detection accuracy, providing a reliable basis for subsequent segmentation and analysis." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.681, + 0.411, + 0.697 + ], + "angle": 0, + "content": "3.3. Inference of the Segmentation Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Using the tuple \\(\\mathcal{T}_{\\mathrm{box}} = (f_i, b_j, \\mathcal{L}_j)\\) obtained from Sec. 3.2, we extract the center point of each bounding box \\(b_j\\). Specifically, given the coordinates of the box \\(b_j = [x_1, y_1, x_2, y_2]\\), the center point is calculated: \\(\\mathbf{c}_j = \\left(\\frac{x_1 + x_2}{2}, \\frac{y_1 + y_2}{2}\\right)\\). Effectively, these prompts are not saved for every frame but are instead collected at regular intervals, specifically every \\(l\\)-th frame, as outlined in Sec. 3.2. This ensures efficient storage and processing while still capturing the necessary information for accurate segmentation [23, 25]. For each selected frame, both the center point \\(\\mathbf{c}_j\\) and the bounding box \\(b_j\\) are stored as prompts. Rather than processing each frame individually, we aggregate prompts from all saved frames \\(\\{f_1, f_{1+l}, f_{1+2l}, \\ldots, f_{N_f}\\}\\), where \\(l\\) denotes the saving in" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.752, + 0.107 + ], + "angle": 0, + "content": "terval. The aggregated prompts are:" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.119, + 0.907, + 0.136 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} = \\left\\{\\left(\\mathbf {c} _ {j}, b _ {j}, f _ {i}\\right) \\mid i = 1, 1 + l, 1 + 2 l, \\dots , N _ {f} \\right\\} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.148, + 0.905, + 0.193 + ], + "angle": 0, + "content": "where \\(\\mathcal{I}\\) represents the set of prompts that includes the center points, bounding box coordinates \\(b_{j}\\), and original frames \\(f_{i}\\) for the selected frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.194, + 0.907, + 0.313 + ], + "angle": 0, + "content": "By inputting this prompt information \\(\\mathcal{I}\\) saved at regular intervals into SAM2, which serves as the prompt-based segmentation model \\(\\mathcal{M}\\), we leverage its feature propagation capabilities to efficiently generate pixel-level segmentation results for the entire video. SAM2 propagates the segmentation results from the prompted frames to the remaining frames, producing pixel-level segmentation across the entire video:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.327, + 0.907, + 0.346 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} = \\mathcal {M} (\\mathcal {I}) = \\left\\{\\mathcal {S} _ {1}, \\mathcal {S} _ {2}, \\dots , \\mathcal {S} _ {N _ {f}} \\right\\} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.356, + 0.907, + 0.432 + ], + "angle": 0, + "content": "where \\( S_{i} = [p_{1}, p_{2}, \\ldots, p_{M}] \\) denotes the segmentation result for frame \\( f_{i} \\), where each pixel \\( p_{j} \\) is either \"normal\" or \"anomalous\". This approach enables the generation of the complete segmentation set \\( S \\), efficiently utilizing contextual information from the prompts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.446, + 0.646, + 0.463 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.47, + 0.86, + 0.487 + ], + "angle": 0, + "content": "4.1. A New Comprehensive VAD Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.907, + 0.689 + ], + "angle": 0, + "content": "Benchmark Design. We present a novel benchmark framework that integrates both pixel-level and object-level evaluation metrics to advance VAD model assessment. This dual-level approach addresses the limitations of using single evaluation metrics. Pixel-level metrics excel at identifying subtle irregularities through independent pixel evaluation, but struggle to capture broader spatial patterns. Object-level metrics complement this by assessing global characteristics like shape and position, enabling coherent anomaly tracking and recognition. By combining these complementary perspectives, our framework provides a comprehensive evaluation of both fine-grained details and holistic anomaly patterns, particularly valuable for real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.84 + ], + "angle": 0, + "content": "Benchmark Datasets. UCSD Ped2 [47] contains 16 training and 12 test videos (240×360 pixels) captured by a fixed overhead camera. The training set consists of normal pedestrian activities, while the test set includes anomalous events such as bikers, skateboarders, and vehicles. ShanghaiTech Campus [30], one of the largest VAD datasets, comprises 330 training and 107 test videos (480×856 pixels). While the training set features normal scenarios, the test set contains various anomalies including robbery, fighting, and unauthorized cycling in pedestrian zones." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Benchmark Metrics. For pixel-level evaluation, we employ four metrics: Pixel-AUROC measures discrimination between normal and anomalous pixels across thresholds, Pixel-AP assesses detection precision by balancing false" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8693" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.089, + 0.882, + 0.105 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison with anomaly detection models adapted for video anomaly detection on the UCSD Ped2 dataset." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.121, + 0.907, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.402, + 0.884, + 0.417 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison on video anomaly detection with pixel-level and object-level metrics on the UCSD Ped2 dataset." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.428, + 0.887, + 0.581 + ], + "angle": 0, + "content": "
MethodPixel-levelObject-level
Pixel-AUROC ↑Pixel-AP ↑Pixel-AUPRO ↑Pixel-F1 ↑RBDC ↑TBDC ↑
AdaCLIP (Zero-shot) [5]51.021.3233.982.615.810.6
AdaCLIP (Fully fine-tuned) [5]53.064.9750.6611.1912.315.5
AnomalyCLIP (Zero-shot) [63]51.6321.2036.345.927.511.2
AnomalyCLIP (Fully fine-tuned) [63]54.2523.7338.597.4813.121.0
DDAD (Fully trained) [37]55.875.6115.122.6718.0113.29
SimpleNet (Fully trained) [33]52.4920.5144.0510.7151.1827.75
DRAEM (Fully trained) [60]69.5830.6335.7810.8944.2670.64
TAO (Partially fine-tuned)75.1150.7872.9764.1283.693.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.604, + 0.483, + 0.633 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison on video anomaly detection with RBDC and TBDC metrics on the UCSD Ped2 dataset." + }, + { + "type": "table", + "bbox": [ + 0.137, + 0.643, + 0.44, + 0.751 + ], + "angle": 0, + "content": "
MethodRBDC ↑TBDC ↑
OCAD [19]52.772.8
SS-VAD [40]62.580.5
SiVL [41]7489.3
AED-SSMTL [10]72.891.2
TAO (Ours)83.693.2
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "positives and negatives, Pixel-AUPRO quantifies segmentation accuracy through region overlap, and Pixel-F1 provides an integrated precision-recall measure. For object-level assessment, we utilize Region-Based Detection Criterion (RBDC) and Track-Based Detection Criterion (TBDC). RBDC measures spatial accuracy using Intersection over Union (IoU) with threshold \\(\\alpha\\), while TBDC evaluates temporal consistency in tracking anomalous regions across" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.604, + 0.907, + 0.633 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison on video anomaly detection with RBDC and TBDC metrics on the ShanghaiTech Campus dataset." + }, + { + "type": "table", + "bbox": [ + 0.56, + 0.644, + 0.864, + 0.768 + ], + "angle": 0, + "content": "
MethodRBDC ↑TBDC ↑
OCAD [19]20.744.5
BAF-AT [9]41.378.8
AED-SSMTL [10]43.284.1
HF2VAD [31]45.484.5
STPT [38]51.684.6
TAO (Ours)62.185.4
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.792, + 0.907, + 0.823 + ], + "angle": 0, + "content": "frames. Together, these metrics provide comprehensive evaluation of both spatial precision and temporal coherence." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.733, + 0.849 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Our framework builds upon SAM2 [42], utilizing an object-level anomaly detection algorithm [43] to extract anomalous bounding boxes as segmentation prompts. We em" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "8694" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.482, + 0.499 + ], + "angle": 0, + "content": "ploy SAM2 [42] as the prompt-based segmentation model, which integrates these prompts with original frames for precise anomaly segmentation. Our implementation adopts a partial fine-tuning strategy, where only the object-level VAD algorithm is fine-tuned on target datasets, while SAM2 retains its pre-trained weights for inference. In the bounding box threshold filtering stage, we assign anomaly scores based on pose and depth features, with thresholds set to \\(\\tau = 1.5\\) for UCSD Ped2 and \\(\\tau = 1.6\\) for ShanghaiTech Campus. The subsequent robustness filtering stage employs a tracking window of \\(k = 5\\), frame match threshold of \\(m = 3\\), and box overlap threshold of \\(h = 0.2\\), with save intervals of \\(l = 5\\) and \\(l = 15\\) for UCSD Ped2 and ShanghaiTech Campus respectively. For segmentation, we utilize SAM2 with hiera base-plus weights on an NVIDIA RTX 4090 GPU. To ensure robust evaluation, overlapping boxes of the same anomaly are merged, and segmentation probability maps are uniformly binarized. For object-level metrics, we derive bounding boxes from segmentation extremal coordinates. Given anomaly pixels \\(P = (x_{1},y_{1}),(x_{2},y_{2}),\\ldots ,(x_{N},y_{N})\\), we compute the box coordinates as: \\(x_{\\mathrm{min}} = \\min_{(x_i,y_i)\\in P}x_i\\), \\(x_{\\mathrm{max}} = \\max_{(x_i,y_i)\\in P}x_i\\), \\(y_{\\mathrm{min}} = \\min_{(x_i,y_i)\\in P}y_i\\), \\(y_{\\mathrm{max}} = \\max_{(x_i,y_i)\\in P}y_i\\), where the resulting bounding box \\((x_{\\mathrm{min}},y_{\\mathrm{min}},x_{\\mathrm{max}},y_{\\mathrm{max}})\\) is used for computing RBDC and TBDC metrics, providing robust spatial evaluation at the object level." + }, + { + "type": "title", + "bbox": [ + 0.094, + 0.514, + 0.452, + 0.529 + ], + "angle": 0, + "content": "4.3. Comparison to Anomaly Detection Models" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.538, + 0.482, + 0.839 + ], + "angle": 0, + "content": "Given the lack of pixel-level segmentation capabilities in current video anomaly detection models, we benchmark several state-of-the-art image anomaly detection algorithms, including AdaCLIP [5], AnomalyCLIP [63], DRAEM [60], DDAD [37], and SimpleNet [33], to evaluate both pixel-level and object-level metrics on the UCSD Ped2 dataset. AdaCLIP and AnomalyCLIP were selected for their pre-trained vision-language architectures, which closely align with SAM2's design and demonstrate strong zero-shot capabilities. To ensure a fair comparison, we assess both AdaCLIP and AnomalyCLIP under zero-shot and fully fine-tuned settings. In contrast, DRAEM, DDAD, and SimpleNet, as fully trained models, provide robust baselines for pixel-level and object-level anomaly detection, representing diverse methodological approaches. For anomaly detection, video streams are processed frame-by-frame through each algorithm, generating heatmaps that identify potential anomalies. Pixel-level segmentation is achieved by applying a threshold to these heatmaps, allowing precise localization of anomalous regions." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.841, + 0.482, + 0.9 + ], + "angle": 0, + "content": "As shown in Tab.1, the experimental results clearly demonstrate the advantages of our model in both pixel-level and object-level anomaly detection. While most benchmarked models struggle to balance these two aspects, of" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.318 + ], + "angle": 0, + "content": "ten excelling in one while underperforming in the other, our model achieves robust performance across both dimensions. This balance reflects its strong and versatile design, enabling precise pixel-level segmentation and maintaining spatial-temporal consistency at the object level. Moreover, our partial fine-tuning approach effectively leverages pretrained features, minimizing the need for extensive dataset-specific training. In contrast, fully trained models such as DRAEM, DDAD, and SimpleNet, as well as vision-language models like AdaCLIP and AnomalyCLIP, fail to achieve comparable results. Fig. 4 further illustrates the strengths of our method. The segmentation results produced by our model are more complete and detailed, accurately delineating object boundaries while preserving the integrity of anomalous objects." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.323, + 0.904, + 0.352 + ], + "angle": 0, + "content": "4.4. Comparison to Conventional Video Anomaly Detection Models" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.365, + 0.905, + 0.545 + ], + "angle": 0, + "content": "To assess the performance of our proposed method, we conducted comprehensive experiments on two widely used datasets, UCSD Ped2 and ShanghaiTech Campus. Our approach was benchmarked against several state-of-the-art video anomaly detection models, including SS-VAD [40], SiVL [41], AED-SSMTL [10], HF2VAD [31], BAF-AT [9], and STPT [38]. Due to the absence of fine-grained segmentation capabilities in traditional methods, the evaluation was focused on object-centric anomaly detection and limited to object-level metrics, TBDC [40] and RBDC [40]. These metrics provide a robust and meaningful basis for comparing the effectiveness of our model with existing approaches." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.55, + 0.905, + 0.76 + ], + "angle": 0, + "content": "Our method achieves state-of-the-art performance on both the UCSD Ped2 and ShanghaiTech Campus datasets. On UCSD Ped2, it outperforms existing methods in TBDC and RBDC metrics, demonstrating precise tracking and localization of anomalous objects across video frames. On the more challenging ShanghaiTech Campus dataset, characterized by complex backgrounds and subtle anomalies, our method consistently surpasses baselines by delivering balanced performance across all evaluation metrics. Unlike existing approaches, which often excel in isolated metrics but fail to generalize, our model effectively integrates object-centric detection and fine-grained segmentation, enabling robust anomaly localization and tracking under diverse real-world conditions, as shown in Tabs. 2 and 3." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.765, + 0.905, + 0.9 + ], + "angle": 0, + "content": "To conclude, our method establishes a new standard in video anomaly detection by combining object-centric detection with fine-grained segmentation. This integration enables precise localization and robust tracking of anomalous objects, addressing the limitations of traditional methods that struggle with balanced performance across metrics. Our approach demonstrates versatility and effectiveness across diverse datasets, highlighting its potential for real-world applications." + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.947, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8695" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.107, + 0.089, + 0.89, + 0.104 + ], + "angle": 0, + "content": "Table 4. Ablation study on proposed key components, evaluated with pixel-level and object-level metrics on the UCSD Ped2 dataset." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.115, + 0.887, + 0.23 + ], + "angle": 0, + "content": "
Video Track ModeBox Robust Filtering AlgorithmPixel-levelObject-level
Pixel-AUROC ↑Pixel-AP ↑Pixel-AUPRO ↑Pixel-F1 ↑RBDC ↑TBDC ↑
××41.7225.3040.5532.9941.071.6
×✓39.0622.8038.0930.5639.371.7
✓×68.9038.4767.2352.1770.876.9
✓✓75.1150.7872.9764.1283.693.2
" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.261, + 0.462, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.452, + 0.483, + 0.494 + ], + "angle": 0, + "content": "Figure 5. Sensitivity Analysis of Anomaly Threshold \\((\\tau)\\). We evaluate performance under varying threshold on ShanghaiTech Campus dataset." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.514, + 0.475, + 0.662 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.703 + ], + "angle": 0, + "content": "Figure 6. Robust Testing of SAM2 Backbones. We compare the performance across different SAM2 model variants." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.728, + 0.327, + 0.745 + ], + "angle": 0, + "content": "4.5. Further Empirical Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Ablation Study. We conduct ablation experiments on the UCSD Ped2 dataset to evaluate two components: video tracking mode and box robust filtering. Video tracking mode, a SAM2 component, enables temporal analysis, while disabling it reduces the model to SAM, performing frame-by-frame segmentation. Disabling box robust filtering uses raw anomaly-thresholded boxes as prompts. Tab. 4 shows the effects of each component on pixel- and object-level metrics. The baseline model, with both components disabled, performs the worst. Enabling box robust filter" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.256, + 0.905, + 0.346 + ], + "angle": 0, + "content": "ing alone yields minimal improvement, while activating video tracking mode significantly boosts pixel-level accuracy. The best performance is achieved with both components enabled, demonstrating their synergistic effect. These results confirm the necessity of both temporal tracking and robust filtering in our framework." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.349, + 0.907, + 0.56 + ], + "angle": 0, + "content": "Sensitivity Analysis. We conduct a sensitivity analysis to assess how variations in the anomaly threshold \\(\\tau\\) affect our framework's performance. Fig. 5 shows the impact of \\(\\tau\\) on four metrics—Pixel-AUROC, Pixel-AP, TBDC, and RBDC—using the ShanghaiTech Campus dataset. As \\(\\tau\\) ranges from 1.0 to 1.9, the framework performs stably, with optimal results at \\(\\tau = 1.6\\). TBDC is most sensitive, peaking at \\(85\\%\\) at \\(\\tau = 1.6\\), while RBDC stays around \\(62\\%\\). Pixel-level metrics, Pixel-AUROC and Pixel-AP, also show steady trends, reflecting the framework's balance of fine-grained anomaly detection and object consistency. This analysis highlights the robustness and adaptability of our framework across threshold variations, ensuring reliable detection in diverse scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.563, + 0.909, + 0.745 + ], + "angle": 0, + "content": "Robustness Testing. We evaluate our framework's robustness through extensive testing across different SAM2 backbone architectures (Fig. 6). Our analysis covers models from the lightweight Tiny variant to the full-scale Large model, with Base+ consistently yielding the best results. This stability suggests that the framework's effectiveness is not dependent on backbone capacity. While larger models show slight improvements in pixel-level metrics, performance remains strong across all variants. These results validate the framework's reliability and robustness, demonstrating its suitability for deployment in a wide range of scenarios, irrespective of computational constraints." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.767, + 0.634, + 0.783 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.903 + ], + "angle": 0, + "content": "This paper presents TAO, a framework integrating coarse object detection with fine-grained anomaly localization in videos. Leveraging SAM2 and our dual-level evaluation, it enables precise segmentation and robust tracking. Experiments on UCSD Ped2 and ShanghaiTech datasets demonstrate state-of-the-art performance in pixel- and object-level metrics, advancing video anomaly detection." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8696" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.222 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 82172033, Grant U19B2031, Grant 61971369, Grant 52105126, Grant 82272071, and Grant 62271430; and in part by the Dreams Foundation of Jianghuai Advance Technology Center and in part by the Open Fund of the National Key Laboratory of Infrared Detection Technologies." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.233, + 0.188, + 0.249 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.259, + 0.484, + 0.314 + ], + "angle": 0, + "content": "[1] Sunghyun Ahn, Youngwan Jo, Kijung Lee, and Sanghyun Park. Videopatchcore: An effective method to memorize normality for video anomaly detection. arXiv preprint arXiv:2409.16225, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.316, + 0.484, + 0.37 + ], + "angle": 0, + "content": "[2] Jinbin Bai, Zhen Dong, Aosong Feng, Xiao Zhang, Tian Ye, Kaicheng Zhou, and Mike Zheng Shou. Integrating view conditions for image synthesis. arXiv preprint arXiv:2310.16002, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.371, + 0.484, + 0.426 + ], + "angle": 0, + "content": "[3] Jinbin Bai, Wei Chow, Ling Yang, Xiangtai Li, Juncheng Li, Hanwang Zhang, and Shuicheng Yan. Humanedit: A high-quality human-rewarded dataset for instruction-based image editing. arXiv preprint arXiv:2412.04280, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.427, + 0.484, + 0.495 + ], + "angle": 0, + "content": "[4] Jinbin Bai, Tian Ye, Wei Chow, Enxin Song, Qing-Guo Chen, Xiangtai Li, Zhen Dong, Lei Zhu, and Shuicheng Yan. Meissonic: Revitalizing Masked Generative Transformers for Efficient High-Resolution Text-to-Image Synthesis. arXiv preprint arXiv:2410.08261, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.497, + 0.484, + 0.564 + ], + "angle": 0, + "content": "[5] Yunkang Cao, Jiangning Zhang, Luca Frittoli, Yuqi Cheng, Weiming Shen, and Giacomo Boracchi. Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly detection. In European Conference on Computer Vision, pages 55-72. Springer, 2025. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.567, + 0.484, + 0.648 + ], + "angle": 0, + "content": "[6] Zhiyuan Ding, Qi Dong, Haote Xu, Chenxin Li, Xinghao Ding, and Yue Huang. Unsupervised anomaly segmentation for brain lesions using dual semantic-manifold reconstruction. In International Conference on Neural Information Processing, pages 133-144. Springer International Publishing Cham, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.65, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[7] Keval Doshi and Yasin Yilmaz. Any-shot sequential anomaly detection in surveillance videos. 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.678, + 0.484, + 0.704 + ], + "angle": 0, + "content": "[8] Keval Doshi and Yasin Yilmaz. Continual learning for anomaly detection in surveillance videos. 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.707, + 0.484, + 0.789 + ], + "angle": 0, + "content": "[9] Mariana-Iuliana Georgescu, RaduTudor Ionescu, Fahad-Shahbaz Khan, Marius Popescu, and Mubarak Shah. A background-agnostic framework with adversarial training for abnormal event detection in video. IEEE Transactions on Pattern Analysis and Machine Intelligence, IEEE Transactions on Pattern Analysis and Machine Intelligence. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.484, + 0.873 + ], + "angle": 0, + "content": "[10] Mariana-Iuliana Georgescu, Antonio Barbalau, Radu Tudor Ionescu, Fahad Shahbaz Khan, Marius Popescu, and Mubarak Shah. Anomaly detection in video via self-supervised and multi-task learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12742-12752, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[11] Mahmudul Hasan, Jonghyun Choi, Jan Neumann, Amit K. Roy-Chowdhury, and Larry S. Davis. Learning temporal" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.259, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "regularity in video sequences. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.906, + 0.19 + ], + "angle": 0, + "content": "[12] Chunming He, Kai Li, Guoxia Xu, Yulun Zhang, Runze Hu, Zhenhua Guo, and Xiu Li. Degradation-resistant unfolding network for heterogeneous image fusion. In ICCV, pages 12611-12621, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.906, + 0.261 + ], + "angle": 0, + "content": "[13] Chunming He, Kai Li, Yachao Zhang, Guoxia Xu, Longxiang Tang, Yulun Zhang, Zhenhua Guo, and Xiu Li. Weakly-supervised concealed object segmentation with sam-based pseudo labeling and multi-scale feature grouping. NeurIPS, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.906, + 0.319 + ], + "angle": 0, + "content": "[14] Chunming He, Kai Li, Yachao Zhang, Yulun Zhang, Zhenhua Guo, Xiu Li, Martin Danelljan, and Fisher Yu. Strategic preys make acute predators: Enhancing camouflaged object detectors by generating camouflaged objects. ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.906, + 0.375 + ], + "angle": 0, + "content": "[15] Chunming He, Chengyu Fang, Yulun Zhang, Kai Li, Longxiang Tang, Chenyu You, Fengyang Xiao, Zhenhua Guo, and Xiu Li. Reti-diff: Illumination degradation image restoration with retinae-based latent diffusion model. *ICLR*, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.906, + 0.433 + ], + "angle": 0, + "content": "[16] Chunming He, Yuqi Shen, Chengyu Fang, Fengyang Xiao, Longxiang Tang, Yulun Zhang, Wangmeng Zuo, Zhenhua Guo, and Xiu Li. Diffusion models in low-level vision: A survey. TPAMI, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.906, + 0.503 + ], + "angle": 0, + "content": "[17] Chunming He, Rihan Zhang, Fengyang Xiao, Chenyu Fang, Longxiang Tang, Yulun Zhang, Linghe Kong, Deng-Ping Fan, Kai Li, and Sina Farsiu. Run: Reversible unfolding network for concealed object segmentation. arXiv preprint arXiv:2501.18783, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.906, + 0.56 + ], + "angle": 0, + "content": "[18] Ryota Hinami, Tao Mei, and Shin'ichi Satoh. Joint detection and recounting of abnormal events by learning deep generic knowledge. In 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.563, + 0.906, + 0.631 + ], + "angle": 0, + "content": "[19] Radu Tudor Ionescu, Fahad Shahbaz Khan, Mariana-Iuliana Georgescu, and Ling Shao. Object-centric auto-encoders and dummy anomalies for abnormal event detection in video. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.906, + 0.689 + ], + "angle": 0, + "content": "[20] Hamza Karim, Keval Doshi, and Yasin Yilmaz. Real-time weakly supervised video anomaly detection. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 6848-6856, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.69, + 0.906, + 0.759 + ], + "angle": 0, + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.906, + 0.815 + ], + "angle": 0, + "content": "[22] Chenxin Li, Yunlong Zhang, Jiongcheng Li, Yue Huang, and Xinghao Ding. Unsupervised anomaly segmentation using image-semantic cycle translation. arXiv preprint arXiv:2103.09094, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[23] Chenxin Li, Yunlong Zhang, Zhehan Liang, Wenao Ma, Yue Huang, and Xinghao Ding. Consistent posterior distributions under vessel-mixing: a regularization for cross-domain retinal artery/vein classification. In 2021 IEEE International Conference on Image Processing (ICIP), pages 61-65. IEEE, 2021. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8697" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[24] Chenxin Li, Wenao Ma, Liyan Sun, Xinghao Ding, Yue Huang, Guisheng Wang, and Yizhou Yu. Hierarchical deep network with uncertainty-aware semi-supervised learning for vessel segmentation. Neural Computing and Applications, pages 1-14, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[25] Chenxin Li, Xinyu Liu, Wuyang Li, Cheng Wang, Hengyu Liu, Yifan Liu, Zhen Chen, and Yixuan Yuan. U-kan makes strong backbone for medical image segmentation and generation. arXiv preprint arXiv:2406.02918, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.286 + ], + "angle": 0, + "content": "[26] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[27] Huangxing Lin, Yunlong Lin, Jingyuan Xia, Linyu Fan, Feifei Li, Yingying Wang, and Xinghao Ding. Fusion2void: Unsupervised multi-focus image fusion based on image inpainting. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[28] Yunlong Lin, Tian Ye, Sixiang Chen, Zhenqi Fu, Yingying Wang, Wenhao Chai, Zhaohu Xing, Lei Zhu, and Xinghao Ding. Aglldiff: Guiding diffusion models towards unsupervised training-free real-world low-light image enhancement. arXiv preprint arXiv:2407.14900, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.482, + 0.469 + ], + "angle": 0, + "content": "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.525 + ], + "angle": 0, + "content": "[30] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection-a new baseline. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6536-6545, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[31] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13588-13597, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[32] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.678, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[33] Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang. Simplenet: A simple network for image anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20402-20411, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[34] Shao-Yuan Lo, Poojan Oza, and Vishal M. Patel. Adversarially robust one-class novelty detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, page 1-12, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[35] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[36] Jakub Micorek, Horst Possegger, Dominik Narnhofer, Horst Bischof, and Mateusz Kozinski. MULDE: Multiscale Log-Density Estimation via Denoising Score Matching for Video" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Anomaly Detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18868-18877, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[37] Arian Mousakhan, Thomas Brox, and Jawad Tayyub. Anomaly detection with conditioned denoising diffusion models. arXiv preprint arXiv:2305.15956, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[38] Yassine Naji, Aleksandr Setkov, Angelique Loesch, Michèle Gouiffes, and Romaric Audigier. Spatio-temporal predictive tasks for abnormal event detection in videos. 2022. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.218, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[39] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[40] Bharathkumar Ramachandra and Michael Jones. Street scene: A new dataset and evaluation protocol for video anomaly detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2569-2578, 2020. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.343, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[41] Bharathkumar Ramachandra, Michael J. Jones, and Ranga Raju Vatsavai. Learning a distance function with a siamese network to localize anomalies in videos. In 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[42] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.483, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[43] Tal Reiss and Yedid Hoshen. Attribute-based representations for accurate and interpretable video anomaly detection. 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.524, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[44] Chenrui Shi, Che Sun, Yuwei Wu, and Yunde Jia. Video anomaly detection via sequentially learning multiple pretext tasks. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[45] Liyan Sun, Chenxin Li, Xinghao Ding, Yue Huang, Zhong Chen, Guisheng Wang, Yizhou Yu, and John Paisley. Few-shot medical image segmentation using a global correlation network with discriminative embedding. Computers in biology and medicine, 140:105067, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.636, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[46] Yu Tian, Guansong Pang, Yuanhong Chen, Rajvinder Singh, Johan W Verjans, and Gustavo Carneiro. Weakly-supervised video anomaly detection with robust temporal feature magnitude learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4975-4986, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[47] Shu Wang and Zhenjiang Miao. Anomaly detection in crowd scene. In IEEE 10th International Conference on Signal Processing Proceedings, pages 1220-1223. IEEE, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[48] Yingying Wang, Yunlong Lin, Ge Meng, Zhenqi Fu, Yuhang Dong, Linyu Fan, Hedeng Yu, Xinghao Ding, and Yue Huang. Learning high-frequency feature enhancement and alignment for pan-sharpening. In Proceedings of the 31st ACM International Conference on Multimedia, pages 358-367, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Yingying Wang, Yunlong Lin, Xuanhua He, Hui Zheng, Keyu Yan, Linyu Fan, Yue Huang, and Xinghao Ding. Learning diffusion high-quality priors for pan-sharpening: A two-stage approach with time-aware adapter fine-tuning. IEEE Transactions on Geoscience and Remote Sensing, 2025. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8698" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[50] Jhih-Ciang Wu, He-Yen Hsieh, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng-Luh Liu. Self-supervised sparse representation for video anomaly detection. In European Conference on Computer Vision, pages 729-745. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[51] Peng Wu, Xuerong Zhou, Guansong Pang, Lingru Zhou, Qingsen Yan, Peng Wang, and Yanning Zhang. Vadclip: Adapting vision-language models for weakly supervised video anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6074-6082, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[52] Fengyang Xiao, Sujie Hu, Yuqi Shen, Chengyu Fang, Jina Huang, Chunming He, Longxiang Tang, Ziyun Yang, and Xiu Li. A survey of camouflaged object detection and beyond. CAIAIR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[53] Zhaohu Xing, Lequan Yu, Liang Wan, Tong Han, and Lei Zhu. Nestedformer: Nested modality-aware transformer for brain tumor segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 140-150. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.396 + ], + "angle": 0, + "content": "[54] Zhaohu Xing, Liang Wan, Huazhu Fu, Guang Yang, and Lei Zhu. Diff-unet: A diffusion embedded network for volumetric segmentation. arXiv preprint arXiv:2303.10326, 2023.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.481 + ], + "angle": 0, + "content": "[55] Zhaohu Xing, Sicheng Yang, Sixiang Chen, Tian Ye, Yijun Yang, Jing Qin, and Lei Zhu. Cross-conditioned diffusion model for medical image to image translation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 201-211. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[56] Zhaohu Xing, Tian Ye, Yijun Yang, Guang Liu, and Lei Zhu. Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 578-588. Springer, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[57] Yuchen Yang, Kwonjoon Lee, Behzad Dariush, Yinzhi Cao, and Shao-Yuan Lo. Follow the rules: Reasoning for video anomaly detection with large language models. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.482, + 0.691 + ], + "angle": 0, + "content": "[58] Guang Yu, Siqi Wang, Zhiping Cai, En Zhu, Chuanfu Xu, Jianping Yin, and Marius Kloft. Cloze test helps: Effective video anomaly detection via learning to complete video events. In Proceedings of the 28th ACM International Conference on Multimedia, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[59] Luca Zanella, Willi Menapace, Massimiliano Mancini, Yiming Wang, and Elisa Ricci. Harnessing large language models for training-free video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18527-18536, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[60] Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8330-8339, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[61] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[62] Yunlong Zhang, Chenxin Li, Xin Lin, Liyan Sun, Yihong Zhuang, Yue Huang, Xinghao Ding, Xiaqing Liu, and" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.176 + ], + "angle": 0, + "content": "Yizhou Yu. Generator versus segmentor: Pseudo-healthy synthesis. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part VI 24, pages 150–160. Springer International Publishing, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[63] Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. Anomalyclip: Object-agnostic prompt learning for zero-shot anomaly detection. arXiv preprint arXiv:2310.18961, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[64] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8699" + } + ] +] \ No newline at end of file diff --git a/2025/Track Any Anomalous Object_A Granular Video Anomaly Detection Pipeline/1ddd1714-9e44-4427-b96d-74620c052e99_origin.pdf b/2025/Track Any Anomalous Object_A Granular Video Anomaly Detection Pipeline/1ddd1714-9e44-4427-b96d-74620c052e99_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b4980a7985f4a95d720c607754892b208ce7e272 --- /dev/null +++ b/2025/Track Any Anomalous Object_A Granular Video Anomaly Detection Pipeline/1ddd1714-9e44-4427-b96d-74620c052e99_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9475527def889719924cb34fd80b7485b54d4a987cdb88e0f42ee2d2ffb5b5e4 +size 2454621 diff --git a/2025/Track Any Anomalous Object_A Granular Video Anomaly Detection Pipeline/full.md b/2025/Track Any Anomalous Object_A Granular Video Anomaly Detection Pipeline/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6c52fe493087ecaf8a202241a8d92160a39e3f42 --- /dev/null +++ b/2025/Track Any Anomalous Object_A Granular Video Anomaly Detection Pipeline/full.md @@ -0,0 +1,309 @@ +# Track Any Anomalous Object: A Granular Video Anomaly Detection Pipeline + +Yuzhi Huang $^{1*}$ , Chenxin Li $^{2*†}$ , Haitao Zhang $^{1}$ , Zixu Lin $^{1}$ , Yunlong Lin $^{1}$ , Hengyu Liu $^{2}$ , Wuyang Li $^{2}$ , Xinyu Liu $^{2}$ , Jiechao Gao $^{3}$ , Yue Huang $^{1†}$ , Xinghao Ding $^{1}$ and Yixuan Yuan $^{2}$ + +1 Key Laboratory of Multimedia Trusted Perception and Efficien Computing, Ministry of Education of China, Xiamen University School of Informatics, Xiamen University + +2 The Chinese University of Hong Kong 3 The University of Virginia + +# Abstract + +Video anomaly detection (VAD) is crucial in scenarios such as surveillance and autonomous driving, where timely detection of unexpected activities is essential. Albeit existing methods have primarily focused on detecting anomalous objects in videos—either by identifying anomalous frames or objects—they often neglect finer-grained analysis, such as anomalous pixels, which limits their ability to capture a broader range of anomalies. To address this challenge, we propose an innovative VAD framework called Track Any Anomalous Object (TAO), which introduces a Granular Video Anomaly Detection Framework that, for the first time, integrates the detection of multiple fine-grained anomalous objects into a unified framework. Unlike methods that assign anomaly scores to every pixel at each moment, our approach transforms the problem into pixel-level tracking of anomalous objects. By linking anomaly scores to subsequent tasks such as image segmentation and video tracking, our method eliminates the need for threshold selection and achieves more precise anomaly localization, even in long and challenging video sequences. Experiments on extensive datasets demonstrate that TAO achieves state-of-the-art performance, setting a new progress for VAD by providing a practical, granular, and holistic solution. For more information, visit the project page at: https://tao-25.github.io/ + +# 1. Introduction + +Video Anomaly Detection (VAD) involves identifying unusual or unexpected activities in surveillance videos and has significant applications in areas such as security monitoring (e.g., detecting violent behavior) and autonomous driving (e.g., recognizing traffic accidents). Current VAD research has developed along two main directions, which + +![](images/e025d956f5427181d2ecc637d75e15e453a7ca7bd5a48ab72235bf197258fcec.jpg) +Figure 1. Analyzing the Limitations of Existing VAD Models. Video anomaly detection (VAD) models are predominantly frame-centric or object-centric. Frame-centric methods detect anomalies in frames without localizing them, while object-centric methods identify anomalous objects but lack pixel-level accuracy. Pixel-centric models address these gaps by providing pixel-level localization, delivering fine-grained segmentation and precise delineation of anomalies, particularly for overlapping objects where traditional methods struggle. + +are reflected in both methodologies and evaluation benchmarks. Frame-centric methods [1, 2, 11, 32, 34, 39, 44, 57] and their corresponding frame-level benchmarks [18, 36, 49, 57] focus on analyzing entire frames to detect global anomalies (e.g., fires or smoke), typically evaluated using frame-level AUC metrics. In contrast, object-centric methods [3, 8-10, 19, 38, 41, 43, 48] and their object-level benchmarks [10, 19, 40] leverage pre-trained feature extractors to detect object-specific anomalies (e.g., human falls or vehicle accidents), evaluated using metrics such as TBDC [40] and RBDC [40], which comprehensively measure temporal consistency and spatial localization precision. + +Despite these advances, a critical gap remains in real-world applications of VAD where fine-grained localization of anomalous objects is essential. As illustrated in Fig. 1, + +current approaches suffer from fundamental shortage as frame-centric methods can only identify the presence of anomalies without localizing specific regions, while object-centric methods, though more precise, lack pixel-level accuracy. These limitations become particularly pronounced in complex scenarios with overlapping anomalies, where accurate delineation of object boundaries and shapes is crucial [4, 12, 14, 20, 28, 46, 50]. This observation motivates us to consider a more challenging yet practical scenario: Can we achieve both object-level structural integrity and pixel-level precision in a granular anomaly detection? + +To address this challenge, we propose enhancing existing VAD benchmarks with pixel-level anomaly evaluation. While a straightforward approach would be to employ an anomaly segmentation task for pixel-wise classification [6, 13, 16, 22, 27, 62], this method falls short when dealing with multiple, potentially overlapping anomalous objects. Instead, inspired by instance segmentation techniques [15, 17, 24, 45, 52], we develop a novel evaluation framework that simultaneously considers both object-level and pixel-level anomaly detection accuracy. This framework ensures that detected anomalies accurately correspond to their true shapes, precise positions, and spatial distributions, while effectively minimizing false positives caused by local feature similarities or noise. + +However, implementing precise pixel-level tracking in videos presents significant challenges, particularly in maintaining semantic and temporal consistency across frames. Traditional supervised approaches require extensive labeled data, which is scarce in existing datasets with pixel-level annotations. To overcome these challenges, we leverage SAM2 [42], a large-scale pre-trained segmentation model capable of processing both static images and video streams in real-time without requiring additional fine-tuning on specific anomaly datasets. Its ability to precisely delineate object contours and structures makes it particularly effective in handling complex scenarios involving overlapping anomalies or occlusions. + +Building upon these foundations, we propose TAO, an integrated framework that combines object-centric anomaly detection algorithms with SAM2 to create a simple yet effective fine-grained video anomaly detection system. For videos containing anomalous frames, our object-centric detection model generates bounding boxes for detected anomalies. After robust filtering, these boxes serve as prompting boxes for SAM2, which then generates segmentation masks for the anomalous objects. Evaluated on our proposed benchmark, our method achieves state-of-the-art performance on two widely-used datasets, demonstrating its effectiveness in bridging the gap between object-level detection and pixel-precise segmentation. The contributions of our paper are as follows: + +- We present a new testing standard that unifies pixel-level + +and object-level assessment in video anomaly detection, addressing the limitations of existing metrics in complex scenarios with multiple or overlapping anomalies. + +- We introduce TAO, a streamlined framework that integrates object-centric anomaly detection with the segmentation capabilities from vision foundation model, enabling precise pixel-level tracking of anomalous objects without requiring additional fine-tuning. +- Through extensive experiments on UCSD Ped2 and ShanghaiTech Campus datasets, we demonstrate that our approach achieves state-of-the-art performance under both traditional metrics and our proposed benchmark. + +# 2. Related Work + +Video Anomaly Detection. Video anomaly detection (VAD) presents significant challenges due to the rarity of anomaly data and the wide variety of abnormal events, which hinder the generalization capabilities of existing models across diverse scenarios. Traditional VAD methods can be broadly classified into frame-centric and object-centric approaches. Frame-centric methods [1, 11, 32, 34, 39, 44, 55-57] analyze entire frames or sequences, often leveraging reconstruction or prediction errors to identify anomalies. These methods are particularly effective for detecting global events, such as fires or smoke. However, they struggle with localized anomalies and are prone to interference from normal regions within a frame, limiting their precision in complex scenarios [11, 32, 34, 39, 53]. Object-centric methods [7-10, 19, 41, 43, 54, 58] focus on specific objects within frames by employing pre-trained object detection models to extract bounding boxes and assess their abnormality. This targeted approach is better suited for detecting anomalies related to individuals or objects, such as human falls or vehicle accidents. By concentrating on specific regions and reducing redundant information, object-centric methods achieve improved accuracy and robustness, especially in complex, multi-object environments. Building on these foundations, our approach is the first to integrate large-scale pre-trained models for pixel-level fine-grained detection in VAD. This advancement enables precise anomaly localization, bridging the gap between frame-level analysis and object-level detection, and significantly enhancing overall detection performance. + +Vision Foundation Models. In recent years, large-scale pre-trained language models (LLMs) and vision-language models (VLMs) have shown remarkable potential in advancing video anomaly detection. Models such as BLIP-2 [26], LLaVA [29], LAVAD [59], VadCLIP [51], VideoChatGPT [35], and Video-LLaMA [61] have significantly improved the understanding of complex visual tasks by integrating visual and linguistic information. Concurrently, prompting techniques have gained traction in large-scale vision models, enabling substantial progress in tasks such + +![](images/d519cad781a79038c645689825d2eb24ba22e10e45c532beeec93b559cf60ddd.jpg) +Figure 2. Pipeline of our proposed TAO. We first generate bounding boxes to identify objects in each frame. Next, we score these boxes using object-centric video anomaly detection algorithms to extract potential anomalous boxes. To ensure robustness, we apply filtering to eliminate redundant boxes. Finally, the filtered boxes and original frames are input into a prompt-based segmentation model to produce pixel-level anomaly segmentation masks. + +![](images/d07cfa6c08efc611527ce94c625577b26a2954cc59c421483f06cfa248da8c35.jpg) + +as image segmentation. By employing semantic prompts (e.g., free-form text) and spatial prompts (e.g., points or bounding boxes), these models achieve highly accurate segmentation guided by input cues. Segmentation Foundation Models (SFMs), such as the Segment Anything Model (SAM) [21] and SEEM [64], have demonstrated exceptional performance, particularly in zero-shot generalization and multi-modal prompting, allowing for cross-task segmentation. Building on these advancements, the Segment Anything Model 2 (SAM2) [42] further improves segmentation, particularly for video-based applications. SAM2 integrates prompting techniques to not only segment static objects with precision but also capture dynamic, complex objects in video contexts. This capability enables SAM2 to deliver finer segmentation and enhanced localization accuracy in video anomaly detection, addressing the challenges posed by real-world variability and complexity. This work leverages SAM2's advanced video segmentation capabilities to tackle the challenge of fine-grained video anomaly detection, achieving precise and robust anomaly identification in + +complex scenarios. + +# 3. Method + +Preliminary. In the proposed video anomaly detection (VAD) benchmark, a video clip is represented as a sequence $\{f_1,f_2,\ldots ,f_{N_f}\} \subseteq \mathcal{C}$ , where $\mathcal{C}$ is the set of all possible video clips, and $N_{f}$ denotes the total number of frames. Each frame $f_{i}$ is expressed as $f_{i} = [p_{i,1},p_{i,2},\dots ,p_{i,N_{i}}]$ , where $N_{i}$ is the total number of pixels in the frame, and $p_{i,j}\in \mathcal{P}$ represents the $j$ -th pixel, with $\mathcal{P}$ as the set of possible pixel values. Each pixel $p_{i,j}$ is classified as either "normal" or "anomalous". + +Overview. To achieve precise anomaly localization in videos, we design a streamlined pipeline for pixel-level segmentation in video anomaly detection. Our pipeline comprises four stages: bounding boxes extraction, anomalous boxes extraction, robust filtering, and segmentation inference. The process begins by generating bounding boxes to identify objects of interest. These boxes are then scored us + +![](images/6adbbf4d76a6cb8125318ce77626324346965b73396d3c8cda8c9bec6039ebd1.jpg) +Figure 3. Impact of Redundant Segmentation on Performance. Potential anomalous boxes are used as prompts for SAM2 segmentation (red masks: true anomalies, green masks: redundant results). Initially, SAM2 tracks true anomalies accurately, but tracking errors accumulate, leading to catastrophic forgetting and a collapse in Pixel-F1 performance. + +![](images/daaff1e3489f33ad5e52695132fea706efa7667581fbb2a3cb06768da8caff8b.jpg) + +# Algorithm 1 Boxes Robustness Filtering + +1: Input: A set of anomalous bounding boxes $\mathcal{B}_{\mathrm{anomaly}} = \{b_i \mid s_i > \tau\}$ for each frame $f_i \in \{f_1, f_2, \ldots, f_n\}$ , where $s_i$ is the anomaly score and $\tau$ is the threshold; parameters $k$ (tracking window size), $h$ (overlap threshold), $l$ (save interval), and $m$ (frame match threshold) +2: Output: Filtered bounding boxes $\mathcal{B}_{\mathrm{filtered}}$ with their corresponding object labels $\mathcal{L}_j$ +3: Initialize the list $\mathcal{B}_i$ to store the filtered bounding boxes from frame $f_{i}$ +4: for each frame $f_{i} \in \{f_{1}, \ldots, f_{n}\}$ do +5: Step 1: Inherit Existing Anomalous Boxes +6: for each bounding box $b_{j} \in \mathcal{B}_{\mathrm{anomaly}}$ and $b_{p}$ from frame $f_{i - k}$ to $f_{i - 1}$ do + +$$ +7: \quad \text {i f} \sum_ {p = i - k} ^ {i - 1} \mathbb {I} (\operatorname {I o U} (b _ {j}, b _ {p}) > h) \geq m \text {t h e n} +$$ + +8: Assign $\mathcal{L}_j = \mathcal{L}_p$ +9: Remove $b_{j}$ from $\mathcal{B}_{i}$ + +10: end if +11: end for +12: Step 2: Assign New Anomalous Boxes +13: for each remaining $b_{j} \in \mathcal{B}_{\mathrm{anomaly}}$ and $b_{p}$ from frame $f_{i + 1}$ to $f_{i + k}$ do + +$$ +\text {i f} \sum_ {p = i + 1} ^ {i + k} \mathbb {I} \left(\operatorname {I o U} \left(b _ {j}, b _ {p}\right) > h\right) \geq m \text {t h e n} \tag {14:} +$$ + +15: Assign a new object label $\mathcal{L}_j = \mathcal{F}(b_j)$ +16: Save the tuple $\mathcal{T}_{\mathrm{box}} = (f_i, b_j, \mathcal{L}_j)$ +17: end if +18: end for +19: Step 3: Save Boxes Every $l$ Frames +20: if $i \bmod l == 0$ then +21: Save all bounding box tuples $\mathcal{T}_{\mathrm{box}} = (f_i, b_j, \mathcal{L}_j)$ in the current frame + +22: end if +23: end for + +ing object-centric anomaly detection algorithms to extract potential anomalies. A tracking-based filtering step follows to eliminate redundant boxes and retain only robust ones. Finally, the filtered boxes and original frames are fed into + +a prompt-based segmentation model to generate pixel-level anomaly masks. The pipeline is illustrated in Fig. 2. + +# 3.1. Anomalous Boxes Extraction + +Bounding Boxes Extraction. To achieve precise pixel-level segmentation in video anomaly detection, pre-trained object detection algorithms are used to generate bounding boxes for objects within a frame, serving as potential prompts for guiding the segmentation model. For a given input frame $f$ , the object detection model $\mathcal{D}$ outputs $m$ bounding boxes $b_{1}, b_{2}, \ldots, b_{m}$ , each paired with a class label $y_{1}, y_{2}, \ldots, y_{m}$ : + +$$ +\{(b _ {1}, y _ {1}), (b _ {2}, y _ {2}), \dots , (b _ {m}, y _ {m}) \} = \mathcal {D} (f) \tag {1} +$$ + +After the extraction of object bounding boxes in each frame, object-centric VAD algorithms compute an anomaly score $s_i$ for each object within its bounding box $b_i$ . This score is derived from features such as pose, speed, and depth, and can be formally represented as: + +$$ +s _ {i} = \mathcal {A} \left(b _ {i}\right) \tag {2} +$$ + +where $\mathcal{A}$ represents the scoring function used in object-centric video anomaly detection algorithms. To isolate anomalous objects, we filter the bounding boxes based on an anomaly score threshold. Specifically, we retain only the bounding boxes where the anomaly score exceeds a predefined threshold $\tau$ , as shown in the following expression: + +$$ +\mathcal {B} _ {\text {a n o m a l y}} = \left\{b _ {i} \mid s _ {i} > \tau \right\} \tag {3} +$$ + +where $\tau$ denotes the anomaly score threshold, $s_i$ is the anomaly score for the $i$ -th object, and $b_i$ is the corresponding bounding box. + +# 3.2. Robust Filtering of Anomalous Boxes + +The anomalous boxes detection from Sec. 3.1 contains numerous false positives due to overlapping characteristics between normal and abnormal boxes. This overlap complicates threshold setting and leads to misclassifications. These redundant boxes can significantly degrade segmentation performance by generating inaccurate results and + +overwhelming the model's tracking capabilities, potentially causing the loss of anomalous targets in subsequent frames. + +As illustrated in Fig. 3, using potential anomalous boxes as SAM2 prompts at fixed intervals reveals the impact of tracking errors. While red masks indicate true anomalies and green masks show redundant results, the model initially tracks anomalous objects effectively. However, tracking errors accumulate over time, leading to an increasing number of incorrectly tracked objects. This deterioration ultimately results in catastrophic forgetting of true anomalies, reflected in steadily declining Pixel-F1 scores. + +To enhance robustness and address redundancy, we present the Boxes Robustness Filtering algorithm, which exploits the temporal consistency of true anomalous boxes. Unlike redundant boxes that appear sporadically, true anomalous boxes maintain consistent tracking of the same target. The algorithm, detailed in Alg. 1, operates in three key steps. First, the inheritance step computes the similarity between each bounding box $b_{j}$ in the current frame $f_{i}$ and bounding boxes $b_{p}$ from the previous $k$ frames. If the Intersection over Union (IoU) exceeds a threshold $h$ for at least $m$ frames, $b_{j}$ inherits the label $\mathcal{L}_{p}$ from earlier frames, ensuring consistent tracking of previously identified anomalies. Second, the assignment step addresses bounding boxes that fail to inherit a label by comparing them with bounding boxes in the subsequent $k$ frames. If the IoU exceeds $h$ for at least $m$ frames, a new label $\mathcal{L}_{j} = \mathcal{F}(b_{j})$ is assigned, effectively labeling newly emerging anomalies. Finally, the saving step records bounding boxes and their labels every $l$ -th frame, systematically capturing all identified anomalies. By iteratively applying these steps, the Boxes Robustness Filtering algorithm achieves robust tracking and localization of anomalous objects. Each bounding box and its label are stored as a tuple $\mathcal{T}_{\mathrm{box}} = (f_i,b_j,\mathcal{L}_j)$ , ensuring spatial and temporal consistency across video frames. This approach reduces redundancy and strengthens anomaly detection accuracy, providing a reliable basis for subsequent segmentation and analysis. + +# 3.3. Inference of the Segmentation Model + +Using the tuple $\mathcal{T}_{\mathrm{box}} = (f_i, b_j, \mathcal{L}_j)$ obtained from Sec. 3.2, we extract the center point of each bounding box $b_j$ . Specifically, given the coordinates of the box $b_j = [x_1, y_1, x_2, y_2]$ , the center point is calculated: $\mathbf{c}_j = \left(\frac{x_1 + x_2}{2}, \frac{y_1 + y_2}{2}\right)$ . Effectively, these prompts are not saved for every frame but are instead collected at regular intervals, specifically every $l$ -th frame, as outlined in Sec. 3.2. This ensures efficient storage and processing while still capturing the necessary information for accurate segmentation [23, 25]. For each selected frame, both the center point $\mathbf{c}_j$ and the bounding box $b_j$ are stored as prompts. Rather than processing each frame individually, we aggregate prompts from all saved frames $\{f_1, f_{1+l}, f_{1+2l}, \ldots, f_{N_f}\}$ , where $l$ denotes the saving in + +terval. The aggregated prompts are: + +$$ +\mathcal {I} = \left\{\left(\mathbf {c} _ {j}, b _ {j}, f _ {i}\right) \mid i = 1, 1 + l, 1 + 2 l, \dots , N _ {f} \right\} \tag {4} +$$ + +where $\mathcal{I}$ represents the set of prompts that includes the center points, bounding box coordinates $b_{j}$ , and original frames $f_{i}$ for the selected frames. + +By inputting this prompt information $\mathcal{I}$ saved at regular intervals into SAM2, which serves as the prompt-based segmentation model $\mathcal{M}$ , we leverage its feature propagation capabilities to efficiently generate pixel-level segmentation results for the entire video. SAM2 propagates the segmentation results from the prompted frames to the remaining frames, producing pixel-level segmentation across the entire video: + +$$ +\mathcal {S} = \mathcal {M} (\mathcal {I}) = \left\{\mathcal {S} _ {1}, \mathcal {S} _ {2}, \dots , \mathcal {S} _ {N _ {f}} \right\} \tag {5} +$$ + +where $S_{i} = [p_{1}, p_{2}, \ldots, p_{M}]$ denotes the segmentation result for frame $f_{i}$ , where each pixel $p_{j}$ is either "normal" or "anomalous". This approach enables the generation of the complete segmentation set $S$ , efficiently utilizing contextual information from the prompts. + +# 4. Experiments + +# 4.1. A New Comprehensive VAD Benchmark + +Benchmark Design. We present a novel benchmark framework that integrates both pixel-level and object-level evaluation metrics to advance VAD model assessment. This dual-level approach addresses the limitations of using single evaluation metrics. Pixel-level metrics excel at identifying subtle irregularities through independent pixel evaluation, but struggle to capture broader spatial patterns. Object-level metrics complement this by assessing global characteristics like shape and position, enabling coherent anomaly tracking and recognition. By combining these complementary perspectives, our framework provides a comprehensive evaluation of both fine-grained details and holistic anomaly patterns, particularly valuable for real-world applications. + +Benchmark Datasets. UCSD Ped2 [47] contains 16 training and 12 test videos (240×360 pixels) captured by a fixed overhead camera. The training set consists of normal pedestrian activities, while the test set includes anomalous events such as bikers, skateboarders, and vehicles. ShanghaiTech Campus [30], one of the largest VAD datasets, comprises 330 training and 107 test videos (480×856 pixels). While the training set features normal scenarios, the test set contains various anomalies including robbery, fighting, and unauthorized cycling in pedestrian zones. + +Benchmark Metrics. For pixel-level evaluation, we employ four metrics: Pixel-AUROC measures discrimination between normal and anomalous pixels across thresholds, Pixel-AP assesses detection precision by balancing false + +![](images/ca47cc552283084344bf2a84c30cf3fe841af99c1dd7705aa375cd0447b02c4e.jpg) +Figure 4. Qualitative comparison with anomaly detection models adapted for video anomaly detection on the UCSD Ped2 dataset. + +Table 1. Quantitative comparison on video anomaly detection with pixel-level and object-level metrics on the UCSD Ped2 dataset. + +
MethodPixel-levelObject-level
Pixel-AUROC ↑Pixel-AP ↑Pixel-AUPRO ↑Pixel-F1 ↑RBDC ↑TBDC ↑
AdaCLIP (Zero-shot) [5]51.021.3233.982.615.810.6
AdaCLIP (Fully fine-tuned) [5]53.064.9750.6611.1912.315.5
AnomalyCLIP (Zero-shot) [63]51.6321.2036.345.927.511.2
AnomalyCLIP (Fully fine-tuned) [63]54.2523.7338.597.4813.121.0
DDAD (Fully trained) [37]55.875.6115.122.6718.0113.29
SimpleNet (Fully trained) [33]52.4920.5144.0510.7151.1827.75
DRAEM (Fully trained) [60]69.5830.6335.7810.8944.2670.64
TAO (Partially fine-tuned)75.1150.7872.9764.1283.693.2
+ +Table 2. Quantitative comparison on video anomaly detection with RBDC and TBDC metrics on the UCSD Ped2 dataset. + +
MethodRBDC ↑TBDC ↑
OCAD [19]52.772.8
SS-VAD [40]62.580.5
SiVL [41]7489.3
AED-SSMTL [10]72.891.2
TAO (Ours)83.693.2
+ +positives and negatives, Pixel-AUPRO quantifies segmentation accuracy through region overlap, and Pixel-F1 provides an integrated precision-recall measure. For object-level assessment, we utilize Region-Based Detection Criterion (RBDC) and Track-Based Detection Criterion (TBDC). RBDC measures spatial accuracy using Intersection over Union (IoU) with threshold $\alpha$ , while TBDC evaluates temporal consistency in tracking anomalous regions across + +Table 3. Quantitative comparison on video anomaly detection with RBDC and TBDC metrics on the ShanghaiTech Campus dataset. + +
MethodRBDC ↑TBDC ↑
OCAD [19]20.744.5
BAF-AT [9]41.378.8
AED-SSMTL [10]43.284.1
HF2VAD [31]45.484.5
STPT [38]51.684.6
TAO (Ours)62.185.4
+ +frames. Together, these metrics provide comprehensive evaluation of both spatial precision and temporal coherence. + +# 4.2. Implementation Details + +Our framework builds upon SAM2 [42], utilizing an object-level anomaly detection algorithm [43] to extract anomalous bounding boxes as segmentation prompts. We em + +ploy SAM2 [42] as the prompt-based segmentation model, which integrates these prompts with original frames for precise anomaly segmentation. Our implementation adopts a partial fine-tuning strategy, where only the object-level VAD algorithm is fine-tuned on target datasets, while SAM2 retains its pre-trained weights for inference. In the bounding box threshold filtering stage, we assign anomaly scores based on pose and depth features, with thresholds set to $\tau = 1.5$ for UCSD Ped2 and $\tau = 1.6$ for ShanghaiTech Campus. The subsequent robustness filtering stage employs a tracking window of $k = 5$ , frame match threshold of $m = 3$ , and box overlap threshold of $h = 0.2$ , with save intervals of $l = 5$ and $l = 15$ for UCSD Ped2 and ShanghaiTech Campus respectively. For segmentation, we utilize SAM2 with hiera base-plus weights on an NVIDIA RTX 4090 GPU. To ensure robust evaluation, overlapping boxes of the same anomaly are merged, and segmentation probability maps are uniformly binarized. For object-level metrics, we derive bounding boxes from segmentation extremal coordinates. Given anomaly pixels $P = (x_{1},y_{1}),(x_{2},y_{2}),\ldots ,(x_{N},y_{N})$ , we compute the box coordinates as: $x_{\mathrm{min}} = \min_{(x_i,y_i)\in P}x_i$ , $x_{\mathrm{max}} = \max_{(x_i,y_i)\in P}x_i$ , $y_{\mathrm{min}} = \min_{(x_i,y_i)\in P}y_i$ , $y_{\mathrm{max}} = \max_{(x_i,y_i)\in P}y_i$ , where the resulting bounding box $(x_{\mathrm{min}},y_{\mathrm{min}},x_{\mathrm{max}},y_{\mathrm{max}})$ is used for computing RBDC and TBDC metrics, providing robust spatial evaluation at the object level. + +# 4.3. Comparison to Anomaly Detection Models + +Given the lack of pixel-level segmentation capabilities in current video anomaly detection models, we benchmark several state-of-the-art image anomaly detection algorithms, including AdaCLIP [5], AnomalyCLIP [63], DRAEM [60], DDAD [37], and SimpleNet [33], to evaluate both pixel-level and object-level metrics on the UCSD Ped2 dataset. AdaCLIP and AnomalyCLIP were selected for their pre-trained vision-language architectures, which closely align with SAM2's design and demonstrate strong zero-shot capabilities. To ensure a fair comparison, we assess both AdaCLIP and AnomalyCLIP under zero-shot and fully fine-tuned settings. In contrast, DRAEM, DDAD, and SimpleNet, as fully trained models, provide robust baselines for pixel-level and object-level anomaly detection, representing diverse methodological approaches. For anomaly detection, video streams are processed frame-by-frame through each algorithm, generating heatmaps that identify potential anomalies. Pixel-level segmentation is achieved by applying a threshold to these heatmaps, allowing precise localization of anomalous regions. + +As shown in Tab.1, the experimental results clearly demonstrate the advantages of our model in both pixel-level and object-level anomaly detection. While most benchmarked models struggle to balance these two aspects, of + +ten excelling in one while underperforming in the other, our model achieves robust performance across both dimensions. This balance reflects its strong and versatile design, enabling precise pixel-level segmentation and maintaining spatial-temporal consistency at the object level. Moreover, our partial fine-tuning approach effectively leverages pretrained features, minimizing the need for extensive dataset-specific training. In contrast, fully trained models such as DRAEM, DDAD, and SimpleNet, as well as vision-language models like AdaCLIP and AnomalyCLIP, fail to achieve comparable results. Fig. 4 further illustrates the strengths of our method. The segmentation results produced by our model are more complete and detailed, accurately delineating object boundaries while preserving the integrity of anomalous objects. + +# 4.4. Comparison to Conventional Video Anomaly Detection Models + +To assess the performance of our proposed method, we conducted comprehensive experiments on two widely used datasets, UCSD Ped2 and ShanghaiTech Campus. Our approach was benchmarked against several state-of-the-art video anomaly detection models, including SS-VAD [40], SiVL [41], AED-SSMTL [10], HF2VAD [31], BAF-AT [9], and STPT [38]. Due to the absence of fine-grained segmentation capabilities in traditional methods, the evaluation was focused on object-centric anomaly detection and limited to object-level metrics, TBDC [40] and RBDC [40]. These metrics provide a robust and meaningful basis for comparing the effectiveness of our model with existing approaches. + +Our method achieves state-of-the-art performance on both the UCSD Ped2 and ShanghaiTech Campus datasets. On UCSD Ped2, it outperforms existing methods in TBDC and RBDC metrics, demonstrating precise tracking and localization of anomalous objects across video frames. On the more challenging ShanghaiTech Campus dataset, characterized by complex backgrounds and subtle anomalies, our method consistently surpasses baselines by delivering balanced performance across all evaluation metrics. Unlike existing approaches, which often excel in isolated metrics but fail to generalize, our model effectively integrates object-centric detection and fine-grained segmentation, enabling robust anomaly localization and tracking under diverse real-world conditions, as shown in Tabs. 2 and 3. + +To conclude, our method establishes a new standard in video anomaly detection by combining object-centric detection with fine-grained segmentation. This integration enables precise localization and robust tracking of anomalous objects, addressing the limitations of traditional methods that struggle with balanced performance across metrics. Our approach demonstrates versatility and effectiveness across diverse datasets, highlighting its potential for real-world applications. + +Table 4. Ablation study on proposed key components, evaluated with pixel-level and object-level metrics on the UCSD Ped2 dataset. + +
Video Track ModeBox Robust Filtering AlgorithmPixel-levelObject-level
Pixel-AUROC ↑Pixel-AP ↑Pixel-AUPRO ↑Pixel-F1 ↑RBDC ↑TBDC ↑
××41.7225.3040.5532.9941.071.6
×✓39.0622.8038.0930.5639.371.7
✓×68.9038.4767.2352.1770.876.9
✓✓75.1150.7872.9764.1283.693.2
+ +![](images/bb0f3153cbed6e9a20805bb3acac1a71f14e06be1d248bafe9dc1aa92d06c558.jpg) +Figure 5. Sensitivity Analysis of Anomaly Threshold $(\tau)$ . We evaluate performance under varying threshold on ShanghaiTech Campus dataset. + +![](images/9469c4c2b498ef09a28583c1d54f6ed4cf3d9f0c9b5d36e37c227dab2b3dccf7.jpg) +Figure 6. Robust Testing of SAM2 Backbones. We compare the performance across different SAM2 model variants. + +# 4.5. Further Empirical Study + +Ablation Study. We conduct ablation experiments on the UCSD Ped2 dataset to evaluate two components: video tracking mode and box robust filtering. Video tracking mode, a SAM2 component, enables temporal analysis, while disabling it reduces the model to SAM, performing frame-by-frame segmentation. Disabling box robust filtering uses raw anomaly-thresholded boxes as prompts. Tab. 4 shows the effects of each component on pixel- and object-level metrics. The baseline model, with both components disabled, performs the worst. Enabling box robust filter + +ing alone yields minimal improvement, while activating video tracking mode significantly boosts pixel-level accuracy. The best performance is achieved with both components enabled, demonstrating their synergistic effect. These results confirm the necessity of both temporal tracking and robust filtering in our framework. + +Sensitivity Analysis. We conduct a sensitivity analysis to assess how variations in the anomaly threshold $\tau$ affect our framework's performance. Fig. 5 shows the impact of $\tau$ on four metrics—Pixel-AUROC, Pixel-AP, TBDC, and RBDC—using the ShanghaiTech Campus dataset. As $\tau$ ranges from 1.0 to 1.9, the framework performs stably, with optimal results at $\tau = 1.6$ . TBDC is most sensitive, peaking at $85\%$ at $\tau = 1.6$ , while RBDC stays around $62\%$ . Pixel-level metrics, Pixel-AUROC and Pixel-AP, also show steady trends, reflecting the framework's balance of fine-grained anomaly detection and object consistency. This analysis highlights the robustness and adaptability of our framework across threshold variations, ensuring reliable detection in diverse scenarios. + +Robustness Testing. We evaluate our framework's robustness through extensive testing across different SAM2 backbone architectures (Fig. 6). Our analysis covers models from the lightweight Tiny variant to the full-scale Large model, with Base+ consistently yielding the best results. This stability suggests that the framework's effectiveness is not dependent on backbone capacity. While larger models show slight improvements in pixel-level metrics, performance remains strong across all variants. These results validate the framework's reliability and robustness, demonstrating its suitability for deployment in a wide range of scenarios, irrespective of computational constraints. + +# 5. Conclusion + +This paper presents TAO, a framework integrating coarse object detection with fine-grained anomaly localization in videos. Leveraging SAM2 and our dual-level evaluation, it enables precise segmentation and robust tracking. Experiments on UCSD Ped2 and ShanghaiTech datasets demonstrate state-of-the-art performance in pixel- and object-level metrics, advancing video anomaly detection. + +# 6. Acknowledgments + +This work was supported in part by the National Natural Science Foundation of China under Grant 82172033, Grant U19B2031, Grant 61971369, Grant 52105126, Grant 82272071, and Grant 62271430; and in part by the Dreams Foundation of Jianghuai Advance Technology Center and in part by the Open Fund of the National Key Laboratory of Infrared Detection Technologies. + +# References + +[1] Sunghyun Ahn, Youngwan Jo, Kijung Lee, and Sanghyun Park. Videopatchcore: An effective method to memorize normality for video anomaly detection. arXiv preprint arXiv:2409.16225, 2024. 1, 2 +[2] Jinbin Bai, Zhen Dong, Aosong Feng, Xiao Zhang, Tian Ye, Kaicheng Zhou, and Mike Zheng Shou. Integrating view conditions for image synthesis. arXiv preprint arXiv:2310.16002, 2023. 1 +[3] Jinbin Bai, Wei Chow, Ling Yang, Xiangtai Li, Juncheng Li, Hanwang Zhang, and Shuicheng Yan. Humanedit: A high-quality human-rewarded dataset for instruction-based image editing. arXiv preprint arXiv:2412.04280, 2024. 1 +[4] Jinbin Bai, Tian Ye, Wei Chow, Enxin Song, Qing-Guo Chen, Xiangtai Li, Zhen Dong, Lei Zhu, and Shuicheng Yan. Meissonic: Revitalizing Masked Generative Transformers for Efficient High-Resolution Text-to-Image Synthesis. arXiv preprint arXiv:2410.08261, 2024. 2 +[5] Yunkang Cao, Jiangning Zhang, Luca Frittoli, Yuqi Cheng, Weiming Shen, and Giacomo Boracchi. Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly detection. In European Conference on Computer Vision, pages 55-72. Springer, 2025. 6, 7 +[6] Zhiyuan Ding, Qi Dong, Haote Xu, Chenxin Li, Xinghao Ding, and Yue Huang. Unsupervised anomaly segmentation for brain lesions using dual semantic-manifold reconstruction. In International Conference on Neural Information Processing, pages 133-144. Springer International Publishing Cham, 2022. 2 +[7] Keval Doshi and Yasin Yilmaz. Any-shot sequential anomaly detection in surveillance videos. 2020. 2 +[8] Keval Doshi and Yasin Yilmaz. Continual learning for anomaly detection in surveillance videos. 2020. 1 +[9] Mariana-Iuliana Georgescu, RaduTudor Ionescu, Fahad-Shahbaz Khan, Marius Popescu, and Mubarak Shah. A background-agnostic framework with adversarial training for abnormal event detection in video. IEEE Transactions on Pattern Analysis and Machine Intelligence, IEEE Transactions on Pattern Analysis and Machine Intelligence. 6, 7 +[10] Mariana-Iuliana Georgescu, Antonio Barbalau, Radu Tudor Ionescu, Fahad Shahbaz Khan, Marius Popescu, and Mubarak Shah. Anomaly detection in video via self-supervised and multi-task learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12742-12752, 2021. 1, 2, 6, 7 +[11] Mahmudul Hasan, Jonghyun Choi, Jan Neumann, Amit K. Roy-Chowdhury, and Larry S. Davis. Learning temporal + +regularity in video sequences. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 2 +[12] Chunming He, Kai Li, Guoxia Xu, Yulun Zhang, Runze Hu, Zhenhua Guo, and Xiu Li. Degradation-resistant unfolding network for heterogeneous image fusion. In ICCV, pages 12611-12621, 2023. 2 +[13] Chunming He, Kai Li, Yachao Zhang, Guoxia Xu, Longxiang Tang, Yulun Zhang, Zhenhua Guo, and Xiu Li. Weakly-supervised concealed object segmentation with sam-based pseudo labeling and multi-scale feature grouping. NeurIPS, 36, 2024. 2 +[14] Chunming He, Kai Li, Yachao Zhang, Yulun Zhang, Zhenhua Guo, Xiu Li, Martin Danelljan, and Fisher Yu. Strategic preys make acute predators: Enhancing camouflaged object detectors by generating camouflaged objects. ICLR, 2024. 2 +[15] Chunming He, Chengyu Fang, Yulun Zhang, Kai Li, Longxiang Tang, Chenyu You, Fengyang Xiao, Zhenhua Guo, and Xiu Li. Reti-diff: Illumination degradation image restoration with retinae-based latent diffusion model. *ICLR*, 2025. 2 +[16] Chunming He, Yuqi Shen, Chengyu Fang, Fengyang Xiao, Longxiang Tang, Yulun Zhang, Wangmeng Zuo, Zhenhua Guo, and Xiu Li. Diffusion models in low-level vision: A survey. TPAMI, 2025. 2 +[17] Chunming He, Rihan Zhang, Fengyang Xiao, Chenyu Fang, Longxiang Tang, Yulun Zhang, Linghe Kong, Deng-Ping Fan, Kai Li, and Sina Farsiu. Run: Reversible unfolding network for concealed object segmentation. arXiv preprint arXiv:2501.18783, 2025. 2 +[18] Ryota Hinami, Tao Mei, and Shin'ichi Satoh. Joint detection and recounting of abnormal events by learning deep generic knowledge. In 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 1 +[19] Radu Tudor Ionescu, Fahad Shahbaz Khan, Mariana-Iuliana Georgescu, and Ling Shao. Object-centric auto-encoders and dummy anomalies for abnormal event detection in video. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 6 +[20] Hamza Karim, Keval Doshi, and Yasin Yilmaz. Real-time weakly supervised video anomaly detection. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 6848-6856, 2024. 2 +[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 3 +[22] Chenxin Li, Yunlong Zhang, Jiongcheng Li, Yue Huang, and Xinghao Ding. Unsupervised anomaly segmentation using image-semantic cycle translation. arXiv preprint arXiv:2103.09094, 2021. 2 +[23] Chenxin Li, Yunlong Zhang, Zhehan Liang, Wenao Ma, Yue Huang, and Xinghao Ding. Consistent posterior distributions under vessel-mixing: a regularization for cross-domain retinal artery/vein classification. In 2021 IEEE International Conference on Image Processing (ICIP), pages 61-65. IEEE, 2021. 5 + +[24] Chenxin Li, Wenao Ma, Liyan Sun, Xinghao Ding, Yue Huang, Guisheng Wang, and Yizhou Yu. Hierarchical deep network with uncertainty-aware semi-supervised learning for vessel segmentation. Neural Computing and Applications, pages 1-14, 2022. 2 +[25] Chenxin Li, Xinyu Liu, Wuyang Li, Cheng Wang, Hengyu Liu, Yifan Liu, Zhen Chen, and Yixuan Yuan. U-kan makes strong backbone for medical image segmentation and generation. arXiv preprint arXiv:2406.02918, 2024. 5 +[26] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2 +[27] Huangxing Lin, Yunlong Lin, Jingyuan Xia, Linyu Fan, Feifei Li, Yingying Wang, and Xinghao Ding. Fusion2void: Unsupervised multi-focus image fusion based on image inpainting. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2 +[28] Yunlong Lin, Tian Ye, Sixiang Chen, Zhenqi Fu, Yingying Wang, Wenhao Chai, Zhaohu Xing, Lei Zhu, and Xinghao Ding. Aglldiff: Guiding diffusion models towards unsupervised training-free real-world low-light image enhancement. arXiv preprint arXiv:2407.14900, 2024. 2 +[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2 +[30] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection-a new baseline. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6536-6545, 2018. 5 +[31] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13588-13597, 2021. 6, 7 +[32] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2 +[33] Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang. Simplenet: A simple network for image anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20402-20411, 2023. 6, 7 +[34] Shao-Yuan Lo, Poojan Oza, and Vishal M. Patel. Adversarially robust one-class novelty detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, page 1-12, 2022. 1, 2 +[35] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 2 +[36] Jakub Micorek, Horst Possegger, Dominik Narnhofer, Horst Bischof, and Mateusz Kozinski. MULDE: Multiscale Log-Density Estimation via Denoising Score Matching for Video + +Anomaly Detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18868-18877, 2024. 1 +[37] Arian Mousakhan, Thomas Brox, and Jawad Tayyub. Anomaly detection with conditioned denoising diffusion models. arXiv preprint arXiv:2305.15956, 2023. 6, 7 +[38] Yassine Naji, Aleksandr Setkov, Angelique Loesch, Michèle Gouiffes, and Romaric Audigier. Spatio-temporal predictive tasks for abnormal event detection in videos. 2022. 1, 6, 7 +[39] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2 +[40] Bharathkumar Ramachandra and Michael Jones. Street scene: A new dataset and evaluation protocol for video anomaly detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2569-2578, 2020. 1, 6, 7 +[41] Bharathkumar Ramachandra, Michael J. Jones, and Ranga Raju Vatsavai. Learning a distance function with a siamese network to localize anomalies in videos. In 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), 2020. 1, 2, 6, 7 +[42] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2, 3, 6, 7 +[43] Tal Reiss and Yedid Hoshen. Attribute-based representations for accurate and interpretable video anomaly detection. 2022. 1, 2, 6 +[44] Chenrui Shi, Che Sun, Yuwei Wu, and Yunde Jia. Video anomaly detection via sequentially learning multiple pretext tasks. 1, 2 +[45] Liyan Sun, Chenxin Li, Xinghao Ding, Yue Huang, Zhong Chen, Guisheng Wang, Yizhou Yu, and John Paisley. Few-shot medical image segmentation using a global correlation network with discriminative embedding. Computers in biology and medicine, 140:105067, 2022. 2 +[46] Yu Tian, Guansong Pang, Yuanhong Chen, Rajvinder Singh, Johan W Verjans, and Gustavo Carneiro. Weakly-supervised video anomaly detection with robust temporal feature magnitude learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4975-4986, 2021. 2 +[47] Shu Wang and Zhenjiang Miao. Anomaly detection in crowd scene. In IEEE 10th International Conference on Signal Processing Proceedings, pages 1220-1223. IEEE, 2010. 5 +[48] Yingying Wang, Yunlong Lin, Ge Meng, Zhenqi Fu, Yuhang Dong, Linyu Fan, Hedeng Yu, Xinghao Ding, and Yue Huang. Learning high-frequency feature enhancement and alignment for pan-sharpening. In Proceedings of the 31st ACM International Conference on Multimedia, pages 358-367, 2023. 1 +[49] Yingying Wang, Yunlong Lin, Xuanhua He, Hui Zheng, Keyu Yan, Linyu Fan, Yue Huang, and Xinghao Ding. Learning diffusion high-quality priors for pan-sharpening: A two-stage approach with time-aware adapter fine-tuning. IEEE Transactions on Geoscience and Remote Sensing, 2025. 1 + +[50] Jhih-Ciang Wu, He-Yen Hsieh, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng-Luh Liu. Self-supervised sparse representation for video anomaly detection. In European Conference on Computer Vision, pages 729-745. Springer, 2022. 2 +[51] Peng Wu, Xuerong Zhou, Guansong Pang, Lingru Zhou, Qingsen Yan, Peng Wang, and Yanning Zhang. Vadclip: Adapting vision-language models for weakly supervised video anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6074-6082, 2024. 2 +[52] Fengyang Xiao, Sujie Hu, Yuqi Shen, Chengyu Fang, Jina Huang, Chunming He, Longxiang Tang, Ziyun Yang, and Xiu Li. A survey of camouflaged object detection and beyond. CAIAIR, 2024. 2 +[53] Zhaohu Xing, Lequan Yu, Liang Wan, Tong Han, and Lei Zhu. Nestedformer: Nested modality-aware transformer for brain tumor segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 140-150. Springer, 2022. 2 +[54] Zhaohu Xing, Liang Wan, Huazhu Fu, Guang Yang, and Lei Zhu. Diff-unet: A diffusion embedded network for volumetric segmentation. arXiv preprint arXiv:2303.10326, 2023.2 +[55] Zhaohu Xing, Sicheng Yang, Sixiang Chen, Tian Ye, Yijun Yang, Jing Qin, and Lei Zhu. Cross-conditioned diffusion model for medical image to image translation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 201-211. Springer, 2024. 2 +[56] Zhaohu Xing, Tian Ye, Yijun Yang, Guang Liu, and Lei Zhu. Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 578-588. Springer, 2024. +[57] Yuchen Yang, Kwonjoon Lee, Behzad Dariush, Yinzhi Cao, and Shao-Yuan Lo. Follow the rules: Reasoning for video anomaly detection with large language models. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 1, 2 +[58] Guang Yu, Siqi Wang, Zhiping Cai, En Zhu, Chuanfu Xu, Jianping Yin, and Marius Kloft. Cloze test helps: Effective video anomaly detection via learning to complete video events. In Proceedings of the 28th ACM International Conference on Multimedia, 2020. 2 +[59] Luca Zanella, Willi Menapace, Massimiliano Mancini, Yiming Wang, and Elisa Ricci. Harnessing large language models for training-free video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18527-18536, 2024. 2 +[60] Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8330-8339, 2021. 6, 7 +[61] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 2 +[62] Yunlong Zhang, Chenxin Li, Xin Lin, Liyan Sun, Yihong Zhuang, Yue Huang, Xinghao Ding, Xiaqing Liu, and + +Yizhou Yu. Generator versus segmentor: Pseudo-healthy synthesis. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part VI 24, pages 150–160. Springer International Publishing, 2021. 2 +[63] Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. Anomalyclip: Object-agnostic prompt learning for zero-shot anomaly detection. arXiv preprint arXiv:2310.18961, 2023. 6, 7 +[64] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36, 2024. 3 \ No newline at end of file diff --git a/2025/Track Any Anomalous Object_A Granular Video Anomaly Detection Pipeline/images.zip b/2025/Track Any Anomalous Object_A Granular Video Anomaly Detection Pipeline/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..58c9ad003b4fd16b3e9595d4705de3192043dbfc --- /dev/null +++ b/2025/Track Any Anomalous Object_A Granular Video Anomaly Detection Pipeline/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f37abce2b8561d04cef150356819b1548b696dfdb2da17260414faae79cf2a2 +size 562253 diff --git a/2025/Track Any Anomalous Object_A Granular Video Anomaly Detection Pipeline/layout.json b/2025/Track Any Anomalous Object_A Granular Video Anomaly Detection Pipeline/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..aa771c301c80622f6c95736d446c2eb3359b2630 --- /dev/null +++ b/2025/Track Any Anomalous Object_A Granular Video Anomaly Detection Pipeline/layout.json @@ -0,0 +1,9472 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 103, + 547, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 103, + 547, + 121 + ], + "spans": [ + { + "bbox": [ + 64, + 103, + 547, + 121 + ], + "type": "text", + "content": "Track Any Anomalous Object: A Granular Video Anomaly Detection Pipeline" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 148, + 558, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 558, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 558, + 178 + ], + "type": "text", + "content": "Yuzhi Huang" + }, + { + "bbox": [ + 61, + 148, + 558, + 178 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 61, + 148, + 558, + 178 + ], + "type": "text", + "content": ", Chenxin Li" + }, + { + "bbox": [ + 61, + 148, + 558, + 178 + ], + "type": "inline_equation", + "content": "^{2*†}" + }, + { + "bbox": [ + 61, + 148, + 558, + 178 + ], + "type": "text", + "content": ", Haitao Zhang" + }, + { + "bbox": [ + 61, + 148, + 558, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 148, + 558, + 178 + ], + "type": "text", + "content": ", Zixu Lin" + }, + { + "bbox": [ + 61, + 148, + 558, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 148, + 558, + 178 + ], + "type": "text", + "content": ", Yunlong Lin" + }, + { + "bbox": [ + 61, + 148, + 558, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 148, + 558, + 178 + ], + "type": "text", + "content": ", Hengyu Liu" + }, + { + "bbox": [ + 61, + 148, + 558, + 178 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 148, + 558, + 178 + ], + "type": "text", + "content": ", Wuyang Li" + }, + { + "bbox": [ + 61, + 148, + 558, + 178 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 148, + 558, + 178 + ], + "type": "text", + "content": ", Xinyu Liu" + }, + { + "bbox": [ + 61, + 148, + 558, + 178 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 148, + 558, + 178 + ], + "type": "text", + "content": ", Jiechao Gao" + }, + { + "bbox": [ + 61, + 148, + 558, + 178 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 148, + 558, + 178 + ], + "type": "text", + "content": ", Yue Huang" + }, + { + "bbox": [ + 61, + 148, + 558, + 178 + ], + "type": "inline_equation", + "content": "^{1†}" + }, + { + "bbox": [ + 61, + 148, + 558, + 178 + ], + "type": "text", + "content": ", Xinghao Ding" + }, + { + "bbox": [ + 61, + 148, + 558, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 148, + 558, + 178 + ], + "type": "text", + "content": " and Yixuan Yuan" + }, + { + "bbox": [ + 61, + 148, + 558, + 178 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 123, + 178, + 496, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 178, + 496, + 221 + ], + "spans": [ + { + "bbox": [ + 123, + 178, + 496, + 221 + ], + "type": "text", + "content": "1 Key Laboratory of Multimedia Trusted Perception and Efficien Computing, Ministry of Education of China, Xiamen University School of Informatics, Xiamen University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 220, + 483, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 220, + 483, + 236 + ], + "spans": [ + { + "bbox": [ + 137, + 220, + 483, + 236 + ], + "type": "text", + "content": "2 The Chinese University of Hong Kong 3 The University of Virginia" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 274, + 200, + 287 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 274, + 200, + 287 + ], + "spans": [ + { + "bbox": [ + 152, + 274, + 200, + 287 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 300, + 297, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 300, + 297, + 587 + ], + "spans": [ + { + "bbox": [ + 54, + 300, + 297, + 587 + ], + "type": "text", + "content": "Video anomaly detection (VAD) is crucial in scenarios such as surveillance and autonomous driving, where timely detection of unexpected activities is essential. Albeit existing methods have primarily focused on detecting anomalous objects in videos—either by identifying anomalous frames or objects—they often neglect finer-grained analysis, such as anomalous pixels, which limits their ability to capture a broader range of anomalies. To address this challenge, we propose an innovative VAD framework called Track Any Anomalous Object (TAO), which introduces a Granular Video Anomaly Detection Framework that, for the first time, integrates the detection of multiple fine-grained anomalous objects into a unified framework. Unlike methods that assign anomaly scores to every pixel at each moment, our approach transforms the problem into pixel-level tracking of anomalous objects. By linking anomaly scores to subsequent tasks such as image segmentation and video tracking, our method eliminates the need for threshold selection and achieves more precise anomaly localization, even in long and challenging video sequences. Experiments on extensive datasets demonstrate that TAO achieves state-of-the-art performance, setting a new progress for VAD by providing a practical, granular, and holistic solution. For more information, visit the project page at: https://tao-25.github.io/" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 135, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 135, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 135, + 605 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 613, + 296, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 296, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 296, + 686 + ], + "type": "text", + "content": "Video Anomaly Detection (VAD) involves identifying unusual or unexpected activities in surveillance videos and has significant applications in areas such as security monitoring (e.g., detecting violent behavior) and autonomous driving (e.g., recognizing traffic accidents). Current VAD research has developed along two main directions, which" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 321, + 263, + 552, + 404 + ], + "blocks": [ + { + "bbox": [ + 321, + 263, + 552, + 404 + ], + "lines": [ + { + "bbox": [ + 321, + 263, + 552, + 404 + ], + "spans": [ + { + "bbox": [ + 321, + 263, + 552, + 404 + ], + "type": "image", + "image_path": "e025d956f5427181d2ecc637d75e15e453a7ca7bd5a48ab72235bf197258fcec.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 413, + 555, + 513 + ], + "lines": [ + { + "bbox": [ + 313, + 413, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 555, + 513 + ], + "type": "text", + "content": "Figure 1. Analyzing the Limitations of Existing VAD Models. Video anomaly detection (VAD) models are predominantly frame-centric or object-centric. Frame-centric methods detect anomalies in frames without localizing them, while object-centric methods identify anomalous objects but lack pixel-level accuracy. Pixel-centric models address these gaps by providing pixel-level localization, delivering fine-grained segmentation and precise delineation of anomalies, particularly for overlapping objects where traditional methods struggle." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 533, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 677 + ], + "type": "text", + "content": "are reflected in both methodologies and evaluation benchmarks. Frame-centric methods [1, 2, 11, 32, 34, 39, 44, 57] and their corresponding frame-level benchmarks [18, 36, 49, 57] focus on analyzing entire frames to detect global anomalies (e.g., fires or smoke), typically evaluated using frame-level AUC metrics. In contrast, object-centric methods [3, 8-10, 19, 38, 41, 43, 48] and their object-level benchmarks [10, 19, 40] leverage pre-trained feature extractors to detect object-specific anomalies (e.g., human falls or vehicle accidents), evaluated using metrics such as TBDC [40] and RBDC [40], which comprehensively measure temporal consistency and spatial localization precision." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Despite these advances, a critical gap remains in real-world applications of VAD where fine-grained localization of anomalous objects is essential. As illustrated in Fig. 1," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 693, + 211, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 211, + 704 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 211, + 704 + ], + "type": "text", + "content": "* Equal contribution † Corresponding author" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8689" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "current approaches suffer from fundamental shortage as frame-centric methods can only identify the presence of anomalies without localizing specific regions, while object-centric methods, though more precise, lack pixel-level accuracy. These limitations become particularly pronounced in complex scenarios with overlapping anomalies, where accurate delineation of object boundaries and shapes is crucial [4, 12, 14, 20, 28, 46, 50]. This observation motivates us to consider a more challenging yet practical scenario: Can we achieve both object-level structural integrity and pixel-level precision in a granular anomaly detection?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 205, + 295, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 295, + 373 + ], + "type": "text", + "content": "To address this challenge, we propose enhancing existing VAD benchmarks with pixel-level anomaly evaluation. While a straightforward approach would be to employ an anomaly segmentation task for pixel-wise classification [6, 13, 16, 22, 27, 62], this method falls short when dealing with multiple, potentially overlapping anomalous objects. Instead, inspired by instance segmentation techniques [15, 17, 24, 45, 52], we develop a novel evaluation framework that simultaneously considers both object-level and pixel-level anomaly detection accuracy. This framework ensures that detected anomalies accurately correspond to their true shapes, precise positions, and spatial distributions, while effectively minimizing false positives caused by local feature similarities or noise." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 375, + 295, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 295, + 530 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 295, + 530 + ], + "type": "text", + "content": "However, implementing precise pixel-level tracking in videos presents significant challenges, particularly in maintaining semantic and temporal consistency across frames. Traditional supervised approaches require extensive labeled data, which is scarce in existing datasets with pixel-level annotations. To overcome these challenges, we leverage SAM2 [42], a large-scale pre-trained segmentation model capable of processing both static images and video streams in real-time without requiring additional fine-tuning on specific anomaly datasets. Its ability to precisely delineate object contours and structures makes it particularly effective in handling complex scenarios involving overlapping anomalies or occlusions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 532, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 700 + ], + "type": "text", + "content": "Building upon these foundations, we propose TAO, an integrated framework that combines object-centric anomaly detection algorithms with SAM2 to create a simple yet effective fine-grained video anomaly detection system. For videos containing anomalous frames, our object-centric detection model generates bounding boxes for detected anomalies. After robust filtering, these boxes serve as prompting boxes for SAM2, which then generates segmentation masks for the anomalous objects. Evaluated on our proposed benchmark, our method achieves state-of-the-art performance on two widely-used datasets, demonstrating its effectiveness in bridging the gap between object-level detection and pixel-precise segmentation. The contributions of our paper are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "- We present a new testing standard that unifies pixel-level" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 322, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 108 + ], + "type": "text", + "content": "and object-level assessment in video anomaly detection, addressing the limitations of existing metrics in complex scenarios with multiple or overlapping anomalies." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 108, + 553, + 216 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 313, + 108, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 553, + 167 + ], + "type": "text", + "content": "- We introduce TAO, a streamlined framework that integrates object-centric anomaly detection with the segmentation capabilities from vision foundation model, enabling precise pixel-level tracking of anomalous objects without requiring additional fine-tuning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 167, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 167, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 553, + 216 + ], + "type": "text", + "content": "- Through extensive experiments on UCSD Ped2 and ShanghaiTech Campus datasets, we demonstrate that our approach achieves state-of-the-art performance under both traditional metrics and our proposed benchmark." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 227, + 400, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 227, + 400, + 239 + ], + "spans": [ + { + "bbox": [ + 314, + 227, + 400, + 239 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 247, + 553, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 553, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 553, + 593 + ], + "type": "text", + "content": "Video Anomaly Detection. Video anomaly detection (VAD) presents significant challenges due to the rarity of anomaly data and the wide variety of abnormal events, which hinder the generalization capabilities of existing models across diverse scenarios. Traditional VAD methods can be broadly classified into frame-centric and object-centric approaches. Frame-centric methods [1, 11, 32, 34, 39, 44, 55-57] analyze entire frames or sequences, often leveraging reconstruction or prediction errors to identify anomalies. These methods are particularly effective for detecting global events, such as fires or smoke. However, they struggle with localized anomalies and are prone to interference from normal regions within a frame, limiting their precision in complex scenarios [11, 32, 34, 39, 53]. Object-centric methods [7-10, 19, 41, 43, 54, 58] focus on specific objects within frames by employing pre-trained object detection models to extract bounding boxes and assess their abnormality. This targeted approach is better suited for detecting anomalies related to individuals or objects, such as human falls or vehicle accidents. By concentrating on specific regions and reducing redundant information, object-centric methods achieve improved accuracy and robustness, especially in complex, multi-object environments. Building on these foundations, our approach is the first to integrate large-scale pre-trained models for pixel-level fine-grained detection in VAD. This advancement enables precise anomaly localization, bridging the gap between frame-level analysis and object-level detection, and significantly enhancing overall detection performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "type": "text", + "content": "Vision Foundation Models. In recent years, large-scale pre-trained language models (LLMs) and vision-language models (VLMs) have shown remarkable potential in advancing video anomaly detection. Models such as BLIP-2 [26], LLaVA [29], LAVAD [59], VadCLIP [51], VideoChatGPT [35], and Video-LLaMA [61] have significantly improved the understanding of complex visual tasks by integrating visual and linguistic information. Concurrently, prompting techniques have gained traction in large-scale vision models, enabling substantial progress in tasks such" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8690" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 76, + 345, + 395 + ], + "blocks": [ + { + "bbox": [ + 62, + 76, + 345, + 395 + ], + "lines": [ + { + "bbox": [ + 62, + 76, + 345, + 395 + ], + "spans": [ + { + "bbox": [ + 62, + 76, + 345, + 395 + ], + "type": "image", + "image_path": "d519cad781a79038c645689825d2eb24ba22e10e45c532beeec93b559cf60ddd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 408, + 555, + 453 + ], + "lines": [ + { + "bbox": [ + 54, + 408, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 54, + 408, + 555, + 453 + ], + "type": "text", + "content": "Figure 2. Pipeline of our proposed TAO. We first generate bounding boxes to identify objects in each frame. Next, we score these boxes using object-centric video anomaly detection algorithms to extract potential anomalous boxes. To ensure robustness, we apply filtering to eliminate redundant boxes. Finally, the filtered boxes and original frames are input into a prompt-based segmentation model to produce pixel-level anomaly segmentation masks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 347, + 76, + 552, + 396 + ], + "blocks": [ + { + "bbox": [ + 347, + 76, + 552, + 396 + ], + "lines": [ + { + "bbox": [ + 347, + 76, + 552, + 396 + ], + "spans": [ + { + "bbox": [ + 347, + 76, + 552, + 396 + ], + "type": "image", + "image_path": "d07cfa6c08efc611527ce94c625577b26a2954cc59c421483f06cfa248da8c35.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 475, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 297, + 713 + ], + "type": "text", + "content": "as image segmentation. By employing semantic prompts (e.g., free-form text) and spatial prompts (e.g., points or bounding boxes), these models achieve highly accurate segmentation guided by input cues. Segmentation Foundation Models (SFMs), such as the Segment Anything Model (SAM) [21] and SEEM [64], have demonstrated exceptional performance, particularly in zero-shot generalization and multi-modal prompting, allowing for cross-task segmentation. Building on these advancements, the Segment Anything Model 2 (SAM2) [42] further improves segmentation, particularly for video-based applications. SAM2 integrates prompting techniques to not only segment static objects with precision but also capture dynamic, complex objects in video contexts. This capability enables SAM2 to deliver finer segmentation and enhanced localization accuracy in video anomaly detection, addressing the challenges posed by real-world variability and complexity. This work leverages SAM2's advanced video segmentation capabilities to tackle the challenge of fine-grained video anomaly detection, achieving precise and robust anomaly identification in" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 475, + 394, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 475, + 394, + 486 + ], + "spans": [ + { + "bbox": [ + 314, + 475, + 394, + 486 + ], + "type": "text", + "content": "complex scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 500, + 370, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 500, + 370, + 511 + ], + "spans": [ + { + "bbox": [ + 314, + 500, + 370, + 511 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "text", + "content": "Preliminary. In the proposed video anomaly detection (VAD) benchmark, a video clip is represented as a sequence " + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\{f_1,f_2,\\ldots ,f_{N_f}\\} \\subseteq \\mathcal{C}" + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "text", + "content": " is the set of all possible video clips, and " + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "inline_equation", + "content": "N_{f}" + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "text", + "content": " denotes the total number of frames. Each frame " + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "text", + "content": " is expressed as " + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "inline_equation", + "content": "f_{i} = [p_{i,1},p_{i,2},\\dots ,p_{i,N_{i}}]" + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "inline_equation", + "content": "N_{i}" + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "text", + "content": " is the total number of pixels in the frame, and " + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "inline_equation", + "content": "p_{i,j}\\in \\mathcal{P}" + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "text", + "content": "-th pixel, with " + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "text", + "content": " as the set of possible pixel values. Each pixel " + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "inline_equation", + "content": "p_{i,j}" + }, + { + "bbox": [ + 313, + 521, + 555, + 628 + ], + "type": "text", + "content": " is classified as either \"normal\" or \"anomalous\"." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Overview. To achieve precise anomaly localization in videos, we design a streamlined pipeline for pixel-level segmentation in video anomaly detection. Our pipeline comprises four stages: bounding boxes extraction, anomalous boxes extraction, robust filtering, and segmentation inference. The process begins by generating bounding boxes to identify objects of interest. These boxes are then scored us" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8691" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 72, + 423, + 150 + ], + "blocks": [ + { + "bbox": [ + 63, + 72, + 423, + 150 + ], + "lines": [ + { + "bbox": [ + 63, + 72, + 423, + 150 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 423, + 150 + ], + "type": "image", + "image_path": "6adbbf4d76a6cb8125318ce77626324346965b73396d3c8cda8c9bec6039ebd1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 161, + 555, + 195 + ], + "lines": [ + { + "bbox": [ + 55, + 161, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 55, + 161, + 555, + 195 + ], + "type": "text", + "content": "Figure 3. Impact of Redundant Segmentation on Performance. Potential anomalous boxes are used as prompts for SAM2 segmentation (red masks: true anomalies, green masks: redundant results). Initially, SAM2 tracks true anomalies accurately, but tracking errors accumulate, leading to catastrophic forgetting and a collapse in Pixel-F1 performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 424, + 72, + 550, + 151 + ], + "blocks": [ + { + "bbox": [ + 424, + 72, + 550, + 151 + ], + "lines": [ + { + "bbox": [ + 424, + 72, + 550, + 151 + ], + "spans": [ + { + "bbox": [ + 424, + 72, + 550, + 151 + ], + "type": "image", + "image_path": "daaff1e3489f33ad5e52695132fea706efa7667581fbb2a3cb06768da8caff8b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 216, + 223, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 216, + 223, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 216, + 223, + 227 + ], + "type": "text", + "content": "Algorithm 1 Boxes Robustness Filtering" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 230, + 296, + 382 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 61, + 230, + 296, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 230, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 61, + 230, + 296, + 289 + ], + "type": "text", + "content": "1: Input: A set of anomalous bounding boxes " + }, + { + "bbox": [ + 61, + 230, + 296, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{anomaly}} = \\{b_i \\mid s_i > \\tau\\}" + }, + { + "bbox": [ + 61, + 230, + 296, + 289 + ], + "type": "text", + "content": " for each frame " + }, + { + "bbox": [ + 61, + 230, + 296, + 289 + ], + "type": "inline_equation", + "content": "f_i \\in \\{f_1, f_2, \\ldots, f_n\\}" + }, + { + "bbox": [ + 61, + 230, + 296, + 289 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 61, + 230, + 296, + 289 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 61, + 230, + 296, + 289 + ], + "type": "text", + "content": " is the anomaly score and " + }, + { + "bbox": [ + 61, + 230, + 296, + 289 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 61, + 230, + 296, + 289 + ], + "type": "text", + "content": " is the threshold; parameters " + }, + { + "bbox": [ + 61, + 230, + 296, + 289 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 61, + 230, + 296, + 289 + ], + "type": "text", + "content": " (tracking window size), " + }, + { + "bbox": [ + 61, + 230, + 296, + 289 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 61, + 230, + 296, + 289 + ], + "type": "text", + "content": " (overlap threshold), " + }, + { + "bbox": [ + 61, + 230, + 296, + 289 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 61, + 230, + 296, + 289 + ], + "type": "text", + "content": " (save interval), and " + }, + { + "bbox": [ + 61, + 230, + 296, + 289 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 61, + 230, + 296, + 289 + ], + "type": "text", + "content": " (frame match threshold)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 289, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 289, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 61, + 289, + 295, + 312 + ], + "type": "text", + "content": "2: Output: Filtered bounding boxes " + }, + { + "bbox": [ + 61, + 289, + 295, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{filtered}}" + }, + { + "bbox": [ + 61, + 289, + 295, + 312 + ], + "type": "text", + "content": " with their corresponding object labels " + }, + { + "bbox": [ + 61, + 289, + 295, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_j" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 313, + 294, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 313, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 62, + 313, + 294, + 335 + ], + "type": "text", + "content": "3: Initialize the list " + }, + { + "bbox": [ + 62, + 313, + 294, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_i" + }, + { + "bbox": [ + 62, + 313, + 294, + 335 + ], + "type": "text", + "content": " to store the filtered bounding boxes from frame " + }, + { + "bbox": [ + 62, + 313, + 294, + 335 + ], + "type": "inline_equation", + "content": "f_{i}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 335, + 223, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 223, + 347 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 223, + 347 + ], + "type": "text", + "content": "4: for each frame " + }, + { + "bbox": [ + 62, + 335, + 223, + 347 + ], + "type": "inline_equation", + "content": "f_{i} \\in \\{f_{1}, \\ldots, f_{n}\\}" + }, + { + "bbox": [ + 62, + 335, + 223, + 347 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 347, + 269, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 269, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 269, + 357 + ], + "type": "text", + "content": "5: Step 1: Inherit Existing Anomalous Boxes" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 358, + 295, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 358, + 295, + 382 + ], + "spans": [ + { + "bbox": [ + 61, + 358, + 295, + 382 + ], + "type": "text", + "content": "6: for each bounding box " + }, + { + "bbox": [ + 61, + 358, + 295, + 382 + ], + "type": "inline_equation", + "content": "b_{j} \\in \\mathcal{B}_{\\mathrm{anomaly}}" + }, + { + "bbox": [ + 61, + 358, + 295, + 382 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 61, + 358, + 295, + 382 + ], + "type": "inline_equation", + "content": "b_{p}" + }, + { + "bbox": [ + 61, + 358, + 295, + 382 + ], + "type": "text", + "content": " from frame " + }, + { + "bbox": [ + 61, + 358, + 295, + 382 + ], + "type": "inline_equation", + "content": "f_{i - k}" + }, + { + "bbox": [ + 61, + 358, + 295, + 382 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 61, + 358, + 295, + 382 + ], + "type": "inline_equation", + "content": "f_{i - 1}" + }, + { + "bbox": [ + 61, + 358, + 295, + 382 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 62, + 382, + 258, + 408 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 382, + 258, + 408 + ], + "spans": [ + { + "bbox": [ + 62, + 382, + 258, + 408 + ], + "type": "interline_equation", + "content": "7: \\quad \\text {i f} \\sum_ {p = i - k} ^ {i - 1} \\mathbb {I} (\\operatorname {I o U} (b _ {j}, b _ {p}) > h) \\geq m \\text {t h e n}", + "image_path": "c5401000b5da27c76498a182f3a83d22796de0bf80ce0e6a0a25b176dbfafa9f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 408, + 198, + 432 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 62, + 408, + 186, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 408, + 186, + 420 + ], + "spans": [ + { + "bbox": [ + 62, + 408, + 186, + 420 + ], + "type": "text", + "content": "8: Assign " + }, + { + "bbox": [ + 62, + 408, + 186, + 420 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_j = \\mathcal{L}_p" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 420, + 198, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 198, + 432 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 198, + 432 + ], + "type": "text", + "content": "9: Remove " + }, + { + "bbox": [ + 62, + 420, + 198, + 432 + ], + "type": "inline_equation", + "content": "b_{j}" + }, + { + "bbox": [ + 62, + 420, + 198, + 432 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 62, + 420, + 198, + 432 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{i}" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 59, + 433, + 295, + 491 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 59, + 433, + 130, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 433, + 130, + 442 + ], + "spans": [ + { + "bbox": [ + 59, + 433, + 130, + 442 + ], + "type": "text", + "content": "10: end if" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 59, + 445, + 121, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 445, + 121, + 454 + ], + "spans": [ + { + "bbox": [ + 59, + 445, + 121, + 454 + ], + "type": "text", + "content": "11: end for" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 59, + 455, + 251, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 455, + 251, + 466 + ], + "spans": [ + { + "bbox": [ + 59, + 455, + 251, + 466 + ], + "type": "text", + "content": "12: Step 2: Assign New Anomalous Boxes" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 59, + 466, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 466, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 59, + 466, + 295, + 491 + ], + "type": "text", + "content": "13: for each remaining " + }, + { + "bbox": [ + 59, + 466, + 295, + 491 + ], + "type": "inline_equation", + "content": "b_{j} \\in \\mathcal{B}_{\\mathrm{anomaly}}" + }, + { + "bbox": [ + 59, + 466, + 295, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 59, + 466, + 295, + 491 + ], + "type": "inline_equation", + "content": "b_{p}" + }, + { + "bbox": [ + 59, + 466, + 295, + 491 + ], + "type": "text", + "content": " from frame " + }, + { + "bbox": [ + 59, + 466, + 295, + 491 + ], + "type": "inline_equation", + "content": "f_{i + 1}" + }, + { + "bbox": [ + 59, + 466, + 295, + 491 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 59, + 466, + 295, + 491 + ], + "type": "inline_equation", + "content": "f_{i + k}" + }, + { + "bbox": [ + 59, + 466, + 295, + 491 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 61, + 491, + 258, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 491, + 258, + 517 + ], + "spans": [ + { + "bbox": [ + 61, + 491, + 258, + 517 + ], + "type": "interline_equation", + "content": "\\text {i f} \\sum_ {p = i + 1} ^ {i + k} \\mathbb {I} \\left(\\operatorname {I o U} \\left(b _ {j}, b _ {p}\\right) > h\\right) \\geq m \\text {t h e n} \\tag {14:}", + "image_path": "a5aa9f05b7ac484512f95ef0dfa40c7f40b98433421bb69b2b3cae5ce6e02f39.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 58, + 491, + 294, + 610 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 59, + 517, + 272, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 517, + 272, + 529 + ], + "spans": [ + { + "bbox": [ + 59, + 517, + 272, + 529 + ], + "type": "text", + "content": "15: Assign a new object label " + }, + { + "bbox": [ + 59, + 517, + 272, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_j = \\mathcal{F}(b_j)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 61, + 529, + 252, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 529, + 252, + 541 + ], + "spans": [ + { + "bbox": [ + 61, + 529, + 252, + 541 + ], + "type": "text", + "content": "16: Save the tuple " + }, + { + "bbox": [ + 61, + 529, + 252, + 541 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{box}} = (f_i, b_j, \\mathcal{L}_j)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 59, + 541, + 130, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 541, + 130, + 551 + ], + "spans": [ + { + "bbox": [ + 59, + 541, + 130, + 551 + ], + "type": "text", + "content": "17: end if" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 59, + 553, + 121, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 553, + 121, + 563 + ], + "spans": [ + { + "bbox": [ + 59, + 553, + 121, + 563 + ], + "type": "text", + "content": "18: end for" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 59, + 563, + 238, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 563, + 238, + 575 + ], + "spans": [ + { + "bbox": [ + 59, + 563, + 238, + 575 + ], + "type": "text", + "content": "19: Step 3: Save Boxes Every " + }, + { + "bbox": [ + 59, + 563, + 238, + 575 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 59, + 563, + 238, + 575 + ], + "type": "text", + "content": " Frames" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 59, + 576, + 178, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 576, + 178, + 586 + ], + "spans": [ + { + "bbox": [ + 59, + 576, + 178, + 586 + ], + "type": "text", + "content": "20: if " + }, + { + "bbox": [ + 59, + 576, + 178, + 586 + ], + "type": "inline_equation", + "content": "i \\bmod l == 0" + }, + { + "bbox": [ + 59, + 576, + 178, + 586 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 58, + 587, + 294, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 587, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 58, + 587, + 294, + 610 + ], + "type": "text", + "content": "21: Save all bounding box tuples " + }, + { + "bbox": [ + 58, + 587, + 294, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{box}} = (f_i, b_j, \\mathcal{L}_j)" + }, + { + "bbox": [ + 58, + 587, + 294, + 610 + ], + "type": "text", + "content": " in the current frame" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 59, + 612, + 116, + 633 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 59, + 612, + 116, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 612, + 116, + 622 + ], + "spans": [ + { + "bbox": [ + 59, + 612, + 116, + 622 + ], + "type": "text", + "content": "22: end if" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 59, + 623, + 106, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 623, + 106, + 633 + ], + "spans": [ + { + "bbox": [ + 59, + 623, + 106, + 633 + ], + "type": "text", + "content": "23: end for" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "ing object-centric anomaly detection algorithms to extract potential anomalies. A tracking-based filtering step follows to eliminate redundant boxes and retain only robust ones. Finally, the filtered boxes and original frames are fed into" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 216, + 555, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 240 + ], + "type": "text", + "content": "a prompt-based segmentation model to generate pixel-level anomaly masks. The pipeline is illustrated in Fig. 2." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 247, + 473, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 473, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 473, + 258 + ], + "type": "text", + "content": "3.1. Anomalous Boxes Extraction" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 264, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 555, + 361 + ], + "type": "text", + "content": "Bounding Boxes Extraction. To achieve precise pixel-level segmentation in video anomaly detection, pre-trained object detection algorithms are used to generate bounding boxes for objects within a frame, serving as potential prompts for guiding the segmentation model. For a given input frame " + }, + { + "bbox": [ + 313, + 264, + 555, + 361 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 264, + 555, + 361 + ], + "type": "text", + "content": ", the object detection model " + }, + { + "bbox": [ + 313, + 264, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 264, + 555, + 361 + ], + "type": "text", + "content": " outputs " + }, + { + "bbox": [ + 313, + 264, + 555, + 361 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 264, + 555, + 361 + ], + "type": "text", + "content": " bounding boxes " + }, + { + "bbox": [ + 313, + 264, + 555, + 361 + ], + "type": "inline_equation", + "content": "b_{1}, b_{2}, \\ldots, b_{m}" + }, + { + "bbox": [ + 313, + 264, + 555, + 361 + ], + "type": "text", + "content": ", each paired with a class label " + }, + { + "bbox": [ + 313, + 264, + 555, + 361 + ], + "type": "inline_equation", + "content": "y_{1}, y_{2}, \\ldots, y_{m}" + }, + { + "bbox": [ + 313, + 264, + 555, + 361 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 348, + 367, + 554, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 367, + 554, + 380 + ], + "spans": [ + { + "bbox": [ + 348, + 367, + 554, + 380 + ], + "type": "interline_equation", + "content": "\\{(b _ {1}, y _ {1}), (b _ {2}, y _ {2}), \\dots , (b _ {m}, y _ {m}) \\} = \\mathcal {D} (f) \\tag {1}", + "image_path": "d5d02df777bee6139596fb9c2217c7f24078bb04f7d37342f83e716eeb1f10d0.jpg" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 385, + 554, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 554, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 554, + 445 + ], + "type": "text", + "content": "After the extraction of object bounding boxes in each frame, object-centric VAD algorithms compute an anomaly score " + }, + { + "bbox": [ + 313, + 385, + 554, + 445 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 313, + 385, + 554, + 445 + ], + "type": "text", + "content": " for each object within its bounding box " + }, + { + "bbox": [ + 313, + 385, + 554, + 445 + ], + "type": "inline_equation", + "content": "b_i" + }, + { + "bbox": [ + 313, + 385, + 554, + 445 + ], + "type": "text", + "content": ". This score is derived from features such as pose, speed, and depth, and can be formally represented as:" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 410, + 455, + 553, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 455, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 410, + 455, + 553, + 468 + ], + "type": "interline_equation", + "content": "s _ {i} = \\mathcal {A} \\left(b _ {i}\\right) \\tag {2}", + "image_path": "73a0de432a3426b2121de90f7eabb12bf5160bd02fbb9e1448ba8ccf8f73978d.jpg" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 477, + 554, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 477, + 554, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 554, + 549 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 477, + 554, + 549 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 477, + 554, + 549 + ], + "type": "text", + "content": " represents the scoring function used in object-centric video anomaly detection algorithms. To isolate anomalous objects, we filter the bounding boxes based on an anomaly score threshold. Specifically, we retain only the bounding boxes where the anomaly score exceeds a predefined threshold " + }, + { + "bbox": [ + 313, + 477, + 554, + 549 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 477, + 554, + 549 + ], + "type": "text", + "content": ", as shown in the following expression:" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 383, + 558, + 553, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 558, + 553, + 571 + ], + "spans": [ + { + "bbox": [ + 383, + 558, + 553, + 571 + ], + "type": "interline_equation", + "content": "\\mathcal {B} _ {\\text {a n o m a l y}} = \\left\\{b _ {i} \\mid s _ {i} > \\tau \\right\\} \\tag {3}", + "image_path": "d077ca2a8f188b0924baeb5043669fab334f0e18b5f71cdd5207a7170bb2ffd3.jpg" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 581, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 554, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 581, + 554, + 616 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 581, + 554, + 616 + ], + "type": "text", + "content": " denotes the anomaly score threshold, " + }, + { + "bbox": [ + 313, + 581, + 554, + 616 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 313, + 581, + 554, + 616 + ], + "type": "text", + "content": " is the anomaly score for the " + }, + { + "bbox": [ + 313, + 581, + 554, + 616 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 581, + 554, + 616 + ], + "type": "text", + "content": "-th object, and " + }, + { + "bbox": [ + 313, + 581, + 554, + 616 + ], + "type": "inline_equation", + "content": "b_i" + }, + { + "bbox": [ + 313, + 581, + 554, + 616 + ], + "type": "text", + "content": " is the corresponding bounding box." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 313, + 624, + 511, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 511, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 511, + 636 + ], + "type": "text", + "content": "3.2. Robust Filtering of Anomalous Boxes" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": "The anomalous boxes detection from Sec. 3.1 contains numerous false positives due to overlapping characteristics between normal and abnormal boxes. This overlap complicates threshold setting and leads to misclassifications. These redundant boxes can significantly degrade segmentation performance by generating inaccurate results and" + } + ] + } + ], + "index": 43 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8692" + } + ] + } + ], + "index": 44 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "overwhelming the model's tracking capabilities, potentially causing the loss of anomalous targets in subsequent frames." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 97, + 295, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 97, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 54, + 97, + 295, + 205 + ], + "type": "text", + "content": "As illustrated in Fig. 3, using potential anomalous boxes as SAM2 prompts at fixed intervals reveals the impact of tracking errors. While red masks indicate true anomalies and green masks show redundant results, the model initially tracks anomalous objects effectively. However, tracking errors accumulate over time, leading to an increasing number of incorrectly tracked objects. This deterioration ultimately results in catastrophic forgetting of true anomalies, reflected in steadily declining Pixel-F1 scores." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "text", + "content": "To enhance robustness and address redundancy, we present the Boxes Robustness Filtering algorithm, which exploits the temporal consistency of true anomalous boxes. Unlike redundant boxes that appear sporadically, true anomalous boxes maintain consistent tracking of the same target. The algorithm, detailed in Alg. 1, operates in three key steps. First, the inheritance step computes the similarity between each bounding box " + }, + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "inline_equation", + "content": "b_{j}" + }, + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "text", + "content": " in the current frame " + }, + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "text", + "content": " and bounding boxes " + }, + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "inline_equation", + "content": "b_{p}" + }, + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "text", + "content": " from the previous " + }, + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "text", + "content": " frames. If the Intersection over Union (IoU) exceeds a threshold " + }, + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "text", + "content": " for at least " + }, + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "text", + "content": " frames, " + }, + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "inline_equation", + "content": "b_{j}" + }, + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "text", + "content": " inherits the label " + }, + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{p}" + }, + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "text", + "content": " from earlier frames, ensuring consistent tracking of previously identified anomalies. Second, the assignment step addresses bounding boxes that fail to inherit a label by comparing them with bounding boxes in the subsequent " + }, + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "text", + "content": " frames. If the IoU exceeds " + }, + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "text", + "content": " for at least " + }, + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "text", + "content": " frames, a new label " + }, + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{j} = \\mathcal{F}(b_{j})" + }, + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "text", + "content": " is assigned, effectively labeling newly emerging anomalies. Finally, the saving step records bounding boxes and their labels every " + }, + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "text", + "content": "-th frame, systematically capturing all identified anomalies. By iteratively applying these steps, the Boxes Robustness Filtering algorithm achieves robust tracking and localization of anomalous objects. Each bounding box and its label are stored as a tuple " + }, + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{box}} = (f_i,b_j,\\mathcal{L}_j)" + }, + { + "bbox": [ + 56, + 207, + 297, + 529 + ], + "type": "text", + "content": ", ensuring spatial and temporal consistency across video frames. This approach reduces redundancy and strengthens anomaly detection accuracy, providing a reliable basis for subsequent segmentation and analysis." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 539, + 251, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 251, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 251, + 552 + ], + "type": "text", + "content": "3.3. Inference of the Segmentation Model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": "Using the tuple " + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{box}} = (f_i, b_j, \\mathcal{L}_j)" + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": " obtained from Sec. 3.2, we extract the center point of each bounding box " + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "inline_equation", + "content": "b_j" + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": ". Specifically, given the coordinates of the box " + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "inline_equation", + "content": "b_j = [x_1, y_1, x_2, y_2]" + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": ", the center point is calculated: " + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_j = \\left(\\frac{x_1 + x_2}{2}, \\frac{y_1 + y_2}{2}\\right)" + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": ". Effectively, these prompts are not saved for every frame but are instead collected at regular intervals, specifically every " + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": "-th frame, as outlined in Sec. 3.2. This ensures efficient storage and processing while still capturing the necessary information for accurate segmentation [23, 25]. For each selected frame, both the center point " + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_j" + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": " and the bounding box " + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "inline_equation", + "content": "b_j" + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": " are stored as prompts. Rather than processing each frame individually, we aggregate prompts from all saved frames " + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\{f_1, f_{1+l}, f_{1+2l}, \\ldots, f_{N_f}\\}" + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": " denotes the saving in" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 460, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 460, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 460, + 84 + ], + "type": "text", + "content": "terval. The aggregated prompts are:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 329, + 94, + 555, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 94, + 555, + 107 + ], + "spans": [ + { + "bbox": [ + 329, + 94, + 555, + 107 + ], + "type": "interline_equation", + "content": "\\mathcal {I} = \\left\\{\\left(\\mathbf {c} _ {j}, b _ {j}, f _ {i}\\right) \\mid i = 1, 1 + l, 1 + 2 l, \\dots , N _ {f} \\right\\} \\tag {4}", + "image_path": "77852676c65a5498fcea5e0a980d1b353a2219a3255c16ff65f808ddcaade969.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 117, + 553, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 117, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 117, + 553, + 152 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 117, + 553, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 117, + 553, + 152 + ], + "type": "text", + "content": " represents the set of prompts that includes the center points, bounding box coordinates " + }, + { + "bbox": [ + 313, + 117, + 553, + 152 + ], + "type": "inline_equation", + "content": "b_{j}" + }, + { + "bbox": [ + 313, + 117, + 553, + 152 + ], + "type": "text", + "content": ", and original frames " + }, + { + "bbox": [ + 313, + 117, + 553, + 152 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 313, + 117, + 553, + 152 + ], + "type": "text", + "content": " for the selected frames." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 153, + 555, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 153, + 555, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 153, + 555, + 247 + ], + "type": "text", + "content": "By inputting this prompt information " + }, + { + "bbox": [ + 313, + 153, + 555, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 153, + 555, + 247 + ], + "type": "text", + "content": " saved at regular intervals into SAM2, which serves as the prompt-based segmentation model " + }, + { + "bbox": [ + 313, + 153, + 555, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 153, + 555, + 247 + ], + "type": "text", + "content": ", we leverage its feature propagation capabilities to efficiently generate pixel-level segmentation results for the entire video. SAM2 propagates the segmentation results from the prompted frames to the remaining frames, producing pixel-level segmentation across the entire video:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 365, + 258, + 555, + 274 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 258, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 365, + 258, + 555, + 274 + ], + "type": "interline_equation", + "content": "\\mathcal {S} = \\mathcal {M} (\\mathcal {I}) = \\left\\{\\mathcal {S} _ {1}, \\mathcal {S} _ {2}, \\dots , \\mathcal {S} _ {N _ {f}} \\right\\} \\tag {5}", + "image_path": "a4678e03626a2a73d17ef9c4eb0ac6665368610e5fe8a730962064e5626a51ef.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 281, + 555, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 342 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 281, + 555, + 342 + ], + "type": "inline_equation", + "content": "S_{i} = [p_{1}, p_{2}, \\ldots, p_{M}]" + }, + { + "bbox": [ + 313, + 281, + 555, + 342 + ], + "type": "text", + "content": " denotes the segmentation result for frame " + }, + { + "bbox": [ + 313, + 281, + 555, + 342 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 313, + 281, + 555, + 342 + ], + "type": "text", + "content": ", where each pixel " + }, + { + "bbox": [ + 313, + 281, + 555, + 342 + ], + "type": "inline_equation", + "content": "p_{j}" + }, + { + "bbox": [ + 313, + 281, + 555, + 342 + ], + "type": "text", + "content": " is either \"normal\" or \"anomalous\". This approach enables the generation of the complete segmentation set " + }, + { + "bbox": [ + 313, + 281, + 555, + 342 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 281, + 555, + 342 + ], + "type": "text", + "content": ", efficiently utilizing contextual information from the prompts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 353, + 395, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 395, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 395, + 366 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 372, + 526, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 526, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 526, + 385 + ], + "type": "text", + "content": "4.1. A New Comprehensive VAD Benchmark" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 389, + 555, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 545 + ], + "type": "text", + "content": "Benchmark Design. We present a novel benchmark framework that integrates both pixel-level and object-level evaluation metrics to advance VAD model assessment. This dual-level approach addresses the limitations of using single evaluation metrics. Pixel-level metrics excel at identifying subtle irregularities through independent pixel evaluation, but struggle to capture broader spatial patterns. Object-level metrics complement this by assessing global characteristics like shape and position, enabling coherent anomaly tracking and recognition. By combining these complementary perspectives, our framework provides a comprehensive evaluation of both fine-grained details and holistic anomaly patterns, particularly valuable for real-world applications." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 545, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 665 + ], + "type": "text", + "content": "Benchmark Datasets. UCSD Ped2 [47] contains 16 training and 12 test videos (240×360 pixels) captured by a fixed overhead camera. The training set consists of normal pedestrian activities, while the test set includes anomalous events such as bikers, skateboarders, and vehicles. ShanghaiTech Campus [30], one of the largest VAD datasets, comprises 330 training and 107 test videos (480×856 pixels). While the training set features normal scenarios, the test set contains various anomalies including robbery, fighting, and unauthorized cycling in pedestrian zones." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": "Benchmark Metrics. For pixel-level evaluation, we employ four metrics: Pixel-AUROC measures discrimination between normal and anomalous pixels across thresholds, Pixel-AP assesses detection precision by balancing false" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8693" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 95, + 555, + 308 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 539, + 83 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 539, + 83 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 539, + 83 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison with anomaly detection models adapted for video anomaly detection on the UCSD Ped2 dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 95, + 555, + 308 + ], + "lines": [ + { + "bbox": [ + 63, + 95, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 63, + 95, + 555, + 308 + ], + "type": "image", + "image_path": "ca47cc552283084344bf2a84c30cf3fe841af99c1dd7705aa375cd0447b02c4e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 71, + 338, + 542, + 460 + ], + "blocks": [ + { + "bbox": [ + 69, + 318, + 541, + 330 + ], + "lines": [ + { + "bbox": [ + 69, + 318, + 541, + 330 + ], + "spans": [ + { + "bbox": [ + 69, + 318, + 541, + 330 + ], + "type": "text", + "content": "Table 1. Quantitative comparison on video anomaly detection with pixel-level and object-level metrics on the UCSD Ped2 dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 71, + 338, + 542, + 460 + ], + "lines": [ + { + "bbox": [ + 71, + 338, + 542, + 460 + ], + "spans": [ + { + "bbox": [ + 71, + 338, + 542, + 460 + ], + "type": "table", + "html": "
MethodPixel-levelObject-level
Pixel-AUROC ↑Pixel-AP ↑Pixel-AUPRO ↑Pixel-F1 ↑RBDC ↑TBDC ↑
AdaCLIP (Zero-shot) [5]51.021.3233.982.615.810.6
AdaCLIP (Fully fine-tuned) [5]53.064.9750.6611.1912.315.5
AnomalyCLIP (Zero-shot) [63]51.6321.2036.345.927.511.2
AnomalyCLIP (Fully fine-tuned) [63]54.2523.7338.597.4813.121.0
DDAD (Fully trained) [37]55.875.6115.122.6718.0113.29
SimpleNet (Fully trained) [33]52.4920.5144.0510.7151.1827.75
DRAEM (Fully trained) [60]69.5830.6335.7810.8944.2670.64
TAO (Partially fine-tuned)75.1150.7872.9764.1283.693.2
", + "image_path": "9ce8179e09099c3d993f6cb6c8dc05eaef753ba741631cd15e6d6eb15d15bc16.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 83, + 509, + 269, + 594 + ], + "blocks": [ + { + "bbox": [ + 55, + 478, + 295, + 501 + ], + "lines": [ + { + "bbox": [ + 55, + 478, + 295, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 295, + 501 + ], + "type": "text", + "content": "Table 2. Quantitative comparison on video anomaly detection with RBDC and TBDC metrics on the UCSD Ped2 dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 509, + 269, + 594 + ], + "lines": [ + { + "bbox": [ + 83, + 509, + 269, + 594 + ], + "spans": [ + { + "bbox": [ + 83, + 509, + 269, + 594 + ], + "type": "table", + "html": "
MethodRBDC ↑TBDC ↑
OCAD [19]52.772.8
SS-VAD [40]62.580.5
SiVL [41]7489.3
AED-SSMTL [10]72.891.2
TAO (Ours)83.693.2
", + "image_path": "c19c8f93b5ffbdad881219fcc7054e37868c49e658c49616437688764a31b4b3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "positives and negatives, Pixel-AUPRO quantifies segmentation accuracy through region overlap, and Pixel-F1 provides an integrated precision-recall measure. For object-level assessment, we utilize Region-Based Detection Criterion (RBDC) and Track-Based Detection Criterion (TBDC). RBDC measures spatial accuracy using Intersection over Union (IoU) with threshold " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": ", while TBDC evaluates temporal consistency in tracking anomalous regions across" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 342, + 510, + 528, + 608 + ], + "blocks": [ + { + "bbox": [ + 313, + 478, + 555, + 501 + ], + "lines": [ + { + "bbox": [ + 313, + 478, + 555, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 555, + 501 + ], + "type": "text", + "content": "Table 3. Quantitative comparison on video anomaly detection with RBDC and TBDC metrics on the ShanghaiTech Campus dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 342, + 510, + 528, + 608 + ], + "lines": [ + { + "bbox": [ + 342, + 510, + 528, + 608 + ], + "spans": [ + { + "bbox": [ + 342, + 510, + 528, + 608 + ], + "type": "table", + "html": "
MethodRBDC ↑TBDC ↑
OCAD [19]20.744.5
BAF-AT [9]41.378.8
AED-SSMTL [10]43.284.1
HF2VAD [31]45.484.5
STPT [38]51.684.6
TAO (Ours)62.185.4
", + "image_path": "fefe2174b3d0a8f56ca09e83f6b73f5d78410db33c69988bda16b7e8fbdd86b5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 627, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 627, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 627, + 555, + 651 + ], + "type": "text", + "content": "frames. Together, these metrics provide comprehensive evaluation of both spatial precision and temporal coherence." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 659, + 448, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 448, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 448, + 672 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": "Our framework builds upon SAM2 [42], utilizing an object-level anomaly detection algorithm [43] to extract anomalous bounding boxes as segmentation prompts. We em" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "8694" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 395 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 395 + ], + "type": "text", + "content": "ploy SAM2 [42] as the prompt-based segmentation model, which integrates these prompts with original frames for precise anomaly segmentation. Our implementation adopts a partial fine-tuning strategy, where only the object-level VAD algorithm is fine-tuned on target datasets, while SAM2 retains its pre-trained weights for inference. In the bounding box threshold filtering stage, we assign anomaly scores based on pose and depth features, with thresholds set to " + }, + { + "bbox": [ + 57, + 72, + 294, + 395 + ], + "type": "inline_equation", + "content": "\\tau = 1.5" + }, + { + "bbox": [ + 57, + 72, + 294, + 395 + ], + "type": "text", + "content": " for UCSD Ped2 and " + }, + { + "bbox": [ + 57, + 72, + 294, + 395 + ], + "type": "inline_equation", + "content": "\\tau = 1.6" + }, + { + "bbox": [ + 57, + 72, + 294, + 395 + ], + "type": "text", + "content": " for ShanghaiTech Campus. The subsequent robustness filtering stage employs a tracking window of " + }, + { + "bbox": [ + 57, + 72, + 294, + 395 + ], + "type": "inline_equation", + "content": "k = 5" + }, + { + "bbox": [ + 57, + 72, + 294, + 395 + ], + "type": "text", + "content": ", frame match threshold of " + }, + { + "bbox": [ + 57, + 72, + 294, + 395 + ], + "type": "inline_equation", + "content": "m = 3" + }, + { + "bbox": [ + 57, + 72, + 294, + 395 + ], + "type": "text", + "content": ", and box overlap threshold of " + }, + { + "bbox": [ + 57, + 72, + 294, + 395 + ], + "type": "inline_equation", + "content": "h = 0.2" + }, + { + "bbox": [ + 57, + 72, + 294, + 395 + ], + "type": "text", + "content": ", with save intervals of " + }, + { + "bbox": [ + 57, + 72, + 294, + 395 + ], + "type": "inline_equation", + "content": "l = 5" + }, + { + "bbox": [ + 57, + 72, + 294, + 395 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 57, + 72, + 294, + 395 + ], + "type": "inline_equation", + "content": "l = 15" + }, + { + "bbox": [ + 57, + 72, + 294, + 395 + ], + "type": "text", + "content": " for UCSD Ped2 and ShanghaiTech Campus respectively. For segmentation, we utilize SAM2 with hiera base-plus weights on an NVIDIA RTX 4090 GPU. To ensure robust evaluation, overlapping boxes of the same anomaly are merged, and segmentation probability maps are uniformly binarized. For object-level metrics, we derive bounding boxes from segmentation extremal coordinates. Given anomaly pixels " + }, + { + "bbox": [ + 57, + 72, + 294, + 395 + ], + "type": "inline_equation", + "content": "P = (x_{1},y_{1}),(x_{2},y_{2}),\\ldots ,(x_{N},y_{N})" + }, + { + "bbox": [ + 57, + 72, + 294, + 395 + ], + "type": "text", + "content": ", we compute the box coordinates as: " + }, + { + "bbox": [ + 57, + 72, + 294, + 395 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{min}} = \\min_{(x_i,y_i)\\in P}x_i" + }, + { + "bbox": [ + 57, + 72, + 294, + 395 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 57, + 72, + 294, + 395 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{max}} = \\max_{(x_i,y_i)\\in P}x_i" + }, + { + "bbox": [ + 57, + 72, + 294, + 395 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 57, + 72, + 294, + 395 + ], + "type": "inline_equation", + "content": "y_{\\mathrm{min}} = \\min_{(x_i,y_i)\\in P}y_i" + }, + { + "bbox": [ + 57, + 72, + 294, + 395 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 57, + 72, + 294, + 395 + ], + "type": "inline_equation", + "content": "y_{\\mathrm{max}} = \\max_{(x_i,y_i)\\in P}y_i" + }, + { + "bbox": [ + 57, + 72, + 294, + 395 + ], + "type": "text", + "content": ", where the resulting bounding box " + }, + { + "bbox": [ + 57, + 72, + 294, + 395 + ], + "type": "inline_equation", + "content": "(x_{\\mathrm{min}},y_{\\mathrm{min}},x_{\\mathrm{max}},y_{\\mathrm{max}})" + }, + { + "bbox": [ + 57, + 72, + 294, + 395 + ], + "type": "text", + "content": " is used for computing RBDC and TBDC metrics, providing robust spatial evaluation at the object level." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 407, + 276, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 407, + 276, + 418 + ], + "spans": [ + { + "bbox": [ + 57, + 407, + 276, + 418 + ], + "type": "text", + "content": "4.3. Comparison to Anomaly Detection Models" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 426, + 294, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 426, + 294, + 664 + ], + "spans": [ + { + "bbox": [ + 57, + 426, + 294, + 664 + ], + "type": "text", + "content": "Given the lack of pixel-level segmentation capabilities in current video anomaly detection models, we benchmark several state-of-the-art image anomaly detection algorithms, including AdaCLIP [5], AnomalyCLIP [63], DRAEM [60], DDAD [37], and SimpleNet [33], to evaluate both pixel-level and object-level metrics on the UCSD Ped2 dataset. AdaCLIP and AnomalyCLIP were selected for their pre-trained vision-language architectures, which closely align with SAM2's design and demonstrate strong zero-shot capabilities. To ensure a fair comparison, we assess both AdaCLIP and AnomalyCLIP under zero-shot and fully fine-tuned settings. In contrast, DRAEM, DDAD, and SimpleNet, as fully trained models, provide robust baselines for pixel-level and object-level anomaly detection, representing diverse methodological approaches. For anomaly detection, video streams are processed frame-by-frame through each algorithm, generating heatmaps that identify potential anomalies. Pixel-level segmentation is achieved by applying a threshold to these heatmaps, allowing precise localization of anomalous regions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 666, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 666, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 666, + 294, + 712 + ], + "type": "text", + "content": "As shown in Tab.1, the experimental results clearly demonstrate the advantages of our model in both pixel-level and object-level anomaly detection. While most benchmarked models struggle to balance these two aspects, of" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 316, + 72, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 251 + ], + "type": "text", + "content": "ten excelling in one while underperforming in the other, our model achieves robust performance across both dimensions. This balance reflects its strong and versatile design, enabling precise pixel-level segmentation and maintaining spatial-temporal consistency at the object level. Moreover, our partial fine-tuning approach effectively leverages pretrained features, minimizing the need for extensive dataset-specific training. In contrast, fully trained models such as DRAEM, DDAD, and SimpleNet, as well as vision-language models like AdaCLIP and AnomalyCLIP, fail to achieve comparable results. Fig. 4 further illustrates the strengths of our method. The segmentation results produced by our model are more complete and detailed, accurately delineating object boundaries while preserving the integrity of anomalous objects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 316, + 255, + 553, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 255, + 553, + 278 + ], + "spans": [ + { + "bbox": [ + 316, + 255, + 553, + 278 + ], + "type": "text", + "content": "4.4. Comparison to Conventional Video Anomaly Detection Models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 316, + 289, + 553, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 289, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 289, + 553, + 431 + ], + "type": "text", + "content": "To assess the performance of our proposed method, we conducted comprehensive experiments on two widely used datasets, UCSD Ped2 and ShanghaiTech Campus. Our approach was benchmarked against several state-of-the-art video anomaly detection models, including SS-VAD [40], SiVL [41], AED-SSMTL [10], HF2VAD [31], BAF-AT [9], and STPT [38]. Due to the absence of fine-grained segmentation capabilities in traditional methods, the evaluation was focused on object-centric anomaly detection and limited to object-level metrics, TBDC [40] and RBDC [40]. These metrics provide a robust and meaningful basis for comparing the effectiveness of our model with existing approaches." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 316, + 435, + 553, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 435, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 435, + 553, + 601 + ], + "type": "text", + "content": "Our method achieves state-of-the-art performance on both the UCSD Ped2 and ShanghaiTech Campus datasets. On UCSD Ped2, it outperforms existing methods in TBDC and RBDC metrics, demonstrating precise tracking and localization of anomalous objects across video frames. On the more challenging ShanghaiTech Campus dataset, characterized by complex backgrounds and subtle anomalies, our method consistently surpasses baselines by delivering balanced performance across all evaluation metrics. Unlike existing approaches, which often excel in isolated metrics but fail to generalize, our model effectively integrates object-centric detection and fine-grained segmentation, enabling robust anomaly localization and tracking under diverse real-world conditions, as shown in Tabs. 2 and 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 605, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 605, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 605, + 553, + 712 + ], + "type": "text", + "content": "To conclude, our method establishes a new standard in video anomaly detection by combining object-centric detection with fine-grained segmentation. This integration enables precise localization and robust tracking of anomalous objects, addressing the limitations of traditional methods that struggle with balanced performance across metrics. Our approach demonstrates versatility and effectiveness across diverse datasets, highlighting its potential for real-world applications." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "text", + "content": "8695" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 91, + 542, + 182 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 544, + 82 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 544, + 82 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 544, + 82 + ], + "type": "text", + "content": "Table 4. Ablation study on proposed key components, evaluated with pixel-level and object-level metrics on the UCSD Ped2 dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 91, + 542, + 182 + ], + "lines": [ + { + "bbox": [ + 69, + 91, + 542, + 182 + ], + "spans": [ + { + "bbox": [ + 69, + 91, + 542, + 182 + ], + "type": "table", + "html": "
Video Track ModeBox Robust Filtering AlgorithmPixel-levelObject-level
Pixel-AUROC ↑Pixel-AP ↑Pixel-AUPRO ↑Pixel-F1 ↑RBDC ↑TBDC ↑
××41.7225.3040.5532.9941.071.6
×✓39.0622.8038.0930.5639.371.7
✓×68.9038.4767.2352.1770.876.9
✓✓75.1150.7872.9764.1283.693.2
", + "image_path": "b3857c62f07e8040ec03a99059073408460f68da0319f50de03db886c411a6a8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 62, + 206, + 282, + 345 + ], + "blocks": [ + { + "bbox": [ + 62, + 206, + 282, + 345 + ], + "lines": [ + { + "bbox": [ + 62, + 206, + 282, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 206, + 282, + 345 + ], + "type": "image", + "image_path": "bb0f3153cbed6e9a20805bb3acac1a71f14e06be1d248bafe9dc1aa92d06c558.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 357, + 295, + 391 + ], + "lines": [ + { + "bbox": [ + 55, + 357, + 295, + 391 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 295, + 391 + ], + "type": "text", + "content": "Figure 5. Sensitivity Analysis of Anomaly Threshold " + }, + { + "bbox": [ + 55, + 357, + 295, + 391 + ], + "type": "inline_equation", + "content": "(\\tau)" + }, + { + "bbox": [ + 55, + 357, + 295, + 391 + ], + "type": "text", + "content": ". We evaluate performance under varying threshold on ShanghaiTech Campus dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 62, + 407, + 290, + 524 + ], + "blocks": [ + { + "bbox": [ + 62, + 407, + 290, + 524 + ], + "lines": [ + { + "bbox": [ + 62, + 407, + 290, + 524 + ], + "spans": [ + { + "bbox": [ + 62, + 407, + 290, + 524 + ], + "type": "image", + "image_path": "9469c4c2b498ef09a28583c1d54f6ed4cf3d9f0c9b5d36e37c227dab2b3dccf7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "type": "text", + "content": "Figure 6. Robust Testing of SAM2 Backbones. We compare the performance across different SAM2 model variants." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 576, + 200, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 200, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 200, + 590 + ], + "type": "text", + "content": "4.5. Further Empirical Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Ablation Study. We conduct ablation experiments on the UCSD Ped2 dataset to evaluate two components: video tracking mode and box robust filtering. Video tracking mode, a SAM2 component, enables temporal analysis, while disabling it reduces the model to SAM, performing frame-by-frame segmentation. Disabling box robust filtering uses raw anomaly-thresholded boxes as prompts. Tab. 4 shows the effects of each component on pixel- and object-level metrics. The baseline model, with both components disabled, performs the worst. Enabling box robust filter" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 202, + 553, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 202, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 553, + 274 + ], + "type": "text", + "content": "ing alone yields minimal improvement, while activating video tracking mode significantly boosts pixel-level accuracy. The best performance is achieved with both components enabled, demonstrating their synergistic effect. These results confirm the necessity of both temporal tracking and robust filtering in our framework." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 276, + 555, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 555, + 443 + ], + "type": "text", + "content": "Sensitivity Analysis. We conduct a sensitivity analysis to assess how variations in the anomaly threshold " + }, + { + "bbox": [ + 313, + 276, + 555, + 443 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 276, + 555, + 443 + ], + "type": "text", + "content": " affect our framework's performance. Fig. 5 shows the impact of " + }, + { + "bbox": [ + 313, + 276, + 555, + 443 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 276, + 555, + 443 + ], + "type": "text", + "content": " on four metrics—Pixel-AUROC, Pixel-AP, TBDC, and RBDC—using the ShanghaiTech Campus dataset. As " + }, + { + "bbox": [ + 313, + 276, + 555, + 443 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 276, + 555, + 443 + ], + "type": "text", + "content": " ranges from 1.0 to 1.9, the framework performs stably, with optimal results at " + }, + { + "bbox": [ + 313, + 276, + 555, + 443 + ], + "type": "inline_equation", + "content": "\\tau = 1.6" + }, + { + "bbox": [ + 313, + 276, + 555, + 443 + ], + "type": "text", + "content": ". TBDC is most sensitive, peaking at " + }, + { + "bbox": [ + 313, + 276, + 555, + 443 + ], + "type": "inline_equation", + "content": "85\\%" + }, + { + "bbox": [ + 313, + 276, + 555, + 443 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 313, + 276, + 555, + 443 + ], + "type": "inline_equation", + "content": "\\tau = 1.6" + }, + { + "bbox": [ + 313, + 276, + 555, + 443 + ], + "type": "text", + "content": ", while RBDC stays around " + }, + { + "bbox": [ + 313, + 276, + 555, + 443 + ], + "type": "inline_equation", + "content": "62\\%" + }, + { + "bbox": [ + 313, + 276, + 555, + 443 + ], + "type": "text", + "content": ". Pixel-level metrics, Pixel-AUROC and Pixel-AP, also show steady trends, reflecting the framework's balance of fine-grained anomaly detection and object consistency. This analysis highlights the robustness and adaptability of our framework across threshold variations, ensuring reliable detection in diverse scenarios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 445, + 556, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 556, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 556, + 590 + ], + "type": "text", + "content": "Robustness Testing. We evaluate our framework's robustness through extensive testing across different SAM2 backbone architectures (Fig. 6). Our analysis covers models from the lightweight Tiny variant to the full-scale Large model, with Base+ consistently yielding the best results. This stability suggests that the framework's effectiveness is not dependent on backbone capacity. While larger models show slight improvements in pixel-level metrics, performance remains strong across all variants. These results validate the framework's reliability and robustness, demonstrating its suitability for deployment in a wide range of scenarios, irrespective of computational constraints." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 607, + 388, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 607, + 388, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 607, + 388, + 620 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "This paper presents TAO, a framework integrating coarse object detection with fine-grained anomaly localization in videos. Leveraging SAM2 and our dual-level evaluation, it enables precise segmentation and robust tracking. Experiments on UCSD Ped2 and ShanghaiTech datasets demonstrate state-of-the-art performance in pixel- and object-level metrics, advancing video anomaly detection." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8696" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 175 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 82172033, Grant U19B2031, Grant 61971369, Grant 52105126, Grant 82272071, and Grant 62271430; and in part by the Dreams Foundation of Jianghuai Advance Technology Center and in part by the Open Fund of the National Key Laboratory of Infrared Detection Technologies." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 115, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 115, + 197 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 115, + 197 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 205, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 205, + 296, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 205, + 296, + 248 + ], + "spans": [ + { + "bbox": [ + 61, + 205, + 296, + 248 + ], + "type": "text", + "content": "[1] Sunghyun Ahn, Youngwan Jo, Kijung Lee, and Sanghyun Park. Videopatchcore: An effective method to memorize normality for video anomaly detection. arXiv preprint arXiv:2409.16225, 2024. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 250, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 250, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 61, + 250, + 296, + 293 + ], + "type": "text", + "content": "[2] Jinbin Bai, Zhen Dong, Aosong Feng, Xiao Zhang, Tian Ye, Kaicheng Zhou, and Mike Zheng Shou. Integrating view conditions for image synthesis. arXiv preprint arXiv:2310.16002, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 293, + 296, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 293, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 62, + 293, + 296, + 337 + ], + "type": "text", + "content": "[3] Jinbin Bai, Wei Chow, Ling Yang, Xiangtai Li, Juncheng Li, Hanwang Zhang, and Shuicheng Yan. Humanedit: A high-quality human-rewarded dataset for instruction-based image editing. arXiv preprint arXiv:2412.04280, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 338, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 338, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 338, + 296, + 392 + ], + "type": "text", + "content": "[4] Jinbin Bai, Tian Ye, Wei Chow, Enxin Song, Qing-Guo Chen, Xiangtai Li, Zhen Dong, Lei Zhu, and Shuicheng Yan. Meissonic: Revitalizing Masked Generative Transformers for Efficient High-Resolution Text-to-Image Synthesis. arXiv preprint arXiv:2410.08261, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 393, + 296, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 393, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 393, + 296, + 446 + ], + "type": "text", + "content": "[5] Yunkang Cao, Jiangning Zhang, Luca Frittoli, Yuqi Cheng, Weiming Shen, and Giacomo Boracchi. Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly detection. In European Conference on Computer Vision, pages 55-72. Springer, 2025. 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 449, + 296, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 296, + 513 + ], + "type": "text", + "content": "[6] Zhiyuan Ding, Qi Dong, Haote Xu, Chenxin Li, Xinghao Ding, and Yue Huang. Unsupervised anomaly segmentation for brain lesions using dual semantic-manifold reconstruction. In International Conference on Neural Information Processing, pages 133-144. Springer International Publishing Cham, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 514, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 294, + 535 + ], + "type": "text", + "content": "[7] Keval Doshi and Yasin Yilmaz. Any-shot sequential anomaly detection in surveillance videos. 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 536, + 296, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 296, + 557 + ], + "type": "text", + "content": "[8] Keval Doshi and Yasin Yilmaz. Continual learning for anomaly detection in surveillance videos. 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 559, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 296, + 624 + ], + "type": "text", + "content": "[9] Mariana-Iuliana Georgescu, RaduTudor Ionescu, Fahad-Shahbaz Khan, Marius Popescu, and Mubarak Shah. A background-agnostic framework with adversarial training for abnormal event detection in video. IEEE Transactions on Pattern Analysis and Machine Intelligence, IEEE Transactions on Pattern Analysis and Machine Intelligence. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 625, + 296, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 296, + 691 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 296, + 691 + ], + "type": "text", + "content": "[10] Mariana-Iuliana Georgescu, Antonio Barbalau, Radu Tudor Ionescu, Fahad Shahbaz Khan, Marius Popescu, and Mubarak Shah. Anomaly detection in video via self-supervised and multi-task learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12742-12752, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "type": "text", + "content": "[11] Mahmudul Hasan, Jonghyun Choi, Jan Neumann, Amit K. Roy-Chowdhury, and Larry S. Davis. Learning temporal" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 555, + 106 + ], + "type": "text", + "content": "regularity in video sequences. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 554, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 554, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 554, + 150 + ], + "type": "text", + "content": "[12] Chunming He, Kai Li, Guoxia Xu, Yulun Zhang, Runze Hu, Zhenhua Guo, and Xiu Li. Degradation-resistant unfolding network for heterogeneous image fusion. In ICCV, pages 12611-12621, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 554, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 554, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 554, + 206 + ], + "type": "text", + "content": "[13] Chunming He, Kai Li, Yachao Zhang, Guoxia Xu, Longxiang Tang, Yulun Zhang, Zhenhua Guo, and Xiu Li. Weakly-supervised concealed object segmentation with sam-based pseudo labeling and multi-scale feature grouping. NeurIPS, 36, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 209, + 554, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 554, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 554, + 252 + ], + "type": "text", + "content": "[14] Chunming He, Kai Li, Yachao Zhang, Yulun Zhang, Zhenhua Guo, Xiu Li, Martin Danelljan, and Fisher Yu. Strategic preys make acute predators: Enhancing camouflaged object detectors by generating camouflaged objects. ICLR, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 254, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 554, + 297 + ], + "type": "text", + "content": "[15] Chunming He, Chengyu Fang, Yulun Zhang, Kai Li, Longxiang Tang, Chenyu You, Fengyang Xiao, Zhenhua Guo, and Xiu Li. Reti-diff: Illumination degradation image restoration with retinae-based latent diffusion model. *ICLR*, 2025. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 299, + 554, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 554, + 342 + ], + "type": "text", + "content": "[16] Chunming He, Yuqi Shen, Chengyu Fang, Fengyang Xiao, Longxiang Tang, Yulun Zhang, Wangmeng Zuo, Zhenhua Guo, and Xiu Li. Diffusion models in low-level vision: A survey. TPAMI, 2025. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 344, + 554, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 554, + 398 + ], + "type": "text", + "content": "[17] Chunming He, Rihan Zhang, Fengyang Xiao, Chenyu Fang, Longxiang Tang, Yulun Zhang, Linghe Kong, Deng-Ping Fan, Kai Li, and Sina Farsiu. Run: Reversible unfolding network for concealed object segmentation. arXiv preprint arXiv:2501.18783, 2025. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 399, + 554, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 554, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 554, + 443 + ], + "type": "text", + "content": "[18] Ryota Hinami, Tao Mei, and Shin'ichi Satoh. Joint detection and recounting of abnormal events by learning deep generic knowledge. In 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 445, + 554, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 554, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 554, + 499 + ], + "type": "text", + "content": "[19] Radu Tudor Ionescu, Fahad Shahbaz Khan, Mariana-Iuliana Georgescu, and Ling Shao. Object-centric auto-encoders and dummy anomalies for abnormal event detection in video. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 501, + 554, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 554, + 545 + ], + "type": "text", + "content": "[20] Hamza Karim, Keval Doshi, and Yasin Yilmaz. Real-time weakly supervised video anomaly detection. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 6848-6856, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 546, + 554, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 546, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 546, + 554, + 601 + ], + "type": "text", + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 602, + 554, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 554, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 554, + 645 + ], + "type": "text", + "content": "[22] Chenxin Li, Yunlong Zhang, Jiongcheng Li, Yue Huang, and Xinghao Ding. Unsupervised anomaly segmentation using image-semantic cycle translation. arXiv preprint arXiv:2103.09094, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 712 + ], + "type": "text", + "content": "[23] Chenxin Li, Yunlong Zhang, Zhehan Liang, Wenao Ma, Yue Huang, and Xinghao Ding. Consistent posterior distributions under vessel-mixing: a regularization for cross-domain retinal artery/vein classification. In 2021 IEEE International Conference on Image Processing (ICIP), pages 61-65. IEEE, 2021. 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8697" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[24] Chenxin Li, Wenao Ma, Liyan Sun, Xinghao Ding, Yue Huang, Guisheng Wang, and Yizhou Yu. Hierarchical deep network with uncertainty-aware semi-supervised learning for vessel segmentation. Neural Computing and Applications, pages 1-14, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "type": "text", + "content": "[25] Chenxin Li, Xinyu Liu, Wuyang Li, Cheng Wang, Hengyu Liu, Yifan Liu, Zhen Chen, and Yixuan Yuan. U-kan makes strong backbone for medical image segmentation and generation. arXiv preprint arXiv:2406.02918, 2024. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 294, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 226 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 226 + ], + "type": "text", + "content": "[26] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 282 + ], + "type": "text", + "content": "[27] Huangxing Lin, Yunlong Lin, Jingyuan Xia, Linyu Fan, Feifei Li, Yingying Wang, and Xinghao Ding. Fusion2void: Unsupervised multi-focus image fusion based on image inpainting. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 283, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 337 + ], + "type": "text", + "content": "[28] Yunlong Lin, Tian Ye, Sixiang Chen, Zhenqi Fu, Yingying Wang, Wenhao Chai, Zhaohu Xing, Lei Zhu, and Xinghao Ding. Aglldiff: Guiding diffusion models towards unsupervised training-free real-world low-light image enhancement. arXiv preprint arXiv:2407.14900, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 294, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 371 + ], + "type": "text", + "content": "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "type": "text", + "content": "[30] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection-a new baseline. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6536-6545, 2018. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 480 + ], + "type": "text", + "content": "[31] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13588-13597, 2021. 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 481, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 294, + 536 + ], + "type": "text", + "content": "[32] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 536, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 536, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 536, + 294, + 591 + ], + "type": "text", + "content": "[33] Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang. Simplenet: A simple network for image anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20402-20411, 2023. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "text", + "content": "[34] Shao-Yuan Lo, Poojan Oza, and Vishal M. Patel. Adversarially robust one-class novelty detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, page 1-12, 2022. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 680 + ], + "type": "text", + "content": "[35] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[36] Jakub Micorek, Horst Possegger, Dominik Narnhofer, Horst Bischof, and Mateusz Kozinski. MULDE: Multiscale Log-Density Estimation via Denoising Score Matching for Video" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "type": "text", + "content": "Anomaly Detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18868-18877, 2024. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 106, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 553, + 139 + ], + "type": "text", + "content": "[37] Arian Mousakhan, Thomas Brox, and Jawad Tayyub. Anomaly detection with conditioned denoising diffusion models. arXiv preprint arXiv:2305.15956, 2023. 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "type": "text", + "content": "[38] Yassine Naji, Aleksandr Setkov, Angelique Loesch, Michèle Gouiffes, and Romaric Audigier. Spatio-temporal predictive tasks for abnormal event detection in videos. 2022. 1, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 172, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 553, + 217 + ], + "type": "text", + "content": "[39] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "type": "text", + "content": "[40] Bharathkumar Ramachandra and Michael Jones. Street scene: A new dataset and evaluation protocol for video anomaly detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2569-2578, 2020. 1, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 271, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 271, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 271, + 553, + 326 + ], + "type": "text", + "content": "[41] Bharathkumar Ramachandra, Michael J. Jones, and Ranga Raju Vatsavai. Learning a distance function with a siamese network to localize anomalies in videos. In 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "text", + "content": "[42] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 382, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 382, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 382, + 553, + 414 + ], + "type": "text", + "content": "[43] Tal Reiss and Yedid Hoshen. Attribute-based representations for accurate and interpretable video anomaly detection. 2022. 1, 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 415, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 448 + ], + "type": "text", + "content": "[44] Chenrui Shi, Che Sun, Yuwei Wu, and Yunde Jia. Video anomaly detection via sequentially learning multiple pretext tasks. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 449, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 503 + ], + "type": "text", + "content": "[45] Liyan Sun, Chenxin Li, Xinghao Ding, Yue Huang, Zhong Chen, Guisheng Wang, Yizhou Yu, and John Paisley. Few-shot medical image segmentation using a global correlation network with discriminative embedding. Computers in biology and medicine, 140:105067, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 503, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 503, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 503, + 553, + 558 + ], + "type": "text", + "content": "[46] Yu Tian, Guansong Pang, Yuanhong Chen, Rajvinder Singh, Johan W Verjans, and Gustavo Carneiro. Weakly-supervised video anomaly detection with robust temporal feature magnitude learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4975-4986, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "type": "text", + "content": "[47] Shu Wang and Zhenjiang Miao. Anomaly detection in crowd scene. In IEEE 10th International Conference on Signal Processing Proceedings, pages 1220-1223. IEEE, 2010. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 592, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 657 + ], + "type": "text", + "content": "[48] Yingying Wang, Yunlong Lin, Ge Meng, Zhenqi Fu, Yuhang Dong, Linyu Fan, Hedeng Yu, Xinghao Ding, and Yue Huang. Learning high-frequency feature enhancement and alignment for pan-sharpening. In Proceedings of the 31st ACM International Conference on Multimedia, pages 358-367, 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[49] Yingying Wang, Yunlong Lin, Xuanhua He, Hui Zheng, Keyu Yan, Linyu Fan, Yue Huang, and Xinghao Ding. Learning diffusion high-quality priors for pan-sharpening: A two-stage approach with time-aware adapter fine-tuning. IEEE Transactions on Geoscience and Remote Sensing, 2025. 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8698" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[50] Jhih-Ciang Wu, He-Yen Hsieh, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng-Luh Liu. Self-supervised sparse representation for video anomaly detection. In European Conference on Computer Vision, pages 729-745. Springer, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "text", + "content": "[51] Peng Wu, Xuerong Zhou, Guansong Pang, Lingru Zhou, Qingsen Yan, Peng Wang, and Yanning Zhang. Vadclip: Adapting vision-language models for weakly supervised video anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6074-6082, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 216 + ], + "type": "text", + "content": "[52] Fengyang Xiao, Sujie Hu, Yuqi Shen, Chengyu Fang, Jina Huang, Chunming He, Longxiang Tang, Ziyun Yang, and Xiu Li. A survey of camouflaged object detection and beyond. CAIAIR, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "type": "text", + "content": "[53] Zhaohu Xing, Lequan Yu, Liang Wan, Tong Han, and Lei Zhu. Nestedformer: Nested modality-aware transformer for brain tumor segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 140-150. Springer, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 294, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 313 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 313 + ], + "type": "text", + "content": "[54] Zhaohu Xing, Liang Wan, Huazhu Fu, Guang Yang, and Lei Zhu. Diff-unet: A diffusion embedded network for volumetric segmentation. arXiv preprint arXiv:2303.10326, 2023.2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 380 + ], + "type": "text", + "content": "[55] Zhaohu Xing, Sicheng Yang, Sixiang Chen, Tian Ye, Yijun Yang, Jing Qin, and Lei Zhu. Cross-conditioned diffusion model for medical image to image translation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 201-211. Springer, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 382, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 294, + 437 + ], + "type": "text", + "content": "[56] Zhaohu Xing, Tian Ye, Yijun Yang, Guang Liu, and Lei Zhu. Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 578-588. Springer, 2024." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "text", + "content": "[57] Yuchen Yang, Kwonjoon Lee, Behzad Dariush, Yinzhi Cao, and Shao-Yuan Lo. Follow the rules: Reasoning for video anomaly detection with large language models. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 492, + 294, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 294, + 547 + ], + "type": "text", + "content": "[58] Guang Yu, Siqi Wang, Zhiping Cai, En Zhu, Chuanfu Xu, Jianping Yin, and Marius Kloft. Cloze test helps: Effective video anomaly detection via learning to complete video events. In Proceedings of the 28th ACM International Conference on Multimedia, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "text", + "content": "[59] Luca Zanella, Willi Menapace, Massimiliano Mancini, Yiming Wang, and Elisa Ricci. Harnessing large language models for training-free video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18527-18536, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 603, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 657 + ], + "type": "text", + "content": "[60] Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8330-8339, 2021. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 690 + ], + "type": "text", + "content": "[61] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[62] Yunlong Zhang, Chenxin Li, Xin Lin, Liyan Sun, Yihong Zhuang, Yue Huang, Xinghao Ding, Xiaqing Liu, and" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 228 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 139 + ], + "type": "text", + "content": "Yizhou Yu. Generator versus segmentor: Pseudo-healthy synthesis. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part VI 24, pages 150–160. Springer International Publishing, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "type": "text", + "content": "[63] Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. Anomalyclip: Object-agnostic prompt learning for zero-shot anomaly detection. arXiv preprint arXiv:2310.18961, 2023. 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 185, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 228 + ], + "type": "text", + "content": "[64] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8699" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Track4Gen_ Teaching Video Diffusion Models to Track Points Improves Video Generation/52b14d4e-8da8-4b5e-a32c-278d64dd36e6_content_list.json b/2025/Track4Gen_ Teaching Video Diffusion Models to Track Points Improves Video Generation/52b14d4e-8da8-4b5e-a32c-278d64dd36e6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b8b6112bc6fa63d7a3bbcbc7f86f085a638486f5 --- /dev/null +++ b/2025/Track4Gen_ Teaching Video Diffusion Models to Track Points Improves Video Generation/52b14d4e-8da8-4b5e-a32c-278d64dd36e6_content_list.json @@ -0,0 +1,1598 @@ +[ + { + "type": "text", + "text": "Track4Gen: Teaching Video Diffusion Models to Track Points Improves Video Generation", + "text_level": 1, + "bbox": [ + 187, + 130, + 810, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hyeonho Jeong $^{1,2,*}$ Chun-Hao P. Huang $^{1}$ Jong Chul Ye $^{2}$ Niloy J. Mitra $^{1,3}$ Duygu Ceylan $^{1}$", + "bbox": [ + 104, + 202, + 885, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Adobe Research $^{2}$ KAIST $^{3}$ University College London", + "bbox": [ + 241, + 227, + 754, + 248 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 282, + 326, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While recent foundational video generators produce visually rich output, they still struggle with appearance drift, where objects gradually degrade or change inconsistently across frames, breaking visual coherence. We hypothesize that this is because there is no explicit supervision in terms of spatial tracking at the feature level. We propose Track4Gen, a spatially aware video generator that combines video diffusion loss with point tracking across frames, providing enhanced spatial supervision on the diffusion features. Track4Gen merges the video generation and point tracking tasks into a single network by making minimal changes to existing video generation architectures. Using Stable Video Diffusion [4] as a backbone, Track4Gen demonstrates that it is possible to unify video generation and point tracking, which are typically handled as separate tasks. Our extensive evaluations show that Track4Gen effectively reduces appearance drift, resulting in temporally stable and visually coherent video generation. Project page: hyeonho99.github.io/track4gen", + "bbox": [ + 89, + 313, + 483, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 627, + 220, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion-based video generators [4, 6, 47] are making rapid strides in creating temporally consistent and visually rich video content. This progress marks a significant shift, as the unification of generation and control has the potential to transform the traditional workflow of first capturing and then digitally editing video.", + "bbox": [ + 89, + 652, + 482, + 743 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite impressive capabilities, video generators often suffer from appearance drift, where visual elements gradually change, mutate, or degrade over time, causing inconsistencies in the objects. For example, in Fig. 1, we observe the horns of the cow distorting and morphing unrealistically over time, breaking the plausibility of the generated content. This is in striking contrast to humans, who develop a sense of appearance constancy as early as infancy through observation and interaction with the world [72].", + "bbox": [ + 89, + 744, + 482, + 880 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a0e63e70d3633171971306d5156655df576d3d1b100c6d09147b576a1b71006a.jpg", + "image_caption": [ + "Stable Video Diffusion", + "Figure 1. Motivation. Videos generated by Stable Video Diffusion [4] suffer from appearance drift, while those from our method, Track4Gen, are free from such appearance inconsistency issues." + ], + "image_footnote": [], + "bbox": [ + 516, + 292, + 709, + 544 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6c2318d3dfdc5acd100c9a18f08f1d83039ad5d056e1fa375206c8c9a1f94514.jpg", + "image_caption": [ + "Track4Gen (Ours)" + ], + "image_footnote": [], + "bbox": [ + 710, + 292, + 903, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unfortunately, appearance drift remains a persistent issue in current video models, even with increased training data and more advanced architectures. We speculate that this limitation arises from supervision being based solely on video diffusion loss (i.e., denoising score matching [64]) in the pixel/latent space, without explicit spatial awareness guidance in the feature space. Hence, in this paper, we ask if and how we can empower video diffusion models with appearance constancy by providing additional supervision.", + "bbox": [ + 509, + 625, + 906, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present Track4Gen as a spatially aware video generator that receives supervision both in terms of the original diffusion-based objective as well as (dense) point correspondence across frames, which we refer to as tracks. We demonstrate that it is possible to provide such track-level supervision in the diffusion feature space by making minimal architecture changes. Our generated videos do not suffer from degradation of video quality (according to the usual video generation metrics), while being significantly more", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during internship at Adobe.", + "bbox": [ + 114, + 887, + 331, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7276", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "spatially coherent as the highlight cow in Fig. 1.", + "bbox": [ + 89, + 90, + 410, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We train Track4Gen using the latest Stable Video Diffusion [4] as the backbone and evaluate on the publicly available VBench dataset [30]. We report significant improvement in terms of appearance constancy of subjects, both in quantitative and qualitative (i.e., via user studies) evaluations. In summary, we demonstrate that it is possible to upgrade existing video generators, by supervising them with additional correspondence tracking loss, to produce videos without significant appearance drifts, a problem commonly encountered in diffusion-based video generators.", + "bbox": [ + 88, + 106, + 480, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 268, + 232, + 284 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion-based video generation. Building on the success of diffusion models in image synthesis [11, 51], diffusion-based video generators have seen significant advancements [4, 6, 29, 47]. A commonly adopted approach is to extend text-to-image models to the video domain by incorporating temporal layers to facilitate interactions across video frames [5, 22, 54]. While some works have adopted cascaded approaches to produce both spatially and temporally high-resolution videos [28, 47, 54, 67, 76, 78], others have utilized lower-dimensional latent space modeling to reduce computational demands [5, 8, 24, 80]. We build on top of one such approach, Stable Video Diffusion (SVD, [4]), which introduces a latent image-to-video diffusion model trained on a large-scale and curated video data.", + "bbox": [ + 91, + 294, + 482, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With advances in generation, systematic evaluation of generation quality has become crucial. Traditionally, metrics such as Fréchet Inception Distance (FID, [26]), Fréchet Video Distance (FVD, [63]), and CLIPSIM [49] are used. Additionally, comprehensive benchmark suites [30, 68] have been introduced to provide a more robust evaluation aligned with human perception. Inspired by such work, we thoroughly evaluate our approach and demonstrate improved video generation quality with respect to both conventional metrics and the recent VBenchmark metrics [30].", + "bbox": [ + 89, + 506, + 482, + 656 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Foundational models as feature extractors. Various foundational models such as vision transformers [15] or diffusion-based generators [50] have been utilized as feature extractors for various tasks including semantic matching [16, 25, 41], classification [38], segmentation [66, 71], and editing [19, 21, 61]. There have been efforts to boost their performance by post-processing the feature maps obtained from the pre-trained models, e.g., by upsampling [18, 58]. In a recent effort, Yue et al. [75] lift semantic per-frame features from a foundational model into a 3D Gaussian representation. They fine-tune the foundational model with such 3D-aware features resulting in improved performance in downstream tasks. Similarly, Sundaram et al. [57] fine-tune state-of-the-art foundational models on human similarity judgments yielding improved representations across", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "downstream tasks. In a concurrent effort, Yu et al. [74] propose to align the internal features of an image generation model with external discriminative features [45], which results in more effective training of the generator.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work also enhances the internal feature representation of a foundational generation model but with significant differences compared to previous literature. First, unlike most previous work that focus on image level foundational models, we exploit the power of recently emerging video models. Second, instead of post-processing, we enhance the spatial awareness of the intermediate features by training the generator to jointly perform an additional tracking task. We show that this joint training boosts the performance of intermediate features in correspondence tracking, leading to improved video generation quality.", + "bbox": [ + 511, + 152, + 906, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Tracking any point in a video. The task involves following any arbitrary query point across a long video sequence. First introduced by PIPs [23] and later re-framed by TAPVid [12], several methods have emerged in recent years to tackle long-term point tracking. PIPs [23] revisits the classical particle-based representation [53] and introduces MLP-based networks that predict point tracks within an 8-frame window. Subsequent works have improved performance by capturing longer temporal context through advanced architectures [2, 13, 23, 33], as well as by enabling the simultaneous tracking of multiple queries [10, 33]. More recent training-based trackers [10, 34, 40, 70] have achieved remarkable performance by leveraging high-capacity neural networks to learn robust priors from large-scale training data. While high-quality data is crucial for accurate tracking, manually annotating point tracks is prohibitively expensive. Hence, synthetic videos [20] with automatic annotations, have become an alternative and have demonstrated effectiveness in real-world video tracking. An alternative approach is self-supervised adaptation at test time, where tracking is learned without ground-truth labels [32, 62, 65]. In a recent study, Aydemir et al. [1] evaluate the effectiveness of several image foundational model features for point tracking both in zero-shot setting as well as with supervised training using low-rank adapter layers. To the best of our knowledge, we are the first to exploit the features of a foundational video diffusion model for dense point tracking.", + "bbox": [ + 511, + 335, + 906, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 753, + 604, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we provide a comprehensive discussion of the Track4Gen framework. We begin with a concise overview of latent video diffusion models (Sec. 3.1). Next, we discuss how video diffusion features relate to temporal correspondences both for real and generated videos (Sec. 3.2). Finally, we detail the design of Track4Gen both in terms of network architecture and the employed supervision signals (Sec. 3.3). An overview is depicted in Fig. 2.", + "bbox": [ + 511, + 780, + 903, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7277", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bc7f5538b0f54c4bee03cb71c68aa9cc4788c40be294ff6445d714b2d5172ec1.jpg", + "image_caption": [ + "Figure 2. Track4Gen overview. Red-colored blocks represent layers optimized by the diffusion loss $\\mathcal{L}_{\\mathrm{diff}}$ , while green blocks are optimized by the correspondence loss $\\mathcal{L}_{\\mathrm{corr}}$ . Blocks colored both red and green are influenced by the joint loss, $\\mathcal{L}_{\\mathrm{diff}} + \\lambda \\mathcal{L}_{\\mathrm{corr}}$ . See text for details." + ], + "image_footnote": [], + "bbox": [ + 101, + 85, + 893, + 280 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Background: Stable Video Diffusion", + "text_level": 1, + "bbox": [ + 89, + 340, + 405, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Starting from random Gaussian noise, diffusion models aim to generate clean images or videos via an iterative denoising process [27, 55]. This process reverses a fixed, time-dependent diffusion forward process, which gradually corrupts the data by adding Gaussian noise. While our method is applicable to general video diffusion models, in this paper, we design our architecture based on Stable Video Diffusion (SVD), a latent video diffusion model which employs the EDM-framework [35]. The diffusion process operates in the lower-dimensional latent space of a pre-trained VAE [37], consisting of an encoder $\\mathcal{E}(\\cdot)$ and a decoder $\\mathcal{D}(\\cdot)$ .", + "bbox": [ + 88, + 364, + 482, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a clean sample $\\pmb{x}_0^{1:N} \\sim p_{\\mathrm{data}}(\\pmb{x})$ of an $N$ -frame video sequence, the frames are first encoded into the latent space as $z_0^{1:N} = \\mathcal{E}(x_0^{1:N})$ . Gaussian noise $\\epsilon \\sim \\mathcal{N}(0,I)$ is then added to the latents to produce the intermediate noisy latents via the forward process $z_t^{1:N} = \\alpha_t z_0^{1:N} + \\sigma_t \\epsilon$ , where $t$ represents the diffusion timestep, and $\\alpha_t, \\sigma_t$ are the discretized noise scheduler parameters. The diffusion denoiser $f_\\theta$ is trained by minimizing the v-prediction loss:", + "bbox": [ + 89, + 534, + 483, + 656 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {\\boldsymbol {\\epsilon} \\sim \\mathcal {N} (0, I), t \\sim U [ 1, T ]} \\left[ \\left\\| \\boldsymbol {f} _ {\\theta} \\left(\\boldsymbol {z} _ {t} ^ {1: N}, t, c\\right) - \\boldsymbol {y} \\right\\| _ {2} ^ {2} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 672, + 482, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pmb{y}$ is defined as $\\pmb{y} = \\alpha_{t}\\pmb{\\epsilon} - \\sigma_{t}\\pmb{z}_{0}^{1:N}$ . In the image-to-video variant of SVD, the condition $c$ refers to the CLIP image embedding [49], replacing the typical text embeddings. For the remainder of this paper, we will refer to Eq. 1 as the video diffusion loss $\\mathcal{L}_{\\mathrm{diff}}$ .", + "bbox": [ + 89, + 715, + 482, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Once trained, the diffusion model generates videos by iteratively denoising a noisy latent $\\pmb{z}_T^{1:N}$ sequence sampled from pure Gaussian distribution. At each diffusion step, the model predicts the noise in the input latent. Once the clean latent $\\pmb{z}_0^{1:N}$ is obtained, the decoder $\\mathcal{D}$ maps it to the higher-dimensional pixel space $\\pmb{x}_0^{1:N} = \\mathcal{D}(\\pmb{z}_0^{1:N})$ . For further details, we refer to the Appendix D of [4].", + "bbox": [ + 89, + 795, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Video Diffusion Features", + "text_level": 1, + "bbox": [ + 513, + 340, + 740, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous studies have demonstrated that image diffusion models learn discriminative features in their hidden states that are effective for various analysis tasks and propose methods for improving the representation power of such features [9, 69, 73, 74]. Similarly, we argue that while also being powerful, internal representations of pre-trained video diffusion models may not be fully temporally consistent, resulting in appearance drift in generated videos.", + "bbox": [ + 511, + 364, + 906, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To better investigate this hypothesis, we first evaluate the long-term video tracking capabilities of U-Net-based video diffusion models [4, 56, 78]. Specifically, we evaluate the effectiveness of the features from each block of the U-Net for the task of point tracking. Given a real-world video, we add a small amount of noise and extract feature maps from each layer in each block. We perform a cosine-similarity-based nearest-neighbor search [44, 59] over these feature maps for a given set of fixed query points on the first frame (we use a similarity threshold of 0.6 [62] in our experiments). We also perform a similar analysis for generated videos where we extract the feature maps corresponding to diffusion steps with small amount of noise.", + "bbox": [ + 511, + 489, + 908, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on this feature analysis, we make some important observations. Notably, regardless of the model (we analyze both Zeroscope T2V [56] and SVD I2V [4]), we find out that output features from the upsampler layer of the third decoder block consistently yield stronger temporal correspondences, as shown in Fig. 3. Hence, we use this block when extracting features for the remainder of our experiments. Furthermore, when we analyze generated videos and point tracks estimated based on the feature maps (as shown in Fig. 4), we observe that there is a correlation between tracking failures that reveal feature-space inconsistencies and appearance drifts that reveal pixel-space inconsistencies. Hence, we hypothesize that enriching feature consistency can help mitigate such appearance drifts. Next, we", + "bbox": [ + 511, + 688, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7278", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c3d26ff034dd6c8539f4a1070898728c77377c605e1a87b947bd6d0e60718dfa.jpg", + "image_caption": [ + "Figure 3. Real-world video tracking using different video diffusion features. Given color-coded query points on the first frame (Leftmost column), we display tracked points on target frames using features from different blocks (right columns). The 13th frame (first row) and 8th frame (second row) are shown as target frames. Full results are available in the supplementary and on our page." + ], + "image_footnote": [], + "bbox": [ + 99, + 85, + 897, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "introduce Track4Gen where we accomplish this goal by supervising video diffusion models with a joint tracking loss.", + "bbox": [ + 89, + 323, + 482, + 356 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Track4Gen", + "text_level": 1, + "bbox": [ + 89, + 366, + 215, + 382 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Track4Gen aims to utilize point tracking as an additional supervision signal to enhance the spatial-awareness of video diffusion features. Given that we build on top of a pretrained video generation model, to retain the prior knowledge and avoid tampering the original features directly, we propose a novel architecture change as shown in Fig. 2. Specifically, instead of directly using the raw diffusion features for correspondence estimation, we propose a trainable refiner module $R_{\\phi}$ , which is designed to refine the raw features by projecting them into a correspondence-rich feature space. The refined features, which are spatially-aware, are then both used to estimate point tracks with an explicit supervision as well as feeding back to the generation backbone. We empirically find out that this design is more effective compared to fine-tuning the original model with no refinement module (see Sec. 4.2).", + "bbox": [ + 88, + 388, + 482, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an $N$ -frame video sequence $\\pmb{x}_0^{1:N}$ , its correspond-", + "bbox": [ + 109, + 632, + 482, + 648 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/83341f49e67f07a973752f72dd26e68bfee31c1bf8a66a2b4c327ef72925575a.jpg", + "image_caption": [ + "Feature-space Inconsistency (SVD)", + "Figure 4. Generated video tracking using video diffusion features. Tracks based on diffusion features are annotated on the generated videos. Track4Gen generates more consistent results." + ], + "image_footnote": [], + "bbox": [ + 93, + 683, + 287, + 851 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0244d30198704051d756247245a1eea0793ec1bb096c52e163b987358ded1e32.jpg", + "image_caption": [ + "Feature-space Consistency (Track4Gen)" + ], + "image_footnote": [], + "bbox": [ + 287, + 683, + 480, + 851 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing latent $z_0^{1:N}$ , and a diffusion timestep $t$ , in order to train Track4Gen we continue to utilize the standard diffusion training loss as defined in Eq. (1), where we adopt the velocity prediction objective [4, 35, 52] for $\\mathcal{L}_{\\mathrm{diff}}$ .", + "bbox": [ + 511, + 323, + 903, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enable tracking supervision, we assume access to a dense set of point trajectories $\\Omega = \\{(x^i,x^j)\\}$ across frames where a point $\\mathbf{x}^i$ in frame $i$ corresponds to a matching point $\\mathbf{x}^j$ in frame $j$ and vice versa. Given the corresponding noisy video latent sequence $\\pmb{z}_t^{1:N}$ , we first extract raw diffusion features as the hidden states $\\pmb{h}^{1:N} \\in \\mathbb{R}^{N \\times H \\times W \\times C}$ from a specific block $b^k$ within the U-Net, where $b^k$ is set to the upsampler layer of the third decoder block (see Sec. 3.2). We then pass these features through the refiner module to obtain the refined feature map $\\tilde{\\pmb{h}}^{1:N} = \\pmb{R}_{\\phi}(\\pmb{h}^{1:N})$ .", + "bbox": [ + 511, + 385, + 905, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We sample a query point $\\mathbf{x}_{\\mathrm{q}}^{i}$ along with its ground-truth target point $\\mathbf{x}_{\\mathrm{trg}}^{j}$ from the correspondence set $\\Omega$ . Given the query point feature $\\tilde{\\pmb{h}}^i (\\mathrm{x_q})\\in \\mathbb{R}^{1\\times 1\\times C}$ and the target feature map $\\tilde{\\pmb{h}}^j\\in \\mathbb{R}^{H\\times W\\times C}$ , we calculate the cost volume $S\\in$ $\\mathbb{R}^{H\\times W\\times 1}$ as follows:", + "bbox": [ + 511, + 541, + 905, + 628 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {S} (\\mathrm {p}) = \\cos - \\operatorname {s i m} \\left(\\tilde {\\boldsymbol {h}} ^ {i} \\left(\\mathrm {x} _ {\\mathrm {q}}\\right), \\tilde {\\boldsymbol {h}} ^ {j} (\\mathrm {p})\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 640, + 903, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where cos-sim denotes cosine similarity. The predicted target point $\\hat{\\mathbf{x}}_{\\mathrm{trg}}$ is then determined using the differentiable soft-argmax operation:", + "bbox": [ + 511, + 672, + 906, + 720 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} _ {\\mathrm {t r g}} = \\frac {\\sum_ {\\mathrm {p} \\in \\Omega^ {\\prime}} \\boldsymbol {S} (\\mathrm {p}) \\cdot \\mathrm {x} _ {\\mathrm {p}}}{\\sum_ {\\mathrm {p} \\in \\Omega^ {\\prime}} \\boldsymbol {S} (\\mathrm {p})}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 731, + 903, + 770 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Omega^{\\prime} = \\{p:\\| \\mathrm{x_p - x_{p_{max}}}\\| _2\\leq R\\} ^1$ . Thus, the target point prediction can be expressed as $\\hat{\\mathbf{x}}_{\\mathrm{trg}} = \\xi (\\mathrm{x}_{\\mathrm{q}}^{i},j,\\tilde{\\boldsymbol{h}}^{1:N})$ and the predicted tracklet for $\\mathrm{x_q^i}$ is given by $\\mathcal{T}_{\\mathrm{x_q^i}} = \\{\\hat{\\mathbf{x}}_n$ .. $\\hat{\\mathbf{x}}_n = \\xi (\\mathbf{x}_q^i,n,\\tilde{\\pmb{h}}^{1:N}),n = 1,\\dots,N\\}$ Finally, the correspond", + "bbox": [ + 511, + 782, + 906, + 861 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "The feature maps have a resolution of $44 \\times 81$ for an input video resolution of $320 \\times 576$ , and we set $R = 35$ .", + "bbox": [ + 511, + 875, + 906, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7279", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bd21d1d1e4725d6bd405f75ea672a22dcd9fc971cfe51f587d57661f318b1c08.jpg", + "image_caption": [ + "Input Image", + "Figure 5. Image-to-video generation results of the original SVD and Track4Gen. Please visit our page for full video view." + ], + "image_footnote": [], + "bbox": [ + 94, + 99, + 254, + 449 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b4bd58864b5e72ec37a126dcb9983fb8dbf6ad4feed6c7871a53498565368a84.jpg", + "image_caption": [ + "Stable Video Diffusion (SVD)" + ], + "image_footnote": [], + "bbox": [ + 256, + 99, + 578, + 449 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dd5cbacaef5caf25b8cd78c19c35e4fee690de50cbcc63a9bc69e13c75217f69.jpg", + "image_caption": [ + "Track4Gen (Ours)" + ], + "image_footnote": [], + "bbox": [ + 580, + 99, + 903, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "dence loss $\\mathcal{L}_{\\mathrm{corr}}$ is computed using the Huber loss $L_{H}$ [31]:", + "bbox": [ + 89, + 493, + 482, + 508 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o r r}} \\left(\\tilde {\\boldsymbol {h}} ^ {1: N}, \\Omega\\right) = \\sum_ {\\left(\\mathrm {x} _ {\\mathrm {q}} ^ {i}, \\mathrm {x} _ {\\text {t r g}} ^ {j}\\right) \\in \\Omega} L _ {H} \\left(\\xi \\left(\\mathrm {x} _ {\\mathrm {q}} ^ {i}, j, \\tilde {\\boldsymbol {h}} ^ {1: N}\\right), \\mathrm {x} _ {\\text {t r g}} ^ {j}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 516, + 482, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When training Track4Gen, we initialize the refiner module as an identity mapping to fully leverage the prior of the base model at the start of finetuning. To re-route the refined features to the backbone generator, we introduce a trainable zero convolution layer [77], denoted as $\\zeta_{\\psi}$ . While the diffusion loss $\\mathcal{L}_{\\mathrm{diff}}$ back-propagates to all the blocks of the video diffusion model, we detach the gradients of $\\tilde{\\pmb{h}}^{1:N}$ before passing into $\\zeta_{\\psi}$ such that refiner module can solely focus on acquiring the correspondence prior. Hence, given that the output of block $b^{k}$ is $h^{1:N}$ , the input to the subsequent block $b^{k+1}$ is computed as $h^{1:N} + \\zeta_{\\psi}$ (stop-gradient $(R_{\\phi}(h^{1:N}))$ ).", + "bbox": [ + 89, + 565, + 483, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 2 visualizes this architecture design, with red and green colors indicating the objective that optimizes each module.", + "bbox": [ + 89, + 734, + 482, + 765 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 777, + 223, + 795 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 801, + 307, + 818 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To train Track4Gen, we construct a training dataset consisting of 567 video-trajectory pairs, with each video having a resolution of $320 \\times 576$ and a duration of 24 frames. Since no real-world video with (dense) ground-truth trajectory annotations exist at the time of this work, we utilize", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "optical flow to generate trajectory annotations. A key challenge is the need for accurate video segmentation maps to ensure a balanced distribution of trajectory points between foreground objects and the background [12]. To address this, we utilize public video datasets paired with ground-truth segmentation maps [7, 17, 39, 46, 48], where we split longer videos into 24-frame segments.", + "bbox": [ + 511, + 493, + 906, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use Stable Video Diffusion (SVD) image-to-video pretrained checkpoints $^2$ as the base video generator. Our proposed refiner module consists of eight stacked 2D convolution layers and is attached to the third decoder block of the SVD UNet. The refiner module preserves the shape of the hidden states throughout and is initialized as the identity mapping. Further details are provided in the supplementary. We finetune this enhanced video generator architecture for $20\\mathrm{K}$ steps with our joint loss $\\mathcal{L}_{\\mathrm{diff}} + \\lambda \\mathcal{L}_{\\mathrm{corr}}$ , where $\\lambda$ is set to 8. Rather than finetuning the entire model, we finetune only the temporal transformer blocks, the refiner module $R_{\\phi}$ , and the zero convolution $\\zeta_{\\psi}$ . In each iteration, we sample 512 correspondence pairs from the precomputed trajectories. We use the AdamW optimizer [43] with a learning rate of $1e - 5$ , $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.999$ , and a weight decay of $1e - 2$ . We train the model on $4\\times$ H100 GPUs with a total batch size of 4. For sampling new videos, we apply the", + "bbox": [ + 511, + 599, + 908, + 859 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2https://huggingface.co/stabilityai/stable-video-diffusion-image2vid-xt", + "bbox": [ + 511, + 875, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7280", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2007efe96eb2e6f569980b87b25183468361e418d42b4a9f51fa2c9b6ddf40b8.jpg", + "image_caption": [ + "Figure 6. Qualitative ablation on video generation. Track4Gen is compared with finetuned SVD (SVD finetuned on the same training videos without any correspondence supervision) and Track4Gen trained without the Refiner module." + ], + "image_footnote": [], + "bbox": [ + 94, + 85, + 906, + 392 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "default settings using 30 steps with the EDM sampler [35], motion bucket id =127, and fps=7.", + "bbox": [ + 89, + 450, + 482, + 481 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fc64a6199491e85adc14954283930459f20aa27ea8a6506e901d8ec729370e25.jpg", + "table_caption": [ + "Table 1. Quantitative comparison on video generation performance. We compare Track4Gen to the pre-trained SVD $\\star$ as well as a finetuned SVD on the same dataset (finetuned SVD). We also train a variant of Track4Gen without the refiner module. All videos are generated at $320\\mathrm{x}576$ resolution, except SVD $\\star$ (576p) which operates at $576\\mathrm{x}1024$ resolution." + ], + "table_footnote": [], + "table_body": "
Subject ConsistencyTemporal FlickeringMotion SmoothnessImaging QualityVideo-Image AlignmentFIDFVD
SVD*0.95350.94640.97740.66480.953929.0776
finetuned SVD0.96650.98000.99090.67660.977127.0735
Track4Gen w/o refiner0.95060.97250.97910.66530.961427.1718
Track4Gen0.97460.98060.99210.68350.981426.6724
SVD* (576p)0.95760.94780.97950.68120.9582--
", + "bbox": [ + 91, + 579, + 483, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Track4Gen for Video Generation", + "text_level": 1, + "bbox": [ + 89, + 667, + 382, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate Track4Gen for the image-to-video generation task via a series of experiments using multiple datasets, automated metrics, and human evaluations.", + "bbox": [ + 89, + 689, + 482, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Setup. We compare Track4Gen against the original SVD $(\\mathrm{SVD}^*)$ [4], as well as a version of SVD that is finetuned on the same videos as Track4Gen (finetuned SVD). Furthermore, we train a variant of Track4Gen without the refiner module. For VBenchmark metrics [30], evaluations are conducted on the VBenchmark-I2V dataset, containing 355 diverse images. FID and FVD are measured using the DAVIS [48] dataset as reference. We generate 24-frame videos conditioned on each input image.", + "bbox": [ + 89, + 734, + 482, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Automatic metrics. We first report five key metrics from VBench [30]: (1) Subject Consistency—assesses", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "subject appearance consistency of the video by computing the similarity of DINO [45] features. (2) Temporal Flickering—detects temporal consistency by taking static frames and calculating the mean absolute difference across frames. (3) Motion Smoothness—measures smoothness of motion, and how well it adheres to real-world physics, using video frame interpolation priors [42]. (4) Image Quality—evaluates distortions (e.g., noise, blur) using a pretrained, multi-scale image quality predictor [36]. (5) Video-Image Alignment—measures alignment between the subject in the input image and in the generated video using DINO features. We additionally report FID [26] and FVD [63].", + "bbox": [ + 511, + 450, + 906, + 632 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Human evaluation. We further evaluate Track4Gen against baselines through a user study. We ask 64 participants to compare our results with a randomly selected baseline. We ask the users to evaluate how consistent main objects appear across the frames in a generated video as well as how natural the depicted motion is. We provide further details of the user study in the supplementary material.", + "bbox": [ + 511, + 637, + 906, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative results. Qualitative comparisons with the base SVD are shown in Fig. 5. As illustrated, Track4Gen generates videos with strong appearance consistency, avoiding issues of appearance drift. In contrast, videos produced by the original SVD exhibit noticeable inconsistencies: the sheep's head (row 1) mutates, the plane's wing (row 2) shows unnatural transitions, and the cars (row 3) disappear. Further comparisons with finetuned SVD and Track4Gen without the refiner module are shown in Fig. 6 and highlight the superior visual coherence of the proposed Track4Gen.", + "bbox": [ + 511, + 750, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7281", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/89b328905f0a6f1472a28dcc4c4c46facf48d8876c515a62ea87af911aa790c0.jpg", + "image_caption": [ + "Query Points", + "Figure 7. Qualitative comparison of Track4Gen and baselines for real-world video tracking. The leftmost column displays query points in the first frame, while the following three columns show tracking results using features from each model." + ], + "image_footnote": [], + "bbox": [ + 93, + 99, + 295, + 345 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/471eabf03fc07b9cf3fa0b5ed52d1690cf1344f5750cc99bd0ccacfafb3bb601.jpg", + "image_caption": [ + "Track4Gen (Ours)" + ], + "image_footnote": [], + "bbox": [ + 297, + 99, + 500, + 345 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3e9dc7be668900098ba8e5d2da1ed158065ecc773d674a1a4be1d5d83d3a6ecc.jpg", + "image_caption": [ + "SVD 12V" + ], + "image_footnote": [], + "bbox": [ + 500, + 99, + 700, + 345 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7faa10865074e5889f0db1d531836c40a01d254c72c9fc1fa9c994d92ac8a5e2.jpg", + "image_caption": [ + "ZeroScope T2V" + ], + "image_footnote": [], + "bbox": [ + 702, + 99, + 903, + 345 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0f2ac8d7130d7332c0140967b0d794c2808e1e2f98140b56f91d304a3b989d40.jpg", + "image_caption": [ + "(a) Identity preservation", + "Figure 8. User study results. Our study shows that Track4Gen better preserves object identity and produces more natural motion." + ], + "image_footnote": [], + "bbox": [ + 93, + 388, + 279, + 489 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d6b4d5d15a2c235187bff389aad213516d46ebb79126cc780c35ca89eb6ba6b4.jpg", + "image_caption": [ + "(b) Motion naturalness" + ], + "image_footnote": [], + "bbox": [ + 297, + 391, + 480, + 488 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative results. As shown in Tab. 1, our method achieves the highest scores across all 5 metrics from VBench, along with the lowest FID and second-lowest FVD values, outperforming the base SVD by substantial margins. Fig. 8 provides the user study results where the majority of the participants agreed that Track4Gen is superior both in terms of identity preservation and naturalness of motion.", + "bbox": [ + 89, + 546, + 483, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Track4Gen for Video Tracking", + "text_level": 1, + "bbox": [ + 89, + 660, + 364, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate Track4Gen's capability to track any point in real videos by adding a small amount of noise to the input video [59] and passing it through the video denoiser $\\pmb{f}_{\\theta}$ to extract feature maps. We first compare tracking results with such features against other raw features [4, 56, 60] in Sec. 4.3.1. In Sec. 4.3.2, we utilize Track4Gen's features in a test-time optimization method [62] and compare to both self-supervised and fully supervised video trackers.", + "bbox": [ + 89, + 680, + 482, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.1. Zero-shot Feature Comparison", + "text_level": 1, + "bbox": [ + 89, + 806, + 352, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the precision of predicted tracks using the features from Track4Gen, the original SVD model $(\\mathrm{SVD}^{\\star})$ and RAFT [60]. We also test another text-to-video model, ZeroScope T2V [56], to demonstrate how raw features from pre-trained video generators typically work out of the box.", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f8f9f1546315dac740978bb4462735c599c7c2b62193d532e418f98078c404f2.jpg", + "table_caption": [ + "Table 2. Quantitative zero-shot feature comparison on video tracking benchmarks. Track4Gen features are compared to the features of SVD* [4], ZeroScope [56], and RAFT [60]. For all the metrics, higher values indicate better performance." + ], + "table_footnote": [], + "table_body": "
MethodDAVIS-480p(24-frame)BADJA(24-frame)DAVIS-480p(whole duration)BADJA(whole duration)
\\( \\delta^{x}_{avg} \\)OAAJ\\( \\delta^{seg} \\)\\( \\delta^{3px} \\)\\( \\delta^{x}_{avg} \\)OAAJ\\( \\delta^{seg} \\)\\( \\delta^{3px} \\)
ZeroScope46.267.039.427.52.837.259.527.819.92.0
SVD*42.479.736.426.22.935.470.126.519.42.2
Track4Gen69.785.856.552.37.758.978.440.240.45.0
RAFT73.3--54.88.766.7--45.05.8
", + "bbox": [ + 514, + 450, + 906, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For RAFT, tracking is achieved by chaining optical flow displacements, while the others use nearest neighbor matching between its encoded features.", + "bbox": [ + 511, + 546, + 905, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets. We use TAP-Vid DAVIS [12] and BADJA [3] as benchmark datasets. Additionally, we include two shorter benchmarks, DAVIS (24-frame) and BADJA (24-frame), which focus on the first 24 frames with query and target points within this range. Details on encoding long videos with the video models are in the supplemental.", + "bbox": [ + 511, + 595, + 905, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics. For evaluating the TAP-Vid benchmarks, we use the following metrics: (i) Position Accuracy $(\\delta_{avg}^{x})$ evaluates the average accuracy of visible points, where each $\\delta^x$ represents the fraction of predicted points that lie within $x$ pixels of the ground-truth position, with $x \\in \\{1, 2, 4, 8, 16\\}$ . (ii) Occlusion Accuracy (OA) evaluates the correctness of occlusion predictions. (iii) Average Jaccard (AJ) jointly assesses both position and occlusion accuracy. For the BADJA dataset, we report $\\delta^{seg}$ , which measures the accuracy of tracked keypoints within a distance of $0.2\\sqrt{A}$ from the ground-truth annotation, where $A$ is the area of the foreground object. We also report $\\delta^{3px}$ , which assesses accuracy within a 3-pixel threshold. A cosine similarity threshold of 0.6 is used for occlusion prediction.", + "bbox": [ + 511, + 689, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7282", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/620611b3f3f85c7a27be961227007156bb3ec0f3fa0572c87926f784cd601428.jpg", + "image_caption": [ + "First Frame with Query Points", + "Figure 9. Extending Track4Gen with test-time adaptation [62]." + ], + "image_footnote": [], + "bbox": [ + 93, + 95, + 222, + 263 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c67742c215bddb707aac03f8fff1c06e96eeb62ceba6f34f040bdba252afead8.jpg", + "image_caption": [ + "Tracking Results in Subsequent Frames" + ], + "image_footnote": [], + "bbox": [ + 222, + 95, + 480, + 263 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/52351e7a6e44c7fbb0938716533d442c069954680d873afd0292309b3fd11e59.jpg", + "table_caption": [ + "Table 3. Quantitative comparison with video trackers. Although primarily designed for video generation, Track4Gen combined with a test-time optimization method [62] achieves performance comparable to dedicated video tracking frameworks, even when compared to supervised methods." + ], + "table_footnote": [ + "\\* - supervised. -test-time training." + ], + "table_body": "
MethodDAVIS-480BADJA
δxavgOAAJδsegδ3px
TAP-Net*66.479.046.045.49.6
PIPs++*73.6--59.09.8
TAPIR*77.389.565.768.710.5
Omnimotion†74.184.558.445.26.9
DINO-Tracker†80.488.164.672.414.3
DINO-Tracker w/ Track4Gen†72.584.555.748.410.9
", + "bbox": [ + 91, + 368, + 480, + 477 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/beb2795496a6d71f289ac6415c55f7d6bcc18afcaf8eaa2e2f1075d0b0b2746d.jpg", + "table_caption": [ + "Table 4. Ablation on trainable modules and refiner." + ], + "table_footnote": [], + "table_body": "
Trainable modulesSubject ConsistencyTemporal FlickeringMotion SmoothnessImaging QualityVideo-Image Alignment
spatial + temporal0.97340.98110.99170.68630.9807
spatial0.97260.98010.99190.68520.9811
temporal0.97460.98060.99210.68350.9814
Refiner architecture
2D convolutions0.97460.98060.99210.68350.9814
3D convolutions0.96870.97340.99040.68330.9820
", + "bbox": [ + 91, + 513, + 480, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. We present the qualitative results in Fig. 7 and the quantitative results in Tab. 2. Although primarily designed for video generation, Track4Gen boosts the poor performance of the pre-trained video models significantly, approaching the accuracy of RAFT optical flow chaining.", + "bbox": [ + 89, + 627, + 482, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.2. Extending Track4Gen with Test-time Adaptation", + "text_level": 1, + "bbox": [ + 89, + 713, + 477, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further evaluate Track4Gen's long-term tracking capabilities, we integrate our features with test-time adaptation algorithm of DINO-Tracker [62], where a per-video optimization is performed using optical flow supervision. We replace the originally used DINOv2 [45] with the features from Track4Gen. We evaluate using the same datasets and metrics outlined in Sec. 4.3.1, against both fully-supervised trackers [12, 13, 79] and self-supervised methods [62, 65].", + "bbox": [ + 89, + 733, + 482, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 3 shows that Track4Gen features optimized with [62] achieve performance comparable to dedicated trackers. Qualitative results are in Fig. 9 and in the supplemental.", + "bbox": [ + 89, + 854, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5ce062a81c9ee0d223a67b8e25a6535a1d61d5e614c05b93b249ef347120efad.jpg", + "table_caption": [ + "Table 5. Quantitative ablation on using annotated, but synthetic videos [20]. Left: Video generation metrics. Right: Video tracking metrics." + ], + "table_footnote": [], + "table_body": "
Dataset compositionSubject ConsistencyMotion SmoothnessImaging QualityBADJA
δsegδ3px
real videos0.97470.99210.683340.45.0
real + synthetic videos0.97080.98920.679342.14.8
", + "bbox": [ + 514, + 133, + 903, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 204, + 678, + 218 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present an ablation study in Tab. 4 where we train different set of modules. Each spatio-temporal block of SVD includes both spatial and temporal transformers. We compare training only spatial transformers, only temporal transformers, or both. We also ablate the architecture of the refiner module using either 2D or 3D convolution layers. Our analysis shows that while results are similar across settings, training only the temporal transformers in SVD with 2D convolutions as the refiner module yields optimal video generation quality. We further analyze our training dataset by additionally incorporating Kubric [20] simulated videos (1K video-track pairs from the Panning MOVi-E data [10, 13]) with automatically annotated trajectories into training. As shown in Tab. 5, optical flow-chained tracklets from real provides provide as effective correspondence guidance as tracklets from synthetic data, while synthetic videos negatively impact the video generation quality.", + "bbox": [ + 511, + 226, + 906, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 511, + 497, + 782, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented the first unified framework that bridges two distinct tasks: video generation and dense point tracking. We demonstrated that this produces temporally consistent feature representations and appearance-consistent videos. As for limitations, videos generated by Track4Gen tend to exhibit less dynamic motion compared to those from other video generators. Additionally, failure cases are included in the supplementary material.", + "bbox": [ + 511, + 522, + 905, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Future work. Recently, cutting-edge video trackers [10, 14, 34] have emerged, enabling dense, accurate, and long-term tracking, especially with better handling of occlusions. This opens up promising future directions for extending our work to utilize real-world videos, automatically annotated by these advanced trackers.", + "bbox": [ + 511, + 643, + 905, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. We thank Seokju Cho and Narek Tumanyan for their invaluable feedback on video point tracking. We also extend our gratitude to Mingi Kwon, Joon-Young Lee, and Gabriel Huang for their insightful discussions. Hyeonho Jeong and Jong Chul Ye are supported by the National Research Foundation of Korea (NRF) under Grants RS-2024-00336454 and RS-2023-00262527, and by the Institute for Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (RS-2019-II190075, Artificial Intelligence Graduate School Program, KAIST).", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7283", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Gorkay Aydemir, Weidi Xie, and Fatma Güney. Can visual foundation models achieve long-term point tracking? arXiv preprint arXiv:2408.13575, 2024. 2", + "[2] Weikang Bian, Zhaoyang Huang, Xiaoyu Shi, Yitong Dong, Yijin Li, and Hongsheng Li. Context-tap: Tracking any point demands spatial context features. arXiv preprint arXiv:2306.02000, 3, 2023. 2", + "[3] Benjamin Biggs, Thomas Roddick, Andrew Fitzgibbon, and Roberto Cipolla. Creatures great and small: Recovering the shape and motion of animals from video. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part V 14, pages 3-19. Springer, 2019. 7", + "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1, 2, 3, 4, 6, 7", + "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2", + "[6] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 1, 2", + "[7] Sergi Caelles, Alberto Montes, Kevis-Kokitsi Maninis, Yuhua Chen, Luc Van Gool, Federico Perazzi, and Jordi Pont-Tuset. The 2018 davis challenge on video object segmentation. arXiv preprint arXiv:1803.00557, 2018. 5", + "[8] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2", + "[9] X Chen, Z Liu, S Xie, and K He. Deconstructing denoising diffusion models for self-supervised learning. arxiv 2024. arXiv preprint arXiv:2401.14404. 3", + "[10] Seokju Cho, Jiahui Huang, Jisu Nam, Honggyu An, Seun-gryong Kim, and Joon-Young Lee. Local all-pair correspondence for point tracking. arXiv preprint arXiv:2407.15420, 2024. 2, 8", + "[11] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2", + "[12] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610-13626, 2022. 2, 5, 7, 8", + "[13] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman." + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Tapir: Tracking any point with per-frame initialization and temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10061-10072, 2023. 2, 8", + "[14] Carl Doersch, Yi Yang, Dilara Gokay, Pauline Luc, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ross Goroshin, João Carreira, and Andrew Zisserman. Bootstrap: Bootstrapped training for tracking-any-point. arXiv preprint arXiv:2402.00847, 2024. 8", + "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2", + "[16] Niladri Shekhar Dutt, Sanjeev Muralikrishnan, and Niloy J. Mitra. Diffusion 3d features (diff3f): Decorating textured shapes with distilled semantic features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4494-4504, 2024. 2", + "[17] Qingnan Fan, Fan Zhong, Dani Lischinski, Daniel Cohen-Or, and Baoquan Chen. Jumpcut: non-successive mask transfer and interpolation for video cutout. ACM Trans. Graph., 34 (6):195-1, 2015. 5", + "[18] Stephanie Fu, Mark Hamilton, Laura E. Brandt, Axel Feldmann, Zhoutong Zhang, and William T. Freeman. Feuput: A model-agnostic framework for features at any resolution. In ICLR, 2024. 2", + "[19] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 2", + "[20] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3749-3761, 2022. 2, 8", + "[21] Yuchao Gu, Yipin Zhou, Bichen Wu, Licheng Yu, Jia-Wei Liu, Rui Zhao, Jay Zhangjie Wu, David Junhao Zhang, Mike Zheng Shou, and Kevin Tang. Videoswap: Customized video subject swapping with interactive semantic point correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7621–7630, 2024. 2", + "[22] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2", + "[23] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In European Conference on Computer Vision, pages 59-75. Springer, 2022. 2", + "[24] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221, 2022. 2" + ], + "bbox": [ + 517, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7284", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Eric Hedlin, Gopal Sharma, Shweta Mahajan, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. Unsupervised semantic correspondence using stable diffusion. In NIPS, 2023. 2", + "[26] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 2, 6", + "[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3", + "[28] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2", + "[29] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2", + "[30] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 2, 6", + "[31] Peter J Huber. Robust estimation of a location parameter. In Breakthroughs in statistics: Methodology and distribution, pages 492-518. Springer, 1992. 5", + "[32] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances in neural information processing systems, 33:19545-19560, 2020. 2", + "[33] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv preprint arXiv:2307.07635, 2023. 2", + "[34] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker3: Simpler and better point tracking by pseudolabelling real videos. arXiv preprint arXiv:2410.11831, 2024. 2, 8", + "[35] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. 3, 4, 6", + "[36] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 6", + "[37] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3", + "[38] Alexander C. Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your diffusion model is secretly a zero-shot classifier. In ICCV, 2023. 2" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Fuxin Li, Taeyoung Kim, Ahmad Humayun, David Tsai, and James M Rehg. Video segmentation by tracking many figure-ground segments. In Proceedings of the IEEE international conference on computer vision, pages 2192-2199, 2013. 5", + "[40] Hongyang Li, Hao Zhang, Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, and Lei Zhang. Taptr: Tracking any point with transformers as detection. arXiv preprint arXiv:2403.13042, 2024. 2", + "[41] Xinghui Li, Jingyi Lu, Kai Han, and Victor Prisacariu. Sd4match: Learning to prompt stable diffusion model for semantic matching. In CVPR, 2023. 2", + "[42] Zhen Li, Zuo-Liang Zhu, Ling-Hao Han, Qibin Hou, Chun-Le Guo, and Ming-Ming Cheng. Amt: All-pairs multi-field transforms for efficient frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9801-9810, 2023. 6", + "[43] Ilya Loshchilov, Frank Hutter, et al. Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101, 5, 2017. 5", + "[44] Grace Luo, Lisa Dunlap, Dong Huk Park, Aleksander Holynski, and Trevor Darrell. Diffusion hyperfeatures: Searching through time and space for semantic correspondence. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[45] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 6, 8", + "[46] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 5", + "[47] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 1, 2", + "[48] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 5, 6", + "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3", + "[50] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2021. 2", + "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + ], + "bbox": [ + 516, + 90, + 905, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7285", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 4", + "[53] Peter Sand and Seth Teller. Particle video: Long-range motion estimation using point trajectories. International journal of computer vision, 80:72-91, 2008. 2", + "[54] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2", + "[55] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3", + "[56] Spencer Sterling. Zeroscope, 2023. https://huggingface.co/cerspense/zeroscope_v2_576w.3.7", + "[57] Shobhita Sundaram, Stephanie Fu, Lukas Mutenthaler, Netanel Y. Tamir, Lucy Chai, Simon Kornblith, Trevor Darrell, and Phillip Isola. When does perceptual alignment benefit vision representations? In NIPS, 2024. 2", + "[58] Saksham Suri, Matthew Walmer, Kamal Gupta, and Abhinav Shrivastava. Lift: A surprisingly simple lightweight feature transform for dense vit descriptors. In ECCV, pages 110-128. Springer, 2024. 2", + "[59] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. Advances in Neural Information Processing Systems, 36:1363-1389, 2023. 3, 7", + "[60] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 7", + "[61] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 2", + "[62] Narek Tumanyan, Assaf Singer, Shai Bagon, and Tali Dekel. Dino-tracker: Taming dino for self-supervised point tracking in a single video. arXiv preprint arXiv:2403.14548, 2024. 2, 3, 7, 8", + "[63] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 2, 6", + "[64] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural computation, 23(7):1661-1674, 2011. 1", + "[65] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19795-19806, 2023. 2, 8" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[66] Qian Wang, Abdelrahman Eldosokey, Mohit Mendiratta, Fangneng Zhan, Adam Kortylewski, Christian Theobalt, and Peter Wonka. Zero-shot video semantic segmentation based on pre-trained diffusion models, 2024. 2", + "[67] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2", + "[68] Jay Zhangjie Wu, Guian Fang, Haoning Wu, Xintao Wang, Yixiao Ge, Xiaodong Cun, David Junhao Zhang, Jia-Wei Liu, Yuchao Gu, Rui Zhao, et al. Towards a better metric for text-to-video generation. arXiv preprint arXiv:2401.07781, 2024. 2", + "[69] Weilai Xiang, Hongyu Yang, Di Huang, and Yunhong Wang. Denoising diffusion autoencoders are unified self-supervised learners. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15802-15812, 2023. 3", + "[70] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20406-20417, 2024. 2", + "[71] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. ODISE: Open-Vocabulary Panoptic Segmentation with Text-to-Image Diffusion Models. In CVPR, 2023. 2", + "[72] Jiale Yang, So Kanazawa, Masami K Yamaguchi, and Isamu Motoyoshi. Pre-constancy vision in infants. Current Biology, 25(24):3209-3212, 2015. 1", + "[73] Xingyi Yang and Xinchao Wang. Diffusion model as representation learner. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18938-18949, 2023. 3", + "[74] Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is easier than you think. arXiv preprint arXiv:2410.06940, 2024. 2, 3", + "[75] Yuanwen Yue, Anurag Das, Francis Engelmann, Siyu Tang, and Jan Eric Lenssen. Improving 2D Feature Representations by 3D-Aware Fine-Tuning. In ECCV, 2024. 2", + "[76] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818, 2023. 2", + "[77] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 5", + "[78] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145, 2023. 2, 3" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7286", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[79] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19855-19865, 2023. 8", + "[80] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2" + ], + "bbox": [ + 91, + 92, + 480, + 217 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "7287", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Track4Gen_ Teaching Video Diffusion Models to Track Points Improves Video Generation/52b14d4e-8da8-4b5e-a32c-278d64dd36e6_model.json b/2025/Track4Gen_ Teaching Video Diffusion Models to Track Points Improves Video Generation/52b14d4e-8da8-4b5e-a32c-278d64dd36e6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b13837355fbf77809642454bbef8d3ef4b29f413 --- /dev/null +++ b/2025/Track4Gen_ Teaching Video Diffusion Models to Track Points Improves Video Generation/52b14d4e-8da8-4b5e-a32c-278d64dd36e6_model.json @@ -0,0 +1,2622 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.189, + 0.131, + 0.812, + 0.175 + ], + "angle": 0, + "content": "Track4Gen: Teaching Video Diffusion Models to Track Points Improves Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.203, + 0.886, + 0.224 + ], + "angle": 0, + "content": "Hyeonho Jeong\\(^{1,2,*}\\) Chun-Hao P. Huang\\(^{1}\\) Jong Chul Ye\\(^{2}\\) Niloy J. Mitra\\(^{1,3}\\) Duygu Ceylan\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.228, + 0.756, + 0.249 + ], + "angle": 0, + "content": "\\(^{1}\\)Adobe Research \\(^{2}\\)KAIST \\(^{3}\\)University College London" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.283, + 0.327, + 0.299 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.314, + 0.485, + 0.603 + ], + "angle": 0, + "content": "While recent foundational video generators produce visually rich output, they still struggle with appearance drift, where objects gradually degrade or change inconsistently across frames, breaking visual coherence. We hypothesize that this is because there is no explicit supervision in terms of spatial tracking at the feature level. We propose Track4Gen, a spatially aware video generator that combines video diffusion loss with point tracking across frames, providing enhanced spatial supervision on the diffusion features. Track4Gen merges the video generation and point tracking tasks into a single network by making minimal changes to existing video generation architectures. Using Stable Video Diffusion [4] as a backbone, Track4Gen demonstrates that it is possible to unify video generation and point tracking, which are typically handled as separate tasks. Our extensive evaluations show that Track4Gen effectively reduces appearance drift, resulting in temporally stable and visually coherent video generation. Project page: hyeonho99.github.io/track4gen" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.628, + 0.222, + 0.645 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.483, + 0.744 + ], + "angle": 0, + "content": "Diffusion-based video generators [4, 6, 47] are making rapid strides in creating temporally consistent and visually rich video content. This progress marks a significant shift, as the unification of generation and control has the potential to transform the traditional workflow of first capturing and then digitally editing video." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.745, + 0.483, + 0.881 + ], + "angle": 0, + "content": "Despite impressive capabilities, video generators often suffer from appearance drift, where visual elements gradually change, mutate, or degrade over time, causing inconsistencies in the objects. For example, in Fig. 1, we observe the horns of the cow distorting and morphing unrealistically over time, breaking the plausibility of the generated content. This is in striking contrast to humans, who develop a sense of appearance constancy as early as infancy through observation and interaction with the world [72]." + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.281, + 0.679, + 0.292 + ], + "angle": 0, + "content": "Stable Video Diffusion" + }, + { + "type": "image_caption", + "bbox": [ + 0.755, + 0.281, + 0.862, + 0.293 + ], + "angle": 0, + "content": "Track4Gen (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.294, + 0.71, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.294, + 0.905, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.547, + 0.907, + 0.59 + ], + "angle": 0, + "content": "Figure 1. Motivation. Videos generated by Stable Video Diffusion [4] suffer from appearance drift, while those from our method, Track4Gen, are free from such appearance inconsistency issues." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.625, + 0.907, + 0.761 + ], + "angle": 0, + "content": "Unfortunately, appearance drift remains a persistent issue in current video models, even with increased training data and more advanced architectures. We speculate that this limitation arises from supervision being based solely on video diffusion loss (i.e., denoising score matching [64]) in the pixel/latent space, without explicit spatial awareness guidance in the feature space. Hence, in this paper, we ask if and how we can empower video diffusion models with appearance constancy by providing additional supervision." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We present Track4Gen as a spatially aware video generator that receives supervision both in terms of the original diffusion-based objective as well as (dense) point correspondence across frames, which we refer to as tracks. We demonstrate that it is possible to provide such track-level supervision in the diffusion feature space by making minimal architecture changes. Our generated videos do not suffer from degradation of video quality (according to the usual video generation metrics), while being significantly more" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.332, + 0.901 + ], + "angle": 0, + "content": "*Work done during internship at Adobe." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7276" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.411, + 0.106 + ], + "angle": 0, + "content": "spatially coherent as the highlight cow in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.107, + 0.482, + 0.258 + ], + "angle": 0, + "content": "We train Track4Gen using the latest Stable Video Diffusion [4] as the backbone and evaluate on the publicly available VBench dataset [30]. We report significant improvement in terms of appearance constancy of subjects, both in quantitative and qualitative (i.e., via user studies) evaluations. In summary, we demonstrate that it is possible to upgrade existing video generators, by supervising them with additional correspondence tracking loss, to produce videos without significant appearance drifts, a problem commonly encountered in diffusion-based video generators." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.269, + 0.233, + 0.285 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.295, + 0.483, + 0.505 + ], + "angle": 0, + "content": "Diffusion-based video generation. Building on the success of diffusion models in image synthesis [11, 51], diffusion-based video generators have seen significant advancements [4, 6, 29, 47]. A commonly adopted approach is to extend text-to-image models to the video domain by incorporating temporal layers to facilitate interactions across video frames [5, 22, 54]. While some works have adopted cascaded approaches to produce both spatially and temporally high-resolution videos [28, 47, 54, 67, 76, 78], others have utilized lower-dimensional latent space modeling to reduce computational demands [5, 8, 24, 80]. We build on top of one such approach, Stable Video Diffusion (SVD, [4]), which introduces a latent image-to-video diffusion model trained on a large-scale and curated video data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.483, + 0.657 + ], + "angle": 0, + "content": "With advances in generation, systematic evaluation of generation quality has become crucial. Traditionally, metrics such as Fréchet Inception Distance (FID, [26]), Fréchet Video Distance (FVD, [63]), and CLIPSIM [49] are used. Additionally, comprehensive benchmark suites [30, 68] have been introduced to provide a more robust evaluation aligned with human perception. Inspired by such work, we thoroughly evaluate our approach and demonstrate improved video generation quality with respect to both conventional metrics and the recent VBenchmark metrics [30]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Foundational models as feature extractors. Various foundational models such as vision transformers [15] or diffusion-based generators [50] have been utilized as feature extractors for various tasks including semantic matching [16, 25, 41], classification [38], segmentation [66, 71], and editing [19, 21, 61]. There have been efforts to boost their performance by post-processing the feature maps obtained from the pre-trained models, e.g., by upsampling [18, 58]. In a recent effort, Yue et al. [75] lift semantic per-frame features from a foundational model into a 3D Gaussian representation. They fine-tune the foundational model with such 3D-aware features resulting in improved performance in downstream tasks. Similarly, Sundaram et al. [57] fine-tune state-of-the-art foundational models on human similarity judgments yielding improved representations across" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "downstream tasks. In a concurrent effort, Yu et al. [74] propose to align the internal features of an image generation model with external discriminative features [45], which results in more effective training of the generator." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.907, + 0.319 + ], + "angle": 0, + "content": "Our work also enhances the internal feature representation of a foundational generation model but with significant differences compared to previous literature. First, unlike most previous work that focus on image level foundational models, we exploit the power of recently emerging video models. Second, instead of post-processing, we enhance the spatial awareness of the intermediate features by training the generator to jointly perform an additional tracking task. We show that this joint training boosts the performance of intermediate features in correspondence tracking, leading to improved video generation quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.336, + 0.908, + 0.744 + ], + "angle": 0, + "content": "Tracking any point in a video. The task involves following any arbitrary query point across a long video sequence. First introduced by PIPs [23] and later re-framed by TAPVid [12], several methods have emerged in recent years to tackle long-term point tracking. PIPs [23] revisits the classical particle-based representation [53] and introduces MLP-based networks that predict point tracks within an 8-frame window. Subsequent works have improved performance by capturing longer temporal context through advanced architectures [2, 13, 23, 33], as well as by enabling the simultaneous tracking of multiple queries [10, 33]. More recent training-based trackers [10, 34, 40, 70] have achieved remarkable performance by leveraging high-capacity neural networks to learn robust priors from large-scale training data. While high-quality data is crucial for accurate tracking, manually annotating point tracks is prohibitively expensive. Hence, synthetic videos [20] with automatic annotations, have become an alternative and have demonstrated effectiveness in real-world video tracking. An alternative approach is self-supervised adaptation at test time, where tracking is learned without ground-truth labels [32, 62, 65]. In a recent study, Aydemir et al. [1] evaluate the effectiveness of several image foundational model features for point tracking both in zero-shot setting as well as with supervised training using low-rank adapter layers. To the best of our knowledge, we are the first to exploit the features of a foundational video diffusion model for dense point tracking." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.755, + 0.605, + 0.77 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.905, + 0.903 + ], + "angle": 0, + "content": "In this section, we provide a comprehensive discussion of the Track4Gen framework. We begin with a concise overview of latent video diffusion models (Sec. 3.1). Next, we discuss how video diffusion features relate to temporal correspondences both for real and generated videos (Sec. 3.2). Finally, we detail the design of Track4Gen both in terms of network architecture and the employed supervision signals (Sec. 3.3). An overview is depicted in Fig. 2." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7277" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.087, + 0.895, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.287, + 0.908, + 0.318 + ], + "angle": 0, + "content": "Figure 2. Track4Gen overview. Red-colored blocks represent layers optimized by the diffusion loss \\(\\mathcal{L}_{\\mathrm{diff}}\\), while green blocks are optimized by the correspondence loss \\(\\mathcal{L}_{\\mathrm{corr}}\\). Blocks colored both red and green are influenced by the joint loss, \\(\\mathcal{L}_{\\mathrm{diff}} + \\lambda \\mathcal{L}_{\\mathrm{corr}}\\). See text for details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.341, + 0.406, + 0.358 + ], + "angle": 0, + "content": "3.1. Background: Stable Video Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.366, + 0.483, + 0.532 + ], + "angle": 0, + "content": "Starting from random Gaussian noise, diffusion models aim to generate clean images or videos via an iterative denoising process [27, 55]. This process reverses a fixed, time-dependent diffusion forward process, which gradually corrupts the data by adding Gaussian noise. While our method is applicable to general video diffusion models, in this paper, we design our architecture based on Stable Video Diffusion (SVD), a latent video diffusion model which employs the EDM-framework [35]. The diffusion process operates in the lower-dimensional latent space of a pre-trained VAE [37], consisting of an encoder \\(\\mathcal{E}(\\cdot)\\) and a decoder \\(\\mathcal{D}(\\cdot)\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.484, + 0.657 + ], + "angle": 0, + "content": "Given a clean sample \\( \\pmb{x}_0^{1:N} \\sim p_{\\mathrm{data}}(\\pmb{x}) \\) of an \\( N \\)-frame video sequence, the frames are first encoded into the latent space as \\( z_0^{1:N} = \\mathcal{E}(x_0^{1:N}) \\). Gaussian noise \\( \\epsilon \\sim \\mathcal{N}(0,I) \\) is then added to the latents to produce the intermediate noisy latents via the forward process \\( z_t^{1:N} = \\alpha_t z_0^{1:N} + \\sigma_t \\epsilon \\), where \\( t \\) represents the diffusion timestep, and \\( \\alpha_t, \\sigma_t \\) are the discretized noise scheduler parameters. The diffusion denoiser \\( f_\\theta \\) is trained by minimizing the v-prediction loss:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.673, + 0.483, + 0.7 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {\\boldsymbol {\\epsilon} \\sim \\mathcal {N} (0, I), t \\sim U [ 1, T ]} \\left[ \\left\\| \\boldsymbol {f} _ {\\theta} \\left(\\boldsymbol {z} _ {t} ^ {1: N}, t, c\\right) - \\boldsymbol {y} \\right\\| _ {2} ^ {2} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.716, + 0.483, + 0.793 + ], + "angle": 0, + "content": "where \\(\\pmb{y}\\) is defined as \\(\\pmb{y} = \\alpha_{t}\\pmb{\\epsilon} - \\sigma_{t}\\pmb{z}_{0}^{1:N}\\). In the image-to-video variant of SVD, the condition \\(c\\) refers to the CLIP image embedding [49], replacing the typical text embeddings. For the remainder of this paper, we will refer to Eq. 1 as the video diffusion loss \\(\\mathcal{L}_{\\mathrm{diff}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Once trained, the diffusion model generates videos by iteratively denoising a noisy latent \\( \\pmb{z}_T^{1:N} \\) sequence sampled from pure Gaussian distribution. At each diffusion step, the model predicts the noise in the input latent. Once the clean latent \\( \\pmb{z}_0^{1:N} \\) is obtained, the decoder \\( \\mathcal{D} \\) maps it to the higher-dimensional pixel space \\( \\pmb{x}_0^{1:N} = \\mathcal{D}(\\pmb{z}_0^{1:N}) \\). For further details, we refer to the Appendix D of [4]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.341, + 0.741, + 0.357 + ], + "angle": 0, + "content": "3.2. Video Diffusion Features" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.366, + 0.907, + 0.488 + ], + "angle": 0, + "content": "Previous studies have demonstrated that image diffusion models learn discriminative features in their hidden states that are effective for various analysis tasks and propose methods for improving the representation power of such features [9, 69, 73, 74]. Similarly, we argue that while also being powerful, internal representations of pre-trained video diffusion models may not be fully temporally consistent, resulting in appearance drift in generated videos." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.909, + 0.687 + ], + "angle": 0, + "content": "To better investigate this hypothesis, we first evaluate the long-term video tracking capabilities of U-Net-based video diffusion models [4, 56, 78]. Specifically, we evaluate the effectiveness of the features from each block of the U-Net for the task of point tracking. Given a real-world video, we add a small amount of noise and extract feature maps from each layer in each block. We perform a cosine-similarity-based nearest-neighbor search [44, 59] over these feature maps for a given set of fixed query points on the first frame (we use a similarity threshold of 0.6 [62] in our experiments). We also perform a similar analysis for generated videos where we extract the feature maps corresponding to diffusion steps with small amount of noise." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Based on this feature analysis, we make some important observations. Notably, regardless of the model (we analyze both Zeroscope T2V [56] and SVD I2V [4]), we find out that output features from the upsampler layer of the third decoder block consistently yield stronger temporal correspondences, as shown in Fig. 3. Hence, we use this block when extracting features for the remainder of our experiments. Furthermore, when we analyze generated videos and point tracks estimated based on the feature maps (as shown in Fig. 4), we observe that there is a correlation between tracking failures that reveal feature-space inconsistencies and appearance drifts that reveal pixel-space inconsistencies. Hence, we hypothesize that enriching feature consistency can help mitigate such appearance drifts. Next, we" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7278" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.087, + 0.898, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.256, + 0.908, + 0.3 + ], + "angle": 0, + "content": "Figure 3. Real-world video tracking using different video diffusion features. Given color-coded query points on the first frame (Leftmost column), we display tracked points on target frames using features from different blocks (right columns). The 13th frame (first row) and 8th frame (second row) are shown as target frames. Full results are available in the supplementary and on our page." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.483, + 0.357 + ], + "angle": 0, + "content": "introduce Track4Gen where we accomplish this goal by supervising video diffusion models with a joint tracking loss." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.367, + 0.216, + 0.383 + ], + "angle": 0, + "content": "3.3. Track4Gen" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.39, + 0.483, + 0.631 + ], + "angle": 0, + "content": "Track4Gen aims to utilize point tracking as an additional supervision signal to enhance the spatial-awareness of video diffusion features. Given that we build on top of a pretrained video generation model, to retain the prior knowledge and avoid tampering the original features directly, we propose a novel architecture change as shown in Fig. 2. Specifically, instead of directly using the raw diffusion features for correspondence estimation, we propose a trainable refiner module \\( R_{\\phi} \\), which is designed to refine the raw features by projecting them into a correspondence-rich feature space. The refined features, which are spatially-aware, are then both used to estimate point tracks with an explicit supervision as well as feeding back to the generation backbone. We empirically find out that this design is more effective compared to fine-tuning the original model with no refinement module (see Sec. 4.2)." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.633, + 0.483, + 0.65 + ], + "angle": 0, + "content": "Given an \\(N\\)-frame video sequence \\(\\pmb{x}_0^{1:N}\\), its correspond-" + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.673, + 0.269, + 0.684 + ], + "angle": 0, + "content": "Feature-space Inconsistency (SVD)" + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.674, + 0.473, + 0.683 + ], + "angle": 0, + "content": "Feature-space Consistency (Track4Gen)" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.684, + 0.288, + 0.852 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.684, + 0.481, + 0.852 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.855, + 0.483, + 0.897 + ], + "angle": 0, + "content": "Figure 4. Generated video tracking using video diffusion features. Tracks based on diffusion features are annotated on the generated videos. Track4Gen generates more consistent results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.324, + 0.905, + 0.385 + ], + "angle": 0, + "content": "ing latent \\( z_0^{1:N} \\), and a diffusion timestep \\( t \\), in order to train Track4Gen we continue to utilize the standard diffusion training loss as defined in Eq. (1), where we adopt the velocity prediction objective [4, 35, 52] for \\( \\mathcal{L}_{\\mathrm{diff}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.386, + 0.906, + 0.541 + ], + "angle": 0, + "content": "To enable tracking supervision, we assume access to a dense set of point trajectories \\(\\Omega = \\{(x^i,x^j)\\}\\) across frames where a point \\(\\mathbf{x}^i\\) in frame \\(i\\) corresponds to a matching point \\(\\mathbf{x}^j\\) in frame \\(j\\) and vice versa. Given the corresponding noisy video latent sequence \\(\\pmb{z}_t^{1:N}\\), we first extract raw diffusion features as the hidden states \\(\\pmb{h}^{1:N} \\in \\mathbb{R}^{N \\times H \\times W \\times C}\\) from a specific block \\(b^k\\) within the U-Net, where \\(b^k\\) is set to the upsampler layer of the third decoder block (see Sec. 3.2). We then pass these features through the refiner module to obtain the refined feature map \\(\\tilde{\\pmb{h}}^{1:N} = \\pmb{R}_{\\phi}(\\pmb{h}^{1:N})\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.542, + 0.906, + 0.629 + ], + "angle": 0, + "content": "We sample a query point \\(\\mathbf{x}_{\\mathrm{q}}^{i}\\) along with its ground-truth target point \\(\\mathbf{x}_{\\mathrm{trg}}^{j}\\) from the correspondence set \\(\\Omega\\). Given the query point feature \\(\\tilde{\\pmb{h}}^i (\\mathrm{x_q})\\in \\mathbb{R}^{1\\times 1\\times C}\\) and the target feature map \\(\\tilde{\\pmb{h}}^j\\in \\mathbb{R}^{H\\times W\\times C}\\), we calculate the cost volume \\(S\\in\\) \\(\\mathbb{R}^{H\\times W\\times 1}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.641, + 0.905, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {S} (\\mathrm {p}) = \\cos - \\operatorname {s i m} \\left(\\tilde {\\boldsymbol {h}} ^ {i} \\left(\\mathrm {x} _ {\\mathrm {q}}\\right), \\tilde {\\boldsymbol {h}} ^ {j} (\\mathrm {p})\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.908, + 0.721 + ], + "angle": 0, + "content": "where cos-sim denotes cosine similarity. The predicted target point \\(\\hat{\\mathbf{x}}_{\\mathrm{trg}}\\) is then determined using the differentiable soft-argmax operation:" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.732, + 0.905, + 0.771 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} _ {\\mathrm {t r g}} = \\frac {\\sum_ {\\mathrm {p} \\in \\Omega^ {\\prime}} \\boldsymbol {S} (\\mathrm {p}) \\cdot \\mathrm {x} _ {\\mathrm {p}}}{\\sum_ {\\mathrm {p} \\in \\Omega^ {\\prime}} \\boldsymbol {S} (\\mathrm {p})}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.783, + 0.907, + 0.862 + ], + "angle": 0, + "content": "where \\(\\Omega^{\\prime} = \\{p:\\| \\mathrm{x_p - x_{p_{max}}}\\| _2\\leq R\\} ^1\\) . Thus, the target point prediction can be expressed as \\(\\hat{\\mathbf{x}}_{\\mathrm{trg}} = \\xi (\\mathrm{x}_{\\mathrm{q}}^{i},j,\\tilde{\\boldsymbol{h}}^{1:N})\\) and the predicted tracklet for \\(\\mathrm{x_q^i}\\) is given by \\(\\mathcal{T}_{\\mathrm{x_q^i}} = \\{\\hat{\\mathbf{x}}_n\\) .. \\(\\hat{\\mathbf{x}}_n = \\xi (\\mathbf{x}_q^i,n,\\tilde{\\pmb{h}}^{1:N}),n = 1,\\dots,N\\}\\) Finally, the correspond" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.908, + 0.9 + ], + "angle": 0, + "content": "The feature maps have a resolution of \\(44 \\times 81\\) for an input video resolution of \\(320 \\times 576\\), and we set \\(R = 35\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7279" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.089, + 0.207, + 0.1 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.089, + 0.493, + 0.099 + ], + "angle": 0, + "content": "Stable Video Diffusion (SVD)" + }, + { + "type": "image_caption", + "bbox": [ + 0.693, + 0.089, + 0.788, + 0.099 + ], + "angle": 0, + "content": "Track4Gen (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.101, + 0.256, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.101, + 0.58, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.101, + 0.904, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.453, + 0.871, + 0.468 + ], + "angle": 0, + "content": "Figure 5. Image-to-video generation results of the original SVD and Track4Gen. Please visit our page for full video view." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.494, + 0.483, + 0.51 + ], + "angle": 0, + "content": "dence loss \\(\\mathcal{L}_{\\mathrm{corr}}\\) is computed using the Huber loss \\(L_{H}\\) [31]:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.517, + 0.483, + 0.558 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o r r}} \\left(\\tilde {\\boldsymbol {h}} ^ {1: N}, \\Omega\\right) = \\sum_ {\\left(\\mathrm {x} _ {\\mathrm {q}} ^ {i}, \\mathrm {x} _ {\\text {t r g}} ^ {j}\\right) \\in \\Omega} L _ {H} \\left(\\xi \\left(\\mathrm {x} _ {\\mathrm {q}} ^ {i}, j, \\tilde {\\boldsymbol {h}} ^ {1: N}\\right), \\mathrm {x} _ {\\text {t r g}} ^ {j}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.484, + 0.736 + ], + "angle": 0, + "content": "When training Track4Gen, we initialize the refiner module as an identity mapping to fully leverage the prior of the base model at the start of finetuning. To re-route the refined features to the backbone generator, we introduce a trainable zero convolution layer [77], denoted as \\(\\zeta_{\\psi}\\). While the diffusion loss \\(\\mathcal{L}_{\\mathrm{diff}}\\) back-propagates to all the blocks of the video diffusion model, we detach the gradients of \\(\\tilde{\\pmb{h}}^{1:N}\\) before passing into \\(\\zeta_{\\psi}\\) such that refiner module can solely focus on acquiring the correspondence prior. Hence, given that the output of block \\(b^{k}\\) is \\(h^{1:N}\\), the input to the subsequent block \\(b^{k+1}\\) is computed as \\(h^{1:N} + \\zeta_{\\psi}\\) (stop-gradient \\((R_{\\phi}(h^{1:N}))\\))." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.736, + 0.483, + 0.766 + ], + "angle": 0, + "content": "Fig. 2 visualizes this architecture design, with red and green colors indicating the objective that optimizes each module." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.779, + 0.224, + 0.796 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.308, + 0.819 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To train Track4Gen, we construct a training dataset consisting of 567 video-trajectory pairs, with each video having a resolution of \\(320 \\times 576\\) and a duration of 24 frames. Since no real-world video with (dense) ground-truth trajectory annotations exist at the time of this work, we utilize" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.907, + 0.6 + ], + "angle": 0, + "content": "optical flow to generate trajectory annotations. A key challenge is the need for accurate video segmentation maps to ensure a balanced distribution of trajectory points between foreground objects and the background [12]. To address this, we utilize public video datasets paired with ground-truth segmentation maps [7, 17, 39, 46, 48], where we split longer videos into 24-frame segments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.601, + 0.909, + 0.86 + ], + "angle": 0, + "content": "We use Stable Video Diffusion (SVD) image-to-video pretrained checkpoints\\(^2\\) as the base video generator. Our proposed refiner module consists of eight stacked 2D convolution layers and is attached to the third decoder block of the SVD UNet. The refiner module preserves the shape of the hidden states throughout and is initialized as the identity mapping. Further details are provided in the supplementary. We finetune this enhanced video generator architecture for \\(20\\mathrm{K}\\) steps with our joint loss \\(\\mathcal{L}_{\\mathrm{diff}} + \\lambda \\mathcal{L}_{\\mathrm{corr}}\\), where \\(\\lambda\\) is set to 8. Rather than finetuning the entire model, we finetune only the temporal transformer blocks, the refiner module \\(R_{\\phi}\\), and the zero convolution \\(\\zeta_{\\psi}\\). In each iteration, we sample 512 correspondence pairs from the precomputed trajectories. We use the AdamW optimizer [43] with a learning rate of \\(1e - 5\\), \\(\\beta_{1} = 0.9\\), \\(\\beta_{2} = 0.999\\), and a weight decay of \\(1e - 2\\). We train the model on \\(4\\times\\) H100 GPUs with a total batch size of 4. For sampling new videos, we apply the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.905, + 0.901 + ], + "angle": 0, + "content": "2https://huggingface.co/stabilityai/stable-video-diffusion-image2vid-xt" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7280" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.087, + 0.907, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.396, + 0.907, + 0.425 + ], + "angle": 0, + "content": "Figure 6. Qualitative ablation on video generation. Track4Gen is compared with finetuned SVD (SVD finetuned on the same training videos without any correspondence supervision) and Track4Gen trained without the Refiner module." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.451, + 0.483, + 0.482 + ], + "angle": 0, + "content": "default settings using 30 steps with the EDM sampler [35], motion bucket id =127, and fps=7." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.494, + 0.484, + 0.577 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison on video generation performance. We compare Track4Gen to the pre-trained SVD \\(\\star\\) as well as a finetuned SVD on the same dataset (finetuned SVD). We also train a variant of Track4Gen without the refiner module. All videos are generated at \\(320\\mathrm{x}576\\) resolution, except SVD \\(\\star\\) (576p) which operates at \\(576\\mathrm{x}1024\\) resolution." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.58, + 0.484, + 0.643 + ], + "angle": 0, + "content": "
Subject ConsistencyTemporal FlickeringMotion SmoothnessImaging QualityVideo-Image AlignmentFIDFVD
SVD*0.95350.94640.97740.66480.953929.0776
finetuned SVD0.96650.98000.99090.67660.977127.0735
Track4Gen w/o refiner0.95060.97250.97910.66530.961427.1718
Track4Gen0.97460.98060.99210.68350.981426.6724
SVD* (576p)0.95760.94780.97950.68120.9582--
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.668, + 0.383, + 0.683 + ], + "angle": 0, + "content": "4.2. Track4Gen for Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.734 + ], + "angle": 0, + "content": "We evaluate Track4Gen for the image-to-video generation task via a series of experiments using multiple datasets, automated metrics, and human evaluations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Evaluation Setup. We compare Track4Gen against the original SVD \\((\\mathrm{SVD}^*)\\) [4], as well as a version of SVD that is finetuned on the same videos as Track4Gen (finetuned SVD). Furthermore, we train a variant of Track4Gen without the refiner module. For VBenchmark metrics [30], evaluations are conducted on the VBenchmark-I2V dataset, containing 355 diverse images. FID and FVD are measured using the DAVIS [48] dataset as reference. We generate 24-frame videos conditioned on each input image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Automatic metrics. We first report five key metrics from VBench [30]: (1) Subject Consistency—assesses" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.451, + 0.907, + 0.633 + ], + "angle": 0, + "content": "subject appearance consistency of the video by computing the similarity of DINO [45] features. (2) Temporal Flickering—detects temporal consistency by taking static frames and calculating the mean absolute difference across frames. (3) Motion Smoothness—measures smoothness of motion, and how well it adheres to real-world physics, using video frame interpolation priors [42]. (4) Image Quality—evaluates distortions (e.g., noise, blur) using a pretrained, multi-scale image quality predictor [36]. (5) Video-Image Alignment—measures alignment between the subject in the input image and in the generated video using DINO features. We additionally report FID [26] and FVD [63]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.638, + 0.907, + 0.745 + ], + "angle": 0, + "content": "Human evaluation. We further evaluate Track4Gen against baselines through a user study. We ask 64 participants to compare our results with a randomly selected baseline. We ask the users to evaluate how consistent main objects appear across the frames in a generated video as well as how natural the depicted motion is. We provide further details of the user study in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Qualitative results. Qualitative comparisons with the base SVD are shown in Fig. 5. As illustrated, Track4Gen generates videos with strong appearance consistency, avoiding issues of appearance drift. In contrast, videos produced by the original SVD exhibit noticeable inconsistencies: the sheep's head (row 1) mutates, the plane's wing (row 2) shows unnatural transitions, and the cars (row 3) disappear. Further comparisons with finetuned SVD and Track4Gen without the refiner module are shown in Fig. 6 and highlight the superior visual coherence of the proposed Track4Gen." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.958 + ], + "angle": 0, + "content": "7281" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.158, + 0.089, + 0.232, + 0.1 + ], + "angle": 0, + "content": "Query Points" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.101, + 0.296, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.348, + 0.089, + 0.45, + 0.1 + ], + "angle": 0, + "content": "Track4Gen (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.101, + 0.5, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.089, + 0.627, + 0.099 + ], + "angle": 0, + "content": "SVD 12V" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.101, + 0.702, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.758, + 0.089, + 0.848, + 0.099 + ], + "angle": 0, + "content": "ZeroScope T2V" + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.101, + 0.905, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.349, + 0.905, + 0.378 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison of Track4Gen and baselines for real-world video tracking. The leftmost column displays query points in the first frame, while the following three columns show tracking results using features from each model." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.39, + 0.28, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.491, + 0.25, + 0.503 + ], + "angle": 0, + "content": "(a) Identity preservation" + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.392, + 0.481, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.33, + 0.491, + 0.451, + 0.502 + ], + "angle": 0, + "content": "(b) Motion naturalness" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.506, + 0.483, + 0.534 + ], + "angle": 0, + "content": "Figure 8. User study results. Our study shows that Track4Gen better preserves object identity and produces more natural motion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.547, + 0.484, + 0.653 + ], + "angle": 0, + "content": "Quantitative results. As shown in Tab. 1, our method achieves the highest scores across all 5 metrics from VBench, along with the lowest FID and second-lowest FVD values, outperforming the base SVD by substantial margins. Fig. 8 provides the user study results where the majority of the participants agreed that Track4Gen is superior both in terms of identity preservation and naturalness of motion." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.661, + 0.365, + 0.678 + ], + "angle": 0, + "content": "4.3. Track4Gen for Video Tracking" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.681, + 0.483, + 0.803 + ], + "angle": 0, + "content": "We evaluate Track4Gen's capability to track any point in real videos by adding a small amount of noise to the input video [59] and passing it through the video denoiser \\( \\pmb{f}_{\\theta} \\) to extract feature maps. We first compare tracking results with such features against other raw features [4, 56, 60] in Sec. 4.3.1. In Sec. 4.3.2, we utilize Track4Gen's features in a test-time optimization method [62] and compare to both self-supervised and fully supervised video trackers." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.807, + 0.353, + 0.823 + ], + "angle": 0, + "content": "4.3.1. Zero-shot Feature Comparison" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We evaluate the precision of predicted tracks using the features from Track4Gen, the original SVD model \\((\\mathrm{SVD}^{\\star})\\) and RAFT [60]. We also test another text-to-video model, ZeroScope T2V [56], to demonstrate how raw features from pre-trained video generators typically work out of the box." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.393, + 0.906, + 0.449 + ], + "angle": 0, + "content": "Table 2. Quantitative zero-shot feature comparison on video tracking benchmarks. Track4Gen features are compared to the features of SVD* [4], ZeroScope [56], and RAFT [60]. For all the metrics, higher values indicate better performance." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.452, + 0.907, + 0.527 + ], + "angle": 0, + "content": "
MethodDAVIS-480p(24-frame)BADJA(24-frame)DAVIS-480p(whole duration)BADJA(whole duration)
\\( \\delta^{x}_{avg} \\)OAAJ\\( \\delta^{seg} \\)\\( \\delta^{3px} \\)\\( \\delta^{x}_{avg} \\)OAAJ\\( \\delta^{seg} \\)\\( \\delta^{3px} \\)
ZeroScope46.267.039.427.52.837.259.527.819.92.0
SVD*42.479.736.426.22.935.470.126.519.42.2
Track4Gen69.785.856.552.37.758.978.440.240.45.0
RAFT73.3--54.88.766.7--45.05.8
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.906, + 0.592 + ], + "angle": 0, + "content": "For RAFT, tracking is achieved by chaining optical flow displacements, while the others use nearest neighbor matching between its encoded features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.906, + 0.687 + ], + "angle": 0, + "content": "Datasets. We use TAP-Vid DAVIS [12] and BADJA [3] as benchmark datasets. Additionally, we include two shorter benchmarks, DAVIS (24-frame) and BADJA (24-frame), which focus on the first 24 frames with query and target points within this range. Details on encoding long videos with the video models are in the supplemental." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Metrics. For evaluating the TAP-Vid benchmarks, we use the following metrics: (i) Position Accuracy \\((\\delta_{avg}^{x})\\) evaluates the average accuracy of visible points, where each \\(\\delta^x\\) represents the fraction of predicted points that lie within \\(x\\) pixels of the ground-truth position, with \\(x \\in \\{1, 2, 4, 8, 16\\}\\). (ii) Occlusion Accuracy (OA) evaluates the correctness of occlusion predictions. (iii) Average Jaccard (AJ) jointly assesses both position and occlusion accuracy. For the BADJA dataset, we report \\(\\delta^{seg}\\), which measures the accuracy of tracked keypoints within a distance of \\(0.2\\sqrt{A}\\) from the ground-truth annotation, where \\(A\\) is the area of the foreground object. We also report \\(\\delta^{3px}\\), which assesses accuracy within a 3-pixel threshold. A cosine similarity threshold of 0.6 is used for occlusion prediction." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7282" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.089, + 0.209, + 0.097 + ], + "angle": 0, + "content": "First Frame with Query Points" + }, + { + "type": "image_caption", + "bbox": [ + 0.286, + 0.089, + 0.418, + 0.096 + ], + "angle": 0, + "content": "Tracking Results in Subsequent Frames" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.097, + 0.223, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.097, + 0.481, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.268, + 0.482, + 0.283 + ], + "angle": 0, + "content": "Figure 9. Extending Track4Gen with test-time adaptation [62]." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.295, + 0.483, + 0.366 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison with video trackers. Although primarily designed for video generation, Track4Gen combined with a test-time optimization method [62] achieves performance comparable to dedicated video tracking frameworks, even when compared to supervised methods." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.369, + 0.482, + 0.478 + ], + "angle": 0, + "content": "
MethodDAVIS-480BADJA
δxavgOAAJδsegδ3px
TAP-Net*66.479.046.045.49.6
PIPs++*73.6--59.09.8
TAPIR*77.389.565.768.710.5
Omnimotion†74.184.558.445.26.9
DINO-Tracker†80.488.164.672.414.3
DINO-Tracker w/ Track4Gen†72.584.555.748.410.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.094, + 0.479, + 0.28, + 0.49 + ], + "angle": 0, + "content": "\\* - supervised. -test-time training." + }, + { + "type": "table_caption", + "bbox": [ + 0.125, + 0.498, + 0.447, + 0.512 + ], + "angle": 0, + "content": "Table 4. Ablation on trainable modules and refiner." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.515, + 0.482, + 0.603 + ], + "angle": 0, + "content": "
Trainable modulesSubject ConsistencyTemporal FlickeringMotion SmoothnessImaging QualityVideo-Image Alignment
spatial + temporal0.97340.98110.99170.68630.9807
spatial0.97260.98010.99190.68520.9811
temporal0.97460.98060.99210.68350.9814
Refiner architecture
2D convolutions0.97460.98060.99210.68350.9814
3D convolutions0.96870.97340.99040.68330.9820
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.483, + 0.704 + ], + "angle": 0, + "content": "Results. We present the qualitative results in Fig. 7 and the quantitative results in Tab. 2. Although primarily designed for video generation, Track4Gen boosts the poor performance of the pre-trained video models significantly, approaching the accuracy of RAFT optical flow chaining." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.714, + 0.478, + 0.73 + ], + "angle": 0, + "content": "4.3.2. Extending Track4Gen with Test-time Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.483, + 0.855 + ], + "angle": 0, + "content": "To further evaluate Track4Gen's long-term tracking capabilities, we integrate our features with test-time adaptation algorithm of DINO-Tracker [62], where a per-video optimization is performed using optical flow supervision. We replace the originally used DINOv2 [45] with the features from Track4Gen. We evaluate using the same datasets and metrics outlined in Sec. 4.3.1, against both fully-supervised trackers [12, 13, 79] and self-supervised methods [62, 65]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Tab. 3 shows that Track4Gen features optimized with [62] achieve performance comparable to dedicated trackers. Qualitative results are in Fig. 9 and in the supplemental." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.131 + ], + "angle": 0, + "content": "Table 5. Quantitative ablation on using annotated, but synthetic videos [20]. Left: Video generation metrics. Right: Video tracking metrics." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.135, + 0.905, + 0.187 + ], + "angle": 0, + "content": "
Dataset compositionSubject ConsistencyMotion SmoothnessImaging QualityBADJA
δsegδ3px
real videos0.97470.99210.683340.45.0
real + synthetic videos0.97080.98920.679342.14.8
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.205, + 0.679, + 0.219 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.227, + 0.907, + 0.485 + ], + "angle": 0, + "content": "We present an ablation study in Tab. 4 where we train different set of modules. Each spatio-temporal block of SVD includes both spatial and temporal transformers. We compare training only spatial transformers, only temporal transformers, or both. We also ablate the architecture of the refiner module using either 2D or 3D convolution layers. Our analysis shows that while results are similar across settings, training only the temporal transformers in SVD with 2D convolutions as the refiner module yields optimal video generation quality. We further analyze our training dataset by additionally incorporating Kubric [20] simulated videos (1K video-track pairs from the Panning MOVi-E data [10, 13]) with automatically annotated trajectories into training. As shown in Tab. 5, optical flow-chained tracklets from real provides provide as effective correspondence guidance as tracklets from synthetic data, while synthetic videos negatively impact the video generation quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.498, + 0.783, + 0.513 + ], + "angle": 0, + "content": "5. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.906, + 0.644 + ], + "angle": 0, + "content": "We have presented the first unified framework that bridges two distinct tasks: video generation and dense point tracking. We demonstrated that this produces temporally consistent feature representations and appearance-consistent videos. As for limitations, videos generated by Track4Gen tend to exhibit less dynamic motion compared to those from other video generators. Additionally, failure cases are included in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.644, + 0.906, + 0.734 + ], + "angle": 0, + "content": "Future work. Recently, cutting-edge video trackers [10, 14, 34] have emerged, enabling dense, accurate, and long-term tracking, especially with better handling of occlusions. This opens up promising future directions for extending our work to utilize real-world videos, automatically annotated by these advanced trackers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments. We thank Seokju Cho and Narek Tumanyan for their invaluable feedback on video point tracking. We also extend our gratitude to Mingi Kwon, Joon-Young Lee, and Gabriel Huang for their insightful discussions. Hyeonho Jeong and Jong Chul Ye are supported by the National Research Foundation of Korea (NRF) under Grants RS-2024-00336454 and RS-2023-00262527, and by the Institute for Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (RS-2019-II190075, Artificial Intelligence Graduate School Program, KAIST)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7283" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.156 + ], + "angle": 0, + "content": "[1] Gorkay Aydemir, Weidi Xie, and Fatma Güney. Can visual foundation models achieve long-term point tracking? arXiv preprint arXiv:2408.13575, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.158, + 0.484, + 0.212 + ], + "angle": 0, + "content": "[2] Weikang Bian, Zhaoyang Huang, Xiaoyu Shi, Yitong Dong, Yijin Li, and Hongsheng Li. Context-tap: Tracking any point demands spatial context features. arXiv preprint arXiv:2306.02000, 3, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.215, + 0.484, + 0.297 + ], + "angle": 0, + "content": "[3] Benjamin Biggs, Thomas Roddick, Andrew Fitzgibbon, and Roberto Cipolla. Creatures great and small: Recovering the shape and motion of animals from video. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part V 14, pages 3-19. Springer, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.483, + 0.38 + ], + "angle": 0, + "content": "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.483, + 0.464 + ], + "angle": 0, + "content": "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.466, + 0.483, + 0.534 + ], + "angle": 0, + "content": "[6] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.536, + 0.483, + 0.591 + ], + "angle": 0, + "content": "[7] Sergi Caelles, Alberto Montes, Kevis-Kokitsi Maninis, Yuhua Chen, Luc Van Gool, Federico Perazzi, and Jordi Pont-Tuset. The 2018 davis challenge on video object segmentation. arXiv preprint arXiv:1803.00557, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.483, + 0.661 + ], + "angle": 0, + "content": "[8] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.663, + 0.483, + 0.703 + ], + "angle": 0, + "content": "[9] X Chen, Z Liu, S Xie, and K He. Deconstructing denoising diffusion models for self-supervised learning. arxiv 2024. arXiv preprint arXiv:2401.14404. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[10] Seokju Cho, Jiahui Huang, Jisu Nam, Honggyu An, Seun-gryong Kim, and Joon-Young Lee. Local all-pair correspondence for point tracking. arXiv preprint arXiv:2407.15420, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[11] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[12] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610-13626, 2022. 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "Tapir: Tracking any point with per-frame initialization and temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10061-10072, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[14] Carl Doersch, Yi Yang, Dilara Gokay, Pauline Luc, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ross Goroshin, João Carreira, and Andrew Zisserman. Bootstrap: Bootstrapped training for tracking-any-point. arXiv preprint arXiv:2402.00847, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.906, + 0.303 + ], + "angle": 0, + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.305, + 0.906, + 0.375 + ], + "angle": 0, + "content": "[16] Niladri Shekhar Dutt, Sanjeev Muralikrishnan, and Niloy J. Mitra. Diffusion 3d features (diff3f): Decorating textured shapes with distilled semantic features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4494-4504, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.377, + 0.906, + 0.431 + ], + "angle": 0, + "content": "[17] Qingnan Fan, Fan Zhong, Dani Lischinski, Daniel Cohen-Or, and Baoquan Chen. Jumpcut: non-successive mask transfer and interpolation for video cutout. ACM Trans. Graph., 34 (6):195-1, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.434, + 0.906, + 0.488 + ], + "angle": 0, + "content": "[18] Stephanie Fu, Mark Hamilton, Laura E. Brandt, Axel Feldmann, Zhoutong Zhang, and William T. Freeman. Feuput: A model-agnostic framework for features at any resolution. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.491, + 0.906, + 0.532 + ], + "angle": 0, + "content": "[19] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.534, + 0.906, + 0.617 + ], + "angle": 0, + "content": "[20] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3749-3761, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.619, + 0.906, + 0.715 + ], + "angle": 0, + "content": "[21] Yuchao Gu, Yipin Zhou, Bichen Wu, Licheng Yu, Jia-Wei Liu, Rui Zhao, Jay Zhangjie Wu, David Junhao Zhang, Mike Zheng Shou, and Kevin Tang. Videoswap: Customized video subject swapping with interactive semantic point correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7621–7630, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.906, + 0.787 + ], + "angle": 0, + "content": "[22] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.906, + 0.844 + ], + "angle": 0, + "content": "[23] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In European Conference on Computer Vision, pages 59-75. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.846, + 0.906, + 0.899 + ], + "angle": 0, + "content": "[24] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.906, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7284" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[25] Eric Hedlin, Gopal Sharma, Shweta Mahajan, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. Unsupervised semantic correspondence using stable diffusion. In NIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[26] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.263 + ], + "angle": 0, + "content": "[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.267, + 0.482, + 0.335 + ], + "angle": 0, + "content": "[28] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.338, + 0.482, + 0.393 + ], + "angle": 0, + "content": "[29] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.396, + 0.482, + 0.479 + ], + "angle": 0, + "content": "[30] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.482, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[31] Peter J Huber. Robust estimation of a location parameter. In Breakthroughs in statistics: Methodology and distribution, pages 492-518. Springer, 1992. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[32] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances in neural information processing systems, 33:19545-19560, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.584, + 0.482, + 0.638 + ], + "angle": 0, + "content": "[33] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv preprint arXiv:2307.07635, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.642, + 0.482, + 0.709 + ], + "angle": 0, + "content": "[34] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker3: Simpler and better point tracking by pseudolabelling real videos. arXiv preprint arXiv:2410.11831, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.714, + 0.482, + 0.768 + ], + "angle": 0, + "content": "[35] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.771, + 0.482, + 0.827 + ], + "angle": 0, + "content": "[36] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.829, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[37] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[38] Alexander C. Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your diffusion model is secretly a zero-shot classifier. In ICCV, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[39] Fuxin Li, Taeyoung Kim, Ahmad Humayun, David Tsai, and James M Rehg. Video segmentation by tracking many figure-ground segments. In Proceedings of the IEEE international conference on computer vision, pages 2192-2199, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.906, + 0.203 + ], + "angle": 0, + "content": "[40] Hongyang Li, Hao Zhang, Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, and Lei Zhang. Taptr: Tracking any point with transformers as detection. arXiv preprint arXiv:2403.13042, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.906, + 0.245 + ], + "angle": 0, + "content": "[41] Xinghui Li, Jingyi Lu, Kai Han, and Victor Prisacariu. Sd4match: Learning to prompt stable diffusion model for semantic matching. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.906, + 0.315 + ], + "angle": 0, + "content": "[42] Zhen Li, Zuo-Liang Zhu, Ling-Hao Han, Qibin Hou, Chun-Le Guo, and Ming-Ming Cheng. Amt: All-pairs multi-field transforms for efficient frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9801-9810, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.906, + 0.355 + ], + "angle": 0, + "content": "[43] Ilya Loshchilov, Frank Hutter, et al. Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101, 5, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.906, + 0.425 + ], + "angle": 0, + "content": "[44] Grace Luo, Lisa Dunlap, Dong Huk Park, Aleksander Holynski, and Trevor Darrell. Diffusion hyperfeatures: Searching through time and space for semantic correspondence. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.906, + 0.496 + ], + "angle": 0, + "content": "[45] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.906, + 0.579 + ], + "angle": 0, + "content": "[46] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.906, + 0.648 + ], + "angle": 0, + "content": "[47] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.906, + 0.704 + ], + "angle": 0, + "content": "[48] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.906, + 0.787 + ], + "angle": 0, + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.906, + 0.83 + ], + "angle": 0, + "content": "[50] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7285" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[52] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[53] Peter Sand and Seth Teller. Particle video: Long-range motion estimation using point trajectories. International journal of computer vision, 80:72-91, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[54] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[55] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.483, + 0.36 + ], + "angle": 0, + "content": "[56] Spencer Sterling. Zeroscope, 2023. https://huggingface.co/cerspense/zeroscope_v2_576w.3.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[57] Shobhita Sundaram, Stephanie Fu, Lukas Mutenthaler, Netanel Y. Tamir, Lucy Chai, Simon Kornblith, Trevor Darrell, and Phillip Isola. When does perceptual alignment benefit vision representations? In NIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[58] Saksham Suri, Matthew Walmer, Kamal Gupta, and Abhinav Shrivastava. Lift: A surprisingly simple lightweight feature transform for dense vit descriptors. In ECCV, pages 110-128. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.478, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[59] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. Advances in Neural Information Processing Systems, 36:1363-1389, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.483, + 0.601 + ], + "angle": 0, + "content": "[60] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.605, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[61] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[62] Narek Tumanyan, Assaf Singer, Shai Bagon, and Tali Dekel. Dino-tracker: Taming dino for self-supervised point tracking in a single video. arXiv preprint arXiv:2403.14548, 2024. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[63] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[64] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural computation, 23(7):1661-1674, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[65] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19795-19806, 2023. 2, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[66] Qian Wang, Abdelrahman Eldosokey, Mohit Mendiratta, Fangneng Zhan, Adam Kortylewski, Christian Theobalt, and Peter Wonka. Zero-shot video semantic segmentation based on pre-trained diffusion models, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[67] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[68] Jay Zhangjie Wu, Guian Fang, Haoning Wu, Xintao Wang, Yixiao Ge, Xiaodong Cun, David Junhao Zhang, Jia-Wei Liu, Yuchao Gu, Rui Zhao, et al. Towards a better metric for text-to-video generation. arXiv preprint arXiv:2401.07781, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.358 + ], + "angle": 0, + "content": "[69] Weilai Xiang, Hongyu Yang, Di Huang, and Yunhong Wang. Denoising diffusion autoencoders are unified self-supervised learners. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15802-15812, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[70] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20406-20417, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[71] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. ODISE: Open-Vocabulary Panoptic Segmentation with Text-to-Image Diffusion Models. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[72] Jiale Yang, So Kanazawa, Masami K Yamaguchi, and Isamu Motoyoshi. Pre-constancy vision in infants. Current Biology, 25(24):3209-3212, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[73] Xingyi Yang and Xinchao Wang. Diffusion model as representation learner. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18938-18949, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[74] Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is easier than you think. arXiv preprint arXiv:2410.06940, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[75] Yuanwen Yue, Anurag Das, Francis Engelmann, Siyu Tang, and Jan Eric Lenssen. Improving 2D Feature Representations by 3D-Aware Fine-Tuning. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[76] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[77] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[78] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7286" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[79] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19855-19865, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[80] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "7287" + } + ] +] \ No newline at end of file diff --git a/2025/Track4Gen_ Teaching Video Diffusion Models to Track Points Improves Video Generation/52b14d4e-8da8-4b5e-a32c-278d64dd36e6_origin.pdf b/2025/Track4Gen_ Teaching Video Diffusion Models to Track Points Improves Video Generation/52b14d4e-8da8-4b5e-a32c-278d64dd36e6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9aa2ac35b85670f6323e0d3ac7a2141b62351bc3 --- /dev/null +++ b/2025/Track4Gen_ Teaching Video Diffusion Models to Track Points Improves Video Generation/52b14d4e-8da8-4b5e-a32c-278d64dd36e6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a6b0b91d905de8b27688eb52933ad6411b0c134ce673a598fa4140b64b0b8feb +size 5523491 diff --git a/2025/Track4Gen_ Teaching Video Diffusion Models to Track Points Improves Video Generation/full.md b/2025/Track4Gen_ Teaching Video Diffusion Models to Track Points Improves Video Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fa3cbecd541f4f207cdfd51a8056c533a353078e --- /dev/null +++ b/2025/Track4Gen_ Teaching Video Diffusion Models to Track Points Improves Video Generation/full.md @@ -0,0 +1,333 @@ +# Track4Gen: Teaching Video Diffusion Models to Track Points Improves Video Generation + +Hyeonho Jeong $^{1,2,*}$ Chun-Hao P. Huang $^{1}$ Jong Chul Ye $^{2}$ Niloy J. Mitra $^{1,3}$ Duygu Ceylan $^{1}$ + +$^{1}$ Adobe Research $^{2}$ KAIST $^{3}$ University College London + +# Abstract + +While recent foundational video generators produce visually rich output, they still struggle with appearance drift, where objects gradually degrade or change inconsistently across frames, breaking visual coherence. We hypothesize that this is because there is no explicit supervision in terms of spatial tracking at the feature level. We propose Track4Gen, a spatially aware video generator that combines video diffusion loss with point tracking across frames, providing enhanced spatial supervision on the diffusion features. Track4Gen merges the video generation and point tracking tasks into a single network by making minimal changes to existing video generation architectures. Using Stable Video Diffusion [4] as a backbone, Track4Gen demonstrates that it is possible to unify video generation and point tracking, which are typically handled as separate tasks. Our extensive evaluations show that Track4Gen effectively reduces appearance drift, resulting in temporally stable and visually coherent video generation. Project page: hyeonho99.github.io/track4gen + +# 1. Introduction + +Diffusion-based video generators [4, 6, 47] are making rapid strides in creating temporally consistent and visually rich video content. This progress marks a significant shift, as the unification of generation and control has the potential to transform the traditional workflow of first capturing and then digitally editing video. + +Despite impressive capabilities, video generators often suffer from appearance drift, where visual elements gradually change, mutate, or degrade over time, causing inconsistencies in the objects. For example, in Fig. 1, we observe the horns of the cow distorting and morphing unrealistically over time, breaking the plausibility of the generated content. This is in striking contrast to humans, who develop a sense of appearance constancy as early as infancy through observation and interaction with the world [72]. + +![](images/a0e63e70d3633171971306d5156655df576d3d1b100c6d09147b576a1b71006a.jpg) +Stable Video Diffusion +Figure 1. Motivation. Videos generated by Stable Video Diffusion [4] suffer from appearance drift, while those from our method, Track4Gen, are free from such appearance inconsistency issues. + +![](images/6c2318d3dfdc5acd100c9a18f08f1d83039ad5d056e1fa375206c8c9a1f94514.jpg) +Track4Gen (Ours) + +Unfortunately, appearance drift remains a persistent issue in current video models, even with increased training data and more advanced architectures. We speculate that this limitation arises from supervision being based solely on video diffusion loss (i.e., denoising score matching [64]) in the pixel/latent space, without explicit spatial awareness guidance in the feature space. Hence, in this paper, we ask if and how we can empower video diffusion models with appearance constancy by providing additional supervision. + +We present Track4Gen as a spatially aware video generator that receives supervision both in terms of the original diffusion-based objective as well as (dense) point correspondence across frames, which we refer to as tracks. We demonstrate that it is possible to provide such track-level supervision in the diffusion feature space by making minimal architecture changes. Our generated videos do not suffer from degradation of video quality (according to the usual video generation metrics), while being significantly more + +spatially coherent as the highlight cow in Fig. 1. + +We train Track4Gen using the latest Stable Video Diffusion [4] as the backbone and evaluate on the publicly available VBench dataset [30]. We report significant improvement in terms of appearance constancy of subjects, both in quantitative and qualitative (i.e., via user studies) evaluations. In summary, we demonstrate that it is possible to upgrade existing video generators, by supervising them with additional correspondence tracking loss, to produce videos without significant appearance drifts, a problem commonly encountered in diffusion-based video generators. + +# 2. Related Work + +Diffusion-based video generation. Building on the success of diffusion models in image synthesis [11, 51], diffusion-based video generators have seen significant advancements [4, 6, 29, 47]. A commonly adopted approach is to extend text-to-image models to the video domain by incorporating temporal layers to facilitate interactions across video frames [5, 22, 54]. While some works have adopted cascaded approaches to produce both spatially and temporally high-resolution videos [28, 47, 54, 67, 76, 78], others have utilized lower-dimensional latent space modeling to reduce computational demands [5, 8, 24, 80]. We build on top of one such approach, Stable Video Diffusion (SVD, [4]), which introduces a latent image-to-video diffusion model trained on a large-scale and curated video data. + +With advances in generation, systematic evaluation of generation quality has become crucial. Traditionally, metrics such as Fréchet Inception Distance (FID, [26]), Fréchet Video Distance (FVD, [63]), and CLIPSIM [49] are used. Additionally, comprehensive benchmark suites [30, 68] have been introduced to provide a more robust evaluation aligned with human perception. Inspired by such work, we thoroughly evaluate our approach and demonstrate improved video generation quality with respect to both conventional metrics and the recent VBenchmark metrics [30]. + +Foundational models as feature extractors. Various foundational models such as vision transformers [15] or diffusion-based generators [50] have been utilized as feature extractors for various tasks including semantic matching [16, 25, 41], classification [38], segmentation [66, 71], and editing [19, 21, 61]. There have been efforts to boost their performance by post-processing the feature maps obtained from the pre-trained models, e.g., by upsampling [18, 58]. In a recent effort, Yue et al. [75] lift semantic per-frame features from a foundational model into a 3D Gaussian representation. They fine-tune the foundational model with such 3D-aware features resulting in improved performance in downstream tasks. Similarly, Sundaram et al. [57] fine-tune state-of-the-art foundational models on human similarity judgments yielding improved representations across + +downstream tasks. In a concurrent effort, Yu et al. [74] propose to align the internal features of an image generation model with external discriminative features [45], which results in more effective training of the generator. + +Our work also enhances the internal feature representation of a foundational generation model but with significant differences compared to previous literature. First, unlike most previous work that focus on image level foundational models, we exploit the power of recently emerging video models. Second, instead of post-processing, we enhance the spatial awareness of the intermediate features by training the generator to jointly perform an additional tracking task. We show that this joint training boosts the performance of intermediate features in correspondence tracking, leading to improved video generation quality. + +Tracking any point in a video. The task involves following any arbitrary query point across a long video sequence. First introduced by PIPs [23] and later re-framed by TAPVid [12], several methods have emerged in recent years to tackle long-term point tracking. PIPs [23] revisits the classical particle-based representation [53] and introduces MLP-based networks that predict point tracks within an 8-frame window. Subsequent works have improved performance by capturing longer temporal context through advanced architectures [2, 13, 23, 33], as well as by enabling the simultaneous tracking of multiple queries [10, 33]. More recent training-based trackers [10, 34, 40, 70] have achieved remarkable performance by leveraging high-capacity neural networks to learn robust priors from large-scale training data. While high-quality data is crucial for accurate tracking, manually annotating point tracks is prohibitively expensive. Hence, synthetic videos [20] with automatic annotations, have become an alternative and have demonstrated effectiveness in real-world video tracking. An alternative approach is self-supervised adaptation at test time, where tracking is learned without ground-truth labels [32, 62, 65]. In a recent study, Aydemir et al. [1] evaluate the effectiveness of several image foundational model features for point tracking both in zero-shot setting as well as with supervised training using low-rank adapter layers. To the best of our knowledge, we are the first to exploit the features of a foundational video diffusion model for dense point tracking. + +# 3. Method + +In this section, we provide a comprehensive discussion of the Track4Gen framework. We begin with a concise overview of latent video diffusion models (Sec. 3.1). Next, we discuss how video diffusion features relate to temporal correspondences both for real and generated videos (Sec. 3.2). Finally, we detail the design of Track4Gen both in terms of network architecture and the employed supervision signals (Sec. 3.3). An overview is depicted in Fig. 2. + +![](images/bc7f5538b0f54c4bee03cb71c68aa9cc4788c40be294ff6445d714b2d5172ec1.jpg) +Figure 2. Track4Gen overview. Red-colored blocks represent layers optimized by the diffusion loss $\mathcal{L}_{\mathrm{diff}}$ , while green blocks are optimized by the correspondence loss $\mathcal{L}_{\mathrm{corr}}$ . Blocks colored both red and green are influenced by the joint loss, $\mathcal{L}_{\mathrm{diff}} + \lambda \mathcal{L}_{\mathrm{corr}}$ . See text for details. + +# 3.1. Background: Stable Video Diffusion + +Starting from random Gaussian noise, diffusion models aim to generate clean images or videos via an iterative denoising process [27, 55]. This process reverses a fixed, time-dependent diffusion forward process, which gradually corrupts the data by adding Gaussian noise. While our method is applicable to general video diffusion models, in this paper, we design our architecture based on Stable Video Diffusion (SVD), a latent video diffusion model which employs the EDM-framework [35]. The diffusion process operates in the lower-dimensional latent space of a pre-trained VAE [37], consisting of an encoder $\mathcal{E}(\cdot)$ and a decoder $\mathcal{D}(\cdot)$ . + +Given a clean sample $\pmb{x}_0^{1:N} \sim p_{\mathrm{data}}(\pmb{x})$ of an $N$ -frame video sequence, the frames are first encoded into the latent space as $z_0^{1:N} = \mathcal{E}(x_0^{1:N})$ . Gaussian noise $\epsilon \sim \mathcal{N}(0,I)$ is then added to the latents to produce the intermediate noisy latents via the forward process $z_t^{1:N} = \alpha_t z_0^{1:N} + \sigma_t \epsilon$ , where $t$ represents the diffusion timestep, and $\alpha_t, \sigma_t$ are the discretized noise scheduler parameters. The diffusion denoiser $f_\theta$ is trained by minimizing the v-prediction loss: + +$$ +\min _ {\theta} \mathbb {E} _ {\boldsymbol {\epsilon} \sim \mathcal {N} (0, I), t \sim U [ 1, T ]} \left[ \left\| \boldsymbol {f} _ {\theta} \left(\boldsymbol {z} _ {t} ^ {1: N}, t, c\right) - \boldsymbol {y} \right\| _ {2} ^ {2} \right], \tag {1} +$$ + +where $\pmb{y}$ is defined as $\pmb{y} = \alpha_{t}\pmb{\epsilon} - \sigma_{t}\pmb{z}_{0}^{1:N}$ . In the image-to-video variant of SVD, the condition $c$ refers to the CLIP image embedding [49], replacing the typical text embeddings. For the remainder of this paper, we will refer to Eq. 1 as the video diffusion loss $\mathcal{L}_{\mathrm{diff}}$ . + +Once trained, the diffusion model generates videos by iteratively denoising a noisy latent $\pmb{z}_T^{1:N}$ sequence sampled from pure Gaussian distribution. At each diffusion step, the model predicts the noise in the input latent. Once the clean latent $\pmb{z}_0^{1:N}$ is obtained, the decoder $\mathcal{D}$ maps it to the higher-dimensional pixel space $\pmb{x}_0^{1:N} = \mathcal{D}(\pmb{z}_0^{1:N})$ . For further details, we refer to the Appendix D of [4]. + +# 3.2. Video Diffusion Features + +Previous studies have demonstrated that image diffusion models learn discriminative features in their hidden states that are effective for various analysis tasks and propose methods for improving the representation power of such features [9, 69, 73, 74]. Similarly, we argue that while also being powerful, internal representations of pre-trained video diffusion models may not be fully temporally consistent, resulting in appearance drift in generated videos. + +To better investigate this hypothesis, we first evaluate the long-term video tracking capabilities of U-Net-based video diffusion models [4, 56, 78]. Specifically, we evaluate the effectiveness of the features from each block of the U-Net for the task of point tracking. Given a real-world video, we add a small amount of noise and extract feature maps from each layer in each block. We perform a cosine-similarity-based nearest-neighbor search [44, 59] over these feature maps for a given set of fixed query points on the first frame (we use a similarity threshold of 0.6 [62] in our experiments). We also perform a similar analysis for generated videos where we extract the feature maps corresponding to diffusion steps with small amount of noise. + +Based on this feature analysis, we make some important observations. Notably, regardless of the model (we analyze both Zeroscope T2V [56] and SVD I2V [4]), we find out that output features from the upsampler layer of the third decoder block consistently yield stronger temporal correspondences, as shown in Fig. 3. Hence, we use this block when extracting features for the remainder of our experiments. Furthermore, when we analyze generated videos and point tracks estimated based on the feature maps (as shown in Fig. 4), we observe that there is a correlation between tracking failures that reveal feature-space inconsistencies and appearance drifts that reveal pixel-space inconsistencies. Hence, we hypothesize that enriching feature consistency can help mitigate such appearance drifts. Next, we + +![](images/c3d26ff034dd6c8539f4a1070898728c77377c605e1a87b947bd6d0e60718dfa.jpg) +Figure 3. Real-world video tracking using different video diffusion features. Given color-coded query points on the first frame (Leftmost column), we display tracked points on target frames using features from different blocks (right columns). The 13th frame (first row) and 8th frame (second row) are shown as target frames. Full results are available in the supplementary and on our page. + +introduce Track4Gen where we accomplish this goal by supervising video diffusion models with a joint tracking loss. + +# 3.3. Track4Gen + +Track4Gen aims to utilize point tracking as an additional supervision signal to enhance the spatial-awareness of video diffusion features. Given that we build on top of a pretrained video generation model, to retain the prior knowledge and avoid tampering the original features directly, we propose a novel architecture change as shown in Fig. 2. Specifically, instead of directly using the raw diffusion features for correspondence estimation, we propose a trainable refiner module $R_{\phi}$ , which is designed to refine the raw features by projecting them into a correspondence-rich feature space. The refined features, which are spatially-aware, are then both used to estimate point tracks with an explicit supervision as well as feeding back to the generation backbone. We empirically find out that this design is more effective compared to fine-tuning the original model with no refinement module (see Sec. 4.2). + +Given an $N$ -frame video sequence $\pmb{x}_0^{1:N}$ , its correspond- + +![](images/83341f49e67f07a973752f72dd26e68bfee31c1bf8a66a2b4c327ef72925575a.jpg) +Feature-space Inconsistency (SVD) +Figure 4. Generated video tracking using video diffusion features. Tracks based on diffusion features are annotated on the generated videos. Track4Gen generates more consistent results. + +![](images/0244d30198704051d756247245a1eea0793ec1bb096c52e163b987358ded1e32.jpg) +Feature-space Consistency (Track4Gen) + +ing latent $z_0^{1:N}$ , and a diffusion timestep $t$ , in order to train Track4Gen we continue to utilize the standard diffusion training loss as defined in Eq. (1), where we adopt the velocity prediction objective [4, 35, 52] for $\mathcal{L}_{\mathrm{diff}}$ . + +To enable tracking supervision, we assume access to a dense set of point trajectories $\Omega = \{(x^i,x^j)\}$ across frames where a point $\mathbf{x}^i$ in frame $i$ corresponds to a matching point $\mathbf{x}^j$ in frame $j$ and vice versa. Given the corresponding noisy video latent sequence $\pmb{z}_t^{1:N}$ , we first extract raw diffusion features as the hidden states $\pmb{h}^{1:N} \in \mathbb{R}^{N \times H \times W \times C}$ from a specific block $b^k$ within the U-Net, where $b^k$ is set to the upsampler layer of the third decoder block (see Sec. 3.2). We then pass these features through the refiner module to obtain the refined feature map $\tilde{\pmb{h}}^{1:N} = \pmb{R}_{\phi}(\pmb{h}^{1:N})$ . + +We sample a query point $\mathbf{x}_{\mathrm{q}}^{i}$ along with its ground-truth target point $\mathbf{x}_{\mathrm{trg}}^{j}$ from the correspondence set $\Omega$ . Given the query point feature $\tilde{\pmb{h}}^i (\mathrm{x_q})\in \mathbb{R}^{1\times 1\times C}$ and the target feature map $\tilde{\pmb{h}}^j\in \mathbb{R}^{H\times W\times C}$ , we calculate the cost volume $S\in$ $\mathbb{R}^{H\times W\times 1}$ as follows: + +$$ +\boldsymbol {S} (\mathrm {p}) = \cos - \operatorname {s i m} \left(\tilde {\boldsymbol {h}} ^ {i} \left(\mathrm {x} _ {\mathrm {q}}\right), \tilde {\boldsymbol {h}} ^ {j} (\mathrm {p})\right), \tag {2} +$$ + +where cos-sim denotes cosine similarity. The predicted target point $\hat{\mathbf{x}}_{\mathrm{trg}}$ is then determined using the differentiable soft-argmax operation: + +$$ +\hat {\mathbf {x}} _ {\mathrm {t r g}} = \frac {\sum_ {\mathrm {p} \in \Omega^ {\prime}} \boldsymbol {S} (\mathrm {p}) \cdot \mathrm {x} _ {\mathrm {p}}}{\sum_ {\mathrm {p} \in \Omega^ {\prime}} \boldsymbol {S} (\mathrm {p})}, \tag {3} +$$ + +where $\Omega^{\prime} = \{p:\| \mathrm{x_p - x_{p_{max}}}\| _2\leq R\} ^1$ . Thus, the target point prediction can be expressed as $\hat{\mathbf{x}}_{\mathrm{trg}} = \xi (\mathrm{x}_{\mathrm{q}}^{i},j,\tilde{\boldsymbol{h}}^{1:N})$ and the predicted tracklet for $\mathrm{x_q^i}$ is given by $\mathcal{T}_{\mathrm{x_q^i}} = \{\hat{\mathbf{x}}_n$ .. $\hat{\mathbf{x}}_n = \xi (\mathbf{x}_q^i,n,\tilde{\pmb{h}}^{1:N}),n = 1,\dots,N\}$ Finally, the correspond + +![](images/bd21d1d1e4725d6bd405f75ea672a22dcd9fc971cfe51f587d57661f318b1c08.jpg) +Input Image +Figure 5. Image-to-video generation results of the original SVD and Track4Gen. Please visit our page for full video view. + +![](images/b4bd58864b5e72ec37a126dcb9983fb8dbf6ad4feed6c7871a53498565368a84.jpg) +Stable Video Diffusion (SVD) + +![](images/dd5cbacaef5caf25b8cd78c19c35e4fee690de50cbcc63a9bc69e13c75217f69.jpg) +Track4Gen (Ours) + +dence loss $\mathcal{L}_{\mathrm{corr}}$ is computed using the Huber loss $L_{H}$ [31]: + +$$ +\mathcal {L} _ {\text {c o r r}} \left(\tilde {\boldsymbol {h}} ^ {1: N}, \Omega\right) = \sum_ {\left(\mathrm {x} _ {\mathrm {q}} ^ {i}, \mathrm {x} _ {\text {t r g}} ^ {j}\right) \in \Omega} L _ {H} \left(\xi \left(\mathrm {x} _ {\mathrm {q}} ^ {i}, j, \tilde {\boldsymbol {h}} ^ {1: N}\right), \mathrm {x} _ {\text {t r g}} ^ {j}\right) \tag {4} +$$ + +When training Track4Gen, we initialize the refiner module as an identity mapping to fully leverage the prior of the base model at the start of finetuning. To re-route the refined features to the backbone generator, we introduce a trainable zero convolution layer [77], denoted as $\zeta_{\psi}$ . While the diffusion loss $\mathcal{L}_{\mathrm{diff}}$ back-propagates to all the blocks of the video diffusion model, we detach the gradients of $\tilde{\pmb{h}}^{1:N}$ before passing into $\zeta_{\psi}$ such that refiner module can solely focus on acquiring the correspondence prior. Hence, given that the output of block $b^{k}$ is $h^{1:N}$ , the input to the subsequent block $b^{k+1}$ is computed as $h^{1:N} + \zeta_{\psi}$ (stop-gradient $(R_{\phi}(h^{1:N}))$ ). + +Fig. 2 visualizes this architecture design, with red and green colors indicating the objective that optimizes each module. + +# 4. Experiments + +# 4.1. Implementation Details + +To train Track4Gen, we construct a training dataset consisting of 567 video-trajectory pairs, with each video having a resolution of $320 \times 576$ and a duration of 24 frames. Since no real-world video with (dense) ground-truth trajectory annotations exist at the time of this work, we utilize + +optical flow to generate trajectory annotations. A key challenge is the need for accurate video segmentation maps to ensure a balanced distribution of trajectory points between foreground objects and the background [12]. To address this, we utilize public video datasets paired with ground-truth segmentation maps [7, 17, 39, 46, 48], where we split longer videos into 24-frame segments. + +We use Stable Video Diffusion (SVD) image-to-video pretrained checkpoints $^2$ as the base video generator. Our proposed refiner module consists of eight stacked 2D convolution layers and is attached to the third decoder block of the SVD UNet. The refiner module preserves the shape of the hidden states throughout and is initialized as the identity mapping. Further details are provided in the supplementary. We finetune this enhanced video generator architecture for $20\mathrm{K}$ steps with our joint loss $\mathcal{L}_{\mathrm{diff}} + \lambda \mathcal{L}_{\mathrm{corr}}$ , where $\lambda$ is set to 8. Rather than finetuning the entire model, we finetune only the temporal transformer blocks, the refiner module $R_{\phi}$ , and the zero convolution $\zeta_{\psi}$ . In each iteration, we sample 512 correspondence pairs from the precomputed trajectories. We use the AdamW optimizer [43] with a learning rate of $1e - 5$ , $\beta_{1} = 0.9$ , $\beta_{2} = 0.999$ , and a weight decay of $1e - 2$ . We train the model on $4\times$ H100 GPUs with a total batch size of 4. For sampling new videos, we apply the + +![](images/2007efe96eb2e6f569980b87b25183468361e418d42b4a9f51fa2c9b6ddf40b8.jpg) +Figure 6. Qualitative ablation on video generation. Track4Gen is compared with finetuned SVD (SVD finetuned on the same training videos without any correspondence supervision) and Track4Gen trained without the Refiner module. + +default settings using 30 steps with the EDM sampler [35], motion bucket id =127, and fps=7. + +Table 1. Quantitative comparison on video generation performance. We compare Track4Gen to the pre-trained SVD $\star$ as well as a finetuned SVD on the same dataset (finetuned SVD). We also train a variant of Track4Gen without the refiner module. All videos are generated at $320\mathrm{x}576$ resolution, except SVD $\star$ (576p) which operates at $576\mathrm{x}1024$ resolution. + +
Subject ConsistencyTemporal FlickeringMotion SmoothnessImaging QualityVideo-Image AlignmentFIDFVD
SVD*0.95350.94640.97740.66480.953929.0776
finetuned SVD0.96650.98000.99090.67660.977127.0735
Track4Gen w/o refiner0.95060.97250.97910.66530.961427.1718
Track4Gen0.97460.98060.99210.68350.981426.6724
SVD* (576p)0.95760.94780.97950.68120.9582--
+ +# 4.2. Track4Gen for Video Generation + +We evaluate Track4Gen for the image-to-video generation task via a series of experiments using multiple datasets, automated metrics, and human evaluations. + +Evaluation Setup. We compare Track4Gen against the original SVD $(\mathrm{SVD}^*)$ [4], as well as a version of SVD that is finetuned on the same videos as Track4Gen (finetuned SVD). Furthermore, we train a variant of Track4Gen without the refiner module. For VBenchmark metrics [30], evaluations are conducted on the VBenchmark-I2V dataset, containing 355 diverse images. FID and FVD are measured using the DAVIS [48] dataset as reference. We generate 24-frame videos conditioned on each input image. + +Automatic metrics. We first report five key metrics from VBench [30]: (1) Subject Consistency—assesses + +subject appearance consistency of the video by computing the similarity of DINO [45] features. (2) Temporal Flickering—detects temporal consistency by taking static frames and calculating the mean absolute difference across frames. (3) Motion Smoothness—measures smoothness of motion, and how well it adheres to real-world physics, using video frame interpolation priors [42]. (4) Image Quality—evaluates distortions (e.g., noise, blur) using a pretrained, multi-scale image quality predictor [36]. (5) Video-Image Alignment—measures alignment between the subject in the input image and in the generated video using DINO features. We additionally report FID [26] and FVD [63]. + +Human evaluation. We further evaluate Track4Gen against baselines through a user study. We ask 64 participants to compare our results with a randomly selected baseline. We ask the users to evaluate how consistent main objects appear across the frames in a generated video as well as how natural the depicted motion is. We provide further details of the user study in the supplementary material. + +Qualitative results. Qualitative comparisons with the base SVD are shown in Fig. 5. As illustrated, Track4Gen generates videos with strong appearance consistency, avoiding issues of appearance drift. In contrast, videos produced by the original SVD exhibit noticeable inconsistencies: the sheep's head (row 1) mutates, the plane's wing (row 2) shows unnatural transitions, and the cars (row 3) disappear. Further comparisons with finetuned SVD and Track4Gen without the refiner module are shown in Fig. 6 and highlight the superior visual coherence of the proposed Track4Gen. + +![](images/89b328905f0a6f1472a28dcc4c4c46facf48d8876c515a62ea87af911aa790c0.jpg) +Query Points +Figure 7. Qualitative comparison of Track4Gen and baselines for real-world video tracking. The leftmost column displays query points in the first frame, while the following three columns show tracking results using features from each model. + +![](images/471eabf03fc07b9cf3fa0b5ed52d1690cf1344f5750cc99bd0ccacfafb3bb601.jpg) +Track4Gen (Ours) + +![](images/3e9dc7be668900098ba8e5d2da1ed158065ecc773d674a1a4be1d5d83d3a6ecc.jpg) +SVD 12V + +![](images/7faa10865074e5889f0db1d531836c40a01d254c72c9fc1fa9c994d92ac8a5e2.jpg) +ZeroScope T2V + +![](images/0f2ac8d7130d7332c0140967b0d794c2808e1e2f98140b56f91d304a3b989d40.jpg) +(a) Identity preservation +Figure 8. User study results. Our study shows that Track4Gen better preserves object identity and produces more natural motion. + +![](images/d6b4d5d15a2c235187bff389aad213516d46ebb79126cc780c35ca89eb6ba6b4.jpg) +(b) Motion naturalness + +Quantitative results. As shown in Tab. 1, our method achieves the highest scores across all 5 metrics from VBench, along with the lowest FID and second-lowest FVD values, outperforming the base SVD by substantial margins. Fig. 8 provides the user study results where the majority of the participants agreed that Track4Gen is superior both in terms of identity preservation and naturalness of motion. + +# 4.3. Track4Gen for Video Tracking + +We evaluate Track4Gen's capability to track any point in real videos by adding a small amount of noise to the input video [59] and passing it through the video denoiser $\pmb{f}_{\theta}$ to extract feature maps. We first compare tracking results with such features against other raw features [4, 56, 60] in Sec. 4.3.1. In Sec. 4.3.2, we utilize Track4Gen's features in a test-time optimization method [62] and compare to both self-supervised and fully supervised video trackers. + +# 4.3.1. Zero-shot Feature Comparison + +We evaluate the precision of predicted tracks using the features from Track4Gen, the original SVD model $(\mathrm{SVD}^{\star})$ and RAFT [60]. We also test another text-to-video model, ZeroScope T2V [56], to demonstrate how raw features from pre-trained video generators typically work out of the box. + +Table 2. Quantitative zero-shot feature comparison on video tracking benchmarks. Track4Gen features are compared to the features of SVD* [4], ZeroScope [56], and RAFT [60]. For all the metrics, higher values indicate better performance. + +
MethodDAVIS-480p(24-frame)BADJA(24-frame)DAVIS-480p(whole duration)BADJA(whole duration)
\( \delta^{x}_{avg} \)OAAJ\( \delta^{seg} \)\( \delta^{3px} \)\( \delta^{x}_{avg} \)OAAJ\( \delta^{seg} \)\( \delta^{3px} \)
ZeroScope46.267.039.427.52.837.259.527.819.92.0
SVD*42.479.736.426.22.935.470.126.519.42.2
Track4Gen69.785.856.552.37.758.978.440.240.45.0
RAFT73.3--54.88.766.7--45.05.8
+ +For RAFT, tracking is achieved by chaining optical flow displacements, while the others use nearest neighbor matching between its encoded features. + +Datasets. We use TAP-Vid DAVIS [12] and BADJA [3] as benchmark datasets. Additionally, we include two shorter benchmarks, DAVIS (24-frame) and BADJA (24-frame), which focus on the first 24 frames with query and target points within this range. Details on encoding long videos with the video models are in the supplemental. + +Metrics. For evaluating the TAP-Vid benchmarks, we use the following metrics: (i) Position Accuracy $(\delta_{avg}^{x})$ evaluates the average accuracy of visible points, where each $\delta^x$ represents the fraction of predicted points that lie within $x$ pixels of the ground-truth position, with $x \in \{1, 2, 4, 8, 16\}$ . (ii) Occlusion Accuracy (OA) evaluates the correctness of occlusion predictions. (iii) Average Jaccard (AJ) jointly assesses both position and occlusion accuracy. For the BADJA dataset, we report $\delta^{seg}$ , which measures the accuracy of tracked keypoints within a distance of $0.2\sqrt{A}$ from the ground-truth annotation, where $A$ is the area of the foreground object. We also report $\delta^{3px}$ , which assesses accuracy within a 3-pixel threshold. A cosine similarity threshold of 0.6 is used for occlusion prediction. + +![](images/620611b3f3f85c7a27be961227007156bb3ec0f3fa0572c87926f784cd601428.jpg) +First Frame with Query Points +Figure 9. Extending Track4Gen with test-time adaptation [62]. + +![](images/c67742c215bddb707aac03f8fff1c06e96eeb62ceba6f34f040bdba252afead8.jpg) +Tracking Results in Subsequent Frames + +Table 3. Quantitative comparison with video trackers. Although primarily designed for video generation, Track4Gen combined with a test-time optimization method [62] achieves performance comparable to dedicated video tracking frameworks, even when compared to supervised methods. + +
MethodDAVIS-480BADJA
δxavgOAAJδsegδ3px
TAP-Net*66.479.046.045.49.6
PIPs++*73.6--59.09.8
TAPIR*77.389.565.768.710.5
Omnimotion†74.184.558.445.26.9
DINO-Tracker†80.488.164.672.414.3
DINO-Tracker w/ Track4Gen†72.584.555.748.410.9
+ +\* - supervised. -test-time training. + +Table 4. Ablation on trainable modules and refiner. + +
Trainable modulesSubject ConsistencyTemporal FlickeringMotion SmoothnessImaging QualityVideo-Image Alignment
spatial + temporal0.97340.98110.99170.68630.9807
spatial0.97260.98010.99190.68520.9811
temporal0.97460.98060.99210.68350.9814
Refiner architecture
2D convolutions0.97460.98060.99210.68350.9814
3D convolutions0.96870.97340.99040.68330.9820
+ +Results. We present the qualitative results in Fig. 7 and the quantitative results in Tab. 2. Although primarily designed for video generation, Track4Gen boosts the poor performance of the pre-trained video models significantly, approaching the accuracy of RAFT optical flow chaining. + +# 4.3.2. Extending Track4Gen with Test-time Adaptation + +To further evaluate Track4Gen's long-term tracking capabilities, we integrate our features with test-time adaptation algorithm of DINO-Tracker [62], where a per-video optimization is performed using optical flow supervision. We replace the originally used DINOv2 [45] with the features from Track4Gen. We evaluate using the same datasets and metrics outlined in Sec. 4.3.1, against both fully-supervised trackers [12, 13, 79] and self-supervised methods [62, 65]. + +Tab. 3 shows that Track4Gen features optimized with [62] achieve performance comparable to dedicated trackers. Qualitative results are in Fig. 9 and in the supplemental. + +Table 5. Quantitative ablation on using annotated, but synthetic videos [20]. Left: Video generation metrics. Right: Video tracking metrics. + +
Dataset compositionSubject ConsistencyMotion SmoothnessImaging QualityBADJA
δsegδ3px
real videos0.97470.99210.683340.45.0
real + synthetic videos0.97080.98920.679342.14.8
+ +# 4.4. Ablation Studies + +We present an ablation study in Tab. 4 where we train different set of modules. Each spatio-temporal block of SVD includes both spatial and temporal transformers. We compare training only spatial transformers, only temporal transformers, or both. We also ablate the architecture of the refiner module using either 2D or 3D convolution layers. Our analysis shows that while results are similar across settings, training only the temporal transformers in SVD with 2D convolutions as the refiner module yields optimal video generation quality. We further analyze our training dataset by additionally incorporating Kubric [20] simulated videos (1K video-track pairs from the Panning MOVi-E data [10, 13]) with automatically annotated trajectories into training. As shown in Tab. 5, optical flow-chained tracklets from real provides provide as effective correspondence guidance as tracklets from synthetic data, while synthetic videos negatively impact the video generation quality. + +# 5. Conclusion and Future Work + +We have presented the first unified framework that bridges two distinct tasks: video generation and dense point tracking. We demonstrated that this produces temporally consistent feature representations and appearance-consistent videos. As for limitations, videos generated by Track4Gen tend to exhibit less dynamic motion compared to those from other video generators. Additionally, failure cases are included in the supplementary material. + +Future work. Recently, cutting-edge video trackers [10, 14, 34] have emerged, enabling dense, accurate, and long-term tracking, especially with better handling of occlusions. This opens up promising future directions for extending our work to utilize real-world videos, automatically annotated by these advanced trackers. + +Acknowledgments. We thank Seokju Cho and Narek Tumanyan for their invaluable feedback on video point tracking. We also extend our gratitude to Mingi Kwon, Joon-Young Lee, and Gabriel Huang for their insightful discussions. Hyeonho Jeong and Jong Chul Ye are supported by the National Research Foundation of Korea (NRF) under Grants RS-2024-00336454 and RS-2023-00262527, and by the Institute for Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (RS-2019-II190075, Artificial Intelligence Graduate School Program, KAIST). + +# References + +[1] Gorkay Aydemir, Weidi Xie, and Fatma Güney. Can visual foundation models achieve long-term point tracking? arXiv preprint arXiv:2408.13575, 2024. 2 +[2] Weikang Bian, Zhaoyang Huang, Xiaoyu Shi, Yitong Dong, Yijin Li, and Hongsheng Li. Context-tap: Tracking any point demands spatial context features. arXiv preprint arXiv:2306.02000, 3, 2023. 2 +[3] Benjamin Biggs, Thomas Roddick, Andrew Fitzgibbon, and Roberto Cipolla. Creatures great and small: Recovering the shape and motion of animals from video. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part V 14, pages 3-19. Springer, 2019. 7 +[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1, 2, 3, 4, 6, 7 +[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2 +[6] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 1, 2 +[7] Sergi Caelles, Alberto Montes, Kevis-Kokitsi Maninis, Yuhua Chen, Luc Van Gool, Federico Perazzi, and Jordi Pont-Tuset. The 2018 davis challenge on video object segmentation. arXiv preprint arXiv:1803.00557, 2018. 5 +[8] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2 +[9] X Chen, Z Liu, S Xie, and K He. Deconstructing denoising diffusion models for self-supervised learning. arxiv 2024. arXiv preprint arXiv:2401.14404. 3 +[10] Seokju Cho, Jiahui Huang, Jisu Nam, Honggyu An, Seun-gryong Kim, and Joon-Young Lee. Local all-pair correspondence for point tracking. arXiv preprint arXiv:2407.15420, 2024. 2, 8 +[11] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2 +[12] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610-13626, 2022. 2, 5, 7, 8 +[13] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. + +Tapir: Tracking any point with per-frame initialization and temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10061-10072, 2023. 2, 8 +[14] Carl Doersch, Yi Yang, Dilara Gokay, Pauline Luc, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ross Goroshin, João Carreira, and Andrew Zisserman. Bootstrap: Bootstrapped training for tracking-any-point. arXiv preprint arXiv:2402.00847, 2024. 8 +[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2 +[16] Niladri Shekhar Dutt, Sanjeev Muralikrishnan, and Niloy J. Mitra. Diffusion 3d features (diff3f): Decorating textured shapes with distilled semantic features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4494-4504, 2024. 2 +[17] Qingnan Fan, Fan Zhong, Dani Lischinski, Daniel Cohen-Or, and Baoquan Chen. Jumpcut: non-successive mask transfer and interpolation for video cutout. ACM Trans. Graph., 34 (6):195-1, 2015. 5 +[18] Stephanie Fu, Mark Hamilton, Laura E. Brandt, Axel Feldmann, Zhoutong Zhang, and William T. Freeman. Feuput: A model-agnostic framework for features at any resolution. In ICLR, 2024. 2 +[19] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 2 +[20] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3749-3761, 2022. 2, 8 +[21] Yuchao Gu, Yipin Zhou, Bichen Wu, Licheng Yu, Jia-Wei Liu, Rui Zhao, Jay Zhangjie Wu, David Junhao Zhang, Mike Zheng Shou, and Kevin Tang. Videoswap: Customized video subject swapping with interactive semantic point correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7621–7630, 2024. 2 +[22] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2 +[23] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In European Conference on Computer Vision, pages 59-75. Springer, 2022. 2 +[24] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221, 2022. 2 + +[25] Eric Hedlin, Gopal Sharma, Shweta Mahajan, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. Unsupervised semantic correspondence using stable diffusion. In NIPS, 2023. 2 +[26] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 2, 6 +[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3 +[28] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2 +[29] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2 +[30] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 2, 6 +[31] Peter J Huber. Robust estimation of a location parameter. In Breakthroughs in statistics: Methodology and distribution, pages 492-518. Springer, 1992. 5 +[32] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances in neural information processing systems, 33:19545-19560, 2020. 2 +[33] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv preprint arXiv:2307.07635, 2023. 2 +[34] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker3: Simpler and better point tracking by pseudolabelling real videos. arXiv preprint arXiv:2410.11831, 2024. 2, 8 +[35] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. 3, 4, 6 +[36] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 6 +[37] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3 +[38] Alexander C. Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your diffusion model is secretly a zero-shot classifier. In ICCV, 2023. 2 + +[39] Fuxin Li, Taeyoung Kim, Ahmad Humayun, David Tsai, and James M Rehg. Video segmentation by tracking many figure-ground segments. In Proceedings of the IEEE international conference on computer vision, pages 2192-2199, 2013. 5 +[40] Hongyang Li, Hao Zhang, Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, and Lei Zhang. Taptr: Tracking any point with transformers as detection. arXiv preprint arXiv:2403.13042, 2024. 2 +[41] Xinghui Li, Jingyi Lu, Kai Han, and Victor Prisacariu. Sd4match: Learning to prompt stable diffusion model for semantic matching. In CVPR, 2023. 2 +[42] Zhen Li, Zuo-Liang Zhu, Ling-Hao Han, Qibin Hou, Chun-Le Guo, and Ming-Ming Cheng. Amt: All-pairs multi-field transforms for efficient frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9801-9810, 2023. 6 +[43] Ilya Loshchilov, Frank Hutter, et al. Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101, 5, 2017. 5 +[44] Grace Luo, Lisa Dunlap, Dong Huk Park, Aleksander Holynski, and Trevor Darrell. Diffusion hyperfeatures: Searching through time and space for semantic correspondence. Advances in Neural Information Processing Systems, 36, 2024. 3 +[45] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 6, 8 +[46] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 5 +[47] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 1, 2 +[48] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 5, 6 +[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3 +[50] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2021. 2 +[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2 + +[52] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 4 +[53] Peter Sand and Seth Teller. Particle video: Long-range motion estimation using point trajectories. International journal of computer vision, 80:72-91, 2008. 2 +[54] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2 +[55] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3 +[56] Spencer Sterling. Zeroscope, 2023. https://huggingface.co/cerspense/zeroscope_v2_576w.3.7 +[57] Shobhita Sundaram, Stephanie Fu, Lukas Mutenthaler, Netanel Y. Tamir, Lucy Chai, Simon Kornblith, Trevor Darrell, and Phillip Isola. When does perceptual alignment benefit vision representations? In NIPS, 2024. 2 +[58] Saksham Suri, Matthew Walmer, Kamal Gupta, and Abhinav Shrivastava. Lift: A surprisingly simple lightweight feature transform for dense vit descriptors. In ECCV, pages 110-128. Springer, 2024. 2 +[59] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. Advances in Neural Information Processing Systems, 36:1363-1389, 2023. 3, 7 +[60] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 7 +[61] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 2 +[62] Narek Tumanyan, Assaf Singer, Shai Bagon, and Tali Dekel. Dino-tracker: Taming dino for self-supervised point tracking in a single video. arXiv preprint arXiv:2403.14548, 2024. 2, 3, 7, 8 +[63] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 2, 6 +[64] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural computation, 23(7):1661-1674, 2011. 1 +[65] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19795-19806, 2023. 2, 8 + +[66] Qian Wang, Abdelrahman Eldosokey, Mohit Mendiratta, Fangneng Zhan, Adam Kortylewski, Christian Theobalt, and Peter Wonka. Zero-shot video semantic segmentation based on pre-trained diffusion models, 2024. 2 +[67] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2 +[68] Jay Zhangjie Wu, Guian Fang, Haoning Wu, Xintao Wang, Yixiao Ge, Xiaodong Cun, David Junhao Zhang, Jia-Wei Liu, Yuchao Gu, Rui Zhao, et al. Towards a better metric for text-to-video generation. arXiv preprint arXiv:2401.07781, 2024. 2 +[69] Weilai Xiang, Hongyu Yang, Di Huang, and Yunhong Wang. Denoising diffusion autoencoders are unified self-supervised learners. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15802-15812, 2023. 3 +[70] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20406-20417, 2024. 2 +[71] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. ODISE: Open-Vocabulary Panoptic Segmentation with Text-to-Image Diffusion Models. In CVPR, 2023. 2 +[72] Jiale Yang, So Kanazawa, Masami K Yamaguchi, and Isamu Motoyoshi. Pre-constancy vision in infants. Current Biology, 25(24):3209-3212, 2015. 1 +[73] Xingyi Yang and Xinchao Wang. Diffusion model as representation learner. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18938-18949, 2023. 3 +[74] Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is easier than you think. arXiv preprint arXiv:2410.06940, 2024. 2, 3 +[75] Yuanwen Yue, Anurag Das, Francis Engelmann, Siyu Tang, and Jan Eric Lenssen. Improving 2D Feature Representations by 3D-Aware Fine-Tuning. In ECCV, 2024. 2 +[76] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818, 2023. 2 +[77] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 5 +[78] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145, 2023. 2, 3 + +[79] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19855-19865, 2023. 8 +[80] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2 \ No newline at end of file diff --git a/2025/Track4Gen_ Teaching Video Diffusion Models to Track Points Improves Video Generation/images.zip b/2025/Track4Gen_ Teaching Video Diffusion Models to Track Points Improves Video Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2253c64f20f7fba6ad3ab41d4c0e7ffb943819a5 --- /dev/null +++ b/2025/Track4Gen_ Teaching Video Diffusion Models to Track Points Improves Video Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:754ff416e19b504a28075a60eb0199a123c8a75515cf96b6f7dd7e4233ed13d0 +size 1043739 diff --git a/2025/Track4Gen_ Teaching Video Diffusion Models to Track Points Improves Video Generation/layout.json b/2025/Track4Gen_ Teaching Video Diffusion Models to Track Points Improves Video Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d8793e0ff1e40eb80b402a9702ff850921640249 --- /dev/null +++ b/2025/Track4Gen_ Teaching Video Diffusion Models to Track Points Improves Video Generation/layout.json @@ -0,0 +1,10028 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 115, + 103, + 496, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 103, + 496, + 138 + ], + "spans": [ + { + "bbox": [ + 115, + 103, + 496, + 138 + ], + "type": "text", + "content": "Track4Gen: Teaching Video Diffusion Models to Track Points Improves Video Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 160, + 542, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 160, + 542, + 177 + ], + "spans": [ + { + "bbox": [ + 64, + 160, + 542, + 177 + ], + "type": "text", + "content": "Hyeonho Jeong" + }, + { + "bbox": [ + 64, + 160, + 542, + 177 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 64, + 160, + 542, + 177 + ], + "type": "text", + "content": " Chun-Hao P. Huang" + }, + { + "bbox": [ + 64, + 160, + 542, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 160, + 542, + 177 + ], + "type": "text", + "content": " Jong Chul Ye" + }, + { + "bbox": [ + 64, + 160, + 542, + 177 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 160, + 542, + 177 + ], + "type": "text", + "content": " Niloy J. Mitra" + }, + { + "bbox": [ + 64, + 160, + 542, + 177 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 64, + 160, + 542, + 177 + ], + "type": "text", + "content": " Duygu Ceylan" + }, + { + "bbox": [ + 64, + 160, + 542, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 148, + 180, + 462, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 180, + 462, + 197 + ], + "spans": [ + { + "bbox": [ + 148, + 180, + 462, + 197 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 180, + 462, + 197 + ], + "type": "text", + "content": "Adobe Research " + }, + { + "bbox": [ + 148, + 180, + 462, + 197 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 148, + 180, + 462, + 197 + ], + "type": "text", + "content": "KAIST " + }, + { + "bbox": [ + 148, + 180, + 462, + 197 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 148, + 180, + 462, + 197 + ], + "type": "text", + "content": "University College London" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 224, + 200, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 224, + 200, + 236 + ], + "spans": [ + { + "bbox": [ + 151, + 224, + 200, + 236 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 248, + 296, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 248, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 296, + 477 + ], + "type": "text", + "content": "While recent foundational video generators produce visually rich output, they still struggle with appearance drift, where objects gradually degrade or change inconsistently across frames, breaking visual coherence. We hypothesize that this is because there is no explicit supervision in terms of spatial tracking at the feature level. We propose Track4Gen, a spatially aware video generator that combines video diffusion loss with point tracking across frames, providing enhanced spatial supervision on the diffusion features. Track4Gen merges the video generation and point tracking tasks into a single network by making minimal changes to existing video generation architectures. Using Stable Video Diffusion [4] as a backbone, Track4Gen demonstrates that it is possible to unify video generation and point tracking, which are typically handled as separate tasks. Our extensive evaluations show that Track4Gen effectively reduces appearance drift, resulting in temporally stable and visually coherent video generation. Project page: hyeonho99.github.io/track4gen" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 497, + 135, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 135, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 135, + 510 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 517, + 295, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 589 + ], + "type": "text", + "content": "Diffusion-based video generators [4, 6, 47] are making rapid strides in creating temporally consistent and visually rich video content. This progress marks a significant shift, as the unification of generation and control has the potential to transform the traditional workflow of first capturing and then digitally editing video." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 590, + 295, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 295, + 697 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 295, + 697 + ], + "type": "text", + "content": "Despite impressive capabilities, video generators often suffer from appearance drift, where visual elements gradually change, mutate, or degrade over time, causing inconsistencies in the objects. For example, in Fig. 1, we observe the horns of the cow distorting and morphing unrealistically over time, breaking the plausibility of the generated content. This is in striking contrast to humans, who develop a sense of appearance constancy as early as infancy through observation and interaction with the world [72]." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 232, + 434, + 431 + ], + "blocks": [ + { + "bbox": [ + 335, + 222, + 415, + 231 + ], + "lines": [ + { + "bbox": [ + 335, + 222, + 415, + 231 + ], + "spans": [ + { + "bbox": [ + 335, + 222, + 415, + 231 + ], + "type": "text", + "content": "Stable Video Diffusion" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 232, + 434, + 431 + ], + "lines": [ + { + "bbox": [ + 316, + 232, + 434, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 434, + 431 + ], + "type": "image", + "image_path": "a0e63e70d3633171971306d5156655df576d3d1b100c6d09147b576a1b71006a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 433, + 555, + 467 + ], + "lines": [ + { + "bbox": [ + 313, + 433, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 555, + 467 + ], + "type": "text", + "content": "Figure 1. Motivation. Videos generated by Stable Video Diffusion [4] suffer from appearance drift, while those from our method, Track4Gen, are free from such appearance inconsistency issues." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 435, + 232, + 553, + 430 + ], + "blocks": [ + { + "bbox": [ + 462, + 222, + 527, + 232 + ], + "lines": [ + { + "bbox": [ + 462, + 222, + 527, + 232 + ], + "spans": [ + { + "bbox": [ + 462, + 222, + 527, + 232 + ], + "type": "text", + "content": "Track4Gen (Ours)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 435, + 232, + 553, + 430 + ], + "lines": [ + { + "bbox": [ + 435, + 232, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 435, + 232, + 553, + 430 + ], + "type": "image", + "image_path": "6c2318d3dfdc5acd100c9a18f08f1d83039ad5d056e1fa375206c8c9a1f94514.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 495, + 555, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 495, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 312, + 495, + 555, + 602 + ], + "type": "text", + "content": "Unfortunately, appearance drift remains a persistent issue in current video models, even with increased training data and more advanced architectures. We speculate that this limitation arises from supervision being based solely on video diffusion loss (i.e., denoising score matching [64]) in the pixel/latent space, without explicit spatial awareness guidance in the feature space. Hence, in this paper, we ask if and how we can empower video diffusion models with appearance constancy by providing additional supervision." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "We present Track4Gen as a spatially aware video generator that receives supervision both in terms of the original diffusion-based objective as well as (dense) point correspondence across frames, which we refer to as tracks. We demonstrate that it is possible to provide such track-level supervision in the diffusion feature space by making minimal architecture changes. Our generated videos do not suffer from degradation of video quality (according to the usual video generation metrics), while being significantly more" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 203, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 203, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 203, + 713 + ], + "type": "text", + "content": "*Work done during internship at Adobe." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7276" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 251, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 251, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 251, + 83 + ], + "type": "text", + "content": "spatially coherent as the highlight cow in Fig. 1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 84, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 84, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 54, + 84, + 294, + 204 + ], + "type": "text", + "content": "We train Track4Gen using the latest Stable Video Diffusion [4] as the backbone and evaluate on the publicly available VBench dataset [30]. We report significant improvement in terms of appearance constancy of subjects, both in quantitative and qualitative (i.e., via user studies) evaluations. In summary, we demonstrate that it is possible to upgrade existing video generators, by supervising them with additional correspondence tracking loss, to produce videos without significant appearance drifts, a problem commonly encountered in diffusion-based video generators." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 213, + 142, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 142, + 225 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 142, + 225 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 233, + 295, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 233, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 233, + 295, + 399 + ], + "type": "text", + "content": "Diffusion-based video generation. Building on the success of diffusion models in image synthesis [11, 51], diffusion-based video generators have seen significant advancements [4, 6, 29, 47]. A commonly adopted approach is to extend text-to-image models to the video domain by incorporating temporal layers to facilitate interactions across video frames [5, 22, 54]. While some works have adopted cascaded approaches to produce both spatially and temporally high-resolution videos [28, 47, 54, 67, 76, 78], others have utilized lower-dimensional latent space modeling to reduce computational demands [5, 8, 24, 80]. We build on top of one such approach, Stable Video Diffusion (SVD, [4]), which introduces a latent image-to-video diffusion model trained on a large-scale and curated video data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 401, + 295, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 295, + 520 + ], + "type": "text", + "content": "With advances in generation, systematic evaluation of generation quality has become crucial. Traditionally, metrics such as Fréchet Inception Distance (FID, [26]), Fréchet Video Distance (FVD, [63]), and CLIPSIM [49] are used. Additionally, comprehensive benchmark suites [30, 68] have been introduced to provide a more robust evaluation aligned with human perception. Inspired by such work, we thoroughly evaluate our approach and demonstrate improved video generation quality with respect to both conventional metrics and the recent VBenchmark metrics [30]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "Foundational models as feature extractors. Various foundational models such as vision transformers [15] or diffusion-based generators [50] have been utilized as feature extractors for various tasks including semantic matching [16, 25, 41], classification [38], segmentation [66, 71], and editing [19, 21, 61]. There have been efforts to boost their performance by post-processing the feature maps obtained from the pre-trained models, e.g., by upsampling [18, 58]. In a recent effort, Yue et al. [75] lift semantic per-frame features from a foundational model into a 3D Gaussian representation. They fine-tune the foundational model with such 3D-aware features resulting in improved performance in downstream tasks. Similarly, Sundaram et al. [57] fine-tune state-of-the-art foundational models on human similarity judgments yielding improved representations across" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "downstream tasks. In a concurrent effort, Yu et al. [74] propose to align the internal features of an image generation model with external discriminative features [45], which results in more effective training of the generator." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 121, + 555, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 252 + ], + "type": "text", + "content": "Our work also enhances the internal feature representation of a foundational generation model but with significant differences compared to previous literature. First, unlike most previous work that focus on image level foundational models, we exploit the power of recently emerging video models. Second, instead of post-processing, we enhance the spatial awareness of the intermediate features by training the generator to jointly perform an additional tracking task. We show that this joint training boosts the performance of intermediate features in correspondence tracking, leading to improved video generation quality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 266, + 555, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 555, + 589 + ], + "type": "text", + "content": "Tracking any point in a video. The task involves following any arbitrary query point across a long video sequence. First introduced by PIPs [23] and later re-framed by TAPVid [12], several methods have emerged in recent years to tackle long-term point tracking. PIPs [23] revisits the classical particle-based representation [53] and introduces MLP-based networks that predict point tracks within an 8-frame window. Subsequent works have improved performance by capturing longer temporal context through advanced architectures [2, 13, 23, 33], as well as by enabling the simultaneous tracking of multiple queries [10, 33]. More recent training-based trackers [10, 34, 40, 70] have achieved remarkable performance by leveraging high-capacity neural networks to learn robust priors from large-scale training data. While high-quality data is crucial for accurate tracking, manually annotating point tracks is prohibitively expensive. Hence, synthetic videos [20] with automatic annotations, have become an alternative and have demonstrated effectiveness in real-world video tracking. An alternative approach is self-supervised adaptation at test time, where tracking is learned without ground-truth labels [32, 62, 65]. In a recent study, Aydemir et al. [1] evaluate the effectiveness of several image foundational model features for point tracking both in zero-shot setting as well as with supervised training using low-rank adapter layers. To the best of our knowledge, we are the first to exploit the features of a foundational video diffusion model for dense point tracking." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 597, + 370, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 597, + 370, + 609 + ], + "spans": [ + { + "bbox": [ + 314, + 597, + 370, + 609 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 618, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 553, + 715 + ], + "type": "text", + "content": "In this section, we provide a comprehensive discussion of the Track4Gen framework. We begin with a concise overview of latent video diffusion models (Sec. 3.1). Next, we discuss how video diffusion features relate to temporal correspondences both for real and generated videos (Sec. 3.2). Finally, we detail the design of Track4Gen both in terms of network architecture and the employed supervision signals (Sec. 3.3). An overview is depicted in Fig. 2." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7277" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 68, + 547, + 222 + ], + "blocks": [ + { + "bbox": [ + 62, + 68, + 547, + 222 + ], + "lines": [ + { + "bbox": [ + 62, + 68, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 62, + 68, + 547, + 222 + ], + "type": "image", + "image_path": "bc7f5538b0f54c4bee03cb71c68aa9cc4788c40be294ff6445d714b2d5172ec1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 227, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 54, + 227, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 54, + 227, + 555, + 251 + ], + "type": "text", + "content": "Figure 2. Track4Gen overview. Red-colored blocks represent layers optimized by the diffusion loss " + }, + { + "bbox": [ + 54, + 227, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{diff}}" + }, + { + "bbox": [ + 54, + 227, + 555, + 251 + ], + "type": "text", + "content": ", while green blocks are optimized by the correspondence loss " + }, + { + "bbox": [ + 54, + 227, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{corr}}" + }, + { + "bbox": [ + 54, + 227, + 555, + 251 + ], + "type": "text", + "content": ". Blocks colored both red and green are influenced by the joint loss, " + }, + { + "bbox": [ + 54, + 227, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{diff}} + \\lambda \\mathcal{L}_{\\mathrm{corr}}" + }, + { + "bbox": [ + 54, + 227, + 555, + 251 + ], + "type": "text", + "content": ". See text for details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 270, + 248, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 248, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 248, + 283 + ], + "type": "text", + "content": "3.1. Background: Stable Video Diffusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 289, + 295, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 289, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 54, + 289, + 295, + 421 + ], + "type": "text", + "content": "Starting from random Gaussian noise, diffusion models aim to generate clean images or videos via an iterative denoising process [27, 55]. This process reverses a fixed, time-dependent diffusion forward process, which gradually corrupts the data by adding Gaussian noise. While our method is applicable to general video diffusion models, in this paper, we design our architecture based on Stable Video Diffusion (SVD), a latent video diffusion model which employs the EDM-framework [35]. The diffusion process operates in the lower-dimensional latent space of a pre-trained VAE [37], consisting of an encoder " + }, + { + "bbox": [ + 54, + 289, + 295, + 421 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\cdot)" + }, + { + "bbox": [ + 54, + 289, + 295, + 421 + ], + "type": "text", + "content": " and a decoder " + }, + { + "bbox": [ + 54, + 289, + 295, + 421 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\cdot)" + }, + { + "bbox": [ + 54, + 289, + 295, + 421 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 423, + 296, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 296, + 520 + ], + "type": "text", + "content": "Given a clean sample " + }, + { + "bbox": [ + 55, + 423, + 296, + 520 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0^{1:N} \\sim p_{\\mathrm{data}}(\\pmb{x})" + }, + { + "bbox": [ + 55, + 423, + 296, + 520 + ], + "type": "text", + "content": " of an " + }, + { + "bbox": [ + 55, + 423, + 296, + 520 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 423, + 296, + 520 + ], + "type": "text", + "content": "-frame video sequence, the frames are first encoded into the latent space as " + }, + { + "bbox": [ + 55, + 423, + 296, + 520 + ], + "type": "inline_equation", + "content": "z_0^{1:N} = \\mathcal{E}(x_0^{1:N})" + }, + { + "bbox": [ + 55, + 423, + 296, + 520 + ], + "type": "text", + "content": ". Gaussian noise " + }, + { + "bbox": [ + 55, + 423, + 296, + 520 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 55, + 423, + 296, + 520 + ], + "type": "text", + "content": " is then added to the latents to produce the intermediate noisy latents via the forward process " + }, + { + "bbox": [ + 55, + 423, + 296, + 520 + ], + "type": "inline_equation", + "content": "z_t^{1:N} = \\alpha_t z_0^{1:N} + \\sigma_t \\epsilon" + }, + { + "bbox": [ + 55, + 423, + 296, + 520 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 423, + 296, + 520 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 423, + 296, + 520 + ], + "type": "text", + "content": " represents the diffusion timestep, and " + }, + { + "bbox": [ + 55, + 423, + 296, + 520 + ], + "type": "inline_equation", + "content": "\\alpha_t, \\sigma_t" + }, + { + "bbox": [ + 55, + 423, + 296, + 520 + ], + "type": "text", + "content": " are the discretized noise scheduler parameters. The diffusion denoiser " + }, + { + "bbox": [ + 55, + 423, + 296, + 520 + ], + "type": "inline_equation", + "content": "f_\\theta" + }, + { + "bbox": [ + 55, + 423, + 296, + 520 + ], + "type": "text", + "content": " is trained by minimizing the v-prediction loss:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 533, + 295, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 533, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 69, + 533, + 295, + 554 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {\\boldsymbol {\\epsilon} \\sim \\mathcal {N} (0, I), t \\sim U [ 1, T ]} \\left[ \\left\\| \\boldsymbol {f} _ {\\theta} \\left(\\boldsymbol {z} _ {t} ^ {1: N}, t, c\\right) - \\boldsymbol {y} \\right\\| _ {2} ^ {2} \\right], \\tag {1}", + "image_path": "d50d96be3d9937a5b7efef75951fa189dd485b2260efc2db67d1d6a3b42e4b3d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 567, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 295, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 567, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 55, + 567, + 295, + 628 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 55, + 567, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\pmb{y} = \\alpha_{t}\\pmb{\\epsilon} - \\sigma_{t}\\pmb{z}_{0}^{1:N}" + }, + { + "bbox": [ + 55, + 567, + 295, + 628 + ], + "type": "text", + "content": ". In the image-to-video variant of SVD, the condition " + }, + { + "bbox": [ + 55, + 567, + 295, + 628 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 567, + 295, + 628 + ], + "type": "text", + "content": " refers to the CLIP image embedding [49], replacing the typical text embeddings. For the remainder of this paper, we will refer to Eq. 1 as the video diffusion loss " + }, + { + "bbox": [ + 55, + 567, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{diff}}" + }, + { + "bbox": [ + 55, + 567, + 295, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": "Once trained, the diffusion model generates videos by iteratively denoising a noisy latent " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{z}_T^{1:N}" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": " sequence sampled from pure Gaussian distribution. At each diffusion step, the model predicts the noise in the input latent. Once the clean latent " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{z}_0^{1:N}" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": " is obtained, the decoder " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": " maps it to the higher-dimensional pixel space " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0^{1:N} = \\mathcal{D}(\\pmb{z}_0^{1:N})" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": ". For further details, we refer to the Appendix D of [4]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 270, + 453, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 270, + 453, + 282 + ], + "spans": [ + { + "bbox": [ + 314, + 270, + 453, + 282 + ], + "type": "text", + "content": "3.2. Video Diffusion Features" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 289, + 555, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 386 + ], + "type": "text", + "content": "Previous studies have demonstrated that image diffusion models learn discriminative features in their hidden states that are effective for various analysis tasks and propose methods for improving the representation power of such features [9, 69, 73, 74]. Similarly, we argue that while also being powerful, internal representations of pre-trained video diffusion models may not be fully temporally consistent, resulting in appearance drift in generated videos." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 388, + 556, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 556, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 556, + 544 + ], + "type": "text", + "content": "To better investigate this hypothesis, we first evaluate the long-term video tracking capabilities of U-Net-based video diffusion models [4, 56, 78]. Specifically, we evaluate the effectiveness of the features from each block of the U-Net for the task of point tracking. Given a real-world video, we add a small amount of noise and extract feature maps from each layer in each block. We perform a cosine-similarity-based nearest-neighbor search [44, 59] over these feature maps for a given set of fixed query points on the first frame (we use a similarity threshold of 0.6 [62] in our experiments). We also perform a similar analysis for generated videos where we extract the feature maps corresponding to diffusion steps with small amount of noise." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 545, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 714 + ], + "type": "text", + "content": "Based on this feature analysis, we make some important observations. Notably, regardless of the model (we analyze both Zeroscope T2V [56] and SVD I2V [4]), we find out that output features from the upsampler layer of the third decoder block consistently yield stronger temporal correspondences, as shown in Fig. 3. Hence, we use this block when extracting features for the remainder of our experiments. Furthermore, when we analyze generated videos and point tracks estimated based on the feature maps (as shown in Fig. 4), we observe that there is a correlation between tracking failures that reveal feature-space inconsistencies and appearance drifts that reveal pixel-space inconsistencies. Hence, we hypothesize that enriching feature consistency can help mitigate such appearance drifts. Next, we" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7278" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 68, + 549, + 201 + ], + "blocks": [ + { + "bbox": [ + 61, + 68, + 549, + 201 + ], + "lines": [ + { + "bbox": [ + 61, + 68, + 549, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 68, + 549, + 201 + ], + "type": "image", + "image_path": "c3d26ff034dd6c8539f4a1070898728c77377c605e1a87b947bd6d0e60718dfa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 202, + 555, + 237 + ], + "lines": [ + { + "bbox": [ + 55, + 202, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 555, + 237 + ], + "type": "text", + "content": "Figure 3. Real-world video tracking using different video diffusion features. Given color-coded query points on the first frame (Leftmost column), we display tracked points on target frames using features from different blocks (right columns). The 13th frame (first row) and 8th frame (second row) are shown as target frames. Full results are available in the supplementary and on our page." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 256, + 295, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 295, + 282 + ], + "type": "text", + "content": "introduce Track4Gen where we accomplish this goal by supervising video diffusion models with a joint tracking loss." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 290, + 132, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 132, + 303 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 132, + 303 + ], + "type": "text", + "content": "3.3. Track4Gen" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 308, + 295, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 308, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 54, + 308, + 295, + 499 + ], + "type": "text", + "content": "Track4Gen aims to utilize point tracking as an additional supervision signal to enhance the spatial-awareness of video diffusion features. Given that we build on top of a pretrained video generation model, to retain the prior knowledge and avoid tampering the original features directly, we propose a novel architecture change as shown in Fig. 2. Specifically, instead of directly using the raw diffusion features for correspondence estimation, we propose a trainable refiner module " + }, + { + "bbox": [ + 54, + 308, + 295, + 499 + ], + "type": "inline_equation", + "content": "R_{\\phi}" + }, + { + "bbox": [ + 54, + 308, + 295, + 499 + ], + "type": "text", + "content": ", which is designed to refine the raw features by projecting them into a correspondence-rich feature space. The refined features, which are spatially-aware, are then both used to estimate point tracks with an explicit supervision as well as feeding back to the generation backbone. We empirically find out that this design is more effective compared to fine-tuning the original model with no refinement module (see Sec. 4.2)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 501, + 295, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 501, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 67, + 501, + 295, + 514 + ], + "type": "text", + "content": "Given an " + }, + { + "bbox": [ + 67, + 501, + 295, + 514 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 67, + 501, + 295, + 514 + ], + "type": "text", + "content": "-frame video sequence " + }, + { + "bbox": [ + 67, + 501, + 295, + 514 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0^{1:N}" + }, + { + "bbox": [ + 67, + 501, + 295, + 514 + ], + "type": "text", + "content": ", its correspond-" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 57, + 541, + 176, + 674 + ], + "blocks": [ + { + "bbox": [ + 69, + 533, + 164, + 541 + ], + "lines": [ + { + "bbox": [ + 69, + 533, + 164, + 541 + ], + "spans": [ + { + "bbox": [ + 69, + 533, + 164, + 541 + ], + "type": "text", + "content": "Feature-space Inconsistency (SVD)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 541, + 176, + 674 + ], + "lines": [ + { + "bbox": [ + 57, + 541, + 176, + 674 + ], + "spans": [ + { + "bbox": [ + 57, + 541, + 176, + 674 + ], + "type": "image", + "image_path": "83341f49e67f07a973752f72dd26e68bfee31c1bf8a66a2b4c327ef72925575a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 677, + 295, + 710 + ], + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 710 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 710 + ], + "type": "text", + "content": "Figure 4. Generated video tracking using video diffusion features. Tracks based on diffusion features are annotated on the generated videos. Track4Gen generates more consistent results." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 176, + 541, + 294, + 674 + ], + "blocks": [ + { + "bbox": [ + 179, + 533, + 289, + 540 + ], + "lines": [ + { + "bbox": [ + 179, + 533, + 289, + 540 + ], + "spans": [ + { + "bbox": [ + 179, + 533, + 289, + 540 + ], + "type": "text", + "content": "Feature-space Consistency (Track4Gen)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 176, + 541, + 294, + 674 + ], + "lines": [ + { + "bbox": [ + 176, + 541, + 294, + 674 + ], + "spans": [ + { + "bbox": [ + 176, + 541, + 294, + 674 + ], + "type": "image", + "image_path": "0244d30198704051d756247245a1eea0793ec1bb096c52e163b987358ded1e32.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 256, + 553, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 553, + 304 + ], + "type": "text", + "content": "ing latent " + }, + { + "bbox": [ + 313, + 256, + 553, + 304 + ], + "type": "inline_equation", + "content": "z_0^{1:N}" + }, + { + "bbox": [ + 313, + 256, + 553, + 304 + ], + "type": "text", + "content": ", and a diffusion timestep " + }, + { + "bbox": [ + 313, + 256, + 553, + 304 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 256, + 553, + 304 + ], + "type": "text", + "content": ", in order to train Track4Gen we continue to utilize the standard diffusion training loss as defined in Eq. (1), where we adopt the velocity prediction objective [4, 35, 52] for " + }, + { + "bbox": [ + 313, + 256, + 553, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{diff}}" + }, + { + "bbox": [ + 313, + 256, + 553, + 304 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 305, + 554, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 554, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 554, + 428 + ], + "type": "text", + "content": "To enable tracking supervision, we assume access to a dense set of point trajectories " + }, + { + "bbox": [ + 313, + 305, + 554, + 428 + ], + "type": "inline_equation", + "content": "\\Omega = \\{(x^i,x^j)\\}" + }, + { + "bbox": [ + 313, + 305, + 554, + 428 + ], + "type": "text", + "content": " across frames where a point " + }, + { + "bbox": [ + 313, + 305, + 554, + 428 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^i" + }, + { + "bbox": [ + 313, + 305, + 554, + 428 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 313, + 305, + 554, + 428 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 305, + 554, + 428 + ], + "type": "text", + "content": " corresponds to a matching point " + }, + { + "bbox": [ + 313, + 305, + 554, + 428 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^j" + }, + { + "bbox": [ + 313, + 305, + 554, + 428 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 313, + 305, + 554, + 428 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 305, + 554, + 428 + ], + "type": "text", + "content": " and vice versa. Given the corresponding noisy video latent sequence " + }, + { + "bbox": [ + 313, + 305, + 554, + 428 + ], + "type": "inline_equation", + "content": "\\pmb{z}_t^{1:N}" + }, + { + "bbox": [ + 313, + 305, + 554, + 428 + ], + "type": "text", + "content": ", we first extract raw diffusion features as the hidden states " + }, + { + "bbox": [ + 313, + 305, + 554, + 428 + ], + "type": "inline_equation", + "content": "\\pmb{h}^{1:N} \\in \\mathbb{R}^{N \\times H \\times W \\times C}" + }, + { + "bbox": [ + 313, + 305, + 554, + 428 + ], + "type": "text", + "content": " from a specific block " + }, + { + "bbox": [ + 313, + 305, + 554, + 428 + ], + "type": "inline_equation", + "content": "b^k" + }, + { + "bbox": [ + 313, + 305, + 554, + 428 + ], + "type": "text", + "content": " within the U-Net, where " + }, + { + "bbox": [ + 313, + 305, + 554, + 428 + ], + "type": "inline_equation", + "content": "b^k" + }, + { + "bbox": [ + 313, + 305, + 554, + 428 + ], + "type": "text", + "content": " is set to the upsampler layer of the third decoder block (see Sec. 3.2). We then pass these features through the refiner module to obtain the refined feature map " + }, + { + "bbox": [ + 313, + 305, + 554, + 428 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{h}}^{1:N} = \\pmb{R}_{\\phi}(\\pmb{h}^{1:N})" + }, + { + "bbox": [ + 313, + 305, + 554, + 428 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 429, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 554, + 498 + ], + "type": "text", + "content": "We sample a query point " + }, + { + "bbox": [ + 313, + 429, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{q}}^{i}" + }, + { + "bbox": [ + 313, + 429, + 554, + 498 + ], + "type": "text", + "content": " along with its ground-truth target point " + }, + { + "bbox": [ + 313, + 429, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{trg}}^{j}" + }, + { + "bbox": [ + 313, + 429, + 554, + 498 + ], + "type": "text", + "content": " from the correspondence set " + }, + { + "bbox": [ + 313, + 429, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 429, + 554, + 498 + ], + "type": "text", + "content": ". Given the query point feature " + }, + { + "bbox": [ + 313, + 429, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{h}}^i (\\mathrm{x_q})\\in \\mathbb{R}^{1\\times 1\\times C}" + }, + { + "bbox": [ + 313, + 429, + 554, + 498 + ], + "type": "text", + "content": " and the target feature map " + }, + { + "bbox": [ + 313, + 429, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{h}}^j\\in \\mathbb{R}^{H\\times W\\times C}" + }, + { + "bbox": [ + 313, + 429, + 554, + 498 + ], + "type": "text", + "content": ", we calculate the cost volume " + }, + { + "bbox": [ + 313, + 429, + 554, + 498 + ], + "type": "inline_equation", + "content": "S\\in" + }, + { + "bbox": [ + 313, + 429, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{H\\times W\\times 1}" + }, + { + "bbox": [ + 313, + 429, + 554, + 498 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 367, + 507, + 553, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 507, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 367, + 507, + 553, + 525 + ], + "type": "interline_equation", + "content": "\\boldsymbol {S} (\\mathrm {p}) = \\cos - \\operatorname {s i m} \\left(\\tilde {\\boldsymbol {h}} ^ {i} \\left(\\mathrm {x} _ {\\mathrm {q}}\\right), \\tilde {\\boldsymbol {h}} ^ {j} (\\mathrm {p})\\right), \\tag {2}", + "image_path": "fd0f28ae0d82f891aa06fb0042658ca888ac35b833e4eefba085a63b18eb5d08.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 533, + 555, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 571 + ], + "type": "text", + "content": "where cos-sim denotes cosine similarity. The predicted target point " + }, + { + "bbox": [ + 313, + 533, + 555, + 571 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{\\mathrm{trg}}" + }, + { + "bbox": [ + 313, + 533, + 555, + 571 + ], + "type": "text", + "content": " is then determined using the differentiable soft-argmax operation:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 382, + 579, + 553, + 610 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 579, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 382, + 579, + 553, + 610 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} _ {\\mathrm {t r g}} = \\frac {\\sum_ {\\mathrm {p} \\in \\Omega^ {\\prime}} \\boldsymbol {S} (\\mathrm {p}) \\cdot \\mathrm {x} _ {\\mathrm {p}}}{\\sum_ {\\mathrm {p} \\in \\Omega^ {\\prime}} \\boldsymbol {S} (\\mathrm {p})}, \\tag {3}", + "image_path": "2fcf33f7fd7d2829c687f6454fe1f274fe56e94b7a68d1831f171601848fbb73.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 620, + 555, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 620, + 555, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 620, + 555, + 682 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 620, + 555, + 682 + ], + "type": "inline_equation", + "content": "\\Omega^{\\prime} = \\{p:\\| \\mathrm{x_p - x_{p_{max}}}\\| _2\\leq R\\} ^1" + }, + { + "bbox": [ + 313, + 620, + 555, + 682 + ], + "type": "text", + "content": " . Thus, the target point prediction can be expressed as " + }, + { + "bbox": [ + 313, + 620, + 555, + 682 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{\\mathrm{trg}} = \\xi (\\mathrm{x}_{\\mathrm{q}}^{i},j,\\tilde{\\boldsymbol{h}}^{1:N})" + }, + { + "bbox": [ + 313, + 620, + 555, + 682 + ], + "type": "text", + "content": " and the predicted tracklet for " + }, + { + "bbox": [ + 313, + 620, + 555, + 682 + ], + "type": "inline_equation", + "content": "\\mathrm{x_q^i}" + }, + { + "bbox": [ + 313, + 620, + 555, + 682 + ], + "type": "text", + "content": " is given by " + }, + { + "bbox": [ + 313, + 620, + 555, + 682 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{x_q^i}} = \\{\\hat{\\mathbf{x}}_n" + }, + { + "bbox": [ + 313, + 620, + 555, + 682 + ], + "type": "text", + "content": " .. " + }, + { + "bbox": [ + 313, + 620, + 555, + 682 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_n = \\xi (\\mathbf{x}_q^i,n,\\tilde{\\pmb{h}}^{1:N}),n = 1,\\dots,N\\}" + }, + { + "bbox": [ + 313, + 620, + 555, + 682 + ], + "type": "text", + "content": " Finally, the correspond" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 555, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 555, + 712 + ], + "type": "text", + "content": "The feature maps have a resolution of " + }, + { + "bbox": [ + 313, + 693, + 555, + 712 + ], + "type": "inline_equation", + "content": "44 \\times 81" + }, + { + "bbox": [ + 313, + 693, + 555, + 712 + ], + "type": "text", + "content": " for an input video resolution of " + }, + { + "bbox": [ + 313, + 693, + 555, + 712 + ], + "type": "inline_equation", + "content": "320 \\times 576" + }, + { + "bbox": [ + 313, + 693, + 555, + 712 + ], + "type": "text", + "content": ", and we set " + }, + { + "bbox": [ + 313, + 693, + 555, + 712 + ], + "type": "inline_equation", + "content": "R = 35" + }, + { + "bbox": [ + 313, + 693, + 555, + 712 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7279" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 79, + 156, + 356 + ], + "blocks": [ + { + "bbox": [ + 86, + 70, + 126, + 79 + ], + "lines": [ + { + "bbox": [ + 86, + 70, + 126, + 79 + ], + "spans": [ + { + "bbox": [ + 86, + 70, + 126, + 79 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 79, + 156, + 356 + ], + "lines": [ + { + "bbox": [ + 58, + 79, + 156, + 356 + ], + "spans": [ + { + "bbox": [ + 58, + 79, + 156, + 356 + ], + "type": "image", + "image_path": "bd21d1d1e4725d6bd405f75ea672a22dcd9fc971cfe51f587d57661f318b1c08.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 358, + 533, + 370 + ], + "lines": [ + { + "bbox": [ + 77, + 358, + 533, + 370 + ], + "spans": [ + { + "bbox": [ + 77, + 358, + 533, + 370 + ], + "type": "text", + "content": "Figure 5. Image-to-video generation results of the original SVD and Track4Gen. Please visit our page for full video view." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 157, + 79, + 354, + 356 + ], + "blocks": [ + { + "bbox": [ + 209, + 70, + 301, + 78 + ], + "lines": [ + { + "bbox": [ + 209, + 70, + 301, + 78 + ], + "spans": [ + { + "bbox": [ + 209, + 70, + 301, + 78 + ], + "type": "text", + "content": "Stable Video Diffusion (SVD)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 157, + 79, + 354, + 356 + ], + "lines": [ + { + "bbox": [ + 157, + 79, + 354, + 356 + ], + "spans": [ + { + "bbox": [ + 157, + 79, + 354, + 356 + ], + "type": "image", + "image_path": "b4bd58864b5e72ec37a126dcb9983fb8dbf6ad4feed6c7871a53498565368a84.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 355, + 79, + 553, + 356 + ], + "blocks": [ + { + "bbox": [ + 424, + 70, + 482, + 78 + ], + "lines": [ + { + "bbox": [ + 424, + 70, + 482, + 78 + ], + "spans": [ + { + "bbox": [ + 424, + 70, + 482, + 78 + ], + "type": "text", + "content": "Track4Gen (Ours)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 355, + 79, + 553, + 356 + ], + "lines": [ + { + "bbox": [ + 355, + 79, + 553, + 356 + ], + "spans": [ + { + "bbox": [ + 355, + 79, + 553, + 356 + ], + "type": "image", + "image_path": "dd5cbacaef5caf25b8cd78c19c35e4fee690de50cbcc63a9bc69e13c75217f69.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 391, + 295, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 295, + 403 + ], + "type": "text", + "content": "dence loss " + }, + { + "bbox": [ + 55, + 391, + 295, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{corr}}" + }, + { + "bbox": [ + 55, + 391, + 295, + 403 + ], + "type": "text", + "content": " is computed using the Huber loss " + }, + { + "bbox": [ + 55, + 391, + 295, + 403 + ], + "type": "inline_equation", + "content": "L_{H}" + }, + { + "bbox": [ + 55, + 391, + 295, + 403 + ], + "type": "text", + "content": " [31]:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 409, + 295, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 409, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 66, + 409, + 295, + 441 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o r r}} \\left(\\tilde {\\boldsymbol {h}} ^ {1: N}, \\Omega\\right) = \\sum_ {\\left(\\mathrm {x} _ {\\mathrm {q}} ^ {i}, \\mathrm {x} _ {\\text {t r g}} ^ {j}\\right) \\in \\Omega} L _ {H} \\left(\\xi \\left(\\mathrm {x} _ {\\mathrm {q}} ^ {i}, j, \\tilde {\\boldsymbol {h}} ^ {1: N}\\right), \\mathrm {x} _ {\\text {t r g}} ^ {j}\\right) \\tag {4}", + "image_path": "8391cdb3f2d203ca7c4a29b7775c3f53e7c16320bed07fbdb4cd5d10c82afbe8.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 448, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 296, + 582 + ], + "type": "text", + "content": "When training Track4Gen, we initialize the refiner module as an identity mapping to fully leverage the prior of the base model at the start of finetuning. To re-route the refined features to the backbone generator, we introduce a trainable zero convolution layer [77], denoted as " + }, + { + "bbox": [ + 55, + 448, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\zeta_{\\psi}" + }, + { + "bbox": [ + 55, + 448, + 296, + 582 + ], + "type": "text", + "content": ". While the diffusion loss " + }, + { + "bbox": [ + 55, + 448, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{diff}}" + }, + { + "bbox": [ + 55, + 448, + 296, + 582 + ], + "type": "text", + "content": " back-propagates to all the blocks of the video diffusion model, we detach the gradients of " + }, + { + "bbox": [ + 55, + 448, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{h}}^{1:N}" + }, + { + "bbox": [ + 55, + 448, + 296, + 582 + ], + "type": "text", + "content": " before passing into " + }, + { + "bbox": [ + 55, + 448, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\zeta_{\\psi}" + }, + { + "bbox": [ + 55, + 448, + 296, + 582 + ], + "type": "text", + "content": " such that refiner module can solely focus on acquiring the correspondence prior. Hence, given that the output of block " + }, + { + "bbox": [ + 55, + 448, + 296, + 582 + ], + "type": "inline_equation", + "content": "b^{k}" + }, + { + "bbox": [ + 55, + 448, + 296, + 582 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 55, + 448, + 296, + 582 + ], + "type": "inline_equation", + "content": "h^{1:N}" + }, + { + "bbox": [ + 55, + 448, + 296, + 582 + ], + "type": "text", + "content": ", the input to the subsequent block " + }, + { + "bbox": [ + 55, + 448, + 296, + 582 + ], + "type": "inline_equation", + "content": "b^{k+1}" + }, + { + "bbox": [ + 55, + 448, + 296, + 582 + ], + "type": "text", + "content": " is computed as " + }, + { + "bbox": [ + 55, + 448, + 296, + 582 + ], + "type": "inline_equation", + "content": "h^{1:N} + \\zeta_{\\psi}" + }, + { + "bbox": [ + 55, + 448, + 296, + 582 + ], + "type": "text", + "content": " (stop-gradient " + }, + { + "bbox": [ + 55, + 448, + 296, + 582 + ], + "type": "inline_equation", + "content": "(R_{\\phi}(h^{1:N}))" + }, + { + "bbox": [ + 55, + 448, + 296, + 582 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 582, + 295, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 606 + ], + "type": "text", + "content": "Fig. 2 visualizes this architecture design, with red and green colors indicating the objective that optimizes each module." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 616, + 137, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 137, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 137, + 630 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 635, + 188, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 188, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 188, + 648 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "To train Track4Gen, we construct a training dataset consisting of 567 video-trajectory pairs, with each video having a resolution of " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "320 \\times 576" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": " and a duration of 24 frames. Since no real-world video with (dense) ground-truth trajectory annotations exist at the time of this work, we utilize" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 391, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 555, + 475 + ], + "type": "text", + "content": "optical flow to generate trajectory annotations. A key challenge is the need for accurate video segmentation maps to ensure a balanced distribution of trajectory points between foreground objects and the background [12]. To address this, we utilize public video datasets paired with ground-truth segmentation maps [7, 17, 39, 46, 48], where we split longer videos into 24-frame segments." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 475, + 556, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 556, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 556, + 681 + ], + "type": "text", + "content": "We use Stable Video Diffusion (SVD) image-to-video pretrained checkpoints" + }, + { + "bbox": [ + 313, + 475, + 556, + 681 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 313, + 475, + 556, + 681 + ], + "type": "text", + "content": " as the base video generator. Our proposed refiner module consists of eight stacked 2D convolution layers and is attached to the third decoder block of the SVD UNet. The refiner module preserves the shape of the hidden states throughout and is initialized as the identity mapping. Further details are provided in the supplementary. We finetune this enhanced video generator architecture for " + }, + { + "bbox": [ + 313, + 475, + 556, + 681 + ], + "type": "inline_equation", + "content": "20\\mathrm{K}" + }, + { + "bbox": [ + 313, + 475, + 556, + 681 + ], + "type": "text", + "content": " steps with our joint loss " + }, + { + "bbox": [ + 313, + 475, + 556, + 681 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{diff}} + \\lambda \\mathcal{L}_{\\mathrm{corr}}" + }, + { + "bbox": [ + 313, + 475, + 556, + 681 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 475, + 556, + 681 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 475, + 556, + 681 + ], + "type": "text", + "content": " is set to 8. Rather than finetuning the entire model, we finetune only the temporal transformer blocks, the refiner module " + }, + { + "bbox": [ + 313, + 475, + 556, + 681 + ], + "type": "inline_equation", + "content": "R_{\\phi}" + }, + { + "bbox": [ + 313, + 475, + 556, + 681 + ], + "type": "text", + "content": ", and the zero convolution " + }, + { + "bbox": [ + 313, + 475, + 556, + 681 + ], + "type": "inline_equation", + "content": "\\zeta_{\\psi}" + }, + { + "bbox": [ + 313, + 475, + 556, + 681 + ], + "type": "text", + "content": ". In each iteration, we sample 512 correspondence pairs from the precomputed trajectories. We use the AdamW optimizer [43] with a learning rate of " + }, + { + "bbox": [ + 313, + 475, + 556, + 681 + ], + "type": "inline_equation", + "content": "1e - 5" + }, + { + "bbox": [ + 313, + 475, + 556, + 681 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 475, + 556, + 681 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 313, + 475, + 556, + 681 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 475, + 556, + 681 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 313, + 475, + 556, + 681 + ], + "type": "text", + "content": ", and a weight decay of " + }, + { + "bbox": [ + 313, + 475, + 556, + 681 + ], + "type": "inline_equation", + "content": "1e - 2" + }, + { + "bbox": [ + 313, + 475, + 556, + 681 + ], + "type": "text", + "content": ". We train the model on " + }, + { + "bbox": [ + 313, + 475, + 556, + 681 + ], + "type": "inline_equation", + "content": "4\\times" + }, + { + "bbox": [ + 313, + 475, + 556, + 681 + ], + "type": "text", + "content": " H100 GPUs with a total batch size of 4. For sampling new videos, we apply the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "text", + "content": "2https://huggingface.co/stabilityai/stable-video-diffusion-image2vid-xt" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7280" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 68, + 555, + 311 + ], + "blocks": [ + { + "bbox": [ + 58, + 68, + 555, + 311 + ], + "lines": [ + { + "bbox": [ + 58, + 68, + 555, + 311 + ], + "spans": [ + { + "bbox": [ + 58, + 68, + 555, + 311 + ], + "type": "image", + "image_path": "2007efe96eb2e6f569980b87b25183468361e418d42b4a9f51fa2c9b6ddf40b8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 313, + 555, + 336 + ], + "lines": [ + { + "bbox": [ + 55, + 313, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 555, + 336 + ], + "type": "text", + "content": "Figure 6. Qualitative ablation on video generation. Track4Gen is compared with finetuned SVD (SVD finetuned on the same training videos without any correspondence supervision) and Track4Gen trained without the Refiner module." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 357, + 295, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 295, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 295, + 381 + ], + "type": "text", + "content": "default settings using 30 steps with the EDM sampler [35], motion bucket id =127, and fps=7." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 459, + 296, + 509 + ], + "blocks": [ + { + "bbox": [ + 55, + 391, + 296, + 456 + ], + "lines": [ + { + "bbox": [ + 55, + 391, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 296, + 456 + ], + "type": "text", + "content": "Table 1. Quantitative comparison on video generation performance. We compare Track4Gen to the pre-trained SVD " + }, + { + "bbox": [ + 55, + 391, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 55, + 391, + 296, + 456 + ], + "type": "text", + "content": " as well as a finetuned SVD on the same dataset (finetuned SVD). We also train a variant of Track4Gen without the refiner module. All videos are generated at " + }, + { + "bbox": [ + 55, + 391, + 296, + 456 + ], + "type": "inline_equation", + "content": "320\\mathrm{x}576" + }, + { + "bbox": [ + 55, + 391, + 296, + 456 + ], + "type": "text", + "content": " resolution, except SVD " + }, + { + "bbox": [ + 55, + 391, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 55, + 391, + 296, + 456 + ], + "type": "text", + "content": " (576p) which operates at " + }, + { + "bbox": [ + 55, + 391, + 296, + 456 + ], + "type": "inline_equation", + "content": "576\\mathrm{x}1024" + }, + { + "bbox": [ + 55, + 391, + 296, + 456 + ], + "type": "text", + "content": " resolution." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 459, + 296, + 509 + ], + "lines": [ + { + "bbox": [ + 56, + 459, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 459, + 296, + 509 + ], + "type": "table", + "html": "
Subject ConsistencyTemporal FlickeringMotion SmoothnessImaging QualityVideo-Image AlignmentFIDFVD
SVD*0.95350.94640.97740.66480.953929.0776
finetuned SVD0.96650.98000.99090.67660.977127.0735
Track4Gen w/o refiner0.95060.97250.97910.66530.961427.1718
Track4Gen0.97460.98060.99210.68350.981426.6724
SVD* (576p)0.95760.94780.97950.68120.9582--
", + "image_path": "fc64a6199491e85adc14954283930459f20aa27ea8a6506e901d8ec729370e25.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 529, + 234, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 234, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 234, + 540 + ], + "type": "text", + "content": "4.2. Track4Gen for Video Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 546, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 581 + ], + "type": "text", + "content": "We evaluate Track4Gen for the image-to-video generation task via a series of experiments using multiple datasets, automated metrics, and human evaluations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "content": "Evaluation Setup. We compare Track4Gen against the original SVD " + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "inline_equation", + "content": "(\\mathrm{SVD}^*)" + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "content": " [4], as well as a version of SVD that is finetuned on the same videos as Track4Gen (finetuned SVD). Furthermore, we train a variant of Track4Gen without the refiner module. For VBenchmark metrics [30], evaluations are conducted on the VBenchmark-I2V dataset, containing 355 diverse images. FID and FVD are measured using the DAVIS [48] dataset as reference. We generate 24-frame videos conditioned on each input image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Automatic metrics. We first report five key metrics from VBench [30]: (1) Subject Consistency—assesses" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "text", + "content": "subject appearance consistency of the video by computing the similarity of DINO [45] features. (2) Temporal Flickering—detects temporal consistency by taking static frames and calculating the mean absolute difference across frames. (3) Motion Smoothness—measures smoothness of motion, and how well it adheres to real-world physics, using video frame interpolation priors [42]. (4) Image Quality—evaluates distortions (e.g., noise, blur) using a pretrained, multi-scale image quality predictor [36]. (5) Video-Image Alignment—measures alignment between the subject in the input image and in the generated video using DINO features. We additionally report FID [26] and FVD [63]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 505, + 555, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 555, + 590 + ], + "type": "text", + "content": "Human evaluation. We further evaluate Track4Gen against baselines through a user study. We ask 64 participants to compare our results with a randomly selected baseline. We ask the users to evaluate how consistent main objects appear across the frames in a generated video as well as how natural the depicted motion is. We provide further details of the user study in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": "Qualitative results. Qualitative comparisons with the base SVD are shown in Fig. 5. As illustrated, Track4Gen generates videos with strong appearance consistency, avoiding issues of appearance drift. In contrast, videos produced by the original SVD exhibit noticeable inconsistencies: the sheep's head (row 1) mutates, the plane's wing (row 2) shows unnatural transitions, and the cars (row 3) disappear. Further comparisons with finetuned SVD and Track4Gen without the refiner module are shown in Fig. 6 and highlight the superior visual coherence of the proposed Track4Gen." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "7281" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 79, + 181, + 274 + ], + "blocks": [ + { + "bbox": [ + 96, + 70, + 141, + 79 + ], + "lines": [ + { + "bbox": [ + 96, + 70, + 141, + 79 + ], + "spans": [ + { + "bbox": [ + 96, + 70, + 141, + 79 + ], + "type": "text", + "content": "Query Points" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 79, + 181, + 274 + ], + "lines": [ + { + "bbox": [ + 57, + 79, + 181, + 274 + ], + "spans": [ + { + "bbox": [ + 57, + 79, + 181, + 274 + ], + "type": "image", + "image_path": "89b328905f0a6f1472a28dcc4c4c46facf48d8876c515a62ea87af911aa790c0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 276, + 553, + 299 + ], + "lines": [ + { + "bbox": [ + 55, + 276, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 553, + 299 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison of Track4Gen and baselines for real-world video tracking. The leftmost column displays query points in the first frame, while the following three columns show tracking results using features from each model." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 182, + 79, + 306, + 274 + ], + "blocks": [ + { + "bbox": [ + 212, + 70, + 275, + 79 + ], + "lines": [ + { + "bbox": [ + 212, + 70, + 275, + 79 + ], + "spans": [ + { + "bbox": [ + 212, + 70, + 275, + 79 + ], + "type": "text", + "content": "Track4Gen (Ours)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 182, + 79, + 306, + 274 + ], + "lines": [ + { + "bbox": [ + 182, + 79, + 306, + 274 + ], + "spans": [ + { + "bbox": [ + 182, + 79, + 306, + 274 + ], + "type": "image", + "image_path": "471eabf03fc07b9cf3fa0b5ed52d1690cf1344f5750cc99bd0ccacfafb3bb601.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 306, + 79, + 429, + 274 + ], + "blocks": [ + { + "bbox": [ + 352, + 70, + 383, + 78 + ], + "lines": [ + { + "bbox": [ + 352, + 70, + 383, + 78 + ], + "spans": [ + { + "bbox": [ + 352, + 70, + 383, + 78 + ], + "type": "text", + "content": "SVD 12V" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 79, + 429, + 274 + ], + "lines": [ + { + "bbox": [ + 306, + 79, + 429, + 274 + ], + "spans": [ + { + "bbox": [ + 306, + 79, + 429, + 274 + ], + "type": "image", + "image_path": "3e9dc7be668900098ba8e5d2da1ed158065ecc773d674a1a4be1d5d83d3a6ecc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 430, + 79, + 553, + 274 + ], + "blocks": [ + { + "bbox": [ + 463, + 70, + 518, + 78 + ], + "lines": [ + { + "bbox": [ + 463, + 70, + 518, + 78 + ], + "spans": [ + { + "bbox": [ + 463, + 70, + 518, + 78 + ], + "type": "text", + "content": "ZeroScope T2V" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 430, + 79, + 553, + 274 + ], + "lines": [ + { + "bbox": [ + 430, + 79, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 430, + 79, + 553, + 274 + ], + "type": "image", + "image_path": "7faa10865074e5889f0db1d531836c40a01d254c72c9fc1fa9c994d92ac8a5e2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 57, + 308, + 171, + 388 + ], + "blocks": [ + { + "bbox": [ + 57, + 308, + 171, + 388 + ], + "lines": [ + { + "bbox": [ + 57, + 308, + 171, + 388 + ], + "spans": [ + { + "bbox": [ + 57, + 308, + 171, + 388 + ], + "type": "image", + "image_path": "0f2ac8d7130d7332c0140967b0d794c2808e1e2f98140b56f91d304a3b989d40.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 388, + 153, + 398 + ], + "lines": [ + { + "bbox": [ + 74, + 388, + 153, + 398 + ], + "spans": [ + { + "bbox": [ + 74, + 388, + 153, + 398 + ], + "type": "text", + "content": "(a) Identity preservation" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 400, + 295, + 422 + ], + "lines": [ + { + "bbox": [ + 55, + 400, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 295, + 422 + ], + "type": "text", + "content": "Figure 8. User study results. Our study shows that Track4Gen better preserves object identity and produces more natural motion." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 182, + 310, + 294, + 387 + ], + "blocks": [ + { + "bbox": [ + 182, + 310, + 294, + 387 + ], + "lines": [ + { + "bbox": [ + 182, + 310, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 182, + 310, + 294, + 387 + ], + "type": "image", + "image_path": "d6b4d5d15a2c235187bff389aad213516d46ebb79126cc780c35ca89eb6ba6b4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 388, + 276, + 397 + ], + "lines": [ + { + "bbox": [ + 201, + 388, + 276, + 397 + ], + "spans": [ + { + "bbox": [ + 201, + 388, + 276, + 397 + ], + "type": "text", + "content": "(b) Motion naturalness" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 433, + 296, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 296, + 517 + ], + "type": "text", + "content": "Quantitative results. As shown in Tab. 1, our method achieves the highest scores across all 5 metrics from VBench, along with the lowest FID and second-lowest FVD values, outperforming the base SVD by substantial margins. Fig. 8 provides the user study results where the majority of the participants agreed that Track4Gen is superior both in terms of identity preservation and naturalness of motion." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 523, + 223, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 523, + 223, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 523, + 223, + 536 + ], + "type": "text", + "content": "4.3. Track4Gen for Video Tracking" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 539, + 295, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 295, + 635 + ], + "type": "text", + "content": "We evaluate Track4Gen's capability to track any point in real videos by adding a small amount of noise to the input video [59] and passing it through the video denoiser " + }, + { + "bbox": [ + 55, + 539, + 295, + 635 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{\\theta}" + }, + { + "bbox": [ + 55, + 539, + 295, + 635 + ], + "type": "text", + "content": " to extract feature maps. We first compare tracking results with such features against other raw features [4, 56, 60] in Sec. 4.3.1. In Sec. 4.3.2, we utilize Track4Gen's features in a test-time optimization method [62] and compare to both self-supervised and fully supervised video trackers." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 639, + 216, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 216, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 216, + 651 + ], + "type": "text", + "content": "4.3.1. Zero-shot Feature Comparison" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "We evaluate the precision of predicted tracks using the features from Track4Gen, the original SVD model " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\mathrm{SVD}^{\\star})" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " and RAFT [60]. We also test another text-to-video model, ZeroScope T2V [56], to demonstrate how raw features from pre-trained video generators typically work out of the box." + } + ] + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 315, + 357, + 555, + 417 + ], + "blocks": [ + { + "bbox": [ + 313, + 311, + 554, + 355 + ], + "lines": [ + { + "bbox": [ + 313, + 311, + 554, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 554, + 355 + ], + "type": "text", + "content": "Table 2. Quantitative zero-shot feature comparison on video tracking benchmarks. Track4Gen features are compared to the features of SVD* [4], ZeroScope [56], and RAFT [60]. For all the metrics, higher values indicate better performance." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 357, + 555, + 417 + ], + "lines": [ + { + "bbox": [ + 315, + 357, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 315, + 357, + 555, + 417 + ], + "type": "table", + "html": "
MethodDAVIS-480p(24-frame)BADJA(24-frame)DAVIS-480p(whole duration)BADJA(whole duration)
\\( \\delta^{x}_{avg} \\)OAAJ\\( \\delta^{seg} \\)\\( \\delta^{3px} \\)\\( \\delta^{x}_{avg} \\)OAAJ\\( \\delta^{seg} \\)\\( \\delta^{3px} \\)
ZeroScope46.267.039.427.52.837.259.527.819.92.0
SVD*42.479.736.426.22.935.470.126.519.42.2
Track4Gen69.785.856.552.37.758.978.440.240.45.0
RAFT73.3--54.88.766.7--45.05.8
", + "image_path": "f8f9f1546315dac740978bb4462735c599c7c2b62193d532e418f98078c404f2.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 433, + 554, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 554, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 554, + 468 + ], + "type": "text", + "content": "For RAFT, tracking is achieved by chaining optical flow displacements, while the others use nearest neighbor matching between its encoded features." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 472, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 554, + 544 + ], + "type": "text", + "content": "Datasets. We use TAP-Vid DAVIS [12] and BADJA [3] as benchmark datasets. Additionally, we include two shorter benchmarks, DAVIS (24-frame) and BADJA (24-frame), which focus on the first 24 frames with query and target points within this range. Details on encoding long videos with the video models are in the supplemental." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": "Metrics. For evaluating the TAP-Vid benchmarks, we use the following metrics: (i) Position Accuracy " + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "inline_equation", + "content": "(\\delta_{avg}^{x})" + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": " evaluates the average accuracy of visible points, where each " + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\delta^x" + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": " represents the fraction of predicted points that lie within " + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": " pixels of the ground-truth position, with " + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "inline_equation", + "content": "x \\in \\{1, 2, 4, 8, 16\\}" + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": ". (ii) Occlusion Accuracy (OA) evaluates the correctness of occlusion predictions. (iii) Average Jaccard (AJ) jointly assesses both position and occlusion accuracy. For the BADJA dataset, we report " + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\delta^{seg}" + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": ", which measures the accuracy of tracked keypoints within a distance of " + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "inline_equation", + "content": "0.2\\sqrt{A}" + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": " from the ground-truth annotation, where " + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": " is the area of the foreground object. We also report " + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\delta^{3px}" + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": ", which assesses accuracy within a 3-pixel threshold. A cosine similarity threshold of 0.6 is used for occlusion prediction." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7282" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 76, + 136, + 209 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 127, + 76 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 127, + 76 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 127, + 76 + ], + "type": "text", + "content": "First Frame with Query Points" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 76, + 136, + 209 + ], + "lines": [ + { + "bbox": [ + 57, + 76, + 136, + 209 + ], + "spans": [ + { + "bbox": [ + 57, + 76, + 136, + 209 + ], + "type": "image", + "image_path": "620611b3f3f85c7a27be961227007156bb3ec0f3fa0572c87926f784cd601428.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 212, + 294, + 224 + ], + "lines": [ + { + "bbox": [ + 56, + 212, + 294, + 224 + ], + "spans": [ + { + "bbox": [ + 56, + 212, + 294, + 224 + ], + "type": "text", + "content": "Figure 9. Extending Track4Gen with test-time adaptation [62]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 136, + 76, + 294, + 209 + ], + "blocks": [ + { + "bbox": [ + 175, + 70, + 255, + 76 + ], + "lines": [ + { + "bbox": [ + 175, + 70, + 255, + 76 + ], + "spans": [ + { + "bbox": [ + 175, + 70, + 255, + 76 + ], + "type": "text", + "content": "Tracking Results in Subsequent Frames" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 136, + 76, + 294, + 209 + ], + "lines": [ + { + "bbox": [ + 136, + 76, + 294, + 209 + ], + "spans": [ + { + "bbox": [ + 136, + 76, + 294, + 209 + ], + "type": "image", + "image_path": "c67742c215bddb707aac03f8fff1c06e96eeb62ceba6f34f040bdba252afead8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 56, + 292, + 294, + 378 + ], + "blocks": [ + { + "bbox": [ + 55, + 233, + 295, + 289 + ], + "lines": [ + { + "bbox": [ + 55, + 233, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 295, + 289 + ], + "type": "text", + "content": "Table 3. Quantitative comparison with video trackers. Although primarily designed for video generation, Track4Gen combined with a test-time optimization method [62] achieves performance comparable to dedicated video tracking frameworks, even when compared to supervised methods." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 292, + 294, + 378 + ], + "lines": [ + { + "bbox": [ + 56, + 292, + 294, + 378 + ], + "spans": [ + { + "bbox": [ + 56, + 292, + 294, + 378 + ], + "type": "table", + "html": "
MethodDAVIS-480BADJA
δxavgOAAJδsegδ3px
TAP-Net*66.479.046.045.49.6
PIPs++*73.6--59.09.8
TAPIR*77.389.565.768.710.5
Omnimotion†74.184.558.445.26.9
DINO-Tracker†80.488.164.672.414.3
DINO-Tracker w/ Track4Gen†72.584.555.748.410.9
", + "image_path": "52351e7a6e44c7fbb0938716533d442c069954680d873afd0292309b3fd11e59.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 57, + 379, + 171, + 388 + ], + "lines": [ + { + "bbox": [ + 57, + 379, + 171, + 388 + ], + "spans": [ + { + "bbox": [ + 57, + 379, + 171, + 388 + ], + "type": "text", + "content": "\\* - supervised. -test-time training." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 56, + 407, + 294, + 477 + ], + "blocks": [ + { + "bbox": [ + 76, + 394, + 273, + 405 + ], + "lines": [ + { + "bbox": [ + 76, + 394, + 273, + 405 + ], + "spans": [ + { + "bbox": [ + 76, + 394, + 273, + 405 + ], + "type": "text", + "content": "Table 4. Ablation on trainable modules and refiner." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 407, + 294, + 477 + ], + "lines": [ + { + "bbox": [ + 56, + 407, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 407, + 294, + 477 + ], + "type": "table", + "html": "
Trainable modulesSubject ConsistencyTemporal FlickeringMotion SmoothnessImaging QualityVideo-Image Alignment
spatial + temporal0.97340.98110.99170.68630.9807
spatial0.97260.98010.99190.68520.9811
temporal0.97460.98060.99210.68350.9814
Refiner architecture
2D convolutions0.97460.98060.99210.68350.9814
3D convolutions0.96870.97340.99040.68330.9820
", + "image_path": "beb2795496a6d71f289ac6415c55f7d6bcc18afcaf8eaa2e2f1075d0b0b2746d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 497, + 295, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 295, + 557 + ], + "type": "text", + "content": "Results. We present the qualitative results in Fig. 7 and the quantitative results in Tab. 2. Although primarily designed for video generation, Track4Gen boosts the poor performance of the pre-trained video models significantly, approaching the accuracy of RAFT optical flow chaining." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 565, + 292, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 292, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 292, + 578 + ], + "type": "text", + "content": "4.3.2. Extending Track4Gen with Test-time Adaptation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 581, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 295, + 677 + ], + "type": "text", + "content": "To further evaluate Track4Gen's long-term tracking capabilities, we integrate our features with test-time adaptation algorithm of DINO-Tracker [62], where a per-video optimization is performed using optical flow supervision. We replace the originally used DINOv2 [45] with the features from Track4Gen. We evaluate using the same datasets and metrics outlined in Sec. 4.3.1, against both fully-supervised trackers [12, 13, 79] and self-supervised methods [62, 65]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "content": "Tab. 3 shows that Track4Gen features optimized with [62] achieve performance comparable to dedicated trackers. Qualitative results are in Fig. 9 and in the supplemental." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 315, + 106, + 553, + 148 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "type": "text", + "content": "Table 5. Quantitative ablation on using annotated, but synthetic videos [20]. Left: Video generation metrics. Right: Video tracking metrics." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 106, + 553, + 148 + ], + "lines": [ + { + "bbox": [ + 315, + 106, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 315, + 106, + 553, + 148 + ], + "type": "table", + "html": "
Dataset compositionSubject ConsistencyMotion SmoothnessImaging QualityBADJA
δsegδ3px
real videos0.97470.99210.683340.45.0
real + synthetic videos0.97080.98920.679342.14.8
", + "image_path": "5ce062a81c9ee0d223a67b8e25a6535a1d61d5e614c05b93b249ef347120efad.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 162, + 415, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 162, + 415, + 173 + ], + "spans": [ + { + "bbox": [ + 313, + 162, + 415, + 173 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 179, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 179, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 179, + 555, + 384 + ], + "type": "text", + "content": "We present an ablation study in Tab. 4 where we train different set of modules. Each spatio-temporal block of SVD includes both spatial and temporal transformers. We compare training only spatial transformers, only temporal transformers, or both. We also ablate the architecture of the refiner module using either 2D or 3D convolution layers. Our analysis shows that while results are similar across settings, training only the temporal transformers in SVD with 2D convolutions as the refiner module yields optimal video generation quality. We further analyze our training dataset by additionally incorporating Kubric [20] simulated videos (1K video-track pairs from the Panning MOVi-E data [10, 13]) with automatically annotated trajectories into training. As shown in Tab. 5, optical flow-chained tracklets from real provides provide as effective correspondence guidance as tracklets from synthetic data, while synthetic videos negatively impact the video generation quality." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 394, + 479, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 479, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 479, + 406 + ], + "type": "text", + "content": "5. Conclusion and Future Work" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 414, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 554, + 510 + ], + "type": "text", + "content": "We have presented the first unified framework that bridges two distinct tasks: video generation and dense point tracking. We demonstrated that this produces temporally consistent feature representations and appearance-consistent videos. As for limitations, videos generated by Track4Gen tend to exhibit less dynamic motion compared to those from other video generators. Additionally, failure cases are included in the supplementary material." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 510, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 581 + ], + "type": "text", + "content": "Future work. Recently, cutting-edge video trackers [10, 14, 34] have emerged, enabling dense, accurate, and long-term tracking, especially with better handling of occlusions. This opens up promising future directions for extending our work to utilize real-world videos, automatically annotated by these advanced trackers." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "Acknowledgments. We thank Seokju Cho and Narek Tumanyan for their invaluable feedback on video point tracking. We also extend our gratitude to Mingi Kwon, Joon-Young Lee, and Gabriel Huang for their insightful discussions. Hyeonho Jeong and Jong Chul Ye are supported by the National Research Foundation of Korea (NRF) under Grants RS-2024-00336454 and RS-2023-00262527, and by the Institute for Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (RS-2019-II190075, Artificial Intelligence Graduate School Program, KAIST)." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7283" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "text", + "content": "[1] Gorkay Aydemir, Weidi Xie, and Fatma Güney. Can visual foundation models achieve long-term point tracking? arXiv preprint arXiv:2408.13575, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 167 + ], + "type": "text", + "content": "[2] Weikang Bian, Zhaoyang Huang, Xiaoyu Shi, Yitong Dong, Yijin Li, and Hongsheng Li. Context-tap: Tracking any point demands spatial context features. arXiv preprint arXiv:2306.02000, 3, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 170, + 296, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 170, + 296, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 170, + 296, + 235 + ], + "type": "text", + "content": "[3] Benjamin Biggs, Thomas Roddick, Andrew Fitzgibbon, and Roberto Cipolla. Creatures great and small: Recovering the shape and motion of animals from video. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part V 14, pages 3-19. Springer, 2019. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 295, + 300 + ], + "type": "text", + "content": "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 302, + 295, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 295, + 367 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 295, + 367 + ], + "type": "text", + "content": "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 369, + 295, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 295, + 422 + ], + "type": "text", + "content": "[6] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 424, + 295, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 424, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 424, + 295, + 468 + ], + "type": "text", + "content": "[7] Sergi Caelles, Alberto Montes, Kevis-Kokitsi Maninis, Yuhua Chen, Luc Van Gool, Federico Perazzi, and Jordi Pont-Tuset. The 2018 davis challenge on video object segmentation. arXiv preprint arXiv:1803.00557, 2018. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 469, + 295, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 295, + 523 + ], + "type": "text", + "content": "[8] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 525, + 295, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 295, + 556 + ], + "type": "text", + "content": "[9] X Chen, Z Liu, S Xie, and K He. Deconstructing denoising diffusion models for self-supervised learning. arxiv 2024. arXiv preprint arXiv:2401.14404. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 558, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 295, + 601 + ], + "type": "text", + "content": "[10] Seokju Cho, Jiahui Huang, Jisu Nam, Honggyu An, Seun-gryong Kim, and Joon-Young Lee. Local all-pair correspondence for point tracking. arXiv preprint arXiv:2407.15420, 2024. 2, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 602, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 635 + ], + "type": "text", + "content": "[11] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 636, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 689 + ], + "type": "text", + "content": "[12] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610-13626, 2022. 2, 5, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[13] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 554, + 116 + ], + "type": "text", + "content": "Tapir: Tracking any point with per-frame initialization and temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10061-10072, 2023. 2, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 173 + ], + "type": "text", + "content": "[14] Carl Doersch, Yi Yang, Dilara Gokay, Pauline Luc, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ross Goroshin, João Carreira, and Andrew Zisserman. Bootstrap: Bootstrapped training for tracking-any-point. arXiv preprint arXiv:2402.00847, 2024. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 175, + 554, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 554, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 554, + 239 + ], + "type": "text", + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 241, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 241, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 241, + 554, + 297 + ], + "type": "text", + "content": "[16] Niladri Shekhar Dutt, Sanjeev Muralikrishnan, and Niloy J. Mitra. Diffusion 3d features (diff3f): Decorating textured shapes with distilled semantic features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4494-4504, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 298, + 554, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 298, + 554, + 341 + ], + "spans": [ + { + "bbox": [ + 317, + 298, + 554, + 341 + ], + "type": "text", + "content": "[17] Qingnan Fan, Fan Zhong, Dani Lischinski, Daniel Cohen-Or, and Baoquan Chen. Jumpcut: non-successive mask transfer and interpolation for video cutout. ACM Trans. Graph., 34 (6):195-1, 2015. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 343, + 554, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 343, + 554, + 386 + ], + "spans": [ + { + "bbox": [ + 317, + 343, + 554, + 386 + ], + "type": "text", + "content": "[18] Stephanie Fu, Mark Hamilton, Laura E. Brandt, Axel Feldmann, Zhoutong Zhang, and William T. Freeman. Feuput: A model-agnostic framework for features at any resolution. In ICLR, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 388, + 554, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 554, + 421 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 554, + 421 + ], + "type": "text", + "content": "[19] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 422, + 554, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 554, + 488 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 554, + 488 + ], + "type": "text", + "content": "[20] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3749-3761, 2022. 2, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 490, + 554, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 490, + 554, + 566 + ], + "spans": [ + { + "bbox": [ + 317, + 490, + 554, + 566 + ], + "type": "text", + "content": "[21] Yuchao Gu, Yipin Zhou, Bichen Wu, Licheng Yu, Jia-Wei Liu, Rui Zhao, Jay Zhangjie Wu, David Junhao Zhang, Mike Zheng Shou, and Kevin Tang. Videoswap: Customized video subject swapping with interactive semantic point correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7621–7630, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 568, + 554, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 554, + 623 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 554, + 623 + ], + "type": "text", + "content": "[22] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 624, + 554, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 554, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 554, + 668 + ], + "type": "text", + "content": "[23] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In European Conference on Computer Vision, pages 59-75. Springer, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 670, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 554, + 712 + ], + "type": "text", + "content": "[24] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221, 2022. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7284" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[25] Eric Hedlin, Gopal Sharma, Shweta Mahajan, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. Unsupervised semantic correspondence using stable diffusion. In NIPS, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 119, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 294, + 173 + ], + "type": "text", + "content": "[26] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 208 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 208 + ], + "type": "text", + "content": "[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 211, + 294, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 211, + 294, + 265 + ], + "spans": [ + { + "bbox": [ + 56, + 211, + 294, + 265 + ], + "type": "text", + "content": "[28] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 267, + 294, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 267, + 294, + 311 + ], + "spans": [ + { + "bbox": [ + 56, + 267, + 294, + 311 + ], + "type": "text", + "content": "[29] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 313, + 294, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 313, + 294, + 379 + ], + "spans": [ + { + "bbox": [ + 56, + 313, + 294, + 379 + ], + "type": "text", + "content": "[30] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 381, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 381, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 381, + 294, + 414 + ], + "type": "text", + "content": "[31] Peter J Huber. Robust estimation of a location parameter. In Breakthroughs in statistics: Methodology and distribution, pages 492-518. Springer, 1992. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 416, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 294, + 458 + ], + "type": "text", + "content": "[32] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances in neural information processing systems, 33:19545-19560, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 462, + 294, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 462, + 294, + 505 + ], + "spans": [ + { + "bbox": [ + 56, + 462, + 294, + 505 + ], + "type": "text", + "content": "[33] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv preprint arXiv:2307.07635, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 508, + 294, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 508, + 294, + 561 + ], + "spans": [ + { + "bbox": [ + 56, + 508, + 294, + 561 + ], + "type": "text", + "content": "[34] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker3: Simpler and better point tracking by pseudolabelling real videos. arXiv preprint arXiv:2410.11831, 2024. 2, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 565, + 294, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 565, + 294, + 608 + ], + "spans": [ + { + "bbox": [ + 56, + 565, + 294, + 608 + ], + "type": "text", + "content": "[35] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. 3, 4, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 610, + 294, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 610, + 294, + 654 + ], + "spans": [ + { + "bbox": [ + 56, + 610, + 294, + 654 + ], + "type": "text", + "content": "[36] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 656, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 656, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 656, + 294, + 677 + ], + "type": "text", + "content": "[37] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "text", + "content": "[38] Alexander C. Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your diffusion model is secretly a zero-shot classifier. In ICCV, 2023. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "type": "text", + "content": "[39] Fuxin Li, Taeyoung Kim, Ahmad Humayun, David Tsai, and James M Rehg. Video segmentation by tracking many figure-ground segments. In Proceedings of the IEEE international conference on computer vision, pages 2192-2199, 2013. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 554, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 160 + ], + "type": "text", + "content": "[40] Hongyang Li, Hao Zhang, Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, and Lei Zhang. Taptr: Tracking any point with transformers as detection. arXiv preprint arXiv:2403.13042, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 162, + 554, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 554, + 194 + ], + "type": "text", + "content": "[41] Xinghui Li, Jingyi Lu, Kai Han, and Victor Prisacariu. Sd4match: Learning to prompt stable diffusion model for semantic matching. In CVPR, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 195, + 554, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 554, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 554, + 249 + ], + "type": "text", + "content": "[42] Zhen Li, Zuo-Liang Zhu, Ling-Hao Han, Qibin Hou, Chun-Le Guo, and Ming-Ming Cheng. Amt: All-pairs multi-field transforms for efficient frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9801-9810, 2023. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 251, + 554, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 554, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 554, + 281 + ], + "type": "text", + "content": "[43] Ilya Loshchilov, Frank Hutter, et al. Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101, 5, 2017. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 283, + 554, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 554, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 554, + 336 + ], + "type": "text", + "content": "[44] Grace Luo, Lisa Dunlap, Dong Huk Park, Aleksander Holynski, and Trevor Darrell. Diffusion hyperfeatures: Searching through time and space for semantic correspondence. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 338, + 554, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 554, + 392 + ], + "type": "text", + "content": "[45] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 394, + 554, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 554, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 554, + 458 + ], + "type": "text", + "content": "[46] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 460, + 554, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 554, + 513 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 554, + 513 + ], + "type": "text", + "content": "[47] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 515, + 554, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 554, + 557 + ], + "type": "text", + "content": "[48] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 559, + 554, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 554, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 554, + 623 + ], + "type": "text", + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 625, + 554, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 554, + 657 + ], + "type": "text", + "content": "[50] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "type": "text", + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7285" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[52] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "text", + "content": "[53] Peter Sand and Seth Teller. Particle video: Long-range motion estimation using point trajectories. International journal of computer vision, 80:72-91, 2008. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 195 + ], + "type": "text", + "content": "[54] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "type": "text", + "content": "[55] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 295, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 295, + 285 + ], + "type": "text", + "content": "[56] Spencer Sterling. Zeroscope, 2023. https://huggingface.co/cerspense/zeroscope_v2_576w.3.7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 287, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 294, + 331 + ], + "type": "text", + "content": "[57] Shobhita Sundaram, Stephanie Fu, Lukas Mutenthaler, Netanel Y. Tamir, Lucy Chai, Simon Kornblith, Trevor Darrell, and Phillip Isola. When does perceptual alignment benefit vision representations? In NIPS, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "text", + "content": "[58] Saksham Suri, Matthew Walmer, Kamal Gupta, and Abhinav Shrivastava. Lift: A surprisingly simple lightweight feature transform for dense vit descriptors. In ECCV, pages 110-128. Springer, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 378, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 378, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 378, + 294, + 421 + ], + "type": "text", + "content": "[59] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. Advances in Neural Information Processing Systems, 36:1363-1389, 2023. 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 295, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 475 + ], + "type": "text", + "content": "[60] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "type": "text", + "content": "[61] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "text", + "content": "[62] Narek Tumanyan, Assaf Singer, Shai Bagon, and Tali Dekel. Dino-tracker: Taming dino for self-supervised point tracking in a single video. arXiv preprint arXiv:2403.14548, 2024. 2, 3, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 580, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 294, + 624 + ], + "type": "text", + "content": "[63] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 656 + ], + "type": "text", + "content": "[64] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural computation, 23(7):1661-1674, 2011. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[65] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19795-19806, 2023. 2, 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[66] Qian Wang, Abdelrahman Eldosokey, Mohit Mendiratta, Fangneng Zhan, Adam Kortylewski, Christian Theobalt, and Peter Wonka. Zero-shot video semantic segmentation based on pre-trained diffusion models, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "type": "text", + "content": "[67] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. arXiv preprint arXiv:2309.15103, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "text", + "content": "[68] Jay Zhangjie Wu, Guian Fang, Haoning Wu, Xintao Wang, Yixiao Ge, Xiaodong Cun, David Junhao Zhang, Jia-Wei Liu, Yuchao Gu, Rui Zhao, et al. Towards a better metric for text-to-video generation. arXiv preprint arXiv:2401.07781, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 230, + 553, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 283 + ], + "type": "text", + "content": "[69] Weilai Xiang, Hongyu Yang, Di Huang, and Yunhong Wang. Denoising diffusion autoencoders are unified self-supervised learners. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15802-15812, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "type": "text", + "content": "[70] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20406-20417, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "type": "text", + "content": "[71] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. ODISE: Open-Vocabulary Panoptic Segmentation with Text-to-Image Diffusion Models. In CVPR, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 388, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 419 + ], + "type": "text", + "content": "[72] Jiale Yang, So Kanazawa, Masami K Yamaguchi, and Isamu Motoyoshi. Pre-constancy vision in infants. Current Biology, 25(24):3209-3212, 2015. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 422, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 464 + ], + "type": "text", + "content": "[73] Xingyi Yang and Xinchao Wang. Diffusion model as representation learner. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18938-18949, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 468, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 553, + 521 + ], + "type": "text", + "content": "[74] Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is easier than you think. arXiv preprint arXiv:2410.06940, 2024. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "text", + "content": "[75] Yuanwen Yue, Anurag Das, Francis Engelmann, Siyu Tang, and Jan Eric Lenssen. Improving 2D Feature Representations by 3D-Aware Fine-Tuning. In ECCV, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 557, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 610 + ], + "type": "text", + "content": "[76] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 613, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 657 + ], + "type": "text", + "content": "[77] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[78] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. arXiv preprint arXiv:2311.04145, 2023. 2, 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7286" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 172 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[79] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19855-19865, 2023. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "text", + "content": "[80] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 2" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7287" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Tracktention_ Leveraging Point Tracking to Attend Videos Faster and Better/55c83ffe-0183-42a7-a82c-e5f5408552e5_content_list.json b/2025/Tracktention_ Leveraging Point Tracking to Attend Videos Faster and Better/55c83ffe-0183-42a7-a82c-e5f5408552e5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..424024cbdb93a7c422bea71697cebe9bd906a9be --- /dev/null +++ b/2025/Tracktention_ Leveraging Point Tracking to Attend Videos Faster and Better/55c83ffe-0183-42a7-a82c-e5f5408552e5_content_list.json @@ -0,0 +1,1666 @@ +[ + { + "type": "text", + "text": "Tracktention: Leveraging Point Tracking to Attend Videos Faster and Better", + "text_level": 1, + "bbox": [ + 112, + 130, + 883, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zihang Lai Andrea Vedaldi Visual Geometry Group (VGG), University of Oxford", + "bbox": [ + 285, + 180, + 710, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zlai,vedaldi}@robots.ox.ac.uk", + "bbox": [ + 362, + 219, + 627, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/61729fc2a75a3e95485e8cb505c4548e0081624ed3e140e55dbf2408538c8dab.jpg", + "image_caption": [ + "Figure 1. Left: The Tracktention Layer is a plug-and-play module that can convert an image-based network (e.g., for monocular depth prediction) into a state-of-the-art video network (e.g., for video depth prediction). It does so by integrating the output of any off-the-shelf, modern, and powerful point trackers via track cross-attention. Right: For example, Tracktention achieves state-of-the-art and efficient video depth prediction by transforming Depth Anything into a video depth model. See Tab. 2 for detailed results. *Single-image models." + ], + "image_footnote": [], + "bbox": [ + 98, + 268, + 189, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/86a95c21ef9b45cd361f1edd74c0707a424ef5eea5ebf14f86c8f43a42c4ba0a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 191, + 268, + 617, + 411 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/07323bda3f0f1e56205a472b8a38445d51bf2a43b7ac149e4ab40573663087f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 271, + 895, + 412 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 508, + 326, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Temporal consistency is critical in video prediction to ensure that outputs are coherent and free of artifacts. Traditional methods, such as temporal attention and 3D convolution, may struggle with significant object motion and may not capture long-range temporal dependencies in dynamic scenes. To address this gap, we propose the Tracktention Layer, a novel architectural component that explicitly integrates motion information using point tracks, i.e., sequences of corresponding points across frames. By incorporating these motion cues, the Tracktention Layer enhances temporal alignment and effectively handles complex object motions, maintaining consistent feature representations over time. Our approach is computationally efficient and can be seamlessly integrated into existing models, such as Vision Transformers, with minimal modification. It can be used to upgrade image-only models to state-of-the-art video ones, sometimes outperforming models natively designed for video prediction. We demonstrate this on video depth prediction and video colorization, where models augmented with the Tracktention Layer exhibit significantly improved temporal consistency compared to baselines. Project website: zlai0.github.io/TrackTention.", + "bbox": [ + 86, + 551, + 485, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 508, + 643, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Compared to image analysis tasks, video analysis tasks, such as video segmentation, video depth estimation, and video colorization, pose additional challenges due to the temporal dimension inherent in video data. A particularly important challenge is ensuring that outputs are temporally consistent across frames, which is necessary to produce coherent and artifact-free results. Temporal inconsistencies, such as flickering or abrupt changes in predicted object attributes, are especially noticeable in applications like colorization, where the output is a new video.", + "bbox": [ + 511, + 535, + 906, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "An aspect of video analysis that has seen significant progress in recent years is point tracking. New trackers such as PIPs [19], TAPIR [12], BootsTAP [13], and CoTracker [26] can now track quasi-dense collections of points across long video sequences with high reliability and efficiency. So far, these trackers have been primarily used in selected applications like 3D reconstruction and object tracking through ad-hoc algorithms. In this paper, we ask whether progress in point tracking can benefit a much wider range of video analysis tasks.", + "bbox": [ + 511, + 688, + 908, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To answer this question, we propose the Tracktention Layer, a novel architectural component designed to enhance temporal consistency in video transformers. Tracktention integrates seamlessly into vision transformers as a new at", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22809", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1df24ea7727c5d35d33c2a108f07d3ff05cd5ed41f342ba90eb524ebb2d46ce8.jpg", + "image_caption": [ + "Figure 2. Overview of Tracktention. We begin by using an off-the-shelf point tracker to extract a number of video tracks. Given these, we first sample image tokens at the track locations, obtaining corresponding track tokens (Sec. 3.1). Next, we use a Track Transformer to update these tokens, propagating information temporally at consistent spatial locations (Sec. 3.2). Finally, we splat the information back to the image tokens (Sec. 3.3). By explicitly incorporating motion information through point tracks, Tracktention improves temporal alignment, effectively captures complex object movements, and ensures stable feature representations over time." + ], + "image_footnote": [], + "bbox": [ + 133, + 85, + 861, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tention layer, making it compatible with a wide range of existing models. The key innovation of Tracktention lies in leveraging pre-extracted point tracks to inform the model about temporal correspondences between image tokens in different frames (Fig. 2). This is achieved by sampling the image tokens at the track points. The information is then propagated along the tracks using a transformer layer, which justifies the name \"Tracktention\". Finally, the information is splatted back to the image tokens, allowing the computation to progress as normal.", + "bbox": [ + 88, + 325, + 480, + 477 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By explicitly incorporating tracking information, the model can establish direct correspondences, enabling precise temporal alignment regardless of how much objects move. This explicit temporal alignment allows the model to maintain consistent feature representations over time, smoothing features, and reducing temporal artifacts.", + "bbox": [ + 89, + 477, + 482, + 568 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The problem of propagating information in video neural networks has been extensively studied in the literature. Many approaches are implicit, in the sense that they do not explicitly estimate or account for the motion observed in the scene. Examples include 3D convolution [6, 53] and space-time attention [1, 3, 14, 40, 58]. However, to maintain manageable computational costs, these methods often reduce the spatial resolution of features and limit the temporal range they can model—either by using small convolution kernels or by decoupling spatial and temporal attention. The resulting networks may struggle to represent motion precisely and to keep up with large object displacements.", + "bbox": [ + 89, + 568, + 482, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, Tracktention explicitly accounts for motion at the resolution of point tracks, which is much finer than that of typical image features. By pooling and splatting information at image locations that are already put in correspondence by the tracker, it only needs to propagate information along the tracks, obviating the need to implicitly establish such correspondences via space-time attention.", + "bbox": [ + 89, + 750, + 482, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There are also approaches to video analysis that, like Tracktention, incorporate motion information explicitly. They often work by feeding optical flow to the network,", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "capturing pixel-level motion between frames [42]. While optical flow can be effective in certain contexts, it may struggle with occlusions or large displacements. Building on the power of point trackers, Tracktention can handle occlusions, large displacements, and long temporal dependencies more effectively.", + "bbox": [ + 511, + 325, + 903, + 416 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, Tracktention offers several advantages over traditional video network architectures: (1) It leverages the power of modern point tracker models, which are excellent motion experts, and provides a general mechanism to inject this knowledge into any video transformer. In particular, it can handle complex and large object motions, limited only by the capabilities of the point trackers. (2) It directly establishes long-term space-time correspondences, unlike 3D convolutions that are local and unlike separated space-time attention, which can only do so indirectly. (3) It avoids the need to repeatedly calculate such correspondences explicitly, as full spatio-temporal attention does, achieving much greater efficiency and operating at a much finer resolution.", + "bbox": [ + 511, + 417, + 905, + 616 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We deliver Tracktention as a plugin layer (Fig. 1) that can be added to single-image neural network architectures to extend them to powerful, temporally-aware video models. For example, we demonstrate that a pre-trained single-image depth predictor can be extended to a corresponding state-of-the-art video depth predictor by integrating our Tracktention Layer. We achieve similar results for automatic video colorization, outperforming models natively designed to process videos while also being more efficient.", + "bbox": [ + 511, + 614, + 905, + 752 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 768, + 656, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Temporal Modeling for Videos. Temporal modeling is a cornerstone of video understanding tasks. Over the years, various techniques have been proposed to capture temporal dependencies in video data. 3D CNNs [6, 16, 53] extend 2D convolutions into the temporal dimension, allowing for direct spatiotemporal feature extraction from video data. Temporal attention mechanisms provide a new approach to", + "bbox": [ + 511, + 794, + 905, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22810", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "effectively capture long-range dependencies across frames. Video transformers [1, 3, 15, 37, 42] apply self-attention mechanisms over temporal and spatial dimensions, typically separately or hierarchically, to reduce computational complexity while modeling temporal dependencies effectively. Recent video processing models [4, 5, 50] also use a hybrid of these approaches. However, these methods have notable limitations: temporal attention struggles with rapid object movements, while 3D convolutions assume local spatiotemporal correlations, failing with unpredictable motion. In this work, we explicitly leverage tracking information to precisely align temporal features regardless of object movement.", + "bbox": [ + 89, + 90, + 483, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consistent Video Prediction. Addressing temporal inconsistency is critical for ensuring smooth and coherent video outputs. Optical flow-based methods [69] enforce temporal coherence by aligning features or predictions across frames, propagating features along motion paths estimated by optical flow. Liu et al. [35] utilize convolutional LSTMs to capture temporal information for consistent video frame synthesis, modeling temporal dependencies through recurrent connections. Lai et al. [30] propose a post-processing model with a temporal consistency objective for video prediction. Similarly, Luo et al. [38] and Kopf et al. [29] present test-time training post-processing methods specifically designed to stabilize video depth predictions using geometric consistency cues. In contrast, our work aims to stabilize features by allowing them to attend to corresponding areas across time based on point tracks. This approach inherently accounts for object movement and embeds temporal alignment efficiently within the model.", + "bbox": [ + 91, + 291, + 483, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point Tracking. Point tracking is the task of identifying and continuously tracking specific points across frames in a video. PIPs [18] pioneered a neural network-based approach that updates tracked points iteratively by extracting correlation maps between frames, inspired by the RAFT model for optical flow [52]. TAP-Vid [10] proposed a comprehensive benchmark and TAP-Net, a new model for point tracking. TAPIR [11] further improved performance by combining global matching capabilities with PIPs, enhancing the accuracy of tracked points even in complex motion scenarios. OmniMotion [55] tracks points by leveraging a quasi-3D canonical volume but is approximately $10^{3}$ times slower than CoTracker due to its test-time optimization. CoTracker [25] leveraged a transformer-based architecture to track multiple points jointly, yielding improvements particularly for occluded points. Recent methods such as LocoTrack [7], which introduced 4D correlation features, and TAPTR [32] further improved precision.", + "bbox": [ + 89, + 566, + 483, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we do not propose new models for point tracking. Instead, we use existing trackers to establish direct token correspondences, ensuring precise temporal alignment despite object motion.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 89, + 604, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce Tracktention, a novel transformer layer designed to aggregate and redistribute feature information in video transformers in a motion-aware manner (Fig. 3). By leveraging motion information captured by the output of a point tracker, this layer improves the temporal consistency and motion awareness of video features.", + "bbox": [ + 511, + 114, + 903, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Tracktention comprises an Attentional Sampling block (Sec. 3.1) that pools information from the video features into the tracks, a Track Transformer (Sec. 3.2) that updates the track tokens, and an Attentional Splatting block (Sec. 3.3) that pushes the updated track tokens back to the video features. We describe these components next.", + "bbox": [ + 511, + 205, + 903, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Attentional Sampling", + "text_level": 1, + "bbox": [ + 511, + 305, + 714, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Attentional Sampling module is the first stage in Track-attention. It uses cross-attention to pool information from the video features into the track tokens.", + "bbox": [ + 511, + 328, + 903, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $F\\in \\mathbb{R}^{T\\times HW\\times D_f}$ denote the input feature map, where $T$ is the number of time steps, $H$ and $W$ are the spatial dimensions, and $D_{f}$ is the feature dimension. Let $\\mathbf{P}\\in \\mathbb{R}^{T\\times M\\times 2}$ represent a set of $M$ tracks, each of which is a sequence of 2D points. We first apply a positional embedding to these 2D points, obtaining one token per track and per frame, and arrange these tokens into a tensor $\\mathcal{T}\\in \\mathbb{R}^{T\\times M\\times D_f}$ . We then project the track tokens into queries $Q\\in \\mathbb{R}^{T\\times M\\times D_k}$ and the feature map into keys $K\\in \\mathbb{R}^{T\\times HW\\times D_k}$ and values $V\\in \\mathbb{R}^{T\\times HW\\times D_f}$ as:", + "bbox": [ + 511, + 373, + 905, + 523 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {t} = \\mathcal {T} _ {t} W _ {Q}, \\quad K _ {t} = F _ {t} W _ {K}, \\quad V _ {t} = F _ {t},\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 534, + 846, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $W_{Q}$ and $W_{K}$ are learnable weight matrices, and the subscript $t \\in \\{1, \\dots, T\\}$ indexes the first dimensions of the tensors, extracting matrices. Note that we avoid projecting the values to preserve the original features during sampling.", + "bbox": [ + 511, + 559, + 905, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For each time $t$ , the attention weights $A_{t} \\in \\mathbb{R}^{M \\times HW}$ are computed using the scaled dot-product attention mechanism:", + "bbox": [ + 511, + 619, + 903, + 662 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nA _ {t} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {t} K _ {t} ^ {\\top}}{\\sqrt {D _ {k}}} + B _ {t}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 661, + 903, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $B_{t}\\in \\mathbb{R}^{M\\times HW}$ is a bias term that encourages attention to align with the tracks. It is defined as:", + "bbox": [ + 511, + 699, + 903, + 731 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nB _ {t i j} = - \\frac {\\| P _ {t i} - \\mathrm {p o s} _ {F _ {t}} (j) \\| ^ {2}}{2 \\sigma^ {2}},\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 739, + 808, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $i\\in \\{1,\\dots ,M\\}$ $j\\in \\{1,\\ldots ,HW\\}$ $P_{ti}\\in \\mathbb{R}^2$ is the spatial location of track $i$ at time $t$ and $\\mathrm{pos}_{F_t}(j)\\in \\mathbb{R}^2$ is the spatial location of the $j$ -th feature map token at time $t$", + "bbox": [ + 511, + 779, + 903, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the softmax operator of Eq. (1), this bias is analogous to multiplying the attention map by a Gaussian window that decays exponentially as the feature position deviates from the track position. The parameter $\\sigma$ , which we set to $1/2$ , determines how fast the window decays.", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22811", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9dd77097f4f8489e4bd92f770a6f3256a0bfae7209f1718b88d3509302743314.jpg", + "image_caption": [ + "(A) Tracktention Architecture" + ], + "image_footnote": [], + "bbox": [ + 106, + 111, + 666, + 330 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/059f4f5464bd5a27fc712040548c59b8177ff89f3b912a192b93f8157f26c20e.jpg", + "image_caption": [ + "Figure 3. Left: the Tracktention architecture comprises Attentional Sampling, pooling information from images to track, Track Transformer, processing this information temporally, and Attentional Splatting, moving the processed information back to the images. Right: Tracktention is easily integrated in ViTs and ConvNets to make video networks out image ones." + ], + "image_footnote": [], + "bbox": [ + 669, + 95, + 893, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use the attention to \"sample\" track features $S \\in \\mathbb{R}^{T \\times M \\times D_f}$ by setting $S_t = A_t V_t = A_t F_t \\in \\mathbb{R}^{M \\times D_f}$ for each time $t$ . Alternatively, we could have obtained $S$ by sampling the features $F$ at the given track locations, for instance, by using bilinear sampling. This can also be expressed as $A_t F_t$ for a specific matrix $A_t$ , so sampling is a special case of our Attentional Sampling block. The advantage of Attentional Sampling is that it allows the model to learn a better sampling strategy.", + "bbox": [ + 88, + 390, + 482, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Relative Spatial Information. To capture the relative spatial relationships between track tokens and feature map locations, we incorporate Rotational Position Encodings (RoPE) [51] into the keys and values derived from the feature map. RoPE is a form of positional encoding that allows attention mechanisms to be aware of the relative positions of points defined in a continuous space.", + "bbox": [ + 89, + 529, + 483, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To compute the RoPE embedding $\\mathrm{RoPE}(\\mathbf{f})$ of a feature vector $\\mathbf{f}$ at spatial location $(x,y)$ , we use the first $d / 2$ channels to encode the $x$ position. For $i = 1,\\dots ,d / 4$ , we define:", + "bbox": [ + 89, + 633, + 483, + 694 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{r} \\mathrm {R o P E} (\\mathbf {f}) [ 2 i ] = \\mathbf {f} [ 2 i ] \\cdot \\cos (\\theta_ {i} x) - \\mathbf {f} [ 2 i + 1 ] \\cdot \\sin (\\theta_ {i} x), \\\\ \\mathrm {R o P E} (\\mathbf {f}) [ 2 i + 1 ] = \\mathbf {f} [ 2 i ] \\cdot \\sin (\\theta_ {i} x) + \\mathbf {f} [ 2 i + 1 ] \\cdot \\cos (\\theta_ {i} x), \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 93, + 707, + 478, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta_{i} = 100^{\\frac{-2(i - 1)}{d / 2}}$ . The second $d / 2$ channels encode the $y$ position analogously by replacing $x$ with $y$ in the above equations.", + "bbox": [ + 89, + 756, + 482, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Implementation Details. To stabilize the training process, especially due to the difference in scale between the track token embeddings and the feature map embeddings (which tend to have larger norms), we apply QK-normalization [9] to the queries and keys before computing the attention weights. This normalization ensures that the scale of the", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "embeddings is consistent, preventing issues related to gradient instability. We also use $H$ attention heads, allowing the model to capture diverse contextual relationships across different subspaces.", + "bbox": [ + 511, + 390, + 906, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Track Transformer", + "text_level": 1, + "bbox": [ + 511, + 463, + 697, + 478 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Once the track tokens $S$ are obtained, the Track Transformer processes them to propagate information along the tracks, making them more temporally consistent and smoothing out any irregularities. It is designed as a transformer that operates along the time dimension of the point tracks and outputs updated track tokens $\\mathcal{T}'$ .", + "bbox": [ + 511, + 487, + 906, + 577 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In more detail, recall that the track tokens $S$ are of shape $T \\times M \\times D_f$ . We swap the first two dimensions to obtain a tensor of shape $M \\times T \\times D_f$ . Then, we apply a transformer block, interpreting the first dimension $M$ as the batch dimension. The effect is to apply attention along the temporal dimension, allowing the model to learn temporal dependencies along each track.", + "bbox": [ + 511, + 579, + 905, + 684 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This approach ensures that information is not exchanged across tracks, meaning each track's temporal sequence is processed independently. We considered a variant of the Track Transformer that included attention across tracks as well, but found that this approach was slightly worse and more computationally expensive. Our interpretation is that information is already exchanged spatially by the Vision Transformer paired with Tracktention, making additional exchange of information across locations redundant.", + "bbox": [ + 511, + 686, + 905, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Architecture. For each track, we utilize a 2-layer transformer encoder along the temporal dimension, structured as a standard transformer model. Each layer consists of Multi-Head Self-Attention, which allows each time step within a track to attend to all other time steps within that same track,", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22812", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and Position-Wise Feed-Forward Networks, which apply nonlinear transformations independently to each time step, enhancing the model's representational capacity. We also incorporate sinusoidal positional encodings to provide the model with a sense of temporal order. These encodings are added to the input features prior to the transformer layers, ensuring the model can distinguish between different time steps and understand the temporal structure of the sequence.", + "bbox": [ + 89, + 90, + 480, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Attentional Splatting", + "text_level": 1, + "bbox": [ + 89, + 220, + 290, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Attentional Splitting module maps the updated track tokens $\\mathcal{T}'$ back onto the feature maps $F$ . This is achieved by reversing the roles of the track tokens and the feature tokens in the various expressions in Sec. 3.1. This means the queries are derived from the grid coordinates of the output feature map, while the keys and values are generated from the track tokens. The same bias term $B_{t}$ is used as before (but now the matrix is transposed), as well as the same $QK$ -normalization and RoPE encodings, this time obtaining a matrix $A_{t}^{\\prime} \\in \\mathbb{R}^{HW \\times M}$ . The final output Tracktention $(F)$ is then computed as:", + "bbox": [ + 89, + 243, + 482, + 409 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\operatorname {T r a c k t e n t i o n} (F) \\right] _ {t} = W _ {\\text {o u t}} A _ {t} ^ {\\prime} \\left[ \\mathcal {T} ^ {\\prime} \\right] _ {t},\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 417, + 408, + 435 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $W_{\\mathrm{out}}$ is a final output projection. By adopting a symmetric design for Attentional Sampling and Splitting, we ensure that information is handled in a consistent manner.", + "bbox": [ + 89, + 444, + 482, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Point Tracker Pre-Processor", + "text_level": 1, + "bbox": [ + 89, + 498, + 344, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To apply the Tracktention layer to a video, we first need to extract the point tracks $\\mathbf{P}$ . We use CoTracker3 [24] due to its robustness in handling complex motions and occlusions in videos. Trackers require a seed point for each track, called a query. We sample 576 points uniformly at random from the spatio-temporal video volume and track each forward and backward in time. Compared to initializing tracks on a grid in the first frame of the video, this simple initialization scheme encourages tracks to cover the video well, despite camera and object motion.", + "bbox": [ + 89, + 521, + 482, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Integration into Image-Based Models", + "text_level": 1, + "bbox": [ + 89, + 681, + 413, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A key application of Tracktention is to build video neural networks from image-based ones. Integrating Tracktention into existing image-based models, such as a ViT or ConvNet, is straightforward. For example, monocular depth estimation models like Depth Anything, which are designed for single images, may lack temporal consistency when applied to video frame by frame. Adding Tracktention layers after each transformer block and fine-tuning the model on videos can increase temporal stability and significantly improve performance in video-based applications.", + "bbox": [ + 89, + 703, + 482, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To incorporate Tracktention into an existing network backbone, we insert it immediately after all or some transformer or convolutional blocks. We also include a residual", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "connection to preserve the original information flow, setting $F' = F + \\text{Tracktention}(F)$ , where $F$ represents the output feature map from the preceding block, and $F'$ is the updated feature map after the Tracktention module.", + "bbox": [ + 511, + 90, + 903, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Furthermore, the Tracktention output projection, $W_{\\mathrm{out}}$ , is initialized to zero, preserving the original network output at training onset. This strategy retains benefits of pre-trained weights while allowing the model to gradually adapt to temporal updates. Consequently, Tracktention modules capture fine spatial details and ensure consistent temporal modeling across layers.", + "bbox": [ + 511, + 151, + 906, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 270, + 645, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We show that Tracktention can turn image-based models into state-of-the-art ones for video depth estimation and colorization (Secs. 4.1 and 4.2) and ablate its design (Sec. 4.3).", + "bbox": [ + 511, + 295, + 905, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Video Depth Estimation", + "text_level": 1, + "bbox": [ + 511, + 349, + 733, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To apply Tracktention to video depth estimation, we start from an image-based model, Depth Anything [61], and upgrade it to a video-based model using the Tracktention module. The Depth Anything model has a DINO ViT backbone consisting of 12 transformer blocks. We insert our Tracktention module after each of the last 6 blocks.", + "bbox": [ + 511, + 373, + 905, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We fine-tune the model on a large dataset of videos, using a combination of synthetic and real data. This data comes with ground-truth depth, which we fit using a scale-and shift-invariant loss, as Depth Anything predicts depth up to an affine transformation. However, we share the same calibration parameters for all frames in a video, as calibrating frames independently would mask temporal inconsistencies, which we want the model to learn to correct.", + "bbox": [ + 511, + 463, + 905, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To reduce the number of new parameters added to the model, we share them between all Tracktention modules. During training, the original model is kept frozen, and only the Tracktention modules are updated. We use the AdamW optimizer with an initial learning rate of $1.6 \\times 10^{-5}$ and cosine learning rate decay. Training is conducted for 4 epochs with a batch size of 4 videos, each containing a randomly sampled number of frames between 8 and 16.", + "bbox": [ + 511, + 584, + 905, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data and evaluation metrics. For training, we use a combination of datasets containing both synthetic and real videos. We use standard depth estimation metrics [62]. Please refer to Appendix A.3.1 for details.", + "bbox": [ + 511, + 707, + 905, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.1 Quantitative Results", + "text_level": 1, + "bbox": [ + 511, + 786, + 710, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Tab. 1, we evaluate our model on four video collections in the DepthCrafter benchmark [20] against several alternatives. Tracktention improves the image-based model it augments, i.e., Depth Anything [62], substantially: Sintel AbsRel improves by $9.2\\%$ and $\\delta_{1}$ by $13.5\\%$ , KITTI by $26.8\\%$ and $12.5\\%$ , Bonn by $15.4\\%$ and $3.4\\%$ , and on average by", + "bbox": [ + 511, + 809, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22813", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6580a0dc45bfed4ac426b258ba8039bd1d10b42d8269ba8e0209685f82cfffd6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodType#ParamsSintel (~50 frames)Scannet (90 frames)KITTI (110 frames)Bonn (110 frames)Average
AbsRel ↓δ1.25 ↑AbsRel ↓δ1.25 ↑AbsRel ↓δ1.25 ↑AbsRel ↓δ1.25 ↑AbsRel ↓δ1.25 ↑
DUS3R [56]Vid.578M0.6280.3930.1940.6940.2920.4560.2500.5880.3410.533
NVDS [59]Vid.430M0.4080.4830.1870.6770.2530.5880.1670.7660.2540.629
ChronoDepth [49]Vid.1521M0.5870.4860.1590.7830.1670.7590.1000.9110.2530.735
DepthCrafter [20]Vid.1521M0.343†0.673†0.1250.8480.1100.8810.0750.9710.1630.843
Marigold [27]Im.865M0.5320.5150.1660.7690.1490.7960.0910.9310.2350.753
DepthAny. [62] (*)Im.343M0.3250.5640.1300.8380.1420.8030.0780.9390.1690.786
DepthAny.-V2 [63]Im.343M0.3670.5540.1350.8220.1400.8040.1060.9210.1870.775
OursVid.140M0.2950.6400.0870.9330.1040.9030.0660.9710.1380.862
", + "bbox": [ + 91, + 88, + 906, + 237 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ff572a2f012f6485f270a5a19208d689023cf72fbae16534646c360ff3202358.jpg", + "table_caption": [ + "Table 1. DepthCrafter [20] video depth benchmark. Our model upgrades DepthAnything (*) to a video depth predictor, outperforming all baselines (image- or video-based) while having the smallest parameter count (140M). We use DepthAnything-Base (97M) as the base model and still outperform the DepthAnything-Large (343M) model. $\\dagger$ Reproduced results confirmed by original authors." + ], + "table_footnote": [], + "table_body": "
MethodsGT PoseGT Intr.AlignKITTIScanNetETH3DDTUT&TAverageRuntime (s)
rel ↓δ1.03 ↑rel ↓δ1.03 ↑rel ↓δ1.03 ↑rel ↓δ1.03 ↑rel ↓δ1.03 ↑rel ↓δ1.03 ↑
COLMAP [45, 46]#✓✓×12.058.214.634.216.455.10.796.52.795.09.367.8≈3 min
COLMAP Dense [45, 46]✓✓×26.952.738.022.589.823.220.869.325.776.440.248.8≈3 min
NVDS [54]××LSvld17.413.913.017.121.411.97.031.615.514.414.917.853.0
ChronoDepth [49]××LSvld15.314.814.714.521.111.37.031.014.316.814.517.77.0
DepthCrafter [20]××LSvld10.222.26.634.615.414.84.546.912.918.49.927.438.1
DepthAny. [62] (*)××LSvld8.428.15.244.313.918.63.755.110.425.38.334.30.1
DepthAny.V2 [63]××LSvld9.126.34.650.016.116.03.261.610.126.58.636.10.1
DUSt3R [56]××LSvld7.231.74.650.013.419.03.459.29.328.47.637.74.3
UniDepth [44]××LSvld(3.9)(58.2)(1.5)(90.1)12.718.97.030.710.623.27.144.20.5
Ours××LSvld6.932.44.550.89.227.13.262.47.332.96.241.10.9‡
", + "bbox": [ + 91, + 285, + 903, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. RobustMVD [48] multi-view depth benchmark, analogous to Tab. 1. Parentheses indicate that a model was trained on the same benchmark used for evaluation, which provides an advantage compared to training on different data. $\\ddagger$ Includes approximately 0.4 seconds for tracking. $^\\ddagger$ Metrics computed only for pixels with a valid prediction, which is an easier task.", + "bbox": [ + 89, + 452, + 906, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "18.3% and 9.7%. Second, Tracktention performs better than all other baselines, including DepthCrafter [20], which is a video depth model, by 15.3% and 2.3% on AbsRel and $\\delta_{1.25}$ , respectively. This demonstrates that Tracktention can turn an image-based model into a video-based one, outperforming models designed for video data from the start.", + "bbox": [ + 89, + 506, + 480, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 reports results on the RobustMVD benchmark [48], which features shorter video clips with potentially larger camera motions. Our model achieved the lowest average relative error of 6.2 and the highest average inlier ratio of $41.1\\%$ , surpassing all baselines. In particular, it improves the base model Depth Anything by $33.8\\%$ on ETH3D and by $29.8\\%$ on Tanks and Temples.", + "bbox": [ + 89, + 597, + 482, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our model is also efficient, utilizing only 140 million parameters — only 17.1 million for the Tracktention module. It operates with a runtime of only 0.9 seconds per sequence, of which 0.4 seconds are attributed to the tracker. Despite its smaller size, our model outperforms DepthCrafter [20], which relies on video diffusion models with over 1.5 billion parameters. Tracktention will further benefit from future accuracy and speed improvements in point trackers.", + "bbox": [ + 89, + 703, + 482, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.2 Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 845, + 277, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 4 shows a qualitative comparison of depth estimation by our model, DepthCrafter, and DUSt3R in dynamic", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "scenes, complex textures, and cluttered environments. Our model produces stable, coherent depth maps across all frames, demonstrating strong temporal consistency. In contrast, DepthCrafter shows significant depth estimation errors in certain regions (blue box), especially in scenes with complex structures. DUSt3R, relying on implicit triangulation, struggles with dynamic content, resulting in inconsistent depth maps across time.", + "bbox": [ + 511, + 506, + 906, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 6 illustrates the Attentional Sampling block, which pools information along the track by attending to image features in correspondence with the track locations.", + "bbox": [ + 511, + 628, + 906, + 675 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Automatic Video Colorization", + "text_level": 1, + "bbox": [ + 511, + 691, + 779, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The goal of video colorization is to add colors to an input grayscale video. Colors should be realistic, vibrant, and consistent across frames to avoid flickering.", + "bbox": [ + 511, + 715, + 905, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We enhance four models — CIC [66], IDC [67], Colorformer [21], and DDColor [22] — by incorporating our Tracktention module into layers of each model's architecture. For instance, for DDColor, which has a ConvNet-based architecture with a ConvNeXt backbone [36] and a U-Net decoder, we insert our Tracktention module before each decoder block. We fine-tune the model using AdamW on the YouTube-VIS training set [60]. For testing, we use the validation sets of both DAVIS [43] and Video [30].", + "bbox": [ + 511, + 763, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22814", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/06c1fb5114cc00995fd1eafdc60c0f23969fe68ad34da647e15834e518a2881d.jpg", + "image_caption": [ + "Figure 4. Video depth prediction, comparing Tracktention (+DepthAnything), DepthCrafter [20], and DUS3R [56]. We visualize a column of pixels (highlighted in red) over time to illustrate temporal variation. Our model shows stable, coherent depth estimation over time, while DepthCrafter exhibits significant errors in certain regions (blue box). DUS3R struggles with dynamic content (green box)." + ], + "image_footnote": [], + "bbox": [ + 106, + 90, + 885, + 338 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6b571058b25d6ece888a668a281c9178d228c77eec8273c173d0e3d8434ea67b.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison of video colorization methods. Tracktention (+DDColor) yields vibrant, realistic, and consistent colors. In contrast, TCVC appears less vibrant (21.7 vs. 29.5 Colorfulness), while DDColor lacks temporal consistency." + ], + "image_footnote": [], + "bbox": [ + 106, + 386, + 890, + 623 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation metrics. We evaluate colorization results using standard metrics. Please refer to Appendix A.3.2 for details.", + "bbox": [ + 89, + 664, + 482, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Results. Figure 5 reports video colorization results on DAVIS and Video. We apply Tracktention to several image colorization models, consistently improving their performance when applied to videos, particularly in color consistency. Specifically, CDC improves by $22.8\\%$ on DAVIS and $28.3\\%$ on Video for CIC [66], $41.3\\%$ and $28.8\\%$ for Colorformer [21], and $46.5\\%$ and $45.7\\%$ for DD-Color [22]. Tracktention also improves FID and Colorfulness, indicating overall improvements in visual quality. We also compare to native video colorization networks (top rows), often outperforming these as well.", + "bbox": [ + 89, + 699, + 482, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. In Fig. 5, our method delivers more vibrant, realistic, and consistent colors compared to TCVC", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and DDColor. TCVC's colors are less vibrant, with a colorfulness score of 21.7 compared to our 29.5, and DDColor exhibits inconsistencies across frames, with colors sometimes changing dramatically over time.", + "bbox": [ + 511, + 664, + 906, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 738, + 679, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, we measure the impact of various design decisions in Tracktention using the AbsRel metric on depth prediction in Sintel, ScanNet, and Tanks and Temples.", + "bbox": [ + 511, + 763, + 906, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4a assesses the impact of the number of tracks used in training: even a single track per video can lead to improvements, and the optimal number is around $24^{2}$ tracks. Table 4b assesses the number of tracks used at testing time, with similar results. Table 4c compares different tracking strategies: eschewing the tracker and simply setting tracks", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22815", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4bd26969699b6ba9b2aee43a62f5af61bb319443ed377b0f85c2ee534a46f474.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDAVIS (medium frame length)Video (long frame length)
FID↓CF↑CDC↓PSNR↑FID↓CF↑CDC↓PSNR↑
Specialized video colorization models:
FAVC [31]—18.554.2224.38—16.281.8824.81
TCVC [34]46.5121.703.7325.5039.5819.071.6425.43
CIC [66]44.2330.346.1823.1939.7829.193.5922.51
[66]+Ours43.1334.224.77 (-22.8%)21.5837.0032.732.57 (-28.3%)22.64
IDC [67]42.9921.705.0125.4237.2519.072.5725.35
[67]+Ours33.4728.104.45 (-11.1%)23.5133.1325.222.42 (-6.0%)24.04
ColorFormer [21]40.7129.618.2923.0340.1229.764.9623.08
[21]+Ours29.7332.244.86 (-41.3%)23.1827.2529.653.53 (-28.8%)23.31
DDColor [22]26.8130.618.6423.8120.2730.385.6024.34
[22]+Ours24.6129.534.62 (-46.5%)23.8522.7826.063.04 (-45.7%)24.39
", + "bbox": [ + 93, + 88, + 624, + 291 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fdfa61bbfe90d7b9c917455b2e51420664a04b0c5ed17b5e91d70e166700f2fa.jpg", + "table_caption": [ + "Table 3. Quantitative comparison of video colorization methods on the DAVIS and Videvo datasets. Our method, when augmented onto four different baseline models, consistently improves the Color Distribution Consistency (CDC) metric across both datasets." + ], + "table_footnote": [], + "table_body": "
#tracksSintel↓ScanNet↓T&T↓
0 (base)0.3250.1300.104
10.3470.0940.087
12²0.3190.0860.080
24²0.2950.0870.073
", + "bbox": [ + 99, + 340, + 354, + 417 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9ac69fa1aa79a05762be34a5fa4a6bec6fb701b4b5c4ae38416703c4e9579550.jpg", + "table_caption": [ + "(a) Number of training tracks: Increasing training tracks consistently improved depth estimation accuracy, with optimal results at $24^{2}$ tracks." + ], + "table_footnote": [], + "table_body": "
MethodSintel↓ScanNet↓T&T↓
3D Conv.0.3640.1120.096
Time Attn.0.3630.0970.081
Tracktention0.2950.0870.073
", + "bbox": [ + 99, + 459, + 354, + 523 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/dc2e3363cf942e6b31b1cb3b267c5e68743989a7908f073a17507feeb865ff07.jpg", + "table_caption": [ + "(d) Temporal modeling: Tracktention outperformed 3D conv. and time attention by leveraging tracking data for better temporal alignment." + ], + "table_footnote": [], + "table_body": "
#tracksSintel↓ScanNet↓T&T↓
0 (base)0.3250.1300.104
10.3180.1180.108
12²0.3210.0870.075
24²0.2950.0870.073
", + "bbox": [ + 374, + 340, + 630, + 417 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/930356aadebf9387afee63d8bd2494d2cbd536b8b9bf2f9a79adaa82b570e4bf.jpg", + "table_caption": [ + "(b) Number of testing tracks: Increasing testing tracks significantly boosted performance, indicating better temporal capture with more tracks." + ], + "table_footnote": [], + "table_body": "
#LayersSintel↓ScanNet↓T&T↓
10.3160.0890.074
20.2950.0870.073
30.3170.0830.073
", + "bbox": [ + 374, + 459, + 630, + 523 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f01fabddce6f52c7beefcc580b00dd92b0e88ea4dafb4bf1ef3f689e8fcf30c3.jpg", + "image_caption": [ + "Figure 6. Attentional Sampling module. The module attends to image features in correspondence with each track, as shown by the attention maps next to each frame." + ], + "image_footnote": [], + "bbox": [ + 638, + 90, + 903, + 267 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/517e742601b64cfa75a4ccc14ee60fe5db226789d32ebb9d4a05456baa443907.jpg", + "table_caption": [ + "(e) Number of Time Attention Layers: Two layers balanced performance and complexity, while additional layers offered no consistent benefits." + ], + "table_footnote": [], + "table_body": "
TypeSintel↓ScanNet↓T&T↓
Constant0.3520.0990.085
Grid, T00.3250.0960.073
Rand., T00.3200.0950.073
Rand., all0.2950.0870.073
", + "bbox": [ + 648, + 340, + 903, + 417 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f7b9c09a570006d83c3244d035f9c11bb3a2b8e0c5d1eb0c4b96a89040ac2a70.jpg", + "table_caption": [ + "(c) Type of tracks: Using a point tracker performs better than using static pseudo-tracks. Furthermore, querying tracks uniformly is better." + ], + "table_footnote": [], + "table_body": "
#LayersSintel↓ScanNet↓T&T↓
10.3440.0900.087
60.2950.0870.073
120.3330.1050.075
", + "bbox": [ + 648, + 459, + 903, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(f) Number of Tracktention layers: 6 layers improved temporal representation, but further layers led to diminishing returns due to overfitting risks.", + "bbox": [ + 640, + 525, + 903, + 563 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablations on video depth estimation results. Each table examines the impact of different factors, such as the number and type of tracks used during training and testing. Results are reported using the AbsRel error metric, where $\\downarrow$ indicates that lower values are better.", + "bbox": [ + 89, + 566, + 906, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to be constant across all frames, and initializing the tracks on a grid in the first frame, at random in the first frame, or at random throughout the video. The latter results in much better performance, showing the importance of tracking and ensuring that tracks cover the video well. Table 4d tests replacing Tracktention Layers with 3D convolutions and time attention, which, by comparison, barely improve results. Table 4e shows the impact of increasing the number of transformer layers in the Track Transformer, where two layers are optimal in most cases. Finally, Tab. 4f shows the impact of adding a different number of Tracktention layers, concluding that six is about optimal.", + "bbox": [ + 88, + 603, + 485, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Limitations", + "text_level": 1, + "bbox": [ + 89, + 800, + 217, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tracktention's success depends on the quality of the extracted tracks. While modern point trackers are robust and efficient, they may still fail, particularly when tracks are off-camera for extended periods. Furthermore, running a point tracker in addition to the primary neural network adds com", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "plexity and computational cost (although this is often offset by avoiding spatio-temporal attention).", + "bbox": [ + 511, + 603, + 906, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 655, + 633, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have introduced Tracktention, a plug-and-play layer that upgrades image predictor networks into video predictor ones. It leverages the power of modern point trackers, transferring their understanding of motion to new applications such as depth prediction and colorization. The module is lightweight, simple to integrate, and compatible with a variety of architectures, including Vision Transformers and ConvNets. Thanks to the efficiency of recent point trackers like CoTracker3, it incurs only a small computational overhead. Remarkably, the resulting upgraded image models are overall faster and more accurate than some state-of-the-art methods specifically designed for video prediction.", + "bbox": [ + 509, + 683, + 906, + 866 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This research was supported by ERC 101001212-UNION.", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22816", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6836-6846, 2021. 2, 3", + "[2] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, et al. Arkitsscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. arXiv preprint arXiv:2111.08897, 2021. 14", + "[3] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 2, 3", + "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3", + "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3", + "[6] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 2", + "[7] Seokju Cho, Jiahui Huang, Jisu Nam, Honggyu An, Seun-gryong Kim, and Joon-Young Lee. Local all-pair correspondence for point tracking. arXiv preprint arXiv:2407.15420, 2024.3", + "[8] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 14", + "[9] Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Peter Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, et al. Scaling vision transformers to 22 billion parameters. In International Conference on Machine Learning, pages 7480-7512. PMLR, 2023. 4", + "[10] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610–13626, 2022. 3", + "[11] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10061-10072, 2023. 3" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. TAPIR: tracking any point with per-frame initialization and temporal refinement. In Proc. CVPR, 2023. 1", + "[13] Carl Doersch, Yi Yang, Dilara Gokay, Pauline Luc, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ross Goroshin, João Carreira, and Andrew Zisserman. BootsTAP: Bootstrapped training for tracking-any-point. arXiv, 2402.00847, 2024. 1", + "[14] H Fan, B Xiong, K Mangalam, Y Li, Z Yan, and J Malik.... Multiscale vision transformers. In Proc. CVPR, 2021. 2", + "[15] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6824-6835, 2021. 3", + "[16] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 2", + "[17] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 14", + "[18] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In European Conference on Computer Vision, pages 59-75. Springer, 2022. 3", + "[19] Adam W. Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle videos revisited: Tracking through occlusions using point trajectories. In Proc. ECCV, 2022. 1", + "[20] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024. 5, 6, 7, 14", + "[21] Xiaozhong Ji, Boyuan Jiang, Donghao Luo, Guangpin Tao, Wenqing Chu, Zhifeng Xie, Chengjie Wang, and Ying Tai. Colorformer: Image colorization via color memory assisted hybrid-attention transformer. In European Conference on Computer Vision, pages 20–36. Springer, 2022. 6, 7, 8, 15", + "[22] Xiaoyang Kang, Tao Yang, Wenqi Ouyang, Peiran Ren, Lingzhi Li, and Xuansong Xie. Ddcolor: Towards photorealistic image colorization via dual decoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 328-338, 2023. 6, 7, 8, 12, 15", + "[23] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Dynamic stereo: Consistent dynamic depth from stereo videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13229-13239, 2023. 14", + "[24] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker3: Simpler and better point tracking by pseudolabelling real videos. arxiv, 2024. 5, 13", + "[25] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22817", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tracker: It is better to track together. In Proc. ECCV, 2024. 3", + "[26] Nikita Karaev, Ignacio Rocco, Ben Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker: It is better to track together. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 1", + "[27] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 6", + "[28] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 14", + "[29] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. 3, 14", + "[30] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In Proceedings of the European conference on computer vision (ECCV), pages 170-185, 2018. 3, 6", + "[31] Chenyang Lei and Qifeng Chen. Fully automatic video colorization with self-regularization and diversity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3753–3761, 2019. 8, 15", + "[32] Hongyang Li, Hao Zhang, Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, and Lei Zhang. Taptr: Tracking any point with transformers as detection. In European Conference on Computer Vision, pages 57-75. Springer, 2025. 3", + "[33] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. 14", + "[34] Yihao Liu, Hengyuan Zhao, Kelvin CK Chan, Xintao Wang, Chen Change Loy, Yu Qiao, and Chao Dong. Temporally consistent video colorization with deep feature propagation and self-regularization learning. arXiv preprint arXiv:2110.04562, 2021.8, 15", + "[35] Ziwei Liu, Raymond A Yeh, Xiaou Tang, Yiming Liu, and Aseem Agarwala. Video frame synthesis using deep voxel flow. In Proceedings of the IEEE international conference on computer vision, pages 4463-4471, 2017. 3", + "[36] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 6, 15", + "[37] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3202-3211, 2022. 3" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4):71-1, 2020. 3, 14", + "[39] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4040-4048, 2016. 14", + "[40] Daniel Neimark, Omri Bar, Maya Zohar, and Dotan Asselmann. Video transformer network. In Proc. ICCV Workshops, 2021. 2", + "[41] E. Palazzolo, J. Behley, P. Lottes, P. Giguère, and C. Stachniss. ReFusion: 3D Reconstruction in Dynamic Environments for RGB-D Cameras Exploiting Residuals. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2019. 14", + "[42] Mandela Patrick, Dylan Campbell, Yuki Markus Asano, Ishan Misra, Florian Metze, Christoph Feichtenhofer, Andrea Vedaldi, and João F. Henriques. Keeping your eye on the ball: Trajectory attention in video transformers. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2021. 2, 3", + "[43] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 6", + "[44] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. UniDepth: universal monocular metric depth estimation. In Proc. CVPR, 2024. 6", + "[45] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proc. CVPR, 2016. 6", + "[46] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In Proc. ECCV, 2016. 6", + "[47] Thomas Schops, Johannes L. Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 14", + "[48] Philipp Schröppel, Jan Bechtold, Artemij Amiranashvili, and Thomas Brox. A benchmark and a baseline for robust multiview depth estimation. In 2022 International Conference on 3D Vision (3DV), pages 637-645. IEEE, 2022. 6, 14", + "[49] Jiahao Shao, Yuanbo Yang, Hongyu Zhou, Youmin Zhang, Yujun Shen, Matteo Poggi, and Yiyi Liao. Learning temporally consistent video depth from video diffusion priors. arXiv, 2406.01493, 2024. 6, 14", + "[50] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.3" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22818", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Jianlin Su, Yu Lu, Shengfeng Pan, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding, 2021. 4", + "[52] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 3", + "[53] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. 2", + "[54] Liao Wang, Qiang Hu, Qihan He, Ziyu Wang, Jingyi Yu, Tinne Tuytelaars, Lan Xu, and Minye Wu. Neural residual radiance fields for streamably free-viewpoint videos. In Proc. CVPR, 2023. 6", + "[55] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19795-19806, 2023. 3", + "[56] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. DUS3R: Geometric 3D vision made easy. In Proc. CVPR, 2024. 6, 7", + "[57] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4909-4916. IEEE, 2020. 14", + "[58] Xiaolong Wang, Ross B. Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proc. CVPR, 2018. 2", + "[59] Yiran Wang, Min Shi, Jiaqi Li, Zihao Huang, Zhiguo Cao, Jianming Zhang, Ke Xian, and Guosheng Lin. Neural video depth stabilizer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9466-9476, 2023. 6, 14", + "[60] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 6, 15", + "[61] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 5", + "[62] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proc. CVPR, 2024. 5, 6, 12, 14", + "[63] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything V2. arXiv, 2406.09414, 2024. 6", + "[64] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 14" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 14", + "[66] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 649-666. Springer, 2016. 6, 7, 8, 15", + "[67] Richard Zhang, Jun-Yan Zhu, Phillip Isola, Xinyang Geng, Angela S Lin, Tianhe Yu, and Alexei A Efros. Real-time user-guided image colorization with learned deep priors. arXiv preprint arXiv:1705.02999, 2017. 6, 8, 15", + "[68] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19855-19865, 2023. 14", + "[69] Xizhou Zhu, Yuwen Xiong, Jifeng Dai, Lu Yuan, and Yichen Wei. Deep feature flow for video recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2349-2358, 2017. 3" + ], + "bbox": [ + 516, + 90, + 905, + 402 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22819", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Tracktention_ Leveraging Point Tracking to Attend Videos Faster and Better/55c83ffe-0183-42a7-a82c-e5f5408552e5_model.json b/2025/Tracktention_ Leveraging Point Tracking to Attend Videos Faster and Better/55c83ffe-0183-42a7-a82c-e5f5408552e5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a7161b8edc3019b480411842c0601597bd449d3f --- /dev/null +++ b/2025/Tracktention_ Leveraging Point Tracking to Attend Videos Faster and Better/55c83ffe-0183-42a7-a82c-e5f5408552e5_model.json @@ -0,0 +1,2433 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.131, + 0.885, + 0.153 + ], + "angle": 0, + "content": "Tracktention: Leveraging Point Tracking to Attend Videos Faster and Better" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.181, + 0.712, + 0.217 + ], + "angle": 0, + "content": "Zihang Lai Andrea Vedaldi Visual Geometry Group (VGG), University of Oxford" + }, + { + "type": "text", + "bbox": [ + 0.364, + 0.22, + 0.629, + 0.234 + ], + "angle": 0, + "content": "{zlai,vedaldi}@robots.ox.ac.uk" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.27, + 0.191, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.27, + 0.619, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.272, + 0.896, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.416, + 0.907, + 0.472 + ], + "angle": 0, + "content": "Figure 1. Left: The Tracktention Layer is a plug-and-play module that can convert an image-based network (e.g., for monocular depth prediction) into a state-of-the-art video network (e.g., for video depth prediction). It does so by integrating the output of any off-the-shelf, modern, and powerful point trackers via track cross-attention. Right: For example, Tracktention achieves state-of-the-art and efficient video depth prediction by transforming Depth Anything into a video depth model. See Tab. 2 for detailed results. *Single-image models." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.51, + 0.327, + 0.525 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.552, + 0.486, + 0.884 + ], + "angle": 0, + "content": "Temporal consistency is critical in video prediction to ensure that outputs are coherent and free of artifacts. Traditional methods, such as temporal attention and 3D convolution, may struggle with significant object motion and may not capture long-range temporal dependencies in dynamic scenes. To address this gap, we propose the Tracktention Layer, a novel architectural component that explicitly integrates motion information using point tracks, i.e., sequences of corresponding points across frames. By incorporating these motion cues, the Tracktention Layer enhances temporal alignment and effectively handles complex object motions, maintaining consistent feature representations over time. Our approach is computationally efficient and can be seamlessly integrated into existing models, such as Vision Transformers, with minimal modification. It can be used to upgrade image-only models to state-of-the-art video ones, sometimes outperforming models natively designed for video prediction. We demonstrate this on video depth prediction and video colorization, where models augmented with the Tracktention Layer exhibit significantly improved temporal consistency compared to baselines. Project website: zlai0.github.io/TrackTention." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.51, + 0.644, + 0.525 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.536, + 0.907, + 0.687 + ], + "angle": 0, + "content": "Compared to image analysis tasks, video analysis tasks, such as video segmentation, video depth estimation, and video colorization, pose additional challenges due to the temporal dimension inherent in video data. A particularly important challenge is ensuring that outputs are temporally consistent across frames, which is necessary to produce coherent and artifact-free results. Temporal inconsistencies, such as flickering or abrupt changes in predicted object attributes, are especially noticeable in applications like colorization, where the output is a new video." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.839 + ], + "angle": 0, + "content": "An aspect of video analysis that has seen significant progress in recent years is point tracking. New trackers such as PIPs [19], TAPIR [12], BootsTAP [13], and CoTracker [26] can now track quasi-dense collections of points across long video sequences with high reliability and efficiency. So far, these trackers have been primarily used in selected applications like 3D reconstruction and object tracking through ad-hoc algorithms. In this paper, we ask whether progress in point tracking can benefit a much wider range of video analysis tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To answer this question, we propose the Tracktention Layer, a novel architectural component designed to enhance temporal consistency in video transformers. Tracktention integrates seamlessly into vision transformers as a new at" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22809" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.135, + 0.086, + 0.862, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.243, + 0.908, + 0.314 + ], + "angle": 0, + "content": "Figure 2. Overview of Tracktention. We begin by using an off-the-shelf point tracker to extract a number of video tracks. Given these, we first sample image tokens at the track locations, obtaining corresponding track tokens (Sec. 3.1). Next, we use a Track Transformer to update these tokens, propagating information temporally at consistent spatial locations (Sec. 3.2). Finally, we splat the information back to the image tokens (Sec. 3.3). By explicitly incorporating motion information through point tracks, Tracktention improves temporal alignment, effectively captures complex object movements, and ensures stable feature representations over time." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.327, + 0.482, + 0.478 + ], + "angle": 0, + "content": "tention layer, making it compatible with a wide range of existing models. The key innovation of Tracktention lies in leveraging pre-extracted point tracks to inform the model about temporal correspondences between image tokens in different frames (Fig. 2). This is achieved by sampling the image tokens at the track points. The information is then propagated along the tracks using a transformer layer, which justifies the name \"Tracktention\". Finally, the information is splatted back to the image tokens, allowing the computation to progress as normal." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.483, + 0.569 + ], + "angle": 0, + "content": "By explicitly incorporating tracking information, the model can establish direct correspondences, enabling precise temporal alignment regardless of how much objects move. This explicit temporal alignment allows the model to maintain consistent feature representations over time, smoothing features, and reducing temporal artifacts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.75 + ], + "angle": 0, + "content": "The problem of propagating information in video neural networks has been extensively studied in the literature. Many approaches are implicit, in the sense that they do not explicitly estimate or account for the motion observed in the scene. Examples include 3D convolution [6, 53] and space-time attention [1, 3, 14, 40, 58]. However, to maintain manageable computational costs, these methods often reduce the spatial resolution of features and limit the temporal range they can model—either by using small convolution kernels or by decoupling spatial and temporal attention. The resulting networks may struggle to represent motion precisely and to keep up with large object displacements." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.856 + ], + "angle": 0, + "content": "In contrast, Tracktention explicitly accounts for motion at the resolution of point tracks, which is much finer than that of typical image features. By pooling and splatting information at image locations that are already put in correspondence by the tracker, it only needs to propagate information along the tracks, obviating the need to implicitly establish such correspondences via space-time attention." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "There are also approaches to video analysis that, like Tracktention, incorporate motion information explicitly. They often work by feeding optical flow to the network," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.327, + 0.905, + 0.417 + ], + "angle": 0, + "content": "capturing pixel-level motion between frames [42]. While optical flow can be effective in certain contexts, it may struggle with occlusions or large displacements. Building on the power of point trackers, Tracktention can handle occlusions, large displacements, and long temporal dependencies more effectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.906, + 0.617 + ], + "angle": 0, + "content": "In summary, Tracktention offers several advantages over traditional video network architectures: (1) It leverages the power of modern point tracker models, which are excellent motion experts, and provides a general mechanism to inject this knowledge into any video transformer. In particular, it can handle complex and large object motions, limited only by the capabilities of the point trackers. (2) It directly establishes long-term space-time correspondences, unlike 3D convolutions that are local and unlike separated space-time attention, which can only do so indirectly. (3) It avoids the need to repeatedly calculate such correspondences explicitly, as full spatio-temporal attention does, achieving much greater efficiency and operating at a much finer resolution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.616, + 0.906, + 0.753 + ], + "angle": 0, + "content": "We deliver Tracktention as a plugin layer (Fig. 1) that can be added to single-image neural network architectures to extend them to powerful, temporally-aware video models. For example, we demonstrate that a pre-trained single-image depth predictor can be extended to a corresponding state-of-the-art video depth predictor by integrating our Tracktention Layer. We achieve similar results for automatic video colorization, outperforming models natively designed to process videos while also being more efficient." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.769, + 0.657, + 0.785 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Temporal Modeling for Videos. Temporal modeling is a cornerstone of video understanding tasks. Over the years, various techniques have been proposed to capture temporal dependencies in video data. 3D CNNs [6, 16, 53] extend 2D convolutions into the temporal dimension, allowing for direct spatiotemporal feature extraction from video data. Temporal attention mechanisms provide a new approach to" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22810" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.288 + ], + "angle": 0, + "content": "effectively capture long-range dependencies across frames. Video transformers [1, 3, 15, 37, 42] apply self-attention mechanisms over temporal and spatial dimensions, typically separately or hierarchically, to reduce computational complexity while modeling temporal dependencies effectively. Recent video processing models [4, 5, 50] also use a hybrid of these approaches. However, these methods have notable limitations: temporal attention struggles with rapid object movements, while 3D convolutions assume local spatiotemporal correlations, failing with unpredictable motion. In this work, we explicitly leverage tracking information to precisely align temporal features regardless of object movement." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.292, + 0.484, + 0.565 + ], + "angle": 0, + "content": "Consistent Video Prediction. Addressing temporal inconsistency is critical for ensuring smooth and coherent video outputs. Optical flow-based methods [69] enforce temporal coherence by aligning features or predictions across frames, propagating features along motion paths estimated by optical flow. Liu et al. [35] utilize convolutional LSTMs to capture temporal information for consistent video frame synthesis, modeling temporal dependencies through recurrent connections. Lai et al. [30] propose a post-processing model with a temporal consistency objective for video prediction. Similarly, Luo et al. [38] and Kopf et al. [29] present test-time training post-processing methods specifically designed to stabilize video depth predictions using geometric consistency cues. In contrast, our work aims to stabilize features by allowing them to attend to corresponding areas across time based on point tracks. This approach inherently accounts for object movement and embeds temporal alignment efficiently within the model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.484, + 0.839 + ], + "angle": 0, + "content": "Point Tracking. Point tracking is the task of identifying and continuously tracking specific points across frames in a video. PIPs [18] pioneered a neural network-based approach that updates tracked points iteratively by extracting correlation maps between frames, inspired by the RAFT model for optical flow [52]. TAP-Vid [10] proposed a comprehensive benchmark and TAP-Net, a new model for point tracking. TAPIR [11] further improved performance by combining global matching capabilities with PIPs, enhancing the accuracy of tracked points even in complex motion scenarios. OmniMotion [55] tracks points by leveraging a quasi-3D canonical volume but is approximately \\(10^{3}\\) times slower than CoTracker due to its test-time optimization. CoTracker [25] leveraged a transformer-based architecture to track multiple points jointly, yielding improvements particularly for occluded points. Recent methods such as LocoTrack [7], which introduced 4D correlation features, and TAPTR [32] further improved precision." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In this work, we do not propose new models for point tracking. Instead, we use existing trackers to establish direct token correspondences, ensuring precise temporal alignment despite object motion." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.605, + 0.105 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.905, + 0.206 + ], + "angle": 0, + "content": "In this section, we introduce Tracktention, a novel transformer layer designed to aggregate and redistribute feature information in video transformers in a motion-aware manner (Fig. 3). By leveraging motion information captured by the output of a point tracker, this layer improves the temporal consistency and motion awareness of video features." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.207, + 0.905, + 0.298 + ], + "angle": 0, + "content": "Tracktention comprises an Attentional Sampling block (Sec. 3.1) that pools information from the video features into the tracks, a Track Transformer (Sec. 3.2) that updates the track tokens, and an Attentional Splatting block (Sec. 3.3) that pushes the updated track tokens back to the video features. We describe these components next." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.306, + 0.715, + 0.322 + ], + "angle": 0, + "content": "3.1. Attentional Sampling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.329, + 0.905, + 0.373 + ], + "angle": 0, + "content": "The Attentional Sampling module is the first stage in Track-attention. It uses cross-attention to pool information from the video features into the track tokens." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.374, + 0.906, + 0.525 + ], + "angle": 0, + "content": "Let \\(F\\in \\mathbb{R}^{T\\times HW\\times D_f}\\) denote the input feature map, where \\(T\\) is the number of time steps, \\(H\\) and \\(W\\) are the spatial dimensions, and \\(D_{f}\\) is the feature dimension. Let \\(\\mathbf{P}\\in \\mathbb{R}^{T\\times M\\times 2}\\) represent a set of \\(M\\) tracks, each of which is a sequence of 2D points. We first apply a positional embedding to these 2D points, obtaining one token per track and per frame, and arrange these tokens into a tensor \\(\\mathcal{T}\\in \\mathbb{R}^{T\\times M\\times D_f}\\). We then project the track tokens into queries \\(Q\\in \\mathbb{R}^{T\\times M\\times D_k}\\) and the feature map into keys \\(K\\in \\mathbb{R}^{T\\times HW\\times D_k}\\) and values \\(V\\in \\mathbb{R}^{T\\times HW\\times D_f}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.535, + 0.848, + 0.551 + ], + "angle": 0, + "content": "\\[\nQ _ {t} = \\mathcal {T} _ {t} W _ {Q}, \\quad K _ {t} = F _ {t} W _ {K}, \\quad V _ {t} = F _ {t},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.56, + 0.906, + 0.62 + ], + "angle": 0, + "content": "where \\( W_{Q} \\) and \\( W_{K} \\) are learnable weight matrices, and the subscript \\( t \\in \\{1, \\dots, T\\} \\) indexes the first dimensions of the tensors, extracting matrices. Note that we avoid projecting the values to preserve the original features during sampling." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.62, + 0.905, + 0.663 + ], + "angle": 0, + "content": "For each time \\(t\\), the attention weights \\(A_{t} \\in \\mathbb{R}^{M \\times HW}\\) are computed using the scaled dot-product attention mechanism:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.662, + 0.905, + 0.697 + ], + "angle": 0, + "content": "\\[\nA _ {t} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {t} K _ {t} ^ {\\top}}{\\sqrt {D _ {k}}} + B _ {t}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.7, + 0.905, + 0.732 + ], + "angle": 0, + "content": "where \\(B_{t}\\in \\mathbb{R}^{M\\times HW}\\) is a bias term that encourages attention to align with the tracks. It is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.74, + 0.81, + 0.772 + ], + "angle": 0, + "content": "\\[\nB _ {t i j} = - \\frac {\\| P _ {t i} - \\mathrm {p o s} _ {F _ {t}} (j) \\| ^ {2}}{2 \\sigma^ {2}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.905, + 0.826 + ], + "angle": 0, + "content": "where \\(i\\in \\{1,\\dots ,M\\}\\) \\(j\\in \\{1,\\ldots ,HW\\}\\) \\(P_{ti}\\in \\mathbb{R}^2\\) is the spatial location of track \\(i\\) at time \\(t\\) and \\(\\mathrm{pos}_{F_t}(j)\\in \\mathbb{R}^2\\) is the spatial location of the \\(j\\) -th feature map token at time \\(t\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In the softmax operator of Eq. (1), this bias is analogous to multiplying the attention map by a Gaussian window that decays exponentially as the feature position deviates from the track position. The parameter \\(\\sigma\\), which we set to \\(1/2\\), determines how fast the window decays." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22811" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.098, + 0.256, + 0.109 + ], + "angle": 0, + "content": "(A) Tracktention Architecture" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.112, + 0.667, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.096, + 0.895, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.335, + 0.905, + 0.378 + ], + "angle": 0, + "content": "Figure 3. Left: the Tracktention architecture comprises Attentional Sampling, pooling information from images to track, Track Transformer, processing this information temporally, and Attentional Splatting, moving the processed information back to the images. Right: Tracktention is easily integrated in ViTs and ConvNets to make video networks out image ones." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.391, + 0.483, + 0.527 + ], + "angle": 0, + "content": "We use the attention to \"sample\" track features \\( S \\in \\mathbb{R}^{T \\times M \\times D_f} \\) by setting \\( S_t = A_t V_t = A_t F_t \\in \\mathbb{R}^{M \\times D_f} \\) for each time \\( t \\). Alternatively, we could have obtained \\( S \\) by sampling the features \\( F \\) at the given track locations, for instance, by using bilinear sampling. This can also be expressed as \\( A_t F_t \\) for a specific matrix \\( A_t \\), so sampling is a special case of our Attentional Sampling block. The advantage of Attentional Sampling is that it allows the model to learn a better sampling strategy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.635 + ], + "angle": 0, + "content": "Relative Spatial Information. To capture the relative spatial relationships between track tokens and feature map locations, we incorporate Rotational Position Encodings (RoPE) [51] into the keys and values derived from the feature map. RoPE is a form of positional encoding that allows attention mechanisms to be aware of the relative positions of points defined in a continuous space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.635, + 0.484, + 0.695 + ], + "angle": 0, + "content": "To compute the RoPE embedding \\(\\mathrm{RoPE}(\\mathbf{f})\\) of a feature vector \\(\\mathbf{f}\\) at spatial location \\((x,y)\\), we use the first \\(d / 2\\) channels to encode the \\(x\\) position. For \\(i = 1,\\dots ,d / 4\\), we define:" + }, + { + "type": "equation", + "bbox": [ + 0.094, + 0.708, + 0.48, + 0.748 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{r} \\mathrm {R o P E} (\\mathbf {f}) [ 2 i ] = \\mathbf {f} [ 2 i ] \\cdot \\cos (\\theta_ {i} x) - \\mathbf {f} [ 2 i + 1 ] \\cdot \\sin (\\theta_ {i} x), \\\\ \\mathrm {R o P E} (\\mathbf {f}) [ 2 i + 1 ] = \\mathbf {f} [ 2 i ] \\cdot \\sin (\\theta_ {i} x) + \\mathbf {f} [ 2 i + 1 ] \\cdot \\cos (\\theta_ {i} x), \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.757, + 0.483, + 0.808 + ], + "angle": 0, + "content": "where \\(\\theta_{i} = 100^{\\frac{-2(i - 1)}{d / 2}}\\). The second \\(d / 2\\) channels encode the \\(y\\) position analogously by replacing \\(x\\) with \\(y\\) in the above equations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. To stabilize the training process, especially due to the difference in scale between the track token embeddings and the feature map embeddings (which tend to have larger norms), we apply QK-normalization [9] to the queries and keys before computing the attention weights. This normalization ensures that the scale of the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.391, + 0.907, + 0.452 + ], + "angle": 0, + "content": "embeddings is consistent, preventing issues related to gradient instability. We also use \\( H \\) attention heads, allowing the model to capture diverse contextual relationships across different subspaces." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.464, + 0.699, + 0.479 + ], + "angle": 0, + "content": "3.2. Track Transformer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.907, + 0.578 + ], + "angle": 0, + "content": "Once the track tokens \\( S \\) are obtained, the Track Transformer processes them to propagate information along the tracks, making them more temporally consistent and smoothing out any irregularities. It is designed as a transformer that operates along the time dimension of the point tracks and outputs updated track tokens \\( \\mathcal{T}' \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.906, + 0.685 + ], + "angle": 0, + "content": "In more detail, recall that the track tokens \\( S \\) are of shape \\( T \\times M \\times D_f \\). We swap the first two dimensions to obtain a tensor of shape \\( M \\times T \\times D_f \\). Then, we apply a transformer block, interpreting the first dimension \\( M \\) as the batch dimension. The effect is to apply attention along the temporal dimension, allowing the model to learn temporal dependencies along each track." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.906, + 0.822 + ], + "angle": 0, + "content": "This approach ensures that information is not exchanged across tracks, meaning each track's temporal sequence is processed independently. We considered a variant of the Track Transformer that included attention across tracks as well, but found that this approach was slightly worse and more computationally expensive. Our interpretation is that information is already exchanged spatially by the Vision Transformer paired with Tracktention, making additional exchange of information across locations redundant." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Architecture. For each track, we utilize a 2-layer transformer encoder along the temporal dimension, structured as a standard transformer model. Each layer consists of Multi-Head Self-Attention, which allows each time step within a track to attend to all other time steps within that same track," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22812" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.213 + ], + "angle": 0, + "content": "and Position-Wise Feed-Forward Networks, which apply nonlinear transformations independently to each time step, enhancing the model's representational capacity. We also incorporate sinusoidal positional encodings to provide the model with a sense of temporal order. These encodings are added to the input features prior to the transformer layers, ensuring the model can distinguish between different time steps and understand the temporal structure of the sequence." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.221, + 0.291, + 0.238 + ], + "angle": 0, + "content": "3.3. Attentional Splatting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.244, + 0.483, + 0.41 + ], + "angle": 0, + "content": "The Attentional Splitting module maps the updated track tokens \\(\\mathcal{T}'\\) back onto the feature maps \\(F\\). This is achieved by reversing the roles of the track tokens and the feature tokens in the various expressions in Sec. 3.1. This means the queries are derived from the grid coordinates of the output feature map, while the keys and values are generated from the track tokens. The same bias term \\(B_{t}\\) is used as before (but now the matrix is transposed), as well as the same \\(QK\\)-normalization and RoPE encodings, this time obtaining a matrix \\(A_{t}^{\\prime} \\in \\mathbb{R}^{HW \\times M}\\). The final output Tracktention \\((F)\\) is then computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.419, + 0.409, + 0.436 + ], + "angle": 0, + "content": "\\[\n\\left[ \\operatorname {T r a c k t e n t i o n} (F) \\right] _ {t} = W _ {\\text {o u t}} A _ {t} ^ {\\prime} \\left[ \\mathcal {T} ^ {\\prime} \\right] _ {t},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.483, + 0.492 + ], + "angle": 0, + "content": "where \\( W_{\\mathrm{out}} \\) is a final output projection. By adopting a symmetric design for Attentional Sampling and Splitting, we ensure that information is handled in a consistent manner." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.499, + 0.345, + 0.513 + ], + "angle": 0, + "content": "3.4. Point Tracker Pre-Processor" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.483, + 0.673 + ], + "angle": 0, + "content": "To apply the Tracktention layer to a video, we first need to extract the point tracks \\(\\mathbf{P}\\). We use CoTracker3 [24] due to its robustness in handling complex motions and occlusions in videos. Trackers require a seed point for each track, called a query. We sample 576 points uniformly at random from the spatio-temporal video volume and track each forward and backward in time. Compared to initializing tracks on a grid in the first frame of the video, this simple initialization scheme encourages tracks to cover the video well, despite camera and object motion." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.414, + 0.697 + ], + "angle": 0, + "content": "3.5. Integration into Image-Based Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.856 + ], + "angle": 0, + "content": "A key application of Tracktention is to build video neural networks from image-based ones. Integrating Tracktention into existing image-based models, such as a ViT or ConvNet, is straightforward. For example, monocular depth estimation models like Depth Anything, which are designed for single images, may lack temporal consistency when applied to video frame by frame. Adding Tracktention layers after each transformer block and fine-tuning the model on videos can increase temporal stability and significantly improve performance in video-based applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To incorporate Tracktention into an existing network backbone, we insert it immediately after all or some transformer or convolutional blocks. We also include a residual" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.151 + ], + "angle": 0, + "content": "connection to preserve the original information flow, setting \\( F' = F + \\text{Tracktention}(F) \\), where \\( F \\) represents the output feature map from the preceding block, and \\( F' \\) is the updated feature map after the Tracktention module." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.152, + 0.907, + 0.258 + ], + "angle": 0, + "content": "Furthermore, the Tracktention output projection, \\( W_{\\mathrm{out}} \\), is initialized to zero, preserving the original network output at training onset. This strategy retains benefits of pre-trained weights while allowing the model to gradually adapt to temporal updates. Consequently, Tracktention modules capture fine spatial details and ensure consistent temporal modeling across layers." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.271, + 0.646, + 0.287 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.296, + 0.906, + 0.342 + ], + "angle": 0, + "content": "We show that Tracktention can turn image-based models into state-of-the-art ones for video depth estimation and colorization (Secs. 4.1 and 4.2) and ablate its design (Sec. 4.3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.35, + 0.734, + 0.366 + ], + "angle": 0, + "content": "4.1. Video Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.374, + 0.906, + 0.463 + ], + "angle": 0, + "content": "To apply Tracktention to video depth estimation, we start from an image-based model, Depth Anything [61], and upgrade it to a video-based model using the Tracktention module. The Depth Anything model has a DINO ViT backbone consisting of 12 transformer blocks. We insert our Tracktention module after each of the last 6 blocks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.906, + 0.584 + ], + "angle": 0, + "content": "We fine-tune the model on a large dataset of videos, using a combination of synthetic and real data. This data comes with ground-truth depth, which we fit using a scale-and shift-invariant loss, as Depth Anything predicts depth up to an affine transformation. However, we share the same calibration parameters for all frames in a video, as calibrating frames independently would mask temporal inconsistencies, which we want the model to learn to correct." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.906, + 0.706 + ], + "angle": 0, + "content": "To reduce the number of new parameters added to the model, we share them between all Tracktention modules. During training, the original model is kept frozen, and only the Tracktention modules are updated. We use the AdamW optimizer with an initial learning rate of \\(1.6 \\times 10^{-5}\\) and cosine learning rate decay. Training is conducted for 4 epochs with a batch size of 4 videos, each containing a randomly sampled number of frames between 8 and 16." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.708, + 0.906, + 0.769 + ], + "angle": 0, + "content": "Data and evaluation metrics. For training, we use a combination of datasets containing both synthetic and real videos. We use standard depth estimation metrics [62]. Please refer to Appendix A.3.1 for details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.787, + 0.712, + 0.802 + ], + "angle": 0, + "content": "4.1.1 Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.902 + ], + "angle": 0, + "content": "In Tab. 1, we evaluate our model on four video collections in the DepthCrafter benchmark [20] against several alternatives. Tracktention improves the image-based model it augments, i.e., Depth Anything [62], substantially: Sintel AbsRel improves by \\(9.2\\%\\) and \\(\\delta_{1}\\) by \\(13.5\\%\\), KITTI by \\(26.8\\%\\) and \\(12.5\\%\\), Bonn by \\(15.4\\%\\) and \\(3.4\\%\\), and on average by" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22813" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.908, + 0.238 + ], + "angle": 0, + "content": "
MethodType#ParamsSintel (~50 frames)Scannet (90 frames)KITTI (110 frames)Bonn (110 frames)Average
AbsRel ↓δ1.25 ↑AbsRel ↓δ1.25 ↑AbsRel ↓δ1.25 ↑AbsRel ↓δ1.25 ↑AbsRel ↓δ1.25 ↑
DUS3R [56]Vid.578M0.6280.3930.1940.6940.2920.4560.2500.5880.3410.533
NVDS [59]Vid.430M0.4080.4830.1870.6770.2530.5880.1670.7660.2540.629
ChronoDepth [49]Vid.1521M0.5870.4860.1590.7830.1670.7590.1000.9110.2530.735
DepthCrafter [20]Vid.1521M0.343†0.673†0.1250.8480.1100.8810.0750.9710.1630.843
Marigold [27]Im.865M0.5320.5150.1660.7690.1490.7960.0910.9310.2350.753
DepthAny. [62] (*)Im.343M0.3250.5640.1300.8380.1420.8030.0780.9390.1690.786
DepthAny.-V2 [63]Im.343M0.3670.5540.1350.8220.1400.8040.1060.9210.1870.775
OursVid.140M0.2950.6400.0870.9330.1040.9030.0660.9710.1380.862
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.24, + 0.907, + 0.283 + ], + "angle": 0, + "content": "Table 1. DepthCrafter [20] video depth benchmark. Our model upgrades DepthAnything (*) to a video depth predictor, outperforming all baselines (image- or video-based) while having the smallest parameter count (140M). We use DepthAnything-Base (97M) as the base model and still outperform the DepthAnything-Large (343M) model. \\(\\dagger\\)Reproduced results confirmed by original authors." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.286, + 0.905, + 0.452 + ], + "angle": 0, + "content": "
MethodsGT PoseGT Intr.AlignKITTIScanNetETH3DDTUT&TAverageRuntime (s)
rel ↓δ1.03 ↑rel ↓δ1.03 ↑rel ↓δ1.03 ↑rel ↓δ1.03 ↑rel ↓δ1.03 ↑rel ↓δ1.03 ↑
COLMAP [45, 46]#✓✓×12.058.214.634.216.455.10.796.52.795.09.367.8≈3 min
COLMAP Dense [45, 46]✓✓×26.952.738.022.589.823.220.869.325.776.440.248.8≈3 min
NVDS [54]××LSvld17.413.913.017.121.411.97.031.615.514.414.917.853.0
ChronoDepth [49]××LSvld15.314.814.714.521.111.37.031.014.316.814.517.77.0
DepthCrafter [20]××LSvld10.222.26.634.615.414.84.546.912.918.49.927.438.1
DepthAny. [62] (*)××LSvld8.428.15.244.313.918.63.755.110.425.38.334.30.1
DepthAny.V2 [63]××LSvld9.126.34.650.016.116.03.261.610.126.58.636.10.1
DUSt3R [56]××LSvld7.231.74.650.013.419.03.459.29.328.47.637.74.3
UniDepth [44]××LSvld(3.9)(58.2)(1.5)(90.1)12.718.97.030.710.623.27.144.20.5
Ours××LSvld6.932.44.550.89.227.13.262.47.332.96.241.10.9‡
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.453, + 0.907, + 0.496 + ], + "angle": 0, + "content": "Table 2. RobustMVD [48] multi-view depth benchmark, analogous to Tab. 1. Parentheses indicate that a model was trained on the same benchmark used for evaluation, which provides an advantage compared to training on different data. \\(\\ddagger\\)Includes approximately 0.4 seconds for tracking. \\(^\\ddagger\\)Metrics computed only for pixels with a valid prediction, which is an easier task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.482, + 0.597 + ], + "angle": 0, + "content": "18.3% and 9.7%. Second, Tracktention performs better than all other baselines, including DepthCrafter [20], which is a video depth model, by 15.3% and 2.3% on AbsRel and \\(\\delta_{1.25}\\), respectively. This demonstrates that Tracktention can turn an image-based model into a video-based one, outperforming models designed for video data from the start." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.703 + ], + "angle": 0, + "content": "Table 2 reports results on the RobustMVD benchmark [48], which features shorter video clips with potentially larger camera motions. Our model achieved the lowest average relative error of 6.2 and the highest average inlier ratio of \\(41.1\\%\\), surpassing all baselines. In particular, it improves the base model Depth Anything by \\(33.8\\%\\) on ETH3D and by \\(29.8\\%\\) on Tanks and Temples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Our model is also efficient, utilizing only 140 million parameters — only 17.1 million for the Tracktention module. It operates with a runtime of only 0.9 seconds per sequence, of which 0.4 seconds are attributed to the tracker. Despite its smaller size, our model outperforms DepthCrafter [20], which relies on video diffusion models with over 1.5 billion parameters. Tracktention will further benefit from future accuracy and speed improvements in point trackers." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.847, + 0.279, + 0.862 + ], + "angle": 0, + "content": "4.1.2 Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Figure 4 shows a qualitative comparison of depth estimation by our model, DepthCrafter, and DUSt3R in dynamic" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.907, + 0.627 + ], + "angle": 0, + "content": "scenes, complex textures, and cluttered environments. Our model produces stable, coherent depth maps across all frames, demonstrating strong temporal consistency. In contrast, DepthCrafter shows significant depth estimation errors in certain regions (blue box), especially in scenes with complex structures. DUSt3R, relying on implicit triangulation, struggles with dynamic content, resulting in inconsistent depth maps across time." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.63, + 0.907, + 0.676 + ], + "angle": 0, + "content": "Figure 6 illustrates the Attentional Sampling block, which pools information along the track by attending to image features in correspondence with the track locations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.692, + 0.78, + 0.707 + ], + "angle": 0, + "content": "4.2. Automatic Video Colorization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.717, + 0.906, + 0.763 + ], + "angle": 0, + "content": "The goal of video colorization is to add colors to an input grayscale video. Colors should be realistic, vibrant, and consistent across frames to avoid flickering." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We enhance four models — CIC [66], IDC [67], Colorformer [21], and DDColor [22] — by incorporating our Tracktention module into layers of each model's architecture. For instance, for DDColor, which has a ConvNet-based architecture with a ConvNeXt backbone [36] and a U-Net decoder, we insert our Tracktention module before each decoder block. We fine-tune the model using AdamW on the YouTube-VIS training set [60]. For testing, we use the validation sets of both DAVIS [43] and Video [30]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22814" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.092, + 0.887, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.34, + 0.907, + 0.381 + ], + "angle": 0, + "content": "Figure 4. Video depth prediction, comparing Tracktention (+DepthAnything), DepthCrafter [20], and DUS3R [56]. We visualize a column of pixels (highlighted in red) over time to illustrate temporal variation. Our model shows stable, coherent depth estimation over time, while DepthCrafter exhibits significant errors in certain regions (blue box). DUS3R struggles with dynamic content (green box)." + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.387, + 0.891, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.624, + 0.907, + 0.651 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of video colorization methods. Tracktention (+DDColor) yields vibrant, realistic, and consistent colors. In contrast, TCVC appears less vibrant (21.7 vs. 29.5 Colorfulness), while DDColor lacks temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.665, + 0.483, + 0.695 + ], + "angle": 0, + "content": "Evaluation metrics. We evaluate colorization results using standard metrics. Please refer to Appendix A.3.2 for details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.7, + 0.483, + 0.866 + ], + "angle": 0, + "content": "Quantitative Results. Figure 5 reports video colorization results on DAVIS and Video. We apply Tracktention to several image colorization models, consistently improving their performance when applied to videos, particularly in color consistency. Specifically, CDC improves by \\(22.8\\%\\) on DAVIS and \\(28.3\\%\\) on Video for CIC [66], \\(41.3\\%\\) and \\(28.8\\%\\) for Colorformer [21], and \\(46.5\\%\\) and \\(45.7\\%\\) for DD-Color [22]. Tracktention also improves FID and Colorfulness, indicating overall improvements in visual quality. We also compare to native video colorization networks (top rows), often outperforming these as well." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Qualitative Results. In Fig. 5, our method delivers more vibrant, realistic, and consistent colors compared to TCVC" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.665, + 0.907, + 0.726 + ], + "angle": 0, + "content": "and DDColor. TCVC's colors are less vibrant, with a colorfulness score of 21.7 compared to our 29.5, and DDColor exhibits inconsistencies across frames, with colors sometimes changing dramatically over time." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.739, + 0.681, + 0.754 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.907, + 0.809 + ], + "angle": 0, + "content": "Here, we measure the impact of various design decisions in Tracktention using the AbsRel metric on depth prediction in Sintel, ScanNet, and Tanks and Temples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Table 4a assesses the impact of the number of tracks used in training: even a single track per video can lead to improvements, and the optimal number is around \\(24^{2}\\) tracks. Table 4b assesses the number of tracks used at testing time, with similar results. Table 4c compares different tracking strategies: eschewing the tracker and simply setting tracks" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22815" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.625, + 0.292 + ], + "angle": 0, + "content": "
MethodDAVIS (medium frame length)Video (long frame length)
FID↓CF↑CDC↓PSNR↑FID↓CF↑CDC↓PSNR↑
Specialized video colorization models:
FAVC [31]—18.554.2224.38—16.281.8824.81
TCVC [34]46.5121.703.7325.5039.5819.071.6425.43
CIC [66]44.2330.346.1823.1939.7829.193.5922.51
[66]+Ours43.1334.224.77 (-22.8%)21.5837.0032.732.57 (-28.3%)22.64
IDC [67]42.9921.705.0125.4237.2519.072.5725.35
[67]+Ours33.4728.104.45 (-11.1%)23.5133.1325.222.42 (-6.0%)24.04
ColorFormer [21]40.7129.618.2923.0340.1229.764.9623.08
[21]+Ours29.7332.244.86 (-41.3%)23.1827.2529.653.53 (-28.8%)23.31
DDColor [22]26.8130.618.6423.8120.2730.385.6024.34
[22]+Ours24.6129.534.62 (-46.5%)23.8522.7826.063.04 (-45.7%)24.39
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.293, + 0.625, + 0.337 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison of video colorization methods on the DAVIS and Videvo datasets. Our method, when augmented onto four different baseline models, consistently improves the Color Distribution Consistency (CDC) metric across both datasets." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.342, + 0.356, + 0.419 + ], + "angle": 0, + "content": "
#tracksSintel↓ScanNet↓T&T↓
0 (base)0.3250.1300.104
10.3470.0940.087
12²0.3190.0860.080
24²0.2950.0870.073
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.421, + 0.356, + 0.457 + ], + "angle": 0, + "content": "(a) Number of training tracks: Increasing training tracks consistently improved depth estimation accuracy, with optimal results at \\(24^{2}\\) tracks." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.46, + 0.356, + 0.524 + ], + "angle": 0, + "content": "
MethodSintel↓ScanNet↓T&T↓
3D Conv.0.3640.1120.096
Time Attn.0.3630.0970.081
Tracktention0.2950.0870.073
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.526, + 0.356, + 0.564 + ], + "angle": 0, + "content": "(d) Temporal modeling: Tracktention outperformed 3D conv. and time attention by leveraging tracking data for better temporal alignment." + }, + { + "type": "table", + "bbox": [ + 0.375, + 0.342, + 0.631, + 0.419 + ], + "angle": 0, + "content": "
#tracksSintel↓ScanNet↓T&T↓
0 (base)0.3250.1300.104
10.3180.1180.108
12²0.3210.0870.075
24²0.2950.0870.073
" + }, + { + "type": "table_caption", + "bbox": [ + 0.368, + 0.42, + 0.63, + 0.457 + ], + "angle": 0, + "content": "(b) Number of testing tracks: Increasing testing tracks significantly boosted performance, indicating better temporal capture with more tracks." + }, + { + "type": "table", + "bbox": [ + 0.375, + 0.46, + 0.631, + 0.524 + ], + "angle": 0, + "content": "
#LayersSintel↓ScanNet↓T&T↓
10.3160.0890.074
20.2950.0870.073
30.3170.0830.073
" + }, + { + "type": "table_caption", + "bbox": [ + 0.367, + 0.526, + 0.63, + 0.564 + ], + "angle": 0, + "content": "(e) Number of Time Attention Layers: Two layers balanced performance and complexity, while additional layers offered no consistent benefits." + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.092, + 0.905, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.632, + 0.279, + 0.907, + 0.337 + ], + "angle": 0, + "content": "Figure 6. Attentional Sampling module. The module attends to image features in correspondence with each track, as shown by the attention maps next to each frame." + }, + { + "type": "table", + "bbox": [ + 0.649, + 0.342, + 0.905, + 0.419 + ], + "angle": 0, + "content": "
TypeSintel↓ScanNet↓T&T↓
Constant0.3520.0990.085
Grid, T00.3250.0960.073
Rand., T00.3200.0950.073
Rand., all0.2950.0870.073
" + }, + { + "type": "table_caption", + "bbox": [ + 0.642, + 0.42, + 0.905, + 0.457 + ], + "angle": 0, + "content": "(c) Type of tracks: Using a point tracker performs better than using static pseudo-tracks. Furthermore, querying tracks uniformly is better." + }, + { + "type": "table", + "bbox": [ + 0.649, + 0.46, + 0.905, + 0.524 + ], + "angle": 0, + "content": "
#LayersSintel↓ScanNet↓T&T↓
10.3440.0900.087
60.2950.0870.073
120.3330.1050.075
" + }, + { + "type": "table_caption", + "bbox": [ + 0.642, + 0.526, + 0.905, + 0.564 + ], + "angle": 0, + "content": "(f) Number of Tracktention layers: 6 layers improved temporal representation, but further layers led to diminishing returns due to overfitting risks." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.568, + 0.908, + 0.597 + ], + "angle": 0, + "content": "Table 4. Ablations on video depth estimation results. Each table examines the impact of different factors, such as the number and type of tracks used during training and testing. Results are reported using the AbsRel error metric, where \\( \\downarrow \\) indicates that lower values are better." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.604, + 0.486, + 0.786 + ], + "angle": 0, + "content": "to be constant across all frames, and initializing the tracks on a grid in the first frame, at random in the first frame, or at random throughout the video. The latter results in much better performance, showing the importance of tracking and ensuring that tracks cover the video well. Table 4d tests replacing Tracktention Layers with 3D convolutions and time attention, which, by comparison, barely improve results. Table 4e shows the impact of increasing the number of transformer layers in the Track Transformer, where two layers are optimal in most cases. Finally, Tab. 4f shows the impact of adding a different number of Tracktention layers, concluding that six is about optimal." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.801, + 0.218, + 0.815 + ], + "angle": 0, + "content": "4.4. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Tracktention's success depends on the quality of the extracted tracks. While modern point trackers are robust and efficient, they may still fail, particularly when tracks are off-camera for extended periods. Furthermore, running a point tracker in addition to the primary neural network adds com" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.604, + 0.907, + 0.636 + ], + "angle": 0, + "content": "plexity and computational cost (although this is often offset by avoiding spatio-temporal attention)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.656, + 0.634, + 0.672 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.684, + 0.907, + 0.867 + ], + "angle": 0, + "content": "We have introduced Tracktention, a plug-and-play layer that upgrades image predictor networks into video predictor ones. It leverages the power of modern point trackers, transferring their understanding of motion to new applications such as depth prediction and colorization. The module is lightweight, simple to integrate, and compatible with a variety of architectures, including Vision Transformers and ConvNets. Thanks to the efficiency of recent point trackers like CoTracker3, it incurs only a small computational overhead. Remarkably, the resulting upgraded image models are overall faster and more accurate than some state-of-the-art methods specifically designed for video prediction." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments. This research was supported by ERC 101001212-UNION." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22816" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6836-6846, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.269 + ], + "angle": 0, + "content": "[2] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, et al. Arkitsscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. arXiv preprint arXiv:2111.08897, 2021. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.273, + 0.483, + 0.315 + ], + "angle": 0, + "content": "[3] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.317, + 0.483, + 0.386 + ], + "angle": 0, + "content": "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.388, + 0.483, + 0.47 + ], + "angle": 0, + "content": "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.473, + 0.483, + 0.53 + ], + "angle": 0, + "content": "[6] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.483, + 0.585 + ], + "angle": 0, + "content": "[7] Seokju Cho, Jiahui Huang, Jisu Nam, Honggyu An, Seun-gryong Kim, and Joon-Young Lee. Local all-pair correspondence for point tracking. arXiv preprint arXiv:2407.15420, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.589, + 0.483, + 0.658 + ], + "angle": 0, + "content": "[8] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.661, + 0.483, + 0.744 + ], + "angle": 0, + "content": "[9] Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Peter Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, et al. Scaling vision transformers to 22 billion parameters. In International Conference on Machine Learning, pages 7480-7512. PMLR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[10] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610–13626, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[11] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10061-10072, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[12] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. TAPIR: tracking any point with per-frame initialization and temporal refinement. In Proc. CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[13] Carl Doersch, Yi Yang, Dilara Gokay, Pauline Luc, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ross Goroshin, João Carreira, and Andrew Zisserman. BootsTAP: Bootstrapped training for tracking-any-point. arXiv, 2402.00847, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[14] H Fan, B Xiong, K Mangalam, Y Li, Z Yan, and J Malik.... Multiscale vision transformers. In Proc. CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.25, + 0.908, + 0.318 + ], + "angle": 0, + "content": "[15] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6824-6835, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[16] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.378, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[17] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.434, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[18] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In European Conference on Computer Vision, pages 59-75. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.491, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[19] Adam W. Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle videos revisited: Tracking through occlusions using point trajectories. In Proc. ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.534, + 0.905, + 0.589 + ], + "angle": 0, + "content": "[20] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024. 5, 6, 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[21] Xiaozhong Ji, Boyuan Jiang, Donghao Luo, Guangpin Tao, Wenqing Chu, Zhifeng Xie, Chengjie Wang, and Ying Tai. Colorformer: Image colorization via color memory assisted hybrid-attention transformer. In European Conference on Computer Vision, pages 20–36. Springer, 2022. 6, 7, 8, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[22] Xiaoyang Kang, Tao Yang, Wenqi Ouyang, Peiran Ren, Lingzhi Li, and Xuansong Xie. Ddcolor: Towards photorealistic image colorization via dual decoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 328-338, 2023. 6, 7, 8, 12, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[23] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Dynamic stereo: Consistent dynamic depth from stereo videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13229-13239, 2023. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[24] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker3: Simpler and better point tracking by pseudolabelling real videos. arxiv, 2024. 5, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22817" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "tracker: It is better to track together. In Proc. ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[26] Nikita Karaev, Ignacio Rocco, Ben Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker: It is better to track together. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[27] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[28] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.483, + 0.376 + ], + "angle": 0, + "content": "[29] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. 3, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[30] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In Proceedings of the European conference on computer vision (ECCV), pages 170-185, 2018. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.483, + 0.504 + ], + "angle": 0, + "content": "[31] Chenyang Lei and Qifeng Chen. Fully automatic video colorization with self-regularization and diversity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3753–3761, 2019. 8, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.506, + 0.483, + 0.561 + ], + "angle": 0, + "content": "[32] Hongyang Li, Hao Zhang, Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, and Lei Zhang. Taptr: Tracking any point with transformers as detection. In European Conference on Computer Vision, pages 57-75. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.483, + 0.646 + ], + "angle": 0, + "content": "[33] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.483, + 0.716 + ], + "angle": 0, + "content": "[34] Yihao Liu, Hengyuan Zhao, Kelvin CK Chan, Xintao Wang, Chen Change Loy, Yu Qiao, and Chao Dong. Temporally consistent video colorization with deep feature propagation and self-regularization learning. arXiv preprint arXiv:2110.04562, 2021.8, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[35] Ziwei Liu, Raymond A Yeh, Xiaou Tang, Yiming Liu, and Aseem Agarwala. Video frame synthesis using deep voxel flow. In Proceedings of the IEEE international conference on computer vision, pages 4463-4471, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[36] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 6, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[37] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3202-3211, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.134 + ], + "angle": 0, + "content": "[38] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4):71-1, 2020. 3, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.907, + 0.219 + ], + "angle": 0, + "content": "[39] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4040-4048, 2016. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.906, + 0.262 + ], + "angle": 0, + "content": "[40] Daniel Neimark, Omri Bar, Maya Zohar, and Dotan Asselmann. Video transformer network. In Proc. ICCV Workshops, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.906, + 0.332 + ], + "angle": 0, + "content": "[41] E. Palazzolo, J. Behley, P. Lottes, P. Giguère, and C. Stachniss. ReFusion: 3D Reconstruction in Dynamic Environments for RGB-D Cameras Exploiting Residuals. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2019. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.906, + 0.417 + ], + "angle": 0, + "content": "[42] Mandela Patrick, Dylan Campbell, Yuki Markus Asano, Ishan Misra, Florian Metze, Christoph Feichtenhofer, Andrea Vedaldi, and João F. Henriques. Keeping your eye on the ball: Trajectory attention in video transformers. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.906, + 0.501 + ], + "angle": 0, + "content": "[43] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.906, + 0.558 + ], + "angle": 0, + "content": "[44] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. UniDepth: universal monocular metric depth estimation. In Proc. CVPR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.906, + 0.587 + ], + "angle": 0, + "content": "[45] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proc. CVPR, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.59, + 0.906, + 0.631 + ], + "angle": 0, + "content": "[46] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In Proc. ECCV, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.906, + 0.716 + ], + "angle": 0, + "content": "[47] Thomas Schops, Johannes L. Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.906, + 0.773 + ], + "angle": 0, + "content": "[48] Philipp Schröppel, Jan Bechtold, Artemij Amiranashvili, and Thomas Brox. A benchmark and a baseline for robust multiview depth estimation. In 2022 International Conference on 3D Vision (3DV), pages 637-645. IEEE, 2022. 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.906, + 0.829 + ], + "angle": 0, + "content": "[49] Jiahao Shao, Yuanbo Yang, Hongyu Zhou, Youmin Zhang, Yujun Shen, Matteo Poggi, and Yiyi Liao. Learning temporally consistent video depth from video diffusion priors. arXiv, 2406.01493, 2024. 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.906, + 0.899 + ], + "angle": 0, + "content": "[50] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22818" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[51] Jianlin Su, Yu Lu, Shengfeng Pan, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[52] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[53] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[54] Liao Wang, Qiang Hu, Qihan He, Ziyu Wang, Jingyi Yu, Tinne Tuytelaars, Lan Xu, and Minye Wu. Neural residual radiance fields for streamably free-viewpoint videos. In Proc. CVPR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[55] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19795-19806, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[56] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. DUS3R: Geometric 3D vision made easy. In Proc. CVPR, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[57] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4909-4916. IEEE, 2020. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[58] Xiaolong Wang, Ross B. Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proc. CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[59] Yiran Wang, Min Shi, Jiaqi Li, Zihao Huang, Zhiguo Cao, Jianming Zhang, Ke Xian, and Guosheng Lin. Neural video depth stabilizer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9466-9476, 2023. 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[60] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 6, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[61] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[62] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proc. CVPR, 2024. 5, 6, 12, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[63] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything V2. arXiv, 2406.09414, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[64] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 14" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[65] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[66] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 649-666. Springer, 2016. 6, 7, 8, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[67] Richard Zhang, Jun-Yan Zhu, Phillip Isola, Xinyang Geng, Angela S Lin, Tianhe Yu, and Alexei A Efros. Real-time user-guided image colorization with learned deep priors. arXiv preprint arXiv:1705.02999, 2017. 6, 8, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.345 + ], + "angle": 0, + "content": "[68] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19855-19865, 2023. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[69] Xizhou Zhu, Yuwen Xiong, Jifeng Dai, Lu Yuan, and Yichen Wei. Deep feature flow for video recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2349-2358, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22819" + } + ] +] \ No newline at end of file diff --git a/2025/Tracktention_ Leveraging Point Tracking to Attend Videos Faster and Better/55c83ffe-0183-42a7-a82c-e5f5408552e5_origin.pdf b/2025/Tracktention_ Leveraging Point Tracking to Attend Videos Faster and Better/55c83ffe-0183-42a7-a82c-e5f5408552e5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b8b18c94bc263b6f1158e2413a20118682204d53 --- /dev/null +++ b/2025/Tracktention_ Leveraging Point Tracking to Attend Videos Faster and Better/55c83ffe-0183-42a7-a82c-e5f5408552e5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a6fd0a28e982d7f485aabcafbe09ca851c2456a270d0584694b5466251f3740 +size 2724515 diff --git a/2025/Tracktention_ Leveraging Point Tracking to Attend Videos Faster and Better/full.md b/2025/Tracktention_ Leveraging Point Tracking to Attend Videos Faster and Better/full.md new file mode 100644 index 0000000000000000000000000000000000000000..47fc673728c8e665a5d650fa66ed7e80f299633b --- /dev/null +++ b/2025/Tracktention_ Leveraging Point Tracking to Attend Videos Faster and Better/full.md @@ -0,0 +1,338 @@ +# Tracktention: Leveraging Point Tracking to Attend Videos Faster and Better + +Zihang Lai Andrea Vedaldi Visual Geometry Group (VGG), University of Oxford + +{zlai,vedaldi}@robots.ox.ac.uk + +![](images/61729fc2a75a3e95485e8cb505c4548e0081624ed3e140e55dbf2408538c8dab.jpg) +Figure 1. Left: The Tracktention Layer is a plug-and-play module that can convert an image-based network (e.g., for monocular depth prediction) into a state-of-the-art video network (e.g., for video depth prediction). It does so by integrating the output of any off-the-shelf, modern, and powerful point trackers via track cross-attention. Right: For example, Tracktention achieves state-of-the-art and efficient video depth prediction by transforming Depth Anything into a video depth model. See Tab. 2 for detailed results. *Single-image models. + +![](images/86a95c21ef9b45cd361f1edd74c0707a424ef5eea5ebf14f86c8f43a42c4ba0a.jpg) + +![](images/07323bda3f0f1e56205a472b8a38445d51bf2a43b7ac149e4ab40573663087f5.jpg) + +# Abstract + +Temporal consistency is critical in video prediction to ensure that outputs are coherent and free of artifacts. Traditional methods, such as temporal attention and 3D convolution, may struggle with significant object motion and may not capture long-range temporal dependencies in dynamic scenes. To address this gap, we propose the Tracktention Layer, a novel architectural component that explicitly integrates motion information using point tracks, i.e., sequences of corresponding points across frames. By incorporating these motion cues, the Tracktention Layer enhances temporal alignment and effectively handles complex object motions, maintaining consistent feature representations over time. Our approach is computationally efficient and can be seamlessly integrated into existing models, such as Vision Transformers, with minimal modification. It can be used to upgrade image-only models to state-of-the-art video ones, sometimes outperforming models natively designed for video prediction. We demonstrate this on video depth prediction and video colorization, where models augmented with the Tracktention Layer exhibit significantly improved temporal consistency compared to baselines. Project website: zlai0.github.io/TrackTention. + +# 1. Introduction + +Compared to image analysis tasks, video analysis tasks, such as video segmentation, video depth estimation, and video colorization, pose additional challenges due to the temporal dimension inherent in video data. A particularly important challenge is ensuring that outputs are temporally consistent across frames, which is necessary to produce coherent and artifact-free results. Temporal inconsistencies, such as flickering or abrupt changes in predicted object attributes, are especially noticeable in applications like colorization, where the output is a new video. + +An aspect of video analysis that has seen significant progress in recent years is point tracking. New trackers such as PIPs [19], TAPIR [12], BootsTAP [13], and CoTracker [26] can now track quasi-dense collections of points across long video sequences with high reliability and efficiency. So far, these trackers have been primarily used in selected applications like 3D reconstruction and object tracking through ad-hoc algorithms. In this paper, we ask whether progress in point tracking can benefit a much wider range of video analysis tasks. + +To answer this question, we propose the Tracktention Layer, a novel architectural component designed to enhance temporal consistency in video transformers. Tracktention integrates seamlessly into vision transformers as a new at + +![](images/1df24ea7727c5d35d33c2a108f07d3ff05cd5ed41f342ba90eb524ebb2d46ce8.jpg) +Figure 2. Overview of Tracktention. We begin by using an off-the-shelf point tracker to extract a number of video tracks. Given these, we first sample image tokens at the track locations, obtaining corresponding track tokens (Sec. 3.1). Next, we use a Track Transformer to update these tokens, propagating information temporally at consistent spatial locations (Sec. 3.2). Finally, we splat the information back to the image tokens (Sec. 3.3). By explicitly incorporating motion information through point tracks, Tracktention improves temporal alignment, effectively captures complex object movements, and ensures stable feature representations over time. + +tention layer, making it compatible with a wide range of existing models. The key innovation of Tracktention lies in leveraging pre-extracted point tracks to inform the model about temporal correspondences between image tokens in different frames (Fig. 2). This is achieved by sampling the image tokens at the track points. The information is then propagated along the tracks using a transformer layer, which justifies the name "Tracktention". Finally, the information is splatted back to the image tokens, allowing the computation to progress as normal. + +By explicitly incorporating tracking information, the model can establish direct correspondences, enabling precise temporal alignment regardless of how much objects move. This explicit temporal alignment allows the model to maintain consistent feature representations over time, smoothing features, and reducing temporal artifacts. + +The problem of propagating information in video neural networks has been extensively studied in the literature. Many approaches are implicit, in the sense that they do not explicitly estimate or account for the motion observed in the scene. Examples include 3D convolution [6, 53] and space-time attention [1, 3, 14, 40, 58]. However, to maintain manageable computational costs, these methods often reduce the spatial resolution of features and limit the temporal range they can model—either by using small convolution kernels or by decoupling spatial and temporal attention. The resulting networks may struggle to represent motion precisely and to keep up with large object displacements. + +In contrast, Tracktention explicitly accounts for motion at the resolution of point tracks, which is much finer than that of typical image features. By pooling and splatting information at image locations that are already put in correspondence by the tracker, it only needs to propagate information along the tracks, obviating the need to implicitly establish such correspondences via space-time attention. + +There are also approaches to video analysis that, like Tracktention, incorporate motion information explicitly. They often work by feeding optical flow to the network, + +capturing pixel-level motion between frames [42]. While optical flow can be effective in certain contexts, it may struggle with occlusions or large displacements. Building on the power of point trackers, Tracktention can handle occlusions, large displacements, and long temporal dependencies more effectively. + +In summary, Tracktention offers several advantages over traditional video network architectures: (1) It leverages the power of modern point tracker models, which are excellent motion experts, and provides a general mechanism to inject this knowledge into any video transformer. In particular, it can handle complex and large object motions, limited only by the capabilities of the point trackers. (2) It directly establishes long-term space-time correspondences, unlike 3D convolutions that are local and unlike separated space-time attention, which can only do so indirectly. (3) It avoids the need to repeatedly calculate such correspondences explicitly, as full spatio-temporal attention does, achieving much greater efficiency and operating at a much finer resolution. + +We deliver Tracktention as a plugin layer (Fig. 1) that can be added to single-image neural network architectures to extend them to powerful, temporally-aware video models. For example, we demonstrate that a pre-trained single-image depth predictor can be extended to a corresponding state-of-the-art video depth predictor by integrating our Tracktention Layer. We achieve similar results for automatic video colorization, outperforming models natively designed to process videos while also being more efficient. + +# 2. Related Work + +Temporal Modeling for Videos. Temporal modeling is a cornerstone of video understanding tasks. Over the years, various techniques have been proposed to capture temporal dependencies in video data. 3D CNNs [6, 16, 53] extend 2D convolutions into the temporal dimension, allowing for direct spatiotemporal feature extraction from video data. Temporal attention mechanisms provide a new approach to + +effectively capture long-range dependencies across frames. Video transformers [1, 3, 15, 37, 42] apply self-attention mechanisms over temporal and spatial dimensions, typically separately or hierarchically, to reduce computational complexity while modeling temporal dependencies effectively. Recent video processing models [4, 5, 50] also use a hybrid of these approaches. However, these methods have notable limitations: temporal attention struggles with rapid object movements, while 3D convolutions assume local spatiotemporal correlations, failing with unpredictable motion. In this work, we explicitly leverage tracking information to precisely align temporal features regardless of object movement. + +Consistent Video Prediction. Addressing temporal inconsistency is critical for ensuring smooth and coherent video outputs. Optical flow-based methods [69] enforce temporal coherence by aligning features or predictions across frames, propagating features along motion paths estimated by optical flow. Liu et al. [35] utilize convolutional LSTMs to capture temporal information for consistent video frame synthesis, modeling temporal dependencies through recurrent connections. Lai et al. [30] propose a post-processing model with a temporal consistency objective for video prediction. Similarly, Luo et al. [38] and Kopf et al. [29] present test-time training post-processing methods specifically designed to stabilize video depth predictions using geometric consistency cues. In contrast, our work aims to stabilize features by allowing them to attend to corresponding areas across time based on point tracks. This approach inherently accounts for object movement and embeds temporal alignment efficiently within the model. + +Point Tracking. Point tracking is the task of identifying and continuously tracking specific points across frames in a video. PIPs [18] pioneered a neural network-based approach that updates tracked points iteratively by extracting correlation maps between frames, inspired by the RAFT model for optical flow [52]. TAP-Vid [10] proposed a comprehensive benchmark and TAP-Net, a new model for point tracking. TAPIR [11] further improved performance by combining global matching capabilities with PIPs, enhancing the accuracy of tracked points even in complex motion scenarios. OmniMotion [55] tracks points by leveraging a quasi-3D canonical volume but is approximately $10^{3}$ times slower than CoTracker due to its test-time optimization. CoTracker [25] leveraged a transformer-based architecture to track multiple points jointly, yielding improvements particularly for occluded points. Recent methods such as LocoTrack [7], which introduced 4D correlation features, and TAPTR [32] further improved precision. + +In this work, we do not propose new models for point tracking. Instead, we use existing trackers to establish direct token correspondences, ensuring precise temporal alignment despite object motion. + +# 3. Method + +In this section, we introduce Tracktention, a novel transformer layer designed to aggregate and redistribute feature information in video transformers in a motion-aware manner (Fig. 3). By leveraging motion information captured by the output of a point tracker, this layer improves the temporal consistency and motion awareness of video features. + +Tracktention comprises an Attentional Sampling block (Sec. 3.1) that pools information from the video features into the tracks, a Track Transformer (Sec. 3.2) that updates the track tokens, and an Attentional Splatting block (Sec. 3.3) that pushes the updated track tokens back to the video features. We describe these components next. + +# 3.1. Attentional Sampling + +The Attentional Sampling module is the first stage in Track-attention. It uses cross-attention to pool information from the video features into the track tokens. + +Let $F\in \mathbb{R}^{T\times HW\times D_f}$ denote the input feature map, where $T$ is the number of time steps, $H$ and $W$ are the spatial dimensions, and $D_{f}$ is the feature dimension. Let $\mathbf{P}\in \mathbb{R}^{T\times M\times 2}$ represent a set of $M$ tracks, each of which is a sequence of 2D points. We first apply a positional embedding to these 2D points, obtaining one token per track and per frame, and arrange these tokens into a tensor $\mathcal{T}\in \mathbb{R}^{T\times M\times D_f}$ . We then project the track tokens into queries $Q\in \mathbb{R}^{T\times M\times D_k}$ and the feature map into keys $K\in \mathbb{R}^{T\times HW\times D_k}$ and values $V\in \mathbb{R}^{T\times HW\times D_f}$ as: + +$$ +Q _ {t} = \mathcal {T} _ {t} W _ {Q}, \quad K _ {t} = F _ {t} W _ {K}, \quad V _ {t} = F _ {t}, +$$ + +where $W_{Q}$ and $W_{K}$ are learnable weight matrices, and the subscript $t \in \{1, \dots, T\}$ indexes the first dimensions of the tensors, extracting matrices. Note that we avoid projecting the values to preserve the original features during sampling. + +For each time $t$ , the attention weights $A_{t} \in \mathbb{R}^{M \times HW}$ are computed using the scaled dot-product attention mechanism: + +$$ +A _ {t} = \operatorname {s o f t m a x} \left(\frac {Q _ {t} K _ {t} ^ {\top}}{\sqrt {D _ {k}}} + B _ {t}\right), \tag {1} +$$ + +where $B_{t}\in \mathbb{R}^{M\times HW}$ is a bias term that encourages attention to align with the tracks. It is defined as: + +$$ +B _ {t i j} = - \frac {\| P _ {t i} - \mathrm {p o s} _ {F _ {t}} (j) \| ^ {2}}{2 \sigma^ {2}}, +$$ + +where $i\in \{1,\dots ,M\}$ $j\in \{1,\ldots ,HW\}$ $P_{ti}\in \mathbb{R}^2$ is the spatial location of track $i$ at time $t$ and $\mathrm{pos}_{F_t}(j)\in \mathbb{R}^2$ is the spatial location of the $j$ -th feature map token at time $t$ + +In the softmax operator of Eq. (1), this bias is analogous to multiplying the attention map by a Gaussian window that decays exponentially as the feature position deviates from the track position. The parameter $\sigma$ , which we set to $1/2$ , determines how fast the window decays. + +![](images/9dd77097f4f8489e4bd92f770a6f3256a0bfae7209f1718b88d3509302743314.jpg) +(A) Tracktention Architecture + +![](images/059f4f5464bd5a27fc712040548c59b8177ff89f3b912a192b93f8157f26c20e.jpg) +Figure 3. Left: the Tracktention architecture comprises Attentional Sampling, pooling information from images to track, Track Transformer, processing this information temporally, and Attentional Splatting, moving the processed information back to the images. Right: Tracktention is easily integrated in ViTs and ConvNets to make video networks out image ones. + +We use the attention to "sample" track features $S \in \mathbb{R}^{T \times M \times D_f}$ by setting $S_t = A_t V_t = A_t F_t \in \mathbb{R}^{M \times D_f}$ for each time $t$ . Alternatively, we could have obtained $S$ by sampling the features $F$ at the given track locations, for instance, by using bilinear sampling. This can also be expressed as $A_t F_t$ for a specific matrix $A_t$ , so sampling is a special case of our Attentional Sampling block. The advantage of Attentional Sampling is that it allows the model to learn a better sampling strategy. + +Relative Spatial Information. To capture the relative spatial relationships between track tokens and feature map locations, we incorporate Rotational Position Encodings (RoPE) [51] into the keys and values derived from the feature map. RoPE is a form of positional encoding that allows attention mechanisms to be aware of the relative positions of points defined in a continuous space. + +To compute the RoPE embedding $\mathrm{RoPE}(\mathbf{f})$ of a feature vector $\mathbf{f}$ at spatial location $(x,y)$ , we use the first $d / 2$ channels to encode the $x$ position. For $i = 1,\dots ,d / 4$ , we define: + +$$ +\begin{array}{r} \mathrm {R o P E} (\mathbf {f}) [ 2 i ] = \mathbf {f} [ 2 i ] \cdot \cos (\theta_ {i} x) - \mathbf {f} [ 2 i + 1 ] \cdot \sin (\theta_ {i} x), \\ \mathrm {R o P E} (\mathbf {f}) [ 2 i + 1 ] = \mathbf {f} [ 2 i ] \cdot \sin (\theta_ {i} x) + \mathbf {f} [ 2 i + 1 ] \cdot \cos (\theta_ {i} x), \end{array} +$$ + +where $\theta_{i} = 100^{\frac{-2(i - 1)}{d / 2}}$ . The second $d / 2$ channels encode the $y$ position analogously by replacing $x$ with $y$ in the above equations. + +Implementation Details. To stabilize the training process, especially due to the difference in scale between the track token embeddings and the feature map embeddings (which tend to have larger norms), we apply QK-normalization [9] to the queries and keys before computing the attention weights. This normalization ensures that the scale of the + +embeddings is consistent, preventing issues related to gradient instability. We also use $H$ attention heads, allowing the model to capture diverse contextual relationships across different subspaces. + +# 3.2. Track Transformer + +Once the track tokens $S$ are obtained, the Track Transformer processes them to propagate information along the tracks, making them more temporally consistent and smoothing out any irregularities. It is designed as a transformer that operates along the time dimension of the point tracks and outputs updated track tokens $\mathcal{T}'$ . + +In more detail, recall that the track tokens $S$ are of shape $T \times M \times D_f$ . We swap the first two dimensions to obtain a tensor of shape $M \times T \times D_f$ . Then, we apply a transformer block, interpreting the first dimension $M$ as the batch dimension. The effect is to apply attention along the temporal dimension, allowing the model to learn temporal dependencies along each track. + +This approach ensures that information is not exchanged across tracks, meaning each track's temporal sequence is processed independently. We considered a variant of the Track Transformer that included attention across tracks as well, but found that this approach was slightly worse and more computationally expensive. Our interpretation is that information is already exchanged spatially by the Vision Transformer paired with Tracktention, making additional exchange of information across locations redundant. + +Architecture. For each track, we utilize a 2-layer transformer encoder along the temporal dimension, structured as a standard transformer model. Each layer consists of Multi-Head Self-Attention, which allows each time step within a track to attend to all other time steps within that same track, + +and Position-Wise Feed-Forward Networks, which apply nonlinear transformations independently to each time step, enhancing the model's representational capacity. We also incorporate sinusoidal positional encodings to provide the model with a sense of temporal order. These encodings are added to the input features prior to the transformer layers, ensuring the model can distinguish between different time steps and understand the temporal structure of the sequence. + +# 3.3. Attentional Splatting + +The Attentional Splitting module maps the updated track tokens $\mathcal{T}'$ back onto the feature maps $F$ . This is achieved by reversing the roles of the track tokens and the feature tokens in the various expressions in Sec. 3.1. This means the queries are derived from the grid coordinates of the output feature map, while the keys and values are generated from the track tokens. The same bias term $B_{t}$ is used as before (but now the matrix is transposed), as well as the same $QK$ -normalization and RoPE encodings, this time obtaining a matrix $A_{t}^{\prime} \in \mathbb{R}^{HW \times M}$ . The final output Tracktention $(F)$ is then computed as: + +$$ +\left[ \operatorname {T r a c k t e n t i o n} (F) \right] _ {t} = W _ {\text {o u t}} A _ {t} ^ {\prime} \left[ \mathcal {T} ^ {\prime} \right] _ {t}, +$$ + +where $W_{\mathrm{out}}$ is a final output projection. By adopting a symmetric design for Attentional Sampling and Splitting, we ensure that information is handled in a consistent manner. + +# 3.4. Point Tracker Pre-Processor + +To apply the Tracktention layer to a video, we first need to extract the point tracks $\mathbf{P}$ . We use CoTracker3 [24] due to its robustness in handling complex motions and occlusions in videos. Trackers require a seed point for each track, called a query. We sample 576 points uniformly at random from the spatio-temporal video volume and track each forward and backward in time. Compared to initializing tracks on a grid in the first frame of the video, this simple initialization scheme encourages tracks to cover the video well, despite camera and object motion. + +# 3.5. Integration into Image-Based Models + +A key application of Tracktention is to build video neural networks from image-based ones. Integrating Tracktention into existing image-based models, such as a ViT or ConvNet, is straightforward. For example, monocular depth estimation models like Depth Anything, which are designed for single images, may lack temporal consistency when applied to video frame by frame. Adding Tracktention layers after each transformer block and fine-tuning the model on videos can increase temporal stability and significantly improve performance in video-based applications. + +To incorporate Tracktention into an existing network backbone, we insert it immediately after all or some transformer or convolutional blocks. We also include a residual + +connection to preserve the original information flow, setting $F' = F + \text{Tracktention}(F)$ , where $F$ represents the output feature map from the preceding block, and $F'$ is the updated feature map after the Tracktention module. + +Furthermore, the Tracktention output projection, $W_{\mathrm{out}}$ , is initialized to zero, preserving the original network output at training onset. This strategy retains benefits of pre-trained weights while allowing the model to gradually adapt to temporal updates. Consequently, Tracktention modules capture fine spatial details and ensure consistent temporal modeling across layers. + +# 4. Experiments + +We show that Tracktention can turn image-based models into state-of-the-art ones for video depth estimation and colorization (Secs. 4.1 and 4.2) and ablate its design (Sec. 4.3). + +# 4.1. Video Depth Estimation + +To apply Tracktention to video depth estimation, we start from an image-based model, Depth Anything [61], and upgrade it to a video-based model using the Tracktention module. The Depth Anything model has a DINO ViT backbone consisting of 12 transformer blocks. We insert our Tracktention module after each of the last 6 blocks. + +We fine-tune the model on a large dataset of videos, using a combination of synthetic and real data. This data comes with ground-truth depth, which we fit using a scale-and shift-invariant loss, as Depth Anything predicts depth up to an affine transformation. However, we share the same calibration parameters for all frames in a video, as calibrating frames independently would mask temporal inconsistencies, which we want the model to learn to correct. + +To reduce the number of new parameters added to the model, we share them between all Tracktention modules. During training, the original model is kept frozen, and only the Tracktention modules are updated. We use the AdamW optimizer with an initial learning rate of $1.6 \times 10^{-5}$ and cosine learning rate decay. Training is conducted for 4 epochs with a batch size of 4 videos, each containing a randomly sampled number of frames between 8 and 16. + +Data and evaluation metrics. For training, we use a combination of datasets containing both synthetic and real videos. We use standard depth estimation metrics [62]. Please refer to Appendix A.3.1 for details. + +# 4.1.1 Quantitative Results + +In Tab. 1, we evaluate our model on four video collections in the DepthCrafter benchmark [20] against several alternatives. Tracktention improves the image-based model it augments, i.e., Depth Anything [62], substantially: Sintel AbsRel improves by $9.2\%$ and $\delta_{1}$ by $13.5\%$ , KITTI by $26.8\%$ and $12.5\%$ , Bonn by $15.4\%$ and $3.4\%$ , and on average by + +
MethodType#ParamsSintel (~50 frames)Scannet (90 frames)KITTI (110 frames)Bonn (110 frames)Average
AbsRel ↓δ1.25 ↑AbsRel ↓δ1.25 ↑AbsRel ↓δ1.25 ↑AbsRel ↓δ1.25 ↑AbsRel ↓δ1.25 ↑
DUS3R [56]Vid.578M0.6280.3930.1940.6940.2920.4560.2500.5880.3410.533
NVDS [59]Vid.430M0.4080.4830.1870.6770.2530.5880.1670.7660.2540.629
ChronoDepth [49]Vid.1521M0.5870.4860.1590.7830.1670.7590.1000.9110.2530.735
DepthCrafter [20]Vid.1521M0.343†0.673†0.1250.8480.1100.8810.0750.9710.1630.843
Marigold [27]Im.865M0.5320.5150.1660.7690.1490.7960.0910.9310.2350.753
DepthAny. [62] (*)Im.343M0.3250.5640.1300.8380.1420.8030.0780.9390.1690.786
DepthAny.-V2 [63]Im.343M0.3670.5540.1350.8220.1400.8040.1060.9210.1870.775
OursVid.140M0.2950.6400.0870.9330.1040.9030.0660.9710.1380.862
+ +Table 1. DepthCrafter [20] video depth benchmark. Our model upgrades DepthAnything (*) to a video depth predictor, outperforming all baselines (image- or video-based) while having the smallest parameter count (140M). We use DepthAnything-Base (97M) as the base model and still outperform the DepthAnything-Large (343M) model. $\dagger$ Reproduced results confirmed by original authors. + +
MethodsGT PoseGT Intr.AlignKITTIScanNetETH3DDTUT&TAverageRuntime (s)
rel ↓δ1.03 ↑rel ↓δ1.03 ↑rel ↓δ1.03 ↑rel ↓δ1.03 ↑rel ↓δ1.03 ↑rel ↓δ1.03 ↑
COLMAP [45, 46]#✓✓×12.058.214.634.216.455.10.796.52.795.09.367.8≈3 min
COLMAP Dense [45, 46]✓✓×26.952.738.022.589.823.220.869.325.776.440.248.8≈3 min
NVDS [54]××LSvld17.413.913.017.121.411.97.031.615.514.414.917.853.0
ChronoDepth [49]××LSvld15.314.814.714.521.111.37.031.014.316.814.517.77.0
DepthCrafter [20]××LSvld10.222.26.634.615.414.84.546.912.918.49.927.438.1
DepthAny. [62] (*)××LSvld8.428.15.244.313.918.63.755.110.425.38.334.30.1
DepthAny.V2 [63]××LSvld9.126.34.650.016.116.03.261.610.126.58.636.10.1
DUSt3R [56]××LSvld7.231.74.650.013.419.03.459.29.328.47.637.74.3
UniDepth [44]××LSvld(3.9)(58.2)(1.5)(90.1)12.718.97.030.710.623.27.144.20.5
Ours××LSvld6.932.44.550.89.227.13.262.47.332.96.241.10.9‡
+ +Table 2. RobustMVD [48] multi-view depth benchmark, analogous to Tab. 1. Parentheses indicate that a model was trained on the same benchmark used for evaluation, which provides an advantage compared to training on different data. $\ddagger$ Includes approximately 0.4 seconds for tracking. $^\ddagger$ Metrics computed only for pixels with a valid prediction, which is an easier task. + +18.3% and 9.7%. Second, Tracktention performs better than all other baselines, including DepthCrafter [20], which is a video depth model, by 15.3% and 2.3% on AbsRel and $\delta_{1.25}$ , respectively. This demonstrates that Tracktention can turn an image-based model into a video-based one, outperforming models designed for video data from the start. + +Table 2 reports results on the RobustMVD benchmark [48], which features shorter video clips with potentially larger camera motions. Our model achieved the lowest average relative error of 6.2 and the highest average inlier ratio of $41.1\%$ , surpassing all baselines. In particular, it improves the base model Depth Anything by $33.8\%$ on ETH3D and by $29.8\%$ on Tanks and Temples. + +Our model is also efficient, utilizing only 140 million parameters — only 17.1 million for the Tracktention module. It operates with a runtime of only 0.9 seconds per sequence, of which 0.4 seconds are attributed to the tracker. Despite its smaller size, our model outperforms DepthCrafter [20], which relies on video diffusion models with over 1.5 billion parameters. Tracktention will further benefit from future accuracy and speed improvements in point trackers. + +# 4.1.2 Qualitative Results + +Figure 4 shows a qualitative comparison of depth estimation by our model, DepthCrafter, and DUSt3R in dynamic + +scenes, complex textures, and cluttered environments. Our model produces stable, coherent depth maps across all frames, demonstrating strong temporal consistency. In contrast, DepthCrafter shows significant depth estimation errors in certain regions (blue box), especially in scenes with complex structures. DUSt3R, relying on implicit triangulation, struggles with dynamic content, resulting in inconsistent depth maps across time. + +Figure 6 illustrates the Attentional Sampling block, which pools information along the track by attending to image features in correspondence with the track locations. + +# 4.2. Automatic Video Colorization + +The goal of video colorization is to add colors to an input grayscale video. Colors should be realistic, vibrant, and consistent across frames to avoid flickering. + +We enhance four models — CIC [66], IDC [67], Colorformer [21], and DDColor [22] — by incorporating our Tracktention module into layers of each model's architecture. For instance, for DDColor, which has a ConvNet-based architecture with a ConvNeXt backbone [36] and a U-Net decoder, we insert our Tracktention module before each decoder block. We fine-tune the model using AdamW on the YouTube-VIS training set [60]. For testing, we use the validation sets of both DAVIS [43] and Video [30]. + +![](images/06c1fb5114cc00995fd1eafdc60c0f23969fe68ad34da647e15834e518a2881d.jpg) +Figure 4. Video depth prediction, comparing Tracktention (+DepthAnything), DepthCrafter [20], and DUS3R [56]. We visualize a column of pixels (highlighted in red) over time to illustrate temporal variation. Our model shows stable, coherent depth estimation over time, while DepthCrafter exhibits significant errors in certain regions (blue box). DUS3R struggles with dynamic content (green box). + +![](images/6b571058b25d6ece888a668a281c9178d228c77eec8273c173d0e3d8434ea67b.jpg) +Figure 5. Qualitative comparison of video colorization methods. Tracktention (+DDColor) yields vibrant, realistic, and consistent colors. In contrast, TCVC appears less vibrant (21.7 vs. 29.5 Colorfulness), while DDColor lacks temporal consistency. + +Evaluation metrics. We evaluate colorization results using standard metrics. Please refer to Appendix A.3.2 for details. + +Quantitative Results. Figure 5 reports video colorization results on DAVIS and Video. We apply Tracktention to several image colorization models, consistently improving their performance when applied to videos, particularly in color consistency. Specifically, CDC improves by $22.8\%$ on DAVIS and $28.3\%$ on Video for CIC [66], $41.3\%$ and $28.8\%$ for Colorformer [21], and $46.5\%$ and $45.7\%$ for DD-Color [22]. Tracktention also improves FID and Colorfulness, indicating overall improvements in visual quality. We also compare to native video colorization networks (top rows), often outperforming these as well. + +Qualitative Results. In Fig. 5, our method delivers more vibrant, realistic, and consistent colors compared to TCVC + +and DDColor. TCVC's colors are less vibrant, with a colorfulness score of 21.7 compared to our 29.5, and DDColor exhibits inconsistencies across frames, with colors sometimes changing dramatically over time. + +# 4.3. Ablation Studies + +Here, we measure the impact of various design decisions in Tracktention using the AbsRel metric on depth prediction in Sintel, ScanNet, and Tanks and Temples. + +Table 4a assesses the impact of the number of tracks used in training: even a single track per video can lead to improvements, and the optimal number is around $24^{2}$ tracks. Table 4b assesses the number of tracks used at testing time, with similar results. Table 4c compares different tracking strategies: eschewing the tracker and simply setting tracks + +
MethodDAVIS (medium frame length)Video (long frame length)
FID↓CF↑CDC↓PSNR↑FID↓CF↑CDC↓PSNR↑
Specialized video colorization models:
FAVC [31]—18.554.2224.38—16.281.8824.81
TCVC [34]46.5121.703.7325.5039.5819.071.6425.43
CIC [66]44.2330.346.1823.1939.7829.193.5922.51
[66]+Ours43.1334.224.77 (-22.8%)21.5837.0032.732.57 (-28.3%)22.64
IDC [67]42.9921.705.0125.4237.2519.072.5725.35
[67]+Ours33.4728.104.45 (-11.1%)23.5133.1325.222.42 (-6.0%)24.04
ColorFormer [21]40.7129.618.2923.0340.1229.764.9623.08
[21]+Ours29.7332.244.86 (-41.3%)23.1827.2529.653.53 (-28.8%)23.31
DDColor [22]26.8130.618.6423.8120.2730.385.6024.34
[22]+Ours24.6129.534.62 (-46.5%)23.8522.7826.063.04 (-45.7%)24.39
+ +Table 3. Quantitative comparison of video colorization methods on the DAVIS and Videvo datasets. Our method, when augmented onto four different baseline models, consistently improves the Color Distribution Consistency (CDC) metric across both datasets. + +
#tracksSintel↓ScanNet↓T&T↓
0 (base)0.3250.1300.104
10.3470.0940.087
12²0.3190.0860.080
24²0.2950.0870.073
+ +(a) Number of training tracks: Increasing training tracks consistently improved depth estimation accuracy, with optimal results at $24^{2}$ tracks. + +
MethodSintel↓ScanNet↓T&T↓
3D Conv.0.3640.1120.096
Time Attn.0.3630.0970.081
Tracktention0.2950.0870.073
+ +(d) Temporal modeling: Tracktention outperformed 3D conv. and time attention by leveraging tracking data for better temporal alignment. + +
#tracksSintel↓ScanNet↓T&T↓
0 (base)0.3250.1300.104
10.3180.1180.108
12²0.3210.0870.075
24²0.2950.0870.073
+ +(b) Number of testing tracks: Increasing testing tracks significantly boosted performance, indicating better temporal capture with more tracks. + +
#LayersSintel↓ScanNet↓T&T↓
10.3160.0890.074
20.2950.0870.073
30.3170.0830.073
+ +![](images/f01fabddce6f52c7beefcc580b00dd92b0e88ea4dafb4bf1ef3f689e8fcf30c3.jpg) +Figure 6. Attentional Sampling module. The module attends to image features in correspondence with each track, as shown by the attention maps next to each frame. + +(e) Number of Time Attention Layers: Two layers balanced performance and complexity, while additional layers offered no consistent benefits. + +
TypeSintel↓ScanNet↓T&T↓
Constant0.3520.0990.085
Grid, T00.3250.0960.073
Rand., T00.3200.0950.073
Rand., all0.2950.0870.073
+ +(c) Type of tracks: Using a point tracker performs better than using static pseudo-tracks. Furthermore, querying tracks uniformly is better. + +
#LayersSintel↓ScanNet↓T&T↓
10.3440.0900.087
60.2950.0870.073
120.3330.1050.075
+ +(f) Number of Tracktention layers: 6 layers improved temporal representation, but further layers led to diminishing returns due to overfitting risks. + +Table 4. Ablations on video depth estimation results. Each table examines the impact of different factors, such as the number and type of tracks used during training and testing. Results are reported using the AbsRel error metric, where $\downarrow$ indicates that lower values are better. + +to be constant across all frames, and initializing the tracks on a grid in the first frame, at random in the first frame, or at random throughout the video. The latter results in much better performance, showing the importance of tracking and ensuring that tracks cover the video well. Table 4d tests replacing Tracktention Layers with 3D convolutions and time attention, which, by comparison, barely improve results. Table 4e shows the impact of increasing the number of transformer layers in the Track Transformer, where two layers are optimal in most cases. Finally, Tab. 4f shows the impact of adding a different number of Tracktention layers, concluding that six is about optimal. + +# 4.4. Limitations + +Tracktention's success depends on the quality of the extracted tracks. While modern point trackers are robust and efficient, they may still fail, particularly when tracks are off-camera for extended periods. Furthermore, running a point tracker in addition to the primary neural network adds com + +plexity and computational cost (although this is often offset by avoiding spatio-temporal attention). + +# 5. Conclusion + +We have introduced Tracktention, a plug-and-play layer that upgrades image predictor networks into video predictor ones. It leverages the power of modern point trackers, transferring their understanding of motion to new applications such as depth prediction and colorization. The module is lightweight, simple to integrate, and compatible with a variety of architectures, including Vision Transformers and ConvNets. Thanks to the efficiency of recent point trackers like CoTracker3, it incurs only a small computational overhead. Remarkably, the resulting upgraded image models are overall faster and more accurate than some state-of-the-art methods specifically designed for video prediction. + +Acknowledgments. This research was supported by ERC 101001212-UNION. + +# References + +[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6836-6846, 2021. 2, 3 +[2] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, et al. Arkitsscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. arXiv preprint arXiv:2111.08897, 2021. 14 +[3] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 2, 3 +[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3 +[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3 +[6] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 2 +[7] Seokju Cho, Jiahui Huang, Jisu Nam, Honggyu An, Seun-gryong Kim, and Joon-Young Lee. Local all-pair correspondence for point tracking. arXiv preprint arXiv:2407.15420, 2024.3 +[8] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 14 +[9] Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Peter Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, et al. Scaling vision transformers to 22 billion parameters. In International Conference on Machine Learning, pages 7480-7512. PMLR, 2023. 4 +[10] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610–13626, 2022. 3 +[11] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10061-10072, 2023. 3 + +[12] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. TAPIR: tracking any point with per-frame initialization and temporal refinement. In Proc. CVPR, 2023. 1 +[13] Carl Doersch, Yi Yang, Dilara Gokay, Pauline Luc, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ross Goroshin, João Carreira, and Andrew Zisserman. BootsTAP: Bootstrapped training for tracking-any-point. arXiv, 2402.00847, 2024. 1 +[14] H Fan, B Xiong, K Mangalam, Y Li, Z Yan, and J Malik.... Multiscale vision transformers. In Proc. CVPR, 2021. 2 +[15] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6824-6835, 2021. 3 +[16] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 2 +[17] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 14 +[18] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In European Conference on Computer Vision, pages 59-75. Springer, 2022. 3 +[19] Adam W. Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle videos revisited: Tracking through occlusions using point trajectories. In Proc. ECCV, 2022. 1 +[20] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024. 5, 6, 7, 14 +[21] Xiaozhong Ji, Boyuan Jiang, Donghao Luo, Guangpin Tao, Wenqing Chu, Zhifeng Xie, Chengjie Wang, and Ying Tai. Colorformer: Image colorization via color memory assisted hybrid-attention transformer. In European Conference on Computer Vision, pages 20–36. Springer, 2022. 6, 7, 8, 15 +[22] Xiaoyang Kang, Tao Yang, Wenqi Ouyang, Peiran Ren, Lingzhi Li, and Xuansong Xie. Ddcolor: Towards photorealistic image colorization via dual decoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 328-338, 2023. 6, 7, 8, 12, 15 +[23] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Dynamic stereo: Consistent dynamic depth from stereo videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13229-13239, 2023. 14 +[24] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker3: Simpler and better point tracking by pseudolabelling real videos. arxiv, 2024. 5, 13 +[25] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co + +tracker: It is better to track together. In Proc. ECCV, 2024. 3 +[26] Nikita Karaev, Ignacio Rocco, Ben Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker: It is better to track together. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 1 +[27] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 6 +[28] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 14 +[29] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. 3, 14 +[30] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In Proceedings of the European conference on computer vision (ECCV), pages 170-185, 2018. 3, 6 +[31] Chenyang Lei and Qifeng Chen. Fully automatic video colorization with self-regularization and diversity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3753–3761, 2019. 8, 15 +[32] Hongyang Li, Hao Zhang, Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, and Lei Zhang. Taptr: Tracking any point with transformers as detection. In European Conference on Computer Vision, pages 57-75. Springer, 2025. 3 +[33] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. 14 +[34] Yihao Liu, Hengyuan Zhao, Kelvin CK Chan, Xintao Wang, Chen Change Loy, Yu Qiao, and Chao Dong. Temporally consistent video colorization with deep feature propagation and self-regularization learning. arXiv preprint arXiv:2110.04562, 2021.8, 15 +[35] Ziwei Liu, Raymond A Yeh, Xiaou Tang, Yiming Liu, and Aseem Agarwala. Video frame synthesis using deep voxel flow. In Proceedings of the IEEE international conference on computer vision, pages 4463-4471, 2017. 3 +[36] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 6, 15 +[37] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3202-3211, 2022. 3 + +[38] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4):71-1, 2020. 3, 14 +[39] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4040-4048, 2016. 14 +[40] Daniel Neimark, Omri Bar, Maya Zohar, and Dotan Asselmann. Video transformer network. In Proc. ICCV Workshops, 2021. 2 +[41] E. Palazzolo, J. Behley, P. Lottes, P. Giguère, and C. Stachniss. ReFusion: 3D Reconstruction in Dynamic Environments for RGB-D Cameras Exploiting Residuals. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2019. 14 +[42] Mandela Patrick, Dylan Campbell, Yuki Markus Asano, Ishan Misra, Florian Metze, Christoph Feichtenhofer, Andrea Vedaldi, and João F. Henriques. Keeping your eye on the ball: Trajectory attention in video transformers. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2021. 2, 3 +[43] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 6 +[44] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. UniDepth: universal monocular metric depth estimation. In Proc. CVPR, 2024. 6 +[45] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proc. CVPR, 2016. 6 +[46] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In Proc. ECCV, 2016. 6 +[47] Thomas Schops, Johannes L. Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 14 +[48] Philipp Schröppel, Jan Bechtold, Artemij Amiranashvili, and Thomas Brox. A benchmark and a baseline for robust multiview depth estimation. In 2022 International Conference on 3D Vision (3DV), pages 637-645. IEEE, 2022. 6, 14 +[49] Jiahao Shao, Yuanbo Yang, Hongyu Zhou, Youmin Zhang, Yujun Shen, Matteo Poggi, and Yiyi Liao. Learning temporally consistent video depth from video diffusion priors. arXiv, 2406.01493, 2024. 6, 14 +[50] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.3 + +[51] Jianlin Su, Yu Lu, Shengfeng Pan, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding, 2021. 4 +[52] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 3 +[53] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. 2 +[54] Liao Wang, Qiang Hu, Qihan He, Ziyu Wang, Jingyi Yu, Tinne Tuytelaars, Lan Xu, and Minye Wu. Neural residual radiance fields for streamably free-viewpoint videos. In Proc. CVPR, 2023. 6 +[55] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19795-19806, 2023. 3 +[56] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. DUS3R: Geometric 3D vision made easy. In Proc. CVPR, 2024. 6, 7 +[57] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4909-4916. IEEE, 2020. 14 +[58] Xiaolong Wang, Ross B. Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proc. CVPR, 2018. 2 +[59] Yiran Wang, Min Shi, Jiaqi Li, Zihao Huang, Zhiguo Cao, Jianming Zhang, Ke Xian, and Guosheng Lin. Neural video depth stabilizer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9466-9476, 2023. 6, 14 +[60] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 6, 15 +[61] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 5 +[62] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proc. CVPR, 2024. 5, 6, 12, 14 +[63] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything V2. arXiv, 2406.09414, 2024. 6 +[64] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 14 + +[65] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 14 +[66] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 649-666. Springer, 2016. 6, 7, 8, 15 +[67] Richard Zhang, Jun-Yan Zhu, Phillip Isola, Xinyang Geng, Angela S Lin, Tianhe Yu, and Alexei A Efros. Real-time user-guided image colorization with learned deep priors. arXiv preprint arXiv:1705.02999, 2017. 6, 8, 15 +[68] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19855-19865, 2023. 14 +[69] Xizhou Zhu, Yuwen Xiong, Jifeng Dai, Lu Yuan, and Yichen Wei. Deep feature flow for video recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2349-2358, 2017. 3 \ No newline at end of file diff --git a/2025/Tracktention_ Leveraging Point Tracking to Attend Videos Faster and Better/images.zip b/2025/Tracktention_ Leveraging Point Tracking to Attend Videos Faster and Better/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..12134202f973a2fae0073f314ffc997a09e06fe9 --- /dev/null +++ b/2025/Tracktention_ Leveraging Point Tracking to Attend Videos Faster and Better/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed9c6ffcac161231febee528edd1e1a0d726a833f332d7f2aa70cb60d45e250c +size 890570 diff --git a/2025/Tracktention_ Leveraging Point Tracking to Attend Videos Faster and Better/layout.json b/2025/Tracktention_ Leveraging Point Tracking to Attend Videos Faster and Better/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c09db67d03aca3990ff61867d7b5e89f77860b3c --- /dev/null +++ b/2025/Tracktention_ Leveraging Point Tracking to Attend Videos Faster and Better/layout.json @@ -0,0 +1,9220 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 103, + 541, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 103, + 541, + 121 + ], + "spans": [ + { + "bbox": [ + 69, + 103, + 541, + 121 + ], + "type": "text", + "content": "Tracktention: Leveraging Point Tracking to Attend Videos Faster and Better" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 175, + 143, + 435, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 143, + 435, + 171 + ], + "spans": [ + { + "bbox": [ + 175, + 143, + 435, + 171 + ], + "type": "text", + "content": "Zihang Lai Andrea Vedaldi Visual Geometry Group (VGG), University of Oxford" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 222, + 174, + 384, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 174, + 384, + 185 + ], + "spans": [ + { + "bbox": [ + 222, + 174, + 384, + 185 + ], + "type": "text", + "content": "{zlai,vedaldi}@robots.ox.ac.uk" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 60, + 213, + 116, + 327 + ], + "blocks": [ + { + "bbox": [ + 60, + 213, + 116, + 327 + ], + "lines": [ + { + "bbox": [ + 60, + 213, + 116, + 327 + ], + "spans": [ + { + "bbox": [ + 60, + 213, + 116, + 327 + ], + "type": "image", + "image_path": "61729fc2a75a3e95485e8cb505c4548e0081624ed3e140e55dbf2408538c8dab.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 329, + 555, + 373 + ], + "lines": [ + { + "bbox": [ + 55, + 329, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 555, + 373 + ], + "type": "text", + "content": "Figure 1. Left: The Tracktention Layer is a plug-and-play module that can convert an image-based network (e.g., for monocular depth prediction) into a state-of-the-art video network (e.g., for video depth prediction). It does so by integrating the output of any off-the-shelf, modern, and powerful point trackers via track cross-attention. Right: For example, Tracktention achieves state-of-the-art and efficient video depth prediction by transforming Depth Anything into a video depth model. See Tab. 2 for detailed results. *Single-image models." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 117, + 213, + 378, + 326 + ], + "blocks": [ + { + "bbox": [ + 117, + 213, + 378, + 326 + ], + "lines": [ + { + "bbox": [ + 117, + 213, + 378, + 326 + ], + "spans": [ + { + "bbox": [ + 117, + 213, + 378, + 326 + ], + "type": "image", + "image_path": "86a95c21ef9b45cd361f1edd74c0707a424ef5eea5ebf14f86c8f43a42c4ba0a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 383, + 215, + 548, + 327 + ], + "blocks": [ + { + "bbox": [ + 383, + 215, + 548, + 327 + ], + "lines": [ + { + "bbox": [ + 383, + 215, + 548, + 327 + ], + "spans": [ + { + "bbox": [ + 383, + 215, + 548, + 327 + ], + "type": "image", + "image_path": "07323bda3f0f1e56205a472b8a38445d51bf2a43b7ac149e4ab40573663087f5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 152, + 403, + 200, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 403, + 200, + 415 + ], + "spans": [ + { + "bbox": [ + 152, + 403, + 200, + 415 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 437, + 297, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 437, + 297, + 700 + ], + "spans": [ + { + "bbox": [ + 53, + 437, + 297, + 700 + ], + "type": "text", + "content": "Temporal consistency is critical in video prediction to ensure that outputs are coherent and free of artifacts. Traditional methods, such as temporal attention and 3D convolution, may struggle with significant object motion and may not capture long-range temporal dependencies in dynamic scenes. To address this gap, we propose the Tracktention Layer, a novel architectural component that explicitly integrates motion information using point tracks, i.e., sequences of corresponding points across frames. By incorporating these motion cues, the Tracktention Layer enhances temporal alignment and effectively handles complex object motions, maintaining consistent feature representations over time. Our approach is computationally efficient and can be seamlessly integrated into existing models, such as Vision Transformers, with minimal modification. It can be used to upgrade image-only models to state-of-the-art video ones, sometimes outperforming models natively designed for video prediction. We demonstrate this on video depth prediction and video colorization, where models augmented with the Tracktention Layer exhibit significantly improved temporal consistency compared to baselines. Project website: zlai0.github.io/TrackTention." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 403, + 394, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 403, + 394, + 415 + ], + "spans": [ + { + "bbox": [ + 315, + 403, + 394, + 415 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 424, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 555, + 544 + ], + "type": "text", + "content": "Compared to image analysis tasks, video analysis tasks, such as video segmentation, video depth estimation, and video colorization, pose additional challenges due to the temporal dimension inherent in video data. A particularly important challenge is ensuring that outputs are temporally consistent across frames, which is necessary to produce coherent and artifact-free results. Temporal inconsistencies, such as flickering or abrupt changes in predicted object attributes, are especially noticeable in applications like colorization, where the output is a new video." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 545, + 556, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 664 + ], + "type": "text", + "content": "An aspect of video analysis that has seen significant progress in recent years is point tracking. New trackers such as PIPs [19], TAPIR [12], BootsTAP [13], and CoTracker [26] can now track quasi-dense collections of points across long video sequences with high reliability and efficiency. So far, these trackers have been primarily used in selected applications like 3D reconstruction and object tracking through ad-hoc algorithms. In this paper, we ask whether progress in point tracking can benefit a much wider range of video analysis tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "To answer this question, we propose the Tracktention Layer, a novel architectural component designed to enhance temporal consistency in video transformers. Tracktention integrates seamlessly into vision transformers as a new at" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22809" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 68, + 527, + 190 + ], + "blocks": [ + { + "bbox": [ + 82, + 68, + 527, + 190 + ], + "lines": [ + { + "bbox": [ + 82, + 68, + 527, + 190 + ], + "spans": [ + { + "bbox": [ + 82, + 68, + 527, + 190 + ], + "type": "image", + "image_path": "1df24ea7727c5d35d33c2a108f07d3ff05cd5ed41f342ba90eb524ebb2d46ce8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 192, + 555, + 248 + ], + "lines": [ + { + "bbox": [ + 54, + 192, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 54, + 192, + 555, + 248 + ], + "type": "text", + "content": "Figure 2. Overview of Tracktention. We begin by using an off-the-shelf point tracker to extract a number of video tracks. Given these, we first sample image tokens at the track locations, obtaining corresponding track tokens (Sec. 3.1). Next, we use a Track Transformer to update these tokens, propagating information temporally at consistent spatial locations (Sec. 3.2). Finally, we splat the information back to the image tokens (Sec. 3.3). By explicitly incorporating motion information through point tracks, Tracktention improves temporal alignment, effectively captures complex object movements, and ensures stable feature representations over time." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 258, + 294, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 258, + 294, + 378 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 294, + 378 + ], + "type": "text", + "content": "tention layer, making it compatible with a wide range of existing models. The key innovation of Tracktention lies in leveraging pre-extracted point tracks to inform the model about temporal correspondences between image tokens in different frames (Fig. 2). This is achieved by sampling the image tokens at the track points. The information is then propagated along the tracks using a transformer layer, which justifies the name \"Tracktention\". Finally, the information is splatted back to the image tokens, allowing the computation to progress as normal." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 378, + 295, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 295, + 450 + ], + "type": "text", + "content": "By explicitly incorporating tracking information, the model can establish direct correspondences, enabling precise temporal alignment regardless of how much objects move. This explicit temporal alignment allows the model to maintain consistent feature representations over time, smoothing features, and reducing temporal artifacts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 450, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 594 + ], + "type": "text", + "content": "The problem of propagating information in video neural networks has been extensively studied in the literature. Many approaches are implicit, in the sense that they do not explicitly estimate or account for the motion observed in the scene. Examples include 3D convolution [6, 53] and space-time attention [1, 3, 14, 40, 58]. However, to maintain manageable computational costs, these methods often reduce the spatial resolution of features and limit the temporal range they can model—either by using small convolution kernels or by decoupling spatial and temporal attention. The resulting networks may struggle to represent motion precisely and to keep up with large object displacements." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "text", + "content": "In contrast, Tracktention explicitly accounts for motion at the resolution of point tracks, which is much finer than that of typical image features. By pooling and splatting information at image locations that are already put in correspondence by the tracker, it only needs to propagate information along the tracks, obviating the need to implicitly establish such correspondences via space-time attention." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "There are also approaches to video analysis that, like Tracktention, incorporate motion information explicitly. They often work by feeding optical flow to the network," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 258, + 553, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 553, + 330 + ], + "type": "text", + "content": "capturing pixel-level motion between frames [42]. While optical flow can be effective in certain contexts, it may struggle with occlusions or large displacements. Building on the power of point trackers, Tracktention can handle occlusions, large displacements, and long temporal dependencies more effectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 331, + 554, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 554, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 554, + 488 + ], + "type": "text", + "content": "In summary, Tracktention offers several advantages over traditional video network architectures: (1) It leverages the power of modern point tracker models, which are excellent motion experts, and provides a general mechanism to inject this knowledge into any video transformer. In particular, it can handle complex and large object motions, limited only by the capabilities of the point trackers. (2) It directly establishes long-term space-time correspondences, unlike 3D convolutions that are local and unlike separated space-time attention, which can only do so indirectly. (3) It avoids the need to repeatedly calculate such correspondences explicitly, as full spatio-temporal attention does, achieving much greater efficiency and operating at a much finer resolution." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 487, + 554, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 554, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 554, + 596 + ], + "type": "text", + "content": "We deliver Tracktention as a plugin layer (Fig. 1) that can be added to single-image neural network architectures to extend them to powerful, temporally-aware video models. For example, we demonstrate that a pre-trained single-image depth predictor can be extended to a corresponding state-of-the-art video depth predictor by integrating our Tracktention Layer. We achieve similar results for automatic video colorization, outperforming models natively designed to process videos while also being more efficient." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 609, + 402, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 609, + 402, + 621 + ], + "spans": [ + { + "bbox": [ + 314, + 609, + 402, + 621 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": "Temporal Modeling for Videos. Temporal modeling is a cornerstone of video understanding tasks. Over the years, various techniques have been proposed to capture temporal dependencies in video data. 3D CNNs [6, 16, 53] extend 2D convolutions into the temporal dimension, allowing for direct spatiotemporal feature extraction from video data. Temporal attention mechanisms provide a new approach to" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22810" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": "effectively capture long-range dependencies across frames. Video transformers [1, 3, 15, 37, 42] apply self-attention mechanisms over temporal and spatial dimensions, typically separately or hierarchically, to reduce computational complexity while modeling temporal dependencies effectively. Recent video processing models [4, 5, 50] also use a hybrid of these approaches. However, these methods have notable limitations: temporal attention struggles with rapid object movements, while 3D convolutions assume local spatiotemporal correlations, failing with unpredictable motion. In this work, we explicitly leverage tracking information to precisely align temporal features regardless of object movement." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 231, + 296, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 296, + 447 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 296, + 447 + ], + "type": "text", + "content": "Consistent Video Prediction. Addressing temporal inconsistency is critical for ensuring smooth and coherent video outputs. Optical flow-based methods [69] enforce temporal coherence by aligning features or predictions across frames, propagating features along motion paths estimated by optical flow. Liu et al. [35] utilize convolutional LSTMs to capture temporal information for consistent video frame synthesis, modeling temporal dependencies through recurrent connections. Lai et al. [30] propose a post-processing model with a temporal consistency objective for video prediction. Similarly, Luo et al. [38] and Kopf et al. [29] present test-time training post-processing methods specifically designed to stabilize video depth predictions using geometric consistency cues. In contrast, our work aims to stabilize features by allowing them to attend to corresponding areas across time based on point tracks. This approach inherently accounts for object movement and embeds temporal alignment efficiently within the model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 449, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 664 + ], + "type": "text", + "content": "Point Tracking. Point tracking is the task of identifying and continuously tracking specific points across frames in a video. PIPs [18] pioneered a neural network-based approach that updates tracked points iteratively by extracting correlation maps between frames, inspired by the RAFT model for optical flow [52]. TAP-Vid [10] proposed a comprehensive benchmark and TAP-Net, a new model for point tracking. TAPIR [11] further improved performance by combining global matching capabilities with PIPs, enhancing the accuracy of tracked points even in complex motion scenarios. OmniMotion [55] tracks points by leveraging a quasi-3D canonical volume but is approximately " + }, + { + "bbox": [ + 55, + 449, + 296, + 664 + ], + "type": "inline_equation", + "content": "10^{3}" + }, + { + "bbox": [ + 55, + 449, + 296, + 664 + ], + "type": "text", + "content": " times slower than CoTracker due to its test-time optimization. CoTracker [25] leveraged a transformer-based architecture to track multiple points jointly, yielding improvements particularly for occluded points. Recent methods such as LocoTrack [7], which introduced 4D correlation features, and TAPTR [32] further improved precision." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "In this work, we do not propose new models for point tracking. Instead, we use existing trackers to establish direct token correspondences, ensuring precise temporal alignment despite object motion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 71, + 370, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 370, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 370, + 83 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 91, + 553, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 163 + ], + "type": "text", + "content": "In this section, we introduce Tracktention, a novel transformer layer designed to aggregate and redistribute feature information in video transformers in a motion-aware manner (Fig. 3). By leveraging motion information captured by the output of a point tracker, this layer improves the temporal consistency and motion awareness of video features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 163, + 553, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 553, + 236 + ], + "type": "text", + "content": "Tracktention comprises an Attentional Sampling block (Sec. 3.1) that pools information from the video features into the tracks, a Track Transformer (Sec. 3.2) that updates the track tokens, and an Attentional Splatting block (Sec. 3.3) that pushes the updated track tokens back to the video features. We describe these components next." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 242, + 437, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 437, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 437, + 255 + ], + "type": "text", + "content": "3.1. Attentional Sampling" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 260, + 553, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 553, + 295 + ], + "type": "text", + "content": "The Attentional Sampling module is the first stage in Track-attention. It uses cross-attention to pool information from the video features into the track tokens." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 296, + 554, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 296, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 554, + 415 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 296, + 554, + 415 + ], + "type": "inline_equation", + "content": "F\\in \\mathbb{R}^{T\\times HW\\times D_f}" + }, + { + "bbox": [ + 313, + 296, + 554, + 415 + ], + "type": "text", + "content": " denote the input feature map, where " + }, + { + "bbox": [ + 313, + 296, + 554, + 415 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 296, + 554, + 415 + ], + "type": "text", + "content": " is the number of time steps, " + }, + { + "bbox": [ + 313, + 296, + 554, + 415 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 296, + 554, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 296, + 554, + 415 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 296, + 554, + 415 + ], + "type": "text", + "content": " are the spatial dimensions, and " + }, + { + "bbox": [ + 313, + 296, + 554, + 415 + ], + "type": "inline_equation", + "content": "D_{f}" + }, + { + "bbox": [ + 313, + 296, + 554, + 415 + ], + "type": "text", + "content": " is the feature dimension. Let " + }, + { + "bbox": [ + 313, + 296, + 554, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{P}\\in \\mathbb{R}^{T\\times M\\times 2}" + }, + { + "bbox": [ + 313, + 296, + 554, + 415 + ], + "type": "text", + "content": " represent a set of " + }, + { + "bbox": [ + 313, + 296, + 554, + 415 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 296, + 554, + 415 + ], + "type": "text", + "content": " tracks, each of which is a sequence of 2D points. We first apply a positional embedding to these 2D points, obtaining one token per track and per frame, and arrange these tokens into a tensor " + }, + { + "bbox": [ + 313, + 296, + 554, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{T}\\in \\mathbb{R}^{T\\times M\\times D_f}" + }, + { + "bbox": [ + 313, + 296, + 554, + 415 + ], + "type": "text", + "content": ". We then project the track tokens into queries " + }, + { + "bbox": [ + 313, + 296, + 554, + 415 + ], + "type": "inline_equation", + "content": "Q\\in \\mathbb{R}^{T\\times M\\times D_k}" + }, + { + "bbox": [ + 313, + 296, + 554, + 415 + ], + "type": "text", + "content": " and the feature map into keys " + }, + { + "bbox": [ + 313, + 296, + 554, + 415 + ], + "type": "inline_equation", + "content": "K\\in \\mathbb{R}^{T\\times HW\\times D_k}" + }, + { + "bbox": [ + 313, + 296, + 554, + 415 + ], + "type": "text", + "content": " and values " + }, + { + "bbox": [ + 313, + 296, + 554, + 415 + ], + "type": "inline_equation", + "content": "V\\in \\mathbb{R}^{T\\times HW\\times D_f}" + }, + { + "bbox": [ + 313, + 296, + 554, + 415 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 349, + 423, + 518, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 423, + 518, + 436 + ], + "spans": [ + { + "bbox": [ + 349, + 423, + 518, + 436 + ], + "type": "interline_equation", + "content": "Q _ {t} = \\mathcal {T} _ {t} W _ {Q}, \\quad K _ {t} = F _ {t} W _ {K}, \\quad V _ {t} = F _ {t},", + "image_path": "4b017e006b47c560644e1486462e4446b7fdfa3677a9314ca0c8d4494d1c1645.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 443, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 554, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 443, + 554, + 491 + ], + "type": "inline_equation", + "content": "W_{Q}" + }, + { + "bbox": [ + 313, + 443, + 554, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 443, + 554, + 491 + ], + "type": "inline_equation", + "content": "W_{K}" + }, + { + "bbox": [ + 313, + 443, + 554, + 491 + ], + "type": "text", + "content": " are learnable weight matrices, and the subscript " + }, + { + "bbox": [ + 313, + 443, + 554, + 491 + ], + "type": "inline_equation", + "content": "t \\in \\{1, \\dots, T\\}" + }, + { + "bbox": [ + 313, + 443, + 554, + 491 + ], + "type": "text", + "content": " indexes the first dimensions of the tensors, extracting matrices. Note that we avoid projecting the values to preserve the original features during sampling." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 491, + 553, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 553, + 525 + ], + "type": "text", + "content": "For each time " + }, + { + "bbox": [ + 313, + 491, + 553, + 525 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 491, + 553, + 525 + ], + "type": "text", + "content": ", the attention weights " + }, + { + "bbox": [ + 313, + 491, + 553, + 525 + ], + "type": "inline_equation", + "content": "A_{t} \\in \\mathbb{R}^{M \\times HW}" + }, + { + "bbox": [ + 313, + 491, + 553, + 525 + ], + "type": "text", + "content": " are computed using the scaled dot-product attention mechanism:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 366, + 524, + 553, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 524, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 366, + 524, + 553, + 552 + ], + "type": "interline_equation", + "content": "A _ {t} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {t} K _ {t} ^ {\\top}}{\\sqrt {D _ {k}}} + B _ {t}\\right), \\tag {1}", + "image_path": "6dd1f58f1cdd6ae19573823d680e7325ca50abf6ff0dad4c6e462ae9dd3edcc1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 554, + 553, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 553, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 553, + 579 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 554, + 553, + 579 + ], + "type": "inline_equation", + "content": "B_{t}\\in \\mathbb{R}^{M\\times HW}" + }, + { + "bbox": [ + 313, + 554, + 553, + 579 + ], + "type": "text", + "content": " is a bias term that encourages attention to align with the tracks. It is defined as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 373, + 586, + 495, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 586, + 495, + 611 + ], + "spans": [ + { + "bbox": [ + 373, + 586, + 495, + 611 + ], + "type": "interline_equation", + "content": "B _ {t i j} = - \\frac {\\| P _ {t i} - \\mathrm {p o s} _ {F _ {t}} (j) \\| ^ {2}}{2 \\sigma^ {2}},", + "image_path": "2b32e353e8008248bbfda261b6989a94dce2f0d77ce6401ebe669995640d2393.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 617, + 553, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 553, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 617, + 553, + 654 + ], + "type": "inline_equation", + "content": "i\\in \\{1,\\dots ,M\\}" + }, + { + "bbox": [ + 313, + 617, + 553, + 654 + ], + "type": "inline_equation", + "content": "j\\in \\{1,\\ldots ,HW\\}" + }, + { + "bbox": [ + 313, + 617, + 553, + 654 + ], + "type": "inline_equation", + "content": "P_{ti}\\in \\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 617, + 553, + 654 + ], + "type": "text", + "content": " is the spatial location of track " + }, + { + "bbox": [ + 313, + 617, + 553, + 654 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 617, + 553, + 654 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 313, + 617, + 553, + 654 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 617, + 553, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 617, + 553, + 654 + ], + "type": "inline_equation", + "content": "\\mathrm{pos}_{F_t}(j)\\in \\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 617, + 553, + 654 + ], + "type": "text", + "content": " is the spatial location of the " + }, + { + "bbox": [ + 313, + 617, + 553, + 654 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 617, + 553, + 654 + ], + "type": "text", + "content": " -th feature map token at time " + }, + { + "bbox": [ + 313, + 617, + 553, + 654 + ], + "type": "inline_equation", + "content": "t" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "In the softmax operator of Eq. (1), this bias is analogous to multiplying the attention map by a Gaussian window that decays exponentially as the feature position deviates from the track position. The parameter " + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": ", which we set to " + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "inline_equation", + "content": "1/2" + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": ", determines how fast the window decays." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22811" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 88, + 408, + 262 + ], + "blocks": [ + { + "bbox": [ + 69, + 77, + 156, + 86 + ], + "lines": [ + { + "bbox": [ + 69, + 77, + 156, + 86 + ], + "spans": [ + { + "bbox": [ + 69, + 77, + 156, + 86 + ], + "type": "text", + "content": "(A) Tracktention Architecture" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 88, + 408, + 262 + ], + "lines": [ + { + "bbox": [ + 65, + 88, + 408, + 262 + ], + "spans": [ + { + "bbox": [ + 65, + 88, + 408, + 262 + ], + "type": "image", + "image_path": "9dd77097f4f8489e4bd92f770a6f3256a0bfae7209f1718b88d3509302743314.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 410, + 76, + 547, + 262 + ], + "blocks": [ + { + "bbox": [ + 410, + 76, + 547, + 262 + ], + "lines": [ + { + "bbox": [ + 410, + 76, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 410, + 76, + 547, + 262 + ], + "type": "image", + "image_path": "059f4f5464bd5a27fc712040548c59b8177ff89f3b912a192b93f8157f26c20e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 265, + 553, + 299 + ], + "lines": [ + { + "bbox": [ + 55, + 265, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 553, + 299 + ], + "type": "text", + "content": "Figure 3. Left: the Tracktention architecture comprises Attentional Sampling, pooling information from images to track, Track Transformer, processing this information temporally, and Attentional Splatting, moving the processed information back to the images. Right: Tracktention is easily integrated in ViTs and ConvNets to make video networks out image ones." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 309, + 295, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 309, + 295, + 417 + ], + "spans": [ + { + "bbox": [ + 54, + 309, + 295, + 417 + ], + "type": "text", + "content": "We use the attention to \"sample\" track features " + }, + { + "bbox": [ + 54, + 309, + 295, + 417 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{T \\times M \\times D_f}" + }, + { + "bbox": [ + 54, + 309, + 295, + 417 + ], + "type": "text", + "content": " by setting " + }, + { + "bbox": [ + 54, + 309, + 295, + 417 + ], + "type": "inline_equation", + "content": "S_t = A_t V_t = A_t F_t \\in \\mathbb{R}^{M \\times D_f}" + }, + { + "bbox": [ + 54, + 309, + 295, + 417 + ], + "type": "text", + "content": " for each time " + }, + { + "bbox": [ + 54, + 309, + 295, + 417 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 309, + 295, + 417 + ], + "type": "text", + "content": ". Alternatively, we could have obtained " + }, + { + "bbox": [ + 54, + 309, + 295, + 417 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 54, + 309, + 295, + 417 + ], + "type": "text", + "content": " by sampling the features " + }, + { + "bbox": [ + 54, + 309, + 295, + 417 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 54, + 309, + 295, + 417 + ], + "type": "text", + "content": " at the given track locations, for instance, by using bilinear sampling. This can also be expressed as " + }, + { + "bbox": [ + 54, + 309, + 295, + 417 + ], + "type": "inline_equation", + "content": "A_t F_t" + }, + { + "bbox": [ + 54, + 309, + 295, + 417 + ], + "type": "text", + "content": " for a specific matrix " + }, + { + "bbox": [ + 54, + 309, + 295, + 417 + ], + "type": "inline_equation", + "content": "A_t" + }, + { + "bbox": [ + 54, + 309, + 295, + 417 + ], + "type": "text", + "content": ", so sampling is a special case of our Attentional Sampling block. The advantage of Attentional Sampling is that it allows the model to learn a better sampling strategy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 419, + 296, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 502 + ], + "type": "text", + "content": "Relative Spatial Information. To capture the relative spatial relationships between track tokens and feature map locations, we incorporate Rotational Position Encodings (RoPE) [51] into the keys and values derived from the feature map. RoPE is a form of positional encoding that allows attention mechanisms to be aware of the relative positions of points defined in a continuous space." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "type": "text", + "content": "To compute the RoPE embedding " + }, + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "type": "inline_equation", + "content": "\\mathrm{RoPE}(\\mathbf{f})" + }, + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "type": "text", + "content": " of a feature vector " + }, + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "type": "text", + "content": " at spatial location " + }, + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "type": "text", + "content": ", we use the first " + }, + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "type": "inline_equation", + "content": "d / 2" + }, + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "type": "text", + "content": " channels to encode the " + }, + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "type": "text", + "content": " position. For " + }, + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "type": "inline_equation", + "content": "i = 1,\\dots ,d / 4" + }, + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "type": "text", + "content": ", we define:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 560, + 293, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 560, + 293, + 592 + ], + "spans": [ + { + "bbox": [ + 57, + 560, + 293, + 592 + ], + "type": "interline_equation", + "content": "\\begin{array}{r} \\mathrm {R o P E} (\\mathbf {f}) [ 2 i ] = \\mathbf {f} [ 2 i ] \\cdot \\cos (\\theta_ {i} x) - \\mathbf {f} [ 2 i + 1 ] \\cdot \\sin (\\theta_ {i} x), \\\\ \\mathrm {R o P E} (\\mathbf {f}) [ 2 i + 1 ] = \\mathbf {f} [ 2 i ] \\cdot \\sin (\\theta_ {i} x) + \\mathbf {f} [ 2 i + 1 ] \\cdot \\cos (\\theta_ {i} x), \\end{array}", + "image_path": "aae3e6cefa226b4530163a7a4b02a0682fe25c48a89953d9233162fa0d7b3b5c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 599, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 295, + 639 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 599, + 295, + 639 + ], + "type": "inline_equation", + "content": "\\theta_{i} = 100^{\\frac{-2(i - 1)}{d / 2}}" + }, + { + "bbox": [ + 55, + 599, + 295, + 639 + ], + "type": "text", + "content": ". The second " + }, + { + "bbox": [ + 55, + 599, + 295, + 639 + ], + "type": "inline_equation", + "content": "d / 2" + }, + { + "bbox": [ + 55, + 599, + 295, + 639 + ], + "type": "text", + "content": " channels encode the " + }, + { + "bbox": [ + 55, + 599, + 295, + 639 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 599, + 295, + 639 + ], + "type": "text", + "content": " position analogously by replacing " + }, + { + "bbox": [ + 55, + 599, + 295, + 639 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 599, + 295, + 639 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 599, + 295, + 639 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 599, + 295, + 639 + ], + "type": "text", + "content": " in the above equations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Implementation Details. To stabilize the training process, especially due to the difference in scale between the track token embeddings and the feature map embeddings (which tend to have larger norms), we apply QK-normalization [9] to the queries and keys before computing the attention weights. This normalization ensures that the scale of the" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 309, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 555, + 357 + ], + "type": "text", + "content": "embeddings is consistent, preventing issues related to gradient instability. We also use " + }, + { + "bbox": [ + 313, + 309, + 555, + 357 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 309, + 555, + 357 + ], + "type": "text", + "content": " attention heads, allowing the model to capture diverse contextual relationships across different subspaces." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 367, + 427, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 427, + 379 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 427, + 379 + ], + "type": "text", + "content": "3.2. Track Transformer" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 386, + 555, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 555, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 555, + 457 + ], + "type": "text", + "content": "Once the track tokens " + }, + { + "bbox": [ + 313, + 386, + 555, + 457 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 386, + 555, + 457 + ], + "type": "text", + "content": " are obtained, the Track Transformer processes them to propagate information along the tracks, making them more temporally consistent and smoothing out any irregularities. It is designed as a transformer that operates along the time dimension of the point tracks and outputs updated track tokens " + }, + { + "bbox": [ + 313, + 386, + 555, + 457 + ], + "type": "inline_equation", + "content": "\\mathcal{T}'" + }, + { + "bbox": [ + 313, + 386, + 555, + 457 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 459, + 554, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 554, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 554, + 542 + ], + "type": "text", + "content": "In more detail, recall that the track tokens " + }, + { + "bbox": [ + 313, + 459, + 554, + 542 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 459, + 554, + 542 + ], + "type": "text", + "content": " are of shape " + }, + { + "bbox": [ + 313, + 459, + 554, + 542 + ], + "type": "inline_equation", + "content": "T \\times M \\times D_f" + }, + { + "bbox": [ + 313, + 459, + 554, + 542 + ], + "type": "text", + "content": ". We swap the first two dimensions to obtain a tensor of shape " + }, + { + "bbox": [ + 313, + 459, + 554, + 542 + ], + "type": "inline_equation", + "content": "M \\times T \\times D_f" + }, + { + "bbox": [ + 313, + 459, + 554, + 542 + ], + "type": "text", + "content": ". Then, we apply a transformer block, interpreting the first dimension " + }, + { + "bbox": [ + 313, + 459, + 554, + 542 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 459, + 554, + 542 + ], + "type": "text", + "content": " as the batch dimension. The effect is to apply attention along the temporal dimension, allowing the model to learn temporal dependencies along each track." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 544, + 554, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 651 + ], + "type": "text", + "content": "This approach ensures that information is not exchanged across tracks, meaning each track's temporal sequence is processed independently. We considered a variant of the Track Transformer that included attention across tracks as well, but found that this approach was slightly worse and more computationally expensive. Our interpretation is that information is already exchanged spatially by the Vision Transformer paired with Tracktention, making additional exchange of information across locations redundant." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Architecture. For each track, we utilize a 2-layer transformer encoder along the temporal dimension, structured as a standard transformer model. Each layer consists of Multi-Head Self-Attention, which allows each time step within a track to attend to all other time steps within that same track," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22812" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": "and Position-Wise Feed-Forward Networks, which apply nonlinear transformations independently to each time step, enhancing the model's representational capacity. We also incorporate sinusoidal positional encodings to provide the model with a sense of temporal order. These encodings are added to the input features prior to the transformer layers, ensuring the model can distinguish between different time steps and understand the temporal structure of the sequence." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 175, + 178, + 188 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 178, + 188 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 178, + 188 + ], + "type": "text", + "content": "3.3. Attentional Splatting" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 193, + 295, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 193, + 295, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 295, + 324 + ], + "type": "text", + "content": "The Attentional Splitting module maps the updated track tokens " + }, + { + "bbox": [ + 55, + 193, + 295, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{T}'" + }, + { + "bbox": [ + 55, + 193, + 295, + 324 + ], + "type": "text", + "content": " back onto the feature maps " + }, + { + "bbox": [ + 55, + 193, + 295, + 324 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 193, + 295, + 324 + ], + "type": "text", + "content": ". This is achieved by reversing the roles of the track tokens and the feature tokens in the various expressions in Sec. 3.1. This means the queries are derived from the grid coordinates of the output feature map, while the keys and values are generated from the track tokens. The same bias term " + }, + { + "bbox": [ + 55, + 193, + 295, + 324 + ], + "type": "inline_equation", + "content": "B_{t}" + }, + { + "bbox": [ + 55, + 193, + 295, + 324 + ], + "type": "text", + "content": " is used as before (but now the matrix is transposed), as well as the same " + }, + { + "bbox": [ + 55, + 193, + 295, + 324 + ], + "type": "inline_equation", + "content": "QK" + }, + { + "bbox": [ + 55, + 193, + 295, + 324 + ], + "type": "text", + "content": "-normalization and RoPE encodings, this time obtaining a matrix " + }, + { + "bbox": [ + 55, + 193, + 295, + 324 + ], + "type": "inline_equation", + "content": "A_{t}^{\\prime} \\in \\mathbb{R}^{HW \\times M}" + }, + { + "bbox": [ + 55, + 193, + 295, + 324 + ], + "type": "text", + "content": ". The final output Tracktention " + }, + { + "bbox": [ + 55, + 193, + 295, + 324 + ], + "type": "inline_equation", + "content": "(F)" + }, + { + "bbox": [ + 55, + 193, + 295, + 324 + ], + "type": "text", + "content": " is then computed as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 101, + 331, + 250, + 345 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 331, + 250, + 345 + ], + "spans": [ + { + "bbox": [ + 101, + 331, + 250, + 345 + ], + "type": "interline_equation", + "content": "\\left[ \\operatorname {T r a c k t e n t i o n} (F) \\right] _ {t} = W _ {\\text {o u t}} A _ {t} ^ {\\prime} \\left[ \\mathcal {T} ^ {\\prime} \\right] _ {t},", + "image_path": "5282c790c401398e31d2f1b317527dc38d78c31d0ba80277508b0a7fbdc04939.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 352, + 295, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 295, + 389 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 352, + 295, + 389 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{out}}" + }, + { + "bbox": [ + 55, + 352, + 295, + 389 + ], + "type": "text", + "content": " is a final output projection. By adopting a symmetric design for Attentional Sampling and Splitting, we ensure that information is handled in a consistent manner." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 395, + 211, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 395, + 211, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 211, + 406 + ], + "type": "text", + "content": "3.4. Point Tracker Pre-Processor" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 413, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 295, + 533 + ], + "type": "text", + "content": "To apply the Tracktention layer to a video, we first need to extract the point tracks " + }, + { + "bbox": [ + 55, + 413, + 295, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 55, + 413, + 295, + 533 + ], + "type": "text", + "content": ". We use CoTracker3 [24] due to its robustness in handling complex motions and occlusions in videos. Trackers require a seed point for each track, called a query. We sample 576 points uniformly at random from the spatio-temporal video volume and track each forward and backward in time. Compared to initializing tracks on a grid in the first frame of the video, this simple initialization scheme encourages tracks to cover the video well, despite camera and object motion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 540, + 253, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 253, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 253, + 552 + ], + "type": "text", + "content": "3.5. Integration into Image-Based Models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 557, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 677 + ], + "type": "text", + "content": "A key application of Tracktention is to build video neural networks from image-based ones. Integrating Tracktention into existing image-based models, such as a ViT or ConvNet, is straightforward. For example, monocular depth estimation models like Depth Anything, which are designed for single images, may lack temporal consistency when applied to video frame by frame. Adding Tracktention layers after each transformer block and fine-tuning the model on videos can increase temporal stability and significantly improve performance in video-based applications." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "To incorporate Tracktention into an existing network backbone, we insert it immediately after all or some transformer or convolutional blocks. We also include a residual" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "content": "connection to preserve the original information flow, setting " + }, + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "inline_equation", + "content": "F' = F + \\text{Tracktention}(F)" + }, + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "content": " represents the output feature map from the preceding block, and " + }, + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "inline_equation", + "content": "F'" + }, + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "content": " is the updated feature map after the Tracktention module." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 120, + 555, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 120, + 555, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 120, + 555, + 204 + ], + "type": "text", + "content": "Furthermore, the Tracktention output projection, " + }, + { + "bbox": [ + 313, + 120, + 555, + 204 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{out}}" + }, + { + "bbox": [ + 313, + 120, + 555, + 204 + ], + "type": "text", + "content": ", is initialized to zero, preserving the original network output at training onset. This strategy retains benefits of pre-trained weights while allowing the model to gradually adapt to temporal updates. Consequently, Tracktention modules capture fine spatial details and ensure consistent temporal modeling across layers." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 214, + 395, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 214, + 395, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 395, + 227 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 234, + 554, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 554, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 554, + 270 + ], + "type": "text", + "content": "We show that Tracktention can turn image-based models into state-of-the-art ones for video depth estimation and colorization (Secs. 4.1 and 4.2) and ablate its design (Sec. 4.3)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 277, + 449, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 449, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 449, + 289 + ], + "type": "text", + "content": "4.1. Video Depth Estimation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 296, + 554, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 296, + 554, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 554, + 366 + ], + "type": "text", + "content": "To apply Tracktention to video depth estimation, we start from an image-based model, Depth Anything [61], and upgrade it to a video-based model using the Tracktention module. The Depth Anything model has a DINO ViT backbone consisting of 12 transformer blocks. We insert our Tracktention module after each of the last 6 blocks." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 367, + 554, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 554, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 554, + 462 + ], + "type": "text", + "content": "We fine-tune the model on a large dataset of videos, using a combination of synthetic and real data. This data comes with ground-truth depth, which we fit using a scale-and shift-invariant loss, as Depth Anything predicts depth up to an affine transformation. However, we share the same calibration parameters for all frames in a video, as calibrating frames independently would mask temporal inconsistencies, which we want the model to learn to correct." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 463, + 554, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 554, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 554, + 559 + ], + "type": "text", + "content": "To reduce the number of new parameters added to the model, we share them between all Tracktention modules. During training, the original model is kept frozen, and only the Tracktention modules are updated. We use the AdamW optimizer with an initial learning rate of " + }, + { + "bbox": [ + 313, + 463, + 554, + 559 + ], + "type": "inline_equation", + "content": "1.6 \\times 10^{-5}" + }, + { + "bbox": [ + 313, + 463, + 554, + 559 + ], + "type": "text", + "content": " and cosine learning rate decay. Training is conducted for 4 epochs with a batch size of 4 videos, each containing a randomly sampled number of frames between 8 and 16." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 560, + 554, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 554, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 554, + 609 + ], + "type": "text", + "content": "Data and evaluation metrics. For training, we use a combination of datasets containing both synthetic and real videos. We use standard depth estimation metrics [62]. Please refer to Appendix A.3.1 for details." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 623, + 435, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 435, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 435, + 635 + ], + "type": "text", + "content": "4.1.1 Quantitative Results" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": "In Tab. 1, we evaluate our model on four video collections in the DepthCrafter benchmark [20] against several alternatives. Tracktention improves the image-based model it augments, i.e., Depth Anything [62], substantially: Sintel AbsRel improves by " + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "inline_equation", + "content": "9.2\\%" + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\delta_{1}" + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "inline_equation", + "content": "13.5\\%" + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": ", KITTI by " + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "inline_equation", + "content": "26.8\\%" + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "inline_equation", + "content": "12.5\\%" + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": ", Bonn by " + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "inline_equation", + "content": "15.4\\%" + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": ", and on average by" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22813" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 188 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 188 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 188 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 188 + ], + "type": "table", + "html": "
MethodType#ParamsSintel (~50 frames)Scannet (90 frames)KITTI (110 frames)Bonn (110 frames)Average
AbsRel ↓δ1.25 ↑AbsRel ↓δ1.25 ↑AbsRel ↓δ1.25 ↑AbsRel ↓δ1.25 ↑AbsRel ↓δ1.25 ↑
DUS3R [56]Vid.578M0.6280.3930.1940.6940.2920.4560.2500.5880.3410.533
NVDS [59]Vid.430M0.4080.4830.1870.6770.2530.5880.1670.7660.2540.629
ChronoDepth [49]Vid.1521M0.5870.4860.1590.7830.1670.7590.1000.9110.2530.735
DepthCrafter [20]Vid.1521M0.343†0.673†0.1250.8480.1100.8810.0750.9710.1630.843
Marigold [27]Im.865M0.5320.5150.1660.7690.1490.7960.0910.9310.2350.753
DepthAny. [62] (*)Im.343M0.3250.5640.1300.8380.1420.8030.0780.9390.1690.786
DepthAny.-V2 [63]Im.343M0.3670.5540.1350.8220.1400.8040.1060.9210.1870.775
OursVid.140M0.2950.6400.0870.9330.1040.9030.0660.9710.1380.862
", + "image_path": "6580a0dc45bfed4ac426b258ba8039bd1d10b42d8269ba8e0209685f82cfffd6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 226, + 553, + 357 + ], + "blocks": [ + { + "bbox": [ + 55, + 190, + 555, + 224 + ], + "lines": [ + { + "bbox": [ + 55, + 190, + 555, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 555, + 224 + ], + "type": "text", + "content": "Table 1. DepthCrafter [20] video depth benchmark. Our model upgrades DepthAnything (*) to a video depth predictor, outperforming all baselines (image- or video-based) while having the smallest parameter count (140M). We use DepthAnything-Base (97M) as the base model and still outperform the DepthAnything-Large (343M) model. " + }, + { + "bbox": [ + 55, + 190, + 555, + 224 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 190, + 555, + 224 + ], + "type": "text", + "content": "Reproduced results confirmed by original authors." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 226, + 553, + 357 + ], + "lines": [ + { + "bbox": [ + 56, + 226, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 56, + 226, + 553, + 357 + ], + "type": "table", + "html": "
MethodsGT PoseGT Intr.AlignKITTIScanNetETH3DDTUT&TAverageRuntime (s)
rel ↓δ1.03 ↑rel ↓δ1.03 ↑rel ↓δ1.03 ↑rel ↓δ1.03 ↑rel ↓δ1.03 ↑rel ↓δ1.03 ↑
COLMAP [45, 46]#✓✓×12.058.214.634.216.455.10.796.52.795.09.367.8≈3 min
COLMAP Dense [45, 46]✓✓×26.952.738.022.589.823.220.869.325.776.440.248.8≈3 min
NVDS [54]××LSvld17.413.913.017.121.411.97.031.615.514.414.917.853.0
ChronoDepth [49]××LSvld15.314.814.714.521.111.37.031.014.316.814.517.77.0
DepthCrafter [20]××LSvld10.222.26.634.615.414.84.546.912.918.49.927.438.1
DepthAny. [62] (*)××LSvld8.428.15.244.313.918.63.755.110.425.38.334.30.1
DepthAny.V2 [63]××LSvld9.126.34.650.016.116.03.261.610.126.58.636.10.1
DUSt3R [56]××LSvld7.231.74.650.013.419.03.459.29.328.47.637.74.3
UniDepth [44]××LSvld(3.9)(58.2)(1.5)(90.1)12.718.97.030.710.623.27.144.20.5
Ours××LSvld6.932.44.550.89.227.13.262.47.332.96.241.10.9‡
", + "image_path": "ff572a2f012f6485f270a5a19208d689023cf72fbae16534646c360ff3202358.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 358, + 555, + 392 + ], + "lines": [ + { + "bbox": [ + 55, + 358, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 555, + 392 + ], + "type": "text", + "content": "Table 2. RobustMVD [48] multi-view depth benchmark, analogous to Tab. 1. Parentheses indicate that a model was trained on the same benchmark used for evaluation, which provides an advantage compared to training on different data. " + }, + { + "bbox": [ + 55, + 358, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 55, + 358, + 555, + 392 + ], + "type": "text", + "content": "Includes approximately 0.4 seconds for tracking. " + }, + { + "bbox": [ + 55, + 358, + 555, + 392 + ], + "type": "inline_equation", + "content": "^\\ddagger" + }, + { + "bbox": [ + 55, + 358, + 555, + 392 + ], + "type": "text", + "content": "Metrics computed only for pixels with a valid prediction, which is an easier task." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 401, + 294, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 294, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 294, + 472 + ], + "type": "text", + "content": "18.3% and 9.7%. Second, Tracktention performs better than all other baselines, including DepthCrafter [20], which is a video depth model, by 15.3% and 2.3% on AbsRel and " + }, + { + "bbox": [ + 55, + 401, + 294, + 472 + ], + "type": "inline_equation", + "content": "\\delta_{1.25}" + }, + { + "bbox": [ + 55, + 401, + 294, + 472 + ], + "type": "text", + "content": ", respectively. This demonstrates that Tracktention can turn an image-based model into a video-based one, outperforming models designed for video data from the start." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 473, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 556 + ], + "type": "text", + "content": "Table 2 reports results on the RobustMVD benchmark [48], which features shorter video clips with potentially larger camera motions. Our model achieved the lowest average relative error of 6.2 and the highest average inlier ratio of " + }, + { + "bbox": [ + 55, + 473, + 295, + 556 + ], + "type": "inline_equation", + "content": "41.1\\%" + }, + { + "bbox": [ + 55, + 473, + 295, + 556 + ], + "type": "text", + "content": ", surpassing all baselines. In particular, it improves the base model Depth Anything by " + }, + { + "bbox": [ + 55, + 473, + 295, + 556 + ], + "type": "inline_equation", + "content": "33.8\\%" + }, + { + "bbox": [ + 55, + 473, + 295, + 556 + ], + "type": "text", + "content": " on ETH3D and by " + }, + { + "bbox": [ + 55, + 473, + 295, + 556 + ], + "type": "inline_equation", + "content": "29.8\\%" + }, + { + "bbox": [ + 55, + 473, + 295, + 556 + ], + "type": "text", + "content": " on Tanks and Temples." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 557, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 653 + ], + "type": "text", + "content": "Our model is also efficient, utilizing only 140 million parameters — only 17.1 million for the Tracktention module. It operates with a runtime of only 0.9 seconds per sequence, of which 0.4 seconds are attributed to the tracker. Despite its smaller size, our model outperforms DepthCrafter [20], which relies on video diffusion models with over 1.5 billion parameters. Tracktention will further benefit from future accuracy and speed improvements in point trackers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 670, + 170, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 170, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 170, + 682 + ], + "type": "text", + "content": "4.1.2 Qualitative Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Figure 4 shows a qualitative comparison of depth estimation by our model, DepthCrafter, and DUSt3R in dynamic" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 401, + 555, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 555, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 555, + 496 + ], + "type": "text", + "content": "scenes, complex textures, and cluttered environments. Our model produces stable, coherent depth maps across all frames, demonstrating strong temporal consistency. In contrast, DepthCrafter shows significant depth estimation errors in certain regions (blue box), especially in scenes with complex structures. DUSt3R, relying on implicit triangulation, struggles with dynamic content, resulting in inconsistent depth maps across time." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 498, + 555, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 535 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 535 + ], + "type": "text", + "content": "Figure 6 illustrates the Attentional Sampling block, which pools information along the track by attending to image features in correspondence with the track locations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 548, + 477, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 477, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 477, + 559 + ], + "type": "text", + "content": "4.2. Automatic Video Colorization" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 567, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 554, + 604 + ], + "type": "text", + "content": "The goal of video colorization is to add colors to an input grayscale video. Colors should be realistic, vibrant, and consistent across frames to avoid flickering." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": "We enhance four models — CIC [66], IDC [67], Colorformer [21], and DDColor [22] — by incorporating our Tracktention module into layers of each model's architecture. For instance, for DDColor, which has a ConvNet-based architecture with a ConvNeXt backbone [36] and a U-Net decoder, we insert our Tracktention module before each decoder block. We fine-tune the model using AdamW on the YouTube-VIS training set [60]. For testing, we use the validation sets of both DAVIS [43] and Video [30]." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22814" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 72, + 542, + 268 + ], + "blocks": [ + { + "bbox": [ + 65, + 72, + 542, + 268 + ], + "lines": [ + { + "bbox": [ + 65, + 72, + 542, + 268 + ], + "spans": [ + { + "bbox": [ + 65, + 72, + 542, + 268 + ], + "type": "image", + "image_path": "06c1fb5114cc00995fd1eafdc60c0f23969fe68ad34da647e15834e518a2881d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 269, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 55, + 269, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 555, + 301 + ], + "type": "text", + "content": "Figure 4. Video depth prediction, comparing Tracktention (+DepthAnything), DepthCrafter [20], and DUS3R [56]. We visualize a column of pixels (highlighted in red) over time to illustrate temporal variation. Our model shows stable, coherent depth estimation over time, while DepthCrafter exhibits significant errors in certain regions (blue box). DUS3R struggles with dynamic content (green box)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 65, + 306, + 545, + 494 + ], + "blocks": [ + { + "bbox": [ + 65, + 306, + 545, + 494 + ], + "lines": [ + { + "bbox": [ + 65, + 306, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 65, + 306, + 545, + 494 + ], + "type": "image", + "image_path": "6b571058b25d6ece888a668a281c9178d228c77eec8273c173d0e3d8434ea67b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 494, + 555, + 515 + ], + "lines": [ + { + "bbox": [ + 55, + 494, + 555, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 555, + 515 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of video colorization methods. Tracktention (+DDColor) yields vibrant, realistic, and consistent colors. In contrast, TCVC appears less vibrant (21.7 vs. 29.5 Colorfulness), while DDColor lacks temporal consistency." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 526, + 295, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 295, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 295, + 550 + ], + "type": "text", + "content": "Evaluation metrics. We evaluate colorization results using standard metrics. Please refer to Appendix A.3.2 for details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 554, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 295, + 685 + ], + "type": "text", + "content": "Quantitative Results. Figure 5 reports video colorization results on DAVIS and Video. We apply Tracktention to several image colorization models, consistently improving their performance when applied to videos, particularly in color consistency. Specifically, CDC improves by " + }, + { + "bbox": [ + 55, + 554, + 295, + 685 + ], + "type": "inline_equation", + "content": "22.8\\%" + }, + { + "bbox": [ + 55, + 554, + 295, + 685 + ], + "type": "text", + "content": " on DAVIS and " + }, + { + "bbox": [ + 55, + 554, + 295, + 685 + ], + "type": "inline_equation", + "content": "28.3\\%" + }, + { + "bbox": [ + 55, + 554, + 295, + 685 + ], + "type": "text", + "content": " on Video for CIC [66], " + }, + { + "bbox": [ + 55, + 554, + 295, + 685 + ], + "type": "inline_equation", + "content": "41.3\\%" + }, + { + "bbox": [ + 55, + 554, + 295, + 685 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 554, + 295, + 685 + ], + "type": "inline_equation", + "content": "28.8\\%" + }, + { + "bbox": [ + 55, + 554, + 295, + 685 + ], + "type": "text", + "content": " for Colorformer [21], and " + }, + { + "bbox": [ + 55, + 554, + 295, + 685 + ], + "type": "inline_equation", + "content": "46.5\\%" + }, + { + "bbox": [ + 55, + 554, + 295, + 685 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 554, + 295, + 685 + ], + "type": "inline_equation", + "content": "45.7\\%" + }, + { + "bbox": [ + 55, + 554, + 295, + 685 + ], + "type": "text", + "content": " for DD-Color [22]. Tracktention also improves FID and Colorfulness, indicating overall improvements in visual quality. We also compare to native video colorization networks (top rows), often outperforming these as well." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Qualitative Results. In Fig. 5, our method delivers more vibrant, realistic, and consistent colors compared to TCVC" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 526, + 555, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 555, + 574 + ], + "type": "text", + "content": "and DDColor. TCVC's colors are less vibrant, with a colorfulness score of 21.7 compared to our 29.5, and DDColor exhibits inconsistencies across frames, with colors sometimes changing dramatically over time." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 585, + 416, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 585, + 416, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 585, + 416, + 597 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 605, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 640 + ], + "type": "text", + "content": "Here, we measure the impact of various design decisions in Tracktention using the AbsRel metric on depth prediction in Sintel, ScanNet, and Tanks and Temples." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Table 4a assesses the impact of the number of tracks used in training: even a single track per video can lead to improvements, and the optimal number is around " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "24^{2}" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " tracks. Table 4b assesses the number of tracks used at testing time, with similar results. Table 4c compares different tracking strategies: eschewing the tracker and simply setting tracks" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22815" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 382, + 231 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 382, + 231 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 382, + 231 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 382, + 231 + ], + "type": "table", + "html": "
MethodDAVIS (medium frame length)Video (long frame length)
FID↓CF↑CDC↓PSNR↑FID↓CF↑CDC↓PSNR↑
Specialized video colorization models:
FAVC [31]—18.554.2224.38—16.281.8824.81
TCVC [34]46.5121.703.7325.5039.5819.071.6425.43
CIC [66]44.2330.346.1823.1939.7829.193.5922.51
[66]+Ours43.1334.224.77 (-22.8%)21.5837.0032.732.57 (-28.3%)22.64
IDC [67]42.9921.705.0125.4237.2519.072.5725.35
[67]+Ours33.4728.104.45 (-11.1%)23.5133.1325.222.42 (-6.0%)24.04
ColorFormer [21]40.7129.618.2923.0340.1229.764.9623.08
[21]+Ours29.7332.244.86 (-41.3%)23.1827.2529.653.53 (-28.8%)23.31
DDColor [22]26.8130.618.6423.8120.2730.385.6024.34
[22]+Ours24.6129.534.62 (-46.5%)23.8522.7826.063.04 (-45.7%)24.39
", + "image_path": "4bd26969699b6ba9b2aee43a62f5af61bb319443ed377b0f85c2ee534a46f474.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 61, + 270, + 217, + 331 + ], + "blocks": [ + { + "bbox": [ + 55, + 232, + 382, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 382, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 382, + 266 + ], + "type": "text", + "content": "Table 3. Quantitative comparison of video colorization methods on the DAVIS and Videvo datasets. Our method, when augmented onto four different baseline models, consistently improves the Color Distribution Consistency (CDC) metric across both datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 270, + 217, + 331 + ], + "lines": [ + { + "bbox": [ + 61, + 270, + 217, + 331 + ], + "spans": [ + { + "bbox": [ + 61, + 270, + 217, + 331 + ], + "type": "table", + "html": "
#tracksSintel↓ScanNet↓T&T↓
0 (base)0.3250.1300.104
10.3470.0940.087
12²0.3190.0860.080
24²0.2950.0870.073
", + "image_path": "fdfa61bbfe90d7b9c917455b2e51420664a04b0c5ed17b5e91d70e166700f2fa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 61, + 364, + 217, + 415 + ], + "blocks": [ + { + "bbox": [ + 56, + 333, + 217, + 361 + ], + "lines": [ + { + "bbox": [ + 56, + 333, + 217, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 217, + 361 + ], + "type": "text", + "content": "(a) Number of training tracks: Increasing training tracks consistently improved depth estimation accuracy, with optimal results at " + }, + { + "bbox": [ + 56, + 333, + 217, + 361 + ], + "type": "inline_equation", + "content": "24^{2}" + }, + { + "bbox": [ + 56, + 333, + 217, + 361 + ], + "type": "text", + "content": " tracks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 364, + 217, + 415 + ], + "lines": [ + { + "bbox": [ + 61, + 364, + 217, + 415 + ], + "spans": [ + { + "bbox": [ + 61, + 364, + 217, + 415 + ], + "type": "table", + "html": "
MethodSintel↓ScanNet↓T&T↓
3D Conv.0.3640.1120.096
Time Attn.0.3630.0970.081
Tracktention0.2950.0870.073
", + "image_path": "9ac69fa1aa79a05762be34a5fa4a6bec6fb701b4b5c4ae38416703c4e9579550.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 229, + 270, + 386, + 331 + ], + "blocks": [ + { + "bbox": [ + 56, + 416, + 217, + 446 + ], + "lines": [ + { + "bbox": [ + 56, + 416, + 217, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 217, + 446 + ], + "type": "text", + "content": "(d) Temporal modeling: Tracktention outperformed 3D conv. and time attention by leveraging tracking data for better temporal alignment." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 229, + 270, + 386, + 331 + ], + "lines": [ + { + "bbox": [ + 229, + 270, + 386, + 331 + ], + "spans": [ + { + "bbox": [ + 229, + 270, + 386, + 331 + ], + "type": "table", + "html": "
#tracksSintel↓ScanNet↓T&T↓
0 (base)0.3250.1300.104
10.3180.1180.108
12²0.3210.0870.075
24²0.2950.0870.073
", + "image_path": "dc2e3363cf942e6b31b1cb3b267c5e68743989a7908f073a17507feeb865ff07.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 229, + 364, + 386, + 415 + ], + "blocks": [ + { + "bbox": [ + 225, + 332, + 385, + 361 + ], + "lines": [ + { + "bbox": [ + 225, + 332, + 385, + 361 + ], + "spans": [ + { + "bbox": [ + 225, + 332, + 385, + 361 + ], + "type": "text", + "content": "(b) Number of testing tracks: Increasing testing tracks significantly boosted performance, indicating better temporal capture with more tracks." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 229, + 364, + 386, + 415 + ], + "lines": [ + { + "bbox": [ + 229, + 364, + 386, + 415 + ], + "spans": [ + { + "bbox": [ + 229, + 364, + 386, + 415 + ], + "type": "table", + "html": "
#LayersSintel↓ScanNet↓T&T↓
10.3160.0890.074
20.2950.0870.073
30.3170.0830.073
", + "image_path": "930356aadebf9387afee63d8bd2494d2cbd536b8b9bf2f9a79adaa82b570e4bf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 391, + 72, + 553, + 212 + ], + "blocks": [ + { + "bbox": [ + 391, + 72, + 553, + 212 + ], + "lines": [ + { + "bbox": [ + 391, + 72, + 553, + 212 + ], + "spans": [ + { + "bbox": [ + 391, + 72, + 553, + 212 + ], + "type": "image", + "image_path": "f01fabddce6f52c7beefcc580b00dd92b0e88ea4dafb4bf1ef3f689e8fcf30c3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 220, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 386, + 220, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 386, + 220, + 555, + 266 + ], + "type": "text", + "content": "Figure 6. Attentional Sampling module. The module attends to image features in correspondence with each track, as shown by the attention maps next to each frame." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 397, + 270, + 553, + 331 + ], + "blocks": [ + { + "bbox": [ + 224, + 416, + 385, + 446 + ], + "lines": [ + { + "bbox": [ + 224, + 416, + 385, + 446 + ], + "spans": [ + { + "bbox": [ + 224, + 416, + 385, + 446 + ], + "type": "text", + "content": "(e) Number of Time Attention Layers: Two layers balanced performance and complexity, while additional layers offered no consistent benefits." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 397, + 270, + 553, + 331 + ], + "lines": [ + { + "bbox": [ + 397, + 270, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 397, + 270, + 553, + 331 + ], + "type": "table", + "html": "
TypeSintel↓ScanNet↓T&T↓
Constant0.3520.0990.085
Grid, T00.3250.0960.073
Rand., T00.3200.0950.073
Rand., all0.2950.0870.073
", + "image_path": "517e742601b64cfa75a4ccc14ee60fe5db226789d32ebb9d4a05456baa443907.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 397, + 364, + 553, + 415 + ], + "blocks": [ + { + "bbox": [ + 392, + 332, + 553, + 361 + ], + "lines": [ + { + "bbox": [ + 392, + 332, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 392, + 332, + 553, + 361 + ], + "type": "text", + "content": "(c) Type of tracks: Using a point tracker performs better than using static pseudo-tracks. Furthermore, querying tracks uniformly is better." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 397, + 364, + 553, + 415 + ], + "lines": [ + { + "bbox": [ + 397, + 364, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 397, + 364, + 553, + 415 + ], + "type": "table", + "html": "
#LayersSintel↓ScanNet↓T&T↓
10.3440.0900.087
60.2950.0870.073
120.3330.1050.075
", + "image_path": "f7b9c09a570006d83c3244d035f9c11bb3a2b8e0c5d1eb0c4b96a89040ac2a70.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 392, + 416, + 553, + 446 + ], + "lines": [ + { + "bbox": [ + 392, + 416, + 553, + 446 + ], + "spans": [ + { + "bbox": [ + 392, + 416, + 553, + 446 + ], + "type": "text", + "content": "(f) Number of Tracktention layers: 6 layers improved temporal representation, but further layers led to diminishing returns due to overfitting risks." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 449, + 555, + 472 + ], + "lines": [ + { + "bbox": [ + 55, + 449, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 555, + 472 + ], + "type": "text", + "content": "Table 4. Ablations on video depth estimation results. Each table examines the impact of different factors, such as the number and type of tracks used during training and testing. Results are reported using the AbsRel error metric, where " + }, + { + "bbox": [ + 55, + 449, + 555, + 472 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 55, + 449, + 555, + 472 + ], + "type": "text", + "content": " indicates that lower values are better." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 478, + 297, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 478, + 297, + 622 + ], + "spans": [ + { + "bbox": [ + 54, + 478, + 297, + 622 + ], + "type": "text", + "content": "to be constant across all frames, and initializing the tracks on a grid in the first frame, at random in the first frame, or at random throughout the video. The latter results in much better performance, showing the importance of tracking and ensuring that tracks cover the video well. Table 4d tests replacing Tracktention Layers with 3D convolutions and time attention, which, by comparison, barely improve results. Table 4e shows the impact of increasing the number of transformer layers in the Track Transformer, where two layers are optimal in most cases. Finally, Tab. 4f shows the impact of adding a different number of Tracktention layers, concluding that six is about optimal." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 634, + 133, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 133, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 133, + 645 + ], + "type": "text", + "content": "4.4. Limitations" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Tracktention's success depends on the quality of the extracted tracks. While modern point trackers are robust and efficient, they may still fail, particularly when tracks are off-camera for extended periods. Furthermore, running a point tracker in addition to the primary neural network adds com" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 478, + 555, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 555, + 503 + ], + "type": "text", + "content": "plexity and computational cost (although this is often offset by avoiding spatio-temporal attention)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 519, + 388, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 388, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 388, + 532 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 541, + 555, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 541, + 555, + 686 + ], + "spans": [ + { + "bbox": [ + 312, + 541, + 555, + 686 + ], + "type": "text", + "content": "We have introduced Tracktention, a plug-and-play layer that upgrades image predictor networks into video predictor ones. It leverages the power of modern point trackers, transferring their understanding of motion to new applications such as depth prediction and colorization. The module is lightweight, simple to integrate, and compatible with a variety of architectures, including Vision Transformers and ConvNets. Thanks to the efficiency of recent point trackers like CoTracker3, it incurs only a small computational overhead. Remarkably, the resulting upgraded image models are overall faster and more accurate than some state-of-the-art methods specifically designed for video prediction." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Acknowledgments. This research was supported by ERC 101001212-UNION." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22816" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6836-6846, 2021. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 213 + ], + "type": "text", + "content": "[2] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, et al. Arkitsscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. arXiv preprint arXiv:2111.08897, 2021. 14" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 216, + 295, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 216, + 295, + 249 + ], + "spans": [ + { + "bbox": [ + 62, + 216, + 295, + 249 + ], + "type": "text", + "content": "[3] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 251, + 295, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 295, + 305 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 295, + 305 + ], + "type": "text", + "content": "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 307, + 295, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 307, + 295, + 372 + ], + "spans": [ + { + "bbox": [ + 62, + 307, + 295, + 372 + ], + "type": "text", + "content": "[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 374, + 295, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 374, + 295, + 419 + ], + "spans": [ + { + "bbox": [ + 62, + 374, + 295, + 419 + ], + "type": "text", + "content": "[6] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 420, + 295, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 295, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 295, + 463 + ], + "type": "text", + "content": "[7] Seokju Cho, Jiahui Huang, Jisu Nam, Honggyu An, Seun-gryong Kim, and Joon-Young Lee. Local all-pair correspondence for point tracking. arXiv preprint arXiv:2407.15420, 2024.3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 466, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 466, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 466, + 295, + 521 + ], + "type": "text", + "content": "[8] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 14" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 523, + 295, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 523, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 62, + 523, + 295, + 589 + ], + "type": "text", + "content": "[9] Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Peter Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, et al. Scaling vision transformers to 22 billion parameters. In International Conference on Machine Learning, pages 7480-7512. PMLR, 2023. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 590, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 295, + 645 + ], + "type": "text", + "content": "[10] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610–13626, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 647, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 712 + ], + "type": "text", + "content": "[11] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10061-10072, 2023. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[12] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. TAPIR: tracking any point with per-frame initialization and temporal refinement. In Proc. CVPR, 2023. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "type": "text", + "content": "[13] Carl Doersch, Yi Yang, Dilara Gokay, Pauline Luc, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ross Goroshin, João Carreira, and Andrew Zisserman. BootsTAP: Bootstrapped training for tracking-any-point. arXiv, 2402.00847, 2024. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 175, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 195 + ], + "type": "text", + "content": "[14] H Fan, B Xiong, K Mangalam, Y Li, Z Yan, and J Malik.... Multiscale vision transformers. In Proc. CVPR, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 198, + 555, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 555, + 251 + ], + "type": "text", + "content": "[15] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6824-6835, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "type": "text", + "content": "[16] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 299, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 299, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 317, + 299, + 553, + 341 + ], + "type": "text", + "content": "[17] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 14" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 343, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 343, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 317, + 343, + 553, + 387 + ], + "type": "text", + "content": "[18] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In European Conference on Computer Vision, pages 59-75. Springer, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 388, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 553, + 421 + ], + "type": "text", + "content": "[19] Adam W. Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle videos revisited: Tracking through occlusions using point trajectories. In Proc. ECCV, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 422, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 466 + ], + "type": "text", + "content": "[20] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024. 5, 6, 7, 14" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "type": "text", + "content": "[21] Xiaozhong Ji, Boyuan Jiang, Donghao Luo, Guangpin Tao, Wenqing Chu, Zhifeng Xie, Chengjie Wang, and Ying Tai. Colorformer: Image colorization via color memory assisted hybrid-attention transformer. In European Conference on Computer Vision, pages 20–36. Springer, 2022. 6, 7, 8, 15" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 523, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 553, + 578 + ], + "type": "text", + "content": "[22] Xiaoyang Kang, Tao Yang, Wenqi Ouyang, Peiran Ren, Lingzhi Li, and Xuansong Xie. Ddcolor: Towards photorealistic image colorization via dual decoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 328-338, 2023. 6, 7, 8, 12, 15" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 580, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 553, + 643 + ], + "type": "text", + "content": "[23] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Dynamic stereo: Consistent dynamic depth from stereo videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13229-13239, 2023. 14" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "type": "text", + "content": "[24] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker3: Simpler and better point tracking by pseudolabelling real videos. arxiv, 2024. 5, 13" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[25] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22817" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "tracker: It is better to track together. In Proc. ECCV, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[26] Nikita Karaev, Ignacio Rocco, Ben Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker: It is better to track together. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 206 + ], + "type": "text", + "content": "[27] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "text", + "content": "[28] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36 (4):1-13, 2017. 14" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 295, + 297 + ], + "type": "text", + "content": "[29] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. 3, 14" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 353 + ], + "type": "text", + "content": "[30] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In Proceedings of the European conference on computer vision (ECCV), pages 170-185, 2018. 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 355, + 295, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 295, + 399 + ], + "type": "text", + "content": "[31] Chenyang Lei and Qifeng Chen. Fully automatic video colorization with self-regularization and diversity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3753–3761, 2019. 8, 15" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 400, + 295, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 400, + 295, + 444 + ], + "spans": [ + { + "bbox": [ + 56, + 400, + 295, + 444 + ], + "type": "text", + "content": "[32] Hongyang Li, Hao Zhang, Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, and Lei Zhang. Taptr: Tracking any point with transformers as detection. In European Conference on Computer Vision, pages 57-75. Springer, 2025. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 295, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 295, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 295, + 511 + ], + "type": "text", + "content": "[33] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. 14" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 295, + 567 + ], + "type": "text", + "content": "[34] Yihao Liu, Hengyuan Zhao, Kelvin CK Chan, Xintao Wang, Chen Change Loy, Yu Qiao, and Chao Dong. Temporally consistent video colorization with deep feature propagation and self-regularization learning. arXiv preprint arXiv:2110.04562, 2021.8, 15" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 295, + 612 + ], + "type": "text", + "content": "[35] Ziwei Liu, Raymond A Yeh, Xiaou Tang, Yiming Liu, and Aseem Agarwala. Video frame synthesis using deep voxel flow. In Proceedings of the IEEE international conference on computer vision, pages 4463-4471, 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 295, + 667 + ], + "type": "text", + "content": "[36] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 6, 15" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "text", + "content": "[37] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3202-3211, 2022. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "text", + "content": "[38] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4):71-1, 2020. 3, 14" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 107, + 555, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 555, + 173 + ], + "type": "text", + "content": "[39] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4040-4048, 2016. 14" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 554, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 554, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 554, + 207 + ], + "type": "text", + "content": "[40] Daniel Neimark, Omri Bar, Maya Zohar, and Dotan Asselmann. Video transformer network. In Proc. ICCV Workshops, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 209, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 554, + 262 + ], + "type": "text", + "content": "[41] E. Palazzolo, J. Behley, P. Lottes, P. Giguère, and C. Stachniss. ReFusion: 3D Reconstruction in Dynamic Environments for RGB-D Cameras Exploiting Residuals. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2019. 14" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 554, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 554, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 554, + 330 + ], + "type": "text", + "content": "[42] Mandela Patrick, Dylan Campbell, Yuki Markus Asano, Ishan Misra, Florian Metze, Christoph Feichtenhofer, Andrea Vedaldi, and João F. Henriques. Keeping your eye on the ball: Trajectory attention in video transformers. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2021. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 331, + 554, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 554, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 554, + 396 + ], + "type": "text", + "content": "[43] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 399, + 554, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 554, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 554, + 441 + ], + "type": "text", + "content": "[44] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. UniDepth: universal monocular metric depth estimation. In Proc. CVPR, 2024. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 444, + 554, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 554, + 464 + ], + "type": "text", + "content": "[45] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proc. CVPR, 2016. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 467, + 554, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 467, + 554, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 467, + 554, + 499 + ], + "type": "text", + "content": "[46] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In Proc. ECCV, 2016. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 501, + 554, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 554, + 567 + ], + "type": "text", + "content": "[47] Thomas Schops, Johannes L. Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 14" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 568, + 554, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 554, + 612 + ], + "type": "text", + "content": "[48] Philipp Schröppel, Jan Bechtold, Artemij Amiranashvili, and Thomas Brox. A benchmark and a baseline for robust multiview depth estimation. In 2022 International Conference on 3D Vision (3DV), pages 637-645. IEEE, 2022. 6, 14" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 613, + 554, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 554, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 554, + 656 + ], + "type": "text", + "content": "[49] Jiahao Shao, Yuanbo Yang, Hongyu Zhou, Youmin Zhang, Yujun Shen, Matteo Poggi, and Yiyi Liao. Learning temporally consistent video depth from video diffusion priors. arXiv, 2406.01493, 2024. 6, 14" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "type": "text", + "content": "[50] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22818" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[51] Jianlin Su, Yu Lu, Shengfeng Pan, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding, 2021. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "text", + "content": "[52] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "text", + "content": "[53] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "text", + "content": "[54] Liao Wang, Qiang Hu, Qihan He, Ziyu Wang, Jingyi Yu, Tinne Tuytelaars, Lan Xu, and Minye Wu. Neural residual radiance fields for streamably free-viewpoint videos. In Proc. CVPR, 2023. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "text", + "content": "[55] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19795-19806, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 353 + ], + "type": "text", + "content": "[56] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. DUS3R: Geometric 3D vision made easy. In Proc. CVPR, 2024. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 419 + ], + "type": "text", + "content": "[57] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4909-4916. IEEE, 2020. 14" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 421, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 294, + 453 + ], + "type": "text", + "content": "[58] Xiaolong Wang, Ross B. Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proc. CVPR, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 455, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 509 + ], + "type": "text", + "content": "[59] Yiran Wang, Min Shi, Jiaqi Li, Zihao Huang, Zhiguo Cao, Jianming Zhang, Ke Xian, and Guosheng Lin. Neural video depth stabilizer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9466-9476, 2023. 6, 14" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 511, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 294, + 533 + ], + "type": "text", + "content": "[60] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 6, 15" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 534, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 589 + ], + "type": "text", + "content": "[61] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "type": "text", + "content": "[62] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proc. CVPR, 2024. 5, 6, 12, 14" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "type": "text", + "content": "[63] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything V2. arXiv, 2406.09414, 2024. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[64] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 14" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 319 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 117 + ], + "type": "text", + "content": "[65] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 14" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 173 + ], + "type": "text", + "content": "[66] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 649-666. Springer, 2016. 6, 7, 8, 15" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "text", + "content": "[67] Richard Zhang, Jun-Yan Zhu, Phillip Isola, Xinyang Geng, Angela S Lin, Tianhe Yu, and Alexei A Efros. Real-time user-guided image colorization with learned deep priors. arXiv preprint arXiv:1705.02999, 2017. 6, 8, 15" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 219, + 553, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 273 + ], + "type": "text", + "content": "[68] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19855-19865, 2023. 14" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 274, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 319 + ], + "type": "text", + "content": "[69] Xizhou Zhu, Yuwen Xiong, Jifeng Dai, Lu Yuan, and Yichen Wei. Deep feature flow for video recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2349-2358, 2017. 3" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22819" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Training Data Provenance Verification_ Did Your Model Use Synthetic Data from My Generative Model for Training_/95a43e45-ca4b-40d2-9bb6-79d7e388dedb_content_list.json b/2025/Training Data Provenance Verification_ Did Your Model Use Synthetic Data from My Generative Model for Training_/95a43e45-ca4b-40d2-9bb6-79d7e388dedb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c08675f0d70fce5eab9f4f5957b2f2a0f5b9a89b --- /dev/null +++ b/2025/Training Data Provenance Verification_ Did Your Model Use Synthetic Data from My Generative Model for Training_/95a43e45-ca4b-40d2-9bb6-79d7e388dedb_content_list.json @@ -0,0 +1,1652 @@ +[ + { + "type": "text", + "text": "Training Data Provenance Verification: Did Your Model Use Synthetic Data from My Generative Model for Training?", + "text_level": 1, + "bbox": [ + 117, + 128, + 880, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuechen Xie $^{1}$ , Jie Song $^{1*}$ , Huiqiong Wang $^{2}$ , Mingli Song $^{1,3,4}$", + "bbox": [ + 261, + 202, + 733, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Zhejiang University", + "bbox": [ + 416, + 227, + 583, + 246 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Ningbo Innovation Center, Zhejiang University", + "bbox": [ + 307, + 246, + 689, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3State Key Laboratory of Blockchain and Security, Zhejiang University", + "bbox": [ + 217, + 263, + 779, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Hangzhou High-Tech Zone (Binjiang) Institute of Blockchain and Data Security", + "bbox": [ + 179, + 281, + 816, + 299 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{xyuechen,sjie,huiqiong-wang,brooksong}@zju.edu.cn", + "bbox": [ + 276, + 308, + 714, + 323 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 358, + 326, + 373 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "High-quality open-source text-to-image models have lowered the threshold for obtaining photorealistic images significantly, but also face potential risks of misuse. Specifically, suspects may use synthetic data generated by these generative models to train models for specific tasks without permission, when lacking real data resources especially. Protecting these generative models is crucial for the wellbeing of their owners. In this work, we propose the first method to this important yet unresolved issue, called Training data Provenance Verification (TrainProVe). The rationale behind TrainProVe is grounded in the principle of generalization error bound, which suggests that, for two models with the same task, if the distance between their training data distributions is smaller, their generalization ability will be closer. We validate the efficacy of TrainProVe across four text-to-image models (Stable Diffusion v1.4, latent consistency model, PixArt-α, and Stable Cascade). The results show that TrainProVe achieves a verification accuracy of over 99% in determining the provenance of suspicious model training data, surpassing all previous methods. Code is available at https://github.com/xieyc99/TrainProVe.", + "bbox": [ + 89, + 391, + 485, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 755, + 220, + 771 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "High-quality open-source text-to-image models based on diffusion models [15] can generate photorealistic images with simple text prompts [4, 28, 34, 38], making them highly significant in terms of lowering creative barriers, enabling personalized customization, etc. Currently, most open-source text-to-image models are only permitted for", + "bbox": [ + 89, + 782, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "academic or educational purposes, and commercial use is prohibited without explicit permission [3, 4, 21, 28, 33, 34, 36, 38]. Therefore, protecting the intellectual property of these generative models is crucial for safeguarding the interests of their owners.", + "bbox": [ + 511, + 359, + 906, + 434 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Currently, one of the major security challenges faced by defenders, i.e., the generative model owners, is the unauthorized use of the synthetic data from their models by suspects [24, 47, 48, 55]. As shown in Fig. 1, based on the suspect's intention, this issue can be divided into three cases:", + "bbox": [ + 511, + 436, + 906, + 513 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Case 1: the suspect steals the synthetic data illegally and claims it as his own personal work [24, 44, 48], implying the intellectual property of the generative model is infringed.", + "bbox": [ + 511, + 515, + 905, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Case 2: the suspect steals the synthetic data from the defender's generative model, and trains another generative model for the same task using knowledge distillation techniques [6, 14, 29, 39, 50]. In this case, the defender cannot access the synthetic data used by the suspect, making it a more challenging defense task compared to Case 1.", + "bbox": [ + 511, + 564, + 906, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Case 3: the suspect steals the synthetic data from the generative model to train another model for a specific task [1, 8, 13, 40, 42, 53, 57]. In this case, the defender is not only unaware of the training data used by the suspicious model, but also cannot compare the similarity between the generative model and the suspicious model directly due to the different tasks, representing an even more arduous defense task compared to Case 2.", + "bbox": [ + 511, + 656, + 906, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For Case 1 and Case 2, there are already some mature methods to address these issues, such as adding watermarks [24, 31, 52, 54, 56] and training data attribution [55], which identify illegal activities by checking the distributions of the suspicious images or the synthetic images from suspicious model. However, for Case 3, although the potential of using synthetic data to train models for other tasks has been recognized gradually in recent years [13, 40, 42, 53]", + "bbox": [ + 511, + 780, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 233, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23817", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f01004709656455b6d0c5b1cc26c89ab12083bcfe969bba33cf233c779edd264.jpg", + "image_caption": [ + "Figure 1. The overview of the three cases. These three cases encompass nearly every conceivable real-world scenario, where a suspect uses synthetic images from the defender's text-to-image model illegally. In this paper, we are committed to addressing the security risks in Case 3." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 450, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "due to the lower collection costs compared to real data, the associated security risks have been overlooked. Methods that work for the first two cases cannot be applied to the third case, because the learning task of the suspicious model is not necessarily image generation. For example, for a suspicious image classification model, there are no synthetic data available for defenders to examine.", + "bbox": [ + 89, + 609, + 483, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we address this important yet unexplored issue. To the best of our knowledge, we propose the first method for protecting the intellectual property of generative models in such scenario, which we call Training data Provenance Verification (TrainProVe). It helps defenders verify whether the suspicious model has been trained using synthetic data from their generative model. Given that suspicious models may use synthetic data for training without authorization, and is used for commercial purposes, we focus on the black-box setting, where the defenders have no information about the training configuration of the suspicious model (e.g., loss function, model architecture, etc.)", + "bbox": [ + 89, + 719, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and can only access the model through Machine Learning as a Service (MLaaS) [37, 43, 49]. This means that the defenders can only obtain prediction results (such as prediction logits or labels) via the suspicious model's API.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The rationale behind TrainProVe is grounded in the principle of generalization error bound, which suggests that, for two models with the same task, if the distance between their training data distributions is smaller, their generalization ability will be closer. Specifically, defenders can train a model for the same task as the suspicious model, using the synthetic data generated by their generative model. By comparing the performances of these two models, defenders can determine whether the suspicious model was trained on synthetic data from their generative model. More concretely, the proposed TrainProVe involves three steps: (1) using the defender's generative model along with different text prompts to generate a shadow dataset and a validation dataset, based on the suspicious model's task; (2) training a shadow model for the same task as the suspicious model using the shadow dataset; (3) performing hypothesis testing on the prediction accuracy of the suspicious model and the shadow model on the validation dataset, in order to determine whether the suspicious model was trained using synthetic data from the defender's generative model.", + "bbox": [ + 511, + 152, + 906, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the principal contributions of this paper are threefold: (1) we introduce the important yet under-explored issue of intellectual property protection for text-to-image models, in which the suspect uses the synthetic data from the defender's generative model illegally to train a specific-task suspicious model, such as an image classification model. For the defender, the suspicious model is a black-box; (2) we discover and demonstrate that the similarity between different models is related to their source domains closely, and based on this, we propose the training data provenance verification technique. To the best of our knowledge, this is the first method to address such intellectual property protection issue; (3) Extensive experiments show that TrainProVe achieves over $99\\%$ accuracy in verifying training data provenance for suspicious models, surpassing all previous studies.", + "bbox": [ + 511, + 457, + 908, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 717, + 653, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-image diffusion model. Diffusion models [15] have revolutionized generative modeling fundamentally, in terms of quality and generalization. Their excellent scalability has led to significant advancements in specific applications, such as text-to-image generation. Text-to-image diffusion models [4, 34, 38] use the diffusion process to transform text descriptions into realistic images. Recently, people have begun to recognize the potential of text-to-image diffusion models for synthesizing dataset [17, 18, 40, 42, 57]. However, so far, no one has realized that suspects", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "aside_text", + "text": "Low Defense Difficulty High", + "bbox": [ + 460, + 90, + 480, + 473 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23818", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "might use these text-to-image diffusion models to synthesize datasets and train their own models without authorization. Therefore, in this paper, we propose the first feasible solution to address this issue, securing and fostering healthy development in this field.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-image model protection. Training generative models requires large amounts of data and computational resources [4, 38], making the intellectual property of these models crucial for their owners. When these models are open-source [4, 28, 34, 38], their intellectual property is vulnerable to infringement by suspects. Specifically, for text-to-image models, suspects have the following three ways primarily to infringe on their intellectual property: (1) stealing synthetic data illegally; (2) illegal knowledge distillation; (3) training suspicious models for specific tasks using synthetic data illegally. For the first two cases, watermarking [24, 27, 56] and training data attribution [55] have been developed to address these issues. However, for the third case, no suitable solution exists currently. This work aims to address this gap.", + "bbox": [ + 89, + 171, + 482, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generalization error bound. Generalization error (upper) bound is used to describe a model's performance on testing data and widely applied in domain adaptation [9, 46], which is a special case of transfer learning [45, 51, 58]. Theorem 1 in [2] states the generalization error upper bound of model $M$ on the target domain $T$ . Let $\\mathcal{H}$ be a hypothesis space with VC dimension $d$ , and let $m$ be the sample size of the dataset in the source domain $S$ . Then, with probability at least $1 - \\delta$ , for every $M \\in \\mathcal{H}$ :", + "bbox": [ + 89, + 402, + 483, + 537 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {T} (M) \\leq \\hat {\\epsilon} _ {S} (M) + \\sqrt {\\frac {4}{m} \\left(d \\log \\frac {2 e m}{d}\\right)} + d _ {\\mathcal {H}} \\left(P _ {S}, P _ {T}\\right) + \\lambda , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 547, + 482, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon_T(M)$ and $\\hat{\\epsilon_S} (M)$ represent the generalization error of $M$ on the target domain and the empirical error of $M$ on the source domain, respectively. $P_{S}$ and $P_{T}$ denote the marginal probability distributions of the source and target domains, respectively. $d_{\\mathcal{H}}(P_S,P_T)$ is the $\\mathcal{H}$ -divergence between the source and target domains, which measures the distance between the two distributions within the hypothesis space $\\mathcal{H}$ . $\\lambda$ is a constant and $e$ is the base of the natural logarithm. The rationale behind TrainProVe is grounded in this theorem, which suggests that, for two models with the same task, if the distance between their training data distributions is smaller, their generalization ability will be closer.", + "bbox": [ + 89, + 599, + 482, + 781 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 792, + 181, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 816, + 294, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this study, we focus on the training data provenance verification task in image classification. This task involves two roles: the defender and the suspect. The defender, embodying the role of text-to-image model provider, endeavors to", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ascertain whether the black-box suspicious model, $M_{sus}$ , has been illicitly trained on synthetic dataset generated by his model $G_{d}$ (as shown in Case 3 of Fig. 1). More specifically, for a black-box suspect model, the defender's objective is to determine whether it is illegal or legal based solely on the labels it outputs.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$M_{sus}$ can be classified as either illegal or legal based on its training datasets: (1) Illegal: $M_{sus}$ is trained on the synthetic data from the defender's model $G_{d}$ , indicating the occurrence of intellectual property infringement; (2) Legal: $M_{sus}$ is trained using a different data source independent of the defender, such as some real data or synthetic data from other generative models, implying the innocence of the suspect. To be precise, when presented with a black-box suspicious model, the defender's objective is to determine whether it is an illegal or legal model.", + "bbox": [ + 511, + 181, + 906, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. The Theoretical Insight", + "text_level": 1, + "bbox": [ + 511, + 343, + 730, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The underlying rationale for TrainProVe is that when two models are assigned the same task, their ability to generalize will be more aligned if the divergence between their training data distributions is diminished. We demonstrate this through a case study of $K$ -way classification, where $\\mathcal{X} = \\mathbb{R}^{C \\times W \\times H}$ and $\\mathcal{Y} = \\{1, \\dots, K\\}$ represent the input and label spaces respectively. We follow the setting of previous domain adaptation works [32, 46], viewing a domain as composed of a feature space $\\mathcal{X}$ and a marginal probability distribution $P(\\boldsymbol{x})$ , i.e., $\\{\\mathcal{X}, P(\\boldsymbol{x})\\}$ , where $\\boldsymbol{x} \\in \\mathcal{X}$ .", + "bbox": [ + 511, + 364, + 905, + 516 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $S_{1} = \\{\\mathcal{X},P_{1}(\\pmb {x})\\}$ and $S_{2} = \\{\\mathcal{X},P_{2}(\\pmb {x})\\}$ be two source domains sharing the same feature space but possessing distinct marginal probability distributions, i.e., $P_{1}(\\pmb {x})\\neq$ $P_{2}(\\pmb {x}).D_{1} = \\{(x_{1}^{i},y_{1}^{i})\\}_{i = 1}^{|D_{1}|}$ and $\\mathcal{D}_2 = \\{(\\pmb {x}_2^j,\\pmb {y}_2^j)\\}_{j = 1}^{|D_2|}$ represent the datasets sampled from the source domains $S_{1}$ and $S_{2}$ respectively, where $\\pmb{x}_1^i\\sim P_1(\\pmb {x})$ and $\\pmb{x}_2^j\\sim P_2(\\pmb {x})$ . $M$ and $\\hat{M}$ are the models trained on $\\mathcal{D}_1$ and $\\mathcal{D}_2$ respectively. Then we have the following theorem:", + "bbox": [ + 511, + 517, + 905, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theorem 1. Assume $M$ is trained on synthetic data generated by a text-to-image model $G$ with a set of text prompts $\\mathcal{T}_1$ , i.e., $P_{1}(\\pmb {x}) = P(\\pmb {x}|G,\\mathcal{T}_{1})$ . $\\hat{M}$ can be trained on either real data or synthetic data generated by any text-to-image model. Based on the generalization errors of $M$ and $\\hat{M}$ on the target domain $T = \\{\\mathcal{X},P(\\pmb {x}|G,\\mathcal{T}_t)\\}$ , we have:", + "bbox": [ + 511, + 652, + 905, + 744 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\sup _ {P _ {2} (\\boldsymbol {x}) = P (\\boldsymbol {x} \\mid G, \\mathcal {T} _ {2})} | \\Delta \\epsilon_ {T} | \\leq \\sup _ {P _ {2} (\\boldsymbol {x}) \\perp G} | \\Delta \\epsilon_ {T} |, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 756, + 903, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Delta \\epsilon_{T}$ represents the difference in generalization error between $M$ and $\\hat{M}$ on the target domain $T$ , expressed as $\\Delta \\epsilon_{T} = \\epsilon_{T}(M) - \\epsilon_{T}(\\hat{M})$ . $P_{2}(\\pmb{x}) \\perp G$ denotes that $P_{2}(\\pmb{x})$ is independent of $G$ , meaning $P_{2}(\\pmb{x}) = P(\\pmb{x}|G^{\\prime},\\mathcal{T}_{2})$ or $P_{2}(\\pmb{x}) = P_{R}(\\pmb{x})$ , where $G^{\\prime}$ is a text-to-image model different from $G$ and $P_{R}(\\pmb{x})$ represents the distribution of real data. $\\mathcal{T}_1, \\mathcal{T}_2$ and $\\mathcal{T}_t$ are distinct sets of text prompts.", + "bbox": [ + 511, + 795, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23819", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6b78a5229db3a7dec00d36cbfb66c8bbe8c062cb07f41a0e37dd762dab0b0c95.jpg", + "image_caption": [ + "Figure 2. The overview of TrainProVe's motivation. $M$ is trained on the dataset whose data distribution is $P(x|G, \\mathcal{T}_1)$ . $\\hat{M}_1$ , $\\hat{M}_2$ , and $\\hat{M}_3$ are $\\hat{M}$ trained on three datasets with different data distributions, respectively. The source and target domain data are sampled from their respective data distributions. $G$ and $G'$ are two different text-to-image models, and $\\mathcal{T}_1$ , $\\mathcal{T}_2$ , and $\\mathcal{T}_t$ are distinct sets of text prompts." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 470, + 357 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/715efa933966fef9d79a188274f5d9efaa7454963eb7c8ba1e8fd9f5c261c52b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 473, + 89, + 903, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Detailed proofs can be found in the supplementary material. Theorem 1 suggests that for two models with distinct (identical) training data sources, their generalization errors differ substantially (minimally). Its contrapositive states that for two models with small (large) discrepancies in generalization error, their training data sources are identical (distinct), which supports our method. Specifically, when $M$ and $\\hat{M}$ are trained on synthetic data from the same text-to-image model, their accuracies on the corresponding synthetic data from this text-to-image model are more akin, compared to when their training data is derived from disparate sources. This aligns with the rationale of TrainProVe, as visually depicted in Fig. 2. The discrepancies in accuracies among these models originate from the disparities in their source domains. The closer the data distributions of the source domain data, the more analogous the accuracies of the trained models on the target domain data.", + "bbox": [ + 88, + 426, + 485, + 684 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. The Proposed TrainProVe", + "text_level": 1, + "bbox": [ + 89, + 695, + 328, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Building upon the above analysis, we propose a method called TrainProVe to protect generative models' intellectual property, as shown in Fig. 3, which consists of three stages:", + "bbox": [ + 89, + 718, + 482, + 762 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) Shadow and Validation Dataset Generation: Utilizing $G_{d}$ to produce a shadow dataset $\\mathcal{D}_{sdw}$ and a validation dataset $\\mathcal{D}_v$ , based on various text prompt sets $\\mathcal{T}_{sdw}$ and $\\mathcal{T}_v$ , tailored to the predicted labels set $\\mathcal{C}$ of the suspect model $M_{sus}$ . Here $G_{d}, \\mathcal{T}_{sdw}$ , and $\\mathcal{T}_v$ correspond to $G, \\mathcal{T}_1,$ and $\\mathcal{T}_t$ in Fig. 2 respectively, while $\\mathcal{D}_{sdw}$ and $\\mathcal{D}_v$ correspond to the training data of $M$ and the target domain data.", + "(2) Shadow Model Training: Using the shadow dataset $\\mathcal{D}_{sdw}$ to train the shadow model $M_{sdw}$ . The shadow model" + ], + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is analogous to $M$ in Fig. 2, whereas the suspect model is denoted as $\\hat{M}$ . The purpose of the shadow model is to compare its performance with the suspect model trained on distinct data, focusing on their generalization capacities (accuracies) on the validation dataset. Specifically, when both models are trained on data from $G_{d}$ , their accuracies on the validation dataset tend to be similar.", + "bbox": [ + 511, + 426, + 906, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(3) Hypothesis Testing: Providing the validation dataset $\\mathcal{D}_v$ to both the shadow model $M_{sdw}$ and the suspect model $M_{sus}$ in the same sequence batch by batch. Subsequently, calculating the accuracy of the two models batch by batch, leading to accuracy sets $\\mathcal{A}_{sdw}$ and $\\mathcal{A}_{sus}$ . Ultimately, utilizing a one-sided Grubbs' hypothesis test [10] to ascertain whether $M_{sus}$ violates the intellectual property of $G_d$ .", + "bbox": [ + 511, + 535, + 908, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the one-sided Grubbs' hypothesis test, the null hypothesis $H_0$ asserts that the mean of the set $\\mathcal{A}_{\\text{sus}}$ , denoted as $\\text{mean}(\\mathcal{A}_{\\text{sus}})$ , is not an unusually low value among the elements of $\\mathcal{A}_{\\text{sdw}}$ . On the contrary, the alternative hypothesis $H_1$ posits that the mean of $\\mathcal{A}_{\\text{sus}}$ is an exceptionally low value within $\\mathcal{A}_{\\text{sdw}}$ . Rejecting the null hypothesis $H_0$ implies that the predictive capabilities of $M_{\\text{sus}}$ are significantly inferior to those of $M_{\\text{sdw}}$ . This suggests that the model $M_{\\text{sus}}$ exhibits notably inferior generalization performance on the validation dataset compared to $M_{\\text{sdw}}$ , indicating that $M_{\\text{sus}}$ has not been exposed to the synthetic data from $G_d$ during training and is therefore compliant. Conversely, if the predictions generated by $M_{\\text{sus}}$ and $M_{\\text{sdw}}$ are similar, it indicates that the generalization performance of $M_{\\text{sus}}$ on the validation dataset closely resembles that of $M_{\\text{sdw}}$ . This similarity implies that $M_{\\text{sus}}$ , much like $M_{\\text{sdw}}$ , has been trained on the synthetic data from $G_d$ , thereby sug-", + "bbox": [ + 511, + 643, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23820", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9bbf7b8e16a032488cd675791d46c51d08569ec6cce0820c24ac10f956b3a3df.jpg", + "image_caption": [ + "1. Shadow and Validation Dataset Generation" + ], + "image_footnote": [], + "bbox": [ + 102, + 112, + 477, + 205 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/78b0c6168b7f39c1ea107b3f8824ec8803f7b8696f11c6f70210db933ebdaaa5.jpg", + "image_caption": [ + "2. Shadow Model Training" + ], + "image_footnote": [], + "bbox": [ + 102, + 228, + 478, + 300 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c562e8cfec122d9aba9ed01058b55243d0cd06254bc1892e6310cc7076b76002.jpg", + "image_caption": [ + "3. Hypothesis Testing", + "Figure 3. The complete process of TrainProVe." + ], + "image_footnote": [], + "bbox": [ + 102, + 321, + 478, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "gesting that $M_{sus}$ violates the intellectual property rights of $G_{d}$ and is considered illicit.", + "bbox": [ + 89, + 481, + 483, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 526, + 223, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, we present the results of DOV4MM on ImageNet [5] and CIFAR [20], ablation experiments, interference resistance analysis, and efficiency analysis. More details and results (hypothesis testing, results on CLIP [35], etc) can be found in the supplementary materials.", + "bbox": [ + 89, + 551, + 483, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 636, + 281, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Models and Datasets. We evaluate the proposed TrainProVe with three visual datasets (CIFAR10 [20], CIFAR100 [20], and ImageNet-100 [40]) and four text-to-image models (Stable Diffusion [38] v1.4, latent consistency model [28], PixArt- $\\alpha$ [4], and Stable Cascade [34]). ImageNet-100 is a randomly selected subset of ImageNet1K [5], encompassing 100 categories and 126,689 images. Stable Cascade is a latent diffusion model constructed on the Würstchen architecture [34]. In the experiments, unless explicitly stated otherwise, all text-to-image models employed are pre-trained models. To distinguish from real datasets, we denote the synthetic datasets of CIFAR10, CIFAR100, and ImageNet-100 as CIFAR10-Syn, CIFAR100-Syn, and ImageNet-100-Syn, respectively.", + "bbox": [ + 89, + 657, + 482, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experiments on CIFAR: The shadow dataset $\\mathcal{D}_{sdw}$ contains 30,000 images with a resolution of $32 \\times 32$ , and the text", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "prompt is $\\mathcal{T}_{sdw} = \\{\\text{“}a\\{c\\}\\text{”}|c\\in \\mathcal{C}\\}$ . The validation dataset $\\mathcal{D}_v$ includes 5,000 images with a resolution of $32\\times 32$ , with a text prompt of $\\mathcal{T}_v = \\{\\text{“}a\\text{ photo of }a\\{c\\}\\text{”}|c\\in \\mathcal{C}\\}$ . For the shadow model $M_{sdw}$ , the architecture is ResNet18 [12], and following previous work [42], we set the batch size for training $M_{sdw}$ to 512, the learning rate to 1e-4, weight decay to 0.9, and the loss function to cross-entropy loss. To simulate the training data for the suspicious model, we generated 60,000 images with a resolution of $32\\times 32$ using the method in [42] as CIFAR10-Syn/CIFAR100-Syn. Moreover, the suspicious models have various architectures and training hyperparameters to simulate different scenarios, specifically including: four architectures (ResNet50 [12], ResNet101 [12], ViT-S/16 [7], ViT-B/16 [7]), two batch sizes (512 and 256), two learning rates (1e-4 and 5e-5), two weight decay values (0.9 and 0.09), and two loss functions (cross-entropy loss and focal loss [22]). For each dataset, we can use these parameters to train 64 ( $C_4^1 C_2^1 C_2^1 C_2^1 C_2^1$ ) suspicious models. Batch size for inference in TrainProVe is 128, and training epoch for all models is set to 100.", + "bbox": [ + 511, + 90, + 903, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experiments on ImageNet-100: For the shadow dataset $\\mathcal{D}_{sdw}$ , validation dataset $\\mathcal{D}_v$ , and shadow model $M_{sdw}$ , all settings are the same as those used for CIFAR, except that $\\mathcal{D}_{sdw}$ contains 50,000 images with a resolution of $256 \\times 256$ and the batch size for model training is 64. The training set for the suspicious model $M_{sus}$ includes 100,000 images generated using the method in [40], with a resolution of $256 \\times 256$ . For the parameter selection of the suspicious model, all settings are the same as those used in the experiments on CIFAR, except that the batch size is 64 or 32. Therefore, we can also obtain 64 suspicious models for each dataset (ImageNet-100 or ImageNet-100-Syn). Additionally, batch size for inference in TrainProVe is 16. The training epoch for all models is set to 100.", + "bbox": [ + 511, + 393, + 903, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. Our goal is to classify the suspicious model $M_{sus}$ as either illegal or legal. Specifically, if the training data for $M_{sus}$ originates from the defender's generative model $G_d$ , then $M_{sus}$ should be classified as illegal; otherwise, $M_{sus}$ is considered legal. Given that this is a classification task, we selected three commonly used metrics for evaluation: accuracy, F1 score, and AUROC.", + "bbox": [ + 511, + 606, + 903, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Baselines", + "text_level": 1, + "bbox": [ + 513, + 724, + 620, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since TrainProVe is the first attempt at addressing the introduced training data provenance verification problem, there are no existing methods for direct comparison. Nonetheless, there are several baselines that can be utilized to showcase the efficacy of the proposed method, as enumerated below:", + "bbox": [ + 511, + 747, + 903, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Random classification: a rather trivial baseline that randomly classify suspicious models as illegal or legal.", + "bbox": [ + 511, + 824, + 903, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Han et al.'s work [11]: This is the closest method to ours, which trains an attributor to classify suspicious models. The underlying assumption of this method is that the defender", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23821", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ada5311c6b21f180f4d0bc2a3b11191e8de759ebad893b70fefa6c45030d3269.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
The Suspect's Data SourcesTPFPFNTNAccuracyF1 ScoreAUROC
The Defender's Generative Model GdData Sources Unrelated to Gd
Stable Diffusion [38] v1.4Latent Consistency Model064
PixArt-α6400640.9880.9700.992
Stable Cascade064
Real Data460
Latent Consistency Model [28]Stable Diffusion v1.4064
PixArt-α6400641.0001.0001.000
Stable Cascade064
Real Data064
PixArt-α [4]Stable Diffusion v1.4064
Latent Consistency Model6400641.0001.0001.000
Stable Cascade064
Real Data064
Stable Cascade [34]Stable Diffusion v1.4064
Latent Consistency Model6400641.0001.0001.000
PixArt-α064
Real Data064
Average Value0.9970.9920.998
", + "bbox": [ + 96, + 88, + 898, + 356 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. The results of TrainProVe on CIFAR10. The \"64\" in the table indicates the 64 suspicious models trained based on a data source, with different training hyperparameters. Suspicious models predicted as illegal/legal are considered positive/negative instances. If the suspicious model's training data source is $G_{d}$ , it is illegal, otherwise, it is innocent.", + "bbox": [ + 89, + 362, + 906, + 405 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "possesses multiple different generative models to train the attributor, while in our scenario, the defender has only a single generative model $G_{d}$ . To evaluate Han et al.'s method, we relaxed the constraints on the defender by allowing the defender to access all four text-to-image models.", + "bbox": [ + 88, + 412, + 480, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TrainProVe-Ent: This method performs hypothesis testing by analyzing the entropy of the predicted logits generated by models $M_{sdw}$ and $M_{sus}$ . Specifically, we analyze $D_v$ in batches using models $M_{sdw}$ and $M_{sus}$ to derive their respective logits, and subsequently compute their entropies $\\mathcal{E}_{sdw}$ and $\\mathcal{E}_{sus}$ . Following this, we employ a one-sided Grubbs' hypothesis test to ascertain whether the average of all elements in $\\mathcal{E}_{sus}$ is an outlier with a high value compared to $\\mathcal{E}_{sdw}$ . This analysis aids in determining the validity of model $M_{sus}$ , with its parameter configurations being consistent with those of TrainProVe.", + "bbox": [ + 88, + 489, + 482, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TrainProVe-Sim: This method conducts hypothesis testing based on the pairwise similarity of logits within the same category from $M_{sdw}$ and $M_{sus}$ . Specifically, $D_v$ is processed in batches using both $M_{sdw}$ and $M_{sus}$ to obtain their respective logits. Subsequently, the pairwise cosine similarity within the same category is computed to obtain the similarity sets $S_{sdw}$ and $S_{sus}$ . Following this, a one-sided Grubbs' hypothesis test is conducted to determine whether the average of all elements in $S_{sus}$ , is an outlier of low value in $S_{sdw}$ , thereby evaluating the validity of $M_{sus}$ . The parameter configurations align with those in TrainProVe.", + "bbox": [ + 88, + 655, + 482, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Experimental Results", + "text_level": 1, + "bbox": [ + 89, + 832, + 294, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this paper, $G_{d}$ is one of the four text-to-image models (Stable Diffusion v1.4, latent consistency model, PixArt-α, and Stable Cascade), and the training dataset of $M_{sus}$", + "bbox": [ + 89, + 854, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "comes from one of five data sources (Stable Diffusion v1.4, latent consistency model, PixArt- $\\alpha$ , Stable Cascade and real data). We first calculate the accuracy, F1 score, and AUROC under the four different $G_{d}$ , and then average the metrics across the four scenarios as the final metrics for comparing different methods. Taking the results of TrainProVe on CIFAR10 as an example (more experimental results can be found in the supplementary material), as shown in Tab. 1. It is evident that TrainProVe can accurately classify suspicious models trained on various data sources.", + "bbox": [ + 511, + 412, + 906, + 564 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results of various methods on different datasets are illustrated in Tab. 2, showcasing the superiority of TrainProVe. In comparison to the research by Han et al., TrainProVe has three advantages: (1) Han et al.'s approach necessitates the defender to have access to multiple different generative models, while TrainProVe is effective even when the defender only has his own generative model $G_{d}$ ; (2) The efficacy of Han et al.'s method consistently lags behind that of TrainProVe across all cases. This is because the former relies on an attributor for classification, which tends to overfit to $M_{sdw}$ 's architecture used by the defender. If the architecture of $M_{sus}$ differs, this method is likely to fail. In contrast, TrainProVe relies on hypothesis testing to classify $M_{sus}$ , thus avoiding this risk; (3) Han et al.'s method requires training a large number of models based on synthetic datasets from each generative model to train the attributor. In contrast, TrainProVe only requires training a single ResNet18 on one synthetic dataset generated by $G_{d}$ .", + "bbox": [ + 511, + 565, + 908, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Moreover, compared to TrainProVe-Ent and TrainProVe-Sim, the efficacy of TrainProVe consistently surpasses that of the others. This is because it compares the predictions of $M_{sus}$ and $M_{sdw}$ with the true labels to obtain accuracy,", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23822", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3f036a589ca4fe0c4dbcdef5917402d7484249077bb0c64802d02628522526fa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetMethodAvg. AccAvg. F1Avg. AUC
CIFAR10Random0.5000.2860.500
Han et al.'s Work0.7680.5790.731
TrainProVe-Sim0.8120.6470.821
TrainProVe-Ent0.8680.7740.902
TrainProVe0.9970.9920.998
CIFAR100Random0.5000.2860.500
Han et al.'s Work0.7300.5480.703
TrainProVe-Sim0.5750.3180.545
TrainProVe-Ent0.3980.4080.624
TrainProVe0.9920.9790.981
ImageNet-100Random0.5000.2860.500
Han et al.'s Work0.6710.4250.645
TrainProVe-Sim0.5550.3050.528
TrainProVe-Ent0.3560.3640.567
TrainProVe0.7860.4950.769
", + "bbox": [ + 91, + 88, + 483, + 319 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/67d256aab1f4fc0c17ab707cde1355de8d1ff8065c5ba9914bef3e8c67d064f0.jpg", + "table_caption": [ + "Table 2. The results of TrainProVe and baselines on different datasets. \"Avg. Acc\", \"Avg. F1\", and \"Avg. AUC\" are the average values of accuracy, F1 score, and AUROC under four different $G_{d}$ , similar to Tab. 1. Bold represents the best results." + ], + "table_footnote": [], + "table_body": "
DatasetMsdwAvg. AccAvg. F1Avg. AUC
CIFAR10ResNet18 [12]0.9970.9920.998
ConvNeXt-B [26]0.9950.9890.997
Swin-B [25]0.9970.9920.998
ImageNet-100ResNet180.7860.4950.769
ConvNeXt-B0.7710.4700.756
Swin-B0.7870.4870.766
", + "bbox": [ + 91, + 387, + 478, + 488 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "using this as the basis for hypothesis testing. In contrast, TrainProVe-Ent and TrainProVe-Sim don't compare model predictions with true labels, resulting in a lack of information from the true labels, which leads to inferior validation performance compared to TrainProVe.", + "bbox": [ + 89, + 537, + 483, + 613 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 625, + 254, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The model architecture of $M_{sdw}$ . We substituted the standard architecture of $M_{sdw}$ , ResNet18, with ConvNeXt-B [26] and Swin-B [25], while leaving all other configurations unaltered. The outcomes are displayed in Tab. 3, demonstrating that TrainProVe remains efficient irrespective of whether the architecture of the shadow model is based on CNNs or transformers.", + "bbox": [ + 89, + 652, + 482, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The training epochs of $M_{sdw}$ . We explored TrainProVe's performance as the $M_{sdw}$ 's training epochs (default is 100) increased from 0 to 200 in increments of 20. The findings on CIFAR10, depicted in Fig. 4, suggest that TrainProVe exhibits a commendable resilience to variations in the number of training epochs for $M_{sdw}$ . It is worth noting that an excessive or insufficient number of training epochs can lead to overfitting or underfitting of $M_{sdw}$ on $\\mathcal{D}_{sdw}$ , consequently resulting in a decline in TrainProVe's performance.", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6689acce4050831022ce1e58c3ac38899680adc6e291df407ef698ca65df3bc6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 87, + 903, + 155 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/60c29107c2a6c838fbf6210cebdfca2c4250deca4f53e067d0ecb805c06186c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 159, + 903, + 227 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d2367a8550fe350ea4dae8c9501bd950d7a038d21d9a2d0cd23289497916b5bc.jpg", + "image_caption": [ + "Figure 4. Changing $M_{sdw}$ 's training epochs. \"Avg. Acc\", \"Avg. F1\", and \"Avg. AUC\" are the average values of accuracy, F1 score, and AUROC under four different $G_d$ ." + ], + "image_footnote": [], + "bbox": [ + 514, + 233, + 901, + 299 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bc61decd7e80bb865991b90b29a96903e0ad9f71a8862d011f130c0fcf4e0cae.jpg", + "table_caption": [ + "Table 3. Changing $M_{sdw}$ 's architecture. \"Avg. Acc\", \"Avg. F1\", and \"Avg. AUC\" are the average values of accuracy, F1 score, and AUROC under four different $G_d$ , similar to Tab. 1." + ], + "table_footnote": [], + "table_body": "
Text PromptsAccuracyF1 ScoreAUROC
Original Prompts0.9880.9700.992
Swapped Prompts0.9880.9700.992
", + "bbox": [ + 531, + 349, + 887, + 397 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Swapping $\\mathcal{T}_{sdw}$ and $\\mathcal{T}_v$ . Note that Stable Diffusion v1.4 is $G_{d}$ . Latent consistency model, PixArt- $\\alpha$ , Stable Cascade, and real data represent other data sources.", + "bbox": [ + 511, + 407, + 905, + 449 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The text prompts of $\\mathcal{D}_{sdw}$ and $\\mathcal{D}_v$ . We swapped the text prompts, setting $\\mathcal{T}_{sdw}' = \\mathcal{T}_v$ and $\\mathcal{T}_v' = \\mathcal{T}_{sdw}$ , and used $\\mathcal{T}_{sdw}'$ and $\\mathcal{T}_v'$ to generate $\\mathcal{D}_{sdw}$ and $\\mathcal{D}_v$ , respectively. The results on CIFAR10, shown in Tab. 4, indicate that TrainProVe is robust to the text prompts of $\\mathcal{D}_{sdw}$ and $\\mathcal{D}_v$ .", + "bbox": [ + 511, + 452, + 905, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The sample sizes of $\\mathcal{D}_{sdw}$ and $\\mathcal{D}_v$ . We generated $\\mathcal{D}_{sdw}$ and $\\mathcal{D}_v$ with different sample sizes for TrainProVe, keeping other settings unchanged. Note that the default sample sizes for $\\mathcal{D}_{sdw}$ and $\\mathcal{D}_v$ are 30,000 and 5,000 respectively, when generating CIFAR10. The results on CIFAR10 shown in Fig. 5 demonstrate that TrainProVe has good robustness to the sample sizes. Moreover, when $\\mathcal{D}_{sdw}$ 's sample size is too small, the generalization performance of $M_{sdw}$ on $\\mathcal{D}_v$ may degrade, leading to a decline in TrainProVe's performance.", + "bbox": [ + 511, + 532, + 906, + 669 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. The Interference Resistance of TrainProVe", + "text_level": 1, + "bbox": [ + 511, + 676, + 877, + 693 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we study that whether TrainProVe remains effective when facing more covert data theft. More experimental details can be found in the supplementary material.", + "bbox": [ + 511, + 700, + 905, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multiple training data sources for $M_{\\text{sus}}$ . We investigated the robustness of TrainProVe when the training dataset of $M_{\\text{sus}}$ has multiple data sources. Specifically, we assume that the training dataset of $M_{\\text{sus}}$ contains both synthetic data from $G_d$ and real data. The results on CIFAR10 are shown in Tab. 5, where the training dataset for $M_{\\text{sus}}$ consists of 120,000 images, with synthetic data and real data each accounting for half. We calculated the accuracy of TrainProVe in classifying $M_{\\text{sus}}$ trained on this mixed dataset. The results indicate that TrainProVe can still accu", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23823", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1b5c18500654067f5d943d907dcd43bd092b42f3c4d0d3b40a9845c51dfabf47.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 85, + 281, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f3af5e167b53f5c4c367d3d557967d57d66578abb99784cb231aaddbc3a9ab1d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 85, + 475, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8eb15b5d9bab2482c6441a65d56889b4c861e9b60b407c044fa302ed847e4ebd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 193, + 281, + 294 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9970fb00da91e6c00d309289c2a107d9fb74c8646c584294d0b8cfd46478d397.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 193, + 475, + 294 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/87135b8a894af98385ddac4a869ff2e93dda05c198959ce56ed95d47f9cdfe85.jpg", + "image_caption": [ + "Figure 5. Changing the sample sizes of $\\mathcal{D}_{sdw}$ and $\\mathcal{D}_v$ . \"Avg. Acc\", \"Avg. F1\", and \"Avg. AUC\" are the average values of accuracy, F1 score, and AUROC under four different $G_{d}$ ." + ], + "image_footnote": [], + "bbox": [ + 89, + 303, + 281, + 402 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3ab8185a87e9ef6857b882fab178a5f164783fd31590823f3b7fd9b00d492afc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 303, + 475, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "rately detect illegal behavior in this more difficult situation.", + "bbox": [ + 89, + 458, + 482, + 473 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fine-tuning $G_{d}$ . When the suspect has access to a small amount of real data, he may fine-tune the defender's open-source model $G_{d}$ and then generate data to train $M_{sus}$ , which can enhance $M_{sus}$ 's generalization ability on real data. Note that the defender only possesses the original $G_{d}$ and has no knowledge of fine-tuning. Specifically, we fine-tuned the pre-trained Stable Diffusion v1.4 using 10,000 ImageNet-100 images (100 per class) and LoRA [16] with 10 epochs, and then generated ImageNet-100-Syn to train $M_{sus}$ . The results are shown in Tab. 6a, which indicate that TrainProVe remains effective in this more arduous scenario.", + "bbox": [ + 89, + 479, + 483, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fine-tuning $M_{sus}$ . Similarly, when the suspect has a small amount of real data, he may also fine-tune $M_{sus}$ with real data, which was trained on synthetic data from $G_{d}$ , to enhance $M_{sus}$ 's generalization ability on real data. Specifically, we fine-tuned $M_{sus}$ using 10,000 ImageNet-100 images (100 per class), where $M_{sus}$ was pre-trained on synthetic data from Stable Diffusion v1.4, with 10 fine-tuning epochs. The results, shown in Tab. 6b, indicate even in this more challenging scenario, TrainProVe remains effective.", + "bbox": [ + 89, + 651, + 482, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Efficiency Analysis", + "text_level": 1, + "bbox": [ + 89, + 801, + 272, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We analyzed the efficiency of TrainProVe and the work of Han et al. on ImageNet-100. Specifically, following the setup in Sec. 4.1, given $M_{sus}$ trained on ImageNet-100 or ImageNet-100-Syn, we performed detection using both methods on an NVIDIA GeForce RTX 4090. $G_{d}$ is Sta", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f0cf2a6124ff3acbe5ecee7377cb8de541f1b8b81c235f22860e2df159f13a8d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
M_sus's Training SourceG_dTPFPAcc
CIFAR10-SynStable Diffusion v1.46401
Latent Consistency Model6401
+CIFAR10PixArt-α6401
Stable Cascade6401
", + "bbox": [ + 514, + 87, + 903, + 162 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f1eb64584a1684f044566acd0bd17bd7a38e61e35c80922f6d91bde26187bfd4.jpg", + "table_caption": [ + "Table 5. Results of TrainProVe when the training dataset of $M_{sus}$ has multiple data sources. The \"64\" in the table indicates the 64 suspicious models with different training hyperparameters." + ], + "table_footnote": [], + "table_body": "
TPFPAcc
w/o fine-tuning6401.00
w/ fine-tuning6220.97
", + "bbox": [ + 514, + 214, + 705, + 261 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7d61640c3ffd53daf75ec08765d529296541c520faac3099f5440aa5ce08aa98.jpg", + "table_caption": [ + "(a) Fine-tuning $G_{d}$" + ], + "table_footnote": [], + "table_body": "
TPFPAcc
w/o fine-tuning6401.00
w/ fine-tuning5770.89
", + "bbox": [ + 715, + 214, + 903, + 262 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9114a6b52beb51222be758a16fd3122044ef7a312441b269e2f803e480199b36.jpg", + "table_caption": [ + "(b) Fine-tuning $M_{sus}$", + "Table 6. Fine-tuning $G_{d}$ or $M_{sus}$ . The \"64\" in the table indicates the 64 suspicious models with different training hyperparameters." + ], + "table_footnote": [], + "table_body": "
MethodτgτtτvTotalAccF1AUC
Han et al.'s Work137.7h8.0h1.2h146.9h0.480.280.48
TrainProVe33.6h0.1h1.4h35.1h0.780.630.84
", + "bbox": [ + 514, + 315, + 903, + 361 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Efficiency analysis. \"Acc\", \"F1\" and \"AUC\" refer to the accuracy, F1 score and AUROC respectively.", + "bbox": [ + 511, + 363, + 903, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ble Cascade. The runtime and results is shown in Tab. 7. For TrainProVe, $\\tau_{g}$ is the time for generating $D_{sdw}$ and $D_v$ , $\\tau_t$ is the time for training $M_{sdw}$ , and $\\tau_v$ is the time for classifying $M_{sus}$ using hypothesis testing. For Han et al.'s method, $\\tau_{g}$ is the time for generating the training and validation dataset, $\\tau_t$ is the time for training the models and the attributor, and $\\tau_v$ is the time for classifying $M_{sus}$ using the attributor. From the results, it shows that the high efficiency of TrainProVe lies in its significantly smaller $\\tau_{g}$ and $\\tau_t$ compared to Han et al.'s method. This is because TrainProVe only requires generating one dataset ( $D_{sdw}$ ) and training one model ( $M_{sdw}$ ). In contrast, Han et al.'s method requires generating $m$ datasets and training $m \\times n$ models, where $m$ is the number of text-to-image models accessible to the defender and $n$ is the number of models trained by the defender on each dataset. These models are then used to train the attributor. Moreover, TrainProVe achieved better classification results, further illustrating its superiority.", + "bbox": [ + 511, + 407, + 906, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 693, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Open-source text-to-image models have reduced the cost of obtaining images. We introduce an important yet largely unexplored security issue: whether a suspicious model's training data originates from a specific text-to-image model? To our knowledge, we are the first to propose a solution to this problem, called TrainProVe. The success of TrainProVe in verifying training data provenance has been demonstrated through theory and experiments. Promising future work includes: (1) extending TrainProVe to other image generative models, such as image-to-image models; (2) adapting TrainProVe to protect other types of data (e.g., text); (3) exploring other privacy risks about text-to-image models.", + "bbox": [ + 511, + 719, + 906, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23824", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 263, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partially supported by the Pioneer R&D Program of Zhejiang (No.2024C01021), and Zhejiang Province High-Level Talents Special Support Program \"Leading Talent of Technological Innovation of Ten-Thousands Talents Program\" (No. 2022R52046).", + "bbox": [ + 89, + 114, + 483, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 205, + 187, + 222 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hritik Bansal and Aditya Grover. Leaving reality to imagination: Robust classification via generated datasets. arXiv preprint arXiv:2302.02503, 2023. 1", + "[2] Shai Ben-David, John Blitzer, Koby Crammer, and Fernando Pereira. Analysis of representations for domain adaptation. Advances in neural information processing systems, 19, 2006. 3, 1", + "[3] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- $\\sigma$ : Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 1", + "[4] Junsong Chen, YU Jincheng, GE Chongjian, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- $\\alpha$ : Fast training of diffusion transformer for photorealistic text-to-image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 5, 6, 4, 7, 8", + "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5, 2", + "[6] Tim Dockhorn, Robin Rombach, Andreas Blatmann, and Yaoliang Yu. Distilling the knowledge in diffusion models. In CVPR Workshop Generative Models for Computer Vision, 2023. 1", + "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5", + "[8] Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7382-7392, 2024. 1", + "[9] Abolfazl Farahani, Sahar Voghoei, Khaled Rasheed, and Hamid R Arabnia. A brief review of domain adaptation. Advances in data science and information engineering: proceedings from ICDATA 2020 and IKE 2020, pages 877-894, 2021. 3", + "[10] Frank Ephraim Grubbs. Sample criteria for testing outlying observations. University of Michigan, 1949. 4", + "[11] Ge Han, Ahmed Salem, Zheng Li, Shanqing Guo, Michael Backes, and Yang Zhang. Detection and attribution of models trained on generated data. In ICASSP 2024-2024 IEEE" + ], + "bbox": [ + 93, + 231, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4875-4879. IEEE, 2024. 5", + "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5, 7", + "[13] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is synthetic data from generative models ready for image recognition? arXiv preprint arXiv:2210.07574, 2022. 1", + "[14] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 1", + "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2", + "[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 8", + "[17] Yongcheng Jing, Yiding Yang, Xinchao Wang, Mingli Song, and Dacheng Tao. Turning frequency to resolution: Video super-resolution via event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7772-7781, 2021. 2", + "[18] Yongcheng Jing, Chongbin Yuan, Li Ju, Yiding Yang, Xinchao Wang, and Dacheng Tao. Deep graph reprogramming. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24345-24354, 2023. 2", + "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2", + "[20] A Krizhevsky. Learning multiple layers of features from tiny images. Master's thesis, University of Tront, 2009. 5, 2", + "[21] Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pretrained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems, 36, 2023. 1", + "[22] T Lin. Focal loss for dense object detection. arXiv preprint arXiv:1708.02002, 2017. 5", + "[23] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2", + "[24] Yugeng Liu, Zheng Li, Michael Backes, Yun Shen, and Yang Zhang. Watermarking diffusion model. arXiv preprint arXiv:2305.12502, 2023. 1, 3", + "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 7", + "[26] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the" + ], + "bbox": [ + 514, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23825", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 7", + "[27] Ge Luo, Junqiang Huang, Manman Zhang, Zhenxing Qian, Sheng Li, and Xinpeng Zhang. Steal my artworks for fine-tuning? a watermarking framework for detecting art theft mimicry in text-to-image models. arXiv preprint arXiv:2311.13619, 2023. 3", + "[28] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 1, 3, 5, 6, 2, 4, 7, 8", + "[29] Kidist Amde Mekonnen, Nicola Dall'Asen, and Paolo Rota. Adv-kd: Adversarial knowledge distillation for faster diffusion sampling. arXiv preprint arXiv:2405.20675, 2024. 1", + "[30] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41, 1995. 2", + "[31] Rui Min, Sen Li, Hongyang Chen, and Minhao Cheng. A watermark-conditioned diffusion model for ip protection. arXiv preprint arXiv:2403.10893, 2024. 1", + "[32] Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions on knowledge and data engineering, 22(10):1345-1359, 2009. 3", + "[33] Suraj Patil, William Berman, Robin Rombach, and Patrick von Platen. amused: An open muse reproduction. arXiv preprint arXiv:2401.01808, 2024. 1", + "[34] Pablo Pernias, Dominic Rampas, Mats L. Richter, Christopher J. Pal, and Marc Aubreville. Wuerstchen: An efficient architecture for large-scale text-to-image diffusion models. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 5, 6, 4, 7, 8", + "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 5, 3", + "[36] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1", + "[37] Mauro Ribeiro, Katarina Grolinger, and Miriam AM Capretz. Mlaas: Machine learning as a service. In 2015 IEEE 14th international conference on machine learning and applications (ICMLA), pages 896-902. IEEE, 2015. 2", + "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3, 5, 6, 4, 7, 8", + "[39] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In International Conference on Learning Representations, 2022. 1", + "[40] Mert Bülent Sariyildüz, Karteek Alahari, Diane Larlus, and Yannis Kalantidis. Fake it till you make it: Learning transferable representations from synthetic imagenet clones. In" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8011-8021, 2023. 1, 2, 5", + "[41] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2", + "[42] Jordan Shipard, Arnold Wiliem, Kien Nguyen Thanh, Wei Xiang, and Clinton Fookes. Diversity is definitely needed: Improving model-agnostic zero-shot classification via stable diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 769–778, 2023. 1, 2, 5", + "[43] Geng Sun, Tingru Cui, Jianming Yong, Jun Shen, and Shiping Chen. Mlaas: a cloud-based system for delivering adaptive micro learning in mobile mooc learning. IEEE Transactions on Services Computing, 11(2):292-305, 2015. 2", + "[44] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2117-2126, 2020. 1", + "[45] Lisa Torrey and Jude Shavlik. Transfer learning. In Handbook of research on machine learning applications and trends: algorithms, methods, and techniques, pages 242-264. IGI global, 2010. 3", + "[46] Mei Wang and Weihong Deng. Deep visual domain adaptation: A survey. Neurocomputing, 312:135-153, 2018. 3", + "[47] Zhenting Wang, Chen Chen, Yuchen Liu, Lingjuan Lyu, Dimitris Metaxas, and Shiqing Ma. How to detect unauthorized data usages in text-to-image diffusion models. arXiv preprint arXiv:2307.03108, 2023. 1", + "[48] Zhenting Wang, Chen Chen, Yi Zeng, Lingjuan Lyu, and Shiqing Ma. Where did i come from? origin attribution of ai-generated images. Advances in neural information processing systems, 36, 2023. 1", + "[49] Qizhen Weng, Wencong Xiao, Yinghao Yu, Wei Wang, Cheng Wang, Jian He, Yong Li, Liping Zhang, Wei Lin, and Yu Ding. Mlaas in the wild: Workload analysis and scheduling in large-scale heterogeneousgpu clusters. In 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI 22), pages 945–960, 2022. 2", + "[50] Qianlong Xiang, Miao Zhang, Yuzhang Shang, Jianlong Wu, Yan Yan, and Liqiang Nie. Dkdm: Data-free knowledge distillation for diffusion models with any architecture. arXiv preprint arXiv:2409.03550, 2024. 1", + "[51] Yuechen Xie, Jie Song, Mengqi Xue, Haofei Zhang, Xingen Wang, Bingde Hu, Genlang Chen, and Mingli Song. Dataset ownership verification in contrastive pre-trained models. arXiv preprint arXiv:2502.07276, 2025.3", + "[52] Cheng Xiong, Chuan Qin, Guorui Feng, and Xinpeng Zhang. Flexible and secure watermarking for latent diffusion model. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1668-1676, 2023. 1", + "[53] Jianhao Yuan, Francesco Pinto, Adam Davies, and Philip Torr. Not just pretty pictures: Toward interventional data" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23826", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "augmentation using text-to-image generators. arXiv preprint arXiv:2212.11237, 2022. 1", + "[54] Zihan Yuan, Li Li, Zichi Wang, and Xinpeng Zhang. Watermarking for stable diffusion models. IEEE Internet of Things Journal, 2024. 1", + "[55] Likun Zhang, Hao Wu, Lingcui Zhang, Fengyuan Xu, Jin Cao, Fenghua Li, and Ben Niu. Training data attribution: Was your model secretly trained on data created by mine? arXiv preprint arXiv:2409.15781, 2024. 1, 3", + "[56] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Ngai-Man Cheung, and Min Lin. A recipe for watermarking diffusion models. arXiv preprint arXiv:2303.10137, 2023. 1, 3", + "[57] Yongchao Zhou, Hshmat Sahak, and Jimmy Ba. Training on thin air: Improve image classification with generated data. arXiv preprint arXiv:2305.15316, 2023. 1, 2", + "[58] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2020. 3" + ], + "bbox": [ + 91, + 92, + 483, + 375 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23827", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Training Data Provenance Verification_ Did Your Model Use Synthetic Data from My Generative Model for Training_/95a43e45-ca4b-40d2-9bb6-79d7e388dedb_model.json b/2025/Training Data Provenance Verification_ Did Your Model Use Synthetic Data from My Generative Model for Training_/95a43e45-ca4b-40d2-9bb6-79d7e388dedb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e266d001d789ebab3c7f530377c2235051be5b7a --- /dev/null +++ b/2025/Training Data Provenance Verification_ Did Your Model Use Synthetic Data from My Generative Model for Training_/95a43e45-ca4b-40d2-9bb6-79d7e388dedb_model.json @@ -0,0 +1,2356 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.119, + 0.13, + 0.882, + 0.177 + ], + "angle": 0, + "content": "Training Data Provenance Verification: Did Your Model Use Synthetic Data from My Generative Model for Training?" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.203, + 0.735, + 0.224 + ], + "angle": 0, + "content": "Yuechen Xie\\(^{1}\\), Jie Song\\(^{1*}\\), Huiqiong Wang\\(^{2}\\), Mingli Song\\(^{1,3,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.417, + 0.228, + 0.584, + 0.247 + ], + "angle": 0, + "content": "\\(^{1}\\)Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.247, + 0.69, + 0.265 + ], + "angle": 0, + "content": "\\(^{2}\\)Ningbo Innovation Center, Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.265, + 0.78, + 0.282 + ], + "angle": 0, + "content": "3State Key Laboratory of Blockchain and Security, Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.181, + 0.282, + 0.817, + 0.3 + ], + "angle": 0, + "content": "\\(^{4}\\)Hangzhou High-Tech Zone (Binjiang) Institute of Blockchain and Data Security" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.309, + 0.715, + 0.324 + ], + "angle": 0, + "content": "{xyuechen,sjie,huiqiong-wang,brooksong}@zju.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.359, + 0.327, + 0.374 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.392, + 0.486, + 0.725 + ], + "angle": 0, + "content": "High-quality open-source text-to-image models have lowered the threshold for obtaining photorealistic images significantly, but also face potential risks of misuse. Specifically, suspects may use synthetic data generated by these generative models to train models for specific tasks without permission, when lacking real data resources especially. Protecting these generative models is crucial for the wellbeing of their owners. In this work, we propose the first method to this important yet unresolved issue, called Training data Provenance Verification (TrainProVe). The rationale behind TrainProVe is grounded in the principle of generalization error bound, which suggests that, for two models with the same task, if the distance between their training data distributions is smaller, their generalization ability will be closer. We validate the efficacy of TrainProVe across four text-to-image models (Stable Diffusion v1.4, latent consistency model, PixArt-α, and Stable Cascade). The results show that TrainProVe achieves a verification accuracy of over 99% in determining the provenance of suspicious model training data, surpassing all previous methods. Code is available at https://github.com/xieyc99/TrainProVe." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.756, + 0.222, + 0.772 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.783, + 0.485, + 0.875 + ], + "angle": 0, + "content": "High-quality open-source text-to-image models based on diffusion models [15] can generate photorealistic images with simple text prompts [4, 28, 34, 38], making them highly significant in terms of lowering creative barriers, enabling personalized customization, etc. Currently, most open-source text-to-image models are only permitted for" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.361, + 0.907, + 0.435 + ], + "angle": 0, + "content": "academic or educational purposes, and commercial use is prohibited without explicit permission [3, 4, 21, 28, 33, 34, 36, 38]. Therefore, protecting the intellectual property of these generative models is crucial for safeguarding the interests of their owners." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.438, + 0.907, + 0.514 + ], + "angle": 0, + "content": "Currently, one of the major security challenges faced by defenders, i.e., the generative model owners, is the unauthorized use of the synthetic data from their models by suspects [24, 47, 48, 55]. As shown in Fig. 1, based on the suspect's intention, this issue can be divided into three cases:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.906, + 0.562 + ], + "angle": 0, + "content": "Case 1: the suspect steals the synthetic data illegally and claims it as his own personal work [24, 44, 48], implying the intellectual property of the generative model is infringed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.907, + 0.655 + ], + "angle": 0, + "content": "Case 2: the suspect steals the synthetic data from the defender's generative model, and trains another generative model for the same task using knowledge distillation techniques [6, 14, 29, 39, 50]. In this case, the defender cannot access the synthetic data used by the suspect, making it a more challenging defense task compared to Case 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.907, + 0.778 + ], + "angle": 0, + "content": "Case 3: the suspect steals the synthetic data from the generative model to train another model for a specific task [1, 8, 13, 40, 42, 53, 57]. In this case, the defender is not only unaware of the training data used by the suspicious model, but also cannot compare the similarity between the generative model and the suspicious model directly due to the different tasks, representing an even more arduous defense task compared to Case 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.903 + ], + "angle": 0, + "content": "For Case 1 and Case 2, there are already some mature methods to address these issues, such as adding watermarks [24, 31, 52, 54, 56] and training data attribution [55], which identify illegal activities by checking the distributions of the suspicious images or the synthetic images from suspicious model. However, for Case 3, although the potential of using synthetic data to train models for other tasks has been recognized gradually in recent years [13, 40, 42, 53]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.235, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23817" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.452, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.522, + 0.484, + 0.592 + ], + "angle": 0, + "content": "Figure 1. The overview of the three cases. These three cases encompass nearly every conceivable real-world scenario, where a suspect uses synthetic images from the defender's text-to-image model illegally. In this paper, we are committed to addressing the security risks in Case 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.484, + 0.716 + ], + "angle": 0, + "content": "due to the lower collection costs compared to real data, the associated security risks have been overlooked. Methods that work for the first two cases cannot be applied to the third case, because the learning task of the suspicious model is not necessarily image generation. For example, for a suspicious image classification model, there are no synthetic data available for defenders to examine." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.903 + ], + "angle": 0, + "content": "In this work, we address this important yet unexplored issue. To the best of our knowledge, we propose the first method for protecting the intellectual property of generative models in such scenario, which we call Training data Provenance Verification (TrainProVe). It helps defenders verify whether the suspicious model has been trained using synthetic data from their generative model. Given that suspicious models may use synthetic data for training without authorization, and is used for commercial purposes, we focus on the black-box setting, where the defenders have no information about the training configuration of the suspicious model (e.g., loss function, model architecture, etc.)" + }, + { + "type": "aside_text", + "bbox": [ + 0.462, + 0.091, + 0.482, + 0.474 + ], + "angle": 90, + "content": "Low Defense Difficulty High" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "and can only access the model through Machine Learning as a Service (MLaaS) [37, 43, 49]. This means that the defenders can only obtain prediction results (such as prediction logits or labels) via the suspicious model's API." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.154, + 0.907, + 0.456 + ], + "angle": 0, + "content": "The rationale behind TrainProVe is grounded in the principle of generalization error bound, which suggests that, for two models with the same task, if the distance between their training data distributions is smaller, their generalization ability will be closer. Specifically, defenders can train a model for the same task as the suspicious model, using the synthetic data generated by their generative model. By comparing the performances of these two models, defenders can determine whether the suspicious model was trained on synthetic data from their generative model. More concretely, the proposed TrainProVe involves three steps: (1) using the defender's generative model along with different text prompts to generate a shadow dataset and a validation dataset, based on the suspicious model's task; (2) training a shadow model for the same task as the suspicious model using the shadow dataset; (3) performing hypothesis testing on the prediction accuracy of the suspicious model and the shadow model on the validation dataset, in order to determine whether the suspicious model was trained using synthetic data from the defender's generative model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.458, + 0.909, + 0.701 + ], + "angle": 0, + "content": "In summary, the principal contributions of this paper are threefold: (1) we introduce the important yet under-explored issue of intellectual property protection for text-to-image models, in which the suspect uses the synthetic data from the defender's generative model illegally to train a specific-task suspicious model, such as an image classification model. For the defender, the suspicious model is a black-box; (2) we discover and demonstrate that the similarity between different models is related to their source domains closely, and based on this, we propose the training data provenance verification technique. To the best of our knowledge, this is the first method to address such intellectual property protection issue; (3) Extensive experiments show that TrainProVe achieves over \\(99\\%\\) accuracy in verifying training data provenance for suspicious models, surpassing all previous studies." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.718, + 0.655, + 0.735 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Text-to-image diffusion model. Diffusion models [15] have revolutionized generative modeling fundamentally, in terms of quality and generalization. Their excellent scalability has led to significant advancements in specific applications, such as text-to-image generation. Text-to-image diffusion models [4, 34, 38] use the diffusion process to transform text descriptions into realistic images. Recently, people have begun to recognize the potential of text-to-image diffusion models for synthesizing dataset [17, 18, 40, 42, 57]. However, so far, no one has realized that suspects" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23818" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "might use these text-to-image diffusion models to synthesize datasets and train their own models without authorization. Therefore, in this paper, we propose the first feasible solution to address this issue, securing and fostering healthy development in this field." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.172, + 0.483, + 0.399 + ], + "angle": 0, + "content": "Text-to-image model protection. Training generative models requires large amounts of data and computational resources [4, 38], making the intellectual property of these models crucial for their owners. When these models are open-source [4, 28, 34, 38], their intellectual property is vulnerable to infringement by suspects. Specifically, for text-to-image models, suspects have the following three ways primarily to infringe on their intellectual property: (1) stealing synthetic data illegally; (2) illegal knowledge distillation; (3) training suspicious models for specific tasks using synthetic data illegally. For the first two cases, watermarking [24, 27, 56] and training data attribution [55] have been developed to address these issues. However, for the third case, no suitable solution exists currently. This work aims to address this gap." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.484, + 0.539 + ], + "angle": 0, + "content": "Generalization error bound. Generalization error (upper) bound is used to describe a model's performance on testing data and widely applied in domain adaptation [9, 46], which is a special case of transfer learning [45, 51, 58]. Theorem 1 in [2] states the generalization error upper bound of model \\(M\\) on the target domain \\(T\\). Let \\(\\mathcal{H}\\) be a hypothesis space with VC dimension \\(d\\), and let \\(m\\) be the sample size of the dataset in the source domain \\(S\\). Then, with probability at least \\(1 - \\delta\\), for every \\(M \\in \\mathcal{H}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.548, + 0.483, + 0.6 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {T} (M) \\leq \\hat {\\epsilon} _ {S} (M) + \\sqrt {\\frac {4}{m} \\left(d \\log \\frac {2 e m}{d}\\right)} + d _ {\\mathcal {H}} \\left(P _ {S}, P _ {T}\\right) + \\lambda , \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.483, + 0.782 + ], + "angle": 0, + "content": "where \\(\\epsilon_T(M)\\) and \\(\\hat{\\epsilon_S} (M)\\) represent the generalization error of \\(M\\) on the target domain and the empirical error of \\(M\\) on the source domain, respectively. \\(P_{S}\\) and \\(P_{T}\\) denote the marginal probability distributions of the source and target domains, respectively. \\(d_{\\mathcal{H}}(P_S,P_T)\\) is the \\(\\mathcal{H}\\)-divergence between the source and target domains, which measures the distance between the two distributions within the hypothesis space \\(\\mathcal{H}\\). \\(\\lambda\\) is a constant and \\(e\\) is the base of the natural logarithm. The rationale behind TrainProVe is grounded in this theorem, which suggests that, for two models with the same task, if the distance between their training data distributions is smaller, their generalization ability will be closer." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.794, + 0.182, + 0.809 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.295, + 0.833 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In this study, we focus on the training data provenance verification task in image classification. This task involves two roles: the defender and the suspect. The defender, embodying the role of text-to-image model provider, endeavors to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "ascertain whether the black-box suspicious model, \\( M_{sus} \\), has been illicitly trained on synthetic dataset generated by his model \\( G_{d} \\) (as shown in Case 3 of Fig. 1). More specifically, for a black-box suspect model, the defender's objective is to determine whether it is illegal or legal based solely on the labels it outputs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.907, + 0.334 + ], + "angle": 0, + "content": "\\(M_{sus}\\) can be classified as either illegal or legal based on its training datasets: (1) Illegal: \\(M_{sus}\\) is trained on the synthetic data from the defender's model \\(G_{d}\\), indicating the occurrence of intellectual property infringement; (2) Legal: \\(M_{sus}\\) is trained using a different data source independent of the defender, such as some real data or synthetic data from other generative models, implying the innocence of the suspect. To be precise, when presented with a black-box suspicious model, the defender's objective is to determine whether it is an illegal or legal model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.344, + 0.731, + 0.36 + ], + "angle": 0, + "content": "3.2. The Theoretical Insight" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.366, + 0.906, + 0.517 + ], + "angle": 0, + "content": "The underlying rationale for TrainProVe is that when two models are assigned the same task, their ability to generalize will be more aligned if the divergence between their training data distributions is diminished. We demonstrate this through a case study of \\( K \\)-way classification, where \\( \\mathcal{X} = \\mathbb{R}^{C \\times W \\times H} \\) and \\( \\mathcal{Y} = \\{1, \\dots, K\\} \\) represent the input and label spaces respectively. We follow the setting of previous domain adaptation works [32, 46], viewing a domain as composed of a feature space \\( \\mathcal{X} \\) and a marginal probability distribution \\( P(\\boldsymbol{x}) \\), i.e., \\( \\{\\mathcal{X}, P(\\boldsymbol{x})\\} \\), where \\( \\boldsymbol{x} \\in \\mathcal{X} \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.518, + 0.906, + 0.643 + ], + "angle": 0, + "content": "Let \\(S_{1} = \\{\\mathcal{X},P_{1}(\\pmb {x})\\}\\) and \\(S_{2} = \\{\\mathcal{X},P_{2}(\\pmb {x})\\}\\) be two source domains sharing the same feature space but possessing distinct marginal probability distributions, i.e., \\(P_{1}(\\pmb {x})\\neq\\) \\(P_{2}(\\pmb {x}).D_{1} = \\{(x_{1}^{i},y_{1}^{i})\\}_{i = 1}^{|D_{1}|}\\) and \\(\\mathcal{D}_2 = \\{(\\pmb {x}_2^j,\\pmb {y}_2^j)\\}_{j = 1}^{|D_2|}\\) represent the datasets sampled from the source domains \\(S_{1}\\) and \\(S_{2}\\) respectively, where \\(\\pmb{x}_1^i\\sim P_1(\\pmb {x})\\) and \\(\\pmb{x}_2^j\\sim P_2(\\pmb {x})\\) . \\(M\\) and \\(\\hat{M}\\) are the models trained on \\(\\mathcal{D}_1\\) and \\(\\mathcal{D}_2\\) respectively. Then we have the following theorem:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.654, + 0.906, + 0.745 + ], + "angle": 0, + "content": "Theorem 1. Assume \\(M\\) is trained on synthetic data generated by a text-to-image model \\(G\\) with a set of text prompts \\(\\mathcal{T}_1\\), i.e., \\(P_{1}(\\pmb {x}) = P(\\pmb {x}|G,\\mathcal{T}_{1})\\). \\(\\hat{M}\\) can be trained on either real data or synthetic data generated by any text-to-image model. Based on the generalization errors of \\(M\\) and \\(\\hat{M}\\) on the target domain \\(T = \\{\\mathcal{X},P(\\pmb {x}|G,\\mathcal{T}_t)\\}\\), we have:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.757, + 0.905, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\sup _ {P _ {2} (\\boldsymbol {x}) = P (\\boldsymbol {x} \\mid G, \\mathcal {T} _ {2})} | \\Delta \\epsilon_ {T} | \\leq \\sup _ {P _ {2} (\\boldsymbol {x}) \\perp G} | \\Delta \\epsilon_ {T} |, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(\\Delta \\epsilon_{T}\\) represents the difference in generalization error between \\(M\\) and \\(\\hat{M}\\) on the target domain \\(T\\), expressed as \\(\\Delta \\epsilon_{T} = \\epsilon_{T}(M) - \\epsilon_{T}(\\hat{M})\\). \\(P_{2}(\\pmb{x}) \\perp G\\) denotes that \\(P_{2}(\\pmb{x})\\) is independent of \\(G\\), meaning \\(P_{2}(\\pmb{x}) = P(\\pmb{x}|G^{\\prime},\\mathcal{T}_{2})\\) or \\(P_{2}(\\pmb{x}) = P_{R}(\\pmb{x})\\), where \\(G^{\\prime}\\) is a text-to-image model different from \\(G\\) and \\(P_{R}(\\pmb{x})\\) represents the distribution of real data. \\(\\mathcal{T}_1, \\mathcal{T}_2\\) and \\(\\mathcal{T}_t\\) are distinct sets of text prompts." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23819" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.472, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.475, + 0.09, + 0.904, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.371, + 0.908, + 0.416 + ], + "angle": 0, + "content": "Figure 2. The overview of TrainProVe's motivation. \\(M\\) is trained on the dataset whose data distribution is \\(P(x|G, \\mathcal{T}_1)\\). \\(\\hat{M}_1\\), \\(\\hat{M}_2\\), and \\(\\hat{M}_3\\) are \\(\\hat{M}\\) trained on three datasets with different data distributions, respectively. The source and target domain data are sampled from their respective data distributions. \\(G\\) and \\(G'\\) are two different text-to-image models, and \\(\\mathcal{T}_1\\), \\(\\mathcal{T}_2\\), and \\(\\mathcal{T}_t\\) are distinct sets of text prompts." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.428, + 0.486, + 0.685 + ], + "angle": 0, + "content": "Detailed proofs can be found in the supplementary material. Theorem 1 suggests that for two models with distinct (identical) training data sources, their generalization errors differ substantially (minimally). Its contrapositive states that for two models with small (large) discrepancies in generalization error, their training data sources are identical (distinct), which supports our method. Specifically, when \\( M \\) and \\( \\hat{M} \\) are trained on synthetic data from the same text-to-image model, their accuracies on the corresponding synthetic data from this text-to-image model are more akin, compared to when their training data is derived from disparate sources. This aligns with the rationale of TrainProVe, as visually depicted in Fig. 2. The discrepancies in accuracies among these models originate from the disparities in their source domains. The closer the data distributions of the source domain data, the more analogous the accuracies of the trained models on the target domain data." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.696, + 0.33, + 0.712 + ], + "angle": 0, + "content": "3.3. The Proposed TrainProVe" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Building upon the above analysis, we propose a method called TrainProVe to protect generative models' intellectual property, as shown in Fig. 3, which consists of three stages:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.869 + ], + "angle": 0, + "content": "(1) Shadow and Validation Dataset Generation: Utilizing \\( G_{d} \\) to produce a shadow dataset \\( \\mathcal{D}_{sdw} \\) and a validation dataset \\( \\mathcal{D}_v \\), based on various text prompt sets \\( \\mathcal{T}_{sdw} \\) and \\( \\mathcal{T}_v \\), tailored to the predicted labels set \\( \\mathcal{C} \\) of the suspect model \\( M_{sus} \\). Here \\( G_{d}, \\mathcal{T}_{sdw} \\), and \\( \\mathcal{T}_v \\) correspond to \\( G, \\mathcal{T}_1, \\) and \\( \\mathcal{T}_t \\) in Fig. 2 respectively, while \\( \\mathcal{D}_{sdw} \\) and \\( \\mathcal{D}_v \\) correspond to the training data of \\( M \\) and the target domain data." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "(2) Shadow Model Training: Using the shadow dataset \\( \\mathcal{D}_{sdw} \\) to train the shadow model \\( M_{sdw} \\). The shadow model" + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.428, + 0.907, + 0.533 + ], + "angle": 0, + "content": "is analogous to \\( M \\) in Fig. 2, whereas the suspect model is denoted as \\( \\hat{M} \\). The purpose of the shadow model is to compare its performance with the suspect model trained on distinct data, focusing on their generalization capacities (accuracies) on the validation dataset. Specifically, when both models are trained on data from \\( G_{d} \\), their accuracies on the validation dataset tend to be similar." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.536, + 0.909, + 0.643 + ], + "angle": 0, + "content": "(3) Hypothesis Testing: Providing the validation dataset \\(\\mathcal{D}_v\\) to both the shadow model \\(M_{sdw}\\) and the suspect model \\(M_{sus}\\) in the same sequence batch by batch. Subsequently, calculating the accuracy of the two models batch by batch, leading to accuracy sets \\(\\mathcal{A}_{sdw}\\) and \\(\\mathcal{A}_{sus}\\). Ultimately, utilizing a one-sided Grubbs' hypothesis test [10] to ascertain whether \\(M_{sus}\\) violates the intellectual property of \\(G_d\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In the one-sided Grubbs' hypothesis test, the null hypothesis \\( H_0 \\) asserts that the mean of the set \\( \\mathcal{A}_{\\text{sus}} \\), denoted as \\( \\text{mean}(\\mathcal{A}_{\\text{sus}}) \\), is not an unusually low value among the elements of \\( \\mathcal{A}_{\\text{sdw}} \\). On the contrary, the alternative hypothesis \\( H_1 \\) posits that the mean of \\( \\mathcal{A}_{\\text{sus}} \\) is an exceptionally low value within \\( \\mathcal{A}_{\\text{sdw}} \\). Rejecting the null hypothesis \\( H_0 \\) implies that the predictive capabilities of \\( M_{\\text{sus}} \\) are significantly inferior to those of \\( M_{\\text{sdw}} \\). This suggests that the model \\( M_{\\text{sus}} \\) exhibits notably inferior generalization performance on the validation dataset compared to \\( M_{\\text{sdw}} \\), indicating that \\( M_{\\text{sus}} \\) has not been exposed to the synthetic data from \\( G_d \\) during training and is therefore compliant. Conversely, if the predictions generated by \\( M_{\\text{sus}} \\) and \\( M_{\\text{sdw}} \\) are similar, it indicates that the generalization performance of \\( M_{\\text{sus}} \\) on the validation dataset closely resembles that of \\( M_{\\text{sdw}} \\). This similarity implies that \\( M_{\\text{sus}} \\), much like \\( M_{\\text{sdw}} \\), has been trained on the synthetic data from \\( G_d \\), thereby sug-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23820" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.095, + 0.424, + 0.109 + ], + "angle": 0, + "content": "1. Shadow and Validation Dataset Generation" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.113, + 0.478, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.213, + 0.293, + 0.227 + ], + "angle": 0, + "content": "2. Shadow Model Training" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.229, + 0.48, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.306, + 0.256, + 0.321 + ], + "angle": 0, + "content": "3. Hypothesis Testing" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.323, + 0.48, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.146, + 0.46, + 0.427, + 0.474 + ], + "angle": 0, + "content": "Figure 3. The complete process of TrainProVe." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.482, + 0.484, + 0.513 + ], + "angle": 0, + "content": "gesting that \\( M_{sus} \\) violates the intellectual property rights of \\( G_{d} \\) and is considered illicit." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.527, + 0.224, + 0.544 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.484, + 0.629 + ], + "angle": 0, + "content": "Here, we present the results of DOV4MM on ImageNet [5] and CIFAR [20], ablation experiments, interference resistance analysis, and efficiency analysis. More details and results (hypothesis testing, results on CLIP [35], etc) can be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.637, + 0.282, + 0.654 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Models and Datasets. We evaluate the proposed TrainProVe with three visual datasets (CIFAR10 [20], CIFAR100 [20], and ImageNet-100 [40]) and four text-to-image models (Stable Diffusion [38] v1.4, latent consistency model [28], PixArt-\\(\\alpha\\) [4], and Stable Cascade [34]). ImageNet-100 is a randomly selected subset of ImageNet1K [5], encompassing 100 categories and 126,689 images. Stable Cascade is a latent diffusion model constructed on the Würstchen architecture [34]. In the experiments, unless explicitly stated otherwise, all text-to-image models employed are pre-trained models. To distinguish from real datasets, we denote the synthetic datasets of CIFAR10, CIFAR100, and ImageNet-100 as CIFAR10-Syn, CIFAR100-Syn, and ImageNet-100-Syn, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Experiments on CIFAR: The shadow dataset \\( \\mathcal{D}_{sdw} \\) contains 30,000 images with a resolution of \\( 32 \\times 32 \\), and the text" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.905, + 0.394 + ], + "angle": 0, + "content": "prompt is \\(\\mathcal{T}_{sdw} = \\{\\text{“}a\\{c\\}\\text{”}|c\\in \\mathcal{C}\\}\\). The validation dataset \\(\\mathcal{D}_v\\) includes 5,000 images with a resolution of \\(32\\times 32\\), with a text prompt of \\(\\mathcal{T}_v = \\{\\text{“}a\\text{ photo of }a\\{c\\}\\text{”}|c\\in \\mathcal{C}\\}\\). For the shadow model \\(M_{sdw}\\), the architecture is ResNet18 [12], and following previous work [42], we set the batch size for training \\(M_{sdw}\\) to 512, the learning rate to 1e-4, weight decay to 0.9, and the loss function to cross-entropy loss. To simulate the training data for the suspicious model, we generated 60,000 images with a resolution of \\(32\\times 32\\) using the method in [42] as CIFAR10-Syn/CIFAR100-Syn. Moreover, the suspicious models have various architectures and training hyperparameters to simulate different scenarios, specifically including: four architectures (ResNet50 [12], ResNet101 [12], ViT-S/16 [7], ViT-B/16 [7]), two batch sizes (512 and 256), two learning rates (1e-4 and 5e-5), two weight decay values (0.9 and 0.09), and two loss functions (cross-entropy loss and focal loss [22]). For each dataset, we can use these parameters to train 64 (\\(C_4^1 C_2^1 C_2^1 C_2^1 C_2^1\\)) suspicious models. Batch size for inference in TrainProVe is 128, and training epoch for all models is set to 100." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.395, + 0.905, + 0.606 + ], + "angle": 0, + "content": "Experiments on ImageNet-100: For the shadow dataset \\( \\mathcal{D}_{sdw} \\), validation dataset \\( \\mathcal{D}_v \\), and shadow model \\( M_{sdw} \\), all settings are the same as those used for CIFAR, except that \\( \\mathcal{D}_{sdw} \\) contains 50,000 images with a resolution of \\( 256 \\times 256 \\) and the batch size for model training is 64. The training set for the suspicious model \\( M_{sus} \\) includes 100,000 images generated using the method in [40], with a resolution of \\( 256 \\times 256 \\). For the parameter selection of the suspicious model, all settings are the same as those used in the experiments on CIFAR, except that the batch size is 64 or 32. Therefore, we can also obtain 64 suspicious models for each dataset (ImageNet-100 or ImageNet-100-Syn). Additionally, batch size for inference in TrainProVe is 16. The training epoch for all models is set to 100." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.905, + 0.713 + ], + "angle": 0, + "content": "Evaluation Metrics. Our goal is to classify the suspicious model \\( M_{sus} \\) as either illegal or legal. Specifically, if the training data for \\( M_{sus} \\) originates from the defender's generative model \\( G_d \\), then \\( M_{sus} \\) should be classified as illegal; otherwise, \\( M_{sus} \\) is considered legal. Given that this is a classification task, we selected three commonly used metrics for evaluation: accuracy, F1 score, and AUROC." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.725, + 0.622, + 0.74 + ], + "angle": 0, + "content": "4.2. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.905, + 0.824 + ], + "angle": 0, + "content": "Since TrainProVe is the first attempt at addressing the introduced training data provenance verification problem, there are no existing methods for direct comparison. Nonetheless, there are several baselines that can be utilized to showcase the efficacy of the proposed method, as enumerated below:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.825, + 0.905, + 0.855 + ], + "angle": 0, + "content": "Random classification: a rather trivial baseline that randomly classify suspicious models as illegal or legal." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Han et al.'s work [11]: This is the closest method to ours, which trains an attributor to classify suspicious models. The underlying assumption of this method is that the defender" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23821" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.9, + 0.357 + ], + "angle": 0, + "content": "
The Suspect's Data SourcesTPFPFNTNAccuracyF1 ScoreAUROC
The Defender's Generative Model GdData Sources Unrelated to Gd
Stable Diffusion [38] v1.4Latent Consistency Model064
PixArt-α6400640.9880.9700.992
Stable Cascade064
Real Data460
Latent Consistency Model [28]Stable Diffusion v1.4064
PixArt-α6400641.0001.0001.000
Stable Cascade064
Real Data064
PixArt-α [4]Stable Diffusion v1.4064
Latent Consistency Model6400641.0001.0001.000
Stable Cascade064
Real Data064
Stable Cascade [34]Stable Diffusion v1.4064
Latent Consistency Model6400641.0001.0001.000
PixArt-α064
Real Data064
Average Value0.9970.9920.998
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.363, + 0.907, + 0.406 + ], + "angle": 0, + "content": "Table 1. The results of TrainProVe on CIFAR10. The \"64\" in the table indicates the 64 suspicious models trained based on a data source, with different training hyperparameters. Suspicious models predicted as illegal/legal are considered positive/negative instances. If the suspicious model's training data source is \\( G_{d} \\), it is illegal, otherwise, it is innocent." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.414, + 0.482, + 0.49 + ], + "angle": 0, + "content": "possesses multiple different generative models to train the attributor, while in our scenario, the defender has only a single generative model \\( G_{d} \\). To evaluate Han et al.'s method, we relaxed the constraints on the defender by allowing the defender to access all four text-to-image models." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.49, + 0.483, + 0.655 + ], + "angle": 0, + "content": "TrainProVe-Ent: This method performs hypothesis testing by analyzing the entropy of the predicted logits generated by models \\( M_{sdw} \\) and \\( M_{sus} \\). Specifically, we analyze \\( D_v \\) in batches using models \\( M_{sdw} \\) and \\( M_{sus} \\) to derive their respective logits, and subsequently compute their entropies \\( \\mathcal{E}_{sdw} \\) and \\( \\mathcal{E}_{sus} \\). Following this, we employ a one-sided Grubbs' hypothesis test to ascertain whether the average of all elements in \\( \\mathcal{E}_{sus} \\) is an outlier with a high value compared to \\( \\mathcal{E}_{sdw} \\). This analysis aids in determining the validity of model \\( M_{sus} \\), with its parameter configurations being consistent with those of TrainProVe." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.656, + 0.483, + 0.822 + ], + "angle": 0, + "content": "TrainProVe-Sim: This method conducts hypothesis testing based on the pairwise similarity of logits within the same category from \\( M_{sdw} \\) and \\( M_{sus} \\). Specifically, \\( D_v \\) is processed in batches using both \\( M_{sdw} \\) and \\( M_{sus} \\) to obtain their respective logits. Subsequently, the pairwise cosine similarity within the same category is computed to obtain the similarity sets \\( S_{sdw} \\) and \\( S_{sus} \\). Following this, a one-sided Grubbs' hypothesis test is conducted to determine whether the average of all elements in \\( S_{sus} \\), is an outlier of low value in \\( S_{sdw} \\), thereby evaluating the validity of \\( M_{sus} \\). The parameter configurations align with those in TrainProVe." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.295, + 0.85 + ], + "angle": 0, + "content": "4.3. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.903 + ], + "angle": 0, + "content": "In this paper, \\( G_{d} \\) is one of the four text-to-image models (Stable Diffusion v1.4, latent consistency model, PixArt-α, and Stable Cascade), and the training dataset of \\( M_{sus} \\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.414, + 0.907, + 0.565 + ], + "angle": 0, + "content": "comes from one of five data sources (Stable Diffusion v1.4, latent consistency model, PixArt-\\(\\alpha\\), Stable Cascade and real data). We first calculate the accuracy, F1 score, and AUROC under the four different \\(G_{d}\\), and then average the metrics across the four scenarios as the final metrics for comparing different methods. Taking the results of TrainProVe on CIFAR10 as an example (more experimental results can be found in the supplementary material), as shown in Tab. 1. It is evident that TrainProVe can accurately classify suspicious models trained on various data sources." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.909, + 0.839 + ], + "angle": 0, + "content": "The results of various methods on different datasets are illustrated in Tab. 2, showcasing the superiority of TrainProVe. In comparison to the research by Han et al., TrainProVe has three advantages: (1) Han et al.'s approach necessitates the defender to have access to multiple different generative models, while TrainProVe is effective even when the defender only has his own generative model \\( G_{d} \\); (2) The efficacy of Han et al.'s method consistently lags behind that of TrainProVe across all cases. This is because the former relies on an attributor for classification, which tends to overfit to \\( M_{sdw} \\)'s architecture used by the defender. If the architecture of \\( M_{sus} \\) differs, this method is likely to fail. In contrast, TrainProVe relies on hypothesis testing to classify \\( M_{sus} \\), thus avoiding this risk; (3) Han et al.'s method requires training a large number of models based on synthetic datasets from each generative model to train the attributor. In contrast, TrainProVe only requires training a single ResNet18 on one synthetic dataset generated by \\( G_{d} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Moreover, compared to TrainProVe-Ent and TrainProVe-Sim, the efficacy of TrainProVe consistently surpasses that of the others. This is because it compares the predictions of \\( M_{sus} \\) and \\( M_{sdw} \\) with the true labels to obtain accuracy," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23822" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.485, + 0.32 + ], + "angle": 0, + "content": "
DatasetMethodAvg. AccAvg. F1Avg. AUC
CIFAR10Random0.5000.2860.500
Han et al.'s Work0.7680.5790.731
TrainProVe-Sim0.8120.6470.821
TrainProVe-Ent0.8680.7740.902
TrainProVe0.9970.9920.998
CIFAR100Random0.5000.2860.500
Han et al.'s Work0.7300.5480.703
TrainProVe-Sim0.5750.3180.545
TrainProVe-Ent0.3980.4080.624
TrainProVe0.9920.9790.981
ImageNet-100Random0.5000.2860.500
Han et al.'s Work0.6710.4250.645
TrainProVe-Sim0.5550.3050.528
TrainProVe-Ent0.3560.3640.567
TrainProVe0.7860.4950.769
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.322, + 0.483, + 0.378 + ], + "angle": 0, + "content": "Table 2. The results of TrainProVe and baselines on different datasets. \"Avg. Acc\", \"Avg. F1\", and \"Avg. AUC\" are the average values of accuracy, F1 score, and AUROC under four different \\( G_{d} \\), similar to Tab. 1. Bold represents the best results." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.388, + 0.48, + 0.489 + ], + "angle": 0, + "content": "
DatasetMsdwAvg. AccAvg. F1Avg. AUC
CIFAR10ResNet18 [12]0.9970.9920.998
ConvNeXt-B [26]0.9950.9890.997
Swin-B [25]0.9970.9920.998
ImageNet-100ResNet180.7860.4950.769
ConvNeXt-B0.7710.4700.756
Swin-B0.7870.4870.766
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.49, + 0.483, + 0.532 + ], + "angle": 0, + "content": "Table 3. Changing \\( M_{sdw} \\)'s architecture. \"Avg. Acc\", \"Avg. F1\", and \"Avg. AUC\" are the average values of accuracy, F1 score, and AUROC under four different \\( G_d \\), similar to Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.484, + 0.614 + ], + "angle": 0, + "content": "using this as the basis for hypothesis testing. In contrast, TrainProVe-Ent and TrainProVe-Sim don't compare model predictions with true labels, resulting in a lack of information from the true labels, which leads to inferior validation performance compared to TrainProVe." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.626, + 0.255, + 0.642 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.483, + 0.759 + ], + "angle": 0, + "content": "The model architecture of \\( M_{sdw} \\). We substituted the standard architecture of \\( M_{sdw} \\), ResNet18, with ConvNeXt-B [26] and Swin-B [25], while leaving all other configurations unaltered. The outcomes are displayed in Tab. 3, demonstrating that TrainProVe remains efficient irrespective of whether the architecture of the shadow model is based on CNNs or transformers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The training epochs of \\( M_{sdw} \\). We explored TrainProVe's performance as the \\( M_{sdw} \\)'s training epochs (default is 100) increased from 0 to 200 in increments of 20. The findings on CIFAR10, depicted in Fig. 4, suggest that TrainProVe exhibits a commendable resilience to variations in the number of training epochs for \\( M_{sdw} \\). It is worth noting that an excessive or insufficient number of training epochs can lead to overfitting or underfitting of \\( M_{sdw} \\) on \\( \\mathcal{D}_{sdw} \\), consequently resulting in a decline in TrainProVe's performance." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.088, + 0.905, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.16, + 0.904, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.234, + 0.903, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.301, + 0.905, + 0.342 + ], + "angle": 0, + "content": "Figure 4. Changing \\( M_{sdw} \\)'s training epochs. \"Avg. Acc\", \"Avg. F1\", and \"Avg. AUC\" are the average values of accuracy, F1 score, and AUROC under four different \\( G_d \\)." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.35, + 0.888, + 0.398 + ], + "angle": 0, + "content": "
Text PromptsAccuracyF1 ScoreAUROC
Original Prompts0.9880.9700.992
Swapped Prompts0.9880.9700.992
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.408, + 0.906, + 0.45 + ], + "angle": 0, + "content": "Table 4. Swapping \\(\\mathcal{T}_{sdw}\\) and \\(\\mathcal{T}_v\\). Note that Stable Diffusion v1.4 is \\(G_{d}\\). Latent consistency model, PixArt-\\(\\alpha\\), Stable Cascade, and real data represent other data sources." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.906, + 0.529 + ], + "angle": 0, + "content": "The text prompts of \\( \\mathcal{D}_{sdw} \\) and \\( \\mathcal{D}_v \\). We swapped the text prompts, setting \\( \\mathcal{T}_{sdw}' = \\mathcal{T}_v \\) and \\( \\mathcal{T}_v' = \\mathcal{T}_{sdw} \\), and used \\( \\mathcal{T}_{sdw}' \\) and \\( \\mathcal{T}_v' \\) to generate \\( \\mathcal{D}_{sdw} \\) and \\( \\mathcal{D}_v \\), respectively. The results on CIFAR10, shown in Tab. 4, indicate that TrainProVe is robust to the text prompts of \\( \\mathcal{D}_{sdw} \\) and \\( \\mathcal{D}_v \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.907, + 0.67 + ], + "angle": 0, + "content": "The sample sizes of \\(\\mathcal{D}_{sdw}\\) and \\(\\mathcal{D}_v\\). We generated \\(\\mathcal{D}_{sdw}\\) and \\(\\mathcal{D}_v\\) with different sample sizes for TrainProVe, keeping other settings unchanged. Note that the default sample sizes for \\(\\mathcal{D}_{sdw}\\) and \\(\\mathcal{D}_v\\) are 30,000 and 5,000 respectively, when generating CIFAR10. The results on CIFAR10 shown in Fig. 5 demonstrate that TrainProVe has good robustness to the sample sizes. Moreover, when \\(\\mathcal{D}_{sdw}\\)'s sample size is too small, the generalization performance of \\(M_{sdw}\\) on \\(\\mathcal{D}_v\\) may degrade, leading to a decline in TrainProVe's performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.678, + 0.878, + 0.694 + ], + "angle": 0, + "content": "4.5. The Interference Resistance of TrainProVe" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.701, + 0.906, + 0.746 + ], + "angle": 0, + "content": "In this section, we study that whether TrainProVe remains effective when facing more covert data theft. More experimental details can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Multiple training data sources for \\( M_{\\text{sus}} \\). We investigated the robustness of TrainProVe when the training dataset of \\( M_{\\text{sus}} \\) has multiple data sources. Specifically, we assume that the training dataset of \\( M_{\\text{sus}} \\) contains both synthetic data from \\( G_d \\) and real data. The results on CIFAR10 are shown in Tab. 5, where the training dataset for \\( M_{\\text{sus}} \\) consists of 120,000 images, with synthetic data and real data each accounting for half. We calculated the accuracy of TrainProVe in classifying \\( M_{\\text{sus}} \\) trained on this mixed dataset. The results indicate that TrainProVe can still accu" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23823" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.282, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.087, + 0.476, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.194, + 0.282, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.194, + 0.476, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.304, + 0.282, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.304, + 0.476, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.406, + 0.483, + 0.448 + ], + "angle": 0, + "content": "Figure 5. Changing the sample sizes of \\(\\mathcal{D}_{sdw}\\) and \\(\\mathcal{D}_v\\). \"Avg. Acc\", \"Avg. F1\", and \"Avg. AUC\" are the average values of accuracy, F1 score, and AUROC under four different \\(G_{d}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.483, + 0.474 + ], + "angle": 0, + "content": "rately detect illegal behavior in this more difficult situation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.48, + 0.484, + 0.646 + ], + "angle": 0, + "content": "Fine-tuning \\( G_{d} \\). When the suspect has access to a small amount of real data, he may fine-tune the defender's open-source model \\( G_{d} \\) and then generate data to train \\( M_{sus} \\), which can enhance \\( M_{sus} \\)'s generalization ability on real data. Note that the defender only possesses the original \\( G_{d} \\) and has no knowledge of fine-tuning. Specifically, we fine-tuned the pre-trained Stable Diffusion v1.4 using 10,000 ImageNet-100 images (100 per class) and LoRA [16] with 10 epochs, and then generated ImageNet-100-Syn to train \\( M_{sus} \\). The results are shown in Tab. 6a, which indicate that TrainProVe remains effective in this more arduous scenario." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.652, + 0.483, + 0.788 + ], + "angle": 0, + "content": "Fine-tuning \\( M_{sus} \\). Similarly, when the suspect has a small amount of real data, he may also fine-tune \\( M_{sus} \\) with real data, which was trained on synthetic data from \\( G_{d} \\), to enhance \\( M_{sus} \\)'s generalization ability on real data. Specifically, we fine-tuned \\( M_{sus} \\) using 10,000 ImageNet-100 images (100 per class), where \\( M_{sus} \\) was pre-trained on synthetic data from Stable Diffusion v1.4, with 10 fine-tuning epochs. The results, shown in Tab. 6b, indicate even in this more challenging scenario, TrainProVe remains effective." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.802, + 0.274, + 0.818 + ], + "angle": 0, + "content": "4.6. Efficiency Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We analyzed the efficiency of TrainProVe and the work of Han et al. on ImageNet-100. Specifically, following the setup in Sec. 4.1, given \\( M_{sus} \\) trained on ImageNet-100 or ImageNet-100-Syn, we performed detection using both methods on an NVIDIA GeForce RTX 4090. \\( G_{d} \\) is Sta" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.088, + 0.905, + 0.163 + ], + "angle": 0, + "content": "
M_sus's Training SourceG_dTPFPAcc
CIFAR10-SynStable Diffusion v1.46401
Latent Consistency Model6401
+CIFAR10PixArt-α6401
Stable Cascade6401
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.165, + 0.905, + 0.207 + ], + "angle": 0, + "content": "Table 5. Results of TrainProVe when the training dataset of \\( M_{sus} \\) has multiple data sources. The \"64\" in the table indicates the 64 suspicious models with different training hyperparameters." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.215, + 0.706, + 0.262 + ], + "angle": 0, + "content": "
TPFPAcc
w/o fine-tuning6401.00
w/ fine-tuning6220.97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.559, + 0.264, + 0.662, + 0.277 + ], + "angle": 0, + "content": "(a) Fine-tuning \\(G_{d}\\)" + }, + { + "type": "table", + "bbox": [ + 0.716, + 0.215, + 0.905, + 0.263 + ], + "angle": 0, + "content": "
TPFPAcc
w/o fine-tuning6401.00
w/ fine-tuning5770.89
" + }, + { + "type": "table_caption", + "bbox": [ + 0.752, + 0.265, + 0.869, + 0.277 + ], + "angle": 0, + "content": "(b) Fine-tuning \\(M_{sus}\\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.279, + 0.905, + 0.308 + ], + "angle": 0, + "content": "Table 6. Fine-tuning \\( G_{d} \\) or \\( M_{sus} \\). The \"64\" in the table indicates the 64 suspicious models with different training hyperparameters." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.316, + 0.905, + 0.362 + ], + "angle": 0, + "content": "
MethodτgτtτvTotalAccF1AUC
Han et al.'s Work137.7h8.0h1.2h146.9h0.480.280.48
TrainProVe33.6h0.1h1.4h35.1h0.780.630.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.364, + 0.905, + 0.392 + ], + "angle": 0, + "content": "Table 7. Efficiency analysis. \"Acc\", \"F1\" and \"AUC\" refer to the accuracy, F1 score and AUROC respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.408, + 0.907, + 0.68 + ], + "angle": 0, + "content": "ble Cascade. The runtime and results is shown in Tab. 7. For TrainProVe, \\(\\tau_{g}\\) is the time for generating \\(D_{sdw}\\) and \\(D_v\\), \\(\\tau_t\\) is the time for training \\(M_{sdw}\\), and \\(\\tau_v\\) is the time for classifying \\(M_{sus}\\) using hypothesis testing. For Han et al.'s method, \\(\\tau_{g}\\) is the time for generating the training and validation dataset, \\(\\tau_t\\) is the time for training the models and the attributor, and \\(\\tau_v\\) is the time for classifying \\(M_{sus}\\) using the attributor. From the results, it shows that the high efficiency of TrainProVe lies in its significantly smaller \\(\\tau_{g}\\) and \\(\\tau_t\\) compared to Han et al.'s method. This is because TrainProVe only requires generating one dataset (\\(D_{sdw}\\)) and training one model (\\(M_{sdw}\\)). In contrast, Han et al.'s method requires generating \\(m\\) datasets and training \\(m \\times n\\) models, where \\(m\\) is the number of text-to-image models accessible to the defender and \\(n\\) is the number of models trained by the defender on each dataset. These models are then used to train the attributor. Moreover, TrainProVe achieved better classification results, further illustrating its superiority." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.694, + 0.634, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Open-source text-to-image models have reduced the cost of obtaining images. We introduce an important yet largely unexplored security issue: whether a suspicious model's training data originates from a specific text-to-image model? To our knowledge, we are the first to propose a solution to this problem, called TrainProVe. The success of TrainProVe in verifying training data provenance has been demonstrated through theory and experiments. Promising future work includes: (1) extending TrainProVe to other image generative models, such as image-to-image models; (2) adapting TrainProVe to protect other types of data (e.g., text); (3) exploring other privacy risks about text-to-image models." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23824" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.264, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.484, + 0.192 + ], + "angle": 0, + "content": "This work was partially supported by the Pioneer R&D Program of Zhejiang (No.2024C01021), and Zhejiang Province High-Level Talents Special Support Program \"Leading Talent of Technological Innovation of Ten-Thousands Talents Program\" (No. 2022R52046)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.207, + 0.188, + 0.223 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.232, + 0.484, + 0.274 + ], + "angle": 0, + "content": "[1] Hritik Bansal and Aditya Grover. Leaving reality to imagination: Robust classification via generated datasets. arXiv preprint arXiv:2302.02503, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.276, + 0.484, + 0.33 + ], + "angle": 0, + "content": "[2] Shai Ben-David, John Blitzer, Koby Crammer, and Fernando Pereira. Analysis of representations for domain adaptation. Advances in neural information processing systems, 19, 2006. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.333, + 0.483, + 0.401 + ], + "angle": 0, + "content": "[3] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- \\(\\sigma\\) : Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.404, + 0.484, + 0.487 + ], + "angle": 0, + "content": "[4] Junsong Chen, YU Jincheng, GE Chongjian, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- \\(\\alpha\\): Fast training of diffusion transformer for photorealistic text-to-image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 5, 6, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.489, + 0.484, + 0.544 + ], + "angle": 0, + "content": "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.546, + 0.484, + 0.6 + ], + "angle": 0, + "content": "[6] Tim Dockhorn, Robin Rombach, Andreas Blatmann, and Yaoliang Yu. Distilling the knowledge in diffusion models. In CVPR Workshop Generative Models for Computer Vision, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.603, + 0.484, + 0.685 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.688, + 0.484, + 0.757 + ], + "angle": 0, + "content": "[8] Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7382-7392, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.759, + 0.484, + 0.827 + ], + "angle": 0, + "content": "[9] Abolfazl Farahani, Sahar Voghoei, Khaled Rasheed, and Hamid R Arabnia. A brief review of domain adaptation. Advances in data science and information engineering: proceedings from ICDATA 2020 and IKE 2020, pages 877-894, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.83, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[10] Frank Ephraim Grubbs. Sample criteria for testing outlying observations. University of Michigan, 1949. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[11] Ge Han, Ahmed Salem, Zheng Li, Shanqing Guo, Michael Backes, and Yang Zhang. Detection and attribution of models trained on generated data. In ICASSP 2024-2024 IEEE" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.232, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.121 + ], + "angle": 0, + "content": "International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4875-4879. IEEE, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.124, + 0.906, + 0.178 + ], + "angle": 0, + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.181, + 0.907, + 0.235 + ], + "angle": 0, + "content": "[13] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is synthetic data from generative models ready for image recognition? arXiv preprint arXiv:2210.07574, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.238, + 0.906, + 0.265 + ], + "angle": 0, + "content": "[14] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.268, + 0.906, + 0.309 + ], + "angle": 0, + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.312, + 0.906, + 0.379 + ], + "angle": 0, + "content": "[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.383, + 0.906, + 0.452 + ], + "angle": 0, + "content": "[17] Yongcheng Jing, Yiding Yang, Xinchao Wang, Mingli Song, and Dacheng Tao. Turning frequency to resolution: Video super-resolution via event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7772-7781, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.454, + 0.906, + 0.522 + ], + "angle": 0, + "content": "[18] Yongcheng Jing, Chongbin Yuan, Li Ju, Yiding Yang, Xinchao Wang, and Dacheng Tao. Deep graph reprogramming. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24345-24354, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.906, + 0.595 + ], + "angle": 0, + "content": "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.598, + 0.906, + 0.624 + ], + "angle": 0, + "content": "[20] A Krizhevsky. Learning multiple layers of features from tiny images. Master's thesis, University of Tront, 2009. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.627, + 0.906, + 0.682 + ], + "angle": 0, + "content": "[21] Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pretrained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems, 36, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.685, + 0.906, + 0.711 + ], + "angle": 0, + "content": "[22] T Lin. Focal loss for dense object detection. arXiv preprint arXiv:1708.02002, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.715, + 0.906, + 0.756 + ], + "angle": 0, + "content": "[23] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.906, + 0.799 + ], + "angle": 0, + "content": "[24] Yugeng Liu, Zheng Li, Michael Backes, Yun Shen, and Yang Zhang. Watermarking diffusion model. arXiv preprint arXiv:2305.12502, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.906, + 0.871 + ], + "angle": 0, + "content": "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.874, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[26] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23825" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.133 + ], + "angle": 0, + "content": "2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.207 + ], + "angle": 0, + "content": "[27] Ge Luo, Junqiang Huang, Manman Zhang, Zhenxing Qian, Sheng Li, and Xinpeng Zhang. Steal my artworks for fine-tuning? a watermarking framework for detecting art theft mimicry in text-to-image models. arXiv preprint arXiv:2311.13619, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.484, + 0.263 + ], + "angle": 0, + "content": "[28] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 1, 3, 5, 6, 2, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.265, + 0.483, + 0.308 + ], + "angle": 0, + "content": "[29] Kidist Amde Mekonnen, Nicola Dall'Asen, and Paolo Rota. Adv-kd: Adversarial knowledge distillation for faster diffusion sampling. arXiv preprint arXiv:2405.20675, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.31, + 0.482, + 0.338 + ], + "angle": 0, + "content": "[30] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41, 1995. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.34, + 0.483, + 0.382 + ], + "angle": 0, + "content": "[31] Rui Min, Sen Li, Hongyang Chen, and Minhao Cheng. A watermark-conditioned diffusion model for ip protection. arXiv preprint arXiv:2403.10893, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.384, + 0.483, + 0.425 + ], + "angle": 0, + "content": "[32] Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions on knowledge and data engineering, 22(10):1345-1359, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.483, + 0.469 + ], + "angle": 0, + "content": "[33] Suraj Patil, William Berman, Robin Rombach, and Patrick von Platen. amused: An open muse reproduction. arXiv preprint arXiv:2401.01808, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.471, + 0.483, + 0.541 + ], + "angle": 0, + "content": "[34] Pablo Pernias, Dominic Rampas, Mats L. Richter, Christopher J. Pal, and Marc Aubreville. Wuerstchen: An efficient architecture for large-scale text-to-image diffusion models. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 5, 6, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.543, + 0.483, + 0.625 + ], + "angle": 0, + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.628, + 0.483, + 0.683 + ], + "angle": 0, + "content": "[36] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.686, + 0.483, + 0.741 + ], + "angle": 0, + "content": "[37] Mauro Ribeiro, Katarina Grolinger, and Miriam AM Capretz. Mlaas: Machine learning as a service. In 2015 IEEE 14th international conference on machine learning and applications (ICMLA), pages 896-902. IEEE, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.744, + 0.483, + 0.813 + ], + "angle": 0, + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3, 5, 6, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[39] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In International Conference on Learning Representations, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[40] Mert Bülent Sariyildüz, Karteek Alahari, Diane Larlus, and Yannis Kalantidis. Fake it till you make it: Learning transferable representations from synthetic imagenet clones. In" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8011-8021, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[41] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[42] Jordan Shipard, Arnold Wiliem, Kien Nguyen Thanh, Wei Xiang, and Clinton Fookes. Diversity is definitely needed: Improving model-agnostic zero-shot classification via stable diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 769–778, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[43] Geng Sun, Tingru Cui, Jianming Yong, Jun Shen, and Shiping Chen. Mlaas: a cloud-based system for delivering adaptive micro learning in mobile mooc learning. IEEE Transactions on Services Computing, 11(2):292-305, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[44] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2117-2126, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[45] Lisa Torrey and Jude Shavlik. Transfer learning. In Handbook of research on machine learning applications and trends: algorithms, methods, and techniques, pages 242-264. IGI global, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.475, + 0.905, + 0.504 + ], + "angle": 0, + "content": "[46] Mei Wang and Weihong Deng. Deep visual domain adaptation: A survey. Neurocomputing, 312:135-153, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[47] Zhenting Wang, Chen Chen, Yuchen Liu, Lingjuan Lyu, Dimitris Metaxas, and Shiqing Ma. How to detect unauthorized data usages in text-to-image diffusion models. arXiv preprint arXiv:2307.03108, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[48] Zhenting Wang, Chen Chen, Yi Zeng, Lingjuan Lyu, and Shiqing Ma. Where did i come from? origin attribution of ai-generated images. Advances in neural information processing systems, 36, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[49] Qizhen Weng, Wencong Xiao, Yinghao Yu, Wei Wang, Cheng Wang, Jian He, Yong Li, Liping Zhang, Wei Lin, and Yu Ding. Mlaas in the wild: Workload analysis and scheduling in large-scale heterogeneousgpu clusters. In 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI 22), pages 945–960, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[50] Qianlong Xiang, Miao Zhang, Yuzhang Shang, Jianlong Wu, Yan Yan, and Liqiang Nie. Dkdm: Data-free knowledge distillation for diffusion models with any architecture. arXiv preprint arXiv:2409.03550, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[51] Yuechen Xie, Jie Song, Mengqi Xue, Haofei Zhang, Xingen Wang, Bingde Hu, Genlang Chen, and Mingli Song. Dataset ownership verification in contrastive pre-trained models. arXiv preprint arXiv:2502.07276, 2025.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[52] Cheng Xiong, Chuan Qin, Guorui Feng, and Xinpeng Zhang. Flexible and secure watermarking for latent diffusion model. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1668-1676, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[53] Jianhao Yuan, Francesco Pinto, Adam Davies, and Philip Torr. Not just pretty pictures: Toward interventional data" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23826" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.484, + 0.12 + ], + "angle": 0, + "content": "augmentation using text-to-image generators. arXiv preprint arXiv:2212.11237, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.162 + ], + "angle": 0, + "content": "[54] Zihan Yuan, Li Li, Zichi Wang, and Xinpeng Zhang. Watermarking for stable diffusion models. IEEE Internet of Things Journal, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.484, + 0.22 + ], + "angle": 0, + "content": "[55] Likun Zhang, Hao Wu, Lingcui Zhang, Fengyuan Xu, Jin Cao, Fenghua Li, and Ben Niu. Training data attribution: Was your model secretly trained on data created by mine? arXiv preprint arXiv:2409.15781, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.484, + 0.274 + ], + "angle": 0, + "content": "[56] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Ngai-Man Cheung, and Min Lin. A recipe for watermarking diffusion models. arXiv preprint arXiv:2303.10137, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.484, + 0.319 + ], + "angle": 0, + "content": "[57] Yongchao Zhou, Hshmat Sahak, and Jimmy Ba. Training on thin air: Improve image classification with generated data. arXiv preprint arXiv:2305.15316, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.484, + 0.375 + ], + "angle": 0, + "content": "[58] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23827" + } + ] +] \ No newline at end of file diff --git a/2025/Training Data Provenance Verification_ Did Your Model Use Synthetic Data from My Generative Model for Training_/95a43e45-ca4b-40d2-9bb6-79d7e388dedb_origin.pdf b/2025/Training Data Provenance Verification_ Did Your Model Use Synthetic Data from My Generative Model for Training_/95a43e45-ca4b-40d2-9bb6-79d7e388dedb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..75bb4ce63d296eeb5842bae4cc5adc604fc8d8cf --- /dev/null +++ b/2025/Training Data Provenance Verification_ Did Your Model Use Synthetic Data from My Generative Model for Training_/95a43e45-ca4b-40d2-9bb6-79d7e388dedb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3d3b2f0504ac37a1c53c5fa0084479282aa0308d660eab9d4c3d78ae060b63b4 +size 7458901 diff --git a/2025/Training Data Provenance Verification_ Did Your Model Use Synthetic Data from My Generative Model for Training_/full.md b/2025/Training Data Provenance Verification_ Did Your Model Use Synthetic Data from My Generative Model for Training_/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0d8a6cabb96b2e7cda0a6f3adb8a6e956f43210a --- /dev/null +++ b/2025/Training Data Provenance Verification_ Did Your Model Use Synthetic Data from My Generative Model for Training_/full.md @@ -0,0 +1,318 @@ +# Training Data Provenance Verification: Did Your Model Use Synthetic Data from My Generative Model for Training? + +Yuechen Xie $^{1}$ , Jie Song $^{1*}$ , Huiqiong Wang $^{2}$ , Mingli Song $^{1,3,4}$ + +$^{1}$ Zhejiang University + +$^{2}$ Ningbo Innovation Center, Zhejiang University + +3State Key Laboratory of Blockchain and Security, Zhejiang University + +$^{4}$ Hangzhou High-Tech Zone (Binjiang) Institute of Blockchain and Data Security + +{xyuechen,sjie,huiqiong-wang,brooksong}@zju.edu.cn + +# Abstract + +High-quality open-source text-to-image models have lowered the threshold for obtaining photorealistic images significantly, but also face potential risks of misuse. Specifically, suspects may use synthetic data generated by these generative models to train models for specific tasks without permission, when lacking real data resources especially. Protecting these generative models is crucial for the wellbeing of their owners. In this work, we propose the first method to this important yet unresolved issue, called Training data Provenance Verification (TrainProVe). The rationale behind TrainProVe is grounded in the principle of generalization error bound, which suggests that, for two models with the same task, if the distance between their training data distributions is smaller, their generalization ability will be closer. We validate the efficacy of TrainProVe across four text-to-image models (Stable Diffusion v1.4, latent consistency model, PixArt-α, and Stable Cascade). The results show that TrainProVe achieves a verification accuracy of over 99% in determining the provenance of suspicious model training data, surpassing all previous methods. Code is available at https://github.com/xieyc99/TrainProVe. + +# 1. Introduction + +High-quality open-source text-to-image models based on diffusion models [15] can generate photorealistic images with simple text prompts [4, 28, 34, 38], making them highly significant in terms of lowering creative barriers, enabling personalized customization, etc. Currently, most open-source text-to-image models are only permitted for + +academic or educational purposes, and commercial use is prohibited without explicit permission [3, 4, 21, 28, 33, 34, 36, 38]. Therefore, protecting the intellectual property of these generative models is crucial for safeguarding the interests of their owners. + +Currently, one of the major security challenges faced by defenders, i.e., the generative model owners, is the unauthorized use of the synthetic data from their models by suspects [24, 47, 48, 55]. As shown in Fig. 1, based on the suspect's intention, this issue can be divided into three cases: + +Case 1: the suspect steals the synthetic data illegally and claims it as his own personal work [24, 44, 48], implying the intellectual property of the generative model is infringed. + +Case 2: the suspect steals the synthetic data from the defender's generative model, and trains another generative model for the same task using knowledge distillation techniques [6, 14, 29, 39, 50]. In this case, the defender cannot access the synthetic data used by the suspect, making it a more challenging defense task compared to Case 1. + +Case 3: the suspect steals the synthetic data from the generative model to train another model for a specific task [1, 8, 13, 40, 42, 53, 57]. In this case, the defender is not only unaware of the training data used by the suspicious model, but also cannot compare the similarity between the generative model and the suspicious model directly due to the different tasks, representing an even more arduous defense task compared to Case 2. + +For Case 1 and Case 2, there are already some mature methods to address these issues, such as adding watermarks [24, 31, 52, 54, 56] and training data attribution [55], which identify illegal activities by checking the distributions of the suspicious images or the synthetic images from suspicious model. However, for Case 3, although the potential of using synthetic data to train models for other tasks has been recognized gradually in recent years [13, 40, 42, 53] + +![](images/f01004709656455b6d0c5b1cc26c89ab12083bcfe969bba33cf233c779edd264.jpg) +Figure 1. The overview of the three cases. These three cases encompass nearly every conceivable real-world scenario, where a suspect uses synthetic images from the defender's text-to-image model illegally. In this paper, we are committed to addressing the security risks in Case 3. + +due to the lower collection costs compared to real data, the associated security risks have been overlooked. Methods that work for the first two cases cannot be applied to the third case, because the learning task of the suspicious model is not necessarily image generation. For example, for a suspicious image classification model, there are no synthetic data available for defenders to examine. + +In this work, we address this important yet unexplored issue. To the best of our knowledge, we propose the first method for protecting the intellectual property of generative models in such scenario, which we call Training data Provenance Verification (TrainProVe). It helps defenders verify whether the suspicious model has been trained using synthetic data from their generative model. Given that suspicious models may use synthetic data for training without authorization, and is used for commercial purposes, we focus on the black-box setting, where the defenders have no information about the training configuration of the suspicious model (e.g., loss function, model architecture, etc.) + +and can only access the model through Machine Learning as a Service (MLaaS) [37, 43, 49]. This means that the defenders can only obtain prediction results (such as prediction logits or labels) via the suspicious model's API. + +The rationale behind TrainProVe is grounded in the principle of generalization error bound, which suggests that, for two models with the same task, if the distance between their training data distributions is smaller, their generalization ability will be closer. Specifically, defenders can train a model for the same task as the suspicious model, using the synthetic data generated by their generative model. By comparing the performances of these two models, defenders can determine whether the suspicious model was trained on synthetic data from their generative model. More concretely, the proposed TrainProVe involves three steps: (1) using the defender's generative model along with different text prompts to generate a shadow dataset and a validation dataset, based on the suspicious model's task; (2) training a shadow model for the same task as the suspicious model using the shadow dataset; (3) performing hypothesis testing on the prediction accuracy of the suspicious model and the shadow model on the validation dataset, in order to determine whether the suspicious model was trained using synthetic data from the defender's generative model. + +In summary, the principal contributions of this paper are threefold: (1) we introduce the important yet under-explored issue of intellectual property protection for text-to-image models, in which the suspect uses the synthetic data from the defender's generative model illegally to train a specific-task suspicious model, such as an image classification model. For the defender, the suspicious model is a black-box; (2) we discover and demonstrate that the similarity between different models is related to their source domains closely, and based on this, we propose the training data provenance verification technique. To the best of our knowledge, this is the first method to address such intellectual property protection issue; (3) Extensive experiments show that TrainProVe achieves over $99\%$ accuracy in verifying training data provenance for suspicious models, surpassing all previous studies. + +# 2. Related Work + +Text-to-image diffusion model. Diffusion models [15] have revolutionized generative modeling fundamentally, in terms of quality and generalization. Their excellent scalability has led to significant advancements in specific applications, such as text-to-image generation. Text-to-image diffusion models [4, 34, 38] use the diffusion process to transform text descriptions into realistic images. Recently, people have begun to recognize the potential of text-to-image diffusion models for synthesizing dataset [17, 18, 40, 42, 57]. However, so far, no one has realized that suspects + +might use these text-to-image diffusion models to synthesize datasets and train their own models without authorization. Therefore, in this paper, we propose the first feasible solution to address this issue, securing and fostering healthy development in this field. + +Text-to-image model protection. Training generative models requires large amounts of data and computational resources [4, 38], making the intellectual property of these models crucial for their owners. When these models are open-source [4, 28, 34, 38], their intellectual property is vulnerable to infringement by suspects. Specifically, for text-to-image models, suspects have the following three ways primarily to infringe on their intellectual property: (1) stealing synthetic data illegally; (2) illegal knowledge distillation; (3) training suspicious models for specific tasks using synthetic data illegally. For the first two cases, watermarking [24, 27, 56] and training data attribution [55] have been developed to address these issues. However, for the third case, no suitable solution exists currently. This work aims to address this gap. + +Generalization error bound. Generalization error (upper) bound is used to describe a model's performance on testing data and widely applied in domain adaptation [9, 46], which is a special case of transfer learning [45, 51, 58]. Theorem 1 in [2] states the generalization error upper bound of model $M$ on the target domain $T$ . Let $\mathcal{H}$ be a hypothesis space with VC dimension $d$ , and let $m$ be the sample size of the dataset in the source domain $S$ . Then, with probability at least $1 - \delta$ , for every $M \in \mathcal{H}$ : + +$$ +\epsilon_ {T} (M) \leq \hat {\epsilon} _ {S} (M) + \sqrt {\frac {4}{m} \left(d \log \frac {2 e m}{d}\right)} + d _ {\mathcal {H}} \left(P _ {S}, P _ {T}\right) + \lambda , \tag {1} +$$ + +where $\epsilon_T(M)$ and $\hat{\epsilon_S} (M)$ represent the generalization error of $M$ on the target domain and the empirical error of $M$ on the source domain, respectively. $P_{S}$ and $P_{T}$ denote the marginal probability distributions of the source and target domains, respectively. $d_{\mathcal{H}}(P_S,P_T)$ is the $\mathcal{H}$ -divergence between the source and target domains, which measures the distance between the two distributions within the hypothesis space $\mathcal{H}$ . $\lambda$ is a constant and $e$ is the base of the natural logarithm. The rationale behind TrainProVe is grounded in this theorem, which suggests that, for two models with the same task, if the distance between their training data distributions is smaller, their generalization ability will be closer. + +# 3. Method + +# 3.1. Problem Formulation + +In this study, we focus on the training data provenance verification task in image classification. This task involves two roles: the defender and the suspect. The defender, embodying the role of text-to-image model provider, endeavors to + +ascertain whether the black-box suspicious model, $M_{sus}$ , has been illicitly trained on synthetic dataset generated by his model $G_{d}$ (as shown in Case 3 of Fig. 1). More specifically, for a black-box suspect model, the defender's objective is to determine whether it is illegal or legal based solely on the labels it outputs. + +$M_{sus}$ can be classified as either illegal or legal based on its training datasets: (1) Illegal: $M_{sus}$ is trained on the synthetic data from the defender's model $G_{d}$ , indicating the occurrence of intellectual property infringement; (2) Legal: $M_{sus}$ is trained using a different data source independent of the defender, such as some real data or synthetic data from other generative models, implying the innocence of the suspect. To be precise, when presented with a black-box suspicious model, the defender's objective is to determine whether it is an illegal or legal model. + +# 3.2. The Theoretical Insight + +The underlying rationale for TrainProVe is that when two models are assigned the same task, their ability to generalize will be more aligned if the divergence between their training data distributions is diminished. We demonstrate this through a case study of $K$ -way classification, where $\mathcal{X} = \mathbb{R}^{C \times W \times H}$ and $\mathcal{Y} = \{1, \dots, K\}$ represent the input and label spaces respectively. We follow the setting of previous domain adaptation works [32, 46], viewing a domain as composed of a feature space $\mathcal{X}$ and a marginal probability distribution $P(\boldsymbol{x})$ , i.e., $\{\mathcal{X}, P(\boldsymbol{x})\}$ , where $\boldsymbol{x} \in \mathcal{X}$ . + +Let $S_{1} = \{\mathcal{X},P_{1}(\pmb {x})\}$ and $S_{2} = \{\mathcal{X},P_{2}(\pmb {x})\}$ be two source domains sharing the same feature space but possessing distinct marginal probability distributions, i.e., $P_{1}(\pmb {x})\neq$ $P_{2}(\pmb {x}).D_{1} = \{(x_{1}^{i},y_{1}^{i})\}_{i = 1}^{|D_{1}|}$ and $\mathcal{D}_2 = \{(\pmb {x}_2^j,\pmb {y}_2^j)\}_{j = 1}^{|D_2|}$ represent the datasets sampled from the source domains $S_{1}$ and $S_{2}$ respectively, where $\pmb{x}_1^i\sim P_1(\pmb {x})$ and $\pmb{x}_2^j\sim P_2(\pmb {x})$ . $M$ and $\hat{M}$ are the models trained on $\mathcal{D}_1$ and $\mathcal{D}_2$ respectively. Then we have the following theorem: + +Theorem 1. Assume $M$ is trained on synthetic data generated by a text-to-image model $G$ with a set of text prompts $\mathcal{T}_1$ , i.e., $P_{1}(\pmb {x}) = P(\pmb {x}|G,\mathcal{T}_{1})$ . $\hat{M}$ can be trained on either real data or synthetic data generated by any text-to-image model. Based on the generalization errors of $M$ and $\hat{M}$ on the target domain $T = \{\mathcal{X},P(\pmb {x}|G,\mathcal{T}_t)\}$ , we have: + +$$ +\sup _ {P _ {2} (\boldsymbol {x}) = P (\boldsymbol {x} \mid G, \mathcal {T} _ {2})} | \Delta \epsilon_ {T} | \leq \sup _ {P _ {2} (\boldsymbol {x}) \perp G} | \Delta \epsilon_ {T} |, \tag {2} +$$ + +where $\Delta \epsilon_{T}$ represents the difference in generalization error between $M$ and $\hat{M}$ on the target domain $T$ , expressed as $\Delta \epsilon_{T} = \epsilon_{T}(M) - \epsilon_{T}(\hat{M})$ . $P_{2}(\pmb{x}) \perp G$ denotes that $P_{2}(\pmb{x})$ is independent of $G$ , meaning $P_{2}(\pmb{x}) = P(\pmb{x}|G^{\prime},\mathcal{T}_{2})$ or $P_{2}(\pmb{x}) = P_{R}(\pmb{x})$ , where $G^{\prime}$ is a text-to-image model different from $G$ and $P_{R}(\pmb{x})$ represents the distribution of real data. $\mathcal{T}_1, \mathcal{T}_2$ and $\mathcal{T}_t$ are distinct sets of text prompts. + +![](images/6b78a5229db3a7dec00d36cbfb66c8bbe8c062cb07f41a0e37dd762dab0b0c95.jpg) +Figure 2. The overview of TrainProVe's motivation. $M$ is trained on the dataset whose data distribution is $P(x|G, \mathcal{T}_1)$ . $\hat{M}_1$ , $\hat{M}_2$ , and $\hat{M}_3$ are $\hat{M}$ trained on three datasets with different data distributions, respectively. The source and target domain data are sampled from their respective data distributions. $G$ and $G'$ are two different text-to-image models, and $\mathcal{T}_1$ , $\mathcal{T}_2$ , and $\mathcal{T}_t$ are distinct sets of text prompts. + +![](images/715efa933966fef9d79a188274f5d9efaa7454963eb7c8ba1e8fd9f5c261c52b.jpg) + +Detailed proofs can be found in the supplementary material. Theorem 1 suggests that for two models with distinct (identical) training data sources, their generalization errors differ substantially (minimally). Its contrapositive states that for two models with small (large) discrepancies in generalization error, their training data sources are identical (distinct), which supports our method. Specifically, when $M$ and $\hat{M}$ are trained on synthetic data from the same text-to-image model, their accuracies on the corresponding synthetic data from this text-to-image model are more akin, compared to when their training data is derived from disparate sources. This aligns with the rationale of TrainProVe, as visually depicted in Fig. 2. The discrepancies in accuracies among these models originate from the disparities in their source domains. The closer the data distributions of the source domain data, the more analogous the accuracies of the trained models on the target domain data. + +# 3.3. The Proposed TrainProVe + +Building upon the above analysis, we propose a method called TrainProVe to protect generative models' intellectual property, as shown in Fig. 3, which consists of three stages: + +(1) Shadow and Validation Dataset Generation: Utilizing $G_{d}$ to produce a shadow dataset $\mathcal{D}_{sdw}$ and a validation dataset $\mathcal{D}_v$ , based on various text prompt sets $\mathcal{T}_{sdw}$ and $\mathcal{T}_v$ , tailored to the predicted labels set $\mathcal{C}$ of the suspect model $M_{sus}$ . Here $G_{d}, \mathcal{T}_{sdw}$ , and $\mathcal{T}_v$ correspond to $G, \mathcal{T}_1,$ and $\mathcal{T}_t$ in Fig. 2 respectively, while $\mathcal{D}_{sdw}$ and $\mathcal{D}_v$ correspond to the training data of $M$ and the target domain data. +(2) Shadow Model Training: Using the shadow dataset $\mathcal{D}_{sdw}$ to train the shadow model $M_{sdw}$ . The shadow model + +is analogous to $M$ in Fig. 2, whereas the suspect model is denoted as $\hat{M}$ . The purpose of the shadow model is to compare its performance with the suspect model trained on distinct data, focusing on their generalization capacities (accuracies) on the validation dataset. Specifically, when both models are trained on data from $G_{d}$ , their accuracies on the validation dataset tend to be similar. + +(3) Hypothesis Testing: Providing the validation dataset $\mathcal{D}_v$ to both the shadow model $M_{sdw}$ and the suspect model $M_{sus}$ in the same sequence batch by batch. Subsequently, calculating the accuracy of the two models batch by batch, leading to accuracy sets $\mathcal{A}_{sdw}$ and $\mathcal{A}_{sus}$ . Ultimately, utilizing a one-sided Grubbs' hypothesis test [10] to ascertain whether $M_{sus}$ violates the intellectual property of $G_d$ . + +In the one-sided Grubbs' hypothesis test, the null hypothesis $H_0$ asserts that the mean of the set $\mathcal{A}_{\text{sus}}$ , denoted as $\text{mean}(\mathcal{A}_{\text{sus}})$ , is not an unusually low value among the elements of $\mathcal{A}_{\text{sdw}}$ . On the contrary, the alternative hypothesis $H_1$ posits that the mean of $\mathcal{A}_{\text{sus}}$ is an exceptionally low value within $\mathcal{A}_{\text{sdw}}$ . Rejecting the null hypothesis $H_0$ implies that the predictive capabilities of $M_{\text{sus}}$ are significantly inferior to those of $M_{\text{sdw}}$ . This suggests that the model $M_{\text{sus}}$ exhibits notably inferior generalization performance on the validation dataset compared to $M_{\text{sdw}}$ , indicating that $M_{\text{sus}}$ has not been exposed to the synthetic data from $G_d$ during training and is therefore compliant. Conversely, if the predictions generated by $M_{\text{sus}}$ and $M_{\text{sdw}}$ are similar, it indicates that the generalization performance of $M_{\text{sus}}$ on the validation dataset closely resembles that of $M_{\text{sdw}}$ . This similarity implies that $M_{\text{sus}}$ , much like $M_{\text{sdw}}$ , has been trained on the synthetic data from $G_d$ , thereby sug- + +![](images/9bbf7b8e16a032488cd675791d46c51d08569ec6cce0820c24ac10f956b3a3df.jpg) +1. Shadow and Validation Dataset Generation + +![](images/78b0c6168b7f39c1ea107b3f8824ec8803f7b8696f11c6f70210db933ebdaaa5.jpg) +2. Shadow Model Training + +![](images/c562e8cfec122d9aba9ed01058b55243d0cd06254bc1892e6310cc7076b76002.jpg) +3. Hypothesis Testing +Figure 3. The complete process of TrainProVe. + +gesting that $M_{sus}$ violates the intellectual property rights of $G_{d}$ and is considered illicit. + +# 4. Experiments + +Here, we present the results of DOV4MM on ImageNet [5] and CIFAR [20], ablation experiments, interference resistance analysis, and efficiency analysis. More details and results (hypothesis testing, results on CLIP [35], etc) can be found in the supplementary materials. + +# 4.1. Experimental Setup + +Models and Datasets. We evaluate the proposed TrainProVe with three visual datasets (CIFAR10 [20], CIFAR100 [20], and ImageNet-100 [40]) and four text-to-image models (Stable Diffusion [38] v1.4, latent consistency model [28], PixArt- $\alpha$ [4], and Stable Cascade [34]). ImageNet-100 is a randomly selected subset of ImageNet1K [5], encompassing 100 categories and 126,689 images. Stable Cascade is a latent diffusion model constructed on the Würstchen architecture [34]. In the experiments, unless explicitly stated otherwise, all text-to-image models employed are pre-trained models. To distinguish from real datasets, we denote the synthetic datasets of CIFAR10, CIFAR100, and ImageNet-100 as CIFAR10-Syn, CIFAR100-Syn, and ImageNet-100-Syn, respectively. + +Experiments on CIFAR: The shadow dataset $\mathcal{D}_{sdw}$ contains 30,000 images with a resolution of $32 \times 32$ , and the text + +prompt is $\mathcal{T}_{sdw} = \{\text{“}a\{c\}\text{”}|c\in \mathcal{C}\}$ . The validation dataset $\mathcal{D}_v$ includes 5,000 images with a resolution of $32\times 32$ , with a text prompt of $\mathcal{T}_v = \{\text{“}a\text{ photo of }a\{c\}\text{”}|c\in \mathcal{C}\}$ . For the shadow model $M_{sdw}$ , the architecture is ResNet18 [12], and following previous work [42], we set the batch size for training $M_{sdw}$ to 512, the learning rate to 1e-4, weight decay to 0.9, and the loss function to cross-entropy loss. To simulate the training data for the suspicious model, we generated 60,000 images with a resolution of $32\times 32$ using the method in [42] as CIFAR10-Syn/CIFAR100-Syn. Moreover, the suspicious models have various architectures and training hyperparameters to simulate different scenarios, specifically including: four architectures (ResNet50 [12], ResNet101 [12], ViT-S/16 [7], ViT-B/16 [7]), two batch sizes (512 and 256), two learning rates (1e-4 and 5e-5), two weight decay values (0.9 and 0.09), and two loss functions (cross-entropy loss and focal loss [22]). For each dataset, we can use these parameters to train 64 ( $C_4^1 C_2^1 C_2^1 C_2^1 C_2^1$ ) suspicious models. Batch size for inference in TrainProVe is 128, and training epoch for all models is set to 100. + +Experiments on ImageNet-100: For the shadow dataset $\mathcal{D}_{sdw}$ , validation dataset $\mathcal{D}_v$ , and shadow model $M_{sdw}$ , all settings are the same as those used for CIFAR, except that $\mathcal{D}_{sdw}$ contains 50,000 images with a resolution of $256 \times 256$ and the batch size for model training is 64. The training set for the suspicious model $M_{sus}$ includes 100,000 images generated using the method in [40], with a resolution of $256 \times 256$ . For the parameter selection of the suspicious model, all settings are the same as those used in the experiments on CIFAR, except that the batch size is 64 or 32. Therefore, we can also obtain 64 suspicious models for each dataset (ImageNet-100 or ImageNet-100-Syn). Additionally, batch size for inference in TrainProVe is 16. The training epoch for all models is set to 100. + +Evaluation Metrics. Our goal is to classify the suspicious model $M_{sus}$ as either illegal or legal. Specifically, if the training data for $M_{sus}$ originates from the defender's generative model $G_d$ , then $M_{sus}$ should be classified as illegal; otherwise, $M_{sus}$ is considered legal. Given that this is a classification task, we selected three commonly used metrics for evaluation: accuracy, F1 score, and AUROC. + +# 4.2. Baselines + +Since TrainProVe is the first attempt at addressing the introduced training data provenance verification problem, there are no existing methods for direct comparison. Nonetheless, there are several baselines that can be utilized to showcase the efficacy of the proposed method, as enumerated below: + +Random classification: a rather trivial baseline that randomly classify suspicious models as illegal or legal. + +Han et al.'s work [11]: This is the closest method to ours, which trains an attributor to classify suspicious models. The underlying assumption of this method is that the defender + +
The Suspect's Data SourcesTPFPFNTNAccuracyF1 ScoreAUROC
The Defender's Generative Model GdData Sources Unrelated to Gd
Stable Diffusion [38] v1.4Latent Consistency Model064
PixArt-α6400640.9880.9700.992
Stable Cascade064
Real Data460
Latent Consistency Model [28]Stable Diffusion v1.4064
PixArt-α6400641.0001.0001.000
Stable Cascade064
Real Data064
PixArt-α [4]Stable Diffusion v1.4064
Latent Consistency Model6400641.0001.0001.000
Stable Cascade064
Real Data064
Stable Cascade [34]Stable Diffusion v1.4064
Latent Consistency Model6400641.0001.0001.000
PixArt-α064
Real Data064
Average Value0.9970.9920.998
+ +Table 1. The results of TrainProVe on CIFAR10. The "64" in the table indicates the 64 suspicious models trained based on a data source, with different training hyperparameters. Suspicious models predicted as illegal/legal are considered positive/negative instances. If the suspicious model's training data source is $G_{d}$ , it is illegal, otherwise, it is innocent. + +possesses multiple different generative models to train the attributor, while in our scenario, the defender has only a single generative model $G_{d}$ . To evaluate Han et al.'s method, we relaxed the constraints on the defender by allowing the defender to access all four text-to-image models. + +TrainProVe-Ent: This method performs hypothesis testing by analyzing the entropy of the predicted logits generated by models $M_{sdw}$ and $M_{sus}$ . Specifically, we analyze $D_v$ in batches using models $M_{sdw}$ and $M_{sus}$ to derive their respective logits, and subsequently compute their entropies $\mathcal{E}_{sdw}$ and $\mathcal{E}_{sus}$ . Following this, we employ a one-sided Grubbs' hypothesis test to ascertain whether the average of all elements in $\mathcal{E}_{sus}$ is an outlier with a high value compared to $\mathcal{E}_{sdw}$ . This analysis aids in determining the validity of model $M_{sus}$ , with its parameter configurations being consistent with those of TrainProVe. + +TrainProVe-Sim: This method conducts hypothesis testing based on the pairwise similarity of logits within the same category from $M_{sdw}$ and $M_{sus}$ . Specifically, $D_v$ is processed in batches using both $M_{sdw}$ and $M_{sus}$ to obtain their respective logits. Subsequently, the pairwise cosine similarity within the same category is computed to obtain the similarity sets $S_{sdw}$ and $S_{sus}$ . Following this, a one-sided Grubbs' hypothesis test is conducted to determine whether the average of all elements in $S_{sus}$ , is an outlier of low value in $S_{sdw}$ , thereby evaluating the validity of $M_{sus}$ . The parameter configurations align with those in TrainProVe. + +# 4.3. Experimental Results + +In this paper, $G_{d}$ is one of the four text-to-image models (Stable Diffusion v1.4, latent consistency model, PixArt-α, and Stable Cascade), and the training dataset of $M_{sus}$ + +comes from one of five data sources (Stable Diffusion v1.4, latent consistency model, PixArt- $\alpha$ , Stable Cascade and real data). We first calculate the accuracy, F1 score, and AUROC under the four different $G_{d}$ , and then average the metrics across the four scenarios as the final metrics for comparing different methods. Taking the results of TrainProVe on CIFAR10 as an example (more experimental results can be found in the supplementary material), as shown in Tab. 1. It is evident that TrainProVe can accurately classify suspicious models trained on various data sources. + +The results of various methods on different datasets are illustrated in Tab. 2, showcasing the superiority of TrainProVe. In comparison to the research by Han et al., TrainProVe has three advantages: (1) Han et al.'s approach necessitates the defender to have access to multiple different generative models, while TrainProVe is effective even when the defender only has his own generative model $G_{d}$ ; (2) The efficacy of Han et al.'s method consistently lags behind that of TrainProVe across all cases. This is because the former relies on an attributor for classification, which tends to overfit to $M_{sdw}$ 's architecture used by the defender. If the architecture of $M_{sus}$ differs, this method is likely to fail. In contrast, TrainProVe relies on hypothesis testing to classify $M_{sus}$ , thus avoiding this risk; (3) Han et al.'s method requires training a large number of models based on synthetic datasets from each generative model to train the attributor. In contrast, TrainProVe only requires training a single ResNet18 on one synthetic dataset generated by $G_{d}$ . + +Moreover, compared to TrainProVe-Ent and TrainProVe-Sim, the efficacy of TrainProVe consistently surpasses that of the others. This is because it compares the predictions of $M_{sus}$ and $M_{sdw}$ with the true labels to obtain accuracy, + +
DatasetMethodAvg. AccAvg. F1Avg. AUC
CIFAR10Random0.5000.2860.500
Han et al.'s Work0.7680.5790.731
TrainProVe-Sim0.8120.6470.821
TrainProVe-Ent0.8680.7740.902
TrainProVe0.9970.9920.998
CIFAR100Random0.5000.2860.500
Han et al.'s Work0.7300.5480.703
TrainProVe-Sim0.5750.3180.545
TrainProVe-Ent0.3980.4080.624
TrainProVe0.9920.9790.981
ImageNet-100Random0.5000.2860.500
Han et al.'s Work0.6710.4250.645
TrainProVe-Sim0.5550.3050.528
TrainProVe-Ent0.3560.3640.567
TrainProVe0.7860.4950.769
+ +Table 2. The results of TrainProVe and baselines on different datasets. "Avg. Acc", "Avg. F1", and "Avg. AUC" are the average values of accuracy, F1 score, and AUROC under four different $G_{d}$ , similar to Tab. 1. Bold represents the best results. + +
DatasetMsdwAvg. AccAvg. F1Avg. AUC
CIFAR10ResNet18 [12]0.9970.9920.998
ConvNeXt-B [26]0.9950.9890.997
Swin-B [25]0.9970.9920.998
ImageNet-100ResNet180.7860.4950.769
ConvNeXt-B0.7710.4700.756
Swin-B0.7870.4870.766
+ +using this as the basis for hypothesis testing. In contrast, TrainProVe-Ent and TrainProVe-Sim don't compare model predictions with true labels, resulting in a lack of information from the true labels, which leads to inferior validation performance compared to TrainProVe. + +# 4.4. Ablation Studies + +The model architecture of $M_{sdw}$ . We substituted the standard architecture of $M_{sdw}$ , ResNet18, with ConvNeXt-B [26] and Swin-B [25], while leaving all other configurations unaltered. The outcomes are displayed in Tab. 3, demonstrating that TrainProVe remains efficient irrespective of whether the architecture of the shadow model is based on CNNs or transformers. + +The training epochs of $M_{sdw}$ . We explored TrainProVe's performance as the $M_{sdw}$ 's training epochs (default is 100) increased from 0 to 200 in increments of 20. The findings on CIFAR10, depicted in Fig. 4, suggest that TrainProVe exhibits a commendable resilience to variations in the number of training epochs for $M_{sdw}$ . It is worth noting that an excessive or insufficient number of training epochs can lead to overfitting or underfitting of $M_{sdw}$ on $\mathcal{D}_{sdw}$ , consequently resulting in a decline in TrainProVe's performance. + +![](images/6689acce4050831022ce1e58c3ac38899680adc6e291df407ef698ca65df3bc6.jpg) + +![](images/60c29107c2a6c838fbf6210cebdfca2c4250deca4f53e067d0ecb805c06186c3.jpg) + +![](images/d2367a8550fe350ea4dae8c9501bd950d7a038d21d9a2d0cd23289497916b5bc.jpg) +Figure 4. Changing $M_{sdw}$ 's training epochs. "Avg. Acc", "Avg. F1", and "Avg. AUC" are the average values of accuracy, F1 score, and AUROC under four different $G_d$ . + +Table 3. Changing $M_{sdw}$ 's architecture. "Avg. Acc", "Avg. F1", and "Avg. AUC" are the average values of accuracy, F1 score, and AUROC under four different $G_d$ , similar to Tab. 1. + +
Text PromptsAccuracyF1 ScoreAUROC
Original Prompts0.9880.9700.992
Swapped Prompts0.9880.9700.992
+ +Table 4. Swapping $\mathcal{T}_{sdw}$ and $\mathcal{T}_v$ . Note that Stable Diffusion v1.4 is $G_{d}$ . Latent consistency model, PixArt- $\alpha$ , Stable Cascade, and real data represent other data sources. + +The text prompts of $\mathcal{D}_{sdw}$ and $\mathcal{D}_v$ . We swapped the text prompts, setting $\mathcal{T}_{sdw}' = \mathcal{T}_v$ and $\mathcal{T}_v' = \mathcal{T}_{sdw}$ , and used $\mathcal{T}_{sdw}'$ and $\mathcal{T}_v'$ to generate $\mathcal{D}_{sdw}$ and $\mathcal{D}_v$ , respectively. The results on CIFAR10, shown in Tab. 4, indicate that TrainProVe is robust to the text prompts of $\mathcal{D}_{sdw}$ and $\mathcal{D}_v$ . + +The sample sizes of $\mathcal{D}_{sdw}$ and $\mathcal{D}_v$ . We generated $\mathcal{D}_{sdw}$ and $\mathcal{D}_v$ with different sample sizes for TrainProVe, keeping other settings unchanged. Note that the default sample sizes for $\mathcal{D}_{sdw}$ and $\mathcal{D}_v$ are 30,000 and 5,000 respectively, when generating CIFAR10. The results on CIFAR10 shown in Fig. 5 demonstrate that TrainProVe has good robustness to the sample sizes. Moreover, when $\mathcal{D}_{sdw}$ 's sample size is too small, the generalization performance of $M_{sdw}$ on $\mathcal{D}_v$ may degrade, leading to a decline in TrainProVe's performance. + +# 4.5. The Interference Resistance of TrainProVe + +In this section, we study that whether TrainProVe remains effective when facing more covert data theft. More experimental details can be found in the supplementary material. + +Multiple training data sources for $M_{\text{sus}}$ . We investigated the robustness of TrainProVe when the training dataset of $M_{\text{sus}}$ has multiple data sources. Specifically, we assume that the training dataset of $M_{\text{sus}}$ contains both synthetic data from $G_d$ and real data. The results on CIFAR10 are shown in Tab. 5, where the training dataset for $M_{\text{sus}}$ consists of 120,000 images, with synthetic data and real data each accounting for half. We calculated the accuracy of TrainProVe in classifying $M_{\text{sus}}$ trained on this mixed dataset. The results indicate that TrainProVe can still accu + +![](images/1b5c18500654067f5d943d907dcd43bd092b42f3c4d0d3b40a9845c51dfabf47.jpg) + +![](images/f3af5e167b53f5c4c367d3d557967d57d66578abb99784cb231aaddbc3a9ab1d.jpg) + +![](images/8eb15b5d9bab2482c6441a65d56889b4c861e9b60b407c044fa302ed847e4ebd.jpg) + +![](images/9970fb00da91e6c00d309289c2a107d9fb74c8646c584294d0b8cfd46478d397.jpg) + +![](images/87135b8a894af98385ddac4a869ff2e93dda05c198959ce56ed95d47f9cdfe85.jpg) +Figure 5. Changing the sample sizes of $\mathcal{D}_{sdw}$ and $\mathcal{D}_v$ . "Avg. Acc", "Avg. F1", and "Avg. AUC" are the average values of accuracy, F1 score, and AUROC under four different $G_{d}$ . + +![](images/3ab8185a87e9ef6857b882fab178a5f164783fd31590823f3b7fd9b00d492afc.jpg) + +rately detect illegal behavior in this more difficult situation. + +Fine-tuning $G_{d}$ . When the suspect has access to a small amount of real data, he may fine-tune the defender's open-source model $G_{d}$ and then generate data to train $M_{sus}$ , which can enhance $M_{sus}$ 's generalization ability on real data. Note that the defender only possesses the original $G_{d}$ and has no knowledge of fine-tuning. Specifically, we fine-tuned the pre-trained Stable Diffusion v1.4 using 10,000 ImageNet-100 images (100 per class) and LoRA [16] with 10 epochs, and then generated ImageNet-100-Syn to train $M_{sus}$ . The results are shown in Tab. 6a, which indicate that TrainProVe remains effective in this more arduous scenario. + +Fine-tuning $M_{sus}$ . Similarly, when the suspect has a small amount of real data, he may also fine-tune $M_{sus}$ with real data, which was trained on synthetic data from $G_{d}$ , to enhance $M_{sus}$ 's generalization ability on real data. Specifically, we fine-tuned $M_{sus}$ using 10,000 ImageNet-100 images (100 per class), where $M_{sus}$ was pre-trained on synthetic data from Stable Diffusion v1.4, with 10 fine-tuning epochs. The results, shown in Tab. 6b, indicate even in this more challenging scenario, TrainProVe remains effective. + +# 4.6. Efficiency Analysis + +We analyzed the efficiency of TrainProVe and the work of Han et al. on ImageNet-100. Specifically, following the setup in Sec. 4.1, given $M_{sus}$ trained on ImageNet-100 or ImageNet-100-Syn, we performed detection using both methods on an NVIDIA GeForce RTX 4090. $G_{d}$ is Sta + +
M_sus's Training SourceG_dTPFPAcc
CIFAR10-SynStable Diffusion v1.46401
Latent Consistency Model6401
+CIFAR10PixArt-α6401
Stable Cascade6401
+ +Table 5. Results of TrainProVe when the training dataset of $M_{sus}$ has multiple data sources. The "64" in the table indicates the 64 suspicious models with different training hyperparameters. + +
TPFPAcc
w/o fine-tuning6401.00
w/ fine-tuning6220.97
+ +(a) Fine-tuning $G_{d}$ + +
TPFPAcc
w/o fine-tuning6401.00
w/ fine-tuning5770.89
+ +(b) Fine-tuning $M_{sus}$ +Table 6. Fine-tuning $G_{d}$ or $M_{sus}$ . The "64" in the table indicates the 64 suspicious models with different training hyperparameters. + +
MethodτgτtτvTotalAccF1AUC
Han et al.'s Work137.7h8.0h1.2h146.9h0.480.280.48
TrainProVe33.6h0.1h1.4h35.1h0.780.630.84
+ +Table 7. Efficiency analysis. "Acc", "F1" and "AUC" refer to the accuracy, F1 score and AUROC respectively. + +ble Cascade. The runtime and results is shown in Tab. 7. For TrainProVe, $\tau_{g}$ is the time for generating $D_{sdw}$ and $D_v$ , $\tau_t$ is the time for training $M_{sdw}$ , and $\tau_v$ is the time for classifying $M_{sus}$ using hypothesis testing. For Han et al.'s method, $\tau_{g}$ is the time for generating the training and validation dataset, $\tau_t$ is the time for training the models and the attributor, and $\tau_v$ is the time for classifying $M_{sus}$ using the attributor. From the results, it shows that the high efficiency of TrainProVe lies in its significantly smaller $\tau_{g}$ and $\tau_t$ compared to Han et al.'s method. This is because TrainProVe only requires generating one dataset ( $D_{sdw}$ ) and training one model ( $M_{sdw}$ ). In contrast, Han et al.'s method requires generating $m$ datasets and training $m \times n$ models, where $m$ is the number of text-to-image models accessible to the defender and $n$ is the number of models trained by the defender on each dataset. These models are then used to train the attributor. Moreover, TrainProVe achieved better classification results, further illustrating its superiority. + +# 5. Conclusion + +Open-source text-to-image models have reduced the cost of obtaining images. We introduce an important yet largely unexplored security issue: whether a suspicious model's training data originates from a specific text-to-image model? To our knowledge, we are the first to propose a solution to this problem, called TrainProVe. The success of TrainProVe in verifying training data provenance has been demonstrated through theory and experiments. Promising future work includes: (1) extending TrainProVe to other image generative models, such as image-to-image models; (2) adapting TrainProVe to protect other types of data (e.g., text); (3) exploring other privacy risks about text-to-image models. + +# 6. Acknowledgment + +This work was partially supported by the Pioneer R&D Program of Zhejiang (No.2024C01021), and Zhejiang Province High-Level Talents Special Support Program "Leading Talent of Technological Innovation of Ten-Thousands Talents Program" (No. 2022R52046). + +# References + +[1] Hritik Bansal and Aditya Grover. Leaving reality to imagination: Robust classification via generated datasets. arXiv preprint arXiv:2302.02503, 2023. 1 +[2] Shai Ben-David, John Blitzer, Koby Crammer, and Fernando Pereira. Analysis of representations for domain adaptation. Advances in neural information processing systems, 19, 2006. 3, 1 +[3] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- $\sigma$ : Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 1 +[4] Junsong Chen, YU Jincheng, GE Chongjian, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- $\alpha$ : Fast training of diffusion transformer for photorealistic text-to-image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 5, 6, 4, 7, 8 +[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5, 2 +[6] Tim Dockhorn, Robin Rombach, Andreas Blatmann, and Yaoliang Yu. Distilling the knowledge in diffusion models. In CVPR Workshop Generative Models for Computer Vision, 2023. 1 +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5 +[8] Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7382-7392, 2024. 1 +[9] Abolfazl Farahani, Sahar Voghoei, Khaled Rasheed, and Hamid R Arabnia. A brief review of domain adaptation. Advances in data science and information engineering: proceedings from ICDATA 2020 and IKE 2020, pages 877-894, 2021. 3 +[10] Frank Ephraim Grubbs. Sample criteria for testing outlying observations. University of Michigan, 1949. 4 +[11] Ge Han, Ahmed Salem, Zheng Li, Shanqing Guo, Michael Backes, and Yang Zhang. Detection and attribution of models trained on generated data. In ICASSP 2024-2024 IEEE + +International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4875-4879. IEEE, 2024. 5 +[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5, 7 +[13] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is synthetic data from generative models ready for image recognition? arXiv preprint arXiv:2210.07574, 2022. 1 +[14] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 1 +[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2 +[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 8 +[17] Yongcheng Jing, Yiding Yang, Xinchao Wang, Mingli Song, and Dacheng Tao. Turning frequency to resolution: Video super-resolution via event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7772-7781, 2021. 2 +[18] Yongcheng Jing, Chongbin Yuan, Li Ju, Yiding Yang, Xinchao Wang, and Dacheng Tao. Deep graph reprogramming. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24345-24354, 2023. 2 +[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2 +[20] A Krizhevsky. Learning multiple layers of features from tiny images. Master's thesis, University of Tront, 2009. 5, 2 +[21] Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pretrained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems, 36, 2023. 1 +[22] T Lin. Focal loss for dense object detection. arXiv preprint arXiv:1708.02002, 2017. 5 +[23] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2 +[24] Yugeng Liu, Zheng Li, Michael Backes, Yun Shen, and Yang Zhang. Watermarking diffusion model. arXiv preprint arXiv:2305.12502, 2023. 1, 3 +[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 7 +[26] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the + +2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 7 +[27] Ge Luo, Junqiang Huang, Manman Zhang, Zhenxing Qian, Sheng Li, and Xinpeng Zhang. Steal my artworks for fine-tuning? a watermarking framework for detecting art theft mimicry in text-to-image models. arXiv preprint arXiv:2311.13619, 2023. 3 +[28] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 1, 3, 5, 6, 2, 4, 7, 8 +[29] Kidist Amde Mekonnen, Nicola Dall'Asen, and Paolo Rota. Adv-kd: Adversarial knowledge distillation for faster diffusion sampling. arXiv preprint arXiv:2405.20675, 2024. 1 +[30] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41, 1995. 2 +[31] Rui Min, Sen Li, Hongyang Chen, and Minhao Cheng. A watermark-conditioned diffusion model for ip protection. arXiv preprint arXiv:2403.10893, 2024. 1 +[32] Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions on knowledge and data engineering, 22(10):1345-1359, 2009. 3 +[33] Suraj Patil, William Berman, Robin Rombach, and Patrick von Platen. amused: An open muse reproduction. arXiv preprint arXiv:2401.01808, 2024. 1 +[34] Pablo Pernias, Dominic Rampas, Mats L. Richter, Christopher J. Pal, and Marc Aubreville. Wuerstchen: An efficient architecture for large-scale text-to-image diffusion models. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 5, 6, 4, 7, 8 +[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 5, 3 +[36] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1 +[37] Mauro Ribeiro, Katarina Grolinger, and Miriam AM Capretz. Mlaas: Machine learning as a service. In 2015 IEEE 14th international conference on machine learning and applications (ICMLA), pages 896-902. IEEE, 2015. 2 +[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3, 5, 6, 4, 7, 8 +[39] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In International Conference on Learning Representations, 2022. 1 +[40] Mert Bülent Sariyildüz, Karteek Alahari, Diane Larlus, and Yannis Kalantidis. Fake it till you make it: Learning transferable representations from synthetic imagenet clones. In + +Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8011-8021, 2023. 1, 2, 5 +[41] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2 +[42] Jordan Shipard, Arnold Wiliem, Kien Nguyen Thanh, Wei Xiang, and Clinton Fookes. Diversity is definitely needed: Improving model-agnostic zero-shot classification via stable diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 769–778, 2023. 1, 2, 5 +[43] Geng Sun, Tingru Cui, Jianming Yong, Jun Shen, and Shiping Chen. Mlaas: a cloud-based system for delivering adaptive micro learning in mobile mooc learning. IEEE Transactions on Services Computing, 11(2):292-305, 2015. 2 +[44] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2117-2126, 2020. 1 +[45] Lisa Torrey and Jude Shavlik. Transfer learning. In Handbook of research on machine learning applications and trends: algorithms, methods, and techniques, pages 242-264. IGI global, 2010. 3 +[46] Mei Wang and Weihong Deng. Deep visual domain adaptation: A survey. Neurocomputing, 312:135-153, 2018. 3 +[47] Zhenting Wang, Chen Chen, Yuchen Liu, Lingjuan Lyu, Dimitris Metaxas, and Shiqing Ma. How to detect unauthorized data usages in text-to-image diffusion models. arXiv preprint arXiv:2307.03108, 2023. 1 +[48] Zhenting Wang, Chen Chen, Yi Zeng, Lingjuan Lyu, and Shiqing Ma. Where did i come from? origin attribution of ai-generated images. Advances in neural information processing systems, 36, 2023. 1 +[49] Qizhen Weng, Wencong Xiao, Yinghao Yu, Wei Wang, Cheng Wang, Jian He, Yong Li, Liping Zhang, Wei Lin, and Yu Ding. Mlaas in the wild: Workload analysis and scheduling in large-scale heterogeneousgpu clusters. In 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI 22), pages 945–960, 2022. 2 +[50] Qianlong Xiang, Miao Zhang, Yuzhang Shang, Jianlong Wu, Yan Yan, and Liqiang Nie. Dkdm: Data-free knowledge distillation for diffusion models with any architecture. arXiv preprint arXiv:2409.03550, 2024. 1 +[51] Yuechen Xie, Jie Song, Mengqi Xue, Haofei Zhang, Xingen Wang, Bingde Hu, Genlang Chen, and Mingli Song. Dataset ownership verification in contrastive pre-trained models. arXiv preprint arXiv:2502.07276, 2025.3 +[52] Cheng Xiong, Chuan Qin, Guorui Feng, and Xinpeng Zhang. Flexible and secure watermarking for latent diffusion model. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1668-1676, 2023. 1 +[53] Jianhao Yuan, Francesco Pinto, Adam Davies, and Philip Torr. Not just pretty pictures: Toward interventional data + +augmentation using text-to-image generators. arXiv preprint arXiv:2212.11237, 2022. 1 +[54] Zihan Yuan, Li Li, Zichi Wang, and Xinpeng Zhang. Watermarking for stable diffusion models. IEEE Internet of Things Journal, 2024. 1 +[55] Likun Zhang, Hao Wu, Lingcui Zhang, Fengyuan Xu, Jin Cao, Fenghua Li, and Ben Niu. Training data attribution: Was your model secretly trained on data created by mine? arXiv preprint arXiv:2409.15781, 2024. 1, 3 +[56] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Ngai-Man Cheung, and Min Lin. A recipe for watermarking diffusion models. arXiv preprint arXiv:2303.10137, 2023. 1, 3 +[57] Yongchao Zhou, Hshmat Sahak, and Jimmy Ba. Training on thin air: Improve image classification with generated data. arXiv preprint arXiv:2305.15316, 2023. 1, 2 +[58] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2020. 3 \ No newline at end of file diff --git a/2025/Training Data Provenance Verification_ Did Your Model Use Synthetic Data from My Generative Model for Training_/images.zip b/2025/Training Data Provenance Verification_ Did Your Model Use Synthetic Data from My Generative Model for Training_/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2b275310c709867577526f4f6568ac21c7963d7d --- /dev/null +++ b/2025/Training Data Provenance Verification_ Did Your Model Use Synthetic Data from My Generative Model for Training_/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:39343ebb553c7d0e174a6ae2b74ac9137209513a6ae9b8b5f0a8b77fa241379d +size 637424 diff --git a/2025/Training Data Provenance Verification_ Did Your Model Use Synthetic Data from My Generative Model for Training_/layout.json b/2025/Training Data Provenance Verification_ Did Your Model Use Synthetic Data from My Generative Model for Training_/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6e0f96fcf004dd74d2b2471e973b5f80ab469e36 --- /dev/null +++ b/2025/Training Data Provenance Verification_ Did Your Model Use Synthetic Data from My Generative Model for Training_/layout.json @@ -0,0 +1,12765 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 72, + 102, + 539, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 102, + 539, + 140 + ], + "spans": [ + { + "bbox": [ + 72, + 102, + 539, + 140 + ], + "type": "text", + "content": "Training Data Provenance Verification: Did Your Model Use Synthetic Data from My Generative Model for Training?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 160, + 449, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 160, + 449, + 177 + ], + "spans": [ + { + "bbox": [ + 160, + 160, + 449, + 177 + ], + "type": "text", + "content": "Yuechen Xie" + }, + { + "bbox": [ + 160, + 160, + 449, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 160, + 160, + 449, + 177 + ], + "type": "text", + "content": ", Jie Song" + }, + { + "bbox": [ + 160, + 160, + 449, + 177 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 160, + 160, + 449, + 177 + ], + "type": "text", + "content": ", Huiqiong Wang" + }, + { + "bbox": [ + 160, + 160, + 449, + 177 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 160, + 160, + 449, + 177 + ], + "type": "text", + "content": ", Mingli Song" + }, + { + "bbox": [ + 160, + 160, + 449, + 177 + ], + "type": "inline_equation", + "content": "^{1,3,4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 255, + 180, + 357, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 180, + 357, + 195 + ], + "spans": [ + { + "bbox": [ + 255, + 180, + 357, + 195 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 255, + 180, + 357, + 195 + ], + "type": "text", + "content": "Zhejiang University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 188, + 195, + 422, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 195, + 422, + 209 + ], + "spans": [ + { + "bbox": [ + 188, + 195, + 422, + 209 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 188, + 195, + 422, + 209 + ], + "type": "text", + "content": "Ningbo Innovation Center, Zhejiang University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 209, + 477, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 209, + 477, + 223 + ], + "spans": [ + { + "bbox": [ + 133, + 209, + 477, + 223 + ], + "type": "text", + "content": "3State Key Laboratory of Blockchain and Security, Zhejiang University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 110, + 223, + 500, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 223, + 500, + 237 + ], + "spans": [ + { + "bbox": [ + 110, + 223, + 500, + 237 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 110, + 223, + 500, + 237 + ], + "type": "text", + "content": "Hangzhou High-Tech Zone (Binjiang) Institute of Blockchain and Data Security" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 169, + 244, + 437, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 244, + 437, + 256 + ], + "spans": [ + { + "bbox": [ + 169, + 244, + 437, + 256 + ], + "type": "text", + "content": "{xyuechen,sjie,huiqiong-wang,brooksong}@zju.edu.cn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 284, + 200, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 284, + 200, + 296 + ], + "spans": [ + { + "bbox": [ + 151, + 284, + 200, + 296 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 310, + 297, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 297, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 297, + 574 + ], + "type": "text", + "content": "High-quality open-source text-to-image models have lowered the threshold for obtaining photorealistic images significantly, but also face potential risks of misuse. Specifically, suspects may use synthetic data generated by these generative models to train models for specific tasks without permission, when lacking real data resources especially. Protecting these generative models is crucial for the wellbeing of their owners. In this work, we propose the first method to this important yet unresolved issue, called Training data Provenance Verification (TrainProVe). The rationale behind TrainProVe is grounded in the principle of generalization error bound, which suggests that, for two models with the same task, if the distance between their training data distributions is smaller, their generalization ability will be closer. We validate the efficacy of TrainProVe across four text-to-image models (Stable Diffusion v1.4, latent consistency model, PixArt-α, and Stable Cascade). The results show that TrainProVe achieves a verification accuracy of over 99% in determining the provenance of suspicious model training data, surpassing all previous methods. Code is available at https://github.com/xieyc99/TrainProVe." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 598, + 135, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 598, + 135, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 598, + 135, + 611 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "type": "text", + "content": "High-quality open-source text-to-image models based on diffusion models [15] can generate photorealistic images with simple text prompts [4, 28, 34, 38], making them highly significant in terms of lowering creative barriers, enabling personalized customization, etc. Currently, most open-source text-to-image models are only permitted for" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 285, + 555, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 344 + ], + "type": "text", + "content": "academic or educational purposes, and commercial use is prohibited without explicit permission [3, 4, 21, 28, 33, 34, 36, 38]. Therefore, protecting the intellectual property of these generative models is crucial for safeguarding the interests of their owners." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 346, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 407 + ], + "type": "text", + "content": "Currently, one of the major security challenges faced by defenders, i.e., the generative model owners, is the unauthorized use of the synthetic data from their models by suspects [24, 47, 48, 55]. As shown in Fig. 1, based on the suspect's intention, this issue can be divided into three cases:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 408, + 554, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 554, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 554, + 445 + ], + "type": "text", + "content": "Case 1: the suspect steals the synthetic data illegally and claims it as his own personal work [24, 44, 48], implying the intellectual property of the generative model is infringed." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 447, + 555, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 518 + ], + "type": "text", + "content": "Case 2: the suspect steals the synthetic data from the defender's generative model, and trains another generative model for the same task using knowledge distillation techniques [6, 14, 29, 39, 50]. In this case, the defender cannot access the synthetic data used by the suspect, making it a more challenging defense task compared to Case 1." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 520, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 555, + 616 + ], + "type": "text", + "content": "Case 3: the suspect steals the synthetic data from the generative model to train another model for a specific task [1, 8, 13, 40, 42, 53, 57]. In this case, the defender is not only unaware of the training data used by the suspicious model, but also cannot compare the similarity between the generative model and the suspicious model directly due to the different tasks, representing an even more arduous defense task compared to Case 2." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "content": "For Case 1 and Case 2, there are already some mature methods to address these issues, such as adding watermarks [24, 31, 52, 54, 56] and training data attribution [55], which identify illegal activities by checking the distributions of the suspicious images or the synthetic images from suspicious model. However, for Case 3, although the potential of using synthetic data to train models for other tasks has been recognized gradually in recent years [13, 40, 42, 53]" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23817" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 276, + 407 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 276, + 407 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 276, + 407 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 276, + 407 + ], + "type": "image", + "image_path": "f01004709656455b6d0c5b1cc26c89ab12083bcfe969bba33cf233c779edd264.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 413, + 296, + 468 + ], + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 468 + ], + "type": "text", + "content": "Figure 1. The overview of the three cases. These three cases encompass nearly every conceivable real-world scenario, where a suspect uses synthetic images from the defender's text-to-image model illegally. In this paper, we are committed to addressing the security risks in Case 3." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 483, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 567 + ], + "type": "text", + "content": "due to the lower collection costs compared to real data, the associated security risks have been overlooked. Methods that work for the first two cases cannot be applied to the third case, because the learning task of the suspicious model is not necessarily image generation. For example, for a suspicious image classification model, there are no synthetic data available for defenders to examine." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": "In this work, we address this important yet unexplored issue. To the best of our knowledge, we propose the first method for protecting the intellectual property of generative models in such scenario, which we call Training data Provenance Verification (TrainProVe). It helps defenders verify whether the suspicious model has been trained using synthetic data from their generative model. Given that suspicious models may use synthetic data for training without authorization, and is used for commercial purposes, we focus on the black-box setting, where the defenders have no information about the training configuration of the suspicious model (e.g., loss function, model architecture, etc.)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "and can only access the model through Machine Learning as a Service (MLaaS) [37, 43, 49]. This means that the defenders can only obtain prediction results (such as prediction logits or labels) via the suspicious model's API." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 121, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 361 + ], + "type": "text", + "content": "The rationale behind TrainProVe is grounded in the principle of generalization error bound, which suggests that, for two models with the same task, if the distance between their training data distributions is smaller, their generalization ability will be closer. Specifically, defenders can train a model for the same task as the suspicious model, using the synthetic data generated by their generative model. By comparing the performances of these two models, defenders can determine whether the suspicious model was trained on synthetic data from their generative model. More concretely, the proposed TrainProVe involves three steps: (1) using the defender's generative model along with different text prompts to generate a shadow dataset and a validation dataset, based on the suspicious model's task; (2) training a shadow model for the same task as the suspicious model using the shadow dataset; (3) performing hypothesis testing on the prediction accuracy of the suspicious model and the shadow model on the validation dataset, in order to determine whether the suspicious model was trained using synthetic data from the defender's generative model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 362, + 556, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 556, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 556, + 555 + ], + "type": "text", + "content": "In summary, the principal contributions of this paper are threefold: (1) we introduce the important yet under-explored issue of intellectual property protection for text-to-image models, in which the suspect uses the synthetic data from the defender's generative model illegally to train a specific-task suspicious model, such as an image classification model. For the defender, the suspicious model is a black-box; (2) we discover and demonstrate that the similarity between different models is related to their source domains closely, and based on this, we propose the training data provenance verification technique. To the best of our knowledge, this is the first method to address such intellectual property protection issue; (3) Extensive experiments show that TrainProVe achieves over " + }, + { + "bbox": [ + 313, + 362, + 556, + 555 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 313, + 362, + 556, + 555 + ], + "type": "text", + "content": " accuracy in verifying training data provenance for suspicious models, surpassing all previous studies." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 568, + 400, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 568, + 400, + 582 + ], + "spans": [ + { + "bbox": [ + 314, + 568, + 400, + 582 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Text-to-image diffusion model. Diffusion models [15] have revolutionized generative modeling fundamentally, in terms of quality and generalization. Their excellent scalability has led to significant advancements in specific applications, such as text-to-image generation. Text-to-image diffusion models [4, 34, 38] use the diffusion process to transform text descriptions into realistic images. Recently, people have begun to recognize the potential of text-to-image diffusion models for synthesizing dataset [17, 18, 40, 42, 57]. However, so far, no one has realized that suspects" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 282, + 72, + 294, + 375 + ], + "type": "aside_text", + "angle": 90, + "lines": [ + { + "bbox": [ + 282, + 72, + 294, + 375 + ], + "spans": [ + { + "bbox": [ + 282, + 72, + 294, + 375 + ], + "type": "text", + "content": "Low Defense Difficulty High" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23818" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "might use these text-to-image diffusion models to synthesize datasets and train their own models without authorization. Therefore, in this paper, we propose the first feasible solution to address this issue, securing and fostering healthy development in this field." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 136, + 295, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 136, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 136, + 295, + 316 + ], + "type": "text", + "content": "Text-to-image model protection. Training generative models requires large amounts of data and computational resources [4, 38], making the intellectual property of these models crucial for their owners. When these models are open-source [4, 28, 34, 38], their intellectual property is vulnerable to infringement by suspects. Specifically, for text-to-image models, suspects have the following three ways primarily to infringe on their intellectual property: (1) stealing synthetic data illegally; (2) illegal knowledge distillation; (3) training suspicious models for specific tasks using synthetic data illegally. For the first two cases, watermarking [24, 27, 56] and training data attribution [55] have been developed to address these issues. However, for the third case, no suitable solution exists currently. This work aims to address this gap." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "type": "text", + "content": "Generalization error bound. Generalization error (upper) bound is used to describe a model's performance on testing data and widely applied in domain adaptation [9, 46], which is a special case of transfer learning [45, 51, 58]. Theorem 1 in [2] states the generalization error upper bound of model " + }, + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "type": "text", + "content": " on the target domain " + }, + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "type": "text", + "content": " be a hypothesis space with VC dimension " + }, + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "type": "text", + "content": ", and let " + }, + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "type": "text", + "content": " be the sample size of the dataset in the source domain " + }, + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "type": "text", + "content": ". Then, with probability at least " + }, + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "type": "inline_equation", + "content": "1 - \\delta" + }, + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "type": "text", + "content": ", for every " + }, + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "type": "inline_equation", + "content": "M \\in \\mathcal{H}" + }, + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 434, + 295, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 475 + ], + "type": "interline_equation", + "content": "\\epsilon_ {T} (M) \\leq \\hat {\\epsilon} _ {S} (M) + \\sqrt {\\frac {4}{m} \\left(d \\log \\frac {2 e m}{d}\\right)} + d _ {\\mathcal {H}} \\left(P _ {S}, P _ {T}\\right) + \\lambda , \\tag {1}", + "image_path": "e768aff3f758a55b62efecdcfd447edba97707fd6686454bb97616604939bf6c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "type": "inline_equation", + "content": "\\epsilon_T(M)" + }, + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon_S} (M)" + }, + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "type": "text", + "content": " represent the generalization error of " + }, + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "type": "text", + "content": " on the target domain and the empirical error of " + }, + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "type": "text", + "content": " on the source domain, respectively. " + }, + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "type": "inline_equation", + "content": "P_{S}" + }, + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "type": "inline_equation", + "content": "P_{T}" + }, + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "type": "text", + "content": " denote the marginal probability distributions of the source and target domains, respectively. " + }, + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "type": "inline_equation", + "content": "d_{\\mathcal{H}}(P_S,P_T)" + }, + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "type": "text", + "content": "-divergence between the source and target domains, which measures the distance between the two distributions within the hypothesis space " + }, + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "type": "text", + "content": " is a constant and " + }, + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "type": "text", + "content": " is the base of the natural logarithm. The rationale behind TrainProVe is grounded in this theorem, which suggests that, for two models with the same task, if the distance between their training data distributions is smaller, their generalization ability will be closer." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 628, + 111, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 111, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 111, + 640 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 647, + 180, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 180, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 180, + 659 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "In this study, we focus on the training data provenance verification task in image classification. This task involves two roles: the defender and the suspect. The defender, embodying the role of text-to-image model provider, endeavors to" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "ascertain whether the black-box suspicious model, " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": ", has been illicitly trained on synthetic dataset generated by his model " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "G_{d}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " (as shown in Case 3 of Fig. 1). More specifically, for a black-box suspect model, the defender's objective is to determine whether it is illegal or legal based solely on the labels it outputs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 144, + 555, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 264 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 313, + 144, + 555, + 264 + ], + "type": "text", + "content": " can be classified as either illegal or legal based on its training datasets: (1) Illegal: " + }, + { + "bbox": [ + 313, + 144, + 555, + 264 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 313, + 144, + 555, + 264 + ], + "type": "text", + "content": " is trained on the synthetic data from the defender's model " + }, + { + "bbox": [ + 313, + 144, + 555, + 264 + ], + "type": "inline_equation", + "content": "G_{d}" + }, + { + "bbox": [ + 313, + 144, + 555, + 264 + ], + "type": "text", + "content": ", indicating the occurrence of intellectual property infringement; (2) Legal: " + }, + { + "bbox": [ + 313, + 144, + 555, + 264 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 313, + 144, + 555, + 264 + ], + "type": "text", + "content": " is trained using a different data source independent of the defender, such as some real data or synthetic data from other generative models, implying the innocence of the suspect. To be precise, when presented with a black-box suspicious model, the defender's objective is to determine whether it is an illegal or legal model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 272, + 447, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 272, + 447, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 272, + 447, + 285 + ], + "type": "text", + "content": "3.2. The Theoretical Insight" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 289, + 554, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 554, + 409 + ], + "type": "text", + "content": "The underlying rationale for TrainProVe is that when two models are assigned the same task, their ability to generalize will be more aligned if the divergence between their training data distributions is diminished. We demonstrate this through a case study of " + }, + { + "bbox": [ + 313, + 289, + 554, + 409 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 289, + 554, + 409 + ], + "type": "text", + "content": "-way classification, where " + }, + { + "bbox": [ + 313, + 289, + 554, + 409 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\mathbb{R}^{C \\times W \\times H}" + }, + { + "bbox": [ + 313, + 289, + 554, + 409 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 289, + 554, + 409 + ], + "type": "inline_equation", + "content": "\\mathcal{Y} = \\{1, \\dots, K\\}" + }, + { + "bbox": [ + 313, + 289, + 554, + 409 + ], + "type": "text", + "content": " represent the input and label spaces respectively. We follow the setting of previous domain adaptation works [32, 46], viewing a domain as composed of a feature space " + }, + { + "bbox": [ + 313, + 289, + 554, + 409 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 289, + 554, + 409 + ], + "type": "text", + "content": " and a marginal probability distribution " + }, + { + "bbox": [ + 313, + 289, + 554, + 409 + ], + "type": "inline_equation", + "content": "P(\\boldsymbol{x})" + }, + { + "bbox": [ + 313, + 289, + 554, + 409 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 289, + 554, + 409 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{X}, P(\\boldsymbol{x})\\}" + }, + { + "bbox": [ + 313, + 289, + 554, + 409 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 289, + 554, + 409 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x} \\in \\mathcal{X}" + }, + { + "bbox": [ + 313, + 289, + 554, + 409 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 410, + 554, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 554, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 554, + 509 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 410, + 554, + 509 + ], + "type": "inline_equation", + "content": "S_{1} = \\{\\mathcal{X},P_{1}(\\pmb {x})\\}" + }, + { + "bbox": [ + 313, + 410, + 554, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 410, + 554, + 509 + ], + "type": "inline_equation", + "content": "S_{2} = \\{\\mathcal{X},P_{2}(\\pmb {x})\\}" + }, + { + "bbox": [ + 313, + 410, + 554, + 509 + ], + "type": "text", + "content": " be two source domains sharing the same feature space but possessing distinct marginal probability distributions, i.e., " + }, + { + "bbox": [ + 313, + 410, + 554, + 509 + ], + "type": "inline_equation", + "content": "P_{1}(\\pmb {x})\\neq" + }, + { + "bbox": [ + 313, + 410, + 554, + 509 + ], + "type": "inline_equation", + "content": "P_{2}(\\pmb {x}).D_{1} = \\{(x_{1}^{i},y_{1}^{i})\\}_{i = 1}^{|D_{1}|}" + }, + { + "bbox": [ + 313, + 410, + 554, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 410, + 554, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_2 = \\{(\\pmb {x}_2^j,\\pmb {y}_2^j)\\}_{j = 1}^{|D_2|}" + }, + { + "bbox": [ + 313, + 410, + 554, + 509 + ], + "type": "text", + "content": " represent the datasets sampled from the source domains " + }, + { + "bbox": [ + 313, + 410, + 554, + 509 + ], + "type": "inline_equation", + "content": "S_{1}" + }, + { + "bbox": [ + 313, + 410, + 554, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 410, + 554, + 509 + ], + "type": "inline_equation", + "content": "S_{2}" + }, + { + "bbox": [ + 313, + 410, + 554, + 509 + ], + "type": "text", + "content": " respectively, where " + }, + { + "bbox": [ + 313, + 410, + 554, + 509 + ], + "type": "inline_equation", + "content": "\\pmb{x}_1^i\\sim P_1(\\pmb {x})" + }, + { + "bbox": [ + 313, + 410, + 554, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 410, + 554, + 509 + ], + "type": "inline_equation", + "content": "\\pmb{x}_2^j\\sim P_2(\\pmb {x})" + }, + { + "bbox": [ + 313, + 410, + 554, + 509 + ], + "type": "text", + "content": " . " + }, + { + "bbox": [ + 313, + 410, + 554, + 509 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 410, + 554, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 410, + 554, + 509 + ], + "type": "inline_equation", + "content": "\\hat{M}" + }, + { + "bbox": [ + 313, + 410, + 554, + 509 + ], + "type": "text", + "content": " are the models trained on " + }, + { + "bbox": [ + 313, + 410, + 554, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_1" + }, + { + "bbox": [ + 313, + 410, + 554, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 410, + 554, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_2" + }, + { + "bbox": [ + 313, + 410, + 554, + 509 + ], + "type": "text", + "content": " respectively. Then we have the following theorem:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 517, + 554, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 554, + 590 + ], + "type": "text", + "content": "Theorem 1. Assume " + }, + { + "bbox": [ + 313, + 517, + 554, + 590 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 517, + 554, + 590 + ], + "type": "text", + "content": " is trained on synthetic data generated by a text-to-image model " + }, + { + "bbox": [ + 313, + 517, + 554, + 590 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 517, + 554, + 590 + ], + "type": "text", + "content": " with a set of text prompts " + }, + { + "bbox": [ + 313, + 517, + 554, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_1" + }, + { + "bbox": [ + 313, + 517, + 554, + 590 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 517, + 554, + 590 + ], + "type": "inline_equation", + "content": "P_{1}(\\pmb {x}) = P(\\pmb {x}|G,\\mathcal{T}_{1})" + }, + { + "bbox": [ + 313, + 517, + 554, + 590 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 517, + 554, + 590 + ], + "type": "inline_equation", + "content": "\\hat{M}" + }, + { + "bbox": [ + 313, + 517, + 554, + 590 + ], + "type": "text", + "content": " can be trained on either real data or synthetic data generated by any text-to-image model. Based on the generalization errors of " + }, + { + "bbox": [ + 313, + 517, + 554, + 590 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 517, + 554, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 517, + 554, + 590 + ], + "type": "inline_equation", + "content": "\\hat{M}" + }, + { + "bbox": [ + 313, + 517, + 554, + 590 + ], + "type": "text", + "content": " on the target domain " + }, + { + "bbox": [ + 313, + 517, + 554, + 590 + ], + "type": "inline_equation", + "content": "T = \\{\\mathcal{X},P(\\pmb {x}|G,\\mathcal{T}_t)\\}" + }, + { + "bbox": [ + 313, + 517, + 554, + 590 + ], + "type": "text", + "content": ", we have:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 351, + 599, + 553, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 599, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 351, + 599, + 553, + 620 + ], + "type": "interline_equation", + "content": "\\sup _ {P _ {2} (\\boldsymbol {x}) = P (\\boldsymbol {x} \\mid G, \\mathcal {T} _ {2})} | \\Delta \\epsilon_ {T} | \\leq \\sup _ {P _ {2} (\\boldsymbol {x}) \\perp G} | \\Delta \\epsilon_ {T} |, \\tag {2}", + "image_path": "c1bf087c5992d694841a13b4df53b0fecc4a86fff550d9dd5a628e1943cf01f5.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\Delta \\epsilon_{T}" + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "content": " represents the difference in generalization error between " + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\hat{M}" + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "content": " on the target domain " + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "content": ", expressed as " + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\Delta \\epsilon_{T} = \\epsilon_{T}(M) - \\epsilon_{T}(\\hat{M})" + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "inline_equation", + "content": "P_{2}(\\pmb{x}) \\perp G" + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "content": " denotes that " + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "inline_equation", + "content": "P_{2}(\\pmb{x})" + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "content": " is independent of " + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "content": ", meaning " + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "inline_equation", + "content": "P_{2}(\\pmb{x}) = P(\\pmb{x}|G^{\\prime},\\mathcal{T}_{2})" + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "inline_equation", + "content": "P_{2}(\\pmb{x}) = P_{R}(\\pmb{x})" + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "inline_equation", + "content": "G^{\\prime}" + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "content": " is a text-to-image model different from " + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "inline_equation", + "content": "P_{R}(\\pmb{x})" + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "content": " represents the distribution of real data. " + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_1, \\mathcal{T}_2" + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t" + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "content": " are distinct sets of text prompts." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23819" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 288, + 283 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 288, + 283 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 288, + 283 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 288, + 283 + ], + "type": "image", + "image_path": "6b78a5229db3a7dec00d36cbfb66c8bbe8c062cb07f41a0e37dd762dab0b0c95.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 293, + 555, + 329 + ], + "lines": [ + { + "bbox": [ + 54, + 293, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 54, + 293, + 555, + 329 + ], + "type": "text", + "content": "Figure 2. The overview of TrainProVe's motivation. " + }, + { + "bbox": [ + 54, + 293, + 555, + 329 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 54, + 293, + 555, + 329 + ], + "type": "text", + "content": " is trained on the dataset whose data distribution is " + }, + { + "bbox": [ + 54, + 293, + 555, + 329 + ], + "type": "inline_equation", + "content": "P(x|G, \\mathcal{T}_1)" + }, + { + "bbox": [ + 54, + 293, + 555, + 329 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 54, + 293, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\hat{M}_1" + }, + { + "bbox": [ + 54, + 293, + 555, + 329 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 293, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\hat{M}_2" + }, + { + "bbox": [ + 54, + 293, + 555, + 329 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 293, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\hat{M}_3" + }, + { + "bbox": [ + 54, + 293, + 555, + 329 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 54, + 293, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\hat{M}" + }, + { + "bbox": [ + 54, + 293, + 555, + 329 + ], + "type": "text", + "content": " trained on three datasets with different data distributions, respectively. The source and target domain data are sampled from their respective data distributions. " + }, + { + "bbox": [ + 54, + 293, + 555, + 329 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 54, + 293, + 555, + 329 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 293, + 555, + 329 + ], + "type": "inline_equation", + "content": "G'" + }, + { + "bbox": [ + 54, + 293, + 555, + 329 + ], + "type": "text", + "content": " are two different text-to-image models, and " + }, + { + "bbox": [ + 54, + 293, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_1" + }, + { + "bbox": [ + 54, + 293, + 555, + 329 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 293, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_2" + }, + { + "bbox": [ + 54, + 293, + 555, + 329 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 293, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t" + }, + { + "bbox": [ + 54, + 293, + 555, + 329 + ], + "type": "text", + "content": " are distinct sets of text prompts." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 290, + 71, + 553, + 283 + ], + "blocks": [ + { + "bbox": [ + 290, + 71, + 553, + 283 + ], + "lines": [ + { + "bbox": [ + 290, + 71, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 290, + 71, + 553, + 283 + ], + "type": "image", + "image_path": "715efa933966fef9d79a188274f5d9efaa7454963eb7c8ba1e8fd9f5c261c52b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 338, + 297, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 338, + 297, + 542 + ], + "spans": [ + { + "bbox": [ + 54, + 338, + 297, + 542 + ], + "type": "text", + "content": "Detailed proofs can be found in the supplementary material. Theorem 1 suggests that for two models with distinct (identical) training data sources, their generalization errors differ substantially (minimally). Its contrapositive states that for two models with small (large) discrepancies in generalization error, their training data sources are identical (distinct), which supports our method. Specifically, when " + }, + { + "bbox": [ + 54, + 338, + 297, + 542 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 54, + 338, + 297, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 338, + 297, + 542 + ], + "type": "inline_equation", + "content": "\\hat{M}" + }, + { + "bbox": [ + 54, + 338, + 297, + 542 + ], + "type": "text", + "content": " are trained on synthetic data from the same text-to-image model, their accuracies on the corresponding synthetic data from this text-to-image model are more akin, compared to when their training data is derived from disparate sources. This aligns with the rationale of TrainProVe, as visually depicted in Fig. 2. The discrepancies in accuracies among these models originate from the disparities in their source domains. The closer the data distributions of the source domain data, the more analogous the accuracies of the trained models on the target domain data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 551, + 201, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 201, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 201, + 563 + ], + "type": "text", + "content": "3.3. The Proposed TrainProVe" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 569, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 604 + ], + "type": "text", + "content": "Building upon the above analysis, we propose a method called TrainProVe to protect generative models' intellectual property, as shown in Fig. 3, which consists of three stages:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "text", + "content": "(1) Shadow and Validation Dataset Generation: Utilizing " + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "inline_equation", + "content": "G_{d}" + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "text", + "content": " to produce a shadow dataset " + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{sdw}" + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "text", + "content": " and a validation dataset " + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_v" + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "text", + "content": ", based on various text prompt sets " + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{sdw}" + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_v" + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "text", + "content": ", tailored to the predicted labels set " + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "text", + "content": " of the suspect model " + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "inline_equation", + "content": "G_{d}, \\mathcal{T}_{sdw}" + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_v" + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "text", + "content": " correspond to " + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "inline_equation", + "content": "G, \\mathcal{T}_1," + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_t" + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "text", + "content": " in Fig. 2 respectively, while " + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{sdw}" + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_v" + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "text", + "content": " correspond to the training data of " + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "text", + "content": " and the target domain data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "(2) Shadow Model Training: Using the shadow dataset " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{sdw}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " to train the shadow model " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "M_{sdw}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": ". The shadow model" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 338, + 555, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 422 + ], + "type": "text", + "content": "is analogous to " + }, + { + "bbox": [ + 313, + 338, + 555, + 422 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 338, + 555, + 422 + ], + "type": "text", + "content": " in Fig. 2, whereas the suspect model is denoted as " + }, + { + "bbox": [ + 313, + 338, + 555, + 422 + ], + "type": "inline_equation", + "content": "\\hat{M}" + }, + { + "bbox": [ + 313, + 338, + 555, + 422 + ], + "type": "text", + "content": ". The purpose of the shadow model is to compare its performance with the suspect model trained on distinct data, focusing on their generalization capacities (accuracies) on the validation dataset. Specifically, when both models are trained on data from " + }, + { + "bbox": [ + 313, + 338, + 555, + 422 + ], + "type": "inline_equation", + "content": "G_{d}" + }, + { + "bbox": [ + 313, + 338, + 555, + 422 + ], + "type": "text", + "content": ", their accuracies on the validation dataset tend to be similar." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 424, + 556, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 556, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 556, + 509 + ], + "type": "text", + "content": "(3) Hypothesis Testing: Providing the validation dataset " + }, + { + "bbox": [ + 313, + 424, + 556, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_v" + }, + { + "bbox": [ + 313, + 424, + 556, + 509 + ], + "type": "text", + "content": " to both the shadow model " + }, + { + "bbox": [ + 313, + 424, + 556, + 509 + ], + "type": "inline_equation", + "content": "M_{sdw}" + }, + { + "bbox": [ + 313, + 424, + 556, + 509 + ], + "type": "text", + "content": " and the suspect model " + }, + { + "bbox": [ + 313, + 424, + 556, + 509 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 313, + 424, + 556, + 509 + ], + "type": "text", + "content": " in the same sequence batch by batch. Subsequently, calculating the accuracy of the two models batch by batch, leading to accuracy sets " + }, + { + "bbox": [ + 313, + 424, + 556, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{sdw}" + }, + { + "bbox": [ + 313, + 424, + 556, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 424, + 556, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{sus}" + }, + { + "bbox": [ + 313, + 424, + 556, + 509 + ], + "type": "text", + "content": ". Ultimately, utilizing a one-sided Grubbs' hypothesis test [10] to ascertain whether " + }, + { + "bbox": [ + 313, + 424, + 556, + 509 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 313, + 424, + 556, + 509 + ], + "type": "text", + "content": " violates the intellectual property of " + }, + { + "bbox": [ + 313, + 424, + 556, + 509 + ], + "type": "inline_equation", + "content": "G_d" + }, + { + "bbox": [ + 313, + 424, + 556, + 509 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": "In the one-sided Grubbs' hypothesis test, the null hypothesis " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "H_0" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " asserts that the mean of the set " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\text{sus}}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\text{mean}(\\mathcal{A}_{\\text{sus}})" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": ", is not an unusually low value among the elements of " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\text{sdw}}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": ". On the contrary, the alternative hypothesis " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "H_1" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " posits that the mean of " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\text{sus}}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " is an exceptionally low value within " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\text{sdw}}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": ". Rejecting the null hypothesis " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "H_0" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " implies that the predictive capabilities of " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "M_{\\text{sus}}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " are significantly inferior to those of " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "M_{\\text{sdw}}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": ". This suggests that the model " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "M_{\\text{sus}}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " exhibits notably inferior generalization performance on the validation dataset compared to " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "M_{\\text{sdw}}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": ", indicating that " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "M_{\\text{sus}}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " has not been exposed to the synthetic data from " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "G_d" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " during training and is therefore compliant. Conversely, if the predictions generated by " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "M_{\\text{sus}}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "M_{\\text{sdw}}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " are similar, it indicates that the generalization performance of " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "M_{\\text{sus}}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " on the validation dataset closely resembles that of " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "M_{\\text{sdw}}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": ". This similarity implies that " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "M_{\\text{sus}}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": ", much like " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "M_{\\text{sdw}}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": ", has been trained on the synthetic data from " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "G_d" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": ", thereby sug-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23820" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 89, + 292, + 163 + ], + "blocks": [ + { + "bbox": [ + 63, + 75, + 259, + 86 + ], + "lines": [ + { + "bbox": [ + 63, + 75, + 259, + 86 + ], + "spans": [ + { + "bbox": [ + 63, + 75, + 259, + 86 + ], + "type": "text", + "content": "1. Shadow and Validation Dataset Generation" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 89, + 292, + 163 + ], + "lines": [ + { + "bbox": [ + 63, + 89, + 292, + 163 + ], + "spans": [ + { + "bbox": [ + 63, + 89, + 292, + 163 + ], + "type": "image", + "image_path": "9bbf7b8e16a032488cd675791d46c51d08569ec6cce0820c24ac10f956b3a3df.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 63, + 181, + 293, + 238 + ], + "blocks": [ + { + "bbox": [ + 65, + 168, + 179, + 179 + ], + "lines": [ + { + "bbox": [ + 65, + 168, + 179, + 179 + ], + "spans": [ + { + "bbox": [ + 65, + 168, + 179, + 179 + ], + "type": "text", + "content": "2. Shadow Model Training" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 181, + 293, + 238 + ], + "lines": [ + { + "bbox": [ + 63, + 181, + 293, + 238 + ], + "spans": [ + { + "bbox": [ + 63, + 181, + 293, + 238 + ], + "type": "image", + "image_path": "78b0c6168b7f39c1ea107b3f8824ec8803f7b8696f11c6f70210db933ebdaaa5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 63, + 255, + 293, + 356 + ], + "blocks": [ + { + "bbox": [ + 65, + 242, + 156, + 254 + ], + "lines": [ + { + "bbox": [ + 65, + 242, + 156, + 254 + ], + "spans": [ + { + "bbox": [ + 65, + 242, + 156, + 254 + ], + "type": "text", + "content": "3. Hypothesis Testing" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 255, + 293, + 356 + ], + "lines": [ + { + "bbox": [ + 63, + 255, + 293, + 356 + ], + "spans": [ + { + "bbox": [ + 63, + 255, + 293, + 356 + ], + "type": "image", + "image_path": "c562e8cfec122d9aba9ed01058b55243d0cd06254bc1892e6310cc7076b76002.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 364, + 261, + 375 + ], + "lines": [ + { + "bbox": [ + 89, + 364, + 261, + 375 + ], + "spans": [ + { + "bbox": [ + 89, + 364, + 261, + 375 + ], + "type": "text", + "content": "Figure 3. The complete process of TrainProVe." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 381, + 296, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 381, + 296, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 296, + 406 + ], + "type": "text", + "content": "gesting that " + }, + { + "bbox": [ + 55, + 381, + 296, + 406 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 55, + 381, + 296, + 406 + ], + "type": "text", + "content": " violates the intellectual property rights of " + }, + { + "bbox": [ + 55, + 381, + 296, + 406 + ], + "type": "inline_equation", + "content": "G_{d}" + }, + { + "bbox": [ + 55, + 381, + 296, + 406 + ], + "type": "text", + "content": " and is considered illicit." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 417, + 137, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 417, + 137, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 137, + 430 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 437, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 498 + ], + "type": "text", + "content": "Here, we present the results of DOV4MM on ImageNet [5] and CIFAR [20], ablation experiments, interference resistance analysis, and efficiency analysis. More details and results (hypothesis testing, results on CLIP [35], etc) can be found in the supplementary materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 504, + 172, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 172, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 172, + 517 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 521, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 689 + ], + "type": "text", + "content": "Models and Datasets. We evaluate the proposed TrainProVe with three visual datasets (CIFAR10 [20], CIFAR100 [20], and ImageNet-100 [40]) and four text-to-image models (Stable Diffusion [38] v1.4, latent consistency model [28], PixArt-" + }, + { + "bbox": [ + 55, + 521, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 521, + 295, + 689 + ], + "type": "text", + "content": " [4], and Stable Cascade [34]). ImageNet-100 is a randomly selected subset of ImageNet1K [5], encompassing 100 categories and 126,689 images. Stable Cascade is a latent diffusion model constructed on the Würstchen architecture [34]. In the experiments, unless explicitly stated otherwise, all text-to-image models employed are pre-trained models. To distinguish from real datasets, we denote the synthetic datasets of CIFAR10, CIFAR100, and ImageNet-100 as CIFAR10-Syn, CIFAR100-Syn, and ImageNet-100-Syn, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Experiments on CIFAR: The shadow dataset " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{sdw}" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " contains 30,000 images with a resolution of " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": ", and the text" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "content": "prompt is " + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{sdw} = \\{\\text{“}a\\{c\\}\\text{”}|c\\in \\mathcal{C}\\}" + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "content": ". The validation dataset " + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_v" + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "content": " includes 5,000 images with a resolution of " + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "inline_equation", + "content": "32\\times 32" + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "content": ", with a text prompt of " + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_v = \\{\\text{“}a\\text{ photo of }a\\{c\\}\\text{”}|c\\in \\mathcal{C}\\}" + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "content": ". For the shadow model " + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "inline_equation", + "content": "M_{sdw}" + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "content": ", the architecture is ResNet18 [12], and following previous work [42], we set the batch size for training " + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "inline_equation", + "content": "M_{sdw}" + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "content": " to 512, the learning rate to 1e-4, weight decay to 0.9, and the loss function to cross-entropy loss. To simulate the training data for the suspicious model, we generated 60,000 images with a resolution of " + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "inline_equation", + "content": "32\\times 32" + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "content": " using the method in [42] as CIFAR10-Syn/CIFAR100-Syn. Moreover, the suspicious models have various architectures and training hyperparameters to simulate different scenarios, specifically including: four architectures (ResNet50 [12], ResNet101 [12], ViT-S/16 [7], ViT-B/16 [7]), two batch sizes (512 and 256), two learning rates (1e-4 and 5e-5), two weight decay values (0.9 and 0.09), and two loss functions (cross-entropy loss and focal loss [22]). For each dataset, we can use these parameters to train 64 (" + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "inline_equation", + "content": "C_4^1 C_2^1 C_2^1 C_2^1 C_2^1" + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "content": ") suspicious models. Batch size for inference in TrainProVe is 128, and training epoch for all models is set to 100." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 312, + 553, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 479 + ], + "type": "text", + "content": "Experiments on ImageNet-100: For the shadow dataset " + }, + { + "bbox": [ + 313, + 312, + 553, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{sdw}" + }, + { + "bbox": [ + 313, + 312, + 553, + 479 + ], + "type": "text", + "content": ", validation dataset " + }, + { + "bbox": [ + 313, + 312, + 553, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_v" + }, + { + "bbox": [ + 313, + 312, + 553, + 479 + ], + "type": "text", + "content": ", and shadow model " + }, + { + "bbox": [ + 313, + 312, + 553, + 479 + ], + "type": "inline_equation", + "content": "M_{sdw}" + }, + { + "bbox": [ + 313, + 312, + 553, + 479 + ], + "type": "text", + "content": ", all settings are the same as those used for CIFAR, except that " + }, + { + "bbox": [ + 313, + 312, + 553, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{sdw}" + }, + { + "bbox": [ + 313, + 312, + 553, + 479 + ], + "type": "text", + "content": " contains 50,000 images with a resolution of " + }, + { + "bbox": [ + 313, + 312, + 553, + 479 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 312, + 553, + 479 + ], + "type": "text", + "content": " and the batch size for model training is 64. The training set for the suspicious model " + }, + { + "bbox": [ + 313, + 312, + 553, + 479 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 313, + 312, + 553, + 479 + ], + "type": "text", + "content": " includes 100,000 images generated using the method in [40], with a resolution of " + }, + { + "bbox": [ + 313, + 312, + 553, + 479 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 312, + 553, + 479 + ], + "type": "text", + "content": ". For the parameter selection of the suspicious model, all settings are the same as those used in the experiments on CIFAR, except that the batch size is 64 or 32. Therefore, we can also obtain 64 suspicious models for each dataset (ImageNet-100 or ImageNet-100-Syn). Additionally, batch size for inference in TrainProVe is 16. The training epoch for all models is set to 100." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 480, + 553, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 553, + 564 + ], + "type": "text", + "content": "Evaluation Metrics. Our goal is to classify the suspicious model " + }, + { + "bbox": [ + 313, + 480, + 553, + 564 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 313, + 480, + 553, + 564 + ], + "type": "text", + "content": " as either illegal or legal. Specifically, if the training data for " + }, + { + "bbox": [ + 313, + 480, + 553, + 564 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 313, + 480, + 553, + 564 + ], + "type": "text", + "content": " originates from the defender's generative model " + }, + { + "bbox": [ + 313, + 480, + 553, + 564 + ], + "type": "inline_equation", + "content": "G_d" + }, + { + "bbox": [ + 313, + 480, + 553, + 564 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 313, + 480, + 553, + 564 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 313, + 480, + 553, + 564 + ], + "type": "text", + "content": " should be classified as illegal; otherwise, " + }, + { + "bbox": [ + 313, + 480, + 553, + 564 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 313, + 480, + 553, + 564 + ], + "type": "text", + "content": " is considered legal. Given that this is a classification task, we selected three commonly used metrics for evaluation: accuracy, F1 score, and AUROC." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 574, + 380, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 574, + 380, + 586 + ], + "spans": [ + { + "bbox": [ + 314, + 574, + 380, + 586 + ], + "type": "text", + "content": "4.2. Baselines" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 592, + 553, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 553, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 553, + 652 + ], + "type": "text", + "content": "Since TrainProVe is the first attempt at addressing the introduced training data provenance verification problem, there are no existing methods for direct comparison. Nonetheless, there are several baselines that can be utilized to showcase the efficacy of the proposed method, as enumerated below:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 653, + 553, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 553, + 677 + ], + "type": "text", + "content": "Random classification: a rather trivial baseline that randomly classify suspicious models as illegal or legal." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Han et al.'s work [11]: This is the closest method to ours, which trains an attributor to classify suspicious models. The underlying assumption of this method is that the defender" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23821" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 550, + 282 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 550, + 282 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 550, + 282 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 550, + 282 + ], + "type": "table", + "html": "
The Suspect's Data SourcesTPFPFNTNAccuracyF1 ScoreAUROC
The Defender's Generative Model GdData Sources Unrelated to Gd
Stable Diffusion [38] v1.4Latent Consistency Model064
PixArt-α6400640.9880.9700.992
Stable Cascade064
Real Data460
Latent Consistency Model [28]Stable Diffusion v1.4064
PixArt-α6400641.0001.0001.000
Stable Cascade064
Real Data064
PixArt-α [4]Stable Diffusion v1.4064
Latent Consistency Model6400641.0001.0001.000
Stable Cascade064
Real Data064
Stable Cascade [34]Stable Diffusion v1.4064
Latent Consistency Model6400641.0001.0001.000
PixArt-α064
Real Data064
Average Value0.9970.9920.998
", + "image_path": "ada5311c6b21f180f4d0bc2a3b11191e8de759ebad893b70fefa6c45030d3269.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 287, + 555, + 321 + ], + "lines": [ + { + "bbox": [ + 55, + 287, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 555, + 321 + ], + "type": "text", + "content": "Table 1. The results of TrainProVe on CIFAR10. The \"64\" in the table indicates the 64 suspicious models trained based on a data source, with different training hyperparameters. Suspicious models predicted as illegal/legal are considered positive/negative instances. If the suspicious model's training data source is " + }, + { + "bbox": [ + 55, + 287, + 555, + 321 + ], + "type": "inline_equation", + "content": "G_{d}" + }, + { + "bbox": [ + 55, + 287, + 555, + 321 + ], + "type": "text", + "content": ", it is illegal, otherwise, it is innocent." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 327, + 294, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 327, + 294, + 388 + ], + "spans": [ + { + "bbox": [ + 54, + 327, + 294, + 388 + ], + "type": "text", + "content": "possesses multiple different generative models to train the attributor, while in our scenario, the defender has only a single generative model " + }, + { + "bbox": [ + 54, + 327, + 294, + 388 + ], + "type": "inline_equation", + "content": "G_{d}" + }, + { + "bbox": [ + 54, + 327, + 294, + 388 + ], + "type": "text", + "content": ". To evaluate Han et al.'s method, we relaxed the constraints on the defender by allowing the defender to access all four text-to-image models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 388, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 388, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 54, + 388, + 295, + 518 + ], + "type": "text", + "content": "TrainProVe-Ent: This method performs hypothesis testing by analyzing the entropy of the predicted logits generated by models " + }, + { + "bbox": [ + 54, + 388, + 295, + 518 + ], + "type": "inline_equation", + "content": "M_{sdw}" + }, + { + "bbox": [ + 54, + 388, + 295, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 388, + 295, + 518 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 54, + 388, + 295, + 518 + ], + "type": "text", + "content": ". Specifically, we analyze " + }, + { + "bbox": [ + 54, + 388, + 295, + 518 + ], + "type": "inline_equation", + "content": "D_v" + }, + { + "bbox": [ + 54, + 388, + 295, + 518 + ], + "type": "text", + "content": " in batches using models " + }, + { + "bbox": [ + 54, + 388, + 295, + 518 + ], + "type": "inline_equation", + "content": "M_{sdw}" + }, + { + "bbox": [ + 54, + 388, + 295, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 388, + 295, + 518 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 54, + 388, + 295, + 518 + ], + "type": "text", + "content": " to derive their respective logits, and subsequently compute their entropies " + }, + { + "bbox": [ + 54, + 388, + 295, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{sdw}" + }, + { + "bbox": [ + 54, + 388, + 295, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 388, + 295, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{sus}" + }, + { + "bbox": [ + 54, + 388, + 295, + 518 + ], + "type": "text", + "content": ". Following this, we employ a one-sided Grubbs' hypothesis test to ascertain whether the average of all elements in " + }, + { + "bbox": [ + 54, + 388, + 295, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{sus}" + }, + { + "bbox": [ + 54, + 388, + 295, + 518 + ], + "type": "text", + "content": " is an outlier with a high value compared to " + }, + { + "bbox": [ + 54, + 388, + 295, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{sdw}" + }, + { + "bbox": [ + 54, + 388, + 295, + 518 + ], + "type": "text", + "content": ". This analysis aids in determining the validity of model " + }, + { + "bbox": [ + 54, + 388, + 295, + 518 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 54, + 388, + 295, + 518 + ], + "type": "text", + "content": ", with its parameter configurations being consistent with those of TrainProVe." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 519, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 519, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 54, + 519, + 295, + 651 + ], + "type": "text", + "content": "TrainProVe-Sim: This method conducts hypothesis testing based on the pairwise similarity of logits within the same category from " + }, + { + "bbox": [ + 54, + 519, + 295, + 651 + ], + "type": "inline_equation", + "content": "M_{sdw}" + }, + { + "bbox": [ + 54, + 519, + 295, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 519, + 295, + 651 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 54, + 519, + 295, + 651 + ], + "type": "text", + "content": ". Specifically, " + }, + { + "bbox": [ + 54, + 519, + 295, + 651 + ], + "type": "inline_equation", + "content": "D_v" + }, + { + "bbox": [ + 54, + 519, + 295, + 651 + ], + "type": "text", + "content": " is processed in batches using both " + }, + { + "bbox": [ + 54, + 519, + 295, + 651 + ], + "type": "inline_equation", + "content": "M_{sdw}" + }, + { + "bbox": [ + 54, + 519, + 295, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 519, + 295, + 651 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 54, + 519, + 295, + 651 + ], + "type": "text", + "content": " to obtain their respective logits. Subsequently, the pairwise cosine similarity within the same category is computed to obtain the similarity sets " + }, + { + "bbox": [ + 54, + 519, + 295, + 651 + ], + "type": "inline_equation", + "content": "S_{sdw}" + }, + { + "bbox": [ + 54, + 519, + 295, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 519, + 295, + 651 + ], + "type": "inline_equation", + "content": "S_{sus}" + }, + { + "bbox": [ + 54, + 519, + 295, + 651 + ], + "type": "text", + "content": ". Following this, a one-sided Grubbs' hypothesis test is conducted to determine whether the average of all elements in " + }, + { + "bbox": [ + 54, + 519, + 295, + 651 + ], + "type": "inline_equation", + "content": "S_{sus}" + }, + { + "bbox": [ + 54, + 519, + 295, + 651 + ], + "type": "text", + "content": ", is an outlier of low value in " + }, + { + "bbox": [ + 54, + 519, + 295, + 651 + ], + "type": "inline_equation", + "content": "S_{sdw}" + }, + { + "bbox": [ + 54, + 519, + 295, + 651 + ], + "type": "text", + "content": ", thereby evaluating the validity of " + }, + { + "bbox": [ + 54, + 519, + 295, + 651 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 54, + 519, + 295, + 651 + ], + "type": "text", + "content": ". The parameter configurations align with those in TrainProVe." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 659, + 180, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 180, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 180, + 673 + ], + "type": "text", + "content": "4.3. Experimental Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": "In this paper, " + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "inline_equation", + "content": "G_{d}" + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": " is one of the four text-to-image models (Stable Diffusion v1.4, latent consistency model, PixArt-α, and Stable Cascade), and the training dataset of " + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "inline_equation", + "content": "M_{sus}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 327, + 555, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 447 + ], + "type": "text", + "content": "comes from one of five data sources (Stable Diffusion v1.4, latent consistency model, PixArt-" + }, + { + "bbox": [ + 313, + 327, + 555, + 447 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 327, + 555, + 447 + ], + "type": "text", + "content": ", Stable Cascade and real data). We first calculate the accuracy, F1 score, and AUROC under the four different " + }, + { + "bbox": [ + 313, + 327, + 555, + 447 + ], + "type": "inline_equation", + "content": "G_{d}" + }, + { + "bbox": [ + 313, + 327, + 555, + 447 + ], + "type": "text", + "content": ", and then average the metrics across the four scenarios as the final metrics for comparing different methods. Taking the results of TrainProVe on CIFAR10 as an example (more experimental results can be found in the supplementary material), as shown in Tab. 1. It is evident that TrainProVe can accurately classify suspicious models trained on various data sources." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 448, + 556, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 556, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 556, + 664 + ], + "type": "text", + "content": "The results of various methods on different datasets are illustrated in Tab. 2, showcasing the superiority of TrainProVe. In comparison to the research by Han et al., TrainProVe has three advantages: (1) Han et al.'s approach necessitates the defender to have access to multiple different generative models, while TrainProVe is effective even when the defender only has his own generative model " + }, + { + "bbox": [ + 313, + 448, + 556, + 664 + ], + "type": "inline_equation", + "content": "G_{d}" + }, + { + "bbox": [ + 313, + 448, + 556, + 664 + ], + "type": "text", + "content": "; (2) The efficacy of Han et al.'s method consistently lags behind that of TrainProVe across all cases. This is because the former relies on an attributor for classification, which tends to overfit to " + }, + { + "bbox": [ + 313, + 448, + 556, + 664 + ], + "type": "inline_equation", + "content": "M_{sdw}" + }, + { + "bbox": [ + 313, + 448, + 556, + 664 + ], + "type": "text", + "content": "'s architecture used by the defender. If the architecture of " + }, + { + "bbox": [ + 313, + 448, + 556, + 664 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 313, + 448, + 556, + 664 + ], + "type": "text", + "content": " differs, this method is likely to fail. In contrast, TrainProVe relies on hypothesis testing to classify " + }, + { + "bbox": [ + 313, + 448, + 556, + 664 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 313, + 448, + 556, + 664 + ], + "type": "text", + "content": ", thus avoiding this risk; (3) Han et al.'s method requires training a large number of models based on synthetic datasets from each generative model to train the attributor. In contrast, TrainProVe only requires training a single ResNet18 on one synthetic dataset generated by " + }, + { + "bbox": [ + 313, + 448, + 556, + 664 + ], + "type": "inline_equation", + "content": "G_{d}" + }, + { + "bbox": [ + 313, + 448, + 556, + 664 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Moreover, compared to TrainProVe-Ent and TrainProVe-Sim, the efficacy of TrainProVe consistently surpasses that of the others. This is because it compares the predictions of " + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "inline_equation", + "content": "M_{sdw}" + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": " with the true labels to obtain accuracy," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23822" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 296, + 253 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 296, + 253 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 296, + 253 + ], + "type": "table", + "html": "
DatasetMethodAvg. AccAvg. F1Avg. AUC
CIFAR10Random0.5000.2860.500
Han et al.'s Work0.7680.5790.731
TrainProVe-Sim0.8120.6470.821
TrainProVe-Ent0.8680.7740.902
TrainProVe0.9970.9920.998
CIFAR100Random0.5000.2860.500
Han et al.'s Work0.7300.5480.703
TrainProVe-Sim0.5750.3180.545
TrainProVe-Ent0.3980.4080.624
TrainProVe0.9920.9790.981
ImageNet-100Random0.5000.2860.500
Han et al.'s Work0.6710.4250.645
TrainProVe-Sim0.5550.3050.528
TrainProVe-Ent0.3560.3640.567
TrainProVe0.7860.4950.769
", + "image_path": "3f036a589ca4fe0c4dbcdef5917402d7484249077bb0c64802d02628522526fa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 307, + 293, + 387 + ], + "blocks": [ + { + "bbox": [ + 55, + 255, + 295, + 299 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 295, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 295, + 299 + ], + "type": "text", + "content": "Table 2. The results of TrainProVe and baselines on different datasets. \"Avg. Acc\", \"Avg. F1\", and \"Avg. AUC\" are the average values of accuracy, F1 score, and AUROC under four different " + }, + { + "bbox": [ + 55, + 255, + 295, + 299 + ], + "type": "inline_equation", + "content": "G_{d}" + }, + { + "bbox": [ + 55, + 255, + 295, + 299 + ], + "type": "text", + "content": ", similar to Tab. 1. Bold represents the best results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 307, + 293, + 387 + ], + "lines": [ + { + "bbox": [ + 56, + 307, + 293, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 307, + 293, + 387 + ], + "type": "table", + "html": "
DatasetMsdwAvg. AccAvg. F1Avg. AUC
CIFAR10ResNet18 [12]0.9970.9920.998
ConvNeXt-B [26]0.9950.9890.997
Swin-B [25]0.9970.9920.998
ImageNet-100ResNet180.7860.4950.769
ConvNeXt-B0.7710.4700.756
Swin-B0.7870.4870.766
", + "image_path": "67d256aab1f4fc0c17ab707cde1355de8d1ff8065c5ba9914bef3e8c67d064f0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "text", + "content": "using this as the basis for hypothesis testing. In contrast, TrainProVe-Ent and TrainProVe-Sim don't compare model predictions with true labels, resulting in a lack of information from the true labels, which leads to inferior validation performance compared to TrainProVe." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 495, + 156, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 156, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 156, + 508 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "type": "text", + "content": "The model architecture of " + }, + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "type": "inline_equation", + "content": "M_{sdw}" + }, + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "type": "text", + "content": ". We substituted the standard architecture of " + }, + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "type": "inline_equation", + "content": "M_{sdw}" + }, + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "type": "text", + "content": ", ResNet18, with ConvNeXt-B [26] and Swin-B [25], while leaving all other configurations unaltered. The outcomes are displayed in Tab. 3, demonstrating that TrainProVe remains efficient irrespective of whether the architecture of the shadow model is based on CNNs or transformers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "The training epochs of " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "M_{sdw}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": ". We explored TrainProVe's performance as the " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "M_{sdw}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "'s training epochs (default is 100) increased from 0 to 200 in increments of 20. The findings on CIFAR10, depicted in Fig. 4, suggest that TrainProVe exhibits a commendable resilience to variations in the number of training epochs for " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "M_{sdw}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": ". It is worth noting that an excessive or insufficient number of training epochs can lead to overfitting or underfitting of " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "M_{sdw}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{sdw}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": ", consequently resulting in a decline in TrainProVe's performance." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 69, + 553, + 123 + ], + "blocks": [ + { + "bbox": [ + 317, + 69, + 553, + 123 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 553, + 123 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 553, + 123 + ], + "type": "image", + "image_path": "6689acce4050831022ce1e58c3ac38899680adc6e291df407ef698ca65df3bc6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 126, + 553, + 180 + ], + "blocks": [ + { + "bbox": [ + 317, + 126, + 553, + 180 + ], + "lines": [ + { + "bbox": [ + 317, + 126, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 317, + 126, + 553, + 180 + ], + "type": "image", + "image_path": "60c29107c2a6c838fbf6210cebdfca2c4250deca4f53e067d0ecb805c06186c3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 185, + 552, + 237 + ], + "blocks": [ + { + "bbox": [ + 315, + 185, + 552, + 237 + ], + "lines": [ + { + "bbox": [ + 315, + 185, + 552, + 237 + ], + "spans": [ + { + "bbox": [ + 315, + 185, + 552, + 237 + ], + "type": "image", + "image_path": "d2367a8550fe350ea4dae8c9501bd950d7a038d21d9a2d0cd23289497916b5bc.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 238, + 553, + 270 + ], + "lines": [ + { + "bbox": [ + 313, + 238, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 553, + 270 + ], + "type": "text", + "content": "Figure 4. Changing " + }, + { + "bbox": [ + 313, + 238, + 553, + 270 + ], + "type": "inline_equation", + "content": "M_{sdw}" + }, + { + "bbox": [ + 313, + 238, + 553, + 270 + ], + "type": "text", + "content": "'s training epochs. \"Avg. Acc\", \"Avg. F1\", and \"Avg. AUC\" are the average values of accuracy, F1 score, and AUROC under four different " + }, + { + "bbox": [ + 313, + 238, + 553, + 270 + ], + "type": "inline_equation", + "content": "G_d" + }, + { + "bbox": [ + 313, + 238, + 553, + 270 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 325, + 277, + 543, + 315 + ], + "blocks": [ + { + "bbox": [ + 55, + 388, + 295, + 421 + ], + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 421 + ], + "type": "text", + "content": "Table 3. Changing " + }, + { + "bbox": [ + 55, + 388, + 295, + 421 + ], + "type": "inline_equation", + "content": "M_{sdw}" + }, + { + "bbox": [ + 55, + 388, + 295, + 421 + ], + "type": "text", + "content": "'s architecture. \"Avg. Acc\", \"Avg. F1\", and \"Avg. AUC\" are the average values of accuracy, F1 score, and AUROC under four different " + }, + { + "bbox": [ + 55, + 388, + 295, + 421 + ], + "type": "inline_equation", + "content": "G_d" + }, + { + "bbox": [ + 55, + 388, + 295, + 421 + ], + "type": "text", + "content": ", similar to Tab. 1." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 277, + 543, + 315 + ], + "lines": [ + { + "bbox": [ + 325, + 277, + 543, + 315 + ], + "spans": [ + { + "bbox": [ + 325, + 277, + 543, + 315 + ], + "type": "table", + "html": "
Text PromptsAccuracyF1 ScoreAUROC
Original Prompts0.9880.9700.992
Swapped Prompts0.9880.9700.992
", + "image_path": "bc61decd7e80bb865991b90b29a96903e0ad9f71a8862d011f130c0fcf4e0cae.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 323, + 554, + 356 + ], + "lines": [ + { + "bbox": [ + 313, + 323, + 554, + 356 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 554, + 356 + ], + "type": "text", + "content": "Table 4. Swapping " + }, + { + "bbox": [ + 313, + 323, + 554, + 356 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{sdw}" + }, + { + "bbox": [ + 313, + 323, + 554, + 356 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 323, + 554, + 356 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_v" + }, + { + "bbox": [ + 313, + 323, + 554, + 356 + ], + "type": "text", + "content": ". Note that Stable Diffusion v1.4 is " + }, + { + "bbox": [ + 313, + 323, + 554, + 356 + ], + "type": "inline_equation", + "content": "G_{d}" + }, + { + "bbox": [ + 313, + 323, + 554, + 356 + ], + "type": "text", + "content": ". Latent consistency model, PixArt-" + }, + { + "bbox": [ + 313, + 323, + 554, + 356 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 323, + 554, + 356 + ], + "type": "text", + "content": ", Stable Cascade, and real data represent other data sources." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 358, + 554, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 554, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 554, + 418 + ], + "type": "text", + "content": "The text prompts of " + }, + { + "bbox": [ + 313, + 358, + 554, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{sdw}" + }, + { + "bbox": [ + 313, + 358, + 554, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 358, + 554, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_v" + }, + { + "bbox": [ + 313, + 358, + 554, + 418 + ], + "type": "text", + "content": ". We swapped the text prompts, setting " + }, + { + "bbox": [ + 313, + 358, + 554, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{sdw}' = \\mathcal{T}_v" + }, + { + "bbox": [ + 313, + 358, + 554, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 358, + 554, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_v' = \\mathcal{T}_{sdw}" + }, + { + "bbox": [ + 313, + 358, + 554, + 418 + ], + "type": "text", + "content": ", and used " + }, + { + "bbox": [ + 313, + 358, + 554, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{sdw}'" + }, + { + "bbox": [ + 313, + 358, + 554, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 358, + 554, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_v'" + }, + { + "bbox": [ + 313, + 358, + 554, + 418 + ], + "type": "text", + "content": " to generate " + }, + { + "bbox": [ + 313, + 358, + 554, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{sdw}" + }, + { + "bbox": [ + 313, + 358, + 554, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 358, + 554, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_v" + }, + { + "bbox": [ + 313, + 358, + 554, + 418 + ], + "type": "text", + "content": ", respectively. The results on CIFAR10, shown in Tab. 4, indicate that TrainProVe is robust to the text prompts of " + }, + { + "bbox": [ + 313, + 358, + 554, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{sdw}" + }, + { + "bbox": [ + 313, + 358, + 554, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 358, + 554, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_v" + }, + { + "bbox": [ + 313, + 358, + 554, + 418 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 422, + 555, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 555, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 555, + 530 + ], + "type": "text", + "content": "The sample sizes of " + }, + { + "bbox": [ + 313, + 422, + 555, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{sdw}" + }, + { + "bbox": [ + 313, + 422, + 555, + 530 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 422, + 555, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_v" + }, + { + "bbox": [ + 313, + 422, + 555, + 530 + ], + "type": "text", + "content": ". We generated " + }, + { + "bbox": [ + 313, + 422, + 555, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{sdw}" + }, + { + "bbox": [ + 313, + 422, + 555, + 530 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 422, + 555, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_v" + }, + { + "bbox": [ + 313, + 422, + 555, + 530 + ], + "type": "text", + "content": " with different sample sizes for TrainProVe, keeping other settings unchanged. Note that the default sample sizes for " + }, + { + "bbox": [ + 313, + 422, + 555, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{sdw}" + }, + { + "bbox": [ + 313, + 422, + 555, + 530 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 422, + 555, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_v" + }, + { + "bbox": [ + 313, + 422, + 555, + 530 + ], + "type": "text", + "content": " are 30,000 and 5,000 respectively, when generating CIFAR10. The results on CIFAR10 shown in Fig. 5 demonstrate that TrainProVe has good robustness to the sample sizes. Moreover, when " + }, + { + "bbox": [ + 313, + 422, + 555, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{sdw}" + }, + { + "bbox": [ + 313, + 422, + 555, + 530 + ], + "type": "text", + "content": "'s sample size is too small, the generalization performance of " + }, + { + "bbox": [ + 313, + 422, + 555, + 530 + ], + "type": "inline_equation", + "content": "M_{sdw}" + }, + { + "bbox": [ + 313, + 422, + 555, + 530 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 422, + 555, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_v" + }, + { + "bbox": [ + 313, + 422, + 555, + 530 + ], + "type": "text", + "content": " may degrade, leading to a decline in TrainProVe's performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 536, + 537, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 537, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 537, + 549 + ], + "type": "text", + "content": "4.5. The Interference Resistance of TrainProVe" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 555, + 554, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 554, + 590 + ], + "type": "text", + "content": "In this section, we study that whether TrainProVe remains effective when facing more covert data theft. More experimental details can be found in the supplementary material." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "Multiple training data sources for " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "M_{\\text{sus}}" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": ". We investigated the robustness of TrainProVe when the training dataset of " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "M_{\\text{sus}}" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " has multiple data sources. Specifically, we assume that the training dataset of " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "M_{\\text{sus}}" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " contains both synthetic data from " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "G_d" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " and real data. The results on CIFAR10 are shown in Tab. 5, where the training dataset for " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "M_{\\text{sus}}" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " consists of 120,000 images, with synthetic data and real data each accounting for half. We calculated the accuracy of TrainProVe in classifying " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "M_{\\text{sus}}" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " trained on this mixed dataset. The results indicate that TrainProVe can still accu" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23823" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 172, + 148 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 172, + 148 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 172, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 172, + 148 + ], + "type": "image", + "image_path": "1b5c18500654067f5d943d907dcd43bd092b42f3c4d0d3b40a9845c51dfabf47.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 68, + 291, + 148 + ], + "blocks": [ + { + "bbox": [ + 173, + 68, + 291, + 148 + ], + "lines": [ + { + "bbox": [ + 173, + 68, + 291, + 148 + ], + "spans": [ + { + "bbox": [ + 173, + 68, + 291, + 148 + ], + "type": "image", + "image_path": "f3af5e167b53f5c4c367d3d557967d57d66578abb99784cb231aaddbc3a9ab1d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 53, + 153, + 172, + 233 + ], + "blocks": [ + { + "bbox": [ + 53, + 153, + 172, + 233 + ], + "lines": [ + { + "bbox": [ + 53, + 153, + 172, + 233 + ], + "spans": [ + { + "bbox": [ + 53, + 153, + 172, + 233 + ], + "type": "image", + "image_path": "8eb15b5d9bab2482c6441a65d56889b4c861e9b60b407c044fa302ed847e4ebd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 173, + 153, + 291, + 233 + ], + "blocks": [ + { + "bbox": [ + 173, + 153, + 291, + 233 + ], + "lines": [ + { + "bbox": [ + 173, + 153, + 291, + 233 + ], + "spans": [ + { + "bbox": [ + 173, + 153, + 291, + 233 + ], + "type": "image", + "image_path": "9970fb00da91e6c00d309289c2a107d9fb74c8646c584294d0b8cfd46478d397.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 55, + 240, + 172, + 319 + ], + "blocks": [ + { + "bbox": [ + 55, + 240, + 172, + 319 + ], + "lines": [ + { + "bbox": [ + 55, + 240, + 172, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 172, + 319 + ], + "type": "image", + "image_path": "87135b8a894af98385ddac4a869ff2e93dda05c198959ce56ed95d47f9cdfe85.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 321, + 295, + 354 + ], + "lines": [ + { + "bbox": [ + 55, + 321, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 321, + 295, + 354 + ], + "type": "text", + "content": "Figure 5. Changing the sample sizes of " + }, + { + "bbox": [ + 55, + 321, + 295, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{sdw}" + }, + { + "bbox": [ + 55, + 321, + 295, + 354 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 321, + 295, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_v" + }, + { + "bbox": [ + 55, + 321, + 295, + 354 + ], + "type": "text", + "content": ". \"Avg. Acc\", \"Avg. F1\", and \"Avg. AUC\" are the average values of accuracy, F1 score, and AUROC under four different " + }, + { + "bbox": [ + 55, + 321, + 295, + 354 + ], + "type": "inline_equation", + "content": "G_{d}" + }, + { + "bbox": [ + 55, + 321, + 295, + 354 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 173, + 240, + 291, + 319 + ], + "blocks": [ + { + "bbox": [ + 173, + 240, + 291, + 319 + ], + "lines": [ + { + "bbox": [ + 173, + 240, + 291, + 319 + ], + "spans": [ + { + "bbox": [ + 173, + 240, + 291, + 319 + ], + "type": "image", + "image_path": "3ab8185a87e9ef6857b882fab178a5f164783fd31590823f3b7fd9b00d492afc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 363, + 295, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 295, + 375 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 295, + 375 + ], + "type": "text", + "content": "rately detect illegal behavior in this more difficult situation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 380, + 296, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 296, + 511 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 296, + 511 + ], + "type": "text", + "content": "Fine-tuning " + }, + { + "bbox": [ + 55, + 380, + 296, + 511 + ], + "type": "inline_equation", + "content": "G_{d}" + }, + { + "bbox": [ + 55, + 380, + 296, + 511 + ], + "type": "text", + "content": ". When the suspect has access to a small amount of real data, he may fine-tune the defender's open-source model " + }, + { + "bbox": [ + 55, + 380, + 296, + 511 + ], + "type": "inline_equation", + "content": "G_{d}" + }, + { + "bbox": [ + 55, + 380, + 296, + 511 + ], + "type": "text", + "content": " and then generate data to train " + }, + { + "bbox": [ + 55, + 380, + 296, + 511 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 55, + 380, + 296, + 511 + ], + "type": "text", + "content": ", which can enhance " + }, + { + "bbox": [ + 55, + 380, + 296, + 511 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 55, + 380, + 296, + 511 + ], + "type": "text", + "content": "'s generalization ability on real data. Note that the defender only possesses the original " + }, + { + "bbox": [ + 55, + 380, + 296, + 511 + ], + "type": "inline_equation", + "content": "G_{d}" + }, + { + "bbox": [ + 55, + 380, + 296, + 511 + ], + "type": "text", + "content": " and has no knowledge of fine-tuning. Specifically, we fine-tuned the pre-trained Stable Diffusion v1.4 using 10,000 ImageNet-100 images (100 per class) and LoRA [16] with 10 epochs, and then generated ImageNet-100-Syn to train " + }, + { + "bbox": [ + 55, + 380, + 296, + 511 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 55, + 380, + 296, + 511 + ], + "type": "text", + "content": ". The results are shown in Tab. 6a, which indicate that TrainProVe remains effective in this more arduous scenario." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 516, + 295, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 295, + 624 + ], + "type": "text", + "content": "Fine-tuning " + }, + { + "bbox": [ + 55, + 516, + 295, + 624 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 55, + 516, + 295, + 624 + ], + "type": "text", + "content": ". Similarly, when the suspect has a small amount of real data, he may also fine-tune " + }, + { + "bbox": [ + 55, + 516, + 295, + 624 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 55, + 516, + 295, + 624 + ], + "type": "text", + "content": " with real data, which was trained on synthetic data from " + }, + { + "bbox": [ + 55, + 516, + 295, + 624 + ], + "type": "inline_equation", + "content": "G_{d}" + }, + { + "bbox": [ + 55, + 516, + 295, + 624 + ], + "type": "text", + "content": ", to enhance " + }, + { + "bbox": [ + 55, + 516, + 295, + 624 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 55, + 516, + 295, + 624 + ], + "type": "text", + "content": "'s generalization ability on real data. Specifically, we fine-tuned " + }, + { + "bbox": [ + 55, + 516, + 295, + 624 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 55, + 516, + 295, + 624 + ], + "type": "text", + "content": " using 10,000 ImageNet-100 images (100 per class), where " + }, + { + "bbox": [ + 55, + 516, + 295, + 624 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 55, + 516, + 295, + 624 + ], + "type": "text", + "content": " was pre-trained on synthetic data from Stable Diffusion v1.4, with 10 fine-tuning epochs. The results, shown in Tab. 6b, indicate even in this more challenging scenario, TrainProVe remains effective." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 635, + 167, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 167, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 167, + 647 + ], + "type": "text", + "content": "4.6. Efficiency Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "We analyzed the efficiency of TrainProVe and the work of Han et al. on ImageNet-100. Specifically, following the setup in Sec. 4.1, given " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": " trained on ImageNet-100 or ImageNet-100-Syn, we performed detection using both methods on an NVIDIA GeForce RTX 4090. " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "G_{d}" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": " is Sta" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 315, + 69, + 553, + 129 + ], + "blocks": [ + { + "bbox": [ + 315, + 69, + 553, + 129 + ], + "lines": [ + { + "bbox": [ + 315, + 69, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 315, + 69, + 553, + 129 + ], + "type": "table", + "html": "
M_sus's Training SourceG_dTPFPAcc
CIFAR10-SynStable Diffusion v1.46401
Latent Consistency Model6401
+CIFAR10PixArt-α6401
Stable Cascade6401
", + "image_path": "f0cf2a6124ff3acbe5ecee7377cb8de541f1b8b81c235f22860e2df159f13a8d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 315, + 170, + 432, + 207 + ], + "blocks": [ + { + "bbox": [ + 313, + 130, + 553, + 163 + ], + "lines": [ + { + "bbox": [ + 313, + 130, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 313, + 130, + 553, + 163 + ], + "type": "text", + "content": "Table 5. Results of TrainProVe when the training dataset of " + }, + { + "bbox": [ + 313, + 130, + 553, + 163 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 313, + 130, + 553, + 163 + ], + "type": "text", + "content": " has multiple data sources. The \"64\" in the table indicates the 64 suspicious models with different training hyperparameters." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 170, + 432, + 207 + ], + "lines": [ + { + "bbox": [ + 315, + 170, + 432, + 207 + ], + "spans": [ + { + "bbox": [ + 315, + 170, + 432, + 207 + ], + "type": "table", + "html": "
TPFPAcc
w/o fine-tuning6401.00
w/ fine-tuning6220.97
", + "image_path": "f1eb64584a1684f044566acd0bd17bd7a38e61e35c80922f6d91bde26187bfd4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 438, + 170, + 553, + 208 + ], + "blocks": [ + { + "bbox": [ + 342, + 209, + 405, + 219 + ], + "lines": [ + { + "bbox": [ + 342, + 209, + 405, + 219 + ], + "spans": [ + { + "bbox": [ + 342, + 209, + 405, + 219 + ], + "type": "text", + "content": "(a) Fine-tuning " + }, + { + "bbox": [ + 342, + 209, + 405, + 219 + ], + "type": "inline_equation", + "content": "G_{d}" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 438, + 170, + 553, + 208 + ], + "lines": [ + { + "bbox": [ + 438, + 170, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 438, + 170, + 553, + 208 + ], + "type": "table", + "html": "
TPFPAcc
w/o fine-tuning6401.00
w/ fine-tuning5770.89
", + "image_path": "7d61640c3ffd53daf75ec08765d529296541c520faac3099f5440aa5ce08aa98.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 315, + 250, + 553, + 286 + ], + "blocks": [ + { + "bbox": [ + 460, + 209, + 531, + 219 + ], + "lines": [ + { + "bbox": [ + 460, + 209, + 531, + 219 + ], + "spans": [ + { + "bbox": [ + 460, + 209, + 531, + 219 + ], + "type": "text", + "content": "(b) Fine-tuning " + }, + { + "bbox": [ + 460, + 209, + 531, + 219 + ], + "type": "inline_equation", + "content": "M_{sus}" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 220, + 553, + 243 + ], + "lines": [ + { + "bbox": [ + 313, + 220, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 553, + 243 + ], + "type": "text", + "content": "Table 6. Fine-tuning " + }, + { + "bbox": [ + 313, + 220, + 553, + 243 + ], + "type": "inline_equation", + "content": "G_{d}" + }, + { + "bbox": [ + 313, + 220, + 553, + 243 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 220, + 553, + 243 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 313, + 220, + 553, + 243 + ], + "type": "text", + "content": ". The \"64\" in the table indicates the 64 suspicious models with different training hyperparameters." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 250, + 553, + 286 + ], + "lines": [ + { + "bbox": [ + 315, + 250, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 315, + 250, + 553, + 286 + ], + "type": "table", + "html": "
MethodτgτtτvTotalAccF1AUC
Han et al.'s Work137.7h8.0h1.2h146.9h0.480.280.48
TrainProVe33.6h0.1h1.4h35.1h0.780.630.84
", + "image_path": "9114a6b52beb51222be758a16fd3122044ef7a312441b269e2f803e480199b36.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 288, + 553, + 310 + ], + "lines": [ + { + "bbox": [ + 313, + 288, + 553, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 553, + 310 + ], + "type": "text", + "content": "Table 7. Efficiency analysis. \"Acc\", \"F1\" and \"AUC\" refer to the accuracy, F1 score and AUROC respectively." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "text", + "content": "ble Cascade. The runtime and results is shown in Tab. 7. For TrainProVe, " + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "inline_equation", + "content": "\\tau_{g}" + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "text", + "content": " is the time for generating " + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "inline_equation", + "content": "D_{sdw}" + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "inline_equation", + "content": "D_v" + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "inline_equation", + "content": "\\tau_t" + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "text", + "content": " is the time for training " + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "inline_equation", + "content": "M_{sdw}" + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "inline_equation", + "content": "\\tau_v" + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "text", + "content": " is the time for classifying " + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "text", + "content": " using hypothesis testing. For Han et al.'s method, " + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "inline_equation", + "content": "\\tau_{g}" + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "text", + "content": " is the time for generating the training and validation dataset, " + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "inline_equation", + "content": "\\tau_t" + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "text", + "content": " is the time for training the models and the attributor, and " + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "inline_equation", + "content": "\\tau_v" + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "text", + "content": " is the time for classifying " + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "inline_equation", + "content": "M_{sus}" + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "text", + "content": " using the attributor. From the results, it shows that the high efficiency of TrainProVe lies in its significantly smaller " + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "inline_equation", + "content": "\\tau_{g}" + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "inline_equation", + "content": "\\tau_t" + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "text", + "content": " compared to Han et al.'s method. This is because TrainProVe only requires generating one dataset (" + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "inline_equation", + "content": "D_{sdw}" + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "text", + "content": ") and training one model (" + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "inline_equation", + "content": "M_{sdw}" + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "text", + "content": "). In contrast, Han et al.'s method requires generating " + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "text", + "content": " datasets and training " + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "inline_equation", + "content": "m \\times n" + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "text", + "content": " models, where " + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "text", + "content": " is the number of text-to-image models accessible to the defender and " + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 323, + 555, + 538 + ], + "type": "text", + "content": " is the number of models trained by the defender on each dataset. These models are then used to train the attributor. Moreover, TrainProVe achieved better classification results, further illustrating its superiority." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 549, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 388, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": "Open-source text-to-image models have reduced the cost of obtaining images. We introduce an important yet largely unexplored security issue: whether a suspicious model's training data originates from a specific text-to-image model? To our knowledge, we are the first to propose a solution to this problem, called TrainProVe. The success of TrainProVe in verifying training data provenance has been demonstrated through theory and experiments. Promising future work includes: (1) extending TrainProVe to other image generative models, such as image-to-image models; (2) adapting TrainProVe to protect other types of data (e.g., text); (3) exploring other privacy risks about text-to-image models." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23824" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 161, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 161, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 161, + 85 + ], + "type": "text", + "content": "6. Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "content": "This work was partially supported by the Pioneer R&D Program of Zhejiang (No.2024C01021), and Zhejiang Province High-Level Talents Special Support Program \"Leading Talent of Technological Innovation of Ten-Thousands Talents Program\" (No. 2022R52046)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 115, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 115, + 176 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 115, + 176 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 183, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 183, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 183, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 61, + 183, + 296, + 217 + ], + "type": "text", + "content": "[1] Hritik Bansal and Aditya Grover. Leaving reality to imagination: Robust classification via generated datasets. arXiv preprint arXiv:2302.02503, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 218, + 296, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 218, + 296, + 261 + ], + "spans": [ + { + "bbox": [ + 61, + 218, + 296, + 261 + ], + "type": "text", + "content": "[2] Shai Ben-David, John Blitzer, Koby Crammer, and Fernando Pereira. Analysis of representations for domain adaptation. Advances in neural information processing systems, 19, 2006. 3, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 263, + 295, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 263, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 61, + 263, + 295, + 317 + ], + "type": "text", + "content": "[3] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- " + }, + { + "bbox": [ + 61, + 263, + 295, + 317 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 61, + 263, + 295, + 317 + ], + "type": "text", + "content": " : Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 319, + 296, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 319, + 296, + 385 + ], + "spans": [ + { + "bbox": [ + 61, + 319, + 296, + 385 + ], + "type": "text", + "content": "[4] Junsong Chen, YU Jincheng, GE Chongjian, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- " + }, + { + "bbox": [ + 61, + 319, + 296, + 385 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 61, + 319, + 296, + 385 + ], + "type": "text", + "content": ": Fast training of diffusion transformer for photorealistic text-to-image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 5, 6, 4, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 387, + 296, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 387, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 387, + 296, + 430 + ], + "type": "text", + "content": "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 432, + 296, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 296, + 475 + ], + "type": "text", + "content": "[6] Tim Dockhorn, Robin Rombach, Andreas Blatmann, and Yaoliang Yu. Distilling the knowledge in diffusion models. In CVPR Workshop Generative Models for Computer Vision, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 477, + 296, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 477, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 62, + 477, + 296, + 542 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 544, + 296, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 544, + 296, + 599 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 296, + 599 + ], + "type": "text", + "content": "[8] Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7382-7392, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 601, + 296, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 601, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 62, + 601, + 296, + 654 + ], + "type": "text", + "content": "[9] Abolfazl Farahani, Sahar Voghoei, Khaled Rasheed, and Hamid R Arabnia. A brief review of domain adaptation. Advances in data science and information engineering: proceedings from ICDATA 2020 and IKE 2020, pages 877-894, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 657, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 657, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 657, + 295, + 678 + ], + "type": "text", + "content": "[10] Frank Ephraim Grubbs. Sample criteria for testing outlying observations. University of Michigan, 1949. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "text", + "content": "[11] Ge Han, Ahmed Salem, Zheng Li, Shanqing Guo, Michael Backes, and Yang Zhang. Detection and attribution of models trained on generated data. In ICASSP 2024-2024 IEEE" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4875-4879. IEEE, 2024. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 98, + 554, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 98, + 554, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 98, + 554, + 140 + ], + "type": "text", + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 143, + 555, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 143, + 555, + 186 + ], + "spans": [ + { + "bbox": [ + 316, + 143, + 555, + 186 + ], + "type": "text", + "content": "[13] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is synthetic data from generative models ready for image recognition? arXiv preprint arXiv:2210.07574, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 188, + 554, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 188, + 554, + 209 + ], + "spans": [ + { + "bbox": [ + 316, + 188, + 554, + 209 + ], + "type": "text", + "content": "[14] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 212, + 554, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 212, + 554, + 244 + ], + "spans": [ + { + "bbox": [ + 316, + 212, + 554, + 244 + ], + "type": "text", + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 247, + 554, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 247, + 554, + 300 + ], + "spans": [ + { + "bbox": [ + 316, + 247, + 554, + 300 + ], + "type": "text", + "content": "[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 303, + 554, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 303, + 554, + 357 + ], + "spans": [ + { + "bbox": [ + 316, + 303, + 554, + 357 + ], + "type": "text", + "content": "[17] Yongcheng Jing, Yiding Yang, Xinchao Wang, Mingli Song, and Dacheng Tao. Turning frequency to resolution: Video super-resolution via event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7772-7781, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 359, + 554, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 359, + 554, + 413 + ], + "spans": [ + { + "bbox": [ + 317, + 359, + 554, + 413 + ], + "type": "text", + "content": "[18] Yongcheng Jing, Chongbin Yuan, Li Ju, Yiding Yang, Xinchao Wang, and Dacheng Tao. Deep graph reprogramming. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24345-24354, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 416, + 554, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 554, + 471 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 554, + 471 + ], + "type": "text", + "content": "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 473, + 554, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 473, + 554, + 494 + ], + "spans": [ + { + "bbox": [ + 316, + 473, + 554, + 494 + ], + "type": "text", + "content": "[20] A Krizhevsky. Learning multiple layers of features from tiny images. Master's thesis, University of Tront, 2009. 5, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 496, + 554, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 496, + 554, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 496, + 554, + 540 + ], + "type": "text", + "content": "[21] Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pretrained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems, 36, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 542, + 554, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 542, + 554, + 563 + ], + "spans": [ + { + "bbox": [ + 315, + 542, + 554, + 563 + ], + "type": "text", + "content": "[22] T Lin. Focal loss for dense object detection. arXiv preprint arXiv:1708.02002, 2017. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 566, + 554, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 566, + 554, + 598 + ], + "spans": [ + { + "bbox": [ + 316, + 566, + 554, + 598 + ], + "type": "text", + "content": "[23] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 601, + 554, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 554, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 554, + 632 + ], + "type": "text", + "content": "[24] Yugeng Liu, Zheng Li, Michael Backes, Yun Shen, and Yang Zhang. Watermarking diffusion model. arXiv preprint arXiv:2305.12502, 2023. 1, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 635, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 554, + 689 + ], + "type": "text", + "content": "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 315, + 692, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 692, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 692, + 554, + 713 + ], + "type": "text", + "content": "[26] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23825" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "type": "text", + "content": "2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 163 + ], + "type": "text", + "content": "[27] Ge Luo, Junqiang Huang, Manman Zhang, Zhenxing Qian, Sheng Li, and Xinpeng Zhang. Steal my artworks for fine-tuning? a watermarking framework for detecting art theft mimicry in text-to-image models. arXiv preprint arXiv:2311.13619, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 296, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 296, + 208 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 296, + 208 + ], + "type": "text", + "content": "[28] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 1, 3, 5, 6, 2, 4, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 295, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 295, + 243 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 295, + 243 + ], + "type": "text", + "content": "[29] Kidist Amde Mekonnen, Nicola Dall'Asen, and Paolo Rota. Adv-kd: Adversarial knowledge distillation for faster diffusion sampling. arXiv preprint arXiv:2405.20675, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 245, + 294, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 245, + 294, + 267 + ], + "spans": [ + { + "bbox": [ + 56, + 245, + 294, + 267 + ], + "type": "text", + "content": "[30] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41, 1995. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 269, + 295, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 269, + 295, + 302 + ], + "spans": [ + { + "bbox": [ + 56, + 269, + 295, + 302 + ], + "type": "text", + "content": "[31] Rui Min, Sen Li, Hongyang Chen, and Minhao Cheng. A watermark-conditioned diffusion model for ip protection. arXiv preprint arXiv:2403.10893, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 304, + 295, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 295, + 336 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 295, + 336 + ], + "type": "text", + "content": "[32] Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions on knowledge and data engineering, 22(10):1345-1359, 2009. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 338, + 295, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 295, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 295, + 371 + ], + "type": "text", + "content": "[33] Suraj Patil, William Berman, Robin Rombach, and Patrick von Platen. amused: An open muse reproduction. arXiv preprint arXiv:2401.01808, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 373, + 295, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 373, + 295, + 428 + ], + "spans": [ + { + "bbox": [ + 56, + 373, + 295, + 428 + ], + "type": "text", + "content": "[34] Pablo Pernias, Dominic Rampas, Mats L. Richter, Christopher J. Pal, and Marc Aubreville. Wuerstchen: An efficient architecture for large-scale text-to-image diffusion models. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 5, 6, 4, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 430, + 295, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 430, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 430, + 295, + 495 + ], + "type": "text", + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 5, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 497, + 295, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 497, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 56, + 497, + 295, + 540 + ], + "type": "text", + "content": "[36] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 543, + 295, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 543, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 543, + 295, + 586 + ], + "type": "text", + "content": "[37] Mauro Ribeiro, Katarina Grolinger, and Miriam AM Capretz. Mlaas: Machine learning as a service. In 2015 IEEE 14th international conference on machine learning and applications (ICMLA), pages 896-902. IEEE, 2015. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 589, + 295, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 589, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 56, + 589, + 295, + 643 + ], + "type": "text", + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3, 5, 6, 4, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 646, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 295, + 678 + ], + "type": "text", + "content": "[39] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In International Conference on Learning Representations, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "type": "text", + "content": "[40] Mert Bülent Sariyildüz, Karteek Alahari, Diane Larlus, and Yannis Kalantidis. Fake it till you make it: Learning transferable representations from synthetic imagenet clones. In" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "type": "text", + "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8011-8021, 2023. 1, 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 173 + ], + "type": "text", + "content": "[41] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 175, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 239 + ], + "type": "text", + "content": "[42] Jordan Shipard, Arnold Wiliem, Kien Nguyen Thanh, Wei Xiang, and Clinton Fookes. Diversity is definitely needed: Improving model-agnostic zero-shot classification via stable diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 769–778, 2023. 1, 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 240, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 285 + ], + "type": "text", + "content": "[43] Geng Sun, Tingru Cui, Jianming Yong, Jun Shen, and Shiping Chen. Mlaas: a cloud-based system for delivering adaptive micro learning in mobile mooc learning. IEEE Transactions on Services Computing, 11(2):292-305, 2015. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "text", + "content": "[44] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2117-2126, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "text", + "content": "[45] Lisa Torrey and Jude Shavlik. Transfer learning. In Handbook of research on machine learning applications and trends: algorithms, methods, and techniques, pages 242-264. IGI global, 2010. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 376, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 399 + ], + "type": "text", + "content": "[46] Mei Wang and Weihong Deng. Deep visual domain adaptation: A survey. Neurocomputing, 312:135-153, 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "text", + "content": "[47] Zhenting Wang, Chen Chen, Yuchen Liu, Lingjuan Lyu, Dimitris Metaxas, and Shiqing Ma. How to detect unauthorized data usages in text-to-image diffusion models. arXiv preprint arXiv:2307.03108, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 444, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 487 + ], + "type": "text", + "content": "[48] Zhenting Wang, Chen Chen, Yi Zeng, Lingjuan Lyu, and Shiqing Ma. Where did i come from? origin attribution of ai-generated images. Advances in neural information processing systems, 36, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 489, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 553, + 555 + ], + "type": "text", + "content": "[49] Qizhen Weng, Wencong Xiao, Yinghao Yu, Wei Wang, Cheng Wang, Jian He, Yong Li, Liping Zhang, Wei Lin, and Yu Ding. Mlaas in the wild: Workload analysis and scheduling in large-scale heterogeneousgpu clusters. In 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI 22), pages 945–960, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 555, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 600 + ], + "type": "text", + "content": "[50] Qianlong Xiang, Miao Zhang, Yuzhang Shang, Jianlong Wu, Yan Yan, and Liqiang Nie. Dkdm: Data-free knowledge distillation for diffusion models with any architecture. arXiv preprint arXiv:2409.03550, 2024. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "text", + "content": "[51] Yuechen Xie, Jie Song, Mengqi Xue, Haofei Zhang, Xingen Wang, Bingde Hu, Genlang Chen, and Mingli Song. Dataset ownership verification in contrastive pre-trained models. arXiv preprint arXiv:2502.07276, 2025.3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "type": "text", + "content": "[52] Cheng Xiong, Chuan Qin, Guorui Feng, and Xinpeng Zhang. Flexible and secure watermarking for latent diffusion model. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1668-1676, 2023. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[53] Jianhao Yuan, Francesco Pinto, Adam Davies, and Philip Torr. Not just pretty pictures: Toward interventional data" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23826" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 297 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "type": "text", + "content": "augmentation using text-to-image generators. arXiv preprint arXiv:2212.11237, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 128 + ], + "type": "text", + "content": "[54] Zihan Yuan, Li Li, Zichi Wang, and Xinpeng Zhang. Watermarking for stable diffusion models. IEEE Internet of Things Journal, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 296, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 296, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 296, + 174 + ], + "type": "text", + "content": "[55] Likun Zhang, Hao Wu, Lingcui Zhang, Fengyuan Xu, Jin Cao, Fenghua Li, and Ben Niu. Training data attribution: Was your model secretly trained on data created by mine? arXiv preprint arXiv:2409.15781, 2024. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 296, + 217 + ], + "type": "text", + "content": "[56] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Ngai-Man Cheung, and Min Lin. A recipe for watermarking diffusion models. arXiv preprint arXiv:2303.10137, 2023. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 219, + 296, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 296, + 252 + ], + "type": "text", + "content": "[57] Yongchao Zhou, Hshmat Sahak, and Jimmy Ba. Training on thin air: Improve image classification with generated data. arXiv preprint arXiv:2305.15316, 2023. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 254, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 296, + 297 + ], + "type": "text", + "content": "[58] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2020. 3" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23827" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Training-free Dense-Aligned Diffusion Guidance for Modular Conditional Image Synthesis/237b334f-3be2-4e40-8ac1-466a51400496_content_list.json b/2025/Training-free Dense-Aligned Diffusion Guidance for Modular Conditional Image Synthesis/237b334f-3be2-4e40-8ac1-466a51400496_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..82e910795f5c3f2bd102d253fea6fa0f57764f57 --- /dev/null +++ b/2025/Training-free Dense-Aligned Diffusion Guidance for Modular Conditional Image Synthesis/237b334f-3be2-4e40-8ac1-466a51400496_content_list.json @@ -0,0 +1,1704 @@ +[ + { + "type": "text", + "text": "Training-free Dense-Aligned Diffusion Guidance for Modular Conditional Image Synthesis", + "text_level": 1, + "bbox": [ + 91, + 128, + 906, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zixuan Wang $^{1}$ Duo Peng $^{2}$ Feng Chen $^{3}$ Yuwei Yang $^{4}$ Yinjie Lei $^{1,\\text{图}}$", + "bbox": [ + 196, + 202, + 797, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Sichuan University $^{2}$ Singapore University of Technology and Design", + "bbox": [ + 215, + 220, + 784, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3The University of Adelaide 4Australian National University", + "bbox": [ + 248, + 239, + 748, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zixuan98@stu.scu.edu.cn, duo_peng@mymail.sutd.edu.sg, chenfeng1271@gmail.com", + "bbox": [ + 163, + 258, + 834, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yuwei.yang@anu.edu.au, yinjie@scu.edu.cn", + "bbox": [ + 316, + 277, + 674, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Conditional image synthesis is a crucial task with broad applications, such as artistic creation and virtual reality. However, current generative methods are often task-oriented with a narrow scope, handling a restricted condition with constrained applicability. In this paper, we propose a novel approach that treats conditional image synthesis as the modular combination of diverse fundamental condition units. Specifically, we divide conditions into three primary units: text, layout, and drag. To enable effective control over these conditions, we design a dedicated alignment module for each. For the text condition, we introduce a Dense Concept Alignment (DCA) module, which achieves dense visual-text alignment by drawing on diverse textual concepts. For the layout condition, we propose a Dense Geometry Alignment (DGA) module to enforce comprehensive geometric constraints that preserve the spatial configuration. For the drag condition, we introduce a Dense Motion Alignment (DMA) module to apply multi-level motion regularization, ensuring that each pixel follows its desired trajectory without visual artifacts. By flexibly inserting and combining these alignment modules, our framework enhances the model's adaptability to diverse conditional generation tasks and greatly expands its application range. Extensive experiments demonstrate the superior performance of our framework across a variety of conditions, including textual description, segmentation mask (bounding box), drag manipulation, and their combinations. Code is available at https://github.com/ZixuanWang0525/DADG", + "bbox": [ + 86, + 357, + 483, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 808, + 223, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Conditional image synthesis aims to generate realistic images based on user-provided requirements, playing a pivotal role in various applications, such as artistic creation [37, 52]", + "bbox": [ + 89, + 834, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/86d7c1b6ed59d84f9a5722e33444e2ce4e186fe46095c8ae16b6e7ba45f47b81.jpg", + "image_caption": [], + "image_footnote": [ + "(a) General pipeline of our proposed modular conditional image synthesis. Overall framework of Training-free Dense-Aligned Diffusion Guidance is illustrated in Fig. 1b." + ], + "bbox": [ + 517, + 325, + 903, + 439 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/de84523cf2da69fef3c48ef29666c32ca12caad66d1e385707236392a6cf5d95.jpg", + "image_caption": [], + "image_footnote": [ + "(b) Illustration of our proposed Training-free Dense-Aligned Diffusion Guidance. We design three dense alignment modules to enable an effective accommodation of each specific condition unit.", + "Figure 1. The overview of our modular conditional image synthesis. Driven by modular combination of fundamental condition units, diverse visual content can be synthesized." + ], + "bbox": [ + 514, + 479, + 901, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and virtual reality [18, 21]. Recent advances in diffusion models have demonstrated promising performance in handling diverse conditional signals, including image descriptions [36, 42], segmentation maps [19, 28, 29], bounding", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding Author: Yinjie Lei (yinjie@scu.edu.cn)", + "bbox": [ + 112, + 886, + 419, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13135", + "bbox": [ + 480, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "boxes [3, 50, 55], and drag information [15, 39], among others. While existing approaches are effective in interpreting a particular form of input from users, they struggle to generalize across a wide range of conditions. This is impractical for flexible generation of complex visual scenes in the real world.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose the Modular Conditional Image Synthesis (MCIS) paradigm, which regulates the synthesis process by selectively applying and combining independent fundamental condition units, as demonstrated in Fig. 1 (a). Specifically, we categorize condition units into: (1) Text. This category refers to natural language descriptions, specifying the semantic content of an image; (2) Layout. This category mainly encompasses segmentation maps and bounding boxes, representing the component arrangement and scene structure; and (3) Drag. This category refers to point pairs (original and destination of drags), encoding the local transformations of an image. However, MCIS presents a non-trivial problem, as it requires ensuring that all conditions are distinctly reflected in the visual content without omission.", + "bbox": [ + 91, + 185, + 483, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We sequentially analyze the key challenges involved in aligning each basic condition unit: (1) Concept Mismatching. Text typically conveys detailed information about the properties of individual objects and their interactions. This requires generative models to match fine-grained region-level and word-level concepts, instead of considering the image and its description as a whole for alignment; (2) Geometry Inconsistency. Layout implicitly encodes complex relationships among multiple objects, such as size ratios and relative positions. Ensuring image realism requires establishing a detailed geometric correspondence between synthesized content and spatial configuration; and (3) Motion Disharmony. Drag often provides displacement vectors indicating the visual regions to be moved. To synthesize visually coherent content, it is crucial to effectively maintain the intended movement of each visual element, while preserving both appearance and semantic consistency.", + "bbox": [ + 91, + 414, + 483, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above issues, we propose a dense-aligned diffusion guidance framework for modular conditional image synthesis. As illustrated in Fig. 1 (b), our framework utilizes plug-and-play guidance modules to allow visual content to independently align with each of the condition units. Specifically, we propose the Dense Concept Alignment (DCA) module, which ensures the consistency between visual content and scene description in a coarse-to-fine manner. Beyond scene-level alignment, we establish fine-grained correspondence from attribute and relational perspectives within a disentangled feature space. We further introduce the Dense Geometry Alignment (DGA) module. Building upon the detection information from the synthesized image, it imposes constraints on object-wise locations, as well as on relative sizes and distances between", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "each pair of objects. This process ensures the separation between different foreground instances while maintaining the realism of scene arrangement. Based on dense displacement fields, we design the Dense Motion Alignment (DMA) module. During diffusion sampling, this module guides the variation of visual content between adjacent timesteps by aligning it with the displacement fields. Meanwhile, it uses pixel-wise color and semantic regularization to achieve photorealistic consistency.", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of our framework in MCIS, we conduct extensive experiments on public benchmarks, including COCO [24], DenseDiffusion [19], DrawBench [38], and DragBench [39]. Quantitative and qualitative results demonstrate that our approach significantly improves the adherence to the text, layout, and drag conditions. Moreover, our guidance modules are compatible to simultaneously control over multiple condition units. In addition, our framework can be seamlessly integrated into various diffusion architectures.", + "bbox": [ + 511, + 227, + 903, + 376 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 531, + 378, + 834, + 391 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a paradigm of modular conditional image synthesis, which expands the application scope of diffusion models by the adaptive combination of fundamental condition units.", + "- We design three plug-and-play dense alignment modules, DCA, DGA, and DMA, to independently achieve effective control over concept, geometry, and motion.", + "- Extensive experiments show that our framework significantly enhances controllability across text, layout, drag, and their combinations." + ], + "bbox": [ + 529, + 393, + 903, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 556, + 663, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Conditional image synthesis is a core research field in AIGC, which aims to generate images explicitly guided by input conditions. From the aspect of framework architecture, current studies can be divided into: (1) Training-based and (2) Training-free condition image synthesis. Crucially, we classify diverse conditions into three fundamental units: text, layout, and drag. Training-based and -free frameworks are, thus, sequentially analyzed in a context of fundamental condition units we define.", + "bbox": [ + 511, + 582, + 906, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Training-based Conditional Image Synthesis", + "text_level": 1, + "bbox": [ + 511, + 727, + 888, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Training-based frameworks often train diffusion models from scratch or lightweight auxiliary modules to accommodate user-provided conditions. Text-driven frameworks, e.g., Stable Diffusion family [8, 33, 36] and Pix-Art family [4, 5], optimize diffusion models from scratch to implicitly aware of embeddings from frozen large-language models, such as CLIP [34] and T5 [35]. Layout-driven approaches incorporate parameter-efficient modules to encode segmentation masks, bounding boxes, and depth maps, among others. For instance, GLIGEN [23] uses a learn-", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13136", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "able gated self-attention module to capture grounding information, i.e., regional bounding boxes and its corresponding phrase. ControlNet family [22, 53, 55] can be viewed as a learnable copy of a U-net encoder, which reuses an off-the-shelf diffusion model as a backbone for handling diverse spatial conditioning vectors. The copy encoder is connected with U-net decoder layers in a frozen diffusion model by zero-convolution modules. Instead of a Siamese network of encoder layers, T2I-Adapter [29] learns various low-complexity adapters according to different conditions, where each adapter is only composed of some residual convolution blocks. However, the demands for high-quality data and considerable computational resources remain the bottlenecks for scalability and widespread adoption of the above approaches.", + "bbox": [ + 89, + 90, + 483, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Training-free Conditional Image Synthesis", + "text_level": 1, + "bbox": [ + 89, + 325, + 454, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training-free works focus on modifying the original diffusion sampling process to drive the conditional image synthesis using feedback from the guidance function. For text condition, mainstream guidance frameworks are classifier and CLIP guidance. Specifically, classifier guidance [7] incorporates class information into diffusion models using gradients of a timestep-aware image classifier. CLIP guidance, such as DOODL [43], replaces classification scores with global CLIP scores, which can make generation process accommodate free-form descriptions. Regarding layout condition, current approaches often implicitly modify distributions of attention scores to achieve guidance. The underlying assumption is: spatial distribution of high-response attention aligns perceptually with the locations of objects in synthesized images [3]. Based on this idea, BoxDiff [50] and R&B [49] impose region and boundary constraints to stored cross-attention maps. DenseDiffusion [19] and A&R [32] further leverage semantic affinity in self-attention maps to ensure more accurate object placement. With respect to drag condition, DragDiffusion [39] and DragNoise [25] employ motion supervision and point tracking steps to optimize diffusion semantic features. These steps enable features of original points to move to destination points as effectively as possible. However, aforementioned approaches are often designed to encode a particular condition form, which restricts their generalization across a wide range of conditions.", + "bbox": [ + 89, + 348, + 483, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 768, + 217, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section provides a brief overview of the core mechanisms in the conditional diffusion model framework, including diffusion sampling and diffusion guidance.", + "bbox": [ + 89, + 794, + 482, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion sampling. Diffusion sampling is designed to generate a clean image by progressively removing noise from a random Gaussian noise image $\\mathbf{x}_T$ [14, 51]. This sampling process is accomplished by a learned denoising", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "neural network $\\epsilon_{\\theta}$ which is conditioned on a given prompt $\\mathbf{y}$ . Concretely, at each time step $t$ , a cleaner image $\\mathbf{x}_{t-1}$ can be generated by subtracting the predicted noise $\\hat{\\epsilon}_t = \\epsilon_{\\theta}(\\mathbf{x}_t; t, \\mathbf{y})$ from the current noise image $\\mathbf{x}_t$ . As an example of denoising diffusion implicit model (DDIM) [40], $\\mathbf{x}_{t-1}$ can be produced via:", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {t - 1} = \\sqrt {\\alpha_ {t - 1}} \\hat {\\mathbf {x}} _ {0} + \\sqrt {1 - \\alpha_ {t - 1}} \\hat {\\epsilon} _ {t}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 195, + 906, + 213 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 513, + 219, + 558, + 231 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} _ {0} = \\frac {\\mathbf {x} _ {t} - \\sqrt {1 - \\alpha_ {t}} \\hat {\\boldsymbol {\\epsilon}} _ {t}}{\\sqrt {\\alpha_ {t}}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 629, + 229, + 906, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where scalar $\\alpha_{t}$ governs denoising intensity at each timestep $t$ . From Eq. 2, the approximation of the clean image $\\hat{\\mathbf{x}}_0$ can be predicted by performing one-step diffusion sampling.", + "bbox": [ + 511, + 268, + 906, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion guidance. The purpose of diffusion guidance is to produce a more reasonable denoising direction based on auxiliary information $\\mathbf{y}'$ during inference time [2]. Specifically, at each timestep $t$ , diffusion guidance enables the denoised image $\\mathbf{x}_{t-1}$ follow the auxiliary information $\\mathbf{y}'$ by modifying the predicted noise $\\hat{\\epsilon}_t$ using the gradient of the energy function $g(\\mathbf{x}_t; t, \\mathbf{y}')$ :", + "bbox": [ + 511, + 313, + 905, + 419 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\epsilon} _ {t} = \\hat {\\epsilon} _ {t} + \\sigma_ {t} \\nabla_ {\\mathbf {x} _ {t}} g (\\mathbf {x} _ {t}; t, \\mathbf {y} ^ {\\prime}), \\qquad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 428, + 906, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\sigma_{t}$ is a weighting schedule.", + "bbox": [ + 511, + 455, + 736, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In practice, the function $g$ could be the approximate energy from any differentiable model, such as probabilities from a classifier [7], similarity scores from a vision-language model [34], position penalties from a detection [6] or segmentation model [45], and motion consistencies from a flow estimator [41]. Recent studies [2, 11] have shown that the predicted clean image $\\hat{\\mathbf{x}}_0$ (See Eq. 2) can be straightforwardly used as the input for the energy function $g$ , demonstrating that we do not require to learn an additional timestep-dependent differentiable model. This paper also computes the energy based on a one step approximate of the clean image $\\hat{\\mathbf{x}}_0$ to produce the gradient $\\nabla_{\\mathbf{x}_t}g(\\hat{\\mathbf{x}}_0;\\mathbf{y}')$ .", + "bbox": [ + 511, + 470, + 906, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 511, + 679, + 648, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our framework is designed for modular conditional image synthesis, generating diverse visual scenarios by flexibly combining three fundamental condition units: text, layout, and drag. Accordingly, it contains three diffusion guidance modules specifically designed for each unit. To clarify each module, we organize the methodology section as follows: in Sec. 4.1, we design a Dense Concept Alignment (DCA) module to make synthesized images follow semantic cues embedded in texts; in Sec. 4.2, we present a Dense Geometry Alignment (DGA) to enforce adherence to spatial configurations within layouts; in Sec. 4.3, we propose a Dense Motion Alignment (DMA) to allow diffusion model to be aware of movement and deformation.", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13137", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e431493c89228e9a530ee47ff5fc6a0ed1ef9a8eda0aaea15e4c1e13ed04de88.jpg", + "image_caption": [ + "Figure 2. The overall pipeline of our Dense Concept Alignment (DCA) module. Beyond scene-level vision-language consistency, this module can guide synthesized image in achieving dense alignment of attribute and relation concepts within textual conditions." + ], + "image_footnote": [], + "bbox": [ + 96, + 87, + 480, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Dense Concept Alignment (DCA)", + "text_level": 1, + "bbox": [ + 89, + 364, + 382, + 382 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a textual description $\\mathcal{T}$ as the condition, our goal is to guide the synthesized image to faithfully reflect its semantic cues. Current approaches [43] often regard the description $\\mathcal{T}$ and predicted clean image $\\hat{\\mathbf{x}}_0$ as a whole for alignment. However, they fail to match compositional region and word information, resulting in semantic drift. Thus, we propose Dense Concept Align (DCA). As shown in Fig. 2, we decouple $\\mathcal{T}$ and $\\hat{\\mathbf{x}}_0$ into intra-object attributes and inter-object relations, and then align these concepts to establish a fine-grained vision-language correspondence.", + "bbox": [ + 89, + 387, + 483, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We begin by deriving textual concepts from the description $\\mathcal{T}$ . Given an example like \"A red car in front of a snowy mountain,\" we employ scene graph parsing [46] to decompose it into textual attribute concepts (e.g., red car, snowy mountain) and relation concepts (e.g., car in front of mountain). Each concept is individually embedded using the CLIP text encoder [34]. An attribute concept is defined as an adjective-noun pair, where the adjective describes the property (e.g., red, snowy) and the noun specifies the category (e.g., car, mountain). We construct its embedding $\\mathbf{w}^{\\mathrm{a},\\mathrm{c}}$ by concatenating the individual embeddings of the property and category: $\\mathbf{w}^{\\mathrm{a},\\mathrm{c}} = \\mathbf{w}^{\\mathrm{a}} \\oplus \\mathbf{w}^{\\mathrm{c}}$ . In contrast, a relation concept is treated as a complete phrase and directly encoded to obtain its contextual representation $\\mathbf{w}^{\\mathrm{r}}$ . This process yields a collection of textual concepts, denoted as $\\mathcal{W}$ :", + "bbox": [ + 89, + 539, + 483, + 765 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {W} = \\left\\{\\left\\{\\mathbf {w} _ {i} ^ {\\mathrm {a}, \\mathrm {c}} \\right\\} _ {i = 1} ^ {I}, \\left\\{\\mathbf {w} _ {k} ^ {\\mathrm {r}} \\right\\} _ {k = 1} ^ {K} \\right\\}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 777, + 482, + 796 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $I$ and $K$ denote the number of parsed objects and relations, respectively.", + "bbox": [ + 89, + 809, + 482, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Simultaneously, we utilize visual region parsing to extract visual concepts from image $\\hat{\\mathbf{x}}_0$ . Concretely, we decouple the corresponding visual region of textual concepts using an open-set detection model [27]. The visual attribute", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "concept is represented by the detected bounding box enclosing an individual object, whereas the relation one is defined by the union of two object boxes. Using these bounding boxes, we crop $\\hat{\\mathbf{x}}_0$ to generate a set of isolated visual regions, which are then independently embedded by CLIP vision encoder[34], forming the visual concept collection $\\mathcal{V}$ :", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {V} = \\left\\{\\left\\{\\mathbf {v} _ {i} ^ {\\mathrm {a}, \\mathrm {c}} \\right\\} _ {i = 1} ^ {I}, \\left\\{\\mathbf {v} _ {k} ^ {\\mathrm {r}} \\right\\} _ {k = 1} ^ {K} \\right\\}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 194, + 903, + 213 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{v}^{\\mathrm{a,c}}$ and $\\mathbf{v}^{\\mathrm{r}}$ represent the visual embeddings of attribute and relation concepts, respectively.", + "bbox": [ + 511, + 226, + 903, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After obtaining embeddings of decoupled textual and visual concepts, beyond scene-level, we achieve finer-grained alignment by:", + "bbox": [ + 511, + 257, + 903, + 303 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D C A}} = (1 - \\gamma) \\mathcal {L} _ {\\text {s c e n e}} + \\gamma (\\mathcal {L} _ {\\text {a t t r}} + \\mathcal {L} _ {\\text {r e l}}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 318, + 903, + 334 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 513, + 349, + 558, + 361 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s c e n e}} = \\cos \\left(\\mathbf {w} ^ {\\mathrm {s}}, \\mathbf {v} ^ {\\mathrm {s}}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 364, + 903, + 382 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a t t r}} (\\mathcal {W}, \\mathcal {V}) = \\frac {1}{I} \\sum_ {i = 1} ^ {I} \\cos \\left(\\mathbf {w} _ {i} ^ {\\mathrm {a}, \\mathrm {c}}, \\mathbf {v} _ {i} ^ {\\mathrm {a}, \\mathrm {c}}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 395, + 903, + 436 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e l} (\\mathcal {W}, \\mathcal {V}) = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\cos \\left(\\mathbf {w} _ {k} ^ {\\mathrm {r}}, \\mathbf {v} _ {k} ^ {r}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 450, + 903, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\cos (\\cdot ,\\cdot)$ is the cosine similarity function. $\\mathbf{w}^{\\mathrm{s}}$ and $\\mathbf{v}^{\\mathrm{s}}$ are the global embeddings of description $\\mathcal{T}$ and approximated image $\\hat{\\mathbf{x}}_0$ , respectively. We set $\\gamma = 0.3$ .", + "bbox": [ + 511, + 500, + 903, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Dense Geometry Alignment (DGA)", + "text_level": 1, + "bbox": [ + 511, + 559, + 818, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To guide the synthesized image to resemble the spatial configuration described by the collection of segmentation masks $\\mathcal{M}$ or bounding boxes $\\mathcal{B}$ , we propose Dense Geometry Alignment (DGA) module, as shown in Fig. 3. Current approaches simply ensure that each object appears in its designated region by enhancing corresponding attention scores [19, 32, 49, 50]. Different from them, our module improves the generation plausibility from the perspectives of the object-wise locations, as well as relative size and distance between different objects. For simplicity, we use the segmentation mask collection $\\mathcal{M} = \\{\\mathbf{m}\\}_{n=1}^{N}$ as an example to illustrate this module.", + "bbox": [ + 511, + 582, + 905, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the segmentation mask collection $\\hat{\\mathcal{M}} = \\{\\hat{\\mathbf{m}}\\}_{n=1}^{N}$ predicted from the approximated clean image $\\hat{\\mathbf{x}}_0$ by leveraging an image segmentation model [44], considering an object with the $n$ -th class, we strive to avoid it overlapping with the other objects and ensure the coverage of the target segmentation mask $\\mathbf{m}_n$ . Specifically, we enhance the IoU between the two segmentation masks $\\hat{\\mathbf{m}}_n$ and $\\mathbf{m}_n$ , and minimize the interaction between $\\hat{\\mathbf{m}}_n$ and other irrelevant masks $\\mathcal{M} - \\{\\mathbf{m}_n\\}$ . Hence, the coverage loss function is", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13138", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b7e9dc5b22fb8ef85041214e448406240d682544b3c8213494b44e243807cd27.jpg", + "image_caption": [ + "Figure 3. The overall pipeline of our Dense Geometry Alignment (DGA) module. On the basis of the detection information from the approximated image, this module can densely align geometric features between the predicted and reference layouts, focusing on object-level location, size, and distance." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 486, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "formulated by:", + "bbox": [ + 89, + 344, + 192, + 359 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {c v g} (\\hat {\\mathcal {M}}, \\mathcal {M}) = \\sum_ {n = 1} ^ {N} \\left(1 - \\frac {\\hat {\\mathbf {m}} _ {n} \\cap \\mathbf {m} _ {n}}{\\hat {\\mathbf {m}} _ {n} \\cup \\mathbf {m} _ {n}}\\right) \\\\ + \\sum_ {n _ {1} = 1} ^ {N} \\sum_ {n _ {2} = 1} ^ {N} \\left(\\hat {\\mathbf {m}} _ {n _ {1}} \\cap \\mathbf {m} _ {n _ {2}}\\right), \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 367, + 480, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where for clarity and brevity, when two segmentation masks are concurrently referenced in an equation, we use the subscript $n_1$ and $n_2$ to describe the class of different boxes for distinguishment.", + "bbox": [ + 89, + 463, + 483, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since maintaining realistic proportions between different objects is crucial for generating visually plausible images, we also restrict the pairwise size ratios of segmentation masks by the proposed size loss. On the basis of the areas computed for the predicted and target segmentation masks, the size loss is formulated by:", + "bbox": [ + 89, + 523, + 483, + 614 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s i z e}} = \\sum_ {n _ {1} = 1} ^ {N} \\sum_ {n _ {2} = 1} ^ {N} \\left(\\frac {\\hat {\\mathbf {s}} _ {n _ {1}}}{\\hat {\\mathbf {s}} _ {n _ {2}}} - \\frac {\\mathbf {s} _ {n _ {1}}}{\\mathbf {s} _ {n _ {2}}}\\right) ^ {2}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 623, + 482, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{s}}_{n_1}(\\hat{\\mathbf{s}}_{n_2})$ and $\\mathbf{s}_{n_1}(\\mathbf{s}_{n_2})$ are areas of the segmentation masks in the collections $\\hat{\\mathcal{M}}$ and $\\mathcal{M}$ , respectively.", + "bbox": [ + 89, + 672, + 482, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition, we enforce the alignment of pairwise distances between multiple objects among the approximated image $\\hat{\\mathbf{x}}_0$ and the given guidance information $\\mathcal{M}$ . In particular, after obtaining centroids of each predicted and target object according to the segmentation mask collections $\\hat{\\mathcal{M}}$ and $\\mathcal{M}$ respectively, we compute the pairwise Euclidean distances within each collection of centroids. By minimizing discrepancy in distances between predicted and target centroid collections, we formulate the distance loss function as:", + "bbox": [ + 89, + 704, + 483, + 854 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i s t}} = \\sum_ {n _ {1} = 1} ^ {N} \\sum_ {n _ {1} = 2} ^ {N} \\left(\\| \\hat {\\mathbf {c}} _ {n _ {1}} - \\hat {\\mathbf {c}} _ {n _ {2}} \\| _ {2} - \\| \\mathbf {c} _ {n _ {1}} - \\mathbf {c} _ {n _ {2}} \\| _ {2}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 859, + 483, + 905 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1cc73fe18bdc4e0ba385e060b9fbb42c83752efb6dc28cda9459fb96722c679a.jpg", + "image_caption": [ + "Figure 4. The overall pipeline of our Dense Motion Alignment (DMA) module. Using our densified reference drag flow, this module aims to guide synthesized image in achieving dense alignment of motion information across drag conditions, encompassing pixel displacement, appearance, and semantic aspects." + ], + "image_footnote": [], + "bbox": [ + 514, + 85, + 910, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{c}}_{n_1}(\\hat{\\mathbf{c}}_{n_2})$ and $\\mathbf{c}_{n_1}(\\mathbf{c}_{n_2})$ are centroids of the segmentation masks in the collections $\\hat{\\mathcal{M}}$ and $\\mathcal{M}$ , respectively.", + "bbox": [ + 511, + 329, + 903, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In summary, our final guidance function for DGA can be defined as follows:", + "bbox": [ + 511, + 361, + 903, + 390 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D G A}} = (1 - \\lambda) \\mathcal {L} _ {c v g} + \\lambda \\left(\\mathcal {L} _ {s i z e} + \\mathcal {L} _ {d i s t}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 406, + 903, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we set $\\lambda = 0.25$", + "bbox": [ + 513, + 436, + 671, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Dense Motion Alignment (DMA)", + "text_level": 1, + "bbox": [ + 511, + 464, + 802, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a collection of $J$ drag signals $\\mathcal{P} = \\{(\\mathbf{p}_j^{\\mathrm{o}},\\mathbf{p}_j^{\\mathrm{d}})\\}_{j = 1}^{J}$ , we aim to move visual contents of each original point $\\mathbf{p}_j^{\\mathrm{o}}$ to its destination $\\mathbf{p}_j^{\\mathrm{d}}$ . Recent studies, e.g., DragDiffusion [39] and DragNoise [25], only regulate feature similarity around $\\mathbf{p}_j^{\\mathrm{o}}$ and $\\mathbf{p}_j^{\\mathrm{d}}$ by implicit optimization in one selected diffusion step. However, on one hand, sparse drag signals do not provide sufficient object awareness, causing incomplete displacement in foreground. On the other hand, one-step feature optimization is unstable. To enable continuous dragging of designated pixels while preserving visual coherence during diffusion sampling, we propose Dense Motion Alignment (DMA), as shown in Fig. 4. Specifically, we predict reference dense drag flow, and enforce pixel correspondence between an image pair in terms of displacement, appearance, and semantics building on this flow. We define $\\hat{\\mathbf{x}}_0^{t + 1}$ and $\\hat{\\mathbf{x}}_0^t$ as original and dragging images, respectively, for each timestep $t$ .", + "bbox": [ + 511, + 487, + 906, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We generate the reference dense drag flow for each individual drag signal using an object-specific displacement field. In particular, we employ SAM [20] to obtain a segmentation mask collection $\\{\\mathbf{m}_j^{\\mathrm{sam}}\\}_{j = 1}^{J}$ , where each mask represents an instance area of $\\mathbf{p}_j^{\\mathrm{o}}$ . Because image pixels closer to $\\mathbf{p}_j^{\\mathrm{o}}$ should have larger displacements [48], we use radial basis functions to produce a collection of 2D Gaussian maps $\\{\\mathbf{g}_j\\}_{j = 1}^{J}$ . Each map depicts gradually diminishing drag influence radiating outward from $\\mathbf{p}_j^{\\mathrm{o}}$ . Based on $\\{\\mathbf{m}_j^{\\mathrm{sam}}\\}_{j = 1}^{J}$ and $\\{\\mathbf{g}_j\\}_{j = 1}^{J}$ , our reference flow can be densi", + "bbox": [ + 511, + 747, + 908, + 904 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13139", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "fied by:", + "bbox": [ + 89, + 90, + 143, + 104 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {u} _ {\\hat {\\mathbf {x}} _ {0} ^ {t + 1} \\rightarrow \\hat {\\mathbf {x}} _ {0} ^ {t}, j} = \\left(\\mathbf {p} _ {j} ^ {\\mathrm {d}} - \\mathbf {p} _ {j} ^ {\\mathrm {o}}\\right) * \\mathbf {m} _ {j} ^ {\\text {s a m}} * \\mathbf {g} _ {j}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 116, + 482, + 140 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{u}_{\\hat{\\mathbf{x}}_0^{t + 1}\\rightarrow \\hat{\\mathbf{x}}_0^t,j}$ is the densified drag flow for $j$ -th signal.", + "bbox": [ + 89, + 147, + 482, + 165 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To rectify discrepancies in pixel displacement direction and magnitude, we enable the estimated optical flow between $\\hat{\\mathbf{x}}_0^{t + 1}$ and $\\hat{\\mathbf{x}}_0^t$ to align with its reference flow by minimizing displacement loss:", + "bbox": [ + 89, + 165, + 483, + 224 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i s p}} = \\sum_ {j = 1} ^ {J} \\left|\\left| \\hat {\\mathbf {u}} _ {\\hat {\\mathbf {x}} _ {0} ^ {t + 1} \\rightarrow \\hat {\\mathbf {x}} _ {0} ^ {t}, j} - \\mathbf {u} _ {\\hat {\\mathbf {x}} _ {0} ^ {t + 1} \\rightarrow \\hat {\\mathbf {x}} _ {0} ^ {t}, j} \\right|\\right| _ {1}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 234, + 482, + 279 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{u}}_{\\hat{\\mathbf{x}}_0^{t + 1}\\rightarrow \\hat{\\mathbf{x}}_0^t,j}$ is an optical flow estimated by a RAFT model [41]. In order to ensure visual coherence, we enforce pixel-level appearance consistency by directly minimizing the differences between the dragging and warped original image, while semantic consistency is preserved by DINO v2 embedding [30]:", + "bbox": [ + 89, + 287, + 483, + 383 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p a p p} = \\sum_ {j = 1} ^ {J} \\sqrt {\\left(\\operatorname {w a r p} \\left(\\hat {\\mathbf {x}} _ {0} ^ {t + 1} , \\mathbf {u} _ {\\hat {\\mathbf {x}} _ {0} ^ {t + 1} \\rightarrow \\hat {\\mathbf {x}} _ {0} ^ {t} , j}\\right) - \\hat {\\mathbf {x}} _ {0} ^ {t}\\right) ^ {2}}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 392, + 482, + 436 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p s e m} = \\sum_ {j = 1} ^ {J} \\sqrt {\\left(\\operatorname {w a r p} \\left(\\hat {\\mathbf {f}} _ {0} ^ {t + 1} , \\mathbf {u} _ {\\hat {\\mathbf {x}} _ {0} ^ {t + 1} \\rightarrow \\hat {\\mathbf {x}} _ {0} ^ {t} , j}\\right) - \\hat {\\mathbf {f}} _ {0} ^ {t}\\right) ^ {2}}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 446, + 482, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{f}}_0^{t + 1}$ and $\\hat{\\mathbf{f}}_0^t$ are semantic embedding calculated by DINO v2 model. For a drag signal, our loss of DMA module is defined as:", + "bbox": [ + 89, + 496, + 483, + 544 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D M A}} = (1 - \\eta) \\mathcal {L} _ {d i s p} + \\eta \\left(\\mathcal {L} _ {p a p p} + \\mathcal {L} _ {p s e m}\\right), \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 555, + 482, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where we set $\\eta = 0.02$ . Like DragDiffusion [39] and Drag-Noise [25], we also update the positions of the original points after each denosing step.", + "bbox": [ + 89, + 583, + 483, + 630 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 642, + 223, + 659 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setups", + "text_level": 1, + "bbox": [ + 89, + 666, + 287, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section sequentially describes used datasets and evaluation metrics in our experiments, and implementation details of our approaches.", + "bbox": [ + 89, + 688, + 482, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. The datasets used in our study are: (1) COCO [24]. We curate the 400 longest image descriptions from COCO validation set, each ranging from 15 to 50 words. This selection is made to better demonstrate our strength in understanding long-form description; (2) DenseDiffusion [19]. This benchmark contains 250 samples $(2\\sim 3$ unique objects of each sample), where each sample has a global image description, a segmentation mask, and local descriptions and class labels for each segment; (3) DrawBench [38]. We use the 'position' subset from it. This subset is composed of 20 samples (2 unique objects of each sample),", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where each sample has a description and some bounding boxes of each object (box coordinates and class labels generated by GPT); and (4) DragBench [39]. This benchmark comprises 205 samples, each of which has a reference image, one or more drag signals, a description, and a mask indicating regions to be modified.", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. For text condition, we use Fréchet Inception Distance (FID) [12] and CLIP score [34]. FID quantifies the similarity between the synthesized and real images, while the CLIP score measures the alignment between the synthesized image and the reference description. For text + layout condition, we use IoU and SOA-I [13] (layout prediction is based on YOLOv7-Seg [44]). IoU quantifies the overlap between the predicted layout from the synthesized image and the reference layout. SOA-I measures whether described objects appear in the synthesized image. For text + drag condition, we use Image Fidelity (IF) [17] and Mean Distance (MD) [31]. IF (1-LPIPS score [54]) measures the image quality based on the global similarity between the reference and dragged images. MD assesses how accurately the post-dragging original point align with its corresponding destination point by measuring Euclidean distance between their positions.", + "bbox": [ + 511, + 189, + 906, + 448 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. We use $T = 50$ steps of the DDIM scheduler [40] and $T = 28$ steps of the Flow Match Euler Discrete scheduler [10]. The weighting schedules for our proposed DCA, DGA, and DMA modules are set to 60, 25, and 90, respectively. Our guidance is only applied during the first $70\\%$ timesteps.", + "bbox": [ + 511, + 455, + 908, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Image Synthesis Results with Text", + "text_level": 1, + "bbox": [ + 511, + 582, + 812, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use Stable Diffusion v1.5 [36] and v3.0 [8] as baseline models to evaluate the effectiveness of our DCA module in aligning concepts within descriptions. Tab. 1 summarizes FID [12] and CLIP scores [34] on long ( $>15$ words) descriptions from COCO dataset [24], our DCA module achieves $12.3\\%$ and $2.97\\%$ improvements in FID and CLIP scores, respectively, building upon Stable Diffusion v3.0 (the most recent baseline). This gain arises from our module's ability to achieve vision-language alignment in a coarse-to-fine manner, effectively leveraging semantic cues embedded in complex sentences. Besides, our module can be applicable to Stable Diffusion v1.5, showing its adaptability across different baseline versions. As shown in Fig. 5, our approach can properly place the dog on a dog bed next to a rocking chair (top) and position the person's head on a blue backpack (middle), indicating its effectiveness in aligning the object interaction concept. Our approach can also generate two people (bottom), showcasing its advantage in understanding the object's numeric information.", + "bbox": [ + 511, + 612, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13140", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7e82b9cbfe60a1b6dd3bef7d3f3c9e36ecc55a38d7dae3db68d3902b2b335719.jpg", + "table_caption": [ + "Table 1. Quantitative comparison under long descriptions from COCO dataset [24]. The evaluation metrics used are FID [12] and CLIPScore [34]. We reimplement all baseline approaches. The best results are highlighted." + ], + "table_footnote": [], + "table_body": "
MethodsVenueFID ↓CLIPScore ↑
ParaDiffusion [47]arXiv202312.960.2731
PIXART-α [5]ICLR202413.970.2724
SDXL [33]ICLR20249.790.2820
ELLA [16]arXiv202413.640.2811
PIXART-Σ [4]ECCV202410.730.2793
Stable Diffusion v1.5 [36]CVPR202211.490.2750
+ DCA-10.290.2882
Stable Diffusion v3.0 [8]ICML202410.510.2861
+ DCA-9.220.2946
", + "bbox": [ + 94, + 155, + 480, + 311 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fbd4e3ec69cc6d17e6d5771f405f9d9a6c86ca7b6bf16480019ada5b19969e46.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DescriptionSD v3.0Ours
A room with filtered light coming through vertical blinds on a doorwall, with a sleeping dog lying on a dog bed next to a rocking chair.
A person covered with an orange sleeping bag rests his head on a blue back pack as he sleeps on a park bench.
In the desert, two people are sitting next to a small propeller plane next to a tent with other planes in the background.
", + "bbox": [ + 89, + 325, + 483, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 5. Visualization comparison between our approach (Stable Diffusion v3.0 [8] + DCA) and its baseline under varying descriptions. Our approach demonstrates superior performance in aligning semantic concepts within descriptions.", + "bbox": [ + 89, + 551, + 483, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Image Synthesis Results with Layout and Text", + "text_level": 1, + "bbox": [ + 89, + 635, + 480, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use Stable Diffusion v1.5 [36], DenseDiffusion [19] and A&R [32] as baseline models to evaluate the effectiveness of our DGA module in aligning spatial configurations represented by segmentation masks and bounding boxes; and our DCA module in aligning geometric concepts within descriptions. Tab. 2 summarizes IoU and SOA-I scores on mask-to-image. Our DGA module brings a $27.1\\%$ improvement in IoU and a $7.55\\%$ increase in SOA-I over A&R baseline, owing to the establishment of object-wise geometric consistency. Besides, incorporating our DCA module on top of DGA leads to an additional improvement, achieving an IoU of 50.61 and an SOA-I of 86.33. This improvement is due to the ability of DCA module to understand descriptions of positions and shapes. We can also observe consistent performance gains across both Stable Diffusion v1.5 and DenseDiffusion architectures. As shown in Fig. 6, our", + "bbox": [ + 88, + 657, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8a61584b0ed3cb225ed2f3341acb439feb5272cb9353e6234c4c441efefb0967.jpg", + "table_caption": [ + "Table 2. Quantitative comparison under descriptions and segmentation masks provided by DenseDiffusion [19]. The evaluation metrics are IoU and SOA-I [13]. We generate four random images of each condition combination for performance evaluation. The best results are highlighted." + ], + "table_footnote": [], + "table_body": "
MethodsVenueIoU ↑SOA-I ↑
Real Images-59.1393.07
Composable Diffusion [26]ECCV2022-55.88±0.78
Structure Diffusion [9]ICLR2023-70.97±1.08
SD-Pww [1]arXiv202223.76±0.5073.92±1.84
Stable Diffusion [36]CVPR2022-73.08±1.54
+ DGA-25.46±1.0274.45±1.46
+ DGA + DCA-27.28±0.9777.02±1.67
Dense Diffusion [19]ICCV202334.99±1.1377.61±1.75
+ DGA-36.72±1.2478.21±1.62
+ DGA + DCA-38.06±1.2678.42±1.17
A&R [32]CVPR202438.97±0.5678.80±1.27
+ DGA-49.52±0.7084.75±2.46
+ DGA + DCA-50.61±0.4786.33±0.88
", + "bbox": [ + 517, + 170, + 906, + 371 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/65af9171ba18b6fa4263469707ca7b188301a7d83a7ddcca921ccfb6ad8e820c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DescriptionLayoutA&ROurs
An orange tabby cat lying in a bathroom sink.
An orange cat laying on top of a black car.
A dog sitting on a bed wearing an orange bowtie.
", + "bbox": [ + 516, + 383, + 906, + 601 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 6. Visualization comparison between our approach (A&R [32] + DGA + DCA) and its baseline, under varying descriptions and segmentation masks. Our approach can adhere to both geometric concepts and positional configuration.", + "bbox": [ + 511, + 611, + 908, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "approach can accurately place a cat inside a bathroom sink (top), position a cat on top of a black car (middle), generate a bowtie around a dog's neck (bottom), demonstrating the advantage of our framework in awareness of spatial relationships. The same conclusion can be observed in Fig. 7, which presents box-to-image visualizations of validation data from the DrawBench benchmark [38].", + "bbox": [ + 511, + 696, + 908, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Image Synthesis Results with Drag and Text", + "text_level": 1, + "bbox": [ + 511, + 816, + 888, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use DragDiffusion [39] as a baseline model to evaluate the effectiveness of our DMA module in motion alignment; and our DCA module for semantic preservation. Fig. 8 presents MD and IF scores on the DragBench benchmark", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13141", + "bbox": [ + 480, + 944, + 517, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/74fdbdee1a5adb7d2c2d8d258d4052db4675ac9ddd2ad10444f8308cb16385b4.jpg", + "image_caption": [ + "Figure 7. Visualization comparison between our approach (A&R [32] + DGA + DCA) and its baseline, under varying descriptions and bounding boxes. Our approach can satisfy both conditions." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 480, + 303 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/76a6c76452ae6710e101e547b4acc7db30821ad773e2c86f0ccdf1e123de64e4.jpg", + "image_caption": [ + "Figure 8. Quantitative comparison under descriptions and drag signals from DragBench benchmark [39]. The evaluation metrics used are Mean Distance $(\\downarrow)$ and Image Fidelity (1-LPIPS) $(\\uparrow)$ . Approaches with the best results appear in the upper-left corner." + ], + "image_footnote": [], + "bbox": [ + 91, + 371, + 478, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "[39]. The combination of our DMA and DCA modules achieves better MD and IF scores, indicating that our framework ensures each pixel follows its desired trajectory with fewer visual artifacts. However, applying our DMA module alone results in a slight drop in IF, compared to DragDiffusion. This is because our motion field may cause oversmoothing in detail. Fortunately, our DCA module can mitigate such an issue to some degree, by maintaining the overall semantic structure unchanged during the diffusion sampling process. As shown in Fig. 9, our framework can effectively synthesize head rotation (top), rigid car displacement (middle), and large-scale deformation of the bus (bottom). This demonstrates that our algorithm enables the desired movement even when handling complex drag signals", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9e5fa1bdbb06071f4fbf2fe3b9301c50d9bb040c619eff466ca74372d97b98b8.jpg", + "image_caption": [ + "Figure 9. Visualization comparison between our approach (DragDiffusion [39] + DMA) and its baseline under varying drag signals. Our approach enables effective synthesis of large-scale movement and deformation." + ], + "image_footnote": [], + "bbox": [ + 516, + 85, + 903, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(multiple drag points).", + "bbox": [ + 513, + 426, + 663, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 454, + 633, + 469 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a plug-and-play modular design for guiding image synthesis by aligning diverse fundamental condition units (text, layout, drag) in combination. Based on this idea, we devise three dense alignment approaches to independently achieve flexible control over concept, geometry and motion. Comprehensive experiments demonstrate the effectiveness of our framework across diverse conditioning scenarios. However, our framework relies on the capability of its foundation model, e.g., SD model. As shown in Fig. 10, counterfactual scenes is hard to generate, e.g., a river flowing upwards, if SD model is unable to synthesis this scene itself. Future work could explore world knowledge-aware generative paradigms to address this challenge.", + "bbox": [ + 511, + 479, + 906, + 676 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3f6afa46f4f503cf30cbdde5146486933b25a48ce7b471e2ada878e9bb03a3fb.jpg", + "image_caption": [ + "Figure 10. Failure cases. As foundation model lacks world knowledge, our approach (Stable Diffusion v3.0 [8] + DCA) fails to generate counterfactual scenes." + ], + "image_footnote": [], + "bbox": [ + 517, + 686, + 903, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement: This work was supported by the National Natural Science Foundation of China (No. U23B2013, 62276176). This work was also partly supported by the SICHUAN Provincial Natural Science Foundation (No. 2024NSFJQ0023).", + "bbox": [ + 511, + 824, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13142", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 7", + "[2] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 843-852, 2023. 3", + "[3] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 2, 3", + "[4] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 2, 7", + "[5] Junsong Chen, YU Jincheng, GE Chongjian, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In Proceedings of the International Conference on Learning Representations (ICLR), pages 1–30, 2024. 2, 7", + "[6] Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xinggang Wang, and Ying Shan. Yolo-world: Real-time open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16901-16911, 2024. 3", + "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems (NeruIPS), 34:8780-8794, 2021. 3", + "[8] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning (ICML), 2024. 2, 6, 7, 8", + "[9] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. In Proceedings of the International Conference on Learning Representations (ICLR), pages 1-13, 2023. 7", + "[10] Itai Gat, Tal Remez, Neta Shaul, Felix Kreuk, Ricky TQ Chen, Gabriel Synnaeve, Yossi Adi, and Yaron Lipman. Discrete flow matching. Advances in Neural Information Processing Systems (NeurIPS), 37:133345-133385, 2024. 6", + "[11] Daniel Geng and Andrew Owens. Motion guidance: Diffusion-based image editing with differentiable motion estimators. In The Twelfth International Conference on Learning Representations (ICLR), pages 1-21. 3", + "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a" + ], + "bbox": [ + 93, + 114, + 483, + 890 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 6, 7", + "[13] Tobias Hinz, Stefan Heinrich, and Stefan Wermter. Semantic object accuracy for generative text-to-image synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(3):1552-1565, 2020. 6, 7", + "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NeurIPS), 33:6840-6851, 2020. 3", + "[15] Xingzhong Hou, Boxiao Liu, Yi Zhang, Jihao Liu, Yu Liu, and Haihang You. Easydrag: Efficient point-based manipulation on diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8404-8413, 2024. 2", + "[16] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135, 2024. 7", + "[17] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6007-6017, 2023. 6", + "[18] Seung Wook Kim, Bradley Brown, Kangxue Yin, Karsten Kreis, Katja Schwarz, Daiqing Li, Robin Rombach, Antonio Torralba, and Sanja Fidler. Neuralfield-ldm: Scene generation with hierarchical latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8496-8506, 2023. 1", + "[19] Yunji Kim, Jiyoung Lee, Jin-Hwa Kim, Jung-Woo Ha, and Jun-Yan Zhu. Dense text-to-image generation with attention modulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7701-7711, 2023. 1, 2, 3, 4, 6, 7", + "[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4015-4026, 2023. 5", + "[21] Yinjie Lei, Zixuan Wang, Feng Chen, Guoqing Wang, Peng Wang, and Yang Yang. Recent advances in multi-modal 3d scene understanding: A comprehensive survey and evaluation. arXiv preprint arXiv:2310.15676, 2023. 1", + "[22] Ming Li, Taojiannan Yang, Huafeng Kuang, Jie Wu, Zhaoning Wang, Xuefeng Xiao, and Chen Chen. Controlnet++: Improving conditional controls with efficient consistency feedback. In European Conference on Computer Vision (ECCV), pages 129-147. Springer, 2025. 3", + "[23] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22511-22521, 2023. 2" + ], + "bbox": [ + 516, + 92, + 903, + 878 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13143", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision (ECCV), pages 740-755. Springer, 2014. 2, 6, 7", + "[25] Haofeng Liu, Chenshu Xu, Yifei Yang, Lihua Zeng, and Shengfeng He. Drag your noise: Interactive point-based editing via diffusion semantic propagation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6743-6752, 2024. 3, 5, 6", + "[26] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In Proceedings of the European Conference on Computer Vision (ECCV), pages 423-439. Springer, 2022. 7", + "[27] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision (ECCV), pages 38-55. Springer, 2024. 4", + "[28] Sicheng Mo, Fangzhou Mu, Kuan Heng Lin, Yanli Liu, Bochen Guan, Yin Li, and Bolei Zhou. Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7465-7475, 2024. 1", + "[29] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 4296-4304, 2024. 1, 3", + "[30] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6", + "[31] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 6", + "[32] Quynh Phung, Songwei Ge, and Jia-Bin Huang. Grounded text-to-image synthesis with attention refocusing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7932-7942, 2024. 3, 4, 7, 8", + "[33] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In Proceedings of the International Conference on Learning Representations (ICLR), pages 1-13, 2024. 2, 7", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PMLR, 2021. 2, 3, 4, 6, 7" + ], + "bbox": [ + 91, + 90, + 482, + 906 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research (JMLR), 21(140):1-67, 2020. 2", + "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 6, 7", + "[37] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6527-6536, 2024. 1", + "[38] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems (NeurIPS), 35:36479-36494, 2022. 2, 6, 7", + "[39] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8839-8849, 2024. 2, 3, 5, 6, 7, 8", + "[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations (ICLR), pages 1-22, 2021. 3, 6", + "[41] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European Conference on Computer Vision (ECCV), pages 402-419. Springer, 2020. 3, 6", + "[42] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1921-1930, 2023. 1", + "[43] Bram Wallace, Akash Gokul, Stefano Ermon, and Nikhil Naik. End-to-end diffusion latent optimization improves classifier guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR), pages 7280-7290, 2023. 3, 4", + "[44] Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7464-7475, 2023. 4, 6", + "[45] Yang Wang, Zhikui Ouyang, Runhua Han, Zhijian Yin, and Zhen Yang. Yolomask: Real-time instance segmentation with integrating yolov5 and orienmask. In IEEE International Conference on Communication Technology (ICCT), pages 1646-1650. IEEE, 2022. 3" + ], + "bbox": [ + 516, + 90, + 906, + 893 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13144", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Hao Wu, Jiayuan Mao, Yufeng Zhang, Yuning Jiang, Lei Li, Weiwei Sun, and Wei-Ying Ma. Unified visual-semantic embeddings: Bridging vision and language with structured meaning representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6609–6618, 2019. 4", + "[47] Weijia Wu, Zhuang Li, Yefei He, Mike Zheng Shou, Chunhua Shen, Lele Cheng, Yan Li, Tingting Gao, Di Zhang, and Zhongyuan Wang. Paragraph-to-image generation with information-enriched diffusion model. arXiv preprint arXiv:2311.14284, 2023. 7", + "[48] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control for anything using entity representation. In European Conference on Computer Vision (ECCV), pages 331-348. Springer, 2025. 5", + "[49] Jiayu Xiao, Henglei Lv, Liang Li, Shuhui Wang, and Qingming Huang. R&b: Region and boundary aware zero-shot grounded text-to-image generation. arXiv preprint arXiv:2310.08872, 2023. 3, 4", + "[50] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7452-7461, 2023. 2, 3, 4" + ], + "bbox": [ + 89, + 92, + 482, + 450 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and Jian Zhang. Freedom: Training-free energy-guided conditional diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 23174-23184, 2023. 3", + "[52] Yu Zeng, Vishal M Patel, Haochen Wang, Xun Huang, Ting-Chun Wang, Ming-Yu Liu, and Yogesh Balaji. Jedi: Joint-image diffusion models for finetuning-free personalized text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6786-6795, 2024. 1", + "[53] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR), pages 3836-3847, 2023. 3", + "[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 6", + "[55] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. Advances in Neural Information Processing Systems (NeurIPS), 36, 2024. 2, 3" + ], + "bbox": [ + 514, + 92, + 903, + 443 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13145", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Training-free Dense-Aligned Diffusion Guidance for Modular Conditional Image Synthesis/237b334f-3be2-4e40-8ac1-466a51400496_model.json b/2025/Training-free Dense-Aligned Diffusion Guidance for Modular Conditional Image Synthesis/237b334f-3be2-4e40-8ac1-466a51400496_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b95512fd76b3fea4109ac6c6196af650113fc0f0 --- /dev/null +++ b/2025/Training-free Dense-Aligned Diffusion Guidance for Modular Conditional Image Synthesis/237b334f-3be2-4e40-8ac1-466a51400496_model.json @@ -0,0 +1,2356 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.13, + 0.908, + 0.178 + ], + "angle": 0, + "content": "Training-free Dense-Aligned Diffusion Guidance for Modular Conditional Image Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.203, + 0.798, + 0.222 + ], + "angle": 0, + "content": "Zixuan Wang\\(^{1}\\) Duo Peng\\(^{2}\\) Feng Chen\\(^{3}\\) Yuwei Yang\\(^{4}\\) Yinjie Lei\\(^{1,\\text{图}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.222, + 0.785, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\)Sichuan University \\(^{2}\\)Singapore University of Technology and Design" + }, + { + "type": "text", + "bbox": [ + 0.249, + 0.24, + 0.749, + 0.258 + ], + "angle": 0, + "content": "3The University of Adelaide 4Australian National University" + }, + { + "type": "text", + "bbox": [ + 0.164, + 0.26, + 0.835, + 0.275 + ], + "angle": 0, + "content": "zixuan98@stu.scu.edu.cn, duo_peng@mymail.sutd.edu.sg, chenfeng1271@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.278, + 0.676, + 0.293 + ], + "angle": 0, + "content": "yuwei.yang@anu.edu.au, yinjie@scu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.358, + 0.485, + 0.783 + ], + "angle": 0, + "content": "Conditional image synthesis is a crucial task with broad applications, such as artistic creation and virtual reality. However, current generative methods are often task-oriented with a narrow scope, handling a restricted condition with constrained applicability. In this paper, we propose a novel approach that treats conditional image synthesis as the modular combination of diverse fundamental condition units. Specifically, we divide conditions into three primary units: text, layout, and drag. To enable effective control over these conditions, we design a dedicated alignment module for each. For the text condition, we introduce a Dense Concept Alignment (DCA) module, which achieves dense visual-text alignment by drawing on diverse textual concepts. For the layout condition, we propose a Dense Geometry Alignment (DGA) module to enforce comprehensive geometric constraints that preserve the spatial configuration. For the drag condition, we introduce a Dense Motion Alignment (DMA) module to apply multi-level motion regularization, ensuring that each pixel follows its desired trajectory without visual artifacts. By flexibly inserting and combining these alignment modules, our framework enhances the model's adaptability to diverse conditional generation tasks and greatly expands its application range. Extensive experiments demonstrate the superior performance of our framework across a variety of conditions, including textual description, segmentation mask (bounding box), drag manipulation, and their combinations. Code is available at https://github.com/ZixuanWang0525/DADG" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.809, + 0.224, + 0.825 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.835, + 0.484, + 0.88 + ], + "angle": 0, + "content": "Conditional image synthesis aims to generate realistic images based on user-provided requirements, playing a pivotal role in various applications, such as artistic creation [37, 52]" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.327, + 0.905, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.516, + 0.442, + 0.904, + 0.478 + ], + "angle": 0, + "content": "(a) General pipeline of our proposed modular conditional image synthesis. Overall framework of Training-free Dense-Aligned Diffusion Guidance is illustrated in Fig. 1b." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.48, + 0.902, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.513, + 0.758, + 0.902, + 0.794 + ], + "angle": 0, + "content": "(b) Illustration of our proposed Training-free Dense-Aligned Diffusion Guidance. We design three dense alignment modules to enable an effective accommodation of each specific condition unit." + }, + { + "type": "image_footnote", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.836 + ], + "angle": 0, + "content": "Figure 1. The overview of our modular conditional image synthesis. Driven by modular combination of fundamental condition units, diverse visual content can be synthesized." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.758, + 0.906, + 0.836 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "and virtual reality [18, 21]. Recent advances in diffusion models have demonstrated promising performance in handling diverse conditional signals, including image descriptions [36, 42], segmentation maps [19, 28, 29], bounding" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.887, + 0.42, + 0.901 + ], + "angle": 0, + "content": "Corresponding Author: Yinjie Lei (yinjie@scu.edu.cn)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "13135" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "boxes [3, 50, 55], and drag information [15, 39], among others. While existing approaches are effective in interpreting a particular form of input from users, they struggle to generalize across a wide range of conditions. This is impractical for flexible generation of complex visual scenes in the real world." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.186, + 0.485, + 0.412 + ], + "angle": 0, + "content": "In this paper, we propose the Modular Conditional Image Synthesis (MCIS) paradigm, which regulates the synthesis process by selectively applying and combining independent fundamental condition units, as demonstrated in Fig. 1 (a). Specifically, we categorize condition units into: (1) Text. This category refers to natural language descriptions, specifying the semantic content of an image; (2) Layout. This category mainly encompasses segmentation maps and bounding boxes, representing the component arrangement and scene structure; and (3) Drag. This category refers to point pairs (original and destination of drags), encoding the local transformations of an image. However, MCIS presents a non-trivial problem, as it requires ensuring that all conditions are distinctly reflected in the visual content without omission." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.415, + 0.484, + 0.672 + ], + "angle": 0, + "content": "We sequentially analyze the key challenges involved in aligning each basic condition unit: (1) Concept Mismatching. Text typically conveys detailed information about the properties of individual objects and their interactions. This requires generative models to match fine-grained region-level and word-level concepts, instead of considering the image and its description as a whole for alignment; (2) Geometry Inconsistency. Layout implicitly encodes complex relationships among multiple objects, such as size ratios and relative positions. Ensuring image realism requires establishing a detailed geometric correspondence between synthesized content and spatial configuration; and (3) Motion Disharmony. Drag often provides displacement vectors indicating the visual regions to be moved. To synthesize visually coherent content, it is crucial to effectively maintain the intended movement of each visual element, while preserving both appearance and semantic consistency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To address the above issues, we propose a dense-aligned diffusion guidance framework for modular conditional image synthesis. As illustrated in Fig. 1 (b), our framework utilizes plug-and-play guidance modules to allow visual content to independently align with each of the condition units. Specifically, we propose the Dense Concept Alignment (DCA) module, which ensures the consistency between visual content and scene description in a coarse-to-fine manner. Beyond scene-level alignment, we establish fine-grained correspondence from attribute and relational perspectives within a disentangled feature space. We further introduce the Dense Geometry Alignment (DGA) module. Building upon the detection information from the synthesized image, it imposes constraints on object-wise locations, as well as on relative sizes and distances between" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "each pair of objects. This process ensures the separation between different foreground instances while maintaining the realism of scene arrangement. Based on dense displacement fields, we design the Dense Motion Alignment (DMA) module. During diffusion sampling, this module guides the variation of visual content between adjacent timesteps by aligning it with the displacement fields. Meanwhile, it uses pixel-wise color and semantic regularization to achieve photorealistic consistency." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.228, + 0.905, + 0.377 + ], + "angle": 0, + "content": "To evaluate the effectiveness of our framework in MCIS, we conduct extensive experiments on public benchmarks, including COCO [24], DenseDiffusion [19], DrawBench [38], and DragBench [39]. Quantitative and qualitative results demonstrate that our approach significantly improves the adherence to the text, layout, and drag conditions. Moreover, our guidance modules are compatible to simultaneously control over multiple condition units. In addition, our framework can be seamlessly integrated into various diffusion architectures." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.379, + 0.836, + 0.392 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.394, + 0.904, + 0.451 + ], + "angle": 0, + "content": "- We propose a paradigm of modular conditional image synthesis, which expands the application scope of diffusion models by the adaptive combination of fundamental condition units." + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.455, + 0.904, + 0.498 + ], + "angle": 0, + "content": "- We design three plug-and-play dense alignment modules, DCA, DGA, and DMA, to independently achieve effective control over concept, geometry, and motion." + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.5, + 0.903, + 0.543 + ], + "angle": 0, + "content": "- Extensive experiments show that our framework significantly enhances controllability across text, layout, drag, and their combinations." + }, + { + "type": "list", + "bbox": [ + 0.531, + 0.394, + 0.904, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.558, + 0.664, + 0.573 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.907, + 0.718 + ], + "angle": 0, + "content": "Conditional image synthesis is a core research field in AIGC, which aims to generate images explicitly guided by input conditions. From the aspect of framework architecture, current studies can be divided into: (1) Training-based and (2) Training-free condition image synthesis. Crucially, we classify diverse conditions into three fundamental units: text, layout, and drag. Training-based and -free frameworks are, thus, sequentially analyzed in a context of fundamental condition units we define." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.728, + 0.89, + 0.744 + ], + "angle": 0, + "content": "2.1. Training-based Conditional Image Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Training-based frameworks often train diffusion models from scratch or lightweight auxiliary modules to accommodate user-provided conditions. Text-driven frameworks, e.g., Stable Diffusion family [8, 33, 36] and Pix-Art family [4, 5], optimize diffusion models from scratch to implicitly aware of embeddings from frozen large-language models, such as CLIP [34] and T5 [35]. Layout-driven approaches incorporate parameter-efficient modules to encode segmentation masks, bounding boxes, and depth maps, among others. For instance, GLIGEN [23] uses a learn-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13136" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.318 + ], + "angle": 0, + "content": "able gated self-attention module to capture grounding information, i.e., regional bounding boxes and its corresponding phrase. ControlNet family [22, 53, 55] can be viewed as a learnable copy of a U-net encoder, which reuses an off-the-shelf diffusion model as a backbone for handling diverse spatial conditioning vectors. The copy encoder is connected with U-net decoder layers in a frozen diffusion model by zero-convolution modules. Instead of a Siamese network of encoder layers, T2I-Adapter [29] learns various low-complexity adapters according to different conditions, where each adapter is only composed of some residual convolution blocks. However, the demands for high-quality data and considerable computational resources remain the bottlenecks for scalability and widespread adoption of the above approaches." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.327, + 0.455, + 0.344 + ], + "angle": 0, + "content": "2.2. Training-free Conditional Image Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.349, + 0.485, + 0.759 + ], + "angle": 0, + "content": "Training-free works focus on modifying the original diffusion sampling process to drive the conditional image synthesis using feedback from the guidance function. For text condition, mainstream guidance frameworks are classifier and CLIP guidance. Specifically, classifier guidance [7] incorporates class information into diffusion models using gradients of a timestep-aware image classifier. CLIP guidance, such as DOODL [43], replaces classification scores with global CLIP scores, which can make generation process accommodate free-form descriptions. Regarding layout condition, current approaches often implicitly modify distributions of attention scores to achieve guidance. The underlying assumption is: spatial distribution of high-response attention aligns perceptually with the locations of objects in synthesized images [3]. Based on this idea, BoxDiff [50] and R&B [49] impose region and boundary constraints to stored cross-attention maps. DenseDiffusion [19] and A&R [32] further leverage semantic affinity in self-attention maps to ensure more accurate object placement. With respect to drag condition, DragDiffusion [39] and DragNoise [25] employ motion supervision and point tracking steps to optimize diffusion semantic features. These steps enable features of original points to move to destination points as effectively as possible. However, aforementioned approaches are often designed to encode a particular condition form, which restricts their generalization across a wide range of conditions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.77, + 0.218, + 0.788 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.841 + ], + "angle": 0, + "content": "This section provides a brief overview of the core mechanisms in the conditional diffusion model framework, including diffusion sampling and diffusion guidance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Diffusion sampling. Diffusion sampling is designed to generate a clean image by progressively removing noise from a random Gaussian noise image \\(\\mathbf{x}_T\\) [14, 51]. This sampling process is accomplished by a learned denoising" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "neural network \\(\\epsilon_{\\theta}\\) which is conditioned on a given prompt \\(\\mathbf{y}\\). Concretely, at each time step \\(t\\), a cleaner image \\(\\mathbf{x}_{t-1}\\) can be generated by subtracting the predicted noise \\(\\hat{\\epsilon}_t = \\epsilon_{\\theta}(\\mathbf{x}_t; t, \\mathbf{y})\\) from the current noise image \\(\\mathbf{x}_t\\). As an example of denoising diffusion implicit model (DDIM) [40], \\(\\mathbf{x}_{t-1}\\) can be produced via:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.196, + 0.907, + 0.214 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {t - 1} = \\sqrt {\\alpha_ {t - 1}} \\hat {\\mathbf {x}} _ {0} + \\sqrt {1 - \\alpha_ {t - 1}} \\hat {\\epsilon} _ {t}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.22, + 0.56, + 0.232 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.63, + 0.23, + 0.907, + 0.265 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} _ {0} = \\frac {\\mathbf {x} _ {t} - \\sqrt {1 - \\alpha_ {t}} \\hat {\\boldsymbol {\\epsilon}} _ {t}}{\\sqrt {\\alpha_ {t}}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.269, + 0.907, + 0.314 + ], + "angle": 0, + "content": "where scalar \\(\\alpha_{t}\\) governs denoising intensity at each timestep \\(t\\). From Eq. 2, the approximation of the clean image \\(\\hat{\\mathbf{x}}_0\\) can be predicted by performing one-step diffusion sampling." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.314, + 0.906, + 0.42 + ], + "angle": 0, + "content": "Diffusion guidance. The purpose of diffusion guidance is to produce a more reasonable denoising direction based on auxiliary information \\(\\mathbf{y}'\\) during inference time [2]. Specifically, at each timestep \\(t\\), diffusion guidance enables the denoised image \\(\\mathbf{x}_{t-1}\\) follow the auxiliary information \\(\\mathbf{y}'\\) by modifying the predicted noise \\(\\hat{\\epsilon}_t\\) using the gradient of the energy function \\(g(\\mathbf{x}_t; t, \\mathbf{y}')\\):" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.429, + 0.907, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\epsilon} _ {t} = \\hat {\\epsilon} _ {t} + \\sigma_ {t} \\nabla_ {\\mathbf {x} _ {t}} g (\\mathbf {x} _ {t}; t, \\mathbf {y} ^ {\\prime}), \\qquad (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.456, + 0.738, + 0.47 + ], + "angle": 0, + "content": "where \\(\\sigma_{t}\\) is a weighting schedule." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.907, + 0.669 + ], + "angle": 0, + "content": "In practice, the function \\( g \\) could be the approximate energy from any differentiable model, such as probabilities from a classifier [7], similarity scores from a vision-language model [34], position penalties from a detection [6] or segmentation model [45], and motion consistencies from a flow estimator [41]. Recent studies [2, 11] have shown that the predicted clean image \\( \\hat{\\mathbf{x}}_0 \\) (See Eq. 2) can be straightforwardly used as the input for the energy function \\( g \\), demonstrating that we do not require to learn an additional timestep-dependent differentiable model. This paper also computes the energy based on a one step approximate of the clean image \\( \\hat{\\mathbf{x}}_0 \\) to produce the gradient \\( \\nabla_{\\mathbf{x}_t}g(\\hat{\\mathbf{x}}_0;\\mathbf{y}') \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.649, + 0.697 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Our framework is designed for modular conditional image synthesis, generating diverse visual scenarios by flexibly combining three fundamental condition units: text, layout, and drag. Accordingly, it contains three diffusion guidance modules specifically designed for each unit. To clarify each module, we organize the methodology section as follows: in Sec. 4.1, we design a Dense Concept Alignment (DCA) module to make synthesized images follow semantic cues embedded in texts; in Sec. 4.2, we present a Dense Geometry Alignment (DGA) to enforce adherence to spatial configurations within layouts; in Sec. 4.3, we propose a Dense Motion Alignment (DMA) to allow diffusion model to be aware of movement and deformation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13137" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.088, + 0.482, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.282, + 0.483, + 0.339 + ], + "angle": 0, + "content": "Figure 2. The overall pipeline of our Dense Concept Alignment (DCA) module. Beyond scene-level vision-language consistency, this module can guide synthesized image in achieving dense alignment of attribute and relation concepts within textual conditions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.366, + 0.383, + 0.383 + ], + "angle": 0, + "content": "4.1. Dense Concept Alignment (DCA)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.388, + 0.484, + 0.539 + ], + "angle": 0, + "content": "Given a textual description \\(\\mathcal{T}\\) as the condition, our goal is to guide the synthesized image to faithfully reflect its semantic cues. Current approaches [43] often regard the description \\(\\mathcal{T}\\) and predicted clean image \\(\\hat{\\mathbf{x}}_0\\) as a whole for alignment. However, they fail to match compositional region and word information, resulting in semantic drift. Thus, we propose Dense Concept Align (DCA). As shown in Fig. 2, we decouple \\(\\mathcal{T}\\) and \\(\\hat{\\mathbf{x}}_0\\) into intra-object attributes and inter-object relations, and then align these concepts to establish a fine-grained vision-language correspondence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.54, + 0.484, + 0.766 + ], + "angle": 0, + "content": "We begin by deriving textual concepts from the description \\(\\mathcal{T}\\). Given an example like \"A red car in front of a snowy mountain,\" we employ scene graph parsing [46] to decompose it into textual attribute concepts (e.g., red car, snowy mountain) and relation concepts (e.g., car in front of mountain). Each concept is individually embedded using the CLIP text encoder [34]. An attribute concept is defined as an adjective-noun pair, where the adjective describes the property (e.g., red, snowy) and the noun specifies the category (e.g., car, mountain). We construct its embedding \\(\\mathbf{w}^{\\mathrm{a},\\mathrm{c}}\\) by concatenating the individual embeddings of the property and category: \\(\\mathbf{w}^{\\mathrm{a},\\mathrm{c}} = \\mathbf{w}^{\\mathrm{a}} \\oplus \\mathbf{w}^{\\mathrm{c}}\\). In contrast, a relation concept is treated as a complete phrase and directly encoded to obtain its contextual representation \\(\\mathbf{w}^{\\mathrm{r}}\\). This process yields a collection of textual concepts, denoted as \\(\\mathcal{W}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.778, + 0.483, + 0.797 + ], + "angle": 0, + "content": "\\[\n\\mathcal {W} = \\left\\{\\left\\{\\mathbf {w} _ {i} ^ {\\mathrm {a}, \\mathrm {c}} \\right\\} _ {i = 1} ^ {I}, \\left\\{\\mathbf {w} _ {k} ^ {\\mathrm {r}} \\right\\} _ {k = 1} ^ {K} \\right\\}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.84 + ], + "angle": 0, + "content": "where \\(I\\) and \\(K\\) denote the number of parsed objects and relations, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Simultaneously, we utilize visual region parsing to extract visual concepts from image \\(\\hat{\\mathbf{x}}_0\\). Concretely, we decouple the corresponding visual region of textual concepts using an open-set detection model [27]. The visual attribute" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.183 + ], + "angle": 0, + "content": "concept is represented by the detected bounding box enclosing an individual object, whereas the relation one is defined by the union of two object boxes. Using these bounding boxes, we crop \\(\\hat{\\mathbf{x}}_0\\) to generate a set of isolated visual regions, which are then independently embedded by CLIP vision encoder[34], forming the visual concept collection \\(\\mathcal{V}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.195, + 0.905, + 0.214 + ], + "angle": 0, + "content": "\\[\n\\mathcal {V} = \\left\\{\\left\\{\\mathbf {v} _ {i} ^ {\\mathrm {a}, \\mathrm {c}} \\right\\} _ {i = 1} ^ {I}, \\left\\{\\mathbf {v} _ {k} ^ {\\mathrm {r}} \\right\\} _ {k = 1} ^ {K} \\right\\}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.227, + 0.905, + 0.258 + ], + "angle": 0, + "content": "where \\(\\mathbf{v}^{\\mathrm{a,c}}\\) and \\(\\mathbf{v}^{\\mathrm{r}}\\) represent the visual embeddings of attribute and relation concepts, respectively." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.258, + 0.905, + 0.304 + ], + "angle": 0, + "content": "After obtaining embeddings of decoupled textual and visual concepts, beyond scene-level, we achieve finer-grained alignment by:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.319, + 0.905, + 0.335 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D C A}} = (1 - \\gamma) \\mathcal {L} _ {\\text {s c e n e}} + \\gamma (\\mathcal {L} _ {\\text {a t t r}} + \\mathcal {L} _ {\\text {r e l}}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.35, + 0.56, + 0.362 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.366, + 0.905, + 0.383 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s c e n e}} = \\cos \\left(\\mathbf {w} ^ {\\mathrm {s}}, \\mathbf {v} ^ {\\mathrm {s}}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.396, + 0.905, + 0.437 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a t t r}} (\\mathcal {W}, \\mathcal {V}) = \\frac {1}{I} \\sum_ {i = 1} ^ {I} \\cos \\left(\\mathbf {w} _ {i} ^ {\\mathrm {a}, \\mathrm {c}}, \\mathbf {v} _ {i} ^ {\\mathrm {a}, \\mathrm {c}}\\right), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.451, + 0.905, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e l} (\\mathcal {W}, \\mathcal {V}) = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\cos \\left(\\mathbf {w} _ {k} ^ {\\mathrm {r}}, \\mathbf {v} _ {k} ^ {r}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.905, + 0.547 + ], + "angle": 0, + "content": "where \\(\\cos (\\cdot ,\\cdot)\\) is the cosine similarity function. \\(\\mathbf{w}^{\\mathrm{s}}\\) and \\(\\mathbf{v}^{\\mathrm{s}}\\) are the global embeddings of description \\(\\mathcal{T}\\) and approximated image \\(\\hat{\\mathbf{x}}_0\\), respectively. We set \\(\\gamma = 0.3\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.56, + 0.819, + 0.576 + ], + "angle": 0, + "content": "4.2. Dense Geometry Alignment (DGA)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.906, + 0.763 + ], + "angle": 0, + "content": "To guide the synthesized image to resemble the spatial configuration described by the collection of segmentation masks \\(\\mathcal{M}\\) or bounding boxes \\(\\mathcal{B}\\), we propose Dense Geometry Alignment (DGA) module, as shown in Fig. 3. Current approaches simply ensure that each object appears in its designated region by enhancing corresponding attention scores [19, 32, 49, 50]. Different from them, our module improves the generation plausibility from the perspectives of the object-wise locations, as well as relative size and distance between different objects. For simplicity, we use the segmentation mask collection \\(\\mathcal{M} = \\{\\mathbf{m}\\}_{n=1}^{N}\\) as an example to illustrate this module." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Given the segmentation mask collection \\(\\hat{\\mathcal{M}} = \\{\\hat{\\mathbf{m}}\\}_{n=1}^{N}\\) predicted from the approximated clean image \\(\\hat{\\mathbf{x}}_0\\) by leveraging an image segmentation model [44], considering an object with the \\(n\\)-th class, we strive to avoid it overlapping with the other objects and ensure the coverage of the target segmentation mask \\(\\mathbf{m}_n\\). Specifically, we enhance the IoU between the two segmentation masks \\(\\hat{\\mathbf{m}}_n\\) and \\(\\mathbf{m}_n\\), and minimize the interaction between \\(\\hat{\\mathbf{m}}_n\\) and other irrelevant masks \\(\\mathcal{M} - \\{\\mathbf{m}_n\\}\\). Hence, the coverage loss function is" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13138" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.488, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.25, + 0.486, + 0.32 + ], + "angle": 0, + "content": "Figure 3. The overall pipeline of our Dense Geometry Alignment (DGA) module. On the basis of the detection information from the approximated image, this module can densely align geometric features between the predicted and reference layouts, focusing on object-level location, size, and distance." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.345, + 0.194, + 0.361 + ], + "angle": 0, + "content": "formulated by:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.368, + 0.482, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {c v g} (\\hat {\\mathcal {M}}, \\mathcal {M}) = \\sum_ {n = 1} ^ {N} \\left(1 - \\frac {\\hat {\\mathbf {m}} _ {n} \\cap \\mathbf {m} _ {n}}{\\hat {\\mathbf {m}} _ {n} \\cup \\mathbf {m} _ {n}}\\right) \\\\ + \\sum_ {n _ {1} = 1} ^ {N} \\sum_ {n _ {2} = 1} ^ {N} \\left(\\hat {\\mathbf {m}} _ {n _ {1}} \\cap \\mathbf {m} _ {n _ {2}}\\right), \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.464, + 0.484, + 0.523 + ], + "angle": 0, + "content": "where for clarity and brevity, when two segmentation masks are concurrently referenced in an equation, we use the subscript \\( n_1 \\) and \\( n_2 \\) to describe the class of different boxes for distinguishment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.484, + 0.615 + ], + "angle": 0, + "content": "Since maintaining realistic proportions between different objects is crucial for generating visually plausible images, we also restrict the pairwise size ratios of segmentation masks by the proposed size loss. On the basis of the areas computed for the predicted and target segmentation masks, the size loss is formulated by:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.624, + 0.483, + 0.667 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s i z e}} = \\sum_ {n _ {1} = 1} ^ {N} \\sum_ {n _ {2} = 1} ^ {N} \\left(\\frac {\\hat {\\mathbf {s}} _ {n _ {1}}}{\\hat {\\mathbf {s}} _ {n _ {2}}} - \\frac {\\mathbf {s} _ {n _ {1}}}{\\mathbf {s} _ {n _ {2}}}\\right) ^ {2}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.704 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{s}}_{n_1}(\\hat{\\mathbf{s}}_{n_2})\\) and \\(\\mathbf{s}_{n_1}(\\mathbf{s}_{n_2})\\) are areas of the segmentation masks in the collections \\(\\hat{\\mathcal{M}}\\) and \\(\\mathcal{M}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.855 + ], + "angle": 0, + "content": "In addition, we enforce the alignment of pairwise distances between multiple objects among the approximated image \\(\\hat{\\mathbf{x}}_0\\) and the given guidance information \\(\\mathcal{M}\\). In particular, after obtaining centroids of each predicted and target object according to the segmentation mask collections \\(\\hat{\\mathcal{M}}\\) and \\(\\mathcal{M}\\) respectively, we compute the pairwise Euclidean distances within each collection of centroids. By minimizing discrepancy in distances between predicted and target centroid collections, we formulate the distance loss function as:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.861, + 0.484, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i s t}} = \\sum_ {n _ {1} = 1} ^ {N} \\sum_ {n _ {1} = 2} ^ {N} \\left(\\| \\hat {\\mathbf {c}} _ {n _ {1}} - \\hat {\\mathbf {c}} _ {n _ {2}} \\| _ {2} - \\| \\mathbf {c} _ {n _ {1}} - \\mathbf {c} _ {n _ {2}} \\| _ {2}\\right), \\tag {12}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.087, + 0.911, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.232, + 0.907, + 0.303 + ], + "angle": 0, + "content": "Figure 4. The overall pipeline of our Dense Motion Alignment (DMA) module. Using our densified reference drag flow, this module aims to guide synthesized image in achieving dense alignment of motion information across drag conditions, encompassing pixel displacement, appearance, and semantic aspects." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.33, + 0.905, + 0.361 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{c}}_{n_1}(\\hat{\\mathbf{c}}_{n_2})\\) and \\(\\mathbf{c}_{n_1}(\\mathbf{c}_{n_2})\\) are centroids of the segmentation masks in the collections \\(\\hat{\\mathcal{M}}\\) and \\(\\mathcal{M}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.362, + 0.905, + 0.391 + ], + "angle": 0, + "content": "In summary, our final guidance function for DGA can be defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.407, + 0.905, + 0.424 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D G A}} = (1 - \\lambda) \\mathcal {L} _ {c v g} + \\lambda \\left(\\mathcal {L} _ {s i z e} + \\mathcal {L} _ {d i s t}\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.438, + 0.673, + 0.452 + ], + "angle": 0, + "content": "where we set \\(\\lambda = 0.25\\)" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.465, + 0.803, + 0.481 + ], + "angle": 0, + "content": "4.3. Dense Motion Alignment (DMA)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.907, + 0.747 + ], + "angle": 0, + "content": "Given a collection of \\(J\\) drag signals \\(\\mathcal{P} = \\{(\\mathbf{p}_j^{\\mathrm{o}},\\mathbf{p}_j^{\\mathrm{d}})\\}_{j = 1}^{J}\\), we aim to move visual contents of each original point \\(\\mathbf{p}_j^{\\mathrm{o}}\\) to its destination \\(\\mathbf{p}_j^{\\mathrm{d}}\\). Recent studies, e.g., DragDiffusion [39] and DragNoise [25], only regulate feature similarity around \\(\\mathbf{p}_j^{\\mathrm{o}}\\) and \\(\\mathbf{p}_j^{\\mathrm{d}}\\) by implicit optimization in one selected diffusion step. However, on one hand, sparse drag signals do not provide sufficient object awareness, causing incomplete displacement in foreground. On the other hand, one-step feature optimization is unstable. To enable continuous dragging of designated pixels while preserving visual coherence during diffusion sampling, we propose Dense Motion Alignment (DMA), as shown in Fig. 4. Specifically, we predict reference dense drag flow, and enforce pixel correspondence between an image pair in terms of displacement, appearance, and semantics building on this flow. We define \\(\\hat{\\mathbf{x}}_0^{t + 1}\\) and \\(\\hat{\\mathbf{x}}_0^t\\) as original and dragging images, respectively, for each timestep \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.909, + 0.905 + ], + "angle": 0, + "content": "We generate the reference dense drag flow for each individual drag signal using an object-specific displacement field. In particular, we employ SAM [20] to obtain a segmentation mask collection \\(\\{\\mathbf{m}_j^{\\mathrm{sam}}\\}_{j = 1}^{J}\\), where each mask represents an instance area of \\(\\mathbf{p}_j^{\\mathrm{o}}\\). Because image pixels closer to \\(\\mathbf{p}_j^{\\mathrm{o}}\\) should have larger displacements [48], we use radial basis functions to produce a collection of 2D Gaussian maps \\(\\{\\mathbf{g}_j\\}_{j = 1}^{J}\\). Each map depicts gradually diminishing drag influence radiating outward from \\(\\mathbf{p}_j^{\\mathrm{o}}\\). Based on \\(\\{\\mathbf{m}_j^{\\mathrm{sam}}\\}_{j = 1}^{J}\\) and \\(\\{\\mathbf{g}_j\\}_{j = 1}^{J}\\), our reference flow can be densi" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13139" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.145, + 0.106 + ], + "angle": 0, + "content": "fied by:" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.117, + 0.483, + 0.141 + ], + "angle": 0, + "content": "\\[\n\\mathbf {u} _ {\\hat {\\mathbf {x}} _ {0} ^ {t + 1} \\rightarrow \\hat {\\mathbf {x}} _ {0} ^ {t}, j} = \\left(\\mathbf {p} _ {j} ^ {\\mathrm {d}} - \\mathbf {p} _ {j} ^ {\\mathrm {o}}\\right) * \\mathbf {m} _ {j} ^ {\\text {s a m}} * \\mathbf {g} _ {j}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.148, + 0.483, + 0.166 + ], + "angle": 0, + "content": "where \\(\\mathbf{u}_{\\hat{\\mathbf{x}}_0^{t + 1}\\rightarrow \\hat{\\mathbf{x}}_0^t,j}\\) is the densified drag flow for \\(j\\) -th signal." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.166, + 0.484, + 0.226 + ], + "angle": 0, + "content": "To rectify discrepancies in pixel displacement direction and magnitude, we enable the estimated optical flow between \\(\\hat{\\mathbf{x}}_0^{t + 1}\\) and \\(\\hat{\\mathbf{x}}_0^t\\) to align with its reference flow by minimizing displacement loss:" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.236, + 0.483, + 0.28 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i s p}} = \\sum_ {j = 1} ^ {J} \\left|\\left| \\hat {\\mathbf {u}} _ {\\hat {\\mathbf {x}} _ {0} ^ {t + 1} \\rightarrow \\hat {\\mathbf {x}} _ {0} ^ {t}, j} - \\mathbf {u} _ {\\hat {\\mathbf {x}} _ {0} ^ {t + 1} \\rightarrow \\hat {\\mathbf {x}} _ {0} ^ {t}, j} \\right|\\right| _ {1}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.485, + 0.385 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{u}}_{\\hat{\\mathbf{x}}_0^{t + 1}\\rightarrow \\hat{\\mathbf{x}}_0^t,j}\\) is an optical flow estimated by a RAFT model [41]. In order to ensure visual coherence, we enforce pixel-level appearance consistency by directly minimizing the differences between the dragging and warped original image, while semantic consistency is preserved by DINO v2 embedding [30]:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.393, + 0.483, + 0.437 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p a p p} = \\sum_ {j = 1} ^ {J} \\sqrt {\\left(\\operatorname {w a r p} \\left(\\hat {\\mathbf {x}} _ {0} ^ {t + 1} , \\mathbf {u} _ {\\hat {\\mathbf {x}} _ {0} ^ {t + 1} \\rightarrow \\hat {\\mathbf {x}} _ {0} ^ {t} , j}\\right) - \\hat {\\mathbf {x}} _ {0} ^ {t}\\right) ^ {2}}, \\tag {16}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.448, + 0.483, + 0.49 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p s e m} = \\sum_ {j = 1} ^ {J} \\sqrt {\\left(\\operatorname {w a r p} \\left(\\hat {\\mathbf {f}} _ {0} ^ {t + 1} , \\mathbf {u} _ {\\hat {\\mathbf {x}} _ {0} ^ {t + 1} \\rightarrow \\hat {\\mathbf {x}} _ {0} ^ {t} , j}\\right) - \\hat {\\mathbf {f}} _ {0} ^ {t}\\right) ^ {2}}, \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.497, + 0.484, + 0.545 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{f}}_0^{t + 1}\\) and \\(\\hat{\\mathbf{f}}_0^t\\) are semantic embedding calculated by DINO v2 model. For a drag signal, our loss of DMA module is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.556, + 0.483, + 0.575 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D M A}} = (1 - \\eta) \\mathcal {L} _ {d i s p} + \\eta \\left(\\mathcal {L} _ {p a p p} + \\mathcal {L} _ {p s e m}\\right), \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.484, + 0.631 + ], + "angle": 0, + "content": "where we set \\(\\eta = 0.02\\). Like DragDiffusion [39] and Drag-Noise [25], we also update the positions of the original points after each denosing step." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.643, + 0.224, + 0.66 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.289, + 0.684 + ], + "angle": 0, + "content": "5.1. Experimental Setups" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.734 + ], + "angle": 0, + "content": "This section sequentially describes used datasets and evaluation metrics in our experiments, and implementation details of our approaches." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Datasets. The datasets used in our study are: (1) COCO [24]. We curate the 400 longest image descriptions from COCO validation set, each ranging from 15 to 50 words. This selection is made to better demonstrate our strength in understanding long-form description; (2) DenseDiffusion [19]. This benchmark contains 250 samples \\((2\\sim 3\\) unique objects of each sample), where each sample has a global image description, a segmentation mask, and local descriptions and class labels for each segment; (3) DrawBench [38]. We use the 'position' subset from it. This subset is composed of 20 samples (2 unique objects of each sample)," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "where each sample has a description and some bounding boxes of each object (box coordinates and class labels generated by GPT); and (4) DragBench [39]. This benchmark comprises 205 samples, each of which has a reference image, one or more drag signals, a description, and a mask indicating regions to be modified." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.19, + 0.907, + 0.449 + ], + "angle": 0, + "content": "Evaluation metrics. For text condition, we use Fréchet Inception Distance (FID) [12] and CLIP score [34]. FID quantifies the similarity between the synthesized and real images, while the CLIP score measures the alignment between the synthesized image and the reference description. For text + layout condition, we use IoU and SOA-I [13] (layout prediction is based on YOLOv7-Seg [44]). IoU quantifies the overlap between the predicted layout from the synthesized image and the reference layout. SOA-I measures whether described objects appear in the synthesized image. For text + drag condition, we use Image Fidelity (IF) [17] and Mean Distance (MD) [31]. IF (1-LPIPS score [54]) measures the image quality based on the global similarity between the reference and dragged images. MD assesses how accurately the post-dragging original point align with its corresponding destination point by measuring Euclidean distance between their positions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.909, + 0.549 + ], + "angle": 0, + "content": "Implementation details. We use \\( T = 50 \\) steps of the DDIM scheduler [40] and \\( T = 28 \\) steps of the Flow Match Euler Discrete scheduler [10]. The weighting schedules for our proposed DCA, DGA, and DMA modules are set to 60, 25, and 90, respectively. Our guidance is only applied during the first \\( 70\\% \\) timesteps." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.583, + 0.813, + 0.6 + ], + "angle": 0, + "content": "5.2. Image Synthesis Results with Text" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We use Stable Diffusion v1.5 [36] and v3.0 [8] as baseline models to evaluate the effectiveness of our DCA module in aligning concepts within descriptions. Tab. 1 summarizes FID [12] and CLIP scores [34] on long (\\(>15\\) words) descriptions from COCO dataset [24], our DCA module achieves \\(12.3\\%\\) and \\(2.97\\%\\) improvements in FID and CLIP scores, respectively, building upon Stable Diffusion v3.0 (the most recent baseline). This gain arises from our module's ability to achieve vision-language alignment in a coarse-to-fine manner, effectively leveraging semantic cues embedded in complex sentences. Besides, our module can be applicable to Stable Diffusion v1.5, showing its adaptability across different baseline versions. As shown in Fig. 5, our approach can properly place the dog on a dog bed next to a rocking chair (top) and position the person's head on a blue backpack (middle), indicating its effectiveness in aligning the object interaction concept. Our approach can also generate two people (bottom), showcasing its advantage in understanding the object's numeric information." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13140" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.148 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison under long descriptions from COCO dataset [24]. The evaluation metrics used are FID [12] and CLIPScore [34]. We reimplement all baseline approaches. The best results are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.156, + 0.482, + 0.312 + ], + "angle": 0, + "content": "
MethodsVenueFID ↓CLIPScore ↑
ParaDiffusion [47]arXiv202312.960.2731
PIXART-α [5]ICLR202413.970.2724
SDXL [33]ICLR20249.790.2820
ELLA [16]arXiv202413.640.2811
PIXART-Σ [4]ECCV202410.730.2793
Stable Diffusion v1.5 [36]CVPR202211.490.2750
+ DCA-10.290.2882
Stable Diffusion v3.0 [8]ICML202410.510.2861
+ DCA-9.220.2946
" + }, + { + "type": "table", + "bbox": [ + 0.091, + 0.326, + 0.484, + 0.543 + ], + "angle": 0, + "content": "
DescriptionSD v3.0Ours
A room with filtered light coming through vertical blinds on a doorwall, with a sleeping dog lying on a dog bed next to a rocking chair.
A person covered with an orange sleeping bag rests his head on a blue back pack as he sleeps on a park bench.
In the desert, two people are sitting next to a small propeller plane next to a tent with other planes in the background.
" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.552, + 0.485, + 0.61 + ], + "angle": 0, + "content": "Figure 5. Visualization comparison between our approach (Stable Diffusion v3.0 [8] + DCA) and its baseline under varying descriptions. Our approach demonstrates superior performance in aligning semantic concepts within descriptions." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.636, + 0.482, + 0.654 + ], + "angle": 0, + "content": "5.3. Image Synthesis Results with Layout and Text" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.659, + 0.486, + 0.903 + ], + "angle": 0, + "content": "We use Stable Diffusion v1.5 [36], DenseDiffusion [19] and A&R [32] as baseline models to evaluate the effectiveness of our DGA module in aligning spatial configurations represented by segmentation masks and bounding boxes; and our DCA module in aligning geometric concepts within descriptions. Tab. 2 summarizes IoU and SOA-I scores on mask-to-image. Our DGA module brings a \\(27.1\\%\\) improvement in IoU and a \\(7.55\\%\\) increase in SOA-I over A&R baseline, owing to the establishment of object-wise geometric consistency. Besides, incorporating our DCA module on top of DGA leads to an additional improvement, achieving an IoU of 50.61 and an SOA-I of 86.33. This improvement is due to the ability of DCA module to understand descriptions of positions and shapes. We can also observe consistent performance gains across both Stable Diffusion v1.5 and DenseDiffusion architectures. As shown in Fig. 6, our" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.161 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison under descriptions and segmentation masks provided by DenseDiffusion [19]. The evaluation metrics are IoU and SOA-I [13]. We generate four random images of each condition combination for performance evaluation. The best results are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.171, + 0.907, + 0.372 + ], + "angle": 0, + "content": "
MethodsVenueIoU ↑SOA-I ↑
Real Images-59.1393.07
Composable Diffusion [26]ECCV2022-55.88±0.78
Structure Diffusion [9]ICLR2023-70.97±1.08
SD-Pww [1]arXiv202223.76±0.5073.92±1.84
Stable Diffusion [36]CVPR2022-73.08±1.54
+ DGA-25.46±1.0274.45±1.46
+ DGA + DCA-27.28±0.9777.02±1.67
Dense Diffusion [19]ICCV202334.99±1.1377.61±1.75
+ DGA-36.72±1.2478.21±1.62
+ DGA + DCA-38.06±1.2678.42±1.17
A&R [32]CVPR202438.97±0.5678.80±1.27
+ DGA-49.52±0.7084.75±2.46
+ DGA + DCA-50.61±0.4786.33±0.88
" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.385, + 0.907, + 0.602 + ], + "angle": 0, + "content": "
DescriptionLayoutA&ROurs
An orange tabby cat lying in a bathroom sink.
An orange cat laying on top of a black car.
A dog sitting on a bed wearing an orange bowtie.
" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.612, + 0.909, + 0.671 + ], + "angle": 0, + "content": "Figure 6. Visualization comparison between our approach (A&R [32] + DGA + DCA) and its baseline, under varying descriptions and segmentation masks. Our approach can adhere to both geometric concepts and positional configuration." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.697, + 0.909, + 0.804 + ], + "angle": 0, + "content": "approach can accurately place a cat inside a bathroom sink (top), position a cat on top of a black car (middle), generate a bowtie around a dog's neck (bottom), demonstrating the advantage of our framework in awareness of spatial relationships. The same conclusion can be observed in Fig. 7, which presents box-to-image visualizations of validation data from the DrawBench benchmark [38]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.817, + 0.89, + 0.835 + ], + "angle": 0, + "content": "5.4. Image Synthesis Results with Drag and Text" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We use DragDiffusion [39] as a baseline model to evaluate the effectiveness of our DMA module in motion alignment; and our DCA module for semantic preservation. Fig. 8 presents MD and IF scores on the DragBench benchmark" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "13141" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.482, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.314, + 0.482, + 0.357 + ], + "angle": 0, + "content": "Figure 7. Visualization comparison between our approach (A&R [32] + DGA + DCA) and its baseline, under varying descriptions and bounding boxes. Our approach can satisfy both conditions." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.372, + 0.48, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.605, + 0.483, + 0.661 + ], + "angle": 0, + "content": "Figure 8. Quantitative comparison under descriptions and drag signals from DragBench benchmark [39]. The evaluation metrics used are Mean Distance \\((\\downarrow)\\) and Image Fidelity (1-LPIPS) \\((\\uparrow)\\). Approaches with the best results appear in the upper-left corner." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[39]. The combination of our DMA and DCA modules achieves better MD and IF scores, indicating that our framework ensures each pixel follows its desired trajectory with fewer visual artifacts. However, applying our DMA module alone results in a slight drop in IF, compared to DragDiffusion. This is because our motion field may cause oversmoothing in detail. Fortunately, our DCA module can mitigate such an issue to some degree, by maintaining the overall semantic structure unchanged during the diffusion sampling process. As shown in Fig. 9, our framework can effectively synthesize head rotation (top), rigid car displacement (middle), and large-scale deformation of the bus (bottom). This demonstrates that our algorithm enables the desired movement even when handling complex drag signals" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.087, + 0.905, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.356, + 0.907, + 0.412 + ], + "angle": 0, + "content": "Figure 9. Visualization comparison between our approach (DragDiffusion [39] + DMA) and its baseline under varying drag signals. Our approach enables effective synthesis of large-scale movement and deformation." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.427, + 0.665, + 0.442 + ], + "angle": 0, + "content": "(multiple drag points)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.455, + 0.634, + 0.47 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.48, + 0.907, + 0.678 + ], + "angle": 0, + "content": "We propose a plug-and-play modular design for guiding image synthesis by aligning diverse fundamental condition units (text, layout, drag) in combination. Based on this idea, we devise three dense alignment approaches to independently achieve flexible control over concept, geometry and motion. Comprehensive experiments demonstrate the effectiveness of our framework across diverse conditioning scenarios. However, our framework relies on the capability of its foundation model, e.g., SD model. As shown in Fig. 10, counterfactual scenes is hard to generate, e.g., a river flowing upwards, if SD model is unable to synthesis this scene itself. Future work could explore world knowledge-aware generative paradigms to address this challenge." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.687, + 0.905, + 0.758 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.768, + 0.907, + 0.809 + ], + "angle": 0, + "content": "Figure 10. Failure cases. As foundation model lacks world knowledge, our approach (Stable Diffusion v3.0 [8] + DCA) fails to generate counterfactual scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement: This work was supported by the National Natural Science Foundation of China (No. U23B2013, 62276176). This work was also partly supported by the SICHUAN Provincial Natural Science Foundation (No. 2024NSFJQ0023)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13142" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.267 + ], + "angle": 0, + "content": "[2] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 843-852, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.268, + 0.483, + 0.322 + ], + "angle": 0, + "content": "[3] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.324, + 0.483, + 0.392 + ], + "angle": 0, + "content": "[4] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.393, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[5] Junsong Chen, YU Jincheng, GE Chongjian, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In Proceedings of the International Conference on Learning Representations (ICLR), pages 1–30, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.476, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[6] Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xinggang Wang, and Ying Shan. Yolo-world: Real-time open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16901-16911, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.545, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems (NeruIPS), 34:8780-8794, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.587, + 0.482, + 0.668 + ], + "angle": 0, + "content": "[8] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning (ICML), 2024. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.669, + 0.482, + 0.751 + ], + "angle": 0, + "content": "[9] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. In Proceedings of the International Conference on Learning Representations (ICLR), pages 1-13, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.753, + 0.483, + 0.807 + ], + "angle": 0, + "content": "[10] Itai Gat, Tal Remez, Neta Shaul, Felix Kreuk, Ricky TQ Chen, Gabriel Synnaeve, Yossi Adi, and Yaron Lipman. Discrete flow matching. Advances in Neural Information Processing Systems (NeurIPS), 37:133345-133385, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.808, + 0.483, + 0.863 + ], + "angle": 0, + "content": "[11] Daniel Geng and Andrew Owens. Motion guidance: Diffusion-based image editing with differentiable motion estimators. In The Twelfth International Conference on Learning Representations (ICLR), pages 1-21. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.864, + 0.483, + 0.891 + ], + "angle": 0, + "content": "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.891 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[13] Tobias Hinz, Stefan Heinrich, and Stefan Wermter. Semantic object accuracy for generative text-to-image synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(3):1552-1565, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NeurIPS), 33:6840-6851, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[15] Xingzhong Hou, Boxiao Liu, Yi Zhang, Jihao Liu, Yu Liu, and Haihang You. Easydrag: Efficient point-based manipulation on diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8404-8413, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[16] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[17] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6007-6017, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.433, + 0.905, + 0.515 + ], + "angle": 0, + "content": "[18] Seung Wook Kim, Bradley Brown, Kangxue Yin, Karsten Kreis, Katja Schwarz, Daiqing Li, Robin Rombach, Antonio Torralba, and Sanja Fidler. Neuralfield-ldm: Scene generation with hierarchical latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8496-8506, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.517, + 0.905, + 0.585 + ], + "angle": 0, + "content": "[19] Yunji Kim, Jiyoung Lee, Jin-Hwa Kim, Jung-Woo Ha, and Jun-Yan Zhu. Dense text-to-image generation with attention modulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7701-7711, 2023. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.587, + 0.905, + 0.668 + ], + "angle": 0, + "content": "[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4015-4026, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.671, + 0.905, + 0.727 + ], + "angle": 0, + "content": "[21] Yinjie Lei, Zixuan Wang, Feng Chen, Guoqing Wang, Peng Wang, and Yang Yang. Recent advances in multi-modal 3d scene understanding: A comprehensive survey and evaluation. arXiv preprint arXiv:2310.15676, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.728, + 0.905, + 0.797 + ], + "angle": 0, + "content": "[22] Ming Li, Taojiannan Yang, Huafeng Kuang, Jie Wu, Zhaoning Wang, Xuefeng Xiao, and Chen Chen. Controlnet++: Improving conditional controls with efficient consistency feedback. In European Conference on Computer Vision (ECCV), pages 129-147. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.799, + 0.905, + 0.879 + ], + "angle": 0, + "content": "[23] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22511-22521, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.879 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13143" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision (ECCV), pages 740-755. Springer, 2014. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.483, + 0.232 + ], + "angle": 0, + "content": "[25] Haofeng Liu, Chenshu Xu, Yifei Yang, Lihua Zeng, and Shengfeng He. Drag your noise: Interactive point-based editing via diffusion semantic propagation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6743-6752, 2024. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.232, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[26] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In Proceedings of the European Conference on Computer Vision (ECCV), pages 423-439. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.3, + 0.482, + 0.383 + ], + "angle": 0, + "content": "[27] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision (ECCV), pages 38-55. Springer, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.383, + 0.482, + 0.466 + ], + "angle": 0, + "content": "[28] Sicheng Mo, Fangzhou Mu, Kuan Heng Lin, Yanli Liu, Bochen Guan, Yin Li, and Bolei Zhou. Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7465-7475, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.466, + 0.482, + 0.536 + ], + "angle": 0, + "content": "[29] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 4296-4304, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.536, + 0.482, + 0.605 + ], + "angle": 0, + "content": "[30] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.605, + 0.482, + 0.674 + ], + "angle": 0, + "content": "[31] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.674, + 0.482, + 0.742 + ], + "angle": 0, + "content": "[32] Quynh Phung, Songwei Ge, and Jia-Bin Huang. Grounded text-to-image synthesis with attention refocusing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7932-7942, 2024. 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.742, + 0.482, + 0.826 + ], + "angle": 0, + "content": "[33] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In Proceedings of the International Conference on Learning Representations (ICLR), pages 1-13, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.826, + 0.482, + 0.907 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PMLR, 2021. 2, 3, 4, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.907 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[35] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research (JMLR), 21(140):1-67, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.907, + 0.233 + ], + "angle": 0, + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[37] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6527-6536, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.318, + 0.905, + 0.415 + ], + "angle": 0, + "content": "[38] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems (NeurIPS), 35:36479-36494, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.416, + 0.905, + 0.5 + ], + "angle": 0, + "content": "[39] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8839-8849, 2024. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.5, + 0.905, + 0.543 + ], + "angle": 0, + "content": "[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations (ICLR), pages 1-22, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.543, + 0.905, + 0.599 + ], + "angle": 0, + "content": "[41] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European Conference on Computer Vision (ECCV), pages 402-419. Springer, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.6, + 0.905, + 0.669 + ], + "angle": 0, + "content": "[42] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1921-1930, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.669, + 0.905, + 0.74 + ], + "angle": 0, + "content": "[43] Bram Wallace, Akash Gokul, Stefano Ermon, and Nikhil Naik. End-to-end diffusion latent optimization improves classifier guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR), pages 7280-7290, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.74, + 0.905, + 0.824 + ], + "angle": 0, + "content": "[44] Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7464-7475, 2023. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.824, + 0.905, + 0.895 + ], + "angle": 0, + "content": "[45] Yang Wang, Zhikui Ouyang, Runhua Han, Zhijian Yin, and Zhen Yang. Yolomask: Real-time instance segmentation with integrating yolov5 and orienmask. In IEEE International Conference on Communication Technology (ICCT), pages 1646-1650. IEEE, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13144" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[46] Hao Wu, Jiayuan Mao, Yufeng Zhang, Yuning Jiang, Lei Li, Weiwei Sun, and Wei-Ying Ma. Unified visual-semantic embeddings: Bridging vision and language with structured meaning representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6609–6618, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.177, + 0.483, + 0.244 + ], + "angle": 0, + "content": "[47] Weijia Wu, Zhuang Li, Yefei He, Mike Zheng Shou, Chunhua Shen, Lele Cheng, Yan Li, Tingting Gao, Di Zhang, and Zhongyuan Wang. Paragraph-to-image generation with information-enriched diffusion model. arXiv preprint arXiv:2311.14284, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.245, + 0.483, + 0.314 + ], + "angle": 0, + "content": "[48] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control for anything using entity representation. In European Conference on Computer Vision (ECCV), pages 331-348. Springer, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.315, + 0.483, + 0.37 + ], + "angle": 0, + "content": "[49] Jiayu Xiao, Henglei Lv, Liang Li, Shuhui Wang, and Qingming Huang. R&b: Region and boundary aware zero-shot grounded text-to-image generation. arXiv preprint arXiv:2310.08872, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.371, + 0.482, + 0.452 + ], + "angle": 0, + "content": "[50] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7452-7461, 2023. 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.483, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[51] Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and Jian Zhang. Freedom: Training-free energy-guided conditional diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 23174-23184, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[52] Yu Zeng, Vishal M Patel, Haochen Wang, Xun Huang, Ting-Chun Wang, Ming-Yu Liu, and Yogesh Balaji. Jedi: Joint-image diffusion models for finetuning-free personalized text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6786-6795, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.248, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[53] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR), pages 3836-3847, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.905, + 0.444 + ], + "angle": 0, + "content": "[55] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. Advances in Neural Information Processing Systems (NeurIPS), 36, 2024. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13145" + } + ] +] \ No newline at end of file diff --git a/2025/Training-free Dense-Aligned Diffusion Guidance for Modular Conditional Image Synthesis/237b334f-3be2-4e40-8ac1-466a51400496_origin.pdf b/2025/Training-free Dense-Aligned Diffusion Guidance for Modular Conditional Image Synthesis/237b334f-3be2-4e40-8ac1-466a51400496_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f7153bf23a4edc66eb7ddeaa95eb49647cbdc2e4 --- /dev/null +++ b/2025/Training-free Dense-Aligned Diffusion Guidance for Modular Conditional Image Synthesis/237b334f-3be2-4e40-8ac1-466a51400496_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7699b8afbe6c8c6443750333979ce855e286b1bc09c0bfacff218bf42652f425 +size 7828678 diff --git a/2025/Training-free Dense-Aligned Diffusion Guidance for Modular Conditional Image Synthesis/full.md b/2025/Training-free Dense-Aligned Diffusion Guidance for Modular Conditional Image Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..93c581e36c787047658a1e8a689624bfcf2cf2f4 --- /dev/null +++ b/2025/Training-free Dense-Aligned Diffusion Guidance for Modular Conditional Image Synthesis/full.md @@ -0,0 +1,353 @@ +# Training-free Dense-Aligned Diffusion Guidance for Modular Conditional Image Synthesis + +Zixuan Wang $^{1}$ Duo Peng $^{2}$ Feng Chen $^{3}$ Yuwei Yang $^{4}$ Yinjie Lei $^{1,\text{图}}$ + +$^{1}$ Sichuan University $^{2}$ Singapore University of Technology and Design + +3The University of Adelaide 4Australian National University + +zixuan98@stu.scu.edu.cn, duo_peng@mymail.sutd.edu.sg, chenfeng1271@gmail.com + +yuwei.yang@anu.edu.au, yinjie@scu.edu.cn + +# Abstract + +Conditional image synthesis is a crucial task with broad applications, such as artistic creation and virtual reality. However, current generative methods are often task-oriented with a narrow scope, handling a restricted condition with constrained applicability. In this paper, we propose a novel approach that treats conditional image synthesis as the modular combination of diverse fundamental condition units. Specifically, we divide conditions into three primary units: text, layout, and drag. To enable effective control over these conditions, we design a dedicated alignment module for each. For the text condition, we introduce a Dense Concept Alignment (DCA) module, which achieves dense visual-text alignment by drawing on diverse textual concepts. For the layout condition, we propose a Dense Geometry Alignment (DGA) module to enforce comprehensive geometric constraints that preserve the spatial configuration. For the drag condition, we introduce a Dense Motion Alignment (DMA) module to apply multi-level motion regularization, ensuring that each pixel follows its desired trajectory without visual artifacts. By flexibly inserting and combining these alignment modules, our framework enhances the model's adaptability to diverse conditional generation tasks and greatly expands its application range. Extensive experiments demonstrate the superior performance of our framework across a variety of conditions, including textual description, segmentation mask (bounding box), drag manipulation, and their combinations. Code is available at https://github.com/ZixuanWang0525/DADG + +# 1. Introduction + +Conditional image synthesis aims to generate realistic images based on user-provided requirements, playing a pivotal role in various applications, such as artistic creation [37, 52] + +![](images/86d7c1b6ed59d84f9a5722e33444e2ce4e186fe46095c8ae16b6e7ba45f47b81.jpg) +(a) General pipeline of our proposed modular conditional image synthesis. Overall framework of Training-free Dense-Aligned Diffusion Guidance is illustrated in Fig. 1b. + +![](images/de84523cf2da69fef3c48ef29666c32ca12caad66d1e385707236392a6cf5d95.jpg) +(b) Illustration of our proposed Training-free Dense-Aligned Diffusion Guidance. We design three dense alignment modules to enable an effective accommodation of each specific condition unit. +Figure 1. The overview of our modular conditional image synthesis. Driven by modular combination of fundamental condition units, diverse visual content can be synthesized. + +and virtual reality [18, 21]. Recent advances in diffusion models have demonstrated promising performance in handling diverse conditional signals, including image descriptions [36, 42], segmentation maps [19, 28, 29], bounding + +boxes [3, 50, 55], and drag information [15, 39], among others. While existing approaches are effective in interpreting a particular form of input from users, they struggle to generalize across a wide range of conditions. This is impractical for flexible generation of complex visual scenes in the real world. + +In this paper, we propose the Modular Conditional Image Synthesis (MCIS) paradigm, which regulates the synthesis process by selectively applying and combining independent fundamental condition units, as demonstrated in Fig. 1 (a). Specifically, we categorize condition units into: (1) Text. This category refers to natural language descriptions, specifying the semantic content of an image; (2) Layout. This category mainly encompasses segmentation maps and bounding boxes, representing the component arrangement and scene structure; and (3) Drag. This category refers to point pairs (original and destination of drags), encoding the local transformations of an image. However, MCIS presents a non-trivial problem, as it requires ensuring that all conditions are distinctly reflected in the visual content without omission. + +We sequentially analyze the key challenges involved in aligning each basic condition unit: (1) Concept Mismatching. Text typically conveys detailed information about the properties of individual objects and their interactions. This requires generative models to match fine-grained region-level and word-level concepts, instead of considering the image and its description as a whole for alignment; (2) Geometry Inconsistency. Layout implicitly encodes complex relationships among multiple objects, such as size ratios and relative positions. Ensuring image realism requires establishing a detailed geometric correspondence between synthesized content and spatial configuration; and (3) Motion Disharmony. Drag often provides displacement vectors indicating the visual regions to be moved. To synthesize visually coherent content, it is crucial to effectively maintain the intended movement of each visual element, while preserving both appearance and semantic consistency. + +To address the above issues, we propose a dense-aligned diffusion guidance framework for modular conditional image synthesis. As illustrated in Fig. 1 (b), our framework utilizes plug-and-play guidance modules to allow visual content to independently align with each of the condition units. Specifically, we propose the Dense Concept Alignment (DCA) module, which ensures the consistency between visual content and scene description in a coarse-to-fine manner. Beyond scene-level alignment, we establish fine-grained correspondence from attribute and relational perspectives within a disentangled feature space. We further introduce the Dense Geometry Alignment (DGA) module. Building upon the detection information from the synthesized image, it imposes constraints on object-wise locations, as well as on relative sizes and distances between + +each pair of objects. This process ensures the separation between different foreground instances while maintaining the realism of scene arrangement. Based on dense displacement fields, we design the Dense Motion Alignment (DMA) module. During diffusion sampling, this module guides the variation of visual content between adjacent timesteps by aligning it with the displacement fields. Meanwhile, it uses pixel-wise color and semantic regularization to achieve photorealistic consistency. + +To evaluate the effectiveness of our framework in MCIS, we conduct extensive experiments on public benchmarks, including COCO [24], DenseDiffusion [19], DrawBench [38], and DragBench [39]. Quantitative and qualitative results demonstrate that our approach significantly improves the adherence to the text, layout, and drag conditions. Moreover, our guidance modules are compatible to simultaneously control over multiple condition units. In addition, our framework can be seamlessly integrated into various diffusion architectures. + +Our contributions are summarized as follows: + +- We propose a paradigm of modular conditional image synthesis, which expands the application scope of diffusion models by the adaptive combination of fundamental condition units. +- We design three plug-and-play dense alignment modules, DCA, DGA, and DMA, to independently achieve effective control over concept, geometry, and motion. +- Extensive experiments show that our framework significantly enhances controllability across text, layout, drag, and their combinations. + +# 2. Related Works + +Conditional image synthesis is a core research field in AIGC, which aims to generate images explicitly guided by input conditions. From the aspect of framework architecture, current studies can be divided into: (1) Training-based and (2) Training-free condition image synthesis. Crucially, we classify diverse conditions into three fundamental units: text, layout, and drag. Training-based and -free frameworks are, thus, sequentially analyzed in a context of fundamental condition units we define. + +# 2.1. Training-based Conditional Image Synthesis + +Training-based frameworks often train diffusion models from scratch or lightweight auxiliary modules to accommodate user-provided conditions. Text-driven frameworks, e.g., Stable Diffusion family [8, 33, 36] and Pix-Art family [4, 5], optimize diffusion models from scratch to implicitly aware of embeddings from frozen large-language models, such as CLIP [34] and T5 [35]. Layout-driven approaches incorporate parameter-efficient modules to encode segmentation masks, bounding boxes, and depth maps, among others. For instance, GLIGEN [23] uses a learn- + +able gated self-attention module to capture grounding information, i.e., regional bounding boxes and its corresponding phrase. ControlNet family [22, 53, 55] can be viewed as a learnable copy of a U-net encoder, which reuses an off-the-shelf diffusion model as a backbone for handling diverse spatial conditioning vectors. The copy encoder is connected with U-net decoder layers in a frozen diffusion model by zero-convolution modules. Instead of a Siamese network of encoder layers, T2I-Adapter [29] learns various low-complexity adapters according to different conditions, where each adapter is only composed of some residual convolution blocks. However, the demands for high-quality data and considerable computational resources remain the bottlenecks for scalability and widespread adoption of the above approaches. + +# 2.2. Training-free Conditional Image Synthesis + +Training-free works focus on modifying the original diffusion sampling process to drive the conditional image synthesis using feedback from the guidance function. For text condition, mainstream guidance frameworks are classifier and CLIP guidance. Specifically, classifier guidance [7] incorporates class information into diffusion models using gradients of a timestep-aware image classifier. CLIP guidance, such as DOODL [43], replaces classification scores with global CLIP scores, which can make generation process accommodate free-form descriptions. Regarding layout condition, current approaches often implicitly modify distributions of attention scores to achieve guidance. The underlying assumption is: spatial distribution of high-response attention aligns perceptually with the locations of objects in synthesized images [3]. Based on this idea, BoxDiff [50] and R&B [49] impose region and boundary constraints to stored cross-attention maps. DenseDiffusion [19] and A&R [32] further leverage semantic affinity in self-attention maps to ensure more accurate object placement. With respect to drag condition, DragDiffusion [39] and DragNoise [25] employ motion supervision and point tracking steps to optimize diffusion semantic features. These steps enable features of original points to move to destination points as effectively as possible. However, aforementioned approaches are often designed to encode a particular condition form, which restricts their generalization across a wide range of conditions. + +# 3. Preliminary + +This section provides a brief overview of the core mechanisms in the conditional diffusion model framework, including diffusion sampling and diffusion guidance. + +Diffusion sampling. Diffusion sampling is designed to generate a clean image by progressively removing noise from a random Gaussian noise image $\mathbf{x}_T$ [14, 51]. This sampling process is accomplished by a learned denoising + +neural network $\epsilon_{\theta}$ which is conditioned on a given prompt $\mathbf{y}$ . Concretely, at each time step $t$ , a cleaner image $\mathbf{x}_{t-1}$ can be generated by subtracting the predicted noise $\hat{\epsilon}_t = \epsilon_{\theta}(\mathbf{x}_t; t, \mathbf{y})$ from the current noise image $\mathbf{x}_t$ . As an example of denoising diffusion implicit model (DDIM) [40], $\mathbf{x}_{t-1}$ can be produced via: + +$$ +\mathbf {x} _ {t - 1} = \sqrt {\alpha_ {t - 1}} \hat {\mathbf {x}} _ {0} + \sqrt {1 - \alpha_ {t - 1}} \hat {\epsilon} _ {t}, \tag {1} +$$ + +where + +$$ +\hat {\mathbf {x}} _ {0} = \frac {\mathbf {x} _ {t} - \sqrt {1 - \alpha_ {t}} \hat {\boldsymbol {\epsilon}} _ {t}}{\sqrt {\alpha_ {t}}}, \tag {2} +$$ + +where scalar $\alpha_{t}$ governs denoising intensity at each timestep $t$ . From Eq. 2, the approximation of the clean image $\hat{\mathbf{x}}_0$ can be predicted by performing one-step diffusion sampling. + +Diffusion guidance. The purpose of diffusion guidance is to produce a more reasonable denoising direction based on auxiliary information $\mathbf{y}'$ during inference time [2]. Specifically, at each timestep $t$ , diffusion guidance enables the denoised image $\mathbf{x}_{t-1}$ follow the auxiliary information $\mathbf{y}'$ by modifying the predicted noise $\hat{\epsilon}_t$ using the gradient of the energy function $g(\mathbf{x}_t; t, \mathbf{y}')$ : + +$$ +\tilde {\epsilon} _ {t} = \hat {\epsilon} _ {t} + \sigma_ {t} \nabla_ {\mathbf {x} _ {t}} g (\mathbf {x} _ {t}; t, \mathbf {y} ^ {\prime}), \qquad (3) +$$ + +where $\sigma_{t}$ is a weighting schedule. + +In practice, the function $g$ could be the approximate energy from any differentiable model, such as probabilities from a classifier [7], similarity scores from a vision-language model [34], position penalties from a detection [6] or segmentation model [45], and motion consistencies from a flow estimator [41]. Recent studies [2, 11] have shown that the predicted clean image $\hat{\mathbf{x}}_0$ (See Eq. 2) can be straightforwardly used as the input for the energy function $g$ , demonstrating that we do not require to learn an additional timestep-dependent differentiable model. This paper also computes the energy based on a one step approximate of the clean image $\hat{\mathbf{x}}_0$ to produce the gradient $\nabla_{\mathbf{x}_t}g(\hat{\mathbf{x}}_0;\mathbf{y}')$ . + +# 4. Methodology + +Our framework is designed for modular conditional image synthesis, generating diverse visual scenarios by flexibly combining three fundamental condition units: text, layout, and drag. Accordingly, it contains three diffusion guidance modules specifically designed for each unit. To clarify each module, we organize the methodology section as follows: in Sec. 4.1, we design a Dense Concept Alignment (DCA) module to make synthesized images follow semantic cues embedded in texts; in Sec. 4.2, we present a Dense Geometry Alignment (DGA) to enforce adherence to spatial configurations within layouts; in Sec. 4.3, we propose a Dense Motion Alignment (DMA) to allow diffusion model to be aware of movement and deformation. + +![](images/e431493c89228e9a530ee47ff5fc6a0ed1ef9a8eda0aaea15e4c1e13ed04de88.jpg) +Figure 2. The overall pipeline of our Dense Concept Alignment (DCA) module. Beyond scene-level vision-language consistency, this module can guide synthesized image in achieving dense alignment of attribute and relation concepts within textual conditions. + +# 4.1. Dense Concept Alignment (DCA) + +Given a textual description $\mathcal{T}$ as the condition, our goal is to guide the synthesized image to faithfully reflect its semantic cues. Current approaches [43] often regard the description $\mathcal{T}$ and predicted clean image $\hat{\mathbf{x}}_0$ as a whole for alignment. However, they fail to match compositional region and word information, resulting in semantic drift. Thus, we propose Dense Concept Align (DCA). As shown in Fig. 2, we decouple $\mathcal{T}$ and $\hat{\mathbf{x}}_0$ into intra-object attributes and inter-object relations, and then align these concepts to establish a fine-grained vision-language correspondence. + +We begin by deriving textual concepts from the description $\mathcal{T}$ . Given an example like "A red car in front of a snowy mountain," we employ scene graph parsing [46] to decompose it into textual attribute concepts (e.g., red car, snowy mountain) and relation concepts (e.g., car in front of mountain). Each concept is individually embedded using the CLIP text encoder [34]. An attribute concept is defined as an adjective-noun pair, where the adjective describes the property (e.g., red, snowy) and the noun specifies the category (e.g., car, mountain). We construct its embedding $\mathbf{w}^{\mathrm{a},\mathrm{c}}$ by concatenating the individual embeddings of the property and category: $\mathbf{w}^{\mathrm{a},\mathrm{c}} = \mathbf{w}^{\mathrm{a}} \oplus \mathbf{w}^{\mathrm{c}}$ . In contrast, a relation concept is treated as a complete phrase and directly encoded to obtain its contextual representation $\mathbf{w}^{\mathrm{r}}$ . This process yields a collection of textual concepts, denoted as $\mathcal{W}$ : + +$$ +\mathcal {W} = \left\{\left\{\mathbf {w} _ {i} ^ {\mathrm {a}, \mathrm {c}} \right\} _ {i = 1} ^ {I}, \left\{\mathbf {w} _ {k} ^ {\mathrm {r}} \right\} _ {k = 1} ^ {K} \right\}, \tag {4} +$$ + +where $I$ and $K$ denote the number of parsed objects and relations, respectively. + +Simultaneously, we utilize visual region parsing to extract visual concepts from image $\hat{\mathbf{x}}_0$ . Concretely, we decouple the corresponding visual region of textual concepts using an open-set detection model [27]. The visual attribute + +concept is represented by the detected bounding box enclosing an individual object, whereas the relation one is defined by the union of two object boxes. Using these bounding boxes, we crop $\hat{\mathbf{x}}_0$ to generate a set of isolated visual regions, which are then independently embedded by CLIP vision encoder[34], forming the visual concept collection $\mathcal{V}$ : + +$$ +\mathcal {V} = \left\{\left\{\mathbf {v} _ {i} ^ {\mathrm {a}, \mathrm {c}} \right\} _ {i = 1} ^ {I}, \left\{\mathbf {v} _ {k} ^ {\mathrm {r}} \right\} _ {k = 1} ^ {K} \right\}, \tag {5} +$$ + +where $\mathbf{v}^{\mathrm{a,c}}$ and $\mathbf{v}^{\mathrm{r}}$ represent the visual embeddings of attribute and relation concepts, respectively. + +After obtaining embeddings of decoupled textual and visual concepts, beyond scene-level, we achieve finer-grained alignment by: + +$$ +\mathcal {L} _ {\mathrm {D C A}} = (1 - \gamma) \mathcal {L} _ {\text {s c e n e}} + \gamma (\mathcal {L} _ {\text {a t t r}} + \mathcal {L} _ {\text {r e l}}), \tag {6} +$$ + +where + +$$ +\mathcal {L} _ {\text {s c e n e}} = \cos \left(\mathbf {w} ^ {\mathrm {s}}, \mathbf {v} ^ {\mathrm {s}}\right), \tag {7} +$$ + +$$ +\mathcal {L} _ {\text {a t t r}} (\mathcal {W}, \mathcal {V}) = \frac {1}{I} \sum_ {i = 1} ^ {I} \cos \left(\mathbf {w} _ {i} ^ {\mathrm {a}, \mathrm {c}}, \mathbf {v} _ {i} ^ {\mathrm {a}, \mathrm {c}}\right), \tag {8} +$$ + +$$ +\mathcal {L} _ {r e l} (\mathcal {W}, \mathcal {V}) = \frac {1}{K} \sum_ {k = 1} ^ {K} \cos \left(\mathbf {w} _ {k} ^ {\mathrm {r}}, \mathbf {v} _ {k} ^ {r}\right), \tag {9} +$$ + +where $\cos (\cdot ,\cdot)$ is the cosine similarity function. $\mathbf{w}^{\mathrm{s}}$ and $\mathbf{v}^{\mathrm{s}}$ are the global embeddings of description $\mathcal{T}$ and approximated image $\hat{\mathbf{x}}_0$ , respectively. We set $\gamma = 0.3$ . + +# 4.2. Dense Geometry Alignment (DGA) + +To guide the synthesized image to resemble the spatial configuration described by the collection of segmentation masks $\mathcal{M}$ or bounding boxes $\mathcal{B}$ , we propose Dense Geometry Alignment (DGA) module, as shown in Fig. 3. Current approaches simply ensure that each object appears in its designated region by enhancing corresponding attention scores [19, 32, 49, 50]. Different from them, our module improves the generation plausibility from the perspectives of the object-wise locations, as well as relative size and distance between different objects. For simplicity, we use the segmentation mask collection $\mathcal{M} = \{\mathbf{m}\}_{n=1}^{N}$ as an example to illustrate this module. + +Given the segmentation mask collection $\hat{\mathcal{M}} = \{\hat{\mathbf{m}}\}_{n=1}^{N}$ predicted from the approximated clean image $\hat{\mathbf{x}}_0$ by leveraging an image segmentation model [44], considering an object with the $n$ -th class, we strive to avoid it overlapping with the other objects and ensure the coverage of the target segmentation mask $\mathbf{m}_n$ . Specifically, we enhance the IoU between the two segmentation masks $\hat{\mathbf{m}}_n$ and $\mathbf{m}_n$ , and minimize the interaction between $\hat{\mathbf{m}}_n$ and other irrelevant masks $\mathcal{M} - \{\mathbf{m}_n\}$ . Hence, the coverage loss function is + +![](images/b7e9dc5b22fb8ef85041214e448406240d682544b3c8213494b44e243807cd27.jpg) +Figure 3. The overall pipeline of our Dense Geometry Alignment (DGA) module. On the basis of the detection information from the approximated image, this module can densely align geometric features between the predicted and reference layouts, focusing on object-level location, size, and distance. + +formulated by: + +$$ +\begin{array}{l} \mathcal {L} _ {c v g} (\hat {\mathcal {M}}, \mathcal {M}) = \sum_ {n = 1} ^ {N} \left(1 - \frac {\hat {\mathbf {m}} _ {n} \cap \mathbf {m} _ {n}}{\hat {\mathbf {m}} _ {n} \cup \mathbf {m} _ {n}}\right) \\ + \sum_ {n _ {1} = 1} ^ {N} \sum_ {n _ {2} = 1} ^ {N} \left(\hat {\mathbf {m}} _ {n _ {1}} \cap \mathbf {m} _ {n _ {2}}\right), \tag {10} \\ \end{array} +$$ + +where for clarity and brevity, when two segmentation masks are concurrently referenced in an equation, we use the subscript $n_1$ and $n_2$ to describe the class of different boxes for distinguishment. + +Since maintaining realistic proportions between different objects is crucial for generating visually plausible images, we also restrict the pairwise size ratios of segmentation masks by the proposed size loss. On the basis of the areas computed for the predicted and target segmentation masks, the size loss is formulated by: + +$$ +\mathcal {L} _ {\text {s i z e}} = \sum_ {n _ {1} = 1} ^ {N} \sum_ {n _ {2} = 1} ^ {N} \left(\frac {\hat {\mathbf {s}} _ {n _ {1}}}{\hat {\mathbf {s}} _ {n _ {2}}} - \frac {\mathbf {s} _ {n _ {1}}}{\mathbf {s} _ {n _ {2}}}\right) ^ {2}, \tag {11} +$$ + +where $\hat{\mathbf{s}}_{n_1}(\hat{\mathbf{s}}_{n_2})$ and $\mathbf{s}_{n_1}(\mathbf{s}_{n_2})$ are areas of the segmentation masks in the collections $\hat{\mathcal{M}}$ and $\mathcal{M}$ , respectively. + +In addition, we enforce the alignment of pairwise distances between multiple objects among the approximated image $\hat{\mathbf{x}}_0$ and the given guidance information $\mathcal{M}$ . In particular, after obtaining centroids of each predicted and target object according to the segmentation mask collections $\hat{\mathcal{M}}$ and $\mathcal{M}$ respectively, we compute the pairwise Euclidean distances within each collection of centroids. By minimizing discrepancy in distances between predicted and target centroid collections, we formulate the distance loss function as: + +$$ +\mathcal {L} _ {\text {d i s t}} = \sum_ {n _ {1} = 1} ^ {N} \sum_ {n _ {1} = 2} ^ {N} \left(\| \hat {\mathbf {c}} _ {n _ {1}} - \hat {\mathbf {c}} _ {n _ {2}} \| _ {2} - \| \mathbf {c} _ {n _ {1}} - \mathbf {c} _ {n _ {2}} \| _ {2}\right), \tag {12} +$$ + +![](images/1cc73fe18bdc4e0ba385e060b9fbb42c83752efb6dc28cda9459fb96722c679a.jpg) +Figure 4. The overall pipeline of our Dense Motion Alignment (DMA) module. Using our densified reference drag flow, this module aims to guide synthesized image in achieving dense alignment of motion information across drag conditions, encompassing pixel displacement, appearance, and semantic aspects. + +where $\hat{\mathbf{c}}_{n_1}(\hat{\mathbf{c}}_{n_2})$ and $\mathbf{c}_{n_1}(\mathbf{c}_{n_2})$ are centroids of the segmentation masks in the collections $\hat{\mathcal{M}}$ and $\mathcal{M}$ , respectively. + +In summary, our final guidance function for DGA can be defined as follows: + +$$ +\mathcal {L} _ {\mathrm {D G A}} = (1 - \lambda) \mathcal {L} _ {c v g} + \lambda \left(\mathcal {L} _ {s i z e} + \mathcal {L} _ {d i s t}\right), \tag {13} +$$ + +where we set $\lambda = 0.25$ + +# 4.3. Dense Motion Alignment (DMA) + +Given a collection of $J$ drag signals $\mathcal{P} = \{(\mathbf{p}_j^{\mathrm{o}},\mathbf{p}_j^{\mathrm{d}})\}_{j = 1}^{J}$ , we aim to move visual contents of each original point $\mathbf{p}_j^{\mathrm{o}}$ to its destination $\mathbf{p}_j^{\mathrm{d}}$ . Recent studies, e.g., DragDiffusion [39] and DragNoise [25], only regulate feature similarity around $\mathbf{p}_j^{\mathrm{o}}$ and $\mathbf{p}_j^{\mathrm{d}}$ by implicit optimization in one selected diffusion step. However, on one hand, sparse drag signals do not provide sufficient object awareness, causing incomplete displacement in foreground. On the other hand, one-step feature optimization is unstable. To enable continuous dragging of designated pixels while preserving visual coherence during diffusion sampling, we propose Dense Motion Alignment (DMA), as shown in Fig. 4. Specifically, we predict reference dense drag flow, and enforce pixel correspondence between an image pair in terms of displacement, appearance, and semantics building on this flow. We define $\hat{\mathbf{x}}_0^{t + 1}$ and $\hat{\mathbf{x}}_0^t$ as original and dragging images, respectively, for each timestep $t$ . + +We generate the reference dense drag flow for each individual drag signal using an object-specific displacement field. In particular, we employ SAM [20] to obtain a segmentation mask collection $\{\mathbf{m}_j^{\mathrm{sam}}\}_{j = 1}^{J}$ , where each mask represents an instance area of $\mathbf{p}_j^{\mathrm{o}}$ . Because image pixels closer to $\mathbf{p}_j^{\mathrm{o}}$ should have larger displacements [48], we use radial basis functions to produce a collection of 2D Gaussian maps $\{\mathbf{g}_j\}_{j = 1}^{J}$ . Each map depicts gradually diminishing drag influence radiating outward from $\mathbf{p}_j^{\mathrm{o}}$ . Based on $\{\mathbf{m}_j^{\mathrm{sam}}\}_{j = 1}^{J}$ and $\{\mathbf{g}_j\}_{j = 1}^{J}$ , our reference flow can be densi + +fied by: + +$$ +\mathbf {u} _ {\hat {\mathbf {x}} _ {0} ^ {t + 1} \rightarrow \hat {\mathbf {x}} _ {0} ^ {t}, j} = \left(\mathbf {p} _ {j} ^ {\mathrm {d}} - \mathbf {p} _ {j} ^ {\mathrm {o}}\right) * \mathbf {m} _ {j} ^ {\text {s a m}} * \mathbf {g} _ {j}, \tag {14} +$$ + +where $\mathbf{u}_{\hat{\mathbf{x}}_0^{t + 1}\rightarrow \hat{\mathbf{x}}_0^t,j}$ is the densified drag flow for $j$ -th signal. + +To rectify discrepancies in pixel displacement direction and magnitude, we enable the estimated optical flow between $\hat{\mathbf{x}}_0^{t + 1}$ and $\hat{\mathbf{x}}_0^t$ to align with its reference flow by minimizing displacement loss: + +$$ +\mathcal {L} _ {\text {d i s p}} = \sum_ {j = 1} ^ {J} \left|\left| \hat {\mathbf {u}} _ {\hat {\mathbf {x}} _ {0} ^ {t + 1} \rightarrow \hat {\mathbf {x}} _ {0} ^ {t}, j} - \mathbf {u} _ {\hat {\mathbf {x}} _ {0} ^ {t + 1} \rightarrow \hat {\mathbf {x}} _ {0} ^ {t}, j} \right|\right| _ {1}, \tag {15} +$$ + +where $\hat{\mathbf{u}}_{\hat{\mathbf{x}}_0^{t + 1}\rightarrow \hat{\mathbf{x}}_0^t,j}$ is an optical flow estimated by a RAFT model [41]. In order to ensure visual coherence, we enforce pixel-level appearance consistency by directly minimizing the differences between the dragging and warped original image, while semantic consistency is preserved by DINO v2 embedding [30]: + +$$ +\mathcal {L} _ {p a p p} = \sum_ {j = 1} ^ {J} \sqrt {\left(\operatorname {w a r p} \left(\hat {\mathbf {x}} _ {0} ^ {t + 1} , \mathbf {u} _ {\hat {\mathbf {x}} _ {0} ^ {t + 1} \rightarrow \hat {\mathbf {x}} _ {0} ^ {t} , j}\right) - \hat {\mathbf {x}} _ {0} ^ {t}\right) ^ {2}}, \tag {16} +$$ + +$$ +\mathcal {L} _ {p s e m} = \sum_ {j = 1} ^ {J} \sqrt {\left(\operatorname {w a r p} \left(\hat {\mathbf {f}} _ {0} ^ {t + 1} , \mathbf {u} _ {\hat {\mathbf {x}} _ {0} ^ {t + 1} \rightarrow \hat {\mathbf {x}} _ {0} ^ {t} , j}\right) - \hat {\mathbf {f}} _ {0} ^ {t}\right) ^ {2}}, \tag {17} +$$ + +where $\hat{\mathbf{f}}_0^{t + 1}$ and $\hat{\mathbf{f}}_0^t$ are semantic embedding calculated by DINO v2 model. For a drag signal, our loss of DMA module is defined as: + +$$ +\mathcal {L} _ {\mathrm {D M A}} = (1 - \eta) \mathcal {L} _ {d i s p} + \eta \left(\mathcal {L} _ {p a p p} + \mathcal {L} _ {p s e m}\right), \tag {18} +$$ + +where we set $\eta = 0.02$ . Like DragDiffusion [39] and Drag-Noise [25], we also update the positions of the original points after each denosing step. + +# 5. Experiments + +# 5.1. Experimental Setups + +This section sequentially describes used datasets and evaluation metrics in our experiments, and implementation details of our approaches. + +Datasets. The datasets used in our study are: (1) COCO [24]. We curate the 400 longest image descriptions from COCO validation set, each ranging from 15 to 50 words. This selection is made to better demonstrate our strength in understanding long-form description; (2) DenseDiffusion [19]. This benchmark contains 250 samples $(2\sim 3$ unique objects of each sample), where each sample has a global image description, a segmentation mask, and local descriptions and class labels for each segment; (3) DrawBench [38]. We use the 'position' subset from it. This subset is composed of 20 samples (2 unique objects of each sample), + +where each sample has a description and some bounding boxes of each object (box coordinates and class labels generated by GPT); and (4) DragBench [39]. This benchmark comprises 205 samples, each of which has a reference image, one or more drag signals, a description, and a mask indicating regions to be modified. + +Evaluation metrics. For text condition, we use Fréchet Inception Distance (FID) [12] and CLIP score [34]. FID quantifies the similarity between the synthesized and real images, while the CLIP score measures the alignment between the synthesized image and the reference description. For text + layout condition, we use IoU and SOA-I [13] (layout prediction is based on YOLOv7-Seg [44]). IoU quantifies the overlap between the predicted layout from the synthesized image and the reference layout. SOA-I measures whether described objects appear in the synthesized image. For text + drag condition, we use Image Fidelity (IF) [17] and Mean Distance (MD) [31]. IF (1-LPIPS score [54]) measures the image quality based on the global similarity between the reference and dragged images. MD assesses how accurately the post-dragging original point align with its corresponding destination point by measuring Euclidean distance between their positions. + +Implementation details. We use $T = 50$ steps of the DDIM scheduler [40] and $T = 28$ steps of the Flow Match Euler Discrete scheduler [10]. The weighting schedules for our proposed DCA, DGA, and DMA modules are set to 60, 25, and 90, respectively. Our guidance is only applied during the first $70\%$ timesteps. + +# 5.2. Image Synthesis Results with Text + +We use Stable Diffusion v1.5 [36] and v3.0 [8] as baseline models to evaluate the effectiveness of our DCA module in aligning concepts within descriptions. Tab. 1 summarizes FID [12] and CLIP scores [34] on long ( $>15$ words) descriptions from COCO dataset [24], our DCA module achieves $12.3\%$ and $2.97\%$ improvements in FID and CLIP scores, respectively, building upon Stable Diffusion v3.0 (the most recent baseline). This gain arises from our module's ability to achieve vision-language alignment in a coarse-to-fine manner, effectively leveraging semantic cues embedded in complex sentences. Besides, our module can be applicable to Stable Diffusion v1.5, showing its adaptability across different baseline versions. As shown in Fig. 5, our approach can properly place the dog on a dog bed next to a rocking chair (top) and position the person's head on a blue backpack (middle), indicating its effectiveness in aligning the object interaction concept. Our approach can also generate two people (bottom), showcasing its advantage in understanding the object's numeric information. + +Table 1. Quantitative comparison under long descriptions from COCO dataset [24]. The evaluation metrics used are FID [12] and CLIPScore [34]. We reimplement all baseline approaches. The best results are highlighted. + +
MethodsVenueFID ↓CLIPScore ↑
ParaDiffusion [47]arXiv202312.960.2731
PIXART-α [5]ICLR202413.970.2724
SDXL [33]ICLR20249.790.2820
ELLA [16]arXiv202413.640.2811
PIXART-Σ [4]ECCV202410.730.2793
Stable Diffusion v1.5 [36]CVPR202211.490.2750
+ DCA-10.290.2882
Stable Diffusion v3.0 [8]ICML202410.510.2861
+ DCA-9.220.2946
+ +
DescriptionSD v3.0Ours
A room with filtered light coming through vertical blinds on a doorwall, with a sleeping dog lying on a dog bed next to a rocking chair.
A person covered with an orange sleeping bag rests his head on a blue back pack as he sleeps on a park bench.
In the desert, two people are sitting next to a small propeller plane next to a tent with other planes in the background.
+ +Figure 5. Visualization comparison between our approach (Stable Diffusion v3.0 [8] + DCA) and its baseline under varying descriptions. Our approach demonstrates superior performance in aligning semantic concepts within descriptions. + +# 5.3. Image Synthesis Results with Layout and Text + +We use Stable Diffusion v1.5 [36], DenseDiffusion [19] and A&R [32] as baseline models to evaluate the effectiveness of our DGA module in aligning spatial configurations represented by segmentation masks and bounding boxes; and our DCA module in aligning geometric concepts within descriptions. Tab. 2 summarizes IoU and SOA-I scores on mask-to-image. Our DGA module brings a $27.1\%$ improvement in IoU and a $7.55\%$ increase in SOA-I over A&R baseline, owing to the establishment of object-wise geometric consistency. Besides, incorporating our DCA module on top of DGA leads to an additional improvement, achieving an IoU of 50.61 and an SOA-I of 86.33. This improvement is due to the ability of DCA module to understand descriptions of positions and shapes. We can also observe consistent performance gains across both Stable Diffusion v1.5 and DenseDiffusion architectures. As shown in Fig. 6, our + +Table 2. Quantitative comparison under descriptions and segmentation masks provided by DenseDiffusion [19]. The evaluation metrics are IoU and SOA-I [13]. We generate four random images of each condition combination for performance evaluation. The best results are highlighted. + +
MethodsVenueIoU ↑SOA-I ↑
Real Images-59.1393.07
Composable Diffusion [26]ECCV2022-55.88±0.78
Structure Diffusion [9]ICLR2023-70.97±1.08
SD-Pww [1]arXiv202223.76±0.5073.92±1.84
Stable Diffusion [36]CVPR2022-73.08±1.54
+ DGA-25.46±1.0274.45±1.46
+ DGA + DCA-27.28±0.9777.02±1.67
Dense Diffusion [19]ICCV202334.99±1.1377.61±1.75
+ DGA-36.72±1.2478.21±1.62
+ DGA + DCA-38.06±1.2678.42±1.17
A&R [32]CVPR202438.97±0.5678.80±1.27
+ DGA-49.52±0.7084.75±2.46
+ DGA + DCA-50.61±0.4786.33±0.88
+ +
DescriptionLayoutA&ROurs
An orange tabby cat lying in a bathroom sink.
An orange cat laying on top of a black car.
A dog sitting on a bed wearing an orange bowtie.
+ +Figure 6. Visualization comparison between our approach (A&R [32] + DGA + DCA) and its baseline, under varying descriptions and segmentation masks. Our approach can adhere to both geometric concepts and positional configuration. + +approach can accurately place a cat inside a bathroom sink (top), position a cat on top of a black car (middle), generate a bowtie around a dog's neck (bottom), demonstrating the advantage of our framework in awareness of spatial relationships. The same conclusion can be observed in Fig. 7, which presents box-to-image visualizations of validation data from the DrawBench benchmark [38]. + +# 5.4. Image Synthesis Results with Drag and Text + +We use DragDiffusion [39] as a baseline model to evaluate the effectiveness of our DMA module in motion alignment; and our DCA module for semantic preservation. Fig. 8 presents MD and IF scores on the DragBench benchmark + +![](images/74fdbdee1a5adb7d2c2d8d258d4052db4675ac9ddd2ad10444f8308cb16385b4.jpg) +Figure 7. Visualization comparison between our approach (A&R [32] + DGA + DCA) and its baseline, under varying descriptions and bounding boxes. Our approach can satisfy both conditions. + +![](images/76a6c76452ae6710e101e547b4acc7db30821ad773e2c86f0ccdf1e123de64e4.jpg) +Figure 8. Quantitative comparison under descriptions and drag signals from DragBench benchmark [39]. The evaluation metrics used are Mean Distance $(\downarrow)$ and Image Fidelity (1-LPIPS) $(\uparrow)$ . Approaches with the best results appear in the upper-left corner. + +[39]. The combination of our DMA and DCA modules achieves better MD and IF scores, indicating that our framework ensures each pixel follows its desired trajectory with fewer visual artifacts. However, applying our DMA module alone results in a slight drop in IF, compared to DragDiffusion. This is because our motion field may cause oversmoothing in detail. Fortunately, our DCA module can mitigate such an issue to some degree, by maintaining the overall semantic structure unchanged during the diffusion sampling process. As shown in Fig. 9, our framework can effectively synthesize head rotation (top), rigid car displacement (middle), and large-scale deformation of the bus (bottom). This demonstrates that our algorithm enables the desired movement even when handling complex drag signals + +![](images/9e5fa1bdbb06071f4fbf2fe3b9301c50d9bb040c619eff466ca74372d97b98b8.jpg) +Figure 9. Visualization comparison between our approach (DragDiffusion [39] + DMA) and its baseline under varying drag signals. Our approach enables effective synthesis of large-scale movement and deformation. + +(multiple drag points). + +# 6. Conclusion + +We propose a plug-and-play modular design for guiding image synthesis by aligning diverse fundamental condition units (text, layout, drag) in combination. Based on this idea, we devise three dense alignment approaches to independently achieve flexible control over concept, geometry and motion. Comprehensive experiments demonstrate the effectiveness of our framework across diverse conditioning scenarios. However, our framework relies on the capability of its foundation model, e.g., SD model. As shown in Fig. 10, counterfactual scenes is hard to generate, e.g., a river flowing upwards, if SD model is unable to synthesis this scene itself. Future work could explore world knowledge-aware generative paradigms to address this challenge. + +![](images/3f6afa46f4f503cf30cbdde5146486933b25a48ce7b471e2ada878e9bb03a3fb.jpg) +Figure 10. Failure cases. As foundation model lacks world knowledge, our approach (Stable Diffusion v3.0 [8] + DCA) fails to generate counterfactual scenes. + +Acknowledgement: This work was supported by the National Natural Science Foundation of China (No. U23B2013, 62276176). This work was also partly supported by the SICHUAN Provincial Natural Science Foundation (No. 2024NSFJQ0023). + +# References + +[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 7 +[2] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 843-852, 2023. 3 +[3] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 2, 3 +[4] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 2, 7 +[5] Junsong Chen, YU Jincheng, GE Chongjian, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In Proceedings of the International Conference on Learning Representations (ICLR), pages 1–30, 2024. 2, 7 +[6] Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xinggang Wang, and Ying Shan. Yolo-world: Real-time open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16901-16911, 2024. 3 +[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems (NeruIPS), 34:8780-8794, 2021. 3 +[8] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning (ICML), 2024. 2, 6, 7, 8 +[9] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. In Proceedings of the International Conference on Learning Representations (ICLR), pages 1-13, 2023. 7 +[10] Itai Gat, Tal Remez, Neta Shaul, Felix Kreuk, Ricky TQ Chen, Gabriel Synnaeve, Yossi Adi, and Yaron Lipman. Discrete flow matching. Advances in Neural Information Processing Systems (NeurIPS), 37:133345-133385, 2024. 6 +[11] Daniel Geng and Andrew Owens. Motion guidance: Diffusion-based image editing with differentiable motion estimators. In The Twelfth International Conference on Learning Representations (ICLR), pages 1-21. 3 +[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a + +two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 6, 7 +[13] Tobias Hinz, Stefan Heinrich, and Stefan Wermter. Semantic object accuracy for generative text-to-image synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(3):1552-1565, 2020. 6, 7 +[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NeurIPS), 33:6840-6851, 2020. 3 +[15] Xingzhong Hou, Boxiao Liu, Yi Zhang, Jihao Liu, Yu Liu, and Haihang You. Easydrag: Efficient point-based manipulation on diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8404-8413, 2024. 2 +[16] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135, 2024. 7 +[17] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6007-6017, 2023. 6 +[18] Seung Wook Kim, Bradley Brown, Kangxue Yin, Karsten Kreis, Katja Schwarz, Daiqing Li, Robin Rombach, Antonio Torralba, and Sanja Fidler. Neuralfield-ldm: Scene generation with hierarchical latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8496-8506, 2023. 1 +[19] Yunji Kim, Jiyoung Lee, Jin-Hwa Kim, Jung-Woo Ha, and Jun-Yan Zhu. Dense text-to-image generation with attention modulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7701-7711, 2023. 1, 2, 3, 4, 6, 7 +[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4015-4026, 2023. 5 +[21] Yinjie Lei, Zixuan Wang, Feng Chen, Guoqing Wang, Peng Wang, and Yang Yang. Recent advances in multi-modal 3d scene understanding: A comprehensive survey and evaluation. arXiv preprint arXiv:2310.15676, 2023. 1 +[22] Ming Li, Taojiannan Yang, Huafeng Kuang, Jie Wu, Zhaoning Wang, Xuefeng Xiao, and Chen Chen. Controlnet++: Improving conditional controls with efficient consistency feedback. In European Conference on Computer Vision (ECCV), pages 129-147. Springer, 2025. 3 +[23] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22511-22521, 2023. 2 + +[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision (ECCV), pages 740-755. Springer, 2014. 2, 6, 7 +[25] Haofeng Liu, Chenshu Xu, Yifei Yang, Lihua Zeng, and Shengfeng He. Drag your noise: Interactive point-based editing via diffusion semantic propagation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6743-6752, 2024. 3, 5, 6 +[26] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In Proceedings of the European Conference on Computer Vision (ECCV), pages 423-439. Springer, 2022. 7 +[27] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision (ECCV), pages 38-55. Springer, 2024. 4 +[28] Sicheng Mo, Fangzhou Mu, Kuan Heng Lin, Yanli Liu, Bochen Guan, Yin Li, and Bolei Zhou. Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7465-7475, 2024. 1 +[29] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 4296-4304, 2024. 1, 3 +[30] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6 +[31] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 6 +[32] Quynh Phung, Songwei Ge, and Jia-Bin Huang. Grounded text-to-image synthesis with attention refocusing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7932-7942, 2024. 3, 4, 7, 8 +[33] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In Proceedings of the International Conference on Learning Representations (ICLR), pages 1-13, 2024. 2, 7 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PMLR, 2021. 2, 3, 4, 6, 7 + +[35] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research (JMLR), 21(140):1-67, 2020. 2 +[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 6, 7 +[37] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6527-6536, 2024. 1 +[38] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems (NeurIPS), 35:36479-36494, 2022. 2, 6, 7 +[39] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8839-8849, 2024. 2, 3, 5, 6, 7, 8 +[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations (ICLR), pages 1-22, 2021. 3, 6 +[41] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European Conference on Computer Vision (ECCV), pages 402-419. Springer, 2020. 3, 6 +[42] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1921-1930, 2023. 1 +[43] Bram Wallace, Akash Gokul, Stefano Ermon, and Nikhil Naik. End-to-end diffusion latent optimization improves classifier guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR), pages 7280-7290, 2023. 3, 4 +[44] Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7464-7475, 2023. 4, 6 +[45] Yang Wang, Zhikui Ouyang, Runhua Han, Zhijian Yin, and Zhen Yang. Yolomask: Real-time instance segmentation with integrating yolov5 and orienmask. In IEEE International Conference on Communication Technology (ICCT), pages 1646-1650. IEEE, 2022. 3 + +[46] Hao Wu, Jiayuan Mao, Yufeng Zhang, Yuning Jiang, Lei Li, Weiwei Sun, and Wei-Ying Ma. Unified visual-semantic embeddings: Bridging vision and language with structured meaning representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6609–6618, 2019. 4 +[47] Weijia Wu, Zhuang Li, Yefei He, Mike Zheng Shou, Chunhua Shen, Lele Cheng, Yan Li, Tingting Gao, Di Zhang, and Zhongyuan Wang. Paragraph-to-image generation with information-enriched diffusion model. arXiv preprint arXiv:2311.14284, 2023. 7 +[48] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control for anything using entity representation. In European Conference on Computer Vision (ECCV), pages 331-348. Springer, 2025. 5 +[49] Jiayu Xiao, Henglei Lv, Liang Li, Shuhui Wang, and Qingming Huang. R&b: Region and boundary aware zero-shot grounded text-to-image generation. arXiv preprint arXiv:2310.08872, 2023. 3, 4 +[50] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7452-7461, 2023. 2, 3, 4 + +[51] Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and Jian Zhang. Freedom: Training-free energy-guided conditional diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 23174-23184, 2023. 3 +[52] Yu Zeng, Vishal M Patel, Haochen Wang, Xun Huang, Ting-Chun Wang, Ming-Yu Liu, and Yogesh Balaji. Jedi: Joint-image diffusion models for finetuning-free personalized text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6786-6795, 2024. 1 +[53] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR), pages 3836-3847, 2023. 3 +[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 6 +[55] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. Advances in Neural Information Processing Systems (NeurIPS), 36, 2024. 2, 3 \ No newline at end of file diff --git a/2025/Training-free Dense-Aligned Diffusion Guidance for Modular Conditional Image Synthesis/images.zip b/2025/Training-free Dense-Aligned Diffusion Guidance for Modular Conditional Image Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..319459fdc4a23a6800c68f47cb69bf0e097e29b1 --- /dev/null +++ b/2025/Training-free Dense-Aligned Diffusion Guidance for Modular Conditional Image Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8b4803dcd809fa1e9f3ddf837a785a43cc8f0ca5ac853b25af97efb1a5b0651 +size 714125 diff --git a/2025/Training-free Dense-Aligned Diffusion Guidance for Modular Conditional Image Synthesis/layout.json b/2025/Training-free Dense-Aligned Diffusion Guidance for Modular Conditional Image Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4293202edf52ec5bf26d9d7708b788dadae677f9 --- /dev/null +++ b/2025/Training-free Dense-Aligned Diffusion Guidance for Modular Conditional Image Synthesis/layout.json @@ -0,0 +1,9385 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 102, + 555, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 102, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 555, + 140 + ], + "type": "text", + "content": "Training-free Dense-Aligned Diffusion Guidance for Modular Conditional Image Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 120, + 160, + 488, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 160, + 488, + 175 + ], + "spans": [ + { + "bbox": [ + 120, + 160, + 488, + 175 + ], + "type": "text", + "content": "Zixuan Wang" + }, + { + "bbox": [ + 120, + 160, + 488, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 120, + 160, + 488, + 175 + ], + "type": "text", + "content": " Duo Peng" + }, + { + "bbox": [ + 120, + 160, + 488, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 120, + 160, + 488, + 175 + ], + "type": "text", + "content": " Feng Chen" + }, + { + "bbox": [ + 120, + 160, + 488, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 120, + 160, + 488, + 175 + ], + "type": "text", + "content": " Yuwei Yang" + }, + { + "bbox": [ + 120, + 160, + 488, + 175 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 120, + 160, + 488, + 175 + ], + "type": "text", + "content": " Yinjie Lei" + }, + { + "bbox": [ + 120, + 160, + 488, + 175 + ], + "type": "inline_equation", + "content": "^{1,\\text{图}}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 175, + 480, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 175, + 480, + 190 + ], + "spans": [ + { + "bbox": [ + 132, + 175, + 480, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 175, + 480, + 190 + ], + "type": "text", + "content": "Sichuan University " + }, + { + "bbox": [ + 132, + 175, + 480, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 175, + 480, + 190 + ], + "type": "text", + "content": "Singapore University of Technology and Design" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 190, + 458, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 190, + 458, + 204 + ], + "spans": [ + { + "bbox": [ + 152, + 190, + 458, + 204 + ], + "type": "text", + "content": "3The University of Adelaide 4Australian National University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 100, + 205, + 511, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 205, + 511, + 217 + ], + "spans": [ + { + "bbox": [ + 100, + 205, + 511, + 217 + ], + "type": "text", + "content": "zixuan98@stu.scu.edu.cn, duo_peng@mymail.sutd.edu.sg, chenfeng1271@gmail.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 194, + 220, + 413, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 220, + 413, + 232 + ], + "spans": [ + { + "bbox": [ + 194, + 220, + 413, + 232 + ], + "type": "text", + "content": "yuwei.yang@anu.edu.au, yinjie@scu.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 283, + 296, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 283, + 296, + 620 + ], + "spans": [ + { + "bbox": [ + 53, + 283, + 296, + 620 + ], + "type": "text", + "content": "Conditional image synthesis is a crucial task with broad applications, such as artistic creation and virtual reality. However, current generative methods are often task-oriented with a narrow scope, handling a restricted condition with constrained applicability. In this paper, we propose a novel approach that treats conditional image synthesis as the modular combination of diverse fundamental condition units. Specifically, we divide conditions into three primary units: text, layout, and drag. To enable effective control over these conditions, we design a dedicated alignment module for each. For the text condition, we introduce a Dense Concept Alignment (DCA) module, which achieves dense visual-text alignment by drawing on diverse textual concepts. For the layout condition, we propose a Dense Geometry Alignment (DGA) module to enforce comprehensive geometric constraints that preserve the spatial configuration. For the drag condition, we introduce a Dense Motion Alignment (DMA) module to apply multi-level motion regularization, ensuring that each pixel follows its desired trajectory without visual artifacts. By flexibly inserting and combining these alignment modules, our framework enhances the model's adaptability to diverse conditional generation tasks and greatly expands its application range. Extensive experiments demonstrate the superior performance of our framework across a variety of conditions, including textual description, segmentation mask (bounding box), drag manipulation, and their combinations. Code is available at https://github.com/ZixuanWang0525/DADG" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 640, + 137, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 640, + 137, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 640, + 137, + 653 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 661, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 661, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 661, + 296, + 696 + ], + "type": "text", + "content": "Conditional image synthesis aims to generate realistic images based on user-provided requirements, playing a pivotal role in various applications, such as artistic creation [37, 52]" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 258, + 553, + 348 + ], + "blocks": [ + { + "bbox": [ + 317, + 258, + 553, + 348 + ], + "lines": [ + { + "bbox": [ + 317, + 258, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 258, + 553, + 348 + ], + "type": "image", + "image_path": "86d7c1b6ed59d84f9a5722e33444e2ce4e186fe46095c8ae16b6e7ba45f47b81.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 350, + 553, + 378 + ], + "lines": [ + { + "bbox": [ + 315, + 350, + 553, + 378 + ], + "spans": [ + { + "bbox": [ + 315, + 350, + 553, + 378 + ], + "type": "text", + "content": "(a) General pipeline of our proposed modular conditional image synthesis. Overall framework of Training-free Dense-Aligned Diffusion Guidance is illustrated in Fig. 1b." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 315, + 380, + 552, + 597 + ], + "blocks": [ + { + "bbox": [ + 315, + 380, + 552, + 597 + ], + "lines": [ + { + "bbox": [ + 315, + 380, + 552, + 597 + ], + "spans": [ + { + "bbox": [ + 315, + 380, + 552, + 597 + ], + "type": "image", + "image_path": "de84523cf2da69fef3c48ef29666c32ca12caad66d1e385707236392a6cf5d95.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 600, + 552, + 628 + ], + "lines": [ + { + "bbox": [ + 313, + 600, + 552, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 552, + 628 + ], + "type": "text", + "content": "(b) Illustration of our proposed Training-free Dense-Aligned Diffusion Guidance. We design three dense alignment modules to enable an effective accommodation of each specific condition unit." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 313, + 629, + 554, + 662 + ], + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 662 + ], + "type": "text", + "content": "Figure 1. The overview of our modular conditional image synthesis. Driven by modular combination of fundamental condition units, diverse visual content can be synthesized." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "and virtual reality [18, 21]. Recent advances in diffusion models have demonstrated promising performance in handling diverse conditional signals, including image descriptions [36, 42], segmentation maps [19, 28, 29], bounding" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 702, + 257, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 702, + 257, + 713 + ], + "spans": [ + { + "bbox": [ + 69, + 702, + 257, + 713 + ], + "type": "text", + "content": "Corresponding Author: Yinjie Lei (yinjie@scu.edu.cn)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "text", + "content": "13135" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "boxes [3, 50, 55], and drag information [15, 39], among others. While existing approaches are effective in interpreting a particular form of input from users, they struggle to generalize across a wide range of conditions. This is impractical for flexible generation of complex visual scenes in the real world." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 147, + 296, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 147, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 147, + 296, + 326 + ], + "type": "text", + "content": "In this paper, we propose the Modular Conditional Image Synthesis (MCIS) paradigm, which regulates the synthesis process by selectively applying and combining independent fundamental condition units, as demonstrated in Fig. 1 (a). Specifically, we categorize condition units into: (1) Text. This category refers to natural language descriptions, specifying the semantic content of an image; (2) Layout. This category mainly encompasses segmentation maps and bounding boxes, representing the component arrangement and scene structure; and (3) Drag. This category refers to point pairs (original and destination of drags), encoding the local transformations of an image. However, MCIS presents a non-trivial problem, as it requires ensuring that all conditions are distinctly reflected in the visual content without omission." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 328, + 296, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 328, + 296, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 328, + 296, + 532 + ], + "type": "text", + "content": "We sequentially analyze the key challenges involved in aligning each basic condition unit: (1) Concept Mismatching. Text typically conveys detailed information about the properties of individual objects and their interactions. This requires generative models to match fine-grained region-level and word-level concepts, instead of considering the image and its description as a whole for alignment; (2) Geometry Inconsistency. Layout implicitly encodes complex relationships among multiple objects, such as size ratios and relative positions. Ensuring image realism requires establishing a detailed geometric correspondence between synthesized content and spatial configuration; and (3) Motion Disharmony. Drag often provides displacement vectors indicating the visual regions to be moved. To synthesize visually coherent content, it is crucial to effectively maintain the intended movement of each visual element, while preserving both appearance and semantic consistency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "To address the above issues, we propose a dense-aligned diffusion guidance framework for modular conditional image synthesis. As illustrated in Fig. 1 (b), our framework utilizes plug-and-play guidance modules to allow visual content to independently align with each of the condition units. Specifically, we propose the Dense Concept Alignment (DCA) module, which ensures the consistency between visual content and scene description in a coarse-to-fine manner. Beyond scene-level alignment, we establish fine-grained correspondence from attribute and relational perspectives within a disentangled feature space. We further introduce the Dense Geometry Alignment (DGA) module. Building upon the detection information from the synthesized image, it imposes constraints on object-wise locations, as well as on relative sizes and distances between" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "each pair of objects. This process ensures the separation between different foreground instances while maintaining the realism of scene arrangement. Based on dense displacement fields, we design the Dense Motion Alignment (DMA) module. During diffusion sampling, this module guides the variation of visual content between adjacent timesteps by aligning it with the displacement fields. Meanwhile, it uses pixel-wise color and semantic regularization to achieve photorealistic consistency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 180, + 553, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 298 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 298 + ], + "type": "text", + "content": "To evaluate the effectiveness of our framework in MCIS, we conduct extensive experiments on public benchmarks, including COCO [24], DenseDiffusion [19], DrawBench [38], and DragBench [39]. Quantitative and qualitative results demonstrate that our approach significantly improves the adherence to the text, layout, and drag conditions. Moreover, our guidance modules are compatible to simultaneously control over multiple condition units. In addition, our framework can be seamlessly integrated into various diffusion architectures." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 325, + 300, + 511, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 300, + 511, + 310 + ], + "spans": [ + { + "bbox": [ + 325, + 300, + 511, + 310 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 324, + 312, + 553, + 430 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 324, + 312, + 553, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 312, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 324, + 312, + 553, + 357 + ], + "type": "text", + "content": "- We propose a paradigm of modular conditional image synthesis, which expands the application scope of diffusion models by the adaptive combination of fundamental condition units." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 324, + 360, + 553, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 360, + 553, + 394 + ], + "spans": [ + { + "bbox": [ + 324, + 360, + 553, + 394 + ], + "type": "text", + "content": "- We design three plug-and-play dense alignment modules, DCA, DGA, and DMA, to independently achieve effective control over concept, geometry, and motion." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 324, + 396, + 552, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 396, + 552, + 430 + ], + "spans": [ + { + "bbox": [ + 324, + 396, + 552, + 430 + ], + "type": "text", + "content": "- Extensive experiments show that our framework significantly enhances controllability across text, layout, drag, and their combinations." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 441, + 406, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 441, + 406, + 453 + ], + "spans": [ + { + "bbox": [ + 314, + 441, + 406, + 453 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 461, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 555, + 568 + ], + "type": "text", + "content": "Conditional image synthesis is a core research field in AIGC, which aims to generate images explicitly guided by input conditions. From the aspect of framework architecture, current studies can be divided into: (1) Training-based and (2) Training-free condition image synthesis. Crucially, we classify diverse conditions into three fundamental units: text, layout, and drag. Training-based and -free frameworks are, thus, sequentially analyzed in a context of fundamental condition units we define." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 576, + 544, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 544, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 544, + 589 + ], + "type": "text", + "content": "2.1. Training-based Conditional Image Synthesis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "Training-based frameworks often train diffusion models from scratch or lightweight auxiliary modules to accommodate user-provided conditions. Text-driven frameworks, e.g., Stable Diffusion family [8, 33, 36] and Pix-Art family [4, 5], optimize diffusion models from scratch to implicitly aware of embeddings from frozen large-language models, such as CLIP [34] and T5 [35]. Layout-driven approaches incorporate parameter-efficient modules to encode segmentation masks, bounding boxes, and depth maps, among others. For instance, GLIGEN [23] uses a learn-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13136" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "type": "text", + "content": "able gated self-attention module to capture grounding information, i.e., regional bounding boxes and its corresponding phrase. ControlNet family [22, 53, 55] can be viewed as a learnable copy of a U-net encoder, which reuses an off-the-shelf diffusion model as a backbone for handling diverse spatial conditioning vectors. The copy encoder is connected with U-net decoder layers in a frozen diffusion model by zero-convolution modules. Instead of a Siamese network of encoder layers, T2I-Adapter [29] learns various low-complexity adapters according to different conditions, where each adapter is only composed of some residual convolution blocks. However, the demands for high-quality data and considerable computational resources remain the bottlenecks for scalability and widespread adoption of the above approaches." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 258, + 278, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 278, + 272 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 278, + 272 + ], + "type": "text", + "content": "2.2. Training-free Conditional Image Synthesis" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 276, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 296, + 601 + ], + "type": "text", + "content": "Training-free works focus on modifying the original diffusion sampling process to drive the conditional image synthesis using feedback from the guidance function. For text condition, mainstream guidance frameworks are classifier and CLIP guidance. Specifically, classifier guidance [7] incorporates class information into diffusion models using gradients of a timestep-aware image classifier. CLIP guidance, such as DOODL [43], replaces classification scores with global CLIP scores, which can make generation process accommodate free-form descriptions. Regarding layout condition, current approaches often implicitly modify distributions of attention scores to achieve guidance. The underlying assumption is: spatial distribution of high-response attention aligns perceptually with the locations of objects in synthesized images [3]. Based on this idea, BoxDiff [50] and R&B [49] impose region and boundary constraints to stored cross-attention maps. DenseDiffusion [19] and A&R [32] further leverage semantic affinity in self-attention maps to ensure more accurate object placement. With respect to drag condition, DragDiffusion [39] and DragNoise [25] employ motion supervision and point tracking steps to optimize diffusion semantic features. These steps enable features of original points to move to destination points as effectively as possible. However, aforementioned approaches are often designed to encode a particular condition form, which restricts their generalization across a wide range of conditions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 609, + 133, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 133, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 133, + 624 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 629, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 666 + ], + "type": "text", + "content": "This section provides a brief overview of the core mechanisms in the conditional diffusion model framework, including diffusion sampling and diffusion guidance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "Diffusion sampling. Diffusion sampling is designed to generate a clean image by progressively removing noise from a random Gaussian noise image " + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": " [14, 51]. This sampling process is accomplished by a learned denoising" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "neural network " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " which is conditioned on a given prompt " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": ". Concretely, at each time step " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": ", a cleaner image " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t-1}" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " can be generated by subtracting the predicted noise " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_t = \\epsilon_{\\theta}(\\mathbf{x}_t; t, \\mathbf{y})" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " from the current noise image " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": ". As an example of denoising diffusion implicit model (DDIM) [40], " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t-1}" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " can be produced via:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 361, + 155, + 555, + 169 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 155, + 555, + 169 + ], + "spans": [ + { + "bbox": [ + 361, + 155, + 555, + 169 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {t - 1} = \\sqrt {\\alpha_ {t - 1}} \\hat {\\mathbf {x}} _ {0} + \\sqrt {1 - \\alpha_ {t - 1}} \\hat {\\epsilon} _ {t}, \\tag {1}", + "image_path": "57513eb8a7c9129dd34b9b05f6b4456b9b58b72cfca68fe8e5530b097e9d5537.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 174, + 342, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 174, + 342, + 183 + ], + "spans": [ + { + "bbox": [ + 314, + 174, + 342, + 183 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 385, + 182, + 555, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 182, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 385, + 182, + 555, + 209 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} _ {0} = \\frac {\\mathbf {x} _ {t} - \\sqrt {1 - \\alpha_ {t}} \\hat {\\boldsymbol {\\epsilon}} _ {t}}{\\sqrt {\\alpha_ {t}}}, \\tag {2}", + "image_path": "367ce40bfb4e0765b3c3d376099d5c1fa067c78d9e41576fcfa5d333616e6453.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 213, + 555, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 248 + ], + "type": "text", + "content": "where scalar " + }, + { + "bbox": [ + 313, + 213, + 555, + 248 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 313, + 213, + 555, + 248 + ], + "type": "text", + "content": " governs denoising intensity at each timestep " + }, + { + "bbox": [ + 313, + 213, + 555, + 248 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 213, + 555, + 248 + ], + "type": "text", + "content": ". From Eq. 2, the approximation of the clean image " + }, + { + "bbox": [ + 313, + 213, + 555, + 248 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0" + }, + { + "bbox": [ + 313, + 213, + 555, + 248 + ], + "type": "text", + "content": " can be predicted by performing one-step diffusion sampling." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 248, + 554, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 554, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 554, + 332 + ], + "type": "text", + "content": "Diffusion guidance. The purpose of diffusion guidance is to produce a more reasonable denoising direction based on auxiliary information " + }, + { + "bbox": [ + 313, + 248, + 554, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{y}'" + }, + { + "bbox": [ + 313, + 248, + 554, + 332 + ], + "type": "text", + "content": " during inference time [2]. Specifically, at each timestep " + }, + { + "bbox": [ + 313, + 248, + 554, + 332 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 248, + 554, + 332 + ], + "type": "text", + "content": ", diffusion guidance enables the denoised image " + }, + { + "bbox": [ + 313, + 248, + 554, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t-1}" + }, + { + "bbox": [ + 313, + 248, + 554, + 332 + ], + "type": "text", + "content": " follow the auxiliary information " + }, + { + "bbox": [ + 313, + 248, + 554, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{y}'" + }, + { + "bbox": [ + 313, + 248, + 554, + 332 + ], + "type": "text", + "content": " by modifying the predicted noise " + }, + { + "bbox": [ + 313, + 248, + 554, + 332 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_t" + }, + { + "bbox": [ + 313, + 248, + 554, + 332 + ], + "type": "text", + "content": " using the gradient of the energy function " + }, + { + "bbox": [ + 313, + 248, + 554, + 332 + ], + "type": "inline_equation", + "content": "g(\\mathbf{x}_t; t, \\mathbf{y}')" + }, + { + "bbox": [ + 313, + 248, + 554, + 332 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 375, + 339, + 555, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 339, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 375, + 339, + 555, + 354 + ], + "type": "interline_equation", + "content": "\\tilde {\\epsilon} _ {t} = \\hat {\\epsilon} _ {t} + \\sigma_ {t} \\nabla_ {\\mathbf {x} _ {t}} g (\\mathbf {x} _ {t}; t, \\mathbf {y} ^ {\\prime}), \\qquad (3)", + "image_path": "0c29aa55dcc2cd52ae0e178bf8f046b0fb5b5b7360e2d0ea52480aa08bc278d5.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 361, + 451, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 451, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 451, + 372 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 361, + 451, + 372 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 313, + 361, + 451, + 372 + ], + "type": "text", + "content": " is a weighting schedule." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 373, + 555, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 529 + ], + "type": "text", + "content": "In practice, the function " + }, + { + "bbox": [ + 313, + 373, + 555, + 529 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 373, + 555, + 529 + ], + "type": "text", + "content": " could be the approximate energy from any differentiable model, such as probabilities from a classifier [7], similarity scores from a vision-language model [34], position penalties from a detection [6] or segmentation model [45], and motion consistencies from a flow estimator [41]. Recent studies [2, 11] have shown that the predicted clean image " + }, + { + "bbox": [ + 313, + 373, + 555, + 529 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0" + }, + { + "bbox": [ + 313, + 373, + 555, + 529 + ], + "type": "text", + "content": " (See Eq. 2) can be straightforwardly used as the input for the energy function " + }, + { + "bbox": [ + 313, + 373, + 555, + 529 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 373, + 555, + 529 + ], + "type": "text", + "content": ", demonstrating that we do not require to learn an additional timestep-dependent differentiable model. This paper also computes the energy based on a one step approximate of the clean image " + }, + { + "bbox": [ + 313, + 373, + 555, + 529 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0" + }, + { + "bbox": [ + 313, + 373, + 555, + 529 + ], + "type": "text", + "content": " to produce the gradient " + }, + { + "bbox": [ + 313, + 373, + 555, + 529 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}_t}g(\\hat{\\mathbf{x}}_0;\\mathbf{y}')" + }, + { + "bbox": [ + 313, + 373, + 555, + 529 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 538, + 397, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 397, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 397, + 552 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "Our framework is designed for modular conditional image synthesis, generating diverse visual scenarios by flexibly combining three fundamental condition units: text, layout, and drag. Accordingly, it contains three diffusion guidance modules specifically designed for each unit. To clarify each module, we organize the methodology section as follows: in Sec. 4.1, we design a Dense Concept Alignment (DCA) module to make synthesized images follow semantic cues embedded in texts; in Sec. 4.2, we present a Dense Geometry Alignment (DGA) to enforce adherence to spatial configurations within layouts; in Sec. 4.3, we propose a Dense Motion Alignment (DMA) to allow diffusion model to be aware of movement and deformation." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13137" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 69, + 294, + 216 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 294, + 216 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 294, + 216 + ], + "type": "image", + "image_path": "e431493c89228e9a530ee47ff5fc6a0ed1ef9a8eda0aaea15e4c1e13ed04de88.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 223, + 295, + 268 + ], + "lines": [ + { + "bbox": [ + 55, + 223, + 295, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 295, + 268 + ], + "type": "text", + "content": "Figure 2. The overall pipeline of our Dense Concept Alignment (DCA) module. Beyond scene-level vision-language consistency, this module can guide synthesized image in achieving dense alignment of attribute and relation concepts within textual conditions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 289, + 234, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 234, + 303 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 234, + 303 + ], + "type": "text", + "content": "4.1. Dense Concept Alignment (DCA)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 307, + 296, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 307, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 296, + 426 + ], + "type": "text", + "content": "Given a textual description " + }, + { + "bbox": [ + 55, + 307, + 296, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 307, + 296, + 426 + ], + "type": "text", + "content": " as the condition, our goal is to guide the synthesized image to faithfully reflect its semantic cues. Current approaches [43] often regard the description " + }, + { + "bbox": [ + 55, + 307, + 296, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 307, + 296, + 426 + ], + "type": "text", + "content": " and predicted clean image " + }, + { + "bbox": [ + 55, + 307, + 296, + 426 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0" + }, + { + "bbox": [ + 55, + 307, + 296, + 426 + ], + "type": "text", + "content": " as a whole for alignment. However, they fail to match compositional region and word information, resulting in semantic drift. Thus, we propose Dense Concept Align (DCA). As shown in Fig. 2, we decouple " + }, + { + "bbox": [ + 55, + 307, + 296, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 307, + 296, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 307, + 296, + 426 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0" + }, + { + "bbox": [ + 55, + 307, + 296, + 426 + ], + "type": "text", + "content": " into intra-object attributes and inter-object relations, and then align these concepts to establish a fine-grained vision-language correspondence." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 427, + 296, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 427, + 296, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 296, + 606 + ], + "type": "text", + "content": "We begin by deriving textual concepts from the description " + }, + { + "bbox": [ + 55, + 427, + 296, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 427, + 296, + 606 + ], + "type": "text", + "content": ". Given an example like \"A red car in front of a snowy mountain,\" we employ scene graph parsing [46] to decompose it into textual attribute concepts (e.g., red car, snowy mountain) and relation concepts (e.g., car in front of mountain). Each concept is individually embedded using the CLIP text encoder [34]. An attribute concept is defined as an adjective-noun pair, where the adjective describes the property (e.g., red, snowy) and the noun specifies the category (e.g., car, mountain). We construct its embedding " + }, + { + "bbox": [ + 55, + 427, + 296, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^{\\mathrm{a},\\mathrm{c}}" + }, + { + "bbox": [ + 55, + 427, + 296, + 606 + ], + "type": "text", + "content": " by concatenating the individual embeddings of the property and category: " + }, + { + "bbox": [ + 55, + 427, + 296, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^{\\mathrm{a},\\mathrm{c}} = \\mathbf{w}^{\\mathrm{a}} \\oplus \\mathbf{w}^{\\mathrm{c}}" + }, + { + "bbox": [ + 55, + 427, + 296, + 606 + ], + "type": "text", + "content": ". In contrast, a relation concept is treated as a complete phrase and directly encoded to obtain its contextual representation " + }, + { + "bbox": [ + 55, + 427, + 296, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^{\\mathrm{r}}" + }, + { + "bbox": [ + 55, + 427, + 296, + 606 + ], + "type": "text", + "content": ". This process yields a collection of textual concepts, denoted as " + }, + { + "bbox": [ + 55, + 427, + 296, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 55, + 427, + 296, + 606 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 616, + 295, + 631 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 616, + 295, + 631 + ], + "spans": [ + { + "bbox": [ + 113, + 616, + 295, + 631 + ], + "type": "interline_equation", + "content": "\\mathcal {W} = \\left\\{\\left\\{\\mathbf {w} _ {i} ^ {\\mathrm {a}, \\mathrm {c}} \\right\\} _ {i = 1} ^ {I}, \\left\\{\\mathbf {w} _ {k} ^ {\\mathrm {r}} \\right\\} _ {k = 1} ^ {K} \\right\\}, \\tag {4}", + "image_path": "b19e1117e8a5a267fd3e6e09b3ea256d8dac807cd349c4cb0acca333b4cd1526.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 641, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 641, + 295, + 665 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 641, + 295, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 641, + 295, + 665 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 641, + 295, + 665 + ], + "type": "text", + "content": " denote the number of parsed objects and relations, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "Simultaneously, we utilize visual region parsing to extract visual concepts from image " + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0" + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": ". Concretely, we decouple the corresponding visual region of textual concepts using an open-set detection model [27]. The visual attribute" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "concept is represented by the detected bounding box enclosing an individual object, whereas the relation one is defined by the union of two object boxes. Using these bounding boxes, we crop " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " to generate a set of isolated visual regions, which are then independently embedded by CLIP vision encoder[34], forming the visual concept collection " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 375, + 154, + 553, + 169 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 154, + 553, + 169 + ], + "spans": [ + { + "bbox": [ + 375, + 154, + 553, + 169 + ], + "type": "interline_equation", + "content": "\\mathcal {V} = \\left\\{\\left\\{\\mathbf {v} _ {i} ^ {\\mathrm {a}, \\mathrm {c}} \\right\\} _ {i = 1} ^ {I}, \\left\\{\\mathbf {v} _ {k} ^ {\\mathrm {r}} \\right\\} _ {k = 1} ^ {K} \\right\\}, \\tag {5}", + "image_path": "4c058619e954c514d3e6ff3a3d9bf1ea6a602426d17bd78bdc0addef11de7a07.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 179, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 179, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 179, + 553, + 204 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 179, + 553, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^{\\mathrm{a,c}}" + }, + { + "bbox": [ + 313, + 179, + 553, + 204 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 179, + 553, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^{\\mathrm{r}}" + }, + { + "bbox": [ + 313, + 179, + 553, + 204 + ], + "type": "text", + "content": " represent the visual embeddings of attribute and relation concepts, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 204, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 553, + 240 + ], + "type": "text", + "content": "After obtaining embeddings of decoupled textual and visual concepts, beyond scene-level, we achieve finer-grained alignment by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 344, + 252, + 553, + 265 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 252, + 553, + 265 + ], + "spans": [ + { + "bbox": [ + 344, + 252, + 553, + 265 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D C A}} = (1 - \\gamma) \\mathcal {L} _ {\\text {s c e n e}} + \\gamma (\\mathcal {L} _ {\\text {a t t r}} + \\mathcal {L} _ {\\text {r e l}}), \\tag {6}", + "image_path": "3bae7cafdbc6ba249162ee6449a4b883936d8d3ca7448cbf626c6bfee54bf222.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 277, + 342, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 277, + 342, + 286 + ], + "spans": [ + { + "bbox": [ + 314, + 277, + 342, + 286 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 388, + 289, + 553, + 303 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 289, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 388, + 289, + 553, + 303 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s c e n e}} = \\cos \\left(\\mathbf {w} ^ {\\mathrm {s}}, \\mathbf {v} ^ {\\mathrm {s}}\\right), \\tag {7}", + "image_path": "40e3ce41cc7469097e5b5dc2d110a53c9cbae0e78f34e6ce8311be18e2b1ff29.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 356, + 313, + 553, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 313, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 356, + 313, + 553, + 346 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a t t r}} (\\mathcal {W}, \\mathcal {V}) = \\frac {1}{I} \\sum_ {i = 1} ^ {I} \\cos \\left(\\mathbf {w} _ {i} ^ {\\mathrm {a}, \\mathrm {c}}, \\mathbf {v} _ {i} ^ {\\mathrm {a}, \\mathrm {c}}\\right), \\tag {8}", + "image_path": "2ae9bb4d162d7fee26ad13384520b218e1ef6e1f829226c05aca865aa1bf27d8.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 361, + 357, + 553, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 357, + 553, + 389 + ], + "spans": [ + { + "bbox": [ + 361, + 357, + 553, + 389 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e l} (\\mathcal {W}, \\mathcal {V}) = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\cos \\left(\\mathbf {w} _ {k} ^ {\\mathrm {r}}, \\mathbf {v} _ {k} ^ {r}\\right), \\tag {9}", + "image_path": "6bbd66140fc6c67657d30cb97eb1f19579280fcec2778a82e8d80f9478571076.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 396, + 553, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 553, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 553, + 433 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 396, + 553, + 433 + ], + "type": "inline_equation", + "content": "\\cos (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 313, + 396, + 553, + 433 + ], + "type": "text", + "content": " is the cosine similarity function. " + }, + { + "bbox": [ + 313, + 396, + 553, + 433 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^{\\mathrm{s}}" + }, + { + "bbox": [ + 313, + 396, + 553, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 396, + 553, + 433 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^{\\mathrm{s}}" + }, + { + "bbox": [ + 313, + 396, + 553, + 433 + ], + "type": "text", + "content": " are the global embeddings of description " + }, + { + "bbox": [ + 313, + 396, + 553, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 396, + 553, + 433 + ], + "type": "text", + "content": " and approximated image " + }, + { + "bbox": [ + 313, + 396, + 553, + 433 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0" + }, + { + "bbox": [ + 313, + 396, + 553, + 433 + ], + "type": "text", + "content": ", respectively. We set " + }, + { + "bbox": [ + 313, + 396, + 553, + 433 + ], + "type": "inline_equation", + "content": "\\gamma = 0.3" + }, + { + "bbox": [ + 313, + 396, + 553, + 433 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 443, + 501, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 501, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 501, + 456 + ], + "type": "text", + "content": "4.2. Dense Geometry Alignment (DGA)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 461, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 554, + 604 + ], + "type": "text", + "content": "To guide the synthesized image to resemble the spatial configuration described by the collection of segmentation masks " + }, + { + "bbox": [ + 313, + 461, + 554, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 461, + 554, + 604 + ], + "type": "text", + "content": " or bounding boxes " + }, + { + "bbox": [ + 313, + 461, + 554, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 461, + 554, + 604 + ], + "type": "text", + "content": ", we propose Dense Geometry Alignment (DGA) module, as shown in Fig. 3. Current approaches simply ensure that each object appears in its designated region by enhancing corresponding attention scores [19, 32, 49, 50]. Different from them, our module improves the generation plausibility from the perspectives of the object-wise locations, as well as relative size and distance between different objects. For simplicity, we use the segmentation mask collection " + }, + { + "bbox": [ + 313, + 461, + 554, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\{\\mathbf{m}\\}_{n=1}^{N}" + }, + { + "bbox": [ + 313, + 461, + 554, + 604 + ], + "type": "text", + "content": " as an example to illustrate this module." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "Given the segmentation mask collection " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}} = \\{\\hat{\\mathbf{m}}\\}_{n=1}^{N}" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": " predicted from the approximated clean image " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": " by leveraging an image segmentation model [44], considering an object with the " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "-th class, we strive to avoid it overlapping with the other objects and ensure the coverage of the target segmentation mask " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_n" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": ". Specifically, we enhance the IoU between the two segmentation masks " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{m}}_n" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_n" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": ", and minimize the interaction between " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{m}}_n" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": " and other irrelevant masks " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M} - \\{\\mathbf{m}_n\\}" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": ". Hence, the coverage loss function is" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13138" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 298, + 190 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 298, + 190 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 298, + 190 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 298, + 190 + ], + "type": "image", + "image_path": "b7e9dc5b22fb8ef85041214e448406240d682544b3c8213494b44e243807cd27.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 198, + 297, + 253 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 297, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 297, + 253 + ], + "type": "text", + "content": "Figure 3. The overall pipeline of our Dense Geometry Alignment (DGA) module. On the basis of the detection information from the approximated image, this module can densely align geometric features between the predicted and reference layouts, focusing on object-level location, size, and distance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 273, + 118, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 273, + 118, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 118, + 285 + ], + "type": "text", + "content": "formulated by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 291, + 294, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 291, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 97, + 291, + 294, + 360 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {c v g} (\\hat {\\mathcal {M}}, \\mathcal {M}) = \\sum_ {n = 1} ^ {N} \\left(1 - \\frac {\\hat {\\mathbf {m}} _ {n} \\cap \\mathbf {m} _ {n}}{\\hat {\\mathbf {m}} _ {n} \\cup \\mathbf {m} _ {n}}\\right) \\\\ + \\sum_ {n _ {1} = 1} ^ {N} \\sum_ {n _ {2} = 1} ^ {N} \\left(\\hat {\\mathbf {m}} _ {n _ {1}} \\cap \\mathbf {m} _ {n _ {2}}\\right), \\tag {10} \\\\ \\end{array}", + "image_path": "41f8d4b8a28bfd6a713f627aede204fa21acfbd0255404ade5cd42118bf7f6c2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 367, + 296, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 296, + 414 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 296, + 414 + ], + "type": "text", + "content": "where for clarity and brevity, when two segmentation masks are concurrently referenced in an equation, we use the subscript " + }, + { + "bbox": [ + 55, + 367, + 296, + 414 + ], + "type": "inline_equation", + "content": "n_1" + }, + { + "bbox": [ + 55, + 367, + 296, + 414 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 367, + 296, + 414 + ], + "type": "inline_equation", + "content": "n_2" + }, + { + "bbox": [ + 55, + 367, + 296, + 414 + ], + "type": "text", + "content": " to describe the class of different boxes for distinguishment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 415, + 296, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 487 + ], + "type": "text", + "content": "Since maintaining realistic proportions between different objects is crucial for generating visually plausible images, we also restrict the pairwise size ratios of segmentation masks by the proposed size loss. On the basis of the areas computed for the predicted and target segmentation masks, the size loss is formulated by:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 494, + 295, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 494, + 295, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 295, + 528 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s i z e}} = \\sum_ {n _ {1} = 1} ^ {N} \\sum_ {n _ {2} = 1} ^ {N} \\left(\\frac {\\hat {\\mathbf {s}} _ {n _ {1}}}{\\hat {\\mathbf {s}} _ {n _ {2}}} - \\frac {\\mathbf {s} _ {n _ {1}}}{\\mathbf {s} _ {n _ {2}}}\\right) ^ {2}, \\tag {11}", + "image_path": "78462b10d95c0e7409ee7bca40ae47bb4fa831f65cc01f4d2be8eb761f00f472.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 533, + 295, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 557 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 533, + 295, + 557 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{s}}_{n_1}(\\hat{\\mathbf{s}}_{n_2})" + }, + { + "bbox": [ + 55, + 533, + 295, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 533, + 295, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{n_1}(\\mathbf{s}_{n_2})" + }, + { + "bbox": [ + 55, + 533, + 295, + 557 + ], + "type": "text", + "content": " are areas of the segmentation masks in the collections " + }, + { + "bbox": [ + 55, + 533, + 295, + 557 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}" + }, + { + "bbox": [ + 55, + 533, + 295, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 533, + 295, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 533, + 295, + 557 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 558, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 677 + ], + "type": "text", + "content": "In addition, we enforce the alignment of pairwise distances between multiple objects among the approximated image " + }, + { + "bbox": [ + 55, + 558, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0" + }, + { + "bbox": [ + 55, + 558, + 296, + 677 + ], + "type": "text", + "content": " and the given guidance information " + }, + { + "bbox": [ + 55, + 558, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 558, + 296, + 677 + ], + "type": "text", + "content": ". In particular, after obtaining centroids of each predicted and target object according to the segmentation mask collections " + }, + { + "bbox": [ + 55, + 558, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}" + }, + { + "bbox": [ + 55, + 558, + 296, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 558, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 558, + 296, + 677 + ], + "type": "text", + "content": " respectively, we compute the pairwise Euclidean distances within each collection of centroids. By minimizing discrepancy in distances between predicted and target centroid collections, we formulate the distance loss function as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 681, + 296, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 681, + 296, + 717 + ], + "spans": [ + { + "bbox": [ + 61, + 681, + 296, + 717 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i s t}} = \\sum_ {n _ {1} = 1} ^ {N} \\sum_ {n _ {1} = 2} ^ {N} \\left(\\| \\hat {\\mathbf {c}} _ {n _ {1}} - \\hat {\\mathbf {c}} _ {n _ {2}} \\| _ {2} - \\| \\mathbf {c} _ {n _ {1}} - \\mathbf {c} _ {n _ {2}} \\| _ {2}\\right), \\tag {12}", + "image_path": "bee282c2638f41b3835c0af5a5a153bc34203abdb3e2a7d05bdc72896ac33b38.jpg" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 68, + 557, + 174 + ], + "blocks": [ + { + "bbox": [ + 315, + 68, + 557, + 174 + ], + "lines": [ + { + "bbox": [ + 315, + 68, + 557, + 174 + ], + "spans": [ + { + "bbox": [ + 315, + 68, + 557, + 174 + ], + "type": "image", + "image_path": "1cc73fe18bdc4e0ba385e060b9fbb42c83752efb6dc28cda9459fb96722c679a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 183, + 555, + 239 + ], + "lines": [ + { + "bbox": [ + 313, + 183, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 555, + 239 + ], + "type": "text", + "content": "Figure 4. The overall pipeline of our Dense Motion Alignment (DMA) module. Using our densified reference drag flow, this module aims to guide synthesized image in achieving dense alignment of motion information across drag conditions, encompassing pixel displacement, appearance, and semantic aspects." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 261, + 553, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 553, + 285 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 261, + 553, + 285 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{c}}_{n_1}(\\hat{\\mathbf{c}}_{n_2})" + }, + { + "bbox": [ + 313, + 261, + 553, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 261, + 553, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{n_1}(\\mathbf{c}_{n_2})" + }, + { + "bbox": [ + 313, + 261, + 553, + 285 + ], + "type": "text", + "content": " are centroids of the segmentation masks in the collections " + }, + { + "bbox": [ + 313, + 261, + 553, + 285 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}" + }, + { + "bbox": [ + 313, + 261, + 553, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 261, + 553, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 261, + 553, + 285 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 286, + 553, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 286, + 553, + 309 + ], + "spans": [ + { + "bbox": [ + 313, + 286, + 553, + 309 + ], + "type": "text", + "content": "In summary, our final guidance function for DGA can be defined as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 337, + 322, + 553, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 322, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 337, + 322, + 553, + 335 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D G A}} = (1 - \\lambda) \\mathcal {L} _ {c v g} + \\lambda \\left(\\mathcal {L} _ {s i z e} + \\mathcal {L} _ {d i s t}\\right), \\tag {13}", + "image_path": "9c97abb427119e40db78ecac9e01c299a53caeaf28a3445550b86cd729f0ff5f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 346, + 411, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 346, + 411, + 357 + ], + "spans": [ + { + "bbox": [ + 314, + 346, + 411, + 357 + ], + "type": "text", + "content": "where we set " + }, + { + "bbox": [ + 314, + 346, + 411, + 357 + ], + "type": "inline_equation", + "content": "\\lambda = 0.25" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 368, + 491, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 491, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 491, + 380 + ], + "type": "text", + "content": "4.3. Dense Motion Alignment (DMA)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 386, + 555, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 555, + 591 + ], + "type": "text", + "content": "Given a collection of " + }, + { + "bbox": [ + 313, + 386, + 555, + 591 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 313, + 386, + 555, + 591 + ], + "type": "text", + "content": " drag signals " + }, + { + "bbox": [ + 313, + 386, + 555, + 591 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{(\\mathbf{p}_j^{\\mathrm{o}},\\mathbf{p}_j^{\\mathrm{d}})\\}_{j = 1}^{J}" + }, + { + "bbox": [ + 313, + 386, + 555, + 591 + ], + "type": "text", + "content": ", we aim to move visual contents of each original point " + }, + { + "bbox": [ + 313, + 386, + 555, + 591 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_j^{\\mathrm{o}}" + }, + { + "bbox": [ + 313, + 386, + 555, + 591 + ], + "type": "text", + "content": " to its destination " + }, + { + "bbox": [ + 313, + 386, + 555, + 591 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_j^{\\mathrm{d}}" + }, + { + "bbox": [ + 313, + 386, + 555, + 591 + ], + "type": "text", + "content": ". Recent studies, e.g., DragDiffusion [39] and DragNoise [25], only regulate feature similarity around " + }, + { + "bbox": [ + 313, + 386, + 555, + 591 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_j^{\\mathrm{o}}" + }, + { + "bbox": [ + 313, + 386, + 555, + 591 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 386, + 555, + 591 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_j^{\\mathrm{d}}" + }, + { + "bbox": [ + 313, + 386, + 555, + 591 + ], + "type": "text", + "content": " by implicit optimization in one selected diffusion step. However, on one hand, sparse drag signals do not provide sufficient object awareness, causing incomplete displacement in foreground. On the other hand, one-step feature optimization is unstable. To enable continuous dragging of designated pixels while preserving visual coherence during diffusion sampling, we propose Dense Motion Alignment (DMA), as shown in Fig. 4. Specifically, we predict reference dense drag flow, and enforce pixel correspondence between an image pair in terms of displacement, appearance, and semantics building on this flow. We define " + }, + { + "bbox": [ + 313, + 386, + 555, + 591 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0^{t + 1}" + }, + { + "bbox": [ + 313, + 386, + 555, + 591 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 386, + 555, + 591 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0^t" + }, + { + "bbox": [ + 313, + 386, + 555, + 591 + ], + "type": "text", + "content": " as original and dragging images, respectively, for each timestep " + }, + { + "bbox": [ + 313, + 386, + 555, + 591 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 386, + 555, + 591 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 592, + 556, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 556, + 716 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 556, + 716 + ], + "type": "text", + "content": "We generate the reference dense drag flow for each individual drag signal using an object-specific displacement field. In particular, we employ SAM [20] to obtain a segmentation mask collection " + }, + { + "bbox": [ + 313, + 592, + 556, + 716 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{m}_j^{\\mathrm{sam}}\\}_{j = 1}^{J}" + }, + { + "bbox": [ + 313, + 592, + 556, + 716 + ], + "type": "text", + "content": ", where each mask represents an instance area of " + }, + { + "bbox": [ + 313, + 592, + 556, + 716 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_j^{\\mathrm{o}}" + }, + { + "bbox": [ + 313, + 592, + 556, + 716 + ], + "type": "text", + "content": ". Because image pixels closer to " + }, + { + "bbox": [ + 313, + 592, + 556, + 716 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_j^{\\mathrm{o}}" + }, + { + "bbox": [ + 313, + 592, + 556, + 716 + ], + "type": "text", + "content": " should have larger displacements [48], we use radial basis functions to produce a collection of 2D Gaussian maps " + }, + { + "bbox": [ + 313, + 592, + 556, + 716 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{g}_j\\}_{j = 1}^{J}" + }, + { + "bbox": [ + 313, + 592, + 556, + 716 + ], + "type": "text", + "content": ". Each map depicts gradually diminishing drag influence radiating outward from " + }, + { + "bbox": [ + 313, + 592, + 556, + 716 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_j^{\\mathrm{o}}" + }, + { + "bbox": [ + 313, + 592, + 556, + 716 + ], + "type": "text", + "content": ". Based on " + }, + { + "bbox": [ + 313, + 592, + 556, + 716 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{m}_j^{\\mathrm{sam}}\\}_{j = 1}^{J}" + }, + { + "bbox": [ + 313, + 592, + 556, + 716 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 592, + 556, + 716 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{g}_j\\}_{j = 1}^{J}" + }, + { + "bbox": [ + 313, + 592, + 556, + 716 + ], + "type": "text", + "content": ", our reference flow can be densi" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13139" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 88, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 88, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 88, + 83 + ], + "type": "text", + "content": "fied by:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 95, + 92, + 295, + 111 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 92, + 295, + 111 + ], + "spans": [ + { + "bbox": [ + 95, + 92, + 295, + 111 + ], + "type": "interline_equation", + "content": "\\mathbf {u} _ {\\hat {\\mathbf {x}} _ {0} ^ {t + 1} \\rightarrow \\hat {\\mathbf {x}} _ {0} ^ {t}, j} = \\left(\\mathbf {p} _ {j} ^ {\\mathrm {d}} - \\mathbf {p} _ {j} ^ {\\mathrm {o}}\\right) * \\mathbf {m} _ {j} ^ {\\text {s a m}} * \\mathbf {g} _ {j}, \\tag {14}", + "image_path": "ad212086dfd32aa484811b538c1be670617b2e80bb7422f6a0744934a7e48d2c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 117, + 295, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 117, + 295, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 117, + 295, + 131 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 117, + 295, + 131 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_{\\hat{\\mathbf{x}}_0^{t + 1}\\rightarrow \\hat{\\mathbf{x}}_0^t,j}" + }, + { + "bbox": [ + 55, + 117, + 295, + 131 + ], + "type": "text", + "content": " is the densified drag flow for " + }, + { + "bbox": [ + 55, + 117, + 295, + 131 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 117, + 295, + 131 + ], + "type": "text", + "content": " -th signal." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 131, + 296, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 131, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 131, + 296, + 178 + ], + "type": "text", + "content": "To rectify discrepancies in pixel displacement direction and magnitude, we enable the estimated optical flow between " + }, + { + "bbox": [ + 55, + 131, + 296, + 178 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0^{t + 1}" + }, + { + "bbox": [ + 55, + 131, + 296, + 178 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 131, + 296, + 178 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0^t" + }, + { + "bbox": [ + 55, + 131, + 296, + 178 + ], + "type": "text", + "content": " to align with its reference flow by minimizing displacement loss:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 78, + 186, + 295, + 221 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 186, + 295, + 221 + ], + "spans": [ + { + "bbox": [ + 78, + 186, + 295, + 221 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i s p}} = \\sum_ {j = 1} ^ {J} \\left|\\left| \\hat {\\mathbf {u}} _ {\\hat {\\mathbf {x}} _ {0} ^ {t + 1} \\rightarrow \\hat {\\mathbf {x}} _ {0} ^ {t}, j} - \\mathbf {u} _ {\\hat {\\mathbf {x}} _ {0} ^ {t + 1} \\rightarrow \\hat {\\mathbf {x}} _ {0} ^ {t}, j} \\right|\\right| _ {1}, \\tag {15}", + "image_path": "4f0552987b2ae32bf2d657d9ebe91372fcd332317324b02d47bfe653c0c64ec2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 228, + 296, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 304 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 228, + 296, + 304 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{u}}_{\\hat{\\mathbf{x}}_0^{t + 1}\\rightarrow \\hat{\\mathbf{x}}_0^t,j}" + }, + { + "bbox": [ + 55, + 228, + 296, + 304 + ], + "type": "text", + "content": " is an optical flow estimated by a RAFT model [41]. In order to ensure visual coherence, we enforce pixel-level appearance consistency by directly minimizing the differences between the dragging and warped original image, while semantic consistency is preserved by DINO v2 embedding [30]:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 65, + 311, + 295, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 311, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 65, + 311, + 295, + 346 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p a p p} = \\sum_ {j = 1} ^ {J} \\sqrt {\\left(\\operatorname {w a r p} \\left(\\hat {\\mathbf {x}} _ {0} ^ {t + 1} , \\mathbf {u} _ {\\hat {\\mathbf {x}} _ {0} ^ {t + 1} \\rightarrow \\hat {\\mathbf {x}} _ {0} ^ {t} , j}\\right) - \\hat {\\mathbf {x}} _ {0} ^ {t}\\right) ^ {2}}, \\tag {16}", + "image_path": "0e0aee590e13dffea67e931fe7fcddfd97400aeb4e00014b125cf8aa9081f05d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 354, + 295, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 354, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 66, + 354, + 295, + 388 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p s e m} = \\sum_ {j = 1} ^ {J} \\sqrt {\\left(\\operatorname {w a r p} \\left(\\hat {\\mathbf {f}} _ {0} ^ {t + 1} , \\mathbf {u} _ {\\hat {\\mathbf {x}} _ {0} ^ {t + 1} \\rightarrow \\hat {\\mathbf {x}} _ {0} ^ {t} , j}\\right) - \\hat {\\mathbf {f}} _ {0} ^ {t}\\right) ^ {2}}, \\tag {17}", + "image_path": "705f8ea5c75f401604785f0bf9f2fe9bbe86a5bab69e654e622fd0d2b3bd56bf.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 393, + 296, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 393, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 296, + 431 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 393, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{f}}_0^{t + 1}" + }, + { + "bbox": [ + 55, + 393, + 296, + 431 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 393, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{f}}_0^t" + }, + { + "bbox": [ + 55, + 393, + 296, + 431 + ], + "type": "text", + "content": " are semantic embedding calculated by DINO v2 model. For a drag signal, our loss of DMA module is defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 74, + 440, + 295, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 440, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 74, + 440, + 295, + 455 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D M A}} = (1 - \\eta) \\mathcal {L} _ {d i s p} + \\eta \\left(\\mathcal {L} _ {p a p p} + \\mathcal {L} _ {p s e m}\\right), \\tag {18}", + "image_path": "15b8a13b7b248c794f2629ff70255705c50564c3c973ae715d155c37956b96ec.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 462, + 296, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 296, + 499 + ], + "type": "text", + "content": "where we set " + }, + { + "bbox": [ + 55, + 462, + 296, + 499 + ], + "type": "inline_equation", + "content": "\\eta = 0.02" + }, + { + "bbox": [ + 55, + 462, + 296, + 499 + ], + "type": "text", + "content": ". Like DragDiffusion [39] and Drag-Noise [25], we also update the positions of the original points after each denosing step." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 509, + 137, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 137, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 137, + 522 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 528, + 176, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 176, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 176, + 541 + ], + "type": "text", + "content": "5.1. Experimental Setups" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 545, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 581 + ], + "type": "text", + "content": "This section sequentially describes used datasets and evaluation metrics in our experiments, and implementation details of our approaches." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "Datasets. The datasets used in our study are: (1) COCO [24]. We curate the 400 longest image descriptions from COCO validation set, each ranging from 15 to 50 words. This selection is made to better demonstrate our strength in understanding long-form description; (2) DenseDiffusion [19]. This benchmark contains 250 samples " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "(2\\sim 3" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " unique objects of each sample), where each sample has a global image description, a segmentation mask, and local descriptions and class labels for each segment; (3) DrawBench [38]. We use the 'position' subset from it. This subset is composed of 20 samples (2 unique objects of each sample)," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "where each sample has a description and some bounding boxes of each object (box coordinates and class labels generated by GPT); and (4) DragBench [39]. This benchmark comprises 205 samples, each of which has a reference image, one or more drag signals, a description, and a mask indicating regions to be modified." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 150, + 555, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 150, + 555, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 150, + 555, + 355 + ], + "type": "text", + "content": "Evaluation metrics. For text condition, we use Fréchet Inception Distance (FID) [12] and CLIP score [34]. FID quantifies the similarity between the synthesized and real images, while the CLIP score measures the alignment between the synthesized image and the reference description. For text + layout condition, we use IoU and SOA-I [13] (layout prediction is based on YOLOv7-Seg [44]). IoU quantifies the overlap between the predicted layout from the synthesized image and the reference layout. SOA-I measures whether described objects appear in the synthesized image. For text + drag condition, we use Image Fidelity (IF) [17] and Mean Distance (MD) [31]. IF (1-LPIPS score [54]) measures the image quality based on the global similarity between the reference and dragged images. MD assesses how accurately the post-dragging original point align with its corresponding destination point by measuring Euclidean distance between their positions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 361, + 556, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 556, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 556, + 434 + ], + "type": "text", + "content": "Implementation details. We use " + }, + { + "bbox": [ + 313, + 361, + 556, + 434 + ], + "type": "inline_equation", + "content": "T = 50" + }, + { + "bbox": [ + 313, + 361, + 556, + 434 + ], + "type": "text", + "content": " steps of the DDIM scheduler [40] and " + }, + { + "bbox": [ + 313, + 361, + 556, + 434 + ], + "type": "inline_equation", + "content": "T = 28" + }, + { + "bbox": [ + 313, + 361, + 556, + 434 + ], + "type": "text", + "content": " steps of the Flow Match Euler Discrete scheduler [10]. The weighting schedules for our proposed DCA, DGA, and DMA modules are set to 60, 25, and 90, respectively. Our guidance is only applied during the first " + }, + { + "bbox": [ + 313, + 361, + 556, + 434 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 313, + 361, + 556, + 434 + ], + "type": "text", + "content": " timesteps." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 461, + 497, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 497, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 497, + 475 + ], + "type": "text", + "content": "5.2. Image Synthesis Results with Text" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 485, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 555, + 714 + ], + "type": "text", + "content": "We use Stable Diffusion v1.5 [36] and v3.0 [8] as baseline models to evaluate the effectiveness of our DCA module in aligning concepts within descriptions. Tab. 1 summarizes FID [12] and CLIP scores [34] on long (" + }, + { + "bbox": [ + 313, + 485, + 555, + 714 + ], + "type": "inline_equation", + "content": ">15" + }, + { + "bbox": [ + 313, + 485, + 555, + 714 + ], + "type": "text", + "content": " words) descriptions from COCO dataset [24], our DCA module achieves " + }, + { + "bbox": [ + 313, + 485, + 555, + 714 + ], + "type": "inline_equation", + "content": "12.3\\%" + }, + { + "bbox": [ + 313, + 485, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 485, + 555, + 714 + ], + "type": "inline_equation", + "content": "2.97\\%" + }, + { + "bbox": [ + 313, + 485, + 555, + 714 + ], + "type": "text", + "content": " improvements in FID and CLIP scores, respectively, building upon Stable Diffusion v3.0 (the most recent baseline). This gain arises from our module's ability to achieve vision-language alignment in a coarse-to-fine manner, effectively leveraging semantic cues embedded in complex sentences. Besides, our module can be applicable to Stable Diffusion v1.5, showing its adaptability across different baseline versions. As shown in Fig. 5, our approach can properly place the dog on a dog bed next to a rocking chair (top) and position the person's head on a blue backpack (middle), indicating its effectiveness in aligning the object interaction concept. Our approach can also generate two people (bottom), showcasing its advantage in understanding the object's numeric information." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13140" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 123, + 294, + 247 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "type": "text", + "content": "Table 1. Quantitative comparison under long descriptions from COCO dataset [24]. The evaluation metrics used are FID [12] and CLIPScore [34]. We reimplement all baseline approaches. The best results are highlighted." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 123, + 294, + 247 + ], + "lines": [ + { + "bbox": [ + 58, + 123, + 294, + 247 + ], + "spans": [ + { + "bbox": [ + 58, + 123, + 294, + 247 + ], + "type": "table", + "html": "
MethodsVenueFID ↓CLIPScore ↑
ParaDiffusion [47]arXiv202312.960.2731
PIXART-α [5]ICLR202413.970.2724
SDXL [33]ICLR20249.790.2820
ELLA [16]arXiv202413.640.2811
PIXART-Σ [4]ECCV202410.730.2793
Stable Diffusion v1.5 [36]CVPR202211.490.2750
+ DCA-10.290.2882
Stable Diffusion v3.0 [8]ICML202410.510.2861
+ DCA-9.220.2946
", + "image_path": "7e82b9cbfe60a1b6dd3bef7d3f3c9e36ecc55a38d7dae3db68d3902b2b335719.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 55, + 258, + 296, + 430 + ], + "blocks": [ + { + "bbox": [ + 55, + 258, + 296, + 430 + ], + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 430 + ], + "type": "table", + "html": "
DescriptionSD v3.0Ours
A room with filtered light coming through vertical blinds on a doorwall, with a sleeping dog lying on a dog bed next to a rocking chair.
A person covered with an orange sleeping bag rests his head on a blue back pack as he sleeps on a park bench.
In the desert, two people are sitting next to a small propeller plane next to a tent with other planes in the background.
", + "image_path": "fbd4e3ec69cc6d17e6d5771f405f9d9a6c86ca7b6bf16480019ada5b19969e46.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 437, + 296, + 483 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 483 + ], + "type": "text", + "content": "Figure 5. Visualization comparison between our approach (Stable Diffusion v3.0 [8] + DCA) and its baseline under varying descriptions. Our approach demonstrates superior performance in aligning semantic concepts within descriptions." + } + ] + } + ], + "index": 3, + "type": "text" + }, + { + "bbox": [ + 55, + 503, + 294, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 294, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 294, + 517 + ], + "type": "text", + "content": "5.3. Image Synthesis Results with Layout and Text" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 521, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 297, + 715 + ], + "type": "text", + "content": "We use Stable Diffusion v1.5 [36], DenseDiffusion [19] and A&R [32] as baseline models to evaluate the effectiveness of our DGA module in aligning spatial configurations represented by segmentation masks and bounding boxes; and our DCA module in aligning geometric concepts within descriptions. Tab. 2 summarizes IoU and SOA-I scores on mask-to-image. Our DGA module brings a " + }, + { + "bbox": [ + 54, + 521, + 297, + 715 + ], + "type": "inline_equation", + "content": "27.1\\%" + }, + { + "bbox": [ + 54, + 521, + 297, + 715 + ], + "type": "text", + "content": " improvement in IoU and a " + }, + { + "bbox": [ + 54, + 521, + 297, + 715 + ], + "type": "inline_equation", + "content": "7.55\\%" + }, + { + "bbox": [ + 54, + 521, + 297, + 715 + ], + "type": "text", + "content": " increase in SOA-I over A&R baseline, owing to the establishment of object-wise geometric consistency. Besides, incorporating our DCA module on top of DGA leads to an additional improvement, achieving an IoU of 50.61 and an SOA-I of 86.33. This improvement is due to the ability of DCA module to understand descriptions of positions and shapes. We can also observe consistent performance gains across both Stable Diffusion v1.5 and DenseDiffusion architectures. As shown in Fig. 6, our" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 135, + 555, + 294 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 127 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 127 + ], + "type": "text", + "content": "Table 2. Quantitative comparison under descriptions and segmentation masks provided by DenseDiffusion [19]. The evaluation metrics are IoU and SOA-I [13]. We generate four random images of each condition combination for performance evaluation. The best results are highlighted." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 135, + 555, + 294 + ], + "lines": [ + { + "bbox": [ + 317, + 135, + 555, + 294 + ], + "spans": [ + { + "bbox": [ + 317, + 135, + 555, + 294 + ], + "type": "table", + "html": "
MethodsVenueIoU ↑SOA-I ↑
Real Images-59.1393.07
Composable Diffusion [26]ECCV2022-55.88±0.78
Structure Diffusion [9]ICLR2023-70.97±1.08
SD-Pww [1]arXiv202223.76±0.5073.92±1.84
Stable Diffusion [36]CVPR2022-73.08±1.54
+ DGA-25.46±1.0274.45±1.46
+ DGA + DCA-27.28±0.9777.02±1.67
Dense Diffusion [19]ICCV202334.99±1.1377.61±1.75
+ DGA-36.72±1.2478.21±1.62
+ DGA + DCA-38.06±1.2678.42±1.17
A&R [32]CVPR202438.97±0.5678.80±1.27
+ DGA-49.52±0.7084.75±2.46
+ DGA + DCA-50.61±0.4786.33±0.88
", + "image_path": "8a61584b0ed3cb225ed2f3341acb439feb5272cb9353e6234c4c441efefb0967.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 304, + 555, + 476 + ], + "blocks": [ + { + "bbox": [ + 316, + 304, + 555, + 476 + ], + "lines": [ + { + "bbox": [ + 316, + 304, + 555, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 555, + 476 + ], + "type": "table", + "html": "
DescriptionLayoutA&ROurs
An orange tabby cat lying in a bathroom sink.
An orange cat laying on top of a black car.
A dog sitting on a bed wearing an orange bowtie.
", + "image_path": "65af9171ba18b6fa4263469707ca7b188301a7d83a7ddcca921ccfb6ad8e820c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 484, + 556, + 531 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 556, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 556, + 531 + ], + "type": "text", + "content": "Figure 6. Visualization comparison between our approach (A&R [32] + DGA + DCA) and its baseline, under varying descriptions and segmentation masks. Our approach can adhere to both geometric concepts and positional configuration." + } + ] + } + ], + "index": 9, + "type": "text" + }, + { + "bbox": [ + 313, + 552, + 556, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 556, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 556, + 636 + ], + "type": "text", + "content": "approach can accurately place a cat inside a bathroom sink (top), position a cat on top of a black car (middle), generate a bowtie around a dog's neck (bottom), demonstrating the advantage of our framework in awareness of spatial relationships. The same conclusion can be observed in Fig. 7, which presents box-to-image visualizations of validation data from the DrawBench benchmark [38]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 647, + 544, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 544, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 544, + 661 + ], + "type": "text", + "content": "5.4. Image Synthesis Results with Drag and Text" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "We use DragDiffusion [39] as a baseline model to evaluate the effectiveness of our DMA module in motion alignment; and our DCA module for semantic preservation. Fig. 8 presents MD and IF scores on the DragBench benchmark" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "text", + "content": "13141" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 294, + 240 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 294, + 240 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 294, + 240 + ], + "type": "image", + "image_path": "74fdbdee1a5adb7d2c2d8d258d4052db4675ac9ddd2ad10444f8308cb16385b4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 248, + 294, + 282 + ], + "lines": [ + { + "bbox": [ + 55, + 248, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 294, + 282 + ], + "type": "text", + "content": "Figure 7. Visualization comparison between our approach (A&R [32] + DGA + DCA) and its baseline, under varying descriptions and bounding boxes. Our approach can satisfy both conditions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 294, + 293, + 470 + ], + "blocks": [ + { + "bbox": [ + 56, + 294, + 293, + 470 + ], + "lines": [ + { + "bbox": [ + 56, + 294, + 293, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 293, + 470 + ], + "type": "image", + "image_path": "76a6c76452ae6710e101e547b4acc7db30821ad773e2c86f0ccdf1e123de64e4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 479, + 295, + 523 + ], + "lines": [ + { + "bbox": [ + 55, + 479, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 295, + 523 + ], + "type": "text", + "content": "Figure 8. Quantitative comparison under descriptions and drag signals from DragBench benchmark [39]. The evaluation metrics used are Mean Distance " + }, + { + "bbox": [ + 55, + 479, + 295, + 523 + ], + "type": "inline_equation", + "content": "(\\downarrow)" + }, + { + "bbox": [ + 55, + 479, + 295, + 523 + ], + "type": "text", + "content": " and Image Fidelity (1-LPIPS) " + }, + { + "bbox": [ + 55, + 479, + 295, + 523 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 55, + 479, + 295, + 523 + ], + "type": "text", + "content": ". Approaches with the best results appear in the upper-left corner." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "[39]. The combination of our DMA and DCA modules achieves better MD and IF scores, indicating that our framework ensures each pixel follows its desired trajectory with fewer visual artifacts. However, applying our DMA module alone results in a slight drop in IF, compared to DragDiffusion. This is because our motion field may cause oversmoothing in detail. Fortunately, our DCA module can mitigate such an issue to some degree, by maintaining the overall semantic structure unchanged during the diffusion sampling process. As shown in Fig. 9, our framework can effectively synthesize head rotation (top), rigid car displacement (middle), and large-scale deformation of the bus (bottom). This demonstrates that our algorithm enables the desired movement even when handling complex drag signals" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 316, + 68, + 553, + 274 + ], + "blocks": [ + { + "bbox": [ + 316, + 68, + 553, + 274 + ], + "lines": [ + { + "bbox": [ + 316, + 68, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 68, + 553, + 274 + ], + "type": "image", + "image_path": "9e5fa1bdbb06071f4fbf2fe3b9301c50d9bb040c619eff466ca74372d97b98b8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 281, + 555, + 326 + ], + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 326 + ], + "type": "text", + "content": "Figure 9. Visualization comparison between our approach (DragDiffusion [39] + DMA) and its baseline under varying drag signals. Our approach enables effective synthesis of large-scale movement and deformation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 338, + 406, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 338, + 406, + 350 + ], + "spans": [ + { + "bbox": [ + 314, + 338, + 406, + 350 + ], + "type": "text", + "content": "(multiple drag points)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 360, + 388, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 360, + 388, + 372 + ], + "spans": [ + { + "bbox": [ + 314, + 360, + 388, + 372 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 380, + 555, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 555, + 536 + ], + "type": "text", + "content": "We propose a plug-and-play modular design for guiding image synthesis by aligning diverse fundamental condition units (text, layout, drag) in combination. Based on this idea, we devise three dense alignment approaches to independently achieve flexible control over concept, geometry and motion. Comprehensive experiments demonstrate the effectiveness of our framework across diverse conditioning scenarios. However, our framework relies on the capability of its foundation model, e.g., SD model. As shown in Fig. 10, counterfactual scenes is hard to generate, e.g., a river flowing upwards, if SD model is unable to synthesis this scene itself. Future work could explore world knowledge-aware generative paradigms to address this challenge." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 544, + 553, + 600 + ], + "blocks": [ + { + "bbox": [ + 317, + 544, + 553, + 600 + ], + "lines": [ + { + "bbox": [ + 317, + 544, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 544, + 553, + 600 + ], + "type": "image", + "image_path": "3f6afa46f4f503cf30cbdde5146486933b25a48ce7b471e2ada878e9bb03a3fb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 608, + 555, + 640 + ], + "lines": [ + { + "bbox": [ + 313, + 608, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 608, + 555, + 640 + ], + "type": "text", + "content": "Figure 10. Failure cases. As foundation model lacks world knowledge, our approach (Stable Diffusion v3.0 [8] + DCA) fails to generate counterfactual scenes." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": "Acknowledgement: This work was supported by the National Natural Science Foundation of China (No. U23B2013, 62276176). This work was also partly supported by the SICHUAN Provincial Natural Science Foundation (No. 2024NSFJQ0023)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13142" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 705 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 211 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 211 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 211 + ], + "type": "text", + "content": "[2] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 843-852, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 212, + 295, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 212, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 62, + 212, + 295, + 255 + ], + "type": "text", + "content": "[3] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 256, + 295, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 256, + 295, + 310 + ], + "spans": [ + { + "bbox": [ + 62, + 256, + 295, + 310 + ], + "type": "text", + "content": "[4] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 311, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 311, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 62, + 311, + 294, + 376 + ], + "type": "text", + "content": "[5] Junsong Chen, YU Jincheng, GE Chongjian, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In Proceedings of the International Conference on Learning Representations (ICLR), pages 1–30, 2024. 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 376, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 294, + 430 + ], + "type": "text", + "content": "[6] Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xinggang Wang, and Ying Shan. Yolo-world: Real-time open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16901-16911, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 431, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 431, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 431, + 294, + 463 + ], + "type": "text", + "content": "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems (NeruIPS), 34:8780-8794, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 464, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 294, + 529 + ], + "type": "text", + "content": "[8] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning (ICML), 2024. 2, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 529, + 294, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 529, + 294, + 594 + ], + "spans": [ + { + "bbox": [ + 62, + 529, + 294, + 594 + ], + "type": "text", + "content": "[9] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. In Proceedings of the International Conference on Learning Representations (ICLR), pages 1-13, 2023. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 596, + 295, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 596, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 57, + 596, + 295, + 639 + ], + "type": "text", + "content": "[10] Itai Gat, Tal Remez, Neta Shaul, Felix Kreuk, Ricky TQ Chen, Gabriel Synnaeve, Yossi Adi, and Yaron Lipman. Discrete flow matching. Advances in Neural Information Processing Systems (NeurIPS), 37:133345-133385, 2024. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 639, + 295, + 683 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 639, + 295, + 683 + ], + "spans": [ + { + "bbox": [ + 57, + 639, + 295, + 683 + ], + "type": "text", + "content": "[11] Daniel Geng and Andrew Owens. Motion guidance: Diffusion-based image editing with differentiable motion estimators. In The Twelfth International Conference on Learning Representations (ICLR), pages 1-21. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 684, + 295, + 705 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 684, + 295, + 705 + ], + "spans": [ + { + "bbox": [ + 57, + 684, + 295, + 705 + ], + "type": "text", + "content": "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 696 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[13] Tobias Hinz, Stefan Heinrich, and Stefan Wermter. Semantic object accuracy for generative text-to-image synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(3):1552-1565, 2020. 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "text", + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NeurIPS), 33:6840-6851, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 186, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 240 + ], + "type": "text", + "content": "[15] Xingzhong Hou, Boxiao Liu, Yi Zhang, Jihao Liu, Yu Liu, and Haihang You. Easydrag: Efficient point-based manipulation on diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8404-8413, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "text", + "content": "[16] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135, 2024. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 286, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 342 + ], + "type": "text", + "content": "[17] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6007-6017, 2023. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 342, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 342, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 317, + 342, + 553, + 407 + ], + "type": "text", + "content": "[18] Seung Wook Kim, Bradley Brown, Kangxue Yin, Karsten Kreis, Katja Schwarz, Daiqing Li, Robin Rombach, Antonio Torralba, and Sanja Fidler. Neuralfield-ldm: Scene generation with hierarchical latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8496-8506, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 409, + 553, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 409, + 553, + 463 + ], + "spans": [ + { + "bbox": [ + 316, + 409, + 553, + 463 + ], + "type": "text", + "content": "[19] Yunji Kim, Jiyoung Lee, Jin-Hwa Kim, Jung-Woo Ha, and Jun-Yan Zhu. Dense text-to-image generation with attention modulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7701-7711, 2023. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 464, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 464, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 317, + 464, + 553, + 529 + ], + "type": "text", + "content": "[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4015-4026, 2023. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 531, + 553, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 531, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 316, + 531, + 553, + 575 + ], + "type": "text", + "content": "[21] Yinjie Lei, Zixuan Wang, Feng Chen, Guoqing Wang, Peng Wang, and Yang Yang. Recent advances in multi-modal 3d scene understanding: A comprehensive survey and evaluation. arXiv preprint arXiv:2310.15676, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 576, + 553, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 576, + 553, + 631 + ], + "spans": [ + { + "bbox": [ + 317, + 576, + 553, + 631 + ], + "type": "text", + "content": "[22] Ming Li, Taojiannan Yang, Huafeng Kuang, Jie Wu, Zhaoning Wang, Xuefeng Xiao, and Chen Chen. Controlnet++: Improving conditional controls with efficient consistency feedback. In European Conference on Computer Vision (ECCV), pages 129-147. Springer, 2025. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 632, + 553, + 696 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 632, + 553, + 696 + ], + "spans": [ + { + "bbox": [ + 317, + 632, + 553, + 696 + ], + "type": "text", + "content": "[23] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22511-22521, 2023. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13143" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 718 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision (ECCV), pages 740-755. Springer, 2014. 2, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 295, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 295, + 183 + ], + "type": "text", + "content": "[25] Haofeng Liu, Chenshu Xu, Yifei Yang, Lihua Zeng, and Shengfeng He. Drag your noise: Interactive point-based editing via diffusion semantic propagation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6743-6752, 2024. 3, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 183, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 183, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 183, + 294, + 237 + ], + "type": "text", + "content": "[26] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In Proceedings of the European Conference on Computer Vision (ECCV), pages 423-439. Springer, 2022. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 237, + 294, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 237, + 294, + 303 + ], + "spans": [ + { + "bbox": [ + 56, + 237, + 294, + 303 + ], + "type": "text", + "content": "[27] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision (ECCV), pages 38-55. Springer, 2024. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 303, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 303, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 303, + 294, + 369 + ], + "type": "text", + "content": "[28] Sicheng Mo, Fangzhou Mu, Kuan Heng Lin, Yanli Liu, Bochen Guan, Yin Li, and Bolei Zhou. Freecontrol: Training-free spatial control of any text-to-image diffusion model with any condition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7465-7475, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 369, + 294, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 369, + 294, + 424 + ], + "spans": [ + { + "bbox": [ + 56, + 369, + 294, + 424 + ], + "type": "text", + "content": "[29] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 4296-4304, 2024. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 424, + 294, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 424, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 56, + 424, + 294, + 479 + ], + "type": "text", + "content": "[30] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "type": "text", + "content": "[31] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 533, + 294, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 294, + 587 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 294, + 587 + ], + "type": "text", + "content": "[32] Quynh Phung, Songwei Ge, and Jia-Bin Huang. Grounded text-to-image synthesis with attention refocusing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7932-7942, 2024. 3, 4, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 587, + 294, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 587, + 294, + 654 + ], + "spans": [ + { + "bbox": [ + 56, + 587, + 294, + 654 + ], + "type": "text", + "content": "[33] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In Proceedings of the International Conference on Learning Representations (ICLR), pages 1-13, 2024. 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 654, + 294, + 718 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 654, + 294, + 718 + ], + "spans": [ + { + "bbox": [ + 56, + 654, + 294, + 718 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PMLR, 2021. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 708 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "text", + "content": "[35] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research (JMLR), 21(140):1-67, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 129, + 555, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 184 + ], + "type": "text", + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 185, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 251 + ], + "type": "text", + "content": "[37] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6527-6536, 2024. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 251, + 553, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 328 + ], + "type": "text", + "content": "[38] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems (NeurIPS), 35:36479-36494, 2022. 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 329, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 329, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 329, + 553, + 396 + ], + "type": "text", + "content": "[39] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8839-8849, 2024. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 396, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 396, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 396, + 553, + 430 + ], + "type": "text", + "content": "[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations (ICLR), pages 1-22, 2021. 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 430, + 553, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 430, + 553, + 474 + ], + "spans": [ + { + "bbox": [ + 316, + 430, + 553, + 474 + ], + "type": "text", + "content": "[41] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In European Conference on Computer Vision (ECCV), pages 402-419. Springer, 2020. 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 475, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 553, + 529 + ], + "type": "text", + "content": "[42] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1921-1930, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 529, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 529, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 529, + 553, + 586 + ], + "type": "text", + "content": "[43] Bram Wallace, Akash Gokul, Stefano Ermon, and Nikhil Naik. End-to-end diffusion latent optimization improves classifier guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR), pages 7280-7290, 2023. 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 586, + 553, + 652 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 586, + 553, + 652 + ], + "spans": [ + { + "bbox": [ + 316, + 586, + 553, + 652 + ], + "type": "text", + "content": "[44] Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7464-7475, 2023. 4, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 652, + 553, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 652, + 553, + 708 + ], + "spans": [ + { + "bbox": [ + 316, + 652, + 553, + 708 + ], + "type": "text", + "content": "[45] Yang Wang, Zhikui Ouyang, Runhua Han, Zhijian Yin, and Zhen Yang. Yolomask: Real-time instance segmentation with integrating yolov5 and orienmask. In IEEE International Conference on Communication Technology (ICCT), pages 1646-1650. IEEE, 2022. 3" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13144" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 295, + 357 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "text", + "content": "[46] Hao Wu, Jiayuan Mao, Yufeng Zhang, Yuning Jiang, Lei Li, Weiwei Sun, and Wei-Ying Ma. Unified visual-semantic embeddings: Bridging vision and language with structured meaning representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6609–6618, 2019. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 140, + 295, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 295, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 295, + 193 + ], + "type": "text", + "content": "[47] Weijia Wu, Zhuang Li, Yefei He, Mike Zheng Shou, Chunhua Shen, Lele Cheng, Yan Li, Tingting Gao, Di Zhang, and Zhongyuan Wang. Paragraph-to-image generation with information-enriched diffusion model. arXiv preprint arXiv:2311.14284, 2023. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 194, + 295, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 194, + 295, + 248 + ], + "spans": [ + { + "bbox": [ + 57, + 194, + 295, + 248 + ], + "type": "text", + "content": "[48] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control for anything using entity representation. In European Conference on Computer Vision (ECCV), pages 331-348. Springer, 2025. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 249, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 249, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 57, + 249, + 295, + 293 + ], + "type": "text", + "content": "[49] Jiayu Xiao, Henglei Lv, Liang Li, Shuhui Wang, and Qingming Huang. R&b: Region and boundary aware zero-shot grounded text-to-image generation. arXiv preprint arXiv:2310.08872, 2023. 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 293, + 294, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 293, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 57, + 293, + 294, + 357 + ], + "type": "text", + "content": "[50] Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wentian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7452-7461, 2023. 2, 3, 4" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 351 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 127 + ], + "type": "text", + "content": "[51] Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and Jian Zhang. Freedom: Training-free energy-guided conditional diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 23174-23184, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "type": "text", + "content": "[52] Yu Zeng, Vishal M Patel, Haochen Wang, Xun Huang, Ting-Chun Wang, Ming-Yu Liu, and Yogesh Balaji. Jedi: Joint-image diffusion models for finetuning-free personalized text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6786-6795, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 196, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 196, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 196, + 553, + 239 + ], + "type": "text", + "content": "[53] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR), pages 3836-3847, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 241, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 295 + ], + "type": "text", + "content": "[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 297, + 553, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 351 + ], + "type": "text", + "content": "[55] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. Advances in Neural Information Processing Systems (NeurIPS), 36, 2024. 2, 3" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13145" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Training-free Neural Architecture Search through Variance of Knowledge of Deep Network Weights/740054ef-86e4-4bac-b7b3-53d57b1f3a66_content_list.json b/2025/Training-free Neural Architecture Search through Variance of Knowledge of Deep Network Weights/740054ef-86e4-4bac-b7b3-53d57b1f3a66_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2eec9266aa082dd90254f81f28332c0ea2a92ec7 --- /dev/null +++ b/2025/Training-free Neural Architecture Search through Variance of Knowledge of Deep Network Weights/740054ef-86e4-4bac-b7b3-53d57b1f3a66_content_list.json @@ -0,0 +1,1815 @@ +[ + { + "type": "text", + "text": "Training-free Neural Architecture Search through Variance of Knowledge of Deep Network Weights", + "text_level": 1, + "bbox": [ + 114, + 128, + 883, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ondrej Tybl Lukas Neumann", + "bbox": [ + 362, + 203, + 671, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "CMP Visual Recognition Group, Faculty of Electrical Engineering, Czech Technical University in Prague", + "bbox": [ + 99, + 220, + 931, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ondrej.tybl@cvut.cz lukas.neumann@cvut.cz", + "bbox": [ + 313, + 242, + 712, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning has revolutionized computer vision, but it achieved its tremendous success using deep network architectures which are mostly hand-crafted and therefore likely suboptimal. Neural Architecture Search (NAS) aims to bridge this gap by following a well-defined optimization paradigm which systematically looks for the best architecture, given objective criterion such as maximal classification accuracy. The main limitation of NAS is however its astronomical computational cost, as it typically requires training each candidate network architecture from scratch.", + "bbox": [ + 89, + 321, + 483, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we aim to alleviate this limitation by proposing a novel training-free proxy for image classification accuracy based on Fisher Information. The proposed proxy has a strong theoretical background in statistics and it allows estimating expected image classification accuracy of a given deep network without training the network, thus significantly reducing computational cost of standard NAS algorithms.", + "bbox": [ + 89, + 474, + 483, + 594 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our training-free proxy achieves state-of-the-art results on three public datasets and in two search spaces, both when evaluated using previously proposed metrics, as well as using a new metric that we propose which we demonstrate is more informative for practical NAS applications. The source code is publicly available at https://www.github.com/ondratybl/VKDNW.", + "bbox": [ + 89, + 595, + 483, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 724, + 220, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In most instances, neural network architectures are designed by authors following the field's \"best-practices\" or their experience, without any formal and repeatable procedure. This is however inconvenient especially in applications on a large scale. Neural Architecture Search (NAS) aims to bridge this gap by following a well-defined optimization paradigm which systematically looks for the best architecture, given objective criterion such as maximal accuracy.", + "bbox": [ + 89, + 750, + 482, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The main limitation of Neural Architecture Search (NAS) is however the computational cost, because in the", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/97fd63a8b192a1c196b654c0292adafa60c228dc90dd62d9cefa4dc0ad195af9.jpg", + "image_caption": [ + "Figure 1. Training-free NAS methods on ImageNet16-120 [10]. Methods are compared by Normalized Discounted Cumulative Gain (see Sec. 3), our method (VKDNW) is the best also measured by Kendall's $\\tau$ and Spearman's $\\rho$ correlations (see Table 1). Also note that simple number of trainable layers (below denoted $\\aleph$ ) is significantly better trivial proxy than the number of FLOPs." + ], + "image_footnote": [], + "bbox": [ + 526, + 292, + 890, + 529 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "most basic NAS setup, it is required to train thousands or more of different deep network architectures from scratch in order just to calculate a single scalar – the objective function value, such as the classification accuracy. This severely limits practical applications of NAS as the size of feasible architecture search space is only a small fraction of the overall space of all networks.", + "bbox": [ + 511, + 643, + 906, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Training-free Neural Architecture Search (TF-NAS) aims to alleviate this limitation by introducing an objective function proxy which – unlike the actual objective function – does not require training the network. As a result, a good proxy allows the TF-NAS algorithm to explore significantly bigger portions of the network architecture search space compared to traditional NAS, and to find the best network architecture without training a single network. The crucial question is however finding an appropriate objective function proxy.", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14881", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we present a novel, principled objective function proxy called Variance of Knowledge of Deep Network Weights (VKDNW) for image classification accuracy, which allows us to find optimal network architectures for image classification without training them first. Our method is, to the best of our knowledge, the first successful application of Fisher Information theory [28] in the context of large deep neural networks, and as such allows us to formally describe and quantify the difficulty of network parameters' estimation process. In other words, given a network architecture, our method estimates how easy or hard it will be to train the network.", + "bbox": [ + 88, + 90, + 483, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, we also observe that the evaluation metrics used in the TF-NAS community are not well-suited to the problem at hand, because it unnecessarily penalizes for bad proxy accuracy for networks which are not interesting, and vice-versa it does not sufficiently reward proxies which are able to accurately pick out good network architectures. Following this observation, we propose that the Normalized Discounted Cumulative Gain should be used in comparison with other TF-NAS metrics, and show indeed that there are significant differences amongst previously proposed TF-NAS methods when the new metric is considered.", + "bbox": [ + 89, + 273, + 483, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, we make the following contributions:", + "bbox": [ + 109, + 441, + 459, + 457 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We introduce a novel algorithm for estimation of Fisher Information Matrix spectrum, which is tractable even for models with large number of parameters such as deep networks, that overcomes the usual problems of numerical stability.", + "2. We introduce a novel principled VKDNW proxy for image classification accuracy. The proxy is based on strong theoretical background and captures uncertainty in weight estimation process. It brings information that is orthogonal to the model size which then allows for efficient combination with previously proposed proxies, leading to state-of-the-art results.", + "3. We propose a new evaluation metric for TF-NAS proxies which is more relevant to the actual NAS objective as it concentrates on ability of given proxy to identify good networks." + ], + "bbox": [ + 89, + 459, + 483, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Method", + "text_level": 1, + "bbox": [ + 89, + 720, + 181, + 736 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our zero-shot proxy for image classification accuracy builds on Fisher Information theory [28], therefore we begin by a thorough analysis of Fisher Information Matrix (FIM) of the network weights estimation problem (see Sec. 2.1). We proceed by discussing challenges involved in practical application of FIM in context of large over-parametrized models that lead to only limited success in previous works and present our contributions to overcome these limitations (see Sec. 2.2). Finally, we propose a novel FIM-based proxy for NAS algorithms (Sec. 2.3).", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Fisher Information", + "text_level": 1, + "bbox": [ + 513, + 90, + 696, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The problem of finding the optimal weights of a neural network $f$ for the task of $C$ -class image classification can be seen as a maximum likelihood estimation with a statistical model", + "bbox": [ + 511, + 114, + 906, + 172 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {\\theta} (c \\mid x) = \\frac {\\exp (\\Psi_ {c} (x , \\theta))}{\\sum_ {d = 1} ^ {C} \\exp (\\Psi_ {d} (x , \\theta))}, \\quad c = 1, \\dots , C \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 176, + 906, + 214 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\sigma_{\\theta}(\\cdot |x)$ denotes the a posteriori distribution of the labels given input image $x$ and $\\Psi (x,\\theta)\\in \\mathbb{R}^{C} = (\\Psi_{1}(x,\\theta),\\dots ,\\Psi_{C}(x,\\theta))$ is the network output (logits) given the weight vector $\\theta \\in \\mathbb{R}^p$ , i.e. the network weights. We describe the process of training as finding the optimal weight vector $\\theta^{*}$ that fits our data and we posit that network architectures should be characterised by how easy it is to estimate their optimal network weights $\\theta^{*}$ . We build upon statistical learning theory and use Fisher Information [28] framework to formally describe expected behaviour of the training process of a given deep network $f$ .", + "bbox": [ + 511, + 219, + 906, + 385 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The Fisher Information Matrix (FIM) encompasses information on the difficulty of the parameter estimation problem and it plays a crucial role in several fundamental results, which we apply below in the context of deep networks. The FIM of a network $f$ and its set of weights $\\theta \\in \\mathbb{R}^p$ is given as", + "bbox": [ + 511, + 386, + 905, + 476 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nF (\\theta) := \\mathbb {E} \\left[ \\nabla_ {\\theta} \\sigma_ {\\theta} (c | x) \\nabla_ {\\theta} \\sigma_ {\\theta} (c | x) ^ {T} \\right] \\in \\mathbb {R} ^ {p \\times p}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 482, + 903, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where we take the expected value $\\mathbb{E}$ with respect to the joint distribution of $(x, c)$ . For more detailed account on Fisher Information theory and its applications in the context of machine learning, we kindly refer reader to [16, 20, 29, 35, 36].", + "bbox": [ + 511, + 506, + 905, + 568 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Cramér-Rao bound. The first part of estimation theory we build upon is the inverse of the FIM, known as the Cramér-Rao bound (see [11]), which is the asymptotic variance of the estimated weights (i.e. the uncertainty coming from the data variability). Thus, the larger the matrix norm of the FIM, the more certain we are about the weights $\\theta$ . More specifically, any data-dependent estimator $\\hat{\\theta}_n = \\hat{\\theta}_n(x_1,\\ldots ,x_n)$ is a random vector with randomness coming from the (independent) choice of input images $x_{1},\\ldots ,x_{n}$ and as such has some variance matrix $\\operatorname{Var}\\left(\\hat{\\theta}_n\\right)$ . The famous result (see [8, 37]) named in honour of H. Cramér and C. R. Rao states that if $\\hat{\\theta}_n$ is unbiased then the variance is bounded from below as", + "bbox": [ + 511, + 570, + 906, + 772 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {V a r} \\left(\\hat {\\theta} _ {n}\\right) \\geq \\frac {1}{n} F ^ {- 1} (\\theta) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 776, + 903, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and for maximum likelihood estimation this bound is attained as the number of input images grows to infinity $n \\to \\infty$ .", + "bbox": [ + 511, + 809, + 905, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We formally show (see Supplementary material) that for each weight $\\theta(j)$ the mean square error of our estimation is controlled by the diagonal element of the FIM inverse as", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14882", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} \\left(\\hat {\\theta} _ {n} (j) - \\theta (j)\\right) ^ {2} \\geq \\frac {1}{n} \\left(F ^ {- 1} (\\theta)\\right) _ {j j}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 109, + 482, + 138 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Therefore, knowing the FIM allows us to evaluate how certain we are about the weight estimates. We can go even further by inspecting the eigenvalues of $F(\\theta)$ . Denoting the largest and smallest eigenvalue of the FIM as $\\lambda_{\\mathrm{min}}$ and $\\lambda_{\\mathrm{max}}$ respectively, then we have that there exist linear combination coefficients $e_{\\mathrm{min}}$ and $e_{\\mathrm{max}}$ of unit size so that", + "bbox": [ + 89, + 143, + 483, + 233 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} \\left(e _ {\\min} ^ {T} \\hat {\\theta} _ {n} - e _ {\\min} ^ {T} \\theta\\right) ^ {2} \\geq \\frac {1}{n \\lambda_ {\\min}},\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 238, + 395, + 268 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} \\left(e _ {\\max } ^ {T} \\hat {\\theta} _ {n} - e _ {\\max } ^ {T} \\theta\\right) ^ {2} \\geq \\frac {1}{n \\lambda_ {\\max }}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 271, + 480, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "indicating that if the difference between the largest and smallest eigenvalue is large then there exist combinations of weights with very large difference in the estimation certainty. Altogether, the more the eigenvalues of the FIM are similar, the more similar is also the variance in the weight estimation across all model weights.", + "bbox": [ + 89, + 306, + 483, + 397 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prediction sensibility. The change of the model prediction measured by the KL-divergence when subject to a small perturbations of the weights is given as", + "bbox": [ + 89, + 401, + 483, + 446 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nD _ {K L} \\left(\\sigma_ {\\theta + \\theta_ {\\delta}} (\\cdot | x), \\sigma_ {\\theta} (\\cdot | x)\\right) \\approx \\frac {1}{2} \\theta_ {\\delta} ^ {T} F (\\theta) \\theta_ {\\delta} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 450, + 482, + 481 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for a small weight perturbation vector $\\theta_{\\delta} \\in \\mathbb{R}^{p}$ . Thus, $F(\\theta)$ measures volatility of the predictions subject to the weight change – if the difference between the largest and the smallest eigenvalues of the FIM is large, then there exist perturbation directions $\\theta_{min}, \\theta_{max}$ corresponding to $\\lambda_{min}, \\lambda_{max}$ such that", + "bbox": [ + 89, + 486, + 483, + 575 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} D _ {K L} \\left(\\sigma_ {\\theta + \\theta_ {m i n}} (\\cdot | x), \\sigma_ {\\theta} (\\cdot | x)\\right) \\approx \\frac {1}{2} \\theta_ {m i n} ^ {T} F (\\theta) \\theta_ {m i n} = \\\\ = \\frac {1}{2} \\lambda_ {\\min } \\| \\theta_ {\\min } \\| ^ {2} \\ll \\frac {1}{2} \\lambda_ {\\max } \\| \\theta_ {\\max } \\| ^ {2} = \\tag {7} \\\\ = \\frac {1}{2} \\theta_ {m a x} ^ {T} F (\\theta) \\theta_ {m a x} \\approx D _ {K L} \\left(\\sigma_ {\\theta + \\theta_ {m a x}} (\\cdot | x), \\sigma_ {\\theta} (\\cdot | x)\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 580, + 493, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and therefore in some directions a small change of the weights has much larger impact on the prediction than in others, making the model less balanced. For further discussion, see the Supplementary material.", + "bbox": [ + 89, + 672, + 483, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Empirical Fisher Information Matrix implementation", + "text_level": 1, + "bbox": [ + 89, + 746, + 483, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When independent and identically distributed sample images $x_{n}$ are available, empirical FIM $\\hat{F} (\\theta)$ is defined as", + "bbox": [ + 89, + 786, + 483, + 816 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {F} (\\theta) := \\frac {1}{n} \\sum_ {n = 1} ^ {N} \\mathbb {E} _ {\\sigma_ {\\theta}} \\left[ \\nabla_ {\\theta} \\sigma_ {\\theta} (c | x _ {n}) \\nabla_ {\\theta} \\sigma_ {\\theta} (c | x _ {n}) ^ {T} \\right] \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 824, + 482, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbb{E}_{\\sigma_{\\theta}}$ now denotes the expectation with respect to the model prediction $\\sigma_{\\theta}$ .", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We would like to first emphasize several crucial aspects and contributions of this paper that lead to the first success of the Fisher Information theory (having e.g. [1] in mind) in the context of Neural Architecture Search, despite Fisher Information being one of the first obvious choices for deep network analysis:", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1. Following [19] we write the empirical FIM (8) as", + "bbox": [ + 532, + 185, + 877, + 200 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {F} (\\theta) = \\frac {1}{n} \\sum_ {n = 1} ^ {N} \\left[ \\nabla_ {\\theta} \\Psi (x _ {n}, \\theta) ^ {T} \\left(\\operatorname {d i a g} \\left(\\sigma_ {\\theta} (\\cdot , x _ {n})\\right) - \\right. \\right. \\\\ \\left. \\sigma_ {\\theta} (\\cdot , x _ {n}) \\sigma_ {\\theta} (\\cdot , x _ {n}) ^ {T}\\right) \\nabla_ {\\theta} \\Psi (x _ {n}, \\theta) ]. \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 207, + 903, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and we further decompose the inner matrix $\\mathrm{diag}(\\sigma_{\\theta}(\\cdot ,x_n)) - \\sigma_{\\theta}(\\cdot ,x_n)\\sigma_{\\theta}(\\cdot ,x_n)^T$ using analytical formulas from [40] to avoid numerical instability of the computation as we arrive at a feasible representation", + "bbox": [ + 511, + 275, + 905, + 335 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {F} (\\theta) = \\frac {1}{n} \\sum_ {n = 1} ^ {N} A _ {n} ^ {T} A _ {n} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 342, + 903, + 382 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for some matrices $A_{n}\\in \\mathbb{R}^{C\\times p}$ . We observed that networks typically yield very imbalanced outputs at initialization (i.e. every network prioritizes few classes over the rest) making it extremely important not to exclude the factor $\\mathrm{diag}(\\sigma_{\\theta}(\\cdot ,x_n)) - \\sigma_{\\theta}(\\cdot ,x_n)\\sigma_{\\theta}(\\cdot ,x_n)^T$ , which is however difficult to compute without underflow/overflow.", + "bbox": [ + 511, + 388, + 905, + 479 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2. The dimension of FIM is equal to $p$ (the number of all trainable parameters) and therefore direct computation of the eigenvalues is numerically intractable and unstable. However, our results show that if a small number of representative parameters is drawn, then computation becomes stable while the discrimination power does not suffer. We used a simple rule where a single parameter from each trainable layer (not including batch normalization) is chosen. Stability with respect to choice of such parameters can be found in Supplementary material.", + "3. FIM of large networks typically suffers from having pathological spectrum, i.e. there usually exist zero eigenvalue and its multiplicity is large (see [15]) and thus the eigenvalues estimation due to a large condition number is imprecise. However, as we deal with a symmetric positive-semidefinite matrix, the eigenvalues actually coincide with the singular values (see [24]), for which the estimation algorithm performs better." + ], + "bbox": [ + 511, + 484, + 905, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let us also emphasize that there is a common misconception in part of the community as it is often mistakenly assumed that one can simply use the true labels of $x_{n}$ in (8) in place of $c$ . However, such definition is then meaningless as it does not approximate the FIM in the classical Monte Carlo sense (see Supplementary material and [19] for a comparison). That means that the empirical FIM does not depend on the true labels and therefore our method does not require real data - we use random input instead.", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14883", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Variance of Knowledge for Deep Network Weights", + "text_level": 1, + "bbox": [ + 89, + 89, + 482, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to characterize properties of the parameter estimation process for a given network $f$ through the lens of Fisher Information theory, we inspect the eigenvalues of the empirical FIM $\\hat{F}(\\theta_{\\mathrm{init}})$ and define the entropy of Variance of Knowledge for Deep Network Weights (VKDNW) as", + "bbox": [ + 89, + 130, + 485, + 205 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {V K D N W} (f) := - \\sum_ {k = 1} ^ {9} \\tilde {\\lambda} _ {k} \\log \\tilde {\\lambda} _ {k} \\\\ \\tilde {\\lambda} _ {k} = \\frac {\\lambda_ {k}}{\\sum_ {j = 1} ^ {9} \\lambda_ {j}}, \\quad k = 1, \\dots , 9, \\tag {11} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 212, + 482, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_{k}$ denotes the $k$ -th decile of the FIM eigenvalues as the representation of the FIM spectrum1, and $\\theta_{\\mathrm{init}}$ denotes network weights at initialization.", + "bbox": [ + 89, + 297, + 482, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our score therefore measures the diversity of the FIM eigenvalues, and from the entropy theory we know that VKDNW attains its maximum exactly when all the eigenvalues $\\lambda_{k}$ are equal and VKDNW gets lower as the eigenvalues become more different. Based on the discussion in Sec. 2.1 we see that VKDNW is high when the uncertainty in all model weight combinations are similar (see Cramér-Rao bound) and there are no directions in the weight space that would influence the network prediction substantially differently than others. Due to the fact that we have normalized both the number of eigenvalues under consideration (by taking a fixed number of representatives irrespective of the network size) and the magnitude of the eigenvalues, VKDNW is independent of network size (number of network weights $p$ ).", + "bbox": [ + 88, + 343, + 483, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let us note, that we are familiar with the fact that even though our motivation was (among others) based on Cramér-Rao bound that assumes evaluation of FIM at the correct weight vector $\\theta_{X}$ that fits the data, which is typically far away from the weights given at initialization. However, our empirical results below support the hypothesis that the evaluation despite being in the wrong point brings valuable information.", + "bbox": [ + 89, + 571, + 483, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ranking networks for NAS. The proposed VKDNW score is independent of network size, which is extremely beneficial to compare individual structures of network architectures. Thus, it does not aim to capture capacity, rather it targets feasibility of the computation graph given number of operations. However, when comparing different network structures and different network sizes together as it is done in NAS, one indeed needs to take network size into account as well, because naturally larger networks have bigger capacity and therefore tend to have higher accuracy. To", + "bbox": [ + 89, + 696, + 483, + 848 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d5f967d8c0886a78528adbb6ab92ca4e59bafe67932a202795c93e69628d8d76.jpg", + "image_caption": [ + "Figure 2. Toy example of two rankings on 10 networks. We plot accuracies ordered by the rankings and evaluation metrics Kendall's $\\tau$ (KT) and Spearman's $\\rho$ (SPR) correlations and Normalized Discounted Cumulative Gain $(\\mathrm{nDCG}_5)$ ." + ], + "image_footnote": [], + "bbox": [ + 521, + 93, + 898, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "capture also the capacity we proxy the network size by the number of layers with weights that we denote as $\\aleph(f)$ for a network $f$ and we introduce the ranking", + "bbox": [ + 511, + 316, + 906, + 362 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {V K D N W} _ {\\text {s i n g l e}} (f) := \\aleph (f) + \\operatorname {V K D N W} (f) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 368, + 903, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here we leverage the fact that VKDNW as an entropy of some quantity is always between 0 and 1 and by summing it with an integer-valued quantity we in fact obtain that we have first grouped networks by our size proxy $\\aleph$ and then within each group of similar networks sizes we order them by VKDNW.", + "bbox": [ + 511, + 392, + 906, + 482 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 513, + 500, + 697, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we dive into the problem of evaluating (image) classification accuracy proxies for training-free NAS (TF-NAS). In NAS algorithms, the proxy actually does not have to predict the classification accuracy in absolute terms – since we are after finding the \"best\" network for given task – it's sufficient that the proxy properly ranks the networks, ordering them from worst to the best.", + "bbox": [ + 511, + 525, + 905, + 631 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Suppose that a collection of $K$ networks with validation accuracies $\\mathrm{acc}_1, \\ldots, \\mathrm{acc}_K$ have been ranked by a TF-NAS proxy as $r_1, \\ldots, r_K \\in \\mathbb{R}$ . The standard way of evaluating [14, 21, 25] how well the ranks correspond to the accuracies is to compute Kendall's $\\tau$ (KT, [17]) and Spearman's $\\rho$ (SPR, [39]) rank correlation coefficients. Kendall's $\\tau$ is given as", + "bbox": [ + 511, + 632, + 905, + 737 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tau := \\frac {n _ {c} - n _ {d}}{\\sqrt {n _ {c} + n _ {d} + n _ {1}} \\sqrt {n _ {c} + n _ {d} + n _ {2}}}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 744, + 906, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $n_c$ denotes the number of concordant pairs $(\\mathrm{acc}_k, r_k)$ , $n_d$ the number of discordant pairs and $n_1$ (resp. $n_2$ ) denote the number of ties in $\\mathrm{acc}_k$ only (resp. in $r_k$ only). In our context, (KT) relates after some normalization to a probability that two randomly chosen network rankings $r_k, r_l$ are ordered correctly according to their accuracies $\\mathrm{acc}_k, \\mathrm{acc}_l$ . In contrast, (SPR) is defined to account more for outliers, i.e. it tends to put a large penalization if", + "bbox": [ + 511, + 780, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1As explained in Sec. 2.2, the smallest eigenvalue $\\lambda_0$ is usually equal to zero and thus we exclude it for stability reasons (similarly with the maximal eigenvalue $\\lambda_{10}$ )", + "bbox": [ + 89, + 862, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14884", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "there are some networks for which the difference between the rank of the accuracy $\\mathrm{acc}_k$ and the rank $r_k$ is large as it is just the classical correlation coefficient however applied on the orders of the assessed quantities.", + "bbox": [ + 88, + 90, + 480, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Normalized Discounted Cumulative Gain. Both (KT) and (SPR) are good evaluation metrics when our interest lies in comparison of ranking proxies when all networks involved are of the same importance. However, in NAS we are primarily interested if a ranking helps us to pick best networks from a given collection and less interested how well the ranking compares two networks with low accuracy. Finding inspiration in information retrieval where one measures the quality of a system that retrieves resources relevant to an input query, we propose to use Normalized Discounted Cumulative Gain (nDCG) [2] as a more relevant metric to measure quality of TF-NAS proxies. The nDCG metric is defined with a key requirement that highly relevant documents are more valuable when they appear earlier in the search engine results (i.e., in higher-ranking positions). This requirement can be reformulated for the task of neural architecture search as networks with high accuracies are more valuable when they appear on higher-ranking positions.", + "bbox": [ + 93, + 154, + 485, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We thus define the metric as follows: first we order the networks by their ranking $r_1, \\ldots, r_K$ so that $r_{k_1} = K, r_{k_2} = K - 1, \\ldots$ and we compute", + "bbox": [ + 89, + 440, + 482, + 487 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {n D C G} _ {P} := \\frac {1}{Z} \\left(\\sum_ {j = 1} ^ {P} \\frac {2 ^ {\\mathrm {a c c} _ {k _ {j}}} - 1}{\\log_ {2} (1 + j)}\\right), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 493, + 482, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $P \\in \\mathbb{N}$ is a parameter determining how many top-ranked networks do we consider (e.g. it corresponds to the population size in an evolution algorithm) and $Z$ is a normalization factor that represents the ideal discounted cumulative gain so that nDCG is equal to one for a perfect fit.", + "bbox": [ + 89, + 545, + 482, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The higher nDCG the better the ranking is as it is a weighted average of transformed top-ranked accuracies. In [47] it is shown that the choice of the discount factor of given by inverse logarithm is a good choice as it nDCG then well separates different ranking systems.", + "bbox": [ + 89, + 621, + 482, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Toy Example. Let us briefly demonstrate the weak ability of (KT) and (SPR) to distinguish rankings that poorly discriminate networks with high accuracy. Suppose a collection of 10 networks with their validation accuracies $100,90,\\dots ,10$ is given and our aim is to rank them from worst to the best. We evaluate two different toy rankings: a) ranking damaged_top which perfectly fits the accuracies, only the two top networks are swapped with the third and fourth, b) ranking perfect_top which fits perfectly for the top-performing networks, however, now the order of the four worst networks is reversed (see Figure 2). The", + "bbox": [ + 89, + 699, + 482, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "damaged_top ranking is an example that should be evaluated worse for the architecture search task compared to perfect_top as in the first case the best networks are not placed first, while in the second they are. On the other hand, even though perfect_top is not a perfect fit, it still discriminates the top networks ideally and only struggles for networks with low accuracy, which are not interesting for NAS which is after the best networks. From the correlation perspectives of (KT), (SPR) damaged_top is better than perfect_top, therefore, if we rely only on these two evaluation metrics we prefer ranking that is worse for the architecture search task as the top networks are ranked worse. On the other hand, for perfect_top we obtain ideal nDCG $_5$ while it drops to 0.5 for damaged_top. We conclude that using nDCG $_5$ we choose ranking that has higher discriminatory power for good networks.", + "bbox": [ + 511, + 90, + 903, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also note that from statistical perspective the (KT) and (SPR) of damaged_top is significantly higher than of a random ranking3, and therefore we'd conclude that damaged_top is not independent of ground truth accuracy. On the other hand, when we perform the same statistical test on damaged_top using nDCG5, we conclude that the damaged_top is not significantly better than ranking networks randomly4.", + "bbox": [ + 511, + 333, + 903, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Related Work", + "text_level": 1, + "bbox": [ + 511, + 470, + 653, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Zero-shot NAS aims to rank given networks in a training-free manner based on their (a priori unknown) final performance, which allows to prune the huge search space with limited costs when seeking for optimal architecture [10, 27, 46] and other configuration [3, 26, 38].", + "bbox": [ + 511, + 491, + 903, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Throughout previous works different approaches for ranking computation can be found. In many of the works the gradient with respect to the network weights is investigated (GradNorm, GraSP, SNIP, Synflow, [1, 23, 41, 42]), i.e. leveraging the first order approximation of the network. In Jacov [1] correlations of the Jacobian matrices among various input samples are compared; in NASWOT [31] the linear maps induced by data points are examined; Zen [26] uses the approximation of gradient with respect to featuremaps; GradSign [49] compares the optimization landscape at the level individual training samples; or in ZiCo [25] the gradients from multiple forward and backward passes preferring large magnitude and low variance.", + "bbox": [ + 511, + 568, + 905, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Other methods aggregate multiple sources aiming at obtaining a better informed ranks: TE-NAS [6] uses both the number of linear regions [12, 44] and the condition number of Neural Tangent Kernel [13, 22]. However, it is well", + "bbox": [ + 511, + 765, + 903, + 825 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3That is, the $p$ -value for the hypothesis that damaged_top is assigned independently of the accuracies is below 0.005.", + "bbox": [ + 511, + 838, + 903, + 863 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "4Running 1000 samples on a random ranking we obtain that damaged_top has nDCG₅ around 75th percentile of such random evaluations and it is not significantly better than a random ranking.", + "bbox": [ + 511, + 863, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2In case of ties we take random ordering within the groups (parameter ignore_ties=True in the scikit-learn implementation).", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14885", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4d18da440a9408c83eac99f52639f21bc648ac987fd0daacff6757ded5e0f812.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypeCIFAR-10CIFAR-100ImageNet16-120
KTSPRnDCGKTSPRnDCGKTSPRnDCG
Simple rankings
FLOPsS0.6230.7990.7450.5860.7630.5760.5450.7180.403
GradNorm [1]S0.3280.4380.5090.3410.4510.2780.3100.4180.265
GraSP [1, 42]S0.3520.5050.5180.3490.4980.2840.3590.5020.281
SNIP [1, 23]S0.4310.5910.5130.4400.5970.2860.3890.5210.286
SynFlow [1, 41]S0.5610.7580.7090.5530.7500.5940.5310.7190.511
Jacov [1]S0.6160.8000.5400.6390.8200.4020.6020.7790.356
NASWOT [31]S0.5710.7620.6070.6070.7990.4750.6050.7940.490
ZenNAS [26]S0.1020.1030.1200.0790.0720.1150.0910.1090.073
GradSign† [49]S·0.765··0.793··0.783·
ZiCo [25]S0.6070.8020.7510.6140.8090.6070.5870.7790.523
TE-NAS [6]A0.5360.7220.6020.5370.7230.3270.5230.7090.330
AZ-NAS [21]A0.7120.8920.7490.6960.8800.5490.6730.8590.534
No. of trainable layers (N)S0.6260.7670.6710.6460.7870.5250.6230.7640.497
VKDNWsingle (ours)S0.6180.8150.7510.6340.8290.6170.6220.8140.608
VKDNWagg (ours)A0.7500.9190.7850.7530.9190.6360.7430.9060.664
Model-driven rankings
GRAF [14]A0.8200.9530.9350.8090.9480.8580.7960.9410.828
VKDNWm (ours)A0.6470.8310.7500.6360.8210.6020.6110.7980.575
(VKDNW+ZCS)m (ours)A0.8400.9630.9220.8340.9600.8840.8300.9580.843
(VKDNW+ZCS+GRAF)m (ours)A0.8590.9710.9460.8470.9660.8950.8420.9630.867
", + "bbox": [ + 89, + 87, + 915, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Training-free NAS methods in the NAS-Bench-201 [10] search space, evaluated on three public datasets. Kendall's $\\tau$ (KT), Spearman's $\\rho$ (SPR) and Normalized Discounted Cumulative Gain (nDCG) are reported, results are averages of 5 independent runs (all stds are $< 0.02$ ). The Type column differentiates single (S) and aggregated (A) rankings. Results are reproduced with code published by their authors, except those marked†, where results from the original paper are taken.", + "bbox": [ + 89, + 450, + 906, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "known that the kernel computation is highly computationally demanding [34]. AZ-NAS [21] assesses expressivity, trainability and progressivity via examination of feature distribution across all orientations and the Jacobian. However, despite AZ-NAS outperforming previous works in some metrics, it is worse than [25] in key NAS-related aspects such as the cumulative gain (see Sec. 5.1).", + "bbox": [ + 89, + 515, + 482, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Despite a tremendous effort of the community, it was shown that most of the zero-shot NAS methods perform worse than a simple proxy given just by FLOPs or #params [33, 43]. Thus, there is still a gap for improvements also driven by the need of the ranking explainability that would have a satisfactory theoretical support.", + "bbox": [ + 89, + 623, + 482, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, [14] uses a model-driven ranking that however needs a set of networks for which the validation accuracies are known to train the model and the ability of this ranking to generalize as it is fitted on a specific network collection only is disputable.", + "bbox": [ + 89, + 715, + 482, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, let us discuss the current practice in the NAS methods comparison. Methods are compared either by means of correlations of their scores to accuracy or by reporting the accuracy of the top-ranked network [21, 25]. However, while the first is not tailor-made for the architecture search task and therefore does not assess the desired ranking properties (see Sec. 5.1), comparing performance", + "bbox": [ + 89, + 795, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "just by the accuracy of a single network is too vulnerable e.g. to the random seed choice.", + "bbox": [ + 511, + 515, + 906, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 560, + 648, + 577 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Ranking aggregation", + "text_level": 1, + "bbox": [ + 511, + 585, + 712, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to using the single ranking of Eq. (12), we also experiment with multiple rankings (as in [14, 21, 25]) to order network architectures by their accuracy. We use two different options: non-linear and model-driven aggregation.", + "bbox": [ + 511, + 609, + 906, + 670 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Non-linear aggregation. The aggregation [21] uses multiplication to combine multiple ranks into a single one, which means a network is highly-ranked if and only if it is highly-ranked in all subsidiary rankings, keeping their influence balanced. That is, denoting rankings $\\mathrm{rank}_1,\\ldots ,\\mathrm{rank}_m$ for some $m\\in \\mathbb{N}$ we define the aggregated ranking", + "bbox": [ + 511, + 672, + 905, + 779 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {r a n k} _ {\\text {a g g}} (f) := \\log \\Pi_ {j = 1} ^ {m} \\operatorname {r a n k} _ {j} (f) \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 785, + 903, + 804 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for each network $f$ . This aggregation is therefore possible only in the context of a given network collection, such as in evolutionary search.", + "bbox": [ + 511, + 809, + 906, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our case, $\\mathrm{VKDNW_{agg}}$ aggregates these five proxies:", + "bbox": [ + 532, + 854, + 890, + 869 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- VKDNWsingle (V) ranking (Eq. (12)),", + "- Jacov (J) [1] measures the activations correlation when" + ], + "bbox": [ + 513, + 869, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14886", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0cf8f611ddc7ee1e2f24d6c36dfc6f60993f5686efab1b7dcf90db421be42681.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFLOPsTop-1 acc.TypeSearch cost (GPU days)
NASNet-B [50]488M72.8MS1800
CARS-D [45]496M73.3MS0.4
BN-NAS [5]470M75.7MS0.8
OFA [4]406M77.7OS50
RLNAS [48]473M75.6OS-
DONNA [32]501M78.0OS405
# Params451M63.5ZS0.02
ZiCo [25]448M78.1ZS0.4
AZ-NAS [21]462M78.6ZS0.4
VKDNWagg(ours)480M78.8ZS0.4
", + "bbox": [ + 89, + 88, + 486, + 271 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Results on ImageNet-1K [9] in the MobileNetV2 search space, the size of the model is constrained to $\\approx 450\\mathrm{M}$ FLOPS.", + "bbox": [ + 89, + 281, + 482, + 310 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "exposed to various inputs,", + "bbox": [ + 104, + 318, + 279, + 332 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Expressivity (E) [21] assesses isotropy and uniformity of the features distribution across all orientations,", + "- Trainability (T) [21] captures ability of the network to keep stable gradient propagation between the layers by inspecting the spectrum of the Jacobian matrix,", + "- FLOPs (F) is the number of FLOPs for one forward pass." + ], + "bbox": [ + 89, + 333, + 482, + 424 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Model-driven aggregation. When accuracies of a sufficient number of architectures in the search space are known, model-driven aggregation can be used to train a regression model to combine individual rankings. The trained model is then used to predict accuracy for unseen networks in the same search space. We evaluate three different models:", + "bbox": [ + 89, + 426, + 482, + 516 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- VKDNW $_m$ where the eigenvalues $\\lambda_k$ of the FIM matrix are used directly as features in companion with $\\aleph$ (number of trainable layers) and FLOPs to allow a more complex proxy of the diversity of the eigenvalues and the network complexity than the simple entropy Eq. (11),", + "- $(\\mathrm{VKDNW} + \\mathrm{ZCS})_m$ where we additionally include all other zero-cost scores available (see Table 1),", + "- (VKDNW+ZCS+GRAF) $_m$ where we add network graph features from [14]." + ], + "bbox": [ + 89, + 517, + 482, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Results", + "text_level": 1, + "bbox": [ + 89, + 662, + 184, + 678 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We have conducted experiments in the NAS-Bench-201 [10] and MobileNetV2 [26, 38] search spaces. To obtain easily comparable results, we used 64 randomly generated input images to compute our score as in [21]. For the methods that rely on the knowledge of true labels, input data from the respective datasets were used.", + "bbox": [ + 89, + 685, + 482, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "NAS-Bench-201. The dataset consists of 15,625 networks for which validation accuracies for CIFAR-10, CIFAR-100 [18] and ImageNet16-120 [7] after training for 200 epochs are provided. The networks are characterized by unique cell structures comprising of several types of operation choices. As one of the possible choices is zero operation, it's possible that some computation edges don't receive any input or cannot propagate their results to the out", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8b64c7352464abe41619d6bfc3723ff08879a9dd2ad85c19d55acf23710d5d61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 92, + 705, + 233 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e4f5f01c1888fabdc9b7380bd8dd2f34a99e4c8ec21c3a7473348ecbc2f28be5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 92, + 898, + 233 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d77d22122e5743a9ff75154781478d7c808cfa3c1f380a6c73f447a4dd5c9bd9.jpg", + "image_caption": [ + "Figure 3. Components of AZ-NAS [21] and our VKDNW are compared w.r.t. correlation with $\\aleph$ (number of trainable layers), in the NAS-Bench-201 search space [10] on ImageNet16-120 [7] dataset. Our VKDNW proxy has the lowest correlation, ie. is the most invariant to the size of the model." + ], + "image_footnote": [], + "bbox": [ + 517, + 243, + 702, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c1dcc1c3cc570de06fc5b5da33871e3d569e4802867351dd28e012ea923c7e9b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 243, + 898, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "put, leading to same computation graphs for different architectures, thus duplicating networks. Following the practice from NAS-Bench-101 [46], we report results on 9,445 unique structures (also in [14, 30]) and we refer the reader to Supplementary material for results on all networks. We measure proxies performance via Kendall's $\\tau$ (KT) and Spearman's $\\rho$ (SPR) correlations with validation accuracies together with Normalized Discounted Cumulative Gain $(\\mathrm{nDCG}_{1000})$ , we write nDCG for short) from Sec. 3, averaged over 5 independent runs. For the model-driven aggregation, we trained a random forest model on 1024 networks for 100 iterations, and used the rest for testing.", + "bbox": [ + 511, + 484, + 906, + 665 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 1, we compare our method with existing NAS methods and show that $\\mathrm{VKDNW_{agg}}$ outperforms all methods with a significant margin in all three metrics. The benefit of our method is two-fold: a) it identifies the best networks due to its high performance in nDCG, which is the desirable property in NAS b) the ranking is consistent across the whole network search space due to high KT and SPR correlations. Note that we can only make these observations by combining standard correlation metrics and our newly proposed nDCG. We also show that $\\mathrm{VKDNW_{single}}$ outperforms all other single-rank proxies in all metrics on ImageNet16-120, in SPR on all three datasets, while being among the highest on CIFAR-10 and CIFAR-100.", + "bbox": [ + 511, + 667, + 908, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MobileNetV2. We search for the best network configuration in the MobileNetV2 space [38], while constraining", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14887", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a1cef225d49c3567b59bf248158c844b16ae9762623e10fb25ab5bec6fb8a27e.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Components of VKDNW $_{\\text{agg}}$ rank with non-linear aggregation. Consistency is shown with respect to Kendall's $\\tau$ (KT), Spearman's $\\rho$ (SPR) and Normalized Discounted Cumulative Gain (nDCG) with $P = 1000$ on ImageNet16-120 image dataset [7]. Here V, J, E, T and F stand for VKDNW, Jacov, expressivity, trainability and FLOPs respectively (see Sec. 5.1). Table of all combinations can be found in the Supplementary material." + ], + "table_body": "
VJETF(KT)(SPR)nDCG
✓0.6220.8140.608
✓0.6030.7810.339
✓0.5880.7790.274
✓0.3530.5170.233
✓0.5450.7180.403
✓✓✓✓0.7170.8910.623
✓✓✓✓0.7060.8710.553
✓✓✓✓0.7360.9050.695
✓✓✓✓0.7220.8960.658
✓✓✓✓0.7350.9010.646
✓✓✓✓✓0.7430.9060.664
", + "bbox": [ + 122, + 88, + 452, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the model size to approximately 450M FLOPs. We ran 100,000 iterations of the evolutionary search algorithm [21] for approximately 10 hours, meaning 100,000 different architectures were evaluated. We then took the best network from the search and trained it for 480 epochs on ImageNet-1K [9], using the same hyper-parameter setting as in [21, 25]. The final training of the model took 7 days on 8xNVidia A100 GPUs.", + "bbox": [ + 89, + 376, + 482, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As seen in Table 2, our method $\\mathrm{VKDNW_{agg}}$ outperforms all prior approaches - even train-based approaches (denoted MS and OS) that incur much higher computational costs for the search.", + "bbox": [ + 89, + 497, + 483, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablations", + "text_level": 1, + "bbox": [ + 89, + 568, + 202, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We ablate our method in the NAS-Bench-201 search space [10] on ImageNet16-120 [7] validation accuracies.", + "bbox": [ + 89, + 592, + 482, + 622 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Orthogonality of VKDNW. Our score VKDNW has a desirable property that it is based on information orthogonal to the size of the network: in Figure 3 we can see that unlike previous work, VKDNW is not correlated with the network size measured by $\\aleph$ (number of trainable layers). We believe this property is key when stepping into much larger search spaces, however components of the previous state-of-the-art method AZ-NAS lack such a property (also Fig. 3 and Tab. 4). We conjecture that improvement in the key metrics by VKDNW is caused by this orthogonality feature. In Table 1 we also provide comparison to previous model-driven method and show that adding our feature significantly improves the ranking in all considered metrics.", + "bbox": [ + 89, + 626, + 482, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Components of aggregated rank. Our single-rank variant $\\mathrm{VKDNW}_{\\mathrm{single}}$ is the strongest component of the aggregated $\\mathrm{VKDNW}_{\\mathrm{agg}}$ as can be seen in Table 3 where it outperforms rest of the rankings with especially large margin in nDCG, thus, it strongly discriminates high-accuracy net", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cb7bf85adbcaf5499790f402df0eef7735b89c2f10c6ac0befb9e308c88401ed.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. Kendall's $\\tau$ (KT) of $\\aleph$ (number of trainable layers) with components of AZ-NAS compared to our new method VKDNW on three datasets on NAS-Bench-201 search space [10] and multiple image datasets." + ], + "table_body": "
(KT)
CIFAR-10CIFAR-100ImageNet16-120
VKDNW0.041-0.090-0.107
trainability0.2200.2200.253
expressivity0.5390.5390.539
progressivity0.3980.3980.385
", + "bbox": [ + 516, + 88, + 903, + 176 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/04bdf7d7ea1cdd7aa227334d50ef65bf3925da68f1c02ba7bb191aabfce84c9a.jpg", + "table_caption": [], + "table_footnote": [ + "Table 5. Our method VKDNW evaluated for different batch sizes using either randomly generated input data or real images with respect to Kendall's $\\tau$ (KT), Spearman's $\\rho$ (SPR) and Normalized Discounted Cumulative Gain with $P = 1000$ (nDCG)." + ], + "table_body": "
Batch SizeRandom InputReal Input
KTSPRnDCGKTSPRnDCG
80.6150.8080.6170.6140.8050.599
160.6140.8080.6040.6080.8000.595
320.6190.8120.6110.6120.8040.594
640.6210.8140.6170.6140.8060.595
1280.6210.8140.6160.6150.8060.591
", + "bbox": [ + 517, + 260, + 903, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "works compared to others.", + "bbox": [ + 511, + 441, + 691, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Random or real input. As our method is computed using generated random data (white noise) it does not rely on any real dataset and is therefore applicable also in situations where no reliable data are available. Table 5 shows that by relying just on random data we do not lose any performance with respect to all key metrics. Moreover, the performance remains relatively stable across different batch sizes. We chose batch size 64 as larger batch size does not bring better results.", + "bbox": [ + 511, + 460, + 906, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Due to lack of space, we kindly refer the reader to Supplementary material for further ablations.", + "bbox": [ + 511, + 599, + 905, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 643, + 633, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed a new training-free NAS proxy called Variance of Knowledge of Deep Network Weights (VKDNW). The method has strong theoretical support and achieved state-of-the-art results on three public datasets and in two search spaces.", + "bbox": [ + 511, + 671, + 905, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Based on throughout evaluation and comparison to other existing approaches, it has been shown that it provides an information orthogonal to the network size leading to a zero-cost ranking were contribution of the network size and architecture feasibility are separated. We have also shown that previously used correlation metrics for proxy evaluations do not sufficiently assess the key ability to discriminate top networks and to address this problem we have proposed a new evaluation metric and re-evaluated previous methods with the new metric.", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14888", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The research was supported by Czech Science Foundation Grant No. 24-10738M. The access to the computational infrastructure of the OP VVV funded project CZ.02.1.01/0.0/0.0/16_019/0000765 “Research Center for Informatics” and of the Ministry of Education, Youth and Sports of the Czech Republic project e-INFRA CZ (ID:90254) is also gratefully acknowledged.", + "bbox": [ + 89, + 114, + 483, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 232, + 187, + 248 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mohamed S Abdelfattah, Abhinav Mehrotra, Lukasz Dudziak, and Nicholas D Lane. Zero-cost proxies for lightweight NAS. arXiv preprint arXiv:2101.08134, 2021.", + "[2] Chris Burges, Tal Shaked, Erin Renshaw, Ari Lazier, Matt Deeds, Nicole Hamilton, and Greg Hullender. Learning to rank using gradient descent. In Proceedings of the 22nd international conference on Machine learning, pages 89-96, 2005.", + "[3] Han Cai, Ligeng Zhu, and Song Han. Proxylessnas: Direct neural architecture search on target task and hardware. arXiv preprint arXiv:1812.00332, 2018.", + "[4] Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. Once-for-all: Train one network and specialize it for efficient deployment. arXiv preprint arXiv:1908.09791, 2019.", + "[5] Boyu Chen, Peixia Li, Baopu Li, Chen Lin, Chuming Li, Ming Sun, Junjie Yan, and Wanli Ouyang. BN-NAS: Neural architecture search with batch normalization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 307-316, 2021.", + "[6] Wuyang Chen, Xinyu Gong, and Zhangyang Wang. Neural architecture search on imagenet in fourgpu hours: A theoretically inspired perspective. arXiv preprint arXiv:2102.11535, 2021.", + "[7] Patryk Chrabaszcz, Ilya Loshchilov, and Frank Hutter. A downsampled variant of imagenet as an alternative to the cifar datasets. arXiv preprint arXiv:1707.08819, 2017.", + "[8] Harald Cramér. Mathematical methods of statistics. Princeton university press, 1999.", + "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009.", + "[10] Xuanyi Dong and Yi Yang. Nas-bench-201: Extending the scope of reproducible neural architecture search. arXiv preprint arXiv:2001.00326, 2020.", + "[11] Roy Frieden and Robert A Gatenby. Exploratory data analysis using Fisher information. Springer Science & Business Media, 2010.", + "[12] Boris Hanin and David Rolnick. Complexity of linear regions in deep networks. In International Conference on Machine Learning, pages 2596-2604. PMLR, 2019.", + "[13] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018." + ], + "bbox": [ + 93, + 257, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Gabriela Kadlecova, Jovita Lukasik, Martin Pilát, Petra Vidnerova, Mahmoud Safari, Roman Neruda, and Frank Hutter. Surprisingly strong performance prediction with neural graph features. arXiv preprint arXiv:2404.16551, 2024.", + "[15] Ryo Karakida, Shotaro Akaho, and Shun-ichi Amari. Pathological spectra of the fisher information metric and its variants in deep neural networks. arXiv preprint arXiv:1910.05992, 2019.", + "[16] Ryo Karakida, Shotaro Akaho, and Shun-ichi Amari. Universal statistics of fisher information in deep neural networks: Mean field approach. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 1032–1041. PMLR, 2019.", + "[17] Maurice G Kendall. A new measure of rank correlation. Biometrika, 30(1-2):81-93, 1938.", + "[18] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009.", + "[19] Frederik Kunstner, Philipp Hennig, and Lukas Balles. Limitations of the empirical fisher approximation for natural gradient descent. Advances in neural information processing systems, 32, 2019.", + "[20] Byung-Kwan Lee, Junho Kim, and Yong Man Ro. Masking adversarial damage: Finding adversarial saliency for robust and sparse network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15126-15136, 2022.", + "[21] Junghyup Lee and Bumsub Ham. AZ-NAS: Assembling zero-cost proxies for network architecture search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5893-5903, 2024.", + "[22] Jaehoon Lee, Lechao Xiao, Samuel Schoenholz, Yasaman Bahri, Roman Novak, Jascha Sohl-Dickstein, and Jeffrey Pennington. Wide neural networks of any depth evolve as linear models under gradient descent. Advances in neural information processing systems, 32, 2019.", + "[23] Namhoon Lee, Thalaiyasingam Ajanthan, and Philip HS Torr. Snip: Single-shot network pruning based on connection sensitivity. arXiv preprint arXiv:1810.02340, 2018.", + "[24] Steven J Leon, Lisette De Pillis, and Lisette G De Pillis. Linear algebra with applications. Pearson Prentice Hall Upper Saddle River, NJ, 2006.", + "[25] Guihong Li, Yuedong Yang, Kartikeya Bhardwaj, and Radu Marculescu. ZiCo: Zero-shot NAS via inverse coefficient of variation on gradients. In The Eleventh International Conference on Learning Representations, 2023.", + "[26] Ming Lin, Pichao Wang, Zhenhong Sun, Hesen Chen, Xiuyu Sun, Qi Qian, Hao Li, and Rong Jin. Zen-nas: A zero-shot nas for high-performance image recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 347-356, 2021.", + "[27] Hanxiao Liu, Karen Simonyan, and Yiming Yang. DARTS: Differentiable architecture search. arXiv preprint arXiv:1806.09055, 2018.", + "[28] Alexander Ly, Maarten Marsman, Josine Verhagen, Raoul PPP Grasman, and Eric-Jan Wagenmakers. A tutorial on fisher information. Journal of Mathematical Psychology, 80:40-55, 2017." + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14889", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] James Martens. New insights and perspectives on the natural gradient method. Journal of Machine Learning Research, 21 (146):1-76, 2020.", + "[30] Abhinav Mehrotra, Alberto Gil CP Ramos, Sourav Bhattacharya, Lukasz Dudziak, Ravichander Vipperla, Thomas Chau, Mohamed S Abdelfattah, Samin Ishtiaq, and Nicholas Donald Lane. Nas-bench-asr: Reproducible neural architecture search for speech recognition. In International Conference on Learning Representations, 2021.", + "[31] Joe Mellor, Jack Turner, Amos Storkey, and Elliot J Crowley. Neural architecture search without training. In International conference on machine learning, pages 7588-7598. PMLR, 2021.", + "[32] Bert Moons, Parham Noorzad, Andrii Skliar, Giovanni Mariani, Dushyant Mehta, Chris Lott, and Tijmen Blankevoort. Distilling optimal neural networks: Rapid search in diverse spaces. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12229-12238, 2021.", + "[33] Xuefei Ning, Changcheng Tang, Wenshuo Li, Zixuan Zhou, Shuang Liang, Huazhong Yang, and Yu Wang. Evaluating efficient performance estimators of neural architectures. Advances in Neural Information Processing Systems, 34: 12265-12277, 2021.", + "[34] Roman Novak, Jascha Sohl-Dickstein, and Samuel S Schoenholz. Fast finite width neural tangent kernel. In International Conference on Machine Learning, pages 17018-17044. PMLR, 2022.", + "[35] Hyeyoung Park and Kwanyong Lee. Adaptive natural gradient method for learning of stochastic neural networks in mini-batch mode. Applied Sciences, 9(21):4568, 2019.", + "[36] Jeffrey Pennington and Pratik Worah. The spectrum of the fisher information matrix of a single-hidden-layer neural network. Advances in neural information processing systems, 31, 2018.", + "[37] C Radhakrishna Rao. Information and the accuracy attainable in the estimation of statistical parameters. In Breakthroughs in Statistics: Foundations and basic theory, pages 235-247. Springer, 1992.", + "[38] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *Proceedings of the IEEE conference on computer vision and pattern recognition*, pages 4510-4520, 2018.", + "[39] Charles Spearman. The proof and measurement of association between two things. 1961.", + "[40] Kunio Tanabe and Masahiko Sagae. An exact cholesky decomposition and the generalized inverse of the variance-covariance matrix of the multinomial distribution, with applications. Journal of the Royal Statistical Society: Series B (Methodological), 54(1):211-219, 1992.", + "[41] Hidenori Tanaka, Daniel Kunin, Daniel L Yamins, and Surya Ganguli. Pruning neural networks without any data by iteratively conserving synaptic flow. Advances in neural information processing systems, 33:6377-6389, 2020.", + "[42] Chaoqi Wang, Guodong Zhang, and Roger Grosse. Picking winning tickets before training by preserving gradient flow. arXiv preprint arXiv:2002.07376, 2020." + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Colin White, Mikhail Khodak, Renbo Tu, Shital Shah, Sébastien Bubeck, and Debadeepta Dey. A deeper look at zero-cost proxies for lightweight NAS. *ICLR Blog Track*, 2022.", + "[44] Huan Xiong, Lei Huang, Mengyang Yu, Li Liu, Fan Zhu, and Ling Shao. On the number of linear regions of convolutional neural networks. In International Conference on Machine Learning, pages 10514-10523. PMLR, 2020.", + "[45] Zhaohui Yang, Yunhe Wang, Xinghao Chen, Boxin Shi, Chao Xu, Chunjing Xu, Qi Tian, and Chang Xu. CARS: Continuous evolution for efficient neural architecture search. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1829-1838, 2020.", + "[46] Chris Ying, Aaron Klein, Eric Christiansen, Esteban Real, Kevin Murphy, and Frank Hutter. Nas-bench-101: Towards reproducible neural architecture search. In International conference on machine learning, pages 7105-7114. PMLR, 2019.", + "[47] Wang Yining, Wang Liwei, Li Yuanzhi, He Di, Chen Wei, and Liu Tie-Yan. A theoretical analysis of ndcg ranking measures. In Proceedings of the 26th annual conference on learning theory, 2013.", + "[48] Xuanyang Zhang, Pengfei Hou, Xiangyu Zhang, and Jian Sun. Neural architecture search with random labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10907-10916, 2021.", + "[49] Zhihao Zhang and Zhihao Jia. Gradsign: Model performance inference with theoretical insights. arXiv preprint arXiv:2110.08616, 2021.", + "[50] Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V Le. Learning transferable architectures for scalable image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8697-8710, 2018." + ], + "bbox": [ + 516, + 92, + 903, + 570 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14890", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Training-free Neural Architecture Search through Variance of Knowledge of Deep Network Weights/740054ef-86e4-4bac-b7b3-53d57b1f3a66_model.json b/2025/Training-free Neural Architecture Search through Variance of Knowledge of Deep Network Weights/740054ef-86e4-4bac-b7b3-53d57b1f3a66_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d8e8a420c06a1d72ed5a2c36363eb3bb3d8ae5d1 --- /dev/null +++ b/2025/Training-free Neural Architecture Search through Variance of Knowledge of Deep Network Weights/740054ef-86e4-4bac-b7b3-53d57b1f3a66_model.json @@ -0,0 +1,2453 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.13, + 0.885, + 0.177 + ], + "angle": 0, + "content": "Training-free Neural Architecture Search through Variance of Knowledge of Deep Network Weights" + }, + { + "type": "text", + "bbox": [ + 0.363, + 0.204, + 0.673, + 0.221 + ], + "angle": 0, + "content": "Ondrej Tybl Lukas Neumann" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.222, + 0.932, + 0.24 + ], + "angle": 0, + "content": "CMP Visual Recognition Group, Faculty of Electrical Engineering, Czech Technical University in Prague" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.243, + 0.714, + 0.256 + ], + "angle": 0, + "content": "ondrej.tybl@cvut.cz lukas.neumann@cvut.cz" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.323, + 0.484, + 0.475 + ], + "angle": 0, + "content": "Deep learning has revolutionized computer vision, but it achieved its tremendous success using deep network architectures which are mostly hand-crafted and therefore likely suboptimal. Neural Architecture Search (NAS) aims to bridge this gap by following a well-defined optimization paradigm which systematically looks for the best architecture, given objective criterion such as maximal classification accuracy. The main limitation of NAS is however its astronomical computational cost, as it typically requires training each candidate network architecture from scratch." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.484, + 0.595 + ], + "angle": 0, + "content": "In this paper, we aim to alleviate this limitation by proposing a novel training-free proxy for image classification accuracy based on Fisher Information. The proposed proxy has a strong theoretical background in statistics and it allows estimating expected image classification accuracy of a given deep network without training the network, thus significantly reducing computational cost of standard NAS algorithms." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.484, + 0.702 + ], + "angle": 0, + "content": "Our training-free proxy achieves state-of-the-art results on three public datasets and in two search spaces, both when evaluated using previously proposed metrics, as well as using a new metric that we propose which we demonstrate is more informative for practical NAS applications. The source code is publicly available at https://www.github.com/ondratybl/VKDNW." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.725, + 0.222, + 0.74 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.87 + ], + "angle": 0, + "content": "In most instances, neural network architectures are designed by authors following the field's \"best-practices\" or their experience, without any formal and repeatable procedure. This is however inconvenient especially in applications on a large scale. Neural Architecture Search (NAS) aims to bridge this gap by following a well-defined optimization paradigm which systematically looks for the best architecture, given objective criterion such as maximal accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The main limitation of Neural Architecture Search (NAS) is however the computational cost, because in the" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.293, + 0.891, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.539, + 0.907, + 0.624 + ], + "angle": 0, + "content": "Figure 1. Training-free NAS methods on ImageNet16-120 [10]. Methods are compared by Normalized Discounted Cumulative Gain (see Sec. 3), our method (VKDNW) is the best also measured by Kendall's \\(\\tau\\) and Spearman's \\(\\rho\\) correlations (see Table 1). Also note that simple number of trainable layers (below denoted \\(\\aleph\\)) is significantly better trivial proxy than the number of FLOPs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.749 + ], + "angle": 0, + "content": "most basic NAS setup, it is required to train thousands or more of different deep network architectures from scratch in order just to calculate a single scalar – the objective function value, such as the classification accuracy. This severely limits practical applications of NAS as the size of feasible architecture search space is only a small fraction of the overall space of all networks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Training-free Neural Architecture Search (TF-NAS) aims to alleviate this limitation by introducing an objective function proxy which – unlike the actual objective function – does not require training the network. As a result, a good proxy allows the TF-NAS algorithm to explore significantly bigger portions of the network architecture search space compared to traditional NAS, and to find the best network architecture without training a single network. The crucial question is however finding an appropriate objective function proxy." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14881" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.485, + 0.272 + ], + "angle": 0, + "content": "In this paper, we present a novel, principled objective function proxy called Variance of Knowledge of Deep Network Weights (VKDNW) for image classification accuracy, which allows us to find optimal network architectures for image classification without training them first. Our method is, to the best of our knowledge, the first successful application of Fisher Information theory [28] in the context of large deep neural networks, and as such allows us to formally describe and quantify the difficulty of network parameters' estimation process. In other words, given a network architecture, our method estimates how easy or hard it will be to train the network." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.275, + 0.484, + 0.44 + ], + "angle": 0, + "content": "Additionally, we also observe that the evaluation metrics used in the TF-NAS community are not well-suited to the problem at hand, because it unnecessarily penalizes for bad proxy accuracy for networks which are not interesting, and vice-versa it does not sufficiently reward proxies which are able to accurately pick out good network architectures. Following this observation, we propose that the Normalized Discounted Cumulative Gain should be used in comparison with other TF-NAS metrics, and show indeed that there are significant differences amongst previously proposed TF-NAS methods when the new metric is considered." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.443, + 0.46, + 0.458 + ], + "angle": 0, + "content": "To summarize, we make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.46, + 0.483, + 0.536 + ], + "angle": 0, + "content": "1. We introduce a novel algorithm for estimation of Fisher Information Matrix spectrum, which is tractable even for models with large number of parameters such as deep networks, that overcomes the usual problems of numerical stability." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.536, + 0.484, + 0.642 + ], + "angle": 0, + "content": "2. We introduce a novel principled VKDNW proxy for image classification accuracy. The proxy is based on strong theoretical background and captures uncertainty in weight estimation process. It brings information that is orthogonal to the model size which then allows for efficient combination with previously proposed proxies, leading to state-of-the-art results." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.642, + 0.484, + 0.702 + ], + "angle": 0, + "content": "3. We propose a new evaluation metric for TF-NAS proxies which is more relevant to the actual NAS objective as it concentrates on ability of given proxy to identify good networks." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.46, + 0.484, + 0.702 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.722, + 0.182, + 0.737 + ], + "angle": 0, + "content": "2. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our zero-shot proxy for image classification accuracy builds on Fisher Information theory [28], therefore we begin by a thorough analysis of Fisher Information Matrix (FIM) of the network weights estimation problem (see Sec. 2.1). We proceed by discussing challenges involved in practical application of FIM in context of large over-parametrized models that lead to only limited success in previous works and present our contributions to overcome these limitations (see Sec. 2.2). Finally, we propose a novel FIM-based proxy for NAS algorithms (Sec. 2.3)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.697, + 0.105 + ], + "angle": 0, + "content": "2.1. Fisher Information" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.115, + 0.907, + 0.174 + ], + "angle": 0, + "content": "The problem of finding the optimal weights of a neural network \\( f \\) for the task of \\( C \\)-class image classification can be seen as a maximum likelihood estimation with a statistical model" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.178, + 0.907, + 0.215 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {\\theta} (c \\mid x) = \\frac {\\exp (\\Psi_ {c} (x , \\theta))}{\\sum_ {d = 1} ^ {C} \\exp (\\Psi_ {d} (x , \\theta))}, \\quad c = 1, \\dots , C \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.22, + 0.907, + 0.386 + ], + "angle": 0, + "content": "where \\(\\sigma_{\\theta}(\\cdot |x)\\) denotes the a posteriori distribution of the labels given input image \\(x\\) and \\(\\Psi (x,\\theta)\\in \\mathbb{R}^{C} = (\\Psi_{1}(x,\\theta),\\dots ,\\Psi_{C}(x,\\theta))\\) is the network output (logits) given the weight vector \\(\\theta \\in \\mathbb{R}^p\\), i.e. the network weights. We describe the process of training as finding the optimal weight vector \\(\\theta^{*}\\) that fits our data and we posit that network architectures should be characterised by how easy it is to estimate their optimal network weights \\(\\theta^{*}\\). We build upon statistical learning theory and use Fisher Information [28] framework to formally describe expected behaviour of the training process of a given deep network \\(f\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.906, + 0.477 + ], + "angle": 0, + "content": "The Fisher Information Matrix (FIM) encompasses information on the difficulty of the parameter estimation problem and it plays a crucial role in several fundamental results, which we apply below in the context of deep networks. The FIM of a network \\( f \\) and its set of weights \\( \\theta \\in \\mathbb{R}^p \\) is given as" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.483, + 0.905, + 0.502 + ], + "angle": 0, + "content": "\\[\nF (\\theta) := \\mathbb {E} \\left[ \\nabla_ {\\theta} \\sigma_ {\\theta} (c | x) \\nabla_ {\\theta} \\sigma_ {\\theta} (c | x) ^ {T} \\right] \\in \\mathbb {R} ^ {p \\times p}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.906, + 0.569 + ], + "angle": 0, + "content": "where we take the expected value \\(\\mathbb{E}\\) with respect to the joint distribution of \\((x, c)\\). For more detailed account on Fisher Information theory and its applications in the context of machine learning, we kindly refer reader to [16, 20, 29, 35, 36]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.571, + 0.907, + 0.773 + ], + "angle": 0, + "content": "Cramér-Rao bound. The first part of estimation theory we build upon is the inverse of the FIM, known as the Cramér-Rao bound (see [11]), which is the asymptotic variance of the estimated weights (i.e. the uncertainty coming from the data variability). Thus, the larger the matrix norm of the FIM, the more certain we are about the weights \\(\\theta\\). More specifically, any data-dependent estimator \\(\\hat{\\theta}_n = \\hat{\\theta}_n(x_1,\\ldots ,x_n)\\) is a random vector with randomness coming from the (independent) choice of input images \\(x_{1},\\ldots ,x_{n}\\) and as such has some variance matrix \\(\\operatorname{Var}\\left(\\hat{\\theta}_n\\right)\\). The famous result (see [8, 37]) named in honour of H. Cramér and C. R. Rao states that if \\(\\hat{\\theta}_n\\) is unbiased then the variance is bounded from below as" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.777, + 0.905, + 0.806 + ], + "angle": 0, + "content": "\\[\n\\operatorname {V a r} \\left(\\hat {\\theta} _ {n}\\right) \\geq \\frac {1}{n} F ^ {- 1} (\\theta) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.855 + ], + "angle": 0, + "content": "and for maximum likelihood estimation this bound is attained as the number of input images grows to infinity \\( n \\to \\infty \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We formally show (see Supplementary material) that for each weight \\(\\theta(j)\\) the mean square error of our estimation is controlled by the diagonal element of the FIM inverse as" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14882" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.16, + 0.11, + 0.483, + 0.139 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} \\left(\\hat {\\theta} _ {n} (j) - \\theta (j)\\right) ^ {2} \\geq \\frac {1}{n} \\left(F ^ {- 1} (\\theta)\\right) _ {j j}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.144, + 0.484, + 0.234 + ], + "angle": 0, + "content": "Therefore, knowing the FIM allows us to evaluate how certain we are about the weight estimates. We can go even further by inspecting the eigenvalues of \\(F(\\theta)\\). Denoting the largest and smallest eigenvalue of the FIM as \\(\\lambda_{\\mathrm{min}}\\) and \\(\\lambda_{\\mathrm{max}}\\) respectively, then we have that there exist linear combination coefficients \\(e_{\\mathrm{min}}\\) and \\(e_{\\mathrm{max}}\\) of unit size so that" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.239, + 0.397, + 0.27 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} \\left(e _ {\\min} ^ {T} \\hat {\\theta} _ {n} - e _ {\\min} ^ {T} \\theta\\right) ^ {2} \\geq \\frac {1}{n \\lambda_ {\\min}},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.272, + 0.482, + 0.304 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} \\left(e _ {\\max } ^ {T} \\hat {\\theta} _ {n} - e _ {\\max } ^ {T} \\theta\\right) ^ {2} \\geq \\frac {1}{n \\lambda_ {\\max }}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.307, + 0.484, + 0.398 + ], + "angle": 0, + "content": "indicating that if the difference between the largest and smallest eigenvalue is large then there exist combinations of weights with very large difference in the estimation certainty. Altogether, the more the eigenvalues of the FIM are similar, the more similar is also the variance in the weight estimation across all model weights." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.484, + 0.448 + ], + "angle": 0, + "content": "Prediction sensibility. The change of the model prediction measured by the KL-divergence when subject to a small perturbations of the weights is given as" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.452, + 0.483, + 0.482 + ], + "angle": 0, + "content": "\\[\nD _ {K L} \\left(\\sigma_ {\\theta + \\theta_ {\\delta}} (\\cdot | x), \\sigma_ {\\theta} (\\cdot | x)\\right) \\approx \\frac {1}{2} \\theta_ {\\delta} ^ {T} F (\\theta) \\theta_ {\\delta} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.487, + 0.484, + 0.576 + ], + "angle": 0, + "content": "for a small weight perturbation vector \\(\\theta_{\\delta} \\in \\mathbb{R}^{p}\\). Thus, \\(F(\\theta)\\) measures volatility of the predictions subject to the weight change – if the difference between the largest and the smallest eigenvalues of the FIM is large, then there exist perturbation directions \\(\\theta_{min}, \\theta_{max}\\) corresponding to \\(\\lambda_{min}, \\lambda_{max}\\) such that" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.581, + 0.495, + 0.669 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} D _ {K L} \\left(\\sigma_ {\\theta + \\theta_ {m i n}} (\\cdot | x), \\sigma_ {\\theta} (\\cdot | x)\\right) \\approx \\frac {1}{2} \\theta_ {m i n} ^ {T} F (\\theta) \\theta_ {m i n} = \\\\ = \\frac {1}{2} \\lambda_ {\\min } \\| \\theta_ {\\min } \\| ^ {2} \\ll \\frac {1}{2} \\lambda_ {\\max } \\| \\theta_ {\\max } \\| ^ {2} = \\tag {7} \\\\ = \\frac {1}{2} \\theta_ {m a x} ^ {T} F (\\theta) \\theta_ {m a x} \\approx D _ {K L} \\left(\\sigma_ {\\theta + \\theta_ {m a x}} (\\cdot | x), \\sigma_ {\\theta} (\\cdot | x)\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.484, + 0.735 + ], + "angle": 0, + "content": "and therefore in some directions a small change of the weights has much larger impact on the prediction than in others, making the model less balanced. For further discussion, see the Supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.747, + 0.484, + 0.777 + ], + "angle": 0, + "content": "2.2. Empirical Fisher Information Matrix implementation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.787, + 0.484, + 0.818 + ], + "angle": 0, + "content": "When independent and identically distributed sample images \\(x_{n}\\) are available, empirical FIM \\(\\hat{F} (\\theta)\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.825, + 0.483, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\hat {F} (\\theta) := \\frac {1}{n} \\sum_ {n = 1} ^ {N} \\mathbb {E} _ {\\sigma_ {\\theta}} \\left[ \\nabla_ {\\theta} \\sigma_ {\\theta} (c | x _ {n}) \\nabla_ {\\theta} \\sigma_ {\\theta} (c | x _ {n}) ^ {T} \\right] \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbb{E}_{\\sigma_{\\theta}}\\) now denotes the expectation with respect to the model prediction \\(\\sigma_{\\theta}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "We would like to first emphasize several crucial aspects and contributions of this paper that lead to the first success of the Fisher Information theory (having e.g. [1] in mind) in the context of Neural Architecture Search, despite Fisher Information being one of the first obvious choices for deep network analysis:" + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.186, + 0.878, + 0.202 + ], + "angle": 0, + "content": "1. Following [19] we write the empirical FIM (8) as" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.208, + 0.905, + 0.27 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {F} (\\theta) = \\frac {1}{n} \\sum_ {n = 1} ^ {N} \\left[ \\nabla_ {\\theta} \\Psi (x _ {n}, \\theta) ^ {T} \\left(\\operatorname {d i a g} \\left(\\sigma_ {\\theta} (\\cdot , x _ {n})\\right) - \\right. \\right. \\\\ \\left. \\sigma_ {\\theta} (\\cdot , x _ {n}) \\sigma_ {\\theta} (\\cdot , x _ {n}) ^ {T}\\right) \\nabla_ {\\theta} \\Psi (x _ {n}, \\theta) ]. \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.276, + 0.906, + 0.337 + ], + "angle": 0, + "content": "and we further decompose the inner matrix \\(\\mathrm{diag}(\\sigma_{\\theta}(\\cdot ,x_n)) - \\sigma_{\\theta}(\\cdot ,x_n)\\sigma_{\\theta}(\\cdot ,x_n)^T\\) using analytical formulas from [40] to avoid numerical instability of the computation as we arrive at a feasible representation" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.343, + 0.905, + 0.383 + ], + "angle": 0, + "content": "\\[\n\\hat {F} (\\theta) = \\frac {1}{n} \\sum_ {n = 1} ^ {N} A _ {n} ^ {T} A _ {n} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.39, + 0.906, + 0.481 + ], + "angle": 0, + "content": "for some matrices \\(A_{n}\\in \\mathbb{R}^{C\\times p}\\). We observed that networks typically yield very imbalanced outputs at initialization (i.e. every network prioritizes few classes over the rest) making it extremely important not to exclude the factor \\(\\mathrm{diag}(\\sigma_{\\theta}(\\cdot ,x_n)) - \\sigma_{\\theta}(\\cdot ,x_n)\\sigma_{\\theta}(\\cdot ,x_n)^T\\), which is however difficult to compute without underflow/overflow." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.906, + 0.637 + ], + "angle": 0, + "content": "2. The dimension of FIM is equal to \\( p \\) (the number of all trainable parameters) and therefore direct computation of the eigenvalues is numerically intractable and unstable. However, our results show that if a small number of representative parameters is drawn, then computation becomes stable while the discrimination power does not suffer. We used a simple rule where a single parameter from each trainable layer (not including batch normalization) is chosen. Stability with respect to choice of such parameters can be found in Supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.641, + 0.906, + 0.761 + ], + "angle": 0, + "content": "3. FIM of large networks typically suffers from having pathological spectrum, i.e. there usually exist zero eigenvalue and its multiplicity is large (see [15]) and thus the eigenvalues estimation due to a large condition number is imprecise. However, as we deal with a symmetric positive-semidefinite matrix, the eigenvalues actually coincide with the singular values (see [24]), for which the estimation algorithm performs better." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.485, + 0.906, + 0.761 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Let us also emphasize that there is a common misconception in part of the community as it is often mistakenly assumed that one can simply use the true labels of \\( x_{n} \\) in (8) in place of \\( c \\). However, such definition is then meaningless as it does not approximate the FIM in the classical Monte Carlo sense (see Supplementary material and [19] for a comparison). That means that the empirical FIM does not depend on the true labels and therefore our method does not require real data - we use random input instead." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14883" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.483, + 0.122 + ], + "angle": 0, + "content": "2.3. Variance of Knowledge for Deep Network Weights" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.131, + 0.486, + 0.207 + ], + "angle": 0, + "content": "In order to characterize properties of the parameter estimation process for a given network \\( f \\) through the lens of Fisher Information theory, we inspect the eigenvalues of the empirical FIM \\( \\hat{F}(\\theta_{\\mathrm{init}}) \\) and define the entropy of Variance of Knowledge for Deep Network Weights (VKDNW) as" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.213, + 0.483, + 0.293 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {V K D N W} (f) := - \\sum_ {k = 1} ^ {9} \\tilde {\\lambda} _ {k} \\log \\tilde {\\lambda} _ {k} \\\\ \\tilde {\\lambda} _ {k} = \\frac {\\lambda_ {k}}{\\sum_ {j = 1} ^ {9} \\lambda_ {j}}, \\quad k = 1, \\dots , 9, \\tag {11} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.298, + 0.483, + 0.342 + ], + "angle": 0, + "content": "where \\(\\lambda_{k}\\) denotes the \\(k\\)-th decile of the FIM eigenvalues as the representation of the FIM spectrum1, and \\(\\theta_{\\mathrm{init}}\\) denotes network weights at initialization." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.344, + 0.484, + 0.571 + ], + "angle": 0, + "content": "Our score therefore measures the diversity of the FIM eigenvalues, and from the entropy theory we know that VKDNW attains its maximum exactly when all the eigenvalues \\(\\lambda_{k}\\) are equal and VKDNW gets lower as the eigenvalues become more different. Based on the discussion in Sec. 2.1 we see that VKDNW is high when the uncertainty in all model weight combinations are similar (see Cramér-Rao bound) and there are no directions in the weight space that would influence the network prediction substantially differently than others. Due to the fact that we have normalized both the number of eigenvalues under consideration (by taking a fixed number of representatives irrespective of the network size) and the magnitude of the eigenvalues, VKDNW is independent of network size (number of network weights \\(p\\))." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.572, + 0.484, + 0.693 + ], + "angle": 0, + "content": "Let us note, that we are familiar with the fact that even though our motivation was (among others) based on Cramér-Rao bound that assumes evaluation of FIM at the correct weight vector \\(\\theta_{X}\\) that fits the data, which is typically far away from the weights given at initialization. However, our empirical results below support the hypothesis that the evaluation despite being in the wrong point brings valuable information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.697, + 0.484, + 0.849 + ], + "angle": 0, + "content": "Ranking networks for NAS. The proposed VKDNW score is independent of network size, which is extremely beneficial to compare individual structures of network architectures. Thus, it does not aim to capture capacity, rather it targets feasibility of the computation graph given number of operations. However, when comparing different network structures and different network sizes together as it is done in NAS, one indeed needs to take network size into account as well, because naturally larger networks have bigger capacity and therefore tend to have higher accuracy. To" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.094, + 0.9, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.251, + 0.907, + 0.307 + ], + "angle": 0, + "content": "Figure 2. Toy example of two rankings on 10 networks. We plot accuracies ordered by the rankings and evaluation metrics Kendall's \\(\\tau\\) (KT) and Spearman's \\(\\rho\\) (SPR) correlations and Normalized Discounted Cumulative Gain \\((\\mathrm{nDCG}_5)\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.907, + 0.363 + ], + "angle": 0, + "content": "capture also the capacity we proxy the network size by the number of layers with weights that we denote as \\(\\aleph(f)\\) for a network \\(f\\) and we introduce the ranking" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.369, + 0.905, + 0.387 + ], + "angle": 0, + "content": "\\[\n\\operatorname {V K D N W} _ {\\text {s i n g l e}} (f) := \\aleph (f) + \\operatorname {V K D N W} (f) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.393, + 0.907, + 0.483 + ], + "angle": 0, + "content": "Here we leverage the fact that VKDNW as an entropy of some quantity is always between 0 and 1 and by summing it with an integer-valued quantity we in fact obtain that we have first grouped networks by our size proxy \\(\\aleph\\) and then within each group of similar networks sizes we order them by VKDNW." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.5, + 0.699, + 0.515 + ], + "angle": 0, + "content": "3. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.526, + 0.906, + 0.632 + ], + "angle": 0, + "content": "In this section, we dive into the problem of evaluating (image) classification accuracy proxies for training-free NAS (TF-NAS). In NAS algorithms, the proxy actually does not have to predict the classification accuracy in absolute terms – since we are after finding the \"best\" network for given task – it's sufficient that the proxy properly ranks the networks, ordering them from worst to the best." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.633, + 0.906, + 0.738 + ], + "angle": 0, + "content": "Suppose that a collection of \\(K\\) networks with validation accuracies \\(\\mathrm{acc}_1, \\ldots, \\mathrm{acc}_K\\) have been ranked by a TF-NAS proxy as \\(r_1, \\ldots, r_K \\in \\mathbb{R}\\). The standard way of evaluating [14, 21, 25] how well the ranks correspond to the accuracies is to compute Kendall's \\(\\tau\\) (KT, [17]) and Spearman's \\(\\rho\\) (SPR, [39]) rank correlation coefficients. Kendall's \\(\\tau\\) is given as" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.745, + 0.907, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\tau := \\frac {n _ {c} - n _ {d}}{\\sqrt {n _ {c} + n _ {d} + n _ {1}} \\sqrt {n _ {c} + n _ {d} + n _ {2}}}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.902 + ], + "angle": 0, + "content": "where \\( n_c \\) denotes the number of concordant pairs \\( (\\mathrm{acc}_k, r_k) \\), \\( n_d \\) the number of discordant pairs and \\( n_1 \\) (resp. \\( n_2 \\)) denote the number of ties in \\( \\mathrm{acc}_k \\) only (resp. in \\( r_k \\) only). In our context, (KT) relates after some normalization to a probability that two randomly chosen network rankings \\( r_k, r_l \\) are ordered correctly according to their accuracies \\( \\mathrm{acc}_k, \\mathrm{acc}_l \\). In contrast, (SPR) is defined to account more for outliers, i.e. it tends to put a large penalization if" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1As explained in Sec. 2.2, the smallest eigenvalue \\( \\lambda_0 \\) is usually equal to zero and thus we exclude it for stability reasons (similarly with the maximal eigenvalue \\( \\lambda_{10} \\))" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14884" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "there are some networks for which the difference between the rank of the accuracy \\(\\mathrm{acc}_k\\) and the rank \\(r_k\\) is large as it is just the classical correlation coefficient however applied on the orders of the assessed quantities." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.155, + 0.486, + 0.44 + ], + "angle": 0, + "content": "Normalized Discounted Cumulative Gain. Both (KT) and (SPR) are good evaluation metrics when our interest lies in comparison of ranking proxies when all networks involved are of the same importance. However, in NAS we are primarily interested if a ranking helps us to pick best networks from a given collection and less interested how well the ranking compares two networks with low accuracy. Finding inspiration in information retrieval where one measures the quality of a system that retrieves resources relevant to an input query, we propose to use Normalized Discounted Cumulative Gain (nDCG) [2] as a more relevant metric to measure quality of TF-NAS proxies. The nDCG metric is defined with a key requirement that highly relevant documents are more valuable when they appear earlier in the search engine results (i.e., in higher-ranking positions). This requirement can be reformulated for the task of neural architecture search as networks with high accuracies are more valuable when they appear on higher-ranking positions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.441, + 0.483, + 0.488 + ], + "angle": 0, + "content": "We thus define the metric as follows: first we order the networks by their ranking \\( r_1, \\ldots, r_K \\) so that \\( r_{k_1} = K, r_{k_2} = K - 1, \\ldots \\) and we compute" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.494, + 0.483, + 0.542 + ], + "angle": 0, + "content": "\\[\n\\mathrm {n D C G} _ {P} := \\frac {1}{Z} \\left(\\sum_ {j = 1} ^ {P} \\frac {2 ^ {\\mathrm {a c c} _ {k _ {j}}} - 1}{\\log_ {2} (1 + j)}\\right), \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.546, + 0.483, + 0.621 + ], + "angle": 0, + "content": "where \\( P \\in \\mathbb{N} \\) is a parameter determining how many top-ranked networks do we consider (e.g. it corresponds to the population size in an evolution algorithm) and \\( Z \\) is a normalization factor that represents the ideal discounted cumulative gain so that nDCG is equal to one for a perfect fit." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.622, + 0.483, + 0.697 + ], + "angle": 0, + "content": "The higher nDCG the better the ranking is as it is a weighted average of transformed top-ranked accuracies. In [47] it is shown that the choice of the discount factor of given by inverse logarithm is a good choice as it nDCG then well separates different ranking systems." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.7, + 0.483, + 0.866 + ], + "angle": 0, + "content": "Toy Example. Let us briefly demonstrate the weak ability of (KT) and (SPR) to distinguish rankings that poorly discriminate networks with high accuracy. Suppose a collection of 10 networks with their validation accuracies \\(100,90,\\dots ,10\\) is given and our aim is to rank them from worst to the best. We evaluate two different toy rankings: a) ranking damaged_top which perfectly fits the accuracies, only the two top networks are swapped with the third and fourth, b) ranking perfect_top which fits perfectly for the top-performing networks, however, now the order of the four worst networks is reversed (see Figure 2). The" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.333 + ], + "angle": 0, + "content": "damaged_top ranking is an example that should be evaluated worse for the architecture search task compared to perfect_top as in the first case the best networks are not placed first, while in the second they are. On the other hand, even though perfect_top is not a perfect fit, it still discriminates the top networks ideally and only struggles for networks with low accuracy, which are not interesting for NAS which is after the best networks. From the correlation perspectives of (KT), (SPR) damaged_top is better than perfect_top, therefore, if we rely only on these two evaluation metrics we prefer ranking that is worse for the architecture search task as the top networks are ranked worse. On the other hand, for perfect_top we obtain ideal nDCG\\(_5\\) while it drops to 0.5 for damaged_top. We conclude that using nDCG\\(_5\\) we choose ranking that has higher discriminatory power for good networks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.905, + 0.456 + ], + "angle": 0, + "content": "We also note that from statistical perspective the (KT) and (SPR) of damaged_top is significantly higher than of a random ranking3, and therefore we'd conclude that damaged_top is not independent of ground truth accuracy. On the other hand, when we perform the same statistical test on damaged_top using nDCG5, we conclude that the damaged_top is not significantly better than ranking networks randomly4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.472, + 0.655, + 0.487 + ], + "angle": 0, + "content": "4. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.905, + 0.568 + ], + "angle": 0, + "content": "Zero-shot NAS aims to rank given networks in a training-free manner based on their (a priori unknown) final performance, which allows to prune the huge search space with limited costs when seeking for optimal architecture [10, 27, 46] and other configuration [3, 26, 38]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.906, + 0.765 + ], + "angle": 0, + "content": "Throughout previous works different approaches for ranking computation can be found. In many of the works the gradient with respect to the network weights is investigated (GradNorm, GraSP, SNIP, Synflow, [1, 23, 41, 42]), i.e. leveraging the first order approximation of the network. In Jacov [1] correlations of the Jacobian matrices among various input samples are compared; in NASWOT [31] the linear maps induced by data points are examined; Zen [26] uses the approximation of gradient with respect to featuremaps; GradSign [49] compares the optimization landscape at the level individual training samples; or in ZiCo [25] the gradients from multiple forward and backward passes preferring large magnitude and low variance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.905, + 0.826 + ], + "angle": 0, + "content": "Other methods aggregate multiple sources aiming at obtaining a better informed ranks: TE-NAS [6] uses both the number of linear regions [12, 44] and the condition number of Neural Tangent Kernel [13, 22]. However, it is well" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.839, + 0.905, + 0.864 + ], + "angle": 0, + "content": "3That is, the \\(p\\)-value for the hypothesis that damaged_top is assigned independently of the accuracies is below 0.005." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.864, + 0.905, + 0.901 + ], + "angle": 0, + "content": "4Running 1000 samples on a random ranking we obtain that damaged_top has nDCG₅ around 75th percentile of such random evaluations and it is not significantly better than a random ranking." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.839, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "2In case of ties we take random ordering within the groups (parameter ignore_ties=True in the scikit-learn implementation)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14885" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.088, + 0.916, + 0.442 + ], + "angle": 0, + "content": "
TypeCIFAR-10CIFAR-100ImageNet16-120
KTSPRnDCGKTSPRnDCGKTSPRnDCG
Simple rankings
FLOPsS0.6230.7990.7450.5860.7630.5760.5450.7180.403
GradNorm [1]S0.3280.4380.5090.3410.4510.2780.3100.4180.265
GraSP [1, 42]S0.3520.5050.5180.3490.4980.2840.3590.5020.281
SNIP [1, 23]S0.4310.5910.5130.4400.5970.2860.3890.5210.286
SynFlow [1, 41]S0.5610.7580.7090.5530.7500.5940.5310.7190.511
Jacov [1]S0.6160.8000.5400.6390.8200.4020.6020.7790.356
NASWOT [31]S0.5710.7620.6070.6070.7990.4750.6050.7940.490
ZenNAS [26]S0.1020.1030.1200.0790.0720.1150.0910.1090.073
GradSign† [49]S·0.765··0.793··0.783·
ZiCo [25]S0.6070.8020.7510.6140.8090.6070.5870.7790.523
TE-NAS [6]A0.5360.7220.6020.5370.7230.3270.5230.7090.330
AZ-NAS [21]A0.7120.8920.7490.6960.8800.5490.6730.8590.534
No. of trainable layers (N)S0.6260.7670.6710.6460.7870.5250.6230.7640.497
VKDNWsingle (ours)S0.6180.8150.7510.6340.8290.6170.6220.8140.608
VKDNWagg (ours)A0.7500.9190.7850.7530.9190.6360.7430.9060.664
Model-driven rankings
GRAF [14]A0.8200.9530.9350.8090.9480.8580.7960.9410.828
VKDNWm (ours)A0.6470.8310.7500.6360.8210.6020.6110.7980.575
(VKDNW+ZCS)m (ours)A0.8400.9630.9220.8340.9600.8840.8300.9580.843
(VKDNW+ZCS+GRAF)m (ours)A0.8590.9710.9460.8470.9660.8950.8420.9630.867
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.452, + 0.908, + 0.509 + ], + "angle": 0, + "content": "Table 1. Training-free NAS methods in the NAS-Bench-201 [10] search space, evaluated on three public datasets. Kendall's \\(\\tau\\) (KT), Spearman's \\(\\rho\\) (SPR) and Normalized Discounted Cumulative Gain (nDCG) are reported, results are averages of 5 independent runs (all stds are \\(< 0.02\\)). The Type column differentiates single (S) and aggregated (A) rankings. Results are reproduced with code published by their authors, except those marked†, where results from the original paper are taken." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.483, + 0.621 + ], + "angle": 0, + "content": "known that the kernel computation is highly computationally demanding [34]. AZ-NAS [21] assesses expressivity, trainability and progressivity via examination of feature distribution across all orientations and the Jacobian. However, despite AZ-NAS outperforming previous works in some metrics, it is worse than [25] in key NAS-related aspects such as the cumulative gain (see Sec. 5.1)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.624, + 0.483, + 0.715 + ], + "angle": 0, + "content": "Despite a tremendous effort of the community, it was shown that most of the zero-shot NAS methods perform worse than a simple proxy given just by FLOPs or #params [33, 43]. Thus, there is still a gap for improvements also driven by the need of the ranking explainability that would have a satisfactory theoretical support." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.483, + 0.793 + ], + "angle": 0, + "content": "Furthermore, [14] uses a model-driven ranking that however needs a set of networks for which the validation accuracies are known to train the model and the ability of this ranking to generalize as it is fitted on a specific network collection only is disputable." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Finally, let us discuss the current practice in the NAS methods comparison. Methods are compared either by means of correlations of their scores to accuracy or by reporting the accuracy of the top-ranked network [21, 25]. However, while the first is not tailor-made for the architecture search task and therefore does not assess the desired ranking properties (see Sec. 5.1), comparing performance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.907, + 0.545 + ], + "angle": 0, + "content": "just by the accuracy of a single network is too vulnerable e.g. to the random seed choice." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.561, + 0.649, + 0.578 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.587, + 0.714, + 0.603 + ], + "angle": 0, + "content": "5.1. Ranking aggregation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.907, + 0.671 + ], + "angle": 0, + "content": "In addition to using the single ranking of Eq. (12), we also experiment with multiple rankings (as in [14, 21, 25]) to order network architectures by their accuracy. We use two different options: non-linear and model-driven aggregation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.906, + 0.78 + ], + "angle": 0, + "content": "Non-linear aggregation. The aggregation [21] uses multiplication to combine multiple ranks into a single one, which means a network is highly-ranked if and only if it is highly-ranked in all subsidiary rankings, keeping their influence balanced. That is, denoting rankings \\(\\mathrm{rank}_1,\\ldots ,\\mathrm{rank}_m\\) for some \\(m\\in \\mathbb{N}\\) we define the aggregated ranking" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.786, + 0.905, + 0.805 + ], + "angle": 0, + "content": "\\[\n\\operatorname {r a n k} _ {\\text {a g g}} (f) := \\log \\Pi_ {j = 1} ^ {m} \\operatorname {r a n k} _ {j} (f) \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.854 + ], + "angle": 0, + "content": "for each network \\( f \\). This aggregation is therefore possible only in the context of a given network collection, such as in evolutionary search." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.856, + 0.892, + 0.871 + ], + "angle": 0, + "content": "In our case, \\(\\mathrm{VKDNW_{agg}}\\) aggregates these five proxies:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.871, + 0.775, + 0.886 + ], + "angle": 0, + "content": "- VKDNWsingle (V) ranking (Eq. (12))," + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "- Jacov (J) [1] measures the activations correlation when" + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14886" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.089, + 0.488, + 0.272 + ], + "angle": 0, + "content": "
MethodFLOPsTop-1 acc.TypeSearch cost (GPU days)
NASNet-B [50]488M72.8MS1800
CARS-D [45]496M73.3MS0.4
BN-NAS [5]470M75.7MS0.8
OFA [4]406M77.7OS50
RLNAS [48]473M75.6OS-
DONNA [32]501M78.0OS405
# Params451M63.5ZS0.02
ZiCo [25]448M78.1ZS0.4
AZ-NAS [21]462M78.6ZS0.4
VKDNWagg(ours)480M78.8ZS0.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.282, + 0.483, + 0.311 + ], + "angle": 0, + "content": "Table 2. Results on ImageNet-1K [9] in the MobileNetV2 search space, the size of the model is constrained to \\(\\approx 450\\mathrm{M}\\) FLOPS." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.319, + 0.28, + 0.333 + ], + "angle": 0, + "content": "exposed to various inputs," + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.334, + 0.483, + 0.363 + ], + "angle": 0, + "content": "- Expressivity (E) [21] assesses isotropy and uniformity of the features distribution across all orientations," + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.364, + 0.482, + 0.409 + ], + "angle": 0, + "content": "- Trainability (T) [21] captures ability of the network to keep stable gradient propagation between the layers by inspecting the spectrum of the Jacobian matrix," + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.41, + 0.482, + 0.425 + ], + "angle": 0, + "content": "- FLOPs (F) is the number of FLOPs for one forward pass." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.334, + 0.483, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.483, + 0.517 + ], + "angle": 0, + "content": "Model-driven aggregation. When accuracies of a sufficient number of architectures in the search space are known, model-driven aggregation can be used to train a regression model to combine individual rankings. The trained model is then used to predict accuracy for unseen networks in the same search space. We evaluate three different models:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.518, + 0.483, + 0.594 + ], + "angle": 0, + "content": "- VKDNW\\(_m\\) where the eigenvalues \\(\\lambda_k\\) of the FIM matrix are used directly as features in companion with \\(\\aleph\\) (number of trainable layers) and FLOPs to allow a more complex proxy of the diversity of the eigenvalues and the network complexity than the simple entropy Eq. (11)," + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.593, + 0.483, + 0.623 + ], + "angle": 0, + "content": "- \\((\\mathrm{VKDNW} + \\mathrm{ZCS})_m\\) where we additionally include all other zero-cost scores available (see Table 1)," + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.623, + 0.483, + 0.654 + ], + "angle": 0, + "content": "- (VKDNW+ZCS+GRAF)\\(_m\\) where we add network graph features from [14]." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.518, + 0.483, + 0.654 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.663, + 0.185, + 0.679 + ], + "angle": 0, + "content": "5.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.483, + 0.777 + ], + "angle": 0, + "content": "We have conducted experiments in the NAS-Bench-201 [10] and MobileNetV2 [26, 38] search spaces. To obtain easily comparable results, we used 64 randomly generated input images to compute our score as in [21]. For the methods that rely on the knowledge of true labels, input data from the respective datasets were used." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "NAS-Bench-201. The dataset consists of 15,625 networks for which validation accuracies for CIFAR-10, CIFAR-100 [18] and ImageNet16-120 [7] after training for 200 epochs are provided. The networks are characterized by unique cell structures comprising of several types of operation choices. As one of the possible choices is zero operation, it's possible that some computation edges don't receive any input or cannot propagate their results to the out" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.093, + 0.706, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.093, + 0.899, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.244, + 0.704, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.244, + 0.899, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.4, + 0.907, + 0.47 + ], + "angle": 0, + "content": "Figure 3. Components of AZ-NAS [21] and our VKDNW are compared w.r.t. correlation with \\(\\aleph\\) (number of trainable layers), in the NAS-Bench-201 search space [10] on ImageNet16-120 [7] dataset. Our VKDNW proxy has the lowest correlation, ie. is the most invariant to the size of the model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.907, + 0.666 + ], + "angle": 0, + "content": "put, leading to same computation graphs for different architectures, thus duplicating networks. Following the practice from NAS-Bench-101 [46], we report results on 9,445 unique structures (also in [14, 30]) and we refer the reader to Supplementary material for results on all networks. We measure proxies performance via Kendall's \\(\\tau\\) (KT) and Spearman's \\(\\rho\\) (SPR) correlations with validation accuracies together with Normalized Discounted Cumulative Gain \\((\\mathrm{nDCG}_{1000})\\), we write nDCG for short) from Sec. 3, averaged over 5 independent runs. For the model-driven aggregation, we trained a random forest model on 1024 networks for 100 iterations, and used the rest for testing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.669, + 0.909, + 0.866 + ], + "angle": 0, + "content": "In Table 1, we compare our method with existing NAS methods and show that \\(\\mathrm{VKDNW_{agg}}\\) outperforms all methods with a significant margin in all three metrics. The benefit of our method is two-fold: a) it identifies the best networks due to its high performance in nDCG, which is the desirable property in NAS b) the ranking is consistent across the whole network search space due to high KT and SPR correlations. Note that we can only make these observations by combining standard correlation metrics and our newly proposed nDCG. We also show that \\(\\mathrm{VKDNW_{single}}\\) outperforms all other single-rank proxies in all metrics on ImageNet16-120, in SPR on all three datasets, while being among the highest on CIFAR-10 and CIFAR-100." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "MobileNetV2. We search for the best network configuration in the MobileNetV2 space [38], while constraining" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14887" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.124, + 0.089, + 0.454, + 0.26 + ], + "angle": 0, + "content": "
VJETF(KT)(SPR)nDCG
✓0.6220.8140.608
✓0.6030.7810.339
✓0.5880.7790.274
✓0.3530.5170.233
✓0.5450.7180.403
✓✓✓✓0.7170.8910.623
✓✓✓✓0.7060.8710.553
✓✓✓✓0.7360.9050.695
✓✓✓✓0.7220.8960.658
✓✓✓✓0.7350.9010.646
✓✓✓✓✓0.7430.9060.664
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.271, + 0.483, + 0.368 + ], + "angle": 0, + "content": "Table 3. Components of VKDNW\\(_{\\text{agg}}\\) rank with non-linear aggregation. Consistency is shown with respect to Kendall's \\(\\tau\\) (KT), Spearman's \\(\\rho\\) (SPR) and Normalized Discounted Cumulative Gain (nDCG) with \\(P = 1000\\) on ImageNet16-120 image dataset [7]. Here V, J, E, T and F stand for VKDNW, Jacov, expressivity, trainability and FLOPs respectively (see Sec. 5.1). Table of all combinations can be found in the Supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.377, + 0.483, + 0.497 + ], + "angle": 0, + "content": "the model size to approximately 450M FLOPs. We ran 100,000 iterations of the evolutionary search algorithm [21] for approximately 10 hours, meaning 100,000 different architectures were evaluated. We then took the best network from the search and trained it for 480 epochs on ImageNet-1K [9], using the same hyper-parameter setting as in [21, 25]. The final training of the model took 7 days on 8xNVidia A100 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.498, + 0.484, + 0.558 + ], + "angle": 0, + "content": "As seen in Table 2, our method \\(\\mathrm{VKDNW_{agg}}\\) outperforms all prior approaches - even train-based approaches (denoted MS and OS) that incur much higher computational costs for the search." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.569, + 0.203, + 0.584 + ], + "angle": 0, + "content": "5.3. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.593, + 0.483, + 0.623 + ], + "angle": 0, + "content": "We ablate our method in the NAS-Bench-201 search space [10] on ImageNet16-120 [7] validation accuracies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.822 + ], + "angle": 0, + "content": "Orthogonality of VKDNW. Our score VKDNW has a desirable property that it is based on information orthogonal to the size of the network: in Figure 3 we can see that unlike previous work, VKDNW is not correlated with the network size measured by \\(\\aleph\\) (number of trainable layers). We believe this property is key when stepping into much larger search spaces, however components of the previous state-of-the-art method AZ-NAS lack such a property (also Fig. 3 and Tab. 4). We conjecture that improvement in the key metrics by VKDNW is caused by this orthogonality feature. In Table 1 we also provide comparison to previous model-driven method and show that adding our feature significantly improves the ranking in all considered metrics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Components of aggregated rank. Our single-rank variant \\(\\mathrm{VKDNW}_{\\mathrm{single}}\\) is the strongest component of the aggregated \\(\\mathrm{VKDNW}_{\\mathrm{agg}}\\) as can be seen in Table 3 where it outperforms rest of the rankings with especially large margin in nDCG, thus, it strongly discriminates high-accuracy net" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.089, + 0.905, + 0.177 + ], + "angle": 0, + "content": "
(KT)
CIFAR-10CIFAR-100ImageNet16-120
VKDNW0.041-0.090-0.107
trainability0.2200.2200.253
expressivity0.5390.5390.539
progressivity0.3980.3980.385
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.187, + 0.907, + 0.243 + ], + "angle": 0, + "content": "Table 4. Kendall's \\(\\tau\\) (KT) of \\(\\aleph\\) (number of trainable layers) with components of AZ-NAS compared to our new method VKDNW on three datasets on NAS-Bench-201 search space [10] and multiple image datasets." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.261, + 0.905, + 0.363 + ], + "angle": 0, + "content": "
Batch SizeRandom InputReal Input
KTSPRnDCGKTSPRnDCG
80.6150.8080.6170.6140.8050.599
160.6140.8080.6040.6080.8000.595
320.6190.8120.6110.6120.8040.594
640.6210.8140.6170.6140.8060.595
1280.6210.8140.6160.6150.8060.591
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.373, + 0.907, + 0.429 + ], + "angle": 0, + "content": "Table 5. Our method VKDNW evaluated for different batch sizes using either randomly generated input data or real images with respect to Kendall's \\(\\tau\\) (KT), Spearman's \\(\\rho\\) (SPR) and Normalized Discounted Cumulative Gain with \\(P = 1000\\) (nDCG)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.442, + 0.692, + 0.456 + ], + "angle": 0, + "content": "works compared to others." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.907, + 0.596 + ], + "angle": 0, + "content": "Random or real input. As our method is computed using generated random data (white noise) it does not rely on any real dataset and is therefore applicable also in situations where no reliable data are available. Table 5 shows that by relying just on random data we do not lose any performance with respect to all key metrics. Moreover, the performance remains relatively stable across different batch sizes. We chose batch size 64 as larger batch size does not bring better results." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.6, + 0.906, + 0.631 + ], + "angle": 0, + "content": "Due to lack of space, we kindly refer the reader to Supplementary material for further ablations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.645, + 0.634, + 0.66 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.906, + 0.748 + ], + "angle": 0, + "content": "We proposed a new training-free NAS proxy called Variance of Knowledge of Deep Network Weights (VKDNW). The method has strong theoretical support and achieved state-of-the-art results on three public datasets and in two search spaces." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Based on throughout evaluation and comparison to other existing approaches, it has been shown that it provides an information orthogonal to the network size leading to a zero-cost ranking were contribution of the network size and architecture feasibility are separated. We have also shown that previously used correlation metrics for proxy evaluations do not sufficiently assess the key ability to discriminate top networks and to address this problem we have proposed a new evaluation metric and re-evaluated previous methods with the new metric." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14888" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.222 + ], + "angle": 0, + "content": "The research was supported by Czech Science Foundation Grant No. 24-10738M. The access to the computational infrastructure of the OP VVV funded project CZ.02.1.01/0.0/0.0/16_019/0000765 “Research Center for Informatics” and of the Ministry of Education, Youth and Sports of the Czech Republic project e-INFRA CZ (ID:90254) is also gratefully acknowledged." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.233, + 0.188, + 0.249 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.258, + 0.483, + 0.3 + ], + "angle": 0, + "content": "[1] Mohamed S Abdelfattah, Abhinav Mehrotra, Lukasz Dudziak, and Nicholas D Lane. Zero-cost proxies for lightweight NAS. arXiv preprint arXiv:2101.08134, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.301, + 0.483, + 0.368 + ], + "angle": 0, + "content": "[2] Chris Burges, Tal Shaked, Erin Renshaw, Ari Lazier, Matt Deeds, Nicole Hamilton, and Greg Hullender. Learning to rank using gradient descent. In Proceedings of the 22nd international conference on Machine learning, pages 89-96, 2005." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.37, + 0.483, + 0.412 + ], + "angle": 0, + "content": "[3] Han Cai, Ligeng Zhu, and Song Han. Proxylessnas: Direct neural architecture search on target task and hardware. arXiv preprint arXiv:1812.00332, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.413, + 0.483, + 0.466 + ], + "angle": 0, + "content": "[4] Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. Once-for-all: Train one network and specialize it for efficient deployment. arXiv preprint arXiv:1908.09791, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.468, + 0.484, + 0.537 + ], + "angle": 0, + "content": "[5] Boyu Chen, Peixia Li, Baopu Li, Chen Lin, Chuming Li, Ming Sun, Junjie Yan, and Wanli Ouyang. BN-NAS: Neural architecture search with batch normalization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 307-316, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.538, + 0.483, + 0.592 + ], + "angle": 0, + "content": "[6] Wuyang Chen, Xinyu Gong, and Zhangyang Wang. Neural architecture search on imagenet in fourgpu hours: A theoretically inspired perspective. arXiv preprint arXiv:2102.11535, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.594, + 0.483, + 0.635 + ], + "angle": 0, + "content": "[7] Patryk Chrabaszcz, Ilya Loshchilov, and Frank Hutter. A downsampled variant of imagenet as an alternative to the cifar datasets. arXiv preprint arXiv:1707.08819, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.636, + 0.483, + 0.663 + ], + "angle": 0, + "content": "[8] Harald Cramér. Mathematical methods of statistics. Princeton university press, 1999." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.664, + 0.483, + 0.719 + ], + "angle": 0, + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.72, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[10] Xuanyi Dong and Yi Yang. Nas-bench-201: Extending the scope of reproducible neural architecture search. arXiv preprint arXiv:2001.00326, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[11] Roy Frieden and Robert A Gatenby. Exploratory data analysis using Fisher information. Springer Science & Business Media, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[12] Boris Hanin and David Rolnick. Complexity of linear regions in deep networks. In International Conference on Machine Learning, pages 2596-2604. PMLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[13] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.258, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[14] Gabriela Kadlecova, Jovita Lukasik, Martin Pilát, Petra Vidnerova, Mahmoud Safari, Roman Neruda, and Frank Hutter. Surprisingly strong performance prediction with neural graph features. arXiv preprint arXiv:2404.16551, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[15] Ryo Karakida, Shotaro Akaho, and Shun-ichi Amari. Pathological spectra of the fisher information metric and its variants in deep neural networks. arXiv preprint arXiv:1910.05992, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.206, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[16] Ryo Karakida, Shotaro Akaho, and Shun-ichi Amari. Universal statistics of fisher information in deep neural networks: Mean field approach. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 1032–1041. PMLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.904, + 0.304 + ], + "angle": 0, + "content": "[17] Maurice G Kendall. A new measure of rank correlation. Biometrika, 30(1-2):81-93, 1938." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.334 + ], + "angle": 0, + "content": "[18] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.335, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[19] Frederik Kunstner, Philipp Hennig, and Lukas Balles. Limitations of the empirical fisher approximation for natural gradient descent. Advances in neural information processing systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[20] Byung-Kwan Lee, Junho Kim, and Yong Man Ro. Masking adversarial damage: Finding adversarial saliency for robust and sparse network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15126-15136, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[21] Junghyup Lee and Bumsub Ham. AZ-NAS: Assembling zero-cost proxies for network architecture search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5893-5903, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[22] Jaehoon Lee, Lechao Xiao, Samuel Schoenholz, Yasaman Bahri, Roman Novak, Jascha Sohl-Dickstein, and Jeffrey Pennington. Wide neural networks of any depth evolve as linear models under gradient descent. Advances in neural information processing systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.59, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[23] Namhoon Lee, Thalaiyasingam Ajanthan, and Philip HS Torr. Snip: Single-shot network pruning based on connection sensitivity. arXiv preprint arXiv:1810.02340, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[24] Steven J Leon, Lisette De Pillis, and Lisette G De Pillis. Linear algebra with applications. Pearson Prentice Hall Upper Saddle River, NJ, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[25] Guihong Li, Yuedong Yang, Kartikeya Bhardwaj, and Radu Marculescu. ZiCo: Zero-shot NAS via inverse coefficient of variation on gradients. In The Eleventh International Conference on Learning Representations, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[26] Ming Lin, Pichao Wang, Zhenhong Sun, Hesen Chen, Xiuyu Sun, Qi Qian, Hao Li, and Rong Jin. Zen-nas: A zero-shot nas for high-performance image recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 347-356, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[27] Hanxiao Liu, Karen Simonyan, and Yiming Yang. DARTS: Differentiable architecture search. arXiv preprint arXiv:1806.09055, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[28] Alexander Ly, Maarten Marsman, Josine Verhagen, Raoul PPP Grasman, and Eric-Jan Wagenmakers. A tutorial on fisher information. Journal of Mathematical Psychology, 80:40-55, 2017." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14889" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[29] James Martens. New insights and perspectives on the natural gradient method. Journal of Machine Learning Research, 21 (146):1-76, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[30] Abhinav Mehrotra, Alberto Gil CP Ramos, Sourav Bhattacharya, Lukasz Dudziak, Ravichander Vipperla, Thomas Chau, Mohamed S Abdelfattah, Samin Ishtiaq, and Nicholas Donald Lane. Nas-bench-asr: Reproducible neural architecture search for speech recognition. In International Conference on Learning Representations, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[31] Joe Mellor, Jack Turner, Amos Storkey, and Elliot J Crowley. Neural architecture search without training. In International conference on machine learning, pages 7588-7598. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[32] Bert Moons, Parham Noorzad, Andrii Skliar, Giovanni Mariani, Dushyant Mehta, Chris Lott, and Tijmen Blankevoort. Distilling optimal neural networks: Rapid search in diverse spaces. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12229-12238, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.416 + ], + "angle": 0, + "content": "[33] Xuefei Ning, Changcheng Tang, Wenshuo Li, Zixuan Zhou, Shuang Liang, Huazhong Yang, and Yu Wang. Evaluating efficient performance estimators of neural architectures. Advances in Neural Information Processing Systems, 34: 12265-12277, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[34] Roman Novak, Jascha Sohl-Dickstein, and Samuel S Schoenholz. Fast finite width neural tangent kernel. In International Conference on Machine Learning, pages 17018-17044. PMLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[35] Hyeyoung Park and Kwanyong Lee. Adaptive natural gradient method for learning of stochastic neural networks in mini-batch mode. Applied Sciences, 9(21):4568, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[36] Jeffrey Pennington and Pratik Worah. The spectrum of the fisher information matrix of a single-hidden-layer neural network. Advances in neural information processing systems, 31, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[37] C Radhakrishna Rao. Information and the accuracy attainable in the estimation of statistical parameters. In Breakthroughs in Statistics: Foundations and basic theory, pages 235-247. Springer, 1992." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[38] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *Proceedings of the IEEE conference on computer vision and pattern recognition*, pages 4510-4520, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[39] Charles Spearman. The proof and measurement of association between two things. 1961." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[40] Kunio Tanabe and Masahiko Sagae. An exact cholesky decomposition and the generalized inverse of the variance-covariance matrix of the multinomial distribution, with applications. Journal of the Royal Statistical Society: Series B (Methodological), 54(1):211-219, 1992." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[41] Hidenori Tanaka, Daniel Kunin, Daniel L Yamins, and Surya Ganguli. Pruning neural networks without any data by iteratively conserving synaptic flow. Advances in neural information processing systems, 33:6377-6389, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[42] Chaoqi Wang, Guodong Zhang, and Roger Grosse. Picking winning tickets before training by preserving gradient flow. arXiv preprint arXiv:2002.07376, 2020." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[43] Colin White, Mikhail Khodak, Renbo Tu, Shital Shah, Sébastien Bubeck, and Debadeepta Dey. A deeper look at zero-cost proxies for lightweight NAS. *ICLR Blog Track*, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[44] Huan Xiong, Lei Huang, Mengyang Yu, Li Liu, Fan Zhu, and Ling Shao. On the number of linear regions of convolutional neural networks. In International Conference on Machine Learning, pages 10514-10523. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[45] Zhaohui Yang, Yunhe Wang, Xinghao Chen, Boxin Shi, Chao Xu, Chunjing Xu, Qi Tian, and Chang Xu. CARS: Continuous evolution for efficient neural architecture search. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1829-1838, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[46] Chris Ying, Aaron Klein, Eric Christiansen, Esteban Real, Kevin Murphy, and Frank Hutter. Nas-bench-101: Towards reproducible neural architecture search. In International conference on machine learning, pages 7105-7114. PMLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.905, + 0.402 + ], + "angle": 0, + "content": "[47] Wang Yining, Wang Liwei, Li Yuanzhi, He Di, Chen Wei, and Liu Tie-Yan. A theoretical analysis of ndcg ranking measures. In Proceedings of the 26th annual conference on learning theory, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[48] Xuanyang Zhang, Pengfei Hou, Xiangyu Zhang, and Jian Sun. Neural architecture search with random labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10907-10916, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.5 + ], + "angle": 0, + "content": "[49] Zhihao Zhang and Zhihao Jia. Gradsign: Model performance inference with theoretical insights. arXiv preprint arXiv:2110.08616, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.503, + 0.905, + 0.571 + ], + "angle": 0, + "content": "[50] Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V Le. Learning transferable architectures for scalable image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8697-8710, 2018." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14890" + } + ] +] \ No newline at end of file diff --git a/2025/Training-free Neural Architecture Search through Variance of Knowledge of Deep Network Weights/740054ef-86e4-4bac-b7b3-53d57b1f3a66_origin.pdf b/2025/Training-free Neural Architecture Search through Variance of Knowledge of Deep Network Weights/740054ef-86e4-4bac-b7b3-53d57b1f3a66_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d7f2fcabc0e6a072ec83c18cfa7c60064885c979 --- /dev/null +++ b/2025/Training-free Neural Architecture Search through Variance of Knowledge of Deep Network Weights/740054ef-86e4-4bac-b7b3-53d57b1f3a66_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6850f4b497280992b56505824362a24f2b4b158a88bb201c76691feee1e06158 +size 414713 diff --git a/2025/Training-free Neural Architecture Search through Variance of Knowledge of Deep Network Weights/full.md b/2025/Training-free Neural Architecture Search through Variance of Knowledge of Deep Network Weights/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3a3f00ef4acf2bfdb383b7fd4af61db7246e9865 --- /dev/null +++ b/2025/Training-free Neural Architecture Search through Variance of Knowledge of Deep Network Weights/full.md @@ -0,0 +1,364 @@ +# Training-free Neural Architecture Search through Variance of Knowledge of Deep Network Weights + +Ondrej Tybl Lukas Neumann + +CMP Visual Recognition Group, Faculty of Electrical Engineering, Czech Technical University in Prague + +ondrej.tybl@cvut.cz lukas.neumann@cvut.cz + +# Abstract + +Deep learning has revolutionized computer vision, but it achieved its tremendous success using deep network architectures which are mostly hand-crafted and therefore likely suboptimal. Neural Architecture Search (NAS) aims to bridge this gap by following a well-defined optimization paradigm which systematically looks for the best architecture, given objective criterion such as maximal classification accuracy. The main limitation of NAS is however its astronomical computational cost, as it typically requires training each candidate network architecture from scratch. + +In this paper, we aim to alleviate this limitation by proposing a novel training-free proxy for image classification accuracy based on Fisher Information. The proposed proxy has a strong theoretical background in statistics and it allows estimating expected image classification accuracy of a given deep network without training the network, thus significantly reducing computational cost of standard NAS algorithms. + +Our training-free proxy achieves state-of-the-art results on three public datasets and in two search spaces, both when evaluated using previously proposed metrics, as well as using a new metric that we propose which we demonstrate is more informative for practical NAS applications. The source code is publicly available at https://www.github.com/ondratybl/VKDNW. + +# 1. Introduction + +In most instances, neural network architectures are designed by authors following the field's "best-practices" or their experience, without any formal and repeatable procedure. This is however inconvenient especially in applications on a large scale. Neural Architecture Search (NAS) aims to bridge this gap by following a well-defined optimization paradigm which systematically looks for the best architecture, given objective criterion such as maximal accuracy. + +The main limitation of Neural Architecture Search (NAS) is however the computational cost, because in the + +![](images/97fd63a8b192a1c196b654c0292adafa60c228dc90dd62d9cefa4dc0ad195af9.jpg) +Figure 1. Training-free NAS methods on ImageNet16-120 [10]. Methods are compared by Normalized Discounted Cumulative Gain (see Sec. 3), our method (VKDNW) is the best also measured by Kendall's $\tau$ and Spearman's $\rho$ correlations (see Table 1). Also note that simple number of trainable layers (below denoted $\aleph$ ) is significantly better trivial proxy than the number of FLOPs. + +most basic NAS setup, it is required to train thousands or more of different deep network architectures from scratch in order just to calculate a single scalar – the objective function value, such as the classification accuracy. This severely limits practical applications of NAS as the size of feasible architecture search space is only a small fraction of the overall space of all networks. + +Training-free Neural Architecture Search (TF-NAS) aims to alleviate this limitation by introducing an objective function proxy which – unlike the actual objective function – does not require training the network. As a result, a good proxy allows the TF-NAS algorithm to explore significantly bigger portions of the network architecture search space compared to traditional NAS, and to find the best network architecture without training a single network. The crucial question is however finding an appropriate objective function proxy. + +In this paper, we present a novel, principled objective function proxy called Variance of Knowledge of Deep Network Weights (VKDNW) for image classification accuracy, which allows us to find optimal network architectures for image classification without training them first. Our method is, to the best of our knowledge, the first successful application of Fisher Information theory [28] in the context of large deep neural networks, and as such allows us to formally describe and quantify the difficulty of network parameters' estimation process. In other words, given a network architecture, our method estimates how easy or hard it will be to train the network. + +Additionally, we also observe that the evaluation metrics used in the TF-NAS community are not well-suited to the problem at hand, because it unnecessarily penalizes for bad proxy accuracy for networks which are not interesting, and vice-versa it does not sufficiently reward proxies which are able to accurately pick out good network architectures. Following this observation, we propose that the Normalized Discounted Cumulative Gain should be used in comparison with other TF-NAS metrics, and show indeed that there are significant differences amongst previously proposed TF-NAS methods when the new metric is considered. + +To summarize, we make the following contributions: + +1. We introduce a novel algorithm for estimation of Fisher Information Matrix spectrum, which is tractable even for models with large number of parameters such as deep networks, that overcomes the usual problems of numerical stability. +2. We introduce a novel principled VKDNW proxy for image classification accuracy. The proxy is based on strong theoretical background and captures uncertainty in weight estimation process. It brings information that is orthogonal to the model size which then allows for efficient combination with previously proposed proxies, leading to state-of-the-art results. +3. We propose a new evaluation metric for TF-NAS proxies which is more relevant to the actual NAS objective as it concentrates on ability of given proxy to identify good networks. + +# 2. Method + +Our zero-shot proxy for image classification accuracy builds on Fisher Information theory [28], therefore we begin by a thorough analysis of Fisher Information Matrix (FIM) of the network weights estimation problem (see Sec. 2.1). We proceed by discussing challenges involved in practical application of FIM in context of large over-parametrized models that lead to only limited success in previous works and present our contributions to overcome these limitations (see Sec. 2.2). Finally, we propose a novel FIM-based proxy for NAS algorithms (Sec. 2.3). + +# 2.1. Fisher Information + +The problem of finding the optimal weights of a neural network $f$ for the task of $C$ -class image classification can be seen as a maximum likelihood estimation with a statistical model + +$$ +\sigma_ {\theta} (c \mid x) = \frac {\exp (\Psi_ {c} (x , \theta))}{\sum_ {d = 1} ^ {C} \exp (\Psi_ {d} (x , \theta))}, \quad c = 1, \dots , C \tag {1} +$$ + +where $\sigma_{\theta}(\cdot |x)$ denotes the a posteriori distribution of the labels given input image $x$ and $\Psi (x,\theta)\in \mathbb{R}^{C} = (\Psi_{1}(x,\theta),\dots ,\Psi_{C}(x,\theta))$ is the network output (logits) given the weight vector $\theta \in \mathbb{R}^p$ , i.e. the network weights. We describe the process of training as finding the optimal weight vector $\theta^{*}$ that fits our data and we posit that network architectures should be characterised by how easy it is to estimate their optimal network weights $\theta^{*}$ . We build upon statistical learning theory and use Fisher Information [28] framework to formally describe expected behaviour of the training process of a given deep network $f$ . + +The Fisher Information Matrix (FIM) encompasses information on the difficulty of the parameter estimation problem and it plays a crucial role in several fundamental results, which we apply below in the context of deep networks. The FIM of a network $f$ and its set of weights $\theta \in \mathbb{R}^p$ is given as + +$$ +F (\theta) := \mathbb {E} \left[ \nabla_ {\theta} \sigma_ {\theta} (c | x) \nabla_ {\theta} \sigma_ {\theta} (c | x) ^ {T} \right] \in \mathbb {R} ^ {p \times p}, \tag {2} +$$ + +where we take the expected value $\mathbb{E}$ with respect to the joint distribution of $(x, c)$ . For more detailed account on Fisher Information theory and its applications in the context of machine learning, we kindly refer reader to [16, 20, 29, 35, 36]. + +Cramér-Rao bound. The first part of estimation theory we build upon is the inverse of the FIM, known as the Cramér-Rao bound (see [11]), which is the asymptotic variance of the estimated weights (i.e. the uncertainty coming from the data variability). Thus, the larger the matrix norm of the FIM, the more certain we are about the weights $\theta$ . More specifically, any data-dependent estimator $\hat{\theta}_n = \hat{\theta}_n(x_1,\ldots ,x_n)$ is a random vector with randomness coming from the (independent) choice of input images $x_{1},\ldots ,x_{n}$ and as such has some variance matrix $\operatorname{Var}\left(\hat{\theta}_n\right)$ . The famous result (see [8, 37]) named in honour of H. Cramér and C. R. Rao states that if $\hat{\theta}_n$ is unbiased then the variance is bounded from below as + +$$ +\operatorname {V a r} \left(\hat {\theta} _ {n}\right) \geq \frac {1}{n} F ^ {- 1} (\theta) \tag {3} +$$ + +and for maximum likelihood estimation this bound is attained as the number of input images grows to infinity $n \to \infty$ . + +We formally show (see Supplementary material) that for each weight $\theta(j)$ the mean square error of our estimation is controlled by the diagonal element of the FIM inverse as + +$$ +\mathbb {E} \left(\hat {\theta} _ {n} (j) - \theta (j)\right) ^ {2} \geq \frac {1}{n} \left(F ^ {- 1} (\theta)\right) _ {j j}. \tag {4} +$$ + +Therefore, knowing the FIM allows us to evaluate how certain we are about the weight estimates. We can go even further by inspecting the eigenvalues of $F(\theta)$ . Denoting the largest and smallest eigenvalue of the FIM as $\lambda_{\mathrm{min}}$ and $\lambda_{\mathrm{max}}$ respectively, then we have that there exist linear combination coefficients $e_{\mathrm{min}}$ and $e_{\mathrm{max}}$ of unit size so that + +$$ +\mathbb {E} \left(e _ {\min} ^ {T} \hat {\theta} _ {n} - e _ {\min} ^ {T} \theta\right) ^ {2} \geq \frac {1}{n \lambda_ {\min}}, +$$ + +$$ +\mathbb {E} \left(e _ {\max } ^ {T} \hat {\theta} _ {n} - e _ {\max } ^ {T} \theta\right) ^ {2} \geq \frac {1}{n \lambda_ {\max }}, \tag {5} +$$ + +indicating that if the difference between the largest and smallest eigenvalue is large then there exist combinations of weights with very large difference in the estimation certainty. Altogether, the more the eigenvalues of the FIM are similar, the more similar is also the variance in the weight estimation across all model weights. + +Prediction sensibility. The change of the model prediction measured by the KL-divergence when subject to a small perturbations of the weights is given as + +$$ +D _ {K L} \left(\sigma_ {\theta + \theta_ {\delta}} (\cdot | x), \sigma_ {\theta} (\cdot | x)\right) \approx \frac {1}{2} \theta_ {\delta} ^ {T} F (\theta) \theta_ {\delta} \tag {6} +$$ + +for a small weight perturbation vector $\theta_{\delta} \in \mathbb{R}^{p}$ . Thus, $F(\theta)$ measures volatility of the predictions subject to the weight change – if the difference between the largest and the smallest eigenvalues of the FIM is large, then there exist perturbation directions $\theta_{min}, \theta_{max}$ corresponding to $\lambda_{min}, \lambda_{max}$ such that + +$$ +\begin{array}{l} D _ {K L} \left(\sigma_ {\theta + \theta_ {m i n}} (\cdot | x), \sigma_ {\theta} (\cdot | x)\right) \approx \frac {1}{2} \theta_ {m i n} ^ {T} F (\theta) \theta_ {m i n} = \\ = \frac {1}{2} \lambda_ {\min } \| \theta_ {\min } \| ^ {2} \ll \frac {1}{2} \lambda_ {\max } \| \theta_ {\max } \| ^ {2} = \tag {7} \\ = \frac {1}{2} \theta_ {m a x} ^ {T} F (\theta) \theta_ {m a x} \approx D _ {K L} \left(\sigma_ {\theta + \theta_ {m a x}} (\cdot | x), \sigma_ {\theta} (\cdot | x)\right) \\ \end{array} +$$ + +and therefore in some directions a small change of the weights has much larger impact on the prediction than in others, making the model less balanced. For further discussion, see the Supplementary material. + +# 2.2. Empirical Fisher Information Matrix implementation + +When independent and identically distributed sample images $x_{n}$ are available, empirical FIM $\hat{F} (\theta)$ is defined as + +$$ +\hat {F} (\theta) := \frac {1}{n} \sum_ {n = 1} ^ {N} \mathbb {E} _ {\sigma_ {\theta}} \left[ \nabla_ {\theta} \sigma_ {\theta} (c | x _ {n}) \nabla_ {\theta} \sigma_ {\theta} (c | x _ {n}) ^ {T} \right] \tag {8} +$$ + +where $\mathbb{E}_{\sigma_{\theta}}$ now denotes the expectation with respect to the model prediction $\sigma_{\theta}$ . + +We would like to first emphasize several crucial aspects and contributions of this paper that lead to the first success of the Fisher Information theory (having e.g. [1] in mind) in the context of Neural Architecture Search, despite Fisher Information being one of the first obvious choices for deep network analysis: + +1. Following [19] we write the empirical FIM (8) as + +$$ +\begin{array}{l} \hat {F} (\theta) = \frac {1}{n} \sum_ {n = 1} ^ {N} \left[ \nabla_ {\theta} \Psi (x _ {n}, \theta) ^ {T} \left(\operatorname {d i a g} \left(\sigma_ {\theta} (\cdot , x _ {n})\right) - \right. \right. \\ \left. \sigma_ {\theta} (\cdot , x _ {n}) \sigma_ {\theta} (\cdot , x _ {n}) ^ {T}\right) \nabla_ {\theta} \Psi (x _ {n}, \theta) ]. \tag {9} \\ \end{array} +$$ + +and we further decompose the inner matrix $\mathrm{diag}(\sigma_{\theta}(\cdot ,x_n)) - \sigma_{\theta}(\cdot ,x_n)\sigma_{\theta}(\cdot ,x_n)^T$ using analytical formulas from [40] to avoid numerical instability of the computation as we arrive at a feasible representation + +$$ +\hat {F} (\theta) = \frac {1}{n} \sum_ {n = 1} ^ {N} A _ {n} ^ {T} A _ {n} \tag {10} +$$ + +for some matrices $A_{n}\in \mathbb{R}^{C\times p}$ . We observed that networks typically yield very imbalanced outputs at initialization (i.e. every network prioritizes few classes over the rest) making it extremely important not to exclude the factor $\mathrm{diag}(\sigma_{\theta}(\cdot ,x_n)) - \sigma_{\theta}(\cdot ,x_n)\sigma_{\theta}(\cdot ,x_n)^T$ , which is however difficult to compute without underflow/overflow. + +2. The dimension of FIM is equal to $p$ (the number of all trainable parameters) and therefore direct computation of the eigenvalues is numerically intractable and unstable. However, our results show that if a small number of representative parameters is drawn, then computation becomes stable while the discrimination power does not suffer. We used a simple rule where a single parameter from each trainable layer (not including batch normalization) is chosen. Stability with respect to choice of such parameters can be found in Supplementary material. +3. FIM of large networks typically suffers from having pathological spectrum, i.e. there usually exist zero eigenvalue and its multiplicity is large (see [15]) and thus the eigenvalues estimation due to a large condition number is imprecise. However, as we deal with a symmetric positive-semidefinite matrix, the eigenvalues actually coincide with the singular values (see [24]), for which the estimation algorithm performs better. + +Let us also emphasize that there is a common misconception in part of the community as it is often mistakenly assumed that one can simply use the true labels of $x_{n}$ in (8) in place of $c$ . However, such definition is then meaningless as it does not approximate the FIM in the classical Monte Carlo sense (see Supplementary material and [19] for a comparison). That means that the empirical FIM does not depend on the true labels and therefore our method does not require real data - we use random input instead. + +# 2.3. Variance of Knowledge for Deep Network Weights + +In order to characterize properties of the parameter estimation process for a given network $f$ through the lens of Fisher Information theory, we inspect the eigenvalues of the empirical FIM $\hat{F}(\theta_{\mathrm{init}})$ and define the entropy of Variance of Knowledge for Deep Network Weights (VKDNW) as + +$$ +\begin{array}{l} \operatorname {V K D N W} (f) := - \sum_ {k = 1} ^ {9} \tilde {\lambda} _ {k} \log \tilde {\lambda} _ {k} \\ \tilde {\lambda} _ {k} = \frac {\lambda_ {k}}{\sum_ {j = 1} ^ {9} \lambda_ {j}}, \quad k = 1, \dots , 9, \tag {11} \\ \end{array} +$$ + +where $\lambda_{k}$ denotes the $k$ -th decile of the FIM eigenvalues as the representation of the FIM spectrum1, and $\theta_{\mathrm{init}}$ denotes network weights at initialization. + +Our score therefore measures the diversity of the FIM eigenvalues, and from the entropy theory we know that VKDNW attains its maximum exactly when all the eigenvalues $\lambda_{k}$ are equal and VKDNW gets lower as the eigenvalues become more different. Based on the discussion in Sec. 2.1 we see that VKDNW is high when the uncertainty in all model weight combinations are similar (see Cramér-Rao bound) and there are no directions in the weight space that would influence the network prediction substantially differently than others. Due to the fact that we have normalized both the number of eigenvalues under consideration (by taking a fixed number of representatives irrespective of the network size) and the magnitude of the eigenvalues, VKDNW is independent of network size (number of network weights $p$ ). + +Let us note, that we are familiar with the fact that even though our motivation was (among others) based on Cramér-Rao bound that assumes evaluation of FIM at the correct weight vector $\theta_{X}$ that fits the data, which is typically far away from the weights given at initialization. However, our empirical results below support the hypothesis that the evaluation despite being in the wrong point brings valuable information. + +Ranking networks for NAS. The proposed VKDNW score is independent of network size, which is extremely beneficial to compare individual structures of network architectures. Thus, it does not aim to capture capacity, rather it targets feasibility of the computation graph given number of operations. However, when comparing different network structures and different network sizes together as it is done in NAS, one indeed needs to take network size into account as well, because naturally larger networks have bigger capacity and therefore tend to have higher accuracy. To + +![](images/d5f967d8c0886a78528adbb6ab92ca4e59bafe67932a202795c93e69628d8d76.jpg) +Figure 2. Toy example of two rankings on 10 networks. We plot accuracies ordered by the rankings and evaluation metrics Kendall's $\tau$ (KT) and Spearman's $\rho$ (SPR) correlations and Normalized Discounted Cumulative Gain $(\mathrm{nDCG}_5)$ . + +capture also the capacity we proxy the network size by the number of layers with weights that we denote as $\aleph(f)$ for a network $f$ and we introduce the ranking + +$$ +\operatorname {V K D N W} _ {\text {s i n g l e}} (f) := \aleph (f) + \operatorname {V K D N W} (f) \tag {12} +$$ + +Here we leverage the fact that VKDNW as an entropy of some quantity is always between 0 and 1 and by summing it with an integer-valued quantity we in fact obtain that we have first grouped networks by our size proxy $\aleph$ and then within each group of similar networks sizes we order them by VKDNW. + +# 3. Evaluation Metrics + +In this section, we dive into the problem of evaluating (image) classification accuracy proxies for training-free NAS (TF-NAS). In NAS algorithms, the proxy actually does not have to predict the classification accuracy in absolute terms – since we are after finding the "best" network for given task – it's sufficient that the proxy properly ranks the networks, ordering them from worst to the best. + +Suppose that a collection of $K$ networks with validation accuracies $\mathrm{acc}_1, \ldots, \mathrm{acc}_K$ have been ranked by a TF-NAS proxy as $r_1, \ldots, r_K \in \mathbb{R}$ . The standard way of evaluating [14, 21, 25] how well the ranks correspond to the accuracies is to compute Kendall's $\tau$ (KT, [17]) and Spearman's $\rho$ (SPR, [39]) rank correlation coefficients. Kendall's $\tau$ is given as + +$$ +\tau := \frac {n _ {c} - n _ {d}}{\sqrt {n _ {c} + n _ {d} + n _ {1}} \sqrt {n _ {c} + n _ {d} + n _ {2}}}, \tag {13} +$$ + +where $n_c$ denotes the number of concordant pairs $(\mathrm{acc}_k, r_k)$ , $n_d$ the number of discordant pairs and $n_1$ (resp. $n_2$ ) denote the number of ties in $\mathrm{acc}_k$ only (resp. in $r_k$ only). In our context, (KT) relates after some normalization to a probability that two randomly chosen network rankings $r_k, r_l$ are ordered correctly according to their accuracies $\mathrm{acc}_k, \mathrm{acc}_l$ . In contrast, (SPR) is defined to account more for outliers, i.e. it tends to put a large penalization if + +there are some networks for which the difference between the rank of the accuracy $\mathrm{acc}_k$ and the rank $r_k$ is large as it is just the classical correlation coefficient however applied on the orders of the assessed quantities. + +Normalized Discounted Cumulative Gain. Both (KT) and (SPR) are good evaluation metrics when our interest lies in comparison of ranking proxies when all networks involved are of the same importance. However, in NAS we are primarily interested if a ranking helps us to pick best networks from a given collection and less interested how well the ranking compares two networks with low accuracy. Finding inspiration in information retrieval where one measures the quality of a system that retrieves resources relevant to an input query, we propose to use Normalized Discounted Cumulative Gain (nDCG) [2] as a more relevant metric to measure quality of TF-NAS proxies. The nDCG metric is defined with a key requirement that highly relevant documents are more valuable when they appear earlier in the search engine results (i.e., in higher-ranking positions). This requirement can be reformulated for the task of neural architecture search as networks with high accuracies are more valuable when they appear on higher-ranking positions. + +We thus define the metric as follows: first we order the networks by their ranking $r_1, \ldots, r_K$ so that $r_{k_1} = K, r_{k_2} = K - 1, \ldots$ and we compute + +$$ +\mathrm {n D C G} _ {P} := \frac {1}{Z} \left(\sum_ {j = 1} ^ {P} \frac {2 ^ {\mathrm {a c c} _ {k _ {j}}} - 1}{\log_ {2} (1 + j)}\right), \tag {14} +$$ + +where $P \in \mathbb{N}$ is a parameter determining how many top-ranked networks do we consider (e.g. it corresponds to the population size in an evolution algorithm) and $Z$ is a normalization factor that represents the ideal discounted cumulative gain so that nDCG is equal to one for a perfect fit. + +The higher nDCG the better the ranking is as it is a weighted average of transformed top-ranked accuracies. In [47] it is shown that the choice of the discount factor of given by inverse logarithm is a good choice as it nDCG then well separates different ranking systems. + +Toy Example. Let us briefly demonstrate the weak ability of (KT) and (SPR) to distinguish rankings that poorly discriminate networks with high accuracy. Suppose a collection of 10 networks with their validation accuracies $100,90,\dots ,10$ is given and our aim is to rank them from worst to the best. We evaluate two different toy rankings: a) ranking damaged_top which perfectly fits the accuracies, only the two top networks are swapped with the third and fourth, b) ranking perfect_top which fits perfectly for the top-performing networks, however, now the order of the four worst networks is reversed (see Figure 2). The + +damaged_top ranking is an example that should be evaluated worse for the architecture search task compared to perfect_top as in the first case the best networks are not placed first, while in the second they are. On the other hand, even though perfect_top is not a perfect fit, it still discriminates the top networks ideally and only struggles for networks with low accuracy, which are not interesting for NAS which is after the best networks. From the correlation perspectives of (KT), (SPR) damaged_top is better than perfect_top, therefore, if we rely only on these two evaluation metrics we prefer ranking that is worse for the architecture search task as the top networks are ranked worse. On the other hand, for perfect_top we obtain ideal nDCG $_5$ while it drops to 0.5 for damaged_top. We conclude that using nDCG $_5$ we choose ranking that has higher discriminatory power for good networks. + +We also note that from statistical perspective the (KT) and (SPR) of damaged_top is significantly higher than of a random ranking3, and therefore we'd conclude that damaged_top is not independent of ground truth accuracy. On the other hand, when we perform the same statistical test on damaged_top using nDCG5, we conclude that the damaged_top is not significantly better than ranking networks randomly4. + +# 4. Related Work + +Zero-shot NAS aims to rank given networks in a training-free manner based on their (a priori unknown) final performance, which allows to prune the huge search space with limited costs when seeking for optimal architecture [10, 27, 46] and other configuration [3, 26, 38]. + +Throughout previous works different approaches for ranking computation can be found. In many of the works the gradient with respect to the network weights is investigated (GradNorm, GraSP, SNIP, Synflow, [1, 23, 41, 42]), i.e. leveraging the first order approximation of the network. In Jacov [1] correlations of the Jacobian matrices among various input samples are compared; in NASWOT [31] the linear maps induced by data points are examined; Zen [26] uses the approximation of gradient with respect to featuremaps; GradSign [49] compares the optimization landscape at the level individual training samples; or in ZiCo [25] the gradients from multiple forward and backward passes preferring large magnitude and low variance. + +Other methods aggregate multiple sources aiming at obtaining a better informed ranks: TE-NAS [6] uses both the number of linear regions [12, 44] and the condition number of Neural Tangent Kernel [13, 22]. However, it is well + +
TypeCIFAR-10CIFAR-100ImageNet16-120
KTSPRnDCGKTSPRnDCGKTSPRnDCG
Simple rankings
FLOPsS0.6230.7990.7450.5860.7630.5760.5450.7180.403
GradNorm [1]S0.3280.4380.5090.3410.4510.2780.3100.4180.265
GraSP [1, 42]S0.3520.5050.5180.3490.4980.2840.3590.5020.281
SNIP [1, 23]S0.4310.5910.5130.4400.5970.2860.3890.5210.286
SynFlow [1, 41]S0.5610.7580.7090.5530.7500.5940.5310.7190.511
Jacov [1]S0.6160.8000.5400.6390.8200.4020.6020.7790.356
NASWOT [31]S0.5710.7620.6070.6070.7990.4750.6050.7940.490
ZenNAS [26]S0.1020.1030.1200.0790.0720.1150.0910.1090.073
GradSign† [49]S·0.765··0.793··0.783·
ZiCo [25]S0.6070.8020.7510.6140.8090.6070.5870.7790.523
TE-NAS [6]A0.5360.7220.6020.5370.7230.3270.5230.7090.330
AZ-NAS [21]A0.7120.8920.7490.6960.8800.5490.6730.8590.534
No. of trainable layers (N)S0.6260.7670.6710.6460.7870.5250.6230.7640.497
VKDNWsingle (ours)S0.6180.8150.7510.6340.8290.6170.6220.8140.608
VKDNWagg (ours)A0.7500.9190.7850.7530.9190.6360.7430.9060.664
Model-driven rankings
GRAF [14]A0.8200.9530.9350.8090.9480.8580.7960.9410.828
VKDNWm (ours)A0.6470.8310.7500.6360.8210.6020.6110.7980.575
(VKDNW+ZCS)m (ours)A0.8400.9630.9220.8340.9600.8840.8300.9580.843
(VKDNW+ZCS+GRAF)m (ours)A0.8590.9710.9460.8470.9660.8950.8420.9630.867
+ +Table 1. Training-free NAS methods in the NAS-Bench-201 [10] search space, evaluated on three public datasets. Kendall's $\tau$ (KT), Spearman's $\rho$ (SPR) and Normalized Discounted Cumulative Gain (nDCG) are reported, results are averages of 5 independent runs (all stds are $< 0.02$ ). The Type column differentiates single (S) and aggregated (A) rankings. Results are reproduced with code published by their authors, except those marked†, where results from the original paper are taken. + +known that the kernel computation is highly computationally demanding [34]. AZ-NAS [21] assesses expressivity, trainability and progressivity via examination of feature distribution across all orientations and the Jacobian. However, despite AZ-NAS outperforming previous works in some metrics, it is worse than [25] in key NAS-related aspects such as the cumulative gain (see Sec. 5.1). + +Despite a tremendous effort of the community, it was shown that most of the zero-shot NAS methods perform worse than a simple proxy given just by FLOPs or #params [33, 43]. Thus, there is still a gap for improvements also driven by the need of the ranking explainability that would have a satisfactory theoretical support. + +Furthermore, [14] uses a model-driven ranking that however needs a set of networks for which the validation accuracies are known to train the model and the ability of this ranking to generalize as it is fitted on a specific network collection only is disputable. + +Finally, let us discuss the current practice in the NAS methods comparison. Methods are compared either by means of correlations of their scores to accuracy or by reporting the accuracy of the top-ranked network [21, 25]. However, while the first is not tailor-made for the architecture search task and therefore does not assess the desired ranking properties (see Sec. 5.1), comparing performance + +just by the accuracy of a single network is too vulnerable e.g. to the random seed choice. + +# 5. Experiments + +# 5.1. Ranking aggregation + +In addition to using the single ranking of Eq. (12), we also experiment with multiple rankings (as in [14, 21, 25]) to order network architectures by their accuracy. We use two different options: non-linear and model-driven aggregation. + +Non-linear aggregation. The aggregation [21] uses multiplication to combine multiple ranks into a single one, which means a network is highly-ranked if and only if it is highly-ranked in all subsidiary rankings, keeping their influence balanced. That is, denoting rankings $\mathrm{rank}_1,\ldots ,\mathrm{rank}_m$ for some $m\in \mathbb{N}$ we define the aggregated ranking + +$$ +\operatorname {r a n k} _ {\text {a g g}} (f) := \log \Pi_ {j = 1} ^ {m} \operatorname {r a n k} _ {j} (f) \tag {15} +$$ + +for each network $f$ . This aggregation is therefore possible only in the context of a given network collection, such as in evolutionary search. + +In our case, $\mathrm{VKDNW_{agg}}$ aggregates these five proxies: + +- VKDNWsingle (V) ranking (Eq. (12)), +- Jacov (J) [1] measures the activations correlation when + +
MethodFLOPsTop-1 acc.TypeSearch cost (GPU days)
NASNet-B [50]488M72.8MS1800
CARS-D [45]496M73.3MS0.4
BN-NAS [5]470M75.7MS0.8
OFA [4]406M77.7OS50
RLNAS [48]473M75.6OS-
DONNA [32]501M78.0OS405
# Params451M63.5ZS0.02
ZiCo [25]448M78.1ZS0.4
AZ-NAS [21]462M78.6ZS0.4
VKDNWagg(ours)480M78.8ZS0.4
+ +Table 2. Results on ImageNet-1K [9] in the MobileNetV2 search space, the size of the model is constrained to $\approx 450\mathrm{M}$ FLOPS. + +exposed to various inputs, + +- Expressivity (E) [21] assesses isotropy and uniformity of the features distribution across all orientations, +- Trainability (T) [21] captures ability of the network to keep stable gradient propagation between the layers by inspecting the spectrum of the Jacobian matrix, +- FLOPs (F) is the number of FLOPs for one forward pass. + +Model-driven aggregation. When accuracies of a sufficient number of architectures in the search space are known, model-driven aggregation can be used to train a regression model to combine individual rankings. The trained model is then used to predict accuracy for unseen networks in the same search space. We evaluate three different models: + +- VKDNW $_m$ where the eigenvalues $\lambda_k$ of the FIM matrix are used directly as features in companion with $\aleph$ (number of trainable layers) and FLOPs to allow a more complex proxy of the diversity of the eigenvalues and the network complexity than the simple entropy Eq. (11), +- $(\mathrm{VKDNW} + \mathrm{ZCS})_m$ where we additionally include all other zero-cost scores available (see Table 1), +- (VKDNW+ZCS+GRAF) $_m$ where we add network graph features from [14]. + +# 5.2. Results + +We have conducted experiments in the NAS-Bench-201 [10] and MobileNetV2 [26, 38] search spaces. To obtain easily comparable results, we used 64 randomly generated input images to compute our score as in [21]. For the methods that rely on the knowledge of true labels, input data from the respective datasets were used. + +NAS-Bench-201. The dataset consists of 15,625 networks for which validation accuracies for CIFAR-10, CIFAR-100 [18] and ImageNet16-120 [7] after training for 200 epochs are provided. The networks are characterized by unique cell structures comprising of several types of operation choices. As one of the possible choices is zero operation, it's possible that some computation edges don't receive any input or cannot propagate their results to the out + +![](images/8b64c7352464abe41619d6bfc3723ff08879a9dd2ad85c19d55acf23710d5d61.jpg) + +![](images/e4f5f01c1888fabdc9b7380bd8dd2f34a99e4c8ec21c3a7473348ecbc2f28be5.jpg) + +![](images/d77d22122e5743a9ff75154781478d7c808cfa3c1f380a6c73f447a4dd5c9bd9.jpg) +Figure 3. Components of AZ-NAS [21] and our VKDNW are compared w.r.t. correlation with $\aleph$ (number of trainable layers), in the NAS-Bench-201 search space [10] on ImageNet16-120 [7] dataset. Our VKDNW proxy has the lowest correlation, ie. is the most invariant to the size of the model. + +![](images/c1dcc1c3cc570de06fc5b5da33871e3d569e4802867351dd28e012ea923c7e9b.jpg) + +put, leading to same computation graphs for different architectures, thus duplicating networks. Following the practice from NAS-Bench-101 [46], we report results on 9,445 unique structures (also in [14, 30]) and we refer the reader to Supplementary material for results on all networks. We measure proxies performance via Kendall's $\tau$ (KT) and Spearman's $\rho$ (SPR) correlations with validation accuracies together with Normalized Discounted Cumulative Gain $(\mathrm{nDCG}_{1000})$ , we write nDCG for short) from Sec. 3, averaged over 5 independent runs. For the model-driven aggregation, we trained a random forest model on 1024 networks for 100 iterations, and used the rest for testing. + +In Table 1, we compare our method with existing NAS methods and show that $\mathrm{VKDNW_{agg}}$ outperforms all methods with a significant margin in all three metrics. The benefit of our method is two-fold: a) it identifies the best networks due to its high performance in nDCG, which is the desirable property in NAS b) the ranking is consistent across the whole network search space due to high KT and SPR correlations. Note that we can only make these observations by combining standard correlation metrics and our newly proposed nDCG. We also show that $\mathrm{VKDNW_{single}}$ outperforms all other single-rank proxies in all metrics on ImageNet16-120, in SPR on all three datasets, while being among the highest on CIFAR-10 and CIFAR-100. + +MobileNetV2. We search for the best network configuration in the MobileNetV2 space [38], while constraining + +
VJETF(KT)(SPR)nDCG
✓0.6220.8140.608
✓0.6030.7810.339
✓0.5880.7790.274
✓0.3530.5170.233
✓0.5450.7180.403
✓✓✓✓0.7170.8910.623
✓✓✓✓0.7060.8710.553
✓✓✓✓0.7360.9050.695
✓✓✓✓0.7220.8960.658
✓✓✓✓0.7350.9010.646
✓✓✓✓✓0.7430.9060.664
+ +Table 3. Components of VKDNW $_{\text{agg}}$ rank with non-linear aggregation. Consistency is shown with respect to Kendall's $\tau$ (KT), Spearman's $\rho$ (SPR) and Normalized Discounted Cumulative Gain (nDCG) with $P = 1000$ on ImageNet16-120 image dataset [7]. Here V, J, E, T and F stand for VKDNW, Jacov, expressivity, trainability and FLOPs respectively (see Sec. 5.1). Table of all combinations can be found in the Supplementary material. + +the model size to approximately 450M FLOPs. We ran 100,000 iterations of the evolutionary search algorithm [21] for approximately 10 hours, meaning 100,000 different architectures were evaluated. We then took the best network from the search and trained it for 480 epochs on ImageNet-1K [9], using the same hyper-parameter setting as in [21, 25]. The final training of the model took 7 days on 8xNVidia A100 GPUs. + +As seen in Table 2, our method $\mathrm{VKDNW_{agg}}$ outperforms all prior approaches - even train-based approaches (denoted MS and OS) that incur much higher computational costs for the search. + +# 5.3. Ablations + +We ablate our method in the NAS-Bench-201 search space [10] on ImageNet16-120 [7] validation accuracies. + +Orthogonality of VKDNW. Our score VKDNW has a desirable property that it is based on information orthogonal to the size of the network: in Figure 3 we can see that unlike previous work, VKDNW is not correlated with the network size measured by $\aleph$ (number of trainable layers). We believe this property is key when stepping into much larger search spaces, however components of the previous state-of-the-art method AZ-NAS lack such a property (also Fig. 3 and Tab. 4). We conjecture that improvement in the key metrics by VKDNW is caused by this orthogonality feature. In Table 1 we also provide comparison to previous model-driven method and show that adding our feature significantly improves the ranking in all considered metrics. + +Components of aggregated rank. Our single-rank variant $\mathrm{VKDNW}_{\mathrm{single}}$ is the strongest component of the aggregated $\mathrm{VKDNW}_{\mathrm{agg}}$ as can be seen in Table 3 where it outperforms rest of the rankings with especially large margin in nDCG, thus, it strongly discriminates high-accuracy net + +
(KT)
CIFAR-10CIFAR-100ImageNet16-120
VKDNW0.041-0.090-0.107
trainability0.2200.2200.253
expressivity0.5390.5390.539
progressivity0.3980.3980.385
+ +Table 4. Kendall's $\tau$ (KT) of $\aleph$ (number of trainable layers) with components of AZ-NAS compared to our new method VKDNW on three datasets on NAS-Bench-201 search space [10] and multiple image datasets. + +
Batch SizeRandom InputReal Input
KTSPRnDCGKTSPRnDCG
80.6150.8080.6170.6140.8050.599
160.6140.8080.6040.6080.8000.595
320.6190.8120.6110.6120.8040.594
640.6210.8140.6170.6140.8060.595
1280.6210.8140.6160.6150.8060.591
+ +Table 5. Our method VKDNW evaluated for different batch sizes using either randomly generated input data or real images with respect to Kendall's $\tau$ (KT), Spearman's $\rho$ (SPR) and Normalized Discounted Cumulative Gain with $P = 1000$ (nDCG). + +works compared to others. + +Random or real input. As our method is computed using generated random data (white noise) it does not rely on any real dataset and is therefore applicable also in situations where no reliable data are available. Table 5 shows that by relying just on random data we do not lose any performance with respect to all key metrics. Moreover, the performance remains relatively stable across different batch sizes. We chose batch size 64 as larger batch size does not bring better results. + +Due to lack of space, we kindly refer the reader to Supplementary material for further ablations. + +# 6. Conclusion + +We proposed a new training-free NAS proxy called Variance of Knowledge of Deep Network Weights (VKDNW). The method has strong theoretical support and achieved state-of-the-art results on three public datasets and in two search spaces. + +Based on throughout evaluation and comparison to other existing approaches, it has been shown that it provides an information orthogonal to the network size leading to a zero-cost ranking were contribution of the network size and architecture feasibility are separated. We have also shown that previously used correlation metrics for proxy evaluations do not sufficiently assess the key ability to discriminate top networks and to address this problem we have proposed a new evaluation metric and re-evaluated previous methods with the new metric. + +# Acknowledgement + +The research was supported by Czech Science Foundation Grant No. 24-10738M. The access to the computational infrastructure of the OP VVV funded project CZ.02.1.01/0.0/0.0/16_019/0000765 “Research Center for Informatics” and of the Ministry of Education, Youth and Sports of the Czech Republic project e-INFRA CZ (ID:90254) is also gratefully acknowledged. + +# References + +[1] Mohamed S Abdelfattah, Abhinav Mehrotra, Lukasz Dudziak, and Nicholas D Lane. Zero-cost proxies for lightweight NAS. arXiv preprint arXiv:2101.08134, 2021. +[2] Chris Burges, Tal Shaked, Erin Renshaw, Ari Lazier, Matt Deeds, Nicole Hamilton, and Greg Hullender. Learning to rank using gradient descent. In Proceedings of the 22nd international conference on Machine learning, pages 89-96, 2005. +[3] Han Cai, Ligeng Zhu, and Song Han. Proxylessnas: Direct neural architecture search on target task and hardware. arXiv preprint arXiv:1812.00332, 2018. +[4] Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. Once-for-all: Train one network and specialize it for efficient deployment. arXiv preprint arXiv:1908.09791, 2019. +[5] Boyu Chen, Peixia Li, Baopu Li, Chen Lin, Chuming Li, Ming Sun, Junjie Yan, and Wanli Ouyang. BN-NAS: Neural architecture search with batch normalization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 307-316, 2021. +[6] Wuyang Chen, Xinyu Gong, and Zhangyang Wang. Neural architecture search on imagenet in fourgpu hours: A theoretically inspired perspective. arXiv preprint arXiv:2102.11535, 2021. +[7] Patryk Chrabaszcz, Ilya Loshchilov, and Frank Hutter. A downsampled variant of imagenet as an alternative to the cifar datasets. arXiv preprint arXiv:1707.08819, 2017. +[8] Harald Cramér. Mathematical methods of statistics. Princeton university press, 1999. +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. +[10] Xuanyi Dong and Yi Yang. Nas-bench-201: Extending the scope of reproducible neural architecture search. arXiv preprint arXiv:2001.00326, 2020. +[11] Roy Frieden and Robert A Gatenby. Exploratory data analysis using Fisher information. Springer Science & Business Media, 2010. +[12] Boris Hanin and David Rolnick. Complexity of linear regions in deep networks. In International Conference on Machine Learning, pages 2596-2604. PMLR, 2019. +[13] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018. + +[14] Gabriela Kadlecova, Jovita Lukasik, Martin Pilát, Petra Vidnerova, Mahmoud Safari, Roman Neruda, and Frank Hutter. Surprisingly strong performance prediction with neural graph features. arXiv preprint arXiv:2404.16551, 2024. +[15] Ryo Karakida, Shotaro Akaho, and Shun-ichi Amari. Pathological spectra of the fisher information metric and its variants in deep neural networks. arXiv preprint arXiv:1910.05992, 2019. +[16] Ryo Karakida, Shotaro Akaho, and Shun-ichi Amari. Universal statistics of fisher information in deep neural networks: Mean field approach. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 1032–1041. PMLR, 2019. +[17] Maurice G Kendall. A new measure of rank correlation. Biometrika, 30(1-2):81-93, 1938. +[18] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. +[19] Frederik Kunstner, Philipp Hennig, and Lukas Balles. Limitations of the empirical fisher approximation for natural gradient descent. Advances in neural information processing systems, 32, 2019. +[20] Byung-Kwan Lee, Junho Kim, and Yong Man Ro. Masking adversarial damage: Finding adversarial saliency for robust and sparse network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15126-15136, 2022. +[21] Junghyup Lee and Bumsub Ham. AZ-NAS: Assembling zero-cost proxies for network architecture search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5893-5903, 2024. +[22] Jaehoon Lee, Lechao Xiao, Samuel Schoenholz, Yasaman Bahri, Roman Novak, Jascha Sohl-Dickstein, and Jeffrey Pennington. Wide neural networks of any depth evolve as linear models under gradient descent. Advances in neural information processing systems, 32, 2019. +[23] Namhoon Lee, Thalaiyasingam Ajanthan, and Philip HS Torr. Snip: Single-shot network pruning based on connection sensitivity. arXiv preprint arXiv:1810.02340, 2018. +[24] Steven J Leon, Lisette De Pillis, and Lisette G De Pillis. Linear algebra with applications. Pearson Prentice Hall Upper Saddle River, NJ, 2006. +[25] Guihong Li, Yuedong Yang, Kartikeya Bhardwaj, and Radu Marculescu. ZiCo: Zero-shot NAS via inverse coefficient of variation on gradients. In The Eleventh International Conference on Learning Representations, 2023. +[26] Ming Lin, Pichao Wang, Zhenhong Sun, Hesen Chen, Xiuyu Sun, Qi Qian, Hao Li, and Rong Jin. Zen-nas: A zero-shot nas for high-performance image recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 347-356, 2021. +[27] Hanxiao Liu, Karen Simonyan, and Yiming Yang. DARTS: Differentiable architecture search. arXiv preprint arXiv:1806.09055, 2018. +[28] Alexander Ly, Maarten Marsman, Josine Verhagen, Raoul PPP Grasman, and Eric-Jan Wagenmakers. A tutorial on fisher information. Journal of Mathematical Psychology, 80:40-55, 2017. + +[29] James Martens. New insights and perspectives on the natural gradient method. Journal of Machine Learning Research, 21 (146):1-76, 2020. +[30] Abhinav Mehrotra, Alberto Gil CP Ramos, Sourav Bhattacharya, Lukasz Dudziak, Ravichander Vipperla, Thomas Chau, Mohamed S Abdelfattah, Samin Ishtiaq, and Nicholas Donald Lane. Nas-bench-asr: Reproducible neural architecture search for speech recognition. In International Conference on Learning Representations, 2021. +[31] Joe Mellor, Jack Turner, Amos Storkey, and Elliot J Crowley. Neural architecture search without training. In International conference on machine learning, pages 7588-7598. PMLR, 2021. +[32] Bert Moons, Parham Noorzad, Andrii Skliar, Giovanni Mariani, Dushyant Mehta, Chris Lott, and Tijmen Blankevoort. Distilling optimal neural networks: Rapid search in diverse spaces. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12229-12238, 2021. +[33] Xuefei Ning, Changcheng Tang, Wenshuo Li, Zixuan Zhou, Shuang Liang, Huazhong Yang, and Yu Wang. Evaluating efficient performance estimators of neural architectures. Advances in Neural Information Processing Systems, 34: 12265-12277, 2021. +[34] Roman Novak, Jascha Sohl-Dickstein, and Samuel S Schoenholz. Fast finite width neural tangent kernel. In International Conference on Machine Learning, pages 17018-17044. PMLR, 2022. +[35] Hyeyoung Park and Kwanyong Lee. Adaptive natural gradient method for learning of stochastic neural networks in mini-batch mode. Applied Sciences, 9(21):4568, 2019. +[36] Jeffrey Pennington and Pratik Worah. The spectrum of the fisher information matrix of a single-hidden-layer neural network. Advances in neural information processing systems, 31, 2018. +[37] C Radhakrishna Rao. Information and the accuracy attainable in the estimation of statistical parameters. In Breakthroughs in Statistics: Foundations and basic theory, pages 235-247. Springer, 1992. +[38] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *Proceedings of the IEEE conference on computer vision and pattern recognition*, pages 4510-4520, 2018. +[39] Charles Spearman. The proof and measurement of association between two things. 1961. +[40] Kunio Tanabe and Masahiko Sagae. An exact cholesky decomposition and the generalized inverse of the variance-covariance matrix of the multinomial distribution, with applications. Journal of the Royal Statistical Society: Series B (Methodological), 54(1):211-219, 1992. +[41] Hidenori Tanaka, Daniel Kunin, Daniel L Yamins, and Surya Ganguli. Pruning neural networks without any data by iteratively conserving synaptic flow. Advances in neural information processing systems, 33:6377-6389, 2020. +[42] Chaoqi Wang, Guodong Zhang, and Roger Grosse. Picking winning tickets before training by preserving gradient flow. arXiv preprint arXiv:2002.07376, 2020. + +[43] Colin White, Mikhail Khodak, Renbo Tu, Shital Shah, Sébastien Bubeck, and Debadeepta Dey. A deeper look at zero-cost proxies for lightweight NAS. *ICLR Blog Track*, 2022. +[44] Huan Xiong, Lei Huang, Mengyang Yu, Li Liu, Fan Zhu, and Ling Shao. On the number of linear regions of convolutional neural networks. In International Conference on Machine Learning, pages 10514-10523. PMLR, 2020. +[45] Zhaohui Yang, Yunhe Wang, Xinghao Chen, Boxin Shi, Chao Xu, Chunjing Xu, Qi Tian, and Chang Xu. CARS: Continuous evolution for efficient neural architecture search. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1829-1838, 2020. +[46] Chris Ying, Aaron Klein, Eric Christiansen, Esteban Real, Kevin Murphy, and Frank Hutter. Nas-bench-101: Towards reproducible neural architecture search. In International conference on machine learning, pages 7105-7114. PMLR, 2019. +[47] Wang Yining, Wang Liwei, Li Yuanzhi, He Di, Chen Wei, and Liu Tie-Yan. A theoretical analysis of ndcg ranking measures. In Proceedings of the 26th annual conference on learning theory, 2013. +[48] Xuanyang Zhang, Pengfei Hou, Xiangyu Zhang, and Jian Sun. Neural architecture search with random labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10907-10916, 2021. +[49] Zhihao Zhang and Zhihao Jia. Gradsign: Model performance inference with theoretical insights. arXiv preprint arXiv:2110.08616, 2021. +[50] Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V Le. Learning transferable architectures for scalable image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8697-8710, 2018. \ No newline at end of file diff --git a/2025/Training-free Neural Architecture Search through Variance of Knowledge of Deep Network Weights/images.zip b/2025/Training-free Neural Architecture Search through Variance of Knowledge of Deep Network Weights/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..df391eaf542663c1435ffcd4c9ec52e2e1dfed37 --- /dev/null +++ b/2025/Training-free Neural Architecture Search through Variance of Knowledge of Deep Network Weights/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9da91c1979a13a568939983a080615aa978148df264c5cecd062d9b0202e5c5c +size 512966 diff --git a/2025/Training-free Neural Architecture Search through Variance of Knowledge of Deep Network Weights/layout.json b/2025/Training-free Neural Architecture Search through Variance of Knowledge of Deep Network Weights/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2837e59c9a2f60bd2acd2b2ff541876321ff7c7d --- /dev/null +++ b/2025/Training-free Neural Architecture Search through Variance of Knowledge of Deep Network Weights/layout.json @@ -0,0 +1,9658 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 102, + 541, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 102, + 541, + 140 + ], + "spans": [ + { + "bbox": [ + 70, + 102, + 541, + 140 + ], + "type": "text", + "content": "Training-free Neural Architecture Search through Variance of Knowledge of Deep Network Weights" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 222, + 161, + 411, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 161, + 411, + 175 + ], + "spans": [ + { + "bbox": [ + 222, + 161, + 411, + 175 + ], + "type": "text", + "content": "Ondrej Tybl Lukas Neumann" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 175, + 570, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 175, + 570, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 175, + 570, + 190 + ], + "type": "text", + "content": "CMP Visual Recognition Group, Faculty of Electrical Engineering, Czech Technical University in Prague" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 192, + 192, + 436, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 192, + 436, + 202 + ], + "spans": [ + { + "bbox": [ + 192, + 192, + 436, + 202 + ], + "type": "text", + "content": "ondrej.tybl@cvut.cz lukas.neumann@cvut.cz" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 255, + 296, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 376 + ], + "type": "text", + "content": "Deep learning has revolutionized computer vision, but it achieved its tremendous success using deep network architectures which are mostly hand-crafted and therefore likely suboptimal. Neural Architecture Search (NAS) aims to bridge this gap by following a well-defined optimization paradigm which systematically looks for the best architecture, given objective criterion such as maximal classification accuracy. The main limitation of NAS is however its astronomical computational cost, as it typically requires training each candidate network architecture from scratch." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 376, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 296, + 471 + ], + "type": "text", + "content": "In this paper, we aim to alleviate this limitation by proposing a novel training-free proxy for image classification accuracy based on Fisher Information. The proposed proxy has a strong theoretical background in statistics and it allows estimating expected image classification accuracy of a given deep network without training the network, thus significantly reducing computational cost of standard NAS algorithms." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 472, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 555 + ], + "type": "text", + "content": "Our training-free proxy achieves state-of-the-art results on three public datasets and in two search spaces, both when evaluated using previously proposed metrics, as well as using a new metric that we propose which we demonstrate is more informative for practical NAS applications. The source code is publicly available at https://www.github.com/ondratybl/VKDNW." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 574, + 135, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 135, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 135, + 586 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": "In most instances, neural network architectures are designed by authors following the field's \"best-practices\" or their experience, without any formal and repeatable procedure. This is however inconvenient especially in applications on a large scale. Neural Architecture Search (NAS) aims to bridge this gap by following a well-defined optimization paradigm which systematically looks for the best architecture, given objective criterion such as maximal accuracy." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "The main limitation of Neural Architecture Search (NAS) is however the computational cost, because in the" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 322, + 232, + 545, + 419 + ], + "blocks": [ + { + "bbox": [ + 322, + 232, + 545, + 419 + ], + "lines": [ + { + "bbox": [ + 322, + 232, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 322, + 232, + 545, + 419 + ], + "type": "image", + "image_path": "97fd63a8b192a1c196b654c0292adafa60c228dc90dd62d9cefa4dc0ad195af9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 426, + 555, + 494 + ], + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 494 + ], + "type": "text", + "content": "Figure 1. Training-free NAS methods on ImageNet16-120 [10]. Methods are compared by Normalized Discounted Cumulative Gain (see Sec. 3), our method (VKDNW) is the best also measured by Kendall's " + }, + { + "bbox": [ + 313, + 426, + 555, + 494 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 426, + 555, + 494 + ], + "type": "text", + "content": " and Spearman's " + }, + { + "bbox": [ + 313, + 426, + 555, + 494 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 426, + 555, + 494 + ], + "type": "text", + "content": " correlations (see Table 1). Also note that simple number of trainable layers (below denoted " + }, + { + "bbox": [ + 313, + 426, + 555, + 494 + ], + "type": "inline_equation", + "content": "\\aleph" + }, + { + "bbox": [ + 313, + 426, + 555, + 494 + ], + "type": "text", + "content": ") is significantly better trivial proxy than the number of FLOPs." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 510, + 555, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 593 + ], + "type": "text", + "content": "most basic NAS setup, it is required to train thousands or more of different deep network architectures from scratch in order just to calculate a single scalar – the objective function value, such as the classification accuracy. This severely limits practical applications of NAS as the size of feasible architecture search space is only a small fraction of the overall space of all networks." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Training-free Neural Architecture Search (TF-NAS) aims to alleviate this limitation by introducing an objective function proxy which – unlike the actual objective function – does not require training the network. As a result, a good proxy allows the TF-NAS algorithm to explore significantly bigger portions of the network architecture search space compared to traditional NAS, and to find the best network architecture without training a single network. The crucial question is however finding an appropriate objective function proxy." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14881" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 296, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 296, + 215 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 296, + 215 + ], + "type": "text", + "content": "In this paper, we present a novel, principled objective function proxy called Variance of Knowledge of Deep Network Weights (VKDNW) for image classification accuracy, which allows us to find optimal network architectures for image classification without training them first. Our method is, to the best of our knowledge, the first successful application of Fisher Information theory [28] in the context of large deep neural networks, and as such allows us to formally describe and quantify the difficulty of network parameters' estimation process. In other words, given a network architecture, our method estimates how easy or hard it will be to train the network." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 217, + 296, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 296, + 348 + ], + "type": "text", + "content": "Additionally, we also observe that the evaluation metrics used in the TF-NAS community are not well-suited to the problem at hand, because it unnecessarily penalizes for bad proxy accuracy for networks which are not interesting, and vice-versa it does not sufficiently reward proxies which are able to accurately pick out good network architectures. Following this observation, we propose that the Normalized Discounted Cumulative Gain should be used in comparison with other TF-NAS metrics, and show indeed that there are significant differences amongst previously proposed TF-NAS methods when the new metric is considered." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 350, + 281, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 350, + 281, + 362 + ], + "spans": [ + { + "bbox": [ + 67, + 350, + 281, + 362 + ], + "type": "text", + "content": "To summarize, we make the following contributions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 364, + 296, + 555 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 364, + 295, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 364, + 295, + 424 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 295, + 424 + ], + "type": "text", + "content": "1. We introduce a novel algorithm for estimation of Fisher Information Matrix spectrum, which is tractable even for models with large number of parameters such as deep networks, that overcomes the usual problems of numerical stability." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 424, + 296, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 296, + 508 + ], + "type": "text", + "content": "2. We introduce a novel principled VKDNW proxy for image classification accuracy. The proxy is based on strong theoretical background and captures uncertainty in weight estimation process. It brings information that is orthogonal to the model size which then allows for efficient combination with previously proposed proxies, leading to state-of-the-art results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 508, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 296, + 555 + ], + "type": "text", + "content": "3. We propose a new evaluation metric for TF-NAS proxies which is more relevant to the actual NAS objective as it concentrates on ability of given proxy to identify good networks." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 571, + 111, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 111, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 111, + 583 + ], + "type": "text", + "content": "2. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "Our zero-shot proxy for image classification accuracy builds on Fisher Information theory [28], therefore we begin by a thorough analysis of Fisher Information Matrix (FIM) of the network weights estimation problem (see Sec. 2.1). We proceed by discussing challenges involved in practical application of FIM in context of large over-parametrized models that lead to only limited success in previous works and present our contributions to overcome these limitations (see Sec. 2.2). Finally, we propose a novel FIM-based proxy for NAS algorithms (Sec. 2.3)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 72, + 426, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 426, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 426, + 83 + ], + "type": "text", + "content": "2.1. Fisher Information" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 91, + 555, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 137 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 137 + ], + "type": "text", + "content": "The problem of finding the optimal weights of a neural network " + }, + { + "bbox": [ + 313, + 91, + 555, + 137 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 91, + 555, + 137 + ], + "type": "text", + "content": " for the task of " + }, + { + "bbox": [ + 313, + 91, + 555, + 137 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 91, + 555, + 137 + ], + "type": "text", + "content": "-class image classification can be seen as a maximum likelihood estimation with a statistical model" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 324, + 140, + 555, + 170 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 140, + 555, + 170 + ], + "spans": [ + { + "bbox": [ + 324, + 140, + 555, + 170 + ], + "type": "interline_equation", + "content": "\\sigma_ {\\theta} (c \\mid x) = \\frac {\\exp (\\Psi_ {c} (x , \\theta))}{\\sum_ {d = 1} ^ {C} \\exp (\\Psi_ {d} (x , \\theta))}, \\quad c = 1, \\dots , C \\tag {1}", + "image_path": "6d6210cf7f02402198a27c293ab1b323b1be58712a2734f0b54a3effec50c118.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 174, + 555, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 174, + 555, + 305 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 555, + 305 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 174, + 555, + 305 + ], + "type": "inline_equation", + "content": "\\sigma_{\\theta}(\\cdot |x)" + }, + { + "bbox": [ + 313, + 174, + 555, + 305 + ], + "type": "text", + "content": " denotes the a posteriori distribution of the labels given input image " + }, + { + "bbox": [ + 313, + 174, + 555, + 305 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 174, + 555, + 305 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 174, + 555, + 305 + ], + "type": "inline_equation", + "content": "\\Psi (x,\\theta)\\in \\mathbb{R}^{C} = (\\Psi_{1}(x,\\theta),\\dots ,\\Psi_{C}(x,\\theta))" + }, + { + "bbox": [ + 313, + 174, + 555, + 305 + ], + "type": "text", + "content": " is the network output (logits) given the weight vector " + }, + { + "bbox": [ + 313, + 174, + 555, + 305 + ], + "type": "inline_equation", + "content": "\\theta \\in \\mathbb{R}^p" + }, + { + "bbox": [ + 313, + 174, + 555, + 305 + ], + "type": "text", + "content": ", i.e. the network weights. We describe the process of training as finding the optimal weight vector " + }, + { + "bbox": [ + 313, + 174, + 555, + 305 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 313, + 174, + 555, + 305 + ], + "type": "text", + "content": " that fits our data and we posit that network architectures should be characterised by how easy it is to estimate their optimal network weights " + }, + { + "bbox": [ + 313, + 174, + 555, + 305 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 313, + 174, + 555, + 305 + ], + "type": "text", + "content": ". We build upon statistical learning theory and use Fisher Information [28] framework to formally describe expected behaviour of the training process of a given deep network " + }, + { + "bbox": [ + 313, + 174, + 555, + 305 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 174, + 555, + 305 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 306, + 554, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 554, + 377 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 554, + 377 + ], + "type": "text", + "content": "The Fisher Information Matrix (FIM) encompasses information on the difficulty of the parameter estimation problem and it plays a crucial role in several fundamental results, which we apply below in the context of deep networks. The FIM of a network " + }, + { + "bbox": [ + 313, + 306, + 554, + 377 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 306, + 554, + 377 + ], + "type": "text", + "content": " and its set of weights " + }, + { + "bbox": [ + 313, + 306, + 554, + 377 + ], + "type": "inline_equation", + "content": "\\theta \\in \\mathbb{R}^p" + }, + { + "bbox": [ + 313, + 306, + 554, + 377 + ], + "type": "text", + "content": " is given as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 331, + 382, + 553, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 382, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 331, + 382, + 553, + 397 + ], + "type": "interline_equation", + "content": "F (\\theta) := \\mathbb {E} \\left[ \\nabla_ {\\theta} \\sigma_ {\\theta} (c | x) \\nabla_ {\\theta} \\sigma_ {\\theta} (c | x) ^ {T} \\right] \\in \\mathbb {R} ^ {p \\times p}, \\tag {2}", + "image_path": "6c4dcd68f1e6509a7941facc61c1589fcd23e5b68fa58de955691f08400e47fd.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 401, + 554, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 554, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 554, + 450 + ], + "type": "text", + "content": "where we take the expected value " + }, + { + "bbox": [ + 313, + 401, + 554, + 450 + ], + "type": "inline_equation", + "content": "\\mathbb{E}" + }, + { + "bbox": [ + 313, + 401, + 554, + 450 + ], + "type": "text", + "content": " with respect to the joint distribution of " + }, + { + "bbox": [ + 313, + 401, + 554, + 450 + ], + "type": "inline_equation", + "content": "(x, c)" + }, + { + "bbox": [ + 313, + 401, + 554, + 450 + ], + "type": "text", + "content": ". For more detailed account on Fisher Information theory and its applications in the context of machine learning, we kindly refer reader to [16, 20, 29, 35, 36]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 452, + 555, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 555, + 612 + ], + "type": "text", + "content": "Cramér-Rao bound. The first part of estimation theory we build upon is the inverse of the FIM, known as the Cramér-Rao bound (see [11]), which is the asymptotic variance of the estimated weights (i.e. the uncertainty coming from the data variability). Thus, the larger the matrix norm of the FIM, the more certain we are about the weights " + }, + { + "bbox": [ + 313, + 452, + 555, + 612 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 452, + 555, + 612 + ], + "type": "text", + "content": ". More specifically, any data-dependent estimator " + }, + { + "bbox": [ + 313, + 452, + 555, + 612 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}_n = \\hat{\\theta}_n(x_1,\\ldots ,x_n)" + }, + { + "bbox": [ + 313, + 452, + 555, + 612 + ], + "type": "text", + "content": " is a random vector with randomness coming from the (independent) choice of input images " + }, + { + "bbox": [ + 313, + 452, + 555, + 612 + ], + "type": "inline_equation", + "content": "x_{1},\\ldots ,x_{n}" + }, + { + "bbox": [ + 313, + 452, + 555, + 612 + ], + "type": "text", + "content": " and as such has some variance matrix " + }, + { + "bbox": [ + 313, + 452, + 555, + 612 + ], + "type": "inline_equation", + "content": "\\operatorname{Var}\\left(\\hat{\\theta}_n\\right)" + }, + { + "bbox": [ + 313, + 452, + 555, + 612 + ], + "type": "text", + "content": ". The famous result (see [8, 37]) named in honour of H. Cramér and C. R. Rao states that if " + }, + { + "bbox": [ + 313, + 452, + 555, + 612 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}_n" + }, + { + "bbox": [ + 313, + 452, + 555, + 612 + ], + "type": "text", + "content": " is unbiased then the variance is bounded from below as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 387, + 615, + 553, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 615, + 553, + 638 + ], + "spans": [ + { + "bbox": [ + 387, + 615, + 553, + 638 + ], + "type": "interline_equation", + "content": "\\operatorname {V a r} \\left(\\hat {\\theta} _ {n}\\right) \\geq \\frac {1}{n} F ^ {- 1} (\\theta) \\tag {3}", + "image_path": "b8fdabfa45bfe25b4684085fdaa010c3f2dbd264cd063d60db653c0fc0363e38.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 641, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 677 + ], + "type": "text", + "content": "and for maximum likelihood estimation this bound is attained as the number of input images grows to infinity " + }, + { + "bbox": [ + 313, + 641, + 554, + 677 + ], + "type": "inline_equation", + "content": "n \\to \\infty" + }, + { + "bbox": [ + 313, + 641, + 554, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "We formally show (see Supplementary material) that for each weight " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\theta(j)" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": " the mean square error of our estimation is controlled by the diagonal element of the FIM inverse as" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14882" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 97, + 87, + 295, + 110 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 87, + 295, + 110 + ], + "spans": [ + { + "bbox": [ + 97, + 87, + 295, + 110 + ], + "type": "interline_equation", + "content": "\\mathbb {E} \\left(\\hat {\\theta} _ {n} (j) - \\theta (j)\\right) ^ {2} \\geq \\frac {1}{n} \\left(F ^ {- 1} (\\theta)\\right) _ {j j}. \\tag {4}", + "image_path": "860e6e11b4427774606943578e9ce99542d7abe3fb616a30803b4abb4d163070.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 114, + 296, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 114, + 296, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 114, + 296, + 185 + ], + "type": "text", + "content": "Therefore, knowing the FIM allows us to evaluate how certain we are about the weight estimates. We can go even further by inspecting the eigenvalues of " + }, + { + "bbox": [ + 55, + 114, + 296, + 185 + ], + "type": "inline_equation", + "content": "F(\\theta)" + }, + { + "bbox": [ + 55, + 114, + 296, + 185 + ], + "type": "text", + "content": ". Denoting the largest and smallest eigenvalue of the FIM as " + }, + { + "bbox": [ + 55, + 114, + 296, + 185 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{min}}" + }, + { + "bbox": [ + 55, + 114, + 296, + 185 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 114, + 296, + 185 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{max}}" + }, + { + "bbox": [ + 55, + 114, + 296, + 185 + ], + "type": "text", + "content": " respectively, then we have that there exist linear combination coefficients " + }, + { + "bbox": [ + 55, + 114, + 296, + 185 + ], + "type": "inline_equation", + "content": "e_{\\mathrm{min}}" + }, + { + "bbox": [ + 55, + 114, + 296, + 185 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 114, + 296, + 185 + ], + "type": "inline_equation", + "content": "e_{\\mathrm{max}}" + }, + { + "bbox": [ + 55, + 114, + 296, + 185 + ], + "type": "text", + "content": " of unit size so that" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 110, + 189, + 242, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 189, + 242, + 213 + ], + "spans": [ + { + "bbox": [ + 110, + 189, + 242, + 213 + ], + "type": "interline_equation", + "content": "\\mathbb {E} \\left(e _ {\\min} ^ {T} \\hat {\\theta} _ {n} - e _ {\\min} ^ {T} \\theta\\right) ^ {2} \\geq \\frac {1}{n \\lambda_ {\\min}},", + "image_path": "1fca86646aa4028fa5b24b05b0171eb33c2d882e6fd0a1169a28317d09f37d37.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 215, + 294, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 215, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 107, + 215, + 294, + 240 + ], + "type": "interline_equation", + "content": "\\mathbb {E} \\left(e _ {\\max } ^ {T} \\hat {\\theta} _ {n} - e _ {\\max } ^ {T} \\theta\\right) ^ {2} \\geq \\frac {1}{n \\lambda_ {\\max }}, \\tag {5}", + "image_path": "d76863098daa7926e0d48eb71000f0faf79cefe2c44a7be010ecdb5b829a615b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 243, + 296, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 296, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 296, + 315 + ], + "type": "text", + "content": "indicating that if the difference between the largest and smallest eigenvalue is large then there exist combinations of weights with very large difference in the estimation certainty. Altogether, the more the eigenvalues of the FIM are similar, the more similar is also the variance in the weight estimation across all model weights." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 318, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 296, + 354 + ], + "type": "text", + "content": "Prediction sensibility. The change of the model prediction measured by the KL-divergence when subject to a small perturbations of the weights is given as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 85, + 357, + 295, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 357, + 295, + 381 + ], + "spans": [ + { + "bbox": [ + 85, + 357, + 295, + 381 + ], + "type": "interline_equation", + "content": "D _ {K L} \\left(\\sigma_ {\\theta + \\theta_ {\\delta}} (\\cdot | x), \\sigma_ {\\theta} (\\cdot | x)\\right) \\approx \\frac {1}{2} \\theta_ {\\delta} ^ {T} F (\\theta) \\theta_ {\\delta} \\tag {6}", + "image_path": "171bb0d5129ab2979d29a184b9bec533cb6931f87c686d459280134cc7f10265.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 385, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 296, + 456 + ], + "type": "text", + "content": "for a small weight perturbation vector " + }, + { + "bbox": [ + 55, + 385, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\theta_{\\delta} \\in \\mathbb{R}^{p}" + }, + { + "bbox": [ + 55, + 385, + 296, + 456 + ], + "type": "text", + "content": ". Thus, " + }, + { + "bbox": [ + 55, + 385, + 296, + 456 + ], + "type": "inline_equation", + "content": "F(\\theta)" + }, + { + "bbox": [ + 55, + 385, + 296, + 456 + ], + "type": "text", + "content": " measures volatility of the predictions subject to the weight change – if the difference between the largest and the smallest eigenvalues of the FIM is large, then there exist perturbation directions " + }, + { + "bbox": [ + 55, + 385, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\theta_{min}, \\theta_{max}" + }, + { + "bbox": [ + 55, + 385, + 296, + 456 + ], + "type": "text", + "content": " corresponding to " + }, + { + "bbox": [ + 55, + 385, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\lambda_{min}, \\lambda_{max}" + }, + { + "bbox": [ + 55, + 385, + 296, + 456 + ], + "type": "text", + "content": " such that" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 460, + 302, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 460, + 302, + 529 + ], + "spans": [ + { + "bbox": [ + 58, + 460, + 302, + 529 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} D _ {K L} \\left(\\sigma_ {\\theta + \\theta_ {m i n}} (\\cdot | x), \\sigma_ {\\theta} (\\cdot | x)\\right) \\approx \\frac {1}{2} \\theta_ {m i n} ^ {T} F (\\theta) \\theta_ {m i n} = \\\\ = \\frac {1}{2} \\lambda_ {\\min } \\| \\theta_ {\\min } \\| ^ {2} \\ll \\frac {1}{2} \\lambda_ {\\max } \\| \\theta_ {\\max } \\| ^ {2} = \\tag {7} \\\\ = \\frac {1}{2} \\theta_ {m a x} ^ {T} F (\\theta) \\theta_ {m a x} \\approx D _ {K L} \\left(\\sigma_ {\\theta + \\theta_ {m a x}} (\\cdot | x), \\sigma_ {\\theta} (\\cdot | x)\\right) \\\\ \\end{array}", + "image_path": "ceb5987ae95ee72a4d85b7e921acbb3da3f2905cab58f7a978097d2782f9dc0a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 533, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 296, + 582 + ], + "type": "text", + "content": "and therefore in some directions a small change of the weights has much larger impact on the prediction than in others, making the model less balanced. For further discussion, see the Supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 591, + 296, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 296, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 296, + 615 + ], + "type": "text", + "content": "2.2. Empirical Fisher Information Matrix implementation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 623, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 296, + 647 + ], + "type": "text", + "content": "When independent and identically distributed sample images " + }, + { + "bbox": [ + 55, + 623, + 296, + 647 + ], + "type": "inline_equation", + "content": "x_{n}" + }, + { + "bbox": [ + 55, + 623, + 296, + 647 + ], + "type": "text", + "content": " are available, empirical FIM " + }, + { + "bbox": [ + 55, + 623, + 296, + 647 + ], + "type": "inline_equation", + "content": "\\hat{F} (\\theta)" + }, + { + "bbox": [ + 55, + 623, + 296, + 647 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 69, + 653, + 295, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 653, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 69, + 653, + 295, + 685 + ], + "type": "interline_equation", + "content": "\\hat {F} (\\theta) := \\frac {1}{n} \\sum_ {n = 1} ^ {N} \\mathbb {E} _ {\\sigma_ {\\theta}} \\left[ \\nabla_ {\\theta} \\sigma_ {\\theta} (c | x _ {n}) \\nabla_ {\\theta} \\sigma_ {\\theta} (c | x _ {n}) ^ {T} \\right] \\tag {8}", + "image_path": "29c8a90e98d2b762e138f97b3b4bd01312dfa365d76a37910d9ebb5861809ca2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\sigma_{\\theta}}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " now denotes the expectation with respect to the model prediction " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\sigma_{\\theta}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "We would like to first emphasize several crucial aspects and contributions of this paper that lead to the first success of the Fisher Information theory (having e.g. [1] in mind) in the context of Neural Architecture Search, despite Fisher Information being one of the first obvious choices for deep network analysis:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 326, + 147, + 537, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 147, + 537, + 159 + ], + "spans": [ + { + "bbox": [ + 326, + 147, + 537, + 159 + ], + "type": "text", + "content": "1. Following [19] we write the empirical FIM (8) as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 324, + 164, + 553, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 164, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 324, + 164, + 553, + 213 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {F} (\\theta) = \\frac {1}{n} \\sum_ {n = 1} ^ {N} \\left[ \\nabla_ {\\theta} \\Psi (x _ {n}, \\theta) ^ {T} \\left(\\operatorname {d i a g} \\left(\\sigma_ {\\theta} (\\cdot , x _ {n})\\right) - \\right. \\right. \\\\ \\left. \\sigma_ {\\theta} (\\cdot , x _ {n}) \\sigma_ {\\theta} (\\cdot , x _ {n}) ^ {T}\\right) \\nabla_ {\\theta} \\Psi (x _ {n}, \\theta) ]. \\tag {9} \\\\ \\end{array}", + "image_path": "2c25d4ae0ac83133c48d6187c1929ae218cc9d7c4e7d86b594e19eb911bf055c.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 218, + 554, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 554, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 554, + 266 + ], + "type": "text", + "content": "and we further decompose the inner matrix " + }, + { + "bbox": [ + 313, + 218, + 554, + 266 + ], + "type": "inline_equation", + "content": "\\mathrm{diag}(\\sigma_{\\theta}(\\cdot ,x_n)) - \\sigma_{\\theta}(\\cdot ,x_n)\\sigma_{\\theta}(\\cdot ,x_n)^T" + }, + { + "bbox": [ + 313, + 218, + 554, + 266 + ], + "type": "text", + "content": " using analytical formulas from [40] to avoid numerical instability of the computation as we arrive at a feasible representation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 389, + 271, + 553, + 303 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 271, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 389, + 271, + 553, + 303 + ], + "type": "interline_equation", + "content": "\\hat {F} (\\theta) = \\frac {1}{n} \\sum_ {n = 1} ^ {N} A _ {n} ^ {T} A _ {n} \\tag {10}", + "image_path": "e29d25f247304974537bfa95a9322e3f4e50deacfc4ecf4c0f67de1de8886c5b.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 308, + 554, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 554, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 554, + 380 + ], + "type": "text", + "content": "for some matrices " + }, + { + "bbox": [ + 313, + 308, + 554, + 380 + ], + "type": "inline_equation", + "content": "A_{n}\\in \\mathbb{R}^{C\\times p}" + }, + { + "bbox": [ + 313, + 308, + 554, + 380 + ], + "type": "text", + "content": ". We observed that networks typically yield very imbalanced outputs at initialization (i.e. every network prioritizes few classes over the rest) making it extremely important not to exclude the factor " + }, + { + "bbox": [ + 313, + 308, + 554, + 380 + ], + "type": "inline_equation", + "content": "\\mathrm{diag}(\\sigma_{\\theta}(\\cdot ,x_n)) - \\sigma_{\\theta}(\\cdot ,x_n)\\sigma_{\\theta}(\\cdot ,x_n)^T" + }, + { + "bbox": [ + 313, + 308, + 554, + 380 + ], + "type": "text", + "content": ", which is however difficult to compute without underflow/overflow." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 384, + 554, + 602 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 313, + 384, + 554, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 554, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 554, + 504 + ], + "type": "text", + "content": "2. The dimension of FIM is equal to " + }, + { + "bbox": [ + 313, + 384, + 554, + 504 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 384, + 554, + 504 + ], + "type": "text", + "content": " (the number of all trainable parameters) and therefore direct computation of the eigenvalues is numerically intractable and unstable. However, our results show that if a small number of representative parameters is drawn, then computation becomes stable while the discrimination power does not suffer. We used a simple rule where a single parameter from each trainable layer (not including batch normalization) is chosen. Stability with respect to choice of such parameters can be found in Supplementary material." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 507, + 554, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 554, + 602 + ], + "type": "text", + "content": "3. FIM of large networks typically suffers from having pathological spectrum, i.e. there usually exist zero eigenvalue and its multiplicity is large (see [15]) and thus the eigenvalues estimation due to a large condition number is imprecise. However, as we deal with a symmetric positive-semidefinite matrix, the eigenvalues actually coincide with the singular values (see [24]), for which the estimation algorithm performs better." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "Let us also emphasize that there is a common misconception in part of the community as it is often mistakenly assumed that one can simply use the true labels of " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "x_{n}" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": " in (8) in place of " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": ". However, such definition is then meaningless as it does not approximate the FIM in the classical Monte Carlo sense (see Supplementary material and [19] for a comparison). That means that the empirical FIM does not depend on the true labels and therefore our method does not require real data - we use random input instead." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14883" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 295, + 96 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 295, + 96 + ], + "type": "text", + "content": "2.3. Variance of Knowledge for Deep Network Weights" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 103, + 297, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 103, + 297, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 103, + 297, + 163 + ], + "type": "text", + "content": "In order to characterize properties of the parameter estimation process for a given network " + }, + { + "bbox": [ + 55, + 103, + 297, + 163 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 103, + 297, + 163 + ], + "type": "text", + "content": " through the lens of Fisher Information theory, we inspect the eigenvalues of the empirical FIM " + }, + { + "bbox": [ + 55, + 103, + 297, + 163 + ], + "type": "inline_equation", + "content": "\\hat{F}(\\theta_{\\mathrm{init}})" + }, + { + "bbox": [ + 55, + 103, + 297, + 163 + ], + "type": "text", + "content": " and define the entropy of Variance of Knowledge for Deep Network Weights (VKDNW) as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 78, + 168, + 295, + 232 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 168, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 78, + 168, + 295, + 232 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {V K D N W} (f) := - \\sum_ {k = 1} ^ {9} \\tilde {\\lambda} _ {k} \\log \\tilde {\\lambda} _ {k} \\\\ \\tilde {\\lambda} _ {k} = \\frac {\\lambda_ {k}}{\\sum_ {j = 1} ^ {9} \\lambda_ {j}}, \\quad k = 1, \\dots , 9, \\tag {11} \\\\ \\end{array}", + "image_path": "8882d9ea1d800c1319576b8472cadeadea8f8aaf7389bd6d37f6de6db1b55e11.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 236, + 295, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 295, + 270 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 236, + 295, + 270 + ], + "type": "inline_equation", + "content": "\\lambda_{k}" + }, + { + "bbox": [ + 55, + 236, + 295, + 270 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 55, + 236, + 295, + 270 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 236, + 295, + 270 + ], + "type": "text", + "content": "-th decile of the FIM eigenvalues as the representation of the FIM spectrum1, and " + }, + { + "bbox": [ + 55, + 236, + 295, + 270 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{init}}" + }, + { + "bbox": [ + 55, + 236, + 295, + 270 + ], + "type": "text", + "content": " denotes network weights at initialization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 272, + 296, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 272, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 54, + 272, + 296, + 452 + ], + "type": "text", + "content": "Our score therefore measures the diversity of the FIM eigenvalues, and from the entropy theory we know that VKDNW attains its maximum exactly when all the eigenvalues " + }, + { + "bbox": [ + 54, + 272, + 296, + 452 + ], + "type": "inline_equation", + "content": "\\lambda_{k}" + }, + { + "bbox": [ + 54, + 272, + 296, + 452 + ], + "type": "text", + "content": " are equal and VKDNW gets lower as the eigenvalues become more different. Based on the discussion in Sec. 2.1 we see that VKDNW is high when the uncertainty in all model weight combinations are similar (see Cramér-Rao bound) and there are no directions in the weight space that would influence the network prediction substantially differently than others. Due to the fact that we have normalized both the number of eigenvalues under consideration (by taking a fixed number of representatives irrespective of the network size) and the magnitude of the eigenvalues, VKDNW is independent of network size (number of network weights " + }, + { + "bbox": [ + 54, + 272, + 296, + 452 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 54, + 272, + 296, + 452 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 453, + 296, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 296, + 548 + ], + "type": "text", + "content": "Let us note, that we are familiar with the fact that even though our motivation was (among others) based on Cramér-Rao bound that assumes evaluation of FIM at the correct weight vector " + }, + { + "bbox": [ + 55, + 453, + 296, + 548 + ], + "type": "inline_equation", + "content": "\\theta_{X}" + }, + { + "bbox": [ + 55, + 453, + 296, + 548 + ], + "type": "text", + "content": " that fits the data, which is typically far away from the weights given at initialization. However, our empirical results below support the hypothesis that the evaluation despite being in the wrong point brings valuable information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 552, + 296, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 296, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 296, + 672 + ], + "type": "text", + "content": "Ranking networks for NAS. The proposed VKDNW score is independent of network size, which is extremely beneficial to compare individual structures of network architectures. Thus, it does not aim to capture capacity, rather it targets feasibility of the computation graph given number of operations. However, when comparing different network structures and different network sizes together as it is done in NAS, one indeed needs to take network size into account as well, because naturally larger networks have bigger capacity and therefore tend to have higher accuracy. To" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 319, + 74, + 550, + 192 + ], + "blocks": [ + { + "bbox": [ + 319, + 74, + 550, + 192 + ], + "lines": [ + { + "bbox": [ + 319, + 74, + 550, + 192 + ], + "spans": [ + { + "bbox": [ + 319, + 74, + 550, + 192 + ], + "type": "image", + "image_path": "d5f967d8c0886a78528adbb6ab92ca4e59bafe67932a202795c93e69628d8d76.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 198, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 243 + ], + "type": "text", + "content": "Figure 2. Toy example of two rankings on 10 networks. We plot accuracies ordered by the rankings and evaluation metrics Kendall's " + }, + { + "bbox": [ + 313, + 198, + 555, + 243 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 198, + 555, + 243 + ], + "type": "text", + "content": " (KT) and Spearman's " + }, + { + "bbox": [ + 313, + 198, + 555, + 243 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 198, + 555, + 243 + ], + "type": "text", + "content": " (SPR) correlations and Normalized Discounted Cumulative Gain " + }, + { + "bbox": [ + 313, + 198, + 555, + 243 + ], + "type": "inline_equation", + "content": "(\\mathrm{nDCG}_5)" + }, + { + "bbox": [ + 313, + 198, + 555, + 243 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 251, + 555, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 287 + ], + "type": "text", + "content": "capture also the capacity we proxy the network size by the number of layers with weights that we denote as " + }, + { + "bbox": [ + 313, + 251, + 555, + 287 + ], + "type": "inline_equation", + "content": "\\aleph(f)" + }, + { + "bbox": [ + 313, + 251, + 555, + 287 + ], + "type": "text", + "content": " for a network " + }, + { + "bbox": [ + 313, + 251, + 555, + 287 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 251, + 555, + 287 + ], + "type": "text", + "content": " and we introduce the ranking" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 337, + 292, + 553, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 292, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 337, + 292, + 553, + 306 + ], + "type": "interline_equation", + "content": "\\operatorname {V K D N W} _ {\\text {s i n g l e}} (f) := \\aleph (f) + \\operatorname {V K D N W} (f) \\tag {12}", + "image_path": "f687c494a9e543c91c2aef54ca7149019d0e8de3ccae58883cc326fc2994eaeb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 311, + 555, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 555, + 382 + ], + "type": "text", + "content": "Here we leverage the fact that VKDNW as an entropy of some quantity is always between 0 and 1 and by summing it with an integer-valued quantity we in fact obtain that we have first grouped networks by our size proxy " + }, + { + "bbox": [ + 313, + 311, + 555, + 382 + ], + "type": "inline_equation", + "content": "\\aleph" + }, + { + "bbox": [ + 313, + 311, + 555, + 382 + ], + "type": "text", + "content": " and then within each group of similar networks sizes we order them by VKDNW." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 396, + 427, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 396, + 427, + 407 + ], + "spans": [ + { + "bbox": [ + 314, + 396, + 427, + 407 + ], + "type": "text", + "content": "3. Evaluation Metrics" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 416, + 554, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 554, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 554, + 500 + ], + "type": "text", + "content": "In this section, we dive into the problem of evaluating (image) classification accuracy proxies for training-free NAS (TF-NAS). In NAS algorithms, the proxy actually does not have to predict the classification accuracy in absolute terms – since we are after finding the \"best\" network for given task – it's sufficient that the proxy properly ranks the networks, ordering them from worst to the best." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 501, + 554, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 554, + 584 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 554, + 584 + ], + "type": "text", + "content": "Suppose that a collection of " + }, + { + "bbox": [ + 313, + 501, + 554, + 584 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 501, + 554, + 584 + ], + "type": "text", + "content": " networks with validation accuracies " + }, + { + "bbox": [ + 313, + 501, + 554, + 584 + ], + "type": "inline_equation", + "content": "\\mathrm{acc}_1, \\ldots, \\mathrm{acc}_K" + }, + { + "bbox": [ + 313, + 501, + 554, + 584 + ], + "type": "text", + "content": " have been ranked by a TF-NAS proxy as " + }, + { + "bbox": [ + 313, + 501, + 554, + 584 + ], + "type": "inline_equation", + "content": "r_1, \\ldots, r_K \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 501, + 554, + 584 + ], + "type": "text", + "content": ". The standard way of evaluating [14, 21, 25] how well the ranks correspond to the accuracies is to compute Kendall's " + }, + { + "bbox": [ + 313, + 501, + 554, + 584 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 501, + 554, + 584 + ], + "type": "text", + "content": " (KT, [17]) and Spearman's " + }, + { + "bbox": [ + 313, + 501, + 554, + 584 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 501, + 554, + 584 + ], + "type": "text", + "content": " (SPR, [39]) rank correlation coefficients. Kendall's " + }, + { + "bbox": [ + 313, + 501, + 554, + 584 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 501, + 554, + 584 + ], + "type": "text", + "content": " is given as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 355, + 590, + 555, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 590, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 355, + 590, + 555, + 613 + ], + "type": "interline_equation", + "content": "\\tau := \\frac {n _ {c} - n _ {d}}{\\sqrt {n _ {c} + n _ {d} + n _ {1}} \\sqrt {n _ {c} + n _ {d} + n _ {2}}}, \\tag {13}", + "image_path": "88acd6fc0e8c952b08948fa4edcde0eee1d07c92030f1ce6281ff030b26960d9.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "inline_equation", + "content": "n_c" + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": " denotes the number of concordant pairs " + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "inline_equation", + "content": "(\\mathrm{acc}_k, r_k)" + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "inline_equation", + "content": "n_d" + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": " the number of discordant pairs and " + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "inline_equation", + "content": "n_1" + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": " (resp. " + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "inline_equation", + "content": "n_2" + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": ") denote the number of ties in " + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{acc}_k" + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": " only (resp. in " + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "inline_equation", + "content": "r_k" + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": " only). In our context, (KT) relates after some normalization to a probability that two randomly chosen network rankings " + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "inline_equation", + "content": "r_k, r_l" + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": " are ordered correctly according to their accuracies " + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{acc}_k, \\mathrm{acc}_l" + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": ". In contrast, (SPR) is defined to account more for outliers, i.e. it tends to put a large penalization if" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": "1As explained in Sec. 2.2, the smallest eigenvalue " + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_0" + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": " is usually equal to zero and thus we exclude it for stability reasons (similarly with the maximal eigenvalue " + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{10}" + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": ")" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14884" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 120 + ], + "type": "text", + "content": "there are some networks for which the difference between the rank of the accuracy " + }, + { + "bbox": [ + 54, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "\\mathrm{acc}_k" + }, + { + "bbox": [ + 54, + 72, + 294, + 120 + ], + "type": "text", + "content": " and the rank " + }, + { + "bbox": [ + 54, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "r_k" + }, + { + "bbox": [ + 54, + 72, + 294, + 120 + ], + "type": "text", + "content": " is large as it is just the classical correlation coefficient however applied on the orders of the assessed quantities." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 122, + 297, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 122, + 297, + 348 + ], + "spans": [ + { + "bbox": [ + 57, + 122, + 297, + 348 + ], + "type": "text", + "content": "Normalized Discounted Cumulative Gain. Both (KT) and (SPR) are good evaluation metrics when our interest lies in comparison of ranking proxies when all networks involved are of the same importance. However, in NAS we are primarily interested if a ranking helps us to pick best networks from a given collection and less interested how well the ranking compares two networks with low accuracy. Finding inspiration in information retrieval where one measures the quality of a system that retrieves resources relevant to an input query, we propose to use Normalized Discounted Cumulative Gain (nDCG) [2] as a more relevant metric to measure quality of TF-NAS proxies. The nDCG metric is defined with a key requirement that highly relevant documents are more valuable when they appear earlier in the search engine results (i.e., in higher-ranking positions). This requirement can be reformulated for the task of neural architecture search as networks with high accuracies are more valuable when they appear on higher-ranking positions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 349, + 295, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 295, + 386 + ], + "type": "text", + "content": "We thus define the metric as follows: first we order the networks by their ranking " + }, + { + "bbox": [ + 55, + 349, + 295, + 386 + ], + "type": "inline_equation", + "content": "r_1, \\ldots, r_K" + }, + { + "bbox": [ + 55, + 349, + 295, + 386 + ], + "type": "text", + "content": " so that " + }, + { + "bbox": [ + 55, + 349, + 295, + 386 + ], + "type": "inline_equation", + "content": "r_{k_1} = K, r_{k_2} = K - 1, \\ldots" + }, + { + "bbox": [ + 55, + 349, + 295, + 386 + ], + "type": "text", + "content": " and we compute" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 391, + 295, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 391, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 99, + 391, + 295, + 429 + ], + "type": "interline_equation", + "content": "\\mathrm {n D C G} _ {P} := \\frac {1}{Z} \\left(\\sum_ {j = 1} ^ {P} \\frac {2 ^ {\\mathrm {a c c} _ {k _ {j}}} - 1}{\\log_ {2} (1 + j)}\\right), \\tag {14}", + "image_path": "81a74552a90685940dcfd6f338a566d62741bf2d84e86031c1a80201e1b01bd5.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 432, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 295, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 432, + 295, + 491 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{N}" + }, + { + "bbox": [ + 55, + 432, + 295, + 491 + ], + "type": "text", + "content": " is a parameter determining how many top-ranked networks do we consider (e.g. it corresponds to the population size in an evolution algorithm) and " + }, + { + "bbox": [ + 55, + 432, + 295, + 491 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 55, + 432, + 295, + 491 + ], + "type": "text", + "content": " is a normalization factor that represents the ideal discounted cumulative gain so that nDCG is equal to one for a perfect fit." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 492, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 295, + 552 + ], + "type": "text", + "content": "The higher nDCG the better the ranking is as it is a weighted average of transformed top-ranked accuracies. In [47] it is shown that the choice of the discount factor of given by inverse logarithm is a good choice as it nDCG then well separates different ranking systems." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 554, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 295, + 685 + ], + "type": "text", + "content": "Toy Example. Let us briefly demonstrate the weak ability of (KT) and (SPR) to distinguish rankings that poorly discriminate networks with high accuracy. Suppose a collection of 10 networks with their validation accuracies " + }, + { + "bbox": [ + 55, + 554, + 295, + 685 + ], + "type": "inline_equation", + "content": "100,90,\\dots ,10" + }, + { + "bbox": [ + 55, + 554, + 295, + 685 + ], + "type": "text", + "content": " is given and our aim is to rank them from worst to the best. We evaluate two different toy rankings: a) ranking damaged_top which perfectly fits the accuracies, only the two top networks are swapped with the third and fourth, b) ranking perfect_top which fits perfectly for the top-performing networks, however, now the order of the four worst networks is reversed (see Figure 2). The" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "content": "damaged_top ranking is an example that should be evaluated worse for the architecture search task compared to perfect_top as in the first case the best networks are not placed first, while in the second they are. On the other hand, even though perfect_top is not a perfect fit, it still discriminates the top networks ideally and only struggles for networks with low accuracy, which are not interesting for NAS which is after the best networks. From the correlation perspectives of (KT), (SPR) damaged_top is better than perfect_top, therefore, if we rely only on these two evaluation metrics we prefer ranking that is worse for the architecture search task as the top networks are ranked worse. On the other hand, for perfect_top we obtain ideal nDCG" + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "inline_equation", + "content": "_5" + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "content": " while it drops to 0.5 for damaged_top. We conclude that using nDCG" + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "inline_equation", + "content": "_5" + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "content": " we choose ranking that has higher discriminatory power for good networks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 264, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 553, + 361 + ], + "type": "text", + "content": "We also note that from statistical perspective the (KT) and (SPR) of damaged_top is significantly higher than of a random ranking3, and therefore we'd conclude that damaged_top is not independent of ground truth accuracy. On the other hand, when we perform the same statistical test on damaged_top using nDCG5, we conclude that the damaged_top is not significantly better than ranking networks randomly4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 373, + 400, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 400, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 400, + 385 + ], + "type": "text", + "content": "4. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 389, + 553, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 553, + 449 + ], + "type": "text", + "content": "Zero-shot NAS aims to rank given networks in a training-free manner based on their (a priori unknown) final performance, which allows to prune the huge search space with limited costs when seeking for optimal architecture [10, 27, 46] and other configuration [3, 26, 38]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 450, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 554, + 605 + ], + "type": "text", + "content": "Throughout previous works different approaches for ranking computation can be found. In many of the works the gradient with respect to the network weights is investigated (GradNorm, GraSP, SNIP, Synflow, [1, 23, 41, 42]), i.e. leveraging the first order approximation of the network. In Jacov [1] correlations of the Jacobian matrices among various input samples are compared; in NASWOT [31] the linear maps induced by data points are examined; Zen [26] uses the approximation of gradient with respect to featuremaps; GradSign [49] compares the optimization landscape at the level individual training samples; or in ZiCo [25] the gradients from multiple forward and backward passes preferring large magnitude and low variance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 606, + 553, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 553, + 654 + ], + "type": "text", + "content": "Other methods aggregate multiple sources aiming at obtaining a better informed ranks: TE-NAS [6] uses both the number of linear regions [12, 44] and the condition number of Neural Tangent Kernel [13, 22]. However, it is well" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 664, + 553, + 684 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 664, + 553, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 664, + 553, + 684 + ], + "type": "text", + "content": "3That is, the " + }, + { + "bbox": [ + 313, + 664, + 553, + 684 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 664, + 553, + 684 + ], + "type": "text", + "content": "-value for the hypothesis that damaged_top is assigned independently of the accuracies is below 0.005." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 684, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 684, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 684, + 553, + 713 + ], + "type": "text", + "content": "4Running 1000 samples on a random ranking we obtain that damaged_top has nDCG₅ around 75th percentile of such random evaluations and it is not significantly better than a random ranking." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "2In case of ties we take random ordering within the groups (parameter ignore_ties=True in the scikit-learn implementation)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14885" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 69, + 560, + 350 + ], + "blocks": [ + { + "bbox": [ + 55, + 69, + 560, + 350 + ], + "lines": [ + { + "bbox": [ + 55, + 69, + 560, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 69, + 560, + 350 + ], + "type": "table", + "html": "
TypeCIFAR-10CIFAR-100ImageNet16-120
KTSPRnDCGKTSPRnDCGKTSPRnDCG
Simple rankings
FLOPsS0.6230.7990.7450.5860.7630.5760.5450.7180.403
GradNorm [1]S0.3280.4380.5090.3410.4510.2780.3100.4180.265
GraSP [1, 42]S0.3520.5050.5180.3490.4980.2840.3590.5020.281
SNIP [1, 23]S0.4310.5910.5130.4400.5970.2860.3890.5210.286
SynFlow [1, 41]S0.5610.7580.7090.5530.7500.5940.5310.7190.511
Jacov [1]S0.6160.8000.5400.6390.8200.4020.6020.7790.356
NASWOT [31]S0.5710.7620.6070.6070.7990.4750.6050.7940.490
ZenNAS [26]S0.1020.1030.1200.0790.0720.1150.0910.1090.073
GradSign† [49]S·0.765··0.793··0.783·
ZiCo [25]S0.6070.8020.7510.6140.8090.6070.5870.7790.523
TE-NAS [6]A0.5360.7220.6020.5370.7230.3270.5230.7090.330
AZ-NAS [21]A0.7120.8920.7490.6960.8800.5490.6730.8590.534
No. of trainable layers (N)S0.6260.7670.6710.6460.7870.5250.6230.7640.497
VKDNWsingle (ours)S0.6180.8150.7510.6340.8290.6170.6220.8140.608
VKDNWagg (ours)A0.7500.9190.7850.7530.9190.6360.7430.9060.664
Model-driven rankings
GRAF [14]A0.8200.9530.9350.8090.9480.8580.7960.9410.828
VKDNWm (ours)A0.6470.8310.7500.6360.8210.6020.6110.7980.575
(VKDNW+ZCS)m (ours)A0.8400.9630.9220.8340.9600.8840.8300.9580.843
(VKDNW+ZCS+GRAF)m (ours)A0.8590.9710.9460.8470.9660.8950.8420.9630.867
", + "image_path": "4d18da440a9408c83eac99f52639f21bc648ac987fd0daacff6757ded5e0f812.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 357, + 555, + 403 + ], + "lines": [ + { + "bbox": [ + 55, + 357, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 555, + 403 + ], + "type": "text", + "content": "Table 1. Training-free NAS methods in the NAS-Bench-201 [10] search space, evaluated on three public datasets. Kendall's " + }, + { + "bbox": [ + 55, + 357, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 357, + 555, + 403 + ], + "type": "text", + "content": " (KT), Spearman's " + }, + { + "bbox": [ + 55, + 357, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 357, + 555, + 403 + ], + "type": "text", + "content": " (SPR) and Normalized Discounted Cumulative Gain (nDCG) are reported, results are averages of 5 independent runs (all stds are " + }, + { + "bbox": [ + 55, + 357, + 555, + 403 + ], + "type": "inline_equation", + "content": "< 0.02" + }, + { + "bbox": [ + 55, + 357, + 555, + 403 + ], + "type": "text", + "content": "). The Type column differentiates single (S) and aggregated (A) rankings. Results are reproduced with code published by their authors, except those marked†, where results from the original paper are taken." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 408, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 295, + 491 + ], + "type": "text", + "content": "known that the kernel computation is highly computationally demanding [34]. AZ-NAS [21] assesses expressivity, trainability and progressivity via examination of feature distribution across all orientations and the Jacobian. However, despite AZ-NAS outperforming previous works in some metrics, it is worse than [25] in key NAS-related aspects such as the cumulative gain (see Sec. 5.1)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 494, + 295, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 295, + 566 + ], + "type": "text", + "content": "Despite a tremendous effort of the community, it was shown that most of the zero-shot NAS methods perform worse than a simple proxy given just by FLOPs or #params [33, 43]. Thus, there is still a gap for improvements also driven by the need of the ranking explainability that would have a satisfactory theoretical support." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 567, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 295, + 628 + ], + "type": "text", + "content": "Furthermore, [14] uses a model-driven ranking that however needs a set of networks for which the validation accuracies are known to train the model and the ability of this ranking to generalize as it is fitted on a specific network collection only is disputable." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "Finally, let us discuss the current practice in the NAS methods comparison. Methods are compared either by means of correlations of their scores to accuracy or by reporting the accuracy of the top-ranked network [21, 25]. However, while the first is not tailor-made for the architecture search task and therefore does not assess the desired ranking properties (see Sec. 5.1), comparing performance" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 408, + 555, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 555, + 431 + ], + "type": "text", + "content": "just by the accuracy of a single network is too vulnerable e.g. to the random seed choice." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 444, + 397, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 397, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 397, + 457 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 464, + 436, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 436, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 436, + 477 + ], + "type": "text", + "content": "5.1. Ranking aggregation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 483, + 555, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 531 + ], + "type": "text", + "content": "In addition to using the single ranking of Eq. (12), we also experiment with multiple rankings (as in [14, 21, 25]) to order network architectures by their accuracy. We use two different options: non-linear and model-driven aggregation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 533, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 554, + 617 + ], + "type": "text", + "content": "Non-linear aggregation. The aggregation [21] uses multiplication to combine multiple ranks into a single one, which means a network is highly-ranked if and only if it is highly-ranked in all subsidiary rankings, keeping their influence balanced. That is, denoting rankings " + }, + { + "bbox": [ + 313, + 533, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\mathrm{rank}_1,\\ldots ,\\mathrm{rank}_m" + }, + { + "bbox": [ + 313, + 533, + 554, + 617 + ], + "type": "text", + "content": " for some " + }, + { + "bbox": [ + 313, + 533, + 554, + 617 + ], + "type": "inline_equation", + "content": "m\\in \\mathbb{N}" + }, + { + "bbox": [ + 313, + 533, + 554, + 617 + ], + "type": "text", + "content": " we define the aggregated ranking" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 363, + 622, + 553, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 622, + 553, + 637 + ], + "spans": [ + { + "bbox": [ + 363, + 622, + 553, + 637 + ], + "type": "interline_equation", + "content": "\\operatorname {r a n k} _ {\\text {a g g}} (f) := \\log \\Pi_ {j = 1} ^ {m} \\operatorname {r a n k} _ {j} (f) \\tag {15}", + "image_path": "132ff67e87972044a348c96cfa6f56143a9e7b102c7d7a4d5e45a121e798d7e1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 641, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 676 + ], + "type": "text", + "content": "for each network " + }, + { + "bbox": [ + 313, + 641, + 555, + 676 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 641, + 555, + 676 + ], + "type": "text", + "content": ". This aggregation is therefore possible only in the context of a given network collection, such as in evolutionary search." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 326, + 677, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 677, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 326, + 677, + 545, + 689 + ], + "type": "text", + "content": "In our case, " + }, + { + "bbox": [ + 326, + 677, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathrm{VKDNW_{agg}}" + }, + { + "bbox": [ + 326, + 677, + 545, + 689 + ], + "type": "text", + "content": " aggregates these five proxies:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 689, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 314, + 689, + 474, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 689, + 474, + 701 + ], + "spans": [ + { + "bbox": [ + 314, + 689, + 474, + 701 + ], + "type": "text", + "content": "- VKDNWsingle (V) ranking (Eq. (12))," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 314, + 701, + 553, + 713 + ], + "type": "text", + "content": "- Jacov (J) [1] measures the activations correlation when" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14886" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 70, + 298, + 215 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 298, + 215 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 298, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 298, + 215 + ], + "type": "table", + "html": "
MethodFLOPsTop-1 acc.TypeSearch cost (GPU days)
NASNet-B [50]488M72.8MS1800
CARS-D [45]496M73.3MS0.4
BN-NAS [5]470M75.7MS0.8
OFA [4]406M77.7OS50
RLNAS [48]473M75.6OS-
DONNA [32]501M78.0OS405
# Params451M63.5ZS0.02
ZiCo [25]448M78.1ZS0.4
AZ-NAS [21]462M78.6ZS0.4
VKDNWagg(ours)480M78.8ZS0.4
", + "image_path": "0cf8f611ddc7ee1e2f24d6c36dfc6f60993f5686efab1b7dcf90db421be42681.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 223, + 295, + 246 + ], + "lines": [ + { + "bbox": [ + 55, + 223, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 295, + 246 + ], + "type": "text", + "content": "Table 2. Results on ImageNet-1K [9] in the MobileNetV2 search space, the size of the model is constrained to " + }, + { + "bbox": [ + 55, + 223, + 295, + 246 + ], + "type": "inline_equation", + "content": "\\approx 450\\mathrm{M}" + }, + { + "bbox": [ + 55, + 223, + 295, + 246 + ], + "type": "text", + "content": " FLOPS." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 64, + 252, + 171, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 252, + 171, + 263 + ], + "spans": [ + { + "bbox": [ + 64, + 252, + 171, + 263 + ], + "type": "text", + "content": "exposed to various inputs," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 264, + 295, + 336 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 264, + 295, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 295, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 295, + 287 + ], + "type": "text", + "content": "- Expressivity (E) [21] assesses isotropy and uniformity of the features distribution across all orientations," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 288, + 294, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 294, + 323 + ], + "type": "text", + "content": "- Trainability (T) [21] captures ability of the network to keep stable gradient propagation between the layers by inspecting the spectrum of the Jacobian matrix," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 324, + 294, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 294, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 294, + 336 + ], + "type": "text", + "content": "- FLOPs (F) is the number of FLOPs for one forward pass." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 338, + 295, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 295, + 409 + ], + "type": "text", + "content": "Model-driven aggregation. When accuracies of a sufficient number of architectures in the search space are known, model-driven aggregation can be used to train a regression model to combine individual rankings. The trained model is then used to predict accuracy for unseen networks in the same search space. We evaluate three different models:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 410, + 295, + 517 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 55, + 410, + 295, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 295, + 470 + ], + "type": "text", + "content": "- VKDNW" + }, + { + "bbox": [ + 55, + 410, + 295, + 470 + ], + "type": "inline_equation", + "content": "_m" + }, + { + "bbox": [ + 55, + 410, + 295, + 470 + ], + "type": "text", + "content": " where the eigenvalues " + }, + { + "bbox": [ + 55, + 410, + 295, + 470 + ], + "type": "inline_equation", + "content": "\\lambda_k" + }, + { + "bbox": [ + 55, + 410, + 295, + 470 + ], + "type": "text", + "content": " of the FIM matrix are used directly as features in companion with " + }, + { + "bbox": [ + 55, + 410, + 295, + 470 + ], + "type": "inline_equation", + "content": "\\aleph" + }, + { + "bbox": [ + 55, + 410, + 295, + 470 + ], + "type": "text", + "content": " (number of trainable layers) and FLOPs to allow a more complex proxy of the diversity of the eigenvalues and the network complexity than the simple entropy Eq. (11)," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 469, + 295, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 295, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 295, + 493 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 55, + 469, + 295, + 493 + ], + "type": "inline_equation", + "content": "(\\mathrm{VKDNW} + \\mathrm{ZCS})_m" + }, + { + "bbox": [ + 55, + 469, + 295, + 493 + ], + "type": "text", + "content": " where we additionally include all other zero-cost scores available (see Table 1)," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 493, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 295, + 517 + ], + "type": "text", + "content": "- (VKDNW+ZCS+GRAF)" + }, + { + "bbox": [ + 55, + 493, + 295, + 517 + ], + "type": "inline_equation", + "content": "_m" + }, + { + "bbox": [ + 55, + 493, + 295, + 517 + ], + "type": "text", + "content": " where we add network graph features from [14]." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 525, + 113, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 113, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 113, + 537 + ], + "type": "text", + "content": "5.2. Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 543, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 295, + 615 + ], + "type": "text", + "content": "We have conducted experiments in the NAS-Bench-201 [10] and MobileNetV2 [26, 38] search spaces. To obtain easily comparable results, we used 64 randomly generated input images to compute our score as in [21]. For the methods that rely on the knowledge of true labels, input data from the respective datasets were used." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "NAS-Bench-201. The dataset consists of 15,625 networks for which validation accuracies for CIFAR-10, CIFAR-100 [18] and ImageNet16-120 [7] after training for 200 epochs are provided. The networks are characterized by unique cell structures comprising of several types of operation choices. As one of the possible choices is zero operation, it's possible that some computation edges don't receive any input or cannot propagate their results to the out" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 317, + 73, + 432, + 185 + ], + "blocks": [ + { + "bbox": [ + 317, + 73, + 432, + 185 + ], + "lines": [ + { + "bbox": [ + 317, + 73, + 432, + 185 + ], + "spans": [ + { + "bbox": [ + 317, + 73, + 432, + 185 + ], + "type": "image", + "image_path": "8b64c7352464abe41619d6bfc3723ff08879a9dd2ad85c19d55acf23710d5d61.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 433, + 73, + 550, + 185 + ], + "blocks": [ + { + "bbox": [ + 433, + 73, + 550, + 185 + ], + "lines": [ + { + "bbox": [ + 433, + 73, + 550, + 185 + ], + "spans": [ + { + "bbox": [ + 433, + 73, + 550, + 185 + ], + "type": "image", + "image_path": "e4f5f01c1888fabdc9b7380bd8dd2f34a99e4c8ec21c3a7473348ecbc2f28be5.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 317, + 193, + 430, + 308 + ], + "blocks": [ + { + "bbox": [ + 317, + 193, + 430, + 308 + ], + "lines": [ + { + "bbox": [ + 317, + 193, + 430, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 193, + 430, + 308 + ], + "type": "image", + "image_path": "d77d22122e5743a9ff75154781478d7c808cfa3c1f380a6c73f447a4dd5c9bd9.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "type": "text", + "content": "Figure 3. Components of AZ-NAS [21] and our VKDNW are compared w.r.t. correlation with " + }, + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "type": "inline_equation", + "content": "\\aleph" + }, + { + "bbox": [ + 313, + 316, + 555, + 372 + ], + "type": "text", + "content": " (number of trainable layers), in the NAS-Bench-201 search space [10] on ImageNet16-120 [7] dataset. Our VKDNW proxy has the lowest correlation, ie. is the most invariant to the size of the model." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 433, + 193, + 550, + 308 + ], + "blocks": [ + { + "bbox": [ + 433, + 193, + 550, + 308 + ], + "lines": [ + { + "bbox": [ + 433, + 193, + 550, + 308 + ], + "spans": [ + { + "bbox": [ + 433, + 193, + 550, + 308 + ], + "type": "image", + "image_path": "c1dcc1c3cc570de06fc5b5da33871e3d569e4802867351dd28e012ea923c7e9b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 384, + 555, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 527 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 527 + ], + "type": "text", + "content": "put, leading to same computation graphs for different architectures, thus duplicating networks. Following the practice from NAS-Bench-101 [46], we report results on 9,445 unique structures (also in [14, 30]) and we refer the reader to Supplementary material for results on all networks. We measure proxies performance via Kendall's " + }, + { + "bbox": [ + 313, + 384, + 555, + 527 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 384, + 555, + 527 + ], + "type": "text", + "content": " (KT) and Spearman's " + }, + { + "bbox": [ + 313, + 384, + 555, + 527 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 384, + 555, + 527 + ], + "type": "text", + "content": " (SPR) correlations with validation accuracies together with Normalized Discounted Cumulative Gain " + }, + { + "bbox": [ + 313, + 384, + 555, + 527 + ], + "type": "inline_equation", + "content": "(\\mathrm{nDCG}_{1000})" + }, + { + "bbox": [ + 313, + 384, + 555, + 527 + ], + "type": "text", + "content": ", we write nDCG for short) from Sec. 3, averaged over 5 independent runs. For the model-driven aggregation, we trained a random forest model on 1024 networks for 100 iterations, and used the rest for testing." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 529, + 556, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 556, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 556, + 685 + ], + "type": "text", + "content": "In Table 1, we compare our method with existing NAS methods and show that " + }, + { + "bbox": [ + 313, + 529, + 556, + 685 + ], + "type": "inline_equation", + "content": "\\mathrm{VKDNW_{agg}}" + }, + { + "bbox": [ + 313, + 529, + 556, + 685 + ], + "type": "text", + "content": " outperforms all methods with a significant margin in all three metrics. The benefit of our method is two-fold: a) it identifies the best networks due to its high performance in nDCG, which is the desirable property in NAS b) the ranking is consistent across the whole network search space due to high KT and SPR correlations. Note that we can only make these observations by combining standard correlation metrics and our newly proposed nDCG. We also show that " + }, + { + "bbox": [ + 313, + 529, + 556, + 685 + ], + "type": "inline_equation", + "content": "\\mathrm{VKDNW_{single}}" + }, + { + "bbox": [ + 313, + 529, + 556, + 685 + ], + "type": "text", + "content": " outperforms all other single-rank proxies in all metrics on ImageNet16-120, in SPR on all three datasets, while being among the highest on CIFAR-10 and CIFAR-100." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "MobileNetV2. We search for the best network configuration in the MobileNetV2 space [38], while constraining" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14887" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 70, + 277, + 205 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 277, + 205 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 277, + 205 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 277, + 205 + ], + "type": "table", + "html": "
VJETF(KT)(SPR)nDCG
✓0.6220.8140.608
✓0.6030.7810.339
✓0.5880.7790.274
✓0.3530.5170.233
✓0.5450.7180.403
✓✓✓✓0.7170.8910.623
✓✓✓✓0.7060.8710.553
✓✓✓✓0.7360.9050.695
✓✓✓✓0.7220.8960.658
✓✓✓✓0.7350.9010.646
✓✓✓✓✓0.7430.9060.664
", + "image_path": "a1cef225d49c3567b59bf248158c844b16ae9762623e10fb25ab5bec6fb8a27e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 214, + 295, + 291 + ], + "lines": [ + { + "bbox": [ + 55, + 214, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 214, + 295, + 291 + ], + "type": "text", + "content": "Table 3. Components of VKDNW" + }, + { + "bbox": [ + 55, + 214, + 295, + 291 + ], + "type": "inline_equation", + "content": "_{\\text{agg}}" + }, + { + "bbox": [ + 55, + 214, + 295, + 291 + ], + "type": "text", + "content": " rank with non-linear aggregation. Consistency is shown with respect to Kendall's " + }, + { + "bbox": [ + 55, + 214, + 295, + 291 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 214, + 295, + 291 + ], + "type": "text", + "content": " (KT), Spearman's " + }, + { + "bbox": [ + 55, + 214, + 295, + 291 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 214, + 295, + 291 + ], + "type": "text", + "content": " (SPR) and Normalized Discounted Cumulative Gain (nDCG) with " + }, + { + "bbox": [ + 55, + 214, + 295, + 291 + ], + "type": "inline_equation", + "content": "P = 1000" + }, + { + "bbox": [ + 55, + 214, + 295, + 291 + ], + "type": "text", + "content": " on ImageNet16-120 image dataset [7]. Here V, J, E, T and F stand for VKDNW, Jacov, expressivity, trainability and FLOPs respectively (see Sec. 5.1). Table of all combinations can be found in the Supplementary material." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 298, + 295, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 295, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 295, + 393 + ], + "type": "text", + "content": "the model size to approximately 450M FLOPs. We ran 100,000 iterations of the evolutionary search algorithm [21] for approximately 10 hours, meaning 100,000 different architectures were evaluated. We then took the best network from the search and trained it for 480 epochs on ImageNet-1K [9], using the same hyper-parameter setting as in [21, 25]. The final training of the model took 7 days on 8xNVidia A100 GPUs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 394, + 296, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 296, + 441 + ], + "type": "text", + "content": "As seen in Table 2, our method " + }, + { + "bbox": [ + 55, + 394, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\mathrm{VKDNW_{agg}}" + }, + { + "bbox": [ + 55, + 394, + 296, + 441 + ], + "type": "text", + "content": " outperforms all prior approaches - even train-based approaches (denoted MS and OS) that incur much higher computational costs for the search." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 450, + 124, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 124, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 124, + 462 + ], + "type": "text", + "content": "5.3. Ablations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 469, + 295, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 295, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 295, + 493 + ], + "type": "text", + "content": "We ablate our method in the NAS-Bench-201 search space [10] on ImageNet16-120 [7] validation accuracies." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 496, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 651 + ], + "type": "text", + "content": "Orthogonality of VKDNW. Our score VKDNW has a desirable property that it is based on information orthogonal to the size of the network: in Figure 3 we can see that unlike previous work, VKDNW is not correlated with the network size measured by " + }, + { + "bbox": [ + 55, + 496, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\aleph" + }, + { + "bbox": [ + 55, + 496, + 295, + 651 + ], + "type": "text", + "content": " (number of trainable layers). We believe this property is key when stepping into much larger search spaces, however components of the previous state-of-the-art method AZ-NAS lack such a property (also Fig. 3 and Tab. 4). We conjecture that improvement in the key metrics by VKDNW is caused by this orthogonality feature. In Table 1 we also provide comparison to previous model-driven method and show that adding our feature significantly improves the ranking in all considered metrics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "Components of aggregated rank. Our single-rank variant " + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{VKDNW}_{\\mathrm{single}}" + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": " is the strongest component of the aggregated " + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{VKDNW}_{\\mathrm{agg}}" + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": " as can be seen in Table 3 where it outperforms rest of the rankings with especially large margin in nDCG, thus, it strongly discriminates high-accuracy net" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 70, + 553, + 140 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 553, + 140 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 553, + 140 + ], + "type": "table", + "html": "
(KT)
CIFAR-10CIFAR-100ImageNet16-120
VKDNW0.041-0.090-0.107
trainability0.2200.2200.253
expressivity0.5390.5390.539
progressivity0.3980.3980.385
", + "image_path": "cb7bf85adbcaf5499790f402df0eef7735b89c2f10c6ac0befb9e308c88401ed.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 148, + 555, + 192 + ], + "lines": [ + { + "bbox": [ + 313, + 148, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 555, + 192 + ], + "type": "text", + "content": "Table 4. Kendall's " + }, + { + "bbox": [ + 313, + 148, + 555, + 192 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 148, + 555, + 192 + ], + "type": "text", + "content": " (KT) of " + }, + { + "bbox": [ + 313, + 148, + 555, + 192 + ], + "type": "inline_equation", + "content": "\\aleph" + }, + { + "bbox": [ + 313, + 148, + 555, + 192 + ], + "type": "text", + "content": " (number of trainable layers) with components of AZ-NAS compared to our new method VKDNW on three datasets on NAS-Bench-201 search space [10] and multiple image datasets." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 206, + 553, + 287 + ], + "blocks": [ + { + "bbox": [ + 317, + 206, + 553, + 287 + ], + "lines": [ + { + "bbox": [ + 317, + 206, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 317, + 206, + 553, + 287 + ], + "type": "table", + "html": "
Batch SizeRandom InputReal Input
KTSPRnDCGKTSPRnDCG
80.6150.8080.6170.6140.8050.599
160.6140.8080.6040.6080.8000.595
320.6190.8120.6110.6120.8040.594
640.6210.8140.6170.6140.8060.595
1280.6210.8140.6160.6150.8060.591
", + "image_path": "04bdf7d7ea1cdd7aa227334d50ef65bf3925da68f1c02ba7bb191aabfce84c9a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 295, + 555, + 339 + ], + "lines": [ + { + "bbox": [ + 313, + 295, + 555, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 555, + 339 + ], + "type": "text", + "content": "Table 5. Our method VKDNW evaluated for different batch sizes using either randomly generated input data or real images with respect to Kendall's " + }, + { + "bbox": [ + 313, + 295, + 555, + 339 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 295, + 555, + 339 + ], + "type": "text", + "content": " (KT), Spearman's " + }, + { + "bbox": [ + 313, + 295, + 555, + 339 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 295, + 555, + 339 + ], + "type": "text", + "content": " (SPR) and Normalized Discounted Cumulative Gain with " + }, + { + "bbox": [ + 313, + 295, + 555, + 339 + ], + "type": "inline_equation", + "content": "P = 1000" + }, + { + "bbox": [ + 313, + 295, + 555, + 339 + ], + "type": "text", + "content": " (nDCG)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 350, + 423, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 423, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 423, + 361 + ], + "type": "text", + "content": "works compared to others." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 365, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 472 + ], + "type": "text", + "content": "Random or real input. As our method is computed using generated random data (white noise) it does not rely on any real dataset and is therefore applicable also in situations where no reliable data are available. Table 5 shows that by relying just on random data we do not lose any performance with respect to all key metrics. Moreover, the performance remains relatively stable across different batch sizes. We chose batch size 64 as larger batch size does not bring better results." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 475, + 554, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 554, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 554, + 499 + ], + "type": "text", + "content": "Due to lack of space, we kindly refer the reader to Supplementary material for further ablations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 510, + 388, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 388, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 388, + 522 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 532, + 554, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 554, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 554, + 592 + ], + "type": "text", + "content": "We proposed a new training-free NAS proxy called Variance of Knowledge of Deep Network Weights (VKDNW). The method has strong theoretical support and achieved state-of-the-art results on three public datasets and in two search spaces." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "Based on throughout evaluation and comparison to other existing approaches, it has been shown that it provides an information orthogonal to the network size leading to a zero-cost ranking were contribution of the network size and architecture feasibility are separated. We have also shown that previously used correlation metrics for proxy evaluations do not sufficiently assess the key ability to discriminate top networks and to address this problem we have proposed a new evaluation metric and re-evaluated previous methods with the new metric." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14888" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "content": "The research was supported by Czech Science Foundation Grant No. 24-10738M. The access to the computational infrastructure of the OP VVV funded project CZ.02.1.01/0.0/0.0/16_019/0000765 “Research Center for Informatics” and of the Ministry of Education, Youth and Sports of the Czech Republic project e-INFRA CZ (ID:90254) is also gratefully acknowledged." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 115, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 115, + 197 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 115, + 197 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 204, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 204, + 295, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 204, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 61, + 204, + 295, + 237 + ], + "type": "text", + "content": "[1] Mohamed S Abdelfattah, Abhinav Mehrotra, Lukasz Dudziak, and Nicholas D Lane. Zero-cost proxies for lightweight NAS. arXiv preprint arXiv:2101.08134, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 238, + 295, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 238, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 61, + 238, + 295, + 291 + ], + "type": "text", + "content": "[2] Chris Burges, Tal Shaked, Erin Renshaw, Ari Lazier, Matt Deeds, Nicole Hamilton, and Greg Hullender. Learning to rank using gradient descent. In Proceedings of the 22nd international conference on Machine learning, pages 89-96, 2005." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 293, + 295, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 293, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 62, + 293, + 295, + 326 + ], + "type": "text", + "content": "[3] Han Cai, Ligeng Zhu, and Song Han. Proxylessnas: Direct neural architecture search on target task and hardware. arXiv preprint arXiv:1812.00332, 2018." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 327, + 295, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 327, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 327, + 295, + 369 + ], + "type": "text", + "content": "[4] Han Cai, Chuang Gan, Tianzhe Wang, Zhekai Zhang, and Song Han. Once-for-all: Train one network and specialize it for efficient deployment. arXiv preprint arXiv:1908.09791, 2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 370, + 296, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 370, + 296, + 425 + ], + "spans": [ + { + "bbox": [ + 62, + 370, + 296, + 425 + ], + "type": "text", + "content": "[5] Boyu Chen, Peixia Li, Baopu Li, Chen Lin, Chuming Li, Ming Sun, Junjie Yan, and Wanli Ouyang. BN-NAS: Neural architecture search with batch normalization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 307-316, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 426, + 295, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 426, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 426, + 295, + 468 + ], + "type": "text", + "content": "[6] Wuyang Chen, Xinyu Gong, and Zhangyang Wang. Neural architecture search on imagenet in fourgpu hours: A theoretically inspired perspective. arXiv preprint arXiv:2102.11535, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 470, + 295, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 470, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 62, + 470, + 295, + 502 + ], + "type": "text", + "content": "[7] Patryk Chrabaszcz, Ilya Loshchilov, and Frank Hutter. A downsampled variant of imagenet as an alternative to the cifar datasets. arXiv preprint arXiv:1707.08819, 2017." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 503, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 503, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 503, + 295, + 525 + ], + "type": "text", + "content": "[8] Harald Cramér. Mathematical methods of statistics. Princeton university press, 1999." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 525, + 295, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 295, + 569 + ], + "type": "text", + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 570, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 570, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 570, + 295, + 602 + ], + "type": "text", + "content": "[10] Xuanyi Dong and Yi Yang. Nas-bench-201: Extending the scope of reproducible neural architecture search. arXiv preprint arXiv:2001.00326, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 603, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 295, + 635 + ], + "type": "text", + "content": "[11] Roy Frieden and Robert A Gatenby. Exploratory data analysis using Fisher information. Springer Science & Business Media, 2010." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 636, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 669 + ], + "type": "text", + "content": "[12] Boris Hanin and David Rolnick. Complexity of linear regions in deep networks. In International Conference on Machine Learning, pages 2596-2604. PMLR, 2019." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[13] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[14] Gabriela Kadlecova, Jovita Lukasik, Martin Pilát, Petra Vidnerova, Mahmoud Safari, Roman Neruda, and Frank Hutter. Surprisingly strong performance prediction with neural graph features. arXiv preprint arXiv:2404.16551, 2024." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "text", + "content": "[15] Ryo Karakida, Shotaro Akaho, and Shun-ichi Amari. Pathological spectra of the fisher information metric and its variants in deep neural networks. arXiv preprint arXiv:1910.05992, 2019." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "type": "text", + "content": "[16] Ryo Karakida, Shotaro Akaho, and Shun-ichi Amari. Universal statistics of fisher information in deep neural networks: Mean field approach. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 1032–1041. PMLR, 2019." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 219, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 240 + ], + "type": "text", + "content": "[17] Maurice G Kendall. A new measure of rank correlation. Biometrika, 30(1-2):81-93, 1938." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 241, + 553, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 264 + ], + "type": "text", + "content": "[18] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 265, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 265, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 553, + 308 + ], + "type": "text", + "content": "[19] Frederik Kunstner, Philipp Hennig, and Lukas Balles. Limitations of the empirical fisher approximation for natural gradient descent. Advances in neural information processing systems, 32, 2019." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 310, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 364 + ], + "type": "text", + "content": "[20] Byung-Kwan Lee, Junho Kim, and Yong Man Ro. Masking adversarial damage: Finding adversarial saliency for robust and sparse network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15126-15136, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "type": "text", + "content": "[21] Junghyup Lee and Bumsub Ham. AZ-NAS: Assembling zero-cost proxies for network architecture search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5893-5903, 2024." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 411, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 465 + ], + "type": "text", + "content": "[22] Jaehoon Lee, Lechao Xiao, Samuel Schoenholz, Yasaman Bahri, Roman Novak, Jascha Sohl-Dickstein, and Jeffrey Pennington. Wide neural networks of any depth evolve as linear models under gradient descent. Advances in neural information processing systems, 32, 2019." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 467, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 467, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 467, + 553, + 499 + ], + "type": "text", + "content": "[23] Namhoon Lee, Thalaiyasingam Ajanthan, and Philip HS Torr. Snip: Single-shot network pruning based on connection sensitivity. arXiv preprint arXiv:1810.02340, 2018." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 500, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 533 + ], + "type": "text", + "content": "[24] Steven J Leon, Lisette De Pillis, and Lisette G De Pillis. Linear algebra with applications. Pearson Prentice Hall Upper Saddle River, NJ, 2006." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "type": "text", + "content": "[25] Guihong Li, Yuedong Yang, Kartikeya Bhardwaj, and Radu Marculescu. ZiCo: Zero-shot NAS via inverse coefficient of variation on gradients. In The Eleventh International Conference on Learning Representations, 2023." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 579, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 634 + ], + "type": "text", + "content": "[26] Ming Lin, Pichao Wang, Zhenhong Sun, Hesen Chen, Xiuyu Sun, Qi Qian, Hao Li, and Rong Jin. Zen-nas: A zero-shot nas for high-performance image recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 347-356, 2021." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "text", + "content": "[27] Hanxiao Liu, Karen Simonyan, and Yiming Yang. DARTS: Differentiable architecture search. arXiv preprint arXiv:1806.09055, 2018." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "text", + "content": "[28] Alexander Ly, Maarten Marsman, Josine Verhagen, Raoul PPP Grasman, and Eric-Jan Wagenmakers. A tutorial on fisher information. Journal of Mathematical Psychology, 80:40-55, 2017." + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14889" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[29] James Martens. New insights and perspectives on the natural gradient method. Journal of Machine Learning Research, 21 (146):1-76, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 173 + ], + "type": "text", + "content": "[30] Abhinav Mehrotra, Alberto Gil CP Ramos, Sourav Bhattacharya, Lukasz Dudziak, Ravichander Vipperla, Thomas Chau, Mohamed S Abdelfattah, Samin Ishtiaq, and Nicholas Donald Lane. Nas-bench-asr: Reproducible neural architecture search for speech recognition. In International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "text", + "content": "[31] Joe Mellor, Jack Turner, Amos Storkey, and Elliot J Crowley. Neural architecture search without training. In International conference on machine learning, pages 7588-7598. PMLR, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "type": "text", + "content": "[32] Bert Moons, Parham Noorzad, Andrii Skliar, Giovanni Mariani, Dushyant Mehta, Chris Lott, and Tijmen Blankevoort. Distilling optimal neural networks: Rapid search in diverse spaces. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12229-12238, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 329 + ], + "type": "text", + "content": "[33] Xuefei Ning, Changcheng Tang, Wenshuo Li, Zixuan Zhou, Shuang Liang, Huazhong Yang, and Yu Wang. Evaluating efficient performance estimators of neural architectures. Advances in Neural Information Processing Systems, 34: 12265-12277, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 374 + ], + "type": "text", + "content": "[34] Roman Novak, Jascha Sohl-Dickstein, and Samuel S Schoenholz. Fast finite width neural tangent kernel. In International Conference on Machine Learning, pages 17018-17044. PMLR, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "type": "text", + "content": "[35] Hyeyoung Park and Kwanyong Lee. Adaptive natural gradient method for learning of stochastic neural networks in mini-batch mode. Applied Sciences, 9(21):4568, 2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "text", + "content": "[36] Jeffrey Pennington and Pratik Worah. The spectrum of the fisher information matrix of a single-hidden-layer neural network. Advances in neural information processing systems, 31, 2018." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 498 + ], + "type": "text", + "content": "[37] C Radhakrishna Rao. Information and the accuracy attainable in the estimation of statistical parameters. In Breakthroughs in Statistics: Foundations and basic theory, pages 235-247. Springer, 1992." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "text", + "content": "[38] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *Proceedings of the IEEE conference on computer vision and pattern recognition*, pages 4510-4520, 2018." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 578 + ], + "type": "text", + "content": "[39] Charles Spearman. The proof and measurement of association between two things. 1961." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 580, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 294, + 634 + ], + "type": "text", + "content": "[40] Kunio Tanabe and Masahiko Sagae. An exact cholesky decomposition and the generalized inverse of the variance-covariance matrix of the multinomial distribution, with applications. Journal of the Royal Statistical Society: Series B (Methodological), 54(1):211-219, 1992." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 636, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 679 + ], + "type": "text", + "content": "[41] Hidenori Tanaka, Daniel Kunin, Daniel L Yamins, and Surya Ganguli. Pruning neural networks without any data by iteratively conserving synaptic flow. Advances in neural information processing systems, 33:6377-6389, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[42] Chaoqi Wang, Guodong Zhang, and Roger Grosse. Picking winning tickets before training by preserving gradient flow. arXiv preprint arXiv:2002.07376, 2020." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 452 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[43] Colin White, Mikhail Khodak, Renbo Tu, Shital Shah, Sébastien Bubeck, and Debadeepta Dey. A deeper look at zero-cost proxies for lightweight NAS. *ICLR Blog Track*, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "text", + "content": "[44] Huan Xiong, Lei Huang, Mengyang Yu, Li Liu, Fan Zhu, and Ling Shao. On the number of linear regions of convolutional neural networks. In International Conference on Machine Learning, pages 10514-10523. PMLR, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "text", + "content": "[45] Zhaohui Yang, Yunhe Wang, Xinghao Chen, Boxin Shi, Chao Xu, Chunjing Xu, Qi Tian, and Chang Xu. CARS: Continuous evolution for efficient neural architecture search. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1829-1838, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 271 + ], + "type": "text", + "content": "[46] Chris Ying, Aaron Klein, Eric Christiansen, Esteban Real, Kevin Murphy, and Frank Hutter. Nas-bench-101: Towards reproducible neural architecture search. In International conference on machine learning, pages 7105-7114. PMLR, 2019." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 274, + 553, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 318 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 318 + ], + "type": "text", + "content": "[47] Wang Yining, Wang Liwei, Li Yuanzhi, He Di, Chen Wei, and Liu Tie-Yan. A theoretical analysis of ndcg ranking measures. In Proceedings of the 26th annual conference on learning theory, 2013." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "type": "text", + "content": "[48] Xuanyang Zhang, Pengfei Hou, Xiangyu Zhang, and Jian Sun. Neural architecture search with random labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10907-10916, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 365, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 396 + ], + "type": "text", + "content": "[49] Zhihao Zhang and Zhihao Jia. Gradsign: Model performance inference with theoretical insights. arXiv preprint arXiv:2110.08616, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 398, + 553, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 398, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 316, + 398, + 553, + 452 + ], + "type": "text", + "content": "[50] Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V Le. Learning transferable architectures for scalable image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8697-8710, 2018." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14890" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Trajectory Mamba_ Efficient Attention-Mamba Forecasting Model Based on Selective SSM/14500b01-a80f-4244-a6f9-1c844f2302da_content_list.json b/2025/Trajectory Mamba_ Efficient Attention-Mamba Forecasting Model Based on Selective SSM/14500b01-a80f-4244-a6f9-1c844f2302da_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0d1138a5ac6b5dbea4bddec99b93ec510b49f61f --- /dev/null +++ b/2025/Trajectory Mamba_ Efficient Attention-Mamba Forecasting Model Based on Selective SSM/14500b01-a80f-4244-a6f9-1c844f2302da_content_list.json @@ -0,0 +1,1505 @@ +[ + { + "type": "text", + "text": "Trajectory Mamba: Efficient Attention-Mamba Forecasting Model Based on Selective SSM", + "text_level": 1, + "bbox": [ + 114, + 128, + 883, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yizhou Huang $^{1}$ , Yihua Cheng $^{2}$ , Kezhi Wang $^{1*}$ $^{1}$ Brunel University of London, $^{2}$ University of Birmingham \n{byyh009, Kezhi.Wang}@brunel.ac.uk, y.cheng.2@bham.ac.uk \nhttps://github.com/Yizhou-H/Trajectory-Mamba-CVPR", + "bbox": [ + 238, + 202, + 751, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 309, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Motion prediction is crucial for autonomous driving, as it enables accurate forecasting of future vehicle trajectories based on historical inputs. This paper introduces Trajectory Mamba, a novel efficient trajectory prediction framework based on the selective state-space model (SSM). Conventional attention-based models face the challenge of computational costs that grow quadratically with the number of targets, hindering their application in highly dynamic environments. In response, we leverage the SSM to redesign the self-attention mechanism in the encoder-decoder architecture, thereby achieving linear time complexity. To address the potential reduction in prediction accuracy resulting from modifications to the attention mechanism, we propose a joint polyline encoding strategy to better capture the associations between static and dynamic contexts, ultimately enhancing prediction accuracy. Additionally, to balance prediction accuracy and inference speed, we adopted the decoder that differs entirely from the encoder. Through cross-state space attention, all target agents share the scene context, allowing the SSM to interact with the shared scene representation during decoding, thus inferring different trajectories over the next prediction steps. Our model achieves state-of-the-art results in terms of inference speed and parameter efficiency on both the Argoverse 1 and Argoverse 2 datasets. It demonstrates a four-fold reduction in FLOPs compared to existing methods and reduces parameter count by over $40\\%$ while surpassing the performance of the vast majority of previous methods. These findings validate the effectiveness of Trajectory Mamba in trajectory prediction tasks.", + "bbox": [ + 89, + 340, + 483, + 781 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 806, + 220, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the rapid development of autonomous driving technology [3, 12, 17], accurate motion prediction plays a crucial role by forecasting future trajectories based on historical vehicle", + "bbox": [ + 89, + 832, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0699c5c3ed435b426a368e436b3fd6c42cded92936a34d53a12a03dea185f68e.jpg", + "image_caption": [ + "Figure 1. Illustration of proposed joint polyline encoding strategy, where we consider all factors that affect the movement of motor vehicles, jointly encoding pedestrians and traffic lights rather than categorizing the scene horizontally into static and dynamic context. Additionally, we decompose and encode the interactions between all agents and elements at each time step." + ], + "image_footnote": [], + "bbox": [ + 519, + 309, + 903, + 491 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "movement data. This task is not only vital for enhancing driving safety but also meets the stringent requirements for real-time system response. In recent years, learning-based approaches have been widely applied to motion prediction [5, 16, 29, 39, 40]. However, existing prediction methods often struggle to balance efficiency and accuracy, making it challenging to simultaneously meet performance and efficiency requirements in real-time applications. We attempt to analyze the underlying reasons from the following aspects and explore possible improvements:", + "bbox": [ + 511, + 613, + 906, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(a). Improving Prediction Accuracy: In autonomous driving scenarios [38], there are numerous dynamic and static agents, and the diversity of this data presents significant challenges for motion prediction and model training. However, conventional methods [29, 39] often struggle to effectively fuse this heterogeneous data, particularly when it comes to integrating various features, which leads to suboptimal prediction accuracy. On a positive note, factorized attention-based transformers [21, 40] have made no", + "bbox": [ + 511, + 765, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 886, + 235, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12058", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "table strides in decomposing heterogeneous data. However, these models often use a unified, horizontal approach when dealing with static scenes and dynamic agent features, and they fail to fully consider the deeper relationships between different scene elements and agent information.", + "bbox": [ + 93, + 90, + 480, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(b). Enhancing Response Efficiency: Motion predictions are inherently recursive, implying that each prediction serves as an input for the next time step, leading to the progressive accumulation of errors over time. Notably, the advancement of sparse context encoding techniques [5, 35] has demonstrated significant potential in enhancing model efficiency. Nonetheless, prediction models based on attention mechanisms still face exponential growth in computational complexity and parameter size during recursive tasks due to inherent design constraints, which severely limit their efficiency and real-time capabilities in practical applications. Conversely, methods such as LSTM [18] are relatively more efficient, but numerous studies [20, 23, 36] have shown that they are incapable of achieving the predictive accuracy of transformer-based models.", + "bbox": [ + 93, + 169, + 480, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In response, we propose an Attention-Mamba framework for motion prediction, termed Trajectory Mamba (Tamba), to overcome the limitations discussed above. First, we design a multimodal encoder-decoder architecture based on the state space model [9, 11, 27], comprising three parallel encoders and a decoder. By introducing a selective state space, we redefined the computation of the self-attention mechanism, enabling the model to selectively process input information by parameterizing the input of the state space model (SSM), thereby allowing the model to focus on or ignore specific inputs. Specifically, the selective state space module processes local short-range features by receiving input from a one-dimensional convolutional layer, ensuring that these features can be selected or discarded before entering the SSM. We utilized local state space attention in place of the multi-head attention in the self-attention module, reducing the computational complexity of recursive reasoning from exponential to linear.", + "bbox": [ + 93, + 398, + 480, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Secondly, to better integrate heterogeneous data features, we proposed a joint polyline encoding strategy. Its objective is to separate pedestrian agents from the dynamic agents and jointly encode them with traffic light information from the static scene (see Fig. 1). We observed that in uncertain motion prediction scenarios, the application of traffic regulations exerts a degree of implicit control over the potential movement states of agents. For example, pedestrian-priority traffic rules remain effective even in the absence of traffic lights, influencing vehicles and motorcycles alike. By jointly encoding pedestrians and traffic lights through a shared embedder, this jointly encoding strategy can independently interact with other dynamic agents to enhance the prediction accuracy of vehicle and motorcycle agents.", + "bbox": [ + 93, + 672, + 480, + 882 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Lastly, we designed a prediction weight inference mod", + "bbox": [ + 112, + 886, + 480, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ule. By redesigning selective state space with a cross-attention mechanism, we utilized independent query tensors to recursively infer multimodal candidate trajectories through interactions with the encoder and the keys and values from the previous recursive step. These candidate trajectories were then evaluated using a recurrent neural network to assign trajectory scores.", + "bbox": [ + 516, + 90, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as: (1) We propose the Tamba framework for motion prediction, which leverages the state space model to redesign computations of selfattention, resulting in computational complexity reducing significantly. (2) We propose a joint polyline encoding strategy. By integrating strongly correlated polyline types using independent or shared embeddings, elements that impose similar behavioral patterns can better influence different agents, collectively contributing to the trajectory reasoning of these agents. (3) We redesign the Mamba decoder with crossattention, allowing the joint scene encoding to share a unified scene representation, which further reduces the model's complexity. At the final stage, we introduced a prediction weight inference module to refine the predicted trajectories and enhance prediction accuracy. (4) Experimental results demonstrate that our method achieves state-of-the-art (SOTA) performance on Argoverse 1 and 2 datasets. Compared to advanced methods, our method reduces the number of parameters by $40\\%$ , achieves a four-fold increase in FLOPs. This represents a new level of balance between efficiency and performance in our model.", + "bbox": [ + 516, + 199, + 903, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 516, + 532, + 660, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. State Space Model", + "text_level": 1, + "bbox": [ + 516, + 558, + 689, + 573 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, the Mamba [7] framework has rekindled interest in state-space models (SSMs) [9, 11, 27] due to its introduction of an input-dependent selection mechanism, which enhances their applicability and performance. SSMs are renowned for their capacity to handle long-range dependencies and offer superior memory and computational efficiency compared to transformer-based architectures [30]. Hippo [8], a seminal contribution to SSMs, established the efficacy of these models in long-range sequence modeling tasks. Subsequent research, such as Vision mamba [41], has further refined the mechanisms underlying SSMs, resulting in substantial performance improvements. SSMs have demonstrated their versatility and robustness across a diverse set of domains, including computer vision, video analysis, medical imaging, graph-based learning, point cloud processing, and recommendation systems.", + "bbox": [ + 516, + 582, + 903, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the domain of trajectory prediction, recent studies have indicated that Motion Mamba [37] achieves notable success in predicting human trajectory time series. Building on these advancements, this work integrates the state-space mechanisms of Mamba with attention mechanisms to address tra", + "bbox": [ + 516, + 825, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12059", + "bbox": [ + 480, + 946, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "jectory prediction tasks in dynamic traffic scenarios, thereby leveraging both the efficiency of SSMs and the adaptability of attention-based models.", + "bbox": [ + 89, + 90, + 482, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Attention Trajectory Prediction Model", + "text_level": 1, + "bbox": [ + 89, + 146, + 421, + 162 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the field of trajectory prediction, methods [13, 19] based on the attention mechanism have been widely adopted to enhance the model's ability to capture and predict complex behaviors. These approaches leverage the capability of attention to focus selectively on important parts of the input, thereby capturing key temporal and spatial features that are crucial for predicting future trajectories.", + "bbox": [ + 89, + 170, + 482, + 275 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Subsequently, several advanced approaches [17, 34] have utilized attention mechanisms for trajectory prediction specifically in multi-agent interaction environments, where understanding the interactions between different agents (such as vehicles, pedestrians, and other road users) is critical. These models employ social attention modules to focus on the movements and behaviors of surrounding vehicles, effectively improving the ability of the system to handle intricate interactive behaviors, which are a common aspect of real-world driving scenarios.", + "bbox": [ + 88, + 276, + 482, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, attention mechanisms have also been successfully applied in models that integrate multimodal information [25, 26, 40]. These models dynamically weight and fuse different types of sensory inputs—such as image features from cameras and data from other sensors like LiDAR—through attention modules. This fusion process allows the model to better understand and represent complex traffic environments, effectively utilizing complementary data to enhance trajectory prediction performance.", + "bbox": [ + 89, + 428, + 482, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our Method", + "text_level": 1, + "bbox": [ + 89, + 577, + 220, + 593 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 89, + 603, + 200, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the autonomous driving scenario, the input consists of the predicted dynamic agent, its surrounding agents, and the static scene context. Let the historical state of the $i$ -th agent at time step $t$ be represented as: $s_i^t = (p_i^t,\\theta_i^t,t,\\nu_i^t)$ , where $p_i^t = (p_{i,x}^t,p_{i,y}^t)$ denotes the spatial position, $\\theta_i^t$ represents the heading angle, $t$ refers to the time step, and $\\nu_{i}^{t}$ denotes the velocity. The static scene context is represented by $n$ polygons, which include elements such as lane boundaries, map edges, and sidewalks. The agent states $s_i^t$ over $T$ time steps within the observation window are used to predict $K$ future trajectories over $T^{\\prime}$ time steps.", + "bbox": [ + 89, + 626, + 482, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we propose the Tamba model to optimize the efficiency and accuracy of motion forecasting tasks. Our input includes $s_i^t$ of dynamic and static agents, $\\mathcal{P}$ processed using a heterogeneous polyline encoding strategy to transform $s_i^t$ into higher-dimensional features. We leverage three Tamba encoders to separately decompose (I) attendations in each time step $(t)$ over all elements, (II) attendations between", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d561ba29edd4ce0fbb3275cf902483eacb131baa7b85efa177e812082709735e.jpg", + "image_caption": [ + "Figure 2. Overview of our Tamba encoder. We employ joint polyline encoding strategy to integrate strongly correlated polyline information and used three parallel encoders to interact with and associate these features. We applied a linear projection to map the attention output of Tamba back to the same dimensions as the input and used normalization to enhance the stability of the output." + ], + "image_footnote": [], + "bbox": [ + 516, + 92, + 903, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "all agents (A) and scene (S), and (III) attentions of traffic control elements (pedestrians and traffic lights) to other dynamic agents (motorcycles and cars). Then, concatenate features from encoders as output to a Tamba decoder, where we utilize cross-attention mechanism to generate proposal trajectories. Subsequently, we fused scene encoding with historical information to perform a secondary decoding of the proposal trajectories for refinement. Finally, the refined trajectories were optimized using a \"winner-takes-all strategy\" to obtain trajectory output.", + "bbox": [ + 511, + 398, + 906, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our model contains a novel heterogeneous polyline encoding strategy along with a decomposed feature fusion method for both static and dynamic agents. A redesigned Tamba encoder-decoder architecture. A refinement process to output the final predicted trajectory. In the following sections, we sequentially introduce the components in Tamba.", + "bbox": [ + 511, + 551, + 905, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Joint Polyline Encoding", + "text_level": 1, + "bbox": [ + 511, + 650, + 732, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In multimodal traffic environment prediction tasks, different types of polylines possess unique geometric characteristics and semantic information. This study proposes a novel heterogeneous polyline embedding method to facilitate the fusion of dynamic and static context features, thereby providing a richer and more efficient feature representation. Specifically, we assign either independent or shared embedders to each type of polyline to effectively model various categories of information. Specifically, we jointly encode polylines with strong information associations, such as pedestrian trajectories and traffic light signals, using a shared embedder. This is because agents (pedestrians) and traffic lights typically impose similar behavioral patterns on other agents (vehicles) and often participate together in traffic control. Thus, using a shared embedder reduces model complexity", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12060", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and promotes feature sharing, collectively influencing vehicle trajectory reasoning.", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast, independent embedders are used for polyline types with significant feature differences, such as lane lines and traffic signs in maps. Each embedder consists of a MLP layer and normalization layers to extract high-dimensional feature representations. The shared pedestrian and traffic light embedder uses a cross-category feature fusion layer to enhance the model's sensitivity to different information sources. Polylines retain sequential information through positional encoding before being input to the embedders for feature extraction, followed by cross-category feature interaction via the Tamba encoder.", + "bbox": [ + 89, + 122, + 483, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The mathematical representation of this process is as follows: Given the individual polyline data $P = \\{p_1, p_2, \\ldots, p_n\\}$ , where $p_i \\in \\mathbb{R}^d$ represents a point on the polyline, $n$ is the number of points, and $d$ is the dimensionality, the embedding process is represented as: $\\mathcal{P} = \\operatorname{Embed}(P)$ . For the pedestrian and traffic light categories that share an embedder, the output feature representation $\\mathcal{P}_{\\mathrm{joint}}$ is computed as follows:", + "bbox": [ + 89, + 289, + 483, + 411 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} _ {\\text {j o i n t}} = \\text {F u s i o n} \\left(\\operatorname {E m b e d} \\left(\\mathcal {P} _ {\\text {p e d e s t r i a n}}\\right), \\operatorname {E m b e d} \\left(\\mathcal {P} _ {\\text {t r a f f i c}}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 422, + 483, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{P}_{\\mathrm{pedestrian}}$ and $\\mathcal{P}_{\\mathrm{traffic}}$ are the input polylines for pedestrians and traffic lights, respectively, and Fusion represents the feature fusion operation.", + "bbox": [ + 89, + 454, + 483, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Attention Tamba Encoder", + "text_level": 1, + "bbox": [ + 89, + 512, + 326, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the proposed encoder, the input features $\\mathcal{P}$ are uniformly embedded through joint polyline encoding. The input feature shape of the Tamba encoder is $(n,L,d)$ , where $n$ denotes the number of encoded agents or scene elements, $L$ represents the temporal sequence length, and $d$ indicates the feature dimensionality at each time step.", + "bbox": [ + 89, + 536, + 482, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, the sequence $(L,d)$ undergoes linear transformations to generate the query, key, and value matrices, each with dimensions $(L,d_{k})$ , where $d_{k}$ is the feature dimensionality used in the attention mechanism. The dot-product similarity between the query and key matrices is computed to yield an attention weight matrix of shape $(L,L)$ , which captures the interdependencies between all time steps. Subsequently, these attention weights are employed to compute a weighted sum of the value matrix, yielding an output of shape $(L,d_{k})$ .", + "bbox": [ + 89, + 628, + 483, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As the output feature dimensionality of the Tamba block $(d_k)$ may differ from the input feature dimensionality $(d)$ , a linear projection is employed to align the attention mechanism's output with the input dimensionality, resulting in $(L,d)$ . This alignment ensures compatibility for the residual connection. Subsequently, the original input and the linearly projected attention output are combined via elementwise addition, yielding a resultant tensor of shape $(L,d)$ .", + "bbox": [ + 89, + 780, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further stabilize the output, a normalization operation is applied, preserving the same dimensionality of $(L,d)$ .", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subsequently, the features are processed by a feedforward neural network comprising two fully connected layers. The first layer projects the input to a higher-dimensional space $d_{ff}$ , resulting in an output with dimensions $(L, d_{ff})$ . The second layer then maps the expanded features back to the original dimensionality $d$ , restoring the shape to $(L, d)$ . The output of the feedforward network is combined with the original input through a residual connection, preserving the output shape of $(L, d)$ . A normalization operation is then applied to ensure stability, maintaining the final output shape of $(L, d)$ for the entire encoder module. Thus, within the encoder module, the temporal sequence length $(L)$ remains constant, whereas the feature dimensionality $(d)$ may briefly change during processing but ultimately returns to its initial dimensionality.", + "bbox": [ + 511, + 122, + 903, + 347 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike traditional transformer models, Tamba adjusts the parameters of its state space model (SSM) based on the input data, allowing it to selectively retain or disregard different pieces of information within the sequence. Mathematically, Tamba uses a state space model to capture the temporal evolution of the input. Given an input sequence $\\mathcal{P} = \\{\\mathcal{P}_1,\\mathcal{P}_2,\\dots ,\\mathcal{P}_T\\}$ , where $T$ represents the sequence length and $\\mathcal{P}_t\\in \\mathbb{R}^d$ is the input at each time step $t$ , the SSM is governed by the following state evolution equation:", + "bbox": [ + 511, + 347, + 906, + 484 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} _ {t + 1} = A \\left(\\mathcal {P} _ {t}\\right) \\mathbf {h} _ {t} + B \\left(\\mathcal {P} _ {t}\\right) \\mathbf {u} _ {t}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 493, + 903, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $\\mathbf{h}_t\\in \\mathbb{R}^n$ represents the hidden state at the time step $t$ , while $A(\\mathcal{P}_t)\\in \\mathbb{R}^{n\\times n}$ and $B(\\mathcal{P}_t)\\in \\mathbb{R}^{n\\times m}$ are matrices that depend on the input and evolve over time, and $\\mathbf{u}_t\\in \\mathbb{R}^m$ represents the control input.", + "bbox": [ + 511, + 518, + 903, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The output of the time step $t$ is denoted as $\\mathbf{y}_t\\in \\mathbb{R}^P$ , and is computed as follows:", + "bbox": [ + 511, + 579, + 903, + 609 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} _ {t} = C \\left(\\mathcal {P} _ {t}\\right) \\mathbf {h} _ {t} + D \\left(\\mathcal {P} _ {t}\\right) \\mathbf {u} _ {t}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 618, + 903, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $C(\\mathcal{P}_t) \\in \\mathbb{R}^{p \\times n}$ and $D(\\mathcal{P}_t) \\in \\mathbb{R}^{p \\times m}$ are also matrices that depend on the input. This formulation allows the system to dynamically adjust its parameters based on the current input, facilitating the selective retention and forgetting of information. The computational complexity of this method is linear, as the state update and output computation at each time step involve matrix-vector multiplication. The overall complexity is $O(T \\cdot n^2)$ , where $T$ represents the sequence length and $n$ is the state dimension.", + "bbox": [ + 511, + 643, + 906, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Decomposed Attention Mechanism. We redesigned the self-attention computation, replacing the multi-head attention mechanism with state space model. After receiving the polyline encoding output $\\mathcal{P}$ , Tamba performs the attention decomposition using three parallel encoders. The detailed computation process is as follows:", + "bbox": [ + 511, + 780, + 903, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(I) Spatiotemporal attention. The spatiotemporal attention for each agent $A_{i}$ and scene elements (S) over the", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12061", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1da23a05ea8800aff8817d54e2ffb153e03c1c129cd1c3cb723953140f3db37c.jpg", + "image_caption": [ + "Figure 3. Overview of our Tamba decoding process. The outputs of the three Atten-Tamba encoders carry the context of both the static scene and dynamic agent states, which are concatenated and input into the Cross-Tamba decoder. For predicting the target agent's trajectory, the decoder queries $K$ trajectory modes using independent query tensors, while the encoder's features combine the current state with recursive reasoning from the previous step, allowing all target agents to share a unified scene representation. During the secondary decoding, the predicted state of the proposed trajectory interacts once again with the current scene information, and a recurrent network assigns prediction weights to the proposed trajectories. The final output trajectory is obtained through refinement of these predictions." + ], + "image_footnote": [], + "bbox": [ + 99, + 90, + 903, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathcal{P}$ at each time step $t$ can be represented as: $Z_{\\mathrm{time}} = \\operatorname{softmax}\\left(\\frac{Q_tK_t^\\top}{\\sqrt{d_k}}\\right)V_t$ , where $Q_{t}, K_{t}, V_{t} \\in \\mathbb{R}^{T \\times d_{k}}$ are the query, key, and value matrices at time step $t$ respectively, and $d_{k}$ represents the dimensionality of the key vectors.", + "bbox": [ + 88, + 378, + 480, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(II) Scene attention. For the attention between agent set $A$ and scene elements $S$ , let $A_{i} \\in \\mathbb{R}^{d_{a}}$ represent the embedding of agent $i$ , and $S_{j} \\in \\mathbb{R}^{d_{m}}$ represents the embedding of scene element $j$ . The attention calculation between agent $i$ and scene element $j$ is given by: $\\alpha_{ij} = \\frac{\\exp(A_i^\\top S_j)}{\\sum_k \\exp(A_i^\\top S_k)}$ , where $\\alpha_{ij}$ denotes the attention weight, indicating the relevance of scene elements $j$ to agent $i$ . The output is a weighted sum of the scene elements features denote as:", + "bbox": [ + 88, + 445, + 482, + 571 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {\\text {s c e n e}} = \\sum_ {j} \\alpha_ {i j} \\mathrm {S} _ {j}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 580, + 482, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(III) Traffic Attention. Similar to Eq. 4, we use $\\beta_{ij}$ to represent the attention weight of traffic control elements (pedestrians and traffic lights) to other dynamic agents (e.g., motorcycles and cars). For a agent $i$ , the output feature of traffic attention can be expressed as: $Z_{\\text{traffic}} = \\sum_{j} \\beta_{ij} A_j$ . This output is obtained by calculating the weighted sum of the embedded vectors $A_j$ of all dynamic agents, where the weight $\\beta_{ij}$ reflects the impact of the traffic control elements on the dynamic agents. Through this interaction, the model can effectively focus on the influence of pedestrians on dynamic agent movement in non-traffic light scenarios.", + "bbox": [ + 88, + 625, + 482, + 791 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Cross Tamba Decoder", + "text_level": 1, + "bbox": [ + 89, + 801, + 297, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The core challenge of trajectory decoding lies in addressing the 'one-to-many' problem, that is, how to generate $K$ output results (possibilities) from one set of input features. In designing the decoder architecture, we drew inspiration from DETR's approach [1] to solving the 'one-to-many' problem", + "bbox": [ + 88, + 825, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in object detection tasks. Specifically, the number of $K$ query vectors is initialized to represent different objects or candidate future trajectories. Through multiple iterations of the attention mechanism, each query vector extracts relevant information from the input features to produce $K$ output results.", + "bbox": [ + 511, + 378, + 906, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Crossed Dynamic and Static Attention. In the proposed cross-Tamba decoder, we designed it by using state space model to replace multi-head attention in the cross-attention mechanism. The query vector $Q$ is separated to interact independently with the keys and values from the encoder's output features. Specifically, the query vector is responsible for querying the modes of $K$ trajectories, while $K$ and $V$ incorporate both the encoded features from the encoder and the reasoning from the previous recursive step. This design allows dynamic agents (e.g., vehicles and motorcycles) to better share mixed features (pedestrians and traffic lights), while all target agents share a single static scene representation and iteratively generate $K$ different proposed trajectories over $T'$ future recursive steps. During the training phase, the proposed trajectories are optimized using the MSE loss ( $L_{\\text{proposal}}$ ) based on the provided labels.", + "bbox": [ + 511, + 468, + 908, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Trajectory Proposal. After proposing $K$ predicted trajectories, we introduce a prediction weight inference module to compute the weight for each trajectory. In this process, we also use the Cross-Tamba module, which first fuses scene (S) with historical information to infer the likelihood of each predicted trajectory. Specifically, given each trajectory $p_k$ (where $k = 1,2,\\dots,K$ ), we can represent it as:", + "bbox": [ + 511, + 710, + 908, + 816 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {k} = \\text {C r o s s - T a m b a} (\\text {S c e n e}, \\text {P r e d i c t i o n s}). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 828, + 903, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where Scene represents the feature representation of the current scene and Predictions contains all predicted proposal trajectories. At the head of the weight inference module, we", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12062", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "use a recurrent neural network (RNN) [24] to extract features from the final trajectories in order to generate a score for each trajectory. This RNN module can be represented as:", + "bbox": [ + 89, + 90, + 483, + 137 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nh _ {t} = \\operatorname {R N N} \\left(h _ {t - 1}, p _ {k}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 148, + 482, + 164 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $h_t$ is the hidden state at time step $t$ , $h_{t-1}$ is the previous state. Through this process, the RNN provides a mechanism for constraints across trajectories. When a trajectory conflicts with the environment or when there is an alternative trajectory that better fits the overall distribution, the RNN generates a prior confidence score for these trajectories. The confidence generated by the RNN can be represented as: $C_k = f(h_T)$ , where $C_k$ is the confidence score for trajectory $p_k$ , $h_T$ is the hidden state of the RNN at the final time step, and $f$ is a nonlinear activation function.", + "bbox": [ + 89, + 174, + 482, + 324 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This step ensures greater stability in the optimization process, such that the proposal trajectory considers not only the current prediction results but also the contextual environment and historical information, thereby enhancing the overall performance and robustness of the model.", + "bbox": [ + 89, + 325, + 483, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Trajectory Refinement. We employed the same optimization approach as the baseline methods [39, 40] to perform mixture modeling of the predicted trajectories. The future trajectory of agent $i$ is modeled as a mixture of Laplace distributions, as given by:", + "bbox": [ + 89, + 401, + 483, + 476 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nf \\left(\\left\\{p _ {t} ^ {i} \\right\\} _ {t = 1} ^ {T ^ {\\prime}}\\right) = \\sum_ {k = 1} ^ {K} \\pi_ {i, k} \\prod_ {t = 1} ^ {T ^ {\\prime}} \\operatorname {L a p l a c e} \\left(p _ {t} ^ {i} \\mid \\mu_ {t, i, k}, b _ {t, i, k}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 486, + 482, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\{\\pi_{i,k}\\}_{k = 1}^{K}$ are the mixing coefficients, representing the significance of each mixture component in the overall model. For each time step $t$ , the Laplace density of the $k^{th}$ mixture component is defined by the location parameter $\\mu t,i,k$ and the scale parameter $b_{t,i,k}$ , which reflect the center and spread of that component. We use a classification loss $L_{\\mathrm{cls}}$ to optimize the mixing coefficients in Eq. 7 by minimizing their negative log-likelihood. It is important to note that during this process, we stop the gradients of the location and scale parameters, focusing solely on the optimization of the mixing coefficients. This allows us to effectively focus on the classification task without affecting the trajectory's location and scale settings.", + "bbox": [ + 89, + 537, + 482, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For optimizing the location and scale parameters, we adopt the \"winner takes all\" strategy [15], resulting in only the best-predicted trajectory from the proposal and refinement modules undergoing backpropagation. This approach ensures that the network is adjusted only based on the best predictions during training, thereby enhancing the accuracy and stability of the model. In addition, to prevent overfitting and improve model robustness, the refinement module halts the gradient propagation of the proposed trajectory. This strategy helps to effectively preserve the previously learned information during the adjustment of proposals.", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, our loss function integrates three components: the trajectory proposal loss ( $L_{\\mathrm{proposal}}$ ), the trajectory refinement loss ( $L_{\\mathrm{refine}}$ ), and the classification loss ( $L_{\\mathrm{cls}}$ ), enabling an end-to-end training process. The expression is given as:", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {t o t a l}} = L _ {\\text {p r o p o s a l}} + L _ {\\text {r e f i n e}} + \\lambda L _ {\\text {c l s}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 166, + 903, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 210, + 645, + 228 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 513, + 236, + 614, + 251 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our experiments were conducted on the Argoverse 1 [2] and Argoverse 2 [14, 33] datasets. It contains extensive real-world autonomous driving scenarios and are designed to advance the development of autonomous driving technologies. The Argoverse 1 dataset includes 360-degree sensor data from the Pittsburgh and Miami regions in the United States, covering various road types and complex urban environments. Sampled at a rate of $10\\mathrm{Hz}$ , this dataset provides trajectory data of the ego vehicle over a 5-second time span, with 2 seconds of observation used to predict the path over the next 3 seconds, comprising over 30,000 such samples. Argoverse 2 further expands on the dataset by including richer sensor data, multimodal annotations, and extending the observation window. Specifically, it includes 11-second scenarios, with 5 seconds of observation followed by prediction of trajectories for the next 6 seconds. The dataset contains 250,000 scenarios and features a large object taxonomy comprising 10 non-overlapping classes, which is used for training and validating motion prediction models.", + "bbox": [ + 511, + 260, + 906, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Metrics", + "text_level": 1, + "bbox": [ + 513, + 559, + 609, + 573 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We adopted evaluation methods consistent with prior research, using metrics such as $\\min\\mathrm{ADE}(K)$ , $\\min\\mathrm{FDE}(K)$ , $\\min\\mathrm{FDE}(K)$ , and $\\mathrm{MR}(K)$ to measure model performance in terms of error. Specifically, $\\min\\mathrm{ADE}(K)$ calculates the average $L2$ distance between the ground truth and the most optimal trajectory among the $K$ predicted trajectories across all future time steps, thereby assessing the overall prediction accuracy of the model over multiple time steps. $\\min\\mathrm{FDE}(K)$ focuses on the prediction error at the final time step, measuring the model's ability to predict the final state accurately. $\\mathrm{b-minFDE}(K)$ extends $\\min\\mathrm{FDE}(K)$ by incorporating probability weighting, using $(1 - \\hat{\\pi})^2$ as a weight to account for prediction uncertainty, where $\\hat{\\pi}$ represents the probability score assigned by the model to the optimal predicted trajectory. $\\mathrm{MR}(K)$ measures the proportion of the $K$ predicted trajectories with an error greater than 2 meters, providing an assessment of the model's stability and the frequency of prediction failures. Together, these metrics offer a comprehensive evaluation of the model's prediction accuracy, stability, and uncertainty estimation. For the parameter $K$ , we followed conventional values of 6 and 1. When the", + "bbox": [ + 511, + 583, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12063", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/310c9b50766782ae433c923c83259ff419a010902f9ecbb0b1d43da51381dcd0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methodb-minFDE6↓minADE6↓minFDE6↓MR6↓minADE1↓minFDE1↓MR1↓#Params (M)↓FLOPs(G)↓
THOMAS [6]2.160.881.510.201.954.710.64--
GoRela [4]2.010.761.480.221.824.620.66--
QML [28]1.950.691.390.191.844.980.629.39-
MTR [26]1.980.731.440.151.744.390.5865.78-
GANet [31]1.960.721.340.171.774.480.5961.7315.79
BANet [32]1.920.711.360.191.794.610.609.4911.93
QCNet [40]1.910.651.290.161.694.300.597.6645.3
Tamba (Ours)1.890.641.240.171.664.240.574.5427.3
", + "bbox": [ + 99, + 88, + 898, + 229 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b322e2bfd7020233409f4733ab350f92268f361338a1ac7e2a1d7a1bc9a566da.jpg", + "table_caption": [ + "Table 1. Performance comparison on Argoverse 2 [14, 33] dataset, leaderboard ranked by b-minFDE6. Subscript 6 and 1 represents the number of predictions. Best performance is highlighted in bold. (M) represents the number of parameters in millions; (G) refers to GigaOps per second. - represents we were unable to reproduce the results with complete accuracy. Notably, all methods in the table are direct comparisons of the original approaches. We exclude additional improvements from engineering techniques such as ensembling." + ], + "table_footnote": [], + "table_body": "
Methodb-minFDE6↓minADE6↓minFDE6↓MR6↓#Params (M)↓
LaneGCN [16]2.060.871.360.16-
DenseTNT [10]1.980.881.280.13-
MTR [26]1.930.821.240.13-
SceneTransformer [22]1.890.801.230.1315.30
HiVT [39]1.840.771.170.13-
GANet [31]1.790.811.160.12-
Wayformer [21]1.740.771.160.12-
QCNet [40]1.690.731.070.117.66
Tamba (Ours)1.670.721.070.094.54
", + "bbox": [ + 205, + 309, + 794, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Quantitative results on the Argoverse 1 [2] dataset, leaderboard ranked by b-minFDE6. Subscript 6 represents the number of predictions. Best performance is highlighted in bold. (M) represents the number of parameters in millions. - represents it is unable to reproduce the results with complete accuracy. We exclude additional improvements from engineering techniques such as ensembling.", + "bbox": [ + 89, + 478, + 906, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "number of predicted trajectories exceeds $K$ , the top $K$ trajectories with the highest probabilities are selected. In terms of lightweight evaluation, we used the number of parameters (M) and FLOPs(G) as the metrics.", + "bbox": [ + 89, + 546, + 482, + 607 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Settings", + "text_level": 1, + "bbox": [ + 89, + 617, + 189, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted all experiments under the same environment (batch size, epochs) using eight NVIDIA 3090TI cards with a batch size of 128. We used Adam for optimization, with an initial learning rate of 0.001. Training was conducted for a total of 50 epochs. We employed an adaptive learning rate adjustment strategy based on the validation set's performance during training. Specifically, when the validation accuracy did not improve for five consecutive epochs, the learning rate was reduced to 0.1 times its original value.", + "bbox": [ + 89, + 640, + 483, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Comparison with State of the Art", + "text_level": 1, + "bbox": [ + 89, + 786, + 383, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our method with other SOTA approaches across various evaluation metrics in Argoverse 2 dataset. As shown in Table 1, our method demonstrates superior performance on key metrics such as $\\mathrm{minFDE}_6$ and $\\mathrm{minADE}_6$ . Tamba achieves a b-minFDE $^6$ score of 1.79, compared to 1.91 for QCNet, 1.92 for BANet and 1.98 for MTR. Additionally,", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tamba achieves $(\\mathrm{MR}_1)$ of 0.57, which is lower than that of other methods, such as QCNet (0.59) and MTR (0.58), indicating more stable prediction accuracy. Furthermore, Tamba has a parameter count of $4.54\\mathrm{M}$ , which is significantly lower than QCNet (7.66 M), even compare to non-transformer models such as BANet (9.49M) and GANet (61.73M), Tamba achieving reductions of $52.0\\%$ and $92.7\\%$ respectively, while still outperforming both models. Additionally, Tamba achieves 27.3 GigaOps per second, which is $40\\%$ faster than QCNet's 45.3 GigaOps per second. These results clearly demonstrate that Tamba is competitive in terms of accuracy but also excels in lightweight design and efficiency, providing an excellent solution for complex task.", + "bbox": [ + 511, + 546, + 906, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On the Argoverse 1 dataset, our method outperforms other state-of-the-art (SOTA) methods across the board. On the leaderboard for the b-minFDE $_6$ metric, Tamba achieves a score of 1.67, while the second-best, QCNet, scores 1.69, and the third, BANet, scores 1.74. In terms of parameter count, our method requires 4.54M parameters, compared to 7.66M for QCNet, reflecting a $40\\%$ reduction. This further demonstrates the superiority of our approach in balancing efficiency and performance. For detailed performance metrics, please refer to Table 2.", + "bbox": [ + 511, + 750, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12064", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/381441827755affb18d307400dc2c79b778ec26edde955583934b2a4704f6db6.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 94, + 85, + 496, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/67ae8c454d91ebbbc9766677cd149d3b8f701f1d8c8f2562146f31e7bfe5b2f7.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 496, + 85, + 898, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bf88a6e27dabc9abd2ddb487b13a6e9f589d77a2b3662b782e436c87ea8f87c1.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 96, + 213, + 495, + 318 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d03ef04f7f3755592c01273a099274afdbc85c935acf9d18c898298324f99d8f.jpg", + "image_caption": [ + "Observation", + "Target agent" + ], + "image_footnote": [], + "bbox": [ + 496, + 213, + 898, + 343 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dd347f734b1df551ffceb00f774dfb41a87277e92dc8d5b48b6c98cd0fbb2aa0.jpg", + "image_caption": [ + "Predicted trajectory" + ], + "image_footnote": [], + "bbox": [ + 99, + 321, + 207, + 363 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/36ba9fb57a9ebb7a9c6ffbba7d1b10f74126a40921077a71e6a0d423e6a28bc0.jpg", + "image_caption": [ + "Figure 4. Qualitative results on the Argoverse 2 validation set. The target agent marked as orange with its surrounding agents marked as cold white. We conduct four different traffic scenarios. (a). Multi-agents in straight-road scenario, (b). Multiple agents on roundabout road, (c). Vehicle avoidance after roundabout, (d). Mixed scenario involving pedestrians and vehicles." + ], + "image_footnote": [], + "bbox": [ + 692, + 345, + 736, + 362 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/549d3007e97c93e528411c640f0113b080755a5da0cb5aaa91931ad622d46470.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methodw/ jointw/o joint
minFDE6↓minADE6↓minFDE1↓minADE1↓minFDE6↓minADE6↓minFDE1↓minADE1↓
Ours + Attention1.0940.7633.0371.3761.0930.7283.0351.372
Ours + Mamba1.0910.7242.8991.2981.0830.7652.8211.292
Ours + Tamba0.9510.6732.7421.2650.9550.6802.7451.272
", + "bbox": [ + 153, + 431, + 844, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation study on the components of the Tamba encoder-decoder. Experimental results are conducted on the Argoverse 2 [14, 33] validation set.", + "bbox": [ + 89, + 523, + 906, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Quantization Results", + "text_level": 1, + "bbox": [ + 89, + 578, + 289, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We visualized the prediction results on the validation set. In some challenging sequences, the predictions made by the Tamba model are more reasonable and smoother compared to existing methods, effectively adhering to map constraints. In Figure 4, we present several multimodal prediction cases, the results demonstrate that the trajectories generated by Tamba are smoother and better adapted to complex road topologies, meeting the demands of vehicle navigation under diverse road conditions.", + "bbox": [ + 89, + 602, + 483, + 737 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 753, + 241, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluated the effectiveness of the Tamba module. For comparison, we used both the traditional attention mechanism module and the Mamba module on the validation set. In addition, we assessed the efficacy of our proposed pedestrian-traffic light joint polyline encoding strategy. Table 3 presents the performance comparison of our proposed three methods (including Attention variant, Mamba variant, and Tamba variant) under conditions with and without joint", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "polyline encoding (w/ joint vs. w/o joint). Specifically, the table lists the performance of each method on two evaluation metrics: minFDE and minADE, considering both six predicted trajectories (minFDE $_6$ and minADE $_6$ ) and one predicted trajectory (minFDE $_1$ and minADE $_1$ ).", + "bbox": [ + 511, + 579, + 905, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 686, + 633, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces the Tamba model, an efficient motion prediction model for the real-time and computational needs of autonomous driving. By employing a state space model (SSM) and a novel attention Tamba encoder-decoder, we reduce computational complexity and parameter count while maintaining state-of-the-art accuracy. Experiments on the Argoverse 1 and 2 datasets show that Tamba achieves superior inference speed and efficiency, with a fourfold reduction in FLOPs and over $40\\%$ fewer parameters compared to transformer-based models. These advantages make Tamba highly suitable for high-dynamic environments requiring rapid responses.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12065", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported in part by the Europe Eureka Intelligence to Drive | Move-Save-Win project (with funding from the UKRI Innovate UK project under Grant No. 10071278) as well as the Horizon Europe COVER project, No. 101086228 (with funding from UKRI grant EP/Y028031/1). Kezhi Wang would like to acknowledge the support in part by the Royal Society Industry Fellow scheme.", + "bbox": [ + 89, + 113, + 483, + 224 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 237, + 187, + 253 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision (ECCV), pages 213-229. Springer, 2020. 5", + "[2] Ming-Fang Chang, John W Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, and James Hays. Argoverse: 3d tracking and forecasting with rich maps. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6, 7", + "[3] Yihua Cheng, Yaning Zhu, Zongji Wang, Hongquan Hao, Yongwei Liu, Shiqing Cheng, Xi Wang, and Hyung Jin Chang. What do you see in vehicle? comprehensive vision solution for in-vehicle gaze estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1", + "[4] Alexander Cui, Sergio Casas, Kelvin Wong, Simon Suo, and Raquel Urtasun. Gorela: Go relative for viewpoint-invariant motion forecasting. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 7801-7807. IEEE, 2023. 7", + "[5] Jiyang Gao, Chen Sun, Hang Zhao, Yi Shen, Dragomir Anguelov, Congcong Li, and Cordelia Schmid. Vectornet: Encoding hd maps and agent dynamics from vectorized representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11525-11533, 2020. 1, 2", + "[6] Thomas Gilles, Stefano Sabatini, Dzmitry Tsishkou, Bogdan Stanciulescu, and Fabien Moutarde. Thomas: Trajectory heatmap output with learned multi-agent sampling. In International Conference on Learning Representations (ICLR), 2022. 7", + "[7] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2", + "[8] Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. Hippo: Recurrent memory with optimal polynomial projections. Advances in neural information processing systems, 33:1474-1487, 2020. 2", + "[9] Albert Gu, Karan Goel, Ankit Gupta, and Christopher Ré. On the parameterization and initialization of diagonal state space models. Advances in Neural Information Processing Systems, 35:35971-35983, 2022. 2", + "[10] Junru Gu, Chen Sun, and Hang Zhao. Densetnt: End-to-end trajectory prediction from dense goal sets. In Proceedings of" + ], + "bbox": [ + 93, + 261, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF International Conference on Computer Vision, pages 15303-15312, 2021. 7", + "[11] Ankit Gupta, Albert Gu, and Jonathan Berant. Diagonal state spaces are as effective as structured state spaces. Advances in Neural Information Processing Systems, 35:22982-22994, 2022. 2", + "[12] Yizhou Huang, Yihua Cheng, and Kezhi Wang. Efficient driving behavior narration and reasoning on edge device using large language models. arXiv preprint arXiv:2409.20364, 2024. 1", + "[13] Parth Kothari, Brian Sifringer, and Alexandre Alahi. Interpretable social anchors for human trajectory forecasting in crowds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15556-15566, 2021. 3", + "[14] John Lambert and James Hays. Trust, but verify: Cross-modality fusion for hd map change detection. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks (NeurIPS Datasets and Benchmarks 2021), 2021. 6, 7, 8", + "[15] Stefan Lee, Senthil Purushwalkam Shiva Prakash, Michael Cogswell, Viresh Ranjan, David Crandall, and Dhruv Batra. Stochastic multiple choice learning for training diverse deep ensembles. Advances in Neural Information Processing Systems, 29, 2016. 6", + "[16] Ming Liang, Bin Yang, Rui Hu, Yun Chen, Renjie Liao, Song Feng, and Raquel Urtasun. Learning lane graph representations for motion forecasting. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 541-556. Springer, 2020. 1, 7", + "[17] Mengmeng Liu, Hao Cheng, Lin Chen, Hellward Broszio, Jiangtao Li, Runjiang Zhao, Monika Sester, and Michael Ying Yang. Laformer: Trajectory prediction for autonomous driving with lane-aware scene constraints. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2039-2049, 2024. 1, 3", + "[18] Yuejiang Liu, Qi Yan, and Alexandre Alahi. Social nce: Contrastive learning of socially-aware motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15118-15129, 2021. 2", + "[19] Karttikeya Mangalam, Yang An, Harshayu Girase, and Jitendra Malik. From goals, waypoints & paths to long term human trajectory forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15233-15242, 2021. 3", + "[20] Abduallah Mohamed, Kun Qian, Mohamed Elhoseiny, and Christian Claudel. Social-stgCNN: A social spatio-temporal graph convolutional neural network for human trajectory prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14424-14432, 2020. 2", + "[21] Nigamaa Nayakanti, Rami Al-Rfou, Aurick Zhou, Kratarth Goel, Khaled S Refaat, and Benjamin Sapp. Wayformer: Motion forecasting via simple & efficient attention networks. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2980-2987. IEEE, 2023. 1, 7" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12066", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Jiquan Ngiam, Vijay Vasudevan, Benjamin Caine, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, et al. Scene transformer: A unified architecture for predicting future trajectories of multiple agents. In 2023 IEEE International Conference on Robotics and Automation (ICRA), 2022. 7", + "[23] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory++: Dynamically-feasible trajectory forecasting with heterogeneous data. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVIII 16, pages 683-700. Springer, 2020. 2", + "[24] Alex Sherstinsky. Fundamentals of recurrent neural network (rnn) and long short-term memory (lstm) network. *Physical D: Nonlinear Phenomena*, 404:132306, 2020. 6", + "[25] Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Wei Tang, Nanning Zheng, and Gang Hua. Representing multimodal behaviors with mean location for pedestrian trajectory prediction. IEEE transactions on pattern analysis and machine intelligence, 2023. 3", + "[26] Shaoshuai Shi, Li Jiang, Dengxin Dai, and Bernt Schiele. Mtr++: Multi-agent motion prediction with symmetric scene modeling and guided intention querying. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3, 7", + "[27] Jimmy Smith, Shalini De Mello, Jan Kautz, Scott Linderman, and Wonmin Byeon. Convolutional state space models for long-range spatiotemporal modeling. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[28] Tong Su, Xishun Wang, and Xiaodong Yang. Qml for argoverse 2 motion forecasting challenge. arXiv preprint arXiv:2207.06553, 2022. 7", + "[29] Balakrishnan Varadarajan, Ahmed Hefny, Avikalp Srivastava, Khaled S Refaat, Nigamaa Nayakanti, Andre Cornman, Kan Chen, Bertrand Douillard, Chi Pang Lam, Dragomir Anguelov, et al. Multipath++: Efficient information fusion and trajectory aggregation for behavior prediction. In 2022 International Conference on Robotics and Automation (ICRA), pages 7814-7821. IEEE, 2022. 1", + "[30] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2", + "[31] Mingkun Wang, Xinge Zhu, Changqian Yu, Wei Li, Yuexin Ma, Ruochun Jin, Xiaoguang Ren, Dongchun Ren, Mingxu Wang, and Wenjing Yang. Ganet: Goal area network for motion forecasting. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1609-1615. IEEE, 2023. 7", + "[32] Zhepeng Wang, Feng Chen, Kanokphan Lertniphonphan, Siwei Chen, Jinyao Bao, Pengfei Zheng, Jinbao Zhang, Kaer Huang, and Tao Zhang. Technical report for argoverse challenges on unified sensor-based detection, tracking, and forecasting. arXiv preprint arXiv:2311.15615, 2023. 7", + "[33] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes," + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks (NeurIPS Datasets and Benchmarks 2021), 2021. 6, 7, 8", + "[34] Conghao Wong, Beihao Xia, Ziqian Zou, Yulong Wang, and Xinge You. Socialcircle: Learning the angle-based social interaction representation for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19005-19015, 2024. 3", + "[35] Maosheng Ye, Tongyi Cao, and Qifeng Chen. Tpcn: Temporal point cloud networks for motion forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11318-11327, 2021. 2", + "[36] Ye Yuan, Xinshuo Weng, Yanglan Ou, and Kris M Kitani. Agentformer: Agent-aware transformers for socio-temporal multi-agent forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9813-9823, 2021. 2", + "[37] Zeyu Zhang, Akide Liu, Ian Reid, Richard Hartley, Bohan Zhuang, and Hao Tang. Motion mamba: Efficient and long sequence motion generation. In European Conference on Computer Vision, pages 265-282. Springer, 2025. 2", + "[38] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634-21643, 2024. 1", + "[39] Zikang Zhou, Luyao Ye, Jianping Wang, Kui Wu, and Kejie Lu. Hivt: Hierarchical vector transformer for multi-agent motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8823-8833, 2022. 1, 6, 7", + "[40] Zikang Zhou, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang. Query-centric trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17863-17873, 2023. 1, 3, 6, 7", + "[41] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 696 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12067", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Trajectory Mamba_ Efficient Attention-Mamba Forecasting Model Based on Selective SSM/14500b01-a80f-4244-a6f9-1c844f2302da_model.json b/2025/Trajectory Mamba_ Efficient Attention-Mamba Forecasting Model Based on Selective SSM/14500b01-a80f-4244-a6f9-1c844f2302da_model.json new file mode 100644 index 0000000000000000000000000000000000000000..34526bcd2a65f38aa6844e7d6e3fd9509d5cfa4d --- /dev/null +++ b/2025/Trajectory Mamba_ Efficient Attention-Mamba Forecasting Model Based on Selective SSM/14500b01-a80f-4244-a6f9-1c844f2302da_model.json @@ -0,0 +1,1991 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.116, + 0.13, + 0.885, + 0.175 + ], + "angle": 0, + "content": "Trajectory Mamba: Efficient Attention-Mamba Forecasting Model Based on Selective SSM" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.203, + 0.752, + 0.275 + ], + "angle": 0, + "content": "Yizhou Huang\\(^{1}\\), Yihua Cheng\\(^{2}\\), Kezhi Wang\\(^{1*}\\) \n\\(^{1}\\)Brunel University of London, \\(^{2}\\)University of Birmingham \n{byyh009, Kezhi.Wang}@brunel.ac.uk, y.cheng.2@bham.ac.uk \nhttps://github.com/Yizhou-H/Trajectory-Mamba-CVPR" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.31, + 0.327, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.485, + 0.782 + ], + "angle": 0, + "content": "Motion prediction is crucial for autonomous driving, as it enables accurate forecasting of future vehicle trajectories based on historical inputs. This paper introduces Trajectory Mamba, a novel efficient trajectory prediction framework based on the selective state-space model (SSM). Conventional attention-based models face the challenge of computational costs that grow quadratically with the number of targets, hindering their application in highly dynamic environments. In response, we leverage the SSM to redesign the self-attention mechanism in the encoder-decoder architecture, thereby achieving linear time complexity. To address the potential reduction in prediction accuracy resulting from modifications to the attention mechanism, we propose a joint polyline encoding strategy to better capture the associations between static and dynamic contexts, ultimately enhancing prediction accuracy. Additionally, to balance prediction accuracy and inference speed, we adopted the decoder that differs entirely from the encoder. Through cross-state space attention, all target agents share the scene context, allowing the SSM to interact with the shared scene representation during decoding, thus inferring different trajectories over the next prediction steps. Our model achieves state-of-the-art results in terms of inference speed and parameter efficiency on both the Argoverse 1 and Argoverse 2 datasets. It demonstrates a four-fold reduction in FLOPs compared to existing methods and reduces parameter count by over \\(40\\%\\) while surpassing the performance of the vast majority of previous methods. These findings validate the effectiveness of Trajectory Mamba in trajectory prediction tasks." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.808, + 0.222, + 0.823 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.833, + 0.483, + 0.879 + ], + "angle": 0, + "content": "In the rapid development of autonomous driving technology [3, 12, 17], accurate motion prediction plays a crucial role by forecasting future trajectories based on historical vehicle" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.31, + 0.905, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.502, + 0.907, + 0.586 + ], + "angle": 0, + "content": "Figure 1. Illustration of proposed joint polyline encoding strategy, where we consider all factors that affect the movement of motor vehicles, jointly encoding pedestrians and traffic lights rather than categorizing the scene horizontally into static and dynamic context. Additionally, we decompose and encode the interactions between all agents and elements at each time step." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.765 + ], + "angle": 0, + "content": "movement data. This task is not only vital for enhancing driving safety but also meets the stringent requirements for real-time system response. In recent years, learning-based approaches have been widely applied to motion prediction [5, 16, 29, 39, 40]. However, existing prediction methods often struggle to balance efficiency and accuracy, making it challenging to simultaneously meet performance and efficiency requirements in real-time applications. We attempt to analyze the underlying reasons from the following aspects and explore possible improvements:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.909, + 0.902 + ], + "angle": 0, + "content": "(a). Improving Prediction Accuracy: In autonomous driving scenarios [38], there are numerous dynamic and static agents, and the diversity of this data presents significant challenges for motion prediction and model training. However, conventional methods [29, 39] often struggle to effectively fuse this heterogeneous data, particularly when it comes to integrating various features, which leads to suboptimal prediction accuracy. On a positive note, factorized attention-based transformers [21, 40] have made no" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.236, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12058" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.481, + 0.166 + ], + "angle": 0, + "content": "table strides in decomposing heterogeneous data. However, these models often use a unified, horizontal approach when dealing with static scenes and dynamic agent features, and they fail to fully consider the deeper relationships between different scene elements and agent information." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.17, + 0.482, + 0.395 + ], + "angle": 0, + "content": "(b). Enhancing Response Efficiency: Motion predictions are inherently recursive, implying that each prediction serves as an input for the next time step, leading to the progressive accumulation of errors over time. Notably, the advancement of sparse context encoding techniques [5, 35] has demonstrated significant potential in enhancing model efficiency. Nonetheless, prediction models based on attention mechanisms still face exponential growth in computational complexity and parameter size during recursive tasks due to inherent design constraints, which severely limit their efficiency and real-time capabilities in practical applications. Conversely, methods such as LSTM [18] are relatively more efficient, but numerous studies [20, 23, 36] have shown that they are incapable of achieving the predictive accuracy of transformer-based models." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.399, + 0.482, + 0.669 + ], + "angle": 0, + "content": "In response, we propose an Attention-Mamba framework for motion prediction, termed Trajectory Mamba (Tamba), to overcome the limitations discussed above. First, we design a multimodal encoder-decoder architecture based on the state space model [9, 11, 27], comprising three parallel encoders and a decoder. By introducing a selective state space, we redefined the computation of the self-attention mechanism, enabling the model to selectively process input information by parameterizing the input of the state space model (SSM), thereby allowing the model to focus on or ignore specific inputs. Specifically, the selective state space module processes local short-range features by receiving input from a one-dimensional convolutional layer, ensuring that these features can be selected or discarded before entering the SSM. We utilized local state space attention in place of the multi-head attention in the self-attention module, reducing the computational complexity of recursive reasoning from exponential to linear." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.673, + 0.482, + 0.883 + ], + "angle": 0, + "content": "Secondly, to better integrate heterogeneous data features, we proposed a joint polyline encoding strategy. Its objective is to separate pedestrian agents from the dynamic agents and jointly encode them with traffic light information from the static scene (see Fig. 1). We observed that in uncertain motion prediction scenarios, the application of traffic regulations exerts a degree of implicit control over the potential movement states of agents. For example, pedestrian-priority traffic rules remain effective even in the absence of traffic lights, influencing vehicles and motorcycles alike. By jointly encoding pedestrians and traffic lights through a shared embedder, this jointly encoding strategy can independently interact with other dynamic agents to enhance the prediction accuracy of vehicle and motorcycle agents." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.887, + 0.481, + 0.9 + ], + "angle": 0, + "content": "Lastly, we designed a prediction weight inference mod" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.092, + 0.904, + 0.198 + ], + "angle": 0, + "content": "ule. By redesigning selective state space with a cross-attention mechanism, we utilized independent query tensors to recursively infer multimodal candidate trajectories through interactions with the encoder and the keys and values from the previous recursive step. These candidate trajectories were then evaluated using a recurrent neural network to assign trajectory scores." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.2, + 0.904, + 0.514 + ], + "angle": 0, + "content": "Our contributions are summarized as: (1) We propose the Tamba framework for motion prediction, which leverages the state space model to redesign computations of selfattention, resulting in computational complexity reducing significantly. (2) We propose a joint polyline encoding strategy. By integrating strongly correlated polyline types using independent or shared embeddings, elements that impose similar behavioral patterns can better influence different agents, collectively contributing to the trajectory reasoning of these agents. (3) We redesign the Mamba decoder with crossattention, allowing the joint scene encoding to share a unified scene representation, which further reduces the model's complexity. At the final stage, we introduced a prediction weight inference module to refine the predicted trajectories and enhance prediction accuracy. (4) Experimental results demonstrate that our method achieves state-of-the-art (SOTA) performance on Argoverse 1 and 2 datasets. Compared to advanced methods, our method reduces the number of parameters by \\(40\\%\\), achieves a four-fold increase in FLOPs. This represents a new level of balance between efficiency and performance in our model." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.533, + 0.661, + 0.548 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.559, + 0.691, + 0.574 + ], + "angle": 0, + "content": "2.1. State Space Model" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.583, + 0.904, + 0.824 + ], + "angle": 0, + "content": "Recently, the Mamba [7] framework has rekindled interest in state-space models (SSMs) [9, 11, 27] due to its introduction of an input-dependent selection mechanism, which enhances their applicability and performance. SSMs are renowned for their capacity to handle long-range dependencies and offer superior memory and computational efficiency compared to transformer-based architectures [30]. Hippo [8], a seminal contribution to SSMs, established the efficacy of these models in long-range sequence modeling tasks. Subsequent research, such as Vision mamba [41], has further refined the mechanisms underlying SSMs, resulting in substantial performance improvements. SSMs have demonstrated their versatility and robustness across a diverse set of domains, including computer vision, video analysis, medical imaging, graph-based learning, point cloud processing, and recommendation systems." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.826, + 0.904, + 0.899 + ], + "angle": 0, + "content": "In the domain of trajectory prediction, recent studies have indicated that Motion Mamba [37] achieves notable success in predicting human trajectory time series. Building on these advancements, this work integrates the state-space mechanisms of Mamba with attention mechanisms to address tra" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.947, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12059" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.137 + ], + "angle": 0, + "content": "jectory prediction tasks in dynamic traffic scenarios, thereby leveraging both the efficiency of SSMs and the adaptability of attention-based models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.147, + 0.423, + 0.163 + ], + "angle": 0, + "content": "2.2. Attention Trajectory Prediction Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.171, + 0.483, + 0.276 + ], + "angle": 0, + "content": "In the field of trajectory prediction, methods [13, 19] based on the attention mechanism have been widely adopted to enhance the model's ability to capture and predict complex behaviors. These approaches leverage the capability of attention to focus selectively on important parts of the input, thereby capturing key temporal and spatial features that are crucial for predicting future trajectories." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.277, + 0.483, + 0.428 + ], + "angle": 0, + "content": "Subsequently, several advanced approaches [17, 34] have utilized attention mechanisms for trajectory prediction specifically in multi-agent interaction environments, where understanding the interactions between different agents (such as vehicles, pedestrians, and other road users) is critical. These models employ social attention modules to focus on the movements and behaviors of surrounding vehicles, effectively improving the ability of the system to handle intricate interactive behaviors, which are a common aspect of real-world driving scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.429, + 0.483, + 0.565 + ], + "angle": 0, + "content": "Moreover, attention mechanisms have also been successfully applied in models that integrate multimodal information [25, 26, 40]. These models dynamically weight and fuse different types of sensory inputs—such as image features from cameras and data from other sensors like LiDAR—through attention modules. This fusion process allows the model to better understand and represent complex traffic environments, effectively utilizing complementary data to enhance trajectory prediction performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.578, + 0.221, + 0.594 + ], + "angle": 0, + "content": "3. Our Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.604, + 0.201, + 0.618 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.795 + ], + "angle": 0, + "content": "In the autonomous driving scenario, the input consists of the predicted dynamic agent, its surrounding agents, and the static scene context. Let the historical state of the \\(i\\)-th agent at time step \\(t\\) be represented as: \\(s_i^t = (p_i^t,\\theta_i^t,t,\\nu_i^t)\\), where \\(p_i^t = (p_{i,x}^t,p_{i,y}^t)\\) denotes the spatial position, \\(\\theta_i^t\\) represents the heading angle, \\(t\\) refers to the time step, and \\(\\nu_{i}^{t}\\) denotes the velocity. The static scene context is represented by \\(n\\) polygons, which include elements such as lane boundaries, map edges, and sidewalks. The agent states \\(s_i^t\\) over \\(T\\) time steps within the observation window are used to predict \\(K\\) future trajectories over \\(T^{\\prime}\\) time steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose the Tamba model to optimize the efficiency and accuracy of motion forecasting tasks. Our input includes \\( s_i^t \\) of dynamic and static agents, \\( \\mathcal{P} \\) processed using a heterogeneous polyline encoding strategy to transform \\( s_i^t \\) into higher-dimensional features. We leverage three Tamba encoders to separately decompose (I) attendations in each time step \\( (t) \\) over all elements, (II) attendations between" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.291, + 0.907, + 0.376 + ], + "angle": 0, + "content": "Figure 2. Overview of our Tamba encoder. We employ joint polyline encoding strategy to integrate strongly correlated polyline information and used three parallel encoders to interact with and associate these features. We applied a linear projection to map the attention output of Tamba back to the same dimensions as the input and used normalization to enhance the stability of the output." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.4, + 0.907, + 0.552 + ], + "angle": 0, + "content": "all agents (A) and scene (S), and (III) attentions of traffic control elements (pedestrians and traffic lights) to other dynamic agents (motorcycles and cars). Then, concatenate features from encoders as output to a Tamba decoder, where we utilize cross-attention mechanism to generate proposal trajectories. Subsequently, we fused scene encoding with historical information to perform a secondary decoding of the proposal trajectories for refinement. Finally, the refined trajectories were optimized using a \"winner-takes-all strategy\" to obtain trajectory output." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.906, + 0.643 + ], + "angle": 0, + "content": "Our model contains a novel heterogeneous polyline encoding strategy along with a decomposed feature fusion method for both static and dynamic agents. A redesigned Tamba encoder-decoder architecture. A refinement process to output the final predicted trajectory. In the following sections, we sequentially introduce the components in Tamba." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.651, + 0.733, + 0.668 + ], + "angle": 0, + "content": "3.2. Joint Polyline Encoding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In multimodal traffic environment prediction tasks, different types of polylines possess unique geometric characteristics and semantic information. This study proposes a novel heterogeneous polyline embedding method to facilitate the fusion of dynamic and static context features, thereby providing a richer and more efficient feature representation. Specifically, we assign either independent or shared embedders to each type of polyline to effectively model various categories of information. Specifically, we jointly encode polylines with strong information associations, such as pedestrian trajectories and traffic light signals, using a shared embedder. This is because agents (pedestrians) and traffic lights typically impose similar behavioral patterns on other agents (vehicles) and often participate together in traffic control. Thus, using a shared embedder reduces model complexity" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12060" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "and promotes feature sharing, collectively influencing vehicle trajectory reasoning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.123, + 0.484, + 0.288 + ], + "angle": 0, + "content": "In contrast, independent embedders are used for polyline types with significant feature differences, such as lane lines and traffic signs in maps. Each embedder consists of a MLP layer and normalization layers to extract high-dimensional feature representations. The shared pedestrian and traffic light embedder uses a cross-category feature fusion layer to enhance the model's sensitivity to different information sources. Polylines retain sequential information through positional encoding before being input to the embedders for feature extraction, followed by cross-category feature interaction via the Tamba encoder." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.29, + 0.484, + 0.412 + ], + "angle": 0, + "content": "The mathematical representation of this process is as follows: Given the individual polyline data \\( P = \\{p_1, p_2, \\ldots, p_n\\} \\), where \\( p_i \\in \\mathbb{R}^d \\) represents a point on the polyline, \\( n \\) is the number of points, and \\( d \\) is the dimensionality, the embedding process is represented as: \\( \\mathcal{P} = \\operatorname{Embed}(P) \\). For the pedestrian and traffic light categories that share an embedder, the output feature representation \\( \\mathcal{P}_{\\mathrm{joint}} \\) is computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.424, + 0.484, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} _ {\\text {j o i n t}} = \\text {F u s i o n} \\left(\\operatorname {E m b e d} \\left(\\mathcal {P} _ {\\text {p e d e s t r i a n}}\\right), \\operatorname {E m b e d} \\left(\\mathcal {P} _ {\\text {t r a f f i c}}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.484, + 0.502 + ], + "angle": 0, + "content": "where \\(\\mathcal{P}_{\\mathrm{pedestrian}}\\) and \\(\\mathcal{P}_{\\mathrm{traffic}}\\) are the input polylines for pedestrians and traffic lights, respectively, and Fusion represents the feature fusion operation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.513, + 0.328, + 0.527 + ], + "angle": 0, + "content": "3.3. Attention Tamba Encoder" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.483, + 0.627 + ], + "angle": 0, + "content": "In the proposed encoder, the input features \\(\\mathcal{P}\\) are uniformly embedded through joint polyline encoding. The input feature shape of the Tamba encoder is \\((n,L,d)\\), where \\(n\\) denotes the number of encoded agents or scene elements, \\(L\\) represents the temporal sequence length, and \\(d\\) indicates the feature dimensionality at each time step." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.78 + ], + "angle": 0, + "content": "Specifically, the sequence \\((L,d)\\) undergoes linear transformations to generate the query, key, and value matrices, each with dimensions \\((L,d_{k})\\), where \\(d_{k}\\) is the feature dimensionality used in the attention mechanism. The dot-product similarity between the query and key matrices is computed to yield an attention weight matrix of shape \\((L,L)\\), which captures the interdependencies between all time steps. Subsequently, these attention weights are employed to compute a weighted sum of the value matrix, yielding an output of shape \\((L,d_{k})\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.902 + ], + "angle": 0, + "content": "As the output feature dimensionality of the Tamba block \\((d_k)\\) may differ from the input feature dimensionality \\((d)\\), a linear projection is employed to align the attention mechanism's output with the input dimensionality, resulting in \\((L,d)\\). This alignment ensures compatibility for the residual connection. Subsequently, the original input and the linearly projected attention output are combined via elementwise addition, yielding a resultant tensor of shape \\((L,d)\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "To further stabilize the output, a normalization operation is applied, preserving the same dimensionality of \\((L,d)\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.123, + 0.905, + 0.348 + ], + "angle": 0, + "content": "Subsequently, the features are processed by a feedforward neural network comprising two fully connected layers. The first layer projects the input to a higher-dimensional space \\( d_{ff} \\), resulting in an output with dimensions \\( (L, d_{ff}) \\). The second layer then maps the expanded features back to the original dimensionality \\( d \\), restoring the shape to \\( (L, d) \\). The output of the feedforward network is combined with the original input through a residual connection, preserving the output shape of \\( (L, d) \\). A normalization operation is then applied to ensure stability, maintaining the final output shape of \\( (L, d) \\) for the entire encoder module. Thus, within the encoder module, the temporal sequence length \\( (L) \\) remains constant, whereas the feature dimensionality \\( (d) \\) may briefly change during processing but ultimately returns to its initial dimensionality." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.348, + 0.907, + 0.485 + ], + "angle": 0, + "content": "Unlike traditional transformer models, Tamba adjusts the parameters of its state space model (SSM) based on the input data, allowing it to selectively retain or disregard different pieces of information within the sequence. Mathematically, Tamba uses a state space model to capture the temporal evolution of the input. Given an input sequence \\(\\mathcal{P} = \\{\\mathcal{P}_1,\\mathcal{P}_2,\\dots ,\\mathcal{P}_T\\}\\), where \\(T\\) represents the sequence length and \\(\\mathcal{P}_t\\in \\mathbb{R}^d\\) is the input at each time step \\(t\\), the SSM is governed by the following state evolution equation:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.494, + 0.905, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} _ {t + 1} = A \\left(\\mathcal {P} _ {t}\\right) \\mathbf {h} _ {t} + B \\left(\\mathcal {P} _ {t}\\right) \\mathbf {u} _ {t}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.519, + 0.905, + 0.579 + ], + "angle": 0, + "content": "Here \\(\\mathbf{h}_t\\in \\mathbb{R}^n\\) represents the hidden state at the time step \\(t\\), while \\(A(\\mathcal{P}_t)\\in \\mathbb{R}^{n\\times n}\\) and \\(B(\\mathcal{P}_t)\\in \\mathbb{R}^{n\\times m}\\) are matrices that depend on the input and evolve over time, and \\(\\mathbf{u}_t\\in \\mathbb{R}^m\\) represents the control input." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.58, + 0.905, + 0.61 + ], + "angle": 0, + "content": "The output of the time step \\(t\\) is denoted as \\(\\mathbf{y}_t\\in \\mathbb{R}^P\\), and is computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.619, + 0.905, + 0.637 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} _ {t} = C \\left(\\mathcal {P} _ {t}\\right) \\mathbf {h} _ {t} + D \\left(\\mathcal {P} _ {t}\\right) \\mathbf {u} _ {t}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.78 + ], + "angle": 0, + "content": "where \\( C(\\mathcal{P}_t) \\in \\mathbb{R}^{p \\times n} \\) and \\( D(\\mathcal{P}_t) \\in \\mathbb{R}^{p \\times m} \\) are also matrices that depend on the input. This formulation allows the system to dynamically adjust its parameters based on the current input, facilitating the selective retention and forgetting of information. The computational complexity of this method is linear, as the state update and output computation at each time step involve matrix-vector multiplication. The overall complexity is \\( O(T \\cdot n^2) \\), where \\( T \\) represents the sequence length and \\( n \\) is the state dimension." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.905, + 0.87 + ], + "angle": 0, + "content": "Decomposed Attention Mechanism. We redesigned the self-attention computation, replacing the multi-head attention mechanism with state space model. After receiving the polyline encoding output \\(\\mathcal{P}\\), Tamba performs the attention decomposition using three parallel encoders. The detailed computation process is as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "(I) Spatiotemporal attention. The spatiotemporal attention for each agent \\(A_{i}\\) and scene elements (S) over the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12061" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.092, + 0.905, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.268, + 0.908, + 0.353 + ], + "angle": 0, + "content": "Figure 3. Overview of our Tamba decoding process. The outputs of the three Atten-Tamba encoders carry the context of both the static scene and dynamic agent states, which are concatenated and input into the Cross-Tamba decoder. For predicting the target agent's trajectory, the decoder queries \\( K \\) trajectory modes using independent query tensors, while the encoder's features combine the current state with recursive reasoning from the previous step, allowing all target agents to share a unified scene representation. During the secondary decoding, the predicted state of the proposed trajectory interacts once again with the current scene information, and a recurrent network assigns prediction weights to the proposed trajectories. The final output trajectory is obtained through refinement of these predictions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.379, + 0.482, + 0.446 + ], + "angle": 0, + "content": "\\(\\mathcal{P}\\) at each time step \\(t\\) can be represented as: \\(Z_{\\mathrm{time}} = \\operatorname{softmax}\\left(\\frac{Q_tK_t^\\top}{\\sqrt{d_k}}\\right)V_t\\), where \\(Q_{t}, K_{t}, V_{t} \\in \\mathbb{R}^{T \\times d_{k}}\\) are the query, key, and value matrices at time step \\(t\\) respectively, and \\(d_{k}\\) represents the dimensionality of the key vectors." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.446, + 0.483, + 0.572 + ], + "angle": 0, + "content": "(II) Scene attention. For the attention between agent set \\(A\\) and scene elements \\(S\\), let \\(A_{i} \\in \\mathbb{R}^{d_{a}}\\) represent the embedding of agent \\(i\\), and \\(S_{j} \\in \\mathbb{R}^{d_{m}}\\) represents the embedding of scene element \\(j\\). The attention calculation between agent \\(i\\) and scene element \\(j\\) is given by: \\(\\alpha_{ij} = \\frac{\\exp(A_i^\\top S_j)}{\\sum_k \\exp(A_i^\\top S_k)}\\), where \\(\\alpha_{ij}\\) denotes the attention weight, indicating the relevance of scene elements \\(j\\) to agent \\(i\\). The output is a weighted sum of the scene elements features denote as:" + }, + { + "type": "equation", + "bbox": [ + 0.222, + 0.581, + 0.483, + 0.614 + ], + "angle": 0, + "content": "\\[\nZ _ {\\text {s c e n e}} = \\sum_ {j} \\alpha_ {i j} \\mathrm {S} _ {j}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.626, + 0.483, + 0.792 + ], + "angle": 0, + "content": "(III) Traffic Attention. Similar to Eq. 4, we use \\(\\beta_{ij}\\) to represent the attention weight of traffic control elements (pedestrians and traffic lights) to other dynamic agents (e.g., motorcycles and cars). For a agent \\(i\\), the output feature of traffic attention can be expressed as: \\(Z_{\\text{traffic}} = \\sum_{j} \\beta_{ij} A_j\\). This output is obtained by calculating the weighted sum of the embedded vectors \\(A_j\\) of all dynamic agents, where the weight \\(\\beta_{ij}\\) reflects the impact of the traffic control elements on the dynamic agents. Through this interaction, the model can effectively focus on the influence of pedestrians on dynamic agent movement in non-traffic light scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.802, + 0.299, + 0.817 + ], + "angle": 0, + "content": "3.4. Cross Tamba Decoder" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.826, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The core challenge of trajectory decoding lies in addressing the 'one-to-many' problem, that is, how to generate \\( K \\) output results (possibilities) from one set of input features. In designing the decoder architecture, we drew inspiration from DETR's approach [1] to solving the 'one-to-many' problem" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.907, + 0.468 + ], + "angle": 0, + "content": "in object detection tasks. Specifically, the number of \\( K \\) query vectors is initialized to represent different objects or candidate future trajectories. Through multiple iterations of the attention mechanism, each query vector extracts relevant information from the input features to produce \\( K \\) output results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.909, + 0.711 + ], + "angle": 0, + "content": "Crossed Dynamic and Static Attention. In the proposed cross-Tamba decoder, we designed it by using state space model to replace multi-head attention in the cross-attention mechanism. The query vector \\( Q \\) is separated to interact independently with the keys and values from the encoder's output features. Specifically, the query vector is responsible for querying the modes of \\( K \\) trajectories, while \\( K \\) and \\( V \\) incorporate both the encoded features from the encoder and the reasoning from the previous recursive step. This design allows dynamic agents (e.g., vehicles and motorcycles) to better share mixed features (pedestrians and traffic lights), while all target agents share a single static scene representation and iteratively generate \\( K \\) different proposed trajectories over \\( T' \\) future recursive steps. During the training phase, the proposed trajectories are optimized using the MSE loss (\\( L_{\\text{proposal}} \\)) based on the provided labels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.711, + 0.909, + 0.817 + ], + "angle": 0, + "content": "Trajectory Proposal. After proposing \\( K \\) predicted trajectories, we introduce a prediction weight inference module to compute the weight for each trajectory. In this process, we also use the Cross-Tamba module, which first fuses scene (S) with historical information to infer the likelihood of each predicted trajectory. Specifically, given each trajectory \\( p_k \\) (where \\( k = 1,2,\\dots,K \\)), we can represent it as:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.829, + 0.905, + 0.845 + ], + "angle": 0, + "content": "\\[\np _ {k} = \\text {C r o s s - T a m b a} (\\text {S c e n e}, \\text {P r e d i c t i o n s}). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "where Scene represents the feature representation of the current scene and Predictions contains all predicted proposal trajectories. At the head of the weight inference module, we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12062" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.138 + ], + "angle": 0, + "content": "use a recurrent neural network (RNN) [24] to extract features from the final trajectories in order to generate a score for each trajectory. This RNN module can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.149, + 0.483, + 0.165 + ], + "angle": 0, + "content": "\\[\nh _ {t} = \\operatorname {R N N} \\left(h _ {t - 1}, p _ {k}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.175, + 0.483, + 0.325 + ], + "angle": 0, + "content": "where \\( h_t \\) is the hidden state at time step \\( t \\), \\( h_{t-1} \\) is the previous state. Through this process, the RNN provides a mechanism for constraints across trajectories. When a trajectory conflicts with the environment or when there is an alternative trajectory that better fits the overall distribution, the RNN generates a prior confidence score for these trajectories. The confidence generated by the RNN can be represented as: \\( C_k = f(h_T) \\), where \\( C_k \\) is the confidence score for trajectory \\( p_k \\), \\( h_T \\) is the hidden state of the RNN at the final time step, and \\( f \\) is a nonlinear activation function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.326, + 0.484, + 0.4 + ], + "angle": 0, + "content": "This step ensures greater stability in the optimization process, such that the proposal trajectory considers not only the current prediction results but also the contextual environment and historical information, thereby enhancing the overall performance and robustness of the model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.484, + 0.477 + ], + "angle": 0, + "content": "Trajectory Refinement. We employed the same optimization approach as the baseline methods [39, 40] to perform mixture modeling of the predicted trajectories. The future trajectory of agent \\(i\\) is modeled as a mixture of Laplace distributions, as given by:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.487, + 0.483, + 0.528 + ], + "angle": 0, + "content": "\\[\nf \\left(\\left\\{p _ {t} ^ {i} \\right\\} _ {t = 1} ^ {T ^ {\\prime}}\\right) = \\sum_ {k = 1} ^ {K} \\pi_ {i, k} \\prod_ {t = 1} ^ {T ^ {\\prime}} \\operatorname {L a p l a c e} \\left(p _ {t} ^ {i} \\mid \\mu_ {t, i, k}, b _ {t, i, k}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.483, + 0.735 + ], + "angle": 0, + "content": "where \\(\\{\\pi_{i,k}\\}_{k = 1}^{K}\\) are the mixing coefficients, representing the significance of each mixture component in the overall model. For each time step \\(t\\), the Laplace density of the \\(k^{th}\\) mixture component is defined by the location parameter \\(\\mu t,i,k\\) and the scale parameter \\(b_{t,i,k}\\), which reflect the center and spread of that component. We use a classification loss \\(L_{\\mathrm{cls}}\\) to optimize the mixing coefficients in Eq. 7 by minimizing their negative log-likelihood. It is important to note that during this process, we stop the gradients of the location and scale parameters, focusing solely on the optimization of the mixing coefficients. This allows us to effectively focus on the classification task without affecting the trajectory's location and scale settings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.736, + 0.483, + 0.902 + ], + "angle": 0, + "content": "For optimizing the location and scale parameters, we adopt the \"winner takes all\" strategy [15], resulting in only the best-predicted trajectory from the proposal and refinement modules undergoing backpropagation. This approach ensures that the network is adjusted only based on the best predictions during training, thereby enhancing the accuracy and stability of the model. In addition, to prevent overfitting and improve model robustness, the refinement module halts the gradient propagation of the proposed trajectory. This strategy helps to effectively preserve the previously learned information during the adjustment of proposals." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "Finally, our loss function integrates three components: the trajectory proposal loss (\\(L_{\\mathrm{proposal}}\\)), the trajectory refinement loss (\\(L_{\\mathrm{refine}}\\)), and the classification loss (\\(L_{\\mathrm{cls}}\\)), enabling an end-to-end training process. The expression is given as:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.167, + 0.905, + 0.183 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {t o t a l}} = L _ {\\text {p r o p o s a l}} + L _ {\\text {r e f i n e}} + \\lambda L _ {\\text {c l s}}. \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.212, + 0.646, + 0.229 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.237, + 0.616, + 0.252 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.261, + 0.907, + 0.549 + ], + "angle": 0, + "content": "Our experiments were conducted on the Argoverse 1 [2] and Argoverse 2 [14, 33] datasets. It contains extensive real-world autonomous driving scenarios and are designed to advance the development of autonomous driving technologies. The Argoverse 1 dataset includes 360-degree sensor data from the Pittsburgh and Miami regions in the United States, covering various road types and complex urban environments. Sampled at a rate of \\(10\\mathrm{Hz}\\), this dataset provides trajectory data of the ego vehicle over a 5-second time span, with 2 seconds of observation used to predict the path over the next 3 seconds, comprising over 30,000 such samples. Argoverse 2 further expands on the dataset by including richer sensor data, multimodal annotations, and extending the observation window. Specifically, it includes 11-second scenarios, with 5 seconds of observation followed by prediction of trajectories for the next 6 seconds. The dataset contains 250,000 scenarios and features a large object taxonomy comprising 10 non-overlapping classes, which is used for training and validating motion prediction models." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.56, + 0.61, + 0.574 + ], + "angle": 0, + "content": "4.2. Metrics" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.584, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We adopted evaluation methods consistent with prior research, using metrics such as \\(\\min\\mathrm{ADE}(K)\\), \\(\\min\\mathrm{FDE}(K)\\), \\(\\min\\mathrm{FDE}(K)\\), and \\(\\mathrm{MR}(K)\\) to measure model performance in terms of error. Specifically, \\(\\min\\mathrm{ADE}(K)\\) calculates the average \\(L2\\) distance between the ground truth and the most optimal trajectory among the \\(K\\) predicted trajectories across all future time steps, thereby assessing the overall prediction accuracy of the model over multiple time steps. \\(\\min\\mathrm{FDE}(K)\\) focuses on the prediction error at the final time step, measuring the model's ability to predict the final state accurately. \\(\\mathrm{b-minFDE}(K)\\) extends \\(\\min\\mathrm{FDE}(K)\\) by incorporating probability weighting, using \\((1 - \\hat{\\pi})^2\\) as a weight to account for prediction uncertainty, where \\(\\hat{\\pi}\\) represents the probability score assigned by the model to the optimal predicted trajectory. \\(\\mathrm{MR}(K)\\) measures the proportion of the \\(K\\) predicted trajectories with an error greater than 2 meters, providing an assessment of the model's stability and the frequency of prediction failures. Together, these metrics offer a comprehensive evaluation of the model's prediction accuracy, stability, and uncertainty estimation. For the parameter \\(K\\), we followed conventional values of 6 and 1. When the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12063" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.899, + 0.23 + ], + "angle": 0, + "content": "
Methodb-minFDE6↓minADE6↓minFDE6↓MR6↓minADE1↓minFDE1↓MR1↓#Params (M)↓FLOPs(G)↓
THOMAS [6]2.160.881.510.201.954.710.64--
GoRela [4]2.010.761.480.221.824.620.66--
QML [28]1.950.691.390.191.844.980.629.39-
MTR [26]1.980.731.440.151.744.390.5865.78-
GANet [31]1.960.721.340.171.774.480.5961.7315.79
BANet [32]1.920.711.360.191.794.610.609.4911.93
QCNet [40]1.910.651.290.161.694.300.597.6645.3
Tamba (Ours)1.890.641.240.171.664.240.574.5427.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.241, + 0.908, + 0.298 + ], + "angle": 0, + "content": "Table 1. Performance comparison on Argoverse 2 [14, 33] dataset, leaderboard ranked by b-minFDE6. Subscript 6 and 1 represents the number of predictions. Best performance is highlighted in bold. (M) represents the number of parameters in millions; (G) refers to GigaOps per second. - represents we were unable to reproduce the results with complete accuracy. Notably, all methods in the table are direct comparisons of the original approaches. We exclude additional improvements from engineering techniques such as ensembling." + }, + { + "type": "table", + "bbox": [ + 0.206, + 0.31, + 0.795, + 0.468 + ], + "angle": 0, + "content": "
Methodb-minFDE6↓minADE6↓minFDE6↓MR6↓#Params (M)↓
LaneGCN [16]2.060.871.360.16-
DenseTNT [10]1.980.881.280.13-
MTR [26]1.930.821.240.13-
SceneTransformer [22]1.890.801.230.1315.30
HiVT [39]1.840.771.170.13-
GANet [31]1.790.811.160.12-
Wayformer [21]1.740.771.160.12-
QCNet [40]1.690.731.070.117.66
Tamba (Ours)1.670.721.070.094.54
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.479, + 0.908, + 0.522 + ], + "angle": 0, + "content": "Table 2. Quantitative results on the Argoverse 1 [2] dataset, leaderboard ranked by b-minFDE6. Subscript 6 represents the number of predictions. Best performance is highlighted in bold. (M) represents the number of parameters in millions. - represents it is unable to reproduce the results with complete accuracy. We exclude additional improvements from engineering techniques such as ensembling." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.547, + 0.483, + 0.608 + ], + "angle": 0, + "content": "number of predicted trajectories exceeds \\( K \\), the top \\( K \\) trajectories with the highest probabilities are selected. In terms of lightweight evaluation, we used the number of parameters (M) and FLOPs(G) as the metrics." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.618, + 0.19, + 0.634 + ], + "angle": 0, + "content": "4.3. Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.641, + 0.484, + 0.777 + ], + "angle": 0, + "content": "We conducted all experiments under the same environment (batch size, epochs) using eight NVIDIA 3090TI cards with a batch size of 128. We used Adam for optimization, with an initial learning rate of 0.001. Training was conducted for a total of 50 epochs. We employed an adaptive learning rate adjustment strategy based on the validation set's performance during training. Specifically, when the validation accuracy did not improve for five consecutive epochs, the learning rate was reduced to 0.1 times its original value." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.787, + 0.384, + 0.803 + ], + "angle": 0, + "content": "4.4. Comparison with State of the Art" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We compare our method with other SOTA approaches across various evaluation metrics in Argoverse 2 dataset. As shown in Table 1, our method demonstrates superior performance on key metrics such as \\(\\mathrm{minFDE}_6\\) and \\(\\mathrm{minADE}_6\\). Tamba achieves a b-minFDE \\(^6\\) score of 1.79, compared to 1.91 for QCNet, 1.92 for BANet and 1.98 for MTR. Additionally," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.907, + 0.744 + ], + "angle": 0, + "content": "Tamba achieves \\((\\mathrm{MR}_1)\\) of 0.57, which is lower than that of other methods, such as QCNet (0.59) and MTR (0.58), indicating more stable prediction accuracy. Furthermore, Tamba has a parameter count of \\(4.54\\mathrm{M}\\), which is significantly lower than QCNet (7.66 M), even compare to non-transformer models such as BANet (9.49M) and GANet (61.73M), Tamba achieving reductions of \\(52.0\\%\\) and \\(92.7\\%\\) respectively, while still outperforming both models. Additionally, Tamba achieves 27.3 GigaOps per second, which is \\(40\\%\\) faster than QCNet's 45.3 GigaOps per second. These results clearly demonstrate that Tamba is competitive in terms of accuracy but also excels in lightweight design and efficiency, providing an excellent solution for complex task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.901 + ], + "angle": 0, + "content": "On the Argoverse 1 dataset, our method outperforms other state-of-the-art (SOTA) methods across the board. On the leaderboard for the b-minFDE\\(_6\\) metric, Tamba achieves a score of 1.67, while the second-best, QCNet, scores 1.69, and the third, BANet, scores 1.74. In terms of parameter count, our method requires 4.54M parameters, compared to 7.66M for QCNet, reflecting a \\(40\\%\\) reduction. This further demonstrates the superiority of our approach in balancing efficiency and performance. For detailed performance metrics, please refer to Table 2." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12064" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.087, + 0.497, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.194, + 0.488, + 0.208 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.087, + 0.9, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.876, + 0.194, + 0.898, + 0.208 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.214, + 0.496, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.214, + 0.9, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.323, + 0.208, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.212, + 0.346, + 0.348, + 0.362 + ], + "angle": 0, + "content": "Predicted trajectory" + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.325, + 0.488, + 0.34 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.346, + 0.594, + 0.36 + ], + "angle": 0, + "content": "Observation" + }, + { + "type": "image", + "bbox": [ + 0.694, + 0.346, + 0.738, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.347, + 0.827, + 0.364 + ], + "angle": 0, + "content": "Target agent" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.378, + 0.907, + 0.421 + ], + "angle": 0, + "content": "Figure 4. Qualitative results on the Argoverse 2 validation set. The target agent marked as orange with its surrounding agents marked as cold white. We conduct four different traffic scenarios. (a). Multi-agents in straight-road scenario, (b). Multiple agents on roundabout road, (c). Vehicle avoidance after roundabout, (d). Mixed scenario involving pedestrians and vehicles." + }, + { + "type": "table", + "bbox": [ + 0.155, + 0.433, + 0.846, + 0.515 + ], + "angle": 0, + "content": "
Methodw/ jointw/o joint
minFDE6↓minADE6↓minFDE1↓minADE1↓minFDE6↓minADE6↓minFDE1↓minADE1↓
Ours + Attention1.0940.7633.0371.3761.0930.7283.0351.372
Ours + Mamba1.0910.7242.8991.2981.0830.7652.8211.292
Ours + Tamba0.9510.6732.7421.2650.9550.6802.7451.272
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.525, + 0.907, + 0.553 + ], + "angle": 0, + "content": "Table 3. Ablation study on the components of the Tamba encoder-decoder. Experimental results are conducted on the Argoverse 2 [14, 33] validation set." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.579, + 0.29, + 0.594 + ], + "angle": 0, + "content": "4.5. Quantization Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.603, + 0.484, + 0.738 + ], + "angle": 0, + "content": "We visualized the prediction results on the validation set. In some challenging sequences, the predictions made by the Tamba model are more reasonable and smoother compared to existing methods, effectively adhering to map constraints. In Figure 4, we present several multimodal prediction cases, the results demonstrate that the trajectories generated by Tamba are smoother and better adapted to complex road topologies, meeting the demands of vehicle navigation under diverse road conditions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.755, + 0.242, + 0.772 + ], + "angle": 0, + "content": "5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We evaluated the effectiveness of the Tamba module. For comparison, we used both the traditional attention mechanism module and the Mamba module on the validation set. In addition, we assessed the efficacy of our proposed pedestrian-traffic light joint polyline encoding strategy. Table 3 presents the performance comparison of our proposed three methods (including Attention variant, Mamba variant, and Tamba variant) under conditions with and without joint" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.906, + 0.656 + ], + "angle": 0, + "content": "polyline encoding (w/ joint vs. w/o joint). Specifically, the table lists the performance of each method on two evaluation metrics: minFDE and minADE, considering both six predicted trajectories (minFDE\\(_6\\) and minADE\\(_6\\)) and one predicted trajectory (minFDE\\(_1\\) and minADE\\(_1\\))." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.688, + 0.634, + 0.704 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This paper introduces the Tamba model, an efficient motion prediction model for the real-time and computational needs of autonomous driving. By employing a state space model (SSM) and a novel attention Tamba encoder-decoder, we reduce computational complexity and parameter count while maintaining state-of-the-art accuracy. Experiments on the Argoverse 1 and 2 datasets show that Tamba achieves superior inference speed and efficiency, with a fourfold reduction in FLOPs and over \\(40\\%\\) fewer parameters compared to transformer-based models. These advantages make Tamba highly suitable for high-dynamic environments requiring rapid responses." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12065" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.226 + ], + "angle": 0, + "content": "This work is supported in part by the Europe Eureka Intelligence to Drive | Move-Save-Win project (with funding from the UKRI Innovate UK project under Grant No. 10071278) as well as the Horizon Europe COVER project, No. 101086228 (with funding from UKRI grant EP/Y028031/1). Kezhi Wang would like to acknowledge the support in part by the Royal Society Industry Fellow scheme." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.238, + 0.188, + 0.254 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.262, + 0.484, + 0.331 + ], + "angle": 0, + "content": "[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision (ECCV), pages 213-229. Springer, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.333, + 0.484, + 0.415 + ], + "angle": 0, + "content": "[2] Ming-Fang Chang, John W Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, and James Hays. Argoverse: 3d tracking and forecasting with rich maps. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.416, + 0.485, + 0.498 + ], + "angle": 0, + "content": "[3] Yihua Cheng, Yaning Zhu, Zongji Wang, Hongquan Hao, Yongwei Liu, Shiqing Cheng, Xi Wang, and Hyung Jin Chang. What do you see in vehicle? comprehensive vision solution for in-vehicle gaze estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.499, + 0.484, + 0.567 + ], + "angle": 0, + "content": "[4] Alexander Cui, Sergio Casas, Kelvin Wong, Simon Suo, and Raquel Urtasun. Gorela: Go relative for viewpoint-invariant motion forecasting. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 7801-7807. IEEE, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.568, + 0.484, + 0.65 + ], + "angle": 0, + "content": "[5] Jiyang Gao, Chen Sun, Hang Zhao, Yi Shen, Dragomir Anguelov, Congcong Li, and Cordelia Schmid. Vectornet: Encoding hd maps and agent dynamics from vectorized representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11525-11533, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.651, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[6] Thomas Gilles, Stefano Sabatini, Dzmitry Tsishkou, Bogdan Stanciulescu, and Fabien Moutarde. Thomas: Trajectory heatmap output with learned multi-agent sampling. In International Conference on Learning Representations (ICLR), 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.721, + 0.484, + 0.762 + ], + "angle": 0, + "content": "[7] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.762, + 0.484, + 0.817 + ], + "angle": 0, + "content": "[8] Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. Hippo: Recurrent memory with optimal polynomial projections. Advances in neural information processing systems, 33:1474-1487, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.484, + 0.873 + ], + "angle": 0, + "content": "[9] Albert Gu, Karan Goel, Ankit Gupta, and Christopher Ré. On the parameterization and initialization of diagonal state space models. Advances in Neural Information Processing Systems, 35:35971-35983, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.485, + 0.902 + ], + "angle": 0, + "content": "[10] Junru Gu, Chen Sun, and Hang Zhao. Densetnt: End-to-end trajectory prediction from dense goal sets. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.262, + 0.485, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF International Conference on Computer Vision, pages 15303-15312, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[11] Ankit Gupta, Albert Gu, and Jonathan Berant. Diagonal state spaces are as effective as structured state spaces. Advances in Neural Information Processing Systems, 35:22982-22994, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[12] Yizhou Huang, Yihua Cheng, and Kezhi Wang. Efficient driving behavior narration and reasoning on edge device using large language models. arXiv preprint arXiv:2409.20364, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[13] Parth Kothari, Brian Sifringer, and Alexandre Alahi. Interpretable social anchors for human trajectory forecasting in crowds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15556-15566, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.372 + ], + "angle": 0, + "content": "[14] John Lambert and James Hays. Trust, but verify: Cross-modality fusion for hd map change detection. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks (NeurIPS Datasets and Benchmarks 2021), 2021. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.442 + ], + "angle": 0, + "content": "[15] Stefan Lee, Senthil Purushwalkam Shiva Prakash, Michael Cogswell, Viresh Ranjan, David Crandall, and Dhruv Batra. Stochastic multiple choice learning for training diverse deep ensembles. Advances in Neural Information Processing Systems, 29, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[16] Ming Liang, Bin Yang, Rui Hu, Yun Chen, Renjie Liao, Song Feng, and Raquel Urtasun. Learning lane graph representations for motion forecasting. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 541-556. Springer, 2020. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.908, + 0.621 + ], + "angle": 0, + "content": "[17] Mengmeng Liu, Hao Cheng, Lin Chen, Hellward Broszio, Jiangtao Li, Runjiang Zhao, Monika Sester, and Michael Ying Yang. Laformer: Trajectory prediction for autonomous driving with lane-aware scene constraints. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2039-2049, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.622, + 0.905, + 0.679 + ], + "angle": 0, + "content": "[18] Yuejiang Liu, Qi Yan, and Alexandre Alahi. Social nce: Contrastive learning of socially-aware motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15118-15129, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[19] Karttikeya Mangalam, Yang An, Harshayu Girase, and Jitendra Malik. From goals, waypoints & paths to long term human trajectory forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15233-15242, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[20] Abduallah Mohamed, Kun Qian, Mohamed Elhoseiny, and Christian Claudel. Social-stgCNN: A social spatio-temporal graph convolutional neural network for human trajectory prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14424-14432, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[21] Nigamaa Nayakanti, Rami Al-Rfou, Aurick Zhou, Kratarth Goel, Khaled S Refaat, and Benjamin Sapp. Wayformer: Motion forecasting via simple & efficient attention networks. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2980-2987. IEEE, 2023. 1, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12066" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.188 + ], + "angle": 0, + "content": "[22] Jiquan Ngiam, Vijay Vasudevan, Benjamin Caine, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, et al. Scene transformer: A unified architecture for predicting future trajectories of multiple agents. In 2023 IEEE International Conference on Robotics and Automation (ICRA), 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[23] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory++: Dynamically-feasible trajectory forecasting with heterogeneous data. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVIII 16, pages 683-700. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[24] Alex Sherstinsky. Fundamentals of recurrent neural network (rnn) and long short-term memory (lstm) network. *Physical D: Nonlinear Phenomena*, 404:132306, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[25] Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Wei Tang, Nanning Zheng, and Gang Hua. Representing multimodal behaviors with mean location for pedestrian trajectory prediction. IEEE transactions on pattern analysis and machine intelligence, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[26] Shaoshuai Shi, Li Jiang, Dengxin Dai, and Bernt Schiele. Mtr++: Multi-agent motion prediction with symmetric scene modeling and guided intention querying. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[27] Jimmy Smith, Shalini De Mello, Jan Kautz, Scott Linderman, and Wonmin Byeon. Convolutional state space models for long-range spatiotemporal modeling. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[28] Tong Su, Xishun Wang, and Xiaodong Yang. Qml for argoverse 2 motion forecasting challenge. arXiv preprint arXiv:2207.06553, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.549, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[29] Balakrishnan Varadarajan, Ahmed Hefny, Avikalp Srivastava, Khaled S Refaat, Nigamaa Nayakanti, Andre Cornman, Kan Chen, Bertrand Douillard, Chi Pang Lam, Dragomir Anguelov, et al. Multipath++: Efficient information fusion and trajectory aggregation for behavior prediction. In 2022 International Conference on Robotics and Automation (ICRA), pages 7814-7821. IEEE, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[30] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[31] Mingkun Wang, Xinge Zhu, Changqian Yu, Wei Li, Yuexin Ma, Ruochun Jin, Xiaoguang Ren, Dongchun Ren, Mingxu Wang, and Wenjing Yang. Ganet: Goal area network for motion forecasting. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1609-1615. IEEE, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[32] Zhepeng Wang, Feng Chen, Kanokphan Lertniphonphan, Siwei Chen, Jinyao Bao, Pengfei Zheng, Jinbao Zhang, Kaer Huang, and Tao Zhang. Technical report for argoverse challenges on unified sensor-based detection, tracking, and forecasting. arXiv preprint arXiv:2311.15615, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[33] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks (NeurIPS Datasets and Benchmarks 2021), 2021. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.244 + ], + "angle": 0, + "content": "[34] Conghao Wong, Beihao Xia, Ziqian Zou, Yulong Wang, and Xinge You. Socialcircle: Learning the angle-based social interaction representation for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19005-19015, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.248, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[35] Maosheng Ye, Tongyi Cao, and Qifeng Chen. Tpcn: Temporal point cloud networks for motion forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11318-11327, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.372 + ], + "angle": 0, + "content": "[36] Ye Yuan, Xinshuo Weng, Yanglan Ou, and Kris M Kitani. Agentformer: Agent-aware transformers for socio-temporal multi-agent forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9813-9823, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.905, + 0.43 + ], + "angle": 0, + "content": "[37] Zeyu Zhang, Akide Liu, Ian Reid, Richard Hartley, Bohan Zhuang, and Hao Tang. Motion mamba: Efficient and long sequence motion generation. In European Conference on Computer Vision, pages 265-282. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.432, + 0.905, + 0.513 + ], + "angle": 0, + "content": "[38] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634-21643, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.516, + 0.905, + 0.583 + ], + "angle": 0, + "content": "[39] Zikang Zhou, Luyao Ye, Jianping Wang, Kui Wu, and Kejie Lu. Hivt: Hierarchical vector transformer for multi-agent motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8823-8833, 2022. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.586, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[40] Zikang Zhou, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang. Query-centric trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17863-17873, 2023. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.905, + 0.697 + ], + "angle": 0, + "content": "[41] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12067" + } + ] +] \ No newline at end of file diff --git a/2025/Trajectory Mamba_ Efficient Attention-Mamba Forecasting Model Based on Selective SSM/14500b01-a80f-4244-a6f9-1c844f2302da_origin.pdf b/2025/Trajectory Mamba_ Efficient Attention-Mamba Forecasting Model Based on Selective SSM/14500b01-a80f-4244-a6f9-1c844f2302da_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ab9f5192ccc268e05f60448d3d6cabcf3616a23d --- /dev/null +++ b/2025/Trajectory Mamba_ Efficient Attention-Mamba Forecasting Model Based on Selective SSM/14500b01-a80f-4244-a6f9-1c844f2302da_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f23916b8d7720eb678842607a209360729d291a30ea6c1c239ef8e1d6ae1f1f +size 434801 diff --git a/2025/Trajectory Mamba_ Efficient Attention-Mamba Forecasting Model Based on Selective SSM/full.md b/2025/Trajectory Mamba_ Efficient Attention-Mamba Forecasting Model Based on Selective SSM/full.md new file mode 100644 index 0000000000000000000000000000000000000000..eb194745c3edc9cc645d5d193388e3c2f1b1ea7d --- /dev/null +++ b/2025/Trajectory Mamba_ Efficient Attention-Mamba Forecasting Model Based on Selective SSM/full.md @@ -0,0 +1,292 @@ +# Trajectory Mamba: Efficient Attention-Mamba Forecasting Model Based on Selective SSM + +Yizhou Huang $^{1}$ , Yihua Cheng $^{2}$ , Kezhi Wang $^{1*}$ $^{1}$ Brunel University of London, $^{2}$ University of Birmingham +{byyh009, Kezhi.Wang}@brunel.ac.uk, y.cheng.2@bham.ac.uk +https://github.com/Yizhou-H/Trajectory-Mamba-CVPR + +# Abstract + +Motion prediction is crucial for autonomous driving, as it enables accurate forecasting of future vehicle trajectories based on historical inputs. This paper introduces Trajectory Mamba, a novel efficient trajectory prediction framework based on the selective state-space model (SSM). Conventional attention-based models face the challenge of computational costs that grow quadratically with the number of targets, hindering their application in highly dynamic environments. In response, we leverage the SSM to redesign the self-attention mechanism in the encoder-decoder architecture, thereby achieving linear time complexity. To address the potential reduction in prediction accuracy resulting from modifications to the attention mechanism, we propose a joint polyline encoding strategy to better capture the associations between static and dynamic contexts, ultimately enhancing prediction accuracy. Additionally, to balance prediction accuracy and inference speed, we adopted the decoder that differs entirely from the encoder. Through cross-state space attention, all target agents share the scene context, allowing the SSM to interact with the shared scene representation during decoding, thus inferring different trajectories over the next prediction steps. Our model achieves state-of-the-art results in terms of inference speed and parameter efficiency on both the Argoverse 1 and Argoverse 2 datasets. It demonstrates a four-fold reduction in FLOPs compared to existing methods and reduces parameter count by over $40\%$ while surpassing the performance of the vast majority of previous methods. These findings validate the effectiveness of Trajectory Mamba in trajectory prediction tasks. + +# 1. Introduction + +In the rapid development of autonomous driving technology [3, 12, 17], accurate motion prediction plays a crucial role by forecasting future trajectories based on historical vehicle + +![](images/0699c5c3ed435b426a368e436b3fd6c42cded92936a34d53a12a03dea185f68e.jpg) +Figure 1. Illustration of proposed joint polyline encoding strategy, where we consider all factors that affect the movement of motor vehicles, jointly encoding pedestrians and traffic lights rather than categorizing the scene horizontally into static and dynamic context. Additionally, we decompose and encode the interactions between all agents and elements at each time step. + +movement data. This task is not only vital for enhancing driving safety but also meets the stringent requirements for real-time system response. In recent years, learning-based approaches have been widely applied to motion prediction [5, 16, 29, 39, 40]. However, existing prediction methods often struggle to balance efficiency and accuracy, making it challenging to simultaneously meet performance and efficiency requirements in real-time applications. We attempt to analyze the underlying reasons from the following aspects and explore possible improvements: + +(a). Improving Prediction Accuracy: In autonomous driving scenarios [38], there are numerous dynamic and static agents, and the diversity of this data presents significant challenges for motion prediction and model training. However, conventional methods [29, 39] often struggle to effectively fuse this heterogeneous data, particularly when it comes to integrating various features, which leads to suboptimal prediction accuracy. On a positive note, factorized attention-based transformers [21, 40] have made no + +table strides in decomposing heterogeneous data. However, these models often use a unified, horizontal approach when dealing with static scenes and dynamic agent features, and they fail to fully consider the deeper relationships between different scene elements and agent information. + +(b). Enhancing Response Efficiency: Motion predictions are inherently recursive, implying that each prediction serves as an input for the next time step, leading to the progressive accumulation of errors over time. Notably, the advancement of sparse context encoding techniques [5, 35] has demonstrated significant potential in enhancing model efficiency. Nonetheless, prediction models based on attention mechanisms still face exponential growth in computational complexity and parameter size during recursive tasks due to inherent design constraints, which severely limit their efficiency and real-time capabilities in practical applications. Conversely, methods such as LSTM [18] are relatively more efficient, but numerous studies [20, 23, 36] have shown that they are incapable of achieving the predictive accuracy of transformer-based models. + +In response, we propose an Attention-Mamba framework for motion prediction, termed Trajectory Mamba (Tamba), to overcome the limitations discussed above. First, we design a multimodal encoder-decoder architecture based on the state space model [9, 11, 27], comprising three parallel encoders and a decoder. By introducing a selective state space, we redefined the computation of the self-attention mechanism, enabling the model to selectively process input information by parameterizing the input of the state space model (SSM), thereby allowing the model to focus on or ignore specific inputs. Specifically, the selective state space module processes local short-range features by receiving input from a one-dimensional convolutional layer, ensuring that these features can be selected or discarded before entering the SSM. We utilized local state space attention in place of the multi-head attention in the self-attention module, reducing the computational complexity of recursive reasoning from exponential to linear. + +Secondly, to better integrate heterogeneous data features, we proposed a joint polyline encoding strategy. Its objective is to separate pedestrian agents from the dynamic agents and jointly encode them with traffic light information from the static scene (see Fig. 1). We observed that in uncertain motion prediction scenarios, the application of traffic regulations exerts a degree of implicit control over the potential movement states of agents. For example, pedestrian-priority traffic rules remain effective even in the absence of traffic lights, influencing vehicles and motorcycles alike. By jointly encoding pedestrians and traffic lights through a shared embedder, this jointly encoding strategy can independently interact with other dynamic agents to enhance the prediction accuracy of vehicle and motorcycle agents. + +Lastly, we designed a prediction weight inference mod + +ule. By redesigning selective state space with a cross-attention mechanism, we utilized independent query tensors to recursively infer multimodal candidate trajectories through interactions with the encoder and the keys and values from the previous recursive step. These candidate trajectories were then evaluated using a recurrent neural network to assign trajectory scores. + +Our contributions are summarized as: (1) We propose the Tamba framework for motion prediction, which leverages the state space model to redesign computations of selfattention, resulting in computational complexity reducing significantly. (2) We propose a joint polyline encoding strategy. By integrating strongly correlated polyline types using independent or shared embeddings, elements that impose similar behavioral patterns can better influence different agents, collectively contributing to the trajectory reasoning of these agents. (3) We redesign the Mamba decoder with crossattention, allowing the joint scene encoding to share a unified scene representation, which further reduces the model's complexity. At the final stage, we introduced a prediction weight inference module to refine the predicted trajectories and enhance prediction accuracy. (4) Experimental results demonstrate that our method achieves state-of-the-art (SOTA) performance on Argoverse 1 and 2 datasets. Compared to advanced methods, our method reduces the number of parameters by $40\%$ , achieves a four-fold increase in FLOPs. This represents a new level of balance between efficiency and performance in our model. + +# 2. Related Works + +# 2.1. State Space Model + +Recently, the Mamba [7] framework has rekindled interest in state-space models (SSMs) [9, 11, 27] due to its introduction of an input-dependent selection mechanism, which enhances their applicability and performance. SSMs are renowned for their capacity to handle long-range dependencies and offer superior memory and computational efficiency compared to transformer-based architectures [30]. Hippo [8], a seminal contribution to SSMs, established the efficacy of these models in long-range sequence modeling tasks. Subsequent research, such as Vision mamba [41], has further refined the mechanisms underlying SSMs, resulting in substantial performance improvements. SSMs have demonstrated their versatility and robustness across a diverse set of domains, including computer vision, video analysis, medical imaging, graph-based learning, point cloud processing, and recommendation systems. + +In the domain of trajectory prediction, recent studies have indicated that Motion Mamba [37] achieves notable success in predicting human trajectory time series. Building on these advancements, this work integrates the state-space mechanisms of Mamba with attention mechanisms to address tra + +jectory prediction tasks in dynamic traffic scenarios, thereby leveraging both the efficiency of SSMs and the adaptability of attention-based models. + +# 2.2. Attention Trajectory Prediction Model + +In the field of trajectory prediction, methods [13, 19] based on the attention mechanism have been widely adopted to enhance the model's ability to capture and predict complex behaviors. These approaches leverage the capability of attention to focus selectively on important parts of the input, thereby capturing key temporal and spatial features that are crucial for predicting future trajectories. + +Subsequently, several advanced approaches [17, 34] have utilized attention mechanisms for trajectory prediction specifically in multi-agent interaction environments, where understanding the interactions between different agents (such as vehicles, pedestrians, and other road users) is critical. These models employ social attention modules to focus on the movements and behaviors of surrounding vehicles, effectively improving the ability of the system to handle intricate interactive behaviors, which are a common aspect of real-world driving scenarios. + +Moreover, attention mechanisms have also been successfully applied in models that integrate multimodal information [25, 26, 40]. These models dynamically weight and fuse different types of sensory inputs—such as image features from cameras and data from other sensors like LiDAR—through attention modules. This fusion process allows the model to better understand and represent complex traffic environments, effectively utilizing complementary data to enhance trajectory prediction performance. + +# 3. Our Method + +# 3.1. Overview + +In the autonomous driving scenario, the input consists of the predicted dynamic agent, its surrounding agents, and the static scene context. Let the historical state of the $i$ -th agent at time step $t$ be represented as: $s_i^t = (p_i^t,\theta_i^t,t,\nu_i^t)$ , where $p_i^t = (p_{i,x}^t,p_{i,y}^t)$ denotes the spatial position, $\theta_i^t$ represents the heading angle, $t$ refers to the time step, and $\nu_{i}^{t}$ denotes the velocity. The static scene context is represented by $n$ polygons, which include elements such as lane boundaries, map edges, and sidewalks. The agent states $s_i^t$ over $T$ time steps within the observation window are used to predict $K$ future trajectories over $T^{\prime}$ time steps. + +In this paper, we propose the Tamba model to optimize the efficiency and accuracy of motion forecasting tasks. Our input includes $s_i^t$ of dynamic and static agents, $\mathcal{P}$ processed using a heterogeneous polyline encoding strategy to transform $s_i^t$ into higher-dimensional features. We leverage three Tamba encoders to separately decompose (I) attendations in each time step $(t)$ over all elements, (II) attendations between + +![](images/d561ba29edd4ce0fbb3275cf902483eacb131baa7b85efa177e812082709735e.jpg) +Figure 2. Overview of our Tamba encoder. We employ joint polyline encoding strategy to integrate strongly correlated polyline information and used three parallel encoders to interact with and associate these features. We applied a linear projection to map the attention output of Tamba back to the same dimensions as the input and used normalization to enhance the stability of the output. + +all agents (A) and scene (S), and (III) attentions of traffic control elements (pedestrians and traffic lights) to other dynamic agents (motorcycles and cars). Then, concatenate features from encoders as output to a Tamba decoder, where we utilize cross-attention mechanism to generate proposal trajectories. Subsequently, we fused scene encoding with historical information to perform a secondary decoding of the proposal trajectories for refinement. Finally, the refined trajectories were optimized using a "winner-takes-all strategy" to obtain trajectory output. + +Our model contains a novel heterogeneous polyline encoding strategy along with a decomposed feature fusion method for both static and dynamic agents. A redesigned Tamba encoder-decoder architecture. A refinement process to output the final predicted trajectory. In the following sections, we sequentially introduce the components in Tamba. + +# 3.2. Joint Polyline Encoding + +In multimodal traffic environment prediction tasks, different types of polylines possess unique geometric characteristics and semantic information. This study proposes a novel heterogeneous polyline embedding method to facilitate the fusion of dynamic and static context features, thereby providing a richer and more efficient feature representation. Specifically, we assign either independent or shared embedders to each type of polyline to effectively model various categories of information. Specifically, we jointly encode polylines with strong information associations, such as pedestrian trajectories and traffic light signals, using a shared embedder. This is because agents (pedestrians) and traffic lights typically impose similar behavioral patterns on other agents (vehicles) and often participate together in traffic control. Thus, using a shared embedder reduces model complexity + +and promotes feature sharing, collectively influencing vehicle trajectory reasoning. + +In contrast, independent embedders are used for polyline types with significant feature differences, such as lane lines and traffic signs in maps. Each embedder consists of a MLP layer and normalization layers to extract high-dimensional feature representations. The shared pedestrian and traffic light embedder uses a cross-category feature fusion layer to enhance the model's sensitivity to different information sources. Polylines retain sequential information through positional encoding before being input to the embedders for feature extraction, followed by cross-category feature interaction via the Tamba encoder. + +The mathematical representation of this process is as follows: Given the individual polyline data $P = \{p_1, p_2, \ldots, p_n\}$ , where $p_i \in \mathbb{R}^d$ represents a point on the polyline, $n$ is the number of points, and $d$ is the dimensionality, the embedding process is represented as: $\mathcal{P} = \operatorname{Embed}(P)$ . For the pedestrian and traffic light categories that share an embedder, the output feature representation $\mathcal{P}_{\mathrm{joint}}$ is computed as follows: + +$$ +\mathcal {P} _ {\text {j o i n t}} = \text {F u s i o n} \left(\operatorname {E m b e d} \left(\mathcal {P} _ {\text {p e d e s t r i a n}}\right), \operatorname {E m b e d} \left(\mathcal {P} _ {\text {t r a f f i c}}\right)\right), \tag {1} +$$ + +where $\mathcal{P}_{\mathrm{pedestrian}}$ and $\mathcal{P}_{\mathrm{traffic}}$ are the input polylines for pedestrians and traffic lights, respectively, and Fusion represents the feature fusion operation. + +# 3.3. Attention Tamba Encoder + +In the proposed encoder, the input features $\mathcal{P}$ are uniformly embedded through joint polyline encoding. The input feature shape of the Tamba encoder is $(n,L,d)$ , where $n$ denotes the number of encoded agents or scene elements, $L$ represents the temporal sequence length, and $d$ indicates the feature dimensionality at each time step. + +Specifically, the sequence $(L,d)$ undergoes linear transformations to generate the query, key, and value matrices, each with dimensions $(L,d_{k})$ , where $d_{k}$ is the feature dimensionality used in the attention mechanism. The dot-product similarity between the query and key matrices is computed to yield an attention weight matrix of shape $(L,L)$ , which captures the interdependencies between all time steps. Subsequently, these attention weights are employed to compute a weighted sum of the value matrix, yielding an output of shape $(L,d_{k})$ . + +As the output feature dimensionality of the Tamba block $(d_k)$ may differ from the input feature dimensionality $(d)$ , a linear projection is employed to align the attention mechanism's output with the input dimensionality, resulting in $(L,d)$ . This alignment ensures compatibility for the residual connection. Subsequently, the original input and the linearly projected attention output are combined via elementwise addition, yielding a resultant tensor of shape $(L,d)$ . + +To further stabilize the output, a normalization operation is applied, preserving the same dimensionality of $(L,d)$ . + +Subsequently, the features are processed by a feedforward neural network comprising two fully connected layers. The first layer projects the input to a higher-dimensional space $d_{ff}$ , resulting in an output with dimensions $(L, d_{ff})$ . The second layer then maps the expanded features back to the original dimensionality $d$ , restoring the shape to $(L, d)$ . The output of the feedforward network is combined with the original input through a residual connection, preserving the output shape of $(L, d)$ . A normalization operation is then applied to ensure stability, maintaining the final output shape of $(L, d)$ for the entire encoder module. Thus, within the encoder module, the temporal sequence length $(L)$ remains constant, whereas the feature dimensionality $(d)$ may briefly change during processing but ultimately returns to its initial dimensionality. + +Unlike traditional transformer models, Tamba adjusts the parameters of its state space model (SSM) based on the input data, allowing it to selectively retain or disregard different pieces of information within the sequence. Mathematically, Tamba uses a state space model to capture the temporal evolution of the input. Given an input sequence $\mathcal{P} = \{\mathcal{P}_1,\mathcal{P}_2,\dots ,\mathcal{P}_T\}$ , where $T$ represents the sequence length and $\mathcal{P}_t\in \mathbb{R}^d$ is the input at each time step $t$ , the SSM is governed by the following state evolution equation: + +$$ +\mathbf {h} _ {t + 1} = A \left(\mathcal {P} _ {t}\right) \mathbf {h} _ {t} + B \left(\mathcal {P} _ {t}\right) \mathbf {u} _ {t}, \tag {2} +$$ + +Here $\mathbf{h}_t\in \mathbb{R}^n$ represents the hidden state at the time step $t$ , while $A(\mathcal{P}_t)\in \mathbb{R}^{n\times n}$ and $B(\mathcal{P}_t)\in \mathbb{R}^{n\times m}$ are matrices that depend on the input and evolve over time, and $\mathbf{u}_t\in \mathbb{R}^m$ represents the control input. + +The output of the time step $t$ is denoted as $\mathbf{y}_t\in \mathbb{R}^P$ , and is computed as follows: + +$$ +\mathbf {y} _ {t} = C \left(\mathcal {P} _ {t}\right) \mathbf {h} _ {t} + D \left(\mathcal {P} _ {t}\right) \mathbf {u} _ {t}. \tag {3} +$$ + +where $C(\mathcal{P}_t) \in \mathbb{R}^{p \times n}$ and $D(\mathcal{P}_t) \in \mathbb{R}^{p \times m}$ are also matrices that depend on the input. This formulation allows the system to dynamically adjust its parameters based on the current input, facilitating the selective retention and forgetting of information. The computational complexity of this method is linear, as the state update and output computation at each time step involve matrix-vector multiplication. The overall complexity is $O(T \cdot n^2)$ , where $T$ represents the sequence length and $n$ is the state dimension. + +Decomposed Attention Mechanism. We redesigned the self-attention computation, replacing the multi-head attention mechanism with state space model. After receiving the polyline encoding output $\mathcal{P}$ , Tamba performs the attention decomposition using three parallel encoders. The detailed computation process is as follows: + +(I) Spatiotemporal attention. The spatiotemporal attention for each agent $A_{i}$ and scene elements (S) over the + +![](images/1da23a05ea8800aff8817d54e2ffb153e03c1c129cd1c3cb723953140f3db37c.jpg) +Figure 3. Overview of our Tamba decoding process. The outputs of the three Atten-Tamba encoders carry the context of both the static scene and dynamic agent states, which are concatenated and input into the Cross-Tamba decoder. For predicting the target agent's trajectory, the decoder queries $K$ trajectory modes using independent query tensors, while the encoder's features combine the current state with recursive reasoning from the previous step, allowing all target agents to share a unified scene representation. During the secondary decoding, the predicted state of the proposed trajectory interacts once again with the current scene information, and a recurrent network assigns prediction weights to the proposed trajectories. The final output trajectory is obtained through refinement of these predictions. + +$\mathcal{P}$ at each time step $t$ can be represented as: $Z_{\mathrm{time}} = \operatorname{softmax}\left(\frac{Q_tK_t^\top}{\sqrt{d_k}}\right)V_t$ , where $Q_{t}, K_{t}, V_{t} \in \mathbb{R}^{T \times d_{k}}$ are the query, key, and value matrices at time step $t$ respectively, and $d_{k}$ represents the dimensionality of the key vectors. + +(II) Scene attention. For the attention between agent set $A$ and scene elements $S$ , let $A_{i} \in \mathbb{R}^{d_{a}}$ represent the embedding of agent $i$ , and $S_{j} \in \mathbb{R}^{d_{m}}$ represents the embedding of scene element $j$ . The attention calculation between agent $i$ and scene element $j$ is given by: $\alpha_{ij} = \frac{\exp(A_i^\top S_j)}{\sum_k \exp(A_i^\top S_k)}$ , where $\alpha_{ij}$ denotes the attention weight, indicating the relevance of scene elements $j$ to agent $i$ . The output is a weighted sum of the scene elements features denote as: + +$$ +Z _ {\text {s c e n e}} = \sum_ {j} \alpha_ {i j} \mathrm {S} _ {j}, \tag {4} +$$ + +(III) Traffic Attention. Similar to Eq. 4, we use $\beta_{ij}$ to represent the attention weight of traffic control elements (pedestrians and traffic lights) to other dynamic agents (e.g., motorcycles and cars). For a agent $i$ , the output feature of traffic attention can be expressed as: $Z_{\text{traffic}} = \sum_{j} \beta_{ij} A_j$ . This output is obtained by calculating the weighted sum of the embedded vectors $A_j$ of all dynamic agents, where the weight $\beta_{ij}$ reflects the impact of the traffic control elements on the dynamic agents. Through this interaction, the model can effectively focus on the influence of pedestrians on dynamic agent movement in non-traffic light scenarios. + +# 3.4. Cross Tamba Decoder + +The core challenge of trajectory decoding lies in addressing the 'one-to-many' problem, that is, how to generate $K$ output results (possibilities) from one set of input features. In designing the decoder architecture, we drew inspiration from DETR's approach [1] to solving the 'one-to-many' problem + +in object detection tasks. Specifically, the number of $K$ query vectors is initialized to represent different objects or candidate future trajectories. Through multiple iterations of the attention mechanism, each query vector extracts relevant information from the input features to produce $K$ output results. + +Crossed Dynamic and Static Attention. In the proposed cross-Tamba decoder, we designed it by using state space model to replace multi-head attention in the cross-attention mechanism. The query vector $Q$ is separated to interact independently with the keys and values from the encoder's output features. Specifically, the query vector is responsible for querying the modes of $K$ trajectories, while $K$ and $V$ incorporate both the encoded features from the encoder and the reasoning from the previous recursive step. This design allows dynamic agents (e.g., vehicles and motorcycles) to better share mixed features (pedestrians and traffic lights), while all target agents share a single static scene representation and iteratively generate $K$ different proposed trajectories over $T'$ future recursive steps. During the training phase, the proposed trajectories are optimized using the MSE loss ( $L_{\text{proposal}}$ ) based on the provided labels. + +Trajectory Proposal. After proposing $K$ predicted trajectories, we introduce a prediction weight inference module to compute the weight for each trajectory. In this process, we also use the Cross-Tamba module, which first fuses scene (S) with historical information to infer the likelihood of each predicted trajectory. Specifically, given each trajectory $p_k$ (where $k = 1,2,\dots,K$ ), we can represent it as: + +$$ +p _ {k} = \text {C r o s s - T a m b a} (\text {S c e n e}, \text {P r e d i c t i o n s}). \tag {5} +$$ + +where Scene represents the feature representation of the current scene and Predictions contains all predicted proposal trajectories. At the head of the weight inference module, we + +use a recurrent neural network (RNN) [24] to extract features from the final trajectories in order to generate a score for each trajectory. This RNN module can be represented as: + +$$ +h _ {t} = \operatorname {R N N} \left(h _ {t - 1}, p _ {k}\right), \tag {6} +$$ + +where $h_t$ is the hidden state at time step $t$ , $h_{t-1}$ is the previous state. Through this process, the RNN provides a mechanism for constraints across trajectories. When a trajectory conflicts with the environment or when there is an alternative trajectory that better fits the overall distribution, the RNN generates a prior confidence score for these trajectories. The confidence generated by the RNN can be represented as: $C_k = f(h_T)$ , where $C_k$ is the confidence score for trajectory $p_k$ , $h_T$ is the hidden state of the RNN at the final time step, and $f$ is a nonlinear activation function. + +This step ensures greater stability in the optimization process, such that the proposal trajectory considers not only the current prediction results but also the contextual environment and historical information, thereby enhancing the overall performance and robustness of the model. + +Trajectory Refinement. We employed the same optimization approach as the baseline methods [39, 40] to perform mixture modeling of the predicted trajectories. The future trajectory of agent $i$ is modeled as a mixture of Laplace distributions, as given by: + +$$ +f \left(\left\{p _ {t} ^ {i} \right\} _ {t = 1} ^ {T ^ {\prime}}\right) = \sum_ {k = 1} ^ {K} \pi_ {i, k} \prod_ {t = 1} ^ {T ^ {\prime}} \operatorname {L a p l a c e} \left(p _ {t} ^ {i} \mid \mu_ {t, i, k}, b _ {t, i, k}\right), \tag {7} +$$ + +where $\{\pi_{i,k}\}_{k = 1}^{K}$ are the mixing coefficients, representing the significance of each mixture component in the overall model. For each time step $t$ , the Laplace density of the $k^{th}$ mixture component is defined by the location parameter $\mu t,i,k$ and the scale parameter $b_{t,i,k}$ , which reflect the center and spread of that component. We use a classification loss $L_{\mathrm{cls}}$ to optimize the mixing coefficients in Eq. 7 by minimizing their negative log-likelihood. It is important to note that during this process, we stop the gradients of the location and scale parameters, focusing solely on the optimization of the mixing coefficients. This allows us to effectively focus on the classification task without affecting the trajectory's location and scale settings. + +For optimizing the location and scale parameters, we adopt the "winner takes all" strategy [15], resulting in only the best-predicted trajectory from the proposal and refinement modules undergoing backpropagation. This approach ensures that the network is adjusted only based on the best predictions during training, thereby enhancing the accuracy and stability of the model. In addition, to prevent overfitting and improve model robustness, the refinement module halts the gradient propagation of the proposed trajectory. This strategy helps to effectively preserve the previously learned information during the adjustment of proposals. + +Finally, our loss function integrates three components: the trajectory proposal loss ( $L_{\mathrm{proposal}}$ ), the trajectory refinement loss ( $L_{\mathrm{refine}}$ ), and the classification loss ( $L_{\mathrm{cls}}$ ), enabling an end-to-end training process. The expression is given as: + +$$ +L _ {\text {t o t a l}} = L _ {\text {p r o p o s a l}} + L _ {\text {r e f i n e}} + \lambda L _ {\text {c l s}}. \tag {8} +$$ + +# 4. Experiments + +# 4.1. Datasets + +Our experiments were conducted on the Argoverse 1 [2] and Argoverse 2 [14, 33] datasets. It contains extensive real-world autonomous driving scenarios and are designed to advance the development of autonomous driving technologies. The Argoverse 1 dataset includes 360-degree sensor data from the Pittsburgh and Miami regions in the United States, covering various road types and complex urban environments. Sampled at a rate of $10\mathrm{Hz}$ , this dataset provides trajectory data of the ego vehicle over a 5-second time span, with 2 seconds of observation used to predict the path over the next 3 seconds, comprising over 30,000 such samples. Argoverse 2 further expands on the dataset by including richer sensor data, multimodal annotations, and extending the observation window. Specifically, it includes 11-second scenarios, with 5 seconds of observation followed by prediction of trajectories for the next 6 seconds. The dataset contains 250,000 scenarios and features a large object taxonomy comprising 10 non-overlapping classes, which is used for training and validating motion prediction models. + +# 4.2. Metrics + +We adopted evaluation methods consistent with prior research, using metrics such as $\min\mathrm{ADE}(K)$ , $\min\mathrm{FDE}(K)$ , $\min\mathrm{FDE}(K)$ , and $\mathrm{MR}(K)$ to measure model performance in terms of error. Specifically, $\min\mathrm{ADE}(K)$ calculates the average $L2$ distance between the ground truth and the most optimal trajectory among the $K$ predicted trajectories across all future time steps, thereby assessing the overall prediction accuracy of the model over multiple time steps. $\min\mathrm{FDE}(K)$ focuses on the prediction error at the final time step, measuring the model's ability to predict the final state accurately. $\mathrm{b-minFDE}(K)$ extends $\min\mathrm{FDE}(K)$ by incorporating probability weighting, using $(1 - \hat{\pi})^2$ as a weight to account for prediction uncertainty, where $\hat{\pi}$ represents the probability score assigned by the model to the optimal predicted trajectory. $\mathrm{MR}(K)$ measures the proportion of the $K$ predicted trajectories with an error greater than 2 meters, providing an assessment of the model's stability and the frequency of prediction failures. Together, these metrics offer a comprehensive evaluation of the model's prediction accuracy, stability, and uncertainty estimation. For the parameter $K$ , we followed conventional values of 6 and 1. When the + +
Methodb-minFDE6↓minADE6↓minFDE6↓MR6↓minADE1↓minFDE1↓MR1↓#Params (M)↓FLOPs(G)↓
THOMAS [6]2.160.881.510.201.954.710.64--
GoRela [4]2.010.761.480.221.824.620.66--
QML [28]1.950.691.390.191.844.980.629.39-
MTR [26]1.980.731.440.151.744.390.5865.78-
GANet [31]1.960.721.340.171.774.480.5961.7315.79
BANet [32]1.920.711.360.191.794.610.609.4911.93
QCNet [40]1.910.651.290.161.694.300.597.6645.3
Tamba (Ours)1.890.641.240.171.664.240.574.5427.3
+ +Table 1. Performance comparison on Argoverse 2 [14, 33] dataset, leaderboard ranked by b-minFDE6. Subscript 6 and 1 represents the number of predictions. Best performance is highlighted in bold. (M) represents the number of parameters in millions; (G) refers to GigaOps per second. - represents we were unable to reproduce the results with complete accuracy. Notably, all methods in the table are direct comparisons of the original approaches. We exclude additional improvements from engineering techniques such as ensembling. + +
Methodb-minFDE6↓minADE6↓minFDE6↓MR6↓#Params (M)↓
LaneGCN [16]2.060.871.360.16-
DenseTNT [10]1.980.881.280.13-
MTR [26]1.930.821.240.13-
SceneTransformer [22]1.890.801.230.1315.30
HiVT [39]1.840.771.170.13-
GANet [31]1.790.811.160.12-
Wayformer [21]1.740.771.160.12-
QCNet [40]1.690.731.070.117.66
Tamba (Ours)1.670.721.070.094.54
+ +Table 2. Quantitative results on the Argoverse 1 [2] dataset, leaderboard ranked by b-minFDE6. Subscript 6 represents the number of predictions. Best performance is highlighted in bold. (M) represents the number of parameters in millions. - represents it is unable to reproduce the results with complete accuracy. We exclude additional improvements from engineering techniques such as ensembling. + +number of predicted trajectories exceeds $K$ , the top $K$ trajectories with the highest probabilities are selected. In terms of lightweight evaluation, we used the number of parameters (M) and FLOPs(G) as the metrics. + +# 4.3. Settings + +We conducted all experiments under the same environment (batch size, epochs) using eight NVIDIA 3090TI cards with a batch size of 128. We used Adam for optimization, with an initial learning rate of 0.001. Training was conducted for a total of 50 epochs. We employed an adaptive learning rate adjustment strategy based on the validation set's performance during training. Specifically, when the validation accuracy did not improve for five consecutive epochs, the learning rate was reduced to 0.1 times its original value. + +# 4.4. Comparison with State of the Art + +We compare our method with other SOTA approaches across various evaluation metrics in Argoverse 2 dataset. As shown in Table 1, our method demonstrates superior performance on key metrics such as $\mathrm{minFDE}_6$ and $\mathrm{minADE}_6$ . Tamba achieves a b-minFDE $^6$ score of 1.79, compared to 1.91 for QCNet, 1.92 for BANet and 1.98 for MTR. Additionally, + +Tamba achieves $(\mathrm{MR}_1)$ of 0.57, which is lower than that of other methods, such as QCNet (0.59) and MTR (0.58), indicating more stable prediction accuracy. Furthermore, Tamba has a parameter count of $4.54\mathrm{M}$ , which is significantly lower than QCNet (7.66 M), even compare to non-transformer models such as BANet (9.49M) and GANet (61.73M), Tamba achieving reductions of $52.0\%$ and $92.7\%$ respectively, while still outperforming both models. Additionally, Tamba achieves 27.3 GigaOps per second, which is $40\%$ faster than QCNet's 45.3 GigaOps per second. These results clearly demonstrate that Tamba is competitive in terms of accuracy but also excels in lightweight design and efficiency, providing an excellent solution for complex task. + +On the Argoverse 1 dataset, our method outperforms other state-of-the-art (SOTA) methods across the board. On the leaderboard for the b-minFDE $_6$ metric, Tamba achieves a score of 1.67, while the second-best, QCNet, scores 1.69, and the third, BANet, scores 1.74. In terms of parameter count, our method requires 4.54M parameters, compared to 7.66M for QCNet, reflecting a $40\%$ reduction. This further demonstrates the superiority of our approach in balancing efficiency and performance. For detailed performance metrics, please refer to Table 2. + +![](images/381441827755affb18d307400dc2c79b778ec26edde955583934b2a4704f6db6.jpg) +(a) + +![](images/67ae8c454d91ebbbc9766677cd149d3b8f701f1d8c8f2562146f31e7bfe5b2f7.jpg) +(b) + +![](images/bf88a6e27dabc9abd2ddb487b13a6e9f589d77a2b3662b782e436c87ea8f87c1.jpg) +(c) + +![](images/d03ef04f7f3755592c01273a099274afdbc85c935acf9d18c898298324f99d8f.jpg) +Observation +Target agent + +![](images/dd347f734b1df551ffceb00f774dfb41a87277e92dc8d5b48b6c98cd0fbb2aa0.jpg) +Predicted trajectory + +![](images/36ba9fb57a9ebb7a9c6ffbba7d1b10f74126a40921077a71e6a0d423e6a28bc0.jpg) +Figure 4. Qualitative results on the Argoverse 2 validation set. The target agent marked as orange with its surrounding agents marked as cold white. We conduct four different traffic scenarios. (a). Multi-agents in straight-road scenario, (b). Multiple agents on roundabout road, (c). Vehicle avoidance after roundabout, (d). Mixed scenario involving pedestrians and vehicles. + +
Methodw/ jointw/o joint
minFDE6↓minADE6↓minFDE1↓minADE1↓minFDE6↓minADE6↓minFDE1↓minADE1↓
Ours + Attention1.0940.7633.0371.3761.0930.7283.0351.372
Ours + Mamba1.0910.7242.8991.2981.0830.7652.8211.292
Ours + Tamba0.9510.6732.7421.2650.9550.6802.7451.272
+ +Table 3. Ablation study on the components of the Tamba encoder-decoder. Experimental results are conducted on the Argoverse 2 [14, 33] validation set. + +# 4.5. Quantization Results + +We visualized the prediction results on the validation set. In some challenging sequences, the predictions made by the Tamba model are more reasonable and smoother compared to existing methods, effectively adhering to map constraints. In Figure 4, we present several multimodal prediction cases, the results demonstrate that the trajectories generated by Tamba are smoother and better adapted to complex road topologies, meeting the demands of vehicle navigation under diverse road conditions. + +# 5. Ablation Study + +We evaluated the effectiveness of the Tamba module. For comparison, we used both the traditional attention mechanism module and the Mamba module on the validation set. In addition, we assessed the efficacy of our proposed pedestrian-traffic light joint polyline encoding strategy. Table 3 presents the performance comparison of our proposed three methods (including Attention variant, Mamba variant, and Tamba variant) under conditions with and without joint + +polyline encoding (w/ joint vs. w/o joint). Specifically, the table lists the performance of each method on two evaluation metrics: minFDE and minADE, considering both six predicted trajectories (minFDE $_6$ and minADE $_6$ ) and one predicted trajectory (minFDE $_1$ and minADE $_1$ ). + +# 6. Conclusion + +This paper introduces the Tamba model, an efficient motion prediction model for the real-time and computational needs of autonomous driving. By employing a state space model (SSM) and a novel attention Tamba encoder-decoder, we reduce computational complexity and parameter count while maintaining state-of-the-art accuracy. Experiments on the Argoverse 1 and 2 datasets show that Tamba achieves superior inference speed and efficiency, with a fourfold reduction in FLOPs and over $40\%$ fewer parameters compared to transformer-based models. These advantages make Tamba highly suitable for high-dynamic environments requiring rapid responses. + +# 7. Acknowledgement + +This work is supported in part by the Europe Eureka Intelligence to Drive | Move-Save-Win project (with funding from the UKRI Innovate UK project under Grant No. 10071278) as well as the Horizon Europe COVER project, No. 101086228 (with funding from UKRI grant EP/Y028031/1). Kezhi Wang would like to acknowledge the support in part by the Royal Society Industry Fellow scheme. + +# References + +[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision (ECCV), pages 213-229. Springer, 2020. 5 +[2] Ming-Fang Chang, John W Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, and James Hays. Argoverse: 3d tracking and forecasting with rich maps. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6, 7 +[3] Yihua Cheng, Yaning Zhu, Zongji Wang, Hongquan Hao, Yongwei Liu, Shiqing Cheng, Xi Wang, and Hyung Jin Chang. What do you see in vehicle? comprehensive vision solution for in-vehicle gaze estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1 +[4] Alexander Cui, Sergio Casas, Kelvin Wong, Simon Suo, and Raquel Urtasun. Gorela: Go relative for viewpoint-invariant motion forecasting. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 7801-7807. IEEE, 2023. 7 +[5] Jiyang Gao, Chen Sun, Hang Zhao, Yi Shen, Dragomir Anguelov, Congcong Li, and Cordelia Schmid. Vectornet: Encoding hd maps and agent dynamics from vectorized representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11525-11533, 2020. 1, 2 +[6] Thomas Gilles, Stefano Sabatini, Dzmitry Tsishkou, Bogdan Stanciulescu, and Fabien Moutarde. Thomas: Trajectory heatmap output with learned multi-agent sampling. In International Conference on Learning Representations (ICLR), 2022. 7 +[7] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2 +[8] Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. Hippo: Recurrent memory with optimal polynomial projections. Advances in neural information processing systems, 33:1474-1487, 2020. 2 +[9] Albert Gu, Karan Goel, Ankit Gupta, and Christopher Ré. On the parameterization and initialization of diagonal state space models. Advances in Neural Information Processing Systems, 35:35971-35983, 2022. 2 +[10] Junru Gu, Chen Sun, and Hang Zhao. Densetnt: End-to-end trajectory prediction from dense goal sets. In Proceedings of + +the IEEE/CVF International Conference on Computer Vision, pages 15303-15312, 2021. 7 +[11] Ankit Gupta, Albert Gu, and Jonathan Berant. Diagonal state spaces are as effective as structured state spaces. Advances in Neural Information Processing Systems, 35:22982-22994, 2022. 2 +[12] Yizhou Huang, Yihua Cheng, and Kezhi Wang. Efficient driving behavior narration and reasoning on edge device using large language models. arXiv preprint arXiv:2409.20364, 2024. 1 +[13] Parth Kothari, Brian Sifringer, and Alexandre Alahi. Interpretable social anchors for human trajectory forecasting in crowds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15556-15566, 2021. 3 +[14] John Lambert and James Hays. Trust, but verify: Cross-modality fusion for hd map change detection. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks (NeurIPS Datasets and Benchmarks 2021), 2021. 6, 7, 8 +[15] Stefan Lee, Senthil Purushwalkam Shiva Prakash, Michael Cogswell, Viresh Ranjan, David Crandall, and Dhruv Batra. Stochastic multiple choice learning for training diverse deep ensembles. Advances in Neural Information Processing Systems, 29, 2016. 6 +[16] Ming Liang, Bin Yang, Rui Hu, Yun Chen, Renjie Liao, Song Feng, and Raquel Urtasun. Learning lane graph representations for motion forecasting. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 541-556. Springer, 2020. 1, 7 +[17] Mengmeng Liu, Hao Cheng, Lin Chen, Hellward Broszio, Jiangtao Li, Runjiang Zhao, Monika Sester, and Michael Ying Yang. Laformer: Trajectory prediction for autonomous driving with lane-aware scene constraints. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2039-2049, 2024. 1, 3 +[18] Yuejiang Liu, Qi Yan, and Alexandre Alahi. Social nce: Contrastive learning of socially-aware motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15118-15129, 2021. 2 +[19] Karttikeya Mangalam, Yang An, Harshayu Girase, and Jitendra Malik. From goals, waypoints & paths to long term human trajectory forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15233-15242, 2021. 3 +[20] Abduallah Mohamed, Kun Qian, Mohamed Elhoseiny, and Christian Claudel. Social-stgCNN: A social spatio-temporal graph convolutional neural network for human trajectory prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14424-14432, 2020. 2 +[21] Nigamaa Nayakanti, Rami Al-Rfou, Aurick Zhou, Kratarth Goel, Khaled S Refaat, and Benjamin Sapp. Wayformer: Motion forecasting via simple & efficient attention networks. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2980-2987. IEEE, 2023. 1, 7 + +[22] Jiquan Ngiam, Vijay Vasudevan, Benjamin Caine, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, et al. Scene transformer: A unified architecture for predicting future trajectories of multiple agents. In 2023 IEEE International Conference on Robotics and Automation (ICRA), 2022. 7 +[23] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory++: Dynamically-feasible trajectory forecasting with heterogeneous data. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVIII 16, pages 683-700. Springer, 2020. 2 +[24] Alex Sherstinsky. Fundamentals of recurrent neural network (rnn) and long short-term memory (lstm) network. *Physical D: Nonlinear Phenomena*, 404:132306, 2020. 6 +[25] Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Wei Tang, Nanning Zheng, and Gang Hua. Representing multimodal behaviors with mean location for pedestrian trajectory prediction. IEEE transactions on pattern analysis and machine intelligence, 2023. 3 +[26] Shaoshuai Shi, Li Jiang, Dengxin Dai, and Bernt Schiele. Mtr++: Multi-agent motion prediction with symmetric scene modeling and guided intention querying. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3, 7 +[27] Jimmy Smith, Shalini De Mello, Jan Kautz, Scott Linderman, and Wonmin Byeon. Convolutional state space models for long-range spatiotemporal modeling. Advances in Neural Information Processing Systems, 36, 2024. 2 +[28] Tong Su, Xishun Wang, and Xiaodong Yang. Qml for argoverse 2 motion forecasting challenge. arXiv preprint arXiv:2207.06553, 2022. 7 +[29] Balakrishnan Varadarajan, Ahmed Hefny, Avikalp Srivastava, Khaled S Refaat, Nigamaa Nayakanti, Andre Cornman, Kan Chen, Bertrand Douillard, Chi Pang Lam, Dragomir Anguelov, et al. Multipath++: Efficient information fusion and trajectory aggregation for behavior prediction. In 2022 International Conference on Robotics and Automation (ICRA), pages 7814-7821. IEEE, 2022. 1 +[30] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2 +[31] Mingkun Wang, Xinge Zhu, Changqian Yu, Wei Li, Yuexin Ma, Ruochun Jin, Xiaoguang Ren, Dongchun Ren, Mingxu Wang, and Wenjing Yang. Ganet: Goal area network for motion forecasting. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1609-1615. IEEE, 2023. 7 +[32] Zhepeng Wang, Feng Chen, Kanokphan Lertniphonphan, Siwei Chen, Jinyao Bao, Pengfei Zheng, Jinbao Zhang, Kaer Huang, and Tao Zhang. Technical report for argoverse challenges on unified sensor-based detection, tracking, and forecasting. arXiv preprint arXiv:2311.15615, 2023. 7 +[33] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, + +Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks (NeurIPS Datasets and Benchmarks 2021), 2021. 6, 7, 8 +[34] Conghao Wong, Beihao Xia, Ziqian Zou, Yulong Wang, and Xinge You. Socialcircle: Learning the angle-based social interaction representation for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19005-19015, 2024. 3 +[35] Maosheng Ye, Tongyi Cao, and Qifeng Chen. Tpcn: Temporal point cloud networks for motion forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11318-11327, 2021. 2 +[36] Ye Yuan, Xinshuo Weng, Yanglan Ou, and Kris M Kitani. Agentformer: Agent-aware transformers for socio-temporal multi-agent forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9813-9823, 2021. 2 +[37] Zeyu Zhang, Akide Liu, Ian Reid, Richard Hartley, Bohan Zhuang, and Hao Tang. Motion mamba: Efficient and long sequence motion generation. In European Conference on Computer Vision, pages 265-282. Springer, 2025. 2 +[38] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634-21643, 2024. 1 +[39] Zikang Zhou, Luyao Ye, Jianping Wang, Kui Wu, and Kejie Lu. Hivt: Hierarchical vector transformer for multi-agent motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8823-8833, 2022. 1, 6, 7 +[40] Zikang Zhou, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang. Query-centric trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17863-17873, 2023. 1, 3, 6, 7 +[41] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 2 \ No newline at end of file diff --git a/2025/Trajectory Mamba_ Efficient Attention-Mamba Forecasting Model Based on Selective SSM/images.zip b/2025/Trajectory Mamba_ Efficient Attention-Mamba Forecasting Model Based on Selective SSM/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..137c25ae347612369924a8ffe4383219795c6735 --- /dev/null +++ b/2025/Trajectory Mamba_ Efficient Attention-Mamba Forecasting Model Based on Selective SSM/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:182b1f4051028072f401cc08a4fd2492a14116a09378bf23d805371fb7858399 +size 327958 diff --git a/2025/Trajectory Mamba_ Efficient Attention-Mamba Forecasting Model Based on Selective SSM/layout.json b/2025/Trajectory Mamba_ Efficient Attention-Mamba Forecasting Model Based on Selective SSM/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ad4bcaaacee820da265149ea86cfd7b146b80130 --- /dev/null +++ b/2025/Trajectory Mamba_ Efficient Attention-Mamba Forecasting Model Based on Selective SSM/layout.json @@ -0,0 +1,9282 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 102, + 541, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 102, + 541, + 138 + ], + "spans": [ + { + "bbox": [ + 70, + 102, + 541, + 138 + ], + "type": "text", + "content": "Trajectory Mamba: Efficient Attention-Mamba Forecasting Model Based on Selective SSM" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 160, + 460, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 160, + 460, + 217 + ], + "spans": [ + { + "bbox": [ + 146, + 160, + 460, + 217 + ], + "type": "text", + "content": "Yizhou Huang" + }, + { + "bbox": [ + 146, + 160, + 460, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 160, + 460, + 217 + ], + "type": "text", + "content": ", Yihua Cheng" + }, + { + "bbox": [ + 146, + 160, + 460, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 160, + 460, + 217 + ], + "type": "text", + "content": ", Kezhi Wang" + }, + { + "bbox": [ + 146, + 160, + 460, + 217 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 146, + 160, + 460, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 160, + 460, + 217 + ], + "type": "text", + "content": "Brunel University of London, " + }, + { + "bbox": [ + 146, + 160, + 460, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 160, + 460, + 217 + ], + "type": "text", + "content": "University of Birmingham \n{byyh009, Kezhi.Wang}@brunel.ac.uk, y.cheng.2@bham.ac.uk \nhttps://github.com/Yizhou-H/Trajectory-Mamba-CVPR" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 245, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 245, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 152, + 245, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 270, + 296, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 619 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 619 + ], + "type": "text", + "content": "Motion prediction is crucial for autonomous driving, as it enables accurate forecasting of future vehicle trajectories based on historical inputs. This paper introduces Trajectory Mamba, a novel efficient trajectory prediction framework based on the selective state-space model (SSM). Conventional attention-based models face the challenge of computational costs that grow quadratically with the number of targets, hindering their application in highly dynamic environments. In response, we leverage the SSM to redesign the self-attention mechanism in the encoder-decoder architecture, thereby achieving linear time complexity. To address the potential reduction in prediction accuracy resulting from modifications to the attention mechanism, we propose a joint polyline encoding strategy to better capture the associations between static and dynamic contexts, ultimately enhancing prediction accuracy. Additionally, to balance prediction accuracy and inference speed, we adopted the decoder that differs entirely from the encoder. Through cross-state space attention, all target agents share the scene context, allowing the SSM to interact with the shared scene representation during decoding, thus inferring different trajectories over the next prediction steps. Our model achieves state-of-the-art results in terms of inference speed and parameter efficiency on both the Argoverse 1 and Argoverse 2 datasets. It demonstrates a four-fold reduction in FLOPs compared to existing methods and reduces parameter count by over " + }, + { + "bbox": [ + 55, + 270, + 296, + 619 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 55, + 270, + 296, + 619 + ], + "type": "text", + "content": " while surpassing the performance of the vast majority of previous methods. These findings validate the effectiveness of Trajectory Mamba in trajectory prediction tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 639, + 135, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 639, + 135, + 651 + ], + "spans": [ + { + "bbox": [ + 56, + 639, + 135, + 651 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 659, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 295, + 696 + ], + "type": "text", + "content": "In the rapid development of autonomous driving technology [3, 12, 17], accurate motion prediction plays a crucial role by forecasting future trajectories based on historical vehicle" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 245, + 553, + 389 + ], + "blocks": [ + { + "bbox": [ + 318, + 245, + 553, + 389 + ], + "lines": [ + { + "bbox": [ + 318, + 245, + 553, + 389 + ], + "spans": [ + { + "bbox": [ + 318, + 245, + 553, + 389 + ], + "type": "image", + "image_path": "0699c5c3ed435b426a368e436b3fd6c42cded92936a34d53a12a03dea185f68e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 397, + 555, + 464 + ], + "lines": [ + { + "bbox": [ + 313, + 397, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 555, + 464 + ], + "type": "text", + "content": "Figure 1. Illustration of proposed joint polyline encoding strategy, where we consider all factors that affect the movement of motor vehicles, jointly encoding pedestrians and traffic lights rather than categorizing the scene horizontally into static and dynamic context. Additionally, we decompose and encode the interactions between all agents and elements at each time step." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 486, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 605 + ], + "type": "text", + "content": "movement data. This task is not only vital for enhancing driving safety but also meets the stringent requirements for real-time system response. In recent years, learning-based approaches have been widely applied to motion prediction [5, 16, 29, 39, 40]. However, existing prediction methods often struggle to balance efficiency and accuracy, making it challenging to simultaneously meet performance and efficiency requirements in real-time applications. We attempt to analyze the underlying reasons from the following aspects and explore possible improvements:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "content": "(a). Improving Prediction Accuracy: In autonomous driving scenarios [38], there are numerous dynamic and static agents, and the diversity of this data presents significant challenges for motion prediction and model training. However, conventional methods [29, 39] often struggle to effectively fuse this heterogeneous data, particularly when it comes to integrating various features, which leads to suboptimal prediction accuracy. On a positive note, factorized attention-based transformers [21, 40] have made no" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 144, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 144, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 144, + 713 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12058" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 131 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 131 + ], + "type": "text", + "content": "table strides in decomposing heterogeneous data. However, these models often use a unified, horizontal approach when dealing with static scenes and dynamic agent features, and they fail to fully consider the deeper relationships between different scene elements and agent information." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 134, + 294, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 134, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 57, + 134, + 294, + 312 + ], + "type": "text", + "content": "(b). Enhancing Response Efficiency: Motion predictions are inherently recursive, implying that each prediction serves as an input for the next time step, leading to the progressive accumulation of errors over time. Notably, the advancement of sparse context encoding techniques [5, 35] has demonstrated significant potential in enhancing model efficiency. Nonetheless, prediction models based on attention mechanisms still face exponential growth in computational complexity and parameter size during recursive tasks due to inherent design constraints, which severely limit their efficiency and real-time capabilities in practical applications. Conversely, methods such as LSTM [18] are relatively more efficient, but numerous studies [20, 23, 36] have shown that they are incapable of achieving the predictive accuracy of transformer-based models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 316, + 294, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 316, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 57, + 316, + 294, + 529 + ], + "type": "text", + "content": "In response, we propose an Attention-Mamba framework for motion prediction, termed Trajectory Mamba (Tamba), to overcome the limitations discussed above. First, we design a multimodal encoder-decoder architecture based on the state space model [9, 11, 27], comprising three parallel encoders and a decoder. By introducing a selective state space, we redefined the computation of the self-attention mechanism, enabling the model to selectively process input information by parameterizing the input of the state space model (SSM), thereby allowing the model to focus on or ignore specific inputs. Specifically, the selective state space module processes local short-range features by receiving input from a one-dimensional convolutional layer, ensuring that these features can be selected or discarded before entering the SSM. We utilized local state space attention in place of the multi-head attention in the self-attention module, reducing the computational complexity of recursive reasoning from exponential to linear." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 533, + 294, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 294, + 699 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 294, + 699 + ], + "type": "text", + "content": "Secondly, to better integrate heterogeneous data features, we proposed a joint polyline encoding strategy. Its objective is to separate pedestrian agents from the dynamic agents and jointly encode them with traffic light information from the static scene (see Fig. 1). We observed that in uncertain motion prediction scenarios, the application of traffic regulations exerts a degree of implicit control over the potential movement states of agents. For example, pedestrian-priority traffic rules remain effective even in the absence of traffic lights, influencing vehicles and motorcycles alike. By jointly encoding pedestrians and traffic lights through a shared embedder, this jointly encoding strategy can independently interact with other dynamic agents to enhance the prediction accuracy of vehicle and motorcycle agents." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 702, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 702, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 69, + 702, + 294, + 712 + ], + "type": "text", + "content": "Lastly, we designed a prediction weight inference mod" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 316, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 156 + ], + "type": "text", + "content": "ule. By redesigning selective state space with a cross-attention mechanism, we utilized independent query tensors to recursively infer multimodal candidate trajectories through interactions with the encoder and the keys and values from the previous recursive step. These candidate trajectories were then evaluated using a recurrent neural network to assign trajectory scores." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 316, + 158, + 553, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 158, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 158, + 553, + 407 + ], + "type": "text", + "content": "Our contributions are summarized as: (1) We propose the Tamba framework for motion prediction, which leverages the state space model to redesign computations of selfattention, resulting in computational complexity reducing significantly. (2) We propose a joint polyline encoding strategy. By integrating strongly correlated polyline types using independent or shared embeddings, elements that impose similar behavioral patterns can better influence different agents, collectively contributing to the trajectory reasoning of these agents. (3) We redesign the Mamba decoder with crossattention, allowing the joint scene encoding to share a unified scene representation, which further reduces the model's complexity. At the final stage, we introduced a prediction weight inference module to refine the predicted trajectories and enhance prediction accuracy. (4) Experimental results demonstrate that our method achieves state-of-the-art (SOTA) performance on Argoverse 1 and 2 datasets. Compared to advanced methods, our method reduces the number of parameters by " + }, + { + "bbox": [ + 316, + 158, + 553, + 407 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 316, + 158, + 553, + 407 + ], + "type": "text", + "content": ", achieves a four-fold increase in FLOPs. This represents a new level of balance between efficiency and performance in our model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 316, + 422, + 404, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 404, + 434 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 404, + 434 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 442, + 422, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 442, + 422, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 442, + 422, + 454 + ], + "type": "text", + "content": "2.1. State Space Model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 461, + 553, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 461, + 553, + 652 + ], + "spans": [ + { + "bbox": [ + 316, + 461, + 553, + 652 + ], + "type": "text", + "content": "Recently, the Mamba [7] framework has rekindled interest in state-space models (SSMs) [9, 11, 27] due to its introduction of an input-dependent selection mechanism, which enhances their applicability and performance. SSMs are renowned for their capacity to handle long-range dependencies and offer superior memory and computational efficiency compared to transformer-based architectures [30]. Hippo [8], a seminal contribution to SSMs, established the efficacy of these models in long-range sequence modeling tasks. Subsequent research, such as Vision mamba [41], has further refined the mechanisms underlying SSMs, resulting in substantial performance improvements. SSMs have demonstrated their versatility and robustness across a diverse set of domains, including computer vision, video analysis, medical imaging, graph-based learning, point cloud processing, and recommendation systems." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 654, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 654, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 654, + 553, + 712 + ], + "type": "text", + "content": "In the domain of trajectory prediction, recent studies have indicated that Motion Mamba [37] achieves notable success in predicting human trajectory time series. Building on these advancements, this work integrates the state-space mechanisms of Mamba with attention mechanisms to address tra" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "text", + "content": "12059" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": "jectory prediction tasks in dynamic traffic scenarios, thereby leveraging both the efficiency of SSMs and the adaptability of attention-based models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 116, + 258, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 116, + 258, + 129 + ], + "spans": [ + { + "bbox": [ + 55, + 116, + 258, + 129 + ], + "type": "text", + "content": "2.2. Attention Trajectory Prediction Model" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 135, + 295, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 135, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 135, + 295, + 218 + ], + "type": "text", + "content": "In the field of trajectory prediction, methods [13, 19] based on the attention mechanism have been widely adopted to enhance the model's ability to capture and predict complex behaviors. These approaches leverage the capability of attention to focus selectively on important parts of the input, thereby capturing key temporal and spatial features that are crucial for predicting future trajectories." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 219, + 295, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 219, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 54, + 219, + 295, + 338 + ], + "type": "text", + "content": "Subsequently, several advanced approaches [17, 34] have utilized attention mechanisms for trajectory prediction specifically in multi-agent interaction environments, where understanding the interactions between different agents (such as vehicles, pedestrians, and other road users) is critical. These models employ social attention modules to focus on the movements and behaviors of surrounding vehicles, effectively improving the ability of the system to handle intricate interactive behaviors, which are a common aspect of real-world driving scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 339, + 295, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 295, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 295, + 447 + ], + "type": "text", + "content": "Moreover, attention mechanisms have also been successfully applied in models that integrate multimodal information [25, 26, 40]. These models dynamically weight and fuse different types of sensory inputs—such as image features from cameras and data from other sensors like LiDAR—through attention modules. This fusion process allows the model to better understand and represent complex traffic environments, effectively utilizing complementary data to enhance trajectory prediction performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 457, + 135, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 135, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 135, + 470 + ], + "type": "text", + "content": "3. Our Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 478, + 123, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 123, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 123, + 489 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 496, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 629 + ], + "type": "text", + "content": "In the autonomous driving scenario, the input consists of the predicted dynamic agent, its surrounding agents, and the static scene context. Let the historical state of the " + }, + { + "bbox": [ + 55, + 496, + 295, + 629 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 496, + 295, + 629 + ], + "type": "text", + "content": "-th agent at time step " + }, + { + "bbox": [ + 55, + 496, + 295, + 629 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 496, + 295, + 629 + ], + "type": "text", + "content": " be represented as: " + }, + { + "bbox": [ + 55, + 496, + 295, + 629 + ], + "type": "inline_equation", + "content": "s_i^t = (p_i^t,\\theta_i^t,t,\\nu_i^t)" + }, + { + "bbox": [ + 55, + 496, + 295, + 629 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 496, + 295, + 629 + ], + "type": "inline_equation", + "content": "p_i^t = (p_{i,x}^t,p_{i,y}^t)" + }, + { + "bbox": [ + 55, + 496, + 295, + 629 + ], + "type": "text", + "content": " denotes the spatial position, " + }, + { + "bbox": [ + 55, + 496, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\theta_i^t" + }, + { + "bbox": [ + 55, + 496, + 295, + 629 + ], + "type": "text", + "content": " represents the heading angle, " + }, + { + "bbox": [ + 55, + 496, + 295, + 629 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 496, + 295, + 629 + ], + "type": "text", + "content": " refers to the time step, and " + }, + { + "bbox": [ + 55, + 496, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\nu_{i}^{t}" + }, + { + "bbox": [ + 55, + 496, + 295, + 629 + ], + "type": "text", + "content": " denotes the velocity. The static scene context is represented by " + }, + { + "bbox": [ + 55, + 496, + 295, + 629 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 496, + 295, + 629 + ], + "type": "text", + "content": " polygons, which include elements such as lane boundaries, map edges, and sidewalks. The agent states " + }, + { + "bbox": [ + 55, + 496, + 295, + 629 + ], + "type": "inline_equation", + "content": "s_i^t" + }, + { + "bbox": [ + 55, + 496, + 295, + 629 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 55, + 496, + 295, + 629 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 496, + 295, + 629 + ], + "type": "text", + "content": " time steps within the observation window are used to predict " + }, + { + "bbox": [ + 55, + 496, + 295, + 629 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 496, + 295, + 629 + ], + "type": "text", + "content": " future trajectories over " + }, + { + "bbox": [ + 55, + 496, + 295, + 629 + ], + "type": "inline_equation", + "content": "T^{\\prime}" + }, + { + "bbox": [ + 55, + 496, + 295, + 629 + ], + "type": "text", + "content": " time steps." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "In this paper, we propose the Tamba model to optimize the efficiency and accuracy of motion forecasting tasks. Our input includes " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "s_i^t" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": " of dynamic and static agents, " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": " processed using a heterogeneous polyline encoding strategy to transform " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "s_i^t" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": " into higher-dimensional features. We leverage three Tamba encoders to separately decompose (I) attendations in each time step " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "(t)" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": " over all elements, (II) attendations between" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 73, + 553, + 220 + ], + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 220 + ], + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 220 + ], + "type": "image", + "image_path": "d561ba29edd4ce0fbb3275cf902483eacb131baa7b85efa177e812082709735e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 230, + 555, + 297 + ], + "lines": [ + { + "bbox": [ + 313, + 230, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 555, + 297 + ], + "type": "text", + "content": "Figure 2. Overview of our Tamba encoder. We employ joint polyline encoding strategy to integrate strongly correlated polyline information and used three parallel encoders to interact with and associate these features. We applied a linear projection to map the attention output of Tamba back to the same dimensions as the input and used normalization to enhance the stability of the output." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 316, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 437 + ], + "type": "text", + "content": "all agents (A) and scene (S), and (III) attentions of traffic control elements (pedestrians and traffic lights) to other dynamic agents (motorcycles and cars). Then, concatenate features from encoders as output to a Tamba decoder, where we utilize cross-attention mechanism to generate proposal trajectories. Subsequently, we fused scene encoding with historical information to perform a secondary decoding of the proposal trajectories for refinement. Finally, the refined trajectories were optimized using a \"winner-takes-all strategy\" to obtain trajectory output." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 437, + 554, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 509 + ], + "type": "text", + "content": "Our model contains a novel heterogeneous polyline encoding strategy along with a decomposed feature fusion method for both static and dynamic agents. A redesigned Tamba encoder-decoder architecture. A refinement process to output the final predicted trajectory. In the following sections, we sequentially introduce the components in Tamba." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 515, + 448, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 448, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 448, + 529 + ], + "type": "text", + "content": "3.2. Joint Polyline Encoding" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "In multimodal traffic environment prediction tasks, different types of polylines possess unique geometric characteristics and semantic information. This study proposes a novel heterogeneous polyline embedding method to facilitate the fusion of dynamic and static context features, thereby providing a richer and more efficient feature representation. Specifically, we assign either independent or shared embedders to each type of polyline to effectively model various categories of information. Specifically, we jointly encode polylines with strong information associations, such as pedestrian trajectories and traffic light signals, using a shared embedder. This is because agents (pedestrians) and traffic lights typically impose similar behavioral patterns on other agents (vehicles) and often participate together in traffic control. Thus, using a shared embedder reduces model complexity" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12060" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "and promotes feature sharing, collectively influencing vehicle trajectory reasoning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 97, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 97, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 97, + 296, + 228 + ], + "type": "text", + "content": "In contrast, independent embedders are used for polyline types with significant feature differences, such as lane lines and traffic signs in maps. Each embedder consists of a MLP layer and normalization layers to extract high-dimensional feature representations. The shared pedestrian and traffic light embedder uses a cross-category feature fusion layer to enhance the model's sensitivity to different information sources. Polylines retain sequential information through positional encoding before being input to the embedders for feature extraction, followed by cross-category feature interaction via the Tamba encoder." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 229, + 296, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 229, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 296, + 326 + ], + "type": "text", + "content": "The mathematical representation of this process is as follows: Given the individual polyline data " + }, + { + "bbox": [ + 55, + 229, + 296, + 326 + ], + "type": "inline_equation", + "content": "P = \\{p_1, p_2, \\ldots, p_n\\}" + }, + { + "bbox": [ + 55, + 229, + 296, + 326 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 229, + 296, + 326 + ], + "type": "inline_equation", + "content": "p_i \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 229, + 296, + 326 + ], + "type": "text", + "content": " represents a point on the polyline, " + }, + { + "bbox": [ + 55, + 229, + 296, + 326 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 229, + 296, + 326 + ], + "type": "text", + "content": " is the number of points, and " + }, + { + "bbox": [ + 55, + 229, + 296, + 326 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 229, + 296, + 326 + ], + "type": "text", + "content": " is the dimensionality, the embedding process is represented as: " + }, + { + "bbox": [ + 55, + 229, + 296, + 326 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\operatorname{Embed}(P)" + }, + { + "bbox": [ + 55, + 229, + 296, + 326 + ], + "type": "text", + "content": ". For the pedestrian and traffic light categories that share an embedder, the output feature representation " + }, + { + "bbox": [ + 55, + 229, + 296, + 326 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{joint}}" + }, + { + "bbox": [ + 55, + 229, + 296, + 326 + ], + "type": "text", + "content": " is computed as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 335, + 296, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 335, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 65, + 335, + 296, + 360 + ], + "type": "interline_equation", + "content": "\\mathcal {P} _ {\\text {j o i n t}} = \\text {F u s i o n} \\left(\\operatorname {E m b e d} \\left(\\mathcal {P} _ {\\text {p e d e s t r i a n}}\\right), \\operatorname {E m b e d} \\left(\\mathcal {P} _ {\\text {t r a f f i c}}\\right)\\right), \\tag {1}", + "image_path": "024b619e9f3c769386781f74c66446987c56cb2a46abaeb20d77a8a2ac504ce4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 360, + 296, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 296, + 397 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 360, + 296, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{pedestrian}}" + }, + { + "bbox": [ + 55, + 360, + 296, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 360, + 296, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{traffic}}" + }, + { + "bbox": [ + 55, + 360, + 296, + 397 + ], + "type": "text", + "content": " are the input polylines for pedestrians and traffic lights, respectively, and Fusion represents the feature fusion operation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 406, + 200, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 200, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 200, + 417 + ], + "type": "text", + "content": "3.3. Attention Tamba Encoder" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 425, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 295, + 496 + ], + "type": "text", + "content": "In the proposed encoder, the input features " + }, + { + "bbox": [ + 55, + 425, + 295, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 425, + 295, + 496 + ], + "type": "text", + "content": " are uniformly embedded through joint polyline encoding. The input feature shape of the Tamba encoder is " + }, + { + "bbox": [ + 55, + 425, + 295, + 496 + ], + "type": "inline_equation", + "content": "(n,L,d)" + }, + { + "bbox": [ + 55, + 425, + 295, + 496 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 425, + 295, + 496 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 425, + 295, + 496 + ], + "type": "text", + "content": " denotes the number of encoded agents or scene elements, " + }, + { + "bbox": [ + 55, + 425, + 295, + 496 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 425, + 295, + 496 + ], + "type": "text", + "content": " represents the temporal sequence length, and " + }, + { + "bbox": [ + 55, + 425, + 295, + 496 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 425, + 295, + 496 + ], + "type": "text", + "content": " indicates the feature dimensionality at each time step." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 498, + 296, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 617 + ], + "type": "text", + "content": "Specifically, the sequence " + }, + { + "bbox": [ + 55, + 498, + 296, + 617 + ], + "type": "inline_equation", + "content": "(L,d)" + }, + { + "bbox": [ + 55, + 498, + 296, + 617 + ], + "type": "text", + "content": " undergoes linear transformations to generate the query, key, and value matrices, each with dimensions " + }, + { + "bbox": [ + 55, + 498, + 296, + 617 + ], + "type": "inline_equation", + "content": "(L,d_{k})" + }, + { + "bbox": [ + 55, + 498, + 296, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 498, + 296, + 617 + ], + "type": "inline_equation", + "content": "d_{k}" + }, + { + "bbox": [ + 55, + 498, + 296, + 617 + ], + "type": "text", + "content": " is the feature dimensionality used in the attention mechanism. The dot-product similarity between the query and key matrices is computed to yield an attention weight matrix of shape " + }, + { + "bbox": [ + 55, + 498, + 296, + 617 + ], + "type": "inline_equation", + "content": "(L,L)" + }, + { + "bbox": [ + 55, + 498, + 296, + 617 + ], + "type": "text", + "content": ", which captures the interdependencies between all time steps. Subsequently, these attention weights are employed to compute a weighted sum of the value matrix, yielding an output of shape " + }, + { + "bbox": [ + 55, + 498, + 296, + 617 + ], + "type": "inline_equation", + "content": "(L,d_{k})" + }, + { + "bbox": [ + 55, + 498, + 296, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": "As the output feature dimensionality of the Tamba block " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "(d_k)" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": " may differ from the input feature dimensionality " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "(d)" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": ", a linear projection is employed to align the attention mechanism's output with the input dimensionality, resulting in " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "(L,d)" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": ". This alignment ensures compatibility for the residual connection. Subsequently, the original input and the linearly projected attention output are combined via elementwise addition, yielding a resultant tensor of shape " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "(L,d)" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "To further stabilize the output, a normalization operation is applied, preserving the same dimensionality of " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "(L,d)" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 97, + 553, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 97, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 97, + 553, + 275 + ], + "type": "text", + "content": "Subsequently, the features are processed by a feedforward neural network comprising two fully connected layers. The first layer projects the input to a higher-dimensional space " + }, + { + "bbox": [ + 313, + 97, + 553, + 275 + ], + "type": "inline_equation", + "content": "d_{ff}" + }, + { + "bbox": [ + 313, + 97, + 553, + 275 + ], + "type": "text", + "content": ", resulting in an output with dimensions " + }, + { + "bbox": [ + 313, + 97, + 553, + 275 + ], + "type": "inline_equation", + "content": "(L, d_{ff})" + }, + { + "bbox": [ + 313, + 97, + 553, + 275 + ], + "type": "text", + "content": ". The second layer then maps the expanded features back to the original dimensionality " + }, + { + "bbox": [ + 313, + 97, + 553, + 275 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 97, + 553, + 275 + ], + "type": "text", + "content": ", restoring the shape to " + }, + { + "bbox": [ + 313, + 97, + 553, + 275 + ], + "type": "inline_equation", + "content": "(L, d)" + }, + { + "bbox": [ + 313, + 97, + 553, + 275 + ], + "type": "text", + "content": ". The output of the feedforward network is combined with the original input through a residual connection, preserving the output shape of " + }, + { + "bbox": [ + 313, + 97, + 553, + 275 + ], + "type": "inline_equation", + "content": "(L, d)" + }, + { + "bbox": [ + 313, + 97, + 553, + 275 + ], + "type": "text", + "content": ". A normalization operation is then applied to ensure stability, maintaining the final output shape of " + }, + { + "bbox": [ + 313, + 97, + 553, + 275 + ], + "type": "inline_equation", + "content": "(L, d)" + }, + { + "bbox": [ + 313, + 97, + 553, + 275 + ], + "type": "text", + "content": " for the entire encoder module. Thus, within the encoder module, the temporal sequence length " + }, + { + "bbox": [ + 313, + 97, + 553, + 275 + ], + "type": "inline_equation", + "content": "(L)" + }, + { + "bbox": [ + 313, + 97, + 553, + 275 + ], + "type": "text", + "content": " remains constant, whereas the feature dimensionality " + }, + { + "bbox": [ + 313, + 97, + 553, + 275 + ], + "type": "inline_equation", + "content": "(d)" + }, + { + "bbox": [ + 313, + 97, + 553, + 275 + ], + "type": "text", + "content": " may briefly change during processing but ultimately returns to its initial dimensionality." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 275, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 555, + 384 + ], + "type": "text", + "content": "Unlike traditional transformer models, Tamba adjusts the parameters of its state space model (SSM) based on the input data, allowing it to selectively retain or disregard different pieces of information within the sequence. Mathematically, Tamba uses a state space model to capture the temporal evolution of the input. Given an input sequence " + }, + { + "bbox": [ + 313, + 275, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{\\mathcal{P}_1,\\mathcal{P}_2,\\dots ,\\mathcal{P}_T\\}" + }, + { + "bbox": [ + 313, + 275, + 555, + 384 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 275, + 555, + 384 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 275, + 555, + 384 + ], + "type": "text", + "content": " represents the sequence length and " + }, + { + "bbox": [ + 313, + 275, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 275, + 555, + 384 + ], + "type": "text", + "content": " is the input at each time step " + }, + { + "bbox": [ + 313, + 275, + 555, + 384 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 275, + 555, + 384 + ], + "type": "text", + "content": ", the SSM is governed by the following state evolution equation:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 377, + 391, + 553, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 391, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 377, + 391, + 553, + 403 + ], + "type": "interline_equation", + "content": "\\mathbf {h} _ {t + 1} = A \\left(\\mathcal {P} _ {t}\\right) \\mathbf {h} _ {t} + B \\left(\\mathcal {P} _ {t}\\right) \\mathbf {u} _ {t}, \\tag {2}", + "image_path": "45bc5c6efe3ddafee67d9218f51cea950c67a48c468746581cf3959225afdc63.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 411, + 553, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 553, + 458 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 313, + 411, + 553, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_t\\in \\mathbb{R}^n" + }, + { + "bbox": [ + 313, + 411, + 553, + 458 + ], + "type": "text", + "content": " represents the hidden state at the time step " + }, + { + "bbox": [ + 313, + 411, + 553, + 458 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 411, + 553, + 458 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 313, + 411, + 553, + 458 + ], + "type": "inline_equation", + "content": "A(\\mathcal{P}_t)\\in \\mathbb{R}^{n\\times n}" + }, + { + "bbox": [ + 313, + 411, + 553, + 458 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 411, + 553, + 458 + ], + "type": "inline_equation", + "content": "B(\\mathcal{P}_t)\\in \\mathbb{R}^{n\\times m}" + }, + { + "bbox": [ + 313, + 411, + 553, + 458 + ], + "type": "text", + "content": " are matrices that depend on the input and evolve over time, and " + }, + { + "bbox": [ + 313, + 411, + 553, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_t\\in \\mathbb{R}^m" + }, + { + "bbox": [ + 313, + 411, + 553, + 458 + ], + "type": "text", + "content": " represents the control input." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 459, + 553, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 553, + 483 + ], + "type": "text", + "content": "The output of the time step " + }, + { + "bbox": [ + 313, + 459, + 553, + 483 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 459, + 553, + 483 + ], + "type": "text", + "content": " is denoted as " + }, + { + "bbox": [ + 313, + 459, + 553, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_t\\in \\mathbb{R}^P" + }, + { + "bbox": [ + 313, + 459, + 553, + 483 + ], + "type": "text", + "content": ", and is computed as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 380, + 490, + 553, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 490, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 380, + 490, + 553, + 504 + ], + "type": "interline_equation", + "content": "\\mathbf {y} _ {t} = C \\left(\\mathcal {P} _ {t}\\right) \\mathbf {h} _ {t} + D \\left(\\mathcal {P} _ {t}\\right) \\mathbf {u} _ {t}. \\tag {3}", + "image_path": "7951b9d356c59a155578b67c31d1470d0206009b2c2dd9e870af002607a37793.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "type": "inline_equation", + "content": "C(\\mathcal{P}_t) \\in \\mathbb{R}^{p \\times n}" + }, + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "type": "inline_equation", + "content": "D(\\mathcal{P}_t) \\in \\mathbb{R}^{p \\times m}" + }, + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "type": "text", + "content": " are also matrices that depend on the input. This formulation allows the system to dynamically adjust its parameters based on the current input, facilitating the selective retention and forgetting of information. The computational complexity of this method is linear, as the state update and output computation at each time step involve matrix-vector multiplication. The overall complexity is " + }, + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "type": "inline_equation", + "content": "O(T \\cdot n^2)" + }, + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "type": "text", + "content": " represents the sequence length and " + }, + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "type": "text", + "content": " is the state dimension." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 618, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 553, + 689 + ], + "type": "text", + "content": "Decomposed Attention Mechanism. We redesigned the self-attention computation, replacing the multi-head attention mechanism with state space model. After receiving the polyline encoding output " + }, + { + "bbox": [ + 313, + 618, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 618, + 553, + 689 + ], + "type": "text", + "content": ", Tamba performs the attention decomposition using three parallel encoders. The detailed computation process is as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "(I) Spatiotemporal attention. The spatiotemporal attention for each agent " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "A_{i}" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": " and scene elements (S) over the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12061" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 72, + 553, + 203 + ], + "blocks": [ + { + "bbox": [ + 61, + 72, + 553, + 203 + ], + "lines": [ + { + "bbox": [ + 61, + 72, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 553, + 203 + ], + "type": "image", + "image_path": "1da23a05ea8800aff8817d54e2ffb153e03c1c129cd1c3cb723953140f3db37c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 212, + 555, + 279 + ], + "lines": [ + { + "bbox": [ + 54, + 212, + 555, + 279 + ], + "spans": [ + { + "bbox": [ + 54, + 212, + 555, + 279 + ], + "type": "text", + "content": "Figure 3. Overview of our Tamba decoding process. The outputs of the three Atten-Tamba encoders carry the context of both the static scene and dynamic agent states, which are concatenated and input into the Cross-Tamba decoder. For predicting the target agent's trajectory, the decoder queries " + }, + { + "bbox": [ + 54, + 212, + 555, + 279 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 212, + 555, + 279 + ], + "type": "text", + "content": " trajectory modes using independent query tensors, while the encoder's features combine the current state with recursive reasoning from the previous step, allowing all target agents to share a unified scene representation. During the secondary decoding, the predicted state of the proposed trajectory interacts once again with the current scene information, and a recurrent network assigns prediction weights to the proposed trajectories. The final output trajectory is obtained through refinement of these predictions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 300, + 294, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 300, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 54, + 300, + 294, + 353 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 54, + 300, + 294, + 353 + ], + "type": "text", + "content": " at each time step " + }, + { + "bbox": [ + 54, + 300, + 294, + 353 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 300, + 294, + 353 + ], + "type": "text", + "content": " can be represented as: " + }, + { + "bbox": [ + 54, + 300, + 294, + 353 + ], + "type": "inline_equation", + "content": "Z_{\\mathrm{time}} = \\operatorname{softmax}\\left(\\frac{Q_tK_t^\\top}{\\sqrt{d_k}}\\right)V_t" + }, + { + "bbox": [ + 54, + 300, + 294, + 353 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 300, + 294, + 353 + ], + "type": "inline_equation", + "content": "Q_{t}, K_{t}, V_{t} \\in \\mathbb{R}^{T \\times d_{k}}" + }, + { + "bbox": [ + 54, + 300, + 294, + 353 + ], + "type": "text", + "content": " are the query, key, and value matrices at time step " + }, + { + "bbox": [ + 54, + 300, + 294, + 353 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 300, + 294, + 353 + ], + "type": "text", + "content": " respectively, and " + }, + { + "bbox": [ + 54, + 300, + 294, + 353 + ], + "type": "inline_equation", + "content": "d_{k}" + }, + { + "bbox": [ + 54, + 300, + 294, + 353 + ], + "type": "text", + "content": " represents the dimensionality of the key vectors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 353, + 295, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 353, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 54, + 353, + 295, + 453 + ], + "type": "text", + "content": "(II) Scene attention. For the attention between agent set " + }, + { + "bbox": [ + 54, + 353, + 295, + 453 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 54, + 353, + 295, + 453 + ], + "type": "text", + "content": " and scene elements " + }, + { + "bbox": [ + 54, + 353, + 295, + 453 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 54, + 353, + 295, + 453 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 54, + 353, + 295, + 453 + ], + "type": "inline_equation", + "content": "A_{i} \\in \\mathbb{R}^{d_{a}}" + }, + { + "bbox": [ + 54, + 353, + 295, + 453 + ], + "type": "text", + "content": " represent the embedding of agent " + }, + { + "bbox": [ + 54, + 353, + 295, + 453 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 54, + 353, + 295, + 453 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 353, + 295, + 453 + ], + "type": "inline_equation", + "content": "S_{j} \\in \\mathbb{R}^{d_{m}}" + }, + { + "bbox": [ + 54, + 353, + 295, + 453 + ], + "type": "text", + "content": " represents the embedding of scene element " + }, + { + "bbox": [ + 54, + 353, + 295, + 453 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 54, + 353, + 295, + 453 + ], + "type": "text", + "content": ". The attention calculation between agent " + }, + { + "bbox": [ + 54, + 353, + 295, + 453 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 54, + 353, + 295, + 453 + ], + "type": "text", + "content": " and scene element " + }, + { + "bbox": [ + 54, + 353, + 295, + 453 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 54, + 353, + 295, + 453 + ], + "type": "text", + "content": " is given by: " + }, + { + "bbox": [ + 54, + 353, + 295, + 453 + ], + "type": "inline_equation", + "content": "\\alpha_{ij} = \\frac{\\exp(A_i^\\top S_j)}{\\sum_k \\exp(A_i^\\top S_k)}" + }, + { + "bbox": [ + 54, + 353, + 295, + 453 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 353, + 295, + 453 + ], + "type": "inline_equation", + "content": "\\alpha_{ij}" + }, + { + "bbox": [ + 54, + 353, + 295, + 453 + ], + "type": "text", + "content": " denotes the attention weight, indicating the relevance of scene elements " + }, + { + "bbox": [ + 54, + 353, + 295, + 453 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 54, + 353, + 295, + 453 + ], + "type": "text", + "content": " to agent " + }, + { + "bbox": [ + 54, + 353, + 295, + 453 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 54, + 353, + 295, + 453 + ], + "type": "text", + "content": ". The output is a weighted sum of the scene elements features denote as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 460, + 295, + 486 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 460, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 135, + 460, + 295, + 486 + ], + "type": "interline_equation", + "content": "Z _ {\\text {s c e n e}} = \\sum_ {j} \\alpha_ {i j} \\mathrm {S} _ {j}, \\tag {4}", + "image_path": "4be4eb248666babb3d002586db0de524ef231604e0c84bf6142d2c85ea2d5697.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 495, + 295, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 495, + 295, + 627 + ], + "spans": [ + { + "bbox": [ + 54, + 495, + 295, + 627 + ], + "type": "text", + "content": "(III) Traffic Attention. Similar to Eq. 4, we use " + }, + { + "bbox": [ + 54, + 495, + 295, + 627 + ], + "type": "inline_equation", + "content": "\\beta_{ij}" + }, + { + "bbox": [ + 54, + 495, + 295, + 627 + ], + "type": "text", + "content": " to represent the attention weight of traffic control elements (pedestrians and traffic lights) to other dynamic agents (e.g., motorcycles and cars). For a agent " + }, + { + "bbox": [ + 54, + 495, + 295, + 627 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 54, + 495, + 295, + 627 + ], + "type": "text", + "content": ", the output feature of traffic attention can be expressed as: " + }, + { + "bbox": [ + 54, + 495, + 295, + 627 + ], + "type": "inline_equation", + "content": "Z_{\\text{traffic}} = \\sum_{j} \\beta_{ij} A_j" + }, + { + "bbox": [ + 54, + 495, + 295, + 627 + ], + "type": "text", + "content": ". This output is obtained by calculating the weighted sum of the embedded vectors " + }, + { + "bbox": [ + 54, + 495, + 295, + 627 + ], + "type": "inline_equation", + "content": "A_j" + }, + { + "bbox": [ + 54, + 495, + 295, + 627 + ], + "type": "text", + "content": " of all dynamic agents, where the weight " + }, + { + "bbox": [ + 54, + 495, + 295, + 627 + ], + "type": "inline_equation", + "content": "\\beta_{ij}" + }, + { + "bbox": [ + 54, + 495, + 295, + 627 + ], + "type": "text", + "content": " reflects the impact of the traffic control elements on the dynamic agents. Through this interaction, the model can effectively focus on the influence of pedestrians on dynamic agent movement in non-traffic light scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 635, + 182, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 182, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 182, + 647 + ], + "type": "text", + "content": "3.4. Cross Tamba Decoder" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 654, + 296, + 715 + ], + "type": "text", + "content": "The core challenge of trajectory decoding lies in addressing the 'one-to-many' problem, that is, how to generate " + }, + { + "bbox": [ + 54, + 654, + 296, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 654, + 296, + 715 + ], + "type": "text", + "content": " output results (possibilities) from one set of input features. In designing the decoder architecture, we drew inspiration from DETR's approach [1] to solving the 'one-to-many' problem" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 300, + 555, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 370 + ], + "type": "text", + "content": "in object detection tasks. Specifically, the number of " + }, + { + "bbox": [ + 313, + 300, + 555, + 370 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 300, + 555, + 370 + ], + "type": "text", + "content": " query vectors is initialized to represent different objects or candidate future trajectories. Through multiple iterations of the attention mechanism, each query vector extracts relevant information from the input features to produce " + }, + { + "bbox": [ + 313, + 300, + 555, + 370 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 300, + 555, + 370 + ], + "type": "text", + "content": " output results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 371, + 556, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 556, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 556, + 563 + ], + "type": "text", + "content": "Crossed Dynamic and Static Attention. In the proposed cross-Tamba decoder, we designed it by using state space model to replace multi-head attention in the cross-attention mechanism. The query vector " + }, + { + "bbox": [ + 313, + 371, + 556, + 563 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 371, + 556, + 563 + ], + "type": "text", + "content": " is separated to interact independently with the keys and values from the encoder's output features. Specifically, the query vector is responsible for querying the modes of " + }, + { + "bbox": [ + 313, + 371, + 556, + 563 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 371, + 556, + 563 + ], + "type": "text", + "content": " trajectories, while " + }, + { + "bbox": [ + 313, + 371, + 556, + 563 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 371, + 556, + 563 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 371, + 556, + 563 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 371, + 556, + 563 + ], + "type": "text", + "content": " incorporate both the encoded features from the encoder and the reasoning from the previous recursive step. This design allows dynamic agents (e.g., vehicles and motorcycles) to better share mixed features (pedestrians and traffic lights), while all target agents share a single static scene representation and iteratively generate " + }, + { + "bbox": [ + 313, + 371, + 556, + 563 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 371, + 556, + 563 + ], + "type": "text", + "content": " different proposed trajectories over " + }, + { + "bbox": [ + 313, + 371, + 556, + 563 + ], + "type": "inline_equation", + "content": "T'" + }, + { + "bbox": [ + 313, + 371, + 556, + 563 + ], + "type": "text", + "content": " future recursive steps. During the training phase, the proposed trajectories are optimized using the MSE loss (" + }, + { + "bbox": [ + 313, + 371, + 556, + 563 + ], + "type": "inline_equation", + "content": "L_{\\text{proposal}}" + }, + { + "bbox": [ + 313, + 371, + 556, + 563 + ], + "type": "text", + "content": ") based on the provided labels." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 563, + 556, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 556, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 556, + 647 + ], + "type": "text", + "content": "Trajectory Proposal. After proposing " + }, + { + "bbox": [ + 313, + 563, + 556, + 647 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 563, + 556, + 647 + ], + "type": "text", + "content": " predicted trajectories, we introduce a prediction weight inference module to compute the weight for each trajectory. In this process, we also use the Cross-Tamba module, which first fuses scene (S) with historical information to infer the likelihood of each predicted trajectory. Specifically, given each trajectory " + }, + { + "bbox": [ + 313, + 563, + 556, + 647 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 313, + 563, + 556, + 647 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 313, + 563, + 556, + 647 + ], + "type": "inline_equation", + "content": "k = 1,2,\\dots,K" + }, + { + "bbox": [ + 313, + 563, + 556, + 647 + ], + "type": "text", + "content": "), we can represent it as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 353, + 656, + 553, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 656, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 353, + 656, + 553, + 669 + ], + "type": "interline_equation", + "content": "p _ {k} = \\text {C r o s s - T a m b a} (\\text {S c e n e}, \\text {P r e d i c t i o n s}). \\tag {5}", + "image_path": "47efc5917eadcca98e23e755bf6e2fe37645404d324ac3065a34130b52580c05.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "where Scene represents the feature representation of the current scene and Predictions contains all predicted proposal trajectories. At the head of the weight inference module, we" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12062" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": "use a recurrent neural network (RNN) [24] to extract features from the final trajectories in order to generate a score for each trajectory. This RNN module can be represented as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 131, + 118, + 295, + 130 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 118, + 295, + 130 + ], + "spans": [ + { + "bbox": [ + 131, + 118, + 295, + 130 + ], + "type": "interline_equation", + "content": "h _ {t} = \\operatorname {R N N} \\left(h _ {t - 1}, p _ {k}\\right), \\tag {6}", + "image_path": "b896e049ab5a6789840a9cb4e568ba4672c13244c00b6a61405e9a6353c7723c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 138, + 295, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 138, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 138, + 295, + 257 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 138, + 295, + 257 + ], + "type": "inline_equation", + "content": "h_t" + }, + { + "bbox": [ + 55, + 138, + 295, + 257 + ], + "type": "text", + "content": " is the hidden state at time step " + }, + { + "bbox": [ + 55, + 138, + 295, + 257 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 138, + 295, + 257 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 138, + 295, + 257 + ], + "type": "inline_equation", + "content": "h_{t-1}" + }, + { + "bbox": [ + 55, + 138, + 295, + 257 + ], + "type": "text", + "content": " is the previous state. Through this process, the RNN provides a mechanism for constraints across trajectories. When a trajectory conflicts with the environment or when there is an alternative trajectory that better fits the overall distribution, the RNN generates a prior confidence score for these trajectories. The confidence generated by the RNN can be represented as: " + }, + { + "bbox": [ + 55, + 138, + 295, + 257 + ], + "type": "inline_equation", + "content": "C_k = f(h_T)" + }, + { + "bbox": [ + 55, + 138, + 295, + 257 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 138, + 295, + 257 + ], + "type": "inline_equation", + "content": "C_k" + }, + { + "bbox": [ + 55, + 138, + 295, + 257 + ], + "type": "text", + "content": " is the confidence score for trajectory " + }, + { + "bbox": [ + 55, + 138, + 295, + 257 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 55, + 138, + 295, + 257 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 138, + 295, + 257 + ], + "type": "inline_equation", + "content": "h_T" + }, + { + "bbox": [ + 55, + 138, + 295, + 257 + ], + "type": "text", + "content": " is the hidden state of the RNN at the final time step, and " + }, + { + "bbox": [ + 55, + 138, + 295, + 257 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 138, + 295, + 257 + ], + "type": "text", + "content": " is a nonlinear activation function." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 258, + 296, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 316 + ], + "type": "text", + "content": "This step ensures greater stability in the optimization process, such that the proposal trajectory considers not only the current prediction results but also the contextual environment and historical information, thereby enhancing the overall performance and robustness of the model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 318, + 296, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 296, + 377 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 296, + 377 + ], + "type": "text", + "content": "Trajectory Refinement. We employed the same optimization approach as the baseline methods [39, 40] to perform mixture modeling of the predicted trajectories. The future trajectory of agent " + }, + { + "bbox": [ + 55, + 318, + 296, + 377 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 318, + 296, + 377 + ], + "type": "text", + "content": " is modeled as a mixture of Laplace distributions, as given by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 65, + 385, + 295, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 385, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 65, + 385, + 295, + 418 + ], + "type": "interline_equation", + "content": "f \\left(\\left\\{p _ {t} ^ {i} \\right\\} _ {t = 1} ^ {T ^ {\\prime}}\\right) = \\sum_ {k = 1} ^ {K} \\pi_ {i, k} \\prod_ {t = 1} ^ {T ^ {\\prime}} \\operatorname {L a p l a c e} \\left(p _ {t} ^ {i} \\mid \\mu_ {t, i, k}, b _ {t, i, k}\\right), \\tag {7}", + "image_path": "c199977e87fa4eb338a4e07508a15634dcbd51e867fc33736c334ec71e9ff3e8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 426, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 582 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 426, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\{\\pi_{i,k}\\}_{k = 1}^{K}" + }, + { + "bbox": [ + 55, + 426, + 295, + 582 + ], + "type": "text", + "content": " are the mixing coefficients, representing the significance of each mixture component in the overall model. For each time step " + }, + { + "bbox": [ + 55, + 426, + 295, + 582 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 426, + 295, + 582 + ], + "type": "text", + "content": ", the Laplace density of the " + }, + { + "bbox": [ + 55, + 426, + 295, + 582 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 55, + 426, + 295, + 582 + ], + "type": "text", + "content": " mixture component is defined by the location parameter " + }, + { + "bbox": [ + 55, + 426, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\mu t,i,k" + }, + { + "bbox": [ + 55, + 426, + 295, + 582 + ], + "type": "text", + "content": " and the scale parameter " + }, + { + "bbox": [ + 55, + 426, + 295, + 582 + ], + "type": "inline_equation", + "content": "b_{t,i,k}" + }, + { + "bbox": [ + 55, + 426, + 295, + 582 + ], + "type": "text", + "content": ", which reflect the center and spread of that component. We use a classification loss " + }, + { + "bbox": [ + 55, + 426, + 295, + 582 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{cls}}" + }, + { + "bbox": [ + 55, + 426, + 295, + 582 + ], + "type": "text", + "content": " to optimize the mixing coefficients in Eq. 7 by minimizing their negative log-likelihood. It is important to note that during this process, we stop the gradients of the location and scale parameters, focusing solely on the optimization of the mixing coefficients. This allows us to effectively focus on the classification task without affecting the trajectory's location and scale settings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "For optimizing the location and scale parameters, we adopt the \"winner takes all\" strategy [15], resulting in only the best-predicted trajectory from the proposal and refinement modules undergoing backpropagation. This approach ensures that the network is adjusted only based on the best predictions during training, thereby enhancing the accuracy and stability of the model. In addition, to prevent overfitting and improve model robustness, the refinement module halts the gradient propagation of the proposed trajectory. This strategy helps to effectively preserve the previously learned information during the adjustment of proposals." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "Finally, our loss function integrates three components: the trajectory proposal loss (" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{proposal}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "), the trajectory refinement loss (" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{refine}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "), and the classification loss (" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{cls}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "), enabling an end-to-end training process. The expression is given as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 367, + 132, + 553, + 144 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 132, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 367, + 132, + 553, + 144 + ], + "type": "interline_equation", + "content": "L _ {\\text {t o t a l}} = L _ {\\text {p r o p o s a l}} + L _ {\\text {r e f i n e}} + \\lambda L _ {\\text {c l s}}. \\tag {8}", + "image_path": "d0c53c9ae72fbff49740ae9228232eed0e366fc22f6f67c6683102b9b48f21db.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 167, + 395, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 167, + 395, + 181 + ], + "spans": [ + { + "bbox": [ + 314, + 167, + 395, + 181 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 187, + 376, + 199 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 187, + 376, + 199 + ], + "spans": [ + { + "bbox": [ + 314, + 187, + 376, + 199 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 206, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 206, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 555, + 434 + ], + "type": "text", + "content": "Our experiments were conducted on the Argoverse 1 [2] and Argoverse 2 [14, 33] datasets. It contains extensive real-world autonomous driving scenarios and are designed to advance the development of autonomous driving technologies. The Argoverse 1 dataset includes 360-degree sensor data from the Pittsburgh and Miami regions in the United States, covering various road types and complex urban environments. Sampled at a rate of " + }, + { + "bbox": [ + 313, + 206, + 555, + 434 + ], + "type": "inline_equation", + "content": "10\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 206, + 555, + 434 + ], + "type": "text", + "content": ", this dataset provides trajectory data of the ego vehicle over a 5-second time span, with 2 seconds of observation used to predict the path over the next 3 seconds, comprising over 30,000 such samples. Argoverse 2 further expands on the dataset by including richer sensor data, multimodal annotations, and extending the observation window. Specifically, it includes 11-second scenarios, with 5 seconds of observation followed by prediction of trajectories for the next 6 seconds. The dataset contains 250,000 scenarios and features a large object taxonomy comprising 10 non-overlapping classes, which is used for training and validating motion prediction models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 443, + 373, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 443, + 373, + 454 + ], + "spans": [ + { + "bbox": [ + 314, + 443, + 373, + 454 + ], + "type": "text", + "content": "4.2. Metrics" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "content": "We adopted evaluation methods consistent with prior research, using metrics such as " + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\min\\mathrm{ADE}(K)" + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\min\\mathrm{FDE}(K)" + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\min\\mathrm{FDE}(K)" + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{MR}(K)" + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "content": " to measure model performance in terms of error. Specifically, " + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\min\\mathrm{ADE}(K)" + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "content": " calculates the average " + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "inline_equation", + "content": "L2" + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "content": " distance between the ground truth and the most optimal trajectory among the " + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "content": " predicted trajectories across all future time steps, thereby assessing the overall prediction accuracy of the model over multiple time steps. " + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\min\\mathrm{FDE}(K)" + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "content": " focuses on the prediction error at the final time step, measuring the model's ability to predict the final state accurately. " + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{b-minFDE}(K)" + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "content": " extends " + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\min\\mathrm{FDE}(K)" + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "content": " by incorporating probability weighting, using " + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "inline_equation", + "content": "(1 - \\hat{\\pi})^2" + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "content": " as a weight to account for prediction uncertainty, where " + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\pi}" + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "content": " represents the probability score assigned by the model to the optimal predicted trajectory. " + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{MR}(K)" + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "content": " measures the proportion of the " + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "content": " predicted trajectories with an error greater than 2 meters, providing an assessment of the model's stability and the frequency of prediction failures. Together, these metrics offer a comprehensive evaluation of the model's prediction accuracy, stability, and uncertainty estimation. For the parameter " + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "content": ", we followed conventional values of 6 and 1. When the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12063" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 550, + 182 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 550, + 182 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 550, + 182 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 550, + 182 + ], + "type": "table", + "html": "
Methodb-minFDE6↓minADE6↓minFDE6↓MR6↓minADE1↓minFDE1↓MR1↓#Params (M)↓FLOPs(G)↓
THOMAS [6]2.160.881.510.201.954.710.64--
GoRela [4]2.010.761.480.221.824.620.66--
QML [28]1.950.691.390.191.844.980.629.39-
MTR [26]1.980.731.440.151.744.390.5865.78-
GANet [31]1.960.721.340.171.774.480.5961.7315.79
BANet [32]1.920.711.360.191.794.610.609.4911.93
QCNet [40]1.910.651.290.161.694.300.597.6645.3
Tamba (Ours)1.890.641.240.171.664.240.574.5427.3
", + "image_path": "310c9b50766782ae433c923c83259ff419a010902f9ecbb0b1d43da51381dcd0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 126, + 245, + 486, + 370 + ], + "blocks": [ + { + "bbox": [ + 55, + 190, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 190, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 555, + 236 + ], + "type": "text", + "content": "Table 1. Performance comparison on Argoverse 2 [14, 33] dataset, leaderboard ranked by b-minFDE6. Subscript 6 and 1 represents the number of predictions. Best performance is highlighted in bold. (M) represents the number of parameters in millions; (G) refers to GigaOps per second. - represents we were unable to reproduce the results with complete accuracy. Notably, all methods in the table are direct comparisons of the original approaches. We exclude additional improvements from engineering techniques such as ensembling." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 126, + 245, + 486, + 370 + ], + "lines": [ + { + "bbox": [ + 126, + 245, + 486, + 370 + ], + "spans": [ + { + "bbox": [ + 126, + 245, + 486, + 370 + ], + "type": "table", + "html": "
Methodb-minFDE6↓minADE6↓minFDE6↓MR6↓#Params (M)↓
LaneGCN [16]2.060.871.360.16-
DenseTNT [10]1.980.881.280.13-
MTR [26]1.930.821.240.13-
SceneTransformer [22]1.890.801.230.1315.30
HiVT [39]1.840.771.170.13-
GANet [31]1.790.811.160.12-
Wayformer [21]1.740.771.160.12-
QCNet [40]1.690.731.070.117.66
Tamba (Ours)1.670.721.070.094.54
", + "image_path": "b322e2bfd7020233409f4733ab350f92268f361338a1ac7e2a1d7a1bc9a566da.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 379, + 555, + 413 + ], + "lines": [ + { + "bbox": [ + 55, + 379, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 555, + 413 + ], + "type": "text", + "content": "Table 2. Quantitative results on the Argoverse 1 [2] dataset, leaderboard ranked by b-minFDE6. Subscript 6 represents the number of predictions. Best performance is highlighted in bold. (M) represents the number of parameters in millions. - represents it is unable to reproduce the results with complete accuracy. We exclude additional improvements from engineering techniques such as ensembling." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 433, + 295, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 295, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 295, + 481 + ], + "type": "text", + "content": "number of predicted trajectories exceeds " + }, + { + "bbox": [ + 55, + 433, + 295, + 481 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 433, + 295, + 481 + ], + "type": "text", + "content": ", the top " + }, + { + "bbox": [ + 55, + 433, + 295, + 481 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 433, + 295, + 481 + ], + "type": "text", + "content": " trajectories with the highest probabilities are selected. In terms of lightweight evaluation, we used the number of parameters (M) and FLOPs(G) as the metrics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 489, + 116, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 489, + 116, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 116, + 502 + ], + "type": "text", + "content": "4.3. Settings" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 507, + 296, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 296, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 296, + 615 + ], + "type": "text", + "content": "We conducted all experiments under the same environment (batch size, epochs) using eight NVIDIA 3090TI cards with a batch size of 128. We used Adam for optimization, with an initial learning rate of 0.001. Training was conducted for a total of 50 epochs. We employed an adaptive learning rate adjustment strategy based on the validation set's performance during training. Specifically, when the validation accuracy did not improve for five consecutive epochs, the learning rate was reduced to 0.1 times its original value." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 623, + 235, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 235, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 235, + 635 + ], + "type": "text", + "content": "4.4. Comparison with State of the Art" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "We compare our method with other SOTA approaches across various evaluation metrics in Argoverse 2 dataset. As shown in Table 1, our method demonstrates superior performance on key metrics such as " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{minFDE}_6" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{minADE}_6" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": ". Tamba achieves a b-minFDE " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "^6" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " score of 1.79, compared to 1.91 for QCNet, 1.92 for BANet and 1.98 for MTR. Additionally," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 433, + 555, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 555, + 589 + ], + "type": "text", + "content": "Tamba achieves " + }, + { + "bbox": [ + 313, + 433, + 555, + 589 + ], + "type": "inline_equation", + "content": "(\\mathrm{MR}_1)" + }, + { + "bbox": [ + 313, + 433, + 555, + 589 + ], + "type": "text", + "content": " of 0.57, which is lower than that of other methods, such as QCNet (0.59) and MTR (0.58), indicating more stable prediction accuracy. Furthermore, Tamba has a parameter count of " + }, + { + "bbox": [ + 313, + 433, + 555, + 589 + ], + "type": "inline_equation", + "content": "4.54\\mathrm{M}" + }, + { + "bbox": [ + 313, + 433, + 555, + 589 + ], + "type": "text", + "content": ", which is significantly lower than QCNet (7.66 M), even compare to non-transformer models such as BANet (9.49M) and GANet (61.73M), Tamba achieving reductions of " + }, + { + "bbox": [ + 313, + 433, + 555, + 589 + ], + "type": "inline_equation", + "content": "52.0\\%" + }, + { + "bbox": [ + 313, + 433, + 555, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 433, + 555, + 589 + ], + "type": "inline_equation", + "content": "92.7\\%" + }, + { + "bbox": [ + 313, + 433, + 555, + 589 + ], + "type": "text", + "content": " respectively, while still outperforming both models. Additionally, Tamba achieves 27.3 GigaOps per second, which is " + }, + { + "bbox": [ + 313, + 433, + 555, + 589 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 433, + 555, + 589 + ], + "type": "text", + "content": " faster than QCNet's 45.3 GigaOps per second. These results clearly demonstrate that Tamba is competitive in terms of accuracy but also excels in lightweight design and efficiency, providing an excellent solution for complex task." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": "On the Argoverse 1 dataset, our method outperforms other state-of-the-art (SOTA) methods across the board. On the leaderboard for the b-minFDE" + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "inline_equation", + "content": "_6" + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": " metric, Tamba achieves a score of 1.67, while the second-best, QCNet, scores 1.69, and the third, BANet, scores 1.74. In terms of parameter count, our method requires 4.54M parameters, compared to 7.66M for QCNet, reflecting a " + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": " reduction. This further demonstrates the superiority of our approach in balancing efficiency and performance. For detailed performance metrics, please refer to Table 2." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12064" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 68, + 304, + 152 + ], + "blocks": [ + { + "bbox": [ + 58, + 68, + 304, + 152 + ], + "lines": [ + { + "bbox": [ + 58, + 68, + 304, + 152 + ], + "spans": [ + { + "bbox": [ + 58, + 68, + 304, + 152 + ], + "type": "image", + "image_path": "381441827755affb18d307400dc2c79b778ec26edde955583934b2a4704f6db6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 286, + 153, + 298, + 164 + ], + "lines": [ + { + "bbox": [ + 286, + 153, + 298, + 164 + ], + "spans": [ + { + "bbox": [ + 286, + 153, + 298, + 164 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 304, + 68, + 550, + 156 + ], + "blocks": [ + { + "bbox": [ + 304, + 68, + 550, + 156 + ], + "lines": [ + { + "bbox": [ + 304, + 68, + 550, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 68, + 550, + 156 + ], + "type": "image", + "image_path": "67ae8c454d91ebbbc9766677cd149d3b8f701f1d8c8f2562146f31e7bfe5b2f7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 536, + 153, + 549, + 164 + ], + "lines": [ + { + "bbox": [ + 536, + 153, + 549, + 164 + ], + "spans": [ + { + "bbox": [ + 536, + 153, + 549, + 164 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 169, + 303, + 252 + ], + "blocks": [ + { + "bbox": [ + 59, + 169, + 303, + 252 + ], + "lines": [ + { + "bbox": [ + 59, + 169, + 303, + 252 + ], + "spans": [ + { + "bbox": [ + 59, + 169, + 303, + 252 + ], + "type": "image", + "image_path": "bf88a6e27dabc9abd2ddb487b13a6e9f589d77a2b3662b782e436c87ea8f87c1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 286, + 257, + 298, + 269 + ], + "lines": [ + { + "bbox": [ + 286, + 257, + 298, + 269 + ], + "spans": [ + { + "bbox": [ + 286, + 257, + 298, + 269 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 304, + 169, + 550, + 272 + ], + "blocks": [ + { + "bbox": [ + 304, + 169, + 550, + 272 + ], + "lines": [ + { + "bbox": [ + 304, + 169, + 550, + 272 + ], + "spans": [ + { + "bbox": [ + 304, + 169, + 550, + 272 + ], + "type": "image", + "image_path": "d03ef04f7f3755592c01273a099274afdbc85c935acf9d18c898298324f99d8f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 274, + 363, + 285 + ], + "lines": [ + { + "bbox": [ + 310, + 274, + 363, + 285 + ], + "spans": [ + { + "bbox": [ + 310, + 274, + 363, + 285 + ], + "type": "text", + "content": "Observation" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 453, + 274, + 506, + 288 + ], + "lines": [ + { + "bbox": [ + 453, + 274, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 453, + 274, + 506, + 288 + ], + "type": "text", + "content": "Target agent" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 61, + 255, + 127, + 288 + ], + "blocks": [ + { + "bbox": [ + 61, + 255, + 127, + 288 + ], + "lines": [ + { + "bbox": [ + 61, + 255, + 127, + 288 + ], + "spans": [ + { + "bbox": [ + 61, + 255, + 127, + 288 + ], + "type": "image", + "image_path": "dd347f734b1df551ffceb00f774dfb41a87277e92dc8d5b48b6c98cd0fbb2aa0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 274, + 212, + 286 + ], + "lines": [ + { + "bbox": [ + 129, + 274, + 212, + 286 + ], + "spans": [ + { + "bbox": [ + 129, + 274, + 212, + 286 + ], + "type": "text", + "content": "Predicted trajectory" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 424, + 274, + 451, + 287 + ], + "blocks": [ + { + "bbox": [ + 424, + 274, + 451, + 287 + ], + "lines": [ + { + "bbox": [ + 424, + 274, + 451, + 287 + ], + "spans": [ + { + "bbox": [ + 424, + 274, + 451, + 287 + ], + "type": "image", + "image_path": "36ba9fb57a9ebb7a9c6ffbba7d1b10f74126a40921077a71e6a0d423e6a28bc0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 299, + 555, + 333 + ], + "lines": [ + { + "bbox": [ + 55, + 299, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 555, + 333 + ], + "type": "text", + "content": "Figure 4. Qualitative results on the Argoverse 2 validation set. The target agent marked as orange with its surrounding agents marked as cold white. We conduct four different traffic scenarios. (a). Multi-agents in straight-road scenario, (b). Multiple agents on roundabout road, (c). Vehicle avoidance after roundabout, (d). Mixed scenario involving pedestrians and vehicles." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 94, + 342, + 517, + 407 + ], + "blocks": [ + { + "bbox": [ + 94, + 342, + 517, + 407 + ], + "lines": [ + { + "bbox": [ + 94, + 342, + 517, + 407 + ], + "spans": [ + { + "bbox": [ + 94, + 342, + 517, + 407 + ], + "type": "table", + "html": "
Methodw/ jointw/o joint
minFDE6↓minADE6↓minFDE1↓minADE1↓minFDE6↓minADE6↓minFDE1↓minADE1↓
Ours + Attention1.0940.7633.0371.3761.0930.7283.0351.372
Ours + Mamba1.0910.7242.8991.2981.0830.7652.8211.292
Ours + Tamba0.9510.6732.7421.2650.9550.6802.7451.272
", + "image_path": "549d3007e97c93e528411c640f0113b080755a5da0cb5aaa91931ad622d46470.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 415, + 555, + 437 + ], + "lines": [ + { + "bbox": [ + 55, + 415, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 555, + 437 + ], + "type": "text", + "content": "Table 3. Ablation study on the components of the Tamba encoder-decoder. Experimental results are conducted on the Argoverse 2 [14, 33] validation set." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 458, + 177, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 458, + 177, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 458, + 177, + 470 + ], + "type": "text", + "content": "4.5. Quantization Results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 477, + 296, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 477, + 296, + 584 + ], + "spans": [ + { + "bbox": [ + 55, + 477, + 296, + 584 + ], + "type": "text", + "content": "We visualized the prediction results on the validation set. In some challenging sequences, the predictions made by the Tamba model are more reasonable and smoother compared to existing methods, effectively adhering to map constraints. In Figure 4, we present several multimodal prediction cases, the results demonstrate that the trajectories generated by Tamba are smoother and better adapted to complex road topologies, meeting the demands of vehicle navigation under diverse road conditions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 597, + 148, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 148, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 148, + 611 + ], + "type": "text", + "content": "5. Ablation Study" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "We evaluated the effectiveness of the Tamba module. For comparison, we used both the traditional attention mechanism module and the Mamba module on the validation set. In addition, we assessed the efficacy of our proposed pedestrian-traffic light joint polyline encoding strategy. Table 3 presents the performance comparison of our proposed three methods (including Attention variant, Mamba variant, and Tamba variant) under conditions with and without joint" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 459, + 554, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 554, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 554, + 519 + ], + "type": "text", + "content": "polyline encoding (w/ joint vs. w/o joint). Specifically, the table lists the performance of each method on two evaluation metrics: minFDE and minADE, considering both six predicted trajectories (minFDE" + }, + { + "bbox": [ + 313, + 459, + 554, + 519 + ], + "type": "inline_equation", + "content": "_6" + }, + { + "bbox": [ + 313, + 459, + 554, + 519 + ], + "type": "text", + "content": " and minADE" + }, + { + "bbox": [ + 313, + 459, + 554, + 519 + ], + "type": "inline_equation", + "content": "_6" + }, + { + "bbox": [ + 313, + 459, + 554, + 519 + ], + "type": "text", + "content": ") and one predicted trajectory (minFDE" + }, + { + "bbox": [ + 313, + 459, + 554, + 519 + ], + "type": "inline_equation", + "content": "_1" + }, + { + "bbox": [ + 313, + 459, + 554, + 519 + ], + "type": "text", + "content": " and minADE" + }, + { + "bbox": [ + 313, + 459, + 554, + 519 + ], + "type": "inline_equation", + "content": "_1" + }, + { + "bbox": [ + 313, + 459, + 554, + 519 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 544, + 388, + 557 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 544, + 388, + 557 + ], + "spans": [ + { + "bbox": [ + 314, + 544, + 388, + 557 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "This paper introduces the Tamba model, an efficient motion prediction model for the real-time and computational needs of autonomous driving. By employing a state space model (SSM) and a novel attention Tamba encoder-decoder, we reduce computational complexity and parameter count while maintaining state-of-the-art accuracy. Experiments on the Argoverse 1 and 2 datasets show that Tamba achieves superior inference speed and efficiency, with a fourfold reduction in FLOPs and over " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " fewer parameters compared to transformer-based models. These advantages make Tamba highly suitable for high-dynamic environments requiring rapid responses." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12065" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "7. Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 178 + ], + "type": "text", + "content": "This work is supported in part by the Europe Eureka Intelligence to Drive | Move-Save-Win project (with funding from the UKRI Innovate UK project under Grant No. 10071278) as well as the Horizon Europe COVER project, No. 101086228 (with funding from UKRI grant EP/Y028031/1). Kezhi Wang would like to acknowledge the support in part by the Royal Society Industry Fellow scheme." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 188, + 115, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 188, + 115, + 201 + ], + "spans": [ + { + "bbox": [ + 56, + 188, + 115, + 201 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 207, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 207, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 207, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 61, + 207, + 296, + 262 + ], + "type": "text", + "content": "[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision (ECCV), pages 213-229. Springer, 2020. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 263, + 296, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 263, + 296, + 328 + ], + "spans": [ + { + "bbox": [ + 61, + 263, + 296, + 328 + ], + "type": "text", + "content": "[2] Ming-Fang Chang, John W Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, and James Hays. Argoverse: 3d tracking and forecasting with rich maps. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 329, + 296, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 296, + 394 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 296, + 394 + ], + "type": "text", + "content": "[3] Yihua Cheng, Yaning Zhu, Zongji Wang, Hongquan Hao, Yongwei Liu, Shiqing Cheng, Xi Wang, and Hyung Jin Chang. What do you see in vehicle? comprehensive vision solution for in-vehicle gaze estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 395, + 296, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 395, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 62, + 395, + 296, + 449 + ], + "type": "text", + "content": "[4] Alexander Cui, Sergio Casas, Kelvin Wong, Simon Suo, and Raquel Urtasun. Gorela: Go relative for viewpoint-invariant motion forecasting. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 7801-7807. IEEE, 2023. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 449, + 296, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 296, + 514 + ], + "type": "text", + "content": "[5] Jiyang Gao, Chen Sun, Hang Zhao, Yi Shen, Dragomir Anguelov, Congcong Li, and Cordelia Schmid. Vectornet: Encoding hd maps and agent dynamics from vectorized representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11525-11533, 2020. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 515, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 515, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 62, + 515, + 294, + 569 + ], + "type": "text", + "content": "[6] Thomas Gilles, Stefano Sabatini, Dzmitry Tsishkou, Bogdan Stanciulescu, and Fabien Moutarde. Thomas: Trajectory heatmap output with learned multi-agent sampling. In International Conference on Learning Representations (ICLR), 2022. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 571, + 296, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 571, + 296, + 603 + ], + "spans": [ + { + "bbox": [ + 62, + 571, + 296, + 603 + ], + "type": "text", + "content": "[7] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 603, + 296, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 603, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 62, + 603, + 296, + 647 + ], + "type": "text", + "content": "[8] Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. Hippo: Recurrent memory with optimal polynomial projections. Advances in neural information processing systems, 33:1474-1487, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 647, + 296, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 296, + 691 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 296, + 691 + ], + "type": "text", + "content": "[9] Albert Gu, Karan Goel, Ankit Gupta, and Christopher Ré. On the parameterization and initialization of diagonal state space models. Advances in Neural Information Processing Systems, 35:35971-35983, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "text", + "content": "[10] Junru Gu, Chen Sun, and Hang Zhao. Densetnt: End-to-end trajectory prediction from dense goal sets. In Proceedings of" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "the IEEE/CVF International Conference on Computer Vision, pages 15303-15312, 2021. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[11] Ankit Gupta, Albert Gu, and Jonathan Berant. Diagonal state spaces are as effective as structured state spaces. Advances in Neural Information Processing Systems, 35:22982-22994, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "type": "text", + "content": "[12] Yizhou Huang, Yihua Cheng, and Kezhi Wang. Efficient driving behavior narration and reasoning on edge device using large language models. arXiv preprint arXiv:2409.20364, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 184, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 239 + ], + "type": "text", + "content": "[13] Parth Kothari, Brian Sifringer, and Alexandre Alahi. Interpretable social anchors for human trajectory forecasting in crowds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15556-15566, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 239, + 553, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 294 + ], + "type": "text", + "content": "[14] John Lambert and James Hays. Trust, but verify: Cross-modality fusion for hd map change detection. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks (NeurIPS Datasets and Benchmarks 2021), 2021. 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 294, + 553, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 350 + ], + "type": "text", + "content": "[15] Stefan Lee, Senthil Purushwalkam Shiva Prakash, Michael Cogswell, Viresh Ranjan, David Crandall, and Dhruv Batra. Stochastic multiple choice learning for training diverse deep ensembles. Advances in Neural Information Processing Systems, 29, 2016. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 350, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 415 + ], + "type": "text", + "content": "[16] Ming Liang, Bin Yang, Rui Hu, Yun Chen, Renjie Liao, Song Feng, and Raquel Urtasun. Learning lane graph representations for motion forecasting. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 541-556. Springer, 2020. 1, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 415, + 555, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 555, + 491 + ], + "type": "text", + "content": "[17] Mengmeng Liu, Hao Cheng, Lin Chen, Hellward Broszio, Jiangtao Li, Runjiang Zhao, Monika Sester, and Michael Ying Yang. Laformer: Trajectory prediction for autonomous driving with lane-aware scene constraints. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 2039-2049, 2024. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 492, + 553, + 537 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 492, + 553, + 537 + ], + "spans": [ + { + "bbox": [ + 316, + 492, + 553, + 537 + ], + "type": "text", + "content": "[18] Yuejiang Liu, Qi Yan, and Alexandre Alahi. Social nce: Contrastive learning of socially-aware motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15118-15129, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 537, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 592 + ], + "type": "text", + "content": "[19] Karttikeya Mangalam, Yang An, Harshayu Girase, and Jitendra Malik. From goals, waypoints & paths to long term human trajectory forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15233-15242, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 592, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 657 + ], + "type": "text", + "content": "[20] Abduallah Mohamed, Kun Qian, Mohamed Elhoseiny, and Christian Claudel. Social-stgCNN: A social spatio-temporal graph convolutional neural network for human trajectory prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14424-14432, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 658, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 714 + ], + "type": "text", + "content": "[21] Nigamaa Nayakanti, Rami Al-Rfou, Aurick Zhou, Kratarth Goel, Khaled S Refaat, and Benjamin Sapp. Wayformer: Motion forecasting via simple & efficient attention networks. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2980-2987. IEEE, 2023. 1, 7" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12066" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 148 + ], + "type": "text", + "content": "[22] Jiquan Ngiam, Vijay Vasudevan, Benjamin Caine, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, et al. Scene transformer: A unified architecture for predicting future trajectories of multiple agents. In 2023 IEEE International Conference on Robotics and Automation (ICRA), 2022. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 152, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 217 + ], + "type": "text", + "content": "[23] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory++: Dynamically-feasible trajectory forecasting with heterogeneous data. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVIII 16, pages 683-700. Springer, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 219, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 251 + ], + "type": "text", + "content": "[24] Alex Sherstinsky. Fundamentals of recurrent neural network (rnn) and long short-term memory (lstm) network. *Physical D: Nonlinear Phenomena*, 404:132306, 2020. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 253, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 306 + ], + "type": "text", + "content": "[25] Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Wei Tang, Nanning Zheng, and Gang Hua. Representing multimodal behaviors with mean location for pedestrian trajectory prediction. IEEE transactions on pattern analysis and machine intelligence, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 309, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 352 + ], + "type": "text", + "content": "[26] Shaoshuai Shi, Li Jiang, Dengxin Dai, and Bernt Schiele. Mtr++: Multi-agent motion prediction with symmetric scene modeling and guided intention querying. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 354, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 397 + ], + "type": "text", + "content": "[27] Jimmy Smith, Shalini De Mello, Jan Kautz, Scott Linderman, and Wonmin Byeon. Convolutional state space models for long-range spatiotemporal modeling. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 399, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 431 + ], + "type": "text", + "content": "[28] Tong Su, Xishun Wang, and Xiaodong Yang. Qml for argoverse 2 motion forecasting challenge. arXiv preprint arXiv:2207.06553, 2022. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 510 + ], + "type": "text", + "content": "[29] Balakrishnan Varadarajan, Ahmed Hefny, Avikalp Srivastava, Khaled S Refaat, Nigamaa Nayakanti, Andre Cornman, Kan Chen, Bertrand Douillard, Chi Pang Lam, Dragomir Anguelov, et al. Multipath++: Efficient information fusion and trajectory aggregation for behavior prediction. In 2022 International Conference on Robotics and Automation (ICRA), pages 7814-7821. IEEE, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "type": "text", + "content": "[30] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 557, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 621 + ], + "type": "text", + "content": "[31] Mingkun Wang, Xinge Zhu, Changqian Yu, Wei Li, Yuexin Ma, Ruochun Jin, Xiaoguang Ren, Dongchun Ren, Mingxu Wang, and Wenjing Yang. Ganet: Goal area network for motion forecasting. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1609-1615. IEEE, 2023. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "type": "text", + "content": "[32] Zhepeng Wang, Feng Chen, Kanokphan Lertniphonphan, Siwei Chen, Jinyao Bao, Pengfei Zheng, Jinbao Zhang, Kaer Huang, and Tao Zhang. Technical report for argoverse challenges on unified sensor-based detection, tracking, and forecasting. arXiv preprint arXiv:2311.15615, 2023. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "text", + "content": "[33] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes," + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 552 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 127 + ], + "type": "text", + "content": "Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks (NeurIPS Datasets and Benchmarks 2021), 2021. 6, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 553, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 193 + ], + "type": "text", + "content": "[34] Conghao Wong, Beihao Xia, Ziqian Zou, Yulong Wang, and Xinge You. Socialcircle: Learning the angle-based social interaction representation for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19005-19015, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 196, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 196, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 196, + 553, + 239 + ], + "type": "text", + "content": "[35] Maosheng Ye, Tongyi Cao, and Qifeng Chen. Tpcn: Temporal point cloud networks for motion forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11318-11327, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 241, + 553, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 294 + ], + "type": "text", + "content": "[36] Ye Yuan, Xinshuo Weng, Yanglan Ou, and Kris M Kitani. Agentformer: Agent-aware transformers for socio-temporal multi-agent forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9813-9823, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 297, + 553, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 340 + ], + "type": "text", + "content": "[37] Zeyu Zhang, Akide Liu, Ian Reid, Richard Hartley, Bohan Zhuang, and Hao Tang. Motion mamba: Efficient and long sequence motion generation. In European Conference on Computer Vision, pages 265-282. Springer, 2025. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 342, + 553, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 406 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 406 + ], + "type": "text", + "content": "[38] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634-21643, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 408, + 553, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 408, + 553, + 461 + ], + "spans": [ + { + "bbox": [ + 316, + 408, + 553, + 461 + ], + "type": "text", + "content": "[39] Zikang Zhou, Luyao Ye, Jianping Wang, Kui Wu, and Kejie Lu. Hivt: Hierarchical vector transformer for multi-agent motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8823-8833, 2022. 1, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 464, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 553, + 507 + ], + "type": "text", + "content": "[40] Zikang Zhou, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang. Query-centric trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17863-17873, 2023. 1, 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "type": "text", + "content": "[41] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 2" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12067" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TransPixeler_ Advancing Text-to-Video Generation with Transparency/a57aba62-4ac6-4f75-a668-a4f223b5c05f_content_list.json b/2025/TransPixeler_ Advancing Text-to-Video Generation with Transparency/a57aba62-4ac6-4f75-a668-a4f223b5c05f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c873fa5000f42a194902070e7a288793fb5cfea5 --- /dev/null +++ b/2025/TransPixeler_ Advancing Text-to-Video Generation with Transparency/a57aba62-4ac6-4f75-a668-a4f223b5c05f_content_list.json @@ -0,0 +1,1759 @@ +[ + { + "type": "text", + "text": "TransPixel: Advancing Text-to-Video Generation with Transparency", + "text_level": 1, + "bbox": [ + 142, + 130, + 854, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Luozhou Wang $^{1*}$ Yijun Li $^{3\\dagger}$ Zhifei Chen $^{1}$ Jui-Hsien Wang $^{3}$ Zhifei Zhang $^{3}$ He Zhang $^{3}$ Zhe Lin $^{3}$ Ying-Cong Chen $^{1,2‡}$ HKUST(GZ) ${}^{2}$ HKUST ${}^{3}$ Adobe Research", + "bbox": [ + 174, + 179, + 816, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"A transparent glass of water with ice cubes gently swirling\" \"A statue crumbling to dust as cracks spread across its surface\"", + "bbox": [ + 109, + 262, + 901, + 276 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fa8aca5e592d8e24b12576bf117460f0e6eefa10cc52f7a7789bcb1758982018.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 277, + 903, + 327 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"A squirrel's bushy tail flicking quickly\"", + "bbox": [ + 174, + 325, + 426, + 340 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"A small explosion rapidly expanding and contracting\"", + "bbox": [ + 535, + 325, + 870, + 339 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4a74fac0f2482707f8a80c4f6a995b0f40124e6b1f30dd18977e4f543d21ff5a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 339, + 903, + 386 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"A white dandelion shifting as seen through a magnifying glass\" \"A massive storm forming, with swirling clouds and thunderbolts\"", + "bbox": [ + 96, + 387, + 905, + 401 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2004066e27c07f6898c00ea7d3000c91b2d3555dd3634b34292669e97a1de2fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 401, + 903, + 450 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"A motorcycle drifting and swerving in an enchanted forest\"", + "bbox": [ + 109, + 450, + 486, + 464 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4e7c9e12ea80879b8a98b389c7161ced727c42c023a09982f198405cef2be474.jpg", + "image_caption": [ + "Figure 1. RGBA Video Generation with TransPixeler. By introducing LoRA layers into DiT-based text-to-video model with a novel alpha channel adaptive attention mechanism, our method enables RGBA video generation from text while preserving Text-to-Video quality." + ], + "image_footnote": [], + "bbox": [ + 94, + 464, + 903, + 513 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 564, + 326, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-video generative models have made significant strides, enabling diverse applications in entertainment, advertising, and education. However, generating RGBA video, which includes alpha channels for transparency, remains a challenge due to limited datasets and the difficulty of adapting existing models. Alpha channels are crucial for visual effects (VFX), allowing transparent elements like smoke and reflections to blend seamlessly into scenes. We introduce TransPixeler, a method to extend pretrained video models for RGBA generation while retaining the original RGB capabilities. TransPixeler leverages a diffusion transformer (DiT) architecture, incorporating alpha-specific tokens and using LoRA-based fine-tuning to jointly generate RGB and alpha channels with high consistency. By optimizing attention mechanisms, TransPixeler preserves the strengths of the original RGB model and achieves strong alignment between RGB and alpha channels despite limited training", + "bbox": [ + 88, + 597, + 485, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "data. Our approach effectively generates diverse and consistent RGBA videos, advancing the possibilities for VFX and interactive content creation. The code is available at https://wileewang.github.io/TransPixeler/.", + "bbox": [ + 511, + 566, + 906, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 662, + 645, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-Video generative models have quickly advanced, achieving impressive results [7, 16, 20, 26, 47, 49, 56, 61, 64]. This progress has enabled various applications, such as video editing [11, 13, 32, 39, 52, 55], image animation [2, 14, 15, 38], and motion customization [18, 24, 31, 36, 48, 51, 59]. Diffusion Transformers (DiT) enhance these models by using self-attention to capture long-range dependencies [3, 26, 56, 64]. These models are now widely used in entertainment, advertising, and education, meeting the demand for customizable, dynamic content. Notably, Text-to-RGBA (A denotes Alpha channel) video generation is invaluable for VFX and creative industries. The inclusion of an alpha channel in RGBA formats allows for transparent effects, enabling seamless blending of elements like smoke", + "bbox": [ + 509, + 689, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done during an internship at Adobe Research.", + "bbox": [ + 107, + 863, + 437, + 876 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Project Leader", + "bbox": [ + 109, + 876, + 196, + 888 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ Corresponding author", + "bbox": [ + 109, + 888, + 233, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18229", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and reflections (see Fig. 1). This transparency creates realistic visuals that can integrate smoothly into scenes without modifying the background. Such flexibility is crucial in gaming, virtual reality (VR), and augmented reality (AR), where dynamic and interactive content is in high demand.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Currently, no direct solutions exist for RGBA video generation, which remains a challenging task due to the scarcity of RGBA video data, with only around 484 videos available in [29]. This scarcity will significantly limit the diversity of generated content, resulting in a constrained set of object types and motion patterns. One feasible solution is to use video matting [28, 30, 40] to obtain alpha channels from generated videos. However, these methods are still limited by the scarcity of RGBA video data and struggle to generalize to a wider range of objects, as shown in Fig. 2 (b). Other video segmentation methods, such as SAM-2 [41], may generalize well to different tasks. However, they cannot generate alpha channels and are therefore unsuitable for direct compositing. There have been attempts to generate RGBA at the image level, such as LayerDiffusion [63]. However, adapting its concept directly to a temporal VAE used in video generative models remains challenging.", + "bbox": [ + 89, + 169, + 482, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we explore how to extend pretrained video models to generate corresponding alpha channels while retaining the original capabilities of pretrained models. Our goal is to generate content beyond the limitations of the current RGBA training set. Existing works such as Lotus [19] and Marigold [25] demonstrate that leveraging pretrained generation model weights significantly enhances out-of-distribution in dense prediction, hinting at the potential for predicting alpha channels. However, in the context of RGBA video generation, these approaches typically require generating RGB channels first, followed by separate alpha channel prediction. Consequently, information flows unidirectionally from RGB to alpha, keeping the two processes largely disconnected. Given the limited availability of RGBA video data, this imbalance results in insufficient alpha prediction when challenging objects are generated, as shown in Fig. 2 (c).", + "bbox": [ + 89, + 429, + 482, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose TransPixeler, which effectively adapts the pretrained RGB video models to generate RGB channels and the alpha channel simultaneously. We leverage state-of-the-art DiT-like video generation models [26, 56], and additionally introduce new tokens appended after text and RGB tokens for generating the alpha channels. To facilitate convergence, we reinitialize the positional embeddings for the alpha tokens and introduce a zero-initialized, learnable domain embedding to distinguish alpha tokens from RGB tokens. Furthermore, we employ a LoRA-based fine-tuning scheme [23], applied exclusively to project alpha tokens into the qkv space, thereby maintaining RGB generation quality. With the proposed approach, we extend the modality while preserving the original input", + "bbox": [ + 89, + 689, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/34008102bfd8108f497fd3fd41ebde0fcc0feff8e7d017d9886db7d99c250570.jpg", + "image_caption": [ + "\"A dust cloud expanding after an explosion, covering the area\"", + "(a) Generated RGB" + ], + "image_footnote": [], + "bbox": [ + 517, + 133, + 707, + 179 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/17e662660344896e39aa36eb2d635ff1a4771d6f004070fa93e42f1dceb82f6f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 90, + 901, + 135 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9da6bbff8b484f935d3ab969a2120a863ceeeaad75ede1c3bbf7dd6fcbda69e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 135, + 901, + 178 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d35bffe7a1d10ee01dd24fc1036a8b64a7e4aac9a88501f277919b96949a08e3.jpg", + "image_caption": [ + "(c) with generative prior (MariGold, Lotus)", + "Figure 2. Comparison between Generation-Then-Prediction and our Joint Generation approach. Given the generated RGB in (a), (b) and (c) show the predicted alpha (top) and the composited result (bottom). In (d), the top shows the jointly generated alpha." + ], + "image_footnote": [], + "bbox": [ + 517, + 189, + 707, + 279 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bbbfaad957332e8796d944f08bfba63473a5fc23707ff91956938563e333f6c0.jpg", + "image_caption": [ + "(b) w/o generative prior (Video Matting)", + "(d)Ours" + ], + "image_footnote": [], + "bbox": [ + 710, + 189, + 901, + 279 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "output structure and relying on the existing attention mechanism through LoRA adaptation.", + "bbox": [ + 511, + 366, + 903, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The extended sequence contains text, RGB, and alpha tokens, with self-attention divided into a 3x3 grouped attention matrix involving interactions like Text-attend-to-RGB (Text as query, RGB as key) and others. We also systematically analyze the attention mechanisms for RGBA generation: 1) Text-attend-to-RGB and RGB-attend-to-Text. The interaction between text and RGB tokens represents original model's generation capabilities. Minimizing the impact on text and RGB tokens during these attention computation processes can better retain the original model's performance; 2) RGB-attend-to-Alpha. We reveals a fundamental limitation in conventional methods is the lack of RGB-attend-to-Alpha attention. This attention is necessary to refine RGB tokens based on alpha information, improving RGB-alpha alignment; 3) Text-attend-to-Alpha. We remove this attention mechanism to reduce the risk caused by limited training data, which could degrade the model's performance. This removal also enhances the retention of the model's original capabilities.", + "bbox": [ + 511, + 397, + 905, + 683 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By integrating these techniques, our method achieves diverse RGBA generation with limited training data while maintaining strong RGB-alpha alignment. To summarize, our contributions are as follows:", + "bbox": [ + 511, + 685, + 903, + 744 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose an RGBA video generation framework using DiT models that requires limited data and training parameters, achieving diverse generation with strong alignment.", + "- We analyze the role of each attention component in the generation process, optimize their interactions, and introduce necessary modifications to improve RGBA generation quality.", + "- Our method is validated through extensive experiments, demonstrating its effectiveness across a variety of challenging scenarios." + ], + "bbox": [ + 511, + 747, + 903, + 897 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18230", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 91, + 89, + 232, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-Video Generation. Early video generation models were primarily based on Unet-based latent diffusion models (LDMs) extended from text-to-image models like Stable Diffusion [42]. For example,AnimateDiff [16] introduced a temporal attention module to improve temporal consistency across frames. Subsequent video generation models [5, 7, 8, 47, 61, 62] adopted an alternating approach with 2D spatial and 1D temporal attention, including works like ModelScope, VideoCrafter, Moonshot, and Show-1.", + "bbox": [ + 89, + 116, + 483, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With advancements in large language models (LLMs) and the introduction of Sora [3], attention shifted from Unet architectures to transformer-based architectures (DiT). DiT-based video generation models, such as Latte [37] and OpenSora [64], extended the DiT text-to-image (T2I) model [9] and maintained the 2D and 1D alternating attention approach, achieving promising results. Recently, DiT-based video generation has rapidly progressed, achieving further improvements in quality. Several methods [26, 44, 56] have moved away from the 2D and 1D alternating approach, instead treating video frames as a single long sequence with 3D positional embeddings for encoding. These approaches also prepend text tokens—processed through a text encoder—to the video sequence, creating a streamlined network that relies solely on full self-attention and feedforward layers. Our method builds upon these recent open-source transformer-based video generation models.", + "bbox": [ + 89, + 255, + 483, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Matting. A straightforward approach for RGBA video generation is to extract the alpha channel from generated RGB content, as done with traditional green screen keying or learning-based video matting expert models [28-30]. OmnimatteRF [28] introduces a video matting method that combines dynamic 2D foreground layers with a 3D background model, enabling more realistic scene reconstruction for real-world videos. Robust Video Matting (RVM) [30] proposes a real-time, high-quality human video matting method with a recurrent architecture for improved temporal coherence, achieving state-of-the-art results without auxiliary inputs. Another work presents a high-speed, high-resolution background replacement technique with precise alpha matte extraction, supported by the VideoMatte240K and PhotoMatte13K/85 datasets [29]. Additionally, many image matting methods [4, 6, 27, 50, 57] can be applied for frame-by-frame matting.", + "bbox": [ + 89, + 520, + 483, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Further, several works [19, 25, 53] in image depth estimation adapt pretrained generation models for prediction tasks, achieving strong performance that often surpasses traditional, scratch-trained expert models. Marigold [25] modifies architectures to create image-conditioned generation models, while Lotus [19] explores the role of the diffusion process in this context. Although there is currently no dedicated approach for video matting within video gen", + "bbox": [ + 89, + 779, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "eration models, we replicate and extend these methods to evaluate their performance, allowing us to highlight the limitations of prediction-based pipelines for RGBA generation.", + "bbox": [ + 511, + 90, + 903, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generation beyond RGB. Another category of methods [1, 17, 34, 35, 54, 60, 63] explores expanding generation models to simultaneously generate additional channels, though they are not specifically designed for RGBA video generation. For instance, LayerDiffusion [63] modifies the VAE in latent diffusion models to decode alpha channels. However, VAEs typically lack the semantic understanding required for precise alpha generation, limiting their effectiveness in complex visual scenarios where texture and contour details are critical. In contrast, other approaches [1, 34, 35, 60] modify the denoising model directly to enable joint generation. Wonder3D [34] uses a domain embedding to control the model's generation modality, while methods like IntrinsicDiffusion [35] and RGB $\\leftrightarrow$ X [60] adapt the UNet's input and output layers to jointly produce intrinsic modalities. However, all these methods are designed for image tasks and rely on UNet architectures. When applied to video generation, they face limitations in quality and diversity due to the scarcity of RGBA video data.", + "bbox": [ + 511, + 142, + 906, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 441, + 604, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 511, + 465, + 642, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first introduce the open-sourced state-of-the-art DiT-based video generation models [44, 56]. The core components of DiT-based video models are attention modules, and there are two primary distinctions between these models and previous approaches. On one hand, unlike previous models that alternate between 1D temporal attention and 2D spatial attention [5, 7, 8, 64], current methods typically employ 3D spatio-temporal attention, allowing them to capture spatio-temporal dependencies more effectively. On the other hand, instead of using cross-attention for text conditioning, these models concatenate text tokens $\\mathbf{x}_{\\mathrm{text}}$ with visual tokens $\\mathbf{x}_{\\mathrm{video}}$ into a single long sequence. The shape of video tokens and text tokens are $B\\times L\\times D$ and $B\\times L_{\\mathrm{text}}\\times D$ when $B$ equals to batch size, $L_{\\mathrm{text}}$ equals to the length of text tokens, $L$ equals to the length of video tokens and $D$ equals to the latent dimension of transformer. Full self-attention is then applied across the combined sequence:", + "bbox": [ + 511, + 489, + 906, + 746 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {T}}{\\sqrt {d _ {k}}}\\right) \\mathbf {V}, \\quad \\text {w h e r e} \\\\ \\mathbf {Z}: \\mathbf {Z} \\in \\{\\mathbf {Q}, \\mathbf {K}, \\mathbf {V} \\} \\\\ = \\left[ \\mathbf {W} _ {z: z \\in \\{q, k, v \\}} \\left(\\mathbf {x} _ {\\text {t e x t}}\\right); \\mathbf {f} _ {z: z \\in \\{q, k, v \\}} \\left(\\mathbf {x} _ {\\text {v i d e o}}\\right) \\right] \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 753, + 903, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here $\\mathbf{W}_t$ (for $t \\in \\{q, k, v\\}$ ) represents the projection matrix in the transformer model, and $\\mathbf{f}_t$ (for $t \\in \\{q, k, v\\}$ ) represents a combined operation that incorporates both the projection and positional encoding for visual tokens. There", + "bbox": [ + 511, + 840, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18231", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2a64c8a397ea52e395ebf393c55464d63b2bac93108dde1919ed4c9f786c78fa.jpg", + "image_caption": [ + "Figure 3. Pipeline of TransPixeler. Our method is organized as follows: (1) Left: we extend the input of DiT to include new alpha tokens; (2) Top Center: we initialize alpha tokens with our positional encoding; (3) Bottom Center: we insert a partial LoRA and adjust attention computation during training and inference." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 898, + 391 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "are two commonly used types of positional encoding. One is absolute positional encoding formulated as follows:", + "bbox": [ + 89, + 460, + 482, + 491 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {z: z \\in \\{q, k, v \\}} \\left(\\mathbf {x} _ {\\text {v i d e o}}\\right) := \\mathbf {W} _ {z: z \\in \\{q, k, v \\}} \\left(\\mathbf {x} _ {\\text {v i d e o}} ^ {m} + \\mathbf {p} ^ {m}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 500, + 482, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{p}$ is the positional embedding (e.g., a sinusoidal function) and $m$ denotes the position of each RGB video token. Another approach is the Rotary Position Embedding (RoPE) [43], often used by [44, 56]. This is expressed as", + "bbox": [ + 89, + 525, + 483, + 585 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {z: z \\in \\{q, k \\}} \\left(\\mathbf {x} _ {\\text {v i d e o}}\\right) := \\mathbf {W} _ {z: z \\in \\{q, k \\}} \\left(\\mathbf {x} _ {\\text {v i d e o}} ^ {m}\\right) \\circ e ^ {i m \\theta}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 593, + 482, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $m$ is the positional index, $i$ is the imaginary unit for rotation, and $\\theta$ is the rotation angle.", + "bbox": [ + 89, + 618, + 483, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Our Approach", + "text_level": 1, + "bbox": [ + 89, + 657, + 240, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To jointly generate RGB and alpha videos, we adapt a pretrained RGB video generation model through several modifications. The whole pipeline is visualized in Fig. 3.", + "bbox": [ + 89, + 680, + 482, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Firstly, we double the sequence length of noisy input tokens to enable the model to generate videos of double length, from $\\mathbf{x}_{\\mathrm{video}}^{1:L}$ to $\\mathbf{x}_{\\mathrm{video}}^{1:2*L}$ . Here, $\\mathbf{x}_{\\mathrm{video}}^{1:L}$ will be decoded into the RGB video, while $\\mathbf{x}_{\\mathrm{video}}^{L+1:2*L}$ will be decoded into the corresponding alpha video. The Query(Q), Key(K), Value(V) representations are formulated as:", + "bbox": [ + 89, + 726, + 483, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {Z}: \\mathbf {Z} \\in \\{\\mathbf {Q}, \\mathbf {K}, \\mathbf {V} \\} \\\\ = \\left[ \\mathbf {W} _ {z: z \\in \\{q, k, v \\}} \\left(\\mathbf {x} _ {\\text {t e x t}}\\right); \\mathbf {f} _ {z: z \\in \\{q, k, v \\}} \\left(\\mathbf {x} _ {\\text {v i d e o}} ^ {1: 2 * L}\\right) \\right] \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 825, + 483, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to sequence doubling, we explored increasing batch size or latent dimensions and splitting output into two", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "domains; however, these approaches showed limited effectiveness under constrained datasets, which we discuss later.", + "bbox": [ + 511, + 460, + 903, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Secondly, we modify the positional encoding function $\\mathbf{f}_{t:t\\in \\{q,k,v\\}}(\\cdot)$ , as shown in Fig. 4. Instead of continuously numbering indices, we allow RGB and alpha tokens to share the same positional encoding. Taking absolute positional encoding as an example:", + "bbox": [ + 511, + 491, + 905, + 566 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {f} _ {z: z \\in \\{q, k, v \\}} ^ {*} (\\mathbf {x} _ {\\mathrm {v i d e o}}) \\\\ := \\left\\{ \\begin{array}{l l} \\mathbf {W} _ {z: z \\in \\{q, k, v \\}} \\left(\\mathbf {x} _ {\\text {v i d e o}} ^ {m} + \\mathbf {p} ^ {m}\\right), & \\text {i f} m \\leq L, \\\\ \\mathbf {W} _ {z: z: z \\in \\{q, k, v \\}} ^ {*} \\left(\\mathbf {x} _ {\\text {v i d e o}} ^ {m} + \\mathbf {p} ^ {m - L} + d\\right), & \\text {i f} m > L. \\end{array} \\right. \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 579, + 903, + 641 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here we introduce a domain embedding $d$ , initialized to zero. We make it learnable to help the model adaptively differentiate between RGB ( $m \\leq L$ ) and alpha tokens ( $m > L$ ). The motivation behind this design is we observe that with same positional encoding, even initializing with different noises, the tokens from two domains tend to generate same results. It minimizes spatial-temporal alignment challenges at the very beginning of training and thus accelerates convergence.", + "bbox": [ + 511, + 651, + 906, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next we propose a fine-tuning scheme using LoRA [23], in which the LoRA layer is applied only to alpha domain tokens:", + "bbox": [ + 511, + 787, + 905, + 830 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {W} _ {z: z \\in \\{q, k, v \\}} ^ {*} (\\mathbf {x} _ {\\text {v i d e o}} ^ {m} + \\mathbf {p} ^ {m - L} + d) \\\\ = \\mathbf {W} _ {z: z \\in \\{q, k, v \\}} \\left(\\mathbf {x} _ {\\text {v i d e o}} ^ {m} + \\mathbf {p} ^ {m - L} + d\\right) \\tag {6} \\\\ + \\gamma \\cdot \\operatorname {L o R A} \\left(\\mathbf {x} _ {\\text {v i d e o}} ^ {m} + \\mathbf {p} ^ {m - L} + d\\right), \\quad \\text {i f} m > L, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 842, + 903, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18232", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2dfe9edaadb401c0c7b99995eef3ead0d14062fdf567400f84dc2e31ef8b9f45.jpg", + "image_caption": [ + "Figure 4. Positional Encoding Design for RGBA Generation. Assigning alpha tokens the same positional encoding as RGB yields similar results, resulting in faster convergence after 1000 iterations compared to standard encoding strategies." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 285, + 258 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fa90b265e7bc6f20c1513283d02a7d10f28b187c9ef294c62afbc29f15416826.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 289, + 89, + 483, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma$ controls the residual strength. Additionally, we design an attention mask to block unwanted attention computation. Given a text-video token sequence length $L_{\\mathrm{text}} + 2L$ , where $L_{\\mathrm{text}}$ represents text token length, the mask is defined as:", + "bbox": [ + 89, + 337, + 483, + 411 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {m n} ^ {*} = \\left\\{ \\begin{array}{l l} - \\infty , & \\text {i f} m \\leq L _ {\\text {t e x t}} \\text {a n d} n > L _ {\\text {t e x t}} + L, \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 417, + 483, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Combining these modifications, inference with our method is expressed as:", + "bbox": [ + 89, + 465, + 483, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Attention $(\\mathbf{Q},\\mathbf{K},\\mathbf{V}) =$ softmax $\\left(\\frac{\\mathbf{Q}\\mathbf{K}^T}{\\sqrt{d_k}} +\\mathbf{M}^*\\right)\\mathbf{V},$ where", + "bbox": [ + 89, + 503, + 496, + 539 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {Z}: \\mathbf {Z} \\in \\{\\mathbf {Q}, \\mathbf {K}, \\mathbf {V} \\} \\\\ = \\left[ \\mathbf {W} _ {z: z \\in \\{q, k, v \\}} \\left(\\mathbf {x} _ {\\text {t e x t}}\\right); \\mathbf {f} _ {z: z \\in \\{q, k, v \\}} ^ {*} \\left(\\mathbf {x} _ {\\text {v i d e o}}\\right) \\right] \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 541, + 483, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training is carried out using flow matching [33] or a traditional diffusion process [21].", + "bbox": [ + 89, + 590, + 483, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Analysis", + "text_level": 1, + "bbox": [ + 89, + 628, + 192, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given our goal of maximizing the inherited capabilities of the pretrained video model, enabling it to generate beyond the existing RGBA training set, we analyze the most critical component within our current 3D full attention DiT video generation model: the attention mechanism. The attention matrix, $\\mathbf{Q}\\mathbf{K}^T$ , has dimensions $(L_{\\mathrm{text}} + 2*L)\\times (L_{\\mathrm{text}} + 2*L)$ which we simplify by organizing it into a $3\\times 3$ grouped attention matrix—including Text-attend-to-RGB, RGB-attend-to-Text, and so forth, as illustrated in Fig. 3.", + "bbox": [ + 89, + 652, + 483, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Text-Attend-to-RGB and RGB-Attend-to-Text. These represent the upper-left 2x2 section of and are computations that exist solely in the original RGB generation model. If we ensure that this part of the computation remains unaffected, we can replicate the original RGB generation performance. Therefore, we limit the scope of LoRA's influence, as defined in Eq. (7), by retaining the original QKV", + "bbox": [ + 89, + 795, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c9b8423a27459dac1d0b53393a47298ee96f360a56ce64bdc65a2c93d3fe10b3.jpg", + "image_caption": [ + "Figure 5. Attention Rectification. (a) Eliminating all attention from alpha as a key preserves $100\\%$ RGB generation but leads to poor alignment. (b) Retaining all attention significantly degrades quality, causing a lack of motion in bicycles. (c) Our method achieves an effective balance." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 905, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "values for both text and RGB tokens, thus preserving the pretrained model's behavior in these domains.", + "bbox": [ + 511, + 387, + 905, + 416 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Besides the partial LoRA, the added alpha tokens requires the text and RGB tokens to also act as queries and interact with the alpha tokens as keys, which alters the computation in this 2x2 attention matrix. Therefore, we further analyze two additional attention computations that impact RGB generation, as shown in Fig. 5.", + "bbox": [ + 511, + 417, + 905, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Text-Attend-to-Alpha. We find that this attention is detrimental to the generation quality. Since the model was originally trained with text and RGB data, introducing attention from text to alpha causes interference due to the domain gap between alpha and RGB. Specifically, the alpha modality provides only contour information and lacks the rich texture, color, and semantic details associated with the text prompt, thereby degrading generation quality. To mitigate this, we design the attention mask (Eq. (7)) that blocks this computation.", + "bbox": [ + 511, + 513, + 905, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "RGB-Attend-to-Alpha. In contrast, we identify RGB-to-Alpha as essential for successful joint generation. This attention allows the model to refine RGB tokens by considering alpha information, facilitating alignment between generated RGB and alpha channels. This refinement process is a critical component missing in previous generation-then-prediction pipelines, which lacked a feedback mechanism for RGB refinement based on alpha guidance.", + "bbox": [ + 511, + 672, + 905, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 511, + 799, + 638, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Dataset. We utilize the public VideoMatte240K dataset [29], a comprehensive collection of 484 high-resolution green screen videos consists of 240,709 unique frames of alpha mattes and foregrounds. These frames provide a diverse range of human subjects, clothing styles, and", + "bbox": [ + 511, + 825, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18233", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "\"A coin spinning\"", + "bbox": [ + 243, + 90, + 354, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "\"A forest floor being consumed by spreading magical fire\"", + "bbox": [ + 524, + 90, + 880, + 104 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9ad2feb3bc0b485ab9216134a217a323e3b64244335dabe5c2968eb5ff6aafcc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 106, + 901, + 152 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "\"A parrot flying\"", + "bbox": [ + 246, + 152, + 348, + 166 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "\"An asteroid belt swirling chaotically through space\"", + "bbox": [ + 537, + 152, + 866, + 165 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3aa9427e38388071c24ba8ca88f69ad51f27ef890046422777b14dc2475743c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 164, + 901, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "\"An astronaut running down an alley, spacesuit flapping\"", + "bbox": [ + 117, + 213, + 478, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "\"A cloud of dust erupting and dispersing like an explosion\"", + "bbox": [ + 522, + 213, + 880, + 226 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4c672db82e44dc88d8e75fd27d55a402b8f03225126632b86bb6cee1b8e61837.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 226, + 901, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "\"a woman's long black hair streaming as she runs\"", + "bbox": [ + 143, + 275, + 455, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "\"Water splattering in mid-air\"", + "bbox": [ + 604, + 275, + 797, + 287 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/17e6491a7d7345d1c9fbf1fbdb199912f466c0bb76aa2ad4cab5ed8758ed8037.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 287, + 901, + 337 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Input Image", + "bbox": [ + 101, + 340, + 181, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Generated RGBA Video", + "bbox": [ + 267, + 340, + 411, + 352 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Input Image", + "bbox": [ + 516, + 340, + 594, + 352 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Generated RGBA Video", + "bbox": [ + 683, + 340, + 826, + 352 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cbcf13f23b412511108c5d7945343918b0bf1b15c8623e2ef802f2f894a0843d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 352, + 901, + 400 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/860d0ae2f5c5dc8142886fb2392e10b8db15807f99d218763bd64e876a1a7788.jpg", + "image_caption": [ + "Figure 6. Applications. Top: Text-to-Video with Transparency. Bottom: Image-to-Video generation with transparency." + ], + "image_footnote": [], + "bbox": [ + 94, + 402, + 901, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "poses. We apply fundamental preprocessing steps for them, including color decontamination and background blurring. Prompts are extracted using ShareGPT4V [10].", + "bbox": [ + 89, + 493, + 482, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model. Our RGBA video diffusion models are developed by fine-tuning pre-trained diffusion models. Specifically, we employ two models based on the diffusion transformer architecture: the open-source model CogVideoX [56] and a modified variant of CogVideoX denoted as $J$ . CogVideoX generates RGB videos in 480x720 with 49 frames at 8 FPS, using 50 sampling steps. In contrast, the modified version produces videos in 176x320 with 64 frames at 24 FPS, while also using 50 sampling steps. Additionally, we integrate our method with CogVideoX-I2V (Image-to-Video) to support image-to-video generation with transparency. We set the LoRA rank to 128. For domain embedding, we initialize it with an original shape of $1 \\times D$ and zero values, then expand it to $L \\times D$ through repetition during training. We train these parameters over 5,000 iterations with a batch size of 8 in total, utilizing 8 NVIDIA A100 GPUs.", + "bbox": [ + 89, + 544, + 483, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Applications", + "text_level": 1, + "bbox": [ + 89, + 795, + 223, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We mainly demonstrate two applications shown in Fig. 6:", + "bbox": [ + 89, + 818, + 470, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Text-to-Video with Transparency. Our method is capable of generating moving objects with various types of motion, such as spinning, running, and flying, while also handling transparent properties of bottles and glasses. Additionally,", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "it can produce complex visual effects, including fire, explosions, cracking, and lightning, as well as creative examples.", + "bbox": [ + 511, + 493, + 903, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Image-to-Video with Transparency. Our method can also be integrated with an I2V video generation model-CogVideoX-I2V. Users can provide a single image along with an alpha channel (optional), and then we generate subsequent frames with dynamic effects and automatically propagate or generate alpha channels for these frames.", + "bbox": [ + 511, + 529, + 905, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons", + "text_level": 1, + "bbox": [ + 511, + 628, + 653, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Generation-then-Prediction Pipeline. As shown in Fig. 2, video matting methods [29, 40, 58] struggle with matting non-human objects (see supplementary materials for additional results). Therefore, we selected Lotus [19] and SAM-2 [41] as baselines due to their stronger generalization: Lotus uses pretrained generative models, and SAM-2 is trained on large datasets. Since Lotus was originally designed for single-image depth estimation, we extended it for RGBA videos, denoted as Lotus + RGBA in our comparisons. Qualitative results are shown in Fig. 7. Since ground-truth alpha channels are not available for generated videos, we focus on qualitative comparison.", + "bbox": [ + 511, + 652, + 906, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Joint Generation Pipeline. Since there are currently no existing RGBA video generation models, we integrateAnimateDiff [16] with LayerDiffusion [63] to generate RGBA videos. We use the open-source video generation model", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18234", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5e794d65b0e64c08a69372d77a3984427867574a027fc1bd2c13d2864f7c0cb6.jpg", + "image_caption": [ + "Figure 7. Comparison with Generation-then-Prediction Pipelines. Our method demonstrates superior alignment." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 885, + 420 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c711bae9c19dcde842b045268589df63b130b192213524f7c7345e25b9a12ae7.jpg", + "image_caption": [ + "Figure 8. Comparison with Joint Generation Pipelines. Top: LayerDiffusion +AnimateDiff; Bottom: Ours. Our method achieves better alignment and generates corresponding motion described by prompts." + ], + "image_footnote": [], + "bbox": [ + 94, + 453, + 906, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CogVideoX [56] as the base model for fair comparison. The qualitative results are illustrated in Fig. 8.", + "bbox": [ + 89, + 748, + 485, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User Study. We also conduct a user study with Amazon Mechanical Turk to compare two joint generation methods, as shown in Table. 1. Participants are asked to evaluate two key aspects: 1) whether the RGB and alpha align correctly; and 2) whether the motion in the generated video matches the corresponding text description. A total of 30 videos are generated from distinct text prompts, and 87 users partici", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a7d14d6c5739910c0d57ec8fb202acb5a68c6db25d1c176f7b2329a53092097a.jpg", + "table_caption": [ + "Table 1. User Study." + ], + "table_footnote": [], + "table_body": "
RGBA AlignmentMotion Quality
AnimateDiff [16]+LayerDiff [63]6.7%21.7%
Ours + CogVideoX [56]93.3%78.3%
", + "bbox": [ + 514, + 762, + 903, + 810 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pated in the evaluation. The study shows that our method is obviously favored more by users with higher votes.", + "bbox": [ + 511, + 881, + 906, + 912 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18235", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ac4e7cabe7a36d1cf518758d685e1fef9c367ad148b7172bb82a725a6b8d8f60.jpg", + "image_caption": [ + "Figure 9. Alternative Designs for Joint Generation with DiT. Sequence extension (b) represents our method." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 480, + 276 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 340, + 243, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Fig. 10, we conduct the ablation study across two dimensions: attention rectification and network design.", + "bbox": [ + 89, + 364, + 482, + 395 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d6bf7fd2750b87cfbfd3643198d9261b8fc4726048d0f6b665e2e821ee28ca91.jpg", + "image_caption": [ + "Figure 10. Ablation Study. (a) Ours; (b) Ours without RGB-attend-to-Alpha; (c) Ours with Text-attend-to-alpha; (d) Batch Extension Strategy; (e) Latent Dimension Extension Strategy. Our method maintains high-quality motion generation (e.g., butterflies waving their wings) while achieving good alignment." + ], + "image_footnote": [], + "bbox": [ + 93, + 407, + 478, + 584 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Attention Rectification. By blocking RGB-to-Alpha attention, we first validate the importance of RGB-to-Alpha attention for aligning RGB and alpha channels, a feature lacking in most prediction-based methods. We also examine the effect of removing unnecessary attention to preserve the model's generative capacity, by learning Text-to-Alpha attention only. Without RGB-to-Alpha attention, the alpha channel misaligns with RGB content and the RGB output loses motion quality (e.g., reverse rocket).", + "bbox": [ + 89, + 685, + 482, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Alternative Designs For Joint Generation. Given the transformer's input dimensions $B \\times L \\times D$ , we extend the sequence dimension $L$ to produce RGB and alpha channels, but alternative extensions are possible at the Batch $B$ or Latent Dimension $D$ levels (see Fig. 9). In the Batch Ex", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1a36e93509c667fc9aed9e6448e8b4f1697c0d429cf9ed16f89ce4eb5b956bf3.jpg", + "image_caption": [ + "Figure 11. Quantitative Evaluation. Our approach achieves a good balance between alignment (low flow difference) and preserving generative quality (low FVD)." + ], + "image_footnote": [], + "bbox": [ + 532, + 89, + 883, + 219 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tension approach, a new module enables inter-batch communication, similar to the technique in [46]. For Latent Dimension Extension, we merge video and alpha tokens, project them into the DiT model's latent space, and unmerge post-generation, using learnable linear layers with fine-tuning. Batch Extension shows weaker RGB-alpha alignment, while Latent Dimension Extension, though akin to training from scratch, significantly reduces diversity.", + "bbox": [ + 511, + 281, + 906, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation. In addition to the qualitative comparisons shown in Fig. 10, we also generated a total of 80 videos, each consisting of 64 frames, and evaluated them using two primary metrics: Flow Difference. To measure alignment between the generated RGB and Alpha videos, we use optical flow [22] to focus on motion consistency while ignoring appearance. Specifically, we calculate optical flow with Farneback method [12] and compute the flow difference as the average Euclidean distance between RGB and Alpha flow fields. Frechet Video Distance (FVD). We use FVD [45] to compare the RGB videos generated by each RGBA method against those from the original RGB model, evaluating how well each method preserves the model's original generative quality. A lower FVD indicates that the generated results are closer to the original RGB model in terms of motion coherence and diversity, thus demonstrating a high fidelity to the model's intended generative quality. Results are shown in Fig. 11.", + "bbox": [ + 511, + 407, + 908, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 693, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce a novel Text-to-RGBA video generator that extends RGB diffusion models to produce RGBA output with minimal modifications and retained fidelity. Leveraging DiT backbones and RGBA-aware attention, the method preserves RGB quality while accurately inferring alpha channels. Targeted adjustments—adding alpha tokens, reinitializing positional embeddings, and applying selective LoRA fine-tuning—deliver rich, high-quality RGBA videos even from limited training data. Extensive quantitative and qualitative experiments validate the framework's versatility and robustness across challenging animation, complex compositing, and diverse real-world scenes.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18236", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Fan Bao, Shen Nie, Kaiwen Xue, Chongxuan Li, Shi Pu, Yaole Wang, Gang Yue, Yue Cao, Hang Su, and Jun Zhu. One transformer fits all distributions in multi-modal diffusion at scale. In International Conference on Machine Learning, pages 1692-1717. PMLR, 2023. 3", + "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1", + "[3] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 1, 3", + "[4] Ryan D Burgert, Brian L Price, Jason Kuen, Yijun Li, and Michael S Ryoo. Magick: A large-scale captioned dataset from matting generated images using chroma keying. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22595-22604, 2024. 3", + "[5] cerspense. zeroscope_v2. https://huggingface.co/ cerspense/zeroscope_v2_576w, 2023. Accessed: 2023-02-03. 3", + "[6] Guowei Chen, Yi Liu, Jian Wang, Juncai Peng, Yuying Hao, Lutao Chu, Shiyu Tang, Zewu Wu, Zeyu Chen, Zhiliang Yu, et al. Pp-matting: high-accuracy natural image matting. arXiv preprint arXiv:2204.09433, 2022. 3", + "[7] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 1, 3", + "[8] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocraft2: Overcoming data limitations for high-quality video diffusion models. arXiv preprint arXiv:2401.09047, 2024. 3", + "[9] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 3", + "[10] Lin Chen, Jisong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 6", + "[11] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023. 1", + "[12] Gunnar Farnebäck. Two-frame motion estimation based on polynomial expansion. In Proceedings of the Scandinavian Conference on Image Analysis (SCIA), pages 363-370. Springer, 2003. 8", + "[13] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 1" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Jianzhu Guo, Dingyun Zhang, Xiaogiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Liveportrait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 1", + "[15] Xun Guo, Mingwu Zheng, Liang Hou, Yuan Gao, Yufan Deng, Chongyang Ma, Weiming Hu, Zhengjun Zha, Haibin Huang, Pengfei Wan, et al. I2v-adapter: A general image-to-video adapter for video diffusion models. arXiv preprint arXiv:2312.16693, 2023. 1", + "[16] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 1, 3, 6, 7", + "[17] Hao He, Yixun Liang, Luozhou Wang, Yuanhao Cai, Xinli Xu, Hao-Xiang Guo, Xiang Wen, and Yingcong Chen. *Lucidfusion: Generating 3d gaussians with arbitrary unposed images*, 2024. 3", + "[18] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrl: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 1", + "[19] Jing He, Haodong Li, Wei Yin, Yixun Liang, Leheng Li, Kaiqiang Zhou, Hongbo Liu, Bingbing Liu, and YingCong Chen. Lotus: Diffusion-based visual foundation model for high-quality dense prediction. arXiv preprint arXiv:2409.18124, 2024. 2, 3, 6", + "[20] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 1", + "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 5", + "[22] Berthold KP Horn and Brian G Schunck. Determining optical flow. Artificial intelligence, 17(1-3):185–203, 1981. 8", + "[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 4", + "[24] Hyeonho Jeong, Jinho Chang, Geon Yeong Park, and Jong Chul Ye. Dreammotion: Space-time self-similarity score distillation for zero-shot video editing. arXiv preprint arXiv:2403.12002, 2024. 1", + "[25] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 2, 3", + "[26] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 1, 2, 3", + "[27] Jiachen Li, Jitesh Jain, and Humphrey Shi. Matting anything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1775-1785, 2024. 3", + "[28] Geng Lin, Chen Gao, Jia-Bin Huang, Changil Kim, Yipeng Wang, Matthias Zwicker, and Ayush Saraf. Omnimatterf: Robust omnimatte with 3d background modeling. In" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18237", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23471-23480, 2023. 2, 3", + "[29] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian L Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8762-8771, 2021. 2, 3, 5, 6", + "[30] Shanchuan Lin, Linjie Yang, Imran Saleemi, and Soumyadip Sengupta. Robust high-resolution video matting with temporal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 238-247, 2022. 2, 3", + "[31] Pengyang Ling, Jiazi Bu, Pan Zhang, Xiaoyi Dong, Yuhang Zang, Tong Wu, Huaian Chen, Jiaqi Wang, and Yi Jin. Motionclone: Training-free motion cloning for controllable video generation. arXiv preprint arXiv:2406.05338, 2024. 1", + "[32] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-p2p: Video editing with cross-attention control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8599–8608, 2024. 1", + "[33] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 5", + "[34] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9970-9980, 2024. 3", + "[35] Jundan Luo, Duygu Ceylan, Jae Shin Yoon, Nanxuan Zhao, Julien Philip, Anna Fruhstick, Wenbin Li, Christian Richardt, and Tuanfeng Wang. Intrinsicdiffusion: Joint intrinsic layers from latent diffusion models. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3", + "[36] Wan-Duo Kurt Ma, John P Lewis, and W Bastiaan Kleijn. Trailblazer: Trajectory control for diffusion-based video generation. arXiv preprint arXiv:2401.00896, 2023. 1", + "[37] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 3", + "[38] Muyao Niu, Xiaodong Cun, Xintao Wang, Yong Zhang, Ying Shan, and Yinqiang Zheng. Mofa-video: Controllable image animation via generative motion field adoptions in frozen image-to-video diffusion model. arXiv preprint arXiv:2405.20222, 2024. 1", + "[39] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15932-15942, 2023. 1", + "[40] Haotong Qin, Lei Ke, Xudong Ma, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Xianglong Liu, and Fisher Yu. Bimatting: Efficient video matting via binarization. Advances in Neural Information Processing Systems, 36: 43307-43321, 2023. 2, 6" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dólár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2, 6", + "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3", + "[43] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 4", + "[44] Genmo Team. Mochi, 2024. 3, 4", + "[45] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. 2019. 8", + "[46] Shimon Vainer, Mark Boss, Mathias Parger, Konstantin Kutsy, Dante De Nigris, Ciara Rowles, Nicolas Perony, and Simon Donne. Collaborative control for geometry-conditioned pbr image generation. arXiv preprint arXiv:2402.05919, 2024. 8", + "[47] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1, 3", + "[48] Luozhou Wang, Guibao Shen, Yixun Liang, Xin Tao, Pengfei Wan, Di Zhang, Yijun Li, and Yingcong Chen. Motion inversion for video customization. arXiv preprint arXiv:2403.20193, 2024. 1", + "[49] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 1", + "[50] Zhixiang Wang, Baiang Li, Jian Wang, Yu-Lun Liu, Jinwei Gu, Yung-Yu Chuang, and Shin'Ichi Satoh. Matting by generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3", + "[51] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1", + "[52] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 1", + "[53] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 3" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18238", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Shuai Yang, Zhifei Chen, Pengguang Chen, Xi Fang, Shu Liu, and Yingcong Chen. Defect spectrum: A granular look of large-scale defect datasets with rich semantics, 2023. 3", + "[55] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 1", + "[56] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 3, 4, 6, 7", + "[57] Jingfeng Yao, Xinggang Wang, Shusheng Yang, and Baoyuan Wang. Vitmatte: Boosting image matting with pretrained plain vision transformers. Information Fusion, 103: 102091, 2024. 3", + "[58] Jingfeng Yao, Xinggang Wang, Lang Ye, and Wenyu Liu. Matte anything: Interactive natural image matting with segment anything model. Image and Vision Computing, 147: 105067, 2024. 6", + "[59] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023. 1", + "[60] Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloš Hasan. Rgb ↔ x: Image decomposition and synthesis using material-and lighting-aware diffusion models. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3", + "[61] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818, 2023. 1, 3", + "[62] David Junhao Zhang, Dongxu Li, Hung Le, Mike Zheng Shou, Caiming Xiong, and Doyen Sahoo. Moonshot: Towards controllable video generation and editing with multimodal conditions. arXiv preprint arXiv:2401.01827, 2024. 3", + "[63] Lvmin Zhang and Maneesh Agrawala. Transparent image layer diffusion using latent transparency. arXiv preprint arXiv:2402.17113, 2024. 2, 3, 6, 7", + "[64] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1, 3" + ], + "bbox": [ + 91, + 92, + 480, + 739 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18239", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/TransPixeler_ Advancing Text-to-Video Generation with Transparency/a57aba62-4ac6-4f75-a668-a4f223b5c05f_model.json b/2025/TransPixeler_ Advancing Text-to-Video Generation with Transparency/a57aba62-4ac6-4f75-a668-a4f223b5c05f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..91ee3f88950e06f49f42f911b2a334956a3cfc3a --- /dev/null +++ b/2025/TransPixeler_ Advancing Text-to-Video Generation with Transparency/a57aba62-4ac6-4f75-a668-a4f223b5c05f_model.json @@ -0,0 +1,2543 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.143, + 0.131, + 0.856, + 0.154 + ], + "angle": 0, + "content": "TransPixel: Advancing Text-to-Video Generation with Transparency" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.18, + 0.818, + 0.234 + ], + "angle": 0, + "content": "Luozhou Wang\\(^{1*}\\) Yijun Li\\(^{3\\dagger}\\) Zhifei Chen\\(^{1}\\) Jui-Hsien Wang\\(^{3}\\) Zhifei Zhang\\(^{3}\\) He Zhang\\(^{3}\\) Zhe Lin\\(^{3}\\) Ying-Cong Chen\\(^{1,2‡}\\) HKUST(GZ) \\({}^{2}\\) HKUST \\({}^{3}\\) Adobe Research" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.263, + 0.903, + 0.277 + ], + "angle": 0, + "content": "\"A transparent glass of water with ice cubes gently swirling\" \"A statue crumbling to dust as cracks spread across its surface\"" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.278, + 0.905, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.176, + 0.327, + 0.428, + 0.341 + ], + "angle": 0, + "content": "\"A squirrel's bushy tail flicking quickly\"" + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.327, + 0.872, + 0.34 + ], + "angle": 0, + "content": "\"A small explosion rapidly expanding and contracting\"" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.34, + 0.904, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.388, + 0.906, + 0.402 + ], + "angle": 0, + "content": "\"A white dandelion shifting as seen through a magnifying glass\" \"A massive storm forming, with swirling clouds and thunderbolts\"" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.402, + 0.904, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.452, + 0.488, + 0.465 + ], + "angle": 0, + "content": "\"A motorcycle drifting and swerving in an enchanted forest\"" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.465, + 0.904, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.517, + 0.907, + 0.547 + ], + "angle": 0, + "content": "Figure 1. RGBA Video Generation with TransPixeler. By introducing LoRA layers into DiT-based text-to-video model with a novel alpha channel adaptive attention mechanism, our method enables RGBA video generation from text while preserving Text-to-Video quality." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.565, + 0.327, + 0.581 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.598, + 0.486, + 0.854 + ], + "angle": 0, + "content": "Text-to-video generative models have made significant strides, enabling diverse applications in entertainment, advertising, and education. However, generating RGBA video, which includes alpha channels for transparency, remains a challenge due to limited datasets and the difficulty of adapting existing models. Alpha channels are crucial for visual effects (VFX), allowing transparent elements like smoke and reflections to blend seamlessly into scenes. We introduce TransPixeler, a method to extend pretrained video models for RGBA generation while retaining the original RGB capabilities. TransPixeler leverages a diffusion transformer (DiT) architecture, incorporating alpha-specific tokens and using LoRA-based fine-tuning to jointly generate RGB and alpha channels with high consistency. By optimizing attention mechanisms, TransPixeler preserves the strengths of the original RGB model and achieves strong alignment between RGB and alpha channels despite limited training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.907, + 0.629 + ], + "angle": 0, + "content": "data. Our approach effectively generates diverse and consistent RGBA videos, advancing the possibilities for VFX and interactive content creation. The code is available at https://wileewang.github.io/TransPixeler/." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.663, + 0.646, + 0.679 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.69, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Text-to-Video generative models have quickly advanced, achieving impressive results [7, 16, 20, 26, 47, 49, 56, 61, 64]. This progress has enabled various applications, such as video editing [11, 13, 32, 39, 52, 55], image animation [2, 14, 15, 38], and motion customization [18, 24, 31, 36, 48, 51, 59]. Diffusion Transformers (DiT) enhance these models by using self-attention to capture long-range dependencies [3, 26, 56, 64]. These models are now widely used in entertainment, advertising, and education, meeting the demand for customizable, dynamic content. Notably, Text-to-RGBA (A denotes Alpha channel) video generation is invaluable for VFX and creative industries. The inclusion of an alpha channel in RGBA formats allows for transparent effects, enabling seamless blending of elements like smoke" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.864, + 0.438, + 0.877 + ], + "angle": 0, + "content": "*This work was done during an internship at Adobe Research." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.877, + 0.197, + 0.889 + ], + "angle": 0, + "content": "†Project Leader" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.889, + 0.234, + 0.902 + ], + "angle": 0, + "content": "\\(\\ddagger\\) Corresponding author" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.864, + 0.438, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18229" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "and reflections (see Fig. 1). This transparency creates realistic visuals that can integrate smoothly into scenes without modifying the background. Such flexibility is crucial in gaming, virtual reality (VR), and augmented reality (AR), where dynamic and interactive content is in high demand." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.483, + 0.427 + ], + "angle": 0, + "content": "Currently, no direct solutions exist for RGBA video generation, which remains a challenging task due to the scarcity of RGBA video data, with only around 484 videos available in [29]. This scarcity will significantly limit the diversity of generated content, resulting in a constrained set of object types and motion patterns. One feasible solution is to use video matting [28, 30, 40] to obtain alpha channels from generated videos. However, these methods are still limited by the scarcity of RGBA video data and struggle to generalize to a wider range of objects, as shown in Fig. 2 (b). Other video segmentation methods, such as SAM-2 [41], may generalize well to different tasks. However, they cannot generate alpha channels and are therefore unsuitable for direct compositing. There have been attempts to generate RGBA at the image level, such as LayerDiffusion [63]. However, adapting its concept directly to a temporal VAE used in video generative models remains challenging." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.483, + 0.687 + ], + "angle": 0, + "content": "In this paper, we explore how to extend pretrained video models to generate corresponding alpha channels while retaining the original capabilities of pretrained models. Our goal is to generate content beyond the limitations of the current RGBA training set. Existing works such as Lotus [19] and Marigold [25] demonstrate that leveraging pretrained generation model weights significantly enhances out-of-distribution in dense prediction, hinting at the potential for predicting alpha channels. However, in the context of RGBA video generation, these approaches typically require generating RGB channels first, followed by separate alpha channel prediction. Consequently, information flows unidirectionally from RGB to alpha, keeping the two processes largely disconnected. Given the limited availability of RGBA video data, this imbalance results in insufficient alpha prediction when challenging objects are generated, as shown in Fig. 2 (c)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose TransPixeler, which effectively adapts the pretrained RGB video models to generate RGB channels and the alpha channel simultaneously. We leverage state-of-the-art DiT-like video generation models [26, 56], and additionally introduce new tokens appended after text and RGB tokens for generating the alpha channels. To facilitate convergence, we reinitialize the positional embeddings for the alpha tokens and introduce a zero-initialized, learnable domain embedding to distinguish alpha tokens from RGB tokens. Furthermore, we employ a LoRA-based fine-tuning scheme [23], applied exclusively to project alpha tokens into the qkv space, thereby maintaining RGB generation quality. With the proposed approach, we extend the modality while preserving the original input" + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.101, + 0.697, + 0.125 + ], + "angle": 0, + "content": "\"A dust cloud expanding after an explosion, covering the area\"" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.134, + 0.709, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.18, + 0.652, + 0.189 + ], + "angle": 0, + "content": "(a) Generated RGB" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.091, + 0.903, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.136, + 0.903, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.18, + 0.886, + 0.19 + ], + "angle": 0, + "content": "(b) w/o generative prior (Video Matting)" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.19, + 0.709, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.281, + 0.7, + 0.291 + ], + "angle": 0, + "content": "(c) with generative prior (MariGold, Lotus)" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.19, + 0.903, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.281, + 0.822, + 0.291 + ], + "angle": 0, + "content": "(d)Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.3, + 0.905, + 0.357 + ], + "angle": 0, + "content": "Figure 2. Comparison between Generation-Then-Prediction and our Joint Generation approach. Given the generated RGB in (a), (b) and (c) show the predicted alpha (top) and the composited result (bottom). In (d), the top shows the jointly generated alpha." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.367, + 0.905, + 0.397 + ], + "angle": 0, + "content": "output structure and relying on the existing attention mechanism through LoRA adaptation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.906, + 0.684 + ], + "angle": 0, + "content": "The extended sequence contains text, RGB, and alpha tokens, with self-attention divided into a 3x3 grouped attention matrix involving interactions like Text-attend-to-RGB (Text as query, RGB as key) and others. We also systematically analyze the attention mechanisms for RGBA generation: 1) Text-attend-to-RGB and RGB-attend-to-Text. The interaction between text and RGB tokens represents original model's generation capabilities. Minimizing the impact on text and RGB tokens during these attention computation processes can better retain the original model's performance; 2) RGB-attend-to-Alpha. We reveals a fundamental limitation in conventional methods is the lack of RGB-attend-to-Alpha attention. This attention is necessary to refine RGB tokens based on alpha information, improving RGB-alpha alignment; 3) Text-attend-to-Alpha. We remove this attention mechanism to reduce the risk caused by limited training data, which could degrade the model's performance. This removal also enhances the retention of the model's original capabilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.905, + 0.746 + ], + "angle": 0, + "content": "By integrating these techniques, our method achieves diverse RGBA generation with limited training data while maintaining strong RGB-alpha alignment. To summarize, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.748, + 0.905, + 0.793 + ], + "angle": 0, + "content": "- We propose an RGBA video generation framework using DiT models that requires limited data and training parameters, achieving diverse generation with strong alignment." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.794, + 0.905, + 0.853 + ], + "angle": 0, + "content": "- We analyze the role of each attention component in the generation process, optimize their interactions, and introduce necessary modifications to improve RGBA generation quality." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.854, + 0.905, + 0.898 + ], + "angle": 0, + "content": "- Our method is validated through extensive experiments, demonstrating its effectiveness across a variety of challenging scenarios." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.748, + 0.905, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18230" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.233, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.117, + 0.484, + 0.254 + ], + "angle": 0, + "content": "Text-to-Video Generation. Early video generation models were primarily based on Unet-based latent diffusion models (LDMs) extended from text-to-image models like Stable Diffusion [42]. For example,AnimateDiff [16] introduced a temporal attention module to improve temporal consistency across frames. Subsequent video generation models [5, 7, 8, 47, 61, 62] adopted an alternating approach with 2D spatial and 1D temporal attention, including works like ModelScope, VideoCrafter, Moonshot, and Show-1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.256, + 0.484, + 0.513 + ], + "angle": 0, + "content": "With advancements in large language models (LLMs) and the introduction of Sora [3], attention shifted from Unet architectures to transformer-based architectures (DiT). DiT-based video generation models, such as Latte [37] and OpenSora [64], extended the DiT text-to-image (T2I) model [9] and maintained the 2D and 1D alternating attention approach, achieving promising results. Recently, DiT-based video generation has rapidly progressed, achieving further improvements in quality. Several methods [26, 44, 56] have moved away from the 2D and 1D alternating approach, instead treating video frames as a single long sequence with 3D positional embeddings for encoding. These approaches also prepend text tokens—processed through a text encoder—to the video sequence, creating a streamlined network that relies solely on full self-attention and feedforward layers. Our method builds upon these recent open-source transformer-based video generation models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.484, + 0.779 + ], + "angle": 0, + "content": "Video Matting. A straightforward approach for RGBA video generation is to extract the alpha channel from generated RGB content, as done with traditional green screen keying or learning-based video matting expert models [28-30]. OmnimatteRF [28] introduces a video matting method that combines dynamic 2D foreground layers with a 3D background model, enabling more realistic scene reconstruction for real-world videos. Robust Video Matting (RVM) [30] proposes a real-time, high-quality human video matting method with a recurrent architecture for improved temporal coherence, achieving state-of-the-art results without auxiliary inputs. Another work presents a high-speed, high-resolution background replacement technique with precise alpha matte extraction, supported by the VideoMatte240K and PhotoMatte13K/85 datasets [29]. Additionally, many image matting methods [4, 6, 27, 50, 57] can be applied for frame-by-frame matting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Further, several works [19, 25, 53] in image depth estimation adapt pretrained generation models for prediction tasks, achieving strong performance that often surpasses traditional, scratch-trained expert models. Marigold [25] modifies architectures to create image-conditioned generation models, while Lotus [19] explores the role of the diffusion process in this context. Although there is currently no dedicated approach for video matting within video gen" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.138 + ], + "angle": 0, + "content": "eration models, we replicate and extend these methods to evaluate their performance, allowing us to highlight the limitations of prediction-based pipelines for RGBA generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.143, + 0.907, + 0.43 + ], + "angle": 0, + "content": "Generation beyond RGB. Another category of methods [1, 17, 34, 35, 54, 60, 63] explores expanding generation models to simultaneously generate additional channels, though they are not specifically designed for RGBA video generation. For instance, LayerDiffusion [63] modifies the VAE in latent diffusion models to decode alpha channels. However, VAEs typically lack the semantic understanding required for precise alpha generation, limiting their effectiveness in complex visual scenarios where texture and contour details are critical. In contrast, other approaches [1, 34, 35, 60] modify the denoising model directly to enable joint generation. Wonder3D [34] uses a domain embedding to control the model's generation modality, while methods like IntrinsicDiffusion [35] and RGB \\(\\leftrightarrow\\) X [60] adapt the UNet's input and output layers to jointly produce intrinsic modalities. However, all these methods are designed for image tasks and rely on UNet architectures. When applied to video generation, they face limitations in quality and diversity due to the scarcity of RGBA video data." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.442, + 0.605, + 0.457 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.467, + 0.643, + 0.484 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.907, + 0.747 + ], + "angle": 0, + "content": "We first introduce the open-sourced state-of-the-art DiT-based video generation models [44, 56]. The core components of DiT-based video models are attention modules, and there are two primary distinctions between these models and previous approaches. On one hand, unlike previous models that alternate between 1D temporal attention and 2D spatial attention [5, 7, 8, 64], current methods typically employ 3D spatio-temporal attention, allowing them to capture spatio-temporal dependencies more effectively. On the other hand, instead of using cross-attention for text conditioning, these models concatenate text tokens \\(\\mathbf{x}_{\\mathrm{text}}\\) with visual tokens \\(\\mathbf{x}_{\\mathrm{video}}\\) into a single long sequence. The shape of video tokens and text tokens are \\(B\\times L\\times D\\) and \\(B\\times L_{\\mathrm{text}}\\times D\\) when \\(B\\) equals to batch size, \\(L_{\\mathrm{text}}\\) equals to the length of text tokens, \\(L\\) equals to the length of video tokens and \\(D\\) equals to the latent dimension of transformer. Full self-attention is then applied across the combined sequence:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.755, + 0.905, + 0.841 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {T}}{\\sqrt {d _ {k}}}\\right) \\mathbf {V}, \\quad \\text {w h e r e} \\\\ \\mathbf {Z}: \\mathbf {Z} \\in \\{\\mathbf {Q}, \\mathbf {K}, \\mathbf {V} \\} \\\\ = \\left[ \\mathbf {W} _ {z: z \\in \\{q, k, v \\}} \\left(\\mathbf {x} _ {\\text {t e x t}}\\right); \\mathbf {f} _ {z: z \\in \\{q, k, v \\}} \\left(\\mathbf {x} _ {\\text {v i d e o}}\\right) \\right] \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Here \\(\\mathbf{W}_t\\) (for \\(t \\in \\{q, k, v\\}\\)) represents the projection matrix in the transformer model, and \\(\\mathbf{f}_t\\) (for \\(t \\in \\{q, k, v\\}\\)) represents a combined operation that incorporates both the projection and positional encoding for visual tokens. There" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18231" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.9, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.401, + 0.907, + 0.445 + ], + "angle": 0, + "content": "Figure 3. Pipeline of TransPixeler. Our method is organized as follows: (1) Left: we extend the input of DiT to include new alpha tokens; (2) Top Center: we initialize alpha tokens with our positional encoding; (3) Bottom Center: we insert a partial LoRA and adjust attention computation during training and inference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.461, + 0.483, + 0.492 + ], + "angle": 0, + "content": "are two commonly used types of positional encoding. One is absolute positional encoding formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.5, + 0.483, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {z: z \\in \\{q, k, v \\}} \\left(\\mathbf {x} _ {\\text {v i d e o}}\\right) := \\mathbf {W} _ {z: z \\in \\{q, k, v \\}} \\left(\\mathbf {x} _ {\\text {v i d e o}} ^ {m} + \\mathbf {p} ^ {m}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.526, + 0.484, + 0.587 + ], + "angle": 0, + "content": "where \\(\\mathbf{p}\\) is the positional embedding (e.g., a sinusoidal function) and \\(m\\) denotes the position of each RGB video token. Another approach is the Rotary Position Embedding (RoPE) [43], often used by [44, 56]. This is expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.594, + 0.483, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {z: z \\in \\{q, k \\}} \\left(\\mathbf {x} _ {\\text {v i d e o}}\\right) := \\mathbf {W} _ {z: z \\in \\{q, k \\}} \\left(\\mathbf {x} _ {\\text {v i d e o}} ^ {m}\\right) \\circ e ^ {i m \\theta}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.619, + 0.484, + 0.651 + ], + "angle": 0, + "content": "where \\(m\\) is the positional index, \\(i\\) is the imaginary unit for rotation, and \\(\\theta\\) is the rotation angle." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.658, + 0.241, + 0.674 + ], + "angle": 0, + "content": "3.2. Our Approach" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.681, + 0.483, + 0.726 + ], + "angle": 0, + "content": "To jointly generate RGB and alpha videos, we adapt a pretrained RGB video generation model through several modifications. The whole pipeline is visualized in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.727, + 0.484, + 0.817 + ], + "angle": 0, + "content": "Firstly, we double the sequence length of noisy input tokens to enable the model to generate videos of double length, from \\( \\mathbf{x}_{\\mathrm{video}}^{1:L} \\) to \\( \\mathbf{x}_{\\mathrm{video}}^{1:2*L} \\). Here, \\( \\mathbf{x}_{\\mathrm{video}}^{1:L} \\) will be decoded into the RGB video, while \\( \\mathbf{x}_{\\mathrm{video}}^{L+1:2*L} \\) will be decoded into the corresponding alpha video. The Query(Q), Key(K), Value(V) representations are formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.826, + 0.484, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {Z}: \\mathbf {Z} \\in \\{\\mathbf {Q}, \\mathbf {K}, \\mathbf {V} \\} \\\\ = \\left[ \\mathbf {W} _ {z: z \\in \\{q, k, v \\}} \\left(\\mathbf {x} _ {\\text {t e x t}}\\right); \\mathbf {f} _ {z: z \\in \\{q, k, v \\}} \\left(\\mathbf {x} _ {\\text {v i d e o}} ^ {1: 2 * L}\\right) \\right] \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In addition to sequence doubling, we explored increasing batch size or latent dimensions and splitting output into two" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.905, + 0.491 + ], + "angle": 0, + "content": "domains; however, these approaches showed limited effectiveness under constrained datasets, which we discuss later." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.906, + 0.568 + ], + "angle": 0, + "content": "Secondly, we modify the positional encoding function \\(\\mathbf{f}_{t:t\\in \\{q,k,v\\}}(\\cdot)\\), as shown in Fig. 4. Instead of continuously numbering indices, we allow RGB and alpha tokens to share the same positional encoding. Taking absolute positional encoding as an example:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.58, + 0.905, + 0.642 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {f} _ {z: z \\in \\{q, k, v \\}} ^ {*} (\\mathbf {x} _ {\\mathrm {v i d e o}}) \\\\ := \\left\\{ \\begin{array}{l l} \\mathbf {W} _ {z: z \\in \\{q, k, v \\}} \\left(\\mathbf {x} _ {\\text {v i d e o}} ^ {m} + \\mathbf {p} ^ {m}\\right), & \\text {i f} m \\leq L, \\\\ \\mathbf {W} _ {z: z: z \\in \\{q, k, v \\}} ^ {*} \\left(\\mathbf {x} _ {\\text {v i d e o}} ^ {m} + \\mathbf {p} ^ {m - L} + d\\right), & \\text {i f} m > L. \\end{array} \\right. \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.652, + 0.907, + 0.788 + ], + "angle": 0, + "content": "Here we introduce a domain embedding \\( d \\), initialized to zero. We make it learnable to help the model adaptively differentiate between RGB (\\( m \\leq L \\)) and alpha tokens (\\( m > L \\)). The motivation behind this design is we observe that with same positional encoding, even initializing with different noises, the tokens from two domains tend to generate same results. It minimizes spatial-temporal alignment challenges at the very beginning of training and thus accelerates convergence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.789, + 0.906, + 0.832 + ], + "angle": 0, + "content": "Next we propose a fine-tuning scheme using LoRA [23], in which the LoRA layer is applied only to alpha domain tokens:" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.843, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {W} _ {z: z \\in \\{q, k, v \\}} ^ {*} (\\mathbf {x} _ {\\text {v i d e o}} ^ {m} + \\mathbf {p} ^ {m - L} + d) \\\\ = \\mathbf {W} _ {z: z \\in \\{q, k, v \\}} \\left(\\mathbf {x} _ {\\text {v i d e o}} ^ {m} + \\mathbf {p} ^ {m - L} + d\\right) \\tag {6} \\\\ + \\gamma \\cdot \\operatorname {L o R A} \\left(\\mathbf {x} _ {\\text {v i d e o}} ^ {m} + \\mathbf {p} ^ {m - L} + d\\right), \\quad \\text {i f} m > L, \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18232" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.287, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.09, + 0.484, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.267, + 0.484, + 0.325 + ], + "angle": 0, + "content": "Figure 4. Positional Encoding Design for RGBA Generation. Assigning alpha tokens the same positional encoding as RGB yields similar results, resulting in faster convergence after 1000 iterations compared to standard encoding strategies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.338, + 0.484, + 0.412 + ], + "angle": 0, + "content": "where \\(\\gamma\\) controls the residual strength. Additionally, we design an attention mask to block unwanted attention computation. Given a text-video token sequence length \\(L_{\\mathrm{text}} + 2L\\), where \\(L_{\\mathrm{text}}\\) represents text token length, the mask is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.419, + 0.484, + 0.46 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {m n} ^ {*} = \\left\\{ \\begin{array}{l l} - \\infty , & \\text {i f} m \\leq L _ {\\text {t e x t}} \\text {a n d} n > L _ {\\text {t e x t}} + L, \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.467, + 0.484, + 0.497 + ], + "angle": 0, + "content": "Combining these modifications, inference with our method is expressed as:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.505, + 0.498, + 0.54 + ], + "angle": 0, + "content": "Attention \\((\\mathbf{Q},\\mathbf{K},\\mathbf{V}) =\\) softmax \\(\\left(\\frac{\\mathbf{Q}\\mathbf{K}^T}{\\sqrt{d_k}} +\\mathbf{M}^*\\right)\\mathbf{V},\\) where" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.542, + 0.484, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {Z}: \\mathbf {Z} \\in \\{\\mathbf {Q}, \\mathbf {K}, \\mathbf {V} \\} \\\\ = \\left[ \\mathbf {W} _ {z: z \\in \\{q, k, v \\}} \\left(\\mathbf {x} _ {\\text {t e x t}}\\right); \\mathbf {f} _ {z: z \\in \\{q, k, v \\}} ^ {*} \\left(\\mathbf {x} _ {\\text {v i d e o}}\\right) \\right] \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.592, + 0.484, + 0.623 + ], + "angle": 0, + "content": "Training is carried out using flow matching [33] or a traditional diffusion process [21]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.63, + 0.194, + 0.645 + ], + "angle": 0, + "content": "3.3. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.653, + 0.484, + 0.79 + ], + "angle": 0, + "content": "Given our goal of maximizing the inherited capabilities of the pretrained video model, enabling it to generate beyond the existing RGBA training set, we analyze the most critical component within our current 3D full attention DiT video generation model: the attention mechanism. The attention matrix, \\(\\mathbf{Q}\\mathbf{K}^T\\), has dimensions \\((L_{\\mathrm{text}} + 2*L)\\times (L_{\\mathrm{text}} + 2*L)\\) which we simplify by organizing it into a \\(3\\times 3\\) grouped attention matrix—including Text-attend-to-RGB, RGB-attend-to-Text, and so forth, as illustrated in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Text-Attend-to-RGB and RGB-Attend-to-Text. These represent the upper-left 2x2 section of and are computations that exist solely in the original RGB generation model. If we ensure that this part of the computation remains unaffected, we can replicate the original RGB generation performance. Therefore, we limit the scope of LoRA's influence, as defined in Eq. (7), by retaining the original QKV" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.906, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.3, + 0.907, + 0.368 + ], + "angle": 0, + "content": "Figure 5. Attention Rectification. (a) Eliminating all attention from alpha as a key preserves \\(100\\%\\) RGB generation but leads to poor alignment. (b) Retaining all attention significantly degrades quality, causing a lack of motion in bicycles. (c) Our method achieves an effective balance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.906, + 0.417 + ], + "angle": 0, + "content": "values for both text and RGB tokens, thus preserving the pretrained model's behavior in these domains." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.906, + 0.509 + ], + "angle": 0, + "content": "Besides the partial LoRA, the added alpha tokens requires the text and RGB tokens to also act as queries and interact with the alpha tokens as keys, which alters the computation in this 2x2 attention matrix. Therefore, we further analyze two additional attention computations that impact RGB generation, as shown in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.515, + 0.906, + 0.667 + ], + "angle": 0, + "content": "Text-Attend-to-Alpha. We find that this attention is detrimental to the generation quality. Since the model was originally trained with text and RGB data, introducing attention from text to alpha causes interference due to the domain gap between alpha and RGB. Specifically, the alpha modality provides only contour information and lacks the rich texture, color, and semantic details associated with the text prompt, thereby degrading generation quality. To mitigate this, we design the attention mask (Eq. (7)) that blocks this computation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.906, + 0.794 + ], + "angle": 0, + "content": "RGB-Attend-to-Alpha. In contrast, we identify RGB-to-Alpha as essential for successful joint generation. This attention allows the model to refine RGB tokens by considering alpha information, facilitating alignment between generated RGB and alpha channels. This refinement process is a critical component missing in previous generation-then-prediction pipelines, which lacked a feedback mechanism for RGB refinement based on alpha guidance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.8, + 0.64, + 0.817 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Training Dataset. We utilize the public VideoMatte240K dataset [29], a comprehensive collection of 484 high-resolution green screen videos consists of 240,709 unique frames of alpha mattes and foregrounds. These frames provide a diverse range of human subjects, clothing styles, and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18233" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.245, + 0.092, + 0.356, + 0.106 + ], + "angle": 0, + "content": "\"A coin spinning\"" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.092, + 0.882, + 0.106 + ], + "angle": 0, + "content": "\"A forest floor being consumed by spreading magical fire\"" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.107, + 0.903, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.247, + 0.154, + 0.349, + 0.167 + ], + "angle": 0, + "content": "\"A parrot flying\"" + }, + { + "type": "text", + "bbox": [ + 0.538, + 0.154, + 0.867, + 0.166 + ], + "angle": 0, + "content": "\"An asteroid belt swirling chaotically through space\"" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.165, + 0.903, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.214, + 0.479, + 0.228 + ], + "angle": 0, + "content": "\"An astronaut running down an alley, spacesuit flapping\"" + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.214, + 0.882, + 0.227 + ], + "angle": 0, + "content": "\"A cloud of dust erupting and dispersing like an explosion\"" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.227, + 0.903, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.276, + 0.456, + 0.289 + ], + "angle": 0, + "content": "\"a woman's long black hair streaming as she runs\"" + }, + { + "type": "text", + "bbox": [ + 0.606, + 0.276, + 0.798, + 0.289 + ], + "angle": 0, + "content": "\"Water splattering in mid-air\"" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.289, + 0.903, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.342, + 0.182, + 0.355 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.342, + 0.412, + 0.353 + ], + "angle": 0, + "content": "Generated RGBA Video" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.342, + 0.596, + 0.353 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "text", + "bbox": [ + 0.684, + 0.342, + 0.827, + 0.353 + ], + "angle": 0, + "content": "Generated RGBA Video" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.353, + 0.903, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.403, + 0.903, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.462, + 0.855, + 0.476 + ], + "angle": 0, + "content": "Figure 6. Applications. Top: Text-to-Video with Transparency. Bottom: Image-to-Video generation with transparency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.494, + 0.483, + 0.54 + ], + "angle": 0, + "content": "poses. We apply fundamental preprocessing steps for them, including color decontamination and background blurring. Prompts are extracted using ShareGPT4V [10]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.545, + 0.484, + 0.788 + ], + "angle": 0, + "content": "Model. Our RGBA video diffusion models are developed by fine-tuning pre-trained diffusion models. Specifically, we employ two models based on the diffusion transformer architecture: the open-source model CogVideoX [56] and a modified variant of CogVideoX denoted as \\( J \\). CogVideoX generates RGB videos in 480x720 with 49 frames at 8 FPS, using 50 sampling steps. In contrast, the modified version produces videos in 176x320 with 64 frames at 24 FPS, while also using 50 sampling steps. Additionally, we integrate our method with CogVideoX-I2V (Image-to-Video) to support image-to-video generation with transparency. We set the LoRA rank to 128. For domain embedding, we initialize it with an original shape of \\( 1 \\times D \\) and zero values, then expand it to \\( L \\times D \\) through repetition during training. We train these parameters over 5,000 iterations with a batch size of 8 in total, utilizing 8 NVIDIA A100 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.796, + 0.225, + 0.812 + ], + "angle": 0, + "content": "4.1. Applications" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.819, + 0.472, + 0.835 + ], + "angle": 0, + "content": "We mainly demonstrate two applications shown in Fig. 6:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Text-to-Video with Transparency. Our method is capable of generating moving objects with various types of motion, such as spinning, running, and flying, while also handling transparent properties of bottles and glasses. Additionally," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.905, + 0.525 + ], + "angle": 0, + "content": "it can produce complex visual effects, including fire, explosions, cracking, and lightning, as well as creative examples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.906, + 0.622 + ], + "angle": 0, + "content": "Image-to-Video with Transparency. Our method can also be integrated with an I2V video generation model-CogVideoX-I2V. Users can provide a single image along with an alpha channel (optional), and then we generate subsequent frames with dynamic effects and automatically propagate or generate alpha channels for these frames." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.63, + 0.654, + 0.646 + ], + "angle": 0, + "content": "4.2. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.653, + 0.907, + 0.835 + ], + "angle": 0, + "content": "Generation-then-Prediction Pipeline. As shown in Fig. 2, video matting methods [29, 40, 58] struggle with matting non-human objects (see supplementary materials for additional results). Therefore, we selected Lotus [19] and SAM-2 [41] as baselines due to their stronger generalization: Lotus uses pretrained generative models, and SAM-2 is trained on large datasets. Since Lotus was originally designed for single-image depth estimation, we extended it for RGBA videos, denoted as Lotus + RGBA in our comparisons. Qualitative results are shown in Fig. 7. Since ground-truth alpha channels are not available for generated videos, we focus on qualitative comparison." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Joint Generation Pipeline. Since there are currently no existing RGBA video generation models, we integrateAnimateDiff [16] with LayerDiffusion [63] to generate RGBA videos. We use the open-source video generation model" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18234" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.887, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.428, + 0.842, + 0.444 + ], + "angle": 0, + "content": "Figure 7. Comparison with Generation-then-Prediction Pipelines. Our method demonstrates superior alignment." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.454, + 0.907, + 0.693 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.702, + 0.908, + 0.734 + ], + "angle": 0, + "content": "Figure 8. Comparison with Joint Generation Pipelines. Top: LayerDiffusion +AnimateDiff; Bottom: Ours. Our method achieves better alignment and generates corresponding motion described by prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.486, + 0.78 + ], + "angle": 0, + "content": "CogVideoX [56] as the base model for fair comparison. The qualitative results are illustrated in Fig. 8." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.485, + 0.903 + ], + "angle": 0, + "content": "User Study. We also conduct a user study with Amazon Mechanical Turk to compare two joint generation methods, as shown in Table. 1. Participants are asked to evaluate two key aspects: 1) whether the RGB and alpha align correctly; and 2) whether the motion in the generated video matches the corresponding text description. A total of 30 videos are generated from distinct text prompts, and 87 users partici" + }, + { + "type": "table_caption", + "bbox": [ + 0.645, + 0.747, + 0.774, + 0.761 + ], + "angle": 0, + "content": "Table 1. User Study." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.763, + 0.905, + 0.811 + ], + "angle": 0, + "content": "
RGBA AlignmentMotion Quality
AnimateDiff [16]+LayerDiff [63]6.7%21.7%
Ours + CogVideoX [56]93.3%78.3%
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.882, + 0.907, + 0.913 + ], + "angle": 0, + "content": "pated in the evaluation. The study shows that our method is obviously favored more by users with higher votes." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18235" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.482, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.288, + 0.483, + 0.318 + ], + "angle": 0, + "content": "Figure 9. Alternative Designs for Joint Generation with DiT. Sequence extension (b) represents our method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.341, + 0.245, + 0.358 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.365, + 0.483, + 0.396 + ], + "angle": 0, + "content": "As shown in Fig. 10, we conduct the ablation study across two dimensions: attention rectification and network design." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.409, + 0.48, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.597, + 0.484, + 0.669 + ], + "angle": 0, + "content": "Figure 10. Ablation Study. (a) Ours; (b) Ours without RGB-attend-to-Alpha; (c) Ours with Text-attend-to-alpha; (d) Batch Extension Strategy; (e) Latent Dimension Extension Strategy. Our method maintains high-quality motion generation (e.g., butterflies waving their wings) while achieving good alignment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.483, + 0.822 + ], + "angle": 0, + "content": "Attention Rectification. By blocking RGB-to-Alpha attention, we first validate the importance of RGB-to-Alpha attention for aligning RGB and alpha channels, a feature lacking in most prediction-based methods. We also examine the effect of removing unnecessary attention to preserve the model's generative capacity, by learning Text-to-Alpha attention only. Without RGB-to-Alpha attention, the alpha channel misaligns with RGB content and the RGB output loses motion quality (e.g., reverse rocket)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Alternative Designs For Joint Generation. Given the transformer's input dimensions \\(B \\times L \\times D\\), we extend the sequence dimension \\(L\\) to produce RGB and alpha channels, but alternative extensions are possible at the Batch \\(B\\) or Latent Dimension \\(D\\) levels (see Fig. 9). In the Batch Ex" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.09, + 0.885, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.223, + 0.907, + 0.265 + ], + "angle": 0, + "content": "Figure 11. Quantitative Evaluation. Our approach achieves a good balance between alignment (low flow difference) and preserving generative quality (low FVD)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.282, + 0.907, + 0.403 + ], + "angle": 0, + "content": "tension approach, a new module enables inter-batch communication, similar to the technique in [46]. For Latent Dimension Extension, we merge video and alpha tokens, project them into the DiT model's latent space, and unmerge post-generation, using learnable linear layers with fine-tuning. Batch Extension shows weaker RGB-alpha alignment, while Latent Dimension Extension, though akin to training from scratch, significantly reduces diversity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.409, + 0.909, + 0.681 + ], + "angle": 0, + "content": "Evaluation. In addition to the qualitative comparisons shown in Fig. 10, we also generated a total of 80 videos, each consisting of 64 frames, and evaluated them using two primary metrics: Flow Difference. To measure alignment between the generated RGB and Alpha videos, we use optical flow [22] to focus on motion consistency while ignoring appearance. Specifically, we calculate optical flow with Farneback method [12] and compute the flow difference as the average Euclidean distance between RGB and Alpha flow fields. Frechet Video Distance (FVD). We use FVD [45] to compare the RGB videos generated by each RGBA method against those from the original RGB model, evaluating how well each method preserves the model's original generative quality. A lower FVD indicates that the generated results are closer to the original RGB model in terms of motion coherence and diversity, thus demonstrating a high fidelity to the model's intended generative quality. Results are shown in Fig. 11." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.694, + 0.634, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We introduce a novel Text-to-RGBA video generator that extends RGB diffusion models to produce RGBA output with minimal modifications and retained fidelity. Leveraging DiT backbones and RGBA-aware attention, the method preserves RGB quality while accurately inferring alpha channels. Targeted adjustments—adding alpha tokens, reinitializing positional embeddings, and applying selective LoRA fine-tuning—deliver rich, high-quality RGBA videos even from limited training data. Extensive quantitative and qualitative experiments validate the framework's versatility and robustness across challenging animation, complex compositing, and diverse real-world scenes." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18236" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Fan Bao, Shen Nie, Kaiwen Xue, Chongxuan Li, Shi Pu, Yaole Wang, Gang Yue, Yue Cao, Hang Su, and Jun Zhu. One transformer fits all distributions in multi-modal diffusion at scale. In International Conference on Machine Learning, pages 1692-1717. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.483, + 0.323 + ], + "angle": 0, + "content": "[3] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.326, + 0.483, + 0.395 + ], + "angle": 0, + "content": "[4] Ryan D Burgert, Brian L Price, Jason Kuen, Yijun Li, and Michael S Ryoo. Magick: A large-scale captioned dataset from matting generated images using chroma keying. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22595-22604, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.396, + 0.482, + 0.436 + ], + "angle": 0, + "content": "[5] cerspense. zeroscope_v2. https://huggingface.co/ cerspense/zeroscope_v2_576w, 2023. Accessed: 2023-02-03. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.483, + 0.493 + ], + "angle": 0, + "content": "[6] Guowei Chen, Yi Liu, Jian Wang, Juncai Peng, Yuying Hao, Lutao Chu, Shiyu Tang, Zewu Wu, Zeyu Chen, Zhiliang Yu, et al. Pp-matting: high-accuracy natural image matting. arXiv preprint arXiv:2204.09433, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.494, + 0.483, + 0.564 + ], + "angle": 0, + "content": "[7] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.483, + 0.62 + ], + "angle": 0, + "content": "[8] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocraft2: Overcoming data limitations for high-quality video diffusion models. arXiv preprint arXiv:2401.09047, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.483, + 0.69 + ], + "angle": 0, + "content": "[9] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[10] Lin Chen, Jisong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.803 + ], + "angle": 0, + "content": "[11] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[12] Gunnar Farnebäck. Two-frame motion estimation based on polynomial expansion. In Proceedings of the Scandinavian Conference on Image Analysis (SCIA), pages 363-370. Springer, 2003. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[14] Jianzhu Guo, Dingyun Zhang, Xiaogiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Liveportrait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[15] Xun Guo, Mingwu Zheng, Liang Hou, Yuan Gao, Yufan Deng, Chongyang Ma, Weiming Hu, Zhengjun Zha, Haibin Huang, Pengfei Wan, et al. I2v-adapter: A general image-to-video adapter for video diffusion models. arXiv preprint arXiv:2312.16693, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[16] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[17] Hao He, Yixun Liang, Luozhou Wang, Yuanhao Cai, Xinli Xu, Hao-Xiang Guo, Xiang Wen, and Yingcong Chen. *Lucidfusion: Generating 3d gaussians with arbitrary unposed images*, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[18] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrl: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[19] Jing He, Haodong Li, Wei Yin, Yixun Liang, Leheng Li, Kaiqiang Zhou, Hongbo Liu, Bingbing Liu, and YingCong Chen. Lotus: Diffusion-based visual foundation model for high-quality dense prediction. arXiv preprint arXiv:2409.18124, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[20] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[22] Berthold KP Horn and Brian G Schunck. Determining optical flow. Artificial intelligence, 17(1-3):185–203, 1981. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.59, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[24] Hyeonho Jeong, Jinho Chang, Geon Yeong Park, and Jong Chul Ye. Dreammotion: Space-time self-similarity score distillation for zero-shot video editing. arXiv preprint arXiv:2403.12002, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[25] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[26] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[27] Jiachen Li, Jitesh Jain, and Humphrey Shi. Matting anything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1775-1785, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[28] Geng Lin, Chen Gao, Jia-Bin Huang, Changil Kim, Yipeng Wang, Matthias Zwicker, and Ayush Saraf. Omnimatterf: Robust omnimatte with 3d background modeling. In" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18237" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23471-23480, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[29] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian L Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8762-8771, 2021. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.275 + ], + "angle": 0, + "content": "[30] Shanchuan Lin, Linjie Yang, Imran Saleemi, and Soumyadip Sengupta. Robust high-resolution video matting with temporal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 238-247, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[31] Pengyang Ling, Jiazi Bu, Pan Zhang, Xiaoyi Dong, Yuhang Zang, Tong Wu, Huaian Chen, Jiaqi Wang, and Yi Jin. Motionclone: Training-free motion cloning for controllable video generation. arXiv preprint arXiv:2406.05338, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.483, + 0.39 + ], + "angle": 0, + "content": "[32] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-p2p: Video editing with cross-attention control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8599–8608, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.482, + 0.433 + ], + "angle": 0, + "content": "[33] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.483, + 0.517 + ], + "angle": 0, + "content": "[34] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9970-9980, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.483, + 0.589 + ], + "angle": 0, + "content": "[35] Jundan Luo, Duygu Ceylan, Jae Shin Yoon, Nanxuan Zhao, Julien Philip, Anna Fruhstick, Wenbin Li, Christian Richardt, and Tuanfeng Wang. Intrinsicdiffusion: Joint intrinsic layers from latent diffusion models. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.483, + 0.631 + ], + "angle": 0, + "content": "[36] Wan-Duo Kurt Ma, John P Lewis, and W Bastiaan Kleijn. Trailblazer: Trajectory control for diffusion-based video generation. arXiv preprint arXiv:2401.00896, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.483, + 0.688 + ], + "angle": 0, + "content": "[37] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[38] Muyao Niu, Xiaodong Cun, Xintao Wang, Yong Zhang, Ying Shan, and Yinqiang Zheng. Mofa-video: Controllable image animation via generative motion field adoptions in frozen image-to-video diffusion model. arXiv preprint arXiv:2405.20222, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[39] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15932-15942, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[40] Haotong Qin, Lei Ke, Xudong Ma, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Xianglong Liu, and Fisher Yu. Bimatting: Efficient video matting via binarization. Advances in Neural Information Processing Systems, 36: 43307-43321, 2023. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[41] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dólár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.907, + 0.259 + ], + "angle": 0, + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.313 + ], + "angle": 0, + "content": "[43] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.746, + 0.328 + ], + "angle": 0, + "content": "[44] Genmo Team. Mochi, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[45] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[46] Shimon Vainer, Mark Boss, Mathias Parger, Konstantin Kutsy, Dante De Nigris, Ciara Rowles, Nicolas Perony, and Simon Donne. Collaborative control for geometry-conditioned pbr image generation. arXiv preprint arXiv:2402.05919, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[47] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[48] Luozhou Wang, Guibao Shen, Yixun Liang, Xin Tao, Pengfei Wan, Di Zhang, Yijun Li, and Yingcong Chen. Motion inversion for video customization. arXiv preprint arXiv:2403.20193, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[49] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.676 + ], + "angle": 0, + "content": "[50] Zhixiang Wang, Baiang Li, Jian Wang, Yu-Lun Liu, Jinwei Gu, Yung-Yu Chuang, and Shin'Ichi Satoh. Matting by generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[51] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[52] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[53] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18238" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[54] Shuai Yang, Zhifei Chen, Pengguang Chen, Xi Fang, Shu Liu, and Yingcong Chen. Defect spectrum: A granular look of large-scale defect datasets with rich semantics, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[55] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[56] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[57] Jingfeng Yao, Xinggang Wang, Shusheng Yang, and Baoyuan Wang. Vitmatte: Boosting image matting with pretrained plain vision transformers. Information Fusion, 103: 102091, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[58] Jingfeng Yao, Xinggang Wang, Lang Ye, and Wenyu Liu. Matte anything: Interactive natural image matting with segment anything model. Image and Vision Computing, 147: 105067, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[59] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[60] Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloš Hasan. Rgb ↔ x: Image decomposition and synthesis using material-and lighting-aware diffusion models. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[61] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.574, + 0.482, + 0.641 + ], + "angle": 0, + "content": "[62] David Junhao Zhang, Dongxu Li, Hung Le, Mike Zheng Shou, Caiming Xiong, and Doyen Sahoo. Moonshot: Towards controllable video generation and editing with multimodal conditions. arXiv preprint arXiv:2401.01827, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.644, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[63] Lvmin Zhang and Maneesh Agrawala. Transparent image layer diffusion using latent transparency. arXiv preprint arXiv:2402.17113, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.687, + 0.482, + 0.741 + ], + "angle": 0, + "content": "[64] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18239" + } + ] +] \ No newline at end of file diff --git a/2025/TransPixeler_ Advancing Text-to-Video Generation with Transparency/a57aba62-4ac6-4f75-a668-a4f223b5c05f_origin.pdf b/2025/TransPixeler_ Advancing Text-to-Video Generation with Transparency/a57aba62-4ac6-4f75-a668-a4f223b5c05f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..349253a2f6920605093ad80ca538276e65e282d5 --- /dev/null +++ b/2025/TransPixeler_ Advancing Text-to-Video Generation with Transparency/a57aba62-4ac6-4f75-a668-a4f223b5c05f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c73fbaf3053df1e0f54c9ac77ccf36a0fea287c8501aa85cb267f78acd9eba0d +size 4144689 diff --git a/2025/TransPixeler_ Advancing Text-to-Video Generation with Transparency/full.md b/2025/TransPixeler_ Advancing Text-to-Video Generation with Transparency/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cf80da148b472fc2f829eab3106ac6d86fd435d8 --- /dev/null +++ b/2025/TransPixeler_ Advancing Text-to-Video Generation with Transparency/full.md @@ -0,0 +1,349 @@ +# TransPixel: Advancing Text-to-Video Generation with Transparency + +Luozhou Wang $^{1*}$ Yijun Li $^{3\dagger}$ Zhifei Chen $^{1}$ Jui-Hsien Wang $^{3}$ Zhifei Zhang $^{3}$ He Zhang $^{3}$ Zhe Lin $^{3}$ Ying-Cong Chen $^{1,2‡}$ HKUST(GZ) ${}^{2}$ HKUST ${}^{3}$ Adobe Research + +"A transparent glass of water with ice cubes gently swirling" "A statue crumbling to dust as cracks spread across its surface" + +![](images/fa8aca5e592d8e24b12576bf117460f0e6eefa10cc52f7a7789bcb1758982018.jpg) + +"A squirrel's bushy tail flicking quickly" + +"A small explosion rapidly expanding and contracting" + +![](images/4a74fac0f2482707f8a80c4f6a995b0f40124e6b1f30dd18977e4f543d21ff5a.jpg) + +"A white dandelion shifting as seen through a magnifying glass" "A massive storm forming, with swirling clouds and thunderbolts" + +![](images/2004066e27c07f6898c00ea7d3000c91b2d3555dd3634b34292669e97a1de2fe.jpg) + +"A motorcycle drifting and swerving in an enchanted forest" + +![](images/4e7c9e12ea80879b8a98b389c7161ced727c42c023a09982f198405cef2be474.jpg) +Figure 1. RGBA Video Generation with TransPixeler. By introducing LoRA layers into DiT-based text-to-video model with a novel alpha channel adaptive attention mechanism, our method enables RGBA video generation from text while preserving Text-to-Video quality. + +# Abstract + +Text-to-video generative models have made significant strides, enabling diverse applications in entertainment, advertising, and education. However, generating RGBA video, which includes alpha channels for transparency, remains a challenge due to limited datasets and the difficulty of adapting existing models. Alpha channels are crucial for visual effects (VFX), allowing transparent elements like smoke and reflections to blend seamlessly into scenes. We introduce TransPixeler, a method to extend pretrained video models for RGBA generation while retaining the original RGB capabilities. TransPixeler leverages a diffusion transformer (DiT) architecture, incorporating alpha-specific tokens and using LoRA-based fine-tuning to jointly generate RGB and alpha channels with high consistency. By optimizing attention mechanisms, TransPixeler preserves the strengths of the original RGB model and achieves strong alignment between RGB and alpha channels despite limited training + +data. Our approach effectively generates diverse and consistent RGBA videos, advancing the possibilities for VFX and interactive content creation. The code is available at https://wileewang.github.io/TransPixeler/. + +# 1. Introduction + +Text-to-Video generative models have quickly advanced, achieving impressive results [7, 16, 20, 26, 47, 49, 56, 61, 64]. This progress has enabled various applications, such as video editing [11, 13, 32, 39, 52, 55], image animation [2, 14, 15, 38], and motion customization [18, 24, 31, 36, 48, 51, 59]. Diffusion Transformers (DiT) enhance these models by using self-attention to capture long-range dependencies [3, 26, 56, 64]. These models are now widely used in entertainment, advertising, and education, meeting the demand for customizable, dynamic content. Notably, Text-to-RGBA (A denotes Alpha channel) video generation is invaluable for VFX and creative industries. The inclusion of an alpha channel in RGBA formats allows for transparent effects, enabling seamless blending of elements like smoke + +and reflections (see Fig. 1). This transparency creates realistic visuals that can integrate smoothly into scenes without modifying the background. Such flexibility is crucial in gaming, virtual reality (VR), and augmented reality (AR), where dynamic and interactive content is in high demand. + +Currently, no direct solutions exist for RGBA video generation, which remains a challenging task due to the scarcity of RGBA video data, with only around 484 videos available in [29]. This scarcity will significantly limit the diversity of generated content, resulting in a constrained set of object types and motion patterns. One feasible solution is to use video matting [28, 30, 40] to obtain alpha channels from generated videos. However, these methods are still limited by the scarcity of RGBA video data and struggle to generalize to a wider range of objects, as shown in Fig. 2 (b). Other video segmentation methods, such as SAM-2 [41], may generalize well to different tasks. However, they cannot generate alpha channels and are therefore unsuitable for direct compositing. There have been attempts to generate RGBA at the image level, such as LayerDiffusion [63]. However, adapting its concept directly to a temporal VAE used in video generative models remains challenging. + +In this paper, we explore how to extend pretrained video models to generate corresponding alpha channels while retaining the original capabilities of pretrained models. Our goal is to generate content beyond the limitations of the current RGBA training set. Existing works such as Lotus [19] and Marigold [25] demonstrate that leveraging pretrained generation model weights significantly enhances out-of-distribution in dense prediction, hinting at the potential for predicting alpha channels. However, in the context of RGBA video generation, these approaches typically require generating RGB channels first, followed by separate alpha channel prediction. Consequently, information flows unidirectionally from RGB to alpha, keeping the two processes largely disconnected. Given the limited availability of RGBA video data, this imbalance results in insufficient alpha prediction when challenging objects are generated, as shown in Fig. 2 (c). + +In this work, we propose TransPixeler, which effectively adapts the pretrained RGB video models to generate RGB channels and the alpha channel simultaneously. We leverage state-of-the-art DiT-like video generation models [26, 56], and additionally introduce new tokens appended after text and RGB tokens for generating the alpha channels. To facilitate convergence, we reinitialize the positional embeddings for the alpha tokens and introduce a zero-initialized, learnable domain embedding to distinguish alpha tokens from RGB tokens. Furthermore, we employ a LoRA-based fine-tuning scheme [23], applied exclusively to project alpha tokens into the qkv space, thereby maintaining RGB generation quality. With the proposed approach, we extend the modality while preserving the original input + +![](images/34008102bfd8108f497fd3fd41ebde0fcc0feff8e7d017d9886db7d99c250570.jpg) +"A dust cloud expanding after an explosion, covering the area" +(a) Generated RGB + +![](images/17e662660344896e39aa36eb2d635ff1a4771d6f004070fa93e42f1dceb82f6f.jpg) + +![](images/9da6bbff8b484f935d3ab969a2120a863ceeeaad75ede1c3bbf7dd6fcbda69e2.jpg) + +![](images/d35bffe7a1d10ee01dd24fc1036a8b64a7e4aac9a88501f277919b96949a08e3.jpg) +(c) with generative prior (MariGold, Lotus) +Figure 2. Comparison between Generation-Then-Prediction and our Joint Generation approach. Given the generated RGB in (a), (b) and (c) show the predicted alpha (top) and the composited result (bottom). In (d), the top shows the jointly generated alpha. + +![](images/bbbfaad957332e8796d944f08bfba63473a5fc23707ff91956938563e333f6c0.jpg) +(b) w/o generative prior (Video Matting) +(d)Ours + +output structure and relying on the existing attention mechanism through LoRA adaptation. + +The extended sequence contains text, RGB, and alpha tokens, with self-attention divided into a 3x3 grouped attention matrix involving interactions like Text-attend-to-RGB (Text as query, RGB as key) and others. We also systematically analyze the attention mechanisms for RGBA generation: 1) Text-attend-to-RGB and RGB-attend-to-Text. The interaction between text and RGB tokens represents original model's generation capabilities. Minimizing the impact on text and RGB tokens during these attention computation processes can better retain the original model's performance; 2) RGB-attend-to-Alpha. We reveals a fundamental limitation in conventional methods is the lack of RGB-attend-to-Alpha attention. This attention is necessary to refine RGB tokens based on alpha information, improving RGB-alpha alignment; 3) Text-attend-to-Alpha. We remove this attention mechanism to reduce the risk caused by limited training data, which could degrade the model's performance. This removal also enhances the retention of the model's original capabilities. + +By integrating these techniques, our method achieves diverse RGBA generation with limited training data while maintaining strong RGB-alpha alignment. To summarize, our contributions are as follows: + +- We propose an RGBA video generation framework using DiT models that requires limited data and training parameters, achieving diverse generation with strong alignment. +- We analyze the role of each attention component in the generation process, optimize their interactions, and introduce necessary modifications to improve RGBA generation quality. +- Our method is validated through extensive experiments, demonstrating its effectiveness across a variety of challenging scenarios. + +# 2. Related Work + +Text-to-Video Generation. Early video generation models were primarily based on Unet-based latent diffusion models (LDMs) extended from text-to-image models like Stable Diffusion [42]. For example,AnimateDiff [16] introduced a temporal attention module to improve temporal consistency across frames. Subsequent video generation models [5, 7, 8, 47, 61, 62] adopted an alternating approach with 2D spatial and 1D temporal attention, including works like ModelScope, VideoCrafter, Moonshot, and Show-1. + +With advancements in large language models (LLMs) and the introduction of Sora [3], attention shifted from Unet architectures to transformer-based architectures (DiT). DiT-based video generation models, such as Latte [37] and OpenSora [64], extended the DiT text-to-image (T2I) model [9] and maintained the 2D and 1D alternating attention approach, achieving promising results. Recently, DiT-based video generation has rapidly progressed, achieving further improvements in quality. Several methods [26, 44, 56] have moved away from the 2D and 1D alternating approach, instead treating video frames as a single long sequence with 3D positional embeddings for encoding. These approaches also prepend text tokens—processed through a text encoder—to the video sequence, creating a streamlined network that relies solely on full self-attention and feedforward layers. Our method builds upon these recent open-source transformer-based video generation models. + +Video Matting. A straightforward approach for RGBA video generation is to extract the alpha channel from generated RGB content, as done with traditional green screen keying or learning-based video matting expert models [28-30]. OmnimatteRF [28] introduces a video matting method that combines dynamic 2D foreground layers with a 3D background model, enabling more realistic scene reconstruction for real-world videos. Robust Video Matting (RVM) [30] proposes a real-time, high-quality human video matting method with a recurrent architecture for improved temporal coherence, achieving state-of-the-art results without auxiliary inputs. Another work presents a high-speed, high-resolution background replacement technique with precise alpha matte extraction, supported by the VideoMatte240K and PhotoMatte13K/85 datasets [29]. Additionally, many image matting methods [4, 6, 27, 50, 57] can be applied for frame-by-frame matting. + +Further, several works [19, 25, 53] in image depth estimation adapt pretrained generation models for prediction tasks, achieving strong performance that often surpasses traditional, scratch-trained expert models. Marigold [25] modifies architectures to create image-conditioned generation models, while Lotus [19] explores the role of the diffusion process in this context. Although there is currently no dedicated approach for video matting within video gen + +eration models, we replicate and extend these methods to evaluate their performance, allowing us to highlight the limitations of prediction-based pipelines for RGBA generation. + +Generation beyond RGB. Another category of methods [1, 17, 34, 35, 54, 60, 63] explores expanding generation models to simultaneously generate additional channels, though they are not specifically designed for RGBA video generation. For instance, LayerDiffusion [63] modifies the VAE in latent diffusion models to decode alpha channels. However, VAEs typically lack the semantic understanding required for precise alpha generation, limiting their effectiveness in complex visual scenarios where texture and contour details are critical. In contrast, other approaches [1, 34, 35, 60] modify the denoising model directly to enable joint generation. Wonder3D [34] uses a domain embedding to control the model's generation modality, while methods like IntrinsicDiffusion [35] and RGB $\leftrightarrow$ X [60] adapt the UNet's input and output layers to jointly produce intrinsic modalities. However, all these methods are designed for image tasks and rely on UNet architectures. When applied to video generation, they face limitations in quality and diversity due to the scarcity of RGBA video data. + +# 3. Method + +# 3.1. Preliminary + +We first introduce the open-sourced state-of-the-art DiT-based video generation models [44, 56]. The core components of DiT-based video models are attention modules, and there are two primary distinctions between these models and previous approaches. On one hand, unlike previous models that alternate between 1D temporal attention and 2D spatial attention [5, 7, 8, 64], current methods typically employ 3D spatio-temporal attention, allowing them to capture spatio-temporal dependencies more effectively. On the other hand, instead of using cross-attention for text conditioning, these models concatenate text tokens $\mathbf{x}_{\mathrm{text}}$ with visual tokens $\mathbf{x}_{\mathrm{video}}$ into a single long sequence. The shape of video tokens and text tokens are $B\times L\times D$ and $B\times L_{\mathrm{text}}\times D$ when $B$ equals to batch size, $L_{\mathrm{text}}$ equals to the length of text tokens, $L$ equals to the length of video tokens and $D$ equals to the latent dimension of transformer. Full self-attention is then applied across the combined sequence: + +$$ +\begin{array}{l} \operatorname {A t t e n t i o n} (\mathbf {Q}, \mathbf {K}, \mathbf {V}) = \operatorname {s o f t m a x} \left(\frac {\mathbf {Q} \mathbf {K} ^ {T}}{\sqrt {d _ {k}}}\right) \mathbf {V}, \quad \text {w h e r e} \\ \mathbf {Z}: \mathbf {Z} \in \{\mathbf {Q}, \mathbf {K}, \mathbf {V} \} \\ = \left[ \mathbf {W} _ {z: z \in \{q, k, v \}} \left(\mathbf {x} _ {\text {t e x t}}\right); \mathbf {f} _ {z: z \in \{q, k, v \}} \left(\mathbf {x} _ {\text {v i d e o}}\right) \right] \tag {1} \\ \end{array} +$$ + +Here $\mathbf{W}_t$ (for $t \in \{q, k, v\}$ ) represents the projection matrix in the transformer model, and $\mathbf{f}_t$ (for $t \in \{q, k, v\}$ ) represents a combined operation that incorporates both the projection and positional encoding for visual tokens. There + +![](images/2a64c8a397ea52e395ebf393c55464d63b2bac93108dde1919ed4c9f786c78fa.jpg) +Figure 3. Pipeline of TransPixeler. Our method is organized as follows: (1) Left: we extend the input of DiT to include new alpha tokens; (2) Top Center: we initialize alpha tokens with our positional encoding; (3) Bottom Center: we insert a partial LoRA and adjust attention computation during training and inference. + +are two commonly used types of positional encoding. One is absolute positional encoding formulated as follows: + +$$ +\mathbf {f} _ {z: z \in \{q, k, v \}} \left(\mathbf {x} _ {\text {v i d e o}}\right) := \mathbf {W} _ {z: z \in \{q, k, v \}} \left(\mathbf {x} _ {\text {v i d e o}} ^ {m} + \mathbf {p} ^ {m}\right), \tag {2} +$$ + +where $\mathbf{p}$ is the positional embedding (e.g., a sinusoidal function) and $m$ denotes the position of each RGB video token. Another approach is the Rotary Position Embedding (RoPE) [43], often used by [44, 56]. This is expressed as + +$$ +\mathbf {f} _ {z: z \in \{q, k \}} \left(\mathbf {x} _ {\text {v i d e o}}\right) := \mathbf {W} _ {z: z \in \{q, k \}} \left(\mathbf {x} _ {\text {v i d e o}} ^ {m}\right) \circ e ^ {i m \theta}, \tag {3} +$$ + +where $m$ is the positional index, $i$ is the imaginary unit for rotation, and $\theta$ is the rotation angle. + +# 3.2. Our Approach + +To jointly generate RGB and alpha videos, we adapt a pretrained RGB video generation model through several modifications. The whole pipeline is visualized in Fig. 3. + +Firstly, we double the sequence length of noisy input tokens to enable the model to generate videos of double length, from $\mathbf{x}_{\mathrm{video}}^{1:L}$ to $\mathbf{x}_{\mathrm{video}}^{1:2*L}$ . Here, $\mathbf{x}_{\mathrm{video}}^{1:L}$ will be decoded into the RGB video, while $\mathbf{x}_{\mathrm{video}}^{L+1:2*L}$ will be decoded into the corresponding alpha video. The Query(Q), Key(K), Value(V) representations are formulated as: + +$$ +\begin{array}{l} \mathbf {Z}: \mathbf {Z} \in \{\mathbf {Q}, \mathbf {K}, \mathbf {V} \} \\ = \left[ \mathbf {W} _ {z: z \in \{q, k, v \}} \left(\mathbf {x} _ {\text {t e x t}}\right); \mathbf {f} _ {z: z \in \{q, k, v \}} \left(\mathbf {x} _ {\text {v i d e o}} ^ {1: 2 * L}\right) \right] \tag {4} \\ \end{array} +$$ + +In addition to sequence doubling, we explored increasing batch size or latent dimensions and splitting output into two + +domains; however, these approaches showed limited effectiveness under constrained datasets, which we discuss later. + +Secondly, we modify the positional encoding function $\mathbf{f}_{t:t\in \{q,k,v\}}(\cdot)$ , as shown in Fig. 4. Instead of continuously numbering indices, we allow RGB and alpha tokens to share the same positional encoding. Taking absolute positional encoding as an example: + +$$ +\begin{array}{l} \mathbf {f} _ {z: z \in \{q, k, v \}} ^ {*} (\mathbf {x} _ {\mathrm {v i d e o}}) \\ := \left\{ \begin{array}{l l} \mathbf {W} _ {z: z \in \{q, k, v \}} \left(\mathbf {x} _ {\text {v i d e o}} ^ {m} + \mathbf {p} ^ {m}\right), & \text {i f} m \leq L, \\ \mathbf {W} _ {z: z: z \in \{q, k, v \}} ^ {*} \left(\mathbf {x} _ {\text {v i d e o}} ^ {m} + \mathbf {p} ^ {m - L} + d\right), & \text {i f} m > L. \end{array} \right. \tag {5} \\ \end{array} +$$ + +Here we introduce a domain embedding $d$ , initialized to zero. We make it learnable to help the model adaptively differentiate between RGB ( $m \leq L$ ) and alpha tokens ( $m > L$ ). The motivation behind this design is we observe that with same positional encoding, even initializing with different noises, the tokens from two domains tend to generate same results. It minimizes spatial-temporal alignment challenges at the very beginning of training and thus accelerates convergence. + +Next we propose a fine-tuning scheme using LoRA [23], in which the LoRA layer is applied only to alpha domain tokens: + +$$ +\begin{array}{l} \mathbf {W} _ {z: z \in \{q, k, v \}} ^ {*} (\mathbf {x} _ {\text {v i d e o}} ^ {m} + \mathbf {p} ^ {m - L} + d) \\ = \mathbf {W} _ {z: z \in \{q, k, v \}} \left(\mathbf {x} _ {\text {v i d e o}} ^ {m} + \mathbf {p} ^ {m - L} + d\right) \tag {6} \\ + \gamma \cdot \operatorname {L o R A} \left(\mathbf {x} _ {\text {v i d e o}} ^ {m} + \mathbf {p} ^ {m - L} + d\right), \quad \text {i f} m > L, \\ \end{array} +$$ + +![](images/2dfe9edaadb401c0c7b99995eef3ead0d14062fdf567400f84dc2e31ef8b9f45.jpg) +Figure 4. Positional Encoding Design for RGBA Generation. Assigning alpha tokens the same positional encoding as RGB yields similar results, resulting in faster convergence after 1000 iterations compared to standard encoding strategies. + +![](images/fa90b265e7bc6f20c1513283d02a7d10f28b187c9ef294c62afbc29f15416826.jpg) + +where $\gamma$ controls the residual strength. Additionally, we design an attention mask to block unwanted attention computation. Given a text-video token sequence length $L_{\mathrm{text}} + 2L$ , where $L_{\mathrm{text}}$ represents text token length, the mask is defined as: + +$$ +\mathbf {M} _ {m n} ^ {*} = \left\{ \begin{array}{l l} - \infty , & \text {i f} m \leq L _ {\text {t e x t}} \text {a n d} n > L _ {\text {t e x t}} + L, \\ 0, & \text {o t h e r w i s e .} \end{array} \right. \tag {7} +$$ + +Combining these modifications, inference with our method is expressed as: + +Attention $(\mathbf{Q},\mathbf{K},\mathbf{V}) =$ softmax $\left(\frac{\mathbf{Q}\mathbf{K}^T}{\sqrt{d_k}} +\mathbf{M}^*\right)\mathbf{V},$ where + +$$ +\begin{array}{l} \mathbf {Z}: \mathbf {Z} \in \{\mathbf {Q}, \mathbf {K}, \mathbf {V} \} \\ = \left[ \mathbf {W} _ {z: z \in \{q, k, v \}} \left(\mathbf {x} _ {\text {t e x t}}\right); \mathbf {f} _ {z: z \in \{q, k, v \}} ^ {*} \left(\mathbf {x} _ {\text {v i d e o}}\right) \right] \tag {8} \\ \end{array} +$$ + +Training is carried out using flow matching [33] or a traditional diffusion process [21]. + +# 3.3. Analysis + +Given our goal of maximizing the inherited capabilities of the pretrained video model, enabling it to generate beyond the existing RGBA training set, we analyze the most critical component within our current 3D full attention DiT video generation model: the attention mechanism. The attention matrix, $\mathbf{Q}\mathbf{K}^T$ , has dimensions $(L_{\mathrm{text}} + 2*L)\times (L_{\mathrm{text}} + 2*L)$ which we simplify by organizing it into a $3\times 3$ grouped attention matrix—including Text-attend-to-RGB, RGB-attend-to-Text, and so forth, as illustrated in Fig. 3. + +Text-Attend-to-RGB and RGB-Attend-to-Text. These represent the upper-left 2x2 section of and are computations that exist solely in the original RGB generation model. If we ensure that this part of the computation remains unaffected, we can replicate the original RGB generation performance. Therefore, we limit the scope of LoRA's influence, as defined in Eq. (7), by retaining the original QKV + +![](images/c9b8423a27459dac1d0b53393a47298ee96f360a56ce64bdc65a2c93d3fe10b3.jpg) +Figure 5. Attention Rectification. (a) Eliminating all attention from alpha as a key preserves $100\%$ RGB generation but leads to poor alignment. (b) Retaining all attention significantly degrades quality, causing a lack of motion in bicycles. (c) Our method achieves an effective balance. + +values for both text and RGB tokens, thus preserving the pretrained model's behavior in these domains. + +Besides the partial LoRA, the added alpha tokens requires the text and RGB tokens to also act as queries and interact with the alpha tokens as keys, which alters the computation in this 2x2 attention matrix. Therefore, we further analyze two additional attention computations that impact RGB generation, as shown in Fig. 5. + +Text-Attend-to-Alpha. We find that this attention is detrimental to the generation quality. Since the model was originally trained with text and RGB data, introducing attention from text to alpha causes interference due to the domain gap between alpha and RGB. Specifically, the alpha modality provides only contour information and lacks the rich texture, color, and semantic details associated with the text prompt, thereby degrading generation quality. To mitigate this, we design the attention mask (Eq. (7)) that blocks this computation. + +RGB-Attend-to-Alpha. In contrast, we identify RGB-to-Alpha as essential for successful joint generation. This attention allows the model to refine RGB tokens by considering alpha information, facilitating alignment between generated RGB and alpha channels. This refinement process is a critical component missing in previous generation-then-prediction pipelines, which lacked a feedback mechanism for RGB refinement based on alpha guidance. + +# 4. Experiment + +Training Dataset. We utilize the public VideoMatte240K dataset [29], a comprehensive collection of 484 high-resolution green screen videos consists of 240,709 unique frames of alpha mattes and foregrounds. These frames provide a diverse range of human subjects, clothing styles, and + +"A coin spinning" + +"A forest floor being consumed by spreading magical fire" + +![](images/9ad2feb3bc0b485ab9216134a217a323e3b64244335dabe5c2968eb5ff6aafcc.jpg) + +"A parrot flying" + +"An asteroid belt swirling chaotically through space" + +![](images/3aa9427e38388071c24ba8ca88f69ad51f27ef890046422777b14dc2475743c8.jpg) + +"An astronaut running down an alley, spacesuit flapping" + +"A cloud of dust erupting and dispersing like an explosion" + +![](images/4c672db82e44dc88d8e75fd27d55a402b8f03225126632b86bb6cee1b8e61837.jpg) + +"a woman's long black hair streaming as she runs" + +"Water splattering in mid-air" + +![](images/17e6491a7d7345d1c9fbf1fbdb199912f466c0bb76aa2ad4cab5ed8758ed8037.jpg) + +Input Image + +Generated RGBA Video + +Input Image + +Generated RGBA Video + +![](images/cbcf13f23b412511108c5d7945343918b0bf1b15c8623e2ef802f2f894a0843d.jpg) + +![](images/860d0ae2f5c5dc8142886fb2392e10b8db15807f99d218763bd64e876a1a7788.jpg) +Figure 6. Applications. Top: Text-to-Video with Transparency. Bottom: Image-to-Video generation with transparency. + +poses. We apply fundamental preprocessing steps for them, including color decontamination and background blurring. Prompts are extracted using ShareGPT4V [10]. + +Model. Our RGBA video diffusion models are developed by fine-tuning pre-trained diffusion models. Specifically, we employ two models based on the diffusion transformer architecture: the open-source model CogVideoX [56] and a modified variant of CogVideoX denoted as $J$ . CogVideoX generates RGB videos in 480x720 with 49 frames at 8 FPS, using 50 sampling steps. In contrast, the modified version produces videos in 176x320 with 64 frames at 24 FPS, while also using 50 sampling steps. Additionally, we integrate our method with CogVideoX-I2V (Image-to-Video) to support image-to-video generation with transparency. We set the LoRA rank to 128. For domain embedding, we initialize it with an original shape of $1 \times D$ and zero values, then expand it to $L \times D$ through repetition during training. We train these parameters over 5,000 iterations with a batch size of 8 in total, utilizing 8 NVIDIA A100 GPUs. + +# 4.1. Applications + +We mainly demonstrate two applications shown in Fig. 6: + +Text-to-Video with Transparency. Our method is capable of generating moving objects with various types of motion, such as spinning, running, and flying, while also handling transparent properties of bottles and glasses. Additionally, + +it can produce complex visual effects, including fire, explosions, cracking, and lightning, as well as creative examples. + +Image-to-Video with Transparency. Our method can also be integrated with an I2V video generation model-CogVideoX-I2V. Users can provide a single image along with an alpha channel (optional), and then we generate subsequent frames with dynamic effects and automatically propagate or generate alpha channels for these frames. + +# 4.2. Comparisons + +Generation-then-Prediction Pipeline. As shown in Fig. 2, video matting methods [29, 40, 58] struggle with matting non-human objects (see supplementary materials for additional results). Therefore, we selected Lotus [19] and SAM-2 [41] as baselines due to their stronger generalization: Lotus uses pretrained generative models, and SAM-2 is trained on large datasets. Since Lotus was originally designed for single-image depth estimation, we extended it for RGBA videos, denoted as Lotus + RGBA in our comparisons. Qualitative results are shown in Fig. 7. Since ground-truth alpha channels are not available for generated videos, we focus on qualitative comparison. + +Joint Generation Pipeline. Since there are currently no existing RGBA video generation models, we integrateAnimateDiff [16] with LayerDiffusion [63] to generate RGBA videos. We use the open-source video generation model + +![](images/5e794d65b0e64c08a69372d77a3984427867574a027fc1bd2c13d2864f7c0cb6.jpg) +Figure 7. Comparison with Generation-then-Prediction Pipelines. Our method demonstrates superior alignment. + +![](images/c711bae9c19dcde842b045268589df63b130b192213524f7c7345e25b9a12ae7.jpg) +Figure 8. Comparison with Joint Generation Pipelines. Top: LayerDiffusion +AnimateDiff; Bottom: Ours. Our method achieves better alignment and generates corresponding motion described by prompts. + +CogVideoX [56] as the base model for fair comparison. The qualitative results are illustrated in Fig. 8. + +User Study. We also conduct a user study with Amazon Mechanical Turk to compare two joint generation methods, as shown in Table. 1. Participants are asked to evaluate two key aspects: 1) whether the RGB and alpha align correctly; and 2) whether the motion in the generated video matches the corresponding text description. A total of 30 videos are generated from distinct text prompts, and 87 users partici + +Table 1. User Study. + +
RGBA AlignmentMotion Quality
AnimateDiff [16]+LayerDiff [63]6.7%21.7%
Ours + CogVideoX [56]93.3%78.3%
+ +pated in the evaluation. The study shows that our method is obviously favored more by users with higher votes. + +![](images/ac4e7cabe7a36d1cf518758d685e1fef9c367ad148b7172bb82a725a6b8d8f60.jpg) +Figure 9. Alternative Designs for Joint Generation with DiT. Sequence extension (b) represents our method. + +# 4.3. Ablation Study + +As shown in Fig. 10, we conduct the ablation study across two dimensions: attention rectification and network design. + +![](images/d6bf7fd2750b87cfbfd3643198d9261b8fc4726048d0f6b665e2e821ee28ca91.jpg) +Figure 10. Ablation Study. (a) Ours; (b) Ours without RGB-attend-to-Alpha; (c) Ours with Text-attend-to-alpha; (d) Batch Extension Strategy; (e) Latent Dimension Extension Strategy. Our method maintains high-quality motion generation (e.g., butterflies waving their wings) while achieving good alignment. + +Attention Rectification. By blocking RGB-to-Alpha attention, we first validate the importance of RGB-to-Alpha attention for aligning RGB and alpha channels, a feature lacking in most prediction-based methods. We also examine the effect of removing unnecessary attention to preserve the model's generative capacity, by learning Text-to-Alpha attention only. Without RGB-to-Alpha attention, the alpha channel misaligns with RGB content and the RGB output loses motion quality (e.g., reverse rocket). + +Alternative Designs For Joint Generation. Given the transformer's input dimensions $B \times L \times D$ , we extend the sequence dimension $L$ to produce RGB and alpha channels, but alternative extensions are possible at the Batch $B$ or Latent Dimension $D$ levels (see Fig. 9). In the Batch Ex + +![](images/1a36e93509c667fc9aed9e6448e8b4f1697c0d429cf9ed16f89ce4eb5b956bf3.jpg) +Figure 11. Quantitative Evaluation. Our approach achieves a good balance between alignment (low flow difference) and preserving generative quality (low FVD). + +tension approach, a new module enables inter-batch communication, similar to the technique in [46]. For Latent Dimension Extension, we merge video and alpha tokens, project them into the DiT model's latent space, and unmerge post-generation, using learnable linear layers with fine-tuning. Batch Extension shows weaker RGB-alpha alignment, while Latent Dimension Extension, though akin to training from scratch, significantly reduces diversity. + +Evaluation. In addition to the qualitative comparisons shown in Fig. 10, we also generated a total of 80 videos, each consisting of 64 frames, and evaluated them using two primary metrics: Flow Difference. To measure alignment between the generated RGB and Alpha videos, we use optical flow [22] to focus on motion consistency while ignoring appearance. Specifically, we calculate optical flow with Farneback method [12] and compute the flow difference as the average Euclidean distance between RGB and Alpha flow fields. Frechet Video Distance (FVD). We use FVD [45] to compare the RGB videos generated by each RGBA method against those from the original RGB model, evaluating how well each method preserves the model's original generative quality. A lower FVD indicates that the generated results are closer to the original RGB model in terms of motion coherence and diversity, thus demonstrating a high fidelity to the model's intended generative quality. Results are shown in Fig. 11. + +# 5. Conclusion + +We introduce a novel Text-to-RGBA video generator that extends RGB diffusion models to produce RGBA output with minimal modifications and retained fidelity. Leveraging DiT backbones and RGBA-aware attention, the method preserves RGB quality while accurately inferring alpha channels. Targeted adjustments—adding alpha tokens, reinitializing positional embeddings, and applying selective LoRA fine-tuning—deliver rich, high-quality RGBA videos even from limited training data. Extensive quantitative and qualitative experiments validate the framework's versatility and robustness across challenging animation, complex compositing, and diverse real-world scenes. + +# References + +[1] Fan Bao, Shen Nie, Kaiwen Xue, Chongxuan Li, Shi Pu, Yaole Wang, Gang Yue, Yue Cao, Hang Su, and Jun Zhu. One transformer fits all distributions in multi-modal diffusion at scale. In International Conference on Machine Learning, pages 1692-1717. PMLR, 2023. 3 +[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1 +[3] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 1, 3 +[4] Ryan D Burgert, Brian L Price, Jason Kuen, Yijun Li, and Michael S Ryoo. Magick: A large-scale captioned dataset from matting generated images using chroma keying. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22595-22604, 2024. 3 +[5] cerspense. zeroscope_v2. https://huggingface.co/ cerspense/zeroscope_v2_576w, 2023. Accessed: 2023-02-03. 3 +[6] Guowei Chen, Yi Liu, Jian Wang, Juncai Peng, Yuying Hao, Lutao Chu, Shiyu Tang, Zewu Wu, Zeyu Chen, Zhiliang Yu, et al. Pp-matting: high-accuracy natural image matting. arXiv preprint arXiv:2204.09433, 2022. 3 +[7] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 1, 3 +[8] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocraft2: Overcoming data limitations for high-quality video diffusion models. arXiv preprint arXiv:2401.09047, 2024. 3 +[9] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 3 +[10] Lin Chen, Jisong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 6 +[11] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023. 1 +[12] Gunnar Farnebäck. Two-frame motion estimation based on polynomial expansion. In Proceedings of the Scandinavian Conference on Image Analysis (SCIA), pages 363-370. Springer, 2003. 8 +[13] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 1 + +[14] Jianzhu Guo, Dingyun Zhang, Xiaogiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Liveportrait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 1 +[15] Xun Guo, Mingwu Zheng, Liang Hou, Yuan Gao, Yufan Deng, Chongyang Ma, Weiming Hu, Zhengjun Zha, Haibin Huang, Pengfei Wan, et al. I2v-adapter: A general image-to-video adapter for video diffusion models. arXiv preprint arXiv:2312.16693, 2023. 1 +[16] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 1, 3, 6, 7 +[17] Hao He, Yixun Liang, Luozhou Wang, Yuanhao Cai, Xinli Xu, Hao-Xiang Guo, Xiang Wen, and Yingcong Chen. *Lucidfusion: Generating 3d gaussians with arbitrary unposed images*, 2024. 3 +[18] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrl: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 1 +[19] Jing He, Haodong Li, Wei Yin, Yixun Liang, Leheng Li, Kaiqiang Zhou, Hongbo Liu, Bingbing Liu, and YingCong Chen. Lotus: Diffusion-based visual foundation model for high-quality dense prediction. arXiv preprint arXiv:2409.18124, 2024. 2, 3, 6 +[20] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 1 +[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 5 +[22] Berthold KP Horn and Brian G Schunck. Determining optical flow. Artificial intelligence, 17(1-3):185–203, 1981. 8 +[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 4 +[24] Hyeonho Jeong, Jinho Chang, Geon Yeong Park, and Jong Chul Ye. Dreammotion: Space-time self-similarity score distillation for zero-shot video editing. arXiv preprint arXiv:2403.12002, 2024. 1 +[25] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 2, 3 +[26] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 1, 2, 3 +[27] Jiachen Li, Jitesh Jain, and Humphrey Shi. Matting anything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1775-1785, 2024. 3 +[28] Geng Lin, Chen Gao, Jia-Bin Huang, Changil Kim, Yipeng Wang, Matthias Zwicker, and Ayush Saraf. Omnimatterf: Robust omnimatte with 3d background modeling. In + +Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23471-23480, 2023. 2, 3 +[29] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian L Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8762-8771, 2021. 2, 3, 5, 6 +[30] Shanchuan Lin, Linjie Yang, Imran Saleemi, and Soumyadip Sengupta. Robust high-resolution video matting with temporal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 238-247, 2022. 2, 3 +[31] Pengyang Ling, Jiazi Bu, Pan Zhang, Xiaoyi Dong, Yuhang Zang, Tong Wu, Huaian Chen, Jiaqi Wang, and Yi Jin. Motionclone: Training-free motion cloning for controllable video generation. arXiv preprint arXiv:2406.05338, 2024. 1 +[32] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-p2p: Video editing with cross-attention control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8599–8608, 2024. 1 +[33] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 5 +[34] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9970-9980, 2024. 3 +[35] Jundan Luo, Duygu Ceylan, Jae Shin Yoon, Nanxuan Zhao, Julien Philip, Anna Fruhstick, Wenbin Li, Christian Richardt, and Tuanfeng Wang. Intrinsicdiffusion: Joint intrinsic layers from latent diffusion models. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3 +[36] Wan-Duo Kurt Ma, John P Lewis, and W Bastiaan Kleijn. Trailblazer: Trajectory control for diffusion-based video generation. arXiv preprint arXiv:2401.00896, 2023. 1 +[37] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 3 +[38] Muyao Niu, Xiaodong Cun, Xintao Wang, Yong Zhang, Ying Shan, and Yinqiang Zheng. Mofa-video: Controllable image animation via generative motion field adoptions in frozen image-to-video diffusion model. arXiv preprint arXiv:2405.20222, 2024. 1 +[39] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15932-15942, 2023. 1 +[40] Haotong Qin, Lei Ke, Xudong Ma, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Xianglong Liu, and Fisher Yu. Bimatting: Efficient video matting via binarization. Advances in Neural Information Processing Systems, 36: 43307-43321, 2023. 2, 6 + +[41] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dólár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2, 6 +[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3 +[43] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 4 +[44] Genmo Team. Mochi, 2024. 3, 4 +[45] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. 2019. 8 +[46] Shimon Vainer, Mark Boss, Mathias Parger, Konstantin Kutsy, Dante De Nigris, Ciara Rowles, Nicolas Perony, and Simon Donne. Collaborative control for geometry-conditioned pbr image generation. arXiv preprint arXiv:2402.05919, 2024. 8 +[47] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1, 3 +[48] Luozhou Wang, Guibao Shen, Yixun Liang, Xin Tao, Pengfei Wan, Di Zhang, Yijun Li, and Yingcong Chen. Motion inversion for video customization. arXiv preprint arXiv:2403.20193, 2024. 1 +[49] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 1 +[50] Zhixiang Wang, Baiang Li, Jian Wang, Yu-Lun Liu, Jinwei Gu, Yung-Yu Chuang, and Shin'Ichi Satoh. Matting by generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3 +[51] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1 +[52] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 1 +[53] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 3 + +[54] Shuai Yang, Zhifei Chen, Pengguang Chen, Xi Fang, Shu Liu, and Yingcong Chen. Defect spectrum: A granular look of large-scale defect datasets with rich semantics, 2023. 3 +[55] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 1 +[56] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 3, 4, 6, 7 +[57] Jingfeng Yao, Xinggang Wang, Shusheng Yang, and Baoyuan Wang. Vitmatte: Boosting image matting with pretrained plain vision transformers. Information Fusion, 103: 102091, 2024. 3 +[58] Jingfeng Yao, Xinggang Wang, Lang Ye, and Wenyu Liu. Matte anything: Interactive natural image matting with segment anything model. Image and Vision Computing, 147: 105067, 2024. 6 +[59] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023. 1 +[60] Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloš Hasan. Rgb ↔ x: Image decomposition and synthesis using material-and lighting-aware diffusion models. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3 +[61] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818, 2023. 1, 3 +[62] David Junhao Zhang, Dongxu Li, Hung Le, Mike Zheng Shou, Caiming Xiong, and Doyen Sahoo. Moonshot: Towards controllable video generation and editing with multimodal conditions. arXiv preprint arXiv:2401.01827, 2024. 3 +[63] Lvmin Zhang and Maneesh Agrawala. Transparent image layer diffusion using latent transparency. arXiv preprint arXiv:2402.17113, 2024. 2, 3, 6, 7 +[64] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1, 3 \ No newline at end of file diff --git a/2025/TransPixeler_ Advancing Text-to-Video Generation with Transparency/images.zip b/2025/TransPixeler_ Advancing Text-to-Video Generation with Transparency/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5e12be7e4a1eb62ab749b6edc2f067e7e94e563b --- /dev/null +++ b/2025/TransPixeler_ Advancing Text-to-Video Generation with Transparency/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2881932a8b3e5efedb6a351a38cbfcd6a8f145a405f7f84b83bcf6c3ce0f53a4 +size 1066555 diff --git a/2025/TransPixeler_ Advancing Text-to-Video Generation with Transparency/layout.json b/2025/TransPixeler_ Advancing Text-to-Video Generation with Transparency/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f07e7b2b711509fbd3d0f60542f3da525b8b0538 --- /dev/null +++ b/2025/TransPixeler_ Advancing Text-to-Video Generation with Transparency/layout.json @@ -0,0 +1,8896 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 87, + 103, + 523, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 103, + 523, + 121 + ], + "spans": [ + { + "bbox": [ + 87, + 103, + 523, + 121 + ], + "type": "text", + "content": "TransPixel: Advancing Text-to-Video Generation with Transparency" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 142, + 500, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 142, + 500, + 185 + ], + "spans": [ + { + "bbox": [ + 107, + 142, + 500, + 185 + ], + "type": "text", + "content": "Luozhou Wang" + }, + { + "bbox": [ + 107, + 142, + 500, + 185 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 107, + 142, + 500, + 185 + ], + "type": "text", + "content": " Yijun Li" + }, + { + "bbox": [ + 107, + 142, + 500, + 185 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + }, + { + "bbox": [ + 107, + 142, + 500, + 185 + ], + "type": "text", + "content": " Zhifei Chen" + }, + { + "bbox": [ + 107, + 142, + 500, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 142, + 500, + 185 + ], + "type": "text", + "content": " Jui-Hsien Wang" + }, + { + "bbox": [ + 107, + 142, + 500, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 107, + 142, + 500, + 185 + ], + "type": "text", + "content": " Zhifei Zhang" + }, + { + "bbox": [ + 107, + 142, + 500, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 107, + 142, + 500, + 185 + ], + "type": "text", + "content": " He Zhang" + }, + { + "bbox": [ + 107, + 142, + 500, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 107, + 142, + 500, + 185 + ], + "type": "text", + "content": " Zhe Lin" + }, + { + "bbox": [ + 107, + 142, + 500, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 107, + 142, + 500, + 185 + ], + "type": "text", + "content": " Ying-Cong Chen" + }, + { + "bbox": [ + 107, + 142, + 500, + 185 + ], + "type": "inline_equation", + "content": "^{1,2‡}" + }, + { + "bbox": [ + 107, + 142, + 500, + 185 + ], + "type": "text", + "content": " HKUST(GZ) " + }, + { + "bbox": [ + 107, + 142, + 500, + 185 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 107, + 142, + 500, + 185 + ], + "type": "text", + "content": " HKUST " + }, + { + "bbox": [ + 107, + 142, + 500, + 185 + ], + "type": "inline_equation", + "content": "{}^{3}" + }, + { + "bbox": [ + 107, + 142, + 500, + 185 + ], + "type": "text", + "content": " Adobe Research" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 208, + 552, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 208, + 552, + 219 + ], + "spans": [ + { + "bbox": [ + 67, + 208, + 552, + 219 + ], + "type": "text", + "content": "\"A transparent glass of water with ice cubes gently swirling\" \"A statue crumbling to dust as cracks spread across its surface\"" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 220, + 553, + 259 + ], + "blocks": [ + { + "bbox": [ + 58, + 220, + 553, + 259 + ], + "lines": [ + { + "bbox": [ + 58, + 220, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 58, + 220, + 553, + 259 + ], + "type": "image", + "image_path": "fa8aca5e592d8e24b12576bf117460f0e6eefa10cc52f7a7789bcb1758982018.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 258, + 261, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 258, + 261, + 270 + ], + "spans": [ + { + "bbox": [ + 107, + 258, + 261, + 270 + ], + "type": "text", + "content": "\"A squirrel's bushy tail flicking quickly\"" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 328, + 258, + 533, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 258, + 533, + 269 + ], + "spans": [ + { + "bbox": [ + 328, + 258, + 533, + 269 + ], + "type": "text", + "content": "\"A small explosion rapidly expanding and contracting\"" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 58, + 269, + 553, + 306 + ], + "blocks": [ + { + "bbox": [ + 58, + 269, + 553, + 306 + ], + "lines": [ + { + "bbox": [ + 58, + 269, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 58, + 269, + 553, + 306 + ], + "type": "image", + "image_path": "4a74fac0f2482707f8a80c4f6a995b0f40124e6b1f30dd18977e4f543d21ff5a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 307, + 554, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 307, + 554, + 318 + ], + "spans": [ + { + "bbox": [ + 59, + 307, + 554, + 318 + ], + "type": "text", + "content": "\"A white dandelion shifting as seen through a magnifying glass\" \"A massive storm forming, with swirling clouds and thunderbolts\"" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 58, + 318, + 553, + 357 + ], + "blocks": [ + { + "bbox": [ + 58, + 318, + 553, + 357 + ], + "lines": [ + { + "bbox": [ + 58, + 318, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 58, + 318, + 553, + 357 + ], + "type": "image", + "image_path": "2004066e27c07f6898c00ea7d3000c91b2d3555dd3634b34292669e97a1de2fe.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 357, + 298, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 357, + 298, + 368 + ], + "spans": [ + { + "bbox": [ + 67, + 357, + 298, + 368 + ], + "type": "text", + "content": "\"A motorcycle drifting and swerving in an enchanted forest\"" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 58, + 368, + 553, + 407 + ], + "blocks": [ + { + "bbox": [ + 58, + 368, + 553, + 407 + ], + "lines": [ + { + "bbox": [ + 58, + 368, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 58, + 368, + 553, + 407 + ], + "type": "image", + "image_path": "4e7c9e12ea80879b8a98b389c7161ced727c42c023a09982f198405cef2be474.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 409, + 555, + 433 + ], + "lines": [ + { + "bbox": [ + 55, + 409, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 555, + 433 + ], + "type": "text", + "content": "Figure 1. RGBA Video Generation with TransPixeler. By introducing LoRA layers into DiT-based text-to-video model with a novel alpha channel adaptive attention mechanism, our method enables RGBA video generation from text while preserving Text-to-Video quality." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 152, + 447, + 200, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 447, + 200, + 460 + ], + "spans": [ + { + "bbox": [ + 152, + 447, + 200, + 460 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 54, + 473, + 297, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 473, + 297, + 676 + ], + "spans": [ + { + "bbox": [ + 54, + 473, + 297, + 676 + ], + "type": "text", + "content": "Text-to-video generative models have made significant strides, enabling diverse applications in entertainment, advertising, and education. However, generating RGBA video, which includes alpha channels for transparency, remains a challenge due to limited datasets and the difficulty of adapting existing models. Alpha channels are crucial for visual effects (VFX), allowing transparent elements like smoke and reflections to blend seamlessly into scenes. We introduce TransPixeler, a method to extend pretrained video models for RGBA generation while retaining the original RGB capabilities. TransPixeler leverages a diffusion transformer (DiT) architecture, incorporating alpha-specific tokens and using LoRA-based fine-tuning to jointly generate RGB and alpha channels with high consistency. By optimizing attention mechanisms, TransPixeler preserves the strengths of the original RGB model and achieves strong alignment between RGB and alpha channels despite limited training" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 449, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 498 + ], + "type": "text", + "content": "data. Our approach effectively generates diverse and consistent RGBA videos, advancing the possibilities for VFX and interactive content creation. The code is available at https://wileewang.github.io/TransPixeler/." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 525, + 395, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 525, + 395, + 537 + ], + "spans": [ + { + "bbox": [ + 314, + 525, + 395, + 537 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 546, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 546, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 546, + 555, + 715 + ], + "type": "text", + "content": "Text-to-Video generative models have quickly advanced, achieving impressive results [7, 16, 20, 26, 47, 49, 56, 61, 64]. This progress has enabled various applications, such as video editing [11, 13, 32, 39, 52, 55], image animation [2, 14, 15, 38], and motion customization [18, 24, 31, 36, 48, 51, 59]. Diffusion Transformers (DiT) enhance these models by using self-attention to capture long-range dependencies [3, 26, 56, 64]. These models are now widely used in entertainment, advertising, and education, meeting the demand for customizable, dynamic content. Notably, Text-to-RGBA (A denotes Alpha channel) video generation is invaluable for VFX and creative industries. The inclusion of an alpha channel in RGBA formats allows for transparent effects, enabling seamless blending of elements like smoke" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 684, + 268, + 694 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 684, + 268, + 694 + ], + "spans": [ + { + "bbox": [ + 66, + 684, + 268, + 694 + ], + "type": "text", + "content": "*This work was done during an internship at Adobe Research." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 67, + 694, + 120, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 694, + 120, + 704 + ], + "spans": [ + { + "bbox": [ + 67, + 694, + 120, + 704 + ], + "type": "text", + "content": "†Project Leader" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 67, + 704, + 143, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 704, + 143, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 704, + 143, + 714 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 67, + 704, + 143, + 714 + ], + "type": "text", + "content": " Corresponding author" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18229" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "and reflections (see Fig. 1). This transparency creates realistic visuals that can integrate smoothly into scenes without modifying the background. Such flexibility is crucial in gaming, virtual reality (VR), and augmented reality (AR), where dynamic and interactive content is in high demand." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 134, + 295, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 295, + 338 + ], + "type": "text", + "content": "Currently, no direct solutions exist for RGBA video generation, which remains a challenging task due to the scarcity of RGBA video data, with only around 484 videos available in [29]. This scarcity will significantly limit the diversity of generated content, resulting in a constrained set of object types and motion patterns. One feasible solution is to use video matting [28, 30, 40] to obtain alpha channels from generated videos. However, these methods are still limited by the scarcity of RGBA video data and struggle to generalize to a wider range of objects, as shown in Fig. 2 (b). Other video segmentation methods, such as SAM-2 [41], may generalize well to different tasks. However, they cannot generate alpha channels and are therefore unsuitable for direct compositing. There have been attempts to generate RGBA at the image level, such as LayerDiffusion [63]. However, adapting its concept directly to a temporal VAE used in video generative models remains challenging." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 340, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 295, + 544 + ], + "type": "text", + "content": "In this paper, we explore how to extend pretrained video models to generate corresponding alpha channels while retaining the original capabilities of pretrained models. Our goal is to generate content beyond the limitations of the current RGBA training set. Existing works such as Lotus [19] and Marigold [25] demonstrate that leveraging pretrained generation model weights significantly enhances out-of-distribution in dense prediction, hinting at the potential for predicting alpha channels. However, in the context of RGBA video generation, these approaches typically require generating RGB channels first, followed by separate alpha channel prediction. Consequently, information flows unidirectionally from RGB to alpha, keeping the two processes largely disconnected. Given the limited availability of RGBA video data, this imbalance results in insufficient alpha prediction when challenging objects are generated, as shown in Fig. 2 (c)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": "In this work, we propose TransPixeler, which effectively adapts the pretrained RGB video models to generate RGB channels and the alpha channel simultaneously. We leverage state-of-the-art DiT-like video generation models [26, 56], and additionally introduce new tokens appended after text and RGB tokens for generating the alpha channels. To facilitate convergence, we reinitialize the positional embeddings for the alpha tokens and introduce a zero-initialized, learnable domain embedding to distinguish alpha tokens from RGB tokens. Furthermore, we employ a LoRA-based fine-tuning scheme [23], applied exclusively to project alpha tokens into the qkv space, thereby maintaining RGB generation quality. With the proposed approach, we extend the modality while preserving the original input" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 317, + 106, + 433, + 142 + ], + "blocks": [ + { + "bbox": [ + 326, + 79, + 426, + 99 + ], + "lines": [ + { + "bbox": [ + 326, + 79, + 426, + 99 + ], + "spans": [ + { + "bbox": [ + 326, + 79, + 426, + 99 + ], + "type": "text", + "content": "\"A dust cloud expanding after an explosion, covering the area\"" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 106, + 433, + 142 + ], + "lines": [ + { + "bbox": [ + 317, + 106, + 433, + 142 + ], + "spans": [ + { + "bbox": [ + 317, + 106, + 433, + 142 + ], + "type": "image", + "image_path": "34008102bfd8108f497fd3fd41ebde0fcc0feff8e7d017d9886db7d99c250570.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 142, + 399, + 149 + ], + "lines": [ + { + "bbox": [ + 351, + 142, + 399, + 149 + ], + "spans": [ + { + "bbox": [ + 351, + 142, + 399, + 149 + ], + "type": "text", + "content": "(a) Generated RGB" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 435, + 72, + 552, + 107 + ], + "blocks": [ + { + "bbox": [ + 435, + 72, + 552, + 107 + ], + "lines": [ + { + "bbox": [ + 435, + 72, + 552, + 107 + ], + "spans": [ + { + "bbox": [ + 435, + 72, + 552, + 107 + ], + "type": "image", + "image_path": "17e662660344896e39aa36eb2d635ff1a4771d6f004070fa93e42f1dceb82f6f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 436, + 107, + 552, + 141 + ], + "blocks": [ + { + "bbox": [ + 436, + 107, + 552, + 141 + ], + "lines": [ + { + "bbox": [ + 436, + 107, + 552, + 141 + ], + "spans": [ + { + "bbox": [ + 436, + 107, + 552, + 141 + ], + "type": "image", + "image_path": "9da6bbff8b484f935d3ab969a2120a863ceeeaad75ede1c3bbf7dd6fcbda69e2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 150, + 433, + 221 + ], + "blocks": [ + { + "bbox": [ + 317, + 150, + 433, + 221 + ], + "lines": [ + { + "bbox": [ + 317, + 150, + 433, + 221 + ], + "spans": [ + { + "bbox": [ + 317, + 150, + 433, + 221 + ], + "type": "image", + "image_path": "d35bffe7a1d10ee01dd24fc1036a8b64a7e4aac9a88501f277919b96949a08e3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 222, + 428, + 230 + ], + "lines": [ + { + "bbox": [ + 322, + 222, + 428, + 230 + ], + "spans": [ + { + "bbox": [ + 322, + 222, + 428, + 230 + ], + "type": "text", + "content": "(c) with generative prior (MariGold, Lotus)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 237, + 553, + 282 + ], + "lines": [ + { + "bbox": [ + 313, + 237, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 553, + 282 + ], + "type": "text", + "content": "Figure 2. Comparison between Generation-Then-Prediction and our Joint Generation approach. Given the generated RGB in (a), (b) and (c) show the predicted alpha (top) and the composited result (bottom). In (d), the top shows the jointly generated alpha." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 435, + 150, + 552, + 221 + ], + "blocks": [ + { + "bbox": [ + 443, + 142, + 542, + 150 + ], + "lines": [ + { + "bbox": [ + 443, + 142, + 542, + 150 + ], + "spans": [ + { + "bbox": [ + 443, + 142, + 542, + 150 + ], + "type": "text", + "content": "(b) w/o generative prior (Video Matting)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 435, + 150, + 552, + 221 + ], + "lines": [ + { + "bbox": [ + 435, + 150, + 552, + 221 + ], + "spans": [ + { + "bbox": [ + 435, + 150, + 552, + 221 + ], + "type": "image", + "image_path": "bbbfaad957332e8796d944f08bfba63473a5fc23707ff91956938563e333f6c0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 222, + 503, + 230 + ], + "lines": [ + { + "bbox": [ + 481, + 222, + 503, + 230 + ], + "spans": [ + { + "bbox": [ + 481, + 222, + 503, + 230 + ], + "type": "text", + "content": "(d)Ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 290, + 553, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 553, + 314 + ], + "type": "text", + "content": "output structure and relying on the existing attention mechanism through LoRA adaptation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 315, + 554, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 554, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 554, + 541 + ], + "type": "text", + "content": "The extended sequence contains text, RGB, and alpha tokens, with self-attention divided into a 3x3 grouped attention matrix involving interactions like Text-attend-to-RGB (Text as query, RGB as key) and others. We also systematically analyze the attention mechanisms for RGBA generation: 1) Text-attend-to-RGB and RGB-attend-to-Text. The interaction between text and RGB tokens represents original model's generation capabilities. Minimizing the impact on text and RGB tokens during these attention computation processes can better retain the original model's performance; 2) RGB-attend-to-Alpha. We reveals a fundamental limitation in conventional methods is the lack of RGB-attend-to-Alpha attention. This attention is necessary to refine RGB tokens based on alpha information, improving RGB-alpha alignment; 3) Text-attend-to-Alpha. We remove this attention mechanism to reduce the risk caused by limited training data, which could degrade the model's performance. This removal also enhances the retention of the model's original capabilities." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 543, + 553, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 553, + 590 + ], + "type": "text", + "content": "By integrating these techniques, our method achieves diverse RGBA generation with limited training data while maintaining strong RGB-alpha alignment. To summarize, our contributions are as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 592, + 553, + 711 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 313, + 592, + 553, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 553, + 628 + ], + "type": "text", + "content": "- We propose an RGBA video generation framework using DiT models that requires limited data and training parameters, achieving diverse generation with strong alignment." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 628, + 553, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 553, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 553, + 675 + ], + "type": "text", + "content": "- We analyze the role of each attention component in the generation process, optimize their interactions, and introduce necessary modifications to improve RGBA generation quality." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 676, + 553, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 676, + 553, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 676, + 553, + 711 + ], + "type": "text", + "content": "- Our method is validated through extensive experiments, demonstrating its effectiveness across a variety of challenging scenarios." + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18230" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 296, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 296, + 201 + ], + "type": "text", + "content": "Text-to-Video Generation. Early video generation models were primarily based on Unet-based latent diffusion models (LDMs) extended from text-to-image models like Stable Diffusion [42]. For example,AnimateDiff [16] introduced a temporal attention module to improve temporal consistency across frames. Subsequent video generation models [5, 7, 8, 47, 61, 62] adopted an alternating approach with 2D spatial and 1D temporal attention, including works like ModelScope, VideoCrafter, Moonshot, and Show-1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 202, + 296, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 202, + 296, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 296, + 406 + ], + "type": "text", + "content": "With advancements in large language models (LLMs) and the introduction of Sora [3], attention shifted from Unet architectures to transformer-based architectures (DiT). DiT-based video generation models, such as Latte [37] and OpenSora [64], extended the DiT text-to-image (T2I) model [9] and maintained the 2D and 1D alternating attention approach, achieving promising results. Recently, DiT-based video generation has rapidly progressed, achieving further improvements in quality. Several methods [26, 44, 56] have moved away from the 2D and 1D alternating approach, instead treating video frames as a single long sequence with 3D positional embeddings for encoding. These approaches also prepend text tokens—processed through a text encoder—to the video sequence, creating a streamlined network that relies solely on full self-attention and feedforward layers. Our method builds upon these recent open-source transformer-based video generation models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 412, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 296, + 616 + ], + "type": "text", + "content": "Video Matting. A straightforward approach for RGBA video generation is to extract the alpha channel from generated RGB content, as done with traditional green screen keying or learning-based video matting expert models [28-30]. OmnimatteRF [28] introduces a video matting method that combines dynamic 2D foreground layers with a 3D background model, enabling more realistic scene reconstruction for real-world videos. Robust Video Matting (RVM) [30] proposes a real-time, high-quality human video matting method with a recurrent architecture for improved temporal coherence, achieving state-of-the-art results without auxiliary inputs. Another work presents a high-speed, high-resolution background replacement technique with precise alpha matte extraction, supported by the VideoMatte240K and PhotoMatte13K/85 datasets [29]. Additionally, many image matting methods [4, 6, 27, 50, 57] can be applied for frame-by-frame matting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "Further, several works [19, 25, 53] in image depth estimation adapt pretrained generation models for prediction tasks, achieving strong performance that often surpasses traditional, scratch-trained expert models. Marigold [25] modifies architectures to create image-conditioned generation models, while Lotus [19] explores the role of the diffusion process in this context. Although there is currently no dedicated approach for video matting within video gen" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "content": "eration models, we replicate and extend these methods to evaluate their performance, allowing us to highlight the limitations of prediction-based pipelines for RGBA generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 113, + 555, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 113, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 113, + 555, + 340 + ], + "type": "text", + "content": "Generation beyond RGB. Another category of methods [1, 17, 34, 35, 54, 60, 63] explores expanding generation models to simultaneously generate additional channels, though they are not specifically designed for RGBA video generation. For instance, LayerDiffusion [63] modifies the VAE in latent diffusion models to decode alpha channels. However, VAEs typically lack the semantic understanding required for precise alpha generation, limiting their effectiveness in complex visual scenarios where texture and contour details are critical. In contrast, other approaches [1, 34, 35, 60] modify the denoising model directly to enable joint generation. Wonder3D [34] uses a domain embedding to control the model's generation modality, while methods like IntrinsicDiffusion [35] and RGB " + }, + { + "bbox": [ + 313, + 113, + 555, + 340 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 313, + 113, + 555, + 340 + ], + "type": "text", + "content": " X [60] adapt the UNet's input and output layers to jointly produce intrinsic modalities. However, all these methods are designed for image tasks and rely on UNet architectures. When applied to video generation, they face limitations in quality and diversity due to the scarcity of RGBA video data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 350, + 370, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 350, + 370, + 361 + ], + "spans": [ + { + "bbox": [ + 314, + 350, + 370, + 361 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 369, + 393, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 393, + 383 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 393, + 383 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 388, + 555, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 591 + ], + "type": "text", + "content": "We first introduce the open-sourced state-of-the-art DiT-based video generation models [44, 56]. The core components of DiT-based video models are attention modules, and there are two primary distinctions between these models and previous approaches. On one hand, unlike previous models that alternate between 1D temporal attention and 2D spatial attention [5, 7, 8, 64], current methods typically employ 3D spatio-temporal attention, allowing them to capture spatio-temporal dependencies more effectively. On the other hand, instead of using cross-attention for text conditioning, these models concatenate text tokens " + }, + { + "bbox": [ + 313, + 388, + 555, + 591 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{text}}" + }, + { + "bbox": [ + 313, + 388, + 555, + 591 + ], + "type": "text", + "content": " with visual tokens " + }, + { + "bbox": [ + 313, + 388, + 555, + 591 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{video}}" + }, + { + "bbox": [ + 313, + 388, + 555, + 591 + ], + "type": "text", + "content": " into a single long sequence. The shape of video tokens and text tokens are " + }, + { + "bbox": [ + 313, + 388, + 555, + 591 + ], + "type": "inline_equation", + "content": "B\\times L\\times D" + }, + { + "bbox": [ + 313, + 388, + 555, + 591 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 388, + 555, + 591 + ], + "type": "inline_equation", + "content": "B\\times L_{\\mathrm{text}}\\times D" + }, + { + "bbox": [ + 313, + 388, + 555, + 591 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 313, + 388, + 555, + 591 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 388, + 555, + 591 + ], + "type": "text", + "content": " equals to batch size, " + }, + { + "bbox": [ + 313, + 388, + 555, + 591 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{text}}" + }, + { + "bbox": [ + 313, + 388, + 555, + 591 + ], + "type": "text", + "content": " equals to the length of text tokens, " + }, + { + "bbox": [ + 313, + 388, + 555, + 591 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 388, + 555, + 591 + ], + "type": "text", + "content": " equals to the length of video tokens and " + }, + { + "bbox": [ + 313, + 388, + 555, + 591 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 388, + 555, + 591 + ], + "type": "text", + "content": " equals to the latent dimension of transformer. Full self-attention is then applied across the combined sequence:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 323, + 597, + 553, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 597, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 323, + 597, + 553, + 666 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {T}}{\\sqrt {d _ {k}}}\\right) \\mathbf {V}, \\quad \\text {w h e r e} \\\\ \\mathbf {Z}: \\mathbf {Z} \\in \\{\\mathbf {Q}, \\mathbf {K}, \\mathbf {V} \\} \\\\ = \\left[ \\mathbf {W} _ {z: z \\in \\{q, k, v \\}} \\left(\\mathbf {x} _ {\\text {t e x t}}\\right); \\mathbf {f} _ {z: z \\in \\{q, k, v \\}} \\left(\\mathbf {x} _ {\\text {v i d e o}}\\right) \\right] \\tag {1} \\\\ \\end{array}", + "image_path": "63c3e8105385fc24ac5915cabb194bd59876d7573d86846ed28bd2f3925769a3.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_t" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": " (for " + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "t \\in \\{q, k, v\\}" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": ") represents the projection matrix in the transformer model, and " + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_t" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": " (for " + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "t \\in \\{q, k, v\\}" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": ") represents a combined operation that incorporates both the projection and positional encoding for visual tokens. There" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18231" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 550, + 310 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 550, + 310 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 550, + 310 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 550, + 310 + ], + "type": "image", + "image_path": "2a64c8a397ea52e395ebf393c55464d63b2bac93108dde1919ed4c9f786c78fa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 317, + 555, + 352 + ], + "lines": [ + { + "bbox": [ + 55, + 317, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 555, + 352 + ], + "type": "text", + "content": "Figure 3. Pipeline of TransPixeler. Our method is organized as follows: (1) Left: we extend the input of DiT to include new alpha tokens; (2) Top Center: we initialize alpha tokens with our positional encoding; (3) Bottom Center: we insert a partial LoRA and adjust attention computation during training and inference." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 365, + 295, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 389 + ], + "type": "text", + "content": "are two commonly used types of positional encoding. One is absolute positional encoding formulated as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 396, + 295, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 396, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 66, + 396, + 295, + 410 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {z: z \\in \\{q, k, v \\}} \\left(\\mathbf {x} _ {\\text {v i d e o}}\\right) := \\mathbf {W} _ {z: z \\in \\{q, k, v \\}} \\left(\\mathbf {x} _ {\\text {v i d e o}} ^ {m} + \\mathbf {p} ^ {m}\\right), \\tag {2}", + "image_path": "3a53b4c0929da0d9a6b986acee09cf93b5d889a428a3bb6e36010702bfeeb7f1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 416, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 296, + 464 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 416, + 296, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 55, + 416, + 296, + 464 + ], + "type": "text", + "content": " is the positional embedding (e.g., a sinusoidal function) and " + }, + { + "bbox": [ + 55, + 416, + 296, + 464 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 416, + 296, + 464 + ], + "type": "text", + "content": " denotes the position of each RGB video token. Another approach is the Rotary Position Embedding (RoPE) [43], often used by [44, 56]. This is expressed as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 71, + 470, + 295, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 470, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 71, + 470, + 295, + 485 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {z: z \\in \\{q, k \\}} \\left(\\mathbf {x} _ {\\text {v i d e o}}\\right) := \\mathbf {W} _ {z: z \\in \\{q, k \\}} \\left(\\mathbf {x} _ {\\text {v i d e o}} ^ {m}\\right) \\circ e ^ {i m \\theta}, \\tag {3}", + "image_path": "beb2f68c2fc09cc6daeb3a534384ca45e4f7fdd58a05d1e003135d1939696962.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 490, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 296, + 515 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 490, + 296, + 515 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 490, + 296, + 515 + ], + "type": "text", + "content": " is the positional index, " + }, + { + "bbox": [ + 55, + 490, + 296, + 515 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 490, + 296, + 515 + ], + "type": "text", + "content": " is the imaginary unit for rotation, and " + }, + { + "bbox": [ + 55, + 490, + 296, + 515 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 490, + 296, + 515 + ], + "type": "text", + "content": " is the rotation angle." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 521, + 147, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 147, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 147, + 533 + ], + "type": "text", + "content": "3.2. Our Approach" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 539, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 295, + 574 + ], + "type": "text", + "content": "To jointly generate RGB and alpha videos, we adapt a pretrained RGB video generation model through several modifications. The whole pipeline is visualized in Fig. 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 575, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 296, + 647 + ], + "type": "text", + "content": "Firstly, we double the sequence length of noisy input tokens to enable the model to generate videos of double length, from " + }, + { + "bbox": [ + 55, + 575, + 296, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{video}}^{1:L}" + }, + { + "bbox": [ + 55, + 575, + 296, + 647 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 575, + 296, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{video}}^{1:2*L}" + }, + { + "bbox": [ + 55, + 575, + 296, + 647 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 575, + 296, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{video}}^{1:L}" + }, + { + "bbox": [ + 55, + 575, + 296, + 647 + ], + "type": "text", + "content": " will be decoded into the RGB video, while " + }, + { + "bbox": [ + 55, + 575, + 296, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{video}}^{L+1:2*L}" + }, + { + "bbox": [ + 55, + 575, + 296, + 647 + ], + "type": "text", + "content": " will be decoded into the corresponding alpha video. The Query(Q), Key(K), Value(V) representations are formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 83, + 654, + 296, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 654, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 83, + 654, + 296, + 685 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {Z}: \\mathbf {Z} \\in \\{\\mathbf {Q}, \\mathbf {K}, \\mathbf {V} \\} \\\\ = \\left[ \\mathbf {W} _ {z: z \\in \\{q, k, v \\}} \\left(\\mathbf {x} _ {\\text {t e x t}}\\right); \\mathbf {f} _ {z: z \\in \\{q, k, v \\}} \\left(\\mathbf {x} _ {\\text {v i d e o}} ^ {1: 2 * L}\\right) \\right] \\tag {4} \\\\ \\end{array}", + "image_path": "c2a46a64e6644a03f48f2c35471c6e765294993224be926e7b790b4664ed868f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "In addition to sequence doubling, we explored increasing batch size or latent dimensions and splitting output into two" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 365, + 553, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 553, + 388 + ], + "type": "text", + "content": "domains; however, these approaches showed limited effectiveness under constrained datasets, which we discuss later." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 389, + 554, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 554, + 449 + ], + "type": "text", + "content": "Secondly, we modify the positional encoding function " + }, + { + "bbox": [ + 313, + 389, + 554, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{t:t\\in \\{q,k,v\\}}(\\cdot)" + }, + { + "bbox": [ + 313, + 389, + 554, + 449 + ], + "type": "text", + "content": ", as shown in Fig. 4. Instead of continuously numbering indices, we allow RGB and alpha tokens to share the same positional encoding. Taking absolute positional encoding as an example:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 322, + 459, + 553, + 508 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 459, + 553, + 508 + ], + "spans": [ + { + "bbox": [ + 322, + 459, + 553, + 508 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {f} _ {z: z \\in \\{q, k, v \\}} ^ {*} (\\mathbf {x} _ {\\mathrm {v i d e o}}) \\\\ := \\left\\{ \\begin{array}{l l} \\mathbf {W} _ {z: z \\in \\{q, k, v \\}} \\left(\\mathbf {x} _ {\\text {v i d e o}} ^ {m} + \\mathbf {p} ^ {m}\\right), & \\text {i f} m \\leq L, \\\\ \\mathbf {W} _ {z: z: z \\in \\{q, k, v \\}} ^ {*} \\left(\\mathbf {x} _ {\\text {v i d e o}} ^ {m} + \\mathbf {p} ^ {m - L} + d\\right), & \\text {i f} m > L. \\end{array} \\right. \\tag {5} \\\\ \\end{array}", + "image_path": "c285cd82bd3e55468ecddbd436749396513798fed471bf4fc9cdc72fc244362d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 516, + 555, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 555, + 624 + ], + "type": "text", + "content": "Here we introduce a domain embedding " + }, + { + "bbox": [ + 313, + 516, + 555, + 624 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 516, + 555, + 624 + ], + "type": "text", + "content": ", initialized to zero. We make it learnable to help the model adaptively differentiate between RGB (" + }, + { + "bbox": [ + 313, + 516, + 555, + 624 + ], + "type": "inline_equation", + "content": "m \\leq L" + }, + { + "bbox": [ + 313, + 516, + 555, + 624 + ], + "type": "text", + "content": ") and alpha tokens (" + }, + { + "bbox": [ + 313, + 516, + 555, + 624 + ], + "type": "inline_equation", + "content": "m > L" + }, + { + "bbox": [ + 313, + 516, + 555, + 624 + ], + "type": "text", + "content": "). The motivation behind this design is we observe that with same positional encoding, even initializing with different noises, the tokens from two domains tend to generate same results. It minimizes spatial-temporal alignment challenges at the very beginning of training and thus accelerates convergence." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 624, + 554, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 554, + 658 + ], + "type": "text", + "content": "Next we propose a fine-tuning scheme using LoRA [23], in which the LoRA layer is applied only to alpha domain tokens:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 340, + 667, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 667, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 340, + 667, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {W} _ {z: z \\in \\{q, k, v \\}} ^ {*} (\\mathbf {x} _ {\\text {v i d e o}} ^ {m} + \\mathbf {p} ^ {m - L} + d) \\\\ = \\mathbf {W} _ {z: z \\in \\{q, k, v \\}} \\left(\\mathbf {x} _ {\\text {v i d e o}} ^ {m} + \\mathbf {p} ^ {m - L} + d\\right) \\tag {6} \\\\ + \\gamma \\cdot \\operatorname {L o R A} \\left(\\mathbf {x} _ {\\text {v i d e o}} ^ {m} + \\mathbf {p} ^ {m - L} + d\\right), \\quad \\text {i f} m > L, \\\\ \\end{array}", + "image_path": "c031d656733efe4d69099121f6aeef0339a983a035c95724852cdef57e03ad1b.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "18232" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 175, + 205 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 175, + 205 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 175, + 205 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 175, + 205 + ], + "type": "image", + "image_path": "2dfe9edaadb401c0c7b99995eef3ead0d14062fdf567400f84dc2e31ef8b9f45.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 211, + 296, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 211, + 296, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 296, + 257 + ], + "type": "text", + "content": "Figure 4. Positional Encoding Design for RGBA Generation. Assigning alpha tokens the same positional encoding as RGB yields similar results, resulting in faster convergence after 1000 iterations compared to standard encoding strategies." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 177, + 71, + 296, + 205 + ], + "blocks": [ + { + "bbox": [ + 177, + 71, + 296, + 205 + ], + "lines": [ + { + "bbox": [ + 177, + 71, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 177, + 71, + 296, + 205 + ], + "type": "image", + "image_path": "fa90b265e7bc6f20c1513283d02a7d10f28b187c9ef294c62afbc29f15416826.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 267, + 296, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 267, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 267, + 296, + 326 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 267, + 296, + 326 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 267, + 296, + 326 + ], + "type": "text", + "content": " controls the residual strength. Additionally, we design an attention mask to block unwanted attention computation. Given a text-video token sequence length " + }, + { + "bbox": [ + 55, + 267, + 296, + 326 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{text}} + 2L" + }, + { + "bbox": [ + 55, + 267, + 296, + 326 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 267, + 296, + 326 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{text}}" + }, + { + "bbox": [ + 55, + 267, + 296, + 326 + ], + "type": "text", + "content": " represents text token length, the mask is defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 65, + 331, + 296, + 364 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 331, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 65, + 331, + 296, + 364 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {m n} ^ {*} = \\left\\{ \\begin{array}{l l} - \\infty , & \\text {i f} m \\leq L _ {\\text {t e x t}} \\text {a n d} n > L _ {\\text {t e x t}} + L, \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {7}", + "image_path": "8535c1a557d60003fb5288dbbbdab5df9524e28078b4ae5be786ced8b77e4657.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 369, + 296, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 393 + ], + "type": "text", + "content": "Combining these modifications, inference with our method is expressed as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 399, + 304, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 304, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 304, + 427 + ], + "type": "text", + "content": "Attention " + }, + { + "bbox": [ + 55, + 399, + 304, + 427 + ], + "type": "inline_equation", + "content": "(\\mathbf{Q},\\mathbf{K},\\mathbf{V}) =" + }, + { + "bbox": [ + 55, + 399, + 304, + 427 + ], + "type": "text", + "content": " softmax " + }, + { + "bbox": [ + 55, + 399, + 304, + 427 + ], + "type": "inline_equation", + "content": "\\left(\\frac{\\mathbf{Q}\\mathbf{K}^T}{\\sqrt{d_k}} +\\mathbf{M}^*\\right)\\mathbf{V}," + }, + { + "bbox": [ + 55, + 399, + 304, + 427 + ], + "type": "text", + "content": " where" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 429, + 296, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 429, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 296, + 467 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {Z}: \\mathbf {Z} \\in \\{\\mathbf {Q}, \\mathbf {K}, \\mathbf {V} \\} \\\\ = \\left[ \\mathbf {W} _ {z: z \\in \\{q, k, v \\}} \\left(\\mathbf {x} _ {\\text {t e x t}}\\right); \\mathbf {f} _ {z: z \\in \\{q, k, v \\}} ^ {*} \\left(\\mathbf {x} _ {\\text {v i d e o}}\\right) \\right] \\tag {8} \\\\ \\end{array}", + "image_path": "4bcdb3c805305cd7d5d7e576bdc24ba397d6df29c10b4a6be67fc3cd121b4efb.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 468, + 296, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 493 + ], + "type": "text", + "content": "Training is carried out using flow matching [33] or a traditional diffusion process [21]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 498, + 118, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 118, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 118, + 510 + ], + "type": "text", + "content": "3.3. Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 517, + 296, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 296, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 296, + 625 + ], + "type": "text", + "content": "Given our goal of maximizing the inherited capabilities of the pretrained video model, enabling it to generate beyond the existing RGBA training set, we analyze the most critical component within our current 3D full attention DiT video generation model: the attention mechanism. The attention matrix, " + }, + { + "bbox": [ + 55, + 517, + 296, + 625 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}\\mathbf{K}^T" + }, + { + "bbox": [ + 55, + 517, + 296, + 625 + ], + "type": "text", + "content": ", has dimensions " + }, + { + "bbox": [ + 55, + 517, + 296, + 625 + ], + "type": "inline_equation", + "content": "(L_{\\mathrm{text}} + 2*L)\\times (L_{\\mathrm{text}} + 2*L)" + }, + { + "bbox": [ + 55, + 517, + 296, + 625 + ], + "type": "text", + "content": " which we simplify by organizing it into a " + }, + { + "bbox": [ + 55, + 517, + 296, + 625 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 55, + 517, + 296, + 625 + ], + "type": "text", + "content": " grouped attention matrix—including Text-attend-to-RGB, RGB-attend-to-Text, and so forth, as illustrated in Fig. 3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": "Text-Attend-to-RGB and RGB-Attend-to-Text. These represent the upper-left 2x2 section of and are computations that exist solely in the original RGB generation model. If we ensure that this part of the computation remains unaffected, we can replicate the original RGB generation performance. Therefore, we limit the scope of LoRA's influence, as defined in Eq. (7), by retaining the original QKV" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 554, + 230 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 554, + 230 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 554, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 554, + 230 + ], + "type": "image", + "image_path": "c9b8423a27459dac1d0b53393a47298ee96f360a56ce64bdc65a2c93d3fe10b3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 237, + 555, + 291 + ], + "lines": [ + { + "bbox": [ + 313, + 237, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 555, + 291 + ], + "type": "text", + "content": "Figure 5. Attention Rectification. (a) Eliminating all attention from alpha as a key preserves " + }, + { + "bbox": [ + 313, + 237, + 555, + 291 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 313, + 237, + 555, + 291 + ], + "type": "text", + "content": " RGB generation but leads to poor alignment. (b) Retaining all attention significantly degrades quality, causing a lack of motion in bicycles. (c) Our method achieves an effective balance." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 307, + 554, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 554, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 554, + 330 + ], + "type": "text", + "content": "values for both text and RGB tokens, thus preserving the pretrained model's behavior in these domains." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 331, + 554, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 554, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 554, + 403 + ], + "type": "text", + "content": "Besides the partial LoRA, the added alpha tokens requires the text and RGB tokens to also act as queries and interact with the alpha tokens as keys, which alters the computation in this 2x2 attention matrix. Therefore, we further analyze two additional attention computations that impact RGB generation, as shown in Fig. 5." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 407, + 554, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 554, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 554, + 528 + ], + "type": "text", + "content": "Text-Attend-to-Alpha. We find that this attention is detrimental to the generation quality. Since the model was originally trained with text and RGB data, introducing attention from text to alpha causes interference due to the domain gap between alpha and RGB. Specifically, the alpha modality provides only contour information and lacks the rich texture, color, and semantic details associated with the text prompt, thereby degrading generation quality. To mitigate this, we design the attention mask (Eq. (7)) that blocks this computation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 533, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 554, + 628 + ], + "type": "text", + "content": "RGB-Attend-to-Alpha. In contrast, we identify RGB-to-Alpha as essential for successful joint generation. This attention allows the model to refine RGB tokens by considering alpha information, facilitating alignment between generated RGB and alpha channels. This refinement process is a critical component missing in previous generation-then-prediction pipelines, which lacked a feedback mechanism for RGB refinement based on alpha guidance." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 633, + 391, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 633, + 391, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 633, + 391, + 647 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": "Training Dataset. We utilize the public VideoMatte240K dataset [29], a comprehensive collection of 484 high-resolution green screen videos consists of 240,709 unique frames of alpha mattes and foregrounds. These frames provide a diverse range of human subjects, clothing styles, and" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18233" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 149, + 72, + 217, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 72, + 217, + 83 + ], + "spans": [ + { + "bbox": [ + 149, + 72, + 217, + 83 + ], + "type": "text", + "content": "\"A coin spinning\"" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 321, + 72, + 539, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 72, + 539, + 83 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 539, + 83 + ], + "type": "text", + "content": "\"A forest floor being consumed by spreading magical fire\"" + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 84, + 552, + 121 + ], + "blocks": [ + { + "bbox": [ + 58, + 84, + 552, + 121 + ], + "lines": [ + { + "bbox": [ + 58, + 84, + 552, + 121 + ], + "spans": [ + { + "bbox": [ + 58, + 84, + 552, + 121 + ], + "type": "image", + "image_path": "9ad2feb3bc0b485ab9216134a217a323e3b64244335dabe5c2968eb5ff6aafcc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 151, + 121, + 213, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 121, + 213, + 132 + ], + "spans": [ + { + "bbox": [ + 151, + 121, + 213, + 132 + ], + "type": "text", + "content": "\"A parrot flying\"" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 329, + 121, + 530, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 121, + 530, + 131 + ], + "spans": [ + { + "bbox": [ + 329, + 121, + 530, + 131 + ], + "type": "text", + "content": "\"An asteroid belt swirling chaotically through space\"" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 130, + 552, + 168 + ], + "blocks": [ + { + "bbox": [ + 58, + 130, + 552, + 168 + ], + "lines": [ + { + "bbox": [ + 58, + 130, + 552, + 168 + ], + "spans": [ + { + "bbox": [ + 58, + 130, + 552, + 168 + ], + "type": "image", + "image_path": "3aa9427e38388071c24ba8ca88f69ad51f27ef890046422777b14dc2475743c8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 72, + 169, + 293, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 169, + 293, + 180 + ], + "spans": [ + { + "bbox": [ + 72, + 169, + 293, + 180 + ], + "type": "text", + "content": "\"An astronaut running down an alley, spacesuit flapping\"" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 320, + 169, + 539, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 169, + 539, + 179 + ], + "spans": [ + { + "bbox": [ + 320, + 169, + 539, + 179 + ], + "type": "text", + "content": "\"A cloud of dust erupting and dispersing like an explosion\"" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 58, + 179, + 552, + 217 + ], + "blocks": [ + { + "bbox": [ + 58, + 179, + 552, + 217 + ], + "lines": [ + { + "bbox": [ + 58, + 179, + 552, + 217 + ], + "spans": [ + { + "bbox": [ + 58, + 179, + 552, + 217 + ], + "type": "image", + "image_path": "4c672db82e44dc88d8e75fd27d55a402b8f03225126632b86bb6cee1b8e61837.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 88, + 218, + 279, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 218, + 279, + 228 + ], + "spans": [ + { + "bbox": [ + 88, + 218, + 279, + 228 + ], + "type": "text", + "content": "\"a woman's long black hair streaming as she runs\"" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 370, + 218, + 488, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 218, + 488, + 228 + ], + "spans": [ + { + "bbox": [ + 370, + 218, + 488, + 228 + ], + "type": "text", + "content": "\"Water splattering in mid-air\"" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 58, + 228, + 552, + 267 + ], + "blocks": [ + { + "bbox": [ + 58, + 228, + 552, + 267 + ], + "lines": [ + { + "bbox": [ + 58, + 228, + 552, + 267 + ], + "spans": [ + { + "bbox": [ + 58, + 228, + 552, + 267 + ], + "type": "image", + "image_path": "17e6491a7d7345d1c9fbf1fbdb199912f466c0bb76aa2ad4cab5ed8758ed8037.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 270, + 111, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 111, + 281 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 111, + 281 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 164, + 270, + 252, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 270, + 252, + 279 + ], + "spans": [ + { + "bbox": [ + 164, + 270, + 252, + 279 + ], + "type": "text", + "content": "Generated RGBA Video" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 270, + 364, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 270, + 364, + 279 + ], + "spans": [ + { + "bbox": [ + 316, + 270, + 364, + 279 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 418, + 270, + 506, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 270, + 506, + 279 + ], + "spans": [ + { + "bbox": [ + 418, + 270, + 506, + 279 + ], + "type": "text", + "content": "Generated RGBA Video" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 58, + 279, + 552, + 317 + ], + "blocks": [ + { + "bbox": [ + 58, + 279, + 552, + 317 + ], + "lines": [ + { + "bbox": [ + 58, + 279, + 552, + 317 + ], + "spans": [ + { + "bbox": [ + 58, + 279, + 552, + 317 + ], + "type": "image", + "image_path": "cbcf13f23b412511108c5d7945343918b0bf1b15c8623e2ef802f2f894a0843d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 58, + 319, + 552, + 358 + ], + "blocks": [ + { + "bbox": [ + 58, + 319, + 552, + 358 + ], + "lines": [ + { + "bbox": [ + 58, + 319, + 552, + 358 + ], + "spans": [ + { + "bbox": [ + 58, + 319, + 552, + 358 + ], + "type": "image", + "image_path": "860d0ae2f5c5dc8142886fb2392e10b8db15807f99d218763bd64e876a1a7788.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 365, + 523, + 376 + ], + "lines": [ + { + "bbox": [ + 85, + 365, + 523, + 376 + ], + "spans": [ + { + "bbox": [ + 85, + 365, + 523, + 376 + ], + "type": "text", + "content": "Figure 6. Applications. Top: Text-to-Video with Transparency. Bottom: Image-to-Video generation with transparency." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 391, + 295, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 295, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 295, + 427 + ], + "type": "text", + "content": "poses. We apply fundamental preprocessing steps for them, including color decontamination and background blurring. Prompts are extracted using ShareGPT4V [10]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 431, + 296, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 296, + 624 + ], + "type": "text", + "content": "Model. Our RGBA video diffusion models are developed by fine-tuning pre-trained diffusion models. Specifically, we employ two models based on the diffusion transformer architecture: the open-source model CogVideoX [56] and a modified variant of CogVideoX denoted as " + }, + { + "bbox": [ + 55, + 431, + 296, + 624 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 55, + 431, + 296, + 624 + ], + "type": "text", + "content": ". CogVideoX generates RGB videos in 480x720 with 49 frames at 8 FPS, using 50 sampling steps. In contrast, the modified version produces videos in 176x320 with 64 frames at 24 FPS, while also using 50 sampling steps. Additionally, we integrate our method with CogVideoX-I2V (Image-to-Video) to support image-to-video generation with transparency. We set the LoRA rank to 128. For domain embedding, we initialize it with an original shape of " + }, + { + "bbox": [ + 55, + 431, + 296, + 624 + ], + "type": "inline_equation", + "content": "1 \\times D" + }, + { + "bbox": [ + 55, + 431, + 296, + 624 + ], + "type": "text", + "content": " and zero values, then expand it to " + }, + { + "bbox": [ + 55, + 431, + 296, + 624 + ], + "type": "inline_equation", + "content": "L \\times D" + }, + { + "bbox": [ + 55, + 431, + 296, + 624 + ], + "type": "text", + "content": " through repetition during training. We train these parameters over 5,000 iterations with a batch size of 8 in total, utilizing 8 NVIDIA A100 GPUs." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 630, + 137, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 137, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 137, + 643 + ], + "type": "text", + "content": "4.1. Applications" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 648, + 288, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 288, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 288, + 661 + ], + "type": "text", + "content": "We mainly demonstrate two applications shown in Fig. 6:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Text-to-Video with Transparency. Our method is capable of generating moving objects with various types of motion, such as spinning, running, and flying, while also handling transparent properties of bottles and glasses. Additionally," + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 391, + 553, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 553, + 415 + ], + "type": "text", + "content": "it can produce complex visual effects, including fire, explosions, cracking, and lightning, as well as creative examples." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 419, + 554, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 554, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 554, + 492 + ], + "type": "text", + "content": "Image-to-Video with Transparency. Our method can also be integrated with an I2V video generation model-CogVideoX-I2V. Users can provide a single image along with an alpha channel (optional), and then we generate subsequent frames with dynamic effects and automatically propagate or generate alpha channels for these frames." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 498, + 400, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 400, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 400, + 511 + ], + "type": "text", + "content": "4.2. Comparisons" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 555, + 661 + ], + "type": "text", + "content": "Generation-then-Prediction Pipeline. As shown in Fig. 2, video matting methods [29, 40, 58] struggle with matting non-human objects (see supplementary materials for additional results). Therefore, we selected Lotus [19] and SAM-2 [41] as baselines due to their stronger generalization: Lotus uses pretrained generative models, and SAM-2 is trained on large datasets. Since Lotus was originally designed for single-image depth estimation, we extended it for RGBA videos, denoted as Lotus + RGBA in our comparisons. Qualitative results are shown in Fig. 7. Since ground-truth alpha channels are not available for generated videos, we focus on qualitative comparison." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Joint Generation Pipeline. Since there are currently no existing RGBA video generation models, we integrateAnimateDiff [16] with LayerDiffusion [63] to generate RGBA videos. We use the open-source video generation model" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18234" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 542, + 333 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 542, + 333 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 542, + 333 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 542, + 333 + ], + "type": "image", + "image_path": "5e794d65b0e64c08a69372d77a3984427867574a027fc1bd2c13d2864f7c0cb6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 338, + 515, + 351 + ], + "lines": [ + { + "bbox": [ + 94, + 338, + 515, + 351 + ], + "spans": [ + { + "bbox": [ + 94, + 338, + 515, + 351 + ], + "type": "text", + "content": "Figure 7. Comparison with Generation-then-Prediction Pipelines. Our method demonstrates superior alignment." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 359, + 555, + 548 + ], + "blocks": [ + { + "bbox": [ + 58, + 359, + 555, + 548 + ], + "lines": [ + { + "bbox": [ + 58, + 359, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 58, + 359, + 555, + 548 + ], + "type": "image", + "image_path": "c711bae9c19dcde842b045268589df63b130b192213524f7c7345e25b9a12ae7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 555, + 555, + 581 + ], + "lines": [ + { + "bbox": [ + 55, + 555, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 555, + 581 + ], + "type": "text", + "content": "Figure 8. Comparison with Joint Generation Pipelines. Top: LayerDiffusion +AnimateDiff; Bottom: Ours. Our method achieves better alignment and generates corresponding motion described by prompts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 593, + 297, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 297, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 297, + 617 + ], + "type": "text", + "content": "CogVideoX [56] as the base model for fair comparison. The qualitative results are illustrated in Fig. 8." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "User Study. We also conduct a user study with Amazon Mechanical Turk to compare two joint generation methods, as shown in Table. 1. Participants are asked to evaluate two key aspects: 1) whether the RGB and alpha align correctly; and 2) whether the motion in the generated video matches the corresponding text description. A total of 30 videos are generated from distinct text prompts, and 87 users partici" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 315, + 604, + 553, + 642 + ], + "blocks": [ + { + "bbox": [ + 394, + 591, + 473, + 602 + ], + "lines": [ + { + "bbox": [ + 394, + 591, + 473, + 602 + ], + "spans": [ + { + "bbox": [ + 394, + 591, + 473, + 602 + ], + "type": "text", + "content": "Table 1. User Study." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 604, + 553, + 642 + ], + "lines": [ + { + "bbox": [ + 315, + 604, + 553, + 642 + ], + "spans": [ + { + "bbox": [ + 315, + 604, + 553, + 642 + ], + "type": "table", + "html": "
RGBA AlignmentMotion Quality
AnimateDiff [16]+LayerDiff [63]6.7%21.7%
Ours + CogVideoX [56]93.3%78.3%
", + "image_path": "a7d14d6c5739910c0d57ec8fb202acb5a68c6db25d1c176f7b2329a53092097a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 698, + 555, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 698, + 555, + 723 + ], + "spans": [ + { + "bbox": [ + 313, + 698, + 555, + 723 + ], + "type": "text", + "content": "pated in the evaluation. The study shows that our method is obviously favored more by users with higher votes." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18235" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 294, + 219 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 294, + 219 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 294, + 219 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 294, + 219 + ], + "type": "image", + "image_path": "ac4e7cabe7a36d1cf518758d685e1fef9c367ad148b7172bb82a725a6b8d8f60.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 228, + 295, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 228, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 295, + 251 + ], + "type": "text", + "content": "Figure 9. Alternative Designs for Joint Generation with DiT. Sequence extension (b) represents our method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 270, + 149, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 149, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 149, + 283 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 289, + 295, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 295, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 295, + 313 + ], + "type": "text", + "content": "As shown in Fig. 10, we conduct the ablation study across two dimensions: attention rectification and network design." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 57, + 323, + 293, + 463 + ], + "blocks": [ + { + "bbox": [ + 57, + 323, + 293, + 463 + ], + "lines": [ + { + "bbox": [ + 57, + 323, + 293, + 463 + ], + "spans": [ + { + "bbox": [ + 57, + 323, + 293, + 463 + ], + "type": "image", + "image_path": "d6bf7fd2750b87cfbfd3643198d9261b8fc4726048d0f6b665e2e821ee28ca91.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 472, + 296, + 529 + ], + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 529 + ], + "type": "text", + "content": "Figure 10. Ablation Study. (a) Ours; (b) Ours without RGB-attend-to-Alpha; (c) Ours with Text-attend-to-alpha; (d) Batch Extension Strategy; (e) Latent Dimension Extension Strategy. Our method maintains high-quality motion generation (e.g., butterflies waving their wings) while achieving good alignment." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 543, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 295, + 651 + ], + "type": "text", + "content": "Attention Rectification. By blocking RGB-to-Alpha attention, we first validate the importance of RGB-to-Alpha attention for aligning RGB and alpha channels, a feature lacking in most prediction-based methods. We also examine the effect of removing unnecessary attention to preserve the model's generative capacity, by learning Text-to-Alpha attention only. Without RGB-to-Alpha attention, the alpha channel misaligns with RGB content and the RGB output loses motion quality (e.g., reverse rocket)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Alternative Designs For Joint Generation. Given the transformer's input dimensions " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "B \\times L \\times D" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": ", we extend the sequence dimension " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " to produce RGB and alpha channels, but alternative extensions are possible at the Batch " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " or Latent Dimension " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " levels (see Fig. 9). In the Batch Ex" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 326, + 71, + 541, + 174 + ], + "blocks": [ + { + "bbox": [ + 326, + 71, + 541, + 174 + ], + "lines": [ + { + "bbox": [ + 326, + 71, + 541, + 174 + ], + "spans": [ + { + "bbox": [ + 326, + 71, + 541, + 174 + ], + "type": "image", + "image_path": "1a36e93509c667fc9aed9e6448e8b4f1697c0d429cf9ed16f89ce4eb5b956bf3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 176, + 555, + 209 + ], + "lines": [ + { + "bbox": [ + 313, + 176, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 176, + 555, + 209 + ], + "type": "text", + "content": "Figure 11. Quantitative Evaluation. Our approach achieves a good balance between alignment (low flow difference) and preserving generative quality (low FVD)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 223, + 555, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 555, + 319 + ], + "type": "text", + "content": "tension approach, a new module enables inter-batch communication, similar to the technique in [46]. For Latent Dimension Extension, we merge video and alpha tokens, project them into the DiT model's latent space, and unmerge post-generation, using learnable linear layers with fine-tuning. Batch Extension shows weaker RGB-alpha alignment, while Latent Dimension Extension, though akin to training from scratch, significantly reduces diversity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 323, + 556, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 556, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 556, + 539 + ], + "type": "text", + "content": "Evaluation. In addition to the qualitative comparisons shown in Fig. 10, we also generated a total of 80 videos, each consisting of 64 frames, and evaluated them using two primary metrics: Flow Difference. To measure alignment between the generated RGB and Alpha videos, we use optical flow [22] to focus on motion consistency while ignoring appearance. Specifically, we calculate optical flow with Farneback method [12] and compute the flow difference as the average Euclidean distance between RGB and Alpha flow fields. Frechet Video Distance (FVD). We use FVD [45] to compare the RGB videos generated by each RGBA method against those from the original RGB model, evaluating how well each method preserves the model's original generative quality. A lower FVD indicates that the generated results are closer to the original RGB model in terms of motion coherence and diversity, thus demonstrating a high fidelity to the model's intended generative quality. Results are shown in Fig. 11." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 549, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 388, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "We introduce a novel Text-to-RGBA video generator that extends RGB diffusion models to produce RGBA output with minimal modifications and retained fidelity. Leveraging DiT backbones and RGBA-aware attention, the method preserves RGB quality while accurately inferring alpha channels. Targeted adjustments—adding alpha tokens, reinitializing positional embeddings, and applying selective LoRA fine-tuning—deliver rich, high-quality RGBA videos even from limited training data. Extensive quantitative and qualitative experiments validate the framework's versatility and robustness across challenging animation, complex compositing, and diverse real-world scenes." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18236" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Fan Bao, Shen Nie, Kaiwen Xue, Chongxuan Li, Shi Pu, Yaole Wang, Gang Yue, Yue Cao, Hang Su, and Jun Zhu. One transformer fits all distributions in multi-modal diffusion at scale. In International Conference on Machine Learning, pages 1692-1717. PMLR, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "type": "text", + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 202, + 295, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 295, + 255 + ], + "type": "text", + "content": "[3] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "type": "text", + "content": "[4] Ryan D Burgert, Brian L Price, Jason Kuen, Yijun Li, and Michael S Ryoo. Magick: A large-scale captioned dataset from matting generated images using chroma keying. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22595-22604, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 313, + 294, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 313, + 294, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 313, + 294, + 345 + ], + "type": "text", + "content": "[5] cerspense. zeroscope_v2. https://huggingface.co/ cerspense/zeroscope_v2_576w, 2023. Accessed: 2023-02-03. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 346, + 295, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 295, + 390 + ], + "type": "text", + "content": "[6] Guowei Chen, Yi Liu, Jian Wang, Juncai Peng, Yuying Hao, Lutao Chu, Shiyu Tang, Zewu Wu, Zeyu Chen, Zhiliang Yu, et al. Pp-matting: high-accuracy natural image matting. arXiv preprint arXiv:2204.09433, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 391, + 295, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 391, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 391, + 295, + 446 + ], + "type": "text", + "content": "[7] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 447, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 295, + 491 + ], + "type": "text", + "content": "[8] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocraft2: Overcoming data limitations for high-quality video diffusion models. arXiv preprint arXiv:2401.09047, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 491, + 295, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 295, + 546 + ], + "type": "text", + "content": "[9] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "type": "text", + "content": "[10] Lin Chen, Jisong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "text", + "content": "[11] Weifeng Chen, Yatai Ji, Jie Wu, Hefeng Wu, Pan Xie, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 636, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 680 + ], + "type": "text", + "content": "[12] Gunnar Farnebäck. Two-frame motion estimation based on polynomial expansion. In Proceedings of the Scandinavian Conference on Image Analysis (SCIA), pages 363-370. Springer, 2003. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[13] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arXiv:2307.10373, 2023. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[14] Jianzhu Guo, Dingyun Zhang, Xiaogiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Liveportrait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 172 + ], + "type": "text", + "content": "[15] Xun Guo, Mingwu Zheng, Liang Hou, Yuan Gao, Yufan Deng, Chongyang Ma, Weiming Hu, Zhengjun Zha, Haibin Huang, Pengfei Wan, et al. I2v-adapter: A general image-to-video adapter for video diffusion models. arXiv preprint arXiv:2312.16693, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "type": "text", + "content": "[16] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 1, 3, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "text", + "content": "[17] Hao He, Yixun Liang, Luozhou Wang, Yuanhao Cai, Xinli Xu, Hao-Xiang Guo, Xiang Wen, and Yingcong Chen. *Lucidfusion: Generating 3d gaussians with arbitrary unposed images*, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 264, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 307 + ], + "type": "text", + "content": "[18] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrl: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 308, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 363 + ], + "type": "text", + "content": "[19] Jing He, Haodong Li, Wei Yin, Yixun Liang, Leheng Li, Kaiqiang Zhou, Hongbo Liu, Bingbing Liu, and YingCong Chen. Lotus: Diffusion-based visual foundation model for high-quality dense prediction. arXiv preprint arXiv:2409.18124, 2024. 2, 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 365, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 408 + ], + "type": "text", + "content": "[20] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 410, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 553, + 442 + ], + "type": "text", + "content": "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 444, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 465 + ], + "type": "text", + "content": "[22] Berthold KP Horn and Brian G Schunck. Determining optical flow. Artificial intelligence, 17(1-3):185–203, 1981. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 467, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 467, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 467, + 553, + 510 + ], + "type": "text", + "content": "[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 512, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 555 + ], + "type": "text", + "content": "[24] Hyeonho Jeong, Jinho Chang, Geon Yeong Park, and Jong Chul Ye. Dreammotion: Space-time self-similarity score distillation for zero-shot video editing. arXiv preprint arXiv:2403.12002, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 556, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 553, + 621 + ], + "type": "text", + "content": "[25] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 624, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 644 + ], + "type": "text", + "content": "[26] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 1, 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 680 + ], + "type": "text", + "content": "[27] Jiachen Li, Jitesh Jain, and Humphrey Shi. Matting anything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1775-1785, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[28] Geng Lin, Chen Gao, Jia-Bin Huang, Changil Kim, Yipeng Wang, Matthias Zwicker, and Ayush Saraf. Omnimatterf: Robust omnimatte with 3d background modeling. In" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18237" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23471-23480, 2023. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "text", + "content": "[29] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian L Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8762-8771, 2021. 2, 3, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 217 + ], + "type": "text", + "content": "[30] Shanchuan Lin, Linjie Yang, Imran Saleemi, and Soumyadip Sengupta. Robust high-resolution video matting with temporal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 238-247, 2022. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "text", + "content": "[31] Pengyang Ling, Jiazi Bu, Pan Zhang, Xiaoyi Dong, Yuhang Zang, Tong Wu, Huaian Chen, Jiaqi Wang, and Yi Jin. Motionclone: Training-free motion cloning for controllable video generation. arXiv preprint arXiv:2406.05338, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "type": "text", + "content": "[32] Shaoteng Liu, Yuechen Zhang, Wenbo Li, Zhe Lin, and Jiaya Jia. Video-p2p: Video editing with cross-attention control. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8599–8608, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 294, + 342 + ], + "type": "text", + "content": "[33] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 344, + 295, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 295, + 409 + ], + "type": "text", + "content": "[34] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9970-9980, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 295, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 295, + 466 + ], + "type": "text", + "content": "[35] Jundan Luo, Duygu Ceylan, Jae Shin Yoon, Nanxuan Zhao, Julien Philip, Anna Fruhstick, Wenbin Li, Christian Richardt, and Tuanfeng Wang. Intrinsicdiffusion: Joint intrinsic layers from latent diffusion models. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 468, + 295, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 295, + 499 + ], + "type": "text", + "content": "[36] Wan-Duo Kurt Ma, John P Lewis, and W Bastiaan Kleijn. Trailblazer: Trajectory control for diffusion-based video generation. arXiv preprint arXiv:2401.00896, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 295, + 544 + ], + "type": "text", + "content": "[37] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 546, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 295, + 600 + ], + "type": "text", + "content": "[38] Muyao Niu, Xiaodong Cun, Xintao Wang, Yong Zhang, Ying Shan, and Yinqiang Zheng. Mofa-video: Controllable image animation via generative motion field adoptions in frozen image-to-video diffusion model. arXiv preprint arXiv:2405.20222, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 602, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 295, + 656 + ], + "type": "text", + "content": "[39] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15932-15942, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "text", + "content": "[40] Haotong Qin, Lei Ke, Xudong Ma, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Xianglong Liu, and Fisher Yu. Bimatting: Efficient video matting via binarization. Advances in Neural Information Processing Systems, 36: 43307-43321, 2023. 2, 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 149 + ], + "type": "text", + "content": "[41] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dólár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 150, + 555, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 555, + 205 + ], + "type": "text", + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 206, + 553, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 247 + ], + "type": "text", + "content": "[43] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 250, + 456, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 456, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 456, + 259 + ], + "type": "text", + "content": "[44] Genmo Team. Mochi, 2024. 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 261, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 293 + ], + "type": "text", + "content": "[45] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. 2019. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "type": "text", + "content": "[46] Shimon Vainer, Mark Boss, Mathias Parger, Konstantin Kutsy, Dante De Nigris, Ciara Rowles, Nicolas Perony, and Simon Donne. Collaborative control for geometry-conditioned pbr image generation. arXiv preprint arXiv:2402.05919, 2024. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "text", + "content": "[47] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 393, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 436 + ], + "type": "text", + "content": "[48] Luozhou Wang, Guibao Shen, Yixun Liang, Xin Tao, Pengfei Wan, Di Zhang, Yijun Li, and Yingcong Chen. Motion inversion for video customization. arXiv preprint arXiv:2403.20193, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "type": "text", + "content": "[49] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 493, + 553, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 535 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 535 + ], + "type": "text", + "content": "[50] Zhixiang Wang, Baiang Li, Jian Wang, Yu-Lun Liu, Jinwei Gu, Yung-Yu Chuang, and Shin'Ichi Satoh. Matting by generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 537, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 592 + ], + "type": "text", + "content": "[51] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 593, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 658 + ], + "type": "text", + "content": "[52] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "type": "text", + "content": "[53] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18238" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 586 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[54] Shuai Yang, Zhifei Chen, Pengguang Chen, Xi Fang, Shu Liu, and Yingcong Chen. Defect spectrum: A granular look of large-scale defect datasets with rich semantics, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "text", + "content": "[55] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "type": "text", + "content": "[56] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "text", + "content": "[57] Jingfeng Yao, Xinggang Wang, Shusheng Yang, and Baoyuan Wang. Vitmatte: Boosting image matting with pretrained plain vision transformers. Information Fusion, 103: 102091, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "type": "text", + "content": "[58] Jingfeng Yao, Xinggang Wang, Lang Ye, and Wenyu Liu. Matte anything: Interactive natural image matting with segment anything model. Image and Vision Computing, 147: 105067, 2024. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "type": "text", + "content": "[59] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 397 + ], + "type": "text", + "content": "[60] Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloš Hasan. Rgb ↔ x: Image decomposition and synthesis using material-and lighting-aware diffusion models. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "type": "text", + "content": "[61] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. arXiv preprint arXiv:2309.15818, 2023. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 454, + 294, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 454, + 294, + 507 + ], + "spans": [ + { + "bbox": [ + 56, + 454, + 294, + 507 + ], + "type": "text", + "content": "[62] David Junhao Zhang, Dongxu Li, Hung Le, Mike Zheng Shou, Caiming Xiong, and Doyen Sahoo. Moonshot: Towards controllable video generation and editing with multimodal conditions. arXiv preprint arXiv:2401.01827, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 510, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 294, + 542 + ], + "type": "text", + "content": "[63] Lvmin Zhang and Maneesh Agrawala. Transparent image layer diffusion using latent transparency. arXiv preprint arXiv:2402.17113, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 544, + 294, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 586 + ], + "type": "text", + "content": "[64] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1, 3" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18239" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Transfer Your Perspective_ Controllable 3D Generation from Any Viewpoint in a Driving Scene/0e711492-3ab2-4748-8147-c897ce7d76b8_content_list.json b/2025/Transfer Your Perspective_ Controllable 3D Generation from Any Viewpoint in a Driving Scene/0e711492-3ab2-4748-8147-c897ce7d76b8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8a7a3d77b3bb97b11743c6251bc5558e5f7b8ad3 --- /dev/null +++ b/2025/Transfer Your Perspective_ Controllable 3D Generation from Any Viewpoint in a Driving Scene/0e711492-3ab2-4748-8147-c897ce7d76b8_content_list.json @@ -0,0 +1,1815 @@ +[ + { + "type": "text", + "text": "Transfer Your Perspective: Controllable 3D Generation from Any Viewpoint in a Driving Scene", + "text_level": 1, + "bbox": [ + 156, + 130, + 841, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tai-Yu Pan1, Sooyoung Jeon1, Mengdi Fan1, Jinsu Yoo1, Zhenyang Feng1, Mark Campbell2, Kilian Q. Weinberger2, Bharath Hariharan2, Wei-Lun Chao1", + "bbox": [ + 189, + 202, + 803, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The Ohio State University, 2Cornell University", + "bbox": [ + 305, + 250, + 684, + 268 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c5eae7588b627f3995606de2f3865121e75062d88ec3559091c23eeb26d94532.jpg", + "image_caption": [ + "(a) Ego-centric driving" + ], + "image_footnote": [], + "bbox": [ + 93, + 292, + 295, + 398 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e27f28dbad465dcc1139d5991ac0b7a3ccbfe08dc906e1ba7a7f66b9b36cbfc7.jpg", + "image_caption": [ + "(b) TYP for dynamic agent", + "Figure 1. Illustration of the proposed problem and solution, Transfer Your Perspective (TYP). (a) A given sensory data captured by the ego-car (red triangle). (b) A generated sensory data by TYP, seeing from the viewpoint of another vehicle (green triangle) in the same scene. (c) A generated sensory data, seeing from an imaginary static agent like roadside units (blue icon). (d) Putting all the sensory data together, given or generated, TYP enables the development of collaborative perception with little or no real collaborative driving data." + ], + "image_footnote": [], + "bbox": [ + 297, + 292, + 496, + 398 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3d13adc16fd89ace3f2f1190193f2e2518d328b72661b1d52aa989107ca99dc0.jpg", + "image_caption": [ + "(c) TYP for static agent" + ], + "image_footnote": [], + "bbox": [ + 500, + 292, + 699, + 398 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cd67161cbbb17dfff66d9eb055a4998bbe8690fa8f314a7ce0a6edecc97feb90.jpg", + "image_caption": [ + "(d) Collaborative driving" + ], + "image_footnote": [], + "bbox": [ + 700, + 292, + 903, + 398 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 494, + 326, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Self-driving cars relying solely on ego-centric perception face limitations in sensing, often failing to detect occluded, faraway objects. Collaborative autonomous driving (CAV) seems like a promising direction, but collecting data for development is non-trivial. It requires placing multiple sensor-equipped agents in a real-world driving scene, simultaneously! As such, existing datasets are limited in locations and agents. We introduce a novel surrogate to the rescue, which is to generate realistic perception from different viewpoints in a driving scene, conditioned on a real-world sample—the ego-car's sensory data. This surrogate has huge potential: it could potentially turn any ego-car dataset into a collaborative driving one to scale up the development of CAV. We present the very first solution, using a combination of simulated collaborative data and real ego-car data. Our method Transfer Your Perspective (TYP) learns a conditioned diffusion model whose output samples are not only realistic but also consistent in both semantics and layouts with the given ego-car data. Empirical results demonstrate TYP's effectiveness in aiding in a CAV setting. In particular, TYP enables us to (pre-)train collaborative perception algorithms like early and late fusion with little or no real-world collaborative data, greatly facilitating downstream CAV applications.", + "bbox": [ + 86, + 536, + 486, + 885 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 494, + 643, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Seeing the world from an ego-centric perspective, a self-driving car risks being \"narrow-sighted,\" limiting its ability to respond appropriately in dynamic driving environments. For instance, it should slow down or honk if a pedestrian is about to cross the road or if another vehicle is merging into its lane. However, these actions depend on the car's ability to detect traffic participants, which may be occluded by a large bus or a building at a sharp intersection. This occlusion problem can hardly be addressed by mounting a few more sensors on the car. We argue that a self-driving car must move beyond its ego-centric perspective.", + "bbox": [ + 511, + 521, + 908, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One intuitive way is to collaborate with nearby \"agents\" that observe the scene from different angles, such as other sensor-equipped cars or static devices like roadside units (RSUs). When trained together to address GPS errors and synchronization delays [55], collaborative perception has been shown to significantly improve the accuracy of each agent's perception, particularly in detecting occluded or distant objects [3, 4, 14, 19, 23, 33, 42, 45, 50-52]. However, collecting training data for collaborative perception is never easy. Unlike single-agent data, which can be collected by simply driving a car on the road, collaborative data requires the simultaneous presence of multiple agents in the same driving scene. For dynamic agents like vehicles, precise coordination is needed to ensure they are within communication", + "bbox": [ + 511, + 690, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12027", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "range. These challenges limit existing works in scale and the number of agents (typically just two). While one may leverage simulated data by game engines [7], they often fail to capture the diversity of real-world scenes. We thus ask:", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Can we obtain realistic data for learning collaborative perception with much less effort, ideally as easily as single-agent data?", + "text_level": 1, + "bbox": [ + 107, + 156, + 470, + 200 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, given the vast amount of single-agent LiDAR data already collected across various driving environments, is it possible to convert each of them into collaborative perception data by generating additional point clouds from alternative reference viewpoints within the same scene?", + "bbox": [ + 89, + 205, + 483, + 280 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "At first glance, this question may seem overly ambitious for three reasons. First, for collaborative perception to be effective, the generated point cloud must provide information that the ego-car's point cloud cannot capture, such as occluded surfaces invisible to the ego-car. This creates a chicken-and-egg problem: if the ego-car's point cloud is all we initially know about the scene, how can we deduce any additional information? Second, for training purposes, the generated data must be realistic. It should replicate the constraints and data patterns of a real sensor as if it were positioned at the reference viewpoint, producing no points in occluded areas and fewer points in distant areas. Last but not least, in regions commonly perceivable from both the reference and ego-car's viewpoints, the generated data should align in layouts and semantics with the ego-car's data. On the surface, this may seem as simple as copying the ego-car's data, but doing so would violate the second requirement. In essence, the generated data should resemble a point cloud seen from the reference viewpoint, not the ego-car.", + "bbox": [ + 91, + 281, + 483, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "That said, after a deeper look at the question, we argue that it is achievable with three key insights. First, given semantic information around the ego-car for conditioning, existing research has shown promising progress in generating realistic point clouds [13, 27, 32, 44, 49, 62, 63]. This is especially encouraging, as most existing single-agent datasets provide 3D object labels, and they can be translated to obtain semantic information centered around the reference position. This further implies that we can generate occluded object surfaces as seen from the ego-car if they are visible from the reference viewpoint. Second, semantic information is easily editable: even if the translated object map has noticeable empty areas due to limited ego-car perception, object boxes can be manually added to make the map appear more realistic from the reference viewpoint. Lastly, in commonly perceivable regions where the generated data needs to meet two physical constraints, we could leverage simulators based on computational graphics and optics. Specifically, we can use simulated data from the two views to train a conditioned generative model that maps one viewpoint to the other.", + "bbox": [ + 89, + 568, + 482, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building upon these insights, we propose Transfer Your Perspective (TYP), a novel research problem and the very", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "first solution for generating a realistic point cloud from any viewpoint in a scene, given real ego-car's point cloud and semantic labels as conditions. TYP assumes access to 1) a simulated collaborative driving dataset with multiple agents perceiving the same scenes from different viewpoints and 2) a real single-agent driving dataset; both are labeled. Our solution involves a conditioned latent diffusion model [34] and a dedicated two-stage training process. In the first stage, we consider a single-agent scenario and train the model using real data conditioning only on object locations. This equips the model with the ability to generate diverse and realistic scenes. We denote the learned model by $P(\\pmb{x}|\\pmb{y})$ , where $\\pmb{y}$ stands for the semantic condition and $\\pmb{x}$ for the point cloud. In the second stage, we incorporate the simulated data to learn how to ground generation on data from another agent's viewpoint so that the model can produce a semantically consistent reference point cloud given the ego-car's data. We learn a lightweight conditioning module to turn $P(\\pmb{x}|\\pmb{y})$ into $P(\\pmb{x}_r|\\pmb{x}_e,\\pmb{y}_r)$ , where $e$ and $r$ indicate ego and reference views, respectively. We note that $\\pmb{x}_e$ was pre-translated to center around the reference position.", + "bbox": [ + 511, + 90, + 906, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One challenge in TYP is the domain gap between real and simulated data. Due to differences in sensor configurations and placements, data collection environments, and the sim-to-real gap, the two sets of point clouds inevitably exhibit discrepant distributions, patterns, and densities. To address this, we insert a domain adaptation step between the two training stages. We train a separate encoder-decoder for the simulated data while enforcing a constraint to make the encoded features of simulated and real data indistinguishable [40]. This step allows us to learn $P(\\boldsymbol{x}_r|\\boldsymbol{x}_e,\\boldsymbol{y}_r)$ in a space with reduced domain discrepancy during the second stage.", + "bbox": [ + 511, + 411, + 908, + 577 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Once trained, we pair $P(\\pmb{x}_r|\\pmb{x}_e,\\pmb{y}_r)$ with the real data's encoder-decoder to generate real-style point clouds grounding on real ego-car's data, so that we can develop collaborative perception algorithms without real collaborative data. More specifically, given a real ego-car's perception $\\pmb{x}_e$ and label $\\pmb{y}_e$ , we first translate them to center around the reference position, followed by optionally injecting object labels into $\\pmb{y}_e$ to make it realistic from the reference viewpoint.", + "bbox": [ + 511, + 580, + 908, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We extensively validate TYP on multiple datasets, all in an offline setting. Empirical results demonstrate TYP's effectiveness in generating high-quality reference data to aid in the development of collaborative perception. In particular, we show that a conditioned diffusion model trained solely on simulated data (e.g., OPV2V [51]) can already turn a real single-agent dataset into a real-alike collaborative one. As such, one can train collaborative perception algorithms for real test data without real training data. We further generate the \"ColWaymo\" dataset by training TYP on real single-agent Waymo data [39] and simulated OPV2V data [51], thus turning the former into a collaborative one. Collaborative perception backbones pre-trained on the large-scale, semi-", + "bbox": [ + 511, + 704, + 910, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12028", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "synthetic ColWaymo data demonstrate a remarkable transferability. They notably improve collaborative perception developed in the downstream tasks (e.g., V2V4Real [52]), even in a many-shot setting.", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In sum, our key contributions are:", + "bbox": [ + 109, + 152, + 336, + 165 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a new research direction to aid in the development of collaborative perception, generating real-looking sensory data from potentially any viewpoint in a real driving scene, capable of turning a single-agent dataset into a collaborative one. This has the potential to scale up collaborative autonomous driving (CAV).", + "- We present TYP, a two-stage training recipe with domain adaptation in between to learn the generative model using simulated collaborative data and real single-agent data.", + "- Extensive experiments demonstrate TYP's effectiveness in aiding in collaborative perception across various scenarios." + ], + "bbox": [ + 89, + 167, + 485, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 345, + 230, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Collaborative autonomous driving (CAV) offers significant benefits, including extended perception range by sharing sensor information to detect objects beyond a single vehicle's field of view. Despite these benefits, collecting large-scale, real-world datasets for CAV poses significant challenges due to the complexity and cost of deploying multiple instrumented vehicles in diverse environments. Existing datasets often have limitations that restrict the scope of CAV research. For instance, OPV2V [51] and V2X-Sim [20] rely on simulations using the CARLA simulator [7], allowing controlled scenarios with multiple vehicle agents, yet lacking real-world variability. V2V4Real [52], on the other hand, provides real-world data, yet it only contains two collaborating agents, which restricts the exploration of more complex multi-agent interactions. Similarly, the DAIR-V2X [57] dataset also offers real-world data including vehicle-to-infrastructure (V2I) and V2V, but it mainly focuses on V2I scenarios and still has a limited number of collaborating vehicles. To overcome current limitations, we propose a new research direction—generating realistic point clouds given the ego's perspective. 3D generation. Various works have explored 3D scene generation, such as LiDARGen [62], R2DM [27], LiDM [32], UltraLiDAR [48], and RangeLDM [13]. Additionally, Li-darDM [63] and Text2LiDAR [44] investigate conditional scene generation, with the former relying on hand-crafted map layouts and the latter conditioning on text inputs. However, all these methods focus on ego-centric generation. We propose a new research direction for CAV, aiming to generate realistic and consistent scenes conditioned on the ego agent's real point clouds—an area that remains largely unexplored.", + "bbox": [ + 89, + 371, + 485, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models. Diffusion models have recently advanced generative modeling for high-quality LiDAR point clouds and images. Denoising Diffusion Probabilistic Models (DDPMs) [12] outperform traditional Generative Adversarial Networks (GANs) [10] and further enhance efficiency", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ca03f5ab0b91be56465867e6e1791601e932ceaa1d5f9cc20429fa1f896b143e.jpg", + "image_caption": [ + "Figure 2. Illustration of TYP's conditioned generative model and training process. We propose a two-stage training procedure. The first stage maximizes the generation capability by conditioning solely on object locations (using real single-agent target data), while the second stage grounds the generation on the ego-car's perspective to match semantics and layouts (using simulated CAV data). Additionally, we introduce a discriminator to adapt simulated CAV features to the real target domain, making the trained model readily applicable to the target domain after the second stage." + ], + "image_footnote": [], + "bbox": [ + 514, + 92, + 903, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for LiDAR generation. In autonomous driving applications, methods like RangeLDM [13], LidarDM [63], and LiDAR-Gen [62] apply diffusion models for realistic LiDAR scene generation. In this paper, we utilize its generation capability to study our proposed problem.", + "bbox": [ + 511, + 393, + 906, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Transfer Your Perspective (TYP)", + "text_level": 1, + "bbox": [ + 511, + 482, + 816, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we introduce a new research direction to advance collaborative autonomous driving (CAV): generating LiDAR point clouds from different perspectives within the same scene as the ego agent, aiming to reduce the tedious efforts of collecting data for CAV. We begin by defining the proposed problem in Sec. 3.1. In Sec. 3.2, we discuss the representations of the inputs, including point clouds and semantic information. Sec. 3.3 outlines the pipeline developed to address this problem. Finally, in Sec. 3.4, we demonstrate how this capability can be applied to datasets that only have labeled ego agents, e.g., Waymo Open Dataset (WOD) [39].", + "bbox": [ + 511, + 508, + 908, + 675 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Setup", + "text_level": 1, + "bbox": [ + 511, + 681, + 663, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the perception data $\\pmb{x}_e$ from an ego agent, we aim to build a model $P(\\pmb{x}_r|\\pmb{x}_e)$ to generate new perception data $\\pmb{x}_r$ seen from a different reference location and perspective within a communication range. Here, $\\pmb{x}$ represents a LiDAR point cloud. This task presents several challenges, including potential information gaps between the two views, the need for alignment within the commonly perceivable area, and ensuring realism in the generated data, as discussed in Sec. 1.", + "bbox": [ + 511, + 704, + 906, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address these challenges, we extend the original problem by incorporating semantic information, such as object bounding boxes, represented as $P(\\pmb{x}_r|\\pmb{x}_e,\\pmb{y}_r)$ . We assume the availability of this information around the ego agent, denoted by $\\pmb{y}_e$ . For example, most of the existing single-agent", + "bbox": [ + 511, + 825, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12029", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "datasets provide object labels. This semantic information can be easily translated and edited to become $\\mathbf{y}_r$ by tools like traffic re-players [11], making it a key to bridge the information gap between the ego and reference views, enabling the reference agent to \"see\" those objects and surfaces beyond the ego's view—a core concept in CAV. Additionally, by aligning generated point clouds with object locations, this extension facilitates our goal of scaling up CAV development. In essence, the generated $(\\mathbf{x}_r, \\mathbf{y}_r)$ pairs can be used flexibly alongside $(\\mathbf{x}_e, \\mathbf{y}_e)$ in various CAV applications, like directly using them to train collaborative perception algorithms.", + "bbox": [ + 89, + 90, + 483, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the following sections, we describe how $\\pmb{x}$ and $\\pmb{y}$ are encoded, followed by our approach to tackling the problem.", + "bbox": [ + 89, + 257, + 483, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Representations and Embeddings", + "text_level": 1, + "bbox": [ + 89, + 295, + 385, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LiDAR. There are multiple ways to represent point clouds in continuous 3D space, such as coordinates (i.e., x, y, and z) [30, 31, 37], range images [25, 30, 43], and voxelization [17, 24, 53, 59]. To better align with spatial control from object locations, we follow [49] to voxelize point clouds using a pre-defined grid and record voxel occupancy. [49] also highlights that this representation can naturally handle varying point densities and only minimally impacts LiDAR generation, with some precision trade-offs during voxelization. In short, we represent a point cloud by $\\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times C}$ .", + "bbox": [ + 89, + 316, + 483, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To avoid the computational cost of 3D convolutions, we convert $x$ into Bird's-Eye-View (BEV) images by treating the height dimension (i.e., $C$ ) as feature channels for 2D convolutions. This approach has been widely adopted in self-driving perception [5, 48, 58], allowing the use of 2D image-based model architectures and algorithms.", + "bbox": [ + 89, + 468, + 483, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantic information. Given that the point clouds are represented as BEV images, it is also intuitive to represent object locations in BEV. We create binary object maps by considering only the x and y coordinates of 3D bounding boxes, resulting in $\\pmb{y} \\in \\mathbb{R}^{H \\times W \\times 1}$ . (We could further extend it by including category information.)", + "bbox": [ + 89, + 560, + 483, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Feature embedding. Following the literature [34], we encode input tensors using a Vector Quantized-Variational Autoencoder (VQ-VAE) [41], comprising an encoder $E$ , a quantization function $Q$ for feature vectors on spatial grids, and a decoder $G$ . Formally, we obtain $\\boldsymbol{x}^f = E_x(\\boldsymbol{x}) \\in \\mathbb{R}^{h \\times w \\times c_x}$ by the encoder, map each $c_x$ -dimensional vector to a learnable code to obtain $\\boldsymbol{x}^z = Q_x(\\boldsymbol{x}^f) \\in \\mathbb{R}^{h \\times w \\times c_x}$ , and generate outputs using the decoder $\\hat{\\boldsymbol{x}} = G_x(\\boldsymbol{x}^z) \\in \\mathbb{R}^{H \\times W \\times C}$ , where $h$ and $w$ are spatial resolutions of the feature map downsampled from $H$ and $W$ , and $c_x$ is the number of channels.", + "bbox": [ + 89, + 650, + 483, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The VQ-VAE model is trained end-to-end by minimizing:", + "bbox": [ + 109, + 801, + 483, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {v q}} = \\mathcal {L} _ {\\mathrm {r e c}} + \\| \\operatorname {s g} [ E _ {x} (\\boldsymbol {x}) ] - \\boldsymbol {x} ^ {z} \\| _ {2} ^ {2} + \\| \\operatorname {s g} [ \\boldsymbol {x} ^ {z} ] - E _ {x} (\\boldsymbol {x}) \\| _ {2} ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 828, + 483, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{sg}[\\cdot ]$ and $\\mathcal{L}_{\\mathrm{rec}}$ denote the stop-gradient operation and the reconstruction loss, respectively. As our point cloud representation is binary occupancy $\\pmb {x}\\in \\{0,1\\}^{H\\times W\\times C}$ , binary", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "cross-entropy is a natural choice for $\\mathcal{L}_{\\mathrm{rec}}$ . However, due to the sparsity of point clouds, this results in an imbalanced loss. To address this, we adopt the Focal Loss (FL) [36]:", + "bbox": [ + 511, + 90, + 905, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\mathrm {F L}} \\left(\\boldsymbol {x} _ {i}, \\hat {\\boldsymbol {x}} _ {i}\\right) = \\left\\{ \\begin{array}{l l} - (1 - \\hat {\\boldsymbol {x}} _ {i}) ^ {\\gamma} \\log \\left(\\hat {\\boldsymbol {x}} _ {i}\\right) & \\text {i f} \\boldsymbol {x} _ {i} = 1 \\\\ - \\hat {\\boldsymbol {x}} _ {i} ^ {\\gamma} \\log \\left(1 - \\hat {\\boldsymbol {x}} _ {i}\\right) & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 148, + 906, + 181 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r e c}} = \\sum_ {i = 0} ^ {M} \\ell_ {\\mathrm {F L}} \\left(\\boldsymbol {x} _ {i}, \\hat {\\boldsymbol {x}} _ {i}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 196, + 906, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $i$ is the voxel index and $M$ is the number of voxels.", + "bbox": [ + 511, + 246, + 893, + 260 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For object locations $\\pmb{y}$ , we use the same approach to train a separate VQ-VAE model $E_{y}, Q_{y}$ , and $G_{y}$ . In the subsequent subsections, we use $\\pmb{x}^{f} \\in \\mathbb{R}^{h \\times w \\times c_{x}}$ and $\\pmb{y}^{f} \\in \\mathbb{R}^{h \\times w \\times c_{y}}$ as feature embeddings for learning a latent diffusion model.", + "bbox": [ + 511, + 261, + 905, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Transfer Your Perspective by Generation", + "text_level": 1, + "bbox": [ + 511, + 333, + 864, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We model the distribution $P(\\pmb{x}_r | \\pmb{x}_e, \\pmb{y}_r)$ defined in Sec. 3.1 by a conditioned generative model and train it in two stages. We assume access to $(\\pmb{x}_r, \\pmb{x}_e, \\pmb{y}_r)$ tuples as training data. We detail the training data preparation in Sec. 3.4 and Sec. 4.", + "bbox": [ + 511, + 354, + 906, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the first stage, we aim to maximize the generation capability by providing minimal conditions, i.e., only bounding boxes, without further constraints on the scene. This encourages the model to produce $P(\\boldsymbol{x}|\\boldsymbol{y})$ with high flexibility. Second, building on this model, we incorporate the ego agent's point cloud as an additional cue and ground the generation process on it to ensure semantic and layout consistency between the generated and ego agent's point clouds. Put together, this two-stage training procedure enables the learned $P(\\boldsymbol{x}_r|\\boldsymbol{x}_e,\\boldsymbol{y}_r)$ model to generate perception data at areas beyond the commonly visible regions between two views. The model and training process is illustrated in Fig. 2.", + "bbox": [ + 511, + 416, + 906, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More importantly, this approach enables us to leverage existing labeled single-agent datasets such as KITTI [9], NuScenes [2], and Waymo Open Dataset (WOD) [39], since the first-stage training only requires ego-centric LiDAR point clouds and their corresponding semantic information (see Sec. 3.4). The $(\\pmb{x}_e,\\pmb{x}_r)$ data pairs seen from the ego and the reference agents are needed only in the second stage.", + "bbox": [ + 511, + 598, + 905, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the following, we elaborate on each training stage.", + "bbox": [ + 532, + 704, + 883, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage 1: Generation with semantic information. The goal of this stage is to equip the model with strong generation capabilities of point clouds given spatial conditions, preparing it for the next stage. As discussed in Sec. 3.2, the point clouds and object locations are embedded by VQVAEs [41], denoted as $\\boldsymbol{x}^f = E_x(\\boldsymbol{x}) \\in \\mathbb{R}^{h \\times w \\times c_x}$ and $\\boldsymbol{y}^f = E_x(\\boldsymbol{y}) \\in \\mathbb{R}^{h \\times w \\times c_y}$ . The image-like feature maps allow us to adopt existing generation algorithms for 2D images [6, 12, 16]. In this paper, we apply one of the most popular generative models, the Latent Diffusion Model (LDM) [34], for conditioned generation $P(\\boldsymbol{x} | \\boldsymbol{y})$ . LDM seeks to model a data distribution by iteratively denoising variables", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12030", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "that are initially sampled from a Gaussian. The objective is:", + "bbox": [ + 89, + 90, + 478, + 107 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {L D M}} = \\mathbb {E} _ {\\boldsymbol {x} ^ {f}, \\boldsymbol {y} ^ {f}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\boldsymbol {x} _ {t} ^ {f}, \\boldsymbol {y} ^ {f}, t\\right) \\right\\| _ {2} ^ {2} \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 116, + 483, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where, $\\epsilon_{\\theta}$ is a UNet [35] backbone and $t$ is the timestamp.", + "bbox": [ + 89, + 160, + 473, + 175 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage 2: Generation for new perspectives. In this stage, we aim to ground the generation of point clouds by incorporating the ego agent's perception. To retain the generation capability from stage 1, we freeze the learned $P(\\boldsymbol{x}|\\boldsymbol{y})$ and introduce a learnable lightweight control module to inject additional cues, following T2I-Adapter [26], as shown in Fig. 2.", + "bbox": [ + 89, + 176, + 482, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Formally, let $\\pmb{x}_{e'}$ denote the translated and rotated egoagent's point cloud centered around the reference location and aligned with the reference orientation, as shown in Fig. 2; $\\pmb{x}_{e'}^f = E_x(\\pmb{x}_{e'}) \\in \\mathbb{R}^{h \\times w \\times c_x}$ is the corresponding embedding. Let $F_{AD}$ denote the learnable control module, which takes $\\pmb{x}_{e'}^f$ as input and outputs", + "bbox": [ + 89, + 267, + 483, + 357 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {c} = F _ {A D} \\left(\\boldsymbol {x} _ {e ^ {\\prime}} ^ {f}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 227, + 368, + 483, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{F}_c = \\{\\mathbf{F}_c^1,\\mathbf{F}_c^2,\\mathbf{F}_c^3,\\mathbf{F}_c^4\\}$ matches the size of the multiscale features $\\mathbf{F}_{enc} = \\{\\mathbf{F}_{enc}^{1},\\mathbf{F}_{enc}^{2},\\mathbf{F}_{enc}^{3},\\mathbf{F}_{enc}^{4}\\}$ extracted from the frozen encoder of the UNet $\\epsilon_{\\theta}$ . With these ingredients, T2I-Adapter influences the generation process by", + "bbox": [ + 89, + 398, + 483, + 460 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {e n c} ^ {i} \\leftarrow \\mathbf {F} _ {e n c} ^ {i} + \\mathbf {F} _ {c} ^ {i}, i \\in \\{1, 2, 3, 4 \\}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 473, + 483, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which injects $\\pmb{x}_{e'}^f$ embedding into each of the denoising steps. The second-stage objective for learning $F_{AD}$ is:", + "bbox": [ + 89, + 497, + 483, + 529 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {L D M}} = \\mathbb {E} _ {\\boldsymbol {x} _ {r} ^ {f}, \\boldsymbol {y} _ {r} ^ {f}, \\boldsymbol {x} _ {e ^ {\\prime}} ^ {f}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\boldsymbol {x} _ {r, t} ^ {f}, \\boldsymbol {y} ^ {f}, \\boldsymbol {x} _ {e ^ {\\prime}} ^ {f}, t\\right) \\right\\| _ {2} ^ {2} \\right], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 539, + 483, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the encoder of $\\epsilon_{\\theta}$ is modified as defined in Eq. (6).", + "bbox": [ + 89, + 582, + 467, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. From Single Agent to Collaborative Datasets", + "text_level": 1, + "bbox": [ + 89, + 606, + 468, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "There are many publicly accessible large-scale real-world autonomous driving datasets for ego-centric 3D perception, like WOD [39], NuScenes [2], KITTI-360 [21], and PandaSet [46], whereas much fewer datasets exist for CAV due to the challenges in data collection as discussed in Sec. 1.", + "bbox": [ + 89, + 628, + 483, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This limitation inspired a bold idea: Can we transfer the generation capability learned from paired CAV data to these existing labeled single-agent datasets? Achieving this would scale up CAV datasets substantially, offering immense benefits to the research community. However, this goal is challenging due to domain gaps between datasets, such as variations in LiDAR sensors, point cloud densities, data patterns, and data collection environments. Below, we describe our approach to addressing these challenges.", + "bbox": [ + 89, + 704, + 483, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Problem setup. Our target domain is a single-agent dataset, while the source domain consists of paired perception and semantic information from ego and reference perspectives, denoted as $(\\pmb{x}_r,\\pmb{x}_e,\\pmb{y}_r)$ . We demonstrate that we can transfer", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "knowledge from a simulated CAV dataset (e.g., OPV2V [51]) to a real-world ego-centric dataset (e.g., WOD [39]).", + "bbox": [ + 511, + 90, + 906, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Two-Stage training is the key. As described in Sec. 3.3, we decompose the generative training into two stages: first learning $P(\\pmb{x}|\\pmb{y})$ , followed by $P(\\pmb{x}_r|\\pmb{x}_e,\\pmb{y}_r)$ . The second stage keeps $P(\\pmb{x}|\\pmb{y})$ frozen by adding an adapter. This strategy enables us to use target domain data in the first stage. The resulting $P(\\pmb{x}|\\pmb{y})$ would generate target-like point clouds.", + "bbox": [ + 511, + 121, + 906, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Minimizing domain gaps. In the first stage, we have learned VQ-VAE encoder-decoders and $P(\\boldsymbol{x}|\\boldsymbol{y})$ in the target domain (single-agent). In the second stage, we bring in CAV data to guide the generation on how to condition on ego-agent's perception. To minimize the gap of $\\boldsymbol{x}^f$ between two domains— $\\boldsymbol{x}^f$ is what the generative model is optimized to generate—we adopt a GAN-style discriminator [8, 10, 22, 40]. Specifically, we aim to make feature embeddings extracted from the source data and the target data indistinguishable; namely, to confuse a discriminator $D$ trained to differentiate them. We freeze all parameters of the target domain's encoder $E_{tgt}$ , codebook $Q_{tgt}$ , and decoder $G_{tgt}$ in the VQ-VAE. For the source domain, we initialize its $E_{src}$ and $G_{src}$ by $E_{tgt}$ and $G_{tgt}$ while reusing the frozen codebook $Q_{tgt}$ . We then learn the discriminator $D$ and the source domain's encoder $E_{src}$ and decoder $G_{src}$ in an interleaving fashion by minimizing the following objective functions:", + "bbox": [ + 511, + 212, + 908, + 468 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{r l} \\text {F o r} D: \\mathcal {L} _ {D} = & \\mathbb {E} _ {\\boldsymbol {x} \\sim P _ {t g t}} [ \\max (0, 1 - D (E _ {t g t} (\\boldsymbol {x}))) ] \\\\ & + \\mathbb {E} _ {\\boldsymbol {x} \\sim P _ {s r c}} [ \\max (0, 1 + D (E _ {s r c} (\\boldsymbol {x}))) ]; \\end{array} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 478, + 905, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {F o r V Q - V A E :} \\mathcal {L} _ {\\mathrm {v q}} = \\mathcal {L} _ {\\mathrm {r e c}} + \\mathcal {L} _ {s r c} + \\| \\operatorname {s g} [ \\boldsymbol {x} ^ {z} ] - E _ {s r c} (\\boldsymbol {x}) \\| _ {2} ^ {2}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 525, + 905, + 566 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} \\mathcal {L} _ {s r c} = - \\mathbb {E} _ {\\boldsymbol {x} \\sim P _ {s r c}} D \\left(E _ {s r c} (\\boldsymbol {x})\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 571, + 905, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The other terms in Eq. (9) follow those in Eq. (1), where for $\\pmb{x} \\sim P_{src}$ , $\\hat{\\pmb{x}} = G_{src} \\circ Q_{tgt} \\circ E_{src}(\\pmb{x})$ in $\\mathcal{L}_{rec}$ . We note that $\\mathcal{L}_D$ takes in two streams of data (i.e., source and target) and is defined by the Hinge Loss [22] on $\\pmb{x}^f$ ; $\\mathcal{L}_{\\mathrm{vq}}$ takes in only the source CAV data. We also note that embeddings for object locations does not require adaptation.", + "bbox": [ + 511, + 613, + 905, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After this adaptation step, we proceed into the second stage of generative training, where we drop the discriminator and train the control module defined in Eq. (5) by minimizing Eq. (7). The feature embeddings are produced by $E_{src}$ on top of $(\\pmb{x}_r, \\pmb{x}_e, \\pmb{y}_r) \\sim P_{src}$ .", + "bbox": [ + 511, + 704, + 905, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Enhancement in target domains. With the discriminator and adversarial training, we reduce the gap between source and target domain data, successfully adapting TYP to the single-agent dataset, as shown in Sec. 4.5. To further improve the conditioned generation quality in the target domain, a further fine-tuning stage in it is desired. However, how can we do so without having reference agents' point clouds in the target domain? Here, we propose two simple solutions.", + "bbox": [ + 511, + 780, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12031", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/32b1667105da42b55656c8c0ec5bcd85263a31a5087320a91863b01d07e62b44.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTrain DataAP 0.5
smlall
No Fusionego's gt only0.670.410.130.40
Early Fusion [4]gt (oracle)0.760.420.310.49
TYP (ours)0.750.380.290.46
Late Fusion [51]gt (oracle)0.740.570.360.55
TYP (ours)0.710.490.320.50
AttFuse [51]gt (oracle)0.940.800.620.77
TYP (ours)0.900.730.560.72
V2X-ViT [50]gt (oracle)0.870.710.500.71
TYP (ours)0.840.650.400.65
", + "bbox": [ + 114, + 88, + 459, + 244 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/259fc6e51134ec81f260d71466fd8d7df89b8d55a1d9c6dca094290ed2cb6246.jpg", + "image_caption": [ + "(a) no adaptation", + "Figure 3. Qualitative results on enhancement in the target domain. Generated point cloud (green) has better quality with the enhancement given ego (gray) from Waymo (cf. Sec. 3.4)." + ], + "image_footnote": [], + "bbox": [ + 99, + 292, + 225, + 340 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/51fdfe250768ef030b1013c443464af5359bfc7cbf9bea5b34c62fac5d788731.jpg", + "image_caption": [ + "(b) disc." + ], + "image_footnote": [], + "bbox": [ + 225, + 292, + 346, + 340 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e3b62f81cc90d270c5388e87d359101ffd43b7f09e619da9d8210b2dd9fd2f5f.jpg", + "image_caption": [ + "(c) disc. + enhancement" + ], + "image_footnote": [], + "bbox": [ + 346, + 292, + 470, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Recall that in training $P(\\pmb{x}_r|\\pmb{x}_e,\\pmb{y}_r)$ using the source CAV data, point clouds show up in both the input as conditions and the outputs as supervisions. In other words, the domain gap exists on both sides, and we aim to reduce it by further fine-tuning using the target single-agent dataset. First, to enhance the output generation in the target domain, we set ego's point clouds to empty, i.e., $\\pmb{x}_e = 0^{H\\times W\\times C}$ for $(\\pmb{x}_r,\\pmb{y}_r)\\sim P_{tgt}$ . Then, to enable adaptation on inputs, we apply self-training [18, 28, 60, 61], randomly sampling a reference location in the scene to generate pseudo-point clouds for fine-tuning, i.e., $(\\hat{\\pmb{x}}_r,\\pmb{x}_e,\\pmb{y}_r)$ for $(\\pmb{x}_e,\\pmb{y}_r)\\sim P_{tgt}$ . We use the above target domain data and the source domain data to jointly fine-tune the control module (cf. Sec. 3.3, Eq. (6), and Eq. (7)) further. Empirical results in Sec. 4.5 show the effectiveness of the proposed solution in turning single-agent datasets into collaborative datasets for CAV development.", + "bbox": [ + 88, + 402, + 483, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 657, + 223, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 89, + 681, + 171, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. In this study, we primarily use three datasets: OPV2V [51], V2V4Real [52], and WOD [39]. OPV2V is a simulation-based CAV dataset that includes over 70 diverse driving scenes and more than 11,000 frames. We utilize this dataset initially to validate our proposed research problem and later to guide the generation training on single-agent datasets. V2V4Real is a real-world CAV dataset with approximately 20,000 LiDAR frames, which we use to validate TYP in a real-world CAV context, treating it as a single-agent dataset (i.e., without using any labels to train TYP). Lastly, we adopt WOD to scale up TYP, one of the largest real-world single-agent datasets available.", + "bbox": [ + 89, + 704, + 485, + 885 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. Following standard benchmarks [51,", + "bbox": [ + 89, + 885, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f560e44fcac80bb2aa9f7d4fbb29b08b37623adda285365658a4a3bdfd4c9828.jpg", + "table_caption": [ + "Table 1. Results on OPV2V. We compare the performance using reference generated by TYP versus ground-truth (oracle). The comparable results demonstrate the quality of our generated data." + ], + "table_footnote": [], + "table_body": "
MethodTrain DataAP 0.5
smlall
No Fusionego's gt only0.710.330.080.46
Early Fusion [4]gt (oracle)0.790.420.450.60
TYP (ours)0.760.360.300.53
Late Fusion [51]gt (oracle)0.790.530.560.67
TYP (ours)0.770.480.510.63
AttFuse [51]gt (oracle)0.800.450.360.61
TYP (ours)0.750.450.260.56
V2X-ViT [50]gt (oracle)0.810.490.300.61
TYP (ours)0.760.380.170.48
", + "bbox": [ + 537, + 88, + 882, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Results on V2V4Real. We compare the performance using reference generated by TYP versus ground-truth (oracle). Note that we do not use V2V4Real data to train the generation, treating it as a single-agent dataset. The comparable results demonstrate the quality of our generated data and the transferability of TYP.", + "bbox": [ + 511, + 250, + 905, + 319 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "52], we report Average Precision (AP 0.5) across three distance ranges: 0-30 meters (short, denoted as s), 30-50 meters (medium, m), and 50-100 meters (long, l) and overall (all). Implementations. For TYP training, we voxelize point clouds within the range $[-51.2, 51.2]$ meters in x and y dimensions and 4 meters in z (range is dataset-dependent) to create 3D volumes, $x \\in \\mathbb{R}^{512 \\times 512 \\times 20}$ . This is then encoded to $x^f \\in \\mathbb{R}^{64 \\times 64 \\times 8}$ with VQ-VAE [41]. Similarly, semantic information $y \\in \\mathbb{R}^{512 \\times 512 \\times 1}$ is encoded to $y^f \\in \\mathbb{R}^{64 \\times 64 \\times 3}$ . We train VQ-VAE for $x$ 120 epochs and for $y$ 10 epochs, with a batch size of 6 per GPU. Next, we train LDM with a batch size of 48 per GPU for 200 epochs in the first stage and 80 epochs in the second stage. During the second stage, we freeze the U-Net from the first stage and fine-tune only the control module, following [26]. For adaptation, we initialize the weights of encoder-decoder for CAV data (fake) with the model trained on single-agent data (real), freeze its codebook, and fine-tune it with a discriminator comprising 4 convolutional layers. The model processes an equal amount of fake and real per batch (i.e., 3 samples each). All training is conducted on 4 NVIDIA H100 GPUs. For experiments on benchmarks, we follow the convention to report results with Pointpillars [17]. We train / fine-tune all models with 8 P100, with batch size 4 per GPU. For V2X-ViT, we use 4 H100 due to resources required.", + "bbox": [ + 511, + 325, + 906, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference of generation and post-processing. We use DDIM scheduler [38] for 200 steps. For post-processing, we apply modified Gumbel-Softmax [15] to the logits. Specifically, we define two thresholds, low and high. Scores below the low threshold are dropped, those above the high threshold are retained, and values in between are perturbed with Gaussian noise on logits. Hyperparameters are tuned on a small hold-out set. Finally, we convert the volume back to coordinates by using the centers of the occupied voxels.", + "bbox": [ + 511, + 704, + 908, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Do We Need Collaborative Perception?", + "text_level": 1, + "bbox": [ + 511, + 848, + 849, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To explore this question, we consistently report the performance without collaborative agents (denoted as \"ego's gt", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12032", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9845242fc67e6b238444d31c7165ce1e13b8fab0130a1f033657e5b06422230c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPre-TrainFT. on 0 SceneFT. on 5 SceneFT. on 10 SceneFT. on 32 Scene
smlallsmlallsmlallsmlall
No Fusionego's gt only0.440.160.060.290.660.290.080.420.710.340.080.47
scratch0.440.180.160.300.730.280.240.480.790.420.450.60
Early Fusion [4]OPV2V0.540.200.070.310.320.180.150.250.730.320.310.500.800.470.540.65
ColWaymo (ours)0.500.240.240.350.680.340.320.510.780.390.330.570.830.480.500.65
Late Fusion [51]scratch0.440.240.330.340.720.420.510.580.790.530.560.67
OPV2V0.600.270.280.440.550.260.380.420.730.420.510.590.780.550.580.67
ColWaymo (ours)0.400.190.150.250.600.280.440.470.760.430.510.610.820.580.610.71
AttFuse [51]scratch0.400.150.100.280.700.300.170.470.800.450.360.61
OPV2V0.510.190.050.310.540.220.110.370.770.400.210.540.830.530.400.65
ColWaymo (ours)0.660.350.110.450.650.290.160.460.830.460.330.610.880.580.530.72
V2X-ViT [50]scratch0.430.150.120.310.700.280.170.430.810.490.300.61
OPV2V0.510.240.070.330.480.230.160.350.760.380.220.530.810.490.350.61
ColWaymo (ours)0.600.280.100.340.660.280.220.460.790.480.260.580.840.570.440.67
", + "bbox": [ + 135, + 88, + 859, + 292 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1c310ef162f2f512a5b6fea873dd29a660f70a47194ff47f8b8105a300ed3baa.jpg", + "table_caption": [ + "Table 3. Results with pre-training on collaborative Waymo. We scale up TYP with exiting large-scale single-agent dataset WOD [39], creating its collaborative version \"ColWaymo\". These data are used for pre-training and subsequently fine-tuned on V2V4Real [52]. Results show significant improvement over training from scratch, highlighting the potential to reduce data collection efforts, scale up, and accelerate CAV development. Comparisons to simulated OPV2V [51] pre-training further demonstrate the realism of our generated point clouds." + ], + "table_footnote": [], + "table_body": "
MethodDisc.ST.Dum.JSD (↓)MMD (↓)
No Adaptation0.264.61 × 10-4
✓0.161.10 × 10-4
Adaptation✓✓0.171.30 × 10-4
✓✓✓0.161.17 × 10-4
", + "bbox": [ + 106, + 357, + 467, + 441 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/487d6dc4883983568a93ef2f51e7ed52c51f0ed3d9d2f7f7d2e24c24367ae6b6.jpg", + "table_caption": [ + "Table 4. Ablation of proposed adaption method. Results with adaptation outperform no adaptation, demonstrating an improvement in generation quality. Abbreviations in title refer to discriminator, self-training, and dummy ego, respectively (cf. Sec. 3.4)." + ], + "table_footnote": [], + "table_body": "
Num. StepsEpochs 1stEpochs 2ndJSD (↓)MMD (↓)
1-2800.116.45 × 10-5
2200800.105.28 × 10-5
", + "bbox": [ + 106, + 503, + 467, + 553 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. 2-Stage training. Results demonstrate the benefit of the proposed 2-stage training in generation quality on OPV2V [51].", + "bbox": [ + 89, + 556, + 482, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "only\" in Tab. 1, Tab. 2, and Tab. 3). For example, in the V2V4Real dataset, a well-trained detector struggles with objects beyond 50 meters, achieving only 0.08 AP at that range, though it performs well on nearby objects with an AP of 0.71 (cf. first row of Tab. 3). simply combining the predictions from two under-trained detectors, we observe a notable boost in performance to 0.33 (cf. \"scratch\" in late fusion, Tab. 2), suggesting the significance of CAV.", + "bbox": [ + 89, + 595, + 482, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Point Clouds Generation on OPV2V", + "text_level": 1, + "bbox": [ + 89, + 726, + 405, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Setting. To validate our research concept, we begin with the simulated dataset OPV2V [51]. We split the original training set of 44 scenes into two halves, using only the first 22 scenes to train our generation model. For the remaining 22 scenes, we generate point clouds for reference agents based on the ego agent's point clouds and compare these results to ground-truth point clouds. We follow the benchmark to report performance on the validation set from Culver City.", + "bbox": [ + 89, + 750, + 482, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. Following the literature [51, 52], we adopt three common fusion algorithms for CAV: early, late, and inter", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mediate fusion, which fuse input point clouds, extracted features, and predicted bounding boxes, respectively. For intermediate fusion, we adipot AttFuse [51] and V2X-VIT [50].", + "bbox": [ + 511, + 359, + 906, + 405 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison to using ground-truth LiDAR. In Tab. 1, we show that training CAV with point clouds generated by TYP achieves performance comparable to that obtained with ground-truth (oracle) across all methods (e.g., 0.46 vs. 0.49 with early, 0.50 vs. 0.55 with late, and 0.77 vs. 0.72 with AttFuse [51], etc.) This result highlights the quality of the generated point clouds. In the supplemental material, we further demonstrate that with additional labeled data (from the first half of the split used for training generation), the gap to oracle performance can be further minimized.", + "bbox": [ + 511, + 406, + 908, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Single-Agent to Collaborative Multi-Agent", + "text_level": 1, + "bbox": [ + 511, + 569, + 877, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Setting. Building on our findings with simulated data (cf. Sec. 4.3), we validate our approach on the real-world dataset V2V4Real [52], treating it as a single-agent dataset without access to reference agents' point clouds. Specifically, we use the model from Sec. 4.3 to generate reference car's point clouds from the ego's perspective and compare performance against ground-truth point clouds (oracle). Note that no V2V4Real data is used to train the generation.", + "bbox": [ + 511, + 592, + 908, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on V2V4Real. In Tab. 2, we show the performance using generated reference point clouds is comparable to the oracle, supporting our idea of translating a single-agent dataset to multi-agent CAV. Xu et al. previously explored OPV2V-to-V2V4Real adaptation, noting significant performance drops even with adaptation algorithms. Here, TYP offers a new solution to this challenge – in a generative way, providing a fresh perspective on domain adaptation.", + "bbox": [ + 511, + 714, + 908, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. ColWaymo: Scaling with Large-Scale Data", + "text_level": 1, + "bbox": [ + 511, + 847, + 877, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scaling CAV to the next level. In Sec. 4.4, we show the potential to create realistic reference's point clouds on real-", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12033", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/56023c3d1150e71dbc23b6a087a38fddfb7fd066bd77fbb91652e7a64981f6d8.jpg", + "image_caption": [ + "Figure 4. Qualitative results on Collaborative Waymo. The gray point clouds are from the original single-agent dataset and the green are generated by TYP conditioning on them." + ], + "image_footnote": [], + "bbox": [ + 143, + 88, + 320, + 223 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/752f1bd8c563280c053b82f26f676cbea0dd1f380c2b393c2bcd1f9f8cecf506.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 88, + 496, + 223 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1b05b0743a4f04bdbd509836ecb1bdee5403a67e91a68fbc2f14b5c2c97f81a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 88, + 671, + 223 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/03dfc23b41211493196497bebcb37145a7ac996d23cf359325eacf6e54c73350.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 88, + 851, + 223 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4abe5b5c88316aacf10b2ee4382db43d2c14c062fb5f9619bde0bf0fee779bc8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 138, + 257, + 375, + 358 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/faff2f785d98135f357ec24971c54b3770e890869d48333deb10e417e9116c2e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 377, + 257, + 614, + 357 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/65428d9475507f7325174b44ae1fc7cb81bec176637a2fc906af87f5a3d0feef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 258, + 856, + 357 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/061b00aca9e01923d1aa59b2808ff613fb4d4e254bdd9a12fc66bcc777bbdfdd.jpg", + "image_caption": [ + "Figure 5. Qualitative results. Our proposed TYP is capable of scene editing, by inputting the same point cloud but different object locations. We (b) remove and (c) add a car from (a) the original point cloud. Inspired by the idea of past traversals [56], we apply completely different traffic conditions and generate (d), (e), and (f), to imagine driving through the same intersection." + ], + "image_footnote": [], + "bbox": [ + 138, + 364, + 375, + 465 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ee06003fcf8d90f7ea437d7a4a6f313fb2f5c5a67ed47b563e6e0b56a2a2623d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 377, + 364, + 614, + 464 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b17e2e0fafe06efd15d5aa42ba408d318b3b5301ad4daa26d3932eb5d590cde6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 366, + 856, + 464 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "world dataset V2V4Real [52], through a simulation-guided training of TYP. Here, our goal is broader: to transform an existing large-scale single-agent dataset (e.g., WOD [39]) into a collaborative version. Using the pipeline described in Sec. 3 and illustrated in Fig. 2, we train the generative model on 17,400 samples from the WOD training set. We then use its validation set to generate approximately 36,000 collaborative samples by sampling reference locations from labeled vehicles. These generated samples serve to pre-train a model, which we subsequently fine-tune on V2V4Real [52]. For comparison, we also conduct pre-training on OPV2V [51].", + "bbox": [ + 91, + 522, + 485, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results of fine-tuning on V2V4Real. Following prior work [1, 29, 47, 54], we perform pre-training followed by fine-tuning on limited labeled data to validate the potential of TYP for scaling up CAV development. As shown in Tab. 3, pre-training on our generated data yields significantly better performance than training from scratch. Additionally, compared to pre-training on the simulated dataset OPV2V, ColWaymo consistently achieves higher performance, highlighting the realism of the generated point clouds. These results demonstrate that TYP effectively transforms an ego-only dataset into a collaborative version, and potentially reducing the need for extensive CAV data collection.", + "bbox": [ + 89, + 688, + 485, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generation quality by proposed adaptation. To assess the quality of generated data, we use two common metrics:", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Jensen-Shannon Divergence (JSD) and Maximum Mean Discrepancy (MMD). As shown in Tab. 4, our adaptation method improves JSD from 0.26 to 0.16 and MMD from 4.61 to 1.17 compared to the baseline without adaptation. Additionally, Fig. 3 visually illustrates this improvement.", + "bbox": [ + 511, + 522, + 908, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 607, + 694, + 622 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We illustrate the proposed problem in Fig. 1, which highlights the benefits of collaboration by allowing an agent to \"see\" beyond the limitations of its own sensors. TYP generates realistic point clouds from various viewpoints, enabling the agents to perceive previously occluded objects while preserving the overall scene semantics. Fig. 4 shows our ColWaymo, and Fig. 5 further demonstrates the model's capability for scene editing.", + "bbox": [ + 509, + 628, + 908, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 763, + 633, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce a new research direction in collaborative driving and present the first solution. Empirical results show our approach can significantly reduce data collection and development efforts, advancing safer autonomous systems.", + "bbox": [ + 511, + 779, + 908, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future work. Following existing benchmarks [51, 52], TYP focuses on vehicle-like objects. Future work could extend it to broader objects and static entities (e.g., traffic signs, signals) essential for real-world traffic.", + "bbox": [ + 511, + 840, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12034", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research is supported in part by grants from the National Science Foundation (IIS-2107077, IIS-2107161). We are thankful for the generous support of the computational resources by the Ohio Supercomputer Center.", + "bbox": [ + 89, + 114, + 485, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 190, + 186, + 205 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alexandre Boulch, Corentin Sautier, Björn Michele, Gilles Puy, and Renaud Marlet. Also: Automotive lidar self-supervision by occupancy estimation. In CVPR, 2023. 8", + "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 4, 5", + "[3] Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. In Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, 2019. 1", + "[4] Qi Chen, Sihai Tang, Qing Yang, and Song Fu. Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds. In International Conference on Distributed Computing Systems (ICDCS). IEEE, 2019. 1, 6, 7", + "[5] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In CVPR, 2017. 4", + "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 4", + "[7] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. In CoRL, 2017. 2, 3", + "[8] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks. Journal of machine learning research, 17(59):1-35, 2016. 5", + "[9] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 4", + "[10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 3, 5", + "[11] Cole Gulino, Justin Fu, Wenjie Luo, George Tucker, Eli Bronstein, Yiren Lu, Jean Harb, Xinlei Pan, Yan Wang, Xiangyu Chen, et al. Waymax: An accelerated, data-driven simulator for large-scale autonomous driving research. In NeurIPS, 2024. 4", + "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3, 4", + "[13] Qianjiang Hu, Zhimin Zhang, and Wei Hu. Rangeldm: Fast realistic lidar point cloud generation. In ECCV, 2025. 2, 3", + "[14] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. In NeurIPS, 2022. 1" + ], + "bbox": [ + 93, + 215, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In ICLR, 2017. 6", + "[16] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. In NeurIPS, 2021. 4", + "[17] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, 2019. 4, 6", + "[18] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, 2013. 6", + "[19] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. In NeurlPS, 2021. 1", + "[20] Yiming Li, Dekun Ma, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: Multi-agent collaborative perception dataset and benchmark for autonomous driving. In Robotics and Automation Letters, pages 10914-10921. IEEE, 2022. 3", + "[21] Yiyi Liao, Jun Xie, and Andreas Geiger. Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE TPAMI, 45(3):3292-3310, 2022. 5", + "[22] Jae Hyun Lim and Jong Chul Ye. Geometric gan. arXiv preprint arXiv:1705.02894, 2017. 5", + "[23] Yen-Cheng Liu, Junjiao Tian, Nathaniel Glaser, and Zsolt Kira. When2com: Multi-agent perception via communication graph grouping. In CVPR, 2020. 1", + "[24] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In IROS, 2015. 4", + "[25] Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. Rangenet++: Fast and accurate lidar semantic segmentation. In IROS, 2019. 4", + "[26] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, 2024. 5, 6", + "[27] Kazuto Nakashima and Ryo Kurazume. Lidar data synthesis with denoising diffusion probabilistic models. In ICRA, 2024. 2, 3", + "[28] Tai-Yu Pan, Qing Liu, Wei-Lun Chao, and Brian Price. Towards open-world segmentation of parts. In CVPR, 2023. 6", + "[29] Tai-Yu Pan, Chenyang Ma, Tianle Chen, Cheng Perng Phoo, Katie Z Luo, Yurong You, Mark Campbell, Kilian Q Weinberger, Bharath Hariharan, and Wei-Lun Chao. Pre-training lidar-based 3d object detectors through colorization. In ICLR, 2023. 8", + "[30] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 4", + "[31] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 4", + "[32] Haoxi Ran, Vitor Guizilini, and Yue Wang. Towards realistic scene generation with lidar diffusion models. In CVPR, 2024. 2, 3" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12035", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Zaydoun Yahya Rawashdeh and Zheng Wang. Collaborative automated driving: A machine learning-based method to enhance the accuracy of shared information. In International Conference on Intelligent Transportation Systems (ITSC). IEEE, 2018. 1", + "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 4", + "[35] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 5", + "[36] T-YLPG Ross and GKHP Dólár. Focal loss for dense object detection. In CVPR, 2017. 4", + "[37] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointr-cnn: 3d object proposal generation and detection from point cloud. In CVPR, 2019. 4", + "[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 6", + "[39] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 2, 3, 4, 5, 6, 7, 8", + "[40] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In CVPR, 2017. 2, 5", + "[41] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. In NeurIPS, 2017. 4, 6", + "[42] Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In ECCV, 2020. 1", + "[43] Bichen Wu, Alvin Wan, Xiangyu Yue, and Kurt Keutzer. Squeezeseg: Convolutional neural nets with recurrent crf for real-time road-object segmentation from 3d lidar point cloud. In ICRA, 2018. 4", + "[44] Yang Wu, Kaihua Zhang, Jianjun Qian, Jin Xie, and Jian Yang. Text2lidar: Text-guided lidar point cloud generation via equirectangular transformer. In ECCV. Springer, 2024. 2, 3", + "[45] Hao Xiang, Zhaoliang Zheng, Xin Xia, Runsheng Xu, Letian Gao, Zewei Zhou, Xu Han, Xinkai Ji, Mingxi Li, Zonglin Meng, et al. V2x-real: a largs-scale dataset for vehicle-to-everything cooperative perception. arXiv preprint arXiv:2403.16034, 2024. 1", + "[46] Pengchuan Xiao, Zhenlei Shao, Steven Hao, Zishuo Zhang, Xiaolin Chai, Judy Jiao, Zesong Li, Jian Wu, Kai Sun, Kun Jiang, et al. Pandaset: Advanced sensor suite dataset for autonomous driving. In International Intelligent Transportation Systems Conference (ITSC). IEEE, 2021. 5", + "[47] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In ECCV, 2020. 8", + "[48] Yuwen Xiong, Wei-Chiu Ma, Jingkang Wang, and Raquel Urtasun. Learning compact representations for lidar completion and generation. In CVPR, 2023. 3, 4" + ], + "bbox": [ + 91, + 92, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Yuwen Xiong, Wei-Chiu Ma, Jingkang Wang, and Raquel Urtasun. Ultralidar: Learning compact representations for lidar completion and generation. In CVPR, 2023. 2, 4", + "[50] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In ECCV, 2022. 1, 6, 7", + "[51] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In ICRA, 2022. 2, 3, 5, 6, 7, 8", + "[52] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In CVPR, 2023. 1, 3, 6, 7, 8", + "[53] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 4", + "[54] Junbo Yin, Dingfu Zhou, Liangjun Zhang, Jin Fang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Proposal-contrast: Unsupervised pre-training for lidar-based 3d object detection. In ECCV, 2022.8", + "[55] Jinsu Yoo, Zhenyang Feng, Tai-Yu Pan, Yihong Sun, Cheng Perng Phoo, Xiangyu Chen, Mark Campbell, Kilian Q Weinberger, Bharath Hariharan, and Wei-Lun Chao. Learning 3d perception from others' predictions. In ICLR, 2025. 1", + "[56] Yurong You, Katie Z Luo, Xiangyu Chen, Junan Chen, WeiLun Chao, Wen Sun, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Hindsight is 20/20: Leveraging past traversals to aid 3d perception. In ICLR, 2022. 8", + "[57] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, and Zaiqing Nie. Dair-v2x: A large-scale dataset for vehicleinfrastructure cooperative 3d object detection. In CVPR, 2022. 3", + "[58] Chris Zhang, Wenjie Luo, and Raquel Urtasun. Efficient convolutions for real-time semantic segmentation of 3d point clouds. In 3DV, 2018. 4", + "[59] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In CVPR, 2018. 4", + "[60] Barret Zoph, Golnaz Ghiasi, Tsung-Yi Lin, Yin Cui, Hanxiao Liu, Ekin Dogus Cubuk, and Quoc Le. Rethinking pretraining and self-training. In NeurIPS, 2020. 6", + "[61] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In ECCV, 2018. 6", + "[62] Vlas Zyrianov, Xiyue Zhu, and Shenlong Wang. Learning to generate realistic lidar point clouds. In ECCV, 2022. 2, 3", + "[63] Vlas Zyrianov, Henry Che, Zhijian Liu, and Shenlong Wang. Lidardm: Generative lidar simulation in a generated world. arXiv preprint arXiv:2404.02903, 2024. 2, 3" + ], + "bbox": [ + 516, + 92, + 906, + 829 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12036", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Transfer Your Perspective_ Controllable 3D Generation from Any Viewpoint in a Driving Scene/0e711492-3ab2-4748-8147-c897ce7d76b8_model.json b/2025/Transfer Your Perspective_ Controllable 3D Generation from Any Viewpoint in a Driving Scene/0e711492-3ab2-4748-8147-c897ce7d76b8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..54884168175343e07e5158587345ffe39914ecf0 --- /dev/null +++ b/2025/Transfer Your Perspective_ Controllable 3D Generation from Any Viewpoint in a Driving Scene/0e711492-3ab2-4748-8147-c897ce7d76b8_model.json @@ -0,0 +1,2541 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.157, + 0.131, + 0.842, + 0.177 + ], + "angle": 0, + "content": "Transfer Your Perspective: Controllable 3D Generation from Any Viewpoint in a Driving Scene" + }, + { + "type": "text", + "bbox": [ + 0.19, + 0.203, + 0.805, + 0.241 + ], + "angle": 0, + "content": "Tai-Yu Pan1, Sooyoung Jeon1, Mengdi Fan1, Jinsu Yoo1, Zhenyang Feng1, Mark Campbell2, Kilian Q. Weinberger2, Bharath Hariharan2, Wei-Lun Chao1" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.25, + 0.686, + 0.269 + ], + "angle": 0, + "content": "1The Ohio State University, 2Cornell University" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.294, + 0.296, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.401, + 0.273, + 0.415 + ], + "angle": 0, + "content": "(a) Ego-centric driving" + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.294, + 0.498, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.401, + 0.488, + 0.415 + ], + "angle": 0, + "content": "(b) TYP for dynamic agent" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.294, + 0.7, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.401, + 0.68, + 0.415 + ], + "angle": 0, + "content": "(c) TYP for static agent" + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.294, + 0.905, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.401, + 0.886, + 0.415 + ], + "angle": 0, + "content": "(d) Collaborative driving" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.419, + 0.908, + 0.477 + ], + "angle": 0, + "content": "Figure 1. Illustration of the proposed problem and solution, Transfer Your Perspective (TYP). (a) A given sensory data captured by the ego-car (red triangle). (b) A generated sensory data by TYP, seeing from the viewpoint of another vehicle (green triangle) in the same scene. (c) A generated sensory data, seeing from an imaginary static agent like roadside units (blue icon). (d) Putting all the sensory data together, given or generated, TYP enables the development of collaborative perception with little or no real collaborative driving data." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.495, + 0.327, + 0.51 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.537, + 0.487, + 0.886 + ], + "angle": 0, + "content": "Self-driving cars relying solely on ego-centric perception face limitations in sensing, often failing to detect occluded, faraway objects. Collaborative autonomous driving (CAV) seems like a promising direction, but collecting data for development is non-trivial. It requires placing multiple sensor-equipped agents in a real-world driving scene, simultaneously! As such, existing datasets are limited in locations and agents. We introduce a novel surrogate to the rescue, which is to generate realistic perception from different viewpoints in a driving scene, conditioned on a real-world sample—the ego-car's sensory data. This surrogate has huge potential: it could potentially turn any ego-car dataset into a collaborative driving one to scale up the development of CAV. We present the very first solution, using a combination of simulated collaborative data and real ego-car data. Our method Transfer Your Perspective (TYP) learns a conditioned diffusion model whose output samples are not only realistic but also consistent in both semantics and layouts with the given ego-car data. Empirical results demonstrate TYP's effectiveness in aiding in a CAV setting. In particular, TYP enables us to (pre-)train collaborative perception algorithms like early and late fusion with little or no real-world collaborative data, greatly facilitating downstream CAV applications." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.495, + 0.645, + 0.51 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.909, + 0.689 + ], + "angle": 0, + "content": "Seeing the world from an ego-centric perspective, a self-driving car risks being \"narrow-sighted,\" limiting its ability to respond appropriately in dynamic driving environments. For instance, it should slow down or honk if a pedestrian is about to cross the road or if another vehicle is merging into its lane. However, these actions depend on the car's ability to detect traffic participants, which may be occluded by a large bus or a building at a sharp intersection. This occlusion problem can hardly be addressed by mounting a few more sensors on the car. We argue that a self-driving car must move beyond its ego-centric perspective." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.691, + 0.911, + 0.903 + ], + "angle": 0, + "content": "One intuitive way is to collaborate with nearby \"agents\" that observe the scene from different angles, such as other sensor-equipped cars or static devices like roadside units (RSUs). When trained together to address GPS errors and synchronization delays [55], collaborative perception has been shown to significantly improve the accuracy of each agent's perception, particularly in detecting occluded or distant objects [3, 4, 14, 19, 23, 33, 42, 45, 50-52]. However, collecting training data for collaborative perception is never easy. Unlike single-agent data, which can be collected by simply driving a car on the road, collaborative data requires the simultaneous presence of multiple agents in the same driving scene. For dynamic agents like vehicles, precise coordination is needed to ensure they are within communication" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12027" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "range. These challenges limit existing works in scale and the number of agents (typically just two). While one may leverage simulated data by game engines [7], they often fail to capture the diversity of real-world scenes. We thus ask:" + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.157, + 0.472, + 0.202 + ], + "angle": 0, + "content": "Can we obtain realistic data for learning collaborative perception with much less effort, ideally as easily as single-agent data?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.206, + 0.484, + 0.281 + ], + "angle": 0, + "content": "Specifically, given the vast amount of single-agent LiDAR data already collected across various driving environments, is it possible to convert each of them into collaborative perception data by generating additional point clouds from alternative reference viewpoints within the same scene?" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.282, + 0.484, + 0.568 + ], + "angle": 0, + "content": "At first glance, this question may seem overly ambitious for three reasons. First, for collaborative perception to be effective, the generated point cloud must provide information that the ego-car's point cloud cannot capture, such as occluded surfaces invisible to the ego-car. This creates a chicken-and-egg problem: if the ego-car's point cloud is all we initially know about the scene, how can we deduce any additional information? Second, for training purposes, the generated data must be realistic. It should replicate the constraints and data patterns of a real sensor as if it were positioned at the reference viewpoint, producing no points in occluded areas and fewer points in distant areas. Last but not least, in regions commonly perceivable from both the reference and ego-car's viewpoints, the generated data should align in layouts and semantics with the ego-car's data. On the surface, this may seem as simple as copying the ego-car's data, but doing so would violate the second requirement. In essence, the generated data should resemble a point cloud seen from the reference viewpoint, not the ego-car." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.87 + ], + "angle": 0, + "content": "That said, after a deeper look at the question, we argue that it is achievable with three key insights. First, given semantic information around the ego-car for conditioning, existing research has shown promising progress in generating realistic point clouds [13, 27, 32, 44, 49, 62, 63]. This is especially encouraging, as most existing single-agent datasets provide 3D object labels, and they can be translated to obtain semantic information centered around the reference position. This further implies that we can generate occluded object surfaces as seen from the ego-car if they are visible from the reference viewpoint. Second, semantic information is easily editable: even if the translated object map has noticeable empty areas due to limited ego-car perception, object boxes can be manually added to make the map appear more realistic from the reference viewpoint. Lastly, in commonly perceivable regions where the generated data needs to meet two physical constraints, we could leverage simulators based on computational graphics and optics. Specifically, we can use simulated data from the two views to train a conditioned generative model that maps one viewpoint to the other." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Building upon these insights, we propose Transfer Your Perspective (TYP), a novel research problem and the very" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.41 + ], + "angle": 0, + "content": "first solution for generating a realistic point cloud from any viewpoint in a scene, given real ego-car's point cloud and semantic labels as conditions. TYP assumes access to 1) a simulated collaborative driving dataset with multiple agents perceiving the same scenes from different viewpoints and 2) a real single-agent driving dataset; both are labeled. Our solution involves a conditioned latent diffusion model [34] and a dedicated two-stage training process. In the first stage, we consider a single-agent scenario and train the model using real data conditioning only on object locations. This equips the model with the ability to generate diverse and realistic scenes. We denote the learned model by \\( P(\\pmb{x}|\\pmb{y}) \\), where \\( \\pmb{y} \\) stands for the semantic condition and \\( \\pmb{x} \\) for the point cloud. In the second stage, we incorporate the simulated data to learn how to ground generation on data from another agent's viewpoint so that the model can produce a semantically consistent reference point cloud given the ego-car's data. We learn a lightweight conditioning module to turn \\( P(\\pmb{x}|\\pmb{y}) \\) into \\( P(\\pmb{x}_r|\\pmb{x}_e,\\pmb{y}_r) \\), where \\( e \\) and \\( r \\) indicate ego and reference views, respectively. We note that \\( \\pmb{x}_e \\) was pre-translated to center around the reference position." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.909, + 0.578 + ], + "angle": 0, + "content": "One challenge in TYP is the domain gap between real and simulated data. Due to differences in sensor configurations and placements, data collection environments, and the sim-to-real gap, the two sets of point clouds inevitably exhibit discrepant distributions, patterns, and densities. To address this, we insert a domain adaptation step between the two training stages. We train a separate encoder-decoder for the simulated data while enforcing a constraint to make the encoded features of simulated and real data indistinguishable [40]. This step allows us to learn \\( P(\\boldsymbol{x}_r|\\boldsymbol{x}_e,\\boldsymbol{y}_r) \\) in a space with reduced domain discrepancy during the second stage." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.909, + 0.702 + ], + "angle": 0, + "content": "Once trained, we pair \\( P(\\pmb{x}_r|\\pmb{x}_e,\\pmb{y}_r) \\) with the real data's encoder-decoder to generate real-style point clouds grounding on real ego-car's data, so that we can develop collaborative perception algorithms without real collaborative data. More specifically, given a real ego-car's perception \\( \\pmb{x}_e \\) and label \\( \\pmb{y}_e \\), we first translate them to center around the reference position, followed by optionally injecting object labels into \\( \\pmb{y}_e \\) to make it realistic from the reference viewpoint." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.911, + 0.903 + ], + "angle": 0, + "content": "We extensively validate TYP on multiple datasets, all in an offline setting. Empirical results demonstrate TYP's effectiveness in generating high-quality reference data to aid in the development of collaborative perception. In particular, we show that a conditioned diffusion model trained solely on simulated data (e.g., OPV2V [51]) can already turn a real single-agent dataset into a real-alike collaborative one. As such, one can train collaborative perception algorithms for real test data without real training data. We further generate the \"ColWaymo\" dataset by training TYP on real single-agent Waymo data [39] and simulated OPV2V data [51], thus turning the former into a collaborative one. Collaborative perception backbones pre-trained on the large-scale, semi-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12028" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "synthetic ColWaymo data demonstrate a remarkable transferability. They notably improve collaborative perception developed in the downstream tasks (e.g., V2V4Real [52]), even in a many-shot setting." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.153, + 0.338, + 0.166 + ], + "angle": 0, + "content": "In sum, our key contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.168, + 0.486, + 0.258 + ], + "angle": 0, + "content": "- We propose a new research direction to aid in the development of collaborative perception, generating real-looking sensory data from potentially any viewpoint in a real driving scene, capable of turning a single-agent dataset into a collaborative one. This has the potential to scale up collaborative autonomous driving (CAV)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.258, + 0.483, + 0.303 + ], + "angle": 0, + "content": "- We present TYP, a two-stage training recipe with domain adaptation in between to learn the generative model using simulated collaborative data and real single-agent data." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.304, + 0.484, + 0.334 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate TYP's effectiveness in aiding in collaborative perception across various scenarios." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.168, + 0.486, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.347, + 0.232, + 0.362 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.372, + 0.486, + 0.825 + ], + "angle": 0, + "content": "Collaborative autonomous driving (CAV) offers significant benefits, including extended perception range by sharing sensor information to detect objects beyond a single vehicle's field of view. Despite these benefits, collecting large-scale, real-world datasets for CAV poses significant challenges due to the complexity and cost of deploying multiple instrumented vehicles in diverse environments. Existing datasets often have limitations that restrict the scope of CAV research. For instance, OPV2V [51] and V2X-Sim [20] rely on simulations using the CARLA simulator [7], allowing controlled scenarios with multiple vehicle agents, yet lacking real-world variability. V2V4Real [52], on the other hand, provides real-world data, yet it only contains two collaborating agents, which restricts the exploration of more complex multi-agent interactions. Similarly, the DAIR-V2X [57] dataset also offers real-world data including vehicle-to-infrastructure (V2I) and V2V, but it mainly focuses on V2I scenarios and still has a limited number of collaborating vehicles. To overcome current limitations, we propose a new research direction—generating realistic point clouds given the ego's perspective. 3D generation. Various works have explored 3D scene generation, such as LiDARGen [62], R2DM [27], LiDM [32], UltraLiDAR [48], and RangeLDM [13]. Additionally, Li-darDM [63] and Text2LiDAR [44] investigate conditional scene generation, with the former relying on hand-crafted map layouts and the latter conditioning on text inputs. However, all these methods focus on ego-centric generation. We propose a new research direction for CAV, aiming to generate realistic and consistent scenes conditioned on the ego agent's real point clouds—an area that remains largely unexplored." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Diffusion models. Diffusion models have recently advanced generative modeling for high-quality LiDAR point clouds and images. Denoising Diffusion Probabilistic Models (DDPMs) [12] outperform traditional Generative Adversarial Networks (GANs) [10] and further enhance efficiency" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.264, + 0.908, + 0.389 + ], + "angle": 0, + "content": "Figure 2. Illustration of TYP's conditioned generative model and training process. We propose a two-stage training procedure. The first stage maximizes the generation capability by conditioning solely on object locations (using real single-agent target data), while the second stage grounds the generation on the ego-car's perspective to match semantics and layouts (using simulated CAV data). Additionally, we introduce a discriminator to adapt simulated CAV features to the real target domain, making the trained model readily applicable to the target domain after the second stage." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.908, + 0.472 + ], + "angle": 0, + "content": "for LiDAR generation. In autonomous driving applications, methods like RangeLDM [13], LidarDM [63], and LiDAR-Gen [62] apply diffusion models for realistic LiDAR scene generation. In this paper, we utilize its generation capability to study our proposed problem." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.483, + 0.818, + 0.5 + ], + "angle": 0, + "content": "3. Transfer Your Perspective (TYP)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.909, + 0.676 + ], + "angle": 0, + "content": "In this paper, we introduce a new research direction to advance collaborative autonomous driving (CAV): generating LiDAR point clouds from different perspectives within the same scene as the ego agent, aiming to reduce the tedious efforts of collecting data for CAV. We begin by defining the proposed problem in Sec. 3.1. In Sec. 3.2, we discuss the representations of the inputs, including point clouds and semantic information. Sec. 3.3 outlines the pipeline developed to address this problem. Finally, in Sec. 3.4, we demonstrate how this capability can be applied to datasets that only have labeled ego agents, e.g., Waymo Open Dataset (WOD) [39]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.683, + 0.665, + 0.699 + ], + "angle": 0, + "content": "3.1. Problem Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.825 + ], + "angle": 0, + "content": "Given the perception data \\( \\pmb{x}_e \\) from an ego agent, we aim to build a model \\( P(\\pmb{x}_r|\\pmb{x}_e) \\) to generate new perception data \\( \\pmb{x}_r \\) seen from a different reference location and perspective within a communication range. Here, \\( \\pmb{x} \\) represents a LiDAR point cloud. This task presents several challenges, including potential information gaps between the two views, the need for alignment within the commonly perceivable area, and ensuring realism in the generated data, as discussed in Sec. 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.902 + ], + "angle": 0, + "content": "To address these challenges, we extend the original problem by incorporating semantic information, such as object bounding boxes, represented as \\( P(\\pmb{x}_r|\\pmb{x}_e,\\pmb{y}_r) \\). We assume the availability of this information around the ego agent, denoted by \\( \\pmb{y}_e \\). For example, most of the existing single-agent" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12029" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.258 + ], + "angle": 0, + "content": "datasets provide object labels. This semantic information can be easily translated and edited to become \\( \\mathbf{y}_r \\) by tools like traffic re-players [11], making it a key to bridge the information gap between the ego and reference views, enabling the reference agent to \"see\" those objects and surfaces beyond the ego's view—a core concept in CAV. Additionally, by aligning generated point clouds with object locations, this extension facilitates our goal of scaling up CAV development. In essence, the generated \\( (\\mathbf{x}_r, \\mathbf{y}_r) \\) pairs can be used flexibly alongside \\( (\\mathbf{x}_e, \\mathbf{y}_e) \\) in various CAV applications, like directly using them to train collaborative perception algorithms." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.258, + 0.484, + 0.289 + ], + "angle": 0, + "content": "In the following sections, we describe how \\( \\pmb{x} \\) and \\( \\pmb{y} \\) are encoded, followed by our approach to tackling the problem." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.296, + 0.386, + 0.312 + ], + "angle": 0, + "content": "3.2. Representations and Embeddings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.485, + 0.469 + ], + "angle": 0, + "content": "LiDAR. There are multiple ways to represent point clouds in continuous 3D space, such as coordinates (i.e., x, y, and z) [30, 31, 37], range images [25, 30, 43], and voxelization [17, 24, 53, 59]. To better align with spatial control from object locations, we follow [49] to voxelize point clouds using a pre-defined grid and record voxel occupancy. [49] also highlights that this representation can naturally handle varying point densities and only minimally impacts LiDAR generation, with some precision trade-offs during voxelization. In short, we represent a point cloud by \\( \\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times C} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.484, + 0.559 + ], + "angle": 0, + "content": "To avoid the computational cost of 3D convolutions, we convert \\( x \\) into Bird's-Eye-View (BEV) images by treating the height dimension (i.e., \\( C \\)) as feature channels for 2D convolutions. This approach has been widely adopted in self-driving perception [5, 48, 58], allowing the use of 2D image-based model architectures and algorithms." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.561, + 0.484, + 0.65 + ], + "angle": 0, + "content": "Semantic information. Given that the point clouds are represented as BEV images, it is also intuitive to represent object locations in BEV. We create binary object maps by considering only the x and y coordinates of 3D bounding boxes, resulting in \\(\\pmb{y} \\in \\mathbb{R}^{H \\times W \\times 1}\\). (We could further extend it by including category information.)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.651, + 0.484, + 0.801 + ], + "angle": 0, + "content": "Feature embedding. Following the literature [34], we encode input tensors using a Vector Quantized-Variational Autoencoder (VQ-VAE) [41], comprising an encoder \\( E \\), a quantization function \\( Q \\) for feature vectors on spatial grids, and a decoder \\( G \\). Formally, we obtain \\( \\boldsymbol{x}^f = E_x(\\boldsymbol{x}) \\in \\mathbb{R}^{h \\times w \\times c_x} \\) by the encoder, map each \\( c_x \\)-dimensional vector to a learnable code to obtain \\( \\boldsymbol{x}^z = Q_x(\\boldsymbol{x}^f) \\in \\mathbb{R}^{h \\times w \\times c_x} \\), and generate outputs using the decoder \\( \\hat{\\boldsymbol{x}} = G_x(\\boldsymbol{x}^z) \\in \\mathbb{R}^{H \\times W \\times C} \\), where \\( h \\) and \\( w \\) are spatial resolutions of the feature map downsampled from \\( H \\) and \\( W \\), and \\( c_x \\) is the number of channels." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.802, + 0.484, + 0.817 + ], + "angle": 0, + "content": "The VQ-VAE model is trained end-to-end by minimizing:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.829, + 0.484, + 0.848 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {v q}} = \\mathcal {L} _ {\\mathrm {r e c}} + \\| \\operatorname {s g} [ E _ {x} (\\boldsymbol {x}) ] - \\boldsymbol {x} ^ {z} \\| _ {2} ^ {2} + \\| \\operatorname {s g} [ \\boldsymbol {x} ^ {z} ] - E _ {x} (\\boldsymbol {x}) \\| _ {2} ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathrm{sg}[\\cdot ]\\) and \\(\\mathcal{L}_{\\mathrm{rec}}\\) denote the stop-gradient operation and the reconstruction loss, respectively. As our point cloud representation is binary occupancy \\(\\pmb {x}\\in \\{0,1\\}^{H\\times W\\times C}\\), binary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.138 + ], + "angle": 0, + "content": "cross-entropy is a natural choice for \\(\\mathcal{L}_{\\mathrm{rec}}\\). However, due to the sparsity of point clouds, this results in an imbalanced loss. To address this, we adopt the Focal Loss (FL) [36]:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.149, + 0.907, + 0.183 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\mathrm {F L}} \\left(\\boldsymbol {x} _ {i}, \\hat {\\boldsymbol {x}} _ {i}\\right) = \\left\\{ \\begin{array}{l l} - (1 - \\hat {\\boldsymbol {x}} _ {i}) ^ {\\gamma} \\log \\left(\\hat {\\boldsymbol {x}} _ {i}\\right) & \\text {i f} \\boldsymbol {x} _ {i} = 1 \\\\ - \\hat {\\boldsymbol {x}} _ {i} ^ {\\gamma} \\log \\left(1 - \\hat {\\boldsymbol {x}} _ {i}\\right) & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.632, + 0.198, + 0.907, + 0.24 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r e c}} = \\sum_ {i = 0} ^ {M} \\ell_ {\\mathrm {F L}} \\left(\\boldsymbol {x} _ {i}, \\hat {\\boldsymbol {x}} _ {i}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.247, + 0.895, + 0.261 + ], + "angle": 0, + "content": "where \\(i\\) is the voxel index and \\(M\\) is the number of voxels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.262, + 0.906, + 0.323 + ], + "angle": 0, + "content": "For object locations \\(\\pmb{y}\\), we use the same approach to train a separate VQ-VAE model \\(E_{y}, Q_{y}\\), and \\(G_{y}\\). In the subsequent subsections, we use \\(\\pmb{x}^{f} \\in \\mathbb{R}^{h \\times w \\times c_{x}}\\) and \\(\\pmb{y}^{f} \\in \\mathbb{R}^{h \\times w \\times c_{y}}\\) as feature embeddings for learning a latent diffusion model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.334, + 0.866, + 0.35 + ], + "angle": 0, + "content": "3.3. Transfer Your Perspective by Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.356, + 0.907, + 0.416 + ], + "angle": 0, + "content": "We model the distribution \\( P(\\pmb{x}_r | \\pmb{x}_e, \\pmb{y}_r) \\) defined in Sec. 3.1 by a conditioned generative model and train it in two stages. We assume access to \\( (\\pmb{x}_r, \\pmb{x}_e, \\pmb{y}_r) \\) tuples as training data. We detail the training data preparation in Sec. 3.4 and Sec. 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.908, + 0.598 + ], + "angle": 0, + "content": "In the first stage, we aim to maximize the generation capability by providing minimal conditions, i.e., only bounding boxes, without further constraints on the scene. This encourages the model to produce \\( P(\\boldsymbol{x}|\\boldsymbol{y}) \\) with high flexibility. Second, building on this model, we incorporate the ego agent's point cloud as an additional cue and ground the generation process on it to ensure semantic and layout consistency between the generated and ego agent's point clouds. Put together, this two-stage training procedure enables the learned \\( P(\\boldsymbol{x}_r|\\boldsymbol{x}_e,\\boldsymbol{y}_r) \\) model to generate perception data at areas beyond the commonly visible regions between two views. The model and training process is illustrated in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.704 + ], + "angle": 0, + "content": "More importantly, this approach enables us to leverage existing labeled single-agent datasets such as KITTI [9], NuScenes [2], and Waymo Open Dataset (WOD) [39], since the first-stage training only requires ego-centric LiDAR point clouds and their corresponding semantic information (see Sec. 3.4). The \\((\\pmb{x}_e,\\pmb{x}_r)\\) data pairs seen from the ego and the reference agents are needed only in the second stage." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.705, + 0.884, + 0.719 + ], + "angle": 0, + "content": "In the following, we elaborate on each training stage." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Stage 1: Generation with semantic information. The goal of this stage is to equip the model with strong generation capabilities of point clouds given spatial conditions, preparing it for the next stage. As discussed in Sec. 3.2, the point clouds and object locations are embedded by VQVAEs [41], denoted as \\( \\boldsymbol{x}^f = E_x(\\boldsymbol{x}) \\in \\mathbb{R}^{h \\times w \\times c_x} \\) and \\( \\boldsymbol{y}^f = E_x(\\boldsymbol{y}) \\in \\mathbb{R}^{h \\times w \\times c_y} \\). The image-like feature maps allow us to adopt existing generation algorithms for 2D images [6, 12, 16]. In this paper, we apply one of the most popular generative models, the Latent Diffusion Model (LDM) [34], for conditioned generation \\( P(\\boldsymbol{x} | \\boldsymbol{y}) \\). LDM seeks to model a data distribution by iteratively denoising variables" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12030" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.48, + 0.108 + ], + "angle": 0, + "content": "that are initially sampled from a Gaussian. The objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.117, + 0.485, + 0.151 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {L D M}} = \\mathbb {E} _ {\\boldsymbol {x} ^ {f}, \\boldsymbol {y} ^ {f}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\boldsymbol {x} _ {t} ^ {f}, \\boldsymbol {y} ^ {f}, t\\right) \\right\\| _ {2} ^ {2} \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.161, + 0.474, + 0.176 + ], + "angle": 0, + "content": "where, \\(\\epsilon_{\\theta}\\) is a UNet [35] backbone and \\(t\\) is the timestamp." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.177, + 0.483, + 0.266 + ], + "angle": 0, + "content": "Stage 2: Generation for new perspectives. In this stage, we aim to ground the generation of point clouds by incorporating the ego agent's perception. To retain the generation capability from stage 1, we freeze the learned \\( P(\\boldsymbol{x}|\\boldsymbol{y}) \\) and introduce a learnable lightweight control module to inject additional cues, following T2I-Adapter [26], as shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.268, + 0.484, + 0.358 + ], + "angle": 0, + "content": "Formally, let \\( \\pmb{x}_{e'} \\) denote the translated and rotated egoagent's point cloud centered around the reference location and aligned with the reference orientation, as shown in Fig. 2; \\( \\pmb{x}_{e'}^f = E_x(\\pmb{x}_{e'}) \\in \\mathbb{R}^{h \\times w \\times c_x} \\) is the corresponding embedding. Let \\( F_{AD} \\) denote the learnable control module, which takes \\( \\pmb{x}_{e'}^f \\) as input and outputs" + }, + { + "type": "equation", + "bbox": [ + 0.228, + 0.369, + 0.484, + 0.389 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {c} = F _ {A D} \\left(\\boldsymbol {x} _ {e ^ {\\prime}} ^ {f}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.399, + 0.485, + 0.462 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}_c = \\{\\mathbf{F}_c^1,\\mathbf{F}_c^2,\\mathbf{F}_c^3,\\mathbf{F}_c^4\\}\\) matches the size of the multiscale features \\(\\mathbf{F}_{enc} = \\{\\mathbf{F}_{enc}^{1},\\mathbf{F}_{enc}^{2},\\mathbf{F}_{enc}^{3},\\mathbf{F}_{enc}^{4}\\}\\) extracted from the frozen encoder of the UNet \\(\\epsilon_{\\theta}\\). With these ingredients, T2I-Adapter influences the generation process by" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.474, + 0.484, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {e n c} ^ {i} \\leftarrow \\mathbf {F} _ {e n c} ^ {i} + \\mathbf {F} _ {c} ^ {i}, i \\in \\{1, 2, 3, 4 \\}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.498, + 0.484, + 0.53 + ], + "angle": 0, + "content": "which injects \\( \\pmb{x}_{e'}^f \\) embedding into each of the denoising steps. The second-stage objective for learning \\( F_{AD} \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.54, + 0.484, + 0.582 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {L D M}} = \\mathbb {E} _ {\\boldsymbol {x} _ {r} ^ {f}, \\boldsymbol {y} _ {r} ^ {f}, \\boldsymbol {x} _ {e ^ {\\prime}} ^ {f}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\boldsymbol {x} _ {r, t} ^ {f}, \\boldsymbol {y} ^ {f}, \\boldsymbol {x} _ {e ^ {\\prime}} ^ {f}, t\\right) \\right\\| _ {2} ^ {2} \\right], \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.468, + 0.599 + ], + "angle": 0, + "content": "where the encoder of \\(\\epsilon_{\\theta}\\) is modified as defined in Eq. (6)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.607, + 0.47, + 0.623 + ], + "angle": 0, + "content": "3.4. From Single Agent to Collaborative Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.704 + ], + "angle": 0, + "content": "There are many publicly accessible large-scale real-world autonomous driving datasets for ego-centric 3D perception, like WOD [39], NuScenes [2], KITTI-360 [21], and PandaSet [46], whereas much fewer datasets exist for CAV due to the challenges in data collection as discussed in Sec. 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.84 + ], + "angle": 0, + "content": "This limitation inspired a bold idea: Can we transfer the generation capability learned from paired CAV data to these existing labeled single-agent datasets? Achieving this would scale up CAV datasets substantially, offering immense benefits to the research community. However, this goal is challenging due to domain gaps between datasets, such as variations in LiDAR sensors, point cloud densities, data patterns, and data collection environments. Below, we describe our approach to addressing these challenges." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Problem setup. Our target domain is a single-agent dataset, while the source domain consists of paired perception and semantic information from ego and reference perspectives, denoted as \\((\\pmb{x}_r,\\pmb{x}_e,\\pmb{y}_r)\\). We demonstrate that we can transfer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.121 + ], + "angle": 0, + "content": "knowledge from a simulated CAV dataset (e.g., OPV2V [51]) to a real-world ego-centric dataset (e.g., WOD [39])." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.907, + 0.212 + ], + "angle": 0, + "content": "Two-Stage training is the key. As described in Sec. 3.3, we decompose the generative training into two stages: first learning \\( P(\\pmb{x}|\\pmb{y}) \\), followed by \\( P(\\pmb{x}_r|\\pmb{x}_e,\\pmb{y}_r) \\). The second stage keeps \\( P(\\pmb{x}|\\pmb{y}) \\) frozen by adding an adapter. This strategy enables us to use target domain data in the first stage. The resulting \\( P(\\pmb{x}|\\pmb{y}) \\) would generate target-like point clouds." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.213, + 0.909, + 0.469 + ], + "angle": 0, + "content": "Minimizing domain gaps. In the first stage, we have learned VQ-VAE encoder-decoders and \\( P(\\boldsymbol{x}|\\boldsymbol{y}) \\) in the target domain (single-agent). In the second stage, we bring in CAV data to guide the generation on how to condition on ego-agent's perception. To minimize the gap of \\( \\boldsymbol{x}^f \\) between two domains—\\( \\boldsymbol{x}^f \\) is what the generative model is optimized to generate—we adopt a GAN-style discriminator [8, 10, 22, 40]. Specifically, we aim to make feature embeddings extracted from the source data and the target data indistinguishable; namely, to confuse a discriminator \\( D \\) trained to differentiate them. We freeze all parameters of the target domain's encoder \\( E_{tgt} \\), codebook \\( Q_{tgt} \\), and decoder \\( G_{tgt} \\) in the VQ-VAE. For the source domain, we initialize its \\( E_{src} \\) and \\( G_{src} \\) by \\( E_{tgt} \\) and \\( G_{tgt} \\) while reusing the frozen codebook \\( Q_{tgt} \\). We then learn the discriminator \\( D \\) and the source domain's encoder \\( E_{src} \\) and decoder \\( G_{src} \\) in an interleaving fashion by minimizing the following objective functions:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.479, + 0.906, + 0.515 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{r l} \\text {F o r} D: \\mathcal {L} _ {D} = & \\mathbb {E} _ {\\boldsymbol {x} \\sim P _ {t g t}} [ \\max (0, 1 - D (E _ {t g t} (\\boldsymbol {x}))) ] \\\\ & + \\mathbb {E} _ {\\boldsymbol {x} \\sim P _ {s r c}} [ \\max (0, 1 + D (E _ {s r c} (\\boldsymbol {x}))) ]; \\end{array} \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.526, + 0.906, + 0.568 + ], + "angle": 0, + "content": "\\[\n\\text {F o r V Q - V A E :} \\mathcal {L} _ {\\mathrm {v q}} = \\mathcal {L} _ {\\mathrm {r e c}} + \\mathcal {L} _ {s r c} + \\| \\operatorname {s g} [ \\boldsymbol {x} ^ {z} ] - E _ {s r c} (\\boldsymbol {x}) \\| _ {2} ^ {2}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.573, + 0.906, + 0.603 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} \\mathcal {L} _ {s r c} = - \\mathbb {E} _ {\\boldsymbol {x} \\sim P _ {s r c}} D \\left(E _ {s r c} (\\boldsymbol {x})\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.906, + 0.704 + ], + "angle": 0, + "content": "The other terms in Eq. (9) follow those in Eq. (1), where for \\( \\pmb{x} \\sim P_{src} \\), \\( \\hat{\\pmb{x}} = G_{src} \\circ Q_{tgt} \\circ E_{src}(\\pmb{x}) \\) in \\( \\mathcal{L}_{rec} \\). We note that \\( \\mathcal{L}_D \\) takes in two streams of data (i.e., source and target) and is defined by the Hinge Loss [22] on \\( \\pmb{x}^f \\); \\( \\mathcal{L}_{\\mathrm{vq}} \\) takes in only the source CAV data. We also note that embeddings for object locations does not require adaptation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.78 + ], + "angle": 0, + "content": "After this adaptation step, we proceed into the second stage of generative training, where we drop the discriminator and train the control module defined in Eq. (5) by minimizing Eq. (7). The feature embeddings are produced by \\( E_{src} \\) on top of \\( (\\pmb{x}_r, \\pmb{x}_e, \\pmb{y}_r) \\sim P_{src} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Enhancement in target domains. With the discriminator and adversarial training, we reduce the gap between source and target domain data, successfully adapting TYP to the single-agent dataset, as shown in Sec. 4.5. To further improve the conditioned generation quality in the target domain, a further fine-tuning stage in it is desired. However, how can we do so without having reference agents' point clouds in the target domain? Here, we propose two simple solutions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12031" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.116, + 0.089, + 0.46, + 0.246 + ], + "angle": 0, + "content": "
MethodTrain DataAP 0.5
smlall
No Fusionego's gt only0.670.410.130.40
Early Fusion [4]gt (oracle)0.760.420.310.49
TYP (ours)0.750.380.290.46
Late Fusion [51]gt (oracle)0.740.570.360.55
TYP (ours)0.710.490.320.50
AttFuse [51]gt (oracle)0.940.800.620.77
TYP (ours)0.900.730.560.72
V2X-ViT [50]gt (oracle)0.870.710.500.71
TYP (ours)0.840.650.400.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.25, + 0.483, + 0.292 + ], + "angle": 0, + "content": "Table 1. Results on OPV2V. We compare the performance using reference generated by TYP versus ground-truth (oracle). The comparable results demonstrate the quality of our generated data." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.294, + 0.226, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.342, + 0.199, + 0.351 + ], + "angle": 0, + "content": "(a) no adaptation" + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.294, + 0.348, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.342, + 0.303, + 0.351 + ], + "angle": 0, + "content": "(b) disc." + }, + { + "type": "image", + "bbox": [ + 0.348, + 0.294, + 0.471, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.355, + 0.342, + 0.458, + 0.351 + ], + "angle": 0, + "content": "(c) disc. + enhancement" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.353, + 0.483, + 0.395 + ], + "angle": 0, + "content": "Figure 3. Qualitative results on enhancement in the target domain. Generated point cloud (green) has better quality with the enhancement given ego (gray) from Waymo (cf. Sec. 3.4)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.403, + 0.485, + 0.645 + ], + "angle": 0, + "content": "Recall that in training \\( P(\\pmb{x}_r|\\pmb{x}_e,\\pmb{y}_r) \\) using the source CAV data, point clouds show up in both the input as conditions and the outputs as supervisions. In other words, the domain gap exists on both sides, and we aim to reduce it by further fine-tuning using the target single-agent dataset. First, to enhance the output generation in the target domain, we set ego's point clouds to empty, i.e., \\( \\pmb{x}_e = 0^{H\\times W\\times C} \\) for \\( (\\pmb{x}_r,\\pmb{y}_r)\\sim P_{tgt} \\). Then, to enable adaptation on inputs, we apply self-training [18, 28, 60, 61], randomly sampling a reference location in the scene to generate pseudo-point clouds for fine-tuning, i.e., \\( (\\hat{\\pmb{x}}_r,\\pmb{x}_e,\\pmb{y}_r) \\) for \\( (\\pmb{x}_e,\\pmb{y}_r)\\sim P_{tgt} \\). We use the above target domain data and the source domain data to jointly fine-tune the control module (cf. Sec. 3.3, Eq. (6), and Eq. (7)) further. Empirical results in Sec. 4.5 show the effectiveness of the proposed solution in turning single-agent datasets into collaborative datasets for CAV development." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.658, + 0.224, + 0.675 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.683, + 0.173, + 0.699 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.486, + 0.886 + ], + "angle": 0, + "content": "Datasets. In this study, we primarily use three datasets: OPV2V [51], V2V4Real [52], and WOD [39]. OPV2V is a simulation-based CAV dataset that includes over 70 diverse driving scenes and more than 11,000 frames. We utilize this dataset initially to validate our proposed research problem and later to guide the generation training on single-agent datasets. V2V4Real is a real-world CAV dataset with approximately 20,000 LiDAR frames, which we use to validate TYP in a real-world CAV context, treating it as a single-agent dataset (i.e., without using any labels to train TYP). Lastly, we adopt WOD to scale up TYP, one of the largest real-world single-agent datasets available." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Evaluation metrics. Following standard benchmarks [51," + }, + { + "type": "table", + "bbox": [ + 0.539, + 0.089, + 0.883, + 0.246 + ], + "angle": 0, + "content": "
MethodTrain DataAP 0.5
smlall
No Fusionego's gt only0.710.330.080.46
Early Fusion [4]gt (oracle)0.790.420.450.60
TYP (ours)0.760.360.300.53
Late Fusion [51]gt (oracle)0.790.530.560.67
TYP (ours)0.770.480.510.63
AttFuse [51]gt (oracle)0.800.450.360.61
TYP (ours)0.750.450.260.56
V2X-ViT [50]gt (oracle)0.810.490.300.61
TYP (ours)0.760.380.170.48
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.25, + 0.906, + 0.32 + ], + "angle": 0, + "content": "Table 2. Results on V2V4Real. We compare the performance using reference generated by TYP versus ground-truth (oracle). Note that we do not use V2V4Real data to train the generation, treating it as a single-agent dataset. The comparable results demonstrate the quality of our generated data and the transferability of TYP." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.327, + 0.908, + 0.704 + ], + "angle": 0, + "content": "52], we report Average Precision (AP 0.5) across three distance ranges: 0-30 meters (short, denoted as s), 30-50 meters (medium, m), and 50-100 meters (long, l) and overall (all). Implementations. For TYP training, we voxelize point clouds within the range \\([-51.2, 51.2]\\) meters in x and y dimensions and 4 meters in z (range is dataset-dependent) to create 3D volumes, \\(x \\in \\mathbb{R}^{512 \\times 512 \\times 20}\\). This is then encoded to \\(x^f \\in \\mathbb{R}^{64 \\times 64 \\times 8}\\) with VQ-VAE [41]. Similarly, semantic information \\(y \\in \\mathbb{R}^{512 \\times 512 \\times 1}\\) is encoded to \\(y^f \\in \\mathbb{R}^{64 \\times 64 \\times 3}\\). We train VQ-VAE for \\(x\\) 120 epochs and for \\(y\\) 10 epochs, with a batch size of 6 per GPU. Next, we train LDM with a batch size of 48 per GPU for 200 epochs in the first stage and 80 epochs in the second stage. During the second stage, we freeze the U-Net from the first stage and fine-tune only the control module, following [26]. For adaptation, we initialize the weights of encoder-decoder for CAV data (fake) with the model trained on single-agent data (real), freeze its codebook, and fine-tune it with a discriminator comprising 4 convolutional layers. The model processes an equal amount of fake and real per batch (i.e., 3 samples each). All training is conducted on 4 NVIDIA H100 GPUs. For experiments on benchmarks, we follow the convention to report results with Pointpillars [17]. We train / fine-tune all models with 8 P100, with batch size 4 per GPU. For V2X-ViT, we use 4 H100 due to resources required." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.841 + ], + "angle": 0, + "content": "Inference of generation and post-processing. We use DDIM scheduler [38] for 200 steps. For post-processing, we apply modified Gumbel-Softmax [15] to the logits. Specifically, we define two thresholds, low and high. Scores below the low threshold are dropped, those above the high threshold are retained, and values in between are perturbed with Gaussian noise on logits. Hyperparameters are tuned on a small hold-out set. Finally, we convert the volume back to coordinates by using the centers of the occupied voxels." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.85, + 0.866 + ], + "angle": 0, + "content": "4.2. Do We Need Collaborative Perception?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To explore this question, we consistently report the performance without collaborative agents (denoted as \"ego's gt" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12032" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.137, + 0.089, + 0.861, + 0.293 + ], + "angle": 0, + "content": "
MethodPre-TrainFT. on 0 SceneFT. on 5 SceneFT. on 10 SceneFT. on 32 Scene
smlallsmlallsmlallsmlall
No Fusionego's gt only0.440.160.060.290.660.290.080.420.710.340.080.47
scratch0.440.180.160.300.730.280.240.480.790.420.450.60
Early Fusion [4]OPV2V0.540.200.070.310.320.180.150.250.730.320.310.500.800.470.540.65
ColWaymo (ours)0.500.240.240.350.680.340.320.510.780.390.330.570.830.480.500.65
Late Fusion [51]scratch0.440.240.330.340.720.420.510.580.790.530.560.67
OPV2V0.600.270.280.440.550.260.380.420.730.420.510.590.780.550.580.67
ColWaymo (ours)0.400.190.150.250.600.280.440.470.760.430.510.610.820.580.610.71
AttFuse [51]scratch0.400.150.100.280.700.300.170.470.800.450.360.61
OPV2V0.510.190.050.310.540.220.110.370.770.400.210.540.830.530.400.65
ColWaymo (ours)0.660.350.110.450.650.290.160.460.830.460.330.610.880.580.530.72
V2X-ViT [50]scratch0.430.150.120.310.700.280.170.430.810.490.300.61
OPV2V0.510.240.070.330.480.230.160.350.760.380.220.530.810.490.350.61
ColWaymo (ours)0.600.280.100.340.660.280.220.460.790.480.260.580.840.570.440.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.296, + 0.907, + 0.353 + ], + "angle": 0, + "content": "Table 3. Results with pre-training on collaborative Waymo. We scale up TYP with exiting large-scale single-agent dataset WOD [39], creating its collaborative version \"ColWaymo\". These data are used for pre-training and subsequently fine-tuned on V2V4Real [52]. Results show significant improvement over training from scratch, highlighting the potential to reduce data collection efforts, scale up, and accelerate CAV development. Comparisons to simulated OPV2V [51] pre-training further demonstrate the realism of our generated point clouds." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.358, + 0.468, + 0.442 + ], + "angle": 0, + "content": "
MethodDisc.ST.Dum.JSD (↓)MMD (↓)
No Adaptation0.264.61 × 10-4
✓0.161.10 × 10-4
Adaptation✓✓0.171.30 × 10-4
✓✓✓0.161.17 × 10-4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.446, + 0.485, + 0.502 + ], + "angle": 0, + "content": "Table 4. Ablation of proposed adaption method. Results with adaptation outperform no adaptation, demonstrating an improvement in generation quality. Abbreviations in title refer to discriminator, self-training, and dummy ego, respectively (cf. Sec. 3.4)." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.504, + 0.468, + 0.554 + ], + "angle": 0, + "content": "
Num. StepsEpochs 1stEpochs 2ndJSD (↓)MMD (↓)
1-2800.116.45 × 10-5
2200800.105.28 × 10-5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.558, + 0.483, + 0.587 + ], + "angle": 0, + "content": "Table 5. 2-Stage training. Results demonstrate the benefit of the proposed 2-stage training in generation quality on OPV2V [51]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.483, + 0.716 + ], + "angle": 0, + "content": "only\" in Tab. 1, Tab. 2, and Tab. 3). For example, in the V2V4Real dataset, a well-trained detector struggles with objects beyond 50 meters, achieving only 0.08 AP at that range, though it performs well on nearby objects with an AP of 0.71 (cf. first row of Tab. 3). simply combining the predictions from two under-trained detectors, we observe a notable boost in performance to 0.33 (cf. \"scratch\" in late fusion, Tab. 2), suggesting the significance of CAV." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.727, + 0.406, + 0.743 + ], + "angle": 0, + "content": "4.3. Point Clouds Generation on OPV2V" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.871 + ], + "angle": 0, + "content": "Setting. To validate our research concept, we begin with the simulated dataset OPV2V [51]. We split the original training set of 44 scenes into two halves, using only the first 22 scenes to train our generation model. For the remaining 22 scenes, we generate point clouds for reference agents based on the ego agent's point clouds and compare these results to ground-truth point clouds. We follow the benchmark to report performance on the validation set from Culver City." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Baselines. Following the literature [51, 52], we adopt three common fusion algorithms for CAV: early, late, and inter" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.361, + 0.908, + 0.406 + ], + "angle": 0, + "content": "mediate fusion, which fuse input point clouds, extracted features, and predicted bounding boxes, respectively. For intermediate fusion, we adipot AttFuse [51] and V2X-VIT [50]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.407, + 0.909, + 0.558 + ], + "angle": 0, + "content": "Comparison to using ground-truth LiDAR. In Tab. 1, we show that training CAV with point clouds generated by TYP achieves performance comparable to that obtained with ground-truth (oracle) across all methods (e.g., 0.46 vs. 0.49 with early, 0.50 vs. 0.55 with late, and 0.77 vs. 0.72 with AttFuse [51], etc.) This result highlights the quality of the generated point clouds. In the supplemental material, we further demonstrate that with additional labeled data (from the first half of the split used for training generation), the gap to oracle performance can be further minimized." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.57, + 0.878, + 0.587 + ], + "angle": 0, + "content": "4.4. Single-Agent to Collaborative Multi-Agent" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.593, + 0.909, + 0.713 + ], + "angle": 0, + "content": "Setting. Building on our findings with simulated data (cf. Sec. 4.3), we validate our approach on the real-world dataset V2V4Real [52], treating it as a single-agent dataset without access to reference agents' point clouds. Specifically, we use the model from Sec. 4.3 to generate reference car's point clouds from the ego's perspective and compare performance against ground-truth point clouds (oracle). Note that no V2V4Real data is used to train the generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.715, + 0.909, + 0.837 + ], + "angle": 0, + "content": "Results on V2V4Real. In Tab. 2, we show the performance using generated reference point clouds is comparable to the oracle, supporting our idea of translating a single-agent dataset to multi-agent CAV. Xu et al. previously explored OPV2V-to-V2V4Real adaptation, noting significant performance drops even with adaptation algorithms. Here, TYP offers a new solution to this challenge – in a generative way, providing a fresh perspective on domain adaptation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.878, + 0.865 + ], + "angle": 0, + "content": "4.5. ColWaymo: Scaling with Large-Scale Data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Scaling CAV to the next level. In Sec. 4.4, we show the potential to create realistic reference's point clouds on real-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12033" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.144, + 0.089, + 0.321, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.322, + 0.089, + 0.497, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.089, + 0.673, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.089, + 0.852, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.226, + 0.908, + 0.254 + ], + "angle": 0, + "content": "Figure 4. Qualitative results on Collaborative Waymo. The gray point clouds are from the original single-agent dataset and the green are generated by TYP conditioning on them." + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.258, + 0.377, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.258, + 0.616, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.259, + 0.857, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.365, + 0.376, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.378, + 0.365, + 0.616, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.367, + 0.857, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.472, + 0.908, + 0.515 + ], + "angle": 0, + "content": "Figure 5. Qualitative results. Our proposed TYP is capable of scene editing, by inputting the same point cloud but different object locations. We (b) remove and (c) add a car from (a) the original point cloud. Inspired by the idea of past traversals [56], we apply completely different traffic conditions and generate (d), (e), and (f), to imagine driving through the same intersection." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.523, + 0.486, + 0.688 + ], + "angle": 0, + "content": "world dataset V2V4Real [52], through a simulation-guided training of TYP. Here, our goal is broader: to transform an existing large-scale single-agent dataset (e.g., WOD [39]) into a collaborative version. Using the pipeline described in Sec. 3 and illustrated in Fig. 2, we train the generative model on 17,400 samples from the WOD training set. We then use its validation set to generate approximately 36,000 collaborative samples by sampling reference locations from labeled vehicles. These generated samples serve to pre-train a model, which we subsequently fine-tune on V2V4Real [52]. For comparison, we also conduct pre-training on OPV2V [51]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.486, + 0.869 + ], + "angle": 0, + "content": "Results of fine-tuning on V2V4Real. Following prior work [1, 29, 47, 54], we perform pre-training followed by fine-tuning on limited labeled data to validate the potential of TYP for scaling up CAV development. As shown in Tab. 3, pre-training on our generated data yields significantly better performance than training from scratch. Additionally, compared to pre-training on the simulated dataset OPV2V, ColWaymo consistently achieves higher performance, highlighting the realism of the generated point clouds. These results demonstrate that TYP effectively transforms an ego-only dataset into a collaborative version, and potentially reducing the need for extensive CAV data collection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Generation quality by proposed adaptation. To assess the quality of generated data, we use two common metrics:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.909, + 0.599 + ], + "angle": 0, + "content": "Jensen-Shannon Divergence (JSD) and Maximum Mean Discrepancy (MMD). As shown in Tab. 4, our adaptation method improves JSD from 0.26 to 0.16 and MMD from 4.61 to 1.17 compared to the baseline without adaptation. Additionally, Fig. 3 visually illustrates this improvement." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.608, + 0.696, + 0.623 + ], + "angle": 0, + "content": "4.6. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.629, + 0.909, + 0.751 + ], + "angle": 0, + "content": "We illustrate the proposed problem in Fig. 1, which highlights the benefits of collaboration by allowing an agent to \"see\" beyond the limitations of its own sensors. TYP generates realistic point clouds from various viewpoints, enabling the agents to perceive previously occluded objects while preserving the overall scene semantics. Fig. 4 shows our ColWaymo, and Fig. 5 further demonstrates the model's capability for scene editing." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.764, + 0.634, + 0.779 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.841 + ], + "angle": 0, + "content": "We introduce a new research direction in collaborative driving and present the first solution. Empirical results show our approach can significantly reduce data collection and development efforts, advancing safer autonomous systems." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Limitations and future work. Following existing benchmarks [51, 52], TYP focuses on vehicle-like objects. Future work could extend it to broader objects and static entities (e.g., traffic signs, signals) essential for real-world traffic." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12034" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.486, + 0.178 + ], + "angle": 0, + "content": "This research is supported in part by grants from the National Science Foundation (IIS-2107077, IIS-2107161). We are thankful for the generous support of the computational resources by the Ohio Supercomputer Center." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.191, + 0.187, + 0.207 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.216, + 0.484, + 0.257 + ], + "angle": 0, + "content": "[1] Alexandre Boulch, Corentin Sautier, Björn Michele, Gilles Puy, and Renaud Marlet. Also: Automotive lidar self-supervision by occupancy estimation. In CVPR, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.484, + 0.314 + ], + "angle": 0, + "content": "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.317, + 0.484, + 0.385 + ], + "angle": 0, + "content": "[3] Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. In Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.387, + 0.484, + 0.442 + ], + "angle": 0, + "content": "[4] Qi Chen, Sihai Tang, Qing Yang, and Song Fu. Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds. In International Conference on Distributed Computing Systems (ICDCS). IEEE, 2019. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.444, + 0.484, + 0.485 + ], + "angle": 0, + "content": "[5] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In CVPR, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.487, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.482, + 0.557 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. In CoRL, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.56, + 0.484, + 0.627 + ], + "angle": 0, + "content": "[8] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks. Journal of machine learning research, 17(59):1-35, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.631, + 0.482, + 0.671 + ], + "angle": 0, + "content": "[9] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.673, + 0.484, + 0.728 + ], + "angle": 0, + "content": "[10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.731, + 0.484, + 0.797 + ], + "angle": 0, + "content": "[11] Cole Gulino, Justin Fu, Wenjie Luo, George Tucker, Eli Bronstein, Yiren Lu, Jean Harb, Xinlei Pan, Yan Wang, Xiangyu Chen, et al. Waymax: An accelerated, data-driven simulator for large-scale autonomous driving research. In NeurIPS, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.801, + 0.484, + 0.828 + ], + "angle": 0, + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[13] Qianjiang Hu, Zhimin Zhang, and Wei Hu. Rangeldm: Fast realistic lidar point cloud generation. In ECCV, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[14] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. In NeurIPS, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.216, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.12 + ], + "angle": 0, + "content": "[15] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In ICLR, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[16] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. In NeurIPS, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[17] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, 2019. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.907, + 0.247 + ], + "angle": 0, + "content": "[18] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[19] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. In NeurlPS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.908, + 0.36 + ], + "angle": 0, + "content": "[20] Yiming Li, Dekun Ma, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: Multi-agent collaborative perception dataset and benchmark for autonomous driving. In Robotics and Automation Letters, pages 10914-10921. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[21] Yiyi Liao, Jun Xie, and Andreas Geiger. Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE TPAMI, 45(3):3292-3310, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[22] Jae Hyun Lim and Jong Chul Ye. Geometric gan. arXiv preprint arXiv:1705.02894, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[23] Yen-Cheng Liu, Junjiao Tian, Nathaniel Glaser, and Zsolt Kira. When2com: Multi-agent perception via communication graph grouping. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.907, + 0.517 + ], + "angle": 0, + "content": "[24] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In IROS, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.52, + 0.907, + 0.559 + ], + "angle": 0, + "content": "[25] Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. Rangenet++: Fast and accurate lidar semantic segmentation. In IROS, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.563, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[26] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.619, + 0.908, + 0.658 + ], + "angle": 0, + "content": "[27] Kazuto Nakashima and Ryo Kurazume. Lidar data synthesis with denoising diffusion probabilistic models. In ICRA, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.662, + 0.908, + 0.701 + ], + "angle": 0, + "content": "[28] Tai-Yu Pan, Qing Liu, Wei-Lun Chao, and Brian Price. Towards open-world segmentation of parts. In CVPR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.908, + 0.771 + ], + "angle": 0, + "content": "[29] Tai-Yu Pan, Chenyang Ma, Tianle Chen, Cheng Perng Phoo, Katie Z Luo, Yurong You, Mark Campbell, Kilian Q Weinberger, Bharath Hariharan, and Wei-Lun Chao. Pre-training lidar-based 3d object detectors through colorization. In ICLR, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[30] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[31] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.907, + 0.899 + ], + "angle": 0, + "content": "[32] Haoxi Ran, Vitor Guizilini, and Yue Wang. Towards realistic scene generation with lidar diffusion models. In CVPR, 2024. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12035" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[33] Zaydoun Yahya Rawashdeh and Zheng Wang. Collaborative automated driving: A machine learning-based method to enhance the accuracy of shared information. In International Conference on Intelligent Transportation Systems (ITSC). IEEE, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.247 + ], + "angle": 0, + "content": "[35] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.249, + 0.483, + 0.275 + ], + "angle": 0, + "content": "[36] T-YLPG Ross and GKHP Dólár. Focal loss for dense object detection. In CVPR, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.278, + 0.484, + 0.319 + ], + "angle": 0, + "content": "[37] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointr-cnn: 3d object proposal generation and detection from point cloud. In CVPR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.321, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.484, + 0.417 + ], + "angle": 0, + "content": "[39] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.421, + 0.484, + 0.461 + ], + "angle": 0, + "content": "[40] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In CVPR, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.463, + 0.482, + 0.49 + ], + "angle": 0, + "content": "[41] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. In NeurIPS, 2017. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.492, + 0.484, + 0.546 + ], + "angle": 0, + "content": "[42] Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In ECCV, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.549, + 0.484, + 0.603 + ], + "angle": 0, + "content": "[43] Bichen Wu, Alvin Wan, Xiangyu Yue, and Kurt Keutzer. Squeezeseg: Convolutional neural nets with recurrent crf for real-time road-object segmentation from 3d lidar point cloud. In ICRA, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.606, + 0.484, + 0.659 + ], + "angle": 0, + "content": "[44] Yang Wu, Kaihua Zhang, Jianjun Qian, Jin Xie, and Jian Yang. Text2lidar: Text-guided lidar point cloud generation via equirectangular transformer. In ECCV. Springer, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.662, + 0.484, + 0.731 + ], + "angle": 0, + "content": "[45] Hao Xiang, Zhaoliang Zheng, Xin Xia, Runsheng Xu, Letian Gao, Zewei Zhou, Xu Han, Xinkai Ji, Mingxi Li, Zonglin Meng, et al. V2x-real: a largs-scale dataset for vehicle-to-everything cooperative perception. arXiv preprint arXiv:2403.16034, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.484, + 0.801 + ], + "angle": 0, + "content": "[46] Pengchuan Xiao, Zhenlei Shao, Steven Hao, Zishuo Zhang, Xiaolin Chai, Judy Jiao, Zesong Li, Jian Wu, Kai Sun, Kun Jiang, et al. Pandaset: Advanced sensor suite dataset for autonomous driving. In International Intelligent Transportation Systems Conference (ITSC). IEEE, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.484, + 0.856 + ], + "angle": 0, + "content": "[47] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In ECCV, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[48] Yuwen Xiong, Wei-Chiu Ma, Jingkang Wang, and Raquel Urtasun. Learning compact representations for lidar completion and generation. In CVPR, 2023. 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.135 + ], + "angle": 0, + "content": "[49] Yuwen Xiong, Wei-Chiu Ma, Jingkang Wang, and Raquel Urtasun. Ultralidar: Learning compact representations for lidar completion and generation. In CVPR, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[50] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In ECCV, 2022. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.907, + 0.247 + ], + "angle": 0, + "content": "[51] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In ICRA, 2022. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.907, + 0.317 + ], + "angle": 0, + "content": "[52] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In CVPR, 2023. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.907, + 0.358 + ], + "angle": 0, + "content": "[53] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.907, + 0.416 + ], + "angle": 0, + "content": "[54] Junbo Yin, Dingfu Zhou, Liangjun Zhang, Jin Fang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Proposal-contrast: Unsupervised pre-training for lidar-based 3d object detection. In ECCV, 2022.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.906, + 0.473 + ], + "angle": 0, + "content": "[55] Jinsu Yoo, Zhenyang Feng, Tai-Yu Pan, Yihong Sun, Cheng Perng Phoo, Xiangyu Chen, Mark Campbell, Kilian Q Weinberger, Bharath Hariharan, and Wei-Lun Chao. Learning 3d perception from others' predictions. In ICLR, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.476, + 0.907, + 0.531 + ], + "angle": 0, + "content": "[56] Yurong You, Katie Z Luo, Xiangyu Chen, Junan Chen, WeiLun Chao, Wen Sun, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Hindsight is 20/20: Leveraging past traversals to aid 3d perception. In ICLR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.532, + 0.907, + 0.599 + ], + "angle": 0, + "content": "[57] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, and Zaiqing Nie. Dair-v2x: A large-scale dataset for vehicleinfrastructure cooperative 3d object detection. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.602, + 0.906, + 0.643 + ], + "angle": 0, + "content": "[58] Chris Zhang, Wenjie Luo, and Raquel Urtasun. Efficient convolutions for real-time semantic segmentation of 3d point clouds. In 3DV, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.906, + 0.673 + ], + "angle": 0, + "content": "[59] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In CVPR, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.907, + 0.715 + ], + "angle": 0, + "content": "[60] Barret Zoph, Golnaz Ghiasi, Tsung-Yi Lin, Yin Cui, Hanxiao Liu, Ekin Dogus Cubuk, and Quoc Le. Rethinking pretraining and self-training. In NeurIPS, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.907, + 0.758 + ], + "angle": 0, + "content": "[61] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In ECCV, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.906, + 0.788 + ], + "angle": 0, + "content": "[62] Vlas Zyrianov, Xiyue Zhu, and Shenlong Wang. Learning to generate realistic lidar point clouds. In ECCV, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.907, + 0.83 + ], + "angle": 0, + "content": "[63] Vlas Zyrianov, Henry Che, Zhijian Liu, and Shenlong Wang. Lidardm: Generative lidar simulation in a generated world. arXiv preprint arXiv:2404.02903, 2024. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.83 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12036" + } + ] +] \ No newline at end of file diff --git a/2025/Transfer Your Perspective_ Controllable 3D Generation from Any Viewpoint in a Driving Scene/0e711492-3ab2-4748-8147-c897ce7d76b8_origin.pdf b/2025/Transfer Your Perspective_ Controllable 3D Generation from Any Viewpoint in a Driving Scene/0e711492-3ab2-4748-8147-c897ce7d76b8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..43d61c93d735014b3217f376ac57c2def72f0c45 --- /dev/null +++ b/2025/Transfer Your Perspective_ Controllable 3D Generation from Any Viewpoint in a Driving Scene/0e711492-3ab2-4748-8147-c897ce7d76b8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13a37352476c9a6f19bdf92aab93b91f4c44f8bb5e7b019d792fecd9d480e374 +size 5492116 diff --git a/2025/Transfer Your Perspective_ Controllable 3D Generation from Any Viewpoint in a Driving Scene/full.md b/2025/Transfer Your Perspective_ Controllable 3D Generation from Any Viewpoint in a Driving Scene/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5c404e4b0c9a80cb60762c972643dd7c49d0e9e5 --- /dev/null +++ b/2025/Transfer Your Perspective_ Controllable 3D Generation from Any Viewpoint in a Driving Scene/full.md @@ -0,0 +1,370 @@ +# Transfer Your Perspective: Controllable 3D Generation from Any Viewpoint in a Driving Scene + +Tai-Yu Pan1, Sooyoung Jeon1, Mengdi Fan1, Jinsu Yoo1, Zhenyang Feng1, Mark Campbell2, Kilian Q. Weinberger2, Bharath Hariharan2, Wei-Lun Chao1 + +1The Ohio State University, 2Cornell University + +![](images/c5eae7588b627f3995606de2f3865121e75062d88ec3559091c23eeb26d94532.jpg) +(a) Ego-centric driving + +![](images/e27f28dbad465dcc1139d5991ac0b7a3ccbfe08dc906e1ba7a7f66b9b36cbfc7.jpg) +(b) TYP for dynamic agent +Figure 1. Illustration of the proposed problem and solution, Transfer Your Perspective (TYP). (a) A given sensory data captured by the ego-car (red triangle). (b) A generated sensory data by TYP, seeing from the viewpoint of another vehicle (green triangle) in the same scene. (c) A generated sensory data, seeing from an imaginary static agent like roadside units (blue icon). (d) Putting all the sensory data together, given or generated, TYP enables the development of collaborative perception with little or no real collaborative driving data. + +![](images/3d13adc16fd89ace3f2f1190193f2e2518d328b72661b1d52aa989107ca99dc0.jpg) +(c) TYP for static agent + +![](images/cd67161cbbb17dfff66d9eb055a4998bbe8690fa8f314a7ce0a6edecc97feb90.jpg) +(d) Collaborative driving + +# Abstract + +Self-driving cars relying solely on ego-centric perception face limitations in sensing, often failing to detect occluded, faraway objects. Collaborative autonomous driving (CAV) seems like a promising direction, but collecting data for development is non-trivial. It requires placing multiple sensor-equipped agents in a real-world driving scene, simultaneously! As such, existing datasets are limited in locations and agents. We introduce a novel surrogate to the rescue, which is to generate realistic perception from different viewpoints in a driving scene, conditioned on a real-world sample—the ego-car's sensory data. This surrogate has huge potential: it could potentially turn any ego-car dataset into a collaborative driving one to scale up the development of CAV. We present the very first solution, using a combination of simulated collaborative data and real ego-car data. Our method Transfer Your Perspective (TYP) learns a conditioned diffusion model whose output samples are not only realistic but also consistent in both semantics and layouts with the given ego-car data. Empirical results demonstrate TYP's effectiveness in aiding in a CAV setting. In particular, TYP enables us to (pre-)train collaborative perception algorithms like early and late fusion with little or no real-world collaborative data, greatly facilitating downstream CAV applications. + +# 1. Introduction + +Seeing the world from an ego-centric perspective, a self-driving car risks being "narrow-sighted," limiting its ability to respond appropriately in dynamic driving environments. For instance, it should slow down or honk if a pedestrian is about to cross the road or if another vehicle is merging into its lane. However, these actions depend on the car's ability to detect traffic participants, which may be occluded by a large bus or a building at a sharp intersection. This occlusion problem can hardly be addressed by mounting a few more sensors on the car. We argue that a self-driving car must move beyond its ego-centric perspective. + +One intuitive way is to collaborate with nearby "agents" that observe the scene from different angles, such as other sensor-equipped cars or static devices like roadside units (RSUs). When trained together to address GPS errors and synchronization delays [55], collaborative perception has been shown to significantly improve the accuracy of each agent's perception, particularly in detecting occluded or distant objects [3, 4, 14, 19, 23, 33, 42, 45, 50-52]. However, collecting training data for collaborative perception is never easy. Unlike single-agent data, which can be collected by simply driving a car on the road, collaborative data requires the simultaneous presence of multiple agents in the same driving scene. For dynamic agents like vehicles, precise coordination is needed to ensure they are within communication + +range. These challenges limit existing works in scale and the number of agents (typically just two). While one may leverage simulated data by game engines [7], they often fail to capture the diversity of real-world scenes. We thus ask: + +# Can we obtain realistic data for learning collaborative perception with much less effort, ideally as easily as single-agent data? + +Specifically, given the vast amount of single-agent LiDAR data already collected across various driving environments, is it possible to convert each of them into collaborative perception data by generating additional point clouds from alternative reference viewpoints within the same scene? + +At first glance, this question may seem overly ambitious for three reasons. First, for collaborative perception to be effective, the generated point cloud must provide information that the ego-car's point cloud cannot capture, such as occluded surfaces invisible to the ego-car. This creates a chicken-and-egg problem: if the ego-car's point cloud is all we initially know about the scene, how can we deduce any additional information? Second, for training purposes, the generated data must be realistic. It should replicate the constraints and data patterns of a real sensor as if it were positioned at the reference viewpoint, producing no points in occluded areas and fewer points in distant areas. Last but not least, in regions commonly perceivable from both the reference and ego-car's viewpoints, the generated data should align in layouts and semantics with the ego-car's data. On the surface, this may seem as simple as copying the ego-car's data, but doing so would violate the second requirement. In essence, the generated data should resemble a point cloud seen from the reference viewpoint, not the ego-car. + +That said, after a deeper look at the question, we argue that it is achievable with three key insights. First, given semantic information around the ego-car for conditioning, existing research has shown promising progress in generating realistic point clouds [13, 27, 32, 44, 49, 62, 63]. This is especially encouraging, as most existing single-agent datasets provide 3D object labels, and they can be translated to obtain semantic information centered around the reference position. This further implies that we can generate occluded object surfaces as seen from the ego-car if they are visible from the reference viewpoint. Second, semantic information is easily editable: even if the translated object map has noticeable empty areas due to limited ego-car perception, object boxes can be manually added to make the map appear more realistic from the reference viewpoint. Lastly, in commonly perceivable regions where the generated data needs to meet two physical constraints, we could leverage simulators based on computational graphics and optics. Specifically, we can use simulated data from the two views to train a conditioned generative model that maps one viewpoint to the other. + +Building upon these insights, we propose Transfer Your Perspective (TYP), a novel research problem and the very + +first solution for generating a realistic point cloud from any viewpoint in a scene, given real ego-car's point cloud and semantic labels as conditions. TYP assumes access to 1) a simulated collaborative driving dataset with multiple agents perceiving the same scenes from different viewpoints and 2) a real single-agent driving dataset; both are labeled. Our solution involves a conditioned latent diffusion model [34] and a dedicated two-stage training process. In the first stage, we consider a single-agent scenario and train the model using real data conditioning only on object locations. This equips the model with the ability to generate diverse and realistic scenes. We denote the learned model by $P(\pmb{x}|\pmb{y})$ , where $\pmb{y}$ stands for the semantic condition and $\pmb{x}$ for the point cloud. In the second stage, we incorporate the simulated data to learn how to ground generation on data from another agent's viewpoint so that the model can produce a semantically consistent reference point cloud given the ego-car's data. We learn a lightweight conditioning module to turn $P(\pmb{x}|\pmb{y})$ into $P(\pmb{x}_r|\pmb{x}_e,\pmb{y}_r)$ , where $e$ and $r$ indicate ego and reference views, respectively. We note that $\pmb{x}_e$ was pre-translated to center around the reference position. + +One challenge in TYP is the domain gap between real and simulated data. Due to differences in sensor configurations and placements, data collection environments, and the sim-to-real gap, the two sets of point clouds inevitably exhibit discrepant distributions, patterns, and densities. To address this, we insert a domain adaptation step between the two training stages. We train a separate encoder-decoder for the simulated data while enforcing a constraint to make the encoded features of simulated and real data indistinguishable [40]. This step allows us to learn $P(\boldsymbol{x}_r|\boldsymbol{x}_e,\boldsymbol{y}_r)$ in a space with reduced domain discrepancy during the second stage. + +Once trained, we pair $P(\pmb{x}_r|\pmb{x}_e,\pmb{y}_r)$ with the real data's encoder-decoder to generate real-style point clouds grounding on real ego-car's data, so that we can develop collaborative perception algorithms without real collaborative data. More specifically, given a real ego-car's perception $\pmb{x}_e$ and label $\pmb{y}_e$ , we first translate them to center around the reference position, followed by optionally injecting object labels into $\pmb{y}_e$ to make it realistic from the reference viewpoint. + +We extensively validate TYP on multiple datasets, all in an offline setting. Empirical results demonstrate TYP's effectiveness in generating high-quality reference data to aid in the development of collaborative perception. In particular, we show that a conditioned diffusion model trained solely on simulated data (e.g., OPV2V [51]) can already turn a real single-agent dataset into a real-alike collaborative one. As such, one can train collaborative perception algorithms for real test data without real training data. We further generate the "ColWaymo" dataset by training TYP on real single-agent Waymo data [39] and simulated OPV2V data [51], thus turning the former into a collaborative one. Collaborative perception backbones pre-trained on the large-scale, semi- + +synthetic ColWaymo data demonstrate a remarkable transferability. They notably improve collaborative perception developed in the downstream tasks (e.g., V2V4Real [52]), even in a many-shot setting. + +In sum, our key contributions are: + +- We propose a new research direction to aid in the development of collaborative perception, generating real-looking sensory data from potentially any viewpoint in a real driving scene, capable of turning a single-agent dataset into a collaborative one. This has the potential to scale up collaborative autonomous driving (CAV). +- We present TYP, a two-stage training recipe with domain adaptation in between to learn the generative model using simulated collaborative data and real single-agent data. +- Extensive experiments demonstrate TYP's effectiveness in aiding in collaborative perception across various scenarios. + +# 2. Related Work + +Collaborative autonomous driving (CAV) offers significant benefits, including extended perception range by sharing sensor information to detect objects beyond a single vehicle's field of view. Despite these benefits, collecting large-scale, real-world datasets for CAV poses significant challenges due to the complexity and cost of deploying multiple instrumented vehicles in diverse environments. Existing datasets often have limitations that restrict the scope of CAV research. For instance, OPV2V [51] and V2X-Sim [20] rely on simulations using the CARLA simulator [7], allowing controlled scenarios with multiple vehicle agents, yet lacking real-world variability. V2V4Real [52], on the other hand, provides real-world data, yet it only contains two collaborating agents, which restricts the exploration of more complex multi-agent interactions. Similarly, the DAIR-V2X [57] dataset also offers real-world data including vehicle-to-infrastructure (V2I) and V2V, but it mainly focuses on V2I scenarios and still has a limited number of collaborating vehicles. To overcome current limitations, we propose a new research direction—generating realistic point clouds given the ego's perspective. 3D generation. Various works have explored 3D scene generation, such as LiDARGen [62], R2DM [27], LiDM [32], UltraLiDAR [48], and RangeLDM [13]. Additionally, Li-darDM [63] and Text2LiDAR [44] investigate conditional scene generation, with the former relying on hand-crafted map layouts and the latter conditioning on text inputs. However, all these methods focus on ego-centric generation. We propose a new research direction for CAV, aiming to generate realistic and consistent scenes conditioned on the ego agent's real point clouds—an area that remains largely unexplored. + +Diffusion models. Diffusion models have recently advanced generative modeling for high-quality LiDAR point clouds and images. Denoising Diffusion Probabilistic Models (DDPMs) [12] outperform traditional Generative Adversarial Networks (GANs) [10] and further enhance efficiency + +![](images/ca03f5ab0b91be56465867e6e1791601e932ceaa1d5f9cc20429fa1f896b143e.jpg) +Figure 2. Illustration of TYP's conditioned generative model and training process. We propose a two-stage training procedure. The first stage maximizes the generation capability by conditioning solely on object locations (using real single-agent target data), while the second stage grounds the generation on the ego-car's perspective to match semantics and layouts (using simulated CAV data). Additionally, we introduce a discriminator to adapt simulated CAV features to the real target domain, making the trained model readily applicable to the target domain after the second stage. + +for LiDAR generation. In autonomous driving applications, methods like RangeLDM [13], LidarDM [63], and LiDAR-Gen [62] apply diffusion models for realistic LiDAR scene generation. In this paper, we utilize its generation capability to study our proposed problem. + +# 3. Transfer Your Perspective (TYP) + +In this paper, we introduce a new research direction to advance collaborative autonomous driving (CAV): generating LiDAR point clouds from different perspectives within the same scene as the ego agent, aiming to reduce the tedious efforts of collecting data for CAV. We begin by defining the proposed problem in Sec. 3.1. In Sec. 3.2, we discuss the representations of the inputs, including point clouds and semantic information. Sec. 3.3 outlines the pipeline developed to address this problem. Finally, in Sec. 3.4, we demonstrate how this capability can be applied to datasets that only have labeled ego agents, e.g., Waymo Open Dataset (WOD) [39]. + +# 3.1. Problem Setup + +Given the perception data $\pmb{x}_e$ from an ego agent, we aim to build a model $P(\pmb{x}_r|\pmb{x}_e)$ to generate new perception data $\pmb{x}_r$ seen from a different reference location and perspective within a communication range. Here, $\pmb{x}$ represents a LiDAR point cloud. This task presents several challenges, including potential information gaps between the two views, the need for alignment within the commonly perceivable area, and ensuring realism in the generated data, as discussed in Sec. 1. + +To address these challenges, we extend the original problem by incorporating semantic information, such as object bounding boxes, represented as $P(\pmb{x}_r|\pmb{x}_e,\pmb{y}_r)$ . We assume the availability of this information around the ego agent, denoted by $\pmb{y}_e$ . For example, most of the existing single-agent + +datasets provide object labels. This semantic information can be easily translated and edited to become $\mathbf{y}_r$ by tools like traffic re-players [11], making it a key to bridge the information gap between the ego and reference views, enabling the reference agent to "see" those objects and surfaces beyond the ego's view—a core concept in CAV. Additionally, by aligning generated point clouds with object locations, this extension facilitates our goal of scaling up CAV development. In essence, the generated $(\mathbf{x}_r, \mathbf{y}_r)$ pairs can be used flexibly alongside $(\mathbf{x}_e, \mathbf{y}_e)$ in various CAV applications, like directly using them to train collaborative perception algorithms. + +In the following sections, we describe how $\pmb{x}$ and $\pmb{y}$ are encoded, followed by our approach to tackling the problem. + +# 3.2. Representations and Embeddings + +LiDAR. There are multiple ways to represent point clouds in continuous 3D space, such as coordinates (i.e., x, y, and z) [30, 31, 37], range images [25, 30, 43], and voxelization [17, 24, 53, 59]. To better align with spatial control from object locations, we follow [49] to voxelize point clouds using a pre-defined grid and record voxel occupancy. [49] also highlights that this representation can naturally handle varying point densities and only minimally impacts LiDAR generation, with some precision trade-offs during voxelization. In short, we represent a point cloud by $\mathbf{x} \in \mathbb{R}^{H \times W \times C}$ . + +To avoid the computational cost of 3D convolutions, we convert $x$ into Bird's-Eye-View (BEV) images by treating the height dimension (i.e., $C$ ) as feature channels for 2D convolutions. This approach has been widely adopted in self-driving perception [5, 48, 58], allowing the use of 2D image-based model architectures and algorithms. + +Semantic information. Given that the point clouds are represented as BEV images, it is also intuitive to represent object locations in BEV. We create binary object maps by considering only the x and y coordinates of 3D bounding boxes, resulting in $\pmb{y} \in \mathbb{R}^{H \times W \times 1}$ . (We could further extend it by including category information.) + +Feature embedding. Following the literature [34], we encode input tensors using a Vector Quantized-Variational Autoencoder (VQ-VAE) [41], comprising an encoder $E$ , a quantization function $Q$ for feature vectors on spatial grids, and a decoder $G$ . Formally, we obtain $\boldsymbol{x}^f = E_x(\boldsymbol{x}) \in \mathbb{R}^{h \times w \times c_x}$ by the encoder, map each $c_x$ -dimensional vector to a learnable code to obtain $\boldsymbol{x}^z = Q_x(\boldsymbol{x}^f) \in \mathbb{R}^{h \times w \times c_x}$ , and generate outputs using the decoder $\hat{\boldsymbol{x}} = G_x(\boldsymbol{x}^z) \in \mathbb{R}^{H \times W \times C}$ , where $h$ and $w$ are spatial resolutions of the feature map downsampled from $H$ and $W$ , and $c_x$ is the number of channels. + +The VQ-VAE model is trained end-to-end by minimizing: + +$$ +\mathcal {L} _ {\mathrm {v q}} = \mathcal {L} _ {\mathrm {r e c}} + \| \operatorname {s g} [ E _ {x} (\boldsymbol {x}) ] - \boldsymbol {x} ^ {z} \| _ {2} ^ {2} + \| \operatorname {s g} [ \boldsymbol {x} ^ {z} ] - E _ {x} (\boldsymbol {x}) \| _ {2} ^ {2}, \tag {1} +$$ + +where $\mathrm{sg}[\cdot ]$ and $\mathcal{L}_{\mathrm{rec}}$ denote the stop-gradient operation and the reconstruction loss, respectively. As our point cloud representation is binary occupancy $\pmb {x}\in \{0,1\}^{H\times W\times C}$ , binary + +cross-entropy is a natural choice for $\mathcal{L}_{\mathrm{rec}}$ . However, due to the sparsity of point clouds, this results in an imbalanced loss. To address this, we adopt the Focal Loss (FL) [36]: + +$$ +\ell_ {\mathrm {F L}} \left(\boldsymbol {x} _ {i}, \hat {\boldsymbol {x}} _ {i}\right) = \left\{ \begin{array}{l l} - (1 - \hat {\boldsymbol {x}} _ {i}) ^ {\gamma} \log \left(\hat {\boldsymbol {x}} _ {i}\right) & \text {i f} \boldsymbol {x} _ {i} = 1 \\ - \hat {\boldsymbol {x}} _ {i} ^ {\gamma} \log \left(1 - \hat {\boldsymbol {x}} _ {i}\right) & \text {o t h e r w i s e ,} \end{array} \right. \tag {2} +$$ + +$$ +\mathcal {L} _ {\mathrm {r e c}} = \sum_ {i = 0} ^ {M} \ell_ {\mathrm {F L}} \left(\boldsymbol {x} _ {i}, \hat {\boldsymbol {x}} _ {i}\right), \tag {3} +$$ + +where $i$ is the voxel index and $M$ is the number of voxels. + +For object locations $\pmb{y}$ , we use the same approach to train a separate VQ-VAE model $E_{y}, Q_{y}$ , and $G_{y}$ . In the subsequent subsections, we use $\pmb{x}^{f} \in \mathbb{R}^{h \times w \times c_{x}}$ and $\pmb{y}^{f} \in \mathbb{R}^{h \times w \times c_{y}}$ as feature embeddings for learning a latent diffusion model. + +# 3.3. Transfer Your Perspective by Generation + +We model the distribution $P(\pmb{x}_r | \pmb{x}_e, \pmb{y}_r)$ defined in Sec. 3.1 by a conditioned generative model and train it in two stages. We assume access to $(\pmb{x}_r, \pmb{x}_e, \pmb{y}_r)$ tuples as training data. We detail the training data preparation in Sec. 3.4 and Sec. 4. + +In the first stage, we aim to maximize the generation capability by providing minimal conditions, i.e., only bounding boxes, without further constraints on the scene. This encourages the model to produce $P(\boldsymbol{x}|\boldsymbol{y})$ with high flexibility. Second, building on this model, we incorporate the ego agent's point cloud as an additional cue and ground the generation process on it to ensure semantic and layout consistency between the generated and ego agent's point clouds. Put together, this two-stage training procedure enables the learned $P(\boldsymbol{x}_r|\boldsymbol{x}_e,\boldsymbol{y}_r)$ model to generate perception data at areas beyond the commonly visible regions between two views. The model and training process is illustrated in Fig. 2. + +More importantly, this approach enables us to leverage existing labeled single-agent datasets such as KITTI [9], NuScenes [2], and Waymo Open Dataset (WOD) [39], since the first-stage training only requires ego-centric LiDAR point clouds and their corresponding semantic information (see Sec. 3.4). The $(\pmb{x}_e,\pmb{x}_r)$ data pairs seen from the ego and the reference agents are needed only in the second stage. + +In the following, we elaborate on each training stage. + +Stage 1: Generation with semantic information. The goal of this stage is to equip the model with strong generation capabilities of point clouds given spatial conditions, preparing it for the next stage. As discussed in Sec. 3.2, the point clouds and object locations are embedded by VQVAEs [41], denoted as $\boldsymbol{x}^f = E_x(\boldsymbol{x}) \in \mathbb{R}^{h \times w \times c_x}$ and $\boldsymbol{y}^f = E_x(\boldsymbol{y}) \in \mathbb{R}^{h \times w \times c_y}$ . The image-like feature maps allow us to adopt existing generation algorithms for 2D images [6, 12, 16]. In this paper, we apply one of the most popular generative models, the Latent Diffusion Model (LDM) [34], for conditioned generation $P(\boldsymbol{x} | \boldsymbol{y})$ . LDM seeks to model a data distribution by iteratively denoising variables + +that are initially sampled from a Gaussian. The objective is: + +$$ +\mathcal {L} _ {\mathrm {L D M}} = \mathbb {E} _ {\boldsymbol {x} ^ {f}, \boldsymbol {y} ^ {f}, \epsilon \sim \mathcal {N} (0, 1), t} \left[ \left\| \epsilon - \epsilon_ {\theta} \left(\boldsymbol {x} _ {t} ^ {f}, \boldsymbol {y} ^ {f}, t\right) \right\| _ {2} ^ {2} \right], \tag {4} +$$ + +where, $\epsilon_{\theta}$ is a UNet [35] backbone and $t$ is the timestamp. + +Stage 2: Generation for new perspectives. In this stage, we aim to ground the generation of point clouds by incorporating the ego agent's perception. To retain the generation capability from stage 1, we freeze the learned $P(\boldsymbol{x}|\boldsymbol{y})$ and introduce a learnable lightweight control module to inject additional cues, following T2I-Adapter [26], as shown in Fig. 2. + +Formally, let $\pmb{x}_{e'}$ denote the translated and rotated egoagent's point cloud centered around the reference location and aligned with the reference orientation, as shown in Fig. 2; $\pmb{x}_{e'}^f = E_x(\pmb{x}_{e'}) \in \mathbb{R}^{h \times w \times c_x}$ is the corresponding embedding. Let $F_{AD}$ denote the learnable control module, which takes $\pmb{x}_{e'}^f$ as input and outputs + +$$ +\mathbf {F} _ {c} = F _ {A D} \left(\boldsymbol {x} _ {e ^ {\prime}} ^ {f}\right), \tag {5} +$$ + +where $\mathbf{F}_c = \{\mathbf{F}_c^1,\mathbf{F}_c^2,\mathbf{F}_c^3,\mathbf{F}_c^4\}$ matches the size of the multiscale features $\mathbf{F}_{enc} = \{\mathbf{F}_{enc}^{1},\mathbf{F}_{enc}^{2},\mathbf{F}_{enc}^{3},\mathbf{F}_{enc}^{4}\}$ extracted from the frozen encoder of the UNet $\epsilon_{\theta}$ . With these ingredients, T2I-Adapter influences the generation process by + +$$ +\mathbf {F} _ {e n c} ^ {i} \leftarrow \mathbf {F} _ {e n c} ^ {i} + \mathbf {F} _ {c} ^ {i}, i \in \{1, 2, 3, 4 \}, \tag {6} +$$ + +which injects $\pmb{x}_{e'}^f$ embedding into each of the denoising steps. The second-stage objective for learning $F_{AD}$ is: + +$$ +\mathcal {L} _ {\mathrm {L D M}} = \mathbb {E} _ {\boldsymbol {x} _ {r} ^ {f}, \boldsymbol {y} _ {r} ^ {f}, \boldsymbol {x} _ {e ^ {\prime}} ^ {f}, \epsilon \sim \mathcal {N} (0, 1), t} \left[ \left\| \epsilon - \epsilon_ {\theta} \left(\boldsymbol {x} _ {r, t} ^ {f}, \boldsymbol {y} ^ {f}, \boldsymbol {x} _ {e ^ {\prime}} ^ {f}, t\right) \right\| _ {2} ^ {2} \right], \tag {7} +$$ + +where the encoder of $\epsilon_{\theta}$ is modified as defined in Eq. (6). + +# 3.4. From Single Agent to Collaborative Datasets + +There are many publicly accessible large-scale real-world autonomous driving datasets for ego-centric 3D perception, like WOD [39], NuScenes [2], KITTI-360 [21], and PandaSet [46], whereas much fewer datasets exist for CAV due to the challenges in data collection as discussed in Sec. 1. + +This limitation inspired a bold idea: Can we transfer the generation capability learned from paired CAV data to these existing labeled single-agent datasets? Achieving this would scale up CAV datasets substantially, offering immense benefits to the research community. However, this goal is challenging due to domain gaps between datasets, such as variations in LiDAR sensors, point cloud densities, data patterns, and data collection environments. Below, we describe our approach to addressing these challenges. + +Problem setup. Our target domain is a single-agent dataset, while the source domain consists of paired perception and semantic information from ego and reference perspectives, denoted as $(\pmb{x}_r,\pmb{x}_e,\pmb{y}_r)$ . We demonstrate that we can transfer + +knowledge from a simulated CAV dataset (e.g., OPV2V [51]) to a real-world ego-centric dataset (e.g., WOD [39]). + +Two-Stage training is the key. As described in Sec. 3.3, we decompose the generative training into two stages: first learning $P(\pmb{x}|\pmb{y})$ , followed by $P(\pmb{x}_r|\pmb{x}_e,\pmb{y}_r)$ . The second stage keeps $P(\pmb{x}|\pmb{y})$ frozen by adding an adapter. This strategy enables us to use target domain data in the first stage. The resulting $P(\pmb{x}|\pmb{y})$ would generate target-like point clouds. + +Minimizing domain gaps. In the first stage, we have learned VQ-VAE encoder-decoders and $P(\boldsymbol{x}|\boldsymbol{y})$ in the target domain (single-agent). In the second stage, we bring in CAV data to guide the generation on how to condition on ego-agent's perception. To minimize the gap of $\boldsymbol{x}^f$ between two domains— $\boldsymbol{x}^f$ is what the generative model is optimized to generate—we adopt a GAN-style discriminator [8, 10, 22, 40]. Specifically, we aim to make feature embeddings extracted from the source data and the target data indistinguishable; namely, to confuse a discriminator $D$ trained to differentiate them. We freeze all parameters of the target domain's encoder $E_{tgt}$ , codebook $Q_{tgt}$ , and decoder $G_{tgt}$ in the VQ-VAE. For the source domain, we initialize its $E_{src}$ and $G_{src}$ by $E_{tgt}$ and $G_{tgt}$ while reusing the frozen codebook $Q_{tgt}$ . We then learn the discriminator $D$ and the source domain's encoder $E_{src}$ and decoder $G_{src}$ in an interleaving fashion by minimizing the following objective functions: + +$$ +\begin{array}{r l} \text {F o r} D: \mathcal {L} _ {D} = & \mathbb {E} _ {\boldsymbol {x} \sim P _ {t g t}} [ \max (0, 1 - D (E _ {t g t} (\boldsymbol {x}))) ] \\ & + \mathbb {E} _ {\boldsymbol {x} \sim P _ {s r c}} [ \max (0, 1 + D (E _ {s r c} (\boldsymbol {x}))) ]; \end{array} \tag {8} +$$ + +$$ +\text {F o r V Q - V A E :} \mathcal {L} _ {\mathrm {v q}} = \mathcal {L} _ {\mathrm {r e c}} + \mathcal {L} _ {s r c} + \| \operatorname {s g} [ \boldsymbol {x} ^ {z} ] - E _ {s r c} (\boldsymbol {x}) \| _ {2} ^ {2}, \tag {9} +$$ + +$$ +\text {w h e r e} \mathcal {L} _ {s r c} = - \mathbb {E} _ {\boldsymbol {x} \sim P _ {s r c}} D \left(E _ {s r c} (\boldsymbol {x})\right). \tag {10} +$$ + +The other terms in Eq. (9) follow those in Eq. (1), where for $\pmb{x} \sim P_{src}$ , $\hat{\pmb{x}} = G_{src} \circ Q_{tgt} \circ E_{src}(\pmb{x})$ in $\mathcal{L}_{rec}$ . We note that $\mathcal{L}_D$ takes in two streams of data (i.e., source and target) and is defined by the Hinge Loss [22] on $\pmb{x}^f$ ; $\mathcal{L}_{\mathrm{vq}}$ takes in only the source CAV data. We also note that embeddings for object locations does not require adaptation. + +After this adaptation step, we proceed into the second stage of generative training, where we drop the discriminator and train the control module defined in Eq. (5) by minimizing Eq. (7). The feature embeddings are produced by $E_{src}$ on top of $(\pmb{x}_r, \pmb{x}_e, \pmb{y}_r) \sim P_{src}$ . + +Enhancement in target domains. With the discriminator and adversarial training, we reduce the gap between source and target domain data, successfully adapting TYP to the single-agent dataset, as shown in Sec. 4.5. To further improve the conditioned generation quality in the target domain, a further fine-tuning stage in it is desired. However, how can we do so without having reference agents' point clouds in the target domain? Here, we propose two simple solutions. + +
MethodTrain DataAP 0.5
smlall
No Fusionego's gt only0.670.410.130.40
Early Fusion [4]gt (oracle)0.760.420.310.49
TYP (ours)0.750.380.290.46
Late Fusion [51]gt (oracle)0.740.570.360.55
TYP (ours)0.710.490.320.50
AttFuse [51]gt (oracle)0.940.800.620.77
TYP (ours)0.900.730.560.72
V2X-ViT [50]gt (oracle)0.870.710.500.71
TYP (ours)0.840.650.400.65
+ +![](images/259fc6e51134ec81f260d71466fd8d7df89b8d55a1d9c6dca094290ed2cb6246.jpg) +(a) no adaptation +Figure 3. Qualitative results on enhancement in the target domain. Generated point cloud (green) has better quality with the enhancement given ego (gray) from Waymo (cf. Sec. 3.4). + +![](images/51fdfe250768ef030b1013c443464af5359bfc7cbf9bea5b34c62fac5d788731.jpg) +(b) disc. + +![](images/e3b62f81cc90d270c5388e87d359101ffd43b7f09e619da9d8210b2dd9fd2f5f.jpg) +(c) disc. + enhancement + +Recall that in training $P(\pmb{x}_r|\pmb{x}_e,\pmb{y}_r)$ using the source CAV data, point clouds show up in both the input as conditions and the outputs as supervisions. In other words, the domain gap exists on both sides, and we aim to reduce it by further fine-tuning using the target single-agent dataset. First, to enhance the output generation in the target domain, we set ego's point clouds to empty, i.e., $\pmb{x}_e = 0^{H\times W\times C}$ for $(\pmb{x}_r,\pmb{y}_r)\sim P_{tgt}$ . Then, to enable adaptation on inputs, we apply self-training [18, 28, 60, 61], randomly sampling a reference location in the scene to generate pseudo-point clouds for fine-tuning, i.e., $(\hat{\pmb{x}}_r,\pmb{x}_e,\pmb{y}_r)$ for $(\pmb{x}_e,\pmb{y}_r)\sim P_{tgt}$ . We use the above target domain data and the source domain data to jointly fine-tune the control module (cf. Sec. 3.3, Eq. (6), and Eq. (7)) further. Empirical results in Sec. 4.5 show the effectiveness of the proposed solution in turning single-agent datasets into collaborative datasets for CAV development. + +# 4. Experiments + +# 4.1. Setup + +Datasets. In this study, we primarily use three datasets: OPV2V [51], V2V4Real [52], and WOD [39]. OPV2V is a simulation-based CAV dataset that includes over 70 diverse driving scenes and more than 11,000 frames. We utilize this dataset initially to validate our proposed research problem and later to guide the generation training on single-agent datasets. V2V4Real is a real-world CAV dataset with approximately 20,000 LiDAR frames, which we use to validate TYP in a real-world CAV context, treating it as a single-agent dataset (i.e., without using any labels to train TYP). Lastly, we adopt WOD to scale up TYP, one of the largest real-world single-agent datasets available. + +Evaluation metrics. Following standard benchmarks [51, + +Table 1. Results on OPV2V. We compare the performance using reference generated by TYP versus ground-truth (oracle). The comparable results demonstrate the quality of our generated data. + +
MethodTrain DataAP 0.5
smlall
No Fusionego's gt only0.710.330.080.46
Early Fusion [4]gt (oracle)0.790.420.450.60
TYP (ours)0.760.360.300.53
Late Fusion [51]gt (oracle)0.790.530.560.67
TYP (ours)0.770.480.510.63
AttFuse [51]gt (oracle)0.800.450.360.61
TYP (ours)0.750.450.260.56
V2X-ViT [50]gt (oracle)0.810.490.300.61
TYP (ours)0.760.380.170.48
+ +Table 2. Results on V2V4Real. We compare the performance using reference generated by TYP versus ground-truth (oracle). Note that we do not use V2V4Real data to train the generation, treating it as a single-agent dataset. The comparable results demonstrate the quality of our generated data and the transferability of TYP. + +52], we report Average Precision (AP 0.5) across three distance ranges: 0-30 meters (short, denoted as s), 30-50 meters (medium, m), and 50-100 meters (long, l) and overall (all). Implementations. For TYP training, we voxelize point clouds within the range $[-51.2, 51.2]$ meters in x and y dimensions and 4 meters in z (range is dataset-dependent) to create 3D volumes, $x \in \mathbb{R}^{512 \times 512 \times 20}$ . This is then encoded to $x^f \in \mathbb{R}^{64 \times 64 \times 8}$ with VQ-VAE [41]. Similarly, semantic information $y \in \mathbb{R}^{512 \times 512 \times 1}$ is encoded to $y^f \in \mathbb{R}^{64 \times 64 \times 3}$ . We train VQ-VAE for $x$ 120 epochs and for $y$ 10 epochs, with a batch size of 6 per GPU. Next, we train LDM with a batch size of 48 per GPU for 200 epochs in the first stage and 80 epochs in the second stage. During the second stage, we freeze the U-Net from the first stage and fine-tune only the control module, following [26]. For adaptation, we initialize the weights of encoder-decoder for CAV data (fake) with the model trained on single-agent data (real), freeze its codebook, and fine-tune it with a discriminator comprising 4 convolutional layers. The model processes an equal amount of fake and real per batch (i.e., 3 samples each). All training is conducted on 4 NVIDIA H100 GPUs. For experiments on benchmarks, we follow the convention to report results with Pointpillars [17]. We train / fine-tune all models with 8 P100, with batch size 4 per GPU. For V2X-ViT, we use 4 H100 due to resources required. + +Inference of generation and post-processing. We use DDIM scheduler [38] for 200 steps. For post-processing, we apply modified Gumbel-Softmax [15] to the logits. Specifically, we define two thresholds, low and high. Scores below the low threshold are dropped, those above the high threshold are retained, and values in between are perturbed with Gaussian noise on logits. Hyperparameters are tuned on a small hold-out set. Finally, we convert the volume back to coordinates by using the centers of the occupied voxels. + +# 4.2. Do We Need Collaborative Perception? + +To explore this question, we consistently report the performance without collaborative agents (denoted as "ego's gt + +
MethodPre-TrainFT. on 0 SceneFT. on 5 SceneFT. on 10 SceneFT. on 32 Scene
smlallsmlallsmlallsmlall
No Fusionego's gt only0.440.160.060.290.660.290.080.420.710.340.080.47
scratch0.440.180.160.300.730.280.240.480.790.420.450.60
Early Fusion [4]OPV2V0.540.200.070.310.320.180.150.250.730.320.310.500.800.470.540.65
ColWaymo (ours)0.500.240.240.350.680.340.320.510.780.390.330.570.830.480.500.65
Late Fusion [51]scratch0.440.240.330.340.720.420.510.580.790.530.560.67
OPV2V0.600.270.280.440.550.260.380.420.730.420.510.590.780.550.580.67
ColWaymo (ours)0.400.190.150.250.600.280.440.470.760.430.510.610.820.580.610.71
AttFuse [51]scratch0.400.150.100.280.700.300.170.470.800.450.360.61
OPV2V0.510.190.050.310.540.220.110.370.770.400.210.540.830.530.400.65
ColWaymo (ours)0.660.350.110.450.650.290.160.460.830.460.330.610.880.580.530.72
V2X-ViT [50]scratch0.430.150.120.310.700.280.170.430.810.490.300.61
OPV2V0.510.240.070.330.480.230.160.350.760.380.220.530.810.490.350.61
ColWaymo (ours)0.600.280.100.340.660.280.220.460.790.480.260.580.840.570.440.67
+ +Table 3. Results with pre-training on collaborative Waymo. We scale up TYP with exiting large-scale single-agent dataset WOD [39], creating its collaborative version "ColWaymo". These data are used for pre-training and subsequently fine-tuned on V2V4Real [52]. Results show significant improvement over training from scratch, highlighting the potential to reduce data collection efforts, scale up, and accelerate CAV development. Comparisons to simulated OPV2V [51] pre-training further demonstrate the realism of our generated point clouds. + +
MethodDisc.ST.Dum.JSD (↓)MMD (↓)
No Adaptation0.264.61 × 10-4
✓0.161.10 × 10-4
Adaptation✓✓0.171.30 × 10-4
✓✓✓0.161.17 × 10-4
+ +Table 4. Ablation of proposed adaption method. Results with adaptation outperform no adaptation, demonstrating an improvement in generation quality. Abbreviations in title refer to discriminator, self-training, and dummy ego, respectively (cf. Sec. 3.4). + +
Num. StepsEpochs 1stEpochs 2ndJSD (↓)MMD (↓)
1-2800.116.45 × 10-5
2200800.105.28 × 10-5
+ +Table 5. 2-Stage training. Results demonstrate the benefit of the proposed 2-stage training in generation quality on OPV2V [51]. + +only" in Tab. 1, Tab. 2, and Tab. 3). For example, in the V2V4Real dataset, a well-trained detector struggles with objects beyond 50 meters, achieving only 0.08 AP at that range, though it performs well on nearby objects with an AP of 0.71 (cf. first row of Tab. 3). simply combining the predictions from two under-trained detectors, we observe a notable boost in performance to 0.33 (cf. "scratch" in late fusion, Tab. 2), suggesting the significance of CAV. + +# 4.3. Point Clouds Generation on OPV2V + +Setting. To validate our research concept, we begin with the simulated dataset OPV2V [51]. We split the original training set of 44 scenes into two halves, using only the first 22 scenes to train our generation model. For the remaining 22 scenes, we generate point clouds for reference agents based on the ego agent's point clouds and compare these results to ground-truth point clouds. We follow the benchmark to report performance on the validation set from Culver City. + +Baselines. Following the literature [51, 52], we adopt three common fusion algorithms for CAV: early, late, and inter + +mediate fusion, which fuse input point clouds, extracted features, and predicted bounding boxes, respectively. For intermediate fusion, we adipot AttFuse [51] and V2X-VIT [50]. + +Comparison to using ground-truth LiDAR. In Tab. 1, we show that training CAV with point clouds generated by TYP achieves performance comparable to that obtained with ground-truth (oracle) across all methods (e.g., 0.46 vs. 0.49 with early, 0.50 vs. 0.55 with late, and 0.77 vs. 0.72 with AttFuse [51], etc.) This result highlights the quality of the generated point clouds. In the supplemental material, we further demonstrate that with additional labeled data (from the first half of the split used for training generation), the gap to oracle performance can be further minimized. + +# 4.4. Single-Agent to Collaborative Multi-Agent + +Setting. Building on our findings with simulated data (cf. Sec. 4.3), we validate our approach on the real-world dataset V2V4Real [52], treating it as a single-agent dataset without access to reference agents' point clouds. Specifically, we use the model from Sec. 4.3 to generate reference car's point clouds from the ego's perspective and compare performance against ground-truth point clouds (oracle). Note that no V2V4Real data is used to train the generation. + +Results on V2V4Real. In Tab. 2, we show the performance using generated reference point clouds is comparable to the oracle, supporting our idea of translating a single-agent dataset to multi-agent CAV. Xu et al. previously explored OPV2V-to-V2V4Real adaptation, noting significant performance drops even with adaptation algorithms. Here, TYP offers a new solution to this challenge – in a generative way, providing a fresh perspective on domain adaptation. + +# 4.5. ColWaymo: Scaling with Large-Scale Data + +Scaling CAV to the next level. In Sec. 4.4, we show the potential to create realistic reference's point clouds on real- + +![](images/56023c3d1150e71dbc23b6a087a38fddfb7fd066bd77fbb91652e7a64981f6d8.jpg) +Figure 4. Qualitative results on Collaborative Waymo. The gray point clouds are from the original single-agent dataset and the green are generated by TYP conditioning on them. + +![](images/752f1bd8c563280c053b82f26f676cbea0dd1f380c2b393c2bcd1f9f8cecf506.jpg) + +![](images/1b05b0743a4f04bdbd509836ecb1bdee5403a67e91a68fbc2f14b5c2c97f81a0.jpg) + +![](images/03dfc23b41211493196497bebcb37145a7ac996d23cf359325eacf6e54c73350.jpg) + +![](images/4abe5b5c88316aacf10b2ee4382db43d2c14c062fb5f9619bde0bf0fee779bc8.jpg) + +![](images/faff2f785d98135f357ec24971c54b3770e890869d48333deb10e417e9116c2e.jpg) + +![](images/65428d9475507f7325174b44ae1fc7cb81bec176637a2fc906af87f5a3d0feef.jpg) + +![](images/061b00aca9e01923d1aa59b2808ff613fb4d4e254bdd9a12fc66bcc777bbdfdd.jpg) +Figure 5. Qualitative results. Our proposed TYP is capable of scene editing, by inputting the same point cloud but different object locations. We (b) remove and (c) add a car from (a) the original point cloud. Inspired by the idea of past traversals [56], we apply completely different traffic conditions and generate (d), (e), and (f), to imagine driving through the same intersection. + +![](images/ee06003fcf8d90f7ea437d7a4a6f313fb2f5c5a67ed47b563e6e0b56a2a2623d.jpg) + +![](images/b17e2e0fafe06efd15d5aa42ba408d318b3b5301ad4daa26d3932eb5d590cde6.jpg) + +world dataset V2V4Real [52], through a simulation-guided training of TYP. Here, our goal is broader: to transform an existing large-scale single-agent dataset (e.g., WOD [39]) into a collaborative version. Using the pipeline described in Sec. 3 and illustrated in Fig. 2, we train the generative model on 17,400 samples from the WOD training set. We then use its validation set to generate approximately 36,000 collaborative samples by sampling reference locations from labeled vehicles. These generated samples serve to pre-train a model, which we subsequently fine-tune on V2V4Real [52]. For comparison, we also conduct pre-training on OPV2V [51]. + +Results of fine-tuning on V2V4Real. Following prior work [1, 29, 47, 54], we perform pre-training followed by fine-tuning on limited labeled data to validate the potential of TYP for scaling up CAV development. As shown in Tab. 3, pre-training on our generated data yields significantly better performance than training from scratch. Additionally, compared to pre-training on the simulated dataset OPV2V, ColWaymo consistently achieves higher performance, highlighting the realism of the generated point clouds. These results demonstrate that TYP effectively transforms an ego-only dataset into a collaborative version, and potentially reducing the need for extensive CAV data collection. + +Generation quality by proposed adaptation. To assess the quality of generated data, we use two common metrics: + +Jensen-Shannon Divergence (JSD) and Maximum Mean Discrepancy (MMD). As shown in Tab. 4, our adaptation method improves JSD from 0.26 to 0.16 and MMD from 4.61 to 1.17 compared to the baseline without adaptation. Additionally, Fig. 3 visually illustrates this improvement. + +# 4.6. Qualitative Results + +We illustrate the proposed problem in Fig. 1, which highlights the benefits of collaboration by allowing an agent to "see" beyond the limitations of its own sensors. TYP generates realistic point clouds from various viewpoints, enabling the agents to perceive previously occluded objects while preserving the overall scene semantics. Fig. 4 shows our ColWaymo, and Fig. 5 further demonstrates the model's capability for scene editing. + +# 5. Conclusion + +We introduce a new research direction in collaborative driving and present the first solution. Empirical results show our approach can significantly reduce data collection and development efforts, advancing safer autonomous systems. + +Limitations and future work. Following existing benchmarks [51, 52], TYP focuses on vehicle-like objects. Future work could extend it to broader objects and static entities (e.g., traffic signs, signals) essential for real-world traffic. + +# Acknowledgment + +This research is supported in part by grants from the National Science Foundation (IIS-2107077, IIS-2107161). We are thankful for the generous support of the computational resources by the Ohio Supercomputer Center. + +# References + +[1] Alexandre Boulch, Corentin Sautier, Björn Michele, Gilles Puy, and Renaud Marlet. Also: Automotive lidar self-supervision by occupancy estimation. In CVPR, 2023. 8 +[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 4, 5 +[3] Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. In Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, 2019. 1 +[4] Qi Chen, Sihai Tang, Qing Yang, and Song Fu. Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds. In International Conference on Distributed Computing Systems (ICDCS). IEEE, 2019. 1, 6, 7 +[5] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In CVPR, 2017. 4 +[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 4 +[7] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. In CoRL, 2017. 2, 3 +[8] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks. Journal of machine learning research, 17(59):1-35, 2016. 5 +[9] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 4 +[10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 3, 5 +[11] Cole Gulino, Justin Fu, Wenjie Luo, George Tucker, Eli Bronstein, Yiren Lu, Jean Harb, Xinlei Pan, Yan Wang, Xiangyu Chen, et al. Waymax: An accelerated, data-driven simulator for large-scale autonomous driving research. In NeurIPS, 2024. 4 +[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3, 4 +[13] Qianjiang Hu, Zhimin Zhang, and Wei Hu. Rangeldm: Fast realistic lidar point cloud generation. In ECCV, 2025. 2, 3 +[14] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. In NeurIPS, 2022. 1 + +[15] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In ICLR, 2017. 6 +[16] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. In NeurIPS, 2021. 4 +[17] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, 2019. 4, 6 +[18] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, 2013. 6 +[19] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. In NeurlPS, 2021. 1 +[20] Yiming Li, Dekun Ma, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: Multi-agent collaborative perception dataset and benchmark for autonomous driving. In Robotics and Automation Letters, pages 10914-10921. IEEE, 2022. 3 +[21] Yiyi Liao, Jun Xie, and Andreas Geiger. Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE TPAMI, 45(3):3292-3310, 2022. 5 +[22] Jae Hyun Lim and Jong Chul Ye. Geometric gan. arXiv preprint arXiv:1705.02894, 2017. 5 +[23] Yen-Cheng Liu, Junjiao Tian, Nathaniel Glaser, and Zsolt Kira. When2com: Multi-agent perception via communication graph grouping. In CVPR, 2020. 1 +[24] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In IROS, 2015. 4 +[25] Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. Rangenet++: Fast and accurate lidar semantic segmentation. In IROS, 2019. 4 +[26] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, 2024. 5, 6 +[27] Kazuto Nakashima and Ryo Kurazume. Lidar data synthesis with denoising diffusion probabilistic models. In ICRA, 2024. 2, 3 +[28] Tai-Yu Pan, Qing Liu, Wei-Lun Chao, and Brian Price. Towards open-world segmentation of parts. In CVPR, 2023. 6 +[29] Tai-Yu Pan, Chenyang Ma, Tianle Chen, Cheng Perng Phoo, Katie Z Luo, Yurong You, Mark Campbell, Kilian Q Weinberger, Bharath Hariharan, and Wei-Lun Chao. Pre-training lidar-based 3d object detectors through colorization. In ICLR, 2023. 8 +[30] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 4 +[31] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 4 +[32] Haoxi Ran, Vitor Guizilini, and Yue Wang. Towards realistic scene generation with lidar diffusion models. In CVPR, 2024. 2, 3 + +[33] Zaydoun Yahya Rawashdeh and Zheng Wang. Collaborative automated driving: A machine learning-based method to enhance the accuracy of shared information. In International Conference on Intelligent Transportation Systems (ITSC). IEEE, 2018. 1 +[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 4 +[35] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 5 +[36] T-YLPG Ross and GKHP Dólár. Focal loss for dense object detection. In CVPR, 2017. 4 +[37] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointr-cnn: 3d object proposal generation and detection from point cloud. In CVPR, 2019. 4 +[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 6 +[39] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 2, 3, 4, 5, 6, 7, 8 +[40] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In CVPR, 2017. 2, 5 +[41] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. In NeurIPS, 2017. 4, 6 +[42] Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In ECCV, 2020. 1 +[43] Bichen Wu, Alvin Wan, Xiangyu Yue, and Kurt Keutzer. Squeezeseg: Convolutional neural nets with recurrent crf for real-time road-object segmentation from 3d lidar point cloud. In ICRA, 2018. 4 +[44] Yang Wu, Kaihua Zhang, Jianjun Qian, Jin Xie, and Jian Yang. Text2lidar: Text-guided lidar point cloud generation via equirectangular transformer. In ECCV. Springer, 2024. 2, 3 +[45] Hao Xiang, Zhaoliang Zheng, Xin Xia, Runsheng Xu, Letian Gao, Zewei Zhou, Xu Han, Xinkai Ji, Mingxi Li, Zonglin Meng, et al. V2x-real: a largs-scale dataset for vehicle-to-everything cooperative perception. arXiv preprint arXiv:2403.16034, 2024. 1 +[46] Pengchuan Xiao, Zhenlei Shao, Steven Hao, Zishuo Zhang, Xiaolin Chai, Judy Jiao, Zesong Li, Jian Wu, Kai Sun, Kun Jiang, et al. Pandaset: Advanced sensor suite dataset for autonomous driving. In International Intelligent Transportation Systems Conference (ITSC). IEEE, 2021. 5 +[47] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In ECCV, 2020. 8 +[48] Yuwen Xiong, Wei-Chiu Ma, Jingkang Wang, and Raquel Urtasun. Learning compact representations for lidar completion and generation. In CVPR, 2023. 3, 4 + +[49] Yuwen Xiong, Wei-Chiu Ma, Jingkang Wang, and Raquel Urtasun. Ultralidar: Learning compact representations for lidar completion and generation. In CVPR, 2023. 2, 4 +[50] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In ECCV, 2022. 1, 6, 7 +[51] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In ICRA, 2022. 2, 3, 5, 6, 7, 8 +[52] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In CVPR, 2023. 1, 3, 6, 7, 8 +[53] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 4 +[54] Junbo Yin, Dingfu Zhou, Liangjun Zhang, Jin Fang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Proposal-contrast: Unsupervised pre-training for lidar-based 3d object detection. In ECCV, 2022.8 +[55] Jinsu Yoo, Zhenyang Feng, Tai-Yu Pan, Yihong Sun, Cheng Perng Phoo, Xiangyu Chen, Mark Campbell, Kilian Q Weinberger, Bharath Hariharan, and Wei-Lun Chao. Learning 3d perception from others' predictions. In ICLR, 2025. 1 +[56] Yurong You, Katie Z Luo, Xiangyu Chen, Junan Chen, WeiLun Chao, Wen Sun, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Hindsight is 20/20: Leveraging past traversals to aid 3d perception. In ICLR, 2022. 8 +[57] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, and Zaiqing Nie. Dair-v2x: A large-scale dataset for vehicleinfrastructure cooperative 3d object detection. In CVPR, 2022. 3 +[58] Chris Zhang, Wenjie Luo, and Raquel Urtasun. Efficient convolutions for real-time semantic segmentation of 3d point clouds. In 3DV, 2018. 4 +[59] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In CVPR, 2018. 4 +[60] Barret Zoph, Golnaz Ghiasi, Tsung-Yi Lin, Yin Cui, Hanxiao Liu, Ekin Dogus Cubuk, and Quoc Le. Rethinking pretraining and self-training. In NeurIPS, 2020. 6 +[61] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In ECCV, 2018. 6 +[62] Vlas Zyrianov, Xiyue Zhu, and Shenlong Wang. Learning to generate realistic lidar point clouds. In ECCV, 2022. 2, 3 +[63] Vlas Zyrianov, Henry Che, Zhijian Liu, and Shenlong Wang. Lidardm: Generative lidar simulation in a generated world. arXiv preprint arXiv:2404.02903, 2024. 2, 3 \ No newline at end of file diff --git a/2025/Transfer Your Perspective_ Controllable 3D Generation from Any Viewpoint in a Driving Scene/images.zip b/2025/Transfer Your Perspective_ Controllable 3D Generation from Any Viewpoint in a Driving Scene/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..42c026979078a7658d4ccf2178afa9a90c0d9af0 --- /dev/null +++ b/2025/Transfer Your Perspective_ Controllable 3D Generation from Any Viewpoint in a Driving Scene/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d2b65985bdd9de49aafff02a745dd38a7232c183b6d3a73fc722e78e701589e3 +size 575822 diff --git a/2025/Transfer Your Perspective_ Controllable 3D Generation from Any Viewpoint in a Driving Scene/layout.json b/2025/Transfer Your Perspective_ Controllable 3D Generation from Any Viewpoint in a Driving Scene/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6607e0e26d13ae46483901d7a4cc52de16cf64c1 --- /dev/null +++ b/2025/Transfer Your Perspective_ Controllable 3D Generation from Any Viewpoint in a Driving Scene/layout.json @@ -0,0 +1,10109 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 103, + 515, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 103, + 515, + 140 + ], + "spans": [ + { + "bbox": [ + 96, + 103, + 515, + 140 + ], + "type": "text", + "content": "Transfer Your Perspective: Controllable 3D Generation from Any Viewpoint in a Driving Scene" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 160, + 492, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 160, + 492, + 190 + ], + "spans": [ + { + "bbox": [ + 116, + 160, + 492, + 190 + ], + "type": "text", + "content": "Tai-Yu Pan1, Sooyoung Jeon1, Mengdi Fan1, Jinsu Yoo1, Zhenyang Feng1, Mark Campbell2, Kilian Q. Weinberger2, Bharath Hariharan2, Wei-Lun Chao1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 187, + 198, + 419, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 198, + 419, + 213 + ], + "spans": [ + { + "bbox": [ + 187, + 198, + 419, + 213 + ], + "type": "text", + "content": "1The Ohio State University, 2Cornell University" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 57, + 232, + 181, + 316 + ], + "blocks": [ + { + "bbox": [ + 57, + 232, + 181, + 316 + ], + "lines": [ + { + "bbox": [ + 57, + 232, + 181, + 316 + ], + "spans": [ + { + "bbox": [ + 57, + 232, + 181, + 316 + ], + "type": "image", + "image_path": "c5eae7588b627f3995606de2f3865121e75062d88ec3559091c23eeb26d94532.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 317, + 167, + 328 + ], + "lines": [ + { + "bbox": [ + 70, + 317, + 167, + 328 + ], + "spans": [ + { + "bbox": [ + 70, + 317, + 167, + 328 + ], + "type": "text", + "content": "(a) Ego-centric driving" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 182, + 232, + 304, + 316 + ], + "blocks": [ + { + "bbox": [ + 182, + 232, + 304, + 316 + ], + "lines": [ + { + "bbox": [ + 182, + 232, + 304, + 316 + ], + "spans": [ + { + "bbox": [ + 182, + 232, + 304, + 316 + ], + "type": "image", + "image_path": "e27f28dbad465dcc1139d5991ac0b7a3ccbfe08dc906e1ba7a7f66b9b36cbfc7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 317, + 298, + 328 + ], + "lines": [ + { + "bbox": [ + 187, + 317, + 298, + 328 + ], + "spans": [ + { + "bbox": [ + 187, + 317, + 298, + 328 + ], + "type": "text", + "content": "(b) TYP for dynamic agent" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 331, + 555, + 377 + ], + "lines": [ + { + "bbox": [ + 54, + 331, + 555, + 377 + ], + "spans": [ + { + "bbox": [ + 54, + 331, + 555, + 377 + ], + "type": "text", + "content": "Figure 1. Illustration of the proposed problem and solution, Transfer Your Perspective (TYP). (a) A given sensory data captured by the ego-car (red triangle). (b) A generated sensory data by TYP, seeing from the viewpoint of another vehicle (green triangle) in the same scene. (c) A generated sensory data, seeing from an imaginary static agent like roadside units (blue icon). (d) Putting all the sensory data together, given or generated, TYP enables the development of collaborative perception with little or no real collaborative driving data." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 232, + 428, + 316 + ], + "blocks": [ + { + "bbox": [ + 306, + 232, + 428, + 316 + ], + "lines": [ + { + "bbox": [ + 306, + 232, + 428, + 316 + ], + "spans": [ + { + "bbox": [ + 306, + 232, + 428, + 316 + ], + "type": "image", + "image_path": "3d13adc16fd89ace3f2f1190193f2e2518d328b72661b1d52aa989107ca99dc0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 317, + 416, + 328 + ], + "lines": [ + { + "bbox": [ + 317, + 317, + 416, + 328 + ], + "spans": [ + { + "bbox": [ + 317, + 317, + 416, + 328 + ], + "type": "text", + "content": "(c) TYP for static agent" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 429, + 232, + 553, + 316 + ], + "blocks": [ + { + "bbox": [ + 429, + 232, + 553, + 316 + ], + "lines": [ + { + "bbox": [ + 429, + 232, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 429, + 232, + 553, + 316 + ], + "type": "image", + "image_path": "cd67161cbbb17dfff66d9eb055a4998bbe8690fa8f314a7ce0a6edecc97feb90.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 317, + 542, + 328 + ], + "lines": [ + { + "bbox": [ + 440, + 317, + 542, + 328 + ], + "spans": [ + { + "bbox": [ + 440, + 317, + 542, + 328 + ], + "type": "text", + "content": "(d) Collaborative driving" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 152, + 392, + 200, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 392, + 200, + 403 + ], + "spans": [ + { + "bbox": [ + 152, + 392, + 200, + 403 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 425, + 298, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 298, + 701 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 298, + 701 + ], + "type": "text", + "content": "Self-driving cars relying solely on ego-centric perception face limitations in sensing, often failing to detect occluded, faraway objects. Collaborative autonomous driving (CAV) seems like a promising direction, but collecting data for development is non-trivial. It requires placing multiple sensor-equipped agents in a real-world driving scene, simultaneously! As such, existing datasets are limited in locations and agents. We introduce a novel surrogate to the rescue, which is to generate realistic perception from different viewpoints in a driving scene, conditioned on a real-world sample—the ego-car's sensory data. This surrogate has huge potential: it could potentially turn any ego-car dataset into a collaborative driving one to scale up the development of CAV. We present the very first solution, using a combination of simulated collaborative data and real ego-car data. Our method Transfer Your Perspective (TYP) learns a conditioned diffusion model whose output samples are not only realistic but also consistent in both semantics and layouts with the given ego-car data. Empirical results demonstrate TYP's effectiveness in aiding in a CAV setting. In particular, TYP enables us to (pre-)train collaborative perception algorithms like early and late fusion with little or no real-world collaborative data, greatly facilitating downstream CAV applications." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 392, + 394, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 392, + 394, + 403 + ], + "spans": [ + { + "bbox": [ + 315, + 392, + 394, + 403 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 413, + 556, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 556, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 556, + 545 + ], + "type": "text", + "content": "Seeing the world from an ego-centric perspective, a self-driving car risks being \"narrow-sighted,\" limiting its ability to respond appropriately in dynamic driving environments. For instance, it should slow down or honk if a pedestrian is about to cross the road or if another vehicle is merging into its lane. However, these actions depend on the car's ability to detect traffic participants, which may be occluded by a large bus or a building at a sharp intersection. This occlusion problem can hardly be addressed by mounting a few more sensors on the car. We argue that a self-driving car must move beyond its ego-centric perspective." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 547, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 547, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 547, + 557, + 715 + ], + "type": "text", + "content": "One intuitive way is to collaborate with nearby \"agents\" that observe the scene from different angles, such as other sensor-equipped cars or static devices like roadside units (RSUs). When trained together to address GPS errors and synchronization delays [55], collaborative perception has been shown to significantly improve the accuracy of each agent's perception, particularly in detecting occluded or distant objects [3, 4, 14, 19, 23, 33, 42, 45, 50-52]. However, collecting training data for collaborative perception is never easy. Unlike single-agent data, which can be collected by simply driving a car on the road, collaborative data requires the simultaneous presence of multiple agents in the same driving scene. For dynamic agents like vehicles, precise coordination is needed to ensure they are within communication" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12027" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "range. These challenges limit existing works in scale and the number of agents (typically just two). While one may leverage simulated data by game engines [7], they often fail to capture the diversity of real-world scenes. We thus ask:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 124, + 288, + 159 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 124, + 288, + 159 + ], + "spans": [ + { + "bbox": [ + 66, + 124, + 288, + 159 + ], + "type": "text", + "content": "Can we obtain realistic data for learning collaborative perception with much less effort, ideally as easily as single-agent data?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 163, + 296, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 296, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 296, + 222 + ], + "type": "text", + "content": "Specifically, given the vast amount of single-agent LiDAR data already collected across various driving environments, is it possible to convert each of them into collaborative perception data by generating additional point clouds from alternative reference viewpoints within the same scene?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 223, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 223, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 56, + 223, + 296, + 449 + ], + "type": "text", + "content": "At first glance, this question may seem overly ambitious for three reasons. First, for collaborative perception to be effective, the generated point cloud must provide information that the ego-car's point cloud cannot capture, such as occluded surfaces invisible to the ego-car. This creates a chicken-and-egg problem: if the ego-car's point cloud is all we initially know about the scene, how can we deduce any additional information? Second, for training purposes, the generated data must be realistic. It should replicate the constraints and data patterns of a real sensor as if it were positioned at the reference viewpoint, producing no points in occluded areas and fewer points in distant areas. Last but not least, in regions commonly perceivable from both the reference and ego-car's viewpoints, the generated data should align in layouts and semantics with the ego-car's data. On the surface, this may seem as simple as copying the ego-car's data, but doing so would violate the second requirement. In essence, the generated data should resemble a point cloud seen from the reference viewpoint, not the ego-car." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 450, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 689 + ], + "type": "text", + "content": "That said, after a deeper look at the question, we argue that it is achievable with three key insights. First, given semantic information around the ego-car for conditioning, existing research has shown promising progress in generating realistic point clouds [13, 27, 32, 44, 49, 62, 63]. This is especially encouraging, as most existing single-agent datasets provide 3D object labels, and they can be translated to obtain semantic information centered around the reference position. This further implies that we can generate occluded object surfaces as seen from the ego-car if they are visible from the reference viewpoint. Second, semantic information is easily editable: even if the translated object map has noticeable empty areas due to limited ego-car perception, object boxes can be manually added to make the map appear more realistic from the reference viewpoint. Lastly, in commonly perceivable regions where the generated data needs to meet two physical constraints, we could leverage simulators based on computational graphics and optics. Specifically, we can use simulated data from the two views to train a conditioned generative model that maps one viewpoint to the other." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Building upon these insights, we propose Transfer Your Perspective (TYP), a novel research problem and the very" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": "first solution for generating a realistic point cloud from any viewpoint in a scene, given real ego-car's point cloud and semantic labels as conditions. TYP assumes access to 1) a simulated collaborative driving dataset with multiple agents perceiving the same scenes from different viewpoints and 2) a real single-agent driving dataset; both are labeled. Our solution involves a conditioned latent diffusion model [34] and a dedicated two-stage training process. In the first stage, we consider a single-agent scenario and train the model using real data conditioning only on object locations. This equips the model with the ability to generate diverse and realistic scenes. We denote the learned model by " + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "P(\\pmb{x}|\\pmb{y})" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": " stands for the semantic condition and " + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": " for the point cloud. In the second stage, we incorporate the simulated data to learn how to ground generation on data from another agent's viewpoint so that the model can produce a semantically consistent reference point cloud given the ego-car's data. We learn a lightweight conditioning module to turn " + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "P(\\pmb{x}|\\pmb{y})" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "P(\\pmb{x}_r|\\pmb{x}_e,\\pmb{y}_r)" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": " indicate ego and reference views, respectively. We note that " + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\pmb{x}_e" + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": " was pre-translated to center around the reference position." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 326, + 556, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 556, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 556, + 457 + ], + "type": "text", + "content": "One challenge in TYP is the domain gap between real and simulated data. Due to differences in sensor configurations and placements, data collection environments, and the sim-to-real gap, the two sets of point clouds inevitably exhibit discrepant distributions, patterns, and densities. To address this, we insert a domain adaptation step between the two training stages. We train a separate encoder-decoder for the simulated data while enforcing a constraint to make the encoded features of simulated and real data indistinguishable [40]. This step allows us to learn " + }, + { + "bbox": [ + 313, + 326, + 556, + 457 + ], + "type": "inline_equation", + "content": "P(\\boldsymbol{x}_r|\\boldsymbol{x}_e,\\boldsymbol{y}_r)" + }, + { + "bbox": [ + 313, + 326, + 556, + 457 + ], + "type": "text", + "content": " in a space with reduced domain discrepancy during the second stage." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 460, + 556, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 556, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 556, + 555 + ], + "type": "text", + "content": "Once trained, we pair " + }, + { + "bbox": [ + 313, + 460, + 556, + 555 + ], + "type": "inline_equation", + "content": "P(\\pmb{x}_r|\\pmb{x}_e,\\pmb{y}_r)" + }, + { + "bbox": [ + 313, + 460, + 556, + 555 + ], + "type": "text", + "content": " with the real data's encoder-decoder to generate real-style point clouds grounding on real ego-car's data, so that we can develop collaborative perception algorithms without real collaborative data. More specifically, given a real ego-car's perception " + }, + { + "bbox": [ + 313, + 460, + 556, + 555 + ], + "type": "inline_equation", + "content": "\\pmb{x}_e" + }, + { + "bbox": [ + 313, + 460, + 556, + 555 + ], + "type": "text", + "content": " and label " + }, + { + "bbox": [ + 313, + 460, + 556, + 555 + ], + "type": "inline_equation", + "content": "\\pmb{y}_e" + }, + { + "bbox": [ + 313, + 460, + 556, + 555 + ], + "type": "text", + "content": ", we first translate them to center around the reference position, followed by optionally injecting object labels into " + }, + { + "bbox": [ + 313, + 460, + 556, + 555 + ], + "type": "inline_equation", + "content": "\\pmb{y}_e" + }, + { + "bbox": [ + 313, + 460, + 556, + 555 + ], + "type": "text", + "content": " to make it realistic from the reference viewpoint." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 558, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 557, + 715 + ], + "type": "text", + "content": "We extensively validate TYP on multiple datasets, all in an offline setting. Empirical results demonstrate TYP's effectiveness in generating high-quality reference data to aid in the development of collaborative perception. In particular, we show that a conditioned diffusion model trained solely on simulated data (e.g., OPV2V [51]) can already turn a real single-agent dataset into a real-alike collaborative one. As such, one can train collaborative perception algorithms for real test data without real training data. We further generate the \"ColWaymo\" dataset by training TYP on real single-agent Waymo data [39] and simulated OPV2V data [51], thus turning the former into a collaborative one. Collaborative perception backbones pre-trained on the large-scale, semi-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12028" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "synthetic ColWaymo data demonstrate a remarkable transferability. They notably improve collaborative perception developed in the downstream tasks (e.g., V2V4Real [52]), even in a many-shot setting." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 121, + 206, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 121, + 206, + 131 + ], + "spans": [ + { + "bbox": [ + 67, + 121, + 206, + 131 + ], + "type": "text", + "content": "In sum, our key contributions are:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 133, + 297, + 264 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 133, + 297, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 297, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 297, + 204 + ], + "type": "text", + "content": "- We propose a new research direction to aid in the development of collaborative perception, generating real-looking sensory data from potentially any viewpoint in a real driving scene, capable of turning a single-agent dataset into a collaborative one. This has the potential to scale up collaborative autonomous driving (CAV)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 204, + 295, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 295, + 239 + ], + "type": "text", + "content": "- We present TYP, a two-stage training recipe with domain adaptation in between to learn the generative model using simulated collaborative data and real single-agent data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 240, + 296, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 296, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 296, + 264 + ], + "type": "text", + "content": "- Extensive experiments demonstrate TYP's effectiveness in aiding in collaborative perception across various scenarios." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 274, + 141, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 141, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 141, + 286 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 294, + 297, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 297, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 297, + 653 + ], + "type": "text", + "content": "Collaborative autonomous driving (CAV) offers significant benefits, including extended perception range by sharing sensor information to detect objects beyond a single vehicle's field of view. Despite these benefits, collecting large-scale, real-world datasets for CAV poses significant challenges due to the complexity and cost of deploying multiple instrumented vehicles in diverse environments. Existing datasets often have limitations that restrict the scope of CAV research. For instance, OPV2V [51] and V2X-Sim [20] rely on simulations using the CARLA simulator [7], allowing controlled scenarios with multiple vehicle agents, yet lacking real-world variability. V2V4Real [52], on the other hand, provides real-world data, yet it only contains two collaborating agents, which restricts the exploration of more complex multi-agent interactions. Similarly, the DAIR-V2X [57] dataset also offers real-world data including vehicle-to-infrastructure (V2I) and V2V, but it mainly focuses on V2I scenarios and still has a limited number of collaborating vehicles. To overcome current limitations, we propose a new research direction—generating realistic point clouds given the ego's perspective. 3D generation. Various works have explored 3D scene generation, such as LiDARGen [62], R2DM [27], LiDM [32], UltraLiDAR [48], and RangeLDM [13]. Additionally, Li-darDM [63] and Text2LiDAR [44] investigate conditional scene generation, with the former relying on hand-crafted map layouts and the latter conditioning on text inputs. However, all these methods focus on ego-centric generation. We propose a new research direction for CAV, aiming to generate realistic and consistent scenes conditioned on the ego agent's real point clouds—an area that remains largely unexplored." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Diffusion models. Diffusion models have recently advanced generative modeling for high-quality LiDAR point clouds and images. Denoising Diffusion Probabilistic Models (DDPMs) [12] outperform traditional Generative Adversarial Networks (GANs) [10] and further enhance efficiency" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 73, + 553, + 203 + ], + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 203 + ], + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 203 + ], + "type": "image", + "image_path": "ca03f5ab0b91be56465867e6e1791601e932ceaa1d5f9cc20429fa1f896b143e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 209, + 555, + 308 + ], + "lines": [ + { + "bbox": [ + 313, + 209, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 555, + 308 + ], + "type": "text", + "content": "Figure 2. Illustration of TYP's conditioned generative model and training process. We propose a two-stage training procedure. The first stage maximizes the generation capability by conditioning solely on object locations (using real single-agent target data), while the second stage grounds the generation on the ego-car's perspective to match semantics and layouts (using simulated CAV data). Additionally, we introduce a discriminator to adapt simulated CAV features to the real target domain, making the trained model readily applicable to the target domain after the second stage." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 312, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 373 + ], + "type": "text", + "content": "for LiDAR generation. In autonomous driving applications, methods like RangeLDM [13], LidarDM [63], and LiDAR-Gen [62] apply diffusion models for realistic LiDAR scene generation. In this paper, we utilize its generation capability to study our proposed problem." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 382, + 500, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 500, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 500, + 396 + ], + "type": "text", + "content": "3. Transfer Your Perspective (TYP)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 403, + 556, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 556, + 535 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 556, + 535 + ], + "type": "text", + "content": "In this paper, we introduce a new research direction to advance collaborative autonomous driving (CAV): generating LiDAR point clouds from different perspectives within the same scene as the ego agent, aiming to reduce the tedious efforts of collecting data for CAV. We begin by defining the proposed problem in Sec. 3.1. In Sec. 3.2, we discuss the representations of the inputs, including point clouds and semantic information. Sec. 3.3 outlines the pipeline developed to address this problem. Finally, in Sec. 3.4, we demonstrate how this capability can be applied to datasets that only have labeled ego agents, e.g., Waymo Open Dataset (WOD) [39]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 540, + 406, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 406, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 406, + 553 + ], + "type": "text", + "content": "3.1. Problem Setup" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 558, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 653 + ], + "type": "text", + "content": "Given the perception data " + }, + { + "bbox": [ + 313, + 558, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\pmb{x}_e" + }, + { + "bbox": [ + 313, + 558, + 555, + 653 + ], + "type": "text", + "content": " from an ego agent, we aim to build a model " + }, + { + "bbox": [ + 313, + 558, + 555, + 653 + ], + "type": "inline_equation", + "content": "P(\\pmb{x}_r|\\pmb{x}_e)" + }, + { + "bbox": [ + 313, + 558, + 555, + 653 + ], + "type": "text", + "content": " to generate new perception data " + }, + { + "bbox": [ + 313, + 558, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\pmb{x}_r" + }, + { + "bbox": [ + 313, + 558, + 555, + 653 + ], + "type": "text", + "content": " seen from a different reference location and perspective within a communication range. Here, " + }, + { + "bbox": [ + 313, + 558, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 558, + 555, + 653 + ], + "type": "text", + "content": " represents a LiDAR point cloud. This task presents several challenges, including potential information gaps between the two views, the need for alignment within the commonly perceivable area, and ensuring realism in the generated data, as discussed in Sec. 1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "content": "To address these challenges, we extend the original problem by incorporating semantic information, such as object bounding boxes, represented as " + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "inline_equation", + "content": "P(\\pmb{x}_r|\\pmb{x}_e,\\pmb{y}_r)" + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "content": ". We assume the availability of this information around the ego agent, denoted by " + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{y}_e" + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "content": ". For example, most of the existing single-agent" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12029" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": "datasets provide object labels. This semantic information can be easily translated and edited to become " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_r" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": " by tools like traffic re-players [11], making it a key to bridge the information gap between the ego and reference views, enabling the reference agent to \"see\" those objects and surfaces beyond the ego's view—a core concept in CAV. Additionally, by aligning generated point clouds with object locations, this extension facilitates our goal of scaling up CAV development. In essence, the generated " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_r, \\mathbf{y}_r)" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": " pairs can be used flexibly alongside " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_e, \\mathbf{y}_e)" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": " in various CAV applications, like directly using them to train collaborative perception algorithms." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 204, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 296, + 228 + ], + "type": "text", + "content": "In the following sections, we describe how " + }, + { + "bbox": [ + 55, + 204, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 204, + 296, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 204, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 55, + 204, + 296, + 228 + ], + "type": "text", + "content": " are encoded, followed by our approach to tackling the problem." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 234, + 236, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 234, + 236, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 236, + 247 + ], + "type": "text", + "content": "3.2. Representations and Embeddings" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 251, + 296, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 371 + ], + "type": "text", + "content": "LiDAR. There are multiple ways to represent point clouds in continuous 3D space, such as coordinates (i.e., x, y, and z) [30, 31, 37], range images [25, 30, 43], and voxelization [17, 24, 53, 59]. To better align with spatial control from object locations, we follow [49] to voxelize point clouds using a pre-defined grid and record voxel occupancy. [49] also highlights that this representation can naturally handle varying point densities and only minimally impacts LiDAR generation, with some precision trade-offs during voxelization. In short, we represent a point cloud by " + }, + { + "bbox": [ + 55, + 251, + 296, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 55, + 251, + 296, + 371 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 371, + 296, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 296, + 442 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 296, + 442 + ], + "type": "text", + "content": "To avoid the computational cost of 3D convolutions, we convert " + }, + { + "bbox": [ + 55, + 371, + 296, + 442 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 371, + 296, + 442 + ], + "type": "text", + "content": " into Bird's-Eye-View (BEV) images by treating the height dimension (i.e., " + }, + { + "bbox": [ + 55, + 371, + 296, + 442 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 371, + 296, + 442 + ], + "type": "text", + "content": ") as feature channels for 2D convolutions. This approach has been widely adopted in self-driving perception [5, 48, 58], allowing the use of 2D image-based model architectures and algorithms." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 444, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 296, + 514 + ], + "type": "text", + "content": "Semantic information. Given that the point clouds are represented as BEV images, it is also intuitive to represent object locations in BEV. We create binary object maps by considering only the x and y coordinates of 3D bounding boxes, resulting in " + }, + { + "bbox": [ + 55, + 444, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\pmb{y} \\in \\mathbb{R}^{H \\times W \\times 1}" + }, + { + "bbox": [ + 55, + 444, + 296, + 514 + ], + "type": "text", + "content": ". (We could further extend it by including category information.)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 515, + 296, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 296, + 634 + ], + "type": "text", + "content": "Feature embedding. Following the literature [34], we encode input tensors using a Vector Quantized-Variational Autoencoder (VQ-VAE) [41], comprising an encoder " + }, + { + "bbox": [ + 55, + 515, + 296, + 634 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 515, + 296, + 634 + ], + "type": "text", + "content": ", a quantization function " + }, + { + "bbox": [ + 55, + 515, + 296, + 634 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 515, + 296, + 634 + ], + "type": "text", + "content": " for feature vectors on spatial grids, and a decoder " + }, + { + "bbox": [ + 55, + 515, + 296, + 634 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 515, + 296, + 634 + ], + "type": "text", + "content": ". Formally, we obtain " + }, + { + "bbox": [ + 55, + 515, + 296, + 634 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}^f = E_x(\\boldsymbol{x}) \\in \\mathbb{R}^{h \\times w \\times c_x}" + }, + { + "bbox": [ + 55, + 515, + 296, + 634 + ], + "type": "text", + "content": " by the encoder, map each " + }, + { + "bbox": [ + 55, + 515, + 296, + 634 + ], + "type": "inline_equation", + "content": "c_x" + }, + { + "bbox": [ + 55, + 515, + 296, + 634 + ], + "type": "text", + "content": "-dimensional vector to a learnable code to obtain " + }, + { + "bbox": [ + 55, + 515, + 296, + 634 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}^z = Q_x(\\boldsymbol{x}^f) \\in \\mathbb{R}^{h \\times w \\times c_x}" + }, + { + "bbox": [ + 55, + 515, + 296, + 634 + ], + "type": "text", + "content": ", and generate outputs using the decoder " + }, + { + "bbox": [ + 55, + 515, + 296, + 634 + ], + "type": "inline_equation", + "content": "\\hat{\\boldsymbol{x}} = G_x(\\boldsymbol{x}^z) \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 55, + 515, + 296, + 634 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 515, + 296, + 634 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 515, + 296, + 634 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 515, + 296, + 634 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 515, + 296, + 634 + ], + "type": "text", + "content": " are spatial resolutions of the feature map downsampled from " + }, + { + "bbox": [ + 55, + 515, + 296, + 634 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 515, + 296, + 634 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 515, + 296, + 634 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 515, + 296, + 634 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 515, + 296, + 634 + ], + "type": "inline_equation", + "content": "c_x" + }, + { + "bbox": [ + 55, + 515, + 296, + 634 + ], + "type": "text", + "content": " is the number of channels." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 635, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 635, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 67, + 635, + 296, + 647 + ], + "type": "text", + "content": "The VQ-VAE model is trained end-to-end by minimizing:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 656, + 296, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 656, + 296, + 671 + ], + "spans": [ + { + "bbox": [ + 61, + 656, + 296, + 671 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {v q}} = \\mathcal {L} _ {\\mathrm {r e c}} + \\| \\operatorname {s g} [ E _ {x} (\\boldsymbol {x}) ] - \\boldsymbol {x} ^ {z} \\| _ {2} ^ {2} + \\| \\operatorname {s g} [ \\boldsymbol {x} ^ {z} ] - E _ {x} (\\boldsymbol {x}) \\| _ {2} ^ {2}, \\tag {1}", + "image_path": "bf724854ed27400a9aa9130cdcf6237e3c8763909d174a873bdf698a1fb06f11.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{sg}[\\cdot ]" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rec}}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " denote the stop-gradient operation and the reconstruction loss, respectively. As our point cloud representation is binary occupancy " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pmb {x}\\in \\{0,1\\}^{H\\times W\\times C}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ", binary" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": "cross-entropy is a natural choice for " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rec}}" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": ". However, due to the sparsity of point clouds, this results in an imbalanced loss. To address this, we adopt the Focal Loss (FL) [36]:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 323, + 118, + 555, + 144 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 118, + 555, + 144 + ], + "spans": [ + { + "bbox": [ + 323, + 118, + 555, + 144 + ], + "type": "interline_equation", + "content": "\\ell_ {\\mathrm {F L}} \\left(\\boldsymbol {x} _ {i}, \\hat {\\boldsymbol {x}} _ {i}\\right) = \\left\\{ \\begin{array}{l l} - (1 - \\hat {\\boldsymbol {x}} _ {i}) ^ {\\gamma} \\log \\left(\\hat {\\boldsymbol {x}} _ {i}\\right) & \\text {i f} \\boldsymbol {x} _ {i} = 1 \\\\ - \\hat {\\boldsymbol {x}} _ {i} ^ {\\gamma} \\log \\left(1 - \\hat {\\boldsymbol {x}} _ {i}\\right) & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {2}", + "image_path": "cb6018c3b212129ec6a7716d790d01c70c7c02c5dd777c79e471d63b40020efb.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 386, + 156, + 555, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 156, + 555, + 190 + ], + "spans": [ + { + "bbox": [ + 386, + 156, + 555, + 190 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r e c}} = \\sum_ {i = 0} ^ {M} \\ell_ {\\mathrm {F L}} \\left(\\boldsymbol {x} _ {i}, \\hat {\\boldsymbol {x}} _ {i}\\right), \\tag {3}", + "image_path": "588fbfbfbe80863fe59535011f7c3e4c502930fe8cb9391eec95043708949fdc.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 195, + 547, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 195, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 313, + 195, + 547, + 206 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 195, + 547, + 206 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 195, + 547, + 206 + ], + "type": "text", + "content": " is the voxel index and " + }, + { + "bbox": [ + 313, + 195, + 547, + 206 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 195, + 547, + 206 + ], + "type": "text", + "content": " is the number of voxels." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 207, + 554, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 554, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 554, + 255 + ], + "type": "text", + "content": "For object locations " + }, + { + "bbox": [ + 313, + 207, + 554, + 255 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 313, + 207, + 554, + 255 + ], + "type": "text", + "content": ", we use the same approach to train a separate VQ-VAE model " + }, + { + "bbox": [ + 313, + 207, + 554, + 255 + ], + "type": "inline_equation", + "content": "E_{y}, Q_{y}" + }, + { + "bbox": [ + 313, + 207, + 554, + 255 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 207, + 554, + 255 + ], + "type": "inline_equation", + "content": "G_{y}" + }, + { + "bbox": [ + 313, + 207, + 554, + 255 + ], + "type": "text", + "content": ". In the subsequent subsections, we use " + }, + { + "bbox": [ + 313, + 207, + 554, + 255 + ], + "type": "inline_equation", + "content": "\\pmb{x}^{f} \\in \\mathbb{R}^{h \\times w \\times c_{x}}" + }, + { + "bbox": [ + 313, + 207, + 554, + 255 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 207, + 554, + 255 + ], + "type": "inline_equation", + "content": "\\pmb{y}^{f} \\in \\mathbb{R}^{h \\times w \\times c_{y}}" + }, + { + "bbox": [ + 313, + 207, + 554, + 255 + ], + "type": "text", + "content": " as feature embeddings for learning a latent diffusion model." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 264, + 529, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 529, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 529, + 277 + ], + "type": "text", + "content": "3.3. Transfer Your Perspective by Generation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 281, + 555, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 329 + ], + "type": "text", + "content": "We model the distribution " + }, + { + "bbox": [ + 313, + 281, + 555, + 329 + ], + "type": "inline_equation", + "content": "P(\\pmb{x}_r | \\pmb{x}_e, \\pmb{y}_r)" + }, + { + "bbox": [ + 313, + 281, + 555, + 329 + ], + "type": "text", + "content": " defined in Sec. 3.1 by a conditioned generative model and train it in two stages. We assume access to " + }, + { + "bbox": [ + 313, + 281, + 555, + 329 + ], + "type": "inline_equation", + "content": "(\\pmb{x}_r, \\pmb{x}_e, \\pmb{y}_r)" + }, + { + "bbox": [ + 313, + 281, + 555, + 329 + ], + "type": "text", + "content": " tuples as training data. We detail the training data preparation in Sec. 3.4 and Sec. 4." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 330, + 555, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 555, + 473 + ], + "type": "text", + "content": "In the first stage, we aim to maximize the generation capability by providing minimal conditions, i.e., only bounding boxes, without further constraints on the scene. This encourages the model to produce " + }, + { + "bbox": [ + 313, + 330, + 555, + 473 + ], + "type": "inline_equation", + "content": "P(\\boldsymbol{x}|\\boldsymbol{y})" + }, + { + "bbox": [ + 313, + 330, + 555, + 473 + ], + "type": "text", + "content": " with high flexibility. Second, building on this model, we incorporate the ego agent's point cloud as an additional cue and ground the generation process on it to ensure semantic and layout consistency between the generated and ego agent's point clouds. Put together, this two-stage training procedure enables the learned " + }, + { + "bbox": [ + 313, + 330, + 555, + 473 + ], + "type": "inline_equation", + "content": "P(\\boldsymbol{x}_r|\\boldsymbol{x}_e,\\boldsymbol{y}_r)" + }, + { + "bbox": [ + 313, + 330, + 555, + 473 + ], + "type": "text", + "content": " model to generate perception data at areas beyond the commonly visible regions between two views. The model and training process is illustrated in Fig. 2." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 474, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 557 + ], + "type": "text", + "content": "More importantly, this approach enables us to leverage existing labeled single-agent datasets such as KITTI [9], NuScenes [2], and Waymo Open Dataset (WOD) [39], since the first-stage training only requires ego-centric LiDAR point clouds and their corresponding semantic information (see Sec. 3.4). The " + }, + { + "bbox": [ + 313, + 474, + 554, + 557 + ], + "type": "inline_equation", + "content": "(\\pmb{x}_e,\\pmb{x}_r)" + }, + { + "bbox": [ + 313, + 474, + 554, + 557 + ], + "type": "text", + "content": " data pairs seen from the ego and the reference agents are needed only in the second stage." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 326, + 558, + 541, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 558, + 541, + 569 + ], + "spans": [ + { + "bbox": [ + 326, + 558, + 541, + 569 + ], + "type": "text", + "content": "In the following, we elaborate on each training stage." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "Stage 1: Generation with semantic information. The goal of this stage is to equip the model with strong generation capabilities of point clouds given spatial conditions, preparing it for the next stage. As discussed in Sec. 3.2, the point clouds and object locations are embedded by VQVAEs [41], denoted as " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}^f = E_x(\\boldsymbol{x}) \\in \\mathbb{R}^{h \\times w \\times c_x}" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\boldsymbol{y}^f = E_x(\\boldsymbol{y}) \\in \\mathbb{R}^{h \\times w \\times c_y}" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": ". The image-like feature maps allow us to adopt existing generation algorithms for 2D images [6, 12, 16]. In this paper, we apply one of the most popular generative models, the Latent Diffusion Model (LDM) [34], for conditioned generation " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "P(\\boldsymbol{x} | \\boldsymbol{y})" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": ". LDM seeks to model a data distribution by iteratively denoising variables" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12030" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 293, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 293, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 293, + 85 + ], + "type": "text", + "content": "that are initially sampled from a Gaussian. The objective is:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 92, + 296, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 92, + 296, + 119 + ], + "spans": [ + { + "bbox": [ + 62, + 92, + 296, + 119 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {L D M}} = \\mathbb {E} _ {\\boldsymbol {x} ^ {f}, \\boldsymbol {y} ^ {f}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\boldsymbol {x} _ {t} ^ {f}, \\boldsymbol {y} ^ {f}, t\\right) \\right\\| _ {2} ^ {2} \\right], \\tag {4}", + "image_path": "6fcb6d0c051265b5f510c3dfc1a43fb2ba723ffbbacf9a759a90176bc7ed4df2.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 127, + 290, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 127, + 290, + 139 + ], + "spans": [ + { + "bbox": [ + 55, + 127, + 290, + 139 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 55, + 127, + 290, + 139 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 127, + 290, + 139 + ], + "type": "text", + "content": " is a UNet [35] backbone and " + }, + { + "bbox": [ + 55, + 127, + 290, + 139 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 127, + 290, + 139 + ], + "type": "text", + "content": " is the timestamp." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 140, + 295, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 295, + 210 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 295, + 210 + ], + "type": "text", + "content": "Stage 2: Generation for new perspectives. In this stage, we aim to ground the generation of point clouds by incorporating the ego agent's perception. To retain the generation capability from stage 1, we freeze the learned " + }, + { + "bbox": [ + 55, + 140, + 295, + 210 + ], + "type": "inline_equation", + "content": "P(\\boldsymbol{x}|\\boldsymbol{y})" + }, + { + "bbox": [ + 55, + 140, + 295, + 210 + ], + "type": "text", + "content": " and introduce a learnable lightweight control module to inject additional cues, following T2I-Adapter [26], as shown in Fig. 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 212, + 296, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 296, + 283 + ], + "type": "text", + "content": "Formally, let " + }, + { + "bbox": [ + 55, + 212, + 296, + 283 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{e'}" + }, + { + "bbox": [ + 55, + 212, + 296, + 283 + ], + "type": "text", + "content": " denote the translated and rotated egoagent's point cloud centered around the reference location and aligned with the reference orientation, as shown in Fig. 2; " + }, + { + "bbox": [ + 55, + 212, + 296, + 283 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{e'}^f = E_x(\\pmb{x}_{e'}) \\in \\mathbb{R}^{h \\times w \\times c_x}" + }, + { + "bbox": [ + 55, + 212, + 296, + 283 + ], + "type": "text", + "content": " is the corresponding embedding. Let " + }, + { + "bbox": [ + 55, + 212, + 296, + 283 + ], + "type": "inline_equation", + "content": "F_{AD}" + }, + { + "bbox": [ + 55, + 212, + 296, + 283 + ], + "type": "text", + "content": " denote the learnable control module, which takes " + }, + { + "bbox": [ + 55, + 212, + 296, + 283 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{e'}^f" + }, + { + "bbox": [ + 55, + 212, + 296, + 283 + ], + "type": "text", + "content": " as input and outputs" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 292, + 296, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 292, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 139, + 292, + 296, + 308 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {c} = F _ {A D} \\left(\\boldsymbol {x} _ {e ^ {\\prime}} ^ {f}\\right), \\tag {5}", + "image_path": "b8556a9b6358e4d0f21254b82360e7999c680d8173b6e5d39ae5c80f37a8d25a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 316, + 296, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 365 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 316, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_c = \\{\\mathbf{F}_c^1,\\mathbf{F}_c^2,\\mathbf{F}_c^3,\\mathbf{F}_c^4\\}" + }, + { + "bbox": [ + 55, + 316, + 296, + 365 + ], + "type": "text", + "content": " matches the size of the multiscale features " + }, + { + "bbox": [ + 55, + 316, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{enc} = \\{\\mathbf{F}_{enc}^{1},\\mathbf{F}_{enc}^{2},\\mathbf{F}_{enc}^{3},\\mathbf{F}_{enc}^{4}\\}" + }, + { + "bbox": [ + 55, + 316, + 296, + 365 + ], + "type": "text", + "content": " extracted from the frozen encoder of the UNet " + }, + { + "bbox": [ + 55, + 316, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 316, + 296, + 365 + ], + "type": "text", + "content": ". With these ingredients, T2I-Adapter influences the generation process by" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 100, + 375, + 296, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 375, + 296, + 389 + ], + "spans": [ + { + "bbox": [ + 100, + 375, + 296, + 389 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {e n c} ^ {i} \\leftarrow \\mathbf {F} _ {e n c} ^ {i} + \\mathbf {F} _ {c} ^ {i}, i \\in \\{1, 2, 3, 4 \\}, \\tag {6}", + "image_path": "dba7126d1d91f9ebc4c39fbde9a8dcda434a973cf17630e7855ab5879f173d5f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 394, + 296, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 296, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 296, + 419 + ], + "type": "text", + "content": "which injects " + }, + { + "bbox": [ + 55, + 394, + 296, + 419 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{e'}^f" + }, + { + "bbox": [ + 55, + 394, + 296, + 419 + ], + "type": "text", + "content": " embedding into each of the denoising steps. The second-stage objective for learning " + }, + { + "bbox": [ + 55, + 394, + 296, + 419 + ], + "type": "inline_equation", + "content": "F_{AD}" + }, + { + "bbox": [ + 55, + 394, + 296, + 419 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 64, + 427, + 296, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 427, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 64, + 427, + 296, + 460 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {L D M}} = \\mathbb {E} _ {\\boldsymbol {x} _ {r} ^ {f}, \\boldsymbol {y} _ {r} ^ {f}, \\boldsymbol {x} _ {e ^ {\\prime}} ^ {f}, \\epsilon \\sim \\mathcal {N} (0, 1), t} \\left[ \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\boldsymbol {x} _ {r, t} ^ {f}, \\boldsymbol {y} ^ {f}, \\boldsymbol {x} _ {e ^ {\\prime}} ^ {f}, t\\right) \\right\\| _ {2} ^ {2} \\right], \\tag {7}", + "image_path": "a3428cb502c867b1d27c58a2b33abaa45db1bfad94ab4beca52c87b2d3357d3a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 461, + 286, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 286, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 286, + 474 + ], + "type": "text", + "content": "where the encoder of " + }, + { + "bbox": [ + 55, + 461, + 286, + 474 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 461, + 286, + 474 + ], + "type": "text", + "content": " is modified as defined in Eq. (6)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 480, + 287, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 287, + 493 + ], + "type": "text", + "content": "3.4. From Single Agent to Collaborative Datasets" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 498, + 296, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 557 + ], + "type": "text", + "content": "There are many publicly accessible large-scale real-world autonomous driving datasets for ego-centric 3D perception, like WOD [39], NuScenes [2], KITTI-360 [21], and PandaSet [46], whereas much fewer datasets exist for CAV due to the challenges in data collection as discussed in Sec. 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 558, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 665 + ], + "type": "text", + "content": "This limitation inspired a bold idea: Can we transfer the generation capability learned from paired CAV data to these existing labeled single-agent datasets? Achieving this would scale up CAV datasets substantially, offering immense benefits to the research community. However, this goal is challenging due to domain gaps between datasets, such as variations in LiDAR sensors, point cloud densities, data patterns, and data collection environments. Below, we describe our approach to addressing these challenges." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "Problem setup. Our target domain is a single-agent dataset, while the source domain consists of paired perception and semantic information from ego and reference perspectives, denoted as " + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\pmb{x}_r,\\pmb{x}_e,\\pmb{y}_r)" + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": ". We demonstrate that we can transfer" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "content": "knowledge from a simulated CAV dataset (e.g., OPV2V [51]) to a real-world ego-centric dataset (e.g., WOD [39])." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 96, + 555, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 555, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 555, + 167 + ], + "type": "text", + "content": "Two-Stage training is the key. As described in Sec. 3.3, we decompose the generative training into two stages: first learning " + }, + { + "bbox": [ + 313, + 96, + 555, + 167 + ], + "type": "inline_equation", + "content": "P(\\pmb{x}|\\pmb{y})" + }, + { + "bbox": [ + 313, + 96, + 555, + 167 + ], + "type": "text", + "content": ", followed by " + }, + { + "bbox": [ + 313, + 96, + 555, + 167 + ], + "type": "inline_equation", + "content": "P(\\pmb{x}_r|\\pmb{x}_e,\\pmb{y}_r)" + }, + { + "bbox": [ + 313, + 96, + 555, + 167 + ], + "type": "text", + "content": ". The second stage keeps " + }, + { + "bbox": [ + 313, + 96, + 555, + 167 + ], + "type": "inline_equation", + "content": "P(\\pmb{x}|\\pmb{y})" + }, + { + "bbox": [ + 313, + 96, + 555, + 167 + ], + "type": "text", + "content": " frozen by adding an adapter. This strategy enables us to use target domain data in the first stage. The resulting " + }, + { + "bbox": [ + 313, + 96, + 555, + 167 + ], + "type": "inline_equation", + "content": "P(\\pmb{x}|\\pmb{y})" + }, + { + "bbox": [ + 313, + 96, + 555, + 167 + ], + "type": "text", + "content": " would generate target-like point clouds." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "text", + "content": "Minimizing domain gaps. In the first stage, we have learned VQ-VAE encoder-decoders and " + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "inline_equation", + "content": "P(\\boldsymbol{x}|\\boldsymbol{y})" + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "text", + "content": " in the target domain (single-agent). In the second stage, we bring in CAV data to guide the generation on how to condition on ego-agent's perception. To minimize the gap of " + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}^f" + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "text", + "content": " between two domains—" + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}^f" + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "text", + "content": " is what the generative model is optimized to generate—we adopt a GAN-style discriminator [8, 10, 22, 40]. Specifically, we aim to make feature embeddings extracted from the source data and the target data indistinguishable; namely, to confuse a discriminator " + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "text", + "content": " trained to differentiate them. We freeze all parameters of the target domain's encoder " + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "inline_equation", + "content": "E_{tgt}" + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "text", + "content": ", codebook " + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "inline_equation", + "content": "Q_{tgt}" + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "text", + "content": ", and decoder " + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "inline_equation", + "content": "G_{tgt}" + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "text", + "content": " in the VQ-VAE. For the source domain, we initialize its " + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "inline_equation", + "content": "E_{src}" + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "inline_equation", + "content": "G_{src}" + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "inline_equation", + "content": "E_{tgt}" + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "inline_equation", + "content": "G_{tgt}" + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "text", + "content": " while reusing the frozen codebook " + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "inline_equation", + "content": "Q_{tgt}" + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "text", + "content": ". We then learn the discriminator " + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "text", + "content": " and the source domain's encoder " + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "inline_equation", + "content": "E_{src}" + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "inline_equation", + "content": "G_{src}" + }, + { + "bbox": [ + 313, + 168, + 556, + 371 + ], + "type": "text", + "content": " in an interleaving fashion by minimizing the following objective functions:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 319, + 379, + 554, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 379, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 319, + 379, + 554, + 407 + ], + "type": "interline_equation", + "content": "\\begin{array}{r l} \\text {F o r} D: \\mathcal {L} _ {D} = & \\mathbb {E} _ {\\boldsymbol {x} \\sim P _ {t g t}} [ \\max (0, 1 - D (E _ {t g t} (\\boldsymbol {x}))) ] \\\\ & + \\mathbb {E} _ {\\boldsymbol {x} \\sim P _ {s r c}} [ \\max (0, 1 + D (E _ {s r c} (\\boldsymbol {x}))) ]; \\end{array} \\tag {8}", + "image_path": "11fa0cd4732388269d77d3b5e1126e4426a7ea6858ce356847ff0da4665ec362.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 416, + 554, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 416, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 315, + 416, + 554, + 449 + ], + "type": "interline_equation", + "content": "\\text {F o r V Q - V A E :} \\mathcal {L} _ {\\mathrm {v q}} = \\mathcal {L} _ {\\mathrm {r e c}} + \\mathcal {L} _ {s r c} + \\| \\operatorname {s g} [ \\boldsymbol {x} ^ {z} ] - E _ {s r c} (\\boldsymbol {x}) \\| _ {2} ^ {2}, \\tag {9}", + "image_path": "02842dcf3163b96002c977e448ceb518362614b044c8a6b9d571733f9807d416.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 375, + 453, + 554, + 477 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 453, + 554, + 477 + ], + "spans": [ + { + "bbox": [ + 375, + 453, + 554, + 477 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} \\mathcal {L} _ {s r c} = - \\mathbb {E} _ {\\boldsymbol {x} \\sim P _ {s r c}} D \\left(E _ {s r c} (\\boldsymbol {x})\\right). \\tag {10}", + "image_path": "ee865efbbcedd6f7469133a3e5e2c2ab140279b787fc977ca6ad1d9712dc9249.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "text", + "content": "The other terms in Eq. (9) follow those in Eq. (1), where for " + }, + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\sim P_{src}" + }, + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}} = G_{src} \\circ Q_{tgt} \\circ E_{src}(\\pmb{x})" + }, + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "text", + "content": ". We note that " + }, + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "text", + "content": " takes in two streams of data (i.e., source and target) and is defined by the Hinge Loss [22] on " + }, + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "inline_equation", + "content": "\\pmb{x}^f" + }, + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{vq}}" + }, + { + "bbox": [ + 313, + 486, + 554, + 557 + ], + "type": "text", + "content": " takes in only the source CAV data. We also note that embeddings for object locations does not require adaptation." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 558, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 617 + ], + "type": "text", + "content": "After this adaptation step, we proceed into the second stage of generative training, where we drop the discriminator and train the control module defined in Eq. (5) by minimizing Eq. (7). The feature embeddings are produced by " + }, + { + "bbox": [ + 313, + 558, + 554, + 617 + ], + "type": "inline_equation", + "content": "E_{src}" + }, + { + "bbox": [ + 313, + 558, + 554, + 617 + ], + "type": "text", + "content": " on top of " + }, + { + "bbox": [ + 313, + 558, + 554, + 617 + ], + "type": "inline_equation", + "content": "(\\pmb{x}_r, \\pmb{x}_e, \\pmb{y}_r) \\sim P_{src}" + }, + { + "bbox": [ + 313, + 558, + 554, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "content": "Enhancement in target domains. With the discriminator and adversarial training, we reduce the gap between source and target domain data, successfully adapting TYP to the single-agent dataset, as shown in Sec. 4.5. To further improve the conditioned generation quality in the target domain, a further fine-tuning stage in it is desired. However, how can we do so without having reference agents' point clouds in the target domain? Here, we propose two simple solutions." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12031" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 281, + 194 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 281, + 194 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 281, + 194 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 281, + 194 + ], + "type": "table", + "html": "
MethodTrain DataAP 0.5
smlall
No Fusionego's gt only0.670.410.130.40
Early Fusion [4]gt (oracle)0.760.420.310.49
TYP (ours)0.750.380.290.46
Late Fusion [51]gt (oracle)0.740.570.360.55
TYP (ours)0.710.490.320.50
AttFuse [51]gt (oracle)0.940.800.620.77
TYP (ours)0.900.730.560.72
V2X-ViT [50]gt (oracle)0.870.710.500.71
TYP (ours)0.840.650.400.65
", + "image_path": "32b1667105da42b55656c8c0ec5bcd85263a31a5087320a91863b01d07e62b44.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 232, + 138, + 270 + ], + "blocks": [ + { + "bbox": [ + 61, + 232, + 138, + 270 + ], + "lines": [ + { + "bbox": [ + 61, + 232, + 138, + 270 + ], + "spans": [ + { + "bbox": [ + 61, + 232, + 138, + 270 + ], + "type": "image", + "image_path": "259fc6e51134ec81f260d71466fd8d7df89b8d55a1d9c6dca094290ed2cb6246.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 270, + 121, + 277 + ], + "lines": [ + { + "bbox": [ + 77, + 270, + 121, + 277 + ], + "spans": [ + { + "bbox": [ + 77, + 270, + 121, + 277 + ], + "type": "text", + "content": "(a) no adaptation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 279, + 295, + 312 + ], + "lines": [ + { + "bbox": [ + 55, + 279, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 295, + 312 + ], + "type": "text", + "content": "Figure 3. Qualitative results on enhancement in the target domain. Generated point cloud (green) has better quality with the enhancement given ego (gray) from Waymo (cf. Sec. 3.4)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 138, + 232, + 212, + 270 + ], + "blocks": [ + { + "bbox": [ + 138, + 232, + 212, + 270 + ], + "lines": [ + { + "bbox": [ + 138, + 232, + 212, + 270 + ], + "spans": [ + { + "bbox": [ + 138, + 232, + 212, + 270 + ], + "type": "image", + "image_path": "51fdfe250768ef030b1013c443464af5359bfc7cbf9bea5b34c62fac5d788731.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 270, + 185, + 277 + ], + "lines": [ + { + "bbox": [ + 163, + 270, + 185, + 277 + ], + "spans": [ + { + "bbox": [ + 163, + 270, + 185, + 277 + ], + "type": "text", + "content": "(b) disc." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 212, + 232, + 288, + 270 + ], + "blocks": [ + { + "bbox": [ + 212, + 232, + 288, + 270 + ], + "lines": [ + { + "bbox": [ + 212, + 232, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 212, + 232, + 288, + 270 + ], + "type": "image", + "image_path": "e3b62f81cc90d270c5388e87d359101ffd43b7f09e619da9d8210b2dd9fd2f5f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 270, + 280, + 277 + ], + "lines": [ + { + "bbox": [ + 217, + 270, + 280, + 277 + ], + "spans": [ + { + "bbox": [ + 217, + 270, + 280, + 277 + ], + "type": "text", + "content": "(c) disc. + enhancement" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 319, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 319, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 54, + 319, + 296, + 510 + ], + "type": "text", + "content": "Recall that in training " + }, + { + "bbox": [ + 54, + 319, + 296, + 510 + ], + "type": "inline_equation", + "content": "P(\\pmb{x}_r|\\pmb{x}_e,\\pmb{y}_r)" + }, + { + "bbox": [ + 54, + 319, + 296, + 510 + ], + "type": "text", + "content": " using the source CAV data, point clouds show up in both the input as conditions and the outputs as supervisions. In other words, the domain gap exists on both sides, and we aim to reduce it by further fine-tuning using the target single-agent dataset. First, to enhance the output generation in the target domain, we set ego's point clouds to empty, i.e., " + }, + { + "bbox": [ + 54, + 319, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\pmb{x}_e = 0^{H\\times W\\times C}" + }, + { + "bbox": [ + 54, + 319, + 296, + 510 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 54, + 319, + 296, + 510 + ], + "type": "inline_equation", + "content": "(\\pmb{x}_r,\\pmb{y}_r)\\sim P_{tgt}" + }, + { + "bbox": [ + 54, + 319, + 296, + 510 + ], + "type": "text", + "content": ". Then, to enable adaptation on inputs, we apply self-training [18, 28, 60, 61], randomly sampling a reference location in the scene to generate pseudo-point clouds for fine-tuning, i.e., " + }, + { + "bbox": [ + 54, + 319, + 296, + 510 + ], + "type": "inline_equation", + "content": "(\\hat{\\pmb{x}}_r,\\pmb{x}_e,\\pmb{y}_r)" + }, + { + "bbox": [ + 54, + 319, + 296, + 510 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 54, + 319, + 296, + 510 + ], + "type": "inline_equation", + "content": "(\\pmb{x}_e,\\pmb{y}_r)\\sim P_{tgt}" + }, + { + "bbox": [ + 54, + 319, + 296, + 510 + ], + "type": "text", + "content": ". We use the above target domain data and the source domain data to jointly fine-tune the control module (cf. Sec. 3.3, Eq. (6), and Eq. (7)) further. Empirical results in Sec. 4.5 show the effectiveness of the proposed solution in turning single-agent datasets into collaborative datasets for CAV development." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 521, + 137, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 137, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 137, + 534 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 540, + 105, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 105, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 105, + 553 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 558, + 297, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 297, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 297, + 701 + ], + "type": "text", + "content": "Datasets. In this study, we primarily use three datasets: OPV2V [51], V2V4Real [52], and WOD [39]. OPV2V is a simulation-based CAV dataset that includes over 70 diverse driving scenes and more than 11,000 frames. We utilize this dataset initially to validate our proposed research problem and later to guide the generation training on single-agent datasets. V2V4Real is a real-world CAV dataset with approximately 20,000 LiDAR frames, which we use to validate TYP in a real-world CAV context, treating it as a single-agent dataset (i.e., without using any labels to train TYP). Lastly, we adopt WOD to scale up TYP, one of the largest real-world single-agent datasets available." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": "Evaluation metrics. Following standard benchmarks [51," + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 329, + 70, + 540, + 194 + ], + "blocks": [ + { + "bbox": [ + 55, + 198, + 295, + 231 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 295, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 295, + 231 + ], + "type": "text", + "content": "Table 1. Results on OPV2V. We compare the performance using reference generated by TYP versus ground-truth (oracle). The comparable results demonstrate the quality of our generated data." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 329, + 70, + 540, + 194 + ], + "lines": [ + { + "bbox": [ + 329, + 70, + 540, + 194 + ], + "spans": [ + { + "bbox": [ + 329, + 70, + 540, + 194 + ], + "type": "table", + "html": "
MethodTrain DataAP 0.5
smlall
No Fusionego's gt only0.710.330.080.46
Early Fusion [4]gt (oracle)0.790.420.450.60
TYP (ours)0.760.360.300.53
Late Fusion [51]gt (oracle)0.790.530.560.67
TYP (ours)0.770.480.510.63
AttFuse [51]gt (oracle)0.800.450.360.61
TYP (ours)0.750.450.260.56
V2X-ViT [50]gt (oracle)0.810.490.300.61
TYP (ours)0.760.380.170.48
", + "image_path": "f560e44fcac80bb2aa9f7d4fbb29b08b37623adda285365658a4a3bdfd4c9828.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 198, + 554, + 253 + ], + "lines": [ + { + "bbox": [ + 313, + 198, + 554, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 554, + 253 + ], + "type": "text", + "content": "Table 2. Results on V2V4Real. We compare the performance using reference generated by TYP versus ground-truth (oracle). Note that we do not use V2V4Real data to train the generation, treating it as a single-agent dataset. The comparable results demonstrate the quality of our generated data and the transferability of TYP." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 258, + 555, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 557 + ], + "type": "text", + "content": "52], we report Average Precision (AP 0.5) across three distance ranges: 0-30 meters (short, denoted as s), 30-50 meters (medium, m), and 50-100 meters (long, l) and overall (all). Implementations. For TYP training, we voxelize point clouds within the range " + }, + { + "bbox": [ + 313, + 258, + 555, + 557 + ], + "type": "inline_equation", + "content": "[-51.2, 51.2]" + }, + { + "bbox": [ + 313, + 258, + 555, + 557 + ], + "type": "text", + "content": " meters in x and y dimensions and 4 meters in z (range is dataset-dependent) to create 3D volumes, " + }, + { + "bbox": [ + 313, + 258, + 555, + 557 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{512 \\times 512 \\times 20}" + }, + { + "bbox": [ + 313, + 258, + 555, + 557 + ], + "type": "text", + "content": ". This is then encoded to " + }, + { + "bbox": [ + 313, + 258, + 555, + 557 + ], + "type": "inline_equation", + "content": "x^f \\in \\mathbb{R}^{64 \\times 64 \\times 8}" + }, + { + "bbox": [ + 313, + 258, + 555, + 557 + ], + "type": "text", + "content": " with VQ-VAE [41]. Similarly, semantic information " + }, + { + "bbox": [ + 313, + 258, + 555, + 557 + ], + "type": "inline_equation", + "content": "y \\in \\mathbb{R}^{512 \\times 512 \\times 1}" + }, + { + "bbox": [ + 313, + 258, + 555, + 557 + ], + "type": "text", + "content": " is encoded to " + }, + { + "bbox": [ + 313, + 258, + 555, + 557 + ], + "type": "inline_equation", + "content": "y^f \\in \\mathbb{R}^{64 \\times 64 \\times 3}" + }, + { + "bbox": [ + 313, + 258, + 555, + 557 + ], + "type": "text", + "content": ". We train VQ-VAE for " + }, + { + "bbox": [ + 313, + 258, + 555, + 557 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 258, + 555, + 557 + ], + "type": "text", + "content": " 120 epochs and for " + }, + { + "bbox": [ + 313, + 258, + 555, + 557 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 258, + 555, + 557 + ], + "type": "text", + "content": " 10 epochs, with a batch size of 6 per GPU. Next, we train LDM with a batch size of 48 per GPU for 200 epochs in the first stage and 80 epochs in the second stage. During the second stage, we freeze the U-Net from the first stage and fine-tune only the control module, following [26]. For adaptation, we initialize the weights of encoder-decoder for CAV data (fake) with the model trained on single-agent data (real), freeze its codebook, and fine-tune it with a discriminator comprising 4 convolutional layers. The model processes an equal amount of fake and real per batch (i.e., 3 samples each). All training is conducted on 4 NVIDIA H100 GPUs. For experiments on benchmarks, we follow the convention to report results with Pointpillars [17]. We train / fine-tune all models with 8 P100, with batch size 4 per GPU. For V2X-ViT, we use 4 H100 due to resources required." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 558, + 556, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 666 + ], + "type": "text", + "content": "Inference of generation and post-processing. We use DDIM scheduler [38] for 200 steps. For post-processing, we apply modified Gumbel-Softmax [15] to the logits. Specifically, we define two thresholds, low and high. Scores below the low threshold are dropped, those above the high threshold are retained, and values in between are perturbed with Gaussian noise on logits. Hyperparameters are tuned on a small hold-out set. Finally, we convert the volume back to coordinates by using the centers of the occupied voxels." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 672, + 520, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 520, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 520, + 685 + ], + "type": "text", + "content": "4.2. Do We Need Collaborative Perception?" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "To explore this question, we consistently report the performance without collaborative agents (denoted as \"ego's gt" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12032" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 70, + 526, + 232 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 526, + 232 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 526, + 232 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 526, + 232 + ], + "type": "table", + "html": "
MethodPre-TrainFT. on 0 SceneFT. on 5 SceneFT. on 10 SceneFT. on 32 Scene
smlallsmlallsmlallsmlall
No Fusionego's gt only0.440.160.060.290.660.290.080.420.710.340.080.47
scratch0.440.180.160.300.730.280.240.480.790.420.450.60
Early Fusion [4]OPV2V0.540.200.070.310.320.180.150.250.730.320.310.500.800.470.540.65
ColWaymo (ours)0.500.240.240.350.680.340.320.510.780.390.330.570.830.480.500.65
Late Fusion [51]scratch0.440.240.330.340.720.420.510.580.790.530.560.67
OPV2V0.600.270.280.440.550.260.380.420.730.420.510.590.780.550.580.67
ColWaymo (ours)0.400.190.150.250.600.280.440.470.760.430.510.610.820.580.610.71
AttFuse [51]scratch0.400.150.100.280.700.300.170.470.800.450.360.61
OPV2V0.510.190.050.310.540.220.110.370.770.400.210.540.830.530.400.65
ColWaymo (ours)0.660.350.110.450.650.290.160.460.830.460.330.610.880.580.530.72
V2X-ViT [50]scratch0.430.150.120.310.700.280.170.430.810.490.300.61
OPV2V0.510.240.070.330.480.230.160.350.760.380.220.530.810.490.350.61
ColWaymo (ours)0.600.280.100.340.660.280.220.460.790.480.260.580.840.570.440.67
", + "image_path": "9845242fc67e6b238444d31c7165ce1e13b8fab0130a1f033657e5b06422230c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 65, + 283, + 286, + 350 + ], + "blocks": [ + { + "bbox": [ + 55, + 234, + 555, + 279 + ], + "lines": [ + { + "bbox": [ + 55, + 234, + 555, + 279 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 555, + 279 + ], + "type": "text", + "content": "Table 3. Results with pre-training on collaborative Waymo. We scale up TYP with exiting large-scale single-agent dataset WOD [39], creating its collaborative version \"ColWaymo\". These data are used for pre-training and subsequently fine-tuned on V2V4Real [52]. Results show significant improvement over training from scratch, highlighting the potential to reduce data collection efforts, scale up, and accelerate CAV development. Comparisons to simulated OPV2V [51] pre-training further demonstrate the realism of our generated point clouds." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 283, + 286, + 350 + ], + "lines": [ + { + "bbox": [ + 65, + 283, + 286, + 350 + ], + "spans": [ + { + "bbox": [ + 65, + 283, + 286, + 350 + ], + "type": "table", + "html": "
MethodDisc.ST.Dum.JSD (↓)MMD (↓)
No Adaptation0.264.61 × 10-4
✓0.161.10 × 10-4
Adaptation✓✓0.171.30 × 10-4
✓✓✓0.161.17 × 10-4
", + "image_path": "1c310ef162f2f512a5b6fea873dd29a660f70a47194ff47f8b8105a300ed3baa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 65, + 399, + 286, + 438 + ], + "blocks": [ + { + "bbox": [ + 55, + 353, + 296, + 397 + ], + "lines": [ + { + "bbox": [ + 55, + 353, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 296, + 397 + ], + "type": "text", + "content": "Table 4. Ablation of proposed adaption method. Results with adaptation outperform no adaptation, demonstrating an improvement in generation quality. Abbreviations in title refer to discriminator, self-training, and dummy ego, respectively (cf. Sec. 3.4)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 399, + 286, + 438 + ], + "lines": [ + { + "bbox": [ + 65, + 399, + 286, + 438 + ], + "spans": [ + { + "bbox": [ + 65, + 399, + 286, + 438 + ], + "type": "table", + "html": "
Num. StepsEpochs 1stEpochs 2ndJSD (↓)MMD (↓)
1-2800.116.45 × 10-5
2200800.105.28 × 10-5
", + "image_path": "487d6dc4883983568a93ef2f51e7ed52c51f0ed3d9d2f7f7d2e24c24367ae6b6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 441, + 295, + 464 + ], + "lines": [ + { + "bbox": [ + 55, + 441, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 295, + 464 + ], + "type": "text", + "content": "Table 5. 2-Stage training. Results demonstrate the benefit of the proposed 2-stage training in generation quality on OPV2V [51]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 472, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 567 + ], + "type": "text", + "content": "only\" in Tab. 1, Tab. 2, and Tab. 3). For example, in the V2V4Real dataset, a well-trained detector struggles with objects beyond 50 meters, achieving only 0.08 AP at that range, though it performs well on nearby objects with an AP of 0.71 (cf. first row of Tab. 3). simply combining the predictions from two under-trained detectors, we observe a notable boost in performance to 0.33 (cf. \"scratch\" in late fusion, Tab. 2), suggesting the significance of CAV." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 575, + 248, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 248, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 248, + 588 + ], + "type": "text", + "content": "4.3. Point Clouds Generation on OPV2V" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": "Setting. To validate our research concept, we begin with the simulated dataset OPV2V [51]. We split the original training set of 44 scenes into two halves, using only the first 22 scenes to train our generation model. For the remaining 22 scenes, we generate point clouds for reference agents based on the ego agent's point clouds and compare these results to ground-truth point clouds. We follow the benchmark to report performance on the validation set from Culver City." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Baselines. Following the literature [51, 52], we adopt three common fusion algorithms for CAV: early, late, and inter" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 285, + 555, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 321 + ], + "type": "text", + "content": "mediate fusion, which fuse input point clouds, extracted features, and predicted bounding boxes, respectively. For intermediate fusion, we adipot AttFuse [51] and V2X-VIT [50]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 322, + 556, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 556, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 556, + 441 + ], + "type": "text", + "content": "Comparison to using ground-truth LiDAR. In Tab. 1, we show that training CAV with point clouds generated by TYP achieves performance comparable to that obtained with ground-truth (oracle) across all methods (e.g., 0.46 vs. 0.49 with early, 0.50 vs. 0.55 with late, and 0.77 vs. 0.72 with AttFuse [51], etc.) This result highlights the quality of the generated point clouds. In the supplemental material, we further demonstrate that with additional labeled data (from the first half of the split used for training generation), the gap to oracle performance can be further minimized." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 451, + 537, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 537, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 537, + 464 + ], + "type": "text", + "content": "4.4. Single-Agent to Collaborative Multi-Agent" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 469, + 556, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 556, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 556, + 564 + ], + "type": "text", + "content": "Setting. Building on our findings with simulated data (cf. Sec. 4.3), we validate our approach on the real-world dataset V2V4Real [52], treating it as a single-agent dataset without access to reference agents' point clouds. Specifically, we use the model from Sec. 4.3 to generate reference car's point clouds from the ego's perspective and compare performance against ground-truth point clouds (oracle). Note that no V2V4Real data is used to train the generation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 566, + 556, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 566, + 556, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 556, + 662 + ], + "type": "text", + "content": "Results on V2V4Real. In Tab. 2, we show the performance using generated reference point clouds is comparable to the oracle, supporting our idea of translating a single-agent dataset to multi-agent CAV. Xu et al. previously explored OPV2V-to-V2V4Real adaptation, noting significant performance drops even with adaptation algorithms. Here, TYP offers a new solution to this challenge – in a generative way, providing a fresh perspective on domain adaptation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 671, + 537, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 537, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 537, + 685 + ], + "type": "text", + "content": "4.5. ColWaymo: Scaling with Large-Scale Data" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Scaling CAV to the next level. In Sec. 4.4, we show the potential to create realistic reference's point clouds on real-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12033" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 70, + 196, + 177 + ], + "blocks": [ + { + "bbox": [ + 88, + 70, + 196, + 177 + ], + "lines": [ + { + "bbox": [ + 88, + 70, + 196, + 177 + ], + "spans": [ + { + "bbox": [ + 88, + 70, + 196, + 177 + ], + "type": "image", + "image_path": "56023c3d1150e71dbc23b6a087a38fddfb7fd066bd77fbb91652e7a64981f6d8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 178, + 555, + 201 + ], + "lines": [ + { + "bbox": [ + 55, + 178, + 555, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 555, + 201 + ], + "type": "text", + "content": "Figure 4. Qualitative results on Collaborative Waymo. The gray point clouds are from the original single-agent dataset and the green are generated by TYP conditioning on them." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 197, + 70, + 304, + 177 + ], + "blocks": [ + { + "bbox": [ + 197, + 70, + 304, + 177 + ], + "lines": [ + { + "bbox": [ + 197, + 70, + 304, + 177 + ], + "spans": [ + { + "bbox": [ + 197, + 70, + 304, + 177 + ], + "type": "image", + "image_path": "752f1bd8c563280c053b82f26f676cbea0dd1f380c2b393c2bcd1f9f8cecf506.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 304, + 70, + 411, + 177 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 411, + 177 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 411, + 177 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 411, + 177 + ], + "type": "image", + "image_path": "1b05b0743a4f04bdbd509836ecb1bdee5403a67e91a68fbc2f14b5c2c97f81a0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 413, + 70, + 521, + 177 + ], + "blocks": [ + { + "bbox": [ + 413, + 70, + 521, + 177 + ], + "lines": [ + { + "bbox": [ + 413, + 70, + 521, + 177 + ], + "spans": [ + { + "bbox": [ + 413, + 70, + 521, + 177 + ], + "type": "image", + "image_path": "03dfc23b41211493196497bebcb37145a7ac996d23cf359325eacf6e54c73350.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 85, + 204, + 230, + 284 + ], + "blocks": [ + { + "bbox": [ + 85, + 204, + 230, + 284 + ], + "lines": [ + { + "bbox": [ + 85, + 204, + 230, + 284 + ], + "spans": [ + { + "bbox": [ + 85, + 204, + 230, + 284 + ], + "type": "image", + "image_path": "4abe5b5c88316aacf10b2ee4382db43d2c14c062fb5f9619bde0bf0fee779bc8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 231, + 204, + 376, + 283 + ], + "blocks": [ + { + "bbox": [ + 231, + 204, + 376, + 283 + ], + "lines": [ + { + "bbox": [ + 231, + 204, + 376, + 283 + ], + "spans": [ + { + "bbox": [ + 231, + 204, + 376, + 283 + ], + "type": "image", + "image_path": "faff2f785d98135f357ec24971c54b3770e890869d48333deb10e417e9116c2e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 378, + 205, + 524, + 283 + ], + "blocks": [ + { + "bbox": [ + 378, + 205, + 524, + 283 + ], + "lines": [ + { + "bbox": [ + 378, + 205, + 524, + 283 + ], + "spans": [ + { + "bbox": [ + 378, + 205, + 524, + 283 + ], + "type": "image", + "image_path": "65428d9475507f7325174b44ae1fc7cb81bec176637a2fc906af87f5a3d0feef.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 85, + 289, + 230, + 369 + ], + "blocks": [ + { + "bbox": [ + 85, + 289, + 230, + 369 + ], + "lines": [ + { + "bbox": [ + 85, + 289, + 230, + 369 + ], + "spans": [ + { + "bbox": [ + 85, + 289, + 230, + 369 + ], + "type": "image", + "image_path": "061b00aca9e01923d1aa59b2808ff613fb4d4e254bdd9a12fc66bcc777bbdfdd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 373, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 54, + 373, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 555, + 407 + ], + "type": "text", + "content": "Figure 5. Qualitative results. Our proposed TYP is capable of scene editing, by inputting the same point cloud but different object locations. We (b) remove and (c) add a car from (a) the original point cloud. Inspired by the idea of past traversals [56], we apply completely different traffic conditions and generate (d), (e), and (f), to imagine driving through the same intersection." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 231, + 289, + 376, + 368 + ], + "blocks": [ + { + "bbox": [ + 231, + 289, + 376, + 368 + ], + "lines": [ + { + "bbox": [ + 231, + 289, + 376, + 368 + ], + "spans": [ + { + "bbox": [ + 231, + 289, + 376, + 368 + ], + "type": "image", + "image_path": "ee06003fcf8d90f7ea437d7a4a6f313fb2f5c5a67ed47b563e6e0b56a2a2623d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 378, + 290, + 524, + 368 + ], + "blocks": [ + { + "bbox": [ + 378, + 290, + 524, + 368 + ], + "lines": [ + { + "bbox": [ + 378, + 290, + 524, + 368 + ], + "spans": [ + { + "bbox": [ + 378, + 290, + 524, + 368 + ], + "type": "image", + "image_path": "b17e2e0fafe06efd15d5aa42ba408d318b3b5301ad4daa26d3932eb5d590cde6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 414, + 297, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 414, + 297, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 414, + 297, + 544 + ], + "type": "text", + "content": "world dataset V2V4Real [52], through a simulation-guided training of TYP. Here, our goal is broader: to transform an existing large-scale single-agent dataset (e.g., WOD [39]) into a collaborative version. Using the pipeline described in Sec. 3 and illustrated in Fig. 2, we train the generative model on 17,400 samples from the WOD training set. We then use its validation set to generate approximately 36,000 collaborative samples by sampling reference locations from labeled vehicles. These generated samples serve to pre-train a model, which we subsequently fine-tune on V2V4Real [52]. For comparison, we also conduct pre-training on OPV2V [51]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 545, + 297, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 297, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 297, + 688 + ], + "type": "text", + "content": "Results of fine-tuning on V2V4Real. Following prior work [1, 29, 47, 54], we perform pre-training followed by fine-tuning on limited labeled data to validate the potential of TYP for scaling up CAV development. As shown in Tab. 3, pre-training on our generated data yields significantly better performance than training from scratch. Additionally, compared to pre-training on the simulated dataset OPV2V, ColWaymo consistently achieves higher performance, highlighting the realism of the generated point clouds. These results demonstrate that TYP effectively transforms an ego-only dataset into a collaborative version, and potentially reducing the need for extensive CAV data collection." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Generation quality by proposed adaptation. To assess the quality of generated data, we use two common metrics:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 414, + 556, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 556, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 556, + 474 + ], + "type": "text", + "content": "Jensen-Shannon Divergence (JSD) and Maximum Mean Discrepancy (MMD). As shown in Tab. 4, our adaptation method improves JSD from 0.26 to 0.16 and MMD from 4.61 to 1.17 compared to the baseline without adaptation. Additionally, Fig. 3 visually illustrates this improvement." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 481, + 425, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 481, + 425, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 425, + 493 + ], + "type": "text", + "content": "4.6. Qualitative Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 498, + 556, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 498, + 556, + 594 + ], + "spans": [ + { + "bbox": [ + 312, + 498, + 556, + 594 + ], + "type": "text", + "content": "We illustrate the proposed problem in Fig. 1, which highlights the benefits of collaboration by allowing an agent to \"see\" beyond the limitations of its own sensors. TYP generates realistic point clouds from various viewpoints, enabling the agents to perceive previously occluded objects while preserving the overall scene semantics. Fig. 4 shows our ColWaymo, and Fig. 5 further demonstrates the model's capability for scene editing." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 605, + 388, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 388, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 388, + 616 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 617, + 556, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 666 + ], + "type": "text", + "content": "We introduce a new research direction in collaborative driving and present the first solution. Empirical results show our approach can significantly reduce data collection and development efforts, advancing safer autonomous systems." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "content": "Limitations and future work. Following existing benchmarks [51, 52], TYP focuses on vehicle-like objects. Future work could extend it to broader objects and static entities (e.g., traffic signs, signals) essential for real-world traffic." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12034" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 140 + ], + "type": "text", + "content": "This research is supported in part by grants from the National Science Foundation (IIS-2107077, IIS-2107161). We are thankful for the generous support of the computational resources by the Ohio Supercomputer Center." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 114, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 114, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 114, + 163 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 171, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 171, + 296, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 171, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 171, + 296, + 203 + ], + "type": "text", + "content": "[1] Alexandre Boulch, Corentin Sautier, Björn Michele, Gilles Puy, and Renaud Marlet. Also: Automotive lidar self-supervision by occupancy estimation. In CVPR, 2023. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 205, + 296, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 296, + 248 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 296, + 248 + ], + "type": "text", + "content": "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 251, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 296, + 304 + ], + "type": "text", + "content": "[3] Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. In Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, 2019. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 306, + 296, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 306, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 306, + 296, + 350 + ], + "type": "text", + "content": "[4] Qi Chen, Sihai Tang, Qing Yang, and Song Fu. Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds. In International Conference on Distributed Computing Systems (ICDCS). IEEE, 2019. 1, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 351, + 296, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 351, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 351, + 296, + 384 + ], + "type": "text", + "content": "[5] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In CVPR, 2017. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 385, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 385, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 385, + 294, + 407 + ], + "type": "text", + "content": "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 409, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 294, + 441 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. In CoRL, 2017. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 443, + 296, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 443, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 62, + 443, + 296, + 496 + ], + "type": "text", + "content": "[8] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks. Journal of machine learning research, 17(59):1-35, 2016. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 499, + 294, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 499, + 294, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 499, + 294, + 531 + ], + "type": "text", + "content": "[9] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 533, + 296, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 296, + 576 + ], + "type": "text", + "content": "[10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 3, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 578, + 296, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 296, + 631 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 296, + 631 + ], + "type": "text", + "content": "[11] Cole Gulino, Justin Fu, Wenjie Luo, George Tucker, Eli Bronstein, Yiren Lu, Jean Harb, Xinlei Pan, Yan Wang, Xiangyu Chen, et al. Waymax: An accelerated, data-driven simulator for large-scale autonomous driving research. In NeurIPS, 2024. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 634, + 296, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 634, + 296, + 655 + ], + "spans": [ + { + "bbox": [ + 57, + 634, + 296, + 655 + ], + "type": "text", + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 658, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 678 + ], + "type": "text", + "content": "[13] Qianjiang Hu, Zhimin Zhang, and Wei Hu. Rangeldm: Fast realistic lidar point cloud generation. In ECCV, 2025. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "text", + "content": "[14] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. In NeurIPS, 2022. 1" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 95 + ], + "type": "text", + "content": "[15] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In ICLR, 2017. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 96, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 117 + ], + "type": "text", + "content": "[16] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. In NeurIPS, 2021. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 119, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 151 + ], + "type": "text", + "content": "[17] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, 2019. 4, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 152, + 555, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 555, + 195 + ], + "type": "text", + "content": "[18] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, 2013. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 198, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 229 + ], + "type": "text", + "content": "[19] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. In NeurlPS, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 232, + 555, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 555, + 285 + ], + "type": "text", + "content": "[20] Yiming Li, Dekun Ma, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: Multi-agent collaborative perception dataset and benchmark for autonomous driving. In Robotics and Automation Letters, pages 10914-10921. IEEE, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "text", + "content": "[21] Yiyi Liao, Jun Xie, and Andreas Geiger. Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE TPAMI, 45(3):3292-3310, 2022. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 321, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 342 + ], + "type": "text", + "content": "[22] Jae Hyun Lim and Jong Chul Ye. Geometric gan. arXiv preprint arXiv:1705.02894, 2017. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 343, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 376 + ], + "type": "text", + "content": "[23] Yen-Cheng Liu, Junjiao Tian, Nathaniel Glaser, and Zsolt Kira. When2com: Multi-agent perception via communication graph grouping. In CVPR, 2020. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 377, + 555, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 555, + 409 + ], + "type": "text", + "content": "[24] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In IROS, 2015. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 411, + 555, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 555, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 555, + 442 + ], + "type": "text", + "content": "[25] Andres Milioto, Ignacio Vizzo, Jens Behley, and Cyril Stachniss. Rangenet++: Fast and accurate lidar semantic segmentation. In IROS, 2019. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "type": "text", + "content": "[26] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, 2024. 5, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 490, + 555, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 490, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 490, + 555, + 521 + ], + "type": "text", + "content": "[27] Kazuto Nakashima and Ryo Kurazume. Lidar data synthesis with denoising diffusion probabilistic models. In ICRA, 2024. 2, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 524, + 555, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 524, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 524, + 555, + 555 + ], + "type": "text", + "content": "[28] Tai-Yu Pan, Qing Liu, Wei-Lun Chao, and Brian Price. Towards open-world segmentation of parts. In CVPR, 2023. 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 557, + 555, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 555, + 610 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 555, + 610 + ], + "type": "text", + "content": "[29] Tai-Yu Pan, Chenyang Ma, Tianle Chen, Cheng Perng Phoo, Katie Z Luo, Yurong You, Mark Campbell, Kilian Q Weinberger, Bharath Hariharan, and Wei-Lun Chao. Pre-training lidar-based 3d object detectors through colorization. In ICLR, 2023. 8" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 613, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 644 + ], + "type": "text", + "content": "[30] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 4" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "type": "text", + "content": "[31] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 4" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 681, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 555, + 712 + ], + "type": "text", + "content": "[32] Haoxi Ran, Vitor Guizilini, and Yue Wang. Towards realistic scene generation with lidar diffusion models. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12035" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "text", + "content": "[33] Zaydoun Yahya Rawashdeh and Zheng Wang. Collaborative automated driving: A machine learning-based method to enhance the accuracy of shared information. In International Conference on Intelligent Transportation Systems (ITSC). IEEE, 2018. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 162 + ], + "type": "text", + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 195 + ], + "type": "text", + "content": "[35] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 197, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 197, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 57, + 197, + 295, + 217 + ], + "type": "text", + "content": "[36] T-YLPG Ross and GKHP Dólár. Focal loss for dense object detection. In CVPR, 2017. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 220, + 296, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 220, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 57, + 220, + 296, + 252 + ], + "type": "text", + "content": "[37] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointr-cnn: 3d object proposal generation and detection from point cloud. In CVPR, 2019. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 254, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 254, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 57, + 254, + 294, + 275 + ], + "type": "text", + "content": "[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 277, + 296, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 296, + 330 + ], + "type": "text", + "content": "[39] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 333, + 296, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 333, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 57, + 333, + 296, + 365 + ], + "type": "text", + "content": "[40] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In CVPR, 2017. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 366, + 294, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 366, + 294, + 388 + ], + "spans": [ + { + "bbox": [ + 57, + 366, + 294, + 388 + ], + "type": "text", + "content": "[41] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. In NeurIPS, 2017. 4, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 389, + 296, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 389, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 57, + 389, + 296, + 432 + ], + "type": "text", + "content": "[42] Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In ECCV, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 434, + 296, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 434, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 57, + 434, + 296, + 477 + ], + "type": "text", + "content": "[43] Bichen Wu, Alvin Wan, Xiangyu Yue, and Kurt Keutzer. Squeezeseg: Convolutional neural nets with recurrent crf for real-time road-object segmentation from 3d lidar point cloud. In ICRA, 2018. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 479, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 479, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 57, + 479, + 296, + 521 + ], + "type": "text", + "content": "[44] Yang Wu, Kaihua Zhang, Jianjun Qian, Jin Xie, and Jian Yang. Text2lidar: Text-guided lidar point cloud generation via equirectangular transformer. In ECCV. Springer, 2024. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 524, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 524, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 57, + 524, + 296, + 578 + ], + "type": "text", + "content": "[45] Hao Xiang, Zhaoliang Zheng, Xin Xia, Runsheng Xu, Letian Gao, Zewei Zhou, Xu Han, Xinkai Ji, Mingxi Li, Zonglin Meng, et al. V2x-real: a largs-scale dataset for vehicle-to-everything cooperative perception. arXiv preprint arXiv:2403.16034, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 580, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 296, + 634 + ], + "type": "text", + "content": "[46] Pengchuan Xiao, Zhenlei Shao, Steven Hao, Zishuo Zhang, Xiaolin Chai, Judy Jiao, Zesong Li, Jian Wu, Kai Sun, Kun Jiang, et al. Pandaset: Advanced sensor suite dataset for autonomous driving. In International Intelligent Transportation Systems Conference (ITSC). IEEE, 2021. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 636, + 296, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 296, + 677 + ], + "type": "text", + "content": "[47] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In ECCV, 2020. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "text", + "content": "[48] Yuwen Xiong, Wei-Chiu Ma, Jingkang Wang, and Raquel Urtasun. Learning compact representations for lidar completion and generation. In CVPR, 2023. 3, 4" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 657 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "text", + "content": "[49] Yuwen Xiong, Wei-Chiu Ma, Jingkang Wang, and Raquel Urtasun. Ultralidar: Learning compact representations for lidar completion and generation. In CVPR, 2023. 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 108, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 555, + 150 + ], + "type": "text", + "content": "[50] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In ECCV, 2022. 1, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 152, + 555, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 555, + 195 + ], + "type": "text", + "content": "[51] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In ICRA, 2022. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 198, + 555, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 555, + 251 + ], + "type": "text", + "content": "[52] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In CVPR, 2023. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 253, + 555, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 555, + 283 + ], + "type": "text", + "content": "[53] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 286, + 555, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 555, + 329 + ], + "type": "text", + "content": "[54] Junbo Yin, Dingfu Zhou, Liangjun Zhang, Jin Fang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Proposal-contrast: Unsupervised pre-training for lidar-based 3d object detection. In ECCV, 2022.8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 331, + 554, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 554, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 554, + 374 + ], + "type": "text", + "content": "[55] Jinsu Yoo, Zhenyang Feng, Tai-Yu Pan, Yihong Sun, Cheng Perng Phoo, Xiangyu Chen, Mark Campbell, Kilian Q Weinberger, Bharath Hariharan, and Wei-Lun Chao. Learning 3d perception from others' predictions. In ICLR, 2025. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 376, + 555, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 376, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 317, + 376, + 555, + 420 + ], + "type": "text", + "content": "[56] Yurong You, Katie Z Luo, Xiangyu Chen, Junan Chen, WeiLun Chao, Wen Sun, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Hindsight is 20/20: Leveraging past traversals to aid 3d perception. In ICLR, 2022. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 421, + 555, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 421, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 317, + 421, + 555, + 474 + ], + "type": "text", + "content": "[57] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, and Zaiqing Nie. Dair-v2x: A large-scale dataset for vehicleinfrastructure cooperative 3d object detection. In CVPR, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 476, + 554, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 476, + 554, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 476, + 554, + 509 + ], + "type": "text", + "content": "[58] Chris Zhang, Wenjie Luo, and Raquel Urtasun. Efficient convolutions for real-time semantic segmentation of 3d point clouds. In 3DV, 2018. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 511, + 554, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 554, + 533 + ], + "type": "text", + "content": "[59] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In CVPR, 2018. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 534, + 555, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 555, + 566 + ], + "type": "text", + "content": "[60] Barret Zoph, Golnaz Ghiasi, Tsung-Yi Lin, Yin Cui, Hanxiao Liu, Ekin Dogus Cubuk, and Quoc Le. Rethinking pretraining and self-training. In NeurIPS, 2020. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 567, + 555, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 555, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 555, + 600 + ], + "type": "text", + "content": "[61] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In ECCV, 2018. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 601, + 554, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 554, + 624 + ], + "type": "text", + "content": "[62] Vlas Zyrianov, Xiyue Zhu, and Shenlong Wang. Learning to generate realistic lidar point clouds. In ECCV, 2022. 2, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 624, + 555, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 555, + 657 + ], + "type": "text", + "content": "[63] Vlas Zyrianov, Henry Che, Zhijian Liu, and Shenlong Wang. Lidardm: Generative lidar simulation in a generated world. arXiv preprint arXiv:2404.02903, 2024. 2, 3" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12036" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Transformers without Normalization/05ce47c0-7126-423f-aeff-039d314c6fa4_content_list.json b/2025/Transformers without Normalization/05ce47c0-7126-423f-aeff-039d314c6fa4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b3aa88886f53b346104eb583e3cfb4a36baa80d4 --- /dev/null +++ b/2025/Transformers without Normalization/05ce47c0-7126-423f-aeff-039d314c6fa4_content_list.json @@ -0,0 +1,1635 @@ +[ + { + "type": "text", + "text": "Transformers without Normalization", + "text_level": 1, + "bbox": [ + 310, + 130, + 687, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiachen Zhu $^{1,2}$ Xinlei Chen $^{1}$ Kaiming He $^{3}$ Yann LeCun $^{1,2}$ Zhuang Liu $^{1,4,\\dagger}$", + "bbox": [ + 183, + 178, + 812, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$\\dagger$ project lead", + "bbox": [ + 459, + 207, + 539, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ FAIR, Meta $^{2}$ New York University $^{3}$ MIT $^{4}$ Princeton University", + "bbox": [ + 223, + 229, + 772, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 282, + 326, + 299 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Normalization layers are ubiquitous in modern neural networks and have long been considered essential. This work demonstrates that Transformers without normalization can achieve the same or better performance using a remarkably simple technique. We introduce Dynamic Tanh (DyT), an element-wise operation $\\mathrm{DyT}(\\boldsymbol{x}) = \\tanh(\\alpha \\boldsymbol{x})$ , as a drop-in replacement for normalization layers in Transformers. DyT is inspired by the observation that layer normalization in Transformers often produces tanh-like, $S$ -shaped input-output mappings. By incorporating DyT, Transformers without normalization can match or exceed the performance of their normalized counterparts, mostly without hyperparameter tuning. We validate the effectiveness of Transformers with DyT across diverse settings, from recognition to generation, supervised to self-supervised learning, and computer vision to language models. These findings challenge the conventional understanding that normalization layers are indispensable in modern neural networks, and offer new insights into their role in deep networks.", + "bbox": [ + 88, + 315, + 485, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 632, + 220, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Over the past decade, normalization layers have solidified their positions as one of the most fundamental components of modern neural networks. It all traces back to the invention of batch normalization in 2015 [41], which enabled drastically faster and better convergence in visual recognition models and quickly gained momentum in the following years. Since then, many variants of normalization layers have been proposed for different network architectures or domains [5, 81, 83, 89]. Today, virtually all modern networks use normalization layers, with layer normalization (Layer Norm, or LN) [5] being one of the most popular, particularly in the dominant Transformer architecture [20, 82].", + "bbox": [ + 89, + 657, + 482, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The widespread adoption of normalization layers is largely driven by their empirical benefits in optimization [10, 68]. In addition to achieving better results, they help accelerate and stabilize convergence. As neural networks", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/41765ed66bf18a82ea42bdaf38ecff4ef1c83483b704ee0c0ee6722d766516b4.jpg", + "image_caption": [ + "Figure 1. Left: original Transformer block. Right: block with our proposed Dynamic Tanh (DyT) layer. DyT is a straightforward replacement for commonly used Layer Norm [5] (in some cases RMSNorm [89]) layers. Transformers with DyT match or exceed the performance of their normalized counterparts." + ], + "image_footnote": [], + "bbox": [ + 529, + 282, + 890, + 407 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "become wider and deeper, this necessity becomes ever more critical [11, 38]. Consequently, normalization layers are widely regarded as crucial, if not indispensable, for effective training of deep networks. This belief is subtly evidenced by the fact that, in recent years, novel architectures often seek to replace attention or convolution layers [22, 29, 75, 78], but almost always retain normalization.", + "bbox": [ + 511, + 492, + 906, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper challenges this belief by introducing a simple alternative to normalization layers in Transformers. Our exploration starts with the observation that LN layers map their inputs to outputs with tanh-like, $S$ -shaped curves, scaling the input activations while squashing the extreme values. Inspired by this insight, we propose an elementwise operation termed Dynamic Tanh (DyT), defined as: $\\mathrm{DyT}(\\boldsymbol{x}) = \\tanh(\\alpha \\boldsymbol{x})$ , where $\\alpha$ is a learnable parameter. This operation aims to emulate the behavior of LN by learning an appropriate scaling factor through $\\alpha$ and squashing extreme values via the bounded tanh function. Notably, unlike normalization layers, it achieves both effects without the need to compute activation statistics.", + "bbox": [ + 511, + 606, + 908, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Employing DyT is straightforward, as shown in Figure 1: we directly replace existing normalization layers with DyT in architectures such as vision and language Transformers. We empirically demonstrate that models with DyT can train stably and achieve high final performance across a wide range of settings. It often does not require tuning the hy", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14901", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/090957d29502bab3ea90da5d620cdba519e426f606475355bf84712d93e89b94.jpg", + "image_caption": [ + "LN output (y axis) vs. LN input (x axis)", + "Figure 2. Output vs. input of selected layer normalization (LN) layers in Vision Transformer (ViT) [20], wav2vec 2.0 (a Transformer model for speech) [7], and Diffusion Transformer (DiT) [64]. We sample a mini-batch of samples and plot the input / output values of four LN layers in each model. The outputs are before the affine transformation in LN. The $S$ -shaped curves highly resemble that of a tanh function. The more linear shapes in earlier layers can also be captured by the center part of a tanh curve. This motivates us to propose Dynamic Tanh (DyT) as a replacement, with a learnable scalar $\\alpha$ to account for different scales on the $x$ axis." + ], + "image_footnote": [], + "bbox": [ + 86, + 104, + 893, + 445 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "perparameters on the original architecture. Our work challenges the notion that normalization layers are indispensable for training modern neural networks and provides empirical insights into the properties of normalization layers.", + "bbox": [ + 88, + 527, + 482, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background: Normalization Layers", + "text_level": 1, + "bbox": [ + 89, + 599, + 415, + 618 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We begin by reviewing the normalization layers. Most normalization layers share a common formulation. Given an input $\\pmb{x}$ with shape $(B, T, C)$ , where $B$ is the batch size, $T$ is the number of tokens, and $C$ is the embedding dimension per token, the output is generally computed as:", + "bbox": [ + 89, + 626, + 483, + 702 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {n o r m a l i z a t i o n} (\\boldsymbol {x}) = \\gamma * \\left(\\frac {\\boldsymbol {x} - \\boldsymbol {\\mu}}{\\sqrt {\\sigma^ {2} + \\epsilon}}\\right) + \\beta \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 709, + 482, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\epsilon$ is a small constant, and $\\gamma$ and $\\beta$ are learnable vector parameters of shape $(C,)$ . They are \"scaling\" and \"shifting\" affine parameters that allow the output to be in any range. The terms $\\mu$ and $\\sigma^2$ denote the mean and variance of the input. Different methods mainly differ in how these two statistics are computed. This results in $\\mu$ and $\\sigma^2$ having different dimensions, each with appropriate broadcasting needed in the tensor calculation.", + "bbox": [ + 89, + 750, + 482, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Batch normalization (BN) [41] is the first modern normalization layer, and it has been primarily used in Con", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "vNet models [33, 76, 84]. Its introduction represents a major milestone in deep learning architecture designs. BN computes the mean and variance across both the batch and token dimensions, specifically: $\\mu_{k} = \\frac{1}{BT}\\sum_{i,j}x_{ijk}$ and $\\sigma_k^2 = \\frac{1}{BT}\\sum_{i,j}(x_{ijk} - \\mu_k)^2$ . Other normalization layers popular in ConvNets, such as group normalization [83] and instance normalization [81], were initially proposed for specialized tasks such as object detection and image stylization. They share the same overall formulation but differ in the axes and ranges over which the statistics are computed.", + "bbox": [ + 511, + 527, + 906, + 683 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Layer normalization (LN) [5] and root mean square normalization (RMSNorm) [89] are the major two types of normalization layers used in Transformer architectures. LN computes these statistics independently for each token in each sample, where $\\mu_{ij} = \\frac{1}{C}\\sum_k x_{ijk}$ and $\\sigma_{ij}^2 = \\frac{1}{C}\\sum_k(x_{ijk} - \\mu_{ij})^2$ . RMSNorm [89] simplifies LN by removing the mean-centering step and normalizing the input with $\\mu_{ij} = 0$ and $\\sigma_{ij}^2 = \\frac{1}{C}\\sum_k x_{ijk}^2$ . Today, most modern neural networks use LN due to its simplicity and universality. Recently, RMSNorm has gained popularity, particularly in language models like T5 [66], LLaMA [21, 79, 80], Mistral [43], Qwen [8, 87], InternLM [13, 91] and DeepSeek [30, 49]. The Transformers we examine in this work all use LN, except that LLaMA uses RMSNorm.", + "bbox": [ + 511, + 686, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14902", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/41a32d50a2a8c9a3a51d104555d6347f8be77b6e90e8e8694681af64cd8ac114.jpg", + "image_caption": [ + "Figure 3. Output vs. input of two LN layers, with tensor elements colored to indicate different channel and token dimensions. The input tensor has a shape of (samples, tokens, and channels), with elements visualized by assigning consistent colors to the same tokens (left two panels) and channels (right two panels). Left two panels: points representing the same token (same color) form straight lines across different channels, as LN operates linearly across channels for each token. Interestingly, when plotted collectively, these lines form a non-linear tanh-shaped curve. Right two panels: each channel's input spans different ranges on the $x$ -axis, contributing distinct segments to the overall tanh-shaped curve. Certain channels (e.g., red, green, and pink) exhibit more extreme $x$ values, which are squashed by LN." + ], + "image_footnote": [], + "bbox": [ + 84, + 70, + 890, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. What Do Normalization Layers Do?", + "text_level": 1, + "bbox": [ + 89, + 300, + 416, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Analysis setup. We first empirically study the behaviors of normalization layers in trained networks. For this analysis, we take a Vision Transformer model (ViT-B) [20] trained on ImageNet-1K [19], a wav2vec 2.0 Large Transformer model [7] trained on LibriSpeech [63], and a Diffusion Transformer (DiT-XL) [64] trained on ImageNet-1K. In all cases, LN is applied in every Transformer block and before the final linear projection.", + "bbox": [ + 89, + 330, + 483, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For all three trained networks, we sample a mini-batch of samples and do a forward pass through the network. We then measure the input and output for the normalization layers, i.e., tensors immediately before and after the normalization operation, before the learnable affine transformation. Since LN preserves the dimensions of the input tensor, we can establish a one-to-one correspondence between input and output elements, allowing a direct visualization of their relationship. We plot the resulting mappings in Figure 2.", + "bbox": [ + 89, + 454, + 483, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Tanh-like mappings with layer normalization. For all three models, in earlier LN layers (1st column of Figure 2), we find this input-output relationship to be mostly linear, resembling a straight line in an $x - y$ plot. However, deeper LN layers are where we make more intriguing observations.", + "bbox": [ + 89, + 595, + 483, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A striking observation from these deeper layers is that most of these curves' shapes resemble full or partial $S$ -shaped curves represented by a tanh function. One might expect LN layers to linearly transform the input tensor, as subtracting the mean and dividing by standard deviation are linear operations. LN normalizes in a per-token manner, only linearly transforming each token's activations. As tokens have different mean and standard deviation values, the linearity does not hold collectively on all activations of the input tensor. Nonetheless, it is surprising that the non-linear transformation is highly similar to a scaled tanh function.", + "bbox": [ + 89, + 672, + 483, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For such an $S$ -shaped curve, we note that the central part, represented by points with $x$ values close to zero, is mainly in a linear shape. Most points $(\\sim 99\\%)$ fall in this linear range. However, there are many points that clearly fall out", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of this range, which are considered to have \"extreme\" values, e.g., those with $x$ larger than 50 or smaller than -50 in the ViT model. Normalization layers' main effect for these values is to squash them into less extreme values, more in line with the majority of points. This is where normalization layers could not approximated by a simple affine transformation layer. We hypothesize this non-linear and disproportional squashing effect on extreme values is what makes normalization layers important and indispensable.", + "bbox": [ + 511, + 301, + 906, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent findings by Ni et al. [62] similarly highlight the strong non-linearities introduced by LN layers, demonstrating how the non-linearity enhances a model's representational capacity. Moreover, this squashing behavior mirrors the saturation properties of biological neurons for large inputs, a phenomenon observed about a century ago [1-3].", + "bbox": [ + 511, + 439, + 906, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Normalization by tokens and channels. How does an LN layer perform a linear transformation for each token but also squash the extreme values in such a non-linear fashion? To understand this, we visualize the points grouped by tokens and channels, respectively. This is plotted in Figure 3 by taking the second and third subplots for ViT from Figure 2, but with a sampled subset of points for more clarity. When we select the channels to plot, we make sure to include the channels with extreme values.", + "bbox": [ + 511, + 536, + 908, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "On the left two panels of Figure 3, we visualize each token's activations using the same color. We observe that all points from any single token do form a straight line. However, since each token has a different variance, the slopes are different. Tokens with smaller input $x$ ranges tend to have smaller variance, and the normalization layer will divide their activations using a smaller standard deviation, hence producing a larger slope in the straight line. Collectively, they form an $S$ -shaped curve that resembles a tanh function. In the two panels on the right, we color each channel's activations using the same color. We find that different channels tend to have drastically different input ranges, with only a few channels (e.g., red, green, and pink) exhibiting large extreme values. These are the channels that get squashed the most by the normalization layer.", + "bbox": [ + 511, + 674, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14903", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Dynamic Tanh (DyT)", + "text_level": 1, + "bbox": [ + 89, + 89, + 294, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by the similarity between shapes of normalization layers and a scaled tanh function, we propose Dynamic Tanh (DyT) as a replacement for normalization layers. Given input tensor $\\pmb{x}$ , a DyT layer is defined as follows:", + "bbox": [ + 89, + 114, + 482, + 175 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {D y T} (\\boldsymbol {x}) = \\boldsymbol {\\gamma} * \\tanh (\\alpha \\boldsymbol {x}) + \\boldsymbol {\\beta} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 183, + 482, + 199 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ is a learnable scalar parameter that allows scaling the input differently based on its range, accounting for varying $x$ scales (Figure 2). This is also why we name the whole operation \"Dynamic\" Tanh. $\\gamma$ and $\\beta$ are learnable, per-channel vector parameters, the same as those used in all normalization layers—they allow the output to scale back to any scales. This is sometimes considered a separate affine layer; for our purposes, we consider them to be part of the DyT layer, just like how normalization layers also include them. See Algorithm 1 for Pytorch-like pseudocode of DyT.", + "bbox": [ + 89, + 202, + 483, + 354 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Pseudocode of DyT layer." + ], + "code_body": "input x has the shape of [B, T, C] \n# B: batch size, T: tokens, C: dimension \nclass DyT(Module): \n def __init__(self, C, init\\_alpha): \n super().__init() \n self. $\\alpha =$ Parameter(ones(1) $\\star$ init\\_alpha) \n self. $\\gamma =$ Parameter(ones(C)) \n self. $\\beta =$ Parameter(zeros(C)) \ndef forward(self, x): \n x = tanh(self.alpha $\\star$ x) \n return self. $\\gamma =$ x + self. $\\beta$", + "bbox": [ + 109, + 380, + 447, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Integrating DyT layers into an existing architecture is straightforward: one DyT layer replaces one normalization layer (see Figure 1). This applies to normalization layers within attention blocks, FFN blocks, and the final normalization layer. Although DyT may look like or be considered an activation function, this study only uses it to replace normalization layers without altering parts of the activation functions in the original architectures, such as GELU or ReLU. Other parts of the network remain intact. We also observe that there is minimal need to tune the hyperparameters for the original architectures for DyT to perform well.", + "bbox": [ + 89, + 513, + 483, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "On scaling parameters. We always simply initialize $\\gamma$ to an all-one vector and $\\beta$ to an all-zero vector following normalization layers. For the scalar parameter $\\alpha$ , a default initialization of 0.5 is generally sufficient, except for LLM training. A detailed analysis of $\\alpha$ initialization is provided in Section 7. Unless explicitly stated otherwise, $\\alpha$ is initialized to 0.5 in our subsequent experiments.", + "bbox": [ + 89, + 684, + 483, + 790 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Remarks. DyT is not a new type of normalization layer, as it operates on each input element from a tensor independently during a forward pass without computing statistics or other types of aggregations. It does, however, preserve the effect of normalization layers in squashing the extreme values in a non-linear fashion while almost linearly transforming the very central parts of the input.", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 89, + 645, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of DyT, we experiment with Transformers and a few other modern architectures across a diverse range of tasks and domains. In each experiment, we replace the LN or RMSNorm in the original architectures with DyT layers and follow the official open-source protocols to train and test both versions of the models. Detailed instructions for reproducing our results are provided in Appendix A. Notably, to highlight the simplicity of adapting DyT, we use hyperparameters identical to those used by the normalized counterparts. For completeness, additional results regarding tuning of learning rates and initial values of $\\alpha$ are provided in Appendix F.", + "bbox": [ + 511, + 116, + 906, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Supervised learning in vision. We train Vision Transformer (ViT) [20] and ConvNeXt [50] of Base\" and Large\" sizes on the ImageNet-1K classification task [19]. These models are selected due to their popularity and distinct operations: attention in ViT and convolution in ConvNeXt. Table 1 reports the top-1 classification accuracies. DyT performs slightly better than LN across both architectures and model sizes. We further plot the training loss for ViT-B and ConvNeXt-B in Figure 4. The curves show that the convergence behaviors of DyT and LN-based models are aligned.", + "bbox": [ + 511, + 304, + 906, + 455 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/2ccda9a35b38199e81b1cc3d55d34e8871441117f244b50de6ea99c35799eef7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
modelLNDyTchange
ViT-B82.3%82.5%↑0.2%
ViT-L83.1%83.6%↑0.5%
ConvNeXt-B83.7%83.7%-
ConvNeXt-L84.3%84.4%↑0.1%
", + "bbox": [ + 575, + 457, + 844, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Supervised classification accuracy on ImageNet-1K. DyT achieves better or similar performance than LN across both architectures and model sizes.", + "bbox": [ + 511, + 545, + 906, + 585 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b7baae6506173d2c014485304a72c5faa58779d1c2990cd0bf8966405b7f33e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 590, + 880, + 718 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9d2ae29f8036826d3469fa1d6544060abde19d956d0b8893473fb7f88ce49646.jpg", + "image_caption": [ + "Figure 4. Training loss curves for ViT-B and ConvNeXt-B. The curves for both models exhibit similar patterns between LN and DyT, suggesting that they may share similar learning dynamics." + ], + "image_footnote": [], + "bbox": [ + 519, + 720, + 880, + 849 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14904", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bedeb54687c14d9c30c54a7f5f5e4f234f7688487352d02ccdd79750a8c92102.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 89, + 55, + 465, + 186 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fd7954f2a83eb567ee516aa747bbd484bc01286a8ebf046e01d2af713bcfd322.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 55, + 864, + 186 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c85074c4872705decb2ad04005aa7d38d0cbbe528a7411e0be5489e82f7a7e41.jpg", + "image_caption": [ + "Figure 5. LLaMA pretraining loss. The loss curves of DyT and RMSNorm models are closely aligned across model sizes." + ], + "image_footnote": [], + "bbox": [ + 88, + 188, + 460, + 316 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/31f8b8dc47ee5d769dbfb8e71e40d4a7e0888ac2d5e9f80ed9ed4e00230c00e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 186, + 859, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Self-supervised learning in vision. We benchmark with two popular visual self-supervised learning methods: masked autoencoders (MAE) [34] and DINO [14]. Both by default use Vision Transformers as backbones, but have different training objectives: MAE is trained with a reconstruction loss, and DINO uses a joint-embedding loss [47]. Following the standard self-supervised learning protocol, we first pretrain models on ImageNet-1K without using labels and then test the pretrained models by attaching a classification layer and fine-tuning them with labels. The fine-tuning results are presented in Table 2. DyT consistently performs on par with LN in self-supervised learning tasks.", + "bbox": [ + 88, + 345, + 483, + 527 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1d458c9053b3e93e995d37ebd703da2ead850c8cc302d23fc83cf188ec6e922c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
modelLNDyTchange
MAE ViT-B83.2%83.2%-
MAE ViT-L85.5%85.4%↓0.1%
DINO ViT-B (patch size 16)83.2%83.4%↑0.2%
DINO ViT-B (patch size 8)84.1%84.5%↑0.4%
", + "bbox": [ + 112, + 529, + 460, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Diffusion models. We train three Diffusion Transformer (DiT) models [64] of sizes B, L and XL on ImageNet-1K [19]. The patch size is 4, 4, and 2, respectively. Note that in DiT, the LN layers' affine parameters are used for class conditioning in DiT, and we keep them that way in our DyT experiments, only replacing the normalizing transformation with the $\\tanh (\\alpha x)$ function. After training, we evaluate the Fréchet Inception Distance (FID) scores using the standard ImageNet \"reference batch\", as presented in Table 3. DyT achieves comparable or improved FID over LN.", + "bbox": [ + 89, + 652, + 483, + 805 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/313da71622988da81f65608e994fc86d2082a51b85d73daae6bcb48d63e2cb50.jpg", + "table_caption": [ + "Table 2. Self-supervised learning accuracy on ImageNet-1K. DyT performs on par with LN across different methods and sizes." + ], + "table_footnote": [], + "table_body": "
modelLNDyTchange
DiT-B64.963.9↓1.0
DiT-L45.945.7↓0.2
DiT-XL19.920.8↑0.9
", + "bbox": [ + 176, + 806, + 397, + 877 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Large Language Models. We pretrain LLaMA 7B, 13B, 34B, and 70B models [21, 79, 80] to assess DyT performance relative to RMSNorm [89], the default normalization layer used in LLaMA. The models are trained on The Pile dataset [25] with 200B tokens, following the original recipe outlined in LLaMA [80]. On LLaMA with DyT, we add a learnable scalar parameter after the initial embedding layer, and adjust the initial value of $\\alpha$ , as detailed in Section 7. We report the loss value after training and also follow OpenLLaMA [26] to benchmark the models on 15 zero-shot tasks from lm-eval [24]. As shown in Table 4, DyT performs on par with RMSNorm across all four model sizes. Figure 5 illustrates the loss curves, demonstrating similar trends across all model sizes, with training losses closely aligned throughout training.", + "bbox": [ + 511, + 345, + 906, + 573 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/cee3d14dd485dff5d1d08253a9d9540f2625be5426bcc24e0183891e070ef609.jpg", + "table_caption": [ + "Table 3. Image generation quality (FID) on ImageNet. DyT achieves comparable FID scores to LN across various model sizes." + ], + "table_footnote": [], + "table_body": "
score / lossRMSNormDyTchange
LLaMA 7B0.513 / 1.590.513 / 1.60- / ↑0.01
LLaMA 13B0.529 / 1.530.529 / 1.54- / ↑0.01
LLaMA 34B0.536 / 1.500.536 / 1.50- / -
LLaMA 70B0.549 / 1.450.549 / 1.45- / -
", + "bbox": [ + 521, + 575, + 897, + 661 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Self-supervised learning in speech. We pretrain two wav2vec 2.0 Transformer models [7] on the LibriSpeech dataset [63]. We retain the first group normalization layer, as it functions as data normalization to handle the unnormalized input data. Table 5 reports the final validation loss. We observe that DyT performs comparably to LN.", + "bbox": [ + 511, + 717, + 905, + 809 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/df7ae9ed77f60b5b7df0826a6c9830579442c4131de923d4444a2977de1a60a6.jpg", + "table_caption": [ + "Table 4. Language models' training loss and average performance with 15 zero-shot 1m-eval tasks. DyT achieves a comparable zero-shot performance and training loss to RMSNorm." + ], + "table_footnote": [], + "table_body": "
modelLNDyTchange
wav2vec 2.0 Base1.951.95-
wav2vec 2.0 Large1.921.91↓0.01
", + "bbox": [ + 570, + 813, + 849, + 871 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 5. Speech pretraining validation loss on LibriSpeech. DyT performs comparably to LN for both wav2vec 2.0 models.", + "bbox": [ + 511, + 875, + 903, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14905", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3073ff5c508db6f14beaea3b3b12d9fad7c867dad3718995952afcf23c4abac5.jpg", + "image_caption": [ + "Figure 6. Curves of three squashing functions: tanh, hardtanh, and sigmoid. All three functions squash inputs into a bounded range, but tanh achieves the best performance when used in DyT layers. We suspect it is due to its smoothness and zero-centered properties." + ], + "image_footnote": [], + "bbox": [ + 102, + 51, + 401, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/600e1a9390e4d16dd57e59a540132dde226199e7d1f1ab5b1895852b2bbbed28.jpg", + "image_caption": [ + "Figure 7. Left: For two selected DyT layers from the ViT-B model, we track $\\alpha$ and the inverse of the standard deviation (1/std) of activations at the end of each epoch, observing that they evolve together during training. Right: We plot the final $\\alpha$ values of two models, ViT-B and ConvNeXt-B, against the 1/std of the input activations, demonstrating a strong correlation between the two values." + ], + "image_footnote": [], + "bbox": [ + 433, + 50, + 653, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/da316b4b6400f407422da48a2d297effac125bbd60a4442150c2968c50a470e1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 51, + 893, + 172 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DNA sequence modeling. On the long-range DNA sequence modeling task, we pretrain the HyenaDNA model [60] and the Caduceus model [69]. The pretraining uses human reference genome data from [27], and the evaluation is on GenomicBenchmarks [28]. Results are presented in Table 6. DyT maintains performance comparable to LN.", + "bbox": [ + 89, + 263, + 483, + 354 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c173f77b64ba17b7908392ca13e606b1294c7c4fb4d2bc7a2acb4bca483b8644.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
modelLNDyTchange
HyenaDNA [60]85.2%85.2%-
Caduceus [69]86.9%86.9%-
", + "bbox": [ + 143, + 356, + 431, + 414 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Analysis", + "text_level": 1, + "bbox": [ + 89, + 468, + 187, + 486 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct two studies examining the roles of the tanh function and the learnable scale $\\alpha$ .", + "bbox": [ + 89, + 494, + 483, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Ablations of tanh and $\\alpha$", + "text_level": 1, + "bbox": [ + 89, + 534, + 308, + 550 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further investigate the role of tanh and $\\alpha$ in DyT, we conduct experiments to evaluate the model's performance when these components are altered or removed.", + "bbox": [ + 89, + 560, + 483, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Replacing and removing tanh. We replace tanh in DyT layers with alternative squashing functions, specifically hardtanh and sigmoid (Figure 6), while keeping the learnable scalar $\\alpha$ intact. Furthermore, we assess the impact of completely removing tanh by replacing it with the identity function while still retaining $\\alpha$ . As shown in Table 7, the squashing function is essential for stable training. Using the identity function leads to unstable training and divergence,", + "bbox": [ + 89, + 609, + 483, + 731 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e365fac92c6222f805f2cda091f40fda49cc91cdac06a3c1462a8de3407bcd9a.jpg", + "table_caption": [ + "Table 6. DNA classification accuracy on GenomicBenchmarks, averaged over each dataset in GenomicBenchmarks. DyT achieves comparable performance to LN." + ], + "table_footnote": [], + "table_body": "
modelidentitytanhhardtanhsigmoid
ViT-S58.5% → failed80.3%79.9%79.6%
ViT-B61.0% → failed82.5%82.2%81.6%
", + "bbox": [ + 107, + 737, + 467, + 795 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "whereas squashing functions enable stable training. Among the squashing functions, tanh performs the best. This is possibly due to its smoothness and zero-centered properties.", + "bbox": [ + 511, + 263, + 905, + 310 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Removing $\\alpha$ . Next, we evaluate the impact of removing the learnable $\\alpha$ while retaining the squashing functions. As shown in Table 8, removing $\\alpha$ results in performance degradation across all squashing functions, highlighting the critical role of $\\alpha$ in overall model performance.", + "bbox": [ + 511, + 321, + 905, + 398 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/97aaa98aa13e7e2fc8706ed4fc49c7aa9cc88e828a8335294b47f7f23b112c6f.jpg", + "table_caption": [ + "Table 7. ImageNet-1K classification accuracy with different squashing functions. All experiments follow the same training recipe as the original LN-based models. Squashing functions play a crucial role in preventing divergence, with tanh achieving the highest performance among the three functions. “ $\\rightarrow$ failed” indicates that training diverged after some progress, with the preceding number representing the accuracy reached before divergence." + ], + "table_footnote": [], + "table_body": "
modeltanhhardtanhsigmoid
without α81.1%80.7%80.7%
with α82.5%82.2%81.6%
", + "bbox": [ + 573, + 400, + 846, + 457 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 8. ImageNet-1K classification accuracy with ViT-B. All experiments follow the same training recipe as the original LN-based models. $\\alpha$ is essential for enhancing model performance.", + "bbox": [ + 511, + 460, + 905, + 502 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Values of $\\alpha$", + "text_level": 1, + "bbox": [ + 513, + 525, + 635, + 540 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During training. Our analysis reveals that the $\\alpha$ closely tracks the $1 / \\mathrm{std}$ of activations throughout training. As illustrated in the left panel of Figure 7, $\\alpha$ first decrease and then increase during training, but always fluctuate consistently with the standard deviation of activations. This supports the important role of $\\alpha$ in maintaining activations within a suitable range, which leads to stable and effective training.", + "bbox": [ + 511, + 560, + 905, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After training. Our further analysis of the final values of $\\alpha$ in trained networks reveals a strong correlation with the 1/std of the input activations. As shown on the right panel of Figure 7, higher 1/std values generally correspond to larger $\\alpha$ values, and vice versa. Additionally, we observe that deeper layers tend to have activations with larger standard deviations. This trend aligns with characteristics of deep residual networks, as shown in Brock et al. [11] for ConvNets, and Sun et al. [74] for Transformers.", + "bbox": [ + 511, + 680, + 906, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Both analyses suggest that $\\alpha$ acts as a normalization mechanism, learning values close to $1 / \\mathrm{std}$ of the input activations. Unlike LN, which normalizes per token, $\\alpha$ normalizes the entire input collectively and cannot suppress extreme values non-linearly.", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14906", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/796e8ffa5539831c9248f034cd69ad5b6f6382596f3fc0815caee396a7edf903.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 55, + 367, + 178 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4a715c71c0d2236260abcd4575cf0ca5b2eed1c716f991ce94d1ffa7e843645e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 55, + 633, + 178 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f188c861c933703a664b47818539b4dbf8a25f2dad051cbc827bcc9fc106e59d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 55, + 900, + 178 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3373dd2e375fe35c87f2a2861451b4945b27c7a9cc88527e40deb375db3925a1.jpg", + "image_caption": [ + "Figure 8. Performance of different tasks across different $\\alpha_0$ values. We benchmark the performance of all non-LLM tasks used in Section 5 with different initial values of $\\alpha$ . Performance remains stable across a wide range of $\\alpha_0$ values. The only exception is that supervised ViT-L models (top right panel) will diverge for $\\alpha_0$ values larger than 0.6." + ], + "image_footnote": [], + "bbox": [ + 96, + 179, + 367, + 303 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9ec84880ebfc85127aa97cf2522f5bae6a416a96c5087ab102710cd4b97f4e4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 179, + 633, + 301 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8c6b92be8d50717e7d543923b4f9f103605016db23ca45b9d3a381c324fc62ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 179, + 900, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7. Initialization of $\\alpha$", + "text_level": 1, + "bbox": [ + 89, + 358, + 261, + 375 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We find that tuning the initialization of $\\alpha$ (denoted $\\alpha_0$ ) rarely leads to significant performance improvements. The only exception is LLM training, where careful tuning of $\\alpha_0$ yields noticeable performance gains. In this section, we detail our findings on the impact of $\\alpha$ initialization.", + "bbox": [ + 89, + 385, + 483, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7.1. Initialization of $\\alpha$ for Non-LLM Models", + "text_level": 1, + "bbox": [ + 89, + 473, + 429, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Non-LLM models are relatively insensitive to $\\alpha_0$ . Figure 8 shows the effect of varying $\\alpha_0$ on validation performance across different tasks. All experiments follow the original setup and hyperparameters of their respective recipe. We observe that performance remains stable across a wide range of $\\alpha_0$ values, with values between 0.5 and 1.2 generally yielding good results. We observe that adjusting $\\alpha_0$ typically affects only the early stages of the training curves. The main exception is supervised ViT-L experiments, where training becomes unstable and diverges when $\\alpha_0$ exceeds 0.6. In such cases, reducing the learning rate restores stability, as detailed below.", + "bbox": [ + 89, + 502, + 483, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Smaller $\\alpha_0$ results in more stable training. Building on previous observations, we further analyze the factors contributing to training instability. Our findings suggest that increasing either the model size or the learning rate requires lowering $\\alpha_0$ to ensure stable training. Conversely, a higher $\\alpha_0$ requires a lower learning rate to mitigate training instability. Figure 9 shows the ablation of the training stability of supervised ViT with ImageNet-1K dataset. We vary learning rates, model sizes, and $\\alpha_0$ values. Training a larger model is more prone to failure, requiring smaller $\\alpha_0$ values or learning rates for stable training. A similar instability pattern is also observed in LN-based models under comparable conditions, and setting $\\alpha_0 = 0.5$ results in a stability pattern similar to that of LN.", + "bbox": [ + 89, + 689, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Setting $\\alpha_0 = 0.5$ as the default. Based on our findings, we set $\\alpha_0 = 0.5$ as the default value for all non-LLM models. This setting provides training stability comparable to LN while maintaining strong performance.", + "bbox": [ + 511, + 359, + 905, + 421 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7.2. Initialization of $\\alpha$ for LLMs", + "text_level": 1, + "bbox": [ + 511, + 426, + 763, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tuning $\\alpha_0$ enhances LLM performance. As discussed earlier, the default setting of $\\alpha_0 = 0.5$ generally performs well across most tasks. However, we find tuning $\\alpha_0$ can substantially improve LLM performance. We tune $\\alpha_0$ across LLaMA models by pretraining each on 30B tokens and comparing their training losses. Table 9 summarizes the tuned $\\alpha_0$ values for each model. Two key findings emerge:", + "bbox": [ + 511, + 454, + 905, + 560 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Larger models require smaller $\\alpha_0$ values. Once the optimal $\\alpha_0$ is determined for smaller models, the search space for larger models can be reduced accordingly.", + "2. Higher $\\alpha_0$ values for attention blocks improve performance. We find that initializing $\\alpha$ with higher values for DyT layers in attention blocks and lower values for DyT layers in other locations (i.e., within FFN blocks or before the final linear projection) improves performance." + ], + "bbox": [ + 511, + 563, + 905, + 690 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fb597714d0014d926a5f432a428b4d441b4c547466031c50d8db635e5192ad66.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
modelwidthdepthoptimal α0 (attention/other)
LLaMA 7B4096320.8/0.2
LLaMA 13B5120400.6/0.15
LLaMA 34B8196480.2/0.05
LLaMA 70B8196800.2/0.05
", + "bbox": [ + 550, + 696, + 870, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 9. Optimal $\\alpha_0$ for different LLaMA models. Larger models require smaller $\\alpha_0$ values. We find it is important to initialize $\\alpha$ differently in (1) attention blocks (\"attention\"), versus (2) the FFN blocks, and the final DyT layer before outputs (\"other\"). $\\alpha_0$ in attention blocks require larger values.", + "bbox": [ + 511, + 797, + 905, + 867 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Further details and visualizations of the tuning of the initial value of $\\alpha$ of LLMs are provided in the appendix C.", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14907", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/21493d7352e45b062d793c5e17907856f6cf5d55b553580b1afa8b8672c0e646.jpg", + "image_caption": [ + "Figure 9. Stability across varying $\\alpha_0$ values, learning rates, and model sizes. We train supervised ViT models on the ImageNet-1K dataset and observe that larger models are more prone to instability for both LN and DyT models. Lowering the learning rate or reducing $\\alpha_0$ enhances stability. LN shows similar stability to DyT with $\\alpha_0 = 0.5$ ." + ], + "image_footnote": [], + "bbox": [ + 94, + 71, + 903, + 255 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Related Work", + "text_level": 1, + "bbox": [ + 89, + 310, + 232, + 325 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Mechanisms of normalization layers. There has been a rich line of work investigating normalization layers' role in enhancing model performance through various mechanisms. These include stabilizing gradient flow during training [9, 17, 52], reducing sensitivity to weight initialization [18, 70, 90], moderating outlier eigenvalues [10, 44], auto-tuning learning rates [4, 77], and smoothing the loss landscape for more stable optimization [68]. These earlier works focused on studying batch normalization. Recent studies [16, 53, 59] further highlight the connection between normalization layers and sharpness reduction, which contributes to better generalization.", + "bbox": [ + 88, + 342, + 485, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Normalization in Transformers. With the rise of Transformer [82], research has increasingly focused on layer normalization [5], which has proven particularly effective for sequential data in natural language tasks [61, 85, 86]. Recent work [62] reveals that layer normalization introduces strong non-linearity, enhancing the model's representational capacity. Additionally, studies [48, 51] demonstrate that modifying the location of normalization layers within Transformers can improve convergence properties.", + "bbox": [ + 88, + 531, + 483, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Removing normalization. Many studies have explored how to train deep models without normalization layers. Several works [6, 18, 90] explore alternative weight initialization schemes to stabilize training. Self-normalizing networks [45] introduce scaled exponential linear units (SELUs) and a carefully chosen initialization scheme to maintain stable activations and gradient flow without the need for explicit normalization layers. The pioneering work by Brock et al. [11, 12] show that high-performing ResNets can be trained without normalization [72] through combination of initialization techniques [18], weight normalization [37, 65, 67], and adaptive gradient clipping [12]. Additionally, their training strategy incorporates extensive data augmentation [15] and regularization [36, 73]. The studies above are based on various ConvNet models.", + "bbox": [ + 89, + 674, + 485, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Transformer architectures, He and Hofmann [32] explore modifications to Transformer blocks that reduce reliance on normalization layers and skip connections. Jha and Reagen [42] introduce AERO, a Softmax-only LLM that improves inference efficiency and privacy with minimal performance loss. Alternatively, Heimersheim [35] propose a method to gradually remove LN from pretrained networks by fine-tuning the model after removing each normalization layer. Unlike previous approaches, DyT requires minimal modifications to both the architecture and the training recipe. Despite its simplicity, DyT achieves stable training and comparable performance.", + "bbox": [ + 511, + 311, + 906, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "9. Limitations", + "text_level": 1, + "bbox": [ + 511, + 505, + 635, + 520 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct experiments on networks using either LN or RMSNorm because of their popularity in Transformers and other modern architectures. Preliminary experiments (see Appendix E) indicate that DyT struggles to replace BN directly in classic networks like ResNets. It remains to be studied in more depth whether and how DyT can adapt to models with other types of normalization layers.", + "bbox": [ + 511, + 530, + 906, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "10. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 648, + 640, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we demonstrate modern neural networks, in particular Transformers, can be trained without normalization layers. This is done through Dynamic Tanh (DyT), a simple replacement for traditional normalization layers. It adjusts the input activation range via a learnable scaling factor $\\alpha$ and then squashes the extreme values through an $S$ -shaped tanh function. Although a simpler function, it effectively captures the behavior of normalization layers. Under various settings, models with DyT match or exceed the performance of their normalized counterparts. The findings challenge the conventional understanding of the necessity of normalization layers in training modern neural networks. Our study also contributes to understanding the mechanisms of normalization layers, one of the most fundamental building blocks in deep neural networks.", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14908", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Edgar D Adrian. The impulses produced by sensory nerve endings: Part 1. The Journal of Physiology, 1926.", + "[2] Edgar D Adrian and Yngve Zotterman. The impulses produced by sensory nerve-endings: Part 2. the response of a single end-organ. The Journal of Physiology, 1926.", + "[3] Edgar D Adrian and Yngve Zotterman. The impulses produced by sensory nerve endings: Part 3. impulses set up by touch and pressure. The Journal of Physiology, 1926.", + "[4] Sanjeev Arora, Zhiyuan Li, and Kaifeng Lyu. Theoretical analysis of auto rate-tuning by batch normalization. arXiv preprint arXiv:1812.03981, 2018.", + "[5] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016.", + "[6] Thomas Bachlechner, Bodhisattwa Prasad Majumder, Henry Mao, Gary Cottrell, and Julian McAuley. Rezero is all you need: Fast convergence at large depth. In UAI, 2021.", + "[7] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. NeurIPS, 2020.", + "[8] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023.", + "[9] David Balduzzi, Marcus Frean, Lennox Leary, JP Lewis, Kurt Wan-Duo Ma, and Brian McWilliams. The shattered gradients problem: If resnets are the answer, then what is the question? In ICML, 2017.", + "[10] Nils Bjorck, Carla P Gomes, Bart Selman, and Kilian Q Weinberger. Understanding batch normalization. NeurIPS, 2018.", + "[11] Andrew Brock, Soham De, and Samuel L Smith. Characterizing signal propagation to close the performance gap in unnormalized resnets. arXiv preprint arXiv:2101.08692, 2021.", + "[12] Andrew Brock, Soham De, Samuel L Smith, and Karen Simonyan. High-performance large-scale image recognition without normalization. In ICML, 2021.", + "[13] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arXiv preprint arXiv:2403.17297, 2024.", + "[14] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021.", + "[15] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In CVPR Workshops, 2020.", + "[16] Yan Dai, Kwangjun Ahn, and Suvrit Sra. The crucial role of normalization in sharpness-aware minimization. NeurIPS, 2024.", + "[17] Hadi Daneshmand, Jonas Kohler, Francis Bach, Thomas Hofmann, and Aurelien Lucchi. Batch normalization provably avoids ranks collapse for randomly initialised deep networks. NeurIPS, 2020." + ], + "bbox": [ + 93, + 114, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Soham De and Sam Smith. Batch normalization biases residual blocks towards the identity function in deep networks. NeurIPS, 2020.", + "[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009.", + "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.", + "[21] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024.", + "[22] Leo Feng, Frederick Tung, Mohamed Osama Ahmed, Yoshua Bengio, and Hossein Hajimirsadegh. Were rnns all we needed? arXiv preprint arXiv:2410.01201, 2024.", + "[23] Foundation Model Stack. Github: FMS FSDP. https://github.com/foundation-model-stack/fms-fsdp. Accessed: 2025-01-23.", + "[24] Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac'h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lintang Sutawika, Eric Tang, Anish Thite, Ben Wang, Kevin Wang, and Andy Zou. A framework for few-shot language model evaluation.", + "[25] Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, Shawn Presser, and Connor Leahy. The Pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027, 2020.", + "[26] Xinyang Geng and Hao Liu. Openlama: An open reproduction of llama, 2023.", + "[27] Ensembl GRCh38. p13 (genome reference consortium human build 38), insdc assembly, 2013.", + "[28] Katarína Grešová, Vlastimil Martinek, David Čechák, Petr Šimeček, and Panagiotis Alexiou. Genomic benchmarks: a collection of datasets for genomic sequence classification. BMC Genomic Data, 2023.", + "[29] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023.", + "[30] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025.", + "[31] HazyResearch. Github: Hyenadna. https://github.com/HazyResearch/hyena-dna.git. Accessed: 2025-01-23.", + "[32] Bobby He and Thomas Hofmann. Simplifying transformer blocks. arXiv preprint arXiv:2311.01906, 2023." + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14909", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016.", + "[34] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022.", + "[35] Stefan Heimersheim. You can remove gpt2's layernorm by fine-tuning. arXiv preprint arXiv:2409.13710, 2024.", + "[36] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In ECCV, 2016.", + "[37] Lei Huang, Xianglong Liu, Yang Liu, Bo Lang, and Dacheng Tao. Centered weight normalization in accelerating training of deep neural networks. In ICCV, 2017.", + "[38] Lei Huang, Jie Qin, Yi Zhou, Fan Zhu, Li Liu, and Ling Shao. Normalization techniques in training dnns: Methodology, analysis and application. TPAMI, 2023.", + "[39] Xiao Shi Huang, Felipe Perez, Jimmy Ba, and Maksims Volkovs. Improving transformer optimization through better initialization. In ICML, 2020.", + "[40] Hugging Face. Hugging Face: LLaMA 2. https://huggingface.co/meta-llama/Llama-2-7b-hf. Accessed: 2025-01-23.", + "[41] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015.", + "[42] Nandan Kumar Jha and Brandon Reagen. Aero: Softmax-only llms for efficient private inference. arXiv preprint arXiv:2410.13060, 2024.", + "[43] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023.", + "[44] Ryo Karakida, Shotaro Akaho, and Shun-ichi Amari. The normalization method for alleviating pathological sharpness in wide neural networks. NeurIPS, 2019.", + "[45] Günter Klambauer, Thomas Unterthiner, Andreas Mayr, and Sepp Hochreiter. Self-normalizing neural networks. NeurIPS, 2017.", + "[46] Kuleshov Group. Github: Caduceus. https://github.com/kuleshov-group/caduceus.git. Accessed: 2025-01-23.", + "[47] Yann LeCun. A path towards autonomous machine intelligence version 0.9.2, 2022-06-27. Open Review, 2022.", + "[48] Pengxiang Li, Lu Yin, and Shiwei Liu. Mix-ln: Unleashing the power of deeper layers by combining pre-ln and post-ln. arXiv preprint arXiv:2412.13795, 2024.", + "[49] Aixin Liu, Bei Feng, Bin Wang, Bingxuan Wang, Bo Liu, Chenggang Zhao, Chengqi Dengr, Chong Ruan, Damai Dai, Daya Guo, et al. Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model. arXiv preprint arXiv:2405.04434, 2024.", + "[50] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022." + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Ilya Loshchilov, Cheng-Ping Hsieh, Simeng Sun, and Boris Ginsburg. ngpt: Normalized transformer with representation learning on the hypersphere. arXiv preprint arXiv:2410.01131, 2024.", + "[52] Ekdeep S Lubana, Robert Dick, and Hidenori Tanaka. Beyond batchnorm: Towards a unified understanding of normalization in deep learning. NeurIPS, 2021.", + "[53] Kaifeng Lyu, Zhiyuan Li, and Sanjeev Arora. Understanding the generalization benefit of normalization layers: Sharpness reduction. NeurIPS, 2022.", + "[54] Meta Research. Github: ConvNeXt. https://github.com/facebookresearch/ConvNeXt, . Accessed: 2025-01-23.", + "[55] Meta Research. Github: DINO. https://github.com/facebookresearch/dino, . Accessed: 2025-01-23.", + "[56] Meta Research. Github: DiT. https://github.com/facebookresearch/DiT,. Accessed: 2025-01-23.", + "[57] Meta Research. Github: MAE. https://github.com/facebookresearch/mae, . Accessed: 2025-01-23.", + "[58] Meta Research. Github: wav2vec 2.0. https://github.com/facebookresearch/fairseq, . Accessed: 2025-01-23.", + "[59] Maximilian Mueller, Tiffany Vlaar, David Rolnick, and Matthias Hein. Normalization layers are all that sharpness-aware minimization needs. NeurIPS, 2024.", + "[60] Eric Nguyen, Michael Poli, Marjan Faizi, Armin Thomas, Michael Wornow, Callum Birch-Sykes, Stefano Massaroli, Aman Patel, Clayton Rabideau, Yoshua Bengio, et al. Hyenadna: Long-range genomic sequence modeling at single nucleotide resolution. NeurIPS, 2024.", + "[61] Toan Q Nguyen and Julian Salazar. Transformers without tears: Improving the normalization of self-attention. arXiv preprint arXiv:1910.05895, 2019.", + "[62] Yunhao Ni, Yuxin Guo, Junlong Jia, and Lei Huang. On the nonlinearity of layer normalization. arXiv preprint arXiv:2406.01255, 2024.", + "[63] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In ICASSP, 2015.", + "[64] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023.", + "[65] Siyuan Qiao, Huiyu Wang, Chenxi Liu, Wei Shen, and Alan Yuille. Micro-batch training with batch-channel normalization and weight standardization. arXiv preprint arXiv:1903.10520, 2019.", + "[66] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 2020.", + "[67] Tim Salimans and Durk P Kingma. Weight normalization: A simple reparameterization to accelerate training of deep neural networks. NeurIPS, 2016.", + "[68] Shibani Santurkar, Dimitris Tsipras, Andrew Ilyas, and Aleksander Madry. How does batch normalization help optimization? NeurIPS, 2018.", + "[69] Yair Schiff, Chia-Hsiang Kao, Aaron Gokaslan, Tri Dao, Albert Gu, and Volodymyr Kuleshov. Caduceus: Bi" + ], + "bbox": [ + 516, + 90, + 905, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14910", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "directional equivariant long-range dna sequence modeling. arXiv preprint arXiv:2403.03234, 2024.", + "[70] Jie Shao, Kai Hu, Changhu Wang, Xiangyang Xue, and Bhiksha Raj. Is normalization indispensable for training deep neural network? NeurIPS, 2020.", + "[71] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014.", + "[72] Samuel L Smith, Andrew Brock, Leonard Berrada, and Soham De. Convnets match vision transformers at scale. arXiv preprint arXiv:2310.16764, 2023.", + "[73] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. JMLR, 2014.", + "[74] Wenfang Sun, Xinyuan Song, Pengxiang Li, Lu Yin, Yefeng Zheng, and Shiwei Liu. The curse of depth in large language models. arXiv preprint arXiv:2502.05795, 2025.", + "[75] Yu Sun, Xinhao Li, Karan Dalal, Jiarui Xu, Arjun Vikram, Genghan Zhang, Yann Dubois, Xinlei Chen, Xiaolong Wang, Sanmi Koyejo, et al. Learning to (learn at test time): Rnns with expressive hidden states. arXiv preprint arXiv:2407.04620, 2024.", + "[76] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016.", + "[77] Hidenori Tanaka and Daniel Kunin. Noether's learning dynamics: Role of symmetry breaking in neural networks. NeurIPS, 2021.", + "[78] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlp architecture for vision. NeurIPS, 2021.", + "[79] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023.", + "[80] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023.", + "[81] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016.", + "[82] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017.", + "[83] Yuxin Wu and Kaiming He. Group normalization. In ECCV, 2018.", + "[84] Saining Xie, Ross Girshick, Piotr Dólar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, 2017.", + "[85] Ruibin Xiong, Yunchang Yang, Di He, Kai Zheng, Shuxin Zheng, Chen Xing, Huishuai Zhang, Yanyan Lan, Liwei Wang, and Tieyan Liu. On layer normalization in the transformer architecture. In ICML, 2020." + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[86] Jingjing Xu, Xu Sun, Zhiyuan Zhang, Guangxiang Zhao, and Junyang Lin. Understanding and improving layer normalization. NeurIPS, 2019.", + "[87] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024.", + "[88] Shuangfei Zhai, Tatiana Likhomanenko, Etai Littwin, Dan Busbridge, Jason Ramapuram, Yizhe Zhang, Jiatao Gu, and Joshua M Susskind. Stabilizing transformer training by preventing attention entropy collapse. In ICML, 2023.", + "[89] Biao Zhang and Rico Sennrich. Root mean square layer normalization. NeurIPS, 2019.", + "[90] Hongyi Zhang, Yann N Dauphin, and Tengyu Ma. Fixup initialization: Residual learning without normalization. arXiv preprint arXiv:1901.09321, 2019.", + "[91] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024." + ], + "bbox": [ + 516, + 92, + 903, + 388 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14911", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Transformers without Normalization/05ce47c0-7126-423f-aeff-039d314c6fa4_model.json b/2025/Transformers without Normalization/05ce47c0-7126-423f-aeff-039d314c6fa4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a7b430ff4e1f2b7cfcbee4cb48f99992f30a21ac --- /dev/null +++ b/2025/Transformers without Normalization/05ce47c0-7126-423f-aeff-039d314c6fa4_model.json @@ -0,0 +1,2642 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.312, + 0.131, + 0.688, + 0.151 + ], + "angle": 0, + "content": "Transformers without Normalization" + }, + { + "type": "text", + "bbox": [ + 0.184, + 0.179, + 0.813, + 0.201 + ], + "angle": 0, + "content": "Jiachen Zhu\\(^{1,2}\\) Xinlei Chen\\(^{1}\\) Kaiming He\\(^{3}\\) Yann LeCun\\(^{1,2}\\) Zhuang Liu\\(^{1,4,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.46, + 0.208, + 0.54, + 0.224 + ], + "angle": 0, + "content": "\\(\\dagger\\) project lead" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.23, + 0.773, + 0.25 + ], + "angle": 0, + "content": "\\(^{1}\\)FAIR, Meta \\(^{2}\\)New York University \\(^{3}\\)MIT \\(^{4}\\)Princeton University" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.283, + 0.327, + 0.3 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.316, + 0.486, + 0.603 + ], + "angle": 0, + "content": "Normalization layers are ubiquitous in modern neural networks and have long been considered essential. This work demonstrates that Transformers without normalization can achieve the same or better performance using a remarkably simple technique. We introduce Dynamic Tanh (DyT), an element-wise operation \\(\\mathrm{DyT}(\\boldsymbol{x}) = \\tanh(\\alpha \\boldsymbol{x})\\), as a drop-in replacement for normalization layers in Transformers. DyT is inspired by the observation that layer normalization in Transformers often produces tanh-like, \\(S\\)-shaped input-output mappings. By incorporating DyT, Transformers without normalization can match or exceed the performance of their normalized counterparts, mostly without hyperparameter tuning. We validate the effectiveness of Transformers with DyT across diverse settings, from recognition to generation, supervised to self-supervised learning, and computer vision to language models. These findings challenge the conventional understanding that normalization layers are indispensable in modern neural networks, and offer new insights into their role in deep networks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.633, + 0.222, + 0.649 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Over the past decade, normalization layers have solidified their positions as one of the most fundamental components of modern neural networks. It all traces back to the invention of batch normalization in 2015 [41], which enabled drastically faster and better convergence in visual recognition models and quickly gained momentum in the following years. Since then, many variants of normalization layers have been proposed for different network architectures or domains [5, 81, 83, 89]. Today, virtually all modern networks use normalization layers, with layer normalization (Layer Norm, or LN) [5] being one of the most popular, particularly in the dominant Transformer architecture [20, 82]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The widespread adoption of normalization layers is largely driven by their empirical benefits in optimization [10, 68]. In addition to achieving better results, they help accelerate and stabilize convergence. As neural networks" + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.284, + 0.891, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.412, + 0.907, + 0.483 + ], + "angle": 0, + "content": "Figure 1. Left: original Transformer block. Right: block with our proposed Dynamic Tanh (DyT) layer. DyT is a straightforward replacement for commonly used Layer Norm [5] (in some cases RMSNorm [89]) layers. Transformers with DyT match or exceed the performance of their normalized counterparts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.907, + 0.6 + ], + "angle": 0, + "content": "become wider and deeper, this necessity becomes ever more critical [11, 38]. Consequently, normalization layers are widely regarded as crucial, if not indispensable, for effective training of deep networks. This belief is subtly evidenced by the fact that, in recent years, novel architectures often seek to replace attention or convolution layers [22, 29, 75, 78], but almost always retain normalization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.909, + 0.803 + ], + "angle": 0, + "content": "This paper challenges this belief by introducing a simple alternative to normalization layers in Transformers. Our exploration starts with the observation that LN layers map their inputs to outputs with tanh-like, \\( S \\)-shaped curves, scaling the input activations while squashing the extreme values. Inspired by this insight, we propose an elementwise operation termed Dynamic Tanh (DyT), defined as: \\( \\mathrm{DyT}(\\boldsymbol{x}) = \\tanh(\\alpha \\boldsymbol{x}) \\), where \\( \\alpha \\) is a learnable parameter. This operation aims to emulate the behavior of LN by learning an appropriate scaling factor through \\( \\alpha \\) and squashing extreme values via the bounded tanh function. Notably, unlike normalization layers, it achieves both effects without the need to compute activation statistics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Employing DyT is straightforward, as shown in Figure 1: we directly replace existing normalization layers with DyT in architectures such as vision and language Transformers. We empirically demonstrate that models with DyT can train stably and achieve high final performance across a wide range of settings. It often does not require tuning the hy" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14901" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.398, + 0.09, + 0.632, + 0.104 + ], + "angle": 0, + "content": "LN output (y axis) vs. LN input (x axis)" + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.106, + 0.895, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.447, + 0.908, + 0.517 + ], + "angle": 0, + "content": "Figure 2. Output vs. input of selected layer normalization (LN) layers in Vision Transformer (ViT) [20], wav2vec 2.0 (a Transformer model for speech) [7], and Diffusion Transformer (DiT) [64]. We sample a mini-batch of samples and plot the input / output values of four LN layers in each model. The outputs are before the affine transformation in LN. The \\(S\\)-shaped curves highly resemble that of a tanh function. The more linear shapes in earlier layers can also be captured by the center part of a tanh curve. This motivates us to propose Dynamic Tanh (DyT) as a replacement, with a learnable scalar \\(\\alpha\\) to account for different scales on the \\(x\\) axis." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.529, + 0.483, + 0.59 + ], + "angle": 0, + "content": "perparameters on the original architecture. Our work challenges the notion that normalization layers are indispensable for training modern neural networks and provides empirical insights into the properties of normalization layers." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.601, + 0.416, + 0.619 + ], + "angle": 0, + "content": "2. Background: Normalization Layers" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.484, + 0.703 + ], + "angle": 0, + "content": "We begin by reviewing the normalization layers. Most normalization layers share a common formulation. Given an input \\( \\pmb{x} \\) with shape \\( (B, T, C) \\), where \\( B \\) is the batch size, \\( T \\) is the number of tokens, and \\( C \\) is the embedding dimension per token, the output is generally computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.71, + 0.483, + 0.745 + ], + "angle": 0, + "content": "\\[\n\\operatorname {n o r m a l i z a t i o n} (\\boldsymbol {x}) = \\gamma * \\left(\\frac {\\boldsymbol {x} - \\boldsymbol {\\mu}}{\\sqrt {\\sigma^ {2} + \\epsilon}}\\right) + \\beta \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.87 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) is a small constant, and \\(\\gamma\\) and \\(\\beta\\) are learnable vector parameters of shape \\((C,)\\). They are \"scaling\" and \"shifting\" affine parameters that allow the output to be in any range. The terms \\(\\mu\\) and \\(\\sigma^2\\) denote the mean and variance of the input. Different methods mainly differ in how these two statistics are computed. This results in \\(\\mu\\) and \\(\\sigma^2\\) having different dimensions, each with appropriate broadcasting needed in the tensor calculation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Batch normalization (BN) [41] is the first modern normalization layer, and it has been primarily used in Con" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.529, + 0.907, + 0.684 + ], + "angle": 0, + "content": "vNet models [33, 76, 84]. Its introduction represents a major milestone in deep learning architecture designs. BN computes the mean and variance across both the batch and token dimensions, specifically: \\(\\mu_{k} = \\frac{1}{BT}\\sum_{i,j}x_{ijk}\\) and \\(\\sigma_k^2 = \\frac{1}{BT}\\sum_{i,j}(x_{ijk} - \\mu_k)^2\\). Other normalization layers popular in ConvNets, such as group normalization [83] and instance normalization [81], were initially proposed for specialized tasks such as object detection and image stylization. They share the same overall formulation but differ in the axes and ranges over which the statistics are computed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Layer normalization (LN) [5] and root mean square normalization (RMSNorm) [89] are the major two types of normalization layers used in Transformer architectures. LN computes these statistics independently for each token in each sample, where \\(\\mu_{ij} = \\frac{1}{C}\\sum_k x_{ijk}\\) and \\(\\sigma_{ij}^2 = \\frac{1}{C}\\sum_k(x_{ijk} - \\mu_{ij})^2\\). RMSNorm [89] simplifies LN by removing the mean-centering step and normalizing the input with \\(\\mu_{ij} = 0\\) and \\(\\sigma_{ij}^2 = \\frac{1}{C}\\sum_k x_{ijk}^2\\). Today, most modern neural networks use LN due to its simplicity and universality. Recently, RMSNorm has gained popularity, particularly in language models like T5 [66], LLaMA [21, 79, 80], Mistral [43], Qwen [8, 87], InternLM [13, 91] and DeepSeek [30, 49]. The Transformers we examine in this work all use LN, except that LLaMA uses RMSNorm." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14902" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.071, + 0.891, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.207, + 0.908, + 0.293 + ], + "angle": 0, + "content": "Figure 3. Output vs. input of two LN layers, with tensor elements colored to indicate different channel and token dimensions. The input tensor has a shape of (samples, tokens, and channels), with elements visualized by assigning consistent colors to the same tokens (left two panels) and channels (right two panels). Left two panels: points representing the same token (same color) form straight lines across different channels, as LN operates linearly across channels for each token. Interestingly, when plotted collectively, these lines form a non-linear tanh-shaped curve. Right two panels: each channel's input spans different ranges on the \\( x \\)-axis, contributing distinct segments to the overall tanh-shaped curve. Certain channels (e.g., red, green, and pink) exhibit more extreme \\( x \\) values, which are squashed by LN." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.301, + 0.418, + 0.318 + ], + "angle": 0, + "content": "3. What Do Normalization Layers Do?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.332, + 0.484, + 0.452 + ], + "angle": 0, + "content": "Analysis setup. We first empirically study the behaviors of normalization layers in trained networks. For this analysis, we take a Vision Transformer model (ViT-B) [20] trained on ImageNet-1K [19], a wav2vec 2.0 Large Transformer model [7] trained on LibriSpeech [63], and a Diffusion Transformer (DiT-XL) [64] trained on ImageNet-1K. In all cases, LN is applied in every Transformer block and before the final linear projection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.484, + 0.592 + ], + "angle": 0, + "content": "For all three trained networks, we sample a mini-batch of samples and do a forward pass through the network. We then measure the input and output for the normalization layers, i.e., tensors immediately before and after the normalization operation, before the learnable affine transformation. Since LN preserves the dimensions of the input tensor, we can establish a one-to-one correspondence between input and output elements, allowing a direct visualization of their relationship. We plot the resulting mappings in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.484, + 0.673 + ], + "angle": 0, + "content": "Tanh-like mappings with layer normalization. For all three models, in earlier LN layers (1st column of Figure 2), we find this input-output relationship to be mostly linear, resembling a straight line in an \\(x - y\\) plot. However, deeper LN layers are where we make more intriguing observations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.484, + 0.839 + ], + "angle": 0, + "content": "A striking observation from these deeper layers is that most of these curves' shapes resemble full or partial \\(S\\)-shaped curves represented by a tanh function. One might expect LN layers to linearly transform the input tensor, as subtracting the mean and dividing by standard deviation are linear operations. LN normalizes in a per-token manner, only linearly transforming each token's activations. As tokens have different mean and standard deviation values, the linearity does not hold collectively on all activations of the input tensor. Nonetheless, it is surprising that the non-linear transformation is highly similar to a scaled tanh function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "For such an \\(S\\)-shaped curve, we note that the central part, represented by points with \\(x\\) values close to zero, is mainly in a linear shape. Most points \\((\\sim 99\\%)\\) fall in this linear range. However, there are many points that clearly fall out" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.302, + 0.907, + 0.438 + ], + "angle": 0, + "content": "of this range, which are considered to have \"extreme\" values, e.g., those with \\( x \\) larger than 50 or smaller than -50 in the ViT model. Normalization layers' main effect for these values is to squash them into less extreme values, more in line with the majority of points. This is where normalization layers could not approximated by a simple affine transformation layer. We hypothesize this non-linear and disproportional squashing effect on extreme values is what makes normalization layers important and indispensable." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.907, + 0.532 + ], + "angle": 0, + "content": "Recent findings by Ni et al. [62] similarly highlight the strong non-linearities introduced by LN layers, demonstrating how the non-linearity enhances a model's representational capacity. Moreover, this squashing behavior mirrors the saturation properties of biological neurons for large inputs, a phenomenon observed about a century ago [1-3]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.909, + 0.673 + ], + "angle": 0, + "content": "Normalization by tokens and channels. How does an LN layer perform a linear transformation for each token but also squash the extreme values in such a non-linear fashion? To understand this, we visualize the points grouped by tokens and channels, respectively. This is plotted in Figure 3 by taking the second and third subplots for ViT from Figure 2, but with a sampled subset of points for more clarity. When we select the channels to plot, we make sure to include the channels with extreme values." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.902 + ], + "angle": 0, + "content": "On the left two panels of Figure 3, we visualize each token's activations using the same color. We observe that all points from any single token do form a straight line. However, since each token has a different variance, the slopes are different. Tokens with smaller input \\( x \\) ranges tend to have smaller variance, and the normalization layer will divide their activations using a smaller standard deviation, hence producing a larger slope in the straight line. Collectively, they form an \\( S \\)-shaped curve that resembles a tanh function. In the two panels on the right, we color each channel's activations using the same color. We find that different channels tend to have drastically different input ranges, with only a few channels (e.g., red, green, and pink) exhibiting large extreme values. These are the channels that get squashed the most by the normalization layer." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14903" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.295, + 0.108 + ], + "angle": 0, + "content": "4. Dynamic Tanh (DyT)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.483, + 0.176 + ], + "angle": 0, + "content": "Inspired by the similarity between shapes of normalization layers and a scaled tanh function, we propose Dynamic Tanh (DyT) as a replacement for normalization layers. Given input tensor \\( \\pmb{x} \\), a DyT layer is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.184, + 0.483, + 0.2 + ], + "angle": 0, + "content": "\\[\n\\mathrm {D y T} (\\boldsymbol {x}) = \\boldsymbol {\\gamma} * \\tanh (\\alpha \\boldsymbol {x}) + \\boldsymbol {\\beta} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.203, + 0.484, + 0.355 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a learnable scalar parameter that allows scaling the input differently based on its range, accounting for varying \\(x\\) scales (Figure 2). This is also why we name the whole operation \"Dynamic\" Tanh. \\(\\gamma\\) and \\(\\beta\\) are learnable, per-channel vector parameters, the same as those used in all normalization layers—they allow the output to scale back to any scales. This is sometimes considered a separate affine layer; for our purposes, we consider them to be part of the DyT layer, just like how normalization layers also include them. See Algorithm 1 for Pytorch-like pseudocode of DyT." + }, + { + "type": "code_caption", + "bbox": [ + 0.111, + 0.363, + 0.375, + 0.378 + ], + "angle": 0, + "content": "Algorithm 1 Pseudocode of DyT layer." + }, + { + "type": "algorithm", + "bbox": [ + 0.111, + 0.381, + 0.449, + 0.507 + ], + "angle": 0, + "content": "input x has the shape of [B, T, C] \n# B: batch size, T: tokens, C: dimension \nclass DyT(Module): \n def __init__(self, C, init\\_alpha): \n super().__init() \n self.\\(\\alpha =\\) Parameter(ones(1) \\(\\star\\) init\\_alpha) \n self.\\(\\gamma =\\) Parameter(ones(C)) \n self.\\(\\beta =\\) Parameter(zeros(C)) \ndef forward(self, x): \n x = tanh(self.alpha \\(\\star\\) x) \n return self.\\(\\gamma =\\) x + self.\\(\\beta\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.484, + 0.681 + ], + "angle": 0, + "content": "Integrating DyT layers into an existing architecture is straightforward: one DyT layer replaces one normalization layer (see Figure 1). This applies to normalization layers within attention blocks, FFN blocks, and the final normalization layer. Although DyT may look like or be considered an activation function, this study only uses it to replace normalization layers without altering parts of the activation functions in the original architectures, such as GELU or ReLU. Other parts of the network remain intact. We also observe that there is minimal need to tune the hyperparameters for the original architectures for DyT to perform well." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.685, + 0.484, + 0.791 + ], + "angle": 0, + "content": "On scaling parameters. We always simply initialize \\(\\gamma\\) to an all-one vector and \\(\\beta\\) to an all-zero vector following normalization layers. For the scalar parameter \\(\\alpha\\), a default initialization of 0.5 is generally sufficient, except for LLM training. A detailed analysis of \\(\\alpha\\) initialization is provided in Section 7. Unless explicitly stated otherwise, \\(\\alpha\\) is initialized to 0.5 in our subsequent experiments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Remarks. DyT is not a new type of normalization layer, as it operates on each input element from a tensor independently during a forward pass without computing statistics or other types of aggregations. It does, however, preserve the effect of normalization layers in squashing the extreme values in a non-linear fashion while almost linearly transforming the very central parts of the input." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.646, + 0.108 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.117, + 0.907, + 0.299 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of DyT, we experiment with Transformers and a few other modern architectures across a diverse range of tasks and domains. In each experiment, we replace the LN or RMSNorm in the original architectures with DyT layers and follow the official open-source protocols to train and test both versions of the models. Detailed instructions for reproducing our results are provided in Appendix A. Notably, to highlight the simplicity of adapting DyT, we use hyperparameters identical to those used by the normalized counterparts. For completeness, additional results regarding tuning of learning rates and initial values of \\(\\alpha\\) are provided in Appendix F." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.305, + 0.907, + 0.457 + ], + "angle": 0, + "content": "Supervised learning in vision. We train Vision Transformer (ViT) [20] and ConvNeXt [50] of Base\" and Large\" sizes on the ImageNet-1K classification task [19]. These models are selected due to their popularity and distinct operations: attention in ViT and convolution in ConvNeXt. Table 1 reports the top-1 classification accuracies. DyT performs slightly better than LN across both architectures and model sizes. We further plot the training loss for ViT-B and ConvNeXt-B in Figure 4. The curves show that the convergence behaviors of DyT and LN-based models are aligned." + }, + { + "type": "table", + "bbox": [ + 0.576, + 0.458, + 0.846, + 0.542 + ], + "angle": 0, + "content": "
modelLNDyTchange
ViT-B82.3%82.5%↑0.2%
ViT-L83.1%83.6%↑0.5%
ConvNeXt-B83.7%83.7%-
ConvNeXt-L84.3%84.4%↑0.1%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.546, + 0.907, + 0.587 + ], + "angle": 0, + "content": "Table 1. Supervised classification accuracy on ImageNet-1K. DyT achieves better or similar performance than LN across both architectures and model sizes." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.591, + 0.882, + 0.719 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.721, + 0.882, + 0.85 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.853, + 0.907, + 0.896 + ], + "angle": 0, + "content": "Figure 4. Training loss curves for ViT-B and ConvNeXt-B. The curves for both models exhibit similar patterns between LN and DyT, suggesting that they may share similar learning dynamics." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14904" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.056, + 0.466, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.056, + 0.865, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.189, + 0.462, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.188, + 0.861, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.321, + 0.862, + 0.335 + ], + "angle": 0, + "content": "Figure 5. LLaMA pretraining loss. The loss curves of DyT and RMSNorm models are closely aligned across model sizes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.347, + 0.484, + 0.528 + ], + "angle": 0, + "content": "Self-supervised learning in vision. We benchmark with two popular visual self-supervised learning methods: masked autoencoders (MAE) [34] and DINO [14]. Both by default use Vision Transformers as backbones, but have different training objectives: MAE is trained with a reconstruction loss, and DINO uses a joint-embedding loss [47]. Following the standard self-supervised learning protocol, we first pretrain models on ImageNet-1K without using labels and then test the pretrained models by attaching a classification layer and fine-tuning them with labels. The fine-tuning results are presented in Table 2. DyT consistently performs on par with LN in self-supervised learning tasks." + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.53, + 0.462, + 0.615 + ], + "angle": 0, + "content": "
modelLNDyTchange
MAE ViT-B83.2%83.2%-
MAE ViT-L85.5%85.4%↓0.1%
DINO ViT-B (patch size 16)83.2%83.4%↑0.2%
DINO ViT-B (patch size 8)84.1%84.5%↑0.4%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.618, + 0.483, + 0.646 + ], + "angle": 0, + "content": "Table 2. Self-supervised learning accuracy on ImageNet-1K. DyT performs on par with LN across different methods and sizes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.484, + 0.806 + ], + "angle": 0, + "content": "Diffusion models. We train three Diffusion Transformer (DiT) models [64] of sizes B, L and XL on ImageNet-1K [19]. The patch size is 4, 4, and 2, respectively. Note that in DiT, the LN layers' affine parameters are used for class conditioning in DiT, and we keep them that way in our DyT experiments, only replacing the normalizing transformation with the \\(\\tanh (\\alpha x)\\) function. After training, we evaluate the Fréchet Inception Distance (FID) scores using the standard ImageNet \"reference batch\", as presented in Table 3. DyT achieves comparable or improved FID over LN." + }, + { + "type": "table", + "bbox": [ + 0.177, + 0.807, + 0.398, + 0.878 + ], + "angle": 0, + "content": "
modelLNDyTchange
DiT-B64.963.9↓1.0
DiT-L45.945.7↓0.2
DiT-XL19.920.8↑0.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.881, + 0.483, + 0.909 + ], + "angle": 0, + "content": "Table 3. Image generation quality (FID) on ImageNet. DyT achieves comparable FID scores to LN across various model sizes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.347, + 0.907, + 0.574 + ], + "angle": 0, + "content": "Large Language Models. We pretrain LLaMA 7B, 13B, 34B, and 70B models [21, 79, 80] to assess DyT performance relative to RMSNorm [89], the default normalization layer used in LLaMA. The models are trained on The Pile dataset [25] with 200B tokens, following the original recipe outlined in LLaMA [80]. On LLaMA with DyT, we add a learnable scalar parameter after the initial embedding layer, and adjust the initial value of \\(\\alpha\\), as detailed in Section 7. We report the loss value after training and also follow OpenLLaMA [26] to benchmark the models on 15 zero-shot tasks from lm-eval [24]. As shown in Table 4, DyT performs on par with RMSNorm across all four model sizes. Figure 5 illustrates the loss curves, demonstrating similar trends across all model sizes, with training losses closely aligned throughout training." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.577, + 0.898, + 0.662 + ], + "angle": 0, + "content": "
score / lossRMSNormDyTchange
LLaMA 7B0.513 / 1.590.513 / 1.60- / ↑0.01
LLaMA 13B0.529 / 1.530.529 / 1.54- / ↑0.01
LLaMA 34B0.536 / 1.500.536 / 1.50- / -
LLaMA 70B0.549 / 1.450.549 / 1.45- / -
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.665, + 0.906, + 0.708 + ], + "angle": 0, + "content": "Table 4. Language models' training loss and average performance with 15 zero-shot 1m-eval tasks. DyT achieves a comparable zero-shot performance and training loss to RMSNorm." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.906, + 0.81 + ], + "angle": 0, + "content": "Self-supervised learning in speech. We pretrain two wav2vec 2.0 Transformer models [7] on the LibriSpeech dataset [63]. We retain the first group normalization layer, as it functions as data normalization to handle the unnormalized input data. Table 5 reports the final validation loss. We observe that DyT performs comparably to LN." + }, + { + "type": "table", + "bbox": [ + 0.571, + 0.814, + 0.85, + 0.872 + ], + "angle": 0, + "content": "
modelLNDyTchange
wav2vec 2.0 Base1.951.95-
wav2vec 2.0 Large1.921.91↓0.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.875, + 0.905, + 0.903 + ], + "angle": 0, + "content": "Table 5. Speech pretraining validation loss on LibriSpeech. DyT performs comparably to LN for both wav2vec 2.0 models." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14905" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.052, + 0.403, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.175, + 0.414, + 0.248 + ], + "angle": 0, + "content": "Figure 6. Curves of three squashing functions: tanh, hardtanh, and sigmoid. All three functions squash inputs into a bounded range, but tanh achieves the best performance when used in DyT layers. We suspect it is due to its smoothness and zero-centered properties." + }, + { + "type": "image", + "bbox": [ + 0.434, + 0.051, + 0.655, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.672, + 0.053, + 0.895, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.175, + 0.905, + 0.248 + ], + "angle": 0, + "content": "Figure 7. Left: For two selected DyT layers from the ViT-B model, we track \\(\\alpha\\) and the inverse of the standard deviation (1/std) of activations at the end of each epoch, observing that they evolve together during training. Right: We plot the final \\(\\alpha\\) values of two models, ViT-B and ConvNeXt-B, against the 1/std of the input activations, demonstrating a strong correlation between the two values." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.264, + 0.484, + 0.356 + ], + "angle": 0, + "content": "DNA sequence modeling. On the long-range DNA sequence modeling task, we pretrain the HyenaDNA model [60] and the Caduceus model [69]. The pretraining uses human reference genome data from [27], and the evaluation is on GenomicBenchmarks [28]. Results are presented in Table 6. DyT maintains performance comparable to LN." + }, + { + "type": "table", + "bbox": [ + 0.144, + 0.357, + 0.432, + 0.415 + ], + "angle": 0, + "content": "
modelLNDyTchange
HyenaDNA [60]85.2%85.2%-
Caduceus [69]86.9%86.9%-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.417, + 0.484, + 0.46 + ], + "angle": 0, + "content": "Table 6. DNA classification accuracy on GenomicBenchmarks, averaged over each dataset in GenomicBenchmarks. DyT achieves comparable performance to LN." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.469, + 0.188, + 0.487 + ], + "angle": 0, + "content": "6. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.495, + 0.484, + 0.525 + ], + "angle": 0, + "content": "We conduct two studies examining the roles of the tanh function and the learnable scale \\(\\alpha\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.535, + 0.31, + 0.551 + ], + "angle": 0, + "content": "6.1. Ablations of tanh and \\(\\alpha\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.561, + 0.484, + 0.606 + ], + "angle": 0, + "content": "To further investigate the role of tanh and \\(\\alpha\\) in DyT, we conduct experiments to evaluate the model's performance when these components are altered or removed." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.484, + 0.732 + ], + "angle": 0, + "content": "Replacing and removing tanh. We replace tanh in DyT layers with alternative squashing functions, specifically hardtanh and sigmoid (Figure 6), while keeping the learnable scalar \\(\\alpha\\) intact. Furthermore, we assess the impact of completely removing tanh by replacing it with the identity function while still retaining \\(\\alpha\\). As shown in Table 7, the squashing function is essential for stable training. Using the identity function leads to unstable training and divergence," + }, + { + "type": "table", + "bbox": [ + 0.109, + 0.738, + 0.468, + 0.796 + ], + "angle": 0, + "content": "
modelidentitytanhhardtanhsigmoid
ViT-S58.5% → failed80.3%79.9%79.6%
ViT-B61.0% → failed82.5%82.2%81.6%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.798, + 0.484, + 0.897 + ], + "angle": 0, + "content": "Table 7. ImageNet-1K classification accuracy with different squashing functions. All experiments follow the same training recipe as the original LN-based models. Squashing functions play a crucial role in preventing divergence, with tanh achieving the highest performance among the three functions. “\\(\\rightarrow\\) failed” indicates that training diverged after some progress, with the preceding number representing the accuracy reached before divergence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.264, + 0.906, + 0.311 + ], + "angle": 0, + "content": "whereas squashing functions enable stable training. Among the squashing functions, tanh performs the best. This is possibly due to its smoothness and zero-centered properties." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.323, + 0.906, + 0.399 + ], + "angle": 0, + "content": "Removing \\(\\alpha\\). Next, we evaluate the impact of removing the learnable \\(\\alpha\\) while retaining the squashing functions. As shown in Table 8, removing \\(\\alpha\\) results in performance degradation across all squashing functions, highlighting the critical role of \\(\\alpha\\) in overall model performance." + }, + { + "type": "table", + "bbox": [ + 0.575, + 0.401, + 0.847, + 0.458 + ], + "angle": 0, + "content": "
modeltanhhardtanhsigmoid
without α81.1%80.7%80.7%
with α82.5%82.2%81.6%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.461, + 0.906, + 0.503 + ], + "angle": 0, + "content": "Table 8. ImageNet-1K classification accuracy with ViT-B. All experiments follow the same training recipe as the original LN-based models. \\(\\alpha\\) is essential for enhancing model performance." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.526, + 0.636, + 0.541 + ], + "angle": 0, + "content": "6.2. Values of \\(\\alpha\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.561, + 0.906, + 0.669 + ], + "angle": 0, + "content": "During training. Our analysis reveals that the \\(\\alpha\\) closely tracks the \\(1 / \\mathrm{std}\\) of activations throughout training. As illustrated in the left panel of Figure 7, \\(\\alpha\\) first decrease and then increase during training, but always fluctuate consistently with the standard deviation of activations. This supports the important role of \\(\\alpha\\) in maintaining activations within a suitable range, which leads to stable and effective training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.681, + 0.907, + 0.817 + ], + "angle": 0, + "content": "After training. Our further analysis of the final values of \\(\\alpha\\) in trained networks reveals a strong correlation with the 1/std of the input activations. As shown on the right panel of Figure 7, higher 1/std values generally correspond to larger \\(\\alpha\\) values, and vice versa. Additionally, we observe that deeper layers tend to have activations with larger standard deviations. This trend aligns with characteristics of deep residual networks, as shown in Brock et al. [11] for ConvNets, and Sun et al. [74] for Transformers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Both analyses suggest that \\(\\alpha\\) acts as a normalization mechanism, learning values close to \\(1 / \\mathrm{std}\\) of the input activations. Unlike LN, which normalizes per token, \\(\\alpha\\) normalizes the entire input collectively and cannot suppress extreme values non-linearly." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14906" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.056, + 0.368, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.056, + 0.634, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.634, + 0.056, + 0.901, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.18, + 0.368, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.18, + 0.634, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.634, + 0.18, + 0.901, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.307, + 0.908, + 0.351 + ], + "angle": 0, + "content": "Figure 8. Performance of different tasks across different \\(\\alpha_0\\) values. We benchmark the performance of all non-LLM tasks used in Section 5 with different initial values of \\(\\alpha\\). Performance remains stable across a wide range of \\(\\alpha_0\\) values. The only exception is that supervised ViT-L models (top right panel) will diverge for \\(\\alpha_0\\) values larger than 0.6." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.359, + 0.262, + 0.375 + ], + "angle": 0, + "content": "7. Initialization of \\(\\alpha\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.386, + 0.484, + 0.462 + ], + "angle": 0, + "content": "We find that tuning the initialization of \\(\\alpha\\) (denoted \\(\\alpha_0\\)) rarely leads to significant performance improvements. The only exception is LLM training, where careful tuning of \\(\\alpha_0\\) yields noticeable performance gains. In this section, we detail our findings on the impact of \\(\\alpha\\) initialization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.474, + 0.431, + 0.49 + ], + "angle": 0, + "content": "7.1. Initialization of \\(\\alpha\\) for Non-LLM Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.503, + 0.484, + 0.684 + ], + "angle": 0, + "content": "Non-LLM models are relatively insensitive to \\(\\alpha_0\\). Figure 8 shows the effect of varying \\(\\alpha_0\\) on validation performance across different tasks. All experiments follow the original setup and hyperparameters of their respective recipe. We observe that performance remains stable across a wide range of \\(\\alpha_0\\) values, with values between 0.5 and 1.2 generally yielding good results. We observe that adjusting \\(\\alpha_0\\) typically affects only the early stages of the training curves. The main exception is supervised ViT-L experiments, where training becomes unstable and diverges when \\(\\alpha_0\\) exceeds 0.6. In such cases, reducing the learning rate restores stability, as detailed below." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Smaller \\(\\alpha_0\\) results in more stable training. Building on previous observations, we further analyze the factors contributing to training instability. Our findings suggest that increasing either the model size or the learning rate requires lowering \\(\\alpha_0\\) to ensure stable training. Conversely, a higher \\(\\alpha_0\\) requires a lower learning rate to mitigate training instability. Figure 9 shows the ablation of the training stability of supervised ViT with ImageNet-1K dataset. We vary learning rates, model sizes, and \\(\\alpha_0\\) values. Training a larger model is more prone to failure, requiring smaller \\(\\alpha_0\\) values or learning rates for stable training. A similar instability pattern is also observed in LN-based models under comparable conditions, and setting \\(\\alpha_0 = 0.5\\) results in a stability pattern similar to that of LN." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.361, + 0.906, + 0.422 + ], + "angle": 0, + "content": "Setting \\(\\alpha_0 = 0.5\\) as the default. Based on our findings, we set \\(\\alpha_0 = 0.5\\) as the default value for all non-LLM models. This setting provides training stability comparable to LN while maintaining strong performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.428, + 0.764, + 0.443 + ], + "angle": 0, + "content": "7.2. Initialization of \\(\\alpha\\) for LLMs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.455, + 0.906, + 0.561 + ], + "angle": 0, + "content": "Tuning \\(\\alpha_0\\) enhances LLM performance. As discussed earlier, the default setting of \\(\\alpha_0 = 0.5\\) generally performs well across most tasks. However, we find tuning \\(\\alpha_0\\) can substantially improve LLM performance. We tune \\(\\alpha_0\\) across LLaMA models by pretraining each on 30B tokens and comparing their training losses. Table 9 summarizes the tuned \\(\\alpha_0\\) values for each model. Two key findings emerge:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.564, + 0.906, + 0.611 + ], + "angle": 0, + "content": "1. Larger models require smaller \\(\\alpha_0\\) values. Once the optimal \\(\\alpha_0\\) is determined for smaller models, the search space for larger models can be reduced accordingly." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.615, + 0.906, + 0.691 + ], + "angle": 0, + "content": "2. Higher \\(\\alpha_0\\) values for attention blocks improve performance. We find that initializing \\(\\alpha\\) with higher values for DyT layers in attention blocks and lower values for DyT layers in other locations (i.e., within FFN blocks or before the final linear projection) improves performance." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.564, + 0.906, + 0.691 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.551, + 0.697, + 0.871, + 0.795 + ], + "angle": 0, + "content": "
modelwidthdepthoptimal α0 (attention/other)
LLaMA 7B4096320.8/0.2
LLaMA 13B5120400.6/0.15
LLaMA 34B8196480.2/0.05
LLaMA 70B8196800.2/0.05
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.798, + 0.906, + 0.868 + ], + "angle": 0, + "content": "Table 9. Optimal \\(\\alpha_0\\) for different LLaMA models. Larger models require smaller \\(\\alpha_0\\) values. We find it is important to initialize \\(\\alpha\\) differently in (1) attention blocks (\"attention\"), versus (2) the FFN blocks, and the final DyT layer before outputs (\"other\"). \\(\\alpha_0\\) in attention blocks require larger values." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Further details and visualizations of the tuning of the initial value of \\(\\alpha\\) of LLMs are provided in the appendix C." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "14907" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.072, + 0.904, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.259, + 0.908, + 0.302 + ], + "angle": 0, + "content": "Figure 9. Stability across varying \\(\\alpha_0\\) values, learning rates, and model sizes. We train supervised ViT models on the ImageNet-1K dataset and observe that larger models are more prone to instability for both LN and DyT models. Lowering the learning rate or reducing \\(\\alpha_0\\) enhances stability. LN shows similar stability to DyT with \\(\\alpha_0 = 0.5\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.311, + 0.233, + 0.327 + ], + "angle": 0, + "content": "8. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.343, + 0.486, + 0.525 + ], + "angle": 0, + "content": "Mechanisms of normalization layers. There has been a rich line of work investigating normalization layers' role in enhancing model performance through various mechanisms. These include stabilizing gradient flow during training [9, 17, 52], reducing sensitivity to weight initialization [18, 70, 90], moderating outlier eigenvalues [10, 44], auto-tuning learning rates [4, 77], and smoothing the loss landscape for more stable optimization [68]. These earlier works focused on studying batch normalization. Recent studies [16, 53, 59] further highlight the connection between normalization layers and sharpness reduction, which contributes to better generalization." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.532, + 0.485, + 0.669 + ], + "angle": 0, + "content": "Normalization in Transformers. With the rise of Transformer [82], research has increasingly focused on layer normalization [5], which has proven particularly effective for sequential data in natural language tasks [61, 85, 86]. Recent work [62] reveals that layer normalization introduces strong non-linearity, enhancing the model's representational capacity. Additionally, studies [48, 51] demonstrate that modifying the location of normalization layers within Transformers can improve convergence properties." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Removing normalization. Many studies have explored how to train deep models without normalization layers. Several works [6, 18, 90] explore alternative weight initialization schemes to stabilize training. Self-normalizing networks [45] introduce scaled exponential linear units (SELUs) and a carefully chosen initialization scheme to maintain stable activations and gradient flow without the need for explicit normalization layers. The pioneering work by Brock et al. [11, 12] show that high-performing ResNets can be trained without normalization [72] through combination of initialization techniques [18], weight normalization [37, 65, 67], and adaptive gradient clipping [12]. Additionally, their training strategy incorporates extensive data augmentation [15] and regularization [36, 73]. The studies above are based on various ConvNet models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.312, + 0.908, + 0.495 + ], + "angle": 0, + "content": "In Transformer architectures, He and Hofmann [32] explore modifications to Transformer blocks that reduce reliance on normalization layers and skip connections. Jha and Reagen [42] introduce AERO, a Softmax-only LLM that improves inference efficiency and privacy with minimal performance loss. Alternatively, Heimersheim [35] propose a method to gradually remove LN from pretrained networks by fine-tuning the model after removing each normalization layer. Unlike previous approaches, DyT requires minimal modifications to both the architecture and the training recipe. Despite its simplicity, DyT achieves stable training and comparable performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.506, + 0.637, + 0.521 + ], + "angle": 0, + "content": "9. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.531, + 0.907, + 0.639 + ], + "angle": 0, + "content": "We conduct experiments on networks using either LN or RMSNorm because of their popularity in Transformers and other modern architectures. Preliminary experiments (see Appendix E) indicate that DyT struggles to replace BN directly in classic networks like ResNets. It remains to be studied in more depth whether and how DyT can adapt to models with other types of normalization layers." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.649, + 0.642, + 0.665 + ], + "angle": 0, + "content": "10. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this work, we demonstrate modern neural networks, in particular Transformers, can be trained without normalization layers. This is done through Dynamic Tanh (DyT), a simple replacement for traditional normalization layers. It adjusts the input activation range via a learnable scaling factor \\(\\alpha\\) and then squashes the extreme values through an \\(S\\)-shaped tanh function. Although a simpler function, it effectively captures the behavior of normalization layers. Under various settings, models with DyT match or exceed the performance of their normalized counterparts. The findings challenge the conventional understanding of the necessity of normalization layers in training modern neural networks. Our study also contributes to understanding the mechanisms of normalization layers, one of the most fundamental building blocks in deep neural networks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14908" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.143 + ], + "angle": 0, + "content": "[1] Edgar D Adrian. The impulses produced by sensory nerve endings: Part 1. The Journal of Physiology, 1926." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.145, + 0.483, + 0.186 + ], + "angle": 0, + "content": "[2] Edgar D Adrian and Yngve Zotterman. The impulses produced by sensory nerve-endings: Part 2. the response of a single end-organ. The Journal of Physiology, 1926." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.188, + 0.482, + 0.229 + ], + "angle": 0, + "content": "[3] Edgar D Adrian and Yngve Zotterman. The impulses produced by sensory nerve endings: Part 3. impulses set up by touch and pressure. The Journal of Physiology, 1926." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.231, + 0.482, + 0.272 + ], + "angle": 0, + "content": "[4] Sanjeev Arora, Zhiyuan Li, and Kaifeng Lyu. Theoretical analysis of auto rate-tuning by batch normalization. arXiv preprint arXiv:1812.03981, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.274, + 0.482, + 0.314 + ], + "angle": 0, + "content": "[5] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.317, + 0.482, + 0.358 + ], + "angle": 0, + "content": "[6] Thomas Bachlechner, Bodhisattwa Prasad Majumder, Henry Mao, Gary Cottrell, and Julian McAuley. Rezero is all you need: Fast convergence at large depth. In UAI, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.36, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[7] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.403, + 0.482, + 0.457 + ], + "angle": 0, + "content": "[8] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.46, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[9] David Balduzzi, Marcus Frean, Lennox Leary, JP Lewis, Kurt Wan-Duo Ma, and Brian McWilliams. The shattered gradients problem: If resnets are the answer, then what is the question? In ICML, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.517, + 0.482, + 0.556 + ], + "angle": 0, + "content": "[10] Nils Bjorck, Carla P Gomes, Bart Selman, and Kilian Q Weinberger. Understanding batch normalization. NeurIPS, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.56, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[11] Andrew Brock, Soham De, and Samuel L Smith. Characterizing signal propagation to close the performance gap in unnormalized resnets. arXiv preprint arXiv:2101.08692, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.603, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[12] Andrew Brock, Soham De, Samuel L Smith, and Karen Simonyan. High-performance large-scale image recognition without normalization. In ICML, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.646, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[13] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arXiv preprint arXiv:2403.17297, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.703, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[14] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[15] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In CVPR Workshops, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[16] Yan Dai, Kwangjun Ahn, and Suvrit Sra. The crucial role of normalization in sharpness-aware minimization. NeurIPS, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[17] Hadi Daneshmand, Jonas Kohler, Francis Bach, Thomas Hofmann, and Aurelien Lucchi. Batch normalization provably avoids ranks collapse for randomly initialised deep networks. NeurIPS, 2020." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[18] Soham De and Sam Smith. Batch normalization biases residual blocks towards the identity function in deep networks. NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[21] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.334, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[22] Leo Feng, Frederick Tung, Mohamed Osama Ahmed, Yoshua Bengio, and Hossein Hajimirsadegh. Were rnns all we needed? arXiv preprint arXiv:2410.01201, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.377, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[23] Foundation Model Stack. Github: FMS FSDP. https://github.com/foundation-model-stack/fms-fsdp. Accessed: 2025-01-23." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.42, + 0.905, + 0.529 + ], + "angle": 0, + "content": "[24] Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac'h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lintang Sutawika, Eric Tang, Anish Thite, Ben Wang, Kevin Wang, and Andy Zou. A framework for few-shot language model evaluation." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.532, + 0.905, + 0.6 + ], + "angle": 0, + "content": "[25] Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, Shawn Presser, and Connor Leahy. The Pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.602, + 0.905, + 0.628 + ], + "angle": 0, + "content": "[26] Xinyang Geng and Hao Liu. Openlama: An open reproduction of llama, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.631, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[27] Ensembl GRCh38. p13 (genome reference consortium human build 38), insdc assembly, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.66, + 0.905, + 0.714 + ], + "angle": 0, + "content": "[28] Katarína Grešová, Vlastimil Martinek, David Čechák, Petr Šimeček, and Panagiotis Alexiou. Genomic benchmarks: a collection of datasets for genomic sequence classification. BMC Genomic Data, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.717, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[29] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.76, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[30] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.831, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[31] HazyResearch. Github: Hyenadna. https://github.com/HazyResearch/hyena-dna.git. Accessed: 2025-01-23." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[32] Bobby He and Thomas Hofmann. Simplifying transformer blocks. arXiv preprint arXiv:2311.01906, 2023." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14909" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[33] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[34] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.18, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[35] Stefan Heimersheim. You can remove gpt2's layernorm by fine-tuning. arXiv preprint arXiv:2409.13710, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.21, + 0.483, + 0.249 + ], + "angle": 0, + "content": "[36] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In ECCV, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.253, + 0.482, + 0.294 + ], + "angle": 0, + "content": "[37] Lei Huang, Xianglong Liu, Yang Liu, Bo Lang, and Dacheng Tao. Centered weight normalization in accelerating training of deep neural networks. In ICCV, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.297, + 0.482, + 0.338 + ], + "angle": 0, + "content": "[38] Lei Huang, Jie Qin, Yi Zhou, Fan Zhu, Li Liu, and Ling Shao. Normalization techniques in training dnns: Methodology, analysis and application. TPAMI, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.34, + 0.482, + 0.381 + ], + "angle": 0, + "content": "[39] Xiao Shi Huang, Felipe Perez, Jimmy Ba, and Maksims Volkovs. Improving transformer optimization through better initialization. In ICML, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.384, + 0.482, + 0.424 + ], + "angle": 0, + "content": "[40] Hugging Face. Hugging Face: LLaMA 2. https://huggingface.co/meta-llama/Llama-2-7b-hf. Accessed: 2025-01-23." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.427, + 0.482, + 0.467 + ], + "angle": 0, + "content": "[41] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.471, + 0.482, + 0.511 + ], + "angle": 0, + "content": "[42] Nandan Kumar Jha and Brandon Reagen. Aero: Softmax-only llms for efficient private inference. arXiv preprint arXiv:2410.13060, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.513, + 0.482, + 0.581 + ], + "angle": 0, + "content": "[43] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.584, + 0.482, + 0.625 + ], + "angle": 0, + "content": "[44] Ryo Karakida, Shotaro Akaho, and Shun-ichi Amari. The normalization method for alleviating pathological sharpness in wide neural networks. NeurIPS, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.629, + 0.482, + 0.668 + ], + "angle": 0, + "content": "[45] Günter Klambauer, Thomas Unterthiner, Andreas Mayr, and Sepp Hochreiter. Self-normalizing neural networks. NeurIPS, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.672, + 0.482, + 0.712 + ], + "angle": 0, + "content": "[46] Kuleshov Group. Github: Caduceus. https://github.com/kuleshov-group/caduceus.git. Accessed: 2025-01-23." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.716, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[47] Yann LeCun. A path towards autonomous machine intelligence version 0.9.2, 2022-06-27. Open Review, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[48] Pengxiang Li, Lu Yin, and Shiwei Liu. Mix-ln: Unleashing the power of deeper layers by combining pre-ln and post-ln. arXiv preprint arXiv:2412.13795, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[49] Aixin Liu, Bei Feng, Bin Wang, Bingxuan Wang, Bo Liu, Chenggang Zhao, Chengqi Dengr, Chong Ruan, Damai Dai, Daya Guo, et al. Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model. arXiv preprint arXiv:2405.04434, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[50] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[51] Ilya Loshchilov, Cheng-Ping Hsieh, Simeng Sun, and Boris Ginsburg. ngpt: Normalized transformer with representation learning on the hypersphere. arXiv preprint arXiv:2410.01131, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[52] Ekdeep S Lubana, Robert Dick, and Hidenori Tanaka. Beyond batchnorm: Towards a unified understanding of normalization in deep learning. NeurIPS, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[53] Kaifeng Lyu, Zhiyuan Li, and Sanjeev Arora. Understanding the generalization benefit of normalization layers: Sharpness reduction. NeurIPS, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[54] Meta Research. Github: ConvNeXt. https://github.com/facebookresearch/ConvNeXt, . Accessed: 2025-01-23." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.904, + 0.304 + ], + "angle": 0, + "content": "[55] Meta Research. Github: DINO. https://github.com/facebookresearch/dino, . Accessed: 2025-01-23." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.904, + 0.333 + ], + "angle": 0, + "content": "[56] Meta Research. Github: DiT. https://github.com/facebookresearch/DiT,. Accessed: 2025-01-23." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.335, + 0.904, + 0.361 + ], + "angle": 0, + "content": "[57] Meta Research. Github: MAE. https://github.com/facebookresearch/mae, . Accessed: 2025-01-23." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[58] Meta Research. Github: wav2vec 2.0. https://github.com/facebookresearch/fairseq, . Accessed: 2025-01-23." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[59] Maximilian Mueller, Tiffany Vlaar, David Rolnick, and Matthias Hein. Normalization layers are all that sharpness-aware minimization needs. NeurIPS, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.449, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[60] Eric Nguyen, Michael Poli, Marjan Faizi, Armin Thomas, Michael Wornow, Callum Birch-Sykes, Stefano Massaroli, Aman Patel, Clayton Rabideau, Yoshua Bengio, et al. Hyenadna: Long-range genomic sequence modeling at single nucleotide resolution. NeurIPS, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[61] Toan Q Nguyen and Julian Salazar. Transformers without tears: Improving the normalization of self-attention. arXiv preprint arXiv:1910.05895, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[62] Yunhao Ni, Yuxin Guo, Junlong Jia, and Lei Huang. On the nonlinearity of layer normalization. arXiv preprint arXiv:2406.01255, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[63] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In ICASSP, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[64] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[65] Siyuan Qiao, Huiyu Wang, Chenxi Liu, Wei Shen, and Alan Yuille. Micro-batch training with batch-channel normalization and weight standardization. arXiv preprint arXiv:1903.10520, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.732, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[66] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[67] Tim Salimans and Durk P Kingma. Weight normalization: A simple reparameterization to accelerate training of deep neural networks. NeurIPS, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[68] Shibani Santurkar, Dimitris Tsipras, Andrew Ilyas, and Aleksander Madry. How does batch normalization help optimization? NeurIPS, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[69] Yair Schiff, Chia-Hsiang Kao, Aaron Gokaslan, Tri Dao, Albert Gu, and Volodymyr Kuleshov. Caduceus: Bi" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14910" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.481, + 0.12 + ], + "angle": 0, + "content": "directional equivariant long-range dna sequence modeling. arXiv preprint arXiv:2403.03234, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.162 + ], + "angle": 0, + "content": "[70] Jie Shao, Kai Hu, Changhu Wang, Xiangyang Xue, and Bhiksha Raj. Is normalization indispensable for training deep neural network? NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.483, + 0.204 + ], + "angle": 0, + "content": "[71] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[72] Samuel L Smith, Andrew Brock, Leonard Berrada, and Soham De. Convnets match vision transformers at scale. arXiv preprint arXiv:2310.16764, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.248, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[73] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. JMLR, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.289, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[74] Wenfang Sun, Xinyuan Song, Pengxiang Li, Lu Yin, Yefeng Zheng, and Shiwei Liu. The curse of depth in large language models. arXiv preprint arXiv:2502.05795, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[75] Yu Sun, Xinhao Li, Karan Dalal, Jiarui Xu, Arjun Vikram, Genghan Zhang, Yann Dubois, Xinlei Chen, Xiaolong Wang, Sanmi Koyejo, et al. Learning to (learn at test time): Rnns with expressive hidden states. arXiv preprint arXiv:2407.04620, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.401, + 0.482, + 0.441 + ], + "angle": 0, + "content": "[76] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.483 + ], + "angle": 0, + "content": "[77] Hidenori Tanaka and Daniel Kunin. Noether's learning dynamics: Role of symmetry breaking in neural networks. NeurIPS, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[78] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlp architecture for vision. NeurIPS, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.622 + ], + "angle": 0, + "content": "[79] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.692 + ], + "angle": 0, + "content": "[80] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[81] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[82] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[83] Yuxin Wu and Kaiming He. Group normalization. In ECCV, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[84] Saining Xie, Ross Girshick, Piotr Dólar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[85] Ruibin Xiong, Yunchang Yang, Di He, Kai Zheng, Shuxin Zheng, Chen Xing, Huishuai Zhang, Yanyan Lan, Liwei Wang, and Tieyan Liu. On layer normalization in the transformer architecture. In ICML, 2020." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[86] Jingjing Xu, Xu Sun, Zhiyuan Zhang, Guangxiang Zhao, and Junyang Lin. Understanding and improving layer normalization. NeurIPS, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[87] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[88] Shuangfei Zhai, Tatiana Likhomanenko, Etai Littwin, Dan Busbridge, Jason Ramapuram, Yizhe Zhang, Jiatao Gu, and Joshua M Susskind. Stabilizing transformer training by preventing attention entropy collapse. In ICML, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[89] Biao Zhang and Rico Sennrich. Root mean square layer normalization. NeurIPS, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[90] Hongyi Zhang, Yann N Dauphin, and Tengyu Ma. Fixup initialization: Residual learning without normalization. arXiv preprint arXiv:1901.09321, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[91] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "14911" + } + ] +] \ No newline at end of file diff --git a/2025/Transformers without Normalization/05ce47c0-7126-423f-aeff-039d314c6fa4_origin.pdf b/2025/Transformers without Normalization/05ce47c0-7126-423f-aeff-039d314c6fa4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8cc372c2a03e1b8cb70f8190e265f8bd540c046c --- /dev/null +++ b/2025/Transformers without Normalization/05ce47c0-7126-423f-aeff-039d314c6fa4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c712c6429aa0af9b285963aa5767e71b18595eea3413b8f3b594b417d4303f18 +size 2511408 diff --git a/2025/Transformers without Normalization/full.md b/2025/Transformers without Normalization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b28898c379437dae6f9410269e70078724f98fe6 --- /dev/null +++ b/2025/Transformers without Normalization/full.md @@ -0,0 +1,351 @@ +# Transformers without Normalization + +Jiachen Zhu $^{1,2}$ Xinlei Chen $^{1}$ Kaiming He $^{3}$ Yann LeCun $^{1,2}$ Zhuang Liu $^{1,4,\dagger}$ + +$\dagger$ project lead + +$^{1}$ FAIR, Meta $^{2}$ New York University $^{3}$ MIT $^{4}$ Princeton University + +# Abstract + +Normalization layers are ubiquitous in modern neural networks and have long been considered essential. This work demonstrates that Transformers without normalization can achieve the same or better performance using a remarkably simple technique. We introduce Dynamic Tanh (DyT), an element-wise operation $\mathrm{DyT}(\boldsymbol{x}) = \tanh(\alpha \boldsymbol{x})$ , as a drop-in replacement for normalization layers in Transformers. DyT is inspired by the observation that layer normalization in Transformers often produces tanh-like, $S$ -shaped input-output mappings. By incorporating DyT, Transformers without normalization can match or exceed the performance of their normalized counterparts, mostly without hyperparameter tuning. We validate the effectiveness of Transformers with DyT across diverse settings, from recognition to generation, supervised to self-supervised learning, and computer vision to language models. These findings challenge the conventional understanding that normalization layers are indispensable in modern neural networks, and offer new insights into their role in deep networks. + +# 1. Introduction + +Over the past decade, normalization layers have solidified their positions as one of the most fundamental components of modern neural networks. It all traces back to the invention of batch normalization in 2015 [41], which enabled drastically faster and better convergence in visual recognition models and quickly gained momentum in the following years. Since then, many variants of normalization layers have been proposed for different network architectures or domains [5, 81, 83, 89]. Today, virtually all modern networks use normalization layers, with layer normalization (Layer Norm, or LN) [5] being one of the most popular, particularly in the dominant Transformer architecture [20, 82]. + +The widespread adoption of normalization layers is largely driven by their empirical benefits in optimization [10, 68]. In addition to achieving better results, they help accelerate and stabilize convergence. As neural networks + +![](images/41765ed66bf18a82ea42bdaf38ecff4ef1c83483b704ee0c0ee6722d766516b4.jpg) +Figure 1. Left: original Transformer block. Right: block with our proposed Dynamic Tanh (DyT) layer. DyT is a straightforward replacement for commonly used Layer Norm [5] (in some cases RMSNorm [89]) layers. Transformers with DyT match or exceed the performance of their normalized counterparts. + +become wider and deeper, this necessity becomes ever more critical [11, 38]. Consequently, normalization layers are widely regarded as crucial, if not indispensable, for effective training of deep networks. This belief is subtly evidenced by the fact that, in recent years, novel architectures often seek to replace attention or convolution layers [22, 29, 75, 78], but almost always retain normalization. + +This paper challenges this belief by introducing a simple alternative to normalization layers in Transformers. Our exploration starts with the observation that LN layers map their inputs to outputs with tanh-like, $S$ -shaped curves, scaling the input activations while squashing the extreme values. Inspired by this insight, we propose an elementwise operation termed Dynamic Tanh (DyT), defined as: $\mathrm{DyT}(\boldsymbol{x}) = \tanh(\alpha \boldsymbol{x})$ , where $\alpha$ is a learnable parameter. This operation aims to emulate the behavior of LN by learning an appropriate scaling factor through $\alpha$ and squashing extreme values via the bounded tanh function. Notably, unlike normalization layers, it achieves both effects without the need to compute activation statistics. + +Employing DyT is straightforward, as shown in Figure 1: we directly replace existing normalization layers with DyT in architectures such as vision and language Transformers. We empirically demonstrate that models with DyT can train stably and achieve high final performance across a wide range of settings. It often does not require tuning the hy + +![](images/090957d29502bab3ea90da5d620cdba519e426f606475355bf84712d93e89b94.jpg) +LN output (y axis) vs. LN input (x axis) +Figure 2. Output vs. input of selected layer normalization (LN) layers in Vision Transformer (ViT) [20], wav2vec 2.0 (a Transformer model for speech) [7], and Diffusion Transformer (DiT) [64]. We sample a mini-batch of samples and plot the input / output values of four LN layers in each model. The outputs are before the affine transformation in LN. The $S$ -shaped curves highly resemble that of a tanh function. The more linear shapes in earlier layers can also be captured by the center part of a tanh curve. This motivates us to propose Dynamic Tanh (DyT) as a replacement, with a learnable scalar $\alpha$ to account for different scales on the $x$ axis. + +perparameters on the original architecture. Our work challenges the notion that normalization layers are indispensable for training modern neural networks and provides empirical insights into the properties of normalization layers. + +# 2. Background: Normalization Layers + +We begin by reviewing the normalization layers. Most normalization layers share a common formulation. Given an input $\pmb{x}$ with shape $(B, T, C)$ , where $B$ is the batch size, $T$ is the number of tokens, and $C$ is the embedding dimension per token, the output is generally computed as: + +$$ +\operatorname {n o r m a l i z a t i o n} (\boldsymbol {x}) = \gamma * \left(\frac {\boldsymbol {x} - \boldsymbol {\mu}}{\sqrt {\sigma^ {2} + \epsilon}}\right) + \beta \tag {1} +$$ + +where $\epsilon$ is a small constant, and $\gamma$ and $\beta$ are learnable vector parameters of shape $(C,)$ . They are "scaling" and "shifting" affine parameters that allow the output to be in any range. The terms $\mu$ and $\sigma^2$ denote the mean and variance of the input. Different methods mainly differ in how these two statistics are computed. This results in $\mu$ and $\sigma^2$ having different dimensions, each with appropriate broadcasting needed in the tensor calculation. + +Batch normalization (BN) [41] is the first modern normalization layer, and it has been primarily used in Con + +vNet models [33, 76, 84]. Its introduction represents a major milestone in deep learning architecture designs. BN computes the mean and variance across both the batch and token dimensions, specifically: $\mu_{k} = \frac{1}{BT}\sum_{i,j}x_{ijk}$ and $\sigma_k^2 = \frac{1}{BT}\sum_{i,j}(x_{ijk} - \mu_k)^2$ . Other normalization layers popular in ConvNets, such as group normalization [83] and instance normalization [81], were initially proposed for specialized tasks such as object detection and image stylization. They share the same overall formulation but differ in the axes and ranges over which the statistics are computed. + +Layer normalization (LN) [5] and root mean square normalization (RMSNorm) [89] are the major two types of normalization layers used in Transformer architectures. LN computes these statistics independently for each token in each sample, where $\mu_{ij} = \frac{1}{C}\sum_k x_{ijk}$ and $\sigma_{ij}^2 = \frac{1}{C}\sum_k(x_{ijk} - \mu_{ij})^2$ . RMSNorm [89] simplifies LN by removing the mean-centering step and normalizing the input with $\mu_{ij} = 0$ and $\sigma_{ij}^2 = \frac{1}{C}\sum_k x_{ijk}^2$ . Today, most modern neural networks use LN due to its simplicity and universality. Recently, RMSNorm has gained popularity, particularly in language models like T5 [66], LLaMA [21, 79, 80], Mistral [43], Qwen [8, 87], InternLM [13, 91] and DeepSeek [30, 49]. The Transformers we examine in this work all use LN, except that LLaMA uses RMSNorm. + +![](images/41a32d50a2a8c9a3a51d104555d6347f8be77b6e90e8e8694681af64cd8ac114.jpg) +Figure 3. Output vs. input of two LN layers, with tensor elements colored to indicate different channel and token dimensions. The input tensor has a shape of (samples, tokens, and channels), with elements visualized by assigning consistent colors to the same tokens (left two panels) and channels (right two panels). Left two panels: points representing the same token (same color) form straight lines across different channels, as LN operates linearly across channels for each token. Interestingly, when plotted collectively, these lines form a non-linear tanh-shaped curve. Right two panels: each channel's input spans different ranges on the $x$ -axis, contributing distinct segments to the overall tanh-shaped curve. Certain channels (e.g., red, green, and pink) exhibit more extreme $x$ values, which are squashed by LN. + +# 3. What Do Normalization Layers Do? + +Analysis setup. We first empirically study the behaviors of normalization layers in trained networks. For this analysis, we take a Vision Transformer model (ViT-B) [20] trained on ImageNet-1K [19], a wav2vec 2.0 Large Transformer model [7] trained on LibriSpeech [63], and a Diffusion Transformer (DiT-XL) [64] trained on ImageNet-1K. In all cases, LN is applied in every Transformer block and before the final linear projection. + +For all three trained networks, we sample a mini-batch of samples and do a forward pass through the network. We then measure the input and output for the normalization layers, i.e., tensors immediately before and after the normalization operation, before the learnable affine transformation. Since LN preserves the dimensions of the input tensor, we can establish a one-to-one correspondence between input and output elements, allowing a direct visualization of their relationship. We plot the resulting mappings in Figure 2. + +Tanh-like mappings with layer normalization. For all three models, in earlier LN layers (1st column of Figure 2), we find this input-output relationship to be mostly linear, resembling a straight line in an $x - y$ plot. However, deeper LN layers are where we make more intriguing observations. + +A striking observation from these deeper layers is that most of these curves' shapes resemble full or partial $S$ -shaped curves represented by a tanh function. One might expect LN layers to linearly transform the input tensor, as subtracting the mean and dividing by standard deviation are linear operations. LN normalizes in a per-token manner, only linearly transforming each token's activations. As tokens have different mean and standard deviation values, the linearity does not hold collectively on all activations of the input tensor. Nonetheless, it is surprising that the non-linear transformation is highly similar to a scaled tanh function. + +For such an $S$ -shaped curve, we note that the central part, represented by points with $x$ values close to zero, is mainly in a linear shape. Most points $(\sim 99\%)$ fall in this linear range. However, there are many points that clearly fall out + +of this range, which are considered to have "extreme" values, e.g., those with $x$ larger than 50 or smaller than -50 in the ViT model. Normalization layers' main effect for these values is to squash them into less extreme values, more in line with the majority of points. This is where normalization layers could not approximated by a simple affine transformation layer. We hypothesize this non-linear and disproportional squashing effect on extreme values is what makes normalization layers important and indispensable. + +Recent findings by Ni et al. [62] similarly highlight the strong non-linearities introduced by LN layers, demonstrating how the non-linearity enhances a model's representational capacity. Moreover, this squashing behavior mirrors the saturation properties of biological neurons for large inputs, a phenomenon observed about a century ago [1-3]. + +Normalization by tokens and channels. How does an LN layer perform a linear transformation for each token but also squash the extreme values in such a non-linear fashion? To understand this, we visualize the points grouped by tokens and channels, respectively. This is plotted in Figure 3 by taking the second and third subplots for ViT from Figure 2, but with a sampled subset of points for more clarity. When we select the channels to plot, we make sure to include the channels with extreme values. + +On the left two panels of Figure 3, we visualize each token's activations using the same color. We observe that all points from any single token do form a straight line. However, since each token has a different variance, the slopes are different. Tokens with smaller input $x$ ranges tend to have smaller variance, and the normalization layer will divide their activations using a smaller standard deviation, hence producing a larger slope in the straight line. Collectively, they form an $S$ -shaped curve that resembles a tanh function. In the two panels on the right, we color each channel's activations using the same color. We find that different channels tend to have drastically different input ranges, with only a few channels (e.g., red, green, and pink) exhibiting large extreme values. These are the channels that get squashed the most by the normalization layer. + +# 4. Dynamic Tanh (DyT) + +Inspired by the similarity between shapes of normalization layers and a scaled tanh function, we propose Dynamic Tanh (DyT) as a replacement for normalization layers. Given input tensor $\pmb{x}$ , a DyT layer is defined as follows: + +$$ +\mathrm {D y T} (\boldsymbol {x}) = \boldsymbol {\gamma} * \tanh (\alpha \boldsymbol {x}) + \boldsymbol {\beta} \tag {2} +$$ + +where $\alpha$ is a learnable scalar parameter that allows scaling the input differently based on its range, accounting for varying $x$ scales (Figure 2). This is also why we name the whole operation "Dynamic" Tanh. $\gamma$ and $\beta$ are learnable, per-channel vector parameters, the same as those used in all normalization layers—they allow the output to scale back to any scales. This is sometimes considered a separate affine layer; for our purposes, we consider them to be part of the DyT layer, just like how normalization layers also include them. See Algorithm 1 for Pytorch-like pseudocode of DyT. + +Algorithm 1 Pseudocode of DyT layer. +input x has the shape of [B, T, C] +# B: batch size, T: tokens, C: dimension +class DyT(Module): + def __init__(self, C, init\_alpha): + super().__init() + self. $\alpha =$ Parameter(ones(1) $\star$ init\_alpha) + self. $\gamma =$ Parameter(ones(C)) + self. $\beta =$ Parameter(zeros(C)) +def forward(self, x): + x = tanh(self.alpha $\star$ x) + return self. $\gamma =$ x + self. $\beta$ + +Integrating DyT layers into an existing architecture is straightforward: one DyT layer replaces one normalization layer (see Figure 1). This applies to normalization layers within attention blocks, FFN blocks, and the final normalization layer. Although DyT may look like or be considered an activation function, this study only uses it to replace normalization layers without altering parts of the activation functions in the original architectures, such as GELU or ReLU. Other parts of the network remain intact. We also observe that there is minimal need to tune the hyperparameters for the original architectures for DyT to perform well. + +On scaling parameters. We always simply initialize $\gamma$ to an all-one vector and $\beta$ to an all-zero vector following normalization layers. For the scalar parameter $\alpha$ , a default initialization of 0.5 is generally sufficient, except for LLM training. A detailed analysis of $\alpha$ initialization is provided in Section 7. Unless explicitly stated otherwise, $\alpha$ is initialized to 0.5 in our subsequent experiments. + +Remarks. DyT is not a new type of normalization layer, as it operates on each input element from a tensor independently during a forward pass without computing statistics or other types of aggregations. It does, however, preserve the effect of normalization layers in squashing the extreme values in a non-linear fashion while almost linearly transforming the very central parts of the input. + +# 5. Experiments + +To demonstrate the effectiveness of DyT, we experiment with Transformers and a few other modern architectures across a diverse range of tasks and domains. In each experiment, we replace the LN or RMSNorm in the original architectures with DyT layers and follow the official open-source protocols to train and test both versions of the models. Detailed instructions for reproducing our results are provided in Appendix A. Notably, to highlight the simplicity of adapting DyT, we use hyperparameters identical to those used by the normalized counterparts. For completeness, additional results regarding tuning of learning rates and initial values of $\alpha$ are provided in Appendix F. + +Supervised learning in vision. We train Vision Transformer (ViT) [20] and ConvNeXt [50] of Base" and Large" sizes on the ImageNet-1K classification task [19]. These models are selected due to their popularity and distinct operations: attention in ViT and convolution in ConvNeXt. Table 1 reports the top-1 classification accuracies. DyT performs slightly better than LN across both architectures and model sizes. We further plot the training loss for ViT-B and ConvNeXt-B in Figure 4. The curves show that the convergence behaviors of DyT and LN-based models are aligned. + +
modelLNDyTchange
ViT-B82.3%82.5%↑0.2%
ViT-L83.1%83.6%↑0.5%
ConvNeXt-B83.7%83.7%-
ConvNeXt-L84.3%84.4%↑0.1%
+ +Table 1. Supervised classification accuracy on ImageNet-1K. DyT achieves better or similar performance than LN across both architectures and model sizes. + +![](images/b7baae6506173d2c014485304a72c5faa58779d1c2990cd0bf8966405b7f33e5.jpg) + +![](images/9d2ae29f8036826d3469fa1d6544060abde19d956d0b8893473fb7f88ce49646.jpg) +Figure 4. Training loss curves for ViT-B and ConvNeXt-B. The curves for both models exhibit similar patterns between LN and DyT, suggesting that they may share similar learning dynamics. + +![](images/bedeb54687c14d9c30c54a7f5f5e4f234f7688487352d02ccdd79750a8c92102.jpg) + +![](images/fd7954f2a83eb567ee516aa747bbd484bc01286a8ebf046e01d2af713bcfd322.jpg) + +![](images/c85074c4872705decb2ad04005aa7d38d0cbbe528a7411e0be5489e82f7a7e41.jpg) +Figure 5. LLaMA pretraining loss. The loss curves of DyT and RMSNorm models are closely aligned across model sizes. + +![](images/31f8b8dc47ee5d769dbfb8e71e40d4a7e0888ac2d5e9f80ed9ed4e00230c00e4.jpg) + +Self-supervised learning in vision. We benchmark with two popular visual self-supervised learning methods: masked autoencoders (MAE) [34] and DINO [14]. Both by default use Vision Transformers as backbones, but have different training objectives: MAE is trained with a reconstruction loss, and DINO uses a joint-embedding loss [47]. Following the standard self-supervised learning protocol, we first pretrain models on ImageNet-1K without using labels and then test the pretrained models by attaching a classification layer and fine-tuning them with labels. The fine-tuning results are presented in Table 2. DyT consistently performs on par with LN in self-supervised learning tasks. + +
modelLNDyTchange
MAE ViT-B83.2%83.2%-
MAE ViT-L85.5%85.4%↓0.1%
DINO ViT-B (patch size 16)83.2%83.4%↑0.2%
DINO ViT-B (patch size 8)84.1%84.5%↑0.4%
+ +Diffusion models. We train three Diffusion Transformer (DiT) models [64] of sizes B, L and XL on ImageNet-1K [19]. The patch size is 4, 4, and 2, respectively. Note that in DiT, the LN layers' affine parameters are used for class conditioning in DiT, and we keep them that way in our DyT experiments, only replacing the normalizing transformation with the $\tanh (\alpha x)$ function. After training, we evaluate the Fréchet Inception Distance (FID) scores using the standard ImageNet "reference batch", as presented in Table 3. DyT achieves comparable or improved FID over LN. + +Table 2. Self-supervised learning accuracy on ImageNet-1K. DyT performs on par with LN across different methods and sizes. + +
modelLNDyTchange
DiT-B64.963.9↓1.0
DiT-L45.945.7↓0.2
DiT-XL19.920.8↑0.9
+ +Large Language Models. We pretrain LLaMA 7B, 13B, 34B, and 70B models [21, 79, 80] to assess DyT performance relative to RMSNorm [89], the default normalization layer used in LLaMA. The models are trained on The Pile dataset [25] with 200B tokens, following the original recipe outlined in LLaMA [80]. On LLaMA with DyT, we add a learnable scalar parameter after the initial embedding layer, and adjust the initial value of $\alpha$ , as detailed in Section 7. We report the loss value after training and also follow OpenLLaMA [26] to benchmark the models on 15 zero-shot tasks from lm-eval [24]. As shown in Table 4, DyT performs on par with RMSNorm across all four model sizes. Figure 5 illustrates the loss curves, demonstrating similar trends across all model sizes, with training losses closely aligned throughout training. + +Table 3. Image generation quality (FID) on ImageNet. DyT achieves comparable FID scores to LN across various model sizes. + +
score / lossRMSNormDyTchange
LLaMA 7B0.513 / 1.590.513 / 1.60- / ↑0.01
LLaMA 13B0.529 / 1.530.529 / 1.54- / ↑0.01
LLaMA 34B0.536 / 1.500.536 / 1.50- / -
LLaMA 70B0.549 / 1.450.549 / 1.45- / -
+ +Self-supervised learning in speech. We pretrain two wav2vec 2.0 Transformer models [7] on the LibriSpeech dataset [63]. We retain the first group normalization layer, as it functions as data normalization to handle the unnormalized input data. Table 5 reports the final validation loss. We observe that DyT performs comparably to LN. + +Table 4. Language models' training loss and average performance with 15 zero-shot 1m-eval tasks. DyT achieves a comparable zero-shot performance and training loss to RMSNorm. + +
modelLNDyTchange
wav2vec 2.0 Base1.951.95-
wav2vec 2.0 Large1.921.91↓0.01
+ +Table 5. Speech pretraining validation loss on LibriSpeech. DyT performs comparably to LN for both wav2vec 2.0 models. + +![](images/3073ff5c508db6f14beaea3b3b12d9fad7c867dad3718995952afcf23c4abac5.jpg) +Figure 6. Curves of three squashing functions: tanh, hardtanh, and sigmoid. All three functions squash inputs into a bounded range, but tanh achieves the best performance when used in DyT layers. We suspect it is due to its smoothness and zero-centered properties. + +![](images/600e1a9390e4d16dd57e59a540132dde226199e7d1f1ab5b1895852b2bbbed28.jpg) +Figure 7. Left: For two selected DyT layers from the ViT-B model, we track $\alpha$ and the inverse of the standard deviation (1/std) of activations at the end of each epoch, observing that they evolve together during training. Right: We plot the final $\alpha$ values of two models, ViT-B and ConvNeXt-B, against the 1/std of the input activations, demonstrating a strong correlation between the two values. + +![](images/da316b4b6400f407422da48a2d297effac125bbd60a4442150c2968c50a470e1.jpg) + +DNA sequence modeling. On the long-range DNA sequence modeling task, we pretrain the HyenaDNA model [60] and the Caduceus model [69]. The pretraining uses human reference genome data from [27], and the evaluation is on GenomicBenchmarks [28]. Results are presented in Table 6. DyT maintains performance comparable to LN. + +
modelLNDyTchange
HyenaDNA [60]85.2%85.2%-
Caduceus [69]86.9%86.9%-
+ +# 6. Analysis + +We conduct two studies examining the roles of the tanh function and the learnable scale $\alpha$ . + +# 6.1. Ablations of tanh and $\alpha$ + +To further investigate the role of tanh and $\alpha$ in DyT, we conduct experiments to evaluate the model's performance when these components are altered or removed. + +Replacing and removing tanh. We replace tanh in DyT layers with alternative squashing functions, specifically hardtanh and sigmoid (Figure 6), while keeping the learnable scalar $\alpha$ intact. Furthermore, we assess the impact of completely removing tanh by replacing it with the identity function while still retaining $\alpha$ . As shown in Table 7, the squashing function is essential for stable training. Using the identity function leads to unstable training and divergence, + +Table 6. DNA classification accuracy on GenomicBenchmarks, averaged over each dataset in GenomicBenchmarks. DyT achieves comparable performance to LN. + +
modelidentitytanhhardtanhsigmoid
ViT-S58.5% → failed80.3%79.9%79.6%
ViT-B61.0% → failed82.5%82.2%81.6%
+ +whereas squashing functions enable stable training. Among the squashing functions, tanh performs the best. This is possibly due to its smoothness and zero-centered properties. + +Removing $\alpha$ . Next, we evaluate the impact of removing the learnable $\alpha$ while retaining the squashing functions. As shown in Table 8, removing $\alpha$ results in performance degradation across all squashing functions, highlighting the critical role of $\alpha$ in overall model performance. + +Table 7. ImageNet-1K classification accuracy with different squashing functions. All experiments follow the same training recipe as the original LN-based models. Squashing functions play a crucial role in preventing divergence, with tanh achieving the highest performance among the three functions. “ $\rightarrow$ failed” indicates that training diverged after some progress, with the preceding number representing the accuracy reached before divergence. + +
modeltanhhardtanhsigmoid
without α81.1%80.7%80.7%
with α82.5%82.2%81.6%
+ +Table 8. ImageNet-1K classification accuracy with ViT-B. All experiments follow the same training recipe as the original LN-based models. $\alpha$ is essential for enhancing model performance. + +# 6.2. Values of $\alpha$ + +During training. Our analysis reveals that the $\alpha$ closely tracks the $1 / \mathrm{std}$ of activations throughout training. As illustrated in the left panel of Figure 7, $\alpha$ first decrease and then increase during training, but always fluctuate consistently with the standard deviation of activations. This supports the important role of $\alpha$ in maintaining activations within a suitable range, which leads to stable and effective training. + +After training. Our further analysis of the final values of $\alpha$ in trained networks reveals a strong correlation with the 1/std of the input activations. As shown on the right panel of Figure 7, higher 1/std values generally correspond to larger $\alpha$ values, and vice versa. Additionally, we observe that deeper layers tend to have activations with larger standard deviations. This trend aligns with characteristics of deep residual networks, as shown in Brock et al. [11] for ConvNets, and Sun et al. [74] for Transformers. + +Both analyses suggest that $\alpha$ acts as a normalization mechanism, learning values close to $1 / \mathrm{std}$ of the input activations. Unlike LN, which normalizes per token, $\alpha$ normalizes the entire input collectively and cannot suppress extreme values non-linearly. + +![](images/796e8ffa5539831c9248f034cd69ad5b6f6382596f3fc0815caee396a7edf903.jpg) + +![](images/4a715c71c0d2236260abcd4575cf0ca5b2eed1c716f991ce94d1ffa7e843645e.jpg) + +![](images/f188c861c933703a664b47818539b4dbf8a25f2dad051cbc827bcc9fc106e59d.jpg) + +![](images/3373dd2e375fe35c87f2a2861451b4945b27c7a9cc88527e40deb375db3925a1.jpg) +Figure 8. Performance of different tasks across different $\alpha_0$ values. We benchmark the performance of all non-LLM tasks used in Section 5 with different initial values of $\alpha$ . Performance remains stable across a wide range of $\alpha_0$ values. The only exception is that supervised ViT-L models (top right panel) will diverge for $\alpha_0$ values larger than 0.6. + +![](images/9ec84880ebfc85127aa97cf2522f5bae6a416a96c5087ab102710cd4b97f4e4d.jpg) + +![](images/8c6b92be8d50717e7d543923b4f9f103605016db23ca45b9d3a381c324fc62ba.jpg) + +# 7. Initialization of $\alpha$ + +We find that tuning the initialization of $\alpha$ (denoted $\alpha_0$ ) rarely leads to significant performance improvements. The only exception is LLM training, where careful tuning of $\alpha_0$ yields noticeable performance gains. In this section, we detail our findings on the impact of $\alpha$ initialization. + +# 7.1. Initialization of $\alpha$ for Non-LLM Models + +Non-LLM models are relatively insensitive to $\alpha_0$ . Figure 8 shows the effect of varying $\alpha_0$ on validation performance across different tasks. All experiments follow the original setup and hyperparameters of their respective recipe. We observe that performance remains stable across a wide range of $\alpha_0$ values, with values between 0.5 and 1.2 generally yielding good results. We observe that adjusting $\alpha_0$ typically affects only the early stages of the training curves. The main exception is supervised ViT-L experiments, where training becomes unstable and diverges when $\alpha_0$ exceeds 0.6. In such cases, reducing the learning rate restores stability, as detailed below. + +Smaller $\alpha_0$ results in more stable training. Building on previous observations, we further analyze the factors contributing to training instability. Our findings suggest that increasing either the model size or the learning rate requires lowering $\alpha_0$ to ensure stable training. Conversely, a higher $\alpha_0$ requires a lower learning rate to mitigate training instability. Figure 9 shows the ablation of the training stability of supervised ViT with ImageNet-1K dataset. We vary learning rates, model sizes, and $\alpha_0$ values. Training a larger model is more prone to failure, requiring smaller $\alpha_0$ values or learning rates for stable training. A similar instability pattern is also observed in LN-based models under comparable conditions, and setting $\alpha_0 = 0.5$ results in a stability pattern similar to that of LN. + +Setting $\alpha_0 = 0.5$ as the default. Based on our findings, we set $\alpha_0 = 0.5$ as the default value for all non-LLM models. This setting provides training stability comparable to LN while maintaining strong performance. + +# 7.2. Initialization of $\alpha$ for LLMs + +Tuning $\alpha_0$ enhances LLM performance. As discussed earlier, the default setting of $\alpha_0 = 0.5$ generally performs well across most tasks. However, we find tuning $\alpha_0$ can substantially improve LLM performance. We tune $\alpha_0$ across LLaMA models by pretraining each on 30B tokens and comparing their training losses. Table 9 summarizes the tuned $\alpha_0$ values for each model. Two key findings emerge: + +1. Larger models require smaller $\alpha_0$ values. Once the optimal $\alpha_0$ is determined for smaller models, the search space for larger models can be reduced accordingly. +2. Higher $\alpha_0$ values for attention blocks improve performance. We find that initializing $\alpha$ with higher values for DyT layers in attention blocks and lower values for DyT layers in other locations (i.e., within FFN blocks or before the final linear projection) improves performance. + +
modelwidthdepthoptimal α0 (attention/other)
LLaMA 7B4096320.8/0.2
LLaMA 13B5120400.6/0.15
LLaMA 34B8196480.2/0.05
LLaMA 70B8196800.2/0.05
+ +Table 9. Optimal $\alpha_0$ for different LLaMA models. Larger models require smaller $\alpha_0$ values. We find it is important to initialize $\alpha$ differently in (1) attention blocks ("attention"), versus (2) the FFN blocks, and the final DyT layer before outputs ("other"). $\alpha_0$ in attention blocks require larger values. + +Further details and visualizations of the tuning of the initial value of $\alpha$ of LLMs are provided in the appendix C. + +![](images/21493d7352e45b062d793c5e17907856f6cf5d55b553580b1afa8b8672c0e646.jpg) +Figure 9. Stability across varying $\alpha_0$ values, learning rates, and model sizes. We train supervised ViT models on the ImageNet-1K dataset and observe that larger models are more prone to instability for both LN and DyT models. Lowering the learning rate or reducing $\alpha_0$ enhances stability. LN shows similar stability to DyT with $\alpha_0 = 0.5$ . + +# 8. Related Work + +Mechanisms of normalization layers. There has been a rich line of work investigating normalization layers' role in enhancing model performance through various mechanisms. These include stabilizing gradient flow during training [9, 17, 52], reducing sensitivity to weight initialization [18, 70, 90], moderating outlier eigenvalues [10, 44], auto-tuning learning rates [4, 77], and smoothing the loss landscape for more stable optimization [68]. These earlier works focused on studying batch normalization. Recent studies [16, 53, 59] further highlight the connection between normalization layers and sharpness reduction, which contributes to better generalization. + +Normalization in Transformers. With the rise of Transformer [82], research has increasingly focused on layer normalization [5], which has proven particularly effective for sequential data in natural language tasks [61, 85, 86]. Recent work [62] reveals that layer normalization introduces strong non-linearity, enhancing the model's representational capacity. Additionally, studies [48, 51] demonstrate that modifying the location of normalization layers within Transformers can improve convergence properties. + +Removing normalization. Many studies have explored how to train deep models without normalization layers. Several works [6, 18, 90] explore alternative weight initialization schemes to stabilize training. Self-normalizing networks [45] introduce scaled exponential linear units (SELUs) and a carefully chosen initialization scheme to maintain stable activations and gradient flow without the need for explicit normalization layers. The pioneering work by Brock et al. [11, 12] show that high-performing ResNets can be trained without normalization [72] through combination of initialization techniques [18], weight normalization [37, 65, 67], and adaptive gradient clipping [12]. Additionally, their training strategy incorporates extensive data augmentation [15] and regularization [36, 73]. The studies above are based on various ConvNet models. + +In Transformer architectures, He and Hofmann [32] explore modifications to Transformer blocks that reduce reliance on normalization layers and skip connections. Jha and Reagen [42] introduce AERO, a Softmax-only LLM that improves inference efficiency and privacy with minimal performance loss. Alternatively, Heimersheim [35] propose a method to gradually remove LN from pretrained networks by fine-tuning the model after removing each normalization layer. Unlike previous approaches, DyT requires minimal modifications to both the architecture and the training recipe. Despite its simplicity, DyT achieves stable training and comparable performance. + +# 9. Limitations + +We conduct experiments on networks using either LN or RMSNorm because of their popularity in Transformers and other modern architectures. Preliminary experiments (see Appendix E) indicate that DyT struggles to replace BN directly in classic networks like ResNets. It remains to be studied in more depth whether and how DyT can adapt to models with other types of normalization layers. + +# 10. Conclusion + +In this work, we demonstrate modern neural networks, in particular Transformers, can be trained without normalization layers. This is done through Dynamic Tanh (DyT), a simple replacement for traditional normalization layers. It adjusts the input activation range via a learnable scaling factor $\alpha$ and then squashes the extreme values through an $S$ -shaped tanh function. Although a simpler function, it effectively captures the behavior of normalization layers. Under various settings, models with DyT match or exceed the performance of their normalized counterparts. The findings challenge the conventional understanding of the necessity of normalization layers in training modern neural networks. Our study also contributes to understanding the mechanisms of normalization layers, one of the most fundamental building blocks in deep neural networks. + +# References + +[1] Edgar D Adrian. The impulses produced by sensory nerve endings: Part 1. The Journal of Physiology, 1926. +[2] Edgar D Adrian and Yngve Zotterman. The impulses produced by sensory nerve-endings: Part 2. the response of a single end-organ. The Journal of Physiology, 1926. +[3] Edgar D Adrian and Yngve Zotterman. The impulses produced by sensory nerve endings: Part 3. impulses set up by touch and pressure. The Journal of Physiology, 1926. +[4] Sanjeev Arora, Zhiyuan Li, and Kaifeng Lyu. Theoretical analysis of auto rate-tuning by batch normalization. arXiv preprint arXiv:1812.03981, 2018. +[5] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. +[6] Thomas Bachlechner, Bodhisattwa Prasad Majumder, Henry Mao, Gary Cottrell, and Julian McAuley. Rezero is all you need: Fast convergence at large depth. In UAI, 2021. +[7] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. NeurIPS, 2020. +[8] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. +[9] David Balduzzi, Marcus Frean, Lennox Leary, JP Lewis, Kurt Wan-Duo Ma, and Brian McWilliams. The shattered gradients problem: If resnets are the answer, then what is the question? In ICML, 2017. +[10] Nils Bjorck, Carla P Gomes, Bart Selman, and Kilian Q Weinberger. Understanding batch normalization. NeurIPS, 2018. +[11] Andrew Brock, Soham De, and Samuel L Smith. Characterizing signal propagation to close the performance gap in unnormalized resnets. arXiv preprint arXiv:2101.08692, 2021. +[12] Andrew Brock, Soham De, Samuel L Smith, and Karen Simonyan. High-performance large-scale image recognition without normalization. In ICML, 2021. +[13] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arXiv preprint arXiv:2403.17297, 2024. +[14] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. +[15] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In CVPR Workshops, 2020. +[16] Yan Dai, Kwangjun Ahn, and Suvrit Sra. The crucial role of normalization in sharpness-aware minimization. NeurIPS, 2024. +[17] Hadi Daneshmand, Jonas Kohler, Francis Bach, Thomas Hofmann, and Aurelien Lucchi. Batch normalization provably avoids ranks collapse for randomly initialised deep networks. NeurIPS, 2020. + +[18] Soham De and Sam Smith. Batch normalization biases residual blocks towards the identity function in deep networks. NeurIPS, 2020. +[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. +[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. +[21] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. +[22] Leo Feng, Frederick Tung, Mohamed Osama Ahmed, Yoshua Bengio, and Hossein Hajimirsadegh. Were rnns all we needed? arXiv preprint arXiv:2410.01201, 2024. +[23] Foundation Model Stack. Github: FMS FSDP. https://github.com/foundation-model-stack/fms-fsdp. Accessed: 2025-01-23. +[24] Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac'h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lintang Sutawika, Eric Tang, Anish Thite, Ben Wang, Kevin Wang, and Andy Zou. A framework for few-shot language model evaluation. +[25] Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, Shawn Presser, and Connor Leahy. The Pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027, 2020. +[26] Xinyang Geng and Hao Liu. Openlama: An open reproduction of llama, 2023. +[27] Ensembl GRCh38. p13 (genome reference consortium human build 38), insdc assembly, 2013. +[28] Katarína Grešová, Vlastimil Martinek, David Čechák, Petr Šimeček, and Panagiotis Alexiou. Genomic benchmarks: a collection of datasets for genomic sequence classification. BMC Genomic Data, 2023. +[29] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. +[30] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. +[31] HazyResearch. Github: Hyenadna. https://github.com/HazyResearch/hyena-dna.git. Accessed: 2025-01-23. +[32] Bobby He and Thomas Hofmann. Simplifying transformer blocks. arXiv preprint arXiv:2311.01906, 2023. + +[33] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. +[34] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. +[35] Stefan Heimersheim. You can remove gpt2's layernorm by fine-tuning. arXiv preprint arXiv:2409.13710, 2024. +[36] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In ECCV, 2016. +[37] Lei Huang, Xianglong Liu, Yang Liu, Bo Lang, and Dacheng Tao. Centered weight normalization in accelerating training of deep neural networks. In ICCV, 2017. +[38] Lei Huang, Jie Qin, Yi Zhou, Fan Zhu, Li Liu, and Ling Shao. Normalization techniques in training dnns: Methodology, analysis and application. TPAMI, 2023. +[39] Xiao Shi Huang, Felipe Perez, Jimmy Ba, and Maksims Volkovs. Improving transformer optimization through better initialization. In ICML, 2020. +[40] Hugging Face. Hugging Face: LLaMA 2. https://huggingface.co/meta-llama/Llama-2-7b-hf. Accessed: 2025-01-23. +[41] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. +[42] Nandan Kumar Jha and Brandon Reagen. Aero: Softmax-only llms for efficient private inference. arXiv preprint arXiv:2410.13060, 2024. +[43] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. +[44] Ryo Karakida, Shotaro Akaho, and Shun-ichi Amari. The normalization method for alleviating pathological sharpness in wide neural networks. NeurIPS, 2019. +[45] Günter Klambauer, Thomas Unterthiner, Andreas Mayr, and Sepp Hochreiter. Self-normalizing neural networks. NeurIPS, 2017. +[46] Kuleshov Group. Github: Caduceus. https://github.com/kuleshov-group/caduceus.git. Accessed: 2025-01-23. +[47] Yann LeCun. A path towards autonomous machine intelligence version 0.9.2, 2022-06-27. Open Review, 2022. +[48] Pengxiang Li, Lu Yin, and Shiwei Liu. Mix-ln: Unleashing the power of deeper layers by combining pre-ln and post-ln. arXiv preprint arXiv:2412.13795, 2024. +[49] Aixin Liu, Bei Feng, Bin Wang, Bingxuan Wang, Bo Liu, Chenggang Zhao, Chengqi Dengr, Chong Ruan, Damai Dai, Daya Guo, et al. Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model. arXiv preprint arXiv:2405.04434, 2024. +[50] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022. + +[51] Ilya Loshchilov, Cheng-Ping Hsieh, Simeng Sun, and Boris Ginsburg. ngpt: Normalized transformer with representation learning on the hypersphere. arXiv preprint arXiv:2410.01131, 2024. +[52] Ekdeep S Lubana, Robert Dick, and Hidenori Tanaka. Beyond batchnorm: Towards a unified understanding of normalization in deep learning. NeurIPS, 2021. +[53] Kaifeng Lyu, Zhiyuan Li, and Sanjeev Arora. Understanding the generalization benefit of normalization layers: Sharpness reduction. NeurIPS, 2022. +[54] Meta Research. Github: ConvNeXt. https://github.com/facebookresearch/ConvNeXt, . Accessed: 2025-01-23. +[55] Meta Research. Github: DINO. https://github.com/facebookresearch/dino, . Accessed: 2025-01-23. +[56] Meta Research. Github: DiT. https://github.com/facebookresearch/DiT,. Accessed: 2025-01-23. +[57] Meta Research. Github: MAE. https://github.com/facebookresearch/mae, . Accessed: 2025-01-23. +[58] Meta Research. Github: wav2vec 2.0. https://github.com/facebookresearch/fairseq, . Accessed: 2025-01-23. +[59] Maximilian Mueller, Tiffany Vlaar, David Rolnick, and Matthias Hein. Normalization layers are all that sharpness-aware minimization needs. NeurIPS, 2024. +[60] Eric Nguyen, Michael Poli, Marjan Faizi, Armin Thomas, Michael Wornow, Callum Birch-Sykes, Stefano Massaroli, Aman Patel, Clayton Rabideau, Yoshua Bengio, et al. Hyenadna: Long-range genomic sequence modeling at single nucleotide resolution. NeurIPS, 2024. +[61] Toan Q Nguyen and Julian Salazar. Transformers without tears: Improving the normalization of self-attention. arXiv preprint arXiv:1910.05895, 2019. +[62] Yunhao Ni, Yuxin Guo, Junlong Jia, and Lei Huang. On the nonlinearity of layer normalization. arXiv preprint arXiv:2406.01255, 2024. +[63] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In ICASSP, 2015. +[64] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. +[65] Siyuan Qiao, Huiyu Wang, Chenxi Liu, Wei Shen, and Alan Yuille. Micro-batch training with batch-channel normalization and weight standardization. arXiv preprint arXiv:1903.10520, 2019. +[66] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 2020. +[67] Tim Salimans and Durk P Kingma. Weight normalization: A simple reparameterization to accelerate training of deep neural networks. NeurIPS, 2016. +[68] Shibani Santurkar, Dimitris Tsipras, Andrew Ilyas, and Aleksander Madry. How does batch normalization help optimization? NeurIPS, 2018. +[69] Yair Schiff, Chia-Hsiang Kao, Aaron Gokaslan, Tri Dao, Albert Gu, and Volodymyr Kuleshov. Caduceus: Bi + +directional equivariant long-range dna sequence modeling. arXiv preprint arXiv:2403.03234, 2024. +[70] Jie Shao, Kai Hu, Changhu Wang, Xiangyang Xue, and Bhiksha Raj. Is normalization indispensable for training deep neural network? NeurIPS, 2020. +[71] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. +[72] Samuel L Smith, Andrew Brock, Leonard Berrada, and Soham De. Convnets match vision transformers at scale. arXiv preprint arXiv:2310.16764, 2023. +[73] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. JMLR, 2014. +[74] Wenfang Sun, Xinyuan Song, Pengxiang Li, Lu Yin, Yefeng Zheng, and Shiwei Liu. The curse of depth in large language models. arXiv preprint arXiv:2502.05795, 2025. +[75] Yu Sun, Xinhao Li, Karan Dalal, Jiarui Xu, Arjun Vikram, Genghan Zhang, Yann Dubois, Xinlei Chen, Xiaolong Wang, Sanmi Koyejo, et al. Learning to (learn at test time): Rnns with expressive hidden states. arXiv preprint arXiv:2407.04620, 2024. +[76] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016. +[77] Hidenori Tanaka and Daniel Kunin. Noether's learning dynamics: Role of symmetry breaking in neural networks. NeurIPS, 2021. +[78] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlp architecture for vision. NeurIPS, 2021. +[79] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. +[80] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. +[81] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016. +[82] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. +[83] Yuxin Wu and Kaiming He. Group normalization. In ECCV, 2018. +[84] Saining Xie, Ross Girshick, Piotr Dólar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, 2017. +[85] Ruibin Xiong, Yunchang Yang, Di He, Kai Zheng, Shuxin Zheng, Chen Xing, Huishuai Zhang, Yanyan Lan, Liwei Wang, and Tieyan Liu. On layer normalization in the transformer architecture. In ICML, 2020. + +[86] Jingjing Xu, Xu Sun, Zhiyuan Zhang, Guangxiang Zhao, and Junyang Lin. Understanding and improving layer normalization. NeurIPS, 2019. +[87] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. +[88] Shuangfei Zhai, Tatiana Likhomanenko, Etai Littwin, Dan Busbridge, Jason Ramapuram, Yizhe Zhang, Jiatao Gu, and Joshua M Susskind. Stabilizing transformer training by preventing attention entropy collapse. In ICML, 2023. +[89] Biao Zhang and Rico Sennrich. Root mean square layer normalization. NeurIPS, 2019. +[90] Hongyi Zhang, Yann N Dauphin, and Tengyu Ma. Fixup initialization: Residual learning without normalization. arXiv preprint arXiv:1901.09321, 2019. +[91] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. \ No newline at end of file diff --git a/2025/Transformers without Normalization/images.zip b/2025/Transformers without Normalization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7302b60b33350e38b5df18eece672a6244fe6158 --- /dev/null +++ b/2025/Transformers without Normalization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae855030bc39426215cd3df98659749b42f578f373c84d1844bf52751bb84bf3 +size 591536 diff --git a/2025/Transformers without Normalization/layout.json b/2025/Transformers without Normalization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bbeeb9dc75807fd80734e09593a9fb9e72423c28 --- /dev/null +++ b/2025/Transformers without Normalization/layout.json @@ -0,0 +1,10818 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 190, + 103, + 421, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 103, + 421, + 119 + ], + "spans": [ + { + "bbox": [ + 190, + 103, + 421, + 119 + ], + "type": "text", + "content": "Transformers without Normalization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 141, + 497, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 141, + 497, + 159 + ], + "spans": [ + { + "bbox": [ + 112, + 141, + 497, + 159 + ], + "type": "text", + "content": "Jiachen Zhu" + }, + { + "bbox": [ + 112, + 141, + 497, + 159 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 112, + 141, + 497, + 159 + ], + "type": "text", + "content": " Xinlei Chen" + }, + { + "bbox": [ + 112, + 141, + 497, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 112, + 141, + 497, + 159 + ], + "type": "text", + "content": " Kaiming He" + }, + { + "bbox": [ + 112, + 141, + 497, + 159 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 112, + 141, + 497, + 159 + ], + "type": "text", + "content": " Yann LeCun" + }, + { + "bbox": [ + 112, + 141, + 497, + 159 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 112, + 141, + 497, + 159 + ], + "type": "text", + "content": " Zhuang Liu" + }, + { + "bbox": [ + 112, + 141, + 497, + 159 + ], + "type": "inline_equation", + "content": "^{1,4,\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 281, + 164, + 330, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 164, + 330, + 177 + ], + "spans": [ + { + "bbox": [ + 281, + 164, + 330, + 177 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 281, + 164, + 330, + 177 + ], + "type": "text", + "content": " project lead" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 182, + 473, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 182, + 473, + 198 + ], + "spans": [ + { + "bbox": [ + 137, + 182, + 473, + 198 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 182, + 473, + 198 + ], + "type": "text", + "content": "FAIR, Meta " + }, + { + "bbox": [ + 137, + 182, + 473, + 198 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 137, + 182, + 473, + 198 + ], + "type": "text", + "content": "New York University " + }, + { + "bbox": [ + 137, + 182, + 473, + 198 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 137, + 182, + 473, + 198 + ], + "type": "text", + "content": "MIT " + }, + { + "bbox": [ + 137, + 182, + 473, + 198 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 137, + 182, + 473, + 198 + ], + "type": "text", + "content": "Princeton University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 224, + 200, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 224, + 200, + 237 + ], + "spans": [ + { + "bbox": [ + 151, + 224, + 200, + 237 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 250, + 297, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 250, + 297, + 477 + ], + "spans": [ + { + "bbox": [ + 54, + 250, + 297, + 477 + ], + "type": "text", + "content": "Normalization layers are ubiquitous in modern neural networks and have long been considered essential. This work demonstrates that Transformers without normalization can achieve the same or better performance using a remarkably simple technique. We introduce Dynamic Tanh (DyT), an element-wise operation " + }, + { + "bbox": [ + 54, + 250, + 297, + 477 + ], + "type": "inline_equation", + "content": "\\mathrm{DyT}(\\boldsymbol{x}) = \\tanh(\\alpha \\boldsymbol{x})" + }, + { + "bbox": [ + 54, + 250, + 297, + 477 + ], + "type": "text", + "content": ", as a drop-in replacement for normalization layers in Transformers. DyT is inspired by the observation that layer normalization in Transformers often produces tanh-like, " + }, + { + "bbox": [ + 54, + 250, + 297, + 477 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 54, + 250, + 297, + 477 + ], + "type": "text", + "content": "-shaped input-output mappings. By incorporating DyT, Transformers without normalization can match or exceed the performance of their normalized counterparts, mostly without hyperparameter tuning. We validate the effectiveness of Transformers with DyT across diverse settings, from recognition to generation, supervised to self-supervised learning, and computer vision to language models. These findings challenge the conventional understanding that normalization layers are indispensable in modern neural networks, and offer new insights into their role in deep networks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 501, + 135, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 135, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 135, + 514 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 521, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 665 + ], + "type": "text", + "content": "Over the past decade, normalization layers have solidified their positions as one of the most fundamental components of modern neural networks. It all traces back to the invention of batch normalization in 2015 [41], which enabled drastically faster and better convergence in visual recognition models and quickly gained momentum in the following years. Since then, many variants of normalization layers have been proposed for different network architectures or domains [5, 81, 83, 89]. Today, virtually all modern networks use normalization layers, with layer normalization (Layer Norm, or LN) [5] being one of the most popular, particularly in the dominant Transformer architecture [20, 82]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "The widespread adoption of normalization layers is largely driven by their empirical benefits in optimization [10, 68]. In addition to achieving better results, they help accelerate and stabilize convergence. As neural networks" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 324, + 224, + 545, + 323 + ], + "blocks": [ + { + "bbox": [ + 324, + 224, + 545, + 323 + ], + "lines": [ + { + "bbox": [ + 324, + 224, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 324, + 224, + 545, + 323 + ], + "type": "image", + "image_path": "41765ed66bf18a82ea42bdaf38ecff4ef1c83483b704ee0c0ee6722d766516b4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 326, + 555, + 382 + ], + "lines": [ + { + "bbox": [ + 313, + 326, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 555, + 382 + ], + "type": "text", + "content": "Figure 1. Left: original Transformer block. Right: block with our proposed Dynamic Tanh (DyT) layer. DyT is a straightforward replacement for commonly used Layer Norm [5] (in some cases RMSNorm [89]) layers. Transformers with DyT match or exceed the performance of their normalized counterparts." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 390, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 555, + 475 + ], + "type": "text", + "content": "become wider and deeper, this necessity becomes ever more critical [11, 38]. Consequently, normalization layers are widely regarded as crucial, if not indispensable, for effective training of deep networks. This belief is subtly evidenced by the fact that, in recent years, novel architectures often seek to replace attention or convolution layers [22, 29, 75, 78], but almost always retain normalization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 480, + 556, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 556, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 556, + 635 + ], + "type": "text", + "content": "This paper challenges this belief by introducing a simple alternative to normalization layers in Transformers. Our exploration starts with the observation that LN layers map their inputs to outputs with tanh-like, " + }, + { + "bbox": [ + 313, + 480, + 556, + 635 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 480, + 556, + 635 + ], + "type": "text", + "content": "-shaped curves, scaling the input activations while squashing the extreme values. Inspired by this insight, we propose an elementwise operation termed Dynamic Tanh (DyT), defined as: " + }, + { + "bbox": [ + 313, + 480, + 556, + 635 + ], + "type": "inline_equation", + "content": "\\mathrm{DyT}(\\boldsymbol{x}) = \\tanh(\\alpha \\boldsymbol{x})" + }, + { + "bbox": [ + 313, + 480, + 556, + 635 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 480, + 556, + 635 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 480, + 556, + 635 + ], + "type": "text", + "content": " is a learnable parameter. This operation aims to emulate the behavior of LN by learning an appropriate scaling factor through " + }, + { + "bbox": [ + 313, + 480, + 556, + 635 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 480, + 556, + 635 + ], + "type": "text", + "content": " and squashing extreme values via the bounded tanh function. Notably, unlike normalization layers, it achieves both effects without the need to compute activation statistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "Employing DyT is straightforward, as shown in Figure 1: we directly replace existing normalization layers with DyT in architectures such as vision and language Transformers. We empirically demonstrate that models with DyT can train stably and achieve high final performance across a wide range of settings. It often does not require tuning the hy" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14901" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 83, + 547, + 353 + ], + "blocks": [ + { + "bbox": [ + 243, + 71, + 386, + 82 + ], + "lines": [ + { + "bbox": [ + 243, + 71, + 386, + 82 + ], + "spans": [ + { + "bbox": [ + 243, + 71, + 386, + 82 + ], + "type": "text", + "content": "LN output (y axis) vs. LN input (x axis)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 83, + 547, + 353 + ], + "lines": [ + { + "bbox": [ + 53, + 83, + 547, + 353 + ], + "spans": [ + { + "bbox": [ + 53, + 83, + 547, + 353 + ], + "type": "image", + "image_path": "090957d29502bab3ea90da5d620cdba519e426f606475355bf84712d93e89b94.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 354, + 555, + 409 + ], + "lines": [ + { + "bbox": [ + 54, + 354, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 54, + 354, + 555, + 409 + ], + "type": "text", + "content": "Figure 2. Output vs. input of selected layer normalization (LN) layers in Vision Transformer (ViT) [20], wav2vec 2.0 (a Transformer model for speech) [7], and Diffusion Transformer (DiT) [64]. We sample a mini-batch of samples and plot the input / output values of four LN layers in each model. The outputs are before the affine transformation in LN. The " + }, + { + "bbox": [ + 54, + 354, + 555, + 409 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 54, + 354, + 555, + 409 + ], + "type": "text", + "content": "-shaped curves highly resemble that of a tanh function. The more linear shapes in earlier layers can also be captured by the center part of a tanh curve. This motivates us to propose Dynamic Tanh (DyT) as a replacement, with a learnable scalar " + }, + { + "bbox": [ + 54, + 354, + 555, + 409 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 354, + 555, + 409 + ], + "type": "text", + "content": " to account for different scales on the " + }, + { + "bbox": [ + 54, + 354, + 555, + 409 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 54, + 354, + 555, + 409 + ], + "type": "text", + "content": " axis." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 418, + 295, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 418, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 54, + 418, + 295, + 467 + ], + "type": "text", + "content": "perparameters on the original architecture. Our work challenges the notion that normalization layers are indispensable for training modern neural networks and provides empirical insights into the properties of normalization layers." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 475, + 254, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 254, + 490 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 254, + 490 + ], + "type": "text", + "content": "2. Background: Normalization Layers" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 496, + 296, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 296, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 296, + 556 + ], + "type": "text", + "content": "We begin by reviewing the normalization layers. Most normalization layers share a common formulation. Given an input " + }, + { + "bbox": [ + 55, + 496, + 296, + 556 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 496, + 296, + 556 + ], + "type": "text", + "content": " with shape " + }, + { + "bbox": [ + 55, + 496, + 296, + 556 + ], + "type": "inline_equation", + "content": "(B, T, C)" + }, + { + "bbox": [ + 55, + 496, + 296, + 556 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 496, + 296, + 556 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 496, + 296, + 556 + ], + "type": "text", + "content": " is the batch size, " + }, + { + "bbox": [ + 55, + 496, + 296, + 556 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 496, + 296, + 556 + ], + "type": "text", + "content": " is the number of tokens, and " + }, + { + "bbox": [ + 55, + 496, + 296, + 556 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 496, + 296, + 556 + ], + "type": "text", + "content": " is the embedding dimension per token, the output is generally computed as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 85, + 562, + 295, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 562, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 85, + 562, + 295, + 590 + ], + "type": "interline_equation", + "content": "\\operatorname {n o r m a l i z a t i o n} (\\boldsymbol {x}) = \\gamma * \\left(\\frac {\\boldsymbol {x} - \\boldsymbol {\\mu}}{\\sqrt {\\sigma^ {2} + \\epsilon}}\\right) + \\beta \\tag {1}", + "image_path": "3788c75b1913788bc4b23f631b4ab8ebd1a16d944c94648ca3c0023bb2573de3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": " is a small constant, and " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": " are learnable vector parameters of shape " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "(C,)" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": ". They are \"scaling\" and \"shifting\" affine parameters that allow the output to be in any range. The terms " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": " denote the mean and variance of the input. Different methods mainly differ in how these two statistics are computed. This results in " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": " having different dimensions, each with appropriate broadcasting needed in the tensor calculation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Batch normalization (BN) [41] is the first modern normalization layer, and it has been primarily used in Con" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 418, + 555, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 555, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 555, + 541 + ], + "type": "text", + "content": "vNet models [33, 76, 84]. Its introduction represents a major milestone in deep learning architecture designs. BN computes the mean and variance across both the batch and token dimensions, specifically: " + }, + { + "bbox": [ + 313, + 418, + 555, + 541 + ], + "type": "inline_equation", + "content": "\\mu_{k} = \\frac{1}{BT}\\sum_{i,j}x_{ijk}" + }, + { + "bbox": [ + 313, + 418, + 555, + 541 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 418, + 555, + 541 + ], + "type": "inline_equation", + "content": "\\sigma_k^2 = \\frac{1}{BT}\\sum_{i,j}(x_{ijk} - \\mu_k)^2" + }, + { + "bbox": [ + 313, + 418, + 555, + 541 + ], + "type": "text", + "content": ". Other normalization layers popular in ConvNets, such as group normalization [83] and instance normalization [81], were initially proposed for specialized tasks such as object detection and image stylization. They share the same overall formulation but differ in the axes and ranges over which the statistics are computed." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 544, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 556, + 713 + ], + "type": "text", + "content": "Layer normalization (LN) [5] and root mean square normalization (RMSNorm) [89] are the major two types of normalization layers used in Transformer architectures. LN computes these statistics independently for each token in each sample, where " + }, + { + "bbox": [ + 313, + 544, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mu_{ij} = \\frac{1}{C}\\sum_k x_{ijk}" + }, + { + "bbox": [ + 313, + 544, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 544, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\sigma_{ij}^2 = \\frac{1}{C}\\sum_k(x_{ijk} - \\mu_{ij})^2" + }, + { + "bbox": [ + 313, + 544, + 556, + 713 + ], + "type": "text", + "content": ". RMSNorm [89] simplifies LN by removing the mean-centering step and normalizing the input with " + }, + { + "bbox": [ + 313, + 544, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mu_{ij} = 0" + }, + { + "bbox": [ + 313, + 544, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 544, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\sigma_{ij}^2 = \\frac{1}{C}\\sum_k x_{ijk}^2" + }, + { + "bbox": [ + 313, + 544, + 556, + 713 + ], + "type": "text", + "content": ". Today, most modern neural networks use LN due to its simplicity and universality. Recently, RMSNorm has gained popularity, particularly in language models like T5 [66], LLaMA [21, 79, 80], Mistral [43], Qwen [8, 87], InternLM [13, 91] and DeepSeek [30, 49]. The Transformers we examine in this work all use LN, except that LLaMA uses RMSNorm." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14902" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 56, + 545, + 163 + ], + "blocks": [ + { + "bbox": [ + 52, + 56, + 545, + 163 + ], + "lines": [ + { + "bbox": [ + 52, + 56, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 52, + 56, + 545, + 163 + ], + "type": "image", + "image_path": "41a32d50a2a8c9a3a51d104555d6347f8be77b6e90e8e8694681af64cd8ac114.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 163, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 54, + 163, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 54, + 163, + 555, + 232 + ], + "type": "text", + "content": "Figure 3. Output vs. input of two LN layers, with tensor elements colored to indicate different channel and token dimensions. The input tensor has a shape of (samples, tokens, and channels), with elements visualized by assigning consistent colors to the same tokens (left two panels) and channels (right two panels). Left two panels: points representing the same token (same color) form straight lines across different channels, as LN operates linearly across channels for each token. Interestingly, when plotted collectively, these lines form a non-linear tanh-shaped curve. Right two panels: each channel's input spans different ranges on the " + }, + { + "bbox": [ + 54, + 163, + 555, + 232 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 54, + 163, + 555, + 232 + ], + "type": "text", + "content": "-axis, contributing distinct segments to the overall tanh-shaped curve. Certain channels (e.g., red, green, and pink) exhibit more extreme " + }, + { + "bbox": [ + 54, + 163, + 555, + 232 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 54, + 163, + 555, + 232 + ], + "type": "text", + "content": " values, which are squashed by LN." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 238, + 255, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 255, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 255, + 251 + ], + "type": "text", + "content": "3. What Do Normalization Layers Do?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 262, + 296, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 296, + 357 + ], + "type": "text", + "content": "Analysis setup. We first empirically study the behaviors of normalization layers in trained networks. For this analysis, we take a Vision Transformer model (ViT-B) [20] trained on ImageNet-1K [19], a wav2vec 2.0 Large Transformer model [7] trained on LibriSpeech [63], and a Diffusion Transformer (DiT-XL) [64] trained on ImageNet-1K. In all cases, LN is applied in every Transformer block and before the final linear projection." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 360, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 296, + 468 + ], + "type": "text", + "content": "For all three trained networks, we sample a mini-batch of samples and do a forward pass through the network. We then measure the input and output for the normalization layers, i.e., tensors immediately before and after the normalization operation, before the learnable affine transformation. Since LN preserves the dimensions of the input tensor, we can establish a one-to-one correspondence between input and output elements, allowing a direct visualization of their relationship. We plot the resulting mappings in Figure 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 472, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 533 + ], + "type": "text", + "content": "Tanh-like mappings with layer normalization. For all three models, in earlier LN layers (1st column of Figure 2), we find this input-output relationship to be mostly linear, resembling a straight line in an " + }, + { + "bbox": [ + 55, + 472, + 296, + 533 + ], + "type": "inline_equation", + "content": "x - y" + }, + { + "bbox": [ + 55, + 472, + 296, + 533 + ], + "type": "text", + "content": " plot. However, deeper LN layers are where we make more intriguing observations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 533, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 296, + 664 + ], + "type": "text", + "content": "A striking observation from these deeper layers is that most of these curves' shapes resemble full or partial " + }, + { + "bbox": [ + 55, + 533, + 296, + 664 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 533, + 296, + 664 + ], + "type": "text", + "content": "-shaped curves represented by a tanh function. One might expect LN layers to linearly transform the input tensor, as subtracting the mean and dividing by standard deviation are linear operations. LN normalizes in a per-token manner, only linearly transforming each token's activations. As tokens have different mean and standard deviation values, the linearity does not hold collectively on all activations of the input tensor. Nonetheless, it is surprising that the non-linear transformation is highly similar to a scaled tanh function." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "For such an " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "-shaped curve, we note that the central part, represented by points with " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " values close to zero, is mainly in a linear shape. Most points " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\sim 99\\%)" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " fall in this linear range. However, there are many points that clearly fall out" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 239, + 555, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 555, + 346 + ], + "type": "text", + "content": "of this range, which are considered to have \"extreme\" values, e.g., those with " + }, + { + "bbox": [ + 313, + 239, + 555, + 346 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 239, + 555, + 346 + ], + "type": "text", + "content": " larger than 50 or smaller than -50 in the ViT model. Normalization layers' main effect for these values is to squash them into less extreme values, more in line with the majority of points. This is where normalization layers could not approximated by a simple affine transformation layer. We hypothesize this non-linear and disproportional squashing effect on extreme values is what makes normalization layers important and indispensable." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 348, + 555, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 555, + 421 + ], + "type": "text", + "content": "Recent findings by Ni et al. [62] similarly highlight the strong non-linearities introduced by LN layers, demonstrating how the non-linearity enhances a model's representational capacity. Moreover, this squashing behavior mirrors the saturation properties of biological neurons for large inputs, a phenomenon observed about a century ago [1-3]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 425, + 556, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 556, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 556, + 533 + ], + "type": "text", + "content": "Normalization by tokens and channels. How does an LN layer perform a linear transformation for each token but also squash the extreme values in such a non-linear fashion? To understand this, we visualize the points grouped by tokens and channels, respectively. This is plotted in Figure 3 by taking the second and third subplots for ViT from Figure 2, but with a sampled subset of points for more clarity. When we select the channels to plot, we make sure to include the channels with extreme values." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": "On the left two panels of Figure 3, we visualize each token's activations using the same color. We observe that all points from any single token do form a straight line. However, since each token has a different variance, the slopes are different. Tokens with smaller input " + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": " ranges tend to have smaller variance, and the normalization layer will divide their activations using a smaller standard deviation, hence producing a larger slope in the straight line. Collectively, they form an " + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": "-shaped curve that resembles a tanh function. In the two panels on the right, we color each channel's activations using the same color. We find that different channels tend to have drastically different input ranges, with only a few channels (e.g., red, green, and pink) exhibiting large extreme values. These are the channels that get squashed the most by the normalization layer." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14903" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 180, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 180, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 180, + 85 + ], + "type": "text", + "content": "4. Dynamic Tanh (DyT)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 295, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 295, + 139 + ], + "type": "text", + "content": "Inspired by the similarity between shapes of normalization layers and a scaled tanh function, we propose Dynamic Tanh (DyT) as a replacement for normalization layers. Given input tensor " + }, + { + "bbox": [ + 55, + 91, + 295, + 139 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 91, + 295, + 139 + ], + "type": "text", + "content": ", a DyT layer is defined as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 111, + 145, + 295, + 158 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 145, + 295, + 158 + ], + "spans": [ + { + "bbox": [ + 111, + 145, + 295, + 158 + ], + "type": "interline_equation", + "content": "\\mathrm {D y T} (\\boldsymbol {x}) = \\boldsymbol {\\gamma} * \\tanh (\\alpha \\boldsymbol {x}) + \\boldsymbol {\\beta} \\tag {2}", + "image_path": "e2a018d7b4ee13127c77b088dea0be973cc38a7976724773282760de9ba66a62.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 160, + 296, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 160, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 160, + 296, + 281 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 160, + 296, + 281 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 160, + 296, + 281 + ], + "type": "text", + "content": " is a learnable scalar parameter that allows scaling the input differently based on its range, accounting for varying " + }, + { + "bbox": [ + 55, + 160, + 296, + 281 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 160, + 296, + 281 + ], + "type": "text", + "content": " scales (Figure 2). This is also why we name the whole operation \"Dynamic\" Tanh. " + }, + { + "bbox": [ + 55, + 160, + 296, + 281 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 160, + 296, + 281 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 160, + 296, + 281 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 160, + 296, + 281 + ], + "type": "text", + "content": " are learnable, per-channel vector parameters, the same as those used in all normalization layers—they allow the output to scale back to any scales. This is sometimes considered a separate affine layer; for our purposes, we consider them to be part of the DyT layer, just like how normalization layers also include them. See Algorithm 1 for Pytorch-like pseudocode of DyT." + } + ] + } + ], + "index": 3 + }, + { + "type": "code", + "bbox": [ + 67, + 301, + 274, + 401 + ], + "blocks": [ + { + "bbox": [ + 67, + 287, + 229, + 299 + ], + "lines": [ + { + "bbox": [ + 67, + 287, + 229, + 299 + ], + "spans": [ + { + "bbox": [ + 67, + 287, + 229, + 299 + ], + "type": "text", + "content": "Algorithm 1 Pseudocode of DyT layer." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 67, + 301, + 274, + 401 + ], + "lines": [ + { + "bbox": [ + 67, + 301, + 274, + 401 + ], + "spans": [ + { + "bbox": [ + 67, + 301, + 274, + 401 + ], + "type": "text", + "content": "input x has the shape of [B, T, C] \n# B: batch size, T: tokens, C: dimension \nclass DyT(Module): \n def __init__(self, C, init\\_alpha): \n super().__init() \n self." + }, + { + "bbox": [ + 67, + 301, + 274, + 401 + ], + "type": "inline_equation", + "content": "\\alpha =" + }, + { + "bbox": [ + 67, + 301, + 274, + 401 + ], + "type": "text", + "content": " Parameter(ones(1) " + }, + { + "bbox": [ + 67, + 301, + 274, + 401 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 67, + 301, + 274, + 401 + ], + "type": "text", + "content": " init\\_alpha) \n self." + }, + { + "bbox": [ + 67, + 301, + 274, + 401 + ], + "type": "inline_equation", + "content": "\\gamma =" + }, + { + "bbox": [ + 67, + 301, + 274, + 401 + ], + "type": "text", + "content": " Parameter(ones(C)) \n self." + }, + { + "bbox": [ + 67, + 301, + 274, + 401 + ], + "type": "inline_equation", + "content": "\\beta =" + }, + { + "bbox": [ + 67, + 301, + 274, + 401 + ], + "type": "text", + "content": " Parameter(zeros(C)) \ndef forward(self, x): \n x = tanh(self.alpha " + }, + { + "bbox": [ + 67, + 301, + 274, + 401 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 67, + 301, + 274, + 401 + ], + "type": "text", + "content": " x) \n return self." + }, + { + "bbox": [ + 67, + 301, + 274, + 401 + ], + "type": "inline_equation", + "content": "\\gamma =" + }, + { + "bbox": [ + 67, + 301, + 274, + 401 + ], + "type": "text", + "content": " x + self." + }, + { + "bbox": [ + 67, + 301, + 274, + 401 + ], + "type": "inline_equation", + "content": "\\beta" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "code_body" + } + ], + "index": 5, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 407, + 296, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 539 + ], + "type": "text", + "content": "Integrating DyT layers into an existing architecture is straightforward: one DyT layer replaces one normalization layer (see Figure 1). This applies to normalization layers within attention blocks, FFN blocks, and the final normalization layer. Although DyT may look like or be considered an activation function, this study only uses it to replace normalization layers without altering parts of the activation functions in the original architectures, such as GELU or ReLU. Other parts of the network remain intact. We also observe that there is minimal need to tune the hyperparameters for the original architectures for DyT to perform well." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 542, + 296, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 296, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 296, + 626 + ], + "type": "text", + "content": "On scaling parameters. We always simply initialize " + }, + { + "bbox": [ + 55, + 542, + 296, + 626 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 542, + 296, + 626 + ], + "type": "text", + "content": " to an all-one vector and " + }, + { + "bbox": [ + 55, + 542, + 296, + 626 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 542, + 296, + 626 + ], + "type": "text", + "content": " to an all-zero vector following normalization layers. For the scalar parameter " + }, + { + "bbox": [ + 55, + 542, + 296, + 626 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 542, + 296, + 626 + ], + "type": "text", + "content": ", a default initialization of 0.5 is generally sufficient, except for LLM training. A detailed analysis of " + }, + { + "bbox": [ + 55, + 542, + 296, + 626 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 542, + 296, + 626 + ], + "type": "text", + "content": " initialization is provided in Section 7. Unless explicitly stated otherwise, " + }, + { + "bbox": [ + 55, + 542, + 296, + 626 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 542, + 296, + 626 + ], + "type": "text", + "content": " is initialized to 0.5 in our subsequent experiments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "Remarks. DyT is not a new type of normalization layer, as it operates on each input element from a tensor independently during a forward pass without computing statistics or other types of aggregations. It does, however, preserve the effect of normalization layers in squashing the extreme values in a non-linear fashion while almost linearly transforming the very central parts of the input." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 71, + 395, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 395, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 395, + 85 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 92, + 555, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 92, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 92, + 555, + 236 + ], + "type": "text", + "content": "To demonstrate the effectiveness of DyT, we experiment with Transformers and a few other modern architectures across a diverse range of tasks and domains. In each experiment, we replace the LN or RMSNorm in the original architectures with DyT layers and follow the official open-source protocols to train and test both versions of the models. Detailed instructions for reproducing our results are provided in Appendix A. Notably, to highlight the simplicity of adapting DyT, we use hyperparameters identical to those used by the normalized counterparts. For completeness, additional results regarding tuning of learning rates and initial values of " + }, + { + "bbox": [ + 313, + 92, + 555, + 236 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 92, + 555, + 236 + ], + "type": "text", + "content": " are provided in Appendix F." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 241, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 555, + 361 + ], + "type": "text", + "content": "Supervised learning in vision. We train Vision Transformer (ViT) [20] and ConvNeXt [50] of Base\" and Large\" sizes on the ImageNet-1K classification task [19]. These models are selected due to their popularity and distinct operations: attention in ViT and convolution in ConvNeXt. Table 1 reports the top-1 classification accuracies. DyT performs slightly better than LN across both architectures and model sizes. We further plot the training loss for ViT-B and ConvNeXt-B in Figure 4. The curves show that the convergence behaviors of DyT and LN-based models are aligned." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 352, + 362, + 517, + 429 + ], + "blocks": [ + { + "bbox": [ + 352, + 362, + 517, + 429 + ], + "lines": [ + { + "bbox": [ + 352, + 362, + 517, + 429 + ], + "spans": [ + { + "bbox": [ + 352, + 362, + 517, + 429 + ], + "type": "table", + "html": "
modelLNDyTchange
ViT-B82.3%82.5%↑0.2%
ViT-L83.1%83.6%↑0.5%
ConvNeXt-B83.7%83.7%-
ConvNeXt-L84.3%84.4%↑0.1%
", + "image_path": "2ccda9a35b38199e81b1cc3d55d34e8871441117f244b50de6ea99c35799eef7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 432, + 555, + 464 + ], + "lines": [ + { + "bbox": [ + 313, + 432, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 555, + 464 + ], + "type": "text", + "content": "Table 1. Supervised classification accuracy on ImageNet-1K. DyT achieves better or similar performance than LN across both architectures and model sizes." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 318, + 468, + 539, + 569 + ], + "blocks": [ + { + "bbox": [ + 318, + 468, + 539, + 569 + ], + "lines": [ + { + "bbox": [ + 318, + 468, + 539, + 569 + ], + "spans": [ + { + "bbox": [ + 318, + 468, + 539, + 569 + ], + "type": "image", + "image_path": "b7baae6506173d2c014485304a72c5faa58779d1c2990cd0bf8966405b7f33e5.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 318, + 571, + 539, + 673 + ], + "blocks": [ + { + "bbox": [ + 318, + 571, + 539, + 673 + ], + "lines": [ + { + "bbox": [ + 318, + 571, + 539, + 673 + ], + "spans": [ + { + "bbox": [ + 318, + 571, + 539, + 673 + ], + "type": "image", + "image_path": "9d2ae29f8036826d3469fa1d6544060abde19d956d0b8893473fb7f88ce49646.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 675, + 555, + 709 + ], + "lines": [ + { + "bbox": [ + 313, + 675, + 555, + 709 + ], + "spans": [ + { + "bbox": [ + 313, + 675, + 555, + 709 + ], + "type": "text", + "content": "Figure 4. Training loss curves for ViT-B and ConvNeXt-B. The curves for both models exhibit similar patterns between LN and DyT, suggesting that they may share similar learning dynamics." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14904" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 44, + 285, + 148 + ], + "blocks": [ + { + "bbox": [ + 55, + 44, + 285, + 148 + ], + "lines": [ + { + "bbox": [ + 55, + 44, + 285, + 148 + ], + "spans": [ + { + "bbox": [ + 55, + 44, + 285, + 148 + ], + "type": "image", + "image_path": "bedeb54687c14d9c30c54a7f5f5e4f234f7688487352d02ccdd79750a8c92102.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 296, + 44, + 529, + 148 + ], + "blocks": [ + { + "bbox": [ + 296, + 44, + 529, + 148 + ], + "lines": [ + { + "bbox": [ + 296, + 44, + 529, + 148 + ], + "spans": [ + { + "bbox": [ + 296, + 44, + 529, + 148 + ], + "type": "image", + "image_path": "fd7954f2a83eb567ee516aa747bbd484bc01286a8ebf046e01d2af713bcfd322.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 54, + 149, + 282, + 251 + ], + "blocks": [ + { + "bbox": [ + 54, + 149, + 282, + 251 + ], + "lines": [ + { + "bbox": [ + 54, + 149, + 282, + 251 + ], + "spans": [ + { + "bbox": [ + 54, + 149, + 282, + 251 + ], + "type": "image", + "image_path": "c85074c4872705decb2ad04005aa7d38d0cbbe528a7411e0be5489e82f7a7e41.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 254, + 527, + 265 + ], + "lines": [ + { + "bbox": [ + 80, + 254, + 527, + 265 + ], + "spans": [ + { + "bbox": [ + 80, + 254, + 527, + 265 + ], + "type": "text", + "content": "Figure 5. LLaMA pretraining loss. The loss curves of DyT and RMSNorm models are closely aligned across model sizes." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 296, + 148, + 526, + 251 + ], + "blocks": [ + { + "bbox": [ + 296, + 148, + 526, + 251 + ], + "lines": [ + { + "bbox": [ + 296, + 148, + 526, + 251 + ], + "spans": [ + { + "bbox": [ + 296, + 148, + 526, + 251 + ], + "type": "image", + "image_path": "31f8b8dc47ee5d769dbfb8e71e40d4a7e0888ac2d5e9f80ed9ed4e00230c00e4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 274, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 274, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 54, + 274, + 296, + 418 + ], + "type": "text", + "content": "Self-supervised learning in vision. We benchmark with two popular visual self-supervised learning methods: masked autoencoders (MAE) [34] and DINO [14]. Both by default use Vision Transformers as backbones, but have different training objectives: MAE is trained with a reconstruction loss, and DINO uses a joint-embedding loss [47]. Following the standard self-supervised learning protocol, we first pretrain models on ImageNet-1K without using labels and then test the pretrained models by attaching a classification layer and fine-tuning them with labels. The fine-tuning results are presented in Table 2. DyT consistently performs on par with LN in self-supervised learning tasks." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 69, + 419, + 282, + 487 + ], + "blocks": [ + { + "bbox": [ + 69, + 419, + 282, + 487 + ], + "lines": [ + { + "bbox": [ + 69, + 419, + 282, + 487 + ], + "spans": [ + { + "bbox": [ + 69, + 419, + 282, + 487 + ], + "type": "table", + "html": "
modelLNDyTchange
MAE ViT-B83.2%83.2%-
MAE ViT-L85.5%85.4%↓0.1%
DINO ViT-B (patch size 16)83.2%83.4%↑0.2%
DINO ViT-B (patch size 8)84.1%84.5%↑0.4%
", + "image_path": "1d458c9053b3e93e995d37ebd703da2ead850c8cc302d23fc83cf188ec6e922c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 517, + 296, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 296, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 296, + 638 + ], + "type": "text", + "content": "Diffusion models. We train three Diffusion Transformer (DiT) models [64] of sizes B, L and XL on ImageNet-1K [19]. The patch size is 4, 4, and 2, respectively. Note that in DiT, the LN layers' affine parameters are used for class conditioning in DiT, and we keep them that way in our DyT experiments, only replacing the normalizing transformation with the " + }, + { + "bbox": [ + 55, + 517, + 296, + 638 + ], + "type": "inline_equation", + "content": "\\tanh (\\alpha x)" + }, + { + "bbox": [ + 55, + 517, + 296, + 638 + ], + "type": "text", + "content": " function. After training, we evaluate the Fréchet Inception Distance (FID) scores using the standard ImageNet \"reference batch\", as presented in Table 3. DyT achieves comparable or improved FID over LN." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 108, + 639, + 243, + 695 + ], + "blocks": [ + { + "bbox": [ + 55, + 489, + 295, + 511 + ], + "lines": [ + { + "bbox": [ + 55, + 489, + 295, + 511 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 295, + 511 + ], + "type": "text", + "content": "Table 2. Self-supervised learning accuracy on ImageNet-1K. DyT performs on par with LN across different methods and sizes." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 108, + 639, + 243, + 695 + ], + "lines": [ + { + "bbox": [ + 108, + 639, + 243, + 695 + ], + "spans": [ + { + "bbox": [ + 108, + 639, + 243, + 695 + ], + "type": "table", + "html": "
modelLNDyTchange
DiT-B64.963.9↓1.0
DiT-L45.945.7↓0.2
DiT-XL19.920.8↑0.9
", + "image_path": "313da71622988da81f65608e994fc86d2082a51b85d73daae6bcb48d63e2cb50.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 274, + 555, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 555, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 555, + 454 + ], + "type": "text", + "content": "Large Language Models. We pretrain LLaMA 7B, 13B, 34B, and 70B models [21, 79, 80] to assess DyT performance relative to RMSNorm [89], the default normalization layer used in LLaMA. The models are trained on The Pile dataset [25] with 200B tokens, following the original recipe outlined in LLaMA [80]. On LLaMA with DyT, we add a learnable scalar parameter after the initial embedding layer, and adjust the initial value of " + }, + { + "bbox": [ + 313, + 274, + 555, + 454 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 274, + 555, + 454 + ], + "type": "text", + "content": ", as detailed in Section 7. We report the loss value after training and also follow OpenLLaMA [26] to benchmark the models on 15 zero-shot tasks from lm-eval [24]. As shown in Table 4, DyT performs on par with RMSNorm across all four model sizes. Figure 5 illustrates the loss curves, demonstrating similar trends across all model sizes, with training losses closely aligned throughout training." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 319, + 456, + 549, + 524 + ], + "blocks": [ + { + "bbox": [ + 55, + 697, + 295, + 719 + ], + "lines": [ + { + "bbox": [ + 55, + 697, + 295, + 719 + ], + "spans": [ + { + "bbox": [ + 55, + 697, + 295, + 719 + ], + "type": "text", + "content": "Table 3. Image generation quality (FID) on ImageNet. DyT achieves comparable FID scores to LN across various model sizes." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 456, + 549, + 524 + ], + "lines": [ + { + "bbox": [ + 319, + 456, + 549, + 524 + ], + "spans": [ + { + "bbox": [ + 319, + 456, + 549, + 524 + ], + "type": "table", + "html": "
score / lossRMSNormDyTchange
LLaMA 7B0.513 / 1.590.513 / 1.60- / ↑0.01
LLaMA 13B0.529 / 1.530.529 / 1.54- / ↑0.01
LLaMA 34B0.536 / 1.500.536 / 1.50- / -
LLaMA 70B0.549 / 1.450.549 / 1.45- / -
", + "image_path": "cee3d14dd485dff5d1d08253a9d9540f2625be5426bcc24e0183891e070ef609.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 568, + 554, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 554, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 554, + 641 + ], + "type": "text", + "content": "Self-supervised learning in speech. We pretrain two wav2vec 2.0 Transformer models [7] on the LibriSpeech dataset [63]. We retain the first group normalization layer, as it functions as data normalization to handle the unnormalized input data. Table 5 reports the final validation loss. We observe that DyT performs comparably to LN." + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 349, + 644, + 520, + 690 + ], + "blocks": [ + { + "bbox": [ + 313, + 526, + 554, + 560 + ], + "lines": [ + { + "bbox": [ + 313, + 526, + 554, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 554, + 560 + ], + "type": "text", + "content": "Table 4. Language models' training loss and average performance with 15 zero-shot 1m-eval tasks. DyT achieves a comparable zero-shot performance and training loss to RMSNorm." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 349, + 644, + 520, + 690 + ], + "lines": [ + { + "bbox": [ + 349, + 644, + 520, + 690 + ], + "spans": [ + { + "bbox": [ + 349, + 644, + 520, + 690 + ], + "type": "table", + "html": "
modelLNDyTchange
wav2vec 2.0 Base1.951.95-
wav2vec 2.0 Large1.921.91↓0.01
", + "image_path": "df7ae9ed77f60b5b7df0826a6c9830579442c4131de923d4444a2977de1a60a6.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 693, + 553, + 715 + ], + "lines": [ + { + "bbox": [ + 313, + 693, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 553, + 715 + ], + "type": "text", + "content": "Table 5. Speech pretraining validation loss on LibriSpeech. DyT performs comparably to LN for both wav2vec 2.0 models." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14905" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 41, + 246, + 133 + ], + "blocks": [ + { + "bbox": [ + 63, + 41, + 246, + 133 + ], + "lines": [ + { + "bbox": [ + 63, + 41, + 246, + 133 + ], + "spans": [ + { + "bbox": [ + 63, + 41, + 246, + 133 + ], + "type": "image", + "image_path": "3073ff5c508db6f14beaea3b3b12d9fad7c867dad3718995952afcf23c4abac5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 138, + 253, + 196 + ], + "lines": [ + { + "bbox": [ + 55, + 138, + 253, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 138, + 253, + 196 + ], + "type": "text", + "content": "Figure 6. Curves of three squashing functions: tanh, hardtanh, and sigmoid. All three functions squash inputs into a bounded range, but tanh achieves the best performance when used in DyT layers. We suspect it is due to its smoothness and zero-centered properties." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 265, + 40, + 400, + 137 + ], + "blocks": [ + { + "bbox": [ + 265, + 40, + 400, + 137 + ], + "lines": [ + { + "bbox": [ + 265, + 40, + 400, + 137 + ], + "spans": [ + { + "bbox": [ + 265, + 40, + 400, + 137 + ], + "type": "image", + "image_path": "600e1a9390e4d16dd57e59a540132dde226199e7d1f1ab5b1895852b2bbbed28.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 138, + 553, + 196 + ], + "lines": [ + { + "bbox": [ + 260, + 138, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 260, + 138, + 553, + 196 + ], + "type": "text", + "content": "Figure 7. Left: For two selected DyT layers from the ViT-B model, we track " + }, + { + "bbox": [ + 260, + 138, + 553, + 196 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 260, + 138, + 553, + 196 + ], + "type": "text", + "content": " and the inverse of the standard deviation (1/std) of activations at the end of each epoch, observing that they evolve together during training. Right: We plot the final " + }, + { + "bbox": [ + 260, + 138, + 553, + 196 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 260, + 138, + 553, + 196 + ], + "type": "text", + "content": " values of two models, ViT-B and ConvNeXt-B, against the 1/std of the input activations, demonstrating a strong correlation between the two values." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 411, + 41, + 547, + 137 + ], + "blocks": [ + { + "bbox": [ + 411, + 41, + 547, + 137 + ], + "lines": [ + { + "bbox": [ + 411, + 41, + 547, + 137 + ], + "spans": [ + { + "bbox": [ + 411, + 41, + 547, + 137 + ], + "type": "image", + "image_path": "da316b4b6400f407422da48a2d297effac125bbd60a4442150c2968c50a470e1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 209, + 296, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 296, + 281 + ], + "type": "text", + "content": "DNA sequence modeling. On the long-range DNA sequence modeling task, we pretrain the HyenaDNA model [60] and the Caduceus model [69]. The pretraining uses human reference genome data from [27], and the evaluation is on GenomicBenchmarks [28]. Results are presented in Table 6. DyT maintains performance comparable to LN." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 88, + 282, + 264, + 328 + ], + "blocks": [ + { + "bbox": [ + 88, + 282, + 264, + 328 + ], + "lines": [ + { + "bbox": [ + 88, + 282, + 264, + 328 + ], + "spans": [ + { + "bbox": [ + 88, + 282, + 264, + 328 + ], + "type": "table", + "html": "
modelLNDyTchange
HyenaDNA [60]85.2%85.2%-
Caduceus [69]86.9%86.9%-
", + "image_path": "c173f77b64ba17b7908392ca13e606b1294c7c4fb4d2bc7a2acb4bca483b8644.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 371, + 115, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 115, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 115, + 385 + ], + "type": "text", + "content": "6. Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 392, + 296, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 296, + 415 + ], + "type": "text", + "content": "We conduct two studies examining the roles of the tanh function and the learnable scale " + }, + { + "bbox": [ + 55, + 392, + 296, + 415 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 392, + 296, + 415 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 423, + 189, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 189, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 189, + 436 + ], + "type": "text", + "content": "6.1. Ablations of tanh and " + }, + { + "bbox": [ + 55, + 423, + 189, + 436 + ], + "type": "inline_equation", + "content": "\\alpha" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 444, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 296, + 479 + ], + "type": "text", + "content": "To further investigate the role of tanh and " + }, + { + "bbox": [ + 55, + 444, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 444, + 296, + 479 + ], + "type": "text", + "content": " in DyT, we conduct experiments to evaluate the model's performance when these components are altered or removed." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 483, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 579 + ], + "type": "text", + "content": "Replacing and removing tanh. We replace tanh in DyT layers with alternative squashing functions, specifically hardtanh and sigmoid (Figure 6), while keeping the learnable scalar " + }, + { + "bbox": [ + 55, + 483, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 483, + 296, + 579 + ], + "type": "text", + "content": " intact. Furthermore, we assess the impact of completely removing tanh by replacing it with the identity function while still retaining " + }, + { + "bbox": [ + 55, + 483, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 483, + 296, + 579 + ], + "type": "text", + "content": ". As shown in Table 7, the squashing function is essential for stable training. Using the identity function leads to unstable training and divergence," + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 66, + 584, + 286, + 630 + ], + "blocks": [ + { + "bbox": [ + 55, + 330, + 296, + 364 + ], + "lines": [ + { + "bbox": [ + 55, + 330, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 296, + 364 + ], + "type": "text", + "content": "Table 6. DNA classification accuracy on GenomicBenchmarks, averaged over each dataset in GenomicBenchmarks. DyT achieves comparable performance to LN." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 584, + 286, + 630 + ], + "lines": [ + { + "bbox": [ + 66, + 584, + 286, + 630 + ], + "spans": [ + { + "bbox": [ + 66, + 584, + 286, + 630 + ], + "type": "table", + "html": "
modelidentitytanhhardtanhsigmoid
ViT-S58.5% → failed80.3%79.9%79.6%
ViT-B61.0% → failed82.5%82.2%81.6%
", + "image_path": "e365fac92c6222f805f2cda091f40fda49cc91cdac06a3c1462a8de3407bcd9a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 209, + 554, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 554, + 246 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 554, + 246 + ], + "type": "text", + "content": "whereas squashing functions enable stable training. Among the squashing functions, tanh performs the best. This is possibly due to its smoothness and zero-centered properties." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 255, + 554, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 554, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 554, + 316 + ], + "type": "text", + "content": "Removing " + }, + { + "bbox": [ + 313, + 255, + 554, + 316 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 255, + 554, + 316 + ], + "type": "text", + "content": ". Next, we evaluate the impact of removing the learnable " + }, + { + "bbox": [ + 313, + 255, + 554, + 316 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 255, + 554, + 316 + ], + "type": "text", + "content": " while retaining the squashing functions. As shown in Table 8, removing " + }, + { + "bbox": [ + 313, + 255, + 554, + 316 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 255, + 554, + 316 + ], + "type": "text", + "content": " results in performance degradation across all squashing functions, highlighting the critical role of " + }, + { + "bbox": [ + 313, + 255, + 554, + 316 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 255, + 554, + 316 + ], + "type": "text", + "content": " in overall model performance." + } + ] + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 351, + 317, + 518, + 362 + ], + "blocks": [ + { + "bbox": [ + 55, + 632, + 296, + 710 + ], + "lines": [ + { + "bbox": [ + 55, + 632, + 296, + 710 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 296, + 710 + ], + "type": "text", + "content": "Table 7. ImageNet-1K classification accuracy with different squashing functions. All experiments follow the same training recipe as the original LN-based models. Squashing functions play a crucial role in preventing divergence, with tanh achieving the highest performance among the three functions. “" + }, + { + "bbox": [ + 55, + 632, + 296, + 710 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 632, + 296, + 710 + ], + "type": "text", + "content": " failed” indicates that training diverged after some progress, with the preceding number representing the accuracy reached before divergence." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 351, + 317, + 518, + 362 + ], + "lines": [ + { + "bbox": [ + 351, + 317, + 518, + 362 + ], + "spans": [ + { + "bbox": [ + 351, + 317, + 518, + 362 + ], + "type": "table", + "html": "
modeltanhhardtanhsigmoid
without α81.1%80.7%80.7%
with α82.5%82.2%81.6%
", + "image_path": "97aaa98aa13e7e2fc8706ed4fc49c7aa9cc88e828a8335294b47f7f23b112c6f.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 365, + 554, + 398 + ], + "lines": [ + { + "bbox": [ + 313, + 365, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 554, + 398 + ], + "type": "text", + "content": "Table 8. ImageNet-1K classification accuracy with ViT-B. All experiments follow the same training recipe as the original LN-based models. " + }, + { + "bbox": [ + 313, + 365, + 554, + 398 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 365, + 554, + 398 + ], + "type": "text", + "content": " is essential for enhancing model performance." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 416, + 389, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 416, + 389, + 428 + ], + "spans": [ + { + "bbox": [ + 314, + 416, + 389, + 428 + ], + "type": "text", + "content": "6.2. Values of " + }, + { + "bbox": [ + 314, + 416, + 389, + 428 + ], + "type": "inline_equation", + "content": "\\alpha" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 444, + 554, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 554, + 529 + ], + "type": "text", + "content": "During training. Our analysis reveals that the " + }, + { + "bbox": [ + 313, + 444, + 554, + 529 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 444, + 554, + 529 + ], + "type": "text", + "content": " closely tracks the " + }, + { + "bbox": [ + 313, + 444, + 554, + 529 + ], + "type": "inline_equation", + "content": "1 / \\mathrm{std}" + }, + { + "bbox": [ + 313, + 444, + 554, + 529 + ], + "type": "text", + "content": " of activations throughout training. As illustrated in the left panel of Figure 7, " + }, + { + "bbox": [ + 313, + 444, + 554, + 529 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 444, + 554, + 529 + ], + "type": "text", + "content": " first decrease and then increase during training, but always fluctuate consistently with the standard deviation of activations. This supports the important role of " + }, + { + "bbox": [ + 313, + 444, + 554, + 529 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 444, + 554, + 529 + ], + "type": "text", + "content": " in maintaining activations within a suitable range, which leads to stable and effective training." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 539, + 555, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 555, + 647 + ], + "type": "text", + "content": "After training. Our further analysis of the final values of " + }, + { + "bbox": [ + 313, + 539, + 555, + 647 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 539, + 555, + 647 + ], + "type": "text", + "content": " in trained networks reveals a strong correlation with the 1/std of the input activations. As shown on the right panel of Figure 7, higher 1/std values generally correspond to larger " + }, + { + "bbox": [ + 313, + 539, + 555, + 647 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 539, + 555, + 647 + ], + "type": "text", + "content": " values, and vice versa. Additionally, we observe that deeper layers tend to have activations with larger standard deviations. This trend aligns with characteristics of deep residual networks, as shown in Brock et al. [11] for ConvNets, and Sun et al. [74] for Transformers." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Both analyses suggest that " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": " acts as a normalization mechanism, learning values close to " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "1 / \\mathrm{std}" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": " of the input activations. Unlike LN, which normalizes per token, " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": " normalizes the entire input collectively and cannot suppress extreme values non-linearly." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14906" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 44, + 225, + 141 + ], + "blocks": [ + { + "bbox": [ + 59, + 44, + 225, + 141 + ], + "lines": [ + { + "bbox": [ + 59, + 44, + 225, + 141 + ], + "spans": [ + { + "bbox": [ + 59, + 44, + 225, + 141 + ], + "type": "image", + "image_path": "796e8ffa5539831c9248f034cd69ad5b6f6382596f3fc0815caee396a7edf903.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 225, + 44, + 388, + 141 + ], + "blocks": [ + { + "bbox": [ + 225, + 44, + 388, + 141 + ], + "lines": [ + { + "bbox": [ + 225, + 44, + 388, + 141 + ], + "spans": [ + { + "bbox": [ + 225, + 44, + 388, + 141 + ], + "type": "image", + "image_path": "4a715c71c0d2236260abcd4575cf0ca5b2eed1c716f991ce94d1ffa7e843645e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 388, + 44, + 551, + 141 + ], + "blocks": [ + { + "bbox": [ + 388, + 44, + 551, + 141 + ], + "lines": [ + { + "bbox": [ + 388, + 44, + 551, + 141 + ], + "spans": [ + { + "bbox": [ + 388, + 44, + 551, + 141 + ], + "type": "image", + "image_path": "f188c861c933703a664b47818539b4dbf8a25f2dad051cbc827bcc9fc106e59d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 142, + 225, + 240 + ], + "blocks": [ + { + "bbox": [ + 59, + 142, + 225, + 240 + ], + "lines": [ + { + "bbox": [ + 59, + 142, + 225, + 240 + ], + "spans": [ + { + "bbox": [ + 59, + 142, + 225, + 240 + ], + "type": "image", + "image_path": "3373dd2e375fe35c87f2a2861451b4945b27c7a9cc88527e40deb375db3925a1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 243, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 55, + 243, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 555, + 277 + ], + "type": "text", + "content": "Figure 8. Performance of different tasks across different " + }, + { + "bbox": [ + 55, + 243, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 55, + 243, + 555, + 277 + ], + "type": "text", + "content": " values. We benchmark the performance of all non-LLM tasks used in Section 5 with different initial values of " + }, + { + "bbox": [ + 55, + 243, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 243, + 555, + 277 + ], + "type": "text", + "content": ". Performance remains stable across a wide range of " + }, + { + "bbox": [ + 55, + 243, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 55, + 243, + 555, + 277 + ], + "type": "text", + "content": " values. The only exception is that supervised ViT-L models (top right panel) will diverge for " + }, + { + "bbox": [ + 55, + 243, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 55, + 243, + 555, + 277 + ], + "type": "text", + "content": " values larger than 0.6." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 225, + 142, + 388, + 239 + ], + "blocks": [ + { + "bbox": [ + 225, + 142, + 388, + 239 + ], + "lines": [ + { + "bbox": [ + 225, + 142, + 388, + 239 + ], + "spans": [ + { + "bbox": [ + 225, + 142, + 388, + 239 + ], + "type": "image", + "image_path": "9ec84880ebfc85127aa97cf2522f5bae6a416a96c5087ab102710cd4b97f4e4d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 388, + 142, + 551, + 239 + ], + "blocks": [ + { + "bbox": [ + 388, + 142, + 551, + 239 + ], + "lines": [ + { + "bbox": [ + 388, + 142, + 551, + 239 + ], + "spans": [ + { + "bbox": [ + 388, + 142, + 551, + 239 + ], + "type": "image", + "image_path": "8c6b92be8d50717e7d543923b4f9f103605016db23ca45b9d3a381c324fc62ba.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 284, + 160, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 160, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 160, + 297 + ], + "type": "text", + "content": "7. Initialization of " + }, + { + "bbox": [ + 55, + 284, + 160, + 297 + ], + "type": "inline_equation", + "content": "\\alpha" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "type": "text", + "content": "We find that tuning the initialization of " + }, + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "type": "text", + "content": " (denoted " + }, + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "type": "text", + "content": ") rarely leads to significant performance improvements. The only exception is LLM training, where careful tuning of " + }, + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "type": "text", + "content": " yields noticeable performance gains. In this section, we detail our findings on the impact of " + }, + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 305, + 296, + 365 + ], + "type": "text", + "content": " initialization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 375, + 263, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 263, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 263, + 388 + ], + "type": "text", + "content": "7.1. Initialization of " + }, + { + "bbox": [ + 55, + 375, + 263, + 388 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 375, + 263, + 388 + ], + "type": "text", + "content": " for Non-LLM Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 398, + 296, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 296, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 296, + 541 + ], + "type": "text", + "content": "Non-LLM models are relatively insensitive to " + }, + { + "bbox": [ + 55, + 398, + 296, + 541 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 55, + 398, + 296, + 541 + ], + "type": "text", + "content": ". Figure 8 shows the effect of varying " + }, + { + "bbox": [ + 55, + 398, + 296, + 541 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 55, + 398, + 296, + 541 + ], + "type": "text", + "content": " on validation performance across different tasks. All experiments follow the original setup and hyperparameters of their respective recipe. We observe that performance remains stable across a wide range of " + }, + { + "bbox": [ + 55, + 398, + 296, + 541 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 55, + 398, + 296, + 541 + ], + "type": "text", + "content": " values, with values between 0.5 and 1.2 generally yielding good results. We observe that adjusting " + }, + { + "bbox": [ + 55, + 398, + 296, + 541 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 55, + 398, + 296, + 541 + ], + "type": "text", + "content": " typically affects only the early stages of the training curves. The main exception is supervised ViT-L experiments, where training becomes unstable and diverges when " + }, + { + "bbox": [ + 55, + 398, + 296, + 541 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 55, + 398, + 296, + 541 + ], + "type": "text", + "content": " exceeds 0.6. In such cases, reducing the learning rate restores stability, as detailed below." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": "Smaller " + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": " results in more stable training. Building on previous observations, we further analyze the factors contributing to training instability. Our findings suggest that increasing either the model size or the learning rate requires lowering " + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": " to ensure stable training. Conversely, a higher " + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": " requires a lower learning rate to mitigate training instability. Figure 9 shows the ablation of the training stability of supervised ViT with ImageNet-1K dataset. We vary learning rates, model sizes, and " + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": " values. Training a larger model is more prone to failure, requiring smaller " + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": " values or learning rates for stable training. A similar instability pattern is also observed in LN-based models under comparable conditions, and setting " + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_0 = 0.5" + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": " results in a stability pattern similar to that of LN." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 285, + 554, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 554, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 554, + 334 + ], + "type": "text", + "content": "Setting " + }, + { + "bbox": [ + 313, + 285, + 554, + 334 + ], + "type": "inline_equation", + "content": "\\alpha_0 = 0.5" + }, + { + "bbox": [ + 313, + 285, + 554, + 334 + ], + "type": "text", + "content": " as the default. Based on our findings, we set " + }, + { + "bbox": [ + 313, + 285, + 554, + 334 + ], + "type": "inline_equation", + "content": "\\alpha_0 = 0.5" + }, + { + "bbox": [ + 313, + 285, + 554, + 334 + ], + "type": "text", + "content": " as the default value for all non-LLM models. This setting provides training stability comparable to LN while maintaining strong performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 338, + 467, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 467, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 467, + 350 + ], + "type": "text", + "content": "7.2. Initialization of " + }, + { + "bbox": [ + 313, + 338, + 467, + 350 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 338, + 467, + 350 + ], + "type": "text", + "content": " for LLMs" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 360, + 554, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 554, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 554, + 444 + ], + "type": "text", + "content": "Tuning " + }, + { + "bbox": [ + 313, + 360, + 554, + 444 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 313, + 360, + 554, + 444 + ], + "type": "text", + "content": " enhances LLM performance. As discussed earlier, the default setting of " + }, + { + "bbox": [ + 313, + 360, + 554, + 444 + ], + "type": "inline_equation", + "content": "\\alpha_0 = 0.5" + }, + { + "bbox": [ + 313, + 360, + 554, + 444 + ], + "type": "text", + "content": " generally performs well across most tasks. However, we find tuning " + }, + { + "bbox": [ + 313, + 360, + 554, + 444 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 313, + 360, + 554, + 444 + ], + "type": "text", + "content": " can substantially improve LLM performance. We tune " + }, + { + "bbox": [ + 313, + 360, + 554, + 444 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 313, + 360, + 554, + 444 + ], + "type": "text", + "content": " across LLaMA models by pretraining each on 30B tokens and comparing their training losses. Table 9 summarizes the tuned " + }, + { + "bbox": [ + 313, + 360, + 554, + 444 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 313, + 360, + 554, + 444 + ], + "type": "text", + "content": " values for each model. Two key findings emerge:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 446, + 554, + 547 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 314, + 446, + 554, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 446, + 554, + 483 + ], + "spans": [ + { + "bbox": [ + 314, + 446, + 554, + 483 + ], + "type": "text", + "content": "1. Larger models require smaller " + }, + { + "bbox": [ + 314, + 446, + 554, + 483 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 314, + 446, + 554, + 483 + ], + "type": "text", + "content": " values. Once the optimal " + }, + { + "bbox": [ + 314, + 446, + 554, + 483 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 314, + 446, + 554, + 483 + ], + "type": "text", + "content": " is determined for smaller models, the search space for larger models can be reduced accordingly." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 487, + 554, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 554, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 554, + 547 + ], + "type": "text", + "content": "2. Higher " + }, + { + "bbox": [ + 313, + 487, + 554, + 547 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 313, + 487, + 554, + 547 + ], + "type": "text", + "content": " values for attention blocks improve performance. We find that initializing " + }, + { + "bbox": [ + 313, + 487, + 554, + 547 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 487, + 554, + 547 + ], + "type": "text", + "content": " with higher values for DyT layers in attention blocks and lower values for DyT layers in other locations (i.e., within FFN blocks or before the final linear projection) improves performance." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "type": "table", + "bbox": [ + 337, + 552, + 533, + 629 + ], + "blocks": [ + { + "bbox": [ + 337, + 552, + 533, + 629 + ], + "lines": [ + { + "bbox": [ + 337, + 552, + 533, + 629 + ], + "spans": [ + { + "bbox": [ + 337, + 552, + 533, + 629 + ], + "type": "table", + "html": "
modelwidthdepthoptimal α0 (attention/other)
LLaMA 7B4096320.8/0.2
LLaMA 13B5120400.6/0.15
LLaMA 34B8196480.2/0.05
LLaMA 70B8196800.2/0.05
", + "image_path": "fb597714d0014d926a5f432a428b4d441b4c547466031c50d8db635e5192ad66.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 632, + 554, + 687 + ], + "lines": [ + { + "bbox": [ + 313, + 632, + 554, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 554, + 687 + ], + "type": "text", + "content": "Table 9. Optimal " + }, + { + "bbox": [ + 313, + 632, + 554, + 687 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 313, + 632, + 554, + 687 + ], + "type": "text", + "content": " for different LLaMA models. Larger models require smaller " + }, + { + "bbox": [ + 313, + 632, + 554, + 687 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 313, + 632, + 554, + 687 + ], + "type": "text", + "content": " values. We find it is important to initialize " + }, + { + "bbox": [ + 313, + 632, + 554, + 687 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 632, + 554, + 687 + ], + "type": "text", + "content": " differently in (1) attention blocks (\"attention\"), versus (2) the FFN blocks, and the final DyT layer before outputs (\"other\"). " + }, + { + "bbox": [ + 313, + 632, + 554, + 687 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 313, + 632, + 554, + 687 + ], + "type": "text", + "content": " in attention blocks require larger values." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "Further details and visualizations of the tuning of the initial value of " + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": " of LLMs are provided in the appendix C." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14907" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 57, + 553, + 202 + ], + "blocks": [ + { + "bbox": [ + 58, + 57, + 553, + 202 + ], + "lines": [ + { + "bbox": [ + 58, + 57, + 553, + 202 + ], + "spans": [ + { + "bbox": [ + 58, + 57, + 553, + 202 + ], + "type": "image", + "image_path": "21493d7352e45b062d793c5e17907856f6cf5d55b553580b1afa8b8672c0e646.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 205, + 555, + 239 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 555, + 239 + ], + "type": "text", + "content": "Figure 9. Stability across varying " + }, + { + "bbox": [ + 55, + 205, + 555, + 239 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 55, + 205, + 555, + 239 + ], + "type": "text", + "content": " values, learning rates, and model sizes. We train supervised ViT models on the ImageNet-1K dataset and observe that larger models are more prone to instability for both LN and DyT models. Lowering the learning rate or reducing " + }, + { + "bbox": [ + 55, + 205, + 555, + 239 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 55, + 205, + 555, + 239 + ], + "type": "text", + "content": " enhances stability. LN shows similar stability to DyT with " + }, + { + "bbox": [ + 55, + 205, + 555, + 239 + ], + "type": "inline_equation", + "content": "\\alpha_0 = 0.5" + }, + { + "bbox": [ + 55, + 205, + 555, + 239 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 246, + 142, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 246, + 142, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 246, + 142, + 258 + ], + "type": "text", + "content": "8. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 271, + 297, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 271, + 297, + 415 + ], + "spans": [ + { + "bbox": [ + 54, + 271, + 297, + 415 + ], + "type": "text", + "content": "Mechanisms of normalization layers. There has been a rich line of work investigating normalization layers' role in enhancing model performance through various mechanisms. These include stabilizing gradient flow during training [9, 17, 52], reducing sensitivity to weight initialization [18, 70, 90], moderating outlier eigenvalues [10, 44], auto-tuning learning rates [4, 77], and smoothing the loss landscape for more stable optimization [68]. These earlier works focused on studying batch normalization. Recent studies [16, 53, 59] further highlight the connection between normalization layers and sharpness reduction, which contributes to better generalization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 421, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 421, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 54, + 421, + 296, + 529 + ], + "type": "text", + "content": "Normalization in Transformers. With the rise of Transformer [82], research has increasingly focused on layer normalization [5], which has proven particularly effective for sequential data in natural language tasks [61, 85, 86]. Recent work [62] reveals that layer normalization introduces strong non-linearity, enhancing the model's representational capacity. Additionally, studies [48, 51] demonstrate that modifying the location of normalization layers within Transformers can improve convergence properties." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 297, + 713 + ], + "type": "text", + "content": "Removing normalization. Many studies have explored how to train deep models without normalization layers. Several works [6, 18, 90] explore alternative weight initialization schemes to stabilize training. Self-normalizing networks [45] introduce scaled exponential linear units (SELUs) and a carefully chosen initialization scheme to maintain stable activations and gradient flow without the need for explicit normalization layers. The pioneering work by Brock et al. [11, 12] show that high-performing ResNets can be trained without normalization [72] through combination of initialization techniques [18], weight normalization [37, 65, 67], and adaptive gradient clipping [12]. Additionally, their training strategy incorporates extensive data augmentation [15] and regularization [36, 73]. The studies above are based on various ConvNet models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 247, + 555, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 555, + 392 + ], + "type": "text", + "content": "In Transformer architectures, He and Hofmann [32] explore modifications to Transformer blocks that reduce reliance on normalization layers and skip connections. Jha and Reagen [42] introduce AERO, a Softmax-only LLM that improves inference efficiency and privacy with minimal performance loss. Alternatively, Heimersheim [35] propose a method to gradually remove LN from pretrained networks by fine-tuning the model after removing each normalization layer. Unlike previous approaches, DyT requires minimal modifications to both the architecture and the training recipe. Despite its simplicity, DyT achieves stable training and comparable performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 400, + 389, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 389, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 389, + 412 + ], + "type": "text", + "content": "9. Limitations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 420, + 555, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 555, + 506 + ], + "type": "text", + "content": "We conduct experiments on networks using either LN or RMSNorm because of their popularity in Transformers and other modern architectures. Preliminary experiments (see Appendix E) indicate that DyT struggles to replace BN directly in classic networks like ResNets. It remains to be studied in more depth whether and how DyT can adapt to models with other types of normalization layers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 514, + 392, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 514, + 392, + 526 + ], + "spans": [ + { + "bbox": [ + 314, + 514, + 392, + 526 + ], + "type": "text", + "content": "10. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "In this work, we demonstrate modern neural networks, in particular Transformers, can be trained without normalization layers. This is done through Dynamic Tanh (DyT), a simple replacement for traditional normalization layers. It adjusts the input activation range via a learnable scaling factor " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": " and then squashes the extreme values through an " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "-shaped tanh function. Although a simpler function, it effectively captures the behavior of normalization layers. Under various settings, models with DyT match or exceed the performance of their normalized counterparts. The findings challenge the conventional understanding of the necessity of normalization layers in training modern neural networks. Our study also contributes to understanding the mechanisms of normalization layers, one of the most fundamental building blocks in deep neural networks." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14908" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "type": "text", + "content": "[1] Edgar D Adrian. The impulses produced by sensory nerve endings: Part 1. The Journal of Physiology, 1926." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 114, + 295, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 114, + 295, + 147 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 295, + 147 + ], + "type": "text", + "content": "[2] Edgar D Adrian and Yngve Zotterman. The impulses produced by sensory nerve-endings: Part 2. the response of a single end-organ. The Journal of Physiology, 1926." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 148, + 294, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 148, + 294, + 181 + ], + "spans": [ + { + "bbox": [ + 62, + 148, + 294, + 181 + ], + "type": "text", + "content": "[3] Edgar D Adrian and Yngve Zotterman. The impulses produced by sensory nerve endings: Part 3. impulses set up by touch and pressure. The Journal of Physiology, 1926." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 182, + 294, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 294, + 215 + ], + "type": "text", + "content": "[4] Sanjeev Arora, Zhiyuan Li, and Kaifeng Lyu. Theoretical analysis of auto rate-tuning by batch normalization. arXiv preprint arXiv:1812.03981, 2018." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 217, + 294, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 217, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 62, + 217, + 294, + 248 + ], + "type": "text", + "content": "[5] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 251, + 294, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 294, + 283 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 294, + 283 + ], + "type": "text", + "content": "[6] Thomas Bachlechner, Bodhisattwa Prasad Majumder, Henry Mao, Gary Cottrell, and Julian McAuley. Rezero is all you need: Fast convergence at large depth. In UAI, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 285, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 294, + 317 + ], + "type": "text", + "content": "[7] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. NeurIPS, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 319, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 294, + 361 + ], + "type": "text", + "content": "[8] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 364, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 364, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 364, + 294, + 407 + ], + "type": "text", + "content": "[9] David Balduzzi, Marcus Frean, Lennox Leary, JP Lewis, Kurt Wan-Duo Ma, and Brian McWilliams. The shattered gradients problem: If resnets are the answer, then what is the question? In ICML, 2017." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 409, + 294, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 409, + 294, + 440 + ], + "spans": [ + { + "bbox": [ + 57, + 409, + 294, + 440 + ], + "type": "text", + "content": "[10] Nils Bjorck, Carla P Gomes, Bart Selman, and Kilian Q Weinberger. Understanding batch normalization. NeurIPS, 2018." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 443, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 443, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 57, + 443, + 294, + 475 + ], + "type": "text", + "content": "[11] Andrew Brock, Soham De, and Samuel L Smith. Characterizing signal propagation to close the performance gap in unnormalized resnets. arXiv preprint arXiv:2101.08692, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 477, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 477, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 57, + 477, + 294, + 509 + ], + "type": "text", + "content": "[12] Andrew Brock, Soham De, Samuel L Smith, and Karen Simonyan. High-performance large-scale image recognition without normalization. In ICML, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 511, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 511, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 57, + 511, + 294, + 554 + ], + "type": "text", + "content": "[13] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arXiv preprint arXiv:2403.17297, 2024." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 556, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 556, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 556, + 294, + 599 + ], + "type": "text", + "content": "[14] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 601, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 294, + 634 + ], + "type": "text", + "content": "[15] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In CVPR Workshops, 2020." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 635, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 294, + 666 + ], + "type": "text", + "content": "[16] Yan Dai, Kwangjun Ahn, and Suvrit Sra. The crucial role of normalization in sharpness-aware minimization. NeurIPS, 2024." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[17] Hadi Daneshmand, Jonas Kohler, Francis Bach, Thomas Hofmann, and Aurelien Lucchi. Batch normalization provably avoids ranks collapse for randomly initialised deep networks. NeurIPS, 2020." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "text", + "content": "[18] Soham De and Sam Smith. Batch normalization biases residual blocks towards the identity function in deep networks. NeurIPS, 2020." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 107, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 139 + ], + "type": "text", + "content": "[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 141, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 206 + ], + "type": "text", + "content": "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 209, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 261 + ], + "type": "text", + "content": "[21] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 264, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 264, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 317, + 264, + 553, + 296 + ], + "type": "text", + "content": "[22] Leo Feng, Frederick Tung, Mohamed Osama Ahmed, Yoshua Bengio, and Hossein Hajimirsadegh. Were rnns all we needed? arXiv preprint arXiv:2410.01201, 2024." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 298, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 298, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 317, + 298, + 553, + 330 + ], + "type": "text", + "content": "[23] Foundation Model Stack. Github: FMS FSDP. https://github.com/foundation-model-stack/fms-fsdp. Accessed: 2025-01-23." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 332, + 553, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 332, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 317, + 332, + 553, + 418 + ], + "type": "text", + "content": "[24] Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac'h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lintang Sutawika, Eric Tang, Anish Thite, Ben Wang, Kevin Wang, and Andy Zou. A framework for few-shot language model evaluation." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 421, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 421, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 317, + 421, + 553, + 475 + ], + "type": "text", + "content": "[25] Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, Shawn Presser, and Connor Leahy. The Pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027, 2020." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 476, + 553, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 476, + 553, + 497 + ], + "spans": [ + { + "bbox": [ + 317, + 476, + 553, + 497 + ], + "type": "text", + "content": "[26] Xinyang Geng and Hao Liu. Openlama: An open reproduction of llama, 2023." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 499, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 499, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 499, + 553, + 521 + ], + "type": "text", + "content": "[27] Ensembl GRCh38. p13 (genome reference consortium human build 38), insdc assembly, 2013." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 522, + 553, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 522, + 553, + 565 + ], + "spans": [ + { + "bbox": [ + 317, + 522, + 553, + 565 + ], + "type": "text", + "content": "[28] Katarína Grešová, Vlastimil Martinek, David Čechák, Petr Šimeček, and Panagiotis Alexiou. Genomic benchmarks: a collection of datasets for genomic sequence classification. BMC Genomic Data, 2023." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 567, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 567, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 567, + 553, + 599 + ], + "type": "text", + "content": "[29] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 601, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 553, + 655 + ], + "type": "text", + "content": "[30] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 658, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 689 + ], + "type": "text", + "content": "[31] HazyResearch. Github: Hyenadna. https://github.com/HazyResearch/hyena-dna.git. Accessed: 2025-01-23." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[32] Bobby He and Thomas Hofmann. Simplifying transformer blocks. arXiv preprint arXiv:2311.01906, 2023." + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14909" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 105 + ], + "type": "text", + "content": "[33] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "text", + "content": "[34] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 142, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 142, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 57, + 142, + 294, + 163 + ], + "type": "text", + "content": "[35] Stefan Heimersheim. You can remove gpt2's layernorm by fine-tuning. arXiv preprint arXiv:2409.13710, 2024." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 166, + 295, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 166, + 295, + 197 + ], + "spans": [ + { + "bbox": [ + 57, + 166, + 295, + 197 + ], + "type": "text", + "content": "[36] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In ECCV, 2016." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 200, + 294, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 200, + 294, + 232 + ], + "spans": [ + { + "bbox": [ + 57, + 200, + 294, + 232 + ], + "type": "text", + "content": "[37] Lei Huang, Xianglong Liu, Yang Liu, Bo Lang, and Dacheng Tao. Centered weight normalization in accelerating training of deep neural networks. In ICCV, 2017." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 235, + 294, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 235, + 294, + 267 + ], + "spans": [ + { + "bbox": [ + 57, + 235, + 294, + 267 + ], + "type": "text", + "content": "[38] Lei Huang, Jie Qin, Yi Zhou, Fan Zhu, Li Liu, and Ling Shao. Normalization techniques in training dnns: Methodology, analysis and application. TPAMI, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 269, + 294, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 269, + 294, + 301 + ], + "spans": [ + { + "bbox": [ + 57, + 269, + 294, + 301 + ], + "type": "text", + "content": "[39] Xiao Shi Huang, Felipe Perez, Jimmy Ba, and Maksims Volkovs. Improving transformer optimization through better initialization. In ICML, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 304, + 294, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 304, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 57, + 304, + 294, + 335 + ], + "type": "text", + "content": "[40] Hugging Face. Hugging Face: LLaMA 2. https://huggingface.co/meta-llama/Llama-2-7b-hf. Accessed: 2025-01-23." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 338, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 338, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 57, + 338, + 294, + 369 + ], + "type": "text", + "content": "[41] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 373, + 294, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 373, + 294, + 404 + ], + "spans": [ + { + "bbox": [ + 57, + 373, + 294, + 404 + ], + "type": "text", + "content": "[42] Nandan Kumar Jha and Brandon Reagen. Aero: Softmax-only llms for efficient private inference. arXiv preprint arXiv:2410.13060, 2024." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 406, + 294, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 406, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 57, + 406, + 294, + 460 + ], + "type": "text", + "content": "[43] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 462, + 294, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 462, + 294, + 495 + ], + "spans": [ + { + "bbox": [ + 57, + 462, + 294, + 495 + ], + "type": "text", + "content": "[44] Ryo Karakida, Shotaro Akaho, and Shun-ichi Amari. The normalization method for alleviating pathological sharpness in wide neural networks. NeurIPS, 2019." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 498, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 498, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 57, + 498, + 294, + 529 + ], + "type": "text", + "content": "[45] Günter Klambauer, Thomas Unterthiner, Andreas Mayr, and Sepp Hochreiter. Self-normalizing neural networks. NeurIPS, 2017." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 532, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 532, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 57, + 532, + 294, + 563 + ], + "type": "text", + "content": "[46] Kuleshov Group. Github: Caduceus. https://github.com/kuleshov-group/caduceus.git. Accessed: 2025-01-23." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 567, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 294, + 588 + ], + "type": "text", + "content": "[47] Yann LeCun. A path towards autonomous machine intelligence version 0.9.2, 2022-06-27. Open Review, 2022." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 590, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 294, + 622 + ], + "type": "text", + "content": "[48] Pengxiang Li, Lu Yin, and Shiwei Liu. Mix-ln: Unleashing the power of deeper layers by combining pre-ln and post-ln. arXiv preprint arXiv:2412.13795, 2024." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 624, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 677 + ], + "type": "text", + "content": "[49] Aixin Liu, Bei Feng, Bin Wang, Bingxuan Wang, Bo Liu, Chenggang Zhao, Chengqi Dengr, Chong Ruan, Damai Dai, Daya Guo, et al. Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model. arXiv preprint arXiv:2405.04434, 2024." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 680, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 294, + 712 + ], + "type": "text", + "content": "[50] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 38, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "type": "text", + "content": "[51] Ilya Loshchilov, Cheng-Ping Hsieh, Simeng Sun, and Boris Ginsburg. ngpt: Normalized transformer with representation learning on the hypersphere. arXiv preprint arXiv:2410.01131, 2024." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "type": "text", + "content": "[52] Ekdeep S Lubana, Robert Dick, and Hidenori Tanaka. Beyond batchnorm: Towards a unified understanding of normalization in deep learning. NeurIPS, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 152, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 183 + ], + "type": "text", + "content": "[53] Kaifeng Lyu, Zhiyuan Li, and Sanjeev Arora. Understanding the generalization benefit of normalization layers: Sharpness reduction. NeurIPS, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 186, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 217 + ], + "type": "text", + "content": "[54] Meta Research. Github: ConvNeXt. https://github.com/facebookresearch/ConvNeXt, . Accessed: 2025-01-23." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 220, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 240 + ], + "type": "text", + "content": "[55] Meta Research. Github: DINO. https://github.com/facebookresearch/dino, . Accessed: 2025-01-23." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 243, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 263 + ], + "type": "text", + "content": "[56] Meta Research. Github: DiT. https://github.com/facebookresearch/DiT,. Accessed: 2025-01-23." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 265, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 265, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 553, + 285 + ], + "type": "text", + "content": "[57] Meta Research. Github: MAE. https://github.com/facebookresearch/mae, . Accessed: 2025-01-23." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 288, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 553, + 319 + ], + "type": "text", + "content": "[58] Meta Research. Github: wav2vec 2.0. https://github.com/facebookresearch/fairseq, . Accessed: 2025-01-23." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 321, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 353 + ], + "type": "text", + "content": "[59] Maximilian Mueller, Tiffany Vlaar, David Rolnick, and Matthias Hein. Normalization layers are all that sharpness-aware minimization needs. NeurIPS, 2024." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 355, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 355, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 317, + 355, + 553, + 408 + ], + "type": "text", + "content": "[60] Eric Nguyen, Michael Poli, Marjan Faizi, Armin Thomas, Michael Wornow, Callum Birch-Sykes, Stefano Massaroli, Aman Patel, Clayton Rabideau, Yoshua Bengio, et al. Hyenadna: Long-range genomic sequence modeling at single nucleotide resolution. NeurIPS, 2024." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "type": "text", + "content": "[61] Toan Q Nguyen and Julian Salazar. Transformers without tears: Improving the normalization of self-attention. arXiv preprint arXiv:1910.05895, 2019." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 445, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 476 + ], + "type": "text", + "content": "[62] Yunhao Ni, Yuxin Guo, Junlong Jia, and Lei Huang. On the nonlinearity of layer normalization. arXiv preprint arXiv:2406.01255, 2024." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "type": "text", + "content": "[63] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In ICASSP, 2015." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 512, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 533 + ], + "type": "text", + "content": "[64] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 317, + 535, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 577 + ], + "type": "text", + "content": "[65] Siyuan Qiao, Huiyu Wang, Chenxi Liu, Wei Shen, and Alan Yuille. Micro-batch training with batch-channel normalization and weight standardization. arXiv preprint arXiv:1903.10520, 2019." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 317, + 579, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 579, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 579, + 553, + 622 + ], + "type": "text", + "content": "[66] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 2020." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 624, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 655 + ], + "type": "text", + "content": "[67] Tim Salimans and Durk P Kingma. Weight normalization: A simple reparameterization to accelerate training of deep neural networks. NeurIPS, 2016." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "text", + "content": "[68] Shibani Santurkar, Dimitris Tsipras, Andrew Ilyas, and Aleksander Madry. How does batch normalization help optimization? NeurIPS, 2018." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[69] Yair Schiff, Chia-Hsiang Kao, Aaron Gokaslan, Tri Dao, Albert Gu, and Volodymyr Kuleshov. Caduceus: Bi" + } + ] + } + ], + "index": 37 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14910" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "directional equivariant long-range dna sequence modeling. arXiv preprint arXiv:2403.03234, 2024." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 128 + ], + "type": "text", + "content": "[70] Jie Shao, Kai Hu, Changhu Wang, Xiangyang Xue, and Bhiksha Raj. Is normalization indispensable for training deep neural network? NeurIPS, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 295, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 161 + ], + "type": "text", + "content": "[71] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 162, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 195 + ], + "type": "text", + "content": "[72] Samuel L Smith, Andrew Brock, Leonard Berrada, and Soham De. Convnets match vision transformers at scale. arXiv preprint arXiv:2310.16764, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 196, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 294, + 228 + ], + "type": "text", + "content": "[73] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. JMLR, 2014." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 228, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 261 + ], + "type": "text", + "content": "[74] Wenfang Sun, Xinyuan Song, Pengxiang Li, Lu Yin, Yefeng Zheng, and Shiwei Liu. The curse of depth in large language models. arXiv preprint arXiv:2502.05795, 2025." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "type": "text", + "content": "[75] Yu Sun, Xinhao Li, Karan Dalal, Jiarui Xu, Arjun Vikram, Genghan Zhang, Yann Dubois, Xinlei Chen, Xiaolong Wang, Sanmi Koyejo, et al. Learning to (learn at test time): Rnns with expressive hidden states. arXiv preprint arXiv:2407.04620, 2024." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 317, + 294, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 294, + 349 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 294, + 349 + ], + "type": "text", + "content": "[76] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 350, + 294, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 382 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 382 + ], + "type": "text", + "content": "[77] Hidenori Tanaka and Daniel Kunin. Noether's learning dynamics: Role of symmetry breaking in neural networks. NeurIPS, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "type": "text", + "content": "[78] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlp architecture for vision. NeurIPS, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 437, + 294, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 492 + ], + "type": "text", + "content": "[79] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 493, + 294, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 548 + ], + "type": "text", + "content": "[80] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 548, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 581 + ], + "type": "text", + "content": "[81] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv preprint arXiv:1607.08022, 2016." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 582, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 613 + ], + "type": "text", + "content": "[82] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 614, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 635 + ], + "type": "text", + "content": "[83] Yuxin Wu and Kaiming He. Group normalization. In ECCV, 2018." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 636, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 669 + ], + "type": "text", + "content": "[84] Saining Xie, Ross Girshick, Piotr Dólar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In CVPR, 2017." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[85] Ruibin Xiong, Yunchang Yang, Di He, Kai Zheng, Shuxin Zheng, Chen Xing, Huishuai Zhang, Yanyan Lan, Liwei Wang, and Tieyan Liu. On layer normalization in the transformer architecture. In ICML, 2020." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 308 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[86] Jingjing Xu, Xu Sun, Zhiyuan Zhang, Guangxiang Zhao, and Junyang Lin. Understanding and improving layer normalization. NeurIPS, 2019." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "type": "text", + "content": "[87] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "text", + "content": "[88] Shuangfei Zhai, Tatiana Likhomanenko, Etai Littwin, Dan Busbridge, Jason Ramapuram, Yizhe Zhang, Jiatao Gu, and Joshua M Susskind. Stabilizing transformer training by preventing attention entropy collapse. In ICML, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 197, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 217 + ], + "type": "text", + "content": "[89] Biao Zhang and Rico Sennrich. Root mean square layer normalization. NeurIPS, 2019." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 219, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 252 + ], + "type": "text", + "content": "[90] Hongyi Zhang, Yann N Dauphin, and Tengyu Ma. Fixup initialization: Residual learning without normalization. arXiv preprint arXiv:1901.09321, 2019." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "type": "text", + "content": "[91] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024." + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14911" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Traversing Distortion-Perception Tradeoff using a Single Score-Based Generative Model/dcdd686b-eeb1-4343-a7e6-20b5cd329198_content_list.json b/2025/Traversing Distortion-Perception Tradeoff using a Single Score-Based Generative Model/dcdd686b-eeb1-4343-a7e6-20b5cd329198_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1951f7ad9375d79b06ef421a074776cbd19166fe --- /dev/null +++ b/2025/Traversing Distortion-Perception Tradeoff using a Single Score-Based Generative Model/dcdd686b-eeb1-4343-a7e6-20b5cd329198_content_list.json @@ -0,0 +1,1905 @@ +[ + { + "type": "text", + "text": "Traversing Distortion-Perception Tradeoff using a Single Score-Based Generative Model", + "text_level": 1, + "bbox": [ + 148, + 128, + 849, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuhan Wang $^{1}$ , Suzhi Bi $^{2*}$ , Ying-Jun Angela Zhang $^{1}$ , Xiaojun Yuan $^{3}$", + "bbox": [ + 225, + 202, + 769, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The Chinese University of Hong Kong 2Shenzhen University", + "bbox": [ + 254, + 220, + 745, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3University of Electronic Science and Technology of China", + "bbox": [ + 264, + 239, + 733, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wy023, yjzhang}@ie.cuhk.edu.hk, bsz@szu.edu.cn, xjyuan@uestc.edu.cn", + "bbox": [ + 196, + 258, + 797, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The distortion-perception (DP) tradeoff reveals a fundamental conflict between distortion metrics (e.g., MSE and PSNR) and perceptual quality. Recent research has increasingly concentrated on evaluating denoising algorithms within the DP framework. However, existing algorithms either prioritize perceptual quality by sacrificing acceptable distortion, or focus on minimizing MSE for faithful restoration. When the goal shifts or noisy measurements vary, adapting to different points on the DP plane needs retraining or even re-designing the model. Inspired by recent advances in solving inverse problems using score-based generative models, we explore the potential of flexibly and optimally traversing DP tradeoffs using a single pre-trained score-based model. Specifically, we introduce a variance-scaled reverse diffusion process and theoretically characterize the marginal distribution. We then prove that the proposed sample process is an optimal solution to the DP tradeoff for conditional Gaussian distribution. Experimental results on two-dimensional and image datasets illustrate that a single score network can effectively and flexibly traverse the DP tradeoff for general denoising problems.", + "bbox": [ + 89, + 340, + 485, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 89, + 686, + 232, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, we have witnessed rapid progress in image restoration algorithms, especially deep learning-based implementations that have shown remarkable achievements in solving general inverse problems. How to evaluate the performance of emerging algorithms is a crucial yet complicated problem. Traditional full reference distortion metrics, such as Mean Square Error (MSE), focus on the pixel-level accuracy between the original image $X$ and its reconstruction $\\hat{X}$ . On the other hand, perceptual quality, referring to the degree to which an image looks natural rather than algorithmically generated [14], is also an important mea", + "bbox": [ + 89, + 710, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sure. It has been demonstrated that perceptual quality could be associated with the distance between the distributions of natural images $p_{X}$ and generated images $p_{\\hat{X}}$ [2, 3, 12].", + "bbox": [ + 511, + 310, + 906, + 356 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Improvements in distortion measures do not necessarily lead to enhancements in perceptual quality. In fact, [2] has demonstrated that there exists a tradeoff between distortion and perception, termed as distortion-perception (DP) trade-off. Mathematically, it can be modeled as", + "bbox": [ + 511, + 356, + 906, + 431 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\nD(P) = \\min_{p_{\\hat{X} |Y}}\\mathbb{E}[\\Delta (X,\\hat{X})],\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 436, + 800, + 463 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\text {s . t .} d (p _ {X}, p _ {\\tilde {X}}) \\leq P, \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 467, + 785, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "where $Y$ is the degraded observation, $\\Delta : \\mathcal{X} \\times \\hat{\\mathcal{X}} \\to \\mathbb{R}^+$ is a full reference distortion function (e.g., such as square-error), and $d(\\cdot, \\cdot)$ is some divergence between probability distribution, such as Wasserstein distance [16] or Kulback-Leibler (KL) divergence.", + "bbox": [ + 511, + 491, + 905, + 568 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Since the introduction of the DP tradeoff, more and more algorithm designs have focused on performance evaluation on the DP plane for specific denoising tasks, e.g., image deblurring [15, 25] and super-resolution [11, 13, 23]. However, most algorithms tend to seek better perceptual quality at the cost of distortion, or focus on optimizing MSE, a distortion measure, to ensure restoration fidelity. When the restoration objective shifts between distortion and perception, we need to retrain or even re-design the network, which can be computationally intensive. In many practical scenarios, it is significant to flexibly and effectively fulfill diverse task objectives or user demands with a single model at inference time.", + "bbox": [ + 511, + 568, + 906, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There have been explorations into theories and algorithms to traverse the DP tradeoff in inference time [7, 15]. Specifically, in [15], the authors developed a framework based on conditional generative adversarial networks (CGAN) with an additional penalty on posterior expectation. By averaging different numbers of samples or adjusting the noise level injected into the generator, this CGAN-based framework can access various points along the tradeoff. However, using multiple image samples and averaging", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 886, + 236, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2377", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "them may not be efficient during inference, and the method is not provably optimal. Furthermore, the training process relies on pairs of clean and noisy data. Thus, retraining is required for different noise levels and measurements. In [7], the DP tradeoff is studied in Wasserstein space. Theoretical results show that estimators on the DP curve can be constructed by linearly combining two estimators at the extremes: an optimal MSE minimizer and a perfect perception sampler with minimum distortion. This idea is further applied in burst restoration [27]. In practice, a method that achieves a relatively low MSE, alongside another method that achieves good perceptual quality, can serve as approximations of these two extreme estimators for a specific task. Nonetheless, different models must be deployed to address various measurements, rendering this method less flexible to general inverse problems.", + "bbox": [ + 89, + 90, + 483, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent research has demonstrated the generative capability of score-based diffusion models [4, 8, 10, 18, 20, 21] in tackling general inverse problems using a single score network [5, 6, 19, 24, 26, 29]. Score-based diffusion models learn the prior distribution of the data $\\mathbf{x}$ by training a score network to match the gradient of the logarithm density $\\nabla_{\\mathbf{x}}\\log p(\\mathbf{x})$ , referred to as the score. After observing a noisy measurement $\\mathbf{y}$ , sampling from posteriors involves approximating the conditional score $\\nabla_{\\mathbf{x}}\\log p(\\mathbf{x}|\\mathbf{y})$ . Theoretical insights provided in [28] also reveal the potential for recovering Minimum Mean Square Error (MMSE) estimation by propagating the mean of the reverse diffusion process. Inspired by these advancements, we explore the potential of traversing DP tradeoffs for different tasks using a single score network. Our main contributions are summarized below:", + "bbox": [ + 91, + 337, + 483, + 577 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- First, we propose a variance-scaled reverse diffusion process and theoretically characterize the marginal distributions produced by this novel reverse process. It is demonstrated that the mean of our proposed sampling converges towards the MMSE point while the marginal covariance is scaled by a scaling factor. By tuning a parameter that dictates the variance of reverse sampling, we can flexibly navigate from the MMSE point to the posterior distribution, where the two extremes are achieved by setting the parameter to zero and one, respectively.", + "- Subsequently, we show that the reconstruction obtained from the variance-scaled reverse sampling represents the optimal solution to the conditional DP tradeoff for multivariate Gaussian distributions, assessed through MSE and Wasserstein-2 distance metrics.", + "- Finally, we validate our methodology through experiments conducted on various two-dimensional datasets and a real-world image dataset. With a single pre-trained score network, we can traverse the DP tradeoff across different inverse problems with varying noise levels. The results indicate that the variance-scaled reverse diffusion" + ], + "bbox": [ + 89, + 583, + 483, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "process achieves a more complete empirical DP tradeoff curve and better MSE than the GAN-based method and other inverse problem solvers, demonstrating the effectiveness and flexibility of the proposed framework.", + "bbox": [ + 524, + 90, + 906, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Notations: For a random variable $X$ denoted by a capital letter, we use small letter $\\mathbf{x}$ to denote its realizations, and use $p_{X}(\\mathbf{x})$ to denote the distribution over its alphabet $\\mathcal{X}$ . When there is no ambiguity, the distribution can be abbreviated as $p(\\mathbf{x})$ . For a discrete sequence of random variables $X_0, X_1, \\dots, X_T$ , we abbreviate them as $\\{X_k\\}_{k=0}^T$ and use $\\mathbf{x}_{0:T}$ to denote their realizations. The expectation and conditional expectation of $X$ given $Y = \\mathbf{y}$ is $\\mathbb{E}_{p(\\mathbf{x})}[X]$ and $\\mathbb{E}_{p(\\mathbf{x}|\\mathbf{y})}[X]$ . We use $\\mathrm{Cov}_{p(\\mathbf{x})}[X]$ and $\\mathrm{Cov}_{p(\\mathbf{x}|\\mathbf{y})}[X]$ to represent the covariance and conditional covariance of $X$ given $Y = \\mathbf{y}$ . Matrices are denoted by uppercase boldface letters (e.g., $\\boldsymbol{\\Sigma}$ ). $\\mathrm{Tr}(\\boldsymbol{\\Sigma})$ and $\\boldsymbol{\\Sigma}^{-1}$ represent the trace and inverse.", + "bbox": [ + 511, + 152, + 906, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Backgrounds", + "text_level": 1, + "bbox": [ + 511, + 344, + 660, + 359 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 Score-based generative model", + "text_level": 1, + "bbox": [ + 511, + 369, + 785, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Score-based generative models or diffusion models define the generative process as the reverse of a diffusion process. Specifically, the diffusion process (i.e., forward process) is a Markov chain $\\{X_{k}\\}_{k = 0}^{T}$ , with joint distribution", + "bbox": [ + 511, + 392, + 905, + 454 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\mathbf {x} _ {0: T}\\right) := p \\left(\\mathbf {x} _ {0}\\right) \\prod_ {k = 0} ^ {T - 1} p \\left(\\mathbf {x} _ {k + 1} \\mid \\mathbf {x} _ {k}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 459, + 906, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "which gradually diffuses the data $X_0 \\in \\mathbb{R}^d$ with distribution $p(\\mathbf{x}_0) \\triangleq p_{\\mathrm{data}}$ . Variance preserving (VP) diffusion [8] adopts $p(\\mathbf{x}_{k+1}|\\mathbf{x}_k) \\triangleq \\mathcal{N}(\\mathbf{x}_{k+1}; \\sqrt{1 - \\beta_{k+1}}\\mathbf{x}_k, \\beta_{k+1}\\mathbf{I})$ with monotonically increasing variance schedule $\\beta_T \\geq \\beta_{T-1} \\geq \\dots \\geq \\beta_0 = 0$ . When $T \\to \\infty$ , $\\mathbf{x}_T$ converges to an isotropic Gaussian distribution $\\mathcal{N}(\\mathbf{0},\\mathbf{I})$ .", + "bbox": [ + 511, + 507, + 905, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To generate a sample following the data distribution $p(\\mathbf{x}_0)$ , we can start from the standard Gaussian $p_{\\theta}(\\mathbf{x}_T) \\coloneqq \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ and follow the reverse process", + "bbox": [ + 511, + 598, + 905, + 643 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\mathbf {x} _ {0: T}\\right) := p _ {\\theta} \\left(\\mathbf {x} _ {T}\\right) \\prod_ {k = 0} ^ {T - 1} p _ {\\theta} \\left(\\mathbf {x} _ {k} \\mid \\mathbf {x} _ {k + 1}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 650, + 906, + 691 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with transitions $p_{\\theta}(\\mathbf{x}_k|\\mathbf{x}_{k + 1})\\sim \\mathcal{N}(\\mu_{\\theta}(\\mathbf{x}_k,k),\\boldsymbol {\\Sigma}_{\\theta}(\\mathbf{x}_k,k))$ parameterized by $\\theta$ . It will be shown later that the reverse sampling is closely related to the score function for each step, i.e., $\\nabla_{\\mathbf{x}_k}\\log p(\\mathbf{x}_k)$ , for $k = 0,1,\\dots ,T$", + "bbox": [ + 511, + 696, + 905, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In continuous-time, we can describe the diffusion process $(X_{t})_{t\\in [0,T_{c}]}$ , $X_{t}\\in \\mathbb{R}^{d}$ with a stochastic differential equation (SDE) [21]", + "bbox": [ + 511, + 758, + 905, + 803 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nd X _ {t} = - \\frac {1}{2} \\beta (t) X _ {t} d t + \\sqrt {\\beta (t)} d W _ {t}, X _ {0} \\sim p _ {\\text {d a t a}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 808, + 906, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $(W_{t})_{t\\in [0,T_{c}]}$ is a standard Brownian motion, and $\\beta (t)$ is the noise schedule. Let $(p_t)_{t\\in [0:T_c]}$ be the path distribution associated with (3). To perform data sampling following the distribution $p_{\\mathrm{data}}$ we can reverse the SDE (3) and", + "bbox": [ + 511, + 842, + 905, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2378", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "apply discretization. According to [1] and [21], the reverse SDE corresponding to (3) is", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} d \\overleftarrow {X} _ {t} = \\left[ \\frac {1}{2} \\beta \\left(T _ {c} - t\\right) \\overleftarrow {X} _ {t} + \\beta \\left(T _ {c} - t\\right) \\nabla \\log p _ {T _ {c} - t} (\\overleftarrow {X} _ {t}) \\right] d t \\\\ + \\sqrt {\\beta (T _ {c} - t)} d \\tilde {W} _ {t}, \\overleftarrow {X} _ {0} \\sim p _ {T _ {c}}, \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 125, + 482, + 172 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $(\\tilde{W}_t)_{t\\in [0,T_c]}$ is another standard Brownian motion. The reverse SDE produces a time-reverse process $(\\overleftarrow{X}_t)_{t\\in [0,T_c]}$ where $\\overleftarrow{X}_t$ has the same distribution with $X_{T_c - t}$ . Note that the drift function in (4) depends on the score function $\\nabla_{\\mathbf{x}_t}\\log p_t(\\mathbf{x}_t)$ , which can be approximated by a time-aware neural network trained with denoising score matching [22].", + "bbox": [ + 89, + 179, + 483, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Score-based model for posterior sampling", + "text_level": 1, + "bbox": [ + 89, + 295, + 455, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In many application scenarios, we may observe a noisy version of the measurement of the data $X_0$ , given by", + "bbox": [ + 89, + 318, + 483, + 349 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nY = \\mathcal {A} (X _ {0}) + N,\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 356, + 349, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{A}(\\cdot):\\mathbb{R}^d\\to \\mathbb{R}^n$ is a measurement operator and $N\\in \\mathbb{R}^n$ is the measurement noise with $N\\sim \\mathcal{N}(\\mathbf{0},\\sigma_n^2\\mathbf{I})$", + "bbox": [ + 89, + 380, + 483, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Score-based generative models have shown powerful capability in solving general inverse problems [6, 10, 19, 26, 28, 29]. Leveraging the diffusion model as the prior, sampling from the reverse posterior distribution", + "bbox": [ + 89, + 412, + 483, + 472 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\mathbf {x} _ {0: T} \\mid \\mathbf {y}\\right) := p \\left(\\mathbf {x} _ {T} \\mid \\mathbf {y}\\right) \\prod_ {k = 0} ^ {T - 1} p \\left(\\mathbf {x} _ {k} \\mid \\mathbf {x} _ {k + 1}, \\mathbf {y}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 479, + 482, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "requires the knowledge of conditional scores $\\nabla_{\\mathbf{x}_k}\\log p(\\mathbf{x}_k|\\mathbf{y})$ which can be expressed as $\\nabla_{\\mathbf{x}_k}\\log p(\\mathbf{x}_k|\\mathbf{y}) = \\nabla_{\\mathbf{x}_k}\\log p(\\mathbf{x}_k) + \\nabla_{\\mathbf{x}_k}\\log p(\\mathbf{y}|\\mathbf{x}_k)$ . The first term can be approximated by a score network $s_\\theta (\\mathbf{x}_k,k)$ [8, 21], and different methods have been proposed to estimate the second term $\\nabla_{\\mathbf{x}_k}\\log p(\\mathbf{y}|\\mathbf{x}_k)$ [6, 19]. In each step the conditional sampling process in VP diffusion is [6, 19]", + "bbox": [ + 89, + 527, + 483, + 626 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {x} _ {k - 1} \\leftarrow \\frac {1}{\\sqrt {\\alpha_ {k}}} \\left(\\mathbf {x} _ {k} + (1 - \\alpha_ {k}) \\nabla_ {\\mathbf {x} _ {k}} \\log p (\\mathbf {x} _ {k} | \\mathbf {y})\\right) + \\tilde {\\sigma} _ {k} \\mathbf {z}, \\tag {6} \\\\ \\approx \\frac {1}{\\sqrt {\\alpha_ {k}}} (\\mathbf {x} _ {k} + (1 - \\alpha_ {k}) (s _ {\\theta} (\\mathbf {x} _ {k}, k) + \\nabla_ {\\mathbf {x} _ {k}} \\log p (\\mathbf {y} | \\mathbf {x} _ {k}))) + \\tilde {\\sigma} _ {k} \\mathbf {z}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 631, + 486, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{z} \\sim \\mathcal{N}(0, \\mathbf{I})$ and $\\tilde{\\sigma}_k^2$ is the approximated variance of the reverse posterior distribution.", + "bbox": [ + 89, + 696, + 483, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Meanwhile, [28] proposed to directly approximate the reverse conditional $p(\\mathbf{x}_k|\\mathbf{x}_{k + 1},\\mathbf{y})$ and propagate the mean in each step. Specifically, when $T\\to \\infty$ , the mean of $p(\\mathbf{x}_k|\\mathbf{x}_{k + 1},\\mathbf{y})$ in VP-diffusion is given by [28]", + "bbox": [ + 89, + 728, + 483, + 789 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} _ {k} \\left(\\mathbf {x} _ {k + 1}, \\mathbf {y}\\right) = \\mathbf {U} _ {k} \\mathbf {x} _ {k + 1} + \\mathbf {V} _ {k} \\mathbb {E} _ {p \\left(\\mathbf {x} _ {0} \\mid \\mathbf {y}\\right)} [ X _ {0} ], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 796, + 482, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 89, + 819, + 137, + 830 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {U} _ {k} := \\sqrt {\\alpha_ {k + 1}} \\left(\\left(1 - \\bar {\\alpha} _ {k}\\right) \\mathbf {I} + \\bar {\\alpha} _ {k} \\operatorname {C o v} _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ]\\right) \\\\ \\cdot \\left(\\left(1 - \\bar {\\alpha} _ {k + 1}\\right) \\mathbf {I} + \\bar {\\alpha} _ {k + 1} \\operatorname {C o v} _ {p \\left(\\mathbf {x} _ {0} \\mid \\mathbf {y}\\right)} [ X _ {0} ]\\right) ^ {- 1}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 839, + 444, + 877 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {V} _ {k} := \\sqrt {\\bar {\\alpha} _ {k}} \\big ((1 - \\alpha_ {k + 1}) \\mathbf {I} \\big) \\big ((1 - \\bar {\\alpha} _ {k + 1}) \\mathbf {I} + \\bar {\\alpha} _ {k + 1} \\operatorname {C o v} _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ] \\big) ^ {- 1}.\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 880, + 480, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here $\\alpha_{k} = 1 - \\beta_{k}$ and $\\bar{\\alpha}_{k} = \\Pi_{i=0}^{k} \\alpha_{i}$ in VP-diffusion. The authors proved that by propagating the mean at each reverse step (i.e., $\\pmb{\\mu}_{T} \\rightarrow \\pmb{\\mu}_{T-1}(\\mathbf{x}_{T} = \\pmb{\\mu}_{T}, \\mathbf{y}) \\rightarrow \\dots \\rightarrow \\pmb{\\mu}_{0}(\\mathbf{x}_{1} = \\pmb{\\mu}_{1}, \\mathbf{y})$ ), the end point $\\pmb{\\mu}_{0}(\\mathbf{x}_{1} = \\pmb{\\mu}_{1}, \\mathbf{y})$ represents the MMSE estimator $\\mathbb{E}_{p(\\mathbf{x}_0|\\mathbf{y})}[X_0]$ when $\\bar{\\alpha}_{T} \\rightarrow 0$ . The proof of reverse mean propagation converging to the MMSE point [28] and its connection to conditional score are included in Appendix A.", + "bbox": [ + 511, + 90, + 905, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3 Connection between reverse mean and conditional score", + "text_level": 1, + "bbox": [ + 511, + 219, + 905, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The mean derived in (7) can be viewed as an approximation of the mean in (6). Specifically, when $T \\to \\infty$ , $p(\\mathbf{x}_k|\\mathbf{y})$ can be approximately expressed in the form of a Gaussian distribution. We can show that (see details in Appendix A)", + "bbox": [ + 511, + 256, + 905, + 314 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {1}{\\sqrt {\\alpha_ {k}}} (\\mathbf {x} _ {k} + (1 - \\alpha_ {k}) \\nabla_ {\\mathbf {x} _ {k}} \\log p (\\mathbf {x} _ {k} | \\mathbf {y})) = \\mathbf {U} _ {k} \\mathbf {x} _ {k + 1} + \\mathbf {V} _ {k} \\mathbb {E} _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ].\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 319, + 901, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theoretically, when $\\tilde{\\sigma}_k = 0$ in each step (6), the endpoint of the reverse process converges to the MMSE estimator, which achieves the best possible performance on distortion. When $\\tilde{\\sigma}_k$ is the true posterior variance, the reverse process samples from the posterior distribution $p(\\mathbf{x}_{0:T}|\\mathbf{y})$ , which results in a reconstruction with perfect perception measured in conditional distribution. The above observations motivate us to bridge the two extreme points with the score-based generative model by tuning the scale of $\\sigma_k$ in the reverse process. It remains to be answered whether the bridging can provide us with an optimal and flexible estimator to traverse along the distortion-perception plane.", + "bbox": [ + 511, + 362, + 906, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Traversing DP Tradeoff with Scaled Reverse Diffusion", + "text_level": 1, + "bbox": [ + 511, + 555, + 905, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first propose a novel reverse sampling method based on a variance-scaled version of the joint inference distribution (5) and theoretically derive the mean and variance of the marginal distribution. Then, we prove that the resulting end distribution of the variance-scaled reverse sampling provides an optimal DP tradeoff with respect to MSE and Wasserstein-2 distance.", + "bbox": [ + 511, + 598, + 905, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 Reverse mean and variance", + "text_level": 1, + "bbox": [ + 511, + 712, + 769, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider the joint inference distribution with true mean but scaled variance, i.e.,", + "bbox": [ + 511, + 734, + 905, + 765 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\lambda} \\left(\\mathbf {x} _ {k} \\mid \\mathbf {x} _ {k + 1}, \\mathbf {y}\\right) := \\mathcal {N} \\left(\\boldsymbol {\\mu} _ {k} \\left(\\mathbf {x} _ {k + 1}, \\mathbf {y}\\right), \\lambda \\mathbf {C} _ {k}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 772, + 906, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the expectation $\\pmb{\\mu}_k(\\mathbf{x}_{k + 1},\\mathbf{y})$ is given in (7), and the true covariance $\\mathbf{C}_k$ is given by [28]", + "bbox": [ + 511, + 804, + 905, + 834 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {C} _ {k} := \\frac {\\beta_ {k + 1}}{1 - \\beta_ {k + 1}} \\left((1 - \\bar {\\alpha} _ {k}) \\mathbf {I} + \\bar {\\alpha} _ {k} \\operatorname {C o v} _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ]\\right) \\\\ \\cdot \\left(\\left(\\frac {\\beta_ {k + 1}}{1 - \\beta_ {k + 1}} + 1 - \\bar {\\alpha} _ {k}\\right) \\mathbf {I} + \\bar {\\alpha} _ {k} \\operatorname {C o v} _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ]\\right) ^ {- 1}. \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 839, + 905, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2379", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We have the following theorem to characterize the marginal distribution given by the variance-scaled version of the joint inference distribution.", + "bbox": [ + 89, + 90, + 483, + 135 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1. For VP-diffusion and $0 \\leq \\lambda \\leq 1$ , consider the joint inference distribution given by", + "bbox": [ + 88, + 143, + 483, + 175 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\lambda} \\left(\\mathbf {x} _ {0: T} \\mid \\mathbf {y}\\right) = p _ {\\lambda} \\left(\\mathbf {x} _ {T} \\mid y\\right) \\prod_ {k = 0} ^ {T - 1} p _ {\\lambda} \\left(\\mathbf {x} _ {k} \\mid \\mathbf {x} _ {k + 1}, \\mathbf {y}\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 183, + 483, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p_{\\lambda}(\\mathbf{x}_T|\\mathbf{y}) = \\mathcal{N}(0,\\mathbf{I})$ and $p_{\\lambda}(\\mathbf{x}_k|\\mathbf{x}_{k + 1},\\mathbf{y})$ is given by (8). Then, the corresponding margin has the distribution $p_{\\lambda}(\\mathbf{x}_k|\\mathbf{y}) = \\mathcal{N}(\\boldsymbol {\\mu}_k^\\lambda ,\\boldsymbol{\\Sigma}_k^\\lambda)$ where", + "bbox": [ + 89, + 233, + 483, + 277 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {\\mu} _ {k} ^ {\\lambda} = \\left(\\mathbf {U} _ {k} \\left(\\mathbf {U} _ {k + 1} (\\dots (\\mathbf {U} _ {T - 2} \\mathbf {V} _ {T - 1} + \\mathbf {V} _ {T - 2}) \\dots) + \\mathbf {V} _ {k + 1}\\right) + \\mathbf {V} _ {k}\\right) \\\\ = \\sqrt {\\bar {\\alpha} _ {k}} (1 - \\alpha_ {k + 1} \\cdot \\cdot \\cdot \\alpha_ {T}) \\left((1 - \\bar {\\alpha} _ {T}) \\mathbf {I} + \\bar {\\alpha} _ {T} C o v _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ]\\right) ^ {- 1} \\mathbb {E} _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ], \\\\ \\boldsymbol {\\Sigma} _ {k} ^ {\\lambda} = \\boldsymbol {\\Sigma} _ {k} \\Big (\\lambda \\mathbf {I} + (1 - \\lambda) \\alpha_ {k + 1} \\alpha_ {k + 2} \\dots \\alpha_ {T} \\boldsymbol {\\Sigma} _ {T - 1} ^ {- 1} \\boldsymbol {\\Sigma} _ {k} \\Big), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 285, + 486, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{\\Sigma}_k$ is the covariance of $p(\\mathbf{x}_k|\\mathbf{y})$ in VP diffusion. In particular, when $\\bar{\\alpha}_T\\rightarrow 0$ , the variance of $p_{\\lambda}(\\mathbf{x}_0|\\mathbf{y})$ is", + "bbox": [ + 89, + 361, + 483, + 392 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\pmb {\\Sigma} _ {0} ^ {\\lambda} = \\pmb {\\Sigma} _ {0} \\bigg (\\lambda \\mathbf {I} + (1 - \\lambda) \\bar {\\alpha} _ {T} \\pmb {\\Sigma} _ {T - 1} ^ {- 1} \\pmb {\\Sigma} _ {0} \\bigg) \\rightarrow \\lambda C o v _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ],\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 400, + 486, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and the mean is", + "bbox": [ + 89, + 433, + 200, + 446 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mu_ {0} ^ {\\lambda} = \\sqrt {\\bar {\\alpha} _ {0}} \\big ((1 - \\bar {\\alpha} _ {T}) \\mathbf {I} \\big) \\big ((1 - \\bar {\\alpha} _ {T}) \\mathbf {I} + \\bar {\\alpha} _ {T} C o v _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ] \\big) ^ {- 1} \\\\ \\cdot \\mathbb {E} _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ] \\rightarrow \\mathbb {E} _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ]. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 454, + 488, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proof. See the details in Appendix B.", + "bbox": [ + 89, + 502, + 341, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The above theorem shows that the proposed variance-scaled reverse diffusion process will lead to a scaled marginal distribution at each step, which finally results in the posterior mean and scaled posterior variance of the original data distribution at the end of the reverse process. This seemingly intuitive result actually provides the optimal solution to the conditional distortion-perception tradeoff for multivariate Gaussian distribution.", + "bbox": [ + 89, + 527, + 483, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 DP tradeoff for conditional Gaussian case", + "text_level": 1, + "bbox": [ + 89, + 656, + 455, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Consider the $d$ -dimensional source $X \\in \\mathbb{R}^d$ with distribution $p(\\mathbf{x})$ . Let $Y$ be the noisy version of the source data. Denote the conditional expectation and variance given an observation $\\mathbf{y}$ as $\\boldsymbol{\\mu}_{\\mathbf{y}} = \\mathbb{E}_{p(\\mathbf{x}|\\mathbf{y})}[X]$ and $\\boldsymbol{\\Sigma}_{\\mathbf{y}} = \\mathrm{Cov}_{p(\\mathbf{x}|\\mathbf{y})}[X]$ , respectively. The goal of signal restoration is to find a good reconstruction $\\hat{X}$ based on the observed $Y$ .", + "bbox": [ + 89, + 679, + 483, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From the optimization perspective, given a specific observation $Y = \\check{\\mathbf{y}}$ , we can define the conditional distortion-perception function as", + "bbox": [ + 89, + 770, + 483, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} D (P) = \\min _ {p _ {\\hat {X} | Y} (\\hat {\\mathbf {x}} | \\check {\\mathbf {y}})} \\mathbb {E} _ {p _ {X, \\hat {X} | Y} (\\mathbf {x}, \\hat {\\mathbf {x}} | \\check {\\mathbf {y}})} [ \\| X - \\hat {X} \\| _ {2} ^ {2} ] \\\\ \\begin{array}{l} \\text {s . t .} d (p _ {X | Y} (\\mathbf {x} | \\breve {\\mathbf {y}}), p _ {\\hat {X} | Y} (\\hat {\\mathbf {x}} | \\breve {\\mathbf {y}})) \\leq P, \\end{array} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 823, + 452, + 875 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $X, Y$ and $\\hat{X}$ form a Markov chain.", + "bbox": [ + 89, + 883, + 366, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 2. Consider the source $X$ and an observation $\\tilde{\\mathbf{y}}$ that satisfy $p_{X|Y}(\\mathbf{x}|\\tilde{\\mathbf{y}}) \\sim \\mathcal{N}(\\boldsymbol{\\mu}_{\\tilde{\\mathbf{y}}}, \\boldsymbol{\\Sigma}_{\\tilde{\\mathbf{y}}})$ . The optimal distortion-perception tradeoff given $\\tilde{\\mathbf{y}}$ with MSE and Wasserstein-2 distance is", + "bbox": [ + 511, + 90, + 906, + 150 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD (P) = \\left\\{ \\begin{array}{l} \\operatorname {T r} \\left(\\boldsymbol {\\Sigma} _ {\\tilde {\\mathbf {y}}}\\right) + \\left(\\sqrt {\\operatorname {T r} \\left(\\boldsymbol {\\Sigma} _ {\\tilde {\\mathbf {y}}}\\right)} - P\\right) ^ {2}, f o r P \\leq \\sqrt {\\operatorname {T r} \\left(\\boldsymbol {\\Sigma} _ {\\tilde {\\mathbf {y}}}\\right)} \\\\ \\operatorname {T r} \\left(\\boldsymbol {\\Sigma} _ {\\tilde {\\mathbf {y}}}\\right), f o r P > \\sqrt {\\operatorname {T r} \\left(\\boldsymbol {\\Sigma} _ {\\tilde {\\mathbf {y}}}\\right)}. \\end{array} \\right. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 157, + 906, + 214 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Meanwhile, the optimal tradeoff can be achieved by sampling from the joint inference distribution (10) in Theorem 1 and traversing different $0 \\leq \\lambda \\leq 1$ .", + "bbox": [ + 511, + 223, + 906, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proof. See details in Appendix C.", + "bbox": [ + 513, + 276, + 741, + 292 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9cb66dac147096be4a2598fb33943fa5a6d200cee83c024859798864e2583017.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 888, + 277, + 906, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 2 explicitly characterizes the optimal DP tradeoff for conditional Gaussian distribution. When $P = 0$ (i.e., perfect perceptual quality is expected), the best achievable MSE is $2\\mathrm{Tr}(\\Sigma_{\\bar{\\mathbf{y}}})$ . As $P$ increases, the optimal MSE gradually converges to the MMSE value $\\mathrm{Tr}(\\Sigma_{\\bar{\\mathbf{y}}})$ . The proof in Appendix C also shows that the variance-scaled reverse diffusion process can attain the entire optimal DP curve for conditional Gaussian distribution. Although the actual data distribution may not follow the conditional Gaussian distribution (e.g., mixture Gaussian, 2D distribution like S-curve, and image dataset), the forthcoming experiments will show that the proposed variance-scaled reverse sampling can effectively and flexibly traverse the DP tradeoff on general datasets using a single score network.", + "bbox": [ + 511, + 301, + 906, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3 Variance-scaled reverse sampling process", + "text_level": 1, + "bbox": [ + 511, + 521, + 875, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this subsection, we discuss how to perform variance-scaled reverse sampling by estimating the conditional score. With a single pre-trained score network, we can traverse different DP tradeoffs for general inverse problems.", + "bbox": [ + 511, + 545, + 905, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recall that the proposed variance-scaled reverse sampling process is given by $p_{\\lambda}(\\mathbf{x}_{0:T}|\\mathbf{y}) = p_{\\lambda}(\\mathbf{x}_T|y)\\prod_{k = 0}^{T - 1}p_{\\lambda}(\\mathbf{x}_k|\\mathbf{x}_{k + 1},\\mathbf{y})$ where for VP diffusion, we have the posteriors $p_{\\lambda}(\\mathbf{x}_{k - 1}|\\mathbf{x}_k,\\mathbf{y})\\coloneqq \\mathcal{N}\\big(\\boldsymbol{\\mu}_{k - 1}(\\mathbf{x}_k,\\mathbf{y}),\\lambda \\mathbf{C}_{k - 1}\\big)$ . The sampling process in each step can be approximately expressed as", + "bbox": [ + 511, + 604, + 905, + 693 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {k - 1} = \\frac {1}{\\sqrt {\\alpha_ {k}}} \\left(\\mathbf {x} _ {k} + (1 - \\alpha_ {k}) \\nabla_ {\\mathbf {x} _ {k}} \\log p \\left(\\mathbf {x} _ {k} \\mid \\mathbf {y}\\right)\\right) + \\lambda \\mathbf {C} _ {k} \\mathbf {z}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 699, + 905, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{z} \\sim \\mathcal{N}(0, \\mathbf{I})$ . To draw a sample $\\mathbf{x}_{k-1}$ according to (12), we need to compute the conditional score $\\nabla_{\\mathbf{x}_k} \\log p(\\mathbf{x}_k | \\mathbf{y})$ . Specifically, the conditional scores can be decomposed as $\\nabla_{\\mathbf{x}_k} p(\\mathbf{x}_k | \\mathbf{y}) = \\nabla_{\\mathbf{x}_k} p(\\mathbf{x}_k) + \\nabla_{\\mathbf{x}_k} p(\\mathbf{y} | \\mathbf{x}_k)$ by Bayes' rule, where the first term can be approximated by a score network $s_\\theta(\\mathbf{x}_k, k)$ [8, 21]. To deal with the second term, Denoising Posterior Sampling (DPS) [6] proposes to approximate $p(\\mathbf{y} | \\mathbf{x}_k) = \\mathbb{E}_{p(\\mathbf{x}_0 | \\mathbf{x}_k)}[p(\\mathbf{y} | \\mathbf{x}_0)]$ with $p(\\mathbf{y} | \\hat{\\mathbf{x}}_0(\\mathbf{x}_k))$ , where", + "bbox": [ + 511, + 736, + 905, + 864 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} _ {0} \\left(\\mathbf {x} _ {k}\\right) = \\mathbb {E} _ {p \\left(\\mathbf {x} _ {0} \\mid \\mathbf {x} _ {k}\\right)} \\left[ X _ {0} \\right] = \\frac {1}{\\alpha_ {k}} \\left(\\mathbf {x} _ {k} + \\left(1 - \\bar {\\alpha} _ {k}\\right) \\nabla_ {\\mathbf {x} _ {k}} p \\left(\\mathbf {x} _ {k}\\right)\\right) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 869, + 906, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2380", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "by Tweedie's formula. The approximation error here is upper-bounded by the Jensen gap [17]. Note that $p(\\mathbf{y}|\\hat{\\mathbf{x}}_0(\\mathbf{x}_k))$ is analytically tractable when the measurement distribution is given. Specifically, when the measurement nosie is $\\mathcal{N}(0,\\sigma_n^2\\mathbf{I})$ , the second term in the conditional score takes the form of $\\nabla_{\\mathbf{x}_k}\\log p(\\mathbf{y}|\\mathbf{x}_k) = \\frac{1}{\\sigma_n^2}\\nabla_{\\mathbf{x}_k}||\\mathbf{y}-\\mathcal{A}(\\hat{\\mathbf{x}}_0(\\mathbf{x}_k))||_2^2$ . Thus, the conditional scores can be approximated as $\\nabla_{\\mathbf{x}_k}\\log p(\\mathbf{x}_k|\\mathbf{y}) = s_\\theta (\\mathbf{x}_k,k) + \\frac{1}{\\sigma_n^2}\\nabla_{\\mathbf{x}_k}||\\mathbf{y}-\\mathcal{A}(\\hat{\\mathbf{x}}_0(\\mathbf{x}_k))||_2^2$ .", + "bbox": [ + 89, + 90, + 483, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the following experiments, we adopt the DPS framework to approximate the conditional score and sample from (12), which is sketched as Algorithm 1.", + "bbox": [ + 89, + 234, + 483, + 280 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Variance-Scaled Reverse Sampling with DPS" + ], + "code_body": "Require: $T,\\mathbf{y},\\lambda ,\\{\\tilde{\\sigma}_k\\}_{k = 0}^T,\\{\\zeta_{k,\\lambda}\\}_{k = 0}^T$ $\\mathbf{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ \nfor $k = T - 1$ to 1 do do $\\hat{s}\\gets s_{\\theta}(\\mathbf{x}_k,k)$ $\\hat{\\mathbf{x}}_0\\gets \\frac{1}{\\sqrt{\\bar{\\alpha}_k}} (\\mathbf{x}_k + (1 - \\bar{\\alpha}_k)\\hat{s})$ $\\hat{c} (\\hat{\\mathbf{x}}_0) = -\\frac{\\sqrt{\\alpha_k}}{1 - \\alpha_k}\\nabla_{\\mathbf{x}_k}||y - \\mathcal{A}(\\hat{\\mathbf{x}}_0(\\mathbf{x}_k))||_2^2$ $\\mathbf{z}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ $\\mathbf{x}_{k - 1}\\leftarrow \\frac{1}{\\sqrt{\\alpha_k}}\\bigl (\\mathbf{x}_k + (1 - \\alpha_k)\\bigl (\\hat{s} +\\zeta_{k,\\lambda}\\hat{c} (\\hat{\\mathbf{x}}_0)\\bigr)\\bigr) + \\lambda \\tilde{\\sigma}_k\\mathbf{z}$ \nend for \nreturn $\\mathbf{x}_0$", + "bbox": [ + 91, + 314, + 483, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Algorithm 1, $\\tilde{\\sigma}_k$ is set to $\\beta_{k}$ [8, 28], and $\\zeta_{k,\\lambda}$ is a hyperparameter to control the weight of the conditional score, and may differ for different $\\lambda$ . The heuristic choices of $\\zeta_{k,\\lambda}$ and the rationale for fine-tuning have been included in Appendix E.", + "bbox": [ + 89, + 498, + 483, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark 1. In this paper, we adopt the conditional score approximation proposed in [6] for illustration. To implement the variance-scaled reverse sampling process (12), one can employ any approach to directly estimate the conditional score [19] or estimate the posterior mean as a whole [28] for general inverse problems. For example, Pseudoinverse-guided Diffusion Models (ΠGDM) [19] approximated the unconditional posteriors $p(\\mathbf{x}_0|\\mathbf{x}_k)$ with Gaussian, i.e., $p(\\mathbf{x}_0|\\mathbf{x}_k) \\approx \\mathcal{N}(\\hat{\\mathbf{x}}_0,r_k^2\\mathbf{I})$ , where $\\hat{\\mathbf{x}}_0$ is given by (13), and $r_k^2$ is the hyperparameter. Thus, the score term $\\nabla_{\\mathbf{x}_k}p(\\mathbf{y}|\\mathbf{x}_k)$ can be analytically expressed by the pseudoinverse of the measurement model. In [28], the posterior mean is estimated as a whole. Specifically, we have $\\mu_{k - 1}(\\mathbf{x}_k,\\mathbf{y}) = \\mathbf{U}_{k - 1}\\mathbf{x}_k + \\mathbf{V}_{k - 1}\\mathbb{E}_{p(\\mathbf{x}_0|\\mathbf{y})}[X_0]$ . Since the expression involves $\\mathbb{E}_{p(\\mathbf{x}_0|\\mathbf{y})}[X_0]$ and $\\mathrm{Cov}_{p(\\mathbf{x}_0|\\mathbf{y})}[X_0]$ , it can not be computed directly. Instead, the authors approximate the joint conditional posterior with variational Gaussian distributions and apply natural gradient descent to perform the sampling. Note that the gradient calculation also involves the approximation of the conditional score $\\nabla_{\\mathbf{x}_k}\\log p(\\mathbf{y}|\\mathbf{x}_k)$ as in DPS [6].", + "bbox": [ + 89, + 583, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e1ee14930088bcc594c2b6a18eae8fb0a7d44a986c5ff556883074c675e8c058.jpg", + "image_caption": [ + "(a) $p(x_0)$", + "Figure 1. An example of mixture Gaussian, the noisy observation, and the conditional distribution given an observation $y = -0.6$" + ], + "image_footnote": [], + "bbox": [ + 521, + 90, + 640, + 170 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c100d554646e05dc0617d22319c1555cd6b1c6d7a8be2e6e6423ee9d3eb2a635.jpg", + "image_caption": [ + "(b) $p(y)$" + ], + "image_footnote": [], + "bbox": [ + 643, + 90, + 759, + 170 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6cc451c67db869511cf8be3cf9c3918fa38aa66220bb4fe8d9fa51b5a55b464c.jpg", + "image_caption": [ + "(c) $p(x|y = -0.6)$" + ], + "image_footnote": [], + "bbox": [ + 763, + 90, + 895, + 169 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4 An illustrative example", + "text_level": 1, + "bbox": [ + 511, + 233, + 736, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Consider the mixture Gaussian distribution $X_0 \\sim p(x_0)$ with two components, where $p(x_0) = w_1 \\mathcal{N}(\\mu_1, \\sigma_1^2) + w_2 \\mathcal{N}(\\mu_2, \\sigma_2^2)$ . The noisy observation is obtained by $Y = aX_0 + \\sigma_0 \\epsilon$ , where $\\epsilon \\sim \\mathcal{N}(0, 1)$ , i.e., $p(y|x_0) = \\mathcal{N}(ax_0, \\sigma_0^2)$ . The MMSE estimator and posterior distribution can be theoretically derived (see Appendix D for details). Fig. 1 shows an example of mixture Gaussian distribution and the conditional distribution given an observation $y = -0.6$ .", + "bbox": [ + 511, + 256, + 903, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given an observation $y$ and a chosen $\\lambda$ , we iteratively perform the variance scaled reverse process (12) for $k = T, \\dots, 1$ , where the variance is approximated by $\\mathbf{C}_k = \\beta_k$ as the conventional VP-diffusion. Starting from a random sample $x_T \\sim \\mathcal{N}(0,1)$ , Fig. 2 shows the trajectories $x_T \\to \\dots x_k \\to \\dots x_0$ of multiple reconstructions for each $\\lambda$ . We can observe that for $\\lambda = 0$ , given an initial $x_T$ , the trajectories are deterministic and converge to the MMSE point. This phenomenon collaborates with results in [28]. When $\\lambda$ increases, the generated trajectories follow the form of the posterior distribution and show more stochasticity. When $\\lambda = 1$ , the reconstruction distribution matches the ground truth posterior.", + "bbox": [ + 511, + 377, + 905, + 573 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The resulting DP tradeoffs are shown in Fig. 3. It can be observed that as $\\lambda$ increases, the divergence between the true posterior distribution and reconstruction distribution continuously decreases, with an increase of MSE value. The overall curve is convex, and both Wasserstein-2 and KL-divergence come to zero. Meanwhile, the MSE achieved at $\\lambda = 1$ is approximately two times the MSE at $\\lambda = 0$ for both Wasserstein-2 and KL-divergence cases, coinciding with the result in Theorem 2.", + "bbox": [ + 511, + 573, + 905, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 722, + 728, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we conduct a series of experiments on two-dimensional datasets and the FFHQ dataset [9]. It is demonstrated that the variance-scaled reverse sampling shows great flexibility and effectiveness in traversing DP tradeoff compared to the GAN-based approach and other denoising methods.", + "bbox": [ + 511, + 747, + 905, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Two-dimensional points example", + "text_level": 1, + "bbox": [ + 511, + 847, + 808, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, we evaluate the validity and effectiveness of our approach on two-dimensional datasets, including pinwheel, S-", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2381", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/44fd470180d9b8a7fa4c35dafe0a9fd4baf1afd510d8e21b8a790f4d5ea4cb47.jpg", + "image_caption": [ + "Figure 2. Trajectories $x_{T} \\to \\dots x_{k} \\to \\dots \\to x_{0}$ of different reconstructions for $\\lambda = 0, 0.3, 0.8$ and 1. The initial $x_{T}$ is $\\mathcal{N}(0,1)$ ." + ], + "image_footnote": [], + "bbox": [ + 96, + 71, + 903, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/67e673cf40bd32b441cd2a9307d4404578d99a9f7a39d23d9448828cfbf0940e.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 94, + 229, + 285, + 345 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6643881d2c6b59afaab83c3244091818b1463bbc20a10993a504b8116e0820f6.jpg", + "image_caption": [ + "(b)", + "Figure 3. Distortion-perception tradeoff traversed by variance-scaled reverse sampling given different $\\lambda$ 's. (a) tradeoff between Wasserstein-2 distance and MSE; (b) tradeoff between KL-divergence and MSE." + ], + "image_footnote": [], + "bbox": [ + 297, + 229, + 483, + 345 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "curve, and moon-shape data distributions. Taking the pinwheel dataset as an example, the first row of Fig. 4 shows the distribution of pinwheel data points $X_0$ and the noisy observation $Y$ , where $Y = aX + N$ for $N \\sim \\mathcal{N}(0, \\sigma_n^2 \\mathbf{I})$ .", + "bbox": [ + 89, + 440, + 482, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For each dataset, we train a conventional score network to approximate $p(\\mathbf{x}_0)$ on the original distribution. Then, we perform the variance-scaled reverse diffusion process (12) on the noisy observation. As a comparison, we adopt a GAN-based DP-traversing approach PSCGAN [15]. This scheme introduces a penalty of the posterior expectation to the training of the conditional GAN. The authors proposed two strategies to navigate the DP tradeoff during the inference time: (1) PSCGAN- $N$ : This method involves sampling $N$ instances from the generator and then averaging them. As $N$ increases, the averaged image is closer to the conditional expectation, leading to a smaller distortion but larger perception loss. (2) PSCGAN- $z$ : It varies the standard deviation $\\sigma_z$ of the noise injected into the generator to control the stochasticity. For $0 \\leq \\sigma_z \\leq 1$ , a larger $\\sigma_z$ results in a better perceptual quality and higher MSE. Note that the training of PSCGAN relies on noisy data points $Y$ . Thus, the model needs to be retrained for different measurement scenarios and different levels of $\\sigma_n^2$ .", + "bbox": [ + 89, + 503, + 483, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 4 shows the reconstruction of our variance-scaled reverse diffusion process for different $\\lambda$ 's, as well as the reconstruction of PSCGAN when setting $N = 16$ , $\\sigma_z = 1$ . It can be observed that when $\\lambda = 0$ , our sampling process leads to a more concentrated reconstruction, and gradually approaches the true distribution as $\\lambda$ increases.", + "bbox": [ + 89, + 792, + 482, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 5 compares the numerical DP tradeoffs achieved", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7367742c11f73a719eb2cf15c2f2a74b107fabb9b8400d1bf8d20b7f86288a4d.jpg", + "image_caption": [ + "Original distributions" + ], + "image_footnote": [], + "bbox": [ + 598, + 229, + 707, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9fcc744ec54d2fb0e6ea48623cd5ecd146a42bc0f1b99226f7ef6647329c8239.jpg", + "image_caption": [ + "Noisy observations" + ], + "image_footnote": [], + "bbox": [ + 718, + 229, + 823, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0c0f5d0e3cc24a89e68c22f69f79b27cd4f97326eddeb6687fb3defc20da2582.jpg", + "image_caption": [ + "(a) $\\lambda = 0$" + ], + "image_footnote": [], + "bbox": [ + 537, + 308, + 651, + 373 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7c594d082776a16799d3443f4016f40ca0bbce69bdc87eede7ac0d11729a4b53.jpg", + "image_caption": [ + "(b) $\\lambda = 0.3$" + ], + "image_footnote": [], + "bbox": [ + 653, + 309, + 766, + 375 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4ebdcc77cb63da8a32d060e277b5b076206a1c265e66b44ad2a9dd7cfac6f8dc.jpg", + "image_caption": [ + "(c) $\\lambda = 0.5$" + ], + "image_footnote": [], + "bbox": [ + 769, + 309, + 883, + 373 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/39324c923ce22887d83fb8bc80d0dc674cd9f747af6f949d69b1528ded7040d6.jpg", + "image_caption": [ + "(d) $\\lambda = 0.8$", + "Figure 4. Experiments on a two-dimensional dataset. The first row illustrates the original distribution (left) of pinwheel and the noisy observation $Y$ , where $Y = aX + N$ for $N \\sim \\mathcal{N}(0, \\sigma_n^2\\mathbf{I})$ (right). The second and third row shows the reconstructions: (a)-(e) variance-scaled reverse diffusion process with different $\\lambda$ 's; (f) PSCGAN with $N = 16$ , $\\sigma_z = 1$ ." + ], + "image_footnote": [], + "bbox": [ + 537, + 388, + 651, + 453 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4d2d216814d00230e7e35e510af8de5e397cb84a79e82ffbce262d9f516fc56d.jpg", + "image_caption": [ + "(e) $\\lambda = 1$" + ], + "image_footnote": [], + "bbox": [ + 653, + 388, + 767, + 453 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f66f08aa9b98c16df988e8d174f09edf0847f6ca9e9ec251ad974f7d7c136c1c.jpg", + "image_caption": [ + "(f) PSCGAN $N = 16, \\sigma_z = 1$" + ], + "image_footnote": [], + "bbox": [ + 769, + 388, + 883, + 453 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "by our method versus PSCGAN. Our score-based method achieves a much larger range of tradeoffs compared to the GAN-based approach, demonstrating superior performance. Moreover, unlike PSCGAN, our method requires only a single score network to handle varying measurements and noise levels, offering greater flexibility. Additional experimental results on other distributions are provided in Appendix E.", + "bbox": [ + 511, + 577, + 906, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 FFHQ dataset", + "text_level": 1, + "bbox": [ + 511, + 709, + 668, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this subsection, we evaluate our method on the FFHQ $256 \\times 256$ dataset [9]. For different measurement scenarios, including Gaussian blur and super-resolution, we use the same pre-trained score-based model from [6], which was trained from scratch using 49k training data for 1M steps. We consider measurement scenarios with noise levels more severe than those in [6, 28, 29]. Specifically, for Gaussian blur, the images are blurred by a Gaussian blur kernel with size $61 \\times 61$ and a standard deviation of 3.0. For super-resolution, the images are downsampled by a factor of 8.", + "bbox": [ + 511, + 733, + 906, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method against two benchmarks: (1)", + "bbox": [ + 532, + 885, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2382", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e3ff47821395411203f0fcd2c7141da420328371aa43a9658e35d5f360146ce0.jpg", + "image_caption": [ + "Figure 5. DP tradeoff on pinwheel dataset traversed by our variance-scaled reverse diffusion process and PSCGAN." + ], + "image_footnote": [], + "bbox": [ + 147, + 90, + 426, + 258 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare the DP tradeoffs provided by our score-based method and PSCGAN [15] on the Gaussian deblur task. Note that PSCGAN is not designed for other measurement scenarios and requires retraining for different noise levels. (2) We evaluate the performance of DiffPIR [29] to see where it falls within the tradeoff spectrum for different inverse problems. DiffPIR is a plug-and-play image restoration method based on diffusion denoising implicit model (DDIM) [18]. It has shown competitive performance on many denoising problems with fewer diffusion steps.", + "bbox": [ + 89, + 311, + 483, + 464 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.1 Gaussian deblur task", + "text_level": 1, + "bbox": [ + 89, + 483, + 292, + 496 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 6 shows the DP tradeoff on the Gaussian deblur task. We test our variance-scaled reverse diffusion process, PSCGAN [15], and DiffPIR [29] on two different additive noise levels $\\sigma_{n} = 0.3$ and $\\sigma_{n} = 0.5$ . Note that our sampling method and DiffPIR rely on a pre-trained score network on the FFHQ dataset. For PSCGAN, we use the model trained on images with additive Gaussian noise of $\\sigma_{n} = 0.3$ .", + "bbox": [ + 89, + 507, + 482, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For noise level $\\sigma_{n} = 0.3$ , PSCGAN shows suboptimal and more limited tradeoffs compared to our method. To obtain a reconstruction with similar MSE achieved in our $\\lambda = 0.9$ case, PSCGAN requires averaging 64 sampled images. Its distortion degrades rapidly when fewer images are averaged. Our variance-scaled reverse diffusion sampling achieves both better fidelity and a broader range of DP tradeoffs. Moreover, when the models are tested on the higher noise level of $\\sigma_{n} = 0.5$ , the overall tradeoff shifts to the left. Our sampling method maintains effective tradeoff traversal. PSCGAN's performance (trained at $\\sigma_{n} = 0.3$ ) deteriorates significantly, showing much larger MSE and excessively high FID values.", + "bbox": [ + 89, + 613, + 482, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Significant visual improvements can be observed in reconstructed samples in Fig. 7. When $\\lambda = 0$ , the restored images maintain high fidelity and faithfully reconstruct the original images, though appear relatively blurry. As $\\lambda$ increases, the images become progressively sharper, revealing more fine details such as hair strands, eye corners, and back", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ground elements. When $\\lambda = 1$ , while the overall reconstruction appears very natural, subtle details deviate from the original images. In comparison, both PSCGAN and DiffPIR have relatively good perceptual quality but compromise fidelity on crucial details. Notable deviations appear in features like eyes and mouth shapes, ultimately altering the facial expression and overall style of faces. We also report more metrics (PSNR and LPIPS) of Gaussian deblur task with additive noise level $\\sigma_{n} = 0.3$ in Appendix E.", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2 Super-resolution task", + "text_level": 1, + "bbox": [ + 511, + 244, + 718, + 260 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Similar phenomena are observed in the super-resolution (SR) task. By adjusting different $\\lambda$ 's, we can traverse a wide range of points on the DP plane, allowing flexible selection of the most suitable point for specific application scenarios. Fig. 8 shows the numerical results and reconstruction samples given by our sampling process and DiffPIR. Note that DiffPIR performs close to our $\\lambda = 1$ case, and achieves a better FID score but larger distortion than our method. While DiffPIR demonstrates strong generative capabilities in both Gaussian deblur and super-resolution tasks with fewer diffusion steps [29], we examine more challenging degradation scenarios ( $\\sigma_{n} = 0.3$ and 0.5 for deblurring, and $8\\times$ downscaling for SR). Under these severe conditions, DiffPIR's reduced diffusion steps lead to fidelity issues, with generated details less faithful compared to our $\\lambda = 1$ case. Despite these differences, theoretical connections exist between DPS and DiffPIR [29]. We will consider the exploration of DDIM-style sampling for faster tradeoff traversal as future work.", + "bbox": [ + 511, + 268, + 906, + 554 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In summary, using a single pre-trained score network, the proposed sampling method can effectively traverse a more complete DP tradeoff and achieve better MSE than the benchmarks. Experimental details and more results are included in Appendix E.", + "bbox": [ + 511, + 556, + 905, + 632 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5 Conclusions", + "text_level": 1, + "bbox": [ + 511, + 643, + 650, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this paper, we proposed a variance-scaled reverse diffusion process to traverse the DP tradeoff for general inverse problems using a single score-based model. By tuning a parameter that controls the scale of the reverse variance, we can navigate from the MMSE point to the whole posterior distribution, generating a complete DP curve. We proved that the proposed reverse sampling process serves as an optimal solution to the conditional DP tradeoff for multivariate Gaussian distribution. Meanwhile, we conducted experiments on the mixture Gaussian example, two-dimensional datasets, and the FFHQ dataset. Our results show that using a single pre-trained score network, we have achieved a more complete empirical DP tradeoff than the GAN-based method and other inverse problem solvers, demonstrating the effectiveness and flexibility of the proposed framework.", + "bbox": [ + 511, + 670, + 906, + 897 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2383", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d7cfc4e12adeead9a8e29f2db1ad310908f5f7589eba88437e7b4fad9aa99132.jpg", + "image_caption": [ + "Figure 6. DP tradeoff on FFHQ dataset. Note that the PSCGAN model in this figure is trained on images contaminated with additive Gaussian noise of $\\sigma_{n} = 0.3$ . It can be observed that when the noise level in inference time does match the trained model (e.g., when the noise level becomes $\\sigma_{n} = 0.5$ during the test time), PSCGAN fails to provide valid reconstructions. On the other hand, our variance-scaled reverse diffusion process and DiffPIR method rely on a pre-trained score network on the FFHQ dataset, which is not dedicated to any noisy measurement and is robust for different noise levels." + ], + "image_footnote": [], + "bbox": [ + 176, + 90, + 823, + 280 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4e93a5f90a08a1dc6e4f75922c8188002972962e1975fe474b97c7217abac7c2.jpg", + "image_caption": [ + "Figure 7. Samples from our sampling process, PSCGAN, and DiffPIR for Gaussian deblur task with noise level $\\sigma_{n} = 0.3$ ." + ], + "image_footnote": [], + "bbox": [ + 133, + 369, + 867, + 616 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9bba45efe3eafb02fe9d6e7a60271a4a90ed6978e77bc5d4e40c3f2da34c4c7a.jpg", + "image_caption": [ + "Figure 8. Samples from our sampling process, PSCGAN, and DiffPIR for super-resolution task with downsampling scale $\\times 8$ ." + ], + "image_footnote": [], + "bbox": [ + 135, + 659, + 867, + 845 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2384", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The work of Yuhan Wang and Ying-Jun Angela Zhang is supported in part by the General Research Fund (project number 14202421, 14214122, 14202723), Area of Excellence Scheme grant (project number AoE/E-601/22-R), and NSFC/RGC Collaborative Research Scheme (project number CRS_HKUST603/22), all from the Research Grants Council of Hong Kong. The work of Suzhi Bi is supported in part by the Natural Science Foundation of China under Project 62271325; in part by Guangdong Basic and Applied Basic Research Foundation under Project 2024B1515020089; in part by Shenzhen Science and Technology Program under Project 20220810142637001. The work of Xiaojun Yuan is supported in part by Sichuan Science and Technology Program under Grant 2024ZYD0036 and in part by National Key Laboratory of Wireless Communications Foundation under Grant IFN20230204.", + "bbox": [ + 89, + 114, + 485, + 357 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 375, + 187, + 388 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Brian D.O. Anderson. Reverse-time diffusion equation models. Stochastic Processes and their Applications, 12(3):313-326, 1982.", + "[2] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6228-6237, 2018.", + "[3] Yochai Blau and Tomer Michaeli. Rethinking lossy compression: The rate-distortion-perception tradeoff. In Proceedings of the 36th International Conference on Machine Learning (ICML), pages 675-685. PMLR, 2019.", + "[4] Valentin De Bortoli, James Thornton, Jeremy Heng, and Arnaud Doucet. Diffusion schrödinger bridge with applications to score-based generative modeling. In Advances in Neural Information Processing Systems (NIPS), 2021.", + "[5] Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. In Advances in Neural Information Processing Systems (NIPS), 2022.", + "[6] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In The Eleventh International Conference on Learning Representations (ICLR), 2023.", + "[7] Dror Freirich, Tomer Michaeli, and Ron Meir. A theory of the distortion-perception tradeoff in wasserstein space. In Advances in Neural Information Processing Systems (NIPS), 2021.", + "[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Proceedings of the 34th International Conference on Neural Information Processing Systems (NIPS), 2020.", + "[9] Tero Karras, Samuli Laine, and Timo Aila. A Style-Based Generator Architecture for Generative Adversarial Networks. IEEE Transactions on Pattern Analysis & Machine Intelligence, 43(12):4217-4228, 2021." + ], + "bbox": [ + 99, + 398, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. In Advances in Neural Information Processing Systems (NIPS), 2022.", + "[11] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017.", + "[12] Dong Liu, Haochen Zhang, and Zhiwei Xiong. On the classification-distortion-perception tradeoff. In Advances in Neural Information Processing Systems (NIPS). Curran Associates, Inc., 2019.", + "[13] Razvan V Marinescu, Daniel Moyer, and Polina Golland. Bayesian image reconstruction using deep generative models. arXiv preprint, 2021. [Online]. Available: https://arxiv.org/abs/2012.04567.", + "[14] Anush Krishna Moorthy and Alan Conrad Bovik. Blind image quality assessment: From natural scene statistics to perceptual quality. IEEE transactions on Image Processing, 20 (12):3350-3364, 2011.", + "[15] Guy Ohayon, Theo Adrai, Gregory Vaksman, Michael Elad, and Peyman Milanfar. High Perceptual Quality Image Denoising with a Posterior Sampling CGAN. In 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 1805-1813, Los Alamitos, CA, USA, 2021. IEEE Computer Society.", + "[16] Victor M. Panaretos and Yoav Zemel. An Invitation to Statistics in Wasserstein Space. Springer Cham, 2020.", + "[17] Slavko Simic. On a global upper bound for jensen's inequality. Journal of Mathematical Analysis and Applications, 343 (1):414-419, 2008.", + "[18] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations (ICLR), 2021.", + "[19] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In International Conference on Learning Representations (ICLR), 2023.", + "[20] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In Advances in Neural Information Processing Systems (NIPS). Curran Associates, Inc., 2019.", + "[21] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations (ICLR), 2021.", + "[22] Pascal Vincent. A connection between score matching and denoising autoencoders. *Neural Computation*, 23(7):1661-1674, 2011.", + "[23] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In The European Conference on Computer Vision Workshops (ECCVW), 2018.", + "[24] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. In" + ], + "bbox": [ + 514, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2385", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "The Eleventh International Conference on Learning Representations (ICLR), 2023.", + "[25] Jay Whang, Maurizio Delbracio, Hossein Talebi, Chitwan Sahara, Alexandros G. Dimakis, and Peyman Milanfar. Deblurring via Stochastic Refinement. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16272-16282, Los Alamitos, CA, USA, 2022.", + "[26] Xingyu Xu and Yuejie Chi. Provably robust score-based diffusion posterior sampling for plug-and-play image reconstruction. In The Thirty-eighth Annual Conference on Neural Information Processing Systems (NIPS), 2024.", + "[27] Danna Xue, Luis Herranz, Javier Vazquez Corral, and Yanning Zhang. Burst perception-distortion tradeoff: Analysis and evaluation. In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2023.", + "[28] Zhipeng Xue, Penghao Cai, Xiaojun Yuan, and Xiqi Gao. Score-based variational inference for inverse problems. arXiv preprint, 2024. [Online]. Available: https://arxiv.org/abs/2410.05646.", + "[29] Yuanzhi Zhu, Kai Zhang, Jingyun Liang, Jiezhang Cao, Bihan Wen, Radu Timofte, and Luc Van Gool. Denoising Diffusion Models for Plug-and-Play Image Restoration. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1219–1229, Los Alamitos, CA, USA, 2023. IEEE Computer Society." + ], + "bbox": [ + 91, + 90, + 482, + 470 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2386", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Traversing Distortion-Perception Tradeoff using a Single Score-Based Generative Model/dcdd686b-eeb1-4343-a7e6-20b5cd329198_model.json b/2025/Traversing Distortion-Perception Tradeoff using a Single Score-Based Generative Model/dcdd686b-eeb1-4343-a7e6-20b5cd329198_model.json new file mode 100644 index 0000000000000000000000000000000000000000..10675ebcc88bbf375c2e78a72834fb0069b14bad --- /dev/null +++ b/2025/Traversing Distortion-Perception Tradeoff using a Single Score-Based Generative Model/dcdd686b-eeb1-4343-a7e6-20b5cd329198_model.json @@ -0,0 +1,2365 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.149, + 0.13, + 0.851, + 0.175 + ], + "angle": 0, + "content": "Traversing Distortion-Perception Tradeoff using a Single Score-Based Generative Model" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.203, + 0.77, + 0.222 + ], + "angle": 0, + "content": "Yuhan Wang\\(^{1}\\), Suzhi Bi\\(^{2*}\\), Ying-Jun Angela Zhang\\(^{1}\\), Xiaojun Yuan\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.222, + 0.746, + 0.24 + ], + "angle": 0, + "content": "1The Chinese University of Hong Kong 2Shenzhen University" + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.24, + 0.734, + 0.257 + ], + "angle": 0, + "content": "3University of Electronic Science and Technology of China" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.26, + 0.798, + 0.275 + ], + "angle": 0, + "content": "{wy023, yjzhang}@ie.cuhk.edu.hk, bsz@szu.edu.cn, xjyuan@uestc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.341, + 0.486, + 0.659 + ], + "angle": 0, + "content": "The distortion-perception (DP) tradeoff reveals a fundamental conflict between distortion metrics (e.g., MSE and PSNR) and perceptual quality. Recent research has increasingly concentrated on evaluating denoising algorithms within the DP framework. However, existing algorithms either prioritize perceptual quality by sacrificing acceptable distortion, or focus on minimizing MSE for faithful restoration. When the goal shifts or noisy measurements vary, adapting to different points on the DP plane needs retraining or even re-designing the model. Inspired by recent advances in solving inverse problems using score-based generative models, we explore the potential of flexibly and optimally traversing DP tradeoffs using a single pre-trained score-based model. Specifically, we introduce a variance-scaled reverse diffusion process and theoretically characterize the marginal distribution. We then prove that the proposed sample process is an optimal solution to the DP tradeoff for conditional Gaussian distribution. Experimental results on two-dimensional and image datasets illustrate that a single score network can effectively and flexibly traverse the DP tradeoff for general denoising problems." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.687, + 0.233, + 0.703 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.712, + 0.485, + 0.879 + ], + "angle": 0, + "content": "In recent years, we have witnessed rapid progress in image restoration algorithms, especially deep learning-based implementations that have shown remarkable achievements in solving general inverse problems. How to evaluate the performance of emerging algorithms is a crucial yet complicated problem. Traditional full reference distortion metrics, such as Mean Square Error (MSE), focus on the pixel-level accuracy between the original image \\( X \\) and its reconstruction \\( \\hat{X} \\). On the other hand, perceptual quality, referring to the degree to which an image looks natural rather than algorithmically generated [14], is also an important mea" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.907, + 0.357 + ], + "angle": 0, + "content": "sure. It has been demonstrated that perceptual quality could be associated with the distance between the distributions of natural images \\( p_{X} \\) and generated images \\( p_{\\hat{X}} \\) [2, 3, 12]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.907, + 0.432 + ], + "angle": 0, + "content": "Improvements in distortion measures do not necessarily lead to enhancements in perceptual quality. In fact, [2] has demonstrated that there exists a tradeoff between distortion and perception, termed as distortion-perception (DP) trade-off. Mathematically, it can be modeled as" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.437, + 0.802, + 0.464 + ], + "angle": 0, + "content": "\\[\nD(P) = \\min_{p_{\\hat{X} |Y}}\\mathbb{E}[\\Delta (X,\\hat{X})],\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.653, + 0.468, + 0.787, + 0.485 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\text {s . t .} d (p _ {X}, p _ {\\tilde {X}}) \\leq P, \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.906, + 0.569 + ], + "angle": 0, + "content": "where \\(Y\\) is the degraded observation, \\(\\Delta : \\mathcal{X} \\times \\hat{\\mathcal{X}} \\to \\mathbb{R}^+\\) is a full reference distortion function (e.g., such as square-error), and \\(d(\\cdot, \\cdot)\\) is some divergence between probability distribution, such as Wasserstein distance [16] or Kulback-Leibler (KL) divergence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.765 + ], + "angle": 0, + "content": "Since the introduction of the DP tradeoff, more and more algorithm designs have focused on performance evaluation on the DP plane for specific denoising tasks, e.g., image deblurring [15, 25] and super-resolution [11, 13, 23]. However, most algorithms tend to seek better perceptual quality at the cost of distortion, or focus on optimizing MSE, a distortion measure, to ensure restoration fidelity. When the restoration objective shifts between distortion and perception, we need to retrain or even re-design the network, which can be computationally intensive. In many practical scenarios, it is significant to flexibly and effectively fulfill diverse task objectives or user demands with a single model at inference time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "There have been explorations into theories and algorithms to traverse the DP tradeoff in inference time [7, 15]. Specifically, in [15], the authors developed a framework based on conditional generative adversarial networks (CGAN) with an additional penalty on posterior expectation. By averaging different numbers of samples or adjusting the noise level injected into the generator, this CGAN-based framework can access various points along the tradeoff. However, using multiple image samples and averaging" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.238, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2377" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.333 + ], + "angle": 0, + "content": "them may not be efficient during inference, and the method is not provably optimal. Furthermore, the training process relies on pairs of clean and noisy data. Thus, retraining is required for different noise levels and measurements. In [7], the DP tradeoff is studied in Wasserstein space. Theoretical results show that estimators on the DP curve can be constructed by linearly combining two estimators at the extremes: an optimal MSE minimizer and a perfect perception sampler with minimum distortion. This idea is further applied in burst restoration [27]. In practice, a method that achieves a relatively low MSE, alongside another method that achieves good perceptual quality, can serve as approximations of these two extreme estimators for a specific task. Nonetheless, different models must be deployed to address various measurements, rendering this method less flexible to general inverse problems." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.338, + 0.485, + 0.578 + ], + "angle": 0, + "content": "Recent research has demonstrated the generative capability of score-based diffusion models [4, 8, 10, 18, 20, 21] in tackling general inverse problems using a single score network [5, 6, 19, 24, 26, 29]. Score-based diffusion models learn the prior distribution of the data \\(\\mathbf{x}\\) by training a score network to match the gradient of the logarithm density \\(\\nabla_{\\mathbf{x}}\\log p(\\mathbf{x})\\), referred to as the score. After observing a noisy measurement \\(\\mathbf{y}\\), sampling from posteriors involves approximating the conditional score \\(\\nabla_{\\mathbf{x}}\\log p(\\mathbf{x}|\\mathbf{y})\\). Theoretical insights provided in [28] also reveal the potential for recovering Minimum Mean Square Error (MMSE) estimation by propagating the mean of the reverse diffusion process. Inspired by these advancements, we explore the potential of traversing DP tradeoffs for different tasks using a single score network. Our main contributions are summarized below:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.584, + 0.483, + 0.735 + ], + "angle": 0, + "content": "- First, we propose a variance-scaled reverse diffusion process and theoretically characterize the marginal distributions produced by this novel reverse process. It is demonstrated that the mean of our proposed sampling converges towards the MMSE point while the marginal covariance is scaled by a scaling factor. By tuning a parameter that dictates the variance of reverse sampling, we can flexibly navigate from the MMSE point to the posterior distribution, where the two extremes are achieved by setting the parameter to zero and one, respectively." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.736, + 0.483, + 0.809 + ], + "angle": 0, + "content": "- Subsequently, we show that the reconstruction obtained from the variance-scaled reverse sampling represents the optimal solution to the conditional DP tradeoff for multivariate Gaussian distributions, assessed through MSE and Wasserstein-2 distance metrics." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.811, + 0.484, + 0.9 + ], + "angle": 0, + "content": "- Finally, we validate our methodology through experiments conducted on various two-dimensional datasets and a real-world image dataset. With a single pre-trained score network, we can traverse the DP tradeoff across different inverse problems with varying noise levels. The results indicate that the variance-scaled reverse diffusion" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.584, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.092, + 0.907, + 0.152 + ], + "angle": 0, + "content": "process achieves a more complete empirical DP tradeoff curve and better MSE than the GAN-based method and other inverse problem solvers, demonstrating the effectiveness and flexibility of the proposed framework." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.153, + 0.907, + 0.334 + ], + "angle": 0, + "content": "Notations: For a random variable \\(X\\) denoted by a capital letter, we use small letter \\(\\mathbf{x}\\) to denote its realizations, and use \\(p_{X}(\\mathbf{x})\\) to denote the distribution over its alphabet \\(\\mathcal{X}\\). When there is no ambiguity, the distribution can be abbreviated as \\(p(\\mathbf{x})\\). For a discrete sequence of random variables \\(X_0, X_1, \\dots, X_T\\), we abbreviate them as \\(\\{X_k\\}_{k=0}^T\\) and use \\(\\mathbf{x}_{0:T}\\) to denote their realizations. The expectation and conditional expectation of \\(X\\) given \\(Y = \\mathbf{y}\\) is \\(\\mathbb{E}_{p(\\mathbf{x})}[X]\\) and \\(\\mathbb{E}_{p(\\mathbf{x}|\\mathbf{y})}[X]\\). We use \\(\\mathrm{Cov}_{p(\\mathbf{x})}[X]\\) and \\(\\mathrm{Cov}_{p(\\mathbf{x}|\\mathbf{y})}[X]\\) to represent the covariance and conditional covariance of \\(X\\) given \\(Y = \\mathbf{y}\\). Matrices are denoted by uppercase boldface letters (e.g., \\(\\boldsymbol{\\Sigma}\\)). \\(\\mathrm{Tr}(\\boldsymbol{\\Sigma})\\) and \\(\\boldsymbol{\\Sigma}^{-1}\\) represent the trace and inverse." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.345, + 0.661, + 0.361 + ], + "angle": 0, + "content": "2 Backgrounds" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.37, + 0.787, + 0.385 + ], + "angle": 0, + "content": "2.1 Score-based generative model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.393, + 0.906, + 0.455 + ], + "angle": 0, + "content": "Score-based generative models or diffusion models define the generative process as the reverse of a diffusion process. Specifically, the diffusion process (i.e., forward process) is a Markov chain \\(\\{X_{k}\\}_{k = 0}^{T}\\), with joint distribution" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.46, + 0.907, + 0.502 + ], + "angle": 0, + "content": "\\[\np \\left(\\mathbf {x} _ {0: T}\\right) := p \\left(\\mathbf {x} _ {0}\\right) \\prod_ {k = 0} ^ {T - 1} p \\left(\\mathbf {x} _ {k + 1} \\mid \\mathbf {x} _ {k}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.508, + 0.906, + 0.599 + ], + "angle": 0, + "content": "which gradually diffuses the data \\(X_0 \\in \\mathbb{R}^d\\) with distribution \\(p(\\mathbf{x}_0) \\triangleq p_{\\mathrm{data}}\\). Variance preserving (VP) diffusion [8] adopts \\(p(\\mathbf{x}_{k+1}|\\mathbf{x}_k) \\triangleq \\mathcal{N}(\\mathbf{x}_{k+1}; \\sqrt{1 - \\beta_{k+1}}\\mathbf{x}_k, \\beta_{k+1}\\mathbf{I})\\) with monotonically increasing variance schedule \\(\\beta_T \\geq \\beta_{T-1} \\geq \\dots \\geq \\beta_0 = 0\\). When \\(T \\to \\infty\\), \\(\\mathbf{x}_T\\) converges to an isotropic Gaussian distribution \\(\\mathcal{N}(\\mathbf{0},\\mathbf{I})\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.644 + ], + "angle": 0, + "content": "To generate a sample following the data distribution \\( p(\\mathbf{x}_0) \\), we can start from the standard Gaussian \\( p_{\\theta}(\\mathbf{x}_T) \\coloneqq \\mathcal{N}(\\mathbf{0}, \\mathbf{I}) \\) and follow the reverse process" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.651, + 0.907, + 0.693 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\mathbf {x} _ {0: T}\\right) := p _ {\\theta} \\left(\\mathbf {x} _ {T}\\right) \\prod_ {k = 0} ^ {T - 1} p _ {\\theta} \\left(\\mathbf {x} _ {k} \\mid \\mathbf {x} _ {k + 1}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.698, + 0.906, + 0.759 + ], + "angle": 0, + "content": "with transitions \\(p_{\\theta}(\\mathbf{x}_k|\\mathbf{x}_{k + 1})\\sim \\mathcal{N}(\\mu_{\\theta}(\\mathbf{x}_k,k),\\boldsymbol {\\Sigma}_{\\theta}(\\mathbf{x}_k,k))\\) parameterized by \\(\\theta\\) . It will be shown later that the reverse sampling is closely related to the score function for each step, i.e., \\(\\nabla_{\\mathbf{x}_k}\\log p(\\mathbf{x}_k)\\) , for \\(k = 0,1,\\dots ,T\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.759, + 0.906, + 0.804 + ], + "angle": 0, + "content": "In continuous-time, we can describe the diffusion process \\((X_{t})_{t\\in [0,T_{c}]}\\), \\(X_{t}\\in \\mathbb{R}^{d}\\) with a stochastic differential equation (SDE) [21]" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.809, + 0.907, + 0.838 + ], + "angle": 0, + "content": "\\[\nd X _ {t} = - \\frac {1}{2} \\beta (t) X _ {t} d t + \\sqrt {\\beta (t)} d W _ {t}, X _ {0} \\sim p _ {\\text {d a t a}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.843, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\((W_{t})_{t\\in [0,T_{c}]}\\) is a standard Brownian motion, and \\(\\beta (t)\\) is the noise schedule. Let \\((p_t)_{t\\in [0:T_c]}\\) be the path distribution associated with (3). To perform data sampling following the distribution \\(p_{\\mathrm{data}}\\) we can reverse the SDE (3) and" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2378" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "apply discretization. According to [1] and [21], the reverse SDE corresponding to (3) is" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.126, + 0.483, + 0.173 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} d \\overleftarrow {X} _ {t} = \\left[ \\frac {1}{2} \\beta \\left(T _ {c} - t\\right) \\overleftarrow {X} _ {t} + \\beta \\left(T _ {c} - t\\right) \\nabla \\log p _ {T _ {c} - t} (\\overleftarrow {X} _ {t}) \\right] d t \\\\ + \\sqrt {\\beta (T _ {c} - t)} d \\tilde {W} _ {t}, \\overleftarrow {X} _ {0} \\sim p _ {T _ {c}}, \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.18, + 0.484, + 0.289 + ], + "angle": 0, + "content": "where \\((\\tilde{W}_t)_{t\\in [0,T_c]}\\) is another standard Brownian motion. The reverse SDE produces a time-reverse process \\((\\overleftarrow{X}_t)_{t\\in [0,T_c]}\\) where \\(\\overleftarrow{X}_t\\) has the same distribution with \\(X_{T_c - t}\\). Note that the drift function in (4) depends on the score function \\(\\nabla_{\\mathbf{x}_t}\\log p_t(\\mathbf{x}_t)\\), which can be approximated by a time-aware neural network trained with denoising score matching [22]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.296, + 0.457, + 0.313 + ], + "angle": 0, + "content": "2.2 Score-based model for posterior sampling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.484, + 0.351 + ], + "angle": 0, + "content": "In many application scenarios, we may observe a noisy version of the measurement of the data \\(X_0\\), given by" + }, + { + "type": "equation", + "bbox": [ + 0.224, + 0.357, + 0.351, + 0.373 + ], + "angle": 0, + "content": "\\[\nY = \\mathcal {A} (X _ {0}) + N,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.381, + 0.484, + 0.413 + ], + "angle": 0, + "content": "where \\(\\mathcal{A}(\\cdot):\\mathbb{R}^d\\to \\mathbb{R}^n\\) is a measurement operator and \\(N\\in \\mathbb{R}^n\\) is the measurement noise with \\(N\\sim \\mathcal{N}(\\mathbf{0},\\sigma_n^2\\mathbf{I})\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.414, + 0.484, + 0.473 + ], + "angle": 0, + "content": "Score-based generative models have shown powerful capability in solving general inverse problems [6, 10, 19, 26, 28, 29]. Leveraging the diffusion model as the prior, sampling from the reverse posterior distribution" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.48, + 0.483, + 0.522 + ], + "angle": 0, + "content": "\\[\np \\left(\\mathbf {x} _ {0: T} \\mid \\mathbf {y}\\right) := p \\left(\\mathbf {x} _ {T} \\mid \\mathbf {y}\\right) \\prod_ {k = 0} ^ {T - 1} p \\left(\\mathbf {x} _ {k} \\mid \\mathbf {x} _ {k + 1}, \\mathbf {y}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.529, + 0.484, + 0.627 + ], + "angle": 0, + "content": "requires the knowledge of conditional scores \\(\\nabla_{\\mathbf{x}_k}\\log p(\\mathbf{x}_k|\\mathbf{y})\\) which can be expressed as \\(\\nabla_{\\mathbf{x}_k}\\log p(\\mathbf{x}_k|\\mathbf{y}) = \\nabla_{\\mathbf{x}_k}\\log p(\\mathbf{x}_k) + \\nabla_{\\mathbf{x}_k}\\log p(\\mathbf{y}|\\mathbf{x}_k)\\). The first term can be approximated by a score network \\(s_\\theta (\\mathbf{x}_k,k)\\) [8, 21], and different methods have been proposed to estimate the second term \\(\\nabla_{\\mathbf{x}_k}\\log p(\\mathbf{y}|\\mathbf{x}_k)\\) [6, 19]. In each step the conditional sampling process in VP diffusion is [6, 19]" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.632, + 0.488, + 0.693 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {x} _ {k - 1} \\leftarrow \\frac {1}{\\sqrt {\\alpha_ {k}}} \\left(\\mathbf {x} _ {k} + (1 - \\alpha_ {k}) \\nabla_ {\\mathbf {x} _ {k}} \\log p (\\mathbf {x} _ {k} | \\mathbf {y})\\right) + \\tilde {\\sigma} _ {k} \\mathbf {z}, \\tag {6} \\\\ \\approx \\frac {1}{\\sqrt {\\alpha_ {k}}} (\\mathbf {x} _ {k} + (1 - \\alpha_ {k}) (s _ {\\theta} (\\mathbf {x} _ {k}, k) + \\nabla_ {\\mathbf {x} _ {k}} \\log p (\\mathbf {y} | \\mathbf {x} _ {k}))) + \\tilde {\\sigma} _ {k} \\mathbf {z}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.698, + 0.484, + 0.728 + ], + "angle": 0, + "content": "where \\(\\mathbf{z} \\sim \\mathcal{N}(0, \\mathbf{I})\\) and \\(\\tilde{\\sigma}_k^2\\) is the approximated variance of the reverse posterior distribution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.729, + 0.484, + 0.79 + ], + "angle": 0, + "content": "Meanwhile, [28] proposed to directly approximate the reverse conditional \\(p(\\mathbf{x}_k|\\mathbf{x}_{k + 1},\\mathbf{y})\\) and propagate the mean in each step. Specifically, when \\(T\\to \\infty\\) , the mean of \\(p(\\mathbf{x}_k|\\mathbf{x}_{k + 1},\\mathbf{y})\\) in VP-diffusion is given by [28]" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.797, + 0.483, + 0.815 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} _ {k} \\left(\\mathbf {x} _ {k + 1}, \\mathbf {y}\\right) = \\mathbf {U} _ {k} \\mathbf {x} _ {k + 1} + \\mathbf {V} _ {k} \\mathbb {E} _ {p \\left(\\mathbf {x} _ {0} \\mid \\mathbf {y}\\right)} [ X _ {0} ], \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.82, + 0.138, + 0.832 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.84, + 0.445, + 0.878 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {U} _ {k} := \\sqrt {\\alpha_ {k + 1}} \\left(\\left(1 - \\bar {\\alpha} _ {k}\\right) \\mathbf {I} + \\bar {\\alpha} _ {k} \\operatorname {C o v} _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ]\\right) \\\\ \\cdot \\left(\\left(1 - \\bar {\\alpha} _ {k + 1}\\right) \\mathbf {I} + \\bar {\\alpha} _ {k + 1} \\operatorname {C o v} _ {p \\left(\\mathbf {x} _ {0} \\mid \\mathbf {y}\\right)} [ X _ {0} ]\\right) ^ {- 1}, \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.881, + 0.481, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\mathbf {V} _ {k} := \\sqrt {\\bar {\\alpha} _ {k}} \\big ((1 - \\alpha_ {k + 1}) \\mathbf {I} \\big) \\big ((1 - \\bar {\\alpha} _ {k + 1}) \\mathbf {I} + \\bar {\\alpha} _ {k + 1} \\operatorname {C o v} _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ] \\big) ^ {- 1}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.906, + 0.213 + ], + "angle": 0, + "content": "Here \\(\\alpha_{k} = 1 - \\beta_{k}\\) and \\(\\bar{\\alpha}_{k} = \\Pi_{i=0}^{k} \\alpha_{i}\\) in VP-diffusion. The authors proved that by propagating the mean at each reverse step (i.e., \\(\\pmb{\\mu}_{T} \\rightarrow \\pmb{\\mu}_{T-1}(\\mathbf{x}_{T} = \\pmb{\\mu}_{T}, \\mathbf{y}) \\rightarrow \\dots \\rightarrow \\pmb{\\mu}_{0}(\\mathbf{x}_{1} = \\pmb{\\mu}_{1}, \\mathbf{y})\\)), the end point \\(\\pmb{\\mu}_{0}(\\mathbf{x}_{1} = \\pmb{\\mu}_{1}, \\mathbf{y})\\) represents the MMSE estimator \\(\\mathbb{E}_{p(\\mathbf{x}_0|\\mathbf{y})}[X_0]\\) when \\(\\bar{\\alpha}_{T} \\rightarrow 0\\). The proof of reverse mean propagation converging to the MMSE point [28] and its connection to conditional score are included in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.22, + 0.906, + 0.25 + ], + "angle": 0, + "content": "2.3 Connection between reverse mean and conditional score" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.257, + 0.906, + 0.315 + ], + "angle": 0, + "content": "The mean derived in (7) can be viewed as an approximation of the mean in (6). Specifically, when \\( T \\to \\infty \\), \\( p(\\mathbf{x}_k|\\mathbf{y}) \\) can be approximately expressed in the form of a Gaussian distribution. We can show that (see details in Appendix A)" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.32, + 0.903, + 0.35 + ], + "angle": 0, + "content": "\\[\n\\frac {1}{\\sqrt {\\alpha_ {k}}} (\\mathbf {x} _ {k} + (1 - \\alpha_ {k}) \\nabla_ {\\mathbf {x} _ {k}} \\log p (\\mathbf {x} _ {k} | \\mathbf {y})) = \\mathbf {U} _ {k} \\mathbf {x} _ {k + 1} + \\mathbf {V} _ {k} \\mathbb {E} _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ].\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.363, + 0.907, + 0.545 + ], + "angle": 0, + "content": "Theoretically, when \\(\\tilde{\\sigma}_k = 0\\) in each step (6), the endpoint of the reverse process converges to the MMSE estimator, which achieves the best possible performance on distortion. When \\(\\tilde{\\sigma}_k\\) is the true posterior variance, the reverse process samples from the posterior distribution \\(p(\\mathbf{x}_{0:T}|\\mathbf{y})\\), which results in a reconstruction with perfect perception measured in conditional distribution. The above observations motivate us to bridge the two extreme points with the score-based generative model by tuning the scale of \\(\\sigma_k\\) in the reverse process. It remains to be answered whether the bridging can provide us with an optimal and flexible estimator to traverse along the distortion-perception plane." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.556, + 0.906, + 0.59 + ], + "angle": 0, + "content": "3 Traversing DP Tradeoff with Scaled Reverse Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.705 + ], + "angle": 0, + "content": "In this section, we first propose a novel reverse sampling method based on a variance-scaled version of the joint inference distribution (5) and theoretically derive the mean and variance of the marginal distribution. Then, we prove that the resulting end distribution of the variance-scaled reverse sampling provides an optimal DP tradeoff with respect to MSE and Wasserstein-2 distance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.713, + 0.771, + 0.727 + ], + "angle": 0, + "content": "3.1 Reverse mean and variance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.736, + 0.906, + 0.766 + ], + "angle": 0, + "content": "Consider the joint inference distribution with true mean but scaled variance, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.773, + 0.907, + 0.798 + ], + "angle": 0, + "content": "\\[\np _ {\\lambda} \\left(\\mathbf {x} _ {k} \\mid \\mathbf {x} _ {k + 1}, \\mathbf {y}\\right) := \\mathcal {N} \\left(\\boldsymbol {\\mu} _ {k} \\left(\\mathbf {x} _ {k + 1}, \\mathbf {y}\\right), \\lambda \\mathbf {C} _ {k}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.805, + 0.906, + 0.835 + ], + "angle": 0, + "content": "where the expectation \\(\\pmb{\\mu}_k(\\mathbf{x}_{k + 1},\\mathbf{y})\\) is given in (7), and the true covariance \\(\\mathbf{C}_k\\) is given by [28]" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.84, + 0.906, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {C} _ {k} := \\frac {\\beta_ {k + 1}}{1 - \\beta_ {k + 1}} \\left((1 - \\bar {\\alpha} _ {k}) \\mathbf {I} + \\bar {\\alpha} _ {k} \\operatorname {C o v} _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ]\\right) \\\\ \\cdot \\left(\\left(\\frac {\\beta_ {k + 1}}{1 - \\beta_ {k + 1}} + 1 - \\bar {\\alpha} _ {k}\\right) \\mathbf {I} + \\bar {\\alpha} _ {k} \\operatorname {C o v} _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ]\\right) ^ {- 1}. \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2379" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.136 + ], + "angle": 0, + "content": "We have the following theorem to characterize the marginal distribution given by the variance-scaled version of the joint inference distribution." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.145, + 0.485, + 0.176 + ], + "angle": 0, + "content": "Theorem 1. For VP-diffusion and \\(0 \\leq \\lambda \\leq 1\\), consider the joint inference distribution given by" + }, + { + "type": "equation", + "bbox": [ + 0.123, + 0.184, + 0.484, + 0.226 + ], + "angle": 0, + "content": "\\[\np _ {\\lambda} \\left(\\mathbf {x} _ {0: T} \\mid \\mathbf {y}\\right) = p _ {\\lambda} \\left(\\mathbf {x} _ {T} \\mid y\\right) \\prod_ {k = 0} ^ {T - 1} p _ {\\lambda} \\left(\\mathbf {x} _ {k} \\mid \\mathbf {x} _ {k + 1}, \\mathbf {y}\\right) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.234, + 0.484, + 0.279 + ], + "angle": 0, + "content": "where \\(p_{\\lambda}(\\mathbf{x}_T|\\mathbf{y}) = \\mathcal{N}(0,\\mathbf{I})\\) and \\(p_{\\lambda}(\\mathbf{x}_k|\\mathbf{x}_{k + 1},\\mathbf{y})\\) is given by (8). Then, the corresponding margin has the distribution \\(p_{\\lambda}(\\mathbf{x}_k|\\mathbf{y}) = \\mathcal{N}(\\boldsymbol {\\mu}_k^\\lambda ,\\boldsymbol{\\Sigma}_k^\\lambda)\\) where" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.286, + 0.487, + 0.355 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {\\mu} _ {k} ^ {\\lambda} = \\left(\\mathbf {U} _ {k} \\left(\\mathbf {U} _ {k + 1} (\\dots (\\mathbf {U} _ {T - 2} \\mathbf {V} _ {T - 1} + \\mathbf {V} _ {T - 2}) \\dots) + \\mathbf {V} _ {k + 1}\\right) + \\mathbf {V} _ {k}\\right) \\\\ = \\sqrt {\\bar {\\alpha} _ {k}} (1 - \\alpha_ {k + 1} \\cdot \\cdot \\cdot \\alpha_ {T}) \\left((1 - \\bar {\\alpha} _ {T}) \\mathbf {I} + \\bar {\\alpha} _ {T} C o v _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ]\\right) ^ {- 1} \\mathbb {E} _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ], \\\\ \\boldsymbol {\\Sigma} _ {k} ^ {\\lambda} = \\boldsymbol {\\Sigma} _ {k} \\Big (\\lambda \\mathbf {I} + (1 - \\lambda) \\alpha_ {k + 1} \\alpha_ {k + 2} \\dots \\alpha_ {T} \\boldsymbol {\\Sigma} _ {T - 1} ^ {- 1} \\boldsymbol {\\Sigma} _ {k} \\Big), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.362, + 0.484, + 0.393 + ], + "angle": 0, + "content": "where \\(\\pmb{\\Sigma}_k\\) is the covariance of \\(p(\\mathbf{x}_k|\\mathbf{y})\\) in VP diffusion. In particular, when \\(\\bar{\\alpha}_T\\rightarrow 0\\), the variance of \\(p_{\\lambda}(\\mathbf{x}_0|\\mathbf{y})\\) is" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.401, + 0.488, + 0.427 + ], + "angle": 0, + "content": "\\[\n\\pmb {\\Sigma} _ {0} ^ {\\lambda} = \\pmb {\\Sigma} _ {0} \\bigg (\\lambda \\mathbf {I} + (1 - \\lambda) \\bar {\\alpha} _ {T} \\pmb {\\Sigma} _ {T - 1} ^ {- 1} \\pmb {\\Sigma} _ {0} \\bigg) \\rightarrow \\lambda C o v _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.434, + 0.201, + 0.448 + ], + "angle": 0, + "content": "and the mean is" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.455, + 0.49, + 0.495 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mu_ {0} ^ {\\lambda} = \\sqrt {\\bar {\\alpha} _ {0}} \\big ((1 - \\bar {\\alpha} _ {T}) \\mathbf {I} \\big) \\big ((1 - \\bar {\\alpha} _ {T}) \\mathbf {I} + \\bar {\\alpha} _ {T} C o v _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ] \\big) ^ {- 1} \\\\ \\cdot \\mathbb {E} _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ] \\rightarrow \\mathbb {E} _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ]. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.503, + 0.342, + 0.519 + ], + "angle": 0, + "content": "Proof. See the details in Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.528, + 0.484, + 0.649 + ], + "angle": 0, + "content": "The above theorem shows that the proposed variance-scaled reverse diffusion process will lead to a scaled marginal distribution at each step, which finally results in the posterior mean and scaled posterior variance of the original data distribution at the end of the reverse process. This seemingly intuitive result actually provides the optimal solution to the conditional distortion-perception tradeoff for multivariate Gaussian distribution." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.657, + 0.456, + 0.672 + ], + "angle": 0, + "content": "3.2 DP tradeoff for conditional Gaussian case" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.68, + 0.484, + 0.77 + ], + "angle": 0, + "content": "Consider the \\(d\\)-dimensional source \\(X \\in \\mathbb{R}^d\\) with distribution \\(p(\\mathbf{x})\\). Let \\(Y\\) be the noisy version of the source data. Denote the conditional expectation and variance given an observation \\(\\mathbf{y}\\) as \\(\\boldsymbol{\\mu}_{\\mathbf{y}} = \\mathbb{E}_{p(\\mathbf{x}|\\mathbf{y})}[X]\\) and \\(\\boldsymbol{\\Sigma}_{\\mathbf{y}} = \\mathrm{Cov}_{p(\\mathbf{x}|\\mathbf{y})}[X]\\), respectively. The goal of signal restoration is to find a good reconstruction \\(\\hat{X}\\) based on the observed \\(Y\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.771, + 0.484, + 0.817 + ], + "angle": 0, + "content": "From the optimization perspective, given a specific observation \\( Y = \\check{\\mathbf{y}} \\), we can define the conditional distortion-perception function as" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.824, + 0.454, + 0.875 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} D (P) = \\min _ {p _ {\\hat {X} | Y} (\\hat {\\mathbf {x}} | \\check {\\mathbf {y}})} \\mathbb {E} _ {p _ {X, \\hat {X} | Y} (\\mathbf {x}, \\hat {\\mathbf {x}} | \\check {\\mathbf {y}})} [ \\| X - \\hat {X} \\| _ {2} ^ {2} ] \\\\ \\begin{array}{l} \\text {s . t .} d (p _ {X | Y} (\\mathbf {x} | \\breve {\\mathbf {y}}), p _ {\\hat {X} | Y} (\\hat {\\mathbf {x}} | \\breve {\\mathbf {y}})) \\leq P, \\end{array} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.885, + 0.367, + 0.901 + ], + "angle": 0, + "content": "where \\(X, Y\\) and \\(\\hat{X}\\) form a Markov chain." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.151 + ], + "angle": 0, + "content": "Theorem 2. Consider the source \\(X\\) and an observation \\(\\tilde{\\mathbf{y}}\\) that satisfy \\(p_{X|Y}(\\mathbf{x}|\\tilde{\\mathbf{y}}) \\sim \\mathcal{N}(\\boldsymbol{\\mu}_{\\tilde{\\mathbf{y}}}, \\boldsymbol{\\Sigma}_{\\tilde{\\mathbf{y}}})\\). The optimal distortion-perception tradeoff given \\(\\tilde{\\mathbf{y}}\\) with MSE and Wasserstein-2 distance is" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.158, + 0.907, + 0.215 + ], + "angle": 0, + "content": "\\[\nD (P) = \\left\\{ \\begin{array}{l} \\operatorname {T r} \\left(\\boldsymbol {\\Sigma} _ {\\tilde {\\mathbf {y}}}\\right) + \\left(\\sqrt {\\operatorname {T r} \\left(\\boldsymbol {\\Sigma} _ {\\tilde {\\mathbf {y}}}\\right)} - P\\right) ^ {2}, f o r P \\leq \\sqrt {\\operatorname {T r} \\left(\\boldsymbol {\\Sigma} _ {\\tilde {\\mathbf {y}}}\\right)} \\\\ \\operatorname {T r} \\left(\\boldsymbol {\\Sigma} _ {\\tilde {\\mathbf {y}}}\\right), f o r P > \\sqrt {\\operatorname {T r} \\left(\\boldsymbol {\\Sigma} _ {\\tilde {\\mathbf {y}}}\\right)}. \\end{array} \\right. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.224, + 0.907, + 0.269 + ], + "angle": 0, + "content": "Meanwhile, the optimal tradeoff can be achieved by sampling from the joint inference distribution (10) in Theorem 1 and traversing different \\(0 \\leq \\lambda \\leq 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.277, + 0.743, + 0.293 + ], + "angle": 0, + "content": "Proof. See details in Appendix C." + }, + { + "type": "image", + "bbox": [ + 0.889, + 0.278, + 0.907, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.302, + 0.907, + 0.514 + ], + "angle": 0, + "content": "Theorem 2 explicitly characterizes the optimal DP tradeoff for conditional Gaussian distribution. When \\( P = 0 \\) (i.e., perfect perceptual quality is expected), the best achievable MSE is \\( 2\\mathrm{Tr}(\\Sigma_{\\bar{\\mathbf{y}}}) \\). As \\( P \\) increases, the optimal MSE gradually converges to the MMSE value \\( \\mathrm{Tr}(\\Sigma_{\\bar{\\mathbf{y}}}) \\). The proof in Appendix C also shows that the variance-scaled reverse diffusion process can attain the entire optimal DP curve for conditional Gaussian distribution. Although the actual data distribution may not follow the conditional Gaussian distribution (e.g., mixture Gaussian, 2D distribution like S-curve, and image dataset), the forthcoming experiments will show that the proposed variance-scaled reverse sampling can effectively and flexibly traverse the DP tradeoff on general datasets using a single score network." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.522, + 0.877, + 0.539 + ], + "angle": 0, + "content": "3.3 Variance-scaled reverse sampling process" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.906, + 0.605 + ], + "angle": 0, + "content": "In this subsection, we discuss how to perform variance-scaled reverse sampling by estimating the conditional score. With a single pre-trained score network, we can traverse different DP tradeoffs for general inverse problems." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.606, + 0.906, + 0.694 + ], + "angle": 0, + "content": "Recall that the proposed variance-scaled reverse sampling process is given by \\(p_{\\lambda}(\\mathbf{x}_{0:T}|\\mathbf{y}) = p_{\\lambda}(\\mathbf{x}_T|y)\\prod_{k = 0}^{T - 1}p_{\\lambda}(\\mathbf{x}_k|\\mathbf{x}_{k + 1},\\mathbf{y})\\) where for VP diffusion, we have the posteriors \\(p_{\\lambda}(\\mathbf{x}_{k - 1}|\\mathbf{x}_k,\\mathbf{y})\\coloneqq \\mathcal{N}\\big(\\boldsymbol{\\mu}_{k - 1}(\\mathbf{x}_k,\\mathbf{y}),\\lambda \\mathbf{C}_{k - 1}\\big)\\). The sampling process in each step can be approximately expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.7, + 0.906, + 0.731 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {k - 1} = \\frac {1}{\\sqrt {\\alpha_ {k}}} \\left(\\mathbf {x} _ {k} + (1 - \\alpha_ {k}) \\nabla_ {\\mathbf {x} _ {k}} \\log p \\left(\\mathbf {x} _ {k} \\mid \\mathbf {y}\\right)\\right) + \\lambda \\mathbf {C} _ {k} \\mathbf {z}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.737, + 0.906, + 0.865 + ], + "angle": 0, + "content": "where \\(\\mathbf{z} \\sim \\mathcal{N}(0, \\mathbf{I})\\). To draw a sample \\(\\mathbf{x}_{k-1}\\) according to (12), we need to compute the conditional score \\(\\nabla_{\\mathbf{x}_k} \\log p(\\mathbf{x}_k | \\mathbf{y})\\). Specifically, the conditional scores can be decomposed as \\(\\nabla_{\\mathbf{x}_k} p(\\mathbf{x}_k | \\mathbf{y}) = \\nabla_{\\mathbf{x}_k} p(\\mathbf{x}_k) + \\nabla_{\\mathbf{x}_k} p(\\mathbf{y} | \\mathbf{x}_k)\\) by Bayes' rule, where the first term can be approximated by a score network \\(s_\\theta(\\mathbf{x}_k, k)\\) [8, 21]. To deal with the second term, Denoising Posterior Sampling (DPS) [6] proposes to approximate \\(p(\\mathbf{y} | \\mathbf{x}_k) = \\mathbb{E}_{p(\\mathbf{x}_0 | \\mathbf{x}_k)}[p(\\mathbf{y} | \\mathbf{x}_0)]\\) with \\(p(\\mathbf{y} | \\hat{\\mathbf{x}}_0(\\mathbf{x}_k))\\), where" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.871, + 0.907, + 0.9 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} _ {0} \\left(\\mathbf {x} _ {k}\\right) = \\mathbb {E} _ {p \\left(\\mathbf {x} _ {0} \\mid \\mathbf {x} _ {k}\\right)} \\left[ X _ {0} \\right] = \\frac {1}{\\alpha_ {k}} \\left(\\mathbf {x} _ {k} + \\left(1 - \\bar {\\alpha} _ {k}\\right) \\nabla_ {\\mathbf {x} _ {k}} p \\left(\\mathbf {x} _ {k}\\right)\\right) \\tag {13}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2380" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.234 + ], + "angle": 0, + "content": "by Tweedie's formula. The approximation error here is upper-bounded by the Jensen gap [17]. Note that \\(p(\\mathbf{y}|\\hat{\\mathbf{x}}_0(\\mathbf{x}_k))\\) is analytically tractable when the measurement distribution is given. Specifically, when the measurement nosie is \\(\\mathcal{N}(0,\\sigma_n^2\\mathbf{I})\\), the second term in the conditional score takes the form of \\(\\nabla_{\\mathbf{x}_k}\\log p(\\mathbf{y}|\\mathbf{x}_k) = \\frac{1}{\\sigma_n^2}\\nabla_{\\mathbf{x}_k}||\\mathbf{y}-\\mathcal{A}(\\hat{\\mathbf{x}}_0(\\mathbf{x}_k))||_2^2\\). Thus, the conditional scores can be approximated as \\(\\nabla_{\\mathbf{x}_k}\\log p(\\mathbf{x}_k|\\mathbf{y}) = s_\\theta (\\mathbf{x}_k,k) + \\frac{1}{\\sigma_n^2}\\nabla_{\\mathbf{x}_k}||\\mathbf{y}-\\mathcal{A}(\\hat{\\mathbf{x}}_0(\\mathbf{x}_k))||_2^2\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.235, + 0.484, + 0.281 + ], + "angle": 0, + "content": "In the following experiments, we adopt the DPS framework to approximate the conditional score and sample from (12), which is sketched as Algorithm 1." + }, + { + "type": "code_caption", + "bbox": [ + 0.093, + 0.295, + 0.482, + 0.311 + ], + "angle": 0, + "content": "Algorithm 1 Variance-Scaled Reverse Sampling with DPS" + }, + { + "type": "algorithm", + "bbox": [ + 0.093, + 0.315, + 0.484, + 0.478 + ], + "angle": 0, + "content": "Require: \\(T,\\mathbf{y},\\lambda ,\\{\\tilde{\\sigma}_k\\}_{k = 0}^T,\\{\\zeta_{k,\\lambda}\\}_{k = 0}^T\\) \\(\\mathbf{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) \nfor \\(k = T - 1\\) to 1 do do \\(\\hat{s}\\gets s_{\\theta}(\\mathbf{x}_k,k)\\) \\(\\hat{\\mathbf{x}}_0\\gets \\frac{1}{\\sqrt{\\bar{\\alpha}_k}} (\\mathbf{x}_k + (1 - \\bar{\\alpha}_k)\\hat{s})\\) \\(\\hat{c} (\\hat{\\mathbf{x}}_0) = -\\frac{\\sqrt{\\alpha_k}}{1 - \\alpha_k}\\nabla_{\\mathbf{x}_k}||y - \\mathcal{A}(\\hat{\\mathbf{x}}_0(\\mathbf{x}_k))||_2^2\\) \\(\\mathbf{z}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) \\(\\mathbf{x}_{k - 1}\\leftarrow \\frac{1}{\\sqrt{\\alpha_k}}\\bigl (\\mathbf{x}_k + (1 - \\alpha_k)\\bigl (\\hat{s} +\\zeta_{k,\\lambda}\\hat{c} (\\hat{\\mathbf{x}}_0)\\bigr)\\bigr) + \\lambda \\tilde{\\sigma}_k\\mathbf{z}\\) \nend for \nreturn \\(\\mathbf{x}_0\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.499, + 0.484, + 0.575 + ], + "angle": 0, + "content": "In Algorithm 1, \\(\\tilde{\\sigma}_k\\) is set to \\(\\beta_{k}\\) [8, 28], and \\(\\zeta_{k,\\lambda}\\) is a hyperparameter to control the weight of the conditional score, and may differ for different \\(\\lambda\\). The heuristic choices of \\(\\zeta_{k,\\lambda}\\) and the rationale for fine-tuning have been included in Appendix E." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Remark 1. In this paper, we adopt the conditional score approximation proposed in [6] for illustration. To implement the variance-scaled reverse sampling process (12), one can employ any approach to directly estimate the conditional score [19] or estimate the posterior mean as a whole [28] for general inverse problems. For example, Pseudoinverse-guided Diffusion Models (ΠGDM) [19] approximated the unconditional posteriors \\( p(\\mathbf{x}_0|\\mathbf{x}_k) \\) with Gaussian, i.e., \\( p(\\mathbf{x}_0|\\mathbf{x}_k) \\approx \\mathcal{N}(\\hat{\\mathbf{x}}_0,r_k^2\\mathbf{I}) \\), where \\( \\hat{\\mathbf{x}}_0 \\) is given by (13), and \\( r_k^2 \\) is the hyperparameter. Thus, the score term \\( \\nabla_{\\mathbf{x}_k}p(\\mathbf{y}|\\mathbf{x}_k) \\) can be analytically expressed by the pseudoinverse of the measurement model. In [28], the posterior mean is estimated as a whole. Specifically, we have \\( \\mu_{k - 1}(\\mathbf{x}_k,\\mathbf{y}) = \\mathbf{U}_{k - 1}\\mathbf{x}_k + \\mathbf{V}_{k - 1}\\mathbb{E}_{p(\\mathbf{x}_0|\\mathbf{y})}[X_0] \\). Since the expression involves \\( \\mathbb{E}_{p(\\mathbf{x}_0|\\mathbf{y})}[X_0] \\) and \\( \\mathrm{Cov}_{p(\\mathbf{x}_0|\\mathbf{y})}[X_0] \\), it can not be computed directly. Instead, the authors approximate the joint conditional posterior with variational Gaussian distributions and apply natural gradient descent to perform the sampling. Note that the gradient calculation also involves the approximation of the conditional score \\( \\nabla_{\\mathbf{x}_k}\\log p(\\mathbf{y}|\\mathbf{x}_k) \\) as in DPS [6]." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.091, + 0.641, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.563, + 0.171, + 0.617, + 0.185 + ], + "angle": 0, + "content": "(a) \\(p(x_0)\\)" + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.091, + 0.761, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.172, + 0.741, + 0.185 + ], + "angle": 0, + "content": "(b) \\(p(y)\\)" + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.091, + 0.897, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.782, + 0.171, + 0.891, + 0.186 + ], + "angle": 0, + "content": "(c) \\(p(x|y = -0.6)\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.199, + 0.905, + 0.227 + ], + "angle": 0, + "content": "Figure 1. An example of mixture Gaussian, the noisy observation, and the conditional distribution given an observation \\( y = -0.6 \\)" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.234, + 0.738, + 0.25 + ], + "angle": 0, + "content": "3.4 An illustrative example" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.257, + 0.905, + 0.378 + ], + "angle": 0, + "content": "Consider the mixture Gaussian distribution \\(X_0 \\sim p(x_0)\\) with two components, where \\(p(x_0) = w_1 \\mathcal{N}(\\mu_1, \\sigma_1^2) + w_2 \\mathcal{N}(\\mu_2, \\sigma_2^2)\\). The noisy observation is obtained by \\(Y = aX_0 + \\sigma_0 \\epsilon\\), where \\(\\epsilon \\sim \\mathcal{N}(0, 1)\\), i.e., \\(p(y|x_0) = \\mathcal{N}(ax_0, \\sigma_0^2)\\). The MMSE estimator and posterior distribution can be theoretically derived (see Appendix D for details). Fig. 1 shows an example of mixture Gaussian distribution and the conditional distribution given an observation \\(y = -0.6\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.378, + 0.906, + 0.574 + ], + "angle": 0, + "content": "Given an observation \\(y\\) and a chosen \\(\\lambda\\), we iteratively perform the variance scaled reverse process (12) for \\(k = T, \\dots, 1\\), where the variance is approximated by \\(\\mathbf{C}_k = \\beta_k\\) as the conventional VP-diffusion. Starting from a random sample \\(x_T \\sim \\mathcal{N}(0,1)\\), Fig. 2 shows the trajectories \\(x_T \\to \\dots x_k \\to \\dots x_0\\) of multiple reconstructions for each \\(\\lambda\\). We can observe that for \\(\\lambda = 0\\), given an initial \\(x_T\\), the trajectories are deterministic and converge to the MMSE point. This phenomenon collaborates with results in [28]. When \\(\\lambda\\) increases, the generated trajectories follow the form of the posterior distribution and show more stochasticity. When \\(\\lambda = 1\\), the reconstruction distribution matches the ground truth posterior." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.574, + 0.906, + 0.71 + ], + "angle": 0, + "content": "The resulting DP tradeoffs are shown in Fig. 3. It can be observed that as \\(\\lambda\\) increases, the divergence between the true posterior distribution and reconstruction distribution continuously decreases, with an increase of MSE value. The overall curve is convex, and both Wasserstein-2 and KL-divergence come to zero. Meanwhile, the MSE achieved at \\(\\lambda = 1\\) is approximately two times the MSE at \\(\\lambda = 0\\) for both Wasserstein-2 and KL-divergence cases, coinciding with the result in Theorem 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.723, + 0.729, + 0.74 + ], + "angle": 0, + "content": "4 Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.906, + 0.839 + ], + "angle": 0, + "content": "In this section, we conduct a series of experiments on two-dimensional datasets and the FFHQ dataset [9]. It is demonstrated that the variance-scaled reverse sampling shows great flexibility and effectiveness in traversing DP tradeoff compared to the GAN-based approach and other denoising methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.81, + 0.864 + ], + "angle": 0, + "content": "4.1 Two-dimensional points example" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "First, we evaluate the validity and effectiveness of our approach on two-dimensional datasets, including pinwheel, S-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2381" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.073, + 0.904, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.198, + 0.88, + 0.213 + ], + "angle": 0, + "content": "Figure 2. Trajectories \\( x_{T} \\to \\dots x_{k} \\to \\dots \\to x_{0} \\) of different reconstructions for \\( \\lambda = 0, 0.3, 0.8 \\) and 1. The initial \\( x_{T} \\) is \\( \\mathcal{N}(0,1) \\)." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.231, + 0.287, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.185, + 0.35, + 0.197, + 0.359 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.231, + 0.484, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.35, + 0.408, + 0.358 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.373, + 0.484, + 0.429 + ], + "angle": 0, + "content": "Figure 3. Distortion-perception tradeoff traversed by variance-scaled reverse sampling given different \\(\\lambda\\)'s. (a) tradeoff between Wasserstein-2 distance and MSE; (b) tradeoff between KL-divergence and MSE." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.441, + 0.483, + 0.502 + ], + "angle": 0, + "content": "curve, and moon-shape data distributions. Taking the pinwheel dataset as an example, the first row of Fig. 4 shows the distribution of pinwheel data points \\(X_0\\) and the noisy observation \\(Y\\), where \\(Y = aX + N\\) for \\(N \\sim \\mathcal{N}(0, \\sigma_n^2 \\mathbf{I})\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.504, + 0.484, + 0.793 + ], + "angle": 0, + "content": "For each dataset, we train a conventional score network to approximate \\( p(\\mathbf{x}_0) \\) on the original distribution. Then, we perform the variance-scaled reverse diffusion process (12) on the noisy observation. As a comparison, we adopt a GAN-based DP-traversing approach PSCGAN [15]. This scheme introduces a penalty of the posterior expectation to the training of the conditional GAN. The authors proposed two strategies to navigate the DP tradeoff during the inference time: (1) PSCGAN- \\( N \\): This method involves sampling \\( N \\) instances from the generator and then averaging them. As \\( N \\) increases, the averaged image is closer to the conditional expectation, leading to a smaller distortion but larger perception loss. (2) PSCGAN- \\( z \\): It varies the standard deviation \\( \\sigma_z \\) of the noise injected into the generator to control the stochasticity. For \\( 0 \\leq \\sigma_z \\leq 1 \\), a larger \\( \\sigma_z \\) results in a better perceptual quality and higher MSE. Note that the training of PSCGAN relies on noisy data points \\( Y \\). Thus, the model needs to be retrained for different measurement scenarios and different levels of \\( \\sigma_n^2 \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.794, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Fig. 4 shows the reconstruction of our variance-scaled reverse diffusion process for different \\(\\lambda\\)'s, as well as the reconstruction of PSCGAN when setting \\(N = 16\\), \\(\\sigma_z = 1\\). It can be observed that when \\(\\lambda = 0\\), our sampling process leads to a more concentrated reconstruction, and gradually approaches the true distribution as \\(\\lambda\\) increases." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Fig. 5 compares the numerical DP tradeoffs achieved" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.23, + 0.709, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.618, + 0.295, + 0.694, + 0.303 + ], + "angle": 0, + "content": "Original distributions" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.231, + 0.824, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.295, + 0.807, + 0.303 + ], + "angle": 0, + "content": "Noisy observations" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.309, + 0.653, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.578, + 0.377, + 0.612, + 0.385 + ], + "angle": 0, + "content": "(a) \\(\\lambda = 0\\)" + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.31, + 0.767, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.695, + 0.377, + 0.735, + 0.385 + ], + "angle": 0, + "content": "(b) \\(\\lambda = 0.3\\)" + }, + { + "type": "image", + "bbox": [ + 0.77, + 0.31, + 0.884, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.812, + 0.377, + 0.853, + 0.385 + ], + "angle": 0, + "content": "(c) \\(\\lambda = 0.5\\)" + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.389, + 0.652, + 0.454 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.578, + 0.457, + 0.622, + 0.465 + ], + "angle": 0, + "content": "(d) \\(\\lambda = 0.8\\)" + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.389, + 0.768, + 0.454 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.457, + 0.732, + 0.465 + ], + "angle": 0, + "content": "(e) \\(\\lambda = 1\\)" + }, + { + "type": "image", + "bbox": [ + 0.77, + 0.389, + 0.884, + 0.454 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.779, + 0.457, + 0.886, + 0.466 + ], + "angle": 0, + "content": "(f) PSCGAN \\(N = 16, \\sigma_z = 1\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.478, + 0.907, + 0.562 + ], + "angle": 0, + "content": "Figure 4. Experiments on a two-dimensional dataset. The first row illustrates the original distribution (left) of pinwheel and the noisy observation \\( Y \\), where \\( Y = aX + N \\) for \\( N \\sim \\mathcal{N}(0, \\sigma_n^2\\mathbf{I}) \\) (right). The second and third row shows the reconstructions: (a)-(e) variance-scaled reverse diffusion process with different \\( \\lambda \\)'s; (f) PSCGAN with \\( N = 16 \\), \\( \\sigma_z = 1 \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.578, + 0.907, + 0.701 + ], + "angle": 0, + "content": "by our method versus PSCGAN. Our score-based method achieves a much larger range of tradeoffs compared to the GAN-based approach, demonstrating superior performance. Moreover, unlike PSCGAN, our method requires only a single score network to handle varying measurements and noise levels, offering greater flexibility. Additional experimental results on other distributions are provided in Appendix E." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.71, + 0.669, + 0.727 + ], + "angle": 0, + "content": "4.2 FFHQ dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.907, + 0.885 + ], + "angle": 0, + "content": "In this subsection, we evaluate our method on the FFHQ \\(256 \\times 256\\) dataset [9]. For different measurement scenarios, including Gaussian blur and super-resolution, we use the same pre-trained score-based model from [6], which was trained from scratch using 49k training data for 1M steps. We consider measurement scenarios with noise levels more severe than those in [6, 28, 29]. Specifically, for Gaussian blur, the images are blurred by a Gaussian blur kernel with size \\(61 \\times 61\\) and a standard deviation of 3.0. For super-resolution, the images are downsampled by a factor of 8." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.906, + 0.902 + ], + "angle": 0, + "content": "We compare our method against two benchmarks: (1)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2382" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.148, + 0.091, + 0.428, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.274, + 0.483, + 0.304 + ], + "angle": 0, + "content": "Figure 5. DP tradeoff on pinwheel dataset traversed by our variance-scaled reverse diffusion process and PSCGAN." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.312, + 0.484, + 0.465 + ], + "angle": 0, + "content": "We compare the DP tradeoffs provided by our score-based method and PSCGAN [15] on the Gaussian deblur task. Note that PSCGAN is not designed for other measurement scenarios and requires retraining for different noise levels. (2) We evaluate the performance of DiffPIR [29] to see where it falls within the tradeoff spectrum for different inverse problems. DiffPIR is a plug-and-play image restoration method based on diffusion denoising implicit model (DDIM) [18]. It has shown competitive performance on many denoising problems with fewer diffusion steps." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.484, + 0.293, + 0.497 + ], + "angle": 0, + "content": "4.2.1 Gaussian deblur task" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.483, + 0.613 + ], + "angle": 0, + "content": "Fig. 6 shows the DP tradeoff on the Gaussian deblur task. We test our variance-scaled reverse diffusion process, PSCGAN [15], and DiffPIR [29] on two different additive noise levels \\(\\sigma_{n} = 0.3\\) and \\(\\sigma_{n} = 0.5\\). Note that our sampling method and DiffPIR rely on a pre-trained score network on the FFHQ dataset. For PSCGAN, we use the model trained on images with additive Gaussian noise of \\(\\sigma_{n} = 0.3\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.81 + ], + "angle": 0, + "content": "For noise level \\(\\sigma_{n} = 0.3\\), PSCGAN shows suboptimal and more limited tradeoffs compared to our method. To obtain a reconstruction with similar MSE achieved in our \\(\\lambda = 0.9\\) case, PSCGAN requires averaging 64 sampled images. Its distortion degrades rapidly when fewer images are averaged. Our variance-scaled reverse diffusion sampling achieves both better fidelity and a broader range of DP tradeoffs. Moreover, when the models are tested on the higher noise level of \\(\\sigma_{n} = 0.5\\), the overall tradeoff shifts to the left. Our sampling method maintains effective tradeoff traversal. PSCGAN's performance (trained at \\(\\sigma_{n} = 0.3\\)) deteriorates significantly, showing much larger MSE and excessively high FID values." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Significant visual improvements can be observed in reconstructed samples in Fig. 7. When \\(\\lambda = 0\\), the restored images maintain high fidelity and faithfully reconstruct the original images, though appear relatively blurry. As \\(\\lambda\\) increases, the images become progressively sharper, revealing more fine details such as hair strands, eye corners, and back" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "ground elements. When \\(\\lambda = 1\\), while the overall reconstruction appears very natural, subtle details deviate from the original images. In comparison, both PSCGAN and DiffPIR have relatively good perceptual quality but compromise fidelity on crucial details. Notable deviations appear in features like eyes and mouth shapes, ultimately altering the facial expression and overall style of faces. We also report more metrics (PSNR and LPIPS) of Gaussian deblur task with additive noise level \\(\\sigma_{n} = 0.3\\) in Appendix E." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.246, + 0.719, + 0.261 + ], + "angle": 0, + "content": "4.2.2 Super-resolution task" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.269, + 0.907, + 0.555 + ], + "angle": 0, + "content": "Similar phenomena are observed in the super-resolution (SR) task. By adjusting different \\(\\lambda\\)'s, we can traverse a wide range of points on the DP plane, allowing flexible selection of the most suitable point for specific application scenarios. Fig. 8 shows the numerical results and reconstruction samples given by our sampling process and DiffPIR. Note that DiffPIR performs close to our \\(\\lambda = 1\\) case, and achieves a better FID score but larger distortion than our method. While DiffPIR demonstrates strong generative capabilities in both Gaussian deblur and super-resolution tasks with fewer diffusion steps [29], we examine more challenging degradation scenarios (\\(\\sigma_{n} = 0.3\\) and 0.5 for deblurring, and \\(8\\times\\) downscaling for SR). Under these severe conditions, DiffPIR's reduced diffusion steps lead to fidelity issues, with generated details less faithful compared to our \\(\\lambda = 1\\) case. Despite these differences, theoretical connections exist between DPS and DiffPIR [29]. We will consider the exploration of DDIM-style sampling for faster tradeoff traversal as future work." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.557, + 0.906, + 0.633 + ], + "angle": 0, + "content": "In summary, using a single pre-trained score network, the proposed sampling method can effectively traverse a more complete DP tradeoff and achieve better MSE than the benchmarks. Experimental details and more results are included in Appendix E." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.645, + 0.651, + 0.661 + ], + "angle": 0, + "content": "5 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.907, + 0.898 + ], + "angle": 0, + "content": "In this paper, we proposed a variance-scaled reverse diffusion process to traverse the DP tradeoff for general inverse problems using a single score-based model. By tuning a parameter that controls the scale of the reverse variance, we can navigate from the MMSE point to the whole posterior distribution, generating a complete DP curve. We proved that the proposed reverse sampling process serves as an optimal solution to the conditional DP tradeoff for multivariate Gaussian distribution. Meanwhile, we conducted experiments on the mixture Gaussian example, two-dimensional datasets, and the FFHQ dataset. Our results show that using a single pre-trained score network, we have achieved a more complete empirical DP tradeoff than the GAN-based method and other inverse problem solvers, demonstrating the effectiveness and flexibility of the proposed framework." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2383" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.178, + 0.092, + 0.824, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.295, + 0.908, + 0.366 + ], + "angle": 0, + "content": "Figure 6. DP tradeoff on FFHQ dataset. Note that the PSCGAN model in this figure is trained on images contaminated with additive Gaussian noise of \\(\\sigma_{n} = 0.3\\). It can be observed that when the noise level in inference time does match the trained model (e.g., when the noise level becomes \\(\\sigma_{n} = 0.5\\) during the test time), PSCGAN fails to provide valid reconstructions. On the other hand, our variance-scaled reverse diffusion process and DiffPIR method rely on a pre-trained score network on the FFHQ dataset, which is not dedicated to any noisy measurement and is robust for different noise levels." + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.371, + 0.868, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.628, + 0.86, + 0.644 + ], + "angle": 0, + "content": "Figure 7. Samples from our sampling process, PSCGAN, and DiffPIR for Gaussian deblur task with noise level \\(\\sigma_{n} = 0.3\\)." + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.66, + 0.868, + 0.846 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.857, + 0.871, + 0.873 + ], + "angle": 0, + "content": "Figure 8. Samples from our sampling process, PSCGAN, and DiffPIR for super-resolution task with downsampling scale \\(\\times 8\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2384" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.486, + 0.358 + ], + "angle": 0, + "content": "The work of Yuhan Wang and Ying-Jun Angela Zhang is supported in part by the General Research Fund (project number 14202421, 14214122, 14202723), Area of Excellence Scheme grant (project number AoE/E-601/22-R), and NSFC/RGC Collaborative Research Scheme (project number CRS_HKUST603/22), all from the Research Grants Council of Hong Kong. The work of Suzhi Bi is supported in part by the Natural Science Foundation of China under Project 62271325; in part by Guangdong Basic and Applied Basic Research Foundation under Project 2024B1515020089; in part by Shenzhen Science and Technology Program under Project 20220810142637001. The work of Xiaojun Yuan is supported in part by Sichuan Science and Technology Program under Grant 2024ZYD0036 and in part by National Key Laboratory of Wireless Communications Foundation under Grant IFN20230204." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.375, + 0.188, + 0.39 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.399, + 0.483, + 0.44 + ], + "angle": 0, + "content": "[1] Brian D.O. Anderson. Reverse-time diffusion equation models. Stochastic Processes and their Applications, 12(3):313-326, 1982." + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.442, + 0.483, + 0.485 + ], + "angle": 0, + "content": "[2] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6228-6237, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.487, + 0.484, + 0.541 + ], + "angle": 0, + "content": "[3] Yochai Blau and Tomer Michaeli. Rethinking lossy compression: The rate-distortion-perception tradeoff. In Proceedings of the 36th International Conference on Machine Learning (ICML), pages 675-685. PMLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.544, + 0.484, + 0.599 + ], + "angle": 0, + "content": "[4] Valentin De Bortoli, James Thornton, Jeremy Heng, and Arnaud Doucet. Diffusion schrödinger bridge with applications to score-based generative modeling. In Advances in Neural Information Processing Systems (NIPS), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.602, + 0.483, + 0.657 + ], + "angle": 0, + "content": "[5] Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. In Advances in Neural Information Processing Systems (NIPS), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.66, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[6] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In The Eleventh International Conference on Learning Representations (ICLR), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.731, + 0.483, + 0.784 + ], + "angle": 0, + "content": "[7] Dror Freirich, Tomer Michaeli, and Ron Meir. A theory of the distortion-perception tradeoff in wasserstein space. In Advances in Neural Information Processing Systems (NIPS), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.788, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Proceedings of the 34th International Conference on Neural Information Processing Systems (NIPS), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[9] Tero Karras, Samuli Laine, and Timo Aila. A Style-Based Generator Architecture for Generative Adversarial Networks. IEEE Transactions on Pattern Analysis & Machine Intelligence, 43(12):4217-4228, 2021." + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.399, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[10] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. In Advances in Neural Information Processing Systems (NIPS), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[11] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[12] Dong Liu, Haochen Zhang, and Zhiwei Xiong. On the classification-distortion-perception tradeoff. In Advances in Neural Information Processing Systems (NIPS). Curran Associates, Inc., 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.904, + 0.319 + ], + "angle": 0, + "content": "[13] Razvan V Marinescu, Daniel Moyer, and Polina Golland. Bayesian image reconstruction using deep generative models. arXiv preprint, 2021. [Online]. Available: https://arxiv.org/abs/2012.04567." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[14] Anush Krishna Moorthy and Alan Conrad Bovik. Blind image quality assessment: From natural scene statistics to perceptual quality. IEEE transactions on Image Processing, 20 (12):3350-3364, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[15] Guy Ohayon, Theo Adrai, Gregory Vaksman, Michael Elad, and Peyman Milanfar. High Perceptual Quality Image Denoising with a Posterior Sampling CGAN. In 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 1805-1813, Los Alamitos, CA, USA, 2021. IEEE Computer Society." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.904, + 0.489 + ], + "angle": 0, + "content": "[16] Victor M. Panaretos and Yoav Zemel. An Invitation to Statistics in Wasserstein Space. Springer Cham, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.904, + 0.531 + ], + "angle": 0, + "content": "[17] Slavko Simic. On a global upper bound for jensen's inequality. Journal of Mathematical Analysis and Applications, 343 (1):414-419, 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.575 + ], + "angle": 0, + "content": "[18] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations (ICLR), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.577, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[19] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In International Conference on Learning Representations (ICLR), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[20] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In Advances in Neural Information Processing Systems (NIPS). Curran Associates, Inc., 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[21] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations (ICLR), 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.76, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[22] Pascal Vincent. A connection between score matching and denoising autoencoders. *Neural Computation*, 23(7):1661-1674, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[23] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In The European Conference on Computer Vision Workshops (ECCVW), 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. In" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2385" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "The Eleventh International Conference on Learning Representations (ICLR), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.203 + ], + "angle": 0, + "content": "[25] Jay Whang, Maurizio Delbracio, Hossein Talebi, Chitwan Sahara, Alexandros G. Dimakis, and Peyman Milanfar. Deblurring via Stochastic Refinement. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16272-16282, Los Alamitos, CA, USA, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.206, + 0.483, + 0.261 + ], + "angle": 0, + "content": "[26] Xingyu Xu and Yuejie Chi. Provably robust score-based diffusion posterior sampling for plug-and-play image reconstruction. In The Thirty-eighth Annual Conference on Neural Information Processing Systems (NIPS), 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[27] Danna Xue, Luis Herranz, Javier Vazquez Corral, and Yanning Zhang. Burst perception-distortion tradeoff: Analysis and evaluation. In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.388 + ], + "angle": 0, + "content": "[28] Zhipeng Xue, Penghao Cai, Xiaojun Yuan, and Xiqi Gao. Score-based variational inference for inverse problems. arXiv preprint, 2024. [Online]. Available: https://arxiv.org/abs/2410.05646." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.472 + ], + "angle": 0, + "content": "[29] Yuanzhi Zhu, Kai Zhang, Jingyun Liang, Jiezhang Cao, Bihan Wen, Radu Timofte, and Luc Van Gool. Denoising Diffusion Models for Plug-and-Play Image Restoration. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1219–1229, Los Alamitos, CA, USA, 2023. IEEE Computer Society." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2386" + } + ] +] \ No newline at end of file diff --git a/2025/Traversing Distortion-Perception Tradeoff using a Single Score-Based Generative Model/dcdd686b-eeb1-4343-a7e6-20b5cd329198_origin.pdf b/2025/Traversing Distortion-Perception Tradeoff using a Single Score-Based Generative Model/dcdd686b-eeb1-4343-a7e6-20b5cd329198_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9ce7a2ea6b27bed3fe8c523b8c045b3bb46d0a31 --- /dev/null +++ b/2025/Traversing Distortion-Perception Tradeoff using a Single Score-Based Generative Model/dcdd686b-eeb1-4343-a7e6-20b5cd329198_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d33b1e7f95f4cb365d18d78b2506aaa9e2e9271341951a61a736205be1134f12 +size 1671998 diff --git a/2025/Traversing Distortion-Perception Tradeoff using a Single Score-Based Generative Model/full.md b/2025/Traversing Distortion-Perception Tradeoff using a Single Score-Based Generative Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d9b2b2ea7023c457265e61b09e49394c66e443b4 --- /dev/null +++ b/2025/Traversing Distortion-Perception Tradeoff using a Single Score-Based Generative Model/full.md @@ -0,0 +1,389 @@ +# Traversing Distortion-Perception Tradeoff using a Single Score-Based Generative Model + +Yuhan Wang $^{1}$ , Suzhi Bi $^{2*}$ , Ying-Jun Angela Zhang $^{1}$ , Xiaojun Yuan $^{3}$ + +1The Chinese University of Hong Kong 2Shenzhen University + +3University of Electronic Science and Technology of China + +{wy023, yjzhang}@ie.cuhk.edu.hk, bsz@szu.edu.cn, xjyuan@uestc.edu.cn + +# Abstract + +The distortion-perception (DP) tradeoff reveals a fundamental conflict between distortion metrics (e.g., MSE and PSNR) and perceptual quality. Recent research has increasingly concentrated on evaluating denoising algorithms within the DP framework. However, existing algorithms either prioritize perceptual quality by sacrificing acceptable distortion, or focus on minimizing MSE for faithful restoration. When the goal shifts or noisy measurements vary, adapting to different points on the DP plane needs retraining or even re-designing the model. Inspired by recent advances in solving inverse problems using score-based generative models, we explore the potential of flexibly and optimally traversing DP tradeoffs using a single pre-trained score-based model. Specifically, we introduce a variance-scaled reverse diffusion process and theoretically characterize the marginal distribution. We then prove that the proposed sample process is an optimal solution to the DP tradeoff for conditional Gaussian distribution. Experimental results on two-dimensional and image datasets illustrate that a single score network can effectively and flexibly traverse the DP tradeoff for general denoising problems. + +# 1 Introduction + +In recent years, we have witnessed rapid progress in image restoration algorithms, especially deep learning-based implementations that have shown remarkable achievements in solving general inverse problems. How to evaluate the performance of emerging algorithms is a crucial yet complicated problem. Traditional full reference distortion metrics, such as Mean Square Error (MSE), focus on the pixel-level accuracy between the original image $X$ and its reconstruction $\hat{X}$ . On the other hand, perceptual quality, referring to the degree to which an image looks natural rather than algorithmically generated [14], is also an important mea + +sure. It has been demonstrated that perceptual quality could be associated with the distance between the distributions of natural images $p_{X}$ and generated images $p_{\hat{X}}$ [2, 3, 12]. + +Improvements in distortion measures do not necessarily lead to enhancements in perceptual quality. In fact, [2] has demonstrated that there exists a tradeoff between distortion and perception, termed as distortion-perception (DP) trade-off. Mathematically, it can be modeled as + +$$ +D(P) = \min_{p_{\hat{X} |Y}}\mathbb{E}[\Delta (X,\hat{X})], +$$ + +$$ +\begin{array}{l} \text {s . t .} d (p _ {X}, p _ {\tilde {X}}) \leq P, \end{array} +$$ + +where $Y$ is the degraded observation, $\Delta : \mathcal{X} \times \hat{\mathcal{X}} \to \mathbb{R}^+$ is a full reference distortion function (e.g., such as square-error), and $d(\cdot, \cdot)$ is some divergence between probability distribution, such as Wasserstein distance [16] or Kulback-Leibler (KL) divergence. + +Since the introduction of the DP tradeoff, more and more algorithm designs have focused on performance evaluation on the DP plane for specific denoising tasks, e.g., image deblurring [15, 25] and super-resolution [11, 13, 23]. However, most algorithms tend to seek better perceptual quality at the cost of distortion, or focus on optimizing MSE, a distortion measure, to ensure restoration fidelity. When the restoration objective shifts between distortion and perception, we need to retrain or even re-design the network, which can be computationally intensive. In many practical scenarios, it is significant to flexibly and effectively fulfill diverse task objectives or user demands with a single model at inference time. + +There have been explorations into theories and algorithms to traverse the DP tradeoff in inference time [7, 15]. Specifically, in [15], the authors developed a framework based on conditional generative adversarial networks (CGAN) with an additional penalty on posterior expectation. By averaging different numbers of samples or adjusting the noise level injected into the generator, this CGAN-based framework can access various points along the tradeoff. However, using multiple image samples and averaging + +them may not be efficient during inference, and the method is not provably optimal. Furthermore, the training process relies on pairs of clean and noisy data. Thus, retraining is required for different noise levels and measurements. In [7], the DP tradeoff is studied in Wasserstein space. Theoretical results show that estimators on the DP curve can be constructed by linearly combining two estimators at the extremes: an optimal MSE minimizer and a perfect perception sampler with minimum distortion. This idea is further applied in burst restoration [27]. In practice, a method that achieves a relatively low MSE, alongside another method that achieves good perceptual quality, can serve as approximations of these two extreme estimators for a specific task. Nonetheless, different models must be deployed to address various measurements, rendering this method less flexible to general inverse problems. + +Recent research has demonstrated the generative capability of score-based diffusion models [4, 8, 10, 18, 20, 21] in tackling general inverse problems using a single score network [5, 6, 19, 24, 26, 29]. Score-based diffusion models learn the prior distribution of the data $\mathbf{x}$ by training a score network to match the gradient of the logarithm density $\nabla_{\mathbf{x}}\log p(\mathbf{x})$ , referred to as the score. After observing a noisy measurement $\mathbf{y}$ , sampling from posteriors involves approximating the conditional score $\nabla_{\mathbf{x}}\log p(\mathbf{x}|\mathbf{y})$ . Theoretical insights provided in [28] also reveal the potential for recovering Minimum Mean Square Error (MMSE) estimation by propagating the mean of the reverse diffusion process. Inspired by these advancements, we explore the potential of traversing DP tradeoffs for different tasks using a single score network. Our main contributions are summarized below: + +- First, we propose a variance-scaled reverse diffusion process and theoretically characterize the marginal distributions produced by this novel reverse process. It is demonstrated that the mean of our proposed sampling converges towards the MMSE point while the marginal covariance is scaled by a scaling factor. By tuning a parameter that dictates the variance of reverse sampling, we can flexibly navigate from the MMSE point to the posterior distribution, where the two extremes are achieved by setting the parameter to zero and one, respectively. +- Subsequently, we show that the reconstruction obtained from the variance-scaled reverse sampling represents the optimal solution to the conditional DP tradeoff for multivariate Gaussian distributions, assessed through MSE and Wasserstein-2 distance metrics. +- Finally, we validate our methodology through experiments conducted on various two-dimensional datasets and a real-world image dataset. With a single pre-trained score network, we can traverse the DP tradeoff across different inverse problems with varying noise levels. The results indicate that the variance-scaled reverse diffusion + +process achieves a more complete empirical DP tradeoff curve and better MSE than the GAN-based method and other inverse problem solvers, demonstrating the effectiveness and flexibility of the proposed framework. + +Notations: For a random variable $X$ denoted by a capital letter, we use small letter $\mathbf{x}$ to denote its realizations, and use $p_{X}(\mathbf{x})$ to denote the distribution over its alphabet $\mathcal{X}$ . When there is no ambiguity, the distribution can be abbreviated as $p(\mathbf{x})$ . For a discrete sequence of random variables $X_0, X_1, \dots, X_T$ , we abbreviate them as $\{X_k\}_{k=0}^T$ and use $\mathbf{x}_{0:T}$ to denote their realizations. The expectation and conditional expectation of $X$ given $Y = \mathbf{y}$ is $\mathbb{E}_{p(\mathbf{x})}[X]$ and $\mathbb{E}_{p(\mathbf{x}|\mathbf{y})}[X]$ . We use $\mathrm{Cov}_{p(\mathbf{x})}[X]$ and $\mathrm{Cov}_{p(\mathbf{x}|\mathbf{y})}[X]$ to represent the covariance and conditional covariance of $X$ given $Y = \mathbf{y}$ . Matrices are denoted by uppercase boldface letters (e.g., $\boldsymbol{\Sigma}$ ). $\mathrm{Tr}(\boldsymbol{\Sigma})$ and $\boldsymbol{\Sigma}^{-1}$ represent the trace and inverse. + +# 2 Backgrounds + +# 2.1 Score-based generative model + +Score-based generative models or diffusion models define the generative process as the reverse of a diffusion process. Specifically, the diffusion process (i.e., forward process) is a Markov chain $\{X_{k}\}_{k = 0}^{T}$ , with joint distribution + +$$ +p \left(\mathbf {x} _ {0: T}\right) := p \left(\mathbf {x} _ {0}\right) \prod_ {k = 0} ^ {T - 1} p \left(\mathbf {x} _ {k + 1} \mid \mathbf {x} _ {k}\right), \tag {1} +$$ + +which gradually diffuses the data $X_0 \in \mathbb{R}^d$ with distribution $p(\mathbf{x}_0) \triangleq p_{\mathrm{data}}$ . Variance preserving (VP) diffusion [8] adopts $p(\mathbf{x}_{k+1}|\mathbf{x}_k) \triangleq \mathcal{N}(\mathbf{x}_{k+1}; \sqrt{1 - \beta_{k+1}}\mathbf{x}_k, \beta_{k+1}\mathbf{I})$ with monotonically increasing variance schedule $\beta_T \geq \beta_{T-1} \geq \dots \geq \beta_0 = 0$ . When $T \to \infty$ , $\mathbf{x}_T$ converges to an isotropic Gaussian distribution $\mathcal{N}(\mathbf{0},\mathbf{I})$ . + +To generate a sample following the data distribution $p(\mathbf{x}_0)$ , we can start from the standard Gaussian $p_{\theta}(\mathbf{x}_T) \coloneqq \mathcal{N}(\mathbf{0}, \mathbf{I})$ and follow the reverse process + +$$ +p _ {\theta} \left(\mathbf {x} _ {0: T}\right) := p _ {\theta} \left(\mathbf {x} _ {T}\right) \prod_ {k = 0} ^ {T - 1} p _ {\theta} \left(\mathbf {x} _ {k} \mid \mathbf {x} _ {k + 1}\right), \tag {2} +$$ + +with transitions $p_{\theta}(\mathbf{x}_k|\mathbf{x}_{k + 1})\sim \mathcal{N}(\mu_{\theta}(\mathbf{x}_k,k),\boldsymbol {\Sigma}_{\theta}(\mathbf{x}_k,k))$ parameterized by $\theta$ . It will be shown later that the reverse sampling is closely related to the score function for each step, i.e., $\nabla_{\mathbf{x}_k}\log p(\mathbf{x}_k)$ , for $k = 0,1,\dots ,T$ + +In continuous-time, we can describe the diffusion process $(X_{t})_{t\in [0,T_{c}]}$ , $X_{t}\in \mathbb{R}^{d}$ with a stochastic differential equation (SDE) [21] + +$$ +d X _ {t} = - \frac {1}{2} \beta (t) X _ {t} d t + \sqrt {\beta (t)} d W _ {t}, X _ {0} \sim p _ {\text {d a t a}}, \tag {3} +$$ + +where $(W_{t})_{t\in [0,T_{c}]}$ is a standard Brownian motion, and $\beta (t)$ is the noise schedule. Let $(p_t)_{t\in [0:T_c]}$ be the path distribution associated with (3). To perform data sampling following the distribution $p_{\mathrm{data}}$ we can reverse the SDE (3) and + +apply discretization. According to [1] and [21], the reverse SDE corresponding to (3) is + +$$ +\begin{array}{l} d \overleftarrow {X} _ {t} = \left[ \frac {1}{2} \beta \left(T _ {c} - t\right) \overleftarrow {X} _ {t} + \beta \left(T _ {c} - t\right) \nabla \log p _ {T _ {c} - t} (\overleftarrow {X} _ {t}) \right] d t \\ + \sqrt {\beta (T _ {c} - t)} d \tilde {W} _ {t}, \overleftarrow {X} _ {0} \sim p _ {T _ {c}}, \tag {4} \\ \end{array} +$$ + +where $(\tilde{W}_t)_{t\in [0,T_c]}$ is another standard Brownian motion. The reverse SDE produces a time-reverse process $(\overleftarrow{X}_t)_{t\in [0,T_c]}$ where $\overleftarrow{X}_t$ has the same distribution with $X_{T_c - t}$ . Note that the drift function in (4) depends on the score function $\nabla_{\mathbf{x}_t}\log p_t(\mathbf{x}_t)$ , which can be approximated by a time-aware neural network trained with denoising score matching [22]. + +# 2.2 Score-based model for posterior sampling + +In many application scenarios, we may observe a noisy version of the measurement of the data $X_0$ , given by + +$$ +Y = \mathcal {A} (X _ {0}) + N, +$$ + +where $\mathcal{A}(\cdot):\mathbb{R}^d\to \mathbb{R}^n$ is a measurement operator and $N\in \mathbb{R}^n$ is the measurement noise with $N\sim \mathcal{N}(\mathbf{0},\sigma_n^2\mathbf{I})$ + +Score-based generative models have shown powerful capability in solving general inverse problems [6, 10, 19, 26, 28, 29]. Leveraging the diffusion model as the prior, sampling from the reverse posterior distribution + +$$ +p \left(\mathbf {x} _ {0: T} \mid \mathbf {y}\right) := p \left(\mathbf {x} _ {T} \mid \mathbf {y}\right) \prod_ {k = 0} ^ {T - 1} p \left(\mathbf {x} _ {k} \mid \mathbf {x} _ {k + 1}, \mathbf {y}\right) \tag {5} +$$ + +requires the knowledge of conditional scores $\nabla_{\mathbf{x}_k}\log p(\mathbf{x}_k|\mathbf{y})$ which can be expressed as $\nabla_{\mathbf{x}_k}\log p(\mathbf{x}_k|\mathbf{y}) = \nabla_{\mathbf{x}_k}\log p(\mathbf{x}_k) + \nabla_{\mathbf{x}_k}\log p(\mathbf{y}|\mathbf{x}_k)$ . The first term can be approximated by a score network $s_\theta (\mathbf{x}_k,k)$ [8, 21], and different methods have been proposed to estimate the second term $\nabla_{\mathbf{x}_k}\log p(\mathbf{y}|\mathbf{x}_k)$ [6, 19]. In each step the conditional sampling process in VP diffusion is [6, 19] + +$$ +\begin{array}{l} \mathbf {x} _ {k - 1} \leftarrow \frac {1}{\sqrt {\alpha_ {k}}} \left(\mathbf {x} _ {k} + (1 - \alpha_ {k}) \nabla_ {\mathbf {x} _ {k}} \log p (\mathbf {x} _ {k} | \mathbf {y})\right) + \tilde {\sigma} _ {k} \mathbf {z}, \tag {6} \\ \approx \frac {1}{\sqrt {\alpha_ {k}}} (\mathbf {x} _ {k} + (1 - \alpha_ {k}) (s _ {\theta} (\mathbf {x} _ {k}, k) + \nabla_ {\mathbf {x} _ {k}} \log p (\mathbf {y} | \mathbf {x} _ {k}))) + \tilde {\sigma} _ {k} \mathbf {z}, \\ \end{array} +$$ + +where $\mathbf{z} \sim \mathcal{N}(0, \mathbf{I})$ and $\tilde{\sigma}_k^2$ is the approximated variance of the reverse posterior distribution. + +Meanwhile, [28] proposed to directly approximate the reverse conditional $p(\mathbf{x}_k|\mathbf{x}_{k + 1},\mathbf{y})$ and propagate the mean in each step. Specifically, when $T\to \infty$ , the mean of $p(\mathbf{x}_k|\mathbf{x}_{k + 1},\mathbf{y})$ in VP-diffusion is given by [28] + +$$ +\boldsymbol {\mu} _ {k} \left(\mathbf {x} _ {k + 1}, \mathbf {y}\right) = \mathbf {U} _ {k} \mathbf {x} _ {k + 1} + \mathbf {V} _ {k} \mathbb {E} _ {p \left(\mathbf {x} _ {0} \mid \mathbf {y}\right)} [ X _ {0} ], \tag {7} +$$ + +where + +$$ +\begin{array}{l} \mathbf {U} _ {k} := \sqrt {\alpha_ {k + 1}} \left(\left(1 - \bar {\alpha} _ {k}\right) \mathbf {I} + \bar {\alpha} _ {k} \operatorname {C o v} _ {p (\mathbf {x} _ {0} | \mathbf {y})} [ X _ {0} ]\right) \\ \cdot \left(\left(1 - \bar {\alpha} _ {k + 1}\right) \mathbf {I} + \bar {\alpha} _ {k + 1} \operatorname {C o v} _ {p \left(\mathbf {x} _ {0} \mid \mathbf {y}\right)} [ X _ {0} ]\right) ^ {- 1}, \\ \end{array} +$$ + +$$ +\mathbf {V} _ {k} := \sqrt {\bar {\alpha} _ {k}} \big ((1 - \alpha_ {k + 1}) \mathbf {I} \big) \big ((1 - \bar {\alpha} _ {k + 1}) \mathbf {I} + \bar {\alpha} _ {k + 1} \operatorname {C o v} _ {p (\mathbf {x} _ {0} | \mathbf {y})} [ X _ {0} ] \big) ^ {- 1}. +$$ + +Here $\alpha_{k} = 1 - \beta_{k}$ and $\bar{\alpha}_{k} = \Pi_{i=0}^{k} \alpha_{i}$ in VP-diffusion. The authors proved that by propagating the mean at each reverse step (i.e., $\pmb{\mu}_{T} \rightarrow \pmb{\mu}_{T-1}(\mathbf{x}_{T} = \pmb{\mu}_{T}, \mathbf{y}) \rightarrow \dots \rightarrow \pmb{\mu}_{0}(\mathbf{x}_{1} = \pmb{\mu}_{1}, \mathbf{y})$ ), the end point $\pmb{\mu}_{0}(\mathbf{x}_{1} = \pmb{\mu}_{1}, \mathbf{y})$ represents the MMSE estimator $\mathbb{E}_{p(\mathbf{x}_0|\mathbf{y})}[X_0]$ when $\bar{\alpha}_{T} \rightarrow 0$ . The proof of reverse mean propagation converging to the MMSE point [28] and its connection to conditional score are included in Appendix A. + +# 2.3 Connection between reverse mean and conditional score + +The mean derived in (7) can be viewed as an approximation of the mean in (6). Specifically, when $T \to \infty$ , $p(\mathbf{x}_k|\mathbf{y})$ can be approximately expressed in the form of a Gaussian distribution. We can show that (see details in Appendix A) + +$$ +\frac {1}{\sqrt {\alpha_ {k}}} (\mathbf {x} _ {k} + (1 - \alpha_ {k}) \nabla_ {\mathbf {x} _ {k}} \log p (\mathbf {x} _ {k} | \mathbf {y})) = \mathbf {U} _ {k} \mathbf {x} _ {k + 1} + \mathbf {V} _ {k} \mathbb {E} _ {p (\mathbf {x} _ {0} | \mathbf {y})} [ X _ {0} ]. +$$ + +Theoretically, when $\tilde{\sigma}_k = 0$ in each step (6), the endpoint of the reverse process converges to the MMSE estimator, which achieves the best possible performance on distortion. When $\tilde{\sigma}_k$ is the true posterior variance, the reverse process samples from the posterior distribution $p(\mathbf{x}_{0:T}|\mathbf{y})$ , which results in a reconstruction with perfect perception measured in conditional distribution. The above observations motivate us to bridge the two extreme points with the score-based generative model by tuning the scale of $\sigma_k$ in the reverse process. It remains to be answered whether the bridging can provide us with an optimal and flexible estimator to traverse along the distortion-perception plane. + +# 3 Traversing DP Tradeoff with Scaled Reverse Diffusion + +In this section, we first propose a novel reverse sampling method based on a variance-scaled version of the joint inference distribution (5) and theoretically derive the mean and variance of the marginal distribution. Then, we prove that the resulting end distribution of the variance-scaled reverse sampling provides an optimal DP tradeoff with respect to MSE and Wasserstein-2 distance. + +# 3.1 Reverse mean and variance + +Consider the joint inference distribution with true mean but scaled variance, i.e., + +$$ +p _ {\lambda} \left(\mathbf {x} _ {k} \mid \mathbf {x} _ {k + 1}, \mathbf {y}\right) := \mathcal {N} \left(\boldsymbol {\mu} _ {k} \left(\mathbf {x} _ {k + 1}, \mathbf {y}\right), \lambda \mathbf {C} _ {k}\right), \tag {8} +$$ + +where the expectation $\pmb{\mu}_k(\mathbf{x}_{k + 1},\mathbf{y})$ is given in (7), and the true covariance $\mathbf{C}_k$ is given by [28] + +$$ +\begin{array}{l} \mathbf {C} _ {k} := \frac {\beta_ {k + 1}}{1 - \beta_ {k + 1}} \left((1 - \bar {\alpha} _ {k}) \mathbf {I} + \bar {\alpha} _ {k} \operatorname {C o v} _ {p (\mathbf {x} _ {0} | \mathbf {y})} [ X _ {0} ]\right) \\ \cdot \left(\left(\frac {\beta_ {k + 1}}{1 - \beta_ {k + 1}} + 1 - \bar {\alpha} _ {k}\right) \mathbf {I} + \bar {\alpha} _ {k} \operatorname {C o v} _ {p (\mathbf {x} _ {0} | \mathbf {y})} [ X _ {0} ]\right) ^ {- 1}. \tag {9} \\ \end{array} +$$ + +We have the following theorem to characterize the marginal distribution given by the variance-scaled version of the joint inference distribution. + +Theorem 1. For VP-diffusion and $0 \leq \lambda \leq 1$ , consider the joint inference distribution given by + +$$ +p _ {\lambda} \left(\mathbf {x} _ {0: T} \mid \mathbf {y}\right) = p _ {\lambda} \left(\mathbf {x} _ {T} \mid y\right) \prod_ {k = 0} ^ {T - 1} p _ {\lambda} \left(\mathbf {x} _ {k} \mid \mathbf {x} _ {k + 1}, \mathbf {y}\right) \tag {10} +$$ + +where $p_{\lambda}(\mathbf{x}_T|\mathbf{y}) = \mathcal{N}(0,\mathbf{I})$ and $p_{\lambda}(\mathbf{x}_k|\mathbf{x}_{k + 1},\mathbf{y})$ is given by (8). Then, the corresponding margin has the distribution $p_{\lambda}(\mathbf{x}_k|\mathbf{y}) = \mathcal{N}(\boldsymbol {\mu}_k^\lambda ,\boldsymbol{\Sigma}_k^\lambda)$ where + +$$ +\begin{array}{l} \boldsymbol {\mu} _ {k} ^ {\lambda} = \left(\mathbf {U} _ {k} \left(\mathbf {U} _ {k + 1} (\dots (\mathbf {U} _ {T - 2} \mathbf {V} _ {T - 1} + \mathbf {V} _ {T - 2}) \dots) + \mathbf {V} _ {k + 1}\right) + \mathbf {V} _ {k}\right) \\ = \sqrt {\bar {\alpha} _ {k}} (1 - \alpha_ {k + 1} \cdot \cdot \cdot \alpha_ {T}) \left((1 - \bar {\alpha} _ {T}) \mathbf {I} + \bar {\alpha} _ {T} C o v _ {p (\mathbf {x} _ {0} | \mathbf {y})} [ X _ {0} ]\right) ^ {- 1} \mathbb {E} _ {p (\mathbf {x} _ {0} | \mathbf {y})} [ X _ {0} ], \\ \boldsymbol {\Sigma} _ {k} ^ {\lambda} = \boldsymbol {\Sigma} _ {k} \Big (\lambda \mathbf {I} + (1 - \lambda) \alpha_ {k + 1} \alpha_ {k + 2} \dots \alpha_ {T} \boldsymbol {\Sigma} _ {T - 1} ^ {- 1} \boldsymbol {\Sigma} _ {k} \Big), \\ \end{array} +$$ + +where $\pmb{\Sigma}_k$ is the covariance of $p(\mathbf{x}_k|\mathbf{y})$ in VP diffusion. In particular, when $\bar{\alpha}_T\rightarrow 0$ , the variance of $p_{\lambda}(\mathbf{x}_0|\mathbf{y})$ is + +$$ +\pmb {\Sigma} _ {0} ^ {\lambda} = \pmb {\Sigma} _ {0} \bigg (\lambda \mathbf {I} + (1 - \lambda) \bar {\alpha} _ {T} \pmb {\Sigma} _ {T - 1} ^ {- 1} \pmb {\Sigma} _ {0} \bigg) \rightarrow \lambda C o v _ {p (\mathbf {x} _ {0} | \mathbf {y})} [ X _ {0} ], +$$ + +and the mean is + +$$ +\begin{array}{l} \mu_ {0} ^ {\lambda} = \sqrt {\bar {\alpha} _ {0}} \big ((1 - \bar {\alpha} _ {T}) \mathbf {I} \big) \big ((1 - \bar {\alpha} _ {T}) \mathbf {I} + \bar {\alpha} _ {T} C o v _ {p (\mathbf {x} _ {0} | \mathbf {y})} [ X _ {0} ] \big) ^ {- 1} \\ \cdot \mathbb {E} _ {p (\mathbf {x} _ {0} | \mathbf {y})} [ X _ {0} ] \rightarrow \mathbb {E} _ {p (\mathbf {x} _ {0} | \mathbf {y})} [ X _ {0} ]. \\ \end{array} +$$ + +Proof. See the details in Appendix B. + +The above theorem shows that the proposed variance-scaled reverse diffusion process will lead to a scaled marginal distribution at each step, which finally results in the posterior mean and scaled posterior variance of the original data distribution at the end of the reverse process. This seemingly intuitive result actually provides the optimal solution to the conditional distortion-perception tradeoff for multivariate Gaussian distribution. + +# 3.2 DP tradeoff for conditional Gaussian case + +Consider the $d$ -dimensional source $X \in \mathbb{R}^d$ with distribution $p(\mathbf{x})$ . Let $Y$ be the noisy version of the source data. Denote the conditional expectation and variance given an observation $\mathbf{y}$ as $\boldsymbol{\mu}_{\mathbf{y}} = \mathbb{E}_{p(\mathbf{x}|\mathbf{y})}[X]$ and $\boldsymbol{\Sigma}_{\mathbf{y}} = \mathrm{Cov}_{p(\mathbf{x}|\mathbf{y})}[X]$ , respectively. The goal of signal restoration is to find a good reconstruction $\hat{X}$ based on the observed $Y$ . + +From the optimization perspective, given a specific observation $Y = \check{\mathbf{y}}$ , we can define the conditional distortion-perception function as + +$$ +\begin{array}{l} D (P) = \min _ {p _ {\hat {X} | Y} (\hat {\mathbf {x}} | \check {\mathbf {y}})} \mathbb {E} _ {p _ {X, \hat {X} | Y} (\mathbf {x}, \hat {\mathbf {x}} | \check {\mathbf {y}})} [ \| X - \hat {X} \| _ {2} ^ {2} ] \\ \begin{array}{l} \text {s . t .} d (p _ {X | Y} (\mathbf {x} | \breve {\mathbf {y}}), p _ {\hat {X} | Y} (\hat {\mathbf {x}} | \breve {\mathbf {y}})) \leq P, \end{array} \\ \end{array} +$$ + +where $X, Y$ and $\hat{X}$ form a Markov chain. + +Theorem 2. Consider the source $X$ and an observation $\tilde{\mathbf{y}}$ that satisfy $p_{X|Y}(\mathbf{x}|\tilde{\mathbf{y}}) \sim \mathcal{N}(\boldsymbol{\mu}_{\tilde{\mathbf{y}}}, \boldsymbol{\Sigma}_{\tilde{\mathbf{y}}})$ . The optimal distortion-perception tradeoff given $\tilde{\mathbf{y}}$ with MSE and Wasserstein-2 distance is + +$$ +D (P) = \left\{ \begin{array}{l} \operatorname {T r} \left(\boldsymbol {\Sigma} _ {\tilde {\mathbf {y}}}\right) + \left(\sqrt {\operatorname {T r} \left(\boldsymbol {\Sigma} _ {\tilde {\mathbf {y}}}\right)} - P\right) ^ {2}, f o r P \leq \sqrt {\operatorname {T r} \left(\boldsymbol {\Sigma} _ {\tilde {\mathbf {y}}}\right)} \\ \operatorname {T r} \left(\boldsymbol {\Sigma} _ {\tilde {\mathbf {y}}}\right), f o r P > \sqrt {\operatorname {T r} \left(\boldsymbol {\Sigma} _ {\tilde {\mathbf {y}}}\right)}. \end{array} \right. \tag {11} +$$ + +Meanwhile, the optimal tradeoff can be achieved by sampling from the joint inference distribution (10) in Theorem 1 and traversing different $0 \leq \lambda \leq 1$ . + +Proof. See details in Appendix C. + +![](images/9cb66dac147096be4a2598fb33943fa5a6d200cee83c024859798864e2583017.jpg) + +Theorem 2 explicitly characterizes the optimal DP tradeoff for conditional Gaussian distribution. When $P = 0$ (i.e., perfect perceptual quality is expected), the best achievable MSE is $2\mathrm{Tr}(\Sigma_{\bar{\mathbf{y}}})$ . As $P$ increases, the optimal MSE gradually converges to the MMSE value $\mathrm{Tr}(\Sigma_{\bar{\mathbf{y}}})$ . The proof in Appendix C also shows that the variance-scaled reverse diffusion process can attain the entire optimal DP curve for conditional Gaussian distribution. Although the actual data distribution may not follow the conditional Gaussian distribution (e.g., mixture Gaussian, 2D distribution like S-curve, and image dataset), the forthcoming experiments will show that the proposed variance-scaled reverse sampling can effectively and flexibly traverse the DP tradeoff on general datasets using a single score network. + +# 3.3 Variance-scaled reverse sampling process + +In this subsection, we discuss how to perform variance-scaled reverse sampling by estimating the conditional score. With a single pre-trained score network, we can traverse different DP tradeoffs for general inverse problems. + +Recall that the proposed variance-scaled reverse sampling process is given by $p_{\lambda}(\mathbf{x}_{0:T}|\mathbf{y}) = p_{\lambda}(\mathbf{x}_T|y)\prod_{k = 0}^{T - 1}p_{\lambda}(\mathbf{x}_k|\mathbf{x}_{k + 1},\mathbf{y})$ where for VP diffusion, we have the posteriors $p_{\lambda}(\mathbf{x}_{k - 1}|\mathbf{x}_k,\mathbf{y})\coloneqq \mathcal{N}\big(\boldsymbol{\mu}_{k - 1}(\mathbf{x}_k,\mathbf{y}),\lambda \mathbf{C}_{k - 1}\big)$ . The sampling process in each step can be approximately expressed as + +$$ +\mathbf {x} _ {k - 1} = \frac {1}{\sqrt {\alpha_ {k}}} \left(\mathbf {x} _ {k} + (1 - \alpha_ {k}) \nabla_ {\mathbf {x} _ {k}} \log p \left(\mathbf {x} _ {k} \mid \mathbf {y}\right)\right) + \lambda \mathbf {C} _ {k} \mathbf {z}, \tag {12} +$$ + +where $\mathbf{z} \sim \mathcal{N}(0, \mathbf{I})$ . To draw a sample $\mathbf{x}_{k-1}$ according to (12), we need to compute the conditional score $\nabla_{\mathbf{x}_k} \log p(\mathbf{x}_k | \mathbf{y})$ . Specifically, the conditional scores can be decomposed as $\nabla_{\mathbf{x}_k} p(\mathbf{x}_k | \mathbf{y}) = \nabla_{\mathbf{x}_k} p(\mathbf{x}_k) + \nabla_{\mathbf{x}_k} p(\mathbf{y} | \mathbf{x}_k)$ by Bayes' rule, where the first term can be approximated by a score network $s_\theta(\mathbf{x}_k, k)$ [8, 21]. To deal with the second term, Denoising Posterior Sampling (DPS) [6] proposes to approximate $p(\mathbf{y} | \mathbf{x}_k) = \mathbb{E}_{p(\mathbf{x}_0 | \mathbf{x}_k)}[p(\mathbf{y} | \mathbf{x}_0)]$ with $p(\mathbf{y} | \hat{\mathbf{x}}_0(\mathbf{x}_k))$ , where + +$$ +\hat {\mathbf {x}} _ {0} \left(\mathbf {x} _ {k}\right) = \mathbb {E} _ {p \left(\mathbf {x} _ {0} \mid \mathbf {x} _ {k}\right)} \left[ X _ {0} \right] = \frac {1}{\alpha_ {k}} \left(\mathbf {x} _ {k} + \left(1 - \bar {\alpha} _ {k}\right) \nabla_ {\mathbf {x} _ {k}} p \left(\mathbf {x} _ {k}\right)\right) \tag {13} +$$ + +by Tweedie's formula. The approximation error here is upper-bounded by the Jensen gap [17]. Note that $p(\mathbf{y}|\hat{\mathbf{x}}_0(\mathbf{x}_k))$ is analytically tractable when the measurement distribution is given. Specifically, when the measurement nosie is $\mathcal{N}(0,\sigma_n^2\mathbf{I})$ , the second term in the conditional score takes the form of $\nabla_{\mathbf{x}_k}\log p(\mathbf{y}|\mathbf{x}_k) = \frac{1}{\sigma_n^2}\nabla_{\mathbf{x}_k}||\mathbf{y}-\mathcal{A}(\hat{\mathbf{x}}_0(\mathbf{x}_k))||_2^2$ . Thus, the conditional scores can be approximated as $\nabla_{\mathbf{x}_k}\log p(\mathbf{x}_k|\mathbf{y}) = s_\theta (\mathbf{x}_k,k) + \frac{1}{\sigma_n^2}\nabla_{\mathbf{x}_k}||\mathbf{y}-\mathcal{A}(\hat{\mathbf{x}}_0(\mathbf{x}_k))||_2^2$ . + +In the following experiments, we adopt the DPS framework to approximate the conditional score and sample from (12), which is sketched as Algorithm 1. + +Algorithm 1 Variance-Scaled Reverse Sampling with DPS +Require: $T,\mathbf{y},\lambda ,\{\tilde{\sigma}_k\}_{k = 0}^T,\{\zeta_{k,\lambda}\}_{k = 0}^T$ $\mathbf{x}_T\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ +for $k = T - 1$ to 1 do do $\hat{s}\gets s_{\theta}(\mathbf{x}_k,k)$ $\hat{\mathbf{x}}_0\gets \frac{1}{\sqrt{\bar{\alpha}_k}} (\mathbf{x}_k + (1 - \bar{\alpha}_k)\hat{s})$ $\hat{c} (\hat{\mathbf{x}}_0) = -\frac{\sqrt{\alpha_k}}{1 - \alpha_k}\nabla_{\mathbf{x}_k}||y - \mathcal{A}(\hat{\mathbf{x}}_0(\mathbf{x}_k))||_2^2$ $\mathbf{z}\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ $\mathbf{x}_{k - 1}\leftarrow \frac{1}{\sqrt{\alpha_k}}\bigl (\mathbf{x}_k + (1 - \alpha_k)\bigl (\hat{s} +\zeta_{k,\lambda}\hat{c} (\hat{\mathbf{x}}_0)\bigr)\bigr) + \lambda \tilde{\sigma}_k\mathbf{z}$ +end for +return $\mathbf{x}_0$ + +In Algorithm 1, $\tilde{\sigma}_k$ is set to $\beta_{k}$ [8, 28], and $\zeta_{k,\lambda}$ is a hyperparameter to control the weight of the conditional score, and may differ for different $\lambda$ . The heuristic choices of $\zeta_{k,\lambda}$ and the rationale for fine-tuning have been included in Appendix E. + +Remark 1. In this paper, we adopt the conditional score approximation proposed in [6] for illustration. To implement the variance-scaled reverse sampling process (12), one can employ any approach to directly estimate the conditional score [19] or estimate the posterior mean as a whole [28] for general inverse problems. For example, Pseudoinverse-guided Diffusion Models (ΠGDM) [19] approximated the unconditional posteriors $p(\mathbf{x}_0|\mathbf{x}_k)$ with Gaussian, i.e., $p(\mathbf{x}_0|\mathbf{x}_k) \approx \mathcal{N}(\hat{\mathbf{x}}_0,r_k^2\mathbf{I})$ , where $\hat{\mathbf{x}}_0$ is given by (13), and $r_k^2$ is the hyperparameter. Thus, the score term $\nabla_{\mathbf{x}_k}p(\mathbf{y}|\mathbf{x}_k)$ can be analytically expressed by the pseudoinverse of the measurement model. In [28], the posterior mean is estimated as a whole. Specifically, we have $\mu_{k - 1}(\mathbf{x}_k,\mathbf{y}) = \mathbf{U}_{k - 1}\mathbf{x}_k + \mathbf{V}_{k - 1}\mathbb{E}_{p(\mathbf{x}_0|\mathbf{y})}[X_0]$ . Since the expression involves $\mathbb{E}_{p(\mathbf{x}_0|\mathbf{y})}[X_0]$ and $\mathrm{Cov}_{p(\mathbf{x}_0|\mathbf{y})}[X_0]$ , it can not be computed directly. Instead, the authors approximate the joint conditional posterior with variational Gaussian distributions and apply natural gradient descent to perform the sampling. Note that the gradient calculation also involves the approximation of the conditional score $\nabla_{\mathbf{x}_k}\log p(\mathbf{y}|\mathbf{x}_k)$ as in DPS [6]. + +![](images/e1ee14930088bcc594c2b6a18eae8fb0a7d44a986c5ff556883074c675e8c058.jpg) +(a) $p(x_0)$ +Figure 1. An example of mixture Gaussian, the noisy observation, and the conditional distribution given an observation $y = -0.6$ + +![](images/c100d554646e05dc0617d22319c1555cd6b1c6d7a8be2e6e6423ee9d3eb2a635.jpg) +(b) $p(y)$ + +![](images/6cc451c67db869511cf8be3cf9c3918fa38aa66220bb4fe8d9fa51b5a55b464c.jpg) +(c) $p(x|y = -0.6)$ + +# 3.4 An illustrative example + +Consider the mixture Gaussian distribution $X_0 \sim p(x_0)$ with two components, where $p(x_0) = w_1 \mathcal{N}(\mu_1, \sigma_1^2) + w_2 \mathcal{N}(\mu_2, \sigma_2^2)$ . The noisy observation is obtained by $Y = aX_0 + \sigma_0 \epsilon$ , where $\epsilon \sim \mathcal{N}(0, 1)$ , i.e., $p(y|x_0) = \mathcal{N}(ax_0, \sigma_0^2)$ . The MMSE estimator and posterior distribution can be theoretically derived (see Appendix D for details). Fig. 1 shows an example of mixture Gaussian distribution and the conditional distribution given an observation $y = -0.6$ . + +Given an observation $y$ and a chosen $\lambda$ , we iteratively perform the variance scaled reverse process (12) for $k = T, \dots, 1$ , where the variance is approximated by $\mathbf{C}_k = \beta_k$ as the conventional VP-diffusion. Starting from a random sample $x_T \sim \mathcal{N}(0,1)$ , Fig. 2 shows the trajectories $x_T \to \dots x_k \to \dots x_0$ of multiple reconstructions for each $\lambda$ . We can observe that for $\lambda = 0$ , given an initial $x_T$ , the trajectories are deterministic and converge to the MMSE point. This phenomenon collaborates with results in [28]. When $\lambda$ increases, the generated trajectories follow the form of the posterior distribution and show more stochasticity. When $\lambda = 1$ , the reconstruction distribution matches the ground truth posterior. + +The resulting DP tradeoffs are shown in Fig. 3. It can be observed that as $\lambda$ increases, the divergence between the true posterior distribution and reconstruction distribution continuously decreases, with an increase of MSE value. The overall curve is convex, and both Wasserstein-2 and KL-divergence come to zero. Meanwhile, the MSE achieved at $\lambda = 1$ is approximately two times the MSE at $\lambda = 0$ for both Wasserstein-2 and KL-divergence cases, coinciding with the result in Theorem 2. + +# 4 Experimental Results + +In this section, we conduct a series of experiments on two-dimensional datasets and the FFHQ dataset [9]. It is demonstrated that the variance-scaled reverse sampling shows great flexibility and effectiveness in traversing DP tradeoff compared to the GAN-based approach and other denoising methods. + +# 4.1 Two-dimensional points example + +First, we evaluate the validity and effectiveness of our approach on two-dimensional datasets, including pinwheel, S- + +![](images/44fd470180d9b8a7fa4c35dafe0a9fd4baf1afd510d8e21b8a790f4d5ea4cb47.jpg) +Figure 2. Trajectories $x_{T} \to \dots x_{k} \to \dots \to x_{0}$ of different reconstructions for $\lambda = 0, 0.3, 0.8$ and 1. The initial $x_{T}$ is $\mathcal{N}(0,1)$ . + +![](images/67e673cf40bd32b441cd2a9307d4404578d99a9f7a39d23d9448828cfbf0940e.jpg) +(a) + +![](images/6643881d2c6b59afaab83c3244091818b1463bbc20a10993a504b8116e0820f6.jpg) +(b) +Figure 3. Distortion-perception tradeoff traversed by variance-scaled reverse sampling given different $\lambda$ 's. (a) tradeoff between Wasserstein-2 distance and MSE; (b) tradeoff between KL-divergence and MSE. + +curve, and moon-shape data distributions. Taking the pinwheel dataset as an example, the first row of Fig. 4 shows the distribution of pinwheel data points $X_0$ and the noisy observation $Y$ , where $Y = aX + N$ for $N \sim \mathcal{N}(0, \sigma_n^2 \mathbf{I})$ . + +For each dataset, we train a conventional score network to approximate $p(\mathbf{x}_0)$ on the original distribution. Then, we perform the variance-scaled reverse diffusion process (12) on the noisy observation. As a comparison, we adopt a GAN-based DP-traversing approach PSCGAN [15]. This scheme introduces a penalty of the posterior expectation to the training of the conditional GAN. The authors proposed two strategies to navigate the DP tradeoff during the inference time: (1) PSCGAN- $N$ : This method involves sampling $N$ instances from the generator and then averaging them. As $N$ increases, the averaged image is closer to the conditional expectation, leading to a smaller distortion but larger perception loss. (2) PSCGAN- $z$ : It varies the standard deviation $\sigma_z$ of the noise injected into the generator to control the stochasticity. For $0 \leq \sigma_z \leq 1$ , a larger $\sigma_z$ results in a better perceptual quality and higher MSE. Note that the training of PSCGAN relies on noisy data points $Y$ . Thus, the model needs to be retrained for different measurement scenarios and different levels of $\sigma_n^2$ . + +Fig. 4 shows the reconstruction of our variance-scaled reverse diffusion process for different $\lambda$ 's, as well as the reconstruction of PSCGAN when setting $N = 16$ , $\sigma_z = 1$ . It can be observed that when $\lambda = 0$ , our sampling process leads to a more concentrated reconstruction, and gradually approaches the true distribution as $\lambda$ increases. + +Fig. 5 compares the numerical DP tradeoffs achieved + +![](images/7367742c11f73a719eb2cf15c2f2a74b107fabb9b8400d1bf8d20b7f86288a4d.jpg) +Original distributions + +![](images/9fcc744ec54d2fb0e6ea48623cd5ecd146a42bc0f1b99226f7ef6647329c8239.jpg) +Noisy observations + +![](images/0c0f5d0e3cc24a89e68c22f69f79b27cd4f97326eddeb6687fb3defc20da2582.jpg) +(a) $\lambda = 0$ + +![](images/7c594d082776a16799d3443f4016f40ca0bbce69bdc87eede7ac0d11729a4b53.jpg) +(b) $\lambda = 0.3$ + +![](images/4ebdcc77cb63da8a32d060e277b5b076206a1c265e66b44ad2a9dd7cfac6f8dc.jpg) +(c) $\lambda = 0.5$ + +![](images/39324c923ce22887d83fb8bc80d0dc674cd9f747af6f949d69b1528ded7040d6.jpg) +(d) $\lambda = 0.8$ +Figure 4. Experiments on a two-dimensional dataset. The first row illustrates the original distribution (left) of pinwheel and the noisy observation $Y$ , where $Y = aX + N$ for $N \sim \mathcal{N}(0, \sigma_n^2\mathbf{I})$ (right). The second and third row shows the reconstructions: (a)-(e) variance-scaled reverse diffusion process with different $\lambda$ 's; (f) PSCGAN with $N = 16$ , $\sigma_z = 1$ . + +![](images/4d2d216814d00230e7e35e510af8de5e397cb84a79e82ffbce262d9f516fc56d.jpg) +(e) $\lambda = 1$ + +![](images/f66f08aa9b98c16df988e8d174f09edf0847f6ca9e9ec251ad974f7d7c136c1c.jpg) +(f) PSCGAN $N = 16, \sigma_z = 1$ + +by our method versus PSCGAN. Our score-based method achieves a much larger range of tradeoffs compared to the GAN-based approach, demonstrating superior performance. Moreover, unlike PSCGAN, our method requires only a single score network to handle varying measurements and noise levels, offering greater flexibility. Additional experimental results on other distributions are provided in Appendix E. + +# 4.2 FFHQ dataset + +In this subsection, we evaluate our method on the FFHQ $256 \times 256$ dataset [9]. For different measurement scenarios, including Gaussian blur and super-resolution, we use the same pre-trained score-based model from [6], which was trained from scratch using 49k training data for 1M steps. We consider measurement scenarios with noise levels more severe than those in [6, 28, 29]. Specifically, for Gaussian blur, the images are blurred by a Gaussian blur kernel with size $61 \times 61$ and a standard deviation of 3.0. For super-resolution, the images are downsampled by a factor of 8. + +We compare our method against two benchmarks: (1) + +![](images/e3ff47821395411203f0fcd2c7141da420328371aa43a9658e35d5f360146ce0.jpg) +Figure 5. DP tradeoff on pinwheel dataset traversed by our variance-scaled reverse diffusion process and PSCGAN. + +We compare the DP tradeoffs provided by our score-based method and PSCGAN [15] on the Gaussian deblur task. Note that PSCGAN is not designed for other measurement scenarios and requires retraining for different noise levels. (2) We evaluate the performance of DiffPIR [29] to see where it falls within the tradeoff spectrum for different inverse problems. DiffPIR is a plug-and-play image restoration method based on diffusion denoising implicit model (DDIM) [18]. It has shown competitive performance on many denoising problems with fewer diffusion steps. + +# 4.2.1 Gaussian deblur task + +Fig. 6 shows the DP tradeoff on the Gaussian deblur task. We test our variance-scaled reverse diffusion process, PSCGAN [15], and DiffPIR [29] on two different additive noise levels $\sigma_{n} = 0.3$ and $\sigma_{n} = 0.5$ . Note that our sampling method and DiffPIR rely on a pre-trained score network on the FFHQ dataset. For PSCGAN, we use the model trained on images with additive Gaussian noise of $\sigma_{n} = 0.3$ . + +For noise level $\sigma_{n} = 0.3$ , PSCGAN shows suboptimal and more limited tradeoffs compared to our method. To obtain a reconstruction with similar MSE achieved in our $\lambda = 0.9$ case, PSCGAN requires averaging 64 sampled images. Its distortion degrades rapidly when fewer images are averaged. Our variance-scaled reverse diffusion sampling achieves both better fidelity and a broader range of DP tradeoffs. Moreover, when the models are tested on the higher noise level of $\sigma_{n} = 0.5$ , the overall tradeoff shifts to the left. Our sampling method maintains effective tradeoff traversal. PSCGAN's performance (trained at $\sigma_{n} = 0.3$ ) deteriorates significantly, showing much larger MSE and excessively high FID values. + +Significant visual improvements can be observed in reconstructed samples in Fig. 7. When $\lambda = 0$ , the restored images maintain high fidelity and faithfully reconstruct the original images, though appear relatively blurry. As $\lambda$ increases, the images become progressively sharper, revealing more fine details such as hair strands, eye corners, and back + +ground elements. When $\lambda = 1$ , while the overall reconstruction appears very natural, subtle details deviate from the original images. In comparison, both PSCGAN and DiffPIR have relatively good perceptual quality but compromise fidelity on crucial details. Notable deviations appear in features like eyes and mouth shapes, ultimately altering the facial expression and overall style of faces. We also report more metrics (PSNR and LPIPS) of Gaussian deblur task with additive noise level $\sigma_{n} = 0.3$ in Appendix E. + +# 4.2.2 Super-resolution task + +Similar phenomena are observed in the super-resolution (SR) task. By adjusting different $\lambda$ 's, we can traverse a wide range of points on the DP plane, allowing flexible selection of the most suitable point for specific application scenarios. Fig. 8 shows the numerical results and reconstruction samples given by our sampling process and DiffPIR. Note that DiffPIR performs close to our $\lambda = 1$ case, and achieves a better FID score but larger distortion than our method. While DiffPIR demonstrates strong generative capabilities in both Gaussian deblur and super-resolution tasks with fewer diffusion steps [29], we examine more challenging degradation scenarios ( $\sigma_{n} = 0.3$ and 0.5 for deblurring, and $8\times$ downscaling for SR). Under these severe conditions, DiffPIR's reduced diffusion steps lead to fidelity issues, with generated details less faithful compared to our $\lambda = 1$ case. Despite these differences, theoretical connections exist between DPS and DiffPIR [29]. We will consider the exploration of DDIM-style sampling for faster tradeoff traversal as future work. + +In summary, using a single pre-trained score network, the proposed sampling method can effectively traverse a more complete DP tradeoff and achieve better MSE than the benchmarks. Experimental details and more results are included in Appendix E. + +# 5 Conclusions + +In this paper, we proposed a variance-scaled reverse diffusion process to traverse the DP tradeoff for general inverse problems using a single score-based model. By tuning a parameter that controls the scale of the reverse variance, we can navigate from the MMSE point to the whole posterior distribution, generating a complete DP curve. We proved that the proposed reverse sampling process serves as an optimal solution to the conditional DP tradeoff for multivariate Gaussian distribution. Meanwhile, we conducted experiments on the mixture Gaussian example, two-dimensional datasets, and the FFHQ dataset. Our results show that using a single pre-trained score network, we have achieved a more complete empirical DP tradeoff than the GAN-based method and other inverse problem solvers, demonstrating the effectiveness and flexibility of the proposed framework. + +![](images/d7cfc4e12adeead9a8e29f2db1ad310908f5f7589eba88437e7b4fad9aa99132.jpg) +Figure 6. DP tradeoff on FFHQ dataset. Note that the PSCGAN model in this figure is trained on images contaminated with additive Gaussian noise of $\sigma_{n} = 0.3$ . It can be observed that when the noise level in inference time does match the trained model (e.g., when the noise level becomes $\sigma_{n} = 0.5$ during the test time), PSCGAN fails to provide valid reconstructions. On the other hand, our variance-scaled reverse diffusion process and DiffPIR method rely on a pre-trained score network on the FFHQ dataset, which is not dedicated to any noisy measurement and is robust for different noise levels. + +![](images/4e93a5f90a08a1dc6e4f75922c8188002972962e1975fe474b97c7217abac7c2.jpg) +Figure 7. Samples from our sampling process, PSCGAN, and DiffPIR for Gaussian deblur task with noise level $\sigma_{n} = 0.3$ . + +![](images/9bba45efe3eafb02fe9d6e7a60271a4a90ed6978e77bc5d4e40c3f2da34c4c7a.jpg) +Figure 8. Samples from our sampling process, PSCGAN, and DiffPIR for super-resolution task with downsampling scale $\times 8$ . + +# Acknowledgements + +The work of Yuhan Wang and Ying-Jun Angela Zhang is supported in part by the General Research Fund (project number 14202421, 14214122, 14202723), Area of Excellence Scheme grant (project number AoE/E-601/22-R), and NSFC/RGC Collaborative Research Scheme (project number CRS_HKUST603/22), all from the Research Grants Council of Hong Kong. The work of Suzhi Bi is supported in part by the Natural Science Foundation of China under Project 62271325; in part by Guangdong Basic and Applied Basic Research Foundation under Project 2024B1515020089; in part by Shenzhen Science and Technology Program under Project 20220810142637001. The work of Xiaojun Yuan is supported in part by Sichuan Science and Technology Program under Grant 2024ZYD0036 and in part by National Key Laboratory of Wireless Communications Foundation under Grant IFN20230204. + +# References + +[1] Brian D.O. Anderson. Reverse-time diffusion equation models. Stochastic Processes and their Applications, 12(3):313-326, 1982. +[2] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6228-6237, 2018. +[3] Yochai Blau and Tomer Michaeli. Rethinking lossy compression: The rate-distortion-perception tradeoff. In Proceedings of the 36th International Conference on Machine Learning (ICML), pages 675-685. PMLR, 2019. +[4] Valentin De Bortoli, James Thornton, Jeremy Heng, and Arnaud Doucet. Diffusion schrödinger bridge with applications to score-based generative modeling. In Advances in Neural Information Processing Systems (NIPS), 2021. +[5] Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. In Advances in Neural Information Processing Systems (NIPS), 2022. +[6] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In The Eleventh International Conference on Learning Representations (ICLR), 2023. +[7] Dror Freirich, Tomer Michaeli, and Ron Meir. A theory of the distortion-perception tradeoff in wasserstein space. In Advances in Neural Information Processing Systems (NIPS), 2021. +[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Proceedings of the 34th International Conference on Neural Information Processing Systems (NIPS), 2020. +[9] Tero Karras, Samuli Laine, and Timo Aila. A Style-Based Generator Architecture for Generative Adversarial Networks. IEEE Transactions on Pattern Analysis & Machine Intelligence, 43(12):4217-4228, 2021. + +[10] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. In Advances in Neural Information Processing Systems (NIPS), 2022. +[11] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017. +[12] Dong Liu, Haochen Zhang, and Zhiwei Xiong. On the classification-distortion-perception tradeoff. In Advances in Neural Information Processing Systems (NIPS). Curran Associates, Inc., 2019. +[13] Razvan V Marinescu, Daniel Moyer, and Polina Golland. Bayesian image reconstruction using deep generative models. arXiv preprint, 2021. [Online]. Available: https://arxiv.org/abs/2012.04567. +[14] Anush Krishna Moorthy and Alan Conrad Bovik. Blind image quality assessment: From natural scene statistics to perceptual quality. IEEE transactions on Image Processing, 20 (12):3350-3364, 2011. +[15] Guy Ohayon, Theo Adrai, Gregory Vaksman, Michael Elad, and Peyman Milanfar. High Perceptual Quality Image Denoising with a Posterior Sampling CGAN. In 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 1805-1813, Los Alamitos, CA, USA, 2021. IEEE Computer Society. +[16] Victor M. Panaretos and Yoav Zemel. An Invitation to Statistics in Wasserstein Space. Springer Cham, 2020. +[17] Slavko Simic. On a global upper bound for jensen's inequality. Journal of Mathematical Analysis and Applications, 343 (1):414-419, 2008. +[18] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations (ICLR), 2021. +[19] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In International Conference on Learning Representations (ICLR), 2023. +[20] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In Advances in Neural Information Processing Systems (NIPS). Curran Associates, Inc., 2019. +[21] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations (ICLR), 2021. +[22] Pascal Vincent. A connection between score matching and denoising autoencoders. *Neural Computation*, 23(7):1661-1674, 2011. +[23] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In The European Conference on Computer Vision Workshops (ECCVW), 2018. +[24] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. In + +The Eleventh International Conference on Learning Representations (ICLR), 2023. +[25] Jay Whang, Maurizio Delbracio, Hossein Talebi, Chitwan Sahara, Alexandros G. Dimakis, and Peyman Milanfar. Deblurring via Stochastic Refinement. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16272-16282, Los Alamitos, CA, USA, 2022. +[26] Xingyu Xu and Yuejie Chi. Provably robust score-based diffusion posterior sampling for plug-and-play image reconstruction. In The Thirty-eighth Annual Conference on Neural Information Processing Systems (NIPS), 2024. +[27] Danna Xue, Luis Herranz, Javier Vazquez Corral, and Yanning Zhang. Burst perception-distortion tradeoff: Analysis and evaluation. In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2023. +[28] Zhipeng Xue, Penghao Cai, Xiaojun Yuan, and Xiqi Gao. Score-based variational inference for inverse problems. arXiv preprint, 2024. [Online]. Available: https://arxiv.org/abs/2410.05646. +[29] Yuanzhi Zhu, Kai Zhang, Jingyun Liang, Jiezhang Cao, Bihan Wen, Radu Timofte, and Luc Van Gool. Denoising Diffusion Models for Plug-and-Play Image Restoration. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1219–1229, Los Alamitos, CA, USA, 2023. IEEE Computer Society. \ No newline at end of file diff --git a/2025/Traversing Distortion-Perception Tradeoff using a Single Score-Based Generative Model/images.zip b/2025/Traversing Distortion-Perception Tradeoff using a Single Score-Based Generative Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..459a6ad70c949324de62e64a1912a2b00302166f --- /dev/null +++ b/2025/Traversing Distortion-Perception Tradeoff using a Single Score-Based Generative Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c67e88dafb51466f50d2ef4c239c50892979b117ff9aaa62f451d74daaad994b +size 600062 diff --git a/2025/Traversing Distortion-Perception Tradeoff using a Single Score-Based Generative Model/layout.json b/2025/Traversing Distortion-Perception Tradeoff using a Single Score-Based Generative Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f36dba86643941534fe98c52c63c2ebf1842dc97 --- /dev/null +++ b/2025/Traversing Distortion-Perception Tradeoff using a Single Score-Based Generative Model/layout.json @@ -0,0 +1,11803 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 102, + 520, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 102, + 520, + 138 + ], + "spans": [ + { + "bbox": [ + 91, + 102, + 520, + 138 + ], + "type": "text", + "content": "Traversing Distortion-Perception Tradeoff using a Single Score-Based Generative Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 160, + 471, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 160, + 471, + 175 + ], + "spans": [ + { + "bbox": [ + 138, + 160, + 471, + 175 + ], + "type": "text", + "content": "Yuhan Wang" + }, + { + "bbox": [ + 138, + 160, + 471, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 160, + 471, + 175 + ], + "type": "text", + "content": ", Suzhi Bi" + }, + { + "bbox": [ + 138, + 160, + 471, + 175 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 138, + 160, + 471, + 175 + ], + "type": "text", + "content": ", Ying-Jun Angela Zhang" + }, + { + "bbox": [ + 138, + 160, + 471, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 160, + 471, + 175 + ], + "type": "text", + "content": ", Xiaojun Yuan" + }, + { + "bbox": [ + 138, + 160, + 471, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 156, + 175, + 456, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 175, + 456, + 190 + ], + "spans": [ + { + "bbox": [ + 156, + 175, + 456, + 190 + ], + "type": "text", + "content": "1The Chinese University of Hong Kong 2Shenzhen University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 162, + 190, + 449, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 190, + 449, + 203 + ], + "spans": [ + { + "bbox": [ + 162, + 190, + 449, + 203 + ], + "type": "text", + "content": "3University of Electronic Science and Technology of China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 120, + 205, + 488, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 205, + 488, + 217 + ], + "spans": [ + { + "bbox": [ + 120, + 205, + 488, + 217 + ], + "type": "text", + "content": "{wy023, yjzhang}@ie.cuhk.edu.hk, bsz@szu.edu.cn, xjyuan@uestc.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 270, + 297, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 297, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 297, + 521 + ], + "type": "text", + "content": "The distortion-perception (DP) tradeoff reveals a fundamental conflict between distortion metrics (e.g., MSE and PSNR) and perceptual quality. Recent research has increasingly concentrated on evaluating denoising algorithms within the DP framework. However, existing algorithms either prioritize perceptual quality by sacrificing acceptable distortion, or focus on minimizing MSE for faithful restoration. When the goal shifts or noisy measurements vary, adapting to different points on the DP plane needs retraining or even re-designing the model. Inspired by recent advances in solving inverse problems using score-based generative models, we explore the potential of flexibly and optimally traversing DP tradeoffs using a single pre-trained score-based model. Specifically, we introduce a variance-scaled reverse diffusion process and theoretically characterize the marginal distribution. We then prove that the proposed sample process is an optimal solution to the DP tradeoff for conditional Gaussian distribution. Experimental results on two-dimensional and image datasets illustrate that a single score network can effectively and flexibly traverse the DP tradeoff for general denoising problems." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 544, + 142, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 142, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 142, + 556 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 563, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 296, + 696 + ], + "type": "text", + "content": "In recent years, we have witnessed rapid progress in image restoration algorithms, especially deep learning-based implementations that have shown remarkable achievements in solving general inverse problems. How to evaluate the performance of emerging algorithms is a crucial yet complicated problem. Traditional full reference distortion metrics, such as Mean Square Error (MSE), focus on the pixel-level accuracy between the original image " + }, + { + "bbox": [ + 55, + 563, + 296, + 696 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 563, + 296, + 696 + ], + "type": "text", + "content": " and its reconstruction " + }, + { + "bbox": [ + 55, + 563, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\hat{X}" + }, + { + "bbox": [ + 55, + 563, + 296, + 696 + ], + "type": "text", + "content": ". On the other hand, perceptual quality, referring to the degree to which an image looks natural rather than algorithmically generated [14], is also an important mea" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 246, + 555, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 555, + 282 + ], + "type": "text", + "content": "sure. It has been demonstrated that perceptual quality could be associated with the distance between the distributions of natural images " + }, + { + "bbox": [ + 313, + 246, + 555, + 282 + ], + "type": "inline_equation", + "content": "p_{X}" + }, + { + "bbox": [ + 313, + 246, + 555, + 282 + ], + "type": "text", + "content": " and generated images " + }, + { + "bbox": [ + 313, + 246, + 555, + 282 + ], + "type": "inline_equation", + "content": "p_{\\hat{X}}" + }, + { + "bbox": [ + 313, + 246, + 555, + 282 + ], + "type": "text", + "content": " [2, 3, 12]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 282, + 555, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 555, + 342 + ], + "type": "text", + "content": "Improvements in distortion measures do not necessarily lead to enhancements in perceptual quality. In fact, [2] has demonstrated that there exists a tradeoff between distortion and perception, termed as distortion-perception (DP) trade-off. Mathematically, it can be modeled as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 377, + 346, + 490, + 367 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 346, + 490, + 367 + ], + "spans": [ + { + "bbox": [ + 377, + 346, + 490, + 367 + ], + "type": "interline_equation", + "content": "D(P) = \\min_{p_{\\hat{X} |Y}}\\mathbb{E}[\\Delta (X,\\hat{X})],", + "image_path": "2e064583238daf3b6157688e9f8430cdfe329a3800781c79714e47a24e94a970.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 399, + 370, + 481, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 370, + 481, + 384 + ], + "spans": [ + { + "bbox": [ + 399, + 370, + 481, + 384 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\text {s . t .} d (p _ {X}, p _ {\\tilde {X}}) \\leq P, \\end{array}", + "image_path": "ad602a275604f69fcf8d53ed18ec2da5788326713bda3144dca031fd702b658e.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 389, + 554, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 554, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 554, + 450 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 389, + 554, + 450 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 313, + 389, + 554, + 450 + ], + "type": "text", + "content": " is the degraded observation, " + }, + { + "bbox": [ + 313, + 389, + 554, + 450 + ], + "type": "inline_equation", + "content": "\\Delta : \\mathcal{X} \\times \\hat{\\mathcal{X}} \\to \\mathbb{R}^+" + }, + { + "bbox": [ + 313, + 389, + 554, + 450 + ], + "type": "text", + "content": " is a full reference distortion function (e.g., such as square-error), and " + }, + { + "bbox": [ + 313, + 389, + 554, + 450 + ], + "type": "inline_equation", + "content": "d(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 389, + 554, + 450 + ], + "type": "text", + "content": " is some divergence between probability distribution, such as Wasserstein distance [16] or Kulback-Leibler (KL) divergence." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 450, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 605 + ], + "type": "text", + "content": "Since the introduction of the DP tradeoff, more and more algorithm designs have focused on performance evaluation on the DP plane for specific denoising tasks, e.g., image deblurring [15, 25] and super-resolution [11, 13, 23]. However, most algorithms tend to seek better perceptual quality at the cost of distortion, or focus on optimizing MSE, a distortion measure, to ensure restoration fidelity. When the restoration objective shifts between distortion and perception, we need to retrain or even re-design the network, which can be computationally intensive. In many practical scenarios, it is significant to flexibly and effectively fulfill diverse task objectives or user demands with a single model at inference time." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "There have been explorations into theories and algorithms to traverse the DP tradeoff in inference time [7, 15]. Specifically, in [15], the authors developed a framework based on conditional generative adversarial networks (CGAN) with an additional penalty on posterior expectation. By averaging different numbers of samples or adjusting the noise level injected into the generator, this CGAN-based framework can access various points along the tradeoff. However, using multiple image samples and averaging" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 145, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 145, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 145, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2377" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 263 + ], + "type": "text", + "content": "them may not be efficient during inference, and the method is not provably optimal. Furthermore, the training process relies on pairs of clean and noisy data. Thus, retraining is required for different noise levels and measurements. In [7], the DP tradeoff is studied in Wasserstein space. Theoretical results show that estimators on the DP curve can be constructed by linearly combining two estimators at the extremes: an optimal MSE minimizer and a perfect perception sampler with minimum distortion. This idea is further applied in burst restoration [27]. In practice, a method that achieves a relatively low MSE, alongside another method that achieves good perceptual quality, can serve as approximations of these two extreme estimators for a specific task. Nonetheless, different models must be deployed to address various measurements, rendering this method less flexible to general inverse problems." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 267, + 296, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 267, + 296, + 457 + ], + "spans": [ + { + "bbox": [ + 56, + 267, + 296, + 457 + ], + "type": "text", + "content": "Recent research has demonstrated the generative capability of score-based diffusion models [4, 8, 10, 18, 20, 21] in tackling general inverse problems using a single score network [5, 6, 19, 24, 26, 29]. Score-based diffusion models learn the prior distribution of the data " + }, + { + "bbox": [ + 56, + 267, + 296, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 56, + 267, + 296, + 457 + ], + "type": "text", + "content": " by training a score network to match the gradient of the logarithm density " + }, + { + "bbox": [ + 56, + 267, + 296, + 457 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}}\\log p(\\mathbf{x})" + }, + { + "bbox": [ + 56, + 267, + 296, + 457 + ], + "type": "text", + "content": ", referred to as the score. After observing a noisy measurement " + }, + { + "bbox": [ + 56, + 267, + 296, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 56, + 267, + 296, + 457 + ], + "type": "text", + "content": ", sampling from posteriors involves approximating the conditional score " + }, + { + "bbox": [ + 56, + 267, + 296, + 457 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}}\\log p(\\mathbf{x}|\\mathbf{y})" + }, + { + "bbox": [ + 56, + 267, + 296, + 457 + ], + "type": "text", + "content": ". Theoretical insights provided in [28] also reveal the potential for recovering Minimum Mean Square Error (MMSE) estimation by propagating the mean of the reverse diffusion process. Inspired by these advancements, we explore the potential of traversing DP tradeoffs for different tasks using a single score network. Our main contributions are summarized below:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 462, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 462, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 582 + ], + "type": "text", + "content": "- First, we propose a variance-scaled reverse diffusion process and theoretically characterize the marginal distributions produced by this novel reverse process. It is demonstrated that the mean of our proposed sampling converges towards the MMSE point while the marginal covariance is scaled by a scaling factor. By tuning a parameter that dictates the variance of reverse sampling, we can flexibly navigate from the MMSE point to the posterior distribution, where the two extremes are achieved by setting the parameter to zero and one, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 582, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 640 + ], + "type": "text", + "content": "- Subsequently, we show that the reconstruction obtained from the variance-scaled reverse sampling represents the optimal solution to the conditional DP tradeoff for multivariate Gaussian distributions, assessed through MSE and Wasserstein-2 distance metrics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 712 + ], + "type": "text", + "content": "- Finally, we validate our methodology through experiments conducted on various two-dimensional datasets and a real-world image dataset. With a single pre-trained score network, we can traverse the DP tradeoff across different inverse problems with varying noise levels. The results indicate that the variance-scaled reverse diffusion" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 321, + 72, + 555, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 72, + 555, + 120 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 555, + 120 + ], + "type": "text", + "content": "process achieves a more complete empirical DP tradeoff curve and better MSE than the GAN-based method and other inverse problem solvers, demonstrating the effectiveness and flexibility of the proposed framework." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "text", + "content": "Notations: For a random variable " + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "text", + "content": " denoted by a capital letter, we use small letter " + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "text", + "content": " to denote its realizations, and use " + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "inline_equation", + "content": "p_{X}(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "text", + "content": " to denote the distribution over its alphabet " + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "text", + "content": ". When there is no ambiguity, the distribution can be abbreviated as " + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "text", + "content": ". For a discrete sequence of random variables " + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "inline_equation", + "content": "X_0, X_1, \\dots, X_T" + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "text", + "content": ", we abbreviate them as " + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "inline_equation", + "content": "\\{X_k\\}_{k=0}^T" + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "text", + "content": " and use " + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{0:T}" + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "text", + "content": " to denote their realizations. The expectation and conditional expectation of " + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "inline_equation", + "content": "Y = \\mathbf{y}" + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{p(\\mathbf{x})}[X]" + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{p(\\mathbf{x}|\\mathbf{y})}[X]" + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "text", + "content": ". We use " + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "inline_equation", + "content": "\\mathrm{Cov}_{p(\\mathbf{x})}[X]" + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "inline_equation", + "content": "\\mathrm{Cov}_{p(\\mathbf{x}|\\mathbf{y})}[X]" + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "text", + "content": " to represent the covariance and conditional covariance of " + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "inline_equation", + "content": "Y = \\mathbf{y}" + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "text", + "content": ". Matrices are denoted by uppercase boldface letters (e.g., " + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}" + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "text", + "content": "). " + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "inline_equation", + "content": "\\mathrm{Tr}(\\boldsymbol{\\Sigma})" + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}^{-1}" + }, + { + "bbox": [ + 313, + 121, + 555, + 264 + ], + "type": "text", + "content": " represent the trace and inverse." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 273, + 404, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 404, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 404, + 285 + ], + "type": "text", + "content": "2 Backgrounds" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 293, + 481, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 481, + 304 + ], + "type": "text", + "content": "2.1 Score-based generative model" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 311, + 554, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 554, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 554, + 360 + ], + "type": "text", + "content": "Score-based generative models or diffusion models define the generative process as the reverse of a diffusion process. Specifically, the diffusion process (i.e., forward process) is a Markov chain " + }, + { + "bbox": [ + 313, + 311, + 554, + 360 + ], + "type": "inline_equation", + "content": "\\{X_{k}\\}_{k = 0}^{T}" + }, + { + "bbox": [ + 313, + 311, + 554, + 360 + ], + "type": "text", + "content": ", with joint distribution" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 361, + 364, + 555, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 364, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 361, + 364, + 555, + 397 + ], + "type": "interline_equation", + "content": "p \\left(\\mathbf {x} _ {0: T}\\right) := p \\left(\\mathbf {x} _ {0}\\right) \\prod_ {k = 0} ^ {T - 1} p \\left(\\mathbf {x} _ {k + 1} \\mid \\mathbf {x} _ {k}\\right), \\tag {1}", + "image_path": "f2369f7c1a4fca6ecf74c6ac9ed7c9bebe8860f5afae0d95102b7647ab74915e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "text", + "content": "which gradually diffuses the data " + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "inline_equation", + "content": "X_0 \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "text", + "content": " with distribution " + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0) \\triangleq p_{\\mathrm{data}}" + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "text", + "content": ". Variance preserving (VP) diffusion [8] adopts " + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_{k+1}|\\mathbf{x}_k) \\triangleq \\mathcal{N}(\\mathbf{x}_{k+1}; \\sqrt{1 - \\beta_{k+1}}\\mathbf{x}_k, \\beta_{k+1}\\mathbf{I})" + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "text", + "content": " with monotonically increasing variance schedule " + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "inline_equation", + "content": "\\beta_T \\geq \\beta_{T-1} \\geq \\dots \\geq \\beta_0 = 0" + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "inline_equation", + "content": "T \\to \\infty" + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "text", + "content": " converges to an isotropic Gaussian distribution " + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 313, + 402, + 554, + 474 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 474, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 510 + ], + "type": "text", + "content": "To generate a sample following the data distribution " + }, + { + "bbox": [ + 313, + 474, + 554, + 510 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0)" + }, + { + "bbox": [ + 313, + 474, + 554, + 510 + ], + "type": "text", + "content": ", we can start from the standard Gaussian " + }, + { + "bbox": [ + 313, + 474, + 554, + 510 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_T) \\coloneqq \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 313, + 474, + 554, + 510 + ], + "type": "text", + "content": " and follow the reverse process" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 353, + 515, + 555, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 515, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 353, + 515, + 555, + 548 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\mathbf {x} _ {0: T}\\right) := p _ {\\theta} \\left(\\mathbf {x} _ {T}\\right) \\prod_ {k = 0} ^ {T - 1} p _ {\\theta} \\left(\\mathbf {x} _ {k} \\mid \\mathbf {x} _ {k + 1}\\right), \\tag {2}", + "image_path": "1d9e7988fb8ee399819ce496cf756c6acb972db0bd8c0c84920aced8f8033e2f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 552, + 554, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 554, + 601 + ], + "type": "text", + "content": "with transitions " + }, + { + "bbox": [ + 313, + 552, + 554, + 601 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_k|\\mathbf{x}_{k + 1})\\sim \\mathcal{N}(\\mu_{\\theta}(\\mathbf{x}_k,k),\\boldsymbol {\\Sigma}_{\\theta}(\\mathbf{x}_k,k))" + }, + { + "bbox": [ + 313, + 552, + 554, + 601 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 313, + 552, + 554, + 601 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 552, + 554, + 601 + ], + "type": "text", + "content": " . It will be shown later that the reverse sampling is closely related to the score function for each step, i.e., " + }, + { + "bbox": [ + 313, + 552, + 554, + 601 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}_k}\\log p(\\mathbf{x}_k)" + }, + { + "bbox": [ + 313, + 552, + 554, + 601 + ], + "type": "text", + "content": " , for " + }, + { + "bbox": [ + 313, + 552, + 554, + 601 + ], + "type": "inline_equation", + "content": "k = 0,1,\\dots ,T" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 601, + 554, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 554, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 554, + 636 + ], + "type": "text", + "content": "In continuous-time, we can describe the diffusion process " + }, + { + "bbox": [ + 313, + 601, + 554, + 636 + ], + "type": "inline_equation", + "content": "(X_{t})_{t\\in [0,T_{c}]}" + }, + { + "bbox": [ + 313, + 601, + 554, + 636 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 601, + 554, + 636 + ], + "type": "inline_equation", + "content": "X_{t}\\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 313, + 601, + 554, + 636 + ], + "type": "text", + "content": " with a stochastic differential equation (SDE) [21]" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 329, + 640, + 555, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 640, + 555, + 663 + ], + "spans": [ + { + "bbox": [ + 329, + 640, + 555, + 663 + ], + "type": "interline_equation", + "content": "d X _ {t} = - \\frac {1}{2} \\beta (t) X _ {t} d t + \\sqrt {\\beta (t)} d W _ {t}, X _ {0} \\sim p _ {\\text {d a t a}}, \\tag {3}", + "image_path": "6445a0d967ae58d091324c16ace8b2bdb80808aac1d6dca42fc5f3e3af826435.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 667, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 667, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 667, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 667, + 554, + 714 + ], + "type": "inline_equation", + "content": "(W_{t})_{t\\in [0,T_{c}]}" + }, + { + "bbox": [ + 313, + 667, + 554, + 714 + ], + "type": "text", + "content": " is a standard Brownian motion, and " + }, + { + "bbox": [ + 313, + 667, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\beta (t)" + }, + { + "bbox": [ + 313, + 667, + 554, + 714 + ], + "type": "text", + "content": " is the noise schedule. Let " + }, + { + "bbox": [ + 313, + 667, + 554, + 714 + ], + "type": "inline_equation", + "content": "(p_t)_{t\\in [0:T_c]}" + }, + { + "bbox": [ + 313, + 667, + 554, + 714 + ], + "type": "text", + "content": " be the path distribution associated with (3). To perform data sampling following the distribution " + }, + { + "bbox": [ + 313, + 667, + 554, + 714 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{data}}" + }, + { + "bbox": [ + 313, + 667, + 554, + 714 + ], + "type": "text", + "content": " we can reverse the SDE (3) and" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "2378" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "apply discretization. According to [1] and [21], the reverse SDE corresponding to (3) is" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 99, + 295, + 137 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 99, + 295, + 137 + ], + "spans": [ + { + "bbox": [ + 69, + 99, + 295, + 137 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} d \\overleftarrow {X} _ {t} = \\left[ \\frac {1}{2} \\beta \\left(T _ {c} - t\\right) \\overleftarrow {X} _ {t} + \\beta \\left(T _ {c} - t\\right) \\nabla \\log p _ {T _ {c} - t} (\\overleftarrow {X} _ {t}) \\right] d t \\\\ + \\sqrt {\\beta (T _ {c} - t)} d \\tilde {W} _ {t}, \\overleftarrow {X} _ {0} \\sim p _ {T _ {c}}, \\tag {4} \\\\ \\end{array}", + "image_path": "cb2bc0f637f3c1461ca00cd32a3289b5a7e746435eb94877d0694b006d9fb8fe.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 142, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 142, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 142, + 296, + 228 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 142, + 296, + 228 + ], + "type": "inline_equation", + "content": "(\\tilde{W}_t)_{t\\in [0,T_c]}" + }, + { + "bbox": [ + 55, + 142, + 296, + 228 + ], + "type": "text", + "content": " is another standard Brownian motion. The reverse SDE produces a time-reverse process " + }, + { + "bbox": [ + 55, + 142, + 296, + 228 + ], + "type": "inline_equation", + "content": "(\\overleftarrow{X}_t)_{t\\in [0,T_c]}" + }, + { + "bbox": [ + 55, + 142, + 296, + 228 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 142, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\overleftarrow{X}_t" + }, + { + "bbox": [ + 55, + 142, + 296, + 228 + ], + "type": "text", + "content": " has the same distribution with " + }, + { + "bbox": [ + 55, + 142, + 296, + 228 + ], + "type": "inline_equation", + "content": "X_{T_c - t}" + }, + { + "bbox": [ + 55, + 142, + 296, + 228 + ], + "type": "text", + "content": ". Note that the drift function in (4) depends on the score function " + }, + { + "bbox": [ + 55, + 142, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}_t}\\log p_t(\\mathbf{x}_t)" + }, + { + "bbox": [ + 55, + 142, + 296, + 228 + ], + "type": "text", + "content": ", which can be approximated by a time-aware neural network trained with denoising score matching [22]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 234, + 279, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 234, + 279, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 279, + 247 + ], + "type": "text", + "content": "2.2 Score-based model for posterior sampling" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 252, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 296, + 277 + ], + "type": "text", + "content": "In many application scenarios, we may observe a noisy version of the measurement of the data " + }, + { + "bbox": [ + 55, + 252, + 296, + 277 + ], + "type": "inline_equation", + "content": "X_0" + }, + { + "bbox": [ + 55, + 252, + 296, + 277 + ], + "type": "text", + "content": ", given by" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 282, + 214, + 295 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 282, + 214, + 295 + ], + "spans": [ + { + "bbox": [ + 137, + 282, + 214, + 295 + ], + "type": "interline_equation", + "content": "Y = \\mathcal {A} (X _ {0}) + N,", + "image_path": "e7668e763a219d39d69a95995839d93ae475e2e17244c5506773af6a0ff4c905.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 301, + 296, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 296, + 327 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 301, + 296, + 327 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(\\cdot):\\mathbb{R}^d\\to \\mathbb{R}^n" + }, + { + "bbox": [ + 55, + 301, + 296, + 327 + ], + "type": "text", + "content": " is a measurement operator and " + }, + { + "bbox": [ + 55, + 301, + 296, + 327 + ], + "type": "inline_equation", + "content": "N\\in \\mathbb{R}^n" + }, + { + "bbox": [ + 55, + 301, + 296, + 327 + ], + "type": "text", + "content": " is the measurement noise with " + }, + { + "bbox": [ + 55, + 301, + 296, + 327 + ], + "type": "inline_equation", + "content": "N\\sim \\mathcal{N}(\\mathbf{0},\\sigma_n^2\\mathbf{I})" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 327, + 296, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 374 + ], + "type": "text", + "content": "Score-based generative models have shown powerful capability in solving general inverse problems [6, 10, 19, 26, 28, 29]. Leveraging the diffusion model as the prior, sampling from the reverse posterior distribution" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 89, + 380, + 295, + 413 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 380, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 89, + 380, + 295, + 413 + ], + "type": "interline_equation", + "content": "p \\left(\\mathbf {x} _ {0: T} \\mid \\mathbf {y}\\right) := p \\left(\\mathbf {x} _ {T} \\mid \\mathbf {y}\\right) \\prod_ {k = 0} ^ {T - 1} p \\left(\\mathbf {x} _ {k} \\mid \\mathbf {x} _ {k + 1}, \\mathbf {y}\\right) \\tag {5}", + "image_path": "a7eed03c5c903890891f06b0dcde20ad0a80e222b6b5a4a0e3367c29c3a90e7d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 418, + 296, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 496 + ], + "type": "text", + "content": "requires the knowledge of conditional scores " + }, + { + "bbox": [ + 55, + 418, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}_k}\\log p(\\mathbf{x}_k|\\mathbf{y})" + }, + { + "bbox": [ + 55, + 418, + 296, + 496 + ], + "type": "text", + "content": " which can be expressed as " + }, + { + "bbox": [ + 55, + 418, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}_k}\\log p(\\mathbf{x}_k|\\mathbf{y}) = \\nabla_{\\mathbf{x}_k}\\log p(\\mathbf{x}_k) + \\nabla_{\\mathbf{x}_k}\\log p(\\mathbf{y}|\\mathbf{x}_k)" + }, + { + "bbox": [ + 55, + 418, + 296, + 496 + ], + "type": "text", + "content": ". The first term can be approximated by a score network " + }, + { + "bbox": [ + 55, + 418, + 296, + 496 + ], + "type": "inline_equation", + "content": "s_\\theta (\\mathbf{x}_k,k)" + }, + { + "bbox": [ + 55, + 418, + 296, + 496 + ], + "type": "text", + "content": " [8, 21], and different methods have been proposed to estimate the second term " + }, + { + "bbox": [ + 55, + 418, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}_k}\\log p(\\mathbf{y}|\\mathbf{x}_k)" + }, + { + "bbox": [ + 55, + 418, + 296, + 496 + ], + "type": "text", + "content": " [6, 19]. In each step the conditional sampling process in VP diffusion is [6, 19]" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 500, + 298, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 500, + 298, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 298, + 548 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {x} _ {k - 1} \\leftarrow \\frac {1}{\\sqrt {\\alpha_ {k}}} \\left(\\mathbf {x} _ {k} + (1 - \\alpha_ {k}) \\nabla_ {\\mathbf {x} _ {k}} \\log p (\\mathbf {x} _ {k} | \\mathbf {y})\\right) + \\tilde {\\sigma} _ {k} \\mathbf {z}, \\tag {6} \\\\ \\approx \\frac {1}{\\sqrt {\\alpha_ {k}}} (\\mathbf {x} _ {k} + (1 - \\alpha_ {k}) (s _ {\\theta} (\\mathbf {x} _ {k}, k) + \\nabla_ {\\mathbf {x} _ {k}} \\log p (\\mathbf {y} | \\mathbf {x} _ {k}))) + \\tilde {\\sigma} _ {k} \\mathbf {z}, \\\\ \\end{array}", + "image_path": "a5196412c8b0197dfebe2f5b8fcfe9f3010e6032e875808e2857c13c8d73e1e4.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 552, + 296, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 296, + 576 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 552, + 296, + 576 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 55, + 552, + 296, + 576 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 552, + 296, + 576 + ], + "type": "inline_equation", + "content": "\\tilde{\\sigma}_k^2" + }, + { + "bbox": [ + 55, + 552, + 296, + 576 + ], + "type": "text", + "content": " is the approximated variance of the reverse posterior distribution." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 577, + 296, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 577, + 296, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 577, + 296, + 625 + ], + "type": "text", + "content": "Meanwhile, [28] proposed to directly approximate the reverse conditional " + }, + { + "bbox": [ + 55, + 577, + 296, + 625 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_k|\\mathbf{x}_{k + 1},\\mathbf{y})" + }, + { + "bbox": [ + 55, + 577, + 296, + 625 + ], + "type": "text", + "content": " and propagate the mean in each step. Specifically, when " + }, + { + "bbox": [ + 55, + 577, + 296, + 625 + ], + "type": "inline_equation", + "content": "T\\to \\infty" + }, + { + "bbox": [ + 55, + 577, + 296, + 625 + ], + "type": "text", + "content": " , the mean of " + }, + { + "bbox": [ + 55, + 577, + 296, + 625 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_k|\\mathbf{x}_{k + 1},\\mathbf{y})" + }, + { + "bbox": [ + 55, + 577, + 296, + 625 + ], + "type": "text", + "content": " in VP-diffusion is given by [28]" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 83, + 631, + 295, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 631, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 83, + 631, + 295, + 645 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} _ {k} \\left(\\mathbf {x} _ {k + 1}, \\mathbf {y}\\right) = \\mathbf {U} _ {k} \\mathbf {x} _ {k + 1} + \\mathbf {V} _ {k} \\mathbb {E} _ {p \\left(\\mathbf {x} _ {0} \\mid \\mathbf {y}\\right)} [ X _ {0} ], \\tag {7}", + "image_path": "019c3213c8528743f06689c75c3d12be53089409e62cac798de5c7a229d12c25.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 649, + 84, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 649, + 84, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 649, + 84, + 658 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 58, + 665, + 272, + 695 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 665, + 272, + 695 + ], + "spans": [ + { + "bbox": [ + 58, + 665, + 272, + 695 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {U} _ {k} := \\sqrt {\\alpha_ {k + 1}} \\left(\\left(1 - \\bar {\\alpha} _ {k}\\right) \\mathbf {I} + \\bar {\\alpha} _ {k} \\operatorname {C o v} _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ]\\right) \\\\ \\cdot \\left(\\left(1 - \\bar {\\alpha} _ {k + 1}\\right) \\mathbf {I} + \\bar {\\alpha} _ {k + 1} \\operatorname {C o v} _ {p \\left(\\mathbf {x} _ {0} \\mid \\mathbf {y}\\right)} [ X _ {0} ]\\right) ^ {- 1}, \\\\ \\end{array}", + "image_path": "0cb8072c9020033b55264c08cb04968f4b27795a41b1817c3469e709db952694.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 58, + 697, + 294, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 697, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 697, + 294, + 712 + ], + "type": "interline_equation", + "content": "\\mathbf {V} _ {k} := \\sqrt {\\bar {\\alpha} _ {k}} \\big ((1 - \\alpha_ {k + 1}) \\mathbf {I} \\big) \\big ((1 - \\bar {\\alpha} _ {k + 1}) \\mathbf {I} + \\bar {\\alpha} _ {k + 1} \\operatorname {C o v} _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ] \\big) ^ {- 1}.", + "image_path": "c5a932f02af8f646d253ed3ba48b1bd8fcd3ae9b6fddd0825a8f37e10b5d593b.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "inline_equation", + "content": "\\alpha_{k} = 1 - \\beta_{k}" + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{k} = \\Pi_{i=0}^{k} \\alpha_{i}" + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": " in VP-diffusion. The authors proved that by propagating the mean at each reverse step (i.e., " + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{T} \\rightarrow \\pmb{\\mu}_{T-1}(\\mathbf{x}_{T} = \\pmb{\\mu}_{T}, \\mathbf{y}) \\rightarrow \\dots \\rightarrow \\pmb{\\mu}_{0}(\\mathbf{x}_{1} = \\pmb{\\mu}_{1}, \\mathbf{y})" + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": "), the end point " + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{0}(\\mathbf{x}_{1} = \\pmb{\\mu}_{1}, \\mathbf{y})" + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": " represents the MMSE estimator " + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{p(\\mathbf{x}_0|\\mathbf{y})}[X_0]" + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{T} \\rightarrow 0" + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": ". The proof of reverse mean propagation converging to the MMSE point [28] and its connection to conditional score are included in Appendix A." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 174, + 554, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 174, + 554, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 554, + 198 + ], + "type": "text", + "content": "2.3 Connection between reverse mean and conditional score" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 203, + 554, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 203, + 554, + 249 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 554, + 249 + ], + "type": "text", + "content": "The mean derived in (7) can be viewed as an approximation of the mean in (6). Specifically, when " + }, + { + "bbox": [ + 313, + 203, + 554, + 249 + ], + "type": "inline_equation", + "content": "T \\to \\infty" + }, + { + "bbox": [ + 313, + 203, + 554, + 249 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 203, + 554, + 249 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_k|\\mathbf{y})" + }, + { + "bbox": [ + 313, + 203, + 554, + 249 + ], + "type": "text", + "content": " can be approximately expressed in the form of a Gaussian distribution. We can show that (see details in Appendix A)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 253, + 552, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 253, + 552, + 277 + ], + "spans": [ + { + "bbox": [ + 317, + 253, + 552, + 277 + ], + "type": "interline_equation", + "content": "\\frac {1}{\\sqrt {\\alpha_ {k}}} (\\mathbf {x} _ {k} + (1 - \\alpha_ {k}) \\nabla_ {\\mathbf {x} _ {k}} \\log p (\\mathbf {x} _ {k} | \\mathbf {y})) = \\mathbf {U} _ {k} \\mathbf {x} _ {k + 1} + \\mathbf {V} _ {k} \\mathbb {E} _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ].", + "image_path": "b172d23623742dc2f37f6ea98446352a6634db4c81f7f9590ea18d98339b8e64.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "text", + "content": "Theoretically, when " + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "inline_equation", + "content": "\\tilde{\\sigma}_k = 0" + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "text", + "content": " in each step (6), the endpoint of the reverse process converges to the MMSE estimator, which achieves the best possible performance on distortion. When " + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "inline_equation", + "content": "\\tilde{\\sigma}_k" + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "text", + "content": " is the true posterior variance, the reverse process samples from the posterior distribution " + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_{0:T}|\\mathbf{y})" + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "text", + "content": ", which results in a reconstruction with perfect perception measured in conditional distribution. The above observations motivate us to bridge the two extreme points with the score-based generative model by tuning the scale of " + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "inline_equation", + "content": "\\sigma_k" + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "text", + "content": " in the reverse process. It remains to be answered whether the bridging can provide us with an optimal and flexible estimator to traverse along the distortion-perception plane." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 440, + 554, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 440, + 554, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 554, + 467 + ], + "type": "text", + "content": "3 Traversing DP Tradeoff with Scaled Reverse Diffusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 474, + 554, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 558 + ], + "type": "text", + "content": "In this section, we first propose a novel reverse sampling method based on a variance-scaled version of the joint inference distribution (5) and theoretically derive the mean and variance of the marginal distribution. Then, we prove that the resulting end distribution of the variance-scaled reverse sampling provides an optimal DP tradeoff with respect to MSE and Wasserstein-2 distance." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 564, + 471, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 471, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 471, + 575 + ], + "type": "text", + "content": "3.1 Reverse mean and variance" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 582, + 554, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 606 + ], + "type": "text", + "content": "Consider the joint inference distribution with true mean but scaled variance, i.e.," + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 341, + 612, + 555, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 612, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 341, + 612, + 555, + 632 + ], + "type": "interline_equation", + "content": "p _ {\\lambda} \\left(\\mathbf {x} _ {k} \\mid \\mathbf {x} _ {k + 1}, \\mathbf {y}\\right) := \\mathcal {N} \\left(\\boldsymbol {\\mu} _ {k} \\left(\\mathbf {x} _ {k + 1}, \\mathbf {y}\\right), \\lambda \\mathbf {C} _ {k}\\right), \\tag {8}", + "image_path": "82eec3b1af9cd739c72c0548144e728ac12e0886bdb79e5622961a9f37aa2982.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 637, + 554, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 637, + 554, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 637, + 554, + 661 + ], + "type": "text", + "content": "where the expectation " + }, + { + "bbox": [ + 313, + 637, + 554, + 661 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_k(\\mathbf{x}_{k + 1},\\mathbf{y})" + }, + { + "bbox": [ + 313, + 637, + 554, + 661 + ], + "type": "text", + "content": " is given in (7), and the true covariance " + }, + { + "bbox": [ + 313, + 637, + 554, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_k" + }, + { + "bbox": [ + 313, + 637, + 554, + 661 + ], + "type": "text", + "content": " is given by [28]" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 336, + 665, + 554, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 665, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 336, + 665, + 554, + 712 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {C} _ {k} := \\frac {\\beta_ {k + 1}}{1 - \\beta_ {k + 1}} \\left((1 - \\bar {\\alpha} _ {k}) \\mathbf {I} + \\bar {\\alpha} _ {k} \\operatorname {C o v} _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ]\\right) \\\\ \\cdot \\left(\\left(\\frac {\\beta_ {k + 1}}{1 - \\beta_ {k + 1}} + 1 - \\bar {\\alpha} _ {k}\\right) \\mathbf {I} + \\bar {\\alpha} _ {k} \\operatorname {C o v} _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ]\\right) ^ {- 1}. \\tag {9} \\\\ \\end{array}", + "image_path": "e7307b3e729331e5b8eb2882cf0e4bb127cc33992308a0f3b07f5d63c9402d0f.jpg" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "2379" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 107 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 107 + ], + "type": "text", + "content": "We have the following theorem to characterize the marginal distribution given by the variance-scaled version of the joint inference distribution." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 114, + 296, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 114, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 54, + 114, + 296, + 139 + ], + "type": "text", + "content": "Theorem 1. For VP-diffusion and " + }, + { + "bbox": [ + 54, + 114, + 296, + 139 + ], + "type": "inline_equation", + "content": "0 \\leq \\lambda \\leq 1" + }, + { + "bbox": [ + 54, + 114, + 296, + 139 + ], + "type": "text", + "content": ", consider the joint inference distribution given by" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 75, + 145, + 296, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 145, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 75, + 145, + 296, + 178 + ], + "type": "interline_equation", + "content": "p _ {\\lambda} \\left(\\mathbf {x} _ {0: T} \\mid \\mathbf {y}\\right) = p _ {\\lambda} \\left(\\mathbf {x} _ {T} \\mid y\\right) \\prod_ {k = 0} ^ {T - 1} p _ {\\lambda} \\left(\\mathbf {x} _ {k} \\mid \\mathbf {x} _ {k + 1}, \\mathbf {y}\\right) \\tag {10}", + "image_path": "2219b6150203117ebe06f0f140cfb40a81f9cc346023d3bff4738be8903e0c87.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 185, + 296, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 185, + 296, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 185, + 296, + 220 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 185, + 296, + 220 + ], + "type": "inline_equation", + "content": "p_{\\lambda}(\\mathbf{x}_T|\\mathbf{y}) = \\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 55, + 185, + 296, + 220 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 185, + 296, + 220 + ], + "type": "inline_equation", + "content": "p_{\\lambda}(\\mathbf{x}_k|\\mathbf{x}_{k + 1},\\mathbf{y})" + }, + { + "bbox": [ + 55, + 185, + 296, + 220 + ], + "type": "text", + "content": " is given by (8). Then, the corresponding margin has the distribution " + }, + { + "bbox": [ + 55, + 185, + 296, + 220 + ], + "type": "inline_equation", + "content": "p_{\\lambda}(\\mathbf{x}_k|\\mathbf{y}) = \\mathcal{N}(\\boldsymbol {\\mu}_k^\\lambda ,\\boldsymbol{\\Sigma}_k^\\lambda)" + }, + { + "bbox": [ + 55, + 185, + 296, + 220 + ], + "type": "text", + "content": " where" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 226, + 298, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 226, + 298, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 226, + 298, + 281 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {\\mu} _ {k} ^ {\\lambda} = \\left(\\mathbf {U} _ {k} \\left(\\mathbf {U} _ {k + 1} (\\dots (\\mathbf {U} _ {T - 2} \\mathbf {V} _ {T - 1} + \\mathbf {V} _ {T - 2}) \\dots) + \\mathbf {V} _ {k + 1}\\right) + \\mathbf {V} _ {k}\\right) \\\\ = \\sqrt {\\bar {\\alpha} _ {k}} (1 - \\alpha_ {k + 1} \\cdot \\cdot \\cdot \\alpha_ {T}) \\left((1 - \\bar {\\alpha} _ {T}) \\mathbf {I} + \\bar {\\alpha} _ {T} C o v _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ]\\right) ^ {- 1} \\mathbb {E} _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ], \\\\ \\boldsymbol {\\Sigma} _ {k} ^ {\\lambda} = \\boldsymbol {\\Sigma} _ {k} \\Big (\\lambda \\mathbf {I} + (1 - \\lambda) \\alpha_ {k + 1} \\alpha_ {k + 2} \\dots \\alpha_ {T} \\boldsymbol {\\Sigma} _ {T - 1} ^ {- 1} \\boldsymbol {\\Sigma} _ {k} \\Big), \\\\ \\end{array}", + "image_path": "53a09e236a61a57bd133e14d140309c720a4ecbe3aebc490346ee522f8916990.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 286, + 296, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 296, + 311 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 296, + 311 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 286, + 296, + 311 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_k" + }, + { + "bbox": [ + 55, + 286, + 296, + 311 + ], + "type": "text", + "content": " is the covariance of " + }, + { + "bbox": [ + 55, + 286, + 296, + 311 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_k|\\mathbf{y})" + }, + { + "bbox": [ + 55, + 286, + 296, + 311 + ], + "type": "text", + "content": " in VP diffusion. In particular, when " + }, + { + "bbox": [ + 55, + 286, + 296, + 311 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_T\\rightarrow 0" + }, + { + "bbox": [ + 55, + 286, + 296, + 311 + ], + "type": "text", + "content": ", the variance of " + }, + { + "bbox": [ + 55, + 286, + 296, + 311 + ], + "type": "inline_equation", + "content": "p_{\\lambda}(\\mathbf{x}_0|\\mathbf{y})" + }, + { + "bbox": [ + 55, + 286, + 296, + 311 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 317, + 298, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 298, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 298, + 338 + ], + "type": "interline_equation", + "content": "\\pmb {\\Sigma} _ {0} ^ {\\lambda} = \\pmb {\\Sigma} _ {0} \\bigg (\\lambda \\mathbf {I} + (1 - \\lambda) \\bar {\\alpha} _ {T} \\pmb {\\Sigma} _ {T - 1} ^ {- 1} \\pmb {\\Sigma} _ {0} \\bigg) \\rightarrow \\lambda C o v _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ],", + "image_path": "878a85b5a97ed967bbffef96de15f7364fd77596339e006a3caf69b3460d9020.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 343, + 123, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 123, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 123, + 354 + ], + "type": "text", + "content": "and the mean is" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 360, + 299, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 299, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 299, + 392 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mu_ {0} ^ {\\lambda} = \\sqrt {\\bar {\\alpha} _ {0}} \\big ((1 - \\bar {\\alpha} _ {T}) \\mathbf {I} \\big) \\big ((1 - \\bar {\\alpha} _ {T}) \\mathbf {I} + \\bar {\\alpha} _ {T} C o v _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ] \\big) ^ {- 1} \\\\ \\cdot \\mathbb {E} _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ] \\rightarrow \\mathbb {E} _ {p (\\mathbf {x} _ {0} | \\mathbf {y})} [ X _ {0} ]. \\\\ \\end{array}", + "image_path": "a5a238b8ff66d5b6d690989c9743504684e01267b90017b67f2748d833cd3d1e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 398, + 209, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 209, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 209, + 411 + ], + "type": "text", + "content": "Proof. See the details in Appendix B." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 418, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 514 + ], + "type": "text", + "content": "The above theorem shows that the proposed variance-scaled reverse diffusion process will lead to a scaled marginal distribution at each step, which finally results in the posterior mean and scaled posterior variance of the original data distribution at the end of the reverse process. This seemingly intuitive result actually provides the optimal solution to the conditional distortion-perception tradeoff for multivariate Gaussian distribution." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 520, + 279, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 279, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 279, + 532 + ], + "type": "text", + "content": "3.2 DP tradeoff for conditional Gaussian case" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 538, + 296, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 296, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 296, + 609 + ], + "type": "text", + "content": "Consider the " + }, + { + "bbox": [ + 55, + 538, + 296, + 609 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 538, + 296, + 609 + ], + "type": "text", + "content": "-dimensional source " + }, + { + "bbox": [ + 55, + 538, + 296, + 609 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 538, + 296, + 609 + ], + "type": "text", + "content": " with distribution " + }, + { + "bbox": [ + 55, + 538, + 296, + 609 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 538, + 296, + 609 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 55, + 538, + 296, + 609 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 55, + 538, + 296, + 609 + ], + "type": "text", + "content": " be the noisy version of the source data. Denote the conditional expectation and variance given an observation " + }, + { + "bbox": [ + 55, + 538, + 296, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 55, + 538, + 296, + 609 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 538, + 296, + 609 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\mu}_{\\mathbf{y}} = \\mathbb{E}_{p(\\mathbf{x}|\\mathbf{y})}[X]" + }, + { + "bbox": [ + 55, + 538, + 296, + 609 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 538, + 296, + 609 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}_{\\mathbf{y}} = \\mathrm{Cov}_{p(\\mathbf{x}|\\mathbf{y})}[X]" + }, + { + "bbox": [ + 55, + 538, + 296, + 609 + ], + "type": "text", + "content": ", respectively. The goal of signal restoration is to find a good reconstruction " + }, + { + "bbox": [ + 55, + 538, + 296, + 609 + ], + "type": "inline_equation", + "content": "\\hat{X}" + }, + { + "bbox": [ + 55, + 538, + 296, + 609 + ], + "type": "text", + "content": " based on the observed " + }, + { + "bbox": [ + 55, + 538, + 296, + 609 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 55, + 538, + 296, + 609 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 610, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 296, + 647 + ], + "type": "text", + "content": "From the optimization perspective, given a specific observation " + }, + { + "bbox": [ + 55, + 610, + 296, + 647 + ], + "type": "inline_equation", + "content": "Y = \\check{\\mathbf{y}}" + }, + { + "bbox": [ + 55, + 610, + 296, + 647 + ], + "type": "text", + "content": ", we can define the conditional distortion-perception function as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 72, + 652, + 277, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 652, + 277, + 693 + ], + "spans": [ + { + "bbox": [ + 72, + 652, + 277, + 693 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} D (P) = \\min _ {p _ {\\hat {X} | Y} (\\hat {\\mathbf {x}} | \\check {\\mathbf {y}})} \\mathbb {E} _ {p _ {X, \\hat {X} | Y} (\\mathbf {x}, \\hat {\\mathbf {x}} | \\check {\\mathbf {y}})} [ \\| X - \\hat {X} \\| _ {2} ^ {2} ] \\\\ \\begin{array}{l} \\text {s . t .} d (p _ {X | Y} (\\mathbf {x} | \\breve {\\mathbf {y}}), p _ {\\hat {X} | Y} (\\hat {\\mathbf {x}} | \\breve {\\mathbf {y}})) \\leq P, \\end{array} \\\\ \\end{array}", + "image_path": "b865d3ecbd2d334f51fb60bd6cb40699722ffd01b27df368209418649c9a7eb7.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 700, + 224, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 700, + 224, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 700, + 224, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 700, + 224, + 713 + ], + "type": "inline_equation", + "content": "X, Y" + }, + { + "bbox": [ + 55, + 700, + 224, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 700, + 224, + 713 + ], + "type": "inline_equation", + "content": "\\hat{X}" + }, + { + "bbox": [ + 55, + 700, + 224, + 713 + ], + "type": "text", + "content": " form a Markov chain." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 72, + 555, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 119 + ], + "type": "text", + "content": "Theorem 2. Consider the source " + }, + { + "bbox": [ + 313, + 72, + 555, + 119 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 72, + 555, + 119 + ], + "type": "text", + "content": " and an observation " + }, + { + "bbox": [ + 313, + 72, + 555, + 119 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{y}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 119 + ], + "type": "text", + "content": " that satisfy " + }, + { + "bbox": [ + 313, + 72, + 555, + 119 + ], + "type": "inline_equation", + "content": "p_{X|Y}(\\mathbf{x}|\\tilde{\\mathbf{y}}) \\sim \\mathcal{N}(\\boldsymbol{\\mu}_{\\tilde{\\mathbf{y}}}, \\boldsymbol{\\Sigma}_{\\tilde{\\mathbf{y}}})" + }, + { + "bbox": [ + 313, + 72, + 555, + 119 + ], + "type": "text", + "content": ". The optimal distortion-perception tradeoff given " + }, + { + "bbox": [ + 313, + 72, + 555, + 119 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{y}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 119 + ], + "type": "text", + "content": " with MSE and Wasserstein-2 distance is" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 125, + 555, + 170 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 125, + 555, + 170 + ], + "spans": [ + { + "bbox": [ + 314, + 125, + 555, + 170 + ], + "type": "interline_equation", + "content": "D (P) = \\left\\{ \\begin{array}{l} \\operatorname {T r} \\left(\\boldsymbol {\\Sigma} _ {\\tilde {\\mathbf {y}}}\\right) + \\left(\\sqrt {\\operatorname {T r} \\left(\\boldsymbol {\\Sigma} _ {\\tilde {\\mathbf {y}}}\\right)} - P\\right) ^ {2}, f o r P \\leq \\sqrt {\\operatorname {T r} \\left(\\boldsymbol {\\Sigma} _ {\\tilde {\\mathbf {y}}}\\right)} \\\\ \\operatorname {T r} \\left(\\boldsymbol {\\Sigma} _ {\\tilde {\\mathbf {y}}}\\right), f o r P > \\sqrt {\\operatorname {T r} \\left(\\boldsymbol {\\Sigma} _ {\\tilde {\\mathbf {y}}}\\right)}. \\end{array} \\right. \\tag {11}", + "image_path": "e1fda689edc8df8f5f6c34815d28165b68be0fffae1cd74aeb01a84d05e5e31b.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 177, + 555, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 177, + 555, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 555, + 213 + ], + "type": "text", + "content": "Meanwhile, the optimal tradeoff can be achieved by sampling from the joint inference distribution (10) in Theorem 1 and traversing different " + }, + { + "bbox": [ + 313, + 177, + 555, + 213 + ], + "type": "inline_equation", + "content": "0 \\leq \\lambda \\leq 1" + }, + { + "bbox": [ + 313, + 177, + 555, + 213 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 219, + 454, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 219, + 454, + 232 + ], + "spans": [ + { + "bbox": [ + 314, + 219, + 454, + 232 + ], + "type": "text", + "content": "Proof. See details in Appendix C." + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 544, + 220, + 555, + 230 + ], + "blocks": [ + { + "bbox": [ + 544, + 220, + 555, + 230 + ], + "lines": [ + { + "bbox": [ + 544, + 220, + 555, + 230 + ], + "spans": [ + { + "bbox": [ + 544, + 220, + 555, + 230 + ], + "type": "image", + "image_path": "9cb66dac147096be4a2598fb33943fa5a6d200cee83c024859798864e2583017.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 239, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 555, + 407 + ], + "type": "text", + "content": "Theorem 2 explicitly characterizes the optimal DP tradeoff for conditional Gaussian distribution. When " + }, + { + "bbox": [ + 313, + 239, + 555, + 407 + ], + "type": "inline_equation", + "content": "P = 0" + }, + { + "bbox": [ + 313, + 239, + 555, + 407 + ], + "type": "text", + "content": " (i.e., perfect perceptual quality is expected), the best achievable MSE is " + }, + { + "bbox": [ + 313, + 239, + 555, + 407 + ], + "type": "inline_equation", + "content": "2\\mathrm{Tr}(\\Sigma_{\\bar{\\mathbf{y}}})" + }, + { + "bbox": [ + 313, + 239, + 555, + 407 + ], + "type": "text", + "content": ". As " + }, + { + "bbox": [ + 313, + 239, + 555, + 407 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 239, + 555, + 407 + ], + "type": "text", + "content": " increases, the optimal MSE gradually converges to the MMSE value " + }, + { + "bbox": [ + 313, + 239, + 555, + 407 + ], + "type": "inline_equation", + "content": "\\mathrm{Tr}(\\Sigma_{\\bar{\\mathbf{y}}})" + }, + { + "bbox": [ + 313, + 239, + 555, + 407 + ], + "type": "text", + "content": ". The proof in Appendix C also shows that the variance-scaled reverse diffusion process can attain the entire optimal DP curve for conditional Gaussian distribution. Although the actual data distribution may not follow the conditional Gaussian distribution (e.g., mixture Gaussian, 2D distribution like S-curve, and image dataset), the forthcoming experiments will show that the proposed variance-scaled reverse sampling can effectively and flexibly traverse the DP tradeoff on general datasets using a single score network." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 413, + 536, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 536, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 536, + 426 + ], + "type": "text", + "content": "3.3 Variance-scaled reverse sampling process" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 432, + 554, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 554, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 554, + 479 + ], + "type": "text", + "content": "In this subsection, we discuss how to perform variance-scaled reverse sampling by estimating the conditional score. With a single pre-trained score network, we can traverse different DP tradeoffs for general inverse problems." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 479, + 554, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 554, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 554, + 549 + ], + "type": "text", + "content": "Recall that the proposed variance-scaled reverse sampling process is given by " + }, + { + "bbox": [ + 313, + 479, + 554, + 549 + ], + "type": "inline_equation", + "content": "p_{\\lambda}(\\mathbf{x}_{0:T}|\\mathbf{y}) = p_{\\lambda}(\\mathbf{x}_T|y)\\prod_{k = 0}^{T - 1}p_{\\lambda}(\\mathbf{x}_k|\\mathbf{x}_{k + 1},\\mathbf{y})" + }, + { + "bbox": [ + 313, + 479, + 554, + 549 + ], + "type": "text", + "content": " where for VP diffusion, we have the posteriors " + }, + { + "bbox": [ + 313, + 479, + 554, + 549 + ], + "type": "inline_equation", + "content": "p_{\\lambda}(\\mathbf{x}_{k - 1}|\\mathbf{x}_k,\\mathbf{y})\\coloneqq \\mathcal{N}\\big(\\boldsymbol{\\mu}_{k - 1}(\\mathbf{x}_k,\\mathbf{y}),\\lambda \\mathbf{C}_{k - 1}\\big)" + }, + { + "bbox": [ + 313, + 479, + 554, + 549 + ], + "type": "text", + "content": ". The sampling process in each step can be approximately expressed as" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 322, + 554, + 554, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 554, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 322, + 554, + 554, + 578 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {k - 1} = \\frac {1}{\\sqrt {\\alpha_ {k}}} \\left(\\mathbf {x} _ {k} + (1 - \\alpha_ {k}) \\nabla_ {\\mathbf {x} _ {k}} \\log p \\left(\\mathbf {x} _ {k} \\mid \\mathbf {y}\\right)\\right) + \\lambda \\mathbf {C} _ {k} \\mathbf {z}, \\tag {12}", + "image_path": "34afddbd3ccfabe24bdd584256f4db0776aeba2a11bc6cde1b79788bdd4f0465.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 583, + 554, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 583, + 554, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 583, + 554, + 685 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 583, + 554, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 313, + 583, + 554, + 685 + ], + "type": "text", + "content": ". To draw a sample " + }, + { + "bbox": [ + 313, + 583, + 554, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{k-1}" + }, + { + "bbox": [ + 313, + 583, + 554, + 685 + ], + "type": "text", + "content": " according to (12), we need to compute the conditional score " + }, + { + "bbox": [ + 313, + 583, + 554, + 685 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}_k} \\log p(\\mathbf{x}_k | \\mathbf{y})" + }, + { + "bbox": [ + 313, + 583, + 554, + 685 + ], + "type": "text", + "content": ". Specifically, the conditional scores can be decomposed as " + }, + { + "bbox": [ + 313, + 583, + 554, + 685 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}_k} p(\\mathbf{x}_k | \\mathbf{y}) = \\nabla_{\\mathbf{x}_k} p(\\mathbf{x}_k) + \\nabla_{\\mathbf{x}_k} p(\\mathbf{y} | \\mathbf{x}_k)" + }, + { + "bbox": [ + 313, + 583, + 554, + 685 + ], + "type": "text", + "content": " by Bayes' rule, where the first term can be approximated by a score network " + }, + { + "bbox": [ + 313, + 583, + 554, + 685 + ], + "type": "inline_equation", + "content": "s_\\theta(\\mathbf{x}_k, k)" + }, + { + "bbox": [ + 313, + 583, + 554, + 685 + ], + "type": "text", + "content": " [8, 21]. To deal with the second term, Denoising Posterior Sampling (DPS) [6] proposes to approximate " + }, + { + "bbox": [ + 313, + 583, + 554, + 685 + ], + "type": "inline_equation", + "content": "p(\\mathbf{y} | \\mathbf{x}_k) = \\mathbb{E}_{p(\\mathbf{x}_0 | \\mathbf{x}_k)}[p(\\mathbf{y} | \\mathbf{x}_0)]" + }, + { + "bbox": [ + 313, + 583, + 554, + 685 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 583, + 554, + 685 + ], + "type": "inline_equation", + "content": "p(\\mathbf{y} | \\hat{\\mathbf{x}}_0(\\mathbf{x}_k))" + }, + { + "bbox": [ + 313, + 583, + 554, + 685 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 689, + 555, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 689, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 321, + 689, + 555, + 712 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} _ {0} \\left(\\mathbf {x} _ {k}\\right) = \\mathbb {E} _ {p \\left(\\mathbf {x} _ {0} \\mid \\mathbf {x} _ {k}\\right)} \\left[ X _ {0} \\right] = \\frac {1}{\\alpha_ {k}} \\left(\\mathbf {x} _ {k} + \\left(1 - \\bar {\\alpha} _ {k}\\right) \\nabla_ {\\mathbf {x} _ {k}} p \\left(\\mathbf {x} _ {k}\\right)\\right) \\tag {13}", + "image_path": "e119072e664c30f46e20e4e9de8927174a84484811502f644b02022d3e0d4774.jpg" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2380" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 185 + ], + "type": "text", + "content": "by Tweedie's formula. The approximation error here is upper-bounded by the Jensen gap [17]. Note that " + }, + { + "bbox": [ + 55, + 72, + 296, + 185 + ], + "type": "inline_equation", + "content": "p(\\mathbf{y}|\\hat{\\mathbf{x}}_0(\\mathbf{x}_k))" + }, + { + "bbox": [ + 55, + 72, + 296, + 185 + ], + "type": "text", + "content": " is analytically tractable when the measurement distribution is given. Specifically, when the measurement nosie is " + }, + { + "bbox": [ + 55, + 72, + 296, + 185 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,\\sigma_n^2\\mathbf{I})" + }, + { + "bbox": [ + 55, + 72, + 296, + 185 + ], + "type": "text", + "content": ", the second term in the conditional score takes the form of " + }, + { + "bbox": [ + 55, + 72, + 296, + 185 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}_k}\\log p(\\mathbf{y}|\\mathbf{x}_k) = \\frac{1}{\\sigma_n^2}\\nabla_{\\mathbf{x}_k}||\\mathbf{y}-\\mathcal{A}(\\hat{\\mathbf{x}}_0(\\mathbf{x}_k))||_2^2" + }, + { + "bbox": [ + 55, + 72, + 296, + 185 + ], + "type": "text", + "content": ". Thus, the conditional scores can be approximated as " + }, + { + "bbox": [ + 55, + 72, + 296, + 185 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}_k}\\log p(\\mathbf{x}_k|\\mathbf{y}) = s_\\theta (\\mathbf{x}_k,k) + \\frac{1}{\\sigma_n^2}\\nabla_{\\mathbf{x}_k}||\\mathbf{y}-\\mathcal{A}(\\hat{\\mathbf{x}}_0(\\mathbf{x}_k))||_2^2" + }, + { + "bbox": [ + 55, + 72, + 296, + 185 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 186, + 296, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 186, + 296, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 296, + 222 + ], + "type": "text", + "content": "In the following experiments, we adopt the DPS framework to approximate the conditional score and sample from (12), which is sketched as Algorithm 1." + } + ] + } + ], + "index": 1 + }, + { + "type": "code", + "bbox": [ + 56, + 249, + 296, + 378 + ], + "blocks": [ + { + "bbox": [ + 56, + 233, + 294, + 246 + ], + "lines": [ + { + "bbox": [ + 56, + 233, + 294, + 246 + ], + "spans": [ + { + "bbox": [ + 56, + 233, + 294, + 246 + ], + "type": "text", + "content": "Algorithm 1 Variance-Scaled Reverse Sampling with DPS" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 56, + 249, + 296, + 378 + ], + "lines": [ + { + "bbox": [ + 56, + 249, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 56, + 249, + 296, + 378 + ], + "type": "text", + "content": "Require: " + }, + { + "bbox": [ + 56, + 249, + 296, + 378 + ], + "type": "inline_equation", + "content": "T,\\mathbf{y},\\lambda ,\\{\\tilde{\\sigma}_k\\}_{k = 0}^T,\\{\\zeta_{k,\\lambda}\\}_{k = 0}^T" + }, + { + "bbox": [ + 56, + 249, + 296, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 56, + 249, + 296, + 378 + ], + "type": "text", + "content": " \nfor " + }, + { + "bbox": [ + 56, + 249, + 296, + 378 + ], + "type": "inline_equation", + "content": "k = T - 1" + }, + { + "bbox": [ + 56, + 249, + 296, + 378 + ], + "type": "text", + "content": " to 1 do do " + }, + { + "bbox": [ + 56, + 249, + 296, + 378 + ], + "type": "inline_equation", + "content": "\\hat{s}\\gets s_{\\theta}(\\mathbf{x}_k,k)" + }, + { + "bbox": [ + 56, + 249, + 296, + 378 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0\\gets \\frac{1}{\\sqrt{\\bar{\\alpha}_k}} (\\mathbf{x}_k + (1 - \\bar{\\alpha}_k)\\hat{s})" + }, + { + "bbox": [ + 56, + 249, + 296, + 378 + ], + "type": "inline_equation", + "content": "\\hat{c} (\\hat{\\mathbf{x}}_0) = -\\frac{\\sqrt{\\alpha_k}}{1 - \\alpha_k}\\nabla_{\\mathbf{x}_k}||y - \\mathcal{A}(\\hat{\\mathbf{x}}_0(\\mathbf{x}_k))||_2^2" + }, + { + "bbox": [ + 56, + 249, + 296, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{z}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 56, + 249, + 296, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{k - 1}\\leftarrow \\frac{1}{\\sqrt{\\alpha_k}}\\bigl (\\mathbf{x}_k + (1 - \\alpha_k)\\bigl (\\hat{s} +\\zeta_{k,\\lambda}\\hat{c} (\\hat{\\mathbf{x}}_0)\\bigr)\\bigr) + \\lambda \\tilde{\\sigma}_k\\mathbf{z}" + }, + { + "bbox": [ + 56, + 249, + 296, + 378 + ], + "type": "text", + "content": " \nend for \nreturn " + }, + { + "bbox": [ + 56, + 249, + 296, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 395, + 296, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 395, + 296, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 296, + 455 + ], + "type": "text", + "content": "In Algorithm 1, " + }, + { + "bbox": [ + 55, + 395, + 296, + 455 + ], + "type": "inline_equation", + "content": "\\tilde{\\sigma}_k" + }, + { + "bbox": [ + 55, + 395, + 296, + 455 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 55, + 395, + 296, + 455 + ], + "type": "inline_equation", + "content": "\\beta_{k}" + }, + { + "bbox": [ + 55, + 395, + 296, + 455 + ], + "type": "text", + "content": " [8, 28], and " + }, + { + "bbox": [ + 55, + 395, + 296, + 455 + ], + "type": "inline_equation", + "content": "\\zeta_{k,\\lambda}" + }, + { + "bbox": [ + 55, + 395, + 296, + 455 + ], + "type": "text", + "content": " is a hyperparameter to control the weight of the conditional score, and may differ for different " + }, + { + "bbox": [ + 55, + 395, + 296, + 455 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 395, + 296, + 455 + ], + "type": "text", + "content": ". The heuristic choices of " + }, + { + "bbox": [ + 55, + 395, + 296, + 455 + ], + "type": "inline_equation", + "content": "\\zeta_{k,\\lambda}" + }, + { + "bbox": [ + 55, + 395, + 296, + 455 + ], + "type": "text", + "content": " and the rationale for fine-tuning have been included in Appendix E." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 462, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 296, + 715 + ], + "type": "text", + "content": "Remark 1. In this paper, we adopt the conditional score approximation proposed in [6] for illustration. To implement the variance-scaled reverse sampling process (12), one can employ any approach to directly estimate the conditional score [19] or estimate the posterior mean as a whole [28] for general inverse problems. For example, Pseudoinverse-guided Diffusion Models (ΠGDM) [19] approximated the unconditional posteriors " + }, + { + "bbox": [ + 55, + 462, + 296, + 715 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0|\\mathbf{x}_k)" + }, + { + "bbox": [ + 55, + 462, + 296, + 715 + ], + "type": "text", + "content": " with Gaussian, i.e., " + }, + { + "bbox": [ + 55, + 462, + 296, + 715 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0|\\mathbf{x}_k) \\approx \\mathcal{N}(\\hat{\\mathbf{x}}_0,r_k^2\\mathbf{I})" + }, + { + "bbox": [ + 55, + 462, + 296, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 462, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0" + }, + { + "bbox": [ + 55, + 462, + 296, + 715 + ], + "type": "text", + "content": " is given by (13), and " + }, + { + "bbox": [ + 55, + 462, + 296, + 715 + ], + "type": "inline_equation", + "content": "r_k^2" + }, + { + "bbox": [ + 55, + 462, + 296, + 715 + ], + "type": "text", + "content": " is the hyperparameter. Thus, the score term " + }, + { + "bbox": [ + 55, + 462, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}_k}p(\\mathbf{y}|\\mathbf{x}_k)" + }, + { + "bbox": [ + 55, + 462, + 296, + 715 + ], + "type": "text", + "content": " can be analytically expressed by the pseudoinverse of the measurement model. In [28], the posterior mean is estimated as a whole. Specifically, we have " + }, + { + "bbox": [ + 55, + 462, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mu_{k - 1}(\\mathbf{x}_k,\\mathbf{y}) = \\mathbf{U}_{k - 1}\\mathbf{x}_k + \\mathbf{V}_{k - 1}\\mathbb{E}_{p(\\mathbf{x}_0|\\mathbf{y})}[X_0]" + }, + { + "bbox": [ + 55, + 462, + 296, + 715 + ], + "type": "text", + "content": ". Since the expression involves " + }, + { + "bbox": [ + 55, + 462, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{p(\\mathbf{x}_0|\\mathbf{y})}[X_0]" + }, + { + "bbox": [ + 55, + 462, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 462, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{Cov}_{p(\\mathbf{x}_0|\\mathbf{y})}[X_0]" + }, + { + "bbox": [ + 55, + 462, + 296, + 715 + ], + "type": "text", + "content": ", it can not be computed directly. Instead, the authors approximate the joint conditional posterior with variational Gaussian distributions and apply natural gradient descent to perform the sampling. Note that the gradient calculation also involves the approximation of the conditional score " + }, + { + "bbox": [ + 55, + 462, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}_k}\\log p(\\mathbf{y}|\\mathbf{x}_k)" + }, + { + "bbox": [ + 55, + 462, + 296, + 715 + ], + "type": "text", + "content": " as in DPS [6]." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 319, + 72, + 392, + 135 + ], + "blocks": [ + { + "bbox": [ + 319, + 72, + 392, + 135 + ], + "lines": [ + { + "bbox": [ + 319, + 72, + 392, + 135 + ], + "spans": [ + { + "bbox": [ + 319, + 72, + 392, + 135 + ], + "type": "image", + "image_path": "e1ee14930088bcc594c2b6a18eae8fb0a7d44a986c5ff556883074c675e8c058.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 344, + 135, + 377, + 146 + ], + "lines": [ + { + "bbox": [ + 344, + 135, + 377, + 146 + ], + "spans": [ + { + "bbox": [ + 344, + 135, + 377, + 146 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 344, + 135, + 377, + 146 + ], + "type": "inline_equation", + "content": "p(x_0)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 157, + 553, + 179 + ], + "lines": [ + { + "bbox": [ + 313, + 157, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 553, + 179 + ], + "type": "text", + "content": "Figure 1. An example of mixture Gaussian, the noisy observation, and the conditional distribution given an observation " + }, + { + "bbox": [ + 313, + 157, + 553, + 179 + ], + "type": "inline_equation", + "content": "y = -0.6" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 394, + 72, + 465, + 135 + ], + "blocks": [ + { + "bbox": [ + 394, + 72, + 465, + 135 + ], + "lines": [ + { + "bbox": [ + 394, + 72, + 465, + 135 + ], + "spans": [ + { + "bbox": [ + 394, + 72, + 465, + 135 + ], + "type": "image", + "image_path": "c100d554646e05dc0617d22319c1555cd6b1c6d7a8be2e6e6423ee9d3eb2a635.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 423, + 136, + 453, + 146 + ], + "lines": [ + { + "bbox": [ + 423, + 136, + 453, + 146 + ], + "spans": [ + { + "bbox": [ + 423, + 136, + 453, + 146 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 423, + 136, + 453, + 146 + ], + "type": "inline_equation", + "content": "p(y)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 467, + 72, + 548, + 134 + ], + "blocks": [ + { + "bbox": [ + 467, + 72, + 548, + 134 + ], + "lines": [ + { + "bbox": [ + 467, + 72, + 548, + 134 + ], + "spans": [ + { + "bbox": [ + 467, + 72, + 548, + 134 + ], + "type": "image", + "image_path": "6cc451c67db869511cf8be3cf9c3918fa38aa66220bb4fe8d9fa51b5a55b464c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 478, + 135, + 545, + 147 + ], + "lines": [ + { + "bbox": [ + 478, + 135, + 545, + 147 + ], + "spans": [ + { + "bbox": [ + 478, + 135, + 545, + 147 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 478, + 135, + 545, + 147 + ], + "type": "inline_equation", + "content": "p(x|y = -0.6)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 185, + 451, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 185, + 451, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 185, + 451, + 198 + ], + "type": "text", + "content": "3.4 An illustrative example" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 203, + 553, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 203, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 553, + 299 + ], + "type": "text", + "content": "Consider the mixture Gaussian distribution " + }, + { + "bbox": [ + 313, + 203, + 553, + 299 + ], + "type": "inline_equation", + "content": "X_0 \\sim p(x_0)" + }, + { + "bbox": [ + 313, + 203, + 553, + 299 + ], + "type": "text", + "content": " with two components, where " + }, + { + "bbox": [ + 313, + 203, + 553, + 299 + ], + "type": "inline_equation", + "content": "p(x_0) = w_1 \\mathcal{N}(\\mu_1, \\sigma_1^2) + w_2 \\mathcal{N}(\\mu_2, \\sigma_2^2)" + }, + { + "bbox": [ + 313, + 203, + 553, + 299 + ], + "type": "text", + "content": ". The noisy observation is obtained by " + }, + { + "bbox": [ + 313, + 203, + 553, + 299 + ], + "type": "inline_equation", + "content": "Y = aX_0 + \\sigma_0 \\epsilon" + }, + { + "bbox": [ + 313, + 203, + 553, + 299 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 203, + 553, + 299 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0, 1)" + }, + { + "bbox": [ + 313, + 203, + 553, + 299 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 203, + 553, + 299 + ], + "type": "inline_equation", + "content": "p(y|x_0) = \\mathcal{N}(ax_0, \\sigma_0^2)" + }, + { + "bbox": [ + 313, + 203, + 553, + 299 + ], + "type": "text", + "content": ". The MMSE estimator and posterior distribution can be theoretically derived (see Appendix D for details). Fig. 1 shows an example of mixture Gaussian distribution and the conditional distribution given an observation " + }, + { + "bbox": [ + 313, + 203, + 553, + 299 + ], + "type": "inline_equation", + "content": "y = -0.6" + }, + { + "bbox": [ + 313, + 203, + 553, + 299 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 299, + 554, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 554, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 554, + 454 + ], + "type": "text", + "content": "Given an observation " + }, + { + "bbox": [ + 313, + 299, + 554, + 454 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 299, + 554, + 454 + ], + "type": "text", + "content": " and a chosen " + }, + { + "bbox": [ + 313, + 299, + 554, + 454 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 299, + 554, + 454 + ], + "type": "text", + "content": ", we iteratively perform the variance scaled reverse process (12) for " + }, + { + "bbox": [ + 313, + 299, + 554, + 454 + ], + "type": "inline_equation", + "content": "k = T, \\dots, 1" + }, + { + "bbox": [ + 313, + 299, + 554, + 454 + ], + "type": "text", + "content": ", where the variance is approximated by " + }, + { + "bbox": [ + 313, + 299, + 554, + 454 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_k = \\beta_k" + }, + { + "bbox": [ + 313, + 299, + 554, + 454 + ], + "type": "text", + "content": " as the conventional VP-diffusion. Starting from a random sample " + }, + { + "bbox": [ + 313, + 299, + 554, + 454 + ], + "type": "inline_equation", + "content": "x_T \\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 313, + 299, + 554, + 454 + ], + "type": "text", + "content": ", Fig. 2 shows the trajectories " + }, + { + "bbox": [ + 313, + 299, + 554, + 454 + ], + "type": "inline_equation", + "content": "x_T \\to \\dots x_k \\to \\dots x_0" + }, + { + "bbox": [ + 313, + 299, + 554, + 454 + ], + "type": "text", + "content": " of multiple reconstructions for each " + }, + { + "bbox": [ + 313, + 299, + 554, + 454 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 299, + 554, + 454 + ], + "type": "text", + "content": ". We can observe that for " + }, + { + "bbox": [ + 313, + 299, + 554, + 454 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 313, + 299, + 554, + 454 + ], + "type": "text", + "content": ", given an initial " + }, + { + "bbox": [ + 313, + 299, + 554, + 454 + ], + "type": "inline_equation", + "content": "x_T" + }, + { + "bbox": [ + 313, + 299, + 554, + 454 + ], + "type": "text", + "content": ", the trajectories are deterministic and converge to the MMSE point. This phenomenon collaborates with results in [28]. When " + }, + { + "bbox": [ + 313, + 299, + 554, + 454 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 299, + 554, + 454 + ], + "type": "text", + "content": " increases, the generated trajectories follow the form of the posterior distribution and show more stochasticity. When " + }, + { + "bbox": [ + 313, + 299, + 554, + 454 + ], + "type": "inline_equation", + "content": "\\lambda = 1" + }, + { + "bbox": [ + 313, + 299, + 554, + 454 + ], + "type": "text", + "content": ", the reconstruction distribution matches the ground truth posterior." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 454, + 554, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 554, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 554, + 562 + ], + "type": "text", + "content": "The resulting DP tradeoffs are shown in Fig. 3. It can be observed that as " + }, + { + "bbox": [ + 313, + 454, + 554, + 562 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 454, + 554, + 562 + ], + "type": "text", + "content": " increases, the divergence between the true posterior distribution and reconstruction distribution continuously decreases, with an increase of MSE value. The overall curve is convex, and both Wasserstein-2 and KL-divergence come to zero. Meanwhile, the MSE achieved at " + }, + { + "bbox": [ + 313, + 454, + 554, + 562 + ], + "type": "inline_equation", + "content": "\\lambda = 1" + }, + { + "bbox": [ + 313, + 454, + 554, + 562 + ], + "type": "text", + "content": " is approximately two times the MSE at " + }, + { + "bbox": [ + 313, + 454, + 554, + 562 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 313, + 454, + 554, + 562 + ], + "type": "text", + "content": " for both Wasserstein-2 and KL-divergence cases, coinciding with the result in Theorem 2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 572, + 446, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 572, + 446, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 572, + 446, + 586 + ], + "type": "text", + "content": "4 Experimental Results" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 592, + 554, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 554, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 554, + 664 + ], + "type": "text", + "content": "In this section, we conduct a series of experiments on two-dimensional datasets and the FFHQ dataset [9]. It is demonstrated that the variance-scaled reverse sampling shows great flexibility and effectiveness in traversing DP tradeoff compared to the GAN-based approach and other denoising methods." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 671, + 495, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 495, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 495, + 684 + ], + "type": "text", + "content": "4.1 Two-dimensional points example" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "First, we evaluate the validity and effectiveness of our approach on two-dimensional datasets, including pinwheel, S-" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2381" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 57, + 553, + 148 + ], + "blocks": [ + { + "bbox": [ + 59, + 57, + 553, + 148 + ], + "lines": [ + { + "bbox": [ + 59, + 57, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 59, + 57, + 553, + 148 + ], + "type": "image", + "image_path": "44fd470180d9b8a7fa4c35dafe0a9fd4baf1afd510d8e21b8a790f4d5ea4cb47.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 156, + 538, + 168 + ], + "lines": [ + { + "bbox": [ + 70, + 156, + 538, + 168 + ], + "spans": [ + { + "bbox": [ + 70, + 156, + 538, + 168 + ], + "type": "text", + "content": "Figure 2. Trajectories " + }, + { + "bbox": [ + 70, + 156, + 538, + 168 + ], + "type": "inline_equation", + "content": "x_{T} \\to \\dots x_{k} \\to \\dots \\to x_{0}" + }, + { + "bbox": [ + 70, + 156, + 538, + 168 + ], + "type": "text", + "content": " of different reconstructions for " + }, + { + "bbox": [ + 70, + 156, + 538, + 168 + ], + "type": "inline_equation", + "content": "\\lambda = 0, 0.3, 0.8" + }, + { + "bbox": [ + 70, + 156, + 538, + 168 + ], + "type": "text", + "content": " and 1. The initial " + }, + { + "bbox": [ + 70, + 156, + 538, + 168 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 70, + 156, + 538, + 168 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 70, + 156, + 538, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,1)" + }, + { + "bbox": [ + 70, + 156, + 538, + 168 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 182, + 175, + 274 + ], + "blocks": [ + { + "bbox": [ + 58, + 182, + 175, + 274 + ], + "lines": [ + { + "bbox": [ + 58, + 182, + 175, + 274 + ], + "spans": [ + { + "bbox": [ + 58, + 182, + 175, + 274 + ], + "type": "image", + "image_path": "67e673cf40bd32b441cd2a9307d4404578d99a9f7a39d23d9448828cfbf0940e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 277, + 120, + 284 + ], + "lines": [ + { + "bbox": [ + 113, + 277, + 120, + 284 + ], + "spans": [ + { + "bbox": [ + 113, + 277, + 120, + 284 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 182, + 182, + 296, + 274 + ], + "blocks": [ + { + "bbox": [ + 182, + 182, + 296, + 274 + ], + "lines": [ + { + "bbox": [ + 182, + 182, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 182, + 182, + 296, + 274 + ], + "type": "image", + "image_path": "6643881d2c6b59afaab83c3244091818b1463bbc20a10993a504b8116e0820f6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 277, + 249, + 283 + ], + "lines": [ + { + "bbox": [ + 241, + 277, + 249, + 283 + ], + "spans": [ + { + "bbox": [ + 241, + 277, + 249, + 283 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 295, + 296, + 339 + ], + "lines": [ + { + "bbox": [ + 55, + 295, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 296, + 339 + ], + "type": "text", + "content": "Figure 3. Distortion-perception tradeoff traversed by variance-scaled reverse sampling given different " + }, + { + "bbox": [ + 55, + 295, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 295, + 296, + 339 + ], + "type": "text", + "content": "'s. (a) tradeoff between Wasserstein-2 distance and MSE; (b) tradeoff between KL-divergence and MSE." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 349, + 295, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 295, + 397 + ], + "type": "text", + "content": "curve, and moon-shape data distributions. Taking the pinwheel dataset as an example, the first row of Fig. 4 shows the distribution of pinwheel data points " + }, + { + "bbox": [ + 55, + 349, + 295, + 397 + ], + "type": "inline_equation", + "content": "X_0" + }, + { + "bbox": [ + 55, + 349, + 295, + 397 + ], + "type": "text", + "content": " and the noisy observation " + }, + { + "bbox": [ + 55, + 349, + 295, + 397 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 55, + 349, + 295, + 397 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 349, + 295, + 397 + ], + "type": "inline_equation", + "content": "Y = aX + N" + }, + { + "bbox": [ + 55, + 349, + 295, + 397 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 349, + 295, + 397 + ], + "type": "inline_equation", + "content": "N \\sim \\mathcal{N}(0, \\sigma_n^2 \\mathbf{I})" + }, + { + "bbox": [ + 55, + 349, + 295, + 397 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 399, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 628 + ], + "type": "text", + "content": "For each dataset, we train a conventional score network to approximate " + }, + { + "bbox": [ + 55, + 399, + 296, + 628 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0)" + }, + { + "bbox": [ + 55, + 399, + 296, + 628 + ], + "type": "text", + "content": " on the original distribution. Then, we perform the variance-scaled reverse diffusion process (12) on the noisy observation. As a comparison, we adopt a GAN-based DP-traversing approach PSCGAN [15]. This scheme introduces a penalty of the posterior expectation to the training of the conditional GAN. The authors proposed two strategies to navigate the DP tradeoff during the inference time: (1) PSCGAN- " + }, + { + "bbox": [ + 55, + 399, + 296, + 628 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 399, + 296, + 628 + ], + "type": "text", + "content": ": This method involves sampling " + }, + { + "bbox": [ + 55, + 399, + 296, + 628 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 399, + 296, + 628 + ], + "type": "text", + "content": " instances from the generator and then averaging them. As " + }, + { + "bbox": [ + 55, + 399, + 296, + 628 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 399, + 296, + 628 + ], + "type": "text", + "content": " increases, the averaged image is closer to the conditional expectation, leading to a smaller distortion but larger perception loss. (2) PSCGAN- " + }, + { + "bbox": [ + 55, + 399, + 296, + 628 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 399, + 296, + 628 + ], + "type": "text", + "content": ": It varies the standard deviation " + }, + { + "bbox": [ + 55, + 399, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\sigma_z" + }, + { + "bbox": [ + 55, + 399, + 296, + 628 + ], + "type": "text", + "content": " of the noise injected into the generator to control the stochasticity. For " + }, + { + "bbox": [ + 55, + 399, + 296, + 628 + ], + "type": "inline_equation", + "content": "0 \\leq \\sigma_z \\leq 1" + }, + { + "bbox": [ + 55, + 399, + 296, + 628 + ], + "type": "text", + "content": ", a larger " + }, + { + "bbox": [ + 55, + 399, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\sigma_z" + }, + { + "bbox": [ + 55, + 399, + 296, + 628 + ], + "type": "text", + "content": " results in a better perceptual quality and higher MSE. Note that the training of PSCGAN relies on noisy data points " + }, + { + "bbox": [ + 55, + 399, + 296, + 628 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 55, + 399, + 296, + 628 + ], + "type": "text", + "content": ". Thus, the model needs to be retrained for different measurement scenarios and different levels of " + }, + { + "bbox": [ + 55, + 399, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\sigma_n^2" + }, + { + "bbox": [ + 55, + 399, + 296, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 628, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 295, + 700 + ], + "type": "text", + "content": "Fig. 4 shows the reconstruction of our variance-scaled reverse diffusion process for different " + }, + { + "bbox": [ + 55, + 628, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 628, + 295, + 700 + ], + "type": "text", + "content": "'s, as well as the reconstruction of PSCGAN when setting " + }, + { + "bbox": [ + 55, + 628, + 295, + 700 + ], + "type": "inline_equation", + "content": "N = 16" + }, + { + "bbox": [ + 55, + 628, + 295, + 700 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 628, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\sigma_z = 1" + }, + { + "bbox": [ + 55, + 628, + 295, + 700 + ], + "type": "text", + "content": ". It can be observed that when " + }, + { + "bbox": [ + 55, + 628, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 55, + 628, + 295, + 700 + ], + "type": "text", + "content": ", our sampling process leads to a more concentrated reconstruction, and gradually approaches the true distribution as " + }, + { + "bbox": [ + 55, + 628, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 628, + 295, + 700 + ], + "type": "text", + "content": " increases." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "Fig. 5 compares the numerical DP tradeoffs achieved" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 366, + 182, + 433, + 232 + ], + "blocks": [ + { + "bbox": [ + 366, + 182, + 433, + 232 + ], + "lines": [ + { + "bbox": [ + 366, + 182, + 433, + 232 + ], + "spans": [ + { + "bbox": [ + 366, + 182, + 433, + 232 + ], + "type": "image", + "image_path": "7367742c11f73a719eb2cf15c2f2a74b107fabb9b8400d1bf8d20b7f86288a4d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 233, + 424, + 239 + ], + "lines": [ + { + "bbox": [ + 378, + 233, + 424, + 239 + ], + "spans": [ + { + "bbox": [ + 378, + 233, + 424, + 239 + ], + "type": "text", + "content": "Original distributions" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 440, + 182, + 504, + 232 + ], + "blocks": [ + { + "bbox": [ + 440, + 182, + 504, + 232 + ], + "lines": [ + { + "bbox": [ + 440, + 182, + 504, + 232 + ], + "spans": [ + { + "bbox": [ + 440, + 182, + 504, + 232 + ], + "type": "image", + "image_path": "9fcc744ec54d2fb0e6ea48623cd5ecd146a42bc0f1b99226f7ef6647329c8239.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 233, + 493, + 239 + ], + "lines": [ + { + "bbox": [ + 453, + 233, + 493, + 239 + ], + "spans": [ + { + "bbox": [ + 453, + 233, + 493, + 239 + ], + "type": "text", + "content": "Noisy observations" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 329, + 244, + 399, + 296 + ], + "blocks": [ + { + "bbox": [ + 329, + 244, + 399, + 296 + ], + "lines": [ + { + "bbox": [ + 329, + 244, + 399, + 296 + ], + "spans": [ + { + "bbox": [ + 329, + 244, + 399, + 296 + ], + "type": "image", + "image_path": "0c0f5d0e3cc24a89e68c22f69f79b27cd4f97326eddeb6687fb3defc20da2582.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 298, + 374, + 304 + ], + "lines": [ + { + "bbox": [ + 353, + 298, + 374, + 304 + ], + "spans": [ + { + "bbox": [ + 353, + 298, + 374, + 304 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 353, + 298, + 374, + 304 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 400, + 245, + 469, + 297 + ], + "blocks": [ + { + "bbox": [ + 400, + 245, + 469, + 297 + ], + "lines": [ + { + "bbox": [ + 400, + 245, + 469, + 297 + ], + "spans": [ + { + "bbox": [ + 400, + 245, + 469, + 297 + ], + "type": "image", + "image_path": "7c594d082776a16799d3443f4016f40ca0bbce69bdc87eede7ac0d11729a4b53.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 298, + 449, + 304 + ], + "lines": [ + { + "bbox": [ + 425, + 298, + 449, + 304 + ], + "spans": [ + { + "bbox": [ + 425, + 298, + 449, + 304 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 425, + 298, + 449, + 304 + ], + "type": "inline_equation", + "content": "\\lambda = 0.3" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 471, + 245, + 541, + 296 + ], + "blocks": [ + { + "bbox": [ + 471, + 245, + 541, + 296 + ], + "lines": [ + { + "bbox": [ + 471, + 245, + 541, + 296 + ], + "spans": [ + { + "bbox": [ + 471, + 245, + 541, + 296 + ], + "type": "image", + "image_path": "4ebdcc77cb63da8a32d060e277b5b076206a1c265e66b44ad2a9dd7cfac6f8dc.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 496, + 298, + 522, + 304 + ], + "lines": [ + { + "bbox": [ + 496, + 298, + 522, + 304 + ], + "spans": [ + { + "bbox": [ + 496, + 298, + 522, + 304 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 496, + 298, + 522, + 304 + ], + "type": "inline_equation", + "content": "\\lambda = 0.5" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 329, + 308, + 399, + 359 + ], + "blocks": [ + { + "bbox": [ + 329, + 308, + 399, + 359 + ], + "lines": [ + { + "bbox": [ + 329, + 308, + 399, + 359 + ], + "spans": [ + { + "bbox": [ + 329, + 308, + 399, + 359 + ], + "type": "image", + "image_path": "39324c923ce22887d83fb8bc80d0dc674cd9f747af6f949d69b1528ded7040d6.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 361, + 380, + 368 + ], + "lines": [ + { + "bbox": [ + 353, + 361, + 380, + 368 + ], + "spans": [ + { + "bbox": [ + 353, + 361, + 380, + 368 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 353, + 361, + 380, + 368 + ], + "type": "inline_equation", + "content": "\\lambda = 0.8" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 378, + 555, + 445 + ], + "lines": [ + { + "bbox": [ + 313, + 378, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 555, + 445 + ], + "type": "text", + "content": "Figure 4. Experiments on a two-dimensional dataset. The first row illustrates the original distribution (left) of pinwheel and the noisy observation " + }, + { + "bbox": [ + 313, + 378, + 555, + 445 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 313, + 378, + 555, + 445 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 378, + 555, + 445 + ], + "type": "inline_equation", + "content": "Y = aX + N" + }, + { + "bbox": [ + 313, + 378, + 555, + 445 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 378, + 555, + 445 + ], + "type": "inline_equation", + "content": "N \\sim \\mathcal{N}(0, \\sigma_n^2\\mathbf{I})" + }, + { + "bbox": [ + 313, + 378, + 555, + 445 + ], + "type": "text", + "content": " (right). The second and third row shows the reconstructions: (a)-(e) variance-scaled reverse diffusion process with different " + }, + { + "bbox": [ + 313, + 378, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 378, + 555, + 445 + ], + "type": "text", + "content": "'s; (f) PSCGAN with " + }, + { + "bbox": [ + 313, + 378, + 555, + 445 + ], + "type": "inline_equation", + "content": "N = 16" + }, + { + "bbox": [ + 313, + 378, + 555, + 445 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 378, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\sigma_z = 1" + }, + { + "bbox": [ + 313, + 378, + 555, + 445 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 400, + 308, + 470, + 359 + ], + "blocks": [ + { + "bbox": [ + 400, + 308, + 470, + 359 + ], + "lines": [ + { + "bbox": [ + 400, + 308, + 470, + 359 + ], + "spans": [ + { + "bbox": [ + 400, + 308, + 470, + 359 + ], + "type": "image", + "image_path": "4d2d216814d00230e7e35e510af8de5e397cb84a79e82ffbce262d9f516fc56d.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 426, + 361, + 447, + 368 + ], + "lines": [ + { + "bbox": [ + 426, + 361, + 447, + 368 + ], + "spans": [ + { + "bbox": [ + 426, + 361, + 447, + 368 + ], + "type": "text", + "content": "(e) " + }, + { + "bbox": [ + 426, + 361, + 447, + 368 + ], + "type": "inline_equation", + "content": "\\lambda = 1" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 471, + 308, + 541, + 359 + ], + "blocks": [ + { + "bbox": [ + 471, + 308, + 541, + 359 + ], + "lines": [ + { + "bbox": [ + 471, + 308, + 541, + 359 + ], + "spans": [ + { + "bbox": [ + 471, + 308, + 541, + 359 + ], + "type": "image", + "image_path": "f66f08aa9b98c16df988e8d174f09edf0847f6ca9e9ec251ad974f7d7c136c1c.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 476, + 361, + 542, + 369 + ], + "lines": [ + { + "bbox": [ + 476, + 361, + 542, + 369 + ], + "spans": [ + { + "bbox": [ + 476, + 361, + 542, + 369 + ], + "type": "text", + "content": "(f) PSCGAN " + }, + { + "bbox": [ + 476, + 361, + 542, + 369 + ], + "type": "inline_equation", + "content": "N = 16, \\sigma_z = 1" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 457, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 555, + 555 + ], + "type": "text", + "content": "by our method versus PSCGAN. Our score-based method achieves a much larger range of tradeoffs compared to the GAN-based approach, demonstrating superior performance. Moreover, unlike PSCGAN, our method requires only a single score network to handle varying measurements and noise levels, offering greater flexibility. Additional experimental results on other distributions are provided in Appendix E." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 562, + 409, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 409, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 409, + 575 + ], + "type": "text", + "content": "4.2 FFHQ dataset" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 581, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 555, + 700 + ], + "type": "text", + "content": "In this subsection, we evaluate our method on the FFHQ " + }, + { + "bbox": [ + 313, + 581, + 555, + 700 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 581, + 555, + 700 + ], + "type": "text", + "content": " dataset [9]. For different measurement scenarios, including Gaussian blur and super-resolution, we use the same pre-trained score-based model from [6], which was trained from scratch using 49k training data for 1M steps. We consider measurement scenarios with noise levels more severe than those in [6, 28, 29]. Specifically, for Gaussian blur, the images are blurred by a Gaussian blur kernel with size " + }, + { + "bbox": [ + 313, + 581, + 555, + 700 + ], + "type": "inline_equation", + "content": "61 \\times 61" + }, + { + "bbox": [ + 313, + 581, + 555, + 700 + ], + "type": "text", + "content": " and a standard deviation of 3.0. For super-resolution, the images are downsampled by a factor of 8." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 326, + 701, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 554, + 714 + ], + "type": "text", + "content": "We compare our method against two benchmarks: (1)" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2382" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 90, + 72, + 261, + 205 + ], + "blocks": [ + { + "bbox": [ + 90, + 72, + 261, + 205 + ], + "lines": [ + { + "bbox": [ + 90, + 72, + 261, + 205 + ], + "spans": [ + { + "bbox": [ + 90, + 72, + 261, + 205 + ], + "type": "image", + "image_path": "e3ff47821395411203f0fcd2c7141da420328371aa43a9658e35d5f360146ce0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 217, + 295, + 240 + ], + "lines": [ + { + "bbox": [ + 55, + 217, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 295, + 240 + ], + "type": "text", + "content": "Figure 5. DP tradeoff on pinwheel dataset traversed by our variance-scaled reverse diffusion process and PSCGAN." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 247, + 296, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 368 + ], + "type": "text", + "content": "We compare the DP tradeoffs provided by our score-based method and PSCGAN [15] on the Gaussian deblur task. Note that PSCGAN is not designed for other measurement scenarios and requires retraining for different noise levels. (2) We evaluate the performance of DiffPIR [29] to see where it falls within the tradeoff spectrum for different inverse problems. DiffPIR is a plug-and-play image restoration method based on diffusion denoising implicit model (DDIM) [18]. It has shown competitive performance on many denoising problems with fewer diffusion steps." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 383, + 179, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 179, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 179, + 393 + ], + "type": "text", + "content": "4.2.1 Gaussian deblur task" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 402, + 295, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 295, + 485 + ], + "type": "text", + "content": "Fig. 6 shows the DP tradeoff on the Gaussian deblur task. We test our variance-scaled reverse diffusion process, PSCGAN [15], and DiffPIR [29] on two different additive noise levels " + }, + { + "bbox": [ + 55, + 402, + 295, + 485 + ], + "type": "inline_equation", + "content": "\\sigma_{n} = 0.3" + }, + { + "bbox": [ + 55, + 402, + 295, + 485 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 402, + 295, + 485 + ], + "type": "inline_equation", + "content": "\\sigma_{n} = 0.5" + }, + { + "bbox": [ + 55, + 402, + 295, + 485 + ], + "type": "text", + "content": ". Note that our sampling method and DiffPIR rely on a pre-trained score network on the FFHQ dataset. For PSCGAN, we use the model trained on images with additive Gaussian noise of " + }, + { + "bbox": [ + 55, + 402, + 295, + 485 + ], + "type": "inline_equation", + "content": "\\sigma_{n} = 0.3" + }, + { + "bbox": [ + 55, + 402, + 295, + 485 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 486, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 641 + ], + "type": "text", + "content": "For noise level " + }, + { + "bbox": [ + 55, + 486, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\sigma_{n} = 0.3" + }, + { + "bbox": [ + 55, + 486, + 295, + 641 + ], + "type": "text", + "content": ", PSCGAN shows suboptimal and more limited tradeoffs compared to our method. To obtain a reconstruction with similar MSE achieved in our " + }, + { + "bbox": [ + 55, + 486, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\lambda = 0.9" + }, + { + "bbox": [ + 55, + 486, + 295, + 641 + ], + "type": "text", + "content": " case, PSCGAN requires averaging 64 sampled images. Its distortion degrades rapidly when fewer images are averaged. Our variance-scaled reverse diffusion sampling achieves both better fidelity and a broader range of DP tradeoffs. Moreover, when the models are tested on the higher noise level of " + }, + { + "bbox": [ + 55, + 486, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\sigma_{n} = 0.5" + }, + { + "bbox": [ + 55, + 486, + 295, + 641 + ], + "type": "text", + "content": ", the overall tradeoff shifts to the left. Our sampling method maintains effective tradeoff traversal. PSCGAN's performance (trained at " + }, + { + "bbox": [ + 55, + 486, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\sigma_{n} = 0.3" + }, + { + "bbox": [ + 55, + 486, + 295, + 641 + ], + "type": "text", + "content": ") deteriorates significantly, showing much larger MSE and excessively high FID values." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "Significant visual improvements can be observed in reconstructed samples in Fig. 7. When " + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": ", the restored images maintain high fidelity and faithfully reconstruct the original images, though appear relatively blurry. As " + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": " increases, the images become progressively sharper, revealing more fine details such as hair strands, eye corners, and back" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "ground elements. When " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "\\lambda = 1" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": ", while the overall reconstruction appears very natural, subtle details deviate from the original images. In comparison, both PSCGAN and DiffPIR have relatively good perceptual quality but compromise fidelity on crucial details. Notable deviations appear in features like eyes and mouth shapes, ultimately altering the facial expression and overall style of faces. We also report more metrics (PSNR and LPIPS) of Gaussian deblur task with additive noise level " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "\\sigma_{n} = 0.3" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": " in Appendix E." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 194, + 440, + 206 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 440, + 206 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 440, + 206 + ], + "type": "text", + "content": "4.2.2 Super-resolution task" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 213, + 555, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 439 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 439 + ], + "type": "text", + "content": "Similar phenomena are observed in the super-resolution (SR) task. By adjusting different " + }, + { + "bbox": [ + 313, + 213, + 555, + 439 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 213, + 555, + 439 + ], + "type": "text", + "content": "'s, we can traverse a wide range of points on the DP plane, allowing flexible selection of the most suitable point for specific application scenarios. Fig. 8 shows the numerical results and reconstruction samples given by our sampling process and DiffPIR. Note that DiffPIR performs close to our " + }, + { + "bbox": [ + 313, + 213, + 555, + 439 + ], + "type": "inline_equation", + "content": "\\lambda = 1" + }, + { + "bbox": [ + 313, + 213, + 555, + 439 + ], + "type": "text", + "content": " case, and achieves a better FID score but larger distortion than our method. While DiffPIR demonstrates strong generative capabilities in both Gaussian deblur and super-resolution tasks with fewer diffusion steps [29], we examine more challenging degradation scenarios (" + }, + { + "bbox": [ + 313, + 213, + 555, + 439 + ], + "type": "inline_equation", + "content": "\\sigma_{n} = 0.3" + }, + { + "bbox": [ + 313, + 213, + 555, + 439 + ], + "type": "text", + "content": " and 0.5 for deblurring, and " + }, + { + "bbox": [ + 313, + 213, + 555, + 439 + ], + "type": "inline_equation", + "content": "8\\times" + }, + { + "bbox": [ + 313, + 213, + 555, + 439 + ], + "type": "text", + "content": " downscaling for SR). Under these severe conditions, DiffPIR's reduced diffusion steps lead to fidelity issues, with generated details less faithful compared to our " + }, + { + "bbox": [ + 313, + 213, + 555, + 439 + ], + "type": "inline_equation", + "content": "\\lambda = 1" + }, + { + "bbox": [ + 313, + 213, + 555, + 439 + ], + "type": "text", + "content": " case. Despite these differences, theoretical connections exist between DPS and DiffPIR [29]. We will consider the exploration of DDIM-style sampling for faster tradeoff traversal as future work." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 441, + 554, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 554, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 554, + 501 + ], + "type": "text", + "content": "In summary, using a single pre-trained score network, the proposed sampling method can effectively traverse a more complete DP tradeoff and achieve better MSE than the benchmarks. Experimental details and more results are included in Appendix E." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 510, + 398, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 398, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 398, + 523 + ], + "type": "text", + "content": "5 Conclusions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 531, + 555, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 555, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 555, + 711 + ], + "type": "text", + "content": "In this paper, we proposed a variance-scaled reverse diffusion process to traverse the DP tradeoff for general inverse problems using a single score-based model. By tuning a parameter that controls the scale of the reverse variance, we can navigate from the MMSE point to the whole posterior distribution, generating a complete DP curve. We proved that the proposed reverse sampling process serves as an optimal solution to the conditional DP tradeoff for multivariate Gaussian distribution. Meanwhile, we conducted experiments on the mixture Gaussian example, two-dimensional datasets, and the FFHQ dataset. Our results show that using a single pre-trained score network, we have achieved a more complete empirical DP tradeoff than the GAN-based method and other inverse problem solvers, demonstrating the effectiveness and flexibility of the proposed framework." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2383" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 72, + 504, + 222 + ], + "blocks": [ + { + "bbox": [ + 108, + 72, + 504, + 222 + ], + "lines": [ + { + "bbox": [ + 108, + 72, + 504, + 222 + ], + "spans": [ + { + "bbox": [ + 108, + 72, + 504, + 222 + ], + "type": "image", + "image_path": "d7cfc4e12adeead9a8e29f2db1ad310908f5f7589eba88437e7b4fad9aa99132.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 233, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 54, + 233, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 54, + 233, + 555, + 289 + ], + "type": "text", + "content": "Figure 6. DP tradeoff on FFHQ dataset. Note that the PSCGAN model in this figure is trained on images contaminated with additive Gaussian noise of " + }, + { + "bbox": [ + 54, + 233, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\sigma_{n} = 0.3" + }, + { + "bbox": [ + 54, + 233, + 555, + 289 + ], + "type": "text", + "content": ". It can be observed that when the noise level in inference time does match the trained model (e.g., when the noise level becomes " + }, + { + "bbox": [ + 54, + 233, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\sigma_{n} = 0.5" + }, + { + "bbox": [ + 54, + 233, + 555, + 289 + ], + "type": "text", + "content": " during the test time), PSCGAN fails to provide valid reconstructions. On the other hand, our variance-scaled reverse diffusion process and DiffPIR method rely on a pre-trained score network on the FFHQ dataset, which is not dedicated to any noisy measurement and is robust for different noise levels." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 82, + 293, + 531, + 488 + ], + "blocks": [ + { + "bbox": [ + 82, + 293, + 531, + 488 + ], + "lines": [ + { + "bbox": [ + 82, + 293, + 531, + 488 + ], + "spans": [ + { + "bbox": [ + 82, + 293, + 531, + 488 + ], + "type": "image", + "image_path": "4e93a5f90a08a1dc6e4f75922c8188002972962e1975fe474b97c7217abac7c2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 497, + 526, + 510 + ], + "lines": [ + { + "bbox": [ + 83, + 497, + 526, + 510 + ], + "spans": [ + { + "bbox": [ + 83, + 497, + 526, + 510 + ], + "type": "text", + "content": "Figure 7. Samples from our sampling process, PSCGAN, and DiffPIR for Gaussian deblur task with noise level " + }, + { + "bbox": [ + 83, + 497, + 526, + 510 + ], + "type": "inline_equation", + "content": "\\sigma_{n} = 0.3" + }, + { + "bbox": [ + 83, + 497, + 526, + 510 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 83, + 522, + 531, + 670 + ], + "blocks": [ + { + "bbox": [ + 83, + 522, + 531, + 670 + ], + "lines": [ + { + "bbox": [ + 83, + 522, + 531, + 670 + ], + "spans": [ + { + "bbox": [ + 83, + 522, + 531, + 670 + ], + "type": "image", + "image_path": "9bba45efe3eafb02fe9d6e7a60271a4a90ed6978e77bc5d4e40c3f2da34c4c7a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 678, + 533, + 691 + ], + "lines": [ + { + "bbox": [ + 77, + 678, + 533, + 691 + ], + "spans": [ + { + "bbox": [ + 77, + 678, + 533, + 691 + ], + "type": "text", + "content": "Figure 8. Samples from our sampling process, PSCGAN, and DiffPIR for super-resolution task with downsampling scale " + }, + { + "bbox": [ + 77, + 678, + 533, + 691 + ], + "type": "inline_equation", + "content": "\\times 8" + }, + { + "bbox": [ + 77, + 678, + 533, + 691 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2384" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 283 + ], + "type": "text", + "content": "The work of Yuhan Wang and Ying-Jun Angela Zhang is supported in part by the General Research Fund (project number 14202421, 14214122, 14202723), Area of Excellence Scheme grant (project number AoE/E-601/22-R), and NSFC/RGC Collaborative Research Scheme (project number CRS_HKUST603/22), all from the Research Grants Council of Hong Kong. The work of Suzhi Bi is supported in part by the Natural Science Foundation of China under Project 62271325; in part by Guangdong Basic and Applied Basic Research Foundation under Project 2024B1515020089; in part by Shenzhen Science and Technology Program under Project 20220810142637001. The work of Xiaojun Yuan is supported in part by Sichuan Science and Technology Program under Grant 2024ZYD0036 and in part by National Key Laboratory of Wireless Communications Foundation under Grant IFN20230204." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 297, + 115, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 115, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 115, + 308 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 316, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 316, + 295, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 316, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 61, + 316, + 295, + 348 + ], + "type": "text", + "content": "[1] Brian D.O. Anderson. Reverse-time diffusion equation models. Stochastic Processes and their Applications, 12(3):313-326, 1982." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 350, + 295, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 350, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 61, + 350, + 295, + 384 + ], + "type": "text", + "content": "[2] Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6228-6237, 2018." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 385, + 296, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 385, + 296, + 428 + ], + "spans": [ + { + "bbox": [ + 61, + 385, + 296, + 428 + ], + "type": "text", + "content": "[3] Yochai Blau and Tomer Michaeli. Rethinking lossy compression: The rate-distortion-perception tradeoff. In Proceedings of the 36th International Conference on Machine Learning (ICML), pages 675-685. PMLR, 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 430, + 296, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 430, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 62, + 430, + 296, + 474 + ], + "type": "text", + "content": "[4] Valentin De Bortoli, James Thornton, Jeremy Heng, and Arnaud Doucet. Diffusion schrödinger bridge with applications to score-based generative modeling. In Advances in Neural Information Processing Systems (NIPS), 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 476, + 295, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 476, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 476, + 295, + 520 + ], + "type": "text", + "content": "[5] Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. In Advances in Neural Information Processing Systems (NIPS), 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 522, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 522, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 62, + 522, + 294, + 576 + ], + "type": "text", + "content": "[6] Hyungjin Chung, Jeongsol Kim, Michael Thompson McCann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In The Eleventh International Conference on Learning Representations (ICLR), 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 578, + 295, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 578, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 62, + 578, + 295, + 620 + ], + "type": "text", + "content": "[7] Dror Freirich, Tomer Michaeli, and Ron Meir. A theory of the distortion-perception tradeoff in wasserstein space. In Advances in Neural Information Processing Systems (NIPS), 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 624, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 295, + 667 + ], + "type": "text", + "content": "[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Proceedings of the 34th International Conference on Neural Information Processing Systems (NIPS), 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 295, + 713 + ], + "type": "text", + "content": "[9] Tero Karras, Samuli Laine, and Timo Aila. A Style-Based Generator Architecture for Generative Adversarial Networks. IEEE Transactions on Pattern Analysis & Machine Intelligence, 43(12):4217-4228, 2021." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[10] Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. In Advances in Neural Information Processing Systems (NIPS), 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 108, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 161 + ], + "type": "text", + "content": "[11] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "text", + "content": "[12] Dong Liu, Haochen Zhang, and Zhiwei Xiong. On the classification-distortion-perception tradeoff. In Advances in Neural Information Processing Systems (NIPS). Curran Associates, Inc., 2019." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 208, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 252 + ], + "type": "text", + "content": "[13] Razvan V Marinescu, Daniel Moyer, and Polina Golland. Bayesian image reconstruction using deep generative models. arXiv preprint, 2021. [Online]. Available: https://arxiv.org/abs/2012.04567." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 254, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 296 + ], + "type": "text", + "content": "[14] Anush Krishna Moorthy and Alan Conrad Bovik. Blind image quality assessment: From natural scene statistics to perceptual quality. IEEE transactions on Image Processing, 20 (12):3350-3364, 2011." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 298, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 363 + ], + "type": "text", + "content": "[15] Guy Ohayon, Theo Adrai, Gregory Vaksman, Michael Elad, and Peyman Milanfar. High Perceptual Quality Image Denoising with a Posterior Sampling CGAN. In 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 1805-1813, Los Alamitos, CA, USA, 2021. IEEE Computer Society." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 365, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 387 + ], + "type": "text", + "content": "[16] Victor M. Panaretos and Yoav Zemel. An Invitation to Statistics in Wasserstein Space. Springer Cham, 2020." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "type": "text", + "content": "[17] Slavko Simic. On a global upper bound for jensen's inequality. Journal of Mathematical Analysis and Applications, 343 (1):414-419, 2008." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 422, + 553, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 455 + ], + "type": "text", + "content": "[18] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations (ICLR), 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 456, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 499 + ], + "type": "text", + "content": "[19] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In International Conference on Learning Representations (ICLR), 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "type": "text", + "content": "[20] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In Advances in Neural Information Processing Systems (NIPS). Curran Associates, Inc., 2019." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 545, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 600 + ], + "type": "text", + "content": "[21] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations (ICLR), 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 601, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 601, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 315, + 601, + 553, + 634 + ], + "type": "text", + "content": "[22] Pascal Vincent. A connection between score matching and denoising autoencoders. *Neural Computation*, 23(7):1661-1674, 2011." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[23] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In The European Conference on Computer Vision Workshops (ECCVW), 2018." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 691, + 553, + 713 + ], + "type": "text", + "content": "[24] Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. In" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2385" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 373 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "The Eleventh International Conference on Learning Representations (ICLR), 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 160 + ], + "type": "text", + "content": "[25] Jay Whang, Maurizio Delbracio, Hossein Talebi, Chitwan Sahara, Alexandros G. Dimakis, and Peyman Milanfar. Deblurring via Stochastic Refinement. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16272-16282, Los Alamitos, CA, USA, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 206 + ], + "type": "text", + "content": "[26] Xingyu Xu and Yuejie Chi. Provably robust score-based diffusion posterior sampling for plug-and-play image reconstruction. In The Thirty-eighth Annual Conference on Neural Information Processing Systems (NIPS), 2024." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "text", + "content": "[27] Danna Xue, Luis Herranz, Javier Vazquez Corral, and Yanning Zhang. Burst perception-distortion tradeoff: Analysis and evaluation. In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 307 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 307 + ], + "type": "text", + "content": "[28] Zhipeng Xue, Penghao Cai, Xiaojun Yuan, and Xiqi Gao. Score-based variational inference for inverse problems. arXiv preprint, 2024. [Online]. Available: https://arxiv.org/abs/2410.05646." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 373 + ], + "type": "text", + "content": "[29] Yuanzhi Zhu, Kai Zhang, Jingyun Liang, Jiezhang Cao, Bihan Wen, Radu Timofte, and Luc Van Gool. Denoising Diffusion Models for Plug-and-Play Image Restoration. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1219–1229, Los Alamitos, CA, USA, 2023. IEEE Computer Society." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2386" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TreeMeshGPT_ Artistic Mesh Generation with Autoregressive Tree Sequencing/cc43c3b2-b6bf-472a-9dd6-c74d4f03cee9_content_list.json b/2025/TreeMeshGPT_ Artistic Mesh Generation with Autoregressive Tree Sequencing/cc43c3b2-b6bf-472a-9dd6-c74d4f03cee9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..edea73850ad89220a621fdcc888813b4aad70d7d --- /dev/null +++ b/2025/TreeMeshGPT_ Artistic Mesh Generation with Autoregressive Tree Sequencing/cc43c3b2-b6bf-472a-9dd6-c74d4f03cee9_content_list.json @@ -0,0 +1,2048 @@ +[ + { + "type": "text", + "text": "TreeMeshGPT: Artistic Mesh Generation with Autoregressive Tree Sequencing", + "text_level": 1, + "bbox": [ + 101, + 130, + 897, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stefan Lionar $^{1,2,3}$", + "bbox": [ + 263, + 179, + 403, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiabin Liang $^{1,2,3}$", + "bbox": [ + 439, + 179, + 568, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gim Hee Lee3", + "bbox": [ + 609, + 179, + 727, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Sea AI Lab", + "bbox": [ + 215, + 208, + 318, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Garena", + "bbox": [ + 375, + 208, + 444, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ National University of Singapore", + "bbox": [ + 503, + 208, + 774, + 227 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e748c1b371f1c49429413a1115bc2da6dbe308a8f69547ee0a28b5e1e7f76ff0.jpg", + "image_caption": [ + "Figure 1. Artistic meshes generated by TreeMeshGPT. Our method offers a novel sequencing approach for artistic mesh generation using autoregressive Transformer decoder by retrieving the next token from a dynamically growing tree structure. In our experiment with 7-bit discretization, TreeMeshGPT supports meshes with up to 5,500 triangular faces under strong point cloud conditioning. $^{12}$" + ], + "image_footnote": [], + "bbox": [ + 96, + 243, + 906, + 501 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 565, + 326, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce TreeMeshGPT, an autoregressive Transformer designed to generate high-quality artistic meshes aligned with input point clouds. Instead of the conventional next-token prediction in autoregressive Transformer, we propose a novel Autoregressive Tree Sequencing where the next input token is retrieved from a dynamically growing tree structure that is built upon the triangle adjacency of faces within the mesh. Our sequencing enables the mesh to extend locally from the last generated triangular face at each step, and therefore reduces training difficulty and improves mesh quality. Our approach represents each triangular face with two tokens, achieving a compression rate of approximately $22\\%$ compared to the naive face tokenization. This efficient tokenization enables our model to generate highly detailed artistic meshes with strong point cloud conditioning, surpassing previous methods in both capacity and fi", + "bbox": [ + 88, + 597, + 485, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "delyity. Furthermore, our method generates mesh with strong normal orientation constraints, minimizing flipped normals commonly encountered in previous methods. Our experiments show that TreeMeshGPT enhances the mesh generation quality with refined details and normal orientation consistency.", + "bbox": [ + 511, + 566, + 906, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 691, + 643, + 707 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent advancements in 3D generation, representations such as voxels, point clouds, and implicit functions are often utilized [17, 35]. After the generation process, these representations are converted into meshes using techniques like Marching Cubes [19], which result in dense, overtessellated triangular meshes. However, these dense meshes are unsuitable for applications that require real-time rendering, such as gaming and virtual reality. Although many mesh down-sampling algorithms can reduce the number of triangles, they also degrade mesh quality and produce messy, unstructured wireframes. In contrast, skilled artists can create highly compact meshes with minimal triangles", + "bbox": [ + 509, + 719, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "${}^{1}$ TreeMeshGPT supports a higher face count with finer discretization as it preserves the required manifold connectivity condition.", + "bbox": [ + 89, + 849, + 482, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "2Our code is available at: https://github.com/sail-sg/TreeMeshGPT.", + "bbox": [ + 89, + 875, + 480, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26608", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "while preserving the object's sharp details. Additionally, the wireframes created by artists are more regular, aesthetically pleasing, and better aligned with the object's feature or semantic boundaries, which facilitates further human interactions, such as editing and animation. However, the process of manually creating artist-quality meshes is highly time-consuming and labor-intensive.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This laborous process highlights the need for automated methods that can replicate the quality of artist-created meshes without requiring extensive manual effort. MeshAnything [3] seeks to bridge the gap between advancements in 3D generation and artist-quality mesh creation by adding point cloud condition to the artistic mesh generation Transformer initially proposed by MeshGPT [28]. Point clouds are chosen as the condition because they are either the direct output or can be obtained conveniently from the generated Marching Cubes meshes of the advanced 3D generation techniques.", + "bbox": [ + 89, + 199, + 480, + 366 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "MeshAnything [3] represents each triangular face with 9 latent tokens, leading to long sequences and limiting artistic mesh generation to 800 faces due to the Transformer's quadratic complexity. This constraint poses challenges for real-world applications, which often require meshes with significantly higher face counts to accurately represent complex objects and environments. In the subsequent works, MeshAnything V2 [4] and EdgeRunner [29] leverage triangle adjacency to create shorter sequences to represent the same meshes. Consequently, they are able to generate meshes with up to 1,600 and 4,000 faces, respectively. However, many real-world applications demand meshes with higher face count to accurately represent detailed surface topology. Additionally, challenges remain in generating high-quality meshes free from artifacts such as gaps, missing components, and flipped normals.", + "bbox": [ + 89, + 369, + 482, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To further improve tokenization efficiency and mesh quality, we introduce TreeMeshGPT. Unlike previous methods that rely on conventional next-token prediction in autoregressive Transformers, TreeMeshGPT introduces a novel Autoregressive Tree Sequencing approach. Instead of sequentially predicting next tokens, our method retrieves the next token from a dynamically growing tree structure built upon triangle adjacency within the mesh. This strategy allows the mesh to expand locally from the last generated triangular face at each step, and thus reducing training difficulty and enhancing mesh quality. Our approach represents each triangular face with two tokens, achieving a compression rate of $22\\%$ compared to the naive face tokenization of 9 tokens per face. This efficient tokenization technique pushes the boundary of artistic mesh generation. With 7-bit discretization, it enables the generation of meshes with up to 5,500 triangles under a strong point cloud condition of 2,048 tokens. Furthermore, our method generates meshes with strong normal orientation constraints,", + "bbox": [ + 89, + 613, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "minimizing flipped normals commonly encountered in MeshAnything [3] and MeshAnythingV2 [4].", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 532, + 122, + 834, + 136 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel Autoregressive Tree Sequencing technique that efficiently represents two tokens per triangular face.", + "- Our proposed tokenization enables the training of a 7-bit discretization artistic mesh generative model with strong point cloud condition, capable of generating high-quality meshes with up to 5,500 faces.", + "- Extensive experiments show that our model can generate higher quality meshes and can generalize to real-world 3D scans." + ], + "bbox": [ + 511, + 140, + 903, + 290 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 306, + 653, + 321 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Mesh Extraction", + "text_level": 1, + "bbox": [ + 513, + 330, + 678, + 345 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Constructing a mesh from other 3D representations has been a research focus for decades. Among many successful methods [1, 15, 19], Marching Cubes [19] is the most widely used. It divides a scalar field into cubes and extracts triangles to approximate the isosurface. It is simple yet robust, producing watertight and 2-manifold results. Many improvements such as Dual Contouring [12] and Dual Marching Cubes [24] have been developed to enhance its capabilities. Another well-known method is Poisson reconstruction [13], which uses point clouds and normals as boundary conditions to solve a scalar field defined in 3D space, then applies Marching Cubes to extract the mesh. However, these approaches often focus on representing shapes with dense, over-tessellated meshes, resulting in messy, unstructured wireframes. This makes them unsuitable for downstream workflows that require efficient and structured meshes, such as real-time rendering, animation rigging, and editing. The dense, over-tessellated meshes not only increase computational load but also lack the regularity and semantic alignment necessary for the downstream processes.", + "bbox": [ + 511, + 353, + 906, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2.3D Generation", + "text_level": 1, + "bbox": [ + 513, + 681, + 663, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "After the great success of 2D image generation [26], 3D generation has become a promising research direction. This field focuses on generating 3D assets for industries such as gaming, film, and AR/VR. Due to the limited availability of 3D data, early methods [11, 21, 25, 27] relied on optimizing underlying representations to mimic conditioning from 2D images or multi-view 2D images.", + "bbox": [ + 511, + 703, + 906, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the introduction of large-scale datasets [6, 7], feedforward 3D generation techniques, such as those in [10, 18, 36], have become feasible. These techniques significantly improve generation speed compared to optimization-based methods. However, the resulting meshes often suffer from lower quality and lack diversity.", + "bbox": [ + 511, + 810, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26609", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by the success of 2D diffusion models in image generation, many researchers have attempted to apply diffusion techniques directly to 3D data [5, 23, 32, 38]. For instance, CLAY [38], a transformer-based 3D latent diffusion model, achieves state-of-the-art results in high-quality shape generation. Despite these advances, these generation methods often require post-conversion for downstream applications, which remains a non-trivial challenge.", + "bbox": [ + 89, + 90, + 480, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Autoregressive Mesh Generation", + "text_level": 1, + "bbox": [ + 89, + 223, + 377, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address these limitations, recent approaches have leveraged autoregressive models for direct mesh generation [2, 22, 28, 33]. MeshGPT [28] was the first to tokenize a mesh through face sorting and compress it using a VQ-VAE [16, 30], followed by an autoregressive transformer to predict the compressed token sequence. This method enables the generation of meshes with direct supervision from artist-created topology information, which is often absent in previous approaches.", + "bbox": [ + 89, + 246, + 482, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Subsequent works [2, 3, 33] explored more efficient representations and incorporated input conditioning, such as point clouds and images. However, these methods are limited to generating meshes with fewer than 800 faces due to the long sequence lengths and the quadratic computational cost of transformers. MeshAnythingV2 [4] and EdgeRunner [29] introduced more compact mesh tokenization techniques that leverage triangle adjacency, increasing the maximum face count to 1,600 and 4,000, respectively. Meshtron [9] proposes hourglass architecture and sliding window inference to scale up face count capacity while using the naive tokenization [2]. In a concurrent work, BPT [34] proposes a compact tokenization using block-wise indexing and localized patch aggregation. These methods rely on the next-token prediction commonly used in large language models (LLMs) and autoregressive image generation. In contrast, we offer a novel sequencing strategy based on a dynamically growing tree structure, aiming to increase the maximum face count and improve the overall quality of the generated meshes.", + "bbox": [ + 89, + 383, + 482, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 700, + 181, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Autoregressive Tree Sequencing", + "text_level": 1, + "bbox": [ + 89, + 726, + 372, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Tokenization plays a crucial role in autoregressive models, particularly in complex tasks like 3D mesh generation, where both the quality and efficiency of tokenization significantly affect model performance and scalability. In the context of mesh generation, tokenization involves encoding vertices, edges, or faces into sequential tokens that the model can process step-by-step. Drawing insights from LLMs, previous autoregressive mesh generation methods follow next-token prediction strategy that explicitly uses each output as the input for the subsequent step.", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our approach differs from those methods by using a tree-based traversal scheme to grow the mesh during the generation process. Specifically, in TreeMeshGPT, the input to the Transformer decoder are directed mesh edges, represented as $\\mathbf{I} = \\{(v_1^n,v_2^n)\\}_{n = 1}^N\\in \\mathbb{R}^{N\\times 6}$ , where each $(v_{1}^{n},v_{2}^{n})$ denotes a pair of dequantized vertices for the generation at step $n$ . At each step, the Transformer decoder makes a localized prediction to either add a new vertex $v_{3}^{n}$ to expand the mesh by connecting to the initial pair of vertices $(v_{1}^{n},v_{2}^{n})$ or predict [STOP] label indicating that no further expansion should occur from the input edge. TreeMeshGPT leverages a tree traversal process to construct the sequential input-output pairs for the autoregressive generation, described as follows (Note: the sequence order description below omits auxiliary tokens for simplicity).", + "bbox": [ + 511, + 90, + 903, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sequence order: We utilize a half-edge data structure with depth-first-search (DFS) traversal to construct the sequential input-output pairs, $(\\mathbf{I},\\mathbf{O}) = \\{(I_n,o_n)\\}_{n = 1}^N$ , accompanied by a dynamic stack $\\mathbf{S}$ to manage the traversal process.", + "bbox": [ + 511, + 321, + 903, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The traversal process starts from a directed edge in a mesh, in which we determine if this edge has an opposite vertex that forms a triangle, is a boundary, or if the triangle has already been visited. When a new triangle is formed, the output $o_n$ is defined as the opposite vertex $v_3^n$ . Two new edges are created by connecting $v_3^n$ to the initial edge's vertices $I_n = (v_1^n, v_2^n)$ , resulting in edges $(v_3^n, v_2^n)$ and $(v_1^n, v_3^n)$ . These edges are directed in a counter-clockwise orientation on the potential next adjacent faces, as enforced by the half-edge data structure. The newly created edges are then pushed onto the stack S for continued traversal:", + "bbox": [ + 511, + 382, + 906, + 547 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} := \\mathbf {S} \\odot \\left(v _ {1} ^ {n}, v _ {3} ^ {n}\\right) \\quad \\text {a n d} \\quad \\mathbf {S} := \\mathbf {S} \\odot \\left(v _ {3} ^ {n}, v _ {2} ^ {n}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 555, + 861, + 573 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\odot$ represents the operation of pushing the edge $(v_{i}, v_{j})$ onto the top of stack $\\mathbf{S}$ . Conversely, if the input edge is a boundary or if adding a new vertex forms a previously visited triangle, the output $o_{n}$ is set to the [STOP] label. In this case, no new vertex or edge is added.", + "bbox": [ + 511, + 580, + 903, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The input for the next step, $n + 1$ , is obtained by popping the top edge from the stack:", + "bbox": [ + 511, + 656, + 903, + 686 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {n + 1} = \\left(v _ {1} ^ {n + 1}, v _ {2} ^ {n + 1}\\right) := \\operatorname {t o p} (\\mathbf {S}), \\quad \\mathbf {S} := \\mathbf {S} \\setminus \\operatorname {t o p} (\\mathbf {S})\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 693, + 880, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\operatorname{top}(\\mathbf{S})$ retrieves the current top edge of the stack $\\mathbf{S}$ , and $\\mathbf{S} := \\mathbf{S} \\setminus \\operatorname{top}(\\mathbf{S})$ updates $\\mathbf{S}$ by removing this top edge. We initialize the stack with a directed edge at the lowest position of the mesh and its twin. The traversal then proceeds until all triangles are visited. A simple illustration of this sequencing process is provided in Figure 2.", + "bbox": [ + 511, + 719, + 903, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that a mesh may consist of multiple connected components. A component begins from an initial edge, expands as edges are added to a stack, and is considered fully traversed once the stack is empty. When there are multiple components in a mesh, the traversal of the first component starts from the edge with the lowest position in the mesh", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26610", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2f6c21b0591613b334314798257907ed55f43d7548a29b9eb3dec4589a9eb30f.jpg", + "image_caption": [ + "Figure 2. Illustration of the sequence order in our Autoregressive Tree sequencing. a). A small subset of a triangular mesh. [STOP] indicates boundary. b). An equivalent tree representation of the mesh. In this tree, each node is represented as a directed edge from a pair of vertices. The root is initialized with two child nodes: $(v_0, v_1)$ and its twin $(v_1, v_0)$ . A DFS traversal then proceeds to create the input-output sequence. c). Dynamic stack from the DFS traversal. The stack is initialized with $(v_0, v_1)$ and its twin $(v_1, v_0)$ . The input $I_n$ is always obtained from the top of the stack. Thus, $I_1 = (v_0, v_1)$ at step 1. The opposite vertex of $I_1$ is $v_2$ and consequently, $o_1$ is set to $v_2$ . Two new edges are then formed by connecting the opposite vertex to the initial pair of vertices: $(v_2, v_1)$ and $(v_0, v_2)$ . The direction is enforced to be counter-clockwise on the potential next adjacent faces. At step 2, $I_2 = (v_2, v_1)$ . Since $I_2$ is a boundary, $o_2$ is set to [STOP] label and no new edge is added to the stack. Step 3 and onwards follow the same traversal process. d). Transformer decoder sequence. The sequence in the Transformer decoder follows the input-output pairs from the tree traversal. The auxiliary tokens to initialize the generation of a connected component and the [EOS] are also added to the input-output sequence." + ], + "image_footnote": [], + "bbox": [ + 89, + 88, + 330, + 265 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b87f4dd24584de014f7a67b0c1f2e5cbe2bbdf77cb82fb67b5cf45c3f7d17ee0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 88, + 580, + 266 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/78cb86f35cab0c13cc144d804fe83c8f3f1627a462130bcbfc960a5969392587.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 88, + 903, + 266 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and continues until no edge remains in the stack. Each subsequent component begins from the next available edge positioned at the lowest among the remaining unvisited faces.", + "bbox": [ + 88, + 436, + 482, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Generation process: To initialize the generation of a mesh component, an auxiliary token $[\\mathrm{SOS}]\\in \\mathbb{R}^D$ is used as the input to predict the first vertex $v_{1}$ . For the next step, a second auxiliary token $[\\mathrm{SOS2}]\\in \\mathbb{R}^{C}$ , concatenated with the embedded representation of $v_{1}$ , serves as input to predict the second vertex $v_{2}$ . Once these initial vertices are predicted, the mesh generation proceeds through our Autoregressive Tree Sequencing with a stack initialized by $(v_{1},v_{2})$ and its twin $(v_{2},v_{1})$ . When the stack is empty—indicating the current component is complete—a new component is initialized by reintroducing the [SOS] and [SOS2] tokens. After all components have been generated, the sequence is terminated with an [EOS] label.", + "bbox": [ + 88, + 486, + 482, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The final mesh is constructed by gathering the faces formed from the initial input vertex pairs and their predicted opposite vertices: $\\mathcal{M} = \\bigcup_{n=1}^{N}(v_1^n,v_2^n,v_3^n)$ for the generation steps where $I_n \\notin \\{[\\mathrm{SOS}], [\\mathrm{SOS2}]\\}$ and $o_n \\notin \\{[\\mathrm{STOP}], [\\mathrm{EOS}]\\}$ .", + "bbox": [ + 89, + 684, + 482, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input embedding: We employ a positional embedding function from [37] to encode each vertex into a high-dimensional space, capturing its positional information across multiple frequency bases. This embedding function $\\mathrm{PosEmbed}(.) : \\mathbb{R}^3 \\to \\mathbb{R}^C$ maps 3D coordinates to a $C$ -dimensional embedding. For each edge, the embeddings of its vertex pair are concatenated, creating a representation in $\\mathbb{R}^{2C}$ , which is subsequently passed through an MLP to map it to the Transformer's hidden dimension, $\\mathbb{R}^{2C} \\to \\mathbb{R}^D$ .", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vertex prediction: In prior works on mesh generation, each vertex is represented as a sequence of three tokens corresponding to its quantized $x-$ , $y-$ , and $z$ -coordinates. Specifically, to predict a single vertex's position, those models generate each coordinate independently as separate tokens in sequence. This approach leads to longer sequences, as each vertex requires three distinct tokens. In contrast, our method predicts the vertex' quantized $x-$ , $y-$ , and $z$ -coordinates in one sequence length by using hierarchical MLP heads. This hierarchical approach maintains the sequential nature in predicting the $x-$ , $y-$ , and $z$ -coordinates. Further details can be found in the supplementary material. As shown in the ablation study (Section 4.5), our hierarchical MLP heads result in easier coordinate sampling compared to prediction heads that predict the $x-$ , $y-$ , and $z$ -coordinates simultaneously.", + "bbox": [ + 511, + 436, + 906, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Advantages: Our Autoregressive Tree Sequencing approach adds only two sequence steps per triangular face as each face introduces two new nodes during the tree traversal process. Additionally, since most meshes consist of only a few connected components, our method requires minimal auxiliary tokens. This efficient sequencing achieves a compression rate of approximately $22\\%$ for most meshes compared to naive tokenization, which uses 9 tokens per face. Our compression rate is thus approximately double that of methods like MeshAnythingV2 [4] and EdgeRunner [29]. Additionally, by using a dynamic stack to manage the input sequence, our method allows the Transformer to focus solely on making localized predictions at each step, hence improving training efficiency. Furthermore, our method", + "bbox": [ + 511, + 689, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26611", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "generates mesh with strong normal orientation constraints, minimizing flipped normals commonly encountered in MeshAnything [3] and MeshAnythingV2 [4].", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss function: We aim to train the Transformer decoder $\\theta$ to maximize the likelihood of generating the sequence of outputs $\\{o_n\\}_{n=1}^N$ given the input sequence $\\{I_n\\}_{n=1}^N$ :", + "bbox": [ + 89, + 140, + 482, + 186 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\prod_ {n = 1} ^ {N} P \\left(o _ {n} \\mid I _ {\\leq n}; \\theta\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 186, + 480, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To this end, given the ground truth and predicted values with teacher-forcing across all steps, denoted by $\\mathbf{O} = \\{\\mathbf{O}_x,\\mathbf{O}_y,\\mathbf{O}_z\\}$ and $\\hat{\\mathbf{O}} = \\{\\hat{\\mathbf{O}}_x,\\hat{\\mathbf{O}}_y,\\hat{\\mathbf{O}}_z\\}$ , respectively, where each $\\mathbf{O}_i$ represents the discretized vertex coordinate along a specific axis, we use a loss function defined as the sum of cross-entropy losses for each coordinate:", + "bbox": [ + 89, + 224, + 482, + 316 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {C E}} \\left(\\mathbf {O} _ {x}, \\hat {\\mathbf {O}} _ {x}\\right) + \\mathcal {L} _ {\\mathrm {C E}} \\left(\\mathbf {O} _ {y}, \\hat {\\mathbf {O}} _ {y}\\right) + \\mathcal {L} _ {\\mathrm {C E}} \\left(\\mathbf {O} _ {z}, \\hat {\\mathbf {O}} _ {z}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 324, + 482, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To incorporate stopping conditions, we add the [STOP] and [EOS] labels to the class selection on the height axis, extending it with two additional classes beyond the discretized coordinate classes.", + "bbox": [ + 89, + 351, + 482, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 425, + 223, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 89, + 450, + 186, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use Objaverse [7] meshes as our training dataset. To ensure high-quality meshes, we select meshes that meet the half-edge traversal requirement, i.e., they are manifold and have no flipped normal. All meshes are preprocessed by centering and normalizing them within a cube spanning $[-0.5, 0.5]$ . We apply 7-bit discretization, remove any duplicate triangles, and choose meshes with less than 5,500 faces. Additionally, we perform orthographic projections and exclude meshes where one of the projections has extremely small area or contains more than one cluster. After filtering, we retain a total of 75,000 meshes, of which 1,000 are reserved for validation, with the remainder used for training. To increase data diversity, we apply the following augmentations:", + "bbox": [ + 89, + 472, + 482, + 681 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Scaling: Each axis is scaled independently by a factor randomly chosen from the range [0.75, 0.95].", + "- Rotation: We first apply a $90^{\\circ}$ or $-90^{\\circ}$ rotation along the $x$ - or $y$ -axis with a probability of 0.3. Afterward, we perform a rotation around the $z$ -axis with an angle uniformly sampled from $[-180^{\\circ}, 180^{\\circ}]$ ." + ], + "bbox": [ + 89, + 684, + 482, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 784, + 307, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Point cloud conditioning: Following previous approaches [3, 4, 29], we adopt point cloud conditioning to provide practical guidance for the generation process. To achieve this, we sample a point cloud from the input mesh surface and apply a lightweight encoder [3, 4, 37, 38]. Specifically, we sample 8192 points, denoted as $\\mathbf{X} \\in$", + "bbox": [ + 89, + 809, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathbb{R}^{8192 \\times 3}$ , from the mesh surface. A cross-attention layer is then used to encode these points into a latent code:", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} = \\operatorname {C r o s s A t t} (\\mathbf {Q}, \\operatorname {P o s E m b e d} (\\mathbf {X})) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 131, + 903, + 148 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{Q} \\in \\mathbb{R}^{2048 \\times C}$ represents query embeddings with a shorter sequence length of 2048 and hidden dimension of $C$ , $\\mathrm{PosEmbed}(.)$ is the same input embedding function in our Autoregressive Tree Sequencing, and $\\mathbf{Z} \\in \\mathbb{R}^{2048 \\times L}$ is the resulting latent code. The latent code $\\mathbf{Z}$ is then prepended to the initial [SOS] token in the Transformer decoder to provide point cloud-based conditioning for mesh generation.", + "bbox": [ + 511, + 156, + 903, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Architecture details: Our model employs a Transformer decoder with 24 layers, 16 attention heads, and a hidden dimension of 1024 and adds the sinusoidal positional encoding [31] to encode the token position. We apply a full self-attention for the latent code condition and a causal self-attention mask for the autoregressive decoder. PyTorch's FlexAttention is used to implement this attention mask efficiently. Additionally, we adopt fp16 mixed-precision to optimize computational speed and memory efficiency. We set the hidden dimension $C$ of the positional embedding (PosEmbed) to 512 and use 7-bit quantization to discretize the coordinate output into 128 classes.", + "bbox": [ + 511, + 267, + 906, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training details: We use AdamW [14, 20] with a learning rate of $10^{-4}$ , $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.99$ as the optimizer. Our model is trained with $8 \\times \\mathrm{A100 - 80GB}$ GPUs for 5 days with an effective batch size of 128.", + "bbox": [ + 511, + 452, + 905, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sampling strategy: We use a multinomial sampling strategy with a top- $k$ of 5 during the generation process. Empirically, we find that adjusting the temperature at different stages achieves an optimal balance between diversity and generation quality. Specifically, we set the temperature to 1 when the stack length is below 10, reduce it to 0.4 when the stack length is below 30, and further decrease it to 0.2 beyond that.", + "bbox": [ + 511, + 518, + 905, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference adjustment: We find that a few adjustments during inference can improve the generation performance. First, we keep track of the generated faces for each step. If our model predicts a vertex that would form a triangle that is a duplicate to the previously generated faces, we adjust the prediction to [STOP] and retrieve the next input from the top of the stack. This checking operation is fast since the faces are of discrete tensor.", + "bbox": [ + 511, + 643, + 905, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, we observe that our model often struggles to predict the [EOS] label in longer sequences. In these cases, while the [EOS] label consistently appears among the top 5 logits, it is rarely sampled. To address this, we apply an addition factor to the logit of [EOS], incrementally increasing this factor each time [EOS] appears in the top 5 logits. To avoid early [EOS] prediction after this adjustment, we bypass multinomial top- $k$ sampling for [EOS] and select it only when it becomes the highest logit.", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26612", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Results on Objverse Dataset", + "text_level": 1, + "bbox": [ + 89, + 90, + 352, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.1. Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 112, + 269, + 127 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present the qualitative results of mesh generation conditioned on input point cloud in Figure 3, comparing our method with MeshAnything [3] and MeshAnythingV2 [4] as the baselines. As shown, our method demonstrates a notable improvement to generate meshes with higher face counts and refined details.", + "bbox": [ + 89, + 131, + 483, + 220 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.2. Quantitative Results", + "text_level": 1, + "bbox": [ + 89, + 229, + 279, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the following metrics to assess the quality of the generated meshes:", + "bbox": [ + 89, + 250, + 482, + 279 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Chamfer Distance (CD). It provides an indication of geometric accuracy by measuring the average closest-point distance between points sampled from the source and reference meshes, computed bidirectionally. Lower values of CD indicate better geometric accuracy.", + "- Normal Consistency (NC). It evaluates the surface orientation alignment between the source and reference meshes. Specifically, it measures the cosine similarity between the normals of each face in the source mesh and the nearest face in the reference mesh, computed bidirectionally. Higher NC values signify better normal alignment. We also report absolute Normal Consistency (|NC|), which disregards the sign, focusing solely on the magnitude of similarity. The mathematical details are provided in the supplementary materials." + ], + "bbox": [ + 89, + 282, + 483, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Objaverse evaluation set: We ran the evaluation on 200 samples in our validation set with 1 generation seed for each model. The quantitative results presented in Table 1, indicate that our model generates meshes more faithful to the ground truth meshes compared to the baselines. The NC values of MeshAnything [3] and MeshAnythingV2 [4] are relatively low due to flipped normals, which leads to inconsistencies in the sign of the normals. In contrast, our method generates meshes with consistent normal orientation.", + "bbox": [ + 89, + 515, + 483, + 650 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/55a9155c167c0c9b9c570c1276bc09c280235cce01181bacbb3167b723855313.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelCD↓NC↑|NC| ↑
MeshAnything [3]0.01150.2230.853
MeshAnythingV2 [4]0.01020.1670.843
Ours0.00700.7980.880
", + "bbox": [ + 106, + 664, + 465, + 741 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tokenization effectiveness: The previous trained models and ours differ in dataset composition, point cloud conditioning, and training settings. To accurately evaluate the effectiveness of different tokenization methods, we conduct a controlled experiment on the subset of our filtered dataset of 24k samples with $\\leq 500$ faces. All factors except for the tokenizers are kept the same: 1). 16-layer Transformers with 768 hidden dimension and positional encoding. 2). Point", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cloud condition of 2048 tokens. 3). Training until 40k steps with an effective batch size of 128 and data augmentation.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on a test set of 200 samples, shown in Table 2, strongly indicate our method provides highly effective inductive bias. We use PivotMesh's [33] pretrained VQ-VAE since MeshAnything does not release their VQ-VAE encoder and noise resistant decoder fine-tuning code.", + "bbox": [ + 511, + 121, + 903, + 196 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4858961ad0b1114687bd4d3327ed3dc6ca9bbe15cbd35867ee8c99db28a47950.jpg", + "table_caption": [ + "Table 1. Quantitative results on Objverse evaluation set." + ], + "table_footnote": [], + "table_body": "
TokenizerCD↓NC↑|NC| ↑Sequence Length↓
Naive [2]0.03760.6390.8229Nf
VQ-VAE [33]0.03520.6730.8156Nf
AMT [4]0.0327-0.0690.768±4Nf
Ours0.01000.7340.8742Nf + 2Nc
", + "bbox": [ + 522, + 209, + 893, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Results on GSO Dataset", + "text_level": 1, + "bbox": [ + 511, + 373, + 733, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further conduct a quantitative evaluation on GSO dataset [8], a dataset of real-world 3D scans to test the generalization capabilities of each model. In this experiment, we observe that the baseline models and ours are sensitive to the input point cloud, with results varying large on different sampling seeds. To mitigate this variability, we generate five samples for each model and select the mesh with the lowest Chamfer Distance for evaluation.", + "bbox": [ + 511, + 395, + 905, + 515 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For our model, we sample point clouds directly from the meshes provided in the GSO dataset. However, direct sampling from these high-resolution meshes often results in reconstructions with extremely small faces. To address this, we decimate the meshes to five target faces ranging from 1000 to 2500 faces and then perform uniform sampling on these decimated versions. For MeshAnything and MeshAnythingV2, we sample the input point clouds from Marching Cubes meshes derived from both the original high-resolution and decimated (1500 faces) meshes using an 8-level octree.", + "bbox": [ + 511, + 516, + 905, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 3, our method achieves the best results across the evaluated metrics. Figure 4 presents qualitative comparisons, demonstrating that our model generates meshes more consistent with the input than the baselines. Additionally, Figure 5 compares our output with the decimated mesh, highlighting that our model can generate meshes with the topology of those created by 3D artists.", + "bbox": [ + 511, + 681, + 905, + 787 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b7578df61dbabe98a46909775a973055e9ca5998ca445b1448a09d716f8ac7d6.jpg", + "table_caption": [ + "Table 2. Quantitative results on our controlled experiment. $N_{f}$ and $N_{c}$ are the number of triangular faces and connected components, respectively." + ], + "table_footnote": [], + "table_body": "
ModelCD↓NC↑|NC| ↑
MeshAnything [3]0.01050.4530.869
MeshAnythingV2 [4]0.01160.32690.865
Ours0.00770.8420.897
", + "bbox": [ + 527, + 800, + 887, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Quantitative results on GSO dataset.", + "bbox": [ + 565, + 881, + 851, + 893 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26613", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4c7008871030e9587f1562e7c753c1859aebfa8dbd2440464e906d44f47eaa35.jpg", + "image_caption": [ + "GT Mesh", + "Face $= 4130$" + ], + "image_footnote": [], + "bbox": [ + 138, + 106, + 236, + 227 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e000042daab9dfbfd71f38951f0abccf0c07a0e37e17dbf182bdbbfb46be8809.jpg", + "image_caption": [ + "Face $= 4316$" + ], + "image_footnote": [], + "bbox": [ + 143, + 250, + 228, + 324 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e871f42414d9870a8d4643b49cde75d1c86a7a693b1a5d1383f8758d8190267a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 142, + 344, + 230, + 417 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b619cebc9bdbf251b4009c2c80c6ce016cd4b166184ff6b81bbb98a29d740e75.jpg", + "image_caption": [ + "Face $= 3718$", + "Face $= 3694$" + ], + "image_footnote": [], + "bbox": [ + 142, + 438, + 225, + 513 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/19a46b7910bd4e7cd66425e402881176a6b420c3c69f2eb2cebefa93cad99d2d.jpg", + "image_caption": [ + "Face $= 2442$" + ], + "image_footnote": [], + "bbox": [ + 148, + 537, + 228, + 599 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/11e2efe6de1025af90451e21a33fbd2fca722fc132dea5ecdd2ac48820658026.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 106, + 638, + 267, + 680 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cdbe6b9459c6b14f95837f49bb91bd5b2dbeb5617e853f1a140716b8ce038ff7.jpg", + "image_caption": [ + "Face $= 3638$", + "Face $= 740$", + "Figure 3. Qualitative comparison on Objverse dataset [7]. Our model is able to generate meshes with higher face counts and refined details compared to the baselines. Results from the baselines use point clouds sampled from marching cube meshes with 8-level octree." + ], + "image_footnote": [], + "bbox": [ + 127, + 729, + 250, + 803 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fd9f595fc5252f1f61aa19060e85596a1c4bbee35c5e7a41f3b8a987af0a24ac.jpg", + "image_caption": [ + "MeshAnything[3]", + "Face $= 439$" + ], + "image_footnote": [], + "bbox": [ + 279, + 107, + 375, + 227 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a276835df9ba95c2dabe8c33274b978bc9b1b240c0f34d037779ea43b4dfd867.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 107, + 478, + 227 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9fcf3347e5801d9571c6426c6abea3832ae9baefae866f4f48e7381343562c09.jpg", + "image_caption": [ + "Face $= 800$" + ], + "image_footnote": [], + "bbox": [ + 336, + 257, + 369, + 324 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/408052b671515aab1508f5b14d64a2c1813071d41d86b54ea99ef646359b075c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 387, + 257, + 468, + 324 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4a27f5d0996d0cb5a021242557e6974b7a87a06684499f3abb5b2ad0546e5c15.jpg", + "image_caption": [ + "Face $= 630$" + ], + "image_footnote": [], + "bbox": [ + 281, + 344, + 370, + 417 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3921a78ceb787db33deb87fb9617e51dd7d15b1b72551e928c72e8b1da0e8698.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 387, + 344, + 478, + 417 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/78ce9f926c0d6d3ff5c88067d089687b45750a8084a5788a3e31e0e86656566a.jpg", + "image_caption": [ + "Face $= 576$" + ], + "image_footnote": [], + "bbox": [ + 287, + 443, + 367, + 513 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c6a05f62242284713fe27ac42aa8e5dd8ac18acd2f89b1478c2c0cf3e2965b86.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 444, + 457, + 513 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c77c5f3a535757ccbfacec013b057d3aeefa7cdfeaffb4ce5fc524afb95b409f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 289, + 537, + 369, + 602 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/62515ae8f680c2b2fa841d7fc0dc922a4c90dabd6f9a1b763bc3ca746a523117.jpg", + "image_caption": [ + "Face $= 495$" + ], + "image_footnote": [], + "bbox": [ + 388, + 539, + 468, + 604 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/46c8e94e65892df848358516405217ac41d43bcfca18b52ce8eceea49a773746.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 299, + 642, + 457, + 680 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c7eb6e95203a5c3ed03255e83c3f563fd433748f3b0ae434ea88b8447f8e3301.jpg", + "image_caption": [ + "Face $= 584$", + "Face $= 235$" + ], + "image_footnote": [], + "bbox": [ + 320, + 729, + 441, + 803 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/adbfea0641ad810a1bdbd262f22ec2ad9fb77bea93f3e6a87b52db614d9bba50.jpg", + "image_caption": [ + "MeshAnythingV2 [4]" + ], + "image_footnote": [], + "bbox": [ + 488, + 107, + 687, + 227 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9d45079fca32265b44c19920c3d1fd9bf9419954282437490a6c664cc563ec05.jpg", + "image_caption": [ + "Face = 2247", + "Face = 2246" + ], + "image_footnote": [], + "bbox": [ + 491, + 250, + 678, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2bee1f01fd798c18bd50ec3a0cd1a854dd2d7c5e023e0425438350bf885dadca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 490, + 343, + 684, + 417 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/11c60089c50c4d1a57340382ee4ed6dcd12bcc085ea646d9d7ea60fe961134a9.jpg", + "image_caption": [ + "Face = 2090" + ], + "image_footnote": [], + "bbox": [ + 498, + 438, + 665, + 491 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4ade773c361062907b496d276215643f34764126d0270e2d3c088b6f92518927.jpg", + "image_caption": [ + "Face = 1998", + "Face = 2159" + ], + "image_footnote": [], + "bbox": [ + 491, + 537, + 687, + 609 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0896b0ac5011a8545b038d324d2b5f216f743fac963677c74cb704788e4c1bd7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 640, + 666, + 680 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d6cf071ba12b87287e08c471bba3adec9d42bcfec457a6d53b6a1d952365da49.jpg", + "image_caption": [ + "Face $= 1769$", + "Face $= 265$" + ], + "image_footnote": [], + "bbox": [ + 527, + 729, + 648, + 803 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/edce66028a2128e05a9952f5e8a0f0d42d04f1141d9d8826fb6fad70d91ee1ef.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 696, + 106, + 893, + 227 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/db2cf1a3cc461253e03ad5d69d010bf7187b96345e953a333c045db3f77a5267.jpg", + "image_caption": [ + "Face = 3311" + ], + "image_footnote": [], + "bbox": [ + 700, + 250, + 883, + 324 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1613048410d1a69a5fd9ff7f126338f8b42a3782338aef4fac345cf1293a56e3.jpg", + "image_caption": [ + "Face = 4545", + "Face $= 5122$" + ], + "image_footnote": [], + "bbox": [ + 697, + 343, + 787, + 417 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aa3bcc8bc99211346acf57e6ba1a48d8de0e21a7a23ec8b10a68cc89127c6e2b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 803, + 345, + 893, + 417 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/03154838e179a348065aadc23a79dc505b0253f6eb04521321b28730fb0dec22.jpg", + "image_caption": [ + "Face $= 2708$" + ], + "image_footnote": [], + "bbox": [ + 700, + 440, + 782, + 513 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8b6bb92ae561ded526374b44c20e523b4e594b887f3840aafbd810c930494c19.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 810, + 440, + 877, + 513 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4c3f35e035726190e280c1355f4cd14ead46d6f454a1454e79d88ad18890d727.jpg", + "image_caption": [ + "Face = 2646" + ], + "image_footnote": [], + "bbox": [ + 704, + 537, + 785, + 608 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/17d681719d8eddfdde4af39a245b17fdbf0e00074e5db8a7fc2eb1aedbdb286f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 805, + 539, + 883, + 604 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/debb74a80c1c187d59811371cdfcf0ad3fffad7cfbbf46aafa008ff447466f70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 640, + 874, + 679 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/53de8261f2468831ccacc32c9c31da56c31d904e5596dd946b3a643bd78d75d3.jpg", + "image_caption": [ + "Face = 2114", + "Face $= 301$" + ], + "image_footnote": [], + "bbox": [ + 733, + 729, + 856, + 803 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26614", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7ef7ecfda59a2d78b1dc6fda89929c0f541a3fade4774c3cdd0a9878cad79061.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 133, + 78, + 236, + 276 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/759d5d3706e288a6bc29e93e0b143ebfe0b554b0ee5ab77cec3689744394d944.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison on GSO dataset [8]." + ], + "image_footnote": [], + "bbox": [ + 245, + 78, + 431, + 276 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/402ebf7e1680da38c2dfc0cc1b3eec4c54b07d78efe7e0204a76f6e394c545d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 446, + 78, + 640, + 276 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/89fec65e06197b931633e8e2984eb8d542e02e2d8eb467d736d5e994e9522daa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 78, + 846, + 276 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2503cf9c79ce30a9a12cade462fedff6a975b3748671614482bd2347b8411b57.jpg", + "image_caption": [ + "Figure 5. Comparison between the decimated mesh and our output. Our model is capable of generating meshes with the topology of those created by 3D artists." + ], + "image_footnote": [], + "bbox": [ + 122, + 305, + 225, + 401 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d5265daa069de701561f839cff25793d42131ccf1e7618922ac0cfe87524614f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 235, + 305, + 339, + 401 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f7c2c6610f54db4a3924337477e542097778c882300ef782084d009517360b21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 348, + 305, + 450, + 401 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 460, + 243, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Decoder MLP head: We compare TreeMeshGPT trained with MLP heads that predict the x-y-z coordinates simultaneously and our proposed hierarchical MLP. Predicting the coordinate simultaneously leads to challenging sampling that results in noisy and more incomplete meshes, as shown in Figure 6. Running evaluation on the 200 validation samples used in Table 1 with this head yields $\\mathrm{CD} = 0.0114$ , $\\mathrm{NC} = 0.724$ , $|\\mathrm{NC}| = 0.847$ .", + "bbox": [ + 89, + 486, + 483, + 607 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/709731a44bdd84b5af04b253a655d6faff6741ac1bd7b41d8df22e54918df39d.jpg", + "image_caption": [ + "Figure 6. MLP head ablation. Our hierarchical MLP maintains the sequential nature of the x-y-z coordinates prediction that results in easier sampling compared to simultaenous prediction." + ], + "image_footnote": [], + "bbox": [ + 124, + 614, + 220, + 712 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8f8af4a4df6b9aae4cbca26166892ccfcde95dddbaac58acedce918f58fb5bd7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 614, + 333, + 710 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4b7f5da83b96abae0953f1860cdcf2a984ebcb295915c37dc0da997ea1becc64.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 614, + 444, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tree traversal: We conduct an ablation study on smaller Transformer architecture comparing DFS and breadth-first-search (BFS) traversals in forming input-output sequences for our Autoregressive Tree Sequencing. As shown in the training perplexity plot in Figure 7, DFS traversal enables more efficient Transformer training. This improvement likely stems from the stronger local dependencies introduced by DFS, where each step is more predictable based on its immediate predecessors. In contrast, BFS traversal often", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "introduces dependencies between steps that are spatially or structurally distant, and thus complicating the learning process.", + "bbox": [ + 511, + 306, + 906, + 351 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/015486ac48dca7b28335bdfe1e67404acfa897eda850796d899fde074531d4e0.jpg", + "image_caption": [ + "Figure 7. Training perplexity comparison Between BFS and DFS traversals. DFS traversal shows a better training perplexity compared to BFS. Shown in the plot is the perplexity for $y$ -axis vertex coordinate." + ], + "image_footnote": [], + "bbox": [ + 517, + 354, + 898, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 628, + 633, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present TreeMeshGPT, an autoregressive Transformer designed to generate high-quality artistic meshes aligned with input point clouds. TreeMeshGPT incorporates a novel Autoregressive Tree Sequencing technique instead of the conventional next-token prediction. Our Autoregressive Tree Sequencing represents each face with two tokens, enabling a 7-bit discretization model that can generate up to 5,500 triangular faces with 2,048 point cloud latent tokens. Experiments show that TreeMeshGPT can generate meshes with higher quality compared to the previous methods.", + "bbox": [ + 511, + 655, + 906, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations: Our model has a similar failure mode to the previous methods that the success rate decreases as the sequence length increases. Additionally, while our model has an improved capacity to generate meshes with higher face counts, challenges persist in enforcing an optimal mesh topology.", + "bbox": [ + 511, + 809, + 908, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26615", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This research is supported by the National Research Foundation (NRF) Singapore, under its NRF-Investigatorship Programme (Award ID. NRF-NRFI09-0008).", + "bbox": [ + 89, + 90, + 485, + 151 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 164, + 187, + 179 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Fausto Bernardini, Joshua Mittleman, Holly Rushmeier, Cláudio Silva, and Gabriel Taubin. The ball-pivoting algorithm for surface reconstruction. IEEE Transactions on Visualization and Computer Graphics, 1999. 2", + "[2] Sijin Chen, Xin Chen, Anqi Pang, Xianfang Zeng, Wei Cheng, Yijun Fu, Fukun Yin, Yanru Wang, Zhibin Wang, Chi Zhang, et al. Meshxl: Neural coordinate field for generative 3d foundation models. Advances in Neural Information Processing Systems (NeurIPS), 2024. 3, 6", + "[3] Yiwen Chen, Tong He, Di Huang, Weicai Ye, Sijin Chen, Ji-axiang Tang, Xin Chen, Zhongang Cai, Lei Yang, Gang Yu, et al. Meshanything: Artist-created mesh generation with autoregressive transformers. arXiv preprint arXiv:2406.10163, 2024. 2, 3, 5, 6, 7, 8", + "[4] Yiwen Chen, Yikai Wang, Yihao Luo, Zhengyi Wang, Zilong Chen, Jun Zhu, Chi Zhang, and Guosheng Lin. Meshanything v2: Artist-created mesh generation with adjacent mesh tokenization. arXiv preprint arXiv:2408.02555, 2024. 2, 3, 4, 5, 6, 7, 8", + "[5] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[6] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\\mathrm{m} + 3\\mathrm{d}$ objects. Advances in Neural Information Processing Systems (NeurIPS), 2023. 2", + "[7] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objaverse: A universe of annotated 3d objects. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 5, 7", + "[8] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2022. 6, 8", + "[9] Zekun Hao, David W Romero, Tsung-Yi Lin, and Ming-Yu Liu. Meshtron: High-fidelity, artist-like 3d mesh generation at scale. arXiv preprint arXiv:2412.09548, 2024. 3", + "[10] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. In International Conference on Learning Representations (ICLR), 2024. 2", + "[11] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object gen" + ], + "bbox": [ + 93, + 188, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "eration with dream fields. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[12] Tao Ju, Frank Losasso, Scott Schaefer, and Joe Warren. Dual contouring of hermite data. In Proceedings of the 29th Annual Conference on Computer graphics and Interactive Techniques, 2002. 2", + "[13] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the 4th Eurographics Symposium on Geometry Processing, 2006. 2", + "[14] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 5", + "[15] Kiriakos N Kutulakos and Steven M Seitz. A theory of shape by space carving. International Journal of Computer Vision, 2000. 2", + "[16] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[17] Chenghao Li, Chaoning Zhang, Joseph Cho, Atish Waghwase, Lik-Hang Lee, Francois Rameau, Yang Yang, SungHo Bae, and Choong Seon Hong. Generative ai meets 3d: A survey on text-to-3d in aigc era. arXiv preprint arXiv:2305.06131, 2023. 1", + "[18] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. In International Conference on Learning Representations (ICLR), 2024. 2", + "[19] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM SIGGRAPH Computer Graphics, 1987. 1, 2", + "[20] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. 5", + "[21] Oscar Michel, Roi Bar-On, Richard Liu, Sage Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[22] Charlie Nash, Yaroslav Ganin, SM Ali Eslami, and Peter Battaglia. *Polygen: An autoregressive generative model of 3d meshes*. In International Conference on Machine Learning (ICML), 2020. 3", + "[23] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 3", + "[24] Gregory M Nielson. Dual marching cubes. In IEEE Visualization, 2004. 2", + "[25] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In International Conference on Learning Representations (ICLR), 2023. 2", + "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26616", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. In International Conference on Learning Representations (ICLR), 2024. 2", + "[28] Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3", + "[29] Jiaxiang Tang, Zhaoshuo Li, Zekun Hao, Xian Liu, Gang Zeng, Ming-Yu Liu, and Qinsheng Zhang. Edgerunner: Auto-regressive auto-encoder for artistic mesh generation. arXiv preprint arXiv:2409.18114, 2024. 2, 3, 4, 5", + "[30] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in Neural Information Processing Systems (NeurIPS), 2017. 3", + "[31] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems (NeurIPS), 2017. 5", + "[32] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[33] Haohan Weng, Yikai Wang, Tong Zhang, CL Chen, and Jun Zhu. Pivotmesh: Generic 3d mesh generation via pivot vertices guidance. arXiv preprint arXiv:2405.16890, 2024. 3, 6", + "[34] Haohan Weng, Zibo Zhao, Biwen Lei, Xianghui Yang, Jian Liu, Zeqiang Lai, Zhuo Chen, Yuhong Liu, Jie Jiang, Chunchao Guo, et al. Scaling mesh generation via compressive tokenization. arXiv preprint arXiv:2411.07025, 2024. 3", + "[35] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, 2022. 1", + "[36] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 2", + "[37] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Transactions on Graphics (TOG), 2023. 4, 5", + "[38] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 2024. 3, 5" + ], + "bbox": [ + 91, + 92, + 480, + 852 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26617", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/TreeMeshGPT_ Artistic Mesh Generation with Autoregressive Tree Sequencing/cc43c3b2-b6bf-472a-9dd6-c74d4f03cee9_model.json b/2025/TreeMeshGPT_ Artistic Mesh Generation with Autoregressive Tree Sequencing/cc43c3b2-b6bf-472a-9dd6-c74d4f03cee9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f274e0bd47b8387680072393b55690a4b95a232d --- /dev/null +++ b/2025/TreeMeshGPT_ Artistic Mesh Generation with Autoregressive Tree Sequencing/cc43c3b2-b6bf-472a-9dd6-c74d4f03cee9_model.json @@ -0,0 +1,2772 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.131, + 0.898, + 0.155 + ], + "angle": 0, + "content": "TreeMeshGPT: Artistic Mesh Generation with Autoregressive Tree Sequencing" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.18, + 0.404, + 0.198 + ], + "angle": 0, + "content": "Stefan Lionar\\(^{1,2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.44, + 0.18, + 0.57, + 0.199 + ], + "angle": 0, + "content": "Jiabin Liang\\(^{1,2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.611, + 0.18, + 0.728, + 0.198 + ], + "angle": 0, + "content": "Gim Hee Lee3" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.209, + 0.319, + 0.225 + ], + "angle": 0, + "content": "1Sea AI Lab" + }, + { + "type": "text", + "bbox": [ + 0.377, + 0.209, + 0.446, + 0.225 + ], + "angle": 0, + "content": "\\(^{2}\\)Garena" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.209, + 0.776, + 0.228 + ], + "angle": 0, + "content": "\\(^{3}\\)National University of Singapore" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.244, + 0.907, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.511, + 0.908, + 0.556 + ], + "angle": 0, + "content": "Figure 1. Artistic meshes generated by TreeMeshGPT. Our method offers a novel sequencing approach for artistic mesh generation using autoregressive Transformer decoder by retrieving the next token from a dynamically growing tree structure. In our experiment with 7-bit discretization, TreeMeshGPT supports meshes with up to 5,500 triangular faces under strong point cloud conditioning.\\(^{12}\\)" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.566, + 0.327, + 0.582 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.598, + 0.486, + 0.841 + ], + "angle": 0, + "content": "We introduce TreeMeshGPT, an autoregressive Transformer designed to generate high-quality artistic meshes aligned with input point clouds. Instead of the conventional next-token prediction in autoregressive Transformer, we propose a novel Autoregressive Tree Sequencing where the next input token is retrieved from a dynamically growing tree structure that is built upon the triangle adjacency of faces within the mesh. Our sequencing enables the mesh to extend locally from the last generated triangular face at each step, and therefore reduces training difficulty and improves mesh quality. Our approach represents each triangular face with two tokens, achieving a compression rate of approximately \\(22\\%\\) compared to the naive face tokenization. This efficient tokenization enables our model to generate highly detailed artistic meshes with strong point cloud conditioning, surpassing previous methods in both capacity and fi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.659 + ], + "angle": 0, + "content": "delyity. Furthermore, our method generates mesh with strong normal orientation constraints, minimizing flipped normals commonly encountered in previous methods. Our experiments show that TreeMeshGPT enhances the mesh generation quality with refined details and normal orientation consistency." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.692, + 0.645, + 0.708 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.72, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In recent advancements in 3D generation, representations such as voxels, point clouds, and implicit functions are often utilized [17, 35]. After the generation process, these representations are converted into meshes using techniques like Marching Cubes [19], which result in dense, overtessellated triangular meshes. However, these dense meshes are unsuitable for applications that require real-time rendering, such as gaming and virtual reality. Although many mesh down-sampling algorithms can reduce the number of triangles, they also degrade mesh quality and produce messy, unstructured wireframes. In contrast, skilled artists can create highly compact meshes with minimal triangles" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.851, + 0.483, + 0.875 + ], + "angle": 0, + "content": "\\( {}^{1} \\) TreeMeshGPT supports a higher face count with finer discretization as it preserves the required manifold connectivity condition." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.482, + 0.9 + ], + "angle": 0, + "content": "2Our code is available at: https://github.com/sail-sg/TreeMeshGPT." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.851, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26608" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "while preserving the object's sharp details. Additionally, the wireframes created by artists are more regular, aesthetically pleasing, and better aligned with the object's feature or semantic boundaries, which facilitates further human interactions, such as editing and animation. However, the process of manually creating artist-quality meshes is highly time-consuming and labor-intensive." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.2, + 0.482, + 0.367 + ], + "angle": 0, + "content": "This laborous process highlights the need for automated methods that can replicate the quality of artist-created meshes without requiring extensive manual effort. MeshAnything [3] seeks to bridge the gap between advancements in 3D generation and artist-quality mesh creation by adding point cloud condition to the artistic mesh generation Transformer initially proposed by MeshGPT [28]. Point clouds are chosen as the condition because they are either the direct output or can be obtained conveniently from the generated Marching Cubes meshes of the advanced 3D generation techniques." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.37, + 0.483, + 0.612 + ], + "angle": 0, + "content": "MeshAnything [3] represents each triangular face with 9 latent tokens, leading to long sequences and limiting artistic mesh generation to 800 faces due to the Transformer's quadratic complexity. This constraint poses challenges for real-world applications, which often require meshes with significantly higher face counts to accurately represent complex objects and environments. In the subsequent works, MeshAnything V2 [4] and EdgeRunner [29] leverage triangle adjacency to create shorter sequences to represent the same meshes. Consequently, they are able to generate meshes with up to 1,600 and 4,000 faces, respectively. However, many real-world applications demand meshes with higher face count to accurately represent detailed surface topology. Additionally, challenges remain in generating high-quality meshes free from artifacts such as gaps, missing components, and flipped normals." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To further improve tokenization efficiency and mesh quality, we introduce TreeMeshGPT. Unlike previous methods that rely on conventional next-token prediction in autoregressive Transformers, TreeMeshGPT introduces a novel Autoregressive Tree Sequencing approach. Instead of sequentially predicting next tokens, our method retrieves the next token from a dynamically growing tree structure built upon triangle adjacency within the mesh. This strategy allows the mesh to expand locally from the last generated triangular face at each step, and thus reducing training difficulty and enhancing mesh quality. Our approach represents each triangular face with two tokens, achieving a compression rate of \\(22\\%\\) compared to the naive face tokenization of 9 tokens per face. This efficient tokenization technique pushes the boundary of artistic mesh generation. With 7-bit discretization, it enables the generation of meshes with up to 5,500 triangles under a strong point cloud condition of 2,048 tokens. Furthermore, our method generates meshes with strong normal orientation constraints," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "minimizing flipped normals commonly encountered in MeshAnything [3] and MeshAnythingV2 [4]." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.123, + 0.835, + 0.137 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.141, + 0.905, + 0.185 + ], + "angle": 0, + "content": "- We propose a novel Autoregressive Tree Sequencing technique that efficiently represents two tokens per triangular face." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.186, + 0.905, + 0.245 + ], + "angle": 0, + "content": "- Our proposed tokenization enables the training of a 7-bit discretization artistic mesh generative model with strong point cloud condition, capable of generating high-quality meshes with up to 5,500 faces." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.247, + 0.905, + 0.291 + ], + "angle": 0, + "content": "- Extensive experiments show that our model can generate higher quality meshes and can generalize to real-world 3D scans." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.141, + 0.905, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.307, + 0.655, + 0.323 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.332, + 0.679, + 0.346 + ], + "angle": 0, + "content": "2.1. Mesh Extraction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.354, + 0.908, + 0.672 + ], + "angle": 0, + "content": "Constructing a mesh from other 3D representations has been a research focus for decades. Among many successful methods [1, 15, 19], Marching Cubes [19] is the most widely used. It divides a scalar field into cubes and extracts triangles to approximate the isosurface. It is simple yet robust, producing watertight and 2-manifold results. Many improvements such as Dual Contouring [12] and Dual Marching Cubes [24] have been developed to enhance its capabilities. Another well-known method is Poisson reconstruction [13], which uses point clouds and normals as boundary conditions to solve a scalar field defined in 3D space, then applies Marching Cubes to extract the mesh. However, these approaches often focus on representing shapes with dense, over-tessellated meshes, resulting in messy, unstructured wireframes. This makes them unsuitable for downstream workflows that require efficient and structured meshes, such as real-time rendering, animation rigging, and editing. The dense, over-tessellated meshes not only increase computational load but also lack the regularity and semantic alignment necessary for the downstream processes." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.682, + 0.664, + 0.697 + ], + "angle": 0, + "content": "2.2.3D Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.81 + ], + "angle": 0, + "content": "After the great success of 2D image generation [26], 3D generation has become a promising research direction. This field focuses on generating 3D assets for industries such as gaming, film, and AR/VR. Due to the limited availability of 3D data, early methods [11, 21, 25, 27] relied on optimizing underlying representations to mimic conditioning from 2D images or multi-view 2D images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.902 + ], + "angle": 0, + "content": "With the introduction of large-scale datasets [6, 7], feedforward 3D generation techniques, such as those in [10, 18, 36], have become feasible. These techniques significantly improve generation speed compared to optimization-based methods. However, the resulting meshes often suffer from lower quality and lack diversity." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26609" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.213 + ], + "angle": 0, + "content": "Inspired by the success of 2D diffusion models in image generation, many researchers have attempted to apply diffusion techniques directly to 3D data [5, 23, 32, 38]. For instance, CLAY [38], a transformer-based 3D latent diffusion model, achieves state-of-the-art results in high-quality shape generation. Despite these advances, these generation methods often require post-conversion for downstream applications, which remains a non-trivial challenge." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.224, + 0.379, + 0.24 + ], + "angle": 0, + "content": "2.3. Autoregressive Mesh Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.247, + 0.483, + 0.382 + ], + "angle": 0, + "content": "To address these limitations, recent approaches have leveraged autoregressive models for direct mesh generation [2, 22, 28, 33]. MeshGPT [28] was the first to tokenize a mesh through face sorting and compress it using a VQ-VAE [16, 30], followed by an autoregressive transformer to predict the compressed token sequence. This method enables the generation of meshes with direct supervision from artist-created topology information, which is often absent in previous approaches." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.384, + 0.483, + 0.685 + ], + "angle": 0, + "content": "Subsequent works [2, 3, 33] explored more efficient representations and incorporated input conditioning, such as point clouds and images. However, these methods are limited to generating meshes with fewer than 800 faces due to the long sequence lengths and the quadratic computational cost of transformers. MeshAnythingV2 [4] and EdgeRunner [29] introduced more compact mesh tokenization techniques that leverage triangle adjacency, increasing the maximum face count to 1,600 and 4,000, respectively. Meshtron [9] proposes hourglass architecture and sliding window inference to scale up face count capacity while using the naive tokenization [2]. In a concurrent work, BPT [34] proposes a compact tokenization using block-wise indexing and localized patch aggregation. These methods rely on the next-token prediction commonly used in large language models (LLMs) and autoregressive image generation. In contrast, we offer a novel sequencing strategy based on a dynamically growing tree structure, aiming to increase the maximum face count and improve the overall quality of the generated meshes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.702, + 0.182, + 0.717 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.373, + 0.744 + ], + "angle": 0, + "content": "3.1. Autoregressive Tree Sequencing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Tokenization plays a crucial role in autoregressive models, particularly in complex tasks like 3D mesh generation, where both the quality and efficiency of tokenization significantly affect model performance and scalability. In the context of mesh generation, tokenization involves encoding vertices, edges, or faces into sequential tokens that the model can process step-by-step. Drawing insights from LLMs, previous autoregressive mesh generation methods follow next-token prediction strategy that explicitly uses each output as the input for the subsequent step." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.319 + ], + "angle": 0, + "content": "Our approach differs from those methods by using a tree-based traversal scheme to grow the mesh during the generation process. Specifically, in TreeMeshGPT, the input to the Transformer decoder are directed mesh edges, represented as \\(\\mathbf{I} = \\{(v_1^n,v_2^n)\\}_{n = 1}^N\\in \\mathbb{R}^{N\\times 6}\\), where each \\((v_{1}^{n},v_{2}^{n})\\) denotes a pair of dequantized vertices for the generation at step \\(n\\). At each step, the Transformer decoder makes a localized prediction to either add a new vertex \\(v_{3}^{n}\\) to expand the mesh by connecting to the initial pair of vertices \\((v_{1}^{n},v_{2}^{n})\\) or predict [STOP] label indicating that no further expansion should occur from the input edge. TreeMeshGPT leverages a tree traversal process to construct the sequential input-output pairs for the autoregressive generation, described as follows (Note: the sequence order description below omits auxiliary tokens for simplicity)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.322, + 0.905, + 0.382 + ], + "angle": 0, + "content": "Sequence order: We utilize a half-edge data structure with depth-first-search (DFS) traversal to construct the sequential input-output pairs, \\((\\mathbf{I},\\mathbf{O}) = \\{(I_n,o_n)\\}_{n = 1}^N\\), accompanied by a dynamic stack \\(\\mathbf{S}\\) to manage the traversal process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.907, + 0.548 + ], + "angle": 0, + "content": "The traversal process starts from a directed edge in a mesh, in which we determine if this edge has an opposite vertex that forms a triangle, is a boundary, or if the triangle has already been visited. When a new triangle is formed, the output \\( o_n \\) is defined as the opposite vertex \\( v_3^n \\). Two new edges are created by connecting \\( v_3^n \\) to the initial edge's vertices \\( I_n = (v_1^n, v_2^n) \\), resulting in edges \\( (v_3^n, v_2^n) \\) and \\( (v_1^n, v_3^n) \\). These edges are directed in a counter-clockwise orientation on the potential next adjacent faces, as enforced by the half-edge data structure. The newly created edges are then pushed onto the stack S for continued traversal:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.556, + 0.862, + 0.574 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} := \\mathbf {S} \\odot \\left(v _ {1} ^ {n}, v _ {3} ^ {n}\\right) \\quad \\text {a n d} \\quad \\mathbf {S} := \\mathbf {S} \\odot \\left(v _ {3} ^ {n}, v _ {2} ^ {n}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.581, + 0.905, + 0.656 + ], + "angle": 0, + "content": "where \\(\\odot\\) represents the operation of pushing the edge \\((v_{i}, v_{j})\\) onto the top of stack \\(\\mathbf{S}\\). Conversely, if the input edge is a boundary or if adding a new vertex forms a previously visited triangle, the output \\(o_{n}\\) is set to the [STOP] label. In this case, no new vertex or edge is added." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.657, + 0.905, + 0.687 + ], + "angle": 0, + "content": "The input for the next step, \\( n + 1 \\), is obtained by popping the top edge from the stack:" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.694, + 0.882, + 0.712 + ], + "angle": 0, + "content": "\\[\nI _ {n + 1} = \\left(v _ {1} ^ {n + 1}, v _ {2} ^ {n + 1}\\right) := \\operatorname {t o p} (\\mathbf {S}), \\quad \\mathbf {S} := \\mathbf {S} \\setminus \\operatorname {t o p} (\\mathbf {S})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.905, + 0.81 + ], + "angle": 0, + "content": "where \\(\\operatorname{top}(\\mathbf{S})\\) retrieves the current top edge of the stack \\(\\mathbf{S}\\), and \\(\\mathbf{S} := \\mathbf{S} \\setminus \\operatorname{top}(\\mathbf{S})\\) updates \\(\\mathbf{S}\\) by removing this top edge. We initialize the stack with a directed edge at the lowest position of the mesh and its twin. The traversal then proceeds until all triangles are visited. A simple illustration of this sequencing process is provided in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Note that a mesh may consist of multiple connected components. A component begins from an initial edge, expands as edges are added to a stack, and is considered fully traversed once the stack is empty. When there are multiple components in a mesh, the traversal of the first component starts from the edge with the lowest position in the mesh" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26610" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.089, + 0.331, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.089, + 0.581, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.089, + 0.905, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.278, + 0.907, + 0.418 + ], + "angle": 0, + "content": "Figure 2. Illustration of the sequence order in our Autoregressive Tree sequencing. a). A small subset of a triangular mesh. [STOP] indicates boundary. b). An equivalent tree representation of the mesh. In this tree, each node is represented as a directed edge from a pair of vertices. The root is initialized with two child nodes: \\((v_0, v_1)\\) and its twin \\((v_1, v_0)\\). A DFS traversal then proceeds to create the input-output sequence. c). Dynamic stack from the DFS traversal. The stack is initialized with \\((v_0, v_1)\\) and its twin \\((v_1, v_0)\\). The input \\(I_n\\) is always obtained from the top of the stack. Thus, \\(I_1 = (v_0, v_1)\\) at step 1. The opposite vertex of \\(I_1\\) is \\(v_2\\) and consequently, \\(o_1\\) is set to \\(v_2\\). Two new edges are then formed by connecting the opposite vertex to the initial pair of vertices: \\((v_2, v_1)\\) and \\((v_0, v_2)\\). The direction is enforced to be counter-clockwise on the potential next adjacent faces. At step 2, \\(I_2 = (v_2, v_1)\\). Since \\(I_2\\) is a boundary, \\(o_2\\) is set to [STOP] label and no new edge is added to the stack. Step 3 and onwards follow the same traversal process. d). Transformer decoder sequence. The sequence in the Transformer decoder follows the input-output pairs from the tree traversal. The auxiliary tokens to initialize the generation of a connected component and the [EOS] are also added to the input-output sequence." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.437, + 0.483, + 0.482 + ], + "angle": 0, + "content": "and continues until no edge remains in the stack. Each subsequent component begins from the next available edge positioned at the lowest among the remaining unvisited faces." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.487, + 0.483, + 0.682 + ], + "angle": 0, + "content": "Generation process: To initialize the generation of a mesh component, an auxiliary token \\([\\mathrm{SOS}]\\in \\mathbb{R}^D\\) is used as the input to predict the first vertex \\(v_{1}\\). For the next step, a second auxiliary token \\([\\mathrm{SOS2}]\\in \\mathbb{R}^{C}\\), concatenated with the embedded representation of \\(v_{1}\\), serves as input to predict the second vertex \\(v_{2}\\). Once these initial vertices are predicted, the mesh generation proceeds through our Autoregressive Tree Sequencing with a stack initialized by \\((v_{1},v_{2})\\) and its twin \\((v_{2},v_{1})\\). When the stack is empty—indicating the current component is complete—a new component is initialized by reintroducing the [SOS] and [SOS2] tokens. After all components have been generated, the sequence is terminated with an [EOS] label." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.685, + 0.483, + 0.761 + ], + "angle": 0, + "content": "The final mesh is constructed by gathering the faces formed from the initial input vertex pairs and their predicted opposite vertices: \\(\\mathcal{M} = \\bigcup_{n=1}^{N}(v_1^n,v_2^n,v_3^n)\\) for the generation steps where \\(I_n \\notin \\{[\\mathrm{SOS}], [\\mathrm{SOS2}]\\}\\) and \\(o_n \\notin \\{[\\mathrm{STOP}], [\\mathrm{EOS}]\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Input embedding: We employ a positional embedding function from [37] to encode each vertex into a high-dimensional space, capturing its positional information across multiple frequency bases. This embedding function \\(\\mathrm{PosEmbed}(.) : \\mathbb{R}^3 \\to \\mathbb{R}^C\\) maps 3D coordinates to a \\(C\\)-dimensional embedding. For each edge, the embeddings of its vertex pair are concatenated, creating a representation in \\(\\mathbb{R}^{2C}\\), which is subsequently passed through an MLP to map it to the Transformer's hidden dimension, \\(\\mathbb{R}^{2C} \\to \\mathbb{R}^D\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.437, + 0.907, + 0.679 + ], + "angle": 0, + "content": "Vertex prediction: In prior works on mesh generation, each vertex is represented as a sequence of three tokens corresponding to its quantized \\( x- \\), \\( y- \\), and \\( z \\)-coordinates. Specifically, to predict a single vertex's position, those models generate each coordinate independently as separate tokens in sequence. This approach leads to longer sequences, as each vertex requires three distinct tokens. In contrast, our method predicts the vertex' quantized \\( x- \\), \\( y- \\), and \\( z \\)-coordinates in one sequence length by using hierarchical MLP heads. This hierarchical approach maintains the sequential nature in predicting the \\( x- \\), \\( y- \\), and \\( z \\)-coordinates. Further details can be found in the supplementary material. As shown in the ablation study (Section 4.5), our hierarchical MLP heads result in easier coordinate sampling compared to prediction heads that predict the \\( x- \\), \\( y- \\), and \\( z \\)-coordinates simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Advantages: Our Autoregressive Tree Sequencing approach adds only two sequence steps per triangular face as each face introduces two new nodes during the tree traversal process. Additionally, since most meshes consist of only a few connected components, our method requires minimal auxiliary tokens. This efficient sequencing achieves a compression rate of approximately \\(22\\%\\) for most meshes compared to naive tokenization, which uses 9 tokens per face. Our compression rate is thus approximately double that of methods like MeshAnythingV2 [4] and EdgeRunner [29]. Additionally, by using a dynamic stack to manage the input sequence, our method allows the Transformer to focus solely on making localized predictions at each step, hence improving training efficiency. Furthermore, our method" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26611" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "generates mesh with strong normal orientation constraints, minimizing flipped normals commonly encountered in MeshAnything [3] and MeshAnythingV2 [4]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.141, + 0.483, + 0.187 + ], + "angle": 0, + "content": "Loss function: We aim to train the Transformer decoder \\(\\theta\\) to maximize the likelihood of generating the sequence of outputs \\(\\{o_n\\}_{n=1}^N\\) given the input sequence \\(\\{I_n\\}_{n=1}^N\\):" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.187, + 0.482, + 0.224 + ], + "angle": 0, + "content": "\\[\n\\prod_ {n = 1} ^ {N} P \\left(o _ {n} \\mid I _ {\\leq n}; \\theta\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.226, + 0.483, + 0.317 + ], + "angle": 0, + "content": "To this end, given the ground truth and predicted values with teacher-forcing across all steps, denoted by \\(\\mathbf{O} = \\{\\mathbf{O}_x,\\mathbf{O}_y,\\mathbf{O}_z\\}\\) and \\(\\hat{\\mathbf{O}} = \\{\\hat{\\mathbf{O}}_x,\\hat{\\mathbf{O}}_y,\\hat{\\mathbf{O}}_z\\}\\), respectively, where each \\(\\mathbf{O}_i\\) represents the discretized vertex coordinate along a specific axis, we use a loss function defined as the sum of cross-entropy losses for each coordinate:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.325, + 0.483, + 0.344 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {C E}} \\left(\\mathbf {O} _ {x}, \\hat {\\mathbf {O}} _ {x}\\right) + \\mathcal {L} _ {\\mathrm {C E}} \\left(\\mathbf {O} _ {y}, \\hat {\\mathbf {O}} _ {y}\\right) + \\mathcal {L} _ {\\mathrm {C E}} \\left(\\mathbf {O} _ {z}, \\hat {\\mathbf {O}} _ {z}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.483, + 0.412 + ], + "angle": 0, + "content": "To incorporate stopping conditions, we add the [STOP] and [EOS] labels to the class selection on the height axis, extending it with two additional classes beyond the discretized coordinate classes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.426, + 0.224, + 0.442 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.451, + 0.187, + 0.465 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.473, + 0.483, + 0.683 + ], + "angle": 0, + "content": "We use Objaverse [7] meshes as our training dataset. To ensure high-quality meshes, we select meshes that meet the half-edge traversal requirement, i.e., they are manifold and have no flipped normal. All meshes are preprocessed by centering and normalizing them within a cube spanning \\([-0.5, 0.5]\\). We apply 7-bit discretization, remove any duplicate triangles, and choose meshes with less than 5,500 faces. Additionally, we perform orthographic projections and exclude meshes where one of the projections has extremely small area or contains more than one cluster. After filtering, we retain a total of 75,000 meshes, of which 1,000 are reserved for validation, with the remainder used for training. To increase data diversity, we apply the following augmentations:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.685, + 0.483, + 0.715 + ], + "angle": 0, + "content": "- Scaling: Each axis is scaled independently by a factor randomly chosen from the range [0.75, 0.95]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.715, + 0.483, + 0.775 + ], + "angle": 0, + "content": "- Rotation: We first apply a \\(90^{\\circ}\\) or \\(-90^{\\circ}\\) rotation along the \\(x\\)- or \\(y\\)-axis with a probability of 0.3. Afterward, we perform a rotation around the \\(z\\)-axis with an angle uniformly sampled from \\([-180^{\\circ}, 180^{\\circ}]\\)." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.685, + 0.483, + 0.775 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.785, + 0.308, + 0.801 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Point cloud conditioning: Following previous approaches [3, 4, 29], we adopt point cloud conditioning to provide practical guidance for the generation process. To achieve this, we sample a point cloud from the input mesh surface and apply a lightweight encoder [3, 4, 37, 38]. Specifically, we sample 8192 points, denoted as \\(\\mathbf{X} \\in\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.091, + 0.905, + 0.122 + ], + "angle": 0, + "content": "\\(\\mathbb{R}^{8192 \\times 3}\\), from the mesh surface. A cross-attention layer is then used to encode these points into a latent code:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.132, + 0.905, + 0.15 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} = \\operatorname {C r o s s A t t} (\\mathbf {Q}, \\operatorname {P o s E m b e d} (\\mathbf {X})) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.157, + 0.905, + 0.264 + ], + "angle": 0, + "content": "where \\(\\mathbf{Q} \\in \\mathbb{R}^{2048 \\times C}\\) represents query embeddings with a shorter sequence length of 2048 and hidden dimension of \\(C\\), \\(\\mathrm{PosEmbed}(.)\\) is the same input embedding function in our Autoregressive Tree Sequencing, and \\(\\mathbf{Z} \\in \\mathbb{R}^{2048 \\times L}\\) is the resulting latent code. The latent code \\(\\mathbf{Z}\\) is then prepended to the initial [SOS] token in the Transformer decoder to provide point cloud-based conditioning for mesh generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.268, + 0.907, + 0.449 + ], + "angle": 0, + "content": "Architecture details: Our model employs a Transformer decoder with 24 layers, 16 attention heads, and a hidden dimension of 1024 and adds the sinusoidal positional encoding [31] to encode the token position. We apply a full self-attention for the latent code condition and a causal self-attention mask for the autoregressive decoder. PyTorch's FlexAttention is used to implement this attention mask efficiently. Additionally, we adopt fp16 mixed-precision to optimize computational speed and memory efficiency. We set the hidden dimension \\( C \\) of the positional embedding (PosEmbed) to 512 and use 7-bit quantization to discretize the coordinate output into 128 classes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.453, + 0.906, + 0.514 + ], + "angle": 0, + "content": "Training details: We use AdamW [14, 20] with a learning rate of \\(10^{-4}\\), \\(\\beta_{1} = 0.9\\) and \\(\\beta_{2} = 0.99\\) as the optimizer. Our model is trained with \\(8 \\times \\mathrm{A100 - 80GB}\\) GPUs for 5 days with an effective batch size of 128." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.906, + 0.639 + ], + "angle": 0, + "content": "Sampling strategy: We use a multinomial sampling strategy with a top-\\(k\\) of 5 during the generation process. Empirically, we find that adjusting the temperature at different stages achieves an optimal balance between diversity and generation quality. Specifically, we set the temperature to 1 when the stack length is below 10, reduce it to 0.4 when the stack length is below 30, and further decrease it to 0.2 beyond that." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.906, + 0.763 + ], + "angle": 0, + "content": "Inference adjustment: We find that a few adjustments during inference can improve the generation performance. First, we keep track of the generated faces for each step. If our model predicts a vertex that would form a triangle that is a duplicate to the previously generated faces, we adjust the prediction to [STOP] and retrieve the next input from the top of the stack. This checking operation is fast since the faces are of discrete tensor." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Additionally, we observe that our model often struggles to predict the [EOS] label in longer sequences. In these cases, while the [EOS] label consistently appears among the top 5 logits, it is rarely sampled. To address this, we apply an addition factor to the logit of [EOS], incrementally increasing this factor each time [EOS] appears in the top 5 logits. To avoid early [EOS] prediction after this adjustment, we bypass multinomial top-\\(k\\) sampling for [EOS] and select it only when it becomes the highest logit." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26612" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.353, + 0.108 + ], + "angle": 0, + "content": "4.3. Results on Objverse Dataset" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.113, + 0.271, + 0.128 + ], + "angle": 0, + "content": "4.3.1. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.132, + 0.484, + 0.222 + ], + "angle": 0, + "content": "We present the qualitative results of mesh generation conditioned on input point cloud in Figure 3, comparing our method with MeshAnything [3] and MeshAnythingV2 [4] as the baselines. As shown, our method demonstrates a notable improvement to generate meshes with higher face counts and refined details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.231, + 0.281, + 0.245 + ], + "angle": 0, + "content": "4.3.2. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.25, + 0.483, + 0.28 + ], + "angle": 0, + "content": "We use the following metrics to assess the quality of the generated meshes:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.283, + 0.484, + 0.359 + ], + "angle": 0, + "content": "- Chamfer Distance (CD). It provides an indication of geometric accuracy by measuring the average closest-point distance between points sampled from the source and reference meshes, computed bidirectionally. Lower values of CD indicate better geometric accuracy." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.359, + 0.484, + 0.51 + ], + "angle": 0, + "content": "- Normal Consistency (NC). It evaluates the surface orientation alignment between the source and reference meshes. Specifically, it measures the cosine similarity between the normals of each face in the source mesh and the nearest face in the reference mesh, computed bidirectionally. Higher NC values signify better normal alignment. We also report absolute Normal Consistency (|NC|), which disregards the sign, focusing solely on the magnitude of similarity. The mathematical details are provided in the supplementary materials." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.283, + 0.484, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.484, + 0.651 + ], + "angle": 0, + "content": "Objaverse evaluation set: We ran the evaluation on 200 samples in our validation set with 1 generation seed for each model. The quantitative results presented in Table 1, indicate that our model generates meshes more faithful to the ground truth meshes compared to the baselines. The NC values of MeshAnything [3] and MeshAnythingV2 [4] are relatively low due to flipped normals, which leads to inconsistencies in the sign of the normals. In contrast, our method generates meshes with consistent normal orientation." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.665, + 0.466, + 0.742 + ], + "angle": 0, + "content": "
ModelCD↓NC↑|NC| ↑
MeshAnything [3]0.01150.2230.853
MeshAnythingV2 [4]0.01020.1670.843
Ours0.00700.7980.880
" + }, + { + "type": "table_caption", + "bbox": [ + 0.107, + 0.745, + 0.466, + 0.759 + ], + "angle": 0, + "content": "Table 1. Quantitative results on Objverse evaluation set." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Tokenization effectiveness: The previous trained models and ours differ in dataset composition, point cloud conditioning, and training settings. To accurately evaluate the effectiveness of different tokenization methods, we conduct a controlled experiment on the subset of our filtered dataset of 24k samples with \\(\\leq 500\\) faces. All factors except for the tokenizers are kept the same: 1). 16-layer Transformers with 768 hidden dimension and positional encoding. 2). Point" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "cloud condition of 2048 tokens. 3). Training until 40k steps with an effective batch size of 128 and data augmentation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.905, + 0.198 + ], + "angle": 0, + "content": "Results on a test set of 200 samples, shown in Table 2, strongly indicate our method provides highly effective inductive bias. We use PivotMesh's [33] pretrained VQ-VAE since MeshAnything does not release their VQ-VAE encoder and noise resistant decoder fine-tuning code." + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.21, + 0.894, + 0.317 + ], + "angle": 0, + "content": "
TokenizerCD↓NC↑|NC| ↑Sequence Length↓
Naive [2]0.03760.6390.8229Nf
VQ-VAE [33]0.03520.6730.8156Nf
AMT [4]0.0327-0.0690.768±4Nf
Ours0.01000.7340.8742Nf + 2Nc
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.321, + 0.905, + 0.363 + ], + "angle": 0, + "content": "Table 2. Quantitative results on our controlled experiment. \\( N_{f} \\) and \\( N_{c} \\) are the number of triangular faces and connected components, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.374, + 0.735, + 0.389 + ], + "angle": 0, + "content": "4.4. Results on GSO Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.396, + 0.906, + 0.516 + ], + "angle": 0, + "content": "We further conduct a quantitative evaluation on GSO dataset [8], a dataset of real-world 3D scans to test the generalization capabilities of each model. In this experiment, we observe that the baseline models and ours are sensitive to the input point cloud, with results varying large on different sampling seeds. To mitigate this variability, we generate five samples for each model and select the mesh with the lowest Chamfer Distance for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.517, + 0.906, + 0.682 + ], + "angle": 0, + "content": "For our model, we sample point clouds directly from the meshes provided in the GSO dataset. However, direct sampling from these high-resolution meshes often results in reconstructions with extremely small faces. To address this, we decimate the meshes to five target faces ranging from 1000 to 2500 faces and then perform uniform sampling on these decimated versions. For MeshAnything and MeshAnythingV2, we sample the input point clouds from Marching Cubes meshes derived from both the original high-resolution and decimated (1500 faces) meshes using an 8-level octree." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.683, + 0.906, + 0.789 + ], + "angle": 0, + "content": "As shown in Table 3, our method achieves the best results across the evaluated metrics. Figure 4 presents qualitative comparisons, demonstrating that our model generates meshes more consistent with the input than the baselines. Additionally, Figure 5 compares our output with the decimated mesh, highlighting that our model can generate meshes with the topology of those created by 3D artists." + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.801, + 0.888, + 0.879 + ], + "angle": 0, + "content": "
ModelCD↓NC↑|NC| ↑
MeshAnything [3]0.01050.4530.869
MeshAnythingV2 [4]0.01160.32690.865
Ours0.00770.8420.897
" + }, + { + "type": "table_caption", + "bbox": [ + 0.566, + 0.882, + 0.852, + 0.895 + ], + "angle": 0, + "content": "Table 3. Quantitative results on GSO dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26613" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.154, + 0.083, + 0.222, + 0.096 + ], + "angle": 0, + "content": "GT Mesh" + }, + { + "type": "image", + "bbox": [ + 0.14, + 0.107, + 0.237, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.234, + 0.234, + 0.247 + ], + "angle": 0, + "content": "Face \\(= 4130\\)" + }, + { + "type": "image", + "bbox": [ + 0.145, + 0.25, + 0.23, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.329, + 0.234, + 0.342 + ], + "angle": 0, + "content": "Face \\(= 4316\\)" + }, + { + "type": "image", + "bbox": [ + 0.143, + 0.345, + 0.232, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.423, + 0.234, + 0.436 + ], + "angle": 0, + "content": "Face \\(= 3718\\)" + }, + { + "type": "image", + "bbox": [ + 0.143, + 0.439, + 0.227, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.518, + 0.234, + 0.531 + ], + "angle": 0, + "content": "Face \\(= 3694\\)" + }, + { + "type": "image", + "bbox": [ + 0.149, + 0.538, + 0.229, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.612, + 0.234, + 0.626 + ], + "angle": 0, + "content": "Face \\(= 2442\\)" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.64, + 0.268, + 0.681 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.707, + 0.234, + 0.72 + ], + "angle": 0, + "content": "Face \\(= 3638\\)" + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.731, + 0.25, + 0.804 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.146, + 0.818, + 0.23, + 0.831 + ], + "angle": 0, + "content": "Face \\(= 740\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.082, + 0.444, + 0.098 + ], + "angle": 0, + "content": "MeshAnything[3]" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.108, + 0.377, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.234, + 0.422, + 0.247 + ], + "angle": 0, + "content": "Face \\(= 439\\)" + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.108, + 0.48, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.258, + 0.37, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.258, + 0.47, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.329, + 0.422, + 0.342 + ], + "angle": 0, + "content": "Face \\(= 800\\)" + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.345, + 0.372, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.345, + 0.48, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.423, + 0.422, + 0.436 + ], + "angle": 0, + "content": "Face \\(= 630\\)" + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.444, + 0.369, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.397, + 0.445, + 0.459, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.518, + 0.422, + 0.531 + ], + "angle": 0, + "content": "Face \\(= 576\\)" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.538, + 0.37, + 0.603 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.54, + 0.469, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.612, + 0.422, + 0.625 + ], + "angle": 0, + "content": "Face \\(= 495\\)" + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.643, + 0.459, + 0.681 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.707, + 0.422, + 0.72 + ], + "angle": 0, + "content": "Face \\(= 584\\)" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.731, + 0.442, + 0.804 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.818, + 0.422, + 0.831 + ], + "angle": 0, + "content": "Face \\(= 235\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.082, + 0.662, + 0.098 + ], + "angle": 0, + "content": "MeshAnythingV2 [4]" + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.108, + 0.688, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.234, + 0.634, + 0.247 + ], + "angle": 0, + "content": "Face = 2247" + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.25, + 0.679, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.329, + 0.634, + 0.342 + ], + "angle": 0, + "content": "Face = 2246" + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.344, + 0.685, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.423, + 0.634, + 0.436 + ], + "angle": 0, + "content": "Face = 2090" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.439, + 0.666, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.518, + 0.634, + 0.531 + ], + "angle": 0, + "content": "Face = 1998" + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.538, + 0.689, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.612, + 0.634, + 0.625 + ], + "angle": 0, + "content": "Face = 2159" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.641, + 0.667, + 0.681 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.707, + 0.634, + 0.72 + ], + "angle": 0, + "content": "Face \\(= 1769\\)" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.73, + 0.65, + 0.804 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.818, + 0.629, + 0.831 + ], + "angle": 0, + "content": "Face \\(= 265\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.083, + 0.816, + 0.096 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.107, + 0.895, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.234, + 0.84, + 0.247 + ], + "angle": 0, + "content": "Face = 3311" + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.25, + 0.885, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.329, + 0.841, + 0.342 + ], + "angle": 0, + "content": "Face = 4545" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.344, + 0.788, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.804, + 0.346, + 0.894, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.423, + 0.841, + 0.436 + ], + "angle": 0, + "content": "Face \\(= 5122\\)" + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.441, + 0.783, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.812, + 0.441, + 0.879, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.518, + 0.841, + 0.53 + ], + "angle": 0, + "content": "Face \\(= 2708\\)" + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.538, + 0.786, + 0.609 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.54, + 0.884, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.612, + 0.841, + 0.625 + ], + "angle": 0, + "content": "Face = 2646" + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.641, + 0.875, + 0.68 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.707, + 0.841, + 0.72 + ], + "angle": 0, + "content": "Face = 2114" + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.731, + 0.857, + 0.804 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.754, + 0.818, + 0.835, + 0.83 + ], + "angle": 0, + "content": "Face \\(= 301\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.845, + 0.907, + 0.873 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison on Objverse dataset [7]. Our model is able to generate meshes with higher face counts and refined details compared to the baselines. Results from the baselines use point clouds sampled from marching cube meshes with 8-level octree." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26614" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.079, + 0.237, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.079, + 0.432, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.079, + 0.642, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.079, + 0.847, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.33, + 0.281, + 0.667, + 0.296 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison on GSO dataset [8]." + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.306, + 0.227, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.306, + 0.34, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.306, + 0.452, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.408, + 0.484, + 0.451 + ], + "angle": 0, + "content": "Figure 5. Comparison between the decimated mesh and our output. Our model is capable of generating meshes with the topology of those created by 3D artists." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.462, + 0.245, + 0.478 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.487, + 0.484, + 0.608 + ], + "angle": 0, + "content": "Decoder MLP head: We compare TreeMeshGPT trained with MLP heads that predict the x-y-z coordinates simultaneously and our proposed hierarchical MLP. Predicting the coordinate simultaneously leads to challenging sampling that results in noisy and more incomplete meshes, as shown in Figure 6. Running evaluation on the 200 validation samples used in Table 1 with this head yields \\(\\mathrm{CD} = 0.0114\\), \\(\\mathrm{NC} = 0.724\\), \\(|\\mathrm{NC}| = 0.847\\)." + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.615, + 0.222, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.615, + 0.334, + 0.712 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.615, + 0.445, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.718, + 0.483, + 0.761 + ], + "angle": 0, + "content": "Figure 6. MLP head ablation. Our hierarchical MLP maintains the sequential nature of the x-y-z coordinates prediction that results in easier sampling compared to simultaenous prediction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Tree traversal: We conduct an ablation study on smaller Transformer architecture comparing DFS and breadth-first-search (BFS) traversals in forming input-output sequences for our Autoregressive Tree Sequencing. As shown in the training perplexity plot in Figure 7, DFS traversal enables more efficient Transformer training. This improvement likely stems from the stronger local dependencies introduced by DFS, where each step is more predictable based on its immediate predecessors. In contrast, BFS traversal often" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.308, + 0.907, + 0.352 + ], + "angle": 0, + "content": "introduces dependencies between steps that are spatially or structurally distant, and thus complicating the learning process." + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.356, + 0.9, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.562, + 0.907, + 0.616 + ], + "angle": 0, + "content": "Figure 7. Training perplexity comparison Between BFS and DFS traversals. DFS traversal shows a better training perplexity compared to BFS. Shown in the plot is the perplexity for \\( y \\) -axis vertex coordinate." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.63, + 0.634, + 0.645 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.907, + 0.807 + ], + "angle": 0, + "content": "We present TreeMeshGPT, an autoregressive Transformer designed to generate high-quality artistic meshes aligned with input point clouds. TreeMeshGPT incorporates a novel Autoregressive Tree Sequencing technique instead of the conventional next-token prediction. Our Autoregressive Tree Sequencing represents each face with two tokens, enabling a 7-bit discretization model that can generate up to 5,500 triangular faces with 2,048 point cloud latent tokens. Experiments show that TreeMeshGPT can generate meshes with higher quality compared to the previous methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Limitations: Our model has a similar failure mode to the previous methods that the success rate decreases as the sequence length increases. Additionally, while our model has an improved capacity to generate meshes with higher face counts, challenges persist in enforcing an optimal mesh topology." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26615" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.152 + ], + "angle": 0, + "content": "Acknowledgment. This research is supported by the National Research Foundation (NRF) Singapore, under its NRF-Investigatorship Programme (Award ID. NRF-NRFI09-0008)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.165, + 0.188, + 0.18 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.189, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[1] Fausto Bernardini, Joshua Mittleman, Holly Rushmeier, Cláudio Silva, and Gabriel Taubin. The ball-pivoting algorithm for surface reconstruction. IEEE Transactions on Visualization and Computer Graphics, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.246, + 0.483, + 0.314 + ], + "angle": 0, + "content": "[2] Sijin Chen, Xin Chen, Anqi Pang, Xianfang Zeng, Wei Cheng, Yijun Fu, Fukun Yin, Yanru Wang, Zhibin Wang, Chi Zhang, et al. Meshxl: Neural coordinate field for generative 3d foundation models. Advances in Neural Information Processing Systems (NeurIPS), 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.316, + 0.483, + 0.383 + ], + "angle": 0, + "content": "[3] Yiwen Chen, Tong He, Di Huang, Weicai Ye, Sijin Chen, Ji-axiang Tang, Xin Chen, Zhongang Cai, Lei Yang, Gang Yu, et al. Meshanything: Artist-created mesh generation with autoregressive transformers. arXiv preprint arXiv:2406.10163, 2024. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.385, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[4] Yiwen Chen, Yikai Wang, Yihao Luo, Zhengyi Wang, Zilong Chen, Jun Zhu, Chi Zhang, and Guosheng Lin. Meshanything v2: Artist-created mesh generation with adjacent mesh tokenization. arXiv preprint arXiv:2408.02555, 2024. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.455, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[5] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.525, + 0.482, + 0.593 + ], + "angle": 0, + "content": "[6] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of \\(10\\mathrm{m} + 3\\mathrm{d}\\) objects. Advances in Neural Information Processing Systems (NeurIPS), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.595, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[7] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objaverse: A universe of annotated 3d objects. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.679, + 0.482, + 0.76 + ], + "angle": 0, + "content": "[8] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2022. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.762, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[9] Zekun Hao, David W Romero, Tsung-Yi Lin, and Ming-Yu Liu. Meshtron: High-fidelity, artist-like 3d mesh generation at scale. arXiv preprint arXiv:2412.09548, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[10] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. In International Conference on Learning Representations (ICLR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[11] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object gen" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.189, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.12 + ], + "angle": 0, + "content": "eration with dream fields. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.906, + 0.176 + ], + "angle": 0, + "content": "[12] Tao Ju, Frank Losasso, Scott Schaefer, and Joe Warren. Dual contouring of hermite data. In Proceedings of the 29th Annual Conference on Computer graphics and Interactive Techniques, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[13] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the 4th Eurographics Symposium on Geometry Processing, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[14] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.904, + 0.3 + ], + "angle": 0, + "content": "[15] Kiriakos N Kutulakos and Steven M Seitz. A theory of shape by space carving. International Journal of Computer Vision, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[16] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[17] Chenghao Li, Chaoning Zhang, Joseph Cho, Atish Waghwase, Lik-Hang Lee, Francois Rameau, Yang Yang, SungHo Bae, and Choong Seon Hong. Generative ai meets 3d: A survey on text-to-3d in aigc era. arXiv preprint arXiv:2305.06131, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.905, + 0.509 + ], + "angle": 0, + "content": "[18] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. In International Conference on Learning Representations (ICLR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[19] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM SIGGRAPH Computer Graphics, 1987. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.904, + 0.594 + ], + "angle": 0, + "content": "[20] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[21] Oscar Michel, Roi Bar-On, Richard Liu, Sage Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[22] Charlie Nash, Yaroslav Ganin, SM Ali Eslami, and Peter Battaglia. *Polygen: An autoregressive generative model of 3d meshes*. In International Conference on Machine Learning (ICML), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.905, + 0.76 + ], + "angle": 0, + "content": "[23] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.904, + 0.788 + ], + "angle": 0, + "content": "[24] Gregory M Nielson. Dual marching cubes. In IEEE Visualization, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[25] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In International Conference on Learning Representations (ICLR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26616" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[27] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. In International Conference on Learning Representations (ICLR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[28] Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[29] Jiaxiang Tang, Zhaoshuo Li, Zekun Hao, Xian Liu, Gang Zeng, Ming-Yu Liu, and Qinsheng Zhang. Edgerunner: Auto-regressive auto-encoder for artistic mesh generation. arXiv preprint arXiv:2409.18114, 2024. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[30] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in Neural Information Processing Systems (NeurIPS), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.388 + ], + "angle": 0, + "content": "[31] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems (NeurIPS), 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.471 + ], + "angle": 0, + "content": "[32] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.474, + 0.482, + 0.528 + ], + "angle": 0, + "content": "[33] Haohan Weng, Yikai Wang, Tong Zhang, CL Chen, and Jun Zhu. Pivotmesh: Generic 3d mesh generation via pivot vertices guidance. arXiv preprint arXiv:2405.16890, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.531, + 0.482, + 0.586 + ], + "angle": 0, + "content": "[34] Haohan Weng, Zibo Zhao, Biwen Lei, Xianghui Yang, Jian Liu, Zeqiang Lai, Zhuo Chen, Yuhong Liu, Jie Jiang, Chunchao Guo, et al. Scaling mesh generation via compressive tokenization. arXiv preprint arXiv:2411.07025, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.588, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[35] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.482, + 0.725 + ], + "angle": 0, + "content": "[36] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.728, + 0.482, + 0.783 + ], + "angle": 0, + "content": "[37] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Transactions on Graphics (TOG), 2023. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.785, + 0.482, + 0.853 + ], + "angle": 0, + "content": "[38] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 2024. 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.853 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26617" + } + ] +] \ No newline at end of file diff --git a/2025/TreeMeshGPT_ Artistic Mesh Generation with Autoregressive Tree Sequencing/cc43c3b2-b6bf-472a-9dd6-c74d4f03cee9_origin.pdf b/2025/TreeMeshGPT_ Artistic Mesh Generation with Autoregressive Tree Sequencing/cc43c3b2-b6bf-472a-9dd6-c74d4f03cee9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..93b715af1bb1d4c14e7a6408e76de05acddc1107 --- /dev/null +++ b/2025/TreeMeshGPT_ Artistic Mesh Generation with Autoregressive Tree Sequencing/cc43c3b2-b6bf-472a-9dd6-c74d4f03cee9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f8d808fe2c2ce70895bdf9d19f2501b5b2a33201e78fcbe6b48da7f0cdd7b734 +size 7457429 diff --git a/2025/TreeMeshGPT_ Artistic Mesh Generation with Autoregressive Tree Sequencing/full.md b/2025/TreeMeshGPT_ Artistic Mesh Generation with Autoregressive Tree Sequencing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..001e2f8eed0c2bdbf76ced7e6bec2beef04273da --- /dev/null +++ b/2025/TreeMeshGPT_ Artistic Mesh Generation with Autoregressive Tree Sequencing/full.md @@ -0,0 +1,388 @@ +# TreeMeshGPT: Artistic Mesh Generation with Autoregressive Tree Sequencing + +Stefan Lionar $^{1,2,3}$ + +Jiabin Liang $^{1,2,3}$ + +Gim Hee Lee3 + +1Sea AI Lab + +$^{2}$ Garena + +$^{3}$ National University of Singapore + +![](images/e748c1b371f1c49429413a1115bc2da6dbe308a8f69547ee0a28b5e1e7f76ff0.jpg) +Figure 1. Artistic meshes generated by TreeMeshGPT. Our method offers a novel sequencing approach for artistic mesh generation using autoregressive Transformer decoder by retrieving the next token from a dynamically growing tree structure. In our experiment with 7-bit discretization, TreeMeshGPT supports meshes with up to 5,500 triangular faces under strong point cloud conditioning. $^{12}$ + +# Abstract + +We introduce TreeMeshGPT, an autoregressive Transformer designed to generate high-quality artistic meshes aligned with input point clouds. Instead of the conventional next-token prediction in autoregressive Transformer, we propose a novel Autoregressive Tree Sequencing where the next input token is retrieved from a dynamically growing tree structure that is built upon the triangle adjacency of faces within the mesh. Our sequencing enables the mesh to extend locally from the last generated triangular face at each step, and therefore reduces training difficulty and improves mesh quality. Our approach represents each triangular face with two tokens, achieving a compression rate of approximately $22\%$ compared to the naive face tokenization. This efficient tokenization enables our model to generate highly detailed artistic meshes with strong point cloud conditioning, surpassing previous methods in both capacity and fi + +delyity. Furthermore, our method generates mesh with strong normal orientation constraints, minimizing flipped normals commonly encountered in previous methods. Our experiments show that TreeMeshGPT enhances the mesh generation quality with refined details and normal orientation consistency. + +# 1. Introduction + +In recent advancements in 3D generation, representations such as voxels, point clouds, and implicit functions are often utilized [17, 35]. After the generation process, these representations are converted into meshes using techniques like Marching Cubes [19], which result in dense, overtessellated triangular meshes. However, these dense meshes are unsuitable for applications that require real-time rendering, such as gaming and virtual reality. Although many mesh down-sampling algorithms can reduce the number of triangles, they also degrade mesh quality and produce messy, unstructured wireframes. In contrast, skilled artists can create highly compact meshes with minimal triangles + +while preserving the object's sharp details. Additionally, the wireframes created by artists are more regular, aesthetically pleasing, and better aligned with the object's feature or semantic boundaries, which facilitates further human interactions, such as editing and animation. However, the process of manually creating artist-quality meshes is highly time-consuming and labor-intensive. + +This laborous process highlights the need for automated methods that can replicate the quality of artist-created meshes without requiring extensive manual effort. MeshAnything [3] seeks to bridge the gap between advancements in 3D generation and artist-quality mesh creation by adding point cloud condition to the artistic mesh generation Transformer initially proposed by MeshGPT [28]. Point clouds are chosen as the condition because they are either the direct output or can be obtained conveniently from the generated Marching Cubes meshes of the advanced 3D generation techniques. + +MeshAnything [3] represents each triangular face with 9 latent tokens, leading to long sequences and limiting artistic mesh generation to 800 faces due to the Transformer's quadratic complexity. This constraint poses challenges for real-world applications, which often require meshes with significantly higher face counts to accurately represent complex objects and environments. In the subsequent works, MeshAnything V2 [4] and EdgeRunner [29] leverage triangle adjacency to create shorter sequences to represent the same meshes. Consequently, they are able to generate meshes with up to 1,600 and 4,000 faces, respectively. However, many real-world applications demand meshes with higher face count to accurately represent detailed surface topology. Additionally, challenges remain in generating high-quality meshes free from artifacts such as gaps, missing components, and flipped normals. + +To further improve tokenization efficiency and mesh quality, we introduce TreeMeshGPT. Unlike previous methods that rely on conventional next-token prediction in autoregressive Transformers, TreeMeshGPT introduces a novel Autoregressive Tree Sequencing approach. Instead of sequentially predicting next tokens, our method retrieves the next token from a dynamically growing tree structure built upon triangle adjacency within the mesh. This strategy allows the mesh to expand locally from the last generated triangular face at each step, and thus reducing training difficulty and enhancing mesh quality. Our approach represents each triangular face with two tokens, achieving a compression rate of $22\%$ compared to the naive face tokenization of 9 tokens per face. This efficient tokenization technique pushes the boundary of artistic mesh generation. With 7-bit discretization, it enables the generation of meshes with up to 5,500 triangles under a strong point cloud condition of 2,048 tokens. Furthermore, our method generates meshes with strong normal orientation constraints, + +minimizing flipped normals commonly encountered in MeshAnything [3] and MeshAnythingV2 [4]. + +In summary, our contributions are as follows: + +- We propose a novel Autoregressive Tree Sequencing technique that efficiently represents two tokens per triangular face. +- Our proposed tokenization enables the training of a 7-bit discretization artistic mesh generative model with strong point cloud condition, capable of generating high-quality meshes with up to 5,500 faces. +- Extensive experiments show that our model can generate higher quality meshes and can generalize to real-world 3D scans. + +# 2. Related Work + +# 2.1. Mesh Extraction + +Constructing a mesh from other 3D representations has been a research focus for decades. Among many successful methods [1, 15, 19], Marching Cubes [19] is the most widely used. It divides a scalar field into cubes and extracts triangles to approximate the isosurface. It is simple yet robust, producing watertight and 2-manifold results. Many improvements such as Dual Contouring [12] and Dual Marching Cubes [24] have been developed to enhance its capabilities. Another well-known method is Poisson reconstruction [13], which uses point clouds and normals as boundary conditions to solve a scalar field defined in 3D space, then applies Marching Cubes to extract the mesh. However, these approaches often focus on representing shapes with dense, over-tessellated meshes, resulting in messy, unstructured wireframes. This makes them unsuitable for downstream workflows that require efficient and structured meshes, such as real-time rendering, animation rigging, and editing. The dense, over-tessellated meshes not only increase computational load but also lack the regularity and semantic alignment necessary for the downstream processes. + +# 2.2.3D Generation + +After the great success of 2D image generation [26], 3D generation has become a promising research direction. This field focuses on generating 3D assets for industries such as gaming, film, and AR/VR. Due to the limited availability of 3D data, early methods [11, 21, 25, 27] relied on optimizing underlying representations to mimic conditioning from 2D images or multi-view 2D images. + +With the introduction of large-scale datasets [6, 7], feedforward 3D generation techniques, such as those in [10, 18, 36], have become feasible. These techniques significantly improve generation speed compared to optimization-based methods. However, the resulting meshes often suffer from lower quality and lack diversity. + +Inspired by the success of 2D diffusion models in image generation, many researchers have attempted to apply diffusion techniques directly to 3D data [5, 23, 32, 38]. For instance, CLAY [38], a transformer-based 3D latent diffusion model, achieves state-of-the-art results in high-quality shape generation. Despite these advances, these generation methods often require post-conversion for downstream applications, which remains a non-trivial challenge. + +# 2.3. Autoregressive Mesh Generation + +To address these limitations, recent approaches have leveraged autoregressive models for direct mesh generation [2, 22, 28, 33]. MeshGPT [28] was the first to tokenize a mesh through face sorting and compress it using a VQ-VAE [16, 30], followed by an autoregressive transformer to predict the compressed token sequence. This method enables the generation of meshes with direct supervision from artist-created topology information, which is often absent in previous approaches. + +Subsequent works [2, 3, 33] explored more efficient representations and incorporated input conditioning, such as point clouds and images. However, these methods are limited to generating meshes with fewer than 800 faces due to the long sequence lengths and the quadratic computational cost of transformers. MeshAnythingV2 [4] and EdgeRunner [29] introduced more compact mesh tokenization techniques that leverage triangle adjacency, increasing the maximum face count to 1,600 and 4,000, respectively. Meshtron [9] proposes hourglass architecture and sliding window inference to scale up face count capacity while using the naive tokenization [2]. In a concurrent work, BPT [34] proposes a compact tokenization using block-wise indexing and localized patch aggregation. These methods rely on the next-token prediction commonly used in large language models (LLMs) and autoregressive image generation. In contrast, we offer a novel sequencing strategy based on a dynamically growing tree structure, aiming to increase the maximum face count and improve the overall quality of the generated meshes. + +# 3. Method + +# 3.1. Autoregressive Tree Sequencing + +Tokenization plays a crucial role in autoregressive models, particularly in complex tasks like 3D mesh generation, where both the quality and efficiency of tokenization significantly affect model performance and scalability. In the context of mesh generation, tokenization involves encoding vertices, edges, or faces into sequential tokens that the model can process step-by-step. Drawing insights from LLMs, previous autoregressive mesh generation methods follow next-token prediction strategy that explicitly uses each output as the input for the subsequent step. + +Our approach differs from those methods by using a tree-based traversal scheme to grow the mesh during the generation process. Specifically, in TreeMeshGPT, the input to the Transformer decoder are directed mesh edges, represented as $\mathbf{I} = \{(v_1^n,v_2^n)\}_{n = 1}^N\in \mathbb{R}^{N\times 6}$ , where each $(v_{1}^{n},v_{2}^{n})$ denotes a pair of dequantized vertices for the generation at step $n$ . At each step, the Transformer decoder makes a localized prediction to either add a new vertex $v_{3}^{n}$ to expand the mesh by connecting to the initial pair of vertices $(v_{1}^{n},v_{2}^{n})$ or predict [STOP] label indicating that no further expansion should occur from the input edge. TreeMeshGPT leverages a tree traversal process to construct the sequential input-output pairs for the autoregressive generation, described as follows (Note: the sequence order description below omits auxiliary tokens for simplicity). + +Sequence order: We utilize a half-edge data structure with depth-first-search (DFS) traversal to construct the sequential input-output pairs, $(\mathbf{I},\mathbf{O}) = \{(I_n,o_n)\}_{n = 1}^N$ , accompanied by a dynamic stack $\mathbf{S}$ to manage the traversal process. + +The traversal process starts from a directed edge in a mesh, in which we determine if this edge has an opposite vertex that forms a triangle, is a boundary, or if the triangle has already been visited. When a new triangle is formed, the output $o_n$ is defined as the opposite vertex $v_3^n$ . Two new edges are created by connecting $v_3^n$ to the initial edge's vertices $I_n = (v_1^n, v_2^n)$ , resulting in edges $(v_3^n, v_2^n)$ and $(v_1^n, v_3^n)$ . These edges are directed in a counter-clockwise orientation on the potential next adjacent faces, as enforced by the half-edge data structure. The newly created edges are then pushed onto the stack S for continued traversal: + +$$ +\mathbf {S} := \mathbf {S} \odot \left(v _ {1} ^ {n}, v _ {3} ^ {n}\right) \quad \text {a n d} \quad \mathbf {S} := \mathbf {S} \odot \left(v _ {3} ^ {n}, v _ {2} ^ {n}\right) +$$ + +where $\odot$ represents the operation of pushing the edge $(v_{i}, v_{j})$ onto the top of stack $\mathbf{S}$ . Conversely, if the input edge is a boundary or if adding a new vertex forms a previously visited triangle, the output $o_{n}$ is set to the [STOP] label. In this case, no new vertex or edge is added. + +The input for the next step, $n + 1$ , is obtained by popping the top edge from the stack: + +$$ +I _ {n + 1} = \left(v _ {1} ^ {n + 1}, v _ {2} ^ {n + 1}\right) := \operatorname {t o p} (\mathbf {S}), \quad \mathbf {S} := \mathbf {S} \setminus \operatorname {t o p} (\mathbf {S}) +$$ + +where $\operatorname{top}(\mathbf{S})$ retrieves the current top edge of the stack $\mathbf{S}$ , and $\mathbf{S} := \mathbf{S} \setminus \operatorname{top}(\mathbf{S})$ updates $\mathbf{S}$ by removing this top edge. We initialize the stack with a directed edge at the lowest position of the mesh and its twin. The traversal then proceeds until all triangles are visited. A simple illustration of this sequencing process is provided in Figure 2. + +Note that a mesh may consist of multiple connected components. A component begins from an initial edge, expands as edges are added to a stack, and is considered fully traversed once the stack is empty. When there are multiple components in a mesh, the traversal of the first component starts from the edge with the lowest position in the mesh + +![](images/2f6c21b0591613b334314798257907ed55f43d7548a29b9eb3dec4589a9eb30f.jpg) +Figure 2. Illustration of the sequence order in our Autoregressive Tree sequencing. a). A small subset of a triangular mesh. [STOP] indicates boundary. b). An equivalent tree representation of the mesh. In this tree, each node is represented as a directed edge from a pair of vertices. The root is initialized with two child nodes: $(v_0, v_1)$ and its twin $(v_1, v_0)$ . A DFS traversal then proceeds to create the input-output sequence. c). Dynamic stack from the DFS traversal. The stack is initialized with $(v_0, v_1)$ and its twin $(v_1, v_0)$ . The input $I_n$ is always obtained from the top of the stack. Thus, $I_1 = (v_0, v_1)$ at step 1. The opposite vertex of $I_1$ is $v_2$ and consequently, $o_1$ is set to $v_2$ . Two new edges are then formed by connecting the opposite vertex to the initial pair of vertices: $(v_2, v_1)$ and $(v_0, v_2)$ . The direction is enforced to be counter-clockwise on the potential next adjacent faces. At step 2, $I_2 = (v_2, v_1)$ . Since $I_2$ is a boundary, $o_2$ is set to [STOP] label and no new edge is added to the stack. Step 3 and onwards follow the same traversal process. d). Transformer decoder sequence. The sequence in the Transformer decoder follows the input-output pairs from the tree traversal. The auxiliary tokens to initialize the generation of a connected component and the [EOS] are also added to the input-output sequence. + +![](images/b87f4dd24584de014f7a67b0c1f2e5cbe2bbdf77cb82fb67b5cf45c3f7d17ee0.jpg) + +![](images/78cb86f35cab0c13cc144d804fe83c8f3f1627a462130bcbfc960a5969392587.jpg) + +and continues until no edge remains in the stack. Each subsequent component begins from the next available edge positioned at the lowest among the remaining unvisited faces. + +Generation process: To initialize the generation of a mesh component, an auxiliary token $[\mathrm{SOS}]\in \mathbb{R}^D$ is used as the input to predict the first vertex $v_{1}$ . For the next step, a second auxiliary token $[\mathrm{SOS2}]\in \mathbb{R}^{C}$ , concatenated with the embedded representation of $v_{1}$ , serves as input to predict the second vertex $v_{2}$ . Once these initial vertices are predicted, the mesh generation proceeds through our Autoregressive Tree Sequencing with a stack initialized by $(v_{1},v_{2})$ and its twin $(v_{2},v_{1})$ . When the stack is empty—indicating the current component is complete—a new component is initialized by reintroducing the [SOS] and [SOS2] tokens. After all components have been generated, the sequence is terminated with an [EOS] label. + +The final mesh is constructed by gathering the faces formed from the initial input vertex pairs and their predicted opposite vertices: $\mathcal{M} = \bigcup_{n=1}^{N}(v_1^n,v_2^n,v_3^n)$ for the generation steps where $I_n \notin \{[\mathrm{SOS}], [\mathrm{SOS2}]\}$ and $o_n \notin \{[\mathrm{STOP}], [\mathrm{EOS}]\}$ . + +Input embedding: We employ a positional embedding function from [37] to encode each vertex into a high-dimensional space, capturing its positional information across multiple frequency bases. This embedding function $\mathrm{PosEmbed}(.) : \mathbb{R}^3 \to \mathbb{R}^C$ maps 3D coordinates to a $C$ -dimensional embedding. For each edge, the embeddings of its vertex pair are concatenated, creating a representation in $\mathbb{R}^{2C}$ , which is subsequently passed through an MLP to map it to the Transformer's hidden dimension, $\mathbb{R}^{2C} \to \mathbb{R}^D$ . + +Vertex prediction: In prior works on mesh generation, each vertex is represented as a sequence of three tokens corresponding to its quantized $x-$ , $y-$ , and $z$ -coordinates. Specifically, to predict a single vertex's position, those models generate each coordinate independently as separate tokens in sequence. This approach leads to longer sequences, as each vertex requires three distinct tokens. In contrast, our method predicts the vertex' quantized $x-$ , $y-$ , and $z$ -coordinates in one sequence length by using hierarchical MLP heads. This hierarchical approach maintains the sequential nature in predicting the $x-$ , $y-$ , and $z$ -coordinates. Further details can be found in the supplementary material. As shown in the ablation study (Section 4.5), our hierarchical MLP heads result in easier coordinate sampling compared to prediction heads that predict the $x-$ , $y-$ , and $z$ -coordinates simultaneously. + +Advantages: Our Autoregressive Tree Sequencing approach adds only two sequence steps per triangular face as each face introduces two new nodes during the tree traversal process. Additionally, since most meshes consist of only a few connected components, our method requires minimal auxiliary tokens. This efficient sequencing achieves a compression rate of approximately $22\%$ for most meshes compared to naive tokenization, which uses 9 tokens per face. Our compression rate is thus approximately double that of methods like MeshAnythingV2 [4] and EdgeRunner [29]. Additionally, by using a dynamic stack to manage the input sequence, our method allows the Transformer to focus solely on making localized predictions at each step, hence improving training efficiency. Furthermore, our method + +generates mesh with strong normal orientation constraints, minimizing flipped normals commonly encountered in MeshAnything [3] and MeshAnythingV2 [4]. + +Loss function: We aim to train the Transformer decoder $\theta$ to maximize the likelihood of generating the sequence of outputs $\{o_n\}_{n=1}^N$ given the input sequence $\{I_n\}_{n=1}^N$ : + +$$ +\prod_ {n = 1} ^ {N} P \left(o _ {n} \mid I _ {\leq n}; \theta\right). \tag {1} +$$ + +To this end, given the ground truth and predicted values with teacher-forcing across all steps, denoted by $\mathbf{O} = \{\mathbf{O}_x,\mathbf{O}_y,\mathbf{O}_z\}$ and $\hat{\mathbf{O}} = \{\hat{\mathbf{O}}_x,\hat{\mathbf{O}}_y,\hat{\mathbf{O}}_z\}$ , respectively, where each $\mathbf{O}_i$ represents the discretized vertex coordinate along a specific axis, we use a loss function defined as the sum of cross-entropy losses for each coordinate: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {C E}} \left(\mathbf {O} _ {x}, \hat {\mathbf {O}} _ {x}\right) + \mathcal {L} _ {\mathrm {C E}} \left(\mathbf {O} _ {y}, \hat {\mathbf {O}} _ {y}\right) + \mathcal {L} _ {\mathrm {C E}} \left(\mathbf {O} _ {z}, \hat {\mathbf {O}} _ {z}\right). \tag {2} +$$ + +To incorporate stopping conditions, we add the [STOP] and [EOS] labels to the class selection on the height axis, extending it with two additional classes beyond the discretized coordinate classes. + +# 4. Experiments + +# 4.1. Dataset + +We use Objaverse [7] meshes as our training dataset. To ensure high-quality meshes, we select meshes that meet the half-edge traversal requirement, i.e., they are manifold and have no flipped normal. All meshes are preprocessed by centering and normalizing them within a cube spanning $[-0.5, 0.5]$ . We apply 7-bit discretization, remove any duplicate triangles, and choose meshes with less than 5,500 faces. Additionally, we perform orthographic projections and exclude meshes where one of the projections has extremely small area or contains more than one cluster. After filtering, we retain a total of 75,000 meshes, of which 1,000 are reserved for validation, with the remainder used for training. To increase data diversity, we apply the following augmentations: + +- Scaling: Each axis is scaled independently by a factor randomly chosen from the range [0.75, 0.95]. +- Rotation: We first apply a $90^{\circ}$ or $-90^{\circ}$ rotation along the $x$ - or $y$ -axis with a probability of 0.3. Afterward, we perform a rotation around the $z$ -axis with an angle uniformly sampled from $[-180^{\circ}, 180^{\circ}]$ . + +# 4.2. Implementation Details + +Point cloud conditioning: Following previous approaches [3, 4, 29], we adopt point cloud conditioning to provide practical guidance for the generation process. To achieve this, we sample a point cloud from the input mesh surface and apply a lightweight encoder [3, 4, 37, 38]. Specifically, we sample 8192 points, denoted as $\mathbf{X} \in$ + +$\mathbb{R}^{8192 \times 3}$ , from the mesh surface. A cross-attention layer is then used to encode these points into a latent code: + +$$ +\mathbf {Z} = \operatorname {C r o s s A t t} (\mathbf {Q}, \operatorname {P o s E m b e d} (\mathbf {X})) \tag {3} +$$ + +where $\mathbf{Q} \in \mathbb{R}^{2048 \times C}$ represents query embeddings with a shorter sequence length of 2048 and hidden dimension of $C$ , $\mathrm{PosEmbed}(.)$ is the same input embedding function in our Autoregressive Tree Sequencing, and $\mathbf{Z} \in \mathbb{R}^{2048 \times L}$ is the resulting latent code. The latent code $\mathbf{Z}$ is then prepended to the initial [SOS] token in the Transformer decoder to provide point cloud-based conditioning for mesh generation. + +Architecture details: Our model employs a Transformer decoder with 24 layers, 16 attention heads, and a hidden dimension of 1024 and adds the sinusoidal positional encoding [31] to encode the token position. We apply a full self-attention for the latent code condition and a causal self-attention mask for the autoregressive decoder. PyTorch's FlexAttention is used to implement this attention mask efficiently. Additionally, we adopt fp16 mixed-precision to optimize computational speed and memory efficiency. We set the hidden dimension $C$ of the positional embedding (PosEmbed) to 512 and use 7-bit quantization to discretize the coordinate output into 128 classes. + +Training details: We use AdamW [14, 20] with a learning rate of $10^{-4}$ , $\beta_{1} = 0.9$ and $\beta_{2} = 0.99$ as the optimizer. Our model is trained with $8 \times \mathrm{A100 - 80GB}$ GPUs for 5 days with an effective batch size of 128. + +Sampling strategy: We use a multinomial sampling strategy with a top- $k$ of 5 during the generation process. Empirically, we find that adjusting the temperature at different stages achieves an optimal balance between diversity and generation quality. Specifically, we set the temperature to 1 when the stack length is below 10, reduce it to 0.4 when the stack length is below 30, and further decrease it to 0.2 beyond that. + +Inference adjustment: We find that a few adjustments during inference can improve the generation performance. First, we keep track of the generated faces for each step. If our model predicts a vertex that would form a triangle that is a duplicate to the previously generated faces, we adjust the prediction to [STOP] and retrieve the next input from the top of the stack. This checking operation is fast since the faces are of discrete tensor. + +Additionally, we observe that our model often struggles to predict the [EOS] label in longer sequences. In these cases, while the [EOS] label consistently appears among the top 5 logits, it is rarely sampled. To address this, we apply an addition factor to the logit of [EOS], incrementally increasing this factor each time [EOS] appears in the top 5 logits. To avoid early [EOS] prediction after this adjustment, we bypass multinomial top- $k$ sampling for [EOS] and select it only when it becomes the highest logit. + +# 4.3. Results on Objverse Dataset + +# 4.3.1. Qualitative Results + +We present the qualitative results of mesh generation conditioned on input point cloud in Figure 3, comparing our method with MeshAnything [3] and MeshAnythingV2 [4] as the baselines. As shown, our method demonstrates a notable improvement to generate meshes with higher face counts and refined details. + +# 4.3.2. Quantitative Results + +We use the following metrics to assess the quality of the generated meshes: + +- Chamfer Distance (CD). It provides an indication of geometric accuracy by measuring the average closest-point distance between points sampled from the source and reference meshes, computed bidirectionally. Lower values of CD indicate better geometric accuracy. +- Normal Consistency (NC). It evaluates the surface orientation alignment between the source and reference meshes. Specifically, it measures the cosine similarity between the normals of each face in the source mesh and the nearest face in the reference mesh, computed bidirectionally. Higher NC values signify better normal alignment. We also report absolute Normal Consistency (|NC|), which disregards the sign, focusing solely on the magnitude of similarity. The mathematical details are provided in the supplementary materials. + +Objaverse evaluation set: We ran the evaluation on 200 samples in our validation set with 1 generation seed for each model. The quantitative results presented in Table 1, indicate that our model generates meshes more faithful to the ground truth meshes compared to the baselines. The NC values of MeshAnything [3] and MeshAnythingV2 [4] are relatively low due to flipped normals, which leads to inconsistencies in the sign of the normals. In contrast, our method generates meshes with consistent normal orientation. + +
ModelCD↓NC↑|NC| ↑
MeshAnything [3]0.01150.2230.853
MeshAnythingV2 [4]0.01020.1670.843
Ours0.00700.7980.880
+ +Tokenization effectiveness: The previous trained models and ours differ in dataset composition, point cloud conditioning, and training settings. To accurately evaluate the effectiveness of different tokenization methods, we conduct a controlled experiment on the subset of our filtered dataset of 24k samples with $\leq 500$ faces. All factors except for the tokenizers are kept the same: 1). 16-layer Transformers with 768 hidden dimension and positional encoding. 2). Point + +cloud condition of 2048 tokens. 3). Training until 40k steps with an effective batch size of 128 and data augmentation. + +Results on a test set of 200 samples, shown in Table 2, strongly indicate our method provides highly effective inductive bias. We use PivotMesh's [33] pretrained VQ-VAE since MeshAnything does not release their VQ-VAE encoder and noise resistant decoder fine-tuning code. + +Table 1. Quantitative results on Objverse evaluation set. + +
TokenizerCD↓NC↑|NC| ↑Sequence Length↓
Naive [2]0.03760.6390.8229Nf
VQ-VAE [33]0.03520.6730.8156Nf
AMT [4]0.0327-0.0690.768±4Nf
Ours0.01000.7340.8742Nf + 2Nc
+ +# 4.4. Results on GSO Dataset + +We further conduct a quantitative evaluation on GSO dataset [8], a dataset of real-world 3D scans to test the generalization capabilities of each model. In this experiment, we observe that the baseline models and ours are sensitive to the input point cloud, with results varying large on different sampling seeds. To mitigate this variability, we generate five samples for each model and select the mesh with the lowest Chamfer Distance for evaluation. + +For our model, we sample point clouds directly from the meshes provided in the GSO dataset. However, direct sampling from these high-resolution meshes often results in reconstructions with extremely small faces. To address this, we decimate the meshes to five target faces ranging from 1000 to 2500 faces and then perform uniform sampling on these decimated versions. For MeshAnything and MeshAnythingV2, we sample the input point clouds from Marching Cubes meshes derived from both the original high-resolution and decimated (1500 faces) meshes using an 8-level octree. + +As shown in Table 3, our method achieves the best results across the evaluated metrics. Figure 4 presents qualitative comparisons, demonstrating that our model generates meshes more consistent with the input than the baselines. Additionally, Figure 5 compares our output with the decimated mesh, highlighting that our model can generate meshes with the topology of those created by 3D artists. + +Table 2. Quantitative results on our controlled experiment. $N_{f}$ and $N_{c}$ are the number of triangular faces and connected components, respectively. + +
ModelCD↓NC↑|NC| ↑
MeshAnything [3]0.01050.4530.869
MeshAnythingV2 [4]0.01160.32690.865
Ours0.00770.8420.897
+ +Table 3. Quantitative results on GSO dataset. + +![](images/4c7008871030e9587f1562e7c753c1859aebfa8dbd2440464e906d44f47eaa35.jpg) +GT Mesh +Face $= 4130$ + +![](images/e000042daab9dfbfd71f38951f0abccf0c07a0e37e17dbf182bdbbfb46be8809.jpg) +Face $= 4316$ + +![](images/e871f42414d9870a8d4643b49cde75d1c86a7a693b1a5d1383f8758d8190267a.jpg) + +![](images/b619cebc9bdbf251b4009c2c80c6ce016cd4b166184ff6b81bbb98a29d740e75.jpg) +Face $= 3718$ +Face $= 3694$ + +![](images/19a46b7910bd4e7cd66425e402881176a6b420c3c69f2eb2cebefa93cad99d2d.jpg) +Face $= 2442$ + +![](images/11e2efe6de1025af90451e21a33fbd2fca722fc132dea5ecdd2ac48820658026.jpg) + +![](images/cdbe6b9459c6b14f95837f49bb91bd5b2dbeb5617e853f1a140716b8ce038ff7.jpg) +Face $= 3638$ +Face $= 740$ +Figure 3. Qualitative comparison on Objverse dataset [7]. Our model is able to generate meshes with higher face counts and refined details compared to the baselines. Results from the baselines use point clouds sampled from marching cube meshes with 8-level octree. + +![](images/fd9f595fc5252f1f61aa19060e85596a1c4bbee35c5e7a41f3b8a987af0a24ac.jpg) +MeshAnything[3] +Face $= 439$ + +![](images/a276835df9ba95c2dabe8c33274b978bc9b1b240c0f34d037779ea43b4dfd867.jpg) + +![](images/9fcf3347e5801d9571c6426c6abea3832ae9baefae866f4f48e7381343562c09.jpg) +Face $= 800$ + +![](images/408052b671515aab1508f5b14d64a2c1813071d41d86b54ea99ef646359b075c.jpg) + +![](images/4a27f5d0996d0cb5a021242557e6974b7a87a06684499f3abb5b2ad0546e5c15.jpg) +Face $= 630$ + +![](images/3921a78ceb787db33deb87fb9617e51dd7d15b1b72551e928c72e8b1da0e8698.jpg) + +![](images/78ce9f926c0d6d3ff5c88067d089687b45750a8084a5788a3e31e0e86656566a.jpg) +Face $= 576$ + +![](images/c6a05f62242284713fe27ac42aa8e5dd8ac18acd2f89b1478c2c0cf3e2965b86.jpg) + +![](images/c77c5f3a535757ccbfacec013b057d3aeefa7cdfeaffb4ce5fc524afb95b409f.jpg) + +![](images/62515ae8f680c2b2fa841d7fc0dc922a4c90dabd6f9a1b763bc3ca746a523117.jpg) +Face $= 495$ + +![](images/46c8e94e65892df848358516405217ac41d43bcfca18b52ce8eceea49a773746.jpg) + +![](images/c7eb6e95203a5c3ed03255e83c3f563fd433748f3b0ae434ea88b8447f8e3301.jpg) +Face $= 584$ +Face $= 235$ + +![](images/adbfea0641ad810a1bdbd262f22ec2ad9fb77bea93f3e6a87b52db614d9bba50.jpg) +MeshAnythingV2 [4] + +![](images/9d45079fca32265b44c19920c3d1fd9bf9419954282437490a6c664cc563ec05.jpg) +Face = 2247 +Face = 2246 + +![](images/2bee1f01fd798c18bd50ec3a0cd1a854dd2d7c5e023e0425438350bf885dadca.jpg) + +![](images/11c60089c50c4d1a57340382ee4ed6dcd12bcc085ea646d9d7ea60fe961134a9.jpg) +Face = 2090 + +![](images/4ade773c361062907b496d276215643f34764126d0270e2d3c088b6f92518927.jpg) +Face = 1998 +Face = 2159 + +![](images/0896b0ac5011a8545b038d324d2b5f216f743fac963677c74cb704788e4c1bd7.jpg) + +![](images/d6cf071ba12b87287e08c471bba3adec9d42bcfec457a6d53b6a1d952365da49.jpg) +Face $= 1769$ +Face $= 265$ + +![](images/edce66028a2128e05a9952f5e8a0f0d42d04f1141d9d8826fb6fad70d91ee1ef.jpg) +Ours + +![](images/db2cf1a3cc461253e03ad5d69d010bf7187b96345e953a333c045db3f77a5267.jpg) +Face = 3311 + +![](images/1613048410d1a69a5fd9ff7f126338f8b42a3782338aef4fac345cf1293a56e3.jpg) +Face = 4545 +Face $= 5122$ + +![](images/aa3bcc8bc99211346acf57e6ba1a48d8de0e21a7a23ec8b10a68cc89127c6e2b.jpg) + +![](images/03154838e179a348065aadc23a79dc505b0253f6eb04521321b28730fb0dec22.jpg) +Face $= 2708$ + +![](images/8b6bb92ae561ded526374b44c20e523b4e594b887f3840aafbd810c930494c19.jpg) + +![](images/4c3f35e035726190e280c1355f4cd14ead46d6f454a1454e79d88ad18890d727.jpg) +Face = 2646 + +![](images/17d681719d8eddfdde4af39a245b17fdbf0e00074e5db8a7fc2eb1aedbdb286f.jpg) + +![](images/debb74a80c1c187d59811371cdfcf0ad3fffad7cfbbf46aafa008ff447466f70.jpg) + +![](images/53de8261f2468831ccacc32c9c31da56c31d904e5596dd946b3a643bd78d75d3.jpg) +Face = 2114 +Face $= 301$ + +![](images/7ef7ecfda59a2d78b1dc6fda89929c0f541a3fade4774c3cdd0a9878cad79061.jpg) + +![](images/759d5d3706e288a6bc29e93e0b143ebfe0b554b0ee5ab77cec3689744394d944.jpg) +Figure 4. Qualitative comparison on GSO dataset [8]. + +![](images/402ebf7e1680da38c2dfc0cc1b3eec4c54b07d78efe7e0204a76f6e394c545d2.jpg) + +![](images/89fec65e06197b931633e8e2984eb8d542e02e2d8eb467d736d5e994e9522daa.jpg) + +![](images/2503cf9c79ce30a9a12cade462fedff6a975b3748671614482bd2347b8411b57.jpg) +Figure 5. Comparison between the decimated mesh and our output. Our model is capable of generating meshes with the topology of those created by 3D artists. + +![](images/d5265daa069de701561f839cff25793d42131ccf1e7618922ac0cfe87524614f.jpg) + +![](images/f7c2c6610f54db4a3924337477e542097778c882300ef782084d009517360b21.jpg) + +# 4.5. Ablation Study + +Decoder MLP head: We compare TreeMeshGPT trained with MLP heads that predict the x-y-z coordinates simultaneously and our proposed hierarchical MLP. Predicting the coordinate simultaneously leads to challenging sampling that results in noisy and more incomplete meshes, as shown in Figure 6. Running evaluation on the 200 validation samples used in Table 1 with this head yields $\mathrm{CD} = 0.0114$ , $\mathrm{NC} = 0.724$ , $|\mathrm{NC}| = 0.847$ . + +![](images/709731a44bdd84b5af04b253a655d6faff6741ac1bd7b41d8df22e54918df39d.jpg) +Figure 6. MLP head ablation. Our hierarchical MLP maintains the sequential nature of the x-y-z coordinates prediction that results in easier sampling compared to simultaenous prediction. + +![](images/8f8af4a4df6b9aae4cbca26166892ccfcde95dddbaac58acedce918f58fb5bd7.jpg) + +![](images/4b7f5da83b96abae0953f1860cdcf2a984ebcb295915c37dc0da997ea1becc64.jpg) + +Tree traversal: We conduct an ablation study on smaller Transformer architecture comparing DFS and breadth-first-search (BFS) traversals in forming input-output sequences for our Autoregressive Tree Sequencing. As shown in the training perplexity plot in Figure 7, DFS traversal enables more efficient Transformer training. This improvement likely stems from the stronger local dependencies introduced by DFS, where each step is more predictable based on its immediate predecessors. In contrast, BFS traversal often + +introduces dependencies between steps that are spatially or structurally distant, and thus complicating the learning process. + +![](images/015486ac48dca7b28335bdfe1e67404acfa897eda850796d899fde074531d4e0.jpg) +Figure 7. Training perplexity comparison Between BFS and DFS traversals. DFS traversal shows a better training perplexity compared to BFS. Shown in the plot is the perplexity for $y$ -axis vertex coordinate. + +# 5. Conclusion + +We present TreeMeshGPT, an autoregressive Transformer designed to generate high-quality artistic meshes aligned with input point clouds. TreeMeshGPT incorporates a novel Autoregressive Tree Sequencing technique instead of the conventional next-token prediction. Our Autoregressive Tree Sequencing represents each face with two tokens, enabling a 7-bit discretization model that can generate up to 5,500 triangular faces with 2,048 point cloud latent tokens. Experiments show that TreeMeshGPT can generate meshes with higher quality compared to the previous methods. + +Limitations: Our model has a similar failure mode to the previous methods that the success rate decreases as the sequence length increases. Additionally, while our model has an improved capacity to generate meshes with higher face counts, challenges persist in enforcing an optimal mesh topology. + +Acknowledgment. This research is supported by the National Research Foundation (NRF) Singapore, under its NRF-Investigatorship Programme (Award ID. NRF-NRFI09-0008). + +# References + +[1] Fausto Bernardini, Joshua Mittleman, Holly Rushmeier, Cláudio Silva, and Gabriel Taubin. The ball-pivoting algorithm for surface reconstruction. IEEE Transactions on Visualization and Computer Graphics, 1999. 2 +[2] Sijin Chen, Xin Chen, Anqi Pang, Xianfang Zeng, Wei Cheng, Yijun Fu, Fukun Yin, Yanru Wang, Zhibin Wang, Chi Zhang, et al. Meshxl: Neural coordinate field for generative 3d foundation models. Advances in Neural Information Processing Systems (NeurIPS), 2024. 3, 6 +[3] Yiwen Chen, Tong He, Di Huang, Weicai Ye, Sijin Chen, Ji-axiang Tang, Xin Chen, Zhongang Cai, Lei Yang, Gang Yu, et al. Meshanything: Artist-created mesh generation with autoregressive transformers. arXiv preprint arXiv:2406.10163, 2024. 2, 3, 5, 6, 7, 8 +[4] Yiwen Chen, Yikai Wang, Yihao Luo, Zhengyi Wang, Zilong Chen, Jun Zhu, Chi Zhang, and Guosheng Lin. Meshanything v2: Artist-created mesh generation with adjacent mesh tokenization. arXiv preprint arXiv:2408.02555, 2024. 2, 3, 4, 5, 6, 7, 8 +[5] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[6] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects. Advances in Neural Information Processing Systems (NeurIPS), 2023. 2 +[7] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objaverse: A universe of annotated 3d objects. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 5, 7 +[8] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2022. 6, 8 +[9] Zekun Hao, David W Romero, Tsung-Yi Lin, and Ming-Yu Liu. Meshtron: High-fidelity, artist-like 3d mesh generation at scale. arXiv preprint arXiv:2412.09548, 2024. 3 +[10] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. In International Conference on Learning Representations (ICLR), 2024. 2 +[11] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object gen + +eration with dream fields. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[12] Tao Ju, Frank Losasso, Scott Schaefer, and Joe Warren. Dual contouring of hermite data. In Proceedings of the 29th Annual Conference on Computer graphics and Interactive Techniques, 2002. 2 +[13] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the 4th Eurographics Symposium on Geometry Processing, 2006. 2 +[14] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 5 +[15] Kiriakos N Kutulakos and Steven M Seitz. A theory of shape by space carving. International Journal of Computer Vision, 2000. 2 +[16] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[17] Chenghao Li, Chaoning Zhang, Joseph Cho, Atish Waghwase, Lik-Hang Lee, Francois Rameau, Yang Yang, SungHo Bae, and Choong Seon Hong. Generative ai meets 3d: A survey on text-to-3d in aigc era. arXiv preprint arXiv:2305.06131, 2023. 1 +[18] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. In International Conference on Learning Representations (ICLR), 2024. 2 +[19] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM SIGGRAPH Computer Graphics, 1987. 1, 2 +[20] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. 5 +[21] Oscar Michel, Roi Bar-On, Richard Liu, Sage Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[22] Charlie Nash, Yaroslav Ganin, SM Ali Eslami, and Peter Battaglia. *Polygen: An autoregressive generative model of 3d meshes*. In International Conference on Machine Learning (ICML), 2020. 3 +[23] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 3 +[24] Gregory M Nielson. Dual marching cubes. In IEEE Visualization, 2004. 2 +[25] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In International Conference on Learning Representations (ICLR), 2023. 2 +[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2 + +[27] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. In International Conference on Learning Representations (ICLR), 2024. 2 +[28] Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3 +[29] Jiaxiang Tang, Zhaoshuo Li, Zekun Hao, Xian Liu, Gang Zeng, Ming-Yu Liu, and Qinsheng Zhang. Edgerunner: Auto-regressive auto-encoder for artistic mesh generation. arXiv preprint arXiv:2409.18114, 2024. 2, 3, 4, 5 +[30] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in Neural Information Processing Systems (NeurIPS), 2017. 3 +[31] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems (NeurIPS), 2017. 5 +[32] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[33] Haohan Weng, Yikai Wang, Tong Zhang, CL Chen, and Jun Zhu. Pivotmesh: Generic 3d mesh generation via pivot vertices guidance. arXiv preprint arXiv:2405.16890, 2024. 3, 6 +[34] Haohan Weng, Zibo Zhao, Biwen Lei, Xianghui Yang, Jian Liu, Zeqiang Lai, Zhuo Chen, Yuhong Liu, Jie Jiang, Chunchao Guo, et al. Scaling mesh generation via compressive tokenization. arXiv preprint arXiv:2411.07025, 2024. 3 +[35] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, 2022. 1 +[36] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 2 +[37] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Transactions on Graphics (TOG), 2023. 4, 5 +[38] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 2024. 3, 5 \ No newline at end of file diff --git a/2025/TreeMeshGPT_ Artistic Mesh Generation with Autoregressive Tree Sequencing/images.zip b/2025/TreeMeshGPT_ Artistic Mesh Generation with Autoregressive Tree Sequencing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bfee47d186bcd161dfdfafd172a30769baa23081 --- /dev/null +++ b/2025/TreeMeshGPT_ Artistic Mesh Generation with Autoregressive Tree Sequencing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:77a74ef6ab7c3ae21bed7fa57f1ab70c024db0dabfa2d27cf1d45b64b65f3812 +size 569142 diff --git a/2025/TreeMeshGPT_ Artistic Mesh Generation with Autoregressive Tree Sequencing/layout.json b/2025/TreeMeshGPT_ Artistic Mesh Generation with Autoregressive Tree Sequencing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..46057990f5a383eeae4551c577b25fc4fffaa2e7 --- /dev/null +++ b/2025/TreeMeshGPT_ Artistic Mesh Generation with Autoregressive Tree Sequencing/layout.json @@ -0,0 +1,11125 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 62, + 103, + 549, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 103, + 549, + 122 + ], + "spans": [ + { + "bbox": [ + 62, + 103, + 549, + 122 + ], + "type": "text", + "content": "TreeMeshGPT: Artistic Mesh Generation with Autoregressive Tree Sequencing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 161, + 142, + 247, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 142, + 247, + 156 + ], + "spans": [ + { + "bbox": [ + 161, + 142, + 247, + 156 + ], + "type": "text", + "content": "Stefan Lionar" + }, + { + "bbox": [ + 161, + 142, + 247, + 156 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 269, + 142, + 348, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 142, + 348, + 157 + ], + "spans": [ + { + "bbox": [ + 269, + 142, + 348, + 157 + ], + "type": "text", + "content": "Jiabin Liang" + }, + { + "bbox": [ + 269, + 142, + 348, + 157 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 373, + 142, + 445, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 142, + 445, + 156 + ], + "spans": [ + { + "bbox": [ + 373, + 142, + 445, + 156 + ], + "type": "text", + "content": "Gim Hee Lee3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 165, + 195, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 165, + 195, + 178 + ], + "spans": [ + { + "bbox": [ + 132, + 165, + 195, + 178 + ], + "type": "text", + "content": "1Sea AI Lab" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 230, + 165, + 272, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 165, + 272, + 178 + ], + "spans": [ + { + "bbox": [ + 230, + 165, + 272, + 178 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 230, + 165, + 272, + 178 + ], + "type": "text", + "content": "Garena" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 165, + 474, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 165, + 474, + 180 + ], + "spans": [ + { + "bbox": [ + 308, + 165, + 474, + 180 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 308, + 165, + 474, + 180 + ], + "type": "text", + "content": "National University of Singapore" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 59, + 193, + 555, + 397 + ], + "blocks": [ + { + "bbox": [ + 59, + 193, + 555, + 397 + ], + "lines": [ + { + "bbox": [ + 59, + 193, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 59, + 193, + 555, + 397 + ], + "type": "image", + "image_path": "e748c1b371f1c49429413a1115bc2da6dbe308a8f69547ee0a28b5e1e7f76ff0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 404, + 555, + 440 + ], + "lines": [ + { + "bbox": [ + 55, + 404, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 555, + 440 + ], + "type": "text", + "content": "Figure 1. Artistic meshes generated by TreeMeshGPT. Our method offers a novel sequencing approach for artistic mesh generation using autoregressive Transformer decoder by retrieving the next token from a dynamically growing tree structure. In our experiment with 7-bit discretization, TreeMeshGPT supports meshes with up to 5,500 triangular faces under strong point cloud conditioning." + }, + { + "bbox": [ + 55, + 404, + 555, + 440 + ], + "type": "inline_equation", + "content": "^{12}" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 152, + 448, + 200, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 448, + 200, + 460 + ], + "spans": [ + { + "bbox": [ + 152, + 448, + 200, + 460 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 473, + 297, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 473, + 297, + 666 + ], + "spans": [ + { + "bbox": [ + 54, + 473, + 297, + 666 + ], + "type": "text", + "content": "We introduce TreeMeshGPT, an autoregressive Transformer designed to generate high-quality artistic meshes aligned with input point clouds. Instead of the conventional next-token prediction in autoregressive Transformer, we propose a novel Autoregressive Tree Sequencing where the next input token is retrieved from a dynamically growing tree structure that is built upon the triangle adjacency of faces within the mesh. Our sequencing enables the mesh to extend locally from the last generated triangular face at each step, and therefore reduces training difficulty and improves mesh quality. Our approach represents each triangular face with two tokens, achieving a compression rate of approximately " + }, + { + "bbox": [ + 54, + 473, + 297, + 666 + ], + "type": "inline_equation", + "content": "22\\%" + }, + { + "bbox": [ + 54, + 473, + 297, + 666 + ], + "type": "text", + "content": " compared to the naive face tokenization. This efficient tokenization enables our model to generate highly detailed artistic meshes with strong point cloud conditioning, surpassing previous methods in both capacity and fi" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 449, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 521 + ], + "type": "text", + "content": "delyity. Furthermore, our method generates mesh with strong normal orientation constraints, minimizing flipped normals commonly encountered in previous methods. Our experiments show that TreeMeshGPT enhances the mesh generation quality with refined details and normal orientation consistency." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 548, + 394, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 548, + 394, + 560 + ], + "spans": [ + { + "bbox": [ + 314, + 548, + 394, + 560 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 570, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 570, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 570, + 555, + 715 + ], + "type": "text", + "content": "In recent advancements in 3D generation, representations such as voxels, point clouds, and implicit functions are often utilized [17, 35]. After the generation process, these representations are converted into meshes using techniques like Marching Cubes [19], which result in dense, overtessellated triangular meshes. However, these dense meshes are unsuitable for applications that require real-time rendering, such as gaming and virtual reality. Although many mesh down-sampling algorithms can reduce the number of triangles, they also degrade mesh quality and produce messy, unstructured wireframes. In contrast, skilled artists can create highly compact meshes with minimal triangles" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 673, + 295, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 295, + 693 + ], + "type": "inline_equation", + "content": "{}^{1}" + }, + { + "bbox": [ + 55, + 673, + 295, + 693 + ], + "type": "text", + "content": " TreeMeshGPT supports a higher face count with finer discretization as it preserves the required manifold connectivity condition." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 693, + 294, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 294, + 712 + ], + "type": "text", + "content": "2Our code is available at: https://github.com/sail-sg/TreeMeshGPT." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26608" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "while preserving the object's sharp details. Additionally, the wireframes created by artists are more regular, aesthetically pleasing, and better aligned with the object's feature or semantic boundaries, which facilitates further human interactions, such as editing and animation. However, the process of manually creating artist-quality meshes is highly time-consuming and labor-intensive." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 158, + 294, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 158, + 294, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 294, + 290 + ], + "type": "text", + "content": "This laborous process highlights the need for automated methods that can replicate the quality of artist-created meshes without requiring extensive manual effort. MeshAnything [3] seeks to bridge the gap between advancements in 3D generation and artist-quality mesh creation by adding point cloud condition to the artistic mesh generation Transformer initially proposed by MeshGPT [28]. Point clouds are chosen as the condition because they are either the direct output or can be obtained conveniently from the generated Marching Cubes meshes of the advanced 3D generation techniques." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 293, + 295, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 295, + 484 + ], + "type": "text", + "content": "MeshAnything [3] represents each triangular face with 9 latent tokens, leading to long sequences and limiting artistic mesh generation to 800 faces due to the Transformer's quadratic complexity. This constraint poses challenges for real-world applications, which often require meshes with significantly higher face counts to accurately represent complex objects and environments. In the subsequent works, MeshAnything V2 [4] and EdgeRunner [29] leverage triangle adjacency to create shorter sequences to represent the same meshes. Consequently, they are able to generate meshes with up to 1,600 and 4,000 faces, respectively. However, many real-world applications demand meshes with higher face count to accurately represent detailed surface topology. Additionally, challenges remain in generating high-quality meshes free from artifacts such as gaps, missing components, and flipped normals." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "content": "To further improve tokenization efficiency and mesh quality, we introduce TreeMeshGPT. Unlike previous methods that rely on conventional next-token prediction in autoregressive Transformers, TreeMeshGPT introduces a novel Autoregressive Tree Sequencing approach. Instead of sequentially predicting next tokens, our method retrieves the next token from a dynamically growing tree structure built upon triangle adjacency within the mesh. This strategy allows the mesh to expand locally from the last generated triangular face at each step, and thus reducing training difficulty and enhancing mesh quality. Our approach represents each triangular face with two tokens, achieving a compression rate of " + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "inline_equation", + "content": "22\\%" + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "content": " compared to the naive face tokenization of 9 tokens per face. This efficient tokenization technique pushes the boundary of artistic mesh generation. With 7-bit discretization, it enables the generation of meshes with up to 5,500 triangles under a strong point cloud condition of 2,048 tokens. Furthermore, our method generates meshes with strong normal orientation constraints," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "minimizing flipped normals commonly encountered in MeshAnything [3] and MeshAnythingV2 [4]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 97, + 511, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 97, + 511, + 108 + ], + "spans": [ + { + "bbox": [ + 326, + 97, + 511, + 108 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 111, + 553, + 230 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 111, + 553, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 111, + 553, + 146 + ], + "spans": [ + { + "bbox": [ + 314, + 111, + 553, + 146 + ], + "type": "text", + "content": "- We propose a novel Autoregressive Tree Sequencing technique that efficiently represents two tokens per triangular face." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 147, + 553, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 147, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 147, + 553, + 194 + ], + "type": "text", + "content": "- Our proposed tokenization enables the training of a 7-bit discretization artistic mesh generative model with strong point cloud condition, capable of generating high-quality meshes with up to 5,500 faces." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 195, + 553, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 195, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 313, + 195, + 553, + 230 + ], + "type": "text", + "content": "- Extensive experiments show that our model can generate higher quality meshes and can generalize to real-world 3D scans." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 243, + 400, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 243, + 400, + 255 + ], + "spans": [ + { + "bbox": [ + 314, + 243, + 400, + 255 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 262, + 415, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 262, + 415, + 274 + ], + "spans": [ + { + "bbox": [ + 314, + 262, + 415, + 274 + ], + "type": "text", + "content": "2.1. Mesh Extraction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 280, + 555, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 555, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 555, + 532 + ], + "type": "text", + "content": "Constructing a mesh from other 3D representations has been a research focus for decades. Among many successful methods [1, 15, 19], Marching Cubes [19] is the most widely used. It divides a scalar field into cubes and extracts triangles to approximate the isosurface. It is simple yet robust, producing watertight and 2-manifold results. Many improvements such as Dual Contouring [12] and Dual Marching Cubes [24] have been developed to enhance its capabilities. Another well-known method is Poisson reconstruction [13], which uses point clouds and normals as boundary conditions to solve a scalar field defined in 3D space, then applies Marching Cubes to extract the mesh. However, these approaches often focus on representing shapes with dense, over-tessellated meshes, resulting in messy, unstructured wireframes. This makes them unsuitable for downstream workflows that require efficient and structured meshes, such as real-time rendering, animation rigging, and editing. The dense, over-tessellated meshes not only increase computational load but also lack the regularity and semantic alignment necessary for the downstream processes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 540, + 406, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 540, + 406, + 552 + ], + "spans": [ + { + "bbox": [ + 314, + 540, + 406, + 552 + ], + "type": "text", + "content": "2.2.3D Generation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 557, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 641 + ], + "type": "text", + "content": "After the great success of 2D image generation [26], 3D generation has become a promising research direction. This field focuses on generating 3D assets for industries such as gaming, film, and AR/VR. Due to the limited availability of 3D data, early methods [11, 21, 25, 27] relied on optimizing underlying representations to mimic conditioning from 2D images or multi-view 2D images." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": "With the introduction of large-scale datasets [6, 7], feedforward 3D generation techniques, such as those in [10, 18, 36], have become feasible. These techniques significantly improve generation speed compared to optimization-based methods. However, the resulting meshes often suffer from lower quality and lack diversity." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26609" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": "Inspired by the success of 2D diffusion models in image generation, many researchers have attempted to apply diffusion techniques directly to 3D data [5, 23, 32, 38]. For instance, CLAY [38], a transformer-based 3D latent diffusion model, achieves state-of-the-art results in high-quality shape generation. Despite these advances, these generation methods often require post-conversion for downstream applications, which remains a non-trivial challenge." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 177, + 231, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 177, + 231, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 177, + 231, + 190 + ], + "type": "text", + "content": "2.3. Autoregressive Mesh Generation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 195, + 295, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 195, + 295, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 195, + 295, + 302 + ], + "type": "text", + "content": "To address these limitations, recent approaches have leveraged autoregressive models for direct mesh generation [2, 22, 28, 33]. MeshGPT [28] was the first to tokenize a mesh through face sorting and compress it using a VQ-VAE [16, 30], followed by an autoregressive transformer to predict the compressed token sequence. This method enables the generation of meshes with direct supervision from artist-created topology information, which is often absent in previous approaches." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 304, + 295, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 295, + 542 + ], + "type": "text", + "content": "Subsequent works [2, 3, 33] explored more efficient representations and incorporated input conditioning, such as point clouds and images. However, these methods are limited to generating meshes with fewer than 800 faces due to the long sequence lengths and the quadratic computational cost of transformers. MeshAnythingV2 [4] and EdgeRunner [29] introduced more compact mesh tokenization techniques that leverage triangle adjacency, increasing the maximum face count to 1,600 and 4,000, respectively. Meshtron [9] proposes hourglass architecture and sliding window inference to scale up face count capacity while using the naive tokenization [2]. In a concurrent work, BPT [34] proposes a compact tokenization using block-wise indexing and localized patch aggregation. These methods rely on the next-token prediction commonly used in large language models (LLMs) and autoregressive image generation. In contrast, we offer a novel sequencing strategy based on a dynamically growing tree structure, aiming to increase the maximum face count and improve the overall quality of the generated meshes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 555, + 111, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 111, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 111, + 567 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 575, + 228, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 228, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 228, + 589 + ], + "type": "text", + "content": "3.1. Autoregressive Tree Sequencing" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Tokenization plays a crucial role in autoregressive models, particularly in complex tasks like 3D mesh generation, where both the quality and efficiency of tokenization significantly affect model performance and scalability. In the context of mesh generation, tokenization involves encoding vertices, edges, or faces into sequential tokens that the model can process step-by-step. Drawing insights from LLMs, previous autoregressive mesh generation methods follow next-token prediction strategy that explicitly uses each output as the input for the subsequent step." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "text", + "content": "Our approach differs from those methods by using a tree-based traversal scheme to grow the mesh during the generation process. Specifically, in TreeMeshGPT, the input to the Transformer decoder are directed mesh edges, represented as " + }, + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "inline_equation", + "content": "\\mathbf{I} = \\{(v_1^n,v_2^n)\\}_{n = 1}^N\\in \\mathbb{R}^{N\\times 6}" + }, + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "inline_equation", + "content": "(v_{1}^{n},v_{2}^{n})" + }, + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "text", + "content": " denotes a pair of dequantized vertices for the generation at step " + }, + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "text", + "content": ". At each step, the Transformer decoder makes a localized prediction to either add a new vertex " + }, + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "inline_equation", + "content": "v_{3}^{n}" + }, + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "text", + "content": " to expand the mesh by connecting to the initial pair of vertices " + }, + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "inline_equation", + "content": "(v_{1}^{n},v_{2}^{n})" + }, + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "text", + "content": " or predict [STOP] label indicating that no further expansion should occur from the input edge. TreeMeshGPT leverages a tree traversal process to construct the sequential input-output pairs for the autoregressive generation, described as follows (Note: the sequence order description below omits auxiliary tokens for simplicity)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 255, + 553, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 553, + 302 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 553, + 302 + ], + "type": "text", + "content": "Sequence order: We utilize a half-edge data structure with depth-first-search (DFS) traversal to construct the sequential input-output pairs, " + }, + { + "bbox": [ + 313, + 255, + 553, + 302 + ], + "type": "inline_equation", + "content": "(\\mathbf{I},\\mathbf{O}) = \\{(I_n,o_n)\\}_{n = 1}^N" + }, + { + "bbox": [ + 313, + 255, + 553, + 302 + ], + "type": "text", + "content": ", accompanied by a dynamic stack " + }, + { + "bbox": [ + 313, + 255, + 553, + 302 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 255, + 553, + 302 + ], + "type": "text", + "content": " to manage the traversal process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 303, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 555, + 434 + ], + "type": "text", + "content": "The traversal process starts from a directed edge in a mesh, in which we determine if this edge has an opposite vertex that forms a triangle, is a boundary, or if the triangle has already been visited. When a new triangle is formed, the output " + }, + { + "bbox": [ + 313, + 303, + 555, + 434 + ], + "type": "inline_equation", + "content": "o_n" + }, + { + "bbox": [ + 313, + 303, + 555, + 434 + ], + "type": "text", + "content": " is defined as the opposite vertex " + }, + { + "bbox": [ + 313, + 303, + 555, + 434 + ], + "type": "inline_equation", + "content": "v_3^n" + }, + { + "bbox": [ + 313, + 303, + 555, + 434 + ], + "type": "text", + "content": ". Two new edges are created by connecting " + }, + { + "bbox": [ + 313, + 303, + 555, + 434 + ], + "type": "inline_equation", + "content": "v_3^n" + }, + { + "bbox": [ + 313, + 303, + 555, + 434 + ], + "type": "text", + "content": " to the initial edge's vertices " + }, + { + "bbox": [ + 313, + 303, + 555, + 434 + ], + "type": "inline_equation", + "content": "I_n = (v_1^n, v_2^n)" + }, + { + "bbox": [ + 313, + 303, + 555, + 434 + ], + "type": "text", + "content": ", resulting in edges " + }, + { + "bbox": [ + 313, + 303, + 555, + 434 + ], + "type": "inline_equation", + "content": "(v_3^n, v_2^n)" + }, + { + "bbox": [ + 313, + 303, + 555, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 303, + 555, + 434 + ], + "type": "inline_equation", + "content": "(v_1^n, v_3^n)" + }, + { + "bbox": [ + 313, + 303, + 555, + 434 + ], + "type": "text", + "content": ". These edges are directed in a counter-clockwise orientation on the potential next adjacent faces, as enforced by the half-edge data structure. The newly created edges are then pushed onto the stack S for continued traversal:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 340, + 440, + 527, + 454 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 440, + 527, + 454 + ], + "spans": [ + { + "bbox": [ + 340, + 440, + 527, + 454 + ], + "type": "interline_equation", + "content": "\\mathbf {S} := \\mathbf {S} \\odot \\left(v _ {1} ^ {n}, v _ {3} ^ {n}\\right) \\quad \\text {a n d} \\quad \\mathbf {S} := \\mathbf {S} \\odot \\left(v _ {3} ^ {n}, v _ {2} ^ {n}\\right)", + "image_path": "0167f35168a167d37a69cea3eb2ac7108e350adba193725354c7049eb40cb5c2.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 460, + 553, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 553, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 553, + 519 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 460, + 553, + 519 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 460, + 553, + 519 + ], + "type": "text", + "content": " represents the operation of pushing the edge " + }, + { + "bbox": [ + 313, + 460, + 553, + 519 + ], + "type": "inline_equation", + "content": "(v_{i}, v_{j})" + }, + { + "bbox": [ + 313, + 460, + 553, + 519 + ], + "type": "text", + "content": " onto the top of stack " + }, + { + "bbox": [ + 313, + 460, + 553, + 519 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 460, + 553, + 519 + ], + "type": "text", + "content": ". Conversely, if the input edge is a boundary or if adding a new vertex forms a previously visited triangle, the output " + }, + { + "bbox": [ + 313, + 460, + 553, + 519 + ], + "type": "inline_equation", + "content": "o_{n}" + }, + { + "bbox": [ + 313, + 460, + 553, + 519 + ], + "type": "text", + "content": " is set to the [STOP] label. In this case, no new vertex or edge is added." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 520, + 553, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 553, + 544 + ], + "type": "text", + "content": "The input for the next step, " + }, + { + "bbox": [ + 313, + 520, + 553, + 544 + ], + "type": "inline_equation", + "content": "n + 1" + }, + { + "bbox": [ + 313, + 520, + 553, + 544 + ], + "type": "text", + "content": ", is obtained by popping the top edge from the stack:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 328, + 549, + 539, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 549, + 539, + 563 + ], + "spans": [ + { + "bbox": [ + 328, + 549, + 539, + 563 + ], + "type": "interline_equation", + "content": "I _ {n + 1} = \\left(v _ {1} ^ {n + 1}, v _ {2} ^ {n + 1}\\right) := \\operatorname {t o p} (\\mathbf {S}), \\quad \\mathbf {S} := \\mathbf {S} \\setminus \\operatorname {t o p} (\\mathbf {S})", + "image_path": "2a2eaca15de86a4096727e6f41e7caf172f51398400266133b6a5b3336af65b5.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 553, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 570, + 553, + 641 + ], + "type": "inline_equation", + "content": "\\operatorname{top}(\\mathbf{S})" + }, + { + "bbox": [ + 313, + 570, + 553, + 641 + ], + "type": "text", + "content": " retrieves the current top edge of the stack " + }, + { + "bbox": [ + 313, + 570, + 553, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 570, + 553, + 641 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 570, + 553, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{S} := \\mathbf{S} \\setminus \\operatorname{top}(\\mathbf{S})" + }, + { + "bbox": [ + 313, + 570, + 553, + 641 + ], + "type": "text", + "content": " updates " + }, + { + "bbox": [ + 313, + 570, + 553, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 570, + 553, + 641 + ], + "type": "text", + "content": " by removing this top edge. We initialize the stack with a directed edge at the lowest position of the mesh and its twin. The traversal then proceeds until all triangles are visited. A simple illustration of this sequencing process is provided in Figure 2." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "Note that a mesh may consist of multiple connected components. A component begins from an initial edge, expands as edges are added to a stack, and is considered fully traversed once the stack is empty. When there are multiple components in a mesh, the traversal of the first component starts from the edge with the lowest position in the mesh" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26610" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 202, + 210 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 202, + 210 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 202, + 210 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 202, + 210 + ], + "type": "image", + "image_path": "2f6c21b0591613b334314798257907ed55f43d7548a29b9eb3dec4589a9eb30f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "text", + "content": "Figure 2. Illustration of the sequence order in our Autoregressive Tree sequencing. a). A small subset of a triangular mesh. [STOP] indicates boundary. b). An equivalent tree representation of the mesh. In this tree, each node is represented as a directed edge from a pair of vertices. The root is initialized with two child nodes: " + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "inline_equation", + "content": "(v_0, v_1)" + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "text", + "content": " and its twin " + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "inline_equation", + "content": "(v_1, v_0)" + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "text", + "content": ". A DFS traversal then proceeds to create the input-output sequence. c). Dynamic stack from the DFS traversal. The stack is initialized with " + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "inline_equation", + "content": "(v_0, v_1)" + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "text", + "content": " and its twin " + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "inline_equation", + "content": "(v_1, v_0)" + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "text", + "content": ". The input " + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "inline_equation", + "content": "I_n" + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "text", + "content": " is always obtained from the top of the stack. Thus, " + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "inline_equation", + "content": "I_1 = (v_0, v_1)" + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "text", + "content": " at step 1. The opposite vertex of " + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "inline_equation", + "content": "I_1" + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "inline_equation", + "content": "v_2" + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "text", + "content": " and consequently, " + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "inline_equation", + "content": "o_1" + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "inline_equation", + "content": "v_2" + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "text", + "content": ". Two new edges are then formed by connecting the opposite vertex to the initial pair of vertices: " + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "inline_equation", + "content": "(v_2, v_1)" + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "inline_equation", + "content": "(v_0, v_2)" + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "text", + "content": ". The direction is enforced to be counter-clockwise on the potential next adjacent faces. At step 2, " + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "inline_equation", + "content": "I_2 = (v_2, v_1)" + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "inline_equation", + "content": "I_2" + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "text", + "content": " is a boundary, " + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "inline_equation", + "content": "o_2" + }, + { + "bbox": [ + 55, + 220, + 555, + 331 + ], + "type": "text", + "content": " is set to [STOP] label and no new edge is added to the stack. Step 3 and onwards follow the same traversal process. d). Transformer decoder sequence. The sequence in the Transformer decoder follows the input-output pairs from the tree traversal. The auxiliary tokens to initialize the generation of a connected component and the [EOS] are also added to the input-output sequence." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 205, + 70, + 355, + 211 + ], + "blocks": [ + { + "bbox": [ + 205, + 70, + 355, + 211 + ], + "lines": [ + { + "bbox": [ + 205, + 70, + 355, + 211 + ], + "spans": [ + { + "bbox": [ + 205, + 70, + 355, + 211 + ], + "type": "image", + "image_path": "b87f4dd24584de014f7a67b0c1f2e5cbe2bbdf77cb82fb67b5cf45c3f7d17ee0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 356, + 70, + 553, + 211 + ], + "blocks": [ + { + "bbox": [ + 356, + 70, + 553, + 211 + ], + "lines": [ + { + "bbox": [ + 356, + 70, + 553, + 211 + ], + "spans": [ + { + "bbox": [ + 356, + 70, + 553, + 211 + ], + "type": "image", + "image_path": "78cb86f35cab0c13cc144d804fe83c8f3f1627a462130bcbfc960a5969392587.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 346, + 295, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 346, + 295, + 381 + ], + "spans": [ + { + "bbox": [ + 54, + 346, + 295, + 381 + ], + "type": "text", + "content": "and continues until no edge remains in the stack. Each subsequent component begins from the next available edge positioned at the lowest among the remaining unvisited faces." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 385, + 295, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 385, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 54, + 385, + 295, + 540 + ], + "type": "text", + "content": "Generation process: To initialize the generation of a mesh component, an auxiliary token " + }, + { + "bbox": [ + 54, + 385, + 295, + 540 + ], + "type": "inline_equation", + "content": "[\\mathrm{SOS}]\\in \\mathbb{R}^D" + }, + { + "bbox": [ + 54, + 385, + 295, + 540 + ], + "type": "text", + "content": " is used as the input to predict the first vertex " + }, + { + "bbox": [ + 54, + 385, + 295, + 540 + ], + "type": "inline_equation", + "content": "v_{1}" + }, + { + "bbox": [ + 54, + 385, + 295, + 540 + ], + "type": "text", + "content": ". For the next step, a second auxiliary token " + }, + { + "bbox": [ + 54, + 385, + 295, + 540 + ], + "type": "inline_equation", + "content": "[\\mathrm{SOS2}]\\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 54, + 385, + 295, + 540 + ], + "type": "text", + "content": ", concatenated with the embedded representation of " + }, + { + "bbox": [ + 54, + 385, + 295, + 540 + ], + "type": "inline_equation", + "content": "v_{1}" + }, + { + "bbox": [ + 54, + 385, + 295, + 540 + ], + "type": "text", + "content": ", serves as input to predict the second vertex " + }, + { + "bbox": [ + 54, + 385, + 295, + 540 + ], + "type": "inline_equation", + "content": "v_{2}" + }, + { + "bbox": [ + 54, + 385, + 295, + 540 + ], + "type": "text", + "content": ". Once these initial vertices are predicted, the mesh generation proceeds through our Autoregressive Tree Sequencing with a stack initialized by " + }, + { + "bbox": [ + 54, + 385, + 295, + 540 + ], + "type": "inline_equation", + "content": "(v_{1},v_{2})" + }, + { + "bbox": [ + 54, + 385, + 295, + 540 + ], + "type": "text", + "content": " and its twin " + }, + { + "bbox": [ + 54, + 385, + 295, + 540 + ], + "type": "inline_equation", + "content": "(v_{2},v_{1})" + }, + { + "bbox": [ + 54, + 385, + 295, + 540 + ], + "type": "text", + "content": ". When the stack is empty—indicating the current component is complete—a new component is initialized by reintroducing the [SOS] and [SOS2] tokens. After all components have been generated, the sequence is terminated with an [EOS] label." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 542, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 295, + 602 + ], + "type": "text", + "content": "The final mesh is constructed by gathering the faces formed from the initial input vertex pairs and their predicted opposite vertices: " + }, + { + "bbox": [ + 55, + 542, + 295, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\bigcup_{n=1}^{N}(v_1^n,v_2^n,v_3^n)" + }, + { + "bbox": [ + 55, + 542, + 295, + 602 + ], + "type": "text", + "content": " for the generation steps where " + }, + { + "bbox": [ + 55, + 542, + 295, + 602 + ], + "type": "inline_equation", + "content": "I_n \\notin \\{[\\mathrm{SOS}], [\\mathrm{SOS2}]\\}" + }, + { + "bbox": [ + 55, + 542, + 295, + 602 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 542, + 295, + 602 + ], + "type": "inline_equation", + "content": "o_n \\notin \\{[\\mathrm{STOP}], [\\mathrm{EOS}]\\}" + }, + { + "bbox": [ + 55, + 542, + 295, + 602 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "Input embedding: We employ a positional embedding function from [37] to encode each vertex into a high-dimensional space, capturing its positional information across multiple frequency bases. This embedding function " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{PosEmbed}(.) : \\mathbb{R}^3 \\to \\mathbb{R}^C" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " maps 3D coordinates to a " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "-dimensional embedding. For each edge, the embeddings of its vertex pair are concatenated, creating a representation in " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{2C}" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": ", which is subsequently passed through an MLP to map it to the Transformer's hidden dimension, " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{2C} \\to \\mathbb{R}^D" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 346, + 555, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 537 + ], + "type": "text", + "content": "Vertex prediction: In prior works on mesh generation, each vertex is represented as a sequence of three tokens corresponding to its quantized " + }, + { + "bbox": [ + 313, + 346, + 555, + 537 + ], + "type": "inline_equation", + "content": "x-" + }, + { + "bbox": [ + 313, + 346, + 555, + 537 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 346, + 555, + 537 + ], + "type": "inline_equation", + "content": "y-" + }, + { + "bbox": [ + 313, + 346, + 555, + 537 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 346, + 555, + 537 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 346, + 555, + 537 + ], + "type": "text", + "content": "-coordinates. Specifically, to predict a single vertex's position, those models generate each coordinate independently as separate tokens in sequence. This approach leads to longer sequences, as each vertex requires three distinct tokens. In contrast, our method predicts the vertex' quantized " + }, + { + "bbox": [ + 313, + 346, + 555, + 537 + ], + "type": "inline_equation", + "content": "x-" + }, + { + "bbox": [ + 313, + 346, + 555, + 537 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 346, + 555, + 537 + ], + "type": "inline_equation", + "content": "y-" + }, + { + "bbox": [ + 313, + 346, + 555, + 537 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 346, + 555, + 537 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 346, + 555, + 537 + ], + "type": "text", + "content": "-coordinates in one sequence length by using hierarchical MLP heads. This hierarchical approach maintains the sequential nature in predicting the " + }, + { + "bbox": [ + 313, + 346, + 555, + 537 + ], + "type": "inline_equation", + "content": "x-" + }, + { + "bbox": [ + 313, + 346, + 555, + 537 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 346, + 555, + 537 + ], + "type": "inline_equation", + "content": "y-" + }, + { + "bbox": [ + 313, + 346, + 555, + 537 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 346, + 555, + 537 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 346, + 555, + 537 + ], + "type": "text", + "content": "-coordinates. Further details can be found in the supplementary material. As shown in the ablation study (Section 4.5), our hierarchical MLP heads result in easier coordinate sampling compared to prediction heads that predict the " + }, + { + "bbox": [ + 313, + 346, + 555, + 537 + ], + "type": "inline_equation", + "content": "x-" + }, + { + "bbox": [ + 313, + 346, + 555, + 537 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 346, + 555, + 537 + ], + "type": "inline_equation", + "content": "y-" + }, + { + "bbox": [ + 313, + 346, + 555, + 537 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 346, + 555, + 537 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 346, + 555, + 537 + ], + "type": "text", + "content": "-coordinates simultaneously." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": "Advantages: Our Autoregressive Tree Sequencing approach adds only two sequence steps per triangular face as each face introduces two new nodes during the tree traversal process. Additionally, since most meshes consist of only a few connected components, our method requires minimal auxiliary tokens. This efficient sequencing achieves a compression rate of approximately " + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "inline_equation", + "content": "22\\%" + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": " for most meshes compared to naive tokenization, which uses 9 tokens per face. Our compression rate is thus approximately double that of methods like MeshAnythingV2 [4] and EdgeRunner [29]. Additionally, by using a dynamic stack to manage the input sequence, our method allows the Transformer to focus solely on making localized predictions at each step, hence improving training efficiency. Furthermore, our method" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26611" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "generates mesh with strong normal orientation constraints, minimizing flipped normals commonly encountered in MeshAnything [3] and MeshAnythingV2 [4]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 111, + 295, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 111, + 295, + 148 + ], + "spans": [ + { + "bbox": [ + 55, + 111, + 295, + 148 + ], + "type": "text", + "content": "Loss function: We aim to train the Transformer decoder " + }, + { + "bbox": [ + 55, + 111, + 295, + 148 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 111, + 295, + 148 + ], + "type": "text", + "content": " to maximize the likelihood of generating the sequence of outputs " + }, + { + "bbox": [ + 55, + 111, + 295, + 148 + ], + "type": "inline_equation", + "content": "\\{o_n\\}_{n=1}^N" + }, + { + "bbox": [ + 55, + 111, + 295, + 148 + ], + "type": "text", + "content": " given the input sequence " + }, + { + "bbox": [ + 55, + 111, + 295, + 148 + ], + "type": "inline_equation", + "content": "\\{I_n\\}_{n=1}^N" + }, + { + "bbox": [ + 55, + 111, + 295, + 148 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 135, + 148, + 294, + 177 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 148, + 294, + 177 + ], + "spans": [ + { + "bbox": [ + 135, + 148, + 294, + 177 + ], + "type": "interline_equation", + "content": "\\prod_ {n = 1} ^ {N} P \\left(o _ {n} \\mid I _ {\\leq n}; \\theta\\right). \\tag {1}", + "image_path": "0f98efe52ec93c76b25a5da77fa733ef4a16c874b85b7395bccdc054c44fc048.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 178, + 295, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 178, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 295, + 251 + ], + "type": "text", + "content": "To this end, given the ground truth and predicted values with teacher-forcing across all steps, denoted by " + }, + { + "bbox": [ + 55, + 178, + 295, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{O} = \\{\\mathbf{O}_x,\\mathbf{O}_y,\\mathbf{O}_z\\}" + }, + { + "bbox": [ + 55, + 178, + 295, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 178, + 295, + 251 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{O}} = \\{\\hat{\\mathbf{O}}_x,\\hat{\\mathbf{O}}_y,\\hat{\\mathbf{O}}_z\\}" + }, + { + "bbox": [ + 55, + 178, + 295, + 251 + ], + "type": "text", + "content": ", respectively, where each " + }, + { + "bbox": [ + 55, + 178, + 295, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_i" + }, + { + "bbox": [ + 55, + 178, + 295, + 251 + ], + "type": "text", + "content": " represents the discretized vertex coordinate along a specific axis, we use a loss function defined as the sum of cross-entropy losses for each coordinate:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 257, + 295, + 272 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 257, + 295, + 272 + ], + "spans": [ + { + "bbox": [ + 61, + 257, + 295, + 272 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {C E}} \\left(\\mathbf {O} _ {x}, \\hat {\\mathbf {O}} _ {x}\\right) + \\mathcal {L} _ {\\mathrm {C E}} \\left(\\mathbf {O} _ {y}, \\hat {\\mathbf {O}} _ {y}\\right) + \\mathcal {L} _ {\\mathrm {C E}} \\left(\\mathbf {O} _ {z}, \\hat {\\mathbf {O}} _ {z}\\right). \\tag {2}", + "image_path": "f9cd31c9a5a97eea73a422e09e11ee51432fa03ce7de2072d97d755c03696306.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 278, + 295, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 295, + 326 + ], + "type": "text", + "content": "To incorporate stopping conditions, we add the [STOP] and [EOS] labels to the class selection on the height axis, extending it with two additional classes beyond the discretized coordinate classes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 337, + 137, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 137, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 137, + 350 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 357, + 114, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 114, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 114, + 368 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 374, + 295, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 295, + 540 + ], + "type": "text", + "content": "We use Objaverse [7] meshes as our training dataset. To ensure high-quality meshes, we select meshes that meet the half-edge traversal requirement, i.e., they are manifold and have no flipped normal. All meshes are preprocessed by centering and normalizing them within a cube spanning " + }, + { + "bbox": [ + 55, + 374, + 295, + 540 + ], + "type": "inline_equation", + "content": "[-0.5, 0.5]" + }, + { + "bbox": [ + 55, + 374, + 295, + 540 + ], + "type": "text", + "content": ". We apply 7-bit discretization, remove any duplicate triangles, and choose meshes with less than 5,500 faces. Additionally, we perform orthographic projections and exclude meshes where one of the projections has extremely small area or contains more than one cluster. After filtering, we retain a total of 75,000 meshes, of which 1,000 are reserved for validation, with the remainder used for training. To increase data diversity, we apply the following augmentations:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 542, + 295, + 613 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 55, + 542, + 295, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 295, + 566 + ], + "type": "text", + "content": "- Scaling: Each axis is scaled independently by a factor randomly chosen from the range [0.75, 0.95]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 566, + 295, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 295, + 613 + ], + "type": "text", + "content": "- Rotation: We first apply a " + }, + { + "bbox": [ + 55, + 566, + 295, + 613 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 55, + 566, + 295, + 613 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 566, + 295, + 613 + ], + "type": "inline_equation", + "content": "-90^{\\circ}" + }, + { + "bbox": [ + 55, + 566, + 295, + 613 + ], + "type": "text", + "content": " rotation along the " + }, + { + "bbox": [ + 55, + 566, + 295, + 613 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 566, + 295, + 613 + ], + "type": "text", + "content": "- or " + }, + { + "bbox": [ + 55, + 566, + 295, + 613 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 566, + 295, + 613 + ], + "type": "text", + "content": "-axis with a probability of 0.3. Afterward, we perform a rotation around the " + }, + { + "bbox": [ + 55, + 566, + 295, + 613 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 566, + 295, + 613 + ], + "type": "text", + "content": "-axis with an angle uniformly sampled from " + }, + { + "bbox": [ + 55, + 566, + 295, + 613 + ], + "type": "inline_equation", + "content": "[-180^{\\circ}, 180^{\\circ}]" + }, + { + "bbox": [ + 55, + 566, + 295, + 613 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 621, + 188, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 188, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 188, + 634 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "type": "text", + "content": "Point cloud conditioning: Following previous approaches [3, 4, 29], we adopt point cloud conditioning to provide practical guidance for the generation process. To achieve this, we sample a point cloud from the input mesh surface and apply a lightweight encoder [3, 4, 37, 38]. Specifically, we sample 8192 points, denoted as " + }, + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{8192 \\times 3}" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": ", from the mesh surface. A cross-attention layer is then used to encode these points into a latent code:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 364, + 104, + 553, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 104, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 364, + 104, + 553, + 118 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} = \\operatorname {C r o s s A t t} (\\mathbf {Q}, \\operatorname {P o s E m b e d} (\\mathbf {X})) \\tag {3}", + "image_path": "c01d16e2ed548c1613e2f67c38112d31e35373cd83838ffc1f59b90de98ff9bd.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 124, + 553, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 124, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 124, + 553, + 209 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 124, + 553, + 209 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} \\in \\mathbb{R}^{2048 \\times C}" + }, + { + "bbox": [ + 313, + 124, + 553, + 209 + ], + "type": "text", + "content": " represents query embeddings with a shorter sequence length of 2048 and hidden dimension of " + }, + { + "bbox": [ + 313, + 124, + 553, + 209 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 124, + 553, + 209 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 124, + 553, + 209 + ], + "type": "inline_equation", + "content": "\\mathrm{PosEmbed}(.)" + }, + { + "bbox": [ + 313, + 124, + 553, + 209 + ], + "type": "text", + "content": " is the same input embedding function in our Autoregressive Tree Sequencing, and " + }, + { + "bbox": [ + 313, + 124, + 553, + 209 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} \\in \\mathbb{R}^{2048 \\times L}" + }, + { + "bbox": [ + 313, + 124, + 553, + 209 + ], + "type": "text", + "content": " is the resulting latent code. The latent code " + }, + { + "bbox": [ + 313, + 124, + 553, + 209 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 313, + 124, + 553, + 209 + ], + "type": "text", + "content": " is then prepended to the initial [SOS] token in the Transformer decoder to provide point cloud-based conditioning for mesh generation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 212, + 555, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 555, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 555, + 355 + ], + "type": "text", + "content": "Architecture details: Our model employs a Transformer decoder with 24 layers, 16 attention heads, and a hidden dimension of 1024 and adds the sinusoidal positional encoding [31] to encode the token position. We apply a full self-attention for the latent code condition and a causal self-attention mask for the autoregressive decoder. PyTorch's FlexAttention is used to implement this attention mask efficiently. Additionally, we adopt fp16 mixed-precision to optimize computational speed and memory efficiency. We set the hidden dimension " + }, + { + "bbox": [ + 313, + 212, + 555, + 355 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 212, + 555, + 355 + ], + "type": "text", + "content": " of the positional embedding (PosEmbed) to 512 and use 7-bit quantization to discretize the coordinate output into 128 classes." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "text", + "content": "Training details: We use AdamW [14, 20] with a learning rate of " + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.99" + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "text", + "content": " as the optimizer. Our model is trained with " + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "inline_equation", + "content": "8 \\times \\mathrm{A100 - 80GB}" + }, + { + "bbox": [ + 313, + 358, + 554, + 407 + ], + "type": "text", + "content": " GPUs for 5 days with an effective batch size of 128." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 411, + 554, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 554, + 506 + ], + "type": "text", + "content": "Sampling strategy: We use a multinomial sampling strategy with a top-" + }, + { + "bbox": [ + 313, + 411, + 554, + 506 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 411, + 554, + 506 + ], + "type": "text", + "content": " of 5 during the generation process. Empirically, we find that adjusting the temperature at different stages achieves an optimal balance between diversity and generation quality. Specifically, we set the temperature to 1 when the stack length is below 10, reduce it to 0.4 when the stack length is below 30, and further decrease it to 0.2 beyond that." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 510, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 604 + ], + "type": "text", + "content": "Inference adjustment: We find that a few adjustments during inference can improve the generation performance. First, we keep track of the generated faces for each step. If our model predicts a vertex that would form a triangle that is a duplicate to the previously generated faces, we adjust the prediction to [STOP] and retrieve the next input from the top of the stack. This checking operation is fast since the faces are of discrete tensor." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "Additionally, we observe that our model often struggles to predict the [EOS] label in longer sequences. In these cases, while the [EOS] label consistently appears among the top 5 logits, it is rarely sampled. To address this, we apply an addition factor to the logit of [EOS], incrementally increasing this factor each time [EOS] appears in the top 5 logits. To avoid early [EOS] prediction after this adjustment, we bypass multinomial top-" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": " sampling for [EOS] and select it only when it becomes the highest logit." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26612" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 216, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 216, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 216, + 85 + ], + "type": "text", + "content": "4.3. Results on Objverse Dataset" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 165, + 101 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 165, + 101 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 165, + 101 + ], + "type": "text", + "content": "4.3.1. Qualitative Results" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 104, + 296, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 104, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 104, + 296, + 175 + ], + "type": "text", + "content": "We present the qualitative results of mesh generation conditioned on input point cloud in Figure 3, comparing our method with MeshAnything [3] and MeshAnythingV2 [4] as the baselines. As shown, our method demonstrates a notable improvement to generate meshes with higher face counts and refined details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 182, + 171, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 171, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 171, + 194 + ], + "type": "text", + "content": "4.3.2. Quantitative Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 198, + 295, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 295, + 221 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 295, + 221 + ], + "type": "text", + "content": "We use the following metrics to assess the quality of the generated meshes:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 224, + 296, + 403 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 224, + 296, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 296, + 284 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 296, + 284 + ], + "type": "text", + "content": "- Chamfer Distance (CD). It provides an indication of geometric accuracy by measuring the average closest-point distance between points sampled from the source and reference meshes, computed bidirectionally. Lower values of CD indicate better geometric accuracy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 284, + 296, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 296, + 403 + ], + "type": "text", + "content": "- Normal Consistency (NC). It evaluates the surface orientation alignment between the source and reference meshes. Specifically, it measures the cosine similarity between the normals of each face in the source mesh and the nearest face in the reference mesh, computed bidirectionally. Higher NC values signify better normal alignment. We also report absolute Normal Consistency (|NC|), which disregards the sign, focusing solely on the magnitude of similarity. The mathematical details are provided in the supplementary materials." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 408, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 296, + 515 + ], + "type": "text", + "content": "Objaverse evaluation set: We ran the evaluation on 200 samples in our validation set with 1 generation seed for each model. The quantitative results presented in Table 1, indicate that our model generates meshes more faithful to the ground truth meshes compared to the baselines. The NC values of MeshAnything [3] and MeshAnythingV2 [4] are relatively low due to flipped normals, which leads to inconsistencies in the sign of the normals. In contrast, our method generates meshes with consistent normal orientation." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 65, + 526, + 285, + 587 + ], + "blocks": [ + { + "bbox": [ + 65, + 526, + 285, + 587 + ], + "lines": [ + { + "bbox": [ + 65, + 526, + 285, + 587 + ], + "spans": [ + { + "bbox": [ + 65, + 526, + 285, + 587 + ], + "type": "table", + "html": "
ModelCD↓NC↑|NC| ↑
MeshAnything [3]0.01150.2230.853
MeshAnythingV2 [4]0.01020.1670.843
Ours0.00700.7980.880
", + "image_path": "55a9155c167c0c9b9c570c1276bc09c280235cce01181bacbb3167b723855313.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "Tokenization effectiveness: The previous trained models and ours differ in dataset composition, point cloud conditioning, and training settings. To accurately evaluate the effectiveness of different tokenization methods, we conduct a controlled experiment on the subset of our filtered dataset of 24k samples with " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\leq 500" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " faces. All factors except for the tokenizers are kept the same: 1). 16-layer Transformers with 768 hidden dimension and positional encoding. 2). Point" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "cloud condition of 2048 tokens. 3). Training until 40k steps with an effective batch size of 128 and data augmentation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 96, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 553, + 156 + ], + "type": "text", + "content": "Results on a test set of 200 samples, shown in Table 2, strongly indicate our method provides highly effective inductive bias. We use PivotMesh's [33] pretrained VQ-VAE since MeshAnything does not release their VQ-VAE encoder and noise resistant decoder fine-tuning code." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 320, + 166, + 547, + 251 + ], + "blocks": [ + { + "bbox": [ + 65, + 590, + 285, + 601 + ], + "lines": [ + { + "bbox": [ + 65, + 590, + 285, + 601 + ], + "spans": [ + { + "bbox": [ + 65, + 590, + 285, + 601 + ], + "type": "text", + "content": "Table 1. Quantitative results on Objverse evaluation set." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 166, + 547, + 251 + ], + "lines": [ + { + "bbox": [ + 320, + 166, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 320, + 166, + 547, + 251 + ], + "type": "table", + "html": "
TokenizerCD↓NC↑|NC| ↑Sequence Length↓
Naive [2]0.03760.6390.8229Nf
VQ-VAE [33]0.03520.6730.8156Nf
AMT [4]0.0327-0.0690.768±4Nf
Ours0.01000.7340.8742Nf + 2Nc
", + "image_path": "4858961ad0b1114687bd4d3327ed3dc6ca9bbe15cbd35867ee8c99db28a47950.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 296, + 449, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 296, + 449, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 449, + 308 + ], + "type": "text", + "content": "4.4. Results on GSO Dataset" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 313, + 554, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 554, + 408 + ], + "type": "text", + "content": "We further conduct a quantitative evaluation on GSO dataset [8], a dataset of real-world 3D scans to test the generalization capabilities of each model. In this experiment, we observe that the baseline models and ours are sensitive to the input point cloud, with results varying large on different sampling seeds. To mitigate this variability, we generate five samples for each model and select the mesh with the lowest Chamfer Distance for evaluation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 409, + 554, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 554, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 554, + 540 + ], + "type": "text", + "content": "For our model, we sample point clouds directly from the meshes provided in the GSO dataset. However, direct sampling from these high-resolution meshes often results in reconstructions with extremely small faces. To address this, we decimate the meshes to five target faces ranging from 1000 to 2500 faces and then perform uniform sampling on these decimated versions. For MeshAnything and MeshAnythingV2, we sample the input point clouds from Marching Cubes meshes derived from both the original high-resolution and decimated (1500 faces) meshes using an 8-level octree." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 540, + 554, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 554, + 624 + ], + "type": "text", + "content": "As shown in Table 3, our method achieves the best results across the evaluated metrics. Figure 4 presents qualitative comparisons, demonstrating that our model generates meshes more consistent with the input than the baselines. Additionally, Figure 5 compares our output with the decimated mesh, highlighting that our model can generate meshes with the topology of those created by 3D artists." + } + ] + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 323, + 634, + 543, + 696 + ], + "blocks": [ + { + "bbox": [ + 313, + 254, + 553, + 287 + ], + "lines": [ + { + "bbox": [ + 313, + 254, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 553, + 287 + ], + "type": "text", + "content": "Table 2. Quantitative results on our controlled experiment. " + }, + { + "bbox": [ + 313, + 254, + 553, + 287 + ], + "type": "inline_equation", + "content": "N_{f}" + }, + { + "bbox": [ + 313, + 254, + 553, + 287 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 254, + 553, + 287 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 313, + 254, + 553, + 287 + ], + "type": "text", + "content": " are the number of triangular faces and connected components, respectively." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 634, + 543, + 696 + ], + "lines": [ + { + "bbox": [ + 323, + 634, + 543, + 696 + ], + "spans": [ + { + "bbox": [ + 323, + 634, + 543, + 696 + ], + "type": "table", + "html": "
ModelCD↓NC↑|NC| ↑
MeshAnything [3]0.01050.4530.869
MeshAnythingV2 [4]0.01160.32690.865
Ours0.00770.8420.897
", + "image_path": "b7578df61dbabe98a46909775a973055e9ca5998ca445b1448a09d716f8ac7d6.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 346, + 698, + 521, + 708 + ], + "lines": [ + { + "bbox": [ + 346, + 698, + 521, + 708 + ], + "spans": [ + { + "bbox": [ + 346, + 698, + 521, + 708 + ], + "type": "text", + "content": "Table 3. Quantitative results on GSO dataset." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26613" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 84, + 145, + 180 + ], + "blocks": [ + { + "bbox": [ + 94, + 65, + 135, + 76 + ], + "lines": [ + { + "bbox": [ + 94, + 65, + 135, + 76 + ], + "spans": [ + { + "bbox": [ + 94, + 65, + 135, + 76 + ], + "type": "text", + "content": "GT Mesh" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 85, + 84, + 145, + 180 + ], + "lines": [ + { + "bbox": [ + 85, + 84, + 145, + 180 + ], + "spans": [ + { + "bbox": [ + 85, + 84, + 145, + 180 + ], + "type": "image", + "image_path": "4c7008871030e9587f1562e7c753c1859aebfa8dbd2440464e906d44f47eaa35.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 185, + 143, + 195 + ], + "lines": [ + { + "bbox": [ + 86, + 185, + 143, + 195 + ], + "spans": [ + { + "bbox": [ + 86, + 185, + 143, + 195 + ], + "type": "text", + "content": "Face " + }, + { + "bbox": [ + 86, + 185, + 143, + 195 + ], + "type": "inline_equation", + "content": "= 4130" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 88, + 198, + 140, + 257 + ], + "blocks": [ + { + "bbox": [ + 88, + 198, + 140, + 257 + ], + "lines": [ + { + "bbox": [ + 88, + 198, + 140, + 257 + ], + "spans": [ + { + "bbox": [ + 88, + 198, + 140, + 257 + ], + "type": "image", + "image_path": "e000042daab9dfbfd71f38951f0abccf0c07a0e37e17dbf182bdbbfb46be8809.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 260, + 143, + 270 + ], + "lines": [ + { + "bbox": [ + 86, + 260, + 143, + 270 + ], + "spans": [ + { + "bbox": [ + 86, + 260, + 143, + 270 + ], + "type": "text", + "content": "Face " + }, + { + "bbox": [ + 86, + 260, + 143, + 270 + ], + "type": "inline_equation", + "content": "= 4316" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 87, + 273, + 141, + 331 + ], + "blocks": [ + { + "bbox": [ + 87, + 273, + 141, + 331 + ], + "lines": [ + { + "bbox": [ + 87, + 273, + 141, + 331 + ], + "spans": [ + { + "bbox": [ + 87, + 273, + 141, + 331 + ], + "type": "image", + "image_path": "e871f42414d9870a8d4643b49cde75d1c86a7a693b1a5d1383f8758d8190267a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 87, + 347, + 138, + 407 + ], + "blocks": [ + { + "bbox": [ + 86, + 335, + 143, + 345 + ], + "lines": [ + { + "bbox": [ + 86, + 335, + 143, + 345 + ], + "spans": [ + { + "bbox": [ + 86, + 335, + 143, + 345 + ], + "type": "text", + "content": "Face " + }, + { + "bbox": [ + 86, + 335, + 143, + 345 + ], + "type": "inline_equation", + "content": "= 3718" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 87, + 347, + 138, + 407 + ], + "lines": [ + { + "bbox": [ + 87, + 347, + 138, + 407 + ], + "spans": [ + { + "bbox": [ + 87, + 347, + 138, + 407 + ], + "type": "image", + "image_path": "b619cebc9bdbf251b4009c2c80c6ce016cd4b166184ff6b81bbb98a29d740e75.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 410, + 143, + 420 + ], + "lines": [ + { + "bbox": [ + 86, + 410, + 143, + 420 + ], + "spans": [ + { + "bbox": [ + 86, + 410, + 143, + 420 + ], + "type": "text", + "content": "Face " + }, + { + "bbox": [ + 86, + 410, + 143, + 420 + ], + "type": "inline_equation", + "content": "= 3694" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 91, + 426, + 140, + 475 + ], + "blocks": [ + { + "bbox": [ + 91, + 426, + 140, + 475 + ], + "lines": [ + { + "bbox": [ + 91, + 426, + 140, + 475 + ], + "spans": [ + { + "bbox": [ + 91, + 426, + 140, + 475 + ], + "type": "image", + "image_path": "19a46b7910bd4e7cd66425e402881176a6b420c3c69f2eb2cebefa93cad99d2d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 484, + 143, + 495 + ], + "lines": [ + { + "bbox": [ + 86, + 484, + 143, + 495 + ], + "spans": [ + { + "bbox": [ + 86, + 484, + 143, + 495 + ], + "type": "text", + "content": "Face " + }, + { + "bbox": [ + 86, + 484, + 143, + 495 + ], + "type": "inline_equation", + "content": "= 2442" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 65, + 506, + 164, + 539 + ], + "blocks": [ + { + "bbox": [ + 65, + 506, + 164, + 539 + ], + "lines": [ + { + "bbox": [ + 65, + 506, + 164, + 539 + ], + "spans": [ + { + "bbox": [ + 65, + 506, + 164, + 539 + ], + "type": "image", + "image_path": "11e2efe6de1025af90451e21a33fbd2fca722fc132dea5ecdd2ac48820658026.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 78, + 578, + 153, + 636 + ], + "blocks": [ + { + "bbox": [ + 86, + 559, + 143, + 570 + ], + "lines": [ + { + "bbox": [ + 86, + 559, + 143, + 570 + ], + "spans": [ + { + "bbox": [ + 86, + 559, + 143, + 570 + ], + "type": "text", + "content": "Face " + }, + { + "bbox": [ + 86, + 559, + 143, + 570 + ], + "type": "inline_equation", + "content": "= 3638" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 78, + 578, + 153, + 636 + ], + "lines": [ + { + "bbox": [ + 78, + 578, + 153, + 636 + ], + "spans": [ + { + "bbox": [ + 78, + 578, + 153, + 636 + ], + "type": "image", + "image_path": "cdbe6b9459c6b14f95837f49bb91bd5b2dbeb5617e853f1a140716b8ce038ff7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 647, + 140, + 658 + ], + "lines": [ + { + "bbox": [ + 89, + 647, + 140, + 658 + ], + "spans": [ + { + "bbox": [ + 89, + 647, + 140, + 658 + ], + "type": "text", + "content": "Face " + }, + { + "bbox": [ + 89, + 647, + 140, + 658 + ], + "type": "inline_equation", + "content": "= 740" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 669, + 555, + 691 + ], + "lines": [ + { + "bbox": [ + 55, + 669, + 555, + 691 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 555, + 691 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison on Objverse dataset [7]. Our model is able to generate meshes with higher face counts and refined details compared to the baselines. Results from the baselines use point clouds sampled from marching cube meshes with 8-level octree." + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 171, + 85, + 230, + 180 + ], + "blocks": [ + { + "bbox": [ + 192, + 64, + 271, + 77 + ], + "lines": [ + { + "bbox": [ + 192, + 64, + 271, + 77 + ], + "spans": [ + { + "bbox": [ + 192, + 64, + 271, + 77 + ], + "type": "text", + "content": "MeshAnything[3]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 171, + 85, + 230, + 180 + ], + "lines": [ + { + "bbox": [ + 171, + 85, + 230, + 180 + ], + "spans": [ + { + "bbox": [ + 171, + 85, + 230, + 180 + ], + "type": "image", + "image_path": "fd9f595fc5252f1f61aa19060e85596a1c4bbee35c5e7a41f3b8a987af0a24ac.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 185, + 258, + 195 + ], + "lines": [ + { + "bbox": [ + 206, + 185, + 258, + 195 + ], + "spans": [ + { + "bbox": [ + 206, + 185, + 258, + 195 + ], + "type": "text", + "content": "Face " + }, + { + "bbox": [ + 206, + 185, + 258, + 195 + ], + "type": "inline_equation", + "content": "= 439" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 236, + 85, + 293, + 180 + ], + "blocks": [ + { + "bbox": [ + 236, + 85, + 293, + 180 + ], + "lines": [ + { + "bbox": [ + 236, + 85, + 293, + 180 + ], + "spans": [ + { + "bbox": [ + 236, + 85, + 293, + 180 + ], + "type": "image", + "image_path": "a276835df9ba95c2dabe8c33274b978bc9b1b240c0f34d037779ea43b4dfd867.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 206, + 204, + 226, + 257 + ], + "blocks": [ + { + "bbox": [ + 206, + 204, + 226, + 257 + ], + "lines": [ + { + "bbox": [ + 206, + 204, + 226, + 257 + ], + "spans": [ + { + "bbox": [ + 206, + 204, + 226, + 257 + ], + "type": "image", + "image_path": "9fcf3347e5801d9571c6426c6abea3832ae9baefae866f4f48e7381343562c09.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 260, + 258, + 270 + ], + "lines": [ + { + "bbox": [ + 206, + 260, + 258, + 270 + ], + "spans": [ + { + "bbox": [ + 206, + 260, + 258, + 270 + ], + "type": "text", + "content": "Face " + }, + { + "bbox": [ + 206, + 260, + 258, + 270 + ], + "type": "inline_equation", + "content": "= 800" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 237, + 204, + 287, + 257 + ], + "blocks": [ + { + "bbox": [ + 237, + 204, + 287, + 257 + ], + "lines": [ + { + "bbox": [ + 237, + 204, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 237, + 204, + 287, + 257 + ], + "type": "image", + "image_path": "408052b671515aab1508f5b14d64a2c1813071d41d86b54ea99ef646359b075c.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 172, + 273, + 227, + 331 + ], + "blocks": [ + { + "bbox": [ + 172, + 273, + 227, + 331 + ], + "lines": [ + { + "bbox": [ + 172, + 273, + 227, + 331 + ], + "spans": [ + { + "bbox": [ + 172, + 273, + 227, + 331 + ], + "type": "image", + "image_path": "4a27f5d0996d0cb5a021242557e6974b7a87a06684499f3abb5b2ad0546e5c15.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 335, + 258, + 345 + ], + "lines": [ + { + "bbox": [ + 206, + 335, + 258, + 345 + ], + "spans": [ + { + "bbox": [ + 206, + 335, + 258, + 345 + ], + "type": "text", + "content": "Face " + }, + { + "bbox": [ + 206, + 335, + 258, + 345 + ], + "type": "inline_equation", + "content": "= 630" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 237, + 273, + 293, + 331 + ], + "blocks": [ + { + "bbox": [ + 237, + 273, + 293, + 331 + ], + "lines": [ + { + "bbox": [ + 237, + 273, + 293, + 331 + ], + "spans": [ + { + "bbox": [ + 237, + 273, + 293, + 331 + ], + "type": "image", + "image_path": "3921a78ceb787db33deb87fb9617e51dd7d15b1b72551e928c72e8b1da0e8698.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 176, + 351, + 225, + 407 + ], + "blocks": [ + { + "bbox": [ + 176, + 351, + 225, + 407 + ], + "lines": [ + { + "bbox": [ + 176, + 351, + 225, + 407 + ], + "spans": [ + { + "bbox": [ + 176, + 351, + 225, + 407 + ], + "type": "image", + "image_path": "78ce9f926c0d6d3ff5c88067d089687b45750a8084a5788a3e31e0e86656566a.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 410, + 258, + 420 + ], + "lines": [ + { + "bbox": [ + 206, + 410, + 258, + 420 + ], + "spans": [ + { + "bbox": [ + 206, + 410, + 258, + 420 + ], + "type": "text", + "content": "Face " + }, + { + "bbox": [ + 206, + 410, + 258, + 420 + ], + "type": "inline_equation", + "content": "= 576" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 242, + 352, + 280, + 407 + ], + "blocks": [ + { + "bbox": [ + 242, + 352, + 280, + 407 + ], + "lines": [ + { + "bbox": [ + 242, + 352, + 280, + 407 + ], + "spans": [ + { + "bbox": [ + 242, + 352, + 280, + 407 + ], + "type": "image", + "image_path": "c6a05f62242284713fe27ac42aa8e5dd8ac18acd2f89b1478c2c0cf3e2965b86.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 177, + 426, + 226, + 477 + ], + "blocks": [ + { + "bbox": [ + 177, + 426, + 226, + 477 + ], + "lines": [ + { + "bbox": [ + 177, + 426, + 226, + 477 + ], + "spans": [ + { + "bbox": [ + 177, + 426, + 226, + 477 + ], + "type": "image", + "image_path": "c77c5f3a535757ccbfacec013b057d3aeefa7cdfeaffb4ce5fc524afb95b409f.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 238, + 427, + 287, + 479 + ], + "blocks": [ + { + "bbox": [ + 238, + 427, + 287, + 479 + ], + "lines": [ + { + "bbox": [ + 238, + 427, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 238, + 427, + 287, + 479 + ], + "type": "image", + "image_path": "62515ae8f680c2b2fa841d7fc0dc922a4c90dabd6f9a1b763bc3ca746a523117.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 484, + 258, + 495 + ], + "lines": [ + { + "bbox": [ + 206, + 484, + 258, + 495 + ], + "spans": [ + { + "bbox": [ + 206, + 484, + 258, + 495 + ], + "type": "text", + "content": "Face " + }, + { + "bbox": [ + 206, + 484, + 258, + 495 + ], + "type": "inline_equation", + "content": "= 495" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 183, + 509, + 280, + 539 + ], + "blocks": [ + { + "bbox": [ + 183, + 509, + 280, + 539 + ], + "lines": [ + { + "bbox": [ + 183, + 509, + 280, + 539 + ], + "spans": [ + { + "bbox": [ + 183, + 509, + 280, + 539 + ], + "type": "image", + "image_path": "46c8e94e65892df848358516405217ac41d43bcfca18b52ce8eceea49a773746.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 196, + 578, + 270, + 636 + ], + "blocks": [ + { + "bbox": [ + 206, + 559, + 258, + 570 + ], + "lines": [ + { + "bbox": [ + 206, + 559, + 258, + 570 + ], + "spans": [ + { + "bbox": [ + 206, + 559, + 258, + 570 + ], + "type": "text", + "content": "Face " + }, + { + "bbox": [ + 206, + 559, + 258, + 570 + ], + "type": "inline_equation", + "content": "= 584" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 196, + 578, + 270, + 636 + ], + "lines": [ + { + "bbox": [ + 196, + 578, + 270, + 636 + ], + "spans": [ + { + "bbox": [ + 196, + 578, + 270, + 636 + ], + "type": "image", + "image_path": "c7eb6e95203a5c3ed03255e83c3f563fd433748f3b0ae434ea88b8447f8e3301.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 647, + 258, + 658 + ], + "lines": [ + { + "bbox": [ + 206, + 647, + 258, + 658 + ], + "spans": [ + { + "bbox": [ + 206, + 647, + 258, + 658 + ], + "type": "text", + "content": "Face " + }, + { + "bbox": [ + 206, + 647, + 258, + 658 + ], + "type": "inline_equation", + "content": "= 235" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 299, + 85, + 421, + 180 + ], + "blocks": [ + { + "bbox": [ + 313, + 64, + 405, + 77 + ], + "lines": [ + { + "bbox": [ + 313, + 64, + 405, + 77 + ], + "spans": [ + { + "bbox": [ + 313, + 64, + 405, + 77 + ], + "type": "text", + "content": "MeshAnythingV2 [4]" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 299, + 85, + 421, + 180 + ], + "lines": [ + { + "bbox": [ + 299, + 85, + 421, + 180 + ], + "spans": [ + { + "bbox": [ + 299, + 85, + 421, + 180 + ], + "type": "image", + "image_path": "adbfea0641ad810a1bdbd262f22ec2ad9fb77bea93f3e6a87b52db614d9bba50.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 301, + 198, + 415, + 258 + ], + "blocks": [ + { + "bbox": [ + 331, + 185, + 388, + 195 + ], + "lines": [ + { + "bbox": [ + 331, + 185, + 388, + 195 + ], + "spans": [ + { + "bbox": [ + 331, + 185, + 388, + 195 + ], + "type": "text", + "content": "Face = 2247" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 301, + 198, + 415, + 258 + ], + "lines": [ + { + "bbox": [ + 301, + 198, + 415, + 258 + ], + "spans": [ + { + "bbox": [ + 301, + 198, + 415, + 258 + ], + "type": "image", + "image_path": "9d45079fca32265b44c19920c3d1fd9bf9419954282437490a6c664cc563ec05.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 260, + 388, + 270 + ], + "lines": [ + { + "bbox": [ + 331, + 260, + 388, + 270 + ], + "spans": [ + { + "bbox": [ + 331, + 260, + 388, + 270 + ], + "type": "text", + "content": "Face = 2246" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 300, + 272, + 419, + 331 + ], + "blocks": [ + { + "bbox": [ + 300, + 272, + 419, + 331 + ], + "lines": [ + { + "bbox": [ + 300, + 272, + 419, + 331 + ], + "spans": [ + { + "bbox": [ + 300, + 272, + 419, + 331 + ], + "type": "image", + "image_path": "2bee1f01fd798c18bd50ec3a0cd1a854dd2d7c5e023e0425438350bf885dadca.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 305, + 347, + 407, + 389 + ], + "blocks": [ + { + "bbox": [ + 331, + 335, + 388, + 345 + ], + "lines": [ + { + "bbox": [ + 331, + 335, + 388, + 345 + ], + "spans": [ + { + "bbox": [ + 331, + 335, + 388, + 345 + ], + "type": "text", + "content": "Face = 2090" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 347, + 407, + 389 + ], + "lines": [ + { + "bbox": [ + 305, + 347, + 407, + 389 + ], + "spans": [ + { + "bbox": [ + 305, + 347, + 407, + 389 + ], + "type": "image", + "image_path": "11c60089c50c4d1a57340382ee4ed6dcd12bcc085ea646d9d7ea60fe961134a9.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 301, + 426, + 421, + 483 + ], + "blocks": [ + { + "bbox": [ + 331, + 410, + 388, + 420 + ], + "lines": [ + { + "bbox": [ + 331, + 410, + 388, + 420 + ], + "spans": [ + { + "bbox": [ + 331, + 410, + 388, + 420 + ], + "type": "text", + "content": "Face = 1998" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 301, + 426, + 421, + 483 + ], + "lines": [ + { + "bbox": [ + 301, + 426, + 421, + 483 + ], + "spans": [ + { + "bbox": [ + 301, + 426, + 421, + 483 + ], + "type": "image", + "image_path": "4ade773c361062907b496d276215643f34764126d0270e2d3c088b6f92518927.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 484, + 388, + 495 + ], + "lines": [ + { + "bbox": [ + 331, + 484, + 388, + 495 + ], + "spans": [ + { + "bbox": [ + 331, + 484, + 388, + 495 + ], + "type": "text", + "content": "Face = 2159" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 310, + 507, + 408, + 539 + ], + "blocks": [ + { + "bbox": [ + 310, + 507, + 408, + 539 + ], + "lines": [ + { + "bbox": [ + 310, + 507, + 408, + 539 + ], + "spans": [ + { + "bbox": [ + 310, + 507, + 408, + 539 + ], + "type": "image", + "image_path": "0896b0ac5011a8545b038d324d2b5f216f743fac963677c74cb704788e4c1bd7.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 323, + 578, + 397, + 636 + ], + "blocks": [ + { + "bbox": [ + 331, + 559, + 388, + 570 + ], + "lines": [ + { + "bbox": [ + 331, + 559, + 388, + 570 + ], + "spans": [ + { + "bbox": [ + 331, + 559, + 388, + 570 + ], + "type": "text", + "content": "Face " + }, + { + "bbox": [ + 331, + 559, + 388, + 570 + ], + "type": "inline_equation", + "content": "= 1769" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 578, + 397, + 636 + ], + "lines": [ + { + "bbox": [ + 323, + 578, + 397, + 636 + ], + "spans": [ + { + "bbox": [ + 323, + 578, + 397, + 636 + ], + "type": "image", + "image_path": "d6cf071ba12b87287e08c471bba3adec9d42bcfec457a6d53b6a1d952365da49.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 647, + 384, + 658 + ], + "lines": [ + { + "bbox": [ + 334, + 647, + 384, + 658 + ], + "spans": [ + { + "bbox": [ + 334, + 647, + 384, + 658 + ], + "type": "text", + "content": "Face " + }, + { + "bbox": [ + 334, + 647, + 384, + 658 + ], + "type": "inline_equation", + "content": "= 265" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 426, + 84, + 547, + 180 + ], + "blocks": [ + { + "bbox": [ + 475, + 65, + 499, + 76 + ], + "lines": [ + { + "bbox": [ + 475, + 65, + 499, + 76 + ], + "spans": [ + { + "bbox": [ + 475, + 65, + 499, + 76 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 426, + 84, + 547, + 180 + ], + "lines": [ + { + "bbox": [ + 426, + 84, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 426, + 84, + 547, + 180 + ], + "type": "image", + "image_path": "edce66028a2128e05a9952f5e8a0f0d42d04f1141d9d8826fb6fad70d91ee1ef.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 429, + 198, + 541, + 257 + ], + "blocks": [ + { + "bbox": [ + 459, + 185, + 514, + 195 + ], + "lines": [ + { + "bbox": [ + 459, + 185, + 514, + 195 + ], + "spans": [ + { + "bbox": [ + 459, + 185, + 514, + 195 + ], + "type": "text", + "content": "Face = 3311" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 429, + 198, + 541, + 257 + ], + "lines": [ + { + "bbox": [ + 429, + 198, + 541, + 257 + ], + "spans": [ + { + "bbox": [ + 429, + 198, + 541, + 257 + ], + "type": "image", + "image_path": "db2cf1a3cc461253e03ad5d69d010bf7187b96345e953a333c045db3f77a5267.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 427, + 272, + 482, + 331 + ], + "blocks": [ + { + "bbox": [ + 459, + 260, + 514, + 270 + ], + "lines": [ + { + "bbox": [ + 459, + 260, + 514, + 270 + ], + "spans": [ + { + "bbox": [ + 459, + 260, + 514, + 270 + ], + "type": "text", + "content": "Face = 4545" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 427, + 272, + 482, + 331 + ], + "lines": [ + { + "bbox": [ + 427, + 272, + 482, + 331 + ], + "spans": [ + { + "bbox": [ + 427, + 272, + 482, + 331 + ], + "type": "image", + "image_path": "1613048410d1a69a5fd9ff7f126338f8b42a3782338aef4fac345cf1293a56e3.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 335, + 514, + 345 + ], + "lines": [ + { + "bbox": [ + 459, + 335, + 514, + 345 + ], + "spans": [ + { + "bbox": [ + 459, + 335, + 514, + 345 + ], + "type": "text", + "content": "Face " + }, + { + "bbox": [ + 459, + 335, + 514, + 345 + ], + "type": "inline_equation", + "content": "= 5122" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_caption" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 492, + 274, + 547, + 331 + ], + "blocks": [ + { + "bbox": [ + 492, + 274, + 547, + 331 + ], + "lines": [ + { + "bbox": [ + 492, + 274, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 492, + 274, + 547, + 331 + ], + "type": "image", + "image_path": "aa3bcc8bc99211346acf57e6ba1a48d8de0e21a7a23ec8b10a68cc89127c6e2b.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 429, + 349, + 479, + 407 + ], + "blocks": [ + { + "bbox": [ + 429, + 349, + 479, + 407 + ], + "lines": [ + { + "bbox": [ + 429, + 349, + 479, + 407 + ], + "spans": [ + { + "bbox": [ + 429, + 349, + 479, + 407 + ], + "type": "image", + "image_path": "03154838e179a348065aadc23a79dc505b0253f6eb04521321b28730fb0dec22.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 410, + 514, + 419 + ], + "lines": [ + { + "bbox": [ + 459, + 410, + 514, + 419 + ], + "spans": [ + { + "bbox": [ + 459, + 410, + 514, + 419 + ], + "type": "text", + "content": "Face " + }, + { + "bbox": [ + 459, + 410, + 514, + 419 + ], + "type": "inline_equation", + "content": "= 2708" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_caption" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 496, + 349, + 537, + 407 + ], + "blocks": [ + { + "bbox": [ + 496, + 349, + 537, + 407 + ], + "lines": [ + { + "bbox": [ + 496, + 349, + 537, + 407 + ], + "spans": [ + { + "bbox": [ + 496, + 349, + 537, + 407 + ], + "type": "image", + "image_path": "8b6bb92ae561ded526374b44c20e523b4e594b887f3840aafbd810c930494c19.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + } + ], + "index": 59 + }, + { + "type": "image", + "bbox": [ + 431, + 426, + 481, + 482 + ], + "blocks": [ + { + "bbox": [ + 431, + 426, + 481, + 482 + ], + "lines": [ + { + "bbox": [ + 431, + 426, + 481, + 482 + ], + "spans": [ + { + "bbox": [ + 431, + 426, + 481, + 482 + ], + "type": "image", + "image_path": "4c3f35e035726190e280c1355f4cd14ead46d6f454a1454e79d88ad18890d727.jpg" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 484, + 514, + 495 + ], + "lines": [ + { + "bbox": [ + 459, + 484, + 514, + 495 + ], + "spans": [ + { + "bbox": [ + 459, + 484, + 514, + 495 + ], + "type": "text", + "content": "Face = 2646" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_caption" + } + ], + "index": 61 + }, + { + "type": "image", + "bbox": [ + 493, + 427, + 541, + 479 + ], + "blocks": [ + { + "bbox": [ + 493, + 427, + 541, + 479 + ], + "lines": [ + { + "bbox": [ + 493, + 427, + 541, + 479 + ], + "spans": [ + { + "bbox": [ + 493, + 427, + 541, + 479 + ], + "type": "image", + "image_path": "17d681719d8eddfdde4af39a245b17fdbf0e00074e5db8a7fc2eb1aedbdb286f.jpg" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_body" + } + ], + "index": 62 + }, + { + "type": "image", + "bbox": [ + 437, + 507, + 535, + 538 + ], + "blocks": [ + { + "bbox": [ + 437, + 507, + 535, + 538 + ], + "lines": [ + { + "bbox": [ + 437, + 507, + 535, + 538 + ], + "spans": [ + { + "bbox": [ + 437, + 507, + 535, + 538 + ], + "type": "image", + "image_path": "debb74a80c1c187d59811371cdfcf0ad3fffad7cfbbf46aafa008ff447466f70.jpg" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_body" + } + ], + "index": 64 + }, + { + "type": "image", + "bbox": [ + 449, + 578, + 524, + 636 + ], + "blocks": [ + { + "bbox": [ + 459, + 559, + 514, + 570 + ], + "lines": [ + { + "bbox": [ + 459, + 559, + 514, + 570 + ], + "spans": [ + { + "bbox": [ + 459, + 559, + 514, + 570 + ], + "type": "text", + "content": "Face = 2114" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 449, + 578, + 524, + 636 + ], + "lines": [ + { + "bbox": [ + 449, + 578, + 524, + 636 + ], + "spans": [ + { + "bbox": [ + 449, + 578, + 524, + 636 + ], + "type": "image", + "image_path": "53de8261f2468831ccacc32c9c31da56c31d904e5596dd946b3a643bd78d75d3.jpg" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 461, + 647, + 511, + 657 + ], + "lines": [ + { + "bbox": [ + 461, + 647, + 511, + 657 + ], + "spans": [ + { + "bbox": [ + 461, + 647, + 511, + 657 + ], + "type": "text", + "content": "Face " + }, + { + "bbox": [ + 461, + 647, + 511, + 657 + ], + "type": "inline_equation", + "content": "= 301" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_caption" + } + ], + "index": 66 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26614" + } + ] + } + ], + "index": 69 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 62, + 145, + 219 + ], + "blocks": [ + { + "bbox": [ + 82, + 62, + 145, + 219 + ], + "lines": [ + { + "bbox": [ + 82, + 62, + 145, + 219 + ], + "spans": [ + { + "bbox": [ + 82, + 62, + 145, + 219 + ], + "type": "image", + "image_path": "7ef7ecfda59a2d78b1dc6fda89929c0f541a3fade4774c3cdd0a9878cad79061.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 150, + 62, + 264, + 219 + ], + "blocks": [ + { + "bbox": [ + 150, + 62, + 264, + 219 + ], + "lines": [ + { + "bbox": [ + 150, + 62, + 264, + 219 + ], + "spans": [ + { + "bbox": [ + 150, + 62, + 264, + 219 + ], + "type": "image", + "image_path": "759d5d3706e288a6bc29e93e0b143ebfe0b554b0ee5ab77cec3689744394d944.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 222, + 408, + 234 + ], + "lines": [ + { + "bbox": [ + 201, + 222, + 408, + 234 + ], + "spans": [ + { + "bbox": [ + 201, + 222, + 408, + 234 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison on GSO dataset [8]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 273, + 62, + 392, + 219 + ], + "blocks": [ + { + "bbox": [ + 273, + 62, + 392, + 219 + ], + "lines": [ + { + "bbox": [ + 273, + 62, + 392, + 219 + ], + "spans": [ + { + "bbox": [ + 273, + 62, + 392, + 219 + ], + "type": "image", + "image_path": "402ebf7e1680da38c2dfc0cc1b3eec4c54b07d78efe7e0204a76f6e394c545d2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 399, + 62, + 518, + 219 + ], + "blocks": [ + { + "bbox": [ + 399, + 62, + 518, + 219 + ], + "lines": [ + { + "bbox": [ + 399, + 62, + 518, + 219 + ], + "spans": [ + { + "bbox": [ + 399, + 62, + 518, + 219 + ], + "type": "image", + "image_path": "89fec65e06197b931633e8e2984eb8d542e02e2d8eb467d736d5e994e9522daa.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 75, + 242, + 138, + 318 + ], + "blocks": [ + { + "bbox": [ + 75, + 242, + 138, + 318 + ], + "lines": [ + { + "bbox": [ + 75, + 242, + 138, + 318 + ], + "spans": [ + { + "bbox": [ + 75, + 242, + 138, + 318 + ], + "type": "image", + "image_path": "2503cf9c79ce30a9a12cade462fedff6a975b3748671614482bd2347b8411b57.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 323, + 296, + 357 + ], + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 357 + ], + "type": "text", + "content": "Figure 5. Comparison between the decimated mesh and our output. Our model is capable of generating meshes with the topology of those created by 3D artists." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 144, + 242, + 208, + 318 + ], + "blocks": [ + { + "bbox": [ + 144, + 242, + 208, + 318 + ], + "lines": [ + { + "bbox": [ + 144, + 242, + 208, + 318 + ], + "spans": [ + { + "bbox": [ + 144, + 242, + 208, + 318 + ], + "type": "image", + "image_path": "d5265daa069de701561f839cff25793d42131ccf1e7618922ac0cfe87524614f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 213, + 242, + 276, + 318 + ], + "blocks": [ + { + "bbox": [ + 213, + 242, + 276, + 318 + ], + "lines": [ + { + "bbox": [ + 213, + 242, + 276, + 318 + ], + "spans": [ + { + "bbox": [ + 213, + 242, + 276, + 318 + ], + "type": "image", + "image_path": "f7c2c6610f54db4a3924337477e542097778c882300ef782084d009517360b21.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 365, + 149, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 149, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 149, + 378 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 385, + 296, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 296, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 296, + 481 + ], + "type": "text", + "content": "Decoder MLP head: We compare TreeMeshGPT trained with MLP heads that predict the x-y-z coordinates simultaneously and our proposed hierarchical MLP. Predicting the coordinate simultaneously leads to challenging sampling that results in noisy and more incomplete meshes, as shown in Figure 6. Running evaluation on the 200 validation samples used in Table 1 with this head yields " + }, + { + "bbox": [ + 55, + 385, + 296, + 481 + ], + "type": "inline_equation", + "content": "\\mathrm{CD} = 0.0114" + }, + { + "bbox": [ + 55, + 385, + 296, + 481 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 385, + 296, + 481 + ], + "type": "inline_equation", + "content": "\\mathrm{NC} = 0.724" + }, + { + "bbox": [ + 55, + 385, + 296, + 481 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 385, + 296, + 481 + ], + "type": "inline_equation", + "content": "|\\mathrm{NC}| = 0.847" + }, + { + "bbox": [ + 55, + 385, + 296, + 481 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 76, + 487, + 135, + 564 + ], + "blocks": [ + { + "bbox": [ + 76, + 487, + 135, + 564 + ], + "lines": [ + { + "bbox": [ + 76, + 487, + 135, + 564 + ], + "spans": [ + { + "bbox": [ + 76, + 487, + 135, + 564 + ], + "type": "image", + "image_path": "709731a44bdd84b5af04b253a655d6faff6741ac1bd7b41d8df22e54918df39d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 568, + 295, + 602 + ], + "lines": [ + { + "bbox": [ + 55, + 568, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 295, + 602 + ], + "type": "text", + "content": "Figure 6. MLP head ablation. Our hierarchical MLP maintains the sequential nature of the x-y-z coordinates prediction that results in easier sampling compared to simultaenous prediction." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 145, + 487, + 204, + 563 + ], + "blocks": [ + { + "bbox": [ + 145, + 487, + 204, + 563 + ], + "lines": [ + { + "bbox": [ + 145, + 487, + 204, + 563 + ], + "spans": [ + { + "bbox": [ + 145, + 487, + 204, + 563 + ], + "type": "image", + "image_path": "8f8af4a4df6b9aae4cbca26166892ccfcde95dddbaac58acedce918f58fb5bd7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 214, + 487, + 272, + 564 + ], + "blocks": [ + { + "bbox": [ + 214, + 487, + 272, + 564 + ], + "lines": [ + { + "bbox": [ + 214, + 487, + 272, + 564 + ], + "spans": [ + { + "bbox": [ + 214, + 487, + 272, + 564 + ], + "type": "image", + "image_path": "4b7f5da83b96abae0953f1860cdcf2a984ebcb295915c37dc0da997ea1becc64.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "Tree traversal: We conduct an ablation study on smaller Transformer architecture comparing DFS and breadth-first-search (BFS) traversals in forming input-output sequences for our Autoregressive Tree Sequencing. As shown in the training perplexity plot in Figure 7, DFS traversal enables more efficient Transformer training. This improvement likely stems from the stronger local dependencies introduced by DFS, where each step is more predictable based on its immediate predecessors. In contrast, BFS traversal often" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 243, + 555, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 278 + ], + "type": "text", + "content": "introduces dependencies between steps that are spatially or structurally distant, and thus complicating the learning process." + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 317, + 281, + 550, + 443 + ], + "blocks": [ + { + "bbox": [ + 317, + 281, + 550, + 443 + ], + "lines": [ + { + "bbox": [ + 317, + 281, + 550, + 443 + ], + "spans": [ + { + "bbox": [ + 317, + 281, + 550, + 443 + ], + "type": "image", + "image_path": "015486ac48dca7b28335bdfe1e67404acfa897eda850796d899fde074531d4e0.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 445, + 555, + 487 + ], + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 487 + ], + "type": "text", + "content": "Figure 7. Training perplexity comparison Between BFS and DFS traversals. DFS traversal shows a better training perplexity compared to BFS. Shown in the plot is the perplexity for " + }, + { + "bbox": [ + 313, + 445, + 555, + 487 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 445, + 555, + 487 + ], + "type": "text", + "content": " -axis vertex coordinate." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 498, + 388, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 498, + 388, + 510 + ], + "spans": [ + { + "bbox": [ + 314, + 498, + 388, + 510 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 519, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 555, + 639 + ], + "type": "text", + "content": "We present TreeMeshGPT, an autoregressive Transformer designed to generate high-quality artistic meshes aligned with input point clouds. TreeMeshGPT incorporates a novel Autoregressive Tree Sequencing technique instead of the conventional next-token prediction. Our Autoregressive Tree Sequencing represents each face with two tokens, enabling a 7-bit discretization model that can generate up to 5,500 triangular faces with 2,048 point cloud latent tokens. Experiments show that TreeMeshGPT can generate meshes with higher quality compared to the previous methods." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "Limitations: Our model has a similar failure mode to the previous methods that the success rate decreases as the sequence length increases. Additionally, while our model has an improved capacity to generate meshes with higher face counts, challenges persist in enforcing an optimal mesh topology." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26615" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "content": "Acknowledgment. This research is supported by the National Research Foundation (NRF) Singapore, under its NRF-Investigatorship Programme (Award ID. NRF-NRFI09-0008)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 130, + 115, + 142 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 115, + 142 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 115, + 142 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 149, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 149, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 149, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 61, + 149, + 294, + 194 + ], + "type": "text", + "content": "[1] Fausto Bernardini, Joshua Mittleman, Holly Rushmeier, Cláudio Silva, and Gabriel Taubin. The ball-pivoting algorithm for surface reconstruction. IEEE Transactions on Visualization and Computer Graphics, 1999. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 295, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 248 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 248 + ], + "type": "text", + "content": "[2] Sijin Chen, Xin Chen, Anqi Pang, Xianfang Zeng, Wei Cheng, Yijun Fu, Fukun Yin, Yanru Wang, Zhibin Wang, Chi Zhang, et al. Meshxl: Neural coordinate field for generative 3d foundation models. Advances in Neural Information Processing Systems (NeurIPS), 2024. 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 250, + 295, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 250, + 295, + 303 + ], + "spans": [ + { + "bbox": [ + 62, + 250, + 295, + 303 + ], + "type": "text", + "content": "[3] Yiwen Chen, Tong He, Di Huang, Weicai Ye, Sijin Chen, Ji-axiang Tang, Xin Chen, Zhongang Cai, Lei Yang, Gang Yu, et al. Meshanything: Artist-created mesh generation with autoregressive transformers. arXiv preprint arXiv:2406.10163, 2024. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 304, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 304, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 62, + 304, + 294, + 358 + ], + "type": "text", + "content": "[4] Yiwen Chen, Yikai Wang, Yihao Luo, Zhengyi Wang, Zilong Chen, Jun Zhu, Chi Zhang, and Guosheng Lin. Meshanything v2: Artist-created mesh generation with adjacent mesh tokenization. arXiv preprint arXiv:2408.02555, 2024. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 360, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 360, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 62, + 360, + 294, + 414 + ], + "type": "text", + "content": "[5] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 415, + 294, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 294, + 469 + ], + "type": "text", + "content": "[6] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of " + }, + { + "bbox": [ + 62, + 415, + 294, + 469 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{d}" + }, + { + "bbox": [ + 62, + 415, + 294, + 469 + ], + "type": "text", + "content": " objects. Advances in Neural Information Processing Systems (NeurIPS), 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 471, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 471, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 62, + 471, + 294, + 535 + ], + "type": "text", + "content": "[7] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objaverse: A universe of annotated 3d objects. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 5, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 537, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 537, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 537, + 294, + 601 + ], + "type": "text", + "content": "[8] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2022. 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 603, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 603, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 603, + 294, + 635 + ], + "type": "text", + "content": "[9] Zekun Hao, David W Romero, Tsung-Yi Lin, and Ming-Yu Liu. Meshtron: High-fidelity, artist-like 3d mesh generation at scale. arXiv preprint arXiv:2412.09548, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 636, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 294, + 689 + ], + "type": "text", + "content": "[10] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. In International Conference on Learning Representations (ICLR), 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "text", + "content": "[11] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object gen" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "text", + "content": "eration with dream fields. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 554, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 554, + 139 + ], + "type": "text", + "content": "[12] Tao Ju, Frank Losasso, Scott Schaefer, and Joe Warren. Dual contouring of hermite data. In Proceedings of the 29th Annual Conference on Computer graphics and Interactive Techniques, 2002. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 140, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 172 + ], + "type": "text", + "content": "[13] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the 4th Eurographics Symposium on Geometry Processing, 2006. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "type": "text", + "content": "[14] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 206, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 237 + ], + "type": "text", + "content": "[15] Kiriakos N Kutulakos and Steven M Seitz. A theory of shape by space carving. International Journal of Computer Vision, 2000. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "text", + "content": "[16] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 284, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 337 + ], + "type": "text", + "content": "[17] Chenghao Li, Chaoning Zhang, Joseph Cho, Atish Waghwase, Lik-Hang Lee, Francois Rameau, Yang Yang, SungHo Bae, and Choong Seon Hong. Generative ai meets 3d: A survey on text-to-3d in aigc era. arXiv preprint arXiv:2305.06131, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 338, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 553, + 403 + ], + "type": "text", + "content": "[18] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. In International Conference on Learning Representations (ICLR), 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 404, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 437 + ], + "type": "text", + "content": "[19] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM SIGGRAPH Computer Graphics, 1987. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 438, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 470 + ], + "type": "text", + "content": "[20] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "type": "text", + "content": "[21] Oscar Michel, Roi Bar-On, Richard Liu, Sage Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 515, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 558 + ], + "type": "text", + "content": "[22] Charlie Nash, Yaroslav Ganin, SM Ali Eslami, and Peter Battaglia. *Polygen: An autoregressive generative model of 3d meshes*. In International Conference on Machine Learning (ICML), 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 559, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 601 + ], + "type": "text", + "content": "[23] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 603, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 624 + ], + "type": "text", + "content": "[24] Gregory M Nielson. Dual marching cubes. In IEEE Visualization, 2004. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 625, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 667 + ], + "type": "text", + "content": "[25] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In International Conference on Learning Representations (ICLR), 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26616" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 675 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[27] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. In International Conference on Learning Representations (ICLR), 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 183 + ], + "type": "text", + "content": "[28] Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "text", + "content": "[29] Jiaxiang Tang, Zhaoshuo Li, Zekun Hao, Xian Liu, Gang Zeng, Ming-Yu Liu, and Qinsheng Zhang. Edgerunner: Auto-regressive auto-encoder for artistic mesh generation. arXiv preprint arXiv:2409.18114, 2024. 2, 3, 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "type": "text", + "content": "[30] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in Neural Information Processing Systems (NeurIPS), 2017. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 307 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 307 + ], + "type": "text", + "content": "[31] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems (NeurIPS), 2017. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 373 + ], + "type": "text", + "content": "[32] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 375, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 375, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 375, + 294, + 418 + ], + "type": "text", + "content": "[33] Haohan Weng, Yikai Wang, Tong Zhang, CL Chen, and Jun Zhu. Pivotmesh: Generic 3d mesh generation via pivot vertices guidance. arXiv preprint arXiv:2405.16890, 2024. 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 420, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 420, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 420, + 294, + 464 + ], + "type": "text", + "content": "[34] Haohan Weng, Zibo Zhao, Biwen Lei, Xianghui Yang, Jian Liu, Zeqiang Lai, Zhuo Chen, Yuhong Liu, Jie Jiang, Chunchao Guo, et al. Scaling mesh generation via compressive tokenization. arXiv preprint arXiv:2411.07025, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 465, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 465, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 465, + 294, + 518 + ], + "type": "text", + "content": "[35] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 520, + 294, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 294, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 294, + 574 + ], + "type": "text", + "content": "[36] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 576, + 294, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 576, + 294, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 576, + 294, + 620 + ], + "type": "text", + "content": "[37] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Transactions on Graphics (TOG), 2023. 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 621, + 294, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 621, + 294, + 675 + ], + "spans": [ + { + "bbox": [ + 56, + 621, + 294, + 675 + ], + "type": "text", + "content": "[38] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 2024. 3, 5" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26617" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TriTex_ Learning Texture from a Single Mesh via Triplane Semantic Features/f0b8cf58-1685-4de9-9df0-451922c97e91_content_list.json b/2025/TriTex_ Learning Texture from a Single Mesh via Triplane Semantic Features/f0b8cf58-1685-4de9-9df0-451922c97e91_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8524de71ea345e984c0585fc379482b10f5dcea7 --- /dev/null +++ b/2025/TriTex_ Learning Texture from a Single Mesh via Triplane Semantic Features/f0b8cf58-1685-4de9-9df0-451922c97e91_content_list.json @@ -0,0 +1,1719 @@ +[ + { + "type": "text", + "text": "TriTex: Learning Texture from a Single Mesh via Triplane Semantic Features", + "text_level": 1, + "bbox": [ + 109, + 130, + 888, + 152 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/746f54f9680bfb83d22012db1d92199bd85a1d28f9829e6c65e8024bd14d74b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 107, + 227, + 344, + 396 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4ebda9dd5a6b95cd194b2fd276af872402e33dad9782ae3bcba8f76c5e0c9562.jpg", + "image_caption": [ + "Dana Cohen-Bar1,2", + "Daniel Cohen-Or1 \n1Tel Aviv University" + ], + "image_footnote": [], + "bbox": [ + 359, + 224, + 493, + 387 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/558702c05ec899e19f121491cd4ee1aadc6f988f36124a723b46de325b730729.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 229, + 689, + 401 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/835465995a9e2f432a28c969237005ad46254bc663548612cdf2a03e9f69c751.jpg", + "image_caption": [ + "Gal Chechik2 Yoni Kasten2 NVIDIA", + "Figure 1. TriTex is a method for learning to transfer a texture from a single mesh in a feed-forward manner to other input meshes. The figure shows four examples. In each, given a target geometry (up left) and a source textured mesh (up right) the texture is transferred to the target 3D object (bottom two views of the target object)." + ], + "image_footnote": [], + "bbox": [ + 730, + 222, + 903, + 406 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 473, + 326, + 489 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As 3D content creation continues to grow, transferring semantic textures between 3D meshes remains a significant challenge in computer graphics. While recent methods leverage text-to-image diffusion models for texturing, they often struggle to preserve the appearance of the source texture during texture transfer. We present TRiTEX, a novel approach that learns a volumetric texture field from a single textured mesh by mapping semantic features to surface colors. Using an efficient triplane-based architecture, our method enables semantic-aware texture transfer to a novel target mesh. Despite training on just one example, it generalizes effectively to diverse shapes within the same category. Extensive evaluation on our newly created benchmark dataset shows that TRiTEX achieves superior texture transfer quality and fast inference times compared to existing methods. Our approach advances single-example texture transfer, providing a practical solution for maintaining visual coherence across related 3D models in applications like game development and simulation.", + "bbox": [ + 89, + 507, + 485, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Project page: https://danacohen95.github.io/TriTex/.", + "bbox": [ + 89, + 794, + 437, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 843, + 220, + 858 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Texturing 3D objects is a fundamental task with wideranging applications in game development, simulation, and", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "video production. For example, when generating 3D scenes, style needs to be consistently applied across multiple 3D models that share semantic properties but differ in shape, such as in environments featuring diverse buildings or plants. Improving texture transfer can streamline texturing workflows by maintaining visual coherence across related models and preserving texture details even during mesh modifications, ensuring a consistent appearance across all elements. With the advent of large generative models, new approaches to 3D texturing have emerged, moving beyond traditional procedural methods to enable automatic texturing and texture transfer [13, 45]. However, these techniques often require the availability of large databases of 3D objects and extensive training for each class to achieve high-quality results.", + "bbox": [ + 511, + 474, + 906, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent texturing methods use text-to-image diffusion models to texture meshes without requiring 3D data collection. Some methods [32, 57, 60] apply an SDS loss to optimize the texture map, which tends to be a slow process. Others [5, 8, 9, 25, 28, 40] employ iterative depth-conditioned image inpainting on rendered mesh views, synchronizing across views to maintain consistency. However, when performing texture transfer, these methods struggle to faithfully preserve the appearance of the source texture.", + "bbox": [ + 511, + 702, + 908, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce TRITEx, a novel technique that learns a volumetric texture field from a single textured mesh, mapping its semantic features to RGB colors on its surface. During inference, given a target mesh, this learned field en", + "bbox": [ + 511, + 839, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21403", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ables texture transfer by predicting colors based on the target mesh's semantic features, allowing the source texture to be applied in a way that maintains semantic correspondences between the source and target meshes (see Figure 1). Despite being trained on a single object, our trained volumetric texture generalizes well to novel objects within the same category. Our fast inference time enables the stylization of large scenes More specifically, to learn this texture field, we first extract semantic features using a frozen Diff3F [18] model, then project these features onto a triplane, and process them through a mapping network to produce the final triplane features.", + "bbox": [ + 89, + 90, + 480, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our experiments demonstrate the effectiveness of our texture transfer method, showing superior quality compared to previous approaches while maintaining fast inference times. We evaluate our approach using automated metrics and human evaluation on a benchmark dataset that we created specifically for the texture transfer task. Our method shows superior performance compared to previous approaches.", + "bbox": [ + 89, + 272, + 480, + 391 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our contributions are: (1) Introducing a method for transferring textures from a single textured mesh to new shapes while preserving semantic relations. (2) Leveraging pre-trained 3D semantic features, reprojecting them into a triplane representation, and enabling effective processing through convolutional layers. (3) Demonstrating strong generalization to novel objects with significant shape variations, despite being trained on only a single example. (4) Achieving high-quality texture transfer with fast inference, outperforming previous methods in both speed and quality.", + "bbox": [ + 89, + 393, + 480, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 570, + 232, + 585 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Texture Synthesis Classical methods use sampling approaches to generate 2D textures [14, 19, 23, 51, 63]. More recent approaches rely on deep learning to create textures based on training data [54, 62]. Texturify [45] introduced a Generative Adversarial Network (GAN) for texturing 3D models using an example image, applying face convolutional operators directly on the 3D object's surface. Auv-net [13] predicts UV mapping and texture images from 3D geometry, ensuring consistent alignment across elements within the same category, which enables effective training of generative texture models with this embedding. Both Auv-net and Texturify require large datasets of predefined categories to train their generative models, whereas our approach is trained solely on a single textured mesh and can generalize to a variety of objects with similar semantics.", + "bbox": [ + 89, + 595, + 482, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Leveraging Large Vision-Language Pretrained Models for Texture Generation Recent advances in image-language foundation models have enabled effective texture transfer and synthesis without requiring additional training", + "bbox": [ + 89, + 839, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "data. Text2Mesh [33] and TANGO [12] edit 3D mesh appearance by aligning rendered images with a text prompt in CLIP space. Dream3D [55] optimizes the appearance of generated geometry using a Neural Radiance Field (NeRF) by applying a CLIP loss. Latent-NeRF [32] applies the SDS-loss [38] to rendered images to optimize a NeRF representation in the stable diffusion latent space. More recently, Paint-it [59], Fantasia3D [10], and FlashTex [17] have enhanced SDS-loss-based texturing by incorporating Physically Based Rendering (PBR), BRDF modeling, and illumination control, respectively. These optimization techniques require processing each object individually, leading to slower performance, whereas our method has a much shorter inference time.", + "bbox": [ + 511, + 90, + 903, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Other methods employed optimization-free approaches by directly applying depth-conditioned diffusion models in sequence. Texture [40] applies depth-conditioned inpainting on sequential mesh projections to achieve mesh texturing. Concurrent and follow-up methods [5, 8, 9, 25, 28] introduced improvements, with Text2Tex[8] proposing an automatic per-object view selection scheme, and TexFusion [5], SyncMVD [28], SyncTweedies [25] and MVEdit [9] synchronizing views by operating on intermediate denoising steps. Paint3D [60] additionally trains a model to remove illumination effects from the generated texture map, enabling relighting of the textured meshes. In [3], a geometry-aware multiview diffusion model is used to enhance view consistency and efficiency.", + "bbox": [ + 511, + 303, + 903, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Texture Transfer The problem of texture transfer has always been an area of interest and research, with early work aiming to transfer both stochastic texture [30] and semantic texture [11, 31]. More recently, several methods leverage 3D data for texture transfer tasks. 3DStyleNet [58] transforms a source mesh by applying part-aware low-frequency deformations and generating texture maps, using a target mesh as a style reference. Mesh2Tex [4] learns a texture manifold from a large dataset of 3D objects and images. This enables new images to be projected into the manifold and transferred to 3D objects in trained categories, such as cars and chairs. Personalization techniques for diffusion models [21, 42] are used in methods like TEXTure and TextureDreamer [40, 57] to texture a mesh from a few input images, which can be either natural images or rendered images from another textured mesh. MVEdit [9], EASI-Tex [37] and Paint3D [60] transfer textures from an image to a 3D object using IP-Adapter [56], where EASI-Tex further enhances geometry details by incorporating a rendered mesh edge map as an additional control for image generation. While these methods are efficient, they often fail to faithfully preserve the appearance of the source texture. In contrast, our method maintains both efficiency and accurate texture preservation.", + "bbox": [ + 511, + 537, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21404", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/39d8cb70cb3987e700167c4773a278db52a6d35e48a246a4f33d18ea13ff35b0.jpg", + "image_caption": [ + "Figure 2. Training Pipeline (Top). Given an input textured mesh and its pre-extracted Diff3F features, we project six orthographic views to create an initial triplane. This triplane is processed using triplane-aware convolutional blocks, which, along with the texture MLP, define a coloring neural field. This field, together with the input geometry, is used to render the colored mesh. Appearance losses are applied between the true mesh appearance and the rendered appearance. Inference (Bottom). Given a new mesh (left), our pre-trained model maps its semantic properties to colors (right), transferring the texture from the original textured mesh learned during the training phase." + ], + "image_footnote": [], + "bbox": [ + 106, + 85, + 883, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Auv-net [13] enables texture transfer between meshes through a learned UV mapping and aligned texture representations. Mitchel et al. [34] proposed a diffusion process within a learned latent space on the surface, allowing the transferring of textures to other meshes using an input semantic label map. Nerf Analogies [20] extends texture transfer to neural radiance fields, enabling appearance transfer between source and target NeRFs with semantic similarities, though it requires per-pair training for each set of NeRFs.", + "bbox": [ + 88, + 512, + 486, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Semantic Features and Correspondences Finding corresponding points between images has traditionally relied on local feature descriptors [29] that are invariant to lighting and color. Later methods used features extracted from pretrained classification models [1, 46], providing invariance to local shape deformations and leveraging higher-level semantic relationships. Recently, Amir et al. [2] demonstrated the effectiveness of features from the pre-trained Vision Transformer DINO-ViT [6] as semantic descriptors. Splice [49] uses DINO-ViT features to transfer colors from a source image to a target image structure by preserving the target's local DINO features while adapting its global appearance to match the source. Neural Congealing [36] employs DINO-ViT semantic features to map a set of se", + "bbox": [ + 88, + 688, + 486, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mentically related images into a shared canonical representation, enabling zero-shot joint editing. More recently, DIFT [47] introduced semantic features extracted from a pre-trained diffusion model during denoising. Diff3F [18] mapped both DIFT and DINO features onto a 3D mesh, resulting in 3D semantic features that enable shape correspondences. Our model benefits from these recent advances in semantic features and uses them to train a texture transfer model with a single textured mesh.", + "bbox": [ + 511, + 512, + 908, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Models Trained on a Single Instance Several methods have been developed to train neural networks using a single example. SinGAN [43] trains a model to capture the internal distribution of patches within a single image, enabling the generation of diverse samples with similar visual content. Similarly, Wu et al. [52] and Hertz et al. [24] use GANs to generate shape and geometric texture variations, respectively, from a single mesh. More recent works [26, 35, 50] have employed diffusion models trained on individual images to produce image variations, with SinFusion [35] also extending this approach to single videos to create video variations. Additionally, Sin3DM [53] and [34] apply similar techniques to generate variations of a single textured mesh, either for textured mesh variations [53] or for texture variations [34]. Similarly, we propose training a", + "bbox": [ + 511, + 672, + 910, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21405", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mesh texturing model using just a single textured mesh.", + "bbox": [ + 89, + 90, + 460, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Triplane Representation The triplane representation introduced in EG3D [7] consists of three 2D feature grids aligned with three distinct orthogonal planes in the 3D coordinate system. This setup enables image operations, such as convolutional neural layers, and uses interpolation to define a continuous feature grid suitable for neural fields. Triplane representation has been applied in 3D-aware GANs [7] and in mesh generation using diffusion models [22, 44]. For a detailed technical description, see Sec. 3.1.", + "bbox": [ + 89, + 125, + 483, + 261 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 275, + 181, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Background", + "text_level": 1, + "bbox": [ + 89, + 300, + 220, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Triplane Representation The triplane representation, introduced in EG3D [7], encodes 3D information using three axis-aligned orthogonal feature planes $F_{XY}, F_{XZ}, F_{YZ}$ , each of size $\\mathbb{R}^{W \\times H \\times C}$ , where $W$ and $H$ represent the spatial resolution, and $C$ is the number of channels. To query features at any 3D position $x \\in \\mathbb{R}^3$ , the coordinates are used to sample from each of the three feature planes (XY, XZ, YZ) via bilinear interpolation, yielding three feature vectors $f_{xy}, f_{xz}, f_{yz} \\in \\mathbb{R}^C$ . These feature vectors are then concatenated and passed through a lightweight MLP decoder network to produce the final output. This approach enables efficient feature extraction while maintaining expressiveness through learned feature integration, offering a balance between efficiency and quality.", + "bbox": [ + 89, + 321, + 483, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Importantly, the triplane representation enables processing with standard 2D convolutional layers, which are far more efficient than using 3D convolutional layers on volumetric grids. In EG3D [7], 2D convolutional layers were applied on the three concatenated feature planes. Recently, Sin3DM [53] introduced the triplane-aware convolution block, suggesting a more geometrically integrated approach. Instead of processing the three feature grids $(F_{XY}, F_{XZ}, F_{YZ})$ independently or simply concatenating them, this method aggregates features from each plane into the others. Let $F_{XY}'$ , $F_{XZ}'$ , $F_{YZ}'$ represent the output of a single independent convolutional layer. Each plane is averaged over its axes and replicated along the third axis, before being concatenated into the relevant plane. For example, $F_{XY}'$ is averaged over the $x$ and $y$ axes, producing $F_{Y}'$ and $F_{X}'$ respectively. These are then replicated along the $z$ -axis and concatenated with $F_{YZ}'$ and $F_{XZ}'$ , respectively.", + "bbox": [ + 89, + 534, + 483, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diff3f Semantic Features Our method leverages DIff3F [18] features, which provide robust semantic descriptors for 3D shapes without requiring textured inputs or additional training. DIff3F extracts semantic features by utilizing foundational vision models in a zero-shot manner. The process begins by rendering depth and normal maps", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "from multiple views of the input mesh. These maps serve as conditioning inputs to ControlNet [61] and Stable Diffusion [41] for generating view-dependent features. The features from the diffusion process maintain semantic consistency despite potential visual variations across views, and are aggregated onto the original 3D surface. These features, further enriched with DINO features [6], capture rich semantic information and enable reliable correspondence across shapes with significant geometric variations, making them well-suited for our texture transfer task.", + "bbox": [ + 511, + 90, + 906, + 242 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/191c13ec86289e92de95d4feaf513168b7df132b17e3ccb6a116e817768a7e16.jpg", + "image_caption": [ + "Figure 3. Texture Transfer Results. Given a source mesh (left column), our network transfers its appearance onto target meshes of varying shapes (three right columns). Each row demonstrates texture transfer using a different source mesh." + ], + "image_footnote": [], + "bbox": [ + 519, + 250, + 910, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Our approach: TRiTEX", + "text_level": 1, + "bbox": [ + 511, + 575, + 725, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a textured source mesh, our goal is to train a feedforward texturing function that can transfer that texture to any target mesh while preserving the semantics of the texture. Our pipeline is shown in Fig. 2.", + "bbox": [ + 511, + 598, + 905, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The architecture, based on a deep neural network, processes a 3D mesh with pre-extracted semantic features and a query 3D point, and outputs the corresponding color for the query point. Formally, we define a learnable function: $f(M,\\mathbb{R}^3)\\to [0,1]^3$ , where $M = (V,S,E,F)$ is the input mesh, with $V = \\{v_{1},\\ldots ,v_{n}\\mid v_{i}\\in \\mathbb{R}^{3}\\}$ and $S = \\{s_1,\\dots ,s_n\\mid s_i\\in \\mathbb{R}^D\\}$ , representing the 3D vertices and their corresponding semantic features, pre-extracted by [18]. Here, $E$ and $F$ represent the edges and faces, respectively, defined on the vertices.", + "bbox": [ + 511, + 659, + 905, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The first part of our architecture processes $M$ by orthog.raphically projecting the semantic features into an axisaligned feature triplane $\\mathcal{T}\\in \\mathbb{R}^{3\\times W\\times H\\times 2D}$ , where each feature plane is created by concatenating features from two orthographic projections taken from opposite directions. To address the low feature resolution $(32\\times 32)$ and enable fine", + "bbox": [ + 511, + 810, + 905, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21406", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "detail generation, we incorporate positional encoding into the input. Triplane-aware convolutional 2D blocks [53] (see further details in Sec.3.1) are then applied to the triplane to produce a modified triplane $\\mathcal{T}'\\in \\mathbb{R}^{3\\times W\\times H\\times D'}$ . Finally, to predict the color for a query 3D point, we sample features from the processed planes in $\\mathcal{T}'$ according to the query point's location, concatenate them to form a single feature vector, and pass this vector through the coloring MLP $c:\\mathbb{R}^{3D^{\\prime}}\\to [0,1]^{3}$ , which maps it to the final RGB color for the point.", + "bbox": [ + 89, + 90, + 483, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training We train our model on a single textured mesh using a rendering-based reconstruction loss, comparing rendered views of our predicted textures with ground truth views of the source mesh from randomly sampled camera angles. Let $I_R(\\theta)$ be an image rendered by our pipeline from angle $\\theta$ , with the corresponding colors predicted by the MLP $c$ , generated by querying 3D points on the mesh through intersecting rendered camera rays. This image is conditioned only on the input semantic features defined for the geometry and depends on the differentiable parameters of the triplane-aware convolutional layers and $c$ . To train the learnable parameters, we use the ground truth rendered images, which are known for the single training textured mesh, and apply the MSE loss on the image pixels:", + "bbox": [ + 89, + 285, + 483, + 500 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M S E}} (\\theta) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left\\| I _ {R} (\\theta) _ {i} - I _ {\\mathrm {G T}} (\\theta) _ {i} \\right\\| ^ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 516, + 424, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $I_{\\mathrm{GT}}(\\theta)$ is the ground truth image, and $i$ indexes the pixels in the image. While MSE loss encourages pixel-level accuracy, it may not effectively capture high-level perceptual details. To address this, we incorporate a perceptual loss $\\mathcal{L}_{\\mathrm{app}}$ from Tumanyan et al. [48], which emphasizes high-level semantic features. This improves the alignment of these features between the generated and ground truth images, enhancing texture realism. Our final loss function is then:", + "bbox": [ + 89, + 578, + 483, + 715 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {\\theta} \\left[ \\mathcal {L} _ {\\mathrm {M S E}} (\\theta) + \\delta_ {\\mathrm {a p p}} \\mathcal {L} _ {\\mathrm {a p p}} (\\theta) \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 728, + 400, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbb{E}_{\\theta}$ represents the expectation over all sampled camera angles, and $\\delta_{\\mathrm{app}}$ controls the relative weight of $\\mathcal{L}_{\\mathrm{app}}$ compared to $\\mathcal{L}_{\\mathrm{MSE}}$ . To improve generalization, we apply two levels of augmentation: (1) preprocessing augmentation, where simple 3D transformations are applied to the input mesh and features are extracted for each variant to enrich the learned feature distribution, and (2) training-time augmentation, where translation, scaling, and small rotational perturbations are applied to the mesh during training.", + "bbox": [ + 89, + 763, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/020e986b90bdbcc1eea4a305809a10672b4f170b7f36fb5b960a3b57b2ba9860.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 108, + 633, + 148 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e807f3b405a93ee120320bbac904a13387831f6a142a74910c5740eab62cc84a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 157, + 593, + 237 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cccc49c311f795cc803020a539811c63b394159f1e159245f2124737ae4b8254.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 243, + 627, + 305 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a1bec29323051dd3f141ceafe8c8529f069e52c3b88888bb50396437e02e74d8.jpg", + "image_caption": [ + "Source-textured" + ], + "image_footnote": [], + "bbox": [ + 529, + 330, + 627, + 368 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7bc5a9a3027125bf9343382e026f7b073a24587a5c2316e40321db4463efa9ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 109, + 754, + 147 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7e75364c148897b62396bbbfe168a1a1f1d8809671d04fe86c66c937121c5cb0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 157, + 720, + 233 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c3b13a4ad5e2c3e58d86e91dfb43c30b2f1c3d3f7d5300d77a90dd68ac7fb652.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 241, + 754, + 306 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/02f9d169cf28215295763e0e9ab51e5f16335e522622192b9e2ab4a5337797c5.jpg", + "image_caption": [ + "Target geometry" + ], + "image_footnote": [], + "bbox": [ + 653, + 321, + 751, + 375 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/731d2c7d06f75c86e93e7c515096ab7fda78cace3d705842b135e61449eab528.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 777, + 109, + 880, + 147 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dac90a887d3e9e92046f2a95ac3bce08d3cd19c4b260c6a63ff5a90548bc820e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 818, + 157, + 846, + 233 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c1ae6ced5be668c918f136fdba68d1e2334d4e1a9658b652fffee60aaf7955b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 781, + 241, + 880, + 306 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c63d101a1cf046cc225f3c1b446ec5a2e3cf14bc7bd031e4eeb8222c5e414e9d.jpg", + "image_caption": [ + "Target-textured", + "Figure 4. Additional Qualitative Results. Showing the target geometry and the texture transfer." + ], + "image_footnote": [], + "bbox": [ + 794, + 323, + 890, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments And Results", + "text_level": 1, + "bbox": [ + 511, + 458, + 751, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Details", + "text_level": 1, + "bbox": [ + 511, + 483, + 697, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset: Our experimental dataset consists of 52 3D meshes across 9 categories, curated from Objaverse [15, 16], a large-scale 3D object database. We first generated descriptive captions for each 3D object to identify semantically related objects using BLIP [27]. We then used keyword filtering on these captions to group objects into similar categories. The final selection includes only objects that both share semantic similarities and maintain consistent orientations within their respective categories. Overall, we have 256 different pairs of source and target objects. The full list of source and target objects is provided in the supplemental.", + "bbox": [ + 511, + 506, + 906, + 689 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baseline Methods: We compare our method against three state-of-the-art optimization-free approaches for visually-guided mesh texturing. (1) TEXTure [40], a personalization-based approach, enabling texture transfer from a source mesh to a target mesh. (2) EASI-TEX [37] and (3) MVEdit [9] use IP-Adapter to incorporate image guidance into their texture generation process, allowing direct control over the generated textures through reference images.", + "bbox": [ + 511, + 710, + 908, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics: We evaluate our method using human evaluation and two automated metrics: CLIP similarity score [39]", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21407", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8250877c7d10b6f0cb958cd8a36e759fcf0d57a27983e94c0ca0f5134112356d.jpg", + "image_caption": [ + "Figure 5. Qualitative Comparisons with baselines. A comparison between TRiTEX (right-most column) and state-of-the-art approaches for texture transfer. As observed, TRiTEX produces much more plausible texture transfer results." + ], + "image_footnote": [], + "bbox": [ + 166, + 85, + 826, + 502 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6943373119d693f1e0313c710b9a828476d4597d1f6c2c5a0dbb5ad016298d86.jpg", + "table_caption": [ + "Table 1. Quantitative Evaluation. Quantitative comparison of our method with baseline methods. We evaluate two appearance metrics and running time, showing that our method best preserves the source texture while achieving competitive speed." + ], + "table_footnote": [], + "table_body": "
MethodSIFID ↓CLIP sim. ↑Inference Runtime
TEXTure0.340.845 min.
MVEdit0.380.841 min.
EASI-TEX0.290.8515 min.
TriTex (ours)0.220.871 min.
", + "bbox": [ + 91, + 637, + 478, + 723 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and Single Image Frechet Inception Distance (SIFID). The CLIP score, adopted following TextureDreamer [57], measures the semantic and visual similarity between source and generated textures, indicating how well our method preserves the desired appearance. SIFID was originally proposed by [43] to measure internal patch statistics within single images. Following Sin3DM [53], we adopt this metric to evaluate texture similarity between source and target shapes.", + "bbox": [ + 89, + 748, + 482, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For both metrics, we render each source and target mesh", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "from the same 10 fixed viewpoints and compare the corresponding views against each other. To ensure balanced evaluation across our dataset, we first compute the average score within each object category, then average these category scores to obtain the final metric values.", + "bbox": [ + 511, + 571, + 906, + 647 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Human Evaluation To complement these quantitative metrics, we also conduct a user study on Amazon Mechanical Turk (AMT) with 15 participants to evaluate our results. Each participant viewed a source object and two versions of the textured target mesh from different methods, selecting which result better preserves the texture and the appearance of the reference object. Each comparison received ratings from two independent evaluators.", + "bbox": [ + 511, + 665, + 908, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 511, + 794, + 607, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We provide qualitative examples, qualitative and quantitative comparisons with baselines, and an ablation study with qualitative and quantitative analysis.", + "bbox": [ + 511, + 816, + 906, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Examples Figure 3 provides four examples illustrating how TRITEX transfers texture from the source", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21408", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "shape (left column) to three other shapes. Notice how colors match the semantics of the location, despite significant shape variations. for example the eyes of the fish at the bottom row. Figure 4 shows additional results along with the corresponding target geometries.", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Comparisons: Table 1 compares our TRI-TEX approach with all baseline methods. We find that TRI-TEX outperforms baselines in both evaluation metrics, successfully capturing the appearance of the original objects. We also report the inference time for texturing a new mesh for each method.", + "bbox": [ + 89, + 175, + 483, + 266 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Comparisons: We show qualitative comparisons between our approach and the baselines in Fig. 5. As demonstrated, MVEdit tends to deviate significantly from the input image, only taking vague inspiration from the reference. TEXTure and EASI-TEX produce artifacts due to their iterative painting techniques.", + "bbox": [ + 89, + 275, + 483, + 367 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/31c498d805cea7c48e04923f4e524059c27ed74315707c957884451541e29d44.jpg", + "image_caption": [ + "Figure 6. Results of human evaluation study. TRITEX was compared with three baseline approaches in a 2-alternative forced choice setting over $\\sim 750$ questions. Raters strongly favored TRI-TEX for better transfer of appearance." + ], + "image_footnote": [], + "bbox": [ + 93, + 380, + 480, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User Study: We conducted a user study to evaluate the quality of TRiTEX compared with baselines. For each baseline, we presented the source object and two target shapes, one textured with TRiTEX and one with the baseline, in a 2-alternative forced choice (2AFC) setting. We used Amazon Mechanical Turk (AMT) and paid raters above the minimum wage. Figure 6 shows the results of these comparisons. TRiTEX was strongly preferred by raters over all three competing approaches.", + "bbox": [ + 89, + 595, + 483, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 742, + 254, + 757 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted an ablation study to evaluate the contribution of each component in our method. The quantitative and quantitative results presented in Table 2 and Fig. 7 respectively, demonstrate that each component is essential for optimal performance. Our analysis reveals that excluding $\\mathcal{L}_{\\mathrm{mse}}$ leads to color predictions that deviate significantly from the original texture. Similarly, removing $\\mathcal{L}_{\\mathrm{app}}$ results in blurry outputs that fail to capture complex texture patterns. We further found that omitting data augmentation", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "causes the network to overfit to specific triplane projections, thereby limiting the method's generalization capabilities.", + "bbox": [ + 511, + 90, + 906, + 122 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1020060e01d90da40c21efb441622f18e0b3e7047b9bc5a7b90b93aa0af3f8a1.jpg", + "image_caption": [ + "source shape" + ], + "image_footnote": [], + "bbox": [ + 550, + 146, + 638, + 205 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1af02c454f246e5ab4655d6287ab298d4559e4ce1048236ef3a7969e5bdbea36.jpg", + "image_caption": [ + "w/o network" + ], + "image_footnote": [], + "bbox": [ + 666, + 146, + 750, + 207 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dfb810799e959cbf95d6b37aeb2d70fcad42071fe9d4136d368a0a42d384ac7a.jpg", + "image_caption": [ + "w/o $\\mathcal{L}_{\\mathrm{app}}$" + ], + "image_footnote": [], + "bbox": [ + 774, + 146, + 857, + 207 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9a6c8780c8a050a3f08152540307b729bcf5ca6b4266e415f7497b54f2e37a67.jpg", + "image_caption": [ + "w/o $\\mathcal{L}_{\\mathrm{MSE}}$" + ], + "image_footnote": [], + "bbox": [ + 557, + 237, + 640, + 297 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cb9de220c05d58891571601679b5b2328c320e277c5c719e44624ef4549beb57.jpg", + "image_caption": [ + "w/o augmentations" + ], + "image_footnote": [], + "bbox": [ + 666, + 237, + 750, + 297 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5904ea246bf966148954ad7a074e0c70dec64bbcd01578e248f6df65defa4010.jpg", + "image_caption": [ + "ours", + "Figure 7. Qualitative Ablation Study. We demonstrate the effect of removing each component of our pipeline on a single example. As shown, each component is essential for successful texture transfer." + ], + "image_footnote": [], + "bbox": [ + 774, + 237, + 857, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also ablate the contribution of the neural network in our pipeline. In the \"w/o network\" baseline, we utilize pre-extracted DIFF3f semantic features and identify the nearest neighbor for each target mesh feature within the source mesh features. Using these nearest-neighbor matches, we transfer colors from the source mesh to the target vertices and then render the target mesh with its new colors. The impact of this simplification is presented in Table 2 and Fig.7, demonstrating the degraded quality of this approach in transferring the texture details compared to our full pipeline. Two additional examples are shown in Fig.8.", + "bbox": [ + 511, + 407, + 906, + 574 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4702cb30bb4e27becee8a4cdb3aaaa242395324d3d24d600a3cf1161eb2817ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 602, + 622, + 656 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2037a3d849c10182be11fcb34db3737ad581e2d2cb2f733e82412a9e05242f48.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 648, + 595, + 743, + 655 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d75a3b477558a3e543c7022381db7f719540f7e8dda700729a5f9eff076ba1f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 776, + 595, + 870, + 656 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7fe0935e5b8f369af087b3a3cb5f54ff1c0da46ecf73d1d3cbead6e69c1aa498.jpg", + "image_caption": [ + "source shape", + "Figure 8. Network Ablation. A comparison between our full method (right) and the nearest neighboring baseline (middle), which replaces the neural network in our pipeline and transfers color based on the closest matching neighbors. As seen, this baseline fails to transfer complex texture details, unlike our full method." + ], + "image_footnote": [], + "bbox": [ + 529, + 685, + 622, + 763 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ad86fe1defffaac39c926a2511599734476a18bc1d845a8c9efbe10b2286e1f8.jpg", + "image_caption": [ + "w/o network" + ], + "image_footnote": [], + "bbox": [ + 650, + 686, + 740, + 755 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5007931271fc63995ecd1e73540dfc38cf35b217f66cf2b770e369db9dbc646a.jpg", + "image_caption": [ + "ours" + ], + "image_footnote": [], + "bbox": [ + 776, + 686, + 867, + 755 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21409", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5553b555e1579030e77fc38b505b27c963d2665c90e1530b8434c8aa93c51443.jpg", + "table_caption": [ + "Table 2. Quantitative Ablation Study. We demonstrate the effect of removing component of TRiTEX with quantitative metrics averaged across our evaluation set. The results show that each component is essential for achieving successful texture transfer." + ], + "table_footnote": [], + "table_body": "
MethodSIFID ↓CLIP sim. ↑
w/o network0.230.86
w/o L MSE0.230.86
w/o L app0.280.85
w/o augmentations0.210.85
TriTex (ours)0.220.87
", + "bbox": [ + 135, + 155, + 437, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 89, + 281, + 205, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cross-Category Object Transfer Although our methods is desinged to transfer appearance between objects from the same category, it is interesting to examine the results when transferring appearance between objects with varying degrees of semantic similarity. As shown in Figure 9, when transferring between objects with shared semantic structure (e.g. parts with similar functionality or spatial relationships), TRiTEX preserves meaningful texture patterns aligned with these semantic features. However, when transferring between objects with little semantic overlap, the resulting texture patterns become more stochastic, as the semantic mapping becomes less well-defined.", + "bbox": [ + 89, + 306, + 483, + 489 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9fd75269493e8d5de6cd21b38cdb54af3eb101932ed9c3585551032583e2ccab.jpg", + "image_caption": [ + "Source Shape" + ], + "image_footnote": [], + "bbox": [ + 101, + 527, + 181, + 594 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6553c32abc2bf514ed5be13f4a9760905d2343239d5ad891b19cf5ac9a696887.jpg", + "image_caption": [ + "Output 1" + ], + "image_footnote": [], + "bbox": [ + 233, + 517, + 316, + 602 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/36ca59a3ad5432db8c89afec9ca3b3344a94d95fa94e9c86dcbe701fd1f8a50d.jpg", + "image_caption": [ + "Output 2", + "Figure 9. Cross Category Texture Transfer. Texture transfer between objects with partial semantic correspondence (fish and deer) and versus objects with minimal semantic similarity (fish and vase with flowers)." + ], + "image_footnote": [], + "bbox": [ + 374, + 520, + 434, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations Despite the effectiveness of our approach, it has several inherent limitations. First, our method does not possess generative capabilities, which means it cannot synthesize novel details that are not present in the source texture. This limitation becomes particularly apparent when there is significant geometric variation between the source and target meshes, as our method can only map existing texture patterns rather than create new ones to accommodate the structural differences (Figure 10, top). Second, without leveraging priors from text-to-image models, our approach lacks the ability to compensate for cases where the semantic features fail to capture sufficient detail or establish accurate", + "bbox": [ + 89, + 719, + 485, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "correspondences. This can result in less detailed or less coherent texture transfers in regions where the feature matching is ambiguous or imprecise (Figure 10, bottom). Additionally, our method expects the target shape to be at the same orientation as the source shape. When enabling arbitrary rotations during training, the model becomes invariant to the object's orientation, but the ability to reproduce fine details such as eyes or wheels is reduced.", + "bbox": [ + 511, + 90, + 908, + 212 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2622a33392d9497f9c7894a1dee4ec6ef802ace47403fdb9ee78b0d703111a95.jpg", + "image_caption": [ + "Figure 10. Limitations. (Top) Since our network lacks a generative prior and the source dog has no visible tongue, it cannot synthesize an appropriate color for the target dog's tongue. (Bottom) The network incorrectly transfers the source bed's sheet texture to the target bed frame, as PCA analysis shows their feature representations are similar." + ], + "image_footnote": [], + "bbox": [ + 517, + 229, + 584, + 289 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a72f626d4dccf759b18e42f3476dd6aa7853dc5cd716202871a0fd7a6ae59d74.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 231, + 683, + 289 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/778bfdf5e9598588f9a8b27a0bfc68b274046d050429e125b4931388fb671036.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 232, + 772, + 289 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0647318a04801b83506edae6c1c5661f6dde1298603ddd9f89020cb47bd17fa6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 813, + 232, + 870, + 289 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3be40d431b08c41bb52feb2ad84d6a37ba7de026a9c23d602d3b9a3c02313493.jpg", + "image_caption": [], + "image_footnote": [ + "source shape source shape PCA" + ], + "bbox": [ + 516, + 301, + 601, + 340 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2a45642008f1a08aa894b4180bb3120152be99265466399b5656364826cd662d.jpg", + "image_caption": [], + "image_footnote": [ + "target shape" + ], + "bbox": [ + 614, + 301, + 700, + 340 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ab964b66388e3faa19f2210bfb74e2a3c5a43cbf1fdb1696d3fd845ca67a4684.jpg", + "image_caption": [], + "image_footnote": [ + "target shape PCA" + ], + "bbox": [ + 715, + 306, + 800, + 335 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1b93789865540a605dc0c15f4a55f8bd5d9a4a1c69239a1fc86a1eeb4a7f257e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 812, + 306, + 897, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 467, + 640, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented a method that allows transferring a texture learned from a single exemplar to similar shapes. The learned texture is challenging in the sense that it is a nonstationary texture, where its patterns can be strongly correlated with the semantics of the shapes. Mapping such semantic textures on a target shape, necessarily requires, at least implicitly, extracting its semantics. For that, our model uses pre-trained 3D lifted semantic features.", + "bbox": [ + 511, + 492, + 906, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our key challenge was to process the semantic features in a way that allows generalization from a single instance, in order to perform the texture learning with minimal supervision. To achieve this, we re-rendered them into a triplane representation, which preserves feature proximity and connectivity, allowing for efficient processing and mapping to the corresponding colors. We introduced an effective training method using a single mesh with appearance losses while applying augmentations. The limitation of our technique is that it currently does not possess generative abilities and cannot create details that do not appear in the source texture.", + "bbox": [ + 511, + 613, + 908, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In future work, we aim to extend this feed-forward approach for fast texturing by training the model to predict the mesh texture from a reference image, using cross-attention layers between the generated triplane and the reference image. Additionally, we plan to leverage pre-trained generative models to enhance the mapped textures, increasing both resolution and quality with their learned priors.", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21410", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank Rinon Gal, Yoad Tewel and Or Patashnik for their constructive suggestions and insightful comments.", + "bbox": [ + 89, + 114, + 485, + 146 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 172, + 187, + 188 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kfir Aberman, Jing Liao, Mingyi Shi, Dani Lischinski, Baoquan Chen, and Daniel Cohen-Or. Neural best-buddies: Sparse cross-domain correspondence. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 3", + "[2] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. arXiv e-prints, pages arXiv-2112, 2021. 3", + "[3] Raphael Bensadoun, Yanir Kleiman, Idan Azuri, Omri Harosh, Andrea Vedaldi, Natalia Neverova, and Oran Gafni. Meta 3d texturegen: Fast and consistent texture generation for 3d objects, 2024. 2", + "[4] Alexey Bokhovkin, Shubham Tulsiani, and Angela Dai. Mesh2tex: Generating mesh textures from image queries, 2023. 2", + "[5] Tianshi Cao, Karsten Kreis, Sanja Fidler, Nicholas Sharp, and Kangxue Yin. Texfusion: Synthesizing 3d textures with text-guided image diffusion models, 2023. 1, 2", + "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3, 4", + "[7] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16123-16133, 2022. 4", + "[8] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models, 2023. 1, 2", + "[9] Hansheng Chen, Ruoxi Shi, Yulin Liu, Bokui Shen, Jiayuan Gu, Gordon Wetzstein, Hao Su, and Leonidas Guibas. Generic 3d diffusion adapter using controlled multi-view editing, 2024. 1, 2, 5", + "[10] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation, 2023. 2", + "[11] Xiaobai Chen, Thomas Funkhouser, Dan B Goldman, and Eli Shechtman. Non-parametric texture transfer using MeshMatch. Adobe Technical Report 2012-2, 2012. 2", + "[12] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition, 2022. 2", + "[13] Zhiqin Chen, Kangxue Yin, and Sanja Fidler. Auv-net: Learning aligned uv maps for texture transfer and synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1465-1474, 2022. 1, 2, 3" + ], + "bbox": [ + 93, + 198, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Jeremy S De Bonet. Multiresolution sampling procedure for analysis and synthesis of texture images. In Proceedings of the 24th annual conference on Computer graphics and interactive techniques, pages 361-368, 1997. 2", + "[15] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects, 2022. 5", + "[16] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of $10\\mathrm{m} + 3\\mathrm{d}$ objects, 2023. 5", + "[17] Kangle Deng, Timothy Omernick, Alexander Weiss, Deva Ramanan, Jun-Yan Zhu, Tinghui Zhou, and Maneesh Agrawala. *Flashtex: Fast reconfigurable mesh texturing with lightcontrolnet*, 2024. 2", + "[18] Niladri Shekhar Dutt, Sanjeev Muralikrishnan, and Niloy J. Mitra. Diffusion 3d features (diff3f): Decorating textured shapes with distilled semantic features, 2024. 2, 3, 4", + "[19] Alexei A Efros and Thomas K Leung. Texture synthesis by non-parametric sampling. In Proceedings of the seventh IEEE international conference on computer vision, pages 1033-1038. IEEE, 1999. 2", + "[20] Michael Fischer, Zhengqin Li, Thu Nguyen-Phuoc, Aljaz Bozic, Zhao Dong, Carl Marshall, and Tobias Ritschel. Nerf analogies: Example-based visual attribute transfer for nerfs, 2024. 3", + "[21] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2", + "[22] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oguz. 3dgen: Triplane latent diffusion for textured mesh generation, 2023. 4", + "[23] David J Heeger and James R Bergen. Pyramid-based texture analysis/synthesis. In Proceedings of the 22nd annual conference on Computer graphics and interactive techniques, pages 229–238, 1995. 2", + "[24] Amir Hertz, Rana Hanocka, Raja Giryes, and Daniel Cohen-Or. Deep geometric texture synthesis. ACM Transactions on Graphics, 39(4), 2020. 3", + "[25] Jaihoon Kim, Juil Koo, Kyeongmin Yeo, and Minhyuk Sung. Synctweedies: A general generative framework based on synchronized diffusions, 2024. 1, 2", + "[26] Vladimir Kulikov, Shahar Yadin, Matan Kleiner, and Tomer Michaeli. Sinddm: A single image denoising diffusion model. In International Conference on Machine Learning, pages 17920-17930. PMLR, 2023. 3", + "[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation, 2022. 5", + "[28] Yuxin Liu, Minshan Xie, Hanyuan Liu, and Tien-Tsin Wong. Text-guided texturing by synchronized multi-view diffusion, 2023. 1, 2" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21411", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 3", + "[30] Jianye Lu, Athinodoros S. Georgiades, Andreas Glaser, Hongzhi Wu, Li-Yi Wei, Baining Guo, Julie Dorsey, and Holly Rushmeier. Context-aware textures. 26, 2007. 2", + "[31] Tom Mertens, Jan Kautz, Jiawen Chen, Philippe Bekaert, and Frédo Durand. Texture transfer using geometry correlation. Goslar, DEU, 2006. Eurographics Association. 2", + "[32] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures, 2022. 1, 2", + "[33] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes, 2021. 2", + "[34] Thomas W. Mitchel, Carlos Esteves, and Ameesh Makadia. Single mesh diffusion models with field latents for texture generation, 2024. 3", + "[35] Yaniv Nikankin, Niv Haim, and Michal Irani. Sinfusion: Training diffusion models on a single image or video. arXiv preprint arXiv:2211.11743, 2022. 3", + "[36] Dolev Ofri-Amar, Michal Geyer, Yoni Kasten, and Tali Dekel. Neural congealing: Aligning images to a joint semantic atlas. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19403-19412, 2023. 3", + "[37] Sai Raj Kishore Perla, Yizhi Wang, Ali Mahdavi-Amiri, and Hao Zhang. Easi-tex: Edge-aware mesh texturing from single image, 2024. 2, 5", + "[38] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2", + "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 5", + "[40] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes, 2023. 1, 2, 5", + "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 4", + "[42] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2", + "[43] Tamar Rott Shaham, Tali Dekel, and Tomer Michaeli. Singan: Learning a generative model from a single natural image, 2019. 3, 6", + "[44] J. Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion, 2022. 4", + "[45] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating textures on 3d shape surfaces, 2022. 1, 2" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Edgar Simo-Serra, Eduard Trulls, Luis Ferraz, Iasonas Kokkinos, Pascal Fua, and Francesc Moreno-Noguer. Discriminative learning of deep convolutional feature point descriptors. In Proceedings of the IEEE international conference on computer vision, pages 118-126, 2015. 3", + "[47] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. Advances in Neural Information Processing Systems, 36:1363-1389, 2023. 3", + "[48] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer, 2022. 5", + "[49] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10748-10757, 2022. 3", + "[50] Weilun Wang, Jianmin Bao, Wengang Zhou, Dongdong Chen, Dong Chen, Lu Yuan, and Houqiang Li. Sindiffusion: Learning a diffusion model from a single natural image. arXiv preprint arXiv:2211.12445, 2022. 3", + "[51] Li-Yi Wei and Marc Levoy. Texture synthesis over arbitrary manifold surfaces. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 355-360, 2001. 2", + "[52] Rundi Wu and Changxi Zheng. Learning to generate 3d shapes from a single example. ACM Transactions on Graphics, 41(6):1-19, 2022. 3", + "[53] Rundi Wu, Ruoshi Liu, Carl Vondrick, and Changxi Zheng. Sin3dm: Learning a diffusion model from a single 3d textured shape, 2024. 3, 4, 5, 6, 12", + "[54] Wenqi Xian, Patsorn Sangkloy, Varun Agrawal, Amit Raj, Jingwan Lu, Chen Fang, Fisher Yu, and James Hays. Texturegan: Controlling deep image synthesis with texture patches. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8456-8465, 2018. 2", + "[55] Jiale Xu, Xintao Wang, Weihao Cheng, Yan-Pei Cao, Ying Shan, Xiaohu Qie, and Shenghua Gao. Dream3d: Zero-shot text-to-3d synthesis using 3d shape prior and text-to-image diffusion models, 2023. 2", + "[56] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2", + "[57] Yu-Ying Yeh, Jia-Bin Huang, Changil Kim, Lei Xiao, Thu Nguyen-Phuoc, Numair Khan, Cheng Zhang, Manmohan Chandraker, Carl S Marshall, Zhao Dong, and Zhengqin Li. Texturedreamer: Image-guided texture synthesis through geometry-aware diffusion, 2024. 1, 2, 6", + "[58] Kangxue Yin, Jun Gao, Maria Shugrina, Sameh Khamis, and Sanja Fidler. 3dstylenet: Creating 3d shapes with geometric and texture style variations, 2021. 2", + "[59] Kim Youwang, Tae-Hyun Oh, and Gerard Pons-Moll. Paint-it: Text-to-texture synthesis via deep convolutional texture map optimization and physically-based rendering, 2024. 2", + "[60] Xianfang Zeng, Xin Chen, Zhongqi Qi, Wen Liu, Zibo Zhao, Zhibin Wang, Bin Fu, Yong Liu, and Gang Yu. Paint3d:" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21412", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Paint anything 3d with lighting-less texture diffusion models, 2023. 1, 2", + "[61] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 4", + "[62] Yang Zhou, Zhen Zhu, Xiang Bai, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Non-stationary texture synthesis by adversarial expansion. arXiv preprint arXiv:1805.04487, 2018. 2", + "[63] Song Chun Zhu, Yingnian Wu, and David Mumford. Filters, random fields and maximum entropy (frame): Towards a unified theory for texture modeling. International Journal of Computer Vision, 27:107-126, 1998. 2" + ], + "bbox": [ + 91, + 90, + 482, + 275 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21413", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/TriTex_ Learning Texture from a Single Mesh via Triplane Semantic Features/f0b8cf58-1685-4de9-9df0-451922c97e91_model.json b/2025/TriTex_ Learning Texture from a Single Mesh via Triplane Semantic Features/f0b8cf58-1685-4de9-9df0-451922c97e91_model.json new file mode 100644 index 0000000000000000000000000000000000000000..90e3e0ec7b7ccce7f269c53e5749fe79658bf4fa --- /dev/null +++ b/2025/TriTex_ Learning Texture from a Single Mesh via Triplane Semantic Features/f0b8cf58-1685-4de9-9df0-451922c97e91_model.json @@ -0,0 +1,2565 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.131, + 0.89, + 0.154 + ], + "angle": 0, + "content": "TriTex: Learning Texture from a Single Mesh via Triplane Semantic Features" + }, + { + "type": "image_caption", + "bbox": [ + 0.186, + 0.18, + 0.341, + 0.198 + ], + "angle": 0, + "content": "Dana Cohen-Bar1,2" + }, + { + "type": "image_caption", + "bbox": [ + 0.358, + 0.18, + 0.524, + 0.223 + ], + "angle": 0, + "content": "Daniel Cohen-Or1 \n1Tel Aviv University" + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.18, + 0.807, + 0.222 + ], + "angle": 0, + "content": "Gal Chechik2 Yoni Kasten2 NVIDIA" + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.228, + 0.346, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.225, + 0.495, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.23, + 0.691, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.223, + 0.904, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.419, + 0.908, + 0.463 + ], + "angle": 0, + "content": "Figure 1. TriTex is a method for learning to transfer a texture from a single mesh in a feed-forward manner to other input meshes. The figure shows four examples. In each, given a target geometry (up left) and a source textured mesh (up right) the texture is transferred to the target 3D object (bottom two views of the target object)." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.474, + 0.327, + 0.49 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.486, + 0.794 + ], + "angle": 0, + "content": "As 3D content creation continues to grow, transferring semantic textures between 3D meshes remains a significant challenge in computer graphics. While recent methods leverage text-to-image diffusion models for texturing, they often struggle to preserve the appearance of the source texture during texture transfer. We present TRiTEX, a novel approach that learns a volumetric texture field from a single textured mesh by mapping semantic features to surface colors. Using an efficient triplane-based architecture, our method enables semantic-aware texture transfer to a novel target mesh. Despite training on just one example, it generalizes effectively to diverse shapes within the same category. Extensive evaluation on our newly created benchmark dataset shows that TRiTEX achieves superior texture transfer quality and fast inference times compared to existing methods. Our approach advances single-example texture transfer, providing a practical solution for maintaining visual coherence across related 3D models in applications like game development and simulation." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.795, + 0.438, + 0.81 + ], + "angle": 0, + "content": "Project page: https://danacohen95.github.io/TriTex/." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.844, + 0.222, + 0.859 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Texturing 3D objects is a fundamental task with wideranging applications in game development, simulation, and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.907, + 0.702 + ], + "angle": 0, + "content": "video production. For example, when generating 3D scenes, style needs to be consistently applied across multiple 3D models that share semantic properties but differ in shape, such as in environments featuring diverse buildings or plants. Improving texture transfer can streamline texturing workflows by maintaining visual coherence across related models and preserving texture details even during mesh modifications, ensuring a consistent appearance across all elements. With the advent of large generative models, new approaches to 3D texturing have emerged, moving beyond traditional procedural methods to enable automatic texturing and texture transfer [13, 45]. However, these techniques often require the availability of large databases of 3D objects and extensive training for each class to achieve high-quality results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.909, + 0.84 + ], + "angle": 0, + "content": "Recent texturing methods use text-to-image diffusion models to texture meshes without requiring 3D data collection. Some methods [32, 57, 60] apply an SDS loss to optimize the texture map, which tends to be a slow process. Others [5, 8, 9, 25, 28, 40] employ iterative depth-conditioned image inpainting on rendered mesh views, synchronizing across views to maintain consistency. However, when performing texture transfer, these methods struggle to faithfully preserve the appearance of the source texture." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.902 + ], + "angle": 0, + "content": "We introduce TRITEx, a novel technique that learns a volumetric texture field from a single textured mesh, mapping its semantic features to RGB colors on its surface. During inference, given a target mesh, this learned field en" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21403" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.272 + ], + "angle": 0, + "content": "ables texture transfer by predicting colors based on the target mesh's semantic features, allowing the source texture to be applied in a way that maintains semantic correspondences between the source and target meshes (see Figure 1). Despite being trained on a single object, our trained volumetric texture generalizes well to novel objects within the same category. Our fast inference time enables the stylization of large scenes More specifically, to learn this texture field, we first extract semantic features using a frozen Diff3F [18] model, then project these features onto a triplane, and process them through a mapping network to produce the final triplane features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.482, + 0.392 + ], + "angle": 0, + "content": "Our experiments demonstrate the effectiveness of our texture transfer method, showing superior quality compared to previous approaches while maintaining fast inference times. We evaluate our approach using automated metrics and human evaluation on a benchmark dataset that we created specifically for the texture transfer task. Our method shows superior performance compared to previous approaches." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.394, + 0.482, + 0.56 + ], + "angle": 0, + "content": "To summarize, our contributions are: (1) Introducing a method for transferring textures from a single textured mesh to new shapes while preserving semantic relations. (2) Leveraging pre-trained 3D semantic features, reprojecting them into a triplane representation, and enabling effective processing through convolutional layers. (3) Demonstrating strong generalization to novel objects with significant shape variations, despite being trained on only a single example. (4) Achieving high-quality texture transfer with fast inference, outperforming previous methods in both speed and quality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.571, + 0.233, + 0.587 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Texture Synthesis Classical methods use sampling approaches to generate 2D textures [14, 19, 23, 51, 63]. More recent approaches rely on deep learning to create textures based on training data [54, 62]. Texturify [45] introduced a Generative Adversarial Network (GAN) for texturing 3D models using an example image, applying face convolutional operators directly on the 3D object's surface. Auv-net [13] predicts UV mapping and texture images from 3D geometry, ensuring consistent alignment across elements within the same category, which enables effective training of generative texture models with this embedding. Both Auv-net and Texturify require large datasets of predefined categories to train their generative models, whereas our approach is trained solely on a single textured mesh and can generalize to a variety of objects with similar semantics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Leveraging Large Vision-Language Pretrained Models for Texture Generation Recent advances in image-language foundation models have enabled effective texture transfer and synthesis without requiring additional training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.302 + ], + "angle": 0, + "content": "data. Text2Mesh [33] and TANGO [12] edit 3D mesh appearance by aligning rendered images with a text prompt in CLIP space. Dream3D [55] optimizes the appearance of generated geometry using a Neural Radiance Field (NeRF) by applying a CLIP loss. Latent-NeRF [32] applies the SDS-loss [38] to rendered images to optimize a NeRF representation in the stable diffusion latent space. More recently, Paint-it [59], Fantasia3D [10], and FlashTex [17] have enhanced SDS-loss-based texturing by incorporating Physically Based Rendering (PBR), BRDF modeling, and illumination control, respectively. These optimization techniques require processing each object individually, leading to slower performance, whereas our method has a much shorter inference time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.304, + 0.905, + 0.515 + ], + "angle": 0, + "content": "Other methods employed optimization-free approaches by directly applying depth-conditioned diffusion models in sequence. Texture [40] applies depth-conditioned inpainting on sequential mesh projections to achieve mesh texturing. Concurrent and follow-up methods [5, 8, 9, 25, 28] introduced improvements, with Text2Tex[8] proposing an automatic per-object view selection scheme, and TexFusion [5], SyncMVD [28], SyncTweedies [25] and MVEdit [9] synchronizing views by operating on intermediate denoising steps. Paint3D [60] additionally trains a model to remove illumination effects from the generated texture map, enabling relighting of the textured meshes. In [3], a geometry-aware multiview diffusion model is used to enhance view consistency and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Texture Transfer The problem of texture transfer has always been an area of interest and research, with early work aiming to transfer both stochastic texture [30] and semantic texture [11, 31]. More recently, several methods leverage 3D data for texture transfer tasks. 3DStyleNet [58] transforms a source mesh by applying part-aware low-frequency deformations and generating texture maps, using a target mesh as a style reference. Mesh2Tex [4] learns a texture manifold from a large dataset of 3D objects and images. This enables new images to be projected into the manifold and transferred to 3D objects in trained categories, such as cars and chairs. Personalization techniques for diffusion models [21, 42] are used in methods like TEXTure and TextureDreamer [40, 57] to texture a mesh from a few input images, which can be either natural images or rendered images from another textured mesh. MVEdit [9], EASI-Tex [37] and Paint3D [60] transfer textures from an image to a 3D object using IP-Adapter [56], where EASI-Tex further enhances geometry details by incorporating a rendered mesh edge map as an additional control for image generation. While these methods are efficient, they often fail to faithfully preserve the appearance of the source texture. In contrast, our method maintains both efficiency and accurate texture preservation." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21404" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.086, + 0.885, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.426, + 0.908, + 0.498 + ], + "angle": 0, + "content": "Figure 2. Training Pipeline (Top). Given an input textured mesh and its pre-extracted Diff3F features, we project six orthographic views to create an initial triplane. This triplane is processed using triplane-aware convolutional blocks, which, along with the texture MLP, define a coloring neural field. This field, together with the input geometry, is used to render the colored mesh. Appearance losses are applied between the true mesh appearance and the rendered appearance. Inference (Bottom). Given a new mesh (left), our pre-trained model maps its semantic properties to colors (right), transferring the texture from the original textured mesh learned during the training phase." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.513, + 0.487, + 0.664 + ], + "angle": 0, + "content": "Auv-net [13] enables texture transfer between meshes through a learned UV mapping and aligned texture representations. Mitchel et al. [34] proposed a diffusion process within a learned latent space on the surface, allowing the transferring of textures to other meshes using an input semantic label map. Nerf Analogies [20] extends texture transfer to neural radiance fields, enabling appearance transfer between source and target NeRFs with semantic similarities, though it requires per-pair training for each set of NeRFs." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.689, + 0.487, + 0.903 + ], + "angle": 0, + "content": "Semantic Features and Correspondences Finding corresponding points between images has traditionally relied on local feature descriptors [29] that are invariant to lighting and color. Later methods used features extracted from pretrained classification models [1, 46], providing invariance to local shape deformations and leveraging higher-level semantic relationships. Recently, Amir et al. [2] demonstrated the effectiveness of features from the pre-trained Vision Transformer DINO-ViT [6] as semantic descriptors. Splice [49] uses DINO-ViT features to transfer colors from a source image to a target image structure by preserving the target's local DINO features while adapting its global appearance to match the source. Neural Congealing [36] employs DINO-ViT semantic features to map a set of se" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.909, + 0.651 + ], + "angle": 0, + "content": "mentically related images into a shared canonical representation, enabling zero-shot joint editing. More recently, DIFT [47] introduced semantic features extracted from a pre-trained diffusion model during denoising. Diff3F [18] mapped both DIFT and DINO features onto a 3D mesh, resulting in 3D semantic features that enable shape correspondences. Our model benefits from these recent advances in semantic features and uses them to train a texture transfer model with a single textured mesh." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Models Trained on a Single Instance Several methods have been developed to train neural networks using a single example. SinGAN [43] trains a model to capture the internal distribution of patches within a single image, enabling the generation of diverse samples with similar visual content. Similarly, Wu et al. [52] and Hertz et al. [24] use GANs to generate shape and geometric texture variations, respectively, from a single mesh. More recent works [26, 35, 50] have employed diffusion models trained on individual images to produce image variations, with SinFusion [35] also extending this approach to single videos to create video variations. Additionally, Sin3DM [53] and [34] apply similar techniques to generate variations of a single textured mesh, either for textured mesh variations [53] or for texture variations [34]. Similarly, we propose training a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21405" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.461, + 0.108 + ], + "angle": 0, + "content": "mesh texturing model using just a single textured mesh." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.126, + 0.484, + 0.262 + ], + "angle": 0, + "content": "Triplane Representation The triplane representation introduced in EG3D [7] consists of three 2D feature grids aligned with three distinct orthogonal planes in the 3D coordinate system. This setup enables image operations, such as convolutional neural layers, and uses interpolation to define a continuous feature grid suitable for neural fields. Triplane representation has been applied in 3D-aware GANs [7] and in mesh generation using diffusion models [22, 44]. For a detailed technical description, see Sec. 3.1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.276, + 0.182, + 0.291 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.301, + 0.222, + 0.317 + ], + "angle": 0, + "content": "3.1. Background" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.323, + 0.484, + 0.534 + ], + "angle": 0, + "content": "Triplane Representation The triplane representation, introduced in EG3D [7], encodes 3D information using three axis-aligned orthogonal feature planes \\( F_{XY}, F_{XZ}, F_{YZ} \\), each of size \\( \\mathbb{R}^{W \\times H \\times C} \\), where \\( W \\) and \\( H \\) represent the spatial resolution, and \\( C \\) is the number of channels. To query features at any 3D position \\( x \\in \\mathbb{R}^3 \\), the coordinates are used to sample from each of the three feature planes (XY, XZ, YZ) via bilinear interpolation, yielding three feature vectors \\( f_{xy}, f_{xz}, f_{yz} \\in \\mathbb{R}^C \\). These feature vectors are then concatenated and passed through a lightweight MLP decoder network to produce the final output. This approach enables efficient feature extraction while maintaining expressiveness through learned feature integration, offering a balance between efficiency and quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.484, + 0.793 + ], + "angle": 0, + "content": "Importantly, the triplane representation enables processing with standard 2D convolutional layers, which are far more efficient than using 3D convolutional layers on volumetric grids. In EG3D [7], 2D convolutional layers were applied on the three concatenated feature planes. Recently, Sin3DM [53] introduced the triplane-aware convolution block, suggesting a more geometrically integrated approach. Instead of processing the three feature grids \\((F_{XY}, F_{XZ}, F_{YZ})\\) independently or simply concatenating them, this method aggregates features from each plane into the others. Let \\(F_{XY}'\\), \\(F_{XZ}'\\), \\(F_{YZ}'\\) represent the output of a single independent convolutional layer. Each plane is averaged over its axes and replicated along the third axis, before being concatenated into the relevant plane. For example, \\(F_{XY}'\\) is averaged over the \\(x\\) and \\(y\\) axes, producing \\(F_{Y}'\\) and \\(F_{X}'\\) respectively. These are then replicated along the \\(z\\)-axis and concatenated with \\(F_{YZ}'\\) and \\(F_{XZ}'\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Diff3f Semantic Features Our method leverages DIff3F [18] features, which provide robust semantic descriptors for 3D shapes without requiring textured inputs or additional training. DIff3F extracts semantic features by utilizing foundational vision models in a zero-shot manner. The process begins by rendering depth and normal maps" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.243 + ], + "angle": 0, + "content": "from multiple views of the input mesh. These maps serve as conditioning inputs to ControlNet [61] and Stable Diffusion [41] for generating view-dependent features. The features from the diffusion process maintain semantic consistency despite potential visual variations across views, and are aggregated onto the original 3D surface. These features, further enriched with DINO features [6], capture rich semantic information and enable reliable correspondence across shapes with significant geometric variations, making them well-suited for our texture transfer task." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.251, + 0.911, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.513, + 0.907, + 0.57 + ], + "angle": 0, + "content": "Figure 3. Texture Transfer Results. Given a source mesh (left column), our network transfers its appearance onto target meshes of varying shapes (three right columns). Each row demonstrates texture transfer using a different source mesh." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.577, + 0.727, + 0.593 + ], + "angle": 0, + "content": "3.2. Our approach: TRiTEX" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.659 + ], + "angle": 0, + "content": "Given a textured source mesh, our goal is to train a feedforward texturing function that can transfer that texture to any target mesh while preserving the semantics of the texture. Our pipeline is shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.906, + 0.809 + ], + "angle": 0, + "content": "The architecture, based on a deep neural network, processes a 3D mesh with pre-extracted semantic features and a query 3D point, and outputs the corresponding color for the query point. Formally, we define a learnable function: \\( f(M,\\mathbb{R}^3)\\to [0,1]^3 \\), where \\( M = (V,S,E,F) \\) is the input mesh, with \\( V = \\{v_{1},\\ldots ,v_{n}\\mid v_{i}\\in \\mathbb{R}^{3}\\} \\) and \\( S = \\{s_1,\\dots ,s_n\\mid s_i\\in \\mathbb{R}^D\\} \\), representing the 3D vertices and their corresponding semantic features, pre-extracted by [18]. Here, \\( E \\) and \\( F \\) represent the edges and faces, respectively, defined on the vertices." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.9 + ], + "angle": 0, + "content": "The first part of our architecture processes \\(M\\) by orthog.raphically projecting the semantic features into an axisaligned feature triplane \\(\\mathcal{T}\\in \\mathbb{R}^{3\\times W\\times H\\times 2D}\\) , where each feature plane is created by concatenating features from two orthographic projections taken from opposite directions. To address the low feature resolution \\((32\\times 32)\\) and enable fine" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21406" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.244 + ], + "angle": 0, + "content": "detail generation, we incorporate positional encoding into the input. Triplane-aware convolutional 2D blocks [53] (see further details in Sec.3.1) are then applied to the triplane to produce a modified triplane \\(\\mathcal{T}'\\in \\mathbb{R}^{3\\times W\\times H\\times D'}\\). Finally, to predict the color for a query 3D point, we sample features from the processed planes in \\(\\mathcal{T}'\\) according to the query point's location, concatenate them to form a single feature vector, and pass this vector through the coloring MLP \\(c:\\mathbb{R}^{3D^{\\prime}}\\to [0,1]^{3}\\), which maps it to the final RGB color for the point." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.286, + 0.484, + 0.5 + ], + "angle": 0, + "content": "Training We train our model on a single textured mesh using a rendering-based reconstruction loss, comparing rendered views of our predicted textures with ground truth views of the source mesh from randomly sampled camera angles. Let \\( I_R(\\theta) \\) be an image rendered by our pipeline from angle \\( \\theta \\), with the corresponding colors predicted by the MLP \\( c \\), generated by querying 3D points on the mesh through intersecting rendered camera rays. This image is conditioned only on the input semantic features defined for the geometry and depends on the differentiable parameters of the triplane-aware convolutional layers and \\( c \\). To train the learnable parameters, we use the ground truth rendered images, which are known for the single training textured mesh, and apply the MSE loss on the image pixels:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.517, + 0.426, + 0.56 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M S E}} (\\theta) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left\\| I _ {R} (\\theta) _ {i} - I _ {\\mathrm {G T}} (\\theta) _ {i} \\right\\| ^ {2},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.579, + 0.484, + 0.716 + ], + "angle": 0, + "content": "where \\(I_{\\mathrm{GT}}(\\theta)\\) is the ground truth image, and \\(i\\) indexes the pixels in the image. While MSE loss encourages pixel-level accuracy, it may not effectively capture high-level perceptual details. To address this, we incorporate a perceptual loss \\(\\mathcal{L}_{\\mathrm{app}}\\) from Tumanyan et al. [48], which emphasizes high-level semantic features. This improves the alignment of these features between the generated and ground truth images, enhancing texture realism. Our final loss function is then:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.729, + 0.401, + 0.747 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {\\theta} \\left[ \\mathcal {L} _ {\\mathrm {M S E}} (\\theta) + \\delta_ {\\mathrm {a p p}} \\mathcal {L} _ {\\mathrm {a p p}} (\\theta) \\right],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathbb{E}_{\\theta}\\) represents the expectation over all sampled camera angles, and \\(\\delta_{\\mathrm{app}}\\) controls the relative weight of \\(\\mathcal{L}_{\\mathrm{app}}\\) compared to \\(\\mathcal{L}_{\\mathrm{MSE}}\\). To improve generalization, we apply two levels of augmentation: (1) preprocessing augmentation, where simple 3D transformations are applied to the input mesh and features are extracted for each variant to enrich the learned feature distribution, and (2) training-time augmentation, where translation, scaling, and small rotational perturbations are applied to the mesh during training." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.109, + 0.634, + 0.149 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.159, + 0.594, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.244, + 0.629, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.331, + 0.629, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.387, + 0.633, + 0.4 + ], + "angle": 0, + "content": "Source-textured" + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.11, + 0.755, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.158, + 0.721, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.657, + 0.242, + 0.756, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.322, + 0.753, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.387, + 0.76, + 0.401 + ], + "angle": 0, + "content": "Target geometry" + }, + { + "type": "image", + "bbox": [ + 0.778, + 0.11, + 0.882, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.819, + 0.159, + 0.848, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.782, + 0.242, + 0.882, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.324, + 0.892, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.387, + 0.884, + 0.4 + ], + "angle": 0, + "content": "Target-textured" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.404, + 0.907, + 0.432 + ], + "angle": 0, + "content": "Figure 4. Additional Qualitative Results. Showing the target geometry and the texture transfer." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.459, + 0.753, + 0.476 + ], + "angle": 0, + "content": "4. Experiments And Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.484, + 0.699, + 0.5 + ], + "angle": 0, + "content": "4.1. Experiment Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.907, + 0.69 + ], + "angle": 0, + "content": "Dataset: Our experimental dataset consists of 52 3D meshes across 9 categories, curated from Objaverse [15, 16], a large-scale 3D object database. We first generated descriptive captions for each 3D object to identify semantically related objects using BLIP [27]. We then used keyword filtering on these captions to group objects into similar categories. The final selection includes only objects that both share semantic similarities and maintain consistent orientations within their respective categories. Overall, we have 256 different pairs of source and target objects. The full list of source and target objects is provided in the supplemental." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.712, + 0.909, + 0.849 + ], + "angle": 0, + "content": "Baseline Methods: We compare our method against three state-of-the-art optimization-free approaches for visually-guided mesh texturing. (1) TEXTure [40], a personalization-based approach, enabling texture transfer from a source mesh to a target mesh. (2) EASI-TEX [37] and (3) MVEdit [9] use IP-Adapter to incorporate image guidance into their texture generation process, allowing direct control over the generated textures through reference images." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Metrics: We evaluate our method using human evaluation and two automated metrics: CLIP similarity score [39]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21407" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.167, + 0.087, + 0.827, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.517, + 0.907, + 0.547 + ], + "angle": 0, + "content": "Figure 5. Qualitative Comparisons with baselines. A comparison between TRiTEX (right-most column) and state-of-the-art approaches for texture transfer. As observed, TRiTEX produces much more plausible texture transfer results." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.57, + 0.483, + 0.628 + ], + "angle": 0, + "content": "Table 1. Quantitative Evaluation. Quantitative comparison of our method with baseline methods. We evaluate two appearance metrics and running time, showing that our method best preserves the source texture while achieving competitive speed." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.638, + 0.479, + 0.724 + ], + "angle": 0, + "content": "
MethodSIFID ↓CLIP sim. ↑Inference Runtime
TEXTure0.340.845 min.
MVEdit0.380.841 min.
EASI-TEX0.290.8515 min.
TriTex (ours)0.220.871 min.
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.483, + 0.884 + ], + "angle": 0, + "content": "and Single Image Frechet Inception Distance (SIFID). The CLIP score, adopted following TextureDreamer [57], measures the semantic and visual similarity between source and generated textures, indicating how well our method preserves the desired appearance. SIFID was originally proposed by [43] to measure internal patch statistics within single images. Following Sin3DM [53], we adopt this metric to evaluate texture similarity between source and target shapes." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "For both metrics, we render each source and target mesh" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.572, + 0.907, + 0.648 + ], + "angle": 0, + "content": "from the same 10 fixed viewpoints and compare the corresponding views against each other. To ensure balanced evaluation across our dataset, we first compute the average score within each object category, then average these category scores to obtain the final metric values." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.666, + 0.909, + 0.788 + ], + "angle": 0, + "content": "Human Evaluation To complement these quantitative metrics, we also conduct a user study on Amazon Mechanical Turk (AMT) with 15 participants to evaluate our results. Each participant viewed a source object and two versions of the textured target mesh from different methods, selecting which result better preserves the texture and the appearance of the reference object. Each comparison received ratings from two independent evaluators." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.795, + 0.608, + 0.81 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.817, + 0.907, + 0.864 + ], + "angle": 0, + "content": "We provide qualitative examples, qualitative and quantitative comparisons with baselines, and an ablation study with qualitative and quantitative analysis." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Qualitative Examples Figure 3 provides four examples illustrating how TRITEX transfers texture from the source" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21408" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "shape (left column) to three other shapes. Notice how colors match the semantics of the location, despite significant shape variations. for example the eyes of the fish at the bottom row. Figure 4 shows additional results along with the corresponding target geometries." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.176, + 0.484, + 0.267 + ], + "angle": 0, + "content": "Quantitative Comparisons: Table 1 compares our TRI-TEX approach with all baseline methods. We find that TRI-TEX outperforms baselines in both evaluation metrics, successfully capturing the appearance of the original objects. We also report the inference time for texturing a new mesh for each method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.276, + 0.484, + 0.368 + ], + "angle": 0, + "content": "Qualitative Comparisons: We show qualitative comparisons between our approach and the baselines in Fig. 5. As demonstrated, MVEdit tends to deviate significantly from the input image, only taking vague inspiration from the reference. TEXTure and EASI-TEX produce artifacts due to their iterative painting techniques." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.381, + 0.482, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.506, + 0.484, + 0.562 + ], + "angle": 0, + "content": "Figure 6. Results of human evaluation study. TRITEX was compared with three baseline approaches in a 2-alternative forced choice setting over \\(\\sim 750\\) questions. Raters strongly favored TRI-TEX for better transfer of appearance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.484, + 0.734 + ], + "angle": 0, + "content": "User Study: We conducted a user study to evaluate the quality of TRiTEX compared with baselines. For each baseline, we presented the source object and two target shapes, one textured with TRiTEX and one with the baseline, in a 2-alternative forced choice (2AFC) setting. We used Amazon Mechanical Turk (AMT) and paid raters above the minimum wage. Figure 6 shows the results of these comparisons. TRiTEX was strongly preferred by raters over all three competing approaches." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.255, + 0.758 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We conducted an ablation study to evaluate the contribution of each component in our method. The quantitative and quantitative results presented in Table 2 and Fig. 7 respectively, demonstrate that each component is essential for optimal performance. Our analysis reveals that excluding \\(\\mathcal{L}_{\\mathrm{mse}}\\) leads to color predictions that deviate significantly from the original texture. Similarly, removing \\(\\mathcal{L}_{\\mathrm{app}}\\) results in blurry outputs that fail to capture complex texture patterns. We further found that omitting data augmentation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.123 + ], + "angle": 0, + "content": "causes the network to overfit to specific triplane projections, thereby limiting the method's generalization capabilities." + }, + { + "type": "image", + "bbox": [ + 0.551, + 0.147, + 0.639, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.217, + 0.639, + 0.229 + ], + "angle": 0, + "content": "source shape" + }, + { + "type": "image", + "bbox": [ + 0.667, + 0.147, + 0.75, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.216, + 0.747, + 0.228 + ], + "angle": 0, + "content": "w/o network" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.147, + 0.858, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.216, + 0.845, + 0.23 + ], + "angle": 0, + "content": "w/o \\(\\mathcal{L}_{\\mathrm{app}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.238, + 0.641, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.307, + 0.629, + 0.319 + ], + "angle": 0, + "content": "w/o \\(\\mathcal{L}_{\\mathrm{MSE}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.667, + 0.238, + 0.75, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.307, + 0.766, + 0.319 + ], + "angle": 0, + "content": "w/o augmentations" + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.238, + 0.858, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.804, + 0.309, + 0.832, + 0.318 + ], + "angle": 0, + "content": "ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.332, + 0.907, + 0.386 + ], + "angle": 0, + "content": "Figure 7. Qualitative Ablation Study. We demonstrate the effect of removing each component of our pipeline on a single example. As shown, each component is essential for successful texture transfer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.408, + 0.907, + 0.575 + ], + "angle": 0, + "content": "We also ablate the contribution of the neural network in our pipeline. In the \"w/o network\" baseline, we utilize pre-extracted DIFF3f semantic features and identify the nearest neighbor for each target mesh feature within the source mesh features. Using these nearest-neighbor matches, we transfer colors from the source mesh to the target vertices and then render the target mesh with its new colors. The impact of this simplification is presented in Table 2 and Fig.7, demonstrating the degraded quality of this approach in transferring the texture details compared to our full pipeline. Two additional examples are shown in Fig.8." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.603, + 0.623, + 0.657 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.65, + 0.596, + 0.745, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.596, + 0.872, + 0.657 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.686, + 0.624, + 0.765 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.771, + 0.621, + 0.784 + ], + "angle": 0, + "content": "source shape" + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.688, + 0.741, + 0.756 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.771, + 0.746, + 0.783 + ], + "angle": 0, + "content": "w/o network" + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.688, + 0.869, + 0.756 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.821, + 0.772, + 0.849, + 0.782 + ], + "angle": 0, + "content": "ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.797, + 0.907, + 0.879 + ], + "angle": 0, + "content": "Figure 8. Network Ablation. A comparison between our full method (right) and the nearest neighboring baseline (middle), which replaces the neural network in our pipeline and transfers color based on the closest matching neighbors. As seen, this baseline fails to transfer complex texture details, unlike our full method." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21409" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.148 + ], + "angle": 0, + "content": "Table 2. Quantitative Ablation Study. We demonstrate the effect of removing component of TRiTEX with quantitative metrics averaged across our evaluation set. The results show that each component is essential for achieving successful texture transfer." + }, + { + "type": "table", + "bbox": [ + 0.136, + 0.156, + 0.439, + 0.259 + ], + "angle": 0, + "content": "
MethodSIFID ↓CLIP sim. ↑
w/o network0.230.86
w/o L MSE0.230.86
w/o L app0.280.85
w/o augmentations0.210.85
TriTex (ours)0.220.87
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.282, + 0.206, + 0.298 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.308, + 0.485, + 0.491 + ], + "angle": 0, + "content": "Cross-Category Object Transfer Although our methods is desinged to transfer appearance between objects from the same category, it is interesting to examine the results when transferring appearance between objects with varying degrees of semantic similarity. As shown in Figure 9, when transferring between objects with shared semantic structure (e.g. parts with similar functionality or spatial relationships), TRiTEX preserves meaningful texture patterns aligned with these semantic features. However, when transferring between objects with little semantic overlap, the resulting texture patterns become more stochastic, as the semantic mapping becomes less well-defined." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.529, + 0.183, + 0.595 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.612, + 0.201, + 0.626 + ], + "angle": 0, + "content": "Source Shape" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.518, + 0.317, + 0.603 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.612, + 0.313, + 0.625 + ], + "angle": 0, + "content": "Output 1" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.521, + 0.435, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.612, + 0.441, + 0.625 + ], + "angle": 0, + "content": "Output 2" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.638, + 0.486, + 0.695 + ], + "angle": 0, + "content": "Figure 9. Cross Category Texture Transfer. Texture transfer between objects with partial semantic correspondence (fish and deer) and versus objects with minimal semantic similarity (fish and vase with flowers)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Limitations Despite the effectiveness of our approach, it has several inherent limitations. First, our method does not possess generative capabilities, which means it cannot synthesize novel details that are not present in the source texture. This limitation becomes particularly apparent when there is significant geometric variation between the source and target meshes, as our method can only map existing texture patterns rather than create new ones to accommodate the structural differences (Figure 10, top). Second, without leveraging priors from text-to-image models, our approach lacks the ability to compensate for cases where the semantic features fail to capture sufficient detail or establish accurate" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.213 + ], + "angle": 0, + "content": "correspondences. This can result in less detailed or less coherent texture transfers in regions where the feature matching is ambiguous or imprecise (Figure 10, bottom). Additionally, our method expects the target shape to be at the same orientation as the source shape. When enabling arbitrary rotations during training, the model becomes invariant to the object's orientation, but the ability to reproduce fine details such as eyes or wheels is reduced." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.231, + 0.585, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.232, + 0.684, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.233, + 0.774, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.815, + 0.233, + 0.871, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.302, + 0.602, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.302, + 0.701, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.308, + 0.801, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.813, + 0.308, + 0.898, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.521, + 0.345, + 0.715, + 0.358 + ], + "angle": 0, + "content": "source shape source shape PCA" + }, + { + "type": "image_footnote", + "bbox": [ + 0.721, + 0.345, + 0.797, + 0.358 + ], + "angle": 0, + "content": "target shape" + }, + { + "type": "image_footnote", + "bbox": [ + 0.804, + 0.345, + 0.91, + 0.358 + ], + "angle": 0, + "content": "target shape PCA" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.372, + 0.909, + 0.456 + ], + "angle": 0, + "content": "Figure 10. Limitations. (Top) Since our network lacks a generative prior and the source dog has no visible tongue, it cannot synthesize an appropriate color for the target dog's tongue. (Bottom) The network incorrectly transfers the source bed's sheet texture to the target bed frame, as PCA analysis shows their feature representations are similar." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.468, + 0.642, + 0.485 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.907, + 0.613 + ], + "angle": 0, + "content": "We have presented a method that allows transferring a texture learned from a single exemplar to similar shapes. The learned texture is challenging in the sense that it is a nonstationary texture, where its patterns can be strongly correlated with the semantics of the shapes. Mapping such semantic textures on a target shape, necessarily requires, at least implicitly, extracting its semantics. For that, our model uses pre-trained 3D lifted semantic features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.909, + 0.795 + ], + "angle": 0, + "content": "Our key challenge was to process the semantic features in a way that allows generalization from a single instance, in order to perform the texture learning with minimal supervision. To achieve this, we re-rendered them into a triplane representation, which preserves feature proximity and connectivity, allowing for efficient processing and mapping to the corresponding colors. We introduced an effective training method using a single mesh with appearance losses while applying augmentations. The limitation of our technique is that it currently does not possess generative abilities and cannot create details that do not appear in the source texture." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In future work, we aim to extend this feed-forward approach for fast texturing by training the model to predict the mesh texture from a reference image, using cross-attention layers between the generated triplane and the reference image. Additionally, we plan to leverage pre-trained generative models to enhance the mapped textures, increasing both resolution and quality with their learned priors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21410" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.486, + 0.147 + ], + "angle": 0, + "content": "We thank Rinon Gal, Yoad Tewel and Or Patashnik for their constructive suggestions and insightful comments." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.173, + 0.188, + 0.189 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.199, + 0.484, + 0.254 + ], + "angle": 0, + "content": "[1] Kfir Aberman, Jing Liao, Mingyi Shi, Dani Lischinski, Baoquan Chen, and Daniel Cohen-Or. Neural best-buddies: Sparse cross-domain correspondence. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.484, + 0.298 + ], + "angle": 0, + "content": "[2] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. arXiv e-prints, pages arXiv-2112, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.3, + 0.484, + 0.355 + ], + "angle": 0, + "content": "[3] Raphael Bensadoun, Yanir Kleiman, Idan Azuri, Omri Harosh, Andrea Vedaldi, Natalia Neverova, and Oran Gafni. Meta 3d texturegen: Fast and consistent texture generation for 3d objects, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.357, + 0.482, + 0.398 + ], + "angle": 0, + "content": "[4] Alexey Bokhovkin, Shubham Tulsiani, and Angela Dai. Mesh2tex: Generating mesh textures from image queries, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.401, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[5] Tianshi Cao, Karsten Kreis, Sanja Fidler, Nicholas Sharp, and Kangxue Yin. Texfusion: Synthesizing 3d textures with text-guided image diffusion models, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.444, + 0.484, + 0.514 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.516, + 0.484, + 0.599 + ], + "angle": 0, + "content": "[7] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16123-16133, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.601, + 0.482, + 0.642 + ], + "angle": 0, + "content": "[8] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.644, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[9] Hansheng Chen, Ruoxi Shi, Yulin Liu, Bokui Shen, Jiayuan Gu, Gordon Wetzstein, Hao Su, and Leonidas Guibas. Generic 3d diffusion adapter using controlled multi-view editing, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.701, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[10] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.745, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[11] Xiaobai Chen, Thomas Funkhouser, Dan B Goldman, and Eli Shechtman. Non-parametric texture transfer using MeshMatch. Adobe Technical Report 2012-2, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[12] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[13] Zhiqin Chen, Kangxue Yin, and Sanja Fidler. Auv-net: Learning aligned uv maps for texture transfer and synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1465-1474, 2022. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.199, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[14] Jeremy S De Bonet. Multiresolution sampling procedure for analysis and synthesis of texture images. In Proceedings of the 24th annual conference on Computer graphics and interactive techniques, pages 361-368, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[15] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[16] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of \\(10\\mathrm{m} + 3\\mathrm{d}\\) objects, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.291, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[17] Kangle Deng, Timothy Omernick, Alexander Weiss, Deva Ramanan, Jun-Yan Zhu, Tinghui Zhou, and Maneesh Agrawala. *Flashtex: Fast reconfigurable mesh texturing with lightcontrolnet*, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.348, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[18] Niladri Shekhar Dutt, Sanjeev Muralikrishnan, and Niloy J. Mitra. Diffusion 3d features (diff3f): Decorating textured shapes with distilled semantic features, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.391, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[19] Alexei A Efros and Thomas K Leung. Texture synthesis by non-parametric sampling. In Proceedings of the seventh IEEE international conference on computer vision, pages 1033-1038. IEEE, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[20] Michael Fischer, Zhengqin Li, Thu Nguyen-Phuoc, Aljaz Bozic, Zhao Dong, Carl Marshall, and Tobias Ritschel. Nerf analogies: Example-based visual attribute transfer for nerfs, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[21] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[22] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oguz. 3dgen: Triplane latent diffusion for textured mesh generation, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.617, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[23] David J Heeger and James R Bergen. Pyramid-based texture analysis/synthesis. In Proceedings of the 22nd annual conference on Computer graphics and interactive techniques, pages 229–238, 1995. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.675, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[24] Amir Hertz, Rana Hanocka, Raja Giryes, and Daniel Cohen-Or. Deep geometric texture synthesis. ACM Transactions on Graphics, 39(4), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[25] Jaihoon Kim, Juil Koo, Kyeongmin Yeo, and Minhyuk Sung. Synctweedies: A general generative framework based on synchronized diffusions, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.76, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[26] Vladimir Kulikov, Shahar Yadin, Matan Kleiner, and Tomer Michaeli. Sinddm: A single image denoising diffusion model. In International Conference on Machine Learning, pages 17920-17930. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[28] Yuxin Liu, Minshan Xie, Hanyuan Liu, and Tien-Tsin Wong. Text-guided texturing by synchronized multi-view diffusion, 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21411" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[29] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.175 + ], + "angle": 0, + "content": "[30] Jianye Lu, Athinodoros S. Georgiades, Andreas Glaser, Hongzhi Wu, Li-Yi Wei, Baining Guo, Julie Dorsey, and Holly Rushmeier. Context-aware textures. 26, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[31] Tom Mertens, Jan Kautz, Jiawen Chen, Philippe Bekaert, and Frédo Durand. Texture transfer using geometry correlation. Goslar, DEU, 2006. Eurographics Association. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[32] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[33] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[34] Thomas W. Mitchel, Carlos Esteves, and Ameesh Makadia. Single mesh diffusion models with field latents for texture generation, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.385 + ], + "angle": 0, + "content": "[35] Yaniv Nikankin, Niv Haim, and Michal Irani. Sinfusion: Training diffusion models on a single image or video. arXiv preprint arXiv:2211.11743, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[36] Dolev Ofri-Amar, Michal Geyer, Yoni Kasten, and Tali Dekel. Neural congealing: Aligning images to a joint semantic atlas. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19403-19412, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[37] Sai Raj Kishore Perla, Yizhi Wang, Ali Mahdavi-Amiri, and Hao Zhang. Easi-tex: Edge-aware mesh texturing from single image, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[38] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[40] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.691 + ], + "angle": 0, + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[42] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[43] Tamar Rott Shaham, Tali Dekel, and Tomer Michaeli. Singan: Learning a generative model from a single natural image, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[44] J. Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[45] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating textures on 3d shape surfaces, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[46] Edgar Simo-Serra, Eduard Trulls, Luis Ferraz, Iasonas Kokkinos, Pascal Fua, and Francesc Moreno-Noguer. Discriminative learning of deep convolutional feature point descriptors. In Proceedings of the IEEE international conference on computer vision, pages 118-126, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[47] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. Advances in Neural Information Processing Systems, 36:1363-1389, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[48] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[49] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10748-10757, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.332, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[50] Weilun Wang, Jianmin Bao, Wengang Zhou, Dongdong Chen, Dong Chen, Lu Yuan, and Houqiang Li. Sindiffusion: Learning a diffusion model from a single natural image. arXiv preprint arXiv:2211.12445, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[51] Li-Yi Wei and Marc Levoy. Texture synthesis over arbitrary manifold surfaces. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 355-360, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[52] Rundi Wu and Changxi Zheng. Learning to generate 3d shapes from a single example. ACM Transactions on Graphics, 41(6):1-19, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[53] Rundi Wu, Ruoshi Liu, Carl Vondrick, and Changxi Zheng. Sin3dm: Learning a diffusion model from a single 3d textured shape, 2024. 3, 4, 5, 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[54] Wenqi Xian, Patsorn Sangkloy, Varun Agrawal, Amit Raj, Jingwan Lu, Chen Fang, Fisher Yu, and James Hays. Texturegan: Controlling deep image synthesis with texture patches. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8456-8465, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.603, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[55] Jiale Xu, Xintao Wang, Weihao Cheng, Yan-Pei Cao, Ying Shan, Xiaohu Qie, and Shenghua Gao. Dream3d: Zero-shot text-to-3d synthesis using 3d shape prior and text-to-image diffusion models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.714 + ], + "angle": 0, + "content": "[56] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.716, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[57] Yu-Ying Yeh, Jia-Bin Huang, Changil Kim, Lei Xiao, Thu Nguyen-Phuoc, Numair Khan, Cheng Zhang, Manmohan Chandraker, Carl S Marshall, Zhao Dong, and Zhengqin Li. Texturedreamer: Image-guided texture synthesis through geometry-aware diffusion, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[58] Kangxue Yin, Jun Gao, Maria Shugrina, Sameh Khamis, and Sanja Fidler. 3dstylenet: Creating 3d shapes with geometric and texture style variations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.83, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[59] Kim Youwang, Tae-Hyun Oh, and Gerard Pons-Moll. Paint-it: Text-to-texture synthesis via deep convolutional texture map optimization and physically-based rendering, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[60] Xianfang Zeng, Xin Chen, Zhongqi Qi, Wen Liu, Zibo Zhao, Zhibin Wang, Bin Fu, Yong Liu, and Gang Yu. Paint3d:" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21412" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Paint anything 3d with lighting-less texture diffusion models, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.162 + ], + "angle": 0, + "content": "[61] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[62] Yang Zhou, Zhen Zhu, Xiang Bai, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Non-stationary texture synthesis by adversarial expansion. arXiv preprint arXiv:1805.04487, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[63] Song Chun Zhu, Yingnian Wu, and David Mumford. Filters, random fields and maximum entropy (frame): Towards a unified theory for texture modeling. International Journal of Computer Vision, 27:107-126, 1998. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "21413" + } + ] +] \ No newline at end of file diff --git a/2025/TriTex_ Learning Texture from a Single Mesh via Triplane Semantic Features/f0b8cf58-1685-4de9-9df0-451922c97e91_origin.pdf b/2025/TriTex_ Learning Texture from a Single Mesh via Triplane Semantic Features/f0b8cf58-1685-4de9-9df0-451922c97e91_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..41bfaaba214694f0437b5fc56e51cc7bd3a6bc52 --- /dev/null +++ b/2025/TriTex_ Learning Texture from a Single Mesh via Triplane Semantic Features/f0b8cf58-1685-4de9-9df0-451922c97e91_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ef249bcc7f8a1a82132a32a7893165a6819004f6bdeb7d55236ad3528b638633 +size 5209954 diff --git a/2025/TriTex_ Learning Texture from a Single Mesh via Triplane Semantic Features/full.md b/2025/TriTex_ Learning Texture from a Single Mesh via Triplane Semantic Features/full.md new file mode 100644 index 0000000000000000000000000000000000000000..420cca7bb249966bb32304bad9a2be56d2992d21 --- /dev/null +++ b/2025/TriTex_ Learning Texture from a Single Mesh via Triplane Semantic Features/full.md @@ -0,0 +1,340 @@ +# TriTex: Learning Texture from a Single Mesh via Triplane Semantic Features + +![](images/746f54f9680bfb83d22012db1d92199bd85a1d28f9829e6c65e8024bd14d74b3.jpg) + +![](images/4ebda9dd5a6b95cd194b2fd276af872402e33dad9782ae3bcba8f76c5e0c9562.jpg) +Dana Cohen-Bar1,2 +Daniel Cohen-Or1 +1Tel Aviv University + +![](images/558702c05ec899e19f121491cd4ee1aadc6f988f36124a723b46de325b730729.jpg) + +![](images/835465995a9e2f432a28c969237005ad46254bc663548612cdf2a03e9f69c751.jpg) +Gal Chechik2 Yoni Kasten2 NVIDIA +Figure 1. TriTex is a method for learning to transfer a texture from a single mesh in a feed-forward manner to other input meshes. The figure shows four examples. In each, given a target geometry (up left) and a source textured mesh (up right) the texture is transferred to the target 3D object (bottom two views of the target object). + +# Abstract + +As 3D content creation continues to grow, transferring semantic textures between 3D meshes remains a significant challenge in computer graphics. While recent methods leverage text-to-image diffusion models for texturing, they often struggle to preserve the appearance of the source texture during texture transfer. We present TRiTEX, a novel approach that learns a volumetric texture field from a single textured mesh by mapping semantic features to surface colors. Using an efficient triplane-based architecture, our method enables semantic-aware texture transfer to a novel target mesh. Despite training on just one example, it generalizes effectively to diverse shapes within the same category. Extensive evaluation on our newly created benchmark dataset shows that TRiTEX achieves superior texture transfer quality and fast inference times compared to existing methods. Our approach advances single-example texture transfer, providing a practical solution for maintaining visual coherence across related 3D models in applications like game development and simulation. + +Project page: https://danacohen95.github.io/TriTex/. + +# 1. Introduction + +Texturing 3D objects is a fundamental task with wideranging applications in game development, simulation, and + +video production. For example, when generating 3D scenes, style needs to be consistently applied across multiple 3D models that share semantic properties but differ in shape, such as in environments featuring diverse buildings or plants. Improving texture transfer can streamline texturing workflows by maintaining visual coherence across related models and preserving texture details even during mesh modifications, ensuring a consistent appearance across all elements. With the advent of large generative models, new approaches to 3D texturing have emerged, moving beyond traditional procedural methods to enable automatic texturing and texture transfer [13, 45]. However, these techniques often require the availability of large databases of 3D objects and extensive training for each class to achieve high-quality results. + +Recent texturing methods use text-to-image diffusion models to texture meshes without requiring 3D data collection. Some methods [32, 57, 60] apply an SDS loss to optimize the texture map, which tends to be a slow process. Others [5, 8, 9, 25, 28, 40] employ iterative depth-conditioned image inpainting on rendered mesh views, synchronizing across views to maintain consistency. However, when performing texture transfer, these methods struggle to faithfully preserve the appearance of the source texture. + +We introduce TRITEx, a novel technique that learns a volumetric texture field from a single textured mesh, mapping its semantic features to RGB colors on its surface. During inference, given a target mesh, this learned field en + +ables texture transfer by predicting colors based on the target mesh's semantic features, allowing the source texture to be applied in a way that maintains semantic correspondences between the source and target meshes (see Figure 1). Despite being trained on a single object, our trained volumetric texture generalizes well to novel objects within the same category. Our fast inference time enables the stylization of large scenes More specifically, to learn this texture field, we first extract semantic features using a frozen Diff3F [18] model, then project these features onto a triplane, and process them through a mapping network to produce the final triplane features. + +Our experiments demonstrate the effectiveness of our texture transfer method, showing superior quality compared to previous approaches while maintaining fast inference times. We evaluate our approach using automated metrics and human evaluation on a benchmark dataset that we created specifically for the texture transfer task. Our method shows superior performance compared to previous approaches. + +To summarize, our contributions are: (1) Introducing a method for transferring textures from a single textured mesh to new shapes while preserving semantic relations. (2) Leveraging pre-trained 3D semantic features, reprojecting them into a triplane representation, and enabling effective processing through convolutional layers. (3) Demonstrating strong generalization to novel objects with significant shape variations, despite being trained on only a single example. (4) Achieving high-quality texture transfer with fast inference, outperforming previous methods in both speed and quality. + +# 2. Related Work + +Texture Synthesis Classical methods use sampling approaches to generate 2D textures [14, 19, 23, 51, 63]. More recent approaches rely on deep learning to create textures based on training data [54, 62]. Texturify [45] introduced a Generative Adversarial Network (GAN) for texturing 3D models using an example image, applying face convolutional operators directly on the 3D object's surface. Auv-net [13] predicts UV mapping and texture images from 3D geometry, ensuring consistent alignment across elements within the same category, which enables effective training of generative texture models with this embedding. Both Auv-net and Texturify require large datasets of predefined categories to train their generative models, whereas our approach is trained solely on a single textured mesh and can generalize to a variety of objects with similar semantics. + +Leveraging Large Vision-Language Pretrained Models for Texture Generation Recent advances in image-language foundation models have enabled effective texture transfer and synthesis without requiring additional training + +data. Text2Mesh [33] and TANGO [12] edit 3D mesh appearance by aligning rendered images with a text prompt in CLIP space. Dream3D [55] optimizes the appearance of generated geometry using a Neural Radiance Field (NeRF) by applying a CLIP loss. Latent-NeRF [32] applies the SDS-loss [38] to rendered images to optimize a NeRF representation in the stable diffusion latent space. More recently, Paint-it [59], Fantasia3D [10], and FlashTex [17] have enhanced SDS-loss-based texturing by incorporating Physically Based Rendering (PBR), BRDF modeling, and illumination control, respectively. These optimization techniques require processing each object individually, leading to slower performance, whereas our method has a much shorter inference time. + +Other methods employed optimization-free approaches by directly applying depth-conditioned diffusion models in sequence. Texture [40] applies depth-conditioned inpainting on sequential mesh projections to achieve mesh texturing. Concurrent and follow-up methods [5, 8, 9, 25, 28] introduced improvements, with Text2Tex[8] proposing an automatic per-object view selection scheme, and TexFusion [5], SyncMVD [28], SyncTweedies [25] and MVEdit [9] synchronizing views by operating on intermediate denoising steps. Paint3D [60] additionally trains a model to remove illumination effects from the generated texture map, enabling relighting of the textured meshes. In [3], a geometry-aware multiview diffusion model is used to enhance view consistency and efficiency. + +Texture Transfer The problem of texture transfer has always been an area of interest and research, with early work aiming to transfer both stochastic texture [30] and semantic texture [11, 31]. More recently, several methods leverage 3D data for texture transfer tasks. 3DStyleNet [58] transforms a source mesh by applying part-aware low-frequency deformations and generating texture maps, using a target mesh as a style reference. Mesh2Tex [4] learns a texture manifold from a large dataset of 3D objects and images. This enables new images to be projected into the manifold and transferred to 3D objects in trained categories, such as cars and chairs. Personalization techniques for diffusion models [21, 42] are used in methods like TEXTure and TextureDreamer [40, 57] to texture a mesh from a few input images, which can be either natural images or rendered images from another textured mesh. MVEdit [9], EASI-Tex [37] and Paint3D [60] transfer textures from an image to a 3D object using IP-Adapter [56], where EASI-Tex further enhances geometry details by incorporating a rendered mesh edge map as an additional control for image generation. While these methods are efficient, they often fail to faithfully preserve the appearance of the source texture. In contrast, our method maintains both efficiency and accurate texture preservation. + +![](images/39d8cb70cb3987e700167c4773a278db52a6d35e48a246a4f33d18ea13ff35b0.jpg) +Figure 2. Training Pipeline (Top). Given an input textured mesh and its pre-extracted Diff3F features, we project six orthographic views to create an initial triplane. This triplane is processed using triplane-aware convolutional blocks, which, along with the texture MLP, define a coloring neural field. This field, together with the input geometry, is used to render the colored mesh. Appearance losses are applied between the true mesh appearance and the rendered appearance. Inference (Bottom). Given a new mesh (left), our pre-trained model maps its semantic properties to colors (right), transferring the texture from the original textured mesh learned during the training phase. + +Auv-net [13] enables texture transfer between meshes through a learned UV mapping and aligned texture representations. Mitchel et al. [34] proposed a diffusion process within a learned latent space on the surface, allowing the transferring of textures to other meshes using an input semantic label map. Nerf Analogies [20] extends texture transfer to neural radiance fields, enabling appearance transfer between source and target NeRFs with semantic similarities, though it requires per-pair training for each set of NeRFs. + +Semantic Features and Correspondences Finding corresponding points between images has traditionally relied on local feature descriptors [29] that are invariant to lighting and color. Later methods used features extracted from pretrained classification models [1, 46], providing invariance to local shape deformations and leveraging higher-level semantic relationships. Recently, Amir et al. [2] demonstrated the effectiveness of features from the pre-trained Vision Transformer DINO-ViT [6] as semantic descriptors. Splice [49] uses DINO-ViT features to transfer colors from a source image to a target image structure by preserving the target's local DINO features while adapting its global appearance to match the source. Neural Congealing [36] employs DINO-ViT semantic features to map a set of se + +mentically related images into a shared canonical representation, enabling zero-shot joint editing. More recently, DIFT [47] introduced semantic features extracted from a pre-trained diffusion model during denoising. Diff3F [18] mapped both DIFT and DINO features onto a 3D mesh, resulting in 3D semantic features that enable shape correspondences. Our model benefits from these recent advances in semantic features and uses them to train a texture transfer model with a single textured mesh. + +Models Trained on a Single Instance Several methods have been developed to train neural networks using a single example. SinGAN [43] trains a model to capture the internal distribution of patches within a single image, enabling the generation of diverse samples with similar visual content. Similarly, Wu et al. [52] and Hertz et al. [24] use GANs to generate shape and geometric texture variations, respectively, from a single mesh. More recent works [26, 35, 50] have employed diffusion models trained on individual images to produce image variations, with SinFusion [35] also extending this approach to single videos to create video variations. Additionally, Sin3DM [53] and [34] apply similar techniques to generate variations of a single textured mesh, either for textured mesh variations [53] or for texture variations [34]. Similarly, we propose training a + +mesh texturing model using just a single textured mesh. + +Triplane Representation The triplane representation introduced in EG3D [7] consists of three 2D feature grids aligned with three distinct orthogonal planes in the 3D coordinate system. This setup enables image operations, such as convolutional neural layers, and uses interpolation to define a continuous feature grid suitable for neural fields. Triplane representation has been applied in 3D-aware GANs [7] and in mesh generation using diffusion models [22, 44]. For a detailed technical description, see Sec. 3.1. + +# 3. Method + +# 3.1. Background + +Triplane Representation The triplane representation, introduced in EG3D [7], encodes 3D information using three axis-aligned orthogonal feature planes $F_{XY}, F_{XZ}, F_{YZ}$ , each of size $\mathbb{R}^{W \times H \times C}$ , where $W$ and $H$ represent the spatial resolution, and $C$ is the number of channels. To query features at any 3D position $x \in \mathbb{R}^3$ , the coordinates are used to sample from each of the three feature planes (XY, XZ, YZ) via bilinear interpolation, yielding three feature vectors $f_{xy}, f_{xz}, f_{yz} \in \mathbb{R}^C$ . These feature vectors are then concatenated and passed through a lightweight MLP decoder network to produce the final output. This approach enables efficient feature extraction while maintaining expressiveness through learned feature integration, offering a balance between efficiency and quality. + +Importantly, the triplane representation enables processing with standard 2D convolutional layers, which are far more efficient than using 3D convolutional layers on volumetric grids. In EG3D [7], 2D convolutional layers were applied on the three concatenated feature planes. Recently, Sin3DM [53] introduced the triplane-aware convolution block, suggesting a more geometrically integrated approach. Instead of processing the three feature grids $(F_{XY}, F_{XZ}, F_{YZ})$ independently or simply concatenating them, this method aggregates features from each plane into the others. Let $F_{XY}'$ , $F_{XZ}'$ , $F_{YZ}'$ represent the output of a single independent convolutional layer. Each plane is averaged over its axes and replicated along the third axis, before being concatenated into the relevant plane. For example, $F_{XY}'$ is averaged over the $x$ and $y$ axes, producing $F_{Y}'$ and $F_{X}'$ respectively. These are then replicated along the $z$ -axis and concatenated with $F_{YZ}'$ and $F_{XZ}'$ , respectively. + +Diff3f Semantic Features Our method leverages DIff3F [18] features, which provide robust semantic descriptors for 3D shapes without requiring textured inputs or additional training. DIff3F extracts semantic features by utilizing foundational vision models in a zero-shot manner. The process begins by rendering depth and normal maps + +from multiple views of the input mesh. These maps serve as conditioning inputs to ControlNet [61] and Stable Diffusion [41] for generating view-dependent features. The features from the diffusion process maintain semantic consistency despite potential visual variations across views, and are aggregated onto the original 3D surface. These features, further enriched with DINO features [6], capture rich semantic information and enable reliable correspondence across shapes with significant geometric variations, making them well-suited for our texture transfer task. + +![](images/191c13ec86289e92de95d4feaf513168b7df132b17e3ccb6a116e817768a7e16.jpg) +Figure 3. Texture Transfer Results. Given a source mesh (left column), our network transfers its appearance onto target meshes of varying shapes (three right columns). Each row demonstrates texture transfer using a different source mesh. + +# 3.2. Our approach: TRiTEX + +Given a textured source mesh, our goal is to train a feedforward texturing function that can transfer that texture to any target mesh while preserving the semantics of the texture. Our pipeline is shown in Fig. 2. + +The architecture, based on a deep neural network, processes a 3D mesh with pre-extracted semantic features and a query 3D point, and outputs the corresponding color for the query point. Formally, we define a learnable function: $f(M,\mathbb{R}^3)\to [0,1]^3$ , where $M = (V,S,E,F)$ is the input mesh, with $V = \{v_{1},\ldots ,v_{n}\mid v_{i}\in \mathbb{R}^{3}\}$ and $S = \{s_1,\dots ,s_n\mid s_i\in \mathbb{R}^D\}$ , representing the 3D vertices and their corresponding semantic features, pre-extracted by [18]. Here, $E$ and $F$ represent the edges and faces, respectively, defined on the vertices. + +The first part of our architecture processes $M$ by orthog.raphically projecting the semantic features into an axisaligned feature triplane $\mathcal{T}\in \mathbb{R}^{3\times W\times H\times 2D}$ , where each feature plane is created by concatenating features from two orthographic projections taken from opposite directions. To address the low feature resolution $(32\times 32)$ and enable fine + +detail generation, we incorporate positional encoding into the input. Triplane-aware convolutional 2D blocks [53] (see further details in Sec.3.1) are then applied to the triplane to produce a modified triplane $\mathcal{T}'\in \mathbb{R}^{3\times W\times H\times D'}$ . Finally, to predict the color for a query 3D point, we sample features from the processed planes in $\mathcal{T}'$ according to the query point's location, concatenate them to form a single feature vector, and pass this vector through the coloring MLP $c:\mathbb{R}^{3D^{\prime}}\to [0,1]^{3}$ , which maps it to the final RGB color for the point. + +Training We train our model on a single textured mesh using a rendering-based reconstruction loss, comparing rendered views of our predicted textures with ground truth views of the source mesh from randomly sampled camera angles. Let $I_R(\theta)$ be an image rendered by our pipeline from angle $\theta$ , with the corresponding colors predicted by the MLP $c$ , generated by querying 3D points on the mesh through intersecting rendered camera rays. This image is conditioned only on the input semantic features defined for the geometry and depends on the differentiable parameters of the triplane-aware convolutional layers and $c$ . To train the learnable parameters, we use the ground truth rendered images, which are known for the single training textured mesh, and apply the MSE loss on the image pixels: + +$$ +\mathcal {L} _ {\mathrm {M S E}} (\theta) = \frac {1}{N} \sum_ {i = 1} ^ {N} \left\| I _ {R} (\theta) _ {i} - I _ {\mathrm {G T}} (\theta) _ {i} \right\| ^ {2}, +$$ + +where $I_{\mathrm{GT}}(\theta)$ is the ground truth image, and $i$ indexes the pixels in the image. While MSE loss encourages pixel-level accuracy, it may not effectively capture high-level perceptual details. To address this, we incorporate a perceptual loss $\mathcal{L}_{\mathrm{app}}$ from Tumanyan et al. [48], which emphasizes high-level semantic features. This improves the alignment of these features between the generated and ground truth images, enhancing texture realism. Our final loss function is then: + +$$ +\mathcal {L} = \mathbb {E} _ {\theta} \left[ \mathcal {L} _ {\mathrm {M S E}} (\theta) + \delta_ {\mathrm {a p p}} \mathcal {L} _ {\mathrm {a p p}} (\theta) \right], +$$ + +where $\mathbb{E}_{\theta}$ represents the expectation over all sampled camera angles, and $\delta_{\mathrm{app}}$ controls the relative weight of $\mathcal{L}_{\mathrm{app}}$ compared to $\mathcal{L}_{\mathrm{MSE}}$ . To improve generalization, we apply two levels of augmentation: (1) preprocessing augmentation, where simple 3D transformations are applied to the input mesh and features are extracted for each variant to enrich the learned feature distribution, and (2) training-time augmentation, where translation, scaling, and small rotational perturbations are applied to the mesh during training. + +![](images/020e986b90bdbcc1eea4a305809a10672b4f170b7f36fb5b960a3b57b2ba9860.jpg) + +![](images/e807f3b405a93ee120320bbac904a13387831f6a142a74910c5740eab62cc84a.jpg) + +![](images/cccc49c311f795cc803020a539811c63b394159f1e159245f2124737ae4b8254.jpg) + +![](images/a1bec29323051dd3f141ceafe8c8529f069e52c3b88888bb50396437e02e74d8.jpg) +Source-textured + +![](images/7bc5a9a3027125bf9343382e026f7b073a24587a5c2316e40321db4463efa9ef.jpg) + +![](images/7e75364c148897b62396bbbfe168a1a1f1d8809671d04fe86c66c937121c5cb0.jpg) + +![](images/c3b13a4ad5e2c3e58d86e91dfb43c30b2f1c3d3f7d5300d77a90dd68ac7fb652.jpg) + +![](images/02f9d169cf28215295763e0e9ab51e5f16335e522622192b9e2ab4a5337797c5.jpg) +Target geometry + +![](images/731d2c7d06f75c86e93e7c515096ab7fda78cace3d705842b135e61449eab528.jpg) + +![](images/dac90a887d3e9e92046f2a95ac3bce08d3cd19c4b260c6a63ff5a90548bc820e.jpg) + +![](images/c1ae6ced5be668c918f136fdba68d1e2334d4e1a9658b652fffee60aaf7955b5.jpg) + +![](images/c63d101a1cf046cc225f3c1b446ec5a2e3cf14bc7bd031e4eeb8222c5e414e9d.jpg) +Target-textured +Figure 4. Additional Qualitative Results. Showing the target geometry and the texture transfer. + +# 4. Experiments And Results + +# 4.1. Experiment Details + +Dataset: Our experimental dataset consists of 52 3D meshes across 9 categories, curated from Objaverse [15, 16], a large-scale 3D object database. We first generated descriptive captions for each 3D object to identify semantically related objects using BLIP [27]. We then used keyword filtering on these captions to group objects into similar categories. The final selection includes only objects that both share semantic similarities and maintain consistent orientations within their respective categories. Overall, we have 256 different pairs of source and target objects. The full list of source and target objects is provided in the supplemental. + +Baseline Methods: We compare our method against three state-of-the-art optimization-free approaches for visually-guided mesh texturing. (1) TEXTure [40], a personalization-based approach, enabling texture transfer from a source mesh to a target mesh. (2) EASI-TEX [37] and (3) MVEdit [9] use IP-Adapter to incorporate image guidance into their texture generation process, allowing direct control over the generated textures through reference images. + +Metrics: We evaluate our method using human evaluation and two automated metrics: CLIP similarity score [39] + +![](images/8250877c7d10b6f0cb958cd8a36e759fcf0d57a27983e94c0ca0f5134112356d.jpg) +Figure 5. Qualitative Comparisons with baselines. A comparison between TRiTEX (right-most column) and state-of-the-art approaches for texture transfer. As observed, TRiTEX produces much more plausible texture transfer results. + +Table 1. Quantitative Evaluation. Quantitative comparison of our method with baseline methods. We evaluate two appearance metrics and running time, showing that our method best preserves the source texture while achieving competitive speed. + +
MethodSIFID ↓CLIP sim. ↑Inference Runtime
TEXTure0.340.845 min.
MVEdit0.380.841 min.
EASI-TEX0.290.8515 min.
TriTex (ours)0.220.871 min.
+ +and Single Image Frechet Inception Distance (SIFID). The CLIP score, adopted following TextureDreamer [57], measures the semantic and visual similarity between source and generated textures, indicating how well our method preserves the desired appearance. SIFID was originally proposed by [43] to measure internal patch statistics within single images. Following Sin3DM [53], we adopt this metric to evaluate texture similarity between source and target shapes. + +For both metrics, we render each source and target mesh + +from the same 10 fixed viewpoints and compare the corresponding views against each other. To ensure balanced evaluation across our dataset, we first compute the average score within each object category, then average these category scores to obtain the final metric values. + +Human Evaluation To complement these quantitative metrics, we also conduct a user study on Amazon Mechanical Turk (AMT) with 15 participants to evaluate our results. Each participant viewed a source object and two versions of the textured target mesh from different methods, selecting which result better preserves the texture and the appearance of the reference object. Each comparison received ratings from two independent evaluators. + +# 4.2. Results + +We provide qualitative examples, qualitative and quantitative comparisons with baselines, and an ablation study with qualitative and quantitative analysis. + +Qualitative Examples Figure 3 provides four examples illustrating how TRITEX transfers texture from the source + +shape (left column) to three other shapes. Notice how colors match the semantics of the location, despite significant shape variations. for example the eyes of the fish at the bottom row. Figure 4 shows additional results along with the corresponding target geometries. + +Quantitative Comparisons: Table 1 compares our TRI-TEX approach with all baseline methods. We find that TRI-TEX outperforms baselines in both evaluation metrics, successfully capturing the appearance of the original objects. We also report the inference time for texturing a new mesh for each method. + +Qualitative Comparisons: We show qualitative comparisons between our approach and the baselines in Fig. 5. As demonstrated, MVEdit tends to deviate significantly from the input image, only taking vague inspiration from the reference. TEXTure and EASI-TEX produce artifacts due to their iterative painting techniques. + +![](images/31c498d805cea7c48e04923f4e524059c27ed74315707c957884451541e29d44.jpg) +Figure 6. Results of human evaluation study. TRITEX was compared with three baseline approaches in a 2-alternative forced choice setting over $\sim 750$ questions. Raters strongly favored TRI-TEX for better transfer of appearance. + +User Study: We conducted a user study to evaluate the quality of TRiTEX compared with baselines. For each baseline, we presented the source object and two target shapes, one textured with TRiTEX and one with the baseline, in a 2-alternative forced choice (2AFC) setting. We used Amazon Mechanical Turk (AMT) and paid raters above the minimum wage. Figure 6 shows the results of these comparisons. TRiTEX was strongly preferred by raters over all three competing approaches. + +# 4.3. Ablation Studies + +We conducted an ablation study to evaluate the contribution of each component in our method. The quantitative and quantitative results presented in Table 2 and Fig. 7 respectively, demonstrate that each component is essential for optimal performance. Our analysis reveals that excluding $\mathcal{L}_{\mathrm{mse}}$ leads to color predictions that deviate significantly from the original texture. Similarly, removing $\mathcal{L}_{\mathrm{app}}$ results in blurry outputs that fail to capture complex texture patterns. We further found that omitting data augmentation + +causes the network to overfit to specific triplane projections, thereby limiting the method's generalization capabilities. + +![](images/1020060e01d90da40c21efb441622f18e0b3e7047b9bc5a7b90b93aa0af3f8a1.jpg) +source shape + +![](images/1af02c454f246e5ab4655d6287ab298d4559e4ce1048236ef3a7969e5bdbea36.jpg) +w/o network + +![](images/dfb810799e959cbf95d6b37aeb2d70fcad42071fe9d4136d368a0a42d384ac7a.jpg) +w/o $\mathcal{L}_{\mathrm{app}}$ + +![](images/9a6c8780c8a050a3f08152540307b729bcf5ca6b4266e415f7497b54f2e37a67.jpg) +w/o $\mathcal{L}_{\mathrm{MSE}}$ + +![](images/cb9de220c05d58891571601679b5b2328c320e277c5c719e44624ef4549beb57.jpg) +w/o augmentations + +![](images/5904ea246bf966148954ad7a074e0c70dec64bbcd01578e248f6df65defa4010.jpg) +ours +Figure 7. Qualitative Ablation Study. We demonstrate the effect of removing each component of our pipeline on a single example. As shown, each component is essential for successful texture transfer. + +We also ablate the contribution of the neural network in our pipeline. In the "w/o network" baseline, we utilize pre-extracted DIFF3f semantic features and identify the nearest neighbor for each target mesh feature within the source mesh features. Using these nearest-neighbor matches, we transfer colors from the source mesh to the target vertices and then render the target mesh with its new colors. The impact of this simplification is presented in Table 2 and Fig.7, demonstrating the degraded quality of this approach in transferring the texture details compared to our full pipeline. Two additional examples are shown in Fig.8. + +![](images/4702cb30bb4e27becee8a4cdb3aaaa242395324d3d24d600a3cf1161eb2817ef.jpg) + +![](images/2037a3d849c10182be11fcb34db3737ad581e2d2cb2f733e82412a9e05242f48.jpg) + +![](images/d75a3b477558a3e543c7022381db7f719540f7e8dda700729a5f9eff076ba1f6.jpg) + +![](images/7fe0935e5b8f369af087b3a3cb5f54ff1c0da46ecf73d1d3cbead6e69c1aa498.jpg) +source shape +Figure 8. Network Ablation. A comparison between our full method (right) and the nearest neighboring baseline (middle), which replaces the neural network in our pipeline and transfers color based on the closest matching neighbors. As seen, this baseline fails to transfer complex texture details, unlike our full method. + +![](images/ad86fe1defffaac39c926a2511599734476a18bc1d845a8c9efbe10b2286e1f8.jpg) +w/o network + +![](images/5007931271fc63995ecd1e73540dfc38cf35b217f66cf2b770e369db9dbc646a.jpg) +ours + +Table 2. Quantitative Ablation Study. We demonstrate the effect of removing component of TRiTEX with quantitative metrics averaged across our evaluation set. The results show that each component is essential for achieving successful texture transfer. + +
MethodSIFID ↓CLIP sim. ↑
w/o network0.230.86
w/o L MSE0.230.86
w/o L app0.280.85
w/o augmentations0.210.85
TriTex (ours)0.220.87
+ +# 5. Discussion + +Cross-Category Object Transfer Although our methods is desinged to transfer appearance between objects from the same category, it is interesting to examine the results when transferring appearance between objects with varying degrees of semantic similarity. As shown in Figure 9, when transferring between objects with shared semantic structure (e.g. parts with similar functionality or spatial relationships), TRiTEX preserves meaningful texture patterns aligned with these semantic features. However, when transferring between objects with little semantic overlap, the resulting texture patterns become more stochastic, as the semantic mapping becomes less well-defined. + +![](images/9fd75269493e8d5de6cd21b38cdb54af3eb101932ed9c3585551032583e2ccab.jpg) +Source Shape + +![](images/6553c32abc2bf514ed5be13f4a9760905d2343239d5ad891b19cf5ac9a696887.jpg) +Output 1 + +![](images/36ca59a3ad5432db8c89afec9ca3b3344a94d95fa94e9c86dcbe701fd1f8a50d.jpg) +Output 2 +Figure 9. Cross Category Texture Transfer. Texture transfer between objects with partial semantic correspondence (fish and deer) and versus objects with minimal semantic similarity (fish and vase with flowers). + +Limitations Despite the effectiveness of our approach, it has several inherent limitations. First, our method does not possess generative capabilities, which means it cannot synthesize novel details that are not present in the source texture. This limitation becomes particularly apparent when there is significant geometric variation between the source and target meshes, as our method can only map existing texture patterns rather than create new ones to accommodate the structural differences (Figure 10, top). Second, without leveraging priors from text-to-image models, our approach lacks the ability to compensate for cases where the semantic features fail to capture sufficient detail or establish accurate + +correspondences. This can result in less detailed or less coherent texture transfers in regions where the feature matching is ambiguous or imprecise (Figure 10, bottom). Additionally, our method expects the target shape to be at the same orientation as the source shape. When enabling arbitrary rotations during training, the model becomes invariant to the object's orientation, but the ability to reproduce fine details such as eyes or wheels is reduced. + +![](images/2622a33392d9497f9c7894a1dee4ec6ef802ace47403fdb9ee78b0d703111a95.jpg) +Figure 10. Limitations. (Top) Since our network lacks a generative prior and the source dog has no visible tongue, it cannot synthesize an appropriate color for the target dog's tongue. (Bottom) The network incorrectly transfers the source bed's sheet texture to the target bed frame, as PCA analysis shows their feature representations are similar. + +![](images/a72f626d4dccf759b18e42f3476dd6aa7853dc5cd716202871a0fd7a6ae59d74.jpg) + +![](images/778bfdf5e9598588f9a8b27a0bfc68b274046d050429e125b4931388fb671036.jpg) + +![](images/0647318a04801b83506edae6c1c5661f6dde1298603ddd9f89020cb47bd17fa6.jpg) + +![](images/3be40d431b08c41bb52feb2ad84d6a37ba7de026a9c23d602d3b9a3c02313493.jpg) +source shape source shape PCA + +![](images/2a45642008f1a08aa894b4180bb3120152be99265466399b5656364826cd662d.jpg) +target shape + +![](images/ab964b66388e3faa19f2210bfb74e2a3c5a43cbf1fdb1696d3fd845ca67a4684.jpg) +target shape PCA + +![](images/1b93789865540a605dc0c15f4a55f8bd5d9a4a1c69239a1fc86a1eeb4a7f257e.jpg) + +# 6. Conclusions + +We have presented a method that allows transferring a texture learned from a single exemplar to similar shapes. The learned texture is challenging in the sense that it is a nonstationary texture, where its patterns can be strongly correlated with the semantics of the shapes. Mapping such semantic textures on a target shape, necessarily requires, at least implicitly, extracting its semantics. For that, our model uses pre-trained 3D lifted semantic features. + +Our key challenge was to process the semantic features in a way that allows generalization from a single instance, in order to perform the texture learning with minimal supervision. To achieve this, we re-rendered them into a triplane representation, which preserves feature proximity and connectivity, allowing for efficient processing and mapping to the corresponding colors. We introduced an effective training method using a single mesh with appearance losses while applying augmentations. The limitation of our technique is that it currently does not possess generative abilities and cannot create details that do not appear in the source texture. + +In future work, we aim to extend this feed-forward approach for fast texturing by training the model to predict the mesh texture from a reference image, using cross-attention layers between the generated triplane and the reference image. Additionally, we plan to leverage pre-trained generative models to enhance the mapped textures, increasing both resolution and quality with their learned priors. + +# Acknowledgements + +We thank Rinon Gal, Yoad Tewel and Or Patashnik for their constructive suggestions and insightful comments. + +# References + +[1] Kfir Aberman, Jing Liao, Mingyi Shi, Dani Lischinski, Baoquan Chen, and Daniel Cohen-Or. Neural best-buddies: Sparse cross-domain correspondence. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 3 +[2] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. arXiv e-prints, pages arXiv-2112, 2021. 3 +[3] Raphael Bensadoun, Yanir Kleiman, Idan Azuri, Omri Harosh, Andrea Vedaldi, Natalia Neverova, and Oran Gafni. Meta 3d texturegen: Fast and consistent texture generation for 3d objects, 2024. 2 +[4] Alexey Bokhovkin, Shubham Tulsiani, and Angela Dai. Mesh2tex: Generating mesh textures from image queries, 2023. 2 +[5] Tianshi Cao, Karsten Kreis, Sanja Fidler, Nicholas Sharp, and Kangxue Yin. Texfusion: Synthesizing 3d textures with text-guided image diffusion models, 2023. 1, 2 +[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3, 4 +[7] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16123-16133, 2022. 4 +[8] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models, 2023. 1, 2 +[9] Hansheng Chen, Ruoxi Shi, Yulin Liu, Bokui Shen, Jiayuan Gu, Gordon Wetzstein, Hao Su, and Leonidas Guibas. Generic 3d diffusion adapter using controlled multi-view editing, 2024. 1, 2, 5 +[10] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation, 2023. 2 +[11] Xiaobai Chen, Thomas Funkhouser, Dan B Goldman, and Eli Shechtman. Non-parametric texture transfer using MeshMatch. Adobe Technical Report 2012-2, 2012. 2 +[12] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition, 2022. 2 +[13] Zhiqin Chen, Kangxue Yin, and Sanja Fidler. Auv-net: Learning aligned uv maps for texture transfer and synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1465-1474, 2022. 1, 2, 3 + +[14] Jeremy S De Bonet. Multiresolution sampling procedure for analysis and synthesis of texture images. In Proceedings of the 24th annual conference on Computer graphics and interactive techniques, pages 361-368, 1997. 2 +[15] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects, 2022. 5 +[16] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects, 2023. 5 +[17] Kangle Deng, Timothy Omernick, Alexander Weiss, Deva Ramanan, Jun-Yan Zhu, Tinghui Zhou, and Maneesh Agrawala. *Flashtex: Fast reconfigurable mesh texturing with lightcontrolnet*, 2024. 2 +[18] Niladri Shekhar Dutt, Sanjeev Muralikrishnan, and Niloy J. Mitra. Diffusion 3d features (diff3f): Decorating textured shapes with distilled semantic features, 2024. 2, 3, 4 +[19] Alexei A Efros and Thomas K Leung. Texture synthesis by non-parametric sampling. In Proceedings of the seventh IEEE international conference on computer vision, pages 1033-1038. IEEE, 1999. 2 +[20] Michael Fischer, Zhengqin Li, Thu Nguyen-Phuoc, Aljaz Bozic, Zhao Dong, Carl Marshall, and Tobias Ritschel. Nerf analogies: Example-based visual attribute transfer for nerfs, 2024. 3 +[21] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2 +[22] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oguz. 3dgen: Triplane latent diffusion for textured mesh generation, 2023. 4 +[23] David J Heeger and James R Bergen. Pyramid-based texture analysis/synthesis. In Proceedings of the 22nd annual conference on Computer graphics and interactive techniques, pages 229–238, 1995. 2 +[24] Amir Hertz, Rana Hanocka, Raja Giryes, and Daniel Cohen-Or. Deep geometric texture synthesis. ACM Transactions on Graphics, 39(4), 2020. 3 +[25] Jaihoon Kim, Juil Koo, Kyeongmin Yeo, and Minhyuk Sung. Synctweedies: A general generative framework based on synchronized diffusions, 2024. 1, 2 +[26] Vladimir Kulikov, Shahar Yadin, Matan Kleiner, and Tomer Michaeli. Sinddm: A single image denoising diffusion model. In International Conference on Machine Learning, pages 17920-17930. PMLR, 2023. 3 +[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation, 2022. 5 +[28] Yuxin Liu, Minshan Xie, Hanyuan Liu, and Tien-Tsin Wong. Text-guided texturing by synchronized multi-view diffusion, 2023. 1, 2 + +[29] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 3 +[30] Jianye Lu, Athinodoros S. Georgiades, Andreas Glaser, Hongzhi Wu, Li-Yi Wei, Baining Guo, Julie Dorsey, and Holly Rushmeier. Context-aware textures. 26, 2007. 2 +[31] Tom Mertens, Jan Kautz, Jiawen Chen, Philippe Bekaert, and Frédo Durand. Texture transfer using geometry correlation. Goslar, DEU, 2006. Eurographics Association. 2 +[32] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures, 2022. 1, 2 +[33] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes, 2021. 2 +[34] Thomas W. Mitchel, Carlos Esteves, and Ameesh Makadia. Single mesh diffusion models with field latents for texture generation, 2024. 3 +[35] Yaniv Nikankin, Niv Haim, and Michal Irani. Sinfusion: Training diffusion models on a single image or video. arXiv preprint arXiv:2211.11743, 2022. 3 +[36] Dolev Ofri-Amar, Michal Geyer, Yoni Kasten, and Tali Dekel. Neural congealing: Aligning images to a joint semantic atlas. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19403-19412, 2023. 3 +[37] Sai Raj Kishore Perla, Yizhi Wang, Ali Mahdavi-Amiri, and Hao Zhang. Easi-tex: Edge-aware mesh texturing from single image, 2024. 2, 5 +[38] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2 +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 5 +[40] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes, 2023. 1, 2, 5 +[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 4 +[42] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2 +[43] Tamar Rott Shaham, Tali Dekel, and Tomer Michaeli. Singan: Learning a generative model from a single natural image, 2019. 3, 6 +[44] J. Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion, 2022. 4 +[45] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating textures on 3d shape surfaces, 2022. 1, 2 + +[46] Edgar Simo-Serra, Eduard Trulls, Luis Ferraz, Iasonas Kokkinos, Pascal Fua, and Francesc Moreno-Noguer. Discriminative learning of deep convolutional feature point descriptors. In Proceedings of the IEEE international conference on computer vision, pages 118-126, 2015. 3 +[47] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. Advances in Neural Information Processing Systems, 36:1363-1389, 2023. 3 +[48] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer, 2022. 5 +[49] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10748-10757, 2022. 3 +[50] Weilun Wang, Jianmin Bao, Wengang Zhou, Dongdong Chen, Dong Chen, Lu Yuan, and Houqiang Li. Sindiffusion: Learning a diffusion model from a single natural image. arXiv preprint arXiv:2211.12445, 2022. 3 +[51] Li-Yi Wei and Marc Levoy. Texture synthesis over arbitrary manifold surfaces. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 355-360, 2001. 2 +[52] Rundi Wu and Changxi Zheng. Learning to generate 3d shapes from a single example. ACM Transactions on Graphics, 41(6):1-19, 2022. 3 +[53] Rundi Wu, Ruoshi Liu, Carl Vondrick, and Changxi Zheng. Sin3dm: Learning a diffusion model from a single 3d textured shape, 2024. 3, 4, 5, 6, 12 +[54] Wenqi Xian, Patsorn Sangkloy, Varun Agrawal, Amit Raj, Jingwan Lu, Chen Fang, Fisher Yu, and James Hays. Texturegan: Controlling deep image synthesis with texture patches. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8456-8465, 2018. 2 +[55] Jiale Xu, Xintao Wang, Weihao Cheng, Yan-Pei Cao, Ying Shan, Xiaohu Qie, and Shenghua Gao. Dream3d: Zero-shot text-to-3d synthesis using 3d shape prior and text-to-image diffusion models, 2023. 2 +[56] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2 +[57] Yu-Ying Yeh, Jia-Bin Huang, Changil Kim, Lei Xiao, Thu Nguyen-Phuoc, Numair Khan, Cheng Zhang, Manmohan Chandraker, Carl S Marshall, Zhao Dong, and Zhengqin Li. Texturedreamer: Image-guided texture synthesis through geometry-aware diffusion, 2024. 1, 2, 6 +[58] Kangxue Yin, Jun Gao, Maria Shugrina, Sameh Khamis, and Sanja Fidler. 3dstylenet: Creating 3d shapes with geometric and texture style variations, 2021. 2 +[59] Kim Youwang, Tae-Hyun Oh, and Gerard Pons-Moll. Paint-it: Text-to-texture synthesis via deep convolutional texture map optimization and physically-based rendering, 2024. 2 +[60] Xianfang Zeng, Xin Chen, Zhongqi Qi, Wen Liu, Zibo Zhao, Zhibin Wang, Bin Fu, Yong Liu, and Gang Yu. Paint3d: + +Paint anything 3d with lighting-less texture diffusion models, 2023. 1, 2 +[61] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 4 +[62] Yang Zhou, Zhen Zhu, Xiang Bai, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Non-stationary texture synthesis by adversarial expansion. arXiv preprint arXiv:1805.04487, 2018. 2 +[63] Song Chun Zhu, Yingnian Wu, and David Mumford. Filters, random fields and maximum entropy (frame): Towards a unified theory for texture modeling. International Journal of Computer Vision, 27:107-126, 1998. 2 \ No newline at end of file diff --git a/2025/TriTex_ Learning Texture from a Single Mesh via Triplane Semantic Features/images.zip b/2025/TriTex_ Learning Texture from a Single Mesh via Triplane Semantic Features/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f1e78252f6abf2476edd2896c9c97b1a1baaa12e --- /dev/null +++ b/2025/TriTex_ Learning Texture from a Single Mesh via Triplane Semantic Features/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:118d7463dc957001fb00b74a4c4921edde8732d7b480a72650d02dfe51fcb694 +size 417695 diff --git a/2025/TriTex_ Learning Texture from a Single Mesh via Triplane Semantic Features/layout.json b/2025/TriTex_ Learning Texture from a Single Mesh via Triplane Semantic Features/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2bb044b5f2595a34002f6323312bf3c5b7df1e5a --- /dev/null +++ b/2025/TriTex_ Learning Texture from a Single Mesh via Triplane Semantic Features/layout.json @@ -0,0 +1,9139 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 103, + 544, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 103, + 544, + 121 + ], + "spans": [ + { + "bbox": [ + 67, + 103, + 544, + 121 + ], + "type": "text", + "content": "TriTex: Learning Texture from a Single Mesh via Triplane Semantic Features" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 66, + 180, + 211, + 314 + ], + "blocks": [ + { + "bbox": [ + 66, + 180, + 211, + 314 + ], + "lines": [ + { + "bbox": [ + 66, + 180, + 211, + 314 + ], + "spans": [ + { + "bbox": [ + 66, + 180, + 211, + 314 + ], + "type": "image", + "image_path": "746f54f9680bfb83d22012db1d92199bd85a1d28f9829e6c65e8024bd14d74b3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 220, + 178, + 302, + 307 + ], + "blocks": [ + { + "bbox": [ + 113, + 142, + 208, + 156 + ], + "lines": [ + { + "bbox": [ + 113, + 142, + 208, + 156 + ], + "spans": [ + { + "bbox": [ + 113, + 142, + 208, + 156 + ], + "type": "text", + "content": "Dana Cohen-Bar1,2" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 219, + 142, + 320, + 176 + ], + "lines": [ + { + "bbox": [ + 219, + 142, + 320, + 176 + ], + "spans": [ + { + "bbox": [ + 219, + 142, + 320, + 176 + ], + "type": "text", + "content": "Daniel Cohen-Or1 \n1Tel Aviv University" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 220, + 178, + 302, + 307 + ], + "lines": [ + { + "bbox": [ + 220, + 178, + 302, + 307 + ], + "spans": [ + { + "bbox": [ + 220, + 178, + 302, + 307 + ], + "type": "image", + "image_path": "4ebda9dd5a6b95cd194b2fd276af872402e33dad9782ae3bcba8f76c5e0c9562.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 326, + 182, + 422, + 318 + ], + "blocks": [ + { + "bbox": [ + 326, + 182, + 422, + 318 + ], + "lines": [ + { + "bbox": [ + 326, + 182, + 422, + 318 + ], + "spans": [ + { + "bbox": [ + 326, + 182, + 422, + 318 + ], + "type": "image", + "image_path": "558702c05ec899e19f121491cd4ee1aadc6f988f36124a723b46de325b730729.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 447, + 176, + 553, + 322 + ], + "blocks": [ + { + "bbox": [ + 340, + 142, + 493, + 175 + ], + "lines": [ + { + "bbox": [ + 340, + 142, + 493, + 175 + ], + "spans": [ + { + "bbox": [ + 340, + 142, + 493, + 175 + ], + "type": "text", + "content": "Gal Chechik2 Yoni Kasten2 NVIDIA" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 447, + 176, + 553, + 322 + ], + "lines": [ + { + "bbox": [ + 447, + 176, + 553, + 322 + ], + "spans": [ + { + "bbox": [ + 447, + 176, + 553, + 322 + ], + "type": "image", + "image_path": "835465995a9e2f432a28c969237005ad46254bc663548612cdf2a03e9f69c751.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 331, + 555, + 366 + ], + "lines": [ + { + "bbox": [ + 54, + 331, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 54, + 331, + 555, + 366 + ], + "type": "text", + "content": "Figure 1. TriTex is a method for learning to transfer a texture from a single mesh in a feed-forward manner to other input meshes. The figure shows four examples. In each, given a target geometry (up left) and a source textured mesh (up right) the texture is transferred to the target 3D object (bottom two views of the target object)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 151, + 375, + 200, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 375, + 200, + 388 + ], + "spans": [ + { + "bbox": [ + 151, + 375, + 200, + 388 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 402, + 297, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 297, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 297, + 628 + ], + "type": "text", + "content": "As 3D content creation continues to grow, transferring semantic textures between 3D meshes remains a significant challenge in computer graphics. While recent methods leverage text-to-image diffusion models for texturing, they often struggle to preserve the appearance of the source texture during texture transfer. We present TRiTEX, a novel approach that learns a volumetric texture field from a single textured mesh by mapping semantic features to surface colors. Using an efficient triplane-based architecture, our method enables semantic-aware texture transfer to a novel target mesh. Despite training on just one example, it generalizes effectively to diverse shapes within the same category. Extensive evaluation on our newly created benchmark dataset shows that TRiTEX achieves superior texture transfer quality and fast inference times compared to existing methods. Our approach advances single-example texture transfer, providing a practical solution for maintaining visual coherence across related 3D models in applications like game development and simulation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 629, + 268, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 268, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 268, + 641 + ], + "type": "text", + "content": "Project page: https://danacohen95.github.io/TriTex/." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 668, + 135, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 668, + 135, + 680 + ], + "spans": [ + { + "bbox": [ + 55, + 668, + 135, + 680 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Texturing 3D objects is a fundamental task with wideranging applications in game development, simulation, and" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 376, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 555 + ], + "type": "text", + "content": "video production. For example, when generating 3D scenes, style needs to be consistently applied across multiple 3D models that share semantic properties but differ in shape, such as in environments featuring diverse buildings or plants. Improving texture transfer can streamline texturing workflows by maintaining visual coherence across related models and preserving texture details even during mesh modifications, ensuring a consistent appearance across all elements. With the advent of large generative models, new approaches to 3D texturing have emerged, moving beyond traditional procedural methods to enable automatic texturing and texture transfer [13, 45]. However, these techniques often require the availability of large databases of 3D objects and extensive training for each class to achieve high-quality results." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 556, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 556, + 665 + ], + "type": "text", + "content": "Recent texturing methods use text-to-image diffusion models to texture meshes without requiring 3D data collection. Some methods [32, 57, 60] apply an SDS loss to optimize the texture map, which tends to be a slow process. Others [5, 8, 9, 25, 28, 40] employ iterative depth-conditioned image inpainting on rendered mesh views, synchronizing across views to maintain consistency. However, when performing texture transfer, these methods struggle to faithfully preserve the appearance of the source texture." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": "We introduce TRITEx, a novel technique that learns a volumetric texture field from a single textured mesh, mapping its semantic features to RGB colors on its surface. During inference, given a target mesh, this learned field en" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21403" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": "ables texture transfer by predicting colors based on the target mesh's semantic features, allowing the source texture to be applied in a way that maintains semantic correspondences between the source and target meshes (see Figure 1). Despite being trained on a single object, our trained volumetric texture generalizes well to novel objects within the same category. Our fast inference time enables the stylization of large scenes More specifically, to learn this texture field, we first extract semantic features using a frozen Diff3F [18] model, then project these features onto a triplane, and process them through a mapping network to produce the final triplane features." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 216, + 294, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 294, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 294, + 310 + ], + "type": "text", + "content": "Our experiments demonstrate the effectiveness of our texture transfer method, showing superior quality compared to previous approaches while maintaining fast inference times. We evaluate our approach using automated metrics and human evaluation on a benchmark dataset that we created specifically for the texture transfer task. Our method shows superior performance compared to previous approaches." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 312, + 294, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 294, + 443 + ], + "type": "text", + "content": "To summarize, our contributions are: (1) Introducing a method for transferring textures from a single textured mesh to new shapes while preserving semantic relations. (2) Leveraging pre-trained 3D semantic features, reprojecting them into a triplane representation, and enabling effective processing through convolutional layers. (3) Demonstrating strong generalization to novel objects with significant shape variations, despite being trained on only a single example. (4) Achieving high-quality texture transfer with fast inference, outperforming previous methods in both speed and quality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 452, + 142, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 452, + 142, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 452, + 142, + 464 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 472, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 652 + ], + "type": "text", + "content": "Texture Synthesis Classical methods use sampling approaches to generate 2D textures [14, 19, 23, 51, 63]. More recent approaches rely on deep learning to create textures based on training data [54, 62]. Texturify [45] introduced a Generative Adversarial Network (GAN) for texturing 3D models using an example image, applying face convolutional operators directly on the 3D object's surface. Auv-net [13] predicts UV mapping and texture images from 3D geometry, ensuring consistent alignment across elements within the same category, which enables effective training of generative texture models with this embedding. Both Auv-net and Texturify require large datasets of predefined categories to train their generative models, whereas our approach is trained solely on a single textured mesh and can generalize to a variety of objects with similar semantics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 715 + ], + "type": "text", + "content": "Leveraging Large Vision-Language Pretrained Models for Texture Generation Recent advances in image-language foundation models have enabled effective texture transfer and synthesis without requiring additional training" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": "data. Text2Mesh [33] and TANGO [12] edit 3D mesh appearance by aligning rendered images with a text prompt in CLIP space. Dream3D [55] optimizes the appearance of generated geometry using a Neural Radiance Field (NeRF) by applying a CLIP loss. Latent-NeRF [32] applies the SDS-loss [38] to rendered images to optimize a NeRF representation in the stable diffusion latent space. More recently, Paint-it [59], Fantasia3D [10], and FlashTex [17] have enhanced SDS-loss-based texturing by incorporating Physically Based Rendering (PBR), BRDF modeling, and illumination control, respectively. These optimization techniques require processing each object individually, leading to slower performance, whereas our method has a much shorter inference time." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 240, + 553, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 553, + 407 + ], + "type": "text", + "content": "Other methods employed optimization-free approaches by directly applying depth-conditioned diffusion models in sequence. Texture [40] applies depth-conditioned inpainting on sequential mesh projections to achieve mesh texturing. Concurrent and follow-up methods [5, 8, 9, 25, 28] introduced improvements, with Text2Tex[8] proposing an automatic per-object view selection scheme, and TexFusion [5], SyncMVD [28], SyncTweedies [25] and MVEdit [9] synchronizing views by operating on intermediate denoising steps. Paint3D [60] additionally trains a model to remove illumination effects from the generated texture map, enabling relighting of the textured meshes. In [3], a geometry-aware multiview diffusion model is used to enhance view consistency and efficiency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "text", + "content": "Texture Transfer The problem of texture transfer has always been an area of interest and research, with early work aiming to transfer both stochastic texture [30] and semantic texture [11, 31]. More recently, several methods leverage 3D data for texture transfer tasks. 3DStyleNet [58] transforms a source mesh by applying part-aware low-frequency deformations and generating texture maps, using a target mesh as a style reference. Mesh2Tex [4] learns a texture manifold from a large dataset of 3D objects and images. This enables new images to be projected into the manifold and transferred to 3D objects in trained categories, such as cars and chairs. Personalization techniques for diffusion models [21, 42] are used in methods like TEXTure and TextureDreamer [40, 57] to texture a mesh from a few input images, which can be either natural images or rendered images from another textured mesh. MVEdit [9], EASI-Tex [37] and Paint3D [60] transfer textures from an image to a 3D object using IP-Adapter [56], where EASI-Tex further enhances geometry details by incorporating a rendered mesh edge map as an additional control for image generation. While these methods are efficient, they often fail to faithfully preserve the appearance of the source texture. In contrast, our method maintains both efficiency and accurate texture preservation." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21404" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 68, + 541, + 327 + ], + "blocks": [ + { + "bbox": [ + 65, + 68, + 541, + 327 + ], + "lines": [ + { + "bbox": [ + 65, + 68, + 541, + 327 + ], + "spans": [ + { + "bbox": [ + 65, + 68, + 541, + 327 + ], + "type": "image", + "image_path": "39d8cb70cb3987e700167c4773a278db52a6d35e48a246a4f33d18ea13ff35b0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 337, + 555, + 394 + ], + "lines": [ + { + "bbox": [ + 54, + 337, + 555, + 394 + ], + "spans": [ + { + "bbox": [ + 54, + 337, + 555, + 394 + ], + "type": "text", + "content": "Figure 2. Training Pipeline (Top). Given an input textured mesh and its pre-extracted Diff3F features, we project six orthographic views to create an initial triplane. This triplane is processed using triplane-aware convolutional blocks, which, along with the texture MLP, define a coloring neural field. This field, together with the input geometry, is used to render the colored mesh. Appearance losses are applied between the true mesh appearance and the rendered appearance. Inference (Bottom). Given a new mesh (left), our pre-trained model maps its semantic properties to colors (right), transferring the texture from the original textured mesh learned during the training phase." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 406, + 298, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 406, + 298, + 525 + ], + "spans": [ + { + "bbox": [ + 54, + 406, + 298, + 525 + ], + "type": "text", + "content": "Auv-net [13] enables texture transfer between meshes through a learned UV mapping and aligned texture representations. Mitchel et al. [34] proposed a diffusion process within a learned latent space on the surface, allowing the transferring of textures to other meshes using an input semantic label map. Nerf Analogies [20] extends texture transfer to neural radiance fields, enabling appearance transfer between source and target NeRFs with semantic similarities, though it requires per-pair training for each set of NeRFs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 545, + 298, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 545, + 298, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 545, + 298, + 715 + ], + "type": "text", + "content": "Semantic Features and Correspondences Finding corresponding points between images has traditionally relied on local feature descriptors [29] that are invariant to lighting and color. Later methods used features extracted from pretrained classification models [1, 46], providing invariance to local shape deformations and leveraging higher-level semantic relationships. Recently, Amir et al. [2] demonstrated the effectiveness of features from the pre-trained Vision Transformer DINO-ViT [6] as semantic descriptors. Splice [49] uses DINO-ViT features to transfer colors from a source image to a target image structure by preserving the target's local DINO features while adapting its global appearance to match the source. Neural Congealing [36] employs DINO-ViT semantic features to map a set of se" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 406, + 556, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 556, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 556, + 515 + ], + "type": "text", + "content": "mentically related images into a shared canonical representation, enabling zero-shot joint editing. More recently, DIFT [47] introduced semantic features extracted from a pre-trained diffusion model during denoising. Diff3F [18] mapped both DIFT and DINO features onto a 3D mesh, resulting in 3D semantic features that enable shape correspondences. Our model benefits from these recent advances in semantic features and uses them to train a texture transfer model with a single textured mesh." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 533, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 557, + 715 + ], + "type": "text", + "content": "Models Trained on a Single Instance Several methods have been developed to train neural networks using a single example. SinGAN [43] trains a model to capture the internal distribution of patches within a single image, enabling the generation of diverse samples with similar visual content. Similarly, Wu et al. [52] and Hertz et al. [24] use GANs to generate shape and geometric texture variations, respectively, from a single mesh. More recent works [26, 35, 50] have employed diffusion models trained on individual images to produce image variations, with SinFusion [35] also extending this approach to single videos to create video variations. Additionally, Sin3DM [53] and [34] apply similar techniques to generate variations of a single textured mesh, either for textured mesh variations [53] or for texture variations [34]. Similarly, we propose training a" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21405" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 282, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 282, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 282, + 85 + ], + "type": "text", + "content": "mesh texturing model using just a single textured mesh." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 99, + 296, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 99, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 55, + 99, + 296, + 207 + ], + "type": "text", + "content": "Triplane Representation The triplane representation introduced in EG3D [7] consists of three 2D feature grids aligned with three distinct orthogonal planes in the 3D coordinate system. This setup enables image operations, such as convolutional neural layers, and uses interpolation to define a continuous feature grid suitable for neural fields. Triplane representation has been applied in 3D-aware GANs [7] and in mesh generation using diffusion models [22, 44]. For a detailed technical description, see Sec. 3.1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 218, + 111, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 218, + 111, + 230 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 111, + 230 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 238, + 135, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 135, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 135, + 251 + ], + "type": "text", + "content": "3.1. Background" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 255, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 422 + ], + "type": "text", + "content": "Triplane Representation The triplane representation, introduced in EG3D [7], encodes 3D information using three axis-aligned orthogonal feature planes " + }, + { + "bbox": [ + 55, + 255, + 296, + 422 + ], + "type": "inline_equation", + "content": "F_{XY}, F_{XZ}, F_{YZ}" + }, + { + "bbox": [ + 55, + 255, + 296, + 422 + ], + "type": "text", + "content": ", each of size " + }, + { + "bbox": [ + 55, + 255, + 296, + 422 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{W \\times H \\times C}" + }, + { + "bbox": [ + 55, + 255, + 296, + 422 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 255, + 296, + 422 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 255, + 296, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 255, + 296, + 422 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 255, + 296, + 422 + ], + "type": "text", + "content": " represent the spatial resolution, and " + }, + { + "bbox": [ + 55, + 255, + 296, + 422 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 255, + 296, + 422 + ], + "type": "text", + "content": " is the number of channels. To query features at any 3D position " + }, + { + "bbox": [ + 55, + 255, + 296, + 422 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 255, + 296, + 422 + ], + "type": "text", + "content": ", the coordinates are used to sample from each of the three feature planes (XY, XZ, YZ) via bilinear interpolation, yielding three feature vectors " + }, + { + "bbox": [ + 55, + 255, + 296, + 422 + ], + "type": "inline_equation", + "content": "f_{xy}, f_{xz}, f_{yz} \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 55, + 255, + 296, + 422 + ], + "type": "text", + "content": ". These feature vectors are then concatenated and passed through a lightweight MLP decoder network to produce the final output. This approach enables efficient feature extraction while maintaining expressiveness through learned feature integration, offering a balance between efficiency and quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 423, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 296, + 628 + ], + "type": "text", + "content": "Importantly, the triplane representation enables processing with standard 2D convolutional layers, which are far more efficient than using 3D convolutional layers on volumetric grids. In EG3D [7], 2D convolutional layers were applied on the three concatenated feature planes. Recently, Sin3DM [53] introduced the triplane-aware convolution block, suggesting a more geometrically integrated approach. Instead of processing the three feature grids " + }, + { + "bbox": [ + 55, + 423, + 296, + 628 + ], + "type": "inline_equation", + "content": "(F_{XY}, F_{XZ}, F_{YZ})" + }, + { + "bbox": [ + 55, + 423, + 296, + 628 + ], + "type": "text", + "content": " independently or simply concatenating them, this method aggregates features from each plane into the others. Let " + }, + { + "bbox": [ + 55, + 423, + 296, + 628 + ], + "type": "inline_equation", + "content": "F_{XY}'" + }, + { + "bbox": [ + 55, + 423, + 296, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 423, + 296, + 628 + ], + "type": "inline_equation", + "content": "F_{XZ}'" + }, + { + "bbox": [ + 55, + 423, + 296, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 423, + 296, + 628 + ], + "type": "inline_equation", + "content": "F_{YZ}'" + }, + { + "bbox": [ + 55, + 423, + 296, + 628 + ], + "type": "text", + "content": " represent the output of a single independent convolutional layer. Each plane is averaged over its axes and replicated along the third axis, before being concatenated into the relevant plane. For example, " + }, + { + "bbox": [ + 55, + 423, + 296, + 628 + ], + "type": "inline_equation", + "content": "F_{XY}'" + }, + { + "bbox": [ + 55, + 423, + 296, + 628 + ], + "type": "text", + "content": " is averaged over the " + }, + { + "bbox": [ + 55, + 423, + 296, + 628 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 423, + 296, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 423, + 296, + 628 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 423, + 296, + 628 + ], + "type": "text", + "content": " axes, producing " + }, + { + "bbox": [ + 55, + 423, + 296, + 628 + ], + "type": "inline_equation", + "content": "F_{Y}'" + }, + { + "bbox": [ + 55, + 423, + 296, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 423, + 296, + 628 + ], + "type": "inline_equation", + "content": "F_{X}'" + }, + { + "bbox": [ + 55, + 423, + 296, + 628 + ], + "type": "text", + "content": " respectively. These are then replicated along the " + }, + { + "bbox": [ + 55, + 423, + 296, + 628 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 423, + 296, + 628 + ], + "type": "text", + "content": "-axis and concatenated with " + }, + { + "bbox": [ + 55, + 423, + 296, + 628 + ], + "type": "inline_equation", + "content": "F_{YZ}'" + }, + { + "bbox": [ + 55, + 423, + 296, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 423, + 296, + 628 + ], + "type": "inline_equation", + "content": "F_{XZ}'" + }, + { + "bbox": [ + 55, + 423, + 296, + 628 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Diff3f Semantic Features Our method leverages DIff3F [18] features, which provide robust semantic descriptors for 3D shapes without requiring textured inputs or additional training. DIff3F extracts semantic features by utilizing foundational vision models in a zero-shot manner. The process begins by rendering depth and normal maps" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "text", + "content": "from multiple views of the input mesh. These maps serve as conditioning inputs to ControlNet [61] and Stable Diffusion [41] for generating view-dependent features. The features from the diffusion process maintain semantic consistency despite potential visual variations across views, and are aggregated onto the original 3D surface. These features, further enriched with DINO features [6], capture rich semantic information and enable reliable correspondence across shapes with significant geometric variations, making them well-suited for our texture transfer task." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 198, + 557, + 398 + ], + "blocks": [ + { + "bbox": [ + 318, + 198, + 557, + 398 + ], + "lines": [ + { + "bbox": [ + 318, + 198, + 557, + 398 + ], + "spans": [ + { + "bbox": [ + 318, + 198, + 557, + 398 + ], + "type": "image", + "image_path": "191c13ec86289e92de95d4feaf513168b7df132b17e3ccb6a116e817768a7e16.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 406, + 555, + 451 + ], + "lines": [ + { + "bbox": [ + 313, + 406, + 555, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 555, + 451 + ], + "type": "text", + "content": "Figure 3. Texture Transfer Results. Given a source mesh (left column), our network transfers its appearance onto target meshes of varying shapes (three right columns). Each row demonstrates texture transfer using a different source mesh." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 456, + 444, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 444, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 444, + 469 + ], + "type": "text", + "content": "3.2. Our approach: TRiTEX" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 474, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 521 + ], + "type": "text", + "content": "Given a textured source mesh, our goal is to train a feedforward texturing function that can transfer that texture to any target mesh while preserving the semantics of the texture. Our pipeline is shown in Fig. 2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 522, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 554, + 640 + ], + "type": "text", + "content": "The architecture, based on a deep neural network, processes a 3D mesh with pre-extracted semantic features and a query 3D point, and outputs the corresponding color for the query point. Formally, we define a learnable function: " + }, + { + "bbox": [ + 313, + 522, + 554, + 640 + ], + "type": "inline_equation", + "content": "f(M,\\mathbb{R}^3)\\to [0,1]^3" + }, + { + "bbox": [ + 313, + 522, + 554, + 640 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 522, + 554, + 640 + ], + "type": "inline_equation", + "content": "M = (V,S,E,F)" + }, + { + "bbox": [ + 313, + 522, + 554, + 640 + ], + "type": "text", + "content": " is the input mesh, with " + }, + { + "bbox": [ + 313, + 522, + 554, + 640 + ], + "type": "inline_equation", + "content": "V = \\{v_{1},\\ldots ,v_{n}\\mid v_{i}\\in \\mathbb{R}^{3}\\}" + }, + { + "bbox": [ + 313, + 522, + 554, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 522, + 554, + 640 + ], + "type": "inline_equation", + "content": "S = \\{s_1,\\dots ,s_n\\mid s_i\\in \\mathbb{R}^D\\}" + }, + { + "bbox": [ + 313, + 522, + 554, + 640 + ], + "type": "text", + "content": ", representing the 3D vertices and their corresponding semantic features, pre-extracted by [18]. Here, " + }, + { + "bbox": [ + 313, + 522, + 554, + 640 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 313, + 522, + 554, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 522, + 554, + 640 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 522, + 554, + 640 + ], + "type": "text", + "content": " represent the edges and faces, respectively, defined on the vertices." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "text", + "content": "The first part of our architecture processes " + }, + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "text", + "content": " by orthog.raphically projecting the semantic features into an axisaligned feature triplane " + }, + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{T}\\in \\mathbb{R}^{3\\times W\\times H\\times 2D}" + }, + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "text", + "content": " , where each feature plane is created by concatenating features from two orthographic projections taken from opposite directions. To address the low feature resolution " + }, + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "inline_equation", + "content": "(32\\times 32)" + }, + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "text", + "content": " and enable fine" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21406" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": "detail generation, we incorporate positional encoding into the input. Triplane-aware convolutional 2D blocks [53] (see further details in Sec.3.1) are then applied to the triplane to produce a modified triplane " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{T}'\\in \\mathbb{R}^{3\\times W\\times H\\times D'}" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": ". Finally, to predict the color for a query 3D point, we sample features from the processed planes in " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{T}'" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": " according to the query point's location, concatenate them to form a single feature vector, and pass this vector through the coloring MLP " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "c:\\mathbb{R}^{3D^{\\prime}}\\to [0,1]^{3}" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": ", which maps it to the final RGB color for the point." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 226, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 226, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 296, + 396 + ], + "type": "text", + "content": "Training We train our model on a single textured mesh using a rendering-based reconstruction loss, comparing rendered views of our predicted textures with ground truth views of the source mesh from randomly sampled camera angles. Let " + }, + { + "bbox": [ + 55, + 226, + 296, + 396 + ], + "type": "inline_equation", + "content": "I_R(\\theta)" + }, + { + "bbox": [ + 55, + 226, + 296, + 396 + ], + "type": "text", + "content": " be an image rendered by our pipeline from angle " + }, + { + "bbox": [ + 55, + 226, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 226, + 296, + 396 + ], + "type": "text", + "content": ", with the corresponding colors predicted by the MLP " + }, + { + "bbox": [ + 55, + 226, + 296, + 396 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 226, + 296, + 396 + ], + "type": "text", + "content": ", generated by querying 3D points on the mesh through intersecting rendered camera rays. This image is conditioned only on the input semantic features defined for the geometry and depends on the differentiable parameters of the triplane-aware convolutional layers and " + }, + { + "bbox": [ + 55, + 226, + 296, + 396 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 226, + 296, + 396 + ], + "type": "text", + "content": ". To train the learnable parameters, we use the ground truth rendered images, which are known for the single training textured mesh, and apply the MSE loss on the image pixels:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 91, + 409, + 260, + 443 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 409, + 260, + 443 + ], + "spans": [ + { + "bbox": [ + 91, + 409, + 260, + 443 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M S E}} (\\theta) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left\\| I _ {R} (\\theta) _ {i} - I _ {\\mathrm {G T}} (\\theta) _ {i} \\right\\| ^ {2},", + "image_path": "c5a8dc13622bc49ae94494142730c4b8966d9c74b15737b18b95346678299d50.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 458, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 458, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 458, + 296, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 458, + 296, + 567 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{GT}}(\\theta)" + }, + { + "bbox": [ + 55, + 458, + 296, + 567 + ], + "type": "text", + "content": " is the ground truth image, and " + }, + { + "bbox": [ + 55, + 458, + 296, + 567 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 458, + 296, + 567 + ], + "type": "text", + "content": " indexes the pixels in the image. While MSE loss encourages pixel-level accuracy, it may not effectively capture high-level perceptual details. To address this, we incorporate a perceptual loss " + }, + { + "bbox": [ + 55, + 458, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{app}}" + }, + { + "bbox": [ + 55, + 458, + 296, + 567 + ], + "type": "text", + "content": " from Tumanyan et al. [48], which emphasizes high-level semantic features. This improves the alignment of these features between the generated and ground truth images, enhancing texture realism. Our final loss function is then:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 577, + 245, + 591 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 577, + 245, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 245, + 591 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {\\theta} \\left[ \\mathcal {L} _ {\\mathrm {M S E}} (\\theta) + \\delta_ {\\mathrm {a p p}} \\mathcal {L} _ {\\mathrm {a p p}} (\\theta) \\right],", + "image_path": "c89b94859b131fca205e2598742b37039c90e3088086c773e8065bb06f853ba4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\theta}" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " represents the expectation over all sampled camera angles, and " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{app}}" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " controls the relative weight of " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{app}}" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " compared to " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MSE}}" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": ". To improve generalization, we apply two levels of augmentation: (1) preprocessing augmentation, where simple 3D transformations are applied to the input mesh and features are extracted for each variant to enrich the learned feature distribution, and (2) training-time augmentation, where translation, scaling, and small rotational perturbations are applied to the mesh during training." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 319, + 86, + 388, + 118 + ], + "blocks": [ + { + "bbox": [ + 319, + 86, + 388, + 118 + ], + "lines": [ + { + "bbox": [ + 319, + 86, + 388, + 118 + ], + "spans": [ + { + "bbox": [ + 319, + 86, + 388, + 118 + ], + "type": "image", + "image_path": "020e986b90bdbcc1eea4a305809a10672b4f170b7f36fb5b960a3b57b2ba9860.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 342, + 125, + 363, + 188 + ], + "blocks": [ + { + "bbox": [ + 342, + 125, + 363, + 188 + ], + "lines": [ + { + "bbox": [ + 342, + 125, + 363, + 188 + ], + "spans": [ + { + "bbox": [ + 342, + 125, + 363, + 188 + ], + "type": "image", + "image_path": "e807f3b405a93ee120320bbac904a13387831f6a142a74910c5740eab62cc84a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 329, + 193, + 384, + 242 + ], + "blocks": [ + { + "bbox": [ + 329, + 193, + 384, + 242 + ], + "lines": [ + { + "bbox": [ + 329, + 193, + 384, + 242 + ], + "spans": [ + { + "bbox": [ + 329, + 193, + 384, + 242 + ], + "type": "image", + "image_path": "cccc49c311f795cc803020a539811c63b394159f1e159245f2124737ae4b8254.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 324, + 262, + 384, + 292 + ], + "blocks": [ + { + "bbox": [ + 324, + 262, + 384, + 292 + ], + "lines": [ + { + "bbox": [ + 324, + 262, + 384, + 292 + ], + "spans": [ + { + "bbox": [ + 324, + 262, + 384, + 292 + ], + "type": "image", + "image_path": "a1bec29323051dd3f141ceafe8c8529f069e52c3b88888bb50396437e02e74d8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 306, + 387, + 316 + ], + "lines": [ + { + "bbox": [ + 327, + 306, + 387, + 316 + ], + "spans": [ + { + "bbox": [ + 327, + 306, + 387, + 316 + ], + "type": "text", + "content": "Source-textured" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 399, + 87, + 462, + 117 + ], + "blocks": [ + { + "bbox": [ + 399, + 87, + 462, + 117 + ], + "lines": [ + { + "bbox": [ + 399, + 87, + 462, + 117 + ], + "spans": [ + { + "bbox": [ + 399, + 87, + 462, + 117 + ], + "type": "image", + "image_path": "7bc5a9a3027125bf9343382e026f7b073a24587a5c2316e40321db4463efa9ef.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 425, + 125, + 441, + 185 + ], + "blocks": [ + { + "bbox": [ + 425, + 125, + 441, + 185 + ], + "lines": [ + { + "bbox": [ + 425, + 125, + 441, + 185 + ], + "spans": [ + { + "bbox": [ + 425, + 125, + 441, + 185 + ], + "type": "image", + "image_path": "7e75364c148897b62396bbbfe168a1a1f1d8809671d04fe86c66c937121c5cb0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 402, + 191, + 462, + 243 + ], + "blocks": [ + { + "bbox": [ + 402, + 191, + 462, + 243 + ], + "lines": [ + { + "bbox": [ + 402, + 191, + 462, + 243 + ], + "spans": [ + { + "bbox": [ + 402, + 191, + 462, + 243 + ], + "type": "image", + "image_path": "c3b13a4ad5e2c3e58d86e91dfb43c30b2f1c3d3f7d5300d77a90dd68ac7fb652.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 400, + 255, + 460, + 297 + ], + "blocks": [ + { + "bbox": [ + 400, + 255, + 460, + 297 + ], + "lines": [ + { + "bbox": [ + 400, + 255, + 460, + 297 + ], + "spans": [ + { + "bbox": [ + 400, + 255, + 460, + 297 + ], + "type": "image", + "image_path": "02f9d169cf28215295763e0e9ab51e5f16335e522622192b9e2ab4a5337797c5.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 404, + 306, + 465, + 317 + ], + "lines": [ + { + "bbox": [ + 404, + 306, + 465, + 317 + ], + "spans": [ + { + "bbox": [ + 404, + 306, + 465, + 317 + ], + "type": "text", + "content": "Target geometry" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 476, + 87, + 539, + 117 + ], + "blocks": [ + { + "bbox": [ + 476, + 87, + 539, + 117 + ], + "lines": [ + { + "bbox": [ + 476, + 87, + 539, + 117 + ], + "spans": [ + { + "bbox": [ + 476, + 87, + 539, + 117 + ], + "type": "image", + "image_path": "731d2c7d06f75c86e93e7c515096ab7fda78cace3d705842b135e61449eab528.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 501, + 125, + 518, + 185 + ], + "blocks": [ + { + "bbox": [ + 501, + 125, + 518, + 185 + ], + "lines": [ + { + "bbox": [ + 501, + 125, + 518, + 185 + ], + "spans": [ + { + "bbox": [ + 501, + 125, + 518, + 185 + ], + "type": "image", + "image_path": "dac90a887d3e9e92046f2a95ac3bce08d3cd19c4b260c6a63ff5a90548bc820e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 478, + 191, + 539, + 243 + ], + "blocks": [ + { + "bbox": [ + 478, + 191, + 539, + 243 + ], + "lines": [ + { + "bbox": [ + 478, + 191, + 539, + 243 + ], + "spans": [ + { + "bbox": [ + 478, + 191, + 539, + 243 + ], + "type": "image", + "image_path": "c1ae6ced5be668c918f136fdba68d1e2334d4e1a9658b652fffee60aaf7955b5.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 486, + 256, + 545, + 297 + ], + "blocks": [ + { + "bbox": [ + 486, + 256, + 545, + 297 + ], + "lines": [ + { + "bbox": [ + 486, + 256, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 486, + 256, + 545, + 297 + ], + "type": "image", + "image_path": "c63d101a1cf046cc225f3c1b446ec5a2e3cf14bc7bd031e4eeb8222c5e414e9d.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 484, + 306, + 541, + 316 + ], + "lines": [ + { + "bbox": [ + 484, + 306, + 541, + 316 + ], + "spans": [ + { + "bbox": [ + 484, + 306, + 541, + 316 + ], + "type": "text", + "content": "Target-textured" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 319, + 555, + 342 + ], + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 342 + ], + "type": "text", + "content": "Figure 4. Additional Qualitative Results. Showing the target geometry and the texture transfer." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 363, + 460, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 460, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 460, + 376 + ], + "type": "text", + "content": "4. Experiments And Results" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 383, + 427, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 427, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 427, + 396 + ], + "type": "text", + "content": "4.1. Experiment Details" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 401, + 555, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 555, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 555, + 546 + ], + "type": "text", + "content": "Dataset: Our experimental dataset consists of 52 3D meshes across 9 categories, curated from Objaverse [15, 16], a large-scale 3D object database. We first generated descriptive captions for each 3D object to identify semantically related objects using BLIP [27]. We then used keyword filtering on these captions to group objects into similar categories. The final selection includes only objects that both share semantic similarities and maintain consistent orientations within their respective categories. Overall, we have 256 different pairs of source and target objects. The full list of source and target objects is provided in the supplemental." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 563, + 556, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 556, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 556, + 672 + ], + "type": "text", + "content": "Baseline Methods: We compare our method against three state-of-the-art optimization-free approaches for visually-guided mesh texturing. (1) TEXTure [40], a personalization-based approach, enabling texture transfer from a source mesh to a target mesh. (2) EASI-TEX [37] and (3) MVEdit [9] use IP-Adapter to incorporate image guidance into their texture generation process, allowing direct control over the generated textures through reference images." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "Metrics: We evaluate our method using human evaluation and two automated metrics: CLIP similarity score [39]" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21407" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 102, + 68, + 506, + 398 + ], + "blocks": [ + { + "bbox": [ + 102, + 68, + 506, + 398 + ], + "lines": [ + { + "bbox": [ + 102, + 68, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 102, + 68, + 506, + 398 + ], + "type": "image", + "image_path": "8250877c7d10b6f0cb958cd8a36e759fcf0d57a27983e94c0ca0f5134112356d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 409, + 555, + 433 + ], + "lines": [ + { + "bbox": [ + 55, + 409, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 555, + 433 + ], + "type": "text", + "content": "Figure 5. Qualitative Comparisons with baselines. A comparison between TRiTEX (right-most column) and state-of-the-art approaches for texture transfer. As observed, TRiTEX produces much more plausible texture transfer results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 505, + 293, + 573 + ], + "blocks": [ + { + "bbox": [ + 55, + 451, + 295, + 497 + ], + "lines": [ + { + "bbox": [ + 55, + 451, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 295, + 497 + ], + "type": "text", + "content": "Table 1. Quantitative Evaluation. Quantitative comparison of our method with baseline methods. We evaluate two appearance metrics and running time, showing that our method best preserves the source texture while achieving competitive speed." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 505, + 293, + 573 + ], + "lines": [ + { + "bbox": [ + 56, + 505, + 293, + 573 + ], + "spans": [ + { + "bbox": [ + 56, + 505, + 293, + 573 + ], + "type": "table", + "html": "
MethodSIFID ↓CLIP sim. ↑Inference Runtime
TEXTure0.340.845 min.
MVEdit0.380.841 min.
EASI-TEX0.290.8515 min.
TriTex (ours)0.220.871 min.
", + "image_path": "6943373119d693f1e0313c710b9a828476d4597d1f6c2c5a0dbb5ad016298d86.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 593, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 295, + 700 + ], + "type": "text", + "content": "and Single Image Frechet Inception Distance (SIFID). The CLIP score, adopted following TextureDreamer [57], measures the semantic and visual similarity between source and generated textures, indicating how well our method preserves the desired appearance. SIFID was originally proposed by [43] to measure internal patch statistics within single images. Following Sin3DM [53], we adopt this metric to evaluate texture similarity between source and target shapes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "For both metrics, we render each source and target mesh" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 453, + 555, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 555, + 513 + ], + "type": "text", + "content": "from the same 10 fixed viewpoints and compare the corresponding views against each other. To ensure balanced evaluation across our dataset, we first compute the average score within each object category, then average these category scores to obtain the final metric values." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 527, + 556, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 556, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 556, + 624 + ], + "type": "text", + "content": "Human Evaluation To complement these quantitative metrics, we also conduct a user study on Amazon Mechanical Turk (AMT) with 15 participants to evaluate our results. Each participant viewed a source object and two versions of the textured target mesh from different methods, selecting which result better preserves the texture and the appearance of the reference object. Each comparison received ratings from two independent evaluators." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 629, + 372, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 372, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 372, + 641 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 647, + 555, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 555, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 555, + 684 + ], + "type": "text", + "content": "We provide qualitative examples, qualitative and quantitative comparisons with baselines, and an ablation study with qualitative and quantitative analysis." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Qualitative Examples Figure 3 provides four examples illustrating how TRITEX transfers texture from the source" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21408" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "shape (left column) to three other shapes. Notice how colors match the semantics of the location, despite significant shape variations. for example the eyes of the fish at the bottom row. Figure 4 shows additional results along with the corresponding target geometries." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 139, + 296, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 139, + 296, + 211 + ], + "spans": [ + { + "bbox": [ + 55, + 139, + 296, + 211 + ], + "type": "text", + "content": "Quantitative Comparisons: Table 1 compares our TRI-TEX approach with all baseline methods. We find that TRI-TEX outperforms baselines in both evaluation metrics, successfully capturing the appearance of the original objects. We also report the inference time for texturing a new mesh for each method." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 218, + 296, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 218, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 296, + 291 + ], + "type": "text", + "content": "Qualitative Comparisons: We show qualitative comparisons between our approach and the baselines in Fig. 5. As demonstrated, MVEdit tends to deviate significantly from the input image, only taking vague inspiration from the reference. TEXTure and EASI-TEX produce artifacts due to their iterative painting techniques." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 57, + 301, + 294, + 392 + ], + "blocks": [ + { + "bbox": [ + 57, + 301, + 294, + 392 + ], + "lines": [ + { + "bbox": [ + 57, + 301, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 57, + 301, + 294, + 392 + ], + "type": "image", + "image_path": "31c498d805cea7c48e04923f4e524059c27ed74315707c957884451541e29d44.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 400, + 296, + 445 + ], + "lines": [ + { + "bbox": [ + 55, + 400, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 296, + 445 + ], + "type": "text", + "content": "Figure 6. Results of human evaluation study. TRITEX was compared with three baseline approaches in a 2-alternative forced choice setting over " + }, + { + "bbox": [ + 55, + 400, + 296, + 445 + ], + "type": "inline_equation", + "content": "\\sim 750" + }, + { + "bbox": [ + 55, + 400, + 296, + 445 + ], + "type": "text", + "content": " questions. Raters strongly favored TRI-TEX for better transfer of appearance." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 472, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 581 + ], + "type": "text", + "content": "User Study: We conducted a user study to evaluate the quality of TRiTEX compared with baselines. For each baseline, we presented the source object and two target shapes, one textured with TRiTEX and one with the baseline, in a 2-alternative forced choice (2AFC) setting. We used Amazon Mechanical Turk (AMT) and paid raters above the minimum wage. Figure 6 shows the results of these comparisons. TRiTEX was strongly preferred by raters over all three competing approaches." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 588, + 156, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 156, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 156, + 600 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "We conducted an ablation study to evaluate the contribution of each component in our method. The quantitative and quantitative results presented in Table 2 and Fig. 7 respectively, demonstrate that each component is essential for optimal performance. Our analysis reveals that excluding " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mse}}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " leads to color predictions that deviate significantly from the original texture. Similarly, removing " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{app}}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " results in blurry outputs that fail to capture complex texture patterns. We further found that omitting data augmentation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": "causes the network to overfit to specific triplane projections, thereby limiting the method's generalization capabilities." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 337, + 116, + 391, + 163 + ], + "blocks": [ + { + "bbox": [ + 337, + 116, + 391, + 163 + ], + "lines": [ + { + "bbox": [ + 337, + 116, + 391, + 163 + ], + "spans": [ + { + "bbox": [ + 337, + 116, + 391, + 163 + ], + "type": "image", + "image_path": "1020060e01d90da40c21efb441622f18e0b3e7047b9bc5a7b90b93aa0af3f8a1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 171, + 391, + 181 + ], + "lines": [ + { + "bbox": [ + 343, + 171, + 391, + 181 + ], + "spans": [ + { + "bbox": [ + 343, + 171, + 391, + 181 + ], + "type": "text", + "content": "source shape" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 408, + 116, + 459, + 164 + ], + "blocks": [ + { + "bbox": [ + 408, + 116, + 459, + 164 + ], + "lines": [ + { + "bbox": [ + 408, + 116, + 459, + 164 + ], + "spans": [ + { + "bbox": [ + 408, + 116, + 459, + 164 + ], + "type": "image", + "image_path": "1af02c454f246e5ab4655d6287ab298d4559e4ce1048236ef3a7969e5bdbea36.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 171, + 457, + 180 + ], + "lines": [ + { + "bbox": [ + 410, + 171, + 457, + 180 + ], + "spans": [ + { + "bbox": [ + 410, + 171, + 457, + 180 + ], + "type": "text", + "content": "w/o network" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 474, + 116, + 525, + 164 + ], + "blocks": [ + { + "bbox": [ + 474, + 116, + 525, + 164 + ], + "lines": [ + { + "bbox": [ + 474, + 116, + 525, + 164 + ], + "spans": [ + { + "bbox": [ + 474, + 116, + 525, + 164 + ], + "type": "image", + "image_path": "dfb810799e959cbf95d6b37aeb2d70fcad42071fe9d4136d368a0a42d384ac7a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 484, + 171, + 517, + 182 + ], + "lines": [ + { + "bbox": [ + 484, + 171, + 517, + 182 + ], + "spans": [ + { + "bbox": [ + 484, + 171, + 517, + 182 + ], + "type": "text", + "content": "w/o " + }, + { + "bbox": [ + 484, + 171, + 517, + 182 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{app}}" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 341, + 188, + 392, + 236 + ], + "blocks": [ + { + "bbox": [ + 341, + 188, + 392, + 236 + ], + "lines": [ + { + "bbox": [ + 341, + 188, + 392, + 236 + ], + "spans": [ + { + "bbox": [ + 341, + 188, + 392, + 236 + ], + "type": "image", + "image_path": "9a6c8780c8a050a3f08152540307b729bcf5ca6b4266e415f7497b54f2e37a67.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 243, + 384, + 252 + ], + "lines": [ + { + "bbox": [ + 348, + 243, + 384, + 252 + ], + "spans": [ + { + "bbox": [ + 348, + 243, + 384, + 252 + ], + "type": "text", + "content": "w/o " + }, + { + "bbox": [ + 348, + 243, + 384, + 252 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MSE}}" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 408, + 188, + 459, + 236 + ], + "blocks": [ + { + "bbox": [ + 408, + 188, + 459, + 236 + ], + "lines": [ + { + "bbox": [ + 408, + 188, + 459, + 236 + ], + "spans": [ + { + "bbox": [ + 408, + 188, + 459, + 236 + ], + "type": "image", + "image_path": "cb9de220c05d58891571601679b5b2328c320e277c5c719e44624ef4549beb57.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 243, + 468, + 252 + ], + "lines": [ + { + "bbox": [ + 399, + 243, + 468, + 252 + ], + "spans": [ + { + "bbox": [ + 399, + 243, + 468, + 252 + ], + "type": "text", + "content": "w/o augmentations" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 474, + 188, + 525, + 236 + ], + "blocks": [ + { + "bbox": [ + 474, + 188, + 525, + 236 + ], + "lines": [ + { + "bbox": [ + 474, + 188, + 525, + 236 + ], + "spans": [ + { + "bbox": [ + 474, + 188, + 525, + 236 + ], + "type": "image", + "image_path": "5904ea246bf966148954ad7a074e0c70dec64bbcd01578e248f6df65defa4010.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 492, + 244, + 509, + 251 + ], + "lines": [ + { + "bbox": [ + 492, + 244, + 509, + 251 + ], + "spans": [ + { + "bbox": [ + 492, + 244, + 509, + 251 + ], + "type": "text", + "content": "ours" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 262, + 555, + 305 + ], + "lines": [ + { + "bbox": [ + 313, + 262, + 555, + 305 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 555, + 305 + ], + "type": "text", + "content": "Figure 7. Qualitative Ablation Study. We demonstrate the effect of removing each component of our pipeline on a single example. As shown, each component is essential for successful texture transfer." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 323, + 555, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 455 + ], + "type": "text", + "content": "We also ablate the contribution of the neural network in our pipeline. In the \"w/o network\" baseline, we utilize pre-extracted DIFF3f semantic features and identify the nearest neighbor for each target mesh feature within the source mesh features. Using these nearest-neighbor matches, we transfer colors from the source mesh to the target vertices and then render the target mesh with its new colors. The impact of this simplification is presented in Table 2 and Fig.7, demonstrating the degraded quality of this approach in transferring the texture details compared to our full pipeline. Two additional examples are shown in Fig.8." + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 319, + 477, + 381, + 520 + ], + "blocks": [ + { + "bbox": [ + 319, + 477, + 381, + 520 + ], + "lines": [ + { + "bbox": [ + 319, + 477, + 381, + 520 + ], + "spans": [ + { + "bbox": [ + 319, + 477, + 381, + 520 + ], + "type": "image", + "image_path": "4702cb30bb4e27becee8a4cdb3aaaa242395324d3d24d600a3cf1161eb2817ef.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 397, + 472, + 455, + 519 + ], + "blocks": [ + { + "bbox": [ + 397, + 472, + 455, + 519 + ], + "lines": [ + { + "bbox": [ + 397, + 472, + 455, + 519 + ], + "spans": [ + { + "bbox": [ + 397, + 472, + 455, + 519 + ], + "type": "image", + "image_path": "2037a3d849c10182be11fcb34db3737ad581e2d2cb2f733e82412a9e05242f48.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 475, + 472, + 533, + 520 + ], + "blocks": [ + { + "bbox": [ + 475, + 472, + 533, + 520 + ], + "lines": [ + { + "bbox": [ + 475, + 472, + 533, + 520 + ], + "spans": [ + { + "bbox": [ + 475, + 472, + 533, + 520 + ], + "type": "image", + "image_path": "d75a3b477558a3e543c7022381db7f719540f7e8dda700729a5f9eff076ba1f6.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 324, + 543, + 381, + 605 + ], + "blocks": [ + { + "bbox": [ + 324, + 543, + 381, + 605 + ], + "lines": [ + { + "bbox": [ + 324, + 543, + 381, + 605 + ], + "spans": [ + { + "bbox": [ + 324, + 543, + 381, + 605 + ], + "type": "image", + "image_path": "7fe0935e5b8f369af087b3a3cb5f54ff1c0da46ecf73d1d3cbead6e69c1aa498.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 610, + 380, + 620 + ], + "lines": [ + { + "bbox": [ + 332, + 610, + 380, + 620 + ], + "spans": [ + { + "bbox": [ + 332, + 610, + 380, + 620 + ], + "type": "text", + "content": "source shape" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 631, + 555, + 696 + ], + "lines": [ + { + "bbox": [ + 313, + 631, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 631, + 555, + 696 + ], + "type": "text", + "content": "Figure 8. Network Ablation. A comparison between our full method (right) and the nearest neighboring baseline (middle), which replaces the neural network in our pipeline and transfers color based on the closest matching neighbors. As seen, this baseline fails to transfer complex texture details, unlike our full method." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 398, + 544, + 453, + 598 + ], + "blocks": [ + { + "bbox": [ + 398, + 544, + 453, + 598 + ], + "lines": [ + { + "bbox": [ + 398, + 544, + 453, + 598 + ], + "spans": [ + { + "bbox": [ + 398, + 544, + 453, + 598 + ], + "type": "image", + "image_path": "ad86fe1defffaac39c926a2511599734476a18bc1d845a8c9efbe10b2286e1f8.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 610, + 456, + 620 + ], + "lines": [ + { + "bbox": [ + 410, + 610, + 456, + 620 + ], + "spans": [ + { + "bbox": [ + 410, + 610, + 456, + 620 + ], + "type": "text", + "content": "w/o network" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 475, + 544, + 531, + 598 + ], + "blocks": [ + { + "bbox": [ + 475, + 544, + 531, + 598 + ], + "lines": [ + { + "bbox": [ + 475, + 544, + 531, + 598 + ], + "spans": [ + { + "bbox": [ + 475, + 544, + 531, + 598 + ], + "type": "image", + "image_path": "5007931271fc63995ecd1e73540dfc38cf35b217f66cf2b770e369db9dbc646a.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 502, + 611, + 519, + 619 + ], + "lines": [ + { + "bbox": [ + 502, + 611, + 519, + 619 + ], + "spans": [ + { + "bbox": [ + 502, + 611, + 519, + 619 + ], + "type": "text", + "content": "ours" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21409" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 123, + 268, + 205 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "type": "text", + "content": "Table 2. Quantitative Ablation Study. We demonstrate the effect of removing component of TRiTEX with quantitative metrics averaged across our evaluation set. The results show that each component is essential for achieving successful texture transfer." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 123, + 268, + 205 + ], + "lines": [ + { + "bbox": [ + 83, + 123, + 268, + 205 + ], + "spans": [ + { + "bbox": [ + 83, + 123, + 268, + 205 + ], + "type": "table", + "html": "
MethodSIFID ↓CLIP sim. ↑
w/o network0.230.86
w/o L MSE0.230.86
w/o L app0.280.85
w/o augmentations0.210.85
TriTex (ours)0.220.87
", + "image_path": "5553b555e1579030e77fc38b505b27c963d2665c90e1530b8434c8aa93c51443.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 223, + 126, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 223, + 126, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 126, + 236 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 243, + 296, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 296, + 388 + ], + "type": "text", + "content": "Cross-Category Object Transfer Although our methods is desinged to transfer appearance between objects from the same category, it is interesting to examine the results when transferring appearance between objects with varying degrees of semantic similarity. As shown in Figure 9, when transferring between objects with shared semantic structure (e.g. parts with similar functionality or spatial relationships), TRiTEX preserves meaningful texture patterns aligned with these semantic features. However, when transferring between objects with little semantic overlap, the resulting texture patterns become more stochastic, as the semantic mapping becomes less well-defined." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 62, + 418, + 111, + 471 + ], + "blocks": [ + { + "bbox": [ + 62, + 418, + 111, + 471 + ], + "lines": [ + { + "bbox": [ + 62, + 418, + 111, + 471 + ], + "spans": [ + { + "bbox": [ + 62, + 418, + 111, + 471 + ], + "type": "image", + "image_path": "9fd75269493e8d5de6cd21b38cdb54af3eb101932ed9c3585551032583e2ccab.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 484, + 123, + 495 + ], + "lines": [ + { + "bbox": [ + 71, + 484, + 123, + 495 + ], + "spans": [ + { + "bbox": [ + 71, + 484, + 123, + 495 + ], + "type": "text", + "content": "Source Shape" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 143, + 410, + 194, + 477 + ], + "blocks": [ + { + "bbox": [ + 143, + 410, + 194, + 477 + ], + "lines": [ + { + "bbox": [ + 143, + 410, + 194, + 477 + ], + "spans": [ + { + "bbox": [ + 143, + 410, + 194, + 477 + ], + "type": "image", + "image_path": "6553c32abc2bf514ed5be13f4a9760905d2343239d5ad891b19cf5ac9a696887.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 484, + 191, + 495 + ], + "lines": [ + { + "bbox": [ + 157, + 484, + 191, + 495 + ], + "spans": [ + { + "bbox": [ + 157, + 484, + 191, + 495 + ], + "type": "text", + "content": "Output 1" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 229, + 412, + 266, + 479 + ], + "blocks": [ + { + "bbox": [ + 229, + 412, + 266, + 479 + ], + "lines": [ + { + "bbox": [ + 229, + 412, + 266, + 479 + ], + "spans": [ + { + "bbox": [ + 229, + 412, + 266, + 479 + ], + "type": "image", + "image_path": "36ca59a3ad5432db8c89afec9ca3b3344a94d95fa94e9c86dcbe701fd1f8a50d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 484, + 269, + 495 + ], + "lines": [ + { + "bbox": [ + 235, + 484, + 269, + 495 + ], + "spans": [ + { + "bbox": [ + 235, + 484, + 269, + 495 + ], + "type": "text", + "content": "Output 2" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 505, + 297, + 550 + ], + "lines": [ + { + "bbox": [ + 55, + 505, + 297, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 297, + 550 + ], + "type": "text", + "content": "Figure 9. Cross Category Texture Transfer. Texture transfer between objects with partial semantic correspondence (fish and deer) and versus objects with minimal semantic similarity (fish and vase with flowers)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": "Limitations Despite the effectiveness of our approach, it has several inherent limitations. First, our method does not possess generative capabilities, which means it cannot synthesize novel details that are not present in the source texture. This limitation becomes particularly apparent when there is significant geometric variation between the source and target meshes, as our method can only map existing texture patterns rather than create new ones to accommodate the structural differences (Figure 10, top). Second, without leveraging priors from text-to-image models, our approach lacks the ability to compensate for cases where the semantic features fail to capture sufficient detail or establish accurate" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 556, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 168 + ], + "type": "text", + "content": "correspondences. This can result in less detailed or less coherent texture transfers in regions where the feature matching is ambiguous or imprecise (Figure 10, bottom). Additionally, our method expects the target shape to be at the same orientation as the source shape. When enabling arbitrary rotations during training, the model becomes invariant to the object's orientation, but the ability to reproduce fine details such as eyes or wheels is reduced." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 182, + 358, + 229 + ], + "blocks": [ + { + "bbox": [ + 317, + 182, + 358, + 229 + ], + "lines": [ + { + "bbox": [ + 317, + 182, + 358, + 229 + ], + "spans": [ + { + "bbox": [ + 317, + 182, + 358, + 229 + ], + "type": "image", + "image_path": "2622a33392d9497f9c7894a1dee4ec6ef802ace47403fdb9ee78b0d703111a95.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 294, + 556, + 361 + ], + "lines": [ + { + "bbox": [ + 313, + 294, + 556, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 556, + 361 + ], + "type": "text", + "content": "Figure 10. Limitations. (Top) Since our network lacks a generative prior and the source dog has no visible tongue, it cannot synthesize an appropriate color for the target dog's tongue. (Bottom) The network incorrectly transfers the source bed's sheet texture to the target bed frame, as PCA analysis shows their feature representations are similar." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 381, + 183, + 418, + 229 + ], + "blocks": [ + { + "bbox": [ + 381, + 183, + 418, + 229 + ], + "lines": [ + { + "bbox": [ + 381, + 183, + 418, + 229 + ], + "spans": [ + { + "bbox": [ + 381, + 183, + 418, + 229 + ], + "type": "image", + "image_path": "a72f626d4dccf759b18e42f3476dd6aa7853dc5cd716202871a0fd7a6ae59d74.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 440, + 184, + 473, + 229 + ], + "blocks": [ + { + "bbox": [ + 440, + 184, + 473, + 229 + ], + "lines": [ + { + "bbox": [ + 440, + 184, + 473, + 229 + ], + "spans": [ + { + "bbox": [ + 440, + 184, + 473, + 229 + ], + "type": "image", + "image_path": "778bfdf5e9598588f9a8b27a0bfc68b274046d050429e125b4931388fb671036.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 498, + 184, + 533, + 229 + ], + "blocks": [ + { + "bbox": [ + 498, + 184, + 533, + 229 + ], + "lines": [ + { + "bbox": [ + 498, + 184, + 533, + 229 + ], + "spans": [ + { + "bbox": [ + 498, + 184, + 533, + 229 + ], + "type": "image", + "image_path": "0647318a04801b83506edae6c1c5661f6dde1298603ddd9f89020cb47bd17fa6.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 316, + 239, + 368, + 270 + ], + "blocks": [ + { + "bbox": [ + 316, + 239, + 368, + 270 + ], + "lines": [ + { + "bbox": [ + 316, + 239, + 368, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 368, + 270 + ], + "type": "image", + "image_path": "3be40d431b08c41bb52feb2ad84d6a37ba7de026a9c23d602d3b9a3c02313493.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 273, + 437, + 283 + ], + "lines": [ + { + "bbox": [ + 318, + 273, + 437, + 283 + ], + "spans": [ + { + "bbox": [ + 318, + 273, + 437, + 283 + ], + "type": "text", + "content": "source shape source shape PCA" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 376, + 239, + 429, + 270 + ], + "blocks": [ + { + "bbox": [ + 376, + 239, + 429, + 270 + ], + "lines": [ + { + "bbox": [ + 376, + 239, + 429, + 270 + ], + "spans": [ + { + "bbox": [ + 376, + 239, + 429, + 270 + ], + "type": "image", + "image_path": "2a45642008f1a08aa894b4180bb3120152be99265466399b5656364826cd662d.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 273, + 487, + 283 + ], + "lines": [ + { + "bbox": [ + 441, + 273, + 487, + 283 + ], + "spans": [ + { + "bbox": [ + 441, + 273, + 487, + 283 + ], + "type": "text", + "content": "target shape" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 438, + 243, + 490, + 266 + ], + "blocks": [ + { + "bbox": [ + 438, + 243, + 490, + 266 + ], + "lines": [ + { + "bbox": [ + 438, + 243, + 490, + 266 + ], + "spans": [ + { + "bbox": [ + 438, + 243, + 490, + 266 + ], + "type": "image", + "image_path": "ab964b66388e3faa19f2210bfb74e2a3c5a43cbf1fdb1696d3fd845ca67a4684.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 492, + 273, + 556, + 283 + ], + "lines": [ + { + "bbox": [ + 492, + 273, + 556, + 283 + ], + "spans": [ + { + "bbox": [ + 492, + 273, + 556, + 283 + ], + "type": "text", + "content": "target shape PCA" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 497, + 243, + 549, + 266 + ], + "blocks": [ + { + "bbox": [ + 497, + 243, + 549, + 266 + ], + "lines": [ + { + "bbox": [ + 497, + 243, + 549, + 266 + ], + "spans": [ + { + "bbox": [ + 497, + 243, + 549, + 266 + ], + "type": "image", + "image_path": "1b93789865540a605dc0c15f4a55f8bd5d9a4a1c69239a1fc86a1eeb4a7f257e.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 370, + 392, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 392, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 392, + 384 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 390, + 555, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 555, + 485 + ], + "type": "text", + "content": "We have presented a method that allows transferring a texture learned from a single exemplar to similar shapes. The learned texture is challenging in the sense that it is a nonstationary texture, where its patterns can be strongly correlated with the semantics of the shapes. Mapping such semantic textures on a target shape, necessarily requires, at least implicitly, extracting its semantics. For that, our model uses pre-trained 3D lifted semantic features." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 486, + 556, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 629 + ], + "type": "text", + "content": "Our key challenge was to process the semantic features in a way that allows generalization from a single instance, in order to perform the texture learning with minimal supervision. To achieve this, we re-rendered them into a triplane representation, which preserves feature proximity and connectivity, allowing for efficient processing and mapping to the corresponding colors. We introduced an effective training method using a single mesh with appearance losses while applying augmentations. The limitation of our technique is that it currently does not possess generative abilities and cannot create details that do not appear in the source texture." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "In future work, we aim to extend this feed-forward approach for fast texturing by training the model to predict the mesh texture from a reference image, using cross-attention layers between the generated triplane and the reference image. Additionally, we plan to leverage pre-trained generative models to enhance the mapped textures, increasing both resolution and quality with their learned priors." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21410" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 116 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 116 + ], + "type": "text", + "content": "We thank Rinon Gal, Yoad Tewel and Or Patashnik for their constructive suggestions and insightful comments." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 157, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 157, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 157, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 157, + 296, + 201 + ], + "type": "text", + "content": "[1] Kfir Aberman, Jing Liao, Mingyi Shi, Dani Lischinski, Baoquan Chen, and Daniel Cohen-Or. Neural best-buddies: Sparse cross-domain correspondence. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 202, + 296, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 296, + 236 + ], + "type": "text", + "content": "[2] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. arXiv e-prints, pages arXiv-2112, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 237, + 296, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 237, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 62, + 237, + 296, + 281 + ], + "type": "text", + "content": "[3] Raphael Bensadoun, Yanir Kleiman, Idan Azuri, Omri Harosh, Andrea Vedaldi, Natalia Neverova, and Oran Gafni. Meta 3d texturegen: Fast and consistent texture generation for 3d objects, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 282, + 294, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 282, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 62, + 282, + 294, + 315 + ], + "type": "text", + "content": "[4] Alexey Bokhovkin, Shubham Tulsiani, and Angela Dai. Mesh2tex: Generating mesh textures from image queries, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 317, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 317, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 317, + 294, + 350 + ], + "type": "text", + "content": "[5] Tianshi Cao, Karsten Kreis, Sanja Fidler, Nicholas Sharp, and Kangxue Yin. Texfusion: Synthesizing 3d textures with text-guided image diffusion models, 2023. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 351, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 351, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 351, + 296, + 407 + ], + "type": "text", + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 408, + 296, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 408, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 62, + 408, + 296, + 474 + ], + "type": "text", + "content": "[7] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16123-16133, 2022. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 475, + 294, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 475, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 62, + 475, + 294, + 508 + ], + "type": "text", + "content": "[8] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models, 2023. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 510, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 510, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 62, + 510, + 294, + 553 + ], + "type": "text", + "content": "[9] Hansheng Chen, Ruoxi Shi, Yulin Liu, Bokui Shen, Jiayuan Gu, Gordon Wetzstein, Hao Su, and Leonidas Guibas. Generic 3d diffusion adapter using controlled multi-view editing, 2024. 1, 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 555, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 555, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 57, + 555, + 294, + 588 + ], + "type": "text", + "content": "[10] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 590, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 294, + 623 + ], + "type": "text", + "content": "[11] Xiaobai Chen, Thomas Funkhouser, Dan B Goldman, and Eli Shechtman. Non-parametric texture transfer using MeshMatch. Adobe Technical Report 2012-2, 2012. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 624, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 657 + ], + "type": "text", + "content": "[12] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "type": "text", + "content": "[13] Zhiqin Chen, Kangxue Yin, and Sanja Fidler. Auv-net: Learning aligned uv maps for texture transfer and synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1465-1474, 2022. 1, 2, 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "text", + "content": "[14] Jeremy S De Bonet. Multiresolution sampling procedure for analysis and synthesis of texture images. In Proceedings of the 24th annual conference on Computer graphics and interactive techniques, pages 361-368, 1997. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 162 + ], + "type": "text", + "content": "[15] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects, 2022. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 163, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 228 + ], + "type": "text", + "content": "[16] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-xl: A universe of " + }, + { + "bbox": [ + 317, + 163, + 553, + 228 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{d}" + }, + { + "bbox": [ + 317, + 163, + 553, + 228 + ], + "type": "text", + "content": " objects, 2023. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 230, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 230, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 553, + 274 + ], + "type": "text", + "content": "[17] Kangle Deng, Timothy Omernick, Alexander Weiss, Deva Ramanan, Jun-Yan Zhu, Tinghui Zhou, and Maneesh Agrawala. *Flashtex: Fast reconfigurable mesh texturing with lightcontrolnet*, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 275, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 275, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 275, + 553, + 308 + ], + "type": "text", + "content": "[18] Niladri Shekhar Dutt, Sanjeev Muralikrishnan, and Niloy J. Mitra. Diffusion 3d features (diff3f): Decorating textured shapes with distilled semantic features, 2024. 2, 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 309, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 309, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 317, + 309, + 553, + 352 + ], + "type": "text", + "content": "[19] Alexei A Efros and Thomas K Leung. Texture synthesis by non-parametric sampling. In Proceedings of the seventh IEEE international conference on computer vision, pages 1033-1038. IEEE, 1999. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "type": "text", + "content": "[20] Michael Fischer, Zhengqin Li, Thu Nguyen-Phuoc, Aljaz Bozic, Zhao Dong, Carl Marshall, and Tobias Ritschel. Nerf analogies: Example-based visual attribute transfer for nerfs, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "text", + "content": "[21] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 455, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 553, + 487 + ], + "type": "text", + "content": "[22] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oguz. 3dgen: Triplane latent diffusion for textured mesh generation, 2023. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 488, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 488, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 488, + 553, + 533 + ], + "type": "text", + "content": "[23] David J Heeger and James R Bergen. Pyramid-based texture analysis/synthesis. In Proceedings of the 22nd annual conference on Computer graphics and interactive techniques, pages 229–238, 1995. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 534, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 534, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 534, + 553, + 567 + ], + "type": "text", + "content": "[24] Amir Hertz, Rana Hanocka, Raja Giryes, and Daniel Cohen-Or. Deep geometric texture synthesis. ACM Transactions on Graphics, 39(4), 2020. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 568, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 553, + 600 + ], + "type": "text", + "content": "[25] Jaihoon Kim, Juil Koo, Kyeongmin Yeo, and Minhyuk Sung. Synctweedies: A general generative framework based on synchronized diffusions, 2024. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 601, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 553, + 645 + ], + "type": "text", + "content": "[26] Vladimir Kulikov, Shahar Yadin, Matan Kleiner, and Tomer Michaeli. Sinddm: A single image denoising diffusion model. In International Conference on Machine Learning, pages 17920-17930. PMLR, 2023. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "text", + "content": "[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation, 2022. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "text", + "content": "[28] Yuxin Liu, Minshan Xie, Hanyuan Liu, and Tien-Tsin Wong. Text-guided texturing by synchronized multi-view diffusion, 2023. 1, 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21411" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "text", + "content": "[29] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 138 + ], + "type": "text", + "content": "[30] Jianye Lu, Athinodoros S. Georgiades, Andreas Glaser, Hongzhi Wu, Li-Yi Wei, Baining Guo, Julie Dorsey, and Holly Rushmeier. Context-aware textures. 26, 2007. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 172 + ], + "type": "text", + "content": "[31] Tom Mertens, Jan Kautz, Jiawen Chen, Philippe Bekaert, and Frédo Durand. Texture transfer using geometry correlation. Goslar, DEU, 2006. Eurographics Association. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 173, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 205 + ], + "type": "text", + "content": "[32] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures, 2022. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 205, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 237 + ], + "type": "text", + "content": "[33] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 239, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 271 + ], + "type": "text", + "content": "[34] Thomas W. Mitchel, Carlos Esteves, and Ameesh Makadia. Single mesh diffusion models with field latents for texture generation, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 272, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 304 + ], + "type": "text", + "content": "[35] Yaniv Nikankin, Niv Haim, and Michal Irani. Sinfusion: Training diffusion models on a single image or video. arXiv preprint arXiv:2211.11743, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 305, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 294, + 358 + ], + "type": "text", + "content": "[36] Dolev Ofri-Amar, Michal Geyer, Yoni Kasten, and Tali Dekel. Neural congealing: Aligning images to a joint semantic atlas. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19403-19412, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 360, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 392 + ], + "type": "text", + "content": "[37] Sai Raj Kishore Perla, Yizhi Wang, Ali Mahdavi-Amiri, and Hao Zhang. Easi-tex: Edge-aware mesh texturing from single image, 2024. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "type": "text", + "content": "[38] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "type": "text", + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 481, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 294, + 514 + ], + "type": "text", + "content": "[40] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes, 2023. 1, 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 515, + 294, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 547 + ], + "type": "text", + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 548, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 613 + ], + "type": "text", + "content": "[42] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 614, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 646 + ], + "type": "text", + "content": "[43] Tamar Rott Shaham, Tali Dekel, and Tomer Michaeli. Singan: Learning a generative model from a single natural image, 2019. 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "type": "text", + "content": "[44] J. Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion, 2022. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[45] Yawar Siddiqui, Justus Thies, Fangchang Ma, Qi Shan, Matthias Nießner, and Angela Dai. Texturify: Generating textures on 3d shape surfaces, 2022. 1, 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[46] Edgar Simo-Serra, Eduard Trulls, Luis Ferraz, Iasonas Kokkinos, Pascal Fua, and Francesc Moreno-Noguer. Discriminative learning of deep convolutional feature point descriptors. In Proceedings of the IEEE international conference on computer vision, pages 118-126, 2015. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "text", + "content": "[47] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. Advances in Neural Information Processing Systems, 36:1363-1389, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 174, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 205 + ], + "type": "text", + "content": "[48] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer, 2022. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 207, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 553, + 261 + ], + "type": "text", + "content": "[49] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10748-10757, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 262, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 308 + ], + "type": "text", + "content": "[50] Weilun Wang, Jianmin Bao, Wengang Zhou, Dongdong Chen, Dong Chen, Lu Yuan, and Houqiang Li. Sindiffusion: Learning a diffusion model from a single natural image. arXiv preprint arXiv:2211.12445, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "type": "text", + "content": "[51] Li-Yi Wei and Marc Levoy. Texture synthesis over arbitrary manifold surfaces. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 355-360, 2001. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "type": "text", + "content": "[52] Rundi Wu and Changxi Zheng. Learning to generate 3d shapes from a single example. ACM Transactions on Graphics, 41(6):1-19, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "type": "text", + "content": "[53] Rundi Wu, Ruoshi Liu, Carl Vondrick, and Changxi Zheng. Sin3dm: Learning a diffusion model from a single 3d textured shape, 2024. 3, 4, 5, 6, 12" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 421, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 476 + ], + "type": "text", + "content": "[54] Wenqi Xian, Patsorn Sangkloy, Varun Agrawal, Amit Raj, Jingwan Lu, Chen Fang, Fisher Yu, and James Hays. Texturegan: Controlling deep image synthesis with texture patches. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8456-8465, 2018. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 477, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 477, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 477, + 553, + 520 + ], + "type": "text", + "content": "[55] Jiale Xu, Xintao Wang, Weihao Cheng, Yan-Pei Cao, Ying Shan, Xiaohu Qie, and Shenghua Gao. Dream3d: Zero-shot text-to-3d synthesis using 3d shape prior and text-to-image diffusion models, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 522, + 553, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 565 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 565 + ], + "type": "text", + "content": "[56] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 567, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 622 + ], + "type": "text", + "content": "[57] Yu-Ying Yeh, Jia-Bin Huang, Changil Kim, Lei Xiao, Thu Nguyen-Phuoc, Numair Khan, Cheng Zhang, Manmohan Chandraker, Carl S Marshall, Zhao Dong, and Zhengqin Li. Texturedreamer: Image-guided texture synthesis through geometry-aware diffusion, 2024. 1, 2, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 624, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 656 + ], + "type": "text", + "content": "[58] Kangxue Yin, Jun Gao, Maria Shugrina, Sameh Khamis, and Sanja Fidler. 3dstylenet: Creating 3d shapes with geometric and texture style variations, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 657, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 657, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 657, + 553, + 690 + ], + "type": "text", + "content": "[59] Kim Youwang, Tae-Hyun Oh, and Gerard Pons-Moll. Paint-it: Text-to-texture synthesis via deep convolutional texture map optimization and physically-based rendering, 2024. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[60] Xianfang Zeng, Xin Chen, Zhongqi Qi, Wen Liu, Zibo Zhao, Zhibin Wang, Bin Fu, Yong Liu, and Gang Yu. Paint3d:" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21412" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 218 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "Paint anything 3d with lighting-less texture diffusion models, 2023. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 128 + ], + "type": "text", + "content": "[61] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 173 + ], + "type": "text", + "content": "[62] Yang Zhou, Zhen Zhu, Xiang Bai, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Non-stationary texture synthesis by adversarial expansion. arXiv preprint arXiv:1805.04487, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "type": "text", + "content": "[63] Song Chun Zhu, Yingnian Wu, and David Mumford. Filters, random fields and maximum entropy (frame): Towards a unified theory for texture modeling. International Journal of Computer Vision, 27:107-126, 1998. 2" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21413" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning/6894bed0-33d2-4729-8056-07cf7489b89f_content_list.json b/2025/Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning/6894bed0-33d2-4729-8056-07cf7489b89f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..43cccaa773f71944c8a2679c4443f2ab416edaed --- /dev/null +++ b/2025/Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning/6894bed0-33d2-4729-8056-07cf7489b89f_content_list.json @@ -0,0 +1,1410 @@ +[ + { + "type": "text", + "text": "Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning", + "text_level": 1, + "bbox": [ + 98, + 130, + 870, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Juntae Lee Munawar Hayat Sungrack Yun \nQualcomm AI Research†", + "bbox": [ + 297, + 181, + 671, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{juntlee, hayat, sungrack}@qti.QUALCOMM.com", + "bbox": [ + 303, + 219, + 666, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 282, + 313, + 300 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Few-shot class incremental learning (FSCIL) enables the continual learning of new concepts with only a few training examples. In FSCIL, the model undergoes substantial updates, making it prone to forgetting previous concepts and overfitting to the limited new examples. Most recent trend is typically to disentangle the learning of the representation from the classification head of the model. A well-generalized feature extractor on the base classes (many examples and many classes) is learned, and then fixed during incremental learning. Arguing that the fixed feature extractor restricts the model's adaptability to new classes, we introduce a novel FSCIL method to effectively address catastrophic forgetting and overfitting issues. Our method enables to seamlessly update the entire model with a few examples. We mainly propose a tripartite weight-space ensemble (Tri-WE). Tri-WE interpolates the base, immediately previous, and current models in weight-space, especially for the classification heads of the models. Then, it collaboratively maintains knowledge from the base and previous models. In addition, we recognize the challenges of distilling generalized representations from the previous model from scarce data. Hence, we suggest a regularization loss term using amplified data knowledge distillation. Simply intermixing the few-shot data, we can produce richer data enabling the distillation of critical knowledge from the previous model. Consequently, we attain state-of-the-art results on the miniImageNet, CUB200, and CIFAR100 datasets.", + "bbox": [ + 75, + 315, + 473, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 750, + 209, + 766 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The ability of a model to continually adapt with the availability of new data is highly desirable for many real-world applications. Few-shot class-incremental learning (FSCIL) has recently gained significant research attention as evidenced by several studies [17, 25, 29, 42, 45], because it offers a promising learning paradigm. It enables the model", + "bbox": [ + 75, + 775, + 468, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/566cc08011a0b6e754bd42e568a4dbf5b71b71a8b6a794a98be3ecbb0c0619be.jpg", + "image_caption": [ + "Figure 1. Overview of the proposed method (FE: feature extractor, Cls. head: classification head). Through Tri-WE, at $t$ th incremental session, Cls. head is composed by interpolating the weights of the Cls. heads covering the classes until base, $(t - 1)$ th, $t$ th sessions, respectively. For ADKD regularization loss, the input few-shot data are amplified to reliably mine generalized knowledge of the model of the previous $(t - 1)$ th session. After incremental learning, deployed components are outlined in green." + ], + "image_footnote": [], + "bbox": [ + 506, + 271, + 898, + 448 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to flexibly expand its classification capabilities to new concepts using only a few training examples. FSCIL starts by training a model on numerous base classes with abundant labeled data. In each subsequent incremental session, the model encounters only a limited number of labeled data for new classes, along with a single prototype for each of the previously learned classes. Hence, FSCIL faces multiple challenges, including the retention of previously learned knowledge and the prevention of overfitting to the limited examples of new data.", + "bbox": [ + 496, + 597, + 892, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these challenges, the most recent works are based on separating learning the representation from the classification. They rely on constructing a robust feature embedding space using plentiful base class data. In subsequent incremental learning where only a few labeled data for new classes are available, the embedding space remains fixed, and the focus is shifted to learning the classifier for the new classes. This strategy is intended to avoid accumulated model collapse or overfitting due to the limited data. For example, [25, 42] refrain from altering any base", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Qualcomm AI Research is an initiative of Qualcomm Technologies, Inc. and/or its subsidiaries.", + "bbox": [ + 75, + 875, + 467, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15329", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "model parameters and use mean features for each class to define their classifiers. In contrast, [17, 29] employ ensembling multiple classifiers to improve generalization. [45] introduces an extra module to assess the reliability of outputs from the base and incrementally learned models for a weighted sum of the two outputs. Moreover, [1, 40] fixes the feature embedding space and also pre-defines classifier templates, requiring prior knowledge of the total number of incremental classes to be encountered at the outset of the base class learning. However, as indicated in Fig. 2, updating the entire model can lead to a drop in base class performance but may enhance the model's adaptability to new classes. Hence, if we address the catastrophic forgetting and overfitting more effectively, we can harness the benefits of updating the entire model.", + "bbox": [ + 75, + 90, + 472, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose a novel framework for FSCIL as depicted in Fig. 1. Mainly, inspired by Wise-FT [37], we develop tripartite weight-space ensemble (Tri-WE) to continually update the weights of the model. We interpolate the classification heads of models from the base, intermediately previous, and current sessions within the weight space, thereby achieving a balance between adaptability to new classes and retention of knowledge of previous ones. This approach allows for a continual update of knowledge without an abrupt shift in the decision boundary. Moreover, to address catastrophic forgetting, knowledge distillation (KD) has been a common regularization technique in class incremental learning. However, its application in a few-shot scenario is prone to overfitting to the few available examples, and then recent cutting-edge methods have less used KD for FSCIL. To counter this, we also introduce amplified data KD (ADKD) regularization loss where we intermix few-shot training datasets, enriching the data source for distilling the knowledge from the previous model. It simply ensures that the model prevents catastrophic forgetting by transferring the generalized knowledge for previous classes. We achieve state-of-the-art (SOTA) results on three benchmarks: miniImageNet, CUB200, and CIFAR100.", + "bbox": [ + 75, + 327, + 472, + 675 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as three-fold:", + "bbox": [ + 96, + 684, + 415, + 699 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the Tri-WE, wherein the classification heads from base, preceding, and current models are synergistically interpolated within the weight space, promoting a knowledge continuum.", + "- Tackling distillation of biased knowledge resulting from a scarcity of data in KD, we develop ADKD sub-loss term that enhances the generality of the distilled knowledge.", + "- Leveraging these proposed components, we attain SOTA results across on miniImageNet, CUB200, and CIFAR100. Comprehensive experiments demonstrate the significance and efficacy of our contributions." + ], + "bbox": [ + 76, + 709, + 467, + 897 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bf9ef0d970d4412fc36a9a386982c91fe9da3e3460b23c92dba24c4266056228.jpg", + "image_caption": [ + "Figure 2. Few-shot (10-way 5-shot miniImageNet) transfer learning (TL) starting from the base session model on each session. Classification accuracy scores on incremented 10 classes (top) & 40 base classes (bottom) are shown for each session." + ], + "image_footnote": [], + "bbox": [ + 509, + 89, + 880, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 339, + 650, + 356 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Class-incremental learning (CIL) aims to seamlessly integrate new classes without access to previously seen data, preventing the catastrophic forgetting problem. To ensure knowledge retention, rehearsal of selective previous samples [3, 28] has been developed. Also, in [35, 39], the network is dynamically expanded to accommodate new learnings. LwF [22] and iCaRL [27] used KD and nearest class mean (NCM) classifiers, respectively, to mitigate the catastrophic forgetting. UCIR [15] introduced a novel feature distillation loss, and PODNet [12] extended distillation beyond the final embedding output, incorporating the pooled outputs from the model's intermediate layers. In [2, 38, 44], the class imbalance problem is tackled. However, these methods depend on ample training data, an aspect not feasible in FSCIL, necessitating new strategies that adapt swiftly with minimal data while preserving existing knowledge.", + "bbox": [ + 496, + 368, + 890, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "FSCIL fuses CIL and few-shot learning. It gained attention after TOPIC [32] initiated the concept, leading to numerous subsequent methods. Recent works align with two strategies. One approach completely fixes the learnable weights of the model after the base session's many-shot many-class training, and then incrementally extends the classifier by adding mean features of new classes. CEC [42] enhances the base model by simulating the FSCIL scenarios sampled from the base data again, while ALICE [25] employs diverse techniques to boost the generalization capability of the base model. Alternatively, some methods [17, 29, 40, 45] update a part of the feature extractor or the added class-mean-initialized classifiers in each session. SoftNet [18] finds the weights less impactful for the catastrophic forgetting at the base session, and focuses on updating them in the incremental learning. In WaRP [19], the important weights are decided in every session, and they keep frozen in the following sessions. GKEAL [47] updates the weight of the classification head using least square", + "bbox": [ + 496, + 613, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15330", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "solver in the view of analytic learning. S3C [17] employed the stochastic classifier as the extreme ensemble of classifiers. NS-FSCIL [40] and OrCo [1] assume the base session knows the amount of total incremental classes and incremental session starts with mapping each class to a classifier predefined in the base session. NS-FSCIL optimizes the last module of the feature encoder and OrCo does the last module and the selected pre-defined classifier. In SAVC [29], a class is recognized by a set of classifiers, each tailored to a specific input augmentation, resulting in a swift expansion of deployed model size during incremental learning. In BiDistill [45], every prediction is made by the ensemble of the base and last models.", + "bbox": [ + 75, + 90, + 472, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Network ensemble has long been an important technique in deep learning in order to reduce overfitting and improve generalization ability [10, 13]. It involves the training of multiple neural networks and combining them for the final prediction. Various ensemble methods have been developed to get different but collaborative models such as random-weight initialization [21], different training data [30], and capturing multiple snapshots in training schedules [16]. While these methods have consistently demonstrated improved performance over single-model approaches, they often come at the cost of increased computational overhead, both in terms of training and inference. Recently, Model Soup [36] suggested interpolating the weights of a bunch of networks finetuned with different hyperparameters configurations. Also, to mitigate the overfitting of the finetuned CLIP [26], WiseFT [37] combines the weights of fine-tuned and initial CLIP models. These approaches can attain better generalization capability without increasing the computational cost of deployed models.", + "bbox": [ + 75, + 287, + 472, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 76, + 588, + 250, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we describe the proposed Tri-WE and ADKD for FSCIL. The pipeline is illustrated in Fig. 3. We also provide the process of the base session training.", + "bbox": [ + 75, + 613, + 468, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem setting. FSCIL aims to incrementally scale the model's class recognition ability by learning new classes with limited training data while preserving its knowledge of previously learned, old classes. In specific, we suppose that a model $f_{\\theta,\\phi}$ consists of the feature extractor $g_{\\theta}$ and the classification head (cosine classifier) $h_{\\phi}$ where $\\theta$ and $\\phi$ denote learnable weights of them, respectively. Initially, the model is pre-trained on the base training set in the base session $S^{(0)}$ . Then, the model progresses a series of $T$ incremental sessions, $\\{S^{(1)},\\ldots,S^{(T)}\\}$ . Current session $S^{(t)}$ includes a class set $C^{(t)}$ and the corresponding training dataset $D^{(t)} = \\{(x,y)\\}$ where $x$ is a sample and $y$ is its label.", + "bbox": [ + 75, + 659, + 468, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the base session $S^{(0)}$ , the class set $\\mathcal{C}^{(0)}$ includes a substantial number of $N_0$ base classes with ample training data for each class. In contrast, in the incremental session $S^{(t)}$ ( $t \\geq 1$ ), the class set $\\mathcal{C}^{(t)}$ contains $N$ classes with only a", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "small $K$ number of training data available for every class $(N$ -way $K$ -shot setting). Notice that $\\mathcal{C}^{(i)}$ and $\\mathcal{C}^{(j)}$ do not overlap when $i \\neq j, \\forall i, j$ . Following the prior FSCIL works [32, 42], the model at each session $\\mathcal{S}^{(t)}$ only has access to $\\mathcal{D}^{(t)}$ and an additional prototype (the averaged feature of the $K$ examples per class) buffer $\\mathcal{M}$ memorizing the classes of the base and previous $(t - 1)$ sessions. A single prototype is maintained for every class in $\\mathcal{M}$ . Once the incremental learning at the session $\\mathcal{S}^{(t)}$ concludes, the model is evaluated on query data from all the seen $N_{t}$ classes thus far, that is $\\mathcal{C}^{(0:t)} = \\mathcal{C}^{(0)} \\cup \\mathcal{C}^{(1)} \\cdots \\cup \\mathcal{C}^{(t)}$ .", + "bbox": [ + 496, + 90, + 893, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Tripartite weight-space ensemble (Tri-WE)", + "text_level": 1, + "bbox": [ + 498, + 297, + 864, + 314 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The base session model is generalized by numerous classes with sufficient data, and hence its weight space is conducive to adapting to new classes. Although the weight space is somewhat compromised, the intermediate previous session's model has knowledge for all the classes encountered thus far. To capitalize on the strengths of both models, we suggest the weight-space ensemble of three models from the sessions $S^{(0)}$ , $S^{(t-1)}$ , and $S^{(t)}$ .", + "bbox": [ + 496, + 330, + 890, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, the feature extractor, which contains many layers, is prone to overfitting or collapsing when only a few examples are available. Hence, while reducing the learning rate for the feature extractor $g_{\\theta}^{(t)}$ , we only ensemble the classification heads $h_{\\phi_0}$ , $h_{\\phi_{\\mathrm{old}}}^{(t)}$ , $h_{\\phi_{\\mathrm{all}}}^{(t)}$ in the weight space. Here, $h_{\\phi_0}$ is the classification head from the base session model. $h_{\\phi_{\\mathrm{old}}}^{(t)}$ is initialized by the classification head $h_{\\phi}^{(t - 1)}$ from the session $S^{(t - 1)}$ , and then specialized to the old classes $\\mathcal{C}^{(0:t - 1)}$ . $h_{\\phi_{\\mathrm{all}}}^{(t)}$ covers all the target classes $\\mathcal{C}^{(0:t)}$ , where the old class part is initialized similarly to $h_{\\phi_{\\mathrm{old}}}^{(t)}$ . Then, as depicted in Fig. 3, Tri-WE interpolates the classification head weight $\\phi^n$ for $n$ th class by", + "bbox": [ + 496, + 462, + 892, + 661 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\phi^ {n} = \\\\ \\left\\{ \\begin{array}{l l} \\bar {\\alpha} _ {1} \\phi_ {0} ^ {n} + \\bar {\\alpha} _ {2} \\phi_ {\\text {o l d}} ^ {n} + \\bar {\\alpha} _ {3} \\phi_ {\\text {a l l}} ^ {n} & \\text {i f} n \\leq N _ {0}, \\\\ \\bar {\\alpha} _ {4} \\phi_ {\\text {o l d}} ^ {n} + \\bar {\\alpha} _ {5} \\phi_ {\\text {a l l}} ^ {n} & \\text {i f} N _ {0} < n \\leq N _ {t - 1}, \\\\ \\phi_ {\\text {a l l}} ^ {n} & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 694, + 890, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $1 \\leq n \\leq N_t$ , $\\phi^n \\in \\mathbb{R}^d$ , and $d$ is the dimensionality of $\\phi^n$ . Then, the weight of the classification head $h_\\phi^{(t)}$ to classify $N_t$ classes consists of $\\{\\phi^n\\}_{n=1}^{N_t}$ . Notice that, we also add two learnable scalar weights $\\alpha_1$ and $\\alpha_2$ to automatically adjust the trade-offs between the three classification", + "bbox": [ + 496, + 819, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15331", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b5f2cc96519b842abc8f0f82a8207496d3862708e3cb9799d03e4056e431e080.jpg", + "image_caption": [ + "Figure 3. The pipeline of the proposed method. At the $t$ th incremental session $S^{(t)}$ , the model is initialized using the resulting model of session $S^{(t-1)}$ . Then, a set of $N$ -way $K$ -shot few-shot examples $D^{(t)}$ and the prototype buffer $\\mathcal{M}$ are given. The weights $\\phi_0$ , $\\phi_{\\mathrm{old}}$ , $\\phi_{\\mathrm{all}}$ of classification heads $h_{\\phi_0}$ , $h_{\\phi_{\\mathrm{old}}}^{(t)}$ , $h_{\\phi_{\\mathrm{all}}}^{(t)}$ are interpolated with learnable scalars $\\alpha_1$ and $\\alpha_2$ . The main loss $\\mathcal{L}_{\\mathrm{Cls}}$ is computed on the resulting classification head $h_{\\phi}^{(t)}$ 's output. Also, classification loss over the old classes, appeared until the session $S^{(t-1)}$ , is computed on the output of $h_{\\phi_{\\mathrm{old}}}^{(t)}$ . Additionally, $D^{(t)}$ is amplified to $D_{\\mathrm{amp}}^{(t)}$ , and then KD losses $\\mathcal{L}_{\\mathrm{feat}}$ and $\\mathcal{L}_{\\mathrm{logit}}$ are computed on the amplified dataset $D_{\\mathrm{amp}}^{(t)}$ . The learned components are outlined in red." + ], + "image_footnote": [], + "bbox": [ + 117, + 88, + 854, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "heads. Then, $\\{\\bar{\\alpha}_1,\\bar{\\alpha}_2,\\bar{\\alpha}_3,\\bar{\\alpha}_4,\\bar{\\alpha}_5\\}$ are obtained by", + "bbox": [ + 76, + 507, + 406, + 523 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\alpha} _ {1} = \\alpha_ {1} / (\\alpha_ {1} + \\alpha_ {2} + 1),\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 530, + 357, + 546 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\alpha} _ {2} = \\alpha_ {2} / \\left(\\alpha_ {1} + \\alpha_ {2} + 1\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 549, + 357, + 565 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\alpha} _ {3} = 1 / \\left(\\alpha_ {1} + \\alpha_ {2} + 1\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 568, + 348, + 583 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\alpha} _ {4} = \\alpha_ {2} / (\\alpha_ {2} + 1),\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 585, + 320, + 602 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\alpha} _ {5} = 1 / \\left(\\alpha_ {2} + 1\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 606, + 468, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When $t = 1$ (i.e. $\\phi_{\\mathrm{old}} = \\phi_0$ ), we set $\\alpha_{2} = 0$ .", + "bbox": [ + 76, + 628, + 369, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, the main classification loss $\\mathcal{L}_{\\mathrm{Cls}}$ is computed on the output from the weight-space ensembled classification head $h_{\\phi}^{(t)}$ by", + "bbox": [ + 75, + 643, + 468, + 691 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {C l s}} = \\mathbb {E} _ {(x, y) \\sim \\mathcal {D} ^ {(t)}} [ \\mathrm {C E} (h _ {\\phi} ^ {(t)} (g _ {\\theta} ^ {(t)} (x)), y ] \\\\ + \\mathbb {E} _ {(\\mathbf {p}, y) \\sim \\mathcal {M}} [ \\mathrm {C E} \\left(h _ {\\phi} ^ {(t)} (\\mathbf {p}), y \\right], \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 700, + 468, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where CE is the cross-entropy loss function, and $\\mathbf{p}$ denotes the old class prototype in $\\mathcal{M}$ .", + "bbox": [ + 76, + 750, + 468, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using the buffer $\\mathcal{M}$ , we also calculate the old class-specific loss $\\mathcal{L}_{\\mathrm{Cls - Old}}$ for the classification head $h_{\\phi_{\\mathrm{old}}}^{(t)}$ by", + "bbox": [ + 76, + 781, + 468, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C l s - O l d}} = \\mathbb {E} _ {(\\mathbf {p}, y) \\sim \\mathcal {M}} [ \\mathrm {C E} \\left(h _ {\\phi_ {\\text {o l d}}} ^ {(t)} (\\mathbf {p}), y\\right) ]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 821, + 468, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this loss, $h_{\\phi_{\\mathrm{old}}}^{(t)}$ focuses on preserving the discriminative ability regarding all the old classes on top of the current feature extractor $g_{\\theta}^{(t)}$ . Hence, we do not update $h_{\\phi_0}$ and instead", + "bbox": [ + 76, + 851, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "use it to retain the knowledge for the original, well-designed base class distribution from the base session model.", + "bbox": [ + 498, + 507, + 890, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike the naive ensemble approaches where multiple classification heads independently make predictions and combine them in the testing phase, our weight-space ensemble deploys only $h_{\\phi}^{(t)}$ collaboratively utilizing the knowledge from the base and previous session models.", + "bbox": [ + 496, + 541, + 890, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Total training loss", + "text_level": 1, + "bbox": [ + 500, + 638, + 673, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "KD has been a widely used regularization technique in CIL to prevent the catastrophic forgetting problem [27]. Specifically, when new classes are introduced, an additional distillation loss is combined with the main classification loss to ensure that the current model replicates the old class behavior of the previous model. Nevertheless, in FSCIL, the limited data source can lead to the extraction of knowledge biased toward those data, rather than generalized knowledge of the old classes. To mitigate this issue, we simply generate an amplified data source $\\mathcal{D}_{\\mathrm{amp}}^{(t)}$ from the few-shot training data by randomly mixing pairs of data from $\\mathcal{D}^{(t)}$ employing simple data augmentation techniques such as [41, 43]. The ADKD loss is then computed as", + "bbox": [ + 496, + 666, + 890, + 864 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {A D K D}} = \\mathcal {L} _ {\\text {f e a t}} + \\mathcal {L} _ {\\text {l o g i t}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 885, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15332", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{feat}}$ and $\\mathcal{L}_{\\mathrm{logit}}$ are feature-level and logit-level KD losses which are defined by", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {f e a t}} = \\mathbb {E} _ {x \\sim \\mathcal {D} _ {\\mathrm {a m p}} ^ {(t)}} | | g _ {\\theta} ^ {(t - 1)} (x) - g _ {\\theta} ^ {(t)} (x) | | _ {2} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 131, + 468, + 155 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {l o g i t}} = \\mathbb {E} _ {x \\sim \\mathcal {D} _ {\\mathrm {a m p}} ^ {(t)}} - f _ {\\theta , \\phi} ^ {(t - 1)} (x) \\log f _ {\\theta , \\phi^ {0: N _ {t - 1}}} ^ {(t)} (x). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 157, + 468, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The previous session model $f_{\\theta, \\phi}^{(t-1)}$ including $h_{\\phi}^{(t-1)}$ and $g_{\\theta}^{(t-1)}$ is kept frozen during this process, unlike $h_{\\phi_{\\mathrm{old}}}^{(t)}$ .", + "bbox": [ + 76, + 193, + 468, + 231 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the total loss function used to optimize the current model is", + "bbox": [ + 76, + 232, + 467, + 258 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {C l s}} + \\gamma_ {1} \\mathcal {L} _ {\\mathrm {C l s - O l d}} + \\gamma_ {2} \\mathcal {L} _ {\\mathrm {A D K D}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 273, + 468, + 289 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma_{1}$ and $\\gamma_{2}$ are hyper-parameters that balance the respective losses, $\\mathcal{L}_{\\mathrm{Cls - Old}}$ and $\\mathcal{L}_{\\mathrm{ADKD}}$ . We provide the pseudo code of the overall process during an incremental session in the supplementary materials.", + "bbox": [ + 76, + 301, + 468, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Base session training", + "text_level": 1, + "bbox": [ + 76, + 369, + 272, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In FSCIL, it is broadly acknowledged that a model with strong generalization to base classes is beneficial for learning new classes [25, 29, 40, 42]. To ensure fairness when assessing the proposed components for incremental learning, we strive for base session performance on par with the prior SOTA methods [29, 40, 45]. For this purpose, we employ the base session training techniques of ALICE [25] (see the literature for details). On top of it, we also added a geometric classification head in parallel with the main classification head. In specific, an input is geometrically transformed into $B$ types, and we use them as the proxy labels of transformed inputs. Then, the geometric classification head is learned to classify each into one of the $B$ proxy labels. This auxiliary geometric classification helps the model encode the inherent structure of data so that the model's generalization capability on new tasks is improved. This auxiliary geometric classification strategy is not used in the incremental sessions. Through this, as aftermentioned in Sec. 4, the base session model achieves comparable or superior performance to previous SOTA methods on three benchmark datasets. A more detailed description is in the supplementary materials.", + "bbox": [ + 75, + 393, + 468, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 76, + 739, + 282, + 756 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Consistent with the setting of [32], we conduct our experiments on three benchmark datasets: miniImageNet [33], CIFAR100 [20] and CUB200 [34]. miniImageNet is a subset of ImageNet [8] that includes 60,000 images from 100 classes, each with resolution $84 \\times 84$ . CIFAR100 comprises 60,000 tiny images of size $32 \\times 32$ across 100 categories. CUB200, which is focused on a fine-grained classification, contains 200 bird species with similar appearance where the image size is $224 \\times 224$ . We follow", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the splits from [32]. For miniImageNet and CIFAR100, 60 categories are selected as base classes while the remaining are split into 8 incremental sessions with only 5 training examples per novel class (i.e., 5-way 5-shot). As for the CUB200 dataset, 100 categories are designated as the base training set, while the rest are divided into 10-way 5-shot tasks across 10 sessions.", + "bbox": [ + 496, + 90, + 890, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation protocol. In line with the FSCIL literature, we employ the top-1 accuracy (Acc) for all the observed classes at the end of each incremental learning session. We also provide the average of Acc over all the sessions.", + "bbox": [ + 496, + 196, + 890, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. We implement our method using the PyTorch library. In keeping with the previous works [18, 25, 29, 45], ResNet18 [14] serves as our feature extractor. The entire model is trained via the SGD optimizer. For the base session, the learning rates are initialized by 0.01, 0.001, and 0.01 for miniImageNet, CUB200, and CIFAR100 respectively, and decayed by 0.1 at 60 and 70 epochs. In each incremental session, the learning rate for the weight-space ensembled $\\phi_{\\mathrm{all}}$ is 0.1, while it is 0.001 for the rest. Hence, the feature extractor is minimally updated in incremental learning. For the ADKD loss, the given $NK$ training examples are amplified to $16NK$ for each class. Also, the learnable scalars $\\alpha_{1}$ and $\\alpha_{2}$ are initially set to 1.0. Loss weights $\\lambda_{1}$ and $\\lambda_{2}$ are empirically determined to be 1.2 and 10.0, respectively. Further details are in Sec. S-5 of the supplementary materials.", + "bbox": [ + 496, + 258, + 892, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Comparative assessment", + "text_level": 1, + "bbox": [ + 498, + 513, + 723, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "On the three benchmark datasets, we first compare the proposed method with the existing approaches. As indicated in Table 1 on the miniImageNet dataset, our method outperforms the compared in all the sessions, achieving a clear SOTA in terms of the average Acc by a margin of $1.82\\%$ at least. Also, BiDistill and SAVC require the base session model and a considerable number of classifiers per class, respectively, even in the testing phase. NC-FSCIL and OrCo benefit from knowing the total number of incremental classes at the outset of the base session. Whereas, we attain clear SOTA with a smaller-scale deployed model compared to BiDistill and SAVC, and without the information advantage of NC-FSCIL and OrCo.", + "bbox": [ + 496, + 537, + 890, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Moreover, Fig. 4 shows the Acc curve according to sessions on the CUB200 and CIFAR100 datasets. As in the miniImageNet dataset, our method consistently surpasses the compared methods by large gaps in the incremental sessions on the CUB 200 dataset. In specific, despite lower base session Acc than SAVC, our method outperforms it in the last session. Also, while NC-FSCIL merely finetuned the last module of the feature encoder in incremental sessions, OrCo attempted to also optimize the pre-set classifiers. Nevertheless, the additional update of only the classifiers is less effective. On the CIFAR100 dataset, we can", + "bbox": [ + 496, + 734, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15333", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ab3053c90cc09bb66d0de2d7a4c009aa9eafafbeb491eb79b4fbe385f263e244.jpg", + "table_caption": [ + "Table 1. Comparative results on miniImageNet dataset. Session-wise Acc (%) and the average of them (Avg) are reported. In each case, the best and second results are bold-faced and underlined, respectively. Last sess. Impro. denotes the performance improvement of our method at the last session." + ], + "table_footnote": [], + "table_body": "
MethodVenueAcc in each session (%)AvgLast sess. impro.
012345678
iCaRL [27]CVPR 201761.3146.3242.9437.6330.4924.0020.8918.8017.2033.29+43.13
TOPIC [32]CVPR 202061.3150.0945.1741.1637.4835.5232.1929.4624.4239.64+35.71
ERL++ [11]AAAI 202161.7057.5854.6651.7248.6646.2744.6742.8140.7949.87+19.34
IDLVQ [4]ICLR 202064.7759.8755.9352.6249.8847.5544.8343.1441.8451.16+18.29
CEC [42]CVPR 202172.0066.8362.9759.4356.7053.7351.1949.2447.6357.74+12.50
SynthFeat [5]ICCV 202161.4059.8054.2051.6949.4548.0045.2043.8042.1050.63+18.03
MetaFSCIL [6]CVPR 202272.0467.9463.7760.2957.5855.1652.950.7949.1958.85+10.94
FACT [46]CVPR 202272.5669.6366.3862.7760.6057.3354.3452.1650.4960.70+9.64
Replay [23]ECCV 202271.8467.1263.2159.7757.0153.9551.5549.5248.2158.02+11.92
ALICE [25]ECCV 202280.6070.6067.4064.5062.5060.0057.8056.8055.7063.99+4.43
S3C [17]ECCV 202276.5571.7467.6664.5261.5158.0955.3653.0451.3462.20+8.79
WaRP [19]ICLR 202372.9968.1064.3161.3058.6456.0853.4051.7250.6559.69+9.48
SoftNet [18]ICLR 202376.6370.1365.9262.5259.4956.5653.7151.7250.4860.79+9.65
NC-FSCIL [40]ICLR 202384.0276.8072.0067.8366.3564.0461.4659.5458.3167.82+1.82
GKEAL [47]CVPR 202373.5968.9065.3362.2959.3956.7054.2052.5951.3160.48+10.40
BiDistill [45]CVPR 202374.6570.4366.2962.7760.7557.2454.7953.6552.2261.42+7.91
SAVC [29]CVPR 202381.1276.1472.4368.9266.4862.9559.9258.3957.1167.05+3.02
OrCo [1]CVPR 202483.3075.3271.5268.1665.6263.1160.2058.8258.0867.12+2.05
CLOSER [24]ECCV 202476.0271.6167.9964.6961.7058.9456.2354.5253.3362.78+6.80
Ours84.1381.4176.6573.5970.1065.1363.4261.0260.1370.62
", + "bbox": [ + 120, + 141, + 849, + 417 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/be9d2f5bd94ff302ff07b19ff3c3c98a62f8cd9da1d84b65295a4d3614f017c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 434, + 753, + 460 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/10d9b2ecc6b2a2673e4c4b064c54d65b6ce092849db4b10e3f511e6fb2912e96.jpg", + "image_caption": [ + "(a) CUB200" + ], + "image_footnote": [], + "bbox": [ + 96, + 465, + 290, + 612 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/05a5fc2d0452e227a6f8c82e86b080a735b41a885e95e1c215847e1eed607491.jpg", + "image_caption": [ + "Figure 4. Comparative results on CUB200 and CIFAR100 datasets. For each dataset, the curve of the session-wise Acc (%) is first provided. And also, the base, last, average Acc are reported on our method comparing with three recent SOTA methods, NC-FSCIL [40], SAVC [29], YourSelf [31] (Best viewed in color)." + ], + "image_footnote": [], + "bbox": [ + 292, + 467, + 482, + 606 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9e0824fa1e85b63a4eedcb5a36fa5ae07e2e1c4934c2e59b53a376a2b3d62d9c.jpg", + "image_caption": [ + "(b) CIFAR100" + ], + "image_footnote": [], + "bbox": [ + 488, + 465, + 683, + 612 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6df7fd8303699a4f3d92d890b86dacfb9ab7ad71f4cc93eeaaf7e835c776fb61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 683, + 467, + 874, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "see the similar tendency. NC-FSCIL's setting is somewhat favorable, as it allows the pre-definition of classifiers for the incremental classes with knowledge of all class numbers in advance at the base session. This advantage is particularly beneficial on the CIFAR100 dataset, where some classes are semantically similar (e.g., possum and mouse), and the low resolution $(32 \\times 32)$ of the images could otherwise complicate classification. The detailed results are available in Sec. S-4 of the supplementary materials. These findings demonstrate our method's superior ability to adapt to new classes with limited training data while maintaining seamless integration with the old classes.", + "bbox": [ + 75, + 710, + 472, + 893 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation analysis", + "text_level": 1, + "bbox": [ + 500, + 710, + 669, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We extensively analyze the components of the proposed method on the miniImageNet or CUB200 datasets.", + "bbox": [ + 496, + 734, + 890, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tripartite strategy. We conduct the ablation experiments on the Tri-WE in Table 2. Before this, initially, we contrast \"Naive\" with the \"No WE\" configuration, the latter incorporating the ADKD loss. The naive approach, lacking both WE and ADKD loss, exhibits a significant performance decline due to simultaneous updates of the feature extractor and classification head. However, No WE achieves respectable results thanks to the ADKD loss (more discussed in the following sections). Next, as demonstrated", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15334", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/05f438ad826f452223d11bce50fd7dee8f4c3dc81ad3bd408212423e704a2c56.jpg", + "table_caption": [ + "Table 2. Ablation analysis for Tri-WE. Varying the combination of classification heads ensembled in weight space (No WE, Dual-WE, Tri-WE), session-wise and average Acc (%) are provided on miniImageNet. Also, \"Naive\" means that none of WE and ADKD is used." + ], + "table_footnote": [], + "table_body": "
Classification head weightAcc in each session (%)Avg
φ0φoldφall012345678
Naive84.1380.5571.2660.9350.0544.0836.4219.7816.7351.55
No WE✓84.1378.9574.0970.0666.5663.0760.3258.0856.0167.92
Dual-WE✓✓84.1380.2575.4771.5867.6264.5261.1160.1758.9369.31
✓84.1379.5474.9970.1366.6063.4360.8559.8957.3168.54
Tri-WE✓✓✓84.1381.4176.6573.5970.165.1363.4261.0260.1370.62
", + "bbox": [ + 99, + 128, + 874, + 247 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/24730b7ef67aac33f7a18db9dc464fec0d9edea1a015ffa5350138cbc617c008.jpg", + "table_caption": [ + "Table 3. Impact on increasing the layers ensembled in weight space. Applying the proposed Tri-WE to the convolutional blocks of the feature extractor (ResNet18), session-wise and average Acc (%) are provided on miniImageNet." + ], + "table_footnote": [], + "table_body": "
Weight-space ensembledAcc in each session (%)Avg
012345678
Blocks 1-484.1375.2570.2567.5464.8561.3459.7456.3853.4565.88
Blocks 1-4 + clf. head84.1376.0872.1368.4265.8362.6060.1957.7555.2966.94
Blocks 2-4 + clf. head84.1377.8473.4269.2265.8562.7560.5657.3455.3367.38
Blocks 3-4 + clf. head84.1378.5974.4369.8066.2563.1360.4658.0455.6567.83
Block 4 + clf. head84.1378.8074.2169.6466.0163.3261.5058.1755.8867.96
Clf. head only84.1381.4176.6573.5970.165.1363.4261.0260.1370.62
", + "bbox": [ + 161, + 297, + 810, + 416 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e492489135de153352ab97b8a4ab4066973d26ddffee44f146a4bdfea9c34d3a.jpg", + "image_caption": [ + "(a) Fixed $(\\alpha_{1},\\alpha_{2})$", + "Figure 5. Analysis on learnable scalars $\\alpha_{1}$ and $\\alpha_{2}$ on miniImageNet. (a) Comparison with multiple combinations of fixed (not-learnable) $\\alpha_{1}$ and $\\alpha_{2}$ in terms of averaged Acc over all the sessions, (b) session-wise learned $\\alpha_{1}$ and $\\alpha_{2}$ in ours." + ], + "image_footnote": [], + "bbox": [ + 96, + 440, + 285, + 587 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d9c3609745a907048338b9539db934e28b6dbbbc2989d41a7c3dbc047c3e5cf2.jpg", + "image_caption": [ + "(b) Learned $\\alpha_{1},\\alpha_{2}$" + ], + "image_footnote": [], + "bbox": [ + 292, + 440, + 452, + 587 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in Table 2, the absence of the weight-space ensemble for the classification head leads to increased catastrophic forgetting. And, from \"Dual-WE,\" we can see that the model from the intermediate previous session is more important than the base session model, as the reduced learning rate for the feature extractor mitigates drastic changes, helping to preserve the integrity of the feature space. Accordingly, the knowledge for incrementally-appeared classes is more helpful. Lastly, the complete Tri-WE clearly outperforms No WE by $2.70\\%$ in average Acc, effectively retaining the essential knowledge of the base and previous session models.", + "bbox": [ + 75, + 702, + 468, + 882 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "WE in classification head. We also evaluate the effect of", + "bbox": [ + 76, + 885, + 470, + 898 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/157a066153e97062c77a965fab1ec9d4abfbcfeb68c7cc1f12964287ce12b536.jpg", + "table_caption": [ + "Table 4. Analysis on data amplification schemes. The last and average Acc (%) are reported on each scheme on miniImageNet." + ], + "table_footnote": [], + "table_body": "
w/o ADAD scheme
CutOut [9]RandAug [7]MixUp [43]CutMix [41]
Last Acc58.0858.1258.2659.1460.13
Avg Acc69.0569.1369.0970.0370.62
", + "bbox": [ + 503, + 477, + 890, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "weight-space ensemble varying the number of interpolated layers in Table 3. In specific, we apply the proposed TriWE on the convolutional blocks of the feature extractors. Here, we use different learnable scalars depending on the convolutional blocks, rather than sharing them across the convolutional blocks and classification head. We observed a tendency for the decline in performance as the number of weight-space-ensemblied layers increases. This is because that deciding how to merge a lot of layers is difficult with only a few examples. Also, only merging feature extractor (1st row of Table 3) shows the lowest performance. It emphasizes that the head's weight-space is crucial to prevent catastrophic forgetting.", + "bbox": [ + 496, + 566, + 890, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Learnable scalars $\\alpha_{1} \\& \\alpha_{2}$ . Fig. 5(a) shows the averaged Acc on diverse combinations of $(\\alpha_{1}, \\alpha_{2})$ where both are fixed after initialization in all the sessions. Compared to all the fixed cases, the learnable scalars show better result (70.62% in the averaged Acc). Also, as Fig. 5(b), we show the learned $\\alpha_{1}$ and $\\alpha_{2}$ in each session in our method where we initialize both as 1.0 in every session. We can see that the learned $\\alpha_{1}$ and $\\alpha_{2}$ are different depending on sessions. Accordingly, our learnable scalar approach is useful for bal", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15335", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6da194ee50ebfca4b9ace4956292922da1c80391e830f60e24dfc073861a201f.jpg", + "table_caption": [ + "Table 5. Analysis on loss terms of ${\\mathcal{L}}_{\\mathrm{{ADKD}}}$ . The last and average Acc (%) are reported abating the loss terms on miniImageNet." + ], + "table_footnote": [], + "table_body": "
LfeatLlogitLast AccAvg Acc
✓59.4270.01
✓58.7768.41
✓✓60.1370.62
", + "bbox": [ + 178, + 128, + 370, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ancing the trade-off among the three classification heads. We also provide session-wise results on the fixed cases in the supplementary materials.", + "bbox": [ + 75, + 214, + 467, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of data amplification. While Tri-WE already attains SOTA (i.e. w/o AD in Table 4), incorporating our ADKD loss provides further increase (69.05% vs 70.62% in average Acc). Also, to see the influence of the choice of data amplification scheme in the ADKD loss, we compare multiple data augmentation methods for our data amplification. To facilitate the use of ADKD in follow-up works, we consider four widely-used data augmentation methods, CutOut [9], RandAug [7], CutMix [41], MixUp [43]. In the first two, the images are not inter-mixed. Whereas, two images are mixed for data augmentation. In Table 4, we can see that CutOut and RandAug are not effective compared to \"w/o AD\" where no data amplification is considered. With only a few examples, those methods cannot yield a diversity of data. Next, in comparison with MixUp, CutMix shows a better result. MixUp mixes two images smoothly, and then the original semantics are somewhat preserved. This is less useful in KD. Namely, CutMix encourages the model to focus more on general image recognition ability rather than high-level semantics, which is more helpful in preventing the catastrophic forgetting problem in FSCIL. Hence, we employ CutMix as the default option. In addition, the impact of the degree of data amplification is addressed in Sec. S-1 of the supplementary materials.", + "bbox": [ + 75, + 261, + 468, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Two loss terms of $\\mathcal{L}_{\\mathrm{ADKD}}$ . As shown in the first two rows of Table 5, we ablate $\\mathcal{L}_{\\mathrm{feat}}$ and $\\mathcal{L}_{\\mathrm{logit}}$ , respectively. From this study, we can see that $\\mathcal{L}_{\\mathrm{feat}}$ is slightly more important than $\\mathcal{L}_{\\mathrm{logit}}$ since Tri-WE already helps to prevent catastrophic forgetting in the side of the classification head. Then, using both is most effective.", + "bbox": [ + 75, + 625, + 467, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{\\mathrm{Cls - Old}}$ . As we update the feature extractor as well as the classification head in the incremental learning, we also adapt $\\phi_{\\mathrm{old}}$ to the updated feature extractor using the loss $\\mathcal{L}_{\\mathrm{Cls - Old}}$ . To verify this, we conduct ablating the feature extractor update and $\\mathcal{L}_{\\mathrm{Cls - Old}}$ . As shown in Table 6, the performance is degraded without this loss term when the feature extractor is updated. Also, when the feature extractor is frozen, updating $\\phi_{\\mathrm{old}}$ is less meaningful. Finally, updating the feature extractor with $\\phi_{\\mathrm{old}}$ is most beneficial.", + "bbox": [ + 75, + 717, + 467, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "New class adaptability. To see the efficacy of our method in terms of the trade-off between base and newly introduced classes, we provide the individual accuracy of the", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ac886ed8f1a56727db76043a51205099a6b83a972bf1e851cdfeb7403cad5d8c.jpg", + "table_caption": [ + "Table 6. Analysis on ${\\mathcal{L}}_{\\text{Cls-Old }}$ and updating the feature extractor ${g}_{\\theta }$ . The last and average Acc (%) are reported,on miniImageNet." + ], + "table_footnote": [], + "table_body": "
LCls-OldUpdate gθLast AccAvg Acc
✓59.2770.15
✓✓60.1370.62
58.8669.41
✓58.8869.40
", + "bbox": [ + 588, + 128, + 802, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/82e4ec3c0847b3fb628da1b5073e460b37cbff7168326273a7898dcff36d4f07.jpg", + "image_caption": [ + "Figure 6. Analysis on the new class adaptability on CUB200. The average of accuracy scores are provided for base and novel classes, denoted by Base and New respectively. And, harmonic means are also compared." + ], + "image_footnote": [], + "bbox": [ + 503, + 222, + 888, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "base and novel classes, and the harmonic mean (HM). Since most previous works have addressed the joint accuracy of the base and novel classes, we can compare our method with only a few recent methods [17, 33, 42, 45, 46] on the CUB200 dataset, following [17, 46]. Fig. 6 shows that our approach is more adept at integrating new classes without compromising the performance on the base classes.", + "bbox": [ + 498, + 446, + 890, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 566, + 625, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We explored FSCIL which facilitates the continuous acquisition of new concepts from a few examples. To this end, we suggested a novel FSCIL method including Tri-WE and ADKD regularization loss. The proposed method can effectively prevent overfitting to the few data and catastrophic forgetting of old concepts in the model. Tri-WE strategically interpolates the weights of the classification heads from the base, immediate previous, and current models, which collectively preserves the knowledge from the earlier models without overwhelming updates. Enriching the few examples for KD, the ADKD loss facilitates the transfer of essential knowledge from the previous model, rather than the knowledge overfitted to the few examples. We conducted extensive analyses on the proposed components and achieved state-of-the-art performance on the miniImageNet, CUB200, and CIFAR100 datasets.", + "bbox": [ + 498, + 590, + 890, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 847, + 596, + 863 + ], + "page_idx": 7 + }, + { + "type": "ref_text", + "text": "[1] Noor Ahmed, Anna Kukleva, and Bernt Schiele. Orco: Towards better generalization via orthogonality and contrast for", + "bbox": [ + 506, + 872, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15336", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "few-shot class-incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2024.", + "[2] Eden Belouadah and Adrian Popescu. II2m: Class incremental learning with dual memory. In IEEE Conf. Comput. Vis. Pattern Recog., 2019.", + "[3] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. arXiv preprint arXiv:1812.00420, 2018.", + "[4] Kuilin Chen and Chi-Guhn Lee. Incremental few-shot learning via vector quantization in deep embedded space. In Int. Conf. Learn. Represent., 2020.", + "[5] Ali Cheraghian, Shafin Rahman, Sameera Ramasinghe, Pengfei Fang, Christian Simon, Lars Petersson, and Mehrtash Harandi. Synthesized feature based few-shot class-incremental learning on a mixture of subspaces. In Int. Conf. Comput. Vis., 2021.", + "[6] Zhixiang Chi, Li Gu, Huan Liu, Yang Wang, Yuanhao Yu, and Jin Tang. Meta-scil: A meta-learning approach for few-shot class incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2022.", + "[7] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In IEEE Conf. Comput. Vis. Pattern Recog. Worksh., 2020.", + "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conf. Comput. Vis. Pattern Recog., 2009.", + "[9] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017.", + "[10] Thomas G Dietterich. Ensemble methods in machine learning. In International Workshop on Multiple Classifier Systems, 2000.", + "[11] Songlin Dong, Xiaopeng Hong, Xiaoyu Tao, Xinyuan Chang, Xing Wei, and Yihong Gong. Few-shot class incremental learning via relation knowledge distillation. In AAAI, 2021.", + "[12] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In Eur. Conf. Comput. Vis., 2020.", + "[13] Lars Kai Hansen and Peter Salamon. Neural network ensembles. IEEE Trans. Pattern Anal. Mach. Intell., 12(10): 993-1001, 1990.", + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., 2016.", + "[15] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In IEEE Conf. Comput. Vis. Pattern Recog., 2019.", + "[16] Gao Huang, Yixuan Li, Geoff Pleiss, Zhuang Liu, John E Hopcroft, and Kilian Q Weinberger. Snapshot ensembles: Train 1, get m for free. In Int. Conf. Learn. Represent. 2017.", + "[17] Jayateja Kalla and Soma Biswas. S3c: Self-supervised stochastic classifiers for few-shot class-incremental learning. In Eur. Conf. Comput. Vis., 2022." + ], + "bbox": [ + 78, + 92, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Haeyong Kang, Jaehong Yoon, Sultan Rizky Hikmawan Madjid, Sung Ju Hwang, and Chang D Yoo. On the soft-subnetwork for few-shot class incremental learning. In Int. Conf. Learn. Represent., 2023.", + "[19] Do-Yeon Kim, Dong-Jun Han, Jun Seo, and Jaekyun Moon. Warping the space: Weight space rotation for class incremental few-shot learning. In Int. Conf. Learn. Represent., 2023.", + "[20] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009.", + "[21] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Adv. Neural Inform. Process. Syst. 2012.", + "[22] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Trans. Pattern Anal. Mach. Intell., 40(12):2935-2947, 2017.", + "[23] Huan Liu, Li Gu, Zhixiang Chi, Yang Wang, Yuanhao Yu, Jun Chen, and Jin Tang. Few-shot class-incremental learning via entropy-regularized data-free replay. In Eur. Conf. Comput. Vis., 2022.", + "[24] Junghun Oh, Sungyong Baik, and Kyoung Mu Lee. Closer: Towards better representation learning for few-shot class-incremental learning. In Eur. Conf. Comput. Vis., 2024.", + "[25] Can Peng, Kun Zhao, Tianren Wang, Meng Li, and Brian C Lovell. Few-shot class-incremental learning from an open-set perspective. In Eur. Conf. Comput. Vis., 2022.", + "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Int. Conf. Machine Learning, 2021.", + "[27] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2017.", + "[28] David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. Experience replay for continual learning. 2019.", + "[29] Zeyin Song, Yifan Zhao, Yujun Shi, Peixi Peng, Li Yuan, and Yonghong Tian. Learning with fantasy: Semantic-aware virtual contrastive constraint for few-shot class-incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2023.", + "[30] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In IEEE Conf. Comput. Vis. Pattern Recog. 2015.", + "[31] Yu-Ming Tang, Yi-Xing Peng, Jingke Meng, and Wei-Shi Zheng. Rethinking few-shot class-incremental learning: Learning from yourself. In Eur. Conf. Comput. Vis., 2024.", + "[32] Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, and Yihong Gong. Few-shot class-incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2020.", + "[33] Oriol Vinyals, Charles Blundell, Tim Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In NeurIPS, 2016." + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15337", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011.", + "[35] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2022.", + "[36] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Marcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In Int. Conf. Machine Learning, 2022.", + "[37] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, and Ludwig Schmidt. Robust fine-tuning of zero-shot models. In IEEE Conf. Comput. Vis. Pattern Recog., 2022.", + "[38] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2019.", + "[39] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2021.", + "[40] Yibo Yang, Haobo Yuan, Xiangtai Li, Zhouchen Lin, Philip Torr, and Dacheng Tao. Neural collapse inspired feature-classifier alignment for few-shot class incremental learning. In Int. Conf. Learn. Represent., 2023.", + "[41] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Int. Conf. Comput. Vis., 2019.", + "[42] Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, and Yinghui Xu. Few-shot incremental learning with continually evolved classifiers. In IEEE Conf. Comput. Vis. Pattern Recog., 2021.", + "[43] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017.", + "[44] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu-Tao Xia. Maintaining discrimination and fairness in class incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2020.", + "[45] Linglan Zhao, Jing Lu, Yunlu Xu, Zhanzhan Cheng, Dashan Guo, Yi Niu, and Xiangzhong Fang. Few-shot class-incremental learning via class-aware bilateral distillation. In IEEE Conf. Comput. Vis. Pattern Recog., 2023.", + "[46] Da-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, Liang Ma, Shiliang Pu, and De-Chuan Zhan. Forward compatible few-shot class-incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2022.", + "[47] Huiping Zhuang, Zhenyu Weng, Run He, Zhiping Lin, and Ziqian Zeng. Gkeal: Gaussian kernel embedded analytic learning for few-shot class incremental task. In IEEE Conf. Comput. Vis. Pattern Recog., 2023." + ], + "bbox": [ + 78, + 90, + 467, + 897 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15338", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning/6894bed0-33d2-4729-8056-07cf7489b89f_model.json b/2025/Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning/6894bed0-33d2-4729-8056-07cf7489b89f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ee7cfe06feaac9110a52de2de458041f7336ed79 --- /dev/null +++ b/2025/Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning/6894bed0-33d2-4729-8056-07cf7489b89f_model.json @@ -0,0 +1,2002 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.099, + 0.131, + 0.872, + 0.154 + ], + "angle": 0, + "content": "Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.182, + 0.672, + 0.217 + ], + "angle": 0, + "content": "Juntae Lee Munawar Hayat Sungrack Yun \nQualcomm AI Research†" + }, + { + "type": "text", + "bbox": [ + 0.305, + 0.22, + 0.667, + 0.235 + ], + "angle": 0, + "content": "{juntlee, hayat, sungrack}@qti.QUALCOMM.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.284, + 0.314, + 0.301 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.316, + 0.474, + 0.725 + ], + "angle": 0, + "content": "Few-shot class incremental learning (FSCIL) enables the continual learning of new concepts with only a few training examples. In FSCIL, the model undergoes substantial updates, making it prone to forgetting previous concepts and overfitting to the limited new examples. Most recent trend is typically to disentangle the learning of the representation from the classification head of the model. A well-generalized feature extractor on the base classes (many examples and many classes) is learned, and then fixed during incremental learning. Arguing that the fixed feature extractor restricts the model's adaptability to new classes, we introduce a novel FSCIL method to effectively address catastrophic forgetting and overfitting issues. Our method enables to seamlessly update the entire model with a few examples. We mainly propose a tripartite weight-space ensemble (Tri-WE). Tri-WE interpolates the base, immediately previous, and current models in weight-space, especially for the classification heads of the models. Then, it collaboratively maintains knowledge from the base and previous models. In addition, we recognize the challenges of distilling generalized representations from the previous model from scarce data. Hence, we suggest a regularization loss term using amplified data knowledge distillation. Simply intermixing the few-shot data, we can produce richer data enabling the distillation of critical knowledge from the previous model. Consequently, we attain state-of-the-art results on the miniImageNet, CUB200, and CIFAR100 datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.751, + 0.21, + 0.767 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.776, + 0.47, + 0.868 + ], + "angle": 0, + "content": "The ability of a model to continually adapt with the availability of new data is highly desirable for many real-world applications. Few-shot class-incremental learning (FSCIL) has recently gained significant research attention as evidenced by several studies [17, 25, 29, 42, 45], because it offers a promising learning paradigm. It enables the model" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.272, + 0.9, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.46, + 0.895, + 0.572 + ], + "angle": 0, + "content": "Figure 1. Overview of the proposed method (FE: feature extractor, Cls. head: classification head). Through Tri-WE, at \\( t \\)th incremental session, Cls. head is composed by interpolating the weights of the Cls. heads covering the classes until base, \\( (t - 1) \\)th, \\( t \\)th sessions, respectively. For ADKD regularization loss, the input few-shot data are amplified to reliably mine generalized knowledge of the model of the previous \\( (t - 1) \\)th session. After incremental learning, deployed components are outlined in green." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.893, + 0.749 + ], + "angle": 0, + "content": "to flexibly expand its classification capabilities to new concepts using only a few training examples. FSCIL starts by training a model on numerous base classes with abundant labeled data. In each subsequent incremental session, the model encounters only a limited number of labeled data for new classes, along with a single prototype for each of the previously learned classes. Hence, FSCIL faces multiple challenges, including the retention of previously learned knowledge and the prevention of overfitting to the limited examples of new data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.902 + ], + "angle": 0, + "content": "To address these challenges, the most recent works are based on separating learning the representation from the classification. They rely on constructing a robust feature embedding space using plentiful base class data. In subsequent incremental learning where only a few labeled data for new classes are available, the embedding space remains fixed, and the focus is shifted to learning the classifier for the new classes. This strategy is intended to avoid accumulated model collapse or overfitting due to the limited data. For example, [25, 42] refrain from altering any base" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.468, + 0.9 + ], + "angle": 0, + "content": "†Qualcomm AI Research is an initiative of Qualcomm Technologies, Inc. and/or its subsidiaries." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15329" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.318 + ], + "angle": 0, + "content": "model parameters and use mean features for each class to define their classifiers. In contrast, [17, 29] employ ensembling multiple classifiers to improve generalization. [45] introduces an extra module to assess the reliability of outputs from the base and incrementally learned models for a weighted sum of the two outputs. Moreover, [1, 40] fixes the feature embedding space and also pre-defines classifier templates, requiring prior knowledge of the total number of incremental classes to be encountered at the outset of the base class learning. However, as indicated in Fig. 2, updating the entire model can lead to a drop in base class performance but may enhance the model's adaptability to new classes. Hence, if we address the catastrophic forgetting and overfitting more effectively, we can harness the benefits of updating the entire model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.328, + 0.473, + 0.676 + ], + "angle": 0, + "content": "To this end, we propose a novel framework for FSCIL as depicted in Fig. 1. Mainly, inspired by Wise-FT [37], we develop tripartite weight-space ensemble (Tri-WE) to continually update the weights of the model. We interpolate the classification heads of models from the base, intermediately previous, and current sessions within the weight space, thereby achieving a balance between adaptability to new classes and retention of knowledge of previous ones. This approach allows for a continual update of knowledge without an abrupt shift in the decision boundary. Moreover, to address catastrophic forgetting, knowledge distillation (KD) has been a common regularization technique in class incremental learning. However, its application in a few-shot scenario is prone to overfitting to the few available examples, and then recent cutting-edge methods have less used KD for FSCIL. To counter this, we also introduce amplified data KD (ADKD) regularization loss where we intermix few-shot training datasets, enriching the data source for distilling the knowledge from the previous model. It simply ensures that the model prevents catastrophic forgetting by transferring the generalized knowledge for previous classes. We achieve state-of-the-art (SOTA) results on three benchmarks: miniImageNet, CUB200, and CIFAR100." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.685, + 0.416, + 0.7 + ], + "angle": 0, + "content": "Our contributions are summarized as three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.71, + 0.468, + 0.771 + ], + "angle": 0, + "content": "- We introduce the Tri-WE, wherein the classification heads from base, preceding, and current models are synergistically interpolated within the weight space, promoting a knowledge continuum." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.774, + 0.468, + 0.834 + ], + "angle": 0, + "content": "- Tackling distillation of biased knowledge resulting from a scarcity of data in KD, we develop ADKD sub-loss term that enhances the generality of the distilled knowledge." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.838, + 0.468, + 0.898 + ], + "angle": 0, + "content": "- Leveraging these proposed components, we attain SOTA results across on miniImageNet, CUB200, and CIFAR100. Comprehensive experiments demonstrate the significance and efficacy of our contributions." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.71, + 0.468, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.09, + 0.882, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.253, + 0.892, + 0.308 + ], + "angle": 0, + "content": "Figure 2. Few-shot (10-way 5-shot miniImageNet) transfer learning (TL) starting from the base session model on each session. Classification accuracy scores on incremented 10 classes (top) & 40 base classes (bottom) are shown for each session." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.34, + 0.651, + 0.357 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.369, + 0.892, + 0.611 + ], + "angle": 0, + "content": "Class-incremental learning (CIL) aims to seamlessly integrate new classes without access to previously seen data, preventing the catastrophic forgetting problem. To ensure knowledge retention, rehearsal of selective previous samples [3, 28] has been developed. Also, in [35, 39], the network is dynamically expanded to accommodate new learnings. LwF [22] and iCaRL [27] used KD and nearest class mean (NCM) classifiers, respectively, to mitigate the catastrophic forgetting. UCIR [15] introduced a novel feature distillation loss, and PODNet [12] extended distillation beyond the final embedding output, incorporating the pooled outputs from the model's intermediate layers. In [2, 38, 44], the class imbalance problem is tackled. However, these methods depend on ample training data, an aspect not feasible in FSCIL, necessitating new strategies that adapt swiftly with minimal data while preserving existing knowledge." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.901 + ], + "angle": 0, + "content": "FSCIL fuses CIL and few-shot learning. It gained attention after TOPIC [32] initiated the concept, leading to numerous subsequent methods. Recent works align with two strategies. One approach completely fixes the learnable weights of the model after the base session's many-shot many-class training, and then incrementally extends the classifier by adding mean features of new classes. CEC [42] enhances the base model by simulating the FSCIL scenarios sampled from the base data again, while ALICE [25] employs diverse techniques to boost the generalization capability of the base model. Alternatively, some methods [17, 29, 40, 45] update a part of the feature extractor or the added class-mean-initialized classifiers in each session. SoftNet [18] finds the weights less impactful for the catastrophic forgetting at the base session, and focuses on updating them in the incremental learning. In WaRP [19], the important weights are decided in every session, and they keep frozen in the following sessions. GKEAL [47] updates the weight of the classification head using least square" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15330" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.287 + ], + "angle": 0, + "content": "solver in the view of analytic learning. S3C [17] employed the stochastic classifier as the extreme ensemble of classifiers. NS-FSCIL [40] and OrCo [1] assume the base session knows the amount of total incremental classes and incremental session starts with mapping each class to a classifier predefined in the base session. NS-FSCIL optimizes the last module of the feature encoder and OrCo does the last module and the selected pre-defined classifier. In SAVC [29], a class is recognized by a set of classifiers, each tailored to a specific input augmentation, resulting in a swift expansion of deployed model size during incremental learning. In BiDistill [45], every prediction is made by the ensemble of the base and last models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.473, + 0.576 + ], + "angle": 0, + "content": "Network ensemble has long been an important technique in deep learning in order to reduce overfitting and improve generalization ability [10, 13]. It involves the training of multiple neural networks and combining them for the final prediction. Various ensemble methods have been developed to get different but collaborative models such as random-weight initialization [21], different training data [30], and capturing multiple snapshots in training schedules [16]. While these methods have consistently demonstrated improved performance over single-model approaches, they often come at the cost of increased computational overhead, both in terms of training and inference. Recently, Model Soup [36] suggested interpolating the weights of a bunch of networks finetuned with different hyperparameters configurations. Also, to mitigate the overfitting of the finetuned CLIP [26], WiseFT [37] combines the weights of fine-tuned and initial CLIP models. These approaches can attain better generalization capability without increasing the computational cost of deployed models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.589, + 0.251, + 0.605 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.659 + ], + "angle": 0, + "content": "In this section, we describe the proposed Tri-WE and ADKD for FSCIL. The pipeline is illustrated in Fig. 3. We also provide the process of the base session training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.84 + ], + "angle": 0, + "content": "Problem setting. FSCIL aims to incrementally scale the model's class recognition ability by learning new classes with limited training data while preserving its knowledge of previously learned, old classes. In specific, we suppose that a model \\( f_{\\theta,\\phi} \\) consists of the feature extractor \\( g_{\\theta} \\) and the classification head (cosine classifier) \\( h_{\\phi} \\) where \\( \\theta \\) and \\( \\phi \\) denote learnable weights of them, respectively. Initially, the model is pre-trained on the base training set in the base session \\( S^{(0)} \\). Then, the model progresses a series of \\( T \\) incremental sessions, \\( \\{S^{(1)},\\ldots,S^{(T)}\\} \\). Current session \\( S^{(t)} \\) includes a class set \\( C^{(t)} \\) and the corresponding training dataset \\( D^{(t)} = \\{(x,y)\\} \\) where \\( x \\) is a sample and \\( y \\) is its label." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "In the base session \\(S^{(0)}\\), the class set \\(\\mathcal{C}^{(0)}\\) includes a substantial number of \\(N_0\\) base classes with ample training data for each class. In contrast, in the incremental session \\(S^{(t)}\\) (\\(t \\geq 1\\)), the class set \\(\\mathcal{C}^{(t)}\\) contains \\(N\\) classes with only a" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.258 + ], + "angle": 0, + "content": "small \\(K\\) number of training data available for every class \\((N\\)-way \\(K\\)-shot setting). Notice that \\(\\mathcal{C}^{(i)}\\) and \\(\\mathcal{C}^{(j)}\\) do not overlap when \\(i \\neq j, \\forall i, j\\). Following the prior FSCIL works [32, 42], the model at each session \\(\\mathcal{S}^{(t)}\\) only has access to \\(\\mathcal{D}^{(t)}\\) and an additional prototype (the averaged feature of the \\(K\\) examples per class) buffer \\(\\mathcal{M}\\) memorizing the classes of the base and previous \\((t - 1)\\) sessions. A single prototype is maintained for every class in \\(\\mathcal{M}\\). Once the incremental learning at the session \\(\\mathcal{S}^{(t)}\\) concludes, the model is evaluated on query data from all the seen \\(N_{t}\\) classes thus far, that is \\(\\mathcal{C}^{(0:t)} = \\mathcal{C}^{(0)} \\cup \\mathcal{C}^{(1)} \\cdots \\cup \\mathcal{C}^{(t)}\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.298, + 0.866, + 0.315 + ], + "angle": 0, + "content": "3.1. Tripartite weight-space ensemble (Tri-WE)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.332, + 0.892, + 0.453 + ], + "angle": 0, + "content": "The base session model is generalized by numerous classes with sufficient data, and hence its weight space is conducive to adapting to new classes. Although the weight space is somewhat compromised, the intermediate previous session's model has knowledge for all the classes encountered thus far. To capitalize on the strengths of both models, we suggest the weight-space ensemble of three models from the sessions \\( S^{(0)} \\), \\( S^{(t-1)} \\), and \\( S^{(t)} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.893, + 0.662 + ], + "angle": 0, + "content": "However, the feature extractor, which contains many layers, is prone to overfitting or collapsing when only a few examples are available. Hence, while reducing the learning rate for the feature extractor \\( g_{\\theta}^{(t)} \\), we only ensemble the classification heads \\( h_{\\phi_0} \\), \\( h_{\\phi_{\\mathrm{old}}}^{(t)} \\), \\( h_{\\phi_{\\mathrm{all}}}^{(t)} \\) in the weight space. Here, \\( h_{\\phi_0} \\) is the classification head from the base session model. \\( h_{\\phi_{\\mathrm{old}}}^{(t)} \\) is initialized by the classification head \\( h_{\\phi}^{(t - 1)} \\) from the session \\( S^{(t - 1)} \\), and then specialized to the old classes \\( \\mathcal{C}^{(0:t - 1)} \\). \\( h_{\\phi_{\\mathrm{all}}}^{(t)} \\) covers all the target classes \\( \\mathcal{C}^{(0:t)} \\), where the old class part is initialized similarly to \\( h_{\\phi_{\\mathrm{old}}}^{(t)} \\). Then, as depicted in Fig. 3, Tri-WE interpolates the classification head weight \\( \\phi^n \\) for \\( n \\)th class by" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.695, + 0.892, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\phi^ {n} = \\\\ \\left\\{ \\begin{array}{l l} \\bar {\\alpha} _ {1} \\phi_ {0} ^ {n} + \\bar {\\alpha} _ {2} \\phi_ {\\text {o l d}} ^ {n} + \\bar {\\alpha} _ {3} \\phi_ {\\text {a l l}} ^ {n} & \\text {i f} n \\leq N _ {0}, \\\\ \\bar {\\alpha} _ {4} \\phi_ {\\text {o l d}} ^ {n} + \\bar {\\alpha} _ {5} \\phi_ {\\text {a l l}} ^ {n} & \\text {i f} N _ {0} < n \\leq N _ {t - 1}, \\\\ \\phi_ {\\text {a l l}} ^ {n} & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.82, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(1 \\leq n \\leq N_t\\), \\(\\phi^n \\in \\mathbb{R}^d\\), and \\(d\\) is the dimensionality of \\(\\phi^n\\). Then, the weight of the classification head \\(h_\\phi^{(t)}\\) to classify \\(N_t\\) classes consists of \\(\\{\\phi^n\\}_{n=1}^{N_t}\\). Notice that, we also add two learnable scalar weights \\(\\alpha_1\\) and \\(\\alpha_2\\) to automatically adjust the trade-offs between the three classification" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15331" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.089, + 0.855, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.389, + 0.892, + 0.482 + ], + "angle": 0, + "content": "Figure 3. The pipeline of the proposed method. At the \\(t\\)th incremental session \\(S^{(t)}\\), the model is initialized using the resulting model of session \\(S^{(t-1)}\\). Then, a set of \\(N\\)-way \\(K\\)-shot few-shot examples \\(D^{(t)}\\) and the prototype buffer \\(\\mathcal{M}\\) are given. The weights \\(\\phi_0\\), \\(\\phi_{\\mathrm{old}}\\), \\(\\phi_{\\mathrm{all}}\\) of classification heads \\(h_{\\phi_0}\\), \\(h_{\\phi_{\\mathrm{old}}}^{(t)}\\), \\(h_{\\phi_{\\mathrm{all}}}^{(t)}\\) are interpolated with learnable scalars \\(\\alpha_1\\) and \\(\\alpha_2\\). The main loss \\(\\mathcal{L}_{\\mathrm{Cls}}\\) is computed on the resulting classification head \\(h_{\\phi}^{(t)}\\)'s output. Also, classification loss over the old classes, appeared until the session \\(S^{(t-1)}\\), is computed on the output of \\(h_{\\phi_{\\mathrm{old}}}^{(t)}\\). Additionally, \\(D^{(t)}\\) is amplified to \\(D_{\\mathrm{amp}}^{(t)}\\), and then KD losses \\(\\mathcal{L}_{\\mathrm{feat}}\\) and \\(\\mathcal{L}_{\\mathrm{logit}}\\) are computed on the amplified dataset \\(D_{\\mathrm{amp}}^{(t)}\\). The learned components are outlined in red." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.508, + 0.408, + 0.524 + ], + "angle": 0, + "content": "heads. Then, \\(\\{\\bar{\\alpha}_1,\\bar{\\alpha}_2,\\bar{\\alpha}_3,\\bar{\\alpha}_4,\\bar{\\alpha}_5\\}\\) are obtained by" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.531, + 0.359, + 0.547 + ], + "angle": 0, + "content": "\\[\n\\bar {\\alpha} _ {1} = \\alpha_ {1} / (\\alpha_ {1} + \\alpha_ {2} + 1),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.189, + 0.55, + 0.358, + 0.566 + ], + "angle": 0, + "content": "\\[\n\\bar {\\alpha} _ {2} = \\alpha_ {2} / \\left(\\alpha_ {1} + \\alpha_ {2} + 1\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.189, + 0.569, + 0.349, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\bar {\\alpha} _ {3} = 1 / \\left(\\alpha_ {1} + \\alpha_ {2} + 1\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.189, + 0.587, + 0.321, + 0.603 + ], + "angle": 0, + "content": "\\[\n\\bar {\\alpha} _ {4} = \\alpha_ {2} / (\\alpha_ {2} + 1),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.607, + 0.469, + 0.623 + ], + "angle": 0, + "content": "\\[\n\\bar {\\alpha} _ {5} = 1 / \\left(\\alpha_ {2} + 1\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.63, + 0.37, + 0.644 + ], + "angle": 0, + "content": "When \\(t = 1\\) (i.e. \\(\\phi_{\\mathrm{old}} = \\phi_0\\)), we set \\(\\alpha_{2} = 0\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.469, + 0.693 + ], + "angle": 0, + "content": "Then, the main classification loss \\(\\mathcal{L}_{\\mathrm{Cls}}\\) is computed on the output from the weight-space ensembled classification head \\(h_{\\phi}^{(t)}\\) by" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.701, + 0.469, + 0.746 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {C l s}} = \\mathbb {E} _ {(x, y) \\sim \\mathcal {D} ^ {(t)}} [ \\mathrm {C E} (h _ {\\phi} ^ {(t)} (g _ {\\theta} ^ {(t)} (x)), y ] \\\\ + \\mathbb {E} _ {(\\mathbf {p}, y) \\sim \\mathcal {M}} [ \\mathrm {C E} \\left(h _ {\\phi} ^ {(t)} (\\mathbf {p}), y \\right], \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.751, + 0.469, + 0.78 + ], + "angle": 0, + "content": "where CE is the cross-entropy loss function, and \\(\\mathbf{p}\\) denotes the old class prototype in \\(\\mathcal{M}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.782, + 0.469, + 0.817 + ], + "angle": 0, + "content": "Using the buffer \\(\\mathcal{M}\\), we also calculate the old class-specific loss \\(\\mathcal{L}_{\\mathrm{Cls - Old}}\\) for the classification head \\(h_{\\phi_{\\mathrm{old}}}^{(t)}\\) by" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.823, + 0.469, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C l s - O l d}} = \\mathbb {E} _ {(\\mathbf {p}, y) \\sim \\mathcal {M}} [ \\mathrm {C E} \\left(h _ {\\phi_ {\\text {o l d}}} ^ {(t)} (\\mathbf {p}), y\\right) ]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.852, + 0.469, + 0.903 + ], + "angle": 0, + "content": "In this loss, \\( h_{\\phi_{\\mathrm{old}}}^{(t)} \\) focuses on preserving the discriminative ability regarding all the old classes on top of the current feature extractor \\( g_{\\theta}^{(t)} \\). Hence, we do not update \\( h_{\\phi_0} \\) and instead" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.508, + 0.892, + 0.538 + ], + "angle": 0, + "content": "use it to retain the knowledge for the original, well-designed base class distribution from the base session model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.542, + 0.892, + 0.62 + ], + "angle": 0, + "content": "Unlike the naive ensemble approaches where multiple classification heads independently make predictions and combine them in the testing phase, our weight-space ensemble deploys only \\( h_{\\phi}^{(t)} \\) collaboratively utilizing the knowledge from the base and previous session models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.64, + 0.674, + 0.657 + ], + "angle": 0, + "content": "3.2. Total training loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.667, + 0.892, + 0.866 + ], + "angle": 0, + "content": "KD has been a widely used regularization technique in CIL to prevent the catastrophic forgetting problem [27]. Specifically, when new classes are introduced, an additional distillation loss is combined with the main classification loss to ensure that the current model replicates the old class behavior of the previous model. Nevertheless, in FSCIL, the limited data source can lead to the extraction of knowledge biased toward those data, rather than generalized knowledge of the old classes. To mitigate this issue, we simply generate an amplified data source \\(\\mathcal{D}_{\\mathrm{amp}}^{(t)}\\) from the few-shot training data by randomly mixing pairs of data from \\(\\mathcal{D}^{(t)}\\) employing simple data augmentation techniques such as [41, 43]. The ADKD loss is then computed as" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.886, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {A D K D}} = \\mathcal {L} _ {\\text {f e a t}} + \\mathcal {L} _ {\\text {l o g i t}}, \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15332" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{feat}}\\) and \\(\\mathcal{L}_{\\mathrm{logit}}\\) are feature-level and logit-level KD losses which are defined by" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.132, + 0.469, + 0.156 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {f e a t}} = \\mathbb {E} _ {x \\sim \\mathcal {D} _ {\\mathrm {a m p}} ^ {(t)}} | | g _ {\\theta} ^ {(t - 1)} (x) - g _ {\\theta} ^ {(t)} (x) | | _ {2} \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.159, + 0.469, + 0.182 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {l o g i t}} = \\mathbb {E} _ {x \\sim \\mathcal {D} _ {\\mathrm {a m p}} ^ {(t)}} - f _ {\\theta , \\phi} ^ {(t - 1)} (x) \\log f _ {\\theta , \\phi^ {0: N _ {t - 1}}} ^ {(t)} (x). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.194, + 0.469, + 0.232 + ], + "angle": 0, + "content": "The previous session model \\( f_{\\theta, \\phi}^{(t-1)} \\) including \\( h_{\\phi}^{(t-1)} \\) and \\( g_{\\theta}^{(t-1)} \\) is kept frozen during this process, unlike \\( h_{\\phi_{\\mathrm{old}}}^{(t)} \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.233, + 0.468, + 0.26 + ], + "angle": 0, + "content": "Finally, the total loss function used to optimize the current model is" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.274, + 0.469, + 0.29 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {C l s}} + \\gamma_ {1} \\mathcal {L} _ {\\mathrm {C l s - O l d}} + \\gamma_ {2} \\mathcal {L} _ {\\mathrm {A D K D}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.302, + 0.469, + 0.361 + ], + "angle": 0, + "content": "where \\(\\gamma_{1}\\) and \\(\\gamma_{2}\\) are hyper-parameters that balance the respective losses, \\(\\mathcal{L}_{\\mathrm{Cls - Old}}\\) and \\(\\mathcal{L}_{\\mathrm{ADKD}}\\). We provide the pseudo code of the overall process during an incremental session in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.371, + 0.274, + 0.387 + ], + "angle": 0, + "content": "3.3. Base session training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.47, + 0.726 + ], + "angle": 0, + "content": "In FSCIL, it is broadly acknowledged that a model with strong generalization to base classes is beneficial for learning new classes [25, 29, 40, 42]. To ensure fairness when assessing the proposed components for incremental learning, we strive for base session performance on par with the prior SOTA methods [29, 40, 45]. For this purpose, we employ the base session training techniques of ALICE [25] (see the literature for details). On top of it, we also added a geometric classification head in parallel with the main classification head. In specific, an input is geometrically transformed into \\( B \\) types, and we use them as the proxy labels of transformed inputs. Then, the geometric classification head is learned to classify each into one of the \\( B \\) proxy labels. This auxiliary geometric classification helps the model encode the inherent structure of data so that the model's generalization capability on new tasks is improved. This auxiliary geometric classification strategy is not used in the incremental sessions. Through this, as aftermentioned in Sec. 4, the base session model achieves comparable or superior performance to previous SOTA methods on three benchmark datasets. A more detailed description is in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.74, + 0.283, + 0.757 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Datasets. Consistent with the setting of [32], we conduct our experiments on three benchmark datasets: miniImageNet [33], CIFAR100 [20] and CUB200 [34]. miniImageNet is a subset of ImageNet [8] that includes 60,000 images from 100 classes, each with resolution \\(84 \\times 84\\). CIFAR100 comprises 60,000 tiny images of size \\(32 \\times 32\\) across 100 categories. CUB200, which is focused on a fine-grained classification, contains 200 bird species with similar appearance where the image size is \\(224 \\times 224\\). We follow" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.196 + ], + "angle": 0, + "content": "the splits from [32]. For miniImageNet and CIFAR100, 60 categories are selected as base classes while the remaining are split into 8 incremental sessions with only 5 training examples per novel class (i.e., 5-way 5-shot). As for the CUB200 dataset, 100 categories are designated as the base training set, while the rest are divided into 10-way 5-shot tasks across 10 sessions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.892, + 0.258 + ], + "angle": 0, + "content": "Evaluation protocol. In line with the FSCIL literature, we employ the top-1 accuracy (Acc) for all the observed classes at the end of each incremental learning session. We also provide the average of Acc over all the sessions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.26, + 0.893, + 0.501 + ], + "angle": 0, + "content": "Implementation details. We implement our method using the PyTorch library. In keeping with the previous works [18, 25, 29, 45], ResNet18 [14] serves as our feature extractor. The entire model is trained via the SGD optimizer. For the base session, the learning rates are initialized by 0.01, 0.001, and 0.01 for miniImageNet, CUB200, and CIFAR100 respectively, and decayed by 0.1 at 60 and 70 epochs. In each incremental session, the learning rate for the weight-space ensembled \\(\\phi_{\\mathrm{all}}\\) is 0.1, while it is 0.001 for the rest. Hence, the feature extractor is minimally updated in incremental learning. For the ADKD loss, the given \\(NK\\) training examples are amplified to \\(16NK\\) for each class. Also, the learnable scalars \\(\\alpha_{1}\\) and \\(\\alpha_{2}\\) are initially set to 1.0. Loss weights \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are empirically determined to be 1.2 and 10.0, respectively. Further details are in Sec. S-5 of the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.514, + 0.725, + 0.529 + ], + "angle": 0, + "content": "4.1. Comparative assessment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.734 + ], + "angle": 0, + "content": "On the three benchmark datasets, we first compare the proposed method with the existing approaches. As indicated in Table 1 on the miniImageNet dataset, our method outperforms the compared in all the sessions, achieving a clear SOTA in terms of the average Acc by a margin of \\(1.82\\%\\) at least. Also, BiDistill and SAVC require the base session model and a considerable number of classifiers per class, respectively, even in the testing phase. NC-FSCIL and OrCo benefit from knowing the total number of incremental classes at the outset of the base session. Whereas, we attain clear SOTA with a smaller-scale deployed model compared to BiDistill and SAVC, and without the information advantage of NC-FSCIL and OrCo." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Moreover, Fig. 4 shows the Acc curve according to sessions on the CUB200 and CIFAR100 datasets. As in the miniImageNet dataset, our method consistently surpasses the compared methods by large gaps in the incremental sessions on the CUB 200 dataset. In specific, despite lower base session Acc than SAVC, our method outperforms it in the last session. Also, while NC-FSCIL merely finetuned the last module of the feature encoder in incremental sessions, OrCo attempted to also optimize the pre-set classifiers. Nevertheless, the additional update of only the classifiers is less effective. On the CIFAR100 dataset, we can" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15333" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.09, + 0.895, + 0.132 + ], + "angle": 0, + "content": "Table 1. Comparative results on miniImageNet dataset. Session-wise Acc (%) and the average of them (Avg) are reported. In each case, the best and second results are bold-faced and underlined, respectively. Last sess. Impro. denotes the performance improvement of our method at the last session." + }, + { + "type": "table", + "bbox": [ + 0.122, + 0.142, + 0.85, + 0.419 + ], + "angle": 0, + "content": "
MethodVenueAcc in each session (%)AvgLast sess. impro.
012345678
iCaRL [27]CVPR 201761.3146.3242.9437.6330.4924.0020.8918.8017.2033.29+43.13
TOPIC [32]CVPR 202061.3150.0945.1741.1637.4835.5232.1929.4624.4239.64+35.71
ERL++ [11]AAAI 202161.7057.5854.6651.7248.6646.2744.6742.8140.7949.87+19.34
IDLVQ [4]ICLR 202064.7759.8755.9352.6249.8847.5544.8343.1441.8451.16+18.29
CEC [42]CVPR 202172.0066.8362.9759.4356.7053.7351.1949.2447.6357.74+12.50
SynthFeat [5]ICCV 202161.4059.8054.2051.6949.4548.0045.2043.8042.1050.63+18.03
MetaFSCIL [6]CVPR 202272.0467.9463.7760.2957.5855.1652.950.7949.1958.85+10.94
FACT [46]CVPR 202272.5669.6366.3862.7760.6057.3354.3452.1650.4960.70+9.64
Replay [23]ECCV 202271.8467.1263.2159.7757.0153.9551.5549.5248.2158.02+11.92
ALICE [25]ECCV 202280.6070.6067.4064.5062.5060.0057.8056.8055.7063.99+4.43
S3C [17]ECCV 202276.5571.7467.6664.5261.5158.0955.3653.0451.3462.20+8.79
WaRP [19]ICLR 202372.9968.1064.3161.3058.6456.0853.4051.7250.6559.69+9.48
SoftNet [18]ICLR 202376.6370.1365.9262.5259.4956.5653.7151.7250.4860.79+9.65
NC-FSCIL [40]ICLR 202384.0276.8072.0067.8366.3564.0461.4659.5458.3167.82+1.82
GKEAL [47]CVPR 202373.5968.9065.3362.2959.3956.7054.2052.5951.3160.48+10.40
BiDistill [45]CVPR 202374.6570.4366.2962.7760.7557.2454.7953.6552.2261.42+7.91
SAVC [29]CVPR 202381.1276.1472.4368.9266.4862.9559.9258.3957.1167.05+3.02
OrCo [1]CVPR 202483.3075.3271.5268.1665.6263.1160.2058.8258.0867.12+2.05
CLOSER [24]ECCV 202476.0271.6167.9964.6961.7058.9456.2354.5253.3362.78+6.80
Ours84.1381.4176.6573.5970.1065.1363.4261.0260.1370.62
" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.435, + 0.754, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.467, + 0.291, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.468, + 0.483, + 0.607 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.256, + 0.619, + 0.324, + 0.631 + ], + "angle": 0, + "content": "(a) CUB200" + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.467, + 0.684, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.468, + 0.875, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.643, + 0.619, + 0.722, + 0.631 + ], + "angle": 0, + "content": "(b) CIFAR100" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.643, + 0.895, + 0.687 + ], + "angle": 0, + "content": "Figure 4. Comparative results on CUB200 and CIFAR100 datasets. For each dataset, the curve of the session-wise Acc (%) is first provided. And also, the base, last, average Acc are reported on our method comparing with three recent SOTA methods, NC-FSCIL [40], SAVC [29], YourSelf [31] (Best viewed in color)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.473, + 0.894 + ], + "angle": 0, + "content": "see the similar tendency. NC-FSCIL's setting is somewhat favorable, as it allows the pre-definition of classifiers for the incremental classes with knowledge of all class numbers in advance at the base session. This advantage is particularly beneficial on the CIFAR100 dataset, where some classes are semantically similar (e.g., possum and mouse), and the low resolution \\((32 \\times 32)\\) of the images could otherwise complicate classification. The detailed results are available in Sec. S-4 of the supplementary materials. These findings demonstrate our method's superior ability to adapt to new classes with limited training data while maintaining seamless integration with the old classes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.712, + 0.67, + 0.728 + ], + "angle": 0, + "content": "4.2. Ablation analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.765 + ], + "angle": 0, + "content": "We extensively analyze the components of the proposed method on the miniImageNet or CUB200 datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Tripartite strategy. We conduct the ablation experiments on the Tri-WE in Table 2. Before this, initially, we contrast \"Naive\" with the \"No WE\" configuration, the latter incorporating the ADKD loss. The naive approach, lacking both WE and ADKD loss, exhibits a significant performance decline due to simultaneous updates of the feature extractor and classification head. However, No WE achieves respectable results thanks to the ADKD loss (more discussed in the following sections). Next, as demonstrated" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15334" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Table 2. Ablation analysis for Tri-WE. Varying the combination of classification heads ensembled in weight space (No WE, Dual-WE, Tri-WE), session-wise and average Acc (%) are provided on miniImageNet. Also, \"Naive\" means that none of WE and ADKD is used." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.129, + 0.875, + 0.248 + ], + "angle": 0, + "content": "
Classification head weightAcc in each session (%)Avg
φ0φoldφall012345678
Naive84.1380.5571.2660.9350.0544.0836.4219.7816.7351.55
No WE✓84.1378.9574.0970.0666.5663.0760.3258.0856.0167.92
Dual-WE✓✓84.1380.2575.4771.5867.6264.5261.1160.1758.9369.31
✓84.1379.5474.9970.1366.6063.4360.8559.8957.3168.54
Tri-WE✓✓✓84.1381.4176.6573.5970.165.1363.4261.0260.1370.62
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.259, + 0.893, + 0.289 + ], + "angle": 0, + "content": "Table 3. Impact on increasing the layers ensembled in weight space. Applying the proposed Tri-WE to the convolutional blocks of the feature extractor (ResNet18), session-wise and average Acc (%) are provided on miniImageNet." + }, + { + "type": "table", + "bbox": [ + 0.162, + 0.298, + 0.812, + 0.417 + ], + "angle": 0, + "content": "
Weight-space ensembledAcc in each session (%)Avg
012345678
Blocks 1-484.1375.2570.2567.5464.8561.3459.7456.3853.4565.88
Blocks 1-4 + clf. head84.1376.0872.1368.4265.8362.6060.1957.7555.2966.94
Blocks 2-4 + clf. head84.1377.8473.4269.2265.8562.7560.5657.3455.3367.38
Blocks 3-4 + clf. head84.1378.5974.4369.8066.2563.1360.4658.0455.6567.83
Block 4 + clf. head84.1378.8074.2169.6466.0163.3261.5058.1755.8867.96
Clf. head only84.1381.4176.6573.5970.165.1363.4261.0260.1370.62
" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.441, + 0.287, + 0.588 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.593, + 0.241, + 0.606 + ], + "angle": 0, + "content": "(a) Fixed \\((\\alpha_{1},\\alpha_{2})\\)" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.441, + 0.454, + 0.588 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.32, + 0.594, + 0.425, + 0.606 + ], + "angle": 0, + "content": "(b) Learned \\(\\alpha_{1},\\alpha_{2}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.618, + 0.471, + 0.675 + ], + "angle": 0, + "content": "Figure 5. Analysis on learnable scalars \\(\\alpha_{1}\\) and \\(\\alpha_{2}\\) on miniImageNet. (a) Comparison with multiple combinations of fixed (not-learnable) \\(\\alpha_{1}\\) and \\(\\alpha_{2}\\) in terms of averaged Acc over all the sessions, (b) session-wise learned \\(\\alpha_{1}\\) and \\(\\alpha_{2}\\) in ours." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.47, + 0.883 + ], + "angle": 0, + "content": "in Table 2, the absence of the weight-space ensemble for the classification head leads to increased catastrophic forgetting. And, from \"Dual-WE,\" we can see that the model from the intermediate previous session is more important than the base session model, as the reduced learning rate for the feature extractor mitigates drastic changes, helping to preserve the integrity of the feature space. Accordingly, the knowledge for incrementally-appeared classes is more helpful. Lastly, the complete Tri-WE clearly outperforms No WE by \\(2.70\\%\\) in average Acc, effectively retaining the essential knowledge of the base and previous session models." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.471, + 0.9 + ], + "angle": 0, + "content": "WE in classification head. We also evaluate the effect of" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.439, + 0.892, + 0.468 + ], + "angle": 0, + "content": "Table 4. Analysis on data amplification schemes. The last and average Acc (%) are reported on each scheme on miniImageNet." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.478, + 0.891, + 0.539 + ], + "angle": 0, + "content": "
w/o ADAD scheme
CutOut [9]RandAug [7]MixUp [43]CutMix [41]
Last Acc58.0858.1258.2659.1460.13
Avg Acc69.0569.1369.0970.0370.62
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.763 + ], + "angle": 0, + "content": "weight-space ensemble varying the number of interpolated layers in Table 3. In specific, we apply the proposed TriWE on the convolutional blocks of the feature extractors. Here, we use different learnable scalars depending on the convolutional blocks, rather than sharing them across the convolutional blocks and classification head. We observed a tendency for the decline in performance as the number of weight-space-ensemblied layers increases. This is because that deciding how to merge a lot of layers is difficult with only a few examples. Also, only merging feature extractor (1st row of Table 3) shows the lowest performance. It emphasizes that the head's weight-space is crucial to prevent catastrophic forgetting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Learnable scalars \\(\\alpha_{1} \\& \\alpha_{2}\\). Fig. 5(a) shows the averaged Acc on diverse combinations of \\((\\alpha_{1}, \\alpha_{2})\\) where both are fixed after initialization in all the sessions. Compared to all the fixed cases, the learnable scalars show better result (70.62% in the averaged Acc). Also, as Fig. 5(b), we show the learned \\(\\alpha_{1}\\) and \\(\\alpha_{2}\\) in each session in our method where we initialize both as 1.0 in every session. We can see that the learned \\(\\alpha_{1}\\) and \\(\\alpha_{2}\\) are different depending on sessions. Accordingly, our learnable scalar approach is useful for bal" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15335" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.47, + 0.118 + ], + "angle": 0, + "content": "Table 5. Analysis on loss terms of \\( {\\mathcal{L}}_{\\mathrm{{ADKD}}} \\) . The last and average Acc (%) are reported abating the loss terms on miniImageNet." + }, + { + "type": "table", + "bbox": [ + 0.179, + 0.129, + 0.371, + 0.187 + ], + "angle": 0, + "content": "
LfeatLlogitLast AccAvg Acc
✓59.4270.01
✓58.7768.41
✓✓60.1370.62
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.215, + 0.468, + 0.26 + ], + "angle": 0, + "content": "ancing the trade-off among the three classification heads. We also provide session-wise results on the fixed cases in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.47, + 0.625 + ], + "angle": 0, + "content": "Effect of data amplification. While Tri-WE already attains SOTA (i.e. w/o AD in Table 4), incorporating our ADKD loss provides further increase (69.05% vs 70.62% in average Acc). Also, to see the influence of the choice of data amplification scheme in the ADKD loss, we compare multiple data augmentation methods for our data amplification. To facilitate the use of ADKD in follow-up works, we consider four widely-used data augmentation methods, CutOut [9], RandAug [7], CutMix [41], MixUp [43]. In the first two, the images are not inter-mixed. Whereas, two images are mixed for data augmentation. In Table 4, we can see that CutOut and RandAug are not effective compared to \"w/o AD\" where no data amplification is considered. With only a few examples, those methods cannot yield a diversity of data. Next, in comparison with MixUp, CutMix shows a better result. MixUp mixes two images smoothly, and then the original semantics are somewhat preserved. This is less useful in KD. Namely, CutMix encourages the model to focus more on general image recognition ability rather than high-level semantics, which is more helpful in preventing the catastrophic forgetting problem in FSCIL. Hence, we employ CutMix as the default option. In addition, the impact of the degree of data amplification is addressed in Sec. S-1 of the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.468, + 0.716 + ], + "angle": 0, + "content": "Two loss terms of \\(\\mathcal{L}_{\\mathrm{ADKD}}\\). As shown in the first two rows of Table 5, we ablate \\(\\mathcal{L}_{\\mathrm{feat}}\\) and \\(\\mathcal{L}_{\\mathrm{logit}}\\), respectively. From this study, we can see that \\(\\mathcal{L}_{\\mathrm{feat}}\\) is slightly more important than \\(\\mathcal{L}_{\\mathrm{logit}}\\) since Tri-WE already helps to prevent catastrophic forgetting in the side of the classification head. Then, using both is most effective." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.468, + 0.854 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{\\mathrm{Cls - Old}}\\). As we update the feature extractor as well as the classification head in the incremental learning, we also adapt \\(\\phi_{\\mathrm{old}}\\) to the updated feature extractor using the loss \\(\\mathcal{L}_{\\mathrm{Cls - Old}}\\). To verify this, we conduct ablating the feature extractor update and \\(\\mathcal{L}_{\\mathrm{Cls - Old}}\\). As shown in Table 6, the performance is degraded without this loss term when the feature extractor is updated. Also, when the feature extractor is frozen, updating \\(\\phi_{\\mathrm{old}}\\) is less meaningful. Finally, updating the feature extractor with \\(\\phi_{\\mathrm{old}}\\) is most beneficial." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "New class adaptability. To see the efficacy of our method in terms of the trade-off between base and newly introduced classes, we provide the individual accuracy of the" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 6. Analysis on \\( {\\mathcal{L}}_{\\text{Cls-Old }} \\) and updating the feature extractor \\( {g}_{\\theta } \\) . The last and average Acc (%) are reported,on miniImageNet." + }, + { + "type": "table", + "bbox": [ + 0.589, + 0.129, + 0.803, + 0.197 + ], + "angle": 0, + "content": "
LCls-OldUpdate gθLast AccAvg Acc
✓59.2770.15
✓✓60.1370.62
58.8669.41
✓58.8869.40
" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.223, + 0.89, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.382, + 0.892, + 0.438 + ], + "angle": 0, + "content": "Figure 6. Analysis on the new class adaptability on CUB200. The average of accuracy scores are provided for base and novel classes, denoted by Base and New respectively. And, harmonic means are also compared." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.447, + 0.892, + 0.553 + ], + "angle": 0, + "content": "base and novel classes, and the harmonic mean (HM). Since most previous works have addressed the joint accuracy of the base and novel classes, we can compare our method with only a few recent methods [17, 33, 42, 45, 46] on the CUB200 dataset, following [17, 46]. Fig. 6 shows that our approach is more adept at integrating new classes without compromising the performance on the base classes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.567, + 0.627, + 0.582 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.592, + 0.892, + 0.833 + ], + "angle": 0, + "content": "We explored FSCIL which facilitates the continuous acquisition of new concepts from a few examples. To this end, we suggested a novel FSCIL method including Tri-WE and ADKD regularization loss. The proposed method can effectively prevent overfitting to the few data and catastrophic forgetting of old concepts in the model. Tri-WE strategically interpolates the weights of the classification heads from the base, immediate previous, and current models, which collectively preserves the knowledge from the earlier models without overwhelming updates. Enriching the few examples for KD, the ADKD loss facilitates the transfer of essential knowledge from the previous model, rather than the knowledge overfitted to the few examples. We conducted extensive analyses on the proposed components and achieved state-of-the-art performance on the miniImageNet, CUB200, and CIFAR100 datasets." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.848, + 0.597, + 0.864 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[1] Noor Ahmed, Anna Kukleva, and Bernt Schiele. Orco: Towards better generalization via orthogonality and contrast for" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15336" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "few-shot class-incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.123, + 0.468, + 0.164 + ], + "angle": 0, + "content": "[2] Eden Belouadah and Adrian Popescu. II2m: Class incremental learning with dual memory. In IEEE Conf. Comput. Vis. Pattern Recog., 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.166, + 0.468, + 0.208 + ], + "angle": 0, + "content": "[3] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. arXiv preprint arXiv:1812.00420, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.21, + 0.468, + 0.251 + ], + "angle": 0, + "content": "[4] Kuilin Chen and Chi-Guhn Lee. Incremental few-shot learning via vector quantization in deep embedded space. In Int. Conf. Learn. Represent., 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.254, + 0.468, + 0.322 + ], + "angle": 0, + "content": "[5] Ali Cheraghian, Shafin Rahman, Sameera Ramasinghe, Pengfei Fang, Christian Simon, Lars Petersson, and Mehrtash Harandi. Synthesized feature based few-shot class-incremental learning on a mixture of subspaces. In Int. Conf. Comput. Vis., 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.325, + 0.468, + 0.379 + ], + "angle": 0, + "content": "[6] Zhixiang Chi, Li Gu, Huan Liu, Yang Wang, Yuanhao Yu, and Jin Tang. Meta-scil: A meta-learning approach for few-shot class incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.382, + 0.468, + 0.436 + ], + "angle": 0, + "content": "[7] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In IEEE Conf. Comput. Vis. Pattern Recog. Worksh., 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.44, + 0.468, + 0.481 + ], + "angle": 0, + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conf. Comput. Vis. Pattern Recog., 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.483, + 0.468, + 0.525 + ], + "angle": 0, + "content": "[9] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.527, + 0.468, + 0.567 + ], + "angle": 0, + "content": "[10] Thomas G Dietterich. Ensemble methods in machine learning. In International Workshop on Multiple Classifier Systems, 2000." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.571, + 0.468, + 0.624 + ], + "angle": 0, + "content": "[11] Songlin Dong, Xiaopeng Hong, Xiaoyu Tao, Xinyuan Chang, Xing Wei, and Yihong Gong. Few-shot class incremental learning via relation knowledge distillation. In AAAI, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.628, + 0.468, + 0.682 + ], + "angle": 0, + "content": "[12] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In Eur. Conf. Comput. Vis., 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.468, + 0.725 + ], + "angle": 0, + "content": "[13] Lars Kai Hansen and Peter Salamon. Neural network ensembles. IEEE Trans. Pattern Anal. Mach. Intell., 12(10): 993-1001, 1990." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.729, + 0.468, + 0.77 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.468, + 0.814 + ], + "angle": 0, + "content": "[15] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In IEEE Conf. Comput. Vis. Pattern Recog., 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.468, + 0.857 + ], + "angle": 0, + "content": "[16] Gao Huang, Yixuan Li, Geoff Pleiss, Zhuang Liu, John E Hopcroft, and Kilian Q Weinberger. Snapshot ensembles: Train 1, get m for free. In Int. Conf. Learn. Represent. 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[17] Jayateja Kalla and Soma Biswas. S3c: Self-supervised stochastic classifiers for few-shot class-incremental learning. In Eur. Conf. Comput. Vis., 2022." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[18] Haeyong Kang, Jaehong Yoon, Sultan Rizky Hikmawan Madjid, Sung Ju Hwang, and Chang D Yoo. On the soft-subnetwork for few-shot class incremental learning. In Int. Conf. Learn. Represent., 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[19] Do-Yeon Kim, Dong-Jun Han, Jun Seo, and Jaekyun Moon. Warping the space: Weight space rotation for class incremental few-shot learning. In Int. Conf. Learn. Represent., 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[20] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[21] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Adv. Neural Inform. Process. Syst. 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.279, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[22] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Trans. Pattern Anal. Mach. Intell., 40(12):2935-2947, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[23] Huan Liu, Li Gu, Zhixiang Chi, Yang Wang, Yuanhao Yu, Jun Chen, and Jin Tang. Few-shot class-incremental learning via entropy-regularized data-free replay. In Eur. Conf. Comput. Vis., 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.378, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[24] Junghun Oh, Sungyong Baik, and Kyoung Mu Lee. Closer: Towards better representation learning for few-shot class-incremental learning. In Eur. Conf. Comput. Vis., 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.421, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[25] Can Peng, Kun Zhao, Tianren Wang, Meng Li, and Brian C Lovell. Few-shot class-incremental learning from an open-set perspective. In Eur. Conf. Comput. Vis., 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.464, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Int. Conf. Machine Learning, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.534, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[27] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[28] David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. Experience replay for continual learning. 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.634, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[29] Zeyin Song, Yifan Zhao, Yujun Shi, Peixi Peng, Li Yuan, and Yonghong Tian. Learning with fantasy: Semantic-aware virtual contrastive constraint for few-shot class-incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.69, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[30] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In IEEE Conf. Comput. Vis. Pattern Recog. 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[31] Yu-Ming Tang, Yi-Xing Peng, Jingke Meng, and Wei-Shi Zheng. Rethinking few-shot class-incremental learning: Learning from yourself. In Eur. Conf. Comput. Vis., 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[32] Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, and Yihong Gong. Few-shot class-incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[33] Oriol Vinyals, Charles Blundell, Tim Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In NeurIPS, 2016." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15337" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.133 + ], + "angle": 0, + "content": "[34] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[35] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.468, + 0.275 + ], + "angle": 0, + "content": "[36] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Marcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In Int. Conf. Machine Learning, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.277, + 0.468, + 0.358 + ], + "angle": 0, + "content": "[37] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, and Ludwig Schmidt. Robust fine-tuning of zero-shot models. In IEEE Conf. Comput. Vis. Pattern Recog., 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.361, + 0.468, + 0.414 + ], + "angle": 0, + "content": "[38] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.418, + 0.468, + 0.459 + ], + "angle": 0, + "content": "[39] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.461, + 0.468, + 0.514 + ], + "angle": 0, + "content": "[40] Yibo Yang, Haobo Yuan, Xiangtai Li, Zhouchen Lin, Philip Torr, and Dacheng Tao. Neural collapse inspired feature-classifier alignment for few-shot class incremental learning. In Int. Conf. Learn. Represent., 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.517, + 0.468, + 0.572 + ], + "angle": 0, + "content": "[41] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Int. Conf. Comput. Vis., 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.574, + 0.468, + 0.627 + ], + "angle": 0, + "content": "[42] Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, and Yinghui Xu. Few-shot incremental learning with continually evolved classifiers. In IEEE Conf. Comput. Vis. Pattern Recog., 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.63, + 0.468, + 0.671 + ], + "angle": 0, + "content": "[43] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.673, + 0.468, + 0.727 + ], + "angle": 0, + "content": "[44] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu-Tao Xia. Maintaining discrimination and fairness in class incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.73, + 0.468, + 0.784 + ], + "angle": 0, + "content": "[45] Linglan Zhao, Jing Lu, Yunlu Xu, Zhanzhan Cheng, Dashan Guo, Yi Niu, and Xiangzhong Fang. Few-shot class-incremental learning via class-aware bilateral distillation. In IEEE Conf. Comput. Vis. Pattern Recog., 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.786, + 0.468, + 0.84 + ], + "angle": 0, + "content": "[46] Da-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, Liang Ma, Shiliang Pu, and De-Chuan Zhan. Forward compatible few-shot class-incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.843, + 0.468, + 0.898 + ], + "angle": 0, + "content": "[47] Huiping Zhuang, Zhenyu Weng, Run He, Zhiping Lin, and Ziqian Zeng. Gkeal: Gaussian kernel embedded analytic learning for few-shot class incremental task. In IEEE Conf. Comput. Vis. Pattern Recog., 2023." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "15338" + } + ] +] \ No newline at end of file diff --git a/2025/Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning/6894bed0-33d2-4729-8056-07cf7489b89f_origin.pdf b/2025/Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning/6894bed0-33d2-4729-8056-07cf7489b89f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f00f2fddec450dadf223cb20f14e4db970d42d30 --- /dev/null +++ b/2025/Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning/6894bed0-33d2-4729-8056-07cf7489b89f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fcce4ac885a140397dbd37131687f34ceb66ede3bacb73cb00f0c3d44d08b8ed +size 3227951 diff --git a/2025/Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning/full.md b/2025/Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6ec3c0d62a16e5b670edcf589c09d5dede7852f0 --- /dev/null +++ b/2025/Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning/full.md @@ -0,0 +1,294 @@ +# Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning + +Juntae Lee Munawar Hayat Sungrack Yun +Qualcomm AI Research† + +{juntlee, hayat, sungrack}@qti.QUALCOMM.com + +# Abstract + +Few-shot class incremental learning (FSCIL) enables the continual learning of new concepts with only a few training examples. In FSCIL, the model undergoes substantial updates, making it prone to forgetting previous concepts and overfitting to the limited new examples. Most recent trend is typically to disentangle the learning of the representation from the classification head of the model. A well-generalized feature extractor on the base classes (many examples and many classes) is learned, and then fixed during incremental learning. Arguing that the fixed feature extractor restricts the model's adaptability to new classes, we introduce a novel FSCIL method to effectively address catastrophic forgetting and overfitting issues. Our method enables to seamlessly update the entire model with a few examples. We mainly propose a tripartite weight-space ensemble (Tri-WE). Tri-WE interpolates the base, immediately previous, and current models in weight-space, especially for the classification heads of the models. Then, it collaboratively maintains knowledge from the base and previous models. In addition, we recognize the challenges of distilling generalized representations from the previous model from scarce data. Hence, we suggest a regularization loss term using amplified data knowledge distillation. Simply intermixing the few-shot data, we can produce richer data enabling the distillation of critical knowledge from the previous model. Consequently, we attain state-of-the-art results on the miniImageNet, CUB200, and CIFAR100 datasets. + +# 1. Introduction + +The ability of a model to continually adapt with the availability of new data is highly desirable for many real-world applications. Few-shot class-incremental learning (FSCIL) has recently gained significant research attention as evidenced by several studies [17, 25, 29, 42, 45], because it offers a promising learning paradigm. It enables the model + +![](images/566cc08011a0b6e754bd42e568a4dbf5b71b71a8b6a794a98be3ecbb0c0619be.jpg) +Figure 1. Overview of the proposed method (FE: feature extractor, Cls. head: classification head). Through Tri-WE, at $t$ th incremental session, Cls. head is composed by interpolating the weights of the Cls. heads covering the classes until base, $(t - 1)$ th, $t$ th sessions, respectively. For ADKD regularization loss, the input few-shot data are amplified to reliably mine generalized knowledge of the model of the previous $(t - 1)$ th session. After incremental learning, deployed components are outlined in green. + +to flexibly expand its classification capabilities to new concepts using only a few training examples. FSCIL starts by training a model on numerous base classes with abundant labeled data. In each subsequent incremental session, the model encounters only a limited number of labeled data for new classes, along with a single prototype for each of the previously learned classes. Hence, FSCIL faces multiple challenges, including the retention of previously learned knowledge and the prevention of overfitting to the limited examples of new data. + +To address these challenges, the most recent works are based on separating learning the representation from the classification. They rely on constructing a robust feature embedding space using plentiful base class data. In subsequent incremental learning where only a few labeled data for new classes are available, the embedding space remains fixed, and the focus is shifted to learning the classifier for the new classes. This strategy is intended to avoid accumulated model collapse or overfitting due to the limited data. For example, [25, 42] refrain from altering any base + +model parameters and use mean features for each class to define their classifiers. In contrast, [17, 29] employ ensembling multiple classifiers to improve generalization. [45] introduces an extra module to assess the reliability of outputs from the base and incrementally learned models for a weighted sum of the two outputs. Moreover, [1, 40] fixes the feature embedding space and also pre-defines classifier templates, requiring prior knowledge of the total number of incremental classes to be encountered at the outset of the base class learning. However, as indicated in Fig. 2, updating the entire model can lead to a drop in base class performance but may enhance the model's adaptability to new classes. Hence, if we address the catastrophic forgetting and overfitting more effectively, we can harness the benefits of updating the entire model. + +To this end, we propose a novel framework for FSCIL as depicted in Fig. 1. Mainly, inspired by Wise-FT [37], we develop tripartite weight-space ensemble (Tri-WE) to continually update the weights of the model. We interpolate the classification heads of models from the base, intermediately previous, and current sessions within the weight space, thereby achieving a balance between adaptability to new classes and retention of knowledge of previous ones. This approach allows for a continual update of knowledge without an abrupt shift in the decision boundary. Moreover, to address catastrophic forgetting, knowledge distillation (KD) has been a common regularization technique in class incremental learning. However, its application in a few-shot scenario is prone to overfitting to the few available examples, and then recent cutting-edge methods have less used KD for FSCIL. To counter this, we also introduce amplified data KD (ADKD) regularization loss where we intermix few-shot training datasets, enriching the data source for distilling the knowledge from the previous model. It simply ensures that the model prevents catastrophic forgetting by transferring the generalized knowledge for previous classes. We achieve state-of-the-art (SOTA) results on three benchmarks: miniImageNet, CUB200, and CIFAR100. + +Our contributions are summarized as three-fold: + +- We introduce the Tri-WE, wherein the classification heads from base, preceding, and current models are synergistically interpolated within the weight space, promoting a knowledge continuum. +- Tackling distillation of biased knowledge resulting from a scarcity of data in KD, we develop ADKD sub-loss term that enhances the generality of the distilled knowledge. +- Leveraging these proposed components, we attain SOTA results across on miniImageNet, CUB200, and CIFAR100. Comprehensive experiments demonstrate the significance and efficacy of our contributions. + +![](images/bf9ef0d970d4412fc36a9a386982c91fe9da3e3460b23c92dba24c4266056228.jpg) +Figure 2. Few-shot (10-way 5-shot miniImageNet) transfer learning (TL) starting from the base session model on each session. Classification accuracy scores on incremented 10 classes (top) & 40 base classes (bottom) are shown for each session. + +# 2. Related Works + +Class-incremental learning (CIL) aims to seamlessly integrate new classes without access to previously seen data, preventing the catastrophic forgetting problem. To ensure knowledge retention, rehearsal of selective previous samples [3, 28] has been developed. Also, in [35, 39], the network is dynamically expanded to accommodate new learnings. LwF [22] and iCaRL [27] used KD and nearest class mean (NCM) classifiers, respectively, to mitigate the catastrophic forgetting. UCIR [15] introduced a novel feature distillation loss, and PODNet [12] extended distillation beyond the final embedding output, incorporating the pooled outputs from the model's intermediate layers. In [2, 38, 44], the class imbalance problem is tackled. However, these methods depend on ample training data, an aspect not feasible in FSCIL, necessitating new strategies that adapt swiftly with minimal data while preserving existing knowledge. + +FSCIL fuses CIL and few-shot learning. It gained attention after TOPIC [32] initiated the concept, leading to numerous subsequent methods. Recent works align with two strategies. One approach completely fixes the learnable weights of the model after the base session's many-shot many-class training, and then incrementally extends the classifier by adding mean features of new classes. CEC [42] enhances the base model by simulating the FSCIL scenarios sampled from the base data again, while ALICE [25] employs diverse techniques to boost the generalization capability of the base model. Alternatively, some methods [17, 29, 40, 45] update a part of the feature extractor or the added class-mean-initialized classifiers in each session. SoftNet [18] finds the weights less impactful for the catastrophic forgetting at the base session, and focuses on updating them in the incremental learning. In WaRP [19], the important weights are decided in every session, and they keep frozen in the following sessions. GKEAL [47] updates the weight of the classification head using least square + +solver in the view of analytic learning. S3C [17] employed the stochastic classifier as the extreme ensemble of classifiers. NS-FSCIL [40] and OrCo [1] assume the base session knows the amount of total incremental classes and incremental session starts with mapping each class to a classifier predefined in the base session. NS-FSCIL optimizes the last module of the feature encoder and OrCo does the last module and the selected pre-defined classifier. In SAVC [29], a class is recognized by a set of classifiers, each tailored to a specific input augmentation, resulting in a swift expansion of deployed model size during incremental learning. In BiDistill [45], every prediction is made by the ensemble of the base and last models. + +Network ensemble has long been an important technique in deep learning in order to reduce overfitting and improve generalization ability [10, 13]. It involves the training of multiple neural networks and combining them for the final prediction. Various ensemble methods have been developed to get different but collaborative models such as random-weight initialization [21], different training data [30], and capturing multiple snapshots in training schedules [16]. While these methods have consistently demonstrated improved performance over single-model approaches, they often come at the cost of increased computational overhead, both in terms of training and inference. Recently, Model Soup [36] suggested interpolating the weights of a bunch of networks finetuned with different hyperparameters configurations. Also, to mitigate the overfitting of the finetuned CLIP [26], WiseFT [37] combines the weights of fine-tuned and initial CLIP models. These approaches can attain better generalization capability without increasing the computational cost of deployed models. + +# 3. Proposed Method + +In this section, we describe the proposed Tri-WE and ADKD for FSCIL. The pipeline is illustrated in Fig. 3. We also provide the process of the base session training. + +Problem setting. FSCIL aims to incrementally scale the model's class recognition ability by learning new classes with limited training data while preserving its knowledge of previously learned, old classes. In specific, we suppose that a model $f_{\theta,\phi}$ consists of the feature extractor $g_{\theta}$ and the classification head (cosine classifier) $h_{\phi}$ where $\theta$ and $\phi$ denote learnable weights of them, respectively. Initially, the model is pre-trained on the base training set in the base session $S^{(0)}$ . Then, the model progresses a series of $T$ incremental sessions, $\{S^{(1)},\ldots,S^{(T)}\}$ . Current session $S^{(t)}$ includes a class set $C^{(t)}$ and the corresponding training dataset $D^{(t)} = \{(x,y)\}$ where $x$ is a sample and $y$ is its label. + +In the base session $S^{(0)}$ , the class set $\mathcal{C}^{(0)}$ includes a substantial number of $N_0$ base classes with ample training data for each class. In contrast, in the incremental session $S^{(t)}$ ( $t \geq 1$ ), the class set $\mathcal{C}^{(t)}$ contains $N$ classes with only a + +small $K$ number of training data available for every class $(N$ -way $K$ -shot setting). Notice that $\mathcal{C}^{(i)}$ and $\mathcal{C}^{(j)}$ do not overlap when $i \neq j, \forall i, j$ . Following the prior FSCIL works [32, 42], the model at each session $\mathcal{S}^{(t)}$ only has access to $\mathcal{D}^{(t)}$ and an additional prototype (the averaged feature of the $K$ examples per class) buffer $\mathcal{M}$ memorizing the classes of the base and previous $(t - 1)$ sessions. A single prototype is maintained for every class in $\mathcal{M}$ . Once the incremental learning at the session $\mathcal{S}^{(t)}$ concludes, the model is evaluated on query data from all the seen $N_{t}$ classes thus far, that is $\mathcal{C}^{(0:t)} = \mathcal{C}^{(0)} \cup \mathcal{C}^{(1)} \cdots \cup \mathcal{C}^{(t)}$ . + +# 3.1. Tripartite weight-space ensemble (Tri-WE) + +The base session model is generalized by numerous classes with sufficient data, and hence its weight space is conducive to adapting to new classes. Although the weight space is somewhat compromised, the intermediate previous session's model has knowledge for all the classes encountered thus far. To capitalize on the strengths of both models, we suggest the weight-space ensemble of three models from the sessions $S^{(0)}$ , $S^{(t-1)}$ , and $S^{(t)}$ . + +However, the feature extractor, which contains many layers, is prone to overfitting or collapsing when only a few examples are available. Hence, while reducing the learning rate for the feature extractor $g_{\theta}^{(t)}$ , we only ensemble the classification heads $h_{\phi_0}$ , $h_{\phi_{\mathrm{old}}}^{(t)}$ , $h_{\phi_{\mathrm{all}}}^{(t)}$ in the weight space. Here, $h_{\phi_0}$ is the classification head from the base session model. $h_{\phi_{\mathrm{old}}}^{(t)}$ is initialized by the classification head $h_{\phi}^{(t - 1)}$ from the session $S^{(t - 1)}$ , and then specialized to the old classes $\mathcal{C}^{(0:t - 1)}$ . $h_{\phi_{\mathrm{all}}}^{(t)}$ covers all the target classes $\mathcal{C}^{(0:t)}$ , where the old class part is initialized similarly to $h_{\phi_{\mathrm{old}}}^{(t)}$ . Then, as depicted in Fig. 3, Tri-WE interpolates the classification head weight $\phi^n$ for $n$ th class by + +$$ +\begin{array}{l} \phi^ {n} = \\ \left\{ \begin{array}{l l} \bar {\alpha} _ {1} \phi_ {0} ^ {n} + \bar {\alpha} _ {2} \phi_ {\text {o l d}} ^ {n} + \bar {\alpha} _ {3} \phi_ {\text {a l l}} ^ {n} & \text {i f} n \leq N _ {0}, \\ \bar {\alpha} _ {4} \phi_ {\text {o l d}} ^ {n} + \bar {\alpha} _ {5} \phi_ {\text {a l l}} ^ {n} & \text {i f} N _ {0} < n \leq N _ {t - 1}, \\ \phi_ {\text {a l l}} ^ {n} & \text {o t h e r w i s e}, \end{array} \right. \tag {1} \\ \end{array} +$$ + +where $1 \leq n \leq N_t$ , $\phi^n \in \mathbb{R}^d$ , and $d$ is the dimensionality of $\phi^n$ . Then, the weight of the classification head $h_\phi^{(t)}$ to classify $N_t$ classes consists of $\{\phi^n\}_{n=1}^{N_t}$ . Notice that, we also add two learnable scalar weights $\alpha_1$ and $\alpha_2$ to automatically adjust the trade-offs between the three classification + +![](images/b5f2cc96519b842abc8f0f82a8207496d3862708e3cb9799d03e4056e431e080.jpg) +Figure 3. The pipeline of the proposed method. At the $t$ th incremental session $S^{(t)}$ , the model is initialized using the resulting model of session $S^{(t-1)}$ . Then, a set of $N$ -way $K$ -shot few-shot examples $D^{(t)}$ and the prototype buffer $\mathcal{M}$ are given. The weights $\phi_0$ , $\phi_{\mathrm{old}}$ , $\phi_{\mathrm{all}}$ of classification heads $h_{\phi_0}$ , $h_{\phi_{\mathrm{old}}}^{(t)}$ , $h_{\phi_{\mathrm{all}}}^{(t)}$ are interpolated with learnable scalars $\alpha_1$ and $\alpha_2$ . The main loss $\mathcal{L}_{\mathrm{Cls}}$ is computed on the resulting classification head $h_{\phi}^{(t)}$ 's output. Also, classification loss over the old classes, appeared until the session $S^{(t-1)}$ , is computed on the output of $h_{\phi_{\mathrm{old}}}^{(t)}$ . Additionally, $D^{(t)}$ is amplified to $D_{\mathrm{amp}}^{(t)}$ , and then KD losses $\mathcal{L}_{\mathrm{feat}}$ and $\mathcal{L}_{\mathrm{logit}}$ are computed on the amplified dataset $D_{\mathrm{amp}}^{(t)}$ . The learned components are outlined in red. + +heads. Then, $\{\bar{\alpha}_1,\bar{\alpha}_2,\bar{\alpha}_3,\bar{\alpha}_4,\bar{\alpha}_5\}$ are obtained by + +$$ +\bar {\alpha} _ {1} = \alpha_ {1} / (\alpha_ {1} + \alpha_ {2} + 1), +$$ + +$$ +\bar {\alpha} _ {2} = \alpha_ {2} / \left(\alpha_ {1} + \alpha_ {2} + 1\right), +$$ + +$$ +\bar {\alpha} _ {3} = 1 / \left(\alpha_ {1} + \alpha_ {2} + 1\right), +$$ + +$$ +\bar {\alpha} _ {4} = \alpha_ {2} / (\alpha_ {2} + 1), +$$ + +$$ +\bar {\alpha} _ {5} = 1 / \left(\alpha_ {2} + 1\right). \tag {2} +$$ + +When $t = 1$ (i.e. $\phi_{\mathrm{old}} = \phi_0$ ), we set $\alpha_{2} = 0$ . + +Then, the main classification loss $\mathcal{L}_{\mathrm{Cls}}$ is computed on the output from the weight-space ensembled classification head $h_{\phi}^{(t)}$ by + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {C l s}} = \mathbb {E} _ {(x, y) \sim \mathcal {D} ^ {(t)}} [ \mathrm {C E} (h _ {\phi} ^ {(t)} (g _ {\theta} ^ {(t)} (x)), y ] \\ + \mathbb {E} _ {(\mathbf {p}, y) \sim \mathcal {M}} [ \mathrm {C E} \left(h _ {\phi} ^ {(t)} (\mathbf {p}), y \right], \tag {3} \\ \end{array} +$$ + +where CE is the cross-entropy loss function, and $\mathbf{p}$ denotes the old class prototype in $\mathcal{M}$ . + +Using the buffer $\mathcal{M}$ , we also calculate the old class-specific loss $\mathcal{L}_{\mathrm{Cls - Old}}$ for the classification head $h_{\phi_{\mathrm{old}}}^{(t)}$ by + +$$ +\mathcal {L} _ {\mathrm {C l s - O l d}} = \mathbb {E} _ {(\mathbf {p}, y) \sim \mathcal {M}} [ \mathrm {C E} \left(h _ {\phi_ {\text {o l d}}} ^ {(t)} (\mathbf {p}), y\right) ]. \tag {4} +$$ + +In this loss, $h_{\phi_{\mathrm{old}}}^{(t)}$ focuses on preserving the discriminative ability regarding all the old classes on top of the current feature extractor $g_{\theta}^{(t)}$ . Hence, we do not update $h_{\phi_0}$ and instead + +use it to retain the knowledge for the original, well-designed base class distribution from the base session model. + +Unlike the naive ensemble approaches where multiple classification heads independently make predictions and combine them in the testing phase, our weight-space ensemble deploys only $h_{\phi}^{(t)}$ collaboratively utilizing the knowledge from the base and previous session models. + +# 3.2. Total training loss + +KD has been a widely used regularization technique in CIL to prevent the catastrophic forgetting problem [27]. Specifically, when new classes are introduced, an additional distillation loss is combined with the main classification loss to ensure that the current model replicates the old class behavior of the previous model. Nevertheless, in FSCIL, the limited data source can lead to the extraction of knowledge biased toward those data, rather than generalized knowledge of the old classes. To mitigate this issue, we simply generate an amplified data source $\mathcal{D}_{\mathrm{amp}}^{(t)}$ from the few-shot training data by randomly mixing pairs of data from $\mathcal{D}^{(t)}$ employing simple data augmentation techniques such as [41, 43]. The ADKD loss is then computed as + +$$ +\mathcal {L} _ {\mathrm {A D K D}} = \mathcal {L} _ {\text {f e a t}} + \mathcal {L} _ {\text {l o g i t}}, \tag {5} +$$ + +where $\mathcal{L}_{\mathrm{feat}}$ and $\mathcal{L}_{\mathrm{logit}}$ are feature-level and logit-level KD losses which are defined by + +$$ +\mathcal {L} _ {\text {f e a t}} = \mathbb {E} _ {x \sim \mathcal {D} _ {\mathrm {a m p}} ^ {(t)}} | | g _ {\theta} ^ {(t - 1)} (x) - g _ {\theta} ^ {(t)} (x) | | _ {2} \tag {6} +$$ + +$$ +\mathcal {L} _ {\text {l o g i t}} = \mathbb {E} _ {x \sim \mathcal {D} _ {\mathrm {a m p}} ^ {(t)}} - f _ {\theta , \phi} ^ {(t - 1)} (x) \log f _ {\theta , \phi^ {0: N _ {t - 1}}} ^ {(t)} (x). \tag {7} +$$ + +The previous session model $f_{\theta, \phi}^{(t-1)}$ including $h_{\phi}^{(t-1)}$ and $g_{\theta}^{(t-1)}$ is kept frozen during this process, unlike $h_{\phi_{\mathrm{old}}}^{(t)}$ . + +Finally, the total loss function used to optimize the current model is + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {C l s}} + \gamma_ {1} \mathcal {L} _ {\mathrm {C l s - O l d}} + \gamma_ {2} \mathcal {L} _ {\mathrm {A D K D}}, \tag {8} +$$ + +where $\gamma_{1}$ and $\gamma_{2}$ are hyper-parameters that balance the respective losses, $\mathcal{L}_{\mathrm{Cls - Old}}$ and $\mathcal{L}_{\mathrm{ADKD}}$ . We provide the pseudo code of the overall process during an incremental session in the supplementary materials. + +# 3.3. Base session training + +In FSCIL, it is broadly acknowledged that a model with strong generalization to base classes is beneficial for learning new classes [25, 29, 40, 42]. To ensure fairness when assessing the proposed components for incremental learning, we strive for base session performance on par with the prior SOTA methods [29, 40, 45]. For this purpose, we employ the base session training techniques of ALICE [25] (see the literature for details). On top of it, we also added a geometric classification head in parallel with the main classification head. In specific, an input is geometrically transformed into $B$ types, and we use them as the proxy labels of transformed inputs. Then, the geometric classification head is learned to classify each into one of the $B$ proxy labels. This auxiliary geometric classification helps the model encode the inherent structure of data so that the model's generalization capability on new tasks is improved. This auxiliary geometric classification strategy is not used in the incremental sessions. Through this, as aftermentioned in Sec. 4, the base session model achieves comparable or superior performance to previous SOTA methods on three benchmark datasets. A more detailed description is in the supplementary materials. + +# 4. Experimental Results + +Datasets. Consistent with the setting of [32], we conduct our experiments on three benchmark datasets: miniImageNet [33], CIFAR100 [20] and CUB200 [34]. miniImageNet is a subset of ImageNet [8] that includes 60,000 images from 100 classes, each with resolution $84 \times 84$ . CIFAR100 comprises 60,000 tiny images of size $32 \times 32$ across 100 categories. CUB200, which is focused on a fine-grained classification, contains 200 bird species with similar appearance where the image size is $224 \times 224$ . We follow + +the splits from [32]. For miniImageNet and CIFAR100, 60 categories are selected as base classes while the remaining are split into 8 incremental sessions with only 5 training examples per novel class (i.e., 5-way 5-shot). As for the CUB200 dataset, 100 categories are designated as the base training set, while the rest are divided into 10-way 5-shot tasks across 10 sessions. + +Evaluation protocol. In line with the FSCIL literature, we employ the top-1 accuracy (Acc) for all the observed classes at the end of each incremental learning session. We also provide the average of Acc over all the sessions. + +Implementation details. We implement our method using the PyTorch library. In keeping with the previous works [18, 25, 29, 45], ResNet18 [14] serves as our feature extractor. The entire model is trained via the SGD optimizer. For the base session, the learning rates are initialized by 0.01, 0.001, and 0.01 for miniImageNet, CUB200, and CIFAR100 respectively, and decayed by 0.1 at 60 and 70 epochs. In each incremental session, the learning rate for the weight-space ensembled $\phi_{\mathrm{all}}$ is 0.1, while it is 0.001 for the rest. Hence, the feature extractor is minimally updated in incremental learning. For the ADKD loss, the given $NK$ training examples are amplified to $16NK$ for each class. Also, the learnable scalars $\alpha_{1}$ and $\alpha_{2}$ are initially set to 1.0. Loss weights $\lambda_{1}$ and $\lambda_{2}$ are empirically determined to be 1.2 and 10.0, respectively. Further details are in Sec. S-5 of the supplementary materials. + +# 4.1. Comparative assessment + +On the three benchmark datasets, we first compare the proposed method with the existing approaches. As indicated in Table 1 on the miniImageNet dataset, our method outperforms the compared in all the sessions, achieving a clear SOTA in terms of the average Acc by a margin of $1.82\%$ at least. Also, BiDistill and SAVC require the base session model and a considerable number of classifiers per class, respectively, even in the testing phase. NC-FSCIL and OrCo benefit from knowing the total number of incremental classes at the outset of the base session. Whereas, we attain clear SOTA with a smaller-scale deployed model compared to BiDistill and SAVC, and without the information advantage of NC-FSCIL and OrCo. + +Moreover, Fig. 4 shows the Acc curve according to sessions on the CUB200 and CIFAR100 datasets. As in the miniImageNet dataset, our method consistently surpasses the compared methods by large gaps in the incremental sessions on the CUB 200 dataset. In specific, despite lower base session Acc than SAVC, our method outperforms it in the last session. Also, while NC-FSCIL merely finetuned the last module of the feature encoder in incremental sessions, OrCo attempted to also optimize the pre-set classifiers. Nevertheless, the additional update of only the classifiers is less effective. On the CIFAR100 dataset, we can + +Table 1. Comparative results on miniImageNet dataset. Session-wise Acc (%) and the average of them (Avg) are reported. In each case, the best and second results are bold-faced and underlined, respectively. Last sess. Impro. denotes the performance improvement of our method at the last session. + +
MethodVenueAcc in each session (%)AvgLast sess. impro.
012345678
iCaRL [27]CVPR 201761.3146.3242.9437.6330.4924.0020.8918.8017.2033.29+43.13
TOPIC [32]CVPR 202061.3150.0945.1741.1637.4835.5232.1929.4624.4239.64+35.71
ERL++ [11]AAAI 202161.7057.5854.6651.7248.6646.2744.6742.8140.7949.87+19.34
IDLVQ [4]ICLR 202064.7759.8755.9352.6249.8847.5544.8343.1441.8451.16+18.29
CEC [42]CVPR 202172.0066.8362.9759.4356.7053.7351.1949.2447.6357.74+12.50
SynthFeat [5]ICCV 202161.4059.8054.2051.6949.4548.0045.2043.8042.1050.63+18.03
MetaFSCIL [6]CVPR 202272.0467.9463.7760.2957.5855.1652.950.7949.1958.85+10.94
FACT [46]CVPR 202272.5669.6366.3862.7760.6057.3354.3452.1650.4960.70+9.64
Replay [23]ECCV 202271.8467.1263.2159.7757.0153.9551.5549.5248.2158.02+11.92
ALICE [25]ECCV 202280.6070.6067.4064.5062.5060.0057.8056.8055.7063.99+4.43
S3C [17]ECCV 202276.5571.7467.6664.5261.5158.0955.3653.0451.3462.20+8.79
WaRP [19]ICLR 202372.9968.1064.3161.3058.6456.0853.4051.7250.6559.69+9.48
SoftNet [18]ICLR 202376.6370.1365.9262.5259.4956.5653.7151.7250.4860.79+9.65
NC-FSCIL [40]ICLR 202384.0276.8072.0067.8366.3564.0461.4659.5458.3167.82+1.82
GKEAL [47]CVPR 202373.5968.9065.3362.2959.3956.7054.2052.5951.3160.48+10.40
BiDistill [45]CVPR 202374.6570.4366.2962.7760.7557.2454.7953.6552.2261.42+7.91
SAVC [29]CVPR 202381.1276.1472.4368.9266.4862.9559.9258.3957.1167.05+3.02
OrCo [1]CVPR 202483.3075.3271.5268.1665.6263.1160.2058.8258.0867.12+2.05
CLOSER [24]ECCV 202476.0271.6167.9964.6961.7058.9456.2354.5253.3362.78+6.80
Ours84.1381.4176.6573.5970.1065.1363.4261.0260.1370.62
+ +![](images/be9d2f5bd94ff302ff07b19ff3c3c98a62f8cd9da1d84b65295a4d3614f017c9.jpg) + +![](images/10d9b2ecc6b2a2673e4c4b064c54d65b6ce092849db4b10e3f511e6fb2912e96.jpg) +(a) CUB200 + +![](images/05a5fc2d0452e227a6f8c82e86b080a735b41a885e95e1c215847e1eed607491.jpg) +Figure 4. Comparative results on CUB200 and CIFAR100 datasets. For each dataset, the curve of the session-wise Acc (%) is first provided. And also, the base, last, average Acc are reported on our method comparing with three recent SOTA methods, NC-FSCIL [40], SAVC [29], YourSelf [31] (Best viewed in color). + +![](images/9e0824fa1e85b63a4eedcb5a36fa5ae07e2e1c4934c2e59b53a376a2b3d62d9c.jpg) +(b) CIFAR100 + +![](images/6df7fd8303699a4f3d92d890b86dacfb9ab7ad71f4cc93eeaaf7e835c776fb61.jpg) + +see the similar tendency. NC-FSCIL's setting is somewhat favorable, as it allows the pre-definition of classifiers for the incremental classes with knowledge of all class numbers in advance at the base session. This advantage is particularly beneficial on the CIFAR100 dataset, where some classes are semantically similar (e.g., possum and mouse), and the low resolution $(32 \times 32)$ of the images could otherwise complicate classification. The detailed results are available in Sec. S-4 of the supplementary materials. These findings demonstrate our method's superior ability to adapt to new classes with limited training data while maintaining seamless integration with the old classes. + +# 4.2. Ablation analysis + +We extensively analyze the components of the proposed method on the miniImageNet or CUB200 datasets. + +Tripartite strategy. We conduct the ablation experiments on the Tri-WE in Table 2. Before this, initially, we contrast "Naive" with the "No WE" configuration, the latter incorporating the ADKD loss. The naive approach, lacking both WE and ADKD loss, exhibits a significant performance decline due to simultaneous updates of the feature extractor and classification head. However, No WE achieves respectable results thanks to the ADKD loss (more discussed in the following sections). Next, as demonstrated + +Table 2. Ablation analysis for Tri-WE. Varying the combination of classification heads ensembled in weight space (No WE, Dual-WE, Tri-WE), session-wise and average Acc (%) are provided on miniImageNet. Also, "Naive" means that none of WE and ADKD is used. + +
Classification head weightAcc in each session (%)Avg
φ0φoldφall012345678
Naive84.1380.5571.2660.9350.0544.0836.4219.7816.7351.55
No WE✓84.1378.9574.0970.0666.5663.0760.3258.0856.0167.92
Dual-WE✓✓84.1380.2575.4771.5867.6264.5261.1160.1758.9369.31
✓84.1379.5474.9970.1366.6063.4360.8559.8957.3168.54
Tri-WE✓✓✓84.1381.4176.6573.5970.165.1363.4261.0260.1370.62
+ +Table 3. Impact on increasing the layers ensembled in weight space. Applying the proposed Tri-WE to the convolutional blocks of the feature extractor (ResNet18), session-wise and average Acc (%) are provided on miniImageNet. + +
Weight-space ensembledAcc in each session (%)Avg
012345678
Blocks 1-484.1375.2570.2567.5464.8561.3459.7456.3853.4565.88
Blocks 1-4 + clf. head84.1376.0872.1368.4265.8362.6060.1957.7555.2966.94
Blocks 2-4 + clf. head84.1377.8473.4269.2265.8562.7560.5657.3455.3367.38
Blocks 3-4 + clf. head84.1378.5974.4369.8066.2563.1360.4658.0455.6567.83
Block 4 + clf. head84.1378.8074.2169.6466.0163.3261.5058.1755.8867.96
Clf. head only84.1381.4176.6573.5970.165.1363.4261.0260.1370.62
+ +![](images/e492489135de153352ab97b8a4ab4066973d26ddffee44f146a4bdfea9c34d3a.jpg) +(a) Fixed $(\alpha_{1},\alpha_{2})$ +Figure 5. Analysis on learnable scalars $\alpha_{1}$ and $\alpha_{2}$ on miniImageNet. (a) Comparison with multiple combinations of fixed (not-learnable) $\alpha_{1}$ and $\alpha_{2}$ in terms of averaged Acc over all the sessions, (b) session-wise learned $\alpha_{1}$ and $\alpha_{2}$ in ours. + +![](images/d9c3609745a907048338b9539db934e28b6dbbbc2989d41a7c3dbc047c3e5cf2.jpg) +(b) Learned $\alpha_{1},\alpha_{2}$ + +in Table 2, the absence of the weight-space ensemble for the classification head leads to increased catastrophic forgetting. And, from "Dual-WE," we can see that the model from the intermediate previous session is more important than the base session model, as the reduced learning rate for the feature extractor mitigates drastic changes, helping to preserve the integrity of the feature space. Accordingly, the knowledge for incrementally-appeared classes is more helpful. Lastly, the complete Tri-WE clearly outperforms No WE by $2.70\%$ in average Acc, effectively retaining the essential knowledge of the base and previous session models. + +WE in classification head. We also evaluate the effect of + +Table 4. Analysis on data amplification schemes. The last and average Acc (%) are reported on each scheme on miniImageNet. + +
w/o ADAD scheme
CutOut [9]RandAug [7]MixUp [43]CutMix [41]
Last Acc58.0858.1258.2659.1460.13
Avg Acc69.0569.1369.0970.0370.62
+ +weight-space ensemble varying the number of interpolated layers in Table 3. In specific, we apply the proposed TriWE on the convolutional blocks of the feature extractors. Here, we use different learnable scalars depending on the convolutional blocks, rather than sharing them across the convolutional blocks and classification head. We observed a tendency for the decline in performance as the number of weight-space-ensemblied layers increases. This is because that deciding how to merge a lot of layers is difficult with only a few examples. Also, only merging feature extractor (1st row of Table 3) shows the lowest performance. It emphasizes that the head's weight-space is crucial to prevent catastrophic forgetting. + +Learnable scalars $\alpha_{1} \& \alpha_{2}$ . Fig. 5(a) shows the averaged Acc on diverse combinations of $(\alpha_{1}, \alpha_{2})$ where both are fixed after initialization in all the sessions. Compared to all the fixed cases, the learnable scalars show better result (70.62% in the averaged Acc). Also, as Fig. 5(b), we show the learned $\alpha_{1}$ and $\alpha_{2}$ in each session in our method where we initialize both as 1.0 in every session. We can see that the learned $\alpha_{1}$ and $\alpha_{2}$ are different depending on sessions. Accordingly, our learnable scalar approach is useful for bal + +Table 5. Analysis on loss terms of ${\mathcal{L}}_{\mathrm{{ADKD}}}$ . The last and average Acc (%) are reported abating the loss terms on miniImageNet. + +
LfeatLlogitLast AccAvg Acc
✓59.4270.01
✓58.7768.41
✓✓60.1370.62
+ +ancing the trade-off among the three classification heads. We also provide session-wise results on the fixed cases in the supplementary materials. + +Effect of data amplification. While Tri-WE already attains SOTA (i.e. w/o AD in Table 4), incorporating our ADKD loss provides further increase (69.05% vs 70.62% in average Acc). Also, to see the influence of the choice of data amplification scheme in the ADKD loss, we compare multiple data augmentation methods for our data amplification. To facilitate the use of ADKD in follow-up works, we consider four widely-used data augmentation methods, CutOut [9], RandAug [7], CutMix [41], MixUp [43]. In the first two, the images are not inter-mixed. Whereas, two images are mixed for data augmentation. In Table 4, we can see that CutOut and RandAug are not effective compared to "w/o AD" where no data amplification is considered. With only a few examples, those methods cannot yield a diversity of data. Next, in comparison with MixUp, CutMix shows a better result. MixUp mixes two images smoothly, and then the original semantics are somewhat preserved. This is less useful in KD. Namely, CutMix encourages the model to focus more on general image recognition ability rather than high-level semantics, which is more helpful in preventing the catastrophic forgetting problem in FSCIL. Hence, we employ CutMix as the default option. In addition, the impact of the degree of data amplification is addressed in Sec. S-1 of the supplementary materials. + +Two loss terms of $\mathcal{L}_{\mathrm{ADKD}}$ . As shown in the first two rows of Table 5, we ablate $\mathcal{L}_{\mathrm{feat}}$ and $\mathcal{L}_{\mathrm{logit}}$ , respectively. From this study, we can see that $\mathcal{L}_{\mathrm{feat}}$ is slightly more important than $\mathcal{L}_{\mathrm{logit}}$ since Tri-WE already helps to prevent catastrophic forgetting in the side of the classification head. Then, using both is most effective. + +$\mathcal{L}_{\mathrm{Cls - Old}}$ . As we update the feature extractor as well as the classification head in the incremental learning, we also adapt $\phi_{\mathrm{old}}$ to the updated feature extractor using the loss $\mathcal{L}_{\mathrm{Cls - Old}}$ . To verify this, we conduct ablating the feature extractor update and $\mathcal{L}_{\mathrm{Cls - Old}}$ . As shown in Table 6, the performance is degraded without this loss term when the feature extractor is updated. Also, when the feature extractor is frozen, updating $\phi_{\mathrm{old}}$ is less meaningful. Finally, updating the feature extractor with $\phi_{\mathrm{old}}$ is most beneficial. + +New class adaptability. To see the efficacy of our method in terms of the trade-off between base and newly introduced classes, we provide the individual accuracy of the + +Table 6. Analysis on ${\mathcal{L}}_{\text{Cls-Old }}$ and updating the feature extractor ${g}_{\theta }$ . The last and average Acc (%) are reported,on miniImageNet. + +
LCls-OldUpdate gθLast AccAvg Acc
✓59.2770.15
✓✓60.1370.62
58.8669.41
✓58.8869.40
+ +![](images/82e4ec3c0847b3fb628da1b5073e460b37cbff7168326273a7898dcff36d4f07.jpg) +Figure 6. Analysis on the new class adaptability on CUB200. The average of accuracy scores are provided for base and novel classes, denoted by Base and New respectively. And, harmonic means are also compared. + +base and novel classes, and the harmonic mean (HM). Since most previous works have addressed the joint accuracy of the base and novel classes, we can compare our method with only a few recent methods [17, 33, 42, 45, 46] on the CUB200 dataset, following [17, 46]. Fig. 6 shows that our approach is more adept at integrating new classes without compromising the performance on the base classes. + +# 5. Conclusions + +We explored FSCIL which facilitates the continuous acquisition of new concepts from a few examples. To this end, we suggested a novel FSCIL method including Tri-WE and ADKD regularization loss. The proposed method can effectively prevent overfitting to the few data and catastrophic forgetting of old concepts in the model. Tri-WE strategically interpolates the weights of the classification heads from the base, immediate previous, and current models, which collectively preserves the knowledge from the earlier models without overwhelming updates. Enriching the few examples for KD, the ADKD loss facilitates the transfer of essential knowledge from the previous model, rather than the knowledge overfitted to the few examples. We conducted extensive analyses on the proposed components and achieved state-of-the-art performance on the miniImageNet, CUB200, and CIFAR100 datasets. + +# References + +[1] Noor Ahmed, Anna Kukleva, and Bernt Schiele. Orco: Towards better generalization via orthogonality and contrast for + +few-shot class-incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. +[2] Eden Belouadah and Adrian Popescu. II2m: Class incremental learning with dual memory. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. +[3] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. arXiv preprint arXiv:1812.00420, 2018. +[4] Kuilin Chen and Chi-Guhn Lee. Incremental few-shot learning via vector quantization in deep embedded space. In Int. Conf. Learn. Represent., 2020. +[5] Ali Cheraghian, Shafin Rahman, Sameera Ramasinghe, Pengfei Fang, Christian Simon, Lars Petersson, and Mehrtash Harandi. Synthesized feature based few-shot class-incremental learning on a mixture of subspaces. In Int. Conf. Comput. Vis., 2021. +[6] Zhixiang Chi, Li Gu, Huan Liu, Yang Wang, Yuanhao Yu, and Jin Tang. Meta-scil: A meta-learning approach for few-shot class incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. +[7] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In IEEE Conf. Comput. Vis. Pattern Recog. Worksh., 2020. +[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conf. Comput. Vis. Pattern Recog., 2009. +[9] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. +[10] Thomas G Dietterich. Ensemble methods in machine learning. In International Workshop on Multiple Classifier Systems, 2000. +[11] Songlin Dong, Xiaopeng Hong, Xiaoyu Tao, Xinyuan Chang, Xing Wei, and Yihong Gong. Few-shot class incremental learning via relation knowledge distillation. In AAAI, 2021. +[12] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In Eur. Conf. Comput. Vis., 2020. +[13] Lars Kai Hansen and Peter Salamon. Neural network ensembles. IEEE Trans. Pattern Anal. Mach. Intell., 12(10): 993-1001, 1990. +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., 2016. +[15] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. +[16] Gao Huang, Yixuan Li, Geoff Pleiss, Zhuang Liu, John E Hopcroft, and Kilian Q Weinberger. Snapshot ensembles: Train 1, get m for free. In Int. Conf. Learn. Represent. 2017. +[17] Jayateja Kalla and Soma Biswas. S3c: Self-supervised stochastic classifiers for few-shot class-incremental learning. In Eur. Conf. Comput. Vis., 2022. + +[18] Haeyong Kang, Jaehong Yoon, Sultan Rizky Hikmawan Madjid, Sung Ju Hwang, and Chang D Yoo. On the soft-subnetwork for few-shot class incremental learning. In Int. Conf. Learn. Represent., 2023. +[19] Do-Yeon Kim, Dong-Jun Han, Jun Seo, and Jaekyun Moon. Warping the space: Weight space rotation for class incremental few-shot learning. In Int. Conf. Learn. Represent., 2023. +[20] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. +[21] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Adv. Neural Inform. Process. Syst. 2012. +[22] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Trans. Pattern Anal. Mach. Intell., 40(12):2935-2947, 2017. +[23] Huan Liu, Li Gu, Zhixiang Chi, Yang Wang, Yuanhao Yu, Jun Chen, and Jin Tang. Few-shot class-incremental learning via entropy-regularized data-free replay. In Eur. Conf. Comput. Vis., 2022. +[24] Junghun Oh, Sungyong Baik, and Kyoung Mu Lee. Closer: Towards better representation learning for few-shot class-incremental learning. In Eur. Conf. Comput. Vis., 2024. +[25] Can Peng, Kun Zhao, Tianren Wang, Meng Li, and Brian C Lovell. Few-shot class-incremental learning from an open-set perspective. In Eur. Conf. Comput. Vis., 2022. +[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Int. Conf. Machine Learning, 2021. +[27] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2017. +[28] David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. Experience replay for continual learning. 2019. +[29] Zeyin Song, Yifan Zhao, Yujun Shi, Peixi Peng, Li Yuan, and Yonghong Tian. Learning with fantasy: Semantic-aware virtual contrastive constraint for few-shot class-incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. +[30] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In IEEE Conf. Comput. Vis. Pattern Recog. 2015. +[31] Yu-Ming Tang, Yi-Xing Peng, Jingke Meng, and Wei-Shi Zheng. Rethinking few-shot class-incremental learning: Learning from yourself. In Eur. Conf. Comput. Vis., 2024. +[32] Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, and Yihong Gong. Few-shot class-incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2020. +[33] Oriol Vinyals, Charles Blundell, Tim Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In NeurIPS, 2016. + +[34] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. +[35] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. +[36] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Marcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In Int. Conf. Machine Learning, 2022. +[37] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, and Ludwig Schmidt. Robust fine-tuning of zero-shot models. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. +[38] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. +[39] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. +[40] Yibo Yang, Haobo Yuan, Xiangtai Li, Zhouchen Lin, Philip Torr, and Dacheng Tao. Neural collapse inspired feature-classifier alignment for few-shot class incremental learning. In Int. Conf. Learn. Represent., 2023. +[41] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Int. Conf. Comput. Vis., 2019. +[42] Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, and Yinghui Xu. Few-shot incremental learning with continually evolved classifiers. In IEEE Conf. Comput. Vis. Pattern Recog., 2021. +[43] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. +[44] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu-Tao Xia. Maintaining discrimination and fairness in class incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2020. +[45] Linglan Zhao, Jing Lu, Yunlu Xu, Zhanzhan Cheng, Dashan Guo, Yi Niu, and Xiangzhong Fang. Few-shot class-incremental learning via class-aware bilateral distillation. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. +[46] Da-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, Liang Ma, Shiliang Pu, and De-Chuan Zhan. Forward compatible few-shot class-incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. +[47] Huiping Zhuang, Zhenyu Weng, Run He, Zhiping Lin, and Ziqian Zeng. Gkeal: Gaussian kernel embedded analytic learning for few-shot class incremental task. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. \ No newline at end of file diff --git a/2025/Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning/images.zip b/2025/Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f36900f5fe99932487d85f087e05cda198182482 --- /dev/null +++ b/2025/Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3027f7b86b3e30ac12108072f5a4ba5659d16d6025b2ec65ee5ef57507ab41c0 +size 565043 diff --git a/2025/Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning/layout.json b/2025/Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7141f7616036011c6988b5b5d3e2b5a033568d5a --- /dev/null +++ b/2025/Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning/layout.json @@ -0,0 +1,9352 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 60, + 103, + 533, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 103, + 533, + 121 + ], + "spans": [ + { + "bbox": [ + 60, + 103, + 533, + 121 + ], + "type": "text", + "content": "Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 182, + 144, + 411, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 144, + 411, + 171 + ], + "spans": [ + { + "bbox": [ + 182, + 144, + 411, + 171 + ], + "type": "text", + "content": "Juntae Lee Munawar Hayat Sungrack Yun \nQualcomm AI Research†" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 186, + 174, + 408, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 174, + 408, + 186 + ], + "spans": [ + { + "bbox": [ + 186, + 174, + 408, + 186 + ], + "type": "text", + "content": "{juntlee, hayat, sungrack}@qti.QUALCOMM.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 224, + 192, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 224, + 192, + 238 + ], + "spans": [ + { + "bbox": [ + 143, + 224, + 192, + 238 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 250, + 290, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 250, + 290, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 250, + 290, + 574 + ], + "type": "text", + "content": "Few-shot class incremental learning (FSCIL) enables the continual learning of new concepts with only a few training examples. In FSCIL, the model undergoes substantial updates, making it prone to forgetting previous concepts and overfitting to the limited new examples. Most recent trend is typically to disentangle the learning of the representation from the classification head of the model. A well-generalized feature extractor on the base classes (many examples and many classes) is learned, and then fixed during incremental learning. Arguing that the fixed feature extractor restricts the model's adaptability to new classes, we introduce a novel FSCIL method to effectively address catastrophic forgetting and overfitting issues. Our method enables to seamlessly update the entire model with a few examples. We mainly propose a tripartite weight-space ensemble (Tri-WE). Tri-WE interpolates the base, immediately previous, and current models in weight-space, especially for the classification heads of the models. Then, it collaboratively maintains knowledge from the base and previous models. In addition, we recognize the challenges of distilling generalized representations from the previous model from scarce data. Hence, we suggest a regularization loss term using amplified data knowledge distillation. Simply intermixing the few-shot data, we can produce richer data enabling the distillation of critical knowledge from the previous model. Consequently, we attain state-of-the-art results on the miniImageNet, CUB200, and CIFAR100 datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 594, + 128, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 128, + 607 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 128, + 607 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 614, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 614, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 614, + 287, + 687 + ], + "type": "text", + "content": "The ability of a model to continually adapt with the availability of new data is highly desirable for many real-world applications. Few-shot class-incremental learning (FSCIL) has recently gained significant research attention as evidenced by several studies [17, 25, 29, 42, 45], because it offers a promising learning paradigm. It enables the model" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 215, + 550, + 355 + ], + "blocks": [ + { + "bbox": [ + 310, + 215, + 550, + 355 + ], + "lines": [ + { + "bbox": [ + 310, + 215, + 550, + 355 + ], + "spans": [ + { + "bbox": [ + 310, + 215, + 550, + 355 + ], + "type": "image", + "image_path": "566cc08011a0b6e754bd42e568a4dbf5b71b71a8b6a794a98be3ecbb0c0619be.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 364, + 547, + 453 + ], + "lines": [ + { + "bbox": [ + 304, + 364, + 547, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 547, + 453 + ], + "type": "text", + "content": "Figure 1. Overview of the proposed method (FE: feature extractor, Cls. head: classification head). Through Tri-WE, at " + }, + { + "bbox": [ + 304, + 364, + 547, + 453 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 364, + 547, + 453 + ], + "type": "text", + "content": "th incremental session, Cls. head is composed by interpolating the weights of the Cls. heads covering the classes until base, " + }, + { + "bbox": [ + 304, + 364, + 547, + 453 + ], + "type": "inline_equation", + "content": "(t - 1)" + }, + { + "bbox": [ + 304, + 364, + 547, + 453 + ], + "type": "text", + "content": "th, " + }, + { + "bbox": [ + 304, + 364, + 547, + 453 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 364, + 547, + 453 + ], + "type": "text", + "content": "th sessions, respectively. For ADKD regularization loss, the input few-shot data are amplified to reliably mine generalized knowledge of the model of the previous " + }, + { + "bbox": [ + 304, + 364, + 547, + 453 + ], + "type": "inline_equation", + "content": "(t - 1)" + }, + { + "bbox": [ + 304, + 364, + 547, + 453 + ], + "type": "text", + "content": "th session. After incremental learning, deployed components are outlined in green." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 473, + 546, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 546, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 546, + 593 + ], + "type": "text", + "content": "to flexibly expand its classification capabilities to new concepts using only a few training examples. FSCIL starts by training a model on numerous base classes with abundant labeled data. In each subsequent incremental session, the model encounters only a limited number of labeled data for new classes, along with a single prototype for each of the previously learned classes. Hence, FSCIL faces multiple challenges, including the retention of previously learned knowledge and the prevention of overfitting to the limited examples of new data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "To address these challenges, the most recent works are based on separating learning the representation from the classification. They rely on constructing a robust feature embedding space using plentiful base class data. In subsequent incremental learning where only a few labeled data for new classes are available, the embedding space remains fixed, and the focus is shifted to learning the classifier for the new classes. This strategy is intended to avoid accumulated model collapse or overfitting due to the limited data. For example, [25, 42] refrain from altering any base" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 693, + 286, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 286, + 712 + ], + "type": "text", + "content": "†Qualcomm AI Research is an initiative of Qualcomm Technologies, Inc. and/or its subsidiaries." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15329" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": "model parameters and use mean features for each class to define their classifiers. In contrast, [17, 29] employ ensembling multiple classifiers to improve generalization. [45] introduces an extra module to assess the reliability of outputs from the base and incrementally learned models for a weighted sum of the two outputs. Moreover, [1, 40] fixes the feature embedding space and also pre-defines classifier templates, requiring prior knowledge of the total number of incremental classes to be encountered at the outset of the base class learning. However, as indicated in Fig. 2, updating the entire model can lead to a drop in base class performance but may enhance the model's adaptability to new classes. Hence, if we address the catastrophic forgetting and overfitting more effectively, we can harness the benefits of updating the entire model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 259, + 289, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 259, + 289, + 535 + ], + "spans": [ + { + "bbox": [ + 46, + 259, + 289, + 535 + ], + "type": "text", + "content": "To this end, we propose a novel framework for FSCIL as depicted in Fig. 1. Mainly, inspired by Wise-FT [37], we develop tripartite weight-space ensemble (Tri-WE) to continually update the weights of the model. We interpolate the classification heads of models from the base, intermediately previous, and current sessions within the weight space, thereby achieving a balance between adaptability to new classes and retention of knowledge of previous ones. This approach allows for a continual update of knowledge without an abrupt shift in the decision boundary. Moreover, to address catastrophic forgetting, knowledge distillation (KD) has been a common regularization technique in class incremental learning. However, its application in a few-shot scenario is prone to overfitting to the few available examples, and then recent cutting-edge methods have less used KD for FSCIL. To counter this, we also introduce amplified data KD (ADKD) regularization loss where we intermix few-shot training datasets, enriching the data source for distilling the knowledge from the previous model. It simply ensures that the model prevents catastrophic forgetting by transferring the generalized knowledge for previous classes. We achieve state-of-the-art (SOTA) results on three benchmarks: miniImageNet, CUB200, and CIFAR100." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 542, + 254, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 542, + 254, + 554 + ], + "spans": [ + { + "bbox": [ + 59, + 542, + 254, + 554 + ], + "type": "text", + "content": "Our contributions are summarized as three-fold:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 562, + 286, + 711 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 562, + 286, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 286, + 610 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 286, + 610 + ], + "type": "text", + "content": "- We introduce the Tri-WE, wherein the classification heads from base, preceding, and current models are synergistically interpolated within the weight space, promoting a knowledge continuum." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 613, + 286, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 613, + 286, + 660 + ], + "spans": [ + { + "bbox": [ + 47, + 613, + 286, + 660 + ], + "type": "text", + "content": "- Tackling distillation of biased knowledge resulting from a scarcity of data in KD, we develop ADKD sub-loss term that enhances the generality of the distilled knowledge." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 663, + 286, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 663, + 286, + 711 + ], + "spans": [ + { + "bbox": [ + 47, + 663, + 286, + 711 + ], + "type": "text", + "content": "- Leveraging these proposed components, we attain SOTA results across on miniImageNet, CUB200, and CIFAR100. Comprehensive experiments demonstrate the significance and efficacy of our contributions." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 312, + 71, + 539, + 190 + ], + "blocks": [ + { + "bbox": [ + 312, + 71, + 539, + 190 + ], + "lines": [ + { + "bbox": [ + 312, + 71, + 539, + 190 + ], + "spans": [ + { + "bbox": [ + 312, + 71, + 539, + 190 + ], + "type": "image", + "image_path": "bf9ef0d970d4412fc36a9a386982c91fe9da3e3460b23c92dba24c4266056228.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 200, + 545, + 243 + ], + "lines": [ + { + "bbox": [ + 305, + 200, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 305, + 200, + 545, + 243 + ], + "type": "text", + "content": "Figure 2. Few-shot (10-way 5-shot miniImageNet) transfer learning (TL) starting from the base session model on each session. Classification accuracy scores on incremented 10 classes (top) & 40 base classes (bottom) are shown for each session." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 269, + 398, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 269, + 398, + 282 + ], + "spans": [ + { + "bbox": [ + 306, + 269, + 398, + 282 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 292, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 292, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 545, + 483 + ], + "type": "text", + "content": "Class-incremental learning (CIL) aims to seamlessly integrate new classes without access to previously seen data, preventing the catastrophic forgetting problem. To ensure knowledge retention, rehearsal of selective previous samples [3, 28] has been developed. Also, in [35, 39], the network is dynamically expanded to accommodate new learnings. LwF [22] and iCaRL [27] used KD and nearest class mean (NCM) classifiers, respectively, to mitigate the catastrophic forgetting. UCIR [15] introduced a novel feature distillation loss, and PODNet [12] extended distillation beyond the final embedding output, incorporating the pooled outputs from the model's intermediate layers. In [2, 38, 44], the class imbalance problem is tackled. However, these methods depend on ample training data, an aspect not feasible in FSCIL, necessitating new strategies that adapt swiftly with minimal data while preserving existing knowledge." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": "FSCIL fuses CIL and few-shot learning. It gained attention after TOPIC [32] initiated the concept, leading to numerous subsequent methods. Recent works align with two strategies. One approach completely fixes the learnable weights of the model after the base session's many-shot many-class training, and then incrementally extends the classifier by adding mean features of new classes. CEC [42] enhances the base model by simulating the FSCIL scenarios sampled from the base data again, while ALICE [25] employs diverse techniques to boost the generalization capability of the base model. Alternatively, some methods [17, 29, 40, 45] update a part of the feature extractor or the added class-mean-initialized classifiers in each session. SoftNet [18] finds the weights less impactful for the catastrophic forgetting at the base session, and focuses on updating them in the incremental learning. In WaRP [19], the important weights are decided in every session, and they keep frozen in the following sessions. GKEAL [47] updates the weight of the classification head using least square" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15330" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "text", + "content": "solver in the view of analytic learning. S3C [17] employed the stochastic classifier as the extreme ensemble of classifiers. NS-FSCIL [40] and OrCo [1] assume the base session knows the amount of total incremental classes and incremental session starts with mapping each class to a classifier predefined in the base session. NS-FSCIL optimizes the last module of the feature encoder and OrCo does the last module and the selected pre-defined classifier. In SAVC [29], a class is recognized by a set of classifiers, each tailored to a specific input augmentation, resulting in a swift expansion of deployed model size during incremental learning. In BiDistill [45], every prediction is made by the ensemble of the base and last models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 289, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 289, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 289, + 456 + ], + "type": "text", + "content": "Network ensemble has long been an important technique in deep learning in order to reduce overfitting and improve generalization ability [10, 13]. It involves the training of multiple neural networks and combining them for the final prediction. Various ensemble methods have been developed to get different but collaborative models such as random-weight initialization [21], different training data [30], and capturing multiple snapshots in training schedules [16]. While these methods have consistently demonstrated improved performance over single-model approaches, they often come at the cost of increased computational overhead, both in terms of training and inference. Recently, Model Soup [36] suggested interpolating the weights of a bunch of networks finetuned with different hyperparameters configurations. Also, to mitigate the overfitting of the finetuned CLIP [26], WiseFT [37] combines the weights of fine-tuned and initial CLIP models. These approaches can attain better generalization capability without increasing the computational cost of deployed models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 466, + 153, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 466, + 153, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 466, + 153, + 479 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 486, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 521 + ], + "type": "text", + "content": "In this section, we describe the proposed Tri-WE and ADKD for FSCIL. The pipeline is illustrated in Fig. 3. We also provide the process of the base session training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "text", + "content": "Problem setting. FSCIL aims to incrementally scale the model's class recognition ability by learning new classes with limited training data while preserving its knowledge of previously learned, old classes. In specific, we suppose that a model " + }, + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "inline_equation", + "content": "f_{\\theta,\\phi}" + }, + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "text", + "content": " consists of the feature extractor " + }, + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "text", + "content": " and the classification head (cosine classifier) " + }, + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "inline_equation", + "content": "h_{\\phi}" + }, + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "text", + "content": " denote learnable weights of them, respectively. Initially, the model is pre-trained on the base training set in the base session " + }, + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "inline_equation", + "content": "S^{(0)}" + }, + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "text", + "content": ". Then, the model progresses a series of " + }, + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "text", + "content": " incremental sessions, " + }, + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\{S^{(1)},\\ldots,S^{(T)}\\}" + }, + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "text", + "content": ". Current session " + }, + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "inline_equation", + "content": "S^{(t)}" + }, + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "text", + "content": " includes a class set " + }, + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "inline_equation", + "content": "C^{(t)}" + }, + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "text", + "content": " and the corresponding training dataset " + }, + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "inline_equation", + "content": "D^{(t)} = \\{(x,y)\\}" + }, + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "text", + "content": " is a sample and " + }, + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 522, + 287, + 665 + ], + "type": "text", + "content": " is its label." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "In the base session " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "S^{(0)}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": ", the class set " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{(0)}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " includes a substantial number of " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "N_0" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " base classes with ample training data for each class. In contrast, in the incremental session " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "S^{(t)}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "t \\geq 1" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "), the class set " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{(t)}" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " classes with only a" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": "small " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": " number of training data available for every class " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "(N" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": "-way " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": "-shot setting). Notice that " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{(i)}" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{(j)}" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": " do not overlap when " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "i \\neq j, \\forall i, j" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": ". Following the prior FSCIL works [32, 42], the model at each session " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{S}^{(t)}" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": " only has access to " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{(t)}" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": " and an additional prototype (the averaged feature of the " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": " examples per class) buffer " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": " memorizing the classes of the base and previous " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "(t - 1)" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": " sessions. A single prototype is maintained for every class in " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": ". Once the incremental learning at the session " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{S}^{(t)}" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": " concludes, the model is evaluated on query data from all the seen " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "N_{t}" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": " classes thus far, that is " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{(0:t)} = \\mathcal{C}^{(0)} \\cup \\mathcal{C}^{(1)} \\cdots \\cup \\mathcal{C}^{(t)}" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 236, + 529, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 236, + 529, + 249 + ], + "spans": [ + { + "bbox": [ + 305, + 236, + 529, + 249 + ], + "type": "text", + "content": "3.1. Tripartite weight-space ensemble (Tri-WE)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 262, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 545, + 358 + ], + "type": "text", + "content": "The base session model is generalized by numerous classes with sufficient data, and hence its weight space is conducive to adapting to new classes. Although the weight space is somewhat compromised, the intermediate previous session's model has knowledge for all the classes encountered thus far. To capitalize on the strengths of both models, we suggest the weight-space ensemble of three models from the sessions " + }, + { + "bbox": [ + 304, + 262, + 545, + 358 + ], + "type": "inline_equation", + "content": "S^{(0)}" + }, + { + "bbox": [ + 304, + 262, + 545, + 358 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 262, + 545, + 358 + ], + "type": "inline_equation", + "content": "S^{(t-1)}" + }, + { + "bbox": [ + 304, + 262, + 545, + 358 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 262, + 545, + 358 + ], + "type": "inline_equation", + "content": "S^{(t)}" + }, + { + "bbox": [ + 304, + 262, + 545, + 358 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "text", + "content": "However, the feature extractor, which contains many layers, is prone to overfitting or collapsing when only a few examples are available. Hence, while reducing the learning rate for the feature extractor " + }, + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "inline_equation", + "content": "g_{\\theta}^{(t)}" + }, + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "text", + "content": ", we only ensemble the classification heads " + }, + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "inline_equation", + "content": "h_{\\phi_0}" + }, + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "inline_equation", + "content": "h_{\\phi_{\\mathrm{old}}}^{(t)}" + }, + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "inline_equation", + "content": "h_{\\phi_{\\mathrm{all}}}^{(t)}" + }, + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "text", + "content": " in the weight space. Here, " + }, + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "inline_equation", + "content": "h_{\\phi_0}" + }, + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "text", + "content": " is the classification head from the base session model. " + }, + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "inline_equation", + "content": "h_{\\phi_{\\mathrm{old}}}^{(t)}" + }, + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "text", + "content": " is initialized by the classification head " + }, + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "inline_equation", + "content": "h_{\\phi}^{(t - 1)}" + }, + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "text", + "content": " from the session " + }, + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "inline_equation", + "content": "S^{(t - 1)}" + }, + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "text", + "content": ", and then specialized to the old classes " + }, + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{(0:t - 1)}" + }, + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "inline_equation", + "content": "h_{\\phi_{\\mathrm{all}}}^{(t)}" + }, + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "text", + "content": " covers all the target classes " + }, + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{(0:t)}" + }, + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "text", + "content": ", where the old class part is initialized similarly to " + }, + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "inline_equation", + "content": "h_{\\phi_{\\mathrm{old}}}^{(t)}" + }, + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "text", + "content": ". Then, as depicted in Fig. 3, Tri-WE interpolates the classification head weight " + }, + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "inline_equation", + "content": "\\phi^n" + }, + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 366, + 546, + 524 + ], + "type": "text", + "content": "th class by" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 550, + 545, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 550, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 315, + 550, + 545, + 622 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\phi^ {n} = \\\\ \\left\\{ \\begin{array}{l l} \\bar {\\alpha} _ {1} \\phi_ {0} ^ {n} + \\bar {\\alpha} _ {2} \\phi_ {\\text {o l d}} ^ {n} + \\bar {\\alpha} _ {3} \\phi_ {\\text {a l l}} ^ {n} & \\text {i f} n \\leq N _ {0}, \\\\ \\bar {\\alpha} _ {4} \\phi_ {\\text {o l d}} ^ {n} + \\bar {\\alpha} _ {5} \\phi_ {\\text {a l l}} ^ {n} & \\text {i f} N _ {0} < n \\leq N _ {t - 1}, \\\\ \\phi_ {\\text {a l l}} ^ {n} & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {1} \\\\ \\end{array}", + "image_path": "caace501e1e3b864a75336982ebc7e2edf37cf35c789dc5854de6f395fb5e464.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "inline_equation", + "content": "1 \\leq n \\leq N_t" + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\phi^n \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "text", + "content": " is the dimensionality of " + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\phi^n" + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "text", + "content": ". Then, the weight of the classification head " + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "inline_equation", + "content": "h_\\phi^{(t)}" + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "text", + "content": " to classify " + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "inline_equation", + "content": "N_t" + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "text", + "content": " classes consists of " + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{\\phi^n\\}_{n=1}^{N_t}" + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "text", + "content": ". Notice that, we also add two learnable scalar weights " + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_1" + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_2" + }, + { + "bbox": [ + 304, + 649, + 545, + 713 + ], + "type": "text", + "content": " to automatically adjust the trade-offs between the three classification" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15331" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 523, + 295 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 523, + 295 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 523, + 295 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 523, + 295 + ], + "type": "image", + "image_path": "b5f2cc96519b842abc8f0f82a8207496d3862708e3cb9799d03e4056e431e080.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "lines": [ + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "text", + "content": "Figure 3. The pipeline of the proposed method. At the " + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "text", + "content": "th incremental session " + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "inline_equation", + "content": "S^{(t)}" + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "text", + "content": ", the model is initialized using the resulting model of session " + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "inline_equation", + "content": "S^{(t-1)}" + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "text", + "content": ". Then, a set of " + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "text", + "content": "-way " + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "text", + "content": "-shot few-shot examples " + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "inline_equation", + "content": "D^{(t)}" + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "text", + "content": " and the prototype buffer " + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "text", + "content": " are given. The weights " + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "inline_equation", + "content": "\\phi_0" + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{old}}" + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{all}}" + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "text", + "content": " of classification heads " + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "inline_equation", + "content": "h_{\\phi_0}" + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "inline_equation", + "content": "h_{\\phi_{\\mathrm{old}}}^{(t)}" + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "inline_equation", + "content": "h_{\\phi_{\\mathrm{all}}}^{(t)}" + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "text", + "content": " are interpolated with learnable scalars " + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "inline_equation", + "content": "\\alpha_1" + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "inline_equation", + "content": "\\alpha_2" + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "text", + "content": ". The main loss " + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Cls}}" + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "text", + "content": " is computed on the resulting classification head " + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "inline_equation", + "content": "h_{\\phi}^{(t)}" + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "text", + "content": "'s output. Also, classification loss over the old classes, appeared until the session " + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "inline_equation", + "content": "S^{(t-1)}" + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "text", + "content": ", is computed on the output of " + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "inline_equation", + "content": "h_{\\phi_{\\mathrm{old}}}^{(t)}" + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "text", + "content": ". Additionally, " + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "inline_equation", + "content": "D^{(t)}" + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "text", + "content": " is amplified to " + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{amp}}^{(t)}" + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "text", + "content": ", and then KD losses " + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{feat}}" + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{logit}}" + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "text", + "content": " are computed on the amplified dataset " + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{amp}}^{(t)}" + }, + { + "bbox": [ + 46, + 308, + 545, + 381 + ], + "type": "text", + "content": ". The learned components are outlined in red." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 402, + 249, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 402, + 249, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 402, + 249, + 415 + ], + "type": "text", + "content": "heads. Then, " + }, + { + "bbox": [ + 47, + 402, + 249, + 415 + ], + "type": "inline_equation", + "content": "\\{\\bar{\\alpha}_1,\\bar{\\alpha}_2,\\bar{\\alpha}_3,\\bar{\\alpha}_4,\\bar{\\alpha}_5\\}" + }, + { + "bbox": [ + 47, + 402, + 249, + 415 + ], + "type": "text", + "content": " are obtained by" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 420, + 219, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 420, + 219, + 433 + ], + "spans": [ + { + "bbox": [ + 113, + 420, + 219, + 433 + ], + "type": "interline_equation", + "content": "\\bar {\\alpha} _ {1} = \\alpha_ {1} / (\\alpha_ {1} + \\alpha_ {2} + 1),", + "image_path": "d9d9cf63d9a9bf82e6338cc95308fee7b9fcba819322f5ad338a8615f993786b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 435, + 219, + 448 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 435, + 219, + 448 + ], + "spans": [ + { + "bbox": [ + 115, + 435, + 219, + 448 + ], + "type": "interline_equation", + "content": "\\bar {\\alpha} _ {2} = \\alpha_ {2} / \\left(\\alpha_ {1} + \\alpha_ {2} + 1\\right),", + "image_path": "7fa8359249fa2ed93c092b737196dcc5022f044a88e94fb10e291a5bd36be411.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 450, + 213, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 450, + 213, + 462 + ], + "spans": [ + { + "bbox": [ + 115, + 450, + 213, + 462 + ], + "type": "interline_equation", + "content": "\\bar {\\alpha} _ {3} = 1 / \\left(\\alpha_ {1} + \\alpha_ {2} + 1\\right),", + "image_path": "9698c2ff458c53d9e551a603176b37a85b9ac379eb1d118e179edd18913427ed.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 464, + 196, + 477 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 464, + 196, + 477 + ], + "spans": [ + { + "bbox": [ + 115, + 464, + 196, + 477 + ], + "type": "interline_equation", + "content": "\\bar {\\alpha} _ {4} = \\alpha_ {2} / (\\alpha_ {2} + 1),", + "image_path": "869ac38ec1f982eb31e51cfaeb2a8d4f07544319469690257138af8eec37ab36.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 480, + 287, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 480, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 115, + 480, + 287, + 493 + ], + "type": "interline_equation", + "content": "\\bar {\\alpha} _ {5} = 1 / \\left(\\alpha_ {2} + 1\\right). \\tag {2}", + "image_path": "9e5530a037593f9e271ae9c5f6c6e5dbd76fd97d5ca4feee57dead6a14ca8820.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 498, + 226, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 498, + 226, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 226, + 510 + ], + "type": "text", + "content": "When " + }, + { + "bbox": [ + 47, + 498, + 226, + 510 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 47, + 498, + 226, + 510 + ], + "type": "text", + "content": " (i.e. " + }, + { + "bbox": [ + 47, + 498, + 226, + 510 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{old}} = \\phi_0" + }, + { + "bbox": [ + 47, + 498, + 226, + 510 + ], + "type": "text", + "content": "), we set " + }, + { + "bbox": [ + 47, + 498, + 226, + 510 + ], + "type": "inline_equation", + "content": "\\alpha_{2} = 0" + }, + { + "bbox": [ + 47, + 498, + 226, + 510 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 510, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 548 + ], + "type": "text", + "content": "Then, the main classification loss " + }, + { + "bbox": [ + 46, + 510, + 287, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Cls}}" + }, + { + "bbox": [ + 46, + 510, + 287, + 548 + ], + "type": "text", + "content": " is computed on the output from the weight-space ensembled classification head " + }, + { + "bbox": [ + 46, + 510, + 287, + 548 + ], + "type": "inline_equation", + "content": "h_{\\phi}^{(t)}" + }, + { + "bbox": [ + 46, + 510, + 287, + 548 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 555, + 287, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 287, + 590 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {C l s}} = \\mathbb {E} _ {(x, y) \\sim \\mathcal {D} ^ {(t)}} [ \\mathrm {C E} (h _ {\\phi} ^ {(t)} (g _ {\\theta} ^ {(t)} (x)), y ] \\\\ + \\mathbb {E} _ {(\\mathbf {p}, y) \\sim \\mathcal {M}} [ \\mathrm {C E} \\left(h _ {\\phi} ^ {(t)} (\\mathbf {p}), y \\right], \\tag {3} \\\\ \\end{array}", + "image_path": "1d647b58e21324187af41f7dc39e0c96c65a2187ef0a4a25337d7814f43f62c8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 594, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 287, + 617 + ], + "type": "text", + "content": "where CE is the cross-entropy loss function, and " + }, + { + "bbox": [ + 47, + 594, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 47, + 594, + 287, + 617 + ], + "type": "text", + "content": " denotes the old class prototype in " + }, + { + "bbox": [ + 47, + 594, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 47, + 594, + 287, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 619, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 619, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 619, + 287, + 647 + ], + "type": "text", + "content": "Using the buffer " + }, + { + "bbox": [ + 47, + 619, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 47, + 619, + 287, + 647 + ], + "type": "text", + "content": ", we also calculate the old class-specific loss " + }, + { + "bbox": [ + 47, + 619, + 287, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Cls - Old}}" + }, + { + "bbox": [ + 47, + 619, + 287, + 647 + ], + "type": "text", + "content": " for the classification head " + }, + { + "bbox": [ + 47, + 619, + 287, + 647 + ], + "type": "inline_equation", + "content": "h_{\\phi_{\\mathrm{old}}}^{(t)}" + }, + { + "bbox": [ + 47, + 619, + 287, + 647 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 89, + 651, + 287, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 651, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 89, + 651, + 287, + 668 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C l s - O l d}} = \\mathbb {E} _ {(\\mathbf {p}, y) \\sim \\mathcal {M}} [ \\mathrm {C E} \\left(h _ {\\phi_ {\\text {o l d}}} ^ {(t)} (\\mathbf {p}), y\\right) ]. \\tag {4}", + "image_path": "3341983ae7823b5e99675121969b140720e3fc3d3c1286e96aad11414475a5d2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 674, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 674, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 674, + 287, + 715 + ], + "type": "text", + "content": "In this loss, " + }, + { + "bbox": [ + 47, + 674, + 287, + 715 + ], + "type": "inline_equation", + "content": "h_{\\phi_{\\mathrm{old}}}^{(t)}" + }, + { + "bbox": [ + 47, + 674, + 287, + 715 + ], + "type": "text", + "content": " focuses on preserving the discriminative ability regarding all the old classes on top of the current feature extractor " + }, + { + "bbox": [ + 47, + 674, + 287, + 715 + ], + "type": "inline_equation", + "content": "g_{\\theta}^{(t)}" + }, + { + "bbox": [ + 47, + 674, + 287, + 715 + ], + "type": "text", + "content": ". Hence, we do not update " + }, + { + "bbox": [ + 47, + 674, + 287, + 715 + ], + "type": "inline_equation", + "content": "h_{\\phi_0}" + }, + { + "bbox": [ + 47, + 674, + 287, + 715 + ], + "type": "text", + "content": " and instead" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 402, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 402, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 305, + 402, + 545, + 426 + ], + "type": "text", + "content": "use it to retain the knowledge for the original, well-designed base class distribution from the base session model." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 429, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 429, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 545, + 491 + ], + "type": "text", + "content": "Unlike the naive ensemble approaches where multiple classification heads independently make predictions and combine them in the testing phase, our weight-space ensemble deploys only " + }, + { + "bbox": [ + 304, + 429, + 545, + 491 + ], + "type": "inline_equation", + "content": "h_{\\phi}^{(t)}" + }, + { + "bbox": [ + 304, + 429, + 545, + 491 + ], + "type": "text", + "content": " collaboratively utilizing the knowledge from the base and previous session models." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 506, + 412, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 506, + 412, + 520 + ], + "spans": [ + { + "bbox": [ + 306, + 506, + 412, + 520 + ], + "type": "text", + "content": "3.2. Total training loss" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 528, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 528, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 528, + 545, + 685 + ], + "type": "text", + "content": "KD has been a widely used regularization technique in CIL to prevent the catastrophic forgetting problem [27]. Specifically, when new classes are introduced, an additional distillation loss is combined with the main classification loss to ensure that the current model replicates the old class behavior of the previous model. Nevertheless, in FSCIL, the limited data source can lead to the extraction of knowledge biased toward those data, rather than generalized knowledge of the old classes. To mitigate this issue, we simply generate an amplified data source " + }, + { + "bbox": [ + 304, + 528, + 545, + 685 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{amp}}^{(t)}" + }, + { + "bbox": [ + 304, + 528, + 545, + 685 + ], + "type": "text", + "content": " from the few-shot training data by randomly mixing pairs of data from " + }, + { + "bbox": [ + 304, + 528, + 545, + 685 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{(t)}" + }, + { + "bbox": [ + 304, + 528, + 545, + 685 + ], + "type": "text", + "content": " employing simple data augmentation techniques such as [41, 43]. The ADKD loss is then computed as" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 377, + 701, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 701, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 377, + 701, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {A D K D}} = \\mathcal {L} _ {\\text {f e a t}} + \\mathcal {L} _ {\\text {l o g i t}}, \\tag {5}", + "image_path": "03b4d27479f8e80c6416e7b55a9581d336dd45e1d0e9211818bf6478d541fd60.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "15332" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{feat}}" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{logit}}" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": " are feature-level and logit-level KD losses which are defined by" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 104, + 287, + 123 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 104, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 65, + 104, + 287, + 123 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {f e a t}} = \\mathbb {E} _ {x \\sim \\mathcal {D} _ {\\mathrm {a m p}} ^ {(t)}} | | g _ {\\theta} ^ {(t - 1)} (x) - g _ {\\theta} ^ {(t)} (x) | | _ {2} \\tag {6}", + "image_path": "2f6db11a3a0955d76069e7b6a468e459eb582a9408258613ce1b3cf29f1a524c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 63, + 125, + 287, + 144 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 125, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 63, + 125, + 287, + 144 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {l o g i t}} = \\mathbb {E} _ {x \\sim \\mathcal {D} _ {\\mathrm {a m p}} ^ {(t)}} - f _ {\\theta , \\phi} ^ {(t - 1)} (x) \\log f _ {\\theta , \\phi^ {0: N _ {t - 1}}} ^ {(t)} (x). \\tag {7}", + "image_path": "d3f34c4d01ee4c2c4a72423e4984be2af863099266349b23994a9c2d37e30793.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 153, + 287, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 153, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 47, + 153, + 287, + 183 + ], + "type": "text", + "content": "The previous session model " + }, + { + "bbox": [ + 47, + 153, + 287, + 183 + ], + "type": "inline_equation", + "content": "f_{\\theta, \\phi}^{(t-1)}" + }, + { + "bbox": [ + 47, + 153, + 287, + 183 + ], + "type": "text", + "content": " including " + }, + { + "bbox": [ + 47, + 153, + 287, + 183 + ], + "type": "inline_equation", + "content": "h_{\\phi}^{(t-1)}" + }, + { + "bbox": [ + 47, + 153, + 287, + 183 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 153, + 287, + 183 + ], + "type": "inline_equation", + "content": "g_{\\theta}^{(t-1)}" + }, + { + "bbox": [ + 47, + 153, + 287, + 183 + ], + "type": "text", + "content": " is kept frozen during this process, unlike " + }, + { + "bbox": [ + 47, + 153, + 287, + 183 + ], + "type": "inline_equation", + "content": "h_{\\phi_{\\mathrm{old}}}^{(t)}" + }, + { + "bbox": [ + 47, + 153, + 287, + 183 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 184, + 286, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 184, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 184, + 286, + 205 + ], + "type": "text", + "content": "Finally, the total loss function used to optimize the current model is" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 94, + 217, + 287, + 229 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 217, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 94, + 217, + 287, + 229 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {C l s}} + \\gamma_ {1} \\mathcal {L} _ {\\mathrm {C l s - O l d}} + \\gamma_ {2} \\mathcal {L} _ {\\mathrm {A D K D}}, \\tag {8}", + "image_path": "a0cef579bd253a0b1618645af50cdff63f34a9a34203092e95bbdd5ebb07ba22.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 239, + 287, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 287, + 285 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 239, + 287, + 285 + ], + "type": "inline_equation", + "content": "\\gamma_{1}" + }, + { + "bbox": [ + 47, + 239, + 287, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 239, + 287, + 285 + ], + "type": "inline_equation", + "content": "\\gamma_{2}" + }, + { + "bbox": [ + 47, + 239, + 287, + 285 + ], + "type": "text", + "content": " are hyper-parameters that balance the respective losses, " + }, + { + "bbox": [ + 47, + 239, + 287, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Cls - Old}}" + }, + { + "bbox": [ + 47, + 239, + 287, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 239, + 287, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ADKD}}" + }, + { + "bbox": [ + 47, + 239, + 287, + 285 + ], + "type": "text", + "content": ". We provide the pseudo code of the overall process during an incremental session in the supplementary materials." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 293, + 167, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 167, + 306 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 167, + 306 + ], + "type": "text", + "content": "3.3. Base session training" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 312, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 287, + 574 + ], + "type": "text", + "content": "In FSCIL, it is broadly acknowledged that a model with strong generalization to base classes is beneficial for learning new classes [25, 29, 40, 42]. To ensure fairness when assessing the proposed components for incremental learning, we strive for base session performance on par with the prior SOTA methods [29, 40, 45]. For this purpose, we employ the base session training techniques of ALICE [25] (see the literature for details). On top of it, we also added a geometric classification head in parallel with the main classification head. In specific, an input is geometrically transformed into " + }, + { + "bbox": [ + 46, + 312, + 287, + 574 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 312, + 287, + 574 + ], + "type": "text", + "content": " types, and we use them as the proxy labels of transformed inputs. Then, the geometric classification head is learned to classify each into one of the " + }, + { + "bbox": [ + 46, + 312, + 287, + 574 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 312, + 287, + 574 + ], + "type": "text", + "content": " proxy labels. This auxiliary geometric classification helps the model encode the inherent structure of data so that the model's generalization capability on new tasks is improved. This auxiliary geometric classification strategy is not used in the incremental sessions. Through this, as aftermentioned in Sec. 4, the base session model achieves comparable or superior performance to previous SOTA methods on three benchmark datasets. A more detailed description is in the supplementary materials." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 586, + 173, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 173, + 599 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 173, + 599 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "Datasets. Consistent with the setting of [32], we conduct our experiments on three benchmark datasets: miniImageNet [33], CIFAR100 [20] and CUB200 [34]. miniImageNet is a subset of ImageNet [8] that includes 60,000 images from 100 classes, each with resolution " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "84 \\times 84" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": ". CIFAR100 comprises 60,000 tiny images of size " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " across 100 categories. CUB200, which is focused on a fine-grained classification, contains 200 bird species with similar appearance where the image size is " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": ". We follow" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 155 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 155 + ], + "type": "text", + "content": "the splits from [32]. For miniImageNet and CIFAR100, 60 categories are selected as base classes while the remaining are split into 8 incremental sessions with only 5 training examples per novel class (i.e., 5-way 5-shot). As for the CUB200 dataset, 100 categories are designated as the base training set, while the rest are divided into 10-way 5-shot tasks across 10 sessions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 156, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 545, + 204 + ], + "type": "text", + "content": "Evaluation protocol. In line with the FSCIL literature, we employ the top-1 accuracy (Acc) for all the observed classes at the end of each incremental learning session. We also provide the average of Acc over all the sessions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 205, + 546, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 546, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 546, + 396 + ], + "type": "text", + "content": "Implementation details. We implement our method using the PyTorch library. In keeping with the previous works [18, 25, 29, 45], ResNet18 [14] serves as our feature extractor. The entire model is trained via the SGD optimizer. For the base session, the learning rates are initialized by 0.01, 0.001, and 0.01 for miniImageNet, CUB200, and CIFAR100 respectively, and decayed by 0.1 at 60 and 70 epochs. In each incremental session, the learning rate for the weight-space ensembled " + }, + { + "bbox": [ + 304, + 205, + 546, + 396 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{all}}" + }, + { + "bbox": [ + 304, + 205, + 546, + 396 + ], + "type": "text", + "content": " is 0.1, while it is 0.001 for the rest. Hence, the feature extractor is minimally updated in incremental learning. For the ADKD loss, the given " + }, + { + "bbox": [ + 304, + 205, + 546, + 396 + ], + "type": "inline_equation", + "content": "NK" + }, + { + "bbox": [ + 304, + 205, + 546, + 396 + ], + "type": "text", + "content": " training examples are amplified to " + }, + { + "bbox": [ + 304, + 205, + 546, + 396 + ], + "type": "inline_equation", + "content": "16NK" + }, + { + "bbox": [ + 304, + 205, + 546, + 396 + ], + "type": "text", + "content": " for each class. Also, the learnable scalars " + }, + { + "bbox": [ + 304, + 205, + 546, + 396 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 304, + 205, + 546, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 205, + 546, + 396 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 304, + 205, + 546, + 396 + ], + "type": "text", + "content": " are initially set to 1.0. Loss weights " + }, + { + "bbox": [ + 304, + 205, + 546, + 396 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 304, + 205, + 546, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 205, + 546, + 396 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 304, + 205, + 546, + 396 + ], + "type": "text", + "content": " are empirically determined to be 1.2 and 10.0, respectively. Further details are in Sec. S-5 of the supplementary materials." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 407, + 443, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 443, + 418 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 443, + 418 + ], + "type": "text", + "content": "4.1. Comparative assessment" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "content": "On the three benchmark datasets, we first compare the proposed method with the existing approaches. As indicated in Table 1 on the miniImageNet dataset, our method outperforms the compared in all the sessions, achieving a clear SOTA in terms of the average Acc by a margin of " + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "inline_equation", + "content": "1.82\\%" + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "content": " at least. Also, BiDistill and SAVC require the base session model and a considerable number of classifiers per class, respectively, even in the testing phase. NC-FSCIL and OrCo benefit from knowing the total number of incremental classes at the outset of the base session. Whereas, we attain clear SOTA with a smaller-scale deployed model compared to BiDistill and SAVC, and without the information advantage of NC-FSCIL and OrCo." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "type": "text", + "content": "Moreover, Fig. 4 shows the Acc curve according to sessions on the CUB200 and CIFAR100 datasets. As in the miniImageNet dataset, our method consistently surpasses the compared methods by large gaps in the incremental sessions on the CUB 200 dataset. In specific, despite lower base session Acc than SAVC, our method outperforms it in the last session. Also, while NC-FSCIL merely finetuned the last module of the feature encoder in incremental sessions, OrCo attempted to also optimize the pre-set classifiers. Nevertheless, the additional update of only the classifiers is less effective. On the CIFAR100 dataset, we can" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15333" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 74, + 112, + 520, + 331 + ], + "blocks": [ + { + "bbox": [ + 46, + 71, + 547, + 104 + ], + "lines": [ + { + "bbox": [ + 46, + 71, + 547, + 104 + ], + "spans": [ + { + "bbox": [ + 46, + 71, + 547, + 104 + ], + "type": "text", + "content": "Table 1. Comparative results on miniImageNet dataset. Session-wise Acc (%) and the average of them (Avg) are reported. In each case, the best and second results are bold-faced and underlined, respectively. Last sess. Impro. denotes the performance improvement of our method at the last session." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 112, + 520, + 331 + ], + "lines": [ + { + "bbox": [ + 74, + 112, + 520, + 331 + ], + "spans": [ + { + "bbox": [ + 74, + 112, + 520, + 331 + ], + "type": "table", + "html": "
MethodVenueAcc in each session (%)AvgLast sess. impro.
012345678
iCaRL [27]CVPR 201761.3146.3242.9437.6330.4924.0020.8918.8017.2033.29+43.13
TOPIC [32]CVPR 202061.3150.0945.1741.1637.4835.5232.1929.4624.4239.64+35.71
ERL++ [11]AAAI 202161.7057.5854.6651.7248.6646.2744.6742.8140.7949.87+19.34
IDLVQ [4]ICLR 202064.7759.8755.9352.6249.8847.5544.8343.1441.8451.16+18.29
CEC [42]CVPR 202172.0066.8362.9759.4356.7053.7351.1949.2447.6357.74+12.50
SynthFeat [5]ICCV 202161.4059.8054.2051.6949.4548.0045.2043.8042.1050.63+18.03
MetaFSCIL [6]CVPR 202272.0467.9463.7760.2957.5855.1652.950.7949.1958.85+10.94
FACT [46]CVPR 202272.5669.6366.3862.7760.6057.3354.3452.1650.4960.70+9.64
Replay [23]ECCV 202271.8467.1263.2159.7757.0153.9551.5549.5248.2158.02+11.92
ALICE [25]ECCV 202280.6070.6067.4064.5062.5060.0057.8056.8055.7063.99+4.43
S3C [17]ECCV 202276.5571.7467.6664.5261.5158.0955.3653.0451.3462.20+8.79
WaRP [19]ICLR 202372.9968.1064.3161.3058.6456.0853.4051.7250.6559.69+9.48
SoftNet [18]ICLR 202376.6370.1365.9262.5259.4956.5653.7151.7250.4860.79+9.65
NC-FSCIL [40]ICLR 202384.0276.8072.0067.8366.3564.0461.4659.5458.3167.82+1.82
GKEAL [47]CVPR 202373.5968.9065.3362.2959.3956.7054.2052.5951.3160.48+10.40
BiDistill [45]CVPR 202374.6570.4366.2962.7760.7557.2454.7953.6552.2261.42+7.91
SAVC [29]CVPR 202381.1276.1472.4368.9266.4862.9559.9258.3957.1167.05+3.02
OrCo [1]CVPR 202483.3075.3271.5268.1665.6263.1160.2058.8258.0867.12+2.05
CLOSER [24]ECCV 202476.0271.6167.9964.6961.7058.9456.2354.5253.3362.78+6.80
Ours84.1381.4176.6573.5970.1065.1363.4261.0260.1370.62
", + "image_path": "ab3053c90cc09bb66d0de2d7a4c009aa9eafafbeb491eb79b4fbe385f263e244.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 133, + 344, + 461, + 365 + ], + "blocks": [ + { + "bbox": [ + 133, + 344, + 461, + 365 + ], + "lines": [ + { + "bbox": [ + 133, + 344, + 461, + 365 + ], + "spans": [ + { + "bbox": [ + 133, + 344, + 461, + 365 + ], + "type": "image", + "image_path": "be9d2f5bd94ff302ff07b19ff3c3c98a62f8cd9da1d84b65295a4d3614f017c9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 369, + 178, + 485 + ], + "blocks": [ + { + "bbox": [ + 59, + 369, + 178, + 485 + ], + "lines": [ + { + "bbox": [ + 59, + 369, + 178, + 485 + ], + "spans": [ + { + "bbox": [ + 59, + 369, + 178, + 485 + ], + "type": "image", + "image_path": "10d9b2ecc6b2a2673e4c4b064c54d65b6ce092849db4b10e3f511e6fb2912e96.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 490, + 198, + 499 + ], + "lines": [ + { + "bbox": [ + 156, + 490, + 198, + 499 + ], + "spans": [ + { + "bbox": [ + 156, + 490, + 198, + 499 + ], + "type": "text", + "content": "(a) CUB200" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 179, + 370, + 295, + 480 + ], + "blocks": [ + { + "bbox": [ + 179, + 370, + 295, + 480 + ], + "lines": [ + { + "bbox": [ + 179, + 370, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 179, + 370, + 295, + 480 + ], + "type": "image", + "image_path": "05a5fc2d0452e227a6f8c82e86b080a735b41a885e95e1c215847e1eed607491.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 509, + 547, + 544 + ], + "lines": [ + { + "bbox": [ + 46, + 509, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 547, + 544 + ], + "type": "text", + "content": "Figure 4. Comparative results on CUB200 and CIFAR100 datasets. For each dataset, the curve of the session-wise Acc (%) is first provided. And also, the base, last, average Acc are reported on our method comparing with three recent SOTA methods, NC-FSCIL [40], SAVC [29], YourSelf [31] (Best viewed in color)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 299, + 369, + 418, + 485 + ], + "blocks": [ + { + "bbox": [ + 299, + 369, + 418, + 485 + ], + "lines": [ + { + "bbox": [ + 299, + 369, + 418, + 485 + ], + "spans": [ + { + "bbox": [ + 299, + 369, + 418, + 485 + ], + "type": "image", + "image_path": "9e0824fa1e85b63a4eedcb5a36fa5ae07e2e1c4934c2e59b53a376a2b3d62d9c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 393, + 490, + 441, + 499 + ], + "lines": [ + { + "bbox": [ + 393, + 490, + 441, + 499 + ], + "spans": [ + { + "bbox": [ + 393, + 490, + 441, + 499 + ], + "type": "text", + "content": "(b) CIFAR100" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 418, + 370, + 535, + 479 + ], + "blocks": [ + { + "bbox": [ + 418, + 370, + 535, + 479 + ], + "lines": [ + { + "bbox": [ + 418, + 370, + 535, + 479 + ], + "spans": [ + { + "bbox": [ + 418, + 370, + 535, + 479 + ], + "type": "image", + "image_path": "6df7fd8303699a4f3d92d890b86dacfb9ab7ad71f4cc93eeaaf7e835c776fb61.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 563, + 289, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 289, + 708 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 289, + 708 + ], + "type": "text", + "content": "see the similar tendency. NC-FSCIL's setting is somewhat favorable, as it allows the pre-definition of classifiers for the incremental classes with knowledge of all class numbers in advance at the base session. This advantage is particularly beneficial on the CIFAR100 dataset, where some classes are semantically similar (e.g., possum and mouse), and the low resolution " + }, + { + "bbox": [ + 46, + 563, + 289, + 708 + ], + "type": "inline_equation", + "content": "(32 \\times 32)" + }, + { + "bbox": [ + 46, + 563, + 289, + 708 + ], + "type": "text", + "content": " of the images could otherwise complicate classification. The detailed results are available in Sec. S-4 of the supplementary materials. These findings demonstrate our method's superior ability to adapt to new classes with limited training data while maintaining seamless integration with the old classes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 563, + 410, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 410, + 576 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 410, + 576 + ], + "type": "text", + "content": "4.2. Ablation analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 582, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 605 + ], + "type": "text", + "content": "We extensively analyze the components of the proposed method on the miniImageNet or CUB200 datasets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Tripartite strategy. We conduct the ablation experiments on the Tri-WE in Table 2. Before this, initially, we contrast \"Naive\" with the \"No WE\" configuration, the latter incorporating the ADKD loss. The naive approach, lacking both WE and ADKD loss, exhibits a significant performance decline due to simultaneous updates of the feature extractor and classification head. However, No WE achieves respectable results thanks to the ADKD loss (more discussed in the following sections). Next, as demonstrated" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15334" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 102, + 535, + 196 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "type": "text", + "content": "Table 2. Ablation analysis for Tri-WE. Varying the combination of classification heads ensembled in weight space (No WE, Dual-WE, Tri-WE), session-wise and average Acc (%) are provided on miniImageNet. Also, \"Naive\" means that none of WE and ADKD is used." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 102, + 535, + 196 + ], + "lines": [ + { + "bbox": [ + 61, + 102, + 535, + 196 + ], + "spans": [ + { + "bbox": [ + 61, + 102, + 535, + 196 + ], + "type": "table", + "html": "
Classification head weightAcc in each session (%)Avg
φ0φoldφall012345678
Naive84.1380.5571.2660.9350.0544.0836.4219.7816.7351.55
No WE✓84.1378.9574.0970.0666.5663.0760.3258.0856.0167.92
Dual-WE✓✓84.1380.2575.4771.5867.6264.5261.1160.1758.9369.31
✓84.1379.5474.9970.1366.6063.4360.8559.8957.3168.54
Tri-WE✓✓✓84.1381.4176.6573.5970.165.1363.4261.0260.1370.62
", + "image_path": "05f438ad826f452223d11bce50fd7dee8f4c3dc81ad3bd408212423e704a2c56.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 99, + 236, + 496, + 330 + ], + "blocks": [ + { + "bbox": [ + 46, + 205, + 546, + 228 + ], + "lines": [ + { + "bbox": [ + 46, + 205, + 546, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 546, + 228 + ], + "type": "text", + "content": "Table 3. Impact on increasing the layers ensembled in weight space. Applying the proposed Tri-WE to the convolutional blocks of the feature extractor (ResNet18), session-wise and average Acc (%) are provided on miniImageNet." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 99, + 236, + 496, + 330 + ], + "lines": [ + { + "bbox": [ + 99, + 236, + 496, + 330 + ], + "spans": [ + { + "bbox": [ + 99, + 236, + 496, + 330 + ], + "type": "table", + "html": "
Weight-space ensembledAcc in each session (%)Avg
012345678
Blocks 1-484.1375.2570.2567.5464.8561.3459.7456.3853.4565.88
Blocks 1-4 + clf. head84.1376.0872.1368.4265.8362.6060.1957.7555.2966.94
Blocks 2-4 + clf. head84.1377.8473.4269.2265.8562.7560.5657.3455.3367.38
Blocks 3-4 + clf. head84.1378.5974.4369.8066.2563.1360.4658.0455.6567.83
Block 4 + clf. head84.1378.8074.2169.6466.0163.3261.5058.1755.8867.96
Clf. head only84.1381.4176.6573.5970.165.1363.4261.0260.1370.62
", + "image_path": "24730b7ef67aac33f7a18db9dc464fec0d9edea1a015ffa5350138cbc617c008.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 59, + 349, + 175, + 465 + ], + "blocks": [ + { + "bbox": [ + 59, + 349, + 175, + 465 + ], + "lines": [ + { + "bbox": [ + 59, + 349, + 175, + 465 + ], + "spans": [ + { + "bbox": [ + 59, + 349, + 175, + 465 + ], + "type": "image", + "image_path": "e492489135de153352ab97b8a4ab4066973d26ddffee44f146a4bdfea9c34d3a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 469, + 147, + 479 + ], + "lines": [ + { + "bbox": [ + 86, + 469, + 147, + 479 + ], + "spans": [ + { + "bbox": [ + 86, + 469, + 147, + 479 + ], + "type": "text", + "content": "(a) Fixed " + }, + { + "bbox": [ + 86, + 469, + 147, + 479 + ], + "type": "inline_equation", + "content": "(\\alpha_{1},\\alpha_{2})" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 489, + 288, + 534 + ], + "lines": [ + { + "bbox": [ + 46, + 489, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 489, + 288, + 534 + ], + "type": "text", + "content": "Figure 5. Analysis on learnable scalars " + }, + { + "bbox": [ + 46, + 489, + 288, + 534 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 46, + 489, + 288, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 489, + 288, + 534 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 46, + 489, + 288, + 534 + ], + "type": "text", + "content": " on miniImageNet. (a) Comparison with multiple combinations of fixed (not-learnable) " + }, + { + "bbox": [ + 46, + 489, + 288, + 534 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 46, + 489, + 288, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 489, + 288, + 534 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 46, + 489, + 288, + 534 + ], + "type": "text", + "content": " in terms of averaged Acc over all the sessions, (b) session-wise learned " + }, + { + "bbox": [ + 46, + 489, + 288, + 534 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 46, + 489, + 288, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 489, + 288, + 534 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 46, + 489, + 288, + 534 + ], + "type": "text", + "content": " in ours." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 179, + 349, + 277, + 465 + ], + "blocks": [ + { + "bbox": [ + 179, + 349, + 277, + 465 + ], + "lines": [ + { + "bbox": [ + 179, + 349, + 277, + 465 + ], + "spans": [ + { + "bbox": [ + 179, + 349, + 277, + 465 + ], + "type": "image", + "image_path": "d9c3609745a907048338b9539db934e28b6dbbbc2989d41a7c3dbc047c3e5cf2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 470, + 260, + 479 + ], + "lines": [ + { + "bbox": [ + 195, + 470, + 260, + 479 + ], + "spans": [ + { + "bbox": [ + 195, + 470, + 260, + 479 + ], + "type": "text", + "content": "(b) Learned " + }, + { + "bbox": [ + 195, + 470, + 260, + 479 + ], + "type": "inline_equation", + "content": "\\alpha_{1},\\alpha_{2}" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 556, + 287, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 699 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 699 + ], + "type": "text", + "content": "in Table 2, the absence of the weight-space ensemble for the classification head leads to increased catastrophic forgetting. And, from \"Dual-WE,\" we can see that the model from the intermediate previous session is more important than the base session model, as the reduced learning rate for the feature extractor mitigates drastic changes, helping to preserve the integrity of the feature space. Accordingly, the knowledge for incrementally-appeared classes is more helpful. Lastly, the complete Tri-WE clearly outperforms No WE by " + }, + { + "bbox": [ + 46, + 556, + 287, + 699 + ], + "type": "inline_equation", + "content": "2.70\\%" + }, + { + "bbox": [ + 46, + 556, + 287, + 699 + ], + "type": "text", + "content": " in average Acc, effectively retaining the essential knowledge of the base and previous session models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 701, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 288, + 712 + ], + "type": "text", + "content": "WE in classification head. We also evaluate the effect of" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 308, + 378, + 545, + 426 + ], + "blocks": [ + { + "bbox": [ + 306, + 347, + 545, + 370 + ], + "lines": [ + { + "bbox": [ + 306, + 347, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 306, + 347, + 545, + 370 + ], + "type": "text", + "content": "Table 4. Analysis on data amplification schemes. The last and average Acc (%) are reported on each scheme on miniImageNet." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 378, + 545, + 426 + ], + "lines": [ + { + "bbox": [ + 308, + 378, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 378, + 545, + 426 + ], + "type": "table", + "html": "
w/o ADAD scheme
CutOut [9]RandAug [7]MixUp [43]CutMix [41]
Last Acc58.0858.1258.2659.1460.13
Avg Acc69.0569.1369.0970.0370.62
", + "image_path": "157a066153e97062c77a965fab1ec9d4abfbcfeb68c7cc1f12964287ce12b536.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 449, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 604 + ], + "type": "text", + "content": "weight-space ensemble varying the number of interpolated layers in Table 3. In specific, we apply the proposed TriWE on the convolutional blocks of the feature extractors. Here, we use different learnable scalars depending on the convolutional blocks, rather than sharing them across the convolutional blocks and classification head. We observed a tendency for the decline in performance as the number of weight-space-ensemblied layers increases. This is because that deciding how to merge a lot of layers is difficult with only a few examples. Also, only merging feature extractor (1st row of Table 3) shows the lowest performance. It emphasizes that the head's weight-space is crucial to prevent catastrophic forgetting." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Learnable scalars " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{1} \\& \\alpha_{2}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ". Fig. 5(a) shows the averaged Acc on diverse combinations of " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "(\\alpha_{1}, \\alpha_{2})" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " where both are fixed after initialization in all the sessions. Compared to all the fixed cases, the learnable scalars show better result (70.62% in the averaged Acc). Also, as Fig. 5(b), we show the learned " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " in each session in our method where we initialize both as 1.0 in every session. We can see that the learned " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " are different depending on sessions. Accordingly, our learnable scalar approach is useful for bal" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15335" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 109, + 102, + 227, + 148 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "type": "text", + "content": "Table 5. Analysis on loss terms of " + }, + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "type": "inline_equation", + "content": "{\\mathcal{L}}_{\\mathrm{{ADKD}}}" + }, + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "type": "text", + "content": " . The last and average Acc (%) are reported abating the loss terms on miniImageNet." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 109, + 102, + 227, + 148 + ], + "lines": [ + { + "bbox": [ + 109, + 102, + 227, + 148 + ], + "spans": [ + { + "bbox": [ + 109, + 102, + 227, + 148 + ], + "type": "table", + "html": "
LfeatLlogitLast AccAvg Acc
✓59.4270.01
✓58.7768.41
✓✓60.1370.62
", + "image_path": "6da194ee50ebfca4b9ace4956292922da1c80391e830f60e24dfc073861a201f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 170, + 286, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 170, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 286, + 205 + ], + "type": "text", + "content": "ancing the trade-off among the three classification heads. We also provide session-wise results on the fixed cases in the supplementary materials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 207, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 287, + 495 + ], + "type": "text", + "content": "Effect of data amplification. While Tri-WE already attains SOTA (i.e. w/o AD in Table 4), incorporating our ADKD loss provides further increase (69.05% vs 70.62% in average Acc). Also, to see the influence of the choice of data amplification scheme in the ADKD loss, we compare multiple data augmentation methods for our data amplification. To facilitate the use of ADKD in follow-up works, we consider four widely-used data augmentation methods, CutOut [9], RandAug [7], CutMix [41], MixUp [43]. In the first two, the images are not inter-mixed. Whereas, two images are mixed for data augmentation. In Table 4, we can see that CutOut and RandAug are not effective compared to \"w/o AD\" where no data amplification is considered. With only a few examples, those methods cannot yield a diversity of data. Next, in comparison with MixUp, CutMix shows a better result. MixUp mixes two images smoothly, and then the original semantics are somewhat preserved. This is less useful in KD. Namely, CutMix encourages the model to focus more on general image recognition ability rather than high-level semantics, which is more helpful in preventing the catastrophic forgetting problem in FSCIL. Hence, we employ CutMix as the default option. In addition, the impact of the degree of data amplification is addressed in Sec. S-1 of the supplementary materials." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 495, + 286, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 286, + 567 + ], + "type": "text", + "content": "Two loss terms of " + }, + { + "bbox": [ + 46, + 495, + 286, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ADKD}}" + }, + { + "bbox": [ + 46, + 495, + 286, + 567 + ], + "type": "text", + "content": ". As shown in the first two rows of Table 5, we ablate " + }, + { + "bbox": [ + 46, + 495, + 286, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{feat}}" + }, + { + "bbox": [ + 46, + 495, + 286, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 495, + 286, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{logit}}" + }, + { + "bbox": [ + 46, + 495, + 286, + 567 + ], + "type": "text", + "content": ", respectively. From this study, we can see that " + }, + { + "bbox": [ + 46, + 495, + 286, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{feat}}" + }, + { + "bbox": [ + 46, + 495, + 286, + 567 + ], + "type": "text", + "content": " is slightly more important than " + }, + { + "bbox": [ + 46, + 495, + 286, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{logit}}" + }, + { + "bbox": [ + 46, + 495, + 286, + 567 + ], + "type": "text", + "content": " since Tri-WE already helps to prevent catastrophic forgetting in the side of the classification head. Then, using both is most effective." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 568, + 286, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 286, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 286, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Cls - Old}}" + }, + { + "bbox": [ + 46, + 568, + 286, + 676 + ], + "type": "text", + "content": ". As we update the feature extractor as well as the classification head in the incremental learning, we also adapt " + }, + { + "bbox": [ + 46, + 568, + 286, + 676 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{old}}" + }, + { + "bbox": [ + 46, + 568, + 286, + 676 + ], + "type": "text", + "content": " to the updated feature extractor using the loss " + }, + { + "bbox": [ + 46, + 568, + 286, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Cls - Old}}" + }, + { + "bbox": [ + 46, + 568, + 286, + 676 + ], + "type": "text", + "content": ". To verify this, we conduct ablating the feature extractor update and " + }, + { + "bbox": [ + 46, + 568, + 286, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Cls - Old}}" + }, + { + "bbox": [ + 46, + 568, + 286, + 676 + ], + "type": "text", + "content": ". As shown in Table 6, the performance is degraded without this loss term when the feature extractor is updated. Also, when the feature extractor is frozen, updating " + }, + { + "bbox": [ + 46, + 568, + 286, + 676 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{old}}" + }, + { + "bbox": [ + 46, + 568, + 286, + 676 + ], + "type": "text", + "content": " is less meaningful. Finally, updating the feature extractor with " + }, + { + "bbox": [ + 46, + 568, + 286, + 676 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{old}}" + }, + { + "bbox": [ + 46, + 568, + 286, + 676 + ], + "type": "text", + "content": " is most beneficial." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "New class adaptability. To see the efficacy of our method in terms of the trade-off between base and newly introduced classes, we provide the individual accuracy of the" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 360, + 102, + 491, + 156 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 6. Analysis on " + }, + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "inline_equation", + "content": "{\\mathcal{L}}_{\\text{Cls-Old }}" + }, + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "text", + "content": " and updating the feature extractor " + }, + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "inline_equation", + "content": "{g}_{\\theta }" + }, + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "text", + "content": " . The last and average Acc (%) are reported,on miniImageNet." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 360, + 102, + 491, + 156 + ], + "lines": [ + { + "bbox": [ + 360, + 102, + 491, + 156 + ], + "spans": [ + { + "bbox": [ + 360, + 102, + 491, + 156 + ], + "type": "table", + "html": "
LCls-OldUpdate gθLast AccAvg Acc
✓59.2770.15
✓✓60.1370.62
58.8669.41
✓58.8869.40
", + "image_path": "ac886ed8f1a56727db76043a51205099a6b83a972bf1e851cdfeb7403cad5d8c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 176, + 544, + 291 + ], + "blocks": [ + { + "bbox": [ + 308, + 176, + 544, + 291 + ], + "lines": [ + { + "bbox": [ + 308, + 176, + 544, + 291 + ], + "spans": [ + { + "bbox": [ + 308, + 176, + 544, + 291 + ], + "type": "image", + "image_path": "82e4ec3c0847b3fb628da1b5073e460b37cbff7168326273a7898dcff36d4f07.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 302, + 545, + 346 + ], + "lines": [ + { + "bbox": [ + 305, + 302, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 305, + 302, + 545, + 346 + ], + "type": "text", + "content": "Figure 6. Analysis on the new class adaptability on CUB200. The average of accuracy scores are provided for base and novel classes, denoted by Base and New respectively. And, harmonic means are also compared." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 354, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 354, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 305, + 354, + 545, + 437 + ], + "type": "text", + "content": "base and novel classes, and the harmonic mean (HM). Since most previous works have addressed the joint accuracy of the base and novel classes, we can compare our method with only a few recent methods [17, 33, 42, 45, 46] on the CUB200 dataset, following [17, 46]. Fig. 6 shows that our approach is more adept at integrating new classes without compromising the performance on the base classes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 449, + 383, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 449, + 383, + 460 + ], + "spans": [ + { + "bbox": [ + 306, + 449, + 383, + 460 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 468, + 545, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 468, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 305, + 468, + 545, + 659 + ], + "type": "text", + "content": "We explored FSCIL which facilitates the continuous acquisition of new concepts from a few examples. To this end, we suggested a novel FSCIL method including Tri-WE and ADKD regularization loss. The proposed method can effectively prevent overfitting to the few data and catastrophic forgetting of old concepts in the model. Tri-WE strategically interpolates the weights of the classification heads from the base, immediate previous, and current models, which collectively preserves the knowledge from the earlier models without overwhelming updates. Enriching the few examples for KD, the ADKD loss facilitates the transfer of essential knowledge from the previous model, rather than the knowledge overfitted to the few examples. We conducted extensive analyses on the proposed components and achieved state-of-the-art performance on the miniImageNet, CUB200, and CIFAR100 datasets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 671, + 365, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 365, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 365, + 684 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 310, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 310, + 691, + 545, + 713 + ], + "type": "text", + "content": "[1] Noor Ahmed, Anna Kukleva, and Bernt Schiele. Orco: Towards better generalization via orthogonality and contrast for" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15336" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 286, + 95 + ], + "type": "text", + "content": "few-shot class-incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2024." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 97, + 286, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 97, + 286, + 129 + ], + "spans": [ + { + "bbox": [ + 53, + 97, + 286, + 129 + ], + "type": "text", + "content": "[2] Eden Belouadah and Adrian Popescu. II2m: Class incremental learning with dual memory. In IEEE Conf. Comput. Vis. Pattern Recog., 2019." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 131, + 286, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 131, + 286, + 164 + ], + "spans": [ + { + "bbox": [ + 53, + 131, + 286, + 164 + ], + "type": "text", + "content": "[3] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. arXiv preprint arXiv:1812.00420, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 166, + 286, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 166, + 286, + 198 + ], + "spans": [ + { + "bbox": [ + 53, + 166, + 286, + 198 + ], + "type": "text", + "content": "[4] Kuilin Chen and Chi-Guhn Lee. Incremental few-shot learning via vector quantization in deep embedded space. In Int. Conf. Learn. Represent., 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 201, + 286, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 201, + 286, + 255 + ], + "spans": [ + { + "bbox": [ + 53, + 201, + 286, + 255 + ], + "type": "text", + "content": "[5] Ali Cheraghian, Shafin Rahman, Sameera Ramasinghe, Pengfei Fang, Christian Simon, Lars Petersson, and Mehrtash Harandi. Synthesized feature based few-shot class-incremental learning on a mixture of subspaces. In Int. Conf. Comput. Vis., 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 257, + 286, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 257, + 286, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 257, + 286, + 300 + ], + "type": "text", + "content": "[6] Zhixiang Chi, Li Gu, Huan Liu, Yang Wang, Yuanhao Yu, and Jin Tang. Meta-scil: A meta-learning approach for few-shot class incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 302, + 286, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 286, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 286, + 345 + ], + "type": "text", + "content": "[7] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In IEEE Conf. Comput. Vis. Pattern Recog. Worksh., 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 348, + 286, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 348, + 286, + 380 + ], + "spans": [ + { + "bbox": [ + 53, + 348, + 286, + 380 + ], + "type": "text", + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conf. Comput. Vis. Pattern Recog., 2009." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 382, + 286, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 382, + 286, + 415 + ], + "spans": [ + { + "bbox": [ + 53, + 382, + 286, + 415 + ], + "type": "text", + "content": "[9] Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 417, + 286, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 417, + 286, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 417, + 286, + 449 + ], + "type": "text", + "content": "[10] Thomas G Dietterich. Ensemble methods in machine learning. In International Workshop on Multiple Classifier Systems, 2000." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 452, + 286, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 452, + 286, + 494 + ], + "spans": [ + { + "bbox": [ + 48, + 452, + 286, + 494 + ], + "type": "text", + "content": "[11] Songlin Dong, Xiaopeng Hong, Xiaoyu Tao, Xinyuan Chang, Xing Wei, and Yihong Gong. Few-shot class incremental learning via relation knowledge distillation. In AAAI, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 497, + 286, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 497, + 286, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 497, + 286, + 540 + ], + "type": "text", + "content": "[12] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In Eur. Conf. Comput. Vis., 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 543, + 286, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 286, + 574 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 286, + 574 + ], + "type": "text", + "content": "[13] Lars Kai Hansen and Peter Salamon. Neural network ensembles. IEEE Trans. Pattern Anal. Mach. Intell., 12(10): 993-1001, 1990." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 577, + 286, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 577, + 286, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 577, + 286, + 609 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog., 2016." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 612, + 286, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 286, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 286, + 644 + ], + "type": "text", + "content": "[15] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In IEEE Conf. Comput. Vis. Pattern Recog., 2019." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 646, + 286, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 286, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 286, + 678 + ], + "type": "text", + "content": "[16] Gao Huang, Yixuan Li, Geoff Pleiss, Zhuang Liu, John E Hopcroft, and Kilian Q Weinberger. Snapshot ensembles: Train 1, get m for free. In Int. Conf. Learn. Represent. 2017." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "text", + "content": "[17] Jayateja Kalla and Soma Biswas. S3c: Self-supervised stochastic classifiers for few-shot class-incremental learning. In Eur. Conf. Comput. Vis., 2022." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[18] Haeyong Kang, Jaehong Yoon, Sultan Rizky Hikmawan Madjid, Sung Ju Hwang, and Chang D Yoo. On the soft-subnetwork for few-shot class incremental learning. In Int. Conf. Learn. Represent., 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[19] Do-Yeon Kim, Dong-Jun Han, Jun Seo, and Jaekyun Moon. Warping the space: Weight space rotation for class incremental few-shot learning. In Int. Conf. Learn. Represent., 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 163, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 185 + ], + "type": "text", + "content": "[20] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 186, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 217 + ], + "type": "text", + "content": "[21] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Adv. Neural Inform. Process. Syst. 2012." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 220, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 251 + ], + "type": "text", + "content": "[22] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Trans. Pattern Anal. Mach. Intell., 40(12):2935-2947, 2017." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "type": "text", + "content": "[23] Huan Liu, Li Gu, Zhixiang Chi, Yang Wang, Yuanhao Yu, Jun Chen, and Jin Tang. Few-shot class-incremental learning via entropy-regularized data-free replay. In Eur. Conf. Comput. Vis., 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 299, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 299, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 545, + 331 + ], + "type": "text", + "content": "[24] Junghun Oh, Sungyong Baik, and Kyoung Mu Lee. Closer: Towards better representation learning for few-shot class-incremental learning. In Eur. Conf. Comput. Vis., 2024." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 333, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 333, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 333, + 545, + 365 + ], + "type": "text", + "content": "[25] Can Peng, Kun Zhao, Tianren Wang, Meng Li, and Brian C Lovell. Few-shot class-incremental learning from an open-set perspective. In Eur. Conf. Comput. Vis., 2022." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 367, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 367, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 367, + 545, + 421 + ], + "type": "text", + "content": "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Int. Conf. Machine Learning, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 422, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 465 + ], + "type": "text", + "content": "[27] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2017." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 468, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 545, + 499 + ], + "type": "text", + "content": "[28] David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. Experience replay for continual learning. 2019." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 502, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 502, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 502, + 545, + 544 + ], + "type": "text", + "content": "[29] Zeyin Song, Yifan Zhao, Yujun Shi, Peixi Peng, Li Yuan, and Yonghong Tian. Learning with fantasy: Semantic-aware virtual contrastive constraint for few-shot class-incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2023." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 546, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 546, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 546, + 545, + 599 + ], + "type": "text", + "content": "[30] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In IEEE Conf. Comput. Vis. Pattern Recog. 2015." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "type": "text", + "content": "[31] Yu-Ming Tang, Yi-Xing Peng, Jingke Meng, and Wei-Shi Zheng. Rethinking few-shot class-incremental learning: Learning from yourself. In Eur. Conf. Comput. Vis., 2024." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 636, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 545, + 678 + ], + "type": "text", + "content": "[32] Xiaoyu Tao, Xiaopeng Hong, Xinyuan Chang, Songlin Dong, Xing Wei, and Yihong Gong. Few-shot class-incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2020." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "text", + "content": "[33] Oriol Vinyals, Charles Blundell, Tim Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In NeurIPS, 2016." + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15337" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 711 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 105 + ], + "type": "text", + "content": "[34] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 150 + ], + "type": "text", + "content": "[35] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 286, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 286, + 217 + ], + "type": "text", + "content": "[36] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Marcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In Int. Conf. Machine Learning, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 219, + 286, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 219, + 286, + 283 + ], + "spans": [ + { + "bbox": [ + 49, + 219, + 286, + 283 + ], + "type": "text", + "content": "[37] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, and Ludwig Schmidt. Robust fine-tuning of zero-shot models. In IEEE Conf. Comput. Vis. Pattern Recog., 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 285, + 286, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 285, + 286, + 327 + ], + "spans": [ + { + "bbox": [ + 49, + 285, + 286, + 327 + ], + "type": "text", + "content": "[38] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 331, + 286, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 331, + 286, + 363 + ], + "spans": [ + { + "bbox": [ + 49, + 331, + 286, + 363 + ], + "type": "text", + "content": "[39] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 365, + 286, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 365, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 49, + 365, + 286, + 407 + ], + "type": "text", + "content": "[40] Yibo Yang, Haobo Yuan, Xiangtai Li, Zhouchen Lin, Philip Torr, and Dacheng Tao. Neural collapse inspired feature-classifier alignment for few-shot class incremental learning. In Int. Conf. Learn. Represent., 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 409, + 286, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 409, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 49, + 409, + 286, + 453 + ], + "type": "text", + "content": "[41] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Int. Conf. Comput. Vis., 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 454, + 286, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 454, + 286, + 496 + ], + "spans": [ + { + "bbox": [ + 49, + 454, + 286, + 496 + ], + "type": "text", + "content": "[42] Chi Zhang, Nan Song, Guosheng Lin, Yun Zheng, Pan Pan, and Yinghui Xu. Few-shot incremental learning with continually evolved classifiers. In IEEE Conf. Comput. Vis. Pattern Recog., 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 498, + 286, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 498, + 286, + 531 + ], + "spans": [ + { + "bbox": [ + 49, + 498, + 286, + 531 + ], + "type": "text", + "content": "[43] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 533, + 286, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 533, + 286, + 575 + ], + "spans": [ + { + "bbox": [ + 49, + 533, + 286, + 575 + ], + "type": "text", + "content": "[44] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu-Tao Xia. Maintaining discrimination and fairness in class incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 578, + 286, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 578, + 286, + 620 + ], + "spans": [ + { + "bbox": [ + 49, + 578, + 286, + 620 + ], + "type": "text", + "content": "[45] Linglan Zhao, Jing Lu, Yunlu Xu, Zhanzhan Cheng, Dashan Guo, Yi Niu, and Xiangzhong Fang. Few-shot class-incremental learning via class-aware bilateral distillation. In IEEE Conf. Comput. Vis. Pattern Recog., 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 622, + 286, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 622, + 286, + 665 + ], + "spans": [ + { + "bbox": [ + 49, + 622, + 286, + 665 + ], + "type": "text", + "content": "[46] Da-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, Liang Ma, Shiliang Pu, and De-Chuan Zhan. Forward compatible few-shot class-incremental learning. In IEEE Conf. Comput. Vis. Pattern Recog., 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 667, + 286, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 667, + 286, + 711 + ], + "spans": [ + { + "bbox": [ + 49, + 667, + 286, + 711 + ], + "type": "text", + "content": "[47] Huiping Zhuang, Zhenyu Weng, Run He, Zhiping Lin, and Ziqian Zeng. Gkeal: Gaussian kernel embedded analytic learning for few-shot class incremental task. In IEEE Conf. Comput. Vis. Pattern Recog., 2023." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15338" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Tuning the Frequencies_ Robust Training for Sinusoidal Neural Networks/79e8b6d6-50a7-4215-bd2a-fb5524e17539_content_list.json b/2025/Tuning the Frequencies_ Robust Training for Sinusoidal Neural Networks/79e8b6d6-50a7-4215-bd2a-fb5524e17539_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..796f7afba053eeb46f09522c8b81a6e39bff7da4 --- /dev/null +++ b/2025/Tuning the Frequencies_ Robust Training for Sinusoidal Neural Networks/79e8b6d6-50a7-4215-bd2a-fb5524e17539_content_list.json @@ -0,0 +1,1715 @@ +[ + { + "type": "text", + "text": "Tuning the Frequencies: Robust Training for Sinusoidal Neural Networks", + "text_level": 1, + "bbox": [ + 127, + 130, + 870, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tiago Novello $^{1,2,*}$ Diana Aldana $^{1,*}$ Andre Araujo $^{2}$ Luiz Velho $^{1}$ $^{1}$ IMPA ${}^{2}$ Google DeepMind", + "bbox": [ + 264, + 181, + 730, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 258, + 326, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sinusoidal neural networks have been shown effective as implicit neural representations (INRs) of low-dimensional signals, due to their smoothness and high representation capacity. However, initializing and training them remain empirical tasks which lack on deeper understanding to guide the learning process. To fill this gap, our work introduces a theoretical framework that explains the capacity property of sinusoidal networks and offers robust control mechanisms for initialization and training. Our analysis is based on a novel amplitude-phase expansion of the sinusoidal multilayer perceptron, showing how its layer compositions produce a large number of new frequencies expressed as integer combinations of the input frequencies. This relationship can be directly used to initialize the input neurons, as a form of spectral sampling, and to bound the network's spectrum while training. Our method, referred to as TUNER (TUNing sinusoidal nEtwoRks), greatly improves the stability and convergence of sinusoidal INR training, leading to detailed reconstructions, while preventing overfitting.", + "bbox": [ + 88, + 287, + 485, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 585, + 220, + 601 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sinusoidal multilayer perceptrons (MLPs) emerged as powerful implicit neural representations (INRs) for low-dimensional signals [3, 11, 26, 35]. In this context, the INR $f$ should fit the input data $\\{x_{i},\\ell_{i}\\}$ as close as possible, i.e. $f(x_{i})\\approx \\ell_{i}$ , without overfitting, thus encoding the signal implicitly in the MLP parameters. Therefore, two major properties are required: (1) $f$ needs high representation capacity to fit $\\{x_i,\\ell_i\\}$ ; (2) $f$ should have bandlimit control to avoid frequencies bypassing the sampling rate.", + "bbox": [ + 89, + 604, + 483, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Training sinusoidal MLPs to satisfy the above properties is challenging, as their initialization and optimization process often lead to undesirable local minima [15]. Recent work made strides towards more effective learning of these models. For example, SIREN [26] proposed an initialization by projecting the input coordinates to a list of sines with frequencies randomly chosen in a range, similar to the Fourier feature mapping (FFM) approach [28]. This way, the model can reach high capacity, but may lead to over", + "bbox": [ + 89, + 739, + 483, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1e10ba604bbe333e12863f62e01cd1ab65339e0bd38c53310ba84272e6aaa4d2.jpg", + "image_caption": [ + "Figure 1. We present TUNER, a robust and theoretically grounded training technique for sinusoidal MLPs, overcoming challenges in initialization and enabling bandlimiting control. Our experiments showcase TUNER's strong initialization results against ReLU, FFM [28], and SIREN [26] (top), where all models use the same size and training conditions. TUNER achieves both fast and stable convergence (bottom-left) while reconstructing gradients without noise. We also compare with BACON [10] across two bandlimits (bottom-right), enhancing quality and avoiding ringing artifacts." + ], + "image_footnote": [], + "bbox": [ + 519, + 257, + 901, + 621 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "fitting with high frequencies resulting in noisy reconstructions. Defining an effective range for the bandlimit initialization remains mostly empirical and often results in noise, as the role of layer composition in generating frequencies is not fully understood. Additionally, uniform initialization may introduce undesired high frequencies and make it harder to model lower ones. More recently, BACON [10] proposes tighter bandlimit control by applying multiplicative filter networks (MFNs) [7] to limit the signal spectrum", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally to this work.", + "bbox": [ + 107, + 886, + 364, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3071", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "with a box filter, hard-truncating the spectrum. While effective in many cases, this may lead to ringing artifacts. Also, this technique lacks non-linear activations (MFNs are not neural networks), preventing the representation of fine details as efficiently as sinusoidal MLPs with similar sizes.", + "bbox": [ + 89, + 90, + 480, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, initializing and training sinusoidal MLPs remain empirical tasks which lack on deeper understanding to guide the learning. To approach these problems, we present TUNER (TUNing sinusoidal nEtwoRks), a training scheme for sinusoidal MLPs consisting of a robust initialization and bandlimit mechanism that greatly improves the capacity and convergence of INRs. In contrast to previous work, TUNER is grounded on a theoretical framework that guides our learning approach. We develop a new trigonometric identity (Thrm 1), resembling a Fourier series, and prove that such expansion produces a large number of frequencies in terms of the first layer (input frequencies). This explains the frequency generation process when composing layers, highlighting the importance of input layer initialization for improving capacity. Remarkably, the amplitudes of these frequencies are given by complex functions of the hidden weights, introducing a challenging training. To address this, we prove that those are bounded by a term depending only on the hidden weights (Thrm 2). We apply this result to control the bandlimit of the INR during training. Some of TUNER's results are showcased in Fig. 1, in comparison to previous work, illustrating substantial enhancements to sinusoidal INR modeling. In summary, our contributions are:", + "bbox": [ + 91, + 167, + 482, + 513 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel trigonometric identity that expands any hidden neuron into a wide sum of sines with frequencies being integer combinations of the input frequencies (Fig 2). The corresponding amplitudes, determined by the hidden weights, have a useful upper bound, offering effective tools for controlling the resulting signal.", + "- The expansion enables a robust initialization scheme for sinusoidal INRs, resulting in a high capacity MLP that trains significantly faster than previous approaches. First, we initialize the input neurons, as they determine the INR spectrum. Next, we initialize the hidden weights, considering their role in setting the corresponding amplitudes.", + "- We leverage the amplitude's upper bound to control the bandlimit of a sinusoidal INR by designing schemes that bound the hidden weights during training. Together with our initialization, these bounds generate frequencies centered around the input, resulting in more stable training compared to previous approaches." + ], + "bbox": [ + 91, + 513, + 480, + 785 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 89, + 799, + 235, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Implicit neural representations [8, 20, 21, 25, 30, 32] are a trending topic in machine learning, used to learn highly detailed signals in low-dimension domains. Current INR architectures use Fourier feature mappings [28] or sinusoidal activation functions [26] to bypass the spectral bias [17]", + "bbox": [ + 89, + 824, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "common in ReLU multilayer perceptrons. The high representation capacity of sinusoidal INRs has motivated their use to represent a wide range of media objects. Examples include audio [6, 26], images [4, 27], face morphing [22], signed distance functions [10, 13, 23, 26, 31], displacement fields [33], surface animation [12, 14], multi-resolution signals [5, 10, 16, 19, 29], among others. Most of these exploit the derivatives of sinusoidal INRs in the loss functions.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Initialization. Considering sinusoidal activation functions in neural networks is a classical problem [18]; however, these INRs have been regarded as difficult to train [15]. Sitzmann et al. [26] overcome this by presenting a specific initialization scheme that allows training sinusoidal INRs, avoiding instability and ensuring convergence. Despite these advances, in practice, the initialization of such INRs remains an empirical task. In this work, inspired by Fourier series theory, we present a novel initialization method that allows us to train INRs with great convergence. Recently, several works have addressed the representation problem of sinusoidal INRs from different perspectives. Zell et al. [35] approached this problem by observing that the first layer of a sinusoidal INR is similar to a Fourier feature mapping. Here, in addition to improving the initialization of sinusoidal INRs, we present a training scheme for bounding the spectra of these networks.", + "bbox": [ + 511, + 212, + 903, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Control of spectrum. One of the main drawbacks of sinusoidal INRs is the lack of frequency control. SIREN [26] addressed this by initializing the input frequencies uniformly in a range. While this ensures a bandlimit at the start of training, as it progresses, the layer composition introduces higher frequencies, leading to noisy reconstructions. We avoid this by providing a novel initialization for the first layer coupled with a bounding scheme which gives controls for limiting the MLP spectrum.", + "bbox": [ + 511, + 468, + 903, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Other works [5, 10] employed MFNs [7] to control the bandlimit by applying a filter on the spectrum. However, this strategy usually introduces reconstruction artifacts since MFNs hard-truncate the spectrum. BANF [24] employed a grid-based MLP with a spatial filter that leverages grid resolution to constrain the highest frequency in the network. In our experiments, we observe that this is prone to creating higher frequencies which propagate as artifacts. In contrast, TUNER, grounded in a theoretical result (Thrm 2), guarantees a bandlimited MLP, and serves as a soft filter, providing a representation without ringing artifacts.", + "bbox": [ + 511, + 604, + 903, + 771 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Sinusoidal MLPs as Fourier series", + "text_level": 1, + "bbox": [ + 511, + 784, + 823, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This work addresses the problem of deriving an efficient training scheme with controlled bandlimit for sinusoidal MLPs. This section presents the mathematical definitions and novel formulas for approaching this task.", + "bbox": [ + 511, + 809, + 903, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sinusoidal MLPs demonstrated high representational capacity with only a few hidden layers [19, 26, 35]. To un", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3072", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "derstand how layer composition encodes this capacity, we propose to thoroughly investigate the structure of a 3-layer sinusoidal MLP $f: \\mathbb{R}^d \\to \\mathbb{R}$ . For simplicity, we assume the codomain to be $\\mathbb{R}$ , however the same analysis holds for dimension $> 1$ . More precisely, we consider $f(\\mathbf{x}) = \\mathbf{C} \\circ \\mathbf{S} \\circ \\mathbf{D}(\\mathbf{x}) + e$ , with $\\mathbf{D}(\\mathbf{x}) = \\sin (\\omega \\mathbf{x} + \\varphi)$ being the input layer that projects $\\mathbf{x}$ into a list of harmonics (input neurons) $D_{i}(\\mathbf{x}) = \\sin (\\omega_{i}\\mathbf{x} + \\varphi_{i})$ with frequencies $\\omega = (\\omega_{1},\\dots,\\omega_{m}) \\in \\mathbb{R}^{m\\times d}$ and shifts $\\varphi \\in \\mathbb{R}^{m}$ . Layer $\\mathbf{D}$ is then composed with $\\mathbf{S}(\\mathbf{x}) = \\sin (\\mathbf{W}\\mathbf{x} + \\mathbf{b})$ , where $\\mathbf{W} \\in \\mathbb{R}^{n\\times m}$ is the hidden matrix, and $\\mathbf{b} \\in \\mathbb{R}^n$ the bias. Finally, $\\mathbf{C} \\cdot \\mathbf{x} + e$ is an affine transformation with $\\mathbf{C} \\in \\mathbb{R}^{n\\times 1}$ and $e \\in \\mathbb{R}$ .", + "bbox": [ + 89, + 90, + 483, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now present some properties of the sinusoidal layers, which play key roles in the representation, and give a reinterpretation of them in terms of the network parameters. First, note that the hidden neurons $\\mathbf{h}(\\mathbf{x})\\coloneqq \\mathbf{S}\\circ \\mathbf{D}(\\mathbf{x})$ are defined by composing the dictionary $\\mathbf{D}$ with the hidden sinusoidal layer which results in a list with elements", + "bbox": [ + 89, + 272, + 483, + 362 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh _ {i} (\\mathbf {x}) = \\sin \\left(\\sum_ {j = 1} ^ {m} W _ {i j} \\sin \\left(\\omega_ {j} \\mathbf {x} + \\varphi_ {j}\\right) + b _ {i}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 363, + 482, + 401 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The following is a key result of this work, which states that we can linearize a hidden neuron (1) as a sum of sines with frequencies and amplitudes determined by $\\omega$ and $\\mathbf{W}$ .", + "bbox": [ + 89, + 400, + 483, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theorem 1. Each hidden neuron $h_i$ of a 3-layer sinusoidal MLP has an amplitude-phase expansion of the form", + "bbox": [ + 89, + 453, + 483, + 484 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh _ {i} (\\mathbf {x}) = \\sum_ {\\mathbf {k} \\in \\mathbb {Z} ^ {m}} \\alpha_ {\\mathbf {k}} \\sin \\left(\\beta_ {\\mathbf {k}} \\mathbf {x} + \\lambda_ {\\mathbf {k}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 487, + 482, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\beta_{\\mathbf{k}} = \\langle \\mathbf{k},\\omega \\rangle$ $\\lambda_{\\mathbf{k}} = \\langle \\mathbf{k},\\varphi \\rangle +b_{i}$ , and $\\alpha_{\\mathbf{k}} = \\prod_{j}J_{k_j}(W_{ij})$ is the product of the Bessel functions of the first kind.", + "bbox": [ + 89, + 518, + 482, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Bessel functions $J_{k}$ appear in the Fourier series of $\\sin (a\\sin (x))$ [2, Page 361] and Thrm 1 generalizes this result. See the proof and generalization in the Supp. Mat. Yüce et al. [34] present a similar expansion for neurons activated by polynomials, however, they do not derive an amplitude formula. Their expansion for sinusoidal neurons is restricted to a 3-layer MLP without bias, a critical parameter for Fourier representation. Ours incorporates bias and generalizes to deep MLPs.", + "bbox": [ + 89, + 556, + 483, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now list some consequences of (2). First, the layer composition introduces a vast number of frequencies $\\beta_{\\mathbf{k}}$ depending only on $\\omega$ , and shifts given by the input shift $\\varphi$ and the bias $\\mathbf{b}$ . More precisely, truncating the expansion by summing over $|\\mathbf{k}|_{\\infty} \\leq B \\in \\mathbb{N}$ , implies that each hidden neuron $h_i$ could learn $\\frac{(2B + 1)^m - 1}{2}$ non-null frequencies. This frequency generation gives a novel explanation of why composing sinusoidal layers greatly increases the network capacity. We will explore (2) to define a robust initialization for the INR's input neurons in Sec 4.1.", + "bbox": [ + 89, + 693, + 483, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Secondly, note that the weights $\\mathbf{W}$ fully determine the amplitude $\\alpha_{\\mathbf{k}}$ of each harmonic $\\sin (\\beta_{\\mathbf{k}}\\mathbf{x} + \\lambda_{\\mathbf{k}})$ . Thus, to", + "bbox": [ + 89, + 845, + 483, + 878 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "derive our bounding scheme we need to focus only on $\\mathbf{W}$ . Finally, we can derive a sine-cosine form of (2),", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh _ {i} (\\mathbf {x}) = \\sum_ {\\mathbf {k} \\in \\mathbb {Z} ^ {m}} A _ {\\mathbf {k}} \\sin \\left(\\beta_ {\\mathbf {k}} \\mathbf {x}\\right) + B _ {\\mathbf {k}} \\cos \\left(\\beta_ {\\mathbf {k}} \\mathbf {x}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 128, + 903, + 162 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with $A_{\\mathbf{k}} = \\alpha_{\\mathbf{k}}\\cos (\\lambda_{\\mathbf{k}})$ and $B_{\\mathbf{k}} = \\alpha_{\\mathbf{k}}\\sin (\\lambda_{\\mathbf{k}})$ . Note that the generated frequencies are independent of the index $i$ , thus the hidden neurons share the same harmonics: $\\sin (\\beta_{\\mathbf{k}}\\mathbf{x})$ , $\\cos (\\beta_{\\mathbf{k}}\\mathbf{x})$ . Since different combinations of the input frequencies may correspond to a single frequency, (3) isn't (yet) the Fourier transform of the network. In other words, we could have two integer vectors $\\mathbf{k},\\mathbf{l}\\in \\mathbb{Z}^{m}$ such that $\\beta_{\\mathbf{k}} = \\beta_{\\mathbf{l}}$ . In the Supp. Mat., we show how to aggregate those frequencies to obtain the final Fourier transform.", + "bbox": [ + 511, + 170, + 906, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To control the MLP's bandlimit, we need a method to bound the amplitudes of $\\beta_{\\mathbf{k}}$ . Precisely, for a given number $B > 0$ we would need $|\\alpha_{\\mathbf{k}}|$ to be small for $|\\beta_{\\mathbf{k}}| \\geq B$ . For this, we use the formula of $\\alpha_{\\mathbf{k}}$ to determine a bounding:", + "bbox": [ + 511, + 308, + 906, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theorem 2. The magnitude of the amplitudes $\\alpha_{\\mathbf{k}}$ in the expansion (2) is bounded by $\\prod_{j=1}^{m}\\left(\\frac{|W_{ij}|}{2}\\right)^{|k_j|}\\frac{1}{|k_j|!}$ .", + "bbox": [ + 511, + 376, + 905, + 420 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sec 4.2 presents a bandlimit control mechanism during training based on Thrm 2.", + "bbox": [ + 511, + 424, + 905, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Initialization and frequency bounding", + "text_level": 1, + "bbox": [ + 511, + 467, + 856, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present TUNER, an initialization and frequency control scheme for sinusoidal INRs. Our initialization considers integer input frequencies, based on a Fourier series interpretation, and uses Thrm 1 to study the spectrum of layer composition. Then, we use Thrm 2 to control the INR's bandlimit during training. Fig 2 provides an overview of TUNER. Throughout this section we present toy experiments to motivate the method, using $f: \\mathbb{R}^2 \\to \\mathbb{R}^3$ in an image reconstruction setup; later, comprehensive experiments will be given in Sec 5. We report the MLP size using the parameters $m$ , $n$ as they completely determine the model. We use the Kodak dataset's images [1] and train the INR using Adam [9].", + "bbox": [ + 511, + 492, + 906, + 675 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Initialization as spectral sampling", + "text_level": 1, + "bbox": [ + 511, + 681, + 808, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A key challenge in initializing a sinusoidal MLP lies in defining its input frequencies $\\omega$ . Recall that the MLP generates frequencies as integer combinations of $\\omega$ , i.e., $\\beta_{\\mathbf{k}} = \\sum_{i} k_{i} \\omega_{i}$ , with the remaining weights representing their amplitudes $\\alpha_{\\mathbf{k}}$ . This setup presents two main challenges. First, the initialization of $\\omega$ must enable $\\beta_{\\mathbf{k}}$ to cover the signal's spectrum; otherwise, the optimization cannot learn missing frequencies. Fig 3 (left) shows an example where using only odd frequencies leads to poor reconstruction. The second problem is the lack of flexibility for learning new frequencies. For example, randomly initializing $\\omega$ may not be enough since it may produce many high frequencies in order to overfit the signal, see Fig 4 (top).", + "bbox": [ + 511, + 703, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "$1|\\mathbf{k}|_{\\infty}$ denotes $\\max \\{|k_1|,\\dots ,|k_m|\\}$", + "bbox": [ + 107, + 885, + 316, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3073", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0a9150c266f167997b739e415fd292a20d6163031252f38a9c7519f10689866b.jpg", + "image_caption": [ + "Figure 2. Overview of TUNER. To train a sinusoidal MLP (gray model, top-left), we employ two techniques derived from Thrms 1 and 2. First, we initialize the input frequencies $\\omega$ (green, bottom-left) with a dense distribution of low frequencies (red square) and a sparse distribution of higher frequencies (green grid). This initialization gives flexibility to learn the remaining signal frequencies which are simply integer combinations of $\\omega$ (the yellow nodes on the right), a consequence of the amplitude-phase expansion given by Theorem 1. Note that this initialization resembles a frequency sampling since the training generates those new frequencies around $\\omega$ . Second, we bound the coefficients of the hidden layer weights (blue nodes) to ensure that the MLP remains within a specified bandlimit. This approach is effective because the amplitude-phase expansion (shown on the right) of each hidden neuron (purple nodes) indicates that the amplitudes of the generated frequencies have an upper-bound depending only on the hidden weights (blue, bottom-right)." + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 883, + 297 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/416cd5539eac2060604bbdf4c606cd6f8c689f3555d123c83fb057826b9b26b2.jpg", + "image_caption": [ + "Figure 3. Choosing $\\omega$ as the cartesian product of the odd frequencies without (left) or with (right) the frequencies $(1,0),(0,1)$ . Note that adding them prevents sub-periods (see Supp. Mat.). We trained for 3000 epochs on a $256^{2}$ image with network parameters $m = 72$ , $n = 512$ , and $\\theta = 30$ ." + ], + "image_footnote": [], + "bbox": [ + 145, + 422, + 284, + 529 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/58d5f9b65d759c8a987ff00280879f99b71ee245652153557b706075598f8407.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 422, + 424, + 529 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6884eee2b87978eb3e8c96218e154fd69d8dd5903f659e34f6a973962c374ee1.jpg", + "image_caption": [ + "Figure 4. Uniform init. of $\\omega$ (top) and ours (bottom). Grayscale bands show INR's gradient. Ours offers better signal/gradient reconstruction w/o gradient supervision. The MLPs ( $m = 128$ , $n = 256$ ) with bandlimit $\\theta = 87$ were trained for 3000 epochs." + ], + "image_footnote": [], + "bbox": [ + 101, + 604, + 468, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address these challenges, we introduce a novel initialization. First, we restrict our MLP to periodic functions, ensuring the training domain is in a full period. Such", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "functions can be represented by sums of sines with integer frequencies, defining an orthogonal basis (Fourier series); This aligns well with our analysis (Thrm 1). The main challenge is initializing $\\omega$ such that the generated frequencies $\\beta_{\\mathbf{k}}$ cover the full spectrum within a bandlimit $B$ (Fourier basis).", + "bbox": [ + 511, + 436, + 906, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More precisely, we initialize the weights $\\omega$ with integer frequencies, that is $\\omega_{j}\\in \\frac{2\\pi}{p}\\mathbb{Z}^{d}$ , and freeze them during training, with $p > 0$ being the period of the INR. This guarantees that the input neurons are $p$ -periodic and Thrm 1 ensures that such periodicity is preserved over layer composition. Additionally, (3) implies that we can rewrite the INR as:", + "bbox": [ + 511, + 512, + 905, + 602 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf (\\mathbf {x}) = \\sum_ {\\mathbf {k} \\in \\mathbb {Z} ^ {m}} \\left\\langle \\mathbf {C}, A _ {\\mathbf {k}} \\right\\rangle \\sin \\left(\\beta_ {\\mathbf {k}} \\mathbf {x}\\right) + \\left\\langle \\mathbf {C}, B _ {\\mathbf {k}} \\right\\rangle \\cos \\left(\\beta_ {\\mathbf {k}} \\mathbf {x}\\right) + e. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 607, + 906, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the generated frequencies $\\beta_{\\mathbf{k}}$ only depend on the frozen parameters $\\omega$ , the training is responsible for learning the amplitudes of the sine-cosine series in (4).", + "bbox": [ + 511, + 643, + 905, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $B$ be the signal's bandlimit. To ensure that the generated frequencies do not bypass $B$ , we cannot sample $\\omega$ directly on $[-B, B]^d$ since the MLP generates multiples of these frequencies. Therefore, we sample $\\omega_j$ in $\\frac{2\\pi}{p} [-\\theta, \\theta]^d$ , with $\\theta \\in (0, B)$ being a threshold for the input frequencies. Indeed, as the coefficients of hidden matrix $\\mathbf{W}$ are limited by 2 (see Sec 4.2), the magnitudes of $A_{\\mathbf{k}}$ and $B_{\\mathbf{k}}$ decrease as $|\\mathbf{k}|_{\\infty}$ increases. This is a consequence of Thrm 2:", + "bbox": [ + 511, + 690, + 906, + 811 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left| \\alpha_ {\\mathbf {k}} \\right| \\leq \\prod_ {j = 1} ^ {m} \\frac {\\left(\\frac {\\left| W _ {i j} \\right|}{2}\\right) ^ {\\left| k _ {j} \\right|}}{\\left| k _ {j} \\right| !} \\leq \\frac {1}{\\left| k _ {1} \\right| ! \\dots \\left| k _ {m} \\right| !}; i = 1, \\dots , n. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 814, + 903, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus, the generated frequencies $\\beta_{\\mathbf{k}}$ with small $\\mathbf{k}$ have more influence over the expansion (4), mimicking Fourier series.", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3074", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a2b79bedba58f4a9879c8b3418ad7bce8da827d77250787bfe1220414a2d8b03.jpg", + "image_caption": [ + "Figure 5. Image reconstructions with bandlimit $B = 128$ , varying $\\theta = 4, 43, 128$ with a network ( $m = 80, n = 1000$ ) trained over 3000 epochs. The gradient magnitude is shown on the left of each image. Note that smaller $\\theta$ (middle-left) yield better reconstructions, while higher values of $\\theta$ introduce noisy gradients (right)." + ], + "image_footnote": [], + "bbox": [ + 98, + 90, + 222, + 186 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/09c628756c6397fe6edc0030b950ab95f481b73fa5e61500eda2b50ff0e17673.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 90, + 348, + 188 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5dc6289d670196bc6830ffd0a171a48069fec8b6f99f4551f29845591a7f068e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 90, + 473, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For example, if input frequencies include 1 and 20, frequencies like $19 = 20 - 1$ and $21 = 20 + 1$ can be represented through their combinations. Thus, low frequencies give flexibility in expanding around the input ones which motivates initializing more input frequencies near the origin. We found $\\theta = \\frac{B}{3}$ to work well experimentally. Fig 5 shows that it is necessary, as higher $\\theta$ makes it harder to learn the small frequencies, introducing noise.", + "bbox": [ + 89, + 272, + 482, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To initialize $\\omega \\in \\mathbb{R}^{m\\times d}$ , we need to choose $m$ frequencies in $[- \\theta, \\theta]^d$ . Taking into account the importance of the lower frequencies discussed above, we split it into regions $\\mathbf{L} = [-\\ell, \\ell]^d$ of low and $\\mathbf{H} = [-\\theta, \\theta]^d \\setminus \\mathbf{L}$ of high frequencies, $\\ell \\in (0, \\theta)$ . As $\\beta_{\\mathbf{k}}$ only has a significant amplitude when $|\\mathbf{k}|_{\\infty}$ is small (5), initializing with only low input frequencies may not allow the generation of high frequencies. Thus, we initialize some frequencies in $\\mathbf{H}$ to cover the spectrum, and the rest in $\\mathbf{L}$ to cover a neighborhood of the input frequencies. Specifically, we select $m_l < m$ frequencies uniformly from $\\mathbf{L}$ . Then, we choose $m - m_l$ frequencies from $\\mathbf{H}$ , spread in a grid-like fashion to cover $[- \\theta, \\theta]^d$ . In our experiments we considered $m_l = 0.7m$ . Finally, since $\\mathbf{D}(\\mathbf{x}) = \\sin (\\omega \\mathbf{x} + \\varphi)$ , we initialize $\\varphi$ uniformly in $[- \\frac{\\pi}{2}, \\frac{\\pi}{2}]$ .", + "bbox": [ + 89, + 393, + 482, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When $\\omega_{j}\\in \\mathbf{L}$ , small variations in $k_{j}$ results in a frequency near $\\beta_{\\mathbf{k}}$ . Indeed, Fig 6 shows how low frequencies (green square) introduce frequencies around the input frequencies (Fig 6a). Conversely, if $\\omega_{j}\\in \\mathbf{H}$ then $\\beta_{\\mathbf{k} + \\mathbf{e}_j}$ is far away from $\\beta_{\\mathbf{k}}$ ; we note this in Fig 6b-d, where $\\omega_{j}$ is the red point and the arrows show the generated frequencies $k_{j}\\omega_{j}$ .", + "bbox": [ + 89, + 607, + 482, + 698 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e0aef8b4f6a97780f8dc3c30b24b6d57cbfa742af0ec8dae5628a3155f1ac2ad.jpg", + "image_caption": [ + "Figure 6. INR spectrum during training. The input frequencies $\\omega$ in green. During training new frequencies appear in a neighborhood of $\\omega$ . The red dot indicates a high input frequency $\\omega_{j}$ and the red arrows show the generated frequencies $k_{j}\\omega_{j}, k_{j} = 1,2,3$ ." + ], + "image_footnote": [], + "bbox": [ + 93, + 703, + 187, + 777 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/826da1394588ff8d16dfea45c4bf8e4dbcc333346b54852888c3f07be7bcb506.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 189, + 703, + 285, + 777 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/17ff66eee1e89182eb16257218cdac11d36593df270ead08c3e3151964c7c5b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 703, + 383, + 777 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6b6f30254f97045b646ad55426325e5014c40711f7d7844096e6d7de7a4060d7.jpg", + "image_caption": [ + "Figure 7. Training a network $f$ with $m = n = 256$ for 6000 epochs with $\\theta = 30$ and bounds $c = 0.1, 0.2, 0.5$ . The bound over the hidden matrix $\\mathbf{W}$ restricts the appearance of high frequencies." + ], + "image_footnote": [], + "bbox": [ + 383, + 703, + 480, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, when sampling frequencies in $\\mathbf{L}$ , we include $(1,0)$ and $(0,1)$ to guarantee that any frequency could be generated and prevents the appearance of sub-periods as", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "shown at the beginning of this section in Fig 3 (left) (see also Supp. Mat for more details). Fig 3 (right) shows how this problem is solved using our initialization.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proposed initialization overcomes the noisy reconstruction issues that arise with naive uniform sampling in $[- \\theta, \\theta]^d$ , as used in [26] (Fig 4 (top)). Fig 4 (bottom) demonstrates that our scheme significantly improves both signal and gradient reconstruction quality.", + "bbox": [ + 511, + 136, + 903, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Bounding and hidden layer initialization", + "text_level": 1, + "bbox": [ + 511, + 220, + 859, + 236 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in previous sections, sinusoidal INRs can represent detailed signals due to the wide number of frequencies generated by layer composition. However, they could manifest as noise in the reconstruction, see Fig 5 (right). Specifically, high input frequencies may generate even higher ones, bypassing the bandlimit $B$ . To avoid this, we use Thrm 2 to introduce a bounding scheme to limit the amplitudes. Precisely, Thrm 2 implies that in each hidden neuron $h_i$ , the amplitude of a generated frequency $\\beta_{\\mathbf{k}}$ is limited by:", + "bbox": [ + 511, + 242, + 906, + 378 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left| \\alpha_ {\\mathbf {k}} \\right| \\leq \\left(\\frac {\\left| \\mathbf {W} _ {i} \\right| _ {\\infty}}{2}\\right) ^ {\\left| \\mathbf {k} \\right| _ {1}} \\frac {1}{\\prod_ {j = 1} ^ {m} \\left| k _ {j} \\right| !}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 388, + 903, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_i$ is the $i$ -row of $\\mathbf{W}$ . Note that $|\\mathbf{W}|_{\\infty} \\leq 2$ implies that the amplitude of any generated frequency $\\beta_{\\mathbf{k}}$ rapidly decreases when the coordinates of $\\mathbf{k}$ grow. Therefore, we use (6) to define a bounding scheme that only allows the appearance of generated frequencies with small multiples of $\\omega$ which is beneficial as shown in previous sections. For this, we simply clamp the coefficients $W_{ij}$ by a bound parameter $c \\in (0,2]$ at each optimization step, enforcing $|\\mathbf{W}|_{\\infty} \\leq c$ . Fig 7 shows results of bounding $|\\mathbf{W}|$ with different $c$ .", + "bbox": [ + 511, + 436, + 906, + 573 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3f38fc49434962dd9928526679512b7558ee3f4fba8ed076b1b9237eae2c7fa3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 547, + 590, + 651, + 705 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2507cf1002ca90190c6fdd41c8c50c2a8ddafe122d333c044dfb5305122ef734.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 655, + 589, + 758, + 705 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8b45a8e6e57a66380bef200f415b384d883ed9550145f088a2a4abe47ebe4e68.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 589, + 867, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Sec 4.1 we split the input frequencies $\\omega$ in two sets: the low $(|\\omega_{j}|_{\\infty}\\leq \\ell)$ and high $(|\\omega_{j}|_{\\infty} > \\ell)$ frequencies. Additionally, we saw that the lower ones have a main role in the frequency generation, thus, we consider a higher threshold for them. For this, observe that each input frequency $\\omega_{j}$ is associated with the $j$ -column of $\\mathbf{W}$ . Then, bounding the columns of $\\mathbf{W}$ related to high frequencies, we restrict the amplitudes of $\\beta_{\\mathbf{k}}$ that could exceed the Nyquist limit $B$ . Conversely, since low frequencies are used to span", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3075", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "around $\\omega$ , we consider a higher bound for their columns. Precisely, we define two parameters $e_{\\mathbf{L}}$ and $e_{\\mathbf{H}}$ to bound the weights $W_{ij}$ corresponding to $\\omega_{j}$ in $\\mathbf{L}$ and $\\mathbf{H}$ , respectively,", + "bbox": [ + 89, + 90, + 482, + 137 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left| \\omega_ {j} \\right| _ {\\infty} \\leq \\ell \\longrightarrow W _ {i j} = \\operatorname {c l a m p} \\left(W _ {i j}, \\left[ - e _ {\\mathbf {L}}, e _ {\\mathbf {L}} \\right]\\right), \\\\ \\left| \\omega_ {j} \\right| _ {\\infty} > \\theta \\longrightarrow W _ {i j} = \\operatorname {c l a m p} \\left(W _ {i j}, \\left[ - \\boldsymbol {c} _ {\\mathbf {H}}, \\boldsymbol {c} _ {\\mathbf {H}} \\right]\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 141, + 464, + 180 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use these bounds to initialize $\\mathbf{W}$ such that each column is normal distributed with mean 0 and standard deviation $\\frac{c_{\\mathrm{L}}}{3}$ , with $c_{\\mathrm{L}}$ being the bound corresponding to that column. Note that SIREN [26] initializes $\\mathbf{W} \\in \\mathbb{R}^{n \\times m}$ uniformly in $(- \\sqrt{6/m}, \\sqrt{6/m})$ , and as $|\\mathbf{W}|_{\\infty} \\leq \\sqrt{6/m} < 2$ for $m > 1$ , it is bandlimited only at the beginning of training, since $\\mathbf{W}$ is not controlled in any way during the optimization. Conversely, our initialization controls the bandlimit during all training and has faster convergence.", + "bbox": [ + 89, + 183, + 483, + 319 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, we present a scheme to learn the bounding parameters during training. For this, we bound each $j$ -column of the hidden matrix $\\mathbf{W}$ using a tanh activation followed by multiplication with a trainable bounding parameter $e_j$ . Clearly, the resulting column is bounded by $e_j$ . In other words, we replace the hidden layer $\\sin (\\mathbf{W}\\mathbf{x} + \\mathbf{b})$ by", + "bbox": [ + 89, + 319, + 483, + 411 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\sin (\\tanh (\\mathbf {W}) \\mathcal {C} \\mathbf {x} + \\mathbf {b}), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 415, + 480, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{C}$ is a diagonal matrix with each $jj$ -entry being $c_{j}$ . To prevent the bounds from growing too large, we use Thrm 2 to define the regularization term as $\\mathcal{L}_{\\mathrm{reg}} = \\sum |c_j|$ . Otherwise, the hidden weights may grow unbounded, leading to the INR's overfit as in the usual training.", + "bbox": [ + 89, + 436, + 483, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 526, + 223, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section presents experiments regarding the initialization and bounding of frequencies of an INR $f$ considering the images from Kodak dataset [1] resized to $[-1, 1]^2$ . We describe the architecture of $f$ with hidden matrix $\\mathbf{W} \\in \\mathbb{R}^{n \\times m}$ with the parameters $m, n$ . We initialize $f$ following the scheme in Sec 4.1 and optimize it using Adam [9] with learning rate $10^{-4}$ . For the quantitative evaluations, we report the mean and standard deviation of the Peak Signal-to-Noise Ratio (PSNR) evaluated on the test set (10% random pixels). To show that TUNER adds high order regularity to the reconstruction, we compare the analytical gradient $\\nabla f$ to the Sobel filter of the ground truth. All experiments were run in a 12 GB NVIDIA GPU (TITAN X Pascal).", + "bbox": [ + 89, + 551, + 483, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Initialization", + "text_level": 1, + "bbox": [ + 89, + 757, + 225, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First, we compare our initialization of $\\omega$ (36.2dB) against the uniform distribution (32.2dB). We obtain a 4dB improvement of PNSR on average (with a std of 1.6dB). We also present an ablation over the parameter $\\ell$ which splits the square $[- \\ell, \\ell]^2$ in regions of low (L) and high (H) frequencies. Here we use $\\ell = 85$ and $\\ell = 21$ , 42, 64, and sample $\\omega$ with $30\\%$ , $50\\%$ , and $70\\%$ of coordinates in H. Table 1 shows the performance of the INR.", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3843b29c5e4e05e7640b64587ae02fc9cb5ce1689928ff33090c35407c594bd4.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Training an INR with $m = n = 416$ parameters, for 400 epochs. Results indicate that it is better to sample more $\\omega_{j} \\in \\mathbf{L}$ and choose a smaller $\\ell$ ." + ], + "table_body": "
% high freqs30%50%70%
\\( \\ell \\)SignalGradSignalGradSignalGrad
2135.226.935.226.635.026.0
4235.126.334.824.633.822.2
6434.225.233.321.630.815.0
", + "bbox": [ + 516, + 88, + 903, + 156 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This experiment considered all the images in the dataset. Observe that all values of the row $\\ell = 64$ have lower PSNR, showing that an INR with very high frequencies performs worse even for small percentages $(30\\%)$ . Similarly, the performance worsens when we increase the number of high frequencies (last column).", + "bbox": [ + 511, + 212, + 903, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We now test the influence of high frequencies by training an INR of size $m = 360$ , $n = 512$ with $\\theta = 256$ over $90\\%$ of the ground truth pixels. We vary the parameter $\\ell = 60, 120, 220$ that defines low frequencies. Increasing the number of higher frequencies, the INR suffers from noise and overfitting. This can be noted in the reconstruction on the unsupervised pixels and gradients in Fig 8. Finally, in Supp. Mat. we compare our initialization of W against SIREN and observe a faster convergence.", + "bbox": [ + 511, + 303, + 905, + 439 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d152895bb551ad1e678149fb5197ba69982a42fe69ad3d4f8cffbe8b976889d7.jpg", + "image_caption": [ + "Figure 8. Re constructions with big $\\ell$ present overfitting on the $10\\%$ unsupervised points. Gradient on the upper right corner." + ], + "image_footnote": [], + "bbox": [ + 553, + 454, + 653, + 599 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ef215ca3acb12cbbf4fb9fd785a437353c53385489dee7244dda4bea2dd6ece7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 655, + 455, + 756, + 599 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4cac0607a8df965aaae5c75da9ba368b9c9af65f8e0f377c39ff3be2b84168c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 455, + 862, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Bounding", + "text_level": 1, + "bbox": [ + 513, + 648, + 625, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Sec 4.2 we presented a scheme to bound the INR spectrum. Here, we show additional experiments to demonstrate that this improves the signal and has a great benefit over the unsupervised gradient reconstruction. Then, we give a comparison between fixed and trained bounding schemes.", + "bbox": [ + 511, + 672, + 905, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig 9 compares the reconstructions of an image of size $512^2$ w/o/w our fixed bounding scheme. To demonstrate the impact of our scheme, we sample the input frequencies using $\\ell = 256$ and $\\ell = 100$ . Without bounding (left), we effectively reconstruct the signal at the supervised pixels (30.8 dB), but with a noisy gradient (15.8 dB). That is, we are overfitting the signal. Conversely, bounding the hidden weights significantly improves both the signal (35.1 dB) and its gradient (27.8 dB). Our bounding scheme acts like a filter in both signal and gradient reconstruction.", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3076", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/778310baad253422766e06d6928afb5b93abce72d0a58facd8551d044e068e22.jpg", + "image_caption": [ + "Figure 9. Comparison of training $\\mathrm{wo} / \\mathrm{w}$ fixed bounds. Note that the use of bounds preserves high order information of the signal. Left side of the figures show the network gradients." + ], + "image_footnote": [], + "bbox": [ + 112, + 90, + 459, + 217 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare quantitatively our bounding scheme over the dataset, varying the parameters $\\theta$ and $c_{\\mathbf{H}}$ . Table 2 shows an improvement in the signal for all cases. Furthermore, the unsupervised gradient of the image is better learned when using fixed bounds, showing an average increase of 9.6dB against the standard training. Also, the last two lines of the table show the impact of the parameters $c_{\\mathbf{L}}$ , $c_{\\mathbf{H}}$ over the signal and gradient reconstructions. We consider training with $c_{\\mathbf{L}} > c_{\\mathbf{H}}$ since we are prioritizing low input frequencies in the spectrum generation. Otherwise, this may generate high frequencies, possibly bypassing the Nyquist limit and resulting in a worse gradient reconstruction.", + "bbox": [ + 89, + 281, + 483, + 464 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8c6552a5dad013456f80c37aad23164c07195045efa9b4b21d49f1d12723ccd8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
signalgrad
θcLcHwo boundw boundwo boundw bound
2561.50.0532.634.814.525.8
1901.50.0534.035.918.526.1
1901.50.234.034.418.521.0
", + "bbox": [ + 91, + 479, + 486, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Training with bounding improves the signal and grad. reconstruction. The choice of the fixed bound $c_{\\mathbf{H}}$ greatly impacts the grad. fidelity. The std of the difference between each image signal (grad) quality $w / w$ bounding is 1.4dB (1.6dB) on average. Evaluations performed on the test set (10% of pixels).", + "bbox": [ + 89, + 547, + 483, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we show how we can learn each bound $c_{j}$ . As described in Sec 4.2, we can incorporate a learnable $c_{j}$ for each $j$ -column of the hidden matrix through a modified layer (7). Fig 10 shows a comparison between our fixed bound scheme using initial un-tuned $c_{\\mathbf{L}} = 1.0$ and $c_{\\mathbf{H}} = 0.6$ (left) and the learned bound scheme (middle).", + "bbox": [ + 89, + 642, + 482, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As expected, training lowers the bounds for high frequencies (H) while maintaining or increasing the bounds of low frequencies (L). Thus, we achieve better reconstruction in both the signal and gradient. In Fig 10 (right), we note that low frequencies (left of the green line) have significant bounds, showing their importance during training. These schemes allow us to initialize with higher frequencies than $B / 3$ , accelerating convergence. Finally, we performed an ablation study on the regularization term $\\mathcal{L}_{reg}$ to show its importance. Training w/wo regularization yielded 28.5/27.4dB for the RGB and 27.1/24.0dB for the gradient.", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5ef1823b9863aebbefc72d4e3b1f85b01bdc63b3fb68ee3be83ae389fc04298f.jpg", + "image_caption": [ + "Figure 10. Comparison of signal and gradient (grayscale) when training with fixed/learned bounds. The learned bounds adjust the INR spectrum resulting in better reconstructions. Each blue point represents a pair $(|\\omega_j|_{\\infty}, c_j)$ , where $c_j$ is the trained bound of $\\omega_j$ ." + ], + "image_footnote": [], + "bbox": [ + 519, + 90, + 903, + 198 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Comparisons", + "text_level": 1, + "bbox": [ + 511, + 284, + 653, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first compare our method (TUNER) with SIREN, which requires choosing a threshold parameter $\\theta$ , a critical task since higher $\\theta$ implies noisy reconstructions. TUNER also uses $\\theta$ to control the frequency generation, however, it provides more stable training with better reconstruction. To demonstrate this, we perform a comparison by varying $\\theta$ and the number of input ( $m$ ) and hidden ( $n$ ) neurons. First, consider the case where we have no information about the signal bandlimit besides the Nyquist limit $B$ . Fig 11 presents a comparison between SIREN (top) and TUNER (bottom) with $\\theta = B$ . Note that SIREN produces noisy reconstructions, while ours learn faster and does not suffer from noise. This is due to our bounding scheme, which limits the generation of high frequencies and provides stable training (Sec 4.2). In contrast, SIREN initializes its parameters such that the spectrum is initially bounded but does not maintain this guarantee during training.", + "bbox": [ + 511, + 306, + 906, + 565 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/421ae920815b31ddb6f15579eedb9daa82c3912508ad9fd1a99d11fbdc6c4ee7.jpg", + "image_caption": [ + "Figure 11. Comparing SIREN (top) and TUNER (bottom) when we only know the Nyquist limit (i.e, $\\theta = B$ ). Since $\\theta$ is big, SIREN introduces high frequencies at the beginning of training, resulting in very noisy reconstructions. Conversely, ours starts training low frequencies, converging faster and being robust to $\\theta$ ." + ], + "image_footnote": [], + "bbox": [ + 517, + 575, + 908, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we consider a tuned $\\ell$ for SIREN to show that even under this condition, our method provides a better training scheme for sinusoidal INRs. Fig 12 presents the comparison", + "bbox": [ + 511, + 829, + 906, + 876 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$^{2}$ Denoted as $\\omega_0$ in [26].", + "bbox": [ + 531, + 886, + 661, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3077", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "varying the number of input $(m)$ and hidden $(n)$ neurons. We consider the cases $m < n$ (top) and $m > n$ (bottom), and TUNER performed better both in the signal/gradient reconstructions. Also, we note more robust training using TUNER, as shown in Fig 12(right).", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The previous experiment shows that initializing SIREN with a high $\\theta$ results in overfitting, while TUNER provides better reconstruction. On the other hand, with the same architecture and a tuned $\\theta$ , our method also effectively learned the lower frequencies first and performed better (Fig 12).", + "bbox": [ + 89, + 167, + 480, + 243 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1df572d89bb0dc7231d1830ad4041c4b386deef8552365012b69a9f7d4c4d142.jpg", + "image_caption": [ + "Figure 12. Comparison with SIREN when $\\theta = 40$ . TUNER improves training convergence in both cases ( $m < n$ , $m > n$ ) and stability when $m > n$ . INR gradient shown in grayscale." + ], + "image_footnote": [], + "bbox": [ + 91, + 255, + 478, + 401 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Next, we numerically evaluate these experiments, providing a quantitative comparison with SIREN and BACON. We compare different architectures $(m,n)$ and spectrum bandlimits $\\theta$ , evaluating the results in terms of signal and gradient PSNR over the supervised (Train) and unsupervised (Test) pixels. Table 3 presents these evaluations.", + "bbox": [ + 89, + 472, + 483, + 564 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/197b36174bb296bbb8d24f7a0a536f3e25d2fded384cff7a9121bcf7be4e64d4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TrainTest
SignalGradientSignalGradient
θ(m,n)SIRENBACOURSSIRENBACOURSSIRENBACOURSSIRENBACOURS
851028,16232.9-32.926.5-28.831.9-32.225.5-28.2
416,41634.026.434.327.722.428.932.925.433.427.122.328.4
80,202432.2-33.126.6-28.531.3-32.326.0-28.0
1711028,16231.3-33.922.0-27.222.9-32.920.7-26.8
416,41629.229.334.417.426.727.026.726.233.116.226.326.4
80,202426.6-30.814.2-22.623.8-29.513.8-21.3
2561028,16226.8-33.312.8-24.724.1-31.912.7-23.9
416,41624.727.741.212.924.223.022.322.733.213.222.821.2
80,202425.1-30.013.0-15.224.1-27.113.1-14.6
", + "bbox": [ + 91, + 574, + 488, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Comparison between SIREN, BACON (BAC), and our method. We use several architectures $(m,n)$ ,bandlimits $(\\theta)$ ,and evaluate over the train $(90\\%)$ and test $(10\\%)$ sets of signal/gradient.", + "bbox": [ + 89, + 736, + 482, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As expected, SIREN performs well with $\\theta = \\frac{B}{3}$ . In this case, TUNER is perceptually similar to SIREN but improves gradient reconstruction and consistently outperforms BACON. When $\\theta = 171$ , TUNER offers superior quality in both signal and gradient, showing greater robustness to noise and superior performance even with $\\theta \\geq B/3$ . Finally, with $\\theta = 256$ , our method surpasses SIREN in all cases.", + "bbox": [ + 89, + 792, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For BACON, our method achieves better signal reconstruction, with comparable gradients. However, the following experiment shows that BACON's reconstruction has visible defects in this scenario. Note that we do not supervise the training with gradient information.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Bandlimit comparison: Lastly, Fig 13 presents a qualitative comparison between BACON, BANF, and TUNER (Ours) with bandlimits $\\theta = 60, 130$ . We train all methods with 67K parameters. For BANF we use the linear filter and adjust grid resolution to match the bandlimits and the MLP size to approximate 67K parameters. Particularly, the spectra were computed with the same process for all three methods. Note that BACON truncates the spectrum with a hard filter, resulting in ringing artifacts (low bandlimit) and noisy reconstruction (high bandlimit). BANF shows distortions around the eye area for both filters, and the spectra show no clear bandlimit. Conversely, TUNER operates as a soft filter providing better reconstructions for all cases.", + "bbox": [ + 511, + 184, + 906, + 380 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d55aa1a130db4305156402adf5fce5539d522f49acb971b45ad4c0beae64d813.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 382, + 643, + 516 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e598c70896dabb7dbe5c39f301d0361f292391302c5e87ef5d76e3559e647e77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 382, + 764, + 515 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a26d5d0407ef594c50a29fa94e040994c0d5de8bfb2ae6a2df3615075852d5b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 382, + 885, + 515 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f53784834ac65abc3e568fa9d79d8058eeb8a886e6c3b8b67e66c6968df1d1cf.jpg", + "image_caption": [ + "Figure 13. Comparison between BACON, BANF, and TUNER, trained with bandlimits $\\theta = 60, 130$ . Note that BACON uses a box filter, generating ringing artifacts (left). BANF (middle) has artifacts near edges and ambiguous bandlimit. Then, TUNER (right) resembles a soft filter, improving quality." + ], + "image_footnote": [], + "bbox": [ + 524, + 517, + 643, + 651 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4c502bb2d474d7e6fd31efdbd0d9597287831b9cf1db675f569a9568efb9dd88.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 517, + 764, + 651 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cddab5cc49355e18ad562340ff5054ef2fb3f4153d12fe3a514e777c09682106.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 517, + 885, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions and limitations", + "text_level": 1, + "bbox": [ + 511, + 739, + 769, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented a study based on Fourier theory for sinusoidal MLPs, which resulted in novel, robust schemes for both initialization and spectral control of such networks. While our experiments focused on images to validate our theoretical claims, future work includes applying our schemes to other signal types. Additionally, we have focused our studies on 3-layer MLPs, leaving the investigation of deep networks for future research. We consider this work a first step toward unraveling the powerful expressiveness of sinusoidal MLPs.", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3078", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments We would like to thank CAPES, FAPERJ, and Google for partially funding this work.", + "bbox": [ + 89, + 90, + 480, + 121 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 136, + 187, + 151 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kodak dataset: https://r0k.us/graphics/kodak/.3, 6", + "[2] Milton Abramowitz and Irene A Stegun. Handbook of mathematical functions with formulas, graphs, and mathematical tables. US Government printing office, 1964. 3", + "[3] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 1", + "[4] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628-8638, 2021. 2", + "[5] Yishun Dou, Zhong Zheng, Qiaoqiao Jin, and Bingbing Ni. Multiplicative fourier level of detail. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1808-1817, 2023. 2", + "[6] Emilien Dupont, Adam Golinski, Milad Alizadeh, Yee Whye Teh, and Arnaud Doucet. Coin: Compression with implicit neural representations. In International Conference on Learning Representations, 2021. 2", + "[7] Rizal Fathony, Anit Kumar Sahu, Devin Willmott, and J Zico Kolter. Multiplicative filter networks. In International Conference on Learning Representations, 2020. 1, 2", + "[8] Adam Kania, Marko Mihajlovic, Sergey Prokudin, Jacek Tabor, Przemys' Spurek, et al. Fresh: Frequency shifting for accelerated neural representation learning. arXiv preprint arXiv:2410.05050, 2024. 2", + "[9] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.3,6", + "[10] David B Lindell, Dave Van Veen, Jeong Joon Park, and Gordon Wetzstein. Bacon: Band-limited coordinate networks for multiscale scene representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16252-16262, 2022. 1, 2", + "[11] Zhen Liu, Hao Zhu, Qi Zhang, Jingde Fu, Weibing Deng, Zhan Ma, Yanwen Guo, and Xun Cao. Finer: Flexible spectral-bias tuning in implicit neural representation by variable-periodic activation functions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2713-2722, 2024. 1", + "[12] Ishit Mehta, Manmohan Chandraker, and Ravi Ramamoorthi. A level set theory for neural implicit evolution under explicit flows. In European Conference on Computer Vision, pages 711-729. Springer, 2022. 2", + "[13] Tiago Novello, Guilherme Schardong, Luiz Schirmer, Vinicius da Silva, Helio Lopes, and Luiz Velho. Exploring differential geometry in neural implicits. Computers & Graphics, 108:49-60, 2022. 2" + ], + "bbox": [ + 93, + 161, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Tiago Novello, Vinicius da Silva, Guilherme Schardong, Luiz Schirmer, Helio Lopes, and Luiz Velho. Neural implicit surface evolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14279-14289, 2023. 2", + "[15] Giambattista Parascandolo, Heikki Huttunen, and Tuomas Virtanen. Taming the waves: sine as activation function in deep neural networks. 2016. 1, 2", + "[16] Hallison Paz, Daniel Perazzo, Tiago Novello, Guilherme Schardong, Luiz Schirmer, Vinicius da Silva, Daniel Yukimura, Fabio Chagas, Helio Lopes, and Luiz Velho. Mrnet: Multiresolution sinusoidal neural networks. Computers & Graphics, 2023. 2", + "[17] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International Conference on Machine Learning, pages 5301-5310. PMLR, 2019. 2", + "[18] E. Romero, J. Sopena, R. Alquezar, and J. Moliner. Neural networks with periodic and monotonic activation functions: a comparative study in classification problems. 2000. 2", + "[19] Vishwanath Saragadam, Jasper Tan, Guha Balakrishnan, Richard G Baraniuk, and Ashok Veeraraghavan. Miner: Multiscale implicit neural representation. In European Conference on Computer Vision, pages 318-333. Springer, 2022. 2", + "[20] Vishwanath Saragadam, Daniel LeJeune, Jasper Tan, Guha Balakrishnan, Ashok Veeraraghavan, and Richard G Baraniuk. Wire: Wavelet implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18507-18516, 2023. 2", + "[21] Hemanth Saratchandran, Sameera Ramasinghe, Violetta Shevchenko, Alexander Long, and Simon Lucey. A sampling theory perspective on activations for implicit neural representations. In International Conference on Machine Learning, pages 43422-43444. PMLR, 2024. 2", + "[22] Guilherme Schardong, Tiago Novello, Hallison Paz, Iurii Medvedev, Vinícius da Silva, Luiz Velho, and Nuno Gonçalves. Neural implicit morphing of face images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7321-7330, 2024. 2", + "[23] Luiz Schirmer, Tiago Novello, Guilherme Schardong, Vinicius da Silva, Helio Lopes, and Luiz Velho. How to train your (neural) dragon. In Conference on Graphics, Patterns and Images, 36. (SIBGRAPI), 2023. 2", + "[24] Akhmedkhan Shabanov, Shrisudhan Govindarajan, Cody Reading, Lily Goli, Daniel Rebain, Kwang Moo Yi, and Andrea Tagliasacchi. Banf: Band-limited neural fields for levels of detail reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20571-20580, 2024. 2", + "[25] Kathan Shah and Chawin Sitawarin. Spider: Semiperiodic damping-enabled object representation. arXiv preprint arXiv:2306.15242, 2023. 2", + "[26] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in neural" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3079", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "information processing systems, 33:7462-7473, 2020. 1, 2, 5, 6, 7", + "[27] Ying Song, Jiaping Wang, Li-Yi Wei, and Wencheng Wang. Vector regression functions for texture compression. ACM Transactions on Graphics (TOG), 35(1):1-10, 2015. 2", + "[28] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 1, 2", + "[29] Zhijie Wu, Yuhe Jin, and Kwang Moo Yi. Neural fourier filter bank. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14153-14163, 2023. 2", + "[30] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, pages 641-676. Wiley Online Library, 2022. 2", + "[31] Guandao Yang, Serge Belongie, Bharath Hariharan, and Vladlen Koltun. Geometry processing with neural fields. Advances in Neural Information Processing Systems, 34: 22483-22497, 2021. 2", + "[32] Guandao Yang, Sagie Benaim, Varun Jampani, Kyle Genova, Jonathan Barron, Thomas Funkhouser, Bharath Hariharan, and Serge Belongie. Polynomial neural fields for subband decomposition and manipulation. Advances in Neural Information Processing Systems, 35:4401-4415, 2022. 2", + "[33] Wang Yifan, Lukas Rahmann, and Olga Sorkine-hornung. Geometry-consistent neural shape representation with implicit displacement fields. In International Conference on Learning Representations, 2021. 2", + "[34] Gizem Yüce, Guillermo Ortiz-Jiménez, Beril Besbinar, and Pascal Frossard. A structured dictionary perspective on implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19228-19238, 2022. 3", + "[35] Andreas Zell, Nuri Benbarka, Timon Hoefer, and Hamd Ul Moqueet Riaz. Seeing implicit neural representations as fourier series. In 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). IEEE Computer Society, 2022. 1, 2" + ], + "bbox": [ + 91, + 90, + 482, + 698 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3080", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Tuning the Frequencies_ Robust Training for Sinusoidal Neural Networks/79e8b6d6-50a7-4215-bd2a-fb5524e17539_model.json b/2025/Tuning the Frequencies_ Robust Training for Sinusoidal Neural Networks/79e8b6d6-50a7-4215-bd2a-fb5524e17539_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ded2862be0272cc3119edc242696c1a4ad5da363 --- /dev/null +++ b/2025/Tuning the Frequencies_ Robust Training for Sinusoidal Neural Networks/79e8b6d6-50a7-4215-bd2a-fb5524e17539_model.json @@ -0,0 +1,2156 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.129, + 0.131, + 0.872, + 0.154 + ], + "angle": 0, + "content": "Tuning the Frequencies: Robust Training for Sinusoidal Neural Networks" + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.182, + 0.731, + 0.217 + ], + "angle": 0, + "content": "Tiago Novello\\(^{1,2,*}\\) Diana Aldana\\(^{1,*}\\) Andre Araujo\\(^{2}\\) Luiz Velho\\(^{1}\\) \n\\(^{1}\\) IMPA \\({}^{2}\\) Google DeepMind" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.259, + 0.327, + 0.274 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.288, + 0.486, + 0.576 + ], + "angle": 0, + "content": "Sinusoidal neural networks have been shown effective as implicit neural representations (INRs) of low-dimensional signals, due to their smoothness and high representation capacity. However, initializing and training them remain empirical tasks which lack on deeper understanding to guide the learning process. To fill this gap, our work introduces a theoretical framework that explains the capacity property of sinusoidal networks and offers robust control mechanisms for initialization and training. Our analysis is based on a novel amplitude-phase expansion of the sinusoidal multilayer perceptron, showing how its layer compositions produce a large number of new frequencies expressed as integer combinations of the input frequencies. This relationship can be directly used to initialize the input neurons, as a form of spectral sampling, and to bound the network's spectrum while training. Our method, referred to as TUNER (TUNing sinusoidal nEtwoRks), greatly improves the stability and convergence of sinusoidal INR training, leading to detailed reconstructions, while preventing overfitting." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.587, + 0.222, + 0.602 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.605, + 0.484, + 0.741 + ], + "angle": 0, + "content": "Sinusoidal multilayer perceptrons (MLPs) emerged as powerful implicit neural representations (INRs) for low-dimensional signals [3, 11, 26, 35]. In this context, the INR \\( f \\) should fit the input data \\( \\{x_{i},\\ell_{i}\\} \\) as close as possible, i.e. \\( f(x_{i})\\approx \\ell_{i} \\), without overfitting, thus encoding the signal implicitly in the MLP parameters. Therefore, two major properties are required: (1) \\( f \\) needs high representation capacity to fit \\( \\{x_i,\\ell_i\\} \\); (2) \\( f \\) should have bandlimit control to avoid frequencies bypassing the sampling rate." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.741, + 0.484, + 0.877 + ], + "angle": 0, + "content": "Training sinusoidal MLPs to satisfy the above properties is challenging, as their initialization and optimization process often lead to undesirable local minima [15]. Recent work made strides towards more effective learning of these models. For example, SIREN [26] proposed an initialization by projecting the input coordinates to a list of sines with frequencies randomly chosen in a range, similar to the Fourier feature mapping (FFM) approach [28]. This way, the model can reach high capacity, but may lead to over" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.258, + 0.902, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.628, + 0.907, + 0.754 + ], + "angle": 0, + "content": "Figure 1. We present TUNER, a robust and theoretically grounded training technique for sinusoidal MLPs, overcoming challenges in initialization and enabling bandlimiting control. Our experiments showcase TUNER's strong initialization results against ReLU, FFM [28], and SIREN [26] (top), where all models use the same size and training conditions. TUNER achieves both fast and stable convergence (bottom-left) while reconstructing gradients without noise. We also compare with BACON [10] across two bandlimits (bottom-right), enhancing quality and avoiding ringing artifacts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "fitting with high frequencies resulting in noisy reconstructions. Defining an effective range for the bandlimit initialization remains mostly empirical and often results in noise, as the role of layer composition in generating frequencies is not fully understood. Additionally, uniform initialization may introduce undesired high frequencies and make it harder to model lower ones. More recently, BACON [10] proposes tighter bandlimit control by applying multiplicative filter networks (MFNs) [7] to limit the signal spectrum" + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.887, + 0.365, + 0.901 + ], + "angle": 0, + "content": "*These authors contributed equally to this work." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3071" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.166 + ], + "angle": 0, + "content": "with a box filter, hard-truncating the spectrum. While effective in many cases, this may lead to ringing artifacts. Also, this technique lacks non-linear activations (MFNs are not neural networks), preventing the representation of fine details as efficiently as sinusoidal MLPs with similar sizes." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.168, + 0.483, + 0.514 + ], + "angle": 0, + "content": "Overall, initializing and training sinusoidal MLPs remain empirical tasks which lack on deeper understanding to guide the learning. To approach these problems, we present TUNER (TUNing sinusoidal nEtwoRks), a training scheme for sinusoidal MLPs consisting of a robust initialization and bandlimit mechanism that greatly improves the capacity and convergence of INRs. In contrast to previous work, TUNER is grounded on a theoretical framework that guides our learning approach. We develop a new trigonometric identity (Thrm 1), resembling a Fourier series, and prove that such expansion produces a large number of frequencies in terms of the first layer (input frequencies). This explains the frequency generation process when composing layers, highlighting the importance of input layer initialization for improving capacity. Remarkably, the amplitudes of these frequencies are given by complex functions of the hidden weights, introducing a challenging training. To address this, we prove that those are bounded by a term depending only on the hidden weights (Thrm 2). We apply this result to control the bandlimit of the INR during training. Some of TUNER's results are showcased in Fig. 1, in comparison to previous work, illustrating substantial enhancements to sinusoidal INR modeling. In summary, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.515, + 0.482, + 0.606 + ], + "angle": 0, + "content": "- We introduce a novel trigonometric identity that expands any hidden neuron into a wide sum of sines with frequencies being integer combinations of the input frequencies (Fig 2). The corresponding amplitudes, determined by the hidden weights, have a useful upper bound, offering effective tools for controlling the resulting signal." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.606, + 0.482, + 0.695 + ], + "angle": 0, + "content": "- The expansion enables a robust initialization scheme for sinusoidal INRs, resulting in a high capacity MLP that trains significantly faster than previous approaches. First, we initialize the input neurons, as they determine the INR spectrum. Next, we initialize the hidden weights, considering their role in setting the corresponding amplitudes." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.697, + 0.482, + 0.786 + ], + "angle": 0, + "content": "- We leverage the amplitude's upper bound to control the bandlimit of a sinusoidal INR by designing schemes that bound the hidden weights during training. Together with our initialization, these bounds generate frequencies centered around the input, resulting in more stable training compared to previous approaches." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.515, + 0.482, + 0.786 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.8, + 0.236, + 0.815 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Implicit neural representations [8, 20, 21, 25, 30, 32] are a trending topic in machine learning, used to learn highly detailed signals in low-dimension domains. Current INR architectures use Fourier feature mappings [28] or sinusoidal activation functions [26] to bypass the spectral bias [17]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "common in ReLU multilayer perceptrons. The high representation capacity of sinusoidal INRs has motivated their use to represent a wide range of media objects. Examples include audio [6, 26], images [4, 27], face morphing [22], signed distance functions [10, 13, 23, 26, 31], displacement fields [33], surface animation [12, 14], multi-resolution signals [5, 10, 16, 19, 29], among others. Most of these exploit the derivatives of sinusoidal INRs in the loss functions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.213, + 0.905, + 0.468 + ], + "angle": 0, + "content": "Initialization. Considering sinusoidal activation functions in neural networks is a classical problem [18]; however, these INRs have been regarded as difficult to train [15]. Sitzmann et al. [26] overcome this by presenting a specific initialization scheme that allows training sinusoidal INRs, avoiding instability and ensuring convergence. Despite these advances, in practice, the initialization of such INRs remains an empirical task. In this work, inspired by Fourier series theory, we present a novel initialization method that allows us to train INRs with great convergence. Recently, several works have addressed the representation problem of sinusoidal INRs from different perspectives. Zell et al. [35] approached this problem by observing that the first layer of a sinusoidal INR is similar to a Fourier feature mapping. Here, in addition to improving the initialization of sinusoidal INRs, we present a training scheme for bounding the spectra of these networks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.905, + 0.605 + ], + "angle": 0, + "content": "Control of spectrum. One of the main drawbacks of sinusoidal INRs is the lack of frequency control. SIREN [26] addressed this by initializing the input frequencies uniformly in a range. While this ensures a bandlimit at the start of training, as it progresses, the layer composition introduces higher frequencies, leading to noisy reconstructions. We avoid this by providing a novel initialization for the first layer coupled with a bounding scheme which gives controls for limiting the MLP spectrum." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.606, + 0.905, + 0.772 + ], + "angle": 0, + "content": "Other works [5, 10] employed MFNs [7] to control the bandlimit by applying a filter on the spectrum. However, this strategy usually introduces reconstruction artifacts since MFNs hard-truncate the spectrum. BANF [24] employed a grid-based MLP with a spatial filter that leverages grid resolution to constrain the highest frequency in the network. In our experiments, we observe that this is prone to creating higher frequencies which propagate as artifacts. In contrast, TUNER, grounded in a theoretical result (Thrm 2), guarantees a bandlimited MLP, and serves as a soft filter, providing a representation without ringing artifacts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.785, + 0.825, + 0.801 + ], + "angle": 0, + "content": "3. Sinusoidal MLPs as Fourier series" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.905, + 0.87 + ], + "angle": 0, + "content": "This work addresses the problem of deriving an efficient training scheme with controlled bandlimit for sinusoidal MLPs. This section presents the mathematical definitions and novel formulas for approaching this task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Sinusoidal MLPs demonstrated high representational capacity with only a few hidden layers [19, 26, 35]. To un" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3072" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.272 + ], + "angle": 0, + "content": "derstand how layer composition encodes this capacity, we propose to thoroughly investigate the structure of a 3-layer sinusoidal MLP \\( f: \\mathbb{R}^d \\to \\mathbb{R} \\). For simplicity, we assume the codomain to be \\( \\mathbb{R} \\), however the same analysis holds for dimension \\( > 1 \\). More precisely, we consider \\( f(\\mathbf{x}) = \\mathbf{C} \\circ \\mathbf{S} \\circ \\mathbf{D}(\\mathbf{x}) + e \\), with \\( \\mathbf{D}(\\mathbf{x}) = \\sin (\\omega \\mathbf{x} + \\varphi) \\) being the input layer that projects \\( \\mathbf{x} \\) into a list of harmonics (input neurons) \\( D_{i}(\\mathbf{x}) = \\sin (\\omega_{i}\\mathbf{x} + \\varphi_{i}) \\) with frequencies \\( \\omega = (\\omega_{1},\\dots,\\omega_{m}) \\in \\mathbb{R}^{m\\times d} \\) and shifts \\( \\varphi \\in \\mathbb{R}^{m} \\). Layer \\( \\mathbf{D} \\) is then composed with \\( \\mathbf{S}(\\mathbf{x}) = \\sin (\\mathbf{W}\\mathbf{x} + \\mathbf{b}) \\), where \\( \\mathbf{W} \\in \\mathbb{R}^{n\\times m} \\) is the hidden matrix, and \\( \\mathbf{b} \\in \\mathbb{R}^n \\) the bias. Finally, \\( \\mathbf{C} \\cdot \\mathbf{x} + e \\) is an affine transformation with \\( \\mathbf{C} \\in \\mathbb{R}^{n\\times 1} \\) and \\( e \\in \\mathbb{R} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.484, + 0.363 + ], + "angle": 0, + "content": "We now present some properties of the sinusoidal layers, which play key roles in the representation, and give a reinterpretation of them in terms of the network parameters. First, note that the hidden neurons \\(\\mathbf{h}(\\mathbf{x})\\coloneqq \\mathbf{S}\\circ \\mathbf{D}(\\mathbf{x})\\) are defined by composing the dictionary \\(\\mathbf{D}\\) with the hidden sinusoidal layer which results in a list with elements" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.364, + 0.483, + 0.402 + ], + "angle": 0, + "content": "\\[\nh _ {i} (\\mathbf {x}) = \\sin \\left(\\sum_ {j = 1} ^ {m} W _ {i j} \\sin \\left(\\omega_ {j} \\mathbf {x} + \\varphi_ {j}\\right) + b _ {i}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.401, + 0.484, + 0.448 + ], + "angle": 0, + "content": "The following is a key result of this work, which states that we can linearize a hidden neuron (1) as a sum of sines with frequencies and amplitudes determined by \\(\\omega\\) and \\(\\mathbf{W}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.454, + 0.484, + 0.485 + ], + "angle": 0, + "content": "Theorem 1. Each hidden neuron \\( h_i \\) of a 3-layer sinusoidal MLP has an amplitude-phase expansion of the form" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.488, + 0.483, + 0.519 + ], + "angle": 0, + "content": "\\[\nh _ {i} (\\mathbf {x}) = \\sum_ {\\mathbf {k} \\in \\mathbb {Z} ^ {m}} \\alpha_ {\\mathbf {k}} \\sin \\left(\\beta_ {\\mathbf {k}} \\mathbf {x} + \\lambda_ {\\mathbf {k}}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.483, + 0.552 + ], + "angle": 0, + "content": "where \\(\\beta_{\\mathbf{k}} = \\langle \\mathbf{k},\\omega \\rangle\\) \\(\\lambda_{\\mathbf{k}} = \\langle \\mathbf{k},\\varphi \\rangle +b_{i}\\) , and \\(\\alpha_{\\mathbf{k}} = \\prod_{j}J_{k_j}(W_{ij})\\) is the product of the Bessel functions of the first kind." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.558, + 0.484, + 0.693 + ], + "angle": 0, + "content": "The Bessel functions \\( J_{k} \\) appear in the Fourier series of \\( \\sin (a\\sin (x)) \\) [2, Page 361] and Thrm 1 generalizes this result. See the proof and generalization in the Supp. Mat. Yüce et al. [34] present a similar expansion for neurons activated by polynomials, however, they do not derive an amplitude formula. Their expansion for sinusoidal neurons is restricted to a 3-layer MLP without bias, a critical parameter for Fourier representation. Ours incorporates bias and generalizes to deep MLPs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.694, + 0.484, + 0.847 + ], + "angle": 0, + "content": "We now list some consequences of (2). First, the layer composition introduces a vast number of frequencies \\(\\beta_{\\mathbf{k}}\\) depending only on \\(\\omega\\), and shifts given by the input shift \\(\\varphi\\) and the bias \\(\\mathbf{b}\\). More precisely, truncating the expansion by summing over \\(|\\mathbf{k}|_{\\infty} \\leq B \\in \\mathbb{N}\\), implies that each hidden neuron \\(h_i\\) could learn \\(\\frac{(2B + 1)^m - 1}{2}\\) non-null frequencies. This frequency generation gives a novel explanation of why composing sinusoidal layers greatly increases the network capacity. We will explore (2) to define a robust initialization for the INR's input neurons in Sec 4.1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.847, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Secondly, note that the weights \\(\\mathbf{W}\\) fully determine the amplitude \\(\\alpha_{\\mathbf{k}}\\) of each harmonic \\(\\sin (\\beta_{\\mathbf{k}}\\mathbf{x} + \\lambda_{\\mathbf{k}})\\). Thus, to" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "derive our bounding scheme we need to focus only on \\(\\mathbf{W}\\). Finally, we can derive a sine-cosine form of (2)," + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.13, + 0.905, + 0.163 + ], + "angle": 0, + "content": "\\[\nh _ {i} (\\mathbf {x}) = \\sum_ {\\mathbf {k} \\in \\mathbb {Z} ^ {m}} A _ {\\mathbf {k}} \\sin \\left(\\beta_ {\\mathbf {k}} \\mathbf {x}\\right) + B _ {\\mathbf {k}} \\cos \\left(\\beta_ {\\mathbf {k}} \\mathbf {x}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.171, + 0.907, + 0.308 + ], + "angle": 0, + "content": "with \\( A_{\\mathbf{k}} = \\alpha_{\\mathbf{k}}\\cos (\\lambda_{\\mathbf{k}}) \\) and \\( B_{\\mathbf{k}} = \\alpha_{\\mathbf{k}}\\sin (\\lambda_{\\mathbf{k}}) \\). Note that the generated frequencies are independent of the index \\( i \\), thus the hidden neurons share the same harmonics: \\( \\sin (\\beta_{\\mathbf{k}}\\mathbf{x}) \\), \\( \\cos (\\beta_{\\mathbf{k}}\\mathbf{x}) \\). Since different combinations of the input frequencies may correspond to a single frequency, (3) isn't (yet) the Fourier transform of the network. In other words, we could have two integer vectors \\( \\mathbf{k},\\mathbf{l}\\in \\mathbb{Z}^{m} \\) such that \\( \\beta_{\\mathbf{k}} = \\beta_{\\mathbf{l}} \\). In the Supp. Mat., we show how to aggregate those frequencies to obtain the final Fourier transform." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.309, + 0.907, + 0.37 + ], + "angle": 0, + "content": "To control the MLP's bandlimit, we need a method to bound the amplitudes of \\(\\beta_{\\mathbf{k}}\\). Precisely, for a given number \\(B > 0\\) we would need \\(|\\alpha_{\\mathbf{k}}|\\) to be small for \\(|\\beta_{\\mathbf{k}}| \\geq B\\). For this, we use the formula of \\(\\alpha_{\\mathbf{k}}\\) to determine a bounding:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.377, + 0.906, + 0.421 + ], + "angle": 0, + "content": "Theorem 2. The magnitude of the amplitudes \\(\\alpha_{\\mathbf{k}}\\) in the expansion (2) is bounded by \\(\\prod_{j=1}^{m}\\left(\\frac{|W_{ij}|}{2}\\right)^{|k_j|}\\frac{1}{|k_j|!}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.425, + 0.906, + 0.456 + ], + "angle": 0, + "content": "Sec 4.2 presents a bandlimit control mechanism during training based on Thrm 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.468, + 0.857, + 0.485 + ], + "angle": 0, + "content": "4. Initialization and frequency bounding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.907, + 0.676 + ], + "angle": 0, + "content": "We present TUNER, an initialization and frequency control scheme for sinusoidal INRs. Our initialization considers integer input frequencies, based on a Fourier series interpretation, and uses Thrm 1 to study the spectrum of layer composition. Then, we use Thrm 2 to control the INR's bandlimit during training. Fig 2 provides an overview of TUNER. Throughout this section we present toy experiments to motivate the method, using \\( f: \\mathbb{R}^2 \\to \\mathbb{R}^3 \\) in an image reconstruction setup; later, comprehensive experiments will be given in Sec 5. We report the MLP size using the parameters \\( m \\), \\( n \\) as they completely determine the model. We use the Kodak dataset's images [1] and train the INR using Adam [9]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.683, + 0.81, + 0.699 + ], + "angle": 0, + "content": "4.1. Initialization as spectral sampling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.903 + ], + "angle": 0, + "content": "A key challenge in initializing a sinusoidal MLP lies in defining its input frequencies \\(\\omega\\). Recall that the MLP generates frequencies as integer combinations of \\(\\omega\\), i.e., \\(\\beta_{\\mathbf{k}} = \\sum_{i} k_{i} \\omega_{i}\\), with the remaining weights representing their amplitudes \\(\\alpha_{\\mathbf{k}}\\). This setup presents two main challenges. First, the initialization of \\(\\omega\\) must enable \\(\\beta_{\\mathbf{k}}\\) to cover the signal's spectrum; otherwise, the optimization cannot learn missing frequencies. Fig 3 (left) shows an example where using only odd frequencies leads to poor reconstruction. The second problem is the lack of flexibility for learning new frequencies. For example, randomly initializing \\(\\omega\\) may not be enough since it may produce many high frequencies in order to overfit the signal, see Fig 4 (top)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.886, + 0.318, + 0.902 + ], + "angle": 0, + "content": "\\(1|\\mathbf{k}|_{\\infty}\\) denotes \\(\\max \\{|k_1|,\\dots ,|k_m|\\}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3073" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.089, + 0.885, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.3, + 0.907, + 0.412 + ], + "angle": 0, + "content": "Figure 2. Overview of TUNER. To train a sinusoidal MLP (gray model, top-left), we employ two techniques derived from Thrms 1 and 2. First, we initialize the input frequencies \\(\\omega\\) (green, bottom-left) with a dense distribution of low frequencies (red square) and a sparse distribution of higher frequencies (green grid). This initialization gives flexibility to learn the remaining signal frequencies which are simply integer combinations of \\(\\omega\\) (the yellow nodes on the right), a consequence of the amplitude-phase expansion given by Theorem 1. Note that this initialization resembles a frequency sampling since the training generates those new frequencies around \\(\\omega\\). Second, we bound the coefficients of the hidden layer weights (blue nodes) to ensure that the MLP remains within a specified bandlimit. This approach is effective because the amplitude-phase expansion (shown on the right) of each hidden neuron (purple nodes) indicates that the amplitudes of the generated frequencies have an upper-bound depending only on the hidden weights (blue, bottom-right)." + }, + { + "type": "image", + "bbox": [ + 0.147, + 0.424, + 0.285, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.424, + 0.425, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.531, + 0.483, + 0.6 + ], + "angle": 0, + "content": "Figure 3. Choosing \\(\\omega\\) as the cartesian product of the odd frequencies without (left) or with (right) the frequencies \\((1,0),(0,1)\\). Note that adding them prevents sub-periods (see Supp. Mat.). We trained for 3000 epochs on a \\(256^{2}\\) image with network parameters \\(m = 72\\), \\(n = 512\\), and \\(\\theta = 30\\)." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.606, + 0.47, + 0.789 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.79, + 0.483, + 0.847 + ], + "angle": 0, + "content": "Figure 4. Uniform init. of \\(\\omega\\) (top) and ours (bottom). Grayscale bands show INR's gradient. Ours offers better signal/gradient reconstruction w/o gradient supervision. The MLPs (\\(m = 128\\), \\(n = 256\\)) with bandlimit \\(\\theta = 87\\) were trained for 3000 epochs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To address these challenges, we introduce a novel initialization. First, we restrict our MLP to periodic functions, ensuring the training domain is in a full period. Such" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.438, + 0.907, + 0.513 + ], + "angle": 0, + "content": "functions can be represented by sums of sines with integer frequencies, defining an orthogonal basis (Fourier series); This aligns well with our analysis (Thrm 1). The main challenge is initializing \\(\\omega\\) such that the generated frequencies \\(\\beta_{\\mathbf{k}}\\) cover the full spectrum within a bandlimit \\(B\\) (Fourier basis)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.906, + 0.603 + ], + "angle": 0, + "content": "More precisely, we initialize the weights \\(\\omega\\) with integer frequencies, that is \\(\\omega_{j}\\in \\frac{2\\pi}{p}\\mathbb{Z}^{d}\\), and freeze them during training, with \\(p > 0\\) being the period of the INR. This guarantees that the input neurons are \\(p\\)-periodic and Thrm 1 ensures that such periodicity is preserved over layer composition. Additionally, (3) implies that we can rewrite the INR as:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.608, + 0.907, + 0.641 + ], + "angle": 0, + "content": "\\[\nf (\\mathbf {x}) = \\sum_ {\\mathbf {k} \\in \\mathbb {Z} ^ {m}} \\left\\langle \\mathbf {C}, A _ {\\mathbf {k}} \\right\\rangle \\sin \\left(\\beta_ {\\mathbf {k}} \\mathbf {x}\\right) + \\left\\langle \\mathbf {C}, B _ {\\mathbf {k}} \\right\\rangle \\cos \\left(\\beta_ {\\mathbf {k}} \\mathbf {x}\\right) + e. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.906, + 0.691 + ], + "angle": 0, + "content": "Since the generated frequencies \\(\\beta_{\\mathbf{k}}\\) only depend on the frozen parameters \\(\\omega\\), the training is responsible for learning the amplitudes of the sine-cosine series in (4)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.691, + 0.907, + 0.813 + ], + "angle": 0, + "content": "Let \\(B\\) be the signal's bandlimit. To ensure that the generated frequencies do not bypass \\(B\\), we cannot sample \\(\\omega\\) directly on \\([-B, B]^d\\) since the MLP generates multiples of these frequencies. Therefore, we sample \\(\\omega_j\\) in \\(\\frac{2\\pi}{p} [-\\theta, \\theta]^d\\), with \\(\\theta \\in (0, B)\\) being a threshold for the input frequencies. Indeed, as the coefficients of hidden matrix \\(\\mathbf{W}\\) are limited by 2 (see Sec 4.2), the magnitudes of \\(A_{\\mathbf{k}}\\) and \\(B_{\\mathbf{k}}\\) decrease as \\(|\\mathbf{k}|_{\\infty}\\) increases. This is a consequence of Thrm 2:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.815, + 0.905, + 0.867 + ], + "angle": 0, + "content": "\\[\n\\left| \\alpha_ {\\mathbf {k}} \\right| \\leq \\prod_ {j = 1} ^ {m} \\frac {\\left(\\frac {\\left| W _ {i j} \\right|}{2}\\right) ^ {\\left| k _ {j} \\right|}}{\\left| k _ {j} \\right| !} \\leq \\frac {1}{\\left| k _ {1} \\right| ! \\dots \\left| k _ {m} \\right| !}; i = 1, \\dots , n. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Thus, the generated frequencies \\(\\beta_{\\mathbf{k}}\\) with small \\(\\mathbf{k}\\) have more influence over the expansion (4), mimicking Fourier series." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3074" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.091, + 0.223, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.091, + 0.349, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.091, + 0.475, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.195, + 0.483, + 0.266 + ], + "angle": 0, + "content": "Figure 5. Image reconstructions with bandlimit \\( B = 128 \\), varying \\( \\theta = 4, 43, 128 \\) with a network (\\( m = 80, n = 1000 \\)) trained over 3000 epochs. The gradient magnitude is shown on the left of each image. Note that smaller \\( \\theta \\) (middle-left) yield better reconstructions, while higher values of \\( \\theta \\) introduce noisy gradients (right)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.483, + 0.394 + ], + "angle": 0, + "content": "For example, if input frequencies include 1 and 20, frequencies like \\( 19 = 20 - 1 \\) and \\( 21 = 20 + 1 \\) can be represented through their combinations. Thus, low frequencies give flexibility in expanding around the input ones which motivates initializing more input frequencies near the origin. We found \\( \\theta = \\frac{B}{3} \\) to work well experimentally. Fig 5 shows that it is necessary, as higher \\( \\theta \\) makes it harder to learn the small frequencies, introducing noise." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.395, + 0.483, + 0.607 + ], + "angle": 0, + "content": "To initialize \\(\\omega \\in \\mathbb{R}^{m\\times d}\\), we need to choose \\(m\\) frequencies in \\([- \\theta, \\theta]^d\\). Taking into account the importance of the lower frequencies discussed above, we split it into regions \\(\\mathbf{L} = [-\\ell, \\ell]^d\\) of low and \\(\\mathbf{H} = [-\\theta, \\theta]^d \\setminus \\mathbf{L}\\) of high frequencies, \\(\\ell \\in (0, \\theta)\\). As \\(\\beta_{\\mathbf{k}}\\) only has a significant amplitude when \\(|\\mathbf{k}|_{\\infty}\\) is small (5), initializing with only low input frequencies may not allow the generation of high frequencies. Thus, we initialize some frequencies in \\(\\mathbf{H}\\) to cover the spectrum, and the rest in \\(\\mathbf{L}\\) to cover a neighborhood of the input frequencies. Specifically, we select \\(m_l < m\\) frequencies uniformly from \\(\\mathbf{L}\\). Then, we choose \\(m - m_l\\) frequencies from \\(\\mathbf{H}\\), spread in a grid-like fashion to cover \\([- \\theta, \\theta]^d\\). In our experiments we considered \\(m_l = 0.7m\\). Finally, since \\(\\mathbf{D}(\\mathbf{x}) = \\sin (\\omega \\mathbf{x} + \\varphi)\\), we initialize \\(\\varphi\\) uniformly in \\([- \\frac{\\pi}{2}, \\frac{\\pi}{2}]\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.608, + 0.483, + 0.699 + ], + "angle": 0, + "content": "When \\(\\omega_{j}\\in \\mathbf{L}\\), small variations in \\(k_{j}\\) results in a frequency near \\(\\beta_{\\mathbf{k}}\\). Indeed, Fig 6 shows how low frequencies (green square) introduce frequencies around the input frequencies (Fig 6a). Conversely, if \\(\\omega_{j}\\in \\mathbf{H}\\) then \\(\\beta_{\\mathbf{k} + \\mathbf{e}_j}\\) is far away from \\(\\beta_{\\mathbf{k}}\\); we note this in Fig 6b-d, where \\(\\omega_{j}\\) is the red point and the arrows show the generated frequencies \\(k_{j}\\omega_{j}\\)." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.704, + 0.189, + 0.779 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.191, + 0.704, + 0.286, + 0.779 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.704, + 0.384, + 0.779 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.704, + 0.481, + 0.779 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.788, + 0.483, + 0.844 + ], + "angle": 0, + "content": "Figure 6. INR spectrum during training. The input frequencies \\(\\omega\\) in green. During training new frequencies appear in a neighborhood of \\(\\omega\\). The red dot indicates a high input frequency \\(\\omega_{j}\\) and the red arrows show the generated frequencies \\(k_{j}\\omega_{j}, k_{j} = 1,2,3\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Finally, when sampling frequencies in \\(\\mathbf{L}\\), we include \\((1,0)\\) and \\((0,1)\\) to guarantee that any frequency could be generated and prevents the appearance of sub-periods as" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "shown at the beginning of this section in Fig 3 (left) (see also Supp. Mat for more details). Fig 3 (right) shows how this problem is solved using our initialization." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.137, + 0.905, + 0.213 + ], + "angle": 0, + "content": "The proposed initialization overcomes the noisy reconstruction issues that arise with naive uniform sampling in \\([- \\theta, \\theta]^d\\), as used in [26] (Fig 4 (top)). Fig 4 (bottom) demonstrates that our scheme significantly improves both signal and gradient reconstruction quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.221, + 0.861, + 0.237 + ], + "angle": 0, + "content": "4.2. Bounding and hidden layer initialization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.243, + 0.907, + 0.38 + ], + "angle": 0, + "content": "As shown in previous sections, sinusoidal INRs can represent detailed signals due to the wide number of frequencies generated by layer composition. However, they could manifest as noise in the reconstruction, see Fig 5 (right). Specifically, high input frequencies may generate even higher ones, bypassing the bandlimit \\( B \\). To avoid this, we use Thrm 2 to introduce a bounding scheme to limit the amplitudes. Precisely, Thrm 2 implies that in each hidden neuron \\( h_i \\), the amplitude of a generated frequency \\( \\beta_{\\mathbf{k}} \\) is limited by:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.389, + 0.905, + 0.428 + ], + "angle": 0, + "content": "\\[\n\\left| \\alpha_ {\\mathbf {k}} \\right| \\leq \\left(\\frac {\\left| \\mathbf {W} _ {i} \\right| _ {\\infty}}{2}\\right) ^ {\\left| \\mathbf {k} \\right| _ {1}} \\frac {1}{\\prod_ {j = 1} ^ {m} \\left| k _ {j} \\right| !}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.437, + 0.907, + 0.574 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_i\\) is the \\(i\\)-row of \\(\\mathbf{W}\\). Note that \\(|\\mathbf{W}|_{\\infty} \\leq 2\\) implies that the amplitude of any generated frequency \\(\\beta_{\\mathbf{k}}\\) rapidly decreases when the coordinates of \\(\\mathbf{k}\\) grow. Therefore, we use (6) to define a bounding scheme that only allows the appearance of generated frequencies with small multiples of \\(\\omega\\) which is beneficial as shown in previous sections. For this, we simply clamp the coefficients \\(W_{ij}\\) by a bound parameter \\(c \\in (0,2]\\) at each optimization step, enforcing \\(|\\mathbf{W}|_{\\infty} \\leq c\\). Fig 7 shows results of bounding \\(|\\mathbf{W}|\\) with different \\(c\\)." + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.591, + 0.653, + 0.707 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.656, + 0.59, + 0.759, + 0.706 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.59, + 0.868, + 0.706 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.715, + 0.907, + 0.757 + ], + "angle": 0, + "content": "Figure 7. Training a network \\( f \\) with \\( m = n = 256 \\) for 6000 epochs with \\( \\theta = 30 \\) and bounds \\( c = 0.1, 0.2, 0.5 \\). The bound over the hidden matrix \\( \\mathbf{W} \\) restricts the appearance of high frequencies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In Sec 4.1 we split the input frequencies \\(\\omega\\) in two sets: the low \\((|\\omega_{j}|_{\\infty}\\leq \\ell)\\) and high \\((|\\omega_{j}|_{\\infty} > \\ell)\\) frequencies. Additionally, we saw that the lower ones have a main role in the frequency generation, thus, we consider a higher threshold for them. For this, observe that each input frequency \\(\\omega_{j}\\) is associated with the \\(j\\)-column of \\(\\mathbf{W}\\). Then, bounding the columns of \\(\\mathbf{W}\\) related to high frequencies, we restrict the amplitudes of \\(\\beta_{\\mathbf{k}}\\) that could exceed the Nyquist limit \\(B\\). Conversely, since low frequencies are used to span" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3075" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.138 + ], + "angle": 0, + "content": "around \\(\\omega\\), we consider a higher bound for their columns. Precisely, we define two parameters \\(e_{\\mathbf{L}}\\) and \\(e_{\\mathbf{H}}\\) to bound the weights \\(W_{ij}\\) corresponding to \\(\\omega_{j}\\) in \\(\\mathbf{L}\\) and \\(\\mathbf{H}\\), respectively," + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.142, + 0.465, + 0.181 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left| \\omega_ {j} \\right| _ {\\infty} \\leq \\ell \\longrightarrow W _ {i j} = \\operatorname {c l a m p} \\left(W _ {i j}, \\left[ - e _ {\\mathbf {L}}, e _ {\\mathbf {L}} \\right]\\right), \\\\ \\left| \\omega_ {j} \\right| _ {\\infty} > \\theta \\longrightarrow W _ {i j} = \\operatorname {c l a m p} \\left(W _ {i j}, \\left[ - \\boldsymbol {c} _ {\\mathbf {H}}, \\boldsymbol {c} _ {\\mathbf {H}} \\right]\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.184, + 0.484, + 0.32 + ], + "angle": 0, + "content": "We use these bounds to initialize \\(\\mathbf{W}\\) such that each column is normal distributed with mean 0 and standard deviation \\(\\frac{c_{\\mathrm{L}}}{3}\\), with \\(c_{\\mathrm{L}}\\) being the bound corresponding to that column. Note that SIREN [26] initializes \\(\\mathbf{W} \\in \\mathbb{R}^{n \\times m}\\) uniformly in \\((- \\sqrt{6/m}, \\sqrt{6/m})\\), and as \\(|\\mathbf{W}|_{\\infty} \\leq \\sqrt{6/m} < 2\\) for \\(m > 1\\), it is bandlimited only at the beginning of training, since \\(\\mathbf{W}\\) is not controlled in any way during the optimization. Conversely, our initialization controls the bandlimit during all training and has faster convergence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.32, + 0.484, + 0.412 + ], + "angle": 0, + "content": "Additionally, we present a scheme to learn the bounding parameters during training. For this, we bound each \\(j\\)-column of the hidden matrix \\(\\mathbf{W}\\) using a tanh activation followed by multiplication with a trainable bounding parameter \\(e_j\\). Clearly, the resulting column is bounded by \\(e_j\\). In other words, we replace the hidden layer \\(\\sin (\\mathbf{W}\\mathbf{x} + \\mathbf{b})\\) by" + }, + { + "type": "equation", + "bbox": [ + 0.207, + 0.416, + 0.482, + 0.433 + ], + "angle": 0, + "content": "\\[\n\\sin (\\tanh (\\mathbf {W}) \\mathcal {C} \\mathbf {x} + \\mathbf {b}), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.438, + 0.484, + 0.514 + ], + "angle": 0, + "content": "where \\(\\mathcal{C}\\) is a diagonal matrix with each \\(jj\\)-entry being \\(c_{j}\\). To prevent the bounds from growing too large, we use Thrm 2 to define the regularization term as \\(\\mathcal{L}_{\\mathrm{reg}} = \\sum |c_j|\\). Otherwise, the hidden weights may grow unbounded, leading to the INR's overfit as in the usual training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.527, + 0.224, + 0.544 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.484, + 0.749 + ], + "angle": 0, + "content": "This section presents experiments regarding the initialization and bounding of frequencies of an INR \\( f \\) considering the images from Kodak dataset [1] resized to \\( [-1, 1]^2 \\). We describe the architecture of \\( f \\) with hidden matrix \\( \\mathbf{W} \\in \\mathbb{R}^{n \\times m} \\) with the parameters \\( m, n \\). We initialize \\( f \\) following the scheme in Sec 4.1 and optimize it using Adam [9] with learning rate \\( 10^{-4} \\). For the quantitative evaluations, we report the mean and standard deviation of the Peak Signal-to-Noise Ratio (PSNR) evaluated on the test set (10% random pixels). To show that TUNER adds high order regularity to the reconstruction, we compare the analytical gradient \\( \\nabla f \\) to the Sobel filter of the ground truth. All experiments were run in a 12 GB NVIDIA GPU (TITAN X Pascal)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.227, + 0.772 + ], + "angle": 0, + "content": "5.1. Initialization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "First, we compare our initialization of \\(\\omega\\) (36.2dB) against the uniform distribution (32.2dB). We obtain a 4dB improvement of PNSR on average (with a std of 1.6dB). We also present an ablation over the parameter \\(\\ell\\) which splits the square \\([- \\ell, \\ell]^2\\) in regions of low (L) and high (H) frequencies. Here we use \\(\\ell = 85\\) and \\(\\ell = 21\\), 42, 64, and sample \\(\\omega\\) with \\(30\\%\\), \\(50\\%\\), and \\(70\\%\\) of coordinates in H. Table 1 shows the performance of the INR." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.089, + 0.905, + 0.157 + ], + "angle": 0, + "content": "
% high freqs30%50%70%
\\( \\ell \\)SignalGradSignalGradSignalGrad
2135.226.935.226.635.026.0
4235.126.334.824.633.822.2
6434.225.233.321.630.815.0
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.16, + 0.907, + 0.202 + ], + "angle": 0, + "content": "Table 1. Training an INR with \\( m = n = 416 \\) parameters, for 400 epochs. Results indicate that it is better to sample more \\( \\omega_{j} \\in \\mathbf{L} \\) and choose a smaller \\( \\ell \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.905, + 0.303 + ], + "angle": 0, + "content": "This experiment considered all the images in the dataset. Observe that all values of the row \\(\\ell = 64\\) have lower PSNR, showing that an INR with very high frequencies performs worse even for small percentages \\((30\\%)\\). Similarly, the performance worsens when we increase the number of high frequencies (last column)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.304, + 0.906, + 0.44 + ], + "angle": 0, + "content": "We now test the influence of high frequencies by training an INR of size \\( m = 360 \\), \\( n = 512 \\) with \\( \\theta = 256 \\) over \\( 90\\% \\) of the ground truth pixels. We vary the parameter \\( \\ell = 60, 120, 220 \\) that defines low frequencies. Increasing the number of higher frequencies, the INR suffers from noise and overfitting. This can be noted in the reconstruction on the unsupervised pixels and gradients in Fig 8. Finally, in Supp. Mat. we compare our initialization of W against SIREN and observe a faster convergence." + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.455, + 0.655, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.656, + 0.456, + 0.758, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.456, + 0.864, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.606, + 0.906, + 0.634 + ], + "angle": 0, + "content": "Figure 8. Re constructions with big \\(\\ell\\) present overfitting on the \\(10\\%\\) unsupervised points. Gradient on the upper right corner." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.65, + 0.627, + 0.667 + ], + "angle": 0, + "content": "5.2. Bounding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.906, + 0.749 + ], + "angle": 0, + "content": "In Sec 4.2 we presented a scheme to bound the INR spectrum. Here, we show additional experiments to demonstrate that this improves the signal and has a great benefit over the unsupervised gradient reconstruction. Then, we give a comparison between fixed and trained bounding schemes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Fig 9 compares the reconstructions of an image of size \\(512^2\\) w/o/w our fixed bounding scheme. To demonstrate the impact of our scheme, we sample the input frequencies using \\(\\ell = 256\\) and \\(\\ell = 100\\). Without bounding (left), we effectively reconstruct the signal at the supervised pixels (30.8 dB), but with a noisy gradient (15.8 dB). That is, we are overfitting the signal. Conversely, bounding the hidden weights significantly improves both the signal (35.1 dB) and its gradient (27.8 dB). Our bounding scheme acts like a filter in both signal and gradient reconstruction." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3076" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.092, + 0.46, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.22, + 0.483, + 0.262 + ], + "angle": 0, + "content": "Figure 9. Comparison of training \\( \\mathrm{wo} / \\mathrm{w} \\) fixed bounds. Note that the use of bounds preserves high order information of the signal. Left side of the figures show the network gradients." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.282, + 0.484, + 0.465 + ], + "angle": 0, + "content": "We compare quantitatively our bounding scheme over the dataset, varying the parameters \\(\\theta\\) and \\(c_{\\mathbf{H}}\\). Table 2 shows an improvement in the signal for all cases. Furthermore, the unsupervised gradient of the image is better learned when using fixed bounds, showing an average increase of 9.6dB against the standard training. Also, the last two lines of the table show the impact of the parameters \\(c_{\\mathbf{L}}\\), \\(c_{\\mathbf{H}}\\) over the signal and gradient reconstructions. We consider training with \\(c_{\\mathbf{L}} > c_{\\mathbf{H}}\\) since we are prioritizing low input frequencies in the spectrum generation. Otherwise, this may generate high frequencies, possibly bypassing the Nyquist limit and resulting in a worse gradient reconstruction." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.48, + 0.487, + 0.548 + ], + "angle": 0, + "content": "
signalgrad
θcLcHwo boundw boundwo boundw bound
2561.50.0532.634.814.525.8
1901.50.0534.035.918.526.1
1901.50.234.034.418.521.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.549, + 0.484, + 0.62 + ], + "angle": 0, + "content": "Table 2. Training with bounding improves the signal and grad. reconstruction. The choice of the fixed bound \\( c_{\\mathbf{H}} \\) greatly impacts the grad. fidelity. The std of the difference between each image signal (grad) quality \\( w / w \\) bounding is 1.4dB (1.6dB) on average. Evaluations performed on the test set (10% of pixels)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Next, we show how we can learn each bound \\( c_{j} \\). As described in Sec 4.2, we can incorporate a learnable \\( c_{j} \\) for each \\( j \\)-column of the hidden matrix through a modified layer (7). Fig 10 shows a comparison between our fixed bound scheme using initial un-tuned \\( c_{\\mathbf{L}} = 1.0 \\) and \\( c_{\\mathbf{H}} = 0.6 \\) (left) and the learned bound scheme (middle)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "As expected, training lowers the bounds for high frequencies (H) while maintaining or increasing the bounds of low frequencies (L). Thus, we achieve better reconstruction in both the signal and gradient. In Fig 10 (right), we note that low frequencies (left of the green line) have significant bounds, showing their importance during training. These schemes allow us to initialize with higher frequencies than \\( B / 3 \\), accelerating convergence. Finally, we performed an ablation study on the regularization term \\( \\mathcal{L}_{reg} \\) to show its importance. Training w/wo regularization yielded 28.5/27.4dB for the RGB and 27.1/24.0dB for the gradient." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.091, + 0.905, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.201, + 0.908, + 0.26 + ], + "angle": 0, + "content": "Figure 10. Comparison of signal and gradient (grayscale) when training with fixed/learned bounds. The learned bounds adjust the INR spectrum resulting in better reconstructions. Each blue point represents a pair \\((|\\omega_j|_{\\infty}, c_j)\\), where \\(c_j\\) is the trained bound of \\(\\omega_j\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.285, + 0.655, + 0.302 + ], + "angle": 0, + "content": "5.3. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.307, + 0.907, + 0.566 + ], + "angle": 0, + "content": "We first compare our method (TUNER) with SIREN, which requires choosing a threshold parameter \\(\\theta\\), a critical task since higher \\(\\theta\\) implies noisy reconstructions. TUNER also uses \\(\\theta\\) to control the frequency generation, however, it provides more stable training with better reconstruction. To demonstrate this, we perform a comparison by varying \\(\\theta\\) and the number of input (\\(m\\)) and hidden (\\(n\\)) neurons. First, consider the case where we have no information about the signal bandlimit besides the Nyquist limit \\(B\\). Fig 11 presents a comparison between SIREN (top) and TUNER (bottom) with \\(\\theta = B\\). Note that SIREN produces noisy reconstructions, while ours learn faster and does not suffer from noise. This is due to our bounding scheme, which limits the generation of high frequencies and provides stable training (Sec 4.2). In contrast, SIREN initializes its parameters such that the spectrum is initially bounded but does not maintain this guarantee during training." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.576, + 0.909, + 0.744 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.749, + 0.908, + 0.82 + ], + "angle": 0, + "content": "Figure 11. Comparing SIREN (top) and TUNER (bottom) when we only know the Nyquist limit (i.e, \\(\\theta = B\\)). Since \\(\\theta\\) is big, SIREN introduces high frequencies at the beginning of training, resulting in very noisy reconstructions. Conversely, ours starts training low frequencies, converging faster and being robust to \\(\\theta\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.83, + 0.908, + 0.877 + ], + "angle": 0, + "content": "Next, we consider a tuned \\(\\ell\\) for SIREN to show that even under this condition, our method provides a better training scheme for sinusoidal INRs. Fig 12 presents the comparison" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.663, + 0.901 + ], + "angle": 0, + "content": "\\(^{2}\\)Denoted as \\(\\omega_0\\) in [26]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3077" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "varying the number of input \\((m)\\) and hidden \\((n)\\) neurons. We consider the cases \\(m < n\\) (top) and \\(m > n\\) (bottom), and TUNER performed better both in the signal/gradient reconstructions. Also, we note more robust training using TUNER, as shown in Fig 12(right)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.482, + 0.244 + ], + "angle": 0, + "content": "The previous experiment shows that initializing SIREN with a high \\(\\theta\\) results in overfitting, while TUNER provides better reconstruction. On the other hand, with the same architecture and a tuned \\(\\theta\\), our method also effectively learned the lower frequencies first and performed better (Fig 12)." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.256, + 0.48, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.415, + 0.483, + 0.457 + ], + "angle": 0, + "content": "Figure 12. Comparison with SIREN when \\(\\theta = 40\\). TUNER improves training convergence in both cases (\\(m < n\\), \\(m > n\\)) and stability when \\(m > n\\). INR gradient shown in grayscale." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.473, + 0.484, + 0.565 + ], + "angle": 0, + "content": "Next, we numerically evaluate these experiments, providing a quantitative comparison with SIREN and BACON. We compare different architectures \\((m,n)\\) and spectrum bandlimits \\(\\theta\\), evaluating the results in terms of signal and gradient PSNR over the supervised (Train) and unsupervised (Test) pixels. Table 3 presents these evaluations." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.575, + 0.49, + 0.726 + ], + "angle": 0, + "content": "
TrainTest
SignalGradientSignalGradient
θ(m,n)SIRENBACOURSSIRENBACOURSSIRENBACOURSSIRENBACOURS
851028,16232.9-32.926.5-28.831.9-32.225.5-28.2
416,41634.026.434.327.722.428.932.925.433.427.122.328.4
80,202432.2-33.126.6-28.531.3-32.326.0-28.0
1711028,16231.3-33.922.0-27.222.9-32.920.7-26.8
416,41629.229.334.417.426.727.026.726.233.116.226.326.4
80,202426.6-30.814.2-22.623.8-29.513.8-21.3
2561028,16226.8-33.312.8-24.724.1-31.912.7-23.9
416,41624.727.741.212.924.223.022.322.733.213.222.821.2
80,202425.1-30.013.0-15.224.1-27.113.1-14.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.737, + 0.483, + 0.78 + ], + "angle": 0, + "content": "Table 3. Comparison between SIREN, BACON (BAC), and our method. We use several architectures \\((m,n)\\) ,bandlimits \\((\\theta)\\) ,and evaluate over the train \\((90\\%)\\) and test \\((10\\%)\\) sets of signal/gradient." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.794, + 0.484, + 0.901 + ], + "angle": 0, + "content": "As expected, SIREN performs well with \\(\\theta = \\frac{B}{3}\\). In this case, TUNER is perceptually similar to SIREN but improves gradient reconstruction and consistently outperforms BACON. When \\(\\theta = 171\\), TUNER offers superior quality in both signal and gradient, showing greater robustness to noise and superior performance even with \\(\\theta \\geq B/3\\). Finally, with \\(\\theta = 256\\), our method surpasses SIREN in all cases." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "For BACON, our method achieves better signal reconstruction, with comparable gradients. However, the following experiment shows that BACON's reconstruction has visible defects in this scenario. Note that we do not supervise the training with gradient information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.185, + 0.907, + 0.381 + ], + "angle": 0, + "content": "Bandlimit comparison: Lastly, Fig 13 presents a qualitative comparison between BACON, BANF, and TUNER (Ours) with bandlimits \\(\\theta = 60, 130\\). We train all methods with 67K parameters. For BANF we use the linear filter and adjust grid resolution to match the bandlimits and the MLP size to approximate 67K parameters. Particularly, the spectra were computed with the same process for all three methods. Note that BACON truncates the spectrum with a hard filter, resulting in ringing artifacts (low bandlimit) and noisy reconstruction (high bandlimit). BANF shows distortions around the eye area for both filters, and the spectra show no clear bandlimit. Conversely, TUNER operates as a soft filter providing better reconstructions for all cases." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.383, + 0.645, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.383, + 0.766, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.383, + 0.887, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.518, + 0.645, + 0.652 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.518, + 0.766, + 0.652 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.518, + 0.887, + 0.652 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.657, + 0.905, + 0.727 + ], + "angle": 0, + "content": "Figure 13. Comparison between BACON, BANF, and TUNER, trained with bandlimits \\(\\theta = 60, 130\\). Note that BACON uses a box filter, generating ringing artifacts (left). BANF (middle) has artifacts near edges and ambiguous bandlimit. Then, TUNER (right) resembles a soft filter, improving quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.771, + 0.756 + ], + "angle": 0, + "content": "6. Conclusions and limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We presented a study based on Fourier theory for sinusoidal MLPs, which resulted in novel, robust schemes for both initialization and spectral control of such networks. While our experiments focused on images to validate our theoretical claims, future work includes applying our schemes to other signal types. Additionally, we have focused our studies on 3-layer MLPs, leaving the investigation of deep networks for future research. We consider this work a first step toward unraveling the powerful expressiveness of sinusoidal MLPs." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3078" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.482, + 0.122 + ], + "angle": 0, + "content": "Acknowledgments We would like to thank CAPES, FAPERJ, and Google for partially funding this work." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.137, + 0.188, + 0.152 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.162, + 0.483, + 0.188 + ], + "angle": 0, + "content": "[1] Kodak dataset: https://r0k.us/graphics/kodak/.3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.192, + 0.483, + 0.233 + ], + "angle": 0, + "content": "[2] Milton Abramowitz and Irene A Stegun. Handbook of mathematical functions with formulas, graphs, and mathematical tables. US Government printing office, 1964. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.235, + 0.483, + 0.303 + ], + "angle": 0, + "content": "[3] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.305, + 0.483, + 0.373 + ], + "angle": 0, + "content": "[4] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628-8638, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.376, + 0.484, + 0.432 + ], + "angle": 0, + "content": "[5] Yishun Dou, Zhong Zheng, Qiaoqiao Jin, and Bingbing Ni. Multiplicative fourier level of detail. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1808-1817, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.434, + 0.483, + 0.488 + ], + "angle": 0, + "content": "[6] Emilien Dupont, Adam Golinski, Milad Alizadeh, Yee Whye Teh, and Arnaud Doucet. Coin: Compression with implicit neural representations. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.49, + 0.483, + 0.532 + ], + "angle": 0, + "content": "[7] Rizal Fathony, Anit Kumar Sahu, Devin Willmott, and J Zico Kolter. Multiplicative filter networks. In International Conference on Learning Representations, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.534, + 0.483, + 0.587 + ], + "angle": 0, + "content": "[8] Adam Kania, Marko Mihajlovic, Sergey Prokudin, Jacek Tabor, Przemys' Spurek, et al. Fresh: Frequency shifting for accelerated neural representation learning. arXiv preprint arXiv:2410.05050, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.59, + 0.483, + 0.63 + ], + "angle": 0, + "content": "[9] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.3,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.633, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[10] David B Lindell, Dave Van Veen, Jeong Joon Park, and Gordon Wetzstein. Bacon: Band-limited coordinate networks for multiscale scene representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16252-16262, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.704, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[11] Zhen Liu, Hao Zhu, Qi Zhang, Jingde Fu, Weibing Deng, Zhan Ma, Yanwen Guo, and Xun Cao. Finer: Flexible spectral-bias tuning in implicit neural representation by variable-periodic activation functions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2713-2722, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[12] Ishit Mehta, Manmohan Chandraker, and Ravi Ramamoorthi. A level set theory for neural implicit evolution under explicit flows. In European Conference on Computer Vision, pages 711-729. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[13] Tiago Novello, Guilherme Schardong, Luiz Schirmer, Vinicius da Silva, Helio Lopes, and Luiz Velho. Exploring differential geometry in neural implicits. Computers & Graphics, 108:49-60, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.162, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.16 + ], + "angle": 0, + "content": "[14] Tiago Novello, Vinicius da Silva, Guilherme Schardong, Luiz Schirmer, Helio Lopes, and Luiz Velho. Neural implicit surface evolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14279-14289, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[15] Giambattista Parascandolo, Heikki Huttunen, and Tuomas Virtanen. Taming the waves: sine as activation function in deep neural networks. 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[16] Hallison Paz, Daniel Perazzo, Tiago Novello, Guilherme Schardong, Luiz Schirmer, Vinicius da Silva, Daniel Yukimura, Fabio Chagas, Helio Lopes, and Luiz Velho. Mrnet: Multiresolution sinusoidal neural networks. Computers & Graphics, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[17] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International Conference on Machine Learning, pages 5301-5310. PMLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[18] E. Romero, J. Sopena, R. Alquezar, and J. Moliner. Neural networks with periodic and monotonic activation functions: a comparative study in classification problems. 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.392, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[19] Vishwanath Saragadam, Jasper Tan, Guha Balakrishnan, Richard G Baraniuk, and Ashok Veeraraghavan. Miner: Multiscale implicit neural representation. In European Conference on Computer Vision, pages 318-333. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[20] Vishwanath Saragadam, Daniel LeJeune, Jasper Tan, Guha Balakrishnan, Ashok Veeraraghavan, and Richard G Baraniuk. Wire: Wavelet implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18507-18516, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.534, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[21] Hemanth Saratchandran, Sameera Ramasinghe, Violetta Shevchenko, Alexander Long, and Simon Lucey. A sampling theory perspective on activations for implicit neural representations. In International Conference on Machine Learning, pages 43422-43444. PMLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[22] Guilherme Schardong, Tiago Novello, Hallison Paz, Iurii Medvedev, Vinícius da Silva, Luiz Velho, and Nuno Gonçalves. Neural implicit morphing of face images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7321-7330, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[23] Luiz Schirmer, Tiago Novello, Guilherme Schardong, Vinicius da Silva, Helio Lopes, and Luiz Velho. How to train your (neural) dragon. In Conference on Graphics, Patterns and Images, 36. (SIBGRAPI), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[24] Akhmedkhan Shabanov, Shrisudhan Govindarajan, Cody Reading, Lily Goli, Daniel Rebain, Kwang Moo Yi, and Andrea Tagliasacchi. Banf: Band-limited neural fields for levels of detail reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20571-20580, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[25] Kathan Shah and Chawin Sitawarin. Spider: Semiperiodic damping-enabled object representation. arXiv preprint arXiv:2306.15242, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in neural" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3079" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "information processing systems, 33:7462-7473, 2020. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.163 + ], + "angle": 0, + "content": "[27] Ying Song, Jiaping Wang, Li-Yi Wei, and Wencheng Wang. Vector regression functions for texture compression. ACM Transactions on Graphics (TOG), 35(1):1-10, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[28] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[29] Zhijie Wu, Yuhe Jin, and Kwang Moo Yi. Neural fourier filter bank. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14153-14163, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[30] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, pages 641-676. Wiley Online Library, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[31] Guandao Yang, Serge Belongie, Bharath Hariharan, and Vladlen Koltun. Geometry processing with neural fields. Advances in Neural Information Processing Systems, 34: 22483-22497, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.501 + ], + "angle": 0, + "content": "[32] Guandao Yang, Sagie Benaim, Varun Jampani, Kyle Genova, Jonathan Barron, Thomas Funkhouser, Bharath Hariharan, and Serge Belongie. Polynomial neural fields for subband decomposition and manipulation. Advances in Neural Information Processing Systems, 35:4401-4415, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.503, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[33] Wang Yifan, Lukas Rahmann, and Olga Sorkine-hornung. Geometry-consistent neural shape representation with implicit displacement fields. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.56, + 0.482, + 0.628 + ], + "angle": 0, + "content": "[34] Gizem Yüce, Guillermo Ortiz-Jiménez, Beril Besbinar, and Pascal Frossard. A structured dictionary perspective on implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19228-19238, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.63, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[35] Andreas Zell, Nuri Benbarka, Timon Hoefer, and Hamd Ul Moqueet Riaz. Seeing implicit neural representations as fourier series. In 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). IEEE Computer Society, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.699 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3080" + } + ] +] \ No newline at end of file diff --git a/2025/Tuning the Frequencies_ Robust Training for Sinusoidal Neural Networks/79e8b6d6-50a7-4215-bd2a-fb5524e17539_origin.pdf b/2025/Tuning the Frequencies_ Robust Training for Sinusoidal Neural Networks/79e8b6d6-50a7-4215-bd2a-fb5524e17539_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6ad1fee5cecda948d905ad613ab008239f37799c --- /dev/null +++ b/2025/Tuning the Frequencies_ Robust Training for Sinusoidal Neural Networks/79e8b6d6-50a7-4215-bd2a-fb5524e17539_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d90650158e7df888fc96dce272448f75a083674d0a8834dba62a509559bb600 +size 7079339 diff --git a/2025/Tuning the Frequencies_ Robust Training for Sinusoidal Neural Networks/full.md b/2025/Tuning the Frequencies_ Robust Training for Sinusoidal Neural Networks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4ade3ce1f88dd5a7ffb8bd05eca7e99b49cc60a2 --- /dev/null +++ b/2025/Tuning the Frequencies_ Robust Training for Sinusoidal Neural Networks/full.md @@ -0,0 +1,317 @@ +# Tuning the Frequencies: Robust Training for Sinusoidal Neural Networks + +Tiago Novello $^{1,2,*}$ Diana Aldana $^{1,*}$ Andre Araujo $^{2}$ Luiz Velho $^{1}$ $^{1}$ IMPA ${}^{2}$ Google DeepMind + +# Abstract + +Sinusoidal neural networks have been shown effective as implicit neural representations (INRs) of low-dimensional signals, due to their smoothness and high representation capacity. However, initializing and training them remain empirical tasks which lack on deeper understanding to guide the learning process. To fill this gap, our work introduces a theoretical framework that explains the capacity property of sinusoidal networks and offers robust control mechanisms for initialization and training. Our analysis is based on a novel amplitude-phase expansion of the sinusoidal multilayer perceptron, showing how its layer compositions produce a large number of new frequencies expressed as integer combinations of the input frequencies. This relationship can be directly used to initialize the input neurons, as a form of spectral sampling, and to bound the network's spectrum while training. Our method, referred to as TUNER (TUNing sinusoidal nEtwoRks), greatly improves the stability and convergence of sinusoidal INR training, leading to detailed reconstructions, while preventing overfitting. + +# 1. Introduction + +Sinusoidal multilayer perceptrons (MLPs) emerged as powerful implicit neural representations (INRs) for low-dimensional signals [3, 11, 26, 35]. In this context, the INR $f$ should fit the input data $\{x_{i},\ell_{i}\}$ as close as possible, i.e. $f(x_{i})\approx \ell_{i}$ , without overfitting, thus encoding the signal implicitly in the MLP parameters. Therefore, two major properties are required: (1) $f$ needs high representation capacity to fit $\{x_i,\ell_i\}$ ; (2) $f$ should have bandlimit control to avoid frequencies bypassing the sampling rate. + +Training sinusoidal MLPs to satisfy the above properties is challenging, as their initialization and optimization process often lead to undesirable local minima [15]. Recent work made strides towards more effective learning of these models. For example, SIREN [26] proposed an initialization by projecting the input coordinates to a list of sines with frequencies randomly chosen in a range, similar to the Fourier feature mapping (FFM) approach [28]. This way, the model can reach high capacity, but may lead to over + +![](images/1e10ba604bbe333e12863f62e01cd1ab65339e0bd38c53310ba84272e6aaa4d2.jpg) +Figure 1. We present TUNER, a robust and theoretically grounded training technique for sinusoidal MLPs, overcoming challenges in initialization and enabling bandlimiting control. Our experiments showcase TUNER's strong initialization results against ReLU, FFM [28], and SIREN [26] (top), where all models use the same size and training conditions. TUNER achieves both fast and stable convergence (bottom-left) while reconstructing gradients without noise. We also compare with BACON [10] across two bandlimits (bottom-right), enhancing quality and avoiding ringing artifacts. + +fitting with high frequencies resulting in noisy reconstructions. Defining an effective range for the bandlimit initialization remains mostly empirical and often results in noise, as the role of layer composition in generating frequencies is not fully understood. Additionally, uniform initialization may introduce undesired high frequencies and make it harder to model lower ones. More recently, BACON [10] proposes tighter bandlimit control by applying multiplicative filter networks (MFNs) [7] to limit the signal spectrum + +with a box filter, hard-truncating the spectrum. While effective in many cases, this may lead to ringing artifacts. Also, this technique lacks non-linear activations (MFNs are not neural networks), preventing the representation of fine details as efficiently as sinusoidal MLPs with similar sizes. + +Overall, initializing and training sinusoidal MLPs remain empirical tasks which lack on deeper understanding to guide the learning. To approach these problems, we present TUNER (TUNing sinusoidal nEtwoRks), a training scheme for sinusoidal MLPs consisting of a robust initialization and bandlimit mechanism that greatly improves the capacity and convergence of INRs. In contrast to previous work, TUNER is grounded on a theoretical framework that guides our learning approach. We develop a new trigonometric identity (Thrm 1), resembling a Fourier series, and prove that such expansion produces a large number of frequencies in terms of the first layer (input frequencies). This explains the frequency generation process when composing layers, highlighting the importance of input layer initialization for improving capacity. Remarkably, the amplitudes of these frequencies are given by complex functions of the hidden weights, introducing a challenging training. To address this, we prove that those are bounded by a term depending only on the hidden weights (Thrm 2). We apply this result to control the bandlimit of the INR during training. Some of TUNER's results are showcased in Fig. 1, in comparison to previous work, illustrating substantial enhancements to sinusoidal INR modeling. In summary, our contributions are: + +- We introduce a novel trigonometric identity that expands any hidden neuron into a wide sum of sines with frequencies being integer combinations of the input frequencies (Fig 2). The corresponding amplitudes, determined by the hidden weights, have a useful upper bound, offering effective tools for controlling the resulting signal. +- The expansion enables a robust initialization scheme for sinusoidal INRs, resulting in a high capacity MLP that trains significantly faster than previous approaches. First, we initialize the input neurons, as they determine the INR spectrum. Next, we initialize the hidden weights, considering their role in setting the corresponding amplitudes. +- We leverage the amplitude's upper bound to control the bandlimit of a sinusoidal INR by designing schemes that bound the hidden weights during training. Together with our initialization, these bounds generate frequencies centered around the input, resulting in more stable training compared to previous approaches. + +# 2. Related works + +Implicit neural representations [8, 20, 21, 25, 30, 32] are a trending topic in machine learning, used to learn highly detailed signals in low-dimension domains. Current INR architectures use Fourier feature mappings [28] or sinusoidal activation functions [26] to bypass the spectral bias [17] + +common in ReLU multilayer perceptrons. The high representation capacity of sinusoidal INRs has motivated their use to represent a wide range of media objects. Examples include audio [6, 26], images [4, 27], face morphing [22], signed distance functions [10, 13, 23, 26, 31], displacement fields [33], surface animation [12, 14], multi-resolution signals [5, 10, 16, 19, 29], among others. Most of these exploit the derivatives of sinusoidal INRs in the loss functions. + +Initialization. Considering sinusoidal activation functions in neural networks is a classical problem [18]; however, these INRs have been regarded as difficult to train [15]. Sitzmann et al. [26] overcome this by presenting a specific initialization scheme that allows training sinusoidal INRs, avoiding instability and ensuring convergence. Despite these advances, in practice, the initialization of such INRs remains an empirical task. In this work, inspired by Fourier series theory, we present a novel initialization method that allows us to train INRs with great convergence. Recently, several works have addressed the representation problem of sinusoidal INRs from different perspectives. Zell et al. [35] approached this problem by observing that the first layer of a sinusoidal INR is similar to a Fourier feature mapping. Here, in addition to improving the initialization of sinusoidal INRs, we present a training scheme for bounding the spectra of these networks. + +Control of spectrum. One of the main drawbacks of sinusoidal INRs is the lack of frequency control. SIREN [26] addressed this by initializing the input frequencies uniformly in a range. While this ensures a bandlimit at the start of training, as it progresses, the layer composition introduces higher frequencies, leading to noisy reconstructions. We avoid this by providing a novel initialization for the first layer coupled with a bounding scheme which gives controls for limiting the MLP spectrum. + +Other works [5, 10] employed MFNs [7] to control the bandlimit by applying a filter on the spectrum. However, this strategy usually introduces reconstruction artifacts since MFNs hard-truncate the spectrum. BANF [24] employed a grid-based MLP with a spatial filter that leverages grid resolution to constrain the highest frequency in the network. In our experiments, we observe that this is prone to creating higher frequencies which propagate as artifacts. In contrast, TUNER, grounded in a theoretical result (Thrm 2), guarantees a bandlimited MLP, and serves as a soft filter, providing a representation without ringing artifacts. + +# 3. Sinusoidal MLPs as Fourier series + +This work addresses the problem of deriving an efficient training scheme with controlled bandlimit for sinusoidal MLPs. This section presents the mathematical definitions and novel formulas for approaching this task. + +Sinusoidal MLPs demonstrated high representational capacity with only a few hidden layers [19, 26, 35]. To un + +derstand how layer composition encodes this capacity, we propose to thoroughly investigate the structure of a 3-layer sinusoidal MLP $f: \mathbb{R}^d \to \mathbb{R}$ . For simplicity, we assume the codomain to be $\mathbb{R}$ , however the same analysis holds for dimension $> 1$ . More precisely, we consider $f(\mathbf{x}) = \mathbf{C} \circ \mathbf{S} \circ \mathbf{D}(\mathbf{x}) + e$ , with $\mathbf{D}(\mathbf{x}) = \sin (\omega \mathbf{x} + \varphi)$ being the input layer that projects $\mathbf{x}$ into a list of harmonics (input neurons) $D_{i}(\mathbf{x}) = \sin (\omega_{i}\mathbf{x} + \varphi_{i})$ with frequencies $\omega = (\omega_{1},\dots,\omega_{m}) \in \mathbb{R}^{m\times d}$ and shifts $\varphi \in \mathbb{R}^{m}$ . Layer $\mathbf{D}$ is then composed with $\mathbf{S}(\mathbf{x}) = \sin (\mathbf{W}\mathbf{x} + \mathbf{b})$ , where $\mathbf{W} \in \mathbb{R}^{n\times m}$ is the hidden matrix, and $\mathbf{b} \in \mathbb{R}^n$ the bias. Finally, $\mathbf{C} \cdot \mathbf{x} + e$ is an affine transformation with $\mathbf{C} \in \mathbb{R}^{n\times 1}$ and $e \in \mathbb{R}$ . + +We now present some properties of the sinusoidal layers, which play key roles in the representation, and give a reinterpretation of them in terms of the network parameters. First, note that the hidden neurons $\mathbf{h}(\mathbf{x})\coloneqq \mathbf{S}\circ \mathbf{D}(\mathbf{x})$ are defined by composing the dictionary $\mathbf{D}$ with the hidden sinusoidal layer which results in a list with elements + +$$ +h _ {i} (\mathbf {x}) = \sin \left(\sum_ {j = 1} ^ {m} W _ {i j} \sin \left(\omega_ {j} \mathbf {x} + \varphi_ {j}\right) + b _ {i}\right). \tag {1} +$$ + +The following is a key result of this work, which states that we can linearize a hidden neuron (1) as a sum of sines with frequencies and amplitudes determined by $\omega$ and $\mathbf{W}$ . + +Theorem 1. Each hidden neuron $h_i$ of a 3-layer sinusoidal MLP has an amplitude-phase expansion of the form + +$$ +h _ {i} (\mathbf {x}) = \sum_ {\mathbf {k} \in \mathbb {Z} ^ {m}} \alpha_ {\mathbf {k}} \sin \left(\beta_ {\mathbf {k}} \mathbf {x} + \lambda_ {\mathbf {k}}\right), \tag {2} +$$ + +where $\beta_{\mathbf{k}} = \langle \mathbf{k},\omega \rangle$ $\lambda_{\mathbf{k}} = \langle \mathbf{k},\varphi \rangle +b_{i}$ , and $\alpha_{\mathbf{k}} = \prod_{j}J_{k_j}(W_{ij})$ is the product of the Bessel functions of the first kind. + +The Bessel functions $J_{k}$ appear in the Fourier series of $\sin (a\sin (x))$ [2, Page 361] and Thrm 1 generalizes this result. See the proof and generalization in the Supp. Mat. Yüce et al. [34] present a similar expansion for neurons activated by polynomials, however, they do not derive an amplitude formula. Their expansion for sinusoidal neurons is restricted to a 3-layer MLP without bias, a critical parameter for Fourier representation. Ours incorporates bias and generalizes to deep MLPs. + +We now list some consequences of (2). First, the layer composition introduces a vast number of frequencies $\beta_{\mathbf{k}}$ depending only on $\omega$ , and shifts given by the input shift $\varphi$ and the bias $\mathbf{b}$ . More precisely, truncating the expansion by summing over $|\mathbf{k}|_{\infty} \leq B \in \mathbb{N}$ , implies that each hidden neuron $h_i$ could learn $\frac{(2B + 1)^m - 1}{2}$ non-null frequencies. This frequency generation gives a novel explanation of why composing sinusoidal layers greatly increases the network capacity. We will explore (2) to define a robust initialization for the INR's input neurons in Sec 4.1. + +Secondly, note that the weights $\mathbf{W}$ fully determine the amplitude $\alpha_{\mathbf{k}}$ of each harmonic $\sin (\beta_{\mathbf{k}}\mathbf{x} + \lambda_{\mathbf{k}})$ . Thus, to + +derive our bounding scheme we need to focus only on $\mathbf{W}$ . Finally, we can derive a sine-cosine form of (2), + +$$ +h _ {i} (\mathbf {x}) = \sum_ {\mathbf {k} \in \mathbb {Z} ^ {m}} A _ {\mathbf {k}} \sin \left(\beta_ {\mathbf {k}} \mathbf {x}\right) + B _ {\mathbf {k}} \cos \left(\beta_ {\mathbf {k}} \mathbf {x}\right), \tag {3} +$$ + +with $A_{\mathbf{k}} = \alpha_{\mathbf{k}}\cos (\lambda_{\mathbf{k}})$ and $B_{\mathbf{k}} = \alpha_{\mathbf{k}}\sin (\lambda_{\mathbf{k}})$ . Note that the generated frequencies are independent of the index $i$ , thus the hidden neurons share the same harmonics: $\sin (\beta_{\mathbf{k}}\mathbf{x})$ , $\cos (\beta_{\mathbf{k}}\mathbf{x})$ . Since different combinations of the input frequencies may correspond to a single frequency, (3) isn't (yet) the Fourier transform of the network. In other words, we could have two integer vectors $\mathbf{k},\mathbf{l}\in \mathbb{Z}^{m}$ such that $\beta_{\mathbf{k}} = \beta_{\mathbf{l}}$ . In the Supp. Mat., we show how to aggregate those frequencies to obtain the final Fourier transform. + +To control the MLP's bandlimit, we need a method to bound the amplitudes of $\beta_{\mathbf{k}}$ . Precisely, for a given number $B > 0$ we would need $|\alpha_{\mathbf{k}}|$ to be small for $|\beta_{\mathbf{k}}| \geq B$ . For this, we use the formula of $\alpha_{\mathbf{k}}$ to determine a bounding: + +Theorem 2. The magnitude of the amplitudes $\alpha_{\mathbf{k}}$ in the expansion (2) is bounded by $\prod_{j=1}^{m}\left(\frac{|W_{ij}|}{2}\right)^{|k_j|}\frac{1}{|k_j|!}$ . + +Sec 4.2 presents a bandlimit control mechanism during training based on Thrm 2. + +# 4. Initialization and frequency bounding + +We present TUNER, an initialization and frequency control scheme for sinusoidal INRs. Our initialization considers integer input frequencies, based on a Fourier series interpretation, and uses Thrm 1 to study the spectrum of layer composition. Then, we use Thrm 2 to control the INR's bandlimit during training. Fig 2 provides an overview of TUNER. Throughout this section we present toy experiments to motivate the method, using $f: \mathbb{R}^2 \to \mathbb{R}^3$ in an image reconstruction setup; later, comprehensive experiments will be given in Sec 5. We report the MLP size using the parameters $m$ , $n$ as they completely determine the model. We use the Kodak dataset's images [1] and train the INR using Adam [9]. + +# 4.1. Initialization as spectral sampling + +A key challenge in initializing a sinusoidal MLP lies in defining its input frequencies $\omega$ . Recall that the MLP generates frequencies as integer combinations of $\omega$ , i.e., $\beta_{\mathbf{k}} = \sum_{i} k_{i} \omega_{i}$ , with the remaining weights representing their amplitudes $\alpha_{\mathbf{k}}$ . This setup presents two main challenges. First, the initialization of $\omega$ must enable $\beta_{\mathbf{k}}$ to cover the signal's spectrum; otherwise, the optimization cannot learn missing frequencies. Fig 3 (left) shows an example where using only odd frequencies leads to poor reconstruction. The second problem is the lack of flexibility for learning new frequencies. For example, randomly initializing $\omega$ may not be enough since it may produce many high frequencies in order to overfit the signal, see Fig 4 (top). + +![](images/0a9150c266f167997b739e415fd292a20d6163031252f38a9c7519f10689866b.jpg) +Figure 2. Overview of TUNER. To train a sinusoidal MLP (gray model, top-left), we employ two techniques derived from Thrms 1 and 2. First, we initialize the input frequencies $\omega$ (green, bottom-left) with a dense distribution of low frequencies (red square) and a sparse distribution of higher frequencies (green grid). This initialization gives flexibility to learn the remaining signal frequencies which are simply integer combinations of $\omega$ (the yellow nodes on the right), a consequence of the amplitude-phase expansion given by Theorem 1. Note that this initialization resembles a frequency sampling since the training generates those new frequencies around $\omega$ . Second, we bound the coefficients of the hidden layer weights (blue nodes) to ensure that the MLP remains within a specified bandlimit. This approach is effective because the amplitude-phase expansion (shown on the right) of each hidden neuron (purple nodes) indicates that the amplitudes of the generated frequencies have an upper-bound depending only on the hidden weights (blue, bottom-right). + +![](images/416cd5539eac2060604bbdf4c606cd6f8c689f3555d123c83fb057826b9b26b2.jpg) +Figure 3. Choosing $\omega$ as the cartesian product of the odd frequencies without (left) or with (right) the frequencies $(1,0),(0,1)$ . Note that adding them prevents sub-periods (see Supp. Mat.). We trained for 3000 epochs on a $256^{2}$ image with network parameters $m = 72$ , $n = 512$ , and $\theta = 30$ . + +![](images/58d5f9b65d759c8a987ff00280879f99b71ee245652153557b706075598f8407.jpg) + +![](images/6884eee2b87978eb3e8c96218e154fd69d8dd5903f659e34f6a973962c374ee1.jpg) +Figure 4. Uniform init. of $\omega$ (top) and ours (bottom). Grayscale bands show INR's gradient. Ours offers better signal/gradient reconstruction w/o gradient supervision. The MLPs ( $m = 128$ , $n = 256$ ) with bandlimit $\theta = 87$ were trained for 3000 epochs. + +To address these challenges, we introduce a novel initialization. First, we restrict our MLP to periodic functions, ensuring the training domain is in a full period. Such + +functions can be represented by sums of sines with integer frequencies, defining an orthogonal basis (Fourier series); This aligns well with our analysis (Thrm 1). The main challenge is initializing $\omega$ such that the generated frequencies $\beta_{\mathbf{k}}$ cover the full spectrum within a bandlimit $B$ (Fourier basis). + +More precisely, we initialize the weights $\omega$ with integer frequencies, that is $\omega_{j}\in \frac{2\pi}{p}\mathbb{Z}^{d}$ , and freeze them during training, with $p > 0$ being the period of the INR. This guarantees that the input neurons are $p$ -periodic and Thrm 1 ensures that such periodicity is preserved over layer composition. Additionally, (3) implies that we can rewrite the INR as: + +$$ +f (\mathbf {x}) = \sum_ {\mathbf {k} \in \mathbb {Z} ^ {m}} \left\langle \mathbf {C}, A _ {\mathbf {k}} \right\rangle \sin \left(\beta_ {\mathbf {k}} \mathbf {x}\right) + \left\langle \mathbf {C}, B _ {\mathbf {k}} \right\rangle \cos \left(\beta_ {\mathbf {k}} \mathbf {x}\right) + e. \tag {4} +$$ + +Since the generated frequencies $\beta_{\mathbf{k}}$ only depend on the frozen parameters $\omega$ , the training is responsible for learning the amplitudes of the sine-cosine series in (4). + +Let $B$ be the signal's bandlimit. To ensure that the generated frequencies do not bypass $B$ , we cannot sample $\omega$ directly on $[-B, B]^d$ since the MLP generates multiples of these frequencies. Therefore, we sample $\omega_j$ in $\frac{2\pi}{p} [-\theta, \theta]^d$ , with $\theta \in (0, B)$ being a threshold for the input frequencies. Indeed, as the coefficients of hidden matrix $\mathbf{W}$ are limited by 2 (see Sec 4.2), the magnitudes of $A_{\mathbf{k}}$ and $B_{\mathbf{k}}$ decrease as $|\mathbf{k}|_{\infty}$ increases. This is a consequence of Thrm 2: + +$$ +\left| \alpha_ {\mathbf {k}} \right| \leq \prod_ {j = 1} ^ {m} \frac {\left(\frac {\left| W _ {i j} \right|}{2}\right) ^ {\left| k _ {j} \right|}}{\left| k _ {j} \right| !} \leq \frac {1}{\left| k _ {1} \right| ! \dots \left| k _ {m} \right| !}; i = 1, \dots , n. \tag {5} +$$ + +Thus, the generated frequencies $\beta_{\mathbf{k}}$ with small $\mathbf{k}$ have more influence over the expansion (4), mimicking Fourier series. + +![](images/a2b79bedba58f4a9879c8b3418ad7bce8da827d77250787bfe1220414a2d8b03.jpg) +Figure 5. Image reconstructions with bandlimit $B = 128$ , varying $\theta = 4, 43, 128$ with a network ( $m = 80, n = 1000$ ) trained over 3000 epochs. The gradient magnitude is shown on the left of each image. Note that smaller $\theta$ (middle-left) yield better reconstructions, while higher values of $\theta$ introduce noisy gradients (right). + +![](images/09c628756c6397fe6edc0030b950ab95f481b73fa5e61500eda2b50ff0e17673.jpg) + +![](images/5dc6289d670196bc6830ffd0a171a48069fec8b6f99f4551f29845591a7f068e.jpg) + +For example, if input frequencies include 1 and 20, frequencies like $19 = 20 - 1$ and $21 = 20 + 1$ can be represented through their combinations. Thus, low frequencies give flexibility in expanding around the input ones which motivates initializing more input frequencies near the origin. We found $\theta = \frac{B}{3}$ to work well experimentally. Fig 5 shows that it is necessary, as higher $\theta$ makes it harder to learn the small frequencies, introducing noise. + +To initialize $\omega \in \mathbb{R}^{m\times d}$ , we need to choose $m$ frequencies in $[- \theta, \theta]^d$ . Taking into account the importance of the lower frequencies discussed above, we split it into regions $\mathbf{L} = [-\ell, \ell]^d$ of low and $\mathbf{H} = [-\theta, \theta]^d \setminus \mathbf{L}$ of high frequencies, $\ell \in (0, \theta)$ . As $\beta_{\mathbf{k}}$ only has a significant amplitude when $|\mathbf{k}|_{\infty}$ is small (5), initializing with only low input frequencies may not allow the generation of high frequencies. Thus, we initialize some frequencies in $\mathbf{H}$ to cover the spectrum, and the rest in $\mathbf{L}$ to cover a neighborhood of the input frequencies. Specifically, we select $m_l < m$ frequencies uniformly from $\mathbf{L}$ . Then, we choose $m - m_l$ frequencies from $\mathbf{H}$ , spread in a grid-like fashion to cover $[- \theta, \theta]^d$ . In our experiments we considered $m_l = 0.7m$ . Finally, since $\mathbf{D}(\mathbf{x}) = \sin (\omega \mathbf{x} + \varphi)$ , we initialize $\varphi$ uniformly in $[- \frac{\pi}{2}, \frac{\pi}{2}]$ . + +When $\omega_{j}\in \mathbf{L}$ , small variations in $k_{j}$ results in a frequency near $\beta_{\mathbf{k}}$ . Indeed, Fig 6 shows how low frequencies (green square) introduce frequencies around the input frequencies (Fig 6a). Conversely, if $\omega_{j}\in \mathbf{H}$ then $\beta_{\mathbf{k} + \mathbf{e}_j}$ is far away from $\beta_{\mathbf{k}}$ ; we note this in Fig 6b-d, where $\omega_{j}$ is the red point and the arrows show the generated frequencies $k_{j}\omega_{j}$ . + +![](images/e0aef8b4f6a97780f8dc3c30b24b6d57cbfa742af0ec8dae5628a3155f1ac2ad.jpg) +Figure 6. INR spectrum during training. The input frequencies $\omega$ in green. During training new frequencies appear in a neighborhood of $\omega$ . The red dot indicates a high input frequency $\omega_{j}$ and the red arrows show the generated frequencies $k_{j}\omega_{j}, k_{j} = 1,2,3$ . + +![](images/826da1394588ff8d16dfea45c4bf8e4dbcc333346b54852888c3f07be7bcb506.jpg) + +![](images/17ff66eee1e89182eb16257218cdac11d36593df270ead08c3e3151964c7c5b7.jpg) + +![](images/6b6f30254f97045b646ad55426325e5014c40711f7d7844096e6d7de7a4060d7.jpg) +Figure 7. Training a network $f$ with $m = n = 256$ for 6000 epochs with $\theta = 30$ and bounds $c = 0.1, 0.2, 0.5$ . The bound over the hidden matrix $\mathbf{W}$ restricts the appearance of high frequencies. + +Finally, when sampling frequencies in $\mathbf{L}$ , we include $(1,0)$ and $(0,1)$ to guarantee that any frequency could be generated and prevents the appearance of sub-periods as + +shown at the beginning of this section in Fig 3 (left) (see also Supp. Mat for more details). Fig 3 (right) shows how this problem is solved using our initialization. + +The proposed initialization overcomes the noisy reconstruction issues that arise with naive uniform sampling in $[- \theta, \theta]^d$ , as used in [26] (Fig 4 (top)). Fig 4 (bottom) demonstrates that our scheme significantly improves both signal and gradient reconstruction quality. + +# 4.2. Bounding and hidden layer initialization + +As shown in previous sections, sinusoidal INRs can represent detailed signals due to the wide number of frequencies generated by layer composition. However, they could manifest as noise in the reconstruction, see Fig 5 (right). Specifically, high input frequencies may generate even higher ones, bypassing the bandlimit $B$ . To avoid this, we use Thrm 2 to introduce a bounding scheme to limit the amplitudes. Precisely, Thrm 2 implies that in each hidden neuron $h_i$ , the amplitude of a generated frequency $\beta_{\mathbf{k}}$ is limited by: + +$$ +\left| \alpha_ {\mathbf {k}} \right| \leq \left(\frac {\left| \mathbf {W} _ {i} \right| _ {\infty}}{2}\right) ^ {\left| \mathbf {k} \right| _ {1}} \frac {1}{\prod_ {j = 1} ^ {m} \left| k _ {j} \right| !}. \tag {6} +$$ + +where $\mathbf{W}_i$ is the $i$ -row of $\mathbf{W}$ . Note that $|\mathbf{W}|_{\infty} \leq 2$ implies that the amplitude of any generated frequency $\beta_{\mathbf{k}}$ rapidly decreases when the coordinates of $\mathbf{k}$ grow. Therefore, we use (6) to define a bounding scheme that only allows the appearance of generated frequencies with small multiples of $\omega$ which is beneficial as shown in previous sections. For this, we simply clamp the coefficients $W_{ij}$ by a bound parameter $c \in (0,2]$ at each optimization step, enforcing $|\mathbf{W}|_{\infty} \leq c$ . Fig 7 shows results of bounding $|\mathbf{W}|$ with different $c$ . + +![](images/3f38fc49434962dd9928526679512b7558ee3f4fba8ed076b1b9237eae2c7fa3.jpg) + +![](images/2507cf1002ca90190c6fdd41c8c50c2a8ddafe122d333c044dfb5305122ef734.jpg) + +![](images/8b45a8e6e57a66380bef200f415b384d883ed9550145f088a2a4abe47ebe4e68.jpg) + +In Sec 4.1 we split the input frequencies $\omega$ in two sets: the low $(|\omega_{j}|_{\infty}\leq \ell)$ and high $(|\omega_{j}|_{\infty} > \ell)$ frequencies. Additionally, we saw that the lower ones have a main role in the frequency generation, thus, we consider a higher threshold for them. For this, observe that each input frequency $\omega_{j}$ is associated with the $j$ -column of $\mathbf{W}$ . Then, bounding the columns of $\mathbf{W}$ related to high frequencies, we restrict the amplitudes of $\beta_{\mathbf{k}}$ that could exceed the Nyquist limit $B$ . Conversely, since low frequencies are used to span + +around $\omega$ , we consider a higher bound for their columns. Precisely, we define two parameters $e_{\mathbf{L}}$ and $e_{\mathbf{H}}$ to bound the weights $W_{ij}$ corresponding to $\omega_{j}$ in $\mathbf{L}$ and $\mathbf{H}$ , respectively, + +$$ +\begin{array}{l} \left| \omega_ {j} \right| _ {\infty} \leq \ell \longrightarrow W _ {i j} = \operatorname {c l a m p} \left(W _ {i j}, \left[ - e _ {\mathbf {L}}, e _ {\mathbf {L}} \right]\right), \\ \left| \omega_ {j} \right| _ {\infty} > \theta \longrightarrow W _ {i j} = \operatorname {c l a m p} \left(W _ {i j}, \left[ - \boldsymbol {c} _ {\mathbf {H}}, \boldsymbol {c} _ {\mathbf {H}} \right]\right). \\ \end{array} +$$ + +We use these bounds to initialize $\mathbf{W}$ such that each column is normal distributed with mean 0 and standard deviation $\frac{c_{\mathrm{L}}}{3}$ , with $c_{\mathrm{L}}$ being the bound corresponding to that column. Note that SIREN [26] initializes $\mathbf{W} \in \mathbb{R}^{n \times m}$ uniformly in $(- \sqrt{6/m}, \sqrt{6/m})$ , and as $|\mathbf{W}|_{\infty} \leq \sqrt{6/m} < 2$ for $m > 1$ , it is bandlimited only at the beginning of training, since $\mathbf{W}$ is not controlled in any way during the optimization. Conversely, our initialization controls the bandlimit during all training and has faster convergence. + +Additionally, we present a scheme to learn the bounding parameters during training. For this, we bound each $j$ -column of the hidden matrix $\mathbf{W}$ using a tanh activation followed by multiplication with a trainable bounding parameter $e_j$ . Clearly, the resulting column is bounded by $e_j$ . In other words, we replace the hidden layer $\sin (\mathbf{W}\mathbf{x} + \mathbf{b})$ by + +$$ +\sin (\tanh (\mathbf {W}) \mathcal {C} \mathbf {x} + \mathbf {b}), \tag {7} +$$ + +where $\mathcal{C}$ is a diagonal matrix with each $jj$ -entry being $c_{j}$ . To prevent the bounds from growing too large, we use Thrm 2 to define the regularization term as $\mathcal{L}_{\mathrm{reg}} = \sum |c_j|$ . Otherwise, the hidden weights may grow unbounded, leading to the INR's overfit as in the usual training. + +# 5. Experiments + +This section presents experiments regarding the initialization and bounding of frequencies of an INR $f$ considering the images from Kodak dataset [1] resized to $[-1, 1]^2$ . We describe the architecture of $f$ with hidden matrix $\mathbf{W} \in \mathbb{R}^{n \times m}$ with the parameters $m, n$ . We initialize $f$ following the scheme in Sec 4.1 and optimize it using Adam [9] with learning rate $10^{-4}$ . For the quantitative evaluations, we report the mean and standard deviation of the Peak Signal-to-Noise Ratio (PSNR) evaluated on the test set (10% random pixels). To show that TUNER adds high order regularity to the reconstruction, we compare the analytical gradient $\nabla f$ to the Sobel filter of the ground truth. All experiments were run in a 12 GB NVIDIA GPU (TITAN X Pascal). + +# 5.1. Initialization + +First, we compare our initialization of $\omega$ (36.2dB) against the uniform distribution (32.2dB). We obtain a 4dB improvement of PNSR on average (with a std of 1.6dB). We also present an ablation over the parameter $\ell$ which splits the square $[- \ell, \ell]^2$ in regions of low (L) and high (H) frequencies. Here we use $\ell = 85$ and $\ell = 21$ , 42, 64, and sample $\omega$ with $30\%$ , $50\%$ , and $70\%$ of coordinates in H. Table 1 shows the performance of the INR. + +
% high freqs30%50%70%
\( \ell \)SignalGradSignalGradSignalGrad
2135.226.935.226.635.026.0
4235.126.334.824.633.822.2
6434.225.233.321.630.815.0
+ +Table 1. Training an INR with $m = n = 416$ parameters, for 400 epochs. Results indicate that it is better to sample more $\omega_{j} \in \mathbf{L}$ and choose a smaller $\ell$ . + +This experiment considered all the images in the dataset. Observe that all values of the row $\ell = 64$ have lower PSNR, showing that an INR with very high frequencies performs worse even for small percentages $(30\%)$ . Similarly, the performance worsens when we increase the number of high frequencies (last column). + +We now test the influence of high frequencies by training an INR of size $m = 360$ , $n = 512$ with $\theta = 256$ over $90\%$ of the ground truth pixels. We vary the parameter $\ell = 60, 120, 220$ that defines low frequencies. Increasing the number of higher frequencies, the INR suffers from noise and overfitting. This can be noted in the reconstruction on the unsupervised pixels and gradients in Fig 8. Finally, in Supp. Mat. we compare our initialization of W against SIREN and observe a faster convergence. + +![](images/d152895bb551ad1e678149fb5197ba69982a42fe69ad3d4f8cffbe8b976889d7.jpg) +Figure 8. Re constructions with big $\ell$ present overfitting on the $10\%$ unsupervised points. Gradient on the upper right corner. + +![](images/ef215ca3acb12cbbf4fb9fd785a437353c53385489dee7244dda4bea2dd6ece7.jpg) + +![](images/4cac0607a8df965aaae5c75da9ba368b9c9af65f8e0f377c39ff3be2b84168c5.jpg) + +# 5.2. Bounding + +In Sec 4.2 we presented a scheme to bound the INR spectrum. Here, we show additional experiments to demonstrate that this improves the signal and has a great benefit over the unsupervised gradient reconstruction. Then, we give a comparison between fixed and trained bounding schemes. + +Fig 9 compares the reconstructions of an image of size $512^2$ w/o/w our fixed bounding scheme. To demonstrate the impact of our scheme, we sample the input frequencies using $\ell = 256$ and $\ell = 100$ . Without bounding (left), we effectively reconstruct the signal at the supervised pixels (30.8 dB), but with a noisy gradient (15.8 dB). That is, we are overfitting the signal. Conversely, bounding the hidden weights significantly improves both the signal (35.1 dB) and its gradient (27.8 dB). Our bounding scheme acts like a filter in both signal and gradient reconstruction. + +![](images/778310baad253422766e06d6928afb5b93abce72d0a58facd8551d044e068e22.jpg) +Figure 9. Comparison of training $\mathrm{wo} / \mathrm{w}$ fixed bounds. Note that the use of bounds preserves high order information of the signal. Left side of the figures show the network gradients. + +We compare quantitatively our bounding scheme over the dataset, varying the parameters $\theta$ and $c_{\mathbf{H}}$ . Table 2 shows an improvement in the signal for all cases. Furthermore, the unsupervised gradient of the image is better learned when using fixed bounds, showing an average increase of 9.6dB against the standard training. Also, the last two lines of the table show the impact of the parameters $c_{\mathbf{L}}$ , $c_{\mathbf{H}}$ over the signal and gradient reconstructions. We consider training with $c_{\mathbf{L}} > c_{\mathbf{H}}$ since we are prioritizing low input frequencies in the spectrum generation. Otherwise, this may generate high frequencies, possibly bypassing the Nyquist limit and resulting in a worse gradient reconstruction. + +
signalgrad
θcLcHwo boundw boundwo boundw bound
2561.50.0532.634.814.525.8
1901.50.0534.035.918.526.1
1901.50.234.034.418.521.0
+ +Table 2. Training with bounding improves the signal and grad. reconstruction. The choice of the fixed bound $c_{\mathbf{H}}$ greatly impacts the grad. fidelity. The std of the difference between each image signal (grad) quality $w / w$ bounding is 1.4dB (1.6dB) on average. Evaluations performed on the test set (10% of pixels). + +Next, we show how we can learn each bound $c_{j}$ . As described in Sec 4.2, we can incorporate a learnable $c_{j}$ for each $j$ -column of the hidden matrix through a modified layer (7). Fig 10 shows a comparison between our fixed bound scheme using initial un-tuned $c_{\mathbf{L}} = 1.0$ and $c_{\mathbf{H}} = 0.6$ (left) and the learned bound scheme (middle). + +As expected, training lowers the bounds for high frequencies (H) while maintaining or increasing the bounds of low frequencies (L). Thus, we achieve better reconstruction in both the signal and gradient. In Fig 10 (right), we note that low frequencies (left of the green line) have significant bounds, showing their importance during training. These schemes allow us to initialize with higher frequencies than $B / 3$ , accelerating convergence. Finally, we performed an ablation study on the regularization term $\mathcal{L}_{reg}$ to show its importance. Training w/wo regularization yielded 28.5/27.4dB for the RGB and 27.1/24.0dB for the gradient. + +![](images/5ef1823b9863aebbefc72d4e3b1f85b01bdc63b3fb68ee3be83ae389fc04298f.jpg) +Figure 10. Comparison of signal and gradient (grayscale) when training with fixed/learned bounds. The learned bounds adjust the INR spectrum resulting in better reconstructions. Each blue point represents a pair $(|\omega_j|_{\infty}, c_j)$ , where $c_j$ is the trained bound of $\omega_j$ . + +# 5.3. Comparisons + +We first compare our method (TUNER) with SIREN, which requires choosing a threshold parameter $\theta$ , a critical task since higher $\theta$ implies noisy reconstructions. TUNER also uses $\theta$ to control the frequency generation, however, it provides more stable training with better reconstruction. To demonstrate this, we perform a comparison by varying $\theta$ and the number of input ( $m$ ) and hidden ( $n$ ) neurons. First, consider the case where we have no information about the signal bandlimit besides the Nyquist limit $B$ . Fig 11 presents a comparison between SIREN (top) and TUNER (bottom) with $\theta = B$ . Note that SIREN produces noisy reconstructions, while ours learn faster and does not suffer from noise. This is due to our bounding scheme, which limits the generation of high frequencies and provides stable training (Sec 4.2). In contrast, SIREN initializes its parameters such that the spectrum is initially bounded but does not maintain this guarantee during training. + +![](images/421ae920815b31ddb6f15579eedb9daa82c3912508ad9fd1a99d11fbdc6c4ee7.jpg) +Figure 11. Comparing SIREN (top) and TUNER (bottom) when we only know the Nyquist limit (i.e, $\theta = B$ ). Since $\theta$ is big, SIREN introduces high frequencies at the beginning of training, resulting in very noisy reconstructions. Conversely, ours starts training low frequencies, converging faster and being robust to $\theta$ . + +Next, we consider a tuned $\ell$ for SIREN to show that even under this condition, our method provides a better training scheme for sinusoidal INRs. Fig 12 presents the comparison + +varying the number of input $(m)$ and hidden $(n)$ neurons. We consider the cases $m < n$ (top) and $m > n$ (bottom), and TUNER performed better both in the signal/gradient reconstructions. Also, we note more robust training using TUNER, as shown in Fig 12(right). + +The previous experiment shows that initializing SIREN with a high $\theta$ results in overfitting, while TUNER provides better reconstruction. On the other hand, with the same architecture and a tuned $\theta$ , our method also effectively learned the lower frequencies first and performed better (Fig 12). + +![](images/1df572d89bb0dc7231d1830ad4041c4b386deef8552365012b69a9f7d4c4d142.jpg) +Figure 12. Comparison with SIREN when $\theta = 40$ . TUNER improves training convergence in both cases ( $m < n$ , $m > n$ ) and stability when $m > n$ . INR gradient shown in grayscale. + +Next, we numerically evaluate these experiments, providing a quantitative comparison with SIREN and BACON. We compare different architectures $(m,n)$ and spectrum bandlimits $\theta$ , evaluating the results in terms of signal and gradient PSNR over the supervised (Train) and unsupervised (Test) pixels. Table 3 presents these evaluations. + +
TrainTest
SignalGradientSignalGradient
θ(m,n)SIRENBACOURSSIRENBACOURSSIRENBACOURSSIRENBACOURS
851028,16232.9-32.926.5-28.831.9-32.225.5-28.2
416,41634.026.434.327.722.428.932.925.433.427.122.328.4
80,202432.2-33.126.6-28.531.3-32.326.0-28.0
1711028,16231.3-33.922.0-27.222.9-32.920.7-26.8
416,41629.229.334.417.426.727.026.726.233.116.226.326.4
80,202426.6-30.814.2-22.623.8-29.513.8-21.3
2561028,16226.8-33.312.8-24.724.1-31.912.7-23.9
416,41624.727.741.212.924.223.022.322.733.213.222.821.2
80,202425.1-30.013.0-15.224.1-27.113.1-14.6
+ +Table 3. Comparison between SIREN, BACON (BAC), and our method. We use several architectures $(m,n)$ ,bandlimits $(\theta)$ ,and evaluate over the train $(90\%)$ and test $(10\%)$ sets of signal/gradient. + +As expected, SIREN performs well with $\theta = \frac{B}{3}$ . In this case, TUNER is perceptually similar to SIREN but improves gradient reconstruction and consistently outperforms BACON. When $\theta = 171$ , TUNER offers superior quality in both signal and gradient, showing greater robustness to noise and superior performance even with $\theta \geq B/3$ . Finally, with $\theta = 256$ , our method surpasses SIREN in all cases. + +For BACON, our method achieves better signal reconstruction, with comparable gradients. However, the following experiment shows that BACON's reconstruction has visible defects in this scenario. Note that we do not supervise the training with gradient information. + +Bandlimit comparison: Lastly, Fig 13 presents a qualitative comparison between BACON, BANF, and TUNER (Ours) with bandlimits $\theta = 60, 130$ . We train all methods with 67K parameters. For BANF we use the linear filter and adjust grid resolution to match the bandlimits and the MLP size to approximate 67K parameters. Particularly, the spectra were computed with the same process for all three methods. Note that BACON truncates the spectrum with a hard filter, resulting in ringing artifacts (low bandlimit) and noisy reconstruction (high bandlimit). BANF shows distortions around the eye area for both filters, and the spectra show no clear bandlimit. Conversely, TUNER operates as a soft filter providing better reconstructions for all cases. + +![](images/d55aa1a130db4305156402adf5fce5539d522f49acb971b45ad4c0beae64d813.jpg) + +![](images/e598c70896dabb7dbe5c39f301d0361f292391302c5e87ef5d76e3559e647e77.jpg) + +![](images/a26d5d0407ef594c50a29fa94e040994c0d5de8bfb2ae6a2df3615075852d5b2.jpg) + +![](images/f53784834ac65abc3e568fa9d79d8058eeb8a886e6c3b8b67e66c6968df1d1cf.jpg) +Figure 13. Comparison between BACON, BANF, and TUNER, trained with bandlimits $\theta = 60, 130$ . Note that BACON uses a box filter, generating ringing artifacts (left). BANF (middle) has artifacts near edges and ambiguous bandlimit. Then, TUNER (right) resembles a soft filter, improving quality. + +![](images/4c502bb2d474d7e6fd31efdbd0d9597287831b9cf1db675f569a9568efb9dd88.jpg) + +![](images/cddab5cc49355e18ad562340ff5054ef2fb3f4153d12fe3a514e777c09682106.jpg) + +# 6. Conclusions and limitations + +We presented a study based on Fourier theory for sinusoidal MLPs, which resulted in novel, robust schemes for both initialization and spectral control of such networks. While our experiments focused on images to validate our theoretical claims, future work includes applying our schemes to other signal types. Additionally, we have focused our studies on 3-layer MLPs, leaving the investigation of deep networks for future research. We consider this work a first step toward unraveling the powerful expressiveness of sinusoidal MLPs. + +Acknowledgments We would like to thank CAPES, FAPERJ, and Google for partially funding this work. + +# References + +[1] Kodak dataset: https://r0k.us/graphics/kodak/.3, 6 +[2] Milton Abramowitz and Irene A Stegun. Handbook of mathematical functions with formulas, graphs, and mathematical tables. US Government printing office, 1964. 3 +[3] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 1 +[4] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628-8638, 2021. 2 +[5] Yishun Dou, Zhong Zheng, Qiaoqiao Jin, and Bingbing Ni. Multiplicative fourier level of detail. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1808-1817, 2023. 2 +[6] Emilien Dupont, Adam Golinski, Milad Alizadeh, Yee Whye Teh, and Arnaud Doucet. Coin: Compression with implicit neural representations. In International Conference on Learning Representations, 2021. 2 +[7] Rizal Fathony, Anit Kumar Sahu, Devin Willmott, and J Zico Kolter. Multiplicative filter networks. In International Conference on Learning Representations, 2020. 1, 2 +[8] Adam Kania, Marko Mihajlovic, Sergey Prokudin, Jacek Tabor, Przemys' Spurek, et al. Fresh: Frequency shifting for accelerated neural representation learning. arXiv preprint arXiv:2410.05050, 2024. 2 +[9] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.3,6 +[10] David B Lindell, Dave Van Veen, Jeong Joon Park, and Gordon Wetzstein. Bacon: Band-limited coordinate networks for multiscale scene representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16252-16262, 2022. 1, 2 +[11] Zhen Liu, Hao Zhu, Qi Zhang, Jingde Fu, Weibing Deng, Zhan Ma, Yanwen Guo, and Xun Cao. Finer: Flexible spectral-bias tuning in implicit neural representation by variable-periodic activation functions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2713-2722, 2024. 1 +[12] Ishit Mehta, Manmohan Chandraker, and Ravi Ramamoorthi. A level set theory for neural implicit evolution under explicit flows. In European Conference on Computer Vision, pages 711-729. Springer, 2022. 2 +[13] Tiago Novello, Guilherme Schardong, Luiz Schirmer, Vinicius da Silva, Helio Lopes, and Luiz Velho. Exploring differential geometry in neural implicits. Computers & Graphics, 108:49-60, 2022. 2 + +[14] Tiago Novello, Vinicius da Silva, Guilherme Schardong, Luiz Schirmer, Helio Lopes, and Luiz Velho. Neural implicit surface evolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14279-14289, 2023. 2 +[15] Giambattista Parascandolo, Heikki Huttunen, and Tuomas Virtanen. Taming the waves: sine as activation function in deep neural networks. 2016. 1, 2 +[16] Hallison Paz, Daniel Perazzo, Tiago Novello, Guilherme Schardong, Luiz Schirmer, Vinicius da Silva, Daniel Yukimura, Fabio Chagas, Helio Lopes, and Luiz Velho. Mrnet: Multiresolution sinusoidal neural networks. Computers & Graphics, 2023. 2 +[17] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International Conference on Machine Learning, pages 5301-5310. PMLR, 2019. 2 +[18] E. Romero, J. Sopena, R. Alquezar, and J. Moliner. Neural networks with periodic and monotonic activation functions: a comparative study in classification problems. 2000. 2 +[19] Vishwanath Saragadam, Jasper Tan, Guha Balakrishnan, Richard G Baraniuk, and Ashok Veeraraghavan. Miner: Multiscale implicit neural representation. In European Conference on Computer Vision, pages 318-333. Springer, 2022. 2 +[20] Vishwanath Saragadam, Daniel LeJeune, Jasper Tan, Guha Balakrishnan, Ashok Veeraraghavan, and Richard G Baraniuk. Wire: Wavelet implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18507-18516, 2023. 2 +[21] Hemanth Saratchandran, Sameera Ramasinghe, Violetta Shevchenko, Alexander Long, and Simon Lucey. A sampling theory perspective on activations for implicit neural representations. In International Conference on Machine Learning, pages 43422-43444. PMLR, 2024. 2 +[22] Guilherme Schardong, Tiago Novello, Hallison Paz, Iurii Medvedev, Vinícius da Silva, Luiz Velho, and Nuno Gonçalves. Neural implicit morphing of face images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7321-7330, 2024. 2 +[23] Luiz Schirmer, Tiago Novello, Guilherme Schardong, Vinicius da Silva, Helio Lopes, and Luiz Velho. How to train your (neural) dragon. In Conference on Graphics, Patterns and Images, 36. (SIBGRAPI), 2023. 2 +[24] Akhmedkhan Shabanov, Shrisudhan Govindarajan, Cody Reading, Lily Goli, Daniel Rebain, Kwang Moo Yi, and Andrea Tagliasacchi. Banf: Band-limited neural fields for levels of detail reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20571-20580, 2024. 2 +[25] Kathan Shah and Chawin Sitawarin. Spider: Semiperiodic damping-enabled object representation. arXiv preprint arXiv:2306.15242, 2023. 2 +[26] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in neural + +information processing systems, 33:7462-7473, 2020. 1, 2, 5, 6, 7 +[27] Ying Song, Jiaping Wang, Li-Yi Wei, and Wencheng Wang. Vector regression functions for texture compression. ACM Transactions on Graphics (TOG), 35(1):1-10, 2015. 2 +[28] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 1, 2 +[29] Zhijie Wu, Yuhe Jin, and Kwang Moo Yi. Neural fourier filter bank. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14153-14163, 2023. 2 +[30] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, pages 641-676. Wiley Online Library, 2022. 2 +[31] Guandao Yang, Serge Belongie, Bharath Hariharan, and Vladlen Koltun. Geometry processing with neural fields. Advances in Neural Information Processing Systems, 34: 22483-22497, 2021. 2 +[32] Guandao Yang, Sagie Benaim, Varun Jampani, Kyle Genova, Jonathan Barron, Thomas Funkhouser, Bharath Hariharan, and Serge Belongie. Polynomial neural fields for subband decomposition and manipulation. Advances in Neural Information Processing Systems, 35:4401-4415, 2022. 2 +[33] Wang Yifan, Lukas Rahmann, and Olga Sorkine-hornung. Geometry-consistent neural shape representation with implicit displacement fields. In International Conference on Learning Representations, 2021. 2 +[34] Gizem Yüce, Guillermo Ortiz-Jiménez, Beril Besbinar, and Pascal Frossard. A structured dictionary perspective on implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19228-19238, 2022. 3 +[35] Andreas Zell, Nuri Benbarka, Timon Hoefer, and Hamd Ul Moqueet Riaz. Seeing implicit neural representations as fourier series. In 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). IEEE Computer Society, 2022. 1, 2 \ No newline at end of file diff --git a/2025/Tuning the Frequencies_ Robust Training for Sinusoidal Neural Networks/images.zip b/2025/Tuning the Frequencies_ Robust Training for Sinusoidal Neural Networks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..322816464585fd799142ad2ae352e74c2399b38a --- /dev/null +++ b/2025/Tuning the Frequencies_ Robust Training for Sinusoidal Neural Networks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1512b5f7404dae3055703672a55b443a42fcaa473f5613fa0c4f049641f7e8d2 +size 697626 diff --git a/2025/Tuning the Frequencies_ Robust Training for Sinusoidal Neural Networks/layout.json b/2025/Tuning the Frequencies_ Robust Training for Sinusoidal Neural Networks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b7a45f59d28d2c72ca3fda3cacd0dc8da1f21df4 --- /dev/null +++ b/2025/Tuning the Frequencies_ Robust Training for Sinusoidal Neural Networks/layout.json @@ -0,0 +1,12635 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 103, + 533, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 103, + 533, + 121 + ], + "spans": [ + { + "bbox": [ + 78, + 103, + 533, + 121 + ], + "type": "text", + "content": "Tuning the Frequencies: Robust Training for Sinusoidal Neural Networks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 162, + 144, + 447, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 144, + 447, + 171 + ], + "spans": [ + { + "bbox": [ + 162, + 144, + 447, + 171 + ], + "type": "text", + "content": "Tiago Novello" + }, + { + "bbox": [ + 162, + 144, + 447, + 171 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 162, + 144, + 447, + 171 + ], + "type": "text", + "content": " Diana Aldana" + }, + { + "bbox": [ + 162, + 144, + 447, + 171 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 162, + 144, + 447, + 171 + ], + "type": "text", + "content": " Andre Araujo" + }, + { + "bbox": [ + 162, + 144, + 447, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 162, + 144, + 447, + 171 + ], + "type": "text", + "content": " Luiz Velho" + }, + { + "bbox": [ + 162, + 144, + 447, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 162, + 144, + 447, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 162, + 144, + 447, + 171 + ], + "type": "text", + "content": " IMPA " + }, + { + "bbox": [ + 162, + 144, + 447, + 171 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 162, + 144, + 447, + 171 + ], + "type": "text", + "content": " Google DeepMind" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 205, + 200, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 205, + 200, + 217 + ], + "spans": [ + { + "bbox": [ + 151, + 205, + 200, + 217 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 228, + 297, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 228, + 297, + 456 + ], + "spans": [ + { + "bbox": [ + 54, + 228, + 297, + 456 + ], + "type": "text", + "content": "Sinusoidal neural networks have been shown effective as implicit neural representations (INRs) of low-dimensional signals, due to their smoothness and high representation capacity. However, initializing and training them remain empirical tasks which lack on deeper understanding to guide the learning process. To fill this gap, our work introduces a theoretical framework that explains the capacity property of sinusoidal networks and offers robust control mechanisms for initialization and training. Our analysis is based on a novel amplitude-phase expansion of the sinusoidal multilayer perceptron, showing how its layer compositions produce a large number of new frequencies expressed as integer combinations of the input frequencies. This relationship can be directly used to initialize the input neurons, as a form of spectral sampling, and to bound the network's spectrum while training. Our method, referred to as TUNER (TUNing sinusoidal nEtwoRks), greatly improves the stability and convergence of sinusoidal INR training, leading to detailed reconstructions, while preventing overfitting." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 464, + 135, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 135, + 476 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 135, + 476 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 479, + 296, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 296, + 586 + ], + "type": "text", + "content": "Sinusoidal multilayer perceptrons (MLPs) emerged as powerful implicit neural representations (INRs) for low-dimensional signals [3, 11, 26, 35]. In this context, the INR " + }, + { + "bbox": [ + 55, + 479, + 296, + 586 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 479, + 296, + 586 + ], + "type": "text", + "content": " should fit the input data " + }, + { + "bbox": [ + 55, + 479, + 296, + 586 + ], + "type": "inline_equation", + "content": "\\{x_{i},\\ell_{i}\\}" + }, + { + "bbox": [ + 55, + 479, + 296, + 586 + ], + "type": "text", + "content": " as close as possible, i.e. " + }, + { + "bbox": [ + 55, + 479, + 296, + 586 + ], + "type": "inline_equation", + "content": "f(x_{i})\\approx \\ell_{i}" + }, + { + "bbox": [ + 55, + 479, + 296, + 586 + ], + "type": "text", + "content": ", without overfitting, thus encoding the signal implicitly in the MLP parameters. Therefore, two major properties are required: (1) " + }, + { + "bbox": [ + 55, + 479, + 296, + 586 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 479, + 296, + 586 + ], + "type": "text", + "content": " needs high representation capacity to fit " + }, + { + "bbox": [ + 55, + 479, + 296, + 586 + ], + "type": "inline_equation", + "content": "\\{x_i,\\ell_i\\}" + }, + { + "bbox": [ + 55, + 479, + 296, + 586 + ], + "type": "text", + "content": "; (2) " + }, + { + "bbox": [ + 55, + 479, + 296, + 586 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 479, + 296, + 586 + ], + "type": "text", + "content": " should have bandlimit control to avoid frequencies bypassing the sampling rate." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 586, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 296, + 694 + ], + "type": "text", + "content": "Training sinusoidal MLPs to satisfy the above properties is challenging, as their initialization and optimization process often lead to undesirable local minima [15]. Recent work made strides towards more effective learning of these models. For example, SIREN [26] proposed an initialization by projecting the input coordinates to a list of sines with frequencies randomly chosen in a range, similar to the Fourier feature mapping (FFM) approach [28]. This way, the model can reach high capacity, but may lead to over" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 204, + 552, + 492 + ], + "blocks": [ + { + "bbox": [ + 318, + 204, + 552, + 492 + ], + "lines": [ + { + "bbox": [ + 318, + 204, + 552, + 492 + ], + "spans": [ + { + "bbox": [ + 318, + 204, + 552, + 492 + ], + "type": "image", + "image_path": "1e10ba604bbe333e12863f62e01cd1ab65339e0bd38c53310ba84272e6aaa4d2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 497, + 555, + 597 + ], + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 597 + ], + "type": "text", + "content": "Figure 1. We present TUNER, a robust and theoretically grounded training technique for sinusoidal MLPs, overcoming challenges in initialization and enabling bandlimiting control. Our experiments showcase TUNER's strong initialization results against ReLU, FFM [28], and SIREN [26] (top), where all models use the same size and training conditions. TUNER achieves both fast and stable convergence (bottom-left) while reconstructing gradients without noise. We also compare with BACON [10] across two bandlimits (bottom-right), enhancing quality and avoiding ringing artifacts." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "fitting with high frequencies resulting in noisy reconstructions. Defining an effective range for the bandlimit initialization remains mostly empirical and often results in noise, as the role of layer composition in generating frequencies is not fully understood. Additionally, uniform initialization may introduce undesired high frequencies and make it harder to model lower ones. More recently, BACON [10] proposes tighter bandlimit control by applying multiplicative filter networks (MFNs) [7] to limit the signal spectrum" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 223, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 223, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 223, + 713 + ], + "type": "text", + "content": "*These authors contributed equally to this work." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3071" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "content": "with a box filter, hard-truncating the spectrum. While effective in many cases, this may lead to ringing artifacts. Also, this technique lacks non-linear activations (MFNs are not neural networks), preventing the representation of fine details as efficiently as sinusoidal MLPs with similar sizes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 133, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 133, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 133, + 295, + 407 + ], + "type": "text", + "content": "Overall, initializing and training sinusoidal MLPs remain empirical tasks which lack on deeper understanding to guide the learning. To approach these problems, we present TUNER (TUNing sinusoidal nEtwoRks), a training scheme for sinusoidal MLPs consisting of a robust initialization and bandlimit mechanism that greatly improves the capacity and convergence of INRs. In contrast to previous work, TUNER is grounded on a theoretical framework that guides our learning approach. We develop a new trigonometric identity (Thrm 1), resembling a Fourier series, and prove that such expansion produces a large number of frequencies in terms of the first layer (input frequencies). This explains the frequency generation process when composing layers, highlighting the importance of input layer initialization for improving capacity. Remarkably, the amplitudes of these frequencies are given by complex functions of the hidden weights, introducing a challenging training. To address this, we prove that those are bounded by a term depending only on the hidden weights (Thrm 2). We apply this result to control the bandlimit of the INR during training. Some of TUNER's results are showcased in Fig. 1, in comparison to previous work, illustrating substantial enhancements to sinusoidal INR modeling. In summary, our contributions are:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 407, + 294, + 622 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 407, + 294, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 407, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 56, + 407, + 294, + 479 + ], + "type": "text", + "content": "- We introduce a novel trigonometric identity that expands any hidden neuron into a wide sum of sines with frequencies being integer combinations of the input frequencies (Fig 2). The corresponding amplitudes, determined by the hidden weights, have a useful upper bound, offering effective tools for controlling the resulting signal." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 479, + 294, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 550 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 550 + ], + "type": "text", + "content": "- The expansion enables a robust initialization scheme for sinusoidal INRs, resulting in a high capacity MLP that trains significantly faster than previous approaches. First, we initialize the input neurons, as they determine the INR spectrum. Next, we initialize the hidden weights, considering their role in setting the corresponding amplitudes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 552, + 294, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 552, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 552, + 294, + 622 + ], + "type": "text", + "content": "- We leverage the amplitude's upper bound to control the bandlimit of a sinusoidal INR by designing schemes that bound the hidden weights during training. Together with our initialization, these bounds generate frequencies centered around the input, resulting in more stable training compared to previous approaches." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 633, + 144, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 144, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 144, + 645 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 294, + 713 + ], + "type": "text", + "content": "Implicit neural representations [8, 20, 21, 25, 30, 32] are a trending topic in machine learning, used to learn highly detailed signals in low-dimension domains. Current INR architectures use Fourier feature mappings [28] or sinusoidal activation functions [26] to bypass the spectral bias [17]" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "common in ReLU multilayer perceptrons. The high representation capacity of sinusoidal INRs has motivated their use to represent a wide range of media objects. Examples include audio [6, 26], images [4, 27], face morphing [22], signed distance functions [10, 13, 23, 26, 31], displacement fields [33], surface animation [12, 14], multi-resolution signals [5, 10, 16, 19, 29], among others. Most of these exploit the derivatives of sinusoidal INRs in the loss functions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 168, + 553, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 370 + ], + "type": "text", + "content": "Initialization. Considering sinusoidal activation functions in neural networks is a classical problem [18]; however, these INRs have been regarded as difficult to train [15]. Sitzmann et al. [26] overcome this by presenting a specific initialization scheme that allows training sinusoidal INRs, avoiding instability and ensuring convergence. Despite these advances, in practice, the initialization of such INRs remains an empirical task. In this work, inspired by Fourier series theory, we present a novel initialization method that allows us to train INRs with great convergence. Recently, several works have addressed the representation problem of sinusoidal INRs from different perspectives. Zell et al. [35] approached this problem by observing that the first layer of a sinusoidal INR is similar to a Fourier feature mapping. Here, in addition to improving the initialization of sinusoidal INRs, we present a training scheme for bounding the spectra of these networks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 371, + 553, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 553, + 479 + ], + "type": "text", + "content": "Control of spectrum. One of the main drawbacks of sinusoidal INRs is the lack of frequency control. SIREN [26] addressed this by initializing the input frequencies uniformly in a range. While this ensures a bandlimit at the start of training, as it progresses, the layer composition introduces higher frequencies, leading to noisy reconstructions. We avoid this by providing a novel initialization for the first layer coupled with a bounding scheme which gives controls for limiting the MLP spectrum." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 479, + 553, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 553, + 611 + ], + "type": "text", + "content": "Other works [5, 10] employed MFNs [7] to control the bandlimit by applying a filter on the spectrum. However, this strategy usually introduces reconstruction artifacts since MFNs hard-truncate the spectrum. BANF [24] employed a grid-based MLP with a spatial filter that leverages grid resolution to constrain the highest frequency in the network. In our experiments, we observe that this is prone to creating higher frequencies which propagate as artifacts. In contrast, TUNER, grounded in a theoretical result (Thrm 2), guarantees a bandlimited MLP, and serves as a soft filter, providing a representation without ringing artifacts." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 621, + 504, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 504, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 504, + 634 + ], + "type": "text", + "content": "3. Sinusoidal MLPs as Fourier series" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 641, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 553, + 689 + ], + "type": "text", + "content": "This work addresses the problem of deriving an efficient training scheme with controlled bandlimit for sinusoidal MLPs. This section presents the mathematical definitions and novel formulas for approaching this task." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Sinusoidal MLPs demonstrated high representational capacity with only a few hidden layers [19, 26, 35]. To un" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3072" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": "derstand how layer composition encodes this capacity, we propose to thoroughly investigate the structure of a 3-layer sinusoidal MLP " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "f: \\mathbb{R}^d \\to \\mathbb{R}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ". For simplicity, we assume the codomain to be " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "\\mathbb{R}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ", however the same analysis holds for dimension " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "> 1" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ". More precisely, we consider " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x}) = \\mathbf{C} \\circ \\mathbf{S} \\circ \\mathbf{D}(\\mathbf{x}) + e" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{D}(\\mathbf{x}) = \\sin (\\omega \\mathbf{x} + \\varphi)" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": " being the input layer that projects " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": " into a list of harmonics (input neurons) " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "D_{i}(\\mathbf{x}) = \\sin (\\omega_{i}\\mathbf{x} + \\varphi_{i})" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": " with frequencies " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "\\omega = (\\omega_{1},\\dots,\\omega_{m}) \\in \\mathbb{R}^{m\\times d}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": " and shifts " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "\\varphi \\in \\mathbb{R}^{m}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ". Layer " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": " is then composed with " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{S}(\\mathbf{x}) = \\sin (\\mathbf{W}\\mathbf{x} + \\mathbf{b})" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{n\\times m}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": " is the hidden matrix, and " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{b} \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": " the bias. Finally, " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{C} \\cdot \\mathbf{x} + e" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": " is an affine transformation with " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{C} \\in \\mathbb{R}^{n\\times 1}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "e \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 216, + 296, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 296, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 296, + 287 + ], + "type": "text", + "content": "We now present some properties of the sinusoidal layers, which play key roles in the representation, and give a reinterpretation of them in terms of the network parameters. First, note that the hidden neurons " + }, + { + "bbox": [ + 55, + 216, + 296, + 287 + ], + "type": "inline_equation", + "content": "\\mathbf{h}(\\mathbf{x})\\coloneqq \\mathbf{S}\\circ \\mathbf{D}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 216, + 296, + 287 + ], + "type": "text", + "content": " are defined by composing the dictionary " + }, + { + "bbox": [ + 55, + 216, + 296, + 287 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 55, + 216, + 296, + 287 + ], + "type": "text", + "content": " with the hidden sinusoidal layer which results in a list with elements" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 82, + 288, + 295, + 318 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 288, + 295, + 318 + ], + "spans": [ + { + "bbox": [ + 82, + 288, + 295, + 318 + ], + "type": "interline_equation", + "content": "h _ {i} (\\mathbf {x}) = \\sin \\left(\\sum_ {j = 1} ^ {m} W _ {i j} \\sin \\left(\\omega_ {j} \\mathbf {x} + \\varphi_ {j}\\right) + b _ {i}\\right). \\tag {1}", + "image_path": "10bc1a24bdf7eb461a3408acff5c579e4bfa7c7cf0b5728ad08d6e1a1a9ff086.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 317, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 296, + 354 + ], + "type": "text", + "content": "The following is a key result of this work, which states that we can linearize a hidden neuron (1) as a sum of sines with frequencies and amplitudes determined by " + }, + { + "bbox": [ + 55, + 317, + 296, + 354 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 317, + 296, + 354 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 317, + 296, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 55, + 317, + 296, + 354 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 359, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 359, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 359, + 296, + 384 + ], + "type": "text", + "content": "Theorem 1. Each hidden neuron " + }, + { + "bbox": [ + 55, + 359, + 296, + 384 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 55, + 359, + 296, + 384 + ], + "type": "text", + "content": " of a 3-layer sinusoidal MLP has an amplitude-phase expansion of the form" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 386, + 295, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 386, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 295, + 411 + ], + "type": "interline_equation", + "content": "h _ {i} (\\mathbf {x}) = \\sum_ {\\mathbf {k} \\in \\mathbb {Z} ^ {m}} \\alpha_ {\\mathbf {k}} \\sin \\left(\\beta_ {\\mathbf {k}} \\mathbf {x} + \\lambda_ {\\mathbf {k}}\\right), \\tag {2}", + "image_path": "2db4f06b5633cb119caf38363d9e4137383a1fd415f93f0fe1c0cd6266c7c6d7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 411, + 295, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 295, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 411, + 295, + 437 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathbf{k}} = \\langle \\mathbf{k},\\omega \\rangle" + }, + { + "bbox": [ + 55, + 411, + 295, + 437 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathbf{k}} = \\langle \\mathbf{k},\\varphi \\rangle +b_{i}" + }, + { + "bbox": [ + 55, + 411, + 295, + 437 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 55, + 411, + 295, + 437 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathbf{k}} = \\prod_{j}J_{k_j}(W_{ij})" + }, + { + "bbox": [ + 55, + 411, + 295, + 437 + ], + "type": "text", + "content": " is the product of the Bessel functions of the first kind." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 441, + 296, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 296, + 548 + ], + "type": "text", + "content": "The Bessel functions " + }, + { + "bbox": [ + 55, + 441, + 296, + 548 + ], + "type": "inline_equation", + "content": "J_{k}" + }, + { + "bbox": [ + 55, + 441, + 296, + 548 + ], + "type": "text", + "content": " appear in the Fourier series of " + }, + { + "bbox": [ + 55, + 441, + 296, + 548 + ], + "type": "inline_equation", + "content": "\\sin (a\\sin (x))" + }, + { + "bbox": [ + 55, + 441, + 296, + 548 + ], + "type": "text", + "content": " [2, Page 361] and Thrm 1 generalizes this result. See the proof and generalization in the Supp. Mat. Yüce et al. [34] present a similar expansion for neurons activated by polynomials, however, they do not derive an amplitude formula. Their expansion for sinusoidal neurons is restricted to a 3-layer MLP without bias, a critical parameter for Fourier representation. Ours incorporates bias and generalizes to deep MLPs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 549, + 296, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 296, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 296, + 670 + ], + "type": "text", + "content": "We now list some consequences of (2). First, the layer composition introduces a vast number of frequencies " + }, + { + "bbox": [ + 55, + 549, + 296, + 670 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathbf{k}}" + }, + { + "bbox": [ + 55, + 549, + 296, + 670 + ], + "type": "text", + "content": " depending only on " + }, + { + "bbox": [ + 55, + 549, + 296, + 670 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 549, + 296, + 670 + ], + "type": "text", + "content": ", and shifts given by the input shift " + }, + { + "bbox": [ + 55, + 549, + 296, + 670 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 55, + 549, + 296, + 670 + ], + "type": "text", + "content": " and the bias " + }, + { + "bbox": [ + 55, + 549, + 296, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 55, + 549, + 296, + 670 + ], + "type": "text", + "content": ". More precisely, truncating the expansion by summing over " + }, + { + "bbox": [ + 55, + 549, + 296, + 670 + ], + "type": "inline_equation", + "content": "|\\mathbf{k}|_{\\infty} \\leq B \\in \\mathbb{N}" + }, + { + "bbox": [ + 55, + 549, + 296, + 670 + ], + "type": "text", + "content": ", implies that each hidden neuron " + }, + { + "bbox": [ + 55, + 549, + 296, + 670 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 55, + 549, + 296, + 670 + ], + "type": "text", + "content": " could learn " + }, + { + "bbox": [ + 55, + 549, + 296, + 670 + ], + "type": "inline_equation", + "content": "\\frac{(2B + 1)^m - 1}{2}" + }, + { + "bbox": [ + 55, + 549, + 296, + 670 + ], + "type": "text", + "content": " non-null frequencies. This frequency generation gives a novel explanation of why composing sinusoidal layers greatly increases the network capacity. We will explore (2) to define a robust initialization for the INR's input neurons in Sec 4.1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 670, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 296, + 696 + ], + "type": "text", + "content": "Secondly, note that the weights " + }, + { + "bbox": [ + 55, + 670, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 55, + 670, + 296, + 696 + ], + "type": "text", + "content": " fully determine the amplitude " + }, + { + "bbox": [ + 55, + 670, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathbf{k}}" + }, + { + "bbox": [ + 55, + 670, + 296, + 696 + ], + "type": "text", + "content": " of each harmonic " + }, + { + "bbox": [ + 55, + 670, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\sin (\\beta_{\\mathbf{k}}\\mathbf{x} + \\lambda_{\\mathbf{k}})" + }, + { + "bbox": [ + 55, + 670, + 296, + 696 + ], + "type": "text", + "content": ". Thus, to" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "derive our bounding scheme we need to focus only on " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": ". Finally, we can derive a sine-cosine form of (2)," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 353, + 102, + 553, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 102, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 353, + 102, + 553, + 129 + ], + "type": "interline_equation", + "content": "h _ {i} (\\mathbf {x}) = \\sum_ {\\mathbf {k} \\in \\mathbb {Z} ^ {m}} A _ {\\mathbf {k}} \\sin \\left(\\beta_ {\\mathbf {k}} \\mathbf {x}\\right) + B _ {\\mathbf {k}} \\cos \\left(\\beta_ {\\mathbf {k}} \\mathbf {x}\\right), \\tag {3}", + "image_path": "ad046928f18359476019086fd2b8d2ec7b734c8d96c25140b49e0da792a2f2c8.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 135, + 555, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 135, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 135, + 555, + 243 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 313, + 135, + 555, + 243 + ], + "type": "inline_equation", + "content": "A_{\\mathbf{k}} = \\alpha_{\\mathbf{k}}\\cos (\\lambda_{\\mathbf{k}})" + }, + { + "bbox": [ + 313, + 135, + 555, + 243 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 135, + 555, + 243 + ], + "type": "inline_equation", + "content": "B_{\\mathbf{k}} = \\alpha_{\\mathbf{k}}\\sin (\\lambda_{\\mathbf{k}})" + }, + { + "bbox": [ + 313, + 135, + 555, + 243 + ], + "type": "text", + "content": ". Note that the generated frequencies are independent of the index " + }, + { + "bbox": [ + 313, + 135, + 555, + 243 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 135, + 555, + 243 + ], + "type": "text", + "content": ", thus the hidden neurons share the same harmonics: " + }, + { + "bbox": [ + 313, + 135, + 555, + 243 + ], + "type": "inline_equation", + "content": "\\sin (\\beta_{\\mathbf{k}}\\mathbf{x})" + }, + { + "bbox": [ + 313, + 135, + 555, + 243 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 135, + 555, + 243 + ], + "type": "inline_equation", + "content": "\\cos (\\beta_{\\mathbf{k}}\\mathbf{x})" + }, + { + "bbox": [ + 313, + 135, + 555, + 243 + ], + "type": "text", + "content": ". Since different combinations of the input frequencies may correspond to a single frequency, (3) isn't (yet) the Fourier transform of the network. In other words, we could have two integer vectors " + }, + { + "bbox": [ + 313, + 135, + 555, + 243 + ], + "type": "inline_equation", + "content": "\\mathbf{k},\\mathbf{l}\\in \\mathbb{Z}^{m}" + }, + { + "bbox": [ + 313, + 135, + 555, + 243 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 313, + 135, + 555, + 243 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathbf{k}} = \\beta_{\\mathbf{l}}" + }, + { + "bbox": [ + 313, + 135, + 555, + 243 + ], + "type": "text", + "content": ". In the Supp. Mat., we show how to aggregate those frequencies to obtain the final Fourier transform." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 244, + 555, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 244, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 555, + 293 + ], + "type": "text", + "content": "To control the MLP's bandlimit, we need a method to bound the amplitudes of " + }, + { + "bbox": [ + 313, + 244, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathbf{k}}" + }, + { + "bbox": [ + 313, + 244, + 555, + 293 + ], + "type": "text", + "content": ". Precisely, for a given number " + }, + { + "bbox": [ + 313, + 244, + 555, + 293 + ], + "type": "inline_equation", + "content": "B > 0" + }, + { + "bbox": [ + 313, + 244, + 555, + 293 + ], + "type": "text", + "content": " we would need " + }, + { + "bbox": [ + 313, + 244, + 555, + 293 + ], + "type": "inline_equation", + "content": "|\\alpha_{\\mathbf{k}}|" + }, + { + "bbox": [ + 313, + 244, + 555, + 293 + ], + "type": "text", + "content": " to be small for " + }, + { + "bbox": [ + 313, + 244, + 555, + 293 + ], + "type": "inline_equation", + "content": "|\\beta_{\\mathbf{k}}| \\geq B" + }, + { + "bbox": [ + 313, + 244, + 555, + 293 + ], + "type": "text", + "content": ". For this, we use the formula of " + }, + { + "bbox": [ + 313, + 244, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathbf{k}}" + }, + { + "bbox": [ + 313, + 244, + 555, + 293 + ], + "type": "text", + "content": " to determine a bounding:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 298, + 554, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 298, + 554, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 554, + 333 + ], + "type": "text", + "content": "Theorem 2. The magnitude of the amplitudes " + }, + { + "bbox": [ + 313, + 298, + 554, + 333 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathbf{k}}" + }, + { + "bbox": [ + 313, + 298, + 554, + 333 + ], + "type": "text", + "content": " in the expansion (2) is bounded by " + }, + { + "bbox": [ + 313, + 298, + 554, + 333 + ], + "type": "inline_equation", + "content": "\\prod_{j=1}^{m}\\left(\\frac{|W_{ij}|}{2}\\right)^{|k_j|}\\frac{1}{|k_j|!}" + }, + { + "bbox": [ + 313, + 298, + 554, + 333 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 336, + 554, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 554, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 554, + 361 + ], + "type": "text", + "content": "Sec 4.2 presents a bandlimit control mechanism during training based on Thrm 2." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 370, + 524, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 524, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 524, + 384 + ], + "type": "text", + "content": "4. Initialization and frequency bounding" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 390, + 555, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 555, + 535 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 555, + 535 + ], + "type": "text", + "content": "We present TUNER, an initialization and frequency control scheme for sinusoidal INRs. Our initialization considers integer input frequencies, based on a Fourier series interpretation, and uses Thrm 1 to study the spectrum of layer composition. Then, we use Thrm 2 to control the INR's bandlimit during training. Fig 2 provides an overview of TUNER. Throughout this section we present toy experiments to motivate the method, using " + }, + { + "bbox": [ + 313, + 390, + 555, + 535 + ], + "type": "inline_equation", + "content": "f: \\mathbb{R}^2 \\to \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 390, + 555, + 535 + ], + "type": "text", + "content": " in an image reconstruction setup; later, comprehensive experiments will be given in Sec 5. We report the MLP size using the parameters " + }, + { + "bbox": [ + 313, + 390, + 555, + 535 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 390, + 555, + 535 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 390, + 555, + 535 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 390, + 555, + 535 + ], + "type": "text", + "content": " as they completely determine the model. We use the Kodak dataset's images [1] and train the INR using Adam [9]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 540, + 495, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 495, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 495, + 553 + ], + "type": "text", + "content": "4.1. Initialization as spectral sampling" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": "A key challenge in initializing a sinusoidal MLP lies in defining its input frequencies " + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": ". Recall that the MLP generates frequencies as integer combinations of " + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathbf{k}} = \\sum_{i} k_{i} \\omega_{i}" + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": ", with the remaining weights representing their amplitudes " + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathbf{k}}" + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": ". This setup presents two main challenges. First, the initialization of " + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": " must enable " + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathbf{k}}" + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": " to cover the signal's spectrum; otherwise, the optimization cannot learn missing frequencies. Fig 3 (left) shows an example where using only odd frequencies leads to poor reconstruction. The second problem is the lack of flexibility for learning new frequencies. For example, randomly initializing " + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": " may not be enough since it may produce many high frequencies in order to overfit the signal, see Fig 4 (top)." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 701, + 194, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 701, + 194, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 701, + 194, + 714 + ], + "type": "inline_equation", + "content": "1|\\mathbf{k}|_{\\infty}" + }, + { + "bbox": [ + 66, + 701, + 194, + 714 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 66, + 701, + 194, + 714 + ], + "type": "inline_equation", + "content": "\\max \\{|k_1|,\\dots ,|k_m|\\}" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3073" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 541, + 236 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 541, + 236 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 541, + 236 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 541, + 236 + ], + "type": "image", + "image_path": "0a9150c266f167997b739e415fd292a20d6163031252f38a9c7519f10689866b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 237, + 555, + 326 + ], + "lines": [ + { + "bbox": [ + 54, + 237, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 54, + 237, + 555, + 326 + ], + "type": "text", + "content": "Figure 2. Overview of TUNER. To train a sinusoidal MLP (gray model, top-left), we employ two techniques derived from Thrms 1 and 2. First, we initialize the input frequencies " + }, + { + "bbox": [ + 54, + 237, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 54, + 237, + 555, + 326 + ], + "type": "text", + "content": " (green, bottom-left) with a dense distribution of low frequencies (red square) and a sparse distribution of higher frequencies (green grid). This initialization gives flexibility to learn the remaining signal frequencies which are simply integer combinations of " + }, + { + "bbox": [ + 54, + 237, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 54, + 237, + 555, + 326 + ], + "type": "text", + "content": " (the yellow nodes on the right), a consequence of the amplitude-phase expansion given by Theorem 1. Note that this initialization resembles a frequency sampling since the training generates those new frequencies around " + }, + { + "bbox": [ + 54, + 237, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 54, + 237, + 555, + 326 + ], + "type": "text", + "content": ". Second, we bound the coefficients of the hidden layer weights (blue nodes) to ensure that the MLP remains within a specified bandlimit. This approach is effective because the amplitude-phase expansion (shown on the right) of each hidden neuron (purple nodes) indicates that the amplitudes of the generated frequencies have an upper-bound depending only on the hidden weights (blue, bottom-right)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 89, + 335, + 174, + 419 + ], + "blocks": [ + { + "bbox": [ + 89, + 335, + 174, + 419 + ], + "lines": [ + { + "bbox": [ + 89, + 335, + 174, + 419 + ], + "spans": [ + { + "bbox": [ + 89, + 335, + 174, + 419 + ], + "type": "image", + "image_path": "416cd5539eac2060604bbdf4c606cd6f8c689f3555d123c83fb057826b9b26b2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 420, + 295, + 475 + ], + "lines": [ + { + "bbox": [ + 55, + 420, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 295, + 475 + ], + "type": "text", + "content": "Figure 3. Choosing " + }, + { + "bbox": [ + 55, + 420, + 295, + 475 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 420, + 295, + 475 + ], + "type": "text", + "content": " as the cartesian product of the odd frequencies without (left) or with (right) the frequencies " + }, + { + "bbox": [ + 55, + 420, + 295, + 475 + ], + "type": "inline_equation", + "content": "(1,0),(0,1)" + }, + { + "bbox": [ + 55, + 420, + 295, + 475 + ], + "type": "text", + "content": ". Note that adding them prevents sub-periods (see Supp. Mat.). We trained for 3000 epochs on a " + }, + { + "bbox": [ + 55, + 420, + 295, + 475 + ], + "type": "inline_equation", + "content": "256^{2}" + }, + { + "bbox": [ + 55, + 420, + 295, + 475 + ], + "type": "text", + "content": " image with network parameters " + }, + { + "bbox": [ + 55, + 420, + 295, + 475 + ], + "type": "inline_equation", + "content": "m = 72" + }, + { + "bbox": [ + 55, + 420, + 295, + 475 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 420, + 295, + 475 + ], + "type": "inline_equation", + "content": "n = 512" + }, + { + "bbox": [ + 55, + 420, + 295, + 475 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 420, + 295, + 475 + ], + "type": "inline_equation", + "content": "\\theta = 30" + }, + { + "bbox": [ + 55, + 420, + 295, + 475 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 175, + 335, + 260, + 419 + ], + "blocks": [ + { + "bbox": [ + 175, + 335, + 260, + 419 + ], + "lines": [ + { + "bbox": [ + 175, + 335, + 260, + 419 + ], + "spans": [ + { + "bbox": [ + 175, + 335, + 260, + 419 + ], + "type": "image", + "image_path": "58d5f9b65d759c8a987ff00280879f99b71ee245652153557b706075598f8407.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 62, + 479, + 287, + 624 + ], + "blocks": [ + { + "bbox": [ + 62, + 479, + 287, + 624 + ], + "lines": [ + { + "bbox": [ + 62, + 479, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 479, + 287, + 624 + ], + "type": "image", + "image_path": "6884eee2b87978eb3e8c96218e154fd69d8dd5903f659e34f6a973962c374ee1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 625, + 295, + 670 + ], + "lines": [ + { + "bbox": [ + 55, + 625, + 295, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 625, + 295, + 670 + ], + "type": "text", + "content": "Figure 4. Uniform init. of " + }, + { + "bbox": [ + 55, + 625, + 295, + 670 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 625, + 295, + 670 + ], + "type": "text", + "content": " (top) and ours (bottom). Grayscale bands show INR's gradient. Ours offers better signal/gradient reconstruction w/o gradient supervision. The MLPs (" + }, + { + "bbox": [ + 55, + 625, + 295, + 670 + ], + "type": "inline_equation", + "content": "m = 128" + }, + { + "bbox": [ + 55, + 625, + 295, + 670 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 625, + 295, + 670 + ], + "type": "inline_equation", + "content": "n = 256" + }, + { + "bbox": [ + 55, + 625, + 295, + 670 + ], + "type": "text", + "content": ") with bandlimit " + }, + { + "bbox": [ + 55, + 625, + 295, + 670 + ], + "type": "inline_equation", + "content": "\\theta = 87" + }, + { + "bbox": [ + 55, + 625, + 295, + 670 + ], + "type": "text", + "content": " were trained for 3000 epochs." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "To address these challenges, we introduce a novel initialization. First, we restrict our MLP to periodic functions, ensuring the training domain is in a full period. Such" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 346, + 555, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 406 + ], + "type": "text", + "content": "functions can be represented by sums of sines with integer frequencies, defining an orthogonal basis (Fourier series); This aligns well with our analysis (Thrm 1). The main challenge is initializing " + }, + { + "bbox": [ + 313, + 346, + 555, + 406 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 346, + 555, + 406 + ], + "type": "text", + "content": " such that the generated frequencies " + }, + { + "bbox": [ + 313, + 346, + 555, + 406 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathbf{k}}" + }, + { + "bbox": [ + 313, + 346, + 555, + 406 + ], + "type": "text", + "content": " cover the full spectrum within a bandlimit " + }, + { + "bbox": [ + 313, + 346, + 555, + 406 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 346, + 555, + 406 + ], + "type": "text", + "content": " (Fourier basis)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 406, + 554, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 554, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 554, + 477 + ], + "type": "text", + "content": "More precisely, we initialize the weights " + }, + { + "bbox": [ + 313, + 406, + 554, + 477 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 406, + 554, + 477 + ], + "type": "text", + "content": " with integer frequencies, that is " + }, + { + "bbox": [ + 313, + 406, + 554, + 477 + ], + "type": "inline_equation", + "content": "\\omega_{j}\\in \\frac{2\\pi}{p}\\mathbb{Z}^{d}" + }, + { + "bbox": [ + 313, + 406, + 554, + 477 + ], + "type": "text", + "content": ", and freeze them during training, with " + }, + { + "bbox": [ + 313, + 406, + 554, + 477 + ], + "type": "inline_equation", + "content": "p > 0" + }, + { + "bbox": [ + 313, + 406, + 554, + 477 + ], + "type": "text", + "content": " being the period of the INR. This guarantees that the input neurons are " + }, + { + "bbox": [ + 313, + 406, + 554, + 477 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 406, + 554, + 477 + ], + "type": "text", + "content": "-periodic and Thrm 1 ensures that such periodicity is preserved over layer composition. Additionally, (3) implies that we can rewrite the INR as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 319, + 481, + 555, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 481, + 555, + 507 + ], + "spans": [ + { + "bbox": [ + 319, + 481, + 555, + 507 + ], + "type": "interline_equation", + "content": "f (\\mathbf {x}) = \\sum_ {\\mathbf {k} \\in \\mathbb {Z} ^ {m}} \\left\\langle \\mathbf {C}, A _ {\\mathbf {k}} \\right\\rangle \\sin \\left(\\beta_ {\\mathbf {k}} \\mathbf {x}\\right) + \\left\\langle \\mathbf {C}, B _ {\\mathbf {k}} \\right\\rangle \\cos \\left(\\beta_ {\\mathbf {k}} \\mathbf {x}\\right) + e. \\tag {4}", + "image_path": "3ea0c8045862fc6e12fd14755ce3eb34a17b7308ed64618dbba74becf5d860f5.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 510, + 554, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 547 + ], + "type": "text", + "content": "Since the generated frequencies " + }, + { + "bbox": [ + 313, + 510, + 554, + 547 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathbf{k}}" + }, + { + "bbox": [ + 313, + 510, + 554, + 547 + ], + "type": "text", + "content": " only depend on the frozen parameters " + }, + { + "bbox": [ + 313, + 510, + 554, + 547 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 510, + 554, + 547 + ], + "type": "text", + "content": ", the training is responsible for learning the amplitudes of the sine-cosine series in (4)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "type": "text", + "content": " be the signal's bandlimit. To ensure that the generated frequencies do not bypass " + }, + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "type": "text", + "content": ", we cannot sample " + }, + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "type": "text", + "content": " directly on " + }, + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "type": "inline_equation", + "content": "[-B, B]^d" + }, + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "type": "text", + "content": " since the MLP generates multiples of these frequencies. Therefore, we sample " + }, + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\omega_j" + }, + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\frac{2\\pi}{p} [-\\theta, \\theta]^d" + }, + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\theta \\in (0, B)" + }, + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "type": "text", + "content": " being a threshold for the input frequencies. Indeed, as the coefficients of hidden matrix " + }, + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "type": "text", + "content": " are limited by 2 (see Sec 4.2), the magnitudes of " + }, + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "type": "inline_equation", + "content": "A_{\\mathbf{k}}" + }, + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "type": "inline_equation", + "content": "B_{\\mathbf{k}}" + }, + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "type": "text", + "content": " decrease as " + }, + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "type": "inline_equation", + "content": "|\\mathbf{k}|_{\\infty}" + }, + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "type": "text", + "content": " increases. This is a consequence of Thrm 2:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 645, + 553, + 686 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 645, + 553, + 686 + ], + "spans": [ + { + "bbox": [ + 320, + 645, + 553, + 686 + ], + "type": "interline_equation", + "content": "\\left| \\alpha_ {\\mathbf {k}} \\right| \\leq \\prod_ {j = 1} ^ {m} \\frac {\\left(\\frac {\\left| W _ {i j} \\right|}{2}\\right) ^ {\\left| k _ {j} \\right|}}{\\left| k _ {j} \\right| !} \\leq \\frac {1}{\\left| k _ {1} \\right| ! \\dots \\left| k _ {m} \\right| !}; i = 1, \\dots , n. \\tag {5}", + "image_path": "fce3bcd1da52bb2a7d9105e4f25cd63f4e1532f978a74aac598943dacbb6da68.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Thus, the generated frequencies " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathbf{k}}" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " with small " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{k}" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " have more influence over the expansion (4), mimicking Fourier series." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3074" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 72, + 136, + 148 + ], + "blocks": [ + { + "bbox": [ + 60, + 72, + 136, + 148 + ], + "lines": [ + { + "bbox": [ + 60, + 72, + 136, + 148 + ], + "spans": [ + { + "bbox": [ + 60, + 72, + 136, + 148 + ], + "type": "image", + "image_path": "a2b79bedba58f4a9879c8b3418ad7bce8da827d77250787bfe1220414a2d8b03.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 154, + 295, + 210 + ], + "lines": [ + { + "bbox": [ + 55, + 154, + 295, + 210 + ], + "spans": [ + { + "bbox": [ + 55, + 154, + 295, + 210 + ], + "type": "text", + "content": "Figure 5. Image reconstructions with bandlimit " + }, + { + "bbox": [ + 55, + 154, + 295, + 210 + ], + "type": "inline_equation", + "content": "B = 128" + }, + { + "bbox": [ + 55, + 154, + 295, + 210 + ], + "type": "text", + "content": ", varying " + }, + { + "bbox": [ + 55, + 154, + 295, + 210 + ], + "type": "inline_equation", + "content": "\\theta = 4, 43, 128" + }, + { + "bbox": [ + 55, + 154, + 295, + 210 + ], + "type": "text", + "content": " with a network (" + }, + { + "bbox": [ + 55, + 154, + 295, + 210 + ], + "type": "inline_equation", + "content": "m = 80, n = 1000" + }, + { + "bbox": [ + 55, + 154, + 295, + 210 + ], + "type": "text", + "content": ") trained over 3000 epochs. The gradient magnitude is shown on the left of each image. Note that smaller " + }, + { + "bbox": [ + 55, + 154, + 295, + 210 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 154, + 295, + 210 + ], + "type": "text", + "content": " (middle-left) yield better reconstructions, while higher values of " + }, + { + "bbox": [ + 55, + 154, + 295, + 210 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 154, + 295, + 210 + ], + "type": "text", + "content": " introduce noisy gradients (right)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 138, + 72, + 213, + 149 + ], + "blocks": [ + { + "bbox": [ + 138, + 72, + 213, + 149 + ], + "lines": [ + { + "bbox": [ + 138, + 72, + 213, + 149 + ], + "spans": [ + { + "bbox": [ + 138, + 72, + 213, + 149 + ], + "type": "image", + "image_path": "09c628756c6397fe6edc0030b950ab95f481b73fa5e61500eda2b50ff0e17673.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 215, + 72, + 290, + 148 + ], + "blocks": [ + { + "bbox": [ + 215, + 72, + 290, + 148 + ], + "lines": [ + { + "bbox": [ + 215, + 72, + 290, + 148 + ], + "spans": [ + { + "bbox": [ + 215, + 72, + 290, + 148 + ], + "type": "image", + "image_path": "5dc6289d670196bc6830ffd0a171a48069fec8b6f99f4551f29845591a7f068e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 216, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 295, + 312 + ], + "type": "text", + "content": "For example, if input frequencies include 1 and 20, frequencies like " + }, + { + "bbox": [ + 55, + 216, + 295, + 312 + ], + "type": "inline_equation", + "content": "19 = 20 - 1" + }, + { + "bbox": [ + 55, + 216, + 295, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 216, + 295, + 312 + ], + "type": "inline_equation", + "content": "21 = 20 + 1" + }, + { + "bbox": [ + 55, + 216, + 295, + 312 + ], + "type": "text", + "content": " can be represented through their combinations. Thus, low frequencies give flexibility in expanding around the input ones which motivates initializing more input frequencies near the origin. We found " + }, + { + "bbox": [ + 55, + 216, + 295, + 312 + ], + "type": "inline_equation", + "content": "\\theta = \\frac{B}{3}" + }, + { + "bbox": [ + 55, + 216, + 295, + 312 + ], + "type": "text", + "content": " to work well experimentally. Fig 5 shows that it is necessary, as higher " + }, + { + "bbox": [ + 55, + 216, + 295, + 312 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 216, + 295, + 312 + ], + "type": "text", + "content": " makes it harder to learn the small frequencies, introducing noise." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "text", + "content": "To initialize " + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "inline_equation", + "content": "\\omega \\in \\mathbb{R}^{m\\times d}" + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "text", + "content": ", we need to choose " + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "text", + "content": " frequencies in " + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "inline_equation", + "content": "[- \\theta, \\theta]^d" + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "text", + "content": ". Taking into account the importance of the lower frequencies discussed above, we split it into regions " + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{L} = [-\\ell, \\ell]^d" + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "text", + "content": " of low and " + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{H} = [-\\theta, \\theta]^d \\setminus \\mathbf{L}" + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "text", + "content": " of high frequencies, " + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "inline_equation", + "content": "\\ell \\in (0, \\theta)" + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "text", + "content": ". As " + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathbf{k}}" + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "text", + "content": " only has a significant amplitude when " + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "inline_equation", + "content": "|\\mathbf{k}|_{\\infty}" + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "text", + "content": " is small (5), initializing with only low input frequencies may not allow the generation of high frequencies. Thus, we initialize some frequencies in " + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "text", + "content": " to cover the spectrum, and the rest in " + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "text", + "content": " to cover a neighborhood of the input frequencies. Specifically, we select " + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "inline_equation", + "content": "m_l < m" + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "text", + "content": " frequencies uniformly from " + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "text", + "content": ". Then, we choose " + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "inline_equation", + "content": "m - m_l" + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "text", + "content": " frequencies from " + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "text", + "content": ", spread in a grid-like fashion to cover " + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "inline_equation", + "content": "[- \\theta, \\theta]^d" + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "text", + "content": ". In our experiments we considered " + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "inline_equation", + "content": "m_l = 0.7m" + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "text", + "content": ". Finally, since " + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{D}(\\mathbf{x}) = \\sin (\\omega \\mathbf{x} + \\varphi)" + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "text", + "content": ", we initialize " + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "text", + "content": " uniformly in " + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "inline_equation", + "content": "[- \\frac{\\pi}{2}, \\frac{\\pi}{2}]" + }, + { + "bbox": [ + 55, + 312, + 295, + 480 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 481, + 295, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 295, + 553 + ], + "type": "text", + "content": "When " + }, + { + "bbox": [ + 55, + 481, + 295, + 553 + ], + "type": "inline_equation", + "content": "\\omega_{j}\\in \\mathbf{L}" + }, + { + "bbox": [ + 55, + 481, + 295, + 553 + ], + "type": "text", + "content": ", small variations in " + }, + { + "bbox": [ + 55, + 481, + 295, + 553 + ], + "type": "inline_equation", + "content": "k_{j}" + }, + { + "bbox": [ + 55, + 481, + 295, + 553 + ], + "type": "text", + "content": " results in a frequency near " + }, + { + "bbox": [ + 55, + 481, + 295, + 553 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathbf{k}}" + }, + { + "bbox": [ + 55, + 481, + 295, + 553 + ], + "type": "text", + "content": ". Indeed, Fig 6 shows how low frequencies (green square) introduce frequencies around the input frequencies (Fig 6a). Conversely, if " + }, + { + "bbox": [ + 55, + 481, + 295, + 553 + ], + "type": "inline_equation", + "content": "\\omega_{j}\\in \\mathbf{H}" + }, + { + "bbox": [ + 55, + 481, + 295, + 553 + ], + "type": "text", + "content": " then " + }, + { + "bbox": [ + 55, + 481, + 295, + 553 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathbf{k} + \\mathbf{e}_j}" + }, + { + "bbox": [ + 55, + 481, + 295, + 553 + ], + "type": "text", + "content": " is far away from " + }, + { + "bbox": [ + 55, + 481, + 295, + 553 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathbf{k}}" + }, + { + "bbox": [ + 55, + 481, + 295, + 553 + ], + "type": "text", + "content": "; we note this in Fig 6b-d, where " + }, + { + "bbox": [ + 55, + 481, + 295, + 553 + ], + "type": "inline_equation", + "content": "\\omega_{j}" + }, + { + "bbox": [ + 55, + 481, + 295, + 553 + ], + "type": "text", + "content": " is the red point and the arrows show the generated frequencies " + }, + { + "bbox": [ + 55, + 481, + 295, + 553 + ], + "type": "inline_equation", + "content": "k_{j}\\omega_{j}" + }, + { + "bbox": [ + 55, + 481, + 295, + 553 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 57, + 557, + 115, + 616 + ], + "blocks": [ + { + "bbox": [ + 57, + 557, + 115, + 616 + ], + "lines": [ + { + "bbox": [ + 57, + 557, + 115, + 616 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 115, + 616 + ], + "type": "image", + "image_path": "e0aef8b4f6a97780f8dc3c30b24b6d57cbfa742af0ec8dae5628a3155f1ac2ad.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 624, + 295, + 668 + ], + "lines": [ + { + "bbox": [ + 55, + 624, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 295, + 668 + ], + "type": "text", + "content": "Figure 6. INR spectrum during training. The input frequencies " + }, + { + "bbox": [ + 55, + 624, + 295, + 668 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 624, + 295, + 668 + ], + "type": "text", + "content": " in green. During training new frequencies appear in a neighborhood of " + }, + { + "bbox": [ + 55, + 624, + 295, + 668 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 624, + 295, + 668 + ], + "type": "text", + "content": ". The red dot indicates a high input frequency " + }, + { + "bbox": [ + 55, + 624, + 295, + 668 + ], + "type": "inline_equation", + "content": "\\omega_{j}" + }, + { + "bbox": [ + 55, + 624, + 295, + 668 + ], + "type": "text", + "content": " and the red arrows show the generated frequencies " + }, + { + "bbox": [ + 55, + 624, + 295, + 668 + ], + "type": "inline_equation", + "content": "k_{j}\\omega_{j}, k_{j} = 1,2,3" + }, + { + "bbox": [ + 55, + 624, + 295, + 668 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 116, + 557, + 175, + 616 + ], + "blocks": [ + { + "bbox": [ + 116, + 557, + 175, + 616 + ], + "lines": [ + { + "bbox": [ + 116, + 557, + 175, + 616 + ], + "spans": [ + { + "bbox": [ + 116, + 557, + 175, + 616 + ], + "type": "image", + "image_path": "826da1394588ff8d16dfea45c4bf8e4dbcc333346b54852888c3f07be7bcb506.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 175, + 557, + 235, + 616 + ], + "blocks": [ + { + "bbox": [ + 175, + 557, + 235, + 616 + ], + "lines": [ + { + "bbox": [ + 175, + 557, + 235, + 616 + ], + "spans": [ + { + "bbox": [ + 175, + 557, + 235, + 616 + ], + "type": "image", + "image_path": "17ff66eee1e89182eb16257218cdac11d36593df270ead08c3e3151964c7c5b7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 235, + 557, + 294, + 616 + ], + "blocks": [ + { + "bbox": [ + 235, + 557, + 294, + 616 + ], + "lines": [ + { + "bbox": [ + 235, + 557, + 294, + 616 + ], + "spans": [ + { + "bbox": [ + 235, + 557, + 294, + 616 + ], + "type": "image", + "image_path": "6b6f30254f97045b646ad55426325e5014c40711f7d7844096e6d7de7a4060d7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 566, + 555, + 599 + ], + "lines": [ + { + "bbox": [ + 313, + 566, + 555, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 555, + 599 + ], + "type": "text", + "content": "Figure 7. Training a network " + }, + { + "bbox": [ + 313, + 566, + 555, + 599 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 566, + 555, + 599 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 566, + 555, + 599 + ], + "type": "inline_equation", + "content": "m = n = 256" + }, + { + "bbox": [ + 313, + 566, + 555, + 599 + ], + "type": "text", + "content": " for 6000 epochs with " + }, + { + "bbox": [ + 313, + 566, + 555, + 599 + ], + "type": "inline_equation", + "content": "\\theta = 30" + }, + { + "bbox": [ + 313, + 566, + 555, + 599 + ], + "type": "text", + "content": " and bounds " + }, + { + "bbox": [ + 313, + 566, + 555, + 599 + ], + "type": "inline_equation", + "content": "c = 0.1, 0.2, 0.5" + }, + { + "bbox": [ + 313, + 566, + 555, + 599 + ], + "type": "text", + "content": ". The bound over the hidden matrix " + }, + { + "bbox": [ + 313, + 566, + 555, + 599 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 313, + 566, + 555, + 599 + ], + "type": "text", + "content": " restricts the appearance of high frequencies." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Finally, when sampling frequencies in " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ", we include " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "(1,0)" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "(0,1)" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " to guarantee that any frequency could be generated and prevents the appearance of sub-periods as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "shown at the beginning of this section in Fig 3 (left) (see also Supp. Mat for more details). Fig 3 (right) shows how this problem is solved using our initialization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 108, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 553, + 168 + ], + "type": "text", + "content": "The proposed initialization overcomes the noisy reconstruction issues that arise with naive uniform sampling in " + }, + { + "bbox": [ + 313, + 108, + 553, + 168 + ], + "type": "inline_equation", + "content": "[- \\theta, \\theta]^d" + }, + { + "bbox": [ + 313, + 108, + 553, + 168 + ], + "type": "text", + "content": ", as used in [26] (Fig 4 (top)). Fig 4 (bottom) demonstrates that our scheme significantly improves both signal and gradient reconstruction quality." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 175, + 526, + 187 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 526, + 187 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 526, + 187 + ], + "type": "text", + "content": "4.2. Bounding and hidden layer initialization" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 192, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 555, + 300 + ], + "type": "text", + "content": "As shown in previous sections, sinusoidal INRs can represent detailed signals due to the wide number of frequencies generated by layer composition. However, they could manifest as noise in the reconstruction, see Fig 5 (right). Specifically, high input frequencies may generate even higher ones, bypassing the bandlimit " + }, + { + "bbox": [ + 313, + 192, + 555, + 300 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 192, + 555, + 300 + ], + "type": "text", + "content": ". To avoid this, we use Thrm 2 to introduce a bounding scheme to limit the amplitudes. Precisely, Thrm 2 implies that in each hidden neuron " + }, + { + "bbox": [ + 313, + 192, + 555, + 300 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 313, + 192, + 555, + 300 + ], + "type": "text", + "content": ", the amplitude of a generated frequency " + }, + { + "bbox": [ + 313, + 192, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathbf{k}}" + }, + { + "bbox": [ + 313, + 192, + 555, + 300 + ], + "type": "text", + "content": " is limited by:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 364, + 308, + 553, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 308, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 364, + 308, + 553, + 338 + ], + "type": "interline_equation", + "content": "\\left| \\alpha_ {\\mathbf {k}} \\right| \\leq \\left(\\frac {\\left| \\mathbf {W} _ {i} \\right| _ {\\infty}}{2}\\right) ^ {\\left| \\mathbf {k} \\right| _ {1}} \\frac {1}{\\prod_ {j = 1} ^ {m} \\left| k _ {j} \\right| !}. \\tag {6}", + "image_path": "0f61867452f982ab5e10b494098af18ee78c2cec3295230303f4c05f3793e069.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 346, + 555, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 454 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 346, + 555, + 454 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_i" + }, + { + "bbox": [ + 313, + 346, + 555, + 454 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 346, + 555, + 454 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 346, + 555, + 454 + ], + "type": "text", + "content": "-row of " + }, + { + "bbox": [ + 313, + 346, + 555, + 454 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 313, + 346, + 555, + 454 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 313, + 346, + 555, + 454 + ], + "type": "inline_equation", + "content": "|\\mathbf{W}|_{\\infty} \\leq 2" + }, + { + "bbox": [ + 313, + 346, + 555, + 454 + ], + "type": "text", + "content": " implies that the amplitude of any generated frequency " + }, + { + "bbox": [ + 313, + 346, + 555, + 454 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathbf{k}}" + }, + { + "bbox": [ + 313, + 346, + 555, + 454 + ], + "type": "text", + "content": " rapidly decreases when the coordinates of " + }, + { + "bbox": [ + 313, + 346, + 555, + 454 + ], + "type": "inline_equation", + "content": "\\mathbf{k}" + }, + { + "bbox": [ + 313, + 346, + 555, + 454 + ], + "type": "text", + "content": " grow. Therefore, we use (6) to define a bounding scheme that only allows the appearance of generated frequencies with small multiples of " + }, + { + "bbox": [ + 313, + 346, + 555, + 454 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 346, + 555, + 454 + ], + "type": "text", + "content": " which is beneficial as shown in previous sections. For this, we simply clamp the coefficients " + }, + { + "bbox": [ + 313, + 346, + 555, + 454 + ], + "type": "inline_equation", + "content": "W_{ij}" + }, + { + "bbox": [ + 313, + 346, + 555, + 454 + ], + "type": "text", + "content": " by a bound parameter " + }, + { + "bbox": [ + 313, + 346, + 555, + 454 + ], + "type": "inline_equation", + "content": "c \\in (0,2]" + }, + { + "bbox": [ + 313, + 346, + 555, + 454 + ], + "type": "text", + "content": " at each optimization step, enforcing " + }, + { + "bbox": [ + 313, + 346, + 555, + 454 + ], + "type": "inline_equation", + "content": "|\\mathbf{W}|_{\\infty} \\leq c" + }, + { + "bbox": [ + 313, + 346, + 555, + 454 + ], + "type": "text", + "content": ". Fig 7 shows results of bounding " + }, + { + "bbox": [ + 313, + 346, + 555, + 454 + ], + "type": "inline_equation", + "content": "|\\mathbf{W}|" + }, + { + "bbox": [ + 313, + 346, + 555, + 454 + ], + "type": "text", + "content": " with different " + }, + { + "bbox": [ + 313, + 346, + 555, + 454 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 346, + 555, + 454 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 335, + 468, + 399, + 559 + ], + "blocks": [ + { + "bbox": [ + 335, + 468, + 399, + 559 + ], + "lines": [ + { + "bbox": [ + 335, + 468, + 399, + 559 + ], + "spans": [ + { + "bbox": [ + 335, + 468, + 399, + 559 + ], + "type": "image", + "image_path": "3f38fc49434962dd9928526679512b7558ee3f4fba8ed076b1b9237eae2c7fa3.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 401, + 467, + 464, + 559 + ], + "blocks": [ + { + "bbox": [ + 401, + 467, + 464, + 559 + ], + "lines": [ + { + "bbox": [ + 401, + 467, + 464, + 559 + ], + "spans": [ + { + "bbox": [ + 401, + 467, + 464, + 559 + ], + "type": "image", + "image_path": "2507cf1002ca90190c6fdd41c8c50c2a8ddafe122d333c044dfb5305122ef734.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 466, + 467, + 531, + 559 + ], + "blocks": [ + { + "bbox": [ + 466, + 467, + 531, + 559 + ], + "lines": [ + { + "bbox": [ + 466, + 467, + 531, + 559 + ], + "spans": [ + { + "bbox": [ + 466, + 467, + 531, + 559 + ], + "type": "image", + "image_path": "8b45a8e6e57a66380bef200f415b384d883ed9550145f088a2a4abe47ebe4e68.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "In Sec 4.1 we split the input frequencies " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " in two sets: the low " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "(|\\omega_{j}|_{\\infty}\\leq \\ell)" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " and high " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "(|\\omega_{j}|_{\\infty} > \\ell)" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " frequencies. Additionally, we saw that the lower ones have a main role in the frequency generation, thus, we consider a higher threshold for them. For this, observe that each input frequency " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\omega_{j}" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " is associated with the " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "-column of " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": ". Then, bounding the columns of " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " related to high frequencies, we restrict the amplitudes of " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\beta_{\\mathbf{k}}" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " that could exceed the Nyquist limit " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": ". Conversely, since low frequencies are used to span" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3075" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": "around " + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": ", we consider a higher bound for their columns. Precisely, we define two parameters " + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "inline_equation", + "content": "e_{\\mathbf{L}}" + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "inline_equation", + "content": "e_{\\mathbf{H}}" + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": " to bound the weights " + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "inline_equation", + "content": "W_{ij}" + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": " corresponding to " + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "inline_equation", + "content": "\\omega_{j}" + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": ", respectively," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 80, + 112, + 284, + 143 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 112, + 284, + 143 + ], + "spans": [ + { + "bbox": [ + 80, + 112, + 284, + 143 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left| \\omega_ {j} \\right| _ {\\infty} \\leq \\ell \\longrightarrow W _ {i j} = \\operatorname {c l a m p} \\left(W _ {i j}, \\left[ - e _ {\\mathbf {L}}, e _ {\\mathbf {L}} \\right]\\right), \\\\ \\left| \\omega_ {j} \\right| _ {\\infty} > \\theta \\longrightarrow W _ {i j} = \\operatorname {c l a m p} \\left(W _ {i j}, \\left[ - \\boldsymbol {c} _ {\\mathbf {H}}, \\boldsymbol {c} _ {\\mathbf {H}} \\right]\\right). \\\\ \\end{array}", + "image_path": "25961d4b6e4cff83b95da65f3c34d02fc861c75ec3b63c61ad5b1a27b4296043.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 145, + 296, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 145, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 145, + 296, + 253 + ], + "type": "text", + "content": "We use these bounds to initialize " + }, + { + "bbox": [ + 55, + 145, + 296, + 253 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 55, + 145, + 296, + 253 + ], + "type": "text", + "content": " such that each column is normal distributed with mean 0 and standard deviation " + }, + { + "bbox": [ + 55, + 145, + 296, + 253 + ], + "type": "inline_equation", + "content": "\\frac{c_{\\mathrm{L}}}{3}" + }, + { + "bbox": [ + 55, + 145, + 296, + 253 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 145, + 296, + 253 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{L}}" + }, + { + "bbox": [ + 55, + 145, + 296, + 253 + ], + "type": "text", + "content": " being the bound corresponding to that column. Note that SIREN [26] initializes " + }, + { + "bbox": [ + 55, + 145, + 296, + 253 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{n \\times m}" + }, + { + "bbox": [ + 55, + 145, + 296, + 253 + ], + "type": "text", + "content": " uniformly in " + }, + { + "bbox": [ + 55, + 145, + 296, + 253 + ], + "type": "inline_equation", + "content": "(- \\sqrt{6/m}, \\sqrt{6/m})" + }, + { + "bbox": [ + 55, + 145, + 296, + 253 + ], + "type": "text", + "content": ", and as " + }, + { + "bbox": [ + 55, + 145, + 296, + 253 + ], + "type": "inline_equation", + "content": "|\\mathbf{W}|_{\\infty} \\leq \\sqrt{6/m} < 2" + }, + { + "bbox": [ + 55, + 145, + 296, + 253 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 145, + 296, + 253 + ], + "type": "inline_equation", + "content": "m > 1" + }, + { + "bbox": [ + 55, + 145, + 296, + 253 + ], + "type": "text", + "content": ", it is bandlimited only at the beginning of training, since " + }, + { + "bbox": [ + 55, + 145, + 296, + 253 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 55, + 145, + 296, + 253 + ], + "type": "text", + "content": " is not controlled in any way during the optimization. Conversely, our initialization controls the bandlimit during all training and has faster convergence." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 253, + 296, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 296, + 326 + ], + "type": "text", + "content": "Additionally, we present a scheme to learn the bounding parameters during training. For this, we bound each " + }, + { + "bbox": [ + 55, + 253, + 296, + 326 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 253, + 296, + 326 + ], + "type": "text", + "content": "-column of the hidden matrix " + }, + { + "bbox": [ + 55, + 253, + 296, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 55, + 253, + 296, + 326 + ], + "type": "text", + "content": " using a tanh activation followed by multiplication with a trainable bounding parameter " + }, + { + "bbox": [ + 55, + 253, + 296, + 326 + ], + "type": "inline_equation", + "content": "e_j" + }, + { + "bbox": [ + 55, + 253, + 296, + 326 + ], + "type": "text", + "content": ". Clearly, the resulting column is bounded by " + }, + { + "bbox": [ + 55, + 253, + 296, + 326 + ], + "type": "inline_equation", + "content": "e_j" + }, + { + "bbox": [ + 55, + 253, + 296, + 326 + ], + "type": "text", + "content": ". In other words, we replace the hidden layer " + }, + { + "bbox": [ + 55, + 253, + 296, + 326 + ], + "type": "inline_equation", + "content": "\\sin (\\mathbf{W}\\mathbf{x} + \\mathbf{b})" + }, + { + "bbox": [ + 55, + 253, + 296, + 326 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 329, + 294, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 329, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 126, + 329, + 294, + 342 + ], + "type": "interline_equation", + "content": "\\sin (\\tanh (\\mathbf {W}) \\mathcal {C} \\mathbf {x} + \\mathbf {b}), \\tag {7}", + "image_path": "c91628a42de129bebea3e49b729e0b1159c7f782c719fab6470412ed244ea926.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "text", + "content": " is a diagonal matrix with each " + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "inline_equation", + "content": "jj" + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "text", + "content": "-entry being " + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "inline_equation", + "content": "c_{j}" + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "text", + "content": ". To prevent the bounds from growing too large, we use Thrm 2 to define the regularization term as " + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}} = \\sum |c_j|" + }, + { + "bbox": [ + 55, + 346, + 296, + 407 + ], + "type": "text", + "content": ". Otherwise, the hidden weights may grow unbounded, leading to the INR's overfit as in the usual training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 417, + 137, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 417, + 137, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 137, + 430 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 437, + 296, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 593 + ], + "type": "text", + "content": "This section presents experiments regarding the initialization and bounding of frequencies of an INR " + }, + { + "bbox": [ + 55, + 437, + 296, + 593 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 437, + 296, + 593 + ], + "type": "text", + "content": " considering the images from Kodak dataset [1] resized to " + }, + { + "bbox": [ + 55, + 437, + 296, + 593 + ], + "type": "inline_equation", + "content": "[-1, 1]^2" + }, + { + "bbox": [ + 55, + 437, + 296, + 593 + ], + "type": "text", + "content": ". We describe the architecture of " + }, + { + "bbox": [ + 55, + 437, + 296, + 593 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 437, + 296, + 593 + ], + "type": "text", + "content": " with hidden matrix " + }, + { + "bbox": [ + 55, + 437, + 296, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{n \\times m}" + }, + { + "bbox": [ + 55, + 437, + 296, + 593 + ], + "type": "text", + "content": " with the parameters " + }, + { + "bbox": [ + 55, + 437, + 296, + 593 + ], + "type": "inline_equation", + "content": "m, n" + }, + { + "bbox": [ + 55, + 437, + 296, + 593 + ], + "type": "text", + "content": ". We initialize " + }, + { + "bbox": [ + 55, + 437, + 296, + 593 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 437, + 296, + 593 + ], + "type": "text", + "content": " following the scheme in Sec 4.1 and optimize it using Adam [9] with learning rate " + }, + { + "bbox": [ + 55, + 437, + 296, + 593 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 55, + 437, + 296, + 593 + ], + "type": "text", + "content": ". For the quantitative evaluations, we report the mean and standard deviation of the Peak Signal-to-Noise Ratio (PSNR) evaluated on the test set (10% random pixels). To show that TUNER adds high order regularity to the reconstruction, we compare the analytical gradient " + }, + { + "bbox": [ + 55, + 437, + 296, + 593 + ], + "type": "inline_equation", + "content": "\\nabla f" + }, + { + "bbox": [ + 55, + 437, + 296, + 593 + ], + "type": "text", + "content": " to the Sobel filter of the ground truth. All experiments were run in a 12 GB NVIDIA GPU (TITAN X Pascal)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 600, + 138, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 138, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 138, + 611 + ], + "type": "text", + "content": "5.1. Initialization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "First, we compare our initialization of " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " (36.2dB) against the uniform distribution (32.2dB). We obtain a 4dB improvement of PNSR on average (with a std of 1.6dB). We also present an ablation over the parameter " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " which splits the square " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "[- \\ell, \\ell]^2" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " in regions of low (L) and high (H) frequencies. Here we use " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\ell = 85" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\ell = 21" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": ", 42, 64, and sample " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " of coordinates in H. Table 1 shows the performance of the INR." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 316, + 70, + 553, + 124 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 553, + 124 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 553, + 124 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 553, + 124 + ], + "type": "table", + "html": "
% high freqs30%50%70%
\\( \\ell \\)SignalGradSignalGradSignalGrad
2135.226.935.226.635.026.0
4235.126.334.824.633.822.2
6434.225.233.321.630.815.0
", + "image_path": "3843b29c5e4e05e7640b64587ae02fc9cb5ce1689928ff33090c35407c594bd4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 126, + 555, + 159 + ], + "lines": [ + { + "bbox": [ + 313, + 126, + 555, + 159 + ], + "spans": [ + { + "bbox": [ + 313, + 126, + 555, + 159 + ], + "type": "text", + "content": "Table 1. Training an INR with " + }, + { + "bbox": [ + 313, + 126, + 555, + 159 + ], + "type": "inline_equation", + "content": "m = n = 416" + }, + { + "bbox": [ + 313, + 126, + 555, + 159 + ], + "type": "text", + "content": " parameters, for 400 epochs. Results indicate that it is better to sample more " + }, + { + "bbox": [ + 313, + 126, + 555, + 159 + ], + "type": "inline_equation", + "content": "\\omega_{j} \\in \\mathbf{L}" + }, + { + "bbox": [ + 313, + 126, + 555, + 159 + ], + "type": "text", + "content": " and choose a smaller " + }, + { + "bbox": [ + 313, + 126, + 555, + 159 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 313, + 126, + 555, + 159 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 168, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 239 + ], + "type": "text", + "content": "This experiment considered all the images in the dataset. Observe that all values of the row " + }, + { + "bbox": [ + 313, + 168, + 553, + 239 + ], + "type": "inline_equation", + "content": "\\ell = 64" + }, + { + "bbox": [ + 313, + 168, + 553, + 239 + ], + "type": "text", + "content": " have lower PSNR, showing that an INR with very high frequencies performs worse even for small percentages " + }, + { + "bbox": [ + 313, + 168, + 553, + 239 + ], + "type": "inline_equation", + "content": "(30\\%)" + }, + { + "bbox": [ + 313, + 168, + 553, + 239 + ], + "type": "text", + "content": ". Similarly, the performance worsens when we increase the number of high frequencies (last column)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 240, + 554, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 554, + 348 + ], + "type": "text", + "content": "We now test the influence of high frequencies by training an INR of size " + }, + { + "bbox": [ + 313, + 240, + 554, + 348 + ], + "type": "inline_equation", + "content": "m = 360" + }, + { + "bbox": [ + 313, + 240, + 554, + 348 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 240, + 554, + 348 + ], + "type": "inline_equation", + "content": "n = 512" + }, + { + "bbox": [ + 313, + 240, + 554, + 348 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 240, + 554, + 348 + ], + "type": "inline_equation", + "content": "\\theta = 256" + }, + { + "bbox": [ + 313, + 240, + 554, + 348 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 313, + 240, + 554, + 348 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 240, + 554, + 348 + ], + "type": "text", + "content": " of the ground truth pixels. We vary the parameter " + }, + { + "bbox": [ + 313, + 240, + 554, + 348 + ], + "type": "inline_equation", + "content": "\\ell = 60, 120, 220" + }, + { + "bbox": [ + 313, + 240, + 554, + 348 + ], + "type": "text", + "content": " that defines low frequencies. Increasing the number of higher frequencies, the INR suffers from noise and overfitting. This can be noted in the reconstruction on the unsupervised pixels and gradients in Fig 8. Finally, in Supp. Mat. we compare our initialization of W against SIREN and observe a faster convergence." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 339, + 360, + 400, + 475 + ], + "blocks": [ + { + "bbox": [ + 339, + 360, + 400, + 475 + ], + "lines": [ + { + "bbox": [ + 339, + 360, + 400, + 475 + ], + "spans": [ + { + "bbox": [ + 339, + 360, + 400, + 475 + ], + "type": "image", + "image_path": "d152895bb551ad1e678149fb5197ba69982a42fe69ad3d4f8cffbe8b976889d7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 479, + 554, + 502 + ], + "lines": [ + { + "bbox": [ + 313, + 479, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 554, + 502 + ], + "type": "text", + "content": "Figure 8. Re constructions with big " + }, + { + "bbox": [ + 313, + 479, + 554, + 502 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 313, + 479, + 554, + 502 + ], + "type": "text", + "content": " present overfitting on the " + }, + { + "bbox": [ + 313, + 479, + 554, + 502 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 479, + 554, + 502 + ], + "type": "text", + "content": " unsupervised points. Gradient on the upper right corner." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 401, + 361, + 463, + 475 + ], + "blocks": [ + { + "bbox": [ + 401, + 361, + 463, + 475 + ], + "lines": [ + { + "bbox": [ + 401, + 361, + 463, + 475 + ], + "spans": [ + { + "bbox": [ + 401, + 361, + 463, + 475 + ], + "type": "image", + "image_path": "ef215ca3acb12cbbf4fb9fd785a437353c53385489dee7244dda4bea2dd6ece7.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 465, + 361, + 528, + 475 + ], + "blocks": [ + { + "bbox": [ + 465, + 361, + 528, + 475 + ], + "lines": [ + { + "bbox": [ + 465, + 361, + 528, + 475 + ], + "spans": [ + { + "bbox": [ + 465, + 361, + 528, + 475 + ], + "type": "image", + "image_path": "4cac0607a8df965aaae5c75da9ba368b9c9af65f8e0f377c39ff3be2b84168c5.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 514, + 383, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 514, + 383, + 528 + ], + "spans": [ + { + "bbox": [ + 314, + 514, + 383, + 528 + ], + "type": "text", + "content": "5.2. Bounding" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 533, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 554, + 593 + ], + "type": "text", + "content": "In Sec 4.2 we presented a scheme to bound the INR spectrum. Here, we show additional experiments to demonstrate that this improves the signal and has a great benefit over the unsupervised gradient reconstruction. Then, we give a comparison between fixed and trained bounding schemes." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "Fig 9 compares the reconstructions of an image of size " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "512^2" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " w/o/w our fixed bounding scheme. To demonstrate the impact of our scheme, we sample the input frequencies using " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\ell = 256" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\ell = 100" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": ". Without bounding (left), we effectively reconstruct the signal at the supervised pixels (30.8 dB), but with a noisy gradient (15.8 dB). That is, we are overfitting the signal. Conversely, bounding the hidden weights significantly improves both the signal (35.1 dB) and its gradient (27.8 dB). Our bounding scheme acts like a filter in both signal and gradient reconstruction." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3076" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 72, + 281, + 172 + ], + "blocks": [ + { + "bbox": [ + 69, + 72, + 281, + 172 + ], + "lines": [ + { + "bbox": [ + 69, + 72, + 281, + 172 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 281, + 172 + ], + "type": "image", + "image_path": "778310baad253422766e06d6928afb5b93abce72d0a58facd8551d044e068e22.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 174, + 295, + 207 + ], + "lines": [ + { + "bbox": [ + 55, + 174, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 55, + 174, + 295, + 207 + ], + "type": "text", + "content": "Figure 9. Comparison of training " + }, + { + "bbox": [ + 55, + 174, + 295, + 207 + ], + "type": "inline_equation", + "content": "\\mathrm{wo} / \\mathrm{w}" + }, + { + "bbox": [ + 55, + 174, + 295, + 207 + ], + "type": "text", + "content": " fixed bounds. Note that the use of bounds preserves high order information of the signal. Left side of the figures show the network gradients." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 223, + 296, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 223, + 296, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 296, + 368 + ], + "type": "text", + "content": "We compare quantitatively our bounding scheme over the dataset, varying the parameters " + }, + { + "bbox": [ + 55, + 223, + 296, + 368 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 223, + 296, + 368 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 223, + 296, + 368 + ], + "type": "inline_equation", + "content": "c_{\\mathbf{H}}" + }, + { + "bbox": [ + 55, + 223, + 296, + 368 + ], + "type": "text", + "content": ". Table 2 shows an improvement in the signal for all cases. Furthermore, the unsupervised gradient of the image is better learned when using fixed bounds, showing an average increase of 9.6dB against the standard training. Also, the last two lines of the table show the impact of the parameters " + }, + { + "bbox": [ + 55, + 223, + 296, + 368 + ], + "type": "inline_equation", + "content": "c_{\\mathbf{L}}" + }, + { + "bbox": [ + 55, + 223, + 296, + 368 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 223, + 296, + 368 + ], + "type": "inline_equation", + "content": "c_{\\mathbf{H}}" + }, + { + "bbox": [ + 55, + 223, + 296, + 368 + ], + "type": "text", + "content": " over the signal and gradient reconstructions. We consider training with " + }, + { + "bbox": [ + 55, + 223, + 296, + 368 + ], + "type": "inline_equation", + "content": "c_{\\mathbf{L}} > c_{\\mathbf{H}}" + }, + { + "bbox": [ + 55, + 223, + 296, + 368 + ], + "type": "text", + "content": " since we are prioritizing low input frequencies in the spectrum generation. Otherwise, this may generate high frequencies, possibly bypassing the Nyquist limit and resulting in a worse gradient reconstruction." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 380, + 298, + 434 + ], + "blocks": [ + { + "bbox": [ + 56, + 380, + 298, + 434 + ], + "lines": [ + { + "bbox": [ + 56, + 380, + 298, + 434 + ], + "spans": [ + { + "bbox": [ + 56, + 380, + 298, + 434 + ], + "type": "table", + "html": "
signalgrad
θcLcHwo boundw boundwo boundw bound
2561.50.0532.634.814.525.8
1901.50.0534.035.918.526.1
1901.50.234.034.418.521.0
", + "image_path": "8c6552a5dad013456f80c37aad23164c07195045efa9b4b21d49f1d12723ccd8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 434, + 296, + 491 + ], + "lines": [ + { + "bbox": [ + 55, + 434, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 296, + 491 + ], + "type": "text", + "content": "Table 2. Training with bounding improves the signal and grad. reconstruction. The choice of the fixed bound " + }, + { + "bbox": [ + 55, + 434, + 296, + 491 + ], + "type": "inline_equation", + "content": "c_{\\mathbf{H}}" + }, + { + "bbox": [ + 55, + 434, + 296, + 491 + ], + "type": "text", + "content": " greatly impacts the grad. fidelity. The std of the difference between each image signal (grad) quality " + }, + { + "bbox": [ + 55, + 434, + 296, + 491 + ], + "type": "inline_equation", + "content": "w / w" + }, + { + "bbox": [ + 55, + 434, + 296, + 491 + ], + "type": "text", + "content": " bounding is 1.4dB (1.6dB) on average. Evaluations performed on the test set (10% of pixels)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 509, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 581 + ], + "type": "text", + "content": "Next, we show how we can learn each bound " + }, + { + "bbox": [ + 55, + 509, + 295, + 581 + ], + "type": "inline_equation", + "content": "c_{j}" + }, + { + "bbox": [ + 55, + 509, + 295, + 581 + ], + "type": "text", + "content": ". As described in Sec 4.2, we can incorporate a learnable " + }, + { + "bbox": [ + 55, + 509, + 295, + 581 + ], + "type": "inline_equation", + "content": "c_{j}" + }, + { + "bbox": [ + 55, + 509, + 295, + 581 + ], + "type": "text", + "content": " for each " + }, + { + "bbox": [ + 55, + 509, + 295, + 581 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 509, + 295, + 581 + ], + "type": "text", + "content": "-column of the hidden matrix through a modified layer (7). Fig 10 shows a comparison between our fixed bound scheme using initial un-tuned " + }, + { + "bbox": [ + 55, + 509, + 295, + 581 + ], + "type": "inline_equation", + "content": "c_{\\mathbf{L}} = 1.0" + }, + { + "bbox": [ + 55, + 509, + 295, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 509, + 295, + 581 + ], + "type": "inline_equation", + "content": "c_{\\mathbf{H}} = 0.6" + }, + { + "bbox": [ + 55, + 509, + 295, + 581 + ], + "type": "text", + "content": " (left) and the learned bound scheme (middle)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "As expected, training lowers the bounds for high frequencies (H) while maintaining or increasing the bounds of low frequencies (L). Thus, we achieve better reconstruction in both the signal and gradient. In Fig 10 (right), we note that low frequencies (left of the green line) have significant bounds, showing their importance during training. These schemes allow us to initialize with higher frequencies than " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "B / 3" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": ", accelerating convergence. Finally, we performed an ablation study on the regularization term " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": " to show its importance. Training w/wo regularization yielded 28.5/27.4dB for the RGB and 27.1/24.0dB for the gradient." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 72, + 553, + 157 + ], + "blocks": [ + { + "bbox": [ + 318, + 72, + 553, + 157 + ], + "lines": [ + { + "bbox": [ + 318, + 72, + 553, + 157 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 553, + 157 + ], + "type": "image", + "image_path": "5ef1823b9863aebbefc72d4e3b1f85b01bdc63b3fb68ee3be83ae389fc04298f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 159, + 555, + 205 + ], + "lines": [ + { + "bbox": [ + 313, + 159, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 555, + 205 + ], + "type": "text", + "content": "Figure 10. Comparison of signal and gradient (grayscale) when training with fixed/learned bounds. The learned bounds adjust the INR spectrum resulting in better reconstructions. Each blue point represents a pair " + }, + { + "bbox": [ + 313, + 159, + 555, + 205 + ], + "type": "inline_equation", + "content": "(|\\omega_j|_{\\infty}, c_j)" + }, + { + "bbox": [ + 313, + 159, + 555, + 205 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 159, + 555, + 205 + ], + "type": "inline_equation", + "content": "c_j" + }, + { + "bbox": [ + 313, + 159, + 555, + 205 + ], + "type": "text", + "content": " is the trained bound of " + }, + { + "bbox": [ + 313, + 159, + 555, + 205 + ], + "type": "inline_equation", + "content": "\\omega_j" + }, + { + "bbox": [ + 313, + 159, + 555, + 205 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 225, + 400, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 225, + 400, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 400, + 239 + ], + "type": "text", + "content": "5.3. Comparisons" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 243, + 555, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 448 + ], + "type": "text", + "content": "We first compare our method (TUNER) with SIREN, which requires choosing a threshold parameter " + }, + { + "bbox": [ + 313, + 243, + 555, + 448 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 243, + 555, + 448 + ], + "type": "text", + "content": ", a critical task since higher " + }, + { + "bbox": [ + 313, + 243, + 555, + 448 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 243, + 555, + 448 + ], + "type": "text", + "content": " implies noisy reconstructions. TUNER also uses " + }, + { + "bbox": [ + 313, + 243, + 555, + 448 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 243, + 555, + 448 + ], + "type": "text", + "content": " to control the frequency generation, however, it provides more stable training with better reconstruction. To demonstrate this, we perform a comparison by varying " + }, + { + "bbox": [ + 313, + 243, + 555, + 448 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 243, + 555, + 448 + ], + "type": "text", + "content": " and the number of input (" + }, + { + "bbox": [ + 313, + 243, + 555, + 448 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 243, + 555, + 448 + ], + "type": "text", + "content": ") and hidden (" + }, + { + "bbox": [ + 313, + 243, + 555, + 448 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 243, + 555, + 448 + ], + "type": "text", + "content": ") neurons. First, consider the case where we have no information about the signal bandlimit besides the Nyquist limit " + }, + { + "bbox": [ + 313, + 243, + 555, + 448 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 243, + 555, + 448 + ], + "type": "text", + "content": ". Fig 11 presents a comparison between SIREN (top) and TUNER (bottom) with " + }, + { + "bbox": [ + 313, + 243, + 555, + 448 + ], + "type": "inline_equation", + "content": "\\theta = B" + }, + { + "bbox": [ + 313, + 243, + 555, + 448 + ], + "type": "text", + "content": ". Note that SIREN produces noisy reconstructions, while ours learn faster and does not suffer from noise. This is due to our bounding scheme, which limits the generation of high frequencies and provides stable training (Sec 4.2). In contrast, SIREN initializes its parameters such that the spectrum is initially bounded but does not maintain this guarantee during training." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 456, + 556, + 589 + ], + "blocks": [ + { + "bbox": [ + 317, + 456, + 556, + 589 + ], + "lines": [ + { + "bbox": [ + 317, + 456, + 556, + 589 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 556, + 589 + ], + "type": "image", + "image_path": "421ae920815b31ddb6f15579eedb9daa82c3912508ad9fd1a99d11fbdc6c4ee7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 593, + 555, + 649 + ], + "lines": [ + { + "bbox": [ + 313, + 593, + 555, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 555, + 649 + ], + "type": "text", + "content": "Figure 11. Comparing SIREN (top) and TUNER (bottom) when we only know the Nyquist limit (i.e, " + }, + { + "bbox": [ + 313, + 593, + 555, + 649 + ], + "type": "inline_equation", + "content": "\\theta = B" + }, + { + "bbox": [ + 313, + 593, + 555, + 649 + ], + "type": "text", + "content": "). Since " + }, + { + "bbox": [ + 313, + 593, + 555, + 649 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 593, + 555, + 649 + ], + "type": "text", + "content": " is big, SIREN introduces high frequencies at the beginning of training, resulting in very noisy reconstructions. Conversely, ours starts training low frequencies, converging faster and being robust to " + }, + { + "bbox": [ + 313, + 593, + 555, + 649 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 593, + 555, + 649 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 657, + 555, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 657, + 555, + 694 + ], + "spans": [ + { + "bbox": [ + 313, + 657, + 555, + 694 + ], + "type": "text", + "content": "Next, we consider a tuned " + }, + { + "bbox": [ + 313, + 657, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 313, + 657, + 555, + 694 + ], + "type": "text", + "content": " for SIREN to show that even under this condition, our method provides a better training scheme for sinusoidal INRs. Fig 12 presents the comparison" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 405, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 405, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 405, + 713 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 325, + 702, + 405, + 713 + ], + "type": "text", + "content": "Denoted as " + }, + { + "bbox": [ + 325, + 702, + 405, + 713 + ], + "type": "inline_equation", + "content": "\\omega_0" + }, + { + "bbox": [ + 325, + 702, + 405, + 713 + ], + "type": "text", + "content": " in [26]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3077" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "varying the number of input " + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "inline_equation", + "content": "(m)" + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": " and hidden " + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "inline_equation", + "content": "(n)" + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": " neurons. We consider the cases " + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "inline_equation", + "content": "m < n" + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": " (top) and " + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "inline_equation", + "content": "m > n" + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": " (bottom), and TUNER performed better both in the signal/gradient reconstructions. Also, we note more robust training using TUNER, as shown in Fig 12(right)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 294, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 294, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 294, + 193 + ], + "type": "text", + "content": "The previous experiment shows that initializing SIREN with a high " + }, + { + "bbox": [ + 55, + 133, + 294, + 193 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 133, + 294, + 193 + ], + "type": "text", + "content": " results in overfitting, while TUNER provides better reconstruction. On the other hand, with the same architecture and a tuned " + }, + { + "bbox": [ + 55, + 133, + 294, + 193 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 133, + 294, + 193 + ], + "type": "text", + "content": ", our method also effectively learned the lower frequencies first and performed better (Fig 12)." + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 202, + 293, + 318 + ], + "blocks": [ + { + "bbox": [ + 56, + 202, + 293, + 318 + ], + "lines": [ + { + "bbox": [ + 56, + 202, + 293, + 318 + ], + "spans": [ + { + "bbox": [ + 56, + 202, + 293, + 318 + ], + "type": "image", + "image_path": "1df572d89bb0dc7231d1830ad4041c4b386deef8552365012b69a9f7d4c4d142.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 328, + 295, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 328, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 295, + 361 + ], + "type": "text", + "content": "Figure 12. Comparison with SIREN when " + }, + { + "bbox": [ + 55, + 328, + 295, + 361 + ], + "type": "inline_equation", + "content": "\\theta = 40" + }, + { + "bbox": [ + 55, + 328, + 295, + 361 + ], + "type": "text", + "content": ". TUNER improves training convergence in both cases (" + }, + { + "bbox": [ + 55, + 328, + 295, + 361 + ], + "type": "inline_equation", + "content": "m < n" + }, + { + "bbox": [ + 55, + 328, + 295, + 361 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 328, + 295, + 361 + ], + "type": "inline_equation", + "content": "m > n" + }, + { + "bbox": [ + 55, + 328, + 295, + 361 + ], + "type": "text", + "content": ") and stability when " + }, + { + "bbox": [ + 55, + 328, + 295, + 361 + ], + "type": "inline_equation", + "content": "m > n" + }, + { + "bbox": [ + 55, + 328, + 295, + 361 + ], + "type": "text", + "content": ". INR gradient shown in grayscale." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 374, + 296, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 296, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 296, + 447 + ], + "type": "text", + "content": "Next, we numerically evaluate these experiments, providing a quantitative comparison with SIREN and BACON. We compare different architectures " + }, + { + "bbox": [ + 55, + 374, + 296, + 447 + ], + "type": "inline_equation", + "content": "(m,n)" + }, + { + "bbox": [ + 55, + 374, + 296, + 447 + ], + "type": "text", + "content": " and spectrum bandlimits " + }, + { + "bbox": [ + 55, + 374, + 296, + 447 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 374, + 296, + 447 + ], + "type": "text", + "content": ", evaluating the results in terms of signal and gradient PSNR over the supervised (Train) and unsupervised (Test) pixels. Table 3 presents these evaluations." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 56, + 455, + 299, + 574 + ], + "blocks": [ + { + "bbox": [ + 56, + 455, + 299, + 574 + ], + "lines": [ + { + "bbox": [ + 56, + 455, + 299, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 299, + 574 + ], + "type": "table", + "html": "
TrainTest
SignalGradientSignalGradient
θ(m,n)SIRENBACOURSSIRENBACOURSSIRENBACOURSSIRENBACOURS
851028,16232.9-32.926.5-28.831.9-32.225.5-28.2
416,41634.026.434.327.722.428.932.925.433.427.122.328.4
80,202432.2-33.126.6-28.531.3-32.326.0-28.0
1711028,16231.3-33.922.0-27.222.9-32.920.7-26.8
416,41629.229.334.417.426.727.026.726.233.116.226.326.4
80,202426.6-30.814.2-22.623.8-29.513.8-21.3
2561028,16226.8-33.312.8-24.724.1-31.912.7-23.9
416,41624.727.741.212.924.223.022.322.733.213.222.821.2
80,202425.1-30.013.0-15.224.1-27.113.1-14.6
", + "image_path": "197b36174bb296bbb8d24f7a0a536f3e25d2fded384cff7a9121bcf7be4e64d4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 583, + 295, + 617 + ], + "lines": [ + { + "bbox": [ + 55, + 583, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 583, + 295, + 617 + ], + "type": "text", + "content": "Table 3. Comparison between SIREN, BACON (BAC), and our method. We use several architectures " + }, + { + "bbox": [ + 55, + 583, + 295, + 617 + ], + "type": "inline_equation", + "content": "(m,n)" + }, + { + "bbox": [ + 55, + 583, + 295, + 617 + ], + "type": "text", + "content": " ,bandlimits " + }, + { + "bbox": [ + 55, + 583, + 295, + 617 + ], + "type": "inline_equation", + "content": "(\\theta)" + }, + { + "bbox": [ + 55, + 583, + 295, + 617 + ], + "type": "text", + "content": " ,and evaluate over the train " + }, + { + "bbox": [ + 55, + 583, + 295, + 617 + ], + "type": "inline_equation", + "content": "(90\\%)" + }, + { + "bbox": [ + 55, + 583, + 295, + 617 + ], + "type": "text", + "content": " and test " + }, + { + "bbox": [ + 55, + 583, + 295, + 617 + ], + "type": "inline_equation", + "content": "(10\\%)" + }, + { + "bbox": [ + 55, + 583, + 295, + 617 + ], + "type": "text", + "content": " sets of signal/gradient." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "text", + "content": "As expected, SIREN performs well with " + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\theta = \\frac{B}{3}" + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "text", + "content": ". In this case, TUNER is perceptually similar to SIREN but improves gradient reconstruction and consistently outperforms BACON. When " + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\theta = 171" + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "text", + "content": ", TUNER offers superior quality in both signal and gradient, showing greater robustness to noise and superior performance even with " + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\theta \\geq B/3" + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "text", + "content": ". Finally, with " + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\theta = 256" + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "text", + "content": ", our method surpasses SIREN in all cases." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "For BACON, our method achieves better signal reconstruction, with comparable gradients. However, the following experiment shows that BACON's reconstruction has visible defects in this scenario. Note that we do not supervise the training with gradient information." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 146, + 555, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 146, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 146, + 555, + 301 + ], + "type": "text", + "content": "Bandlimit comparison: Lastly, Fig 13 presents a qualitative comparison between BACON, BANF, and TUNER (Ours) with bandlimits " + }, + { + "bbox": [ + 313, + 146, + 555, + 301 + ], + "type": "inline_equation", + "content": "\\theta = 60, 130" + }, + { + "bbox": [ + 313, + 146, + 555, + 301 + ], + "type": "text", + "content": ". We train all methods with 67K parameters. For BANF we use the linear filter and adjust grid resolution to match the bandlimits and the MLP size to approximate 67K parameters. Particularly, the spectra were computed with the same process for all three methods. Note that BACON truncates the spectrum with a hard filter, resulting in ringing artifacts (low bandlimit) and noisy reconstruction (high bandlimit). BANF shows distortions around the eye area for both filters, and the spectra show no clear bandlimit. Conversely, TUNER operates as a soft filter providing better reconstructions for all cases." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 321, + 303, + 394, + 409 + ], + "blocks": [ + { + "bbox": [ + 321, + 303, + 394, + 409 + ], + "lines": [ + { + "bbox": [ + 321, + 303, + 394, + 409 + ], + "spans": [ + { + "bbox": [ + 321, + 303, + 394, + 409 + ], + "type": "image", + "image_path": "d55aa1a130db4305156402adf5fce5539d522f49acb971b45ad4c0beae64d813.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 396, + 303, + 468, + 408 + ], + "blocks": [ + { + "bbox": [ + 396, + 303, + 468, + 408 + ], + "lines": [ + { + "bbox": [ + 396, + 303, + 468, + 408 + ], + "spans": [ + { + "bbox": [ + 396, + 303, + 468, + 408 + ], + "type": "image", + "image_path": "e598c70896dabb7dbe5c39f301d0361f292391302c5e87ef5d76e3559e647e77.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 470, + 303, + 542, + 408 + ], + "blocks": [ + { + "bbox": [ + 470, + 303, + 542, + 408 + ], + "lines": [ + { + "bbox": [ + 470, + 303, + 542, + 408 + ], + "spans": [ + { + "bbox": [ + 470, + 303, + 542, + 408 + ], + "type": "image", + "image_path": "a26d5d0407ef594c50a29fa94e040994c0d5de8bfb2ae6a2df3615075852d5b2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 321, + 410, + 394, + 516 + ], + "blocks": [ + { + "bbox": [ + 321, + 410, + 394, + 516 + ], + "lines": [ + { + "bbox": [ + 321, + 410, + 394, + 516 + ], + "spans": [ + { + "bbox": [ + 321, + 410, + 394, + 516 + ], + "type": "image", + "image_path": "f53784834ac65abc3e568fa9d79d8058eeb8a886e6c3b8b67e66c6968df1d1cf.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 520, + 553, + 575 + ], + "lines": [ + { + "bbox": [ + 313, + 520, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 553, + 575 + ], + "type": "text", + "content": "Figure 13. Comparison between BACON, BANF, and TUNER, trained with bandlimits " + }, + { + "bbox": [ + 313, + 520, + 553, + 575 + ], + "type": "inline_equation", + "content": "\\theta = 60, 130" + }, + { + "bbox": [ + 313, + 520, + 553, + 575 + ], + "type": "text", + "content": ". Note that BACON uses a box filter, generating ringing artifacts (left). BANF (middle) has artifacts near edges and ambiguous bandlimit. Then, TUNER (right) resembles a soft filter, improving quality." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 396, + 410, + 468, + 516 + ], + "blocks": [ + { + "bbox": [ + 396, + 410, + 468, + 516 + ], + "lines": [ + { + "bbox": [ + 396, + 410, + 468, + 516 + ], + "spans": [ + { + "bbox": [ + 396, + 410, + 468, + 516 + ], + "type": "image", + "image_path": "4c502bb2d474d7e6fd31efdbd0d9597287831b9cf1db675f569a9568efb9dd88.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 470, + 410, + 542, + 516 + ], + "blocks": [ + { + "bbox": [ + 470, + 410, + 542, + 516 + ], + "lines": [ + { + "bbox": [ + 470, + 410, + 542, + 516 + ], + "spans": [ + { + "bbox": [ + 470, + 410, + 542, + 516 + ], + "type": "image", + "image_path": "cddab5cc49355e18ad562340ff5054ef2fb3f4153d12fe3a514e777c09682106.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 586, + 471, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 471, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 471, + 598 + ], + "type": "text", + "content": "6. Conclusions and limitations" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "We presented a study based on Fourier theory for sinusoidal MLPs, which resulted in novel, robust schemes for both initialization and spectral control of such networks. While our experiments focused on images to validate our theoretical claims, future work includes applying our schemes to other signal types. Additionally, we have focused our studies on 3-layer MLPs, leaving the investigation of deep networks for future research. We consider this work a first step toward unraveling the powerful expressiveness of sinusoidal MLPs." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3078" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 96 + ], + "type": "text", + "content": "Acknowledgments We would like to thank CAPES, FAPERJ, and Google for partially funding this work." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 108, + 115, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 108, + 115, + 120 + ], + "spans": [ + { + "bbox": [ + 56, + 108, + 115, + 120 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 128, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 128, + 295, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 128, + 295, + 148 + ], + "spans": [ + { + "bbox": [ + 61, + 128, + 295, + 148 + ], + "type": "text", + "content": "[1] Kodak dataset: https://r0k.us/graphics/kodak/.3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 152, + 295, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 152, + 295, + 184 + ], + "spans": [ + { + "bbox": [ + 61, + 152, + 295, + 184 + ], + "type": "text", + "content": "[2] Milton Abramowitz and Irene A Stegun. Handbook of mathematical functions with formulas, graphs, and mathematical tables. US Government printing office, 1964. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 186, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 186, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 61, + 186, + 295, + 239 + ], + "type": "text", + "content": "[3] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 241, + 295, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 241, + 295, + 295 + ], + "spans": [ + { + "bbox": [ + 62, + 241, + 295, + 295 + ], + "type": "text", + "content": "[4] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628-8638, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 297, + 296, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 297, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 62, + 297, + 296, + 342 + ], + "type": "text", + "content": "[5] Yishun Dou, Zhong Zheng, Qiaoqiao Jin, and Bingbing Ni. Multiplicative fourier level of detail. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1808-1817, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 343, + 295, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 343, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 62, + 343, + 295, + 386 + ], + "type": "text", + "content": "[6] Emilien Dupont, Adam Golinski, Milad Alizadeh, Yee Whye Teh, and Arnaud Doucet. Coin: Compression with implicit neural representations. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 388, + 295, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 295, + 421 + ], + "type": "text", + "content": "[7] Rizal Fathony, Anit Kumar Sahu, Devin Willmott, and J Zico Kolter. Multiplicative filter networks. In International Conference on Learning Representations, 2020. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 422, + 295, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 295, + 464 + ], + "type": "text", + "content": "[8] Adam Kania, Marko Mihajlovic, Sergey Prokudin, Jacek Tabor, Przemys' Spurek, et al. Fresh: Frequency shifting for accelerated neural representation learning. arXiv preprint arXiv:2410.05050, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 467, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 467, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 467, + 295, + 498 + ], + "type": "text", + "content": "[9] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.3,6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 501, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 501, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 57, + 501, + 295, + 555 + ], + "type": "text", + "content": "[10] David B Lindell, Dave Van Veen, Jeong Joon Park, and Gordon Wetzstein. Bacon: Band-limited coordinate networks for multiscale scene representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16252-16262, 2022. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 557, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 295, + 623 + ], + "type": "text", + "content": "[11] Zhen Liu, Hao Zhu, Qi Zhang, Jingde Fu, Weibing Deng, Zhan Ma, Yanwen Guo, and Xun Cao. Finer: Flexible spectral-bias tuning in implicit neural representation by variable-periodic activation functions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2713-2722, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 624, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 668 + ], + "type": "text", + "content": "[12] Ishit Mehta, Manmohan Chandraker, and Ravi Ramamoorthi. A level set theory for neural implicit evolution under explicit flows. In European Conference on Computer Vision, pages 711-729. Springer, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[13] Tiago Novello, Guilherme Schardong, Luiz Schirmer, Vinicius da Silva, Helio Lopes, and Luiz Velho. Exploring differential geometry in neural implicits. Computers & Graphics, 108:49-60, 2022. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 126 + ], + "type": "text", + "content": "[14] Tiago Novello, Vinicius da Silva, Guilherme Schardong, Luiz Schirmer, Helio Lopes, and Luiz Velho. Neural implicit surface evolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14279-14289, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 129, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 162 + ], + "type": "text", + "content": "[15] Giambattista Parascandolo, Heikki Huttunen, and Tuomas Virtanen. Taming the waves: sine as activation function in deep neural networks. 2016. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 164, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 217 + ], + "type": "text", + "content": "[16] Hallison Paz, Daniel Perazzo, Tiago Novello, Guilherme Schardong, Luiz Schirmer, Vinicius da Silva, Daniel Yukimura, Fabio Chagas, Helio Lopes, and Luiz Velho. Mrnet: Multiresolution sinusoidal neural networks. Computers & Graphics, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "type": "text", + "content": "[17] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International Conference on Machine Learning, pages 5301-5310. PMLR, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "type": "text", + "content": "[18] E. Romero, J. Sopena, R. Alquezar, and J. Moliner. Neural networks with periodic and monotonic activation functions: a comparative study in classification problems. 2000. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 310, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 310, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 317, + 310, + 553, + 363 + ], + "type": "text", + "content": "[19] Vishwanath Saragadam, Jasper Tan, Guha Balakrishnan, Richard G Baraniuk, and Ashok Veeraraghavan. Miner: Multiscale implicit neural representation. In European Conference on Computer Vision, pages 318-333. Springer, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 366, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 421 + ], + "type": "text", + "content": "[20] Vishwanath Saragadam, Daniel LeJeune, Jasper Tan, Guha Balakrishnan, Ashok Veeraraghavan, and Richard G Baraniuk. Wire: Wavelet implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18507-18516, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 422, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 476 + ], + "type": "text", + "content": "[21] Hemanth Saratchandran, Sameera Ramasinghe, Violetta Shevchenko, Alexander Long, and Simon Lucey. A sampling theory perspective on activations for implicit neural representations. In International Conference on Machine Learning, pages 43422-43444. PMLR, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 478, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 553, + 533 + ], + "type": "text", + "content": "[22] Guilherme Schardong, Tiago Novello, Hallison Paz, Iurii Medvedev, Vinícius da Silva, Luiz Velho, and Nuno Gonçalves. Neural implicit morphing of face images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7321-7330, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "type": "text", + "content": "[23] Luiz Schirmer, Tiago Novello, Guilherme Schardong, Vinicius da Silva, Helio Lopes, and Luiz Velho. How to train your (neural) dragon. In Conference on Graphics, Patterns and Images, 36. (SIBGRAPI), 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 580, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 553, + 644 + ], + "type": "text", + "content": "[24] Akhmedkhan Shabanov, Shrisudhan Govindarajan, Cody Reading, Lily Goli, Daniel Rebain, Kwang Moo Yi, and Andrea Tagliasacchi. Banf: Band-limited neural fields for levels of detail reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20571-20580, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "type": "text", + "content": "[25] Kathan Shah and Chawin Sitawarin. Spider: Semiperiodic damping-enabled object representation. arXiv preprint arXiv:2306.15242, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[26] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in neural" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3079" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 553 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "information processing systems, 33:7462-7473, 2020. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "text", + "content": "[27] Ying Song, Jiaping Wang, Li-Yi Wei, and Wencheng Wang. Vector regression functions for texture compression. ACM Transactions on Graphics (TOG), 35(1):1-10, 2015. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 294, + 195 + ], + "type": "text", + "content": "[28] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "text", + "content": "[29] Zhijie Wu, Yuhe Jin, and Kwang Moo Yi. Neural fourier filter bank. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14153-14163, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 297 + ], + "type": "text", + "content": "[30] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, pages 641-676. Wiley Online Library, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "type": "text", + "content": "[31] Guandao Yang, Serge Belongie, Bharath Hariharan, and Vladlen Koltun. Geometry processing with neural fields. Advances in Neural Information Processing Systems, 34: 22483-22497, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 342, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 396 + ], + "type": "text", + "content": "[32] Guandao Yang, Sagie Benaim, Varun Jampani, Kyle Genova, Jonathan Barron, Thomas Funkhouser, Bharath Hariharan, and Serge Belongie. Polynomial neural fields for subband decomposition and manipulation. Advances in Neural Information Processing Systems, 35:4401-4415, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 398, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 398, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 398, + 294, + 441 + ], + "type": "text", + "content": "[33] Wang Yifan, Lukas Rahmann, and Olga Sorkine-hornung. Geometry-consistent neural shape representation with implicit displacement fields. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 443, + 294, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 443, + 294, + 497 + ], + "spans": [ + { + "bbox": [ + 56, + 443, + 294, + 497 + ], + "type": "text", + "content": "[34] Gizem Yüce, Guillermo Ortiz-Jiménez, Beril Besbinar, and Pascal Frossard. A structured dictionary perspective on implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19228-19238, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 498, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 294, + 553 + ], + "type": "text", + "content": "[35] Andreas Zell, Nuri Benbarka, Timon Hoefer, and Hamd Ul Moqueet Riaz. Seeing implicit neural representations as fourier series. In 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). IEEE Computer Society, 2022. 1, 2" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3080" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Turbo3D_ Ultra-fast Text-to-3D Generation/bd4523a6-3f35-4dc6-9b48-90da3f36ebea_content_list.json b/2025/Turbo3D_ Ultra-fast Text-to-3D Generation/bd4523a6-3f35-4dc6-9b48-90da3f36ebea_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f13471b8ff98c6d187795252bc46ca134fa9f4a4 --- /dev/null +++ b/2025/Turbo3D_ Ultra-fast Text-to-3D Generation/bd4523a6-3f35-4dc6-9b48-90da3f36ebea_content_list.json @@ -0,0 +1,2295 @@ +[ + { + "type": "text", + "text": "Turbo3D: Ultra-fast Text-to-3D Generation", + "text_level": 1, + "bbox": [ + 279, + 130, + 718, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hanzhe Hu", + "bbox": [ + 114, + 179, + 212, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tianwei Yin²", + "bbox": [ + 253, + 180, + 359, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sai Bi3", + "bbox": [ + 279, + 198, + 338, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fujun Luan3", + "bbox": [ + 398, + 180, + 500, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yiwei Hu3", + "bbox": [ + 539, + 180, + 622, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shubham Tulsiani\\*", + "bbox": [ + 403, + 199, + 563, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao Tan3", + "bbox": [ + 663, + 180, + 738, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zexiang $\\mathbf{X}\\mathbf{u}^{3}$", + "bbox": [ + 777, + 180, + 880, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kai Zhang $^{3*}$", + "bbox": [ + 624, + 198, + 723, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Carnegie Mellon University", + "bbox": [ + 127, + 232, + 362, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Massachusetts Institute of Technology", + "bbox": [ + 380, + 232, + 697, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Adobe Research", + "bbox": [ + 718, + 232, + 864, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://turbo-3d.github.io/", + "bbox": [ + 374, + 272, + 614, + 286 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e517bb0932019afa3b6d064b2be84d827a893f280ce57e0ff3e0a18e77382d7f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 316, + 859, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4279294ded69c7d28972ea4ec79bd89e461d7ade553c6f814fec3a8fa209a81c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 463, + 890, + 539 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a panda wearing a chef's hat and kneading bread dough on a countertop", + "bbox": [ + 109, + 542, + 369, + 568 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f94ab1b71cc78cd9b77f90ea0f79db391c63779992bee7d21d5870eacacb684d.jpg", + "image_caption": [ + "a plush toy of a cordi nurse" + ], + "image_footnote": [], + "bbox": [ + 114, + 574, + 173, + 637 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/36163932441e8a06763f9c45ba4180e6edfaf2c25c17f8ee2ccb40f610dea4c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 205, + 574, + 264, + 637 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b79dc2f7e7464eb5107d954e23a323ae8e8c7ddb1ea41bf975898202dbb59a0b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 575, + 359, + 637 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7f0e3292aa9bb02f38ad181b9f4635d07e68a49c733c7c9e53a2189927899fb0.jpg", + "image_caption": [ + "a raccoon stealing a pie" + ], + "image_footnote": [], + "bbox": [ + 380, + 571, + 442, + 637 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8ceb6de2b254accb7f27384fe948e4c52ab759108ac308def4abd800a9043704.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 464, + 571, + 537, + 638 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dc7e2cfa88f59aa1a7f182cad879b1370cd4e5b66f666f45ceb9b9206faaf147.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 553, + 571, + 625, + 640 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6c7e4a00eb82f0144894ba5fc51e4b6184341526b0bccdda3c552ab82b4eb07f.jpg", + "image_caption": [ + "a badger wearing a party hat and blowing out birthday candles on a cake" + ], + "image_footnote": [], + "bbox": [ + 643, + 569, + 710, + 635 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d31a75e94589ea056799ae129d2e46eb6d1fd259d5f602602c067be7629e7843.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 565, + 805, + 633 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ab0e4bd55bb548d4f610b56d92c6e74be0c0335d8c6ab5862346601347ef4245.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 826, + 565, + 888, + 627 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/74cedcfe912707107c8fd02329a4dc14d8f4bf46094c60cb960f6f313870a634.jpg", + "image_caption": [ + "a cake in the shape of a train" + ], + "image_footnote": [], + "bbox": [ + 102, + 674, + 173, + 736 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/71fea974c24db610363a01e4cd0921b13bbb7c0c4f1fc2190549b3e3ed07cd22.jpg", + "image_caption": [ + "Figure 1. Overview of our Turbo3D text-to-3D system. Turbo3D generates high-quality 3D Gaussian Splatting (3DGS) assets from user prompts in less than 1 second on a single A100 GPU. It's a two-stage pipeline consisting of a highly efficient latent-space few-step multi-view (MV) generator and single-step MV reconstructor. Note that we visualize latents as RGB images and 3DGS assets as point clouds in the pipeline figure for clarity." + ], + "image_footnote": [], + "bbox": [ + 186, + 676, + 271, + 737 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9c8c1f7112d652362c2dca86db6a3f6844c498f0d1b27f02f842287aacf0ff29.jpg", + "image_caption": [ + "Campsite with a tent on grassy ground, a backpack against it, a campfire surrounded by stones. A sleeping bag partially unrolled next to the tent, a lantern hanging from a tree branch" + ], + "image_footnote": [], + "bbox": [ + 281, + 674, + 366, + 736 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2c23987bc45820a6509323818894aaee7523488f07e8f198e65441ee0d98c008.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 667, + 454, + 732 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a0460c86b4c0cb4a7512daf422b0239c7d3ff89010d1bd9c4c1d7865d45a18a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 455, + 667, + 550, + 734 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f00753f7bb123a1f3dcc94411851aa062787564c4df60e4beec670c909478467.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 552, + 675, + 638, + 733 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/32e803b35fabdcae2a9691bf90b9badb9ef65ba190ea5addaa0d36edf28c4ecf.jpg", + "image_caption": [ + "a lion reading the newspaper" + ], + "image_footnote": [], + "bbox": [ + 650, + 676, + 727, + 736 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c586f16699d00a7d3b144bcd03349b9156805d01276c433c0ad4a2c6e47d2618.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 730, + 678, + 816, + 736 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9ac1abafbf1804825eeeba62e9044c45142c8404de4632205e99720ac9f4ef35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 820, + 678, + 903, + 736 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal advising.", + "bbox": [ + 109, + 897, + 199, + 910 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23668", + "bbox": [ + 478, + 950, + 517, + 960 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 90, + 326, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We present Turbo3D, an ultra-fast text-to-3D system capable of generating high-quality Gaussian splatting assets in under one second. Turbo3D employs a rapid 4-step, 4-view diffusion generator and an efficient feed-forward Gaussian reconstructor, both operating in latent space, as shown in Fig. 1. The 4-step, 4-view generator is a student model distilled through a novel Dual-Teacher approach, which encourages the student to learn view consistency from a multiview teacher and photo-realism from a single-view teacher. By shifting the Gaussian reconstructor's inputs from pixel space to latent space, we eliminate the extra image decoding time and halve the transformer sequence length for maximum efficiency. Our method demonstrates superior 3D generation results compared to previous baselines, while operating in a fraction of their runtime.", + "bbox": [ + 89, + 121, + 485, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 376, + 222, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The recent advances in image generative models allow users to generate detailed outputs from just a text prompt. While initial denoising diffusion-based methods [7, 34, 36, 37, 44, 50] enabled impressive photo-realistic generation, recent techniques [19, 22, 23, 39, 45, 60, 61] have significantly improved inference efficiency of such models, allowing high-fidelity generation in the blink of an eye. Unfortunately, these advances in generative 2D modeling methods have not yet been matched in the 3D domain, where ultra-fast realistic 3D generation remains a challenge. In this work, we seek to bridge this gap, and present Turbo3D, a text-to-3D generative model that can synthesize detailed 3D outputs in a fraction of a second.", + "bbox": [ + 89, + 402, + 483, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The existing approaches for text-based 3D inference can be categorized as either generative [10, 17, 24, 30, 57, 63] or optimization-driven [18, 32, 55]. The latter class of methods optimize 3D representations by 'distilling' pretrained 2D diffusion models [32]. While this approach can yield decent 3D outputs, it is highly inefficient, typically requiring several minutes or even hours to output a single 3D representation. The alternative paradigm is to learn a generative model that directly outputs 3D representations. While initial methods [10, 28] investigated representations such as point clouds and SDFs, the limited availability of 3D data restricted the generation quality. Recent approaches [9, 17, 24, 40, 41] have instead advocated for learning generative models of multi-view images (followed by deterministic 3D reconstruction), as these can be initialized from pre-trained 2D generative models. While these methods have resulted in impressive generations, the multiview finetuning on synthetic data does inhibit their generation quality. More crucially, their inference efficiency is restricted by the iterative denoising process required for the", + "bbox": [ + 89, + 599, + 485, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "text-conditioned multi-view generation.", + "bbox": [ + 511, + 90, + 777, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we follow the paradigm of text-to-3D via multi-view generation, and aim to improve the efficiency of the underlying components to enable ultra-fast generation, while enhancing the fidelity of the generated outputs. Inspired by the recent progress in reducing inference time by distilling 2D diffusion models into one-step or few-step generators [38, 39, 45, 61], we adapt these for multi-view 3D generation. We first train a many-step text-to-multiview diffusion model and then distill it into a much faster 4-step generator using the distribution matching distillation (DMD) loss [61]. However, we find that this process leads to a significant quality degradation, as it fails to capture the full range of modes present in the multi-view teacher model. To overcome this, we propose to extend the DMD pipeline to incorporate another single-view teacher - a 2D denoising diffusion model trained on a large set of high quality aesthetic images. Our few-step multi-view generator is thus trained with a dual-teacher distillation approach, where the multi-view DMD loss helps our model learn multi-view consistency, and the single-view DMD loss ensures high-fidelity outputs. To further improve the 3D generation efficiency, we note that our multi-view generator outputs latent representations (and not pixels) for the multi-view images. We build on this insight to adapt a prior multi-view to 3D reconstruction approach [62] to instead consume multi-view latents as input, and show that this improves the reconstruction efficiency without any performance loss.", + "bbox": [ + 511, + 107, + 906, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our overall system thus combines an efficient few-step multi-view generator with a multi-view latent to 3D model to enable ultra fast 3D generation. We train our model on the subset of Objaverse dataset [3, 4], which contains about 400k instances with Cap3D text captions [25]. We show that our proposed Turbo3D is able to produce high-quality 3D assets in less than one second (see Fig. 1), and also achieves comparable quality with previous state-of-the-art.", + "bbox": [ + 511, + 513, + 908, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 648, + 656, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we discuss the closely related prior feedforward 3D generation methods. Optimization-based methods [18, 32] is out of this work's scope. We review feedforward 3D generative models in the following three categories: 1) methods that directly generate full 3D representation encoding geometry and appearance; 2) methods that generate shapes and then generate the textures; 3) methods that generate multi-view (MV) images followed by reconstruction. We also review the diffusion distillation literature which our method is built upon.", + "bbox": [ + 511, + 672, + 908, + 825 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Directly Genereating Full 3D Representation", + "text_level": 1, + "bbox": [ + 511, + 833, + 893, + 849 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several prior methods [1, 10, 30, 46, 48, 54, 58] have been introduced that directly generate 3D representations encoding both geometry and texture, e.g., implicit fields [10],", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23669", + "bbox": [ + 478, + 949, + 519, + 960 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "point clouds [30], and triplanes [54]. This line of work typically have to preprocess the source 3D data (meshes or multi-view images) into the target representations for generative models in a lossy fashion, hence limiting their quality and scalability.", + "bbox": [ + 89, + 90, + 480, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Alternative methods [1, 46, 48, 58] have been proposed to get rid of the lossy preprocessing. They implicitly bake a 3D generative process into a multi-view diffusion framework, creating a single-stage 3D generative model. However, trained only on 3D data, they suffer from the issue of limited generalization, lower pixel quality and reduced understanding of complex text prompts, compared with methods leveraging powerful pretrained text-to-image models [17, 21].", + "bbox": [ + 89, + 167, + 480, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method leverages the strong priors in a pretrained image generative model, but we go a step further by distilling the slow teacher model into a fast student generator for ultra-fast text-to-3D.", + "bbox": [ + 89, + 303, + 480, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Shape Generation + Texture Generation", + "text_level": 1, + "bbox": [ + 89, + 375, + 433, + 390 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent advancements in 3D content creation have introduced innovative methods [42, 57, 63] that focus on generating high-quality 3D shapes first, followed by generative texturing [2, 35], rather than creating both simultaneously. CLAY [63] offers a framework using a multiresolution Variational Autoencoder (VAE) and a latent Diffusion Transformer (DiT) to initially create detailed 3D geometries from inputs like text and images, before applying high-resolution physically-based rendering (PBR) textures. Direct3D [57] enhances scalability in image-to-3D generation by improving the performance of 3D VAE and DiT for generating 3D shapes, subsequently adding textures.", + "bbox": [ + 89, + 398, + 480, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, besides the challenge of learning a compact latent space suitable for generation, these methods are also usually slow as a result of the two diffusion models - one for shapes and the other for textures. In contrast, our Turbo3D only have one diffusion model for generating multi-views and is fast once distilled, while the reconstructor is deterministic and hence fast.", + "bbox": [ + 89, + 580, + 480, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. MV Generation + MV Reconstruction", + "text_level": 1, + "bbox": [ + 89, + 696, + 418, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A large body of the 3D generation work [17, 20, 24, 41] focus on leveraging multi-view generation and reconstruction to enhance quality and efficiency. Instant3D [17] offers a fast method for creating high-quality 3D assets from text prompts by combining sparse-view generation with a transformer-based reconstructor [8, 53, 56, 62], significantly reducing inference time. MVDream [41] employs a multi-view diffusion model to improve consistency and stability in 3D generation, integrating 2D and 3D data. SyncDreamer [24] enhances multiview-consistency from a single-view image using a 3D-aware feature attention mechanism. One-2-3-45 [20] introduces an efficient ap", + "bbox": [ + 89, + 719, + 480, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "proach to single image 3D reconstruction without extensive optimization, producing consistent 3D meshes. SV3D [51] utilizes a latent video diffusion model for novel multi-view synthesis, incorporating explicit camera control to improve 3D reconstruction quality.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We follow the same approach of reconstructing generated multi-views to create 3D assets in this work, but we focus on improving the efficiency of such systems. With our novel Dual-Teacher Distillation and Latent GS-LRM components, our Turbo3D manages to be at least an order of magnitude faster than these baselines while maintaining competitive quality. Concurrent work GECO [52] also use diffusion distillation to speed up image-to-3D. However, we differ in our dual-teacher distillation design with focus on text-to-3D. Our pipeline is also simpler and avoids the cumbersome mesh reconstructions for 3D distillation in GECO.", + "bbox": [ + 511, + 167, + 903, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Diffusion Distillation", + "text_level": 1, + "bbox": [ + 513, + 345, + 710, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent progress in diffusion models have focused on improving the efficiency of image generation by reducing the number of sampling steps required, leading to the development of several innovative distillation techniques [26, 27, 29, 38, 39, 45, 59-61]. Methods like Improved Rectified Flows [15] and InstaFlow [23] straighten the ODE trajectories, making them easier to approximate with a one-step student model. Consistency Models [45] train student generators to map any point on the teacher's ODE trajectory to a consistent target, enabling one-step and few-step image generation. Distribution Matching Distillation (DMD) [61] trains a one-step generator by minimizing the reverse KL divergence between the data distribution and the generator's output distribution [5, 26, 55]. Building on this, DMD2 [60] integrates GAN losses [6] and extends the method to multi-step generators, further enhancing generation quality.", + "bbox": [ + 511, + 368, + 903, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, most of existing approaches focus on distilling pretrained 2D image diffusion models. In our work, we adopt the popular DMD [61] approach and extend it to the multi-view domain. Distilling multi-view models presents unique challenges, such as increased mode collapse due to fine-tuning and distillation processes. To address this issue, we introduce a novel dual-teacher distillation technique, enabling swift, photorealistic multi-view generation.", + "bbox": [ + 511, + 609, + 903, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Background", + "text_level": 1, + "bbox": [ + 513, + 747, + 640, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Multi-view Diffusion Model", + "text_level": 1, + "bbox": [ + 511, + 772, + 761, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models generate data from noise by reversing a forward noising process. Given a data sample $x_0 \\sim p(x_0)$ from the data distribution, the forward process progressively adds Gaussian noise over $T$ timesteps. At timestep $t$ , the noised distribution conditioned on $x_0$ is given by:", + "bbox": [ + 511, + 796, + 903, + 871 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {t} \\mid x _ {0}\\right) = \\mathcal {N} \\left(x _ {t}; \\alpha_ {t} x _ {0}, \\beta_ {t} ^ {2} \\mathbf {I}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 883, + 903, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23670", + "bbox": [ + 478, + 949, + 517, + 960 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{t}$ and $\\beta_{t}$ are timestep-dependent constant that are specified by the noise schedule [12, 44]. The diffusion model $\\epsilon_{\\theta}$ is trained to reverse this noising process by learning to predict the added Gaussian noise $\\epsilon$ :", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL (\\theta) = \\mathbb {E} | | \\epsilon - \\epsilon_ {\\theta} (x _ {t}, t) | | ^ {2}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 157, + 482, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There exists other formulations, in which the diffusion model learns to predict the clean input $x_0$ directly [11], or a combination of $x_0$ and $\\epsilon$ [38]. Regardless of the specific prediction target, the outputs of these models can be related to the score function, which is the gradient of the log probability density of the data distribution [44]:", + "bbox": [ + 89, + 181, + 483, + 273 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {\\theta} \\left(x _ {t}, t\\right) = \\nabla_ {x _ {t}} \\log p \\left(x _ {t}\\right) = - \\frac {\\epsilon_ {\\theta} \\left(x _ {t} , t\\right)}{\\beta_ {t}} = - \\frac {x _ {t} - \\alpha_ {t} x _ {\\theta} \\left(x _ {t} , t\\right)}{\\beta_ {t} ^ {2}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 277, + 482, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\theta}$ [7] and $x_{\\theta}$ [11] represents noise and data prediction diffusion models, respectively. In our paper, we adopt the noise prediction scheme but our method generalizes to any formulations.", + "bbox": [ + 89, + 319, + 483, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In multi-view diffusion models, the generation of a 3D scene is conditioned on a text prompt, enabling the joint denoising of multiple views to produce a set of 3D-consistent output images [17, 41]. Specifically, given a text prompt $c$ and a set of multi-view images $\\{x^i\\}_{i=1}^K$ where $K$ is the number of view, the multi-view diffusion model learns to predict the added noise across all views simultaneously. The training objective is formulated as:", + "bbox": [ + 89, + 378, + 483, + 500 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} | | \\epsilon - \\epsilon_ {\\theta} \\left(\\left\\{x _ {t} ^ {i} \\right\\} _ {i = 1} ^ {K}, t, c\\right) | | ^ {2}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 505, + 482, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon$ represents independent Gaussian noise with the same variance applied to each view. At inference time, the generation starts from a set of fully noisy multi-view images $\\{x_{T}^{i}\\}_{i = 1}^{K}$ sampled from a standard Gaussian distribution. The multi-view diffusion model iteratively generate a sequence of cleaner multi-view images. Various diffusion samplers [7, 43] can be employed during this process to optimize generation speed and quality.", + "bbox": [ + 89, + 530, + 483, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Distribution Matching Distillation", + "text_level": 1, + "bbox": [ + 89, + 659, + 387, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Distribution Matching Distillation (DMD) is a widely used diffusion distillation technique that converts teacher diffusion models into a student generator requiring significantly fewer sampling steps [60, 61]. The DMD approach trains the student generator $G_{\\theta}$ by minimizing an approximate reverse KL divergence between the smoothed student's output distribution (denoted as $p_{\\mathrm{fake}}$ ) and the smoothed data distribution (denoted as $p_{\\mathrm{real}}$ ):", + "bbox": [ + 89, + 680, + 483, + 801 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {D M D}} (\\theta) = D _ {\\mathrm {K L}} \\left(p _ {\\text {f a k e}} \\| p _ {\\text {r e a l}}\\right) = \\mathbb {E} _ {x, t} \\left(\\log \\left(\\frac {p _ {\\text {f a k e}} \\left(x _ {t}\\right)}{p _ {\\text {r e a l}} \\left(x _ {t}\\right)}\\right)\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 806, + 482, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let the score functions of the data distribution and the student's output distribution be denoted as $s_{\\mathrm{real}}$ and $s_{\\mathrm{fake}}$ , respectively. The gradient of this KL divergence can be effectively approximated by the difference between these two", + "bbox": [ + 89, + 840, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "score functions:", + "bbox": [ + 513, + 92, + 624, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {\\theta} L _ {\\mathrm {D M D}} (\\theta) \\approx \\mathbb {E} \\left[ - \\int \\left(s _ {\\text {r e a l}} \\left(F \\left(G _ {\\theta} (\\epsilon), t\\right), t\\right) \\right. \\right. \\\\ \\left. - s _ {\\text {f a k e}} \\left(F \\left(G _ {\\theta} (\\epsilon), t\\right), t\\right)\\right) \\frac {d G _ {\\theta} (\\epsilon)}{d \\theta} d \\epsilon \\Bigg ] \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 109, + 903, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $F$ represents the forward diffusion process defined in Eq. 1. The student generator can be adapted for a multi-step generation setting by replacing the pure noise input $\\epsilon$ with a partially noisy image $x_{t}$ [60]. During training, the data distribution's score function is initialized from the teacher diffusion model and kept fixed, while the student's output distribution score function is dynamically trained using the student's output and a denoising loss (Eq. 4).", + "bbox": [ + 511, + 214, + 906, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 513, + 347, + 604, + 363 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we provide the technical details of our Turbo3D text-to-3D system, which features a highly efficient multi-view (MV) generator and reconstructor. We begin by describing our novel DualTeacher Distillation approach; it creates a rapid MV generator by jointly distilling knowledge from both a multi-view (MV) teacher and a single-view (SV) teacher. Following this, we discuss the latent-space GS-LRM that instantly lifts the generated MV latents into high-quality 3D Gaussians.", + "bbox": [ + 511, + 372, + 906, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Dual-teacher Distillation for MV Diffusion", + "text_level": 1, + "bbox": [ + 511, + 516, + 875, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Leveraging a multi-step multi-view diffusion in a text-to-3D generation process can be inefficient due to repeated evaluations of the diffusion denoiser in the sampling process. To speed this up, one approach is to use diffusion distillation [60, 61] to train a single-step or few-step MV generator.", + "bbox": [ + 511, + 537, + 905, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, we find that naively applying diffusion distillation methods to a MV teacher can cause the student model to generate overly simplistic and cartoonish appearance, which closely resembles the 3D Objverse dataset used during MV teacher finetuning and distillation [3] (shown in Fig. 6). We call this phenomenon 'compounding mode collapse'; this happens because both finetuning and distillation sacrifices generation diversity for efficiency. As the MV teacher is already biased towards synthetic Objverse-style appearance, further distilling it will have compound effect of locking the distilled generator in the mode of Objverse data that are far from the modes of photorealistic natural images.", + "bbox": [ + 511, + 614, + 905, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this issue, we propose to use dual teachers in the distillation process: one MV teacher to teach the student model about multi-view consistency, and one SV teacher to teach about each views' photo-realism. We illustrate this Dual-teacher Distillation algorithm in Fig. 2. Concretely, this is formulated as:", + "bbox": [ + 511, + 810, + 905, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23671", + "bbox": [ + 478, + 949, + 517, + 960 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b6c8af2b5b0d9295f570d9c4e21f14d471ce453b0f87792ba39f1cdcdc104285.jpg", + "image_caption": [ + "Figure 2. Dual-teacher distillation framework in our Turbo3D. Note that latents are visualized as RGB images for clarity. We aim to distill a multi-step multi-view teacher generator (right, green) into a few-step multi-view generator (left, blue). Our few-step MV student generator is conditioned on Plücker embeddings for better 3D awareness. Similar to [60], we optimize the student generator using distribution matching objective (DMD loss) and train the fake score function to model the distribution of samples produced by the student generator. In particular, we integrate two teacher models: multi-view teacher and single-view (SV) teacher to enhance both multi-view consistency and photorealism. The MV score functions take a set of images of one object as input and calculate the MV DMD loss, while the SV score functions treat each image separately and calculate the SV DMD loss." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 903, + 292 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {\\mathrm {D M D}} ^ {\\text {D u a l}} (\\theta) = D _ {\\mathrm {K L}} \\left(p _ {\\text {f a k e}} \\left(\\left\\{x _ {t} ^ {i} \\right\\} _ {i = 1} ^ {K}\\right) \\mid \\mid p _ {\\text {r e a l}} ^ {\\mathrm {M V}} \\left(\\left\\{x _ {t} ^ {i} \\right\\} _ {i = 1} ^ {K}\\right)\\right) \\\\ + \\lambda \\cdot \\frac {1}{K} \\sum_ {i = 1} ^ {K} D _ {\\mathrm {K L}} \\left(p _ {\\text {f a k e}} \\left(x _ {t} ^ {i}\\right) \\mid \\mid p _ {\\text {r e a l}} ^ {\\mathrm {S V}} \\left(x _ {t} ^ {i}\\right)\\right), \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 449, + 480, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p_{\\mathrm{real}}^{\\mathrm{MV}}(x_t), p_{\\mathrm{real}}^{\\mathrm{SV}}(x_t)$ represent MV and SV teachers, respectively; $\\lambda$ is the loss weight balancing the influence of MV and SV teachers on the distilled student model; $K = 4$ is the number of views. We set $\\lambda$ to 1 in our experiments. As demonstrated in Fig. 6, having the additional SV teacher in the distillation process effectively addresses the compound mode collapse issue, because it tries to pull each view to look like natural images.", + "bbox": [ + 89, + 521, + 483, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Latent GS-LRM for MV Reconstruction", + "text_level": 1, + "bbox": [ + 89, + 651, + 437, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To reconstruct 3D from the generated MV latents, one straightforward approach is to decode them into multi-view images, and then use the pixel-space GS-LRM [62] to produce the 3D Gaussians. However, such an approach can suffer from efficiency and memory issue when scaling to high resolution due to the poor performance of Conv2D operators in VAE decoder [13].", + "bbox": [ + 89, + 672, + 482, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose to skip the VAE decoding and directly input the generated MV latents to a latent GS-LRM for best efficiency. To train such a model, we supervise the reconstructed Gaussians with pixel-space novel-view rendering losses $(\\ell_{2}$ and perceptual losses as in [62]). We show that replacing pixel-space GS-LRM with a latent one does not affect the quality of generated assets in Tab. 3 and Fig. 3, while being faster.", + "bbox": [ + 89, + 779, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/770fab4cf3e96e8ae624c6d3039b8f00cd3b24ae52269edd1657ee56df65a012.jpg", + "image_caption": [ + "Figure 3. We compare the renderings of pixel GS-LRM and latent GS-LRM. Latent GS-LRM achieves comparable reconstruction quality as pixel GS-LRM." + ], + "image_footnote": [], + "bbox": [ + 514, + 422, + 906, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 638, + 645, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first describe our experimental setup (Section 5.1). We then compare our method with state-of-the-art text-to-3D baselines (Section 5.2). Finally, we ablate each component of our framework to showcase their effectiveness (Section 5.3).", + "bbox": [ + 511, + 666, + 906, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 755, + 702, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We use the Objaverse dataset [3] to train both our multi-view generation model and multi-view reconstruction model. We scale the objects and center them to fit into a cube $[-1, 1]^3$ . For the generation task, we render the dataset at a fixed elevation (20 degrees) and 16 equidistant azimuths to achieve a good coverage of the object. We render using a field of view $50^o$ at a distance of 2.7 and uniform lighting. For the reconstruction task, we render 32 views ran-", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23672", + "bbox": [ + 478, + 949, + 519, + 960 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "domly placed around the object with a random distance in the range of [1.5, 2.8]. We render a total of $730K$ objects.", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We adopt Instant3D [17] and LGM [47] as our baseline text-to-3D methods. However, since the field of fast text-to-3D is relatively underexplored, we also include recent state-of-the-art fast image-to-3D methods TripoSR [49] and SV3D [51] as baseline methods. For a fair comparison, we use a popular few-step text-to-image model Flux [14] to generate an input image first and then feed it to the image-to-3D models. The inference time of image-to-3D models is a summation of the two parts.", + "bbox": [ + 88, + 122, + 480, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We adopt the CLIP score [33] and VQA score [16] to assess the semantic alignment between the generated results and text prompts. We use 400 text prompts from DreamFusion [32] for evaluation. We generate one object for each prompt. For each generated 3D object, we render 10 random views and calculate the average CLIP score and VQA score between the rendered images and the input text. For inference time, we report it with all methods under the same image resolution of 256. Notably, some methods, e.g., Instant3D [17], only support a higher resolution of 512. For a fair comparison, we report their quantitative results under the resolution of 512 and only inference time on the resolution of 256.", + "bbox": [ + 91, + 258, + 483, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. The whole pipeline of our method includes three training phases. We first train a multi-step multi-view diffusion model on the Objverse dataset, by fine-tuning an internal DiT [31] based text-to-image model. We train this model for 30k iterations with 32 80G A100 GPUs using a total batch size of 128 and a learning rate of $3e^{-5}$ . Then we perform distillation to distill the multi-step multi-view generator into a few-step multi-view generator, which takes 10k iterations with 32 80G A100 GPUs with a global batch size of 128 and a learning rate of $5e^{-6}$ . Finally, we train a reconstruction model - latent GS-LRM - from scratch, which takes 80k iterations with 32 80G A100 GPUs with a total batch size of 256 and a learning rate of $4e^{-4}$ .", + "bbox": [ + 91, + 455, + 483, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Evaluation against baselines", + "text_level": 1, + "bbox": [ + 89, + 681, + 344, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparisons. As shown in Fig. 4, our method generates significantly better results compared with LGM [47] and Instant3D [17]. In particular, LGM tends to produce simple 3D assets without geometry and texture details; it also lacks robustness, oftentimes generating low-quality 3D assets that are broken, or suffer from multi-face Janus problem [32], or do not closely follow provided text prompts. Instant3D is much more robust and able to produce plausible 3D assets most of the time. However, its performance on text-image coherence is limited compared with our Turbo3D. For example, in the first example, LGM's rendition appears oversimplified with indistinct egg and yolk boundaries, while Instant3D provides a", + "bbox": [ + 89, + 703, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/41278a1e6349626d5e2eb0fa3061778fcabd7a0f9bd15cc8a82fab3c4f80429b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodClip Score ↑VQA Score ↑Inference Time ↓
TripoSR [49]23.850.571.19s
SV3D [51]24.920.6412.52s
Instant3D [17]26.230.6515.02s
LGM [47]24.730.586.56s
Turbo3D (Ours)27.610.760.35s
", + "bbox": [ + 517, + 89, + 903, + 220 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison against state-of-the-art 3D generation methods. Our Turbo3D generates 3D assets with highest CLIP and VQA scores while using the least amount of time (benchmarked on a A100 GPU).", + "bbox": [ + 511, + 234, + 903, + 290 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "closer match but lacks fine details in the egg's structure and wood texture and fails to capture the concept of 'spill out'. For complex prompts that describes a scene with multiple objects (4th and 6th rows), LGM generates a Janus asset while Instant3D miss a lot of concepts like blanket, basket, wildflowers etc; our generations adhere to the prompts much more closely. These qualitative results highlight our Turbo3D's ability to generate high-quality, text-aligned 3D models with a superior level of detail, realism, and coherence across diverse and complex prompts, outperforming both LGM and Instant3D by a large margin.", + "bbox": [ + 511, + 294, + 906, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Comparisons. Tab. 1 presents a quantitative comparison of our proposed method, Turbo3D, against several state-of-the-art approaches, including TripoSR [49], SV3D [51], Instant3D [17], and LGM [47], on the text-to-3D task. Our proposed Turbo3D achieves the highest CLIP Score of 27.61 and VQA Score of 0.76, outperforming other methods by a significant margin in both quality metrics. In addition to the quality improvement, Turbo3D demonstrates remarkable efficiency with an inference time of only 0.35 seconds, substantially faster than competing methods. Although TripoSR is able to generate a 3D asset in only 1.19s, the quality of the generated results is highly limited. As a result, our proposed Turbo3D is able to achieve outstanding performance in terms of both quality and inference speed compared with state-of-the-art methods.", + "bbox": [ + 511, + 462, + 908, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "User Study. We run a user study by randomly selecting 80 text prompts and asking 56 users to make 1120 pairwise comparisons (users are shown an input text prompt and two generated 3D assets from two anonymous methods). We show results in Fig. 5. Since the quantitative results of TripoSR is not competitive with others, it is not included in the user study. Moreover, SV3D only outputs videos instead of 3D representations of objects, so it is hard to perform a fair comparison. Therefore, we compare our Turbo3D with LGM [47], Instant3D [17], and our MV Teacher. In particular, MV Teacher is the model which our Turbo3D gets distilled from. When compared to LGM, Turbo3D achieved a win rate of $89.8\\%$ , with only $10.2\\%$ of participants favoring LGM. Against Instant3D, Turbo3D also outperformed", + "bbox": [ + 511, + 688, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23673", + "bbox": [ + 478, + 949, + 517, + 960 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dd7c273064bb74e2de53156680c2486bd0fb7e7fb3a1e995c1bc9a9af9c02ea5.jpg", + "image_caption": [ + "A cracked egg with the yolk spilling out on a wooden table" + ], + "image_footnote": [], + "bbox": [ + 258, + 95, + 436, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/79b84fdfa282a3c47faf90b10148611bd13ba19afe4f359b98645d9cd136ad01.jpg", + "image_caption": [ + "An origami hippo in a river" + ], + "image_footnote": [], + "bbox": [ + 259, + 165, + 439, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/17f9a0dbb3f38c3f6bdbe582d1879bfa5bf825a056ddaf299cb146be85340ca9.jpg", + "image_caption": [ + "Wedding dress made of tentacles" + ], + "image_footnote": [], + "bbox": [ + 259, + 241, + 439, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/731bf3ab30d05f460e110f61a89898a6040dee9fb6052404f82de3b7a08ee963.jpg", + "image_caption": [ + "A cozy campfire surrounded by logs for sitting, with a small tent set up nearby. Add a pot hanging over the fire and a few scattered leaves" + ], + "image_footnote": [], + "bbox": [ + 259, + 327, + 444, + 371 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9d73cf11344764951ce8ad6a64202543f797d544ff427b3cd079dba0474f92c8.jpg", + "image_caption": [ + "A chimpanzee looking through a telescope" + ], + "image_footnote": [], + "bbox": [ + 267, + 388, + 444, + 450 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3f1aa7a34d3157c18aa86fcb26e704a7cae3452bb701fb0face8ccad5618fb9e.jpg", + "image_caption": [ + "Picnic scene with a checkered blanket on grass, a wicker basket with sandwiches and fruit, a bottle of lemonade with glasses. An open book, a small bouquet of wildflowers on the blanket", + "LGM (6.56s)" + ], + "image_footnote": [], + "bbox": [ + 267, + 468, + 447, + 508 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/25a0fcdb3fabe4b14e256af2acc145b72a37f4b62e89e4fb5c02601a318a8152.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 452, + 93, + 648, + 161 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/75b9e66296042d17b04eca594b43ad1de34f2cad7a287dbe510221d059b6a977.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 459, + 167, + 663, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aa328635e59287c006911e2f51f0b6db3a1f9d5fdc3f5f84b016084a56157494.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 452, + 251, + 658, + 308 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/425ab4eb59c40b5d37b3eddc27f7399d2a662ed62d4b7fd0fdac0b8ca4783fa3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 457, + 321, + 650, + 373 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a4c4fcda4d11b1a8c6b0c68a886ca6a56ffeac786a2eb6c5bad0653224b77f51.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 392, + 651, + 450 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/130f858e1a4c7f5c41446305cdade90df368422fcc29c2672615cdc12d33de35.jpg", + "image_caption": [ + "Instant3D (15.02s)" + ], + "image_footnote": [], + "bbox": [ + 467, + 463, + 668, + 508 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/446a4a76926606f4e79759f91f976df15c955004eee456a5e17e9af1ba99c31c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 90, + 864, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fad229eb454b2d7b8c139f5736523156725d6735512b45b5548679213e3a50f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 167, + 867, + 227 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f1fe5b925e4af1be6e702545b3be5ca8b3b90b2f9f4f49cc9eefa5a5ccaee05e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 238, + 754, + 304 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0dfc3dcbe4e34172619ae622997cabc2216f5986c10284e30352acf1302a4d0d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 238, + 864, + 304 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b730d47708a28172c7cda6492f28cc1e8d5d473b2eea05b461b9358785307723.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 315, + 767, + 382 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fb856587482a4dcb9a3a4f0f154b0045081a146f5d9335315f4b2ab15c95eac8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 315, + 866, + 382 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dab9d381ead6a661e1962505b65d99c070d4002c1c004bf6c0d9401d9319c5bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 390, + 754, + 452 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f8b6f3b3370d3818c0b6370d45e7503073c96445286fb28d3aa9194afe89e030.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 797, + 388, + 862, + 450 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/02c8665de1cd69726625f24da476647e80db1de7c11175dca078927a8d2188a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 453, + 772, + 513 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c381af4e1ec5aa84b323ff3df4909d0ebbbe3ace19496407f3ef1f8caeb98a60.jpg", + "image_caption": [ + "Turbo3D (0.35s)" + ], + "image_footnote": [], + "bbox": [ + 781, + 453, + 883, + 516 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c696e2855ab09dbcd9814aec49a813d3a2d579cb39f29b015a8062884d6d6ef2.jpg", + "image_caption": [ + "Figure 4. Comparison of our Turbo3D against baselines LGM [47] and Instant3D [17]. Among these methods, Our method generates the most detailed and physically plausible 3D assets, closely adhering to the provided text prompts. In contrast, LGM tends to generate broken assets with Janus issue [32], while Instant3D has poorer text alignment, oftentimes missing some concepts, e.g., 'spilling out' in the first row, 'river' in the second row, etc.", + "Figure 5. User study results comparing our Turbo3D to baseline LGM [47], Instant3D [17], and our slow MV teacher. Our Turbo3D is consistently preferred over baseline LGM and Instant3D, while having on-par preference with our MV teacher. See Fig. 4,6 for visual comparison." + ], + "image_footnote": [], + "bbox": [ + 94, + 633, + 478, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with a $74.9\\%$ win rate, indicating that users consistently found Turbo3D's outputs more aligned with the input text descriptions and more visually compelling. When evaluated against our teacher model, Turbo3D held a close win rate of $50.6\\%$ , with $49.4\\%$ preferring MV T., reflecting compara", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ble quality between the student and teacher models. The result indicates that our Turbo3D not only achieves significant speedup with distillation, but also preserves the generation ability from the teacher model.", + "bbox": [ + 511, + 633, + 906, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 709, + 666, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Single-view Teacher. In Tab. 2, we demonstrate the effectiveness of the dual teacher distillation strategy. The first line is our multi-view (MV) teacher model, which achieves impressive results but runs slowly due to the many diffusion sampling steps required. When performing distillation only with this MV Teacher, the quality drops by a large margin as shown in the second row. When adding a single-view teacher model for distillation, the distilled model achieves much better results compared with the previous one. This configuration approaches the performance of the Multi-step MV Model while maintaining", + "bbox": [ + 511, + 734, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23674", + "bbox": [ + 478, + 949, + 517, + 960 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/14c11aab51a2b79b3b14707f6134539692ef778fc0bc731541c042c1a9f400da.jpg", + "image_caption": [ + "a baby dragon hatching out of a stone egg" + ], + "image_footnote": [], + "bbox": [ + 227, + 90, + 302, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b8a75eb4345c7411a2cf574b66ba1fb290d9237c1f8cf05f8cb0dce7d3a4ccb1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 333, + 90, + 408, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b4a0e75f9e2d85803a89cf93b419fb492d279819a7a51a6888e7e5e42bb71266.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 90, + 514, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/35e5ca8550afdc03748a2a0e9ca5eb6ac891282671ee7e42a3a4b561e55b46e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 544, + 90, + 624, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/221515456723992d465e999b011480f321f742f645c379b9381e8560c46c7969.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 655, + 95, + 736, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/87998b6538d655386135afc1be9221cd7f36f6ae4ae860210a9c5676857b3324.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 95, + 844, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5351ae9224cfdf8588d2fdb2cf06ba23c523833290ab104683d10d652ea79188.jpg", + "image_caption": [ + "a kitten \nstanding on top \nof a giant \ntortoise" + ], + "image_footnote": [], + "bbox": [ + 233, + 179, + 292, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4173eedb7dd2598e8584a115254ad1fe1e880c543257495dddccbbe6dd28edb4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 179, + 408, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6da348535d4671af31fa25b97a7e31cbe88326dc82c62094386658935ba50fc0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 449, + 179, + 509, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a326bc282810c2a4fbb0bb9b37e16a49e55517b373e8ef6419aff1fb6091c3d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 179, + 625, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/76ee5af2d72470e2510d1cff3ba3252db76b0f0956dcb23be217a629713ef14c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 179, + 732, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f546f40024d31519de3334ac10cf67992d2227877687a309f140222f17a80ada.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 179, + 839, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ceaa9e9801aa331355267dec446402868bc2d274aeae513bdfe22a97baa2f32c.jpg", + "image_caption": [ + "a wide angle DSLR photo of a squirrel in samurai armor wielding a katana" + ], + "image_footnote": [], + "bbox": [ + 222, + 268, + 305, + 345 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/084eb55434eec498e2f4ee1aee37e7fecbea33d68cabfcd9bd036c857a6b3c04.jpg", + "image_caption": [ + "MV teacher (10.18s)", + "Figure 6. Ablation of our Dual-teacher distillation algorithm. Naively distilling MV teacher (middle column) causes compound mode collapse (see Sec. 4.1), producing overly smooth synthetic-looking assets. Our dual-teacher distillation (right column) fixes the issue and generates 3D assets that are as photorealistic as, if not more than, the baseline MV teacher (left column). We also include the inference timings for each method; the distilled model is $\\sim 50\\mathrm{x}$ faster than the teacher model." + ], + "image_footnote": [], + "bbox": [ + 325, + 268, + 410, + 348 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cade8740465662074ba3ac19e8e4387a4709a3b6230231096e61f39b3730f896.jpg", + "image_caption": [ + "MV teacher only distillation (0.35s)" + ], + "image_footnote": [], + "bbox": [ + 424, + 266, + 514, + 348 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d638913a9d51230b9845e32ea9c8f108c978159f5ed354372e520e91edfe6b54.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 545, + 265, + 640, + 330 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/79eb8e06658290f62d81bb0f6fd1dc01c89e28c7b488173e89f41021b4b1d192.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 265, + 736, + 347 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0853bc60ea13790ae35eb4300b8725763e8fa5c6a4ff96e8b4f0e79c97f71970.jpg", + "image_caption": [ + "Dual teacher distillation (0.35s)" + ], + "image_footnote": [], + "bbox": [ + 754, + 265, + 859, + 348 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cc30cd6a29de9e8a23b45954878568565fd9e9fbe9607a36a090426fbb3d34b1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AblationCLIP Score ↑VQA Score ↑
Multi-step MV Model28.040.77
Few-step Model (MV Teacher)26.600.69
Few-step Model (Dual Teacher)27.610.76
", + "bbox": [ + 94, + 455, + 480, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the efficiency benefits of the few-step setup, showcasing the advantages of the dual-teacher strategy in our distillation. We also provide visualizations of the three models in Fig. 6. These comparisons show that the dual teacher distillation model strikes a balance between detail retention and stylization, closely replicating the MV teacher model's quality while benefiting from the efficiency gains of distillation.", + "bbox": [ + 89, + 646, + 482, + 751 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Latent GS-LRM. In Tab. 3, we showcase the effectiveness of latent GS-LRM. Compared with the original GS-LRM which operates in pixel space, our latent GS-LRM is able to skip the expensive image decoding process while achieving similar image quality.", + "bbox": [ + 89, + 752, + 482, + 829 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 844, + 209, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose Turbo3D for ultra-fast text-to-3D generation. To enable fast multi-view generation, we pro", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4689361cdb19fc0e9d186e3563272f395eefe6954231812ce5c60b2917b9efdc.jpg", + "table_caption": [ + "Table 2. Ablation of dual teacher distillation. Distillation leads to quality drop compared with the MV teacher model. Compared with naively distilling the single MV teacher, dual-teacher distillation leads to much smaller quality drop. See Fig. 6 for visual comparison." + ], + "table_footnote": [], + "table_body": "
AblationCLIP Score ↑VQA Score ↑Inference Time ↓
Pixel GS-LRM [62]27.620.760.45s
Latent GS-LRM27.610.760.35s
", + "bbox": [ + 517, + 455, + 903, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation for latent GS-LRM. We report the CLIP score, VQA score, and overall text-to-3D inference time for comparison. Our latent GS-LRM achieves similar image quality while enabling better efficiency ( $\\sim$ 22% speedup).", + "bbox": [ + 511, + 551, + 906, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pose to distill a multi-step multi-view generator into a few-step multi-view generator. Moreover, to restore the multiview consistency and photo-realism during distillation, we introduce a novel dual-teacher distillation framework. To further improve the multi-view reconstruction efficiency, we propose a latent GS-LRM which directly reconstructs 3D Gaussians from multi-view latents. Extensive experiments demonstrate that our proposed Turbo3D is able to achieve outstanding performance in terms of both generation quality and inference efficiency.", + "bbox": [ + 511, + 632, + 906, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 513, + 814, + 679, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work began during Hanzhe Hu and Tianwei Yin's internships at Adobe Research. This research was supported in part by NSF award IIS-2345610.", + "bbox": [ + 511, + 844, + 905, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23675", + "bbox": [ + 478, + 949, + 517, + 960 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. Renderdiffusion: Image diffusion for 3d reconstruction, inpainting and generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12608-12618, 2023. 2, 3", + "[2] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18558-18568, 2023. 3", + "[3] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 4, 5, 1", + "[4] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\\mathrm{m} + 3\\mathrm{d}$ objects. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[5] Jean-Yves Franceschi, Mike Gartrell, Ludovic Dos Santos, Thibaut Issenhuth, Emmanuel de Bezenac, Mickaël Chen, and Alain Rakotomamonjy. Unifying gans and score-based diffusion as generative particle models. In NeurIPS, 2023. 3", + "[6] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 3", + "[7] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 4", + "[8] Yicong Hong, Kai Zhang, Jiumiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 3", + "[9] Hanzhe Hu, Zhizhuo Zhou, Varun Jampani, and Shubham Tulsiani. Mvd-fusion: Single-view 3d via depth-consistent multi-view generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9698-9707, 2024. 2", + "[10] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2", + "[11] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 4", + "[12] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. In NeurIPS, 2021. 4", + "[13] MIT HAN Lab. Patch conv: Patch convolution to avoid largegpu memory usage. https://hanlab.mit.edu/blog/patch-conv, 2024. Accessed: 2024-11-14. 5", + "[14] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024. Accessed: 2024-11-14. 6" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Sangyun Lee, Zinan Lin, and Giulia Fanti. Improving the training of rectified flows. arXiv preprint arXiv:2405.20320, 2024. 3", + "[16] Baiqi Li, Zhiqiu Lin, Wenxuan Peng, Jean de Dieu Nyandwi, Daniel Jiang, Zixian Ma, Simran Khanuja, Ranjay Krishna, Graham Neubig, and Deva Ramanan. Naturalbench: Evaluating vision-language models on natural adversarial samples. arXiv preprint arXiv:2410.14669, 2024. 6", + "[17] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214, 2023. 2, 3, 4, 6, 7, 1", + "[18] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 2", + "[19] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 2", + "[20] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[21] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023. 3", + "[22] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 2", + "[23] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In The Twelfth International Conference on Learning Representations, 2023. 2, 3", + "[24] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 2, 3", + "[25] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[26] Weijian Luo, Tianyang Hu, Shifeng Zhang, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang. Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models. In NeurIPS, 2023. 3", + "[27] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In CVPR, 2023. 3", + "[28] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. AutoSDF: Shape priors for 3d completion, reconstruction and generation. In CVPR, 2022. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23676", + "bbox": [ + 478, + 949, + 519, + 960 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Thuan Hoang Nguyen and Anh Tran. Swiftbrush: One-step text-to-image diffusion model with variational score distillation. In CVPR, 2024. 3", + "[30] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2, 3", + "[31] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 6, 1", + "[32] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 6, 7", + "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6", + "[34] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 2", + "[35] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAPH 2023 conference proceedings, pages 1–11, 2023. 3", + "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2", + "[37] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2", + "[38] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 2, 3, 4", + "[39] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. arXiv preprint arXiv:2403.12015, 2024. 2, 3", + "[40] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 2", + "[41] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 3, 4", + "[42] Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "meshes with decoder-only transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19615-19625, 2024. 3", + "[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 4", + "[44] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2, 4", + "[45] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 2, 3", + "[46] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Viewset diffusion:(0-) image-conditioned 3d generative models from 2d data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8863-8873, 2023. 2, 3", + "[47] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1-18. Springer, 2025. 6, 7, 1", + "[48] Ayush Tewari, Tianwei Yin, George Cazenavette, Semon Rezchikov, Josh Tenenbaum, Frédo Durand, Bill Freeman, and Vincent Sitzmann. Diffusion with forward models: Solving stochastic inverse problems without direct supervision. Advances in Neural Information Processing Systems, 36: 12349-12362, 2023. 2, 3", + "[49] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151, 2024. 6, 1", + "[50] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural computation, 23(7):1661-1674, 2011. 2", + "[51] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitrii Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. SV3D: Novel multi-view synthesis and 3D generation from a single image using latent video diffusion. In European Conference on Computer Vision (ECCV), 2024. 3, 6, 1", + "[52] Chen Wang, Jiatao Gu, Xiaoxiao Long, Yuan Liu, and Lingjie Liu. Geco: Generative image-to-3d within a second. arXiv preprint arXiv:2405.20327, 2024. 3", + "[53] Peng Wang, Hao Tan, Sai Bi, Yinghao Xu, Fujun Luan, Kalyan Sunkavalli, Wenping Wang, Zexiang Xu, and Kai Zhang. Pf-lrm: Pose-free large reconstruction model for joint pose and shape prediction. arXiv preprint arXiv:2311.12024, 2023. 3", + "[54] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4563-4573, 2023. 2, 3", + "[55] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23677", + "bbox": [ + 478, + 949, + 517, + 960 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "diverse text-to-3d generation with variational score distillation. In NeurIPS, 2023. 2, 3", + "[56] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Meshlrm: Large reconstruction model for high-quality mesh. arXiv preprint arXiv:2404.12385, 2024. 3", + "[57] Shuang Wu, Youtian Lin, Feihu Zhang, Yifei Zeng, Jingxi Xu, Philip Torr, Xun Cao, and Yao Yao. Direct3d: Scalable image-to-3d generation via 3d latent diffusion transformer. arXiv preprint arXiv:2405.14832, 2024. 2, 3", + "[58] Yinghao Xu, Hao Tan, Fujun Luan, Sai Bi, Peng Wang, Jiahao Li, Zifan Shi, Kalyan Sunkavalli, Gordon Wetzstein, Zexiang Xu, et al. Dmv3d: Denoising multi-view diffusion using 3d large reconstruction model. arXiv preprint arXiv:2311.09217, 2023. 2, 3", + "[59] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image generation via diffusion gans. In CVPR, 2024. 3", + "[60] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. arXiv preprint arXiv:2405.14867, 2024. 2, 3, 4, 5", + "[61] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6613-6623, 2024. 2, 3, 4", + "[62] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 2, 3, 5, 8, 1", + "[63] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 2, 3" + ], + "bbox": [ + 91, + 90, + 482, + 628 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23678", + "bbox": [ + 478, + 949, + 517, + 960 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Turbo3D_ Ultra-fast Text-to-3D Generation/bd4523a6-3f35-4dc6-9b48-90da3f36ebea_model.json b/2025/Turbo3D_ Ultra-fast Text-to-3D Generation/bd4523a6-3f35-4dc6-9b48-90da3f36ebea_model.json new file mode 100644 index 0000000000000000000000000000000000000000..86f8baf53fc966ed9c1a5f392b311515c179cf79 --- /dev/null +++ b/2025/Turbo3D_ Ultra-fast Text-to-3D Generation/bd4523a6-3f35-4dc6-9b48-90da3f36ebea_model.json @@ -0,0 +1,3049 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.28, + 0.131, + 0.72, + 0.151 + ], + "angle": 0, + "content": "Turbo3D: Ultra-fast Text-to-3D Generation" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.18, + 0.214, + 0.198 + ], + "angle": 0, + "content": "Hanzhe Hu" + }, + { + "type": "text", + "bbox": [ + 0.254, + 0.181, + 0.361, + 0.197 + ], + "angle": 0, + "content": "Tianwei Yin²" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.199, + 0.339, + 0.215 + ], + "angle": 0, + "content": "Sai Bi3" + }, + { + "type": "text", + "bbox": [ + 0.4, + 0.181, + 0.501, + 0.199 + ], + "angle": 0, + "content": "Fujun Luan3" + }, + { + "type": "text", + "bbox": [ + 0.54, + 0.181, + 0.624, + 0.198 + ], + "angle": 0, + "content": "Yiwei Hu3" + }, + { + "type": "text", + "bbox": [ + 0.405, + 0.2, + 0.565, + 0.216 + ], + "angle": 0, + "content": "Shubham Tulsiani\\*" + }, + { + "type": "text", + "bbox": [ + 0.664, + 0.181, + 0.74, + 0.198 + ], + "angle": 0, + "content": "Hao Tan3" + }, + { + "type": "text", + "bbox": [ + 0.778, + 0.181, + 0.882, + 0.199 + ], + "angle": 0, + "content": "Zexiang \\(\\mathbf{X}\\mathbf{u}^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.625, + 0.199, + 0.724, + 0.217 + ], + "angle": 0, + "content": "Kai Zhang\\(^{3*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.233, + 0.363, + 0.252 + ], + "angle": 0, + "content": "\\(^{1}\\) Carnegie Mellon University" + }, + { + "type": "text", + "bbox": [ + 0.382, + 0.233, + 0.699, + 0.253 + ], + "angle": 0, + "content": "2 Massachusetts Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.72, + 0.233, + 0.865, + 0.251 + ], + "angle": 0, + "content": "\\(^{3}\\) Adobe Research" + }, + { + "type": "text", + "bbox": [ + 0.375, + 0.273, + 0.616, + 0.287 + ], + "angle": 0, + "content": "https://turbo-3d.github.io/" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.317, + 0.861, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.464, + 0.891, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.543, + 0.37, + 0.569 + ], + "angle": 0, + "content": "a panda wearing a chef's hat and kneading bread dough on a countertop" + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.575, + 0.174, + 0.638 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.16, + 0.646, + 0.307, + 0.659 + ], + "angle": 0, + "content": "a plush toy of a cordi nurse" + }, + { + "type": "image", + "bbox": [ + 0.206, + 0.575, + 0.266, + 0.638 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.576, + 0.36, + 0.638 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.572, + 0.444, + 0.638 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.572, + 0.539, + 0.64 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.555, + 0.572, + 0.627, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.419, + 0.647, + 0.547, + 0.659 + ], + "angle": 0, + "content": "a raccoon stealing a pie" + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.57, + 0.711, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.566, + 0.807, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.827, + 0.566, + 0.889, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.641, + 0.875, + 0.666 + ], + "angle": 0, + "content": "a badger wearing a party hat and blowing out birthday candles on a cake" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.675, + 0.174, + 0.737 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.187, + 0.677, + 0.272, + 0.738 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.163, + 0.749, + 0.317, + 0.762 + ], + "angle": 0, + "content": "a cake in the shape of a train" + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.675, + 0.367, + 0.737 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.668, + 0.455, + 0.733 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.457, + 0.668, + 0.552, + 0.735 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.676, + 0.639, + 0.734 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.678, + 0.728, + 0.737 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.732, + 0.679, + 0.818, + 0.737 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.821, + 0.679, + 0.904, + 0.737 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.372, + 0.739, + 0.65, + 0.788 + ], + "angle": 0, + "content": "Campsite with a tent on grassy ground, a backpack against it, a campfire surrounded by stones. A sleeping bag partially unrolled next to the tent, a lantern hanging from a tree branch" + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.751, + 0.852, + 0.765 + ], + "angle": 0, + "content": "a lion reading the newspaper" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.802, + 0.907, + 0.859 + ], + "angle": 0, + "content": "Figure 1. Overview of our Turbo3D text-to-3D system. Turbo3D generates high-quality 3D Gaussian Splatting (3DGS) assets from user prompts in less than 1 second on a single A100 GPU. It's a two-stage pipeline consisting of a highly efficient latent-space few-step multi-view (MV) generator and single-step MV reconstructor. Note that we visualize latents as RGB images and 3DGS assets as point clouds in the pipeline figure for clarity." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.898, + 0.2, + 0.911 + ], + "angle": 0, + "content": "*Equal advising." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.951, + 0.519, + 0.962 + ], + "angle": 0, + "content": "23668" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.249, + 0.091, + 0.327, + 0.108 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.486, + 0.351 + ], + "angle": 0, + "content": "We present Turbo3D, an ultra-fast text-to-3D system capable of generating high-quality Gaussian splatting assets in under one second. Turbo3D employs a rapid 4-step, 4-view diffusion generator and an efficient feed-forward Gaussian reconstructor, both operating in latent space, as shown in Fig. 1. The 4-step, 4-view generator is a student model distilled through a novel Dual-Teacher approach, which encourages the student to learn view consistency from a multiview teacher and photo-realism from a single-view teacher. By shifting the Gaussian reconstructor's inputs from pixel space to latent space, we eliminate the extra image decoding time and halve the transformer sequence length for maximum efficiency. Our method demonstrates superior 3D generation results compared to previous baselines, while operating in a fraction of their runtime." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.377, + 0.223, + 0.394 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.485, + 0.599 + ], + "angle": 0, + "content": "The recent advances in image generative models allow users to generate detailed outputs from just a text prompt. While initial denoising diffusion-based methods [7, 34, 36, 37, 44, 50] enabled impressive photo-realistic generation, recent techniques [19, 22, 23, 39, 45, 60, 61] have significantly improved inference efficiency of such models, allowing high-fidelity generation in the blink of an eye. Unfortunately, these advances in generative 2D modeling methods have not yet been matched in the 3D domain, where ultra-fast realistic 3D generation remains a challenge. In this work, we seek to bridge this gap, and present Turbo3D, a text-to-3D generative model that can synthesize detailed 3D outputs in a fraction of a second." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.486, + 0.902 + ], + "angle": 0, + "content": "The existing approaches for text-based 3D inference can be categorized as either generative [10, 17, 24, 30, 57, 63] or optimization-driven [18, 32, 55]. The latter class of methods optimize 3D representations by 'distilling' pretrained 2D diffusion models [32]. While this approach can yield decent 3D outputs, it is highly inefficient, typically requiring several minutes or even hours to output a single 3D representation. The alternative paradigm is to learn a generative model that directly outputs 3D representations. While initial methods [10, 28] investigated representations such as point clouds and SDFs, the limited availability of 3D data restricted the generation quality. Recent approaches [9, 17, 24, 40, 41] have instead advocated for learning generative models of multi-view images (followed by deterministic 3D reconstruction), as these can be initialized from pre-trained 2D generative models. While these methods have resulted in impressive generations, the multiview finetuning on synthetic data does inhibit their generation quality. More crucially, their inference efficiency is restricted by the iterative denoising process required for the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.779, + 0.107 + ], + "angle": 0, + "content": "text-conditioned multi-view generation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.108, + 0.907, + 0.514 + ], + "angle": 0, + "content": "In this work, we follow the paradigm of text-to-3D via multi-view generation, and aim to improve the efficiency of the underlying components to enable ultra-fast generation, while enhancing the fidelity of the generated outputs. Inspired by the recent progress in reducing inference time by distilling 2D diffusion models into one-step or few-step generators [38, 39, 45, 61], we adapt these for multi-view 3D generation. We first train a many-step text-to-multiview diffusion model and then distill it into a much faster 4-step generator using the distribution matching distillation (DMD) loss [61]. However, we find that this process leads to a significant quality degradation, as it fails to capture the full range of modes present in the multi-view teacher model. To overcome this, we propose to extend the DMD pipeline to incorporate another single-view teacher - a 2D denoising diffusion model trained on a large set of high quality aesthetic images. Our few-step multi-view generator is thus trained with a dual-teacher distillation approach, where the multi-view DMD loss helps our model learn multi-view consistency, and the single-view DMD loss ensures high-fidelity outputs. To further improve the 3D generation efficiency, we note that our multi-view generator outputs latent representations (and not pixels) for the multi-view images. We build on this insight to adapt a prior multi-view to 3D reconstruction approach [62] to instead consume multi-view latents as input, and show that this improves the reconstruction efficiency without any performance loss." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.515, + 0.909, + 0.637 + ], + "angle": 0, + "content": "Our overall system thus combines an efficient few-step multi-view generator with a multi-view latent to 3D model to enable ultra fast 3D generation. We train our model on the subset of Objaverse dataset [3, 4], which contains about 400k instances with Cap3D text captions [25]. We show that our proposed Turbo3D is able to produce high-quality 3D assets in less than one second (see Fig. 1), and also achieves comparable quality with previous state-of-the-art." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.649, + 0.657, + 0.665 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.674, + 0.909, + 0.826 + ], + "angle": 0, + "content": "In this section, we discuss the closely related prior feedforward 3D generation methods. Optimization-based methods [18, 32] is out of this work's scope. We review feedforward 3D generative models in the following three categories: 1) methods that directly generate full 3D representation encoding geometry and appearance; 2) methods that generate shapes and then generate the textures; 3) methods that generate multi-view (MV) images followed by reconstruction. We also review the diffusion distillation literature which our method is built upon." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.895, + 0.851 + ], + "angle": 0, + "content": "2.1. Directly Genereating Full 3D Representation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Several prior methods [1, 10, 30, 46, 48, 54, 58] have been introduced that directly generate 3D representations encoding both geometry and texture, e.g., implicit fields [10]," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.95, + 0.521, + 0.962 + ], + "angle": 0, + "content": "23669" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.166 + ], + "angle": 0, + "content": "point clouds [30], and triplanes [54]. This line of work typically have to preprocess the source 3D data (meshes or multi-view images) into the target representations for generative models in a lossy fashion, hence limiting their quality and scalability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.482, + 0.303 + ], + "angle": 0, + "content": "Alternative methods [1, 46, 48, 58] have been proposed to get rid of the lossy preprocessing. They implicitly bake a 3D generative process into a multi-view diffusion framework, creating a single-stage 3D generative model. However, trained only on 3D data, they suffer from the issue of limited generalization, lower pixel quality and reduced understanding of complex text prompts, compared with methods leveraging powerful pretrained text-to-image models [17, 21]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.304, + 0.482, + 0.364 + ], + "angle": 0, + "content": "Our method leverages the strong priors in a pretrained image generative model, but we go a step further by distilling the slow teacher model into a fast student generator for ultra-fast text-to-3D." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.376, + 0.434, + 0.391 + ], + "angle": 0, + "content": "2.2. Shape Generation + Texture Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.399, + 0.482, + 0.58 + ], + "angle": 0, + "content": "Recent advancements in 3D content creation have introduced innovative methods [42, 57, 63] that focus on generating high-quality 3D shapes first, followed by generative texturing [2, 35], rather than creating both simultaneously. CLAY [63] offers a framework using a multiresolution Variational Autoencoder (VAE) and a latent Diffusion Transformer (DiT) to initially create detailed 3D geometries from inputs like text and images, before applying high-resolution physically-based rendering (PBR) textures. Direct3D [57] enhances scalability in image-to-3D generation by improving the performance of 3D VAE and DiT for generating 3D shapes, subsequently adding textures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.482, + 0.685 + ], + "angle": 0, + "content": "However, besides the challenge of learning a compact latent space suitable for generation, these methods are also usually slow as a result of the two diffusion models - one for shapes and the other for textures. In contrast, our Turbo3D only have one diffusion model for generating multi-views and is fast once distilled, while the reconstructor is deterministic and hence fast." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.697, + 0.419, + 0.712 + ], + "angle": 0, + "content": "2.3. MV Generation + MV Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.482, + 0.901 + ], + "angle": 0, + "content": "A large body of the 3D generation work [17, 20, 24, 41] focus on leveraging multi-view generation and reconstruction to enhance quality and efficiency. Instant3D [17] offers a fast method for creating high-quality 3D assets from text prompts by combining sparse-view generation with a transformer-based reconstructor [8, 53, 56, 62], significantly reducing inference time. MVDream [41] employs a multi-view diffusion model to improve consistency and stability in 3D generation, integrating 2D and 3D data. SyncDreamer [24] enhances multiview-consistency from a single-view image using a 3D-aware feature attention mechanism. One-2-3-45 [20] introduces an efficient ap" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "proach to single image 3D reconstruction without extensive optimization, producing consistent 3D meshes. SV3D [51] utilizes a latent video diffusion model for novel multi-view synthesis, incorporating explicit camera control to improve 3D reconstruction quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.169, + 0.905, + 0.334 + ], + "angle": 0, + "content": "We follow the same approach of reconstructing generated multi-views to create 3D assets in this work, but we focus on improving the efficiency of such systems. With our novel Dual-Teacher Distillation and Latent GS-LRM components, our Turbo3D manages to be at least an order of magnitude faster than these baselines while maintaining competitive quality. Concurrent work GECO [52] also use diffusion distillation to speed up image-to-3D. However, we differ in our dual-teacher distillation design with focus on text-to-3D. Our pipeline is also simpler and avoids the cumbersome mesh reconstructions for 3D distillation in GECO." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.346, + 0.712, + 0.361 + ], + "angle": 0, + "content": "2.4. Diffusion Distillation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.369, + 0.905, + 0.61 + ], + "angle": 0, + "content": "Recent progress in diffusion models have focused on improving the efficiency of image generation by reducing the number of sampling steps required, leading to the development of several innovative distillation techniques [26, 27, 29, 38, 39, 45, 59-61]. Methods like Improved Rectified Flows [15] and InstaFlow [23] straighten the ODE trajectories, making them easier to approximate with a one-step student model. Consistency Models [45] train student generators to map any point on the teacher's ODE trajectory to a consistent target, enabling one-step and few-step image generation. Distribution Matching Distillation (DMD) [61] trains a one-step generator by minimizing the reverse KL divergence between the data distribution and the generator's output distribution [5, 26, 55]. Building on this, DMD2 [60] integrates GAN losses [6] and extends the method to multi-step generators, further enhancing generation quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.905, + 0.732 + ], + "angle": 0, + "content": "However, most of existing approaches focus on distilling pretrained 2D image diffusion models. In our work, we adopt the popular DMD [61] approach and extend it to the multi-view domain. Distilling multi-view models presents unique challenges, such as increased mode collapse due to fine-tuning and distillation processes. To address this issue, we introduce a novel dual-teacher distillation technique, enabling swift, photorealistic multi-view generation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.748, + 0.642, + 0.765 + ], + "angle": 0, + "content": "3. Background" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.763, + 0.788 + ], + "angle": 0, + "content": "3.1. Multi-view Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.797, + 0.905, + 0.872 + ], + "angle": 0, + "content": "Diffusion models generate data from noise by reversing a forward noising process. Given a data sample \\( x_0 \\sim p(x_0) \\) from the data distribution, the forward process progressively adds Gaussian noise over \\( T \\) timesteps. At timestep \\( t \\), the noised distribution conditioned on \\( x_0 \\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.885, + 0.905, + 0.902 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {t} \\mid x _ {0}\\right) = \\mathcal {N} \\left(x _ {t}; \\alpha_ {t} x _ {0}, \\beta_ {t} ^ {2} \\mathbf {I}\\right), \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.95, + 0.519, + 0.962 + ], + "angle": 0, + "content": "23670" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "where \\(\\alpha_{t}\\) and \\(\\beta_{t}\\) are timestep-dependent constant that are specified by the noise schedule [12, 44]. The diffusion model \\(\\epsilon_{\\theta}\\) is trained to reverse this noising process by learning to predict the added Gaussian noise \\(\\epsilon\\):" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.158, + 0.483, + 0.177 + ], + "angle": 0, + "content": "\\[\nL (\\theta) = \\mathbb {E} | | \\epsilon - \\epsilon_ {\\theta} (x _ {t}, t) | | ^ {2}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.484, + 0.274 + ], + "angle": 0, + "content": "There exists other formulations, in which the diffusion model learns to predict the clean input \\( x_0 \\) directly [11], or a combination of \\( x_0 \\) and \\( \\epsilon \\) [38]. Regardless of the specific prediction target, the outputs of these models can be related to the score function, which is the gradient of the log probability density of the data distribution [44]:" + }, + { + "type": "equation", + "bbox": [ + 0.096, + 0.278, + 0.483, + 0.319 + ], + "angle": 0, + "content": "\\[\ns _ {\\theta} \\left(x _ {t}, t\\right) = \\nabla_ {x _ {t}} \\log p \\left(x _ {t}\\right) = - \\frac {\\epsilon_ {\\theta} \\left(x _ {t} , t\\right)}{\\beta_ {t}} = - \\frac {x _ {t} - \\alpha_ {t} x _ {\\theta} \\left(x _ {t} , t\\right)}{\\beta_ {t} ^ {2}} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.32, + 0.484, + 0.379 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\theta}\\) [7] and \\(x_{\\theta}\\) [11] represents noise and data prediction diffusion models, respectively. In our paper, we adopt the noise prediction scheme but our method generalizes to any formulations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.484, + 0.501 + ], + "angle": 0, + "content": "In multi-view diffusion models, the generation of a 3D scene is conditioned on a text prompt, enabling the joint denoising of multiple views to produce a set of 3D-consistent output images [17, 41]. Specifically, given a text prompt \\( c \\) and a set of multi-view images \\( \\{x^i\\}_{i=1}^K \\) where \\( K \\) is the number of view, the multi-view diffusion model learns to predict the added noise across all views simultaneously. The training objective is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.506, + 0.483, + 0.525 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} | | \\epsilon - \\epsilon_ {\\theta} \\left(\\left\\{x _ {t} ^ {i} \\right\\} _ {i = 1} ^ {K}, t, c\\right) | | ^ {2}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.531, + 0.484, + 0.652 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) represents independent Gaussian noise with the same variance applied to each view. At inference time, the generation starts from a set of fully noisy multi-view images \\(\\{x_{T}^{i}\\}_{i = 1}^{K}\\) sampled from a standard Gaussian distribution. The multi-view diffusion model iteratively generate a sequence of cleaner multi-view images. Various diffusion samplers [7, 43] can be employed during this process to optimize generation speed and quality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.66, + 0.388, + 0.675 + ], + "angle": 0, + "content": "3.2. Distribution Matching Distillation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.681, + 0.484, + 0.803 + ], + "angle": 0, + "content": "Distribution Matching Distillation (DMD) is a widely used diffusion distillation technique that converts teacher diffusion models into a student generator requiring significantly fewer sampling steps [60, 61]. The DMD approach trains the student generator \\( G_{\\theta} \\) by minimizing an approximate reverse KL divergence between the smoothed student's output distribution (denoted as \\( p_{\\mathrm{fake}} \\)) and the smoothed data distribution (denoted as \\( p_{\\mathrm{real}} \\)):" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.807, + 0.483, + 0.838 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {D M D}} (\\theta) = D _ {\\mathrm {K L}} \\left(p _ {\\text {f a k e}} \\| p _ {\\text {r e a l}}\\right) = \\mathbb {E} _ {x, t} \\left(\\log \\left(\\frac {p _ {\\text {f a k e}} \\left(x _ {t}\\right)}{p _ {\\text {r e a l}} \\left(x _ {t}\\right)}\\right)\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Let the score functions of the data distribution and the student's output distribution be denoted as \\( s_{\\mathrm{real}} \\) and \\( s_{\\mathrm{fake}} \\), respectively. The gradient of this KL divergence can be effectively approximated by the difference between these two" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.093, + 0.625, + 0.106 + ], + "angle": 0, + "content": "score functions:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.11, + 0.905, + 0.208 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {\\theta} L _ {\\mathrm {D M D}} (\\theta) \\approx \\mathbb {E} \\left[ - \\int \\left(s _ {\\text {r e a l}} \\left(F \\left(G _ {\\theta} (\\epsilon), t\\right), t\\right) \\right. \\right. \\\\ \\left. - s _ {\\text {f a k e}} \\left(F \\left(G _ {\\theta} (\\epsilon), t\\right), t\\right)\\right) \\frac {d G _ {\\theta} (\\epsilon)}{d \\theta} d \\epsilon \\Bigg ] \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.215, + 0.907, + 0.337 + ], + "angle": 0, + "content": "where \\(F\\) represents the forward diffusion process defined in Eq. 1. The student generator can be adapted for a multi-step generation setting by replacing the pure noise input \\(\\epsilon\\) with a partially noisy image \\(x_{t}\\) [60]. During training, the data distribution's score function is initialized from the teacher diffusion model and kept fixed, while the student's output distribution score function is dynamically trained using the student's output and a denoising loss (Eq. 4)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.348, + 0.605, + 0.364 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.373, + 0.907, + 0.51 + ], + "angle": 0, + "content": "In this section, we provide the technical details of our Turbo3D text-to-3D system, which features a highly efficient multi-view (MV) generator and reconstructor. We begin by describing our novel DualTeacher Distillation approach; it creates a rapid MV generator by jointly distilling knowledge from both a multi-view (MV) teacher and a single-view (SV) teacher. Following this, we discuss the latent-space GS-LRM that instantly lifts the generated MV latents into high-quality 3D Gaussians." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.517, + 0.876, + 0.532 + ], + "angle": 0, + "content": "4.1. Dual-teacher Distillation for MV Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.906, + 0.614 + ], + "angle": 0, + "content": "Leveraging a multi-step multi-view diffusion in a text-to-3D generation process can be inefficient due to repeated evaluations of the diffusion denoiser in the sampling process. To speed this up, one approach is to use diffusion distillation [60, 61] to train a single-step or few-step MV generator." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.615, + 0.906, + 0.81 + ], + "angle": 0, + "content": "However, we find that naively applying diffusion distillation methods to a MV teacher can cause the student model to generate overly simplistic and cartoonish appearance, which closely resembles the 3D Objverse dataset used during MV teacher finetuning and distillation [3] (shown in Fig. 6). We call this phenomenon 'compounding mode collapse'; this happens because both finetuning and distillation sacrifices generation diversity for efficiency. As the MV teacher is already biased towards synthetic Objverse-style appearance, further distilling it will have compound effect of locking the distilled generator in the mode of Objverse data that are far from the modes of photorealistic natural images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.9 + ], + "angle": 0, + "content": "To address this issue, we propose to use dual teachers in the distillation process: one MV teacher to teach the student model about multi-view consistency, and one SV teacher to teach about each views' photo-realism. We illustrate this Dual-teacher Distillation algorithm in Fig. 2. Concretely, this is formulated as:" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.95, + 0.518, + 0.962 + ], + "angle": 0, + "content": "23671" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.905, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.304, + 0.907, + 0.404 + ], + "angle": 0, + "content": "Figure 2. Dual-teacher distillation framework in our Turbo3D. Note that latents are visualized as RGB images for clarity. We aim to distill a multi-step multi-view teacher generator (right, green) into a few-step multi-view generator (left, blue). Our few-step MV student generator is conditioned on Plücker embeddings for better 3D awareness. Similar to [60], we optimize the student generator using distribution matching objective (DMD loss) and train the fake score function to model the distribution of samples produced by the student generator. In particular, we integrate two teacher models: multi-view teacher and single-view (SV) teacher to enhance both multi-view consistency and photorealism. The MV score functions take a set of images of one object as input and calculate the MV DMD loss, while the SV score functions treat each image separately and calculate the SV DMD loss." + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.45, + 0.482, + 0.514 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {\\mathrm {D M D}} ^ {\\text {D u a l}} (\\theta) = D _ {\\mathrm {K L}} \\left(p _ {\\text {f a k e}} \\left(\\left\\{x _ {t} ^ {i} \\right\\} _ {i = 1} ^ {K}\\right) \\mid \\mid p _ {\\text {r e a l}} ^ {\\mathrm {M V}} \\left(\\left\\{x _ {t} ^ {i} \\right\\} _ {i = 1} ^ {K}\\right)\\right) \\\\ + \\lambda \\cdot \\frac {1}{K} \\sum_ {i = 1} ^ {K} D _ {\\mathrm {K L}} \\left(p _ {\\text {f a k e}} \\left(x _ {t} ^ {i}\\right) \\mid \\mid p _ {\\text {r e a l}} ^ {\\mathrm {S V}} \\left(x _ {t} ^ {i}\\right)\\right), \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.484, + 0.645 + ], + "angle": 0, + "content": "where \\( p_{\\mathrm{real}}^{\\mathrm{MV}}(x_t), p_{\\mathrm{real}}^{\\mathrm{SV}}(x_t) \\) represent MV and SV teachers, respectively; \\( \\lambda \\) is the loss weight balancing the influence of MV and SV teachers on the distilled student model; \\( K = 4 \\) is the number of views. We set \\( \\lambda \\) to 1 in our experiments. As demonstrated in Fig. 6, having the additional SV teacher in the distillation process effectively addresses the compound mode collapse issue, because it tries to pull each view to look like natural images." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.652, + 0.439, + 0.669 + ], + "angle": 0, + "content": "4.2. Latent GS-LRM for MV Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.78 + ], + "angle": 0, + "content": "To reconstruct 3D from the generated MV latents, one straightforward approach is to decode them into multi-view images, and then use the pixel-space GS-LRM [62] to produce the 3D Gaussians. However, such an approach can suffer from efficiency and memory issue when scaling to high resolution due to the poor performance of Conv2D operators in VAE decoder [13]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We propose to skip the VAE decoding and directly input the generated MV latents to a latent GS-LRM for best efficiency. To train such a model, we supervise the reconstructed Gaussians with pixel-space novel-view rendering losses \\((\\ell_{2}\\) and perceptual losses as in [62]). We show that replacing pixel-space GS-LRM with a latent one does not affect the quality of generated assets in Tab. 3 and Fig. 3, while being faster." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.424, + 0.907, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.612 + ], + "angle": 0, + "content": "Figure 3. We compare the renderings of pixel GS-LRM and latent GS-LRM. Latent GS-LRM achieves comparable reconstruction quality as pixel GS-LRM." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.639, + 0.646, + 0.657 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.907, + 0.743 + ], + "angle": 0, + "content": "In this section, we first describe our experimental setup (Section 5.1). We then compare our method with state-of-the-art text-to-3D baselines (Section 5.2). Finally, we ablate each component of our framework to showcase their effectiveness (Section 5.3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.756, + 0.704, + 0.773 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Datasets. We use the Objaverse dataset [3] to train both our multi-view generation model and multi-view reconstruction model. We scale the objects and center them to fit into a cube \\([-1, 1]^3\\). For the generation task, we render the dataset at a fixed elevation (20 degrees) and 16 equidistant azimuths to achieve a good coverage of the object. We render using a field of view \\(50^o\\) at a distance of 2.7 and uniform lighting. For the reconstruction task, we render 32 views ran-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.95, + 0.521, + 0.962 + ], + "angle": 0, + "content": "23672" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "domly placed around the object with a random distance in the range of [1.5, 2.8]. We render a total of \\(730K\\) objects." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.123, + 0.482, + 0.258 + ], + "angle": 0, + "content": "Baselines. We adopt Instant3D [17] and LGM [47] as our baseline text-to-3D methods. However, since the field of fast text-to-3D is relatively underexplored, we also include recent state-of-the-art fast image-to-3D methods TripoSR [49] and SV3D [51] as baseline methods. For a fair comparison, we use a popular few-step text-to-image model Flux [14] to generate an input image first and then feed it to the image-to-3D models. The inference time of image-to-3D models is a summation of the two parts." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.26, + 0.485, + 0.456 + ], + "angle": 0, + "content": "Metrics. We adopt the CLIP score [33] and VQA score [16] to assess the semantic alignment between the generated results and text prompts. We use 400 text prompts from DreamFusion [32] for evaluation. We generate one object for each prompt. For each generated 3D object, we render 10 random views and calculate the average CLIP score and VQA score between the rendered images and the input text. For inference time, we report it with all methods under the same image resolution of 256. Notably, some methods, e.g., Instant3D [17], only support a higher resolution of 512. For a fair comparison, we report their quantitative results under the resolution of 512 and only inference time on the resolution of 256." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.457, + 0.484, + 0.669 + ], + "angle": 0, + "content": "Implementation Details. The whole pipeline of our method includes three training phases. We first train a multi-step multi-view diffusion model on the Objverse dataset, by fine-tuning an internal DiT [31] based text-to-image model. We train this model for 30k iterations with 32 80G A100 GPUs using a total batch size of 128 and a learning rate of \\(3e^{-5}\\). Then we perform distillation to distill the multi-step multi-view generator into a few-step multi-view generator, which takes 10k iterations with 32 80G A100 GPUs with a global batch size of 128 and a learning rate of \\(5e^{-6}\\). Finally, we train a reconstruction model - latent GS-LRM - from scratch, which takes 80k iterations with 32 80G A100 GPUs with a total batch size of 256 and a learning rate of \\(4e^{-4}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.345, + 0.697 + ], + "angle": 0, + "content": "5.2. Evaluation against baselines" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Qualitative Comparisons. As shown in Fig. 4, our method generates significantly better results compared with LGM [47] and Instant3D [17]. In particular, LGM tends to produce simple 3D assets without geometry and texture details; it also lacks robustness, oftentimes generating low-quality 3D assets that are broken, or suffer from multi-face Janus problem [32], or do not closely follow provided text prompts. Instant3D is much more robust and able to produce plausible 3D assets most of the time. However, its performance on text-image coherence is limited compared with our Turbo3D. For example, in the first example, LGM's rendition appears oversimplified with indistinct egg and yolk boundaries, while Instant3D provides a" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.09, + 0.905, + 0.221 + ], + "angle": 0, + "content": "
MethodClip Score ↑VQA Score ↑Inference Time ↓
TripoSR [49]23.850.571.19s
SV3D [51]24.920.6412.52s
Instant3D [17]26.230.6515.02s
LGM [47]24.730.586.56s
Turbo3D (Ours)27.610.760.35s
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.236, + 0.905, + 0.291 + ], + "angle": 0, + "content": "Table 1. Comparison against state-of-the-art 3D generation methods. Our Turbo3D generates 3D assets with highest CLIP and VQA scores while using the least amount of time (benchmarked on a A100 GPU)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.295, + 0.907, + 0.462 + ], + "angle": 0, + "content": "closer match but lacks fine details in the egg's structure and wood texture and fails to capture the concept of 'spill out'. For complex prompts that describes a scene with multiple objects (4th and 6th rows), LGM generates a Janus asset while Instant3D miss a lot of concepts like blanket, basket, wildflowers etc; our generations adhere to the prompts much more closely. These qualitative results highlight our Turbo3D's ability to generate high-quality, text-aligned 3D models with a superior level of detail, realism, and coherence across diverse and complex prompts, outperforming both LGM and Instant3D by a large margin." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.909, + 0.688 + ], + "angle": 0, + "content": "Quantitative Comparisons. Tab. 1 presents a quantitative comparison of our proposed method, Turbo3D, against several state-of-the-art approaches, including TripoSR [49], SV3D [51], Instant3D [17], and LGM [47], on the text-to-3D task. Our proposed Turbo3D achieves the highest CLIP Score of 27.61 and VQA Score of 0.76, outperforming other methods by a significant margin in both quality metrics. In addition to the quality improvement, Turbo3D demonstrates remarkable efficiency with an inference time of only 0.35 seconds, substantially faster than competing methods. Although TripoSR is able to generate a 3D asset in only 1.19s, the quality of the generated results is highly limited. As a result, our proposed Turbo3D is able to achieve outstanding performance in terms of both quality and inference speed compared with state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.901 + ], + "angle": 0, + "content": "User Study. We run a user study by randomly selecting 80 text prompts and asking 56 users to make 1120 pairwise comparisons (users are shown an input text prompt and two generated 3D assets from two anonymous methods). We show results in Fig. 5. Since the quantitative results of TripoSR is not competitive with others, it is not included in the user study. Moreover, SV3D only outputs videos instead of 3D representations of objects, so it is hard to perform a fair comparison. Therefore, we compare our Turbo3D with LGM [47], Instant3D [17], and our MV Teacher. In particular, MV Teacher is the model which our Turbo3D gets distilled from. When compared to LGM, Turbo3D achieved a win rate of \\(89.8\\%\\), with only \\(10.2\\%\\) of participants favoring LGM. Against Instant3D, Turbo3D also outperformed" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.95, + 0.518, + 0.962 + ], + "angle": 0, + "content": "23673" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.109, + 0.253, + 0.141 + ], + "angle": 0, + "content": "A cracked egg with the yolk spilling out on a wooden table" + }, + { + "type": "image_caption", + "bbox": [ + 0.118, + 0.18, + 0.225, + 0.201 + ], + "angle": 0, + "content": "An origami hippo in a river" + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.266, + 0.231, + 0.287 + ], + "angle": 0, + "content": "Wedding dress made of tentacles" + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.32, + 0.238, + 0.384 + ], + "angle": 0, + "content": "A cozy campfire surrounded by logs for sitting, with a small tent set up nearby. Add a pot hanging over the fire and a few scattered leaves" + }, + { + "type": "image_caption", + "bbox": [ + 0.118, + 0.403, + 0.227, + 0.425 + ], + "angle": 0, + "content": "A chimpanzee looking through a telescope" + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.45, + 0.25, + 0.536 + ], + "angle": 0, + "content": "Picnic scene with a checkered blanket on grass, a wicker basket with sandwiches and fruit, a bottle of lemonade with glasses. An open book, a small bouquet of wildflowers on the blanket" + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.096, + 0.437, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.166, + 0.44, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.242, + 0.441, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.328, + 0.446, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.39, + 0.446, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.469, + 0.449, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.314, + 0.526, + 0.398, + 0.54 + ], + "angle": 0, + "content": "LGM (6.56s)" + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.094, + 0.65, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.168, + 0.665, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.252, + 0.659, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.322, + 0.651, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.482, + 0.393, + 0.653, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.464, + 0.669, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.526, + 0.635, + 0.54 + ], + "angle": 0, + "content": "Instant3D (15.02s)" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.091, + 0.865, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.169, + 0.869, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.681, + 0.239, + 0.756, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.239, + 0.866, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.667, + 0.316, + 0.768, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.316, + 0.867, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.391, + 0.755, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.39, + 0.863, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.454, + 0.774, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.782, + 0.454, + 0.885, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.526, + 0.825, + 0.54 + ], + "angle": 0, + "content": "Turbo3D (0.35s)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.551, + 0.907, + 0.607 + ], + "angle": 0, + "content": "Figure 4. Comparison of our Turbo3D against baselines LGM [47] and Instant3D [17]. Among these methods, Our method generates the most detailed and physically plausible 3D assets, closely adhering to the provided text prompts. In contrast, LGM tends to generate broken assets with Janus issue [32], while Instant3D has poorer text alignment, oftentimes missing some concepts, e.g., 'spilling out' in the first row, 'river' in the second row, etc." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.634, + 0.48, + 0.74 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.752, + 0.483, + 0.822 + ], + "angle": 0, + "content": "Figure 5. User study results comparing our Turbo3D to baseline LGM [47], Instant3D [17], and our slow MV teacher. Our Turbo3D is consistently preferred over baseline LGM and Instant3D, while having on-par preference with our MV teacher. See Fig. 4,6 for visual comparison." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "with a \\(74.9\\%\\) win rate, indicating that users consistently found Turbo3D's outputs more aligned with the input text descriptions and more visually compelling. When evaluated against our teacher model, Turbo3D held a close win rate of \\(50.6\\%\\), with \\(49.4\\%\\) preferring MV T., reflecting compara" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.907, + 0.695 + ], + "angle": 0, + "content": "ble quality between the student and teacher models. The result indicates that our Turbo3D not only achieves significant speedup with distillation, but also preserves the generation ability from the teacher model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.71, + 0.667, + 0.727 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Effect of Single-view Teacher. In Tab. 2, we demonstrate the effectiveness of the dual teacher distillation strategy. The first line is our multi-view (MV) teacher model, which achieves impressive results but runs slowly due to the many diffusion sampling steps required. When performing distillation only with this MV Teacher, the quality drops by a large margin as shown in the second row. When adding a single-view teacher model for distillation, the distilled model achieves much better results compared with the previous one. This configuration approaches the performance of the Multi-step MV Model while maintaining" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.95, + 0.519, + 0.962 + ], + "angle": 0, + "content": "23674" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.106, + 0.223, + 0.147 + ], + "angle": 0, + "content": "a baby dragon hatching out of a stone egg" + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.092, + 0.303, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.092, + 0.41, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.091, + 0.516, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.091, + 0.625, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.656, + 0.096, + 0.738, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.097, + 0.846, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.189, + 0.23, + 0.241 + ], + "angle": 0, + "content": "a kitten \nstanding on top \nof a giant \ntortoise" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.18, + 0.294, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.18, + 0.41, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.45, + 0.18, + 0.51, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.18, + 0.626, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.663, + 0.18, + 0.733, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.18, + 0.841, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.264, + 0.226, + 0.342 + ], + "angle": 0, + "content": "a wide angle DSLR photo of a squirrel in samurai armor wielding a katana" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.269, + 0.307, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.269, + 0.411, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.267, + 0.516, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.266, + 0.641, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.266, + 0.737, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.266, + 0.861, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.364, + 0.375, + 0.377 + ], + "angle": 0, + "content": "MV teacher (10.18s)" + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.364, + 0.629, + 0.378 + ], + "angle": 0, + "content": "MV teacher only distillation (0.35s)" + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.364, + 0.847, + 0.378 + ], + "angle": 0, + "content": "Dual teacher distillation (0.35s)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.389, + 0.907, + 0.445 + ], + "angle": 0, + "content": "Figure 6. Ablation of our Dual-teacher distillation algorithm. Naively distilling MV teacher (middle column) causes compound mode collapse (see Sec. 4.1), producing overly smooth synthetic-looking assets. Our dual-teacher distillation (right column) fixes the issue and generates 3D assets that are as photorealistic as, if not more than, the baseline MV teacher (left column). We also include the inference timings for each method; the distilled model is \\(\\sim 50\\mathrm{x}\\) faster than the teacher model." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.456, + 0.481, + 0.558 + ], + "angle": 0, + "content": "
AblationCLIP Score ↑VQA Score ↑
Multi-step MV Model28.040.77
Few-step Model (MV Teacher)26.600.69
Few-step Model (Dual Teacher)27.610.76
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.574, + 0.483, + 0.645 + ], + "angle": 0, + "content": "Table 2. Ablation of dual teacher distillation. Distillation leads to quality drop compared with the MV teacher model. Compared with naively distilling the single MV teacher, dual-teacher distillation leads to much smaller quality drop. See Fig. 6 for visual comparison." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.647, + 0.483, + 0.752 + ], + "angle": 0, + "content": "the efficiency benefits of the few-step setup, showcasing the advantages of the dual-teacher strategy in our distillation. We also provide visualizations of the three models in Fig. 6. These comparisons show that the dual teacher distillation model strikes a balance between detail retention and stylization, closely replicating the MV teacher model's quality while benefiting from the efficiency gains of distillation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.753, + 0.483, + 0.83 + ], + "angle": 0, + "content": "Effect of Latent GS-LRM. In Tab. 3, we showcase the effectiveness of latent GS-LRM. Compared with the original GS-LRM which operates in pixel space, our latent GS-LRM is able to skip the expensive image decoding process while achieving similar image quality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.845, + 0.21, + 0.86 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "In this work, we propose Turbo3D for ultra-fast text-to-3D generation. To enable fast multi-view generation, we pro" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.456, + 0.905, + 0.537 + ], + "angle": 0, + "content": "
AblationCLIP Score ↑VQA Score ↑Inference Time ↓
Pixel GS-LRM [62]27.620.760.45s
Latent GS-LRM27.610.760.35s
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.552, + 0.907, + 0.608 + ], + "angle": 0, + "content": "Table 3. Ablation for latent GS-LRM. We report the CLIP score, VQA score, and overall text-to-3D inference time for comparison. Our latent GS-LRM achieves similar image quality while enabling better efficiency (\\(\\sim\\)22% speedup)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.633, + 0.907, + 0.785 + ], + "angle": 0, + "content": "pose to distill a multi-step multi-view generator into a few-step multi-view generator. Moreover, to restore the multiview consistency and photo-realism during distillation, we introduce a novel dual-teacher distillation framework. To further improve the multi-view reconstruction efficiency, we propose a latent GS-LRM which directly reconstructs 3D Gaussians from multi-view latents. Extensive experiments demonstrate that our proposed Turbo3D is able to achieve outstanding performance in terms of both generation quality and inference efficiency." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.815, + 0.681, + 0.831 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.845, + 0.906, + 0.888 + ], + "angle": 0, + "content": "This work began during Hanzhe Hu and Tianwei Yin's internships at Adobe Research. This research was supported in part by NSF award IIS-2345610." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.95, + 0.518, + 0.962 + ], + "angle": 0, + "content": "23675" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.198 + ], + "angle": 0, + "content": "[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. Renderdiffusion: Image diffusion for 3d reconstruction, inpainting and generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12608-12618, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.2, + 0.484, + 0.269 + ], + "angle": 0, + "content": "[2] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18558-18568, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.27, + 0.484, + 0.352 + ], + "angle": 0, + "content": "[3] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 4, 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.483, + 0.422 + ], + "angle": 0, + "content": "[4] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of \\(10\\mathrm{m} + 3\\mathrm{d}\\) objects. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.423, + 0.483, + 0.478 + ], + "angle": 0, + "content": "[5] Jean-Yves Franceschi, Mike Gartrell, Ludovic Dos Santos, Thibaut Issenhuth, Emmanuel de Bezenac, Mickaël Chen, and Alain Rakotomamonjy. Unifying gans and score-based diffusion as generative particle models. In NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.48, + 0.483, + 0.535 + ], + "angle": 0, + "content": "[6] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.536, + 0.483, + 0.577 + ], + "angle": 0, + "content": "[7] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.483, + 0.633 + ], + "angle": 0, + "content": "[8] Yicong Hong, Kai Zhang, Jiumiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[9] Hanzhe Hu, Zhizhuo Zhou, Varun Jampani, and Shubham Tulsiani. Mvd-fusion: Single-view 3d via depth-consistent multi-view generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9698-9707, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.704, + 0.483, + 0.745 + ], + "angle": 0, + "content": "[10] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[11] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[12] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. In NeurIPS, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[13] MIT HAN Lab. Patch conv: Patch convolution to avoid largegpu memory usage. https://hanlab.mit.edu/blog/patch-conv, 2024. Accessed: 2024-11-14. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[14] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024. Accessed: 2024-11-14. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.133 + ], + "angle": 0, + "content": "[15] Sangyun Lee, Zinan Lin, and Giulia Fanti. Improving the training of rectified flows. arXiv preprint arXiv:2405.20320, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[16] Baiqi Li, Zhiqiu Lin, Wenxuan Peng, Jean de Dieu Nyandwi, Daniel Jiang, Zixian Ma, Simran Khanuja, Ranjay Krishna, Graham Neubig, and Deva Ramanan. Naturalbench: Evaluating vision-language models on natural adversarial samples. arXiv preprint arXiv:2410.14669, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.906, + 0.274 + ], + "angle": 0, + "content": "[17] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214, 2023. 2, 3, 4, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.906, + 0.356 + ], + "angle": 0, + "content": "[18] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.906, + 0.399 + ], + "angle": 0, + "content": "[19] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.401, + 0.906, + 0.467 + ], + "angle": 0, + "content": "[20] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.908, + 0.538 + ], + "angle": 0, + "content": "[21] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.906, + 0.58 + ], + "angle": 0, + "content": "[22] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.906, + 0.636 + ], + "angle": 0, + "content": "[23] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In The Twelfth International Conference on Learning Representations, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.906, + 0.692 + ], + "angle": 0, + "content": "[24] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.693, + 0.906, + 0.746 + ], + "angle": 0, + "content": "[25] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.748, + 0.906, + 0.802 + ], + "angle": 0, + "content": "[26] Weijian Luo, Tianyang Hu, Shifeng Zhang, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang. Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models. In NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.906, + 0.858 + ], + "angle": 0, + "content": "[27] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[28] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. AutoSDF: Shape priors for 3d completion, reconstruction and generation. In CVPR, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.95, + 0.52, + 0.962 + ], + "angle": 0, + "content": "23676" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[29] Thuan Hoang Nguyen and Anh Tran. Swiftbrush: One-step text-to-image diffusion model with variational score distillation. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[30] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[31] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.291 + ], + "angle": 0, + "content": "[32] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.433 + ], + "angle": 0, + "content": "[34] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[35] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAPH 2023 conference proceedings, pages 1–11, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[37] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[38] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[39] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. arXiv preprint arXiv:2403.12015, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[40] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[41] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[42] Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.135 + ], + "angle": 0, + "content": "meshes with decoder-only transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19615-19625, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[44] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.25 + ], + "angle": 0, + "content": "[45] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.252, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[46] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Viewset diffusion:(0-) image-conditioned 3d generative models from 2d data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8863-8873, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[47] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1-18. Springer, 2025. 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[48] Ayush Tewari, Tianwei Yin, George Cazenavette, Semon Rezchikov, Josh Tenenbaum, Frédo Durand, Bill Freeman, and Vincent Sitzmann. Diffusion with forward models: Solving stochastic inverse problems without direct supervision. Advances in Neural Information Processing Systems, 36: 12349-12362, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[49] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151, 2024. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[50] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural computation, 23(7):1661-1674, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[51] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitrii Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. SV3D: Novel multi-view synthesis and 3D generation from a single image using latent video diffusion. In European Conference on Computer Vision (ECCV), 2024. 3, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[52] Chen Wang, Jiatao Gu, Xiaoxiao Long, Yuan Liu, and Lingjie Liu. Geco: Generative image-to-3d within a second. arXiv preprint arXiv:2405.20327, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[53] Peng Wang, Hao Tan, Sai Bi, Yinghao Xu, Fujun Luan, Kalyan Sunkavalli, Wenping Wang, Zexiang Xu, and Kai Zhang. Pf-lrm: Pose-free large reconstruction model for joint pose and shape prediction. arXiv preprint arXiv:2311.12024, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[54] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4563-4573, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[55] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.95, + 0.519, + 0.962 + ], + "angle": 0, + "content": "23677" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "diverse text-to-3d generation with variational score distillation. In NeurIPS, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[56] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Meshlrm: Large reconstruction model for high-quality mesh. arXiv preprint arXiv:2404.12385, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[57] Shuang Wu, Youtian Lin, Feihu Zhang, Yifei Zeng, Jingxi Xu, Philip Torr, Xun Cao, and Yao Yao. Direct3d: Scalable image-to-3d generation via 3d latent diffusion transformer. arXiv preprint arXiv:2405.14832, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.483, + 0.303 + ], + "angle": 0, + "content": "[58] Yinghao Xu, Hao Tan, Fujun Luan, Sai Bi, Peng Wang, Jiahao Li, Zifan Shi, Kalyan Sunkavalli, Gordon Wetzstein, Zexiang Xu, et al. Dmv3d: Denoising multi-view diffusion using 3d large reconstruction model. arXiv preprint arXiv:2311.09217, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[59] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image generation via diffusion gans. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[60] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. arXiv preprint arXiv:2405.14867, 2024. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.485 + ], + "angle": 0, + "content": "[61] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6613-6623, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.489, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[62] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 2, 3, 5, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.56, + 0.482, + 0.629 + ], + "angle": 0, + "content": "[63] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.629 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.95, + 0.519, + 0.962 + ], + "angle": 0, + "content": "23678" + } + ] +] \ No newline at end of file diff --git a/2025/Turbo3D_ Ultra-fast Text-to-3D Generation/bd4523a6-3f35-4dc6-9b48-90da3f36ebea_origin.pdf b/2025/Turbo3D_ Ultra-fast Text-to-3D Generation/bd4523a6-3f35-4dc6-9b48-90da3f36ebea_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..79552f3d3c4a6cede920789857b464ccd46868e3 --- /dev/null +++ b/2025/Turbo3D_ Ultra-fast Text-to-3D Generation/bd4523a6-3f35-4dc6-9b48-90da3f36ebea_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e88c64d1b314dbff6935b30ca9609bb49f57d46d6f338133bcb240b5c34d223 +size 2331559 diff --git a/2025/Turbo3D_ Ultra-fast Text-to-3D Generation/full.md b/2025/Turbo3D_ Ultra-fast Text-to-3D Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f26d13347ee06ee03cc0cd3568cc2ad35f8c021e --- /dev/null +++ b/2025/Turbo3D_ Ultra-fast Text-to-3D Generation/full.md @@ -0,0 +1,441 @@ +# Turbo3D: Ultra-fast Text-to-3D Generation + +Hanzhe Hu + +Tianwei Yin² + +Sai Bi3 + +Fujun Luan3 + +Yiwei Hu3 + +Shubham Tulsiani\* + +Hao Tan3 + +Zexiang $\mathbf{X}\mathbf{u}^{3}$ + +Kai Zhang $^{3*}$ + +$^{1}$ Carnegie Mellon University + +2 Massachusetts Institute of Technology + +$^{3}$ Adobe Research + +https://turbo-3d.github.io/ + +![](images/e517bb0932019afa3b6d064b2be84d827a893f280ce57e0ff3e0a18e77382d7f.jpg) + +![](images/4279294ded69c7d28972ea4ec79bd89e461d7ade553c6f814fec3a8fa209a81c.jpg) + +a panda wearing a chef's hat and kneading bread dough on a countertop + +![](images/f94ab1b71cc78cd9b77f90ea0f79db391c63779992bee7d21d5870eacacb684d.jpg) +a plush toy of a cordi nurse + +![](images/36163932441e8a06763f9c45ba4180e6edfaf2c25c17f8ee2ccb40f610dea4c8.jpg) + +![](images/b79dc2f7e7464eb5107d954e23a323ae8e8c7ddb1ea41bf975898202dbb59a0b.jpg) + +![](images/7f0e3292aa9bb02f38ad181b9f4635d07e68a49c733c7c9e53a2189927899fb0.jpg) +a raccoon stealing a pie + +![](images/8ceb6de2b254accb7f27384fe948e4c52ab759108ac308def4abd800a9043704.jpg) + +![](images/dc7e2cfa88f59aa1a7f182cad879b1370cd4e5b66f666f45ceb9b9206faaf147.jpg) + +![](images/6c7e4a00eb82f0144894ba5fc51e4b6184341526b0bccdda3c552ab82b4eb07f.jpg) +a badger wearing a party hat and blowing out birthday candles on a cake + +![](images/d31a75e94589ea056799ae129d2e46eb6d1fd259d5f602602c067be7629e7843.jpg) + +![](images/ab0e4bd55bb548d4f610b56d92c6e74be0c0335d8c6ab5862346601347ef4245.jpg) + +![](images/74cedcfe912707107c8fd02329a4dc14d8f4bf46094c60cb960f6f313870a634.jpg) +a cake in the shape of a train + +![](images/71fea974c24db610363a01e4cd0921b13bbb7c0c4f1fc2190549b3e3ed07cd22.jpg) +Figure 1. Overview of our Turbo3D text-to-3D system. Turbo3D generates high-quality 3D Gaussian Splatting (3DGS) assets from user prompts in less than 1 second on a single A100 GPU. It's a two-stage pipeline consisting of a highly efficient latent-space few-step multi-view (MV) generator and single-step MV reconstructor. Note that we visualize latents as RGB images and 3DGS assets as point clouds in the pipeline figure for clarity. + +![](images/9c8c1f7112d652362c2dca86db6a3f6844c498f0d1b27f02f842287aacf0ff29.jpg) +Campsite with a tent on grassy ground, a backpack against it, a campfire surrounded by stones. A sleeping bag partially unrolled next to the tent, a lantern hanging from a tree branch + +![](images/2c23987bc45820a6509323818894aaee7523488f07e8f198e65441ee0d98c008.jpg) + +![](images/a0460c86b4c0cb4a7512daf422b0239c7d3ff89010d1bd9c4c1d7865d45a18a1.jpg) + +![](images/f00753f7bb123a1f3dcc94411851aa062787564c4df60e4beec670c909478467.jpg) + +![](images/32e803b35fabdcae2a9691bf90b9badb9ef65ba190ea5addaa0d36edf28c4ecf.jpg) +a lion reading the newspaper + +![](images/c586f16699d00a7d3b144bcd03349b9156805d01276c433c0ad4a2c6e47d2618.jpg) + +![](images/9ac1abafbf1804825eeeba62e9044c45142c8404de4632205e99720ac9f4ef35.jpg) + +# Abstract + +We present Turbo3D, an ultra-fast text-to-3D system capable of generating high-quality Gaussian splatting assets in under one second. Turbo3D employs a rapid 4-step, 4-view diffusion generator and an efficient feed-forward Gaussian reconstructor, both operating in latent space, as shown in Fig. 1. The 4-step, 4-view generator is a student model distilled through a novel Dual-Teacher approach, which encourages the student to learn view consistency from a multiview teacher and photo-realism from a single-view teacher. By shifting the Gaussian reconstructor's inputs from pixel space to latent space, we eliminate the extra image decoding time and halve the transformer sequence length for maximum efficiency. Our method demonstrates superior 3D generation results compared to previous baselines, while operating in a fraction of their runtime. + +# 1. Introduction + +The recent advances in image generative models allow users to generate detailed outputs from just a text prompt. While initial denoising diffusion-based methods [7, 34, 36, 37, 44, 50] enabled impressive photo-realistic generation, recent techniques [19, 22, 23, 39, 45, 60, 61] have significantly improved inference efficiency of such models, allowing high-fidelity generation in the blink of an eye. Unfortunately, these advances in generative 2D modeling methods have not yet been matched in the 3D domain, where ultra-fast realistic 3D generation remains a challenge. In this work, we seek to bridge this gap, and present Turbo3D, a text-to-3D generative model that can synthesize detailed 3D outputs in a fraction of a second. + +The existing approaches for text-based 3D inference can be categorized as either generative [10, 17, 24, 30, 57, 63] or optimization-driven [18, 32, 55]. The latter class of methods optimize 3D representations by 'distilling' pretrained 2D diffusion models [32]. While this approach can yield decent 3D outputs, it is highly inefficient, typically requiring several minutes or even hours to output a single 3D representation. The alternative paradigm is to learn a generative model that directly outputs 3D representations. While initial methods [10, 28] investigated representations such as point clouds and SDFs, the limited availability of 3D data restricted the generation quality. Recent approaches [9, 17, 24, 40, 41] have instead advocated for learning generative models of multi-view images (followed by deterministic 3D reconstruction), as these can be initialized from pre-trained 2D generative models. While these methods have resulted in impressive generations, the multiview finetuning on synthetic data does inhibit their generation quality. More crucially, their inference efficiency is restricted by the iterative denoising process required for the + +text-conditioned multi-view generation. + +In this work, we follow the paradigm of text-to-3D via multi-view generation, and aim to improve the efficiency of the underlying components to enable ultra-fast generation, while enhancing the fidelity of the generated outputs. Inspired by the recent progress in reducing inference time by distilling 2D diffusion models into one-step or few-step generators [38, 39, 45, 61], we adapt these for multi-view 3D generation. We first train a many-step text-to-multiview diffusion model and then distill it into a much faster 4-step generator using the distribution matching distillation (DMD) loss [61]. However, we find that this process leads to a significant quality degradation, as it fails to capture the full range of modes present in the multi-view teacher model. To overcome this, we propose to extend the DMD pipeline to incorporate another single-view teacher - a 2D denoising diffusion model trained on a large set of high quality aesthetic images. Our few-step multi-view generator is thus trained with a dual-teacher distillation approach, where the multi-view DMD loss helps our model learn multi-view consistency, and the single-view DMD loss ensures high-fidelity outputs. To further improve the 3D generation efficiency, we note that our multi-view generator outputs latent representations (and not pixels) for the multi-view images. We build on this insight to adapt a prior multi-view to 3D reconstruction approach [62] to instead consume multi-view latents as input, and show that this improves the reconstruction efficiency without any performance loss. + +Our overall system thus combines an efficient few-step multi-view generator with a multi-view latent to 3D model to enable ultra fast 3D generation. We train our model on the subset of Objaverse dataset [3, 4], which contains about 400k instances with Cap3D text captions [25]. We show that our proposed Turbo3D is able to produce high-quality 3D assets in less than one second (see Fig. 1), and also achieves comparable quality with previous state-of-the-art. + +# 2. Related Work + +In this section, we discuss the closely related prior feedforward 3D generation methods. Optimization-based methods [18, 32] is out of this work's scope. We review feedforward 3D generative models in the following three categories: 1) methods that directly generate full 3D representation encoding geometry and appearance; 2) methods that generate shapes and then generate the textures; 3) methods that generate multi-view (MV) images followed by reconstruction. We also review the diffusion distillation literature which our method is built upon. + +# 2.1. Directly Genereating Full 3D Representation + +Several prior methods [1, 10, 30, 46, 48, 54, 58] have been introduced that directly generate 3D representations encoding both geometry and texture, e.g., implicit fields [10], + +point clouds [30], and triplanes [54]. This line of work typically have to preprocess the source 3D data (meshes or multi-view images) into the target representations for generative models in a lossy fashion, hence limiting their quality and scalability. + +Alternative methods [1, 46, 48, 58] have been proposed to get rid of the lossy preprocessing. They implicitly bake a 3D generative process into a multi-view diffusion framework, creating a single-stage 3D generative model. However, trained only on 3D data, they suffer from the issue of limited generalization, lower pixel quality and reduced understanding of complex text prompts, compared with methods leveraging powerful pretrained text-to-image models [17, 21]. + +Our method leverages the strong priors in a pretrained image generative model, but we go a step further by distilling the slow teacher model into a fast student generator for ultra-fast text-to-3D. + +# 2.2. Shape Generation + Texture Generation + +Recent advancements in 3D content creation have introduced innovative methods [42, 57, 63] that focus on generating high-quality 3D shapes first, followed by generative texturing [2, 35], rather than creating both simultaneously. CLAY [63] offers a framework using a multiresolution Variational Autoencoder (VAE) and a latent Diffusion Transformer (DiT) to initially create detailed 3D geometries from inputs like text and images, before applying high-resolution physically-based rendering (PBR) textures. Direct3D [57] enhances scalability in image-to-3D generation by improving the performance of 3D VAE and DiT for generating 3D shapes, subsequently adding textures. + +However, besides the challenge of learning a compact latent space suitable for generation, these methods are also usually slow as a result of the two diffusion models - one for shapes and the other for textures. In contrast, our Turbo3D only have one diffusion model for generating multi-views and is fast once distilled, while the reconstructor is deterministic and hence fast. + +# 2.3. MV Generation + MV Reconstruction + +A large body of the 3D generation work [17, 20, 24, 41] focus on leveraging multi-view generation and reconstruction to enhance quality and efficiency. Instant3D [17] offers a fast method for creating high-quality 3D assets from text prompts by combining sparse-view generation with a transformer-based reconstructor [8, 53, 56, 62], significantly reducing inference time. MVDream [41] employs a multi-view diffusion model to improve consistency and stability in 3D generation, integrating 2D and 3D data. SyncDreamer [24] enhances multiview-consistency from a single-view image using a 3D-aware feature attention mechanism. One-2-3-45 [20] introduces an efficient ap + +proach to single image 3D reconstruction without extensive optimization, producing consistent 3D meshes. SV3D [51] utilizes a latent video diffusion model for novel multi-view synthesis, incorporating explicit camera control to improve 3D reconstruction quality. + +We follow the same approach of reconstructing generated multi-views to create 3D assets in this work, but we focus on improving the efficiency of such systems. With our novel Dual-Teacher Distillation and Latent GS-LRM components, our Turbo3D manages to be at least an order of magnitude faster than these baselines while maintaining competitive quality. Concurrent work GECO [52] also use diffusion distillation to speed up image-to-3D. However, we differ in our dual-teacher distillation design with focus on text-to-3D. Our pipeline is also simpler and avoids the cumbersome mesh reconstructions for 3D distillation in GECO. + +# 2.4. Diffusion Distillation + +Recent progress in diffusion models have focused on improving the efficiency of image generation by reducing the number of sampling steps required, leading to the development of several innovative distillation techniques [26, 27, 29, 38, 39, 45, 59-61]. Methods like Improved Rectified Flows [15] and InstaFlow [23] straighten the ODE trajectories, making them easier to approximate with a one-step student model. Consistency Models [45] train student generators to map any point on the teacher's ODE trajectory to a consistent target, enabling one-step and few-step image generation. Distribution Matching Distillation (DMD) [61] trains a one-step generator by minimizing the reverse KL divergence between the data distribution and the generator's output distribution [5, 26, 55]. Building on this, DMD2 [60] integrates GAN losses [6] and extends the method to multi-step generators, further enhancing generation quality. + +However, most of existing approaches focus on distilling pretrained 2D image diffusion models. In our work, we adopt the popular DMD [61] approach and extend it to the multi-view domain. Distilling multi-view models presents unique challenges, such as increased mode collapse due to fine-tuning and distillation processes. To address this issue, we introduce a novel dual-teacher distillation technique, enabling swift, photorealistic multi-view generation. + +# 3. Background + +# 3.1. Multi-view Diffusion Model + +Diffusion models generate data from noise by reversing a forward noising process. Given a data sample $x_0 \sim p(x_0)$ from the data distribution, the forward process progressively adds Gaussian noise over $T$ timesteps. At timestep $t$ , the noised distribution conditioned on $x_0$ is given by: + +$$ +q \left(x _ {t} \mid x _ {0}\right) = \mathcal {N} \left(x _ {t}; \alpha_ {t} x _ {0}, \beta_ {t} ^ {2} \mathbf {I}\right), \tag {1} +$$ + +where $\alpha_{t}$ and $\beta_{t}$ are timestep-dependent constant that are specified by the noise schedule [12, 44]. The diffusion model $\epsilon_{\theta}$ is trained to reverse this noising process by learning to predict the added Gaussian noise $\epsilon$ : + +$$ +L (\theta) = \mathbb {E} | | \epsilon - \epsilon_ {\theta} (x _ {t}, t) | | ^ {2}. \tag {2} +$$ + +There exists other formulations, in which the diffusion model learns to predict the clean input $x_0$ directly [11], or a combination of $x_0$ and $\epsilon$ [38]. Regardless of the specific prediction target, the outputs of these models can be related to the score function, which is the gradient of the log probability density of the data distribution [44]: + +$$ +s _ {\theta} \left(x _ {t}, t\right) = \nabla_ {x _ {t}} \log p \left(x _ {t}\right) = - \frac {\epsilon_ {\theta} \left(x _ {t} , t\right)}{\beta_ {t}} = - \frac {x _ {t} - \alpha_ {t} x _ {\theta} \left(x _ {t} , t\right)}{\beta_ {t} ^ {2}} \tag {3} +$$ + +where $\epsilon_{\theta}$ [7] and $x_{\theta}$ [11] represents noise and data prediction diffusion models, respectively. In our paper, we adopt the noise prediction scheme but our method generalizes to any formulations. + +In multi-view diffusion models, the generation of a 3D scene is conditioned on a text prompt, enabling the joint denoising of multiple views to produce a set of 3D-consistent output images [17, 41]. Specifically, given a text prompt $c$ and a set of multi-view images $\{x^i\}_{i=1}^K$ where $K$ is the number of view, the multi-view diffusion model learns to predict the added noise across all views simultaneously. The training objective is formulated as: + +$$ +\mathbb {E} | | \epsilon - \epsilon_ {\theta} \left(\left\{x _ {t} ^ {i} \right\} _ {i = 1} ^ {K}, t, c\right) | | ^ {2}. \tag {4} +$$ + +where $\epsilon$ represents independent Gaussian noise with the same variance applied to each view. At inference time, the generation starts from a set of fully noisy multi-view images $\{x_{T}^{i}\}_{i = 1}^{K}$ sampled from a standard Gaussian distribution. The multi-view diffusion model iteratively generate a sequence of cleaner multi-view images. Various diffusion samplers [7, 43] can be employed during this process to optimize generation speed and quality. + +# 3.2. Distribution Matching Distillation + +Distribution Matching Distillation (DMD) is a widely used diffusion distillation technique that converts teacher diffusion models into a student generator requiring significantly fewer sampling steps [60, 61]. The DMD approach trains the student generator $G_{\theta}$ by minimizing an approximate reverse KL divergence between the smoothed student's output distribution (denoted as $p_{\mathrm{fake}}$ ) and the smoothed data distribution (denoted as $p_{\mathrm{real}}$ ): + +$$ +L _ {\mathrm {D M D}} (\theta) = D _ {\mathrm {K L}} \left(p _ {\text {f a k e}} \| p _ {\text {r e a l}}\right) = \mathbb {E} _ {x, t} \left(\log \left(\frac {p _ {\text {f a k e}} \left(x _ {t}\right)}{p _ {\text {r e a l}} \left(x _ {t}\right)}\right)\right). \tag {5} +$$ + +Let the score functions of the data distribution and the student's output distribution be denoted as $s_{\mathrm{real}}$ and $s_{\mathrm{fake}}$ , respectively. The gradient of this KL divergence can be effectively approximated by the difference between these two + +score functions: + +$$ +\begin{array}{l} \nabla_ {\theta} L _ {\mathrm {D M D}} (\theta) \approx \mathbb {E} \left[ - \int \left(s _ {\text {r e a l}} \left(F \left(G _ {\theta} (\epsilon), t\right), t\right) \right. \right. \\ \left. - s _ {\text {f a k e}} \left(F \left(G _ {\theta} (\epsilon), t\right), t\right)\right) \frac {d G _ {\theta} (\epsilon)}{d \theta} d \epsilon \Bigg ] \tag {6} \\ \end{array} +$$ + +where $F$ represents the forward diffusion process defined in Eq. 1. The student generator can be adapted for a multi-step generation setting by replacing the pure noise input $\epsilon$ with a partially noisy image $x_{t}$ [60]. During training, the data distribution's score function is initialized from the teacher diffusion model and kept fixed, while the student's output distribution score function is dynamically trained using the student's output and a denoising loss (Eq. 4). + +# 4. Method + +In this section, we provide the technical details of our Turbo3D text-to-3D system, which features a highly efficient multi-view (MV) generator and reconstructor. We begin by describing our novel DualTeacher Distillation approach; it creates a rapid MV generator by jointly distilling knowledge from both a multi-view (MV) teacher and a single-view (SV) teacher. Following this, we discuss the latent-space GS-LRM that instantly lifts the generated MV latents into high-quality 3D Gaussians. + +# 4.1. Dual-teacher Distillation for MV Diffusion + +Leveraging a multi-step multi-view diffusion in a text-to-3D generation process can be inefficient due to repeated evaluations of the diffusion denoiser in the sampling process. To speed this up, one approach is to use diffusion distillation [60, 61] to train a single-step or few-step MV generator. + +However, we find that naively applying diffusion distillation methods to a MV teacher can cause the student model to generate overly simplistic and cartoonish appearance, which closely resembles the 3D Objverse dataset used during MV teacher finetuning and distillation [3] (shown in Fig. 6). We call this phenomenon 'compounding mode collapse'; this happens because both finetuning and distillation sacrifices generation diversity for efficiency. As the MV teacher is already biased towards synthetic Objverse-style appearance, further distilling it will have compound effect of locking the distilled generator in the mode of Objverse data that are far from the modes of photorealistic natural images. + +To address this issue, we propose to use dual teachers in the distillation process: one MV teacher to teach the student model about multi-view consistency, and one SV teacher to teach about each views' photo-realism. We illustrate this Dual-teacher Distillation algorithm in Fig. 2. Concretely, this is formulated as: + +![](images/b6c8af2b5b0d9295f570d9c4e21f14d471ce453b0f87792ba39f1cdcdc104285.jpg) +Figure 2. Dual-teacher distillation framework in our Turbo3D. Note that latents are visualized as RGB images for clarity. We aim to distill a multi-step multi-view teacher generator (right, green) into a few-step multi-view generator (left, blue). Our few-step MV student generator is conditioned on Plücker embeddings for better 3D awareness. Similar to [60], we optimize the student generator using distribution matching objective (DMD loss) and train the fake score function to model the distribution of samples produced by the student generator. In particular, we integrate two teacher models: multi-view teacher and single-view (SV) teacher to enhance both multi-view consistency and photorealism. The MV score functions take a set of images of one object as input and calculate the MV DMD loss, while the SV score functions treat each image separately and calculate the SV DMD loss. + +$$ +\begin{array}{l} L _ {\mathrm {D M D}} ^ {\text {D u a l}} (\theta) = D _ {\mathrm {K L}} \left(p _ {\text {f a k e}} \left(\left\{x _ {t} ^ {i} \right\} _ {i = 1} ^ {K}\right) \mid \mid p _ {\text {r e a l}} ^ {\mathrm {M V}} \left(\left\{x _ {t} ^ {i} \right\} _ {i = 1} ^ {K}\right)\right) \\ + \lambda \cdot \frac {1}{K} \sum_ {i = 1} ^ {K} D _ {\mathrm {K L}} \left(p _ {\text {f a k e}} \left(x _ {t} ^ {i}\right) \mid \mid p _ {\text {r e a l}} ^ {\mathrm {S V}} \left(x _ {t} ^ {i}\right)\right), \tag {7} \\ \end{array} +$$ + +where $p_{\mathrm{real}}^{\mathrm{MV}}(x_t), p_{\mathrm{real}}^{\mathrm{SV}}(x_t)$ represent MV and SV teachers, respectively; $\lambda$ is the loss weight balancing the influence of MV and SV teachers on the distilled student model; $K = 4$ is the number of views. We set $\lambda$ to 1 in our experiments. As demonstrated in Fig. 6, having the additional SV teacher in the distillation process effectively addresses the compound mode collapse issue, because it tries to pull each view to look like natural images. + +# 4.2. Latent GS-LRM for MV Reconstruction + +To reconstruct 3D from the generated MV latents, one straightforward approach is to decode them into multi-view images, and then use the pixel-space GS-LRM [62] to produce the 3D Gaussians. However, such an approach can suffer from efficiency and memory issue when scaling to high resolution due to the poor performance of Conv2D operators in VAE decoder [13]. + +We propose to skip the VAE decoding and directly input the generated MV latents to a latent GS-LRM for best efficiency. To train such a model, we supervise the reconstructed Gaussians with pixel-space novel-view rendering losses $(\ell_{2}$ and perceptual losses as in [62]). We show that replacing pixel-space GS-LRM with a latent one does not affect the quality of generated assets in Tab. 3 and Fig. 3, while being faster. + +![](images/770fab4cf3e96e8ae624c6d3039b8f00cd3b24ae52269edd1657ee56df65a012.jpg) +Figure 3. We compare the renderings of pixel GS-LRM and latent GS-LRM. Latent GS-LRM achieves comparable reconstruction quality as pixel GS-LRM. + +# 5. Experiments + +In this section, we first describe our experimental setup (Section 5.1). We then compare our method with state-of-the-art text-to-3D baselines (Section 5.2). Finally, we ablate each component of our framework to showcase their effectiveness (Section 5.3). + +# 5.1. Experimental Setup + +Datasets. We use the Objaverse dataset [3] to train both our multi-view generation model and multi-view reconstruction model. We scale the objects and center them to fit into a cube $[-1, 1]^3$ . For the generation task, we render the dataset at a fixed elevation (20 degrees) and 16 equidistant azimuths to achieve a good coverage of the object. We render using a field of view $50^o$ at a distance of 2.7 and uniform lighting. For the reconstruction task, we render 32 views ran- + +domly placed around the object with a random distance in the range of [1.5, 2.8]. We render a total of $730K$ objects. + +Baselines. We adopt Instant3D [17] and LGM [47] as our baseline text-to-3D methods. However, since the field of fast text-to-3D is relatively underexplored, we also include recent state-of-the-art fast image-to-3D methods TripoSR [49] and SV3D [51] as baseline methods. For a fair comparison, we use a popular few-step text-to-image model Flux [14] to generate an input image first and then feed it to the image-to-3D models. The inference time of image-to-3D models is a summation of the two parts. + +Metrics. We adopt the CLIP score [33] and VQA score [16] to assess the semantic alignment between the generated results and text prompts. We use 400 text prompts from DreamFusion [32] for evaluation. We generate one object for each prompt. For each generated 3D object, we render 10 random views and calculate the average CLIP score and VQA score between the rendered images and the input text. For inference time, we report it with all methods under the same image resolution of 256. Notably, some methods, e.g., Instant3D [17], only support a higher resolution of 512. For a fair comparison, we report their quantitative results under the resolution of 512 and only inference time on the resolution of 256. + +Implementation Details. The whole pipeline of our method includes three training phases. We first train a multi-step multi-view diffusion model on the Objverse dataset, by fine-tuning an internal DiT [31] based text-to-image model. We train this model for 30k iterations with 32 80G A100 GPUs using a total batch size of 128 and a learning rate of $3e^{-5}$ . Then we perform distillation to distill the multi-step multi-view generator into a few-step multi-view generator, which takes 10k iterations with 32 80G A100 GPUs with a global batch size of 128 and a learning rate of $5e^{-6}$ . Finally, we train a reconstruction model - latent GS-LRM - from scratch, which takes 80k iterations with 32 80G A100 GPUs with a total batch size of 256 and a learning rate of $4e^{-4}$ . + +# 5.2. Evaluation against baselines + +Qualitative Comparisons. As shown in Fig. 4, our method generates significantly better results compared with LGM [47] and Instant3D [17]. In particular, LGM tends to produce simple 3D assets without geometry and texture details; it also lacks robustness, oftentimes generating low-quality 3D assets that are broken, or suffer from multi-face Janus problem [32], or do not closely follow provided text prompts. Instant3D is much more robust and able to produce plausible 3D assets most of the time. However, its performance on text-image coherence is limited compared with our Turbo3D. For example, in the first example, LGM's rendition appears oversimplified with indistinct egg and yolk boundaries, while Instant3D provides a + +
MethodClip Score ↑VQA Score ↑Inference Time ↓
TripoSR [49]23.850.571.19s
SV3D [51]24.920.6412.52s
Instant3D [17]26.230.6515.02s
LGM [47]24.730.586.56s
Turbo3D (Ours)27.610.760.35s
+ +Table 1. Comparison against state-of-the-art 3D generation methods. Our Turbo3D generates 3D assets with highest CLIP and VQA scores while using the least amount of time (benchmarked on a A100 GPU). + +closer match but lacks fine details in the egg's structure and wood texture and fails to capture the concept of 'spill out'. For complex prompts that describes a scene with multiple objects (4th and 6th rows), LGM generates a Janus asset while Instant3D miss a lot of concepts like blanket, basket, wildflowers etc; our generations adhere to the prompts much more closely. These qualitative results highlight our Turbo3D's ability to generate high-quality, text-aligned 3D models with a superior level of detail, realism, and coherence across diverse and complex prompts, outperforming both LGM and Instant3D by a large margin. + +Quantitative Comparisons. Tab. 1 presents a quantitative comparison of our proposed method, Turbo3D, against several state-of-the-art approaches, including TripoSR [49], SV3D [51], Instant3D [17], and LGM [47], on the text-to-3D task. Our proposed Turbo3D achieves the highest CLIP Score of 27.61 and VQA Score of 0.76, outperforming other methods by a significant margin in both quality metrics. In addition to the quality improvement, Turbo3D demonstrates remarkable efficiency with an inference time of only 0.35 seconds, substantially faster than competing methods. Although TripoSR is able to generate a 3D asset in only 1.19s, the quality of the generated results is highly limited. As a result, our proposed Turbo3D is able to achieve outstanding performance in terms of both quality and inference speed compared with state-of-the-art methods. + +User Study. We run a user study by randomly selecting 80 text prompts and asking 56 users to make 1120 pairwise comparisons (users are shown an input text prompt and two generated 3D assets from two anonymous methods). We show results in Fig. 5. Since the quantitative results of TripoSR is not competitive with others, it is not included in the user study. Moreover, SV3D only outputs videos instead of 3D representations of objects, so it is hard to perform a fair comparison. Therefore, we compare our Turbo3D with LGM [47], Instant3D [17], and our MV Teacher. In particular, MV Teacher is the model which our Turbo3D gets distilled from. When compared to LGM, Turbo3D achieved a win rate of $89.8\%$ , with only $10.2\%$ of participants favoring LGM. Against Instant3D, Turbo3D also outperformed + +![](images/dd7c273064bb74e2de53156680c2486bd0fb7e7fb3a1e995c1bc9a9af9c02ea5.jpg) +A cracked egg with the yolk spilling out on a wooden table + +![](images/79b84fdfa282a3c47faf90b10148611bd13ba19afe4f359b98645d9cd136ad01.jpg) +An origami hippo in a river + +![](images/17f9a0dbb3f38c3f6bdbe582d1879bfa5bf825a056ddaf299cb146be85340ca9.jpg) +Wedding dress made of tentacles + +![](images/731bf3ab30d05f460e110f61a89898a6040dee9fb6052404f82de3b7a08ee963.jpg) +A cozy campfire surrounded by logs for sitting, with a small tent set up nearby. Add a pot hanging over the fire and a few scattered leaves + +![](images/9d73cf11344764951ce8ad6a64202543f797d544ff427b3cd079dba0474f92c8.jpg) +A chimpanzee looking through a telescope + +![](images/3f1aa7a34d3157c18aa86fcb26e704a7cae3452bb701fb0face8ccad5618fb9e.jpg) +Picnic scene with a checkered blanket on grass, a wicker basket with sandwiches and fruit, a bottle of lemonade with glasses. An open book, a small bouquet of wildflowers on the blanket +LGM (6.56s) + +![](images/25a0fcdb3fabe4b14e256af2acc145b72a37f4b62e89e4fb5c02601a318a8152.jpg) + +![](images/75b9e66296042d17b04eca594b43ad1de34f2cad7a287dbe510221d059b6a977.jpg) + +![](images/aa328635e59287c006911e2f51f0b6db3a1f9d5fdc3f5f84b016084a56157494.jpg) + +![](images/425ab4eb59c40b5d37b3eddc27f7399d2a662ed62d4b7fd0fdac0b8ca4783fa3.jpg) + +![](images/a4c4fcda4d11b1a8c6b0c68a886ca6a56ffeac786a2eb6c5bad0653224b77f51.jpg) + +![](images/130f858e1a4c7f5c41446305cdade90df368422fcc29c2672615cdc12d33de35.jpg) +Instant3D (15.02s) + +![](images/446a4a76926606f4e79759f91f976df15c955004eee456a5e17e9af1ba99c31c.jpg) + +![](images/fad229eb454b2d7b8c139f5736523156725d6735512b45b5548679213e3a50f4.jpg) + +![](images/f1fe5b925e4af1be6e702545b3be5ca8b3b90b2f9f4f49cc9eefa5a5ccaee05e.jpg) + +![](images/0dfc3dcbe4e34172619ae622997cabc2216f5986c10284e30352acf1302a4d0d.jpg) + +![](images/b730d47708a28172c7cda6492f28cc1e8d5d473b2eea05b461b9358785307723.jpg) + +![](images/fb856587482a4dcb9a3a4f0f154b0045081a146f5d9335315f4b2ab15c95eac8.jpg) + +![](images/dab9d381ead6a661e1962505b65d99c070d4002c1c004bf6c0d9401d9319c5bb.jpg) + +![](images/f8b6f3b3370d3818c0b6370d45e7503073c96445286fb28d3aa9194afe89e030.jpg) + +![](images/02c8665de1cd69726625f24da476647e80db1de7c11175dca078927a8d2188a4.jpg) + +![](images/c381af4e1ec5aa84b323ff3df4909d0ebbbe3ace19496407f3ef1f8caeb98a60.jpg) +Turbo3D (0.35s) + +![](images/c696e2855ab09dbcd9814aec49a813d3a2d579cb39f29b015a8062884d6d6ef2.jpg) +Figure 4. Comparison of our Turbo3D against baselines LGM [47] and Instant3D [17]. Among these methods, Our method generates the most detailed and physically plausible 3D assets, closely adhering to the provided text prompts. In contrast, LGM tends to generate broken assets with Janus issue [32], while Instant3D has poorer text alignment, oftentimes missing some concepts, e.g., 'spilling out' in the first row, 'river' in the second row, etc. +Figure 5. User study results comparing our Turbo3D to baseline LGM [47], Instant3D [17], and our slow MV teacher. Our Turbo3D is consistently preferred over baseline LGM and Instant3D, while having on-par preference with our MV teacher. See Fig. 4,6 for visual comparison. + +with a $74.9\%$ win rate, indicating that users consistently found Turbo3D's outputs more aligned with the input text descriptions and more visually compelling. When evaluated against our teacher model, Turbo3D held a close win rate of $50.6\%$ , with $49.4\%$ preferring MV T., reflecting compara + +ble quality between the student and teacher models. The result indicates that our Turbo3D not only achieves significant speedup with distillation, but also preserves the generation ability from the teacher model. + +# 5.3. Ablation Study + +Effect of Single-view Teacher. In Tab. 2, we demonstrate the effectiveness of the dual teacher distillation strategy. The first line is our multi-view (MV) teacher model, which achieves impressive results but runs slowly due to the many diffusion sampling steps required. When performing distillation only with this MV Teacher, the quality drops by a large margin as shown in the second row. When adding a single-view teacher model for distillation, the distilled model achieves much better results compared with the previous one. This configuration approaches the performance of the Multi-step MV Model while maintaining + +![](images/14c11aab51a2b79b3b14707f6134539692ef778fc0bc731541c042c1a9f400da.jpg) +a baby dragon hatching out of a stone egg + +![](images/b8a75eb4345c7411a2cf574b66ba1fb290d9237c1f8cf05f8cb0dce7d3a4ccb1.jpg) + +![](images/b4a0e75f9e2d85803a89cf93b419fb492d279819a7a51a6888e7e5e42bb71266.jpg) + +![](images/35e5ca8550afdc03748a2a0e9ca5eb6ac891282671ee7e42a3a4b561e55b46e3.jpg) + +![](images/221515456723992d465e999b011480f321f742f645c379b9381e8560c46c7969.jpg) + +![](images/87998b6538d655386135afc1be9221cd7f36f6ae4ae860210a9c5676857b3324.jpg) + +![](images/5351ae9224cfdf8588d2fdb2cf06ba23c523833290ab104683d10d652ea79188.jpg) +a kitten +standing on top +of a giant +tortoise + +![](images/4173eedb7dd2598e8584a115254ad1fe1e880c543257495dddccbbe6dd28edb4.jpg) + +![](images/6da348535d4671af31fa25b97a7e31cbe88326dc82c62094386658935ba50fc0.jpg) + +![](images/a326bc282810c2a4fbb0bb9b37e16a49e55517b373e8ef6419aff1fb6091c3d7.jpg) + +![](images/76ee5af2d72470e2510d1cff3ba3252db76b0f0956dcb23be217a629713ef14c.jpg) + +![](images/f546f40024d31519de3334ac10cf67992d2227877687a309f140222f17a80ada.jpg) + +![](images/ceaa9e9801aa331355267dec446402868bc2d274aeae513bdfe22a97baa2f32c.jpg) +a wide angle DSLR photo of a squirrel in samurai armor wielding a katana + +![](images/084eb55434eec498e2f4ee1aee37e7fecbea33d68cabfcd9bd036c857a6b3c04.jpg) +MV teacher (10.18s) +Figure 6. Ablation of our Dual-teacher distillation algorithm. Naively distilling MV teacher (middle column) causes compound mode collapse (see Sec. 4.1), producing overly smooth synthetic-looking assets. Our dual-teacher distillation (right column) fixes the issue and generates 3D assets that are as photorealistic as, if not more than, the baseline MV teacher (left column). We also include the inference timings for each method; the distilled model is $\sim 50\mathrm{x}$ faster than the teacher model. + +![](images/cade8740465662074ba3ac19e8e4387a4709a3b6230231096e61f39b3730f896.jpg) +MV teacher only distillation (0.35s) + +![](images/d638913a9d51230b9845e32ea9c8f108c978159f5ed354372e520e91edfe6b54.jpg) + +![](images/79eb8e06658290f62d81bb0f6fd1dc01c89e28c7b488173e89f41021b4b1d192.jpg) + +![](images/0853bc60ea13790ae35eb4300b8725763e8fa5c6a4ff96e8b4f0e79c97f71970.jpg) +Dual teacher distillation (0.35s) + +
AblationCLIP Score ↑VQA Score ↑
Multi-step MV Model28.040.77
Few-step Model (MV Teacher)26.600.69
Few-step Model (Dual Teacher)27.610.76
+ +the efficiency benefits of the few-step setup, showcasing the advantages of the dual-teacher strategy in our distillation. We also provide visualizations of the three models in Fig. 6. These comparisons show that the dual teacher distillation model strikes a balance between detail retention and stylization, closely replicating the MV teacher model's quality while benefiting from the efficiency gains of distillation. + +Effect of Latent GS-LRM. In Tab. 3, we showcase the effectiveness of latent GS-LRM. Compared with the original GS-LRM which operates in pixel space, our latent GS-LRM is able to skip the expensive image decoding process while achieving similar image quality. + +# 6. Conclusion + +In this work, we propose Turbo3D for ultra-fast text-to-3D generation. To enable fast multi-view generation, we pro + +Table 2. Ablation of dual teacher distillation. Distillation leads to quality drop compared with the MV teacher model. Compared with naively distilling the single MV teacher, dual-teacher distillation leads to much smaller quality drop. See Fig. 6 for visual comparison. + +
AblationCLIP Score ↑VQA Score ↑Inference Time ↓
Pixel GS-LRM [62]27.620.760.45s
Latent GS-LRM27.610.760.35s
+ +Table 3. Ablation for latent GS-LRM. We report the CLIP score, VQA score, and overall text-to-3D inference time for comparison. Our latent GS-LRM achieves similar image quality while enabling better efficiency ( $\sim$ 22% speedup). + +pose to distill a multi-step multi-view generator into a few-step multi-view generator. Moreover, to restore the multiview consistency and photo-realism during distillation, we introduce a novel dual-teacher distillation framework. To further improve the multi-view reconstruction efficiency, we propose a latent GS-LRM which directly reconstructs 3D Gaussians from multi-view latents. Extensive experiments demonstrate that our proposed Turbo3D is able to achieve outstanding performance in terms of both generation quality and inference efficiency. + +# Acknowledgements + +This work began during Hanzhe Hu and Tianwei Yin's internships at Adobe Research. This research was supported in part by NSF award IIS-2345610. + +# References + +[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. Renderdiffusion: Image diffusion for 3d reconstruction, inpainting and generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12608-12618, 2023. 2, 3 +[2] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18558-18568, 2023. 3 +[3] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 4, 5, 1 +[4] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects. Advances in Neural Information Processing Systems, 36, 2024. 2 +[5] Jean-Yves Franceschi, Mike Gartrell, Ludovic Dos Santos, Thibaut Issenhuth, Emmanuel de Bezenac, Mickaël Chen, and Alain Rakotomamonjy. Unifying gans and score-based diffusion as generative particle models. In NeurIPS, 2023. 3 +[6] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 3 +[7] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 4 +[8] Yicong Hong, Kai Zhang, Jiumiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 3 +[9] Hanzhe Hu, Zhizhuo Zhou, Varun Jampani, and Shubham Tulsiani. Mvd-fusion: Single-view 3d via depth-consistent multi-view generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9698-9707, 2024. 2 +[10] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2 +[11] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 4 +[12] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. In NeurIPS, 2021. 4 +[13] MIT HAN Lab. Patch conv: Patch convolution to avoid largegpu memory usage. https://hanlab.mit.edu/blog/patch-conv, 2024. Accessed: 2024-11-14. 5 +[14] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024. Accessed: 2024-11-14. 6 + +[15] Sangyun Lee, Zinan Lin, and Giulia Fanti. Improving the training of rectified flows. arXiv preprint arXiv:2405.20320, 2024. 3 +[16] Baiqi Li, Zhiqiu Lin, Wenxuan Peng, Jean de Dieu Nyandwi, Daniel Jiang, Zixian Ma, Simran Khanuja, Ranjay Krishna, Graham Neubig, and Deva Ramanan. Naturalbench: Evaluating vision-language models on natural adversarial samples. arXiv preprint arXiv:2410.14669, 2024. 6 +[17] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214, 2023. 2, 3, 4, 6, 7, 1 +[18] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 2 +[19] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 2 +[20] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems, 36, 2024. 3 +[21] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023. 3 +[22] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 2 +[23] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In The Twelfth International Conference on Learning Representations, 2023. 2, 3 +[24] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 2, 3 +[25] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems, 36, 2024. 2 +[26] Weijian Luo, Tianyang Hu, Shifeng Zhang, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang. Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models. In NeurIPS, 2023. 3 +[27] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In CVPR, 2023. 3 +[28] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. AutoSDF: Shape priors for 3d completion, reconstruction and generation. In CVPR, 2022. 2 + +[29] Thuan Hoang Nguyen and Anh Tran. Swiftbrush: One-step text-to-image diffusion model with variational score distillation. In CVPR, 2024. 3 +[30] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2, 3 +[31] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 6, 1 +[32] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 6, 7 +[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6 +[34] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 2 +[35] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAPH 2023 conference proceedings, pages 1–11, 2023. 3 +[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2 +[37] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2 +[38] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 2, 3, 4 +[39] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. arXiv preprint arXiv:2403.12015, 2024. 2, 3 +[40] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 2 +[41] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 3, 4 +[42] Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle + +meshes with decoder-only transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19615-19625, 2024. 3 +[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 4 +[44] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2, 4 +[45] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 2, 3 +[46] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Viewset diffusion:(0-) image-conditioned 3d generative models from 2d data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8863-8873, 2023. 2, 3 +[47] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1-18. Springer, 2025. 6, 7, 1 +[48] Ayush Tewari, Tianwei Yin, George Cazenavette, Semon Rezchikov, Josh Tenenbaum, Frédo Durand, Bill Freeman, and Vincent Sitzmann. Diffusion with forward models: Solving stochastic inverse problems without direct supervision. Advances in Neural Information Processing Systems, 36: 12349-12362, 2023. 2, 3 +[49] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151, 2024. 6, 1 +[50] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural computation, 23(7):1661-1674, 2011. 2 +[51] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitrii Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. SV3D: Novel multi-view synthesis and 3D generation from a single image using latent video diffusion. In European Conference on Computer Vision (ECCV), 2024. 3, 6, 1 +[52] Chen Wang, Jiatao Gu, Xiaoxiao Long, Yuan Liu, and Lingjie Liu. Geco: Generative image-to-3d within a second. arXiv preprint arXiv:2405.20327, 2024. 3 +[53] Peng Wang, Hao Tan, Sai Bi, Yinghao Xu, Fujun Luan, Kalyan Sunkavalli, Wenping Wang, Zexiang Xu, and Kai Zhang. Pf-lrm: Pose-free large reconstruction model for joint pose and shape prediction. arXiv preprint arXiv:2311.12024, 2023. 3 +[54] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4563-4573, 2023. 2, 3 +[55] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and + +diverse text-to-3d generation with variational score distillation. In NeurIPS, 2023. 2, 3 +[56] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Meshlrm: Large reconstruction model for high-quality mesh. arXiv preprint arXiv:2404.12385, 2024. 3 +[57] Shuang Wu, Youtian Lin, Feihu Zhang, Yifei Zeng, Jingxi Xu, Philip Torr, Xun Cao, and Yao Yao. Direct3d: Scalable image-to-3d generation via 3d latent diffusion transformer. arXiv preprint arXiv:2405.14832, 2024. 2, 3 +[58] Yinghao Xu, Hao Tan, Fujun Luan, Sai Bi, Peng Wang, Jiahao Li, Zifan Shi, Kalyan Sunkavalli, Gordon Wetzstein, Zexiang Xu, et al. Dmv3d: Denoising multi-view diffusion using 3d large reconstruction model. arXiv preprint arXiv:2311.09217, 2023. 2, 3 +[59] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image generation via diffusion gans. In CVPR, 2024. 3 +[60] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. arXiv preprint arXiv:2405.14867, 2024. 2, 3, 4, 5 +[61] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6613-6623, 2024. 2, 3, 4 +[62] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 2, 3, 5, 8, 1 +[63] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 2, 3 \ No newline at end of file diff --git a/2025/Turbo3D_ Ultra-fast Text-to-3D Generation/images.zip b/2025/Turbo3D_ Ultra-fast Text-to-3D Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4634a4e3086fbc6a5b4ccc7d6c973802309eea57 --- /dev/null +++ b/2025/Turbo3D_ Ultra-fast Text-to-3D Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5b662ef4725d5fe75f5b2bd25aee3efffdf9eeb84b9a13b2164702bd39c5f476 +size 599741 diff --git a/2025/Turbo3D_ Ultra-fast Text-to-3D Generation/layout.json b/2025/Turbo3D_ Ultra-fast Text-to-3D Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8500684ede5c1733f4fd2909915aa565895b8ecb --- /dev/null +++ b/2025/Turbo3D_ Ultra-fast Text-to-3D Generation/layout.json @@ -0,0 +1,10849 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 171, + 103, + 440, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 103, + 440, + 119 + ], + "spans": [ + { + "bbox": [ + 171, + 103, + 440, + 119 + ], + "type": "text", + "content": "Turbo3D: Ultra-fast Text-to-3D Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 142, + 130, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 142, + 130, + 156 + ], + "spans": [ + { + "bbox": [ + 70, + 142, + 130, + 156 + ], + "type": "text", + "content": "Hanzhe Hu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 155, + 143, + 220, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 143, + 220, + 156 + ], + "spans": [ + { + "bbox": [ + 155, + 143, + 220, + 156 + ], + "type": "text", + "content": "Tianwei Yin²" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 171, + 157, + 207, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 157, + 207, + 170 + ], + "spans": [ + { + "bbox": [ + 171, + 157, + 207, + 170 + ], + "type": "text", + "content": "Sai Bi3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 244, + 143, + 306, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 143, + 306, + 157 + ], + "spans": [ + { + "bbox": [ + 244, + 143, + 306, + 157 + ], + "type": "text", + "content": "Fujun Luan3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 330, + 143, + 381, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 143, + 381, + 156 + ], + "spans": [ + { + "bbox": [ + 330, + 143, + 381, + 156 + ], + "type": "text", + "content": "Yiwei Hu3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 247, + 158, + 345, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 158, + 345, + 171 + ], + "spans": [ + { + "bbox": [ + 247, + 158, + 345, + 171 + ], + "type": "text", + "content": "Shubham Tulsiani\\*" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 406, + 143, + 452, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 143, + 452, + 156 + ], + "spans": [ + { + "bbox": [ + 406, + 143, + 452, + 156 + ], + "type": "text", + "content": "Hao Tan3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 476, + 143, + 539, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 476, + 143, + 539, + 157 + ], + "spans": [ + { + "bbox": [ + 476, + 143, + 539, + 157 + ], + "type": "text", + "content": "Zexiang " + }, + { + "bbox": [ + 476, + 143, + 539, + 157 + ], + "type": "inline_equation", + "content": "\\mathbf{X}\\mathbf{u}^{3}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 382, + 157, + 443, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 157, + 443, + 171 + ], + "spans": [ + { + "bbox": [ + 382, + 157, + 443, + 171 + ], + "type": "text", + "content": "Kai Zhang" + }, + { + "bbox": [ + 382, + 157, + 443, + 171 + ], + "type": "inline_equation", + "content": "^{3*}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 78, + 184, + 222, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 184, + 222, + 199 + ], + "spans": [ + { + "bbox": [ + 78, + 184, + 222, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 184, + 222, + 199 + ], + "type": "text", + "content": " Carnegie Mellon University" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 233, + 184, + 427, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 184, + 427, + 200 + ], + "spans": [ + { + "bbox": [ + 233, + 184, + 427, + 200 + ], + "type": "text", + "content": "2 Massachusetts Institute of Technology" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 440, + 184, + 529, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 440, + 184, + 529, + 198 + ], + "spans": [ + { + "bbox": [ + 440, + 184, + 529, + 198 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 440, + 184, + 529, + 198 + ], + "type": "text", + "content": " Adobe Research" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 229, + 216, + 376, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 216, + 376, + 227 + ], + "spans": [ + { + "bbox": [ + 229, + 216, + 376, + 227 + ], + "type": "text", + "content": "https://turbo-3d.github.io/" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 59, + 251, + 526, + 361 + ], + "blocks": [ + { + "bbox": [ + 59, + 251, + 526, + 361 + ], + "lines": [ + { + "bbox": [ + 59, + 251, + 526, + 361 + ], + "spans": [ + { + "bbox": [ + 59, + 251, + 526, + 361 + ], + "type": "image", + "image_path": "e517bb0932019afa3b6d064b2be84d827a893f280ce57e0ff3e0a18e77382d7f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 58, + 367, + 545, + 427 + ], + "blocks": [ + { + "bbox": [ + 58, + 367, + 545, + 427 + ], + "lines": [ + { + "bbox": [ + 58, + 367, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 58, + 367, + 545, + 427 + ], + "type": "image", + "image_path": "4279294ded69c7d28972ea4ec79bd89e461d7ade553c6f814fec3a8fa209a81c.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 67, + 430, + 226, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 430, + 226, + 450 + ], + "spans": [ + { + "bbox": [ + 67, + 430, + 226, + 450 + ], + "type": "text", + "content": "a panda wearing a chef's hat and kneading bread dough on a countertop" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 70, + 455, + 106, + 505 + ], + "blocks": [ + { + "bbox": [ + 70, + 455, + 106, + 505 + ], + "lines": [ + { + "bbox": [ + 70, + 455, + 106, + 505 + ], + "spans": [ + { + "bbox": [ + 70, + 455, + 106, + 505 + ], + "type": "image", + "image_path": "f94ab1b71cc78cd9b77f90ea0f79db391c63779992bee7d21d5870eacacb684d.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 511, + 187, + 521 + ], + "lines": [ + { + "bbox": [ + 97, + 511, + 187, + 521 + ], + "spans": [ + { + "bbox": [ + 97, + 511, + 187, + 521 + ], + "type": "text", + "content": "a plush toy of a cordi nurse" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 126, + 455, + 162, + 505 + ], + "blocks": [ + { + "bbox": [ + 126, + 455, + 162, + 505 + ], + "lines": [ + { + "bbox": [ + 126, + 455, + 162, + 505 + ], + "spans": [ + { + "bbox": [ + 126, + 455, + 162, + 505 + ], + "type": "image", + "image_path": "36163932441e8a06763f9c45ba4180e6edfaf2c25c17f8ee2ccb40f610dea4c8.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 182, + 456, + 220, + 505 + ], + "blocks": [ + { + "bbox": [ + 182, + 456, + 220, + 505 + ], + "lines": [ + { + "bbox": [ + 182, + 456, + 220, + 505 + ], + "spans": [ + { + "bbox": [ + 182, + 456, + 220, + 505 + ], + "type": "image", + "image_path": "b79dc2f7e7464eb5107d954e23a323ae8e8c7ddb1ea41bf975898202dbb59a0b.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 233, + 453, + 271, + 505 + ], + "blocks": [ + { + "bbox": [ + 233, + 453, + 271, + 505 + ], + "lines": [ + { + "bbox": [ + 233, + 453, + 271, + 505 + ], + "spans": [ + { + "bbox": [ + 233, + 453, + 271, + 505 + ], + "type": "image", + "image_path": "7f0e3292aa9bb02f38ad181b9f4635d07e68a49c733c7c9e53a2189927899fb0.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 256, + 512, + 334, + 521 + ], + "lines": [ + { + "bbox": [ + 256, + 512, + 334, + 521 + ], + "spans": [ + { + "bbox": [ + 256, + 512, + 334, + 521 + ], + "type": "text", + "content": "a raccoon stealing a pie" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 284, + 453, + 329, + 506 + ], + "blocks": [ + { + "bbox": [ + 284, + 453, + 329, + 506 + ], + "lines": [ + { + "bbox": [ + 284, + 453, + 329, + 506 + ], + "spans": [ + { + "bbox": [ + 284, + 453, + 329, + 506 + ], + "type": "image", + "image_path": "8ceb6de2b254accb7f27384fe948e4c52ab759108ac308def4abd800a9043704.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 339, + 453, + 383, + 507 + ], + "blocks": [ + { + "bbox": [ + 339, + 453, + 383, + 507 + ], + "lines": [ + { + "bbox": [ + 339, + 453, + 383, + 507 + ], + "spans": [ + { + "bbox": [ + 339, + 453, + 383, + 507 + ], + "type": "image", + "image_path": "dc7e2cfa88f59aa1a7f182cad879b1370cd4e5b66f666f45ceb9b9206faaf147.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 394, + 451, + 435, + 503 + ], + "blocks": [ + { + "bbox": [ + 394, + 451, + 435, + 503 + ], + "lines": [ + { + "bbox": [ + 394, + 451, + 435, + 503 + ], + "spans": [ + { + "bbox": [ + 394, + 451, + 435, + 503 + ], + "type": "image", + "image_path": "6c7e4a00eb82f0144894ba5fc51e4b6184341526b0bccdda3c552ab82b4eb07f.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 507, + 535, + 527 + ], + "lines": [ + { + "bbox": [ + 399, + 507, + 535, + 527 + ], + "spans": [ + { + "bbox": [ + 399, + 507, + 535, + 527 + ], + "type": "text", + "content": "a badger wearing a party hat and blowing out birthday candles on a cake" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 451, + 448, + 493, + 502 + ], + "blocks": [ + { + "bbox": [ + 451, + 448, + 493, + 502 + ], + "lines": [ + { + "bbox": [ + 451, + 448, + 493, + 502 + ], + "spans": [ + { + "bbox": [ + 451, + 448, + 493, + 502 + ], + "type": "image", + "image_path": "d31a75e94589ea056799ae129d2e46eb6d1fd259d5f602602c067be7629e7843.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 506, + 448, + 544, + 497 + ], + "blocks": [ + { + "bbox": [ + 506, + 448, + 544, + 497 + ], + "lines": [ + { + "bbox": [ + 506, + 448, + 544, + 497 + ], + "spans": [ + { + "bbox": [ + 506, + 448, + 544, + 497 + ], + "type": "image", + "image_path": "ab0e4bd55bb548d4f610b56d92c6e74be0c0335d8c6ab5862346601347ef4245.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 63, + 534, + 106, + 583 + ], + "blocks": [ + { + "bbox": [ + 63, + 534, + 106, + 583 + ], + "lines": [ + { + "bbox": [ + 63, + 534, + 106, + 583 + ], + "spans": [ + { + "bbox": [ + 63, + 534, + 106, + 583 + ], + "type": "image", + "image_path": "74cedcfe912707107c8fd02329a4dc14d8f4bf46094c60cb960f6f313870a634.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 593, + 194, + 603 + ], + "lines": [ + { + "bbox": [ + 99, + 593, + 194, + 603 + ], + "spans": [ + { + "bbox": [ + 99, + 593, + 194, + 603 + ], + "type": "text", + "content": "a cake in the shape of a train" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 114, + 536, + 166, + 584 + ], + "blocks": [ + { + "bbox": [ + 114, + 536, + 166, + 584 + ], + "lines": [ + { + "bbox": [ + 114, + 536, + 166, + 584 + ], + "spans": [ + { + "bbox": [ + 114, + 536, + 166, + 584 + ], + "type": "image", + "image_path": "71fea974c24db610363a01e4cd0921b13bbb7c0c4f1fc2190549b3e3ed07cd22.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 635, + 555, + 680 + ], + "lines": [ + { + "bbox": [ + 55, + 635, + 555, + 680 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 555, + 680 + ], + "type": "text", + "content": "Figure 1. Overview of our Turbo3D text-to-3D system. Turbo3D generates high-quality 3D Gaussian Splatting (3DGS) assets from user prompts in less than 1 second on a single A100 GPU. It's a two-stage pipeline consisting of a highly efficient latent-space few-step multi-view (MV) generator and single-step MV reconstructor. Note that we visualize latents as RGB images and 3DGS assets as point clouds in the pipeline figure for clarity." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 172, + 534, + 224, + 583 + ], + "blocks": [ + { + "bbox": [ + 172, + 534, + 224, + 583 + ], + "lines": [ + { + "bbox": [ + 172, + 534, + 224, + 583 + ], + "spans": [ + { + "bbox": [ + 172, + 534, + 224, + 583 + ], + "type": "image", + "image_path": "9c8c1f7112d652362c2dca86db6a3f6844c498f0d1b27f02f842287aacf0ff29.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 585, + 397, + 624 + ], + "lines": [ + { + "bbox": [ + 227, + 585, + 397, + 624 + ], + "spans": [ + { + "bbox": [ + 227, + 585, + 397, + 624 + ], + "type": "text", + "content": "Campsite with a tent on grassy ground, a backpack against it, a campfire surrounded by stones. A sleeping bag partially unrolled next to the tent, a lantern hanging from a tree branch" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 227, + 529, + 278, + 580 + ], + "blocks": [ + { + "bbox": [ + 227, + 529, + 278, + 580 + ], + "lines": [ + { + "bbox": [ + 227, + 529, + 278, + 580 + ], + "spans": [ + { + "bbox": [ + 227, + 529, + 278, + 580 + ], + "type": "image", + "image_path": "2c23987bc45820a6509323818894aaee7523488f07e8f198e65441ee0d98c008.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 279, + 529, + 337, + 582 + ], + "blocks": [ + { + "bbox": [ + 279, + 529, + 337, + 582 + ], + "lines": [ + { + "bbox": [ + 279, + 529, + 337, + 582 + ], + "spans": [ + { + "bbox": [ + 279, + 529, + 337, + 582 + ], + "type": "image", + "image_path": "a0460c86b4c0cb4a7512daf422b0239c7d3ff89010d1bd9c4c1d7865d45a18a1.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 338, + 535, + 391, + 581 + ], + "blocks": [ + { + "bbox": [ + 338, + 535, + 391, + 581 + ], + "lines": [ + { + "bbox": [ + 338, + 535, + 391, + 581 + ], + "spans": [ + { + "bbox": [ + 338, + 535, + 391, + 581 + ], + "type": "image", + "image_path": "f00753f7bb123a1f3dcc94411851aa062787564c4df60e4beec670c909478467.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 398, + 536, + 445, + 583 + ], + "blocks": [ + { + "bbox": [ + 398, + 536, + 445, + 583 + ], + "lines": [ + { + "bbox": [ + 398, + 536, + 445, + 583 + ], + "spans": [ + { + "bbox": [ + 398, + 536, + 445, + 583 + ], + "type": "image", + "image_path": "32e803b35fabdcae2a9691bf90b9badb9ef65ba190ea5addaa0d36edf28c4ecf.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 426, + 594, + 521, + 605 + ], + "lines": [ + { + "bbox": [ + 426, + 594, + 521, + 605 + ], + "spans": [ + { + "bbox": [ + 426, + 594, + 521, + 605 + ], + "type": "text", + "content": "a lion reading the newspaper" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 447, + 537, + 500, + 583 + ], + "blocks": [ + { + "bbox": [ + 447, + 537, + 500, + 583 + ], + "lines": [ + { + "bbox": [ + 447, + 537, + 500, + 583 + ], + "spans": [ + { + "bbox": [ + 447, + 537, + 500, + 583 + ], + "type": "image", + "image_path": "c586f16699d00a7d3b144bcd03349b9156805d01276c433c0ad4a2c6e47d2618.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 502, + 537, + 553, + 583 + ], + "blocks": [ + { + "bbox": [ + 502, + 537, + 553, + 583 + ], + "lines": [ + { + "bbox": [ + 502, + 537, + 553, + 583 + ], + "spans": [ + { + "bbox": [ + 502, + 537, + 553, + 583 + ], + "type": "image", + "image_path": "9ac1abafbf1804825eeeba62e9044c45142c8404de4632205e99720ac9f4ef35.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 711, + 122, + 721 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 711, + 122, + 721 + ], + "spans": [ + { + "bbox": [ + 67, + 711, + 122, + 721 + ], + "type": "text", + "content": "*Equal advising." + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 293, + 753, + 317, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 761 + ], + "type": "text", + "content": "23668" + } + ] + } + ], + "index": 45 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 152, + 72, + 200, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 72, + 200, + 85 + ], + "spans": [ + { + "bbox": [ + 152, + 72, + 200, + 85 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 297, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 297, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 297, + 277 + ], + "type": "text", + "content": "We present Turbo3D, an ultra-fast text-to-3D system capable of generating high-quality Gaussian splatting assets in under one second. Turbo3D employs a rapid 4-step, 4-view diffusion generator and an efficient feed-forward Gaussian reconstructor, both operating in latent space, as shown in Fig. 1. The 4-step, 4-view generator is a student model distilled through a novel Dual-Teacher approach, which encourages the student to learn view consistency from a multiview teacher and photo-realism from a single-view teacher. By shifting the Gaussian reconstructor's inputs from pixel space to latent space, we eliminate the extra image decoding time and halve the transformer sequence length for maximum efficiency. Our method demonstrates superior 3D generation results compared to previous baselines, while operating in a fraction of their runtime." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 298, + 136, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 136, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 136, + 312 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 319, + 296, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 474 + ], + "type": "text", + "content": "The recent advances in image generative models allow users to generate detailed outputs from just a text prompt. While initial denoising diffusion-based methods [7, 34, 36, 37, 44, 50] enabled impressive photo-realistic generation, recent techniques [19, 22, 23, 39, 45, 60, 61] have significantly improved inference efficiency of such models, allowing high-fidelity generation in the blink of an eye. Unfortunately, these advances in generative 2D modeling methods have not yet been matched in the 3D domain, where ultra-fast realistic 3D generation remains a challenge. In this work, we seek to bridge this gap, and present Turbo3D, a text-to-3D generative model that can synthesize detailed 3D outputs in a fraction of a second." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 475, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 297, + 714 + ], + "type": "text", + "content": "The existing approaches for text-based 3D inference can be categorized as either generative [10, 17, 24, 30, 57, 63] or optimization-driven [18, 32, 55]. The latter class of methods optimize 3D representations by 'distilling' pretrained 2D diffusion models [32]. While this approach can yield decent 3D outputs, it is highly inefficient, typically requiring several minutes or even hours to output a single 3D representation. The alternative paradigm is to learn a generative model that directly outputs 3D representations. While initial methods [10, 28] investigated representations such as point clouds and SDFs, the limited availability of 3D data restricted the generation quality. Recent approaches [9, 17, 24, 40, 41] have instead advocated for learning generative models of multi-view images (followed by deterministic 3D reconstruction), as these can be initialized from pre-trained 2D generative models. While these methods have resulted in impressive generations, the multiview finetuning on synthetic data does inhibit their generation quality. More crucially, their inference efficiency is restricted by the iterative denoising process required for the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 476, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 476, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 476, + 84 + ], + "type": "text", + "content": "text-conditioned multi-view generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 85, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 85, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 85, + 555, + 407 + ], + "type": "text", + "content": "In this work, we follow the paradigm of text-to-3D via multi-view generation, and aim to improve the efficiency of the underlying components to enable ultra-fast generation, while enhancing the fidelity of the generated outputs. Inspired by the recent progress in reducing inference time by distilling 2D diffusion models into one-step or few-step generators [38, 39, 45, 61], we adapt these for multi-view 3D generation. We first train a many-step text-to-multiview diffusion model and then distill it into a much faster 4-step generator using the distribution matching distillation (DMD) loss [61]. However, we find that this process leads to a significant quality degradation, as it fails to capture the full range of modes present in the multi-view teacher model. To overcome this, we propose to extend the DMD pipeline to incorporate another single-view teacher - a 2D denoising diffusion model trained on a large set of high quality aesthetic images. Our few-step multi-view generator is thus trained with a dual-teacher distillation approach, where the multi-view DMD loss helps our model learn multi-view consistency, and the single-view DMD loss ensures high-fidelity outputs. To further improve the 3D generation efficiency, we note that our multi-view generator outputs latent representations (and not pixels) for the multi-view images. We build on this insight to adapt a prior multi-view to 3D reconstruction approach [62] to instead consume multi-view latents as input, and show that this improves the reconstruction efficiency without any performance loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 407, + 556, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 556, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 556, + 504 + ], + "type": "text", + "content": "Our overall system thus combines an efficient few-step multi-view generator with a multi-view latent to 3D model to enable ultra fast 3D generation. We train our model on the subset of Objaverse dataset [3, 4], which contains about 400k instances with Cap3D text captions [25]. We show that our proposed Turbo3D is able to produce high-quality 3D assets in less than one second (see Fig. 1), and also achieves comparable quality with previous state-of-the-art." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 514, + 402, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 514, + 402, + 526 + ], + "spans": [ + { + "bbox": [ + 314, + 514, + 402, + 526 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 533, + 556, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 654 + ], + "type": "text", + "content": "In this section, we discuss the closely related prior feedforward 3D generation methods. Optimization-based methods [18, 32] is out of this work's scope. We review feedforward 3D generative models in the following three categories: 1) methods that directly generate full 3D representation encoding geometry and appearance; 2) methods that generate shapes and then generate the textures; 3) methods that generate multi-view (MV) images followed by reconstruction. We also review the diffusion distillation literature which our method is built upon." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 660, + 547, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 547, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 547, + 673 + ], + "type": "text", + "content": "2.1. Directly Genereating Full 3D Representation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Several prior methods [1, 10, 30, 46, 48, 54, 58] have been introduced that directly generate 3D representations encoding both geometry and texture, e.g., implicit fields [10]," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "type": "text", + "content": "23669" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "content": "point clouds [30], and triplanes [54]. This line of work typically have to preprocess the source 3D data (meshes or multi-view images) into the target representations for generative models in a lossy fashion, hence limiting their quality and scalability." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 294, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 294, + 239 + ], + "type": "text", + "content": "Alternative methods [1, 46, 48, 58] have been proposed to get rid of the lossy preprocessing. They implicitly bake a 3D generative process into a multi-view diffusion framework, creating a single-stage 3D generative model. However, trained only on 3D data, they suffer from the issue of limited generalization, lower pixel quality and reduced understanding of complex text prompts, compared with methods leveraging powerful pretrained text-to-image models [17, 21]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 240, + 294, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 294, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 294, + 288 + ], + "type": "text", + "content": "Our method leverages the strong priors in a pretrained image generative model, but we go a step further by distilling the slow teacher model into a fast student generator for ultra-fast text-to-3D." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 297, + 265, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 265, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 265, + 309 + ], + "type": "text", + "content": "2.2. Shape Generation + Texture Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 316, + 294, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 294, + 459 + ], + "type": "text", + "content": "Recent advancements in 3D content creation have introduced innovative methods [42, 57, 63] that focus on generating high-quality 3D shapes first, followed by generative texturing [2, 35], rather than creating both simultaneously. CLAY [63] offers a framework using a multiresolution Variational Autoencoder (VAE) and a latent Diffusion Transformer (DiT) to initially create detailed 3D geometries from inputs like text and images, before applying high-resolution physically-based rendering (PBR) textures. Direct3D [57] enhances scalability in image-to-3D generation by improving the performance of 3D VAE and DiT for generating 3D shapes, subsequently adding textures." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 460, + 294, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 294, + 542 + ], + "type": "text", + "content": "However, besides the challenge of learning a compact latent space suitable for generation, these methods are also usually slow as a result of the two diffusion models - one for shapes and the other for textures. In contrast, our Turbo3D only have one diffusion model for generating multi-views and is fast once distilled, while the reconstructor is deterministic and hence fast." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 552, + 256, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 256, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 256, + 563 + ], + "type": "text", + "content": "2.3. MV Generation + MV Reconstruction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 294, + 713 + ], + "type": "text", + "content": "A large body of the 3D generation work [17, 20, 24, 41] focus on leveraging multi-view generation and reconstruction to enhance quality and efficiency. Instant3D [17] offers a fast method for creating high-quality 3D assets from text prompts by combining sparse-view generation with a transformer-based reconstructor [8, 53, 56, 62], significantly reducing inference time. MVDream [41] employs a multi-view diffusion model to improve consistency and stability in 3D generation, integrating 2D and 3D data. SyncDreamer [24] enhances multiview-consistency from a single-view image using a 3D-aware feature attention mechanism. One-2-3-45 [20] introduces an efficient ap" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "proach to single image 3D reconstruction without extensive optimization, producing consistent 3D meshes. SV3D [51] utilizes a latent video diffusion model for novel multi-view synthesis, incorporating explicit camera control to improve 3D reconstruction quality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 133, + 553, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 553, + 264 + ], + "type": "text", + "content": "We follow the same approach of reconstructing generated multi-views to create 3D assets in this work, but we focus on improving the efficiency of such systems. With our novel Dual-Teacher Distillation and Latent GS-LRM components, our Turbo3D manages to be at least an order of magnitude faster than these baselines while maintaining competitive quality. Concurrent work GECO [52] also use diffusion distillation to speed up image-to-3D. However, we differ in our dual-teacher distillation design with focus on text-to-3D. Our pipeline is also simpler and avoids the cumbersome mesh reconstructions for 3D distillation in GECO." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 274, + 435, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 274, + 435, + 285 + ], + "spans": [ + { + "bbox": [ + 314, + 274, + 435, + 285 + ], + "type": "text", + "content": "2.4. Diffusion Distillation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 292, + 553, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 553, + 483 + ], + "type": "text", + "content": "Recent progress in diffusion models have focused on improving the efficiency of image generation by reducing the number of sampling steps required, leading to the development of several innovative distillation techniques [26, 27, 29, 38, 39, 45, 59-61]. Methods like Improved Rectified Flows [15] and InstaFlow [23] straighten the ODE trajectories, making them easier to approximate with a one-step student model. Consistency Models [45] train student generators to map any point on the teacher's ODE trajectory to a consistent target, enabling one-step and few-step image generation. Distribution Matching Distillation (DMD) [61] trains a one-step generator by minimizing the reverse KL divergence between the data distribution and the generator's output distribution [5, 26, 55]. Building on this, DMD2 [60] integrates GAN losses [6] and extends the method to multi-step generators, further enhancing generation quality." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 483, + 553, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 553, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 553, + 579 + ], + "type": "text", + "content": "However, most of existing approaches focus on distilling pretrained 2D image diffusion models. In our work, we adopt the popular DMD [61] approach and extend it to the multi-view domain. Distilling multi-view models presents unique challenges, such as increased mode collapse due to fine-tuning and distillation processes. To address this issue, we introduce a novel dual-teacher distillation technique, enabling swift, photorealistic multi-view generation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 592, + 392, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 592, + 392, + 605 + ], + "spans": [ + { + "bbox": [ + 314, + 592, + 392, + 605 + ], + "type": "text", + "content": "3. Background" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 612, + 466, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 466, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 466, + 624 + ], + "type": "text", + "content": "3.1. Multi-view Diffusion Model" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 631, + 553, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 631, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 313, + 631, + 553, + 690 + ], + "type": "text", + "content": "Diffusion models generate data from noise by reversing a forward noising process. Given a data sample " + }, + { + "bbox": [ + 313, + 631, + 553, + 690 + ], + "type": "inline_equation", + "content": "x_0 \\sim p(x_0)" + }, + { + "bbox": [ + 313, + 631, + 553, + 690 + ], + "type": "text", + "content": " from the data distribution, the forward process progressively adds Gaussian noise over " + }, + { + "bbox": [ + 313, + 631, + 553, + 690 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 631, + 553, + 690 + ], + "type": "text", + "content": " timesteps. At timestep " + }, + { + "bbox": [ + 313, + 631, + 553, + 690 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 631, + 553, + 690 + ], + "type": "text", + "content": ", the noised distribution conditioned on " + }, + { + "bbox": [ + 313, + 631, + 553, + 690 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 631, + 553, + 690 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 372, + 700, + 553, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 700, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 372, + 700, + 553, + 714 + ], + "type": "interline_equation", + "content": "q \\left(x _ {t} \\mid x _ {0}\\right) = \\mathcal {N} \\left(x _ {t}; \\alpha_ {t} x _ {0}, \\beta_ {t} ^ {2} \\mathbf {I}\\right), \\tag {1}", + "image_path": "8c607de8800988cd6d0e712dd737266141f33731ffc8782dfc6f4a6e0ba42dc6.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "text", + "content": "23670" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": " are timestep-dependent constant that are specified by the noise schedule [12, 44]. The diffusion model " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": " is trained to reverse this noising process by learning to predict the added Gaussian noise " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 118, + 125, + 295, + 140 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 125, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 118, + 125, + 295, + 140 + ], + "type": "interline_equation", + "content": "L (\\theta) = \\mathbb {E} | | \\epsilon - \\epsilon_ {\\theta} (x _ {t}, t) | | ^ {2}. \\tag {2}", + "image_path": "e88b983f15d289e37fc4f96991971f1c467b5d455a3a900ad85d9d14db38e212.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 144, + 296, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 296, + 217 + ], + "type": "text", + "content": "There exists other formulations, in which the diffusion model learns to predict the clean input " + }, + { + "bbox": [ + 55, + 144, + 296, + 217 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 144, + 296, + 217 + ], + "type": "text", + "content": " directly [11], or a combination of " + }, + { + "bbox": [ + 55, + 144, + 296, + 217 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 144, + 296, + 217 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 144, + 296, + 217 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 144, + 296, + 217 + ], + "type": "text", + "content": " [38]. Regardless of the specific prediction target, the outputs of these models can be related to the score function, which is the gradient of the log probability density of the data distribution [44]:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 220, + 295, + 252 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 220, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 58, + 220, + 295, + 252 + ], + "type": "interline_equation", + "content": "s _ {\\theta} \\left(x _ {t}, t\\right) = \\nabla_ {x _ {t}} \\log p \\left(x _ {t}\\right) = - \\frac {\\epsilon_ {\\theta} \\left(x _ {t} , t\\right)}{\\beta_ {t}} = - \\frac {x _ {t} - \\alpha_ {t} x _ {\\theta} \\left(x _ {t} , t\\right)}{\\beta_ {t} ^ {2}} \\tag {3}", + "image_path": "63085148322954c8a9937c1d041b82ba8381be50db87b23c9fe9700113ad6488.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 253, + 296, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 296, + 300 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 253, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 253, + 296, + 300 + ], + "type": "text", + "content": " [7] and " + }, + { + "bbox": [ + 55, + 253, + 296, + 300 + ], + "type": "inline_equation", + "content": "x_{\\theta}" + }, + { + "bbox": [ + 55, + 253, + 296, + 300 + ], + "type": "text", + "content": " [11] represents noise and data prediction diffusion models, respectively. In our paper, we adopt the noise prediction scheme but our method generalizes to any formulations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 300, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 396 + ], + "type": "text", + "content": "In multi-view diffusion models, the generation of a 3D scene is conditioned on a text prompt, enabling the joint denoising of multiple views to produce a set of 3D-consistent output images [17, 41]. Specifically, given a text prompt " + }, + { + "bbox": [ + 55, + 300, + 296, + 396 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 300, + 296, + 396 + ], + "type": "text", + "content": " and a set of multi-view images " + }, + { + "bbox": [ + 55, + 300, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\{x^i\\}_{i=1}^K" + }, + { + "bbox": [ + 55, + 300, + 296, + 396 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 300, + 296, + 396 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 300, + 296, + 396 + ], + "type": "text", + "content": " is the number of view, the multi-view diffusion model learns to predict the added noise across all views simultaneously. The training objective is formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 119, + 400, + 295, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 400, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 119, + 400, + 295, + 415 + ], + "type": "interline_equation", + "content": "\\mathbb {E} | | \\epsilon - \\epsilon_ {\\theta} \\left(\\left\\{x _ {t} ^ {i} \\right\\} _ {i = 1} ^ {K}, t, c\\right) | | ^ {2}. \\tag {4}", + "image_path": "b0561762f2b9f9f464c0b326215d7ba67341a88586fdb6d0f0de9248d96ec766.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 420, + 296, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 296, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 296, + 516 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 420, + 296, + 516 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 420, + 296, + 516 + ], + "type": "text", + "content": " represents independent Gaussian noise with the same variance applied to each view. At inference time, the generation starts from a set of fully noisy multi-view images " + }, + { + "bbox": [ + 55, + 420, + 296, + 516 + ], + "type": "inline_equation", + "content": "\\{x_{T}^{i}\\}_{i = 1}^{K}" + }, + { + "bbox": [ + 55, + 420, + 296, + 516 + ], + "type": "text", + "content": " sampled from a standard Gaussian distribution. The multi-view diffusion model iteratively generate a sequence of cleaner multi-view images. Various diffusion samplers [7, 43] can be employed during this process to optimize generation speed and quality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 522, + 237, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 237, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 237, + 534 + ], + "type": "text", + "content": "3.2. Distribution Matching Distillation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "content": "Distribution Matching Distillation (DMD) is a widely used diffusion distillation technique that converts teacher diffusion models into a student generator requiring significantly fewer sampling steps [60, 61]. The DMD approach trains the student generator " + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "content": " by minimizing an approximate reverse KL divergence between the smoothed student's output distribution (denoted as " + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{fake}}" + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "content": ") and the smoothed data distribution (denoted as " + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{real}}" + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "content": "):" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 639, + 295, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 639, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 67, + 639, + 295, + 663 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {D M D}} (\\theta) = D _ {\\mathrm {K L}} \\left(p _ {\\text {f a k e}} \\| p _ {\\text {r e a l}}\\right) = \\mathbb {E} _ {x, t} \\left(\\log \\left(\\frac {p _ {\\text {f a k e}} \\left(x _ {t}\\right)}{p _ {\\text {r e a l}} \\left(x _ {t}\\right)}\\right)\\right). \\tag {5}", + "image_path": "4209209d76d3fe87b5af4a227eca72bada4a7192981b285c7b0444d06e053264.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "content": "Let the score functions of the data distribution and the student's output distribution be denoted as " + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{real}}" + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{fake}}" + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "content": ", respectively. The gradient of this KL divergence can be effectively approximated by the difference between these two" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 73, + 382, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 73, + 382, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 73, + 382, + 83 + ], + "type": "text", + "content": "score functions:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 87, + 553, + 164 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 87, + 553, + 164 + ], + "spans": [ + { + "bbox": [ + 315, + 87, + 553, + 164 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {\\theta} L _ {\\mathrm {D M D}} (\\theta) \\approx \\mathbb {E} \\left[ - \\int \\left(s _ {\\text {r e a l}} \\left(F \\left(G _ {\\theta} (\\epsilon), t\\right), t\\right) \\right. \\right. \\\\ \\left. - s _ {\\text {f a k e}} \\left(F \\left(G _ {\\theta} (\\epsilon), t\\right), t\\right)\\right) \\frac {d G _ {\\theta} (\\epsilon)}{d \\theta} d \\epsilon \\Bigg ] \\tag {6} \\\\ \\end{array}", + "image_path": "f02cc15ee33fb1fd66a27de990a09e704a6a904607d0e03b5c436d63c81850ac.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 170, + 555, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 170, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 555, + 266 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 170, + 555, + 266 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 170, + 555, + 266 + ], + "type": "text", + "content": " represents the forward diffusion process defined in Eq. 1. The student generator can be adapted for a multi-step generation setting by replacing the pure noise input " + }, + { + "bbox": [ + 313, + 170, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 170, + 555, + 266 + ], + "type": "text", + "content": " with a partially noisy image " + }, + { + "bbox": [ + 313, + 170, + 555, + 266 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 170, + 555, + 266 + ], + "type": "text", + "content": " [60]. During training, the data distribution's score function is initialized from the teacher diffusion model and kept fixed, while the student's output distribution score function is dynamically trained using the student's output and a denoising loss (Eq. 4)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 275, + 370, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 275, + 370, + 288 + ], + "spans": [ + { + "bbox": [ + 314, + 275, + 370, + 288 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 295, + 555, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 555, + 403 + ], + "type": "text", + "content": "In this section, we provide the technical details of our Turbo3D text-to-3D system, which features a highly efficient multi-view (MV) generator and reconstructor. We begin by describing our novel DualTeacher Distillation approach; it creates a rapid MV generator by jointly distilling knowledge from both a multi-view (MV) teacher and a single-view (SV) teacher. Following this, we discuss the latent-space GS-LRM that instantly lifts the generated MV latents into high-quality 3D Gaussians." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 409, + 536, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 536, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 536, + 421 + ], + "type": "text", + "content": "4.1. Dual-teacher Distillation for MV Diffusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 426, + 554, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 554, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 554, + 486 + ], + "type": "text", + "content": "Leveraging a multi-step multi-view diffusion in a text-to-3D generation process can be inefficient due to repeated evaluations of the diffusion denoiser in the sampling process. To speed this up, one approach is to use diffusion distillation [60, 61] to train a single-step or few-step MV generator." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 487, + 554, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 554, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 554, + 641 + ], + "type": "text", + "content": "However, we find that naively applying diffusion distillation methods to a MV teacher can cause the student model to generate overly simplistic and cartoonish appearance, which closely resembles the 3D Objverse dataset used during MV teacher finetuning and distillation [3] (shown in Fig. 6). We call this phenomenon 'compounding mode collapse'; this happens because both finetuning and distillation sacrifices generation diversity for efficiency. As the MV teacher is already biased towards synthetic Objverse-style appearance, further distilling it will have compound effect of locking the distilled generator in the mode of Objverse data that are far from the modes of photorealistic natural images." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "text", + "content": "To address this issue, we propose to use dual teachers in the distillation process: one MV teacher to teach the student model about multi-view consistency, and one SV teacher to teach about each views' photo-realism. We illustrate this Dual-teacher Distillation algorithm in Fig. 2. Concretely, this is formulated as:" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "text", + "content": "23671" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 553, + 232 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 232 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 232 + ], + "type": "image", + "image_path": "b6c8af2b5b0d9295f570d9c4e21f14d471ce453b0f87792ba39f1cdcdc104285.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 240, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 54, + 240, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 54, + 240, + 555, + 319 + ], + "type": "text", + "content": "Figure 2. Dual-teacher distillation framework in our Turbo3D. Note that latents are visualized as RGB images for clarity. We aim to distill a multi-step multi-view teacher generator (right, green) into a few-step multi-view generator (left, blue). Our few-step MV student generator is conditioned on Plücker embeddings for better 3D awareness. Similar to [60], we optimize the student generator using distribution matching objective (DMD loss) and train the fake score function to model the distribution of samples produced by the student generator. In particular, we integrate two teacher models: multi-view teacher and single-view (SV) teacher to enhance both multi-view consistency and photorealism. The MV score functions take a set of images of one object as input and calculate the MV DMD loss, while the SV score functions treat each image separately and calculate the SV DMD loss." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 356, + 294, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 356, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 63, + 356, + 294, + 407 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {\\mathrm {D M D}} ^ {\\text {D u a l}} (\\theta) = D _ {\\mathrm {K L}} \\left(p _ {\\text {f a k e}} \\left(\\left\\{x _ {t} ^ {i} \\right\\} _ {i = 1} ^ {K}\\right) \\mid \\mid p _ {\\text {r e a l}} ^ {\\mathrm {M V}} \\left(\\left\\{x _ {t} ^ {i} \\right\\} _ {i = 1} ^ {K}\\right)\\right) \\\\ + \\lambda \\cdot \\frac {1}{K} \\sum_ {i = 1} ^ {K} D _ {\\mathrm {K L}} \\left(p _ {\\text {f a k e}} \\left(x _ {t} ^ {i}\\right) \\mid \\mid p _ {\\text {r e a l}} ^ {\\mathrm {S V}} \\left(x _ {t} ^ {i}\\right)\\right), \\tag {7} \\\\ \\end{array}", + "image_path": "276ee33fdb495d55d8417da4f349d85f37ef913354c22982acfaca81f1a2006f.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 413, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 413, + 296, + 510 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{real}}^{\\mathrm{MV}}(x_t), p_{\\mathrm{real}}^{\\mathrm{SV}}(x_t)" + }, + { + "bbox": [ + 55, + 413, + 296, + 510 + ], + "type": "text", + "content": " represent MV and SV teachers, respectively; " + }, + { + "bbox": [ + 55, + 413, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 413, + 296, + 510 + ], + "type": "text", + "content": " is the loss weight balancing the influence of MV and SV teachers on the distilled student model; " + }, + { + "bbox": [ + 55, + 413, + 296, + 510 + ], + "type": "inline_equation", + "content": "K = 4" + }, + { + "bbox": [ + 55, + 413, + 296, + 510 + ], + "type": "text", + "content": " is the number of views. We set " + }, + { + "bbox": [ + 55, + 413, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 413, + 296, + 510 + ], + "type": "text", + "content": " to 1 in our experiments. As demonstrated in Fig. 6, having the additional SV teacher in the distillation process effectively addresses the compound mode collapse issue, because it tries to pull each view to look like natural images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 516, + 268, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 268, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 268, + 529 + ], + "type": "text", + "content": "4.2. Latent GS-LRM for MV Reconstruction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 533, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 617 + ], + "type": "text", + "content": "To reconstruct 3D from the generated MV latents, one straightforward approach is to decode them into multi-view images, and then use the pixel-space GS-LRM [62] to produce the 3D Gaussians. However, such an approach can suffer from efficiency and memory issue when scaling to high resolution due to the poor performance of Conv2D operators in VAE decoder [13]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "We propose to skip the VAE decoding and directly input the generated MV latents to a latent GS-LRM for best efficiency. To train such a model, we supervise the reconstructed Gaussians with pixel-space novel-view rendering losses " + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\ell_{2}" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": " and perceptual losses as in [62]). We show that replacing pixel-space GS-LRM with a latent one does not affect the quality of generated assets in Tab. 3 and Fig. 3, while being faster." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 315, + 335, + 555, + 443 + ], + "blocks": [ + { + "bbox": [ + 315, + 335, + 555, + 443 + ], + "lines": [ + { + "bbox": [ + 315, + 335, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 315, + 335, + 555, + 443 + ], + "type": "image", + "image_path": "770fab4cf3e96e8ae624c6d3039b8f00cd3b24ae52269edd1657ee56df65a012.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 450, + 555, + 484 + ], + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 484 + ], + "type": "text", + "content": "Figure 3. We compare the renderings of pixel GS-LRM and latent GS-LRM. Latent GS-LRM achieves comparable reconstruction quality as pixel GS-LRM." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 506, + 395, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 395, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 395, + 520 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 528, + 555, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 555, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 555, + 588 + ], + "type": "text", + "content": "In this section, we first describe our experimental setup (Section 5.1). We then compare our method with state-of-the-art text-to-3D baselines (Section 5.2). Finally, we ablate each component of our framework to showcase their effectiveness (Section 5.3)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 598, + 430, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 430, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 430, + 612 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Datasets. We use the Objaverse dataset [3] to train both our multi-view generation model and multi-view reconstruction model. We scale the objects and center them to fit into a cube " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "[-1, 1]^3" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": ". For the generation task, we render the dataset at a fixed elevation (20 degrees) and 16 equidistant azimuths to achieve a good coverage of the object. We render using a field of view " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "50^o" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": " at a distance of 2.7 and uniform lighting. For the reconstruction task, we render 32 views ran-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "type": "text", + "content": "23672" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "domly placed around the object with a random distance in the range of [1.5, 2.8]. We render a total of " + }, + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "inline_equation", + "content": "730K" + }, + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": " objects." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 97, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 97, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 54, + 97, + 294, + 204 + ], + "type": "text", + "content": "Baselines. We adopt Instant3D [17] and LGM [47] as our baseline text-to-3D methods. However, since the field of fast text-to-3D is relatively underexplored, we also include recent state-of-the-art fast image-to-3D methods TripoSR [49] and SV3D [51] as baseline methods. For a fair comparison, we use a popular few-step text-to-image model Flux [14] to generate an input image first and then feed it to the image-to-3D models. The inference time of image-to-3D models is a summation of the two parts." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 205, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 296, + 361 + ], + "type": "text", + "content": "Metrics. We adopt the CLIP score [33] and VQA score [16] to assess the semantic alignment between the generated results and text prompts. We use 400 text prompts from DreamFusion [32] for evaluation. We generate one object for each prompt. For each generated 3D object, we render 10 random views and calculate the average CLIP score and VQA score between the rendered images and the input text. For inference time, we report it with all methods under the same image resolution of 256. Notably, some methods, e.g., Instant3D [17], only support a higher resolution of 512. For a fair comparison, we report their quantitative results under the resolution of 512 and only inference time on the resolution of 256." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 361, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 296, + 529 + ], + "type": "text", + "content": "Implementation Details. The whole pipeline of our method includes three training phases. We first train a multi-step multi-view diffusion model on the Objverse dataset, by fine-tuning an internal DiT [31] based text-to-image model. We train this model for 30k iterations with 32 80G A100 GPUs using a total batch size of 128 and a learning rate of " + }, + { + "bbox": [ + 56, + 361, + 296, + 529 + ], + "type": "inline_equation", + "content": "3e^{-5}" + }, + { + "bbox": [ + 56, + 361, + 296, + 529 + ], + "type": "text", + "content": ". Then we perform distillation to distill the multi-step multi-view generator into a few-step multi-view generator, which takes 10k iterations with 32 80G A100 GPUs with a global batch size of 128 and a learning rate of " + }, + { + "bbox": [ + 56, + 361, + 296, + 529 + ], + "type": "inline_equation", + "content": "5e^{-6}" + }, + { + "bbox": [ + 56, + 361, + 296, + 529 + ], + "type": "text", + "content": ". Finally, we train a reconstruction model - latent GS-LRM - from scratch, which takes 80k iterations with 32 80G A100 GPUs with a total batch size of 256 and a learning rate of " + }, + { + "bbox": [ + 56, + 361, + 296, + 529 + ], + "type": "inline_equation", + "content": "4e^{-4}" + }, + { + "bbox": [ + 56, + 361, + 296, + 529 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 540, + 211, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 211, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 211, + 552 + ], + "type": "text", + "content": "5.2. Evaluation against baselines" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 557, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 713 + ], + "type": "text", + "content": "Qualitative Comparisons. As shown in Fig. 4, our method generates significantly better results compared with LGM [47] and Instant3D [17]. In particular, LGM tends to produce simple 3D assets without geometry and texture details; it also lacks robustness, oftentimes generating low-quality 3D assets that are broken, or suffer from multi-face Janus problem [32], or do not closely follow provided text prompts. Instant3D is much more robust and able to produce plausible 3D assets most of the time. However, its performance on text-image coherence is limited compared with our Turbo3D. For example, in the first example, LGM's rendition appears oversimplified with indistinct egg and yolk boundaries, while Instant3D provides a" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 71, + 553, + 175 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 553, + 175 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 553, + 175 + ], + "type": "table", + "html": "
MethodClip Score ↑VQA Score ↑Inference Time ↓
TripoSR [49]23.850.571.19s
SV3D [51]24.920.6412.52s
Instant3D [17]26.230.6515.02s
LGM [47]24.730.586.56s
Turbo3D (Ours)27.610.760.35s
", + "image_path": "41278a1e6349626d5e2eb0fa3061778fcabd7a0f9bd15cc8a82fab3c4f80429b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 186, + 553, + 230 + ], + "lines": [ + { + "bbox": [ + 313, + 186, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 553, + 230 + ], + "type": "text", + "content": "Table 1. Comparison against state-of-the-art 3D generation methods. Our Turbo3D generates 3D assets with highest CLIP and VQA scores while using the least amount of time (benchmarked on a A100 GPU)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 233, + 555, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 233, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 555, + 365 + ], + "type": "text", + "content": "closer match but lacks fine details in the egg's structure and wood texture and fails to capture the concept of 'spill out'. For complex prompts that describes a scene with multiple objects (4th and 6th rows), LGM generates a Janus asset while Instant3D miss a lot of concepts like blanket, basket, wildflowers etc; our generations adhere to the prompts much more closely. These qualitative results highlight our Turbo3D's ability to generate high-quality, text-aligned 3D models with a superior level of detail, realism, and coherence across diverse and complex prompts, outperforming both LGM and Instant3D by a large margin." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 366, + 556, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 556, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 556, + 544 + ], + "type": "text", + "content": "Quantitative Comparisons. Tab. 1 presents a quantitative comparison of our proposed method, Turbo3D, against several state-of-the-art approaches, including TripoSR [49], SV3D [51], Instant3D [17], and LGM [47], on the text-to-3D task. Our proposed Turbo3D achieves the highest CLIP Score of 27.61 and VQA Score of 0.76, outperforming other methods by a significant margin in both quality metrics. In addition to the quality improvement, Turbo3D demonstrates remarkable efficiency with an inference time of only 0.35 seconds, substantially faster than competing methods. Although TripoSR is able to generate a 3D asset in only 1.19s, the quality of the generated results is highly limited. As a result, our proposed Turbo3D is able to achieve outstanding performance in terms of both quality and inference speed compared with state-of-the-art methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": "User Study. We run a user study by randomly selecting 80 text prompts and asking 56 users to make 1120 pairwise comparisons (users are shown an input text prompt and two generated 3D assets from two anonymous methods). We show results in Fig. 5. Since the quantitative results of TripoSR is not competitive with others, it is not included in the user study. Moreover, SV3D only outputs videos instead of 3D representations of objects, so it is hard to perform a fair comparison. Therefore, we compare our Turbo3D with LGM [47], Instant3D [17], and our MV Teacher. In particular, MV Teacher is the model which our Turbo3D gets distilled from. When compared to LGM, Turbo3D achieved a win rate of " + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "inline_equation", + "content": "89.8\\%" + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": ", with only " + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "inline_equation", + "content": "10.2\\%" + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": " of participants favoring LGM. Against Instant3D, Turbo3D also outperformed" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "text", + "content": "23673" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 158, + 76, + 267, + 129 + ], + "blocks": [ + { + "bbox": [ + 71, + 86, + 154, + 111 + ], + "lines": [ + { + "bbox": [ + 71, + 86, + 154, + 111 + ], + "spans": [ + { + "bbox": [ + 71, + 86, + 154, + 111 + ], + "type": "text", + "content": "A cracked egg with the yolk spilling out on a wooden table" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 158, + 76, + 267, + 129 + ], + "lines": [ + { + "bbox": [ + 158, + 76, + 267, + 129 + ], + "spans": [ + { + "bbox": [ + 158, + 76, + 267, + 129 + ], + "type": "image", + "image_path": "dd7c273064bb74e2de53156680c2486bd0fb7e7fb3a1e995c1bc9a9af9c02ea5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 159, + 131, + 269, + 186 + ], + "blocks": [ + { + "bbox": [ + 72, + 142, + 137, + 159 + ], + "lines": [ + { + "bbox": [ + 72, + 142, + 137, + 159 + ], + "spans": [ + { + "bbox": [ + 72, + 142, + 137, + 159 + ], + "type": "text", + "content": "An origami hippo in a river" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 159, + 131, + 269, + 186 + ], + "lines": [ + { + "bbox": [ + 159, + 131, + 269, + 186 + ], + "spans": [ + { + "bbox": [ + 159, + 131, + 269, + 186 + ], + "type": "image", + "image_path": "79b84fdfa282a3c47faf90b10148611bd13ba19afe4f359b98645d9cd136ad01.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 159, + 191, + 269, + 251 + ], + "blocks": [ + { + "bbox": [ + 71, + 210, + 141, + 227 + ], + "lines": [ + { + "bbox": [ + 71, + 210, + 141, + 227 + ], + "spans": [ + { + "bbox": [ + 71, + 210, + 141, + 227 + ], + "type": "text", + "content": "Wedding dress made of tentacles" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 159, + 191, + 269, + 251 + ], + "lines": [ + { + "bbox": [ + 159, + 191, + 269, + 251 + ], + "spans": [ + { + "bbox": [ + 159, + 191, + 269, + 251 + ], + "type": "image", + "image_path": "17f9a0dbb3f38c3f6bdbe582d1879bfa5bf825a056ddaf299cb146be85340ca9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 159, + 259, + 272, + 294 + ], + "blocks": [ + { + "bbox": [ + 71, + 253, + 145, + 304 + ], + "lines": [ + { + "bbox": [ + 71, + 253, + 145, + 304 + ], + "spans": [ + { + "bbox": [ + 71, + 253, + 145, + 304 + ], + "type": "text", + "content": "A cozy campfire surrounded by logs for sitting, with a small tent set up nearby. Add a pot hanging over the fire and a few scattered leaves" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 159, + 259, + 272, + 294 + ], + "lines": [ + { + "bbox": [ + 159, + 259, + 272, + 294 + ], + "spans": [ + { + "bbox": [ + 159, + 259, + 272, + 294 + ], + "type": "image", + "image_path": "731bf3ab30d05f460e110f61a89898a6040dee9fb6052404f82de3b7a08ee963.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 164, + 308, + 272, + 357 + ], + "blocks": [ + { + "bbox": [ + 72, + 319, + 138, + 336 + ], + "lines": [ + { + "bbox": [ + 72, + 319, + 138, + 336 + ], + "spans": [ + { + "bbox": [ + 72, + 319, + 138, + 336 + ], + "type": "text", + "content": "A chimpanzee looking through a telescope" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 164, + 308, + 272, + 357 + ], + "lines": [ + { + "bbox": [ + 164, + 308, + 272, + 357 + ], + "spans": [ + { + "bbox": [ + 164, + 308, + 272, + 357 + ], + "type": "image", + "image_path": "9d73cf11344764951ce8ad6a64202543f797d544ff427b3cd079dba0474f92c8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 164, + 371, + 274, + 403 + ], + "blocks": [ + { + "bbox": [ + 71, + 356, + 153, + 424 + ], + "lines": [ + { + "bbox": [ + 71, + 356, + 153, + 424 + ], + "spans": [ + { + "bbox": [ + 71, + 356, + 153, + 424 + ], + "type": "text", + "content": "Picnic scene with a checkered blanket on grass, a wicker basket with sandwiches and fruit, a bottle of lemonade with glasses. An open book, a small bouquet of wildflowers on the blanket" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 164, + 371, + 274, + 403 + ], + "lines": [ + { + "bbox": [ + 164, + 371, + 274, + 403 + ], + "spans": [ + { + "bbox": [ + 164, + 371, + 274, + 403 + ], + "type": "image", + "image_path": "3f1aa7a34d3157c18aa86fcb26e704a7cae3452bb701fb0face8ccad5618fb9e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 416, + 243, + 427 + ], + "lines": [ + { + "bbox": [ + 192, + 416, + 243, + 427 + ], + "spans": [ + { + "bbox": [ + 192, + 416, + 243, + 427 + ], + "type": "text", + "content": "LGM (6.56s)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 277, + 74, + 397, + 128 + ], + "blocks": [ + { + "bbox": [ + 277, + 74, + 397, + 128 + ], + "lines": [ + { + "bbox": [ + 277, + 74, + 397, + 128 + ], + "spans": [ + { + "bbox": [ + 277, + 74, + 397, + 128 + ], + "type": "image", + "image_path": "25a0fcdb3fabe4b14e256af2acc145b72a37f4b62e89e4fb5c02601a318a8152.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 281, + 133, + 406, + 182 + ], + "blocks": [ + { + "bbox": [ + 281, + 133, + 406, + 182 + ], + "lines": [ + { + "bbox": [ + 281, + 133, + 406, + 182 + ], + "spans": [ + { + "bbox": [ + 281, + 133, + 406, + 182 + ], + "type": "image", + "image_path": "75b9e66296042d17b04eca594b43ad1de34f2cad7a287dbe510221d059b6a977.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 277, + 199, + 403, + 244 + ], + "blocks": [ + { + "bbox": [ + 277, + 199, + 403, + 244 + ], + "lines": [ + { + "bbox": [ + 277, + 199, + 403, + 244 + ], + "spans": [ + { + "bbox": [ + 277, + 199, + 403, + 244 + ], + "type": "image", + "image_path": "aa328635e59287c006911e2f51f0b6db3a1f9d5fdc3f5f84b016084a56157494.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 280, + 255, + 398, + 296 + ], + "blocks": [ + { + "bbox": [ + 280, + 255, + 398, + 296 + ], + "lines": [ + { + "bbox": [ + 280, + 255, + 398, + 296 + ], + "spans": [ + { + "bbox": [ + 280, + 255, + 398, + 296 + ], + "type": "image", + "image_path": "425ab4eb59c40b5d37b3eddc27f7399d2a662ed62d4b7fd0fdac0b8ca4783fa3.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 294, + 311, + 399, + 357 + ], + "blocks": [ + { + "bbox": [ + 294, + 311, + 399, + 357 + ], + "lines": [ + { + "bbox": [ + 294, + 311, + 399, + 357 + ], + "spans": [ + { + "bbox": [ + 294, + 311, + 399, + 357 + ], + "type": "image", + "image_path": "a4c4fcda4d11b1a8c6b0c68a886ca6a56ffeac786a2eb6c5bad0653224b77f51.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 286, + 367, + 409, + 403 + ], + "blocks": [ + { + "bbox": [ + 286, + 367, + 409, + 403 + ], + "lines": [ + { + "bbox": [ + 286, + 367, + 409, + 403 + ], + "spans": [ + { + "bbox": [ + 286, + 367, + 409, + 403 + ], + "type": "image", + "image_path": "130f858e1a4c7f5c41446305cdade90df368422fcc29c2672615cdc12d33de35.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 416, + 388, + 427 + ], + "lines": [ + { + "bbox": [ + 315, + 416, + 388, + 427 + ], + "spans": [ + { + "bbox": [ + 315, + 416, + 388, + 427 + ], + "type": "text", + "content": "Instant3D (15.02s)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 403, + 72, + 529, + 129 + ], + "blocks": [ + { + "bbox": [ + 403, + 72, + 529, + 129 + ], + "lines": [ + { + "bbox": [ + 403, + 72, + 529, + 129 + ], + "spans": [ + { + "bbox": [ + 403, + 72, + 529, + 129 + ], + "type": "image", + "image_path": "446a4a76926606f4e79759f91f976df15c955004eee456a5e17e9af1ba99c31c.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 410, + 133, + 531, + 180 + ], + "blocks": [ + { + "bbox": [ + 410, + 133, + 531, + 180 + ], + "lines": [ + { + "bbox": [ + 410, + 133, + 531, + 180 + ], + "spans": [ + { + "bbox": [ + 410, + 133, + 531, + 180 + ], + "type": "image", + "image_path": "fad229eb454b2d7b8c139f5736523156725d6735512b45b5548679213e3a50f4.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 416, + 189, + 462, + 241 + ], + "blocks": [ + { + "bbox": [ + 416, + 189, + 462, + 241 + ], + "lines": [ + { + "bbox": [ + 416, + 189, + 462, + 241 + ], + "spans": [ + { + "bbox": [ + 416, + 189, + 462, + 241 + ], + "type": "image", + "image_path": "f1fe5b925e4af1be6e702545b3be5ca8b3b90b2f9f4f49cc9eefa5a5ccaee05e.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 484, + 189, + 529, + 241 + ], + "blocks": [ + { + "bbox": [ + 484, + 189, + 529, + 241 + ], + "lines": [ + { + "bbox": [ + 484, + 189, + 529, + 241 + ], + "spans": [ + { + "bbox": [ + 484, + 189, + 529, + 241 + ], + "type": "image", + "image_path": "0dfc3dcbe4e34172619ae622997cabc2216f5986c10284e30352acf1302a4d0d.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 408, + 250, + 470, + 303 + ], + "blocks": [ + { + "bbox": [ + 408, + 250, + 470, + 303 + ], + "lines": [ + { + "bbox": [ + 408, + 250, + 470, + 303 + ], + "spans": [ + { + "bbox": [ + 408, + 250, + 470, + 303 + ], + "type": "image", + "image_path": "b730d47708a28172c7cda6492f28cc1e8d5d473b2eea05b461b9358785307723.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 471, + 250, + 530, + 303 + ], + "blocks": [ + { + "bbox": [ + 471, + 250, + 530, + 303 + ], + "lines": [ + { + "bbox": [ + 471, + 250, + 530, + 303 + ], + "spans": [ + { + "bbox": [ + 471, + 250, + 530, + 303 + ], + "type": "image", + "image_path": "fb856587482a4dcb9a3a4f0f154b0045081a146f5d9335315f4b2ab15c95eac8.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 417, + 309, + 462, + 358 + ], + "blocks": [ + { + "bbox": [ + 417, + 309, + 462, + 358 + ], + "lines": [ + { + "bbox": [ + 417, + 309, + 462, + 358 + ], + "spans": [ + { + "bbox": [ + 417, + 309, + 462, + 358 + ], + "type": "image", + "image_path": "dab9d381ead6a661e1962505b65d99c070d4002c1c004bf6c0d9401d9319c5bb.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 488, + 308, + 528, + 357 + ], + "blocks": [ + { + "bbox": [ + 488, + 308, + 528, + 357 + ], + "lines": [ + { + "bbox": [ + 488, + 308, + 528, + 357 + ], + "spans": [ + { + "bbox": [ + 488, + 308, + 528, + 357 + ], + "type": "image", + "image_path": "f8b6f3b3370d3818c0b6370d45e7503073c96445286fb28d3aa9194afe89e030.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 410, + 359, + 473, + 407 + ], + "blocks": [ + { + "bbox": [ + 410, + 359, + 473, + 407 + ], + "lines": [ + { + "bbox": [ + 410, + 359, + 473, + 407 + ], + "spans": [ + { + "bbox": [ + 410, + 359, + 473, + 407 + ], + "type": "image", + "image_path": "02c8665de1cd69726625f24da476647e80db1de7c11175dca078927a8d2188a4.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 478, + 359, + 541, + 409 + ], + "blocks": [ + { + "bbox": [ + 478, + 359, + 541, + 409 + ], + "lines": [ + { + "bbox": [ + 478, + 359, + 541, + 409 + ], + "spans": [ + { + "bbox": [ + 478, + 359, + 541, + 409 + ], + "type": "image", + "image_path": "c381af4e1ec5aa84b323ff3df4909d0ebbbe3ace19496407f3ef1f8caeb98a60.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 416, + 504, + 427 + ], + "lines": [ + { + "bbox": [ + 440, + 416, + 504, + 427 + ], + "spans": [ + { + "bbox": [ + 440, + 416, + 504, + 427 + ], + "type": "text", + "content": "Turbo3D (0.35s)" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 58, + 502, + 293, + 586 + ], + "blocks": [ + { + "bbox": [ + 55, + 436, + 555, + 480 + ], + "lines": [ + { + "bbox": [ + 55, + 436, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 555, + 480 + ], + "type": "text", + "content": "Figure 4. Comparison of our Turbo3D against baselines LGM [47] and Instant3D [17]. Among these methods, Our method generates the most detailed and physically plausible 3D assets, closely adhering to the provided text prompts. In contrast, LGM tends to generate broken assets with Janus issue [32], while Instant3D has poorer text alignment, oftentimes missing some concepts, e.g., 'spilling out' in the first row, 'river' in the second row, etc." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 502, + 293, + 586 + ], + "lines": [ + { + "bbox": [ + 58, + 502, + 293, + 586 + ], + "spans": [ + { + "bbox": [ + 58, + 502, + 293, + 586 + ], + "type": "image", + "image_path": "c696e2855ab09dbcd9814aec49a813d3a2d579cb39f29b015a8062884d6d6ef2.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 595, + 295, + 651 + ], + "lines": [ + { + "bbox": [ + 55, + 595, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 595, + 295, + 651 + ], + "type": "text", + "content": "Figure 5. User study results comparing our Turbo3D to baseline LGM [47], Instant3D [17], and our slow MV teacher. Our Turbo3D is consistently preferred over baseline LGM and Instant3D, while having on-par preference with our MV teacher. See Fig. 4,6 for visual comparison." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "with a " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "74.9\\%" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " win rate, indicating that users consistently found Turbo3D's outputs more aligned with the input text descriptions and more visually compelling. When evaluated against our teacher model, Turbo3D held a close win rate of " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "50.6\\%" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "49.4\\%" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " preferring MV T., reflecting compara" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 502, + 555, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 550 + ], + "type": "text", + "content": "ble quality between the student and teacher models. The result indicates that our Turbo3D not only achieves significant speedup with distillation, but also preserves the generation ability from the teacher model." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 562, + 408, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 408, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 408, + 575 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "Effect of Single-view Teacher. In Tab. 2, we demonstrate the effectiveness of the dual teacher distillation strategy. The first line is our multi-view (MV) teacher model, which achieves impressive results but runs slowly due to the many diffusion sampling steps required. When performing distillation only with this MV Teacher, the quality drops by a large margin as shown in the second row. When adding a single-view teacher model for distillation, the distilled model achieves much better results compared with the previous one. This configuration approaches the performance of the Multi-step MV Model while maintaining" + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "text", + "content": "23674" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 139, + 72, + 185, + 137 + ], + "blocks": [ + { + "bbox": [ + 83, + 83, + 136, + 116 + ], + "lines": [ + { + "bbox": [ + 83, + 83, + 136, + 116 + ], + "spans": [ + { + "bbox": [ + 83, + 83, + 136, + 116 + ], + "type": "text", + "content": "a baby dragon hatching out of a stone egg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 139, + 72, + 185, + 137 + ], + "lines": [ + { + "bbox": [ + 139, + 72, + 185, + 137 + ], + "spans": [ + { + "bbox": [ + 139, + 72, + 185, + 137 + ], + "type": "image", + "image_path": "14c11aab51a2b79b3b14707f6134539692ef778fc0bc731541c042c1a9f400da.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 204, + 72, + 250, + 137 + ], + "blocks": [ + { + "bbox": [ + 204, + 72, + 250, + 137 + ], + "lines": [ + { + "bbox": [ + 204, + 72, + 250, + 137 + ], + "spans": [ + { + "bbox": [ + 204, + 72, + 250, + 137 + ], + "type": "image", + "image_path": "b8a75eb4345c7411a2cf574b66ba1fb290d9237c1f8cf05f8cb0dce7d3a4ccb1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 272, + 72, + 315, + 137 + ], + "blocks": [ + { + "bbox": [ + 272, + 72, + 315, + 137 + ], + "lines": [ + { + "bbox": [ + 272, + 72, + 315, + 137 + ], + "spans": [ + { + "bbox": [ + 272, + 72, + 315, + 137 + ], + "type": "image", + "image_path": "b4a0e75f9e2d85803a89cf93b419fb492d279819a7a51a6888e7e5e42bb71266.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 333, + 72, + 382, + 137 + ], + "blocks": [ + { + "bbox": [ + 333, + 72, + 382, + 137 + ], + "lines": [ + { + "bbox": [ + 333, + 72, + 382, + 137 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 382, + 137 + ], + "type": "image", + "image_path": "35e5ca8550afdc03748a2a0e9ca5eb6ac891282671ee7e42a3a4b561e55b46e3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 401, + 76, + 451, + 137 + ], + "blocks": [ + { + "bbox": [ + 401, + 76, + 451, + 137 + ], + "lines": [ + { + "bbox": [ + 401, + 76, + 451, + 137 + ], + "spans": [ + { + "bbox": [ + 401, + 76, + 451, + 137 + ], + "type": "image", + "image_path": "221515456723992d465e999b011480f321f742f645c379b9381e8560c46c7969.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 466, + 76, + 517, + 137 + ], + "blocks": [ + { + "bbox": [ + 466, + 76, + 517, + 137 + ], + "lines": [ + { + "bbox": [ + 466, + 76, + 517, + 137 + ], + "spans": [ + { + "bbox": [ + 466, + 76, + 517, + 137 + ], + "type": "image", + "image_path": "87998b6538d655386135afc1be9221cd7f36f6ae4ae860210a9c5676857b3324.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 143, + 142, + 179, + 205 + ], + "blocks": [ + { + "bbox": [ + 84, + 149, + 140, + 190 + ], + "lines": [ + { + "bbox": [ + 84, + 149, + 140, + 190 + ], + "spans": [ + { + "bbox": [ + 84, + 149, + 140, + 190 + ], + "type": "text", + "content": "a kitten \nstanding on top \nof a giant \ntortoise" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 143, + 142, + 179, + 205 + ], + "lines": [ + { + "bbox": [ + 143, + 142, + 179, + 205 + ], + "spans": [ + { + "bbox": [ + 143, + 142, + 179, + 205 + ], + "type": "image", + "image_path": "5351ae9224cfdf8588d2fdb2cf06ba23c523833290ab104683d10d652ea79188.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 203, + 142, + 250, + 205 + ], + "blocks": [ + { + "bbox": [ + 203, + 142, + 250, + 205 + ], + "lines": [ + { + "bbox": [ + 203, + 142, + 250, + 205 + ], + "spans": [ + { + "bbox": [ + 203, + 142, + 250, + 205 + ], + "type": "image", + "image_path": "4173eedb7dd2598e8584a115254ad1fe1e880c543257495dddccbbe6dd28edb4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 275, + 142, + 312, + 204 + ], + "blocks": [ + { + "bbox": [ + 275, + 142, + 312, + 204 + ], + "lines": [ + { + "bbox": [ + 275, + 142, + 312, + 204 + ], + "spans": [ + { + "bbox": [ + 275, + 142, + 312, + 204 + ], + "type": "image", + "image_path": "6da348535d4671af31fa25b97a7e31cbe88326dc82c62094386658935ba50fc0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 331, + 142, + 383, + 205 + ], + "blocks": [ + { + "bbox": [ + 331, + 142, + 383, + 205 + ], + "lines": [ + { + "bbox": [ + 331, + 142, + 383, + 205 + ], + "spans": [ + { + "bbox": [ + 331, + 142, + 383, + 205 + ], + "type": "image", + "image_path": "a326bc282810c2a4fbb0bb9b37e16a49e55517b373e8ef6419aff1fb6091c3d7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 405, + 142, + 448, + 205 + ], + "blocks": [ + { + "bbox": [ + 405, + 142, + 448, + 205 + ], + "lines": [ + { + "bbox": [ + 405, + 142, + 448, + 205 + ], + "spans": [ + { + "bbox": [ + 405, + 142, + 448, + 205 + ], + "type": "image", + "image_path": "76ee5af2d72470e2510d1cff3ba3252db76b0f0956dcb23be217a629713ef14c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 467, + 142, + 514, + 205 + ], + "blocks": [ + { + "bbox": [ + 467, + 142, + 514, + 205 + ], + "lines": [ + { + "bbox": [ + 467, + 142, + 514, + 205 + ], + "spans": [ + { + "bbox": [ + 467, + 142, + 514, + 205 + ], + "type": "image", + "image_path": "f546f40024d31519de3334ac10cf67992d2227877687a309f140222f17a80ada.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 136, + 213, + 187, + 274 + ], + "blocks": [ + { + "bbox": [ + 83, + 209, + 138, + 270 + ], + "lines": [ + { + "bbox": [ + 83, + 209, + 138, + 270 + ], + "spans": [ + { + "bbox": [ + 83, + 209, + 138, + 270 + ], + "type": "text", + "content": "a wide angle DSLR photo of a squirrel in samurai armor wielding a katana" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 136, + 213, + 187, + 274 + ], + "lines": [ + { + "bbox": [ + 136, + 213, + 187, + 274 + ], + "spans": [ + { + "bbox": [ + 136, + 213, + 187, + 274 + ], + "type": "image", + "image_path": "ceaa9e9801aa331355267dec446402868bc2d274aeae513bdfe22a97baa2f32c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 199, + 213, + 251, + 276 + ], + "blocks": [ + { + "bbox": [ + 199, + 213, + 251, + 276 + ], + "lines": [ + { + "bbox": [ + 199, + 213, + 251, + 276 + ], + "spans": [ + { + "bbox": [ + 199, + 213, + 251, + 276 + ], + "type": "image", + "image_path": "084eb55434eec498e2f4ee1aee37e7fecbea33d68cabfcd9bd036c857a6b3c04.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 288, + 229, + 298 + ], + "lines": [ + { + "bbox": [ + 153, + 288, + 229, + 298 + ], + "spans": [ + { + "bbox": [ + 153, + 288, + 229, + 298 + ], + "type": "text", + "content": "MV teacher (10.18s)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 308, + 555, + 352 + ], + "lines": [ + { + "bbox": [ + 55, + 308, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 555, + 352 + ], + "type": "text", + "content": "Figure 6. Ablation of our Dual-teacher distillation algorithm. Naively distilling MV teacher (middle column) causes compound mode collapse (see Sec. 4.1), producing overly smooth synthetic-looking assets. Our dual-teacher distillation (right column) fixes the issue and generates 3D assets that are as photorealistic as, if not more than, the baseline MV teacher (left column). We also include the inference timings for each method; the distilled model is " + }, + { + "bbox": [ + 55, + 308, + 555, + 352 + ], + "type": "inline_equation", + "content": "\\sim 50\\mathrm{x}" + }, + { + "bbox": [ + 55, + 308, + 555, + 352 + ], + "type": "text", + "content": " faster than the teacher model." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 260, + 211, + 315, + 276 + ], + "blocks": [ + { + "bbox": [ + 260, + 211, + 315, + 276 + ], + "lines": [ + { + "bbox": [ + 260, + 211, + 315, + 276 + ], + "spans": [ + { + "bbox": [ + 260, + 211, + 315, + 276 + ], + "type": "image", + "image_path": "cade8740465662074ba3ac19e8e4387a4709a3b6230231096e61f39b3730f896.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 288, + 384, + 299 + ], + "lines": [ + { + "bbox": [ + 255, + 288, + 384, + 299 + ], + "spans": [ + { + "bbox": [ + 255, + 288, + 384, + 299 + ], + "type": "text", + "content": "MV teacher only distillation (0.35s)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 334, + 210, + 392, + 262 + ], + "blocks": [ + { + "bbox": [ + 334, + 210, + 392, + 262 + ], + "lines": [ + { + "bbox": [ + 334, + 210, + 392, + 262 + ], + "spans": [ + { + "bbox": [ + 334, + 210, + 392, + 262 + ], + "type": "image", + "image_path": "d638913a9d51230b9845e32ea9c8f108c978159f5ed354372e520e91edfe6b54.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 395, + 210, + 451, + 275 + ], + "blocks": [ + { + "bbox": [ + 395, + 210, + 451, + 275 + ], + "lines": [ + { + "bbox": [ + 395, + 210, + 451, + 275 + ], + "spans": [ + { + "bbox": [ + 395, + 210, + 451, + 275 + ], + "type": "image", + "image_path": "79eb8e06658290f62d81bb0f6fd1dc01c89e28c7b488173e89f41021b4b1d192.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 462, + 210, + 526, + 276 + ], + "blocks": [ + { + "bbox": [ + 462, + 210, + 526, + 276 + ], + "lines": [ + { + "bbox": [ + 462, + 210, + 526, + 276 + ], + "spans": [ + { + "bbox": [ + 462, + 210, + 526, + 276 + ], + "type": "image", + "image_path": "0853bc60ea13790ae35eb4300b8725763e8fa5c6a4ff96e8b4f0e79c97f71970.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 404, + 288, + 518, + 299 + ], + "lines": [ + { + "bbox": [ + 404, + 288, + 518, + 299 + ], + "spans": [ + { + "bbox": [ + 404, + 288, + 518, + 299 + ], + "type": "text", + "content": "Dual teacher distillation (0.35s)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "table", + "bbox": [ + 58, + 361, + 294, + 441 + ], + "blocks": [ + { + "bbox": [ + 58, + 361, + 294, + 441 + ], + "lines": [ + { + "bbox": [ + 58, + 361, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 58, + 361, + 294, + 441 + ], + "type": "table", + "html": "
AblationCLIP Score ↑VQA Score ↑
Multi-step MV Model28.040.77
Few-step Model (MV Teacher)26.600.69
Few-step Model (Dual Teacher)27.610.76
", + "image_path": "cc30cd6a29de9e8a23b45954878568565fd9e9fbe9607a36a090426fbb3d34b1.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 512, + 295, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 512, + 295, + 595 + ], + "spans": [ + { + "bbox": [ + 55, + 512, + 295, + 595 + ], + "type": "text", + "content": "the efficiency benefits of the few-step setup, showcasing the advantages of the dual-teacher strategy in our distillation. We also provide visualizations of the three models in Fig. 6. These comparisons show that the dual teacher distillation model strikes a balance between detail retention and stylization, closely replicating the MV teacher model's quality while benefiting from the efficiency gains of distillation." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 596, + 295, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 596, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 295, + 657 + ], + "type": "text", + "content": "Effect of Latent GS-LRM. In Tab. 3, we showcase the effectiveness of latent GS-LRM. Compared with the original GS-LRM which operates in pixel space, our latent GS-LRM is able to skip the expensive image decoding process while achieving similar image quality." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 669, + 128, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 128, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 128, + 681 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "In this work, we propose Turbo3D for ultra-fast text-to-3D generation. To enable fast multi-view generation, we pro" + } + ] + } + ], + "index": 30 + }, + { + "type": "table", + "bbox": [ + 317, + 361, + 553, + 425 + ], + "blocks": [ + { + "bbox": [ + 55, + 454, + 295, + 510 + ], + "lines": [ + { + "bbox": [ + 55, + 454, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 295, + 510 + ], + "type": "text", + "content": "Table 2. Ablation of dual teacher distillation. Distillation leads to quality drop compared with the MV teacher model. Compared with naively distilling the single MV teacher, dual-teacher distillation leads to much smaller quality drop. See Fig. 6 for visual comparison." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 361, + 553, + 425 + ], + "lines": [ + { + "bbox": [ + 317, + 361, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 553, + 425 + ], + "type": "table", + "html": "
AblationCLIP Score ↑VQA Score ↑Inference Time ↓
Pixel GS-LRM [62]27.620.760.45s
Latent GS-LRM27.610.760.35s
", + "image_path": "4689361cdb19fc0e9d186e3563272f395eefe6954231812ce5c60b2917b9efdc.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "table_body" + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 437, + 555, + 481 + ], + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 481 + ], + "type": "text", + "content": "Table 3. Ablation for latent GS-LRM. We report the CLIP score, VQA score, and overall text-to-3D inference time for comparison. Our latent GS-LRM achieves similar image quality while enabling better efficiency (" + }, + { + "bbox": [ + 313, + 437, + 555, + 481 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 313, + 437, + 555, + 481 + ], + "type": "text", + "content": "22% speedup)." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 501, + 555, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 555, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 555, + 621 + ], + "type": "text", + "content": "pose to distill a multi-step multi-view generator into a few-step multi-view generator. Moreover, to restore the multiview consistency and photo-realism during distillation, we introduce a novel dual-teacher distillation framework. To further improve the multi-view reconstruction efficiency, we propose a latent GS-LRM which directly reconstructs 3D Gaussians from multi-view latents. Extensive experiments demonstrate that our proposed Turbo3D is able to achieve outstanding performance in terms of both generation quality and inference efficiency." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 314, + 645, + 416, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 645, + 416, + 658 + ], + "spans": [ + { + "bbox": [ + 314, + 645, + 416, + 658 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 669, + 554, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 554, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 554, + 703 + ], + "type": "text", + "content": "This work began during Hanzhe Hu and Tianwei Yin's internships at Adobe Research. This research was supported in part by NSF award IIS-2345610." + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "text", + "content": "23675" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "text", + "content": "[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. Renderdiffusion: Image diffusion for 3d reconstruction, inpainting and generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12608-12618, 2023. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 158, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 158, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 296, + 213 + ], + "type": "text", + "content": "[2] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18558-18568, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 213, + 296, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 213, + 296, + 278 + ], + "spans": [ + { + "bbox": [ + 62, + 213, + 296, + 278 + ], + "type": "text", + "content": "[3] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 4, 5, 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 280, + 295, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 295, + 334 + ], + "type": "text", + "content": "[4] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of " + }, + { + "bbox": [ + 62, + 280, + 295, + 334 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{d}" + }, + { + "bbox": [ + 62, + 280, + 295, + 334 + ], + "type": "text", + "content": " objects. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 335, + 295, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 295, + 378 + ], + "type": "text", + "content": "[5] Jean-Yves Franceschi, Mike Gartrell, Ludovic Dos Santos, Thibaut Issenhuth, Emmanuel de Bezenac, Mickaël Chen, and Alain Rakotomamonjy. Unifying gans and score-based diffusion as generative particle models. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "type": "text", + "content": "[6] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 424, + 295, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 424, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 424, + 295, + 456 + ], + "type": "text", + "content": "[7] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 458, + 295, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 295, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 295, + 501 + ], + "type": "text", + "content": "[8] Yicong Hong, Kai Zhang, Jiumiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 502, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 295, + 555 + ], + "type": "text", + "content": "[9] Hanzhe Hu, Zhizhuo Zhou, Varun Jampani, and Shubham Tulsiani. Mvd-fusion: Single-view 3d via depth-consistent multi-view generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9698-9707, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 557, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 295, + 590 + ], + "type": "text", + "content": "[10] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 591, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 623 + ], + "type": "text", + "content": "[11] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 646 + ], + "type": "text", + "content": "[12] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. In NeurIPS, 2021. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 680 + ], + "type": "text", + "content": "[13] MIT HAN Lab. Patch conv: Patch convolution to avoid largegpu memory usage. https://hanlab.mit.edu/blog/patch-conv, 2024. Accessed: 2024-11-14. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "type": "text", + "content": "[14] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024. Accessed: 2024-11-14. 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 105 + ], + "type": "text", + "content": "[15] Sangyun Lee, Zinan Lin, and Giulia Fanti. Improving the training of rectified flows. arXiv preprint arXiv:2405.20320, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 107, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 554, + 161 + ], + "type": "text", + "content": "[16] Baiqi Li, Zhiqiu Lin, Wenxuan Peng, Jean de Dieu Nyandwi, Daniel Jiang, Zixian Ma, Simran Khanuja, Ranjay Krishna, Graham Neubig, and Deva Ramanan. Naturalbench: Evaluating vision-language models on natural adversarial samples. arXiv preprint arXiv:2410.14669, 2024. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 162, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 554, + 217 + ], + "type": "text", + "content": "[17] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214, 2023. 2, 3, 4, 6, 7, 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 217, + 554, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 554, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 554, + 281 + ], + "type": "text", + "content": "[18] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 283, + 554, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 554, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 554, + 316 + ], + "type": "text", + "content": "[19] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 317, + 554, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 317, + 554, + 369 + ], + "spans": [ + { + "bbox": [ + 317, + 317, + 554, + 369 + ], + "type": "text", + "content": "[20] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 371, + 555, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 555, + 426 + ], + "type": "text", + "content": "[21] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 426, + 554, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 554, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 554, + 459 + ], + "type": "text", + "content": "[22] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 460, + 554, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 554, + 503 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 554, + 503 + ], + "type": "text", + "content": "[23] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In The Twelfth International Conference on Learning Representations, 2023. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 504, + 554, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 554, + 548 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 554, + 548 + ], + "type": "text", + "content": "[24] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 548, + 554, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 554, + 590 + ], + "type": "text", + "content": "[25] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 592, + 554, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 592, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 592, + 554, + 635 + ], + "type": "text", + "content": "[26] Weijian Luo, Tianyang Hu, Shifeng Zhang, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang. Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 636, + 554, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 554, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 554, + 679 + ], + "type": "text", + "content": "[27] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In CVPR, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[28] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. AutoSDF: Shape priors for 3d completion, reconstruction and generation. In CVPR, 2022. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 318, + 761 + ], + "type": "text", + "content": "23676" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[29] Thuan Hoang Nguyen and Anh Tran. Swiftbrush: One-step text-to-image diffusion model with variational score distillation. In CVPR, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "text", + "content": "[30] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[31] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 6, 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 230 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 230 + ], + "type": "text", + "content": "[32] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 294, + 297 + ], + "type": "text", + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "text", + "content": "[34] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "type": "text", + "content": "[35] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAPH 2023 conference proceedings, pages 1–11, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 443 + ], + "type": "text", + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 510 + ], + "type": "text", + "content": "[37] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "type": "text", + "content": "[38] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 2, 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "type": "text", + "content": "[39] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. arXiv preprint arXiv:2403.12015, 2024. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "type": "text", + "content": "[40] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "type": "text", + "content": "[41] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 3, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[42] Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "meshes with decoder-only transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19615-19625, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 129 + ], + "type": "text", + "content": "[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 130, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 174 + ], + "type": "text", + "content": "[44] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 175, + 553, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 198 + ], + "type": "text", + "content": "[45] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 199, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 199, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 199, + 553, + 252 + ], + "type": "text", + "content": "[46] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Viewset diffusion:(0-) image-conditioned 3d generative models from 2d data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8863-8873, 2023. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 308 + ], + "type": "text", + "content": "[47] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1-18. Springer, 2025. 6, 7, 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 310, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 376 + ], + "type": "text", + "content": "[48] Ayush Tewari, Tianwei Yin, George Cazenavette, Semon Rezchikov, Josh Tenenbaum, Frédo Durand, Bill Freeman, and Vincent Sitzmann. Diffusion with forward models: Solving stochastic inverse problems without direct supervision. Advances in Neural Information Processing Systems, 36: 12349-12362, 2023. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 377, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 431 + ], + "type": "text", + "content": "[49] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151, 2024. 6, 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 434, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 465 + ], + "type": "text", + "content": "[50] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural computation, 23(7):1661-1674, 2011. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 468, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 553, + 533 + ], + "type": "text", + "content": "[51] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitrii Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. SV3D: Novel multi-view synthesis and 3D generation from a single image using latent video diffusion. In European Conference on Computer Vision (ECCV), 2024. 3, 6, 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 535, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 567 + ], + "type": "text", + "content": "[52] Chen Wang, Jiatao Gu, Xiaoxiao Long, Yuan Liu, and Lingjie Liu. Geco: Generative image-to-3d within a second. arXiv preprint arXiv:2405.20327, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 568, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 622 + ], + "type": "text", + "content": "[53] Peng Wang, Hao Tan, Sai Bi, Yinghao Xu, Fujun Luan, Kalyan Sunkavalli, Wenping Wang, Zexiang Xu, and Kai Zhang. Pf-lrm: Pose-free large reconstruction model for joint pose and shape prediction. arXiv preprint arXiv:2311.12024, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 624, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 689 + ], + "type": "text", + "content": "[54] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4563-4573, 2023. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[55] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "text", + "content": "23677" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 498 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "diverse text-to-3d generation with variational score distillation. In NeurIPS, 2023. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[56] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Meshlrm: Large reconstruction model for high-quality mesh. arXiv preprint arXiv:2404.12385, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "text", + "content": "[57] Shuang Wu, Youtian Lin, Feihu Zhang, Yifei Zeng, Jingxi Xu, Philip Torr, Xun Cao, and Yao Yao. Direct3d: Scalable image-to-3d generation via 3d latent diffusion transformer. arXiv preprint arXiv:2405.14832, 2024. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 239 + ], + "type": "text", + "content": "[58] Yinghao Xu, Hao Tan, Fujun Luan, Sai Bi, Peng Wang, Jiahao Li, Zifan Shi, Kalyan Sunkavalli, Gordon Wetzstein, Zexiang Xu, et al. Dmv3d: Denoising multi-view diffusion using 3d large reconstruction model. arXiv preprint arXiv:2311.09217, 2023. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "type": "text", + "content": "[59] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image generation via diffusion gans. In CVPR, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "text", + "content": "[60] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. arXiv preprint arXiv:2405.14867, 2024. 2, 3, 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 320, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 384 + ], + "type": "text", + "content": "[61] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6613-6623, 2024. 2, 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 387, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 294, + 441 + ], + "type": "text", + "content": "[62] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 2, 3, 5, 8, 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 443, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 443, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 443, + 294, + 498 + ], + "type": "text", + "content": "[63] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 2, 3" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "spans": [ + { + "bbox": [ + 293, + 752, + 317, + 761 + ], + "type": "text", + "content": "23678" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TurboFill_ Adapting Few-step Text-to-image Model for Fast Image Inpainting/6302d669-b0e3-4a35-adf2-a42f09f692e9_content_list.json b/2025/TurboFill_ Adapting Few-step Text-to-image Model for Fast Image Inpainting/6302d669-b0e3-4a35-adf2-a42f09f692e9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..13ebb13eb8f79acefceb42e91e506aa45933d95b --- /dev/null +++ b/2025/TurboFill_ Adapting Few-step Text-to-image Model for Fast Image Inpainting/6302d669-b0e3-4a35-adf2-a42f09f692e9_content_list.json @@ -0,0 +1,1662 @@ +[ + { + "type": "text", + "text": "TurboFill: Adapting Few-step Text-to-image Model for Fast Image Inpainting", + "text_level": 1, + "bbox": [ + 107, + 130, + 890, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Liangbin Xie $^{1,4}$ Daniil Pakhomov $^{3}$ Zhonghao Wang $^{3}$ Zongze Wu $^{3}$ Ziyan Chen $^{4}$ Yuqian Zhou $^{3}$ Haitian Zheng $^{3}$ Zhifei Zhang $^{3}$ Zhe Lin $^{3}$ Jiantao Zhou $^{1\\dagger}$ Chao Dong $^{2,4\\dagger}$", + "bbox": [ + 114, + 179, + 880, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ State Key Laboratory of Internet of Things for Smart City, University of Macau $^{2}$ Shenzhen University of Advanced Technology $^{3}$ Adobe $^{4}$ Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences", + "bbox": [ + 109, + 218, + 879, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yellow baby chick on wooden surface.", + "bbox": [ + 192, + 266, + 393, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A miniature blue convertible with its top down", + "bbox": [ + 607, + 266, + 864, + 276 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5e6687e177172a2e3492764e5438725748c2ab485972dd07f6ca401470d5dd6f.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 93, + 277, + 225, + 381 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cf85798f3423a8c63336ef0e7fb3ca0143b92d97ff145226d2dc3ee915783591.jpg", + "image_caption": [ + "BrushNet* (50 steps)" + ], + "image_footnote": [], + "bbox": [ + 228, + 277, + 362, + 381 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/55f52f85a82c858ccd571c89d0524f9c9ba5a270297927b9a3e1758ff8d6da74.jpg", + "image_caption": [ + "TurboFill (4 steps)" + ], + "image_footnote": [], + "bbox": [ + 364, + 277, + 500, + 381 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/05f8888f7625f773512e8f30dec75077631e28de82b334399f7d1de6d10a454a.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 501, + 277, + 635, + 381 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/25027e4998702b86029330b9b3fc906a857b1382f886d0e13e500426466ac479.jpg", + "image_caption": [ + "BrushNet* (50 steps)" + ], + "image_footnote": [], + "bbox": [ + 637, + 277, + 771, + 381 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/95a4da8cca6f387d867ac3ef78f1353c54b5c7eba1b51a9d7d26ee7bdcf40df4.jpg", + "image_caption": [ + "TurboFill (4 steps)" + ], + "image_footnote": [], + "bbox": [ + 774, + 277, + 906, + 381 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A fluffy brown teddy bear wearing a tiny hat. A tinted sunglasses lens reflecting tall palm trees. Man with cowboy hat and bandana in front of wooden wall.", + "bbox": [ + 102, + 405, + 913, + 415 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d6047e0f407158605f419a54b2e168329792a9b2bc5125b86627f589d69ce254.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 93, + 417, + 225, + 521 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bc05916997be740db3f6a871795c007341f28352a8298aee31e15d90a38fb309.jpg", + "image_caption": [ + "TurboFill (4 steps)" + ], + "image_footnote": [], + "bbox": [ + 228, + 417, + 362, + 521 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aca537c86d5eab556bee1c64a0e847af6fd4a5fe53f834e119a82344c07aa6c5.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 364, + 417, + 498, + 521 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/01ea8822a4fc25883e047968a37c3f0fb692f7505fe0c5709bc974fc3ce99899.jpg", + "image_caption": [ + "TurboFill (4 steps)", + "Figure 1. We propose TurboFill, a fast image inpainting method that leverages a 3-step adversarial training scheme. With only four diffusion steps, TurboFill outperforms the multi-step BrushNet* [11], delivering realistic details and textures with remarkable efficiency." + ], + "image_footnote": [], + "bbox": [ + 501, + 417, + 635, + 521 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9d5db6efbb2ed4a309e250980081c7271eb8830fafc8b7802c2c831d85d88a38.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 637, + 417, + 771, + 521 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/477974fc0ee3751fa55b9273381bae05337c978633db4ef010f04012f61362e2.jpg", + "image_caption": [ + "TurboFill (4 steps)" + ], + "image_footnote": [], + "bbox": [ + 774, + 417, + 906, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 587, + 326, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper introduces TurboFill, a fast image inpainting model that enhances a few-step text-to-image diffusion model with an inpainting adapter for high-quality and efficient inpainting. While standard diffusion models generate high-quality results, they incur high computational costs. We overcome this by training an inpainting adapter on a few-step distilled text-to-image model, DMD2, using a novel 3-step adversarial training scheme to ensure realistic, structurally consistent, and visually harmonious inpainted regions. To evaluate TurboFill, we propose two benchmarks: DilationBench, which tests performance across mask sizes, and HumanBench, based on human feedback for complex prompts. Experiments show that TurboFill outperforms both multi-step BrushNet and few-step inpainting methods, setting a new benchmark for high-performance inpainting tasks. The project page is available here.", + "bbox": [ + 86, + 619, + 485, + 861 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 587, + 643, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image inpainting, the task of filling missing regions in an image, has seen significant advancement with the rise of deep generative models [1, 5, 11, 16]. Existing diffusion-based inpainting methods have shown promising results for various applications such as object removal, regeneration, and text-guided object insertion.", + "bbox": [ + 511, + 612, + 906, + 703 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The most straightforward way of training diffusion-based inpainting models is to add the masked image (background conditioning) to the input channels of a pretrained diffusion model and finetune it with inpainting data, as shown by SDXL-inpaint [7]. However, such methods typically require fine-tuning the entire network, entailing large training datasets and significant computational resources. In contrast, adapter-based approaches, like BrushNet [11], utilize a ControlNet [33] to capture conditioning information and inject features into a frozen diffusion U-Net. BrushNet achieves state-of-the-art inpainting results while incurring relatively low training costs in terms of data and computational resources, owing to the adapter's efficiency and the", + "bbox": [ + 509, + 704, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 114, + 886, + 243, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7613", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/078363a1a7eb3661db3f22dc4e2e4436d37c76c5e995a6aa09d7814df2853e0e.jpg", + "image_caption": [ + "Figure 2. (Zoom in for best view) 1. The multi-step adapter achieves high-quality inpainting results but incurs significant inference costs, requiring over 50 diffusion steps. 2. Applying the pre-trained multi-step BrushNet adapter directly to the few-step U-Net (DMD2) results in artifacts, including oversaturated colors and semantic inconsistencies (e.g., generating a dog with two tails). 3. Training the adapter-DMD2 solely with diffusion loss produces blurred outputs with low-quality inpainting results. 4. In contrast, training the adapter-DMD2 using the proposed 3-step adversarial training scheme yields high-quality inpainting results, requiring only four diffusion steps." + ], + "image_footnote": [], + "bbox": [ + 93, + 51, + 903, + 268 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "preservation of the U-Net's original generation capacity.", + "bbox": [ + 89, + 357, + 460, + 372 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the advantage of BrushNet-like models, they still depend on many iterative sampling steps, leading to substantial inference cost in practice. Interestingly, we found that an inpainting adapter trained on a pretrained text-to-image diffusion model can be seamlessly integrated into its corresponding few-step student model without further training, while yielding reasonable quality results (Fig. 2 ②). This can be attributed to the semantic alignment between the teacher and student models. However, this naive migration approach exhibits quality issues such as color shifts, over-saturation, and loss of details.", + "bbox": [ + 88, + 372, + 482, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Alternatively, we can use a two-stage approach by first training a teacher BrushNet and then applying advanced diffusion distillation techniques [13, 15, 20, 26, 27] to the teacher BrushNet to reduce steps while maintaining quality. Although recent distillation methods for diffusion models have demonstrated remarkable performance, we have found that due to BrushNet's large model size, it demands excessive memory and incurs high computational costs, making the training process difficult to manage.", + "bbox": [ + 88, + 537, + 482, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a simpler single-stage training strategy by training an inpainting adapter directly on top of a pretrained few-step distilled text-to-image model (e.g., DMD2). We observe that naively training an inpainting adapter and few-step model solely with diffusion loss results in blurry, low-quality outputs (Fig. 2③). To address these shortcomings and enhance inpainting performance, we introduce a diffusion discriminator to guide optimization of the adapter-DMD2 generator. The diffusion discriminator is trained to differentiate between noisy real images and fake latents, while a combined diffusion loss ensures the model learns meaningful semantic representations.", + "bbox": [ + 89, + 674, + 482, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, TurboFill is trained with a novel 3-step adversarial training scheme with two generator update steps and one discriminator update step. Our model enables high-", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "quality image inpainting with only a few diffusion steps. Compared to the naive approach of training a multi-step adapter (requiring 8 V100 GPUs for 72 hours) followed by distillation (requiring 64 A100 GPUs for 50 hours), TurboFill attains comparable outcomes using merely 8 A100 GPUs within 50 hours. This accomplishment marks a reduction in training time by over a factor of 10 and a substantial enhancement in training efficiency.", + "bbox": [ + 511, + 357, + 906, + 478 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing benchmarks often fall short in simulating real-world user operations. To address this gap, we introduce two new benchmarks: DilationBench and HumanBench. DilationBench provides an objective measure by testing the model on the masks with varying dilation sizes, while HumanBench incorporates human user feedback and complex prompts, offering insights into real-world performance. These benchmarks contribute a robust evaluation framework that can facilitate advancements in inpainting models and more accurately reflect user-oriented performance.", + "bbox": [ + 511, + 479, + 908, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive quantitative and qualitative evaluations demonstrate that TurboFill not only supports complex prompts and user-driven inpainting tasks but also outperforms existing few-step inpainting models in terms of speed and quality. In summary, our contributions are as follows:", + "bbox": [ + 511, + 635, + 906, + 710 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A novel 3-step adversarial training scheme is proposed to train an inpainting adapter directly on top of few-step text-to-image diffusion model, which helps the model to focus on image quality and improves textures, details, and scene harmonization.", + "- New benchmarks, DilationBench and HumanBench, are designed to better evaluate inpainting methods in user-centered scenarios.", + "- We demonstrate that our few-step TurboFill outperforms multi-step BrushNet for the inpainting task and achieves state-of-the-art performance in a few-step setting through extensive experiments." + ], + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7614", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 91, + 89, + 232, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image inpainting, a classic problem in computer vision, focuses on reconstructing missing regions in an image with realistic and natural content. Early methods, including traditional techniques, VAEs, and GANs [2, 3, 9, 28-31], often struggled to produce high-quality results. Recently, diffusion models have gained popularity for their ability to generate high-quality, diverse outputs with fine-grained control, demonstrating significant advancements over earlier approaches [1, 5, 11, 16]. Diffusion models, particularly the Stable Diffusion family, have introduced a robust framework for image generation and inpainting by iteratively denoising latent representations. Stable Diffusion-based methods have achieved impressive results in high-resolution image inpainting, yet are computationally intensive due to multi-step sampling.", + "bbox": [ + 89, + 114, + 483, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inpainting tasks based on diffusion models generally follow two approaches: training an entire diffusion model or incorporating external modules. The first approach, like PowerPaint [5], CATDiffusion [6], and SDXL-Inpainting, focuses on adapting the model itself. PowerPaint introduces learnable task prompts along with tailored fine-tuning strategies to guide the model's focus on specific inpainting targets, while SDXL-Inpainting [7] directly utilizes both the background region and the mask as inputs to fine-tune a new model. In contrast, the second approach, exemplified by methods such as ControlNet [33] or T2I-Adapter [17], leverages external modules to enhance the diffusion model's capabilities without retraining it entirely. Recent methods, such as BrushNet [11], employ mask-specific prompts, refining the generated content within the masked area for improved coherence with the background.", + "bbox": [ + 89, + 340, + 483, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method builds on these developments by leveraging a ControlNet-like adapter for DMD2, integrating both GAN and diffusion losses to enhance image realism and style coherence. By focusing on masked-region-specific prompts and incorporating fine-grained diffusion control, our approach achieves fast and high-quality inpainting results that align closely with user inputs and surrounding context, advancing upon prior work in speed, quality, and flexibility in real-world applications.", + "bbox": [ + 89, + 584, + 483, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 733, + 217, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The forward diffusion process gradually adds Gaussian noise $\\epsilon_t$ to a clean image $x_0$ by", + "bbox": [ + 89, + 758, + 482, + 789 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 803, + 482, + 820 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{\\alpha}_t$ controls noise schedule. In the training phase, a denoising network $\\epsilon_{\\theta}$ is trained to predict $\\epsilon_t$ given $x_t$ , timestep $t$ and text prompt $c$ , with the diffusion loss:", + "bbox": [ + 89, + 827, + 483, + 873 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL \\left(\\hat {\\epsilon} _ {\\theta}\\right) = \\mathbb {E} _ {x _ {0} \\sim q; \\epsilon_ {t} \\sim \\mathcal {N} (0, 1)} [ \\| \\epsilon_ {\\theta} (x _ {t}, c, t) - \\epsilon_ {t} \\| ^ {2} ]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 883, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enable condition generation, adapter-based methods [11, 17, 32] introduce additional conditioning information via a lightweight adapter, and achieve efficient training (since the main diffusion network remains frozen), with a slight inference overhead.", + "bbox": [ + 511, + 90, + 903, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Traditionally, reverse diffusion processes necessitate a computationally intensive multi-step iterative sampling procedure (typically 20-50 steps). Distillation techniques [13, 15, 20, 26] have been developed to reduce this requirement, achieving accelerated sampling with as few as 1-4 steps. However, these distillation processes are resource-intensive in both computational power and time. In contrast, our approach directly trains the adapter with both the multi-step and few-step text-to-image diffusion model, eliminating the need for a separate distillation process, thereby enabling more efficient training and inference.", + "bbox": [ + 511, + 167, + 906, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Our Approach", + "text_level": 1, + "bbox": [ + 511, + 345, + 661, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we propose a fast inpainting model, TurboFill, which directly trains an inpainting adapter on top of a few-step text-to-image diffusion model. As shown in Fig. 3, there are three components: the fast generator, the slow generator and the diffusion discriminator. All these three components aim to enhance the adapter's capability of inpainting masked images. Specifically, the slow generator helps the adapter to denoise the noisy latents towards the most probable direction of realistic image generation. The fast generator enables the adapter to generate a clean image for critique during training. The diffusion discriminator guides the adapter to generate images with better textures, details and content harmonization. Based on these three components, TurboFill follows a 3-step adversarial training scheme. The training alternates between 3 steps throughout the training process. In the following sections, we provide a detailed description of the training procedure for each step and the components.", + "bbox": [ + 511, + 369, + 906, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Training the Adapter in Slow Generator", + "text_level": 1, + "bbox": [ + 511, + 652, + 857, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To capture conditioning information and introduce additional features to the diffusion model, we utilize an inpainting adapter with the same ControlNet architecture as BrushNet [11]. Given that SDXL is a multi-step text-to-image generator, we designate the integration of the inpainting adapter and SDXL as the 'slow generator'. In contrast, DMD2 is a few-step text-to-image generator, so the combination of the inpainting adapter and DMD2 is named the 'fast generator'. In both generators, conditioning information is introduced by concatenating the noisy latent $x_{t}$ , the background image latent $x_{bg}$ , and a downsampled binary mask $m$ . This concatenated input is then processed by the inpainting adapter, while $x_{t}$ is simultaneously fed into the UNet. Then, the features produced by the inpainting adapter are fused with the UNet's features via residual connection.", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7615", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/014f01b28fe85da2eb5385e21f38092ea752b505a51043d421c657ee3f49ed6f.jpg", + "image_caption": [ + "Figure 3. The training of TurboFill alternates between 3 steps: (1) optimizing the adapter using the gradient of $\\mathcal{L}_{\\mathrm{Diff}}^R$ , and (2) $\\mathcal{L}_{\\mathcal{G}}$ and $\\mathcal{L}_{\\mathrm{BG}}$ are employed to update the adapter in fast generator, and (3) $\\mathcal{L}_{\\mathrm{Diff}}^F$ and $\\mathcal{L}_{\\mathcal{D}}$ are jointly applied to update the parameters of the diffusion discriminator module. Note that the adapters in slow generator and fast generator share same weights during training." + ], + "image_footnote": [], + "bbox": [ + 109, + 51, + 885, + 294 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 3 (1), the parameters of the inpainting adapter in slow generator are updated based on the real diffusion loss:", + "bbox": [ + 89, + 356, + 482, + 401 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {D i f f}} ^ {R} = \\min _ {\\epsilon_ {\\theta}} \\mathbb {E} _ {x _ {t}, t \\sim [ 0, T ], \\epsilon \\sim \\mathcal {N} (0, 1)} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 410, + 480, + 433 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| \\epsilon - \\epsilon_ {\\theta} ^ {\\mathrm {S D X L}} (x _ {t}, [ x _ {b g}, x _ {t}, m ], t) \\right\\| _ {2} ^ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 436, + 400, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $T = 1,000$ and DDPM scheduler [10] is employed to generate $x_{t}$ .", + "bbox": [ + 89, + 473, + 482, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Training the Adapter in Fast Generator", + "text_level": 1, + "bbox": [ + 89, + 511, + 429, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given that adversarial training is highly effective in minimizing the divergence between different distributions, we employ adversarial training to minimize the discrepancy between real latent distribution and the fake latent distribution $\\{\\hat{x}_0\\}$ . Specifically, we first generate the clean fake latent $\\hat{x}_0$ from the noisy latent $x_{t}$ with the fast generator in a single step, as described by", + "bbox": [ + 89, + 532, + 483, + 640 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} _ {0} = \\frac {x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\times \\epsilon_ {\\theta} ^ {\\mathrm {D M D 2}} \\left(x _ {t} , \\left[ x _ {b g} , x _ {t} , m \\right] , t\\right)}{\\sqrt {\\bar {\\alpha} _ {t}}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 647, + 482, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\bar{\\alpha}_t$ is determined by the LCM scheduler [15] and $t$ is the sampled timestep. Then, the noisy version $\\hat{x}_t$ originated from $\\hat{x_0}$ is fed into the diffusion discriminator. Meanwhile, to preserve the background region, $\\hat{x_0}$ is used to compute the background loss.", + "bbox": [ + 89, + 688, + 482, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion discriminator. As shown in Fig. 3 (2), the diffusion discriminator $\\mathcal{D}_{\\phi}$ consists of the SDXL encoder, the assistant encoder, and the convolution-based classifier. For assistant encoder, we use the same encoder structure of inpainting adapter. It takes concatenated noisy latent, downsampled binary mask, and masked image latent $x_{bg}$ as the input. The features extracted by the assistant encoder are progressively integrated into the SDXL encoder. The final feature maps are then passed through the classifier and", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mapped to a one-dimensional vector. Based on the vector, we calculate $\\mathcal{L}_{\\mathcal{G}}$ :", + "bbox": [ + 511, + 356, + 906, + 387 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {G}} = \\min _ {\\epsilon_ {\\theta}} - \\mathbb {E} _ {y _ {t}, t \\sim [ 0, T ]} \\left[ \\log \\mathcal {D} _ {\\phi} \\left(\\hat {x} _ {0}, \\left[ x _ {b g}, \\hat {x} _ {0}, m \\right], t\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 398, + 903, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $T = 1,000$", + "bbox": [ + 513, + 435, + 640, + 449 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Background preservation. For the image inpainting task, it is essential to preserve the background region. Therefore, for the clean fake latent $\\hat{x_0}$ , we apply a reconstruction loss to the background region to minimize the distance between $\\hat{x_0}$ and the clean latent $x_0$ :", + "bbox": [ + 511, + 450, + 905, + 525 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {B G}} = \\left\\| x _ {0} \\odot m - \\hat {x} _ {0} \\odot m \\right\\| _ {2} ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 539, + 903, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\odot$ represents point-wise multiplication. Based on $\\lambda_1\\mathcal{L}_{\\mathcal{G}} + \\lambda_2\\mathcal{L}_{\\mathrm{BG}}$ , in this step, only the parameters of the inpainting adapter are updated.", + "bbox": [ + 511, + 564, + 906, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Training the Diffusion Discriminator", + "text_level": 1, + "bbox": [ + 511, + 619, + 831, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When updating the discriminator, we fix the parameters of the inpainting adapter and update the parameters of the assistant encoder and classifier. Specifically, both the real noisy latent $x_{t}$ and the fake noisy latent $\\hat{x}_{t}$ are input into the diffusion discriminator to compute the loss $\\mathcal{L}_{\\mathcal{D}}$ :", + "bbox": [ + 511, + 641, + 906, + 717 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathcal {D}} = - \\mathbb {E} _ {\\hat {x} _ {0, t}, t \\sim [ 0, T ]} \\left[ \\log \\left(1 - \\mathcal {D} _ {\\phi} (\\hat {x} _ {0}, [ x _ {b g}, \\hat {x} _ {0}, m ], t)\\right)\\right) ] \\\\ - \\mathbb {E} _ {x _ {t}, t \\sim [ 0, T ]} \\left[ \\right. \\log \\mathcal {D} _ {\\phi} \\left(x _ {t}, \\left[ x _ {\\mathrm {b g}}, x _ {t}, m \\right], t\\right)\\left. \\right)\\left. \\right]. \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 729, + 903, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although the inpainting adapter optimized only with GAN loss can generate images, the generated images are of inferior quality and commonly incorporate unwanted elements. We attribute this issue to the limitations of the GAN loss, which rarely assists the discriminator in comprehending the image structure. In contrast, diffusion loss excels at scene understanding. Therefore, we introduce the fake diffusion loss to train the discriminator while allowing the", + "bbox": [ + 511, + 780, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7616", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "GAN loss to focus more on enhancing image textures and object details. To compute the diffusion loss for the fake noisy latent $\\hat{x}_t$ , we introduce an assistant decoder, following the assistant encoder. The fake diffusion loss is:", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {D i f f}} ^ {F} = \\min _ {\\mathcal {A} _ {\\psi}} \\mathbb {E} _ {y _ {t}, t \\sim [ 0, T ], \\epsilon \\sim \\mathcal {N} (0, 1)} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 157, + 482, + 184 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| \\epsilon - \\mathcal {A} _ {\\psi} ^ {\\mathrm {S D X L}} (x _ {t}, [ x _ {b g}, \\hat {x} _ {t}, m ], t) \\right\\| _ {2} ^ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 186, + 406, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{A}_{\\psi}$ represents the parameters of assistant encoder and decoder. The combination of $\\mathcal{L}_{\\mathrm{Diff}}^{F} + \\lambda_{3}\\mathcal{L}_{\\mathcal{D}}$ is used as the loss to optimize the assistant encoder, assistant decoder modules, and classifier.", + "bbox": [ + 89, + 220, + 483, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 292, + 223, + 310 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 89, + 318, + 289, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1.1. Training Dataset", + "text_level": 1, + "bbox": [ + 89, + 340, + 254, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enable editing through the input of a local mask and the corresponding prompt for the masked region, we construct a dataset that includes local masks and their associated textual descriptions. Specifically, we crawl approximately 1,200,000 images from Internet and use Florence-2 [25] to perform the dense region caption task with the prompt $<$ DENSE_REGION_CAPTION>. This task identifies the main objects in the images and generates their corresponding short descriptions. These descriptions are then input into SAM2 [19] to obtain the segmentation masks for the objects. We name this dataset as LocalCaptionData.", + "bbox": [ + 89, + 359, + 483, + 525 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1.2. Evaluation Datasets", + "text_level": 1, + "bbox": [ + 89, + 531, + 274, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The commonly used public datasets in the field of image inpainting are primarily OpenImage V6 [12], MSCOCO [14], and BrushBench [11]. Among these, OpenImage V6 and MSCOCO provide precise segmentation masks and bounding box masks, which are used to perform inpainting on the masked regions based on class names. BrushBench only offers accurate segmentation masks, where inpainting is performed on the masked regions using a global image description. However, considering that users typically perform inpainting with rough masks and corresponding descriptions (which may sometimes be complex and contain multiple attributes), these three datasets significantly diverge from real-world scenarios. To better simulate user operations, we construct two evaluation datasets:", + "bbox": [ + 89, + 549, + 483, + 760 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- DilationBench: The masks are created by randomly dilating the segmentation masks, and the prompts describe the content within the masked area of the original image. DilationBench contains 300 pairs of masks and prompts.", + "- HumanBench: The masks are manually labeled, and the prompts are manually written. HumanBench contains 150 pairs of masks and prompts." + ], + "bbox": [ + 89, + 763, + 482, + 868 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The descriptions and visualizations of the two benchmark datasets are included in the supplementary material.", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1.3. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 511, + 90, + 692, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate the quality of inpainted images, we use three metrics: Q-Align [23], CLIPIQA+ [22], TOPIQ [4]. They assess visual quality from different perspectives. Q-Align, which leverages a large language model (LLM) to generate visual scores, is state-of-the-art on various Image Quality Assessment (IQA) leaderboards1. Excluding Q-Align, CLIPIQA+ and TOPIQ rank first in non-reference and image aesthetic benchmarks, respectively, making them reliable indicators of visual quality and providing complementary insights. For a more comprehensive evaluation of inpainting results, we consider not only the masked region but also the entire image, ensuring that both local quality in the inpainted area and overall visual coherence across the entire image are assessed. To measure the consistency between the text and the content generated in the mask region, we use the CLIP Similarity metric [18]. Higher values for all four metrics indicate better performance.", + "bbox": [ + 511, + 109, + 906, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 378, + 720, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2.1. Network Configurations", + "text_level": 1, + "bbox": [ + 511, + 400, + 728, + 416 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inpainting adapter adopts the same architecture as Brush-Net, with modifications to the SDXL model. Specifically, the self-attention and cross-attention modules are removed. Note that our approach does not impose any architectural constraints on inpainting adapter. Assistant also adopts U-Net like structure, for simplicity, we directly adopt the same architecture as inpainting adapter. The classifier is composed of convolutional layers, GroupNorm [24], and SiLU [8] activation, which together map the feature map to a 1-dimensional vector. The details of classifier are described in the supplementary material.", + "bbox": [ + 511, + 420, + 905, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When using SDXL as the base model for noise prediction, we employ a DDIM scheduler [21] and sample randomly from 1,000 timesteps. In contrast, when using DMD2 as the base model, we use an LCM scheduler [15] and randomly sample from four specific timesteps: \\{999, 749, 499, 249\\}.", + "bbox": [ + 511, + 587, + 905, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 686, + 723, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "All experiments are performed on 8 A100 GPUs with 80GB memory. During training, the learning rate for parameter updates is set to 1e-5, with a batch size of 2 and gradient accumulation steps configured to 4. We adopt mixed precision to optimize computational efficiency. $\\lambda_{1},\\lambda_{2},\\lambda_{3}$ are set to $1e^{-3}$ , $1e^{-1}$ , and $1e^{-2}$ , respectively. All models are trained for 40K iterations using the AdamW optimizer.", + "bbox": [ + 511, + 705, + 905, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.3. Quantitative Comparisons", + "text_level": 1, + "bbox": [ + 511, + 823, + 754, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare TurboFill with existing state-of-the-art image inpainting methods, namely BLD [1], HD-Painter [16],", + "bbox": [ + 511, + 845, + 905, + 876 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "https://iqa-pytorch.readthedocs.io/en/latest/benchmark.html", + "bbox": [ + 535, + 887, + 852, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7617", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/96743521a278e94527a3f6d876b6c9df93ccc04878204fc3e48c35c7829b3dc4.jpg", + "image_caption": [ + "' tiger'", + "Figure 4. Comparison of previous inpainting methods and BrushNet on DilationBench. Compared to other methods, TurboFill generates more realistic details and textures in just 4 steps, while achieving good scene harmonization. (Zoom in for best view)" + ], + "image_footnote": [], + "bbox": [ + 109, + 69, + 885, + 647 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SDXL-Inpainting $^1$ , BrushNet-Rand [11], and PowerPaint [5]. All of these methods follow a multi-step approach, requiring 50 steps during inference to generate inpainting results. While BrushNet-Rand, trained with random masks, effectively handles imprecise masks and uses a full image description during inference, the other methods rely only on descriptions of the masked region.", + "bbox": [ + 89, + 696, + 482, + 804 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since we adopt the architecture design of BrushNet, for fair comparison, we also train BrushNet using LocalCaptionData, referring to this configuration as BrushNet*. Notably, BrushNet, as an external module, can be directly", + "bbox": [ + 89, + 804, + 482, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "adapted into a 4-step diffusion model (e.g., DMD2) to form a 4-step BrushNet configuration. In contrast, methods such as HD-Painter and SDXL-Inpainting, which modify the base model structure, cannot integrate with a 4-step diffusion model. PowerPaint V2, an advanced version of BrushNet and SD1.5, incorporates learnable task prompts and specialized fine-tuning strategies across four distinct inpainting tasks. However, since PowerPaint lacks an SDXL version, it cannot be adapted to a 4-step DMD2 model.", + "bbox": [ + 511, + 698, + 906, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparative results for these methods on both Dilation-Bench and HumanBench datasets are presented in Tab. 1 and Tab. 2. Given that HumanBench, like real inpainting tasks, does not provide a global description of the en", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "https://huggingface.co/diffusers/stable-diffusion-x1-1.0-inpainting-0.1", + "bbox": [ + 89, + 875, + 465, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7618", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a6f984202f298aa971368981aaef353f3bac92b642c309d3ceeec47bad019ce2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MetricsMask Region QualityWhole Image QualityText-Align
MethodQ-AlignCLIPIQA+TOPIQQ-AlignCLIPIQATOPIQCLIP Sim
MultiStep (50 steps)BLD4.18360.68745.17814.33200.67915.466224.884
HD-Painter3.97270.66114.93454.48050.64395.456625.585
SDXL-Inpainting4.24610.66665.15334.61720.66955.540924.848
BrushNet-Rand4.26170.67295.17384.62890.68185.577921.634
BrushNet*4.46880.71375.25044.53120.69745.558525.389
PowerPaint V24.77730.77655.54684.72270.73715.647426.256
FewStep (4 steps)BrushNet-Rand4.16020.65384.99274.55470.66085.459921.831
BrushNet*4.18360.65724.99394.44920.64275.396125.341
TurboFill4.57030.73325.27534.71880.71115.539225.352
", + "bbox": [ + 114, + 51, + 880, + 209 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/18b6a3d329ca076662f047aae81b0ad2d0af841758baa07700627ded261478b6.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons among TurboFill and other diffusion based inpainting models in DilationBench. Red and blue indicates the best multi-step and the best few-step performances, respectively." + ], + "table_footnote": [], + "table_body": "
MetricsMask Region QualityWhole Image QualityText-Align
MethodQ-AlignCLIPIQA+TOPIQQ-AlignCLIPIQATOPIQCLIP Sim
MultiStep (50 steps)BLD4.15230.69085.29963.90620.65795.419924.767
HD-Painter3.98440.64935.08734.20700.61195.438125.996
SDXL-Inpainting3.95510.64055.20404.04690.62885.456524.041
BrushNet*4.25780.70635.43404.08980.66945.548025.366
PowerPaint V24.55860.75295.54754.36330.71125.633826.264
FewStep (4 steps)BrushNet*4.05080.63275.10414.14840.60035.350525.473
Ours4.47270.72575.38654.32030.68225.499225.710
", + "bbox": [ + 114, + 253, + 880, + 385 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparisons among TurboFill and other diffusion based inpainting models in HumanBench. Red and blue indicates the best multi-step and the best few-step performances, respectively.", + "bbox": [ + 89, + 388, + 906, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tire image, we exclude BrushNet-Rand from evaluation on this dataset. From Tab. 1, we find that BrushNet*, trained with LocalCaptionData, shows improvements of 3.755 and 3.52 in the text-align metric for the 50-step and 4-step models, respectively, compared to BrushNet. This suggests that training with segmentation masks and paired local descriptions, rather than global descriptions, significantly enhances BrushNet's text alignment capability.", + "bbox": [ + 88, + 422, + 480, + 545 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, we observe that directly inserting BrushNet-Rand and BrushNet* into DMD2 to form a 4-step model results in no significant change in text alignment performance. However, the image quality of the final inpainting results decreases noticeably. This suggests that while replacing SDXL with a few-step model (DMD2) can accelerate BrushNet, it leads to a reduction in quality. In contrast, TurboFill can effectively mitigate this issue, showing significant improvements across all three metrics—mask region quality, whole image quality, and text alignment. Even with 4 steps, our method outperforms the 50-step BrushNet-Rand and BrushNet* in many image quality metrics (e.g., Q-Align, CLIPIQA+). This phenomenon can be observed on both benchmarks.", + "bbox": [ + 88, + 547, + 482, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "PowerPaint V2 currently achieves the best results across these metrics. When comparing our method with PowerPaint, we find that the gap in whole image quality is smaller than in mask image quality. Visual comparisons (Fig. 4) reveal that although PowerPaint V2 generates rich textures, the results in the mask region are less natural, with noticeable style mismatch compared to the background region. This inconsistency leads to a lower performance in whole image quality compared to mask region quality. In contrast,", + "bbox": [ + 88, + 763, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "our method produces more coherent results. Since no existing metrics effectively capture these differences, we conduct a user study, detailed in the supplementary file.", + "bbox": [ + 511, + 424, + 906, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Qualitative Comparison", + "text_level": 1, + "bbox": [ + 511, + 478, + 736, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The qualitative comparison on DilationBench is shown in Fig. 4. The results from SDXL-Inpainting exhibit a yellowish color bias (e.g., tiger, man) and lack detail (e.g., boy, figurine). For PowerPaint V2, the inpainted results are very sharp in texture and rich in color, producing high scores on most IQA metrics. However, these results suffer from severe distortions (e.g., a tiger with two heads, a person with two heads, overly elongated legs) and appear less realistic and natural (e.g., third row). The results generated by BrushNet* exhibit significant artifacts, such as the tiger's fur (1st) and the man's beard (5th), as well as the appearance of strange backgrounds (5th). When BrushNet* is inserted into the 4-step DMD2 model, we observe that the results lack significant detail (e.g., 5th row) and exhibit issues with oversaturation (e.g., tiger, figurine). Compared to these methods, TurboFill generates results with higher quality in visual comparisons. While preserving detail, our results do not exhibit the oversharpness seen in PowerPaint V2 (e.g., tiger), nor the oversaturation issue observed in BrushNet* (4-steps) (e.g., rabbit). The generated content is more realistic (e.g., bird, figurine), leading to a more harmonious image overall.", + "bbox": [ + 511, + 502, + 906, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. Effectiveness of LocalCaptionData", + "text_level": 1, + "bbox": [ + 511, + 847, + 815, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "BrushNet, when trained with a full image description (i.e., global prompt) as input to the base model, exhibits rela", + "bbox": [ + 511, + 869, + 905, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7619", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f0bb17df08fd73b19104df8417c40a990ad27f46606ea7e8f4fd01c95924f935.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingsMask Region QualityWhole Image QualityText-Align
Q-AlignCLIPQA+TOPIQQ-AlignCLIPQA+TOPIQCLIP Sim
TurboFill4.57030.73325.27534.72270.71115.539225.352
-LBG4.36720.68585.02624.54690.65705.382625.335
-LFDiff4.18750.65544.86954.49220.64265.329725.062
-RFDiff4.06640.63224.84974.44920.64125.306224.875
", + "bbox": [ + 207, + 51, + 787, + 137 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. The effectiveness of different losses. From top to down, we progressively remove specific losses.", + "bbox": [ + 184, + 141, + 810, + 156 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b7814c1e23257a81b8cdd09e33ba89392b4f0328dee92ecd698a1fd9a3db7b0e.jpg", + "image_caption": [ + "Figure 5. The effectiveness of LocalCaptionData. All results are obtained based on 4-step DMD2. (Zoom in for best view)" + ], + "image_footnote": [], + "bbox": [ + 96, + 169, + 473, + 503 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/188969e6392e51f795a6e1ffbabeb1e56a7d677838b1656d7fd63f8b1fb153c3.jpg", + "image_caption": [ + "Figure 6. The effectiveness of different losses. From left to right, we progressively remove specific losses. (Zoom in for best view)" + ], + "image_footnote": [], + "bbox": [ + 91, + 563, + 482, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tively weak text alignment capabilities, as shown in Fig. 5. In cases where BrushNet attempts to generate content similar or semantically related to existing objects in the image, it often either defaults to complete the background or shifts focus to other categories in the prompt (e.g., a cat). Furthermore, relying on a global prompt limits BrushNet's effectiveness in handling more complex prompts. By using LocalCaptionData, BrushNet* effectively overcomes these limitations without any changes to the training approach, resulting in a significant improvement in text alignment performance. Additionally, TurboFill demonstrates superior performance in generating fine details.", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.6. Ablation Studies.", + "text_level": 1, + "bbox": [ + 511, + 169, + 681, + 184 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Starting from TurboFill, we remove $\\mathcal{L}_{\\mathrm{BG}}$ , $\\mathcal{L}_{\\mathrm{Diff}}^F$ , and $\\mathcal{L}_{\\mathrm{Diff}}^R$ in sequence, with quantitative and qualitative results shown in Fig. 6 and Tab. 3. When $\\mathcal{L}_{\\mathrm{BG}}$ is removed, there is a clear performance drop in all metrics. In the visualizations, we see that without $\\mathcal{L}_{\\mathrm{BG}}$ , the color of background region changes noticeably, creating a sharp boundary between the fill-in and background regions. Further removing $\\mathcal{L}_{\\mathrm{Diff}}^F$ introduces conflicting elements (i.e., house) in the fill-in region, suggesting the discriminator fails to fully capture the holistic scene. Finally, without $\\mathcal{L}_{\\mathrm{Diff}}^R$ , relying only on GAN loss, the inpainted images exhibit not only inconsistencies with the background but also poor texture and detail. This highlights that GAN loss alone struggles to close the gap between fake and real latents. Only when combining $\\mathcal{L}_{\\mathrm{Diff}}^F$ , $\\mathcal{L}_{\\mathrm{Diff}}^R$ , and $\\mathcal{L}_{\\mathrm{BG}}$ in GAN training does the model achieve enhanced texture, detail, and effective scene harmonization between fill-in and background regions.", + "bbox": [ + 511, + 190, + 906, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 453, + 633, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce TurboFill, a fast inpainting model designed to address the computational and quality challenges in diffusion-based inpainting tasks. By directly training an inpainting adapter on a few-step distilled text-to-image diffusion model, TurboFill achieves state-of-the-art inpainting results with significantly smaller inference costs. The proposed three-step adversarial training scheme, incorporating both GAN and diffusion losses, ensures the generation of realistic and harmonious content with fine textures and structural coherence. To evaluate TurboFill's performance, we introduce two new benchmarks, DilationBench and HumanBench, which provide robust evaluations across varying mask complexities and user-centric scenarios. Experimental results demonstrate TurboFill's ability to outperform existing few-step and multi-step BrushNet, setting a new benchmark for practical and efficient inpainting.", + "bbox": [ + 511, + 478, + 906, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgement", + "text_level": 1, + "bbox": [ + 511, + 724, + 692, + 741 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by Macau Science and Technology Development Fund under SKLIOTSC-2021-2023, 0022/2022/A1, and 0014/2022/AFJ; in part by Research Committee at University of Macau under MYRG-GRG2023-00058-FST-UMDF; in part by the Guangdong Basic and Applied Basic Research Foundation under Grant 2024A1515012536; in part by National Natural Science Foundation of China (Grant No. 62276251), and the Joint Lab of Shenzhen Key Laboratory of Computer Vision and Pattern Recognition.", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7620", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM transactions on graphics (TOG), 42 (4):1-11, 2023. 1, 3, 5", + "[2] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Trans. Graph., 28(3):24, 2009. 3", + "[3] Chenjie Cao, Yunuo Cai, Qiaole Dong, Yikai Wang, and Yanwei Fu. Leftrefill: Filling right canvas based on left reference through generalized text-to-image diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7705-7715, 2024. 3", + "[4] Chaofeng Chen, Jiadi Mo, Jingwen Hou, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Topiq: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing, 2024. 5", + "[5] Li Chen, Yu Wang, Jiali Zhang, and Xingxing Xu. Powerpaint: Adaptive task prompts for generalizable image inpainting. arXiv preprint arXiv:2401.08965, 2024. 1, 3, 6", + "[6] Yifu Chen, Jingwen Chen, Yingwei Pan, Yehao Li, Ting Yao, Zhineng Chen, and Tao Mei. Improving text-guided object inpainting with semantic pre-inpainting. In European Conference on Computer Vision, pages 110-126. Springer, 2025. 3", + "[7] Huggingface Diffusers. Sdxl-inpaint. https : / / huggingface . co / diffusers / stable - diffusion-x1- 1 . 0 - inpainting- 0 . 1, Year. Accessed: YYYYY-MM-DD. 1,3", + "[8] Stefan Elfwing, Eiji Uchibe, and Kenji Doya. Sigmoid-weighted linear units for neural network function approximation in reinforcement learning. Neural networks, 107:3-11, 2018. 5", + "[9] Omar Elharrouss, Noor Almaadeed, Somaya Al-Maadeed, and Younes Akbari. Image inpainting: A review. Neural Processing Letters, 51:2007-2028, 2020. 3", + "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems (NeurIPS), pages 6840-6851, 2020. 4", + "[11] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. arXiv preprint arXiv:2403.06976, 2024. 1, 3, 5, 6", + "[12] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision, 128(7):1956-1981, 2020. 5", + "[13] Shanchuan Lin, Anran Wang, and Xiao Yang. Sdxlightning: Progressive adversarial diffusion distillation. arXiv preprint arXiv:2402.13929, 2024. 2, 3", + "[14] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5", + "[15] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 2, 3, 4, 5", + "[16] Hayk Manukyan, Andranik Sargsyan, Barsegh Atanyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Hd-painter: high-resolution and prompt-faithful text-guided image inpainting with diffusion models. arXiv preprint arXiv:2312.14091, 2023. 1, 3, 5", + "[17] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 3", + "[18] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5", + "[19] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 5", + "[20] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 2, 3", + "[21] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 5", + "[22] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 5", + "[23] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching Imms for visual scoring via discrete text-defined levels. arXiv preprint arXiv:2312.17090, 2023. 5", + "[24] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 5", + "[25] Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, and Lu Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4818–4829, 2024. 5", + "[26] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. arXiv preprint arXiv:2311.18828, 2023. 2, 3" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7621", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. arXiv preprint arXiv:2405.14867, 2024. 2", + "[28] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contextual attention. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5505-5514, 2018. 3", + "[29] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4471-4480, 2019.", + "[30] Yanhong Zeng, Jianlong Fu, Hongyang Chao, and Baining Guo. Learning pyramid-context encoder network for high-quality image inpainting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1486-1494, 2019.", + "[31] Yanhong Zeng, Jianlong Fu, Hongyang Chao, and Baining Guo. Aggregated contextual transformations for high-resolution image inpainting. IEEE Transactions on Visualization and Computer Graphics, 29(7):3266-3280, 2022. 3", + "[32] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 3", + "[33] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 3" + ], + "bbox": [ + 91, + 92, + 480, + 513 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7622", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/TurboFill_ Adapting Few-step Text-to-image Model for Fast Image Inpainting/6302d669-b0e3-4a35-adf2-a42f09f692e9_model.json b/2025/TurboFill_ Adapting Few-step Text-to-image Model for Fast Image Inpainting/6302d669-b0e3-4a35-adf2-a42f09f692e9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..545fcc6f9a05e64f0f53fd022282c5d91f2c8dfe --- /dev/null +++ b/2025/TurboFill_ Adapting Few-step Text-to-image Model for Fast Image Inpainting/6302d669-b0e3-4a35-adf2-a42f09f692e9_model.json @@ -0,0 +1,2167 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.109, + 0.131, + 0.892, + 0.154 + ], + "angle": 0, + "content": "TurboFill: Adapting Few-step Text-to-image Model for Fast Image Inpainting" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.18, + 0.882, + 0.217 + ], + "angle": 0, + "content": "Liangbin Xie\\(^{1,4}\\) Daniil Pakhomov\\(^{3}\\) Zhonghao Wang\\(^{3}\\) Zongze Wu\\(^{3}\\) Ziyan Chen\\(^{4}\\) Yuqian Zhou\\(^{3}\\) Haitian Zheng\\(^{3}\\) Zhifei Zhang\\(^{3}\\) Zhe Lin\\(^{3}\\) Jiantao Zhou\\(^{1\\dagger}\\) Chao Dong\\(^{2,4\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.219, + 0.88, + 0.251 + ], + "angle": 0, + "content": "\\(^{1}\\)State Key Laboratory of Internet of Things for Smart City, University of Macau \\(^{2}\\)Shenzhen University of Advanced Technology \\(^{3}\\)Adobe \\(^{4}\\)Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.267, + 0.394, + 0.277 + ], + "angle": 0, + "content": "Yellow baby chick on wooden surface." + }, + { + "type": "text", + "bbox": [ + 0.608, + 0.267, + 0.865, + 0.277 + ], + "angle": 0, + "content": "A miniature blue convertible with its top down" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.279, + 0.226, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.387, + 0.192, + 0.398 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.279, + 0.363, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.244, + 0.387, + 0.347, + 0.398 + ], + "angle": 0, + "content": "BrushNet* (50 steps)" + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.279, + 0.501, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.387, + 0.48, + 0.398 + ], + "angle": 0, + "content": "TurboFill (4 steps)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.279, + 0.636, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.387, + 0.596, + 0.398 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.279, + 0.772, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.387, + 0.755, + 0.398 + ], + "angle": 0, + "content": "BrushNet* (50 steps)" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.279, + 0.908, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.793, + 0.387, + 0.888, + 0.398 + ], + "angle": 0, + "content": "TurboFill (4 steps)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.406, + 0.914, + 0.416 + ], + "angle": 0, + "content": "A fluffy brown teddy bear wearing a tiny hat. A tinted sunglasses lens reflecting tall palm trees. Man with cowboy hat and bandana in front of wooden wall." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.419, + 0.227, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.526, + 0.191, + 0.538 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.419, + 0.364, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.247, + 0.526, + 0.343, + 0.538 + ], + "angle": 0, + "content": "TurboFill (4 steps)" + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.419, + 0.499, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.4, + 0.526, + 0.461, + 0.538 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.419, + 0.636, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.526, + 0.616, + 0.538 + ], + "angle": 0, + "content": "TurboFill (4 steps)" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.419, + 0.772, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.526, + 0.732, + 0.538 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.419, + 0.908, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.793, + 0.526, + 0.888, + 0.538 + ], + "angle": 0, + "content": "TurboFill (4 steps)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.545, + 0.908, + 0.574 + ], + "angle": 0, + "content": "Figure 1. We propose TurboFill, a fast image inpainting method that leverages a 3-step adversarial training scheme. With only four diffusion steps, TurboFill outperforms the multi-step BrushNet* [11], delivering realistic details and textures with remarkable efficiency." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.588, + 0.327, + 0.604 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.62, + 0.486, + 0.862 + ], + "angle": 0, + "content": "This paper introduces TurboFill, a fast image inpainting model that enhances a few-step text-to-image diffusion model with an inpainting adapter for high-quality and efficient inpainting. While standard diffusion models generate high-quality results, they incur high computational costs. We overcome this by training an inpainting adapter on a few-step distilled text-to-image model, DMD2, using a novel 3-step adversarial training scheme to ensure realistic, structurally consistent, and visually harmonious inpainted regions. To evaluate TurboFill, we propose two benchmarks: DilationBench, which tests performance across mask sizes, and HumanBench, based on human feedback for complex prompts. Experiments show that TurboFill outperforms both multi-step BrushNet and few-step inpainting methods, setting a new benchmark for high-performance inpainting tasks. The project page is available here." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.588, + 0.645, + 0.604 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.907, + 0.704 + ], + "angle": 0, + "content": "Image inpainting, the task of filling missing regions in an image, has seen significant advancement with the rise of deep generative models [1, 5, 11, 16]. Existing diffusion-based inpainting methods have shown promising results for various applications such as object removal, regeneration, and text-guided object insertion." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.705, + 0.909, + 0.902 + ], + "angle": 0, + "content": "The most straightforward way of training diffusion-based inpainting models is to add the masked image (background conditioning) to the input channels of a pretrained diffusion model and finetune it with inpainting data, as shown by SDXL-inpaint [7]. However, such methods typically require fine-tuning the entire network, entailing large training datasets and significant computational resources. In contrast, adapter-based approaches, like BrushNet [11], utilize a ControlNet [33] to capture conditioning information and inject features into a frozen diffusion U-Net. BrushNet achieves state-of-the-art inpainting results while incurring relatively low training costs in terms of data and computational resources, owing to the adapter's efficiency and the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.116, + 0.887, + 0.245, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7613" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.053, + 0.904, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.271, + 0.907, + 0.342 + ], + "angle": 0, + "content": "Figure 2. (Zoom in for best view) 1. The multi-step adapter achieves high-quality inpainting results but incurs significant inference costs, requiring over 50 diffusion steps. 2. Applying the pre-trained multi-step BrushNet adapter directly to the few-step U-Net (DMD2) results in artifacts, including oversaturated colors and semantic inconsistencies (e.g., generating a dog with two tails). 3. Training the adapter-DMD2 solely with diffusion loss produces blurred outputs with low-quality inpainting results. 4. In contrast, training the adapter-DMD2 using the proposed 3-step adversarial training scheme yields high-quality inpainting results, requiring only four diffusion steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.462, + 0.373 + ], + "angle": 0, + "content": "preservation of the U-Net's original generation capacity." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.373, + 0.483, + 0.538 + ], + "angle": 0, + "content": "Despite the advantage of BrushNet-like models, they still depend on many iterative sampling steps, leading to substantial inference cost in practice. Interestingly, we found that an inpainting adapter trained on a pretrained text-to-image diffusion model can be seamlessly integrated into its corresponding few-step student model without further training, while yielding reasonable quality results (Fig. 2 ②). This can be attributed to the semantic alignment between the teacher and student models. However, this naive migration approach exhibits quality issues such as color shifts, over-saturation, and loss of details." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.539, + 0.483, + 0.674 + ], + "angle": 0, + "content": "Alternatively, we can use a two-stage approach by first training a teacher BrushNet and then applying advanced diffusion distillation techniques [13, 15, 20, 26, 27] to the teacher BrushNet to reduce steps while maintaining quality. Although recent distillation methods for diffusion models have demonstrated remarkable performance, we have found that due to BrushNet's large model size, it demands excessive memory and incurs high computational costs, making the training process difficult to manage." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.856 + ], + "angle": 0, + "content": "In this paper, we propose a simpler single-stage training strategy by training an inpainting adapter directly on top of a pretrained few-step distilled text-to-image model (e.g., DMD2). We observe that naively training an inpainting adapter and few-step model solely with diffusion loss results in blurry, low-quality outputs (Fig. 2③). To address these shortcomings and enhance inpainting performance, we introduce a diffusion discriminator to guide optimization of the adapter-DMD2 generator. The diffusion discriminator is trained to differentiate between noisy real images and fake latents, while a combined diffusion loss ensures the model learns meaningful semantic representations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In summary, TurboFill is trained with a novel 3-step adversarial training scheme with two generator update steps and one discriminator update step. Our model enables high-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.907, + 0.479 + ], + "angle": 0, + "content": "quality image inpainting with only a few diffusion steps. Compared to the naive approach of training a multi-step adapter (requiring 8 V100 GPUs for 72 hours) followed by distillation (requiring 64 A100 GPUs for 50 hours), TurboFill attains comparable outcomes using merely 8 A100 GPUs within 50 hours. This accomplishment marks a reduction in training time by over a factor of 10 and a substantial enhancement in training efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.481, + 0.909, + 0.634 + ], + "angle": 0, + "content": "Existing benchmarks often fall short in simulating real-world user operations. To address this gap, we introduce two new benchmarks: DilationBench and HumanBench. DilationBench provides an objective measure by testing the model on the masks with varying dilation sizes, while HumanBench incorporates human user feedback and complex prompts, offering insights into real-world performance. These benchmarks contribute a robust evaluation framework that can facilitate advancements in inpainting models and more accurately reflect user-oriented performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.636, + 0.907, + 0.711 + ], + "angle": 0, + "content": "Extensive quantitative and qualitative evaluations demonstrate that TurboFill not only supports complex prompts and user-driven inpainting tasks but also outperforms existing few-step inpainting models in terms of speed and quality. In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.72, + 0.907, + 0.794 + ], + "angle": 0, + "content": "- A novel 3-step adversarial training scheme is proposed to train an inpainting adapter directly on top of few-step text-to-image diffusion model, which helps the model to focus on image quality and improves textures, details, and scene harmonization." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.796, + 0.905, + 0.839 + ], + "angle": 0, + "content": "- New benchmarks, DilationBench and HumanBench, are designed to better evaluate inpainting methods in user-centered scenarios." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "- We demonstrate that our few-step TurboFill outperforms multi-step BrushNet for the inpainting task and achieves state-of-the-art performance in a few-step setting through extensive experiments." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7614" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.233, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.341 + ], + "angle": 0, + "content": "Image inpainting, a classic problem in computer vision, focuses on reconstructing missing regions in an image with realistic and natural content. Early methods, including traditional techniques, VAEs, and GANs [2, 3, 9, 28-31], often struggled to produce high-quality results. Recently, diffusion models have gained popularity for their ability to generate high-quality, diverse outputs with fine-grained control, demonstrating significant advancements over earlier approaches [1, 5, 11, 16]. Diffusion models, particularly the Stable Diffusion family, have introduced a robust framework for image generation and inpainting by iteratively denoising latent representations. Stable Diffusion-based methods have achieved impressive results in high-resolution image inpainting, yet are computationally intensive due to multi-step sampling." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.484, + 0.584 + ], + "angle": 0, + "content": "Inpainting tasks based on diffusion models generally follow two approaches: training an entire diffusion model or incorporating external modules. The first approach, like PowerPaint [5], CATDiffusion [6], and SDXL-Inpainting, focuses on adapting the model itself. PowerPaint introduces learnable task prompts along with tailored fine-tuning strategies to guide the model's focus on specific inpainting targets, while SDXL-Inpainting [7] directly utilizes both the background region and the mask as inputs to fine-tune a new model. In contrast, the second approach, exemplified by methods such as ControlNet [33] or T2I-Adapter [17], leverages external modules to enhance the diffusion model's capabilities without retraining it entirely. Recent methods, such as BrushNet [11], employ mask-specific prompts, refining the generated content within the masked area for improved coherence with the background." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.484, + 0.721 + ], + "angle": 0, + "content": "Our method builds on these developments by leveraging a ControlNet-like adapter for DMD2, integrating both GAN and diffusion losses to enhance image realism and style coherence. By focusing on masked-region-specific prompts and incorporating fine-grained diffusion control, our approach achieves fast and high-quality inpainting results that align closely with user inputs and surrounding context, advancing upon prior work in speed, quality, and flexibility in real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.734, + 0.218, + 0.751 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.759, + 0.483, + 0.79 + ], + "angle": 0, + "content": "The forward diffusion process gradually adds Gaussian noise \\(\\epsilon_t\\) to a clean image \\(x_0\\) by" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.804, + 0.483, + 0.821 + ], + "angle": 0, + "content": "\\[\nx _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.828, + 0.484, + 0.874 + ], + "angle": 0, + "content": "where \\(\\bar{\\alpha}_t\\) controls noise schedule. In the training phase, a denoising network \\(\\epsilon_{\\theta}\\) is trained to predict \\(\\epsilon_t\\) given \\(x_t\\), timestep \\(t\\) and text prompt \\(c\\), with the diffusion loss:" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.884, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\nL \\left(\\hat {\\epsilon} _ {\\theta}\\right) = \\mathbb {E} _ {x _ {0} \\sim q; \\epsilon_ {t} \\sim \\mathcal {N} (0, 1)} [ \\| \\epsilon_ {\\theta} (x _ {t}, c, t) - \\epsilon_ {t} \\| ^ {2} ]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.166 + ], + "angle": 0, + "content": "To enable condition generation, adapter-based methods [11, 17, 32] introduce additional conditioning information via a lightweight adapter, and achieve efficient training (since the main diffusion network remains frozen), with a slight inference overhead." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.907, + 0.334 + ], + "angle": 0, + "content": "Traditionally, reverse diffusion processes necessitate a computationally intensive multi-step iterative sampling procedure (typically 20-50 steps). Distillation techniques [13, 15, 20, 26] have been developed to reduce this requirement, achieving accelerated sampling with as few as 1-4 steps. However, these distillation processes are resource-intensive in both computational power and time. In contrast, our approach directly trains the adapter with both the multi-step and few-step text-to-image diffusion model, eliminating the need for a separate distillation process, thereby enabling more efficient training and inference." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.347, + 0.662, + 0.364 + ], + "angle": 0, + "content": "4. Our Approach" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.371, + 0.907, + 0.644 + ], + "angle": 0, + "content": "In this work, we propose a fast inpainting model, TurboFill, which directly trains an inpainting adapter on top of a few-step text-to-image diffusion model. As shown in Fig. 3, there are three components: the fast generator, the slow generator and the diffusion discriminator. All these three components aim to enhance the adapter's capability of inpainting masked images. Specifically, the slow generator helps the adapter to denoise the noisy latents towards the most probable direction of realistic image generation. The fast generator enables the adapter to generate a clean image for critique during training. The diffusion discriminator guides the adapter to generate images with better textures, details and content harmonization. Based on these three components, TurboFill follows a 3-step adversarial training scheme. The training alternates between 3 steps throughout the training process. In the following sections, we provide a detailed description of the training procedure for each step and the components." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.653, + 0.859, + 0.669 + ], + "angle": 0, + "content": "4.1. Training the Adapter in Slow Generator" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To capture conditioning information and introduce additional features to the diffusion model, we utilize an inpainting adapter with the same ControlNet architecture as BrushNet [11]. Given that SDXL is a multi-step text-to-image generator, we designate the integration of the inpainting adapter and SDXL as the 'slow generator'. In contrast, DMD2 is a few-step text-to-image generator, so the combination of the inpainting adapter and DMD2 is named the 'fast generator'. In both generators, conditioning information is introduced by concatenating the noisy latent \\( x_{t} \\), the background image latent \\( x_{bg} \\), and a downsampled binary mask \\( m \\). This concatenated input is then processed by the inpainting adapter, while \\( x_{t} \\) is simultaneously fed into the UNet. Then, the features produced by the inpainting adapter are fused with the UNet's features via residual connection." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7615" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.053, + 0.887, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.301, + 0.907, + 0.345 + ], + "angle": 0, + "content": "Figure 3. The training of TurboFill alternates between 3 steps: (1) optimizing the adapter using the gradient of \\(\\mathcal{L}_{\\mathrm{Diff}}^R\\), and (2) \\(\\mathcal{L}_{\\mathcal{G}}\\) and \\(\\mathcal{L}_{\\mathrm{BG}}\\) are employed to update the adapter in fast generator, and (3) \\(\\mathcal{L}_{\\mathrm{Diff}}^F\\) and \\(\\mathcal{L}_{\\mathcal{D}}\\) are jointly applied to update the parameters of the diffusion discriminator module. Note that the adapters in slow generator and fast generator share same weights during training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.357, + 0.483, + 0.402 + ], + "angle": 0, + "content": "As shown in Fig. 3 (1), the parameters of the inpainting adapter in slow generator are updated based on the real diffusion loss:" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.411, + 0.482, + 0.434 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {D i f f}} ^ {R} = \\min _ {\\epsilon_ {\\theta}} \\mathbb {E} _ {x _ {t}, t \\sim [ 0, T ], \\epsilon \\sim \\mathcal {N} (0, 1)} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.438, + 0.401, + 0.468 + ], + "angle": 0, + "content": "\\[\n\\left\\| \\epsilon - \\epsilon_ {\\theta} ^ {\\mathrm {S D X L}} (x _ {t}, [ x _ {b g}, x _ {t}, m ], t) \\right\\| _ {2} ^ {2},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.474, + 0.483, + 0.504 + ], + "angle": 0, + "content": "where \\( T = 1,000 \\) and DDPM scheduler [10] is employed to generate \\( x_{t} \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.512, + 0.431, + 0.529 + ], + "angle": 0, + "content": "4.2. Training the Adapter in Fast Generator" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.484, + 0.641 + ], + "angle": 0, + "content": "Given that adversarial training is highly effective in minimizing the divergence between different distributions, we employ adversarial training to minimize the discrepancy between real latent distribution and the fake latent distribution \\(\\{\\hat{x}_0\\}\\). Specifically, we first generate the clean fake latent \\(\\hat{x}_0\\) from the noisy latent \\(x_{t}\\) with the fast generator in a single step, as described by" + }, + { + "type": "equation", + "bbox": [ + 0.111, + 0.648, + 0.483, + 0.682 + ], + "angle": 0, + "content": "\\[\n\\hat {x} _ {0} = \\frac {x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\times \\epsilon_ {\\theta} ^ {\\mathrm {D M D 2}} \\left(x _ {t} , \\left[ x _ {b g} , x _ {t} , m \\right] , t\\right)}{\\sqrt {\\bar {\\alpha} _ {t}}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.764 + ], + "angle": 0, + "content": "where \\(\\bar{\\alpha}_t\\) is determined by the LCM scheduler [15] and \\(t\\) is the sampled timestep. Then, the noisy version \\(\\hat{x}_t\\) originated from \\(\\hat{x_0}\\) is fed into the diffusion discriminator. Meanwhile, to preserve the background region, \\(\\hat{x_0}\\) is used to compute the background loss." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.764, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Diffusion discriminator. As shown in Fig. 3 (2), the diffusion discriminator \\(\\mathcal{D}_{\\phi}\\) consists of the SDXL encoder, the assistant encoder, and the convolution-based classifier. For assistant encoder, we use the same encoder structure of inpainting adapter. It takes concatenated noisy latent, downsampled binary mask, and masked image latent \\(x_{bg}\\) as the input. The features extracted by the assistant encoder are progressively integrated into the SDXL encoder. The final feature maps are then passed through the classifier and" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.357, + 0.907, + 0.388 + ], + "angle": 0, + "content": "mapped to a one-dimensional vector. Based on the vector, we calculate \\(\\mathcal{L}_{\\mathcal{G}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.399, + 0.905, + 0.435 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {G}} = \\min _ {\\epsilon_ {\\theta}} - \\mathbb {E} _ {y _ {t}, t \\sim [ 0, T ]} \\left[ \\log \\mathcal {D} _ {\\phi} \\left(\\hat {x} _ {0}, \\left[ x _ {b g}, \\hat {x} _ {0}, m \\right], t\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.436, + 0.641, + 0.45 + ], + "angle": 0, + "content": "where \\(T = 1,000\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.451, + 0.906, + 0.526 + ], + "angle": 0, + "content": "Background preservation. For the image inpainting task, it is essential to preserve the background region. Therefore, for the clean fake latent \\( \\hat{x_0} \\), we apply a reconstruction loss to the background region to minimize the distance between \\( \\hat{x_0} \\) and the clean latent \\( x_0 \\):" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.54, + 0.905, + 0.559 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {B G}} = \\left\\| x _ {0} \\odot m - \\hat {x} _ {0} \\odot m \\right\\| _ {2} ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.565, + 0.907, + 0.611 + ], + "angle": 0, + "content": "where \\(\\odot\\) represents point-wise multiplication. Based on \\(\\lambda_1\\mathcal{L}_{\\mathcal{G}} + \\lambda_2\\mathcal{L}_{\\mathrm{BG}}\\), in this step, only the parameters of the inpainting adapter are updated." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.62, + 0.833, + 0.636 + ], + "angle": 0, + "content": "4.3. Training the Diffusion Discriminator" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.907, + 0.718 + ], + "angle": 0, + "content": "When updating the discriminator, we fix the parameters of the inpainting adapter and update the parameters of the assistant encoder and classifier. Specifically, both the real noisy latent \\( x_{t} \\) and the fake noisy latent \\( \\hat{x}_{t} \\) are input into the diffusion discriminator to compute the loss \\( \\mathcal{L}_{\\mathcal{D}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.73, + 0.905, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathcal {D}} = - \\mathbb {E} _ {\\hat {x} _ {0, t}, t \\sim [ 0, T ]} \\left[ \\log \\left(1 - \\mathcal {D} _ {\\phi} (\\hat {x} _ {0}, [ x _ {b g}, \\hat {x} _ {0}, m ], t)\\right)\\right) ] \\\\ - \\mathbb {E} _ {x _ {t}, t \\sim [ 0, T ]} \\left[ \\right. \\log \\mathcal {D} _ {\\phi} \\left(x _ {t}, \\left[ x _ {\\mathrm {b g}}, x _ {t}, m \\right], t\\right)\\left. \\right)\\left. \\right]. \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Although the inpainting adapter optimized only with GAN loss can generate images, the generated images are of inferior quality and commonly incorporate unwanted elements. We attribute this issue to the limitations of the GAN loss, which rarely assists the discriminator in comprehending the image structure. In contrast, diffusion loss excels at scene understanding. Therefore, we introduce the fake diffusion loss to train the discriminator while allowing the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7616" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "GAN loss to focus more on enhancing image textures and object details. To compute the diffusion loss for the fake noisy latent \\(\\hat{x}_t\\), we introduce an assistant decoder, following the assistant encoder. The fake diffusion loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.158, + 0.483, + 0.185 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {D i f f}} ^ {F} = \\min _ {\\mathcal {A} _ {\\psi}} \\mathbb {E} _ {y _ {t}, t \\sim [ 0, T ], \\epsilon \\sim \\mathcal {N} (0, 1)} \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.187, + 0.407, + 0.217 + ], + "angle": 0, + "content": "\\[\n\\left\\| \\epsilon - \\mathcal {A} _ {\\psi} ^ {\\mathrm {S D X L}} (x _ {t}, [ x _ {b g}, \\hat {x} _ {t}, m ], t) \\right\\| _ {2} ^ {2},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.222, + 0.485, + 0.282 + ], + "angle": 0, + "content": "where \\(\\mathcal{A}_{\\psi}\\) represents the parameters of assistant encoder and decoder. The combination of \\(\\mathcal{L}_{\\mathrm{Diff}}^{F} + \\lambda_{3}\\mathcal{L}_{\\mathcal{D}}\\) is used as the loss to optimize the assistant encoder, assistant decoder modules, and classifier." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.294, + 0.224, + 0.311 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.319, + 0.29, + 0.333 + ], + "angle": 0, + "content": "5.1. Datasets and Metrics" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.341, + 0.255, + 0.357 + ], + "angle": 0, + "content": "5.1.1. Training Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.36, + 0.484, + 0.526 + ], + "angle": 0, + "content": "To enable editing through the input of a local mask and the corresponding prompt for the masked region, we construct a dataset that includes local masks and their associated textual descriptions. Specifically, we crawl approximately 1,200,000 images from Internet and use Florence-2 [25] to perform the dense region caption task with the prompt \\(< \\) DENSE_REGION_CAPTION>. This task identifies the main objects in the images and generates their corresponding short descriptions. These descriptions are then input into SAM2 [19] to obtain the segmentation masks for the objects. We name this dataset as LocalCaptionData." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.532, + 0.276, + 0.545 + ], + "angle": 0, + "content": "5.1.2. Evaluation Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.55, + 0.484, + 0.761 + ], + "angle": 0, + "content": "The commonly used public datasets in the field of image inpainting are primarily OpenImage V6 [12], MSCOCO [14], and BrushBench [11]. Among these, OpenImage V6 and MSCOCO provide precise segmentation masks and bounding box masks, which are used to perform inpainting on the masked regions based on class names. BrushBench only offers accurate segmentation masks, where inpainting is performed on the masked regions using a global image description. However, considering that users typically perform inpainting with rough masks and corresponding descriptions (which may sometimes be complex and contain multiple attributes), these three datasets significantly diverge from real-world scenarios. To better simulate user operations, we construct two evaluation datasets:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.764, + 0.483, + 0.824 + ], + "angle": 0, + "content": "- DilationBench: The masks are created by randomly dilating the segmentation masks, and the prompts describe the content within the masked area of the original image. DilationBench contains 300 pairs of masks and prompts." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.824, + 0.483, + 0.869 + ], + "angle": 0, + "content": "- HumanBench: The masks are manually labeled, and the prompts are manually written. HumanBench contains 150 pairs of masks and prompts." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.764, + 0.483, + 0.869 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The descriptions and visualizations of the two benchmark datasets are included in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.092, + 0.694, + 0.106 + ], + "angle": 0, + "content": "5.1.3. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.111, + 0.907, + 0.368 + ], + "angle": 0, + "content": "To evaluate the quality of inpainted images, we use three metrics: Q-Align [23], CLIPIQA+ [22], TOPIQ [4]. They assess visual quality from different perspectives. Q-Align, which leverages a large language model (LLM) to generate visual scores, is state-of-the-art on various Image Quality Assessment (IQA) leaderboards1. Excluding Q-Align, CLIPIQA+ and TOPIQ rank first in non-reference and image aesthetic benchmarks, respectively, making them reliable indicators of visual quality and providing complementary insights. For a more comprehensive evaluation of inpainting results, we consider not only the masked region but also the entire image, ensuring that both local quality in the inpainted area and overall visual coherence across the entire image are assessed. To measure the consistency between the text and the content generated in the mask region, we use the CLIP Similarity metric [18]. Higher values for all four metrics indicate better performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.379, + 0.722, + 0.395 + ], + "angle": 0, + "content": "5.2. Experimental Settings" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.401, + 0.729, + 0.417 + ], + "angle": 0, + "content": "5.2.1. Network Configurations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.421, + 0.906, + 0.587 + ], + "angle": 0, + "content": "Inpainting adapter adopts the same architecture as Brush-Net, with modifications to the SDXL model. Specifically, the self-attention and cross-attention modules are removed. Note that our approach does not impose any architectural constraints on inpainting adapter. Assistant also adopts U-Net like structure, for simplicity, we directly adopt the same architecture as inpainting adapter. The classifier is composed of convolutional layers, GroupNorm [24], and SiLU [8] activation, which together map the feature map to a 1-dimensional vector. The details of classifier are described in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.588, + 0.906, + 0.68 + ], + "angle": 0, + "content": "When using SDXL as the base model for noise prediction, we employ a DDIM scheduler [21] and sample randomly from 1,000 timesteps. In contrast, when using DMD2 as the base model, we use an LCM scheduler [15] and randomly sample from four specific timesteps: \\{999, 749, 499, 249\\}." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.687, + 0.724, + 0.702 + ], + "angle": 0, + "content": "5.2.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.707, + 0.906, + 0.813 + ], + "angle": 0, + "content": "All experiments are performed on 8 A100 GPUs with 80GB memory. During training, the learning rate for parameter updates is set to 1e-5, with a batch size of 2 and gradient accumulation steps configured to 4. We adopt mixed precision to optimize computational efficiency. \\(\\lambda_{1},\\lambda_{2},\\lambda_{3}\\) are set to \\(1e^{-3}\\), \\(1e^{-1}\\), and \\(1e^{-2}\\), respectively. All models are trained for 40K iterations using the AdamW optimizer." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.824, + 0.755, + 0.84 + ], + "angle": 0, + "content": "5.3. Quantitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.846, + 0.906, + 0.877 + ], + "angle": 0, + "content": "We compare TurboFill with existing state-of-the-art image inpainting methods, namely BLD [1], HD-Painter [16]," + }, + { + "type": "page_footnote", + "bbox": [ + 0.537, + 0.888, + 0.854, + 0.901 + ], + "angle": 0, + "content": "https://iqa-pytorch.readthedocs.io/en/latest/benchmark.html" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7617" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.462, + 0.056, + 0.528, + 0.069 + ], + "angle": 0, + "content": "' tiger'" + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.07, + 0.887, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.657, + 0.907, + 0.687 + ], + "angle": 0, + "content": "Figure 4. Comparison of previous inpainting methods and BrushNet on DilationBench. Compared to other methods, TurboFill generates more realistic details and textures in just 4 steps, while achieving good scene harmonization. (Zoom in for best view)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.698, + 0.483, + 0.805 + ], + "angle": 0, + "content": "SDXL-Inpainting\\(^1\\), BrushNet-Rand [11], and PowerPaint [5]. All of these methods follow a multi-step approach, requiring 50 steps during inference to generate inpainting results. While BrushNet-Rand, trained with random masks, effectively handles imprecise masks and uses a full image description during inference, the other methods rely only on descriptions of the masked region." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.805, + 0.483, + 0.866 + ], + "angle": 0, + "content": "Since we adopt the architecture design of BrushNet, for fair comparison, we also train BrushNet using LocalCaptionData, referring to this configuration as BrushNet*. Notably, BrushNet, as an external module, can be directly" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.699, + 0.907, + 0.835 + ], + "angle": 0, + "content": "adapted into a 4-step diffusion model (e.g., DMD2) to form a 4-step BrushNet configuration. In contrast, methods such as HD-Painter and SDXL-Inpainting, which modify the base model structure, cannot integrate with a 4-step diffusion model. PowerPaint V2, an advanced version of BrushNet and SD1.5, incorporates learnable task prompts and specialized fine-tuning strategies across four distinct inpainting tasks. However, since PowerPaint lacks an SDXL version, it cannot be adapted to a 4-step DMD2 model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Comparative results for these methods on both Dilation-Bench and HumanBench datasets are presented in Tab. 1 and Tab. 2. Given that HumanBench, like real inpainting tasks, does not provide a global description of the en" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.466, + 0.899 + ], + "angle": 0, + "content": "https://huggingface.co/diffusers/stable-diffusion-x1-1.0-inpainting-0.1" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7618" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.053, + 0.882, + 0.21 + ], + "angle": 0, + "content": "
MetricsMask Region QualityWhole Image QualityText-Align
MethodQ-AlignCLIPIQA+TOPIQQ-AlignCLIPIQATOPIQCLIP Sim
MultiStep (50 steps)BLD4.18360.68745.17814.33200.67915.466224.884
HD-Painter3.97270.66114.93454.48050.64395.456625.585
SDXL-Inpainting4.24610.66665.15334.61720.66955.540924.848
BrushNet-Rand4.26170.67295.17384.62890.68185.577921.634
BrushNet*4.46880.71375.25044.53120.69745.558525.389
PowerPaint V24.77730.77655.54684.72270.73715.647426.256
FewStep (4 steps)BrushNet-Rand4.16020.65384.99274.55470.66085.459921.831
BrushNet*4.18360.65724.99394.44920.64275.396125.341
TurboFill4.57030.73325.27534.71880.71115.539225.352
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.213, + 0.907, + 0.243 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons among TurboFill and other diffusion based inpainting models in DilationBench. Red and blue indicates the best multi-step and the best few-step performances, respectively." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.255, + 0.882, + 0.386 + ], + "angle": 0, + "content": "
MetricsMask Region QualityWhole Image QualityText-Align
MethodQ-AlignCLIPIQA+TOPIQQ-AlignCLIPIQATOPIQCLIP Sim
MultiStep (50 steps)BLD4.15230.69085.29963.90620.65795.419924.767
HD-Painter3.98440.64935.08734.20700.61195.438125.996
SDXL-Inpainting3.95510.64055.20404.04690.62885.456524.041
BrushNet*4.25780.70635.43404.08980.66945.548025.366
PowerPaint V24.55860.75295.54754.36330.71125.633826.264
FewStep (4 steps)BrushNet*4.05080.63275.10414.14840.60035.350525.473
Ours4.47270.72575.38654.32030.68225.499225.710
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.389, + 0.907, + 0.418 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons among TurboFill and other diffusion based inpainting models in HumanBench. Red and blue indicates the best multi-step and the best few-step performances, respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.424, + 0.482, + 0.546 + ], + "angle": 0, + "content": "tire image, we exclude BrushNet-Rand from evaluation on this dataset. From Tab. 1, we find that BrushNet*, trained with LocalCaptionData, shows improvements of 3.755 and 3.52 in the text-align metric for the 50-step and 4-step models, respectively, compared to BrushNet. This suggests that training with segmentation masks and paired local descriptions, rather than global descriptions, significantly enhances BrushNet's text alignment capability." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.549, + 0.483, + 0.761 + ], + "angle": 0, + "content": "Additionally, we observe that directly inserting BrushNet-Rand and BrushNet* into DMD2 to form a 4-step model results in no significant change in text alignment performance. However, the image quality of the final inpainting results decreases noticeably. This suggests that while replacing SDXL with a few-step model (DMD2) can accelerate BrushNet, it leads to a reduction in quality. In contrast, TurboFill can effectively mitigate this issue, showing significant improvements across all three metrics—mask region quality, whole image quality, and text alignment. Even with 4 steps, our method outperforms the 50-step BrushNet-Rand and BrushNet* in many image quality metrics (e.g., Q-Align, CLIPIQA+). This phenomenon can be observed on both benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.903 + ], + "angle": 0, + "content": "PowerPaint V2 currently achieves the best results across these metrics. When comparing our method with PowerPaint, we find that the gap in whole image quality is smaller than in mask image quality. Visual comparisons (Fig. 4) reveal that although PowerPaint V2 generates rich textures, the results in the mask region are less natural, with noticeable style mismatch compared to the background region. This inconsistency leads to a lower performance in whole image quality compared to mask region quality. In contrast," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.907, + 0.471 + ], + "angle": 0, + "content": "our method produces more coherent results. Since no existing metrics effectively capture these differences, we conduct a user study, detailed in the supplementary file." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.479, + 0.737, + 0.496 + ], + "angle": 0, + "content": "5.4. Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.503, + 0.907, + 0.835 + ], + "angle": 0, + "content": "The qualitative comparison on DilationBench is shown in Fig. 4. The results from SDXL-Inpainting exhibit a yellowish color bias (e.g., tiger, man) and lack detail (e.g., boy, figurine). For PowerPaint V2, the inpainted results are very sharp in texture and rich in color, producing high scores on most IQA metrics. However, these results suffer from severe distortions (e.g., a tiger with two heads, a person with two heads, overly elongated legs) and appear less realistic and natural (e.g., third row). The results generated by BrushNet* exhibit significant artifacts, such as the tiger's fur (1st) and the man's beard (5th), as well as the appearance of strange backgrounds (5th). When BrushNet* is inserted into the 4-step DMD2 model, we observe that the results lack significant detail (e.g., 5th row) and exhibit issues with oversaturation (e.g., tiger, figurine). Compared to these methods, TurboFill generates results with higher quality in visual comparisons. While preserving detail, our results do not exhibit the oversharpness seen in PowerPaint V2 (e.g., tiger), nor the oversaturation issue observed in BrushNet* (4-steps) (e.g., rabbit). The generated content is more realistic (e.g., bird, figurine), leading to a more harmonious image overall." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.816, + 0.864 + ], + "angle": 0, + "content": "5.5. Effectiveness of LocalCaptionData" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.903 + ], + "angle": 0, + "content": "BrushNet, when trained with a full image description (i.e., global prompt) as input to the base model, exhibits rela" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7619" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.208, + 0.053, + 0.788, + 0.138 + ], + "angle": 0, + "content": "
SettingsMask Region QualityWhole Image QualityText-Align
Q-AlignCLIPQA+TOPIQQ-AlignCLIPQA+TOPIQCLIP Sim
TurboFill4.57030.73325.27534.72270.71115.539225.352
-LBG4.36720.68585.02624.54690.65705.382625.335
-LFDiff4.18750.65544.86954.49220.64265.329725.062
-RFDiff4.06640.63224.84974.44920.64125.306224.875
" + }, + { + "type": "table_caption", + "bbox": [ + 0.186, + 0.142, + 0.812, + 0.157 + ], + "angle": 0, + "content": "Table 3. The effectiveness of different losses. From top to down, we progressively remove specific losses." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.17, + 0.475, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.519, + 0.483, + 0.547 + ], + "angle": 0, + "content": "Figure 5. The effectiveness of LocalCaptionData. All results are obtained based on 4-step DMD2. (Zoom in for best view)" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.564, + 0.483, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.657, + 0.483, + 0.686 + ], + "angle": 0, + "content": "Figure 6. The effectiveness of different losses. From left to right, we progressively remove specific losses. (Zoom in for best view)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "tively weak text alignment capabilities, as shown in Fig. 5. In cases where BrushNet attempts to generate content similar or semantically related to existing objects in the image, it often either defaults to complete the background or shifts focus to other categories in the prompt (e.g., a cat). Furthermore, relying on a global prompt limits BrushNet's effectiveness in handling more complex prompts. By using LocalCaptionData, BrushNet* effectively overcomes these limitations without any changes to the training approach, resulting in a significant improvement in text alignment performance. Additionally, TurboFill demonstrates superior performance in generating fine details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.17, + 0.682, + 0.185 + ], + "angle": 0, + "content": "5.6. Ablation Studies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.191, + 0.907, + 0.449 + ], + "angle": 0, + "content": "Starting from TurboFill, we remove \\(\\mathcal{L}_{\\mathrm{BG}}\\), \\(\\mathcal{L}_{\\mathrm{Diff}}^F\\), and \\(\\mathcal{L}_{\\mathrm{Diff}}^R\\) in sequence, with quantitative and qualitative results shown in Fig. 6 and Tab. 3. When \\(\\mathcal{L}_{\\mathrm{BG}}\\) is removed, there is a clear performance drop in all metrics. In the visualizations, we see that without \\(\\mathcal{L}_{\\mathrm{BG}}\\), the color of background region changes noticeably, creating a sharp boundary between the fill-in and background regions. Further removing \\(\\mathcal{L}_{\\mathrm{Diff}}^F\\) introduces conflicting elements (i.e., house) in the fill-in region, suggesting the discriminator fails to fully capture the holistic scene. Finally, without \\(\\mathcal{L}_{\\mathrm{Diff}}^R\\), relying only on GAN loss, the inpainted images exhibit not only inconsistencies with the background but also poor texture and detail. This highlights that GAN loss alone struggles to close the gap between fake and real latents. Only when combining \\(\\mathcal{L}_{\\mathrm{Diff}}^F\\), \\(\\mathcal{L}_{\\mathrm{Diff}}^R\\), and \\(\\mathcal{L}_{\\mathrm{BG}}\\) in GAN training does the model achieve enhanced texture, detail, and effective scene harmonization between fill-in and background regions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.454, + 0.634, + 0.469 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.479, + 0.907, + 0.72 + ], + "angle": 0, + "content": "In this work, we introduce TurboFill, a fast inpainting model designed to address the computational and quality challenges in diffusion-based inpainting tasks. By directly training an inpainting adapter on a few-step distilled text-to-image diffusion model, TurboFill achieves state-of-the-art inpainting results with significantly smaller inference costs. The proposed three-step adversarial training scheme, incorporating both GAN and diffusion losses, ensures the generation of realistic and harmonious content with fine textures and structural coherence. To evaluate TurboFill's performance, we introduce two new benchmarks, DilationBench and HumanBench, which provide robust evaluations across varying mask complexities and user-centric scenarios. Experimental results demonstrate TurboFill's ability to outperform existing few-step and multi-step BrushNet, setting a new benchmark for practical and efficient inpainting." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.725, + 0.694, + 0.742 + ], + "angle": 0, + "content": "7. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "This work was supported in part by Macau Science and Technology Development Fund under SKLIOTSC-2021-2023, 0022/2022/A1, and 0014/2022/AFJ; in part by Research Committee at University of Macau under MYRG-GRG2023-00058-FST-UMDF; in part by the Guangdong Basic and Applied Basic Research Foundation under Grant 2024A1515012536; in part by National Natural Science Foundation of China (Grant No. 62276251), and the Joint Lab of Shenzhen Key Laboratory of Computer Vision and Pattern Recognition." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7620" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.156 + ], + "angle": 0, + "content": "[1] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM transactions on graphics (TOG), 42 (4):1-11, 2023. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.158, + 0.484, + 0.212 + ], + "angle": 0, + "content": "[2] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Trans. Graph., 28(3):24, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.215, + 0.483, + 0.284 + ], + "angle": 0, + "content": "[3] Chenjie Cao, Yunuo Cai, Qiaole Dong, Yikai Wang, and Yanwei Fu. Leftrefill: Filling right canvas based on left reference through generalized text-to-image diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7705-7715, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.285, + 0.483, + 0.352 + ], + "angle": 0, + "content": "[4] Chaofeng Chen, Jiadi Mo, Jingwen Hou, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Topiq: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.483, + 0.396 + ], + "angle": 0, + "content": "[5] Li Chen, Yu Wang, Jiali Zhang, and Xingxing Xu. Powerpaint: Adaptive task prompts for generalizable image inpainting. arXiv preprint arXiv:2401.08965, 2024. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.398, + 0.483, + 0.464 + ], + "angle": 0, + "content": "[6] Yifu Chen, Jingwen Chen, Yingwei Pan, Yehao Li, Ting Yao, Zhineng Chen, and Tao Mei. Improving text-guided object inpainting with semantic pre-inpainting. In European Conference on Computer Vision, pages 110-126. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.483, + 0.522 + ], + "angle": 0, + "content": "[7] Huggingface Diffusers. Sdxl-inpaint. https : / / huggingface . co / diffusers / stable - diffusion-x1- 1 . 0 - inpainting- 0 . 1, Year. Accessed: YYYYY-MM-DD. 1,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.524, + 0.483, + 0.577 + ], + "angle": 0, + "content": "[8] Stefan Elfwing, Eiji Uchibe, and Kenji Doya. Sigmoid-weighted linear units for neural network function approximation in reinforcement learning. Neural networks, 107:3-11, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.483, + 0.62 + ], + "angle": 0, + "content": "[9] Omar Elharrouss, Noor Almaadeed, Somaya Al-Maadeed, and Younes Akbari. Image inpainting: A review. Neural Processing Letters, 51:2007-2028, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.622, + 0.483, + 0.675 + ], + "angle": 0, + "content": "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems (NeurIPS), pages 6840-6851, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.678, + 0.483, + 0.734 + ], + "angle": 0, + "content": "[11] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. arXiv preprint arXiv:2403.06976, 2024. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.735, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[12] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision, 128(7):1956-1981, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.483, + 0.873 + ], + "angle": 0, + "content": "[13] Shanchuan Lin, Anran Wang, and Xiao Yang. Sdxlightning: Progressive adversarial diffusion distillation. arXiv preprint arXiv:2402.13929, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.149 + ], + "angle": 0, + "content": "Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[15] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[16] Hayk Manukyan, Andranik Sargsyan, Barsegh Atanyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Hd-painter: high-resolution and prompt-faithful text-guided image inpainting with diffusion models. arXiv preprint arXiv:2312.14091, 2023. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.278, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[17] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[18] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.433, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[19] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.504, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[20] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.547, + 0.905, + 0.589 + ], + "angle": 0, + "content": "[21] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.905, + 0.646 + ], + "angle": 0, + "content": "[22] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.648, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[23] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching Imms for visual scoring via discrete text-defined levels. arXiv preprint arXiv:2312.17090, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[24] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.761, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[25] Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, and Lu Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4818–4829, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. arXiv preprint arXiv:2311.18828, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7621" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[27] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. arXiv preprint arXiv:2405.14867, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.217 + ], + "angle": 0, + "content": "[28] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contextual attention. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5505-5514, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[29] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4471-4480, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[30] Yanhong Zeng, Jianlong Fu, Hongyang Chao, and Baining Guo. Learning pyramid-context encoder network for high-quality image inpainting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1486-1494, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.346, + 0.482, + 0.402 + ], + "angle": 0, + "content": "[31] Yanhong Zeng, Jianlong Fu, Hongyang Chao, and Baining Guo. Aggregated contextual transformations for high-resolution image inpainting. IEEE Transactions on Visualization and Computer Graphics, 29(7):3266-3280, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.404, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[32] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.46, + 0.482, + 0.515 + ], + "angle": 0, + "content": "[33] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7622" + } + ] +] \ No newline at end of file diff --git a/2025/TurboFill_ Adapting Few-step Text-to-image Model for Fast Image Inpainting/6302d669-b0e3-4a35-adf2-a42f09f692e9_origin.pdf b/2025/TurboFill_ Adapting Few-step Text-to-image Model for Fast Image Inpainting/6302d669-b0e3-4a35-adf2-a42f09f692e9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1d3eedc1813e6c58a331b28b01b0c927d75b1f05 --- /dev/null +++ b/2025/TurboFill_ Adapting Few-step Text-to-image Model for Fast Image Inpainting/6302d669-b0e3-4a35-adf2-a42f09f692e9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6b09b7d488c3d5a7c1bf808961282e56ba8b29cf0c0b70a45aff90cd9aff3030 +size 9367756 diff --git a/2025/TurboFill_ Adapting Few-step Text-to-image Model for Fast Image Inpainting/full.md b/2025/TurboFill_ Adapting Few-step Text-to-image Model for Fast Image Inpainting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..37efa7a1dd05232a90de673627e6b8bae540e5f7 --- /dev/null +++ b/2025/TurboFill_ Adapting Few-step Text-to-image Model for Fast Image Inpainting/full.md @@ -0,0 +1,316 @@ +# TurboFill: Adapting Few-step Text-to-image Model for Fast Image Inpainting + +Liangbin Xie $^{1,4}$ Daniil Pakhomov $^{3}$ Zhonghao Wang $^{3}$ Zongze Wu $^{3}$ Ziyan Chen $^{4}$ Yuqian Zhou $^{3}$ Haitian Zheng $^{3}$ Zhifei Zhang $^{3}$ Zhe Lin $^{3}$ Jiantao Zhou $^{1\dagger}$ Chao Dong $^{2,4\dagger}$ + +$^{1}$ State Key Laboratory of Internet of Things for Smart City, University of Macau $^{2}$ Shenzhen University of Advanced Technology $^{3}$ Adobe $^{4}$ Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences + +Yellow baby chick on wooden surface. + +A miniature blue convertible with its top down + +![](images/5e6687e177172a2e3492764e5438725748c2ab485972dd07f6ca401470d5dd6f.jpg) +Input Image + +![](images/cf85798f3423a8c63336ef0e7fb3ca0143b92d97ff145226d2dc3ee915783591.jpg) +BrushNet* (50 steps) + +![](images/55f52f85a82c858ccd571c89d0524f9c9ba5a270297927b9a3e1758ff8d6da74.jpg) +TurboFill (4 steps) + +![](images/05f8888f7625f773512e8f30dec75077631e28de82b334399f7d1de6d10a454a.jpg) +Input Image + +![](images/25027e4998702b86029330b9b3fc906a857b1382f886d0e13e500426466ac479.jpg) +BrushNet* (50 steps) + +![](images/95a4da8cca6f387d867ac3ef78f1353c54b5c7eba1b51a9d7d26ee7bdcf40df4.jpg) +TurboFill (4 steps) + +A fluffy brown teddy bear wearing a tiny hat. A tinted sunglasses lens reflecting tall palm trees. Man with cowboy hat and bandana in front of wooden wall. + +![](images/d6047e0f407158605f419a54b2e168329792a9b2bc5125b86627f589d69ce254.jpg) +Input Image + +![](images/bc05916997be740db3f6a871795c007341f28352a8298aee31e15d90a38fb309.jpg) +TurboFill (4 steps) + +![](images/aca537c86d5eab556bee1c64a0e847af6fd4a5fe53f834e119a82344c07aa6c5.jpg) +Input Image + +![](images/01ea8822a4fc25883e047968a37c3f0fb692f7505fe0c5709bc974fc3ce99899.jpg) +TurboFill (4 steps) +Figure 1. We propose TurboFill, a fast image inpainting method that leverages a 3-step adversarial training scheme. With only four diffusion steps, TurboFill outperforms the multi-step BrushNet* [11], delivering realistic details and textures with remarkable efficiency. + +![](images/9d5db6efbb2ed4a309e250980081c7271eb8830fafc8b7802c2c831d85d88a38.jpg) +Input Image + +![](images/477974fc0ee3751fa55b9273381bae05337c978633db4ef010f04012f61362e2.jpg) +TurboFill (4 steps) + +# Abstract + +This paper introduces TurboFill, a fast image inpainting model that enhances a few-step text-to-image diffusion model with an inpainting adapter for high-quality and efficient inpainting. While standard diffusion models generate high-quality results, they incur high computational costs. We overcome this by training an inpainting adapter on a few-step distilled text-to-image model, DMD2, using a novel 3-step adversarial training scheme to ensure realistic, structurally consistent, and visually harmonious inpainted regions. To evaluate TurboFill, we propose two benchmarks: DilationBench, which tests performance across mask sizes, and HumanBench, based on human feedback for complex prompts. Experiments show that TurboFill outperforms both multi-step BrushNet and few-step inpainting methods, setting a new benchmark for high-performance inpainting tasks. The project page is available here. + +# 1. Introduction + +Image inpainting, the task of filling missing regions in an image, has seen significant advancement with the rise of deep generative models [1, 5, 11, 16]. Existing diffusion-based inpainting methods have shown promising results for various applications such as object removal, regeneration, and text-guided object insertion. + +The most straightforward way of training diffusion-based inpainting models is to add the masked image (background conditioning) to the input channels of a pretrained diffusion model and finetune it with inpainting data, as shown by SDXL-inpaint [7]. However, such methods typically require fine-tuning the entire network, entailing large training datasets and significant computational resources. In contrast, adapter-based approaches, like BrushNet [11], utilize a ControlNet [33] to capture conditioning information and inject features into a frozen diffusion U-Net. BrushNet achieves state-of-the-art inpainting results while incurring relatively low training costs in terms of data and computational resources, owing to the adapter's efficiency and the + +![](images/078363a1a7eb3661db3f22dc4e2e4436d37c76c5e995a6aa09d7814df2853e0e.jpg) +Figure 2. (Zoom in for best view) 1. The multi-step adapter achieves high-quality inpainting results but incurs significant inference costs, requiring over 50 diffusion steps. 2. Applying the pre-trained multi-step BrushNet adapter directly to the few-step U-Net (DMD2) results in artifacts, including oversaturated colors and semantic inconsistencies (e.g., generating a dog with two tails). 3. Training the adapter-DMD2 solely with diffusion loss produces blurred outputs with low-quality inpainting results. 4. In contrast, training the adapter-DMD2 using the proposed 3-step adversarial training scheme yields high-quality inpainting results, requiring only four diffusion steps. + +preservation of the U-Net's original generation capacity. + +Despite the advantage of BrushNet-like models, they still depend on many iterative sampling steps, leading to substantial inference cost in practice. Interestingly, we found that an inpainting adapter trained on a pretrained text-to-image diffusion model can be seamlessly integrated into its corresponding few-step student model without further training, while yielding reasonable quality results (Fig. 2 ②). This can be attributed to the semantic alignment between the teacher and student models. However, this naive migration approach exhibits quality issues such as color shifts, over-saturation, and loss of details. + +Alternatively, we can use a two-stage approach by first training a teacher BrushNet and then applying advanced diffusion distillation techniques [13, 15, 20, 26, 27] to the teacher BrushNet to reduce steps while maintaining quality. Although recent distillation methods for diffusion models have demonstrated remarkable performance, we have found that due to BrushNet's large model size, it demands excessive memory and incurs high computational costs, making the training process difficult to manage. + +In this paper, we propose a simpler single-stage training strategy by training an inpainting adapter directly on top of a pretrained few-step distilled text-to-image model (e.g., DMD2). We observe that naively training an inpainting adapter and few-step model solely with diffusion loss results in blurry, low-quality outputs (Fig. 2③). To address these shortcomings and enhance inpainting performance, we introduce a diffusion discriminator to guide optimization of the adapter-DMD2 generator. The diffusion discriminator is trained to differentiate between noisy real images and fake latents, while a combined diffusion loss ensures the model learns meaningful semantic representations. + +In summary, TurboFill is trained with a novel 3-step adversarial training scheme with two generator update steps and one discriminator update step. Our model enables high- + +quality image inpainting with only a few diffusion steps. Compared to the naive approach of training a multi-step adapter (requiring 8 V100 GPUs for 72 hours) followed by distillation (requiring 64 A100 GPUs for 50 hours), TurboFill attains comparable outcomes using merely 8 A100 GPUs within 50 hours. This accomplishment marks a reduction in training time by over a factor of 10 and a substantial enhancement in training efficiency. + +Existing benchmarks often fall short in simulating real-world user operations. To address this gap, we introduce two new benchmarks: DilationBench and HumanBench. DilationBench provides an objective measure by testing the model on the masks with varying dilation sizes, while HumanBench incorporates human user feedback and complex prompts, offering insights into real-world performance. These benchmarks contribute a robust evaluation framework that can facilitate advancements in inpainting models and more accurately reflect user-oriented performance. + +Extensive quantitative and qualitative evaluations demonstrate that TurboFill not only supports complex prompts and user-driven inpainting tasks but also outperforms existing few-step inpainting models in terms of speed and quality. In summary, our contributions are as follows: + +- A novel 3-step adversarial training scheme is proposed to train an inpainting adapter directly on top of few-step text-to-image diffusion model, which helps the model to focus on image quality and improves textures, details, and scene harmonization. +- New benchmarks, DilationBench and HumanBench, are designed to better evaluate inpainting methods in user-centered scenarios. +- We demonstrate that our few-step TurboFill outperforms multi-step BrushNet for the inpainting task and achieves state-of-the-art performance in a few-step setting through extensive experiments. + +# 2. Related Work + +Image inpainting, a classic problem in computer vision, focuses on reconstructing missing regions in an image with realistic and natural content. Early methods, including traditional techniques, VAEs, and GANs [2, 3, 9, 28-31], often struggled to produce high-quality results. Recently, diffusion models have gained popularity for their ability to generate high-quality, diverse outputs with fine-grained control, demonstrating significant advancements over earlier approaches [1, 5, 11, 16]. Diffusion models, particularly the Stable Diffusion family, have introduced a robust framework for image generation and inpainting by iteratively denoising latent representations. Stable Diffusion-based methods have achieved impressive results in high-resolution image inpainting, yet are computationally intensive due to multi-step sampling. + +Inpainting tasks based on diffusion models generally follow two approaches: training an entire diffusion model or incorporating external modules. The first approach, like PowerPaint [5], CATDiffusion [6], and SDXL-Inpainting, focuses on adapting the model itself. PowerPaint introduces learnable task prompts along with tailored fine-tuning strategies to guide the model's focus on specific inpainting targets, while SDXL-Inpainting [7] directly utilizes both the background region and the mask as inputs to fine-tune a new model. In contrast, the second approach, exemplified by methods such as ControlNet [33] or T2I-Adapter [17], leverages external modules to enhance the diffusion model's capabilities without retraining it entirely. Recent methods, such as BrushNet [11], employ mask-specific prompts, refining the generated content within the masked area for improved coherence with the background. + +Our method builds on these developments by leveraging a ControlNet-like adapter for DMD2, integrating both GAN and diffusion losses to enhance image realism and style coherence. By focusing on masked-region-specific prompts and incorporating fine-grained diffusion control, our approach achieves fast and high-quality inpainting results that align closely with user inputs and surrounding context, advancing upon prior work in speed, quality, and flexibility in real-world applications. + +# 3. Preliminary + +The forward diffusion process gradually adds Gaussian noise $\epsilon_t$ to a clean image $x_0$ by + +$$ +x _ {t} = \sqrt {\bar {\alpha} _ {t}} x _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {t}, \tag {1} +$$ + +where $\bar{\alpha}_t$ controls noise schedule. In the training phase, a denoising network $\epsilon_{\theta}$ is trained to predict $\epsilon_t$ given $x_t$ , timestep $t$ and text prompt $c$ , with the diffusion loss: + +$$ +L \left(\hat {\epsilon} _ {\theta}\right) = \mathbb {E} _ {x _ {0} \sim q; \epsilon_ {t} \sim \mathcal {N} (0, 1)} [ \| \epsilon_ {\theta} (x _ {t}, c, t) - \epsilon_ {t} \| ^ {2} ]. \tag {2} +$$ + +To enable condition generation, adapter-based methods [11, 17, 32] introduce additional conditioning information via a lightweight adapter, and achieve efficient training (since the main diffusion network remains frozen), with a slight inference overhead. + +Traditionally, reverse diffusion processes necessitate a computationally intensive multi-step iterative sampling procedure (typically 20-50 steps). Distillation techniques [13, 15, 20, 26] have been developed to reduce this requirement, achieving accelerated sampling with as few as 1-4 steps. However, these distillation processes are resource-intensive in both computational power and time. In contrast, our approach directly trains the adapter with both the multi-step and few-step text-to-image diffusion model, eliminating the need for a separate distillation process, thereby enabling more efficient training and inference. + +# 4. Our Approach + +In this work, we propose a fast inpainting model, TurboFill, which directly trains an inpainting adapter on top of a few-step text-to-image diffusion model. As shown in Fig. 3, there are three components: the fast generator, the slow generator and the diffusion discriminator. All these three components aim to enhance the adapter's capability of inpainting masked images. Specifically, the slow generator helps the adapter to denoise the noisy latents towards the most probable direction of realistic image generation. The fast generator enables the adapter to generate a clean image for critique during training. The diffusion discriminator guides the adapter to generate images with better textures, details and content harmonization. Based on these three components, TurboFill follows a 3-step adversarial training scheme. The training alternates between 3 steps throughout the training process. In the following sections, we provide a detailed description of the training procedure for each step and the components. + +# 4.1. Training the Adapter in Slow Generator + +To capture conditioning information and introduce additional features to the diffusion model, we utilize an inpainting adapter with the same ControlNet architecture as BrushNet [11]. Given that SDXL is a multi-step text-to-image generator, we designate the integration of the inpainting adapter and SDXL as the 'slow generator'. In contrast, DMD2 is a few-step text-to-image generator, so the combination of the inpainting adapter and DMD2 is named the 'fast generator'. In both generators, conditioning information is introduced by concatenating the noisy latent $x_{t}$ , the background image latent $x_{bg}$ , and a downsampled binary mask $m$ . This concatenated input is then processed by the inpainting adapter, while $x_{t}$ is simultaneously fed into the UNet. Then, the features produced by the inpainting adapter are fused with the UNet's features via residual connection. + +![](images/014f01b28fe85da2eb5385e21f38092ea752b505a51043d421c657ee3f49ed6f.jpg) +Figure 3. The training of TurboFill alternates between 3 steps: (1) optimizing the adapter using the gradient of $\mathcal{L}_{\mathrm{Diff}}^R$ , and (2) $\mathcal{L}_{\mathcal{G}}$ and $\mathcal{L}_{\mathrm{BG}}$ are employed to update the adapter in fast generator, and (3) $\mathcal{L}_{\mathrm{Diff}}^F$ and $\mathcal{L}_{\mathcal{D}}$ are jointly applied to update the parameters of the diffusion discriminator module. Note that the adapters in slow generator and fast generator share same weights during training. + +As shown in Fig. 3 (1), the parameters of the inpainting adapter in slow generator are updated based on the real diffusion loss: + +$$ +\mathcal {L} _ {\text {D i f f}} ^ {R} = \min _ {\epsilon_ {\theta}} \mathbb {E} _ {x _ {t}, t \sim [ 0, T ], \epsilon \sim \mathcal {N} (0, 1)} \tag {3} +$$ + +$$ +\left\| \epsilon - \epsilon_ {\theta} ^ {\mathrm {S D X L}} (x _ {t}, [ x _ {b g}, x _ {t}, m ], t) \right\| _ {2} ^ {2}, +$$ + +where $T = 1,000$ and DDPM scheduler [10] is employed to generate $x_{t}$ . + +# 4.2. Training the Adapter in Fast Generator + +Given that adversarial training is highly effective in minimizing the divergence between different distributions, we employ adversarial training to minimize the discrepancy between real latent distribution and the fake latent distribution $\{\hat{x}_0\}$ . Specifically, we first generate the clean fake latent $\hat{x}_0$ from the noisy latent $x_{t}$ with the fast generator in a single step, as described by + +$$ +\hat {x} _ {0} = \frac {x _ {t} - \sqrt {1 - \bar {\alpha} _ {t}} \times \epsilon_ {\theta} ^ {\mathrm {D M D 2}} \left(x _ {t} , \left[ x _ {b g} , x _ {t} , m \right] , t\right)}{\sqrt {\bar {\alpha} _ {t}}}, \tag {4} +$$ + +where $\bar{\alpha}_t$ is determined by the LCM scheduler [15] and $t$ is the sampled timestep. Then, the noisy version $\hat{x}_t$ originated from $\hat{x_0}$ is fed into the diffusion discriminator. Meanwhile, to preserve the background region, $\hat{x_0}$ is used to compute the background loss. + +Diffusion discriminator. As shown in Fig. 3 (2), the diffusion discriminator $\mathcal{D}_{\phi}$ consists of the SDXL encoder, the assistant encoder, and the convolution-based classifier. For assistant encoder, we use the same encoder structure of inpainting adapter. It takes concatenated noisy latent, downsampled binary mask, and masked image latent $x_{bg}$ as the input. The features extracted by the assistant encoder are progressively integrated into the SDXL encoder. The final feature maps are then passed through the classifier and + +mapped to a one-dimensional vector. Based on the vector, we calculate $\mathcal{L}_{\mathcal{G}}$ : + +$$ +\mathcal {L} _ {\mathcal {G}} = \min _ {\epsilon_ {\theta}} - \mathbb {E} _ {y _ {t}, t \sim [ 0, T ]} \left[ \log \mathcal {D} _ {\phi} \left(\hat {x} _ {0}, \left[ x _ {b g}, \hat {x} _ {0}, m \right], t\right)\right), \tag {5} +$$ + +where $T = 1,000$ + +Background preservation. For the image inpainting task, it is essential to preserve the background region. Therefore, for the clean fake latent $\hat{x_0}$ , we apply a reconstruction loss to the background region to minimize the distance between $\hat{x_0}$ and the clean latent $x_0$ : + +$$ +\mathcal {L} _ {\mathrm {B G}} = \left\| x _ {0} \odot m - \hat {x} _ {0} \odot m \right\| _ {2} ^ {2}, \tag {6} +$$ + +where $\odot$ represents point-wise multiplication. Based on $\lambda_1\mathcal{L}_{\mathcal{G}} + \lambda_2\mathcal{L}_{\mathrm{BG}}$ , in this step, only the parameters of the inpainting adapter are updated. + +# 4.3. Training the Diffusion Discriminator + +When updating the discriminator, we fix the parameters of the inpainting adapter and update the parameters of the assistant encoder and classifier. Specifically, both the real noisy latent $x_{t}$ and the fake noisy latent $\hat{x}_{t}$ are input into the diffusion discriminator to compute the loss $\mathcal{L}_{\mathcal{D}}$ : + +$$ +\begin{array}{l} \mathcal {L} _ {\mathcal {D}} = - \mathbb {E} _ {\hat {x} _ {0, t}, t \sim [ 0, T ]} \left[ \log \left(1 - \mathcal {D} _ {\phi} (\hat {x} _ {0}, [ x _ {b g}, \hat {x} _ {0}, m ], t)\right)\right) ] \\ - \mathbb {E} _ {x _ {t}, t \sim [ 0, T ]} \left[ \right. \log \mathcal {D} _ {\phi} \left(x _ {t}, \left[ x _ {\mathrm {b g}}, x _ {t}, m \right], t\right)\left. \right)\left. \right]. \tag {7} \\ \end{array} +$$ + +Although the inpainting adapter optimized only with GAN loss can generate images, the generated images are of inferior quality and commonly incorporate unwanted elements. We attribute this issue to the limitations of the GAN loss, which rarely assists the discriminator in comprehending the image structure. In contrast, diffusion loss excels at scene understanding. Therefore, we introduce the fake diffusion loss to train the discriminator while allowing the + +GAN loss to focus more on enhancing image textures and object details. To compute the diffusion loss for the fake noisy latent $\hat{x}_t$ , we introduce an assistant decoder, following the assistant encoder. The fake diffusion loss is: + +$$ +\mathcal {L} _ {\text {D i f f}} ^ {F} = \min _ {\mathcal {A} _ {\psi}} \mathbb {E} _ {y _ {t}, t \sim [ 0, T ], \epsilon \sim \mathcal {N} (0, 1)} \tag {8} +$$ + +$$ +\left\| \epsilon - \mathcal {A} _ {\psi} ^ {\mathrm {S D X L}} (x _ {t}, [ x _ {b g}, \hat {x} _ {t}, m ], t) \right\| _ {2} ^ {2}, +$$ + +where $\mathcal{A}_{\psi}$ represents the parameters of assistant encoder and decoder. The combination of $\mathcal{L}_{\mathrm{Diff}}^{F} + \lambda_{3}\mathcal{L}_{\mathcal{D}}$ is used as the loss to optimize the assistant encoder, assistant decoder modules, and classifier. + +# 5. Experiments + +# 5.1. Datasets and Metrics + +# 5.1.1. Training Dataset + +To enable editing through the input of a local mask and the corresponding prompt for the masked region, we construct a dataset that includes local masks and their associated textual descriptions. Specifically, we crawl approximately 1,200,000 images from Internet and use Florence-2 [25] to perform the dense region caption task with the prompt $<$ DENSE_REGION_CAPTION>. This task identifies the main objects in the images and generates their corresponding short descriptions. These descriptions are then input into SAM2 [19] to obtain the segmentation masks for the objects. We name this dataset as LocalCaptionData. + +# 5.1.2. Evaluation Datasets + +The commonly used public datasets in the field of image inpainting are primarily OpenImage V6 [12], MSCOCO [14], and BrushBench [11]. Among these, OpenImage V6 and MSCOCO provide precise segmentation masks and bounding box masks, which are used to perform inpainting on the masked regions based on class names. BrushBench only offers accurate segmentation masks, where inpainting is performed on the masked regions using a global image description. However, considering that users typically perform inpainting with rough masks and corresponding descriptions (which may sometimes be complex and contain multiple attributes), these three datasets significantly diverge from real-world scenarios. To better simulate user operations, we construct two evaluation datasets: + +- DilationBench: The masks are created by randomly dilating the segmentation masks, and the prompts describe the content within the masked area of the original image. DilationBench contains 300 pairs of masks and prompts. +- HumanBench: The masks are manually labeled, and the prompts are manually written. HumanBench contains 150 pairs of masks and prompts. + +The descriptions and visualizations of the two benchmark datasets are included in the supplementary material. + +# 5.1.3. Evaluation Metrics + +To evaluate the quality of inpainted images, we use three metrics: Q-Align [23], CLIPIQA+ [22], TOPIQ [4]. They assess visual quality from different perspectives. Q-Align, which leverages a large language model (LLM) to generate visual scores, is state-of-the-art on various Image Quality Assessment (IQA) leaderboards1. Excluding Q-Align, CLIPIQA+ and TOPIQ rank first in non-reference and image aesthetic benchmarks, respectively, making them reliable indicators of visual quality and providing complementary insights. For a more comprehensive evaluation of inpainting results, we consider not only the masked region but also the entire image, ensuring that both local quality in the inpainted area and overall visual coherence across the entire image are assessed. To measure the consistency between the text and the content generated in the mask region, we use the CLIP Similarity metric [18]. Higher values for all four metrics indicate better performance. + +# 5.2. Experimental Settings + +# 5.2.1. Network Configurations + +Inpainting adapter adopts the same architecture as Brush-Net, with modifications to the SDXL model. Specifically, the self-attention and cross-attention modules are removed. Note that our approach does not impose any architectural constraints on inpainting adapter. Assistant also adopts U-Net like structure, for simplicity, we directly adopt the same architecture as inpainting adapter. The classifier is composed of convolutional layers, GroupNorm [24], and SiLU [8] activation, which together map the feature map to a 1-dimensional vector. The details of classifier are described in the supplementary material. + +When using SDXL as the base model for noise prediction, we employ a DDIM scheduler [21] and sample randomly from 1,000 timesteps. In contrast, when using DMD2 as the base model, we use an LCM scheduler [15] and randomly sample from four specific timesteps: \{999, 749, 499, 249\}. + +# 5.2.2. Implementation Details + +All experiments are performed on 8 A100 GPUs with 80GB memory. During training, the learning rate for parameter updates is set to 1e-5, with a batch size of 2 and gradient accumulation steps configured to 4. We adopt mixed precision to optimize computational efficiency. $\lambda_{1},\lambda_{2},\lambda_{3}$ are set to $1e^{-3}$ , $1e^{-1}$ , and $1e^{-2}$ , respectively. All models are trained for 40K iterations using the AdamW optimizer. + +# 5.3. Quantitative Comparisons + +We compare TurboFill with existing state-of-the-art image inpainting methods, namely BLD [1], HD-Painter [16], + +![](images/96743521a278e94527a3f6d876b6c9df93ccc04878204fc3e48c35c7829b3dc4.jpg) +' tiger' +Figure 4. Comparison of previous inpainting methods and BrushNet on DilationBench. Compared to other methods, TurboFill generates more realistic details and textures in just 4 steps, while achieving good scene harmonization. (Zoom in for best view) + +SDXL-Inpainting $^1$ , BrushNet-Rand [11], and PowerPaint [5]. All of these methods follow a multi-step approach, requiring 50 steps during inference to generate inpainting results. While BrushNet-Rand, trained with random masks, effectively handles imprecise masks and uses a full image description during inference, the other methods rely only on descriptions of the masked region. + +Since we adopt the architecture design of BrushNet, for fair comparison, we also train BrushNet using LocalCaptionData, referring to this configuration as BrushNet*. Notably, BrushNet, as an external module, can be directly + +adapted into a 4-step diffusion model (e.g., DMD2) to form a 4-step BrushNet configuration. In contrast, methods such as HD-Painter and SDXL-Inpainting, which modify the base model structure, cannot integrate with a 4-step diffusion model. PowerPaint V2, an advanced version of BrushNet and SD1.5, incorporates learnable task prompts and specialized fine-tuning strategies across four distinct inpainting tasks. However, since PowerPaint lacks an SDXL version, it cannot be adapted to a 4-step DMD2 model. + +Comparative results for these methods on both Dilation-Bench and HumanBench datasets are presented in Tab. 1 and Tab. 2. Given that HumanBench, like real inpainting tasks, does not provide a global description of the en + +
MetricsMask Region QualityWhole Image QualityText-Align
MethodQ-AlignCLIPIQA+TOPIQQ-AlignCLIPIQATOPIQCLIP Sim
MultiStep (50 steps)BLD4.18360.68745.17814.33200.67915.466224.884
HD-Painter3.97270.66114.93454.48050.64395.456625.585
SDXL-Inpainting4.24610.66665.15334.61720.66955.540924.848
BrushNet-Rand4.26170.67295.17384.62890.68185.577921.634
BrushNet*4.46880.71375.25044.53120.69745.558525.389
PowerPaint V24.77730.77655.54684.72270.73715.647426.256
FewStep (4 steps)BrushNet-Rand4.16020.65384.99274.55470.66085.459921.831
BrushNet*4.18360.65724.99394.44920.64275.396125.341
TurboFill4.57030.73325.27534.71880.71115.539225.352
+ +Table 1. Quantitative comparisons among TurboFill and other diffusion based inpainting models in DilationBench. Red and blue indicates the best multi-step and the best few-step performances, respectively. + +
MetricsMask Region QualityWhole Image QualityText-Align
MethodQ-AlignCLIPIQA+TOPIQQ-AlignCLIPIQATOPIQCLIP Sim
MultiStep (50 steps)BLD4.15230.69085.29963.90620.65795.419924.767
HD-Painter3.98440.64935.08734.20700.61195.438125.996
SDXL-Inpainting3.95510.64055.20404.04690.62885.456524.041
BrushNet*4.25780.70635.43404.08980.66945.548025.366
PowerPaint V24.55860.75295.54754.36330.71125.633826.264
FewStep (4 steps)BrushNet*4.05080.63275.10414.14840.60035.350525.473
Ours4.47270.72575.38654.32030.68225.499225.710
+ +Table 2. Quantitative comparisons among TurboFill and other diffusion based inpainting models in HumanBench. Red and blue indicates the best multi-step and the best few-step performances, respectively. + +tire image, we exclude BrushNet-Rand from evaluation on this dataset. From Tab. 1, we find that BrushNet*, trained with LocalCaptionData, shows improvements of 3.755 and 3.52 in the text-align metric for the 50-step and 4-step models, respectively, compared to BrushNet. This suggests that training with segmentation masks and paired local descriptions, rather than global descriptions, significantly enhances BrushNet's text alignment capability. + +Additionally, we observe that directly inserting BrushNet-Rand and BrushNet* into DMD2 to form a 4-step model results in no significant change in text alignment performance. However, the image quality of the final inpainting results decreases noticeably. This suggests that while replacing SDXL with a few-step model (DMD2) can accelerate BrushNet, it leads to a reduction in quality. In contrast, TurboFill can effectively mitigate this issue, showing significant improvements across all three metrics—mask region quality, whole image quality, and text alignment. Even with 4 steps, our method outperforms the 50-step BrushNet-Rand and BrushNet* in many image quality metrics (e.g., Q-Align, CLIPIQA+). This phenomenon can be observed on both benchmarks. + +PowerPaint V2 currently achieves the best results across these metrics. When comparing our method with PowerPaint, we find that the gap in whole image quality is smaller than in mask image quality. Visual comparisons (Fig. 4) reveal that although PowerPaint V2 generates rich textures, the results in the mask region are less natural, with noticeable style mismatch compared to the background region. This inconsistency leads to a lower performance in whole image quality compared to mask region quality. In contrast, + +our method produces more coherent results. Since no existing metrics effectively capture these differences, we conduct a user study, detailed in the supplementary file. + +# 5.4. Qualitative Comparison + +The qualitative comparison on DilationBench is shown in Fig. 4. The results from SDXL-Inpainting exhibit a yellowish color bias (e.g., tiger, man) and lack detail (e.g., boy, figurine). For PowerPaint V2, the inpainted results are very sharp in texture and rich in color, producing high scores on most IQA metrics. However, these results suffer from severe distortions (e.g., a tiger with two heads, a person with two heads, overly elongated legs) and appear less realistic and natural (e.g., third row). The results generated by BrushNet* exhibit significant artifacts, such as the tiger's fur (1st) and the man's beard (5th), as well as the appearance of strange backgrounds (5th). When BrushNet* is inserted into the 4-step DMD2 model, we observe that the results lack significant detail (e.g., 5th row) and exhibit issues with oversaturation (e.g., tiger, figurine). Compared to these methods, TurboFill generates results with higher quality in visual comparisons. While preserving detail, our results do not exhibit the oversharpness seen in PowerPaint V2 (e.g., tiger), nor the oversaturation issue observed in BrushNet* (4-steps) (e.g., rabbit). The generated content is more realistic (e.g., bird, figurine), leading to a more harmonious image overall. + +# 5.5. Effectiveness of LocalCaptionData + +BrushNet, when trained with a full image description (i.e., global prompt) as input to the base model, exhibits rela + +
SettingsMask Region QualityWhole Image QualityText-Align
Q-AlignCLIPQA+TOPIQQ-AlignCLIPQA+TOPIQCLIP Sim
TurboFill4.57030.73325.27534.72270.71115.539225.352
-LBG4.36720.68585.02624.54690.65705.382625.335
-LFDiff4.18750.65544.86954.49220.64265.329725.062
-RFDiff4.06640.63224.84974.44920.64125.306224.875
+ +Table 3. The effectiveness of different losses. From top to down, we progressively remove specific losses. + +![](images/b7814c1e23257a81b8cdd09e33ba89392b4f0328dee92ecd698a1fd9a3db7b0e.jpg) +Figure 5. The effectiveness of LocalCaptionData. All results are obtained based on 4-step DMD2. (Zoom in for best view) + +![](images/188969e6392e51f795a6e1ffbabeb1e56a7d677838b1656d7fd63f8b1fb153c3.jpg) +Figure 6. The effectiveness of different losses. From left to right, we progressively remove specific losses. (Zoom in for best view) + +tively weak text alignment capabilities, as shown in Fig. 5. In cases where BrushNet attempts to generate content similar or semantically related to existing objects in the image, it often either defaults to complete the background or shifts focus to other categories in the prompt (e.g., a cat). Furthermore, relying on a global prompt limits BrushNet's effectiveness in handling more complex prompts. By using LocalCaptionData, BrushNet* effectively overcomes these limitations without any changes to the training approach, resulting in a significant improvement in text alignment performance. Additionally, TurboFill demonstrates superior performance in generating fine details. + +# 5.6. Ablation Studies. + +Starting from TurboFill, we remove $\mathcal{L}_{\mathrm{BG}}$ , $\mathcal{L}_{\mathrm{Diff}}^F$ , and $\mathcal{L}_{\mathrm{Diff}}^R$ in sequence, with quantitative and qualitative results shown in Fig. 6 and Tab. 3. When $\mathcal{L}_{\mathrm{BG}}$ is removed, there is a clear performance drop in all metrics. In the visualizations, we see that without $\mathcal{L}_{\mathrm{BG}}$ , the color of background region changes noticeably, creating a sharp boundary between the fill-in and background regions. Further removing $\mathcal{L}_{\mathrm{Diff}}^F$ introduces conflicting elements (i.e., house) in the fill-in region, suggesting the discriminator fails to fully capture the holistic scene. Finally, without $\mathcal{L}_{\mathrm{Diff}}^R$ , relying only on GAN loss, the inpainted images exhibit not only inconsistencies with the background but also poor texture and detail. This highlights that GAN loss alone struggles to close the gap between fake and real latents. Only when combining $\mathcal{L}_{\mathrm{Diff}}^F$ , $\mathcal{L}_{\mathrm{Diff}}^R$ , and $\mathcal{L}_{\mathrm{BG}}$ in GAN training does the model achieve enhanced texture, detail, and effective scene harmonization between fill-in and background regions. + +# 6. Conclusion + +In this work, we introduce TurboFill, a fast inpainting model designed to address the computational and quality challenges in diffusion-based inpainting tasks. By directly training an inpainting adapter on a few-step distilled text-to-image diffusion model, TurboFill achieves state-of-the-art inpainting results with significantly smaller inference costs. The proposed three-step adversarial training scheme, incorporating both GAN and diffusion losses, ensures the generation of realistic and harmonious content with fine textures and structural coherence. To evaluate TurboFill's performance, we introduce two new benchmarks, DilationBench and HumanBench, which provide robust evaluations across varying mask complexities and user-centric scenarios. Experimental results demonstrate TurboFill's ability to outperform existing few-step and multi-step BrushNet, setting a new benchmark for practical and efficient inpainting. + +# 7. Acknowledgement + +This work was supported in part by Macau Science and Technology Development Fund under SKLIOTSC-2021-2023, 0022/2022/A1, and 0014/2022/AFJ; in part by Research Committee at University of Macau under MYRG-GRG2023-00058-FST-UMDF; in part by the Guangdong Basic and Applied Basic Research Foundation under Grant 2024A1515012536; in part by National Natural Science Foundation of China (Grant No. 62276251), and the Joint Lab of Shenzhen Key Laboratory of Computer Vision and Pattern Recognition. + +# References + +[1] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM transactions on graphics (TOG), 42 (4):1-11, 2023. 1, 3, 5 +[2] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Trans. Graph., 28(3):24, 2009. 3 +[3] Chenjie Cao, Yunuo Cai, Qiaole Dong, Yikai Wang, and Yanwei Fu. Leftrefill: Filling right canvas based on left reference through generalized text-to-image diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7705-7715, 2024. 3 +[4] Chaofeng Chen, Jiadi Mo, Jingwen Hou, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Topiq: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing, 2024. 5 +[5] Li Chen, Yu Wang, Jiali Zhang, and Xingxing Xu. Powerpaint: Adaptive task prompts for generalizable image inpainting. arXiv preprint arXiv:2401.08965, 2024. 1, 3, 6 +[6] Yifu Chen, Jingwen Chen, Yingwei Pan, Yehao Li, Ting Yao, Zhineng Chen, and Tao Mei. Improving text-guided object inpainting with semantic pre-inpainting. In European Conference on Computer Vision, pages 110-126. Springer, 2025. 3 +[7] Huggingface Diffusers. Sdxl-inpaint. https : / / huggingface . co / diffusers / stable - diffusion-x1- 1 . 0 - inpainting- 0 . 1, Year. Accessed: YYYYY-MM-DD. 1,3 +[8] Stefan Elfwing, Eiji Uchibe, and Kenji Doya. Sigmoid-weighted linear units for neural network function approximation in reinforcement learning. Neural networks, 107:3-11, 2018. 5 +[9] Omar Elharrouss, Noor Almaadeed, Somaya Al-Maadeed, and Younes Akbari. Image inpainting: A review. Neural Processing Letters, 51:2007-2028, 2020. 3 +[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems (NeurIPS), pages 6840-6851, 2020. 4 +[11] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. arXiv preprint arXiv:2403.06976, 2024. 1, 3, 5, 6 +[12] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision, 128(7):1956-1981, 2020. 5 +[13] Shanchuan Lin, Anran Wang, and Xiao Yang. Sdxlightning: Progressive adversarial diffusion distillation. arXiv preprint arXiv:2402.13929, 2024. 2, 3 +[14] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence + +Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5 +[15] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 2, 3, 4, 5 +[16] Hayk Manukyan, Andranik Sargsyan, Barsegh Atanyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Hd-painter: high-resolution and prompt-faithful text-guided image inpainting with diffusion models. arXiv preprint arXiv:2312.14091, 2023. 1, 3, 5 +[17] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 3 +[18] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 +[19] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 5 +[20] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 2, 3 +[21] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 5 +[22] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 5 +[23] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching Imms for visual scoring via discrete text-defined levels. arXiv preprint arXiv:2312.17090, 2023. 5 +[24] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 5 +[25] Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, and Lu Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4818–4829, 2024. 5 +[26] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. arXiv preprint arXiv:2311.18828, 2023. 2, 3 + +[27] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. arXiv preprint arXiv:2405.14867, 2024. 2 +[28] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contextual attention. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5505-5514, 2018. 3 +[29] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4471-4480, 2019. +[30] Yanhong Zeng, Jianlong Fu, Hongyang Chao, and Baining Guo. Learning pyramid-context encoder network for high-quality image inpainting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1486-1494, 2019. +[31] Yanhong Zeng, Jianlong Fu, Hongyang Chao, and Baining Guo. Aggregated contextual transformations for high-resolution image inpainting. IEEE Transactions on Visualization and Computer Graphics, 29(7):3266-3280, 2022. 3 +[32] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 3 +[33] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 3 \ No newline at end of file diff --git a/2025/TurboFill_ Adapting Few-step Text-to-image Model for Fast Image Inpainting/images.zip b/2025/TurboFill_ Adapting Few-step Text-to-image Model for Fast Image Inpainting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9e82b77d2ec7bd8b05b2cf491bf934d0c0115f47 --- /dev/null +++ b/2025/TurboFill_ Adapting Few-step Text-to-image Model for Fast Image Inpainting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d4a5f7a2367f89d7abd219b4514b9e2c659f1421f0d3fd4d0358d95870df0b8 +size 930815 diff --git a/2025/TurboFill_ Adapting Few-step Text-to-image Model for Fast Image Inpainting/layout.json b/2025/TurboFill_ Adapting Few-step Text-to-image Model for Fast Image Inpainting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ea0b2225e13e1aafe2c1e9e389c0b75bdc476000 --- /dev/null +++ b/2025/TurboFill_ Adapting Few-step Text-to-image Model for Fast Image Inpainting/layout.json @@ -0,0 +1,8139 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 103, + 545, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 103, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 66, + 103, + 545, + 121 + ], + "type": "text", + "content": "TurboFill: Adapting Few-step Text-to-image Model for Fast Image Inpainting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 142, + 539, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 142, + 539, + 171 + ], + "spans": [ + { + "bbox": [ + 70, + 142, + 539, + 171 + ], + "type": "text", + "content": "Liangbin Xie" + }, + { + "bbox": [ + 70, + 142, + 539, + 171 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 70, + 142, + 539, + 171 + ], + "type": "text", + "content": " Daniil Pakhomov" + }, + { + "bbox": [ + 70, + 142, + 539, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 70, + 142, + 539, + 171 + ], + "type": "text", + "content": " Zhonghao Wang" + }, + { + "bbox": [ + 70, + 142, + 539, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 70, + 142, + 539, + 171 + ], + "type": "text", + "content": " Zongze Wu" + }, + { + "bbox": [ + 70, + 142, + 539, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 70, + 142, + 539, + 171 + ], + "type": "text", + "content": " Ziyan Chen" + }, + { + "bbox": [ + 70, + 142, + 539, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 70, + 142, + 539, + 171 + ], + "type": "text", + "content": " Yuqian Zhou" + }, + { + "bbox": [ + 70, + 142, + 539, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 70, + 142, + 539, + 171 + ], + "type": "text", + "content": " Haitian Zheng" + }, + { + "bbox": [ + 70, + 142, + 539, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 70, + 142, + 539, + 171 + ], + "type": "text", + "content": " Zhifei Zhang" + }, + { + "bbox": [ + 70, + 142, + 539, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 70, + 142, + 539, + 171 + ], + "type": "text", + "content": " Zhe Lin" + }, + { + "bbox": [ + 70, + 142, + 539, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 70, + 142, + 539, + 171 + ], + "type": "text", + "content": " Jiantao Zhou" + }, + { + "bbox": [ + 70, + 142, + 539, + 171 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 70, + 142, + 539, + 171 + ], + "type": "text", + "content": " Chao Dong" + }, + { + "bbox": [ + 70, + 142, + 539, + 171 + ], + "type": "inline_equation", + "content": "^{2,4\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 173, + 538, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 173, + 538, + 198 + ], + "spans": [ + { + "bbox": [ + 67, + 173, + 538, + 198 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 173, + 538, + 198 + ], + "type": "text", + "content": "State Key Laboratory of Internet of Things for Smart City, University of Macau " + }, + { + "bbox": [ + 67, + 173, + 538, + 198 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 173, + 538, + 198 + ], + "type": "text", + "content": "Shenzhen University of Advanced Technology " + }, + { + "bbox": [ + 67, + 173, + 538, + 198 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 173, + 538, + 198 + ], + "type": "text", + "content": "Adobe " + }, + { + "bbox": [ + 67, + 173, + 538, + 198 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 67, + 173, + 538, + 198 + ], + "type": "text", + "content": "Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 118, + 211, + 241, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 211, + 241, + 219 + ], + "spans": [ + { + "bbox": [ + 118, + 211, + 241, + 219 + ], + "type": "text", + "content": "Yellow baby chick on wooden surface." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 372, + 211, + 529, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 211, + 529, + 219 + ], + "spans": [ + { + "bbox": [ + 372, + 211, + 529, + 219 + ], + "type": "text", + "content": "A miniature blue convertible with its top down" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 57, + 220, + 138, + 302 + ], + "blocks": [ + { + "bbox": [ + 57, + 220, + 138, + 302 + ], + "lines": [ + { + "bbox": [ + 57, + 220, + 138, + 302 + ], + "spans": [ + { + "bbox": [ + 57, + 220, + 138, + 302 + ], + "type": "image", + "image_path": "5e6687e177172a2e3492764e5438725748c2ab485972dd07f6ca401470d5dd6f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 79, + 306, + 117, + 315 + ], + "lines": [ + { + "bbox": [ + 79, + 306, + 117, + 315 + ], + "spans": [ + { + "bbox": [ + 79, + 306, + 117, + 315 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 140, + 220, + 222, + 302 + ], + "blocks": [ + { + "bbox": [ + 140, + 220, + 222, + 302 + ], + "lines": [ + { + "bbox": [ + 140, + 220, + 222, + 302 + ], + "spans": [ + { + "bbox": [ + 140, + 220, + 222, + 302 + ], + "type": "image", + "image_path": "cf85798f3423a8c63336ef0e7fb3ca0143b92d97ff145226d2dc3ee915783591.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 306, + 212, + 315 + ], + "lines": [ + { + "bbox": [ + 149, + 306, + 212, + 315 + ], + "spans": [ + { + "bbox": [ + 149, + 306, + 212, + 315 + ], + "type": "text", + "content": "BrushNet* (50 steps)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 223, + 220, + 306, + 302 + ], + "blocks": [ + { + "bbox": [ + 223, + 220, + 306, + 302 + ], + "lines": [ + { + "bbox": [ + 223, + 220, + 306, + 302 + ], + "spans": [ + { + "bbox": [ + 223, + 220, + 306, + 302 + ], + "type": "image", + "image_path": "55f52f85a82c858ccd571c89d0524f9c9ba5a270297927b9a3e1758ff8d6da74.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 306, + 293, + 315 + ], + "lines": [ + { + "bbox": [ + 235, + 306, + 293, + 315 + ], + "spans": [ + { + "bbox": [ + 235, + 306, + 293, + 315 + ], + "type": "text", + "content": "TurboFill (4 steps)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 220, + 389, + 302 + ], + "blocks": [ + { + "bbox": [ + 307, + 220, + 389, + 302 + ], + "lines": [ + { + "bbox": [ + 307, + 220, + 389, + 302 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 389, + 302 + ], + "type": "image", + "image_path": "05f8888f7625f773512e8f30dec75077631e28de82b334399f7d1de6d10a454a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 306, + 364, + 315 + ], + "lines": [ + { + "bbox": [ + 328, + 306, + 364, + 315 + ], + "spans": [ + { + "bbox": [ + 328, + 306, + 364, + 315 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 390, + 220, + 472, + 302 + ], + "blocks": [ + { + "bbox": [ + 390, + 220, + 472, + 302 + ], + "lines": [ + { + "bbox": [ + 390, + 220, + 472, + 302 + ], + "spans": [ + { + "bbox": [ + 390, + 220, + 472, + 302 + ], + "type": "image", + "image_path": "25027e4998702b86029330b9b3fc906a857b1382f886d0e13e500426466ac479.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 306, + 462, + 315 + ], + "lines": [ + { + "bbox": [ + 398, + 306, + 462, + 315 + ], + "spans": [ + { + "bbox": [ + 398, + 306, + 462, + 315 + ], + "type": "text", + "content": "BrushNet* (50 steps)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 474, + 220, + 555, + 302 + ], + "blocks": [ + { + "bbox": [ + 474, + 220, + 555, + 302 + ], + "lines": [ + { + "bbox": [ + 474, + 220, + 555, + 302 + ], + "spans": [ + { + "bbox": [ + 474, + 220, + 555, + 302 + ], + "type": "image", + "image_path": "95a4da8cca6f387d867ac3ef78f1353c54b5c7eba1b51a9d7d26ee7bdcf40df4.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 485, + 306, + 543, + 315 + ], + "lines": [ + { + "bbox": [ + 485, + 306, + 543, + 315 + ], + "spans": [ + { + "bbox": [ + 485, + 306, + 543, + 315 + ], + "type": "text", + "content": "TurboFill (4 steps)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 63, + 321, + 559, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 321, + 559, + 329 + ], + "spans": [ + { + "bbox": [ + 63, + 321, + 559, + 329 + ], + "type": "text", + "content": "A fluffy brown teddy bear wearing a tiny hat. A tinted sunglasses lens reflecting tall palm trees. Man with cowboy hat and bandana in front of wooden wall." + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 57, + 331, + 138, + 413 + ], + "blocks": [ + { + "bbox": [ + 57, + 331, + 138, + 413 + ], + "lines": [ + { + "bbox": [ + 57, + 331, + 138, + 413 + ], + "spans": [ + { + "bbox": [ + 57, + 331, + 138, + 413 + ], + "type": "image", + "image_path": "d6047e0f407158605f419a54b2e168329792a9b2bc5125b86627f589d69ce254.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 79, + 416, + 116, + 426 + ], + "lines": [ + { + "bbox": [ + 79, + 416, + 116, + 426 + ], + "spans": [ + { + "bbox": [ + 79, + 416, + 116, + 426 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 140, + 331, + 222, + 413 + ], + "blocks": [ + { + "bbox": [ + 140, + 331, + 222, + 413 + ], + "lines": [ + { + "bbox": [ + 140, + 331, + 222, + 413 + ], + "spans": [ + { + "bbox": [ + 140, + 331, + 222, + 413 + ], + "type": "image", + "image_path": "bc05916997be740db3f6a871795c007341f28352a8298aee31e15d90a38fb309.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 416, + 209, + 426 + ], + "lines": [ + { + "bbox": [ + 151, + 416, + 209, + 426 + ], + "spans": [ + { + "bbox": [ + 151, + 416, + 209, + 426 + ], + "type": "text", + "content": "TurboFill (4 steps)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 223, + 331, + 305, + 413 + ], + "blocks": [ + { + "bbox": [ + 223, + 331, + 305, + 413 + ], + "lines": [ + { + "bbox": [ + 223, + 331, + 305, + 413 + ], + "spans": [ + { + "bbox": [ + 223, + 331, + 305, + 413 + ], + "type": "image", + "image_path": "aca537c86d5eab556bee1c64a0e847af6fd4a5fe53f834e119a82344c07aa6c5.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 416, + 282, + 426 + ], + "lines": [ + { + "bbox": [ + 244, + 416, + 282, + 426 + ], + "spans": [ + { + "bbox": [ + 244, + 416, + 282, + 426 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 307, + 331, + 389, + 413 + ], + "blocks": [ + { + "bbox": [ + 307, + 331, + 389, + 413 + ], + "lines": [ + { + "bbox": [ + 307, + 331, + 389, + 413 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 389, + 413 + ], + "type": "image", + "image_path": "01ea8822a4fc25883e047968a37c3f0fb692f7505fe0c5709bc974fc3ce99899.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 416, + 376, + 426 + ], + "lines": [ + { + "bbox": [ + 318, + 416, + 376, + 426 + ], + "spans": [ + { + "bbox": [ + 318, + 416, + 376, + 426 + ], + "type": "text", + "content": "TurboFill (4 steps)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 431, + 555, + 454 + ], + "lines": [ + { + "bbox": [ + 55, + 431, + 555, + 454 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 555, + 454 + ], + "type": "text", + "content": "Figure 1. We propose TurboFill, a fast image inpainting method that leverages a 3-step adversarial training scheme. With only four diffusion steps, TurboFill outperforms the multi-step BrushNet* [11], delivering realistic details and textures with remarkable efficiency." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 390, + 331, + 472, + 413 + ], + "blocks": [ + { + "bbox": [ + 390, + 331, + 472, + 413 + ], + "lines": [ + { + "bbox": [ + 390, + 331, + 472, + 413 + ], + "spans": [ + { + "bbox": [ + 390, + 331, + 472, + 413 + ], + "type": "image", + "image_path": "9d5db6efbb2ed4a309e250980081c7271eb8830fafc8b7802c2c831d85d88a38.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 416, + 447, + 426 + ], + "lines": [ + { + "bbox": [ + 410, + 416, + 447, + 426 + ], + "spans": [ + { + "bbox": [ + 410, + 416, + 447, + 426 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 474, + 331, + 555, + 413 + ], + "blocks": [ + { + "bbox": [ + 474, + 331, + 555, + 413 + ], + "lines": [ + { + "bbox": [ + 474, + 331, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 474, + 331, + 555, + 413 + ], + "type": "image", + "image_path": "477974fc0ee3751fa55b9273381bae05337c978633db4ef010f04012f61362e2.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 485, + 416, + 543, + 426 + ], + "lines": [ + { + "bbox": [ + 485, + 416, + 543, + 426 + ], + "spans": [ + { + "bbox": [ + 485, + 416, + 543, + 426 + ], + "type": "text", + "content": "TurboFill (4 steps)" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "bbox": [ + 152, + 465, + 200, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 465, + 200, + 478 + ], + "spans": [ + { + "bbox": [ + 152, + 465, + 200, + 478 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 53, + 491, + 297, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 297, + 682 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 297, + 682 + ], + "type": "text", + "content": "This paper introduces TurboFill, a fast image inpainting model that enhances a few-step text-to-image diffusion model with an inpainting adapter for high-quality and efficient inpainting. While standard diffusion models generate high-quality results, they incur high computational costs. We overcome this by training an inpainting adapter on a few-step distilled text-to-image model, DMD2, using a novel 3-step adversarial training scheme to ensure realistic, structurally consistent, and visually harmonious inpainted regions. To evaluate TurboFill, we propose two benchmarks: DilationBench, which tests performance across mask sizes, and HumanBench, based on human feedback for complex prompts. Experiments show that TurboFill outperforms both multi-step BrushNet and few-step inpainting methods, setting a new benchmark for high-performance inpainting tasks. The project page is available here." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 315, + 465, + 394, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 465, + 394, + 478 + ], + "spans": [ + { + "bbox": [ + 315, + 465, + 394, + 478 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 485, + 555, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 555, + 557 + ], + "type": "text", + "content": "Image inpainting, the task of filling missing regions in an image, has seen significant advancement with the rise of deep generative models [1, 5, 11, 16]. Existing diffusion-based inpainting methods have shown promising results for various applications such as object removal, regeneration, and text-guided object insertion." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 312, + 558, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 558, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 558, + 556, + 714 + ], + "type": "text", + "content": "The most straightforward way of training diffusion-based inpainting models is to add the masked image (background conditioning) to the input channels of a pretrained diffusion model and finetune it with inpainting data, as shown by SDXL-inpaint [7]. However, such methods typically require fine-tuning the entire network, entailing large training datasets and significant computational resources. In contrast, adapter-based approaches, like BrushNet [11], utilize a ControlNet [33] to capture conditioning information and inject features into a frozen diffusion U-Net. BrushNet achieves state-of-the-art inpainting results while incurring relatively low training costs in terms of data and computational resources, owing to the adapter's efficiency and the" + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 149, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 149, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 149, + 712 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7613" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 41, + 553, + 213 + ], + "blocks": [ + { + "bbox": [ + 57, + 41, + 553, + 213 + ], + "lines": [ + { + "bbox": [ + 57, + 41, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 57, + 41, + 553, + 213 + ], + "type": "image", + "image_path": "078363a1a7eb3661db3f22dc4e2e4436d37c76c5e995a6aa09d7814df2853e0e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 214, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 214, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 214, + 555, + 270 + ], + "type": "text", + "content": "Figure 2. (Zoom in for best view) 1. The multi-step adapter achieves high-quality inpainting results but incurs significant inference costs, requiring over 50 diffusion steps. 2. Applying the pre-trained multi-step BrushNet adapter directly to the few-step U-Net (DMD2) results in artifacts, including oversaturated colors and semantic inconsistencies (e.g., generating a dog with two tails). 3. Training the adapter-DMD2 solely with diffusion loss produces blurred outputs with low-quality inpainting results. 4. In contrast, training the adapter-DMD2 using the proposed 3-step adversarial training scheme yields high-quality inpainting results, requiring only four diffusion steps." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 283, + 282, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 282, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 282, + 295 + ], + "type": "text", + "content": "preservation of the U-Net's original generation capacity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 295, + 295, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 295, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 54, + 295, + 295, + 426 + ], + "type": "text", + "content": "Despite the advantage of BrushNet-like models, they still depend on many iterative sampling steps, leading to substantial inference cost in practice. Interestingly, we found that an inpainting adapter trained on a pretrained text-to-image diffusion model can be seamlessly integrated into its corresponding few-step student model without further training, while yielding reasonable quality results (Fig. 2 ②). This can be attributed to the semantic alignment between the teacher and student models. However, this naive migration approach exhibits quality issues such as color shifts, over-saturation, and loss of details." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 426, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 426, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 54, + 426, + 295, + 533 + ], + "type": "text", + "content": "Alternatively, we can use a two-stage approach by first training a teacher BrushNet and then applying advanced diffusion distillation techniques [13, 15, 20, 26, 27] to the teacher BrushNet to reduce steps while maintaining quality. Although recent distillation methods for diffusion models have demonstrated remarkable performance, we have found that due to BrushNet's large model size, it demands excessive memory and incurs high computational costs, making the training process difficult to manage." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 677 + ], + "type": "text", + "content": "In this paper, we propose a simpler single-stage training strategy by training an inpainting adapter directly on top of a pretrained few-step distilled text-to-image model (e.g., DMD2). We observe that naively training an inpainting adapter and few-step model solely with diffusion loss results in blurry, low-quality outputs (Fig. 2③). To address these shortcomings and enhance inpainting performance, we introduce a diffusion discriminator to guide optimization of the adapter-DMD2 generator. The diffusion discriminator is trained to differentiate between noisy real images and fake latents, while a combined diffusion loss ensures the model learns meaningful semantic representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "In summary, TurboFill is trained with a novel 3-step adversarial training scheme with two generator update steps and one discriminator update step. Our model enables high-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 283, + 555, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 555, + 379 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 555, + 379 + ], + "type": "text", + "content": "quality image inpainting with only a few diffusion steps. Compared to the naive approach of training a multi-step adapter (requiring 8 V100 GPUs for 72 hours) followed by distillation (requiring 64 A100 GPUs for 50 hours), TurboFill attains comparable outcomes using merely 8 A100 GPUs within 50 hours. This accomplishment marks a reduction in training time by over a factor of 10 and a substantial enhancement in training efficiency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 380, + 556, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 556, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 556, + 502 + ], + "type": "text", + "content": "Existing benchmarks often fall short in simulating real-world user operations. To address this gap, we introduce two new benchmarks: DilationBench and HumanBench. DilationBench provides an objective measure by testing the model on the masks with varying dilation sizes, while HumanBench incorporates human user feedback and complex prompts, offering insights into real-world performance. These benchmarks contribute a robust evaluation framework that can facilitate advancements in inpainting models and more accurately reflect user-oriented performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 503, + 555, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 555, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 555, + 563 + ], + "type": "text", + "content": "Extensive quantitative and qualitative evaluations demonstrate that TurboFill not only supports complex prompts and user-driven inpainting tasks but also outperforms existing few-step inpainting models in terms of speed and quality. In summary, our contributions are as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 313, + 570, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 628 + ], + "type": "text", + "content": "- A novel 3-step adversarial training scheme is proposed to train an inpainting adapter directly on top of few-step text-to-image diffusion model, which helps the model to focus on image quality and improves textures, details, and scene harmonization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 630, + 553, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 664 + ], + "type": "text", + "content": "- New benchmarks, DilationBench and HumanBench, are designed to better evaluate inpainting methods in user-centered scenarios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "- We demonstrate that our few-step TurboFill outperforms multi-step BrushNet for the inpainting task and achieves state-of-the-art performance in a few-step setting through extensive experiments." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7614" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 270 + ], + "type": "text", + "content": "Image inpainting, a classic problem in computer vision, focuses on reconstructing missing regions in an image with realistic and natural content. Early methods, including traditional techniques, VAEs, and GANs [2, 3, 9, 28-31], often struggled to produce high-quality results. Recently, diffusion models have gained popularity for their ability to generate high-quality, diverse outputs with fine-grained control, demonstrating significant advancements over earlier approaches [1, 5, 11, 16]. Diffusion models, particularly the Stable Diffusion family, have introduced a robust framework for image generation and inpainting by iteratively denoising latent representations. Stable Diffusion-based methods have achieved impressive results in high-resolution image inpainting, yet are computationally intensive due to multi-step sampling." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 270, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 462 + ], + "type": "text", + "content": "Inpainting tasks based on diffusion models generally follow two approaches: training an entire diffusion model or incorporating external modules. The first approach, like PowerPaint [5], CATDiffusion [6], and SDXL-Inpainting, focuses on adapting the model itself. PowerPaint introduces learnable task prompts along with tailored fine-tuning strategies to guide the model's focus on specific inpainting targets, while SDXL-Inpainting [7] directly utilizes both the background region and the mask as inputs to fine-tune a new model. In contrast, the second approach, exemplified by methods such as ControlNet [33] or T2I-Adapter [17], leverages external modules to enhance the diffusion model's capabilities without retraining it entirely. Recent methods, such as BrushNet [11], employ mask-specific prompts, refining the generated content within the masked area for improved coherence with the background." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 463, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 296, + 571 + ], + "type": "text", + "content": "Our method builds on these developments by leveraging a ControlNet-like adapter for DMD2, integrating both GAN and diffusion losses to enhance image realism and style coherence. By focusing on masked-region-specific prompts and incorporating fine-grained diffusion control, our approach achieves fast and high-quality inpainting results that align closely with user inputs and surrounding context, advancing upon prior work in speed, quality, and flexibility in real-world applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 581, + 133, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 133, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 133, + 594 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 601, + 295, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 295, + 625 + ], + "type": "text", + "content": "The forward diffusion process gradually adds Gaussian noise " + }, + { + "bbox": [ + 55, + 601, + 295, + 625 + ], + "type": "inline_equation", + "content": "\\epsilon_t" + }, + { + "bbox": [ + 55, + 601, + 295, + 625 + ], + "type": "text", + "content": " to a clean image " + }, + { + "bbox": [ + 55, + 601, + 295, + 625 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 601, + 295, + 625 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 119, + 636, + 295, + 650 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 636, + 295, + 650 + ], + "spans": [ + { + "bbox": [ + 119, + 636, + 295, + 650 + ], + "type": "interline_equation", + "content": "x _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t}, \\tag {1}", + "image_path": "a64b099395293edb52b28c87d30178a18942a7cc367222fde151673bd1da54f1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 655, + 296, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 655, + 296, + 692 + ], + "spans": [ + { + "bbox": [ + 55, + 655, + 296, + 692 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 655, + 296, + 692 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t" + }, + { + "bbox": [ + 55, + 655, + 296, + 692 + ], + "type": "text", + "content": " controls noise schedule. In the training phase, a denoising network " + }, + { + "bbox": [ + 55, + 655, + 296, + 692 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 655, + 296, + 692 + ], + "type": "text", + "content": " is trained to predict " + }, + { + "bbox": [ + 55, + 655, + 296, + 692 + ], + "type": "inline_equation", + "content": "\\epsilon_t" + }, + { + "bbox": [ + 55, + 655, + 296, + 692 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 55, + 655, + 296, + 692 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 55, + 655, + 296, + 692 + ], + "type": "text", + "content": ", timestep " + }, + { + "bbox": [ + 55, + 655, + 296, + 692 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 655, + 296, + 692 + ], + "type": "text", + "content": " and text prompt " + }, + { + "bbox": [ + 55, + 655, + 296, + 692 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 655, + 296, + 692 + ], + "type": "text", + "content": ", with the diffusion loss:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 72, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 72, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "L \\left(\\hat {\\epsilon} _ {\\theta}\\right) = \\mathbb {E} _ {x _ {0} \\sim q; \\epsilon_ {t} \\sim \\mathcal {N} (0, 1)} [ \\| \\epsilon_ {\\theta} (x _ {t}, c, t) - \\epsilon_ {t} \\| ^ {2} ]. \\tag {2}", + "image_path": "bf2a4dbdbea76a5bac7ace2003e4c25631028ff58c03ba732ef35eb30aeeb225.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "text", + "content": "To enable condition generation, adapter-based methods [11, 17, 32] introduce additional conditioning information via a lightweight adapter, and achieve efficient training (since the main diffusion network remains frozen), with a slight inference overhead." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 133, + 555, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 264 + ], + "type": "text", + "content": "Traditionally, reverse diffusion processes necessitate a computationally intensive multi-step iterative sampling procedure (typically 20-50 steps). Distillation techniques [13, 15, 20, 26] have been developed to reduce this requirement, achieving accelerated sampling with as few as 1-4 steps. However, these distillation processes are resource-intensive in both computational power and time. In contrast, our approach directly trains the adapter with both the multi-step and few-step text-to-image diffusion model, eliminating the need for a separate distillation process, thereby enabling more efficient training and inference." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 274, + 405, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 405, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 405, + 288 + ], + "type": "text", + "content": "4. Our Approach" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 293, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 510 + ], + "type": "text", + "content": "In this work, we propose a fast inpainting model, TurboFill, which directly trains an inpainting adapter on top of a few-step text-to-image diffusion model. As shown in Fig. 3, there are three components: the fast generator, the slow generator and the diffusion discriminator. All these three components aim to enhance the adapter's capability of inpainting masked images. Specifically, the slow generator helps the adapter to denoise the noisy latents towards the most probable direction of realistic image generation. The fast generator enables the adapter to generate a clean image for critique during training. The diffusion discriminator guides the adapter to generate images with better textures, details and content harmonization. Based on these three components, TurboFill follows a 3-step adversarial training scheme. The training alternates between 3 steps throughout the training process. In the following sections, we provide a detailed description of the training procedure for each step and the components." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 517, + 525, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 525, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 525, + 529 + ], + "type": "text", + "content": "4.1. Training the Adapter in Slow Generator" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "To capture conditioning information and introduce additional features to the diffusion model, we utilize an inpainting adapter with the same ControlNet architecture as BrushNet [11]. Given that SDXL is a multi-step text-to-image generator, we designate the integration of the inpainting adapter and SDXL as the 'slow generator'. In contrast, DMD2 is a few-step text-to-image generator, so the combination of the inpainting adapter and DMD2 is named the 'fast generator'. In both generators, conditioning information is introduced by concatenating the noisy latent " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": ", the background image latent " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "x_{bg}" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": ", and a downsampled binary mask " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": ". This concatenated input is then processed by the inpainting adapter, while " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " is simultaneously fed into the UNet. Then, the features produced by the inpainting adapter are fused with the UNet's features via residual connection." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7615" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 41, + 542, + 233 + ], + "blocks": [ + { + "bbox": [ + 67, + 41, + 542, + 233 + ], + "lines": [ + { + "bbox": [ + 67, + 41, + 542, + 233 + ], + "spans": [ + { + "bbox": [ + 67, + 41, + 542, + 233 + ], + "type": "image", + "image_path": "014f01b28fe85da2eb5385e21f38092ea752b505a51043d421c657ee3f49ed6f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 238, + 555, + 273 + ], + "lines": [ + { + "bbox": [ + 55, + 238, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 555, + 273 + ], + "type": "text", + "content": "Figure 3. The training of TurboFill alternates between 3 steps: (1) optimizing the adapter using the gradient of " + }, + { + "bbox": [ + 55, + 238, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Diff}}^R" + }, + { + "bbox": [ + 55, + 238, + 555, + 273 + ], + "type": "text", + "content": ", and (2) " + }, + { + "bbox": [ + 55, + 238, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{G}}" + }, + { + "bbox": [ + 55, + 238, + 555, + 273 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 238, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BG}}" + }, + { + "bbox": [ + 55, + 238, + 555, + 273 + ], + "type": "text", + "content": " are employed to update the adapter in fast generator, and (3) " + }, + { + "bbox": [ + 55, + 238, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Diff}}^F" + }, + { + "bbox": [ + 55, + 238, + 555, + 273 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 238, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{D}}" + }, + { + "bbox": [ + 55, + 238, + 555, + 273 + ], + "type": "text", + "content": " are jointly applied to update the parameters of the diffusion discriminator module. Note that the adapters in slow generator and fast generator share same weights during training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 282, + 295, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 295, + 318 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 295, + 318 + ], + "type": "text", + "content": "As shown in Fig. 3 (1), the parameters of the inpainting adapter in slow generator are updated based on the real diffusion loss:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 325, + 294, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 325, + 294, + 343 + ], + "spans": [ + { + "bbox": [ + 112, + 325, + 294, + 343 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {D i f f}} ^ {R} = \\min _ {\\epsilon_ {\\theta}} \\mathbb {E} _ {x _ {t}, t \\sim [ 0, T ], \\epsilon \\sim \\mathcal {N} (0, 1)} \\tag {3}", + "image_path": "9d9d43f3c701416f839a032109329b0aafc7d57be41dd58d98df5814a1c2a751.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 346, + 245, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 346, + 245, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 346, + 245, + 370 + ], + "type": "interline_equation", + "content": "\\left\\| \\epsilon - \\epsilon_ {\\theta} ^ {\\mathrm {S D X L}} (x _ {t}, [ x _ {b g}, x _ {t}, m ], t) \\right\\| _ {2} ^ {2},", + "image_path": "f45b66dd56627146a88a3cfeb92742e8889f2c8871cd3c9025216643d6fbe8e5.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 375, + 295, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 295, + 399 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 375, + 295, + 399 + ], + "type": "inline_equation", + "content": "T = 1,000" + }, + { + "bbox": [ + 55, + 375, + 295, + 399 + ], + "type": "text", + "content": " and DDPM scheduler [10] is employed to generate " + }, + { + "bbox": [ + 55, + 375, + 295, + 399 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 55, + 375, + 295, + 399 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 405, + 263, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 263, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 263, + 418 + ], + "type": "text", + "content": "4.2. Training the Adapter in Fast Generator" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 422, + 296, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 507 + ], + "type": "text", + "content": "Given that adversarial training is highly effective in minimizing the divergence between different distributions, we employ adversarial training to minimize the discrepancy between real latent distribution and the fake latent distribution " + }, + { + "bbox": [ + 55, + 422, + 296, + 507 + ], + "type": "inline_equation", + "content": "\\{\\hat{x}_0\\}" + }, + { + "bbox": [ + 55, + 422, + 296, + 507 + ], + "type": "text", + "content": ". Specifically, we first generate the clean fake latent " + }, + { + "bbox": [ + 55, + 422, + 296, + 507 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 55, + 422, + 296, + 507 + ], + "type": "text", + "content": " from the noisy latent " + }, + { + "bbox": [ + 55, + 422, + 296, + 507 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 55, + 422, + 296, + 507 + ], + "type": "text", + "content": " with the fast generator in a single step, as described by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 513, + 295, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 513, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 67, + 513, + 295, + 540 + ], + "type": "interline_equation", + "content": "\\hat {x} _ {0} = \\frac {x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\times \\epsilon_ {\\theta} ^ {\\mathrm {D M D 2}} \\left(x _ {t} , \\left[ x _ {b g} , x _ {t} , m \\right] , t\\right)}{\\sqrt {\\bar {\\alpha} _ {t}}}, \\tag {4}", + "image_path": "cad4eacb446f5245984abeccc9d753871c2cfd954429949cf928b88150bc0bfc.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t" + }, + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "type": "text", + "content": " is determined by the LCM scheduler [15] and " + }, + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "type": "text", + "content": " is the sampled timestep. Then, the noisy version " + }, + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\hat{x}_t" + }, + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "type": "text", + "content": " originated from " + }, + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\hat{x_0}" + }, + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "type": "text", + "content": " is fed into the diffusion discriminator. Meanwhile, to preserve the background region, " + }, + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\hat{x_0}" + }, + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "type": "text", + "content": " is used to compute the background loss." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "Diffusion discriminator. As shown in Fig. 3 (2), the diffusion discriminator " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\phi}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " consists of the SDXL encoder, the assistant encoder, and the convolution-based classifier. For assistant encoder, we use the same encoder structure of inpainting adapter. It takes concatenated noisy latent, downsampled binary mask, and masked image latent " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "x_{bg}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " as the input. The features extracted by the assistant encoder are progressively integrated into the SDXL encoder. The final feature maps are then passed through the classifier and" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 282, + 555, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 555, + 307 + ], + "type": "text", + "content": "mapped to a one-dimensional vector. Based on the vector, we calculate " + }, + { + "bbox": [ + 313, + 282, + 555, + 307 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{G}}" + }, + { + "bbox": [ + 313, + 282, + 555, + 307 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 323, + 316, + 553, + 344 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 316, + 553, + 344 + ], + "spans": [ + { + "bbox": [ + 323, + 316, + 553, + 344 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {G}} = \\min _ {\\epsilon_ {\\theta}} - \\mathbb {E} _ {y _ {t}, t \\sim [ 0, T ]} \\left[ \\log \\mathcal {D} _ {\\phi} \\left(\\hat {x} _ {0}, \\left[ x _ {b g}, \\hat {x} _ {0}, m \\right], t\\right)\\right), \\tag {5}", + "image_path": "c9bf6d57eab3d65dc3db4dd6e460483e4030fbb8e9b5005de0f2e3559ed5a1fa.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 345, + 392, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 345, + 392, + 356 + ], + "spans": [ + { + "bbox": [ + 314, + 345, + 392, + 356 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 314, + 345, + 392, + 356 + ], + "type": "inline_equation", + "content": "T = 1,000" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 357, + 554, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 554, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 554, + 416 + ], + "type": "text", + "content": "Background preservation. For the image inpainting task, it is essential to preserve the background region. Therefore, for the clean fake latent " + }, + { + "bbox": [ + 313, + 357, + 554, + 416 + ], + "type": "inline_equation", + "content": "\\hat{x_0}" + }, + { + "bbox": [ + 313, + 357, + 554, + 416 + ], + "type": "text", + "content": ", we apply a reconstruction loss to the background region to minimize the distance between " + }, + { + "bbox": [ + 313, + 357, + 554, + 416 + ], + "type": "inline_equation", + "content": "\\hat{x_0}" + }, + { + "bbox": [ + 313, + 357, + 554, + 416 + ], + "type": "text", + "content": " and the clean latent " + }, + { + "bbox": [ + 313, + 357, + 554, + 416 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 357, + 554, + 416 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 369, + 427, + 553, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 427, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 369, + 427, + 553, + 442 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {B G}} = \\left\\| x _ {0} \\odot m - \\hat {x} _ {0} \\odot m \\right\\| _ {2} ^ {2}, \\tag {6}", + "image_path": "34334aac08eccc14a8cc72bf330eaa49184427bf2e3c75541cd1e23e33500377.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "text", + "content": " represents point-wise multiplication. Based on " + }, + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\lambda_1\\mathcal{L}_{\\mathcal{G}} + \\lambda_2\\mathcal{L}_{\\mathrm{BG}}" + }, + { + "bbox": [ + 313, + 447, + 555, + 483 + ], + "type": "text", + "content": ", in this step, only the parameters of the inpainting adapter are updated." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 491, + 509, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 509, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 509, + 503 + ], + "type": "text", + "content": "4.3. Training the Diffusion Discriminator" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 508, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 555, + 568 + ], + "type": "text", + "content": "When updating the discriminator, we fix the parameters of the inpainting adapter and update the parameters of the assistant encoder and classifier. Specifically, both the real noisy latent " + }, + { + "bbox": [ + 313, + 508, + 555, + 568 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 508, + 555, + 568 + ], + "type": "text", + "content": " and the fake noisy latent " + }, + { + "bbox": [ + 313, + 508, + 555, + 568 + ], + "type": "inline_equation", + "content": "\\hat{x}_{t}" + }, + { + "bbox": [ + 313, + 508, + 555, + 568 + ], + "type": "text", + "content": " are input into the diffusion discriminator to compute the loss " + }, + { + "bbox": [ + 313, + 508, + 555, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{D}}" + }, + { + "bbox": [ + 313, + 508, + 555, + 568 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 318, + 578, + 553, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 578, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 318, + 578, + 553, + 617 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathcal {D}} = - \\mathbb {E} _ {\\hat {x} _ {0, t}, t \\sim [ 0, T ]} \\left[ \\log \\left(1 - \\mathcal {D} _ {\\phi} (\\hat {x} _ {0}, [ x _ {b g}, \\hat {x} _ {0}, m ], t)\\right)\\right) ] \\\\ - \\mathbb {E} _ {x _ {t}, t \\sim [ 0, T ]} \\left[ \\right. \\log \\mathcal {D} _ {\\phi} \\left(x _ {t}, \\left[ x _ {\\mathrm {b g}}, x _ {t}, m \\right], t\\right)\\left. \\right)\\left. \\right]. \\tag {7} \\\\ \\end{array}", + "image_path": "d048c18a983900ef201c51129c4751e5ec7704bf160cf912dad82c5655f1082b.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": "Although the inpainting adapter optimized only with GAN loss can generate images, the generated images are of inferior quality and commonly incorporate unwanted elements. We attribute this issue to the limitations of the GAN loss, which rarely assists the discriminator in comprehending the image structure. In contrast, diffusion loss excels at scene understanding. Therefore, we introduce the fake diffusion loss to train the discriminator while allowing the" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7616" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "GAN loss to focus more on enhancing image textures and object details. To compute the diffusion loss for the fake noisy latent " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "\\hat{x}_t" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": ", we introduce an assistant decoder, following the assistant encoder. The fake diffusion loss is:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 111, + 125, + 295, + 146 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 125, + 295, + 146 + ], + "spans": [ + { + "bbox": [ + 111, + 125, + 295, + 146 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {D i f f}} ^ {F} = \\min _ {\\mathcal {A} _ {\\psi}} \\mathbb {E} _ {y _ {t}, t \\sim [ 0, T ], \\epsilon \\sim \\mathcal {N} (0, 1)} \\tag {8}", + "image_path": "aca5f0a222ea90e3fe8628fdbfeff6c22eea9ada85d389e0b40b6e8d237ec1d8.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 148, + 249, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 148, + 249, + 171 + ], + "spans": [ + { + "bbox": [ + 104, + 148, + 249, + 171 + ], + "type": "interline_equation", + "content": "\\left\\| \\epsilon - \\mathcal {A} _ {\\psi} ^ {\\mathrm {S D X L}} (x _ {t}, [ x _ {b g}, \\hat {x} _ {t}, m ], t) \\right\\| _ {2} ^ {2},", + "image_path": "e83e1561c39f6799283fe528f78b276f89f4805dddf3c11209f7d15dc8d00ae1.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 175, + 296, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 296, + 223 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 296, + 223 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 175, + 296, + 223 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\psi}" + }, + { + "bbox": [ + 55, + 175, + 296, + 223 + ], + "type": "text", + "content": " represents the parameters of assistant encoder and decoder. The combination of " + }, + { + "bbox": [ + 55, + 175, + 296, + 223 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Diff}}^{F} + \\lambda_{3}\\mathcal{L}_{\\mathcal{D}}" + }, + { + "bbox": [ + 55, + 175, + 296, + 223 + ], + "type": "text", + "content": " is used as the loss to optimize the assistant encoder, assistant decoder modules, and classifier." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 232, + 137, + 246 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 137, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 137, + 246 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 252, + 177, + 263 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 177, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 177, + 263 + ], + "type": "text", + "content": "5.1. Datasets and Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 270, + 156, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 156, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 156, + 282 + ], + "type": "text", + "content": "5.1.1. Training Dataset" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 285, + 296, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 296, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 296, + 416 + ], + "type": "text", + "content": "To enable editing through the input of a local mask and the corresponding prompt for the masked region, we construct a dataset that includes local masks and their associated textual descriptions. Specifically, we crawl approximately 1,200,000 images from Internet and use Florence-2 [25] to perform the dense region caption task with the prompt " + }, + { + "bbox": [ + 55, + 285, + 296, + 416 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 55, + 285, + 296, + 416 + ], + "type": "text", + "content": " DENSE_REGION_CAPTION>. This task identifies the main objects in the images and generates their corresponding short descriptions. These descriptions are then input into SAM2 [19] to obtain the segmentation masks for the objects. We name this dataset as LocalCaptionData." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 421, + 168, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 168, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 168, + 431 + ], + "type": "text", + "content": "5.1.2. Evaluation Datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 435, + 296, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 296, + 602 + ], + "type": "text", + "content": "The commonly used public datasets in the field of image inpainting are primarily OpenImage V6 [12], MSCOCO [14], and BrushBench [11]. Among these, OpenImage V6 and MSCOCO provide precise segmentation masks and bounding box masks, which are used to perform inpainting on the masked regions based on class names. BrushBench only offers accurate segmentation masks, where inpainting is performed on the masked regions using a global image description. However, considering that users typically perform inpainting with rough masks and corresponding descriptions (which may sometimes be complex and contain multiple attributes), these three datasets significantly diverge from real-world scenarios. To better simulate user operations, we construct two evaluation datasets:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 55, + 605, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 652 + ], + "type": "text", + "content": "- DilationBench: The masks are created by randomly dilating the segmentation masks, and the prompts describe the content within the masked area of the original image. DilationBench contains 300 pairs of masks and prompts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 652, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 295, + 688 + ], + "type": "text", + "content": "- HumanBench: The masks are manually labeled, and the prompts are manually written. HumanBench contains 150 pairs of masks and prompts." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "The descriptions and visualizations of the two benchmark datasets are included in the supplementary material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 424, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 424, + 83 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 424, + 83 + ], + "type": "text", + "content": "5.1.3. Evaluation Metrics" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 87, + 555, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 87, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 87, + 555, + 291 + ], + "type": "text", + "content": "To evaluate the quality of inpainted images, we use three metrics: Q-Align [23], CLIPIQA+ [22], TOPIQ [4]. They assess visual quality from different perspectives. Q-Align, which leverages a large language model (LLM) to generate visual scores, is state-of-the-art on various Image Quality Assessment (IQA) leaderboards1. Excluding Q-Align, CLIPIQA+ and TOPIQ rank first in non-reference and image aesthetic benchmarks, respectively, making them reliable indicators of visual quality and providing complementary insights. For a more comprehensive evaluation of inpainting results, we consider not only the masked region but also the entire image, ensuring that both local quality in the inpainted area and overall visual coherence across the entire image are assessed. To measure the consistency between the text and the content generated in the mask region, we use the CLIP Similarity metric [18]. Higher values for all four metrics indicate better performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 300, + 441, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 441, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 441, + 312 + ], + "type": "text", + "content": "5.2. Experimental Settings" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 317, + 446, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 446, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 446, + 330 + ], + "type": "text", + "content": "5.2.1. Network Configurations" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 333, + 554, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 554, + 464 + ], + "type": "text", + "content": "Inpainting adapter adopts the same architecture as Brush-Net, with modifications to the SDXL model. Specifically, the self-attention and cross-attention modules are removed. Note that our approach does not impose any architectural constraints on inpainting adapter. Assistant also adopts U-Net like structure, for simplicity, we directly adopt the same architecture as inpainting adapter. The classifier is composed of convolutional layers, GroupNorm [24], and SiLU [8] activation, which together map the feature map to a 1-dimensional vector. The details of classifier are described in the supplementary material." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 465, + 554, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 554, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 554, + 538 + ], + "type": "text", + "content": "When using SDXL as the base model for noise prediction, we employ a DDIM scheduler [21] and sample randomly from 1,000 timesteps. In contrast, when using DMD2 as the base model, we use an LCM scheduler [15] and randomly sample from four specific timesteps: \\{999, 749, 499, 249\\}." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 544, + 443, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 443, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 443, + 555 + ], + "type": "text", + "content": "5.2.2. Implementation Details" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "text", + "content": "All experiments are performed on 8 A100 GPUs with 80GB memory. During training, the learning rate for parameter updates is set to 1e-5, with a batch size of 2 and gradient accumulation steps configured to 4. We adopt mixed precision to optimize computational efficiency. " + }, + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "inline_equation", + "content": "\\lambda_{1},\\lambda_{2},\\lambda_{3}" + }, + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "text", + "content": " are set to " + }, + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "inline_equation", + "content": "1e^{-3}" + }, + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "inline_equation", + "content": "1e^{-1}" + }, + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "inline_equation", + "content": "1e^{-2}" + }, + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "text", + "content": ", respectively. All models are trained for 40K iterations using the AdamW optimizer." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 652, + 462, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 462, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 462, + 665 + ], + "type": "text", + "content": "5.3. Quantitative Comparisons" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 670, + 554, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 554, + 694 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 554, + 694 + ], + "type": "text", + "content": "We compare TurboFill with existing state-of-the-art image inpainting methods, namely BLD [1], HD-Painter [16]," + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 328, + 703, + 522, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 703, + 522, + 713 + ], + "spans": [ + { + "bbox": [ + 328, + 703, + 522, + 713 + ], + "type": "text", + "content": "https://iqa-pytorch.readthedocs.io/en/latest/benchmark.html" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7617" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 55, + 542, + 513 + ], + "blocks": [ + { + "bbox": [ + 282, + 44, + 323, + 54 + ], + "lines": [ + { + "bbox": [ + 282, + 44, + 323, + 54 + ], + "spans": [ + { + "bbox": [ + 282, + 44, + 323, + 54 + ], + "type": "text", + "content": "' tiger'" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 55, + 542, + 513 + ], + "lines": [ + { + "bbox": [ + 67, + 55, + 542, + 513 + ], + "spans": [ + { + "bbox": [ + 67, + 55, + 542, + 513 + ], + "type": "image", + "image_path": "96743521a278e94527a3f6d876b6c9df93ccc04878204fc3e48c35c7829b3dc4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 520, + 555, + 544 + ], + "lines": [ + { + "bbox": [ + 55, + 520, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 555, + 544 + ], + "type": "text", + "content": "Figure 4. Comparison of previous inpainting methods and BrushNet on DilationBench. Compared to other methods, TurboFill generates more realistic details and textures in just 4 steps, while achieving good scene harmonization. (Zoom in for best view)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 552, + 295, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 295, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 295, + 637 + ], + "type": "text", + "content": "SDXL-Inpainting" + }, + { + "bbox": [ + 55, + 552, + 295, + 637 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 55, + 552, + 295, + 637 + ], + "type": "text", + "content": ", BrushNet-Rand [11], and PowerPaint [5]. All of these methods follow a multi-step approach, requiring 50 steps during inference to generate inpainting results. While BrushNet-Rand, trained with random masks, effectively handles imprecise masks and uses a full image description during inference, the other methods rely only on descriptions of the masked region." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 637, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 637, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 637, + 295, + 685 + ], + "type": "text", + "content": "Since we adopt the architecture design of BrushNet, for fair comparison, we also train BrushNet using LocalCaptionData, referring to this configuration as BrushNet*. Notably, BrushNet, as an external module, can be directly" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 553, + 555, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 553, + 555, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 553, + 555, + 661 + ], + "type": "text", + "content": "adapted into a 4-step diffusion model (e.g., DMD2) to form a 4-step BrushNet configuration. In contrast, methods such as HD-Painter and SDXL-Inpainting, which modify the base model structure, cannot integrate with a 4-step diffusion model. PowerPaint V2, an advanced version of BrushNet and SD1.5, incorporates learnable task prompts and specialized fine-tuning strategies across four distinct inpainting tasks. However, since PowerPaint lacks an SDXL version, it cannot be adapted to a 4-step DMD2 model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Comparative results for these methods on both Dilation-Bench and HumanBench datasets are presented in Tab. 1 and Tab. 2. Given that HumanBench, like real inpainting tasks, does not provide a global description of the en" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 285, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 285, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 285, + 712 + ], + "type": "text", + "content": "https://huggingface.co/diffusers/stable-diffusion-x1-1.0-inpainting-0.1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7618" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 41, + 539, + 166 + ], + "blocks": [ + { + "bbox": [ + 70, + 41, + 539, + 166 + ], + "lines": [ + { + "bbox": [ + 70, + 41, + 539, + 166 + ], + "spans": [ + { + "bbox": [ + 70, + 41, + 539, + 166 + ], + "type": "table", + "html": "
MetricsMask Region QualityWhole Image QualityText-Align
MethodQ-AlignCLIPIQA+TOPIQQ-AlignCLIPIQATOPIQCLIP Sim
MultiStep (50 steps)BLD4.18360.68745.17814.33200.67915.466224.884
HD-Painter3.97270.66114.93454.48050.64395.456625.585
SDXL-Inpainting4.24610.66665.15334.61720.66955.540924.848
BrushNet-Rand4.26170.67295.17384.62890.68185.577921.634
BrushNet*4.46880.71375.25044.53120.69745.558525.389
PowerPaint V24.77730.77655.54684.72270.73715.647426.256
FewStep (4 steps)BrushNet-Rand4.16020.65384.99274.55470.66085.459921.831
BrushNet*4.18360.65724.99394.44920.64275.396125.341
TurboFill4.57030.73325.27534.71880.71115.539225.352
", + "image_path": "a6f984202f298aa971368981aaef353f3bac92b642c309d3ceeec47bad019ce2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 70, + 201, + 539, + 305 + ], + "blocks": [ + { + "bbox": [ + 55, + 168, + 555, + 192 + ], + "lines": [ + { + "bbox": [ + 55, + 168, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 555, + 192 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons among TurboFill and other diffusion based inpainting models in DilationBench. Red and blue indicates the best multi-step and the best few-step performances, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 201, + 539, + 305 + ], + "lines": [ + { + "bbox": [ + 70, + 201, + 539, + 305 + ], + "spans": [ + { + "bbox": [ + 70, + 201, + 539, + 305 + ], + "type": "table", + "html": "
MetricsMask Region QualityWhole Image QualityText-Align
MethodQ-AlignCLIPIQA+TOPIQQ-AlignCLIPIQATOPIQCLIP Sim
MultiStep (50 steps)BLD4.15230.69085.29963.90620.65795.419924.767
HD-Painter3.98440.64935.08734.20700.61195.438125.996
SDXL-Inpainting3.95510.64055.20404.04690.62885.456524.041
BrushNet*4.25780.70635.43404.08980.66945.548025.366
PowerPaint V24.55860.75295.54754.36330.71125.633826.264
FewStep (4 steps)BrushNet*4.05080.63275.10414.14840.60035.350525.473
Ours4.47270.72575.38654.32030.68225.499225.710
", + "image_path": "18b6a3d329ca076662f047aae81b0ad2d0af841758baa07700627ded261478b6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons among TurboFill and other diffusion based inpainting models in HumanBench. Red and blue indicates the best multi-step and the best few-step performances, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 335, + 294, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 335, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 54, + 335, + 294, + 432 + ], + "type": "text", + "content": "tire image, we exclude BrushNet-Rand from evaluation on this dataset. From Tab. 1, we find that BrushNet*, trained with LocalCaptionData, shows improvements of 3.755 and 3.52 in the text-align metric for the 50-step and 4-step models, respectively, compared to BrushNet. This suggests that training with segmentation masks and paired local descriptions, rather than global descriptions, significantly enhances BrushNet's text alignment capability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 434, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 434, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 54, + 434, + 295, + 602 + ], + "type": "text", + "content": "Additionally, we observe that directly inserting BrushNet-Rand and BrushNet* into DMD2 to form a 4-step model results in no significant change in text alignment performance. However, the image quality of the final inpainting results decreases noticeably. This suggests that while replacing SDXL with a few-step model (DMD2) can accelerate BrushNet, it leads to a reduction in quality. In contrast, TurboFill can effectively mitigate this issue, showing significant improvements across all three metrics—mask region quality, whole image quality, and text alignment. Even with 4 steps, our method outperforms the 50-step BrushNet-Rand and BrushNet* in many image quality metrics (e.g., Q-Align, CLIPIQA+). This phenomenon can be observed on both benchmarks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "content": "PowerPaint V2 currently achieves the best results across these metrics. When comparing our method with PowerPaint, we find that the gap in whole image quality is smaller than in mask image quality. Visual comparisons (Fig. 4) reveal that although PowerPaint V2 generates rich textures, the results in the mask region are less natural, with noticeable style mismatch compared to the background region. This inconsistency leads to a lower performance in whole image quality compared to mask region quality. In contrast," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 336, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 555, + 373 + ], + "type": "text", + "content": "our method produces more coherent results. Since no existing metrics effectively capture these differences, we conduct a user study, detailed in the supplementary file." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 379, + 451, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 451, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 451, + 392 + ], + "type": "text", + "content": "5.4. Qualitative Comparison" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 398, + 555, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 555, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 555, + 661 + ], + "type": "text", + "content": "The qualitative comparison on DilationBench is shown in Fig. 4. The results from SDXL-Inpainting exhibit a yellowish color bias (e.g., tiger, man) and lack detail (e.g., boy, figurine). For PowerPaint V2, the inpainted results are very sharp in texture and rich in color, producing high scores on most IQA metrics. However, these results suffer from severe distortions (e.g., a tiger with two heads, a person with two heads, overly elongated legs) and appear less realistic and natural (e.g., third row). The results generated by BrushNet* exhibit significant artifacts, such as the tiger's fur (1st) and the man's beard (5th), as well as the appearance of strange backgrounds (5th). When BrushNet* is inserted into the 4-step DMD2 model, we observe that the results lack significant detail (e.g., 5th row) and exhibit issues with oversaturation (e.g., tiger, figurine). Compared to these methods, TurboFill generates results with higher quality in visual comparisons. While preserving detail, our results do not exhibit the oversharpness seen in PowerPaint V2 (e.g., tiger), nor the oversaturation issue observed in BrushNet* (4-steps) (e.g., rabbit). The generated content is more realistic (e.g., bird, figurine), leading to a more harmonious image overall." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 671, + 499, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 499, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 499, + 684 + ], + "type": "text", + "content": "5.5. Effectiveness of LocalCaptionData" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "content": "BrushNet, when trained with a full image description (i.e., global prompt) as input to the base model, exhibits rela" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7619" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 127, + 41, + 482, + 109 + ], + "blocks": [ + { + "bbox": [ + 127, + 41, + 482, + 109 + ], + "lines": [ + { + "bbox": [ + 127, + 41, + 482, + 109 + ], + "spans": [ + { + "bbox": [ + 127, + 41, + 482, + 109 + ], + "type": "table", + "html": "
SettingsMask Region QualityWhole Image QualityText-Align
Q-AlignCLIPQA+TOPIQQ-AlignCLIPQA+TOPIQCLIP Sim
TurboFill4.57030.73325.27534.72270.71115.539225.352
-LBG4.36720.68585.02624.54690.65705.382625.335
-LFDiff4.18750.65544.86954.49220.64265.329725.062
-RFDiff4.06640.63224.84974.44920.64125.306224.875
", + "image_path": "f0bb17df08fd73b19104df8417c40a990ad27f46606ea7e8f4fd01c95924f935.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 113, + 112, + 496, + 124 + ], + "lines": [ + { + "bbox": [ + 113, + 112, + 496, + 124 + ], + "spans": [ + { + "bbox": [ + 113, + 112, + 496, + 124 + ], + "type": "text", + "content": "Table 3. The effectiveness of different losses. From top to down, we progressively remove specific losses." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 59, + 134, + 290, + 399 + ], + "blocks": [ + { + "bbox": [ + 59, + 134, + 290, + 399 + ], + "lines": [ + { + "bbox": [ + 59, + 134, + 290, + 399 + ], + "spans": [ + { + "bbox": [ + 59, + 134, + 290, + 399 + ], + "type": "image", + "image_path": "b7814c1e23257a81b8cdd09e33ba89392b4f0328dee92ecd698a1fd9a3db7b0e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 411, + 295, + 433 + ], + "lines": [ + { + "bbox": [ + 55, + 411, + 295, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 295, + 433 + ], + "type": "text", + "content": "Figure 5. The effectiveness of LocalCaptionData. All results are obtained based on 4-step DMD2. (Zoom in for best view)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 56, + 446, + 295, + 509 + ], + "blocks": [ + { + "bbox": [ + 56, + 446, + 295, + 509 + ], + "lines": [ + { + "bbox": [ + 56, + 446, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 446, + 295, + 509 + ], + "type": "image", + "image_path": "188969e6392e51f795a6e1ffbabeb1e56a7d677838b1656d7fd63f8b1fb153c3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 520, + 295, + 543 + ], + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 543 + ], + "type": "text", + "content": "Figure 6. The effectiveness of different losses. From left to right, we progressively remove specific losses. (Zoom in for best view)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "tively weak text alignment capabilities, as shown in Fig. 5. In cases where BrushNet attempts to generate content similar or semantically related to existing objects in the image, it often either defaults to complete the background or shifts focus to other categories in the prompt (e.g., a cat). Furthermore, relying on a global prompt limits BrushNet's effectiveness in handling more complex prompts. By using LocalCaptionData, BrushNet* effectively overcomes these limitations without any changes to the training approach, resulting in a significant improvement in text alignment performance. Additionally, TurboFill demonstrates superior performance in generating fine details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 134, + 417, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 134, + 417, + 146 + ], + "spans": [ + { + "bbox": [ + 313, + 134, + 417, + 146 + ], + "type": "text", + "content": "5.6. Ablation Studies." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 151, + 555, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 151, + 555, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 151, + 555, + 355 + ], + "type": "text", + "content": "Starting from TurboFill, we remove " + }, + { + "bbox": [ + 313, + 151, + 555, + 355 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BG}}" + }, + { + "bbox": [ + 313, + 151, + 555, + 355 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 151, + 555, + 355 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Diff}}^F" + }, + { + "bbox": [ + 313, + 151, + 555, + 355 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 151, + 555, + 355 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Diff}}^R" + }, + { + "bbox": [ + 313, + 151, + 555, + 355 + ], + "type": "text", + "content": " in sequence, with quantitative and qualitative results shown in Fig. 6 and Tab. 3. When " + }, + { + "bbox": [ + 313, + 151, + 555, + 355 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BG}}" + }, + { + "bbox": [ + 313, + 151, + 555, + 355 + ], + "type": "text", + "content": " is removed, there is a clear performance drop in all metrics. In the visualizations, we see that without " + }, + { + "bbox": [ + 313, + 151, + 555, + 355 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BG}}" + }, + { + "bbox": [ + 313, + 151, + 555, + 355 + ], + "type": "text", + "content": ", the color of background region changes noticeably, creating a sharp boundary between the fill-in and background regions. Further removing " + }, + { + "bbox": [ + 313, + 151, + 555, + 355 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Diff}}^F" + }, + { + "bbox": [ + 313, + 151, + 555, + 355 + ], + "type": "text", + "content": " introduces conflicting elements (i.e., house) in the fill-in region, suggesting the discriminator fails to fully capture the holistic scene. Finally, without " + }, + { + "bbox": [ + 313, + 151, + 555, + 355 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Diff}}^R" + }, + { + "bbox": [ + 313, + 151, + 555, + 355 + ], + "type": "text", + "content": ", relying only on GAN loss, the inpainted images exhibit not only inconsistencies with the background but also poor texture and detail. This highlights that GAN loss alone struggles to close the gap between fake and real latents. Only when combining " + }, + { + "bbox": [ + 313, + 151, + 555, + 355 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Diff}}^F" + }, + { + "bbox": [ + 313, + 151, + 555, + 355 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 151, + 555, + 355 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Diff}}^R" + }, + { + "bbox": [ + 313, + 151, + 555, + 355 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 151, + 555, + 355 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BG}}" + }, + { + "bbox": [ + 313, + 151, + 555, + 355 + ], + "type": "text", + "content": " in GAN training does the model achieve enhanced texture, detail, and effective scene harmonization between fill-in and background regions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 359, + 388, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 388, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 388, + 371 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 379, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 555, + 570 + ], + "type": "text", + "content": "In this work, we introduce TurboFill, a fast inpainting model designed to address the computational and quality challenges in diffusion-based inpainting tasks. By directly training an inpainting adapter on a few-step distilled text-to-image diffusion model, TurboFill achieves state-of-the-art inpainting results with significantly smaller inference costs. The proposed three-step adversarial training scheme, incorporating both GAN and diffusion losses, ensures the generation of realistic and harmonious content with fine textures and structural coherence. To evaluate TurboFill's performance, we introduce two new benchmarks, DilationBench and HumanBench, which provide robust evaluations across varying mask complexities and user-centric scenarios. Experimental results demonstrate TurboFill's ability to outperform existing few-step and multi-step BrushNet, setting a new benchmark for practical and efficient inpainting." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 574, + 424, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 424, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 424, + 587 + ], + "type": "text", + "content": "7. Acknowledgement" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "This work was supported in part by Macau Science and Technology Development Fund under SKLIOTSC-2021-2023, 0022/2022/A1, and 0014/2022/AFJ; in part by Research Committee at University of Macau under MYRG-GRG2023-00058-FST-UMDF; in part by the Guangdong Basic and Applied Basic Research Foundation under Grant 2024A1515012536; in part by National Natural Science Foundation of China (Grant No. 62276251), and the Joint Lab of Shenzhen Key Laboratory of Computer Vision and Pattern Recognition." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7620" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "text", + "content": "[1] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM transactions on graphics (TOG), 42 (4):1-11, 2023. 1, 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 167 + ], + "type": "text", + "content": "[2] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Trans. Graph., 28(3):24, 2009. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 170, + 295, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 170, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 62, + 170, + 295, + 224 + ], + "type": "text", + "content": "[3] Chenjie Cao, Yunuo Cai, Qiaole Dong, Yikai Wang, and Yanwei Fu. Leftrefill: Filling right canvas based on left reference through generalized text-to-image diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7705-7715, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 225, + 295, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 225, + 295, + 278 + ], + "spans": [ + { + "bbox": [ + 62, + 225, + 295, + 278 + ], + "type": "text", + "content": "[4] Chaofeng Chen, Jiadi Mo, Jingwen Hou, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Topiq: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing, 2024. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 280, + 295, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 295, + 313 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 295, + 313 + ], + "type": "text", + "content": "[5] Li Chen, Yu Wang, Jiali Zhang, and Xingxing Xu. Powerpaint: Adaptive task prompts for generalizable image inpainting. arXiv preprint arXiv:2401.08965, 2024. 1, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 315, + 295, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 315, + 295, + 367 + ], + "spans": [ + { + "bbox": [ + 62, + 315, + 295, + 367 + ], + "type": "text", + "content": "[6] Yifu Chen, Jingwen Chen, Yingwei Pan, Yehao Li, Ting Yao, Zhineng Chen, and Tao Mei. Improving text-guided object inpainting with semantic pre-inpainting. In European Conference on Computer Vision, pages 110-126. Springer, 2025. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 369, + 295, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 295, + 413 + ], + "type": "text", + "content": "[7] Huggingface Diffusers. Sdxl-inpaint. https : / / huggingface . co / diffusers / stable - diffusion-x1- 1 . 0 - inpainting- 0 . 1, Year. Accessed: YYYYY-MM-DD. 1,3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 415, + 295, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 456 + ], + "type": "text", + "content": "[8] Stefan Elfwing, Eiji Uchibe, and Kenji Doya. Sigmoid-weighted linear units for neural network function approximation in reinforcement learning. Neural networks, 107:3-11, 2018. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 458, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 295, + 491 + ], + "type": "text", + "content": "[9] Omar Elharrouss, Noor Almaadeed, Somaya Al-Maadeed, and Younes Akbari. Image inpainting: A review. Neural Processing Letters, 51:2007-2028, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 492, + 295, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 492, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 57, + 492, + 295, + 534 + ], + "type": "text", + "content": "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems (NeurIPS), pages 6840-6851, 2020. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 536, + 295, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 295, + 581 + ], + "type": "text", + "content": "[11] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. arXiv preprint arXiv:2403.06976, 2024. 1, 3, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 582, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 582, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 582, + 295, + 656 + ], + "type": "text", + "content": "[12] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision, 128(7):1956-1981, 2020. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 295, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 691 + ], + "type": "text", + "content": "[13] Shanchuan Lin, Anran Wang, and Xiao Yang. Sdxlightning: Progressive adversarial diffusion distillation. arXiv preprint arXiv:2402.13929, 2024. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[14] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 118 + ], + "type": "text", + "content": "Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "type": "text", + "content": "[15] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "type": "text", + "content": "[16] Hayk Manukyan, Andranik Sargsyan, Barsegh Atanyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Hd-painter: high-resolution and prompt-faithful text-guided image inpainting with diffusion models. arXiv preprint arXiv:2312.14091, 2023. 1, 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 220, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 220, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 317, + 220, + 553, + 275 + ], + "type": "text", + "content": "[17] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 276, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 553, + 341 + ], + "type": "text", + "content": "[18] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 342, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 342, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 317, + 342, + 553, + 397 + ], + "type": "text", + "content": "[19] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 399, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 431 + ], + "type": "text", + "content": "[20] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 433, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 433, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 317, + 433, + 553, + 466 + ], + "type": "text", + "content": "[21] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 468, + 553, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 511 + ], + "type": "text", + "content": "[22] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 513, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 513, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 513, + 553, + 567 + ], + "type": "text", + "content": "[23] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching Imms for visual scoring via discrete text-defined levels. arXiv preprint arXiv:2312.17090, 2023. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 568, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 553, + 602 + ], + "type": "text", + "content": "[24] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 602, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 602, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 602, + 553, + 667 + ], + "type": "text", + "content": "[25] Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, and Lu Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4818–4829, 2024. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "text", + "content": "[26] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. arXiv preprint arXiv:2311.18828, 2023. 2, 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7621" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 407 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[27] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. arXiv preprint arXiv:2405.14867, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "type": "text", + "content": "[28] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contextual attention. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5505-5514, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 217 + ], + "type": "text", + "content": "[29] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4471-4480, 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 272 + ], + "type": "text", + "content": "[30] Yanhong Zeng, Jianlong Fu, Hongyang Chao, and Baining Guo. Learning pyramid-context encoder network for high-quality image inpainting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1486-1494, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 294, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 318 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 318 + ], + "type": "text", + "content": "[31] Yanhong Zeng, Jianlong Fu, Hongyang Chao, and Baining Guo. Aggregated contextual transformations for high-resolution image inpainting. IEEE Transactions on Visualization and Computer Graphics, 29(7):3266-3280, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 363 + ], + "type": "text", + "content": "[32] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 364, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 364, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 294, + 407 + ], + "type": "text", + "content": "[33] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 3" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7622" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Twinner_ Shining Light on Digital Twins in a Few Snaps/0293a039-5500-41cc-90fa-fe4f8710c6e9_content_list.json b/2025/Twinner_ Shining Light on Digital Twins in a Few Snaps/0293a039-5500-41cc-90fa-fe4f8710c6e9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1e70a5df57996281e2dd1bcc2309a0788771826e --- /dev/null +++ b/2025/Twinner_ Shining Light on Digital Twins in a Few Snaps/0293a039-5500-41cc-90fa-fe4f8710c6e9_content_list.json @@ -0,0 +1,1423 @@ +[ + { + "type": "text", + "text": "Twinner: Shining Light on Digital Twins in a Few Caps", + "text_level": 1, + "bbox": [ + 215, + 130, + 782, + 152 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/25e121c60bbfe31469346c6950294ce48c3797587259d68d4fdf63fefa30c6e8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 178, + 743, + 260 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ac242e6501d8cbc0fe71db3f48326501294e99d57de24953fc5b1db06ad90e9a.jpg", + "image_caption": [ + "Figure 1. We present Twinner, the first large reconstruction model capable of predicting a scene's illumination as well as an object's geometry and material properties from a few posed images. This enables tasks such as realistic relighting of objects in novel scenes in a few seconds by replacing costly per-scene optimizations with a single forward pass." + ], + "image_footnote": [], + "bbox": [ + 91, + 290, + 906, + 513 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 589, + 326, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present Twinner, the first large reconstruction model capable of recovering a scene's illumination as well as an object's geometry and material properties from only a few posed images. Twinner is based on the Large Reconstruction Model and innovates in three key ways: 1) We introduce a memory-efficient voxel-grid transformer whose memory scales only quadratically with the size of the voxel grid. 2) To address the scarcity of high-quality ground-truth PBR-shaded models, we introduce a large, fully synthetic dataset of procedurally generated PBR-textured objects lit with varied illumination. 3) To bridge the synthetic-to-real gap, we finetune the model on real-world datasets using a differentiable physically based shading model, eliminating the need for ground-truth illumination or material properties, which are challenging to obtain in real-world scenarios. We demonstrate the efficacy of our model on the real-world StanfordORB benchmark, where, given a few input views, we achieve reconstruction quality significantly superior to", + "bbox": [ + 88, + 628, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "existing feed-forward reconstruction networks and comparable to slower per-scene optimization methods.", + "bbox": [ + 511, + 590, + 906, + 621 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 648, + 645, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The creation of digital twins, i.e., the conversion of real-world objects into virtual 3D assets, is an important problem in virtual and augmented reality, gaming, conservation, education, advertisement, and general 3D content creation. In recent years, new techniques such as NeRF [31] and 3D Gaussian Splatting [24], along with their feed-forward counterparts [18, 49, 60, 63], have significantly improved the quality, speed, and generality of digital twinning. However, most of these techniques still bake illumination into the scene's appearance as radiance. As a consequence, the resulting assets cannot be relit and appear out of place when visualized under different lighting conditions.", + "bbox": [ + 509, + 672, + 906, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this limitation, one must also reconstruct the field of Bidirectional Reflectance Distribution Functions (BRDF), which describe the optical properties of materi-", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5859", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "als. Estimating the BRDF directly is notoriously challenging due to its high dimensionality [3] and is usually done using 3D scanners that capture thousands of images while controlling for viewpoint and illumination [4-7, 35, 42, 65]. This is often impractical, so authors have considered \"casual\" reconstruction setups, where the object is captured by a conventional camera in uncontrolled conditions. Examples include NVDiffRec [34], NeRD [8], and NeRFactor [59], which simultaneously estimate the illumination, in the form of an environment map, and a simplified BRDF parameterized using Physically-Based Rendering (PBR) with albedo, roughness, and metallicity [10]. These quantities are recovered by minimizing the photometric reconstruction error, comparing the output of a differentiable PBR renderer to the input images. However, this optimization is slow, ill-posed, and requires the introduction of handcrafted priors for regularization.", + "bbox": [ + 89, + 90, + 480, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a new method, Twinner, a feedforward network that outputs shape, PBR textures, and environment illumination. This network is fast, performs well, requires only a small number of images as input, and learns empirically the necessary reconstruction priors.", + "bbox": [ + 89, + 349, + 480, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Siddiqui et al. [45] recently introduced a similar predictor, MetalLRM, as an extension of the LightplaneLRM [11]. MetalLRM requires supervision with a large dataset of 3D assets with high-quality PBR textures, which, unfortunately, are not abundant in existing synthetic datasets like Objaverse [14]. Moreover, even if a large synthetic dataset were available, the synthetic-to-real gap between the artist-created training assets and the target domain of real input images makes MetalLRM unsuitable for digital twinning.", + "bbox": [ + 89, + 428, + 480, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With our new Twinner, we address the scarcity of high-quality training PBR assets and the train/test domain shift. First, we ask whether real images in datasets like CO3Dv2 [40] can be used to supervise the model. Although these datasets lack PBR information, they are abundant and match the target distribution of real images we aim to digitally twin. To use this data, we propose minimizing the photometric error similar to optimization-based models, but with the key difference that the 3D shape, PBR materials, and environment light are all predicted by a large feedforward transformer [19, 45]. During training, each scene is rendered with a differentiable PBR shader that accepts the predicted environment light and the predicted PBR-textured shape. The latter then admits indirect PBR and illumination supervision via photometric losses between the renders and the corresponding real ground-truth images.", + "bbox": [ + 89, + 566, + 482, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, we consider a second approach to tackle data scarcity where, inspired by Zeroverse [58], we procedurally generate PBR-textured objects lit with a known set of environment maps. While the resulting objects are not realistic, their key advantage is that they come with clean ground-truth PBR and illumination information, which can", + "bbox": [ + 89, + 810, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "be used to supervise the corresponding predictors directly instead of indirectly via the photometric loss.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Besides addressing data scarcity, we also improve the architecture of the reconstruction model. We believe in the virtues of a simple and direct voxel-grid representation of the object [47], but note that predicting a number of tokens cubic in the voxel grid resolution is intractable with a transformer. Reducing the number of predicted tokens is the primary reason why recent models [19, 45] factor the voxel grid as a triplane. The question, then, is how to retain the benefits of the direct voxel grid representation and still make computations feasible. We solve this problem by introducing the novel tricolumn representation, where each feature vector (column) in a triplane is obtained by stacking the corresponding cells in the underlying voxel grid. In this manner, the number of tokens becomes quadratic in the voxel grid size, while each voxel is still represented explicitly. We further augment the network to predict a representation of the environment map, capturing illumination.", + "bbox": [ + 511, + 123, + 903, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We demonstrate the efficacy of the Twinner model on the StanfordORB benchmark [26], consisting of casual captures of real-life objects. We achieve reconstruction quality superior to existing feed-forward predictors and comparable to optimization-based methods, while using only a very small number of input images of the object and operating much faster. We also obtain state-of-the-art reconstruction of the environment maps, achieving a $28\\%$ improvement in the predicted illumination's angular error over existing methods.", + "bbox": [ + 511, + 380, + 903, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 549, + 651, + 564 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D reconstruction. 3D reconstruction aims to recover a three-dimensional representation of a scene from a set of posed images. Traditionally, this was accomplished with structure-from-motion [1, 43, 53], multi-view stereo [16, 44], and mesh surface extraction methods [23, 29]. Recently, methods based on neural fields [2, 31-33, 51, 54, 62] have focused on extracting both shape and appearance by learning the scene radiance function. Through volumetric rendering and photometric losses, they are able to implicitly reconstruct a 3D density volume from which a mesh can be extracted. However, these methods infer the parameters of each scene via test-time optimization, making them slow. Authors have thus introduced new feed-forward reconstruction networks, generically called Large Reconstruction Models (LRMs) [19, 56, 58, 67]. LRMs use transformers to predict 3D representations from a small set of input images. Our Twinner predicts not just geometry and radiance, but also material properties and illumination.", + "bbox": [ + 511, + 575, + 903, + 847 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Material reconstruction. 3D reconstruction typically recovers a scene's geometry and radiance under fixed illumination. However, it is also necessary to recover material", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5860", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8fa14c1ea85692e07b84abf9d144bb320c70789d9224a53c44bdb0d95f85f84f.jpg", + "image_caption": [ + "Figure 2. Overview of Twinner. The input views are processed together with their foreground masks and camera poses by an image tokenizer. The resulting tokens are processed by a Diffusion Transformer (DiT) model to predict a 3D volumetric representation of the scene, and a second DiT model to predict the scene's illumination. The 3D representation is then rendered to obtain images of the scene's material properties, normals, and opacity from the target's point of view. Using these rendered images together with the predicted illumination, Twinner renders an approximate physically-shaded image of the scene which then enters a photometric loss." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 908, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "properties so that the 3D assets look realistic in novel virtual scenes. To do so, some works focus on optimizing explicit representations, such as material textures on meshes [17, 34, 48], while other works have added physically-based rendering models into neural fields [8, 9, 21, 27, 30, 64] and Gaussian splats [20, 28]. While these methods achieve high-quality material reconstructions, they require costly per-scene optimization. Some have addressed this issue by learning feed-forward reconstructors of PBR assets [45]. However, these are typically trained with synthetic data due to the complexity of capturing real-world material properties. We bridge the gap between synthetic and real data by training Twinner on real-world datasets.", + "bbox": [ + 88, + 369, + 483, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Illumination prediction. A key to accurate material reconstruction is accurate modeling of the scene's illumination. Single-scene optimization methods achieve this by optimizing an internal representation of illumination using light probe images [17, 34, 59], spherical harmonics [28], spherical Gaussians [21, 66], or MLPs [27, 30, 64]. These methods rely on physical models of lighting to aid the optimization. Others have attempted to predict the illumination of a scene from a single image [38, 52]. They rely on generative models to map the input view to a possible distribution of panorama views representing the environment illumination. Our Twinner lies in between these two lines of work: it is similar to the predictive approach as it predicts illumination from a set of four input views, but it uses physically-based rendering losses for supervision, similar to single-scene optimization methods.", + "bbox": [ + 89, + 573, + 483, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 829, + 225, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce Twinner, a model that reconstructs the shape, materials, and illumination of real-life 3D objects from a small set of posed input views. Twinner extends Light-", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "planeLRM [11], an enhanced version of LRM [19], by incorporating predictors for the object's material properties and scene illumination, as well as a differentiable physically-based renderer. In Sec. 3.1, we begin by reviewing the reconstructor and the PBR model we use. Then, in Secs. 3.2 to 3.4, we describe the main building blocks of Twinner. Finally, in Sec. 3.5, we detail the training protocol, including the losses used to supervise Twinner on real data.", + "bbox": [ + 511, + 369, + 906, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 516, + 653, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Physically-based rendering. To supervise our model with a photometric loss, it is necessary to render the reconstructed scene into an image. The color of each pixel corresponds to the radiance $\\mathbf{L}_o$ emitted by the imaged object point $p\\in \\mathbb{R}^3$ in the direction of the camera. With knowledge of a scene's illumination, geometry, and material properties, it is possible to estimate the outgoing radiance at any such point according to the reflectance equation:", + "bbox": [ + 511, + 537, + 908, + 659 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {L} _ {o} = \\int_ {\\Omega} \\left(k _ {d} \\frac {\\mathbf {a}}{\\pi} + \\mathbf {f} _ {s}\\right) \\mathbf {L} _ {i} \\left\\langle \\boldsymbol {\\omega} _ {i}, \\mathbf {n} \\right\\rangle d \\boldsymbol {\\omega} _ {i}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 669, + 906, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $\\mathbf{L}_i$ is the incoming radiance, representing the scene illumination falling on the imaged point $\\pmb{p}$ . The integral is carried out over a hemisphere $\\Omega$ of incoming directions $\\omega_{i}$ determined by the surface normal $\\mathbf{n}$ at $\\pmb{p}$ . The factor in parentheses is a model of the object's Bidirectional Reflectance Distribution Function (BRDF), which describes how the incoming radiance reflects off the object's surface into various directions. The first term is the diffuse component, parameterized by the object's diffuse albedo $\\mathbf{a}$ and a constant $k_{d}$ , both of which are independent of direction. In the integral, this term averages all incoming radiance over the hemisphere. The second term, $\\mathbf{f}_s$ , is the specular com", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5861", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/194392d4b0cbe5ee68de5ccc42997a79092f846e1d944a55ed3c5925e4abd483.jpg", + "image_caption": [ + "Figure 3. Tricolumn Architecture. The transformer module processes image tokens and a learnable embedding to predict the tricolumn representation, which consists of a grid of $3R^2$ feature vectors with dimensionality $\\mathbb{R}^{CR}$ . These feature vectors are then split into three axis-aligned column grids, each with dimensionality $\\mathbb{R}^{CR}$ . Finally, all axis-aligned column grids are reshaped into voxel grids of size $R\\times R\\times R$ with dimensionality $\\mathbb{R}^C$ and summed into a single voxel grid." + ], + "image_footnote": [], + "bbox": [ + 106, + 95, + 908, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ponent of the BRDF, which captures directional effects such as reflections or highlights. Note that $\\mathbf{L}_i$ is a function of the incoming direction $\\omega_{i}$ , and the specular BRDF $\\mathbf{f}_s$ is a function of both the incoming and outgoing directions $\\omega_{i}$ and $\\omega_{o}$ . Radiances and BRDFs also depend on the light frequency, which we model using three RGB components. Furthermore, all quantities depend on the point position $\\pmb{p}$ . We omit these details for brevity.", + "bbox": [ + 88, + 329, + 480, + 449 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Due to the complexity of the integral in Eq. (1), numerous approximations have been proposed. We rely on the split-sum approximation [22] for image-based lighting, which leverages the scene's illumination represented as a spherical image enveloping the scene at an infinite radius. Following [34], we use a set of differentiable mipmaps to represent the pre-integrated lighting. The BRDF follows the Disney [22] microfacet BRDF model and is parameterized by the albedo $\\mathbf{a}$ , metalness $m$ , and roughness $r$ at each point. We further assume that $\\mathbf{L}_i$ is constant in space (hence we do not account for shadows cast by the object).", + "bbox": [ + 88, + 452, + 482, + 618 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In summary, to apply this model to render an image of the object, we require an estimate of the object's 3D shape and its albedo, metalness, and roughness at its surfaces. We also require a spherical image representing the directional function $\\mathbf{L}_i$ , known as an environment map.", + "bbox": [ + 89, + 619, + 483, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Large Reconstruction Model (LRM). The aim of LRM [19] is to map a few posed views $\\mathcal{I}$ of the scene to a 3D reconstruction $V$ . LRM first extracts features $y$ from the set of input views, utilizing a pre-trained DINO [36] Vision Transformer (ViT) [15]. Following [61], we feed into the ViT the RGB images $\\mathcal{I}$ , the Plücker embeddings [46] $\\mathcal{I}_{\\mathrm{Pl}}$ representing the cameras, and the masks $\\mathcal{M}$ of the foreground object to be reconstructed.", + "bbox": [ + 89, + 702, + 483, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The extracted feature tokens enter, via cross-attention, the transformer architecture from Diffusion Transformer (DiT) [37] $\\Phi$ , which is tasked with reconstructing the scene. We emphasize that we borrow the DiT architecture without any diffusion process. The tokens in this transformer en", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "code the 3D representation of the scene, realizing a mapping $V = \\Phi(V_0|y)$ , where $V_0$ are the learnable initial tokens. A crucial design choice is how to construct the tokens $V$ manipulated by this transformer, which is discussed in Sec. 3.2.", + "bbox": [ + 511, + 329, + 905, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, using the memory-efficient rendering algorithm proposed in Lightplane [11], a target set of views is rendered from the 3D representation. A shaded RGB image $\\hat{\\mathcal{I}}$ , a foreground mask image $\\hat{\\mathcal{M}}$ , and a depth image $\\hat{\\mathcal{I}}_{\\mathrm{d}}$ are rendered for each view and used to supervise the LRM through a combination of the following photometric losses:", + "bbox": [ + 511, + 405, + 905, + 496 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {M} = \\operatorname {B C E} (\\mathcal {M}, \\hat {\\mathcal {M}}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 505, + 903, + 520 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {R} = \\operatorname {L P I P S} \\left(\\mathcal {M} \\mathcal {I}, \\mathcal {M} \\hat {\\mathcal {I}}\\right) + \\left\\| \\mathcal {M} \\mathcal {I} - \\mathcal {M} \\hat {\\mathcal {I}} \\right\\| ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 525, + 903, + 542 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d} = \\left\\| \\mathcal {M I} _ {\\mathrm {d}} - \\mathcal {M} \\hat {\\mathcal {I}} _ {\\mathrm {d}} \\right\\|, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 547, + 903, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{I}_{\\mathrm{d}}$ is the set of ground-truth depth maps corresponding to the input images $\\mathcal{I}$ , and $\\mathcal{M}\\mathcal{I}$ denotes a masked image.", + "bbox": [ + 511, + 573, + 905, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Tricolumn representation", + "text_level": 1, + "bbox": [ + 511, + 612, + 746, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Crucial to the Twinner design is the representation of the 3D scene, which is encoded as a set of volumetric functions expressing opacity and material properties at every point $\\pmb{p} \\in \\mathbb{R}^3$ in space. A simple and effective option [47] to encode a volumetric map $V: \\mathbb{R}^3 \\to \\mathbb{R}^C$ is to use a voxel grid $V \\in \\mathbb{R}^{C \\times D \\times H \\times W}$ . Here, the value $V(\\pmb{p}) \\in \\mathbb{R}^C$ interpolates the voxel values at the nearest grid points. To predict the voxel grid $V$ with the transformer of Sec. 3.1, we could assign one token to each voxel. However, the number $WHD$ of voxels grows cubically with the resolution of the grid, which leads to prohibitive memory demands.", + "bbox": [ + 511, + 633, + 905, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The alternative adopted by models like [19, 45] is to use a triplane representation [12]. In this case, the volumetric function is represented by three 2D grids $V_{xy} \\in \\mathbb{R}^{K \\times H \\times W}$ , $V_{yz} \\in \\mathbb{R}^{K \\times W \\times D}$ , and $V_{zx} \\in \\mathbb{R}^{K \\times D \\times H}$ . The volumetric function at $\\pmb{p} = (x,y,z)$ is then expressed as", + "bbox": [ + 511, + 800, + 905, + 876 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV (x, y, z) = f \\left(V _ {x y} (x, y), V _ {y z} (y, z), V _ {z x} (z, x)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 883, + 903, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5862", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/711809666565476ddb2350ffe020e6716d6988f6aa55345e373cc0d9d6505e46.jpg", + "image_caption": [ + "Figure 4. Examples from our procedural dataset visualizing the shaded image, environment map, materials, and normals from our environment-map-endowed version of Zeroverse [57]." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 486, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f: \\mathbb{R}^{K \\times K \\times K} \\to \\mathbb{R}^C$ is a point-wise mixing function (e.g., an MLP applied to the concatenation of the three $K$ -dimensional feature vectors). Paired with a transformer, this conveniently reduces the number of tokens to $WH + HD + DW$ only. However, triplane mixes information between different voxels, which can lead to artifacts.", + "bbox": [ + 89, + 330, + 483, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We thus propose a novel tricolumn representation instead. This is similar to triplane as it uses a plane instead of a volume of tokens; however, each feature vector is in itself a column of corresponding voxel cells, so that voxel-specific information is represented directly and separately.", + "bbox": [ + 89, + 422, + 483, + 497 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In detail, we consider a plane of feature vectors $V_{xy} \\in \\mathbb{R}^{(CR) \\times R \\times R}$ . Because this is a plane, there are only $R^2$ tokens, which are easily handled by the transformer. However, $V_{xy}$ can be reshaped into a $C \\times R \\times R \\times R$ voxel grid by unflattening the first dimension. We use the symbol $V_{xy}(x,y;z) \\in \\mathbb{R}^C$ to denote the feature vector obtained by interpolating this grid in 3D.", + "bbox": [ + 89, + 498, + 483, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When three such planes are used, the full feature vector at $\\pmb{p}$ is decoded as:", + "bbox": [ + 89, + 604, + 483, + 635 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nV (x, y, z) = f \\left(V _ {x y} (x, y; z), V _ {y z} (y, z; x), V _ {z x} (z, x; y)\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 642, + 483, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that this construction is strictly different from the triplane equation (5) as each term depends on all three spatial coordinates. We opt for using summation followed by an MLP as the mixing function $f$ .", + "bbox": [ + 89, + 674, + 483, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. PBR reconstruction model", + "text_level": 1, + "bbox": [ + 89, + 742, + 331, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The LRM models of [19, 61] only recover the shape and radiance of the scene. Like in [45], we extend Twinner to also recover the material properties. We do so by extending the tricolumn representation to encode the albedo $a$ (instead of radiance) and two additional channels representing the scene's metalness $m$ and roughness $r$ . We also note the importance of accurate normals for relighting. While a normal image could be directly estimated by projecting the rendered depth image into 3D space and taking the gradient", + "bbox": [ + 89, + 763, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of the opacity field, $\\bar{\\mathcal{I}}_{\\mathrm{n}} = \\nabla \\sigma (\\pi^{-1}(\\hat{\\mathcal{I}}_{\\mathrm{d}}))$ , this estimate tends to be noisy. Therefore, we follow previous works [51] and predict three additional channels representing the scene's normals. The predicted normal image $\\hat{\\mathcal{I}}_{\\mathrm{n}}$ is then supervised with the pseudo-ground-truth normals from the opacity field $\\bar{\\mathcal{I}}_{\\mathrm{n}}$ via:", + "bbox": [ + 511, + 90, + 905, + 180 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {N} = 1 - \\langle \\mathcal {M} \\bar {\\mathcal {I}} _ {\\mathrm {n}}, \\mathcal {M} \\hat {\\mathcal {I}} _ {\\mathrm {n}} \\rangle . \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 180, + 905, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Direct supervision using synthetic data. Reconstruction models that predict radiance can be supervised by rendering the predicted 3D scene into images and comparing these to training views using a photometric loss. In our case, however, the model does not predict radiance, but albedo, roughness, and metalness, which cannot be rendered into a 'shaded' image unless illumination is also known.", + "bbox": [ + 511, + 200, + 905, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A simple approach for supervising the model that does not require knowledge of illumination is to render albedo, roughness, and metalness images $\\hat{\\mathcal{I}}_{\\mathrm{a}},\\hat{\\mathcal{I}}_{\\mathrm{r}}$ ,and $\\hat{\\mathcal{I}}_{\\mathrm{m}}$ , and compare these to their corresponding ground-truth renders. We use this approach to supervise our model with synthetic data by applying the following loss on the set of material renders $\\hat{\\mathcal{I}}_{\\mathrm{a}},\\hat{\\mathcal{I}}_{\\mathrm{m}},\\hat{\\mathcal{I}}_{\\mathrm{r}}$ , and also on the normal render $\\hat{\\mathcal{I}}_{\\mathrm{n}}$ ..", + "bbox": [ + 511, + 306, + 905, + 412 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S} = \\ell_ {\\mathcal {M}} \\left(\\mathcal {I} _ {\\mathrm {a}}, \\hat {\\mathcal {I}} _ {\\mathrm {a}}\\right) + \\ell_ {\\mathcal {M}} \\left(\\mathcal {I} _ {\\mathrm {r}}, \\hat {\\mathcal {I}} _ {\\mathrm {r}}\\right) + \\ell_ {\\mathcal {M}} \\left(\\mathcal {I} _ {\\mathrm {m}}, \\hat {\\mathcal {I}} _ {\\mathrm {m}}\\right) + \\ell_ {\\mathcal {M}} \\left(\\mathcal {I} _ {\\mathrm {n}}, \\hat {\\mathcal {I}} _ {\\mathrm {n}}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 417, + 903, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\ell_{\\mathcal{M}}(u,v) = \\| \\mathcal{M}(u - v)\\|_{\\epsilon}$ is the Huber loss $\\| \\cdot \\|_{\\epsilon}$ with cut-off threshold $\\epsilon$ masked by $\\mathcal{M}$ .", + "bbox": [ + 511, + 449, + 903, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training on real data by predicting illumination", + "text_level": 1, + "bbox": [ + 511, + 487, + 905, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The learning formulation of Sec. 3.3 requires ground-truth material properties to be known, which is not the case for real-world data where only shaded images are available. In order to fine-tune our model using real data with a photometric loss, it is necessary to also estimate the scene illumination. With this, we can use the differentiable physically-based shader detailed in Sec. 3.1 to render the shaded scene appearance and compare it to the input images through a photometric loss.", + "bbox": [ + 511, + 508, + 905, + 645 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To estimate the scene's illumination, we propose using an additional DiT module. The latter takes as input the feature tokens $y$ computed by the image feature extractor and predicts a cubemap representation of the scene's lighting. Recall that a cubemap is a function $C \\to \\mathbb{R}^3$ that assigns every point $\\pmb{p}$ of the cube surface $C = \\{\\pmb{p} : \\| \\pmb{p} \\|_{\\infty} = 1\\}$ to the radiance $\\mathbf{L}_i(\\boldsymbol{\\omega}_i)$ incoming from direction $\\boldsymbol{\\omega}_i \\propto -\\pmb{p}$ , and is thus a representation of the environment illumination (this works because we assume that the function $\\mathbf{L}_i$ only depends on direction, but not on position). The cubemap is unfolded to a collection of six 2D images, which are directly predicted by the DiT module.", + "bbox": [ + 511, + 646, + 906, + 827 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Supervising illumination with synthetic data. In synthetic data, the cubemap is known, and its predictor can be supervised directly with the environment-map loss:", + "bbox": [ + 511, + 832, + 905, + 878 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {e n v}} = \\left\\| \\mathcal {I} _ {\\mathrm {e}} - \\hat {\\mathcal {I}} _ {\\mathrm {e}} \\right\\| ^ {2} + \\operatorname {L P I P S} \\left(\\mathcal {I} _ {\\mathrm {e}}, \\hat {\\mathcal {I}} _ {\\mathrm {e}}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 883, + 905, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5863", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5db87447dcdd41a1a4729a3c60c8eb49649c847299d61f8715e8a596f47bfac8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Novel Scene RelightingNovel View SynthesisGeometry EstimationRuntime
MethodPSNR-H↑PSNR-L↑SSIM↑LPIPS↓PSNR-H↑PSNR-L↑SSIM↑LPIPS↓Depth↓Geometry↓Shape↓Time (hr.)↓
Single-Scene Opt.PhySG [66]21.8128.110.9600.05524.2432.150.9740.0471.900.179.28~3
NVDiffRec [34]22.9129.720.9630.03921.9428.440.9690.0300.310.060.62~1
NeRD [8]23.2929.650.9570.05925.8332.610.9630.0541.390.2813.70>20
NeRFactor [59]23.5430.380.9690.04826.0633.470.9730.0460.870.299.53>20
InvRender [69]23.7630.830.9700.04625.9134.010.9770.0420.590.060.44~3
NVDiffRecMC [17]24.4331.600.9720.03628.0336.400.9820.0280.320.040.51~2
Neural-PBIR [48]26.0133.260.9790.02328.8236.800.9860.0190.300.060.43~1
PredictiveMetaLRM [45]21.5028.060.9570.05219.5525.980.9560.0518.6690.061.260.004
w/o ZeroVerse23.5330.310.9650.04821.5328.780.9650.0481.1260.090.740.006
w/o Shading Loss23.8930.850.9670.04221.4229.470.9670.0410.7620.060.300.004
Twinner (ours)24.5531.610.9690.03821.7029.620.9680.0370.7090.060.370.006
", + "bbox": [ + 98, + 88, + 903, + 250 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Relighting evaluation on StanfordORB [26] comparing our method to other feedforward predictors of PBR-textured assets (two ablations and MetalLRM). For completeness, we also include single-scene optimization methods in the top part of the table. Our model outperforms most single-scene optimization baseline methods in terms of novel scene relighting and shape estimation despite being a predictive model and having access to only four input views. More impressively, it only requires a few seconds to reconstruct an object's mesh with PBR materials, while single-scene optimization methods have runtimes in the order of hours.", + "bbox": [ + 89, + 261, + 906, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "comprising the photometric loss from (8) using the predicted cubemap projected into an equirectangular image $\\hat{\\mathcal{I}}_{\\mathrm{e}}$ and the ground-truth equirectangular image $\\mathcal{I}_{\\mathrm{e}}$ . In practice, we use a procedurally generated dataset of synthetic assets detailed in Sec. 3.5 to provide ground-truth illumination.", + "bbox": [ + 88, + 344, + 480, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Supervising illumination with real data. In real data, the cubemap is observed indirectly via the effect it has on the object's appearance according to the reflectance equation (1), which provides only weak supervision. We note, however, that the scene background visible in the input images is, to a first approximation, equal to the function $\\mathbf{L}_i$ , and can thus provide direct supervision of the latter. We can render the background observed from any given camera by assigning to each pixel the color $\\mathbf{L}_i(\\boldsymbol{\\omega}_i) = C(\\boldsymbol{\\omega}_i / \\| \\boldsymbol{\\omega}_i \\|_{\\infty})$ of the incoming radiance from the direction $\\boldsymbol{\\omega}_i$ of the pixel's ray. Since input images are in sRGB color space, the predicted background image is first tone-mapped and then composited with the foreground object prediction to obtain a scene image. The latter allows us to supervise both the illumination prediction and the shaded object prediction through the photometric loss $\\mathcal{L}_R$ of Eq. (3). Here, as training data, we use abundant object-centric video captures from CO3Dv2 [41] as described in the next section.", + "bbox": [ + 91, + 428, + 483, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Training protocol", + "text_level": 1, + "bbox": [ + 89, + 712, + 264, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Twinner's training minimizes the total loss", + "bbox": [ + 89, + 734, + 375, + 750 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t}} = \\mathcal {L} _ {N} + \\mathcal {L} _ {S} + \\mathcal {L} _ {\\mathrm {T V}} + \\mathcal {L} _ {d} + \\mathcal {L} _ {M} + \\mathcal {L} _ {R} + \\mathcal {L} _ {\\mathrm {T V}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 763, + 482, + 781 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{TV}} = \\mathrm{TV}(\\hat{\\mathcal{I}}_{\\mathrm{n}}) + \\mathrm{TV}(\\hat{\\mathcal{I}}_{\\mathrm{a}}) + \\mathrm{TV}(\\hat{\\mathcal{I}}_{\\mathrm{d}})$ is the Total-Variation (TV) regularization on normal, albedo, and depth images promoting smooth solutions. $\\mathcal{L}_{\\mathrm{tot}}$ is optimized with Adam [25] using a learning rate of $10^{-4}$ until convergence. As mentioned before, individual terms of $\\mathcal{L}_{\\mathrm{tot}}$ are switched on or off depending on whether a real or a synthetic training example is present in the batch of training images.", + "bbox": [ + 89, + 792, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training data. Following [45], Twinner is trained on a large dataset of 140k artist-created assets similar to Objaverse. However, this dataset often lacks high-quality material maps where, for instance, distinguishing between missing material images and material images with zero values is impossible. Moreover, there is a large domain gap between the synthetic objects and the real ones we intend to twin.", + "bbox": [ + 511, + 344, + 906, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As such, together with the artist-created assets, we also train on two additional data sources: 1) A large dataset of procedurally generated assets with clean PBR textures and environment maps; 2) Real-object captures from CO3Dv2 [40]. Both datasets are detailed next.", + "bbox": [ + 511, + 455, + 906, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Endowing Zeroverse with environment maps. Following LRM-Zero [57], we generate a set of 10k randomized 3D objects with realistic material properties sampled from the MatSynth [50] dataset. Additionally, to enable the environment-map supervision loss $\\mathcal{L}_{\\mathrm{env}}$ (9), for each rendered scene, we randomly sample one of $\\sim 1\\mathrm{k}$ environment maps from PolyHaven [39]. We render color, normal, material, depth, and foreground mask images for each object from a predetermined set of viewpoints (see examples in Fig. 4). As such, samples from this dataset admit optimizing all terms of the total loss $\\mathcal{L}_{\\mathrm{tot}}$ , including the environment loss $\\mathcal{L}_{\\mathrm{env}}$ , which is crucial for stable training of Twinner.", + "bbox": [ + 511, + 541, + 908, + 723 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Exploiting real object captures. We also train on CO3Dv2 [40], which is a large-scale real-world dataset of \"turntable\" videos capturing objects of common categories. Each frame in the dataset comes with camera and foreground mask annotations. As such, for each CO3Dv2 sample, we can optimize the normal, mask, photometric, and TV losses $\\mathcal{L}_{\\mathrm{tot}}^{\\mathrm{real}} = \\mathcal{L}_N + \\mathcal{L}_M + \\mathcal{L}_R + \\mathcal{L}_{\\mathrm{TV}}$ . Since $\\mathcal{L}_{\\mathrm{tot}}^{\\mathrm{real}}$ is a function of the illumination image $\\hat{\\mathcal{I}}_{\\mathrm{e}}$ and PBR renders $\\hat{\\mathcal{I}}_{\\mathrm{a}}$ , $\\hat{\\mathcal{I}}_{\\mathrm{m}}$ , $\\hat{\\mathcal{I}}_{\\mathrm{r}}$ , the loss provides a valuable supervisory signal for all volumetric predictors. Furthermore, using real training data narrows the domain gap between the training set and", + "bbox": [ + 511, + 733, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5864", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the target domain of Twinner, which are real images.", + "bbox": [ + 89, + 90, + 439, + 106 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ec178c20f5c817dfc35e49cecd77609ad75198bea25d62d559faa2112599a80a.jpg", + "image_caption": [ + "Figure 5. StanfordORB Illumination Visualization. We visualize the illumination estimation of multiple baselines and our method on four different scenes from the StanfordORB dataset." + ], + "image_footnote": [], + "bbox": [ + 91, + 117, + 483, + 265 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/622a218f0827ce7715b4e1df472f609bf427db347ed1a369dea9958049997275.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Illumination prediction results. We provide a quantitative evaluation of the illumination estimation for baselines and our Twinner. Rows are grouped by single-scene optimization models (top) and predictive feed-forward models (bottom)." + ], + "table_body": "
MethodAngular Error ↓RMSE↓Norm. RMSE↓SI RMSE↓Time (hr.)↓
NMF [30]11.240.4870.5240.394~ 2
TensoIR [21]15.080.7520.7980.451~ 2
NVDiffRec [34]12.260.7230.7150.367~ 1
NVDiffRecMC [17]13.250.5740.6140.527~ 2
StyleLight [52]13.050.6400.6890.4400.011
DiffusionLight [38]11.350.5810.5690.5440.317
Twinner (ours)8.090.5700.6020.4720.006
", + "bbox": [ + 91, + 330, + 480, + 419 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 500, + 223, + 517 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation data. We evaluate Twinner on the StanfordORB [26] dataset, which comprises 42 captures of 14 different objects, each imaged under 3 different illuminations.", + "bbox": [ + 89, + 525, + 482, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Each object has a high-quality geometric reconstruction obtained using a 3D scanner and comes with ground-truth camera poses and estimated illumination.", + "bbox": [ + 89, + 571, + 483, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Protocol. For each evaluation scene, we select four views from the set of training images as input frames via furthest camera-origin point sampling. With these input views, Twinner reconstructs a 3D PBR-textured mesh. Each predicted mesh is rendered with Blender [13] to held-out test views while being lit with a given ground-truth environment map. Following [26], the renders are compared to the corresponding ground-truth images using scale-invariant versions of LPIPS [68], SSIM [55], and PSNR.", + "bbox": [ + 89, + 622, + 482, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We compare to existing feed-forward predictors of PBR-textured meshes. As such, we selected the state-of-the-art MetalLRM few-view reconstructor from AssetGen [45] as our main baseline. We use the version of MetalLRM conditioned on four shaded input views, as opposed to the albedo+shaded conditioning used in the AssetGen pipeline. We also compare to two ablations of our method. The first ablation, Twinner w/o Shading Loss, evaluates the contribution of training on real data and omits", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/61b82d5a777513590a3905c372be15ff8005d05268b97e00ce561d11ae6aaf65.jpg", + "image_caption": [ + "Figure 6. Qualitative results on StanfordORB. We visualize the shaded RGB, geometry reconstruction, and material predictions for our baseline, two ablations, and our Twinner. Our model is able to better predict relightable objects thanks to the combination of curated synthetic training data and the shading loss on real data." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 908, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the training images of CO3Dv2. Second, we compare to Twinner w/o Zeroverse, which ablates training of the environment-map and PBR-texture predictor on our version of the procedurally-created Zeroverse. For completeness, we also include several single-scene optimization methods that \"overfit\" a relightable implicit shape in each scene from scratch. Note that, to save GPU memory, our feed-forward predictors randomly sample only four source views from the pool of all available training images of each scene. This puts us at a disadvantage compared to single-scene optimizers, as they leverage all available scene views for each reconstruction.", + "bbox": [ + 511, + 664, + 906, + 844 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Relighting Evaluation. All quantitative results are reported in Tab. 1. Twinner is the best among feed-forward reconstructors, outperforming its own two ablations and", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5865", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/256b653bc1c34a5fbf1ec2bcd5798f5e2d08f59a8e7b5a26ec5e491d98af3cf7.jpg", + "image_caption": [ + "Figure 7. CO3Dv2 Qualitative Results. We visualize predicted PBR meshes from the CO3Dv2 dataset illuminated under direct sunlight as well as their geometry, albedo, and material properties. The shading loss aids at predicting more realistic material properties, while the synthetic data improves the predicted geometry and albedo." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 911, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MetalLRM. For instance, Twinner achieves a $27\\%$ reduction in both novel-scene relighting and novel-view synthesis LPIPS compared to the strongest baseline (MetalLRM), while maintaining comparable inference time. Remarkably, our method is on par with most single-scene optimization methods while being significantly faster, requiring only a few seconds per prediction compared to several hours for optimization-based methods. Examples of our predictions and the baselines are shown in Fig. 6. We also evaluate qualitatively on testing sequences of CO3Dv2 in Fig. 7.", + "bbox": [ + 89, + 489, + 482, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Illumination Evaluation. We evaluate Twinner's ability to predict a scene's illumination from four input images. Since there are no existing baselines for few-view conditioned feed-forward illumination prediction, we compare to environment map estimates output by the single-scene optimizers in Tab. 1, as well as to single-view conditioned predictors. We follow [38] and evaluate illumination in terms of angular error, RMSE, normalized RMSE, and scale-invariant RMSE. Results are given in Tab. 2 and Fig. 5.", + "bbox": [ + 89, + 648, + 482, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tricolumn Ablation. We also evaluate the effect of our tricolumn representation. In Tab. 3, we report the novel-view synthesis performance of the vanilla LightplaneLRM [11] model trained only on our synthetic dataset, utilizing either tricolumn or triplane representations. Tricolumn outperforms triplane significantly, achieving a $22\\%$ improvement in LPIPS, validating this contribution.", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d818ce187e297cc3b41af667793e7ac4c7b5aadcf0616e0a975f0b9abfa18b82.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
3D representationPSNR↑SSIM↑LPIPS↓
Triplane31.200.9720.037
Tricolumn (ours)32.810.9660.029
", + "bbox": [ + 553, + 487, + 867, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Tricolumn ablation evaluating novel view synthesis on the StanfordORB dataset for a LightplaneLRM [11] using the triplane 3D representation or our proposed tricolumn instead. Our tricolumn outperforms triplane on all reconstruction metrics.", + "bbox": [ + 511, + 556, + 906, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 631, + 633, + 646 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented Twinner, a new LRM capable of creating high-quality 3D assets with accurate geometry, material properties, and illumination from only a few posed images in a feed-forward manner. Twinner leverages the novel tricolumn representation, whose memory efficiency enables the processing of voxel grids with large transformers. For training, we demonstrated the power of generating synthetic PBR assets procedurally, addressing the scarcity of high-quality PBR assets in existing synthetic datasets. We also showed that, by predicting illumination, the model can learn from real-life datasets without requiring groundtruth illumination and material properties. We evaluated Twinner on the StanfordORB benchmark, where it outperformed existing feed-forward predictors and achieved comparable quality to optimization-based methods while being orders of magnitude faster.", + "bbox": [ + 511, + 657, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5866", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sameer Agarwal, Noah Snavely, Ian Simon, Steven M. Seitz, and Richard Szeliski. Building rome in a day. In Proc. ICCV, 2009. 2", + "[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In Proc. CVPR, 2022. 2", + "[3] F O Bartell, E. L. Dereniak, and W. L Wolfe. The Theory And Measurement Of Bidirectional Reflectance Distribution Function (Brdf) And Bidirectional Transmittance Distribution Function (BTDF). In Proc. Radiation Scattering in Optical Systems, 1981. 2", + "[4] Sai Bi, Zexiang Xu, Pratul P. Srinivasan, B. Mildenhall, Kalyan Sunkavalli, Milovs Havsan, Yannick Hold-Geoffroy, D. Kriegman, and R. Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv, 2020. 2", + "[5] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Deep reflectance volumes: Relightable reconstructions from multi-view photometric images. In Proc. ECCV, 2020.", + "[6] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, David Kriegman, and Ravi Ramamoorthi. Deep 3D capture: Geometry and reflectance from sparse multi-view images. In Proc. CVPR, 2020.", + "[7] Mark Boss, Varun Jampani, Kihwan Kim, Hendrik P.A. Lensch, and Jan Kautz. Two-shot spatially-varying BRDF and shape estimation. In Proc. CVPR, 2020. 2", + "[8] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. Nerd: Neural reflectance decomposition from image collections. In IEEE International Conference on Computer Vision (ICCV), 2021. 2, 3, 6", + "[9] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan T. Barron, and Hendrik P.A. Lensch. Neuralpil: Neural pre-integrated lighting for reflectance decomposition. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3", + "[10] Brent Burley. Physically-based shading at disney. Technical report, Disney, 2012. 2", + "[11] Ang Cao, Justin Johnson, Andrea Vedaldi, and David Novotny. Lightplane: Highly-scalable components for neural 3d fields. arXiv, 2024. 2, 3, 4, 8", + "[12] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In Proc. CVPR, 2022. 4", + "[13] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 7", + "[14] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In Proc. CVPR, 2023. 2", + "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth $16 \\times 16$ words: Transformers for image recognition at scale. In Proc. ICLR, 2021. 4", + "[16] Yasutaka Furukawa and Jean Ponce. Accurate, dense, and robust multiview stereopsis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(8):1362-1376, 2010. 2", + "[17] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using monte carlo rendering and denoising. In Advances in Neural Information Processing Systems, pages 22856-22869. Curran Associates, Inc., 2022. 3, 6, 7", + "[18] Philipp Henzler, Jeremy Reizenstein, Patrick Labatut, Roman Shapovalov, Tobias Ritschel, Andrea Vedaldi, and David Novotny. Unsupervised learning of 3d object categories from videos in the wild. CVPR, 2021. 1", + "[19] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. In Proc. ICLR, 2024. 2, 3, 4, 5", + "[20] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. GaussianShader: 3D Gaussian splattering with shading functions for reflective surfaces. arXiv.cs, abs/2311.17977, 2023. 3", + "[21] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 7", + "[22] Brian Karis and Epic Games. Real shading in unreal engine 4. Proc. Physically Based Shading Theory Practice, 4(3):1, 2013. 4", + "[23] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the Fourth Eurographics Symposium on Geometry Processing, page 61-70, Goslar, DEU, 2006. Eurographics Association. 2", + "[24] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian Splating for real-time radiance field rendering. Proc. SIGGRAPH, 42(4), 2023. 1", + "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[26] Zhengfei Kuang, Yunzhi Zhang, Hong-Xing Yu, Samir Agarwala, Shangzhe Wu, and Jiajun Wu. Stanford-orb: A real-world 3d object inverse rendering benchmark, 2023. 2, 6, 7", + "[27] Ruofan Liang, Huiting Chen, Chunlin Li, Fan Chen, Selvakumar Panneer, and Nandita Vijaykumar. Envidr: Implicit differentiable renderer with neural environment lighting. In Proc. ICCV, 2023. 3", + "[28] Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. GS-IR: 3D Gaussian splattering for inverse rendering. arXiv.cs, abs/2311.16473, 2023. 3", + "[29] W. Lorensen and H. Cline. Marching cubes: A high resolution 3D surface construction algorithm. ACM Computer Graphocs, 21(24), 1987. 2" + ], + "bbox": [ + 517, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5867", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Alexander Mai, Dor Verbin, Falko Kuester, and Sara Fridovich-Keil. Neural microfacet fields for inverse rendering. In Proc. ICCV, 2023. 3, 7", + "[31] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Proc. ECCV, 2020. 1, 2", + "[32] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P. Srinivasan, and Jonathan T. Barron. Nerf in the dark: High dynamic range view synthesis from noisy raw images. In Proc. CVPR, 2022.", + "[33] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. In Proc. SIGGRAPH, 2022. 2", + "[34] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting Triangular 3D Models, Materials, and Lighting From Images. In Proc. CVPR, 2022. 2, 3, 4, 6, 7", + "[35] Giljoo Nam, J. Lee, D. Gutierrez, and Min H. Kim. Practical SVBRDF acquisition of 3D objects with unstructured flash photography. arXiv, 2018. 2", + "[36] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 4", + "[37] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proc. ICCV, 2023. 4", + "[38] Pakkapon Phongthawee, Worameth Chinchuthakun, Nontaphat Sinsunthithet, Varun Jampani, Amit Raj, Pramook Khungurn, and Supasorn Suwajanakorn. Diffusionlight: Light probes for free by painting a chrome ball. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 98-108, 2024. 3, 7, 8", + "[39] PolyHaven. Polyhaven: curated public asset library for visual effects artists and game designers. https://polyhaven.com/, 2024. Accessed: 2024-11-07. 6", + "[40] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In International Conference on Computer Vision, 2021. 2, 6", + "[41] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common Objects in 3D: Large-scale learning and evaluation of real-life 3D category reconstruction. In Proc. ICCV, 2021. 6", + "[42] Shen Sang and Manmohan Chandraker. Single-shot neural relighting and SVBRDF estimation. In Proc. ECCV, 2020. 2", + "[43] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proc. CVPR, 2016. 2", + "[44] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In Proc. ECCV, 2016. 2" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Yawar Siddiqui, Filippos Kokkinos, Tom Monnier, Mahendra Kariya, Yanir Kleiman, Emilien Garreau, Oran Gafni, Natalia Neverova, Andrea Vedaldi, Roman Shapovalov, and David Novotny. Meta 3D Asset Gen: Text-to-mesh generation with high-quality geometry, texture, and PBR materials. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2024. 2, 3, 4, 5, 6, 7", + "[46] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Frédo Durand. Light field networks: Neural scene representations with single-evaluation rendering. In Proc. NeurIPS, 2021. 4", + "[47] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proc. CVPR, 2022. 2, 4", + "[48] Cheng Sun, Guangyan Cai, Zhengqin Li, Kai Yan, Cheng Zhang, Carl Marshall, Jia-Bin Huang, Shuang Zhao, and Zhao Dong. Neural-pbir reconstruction of shape, material, and illumination. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 18000-18010, 2023. 3, 6", + "[49] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. LGM: Large multi-view Gaussian model for high-resolution 3D content creation. arXiv, 2402.05054, 2024. 1", + "[50] Giuseppe Vecchio and Valentin Deschaintre. Matsynth: A modern pbr materials dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22109-22118, 2024. 6", + "[51] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd E. Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In Proc. CVPR, 2022. 2, 5", + "[52] Guangcong Wang, Yinuo Yang, Chen Change Loy, and Zwei Liu. Stylelight: HDR panorama generation for lighting estimation and editing. In European Conference on Computer Vision (ECCV), 2022. 3, 7", + "[53] Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. VGGsF: visual geometry grounded deep structure from motion. In Proc. CVPR, 2024. 2", + "[54] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv.cs, abs/2106.10689, 2021. 2", + "[55] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. on Image Processing, 13(4), 2004. 7", + "[56] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. MeshLRM: large reconstruction model for high-quality mesh. arXiv, 2404.12385, 2024. 2", + "[57] Desai Xie, Sai Bi, Zhixin Shu, Kai Zhang, Zexiang Xu, Yi Zhou, Soren Pirk, Arie Kaufman, Xin Sun, and Hao Tan. Lrm-zero: Training large reconstruction models with synthesized data, 2024. 5, 6", + "[58] Desai Xie, Sai Bi, Zhixin Shu, Kai Zhang, Zexiang Xu, Yi Zhou, Soren Pirk, Arie Kaufman, Xin Sun, and Hao Tan." + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5868", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "LRM-Zero: training large reconstruction models with synthesized data. arXiv, 2406.09371, 2024. 2", + "[59] Zhang Xiuming, Srinivasan Pratul P., Deng Boyang, Debevec Paul, Freeman William T., and Barron Jonathan T. NeRFactor: neural factorization of shape and reflectance under an unknown illumination. In Proc. SIGGRAPH, 2021. 2, 3, 6", + "[60] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. GRM: Large gaussian reconstruction model for efficient 3D reconstruction and generation. arXiv, 2403.14621, 2024. 1", + "[61] Yinghao Xu, Hao Tan, Fujun Luan, Sai Bi, Peng Wang, Jiahao Li, Zifan Shi, Kalyan Sunkavalli, Gordon Wetzstein, Zexiang Xu, and Kai Zhang. DMV3D: Denoising multiview diffusion using 3D large reconstruction model. In Proc. ICLR, 2024. 4, 5", + "[62] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. arXiv.cs, abs/2106.12052, 2021. 2", + "[63] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. PixelNeRF: Neural radiance fields from one or few images. In Proc. CVPR, 2021. 1", + "[64] Jesus Zarzar and Bernard Ghanem. Splitnerf: Split sum approximation neural field for joint geometry, illumination, and material estimation. arXiv.cs, abs/2311.16671, 2023. 3", + "[65] Jianzhao Zhang, Guojun Chen, Yue Dong, Jian Shi, Bob Zhang, and Enhua Wu. Deep inverse rendering for practical object appearance scan with uncalibrated illumination. In Proc. Advances in Computer Graphics, 2020. 2", + "[66] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snively. PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 6", + "[67] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. GS-LRM: large reconstruction model for 3D Gaussian splatting. arXiv, 2404.19702, 2024. 2", + "[68] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proc. CVPR, pages 586-595, 2018. 7", + "[69] Jingsen Zhu, Fujun Luan, Yuchi Huo, Zihao Lin, Zhihua Zhong, Dianbing Xi, Rui Wang, Hujun Bao, Jiaxiang Zheng, and Rui Tang. Learning-based inverse rendering of complex indoor scenes with differentiable monte carlo raytracing. In SIGGRAPH Asia 2022 Conference Papers. ACM, 2022. 6" + ], + "bbox": [ + 91, + 90, + 482, + 768 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5869", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Twinner_ Shining Light on Digital Twins in a Few Snaps/0293a039-5500-41cc-90fa-fe4f8710c6e9_model.json b/2025/Twinner_ Shining Light on Digital Twins in a Few Snaps/0293a039-5500-41cc-90fa-fe4f8710c6e9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7d10df1156a70afd6be21e6d56bf0113595fcf55 --- /dev/null +++ b/2025/Twinner_ Shining Light on Digital Twins in a Few Snaps/0293a039-5500-41cc-90fa-fe4f8710c6e9_model.json @@ -0,0 +1,2169 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.216, + 0.131, + 0.784, + 0.154 + ], + "angle": 0, + "content": "Twinner: Shining Light on Digital Twins in a Few Caps" + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.179, + 0.744, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.291, + 0.907, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.525, + 0.908, + 0.569 + ], + "angle": 0, + "content": "Figure 1. We present Twinner, the first large reconstruction model capable of predicting a scene's illumination as well as an object's geometry and material properties from a few posed images. This enables tasks such as realistic relighting of objects in novel scenes in a few seconds by replacing costly per-scene optimizations with a single forward pass." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.59, + 0.327, + 0.606 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.629, + 0.486, + 0.902 + ], + "angle": 0, + "content": "We present Twinner, the first large reconstruction model capable of recovering a scene's illumination as well as an object's geometry and material properties from only a few posed images. Twinner is based on the Large Reconstruction Model and innovates in three key ways: 1) We introduce a memory-efficient voxel-grid transformer whose memory scales only quadratically with the size of the voxel grid. 2) To address the scarcity of high-quality ground-truth PBR-shaded models, we introduce a large, fully synthetic dataset of procedurally generated PBR-textured objects lit with varied illumination. 3) To bridge the synthetic-to-real gap, we finetune the model on real-world datasets using a differentiable physically based shading model, eliminating the need for ground-truth illumination or material properties, which are challenging to obtain in real-world scenarios. We demonstrate the efficacy of our model on the real-world StanfordORB benchmark, where, given a few input views, we achieve reconstruction quality significantly superior to" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.591, + 0.907, + 0.622 + ], + "angle": 0, + "content": "existing feed-forward reconstruction networks and comparable to slower per-scene optimization methods." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.649, + 0.646, + 0.665 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.674, + 0.907, + 0.856 + ], + "angle": 0, + "content": "The creation of digital twins, i.e., the conversion of real-world objects into virtual 3D assets, is an important problem in virtual and augmented reality, gaming, conservation, education, advertisement, and general 3D content creation. In recent years, new techniques such as NeRF [31] and 3D Gaussian Splatting [24], along with their feed-forward counterparts [18, 49, 60, 63], have significantly improved the quality, speed, and generality of digital twinning. However, most of these techniques still bake illumination into the scene's appearance as radiance. As a consequence, the resulting assets cannot be relit and appear out of place when visualized under different lighting conditions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To address this limitation, one must also reconstruct the field of Bidirectional Reflectance Distribution Functions (BRDF), which describe the optical properties of materi-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5859" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.348 + ], + "angle": 0, + "content": "als. Estimating the BRDF directly is notoriously challenging due to its high dimensionality [3] and is usually done using 3D scanners that capture thousands of images while controlling for viewpoint and illumination [4-7, 35, 42, 65]. This is often impractical, so authors have considered \"casual\" reconstruction setups, where the object is captured by a conventional camera in uncontrolled conditions. Examples include NVDiffRec [34], NeRD [8], and NeRFactor [59], which simultaneously estimate the illumination, in the form of an environment map, and a simplified BRDF parameterized using Physically-Based Rendering (PBR) with albedo, roughness, and metallicity [10]. These quantities are recovered by minimizing the photometric reconstruction error, comparing the output of a differentiable PBR renderer to the input images. However, this optimization is slow, ill-posed, and requires the introduction of handcrafted priors for regularization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.351, + 0.482, + 0.426 + ], + "angle": 0, + "content": "In this paper, we propose a new method, Twinner, a feedforward network that outputs shape, PBR textures, and environment illumination. This network is fast, performs well, requires only a small number of images as input, and learns empirically the necessary reconstruction priors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.429, + 0.482, + 0.563 + ], + "angle": 0, + "content": "Siddiqui et al. [45] recently introduced a similar predictor, MetalLRM, as an extension of the LightplaneLRM [11]. MetalLRM requires supervision with a large dataset of 3D assets with high-quality PBR textures, which, unfortunately, are not abundant in existing synthetic datasets like Objaverse [14]. Moreover, even if a large synthetic dataset were available, the synthetic-to-real gap between the artist-created training assets and the target domain of real input images makes MetalLRM unsuitable for digital twinning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.483, + 0.809 + ], + "angle": 0, + "content": "With our new Twinner, we address the scarcity of high-quality training PBR assets and the train/test domain shift. First, we ask whether real images in datasets like CO3Dv2 [40] can be used to supervise the model. Although these datasets lack PBR information, they are abundant and match the target distribution of real images we aim to digitally twin. To use this data, we propose minimizing the photometric error similar to optimization-based models, but with the key difference that the 3D shape, PBR materials, and environment light are all predicted by a large feedforward transformer [19, 45]. During training, each scene is rendered with a differentiable PBR shader that accepts the predicted environment light and the predicted PBR-textured shape. The latter then admits indirect PBR and illumination supervision via photometric losses between the renders and the corresponding real ground-truth images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Additionally, we consider a second approach to tackle data scarcity where, inspired by Zeroverse [58], we procedurally generate PBR-textured objects lit with a known set of environment maps. While the resulting objects are not realistic, their key advantage is that they come with clean ground-truth PBR and illumination information, which can" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "be used to supervise the corresponding predictors directly instead of indirectly via the photometric loss." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.124, + 0.905, + 0.379 + ], + "angle": 0, + "content": "Besides addressing data scarcity, we also improve the architecture of the reconstruction model. We believe in the virtues of a simple and direct voxel-grid representation of the object [47], but note that predicting a number of tokens cubic in the voxel grid resolution is intractable with a transformer. Reducing the number of predicted tokens is the primary reason why recent models [19, 45] factor the voxel grid as a triplane. The question, then, is how to retain the benefits of the direct voxel grid representation and still make computations feasible. We solve this problem by introducing the novel tricolumn representation, where each feature vector (column) in a triplane is obtained by stacking the corresponding cells in the underlying voxel grid. In this manner, the number of tokens becomes quadratic in the voxel grid size, while each voxel is still represented explicitly. We further augment the network to predict a representation of the environment map, capturing illumination." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.381, + 0.905, + 0.53 + ], + "angle": 0, + "content": "We demonstrate the efficacy of the Twinner model on the StanfordORB benchmark [26], consisting of casual captures of real-life objects. We achieve reconstruction quality superior to existing feed-forward predictors and comparable to optimization-based methods, while using only a very small number of input images of the object and operating much faster. We also obtain state-of-the-art reconstruction of the environment maps, achieving a \\(28\\%\\) improvement in the predicted illumination's angular error over existing methods." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.55, + 0.652, + 0.565 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.577, + 0.905, + 0.848 + ], + "angle": 0, + "content": "3D reconstruction. 3D reconstruction aims to recover a three-dimensional representation of a scene from a set of posed images. Traditionally, this was accomplished with structure-from-motion [1, 43, 53], multi-view stereo [16, 44], and mesh surface extraction methods [23, 29]. Recently, methods based on neural fields [2, 31-33, 51, 54, 62] have focused on extracting both shape and appearance by learning the scene radiance function. Through volumetric rendering and photometric losses, they are able to implicitly reconstruct a 3D density volume from which a mesh can be extracted. However, these methods infer the parameters of each scene via test-time optimization, making them slow. Authors have thus introduced new feed-forward reconstruction networks, generically called Large Reconstruction Models (LRMs) [19, 56, 58, 67]. LRMs use transformers to predict 3D representations from a small set of input images. Our Twinner predicts not just geometry and radiance, but also material properties and illumination." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Material reconstruction. 3D reconstruction typically recovers a scene's geometry and radiance under fixed illumination. However, it is also necessary to recover material" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5860" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.087, + 0.91, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.287, + 0.908, + 0.359 + ], + "angle": 0, + "content": "Figure 2. Overview of Twinner. The input views are processed together with their foreground masks and camera poses by an image tokenizer. The resulting tokens are processed by a Diffusion Transformer (DiT) model to predict a 3D volumetric representation of the scene, and a second DiT model to predict the scene's illumination. The 3D representation is then rendered to obtain images of the scene's material properties, normals, and opacity from the target's point of view. Using these rendered images together with the predicted illumination, Twinner renders an approximate physically-shaded image of the scene which then enters a photometric loss." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.371, + 0.484, + 0.569 + ], + "angle": 0, + "content": "properties so that the 3D assets look realistic in novel virtual scenes. To do so, some works focus on optimizing explicit representations, such as material textures on meshes [17, 34, 48], while other works have added physically-based rendering models into neural fields [8, 9, 21, 27, 30, 64] and Gaussian splats [20, 28]. While these methods achieve high-quality material reconstructions, they require costly per-scene optimization. Some have addressed this issue by learning feed-forward reconstructors of PBR assets [45]. However, these are typically trained with synthetic data due to the complexity of capturing real-world material properties. We bridge the gap between synthetic and real data by training Twinner on real-world datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.574, + 0.484, + 0.818 + ], + "angle": 0, + "content": "Illumination prediction. A key to accurate material reconstruction is accurate modeling of the scene's illumination. Single-scene optimization methods achieve this by optimizing an internal representation of illumination using light probe images [17, 34, 59], spherical harmonics [28], spherical Gaussians [21, 66], or MLPs [27, 30, 64]. These methods rely on physical models of lighting to aid the optimization. Others have attempted to predict the illumination of a scene from a single image [38, 52]. They rely on generative models to map the input view to a possible distribution of panorama views representing the environment illumination. Our Twinner lies in between these two lines of work: it is similar to the predictive approach as it predicts illumination from a set of four input views, but it uses physically-based rendering losses for supervision, similar to single-scene optimization methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.226, + 0.848 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We introduce Twinner, a model that reconstructs the shape, materials, and illumination of real-life 3D objects from a small set of posed input views. Twinner extends Light-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.371, + 0.907, + 0.507 + ], + "angle": 0, + "content": "planeLRM [11], an enhanced version of LRM [19], by incorporating predictors for the object's material properties and scene illumination, as well as a differentiable physically-based renderer. In Sec. 3.1, we begin by reviewing the reconstructor and the PBR model we use. Then, in Secs. 3.2 to 3.4, we describe the main building blocks of Twinner. Finally, in Sec. 3.5, we detail the training protocol, including the losses used to supervise Twinner on real data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.517, + 0.655, + 0.532 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.909, + 0.66 + ], + "angle": 0, + "content": "Physically-based rendering. To supervise our model with a photometric loss, it is necessary to render the reconstructed scene into an image. The color of each pixel corresponds to the radiance \\(\\mathbf{L}_o\\) emitted by the imaged object point \\(p\\in \\mathbb{R}^3\\) in the direction of the camera. With knowledge of a scene's illumination, geometry, and material properties, it is possible to estimate the outgoing radiance at any such point according to the reflectance equation:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.67, + 0.907, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\mathbf {L} _ {o} = \\int_ {\\Omega} \\left(k _ {d} \\frac {\\mathbf {a}}{\\pi} + \\mathbf {f} _ {s}\\right) \\mathbf {L} _ {i} \\left\\langle \\boldsymbol {\\omega} _ {i}, \\mathbf {n} \\right\\rangle d \\boldsymbol {\\omega} _ {i}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{L}_i\\) is the incoming radiance, representing the scene illumination falling on the imaged point \\(\\pmb{p}\\). The integral is carried out over a hemisphere \\(\\Omega\\) of incoming directions \\(\\omega_{i}\\) determined by the surface normal \\(\\mathbf{n}\\) at \\(\\pmb{p}\\). The factor in parentheses is a model of the object's Bidirectional Reflectance Distribution Function (BRDF), which describes how the incoming radiance reflects off the object's surface into various directions. The first term is the diffuse component, parameterized by the object's diffuse albedo \\(\\mathbf{a}\\) and a constant \\(k_{d}\\), both of which are independent of direction. In the integral, this term averages all incoming radiance over the hemisphere. The second term, \\(\\mathbf{f}_s\\), is the specular com" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5861" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.097, + 0.91, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.26, + 0.908, + 0.318 + ], + "angle": 0, + "content": "Figure 3. Tricolumn Architecture. The transformer module processes image tokens and a learnable embedding to predict the tricolumn representation, which consists of a grid of \\(3R^2\\) feature vectors with dimensionality \\(\\mathbb{R}^{CR}\\). These feature vectors are then split into three axis-aligned column grids, each with dimensionality \\(\\mathbb{R}^{CR}\\). Finally, all axis-aligned column grids are reshaped into voxel grids of size \\(R\\times R\\times R\\) with dimensionality \\(\\mathbb{R}^C\\) and summed into a single voxel grid." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.33, + 0.482, + 0.45 + ], + "angle": 0, + "content": "ponent of the BRDF, which captures directional effects such as reflections or highlights. Note that \\(\\mathbf{L}_i\\) is a function of the incoming direction \\(\\omega_{i}\\), and the specular BRDF \\(\\mathbf{f}_s\\) is a function of both the incoming and outgoing directions \\(\\omega_{i}\\) and \\(\\omega_{o}\\). Radiances and BRDFs also depend on the light frequency, which we model using three RGB components. Furthermore, all quantities depend on the point position \\(\\pmb{p}\\). We omit these details for brevity." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.453, + 0.483, + 0.619 + ], + "angle": 0, + "content": "Due to the complexity of the integral in Eq. (1), numerous approximations have been proposed. We rely on the split-sum approximation [22] for image-based lighting, which leverages the scene's illumination represented as a spherical image enveloping the scene at an infinite radius. Following [34], we use a set of differentiable mipmaps to represent the pre-integrated lighting. The BRDF follows the Disney [22] microfacet BRDF model and is parameterized by the albedo \\( \\mathbf{a} \\), metalness \\( m \\), and roughness \\( r \\) at each point. We further assume that \\( \\mathbf{L}_i \\) is constant in space (hence we do not account for shadows cast by the object)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.62, + 0.484, + 0.696 + ], + "angle": 0, + "content": "In summary, to apply this model to render an image of the object, we require an estimate of the object's 3D shape and its albedo, metalness, and roughness at its surfaces. We also require a spherical image representing the directional function \\(\\mathbf{L}_i\\), known as an environment map." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.484, + 0.824 + ], + "angle": 0, + "content": "Large Reconstruction Model (LRM). The aim of LRM [19] is to map a few posed views \\(\\mathcal{I}\\) of the scene to a 3D reconstruction \\(V\\). LRM first extracts features \\(y\\) from the set of input views, utilizing a pre-trained DINO [36] Vision Transformer (ViT) [15]. Following [61], we feed into the ViT the RGB images \\(\\mathcal{I}\\), the Plücker embeddings [46] \\(\\mathcal{I}_{\\mathrm{Pl}}\\) representing the cameras, and the masks \\(\\mathcal{M}\\) of the foreground object to be reconstructed." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The extracted feature tokens enter, via cross-attention, the transformer architecture from Diffusion Transformer (DiT) [37] \\(\\Phi\\), which is tasked with reconstructing the scene. We emphasize that we borrow the DiT architecture without any diffusion process. The tokens in this transformer en" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.906, + 0.404 + ], + "angle": 0, + "content": "code the 3D representation of the scene, realizing a mapping \\( V = \\Phi(V_0|y) \\), where \\( V_0 \\) are the learnable initial tokens. A crucial design choice is how to construct the tokens \\( V \\) manipulated by this transformer, which is discussed in Sec. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.406, + 0.906, + 0.497 + ], + "angle": 0, + "content": "Finally, using the memory-efficient rendering algorithm proposed in Lightplane [11], a target set of views is rendered from the 3D representation. A shaded RGB image \\(\\hat{\\mathcal{I}}\\), a foreground mask image \\(\\hat{\\mathcal{M}}\\), and a depth image \\(\\hat{\\mathcal{I}}_{\\mathrm{d}}\\) are rendered for each view and used to supervise the LRM through a combination of the following photometric losses:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.506, + 0.905, + 0.521 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {M} = \\operatorname {B C E} (\\mathcal {M}, \\hat {\\mathcal {M}}), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.526, + 0.905, + 0.543 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {R} = \\operatorname {L P I P S} \\left(\\mathcal {M} \\mathcal {I}, \\mathcal {M} \\hat {\\mathcal {I}}\\right) + \\left\\| \\mathcal {M} \\mathcal {I} - \\mathcal {M} \\hat {\\mathcal {I}} \\right\\| ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.548, + 0.905, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d} = \\left\\| \\mathcal {M I} _ {\\mathrm {d}} - \\mathcal {M} \\hat {\\mathcal {I}} _ {\\mathrm {d}} \\right\\|, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.574, + 0.906, + 0.605 + ], + "angle": 0, + "content": "where \\(\\mathcal{I}_{\\mathrm{d}}\\) is the set of ground-truth depth maps corresponding to the input images \\(\\mathcal{I}\\), and \\(\\mathcal{M}\\mathcal{I}\\) denotes a masked image." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.613, + 0.747, + 0.629 + ], + "angle": 0, + "content": "3.2. Tricolumn representation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.635, + 0.906, + 0.8 + ], + "angle": 0, + "content": "Crucial to the Twinner design is the representation of the 3D scene, which is encoded as a set of volumetric functions expressing opacity and material properties at every point \\(\\pmb{p} \\in \\mathbb{R}^3\\) in space. A simple and effective option [47] to encode a volumetric map \\(V: \\mathbb{R}^3 \\to \\mathbb{R}^C\\) is to use a voxel grid \\(V \\in \\mathbb{R}^{C \\times D \\times H \\times W}\\). Here, the value \\(V(\\pmb{p}) \\in \\mathbb{R}^C\\) interpolates the voxel values at the nearest grid points. To predict the voxel grid \\(V\\) with the transformer of Sec. 3.1, we could assign one token to each voxel. However, the number \\(WHD\\) of voxels grows cubically with the resolution of the grid, which leads to prohibitive memory demands." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.801, + 0.906, + 0.877 + ], + "angle": 0, + "content": "The alternative adopted by models like [19, 45] is to use a triplane representation [12]. In this case, the volumetric function is represented by three 2D grids \\( V_{xy} \\in \\mathbb{R}^{K \\times H \\times W} \\), \\( V_{yz} \\in \\mathbb{R}^{K \\times W \\times D} \\), and \\( V_{zx} \\in \\mathbb{R}^{K \\times D \\times H} \\). The volumetric function at \\( \\pmb{p} = (x,y,z) \\) is then expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.885, + 0.905, + 0.903 + ], + "angle": 0, + "content": "\\[\nV (x, y, z) = f \\left(V _ {x y} (x, y), V _ {y z} (y, z), V _ {z x} (z, x)\\right), \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5862" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.488, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.279, + 0.486, + 0.322 + ], + "angle": 0, + "content": "Figure 4. Examples from our procedural dataset visualizing the shaded image, environment map, materials, and normals from our environment-map-endowed version of Zeroverse [57]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.332, + 0.484, + 0.422 + ], + "angle": 0, + "content": "where \\( f: \\mathbb{R}^{K \\times K \\times K} \\to \\mathbb{R}^C \\) is a point-wise mixing function (e.g., an MLP applied to the concatenation of the three \\( K \\)-dimensional feature vectors). Paired with a transformer, this conveniently reduces the number of tokens to \\( WH + HD + DW \\) only. However, triplane mixes information between different voxels, which can lead to artifacts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.484, + 0.498 + ], + "angle": 0, + "content": "We thus propose a novel tricolumn representation instead. This is similar to triplane as it uses a plane instead of a volume of tokens; however, each feature vector is in itself a column of corresponding voxel cells, so that voxel-specific information is represented directly and separately." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.499, + 0.484, + 0.605 + ], + "angle": 0, + "content": "In detail, we consider a plane of feature vectors \\( V_{xy} \\in \\mathbb{R}^{(CR) \\times R \\times R} \\). Because this is a plane, there are only \\( R^2 \\) tokens, which are easily handled by the transformer. However, \\( V_{xy} \\) can be reshaped into a \\( C \\times R \\times R \\times R \\) voxel grid by unflattening the first dimension. We use the symbol \\( V_{xy}(x,y;z) \\in \\mathbb{R}^C \\) to denote the feature vector obtained by interpolating this grid in 3D." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.605, + 0.484, + 0.636 + ], + "angle": 0, + "content": "When three such planes are used, the full feature vector at \\( \\pmb{p} \\) is decoded as:" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.643, + 0.484, + 0.674 + ], + "angle": 0, + "content": "\\[\nV (x, y, z) = f \\left(V _ {x y} (x, y; z), V _ {y z} (y, z; x), V _ {z x} (z, x; y)\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.737 + ], + "angle": 0, + "content": "Note that this construction is strictly different from the triplane equation (5) as each term depends on all three spatial coordinates. We opt for using summation followed by an MLP as the mixing function \\( f \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.332, + 0.758 + ], + "angle": 0, + "content": "3.3. PBR reconstruction model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The LRM models of [19, 61] only recover the shape and radiance of the scene. Like in [45], we extend Twinner to also recover the material properties. We do so by extending the tricolumn representation to encode the albedo \\( a \\) (instead of radiance) and two additional channels representing the scene's metalness \\( m \\) and roughness \\( r \\). We also note the importance of accurate normals for relighting. While a normal image could be directly estimated by projecting the rendered depth image into 3D space and taking the gradient" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.906, + 0.181 + ], + "angle": 0, + "content": "of the opacity field, \\(\\bar{\\mathcal{I}}_{\\mathrm{n}} = \\nabla \\sigma (\\pi^{-1}(\\hat{\\mathcal{I}}_{\\mathrm{d}}))\\), this estimate tends to be noisy. Therefore, we follow previous works [51] and predict three additional channels representing the scene's normals. The predicted normal image \\(\\hat{\\mathcal{I}}_{\\mathrm{n}}\\) is then supervised with the pseudo-ground-truth normals from the opacity field \\(\\bar{\\mathcal{I}}_{\\mathrm{n}}\\) via:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.181, + 0.906, + 0.199 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {N} = 1 - \\langle \\mathcal {M} \\bar {\\mathcal {I}} _ {\\mathrm {n}}, \\mathcal {M} \\hat {\\mathcal {I}} _ {\\mathrm {n}} \\rangle . \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.202, + 0.906, + 0.307 + ], + "angle": 0, + "content": "Direct supervision using synthetic data. Reconstruction models that predict radiance can be supervised by rendering the predicted 3D scene into images and comparing these to training views using a photometric loss. In our case, however, the model does not predict radiance, but albedo, roughness, and metalness, which cannot be rendered into a 'shaded' image unless illumination is also known." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.308, + 0.906, + 0.414 + ], + "angle": 0, + "content": "A simple approach for supervising the model that does not require knowledge of illumination is to render albedo, roughness, and metalness images \\(\\hat{\\mathcal{I}}_{\\mathrm{a}},\\hat{\\mathcal{I}}_{\\mathrm{r}}\\) ,and \\(\\hat{\\mathcal{I}}_{\\mathrm{m}}\\) , and compare these to their corresponding ground-truth renders. We use this approach to supervise our model with synthetic data by applying the following loss on the set of material renders \\(\\hat{\\mathcal{I}}_{\\mathrm{a}},\\hat{\\mathcal{I}}_{\\mathrm{m}},\\hat{\\mathcal{I}}_{\\mathrm{r}}\\) , and also on the normal render \\(\\hat{\\mathcal{I}}_{\\mathrm{n}}\\) .." + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.418, + 0.905, + 0.45 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S} = \\ell_ {\\mathcal {M}} \\left(\\mathcal {I} _ {\\mathrm {a}}, \\hat {\\mathcal {I}} _ {\\mathrm {a}}\\right) + \\ell_ {\\mathcal {M}} \\left(\\mathcal {I} _ {\\mathrm {r}}, \\hat {\\mathcal {I}} _ {\\mathrm {r}}\\right) + \\ell_ {\\mathcal {M}} \\left(\\mathcal {I} _ {\\mathrm {m}}, \\hat {\\mathcal {I}} _ {\\mathrm {m}}\\right) + \\ell_ {\\mathcal {M}} \\left(\\mathcal {I} _ {\\mathrm {n}}, \\hat {\\mathcal {I}} _ {\\mathrm {n}}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.904, + 0.481 + ], + "angle": 0, + "content": "where \\(\\ell_{\\mathcal{M}}(u,v) = \\| \\mathcal{M}(u - v)\\|_{\\epsilon}\\) is the Huber loss \\(\\| \\cdot \\|_{\\epsilon}\\) with cut-off threshold \\(\\epsilon\\) masked by \\(\\mathcal{M}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.488, + 0.906, + 0.505 + ], + "angle": 0, + "content": "3.4. Training on real data by predicting illumination" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.51, + 0.906, + 0.646 + ], + "angle": 0, + "content": "The learning formulation of Sec. 3.3 requires ground-truth material properties to be known, which is not the case for real-world data where only shaded images are available. In order to fine-tune our model using real data with a photometric loss, it is necessary to also estimate the scene illumination. With this, we can use the differentiable physically-based shader detailed in Sec. 3.1 to render the shaded scene appearance and compare it to the input images through a photometric loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.647, + 0.907, + 0.828 + ], + "angle": 0, + "content": "To estimate the scene's illumination, we propose using an additional DiT module. The latter takes as input the feature tokens \\(y\\) computed by the image feature extractor and predicts a cubemap representation of the scene's lighting. Recall that a cubemap is a function \\(C \\to \\mathbb{R}^3\\) that assigns every point \\(\\pmb{p}\\) of the cube surface \\(C = \\{\\pmb{p} : \\| \\pmb{p} \\|_{\\infty} = 1\\}\\) to the radiance \\(\\mathbf{L}_i(\\boldsymbol{\\omega}_i)\\) incoming from direction \\(\\boldsymbol{\\omega}_i \\propto -\\pmb{p}\\), and is thus a representation of the environment illumination (this works because we assume that the function \\(\\mathbf{L}_i\\) only depends on direction, but not on position). The cubemap is unfolded to a collection of six 2D images, which are directly predicted by the DiT module." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.833, + 0.906, + 0.879 + ], + "angle": 0, + "content": "Supervising illumination with synthetic data. In synthetic data, the cubemap is known, and its predictor can be supervised directly with the environment-map loss:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.884, + 0.906, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {e n v}} = \\left\\| \\mathcal {I} _ {\\mathrm {e}} - \\hat {\\mathcal {I}} _ {\\mathrm {e}} \\right\\| ^ {2} + \\operatorname {L P I P S} \\left(\\mathcal {I} _ {\\mathrm {e}}, \\hat {\\mathcal {I}} _ {\\mathrm {e}}\\right), \\tag {9}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5863" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.099, + 0.089, + 0.904, + 0.251 + ], + "angle": 0, + "content": "
Novel Scene RelightingNovel View SynthesisGeometry EstimationRuntime
MethodPSNR-H↑PSNR-L↑SSIM↑LPIPS↓PSNR-H↑PSNR-L↑SSIM↑LPIPS↓Depth↓Geometry↓Shape↓Time (hr.)↓
Single-Scene Opt.PhySG [66]21.8128.110.9600.05524.2432.150.9740.0471.900.179.28~3
NVDiffRec [34]22.9129.720.9630.03921.9428.440.9690.0300.310.060.62~1
NeRD [8]23.2929.650.9570.05925.8332.610.9630.0541.390.2813.70>20
NeRFactor [59]23.5430.380.9690.04826.0633.470.9730.0460.870.299.53>20
InvRender [69]23.7630.830.9700.04625.9134.010.9770.0420.590.060.44~3
NVDiffRecMC [17]24.4331.600.9720.03628.0336.400.9820.0280.320.040.51~2
Neural-PBIR [48]26.0133.260.9790.02328.8236.800.9860.0190.300.060.43~1
PredictiveMetaLRM [45]21.5028.060.9570.05219.5525.980.9560.0518.6690.061.260.004
w/o ZeroVerse23.5330.310.9650.04821.5328.780.9650.0481.1260.090.740.006
w/o Shading Loss23.8930.850.9670.04221.4229.470.9670.0410.7620.060.300.004
Twinner (ours)24.5531.610.9690.03821.7029.620.9680.0370.7090.060.370.006
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.262, + 0.907, + 0.331 + ], + "angle": 0, + "content": "Table 1. Relighting evaluation on StanfordORB [26] comparing our method to other feedforward predictors of PBR-textured assets (two ablations and MetalLRM). For completeness, we also include single-scene optimization methods in the top part of the table. Our model outperforms most single-scene optimization baseline methods in terms of novel scene relighting and shape estimation despite being a predictive model and having access to only four input views. More impressively, it only requires a few seconds to reconstruct an object's mesh with PBR materials, while single-scene optimization methods have runtimes in the order of hours." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.345, + 0.482, + 0.422 + ], + "angle": 0, + "content": "comprising the photometric loss from (8) using the predicted cubemap projected into an equirectangular image \\(\\hat{\\mathcal{I}}_{\\mathrm{e}}\\) and the ground-truth equirectangular image \\(\\mathcal{I}_{\\mathrm{e}}\\). In practice, we use a procedurally generated dataset of synthetic assets detailed in Sec. 3.5 to provide ground-truth illumination." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.429, + 0.485, + 0.701 + ], + "angle": 0, + "content": "Supervising illumination with real data. In real data, the cubemap is observed indirectly via the effect it has on the object's appearance according to the reflectance equation (1), which provides only weak supervision. We note, however, that the scene background visible in the input images is, to a first approximation, equal to the function \\(\\mathbf{L}_i\\), and can thus provide direct supervision of the latter. We can render the background observed from any given camera by assigning to each pixel the color \\(\\mathbf{L}_i(\\boldsymbol{\\omega}_i) = C(\\boldsymbol{\\omega}_i / \\| \\boldsymbol{\\omega}_i \\|_{\\infty})\\) of the incoming radiance from the direction \\(\\boldsymbol{\\omega}_i\\) of the pixel's ray. Since input images are in sRGB color space, the predicted background image is first tone-mapped and then composited with the foreground object prediction to obtain a scene image. The latter allows us to supervise both the illumination prediction and the shaded object prediction through the photometric loss \\(\\mathcal{L}_R\\) of Eq. (3). Here, as training data, we use abundant object-centric video captures from CO3Dv2 [41] as described in the next section." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.265, + 0.729 + ], + "angle": 0, + "content": "3.5. Training protocol" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.736, + 0.376, + 0.751 + ], + "angle": 0, + "content": "Twinner's training minimizes the total loss" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.765, + 0.483, + 0.782 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t}} = \\mathcal {L} _ {N} + \\mathcal {L} _ {S} + \\mathcal {L} _ {\\mathrm {T V}} + \\mathcal {L} _ {d} + \\mathcal {L} _ {M} + \\mathcal {L} _ {R} + \\mathcal {L} _ {\\mathrm {T V}}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.794, + 0.483, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{TV}} = \\mathrm{TV}(\\hat{\\mathcal{I}}_{\\mathrm{n}}) + \\mathrm{TV}(\\hat{\\mathcal{I}}_{\\mathrm{a}}) + \\mathrm{TV}(\\hat{\\mathcal{I}}_{\\mathrm{d}})\\) is the Total-Variation (TV) regularization on normal, albedo, and depth images promoting smooth solutions. \\(\\mathcal{L}_{\\mathrm{tot}}\\) is optimized with Adam [25] using a learning rate of \\(10^{-4}\\) until convergence. As mentioned before, individual terms of \\(\\mathcal{L}_{\\mathrm{tot}}\\) are switched on or off depending on whether a real or a synthetic training example is present in the batch of training images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.345, + 0.907, + 0.452 + ], + "angle": 0, + "content": "Training data. Following [45], Twinner is trained on a large dataset of 140k artist-created assets similar to Objaverse. However, this dataset often lacks high-quality material maps where, for instance, distinguishing between missing material images and material images with zero values is impossible. Moreover, there is a large domain gap between the synthetic objects and the real ones we intend to twin." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.907, + 0.531 + ], + "angle": 0, + "content": "As such, together with the artist-created assets, we also train on two additional data sources: 1) A large dataset of procedurally generated assets with clean PBR textures and environment maps; 2) Real-object captures from CO3Dv2 [40]. Both datasets are detailed next." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.542, + 0.909, + 0.724 + ], + "angle": 0, + "content": "Endowing Zeroverse with environment maps. Following LRM-Zero [57], we generate a set of 10k randomized 3D objects with realistic material properties sampled from the MatSynth [50] dataset. Additionally, to enable the environment-map supervision loss \\(\\mathcal{L}_{\\mathrm{env}}\\) (9), for each rendered scene, we randomly sample one of \\(\\sim 1\\mathrm{k}\\) environment maps from PolyHaven [39]. We render color, normal, material, depth, and foreground mask images for each object from a predetermined set of viewpoints (see examples in Fig. 4). As such, samples from this dataset admit optimizing all terms of the total loss \\(\\mathcal{L}_{\\mathrm{tot}}\\), including the environment loss \\(\\mathcal{L}_{\\mathrm{env}}\\), which is crucial for stable training of Twinner." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Exploiting real object captures. We also train on CO3Dv2 [40], which is a large-scale real-world dataset of \"turntable\" videos capturing objects of common categories. Each frame in the dataset comes with camera and foreground mask annotations. As such, for each CO3Dv2 sample, we can optimize the normal, mask, photometric, and TV losses \\(\\mathcal{L}_{\\mathrm{tot}}^{\\mathrm{real}} = \\mathcal{L}_N + \\mathcal{L}_M + \\mathcal{L}_R + \\mathcal{L}_{\\mathrm{TV}}\\). Since \\(\\mathcal{L}_{\\mathrm{tot}}^{\\mathrm{real}}\\) is a function of the illumination image \\(\\hat{\\mathcal{I}}_{\\mathrm{e}}\\) and PBR renders \\(\\hat{\\mathcal{I}}_{\\mathrm{a}}\\), \\(\\hat{\\mathcal{I}}_{\\mathrm{m}}\\), \\(\\hat{\\mathcal{I}}_{\\mathrm{r}}\\), the loss provides a valuable supervisory signal for all volumetric predictors. Furthermore, using real training data narrows the domain gap between the training set and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5864" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.441, + 0.107 + ], + "angle": 0, + "content": "the target domain of Twinner, which are real images." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.118, + 0.484, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.277, + 0.484, + 0.318 + ], + "angle": 0, + "content": "Figure 5. StanfordORB Illumination Visualization. We visualize the illumination estimation of multiple baselines and our method on four different scenes from the StanfordORB dataset." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.332, + 0.481, + 0.42 + ], + "angle": 0, + "content": "
MethodAngular Error ↓RMSE↓Norm. RMSE↓SI RMSE↓Time (hr.)↓
NMF [30]11.240.4870.5240.394~ 2
TensoIR [21]15.080.7520.7980.451~ 2
NVDiffRec [34]12.260.7230.7150.367~ 1
NVDiffRecMC [17]13.250.5740.6140.527~ 2
StyleLight [52]13.050.6400.6890.4400.011
DiffusionLight [38]11.350.5810.5690.5440.317
Twinner (ours)8.090.5700.6020.4720.006
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.43, + 0.484, + 0.486 + ], + "angle": 0, + "content": "Table 2. Illumination prediction results. We provide a quantitative evaluation of the illumination estimation for baselines and our Twinner. Rows are grouped by single-scene optimization models (top) and predictive feed-forward models (bottom)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.501, + 0.224, + 0.518 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.526, + 0.483, + 0.571 + ], + "angle": 0, + "content": "Evaluation data. We evaluate Twinner on the StanfordORB [26] dataset, which comprises 42 captures of 14 different objects, each imaged under 3 different illuminations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.572, + 0.484, + 0.617 + ], + "angle": 0, + "content": "Each object has a high-quality geometric reconstruction obtained using a 3D scanner and comes with ground-truth camera poses and estimated illumination." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.623, + 0.483, + 0.759 + ], + "angle": 0, + "content": "Evaluation Protocol. For each evaluation scene, we select four views from the set of training images as input frames via furthest camera-origin point sampling. With these input views, Twinner reconstructs a 3D PBR-textured mesh. Each predicted mesh is rendered with Blender [13] to held-out test views while being lit with a given ground-truth environment map. Following [26], the renders are compared to the corresponding ground-truth images using scale-invariant versions of LPIPS [68], SSIM [55], and PSNR." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Baselines. We compare to existing feed-forward predictors of PBR-textured meshes. As such, we selected the state-of-the-art MetalLRM few-view reconstructor from AssetGen [45] as our main baseline. We use the version of MetalLRM conditioned on four shaded input views, as opposed to the albedo+shaded conditioning used in the AssetGen pipeline. We also compare to two ablations of our method. The first ablation, Twinner w/o Shading Loss, evaluates the contribution of training on real data and omits" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.909, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.573, + 0.907, + 0.643 + ], + "angle": 0, + "content": "Figure 6. Qualitative results on StanfordORB. We visualize the shaded RGB, geometry reconstruction, and material predictions for our baseline, two ablations, and our Twinner. Our model is able to better predict relightable objects thanks to the combination of curated synthetic training data and the shading loss on real data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.665, + 0.907, + 0.845 + ], + "angle": 0, + "content": "the training images of CO3Dv2. Second, we compare to Twinner w/o Zeroverse, which ablates training of the environment-map and PBR-texture predictor on our version of the procedurally-created Zeroverse. For completeness, we also include several single-scene optimization methods that \"overfit\" a relightable implicit shape in each scene from scratch. Note that, to save GPU memory, our feed-forward predictors randomly sample only four source views from the pool of all available training images of each scene. This puts us at a disadvantage compared to single-scene optimizers, as they leverage all available scene views for each reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Relighting Evaluation. All quantitative results are reported in Tab. 1. Twinner is the best among feed-forward reconstructors, outperforming its own two ablations and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5865" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.086, + 0.912, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.435, + 0.908, + 0.477 + ], + "angle": 0, + "content": "Figure 7. CO3Dv2 Qualitative Results. We visualize predicted PBR meshes from the CO3Dv2 dataset illuminated under direct sunlight as well as their geometry, albedo, and material properties. The shading loss aids at predicting more realistic material properties, while the synthetic data improves the predicted geometry and albedo." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.491, + 0.483, + 0.643 + ], + "angle": 0, + "content": "MetalLRM. For instance, Twinner achieves a \\(27\\%\\) reduction in both novel-scene relighting and novel-view synthesis LPIPS compared to the strongest baseline (MetalLRM), while maintaining comparable inference time. Remarkably, our method is on par with most single-scene optimization methods while being significantly faster, requiring only a few seconds per prediction compared to several hours for optimization-based methods. Examples of our predictions and the baselines are shown in Fig. 6. We also evaluate qualitatively on testing sequences of CO3Dv2 in Fig. 7." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.65, + 0.483, + 0.788 + ], + "angle": 0, + "content": "Illumination Evaluation. We evaluate Twinner's ability to predict a scene's illumination from four input images. Since there are no existing baselines for few-view conditioned feed-forward illumination prediction, we compare to environment map estimates output by the single-scene optimizers in Tab. 1, as well as to single-view conditioned predictors. We follow [38] and evaluate illumination in terms of angular error, RMSE, normalized RMSE, and scale-invariant RMSE. Results are given in Tab. 2 and Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Tricolumn Ablation. We also evaluate the effect of our tricolumn representation. In Tab. 3, we report the novel-view synthesis performance of the vanilla LightplaneLRM [11] model trained only on our synthetic dataset, utilizing either tricolumn or triplane representations. Tricolumn outperforms triplane significantly, achieving a \\(22\\%\\) improvement in LPIPS, validating this contribution." + }, + { + "type": "table", + "bbox": [ + 0.554, + 0.488, + 0.868, + 0.547 + ], + "angle": 0, + "content": "
3D representationPSNR↑SSIM↑LPIPS↓
Triplane31.200.9720.037
Tricolumn (ours)32.810.9660.029
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.557, + 0.907, + 0.613 + ], + "angle": 0, + "content": "Table 3. Tricolumn ablation evaluating novel view synthesis on the StanfordORB dataset for a LightplaneLRM [11] using the triplane 3D representation or our proposed tricolumn instead. Our tricolumn outperforms triplane on all reconstruction metrics." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.632, + 0.634, + 0.647 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We have presented Twinner, a new LRM capable of creating high-quality 3D assets with accurate geometry, material properties, and illumination from only a few posed images in a feed-forward manner. Twinner leverages the novel tricolumn representation, whose memory efficiency enables the processing of voxel grids with large transformers. For training, we demonstrated the power of generating synthetic PBR assets procedurally, addressing the scarcity of high-quality PBR assets in existing synthetic datasets. We also showed that, by predicting illumination, the model can learn from real-life datasets without requiring groundtruth illumination and material properties. We evaluated Twinner on the StanfordORB benchmark, where it outperformed existing feed-forward predictors and achieved comparable quality to optimization-based methods while being orders of magnitude faster." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5866" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.155 + ], + "angle": 0, + "content": "[1] Sameer Agarwal, Noah Snavely, Ian Simon, Steven M. Seitz, and Richard Szeliski. Building rome in a day. In Proc. ICCV, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.157, + 0.483, + 0.199 + ], + "angle": 0, + "content": "[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In Proc. CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.201, + 0.482, + 0.268 + ], + "angle": 0, + "content": "[3] F O Bartell, E. L. Dereniak, and W. L Wolfe. The Theory And Measurement Of Bidirectional Reflectance Distribution Function (Brdf) And Bidirectional Transmittance Distribution Function (BTDF). In Proc. Radiation Scattering in Optical Systems, 1981. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.482, + 0.325 + ], + "angle": 0, + "content": "[4] Sai Bi, Zexiang Xu, Pratul P. Srinivasan, B. Mildenhall, Kalyan Sunkavalli, Milovs Havsan, Yannick Hold-Geoffroy, D. Kriegman, and R. Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.327, + 0.482, + 0.381 + ], + "angle": 0, + "content": "[5] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Deep reflectance volumes: Relightable reconstructions from multi-view photometric images. In Proc. ECCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.383, + 0.482, + 0.435 + ], + "angle": 0, + "content": "[6] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, David Kriegman, and Ravi Ramamoorthi. Deep 3D capture: Geometry and reflectance from sparse multi-view images. In Proc. CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.482, + 0.479 + ], + "angle": 0, + "content": "[7] Mark Boss, Varun Jampani, Kihwan Kim, Hendrik P.A. Lensch, and Jan Kautz. Two-shot spatially-varying BRDF and shape estimation. In Proc. CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.482, + 0.548 + ], + "angle": 0, + "content": "[8] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. Nerd: Neural reflectance decomposition from image collections. In IEEE International Conference on Computer Vision (ICCV), 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.482, + 0.619 + ], + "angle": 0, + "content": "[9] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan T. Barron, and Hendrik P.A. Lensch. Neuralpil: Neural pre-integrated lighting for reflectance decomposition. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.621, + 0.482, + 0.647 + ], + "angle": 0, + "content": "[10] Brent Burley. Physically-based shading at disney. Technical report, Disney, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.65, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[11] Ang Cao, Justin Johnson, Andrea Vedaldi, and David Novotny. Lightplane: Highly-scalable components for neural 3d fields. arXiv, 2024. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.692, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[12] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In Proc. CVPR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[13] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[14] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In Proc. CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth \\(16 \\times 16\\) words: Transformers for image recognition at scale. In Proc. ICLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[16] Yasutaka Furukawa and Jean Ponce. Accurate, dense, and robust multiview stereopsis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(8):1362-1376, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[17] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using monte carlo rendering and denoising. In Advances in Neural Information Processing Systems, pages 22856-22869. Curran Associates, Inc., 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.278, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[18] Philipp Henzler, Jeremy Reizenstein, Patrick Labatut, Roman Shapovalov, Tobias Ritschel, Andrea Vedaldi, and David Novotny. Unsupervised learning of 3d object categories from videos in the wild. CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.334, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[19] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. In Proc. ICLR, 2024. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.391, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[20] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. GaussianShader: 3D Gaussian splattering with shading functions for reflective surfaces. arXiv.cs, abs/2311.17977, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.448, + 0.907, + 0.516 + ], + "angle": 0, + "content": "[21] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.518, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[22] Brian Karis and Epic Games. Real shading in unreal engine 4. Proc. Physically Based Shading Theory Practice, 4(3):1, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.561, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[23] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the Fourth Eurographics Symposium on Geometry Processing, page 61-70, Goslar, DEU, 2006. Eurographics Association. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.618, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[24] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian Splating for real-time radiance field rendering. Proc. SIGGRAPH, 42(4), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.905, + 0.7 + ], + "angle": 0, + "content": "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[26] Zhengfei Kuang, Yunzhi Zhang, Hong-Xing Yu, Samir Agarwala, Shangzhe Wu, and Jiajun Wu. Stanford-orb: A real-world 3d object inverse rendering benchmark, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.76, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[27] Ruofan Liang, Huiting Chen, Chunlin Li, Fan Chen, Selvakumar Panneer, and Nandita Vijaykumar. Envidr: Implicit differentiable renderer with neural environment lighting. In Proc. ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.817, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[28] Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. GS-IR: 3D Gaussian splattering for inverse rendering. arXiv.cs, abs/2311.16473, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[29] W. Lorensen and H. Cline. Marching cubes: A high resolution 3D surface construction algorithm. ACM Computer Graphocs, 21(24), 1987. 2" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5867" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[30] Alexander Mai, Dor Verbin, Falko Kuester, and Sara Fridovich-Keil. Neural microfacet fields for inverse rendering. In Proc. ICCV, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.189 + ], + "angle": 0, + "content": "[31] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Proc. ECCV, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[32] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P. Srinivasan, and Jonathan T. Barron. Nerf in the dark: High dynamic range view synthesis from noisy raw images. In Proc. CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[33] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. In Proc. SIGGRAPH, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[34] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting Triangular 3D Models, Materials, and Lighting From Images. In Proc. CVPR, 2022. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.385 + ], + "angle": 0, + "content": "[35] Giljoo Nam, J. Lee, D. Gutierrez, and Min H. Kim. Practical SVBRDF acquisition of 3D objects with unstructured flash photography. arXiv, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[36] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.524, + 0.482, + 0.551 + ], + "angle": 0, + "content": "[37] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proc. ICCV, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.552, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[38] Pakkapon Phongthawee, Worameth Chinchuthakun, Nontaphat Sinsunthithet, Varun Jampani, Amit Raj, Pramook Khungurn, and Supasorn Suwajanakorn. Diffusionlight: Light probes for free by painting a chrome ball. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 98-108, 2024. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.482, + 0.677 + ], + "angle": 0, + "content": "[39] PolyHaven. Polyhaven: curated public asset library for visual effects artists and game designers. https://polyhaven.com/, 2024. Accessed: 2024-11-07. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.678, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[40] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In International Conference on Computer Vision, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[41] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common Objects in 3D: Large-scale learning and evaluation of real-life 3D category reconstruction. In Proc. ICCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[42] Shen Sang and Manmohan Chandraker. Single-shot neural relighting and SVBRDF estimation. In Proc. ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[43] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proc. CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[44] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In Proc. ECCV, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[45] Yawar Siddiqui, Filippos Kokkinos, Tom Monnier, Mahendra Kariya, Yanir Kleiman, Emilien Garreau, Oran Gafni, Natalia Neverova, Andrea Vedaldi, Roman Shapovalov, and David Novotny. Meta 3D Asset Gen: Text-to-mesh generation with high-quality geometry, texture, and PBR materials. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2024. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[46] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Frédo Durand. Light field networks: Neural scene representations with single-evaluation rendering. In Proc. NeurIPS, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[47] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proc. CVPR, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.29, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[48] Cheng Sun, Guangyan Cai, Zhengqin Li, Kai Yan, Cheng Zhang, Carl Marshall, Jia-Bin Huang, Shuang Zhao, and Zhao Dong. Neural-pbir reconstruction of shape, material, and illumination. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 18000-18010, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.905, + 0.43 + ], + "angle": 0, + "content": "[49] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. LGM: Large multi-view Gaussian model for high-resolution 3D content creation. arXiv, 2402.05054, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.432, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[50] Giuseppe Vecchio and Valentin Deschaintre. Matsynth: A modern pbr materials dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22109-22118, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.905, + 0.543 + ], + "angle": 0, + "content": "[51] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd E. Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In Proc. CVPR, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.545, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[52] Guangcong Wang, Yinuo Yang, Chen Change Loy, and Zwei Liu. Stylelight: HDR panorama generation for lighting estimation and editing. In European Conference on Computer Vision (ECCV), 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.602, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[53] Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. VGGsF: visual geometry grounded deep structure from motion. In Proc. CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.645, + 0.905, + 0.7 + ], + "angle": 0, + "content": "[54] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv.cs, abs/2106.10689, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[55] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. on Image Processing, 13(4), 2004. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.758, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[56] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. MeshLRM: large reconstruction model for high-quality mesh. arXiv, 2404.12385, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[57] Desai Xie, Sai Bi, Zhixin Shu, Kai Zhang, Zexiang Xu, Yi Zhou, Soren Pirk, Arie Kaufman, Xin Sun, and Hao Tan. Lrm-zero: Training large reconstruction models with synthesized data, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[58] Desai Xie, Sai Bi, Zhixin Shu, Kai Zhang, Zexiang Xu, Yi Zhou, Soren Pirk, Arie Kaufman, Xin Sun, and Hao Tan." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5868" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "LRM-Zero: training large reconstruction models with synthesized data. arXiv, 2406.09371, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[59] Zhang Xiuming, Srinivasan Pratul P., Deng Boyang, Debevec Paul, Freeman William T., and Barron Jonathan T. NeRFactor: neural factorization of shape and reflectance under an unknown illumination. In Proc. SIGGRAPH, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[60] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. GRM: Large gaussian reconstruction model for efficient 3D reconstruction and generation. arXiv, 2403.14621, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[61] Yinghao Xu, Hao Tan, Fujun Luan, Sai Bi, Peng Wang, Jiahao Li, Zifan Shi, Kalyan Sunkavalli, Gordon Wetzstein, Zexiang Xu, and Kai Zhang. DMV3D: Denoising multiview diffusion using 3D large reconstruction model. In Proc. ICLR, 2024. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[62] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. arXiv.cs, abs/2106.12052, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.416 + ], + "angle": 0, + "content": "[63] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. PixelNeRF: Neural radiance fields from one or few images. In Proc. CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[64] Jesus Zarzar and Bernard Ghanem. Splitnerf: Split sum approximation neural field for joint geometry, illumination, and material estimation. arXiv.cs, abs/2311.16671, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[65] Jianzhao Zhang, Guojun Chen, Yue Dong, Jian Shi, Bob Zhang, and Enhua Wu. Deep inverse rendering for practical object appearance scan with uncalibrated illumination. In Proc. Advances in Computer Graphics, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.586 + ], + "angle": 0, + "content": "[66] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snively. PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.482, + 0.642 + ], + "angle": 0, + "content": "[67] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. GS-LRM: large reconstruction model for 3D Gaussian splatting. arXiv, 2404.19702, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.645, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[68] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proc. CVPR, pages 586-595, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.702, + 0.482, + 0.77 + ], + "angle": 0, + "content": "[69] Jingsen Zhu, Fujun Luan, Yuchi Huo, Zihao Lin, Zhihua Zhong, Dianbing Xi, Rui Wang, Hujun Bao, Jiaxiang Zheng, and Rui Tang. Learning-based inverse rendering of complex indoor scenes with differentiable monte carlo raytracing. In SIGGRAPH Asia 2022 Conference Papers. ACM, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.77 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5869" + } + ] +] \ No newline at end of file diff --git a/2025/Twinner_ Shining Light on Digital Twins in a Few Snaps/0293a039-5500-41cc-90fa-fe4f8710c6e9_origin.pdf b/2025/Twinner_ Shining Light on Digital Twins in a Few Snaps/0293a039-5500-41cc-90fa-fe4f8710c6e9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6b5e293ed9d19f600d204dd7e627cfc8b54af7ef --- /dev/null +++ b/2025/Twinner_ Shining Light on Digital Twins in a Few Snaps/0293a039-5500-41cc-90fa-fe4f8710c6e9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c0f79114efe90aa226233a70f9a4a96ad1592fdc86ea02fe1fa06cf3f0448ecb +size 3221948 diff --git a/2025/Twinner_ Shining Light on Digital Twins in a Few Snaps/full.md b/2025/Twinner_ Shining Light on Digital Twins in a Few Snaps/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9223f7eb14e3e0f76fff63da4a90a49ce08ce52b --- /dev/null +++ b/2025/Twinner_ Shining Light on Digital Twins in a Few Snaps/full.md @@ -0,0 +1,300 @@ +# Twinner: Shining Light on Digital Twins in a Few Caps + +![](images/25e121c60bbfe31469346c6950294ce48c3797587259d68d4fdf63fefa30c6e8.jpg) + +![](images/ac242e6501d8cbc0fe71db3f48326501294e99d57de24953fc5b1db06ad90e9a.jpg) +Figure 1. We present Twinner, the first large reconstruction model capable of predicting a scene's illumination as well as an object's geometry and material properties from a few posed images. This enables tasks such as realistic relighting of objects in novel scenes in a few seconds by replacing costly per-scene optimizations with a single forward pass. + +# Abstract + +We present Twinner, the first large reconstruction model capable of recovering a scene's illumination as well as an object's geometry and material properties from only a few posed images. Twinner is based on the Large Reconstruction Model and innovates in three key ways: 1) We introduce a memory-efficient voxel-grid transformer whose memory scales only quadratically with the size of the voxel grid. 2) To address the scarcity of high-quality ground-truth PBR-shaded models, we introduce a large, fully synthetic dataset of procedurally generated PBR-textured objects lit with varied illumination. 3) To bridge the synthetic-to-real gap, we finetune the model on real-world datasets using a differentiable physically based shading model, eliminating the need for ground-truth illumination or material properties, which are challenging to obtain in real-world scenarios. We demonstrate the efficacy of our model on the real-world StanfordORB benchmark, where, given a few input views, we achieve reconstruction quality significantly superior to + +existing feed-forward reconstruction networks and comparable to slower per-scene optimization methods. + +# 1. Introduction + +The creation of digital twins, i.e., the conversion of real-world objects into virtual 3D assets, is an important problem in virtual and augmented reality, gaming, conservation, education, advertisement, and general 3D content creation. In recent years, new techniques such as NeRF [31] and 3D Gaussian Splatting [24], along with their feed-forward counterparts [18, 49, 60, 63], have significantly improved the quality, speed, and generality of digital twinning. However, most of these techniques still bake illumination into the scene's appearance as radiance. As a consequence, the resulting assets cannot be relit and appear out of place when visualized under different lighting conditions. + +To address this limitation, one must also reconstruct the field of Bidirectional Reflectance Distribution Functions (BRDF), which describe the optical properties of materi- + +als. Estimating the BRDF directly is notoriously challenging due to its high dimensionality [3] and is usually done using 3D scanners that capture thousands of images while controlling for viewpoint and illumination [4-7, 35, 42, 65]. This is often impractical, so authors have considered "casual" reconstruction setups, where the object is captured by a conventional camera in uncontrolled conditions. Examples include NVDiffRec [34], NeRD [8], and NeRFactor [59], which simultaneously estimate the illumination, in the form of an environment map, and a simplified BRDF parameterized using Physically-Based Rendering (PBR) with albedo, roughness, and metallicity [10]. These quantities are recovered by minimizing the photometric reconstruction error, comparing the output of a differentiable PBR renderer to the input images. However, this optimization is slow, ill-posed, and requires the introduction of handcrafted priors for regularization. + +In this paper, we propose a new method, Twinner, a feedforward network that outputs shape, PBR textures, and environment illumination. This network is fast, performs well, requires only a small number of images as input, and learns empirically the necessary reconstruction priors. + +Siddiqui et al. [45] recently introduced a similar predictor, MetalLRM, as an extension of the LightplaneLRM [11]. MetalLRM requires supervision with a large dataset of 3D assets with high-quality PBR textures, which, unfortunately, are not abundant in existing synthetic datasets like Objaverse [14]. Moreover, even if a large synthetic dataset were available, the synthetic-to-real gap between the artist-created training assets and the target domain of real input images makes MetalLRM unsuitable for digital twinning. + +With our new Twinner, we address the scarcity of high-quality training PBR assets and the train/test domain shift. First, we ask whether real images in datasets like CO3Dv2 [40] can be used to supervise the model. Although these datasets lack PBR information, they are abundant and match the target distribution of real images we aim to digitally twin. To use this data, we propose minimizing the photometric error similar to optimization-based models, but with the key difference that the 3D shape, PBR materials, and environment light are all predicted by a large feedforward transformer [19, 45]. During training, each scene is rendered with a differentiable PBR shader that accepts the predicted environment light and the predicted PBR-textured shape. The latter then admits indirect PBR and illumination supervision via photometric losses between the renders and the corresponding real ground-truth images. + +Additionally, we consider a second approach to tackle data scarcity where, inspired by Zeroverse [58], we procedurally generate PBR-textured objects lit with a known set of environment maps. While the resulting objects are not realistic, their key advantage is that they come with clean ground-truth PBR and illumination information, which can + +be used to supervise the corresponding predictors directly instead of indirectly via the photometric loss. + +Besides addressing data scarcity, we also improve the architecture of the reconstruction model. We believe in the virtues of a simple and direct voxel-grid representation of the object [47], but note that predicting a number of tokens cubic in the voxel grid resolution is intractable with a transformer. Reducing the number of predicted tokens is the primary reason why recent models [19, 45] factor the voxel grid as a triplane. The question, then, is how to retain the benefits of the direct voxel grid representation and still make computations feasible. We solve this problem by introducing the novel tricolumn representation, where each feature vector (column) in a triplane is obtained by stacking the corresponding cells in the underlying voxel grid. In this manner, the number of tokens becomes quadratic in the voxel grid size, while each voxel is still represented explicitly. We further augment the network to predict a representation of the environment map, capturing illumination. + +We demonstrate the efficacy of the Twinner model on the StanfordORB benchmark [26], consisting of casual captures of real-life objects. We achieve reconstruction quality superior to existing feed-forward predictors and comparable to optimization-based methods, while using only a very small number of input images of the object and operating much faster. We also obtain state-of-the-art reconstruction of the environment maps, achieving a $28\%$ improvement in the predicted illumination's angular error over existing methods. + +# 2. Related work + +3D reconstruction. 3D reconstruction aims to recover a three-dimensional representation of a scene from a set of posed images. Traditionally, this was accomplished with structure-from-motion [1, 43, 53], multi-view stereo [16, 44], and mesh surface extraction methods [23, 29]. Recently, methods based on neural fields [2, 31-33, 51, 54, 62] have focused on extracting both shape and appearance by learning the scene radiance function. Through volumetric rendering and photometric losses, they are able to implicitly reconstruct a 3D density volume from which a mesh can be extracted. However, these methods infer the parameters of each scene via test-time optimization, making them slow. Authors have thus introduced new feed-forward reconstruction networks, generically called Large Reconstruction Models (LRMs) [19, 56, 58, 67]. LRMs use transformers to predict 3D representations from a small set of input images. Our Twinner predicts not just geometry and radiance, but also material properties and illumination. + +Material reconstruction. 3D reconstruction typically recovers a scene's geometry and radiance under fixed illumination. However, it is also necessary to recover material + +![](images/8fa14c1ea85692e07b84abf9d144bb320c70789d9224a53c44bdb0d95f85f84f.jpg) +Figure 2. Overview of Twinner. The input views are processed together with their foreground masks and camera poses by an image tokenizer. The resulting tokens are processed by a Diffusion Transformer (DiT) model to predict a 3D volumetric representation of the scene, and a second DiT model to predict the scene's illumination. The 3D representation is then rendered to obtain images of the scene's material properties, normals, and opacity from the target's point of view. Using these rendered images together with the predicted illumination, Twinner renders an approximate physically-shaded image of the scene which then enters a photometric loss. + +properties so that the 3D assets look realistic in novel virtual scenes. To do so, some works focus on optimizing explicit representations, such as material textures on meshes [17, 34, 48], while other works have added physically-based rendering models into neural fields [8, 9, 21, 27, 30, 64] and Gaussian splats [20, 28]. While these methods achieve high-quality material reconstructions, they require costly per-scene optimization. Some have addressed this issue by learning feed-forward reconstructors of PBR assets [45]. However, these are typically trained with synthetic data due to the complexity of capturing real-world material properties. We bridge the gap between synthetic and real data by training Twinner on real-world datasets. + +Illumination prediction. A key to accurate material reconstruction is accurate modeling of the scene's illumination. Single-scene optimization methods achieve this by optimizing an internal representation of illumination using light probe images [17, 34, 59], spherical harmonics [28], spherical Gaussians [21, 66], or MLPs [27, 30, 64]. These methods rely on physical models of lighting to aid the optimization. Others have attempted to predict the illumination of a scene from a single image [38, 52]. They rely on generative models to map the input view to a possible distribution of panorama views representing the environment illumination. Our Twinner lies in between these two lines of work: it is similar to the predictive approach as it predicts illumination from a set of four input views, but it uses physically-based rendering losses for supervision, similar to single-scene optimization methods. + +# 3. Methodology + +We introduce Twinner, a model that reconstructs the shape, materials, and illumination of real-life 3D objects from a small set of posed input views. Twinner extends Light- + +planeLRM [11], an enhanced version of LRM [19], by incorporating predictors for the object's material properties and scene illumination, as well as a differentiable physically-based renderer. In Sec. 3.1, we begin by reviewing the reconstructor and the PBR model we use. Then, in Secs. 3.2 to 3.4, we describe the main building blocks of Twinner. Finally, in Sec. 3.5, we detail the training protocol, including the losses used to supervise Twinner on real data. + +# 3.1. Preliminaries + +Physically-based rendering. To supervise our model with a photometric loss, it is necessary to render the reconstructed scene into an image. The color of each pixel corresponds to the radiance $\mathbf{L}_o$ emitted by the imaged object point $p\in \mathbb{R}^3$ in the direction of the camera. With knowledge of a scene's illumination, geometry, and material properties, it is possible to estimate the outgoing radiance at any such point according to the reflectance equation: + +$$ +\mathbf {L} _ {o} = \int_ {\Omega} \left(k _ {d} \frac {\mathbf {a}}{\pi} + \mathbf {f} _ {s}\right) \mathbf {L} _ {i} \left\langle \boldsymbol {\omega} _ {i}, \mathbf {n} \right\rangle d \boldsymbol {\omega} _ {i}. \tag {1} +$$ + +Here, $\mathbf{L}_i$ is the incoming radiance, representing the scene illumination falling on the imaged point $\pmb{p}$ . The integral is carried out over a hemisphere $\Omega$ of incoming directions $\omega_{i}$ determined by the surface normal $\mathbf{n}$ at $\pmb{p}$ . The factor in parentheses is a model of the object's Bidirectional Reflectance Distribution Function (BRDF), which describes how the incoming radiance reflects off the object's surface into various directions. The first term is the diffuse component, parameterized by the object's diffuse albedo $\mathbf{a}$ and a constant $k_{d}$ , both of which are independent of direction. In the integral, this term averages all incoming radiance over the hemisphere. The second term, $\mathbf{f}_s$ , is the specular com + +![](images/194392d4b0cbe5ee68de5ccc42997a79092f846e1d944a55ed3c5925e4abd483.jpg) +Figure 3. Tricolumn Architecture. The transformer module processes image tokens and a learnable embedding to predict the tricolumn representation, which consists of a grid of $3R^2$ feature vectors with dimensionality $\mathbb{R}^{CR}$ . These feature vectors are then split into three axis-aligned column grids, each with dimensionality $\mathbb{R}^{CR}$ . Finally, all axis-aligned column grids are reshaped into voxel grids of size $R\times R\times R$ with dimensionality $\mathbb{R}^C$ and summed into a single voxel grid. + +ponent of the BRDF, which captures directional effects such as reflections or highlights. Note that $\mathbf{L}_i$ is a function of the incoming direction $\omega_{i}$ , and the specular BRDF $\mathbf{f}_s$ is a function of both the incoming and outgoing directions $\omega_{i}$ and $\omega_{o}$ . Radiances and BRDFs also depend on the light frequency, which we model using three RGB components. Furthermore, all quantities depend on the point position $\pmb{p}$ . We omit these details for brevity. + +Due to the complexity of the integral in Eq. (1), numerous approximations have been proposed. We rely on the split-sum approximation [22] for image-based lighting, which leverages the scene's illumination represented as a spherical image enveloping the scene at an infinite radius. Following [34], we use a set of differentiable mipmaps to represent the pre-integrated lighting. The BRDF follows the Disney [22] microfacet BRDF model and is parameterized by the albedo $\mathbf{a}$ , metalness $m$ , and roughness $r$ at each point. We further assume that $\mathbf{L}_i$ is constant in space (hence we do not account for shadows cast by the object). + +In summary, to apply this model to render an image of the object, we require an estimate of the object's 3D shape and its albedo, metalness, and roughness at its surfaces. We also require a spherical image representing the directional function $\mathbf{L}_i$ , known as an environment map. + +Large Reconstruction Model (LRM). The aim of LRM [19] is to map a few posed views $\mathcal{I}$ of the scene to a 3D reconstruction $V$ . LRM first extracts features $y$ from the set of input views, utilizing a pre-trained DINO [36] Vision Transformer (ViT) [15]. Following [61], we feed into the ViT the RGB images $\mathcal{I}$ , the Plücker embeddings [46] $\mathcal{I}_{\mathrm{Pl}}$ representing the cameras, and the masks $\mathcal{M}$ of the foreground object to be reconstructed. + +The extracted feature tokens enter, via cross-attention, the transformer architecture from Diffusion Transformer (DiT) [37] $\Phi$ , which is tasked with reconstructing the scene. We emphasize that we borrow the DiT architecture without any diffusion process. The tokens in this transformer en + +code the 3D representation of the scene, realizing a mapping $V = \Phi(V_0|y)$ , where $V_0$ are the learnable initial tokens. A crucial design choice is how to construct the tokens $V$ manipulated by this transformer, which is discussed in Sec. 3.2. + +Finally, using the memory-efficient rendering algorithm proposed in Lightplane [11], a target set of views is rendered from the 3D representation. A shaded RGB image $\hat{\mathcal{I}}$ , a foreground mask image $\hat{\mathcal{M}}$ , and a depth image $\hat{\mathcal{I}}_{\mathrm{d}}$ are rendered for each view and used to supervise the LRM through a combination of the following photometric losses: + +$$ +\mathcal {L} _ {M} = \operatorname {B C E} (\mathcal {M}, \hat {\mathcal {M}}), \tag {2} +$$ + +$$ +\mathcal {L} _ {R} = \operatorname {L P I P S} \left(\mathcal {M} \mathcal {I}, \mathcal {M} \hat {\mathcal {I}}\right) + \left\| \mathcal {M} \mathcal {I} - \mathcal {M} \hat {\mathcal {I}} \right\| ^ {2}, \tag {3} +$$ + +$$ +\mathcal {L} _ {d} = \left\| \mathcal {M I} _ {\mathrm {d}} - \mathcal {M} \hat {\mathcal {I}} _ {\mathrm {d}} \right\|, \tag {4} +$$ + +where $\mathcal{I}_{\mathrm{d}}$ is the set of ground-truth depth maps corresponding to the input images $\mathcal{I}$ , and $\mathcal{M}\mathcal{I}$ denotes a masked image. + +# 3.2. Tricolumn representation + +Crucial to the Twinner design is the representation of the 3D scene, which is encoded as a set of volumetric functions expressing opacity and material properties at every point $\pmb{p} \in \mathbb{R}^3$ in space. A simple and effective option [47] to encode a volumetric map $V: \mathbb{R}^3 \to \mathbb{R}^C$ is to use a voxel grid $V \in \mathbb{R}^{C \times D \times H \times W}$ . Here, the value $V(\pmb{p}) \in \mathbb{R}^C$ interpolates the voxel values at the nearest grid points. To predict the voxel grid $V$ with the transformer of Sec. 3.1, we could assign one token to each voxel. However, the number $WHD$ of voxels grows cubically with the resolution of the grid, which leads to prohibitive memory demands. + +The alternative adopted by models like [19, 45] is to use a triplane representation [12]. In this case, the volumetric function is represented by three 2D grids $V_{xy} \in \mathbb{R}^{K \times H \times W}$ , $V_{yz} \in \mathbb{R}^{K \times W \times D}$ , and $V_{zx} \in \mathbb{R}^{K \times D \times H}$ . The volumetric function at $\pmb{p} = (x,y,z)$ is then expressed as + +$$ +V (x, y, z) = f \left(V _ {x y} (x, y), V _ {y z} (y, z), V _ {z x} (z, x)\right), \tag {5} +$$ + +![](images/711809666565476ddb2350ffe020e6716d6988f6aa55345e373cc0d9d6505e46.jpg) +Figure 4. Examples from our procedural dataset visualizing the shaded image, environment map, materials, and normals from our environment-map-endowed version of Zeroverse [57]. + +where $f: \mathbb{R}^{K \times K \times K} \to \mathbb{R}^C$ is a point-wise mixing function (e.g., an MLP applied to the concatenation of the three $K$ -dimensional feature vectors). Paired with a transformer, this conveniently reduces the number of tokens to $WH + HD + DW$ only. However, triplane mixes information between different voxels, which can lead to artifacts. + +We thus propose a novel tricolumn representation instead. This is similar to triplane as it uses a plane instead of a volume of tokens; however, each feature vector is in itself a column of corresponding voxel cells, so that voxel-specific information is represented directly and separately. + +In detail, we consider a plane of feature vectors $V_{xy} \in \mathbb{R}^{(CR) \times R \times R}$ . Because this is a plane, there are only $R^2$ tokens, which are easily handled by the transformer. However, $V_{xy}$ can be reshaped into a $C \times R \times R \times R$ voxel grid by unflattening the first dimension. We use the symbol $V_{xy}(x,y;z) \in \mathbb{R}^C$ to denote the feature vector obtained by interpolating this grid in 3D. + +When three such planes are used, the full feature vector at $\pmb{p}$ is decoded as: + +$$ +V (x, y, z) = f \left(V _ {x y} (x, y; z), V _ {y z} (y, z; x), V _ {z x} (z, x; y)\right). \tag {6} +$$ + +Note that this construction is strictly different from the triplane equation (5) as each term depends on all three spatial coordinates. We opt for using summation followed by an MLP as the mixing function $f$ . + +# 3.3. PBR reconstruction model + +The LRM models of [19, 61] only recover the shape and radiance of the scene. Like in [45], we extend Twinner to also recover the material properties. We do so by extending the tricolumn representation to encode the albedo $a$ (instead of radiance) and two additional channels representing the scene's metalness $m$ and roughness $r$ . We also note the importance of accurate normals for relighting. While a normal image could be directly estimated by projecting the rendered depth image into 3D space and taking the gradient + +of the opacity field, $\bar{\mathcal{I}}_{\mathrm{n}} = \nabla \sigma (\pi^{-1}(\hat{\mathcal{I}}_{\mathrm{d}}))$ , this estimate tends to be noisy. Therefore, we follow previous works [51] and predict three additional channels representing the scene's normals. The predicted normal image $\hat{\mathcal{I}}_{\mathrm{n}}$ is then supervised with the pseudo-ground-truth normals from the opacity field $\bar{\mathcal{I}}_{\mathrm{n}}$ via: + +$$ +\mathcal {L} _ {N} = 1 - \langle \mathcal {M} \bar {\mathcal {I}} _ {\mathrm {n}}, \mathcal {M} \hat {\mathcal {I}} _ {\mathrm {n}} \rangle . \tag {7} +$$ + +Direct supervision using synthetic data. Reconstruction models that predict radiance can be supervised by rendering the predicted 3D scene into images and comparing these to training views using a photometric loss. In our case, however, the model does not predict radiance, but albedo, roughness, and metalness, which cannot be rendered into a 'shaded' image unless illumination is also known. + +A simple approach for supervising the model that does not require knowledge of illumination is to render albedo, roughness, and metalness images $\hat{\mathcal{I}}_{\mathrm{a}},\hat{\mathcal{I}}_{\mathrm{r}}$ ,and $\hat{\mathcal{I}}_{\mathrm{m}}$ , and compare these to their corresponding ground-truth renders. We use this approach to supervise our model with synthetic data by applying the following loss on the set of material renders $\hat{\mathcal{I}}_{\mathrm{a}},\hat{\mathcal{I}}_{\mathrm{m}},\hat{\mathcal{I}}_{\mathrm{r}}$ , and also on the normal render $\hat{\mathcal{I}}_{\mathrm{n}}$ .. + +$$ +\mathcal {L} _ {S} = \ell_ {\mathcal {M}} \left(\mathcal {I} _ {\mathrm {a}}, \hat {\mathcal {I}} _ {\mathrm {a}}\right) + \ell_ {\mathcal {M}} \left(\mathcal {I} _ {\mathrm {r}}, \hat {\mathcal {I}} _ {\mathrm {r}}\right) + \ell_ {\mathcal {M}} \left(\mathcal {I} _ {\mathrm {m}}, \hat {\mathcal {I}} _ {\mathrm {m}}\right) + \ell_ {\mathcal {M}} \left(\mathcal {I} _ {\mathrm {n}}, \hat {\mathcal {I}} _ {\mathrm {n}}\right) \tag {8} +$$ + +where $\ell_{\mathcal{M}}(u,v) = \| \mathcal{M}(u - v)\|_{\epsilon}$ is the Huber loss $\| \cdot \|_{\epsilon}$ with cut-off threshold $\epsilon$ masked by $\mathcal{M}$ . + +# 3.4. Training on real data by predicting illumination + +The learning formulation of Sec. 3.3 requires ground-truth material properties to be known, which is not the case for real-world data where only shaded images are available. In order to fine-tune our model using real data with a photometric loss, it is necessary to also estimate the scene illumination. With this, we can use the differentiable physically-based shader detailed in Sec. 3.1 to render the shaded scene appearance and compare it to the input images through a photometric loss. + +To estimate the scene's illumination, we propose using an additional DiT module. The latter takes as input the feature tokens $y$ computed by the image feature extractor and predicts a cubemap representation of the scene's lighting. Recall that a cubemap is a function $C \to \mathbb{R}^3$ that assigns every point $\pmb{p}$ of the cube surface $C = \{\pmb{p} : \| \pmb{p} \|_{\infty} = 1\}$ to the radiance $\mathbf{L}_i(\boldsymbol{\omega}_i)$ incoming from direction $\boldsymbol{\omega}_i \propto -\pmb{p}$ , and is thus a representation of the environment illumination (this works because we assume that the function $\mathbf{L}_i$ only depends on direction, but not on position). The cubemap is unfolded to a collection of six 2D images, which are directly predicted by the DiT module. + +Supervising illumination with synthetic data. In synthetic data, the cubemap is known, and its predictor can be supervised directly with the environment-map loss: + +$$ +\mathcal {L} _ {\text {e n v}} = \left\| \mathcal {I} _ {\mathrm {e}} - \hat {\mathcal {I}} _ {\mathrm {e}} \right\| ^ {2} + \operatorname {L P I P S} \left(\mathcal {I} _ {\mathrm {e}}, \hat {\mathcal {I}} _ {\mathrm {e}}\right), \tag {9} +$$ + +
Novel Scene RelightingNovel View SynthesisGeometry EstimationRuntime
MethodPSNR-H↑PSNR-L↑SSIM↑LPIPS↓PSNR-H↑PSNR-L↑SSIM↑LPIPS↓Depth↓Geometry↓Shape↓Time (hr.)↓
Single-Scene Opt.PhySG [66]21.8128.110.9600.05524.2432.150.9740.0471.900.179.28~3
NVDiffRec [34]22.9129.720.9630.03921.9428.440.9690.0300.310.060.62~1
NeRD [8]23.2929.650.9570.05925.8332.610.9630.0541.390.2813.70>20
NeRFactor [59]23.5430.380.9690.04826.0633.470.9730.0460.870.299.53>20
InvRender [69]23.7630.830.9700.04625.9134.010.9770.0420.590.060.44~3
NVDiffRecMC [17]24.4331.600.9720.03628.0336.400.9820.0280.320.040.51~2
Neural-PBIR [48]26.0133.260.9790.02328.8236.800.9860.0190.300.060.43~1
PredictiveMetaLRM [45]21.5028.060.9570.05219.5525.980.9560.0518.6690.061.260.004
w/o ZeroVerse23.5330.310.9650.04821.5328.780.9650.0481.1260.090.740.006
w/o Shading Loss23.8930.850.9670.04221.4229.470.9670.0410.7620.060.300.004
Twinner (ours)24.5531.610.9690.03821.7029.620.9680.0370.7090.060.370.006
+ +Table 1. Relighting evaluation on StanfordORB [26] comparing our method to other feedforward predictors of PBR-textured assets (two ablations and MetalLRM). For completeness, we also include single-scene optimization methods in the top part of the table. Our model outperforms most single-scene optimization baseline methods in terms of novel scene relighting and shape estimation despite being a predictive model and having access to only four input views. More impressively, it only requires a few seconds to reconstruct an object's mesh with PBR materials, while single-scene optimization methods have runtimes in the order of hours. + +comprising the photometric loss from (8) using the predicted cubemap projected into an equirectangular image $\hat{\mathcal{I}}_{\mathrm{e}}$ and the ground-truth equirectangular image $\mathcal{I}_{\mathrm{e}}$ . In practice, we use a procedurally generated dataset of synthetic assets detailed in Sec. 3.5 to provide ground-truth illumination. + +Supervising illumination with real data. In real data, the cubemap is observed indirectly via the effect it has on the object's appearance according to the reflectance equation (1), which provides only weak supervision. We note, however, that the scene background visible in the input images is, to a first approximation, equal to the function $\mathbf{L}_i$ , and can thus provide direct supervision of the latter. We can render the background observed from any given camera by assigning to each pixel the color $\mathbf{L}_i(\boldsymbol{\omega}_i) = C(\boldsymbol{\omega}_i / \| \boldsymbol{\omega}_i \|_{\infty})$ of the incoming radiance from the direction $\boldsymbol{\omega}_i$ of the pixel's ray. Since input images are in sRGB color space, the predicted background image is first tone-mapped and then composited with the foreground object prediction to obtain a scene image. The latter allows us to supervise both the illumination prediction and the shaded object prediction through the photometric loss $\mathcal{L}_R$ of Eq. (3). Here, as training data, we use abundant object-centric video captures from CO3Dv2 [41] as described in the next section. + +# 3.5. Training protocol + +Twinner's training minimizes the total loss + +$$ +\mathcal {L} _ {\text {t o t}} = \mathcal {L} _ {N} + \mathcal {L} _ {S} + \mathcal {L} _ {\mathrm {T V}} + \mathcal {L} _ {d} + \mathcal {L} _ {M} + \mathcal {L} _ {R} + \mathcal {L} _ {\mathrm {T V}}, \tag {10} +$$ + +where $\mathcal{L}_{\mathrm{TV}} = \mathrm{TV}(\hat{\mathcal{I}}_{\mathrm{n}}) + \mathrm{TV}(\hat{\mathcal{I}}_{\mathrm{a}}) + \mathrm{TV}(\hat{\mathcal{I}}_{\mathrm{d}})$ is the Total-Variation (TV) regularization on normal, albedo, and depth images promoting smooth solutions. $\mathcal{L}_{\mathrm{tot}}$ is optimized with Adam [25] using a learning rate of $10^{-4}$ until convergence. As mentioned before, individual terms of $\mathcal{L}_{\mathrm{tot}}$ are switched on or off depending on whether a real or a synthetic training example is present in the batch of training images. + +Training data. Following [45], Twinner is trained on a large dataset of 140k artist-created assets similar to Objaverse. However, this dataset often lacks high-quality material maps where, for instance, distinguishing between missing material images and material images with zero values is impossible. Moreover, there is a large domain gap between the synthetic objects and the real ones we intend to twin. + +As such, together with the artist-created assets, we also train on two additional data sources: 1) A large dataset of procedurally generated assets with clean PBR textures and environment maps; 2) Real-object captures from CO3Dv2 [40]. Both datasets are detailed next. + +Endowing Zeroverse with environment maps. Following LRM-Zero [57], we generate a set of 10k randomized 3D objects with realistic material properties sampled from the MatSynth [50] dataset. Additionally, to enable the environment-map supervision loss $\mathcal{L}_{\mathrm{env}}$ (9), for each rendered scene, we randomly sample one of $\sim 1\mathrm{k}$ environment maps from PolyHaven [39]. We render color, normal, material, depth, and foreground mask images for each object from a predetermined set of viewpoints (see examples in Fig. 4). As such, samples from this dataset admit optimizing all terms of the total loss $\mathcal{L}_{\mathrm{tot}}$ , including the environment loss $\mathcal{L}_{\mathrm{env}}$ , which is crucial for stable training of Twinner. + +Exploiting real object captures. We also train on CO3Dv2 [40], which is a large-scale real-world dataset of "turntable" videos capturing objects of common categories. Each frame in the dataset comes with camera and foreground mask annotations. As such, for each CO3Dv2 sample, we can optimize the normal, mask, photometric, and TV losses $\mathcal{L}_{\mathrm{tot}}^{\mathrm{real}} = \mathcal{L}_N + \mathcal{L}_M + \mathcal{L}_R + \mathcal{L}_{\mathrm{TV}}$ . Since $\mathcal{L}_{\mathrm{tot}}^{\mathrm{real}}$ is a function of the illumination image $\hat{\mathcal{I}}_{\mathrm{e}}$ and PBR renders $\hat{\mathcal{I}}_{\mathrm{a}}$ , $\hat{\mathcal{I}}_{\mathrm{m}}$ , $\hat{\mathcal{I}}_{\mathrm{r}}$ , the loss provides a valuable supervisory signal for all volumetric predictors. Furthermore, using real training data narrows the domain gap between the training set and + +the target domain of Twinner, which are real images. + +![](images/ec178c20f5c817dfc35e49cecd77609ad75198bea25d62d559faa2112599a80a.jpg) +Figure 5. StanfordORB Illumination Visualization. We visualize the illumination estimation of multiple baselines and our method on four different scenes from the StanfordORB dataset. + +
MethodAngular Error ↓RMSE↓Norm. RMSE↓SI RMSE↓Time (hr.)↓
NMF [30]11.240.4870.5240.394~ 2
TensoIR [21]15.080.7520.7980.451~ 2
NVDiffRec [34]12.260.7230.7150.367~ 1
NVDiffRecMC [17]13.250.5740.6140.527~ 2
StyleLight [52]13.050.6400.6890.4400.011
DiffusionLight [38]11.350.5810.5690.5440.317
Twinner (ours)8.090.5700.6020.4720.006
+ +Table 2. Illumination prediction results. We provide a quantitative evaluation of the illumination estimation for baselines and our Twinner. Rows are grouped by single-scene optimization models (top) and predictive feed-forward models (bottom). + +# 4. Experiments + +Evaluation data. We evaluate Twinner on the StanfordORB [26] dataset, which comprises 42 captures of 14 different objects, each imaged under 3 different illuminations. + +Each object has a high-quality geometric reconstruction obtained using a 3D scanner and comes with ground-truth camera poses and estimated illumination. + +Evaluation Protocol. For each evaluation scene, we select four views from the set of training images as input frames via furthest camera-origin point sampling. With these input views, Twinner reconstructs a 3D PBR-textured mesh. Each predicted mesh is rendered with Blender [13] to held-out test views while being lit with a given ground-truth environment map. Following [26], the renders are compared to the corresponding ground-truth images using scale-invariant versions of LPIPS [68], SSIM [55], and PSNR. + +Baselines. We compare to existing feed-forward predictors of PBR-textured meshes. As such, we selected the state-of-the-art MetalLRM few-view reconstructor from AssetGen [45] as our main baseline. We use the version of MetalLRM conditioned on four shaded input views, as opposed to the albedo+shaded conditioning used in the AssetGen pipeline. We also compare to two ablations of our method. The first ablation, Twinner w/o Shading Loss, evaluates the contribution of training on real data and omits + +![](images/61b82d5a777513590a3905c372be15ff8005d05268b97e00ce561d11ae6aaf65.jpg) +Figure 6. Qualitative results on StanfordORB. We visualize the shaded RGB, geometry reconstruction, and material predictions for our baseline, two ablations, and our Twinner. Our model is able to better predict relightable objects thanks to the combination of curated synthetic training data and the shading loss on real data. + +the training images of CO3Dv2. Second, we compare to Twinner w/o Zeroverse, which ablates training of the environment-map and PBR-texture predictor on our version of the procedurally-created Zeroverse. For completeness, we also include several single-scene optimization methods that "overfit" a relightable implicit shape in each scene from scratch. Note that, to save GPU memory, our feed-forward predictors randomly sample only four source views from the pool of all available training images of each scene. This puts us at a disadvantage compared to single-scene optimizers, as they leverage all available scene views for each reconstruction. + +Relighting Evaluation. All quantitative results are reported in Tab. 1. Twinner is the best among feed-forward reconstructors, outperforming its own two ablations and + +![](images/256b653bc1c34a5fbf1ec2bcd5798f5e2d08f59a8e7b5a26ec5e491d98af3cf7.jpg) +Figure 7. CO3Dv2 Qualitative Results. We visualize predicted PBR meshes from the CO3Dv2 dataset illuminated under direct sunlight as well as their geometry, albedo, and material properties. The shading loss aids at predicting more realistic material properties, while the synthetic data improves the predicted geometry and albedo. + +MetalLRM. For instance, Twinner achieves a $27\%$ reduction in both novel-scene relighting and novel-view synthesis LPIPS compared to the strongest baseline (MetalLRM), while maintaining comparable inference time. Remarkably, our method is on par with most single-scene optimization methods while being significantly faster, requiring only a few seconds per prediction compared to several hours for optimization-based methods. Examples of our predictions and the baselines are shown in Fig. 6. We also evaluate qualitatively on testing sequences of CO3Dv2 in Fig. 7. + +Illumination Evaluation. We evaluate Twinner's ability to predict a scene's illumination from four input images. Since there are no existing baselines for few-view conditioned feed-forward illumination prediction, we compare to environment map estimates output by the single-scene optimizers in Tab. 1, as well as to single-view conditioned predictors. We follow [38] and evaluate illumination in terms of angular error, RMSE, normalized RMSE, and scale-invariant RMSE. Results are given in Tab. 2 and Fig. 5. + +Tricolumn Ablation. We also evaluate the effect of our tricolumn representation. In Tab. 3, we report the novel-view synthesis performance of the vanilla LightplaneLRM [11] model trained only on our synthetic dataset, utilizing either tricolumn or triplane representations. Tricolumn outperforms triplane significantly, achieving a $22\%$ improvement in LPIPS, validating this contribution. + +
3D representationPSNR↑SSIM↑LPIPS↓
Triplane31.200.9720.037
Tricolumn (ours)32.810.9660.029
+ +Table 3. Tricolumn ablation evaluating novel view synthesis on the StanfordORB dataset for a LightplaneLRM [11] using the triplane 3D representation or our proposed tricolumn instead. Our tricolumn outperforms triplane on all reconstruction metrics. + +# 5. Conclusion + +We have presented Twinner, a new LRM capable of creating high-quality 3D assets with accurate geometry, material properties, and illumination from only a few posed images in a feed-forward manner. Twinner leverages the novel tricolumn representation, whose memory efficiency enables the processing of voxel grids with large transformers. For training, we demonstrated the power of generating synthetic PBR assets procedurally, addressing the scarcity of high-quality PBR assets in existing synthetic datasets. We also showed that, by predicting illumination, the model can learn from real-life datasets without requiring groundtruth illumination and material properties. We evaluated Twinner on the StanfordORB benchmark, where it outperformed existing feed-forward predictors and achieved comparable quality to optimization-based methods while being orders of magnitude faster. + +# References + +[1] Sameer Agarwal, Noah Snavely, Ian Simon, Steven M. Seitz, and Richard Szeliski. Building rome in a day. In Proc. ICCV, 2009. 2 +[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In Proc. CVPR, 2022. 2 +[3] F O Bartell, E. L. Dereniak, and W. L Wolfe. The Theory And Measurement Of Bidirectional Reflectance Distribution Function (Brdf) And Bidirectional Transmittance Distribution Function (BTDF). In Proc. Radiation Scattering in Optical Systems, 1981. 2 +[4] Sai Bi, Zexiang Xu, Pratul P. Srinivasan, B. Mildenhall, Kalyan Sunkavalli, Milovs Havsan, Yannick Hold-Geoffroy, D. Kriegman, and R. Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv, 2020. 2 +[5] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Deep reflectance volumes: Relightable reconstructions from multi-view photometric images. In Proc. ECCV, 2020. +[6] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, David Kriegman, and Ravi Ramamoorthi. Deep 3D capture: Geometry and reflectance from sparse multi-view images. In Proc. CVPR, 2020. +[7] Mark Boss, Varun Jampani, Kihwan Kim, Hendrik P.A. Lensch, and Jan Kautz. Two-shot spatially-varying BRDF and shape estimation. In Proc. CVPR, 2020. 2 +[8] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. Nerd: Neural reflectance decomposition from image collections. In IEEE International Conference on Computer Vision (ICCV), 2021. 2, 3, 6 +[9] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan T. Barron, and Hendrik P.A. Lensch. Neuralpil: Neural pre-integrated lighting for reflectance decomposition. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3 +[10] Brent Burley. Physically-based shading at disney. Technical report, Disney, 2012. 2 +[11] Ang Cao, Justin Johnson, Andrea Vedaldi, and David Novotny. Lightplane: Highly-scalable components for neural 3d fields. arXiv, 2024. 2, 3, 4, 8 +[12] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In Proc. CVPR, 2022. 4 +[13] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 7 +[14] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In Proc. CVPR, 2023. 2 +[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, + +Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth $16 \times 16$ words: Transformers for image recognition at scale. In Proc. ICLR, 2021. 4 +[16] Yasutaka Furukawa and Jean Ponce. Accurate, dense, and robust multiview stereopsis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(8):1362-1376, 2010. 2 +[17] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using monte carlo rendering and denoising. In Advances in Neural Information Processing Systems, pages 22856-22869. Curran Associates, Inc., 2022. 3, 6, 7 +[18] Philipp Henzler, Jeremy Reizenstein, Patrick Labatut, Roman Shapovalov, Tobias Ritschel, Andrea Vedaldi, and David Novotny. Unsupervised learning of 3d object categories from videos in the wild. CVPR, 2021. 1 +[19] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. In Proc. ICLR, 2024. 2, 3, 4, 5 +[20] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. GaussianShader: 3D Gaussian splattering with shading functions for reflective surfaces. arXiv.cs, abs/2311.17977, 2023. 3 +[21] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 7 +[22] Brian Karis and Epic Games. Real shading in unreal engine 4. Proc. Physically Based Shading Theory Practice, 4(3):1, 2013. 4 +[23] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the Fourth Eurographics Symposium on Geometry Processing, page 61-70, Goslar, DEU, 2006. Eurographics Association. 2 +[24] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian Splating for real-time radiance field rendering. Proc. SIGGRAPH, 42(4), 2023. 1 +[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[26] Zhengfei Kuang, Yunzhi Zhang, Hong-Xing Yu, Samir Agarwala, Shangzhe Wu, and Jiajun Wu. Stanford-orb: A real-world 3d object inverse rendering benchmark, 2023. 2, 6, 7 +[27] Ruofan Liang, Huiting Chen, Chunlin Li, Fan Chen, Selvakumar Panneer, and Nandita Vijaykumar. Envidr: Implicit differentiable renderer with neural environment lighting. In Proc. ICCV, 2023. 3 +[28] Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. GS-IR: 3D Gaussian splattering for inverse rendering. arXiv.cs, abs/2311.16473, 2023. 3 +[29] W. Lorensen and H. Cline. Marching cubes: A high resolution 3D surface construction algorithm. ACM Computer Graphocs, 21(24), 1987. 2 + +[30] Alexander Mai, Dor Verbin, Falko Kuester, and Sara Fridovich-Keil. Neural microfacet fields for inverse rendering. In Proc. ICCV, 2023. 3, 7 +[31] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Proc. ECCV, 2020. 1, 2 +[32] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P. Srinivasan, and Jonathan T. Barron. Nerf in the dark: High dynamic range view synthesis from noisy raw images. In Proc. CVPR, 2022. +[33] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. In Proc. SIGGRAPH, 2022. 2 +[34] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting Triangular 3D Models, Materials, and Lighting From Images. In Proc. CVPR, 2022. 2, 3, 4, 6, 7 +[35] Giljoo Nam, J. Lee, D. Gutierrez, and Min H. Kim. Practical SVBRDF acquisition of 3D objects with unstructured flash photography. arXiv, 2018. 2 +[36] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 4 +[37] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proc. ICCV, 2023. 4 +[38] Pakkapon Phongthawee, Worameth Chinchuthakun, Nontaphat Sinsunthithet, Varun Jampani, Amit Raj, Pramook Khungurn, and Supasorn Suwajanakorn. Diffusionlight: Light probes for free by painting a chrome ball. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 98-108, 2024. 3, 7, 8 +[39] PolyHaven. Polyhaven: curated public asset library for visual effects artists and game designers. https://polyhaven.com/, 2024. Accessed: 2024-11-07. 6 +[40] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In International Conference on Computer Vision, 2021. 2, 6 +[41] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common Objects in 3D: Large-scale learning and evaluation of real-life 3D category reconstruction. In Proc. ICCV, 2021. 6 +[42] Shen Sang and Manmohan Chandraker. Single-shot neural relighting and SVBRDF estimation. In Proc. ECCV, 2020. 2 +[43] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proc. CVPR, 2016. 2 +[44] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In Proc. ECCV, 2016. 2 + +[45] Yawar Siddiqui, Filippos Kokkinos, Tom Monnier, Mahendra Kariya, Yanir Kleiman, Emilien Garreau, Oran Gafni, Natalia Neverova, Andrea Vedaldi, Roman Shapovalov, and David Novotny. Meta 3D Asset Gen: Text-to-mesh generation with high-quality geometry, texture, and PBR materials. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2024. 2, 3, 4, 5, 6, 7 +[46] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Frédo Durand. Light field networks: Neural scene representations with single-evaluation rendering. In Proc. NeurIPS, 2021. 4 +[47] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proc. CVPR, 2022. 2, 4 +[48] Cheng Sun, Guangyan Cai, Zhengqin Li, Kai Yan, Cheng Zhang, Carl Marshall, Jia-Bin Huang, Shuang Zhao, and Zhao Dong. Neural-pbir reconstruction of shape, material, and illumination. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 18000-18010, 2023. 3, 6 +[49] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. LGM: Large multi-view Gaussian model for high-resolution 3D content creation. arXiv, 2402.05054, 2024. 1 +[50] Giuseppe Vecchio and Valentin Deschaintre. Matsynth: A modern pbr materials dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22109-22118, 2024. 6 +[51] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd E. Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In Proc. CVPR, 2022. 2, 5 +[52] Guangcong Wang, Yinuo Yang, Chen Change Loy, and Zwei Liu. Stylelight: HDR panorama generation for lighting estimation and editing. In European Conference on Computer Vision (ECCV), 2022. 3, 7 +[53] Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. VGGsF: visual geometry grounded deep structure from motion. In Proc. CVPR, 2024. 2 +[54] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv.cs, abs/2106.10689, 2021. 2 +[55] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. on Image Processing, 13(4), 2004. 7 +[56] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. MeshLRM: large reconstruction model for high-quality mesh. arXiv, 2404.12385, 2024. 2 +[57] Desai Xie, Sai Bi, Zhixin Shu, Kai Zhang, Zexiang Xu, Yi Zhou, Soren Pirk, Arie Kaufman, Xin Sun, and Hao Tan. Lrm-zero: Training large reconstruction models with synthesized data, 2024. 5, 6 +[58] Desai Xie, Sai Bi, Zhixin Shu, Kai Zhang, Zexiang Xu, Yi Zhou, Soren Pirk, Arie Kaufman, Xin Sun, and Hao Tan. + +LRM-Zero: training large reconstruction models with synthesized data. arXiv, 2406.09371, 2024. 2 +[59] Zhang Xiuming, Srinivasan Pratul P., Deng Boyang, Debevec Paul, Freeman William T., and Barron Jonathan T. NeRFactor: neural factorization of shape and reflectance under an unknown illumination. In Proc. SIGGRAPH, 2021. 2, 3, 6 +[60] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. GRM: Large gaussian reconstruction model for efficient 3D reconstruction and generation. arXiv, 2403.14621, 2024. 1 +[61] Yinghao Xu, Hao Tan, Fujun Luan, Sai Bi, Peng Wang, Jiahao Li, Zifan Shi, Kalyan Sunkavalli, Gordon Wetzstein, Zexiang Xu, and Kai Zhang. DMV3D: Denoising multiview diffusion using 3D large reconstruction model. In Proc. ICLR, 2024. 4, 5 +[62] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. arXiv.cs, abs/2106.12052, 2021. 2 +[63] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. PixelNeRF: Neural radiance fields from one or few images. In Proc. CVPR, 2021. 1 +[64] Jesus Zarzar and Bernard Ghanem. Splitnerf: Split sum approximation neural field for joint geometry, illumination, and material estimation. arXiv.cs, abs/2311.16671, 2023. 3 +[65] Jianzhao Zhang, Guojun Chen, Yue Dong, Jian Shi, Bob Zhang, and Enhua Wu. Deep inverse rendering for practical object appearance scan with uncalibrated illumination. In Proc. Advances in Computer Graphics, 2020. 2 +[66] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snively. PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 6 +[67] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. GS-LRM: large reconstruction model for 3D Gaussian splatting. arXiv, 2404.19702, 2024. 2 +[68] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proc. CVPR, pages 586-595, 2018. 7 +[69] Jingsen Zhu, Fujun Luan, Yuchi Huo, Zihao Lin, Zhihua Zhong, Dianbing Xi, Rui Wang, Hujun Bao, Jiaxiang Zheng, and Rui Tang. Learning-based inverse rendering of complex indoor scenes with differentiable monte carlo raytracing. In SIGGRAPH Asia 2022 Conference Papers. ACM, 2022. 6 \ No newline at end of file diff --git a/2025/Twinner_ Shining Light on Digital Twins in a Few Snaps/images.zip b/2025/Twinner_ Shining Light on Digital Twins in a Few Snaps/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d6932872dbd1f7989af9673002b40535e7338d33 --- /dev/null +++ b/2025/Twinner_ Shining Light on Digital Twins in a Few Snaps/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f7ee147be2fc6818bd7a13f9fb544df92c1672ac1fb150db883262e66fea902 +size 795022 diff --git a/2025/Twinner_ Shining Light on Digital Twins in a Few Snaps/layout.json b/2025/Twinner_ Shining Light on Digital Twins in a Few Snaps/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..75a391733e5f58004d2fd1b45cf703ed54ab380c --- /dev/null +++ b/2025/Twinner_ Shining Light on Digital Twins in a Few Snaps/layout.json @@ -0,0 +1,8772 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 132, + 103, + 479, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 103, + 479, + 121 + ], + "spans": [ + { + "bbox": [ + 132, + 103, + 479, + 121 + ], + "type": "text", + "content": "Twinner: Shining Light on Digital Twins in a Few Caps" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 154, + 141, + 455, + 206 + ], + "blocks": [ + { + "bbox": [ + 154, + 141, + 455, + 206 + ], + "lines": [ + { + "bbox": [ + 154, + 141, + 455, + 206 + ], + "spans": [ + { + "bbox": [ + 154, + 141, + 455, + 206 + ], + "type": "image", + "image_path": "25e121c60bbfe31469346c6950294ce48c3797587259d68d4fdf63fefa30c6e8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 56, + 230, + 555, + 407 + ], + "blocks": [ + { + "bbox": [ + 56, + 230, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 56, + 230, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 555, + 407 + ], + "type": "image", + "image_path": "ac242e6501d8cbc0fe71db3f48326501294e99d57de24953fc5b1db06ad90e9a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 415, + 555, + 450 + ], + "lines": [ + { + "bbox": [ + 54, + 415, + 555, + 450 + ], + "spans": [ + { + "bbox": [ + 54, + 415, + 555, + 450 + ], + "type": "text", + "content": "Figure 1. We present Twinner, the first large reconstruction model capable of predicting a scene's illumination as well as an object's geometry and material properties from a few posed images. This enables tasks such as realistic relighting of objects in novel scenes in a few seconds by replacing costly per-scene optimizations with a single forward pass." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 467, + 200, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 467, + 200, + 479 + ], + "spans": [ + { + "bbox": [ + 151, + 467, + 200, + 479 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 498, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 498, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 498, + 297, + 714 + ], + "type": "text", + "content": "We present Twinner, the first large reconstruction model capable of recovering a scene's illumination as well as an object's geometry and material properties from only a few posed images. Twinner is based on the Large Reconstruction Model and innovates in three key ways: 1) We introduce a memory-efficient voxel-grid transformer whose memory scales only quadratically with the size of the voxel grid. 2) To address the scarcity of high-quality ground-truth PBR-shaded models, we introduce a large, fully synthetic dataset of procedurally generated PBR-textured objects lit with varied illumination. 3) To bridge the synthetic-to-real gap, we finetune the model on real-world datasets using a differentiable physically based shading model, eliminating the need for ground-truth illumination or material properties, which are challenging to obtain in real-world scenarios. We demonstrate the efficacy of our model on the real-world StanfordORB benchmark, where, given a few input views, we achieve reconstruction quality significantly superior to" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 468, + 555, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 555, + 492 + ], + "type": "text", + "content": "existing feed-forward reconstruction networks and comparable to slower per-scene optimization methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 514, + 395, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 514, + 395, + 526 + ], + "spans": [ + { + "bbox": [ + 314, + 514, + 395, + 526 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 533, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 533, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 312, + 533, + 555, + 677 + ], + "type": "text", + "content": "The creation of digital twins, i.e., the conversion of real-world objects into virtual 3D assets, is an important problem in virtual and augmented reality, gaming, conservation, education, advertisement, and general 3D content creation. In recent years, new techniques such as NeRF [31] and 3D Gaussian Splatting [24], along with their feed-forward counterparts [18, 49, 60, 63], have significantly improved the quality, speed, and generality of digital twinning. However, most of these techniques still bake illumination into the scene's appearance as radiance. As a consequence, the resulting assets cannot be relit and appear out of place when visualized under different lighting conditions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "To address this limitation, one must also reconstruct the field of Bidirectional Reflectance Distribution Functions (BRDF), which describe the optical properties of materi-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5859" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "content": "als. Estimating the BRDF directly is notoriously challenging due to its high dimensionality [3] and is usually done using 3D scanners that capture thousands of images while controlling for viewpoint and illumination [4-7, 35, 42, 65]. This is often impractical, so authors have considered \"casual\" reconstruction setups, where the object is captured by a conventional camera in uncontrolled conditions. Examples include NVDiffRec [34], NeRD [8], and NeRFactor [59], which simultaneously estimate the illumination, in the form of an environment map, and a simplified BRDF parameterized using Physically-Based Rendering (PBR) with albedo, roughness, and metallicity [10]. These quantities are recovered by minimizing the photometric reconstruction error, comparing the output of a differentiable PBR renderer to the input images. However, this optimization is slow, ill-posed, and requires the introduction of handcrafted priors for regularization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 277, + 294, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 294, + 337 + ], + "type": "text", + "content": "In this paper, we propose a new method, Twinner, a feedforward network that outputs shape, PBR textures, and environment illumination. This network is fast, performs well, requires only a small number of images as input, and learns empirically the necessary reconstruction priors." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 339, + 294, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 294, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 294, + 445 + ], + "type": "text", + "content": "Siddiqui et al. [45] recently introduced a similar predictor, MetalLRM, as an extension of the LightplaneLRM [11]. MetalLRM requires supervision with a large dataset of 3D assets with high-quality PBR textures, which, unfortunately, are not abundant in existing synthetic datasets like Objaverse [14]. Moreover, even if a large synthetic dataset were available, the synthetic-to-real gap between the artist-created training assets and the target domain of real input images makes MetalLRM unsuitable for digital twinning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 449, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 640 + ], + "type": "text", + "content": "With our new Twinner, we address the scarcity of high-quality training PBR assets and the train/test domain shift. First, we ask whether real images in datasets like CO3Dv2 [40] can be used to supervise the model. Although these datasets lack PBR information, they are abundant and match the target distribution of real images we aim to digitally twin. To use this data, we propose minimizing the photometric error similar to optimization-based models, but with the key difference that the 3D shape, PBR materials, and environment light are all predicted by a large feedforward transformer [19, 45]. During training, each scene is rendered with a differentiable PBR shader that accepts the predicted environment light and the predicted PBR-textured shape. The latter then admits indirect PBR and illumination supervision via photometric losses between the renders and the corresponding real ground-truth images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "type": "text", + "content": "Additionally, we consider a second approach to tackle data scarcity where, inspired by Zeroverse [58], we procedurally generate PBR-textured objects lit with a known set of environment maps. While the resulting objects are not realistic, their key advantage is that they come with clean ground-truth PBR and illumination information, which can" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "be used to supervise the corresponding predictors directly instead of indirectly via the photometric loss." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 98, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 98, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 98, + 553, + 300 + ], + "type": "text", + "content": "Besides addressing data scarcity, we also improve the architecture of the reconstruction model. We believe in the virtues of a simple and direct voxel-grid representation of the object [47], but note that predicting a number of tokens cubic in the voxel grid resolution is intractable with a transformer. Reducing the number of predicted tokens is the primary reason why recent models [19, 45] factor the voxel grid as a triplane. The question, then, is how to retain the benefits of the direct voxel grid representation and still make computations feasible. We solve this problem by introducing the novel tricolumn representation, where each feature vector (column) in a triplane is obtained by stacking the corresponding cells in the underlying voxel grid. In this manner, the number of tokens becomes quadratic in the voxel grid size, while each voxel is still represented explicitly. We further augment the network to predict a representation of the environment map, capturing illumination." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 301, + 553, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 553, + 419 + ], + "type": "text", + "content": "We demonstrate the efficacy of the Twinner model on the StanfordORB benchmark [26], consisting of casual captures of real-life objects. We achieve reconstruction quality superior to existing feed-forward predictors and comparable to optimization-based methods, while using only a very small number of input images of the object and operating much faster. We also obtain state-of-the-art reconstruction of the environment maps, achieving a " + }, + { + "bbox": [ + 313, + 301, + 553, + 419 + ], + "type": "inline_equation", + "content": "28\\%" + }, + { + "bbox": [ + 313, + 301, + 553, + 419 + ], + "type": "text", + "content": " improvement in the predicted illumination's angular error over existing methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 435, + 399, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 435, + 399, + 447 + ], + "spans": [ + { + "bbox": [ + 314, + 435, + 399, + 447 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 456, + 553, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 553, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 553, + 671 + ], + "type": "text", + "content": "3D reconstruction. 3D reconstruction aims to recover a three-dimensional representation of a scene from a set of posed images. Traditionally, this was accomplished with structure-from-motion [1, 43, 53], multi-view stereo [16, 44], and mesh surface extraction methods [23, 29]. Recently, methods based on neural fields [2, 31-33, 51, 54, 62] have focused on extracting both shape and appearance by learning the scene radiance function. Through volumetric rendering and photometric losses, they are able to implicitly reconstruct a 3D density volume from which a mesh can be extracted. However, these methods infer the parameters of each scene via test-time optimization, making them slow. Authors have thus introduced new feed-forward reconstruction networks, generically called Large Reconstruction Models (LRMs) [19, 56, 58, 67]. LRMs use transformers to predict 3D representations from a small set of input images. Our Twinner predicts not just geometry and radiance, but also material properties and illumination." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Material reconstruction. 3D reconstruction typically recovers a scene's geometry and radiance under fixed illumination. However, it is also necessary to recover material" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5860" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 556, + 220 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 556, + 220 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 556, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 556, + 220 + ], + "type": "image", + "image_path": "8fa14c1ea85692e07b84abf9d144bb320c70789d9224a53c44bdb0d95f85f84f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 227, + 555, + 284 + ], + "lines": [ + { + "bbox": [ + 54, + 227, + 555, + 284 + ], + "spans": [ + { + "bbox": [ + 54, + 227, + 555, + 284 + ], + "type": "text", + "content": "Figure 2. Overview of Twinner. The input views are processed together with their foreground masks and camera poses by an image tokenizer. The resulting tokens are processed by a Diffusion Transformer (DiT) model to predict a 3D volumetric representation of the scene, and a second DiT model to predict the scene's illumination. The 3D representation is then rendered to obtain images of the scene's material properties, normals, and opacity from the target's point of view. Using these rendered images together with the predicted illumination, Twinner renders an approximate physically-shaded image of the scene which then enters a photometric loss." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 293, + 296, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 293, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 54, + 293, + 296, + 450 + ], + "type": "text", + "content": "properties so that the 3D assets look realistic in novel virtual scenes. To do so, some works focus on optimizing explicit representations, such as material textures on meshes [17, 34, 48], while other works have added physically-based rendering models into neural fields [8, 9, 21, 27, 30, 64] and Gaussian splats [20, 28]. While these methods achieve high-quality material reconstructions, they require costly per-scene optimization. Some have addressed this issue by learning feed-forward reconstructors of PBR assets [45]. However, these are typically trained with synthetic data due to the complexity of capturing real-world material properties. We bridge the gap between synthetic and real data by training Twinner on real-world datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 454, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 296, + 647 + ], + "type": "text", + "content": "Illumination prediction. A key to accurate material reconstruction is accurate modeling of the scene's illumination. Single-scene optimization methods achieve this by optimizing an internal representation of illumination using light probe images [17, 34, 59], spherical harmonics [28], spherical Gaussians [21, 66], or MLPs [27, 30, 64]. These methods rely on physical models of lighting to aid the optimization. Others have attempted to predict the illumination of a scene from a single image [38, 52]. They rely on generative models to map the input view to a possible distribution of panorama views representing the environment illumination. Our Twinner lies in between these two lines of work: it is similar to the predictive approach as it predicts illumination from a set of four input views, but it uses physically-based rendering losses for supervision, similar to single-scene optimization methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 657, + 138, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 138, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 138, + 671 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "We introduce Twinner, a model that reconstructs the shape, materials, and illumination of real-life 3D objects from a small set of posed input views. Twinner extends Light-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 293, + 555, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 401 + ], + "type": "text", + "content": "planeLRM [11], an enhanced version of LRM [19], by incorporating predictors for the object's material properties and scene illumination, as well as a differentiable physically-based renderer. In Sec. 3.1, we begin by reviewing the reconstructor and the PBR model we use. Then, in Secs. 3.2 to 3.4, we describe the main building blocks of Twinner. Finally, in Sec. 3.5, we detail the training protocol, including the losses used to supervise Twinner on real data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 409, + 400, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 400, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 400, + 421 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 426, + 556, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 522 + ], + "type": "text", + "content": "Physically-based rendering. To supervise our model with a photometric loss, it is necessary to render the reconstructed scene into an image. The color of each pixel corresponds to the radiance " + }, + { + "bbox": [ + 313, + 426, + 556, + 522 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_o" + }, + { + "bbox": [ + 313, + 426, + 556, + 522 + ], + "type": "text", + "content": " emitted by the imaged object point " + }, + { + "bbox": [ + 313, + 426, + 556, + 522 + ], + "type": "inline_equation", + "content": "p\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 426, + 556, + 522 + ], + "type": "text", + "content": " in the direction of the camera. With knowledge of a scene's illumination, geometry, and material properties, it is possible to estimate the outgoing radiance at any such point according to the reflectance equation:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 357, + 530, + 555, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 530, + 555, + 563 + ], + "spans": [ + { + "bbox": [ + 357, + 530, + 555, + 563 + ], + "type": "interline_equation", + "content": "\\mathbf {L} _ {o} = \\int_ {\\Omega} \\left(k _ {d} \\frac {\\mathbf {a}}{\\pi} + \\mathbf {f} _ {s}\\right) \\mathbf {L} _ {i} \\left\\langle \\boldsymbol {\\omega} _ {i}, \\mathbf {n} \\right\\rangle d \\boldsymbol {\\omega} _ {i}. \\tag {1}", + "image_path": "e8ce7e015b134477940e2f75e30ac6e001d0c3edc8dc9ee8861add19c8413e0c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_i" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": " is the incoming radiance, representing the scene illumination falling on the imaged point " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": ". The integral is carried out over a hemisphere " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": " of incoming directions " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\omega_{i}" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": " determined by the surface normal " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": ". The factor in parentheses is a model of the object's Bidirectional Reflectance Distribution Function (BRDF), which describes how the incoming radiance reflects off the object's surface into various directions. The first term is the diffuse component, parameterized by the object's diffuse albedo " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": " and a constant " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "k_{d}" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": ", both of which are independent of direction. In the integral, this term averages all incoming radiance over the hemisphere. The second term, " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_s" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": ", is the specular com" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5861" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 76, + 556, + 191 + ], + "blocks": [ + { + "bbox": [ + 65, + 76, + 556, + 191 + ], + "lines": [ + { + "bbox": [ + 65, + 76, + 556, + 191 + ], + "spans": [ + { + "bbox": [ + 65, + 76, + 556, + 191 + ], + "type": "image", + "image_path": "194392d4b0cbe5ee68de5ccc42997a79092f846e1d944a55ed3c5925e4abd483.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 205, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 54, + 205, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 54, + 205, + 555, + 251 + ], + "type": "text", + "content": "Figure 3. Tricolumn Architecture. The transformer module processes image tokens and a learnable embedding to predict the tricolumn representation, which consists of a grid of " + }, + { + "bbox": [ + 54, + 205, + 555, + 251 + ], + "type": "inline_equation", + "content": "3R^2" + }, + { + "bbox": [ + 54, + 205, + 555, + 251 + ], + "type": "text", + "content": " feature vectors with dimensionality " + }, + { + "bbox": [ + 54, + 205, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{CR}" + }, + { + "bbox": [ + 54, + 205, + 555, + 251 + ], + "type": "text", + "content": ". These feature vectors are then split into three axis-aligned column grids, each with dimensionality " + }, + { + "bbox": [ + 54, + 205, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{CR}" + }, + { + "bbox": [ + 54, + 205, + 555, + 251 + ], + "type": "text", + "content": ". Finally, all axis-aligned column grids are reshaped into voxel grids of size " + }, + { + "bbox": [ + 54, + 205, + 555, + 251 + ], + "type": "inline_equation", + "content": "R\\times R\\times R" + }, + { + "bbox": [ + 54, + 205, + 555, + 251 + ], + "type": "text", + "content": " with dimensionality " + }, + { + "bbox": [ + 54, + 205, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^C" + }, + { + "bbox": [ + 54, + 205, + 555, + 251 + ], + "type": "text", + "content": " and summed into a single voxel grid." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 261, + 294, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 261, + 294, + 356 + ], + "spans": [ + { + "bbox": [ + 54, + 261, + 294, + 356 + ], + "type": "text", + "content": "ponent of the BRDF, which captures directional effects such as reflections or highlights. Note that " + }, + { + "bbox": [ + 54, + 261, + 294, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_i" + }, + { + "bbox": [ + 54, + 261, + 294, + 356 + ], + "type": "text", + "content": " is a function of the incoming direction " + }, + { + "bbox": [ + 54, + 261, + 294, + 356 + ], + "type": "inline_equation", + "content": "\\omega_{i}" + }, + { + "bbox": [ + 54, + 261, + 294, + 356 + ], + "type": "text", + "content": ", and the specular BRDF " + }, + { + "bbox": [ + 54, + 261, + 294, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_s" + }, + { + "bbox": [ + 54, + 261, + 294, + 356 + ], + "type": "text", + "content": " is a function of both the incoming and outgoing directions " + }, + { + "bbox": [ + 54, + 261, + 294, + 356 + ], + "type": "inline_equation", + "content": "\\omega_{i}" + }, + { + "bbox": [ + 54, + 261, + 294, + 356 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 261, + 294, + 356 + ], + "type": "inline_equation", + "content": "\\omega_{o}" + }, + { + "bbox": [ + 54, + 261, + 294, + 356 + ], + "type": "text", + "content": ". Radiances and BRDFs also depend on the light frequency, which we model using three RGB components. Furthermore, all quantities depend on the point position " + }, + { + "bbox": [ + 54, + 261, + 294, + 356 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 54, + 261, + 294, + 356 + ], + "type": "text", + "content": ". We omit these details for brevity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 358, + 295, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 358, + 295, + 490 + ], + "spans": [ + { + "bbox": [ + 54, + 358, + 295, + 490 + ], + "type": "text", + "content": "Due to the complexity of the integral in Eq. (1), numerous approximations have been proposed. We rely on the split-sum approximation [22] for image-based lighting, which leverages the scene's illumination represented as a spherical image enveloping the scene at an infinite radius. Following [34], we use a set of differentiable mipmaps to represent the pre-integrated lighting. The BRDF follows the Disney [22] microfacet BRDF model and is parameterized by the albedo " + }, + { + "bbox": [ + 54, + 358, + 295, + 490 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 54, + 358, + 295, + 490 + ], + "type": "text", + "content": ", metalness " + }, + { + "bbox": [ + 54, + 358, + 295, + 490 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 54, + 358, + 295, + 490 + ], + "type": "text", + "content": ", and roughness " + }, + { + "bbox": [ + 54, + 358, + 295, + 490 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 54, + 358, + 295, + 490 + ], + "type": "text", + "content": " at each point. We further assume that " + }, + { + "bbox": [ + 54, + 358, + 295, + 490 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_i" + }, + { + "bbox": [ + 54, + 358, + 295, + 490 + ], + "type": "text", + "content": " is constant in space (hence we do not account for shadows cast by the object)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 491, + 296, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 296, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 296, + 551 + ], + "type": "text", + "content": "In summary, to apply this model to render an image of the object, we require an estimate of the object's 3D shape and its albedo, metalness, and roughness at its surfaces. We also require a spherical image representing the directional function " + }, + { + "bbox": [ + 55, + 491, + 296, + 551 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_i" + }, + { + "bbox": [ + 55, + 491, + 296, + 551 + ], + "type": "text", + "content": ", known as an environment map." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "text", + "content": "Large Reconstruction Model (LRM). The aim of LRM [19] is to map a few posed views " + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "text", + "content": " of the scene to a 3D reconstruction " + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "text", + "content": ". LRM first extracts features " + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "text", + "content": " from the set of input views, utilizing a pre-trained DINO [36] Vision Transformer (ViT) [15]. Following [61], we feed into the ViT the RGB images " + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "text", + "content": ", the Plücker embeddings [46] " + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathrm{Pl}}" + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "text", + "content": " representing the cameras, and the masks " + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "text", + "content": " of the foreground object to be reconstructed." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "The extracted feature tokens enter, via cross-attention, the transformer architecture from Diffusion Transformer (DiT) [37] " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": ", which is tasked with reconstructing the scene. We emphasize that we borrow the DiT architecture without any diffusion process. The tokens in this transformer en" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 261, + 554, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 554, + 319 + ], + "type": "text", + "content": "code the 3D representation of the scene, realizing a mapping " + }, + { + "bbox": [ + 313, + 261, + 554, + 319 + ], + "type": "inline_equation", + "content": "V = \\Phi(V_0|y)" + }, + { + "bbox": [ + 313, + 261, + 554, + 319 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 261, + 554, + 319 + ], + "type": "inline_equation", + "content": "V_0" + }, + { + "bbox": [ + 313, + 261, + 554, + 319 + ], + "type": "text", + "content": " are the learnable initial tokens. A crucial design choice is how to construct the tokens " + }, + { + "bbox": [ + 313, + 261, + 554, + 319 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 261, + 554, + 319 + ], + "type": "text", + "content": " manipulated by this transformer, which is discussed in Sec. 3.2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 321, + 554, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 554, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 554, + 393 + ], + "type": "text", + "content": "Finally, using the memory-efficient rendering algorithm proposed in Lightplane [11], a target set of views is rendered from the 3D representation. A shaded RGB image " + }, + { + "bbox": [ + 313, + 321, + 554, + 393 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{I}}" + }, + { + "bbox": [ + 313, + 321, + 554, + 393 + ], + "type": "text", + "content": ", a foreground mask image " + }, + { + "bbox": [ + 313, + 321, + 554, + 393 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}" + }, + { + "bbox": [ + 313, + 321, + 554, + 393 + ], + "type": "text", + "content": ", and a depth image " + }, + { + "bbox": [ + 313, + 321, + 554, + 393 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{I}}_{\\mathrm{d}}" + }, + { + "bbox": [ + 313, + 321, + 554, + 393 + ], + "type": "text", + "content": " are rendered for each view and used to supervise the LRM through a combination of the following photometric losses:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 341, + 400, + 553, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 400, + 553, + 412 + ], + "spans": [ + { + "bbox": [ + 341, + 400, + 553, + 412 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {M} = \\operatorname {B C E} (\\mathcal {M}, \\hat {\\mathcal {M}}), \\tag {2}", + "image_path": "5f47fa0751d09b2706dfe10f892e64920c089d9501d441de44d1b4664fcced30.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 344, + 416, + 553, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 416, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 344, + 416, + 553, + 430 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {R} = \\operatorname {L P I P S} \\left(\\mathcal {M} \\mathcal {I}, \\mathcal {M} \\hat {\\mathcal {I}}\\right) + \\left\\| \\mathcal {M} \\mathcal {I} - \\mathcal {M} \\hat {\\mathcal {I}} \\right\\| ^ {2}, \\tag {3}", + "image_path": "4e1933def13858afa4acd63ed353ece6c4ad15b91ee9346f29e73c0507d79440.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 346, + 434, + 553, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 434, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 346, + 434, + 553, + 447 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d} = \\left\\| \\mathcal {M I} _ {\\mathrm {d}} - \\mathcal {M} \\hat {\\mathcal {I}} _ {\\mathrm {d}} \\right\\|, \\tag {4}", + "image_path": "6a780d772dcd191762db1e30254f7a75ea8eaaa377867ec742f4dabe72ceaafa.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 454, + 554, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 554, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 554, + 479 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 454, + 554, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathrm{d}}" + }, + { + "bbox": [ + 313, + 454, + 554, + 479 + ], + "type": "text", + "content": " is the set of ground-truth depth maps corresponding to the input images " + }, + { + "bbox": [ + 313, + 454, + 554, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 454, + 554, + 479 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 454, + 554, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{M}\\mathcal{I}" + }, + { + "bbox": [ + 313, + 454, + 554, + 479 + ], + "type": "text", + "content": " denotes a masked image." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 485, + 457, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 457, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 457, + 498 + ], + "type": "text", + "content": "3.2. Tricolumn representation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 502, + 554, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 554, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 554, + 633 + ], + "type": "text", + "content": "Crucial to the Twinner design is the representation of the 3D scene, which is encoded as a set of volumetric functions expressing opacity and material properties at every point " + }, + { + "bbox": [ + 313, + 502, + 554, + 633 + ], + "type": "inline_equation", + "content": "\\pmb{p} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 502, + 554, + 633 + ], + "type": "text", + "content": " in space. A simple and effective option [47] to encode a volumetric map " + }, + { + "bbox": [ + 313, + 502, + 554, + 633 + ], + "type": "inline_equation", + "content": "V: \\mathbb{R}^3 \\to \\mathbb{R}^C" + }, + { + "bbox": [ + 313, + 502, + 554, + 633 + ], + "type": "text", + "content": " is to use a voxel grid " + }, + { + "bbox": [ + 313, + 502, + 554, + 633 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{C \\times D \\times H \\times W}" + }, + { + "bbox": [ + 313, + 502, + 554, + 633 + ], + "type": "text", + "content": ". Here, the value " + }, + { + "bbox": [ + 313, + 502, + 554, + 633 + ], + "type": "inline_equation", + "content": "V(\\pmb{p}) \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 313, + 502, + 554, + 633 + ], + "type": "text", + "content": " interpolates the voxel values at the nearest grid points. To predict the voxel grid " + }, + { + "bbox": [ + 313, + 502, + 554, + 633 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 502, + 554, + 633 + ], + "type": "text", + "content": " with the transformer of Sec. 3.1, we could assign one token to each voxel. However, the number " + }, + { + "bbox": [ + 313, + 502, + 554, + 633 + ], + "type": "inline_equation", + "content": "WHD" + }, + { + "bbox": [ + 313, + 502, + 554, + 633 + ], + "type": "text", + "content": " of voxels grows cubically with the resolution of the grid, which leads to prohibitive memory demands." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 634, + 554, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 634, + 554, + 694 + ], + "spans": [ + { + "bbox": [ + 313, + 634, + 554, + 694 + ], + "type": "text", + "content": "The alternative adopted by models like [19, 45] is to use a triplane representation [12]. In this case, the volumetric function is represented by three 2D grids " + }, + { + "bbox": [ + 313, + 634, + 554, + 694 + ], + "type": "inline_equation", + "content": "V_{xy} \\in \\mathbb{R}^{K \\times H \\times W}" + }, + { + "bbox": [ + 313, + 634, + 554, + 694 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 634, + 554, + 694 + ], + "type": "inline_equation", + "content": "V_{yz} \\in \\mathbb{R}^{K \\times W \\times D}" + }, + { + "bbox": [ + 313, + 634, + 554, + 694 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 634, + 554, + 694 + ], + "type": "inline_equation", + "content": "V_{zx} \\in \\mathbb{R}^{K \\times D \\times H}" + }, + { + "bbox": [ + 313, + 634, + 554, + 694 + ], + "type": "text", + "content": ". The volumetric function at " + }, + { + "bbox": [ + 313, + 634, + 554, + 694 + ], + "type": "inline_equation", + "content": "\\pmb{p} = (x,y,z)" + }, + { + "bbox": [ + 313, + 634, + 554, + 694 + ], + "type": "text", + "content": " is then expressed as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 330, + 700, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 700, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 330, + 700, + 553, + 715 + ], + "type": "interline_equation", + "content": "V (x, y, z) = f \\left(V _ {x y} (x, y), V _ {y z} (y, z), V _ {z x} (z, x)\\right), \\tag {5}", + "image_path": "e855bf955e037f8dcee658c82f467896169a323ae6557ad83966888a5f986639.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5862" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 298, + 215 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 298, + 215 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 298, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 298, + 215 + ], + "type": "image", + "image_path": "711809666565476ddb2350ffe020e6716d6988f6aa55345e373cc0d9d6505e46.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 220, + 297, + 255 + ], + "lines": [ + { + "bbox": [ + 55, + 220, + 297, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 297, + 255 + ], + "type": "text", + "content": "Figure 4. Examples from our procedural dataset visualizing the shaded image, environment map, materials, and normals from our environment-map-endowed version of Zeroverse [57]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 262, + 296, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 296, + 334 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 262, + 296, + 334 + ], + "type": "inline_equation", + "content": "f: \\mathbb{R}^{K \\times K \\times K} \\to \\mathbb{R}^C" + }, + { + "bbox": [ + 55, + 262, + 296, + 334 + ], + "type": "text", + "content": " is a point-wise mixing function (e.g., an MLP applied to the concatenation of the three " + }, + { + "bbox": [ + 55, + 262, + 296, + 334 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 262, + 296, + 334 + ], + "type": "text", + "content": "-dimensional feature vectors). Paired with a transformer, this conveniently reduces the number of tokens to " + }, + { + "bbox": [ + 55, + 262, + 296, + 334 + ], + "type": "inline_equation", + "content": "WH + HD + DW" + }, + { + "bbox": [ + 55, + 262, + 296, + 334 + ], + "type": "text", + "content": " only. However, triplane mixes information between different voxels, which can lead to artifacts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 335, + 296, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 296, + 394 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 296, + 394 + ], + "type": "text", + "content": "We thus propose a novel tricolumn representation instead. This is similar to triplane as it uses a plane instead of a volume of tokens; however, each feature vector is in itself a column of corresponding voxel cells, so that voxel-specific information is represented directly and separately." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 395, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 395, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 296, + 479 + ], + "type": "text", + "content": "In detail, we consider a plane of feature vectors " + }, + { + "bbox": [ + 55, + 395, + 296, + 479 + ], + "type": "inline_equation", + "content": "V_{xy} \\in \\mathbb{R}^{(CR) \\times R \\times R}" + }, + { + "bbox": [ + 55, + 395, + 296, + 479 + ], + "type": "text", + "content": ". Because this is a plane, there are only " + }, + { + "bbox": [ + 55, + 395, + 296, + 479 + ], + "type": "inline_equation", + "content": "R^2" + }, + { + "bbox": [ + 55, + 395, + 296, + 479 + ], + "type": "text", + "content": " tokens, which are easily handled by the transformer. However, " + }, + { + "bbox": [ + 55, + 395, + 296, + 479 + ], + "type": "inline_equation", + "content": "V_{xy}" + }, + { + "bbox": [ + 55, + 395, + 296, + 479 + ], + "type": "text", + "content": " can be reshaped into a " + }, + { + "bbox": [ + 55, + 395, + 296, + 479 + ], + "type": "inline_equation", + "content": "C \\times R \\times R \\times R" + }, + { + "bbox": [ + 55, + 395, + 296, + 479 + ], + "type": "text", + "content": " voxel grid by unflattening the first dimension. We use the symbol " + }, + { + "bbox": [ + 55, + 395, + 296, + 479 + ], + "type": "inline_equation", + "content": "V_{xy}(x,y;z) \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 55, + 395, + 296, + 479 + ], + "type": "text", + "content": " to denote the feature vector obtained by interpolating this grid in 3D." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 479, + 296, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 296, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 296, + 503 + ], + "type": "text", + "content": "When three such planes are used, the full feature vector at " + }, + { + "bbox": [ + 55, + 479, + 296, + 503 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 55, + 479, + 296, + 503 + ], + "type": "text", + "content": " is decoded as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 509, + 296, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 509, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 63, + 509, + 296, + 533 + ], + "type": "interline_equation", + "content": "V (x, y, z) = f \\left(V _ {x y} (x, y; z), V _ {y z} (y, z; x), V _ {z x} (z, x; y)\\right). \\tag {6}", + "image_path": "ec93aa328e96b778241342ed1302bb269b378f4ff9ab86b765a2727d47c899a3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 534, + 296, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 583 + ], + "type": "text", + "content": "Note that this construction is strictly different from the triplane equation (5) as each term depends on all three spatial coordinates. We opt for using summation followed by an MLP as the mixing function " + }, + { + "bbox": [ + 55, + 534, + 296, + 583 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 534, + 296, + 583 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 588, + 203, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 203, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 203, + 600 + ], + "type": "text", + "content": "3.3. PBR reconstruction model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "The LRM models of [19, 61] only recover the shape and radiance of the scene. Like in [45], we extend Twinner to also recover the material properties. We do so by extending the tricolumn representation to encode the albedo " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " (instead of radiance) and two additional channels representing the scene's metalness " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " and roughness " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": ". We also note the importance of accurate normals for relighting. While a normal image could be directly estimated by projecting the rendered depth image into 3D space and taking the gradient" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 554, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 143 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 143 + ], + "type": "text", + "content": "of the opacity field, " + }, + { + "bbox": [ + 313, + 72, + 554, + 143 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{I}}_{\\mathrm{n}} = \\nabla \\sigma (\\pi^{-1}(\\hat{\\mathcal{I}}_{\\mathrm{d}}))" + }, + { + "bbox": [ + 313, + 72, + 554, + 143 + ], + "type": "text", + "content": ", this estimate tends to be noisy. Therefore, we follow previous works [51] and predict three additional channels representing the scene's normals. The predicted normal image " + }, + { + "bbox": [ + 313, + 72, + 554, + 143 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{I}}_{\\mathrm{n}}" + }, + { + "bbox": [ + 313, + 72, + 554, + 143 + ], + "type": "text", + "content": " is then supervised with the pseudo-ground-truth normals from the opacity field " + }, + { + "bbox": [ + 313, + 72, + 554, + 143 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{I}}_{\\mathrm{n}}" + }, + { + "bbox": [ + 313, + 72, + 554, + 143 + ], + "type": "text", + "content": " via:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 381, + 143, + 554, + 157 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 143, + 554, + 157 + ], + "spans": [ + { + "bbox": [ + 381, + 143, + 554, + 157 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {N} = 1 - \\langle \\mathcal {M} \\bar {\\mathcal {I}} _ {\\mathrm {n}}, \\mathcal {M} \\hat {\\mathcal {I}} _ {\\mathrm {n}} \\rangle . \\tag {7}", + "image_path": "3acebec50e01133f7f357c15973e4d7459289e3b413a060bc7a97fe45576fa4a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 159, + 554, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 159, + 554, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 554, + 243 + ], + "type": "text", + "content": "Direct supervision using synthetic data. Reconstruction models that predict radiance can be supervised by rendering the predicted 3D scene into images and comparing these to training views using a photometric loss. In our case, however, the model does not predict radiance, but albedo, roughness, and metalness, which cannot be rendered into a 'shaded' image unless illumination is also known." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 243, + 554, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 554, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 554, + 327 + ], + "type": "text", + "content": "A simple approach for supervising the model that does not require knowledge of illumination is to render albedo, roughness, and metalness images " + }, + { + "bbox": [ + 313, + 243, + 554, + 327 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{I}}_{\\mathrm{a}},\\hat{\\mathcal{I}}_{\\mathrm{r}}" + }, + { + "bbox": [ + 313, + 243, + 554, + 327 + ], + "type": "text", + "content": " ,and " + }, + { + "bbox": [ + 313, + 243, + 554, + 327 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{I}}_{\\mathrm{m}}" + }, + { + "bbox": [ + 313, + 243, + 554, + 327 + ], + "type": "text", + "content": " , and compare these to their corresponding ground-truth renders. We use this approach to supervise our model with synthetic data by applying the following loss on the set of material renders " + }, + { + "bbox": [ + 313, + 243, + 554, + 327 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{I}}_{\\mathrm{a}},\\hat{\\mathcal{I}}_{\\mathrm{m}},\\hat{\\mathcal{I}}_{\\mathrm{r}}" + }, + { + "bbox": [ + 313, + 243, + 554, + 327 + ], + "type": "text", + "content": " , and also on the normal render " + }, + { + "bbox": [ + 313, + 243, + 554, + 327 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{I}}_{\\mathrm{n}}" + }, + { + "bbox": [ + 313, + 243, + 554, + 327 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 331, + 553, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 331, + 553, + 356 + ], + "spans": [ + { + "bbox": [ + 314, + 331, + 553, + 356 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S} = \\ell_ {\\mathcal {M}} \\left(\\mathcal {I} _ {\\mathrm {a}}, \\hat {\\mathcal {I}} _ {\\mathrm {a}}\\right) + \\ell_ {\\mathcal {M}} \\left(\\mathcal {I} _ {\\mathrm {r}}, \\hat {\\mathcal {I}} _ {\\mathrm {r}}\\right) + \\ell_ {\\mathcal {M}} \\left(\\mathcal {I} _ {\\mathrm {m}}, \\hat {\\mathcal {I}} _ {\\mathrm {m}}\\right) + \\ell_ {\\mathcal {M}} \\left(\\mathcal {I} _ {\\mathrm {n}}, \\hat {\\mathcal {I}} _ {\\mathrm {n}}\\right) \\tag {8}", + "image_path": "db6ec8bd281e3d7a917cb7f9e043929573ca6e479c73611c34124d7e24b898c6.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 356, + 553, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 553, + 380 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 356, + 553, + 380 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathcal{M}}(u,v) = \\| \\mathcal{M}(u - v)\\|_{\\epsilon}" + }, + { + "bbox": [ + 313, + 356, + 553, + 380 + ], + "type": "text", + "content": " is the Huber loss " + }, + { + "bbox": [ + 313, + 356, + 553, + 380 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_{\\epsilon}" + }, + { + "bbox": [ + 313, + 356, + 553, + 380 + ], + "type": "text", + "content": " with cut-off threshold " + }, + { + "bbox": [ + 313, + 356, + 553, + 380 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 356, + 553, + 380 + ], + "type": "text", + "content": " masked by " + }, + { + "bbox": [ + 313, + 356, + 553, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 356, + 553, + 380 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 386, + 554, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 554, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 554, + 399 + ], + "type": "text", + "content": "3.4. Training on real data by predicting illumination" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 403, + 554, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 554, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 554, + 511 + ], + "type": "text", + "content": "The learning formulation of Sec. 3.3 requires ground-truth material properties to be known, which is not the case for real-world data where only shaded images are available. In order to fine-tune our model using real data with a photometric loss, it is necessary to also estimate the scene illumination. With this, we can use the differentiable physically-based shader detailed in Sec. 3.1 to render the shaded scene appearance and compare it to the input images through a photometric loss." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 512, + 555, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 512, + 555, + 655 + ], + "spans": [ + { + "bbox": [ + 313, + 512, + 555, + 655 + ], + "type": "text", + "content": "To estimate the scene's illumination, we propose using an additional DiT module. The latter takes as input the feature tokens " + }, + { + "bbox": [ + 313, + 512, + 555, + 655 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 512, + 555, + 655 + ], + "type": "text", + "content": " computed by the image feature extractor and predicts a cubemap representation of the scene's lighting. Recall that a cubemap is a function " + }, + { + "bbox": [ + 313, + 512, + 555, + 655 + ], + "type": "inline_equation", + "content": "C \\to \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 512, + 555, + 655 + ], + "type": "text", + "content": " that assigns every point " + }, + { + "bbox": [ + 313, + 512, + 555, + 655 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 313, + 512, + 555, + 655 + ], + "type": "text", + "content": " of the cube surface " + }, + { + "bbox": [ + 313, + 512, + 555, + 655 + ], + "type": "inline_equation", + "content": "C = \\{\\pmb{p} : \\| \\pmb{p} \\|_{\\infty} = 1\\}" + }, + { + "bbox": [ + 313, + 512, + 555, + 655 + ], + "type": "text", + "content": " to the radiance " + }, + { + "bbox": [ + 313, + 512, + 555, + 655 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_i(\\boldsymbol{\\omega}_i)" + }, + { + "bbox": [ + 313, + 512, + 555, + 655 + ], + "type": "text", + "content": " incoming from direction " + }, + { + "bbox": [ + 313, + 512, + 555, + 655 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\omega}_i \\propto -\\pmb{p}" + }, + { + "bbox": [ + 313, + 512, + 555, + 655 + ], + "type": "text", + "content": ", and is thus a representation of the environment illumination (this works because we assume that the function " + }, + { + "bbox": [ + 313, + 512, + 555, + 655 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_i" + }, + { + "bbox": [ + 313, + 512, + 555, + 655 + ], + "type": "text", + "content": " only depends on direction, but not on position). The cubemap is unfolded to a collection of six 2D images, which are directly predicted by the DiT module." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 659, + 554, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 554, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 554, + 696 + ], + "type": "text", + "content": "Supervising illumination with synthetic data. In synthetic data, the cubemap is known, and its predictor can be supervised directly with the environment-map loss:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 358, + 700, + 554, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 700, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 358, + 700, + 554, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {e n v}} = \\left\\| \\mathcal {I} _ {\\mathrm {e}} - \\hat {\\mathcal {I}} _ {\\mathrm {e}} \\right\\| ^ {2} + \\operatorname {L P I P S} \\left(\\mathcal {I} _ {\\mathrm {e}}, \\hat {\\mathcal {I}} _ {\\mathrm {e}}\\right), \\tag {9}", + "image_path": "9d1fc7e59a69694a5879b8ad470caff377f10d08973a28c7f640b0cf2305b362.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5863" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 70, + 553, + 198 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 553, + 198 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 553, + 198 + ], + "type": "table", + "html": "
Novel Scene RelightingNovel View SynthesisGeometry EstimationRuntime
MethodPSNR-H↑PSNR-L↑SSIM↑LPIPS↓PSNR-H↑PSNR-L↑SSIM↑LPIPS↓Depth↓Geometry↓Shape↓Time (hr.)↓
Single-Scene Opt.PhySG [66]21.8128.110.9600.05524.2432.150.9740.0471.900.179.28~3
NVDiffRec [34]22.9129.720.9630.03921.9428.440.9690.0300.310.060.62~1
NeRD [8]23.2929.650.9570.05925.8332.610.9630.0541.390.2813.70>20
NeRFactor [59]23.5430.380.9690.04826.0633.470.9730.0460.870.299.53>20
InvRender [69]23.7630.830.9700.04625.9134.010.9770.0420.590.060.44~3
NVDiffRecMC [17]24.4331.600.9720.03628.0336.400.9820.0280.320.040.51~2
Neural-PBIR [48]26.0133.260.9790.02328.8236.800.9860.0190.300.060.43~1
PredictiveMetaLRM [45]21.5028.060.9570.05219.5525.980.9560.0518.6690.061.260.004
w/o ZeroVerse23.5330.310.9650.04821.5328.780.9650.0481.1260.090.740.006
w/o Shading Loss23.8930.850.9670.04221.4229.470.9670.0410.7620.060.300.004
Twinner (ours)24.5531.610.9690.03821.7029.620.9680.0370.7090.060.370.006
", + "image_path": "5db87447dcdd41a1a4729a3c60c8eb49649c847299d61f8715e8a596f47bfac8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 207, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 55, + 207, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 555, + 262 + ], + "type": "text", + "content": "Table 1. Relighting evaluation on StanfordORB [26] comparing our method to other feedforward predictors of PBR-textured assets (two ablations and MetalLRM). For completeness, we also include single-scene optimization methods in the top part of the table. Our model outperforms most single-scene optimization baseline methods in terms of novel scene relighting and shape estimation despite being a predictive model and having access to only four input views. More impressively, it only requires a few seconds to reconstruct an object's mesh with PBR materials, while single-scene optimization methods have runtimes in the order of hours." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 273, + 294, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 273, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 54, + 273, + 294, + 334 + ], + "type": "text", + "content": "comprising the photometric loss from (8) using the predicted cubemap projected into an equirectangular image " + }, + { + "bbox": [ + 54, + 273, + 294, + 334 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{I}}_{\\mathrm{e}}" + }, + { + "bbox": [ + 54, + 273, + 294, + 334 + ], + "type": "text", + "content": " and the ground-truth equirectangular image " + }, + { + "bbox": [ + 54, + 273, + 294, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathrm{e}}" + }, + { + "bbox": [ + 54, + 273, + 294, + 334 + ], + "type": "text", + "content": ". In practice, we use a procedurally generated dataset of synthetic assets detailed in Sec. 3.5 to provide ground-truth illumination." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 339, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 339, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 339, + 296, + 555 + ], + "type": "text", + "content": "Supervising illumination with real data. In real data, the cubemap is observed indirectly via the effect it has on the object's appearance according to the reflectance equation (1), which provides only weak supervision. We note, however, that the scene background visible in the input images is, to a first approximation, equal to the function " + }, + { + "bbox": [ + 56, + 339, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_i" + }, + { + "bbox": [ + 56, + 339, + 296, + 555 + ], + "type": "text", + "content": ", and can thus provide direct supervision of the latter. We can render the background observed from any given camera by assigning to each pixel the color " + }, + { + "bbox": [ + 56, + 339, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_i(\\boldsymbol{\\omega}_i) = C(\\boldsymbol{\\omega}_i / \\| \\boldsymbol{\\omega}_i \\|_{\\infty})" + }, + { + "bbox": [ + 56, + 339, + 296, + 555 + ], + "type": "text", + "content": " of the incoming radiance from the direction " + }, + { + "bbox": [ + 56, + 339, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\omega}_i" + }, + { + "bbox": [ + 56, + 339, + 296, + 555 + ], + "type": "text", + "content": " of the pixel's ray. Since input images are in sRGB color space, the predicted background image is first tone-mapped and then composited with the foreground object prediction to obtain a scene image. The latter allows us to supervise both the illumination prediction and the shaded object prediction through the photometric loss " + }, + { + "bbox": [ + 56, + 339, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_R" + }, + { + "bbox": [ + 56, + 339, + 296, + 555 + ], + "type": "text", + "content": " of Eq. (3). Here, as training data, we use abundant object-centric video captures from CO3Dv2 [41] as described in the next section." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 564, + 162, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 162, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 162, + 577 + ], + "type": "text", + "content": "3.5. Training protocol" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 230, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 230, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 230, + 594 + ], + "type": "text", + "content": "Twinner's training minimizes the total loss" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 605, + 295, + 619 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 605, + 295, + 619 + ], + "spans": [ + { + "bbox": [ + 63, + 605, + 295, + 619 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t}} = \\mathcal {L} _ {N} + \\mathcal {L} _ {S} + \\mathcal {L} _ {\\mathrm {T V}} + \\mathcal {L} _ {d} + \\mathcal {L} _ {M} + \\mathcal {L} _ {R} + \\mathcal {L} _ {\\mathrm {T V}}, \\tag {10}", + "image_path": "8e5e05c6dc2dca89ecad6528bdc2416ca6ab53677e47eb8f386ca1e9f083f27b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 628, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 295, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 628, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{TV}} = \\mathrm{TV}(\\hat{\\mathcal{I}}_{\\mathrm{n}}) + \\mathrm{TV}(\\hat{\\mathcal{I}}_{\\mathrm{a}}) + \\mathrm{TV}(\\hat{\\mathcal{I}}_{\\mathrm{d}})" + }, + { + "bbox": [ + 55, + 628, + 295, + 713 + ], + "type": "text", + "content": " is the Total-Variation (TV) regularization on normal, albedo, and depth images promoting smooth solutions. " + }, + { + "bbox": [ + 55, + 628, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{tot}}" + }, + { + "bbox": [ + 55, + 628, + 295, + 713 + ], + "type": "text", + "content": " is optimized with Adam [25] using a learning rate of " + }, + { + "bbox": [ + 55, + 628, + 295, + 713 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 55, + 628, + 295, + 713 + ], + "type": "text", + "content": " until convergence. As mentioned before, individual terms of " + }, + { + "bbox": [ + 55, + 628, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{tot}}" + }, + { + "bbox": [ + 55, + 628, + 295, + 713 + ], + "type": "text", + "content": " are switched on or off depending on whether a real or a synthetic training example is present in the batch of training images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 273, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 555, + 357 + ], + "type": "text", + "content": "Training data. Following [45], Twinner is trained on a large dataset of 140k artist-created assets similar to Objaverse. However, this dataset often lacks high-quality material maps where, for instance, distinguishing between missing material images and material images with zero values is impossible. Moreover, there is a large domain gap between the synthetic objects and the real ones we intend to twin." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 361, + 555, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 420 + ], + "type": "text", + "content": "As such, together with the artist-created assets, we also train on two additional data sources: 1) A large dataset of procedurally generated assets with clean PBR textures and environment maps; 2) Real-object captures from CO3Dv2 [40]. Both datasets are detailed next." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 429, + 556, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 556, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 556, + 573 + ], + "type": "text", + "content": "Endowing Zeroverse with environment maps. Following LRM-Zero [57], we generate a set of 10k randomized 3D objects with realistic material properties sampled from the MatSynth [50] dataset. Additionally, to enable the environment-map supervision loss " + }, + { + "bbox": [ + 313, + 429, + 556, + 573 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{env}}" + }, + { + "bbox": [ + 313, + 429, + 556, + 573 + ], + "type": "text", + "content": " (9), for each rendered scene, we randomly sample one of " + }, + { + "bbox": [ + 313, + 429, + 556, + 573 + ], + "type": "inline_equation", + "content": "\\sim 1\\mathrm{k}" + }, + { + "bbox": [ + 313, + 429, + 556, + 573 + ], + "type": "text", + "content": " environment maps from PolyHaven [39]. We render color, normal, material, depth, and foreground mask images for each object from a predetermined set of viewpoints (see examples in Fig. 4). As such, samples from this dataset admit optimizing all terms of the total loss " + }, + { + "bbox": [ + 313, + 429, + 556, + 573 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{tot}}" + }, + { + "bbox": [ + 313, + 429, + 556, + 573 + ], + "type": "text", + "content": ", including the environment loss " + }, + { + "bbox": [ + 313, + 429, + 556, + 573 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{env}}" + }, + { + "bbox": [ + 313, + 429, + 556, + 573 + ], + "type": "text", + "content": ", which is crucial for stable training of Twinner." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 581, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 556, + 713 + ], + "type": "text", + "content": "Exploiting real object captures. We also train on CO3Dv2 [40], which is a large-scale real-world dataset of \"turntable\" videos capturing objects of common categories. Each frame in the dataset comes with camera and foreground mask annotations. As such, for each CO3Dv2 sample, we can optimize the normal, mask, photometric, and TV losses " + }, + { + "bbox": [ + 313, + 581, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{tot}}^{\\mathrm{real}} = \\mathcal{L}_N + \\mathcal{L}_M + \\mathcal{L}_R + \\mathcal{L}_{\\mathrm{TV}}" + }, + { + "bbox": [ + 313, + 581, + 556, + 713 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 313, + 581, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{tot}}^{\\mathrm{real}}" + }, + { + "bbox": [ + 313, + 581, + 556, + 713 + ], + "type": "text", + "content": " is a function of the illumination image " + }, + { + "bbox": [ + 313, + 581, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{I}}_{\\mathrm{e}}" + }, + { + "bbox": [ + 313, + 581, + 556, + 713 + ], + "type": "text", + "content": " and PBR renders " + }, + { + "bbox": [ + 313, + 581, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{I}}_{\\mathrm{a}}" + }, + { + "bbox": [ + 313, + 581, + 556, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 581, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{I}}_{\\mathrm{m}}" + }, + { + "bbox": [ + 313, + 581, + 556, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 581, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{I}}_{\\mathrm{r}}" + }, + { + "bbox": [ + 313, + 581, + 556, + 713 + ], + "type": "text", + "content": ", the loss provides a valuable supervisory signal for all volumetric predictors. Furthermore, using real training data narrows the domain gap between the training set and" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5864" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 269, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 269, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 269, + 84 + ], + "type": "text", + "content": "the target domain of Twinner, which are real images." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 93, + 296, + 210 + ], + "blocks": [ + { + "bbox": [ + 56, + 93, + 296, + 210 + ], + "lines": [ + { + "bbox": [ + 56, + 93, + 296, + 210 + ], + "spans": [ + { + "bbox": [ + 56, + 93, + 296, + 210 + ], + "type": "image", + "image_path": "ec178c20f5c817dfc35e49cecd77609ad75198bea25d62d559faa2112599a80a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 219, + 296, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 219, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 219, + 296, + 251 + ], + "type": "text", + "content": "Figure 5. StanfordORB Illumination Visualization. We visualize the illumination estimation of multiple baselines and our method on four different scenes from the StanfordORB dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 262, + 294, + 332 + ], + "blocks": [ + { + "bbox": [ + 56, + 262, + 294, + 332 + ], + "lines": [ + { + "bbox": [ + 56, + 262, + 294, + 332 + ], + "spans": [ + { + "bbox": [ + 56, + 262, + 294, + 332 + ], + "type": "table", + "html": "
MethodAngular Error ↓RMSE↓Norm. RMSE↓SI RMSE↓Time (hr.)↓
NMF [30]11.240.4870.5240.394~ 2
TensoIR [21]15.080.7520.7980.451~ 2
NVDiffRec [34]12.260.7230.7150.367~ 1
NVDiffRecMC [17]13.250.5740.6140.527~ 2
StyleLight [52]13.050.6400.6890.4400.011
DiffusionLight [38]11.350.5810.5690.5440.317
Twinner (ours)8.090.5700.6020.4720.006
", + "image_path": "622a218f0827ce7715b4e1df472f609bf427db347ed1a369dea9958049997275.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 340, + 296, + 384 + ], + "lines": [ + { + "bbox": [ + 55, + 340, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 296, + 384 + ], + "type": "text", + "content": "Table 2. Illumination prediction results. We provide a quantitative evaluation of the illumination estimation for baselines and our Twinner. Rows are grouped by single-scene optimization models (top) and predictive feed-forward models (bottom)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 396, + 137, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 137, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 137, + 410 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 416, + 295, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 295, + 452 + ], + "type": "text", + "content": "Evaluation data. We evaluate Twinner on the StanfordORB [26] dataset, which comprises 42 captures of 14 different objects, each imaged under 3 different illuminations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 453, + 296, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 296, + 488 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 296, + 488 + ], + "type": "text", + "content": "Each object has a high-quality geometric reconstruction obtained using a 3D scanner and comes with ground-truth camera poses and estimated illumination." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 493, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 295, + 601 + ], + "type": "text", + "content": "Evaluation Protocol. For each evaluation scene, we select four views from the set of training images as input frames via furthest camera-origin point sampling. With these input views, Twinner reconstructs a 3D PBR-textured mesh. Each predicted mesh is rendered with Blender [13] to held-out test views while being lit with a given ground-truth environment map. Following [26], the renders are compared to the corresponding ground-truth images using scale-invariant versions of LPIPS [68], SSIM [55], and PSNR." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "Baselines. We compare to existing feed-forward predictors of PBR-textured meshes. As such, we selected the state-of-the-art MetalLRM few-view reconstructor from AssetGen [45] as our main baseline. We use the version of MetalLRM conditioned on four shaded input views, as opposed to the albedo+shaded conditioning used in the AssetGen pipeline. We also compare to two ablations of our method. The first ablation, Twinner w/o Shading Loss, evaluates the contribution of training on real data and omits" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 556, + 443 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 556, + 443 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 556, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 556, + 443 + ], + "type": "image", + "image_path": "61b82d5a777513590a3905c372be15ff8005d05268b97e00ce561d11ae6aaf65.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 453, + 555, + 509 + ], + "lines": [ + { + "bbox": [ + 313, + 453, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 555, + 509 + ], + "type": "text", + "content": "Figure 6. Qualitative results on StanfordORB. We visualize the shaded RGB, geometry reconstruction, and material predictions for our baseline, two ablations, and our Twinner. Our model is able to better predict relightable objects thanks to the combination of curated synthetic training data and the shading loss on real data." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 526, + 555, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 555, + 669 + ], + "type": "text", + "content": "the training images of CO3Dv2. Second, we compare to Twinner w/o Zeroverse, which ablates training of the environment-map and PBR-texture predictor on our version of the procedurally-created Zeroverse. For completeness, we also include several single-scene optimization methods that \"overfit\" a relightable implicit shape in each scene from scratch. Note that, to save GPU memory, our feed-forward predictors randomly sample only four source views from the pool of all available training images of each scene. This puts us at a disadvantage compared to single-scene optimizers, as they leverage all available scene views for each reconstruction." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "Relighting Evaluation. All quantitative results are reported in Tab. 1. Twinner is the best among feed-forward reconstructors, outperforming its own two ablations and" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5865" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 558, + 337 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 558, + 337 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 558, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 558, + 337 + ], + "type": "image", + "image_path": "256b653bc1c34a5fbf1ec2bcd5798f5e2d08f59a8e7b5a26ec5e491d98af3cf7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 344, + 555, + 377 + ], + "lines": [ + { + "bbox": [ + 55, + 344, + 555, + 377 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 555, + 377 + ], + "type": "text", + "content": "Figure 7. CO3Dv2 Qualitative Results. We visualize predicted PBR meshes from the CO3Dv2 dataset illuminated under direct sunlight as well as their geometry, albedo, and material properties. The shading loss aids at predicting more realistic material properties, while the synthetic data improves the predicted geometry and albedo." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 388, + 295, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 509 + ], + "type": "text", + "content": "MetalLRM. For instance, Twinner achieves a " + }, + { + "bbox": [ + 55, + 388, + 295, + 509 + ], + "type": "inline_equation", + "content": "27\\%" + }, + { + "bbox": [ + 55, + 388, + 295, + 509 + ], + "type": "text", + "content": " reduction in both novel-scene relighting and novel-view synthesis LPIPS compared to the strongest baseline (MetalLRM), while maintaining comparable inference time. Remarkably, our method is on par with most single-scene optimization methods while being significantly faster, requiring only a few seconds per prediction compared to several hours for optimization-based methods. Examples of our predictions and the baselines are shown in Fig. 6. We also evaluate qualitatively on testing sequences of CO3Dv2 in Fig. 7." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 514, + 295, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 295, + 624 + ], + "type": "text", + "content": "Illumination Evaluation. We evaluate Twinner's ability to predict a scene's illumination from four input images. Since there are no existing baselines for few-view conditioned feed-forward illumination prediction, we compare to environment map estimates output by the single-scene optimizers in Tab. 1, as well as to single-view conditioned predictors. We follow [38] and evaluate illumination in terms of angular error, RMSE, normalized RMSE, and scale-invariant RMSE. Results are given in Tab. 2 and Fig. 5." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "Tricolumn Ablation. We also evaluate the effect of our tricolumn representation. In Tab. 3, we report the novel-view synthesis performance of the vanilla LightplaneLRM [11] model trained only on our synthetic dataset, utilizing either tricolumn or triplane representations. Tricolumn outperforms triplane significantly, achieving a " + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "inline_equation", + "content": "22\\%" + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": " improvement in LPIPS, validating this contribution." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 339, + 386, + 531, + 433 + ], + "blocks": [ + { + "bbox": [ + 339, + 386, + 531, + 433 + ], + "lines": [ + { + "bbox": [ + 339, + 386, + 531, + 433 + ], + "spans": [ + { + "bbox": [ + 339, + 386, + 531, + 433 + ], + "type": "table", + "html": "
3D representationPSNR↑SSIM↑LPIPS↓
Triplane31.200.9720.037
Tricolumn (ours)32.810.9660.029
", + "image_path": "d818ce187e297cc3b41af667793e7ac4c7b5aadcf0616e0a975f0b9abfa18b82.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 441, + 555, + 485 + ], + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 485 + ], + "type": "text", + "content": "Table 3. Tricolumn ablation evaluating novel view synthesis on the StanfordORB dataset for a LightplaneLRM [11] using the triplane 3D representation or our proposed tricolumn instead. Our tricolumn outperforms triplane on all reconstruction metrics." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 500, + 388, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 500, + 388, + 512 + ], + "spans": [ + { + "bbox": [ + 314, + 500, + 388, + 512 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": "We have presented Twinner, a new LRM capable of creating high-quality 3D assets with accurate geometry, material properties, and illumination from only a few posed images in a feed-forward manner. Twinner leverages the novel tricolumn representation, whose memory efficiency enables the processing of voxel grids with large transformers. For training, we demonstrated the power of generating synthetic PBR assets procedurally, addressing the scarcity of high-quality PBR assets in existing synthetic datasets. We also showed that, by predicting illumination, the model can learn from real-life datasets without requiring groundtruth illumination and material properties. We evaluated Twinner on the StanfordORB benchmark, where it outperformed existing feed-forward predictors and achieved comparable quality to optimization-based methods while being orders of magnitude faster." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5866" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 122 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 122 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 122 + ], + "type": "text", + "content": "[1] Sameer Agarwal, Noah Snavely, Ian Simon, Steven M. Seitz, and Richard Szeliski. Building rome in a day. In Proc. ICCV, 2009. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 124, + 295, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 124, + 295, + 157 + ], + "spans": [ + { + "bbox": [ + 61, + 124, + 295, + 157 + ], + "type": "text", + "content": "[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In Proc. CVPR, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 159, + 294, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 159, + 294, + 212 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 294, + 212 + ], + "type": "text", + "content": "[3] F O Bartell, E. L. Dereniak, and W. L Wolfe. The Theory And Measurement Of Bidirectional Reflectance Distribution Function (Brdf) And Bidirectional Transmittance Distribution Function (BTDF). In Proc. Radiation Scattering in Optical Systems, 1981. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 214, + 294, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 294, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 294, + 257 + ], + "type": "text", + "content": "[4] Sai Bi, Zexiang Xu, Pratul P. Srinivasan, B. Mildenhall, Kalyan Sunkavalli, Milovs Havsan, Yannick Hold-Geoffroy, D. Kriegman, and R. Ramamoorthi. Neural reflectance fields for appearance acquisition. arXiv, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 258, + 294, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 294, + 301 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 294, + 301 + ], + "type": "text", + "content": "[5] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Deep reflectance volumes: Relightable reconstructions from multi-view photometric images. In Proc. ECCV, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 303, + 294, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 303, + 294, + 344 + ], + "spans": [ + { + "bbox": [ + 62, + 303, + 294, + 344 + ], + "type": "text", + "content": "[6] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, David Kriegman, and Ravi Ramamoorthi. Deep 3D capture: Geometry and reflectance from sparse multi-view images. In Proc. CVPR, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 347, + 294, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 294, + 379 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 294, + 379 + ], + "type": "text", + "content": "[7] Mark Boss, Varun Jampani, Kihwan Kim, Hendrik P.A. Lensch, and Jan Kautz. Two-shot spatially-varying BRDF and shape estimation. In Proc. CVPR, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 380, + 294, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 294, + 434 + ], + "type": "text", + "content": "[8] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T. Barron, Ce Liu, and Hendrik P.A. Lensch. Nerd: Neural reflectance decomposition from image collections. In IEEE International Conference on Computer Vision (ICCV), 2021. 2, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 436, + 294, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 294, + 490 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 294, + 490 + ], + "type": "text", + "content": "[9] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan T. Barron, and Hendrik P.A. Lensch. Neuralpil: Neural pre-integrated lighting for reflectance decomposition. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 491, + 294, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 491, + 294, + 512 + ], + "spans": [ + { + "bbox": [ + 57, + 491, + 294, + 512 + ], + "type": "text", + "content": "[10] Brent Burley. Physically-based shading at disney. Technical report, Disney, 2012. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 514, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 294, + 546 + ], + "type": "text", + "content": "[11] Ang Cao, Justin Johnson, Andrea Vedaldi, and David Novotny. Lightplane: Highly-scalable components for neural 3d fields. arXiv, 2024. 2, 3, 4, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 548, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 548, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 548, + 294, + 612 + ], + "type": "text", + "content": "[12] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In Proc. CVPR, 2022. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 614, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 294, + 645 + ], + "type": "text", + "content": "[13] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 647, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 690 + ], + "type": "text", + "content": "[14] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In Proc. CVPR, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "text", + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 116 + ], + "type": "text", + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth " + }, + { + "bbox": [ + 335, + 73, + 553, + 116 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 335, + 73, + 553, + 116 + ], + "type": "text", + "content": " words: Transformers for image recognition at scale. In Proc. ICLR, 2021. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "text", + "content": "[16] Yasutaka Furukawa and Jean Ponce. Accurate, dense, and robust multiview stereopsis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(8):1362-1376, 2010. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "type": "text", + "content": "[17] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using monte carlo rendering and denoising. In Advances in Neural Information Processing Systems, pages 22856-22869. Curran Associates, Inc., 2022. 3, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 220, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 220, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 220, + 553, + 262 + ], + "type": "text", + "content": "[18] Philipp Henzler, Jeremy Reizenstein, Patrick Labatut, Roman Shapovalov, Tobias Ritschel, Andrea Vedaldi, and David Novotny. Unsupervised learning of 3d object categories from videos in the wild. CVPR, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 264, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 264, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 317, + 264, + 553, + 307 + ], + "type": "text", + "content": "[19] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. In Proc. ICLR, 2024. 2, 3, 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 309, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 309, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 317, + 309, + 553, + 352 + ], + "type": "text", + "content": "[20] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. GaussianShader: 3D Gaussian splattering with shading functions for reflective surfaces. arXiv.cs, abs/2311.17977, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 354, + 555, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 555, + 408 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 555, + 408 + ], + "type": "text", + "content": "[21] Haian Jin, Isabella Liu, Peijia Xu, Xiaoshuai Zhang, Songfang Han, Sai Bi, Xiaowei Zhou, Zexiang Xu, and Hao Su. Tensoir: Tensorial inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 410, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 410, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 317, + 410, + 553, + 441 + ], + "type": "text", + "content": "[22] Brian Karis and Epic Games. Real shading in unreal engine 4. Proc. Physically Based Shading Theory Practice, 4(3):1, 2013. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 444, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 444, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 444, + 553, + 487 + ], + "type": "text", + "content": "[23] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the Fourth Eurographics Symposium on Geometry Processing, page 61-70, Goslar, DEU, 2006. Eurographics Association. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 489, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 489, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 489, + 553, + 521 + ], + "type": "text", + "content": "[24] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian Splating for real-time radiance field rendering. Proc. SIGGRAPH, 42(4), 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 523, + 553, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 553, + 554 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 553, + 554 + ], + "type": "text", + "content": "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 557, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 553, + 599 + ], + "type": "text", + "content": "[26] Zhengfei Kuang, Yunzhi Zhang, Hong-Xing Yu, Samir Agarwala, Shangzhe Wu, and Jiajun Wu. Stanford-orb: A real-world 3d object inverse rendering benchmark, 2023. 2, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 601, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 553, + 644 + ], + "type": "text", + "content": "[27] Ruofan Liang, Huiting Chen, Chunlin Li, Fan Chen, Selvakumar Panneer, and Nandita Vijaykumar. Envidr: Implicit differentiable renderer with neural environment lighting. In Proc. ICCV, 2023. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 647, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 678 + ], + "type": "text", + "content": "[28] Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. GS-IR: 3D Gaussian splattering for inverse rendering. arXiv.cs, abs/2311.16473, 2023. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "text", + "content": "[29] W. Lorensen and H. Cline. Marching cubes: A high resolution 3D surface construction algorithm. ACM Computer Graphocs, 21(24), 1987. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5867" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[30] Alexander Mai, Dor Verbin, Falko Kuester, and Sara Fridovich-Keil. Neural microfacet fields for inverse rendering. In Proc. ICCV, 2023. 3, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 149 + ], + "type": "text", + "content": "[31] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Proc. ECCV, 2020. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "text", + "content": "[32] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P. Srinivasan, and Jonathan T. Barron. Nerf in the dark: High dynamic range view synthesis from noisy raw images. In Proc. CVPR, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 194, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 227 + ], + "type": "text", + "content": "[33] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. In Proc. SIGGRAPH, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 271 + ], + "type": "text", + "content": "[34] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting Triangular 3D Models, Materials, and Lighting From Images. In Proc. CVPR, 2022. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 272, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 304 + ], + "type": "text", + "content": "[35] Giljoo Nam, J. Lee, D. Gutierrez, and Min H. Kim. Practical SVBRDF acquisition of 3D objects with unstructured flash photography. arXiv, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 305, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 294, + 414 + ], + "type": "text", + "content": "[36] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 294, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 436 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 436 + ], + "type": "text", + "content": "[37] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proc. ICCV, 2023. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 502 + ], + "type": "text", + "content": "[38] Pakkapon Phongthawee, Worameth Chinchuthakun, Nontaphat Sinsunthithet, Varun Jampani, Amit Raj, Pramook Khungurn, and Supasorn Suwajanakorn. Diffusionlight: Light probes for free by painting a chrome ball. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 98-108, 2024. 3, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 503, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 294, + 536 + ], + "type": "text", + "content": "[39] PolyHaven. Polyhaven: curated public asset library for visual effects artists and game designers. https://polyhaven.com/, 2024. Accessed: 2024-11-07. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 536, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 536, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 536, + 294, + 591 + ], + "type": "text", + "content": "[40] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In International Conference on Computer Vision, 2021. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "text", + "content": "[41] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common Objects in 3D: Large-scale learning and evaluation of real-life 3D category reconstruction. In Proc. ICCV, 2021. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 657 + ], + "type": "text", + "content": "[42] Shen Sang and Manmohan Chandraker. Single-shot neural relighting and SVBRDF estimation. In Proc. ECCV, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 679 + ], + "type": "text", + "content": "[43] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proc. CVPR, 2016. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "text", + "content": "[44] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In Proc. ECCV, 2016. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 149 + ], + "type": "text", + "content": "[45] Yawar Siddiqui, Filippos Kokkinos, Tom Monnier, Mahendra Kariya, Yanir Kleiman, Emilien Garreau, Oran Gafni, Natalia Neverova, Andrea Vedaldi, Roman Shapovalov, and David Novotny. Meta 3D Asset Gen: Text-to-mesh generation with high-quality geometry, texture, and PBR materials. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2024. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 150, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 553, + 194 + ], + "type": "text", + "content": "[46] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Frédo Durand. Light field networks: Neural scene representations with single-evaluation rendering. In Proc. NeurIPS, 2021. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 195, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 228 + ], + "type": "text", + "content": "[47] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proc. CVPR, 2022. 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 229, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 229, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 229, + 553, + 295 + ], + "type": "text", + "content": "[48] Cheng Sun, Guangyan Cai, Zhengqin Li, Kai Yan, Cheng Zhang, Carl Marshall, Jia-Bin Huang, Shuang Zhao, and Zhao Dong. Neural-pbir reconstruction of shape, material, and illumination. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 18000-18010, 2023. 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 297, + 553, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 340 + ], + "type": "text", + "content": "[49] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. LGM: Large multi-view Gaussian model for high-resolution 3D content creation. arXiv, 2402.05054, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "type": "text", + "content": "[50] Giuseppe Vecchio and Valentin Deschaintre. Matsynth: A modern pbr materials dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22109-22118, 2024. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 387, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 553, + 430 + ], + "type": "text", + "content": "[51] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd E. Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In Proc. CVPR, 2022. 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 431, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 431, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 431, + 553, + 475 + ], + "type": "text", + "content": "[52] Guangcong Wang, Yinuo Yang, Chen Change Loy, and Zwei Liu. Stylelight: HDR panorama generation for lighting estimation and editing. In European Conference on Computer Vision (ECCV), 2022. 3, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 476, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 476, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 476, + 553, + 510 + ], + "type": "text", + "content": "[53] Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. VGGsF: visual geometry grounded deep structure from motion. In Proc. CVPR, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 510, + 553, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 554 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 554 + ], + "type": "text", + "content": "[54] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv.cs, abs/2106.10689, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 555, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 599 + ], + "type": "text", + "content": "[55] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. on Image Processing, 13(4), 2004. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 600, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 600, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 600, + 553, + 644 + ], + "type": "text", + "content": "[56] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. MeshLRM: large reconstruction model for high-quality mesh. arXiv, 2404.12385, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "type": "text", + "content": "[57] Desai Xie, Sai Bi, Zhixin Shu, Kai Zhang, Zexiang Xu, Yi Zhou, Soren Pirk, Arie Kaufman, Xin Sun, and Hao Tan. Lrm-zero: Training large reconstruction models with synthesized data, 2024. 5, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 690, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 553, + 712 + ], + "type": "text", + "content": "[58] Desai Xie, Sai Bi, Zhixin Shu, Kai Zhang, Zexiang Xu, Yi Zhou, Soren Pirk, Arie Kaufman, Xin Sun, and Hao Tan." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5868" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 609 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "LRM-Zero: training large reconstruction models with synthesized data. arXiv, 2406.09371, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "text", + "content": "[59] Zhang Xiuming, Srinivasan Pratul P., Deng Boyang, Debevec Paul, Freeman William T., and Barron Jonathan T. NeRFactor: neural factorization of shape and reflectance under an unknown illumination. In Proc. SIGGRAPH, 2021. 2, 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 205 + ], + "type": "text", + "content": "[60] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. GRM: Large gaussian reconstruction model for efficient 3D reconstruction and generation. arXiv, 2403.14621, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "text", + "content": "[61] Yinghao Xu, Hao Tan, Fujun Luan, Sai Bi, Peng Wang, Jiahao Li, Zifan Shi, Kalyan Sunkavalli, Gordon Wetzstein, Zexiang Xu, and Kai Zhang. DMV3D: Denoising multiview diffusion using 3D large reconstruction model. In Proc. ICLR, 2024. 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 296 + ], + "type": "text", + "content": "[62] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. arXiv.cs, abs/2106.12052, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 329 + ], + "type": "text", + "content": "[63] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. PixelNeRF: Neural radiance fields from one or few images. In Proc. CVPR, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 331, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 364 + ], + "type": "text", + "content": "[64] Jesus Zarzar and Bernard Ghanem. Splitnerf: Split sum approximation neural field for joint geometry, illumination, and material estimation. arXiv.cs, abs/2311.16671, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "text", + "content": "[65] Jianzhao Zhang, Guojun Chen, Yue Dong, Jian Shi, Bob Zhang, and Enhua Wu. Deep inverse rendering for practical object appearance scan with uncalibrated illumination. In Proc. Advances in Computer Graphics, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "text", + "content": "[66] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snively. PhySG: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 466, + 294, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 294, + 508 + ], + "type": "text", + "content": "[67] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. GS-LRM: large reconstruction model for 3D Gaussian splatting. arXiv, 2404.19702, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 510, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 294, + 553 + ], + "type": "text", + "content": "[68] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proc. CVPR, pages 586-595, 2018. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 555, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 294, + 609 + ], + "type": "text", + "content": "[69] Jingsen Zhu, Fujun Luan, Yuchi Huo, Zihao Lin, Zhihua Zhong, Dianbing Xi, Rui Wang, Hujun Bao, Jiaxiang Zheng, and Rui Tang. Learning-based inverse rendering of complex indoor scenes with differentiable monte carlo raytracing. In SIGGRAPH Asia 2022 Conference Papers. ACM, 2022. 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5869" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Two by Two_ Learning Multi-Task Pairwise Objects Assembly for Generalizable Robot Manipulation/57fa886e-a784-460f-94a7-de5f04b8ab96_content_list.json b/2025/Two by Two_ Learning Multi-Task Pairwise Objects Assembly for Generalizable Robot Manipulation/57fa886e-a784-460f-94a7-de5f04b8ab96_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..67c5e0708c36177999e828d3980e968adf2c6329 --- /dev/null +++ b/2025/Two by Two_ Learning Multi-Task Pairwise Objects Assembly for Generalizable Robot Manipulation/57fa886e-a784-460f-94a7-de5f04b8ab96_content_list.json @@ -0,0 +1,1645 @@ +[ + { + "type": "text", + "text": "Two by Two: Learning Multi-Task Pairwise Objects Assembly for Generalizable Robot Manipulation", + "text_level": 1, + "bbox": [ + 156, + 122, + 813, + 178 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yu Qi $^{2,1*}$ Yuanchen Ju $^{1,3*}$ Tianming Wei $^{1,4}$ Chi Chu $^{1}$ Lawson L.S. Wong $^{2}$ Huazhe Xu $^{1,3,5\\dagger}$ $^{1}$ Shanghai Qi Zhi Institute $^{2}$ Northeastern University $^{3}$ IIIS, Tsinghua University \n $^{4}$ Shanghai Jiao Tong University $^{5}$ Shanghai AI Laboratory", + "bbox": [ + 86, + 205, + 911, + 260 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a5cbc2fb8e59f02077f4a25321f169f156b49bd3b50d7cf4ec0f05d3794fa9a5.jpg", + "image_caption": [ + "Figure 1. Overview of the 2BY2 Dataset. We propose the first large-scale daily pairwise object assembly dataset 2BY2, which contains 1,034 instances and 517 pairwise objects with pose and symmetry annotations." + ], + "image_footnote": [], + "bbox": [ + 78, + 270, + 890, + 539 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 599, + 313, + 614 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D assembly tasks, such as furniture assembly and component fitting, play a crucial role in daily life and represent essential capabilities for future home robots. Existing benchmarks and datasets predominantly focus on assembling geometric fragments or factory parts, which fall short in addressing the complexities of everyday object interactions and assemblies. To bridge this gap, we present 2BY2, a large-scale annotated dataset for daily pairwise objects assembly, covering 18 fine-grained tasks that reflect real-life scenarios, such as plugging into sockets, arranging flowers in vases, and inserting bread into toasters. 2BY2 dataset includes 1,034 instances and 517 pairwise objects with pose and symmetry annotations, requiring approaches that align geometric shapes while accounting for functional and spatial relationships between objects. Leveraging the 2BY2 dataset, we propose a two-step SE(3) pose estimation method with equivariant features for assembly constraints. Compared to previous shape assembly methods, our approach achieves state-of-the-art", + "bbox": [ + 75, + 628, + 473, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "performance across all 18 tasks in the 2BY2 dataset. Additionally, robot experiments further validate the reliability and generalization ability of our method for complex 3D assembly tasks. More details and demonstrations can be found at https://tea-lab.github.io/TwoByTwo/.", + "bbox": [ + 496, + 599, + 895, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 695, + 632, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Assembly tasks are ubiquitous, such as assembling furniture, repairing household appliances, or putting together electronics. Successfully completing these tasks requires precise reasoning about the spatial relationships between pairs of objects. For robots to assist in these activities, they need to accurately estimate the 6D pose of each objects—including both their orientation and position in space. This capability is essential for domestic robots to help humans with various tasks, as it enables them to interact with their environment in a meaningful way.", + "bbox": [ + 496, + 720, + 893, + 871 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Daily object pairwise assembly not only requires considering the geometric constraints and spatial relationships", + "bbox": [ + 498, + 872, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17383", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "between objects to achieve precise alignment but also needs to exhibit a certain level of generalization. Existing methods and benchmarks for solving assembly problems [31, 32, 49, 50, 63], typically focus on matching local geometric shapes, which often results in suboptimal performance in everyday assembly scenarios which require semantic and spatial alignment. This is because they are primarily trained and tested on existing assembly datasets that consist of large-scale geometric fragments, such as Breaking Bad [41] and Neural Shape Mating [5]. Compared with existing assembly tasks that focus on putting together fractures of objects, daily pairwise assembly tasks are more challenging and hold greater practical significance in human life.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To bridge this gap, we introduce 2BY2, the first large-scale daily pairwise assembly dataset comprising 18 fine-grained tasks, shown as Figure 1. Compared to previous datasets and benchmarks, 2BY2 contains 1,034 instances and 517 pairwise objects with pose and symmetry annotations, covering a variety of pairwise assembly tasks reflecting everyday scenarios, which require approaches that align geometric shapes while accounting for functional and spatial relationships between objects, as shown in Table 1.", + "bbox": [ + 75, + 287, + 472, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on this dataset, we propose a two-step pairwise network architecture for assembly tasks. Mimicking the human assembly process like we firstly put the vase on the table and then arrange flower in it, our approach predicts the pose of each object in a step-by-step manner to assemble them to a predefined canonical space, which refers to a standard coordinate system that aligns with the principles of the human world, with detailed definition in Section 3.2. The network leverages a custom two-scale Vector Neuron DGCNN [9] encoder with spherical convolution [7] to extract SE(3) equivariant and SO(3) invariant features from point cloud inputs. Additionally, a feature fusion module and a two-step training and evaluation strategy are used to improve pose prediction accuracy.", + "bbox": [ + 75, + 426, + 472, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate our approach on 18 tasks in 2BY2 dataset to demonstrate the effectiveness on multi-task object pairwise assembly prediction. Compared to existing baselines, our method achieves an average improvement of 0.046 in translation RMSE and 8.97 in rotation RMSE. Moreover, we validate the effectiveness of our approach on three multicategory task, namely Lid Covering, Inserting and High Precision Placing, as well as the All task, which is defined in Section 2. Besides, real-world robot experiments validate the practical applicability of our approach.", + "bbox": [ + 75, + 638, + 472, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are listed as follows:", + "bbox": [ + 96, + 792, + 392, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. We introduce 2BY2, the first large-scale daily pairwise object assembly dataset. By providing comprehensive pose and symmetry annotations for 517 pairwise objects across 18 fine-grained tasks, 2BY2 pushes the boundaries of real-world 3D assembly challenges and establishes a benchmark for pairwise assembly tasks.", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2. Our two-step pairwise SE(3) pose estimation method, leveraging equivariant geometric features, demonstrates superior performance compared to existing shape assembly methods, significantly reducing translation and rotation errors and enhancing the accuracy of 6D pose estimation.", + "3. Our approach achieves state-of-the-art performance on the benchmark, with real-world robot experiments demonstrating its capability, providing a generalizable solution for robot manipulation using pairwise object assembly." + ], + "bbox": [ + 496, + 90, + 893, + 228 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 498, + 241, + 640, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Object Assembly Benchmarks and Datasets", + "text_level": 1, + "bbox": [ + 498, + 266, + 870, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Object reassembly has led to various datasets in computer vision and robotics. In computer vision, datasets like AutoMate [24] and JoinABLe [51] focus on reassembling fragments using geometric clues, while early datasets [3, 12, 21, 42] were limited in scale. Recent efforts, such as Neural Shape Mating [5] and Breaking Bad [41], generate largescale fractured object data using parametric segmentation. In robotics, benchmarks like Factory [35], RLBench [23], and RoboSuite [68] lack diverse shapes and assembly tasks under varying initial poses. In contrast, our dataset includes over 500 diverse object pairs across 3 categories and 18 assembly tasks, providing a comprehensive benchmark for pairwise object assembly, supporting the development of generalizable methods for real-world applications.", + "bbox": [ + 496, + 289, + 893, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2.3D Shape Assembly", + "text_level": 1, + "bbox": [ + 498, + 510, + 687, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D shape assembly [11, 30, 33, 56, 61], also known as part assembly, involves reconstructing objects from fragments, such as shattered sculptures or disassembled furniture. Existing methods use graphical models [4, 22, 26] and neural networks [6, 25, 28, 47, 52, 57, 58, 60, 64, 66] to capture geometric and semantic relationships. Approaches such as [5, 36, 62, 63] focus on pose estimation and part assembly without relying on predefined semantic information. Few-shot learning has been applied to assembly tasks [29], while jigsaw puzzle techniques [32, 37] leverage shape completion strategies. Recent works [16, 20, 40, 50] utilize diffusion models to refine poses or point clouds for assembly. In contrast, our method introduces a two-step pairwise network for step-by-step assembly, tailored to pairwise object alignment.", + "bbox": [ + 496, + 534, + 893, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3.6D Pose Estimation for Robot Manipulation", + "text_level": 1, + "bbox": [ + 498, + 756, + 872, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "6D pose estimation is crucial in robotics and computer vision for object interaction in unstructured environments [15, 45, 55]. Early handcrafted feature-based methods struggled in cluttered scenes [14, 15], while CNN-based approaches improved performance but lacked generalization [27, 55]. Domain randomization enhances robustness by varying synthetic datasets [44, 45]. In assembly tasks, 6D pose estimation aids manipulation planning with predefined", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17384", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/cbaab6546e4e6c458c1f489d1f5394cc71ecaa7d0b733b9bba15817f0a640d24.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset#OC#OSTask NumberPairTask HierarchyEveryday ScenarioSymmetryAssemble Type
PartNet [34]2426,671-NoNoNoNoSemantic
AutoMate [24]292,5291NoNoNoNoGeometric
JoinABLe [51]68,2511NoNoNoNoSemantic
NSM dataset[5]111,2461YesNoNoNoGeometric
Breaking Bad[41]-10,4741NoNoNoNoGeometric
Factory[35]8608YesNoNoNoGeometric
2BY2 Dataset (Ours)36103418YesYesYesYesGeometric and Semantic
", + "bbox": [ + 81, + 88, + 893, + 204 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Dataset Comparison. We compare 2BY2 dataset with existing datasets and benchmarks. #OC stands for the number of object categories. #OS stands for the number of object shapes. Pair denotes whether the dataset is pairwise. Task Number refers to the number of distinct assembly tasks, with the assembly of fractured pieces considered as a single task. Task Hierarchy stands for the different categories of task from coarse to fine, with ours shown in Section 3.1. Everyday Scenario means whether the assemble task has practical significance in real-world human applications. Symmetry denotes whether the dataset contains part symmetry annotation.", + "bbox": [ + 75, + 214, + 893, + 285 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "objects [39, 46]. Like [13, 17-20, 43, 59, 65, 67], our method leverages equivariant features for efficient 6D pose learning and improved generalization.", + "bbox": [ + 75, + 310, + 470, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2BY2 Dataset", + "text_level": 1, + "bbox": [ + 76, + 368, + 217, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2BY2 Dataset Overview", + "text_level": 1, + "bbox": [ + 76, + 392, + 295, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present the first large-scale 3D pairwise object assembly dataset for everyday scenarios, with detailed annotations for each object pair. The meshes in our dataset come from 3D Warehouse [1], SAPIEN PartNet-Mobility [54], Google SketchUp 3D Challenge [2], and Objaverse [8]. These meshes are manually paired, cleaned, annotated, and scaled uniformly. The 2BY2 dataset contains 517 unique pairs across three main tasks: Lid Covering, Inserting, and High Precision Placing, further subdivided into multiple subcategories, as shown in Table 2.", + "bbox": [ + 75, + 415, + 470, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Data Annotation", + "text_level": 1, + "bbox": [ + 76, + 575, + 241, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To ensure high quality and reliability of our dataset, we conducted systematic cleaning and annotation of the collected meshes. First, we manually segment, integrate, and pair the meshes, classify them into Object $B$ and Object $A$ . Object $B$ is the base or the receiving component, such as the nut, the vase, the postbox. Object $A$ , is the fitting component, such as the bolt, the flower, the mail. This classification aligns with intuitive human assembly logic and supports our network's prediction strategy, such as positioning a nut before the bolt, as detailed in Sections 4 and 5.1. Automated scripts were used to uniformly scale meshes and align each pair to a canonical pose in world frame, defined as the object resting stably on the $XY$ plane with its lowest point aligned to $Z = 0$ . For instance, bottles and vases are aligned as if placed on a table, and mailboxes on the ground.", + "bbox": [ + 75, + 598, + 470, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During point cloud generation, we utilized blue noise sampling method [41] to extract point clouds uniformly from each mesh surface with dimension (1024, 3). We also annotated each object category with its inherent symmetry properties, specifically considering rotational symmetry along the", + "bbox": [ + 75, + 825, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$Z$ -axis, such as bottles, screws, and mirror symmetry along the $X$ -axis, such as bread, letters.", + "bbox": [ + 496, + 310, + 890, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Data Division and Task Diversity Analysis", + "text_level": 1, + "bbox": [ + 498, + 348, + 857, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our dataset provides diverse task coverage across categories, with each further divided into specific sub-categories, see Table 2. Objects within each category vary in shape, size, and type. To enhance generalization, the testing set includes objects with unseen geometric shapes, as shown in Figure 3. We also compute Chamfer Distance on point clouds between training and testing sets to quantify geometry differences, as shown in Figure 2. This diversity ensures generalization ability and applicability in real world scenarios and supports robust 3D matching and assembly tasks.", + "bbox": [ + 496, + 372, + 893, + 523 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fa84678daecd4f114ab87ad5c863115c41885046160521308f804e83832416b9.jpg", + "image_caption": [ + "Figure 2. Chamfer Distance Between Training and Testing Set. We normalize point clouds and compute the Chamfer Distance. For each task we calculate the distance separately between point cloud of Object $A$ and Object $B$ in the training set and test set." + ], + "image_footnote": [], + "bbox": [ + 504, + 535, + 890, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Problem Formulation", + "text_level": 1, + "bbox": [ + 498, + 768, + 705, + 784 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The task takes two point clouds as input, namely $\\mathcal{P}_A$ and $\\mathcal{P}_B$ , each with dimension (1024, 3). These point clouds are derived from objects $\\mathcal{O}_A$ and $\\mathcal{O}_B$ from predefined canonical pose, as detailed in 3.2, respectively, and is randomly augmented with SO(3) rotation and being translated to its centroid. The desired output would be two individual SE(3) pose two assemble $\\mathcal{O}_A$ and $\\mathcal{O}_B$ to the canonical pose.", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17385", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/f0800a33a2e4796a3a8670c3dfe9d6285a9a4319f0afbdf2a9ded4d16a6f7add.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskLid CoveringInsertingHigh Precision Placing
KitBotKetCofCupPluChiLetBreNutCoiKeyUsbBoxTisFloTeaPos
Pair Num248628261614193224202120202520604221
", + "bbox": [ + 81, + 88, + 890, + 154 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2. 2BY2 Dataset Statistics Overview. The figure presents the number of object pairs across all task categories in the 2BY2 dataset, where each pair consists of two unique objects. The first row categorizes tasks into three major groups, while the second row provides a detailed breakdown of specific task categories. Specifically, Kit = Kitchenport, Bot = Bottle, Ket = Kettle, Cof = Coffee machine, Cup = Cup, Plu = Plug into socket, Chi = Children's toy, Let = Letter into mailbox, Bre = Bread into toaster, Nut = Bolt into nut, Coi = Coin into piggy bank, Key = Key into lock, Usb = USB cap, Box = Shoe boxing, Tis = Tissue placement on rack, Flo = Flower into vase, Tea = Teaware arrangement on tray, Pos = Positioning a cup on the coffee machine for coffee dispensing.", + "bbox": [ + 75, + 164, + 893, + 250 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0d3ba6708fc5f95c010232ecbee4aa425fc8af00308e3d5b7b04880c20e67bd3.jpg", + "image_caption": [ + "Figure 3. Task Diversity Visualization. The image shows selected objects from four different tasks: USB, Bottle, Letter, and Plug in Socket. On the left are the objects selected on training set, and on the right is the testing set. As seen in the legend, object geometry varies in both the training and testing set, with the testing set containing novel shapes not seen in the training set." + ], + "image_footnote": [], + "bbox": [ + 81, + 272, + 272, + 464 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/aa6cd26a3d7f7e604524c8ec7094bb194075a6c98b0b8179ba5f33a465875c90.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 272, + 467, + 464 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Method", + "text_level": 1, + "bbox": [ + 76, + 588, + 169, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. Two-step Pairwise Network Architecture", + "text_level": 1, + "bbox": [ + 76, + 614, + 426, + 631 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To effectively learn pairwise object assembly, we propose a two-step pairwise network architecture with two branches: Branch B $(\\mathcal{B}_{\\mathcal{B}})$ and Branch A $(\\mathcal{B}_{\\mathcal{A}})$ , as shown in Figure 4. Branch B predicts the pose of $\\mathcal{P}_{\\mathcal{B}}$ , which is the socket, using a two-scale Vector Neuron DGCNN encoder [9] to extract SE(3) equivariant features, denoted as $\\mathcal{E}_{\\mathcal{B}}$ , followed by MLP-based pose prediction heads for translation and rotation. The transformed $\\mathcal{P}_{\\mathcal{B}}$ and inserter object $\\mathcal{P}_{\\mathcal{A}}$ , which is the plug, are then passed to Branch A, which extracts SE(3) equivariant features, denoted as $\\mathcal{E}_{\\mathcal{A}}$ and SO(3) invariant features $(\\mathcal{I}_{\\mathcal{B}})$ . The features are fused through element-wise multiplication, allowing $\\mathcal{B}_{\\mathcal{A}}$ to predict the pose of $\\mathcal{P}_{\\mathcal{A}}$ using information from both objects. This architecture ensures geometric alignment and matching by leveraging shared feature representations while reducing feature interference.", + "bbox": [ + 75, + 641, + 472, + 867 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our two-step pairwise network is inspired by the human approach to pairwise assembly tasks. For example, when ar-", + "bbox": [ + 76, + 869, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ranging a vase with flowers, one intuitively first positions the vase correctly before placing the flowers inside. Similarly, inserting an envelope into a mailbox requires identifying the mailbox slot's pose first. By mimicking this sequential strategy, our model simulates human decision-making process, enabling more efficient and accurate assembly tasks.", + "bbox": [ + 496, + 273, + 893, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.2. Two-scale SE(3) Equivariant and SO(3) Invariant Feature Extraction", + "text_level": 1, + "bbox": [ + 498, + 376, + 890, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We employ a two-scale SE(3) Vector Neuron DGCNN, an enhanced variant of the original Vector Neuron DGCNN [9], as our encoder to extract SE(3) equivariant and SO(3) invariant features. This architecture leverages equivariance to improve sample efficiency of the model, while incorporating a two-scale information fusion mechanism to capture geometric features at two different scales.", + "bbox": [ + 496, + 415, + 893, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SE(3) Equivariance and SO(3) Invariance. SE(3) equivariance combines SO(3) rotation and T(3) translation equivariance: rotation equivariance ensures that a network's output rotates with the input, while translation equivariance shifts the output accordingly. SO(3) invariance means the network's output remains unchanged under any 3D rotation. By leveraging SE(3) equivariance, the model benefits from improved sample efficiency and generalization. This is particularly advantageous for assembly tasks, where objects may appear in arbitrary poses.", + "bbox": [ + 496, + 521, + 893, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vector Neuron DGCNN. The Vector Neuron Network [9] extends traditional neurons from scalars to 3D vectors, designs vector-based convolutional layers and nonlinear functions like pooling and ReLU to support SO(3) equivariant and SO(3) invariant feature extraction. VNN operates in vector space, captures richer geometric relationships and ensures more robust feature representations for downstream tasks.", + "bbox": [ + 496, + 672, + 893, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Two-scale Vector Neuron DGCNN. We propose a two-scale Vector Neuron DGCNN for extracting SE(3) equivariant and SO(3) invariant features $\\mathcal{E}_{\\mathcal{B}}$ , $\\mathcal{I}_{\\mathcal{B}}$ , and $\\mathcal{E}_{\\mathcal{A}}$ . As shown in Figure 4, the encoder comprises two branches with different $K$ values, each consisting of multiple Vector Neuron convolutional layers followed by pooling. The extracted features from both branches are concatenated and further pro", + "bbox": [ + 496, + 795, + 895, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17386", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/426fdc6ea95e7999c713f3f9b402dc937e672444a66d44b20ab891b96882fccb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 89, + 514, + 262 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ec389a409c21357f0ba92a5925b7ea26c9a8c1ce5489b63d37b43bee8c8b2ea3.jpg", + "image_caption": [ + "Inference Stage" + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 723, + 402 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ee017c079a0e93900d5a1be86390829fd704ea4a7869d6b55c1803bc4b88ddf6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 725, + 90, + 880, + 166 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/becc995c74a834c658d0cd186207979920259cc0fb5358cd4c2df100e69d8e30.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 727, + 170, + 880, + 244 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/15b00f593880eaf4dbb0ba131eed6e52d0012e5b76d1cd0cecc6c8de59b64b83.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 727, + 248, + 880, + 323 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/288ef862b3bdd968b1a0c42c80bfdfd0b3a1d4a8c5910493018c8f8e55f46ec0.jpg", + "image_caption": [ + "Robot Experiments", + "Figure 4. Our Two-Step Pairwise Network. We utilize two-scale VN DGCNN as our encoder to extract equivariant and invariant feature. We first predict the canonical pose of $\\mathcal{O}_B$ and then predict the pose of $\\mathcal{O}_A$ according to it." + ], + "image_footnote": [], + "bbox": [ + 727, + 325, + 880, + 401 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/19b59556fd9937072e7e0aed90b18510e3e2fcba159b3febf5b5fda5dc08897f.jpg", + "image_caption": [ + "Training Stage" + ], + "image_footnote": [], + "bbox": [ + 78, + 263, + 514, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "cessed through an additional Vector Neuron convolutional layer. Point clouds $\\mathcal{P}_{\\mathcal{B}}$ and $\\mathcal{P}_{\\mathcal{A}}$ are independently processed, forming graphs that propagate through both branches.", + "bbox": [ + 75, + 496, + 470, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The SO(3) rotation equivariance of our encoder is ensured by the inherent equivariant properties of the Vector Neuron layers. To achieve T(3) translation equivariance, with an input point cloud $P = (p_{1}, p_{2}, \\dots, p_{n})$ , $p_{i} \\in R^{3}$ , we compute its centroid $x = (\\Sigma_{i=1}^{n} p_{i}) / n$ , and get the input point cloud as $P' = P - x$ . In this way, our prediction is T(3) translation equivariant, i.e., $f$ is our encoder and $\\mathcal{P}$ is the original point cloud.", + "bbox": [ + 75, + 542, + 468, + 661 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf (\\mathcal {P} + \\mathcal {T}) = f (\\mathcal {P}) + \\mathcal {T}, \\quad T \\in R ^ {3} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 672, + 468, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our two-scale VN DGCNN employs dual K-nearest neighbor (KNN) values to extract features across two distinct scales, enhancing its ability to capture both local and global information. This pyramid structure enables the network to simultaneously grasp overall object shapes and fine-grained details, improving feature extraction.", + "bbox": [ + 75, + 702, + 468, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.3. Cross Object Fusion Module", + "text_level": 1, + "bbox": [ + 76, + 801, + 333, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We utilize point-wise multiplication, shown in Figure 4, as our cross object fusion module designed in $B_A$ . We fuse the feature of $\\mathcal{P}_B$ and $\\mathcal{P}_A$ by multiplying $\\mathcal{I}_B$ and $\\mathcal{E}_A$ , so that each point in $\\mathcal{P}_A$ will have the geometry feature of both $\\mathcal{P}_A$ and $\\mathcal{P}_B$ . This approach integrates the geometric feature of", + "bbox": [ + 75, + 824, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathcal{P}_B$ in each point while preserving the rotation equivariance of $\\mathcal{P}_A$ , i.e., $f$ is an equivariant neural network, $R$ is random rotation matrix,", + "bbox": [ + 496, + 496, + 892, + 542 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf \\left(R \\cdot \\left(\\mathcal {I} _ {\\mathcal {B}} * \\mathcal {E} _ {\\mathcal {A}}\\right)\\right) = R \\cdot f \\left(\\mathcal {I} _ {\\mathcal {B}} * \\mathcal {E} _ {\\mathcal {A}}\\right), \\quad R \\in R ^ {3 \\times 3} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 554, + 892, + 573 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.4. Pose Prediction", + "text_level": 1, + "bbox": [ + 500, + 585, + 655, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At both branches, we utilize two separate MLPs as our pose prediction head, to separately predict the translation $T \\in R^3$ and rotation $R \\in R^{3 \\times 3}$ . Compared to predicting translation and rotation within a single prediction head, this approach helps mitigate the issue of differing convergence speeds between the two components.", + "bbox": [ + 496, + 608, + 892, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.5. Training and Evaluation Strategy", + "text_level": 1, + "bbox": [ + 498, + 710, + 794, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adopt a separate training and evaluation strategy for our network. To minimize the impact of pose prediction errors of $\\mathcal{P}_{\\mathcal{B}}$ on $\\mathcal{P}_{\\mathcal{A}}$ , we train $\\mathcal{B}_{\\mathcal{A}}$ and $\\mathcal{B}_{\\mathcal{B}}$ independently. Specifically, for $\\mathcal{B}_{\\mathcal{A}}$ , during training, we utilize $\\mathcal{P}_{\\mathcal{B}}$ under canonical pose, which is our ground truth point cloud of $\\mathcal{P}_{\\mathcal{B}}$ , to train our model. During testing, we first predict the pose of $\\mathcal{P}_{\\mathcal{B}}$ , then use the transformed $\\mathcal{P}_{\\mathcal{B}}$ , along with the initial $\\mathcal{P}_{\\mathcal{A}}$ to predict A's pose, as shown in Figure 4. This phased, two-step training and evaluation strategy reduces errors caused by joint training of object poses, ensuring more accurate predictions.", + "bbox": [ + 496, + 734, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17387", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.6. Loss Function", + "text_level": 1, + "bbox": [ + 76, + 90, + 222, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To train our network to robustly predict poses, we use the following equation as our loss function:", + "bbox": [ + 76, + 113, + 468, + 143 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {\\text {r o t}} \\mathcal {L} _ {\\text {r o t}} + \\lambda_ {\\text {t r a n s}} \\mathcal {L} _ {\\text {t r a n s}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 152, + 468, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, for predicted pose translation $T_{\\mathrm{pred}} \\in R^3$ , rotation $\\mathcal{R}_{\\mathrm{pred}} \\in R^{3 \\times 3}$ and ground truth pose translation $T_{\\mathrm{gt}} \\in R^3$ and rotation $\\mathcal{R}_{\\mathrm{gt}} \\in R^{3 \\times 3}$ , we use $\\mathcal{L}_1$ loss to compute our $\\mathcal{L}_{\\mathrm{trans}}$ :", + "bbox": [ + 76, + 175, + 468, + 236 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t r a n s}} = \\mathcal {L} _ {1} \\left(T _ {\\text {p r e d}}, T _ {\\text {g t}}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 244, + 468, + 261 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As for the rotation, we utilize Geodesic Distance, which measures the shortest path between two rotations on the rotation manifold. It offers a smooth and bounded angular error, ensuring stable gradients and accurately achieving precise rotation alignment.", + "bbox": [ + 76, + 268, + 468, + 344 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r o t}} = \\arccos \\left(\\frac {\\operatorname {t r} \\left(\\mathcal {R} _ {\\mathrm {g t}} \\mathcal {R} _ {\\text {p r e d}} ^ {T}\\right) - 1}{2}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 349, + 468, + 391 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 76, + 398, + 209, + 415 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we present a comprehensive evaluation and analysis of our two-step pairwise network architecture by addressing the following key questions:", + "bbox": [ + 76, + 422, + 468, + 468 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. How does our network perform on 2BY2 tasks compared to existing baseline approaches, including matching-based, graph-network-based, and diffusion-based assembly methods?", + "2. How well does our network generalize across multiple tasks within the 2BY2 dataset? Can our network effectively handle a diverse set of tasks simultaneously?", + "3. Can our network generalize to real-world robot tasks?" + ], + "bbox": [ + 76, + 468, + 468, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1.2BY2 Dataset Main Experiment", + "text_level": 1, + "bbox": [ + 76, + 597, + 359, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1.1 Experiment Set Up", + "text_level": 1, + "bbox": [ + 76, + 619, + 264, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tasks. We divide the 18 assembly tasks in the 2BY2 dataset into training and testing sets individually and compared the performance of our method with various baseline approaches. To further evaluate its cross-task generalization ability, we conducted additional experiments on tasks such as Lid Covering, Insertion, and High Precision Placement, as well as All task, which requires the method to handle all tasks in the entire dataset. See Table 2 for task details.", + "bbox": [ + 75, + 643, + 468, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. Following metrics from datasets like Breaking Bad [41] and Neural Shape Mating [5], we use Root Mean Squared Error (RMSE) to evaluate both rotation and translation of the predicted SE(3) pose. Specifically, rotations are represented using Euler angles with symmetry considerations, see Section 3.2 for symmetry details.", + "bbox": [ + 75, + 763, + 468, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training parameters. We set batch size to be 4, and the initial learning rate of Adam Optimizer [10] to be 1e-4. We train models for 1000 epochs for them to fully converge.", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1.2 Baselines", + "text_level": 1, + "bbox": [ + 500, + 90, + 619, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method with SE-3 assembly [53], Puzzlefusion++ [50], Jigsaw [32] and Neural Shape Mating [5].", + "bbox": [ + 498, + 114, + 893, + 143 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- SE-3 Assembly [53] proposes a network architecture to leverage SE(3) equivariance for representations considering multi-part correlations, and predict the pose of each part jointly.", + "- Puzzlefusion++ [50] proposes an auto-agglomerative 3D fracture assembly framework. It iteratively aligns and merges fragments using a diffusion model for 6-DoF alignment and a transformer model for verification.", + "- Jigsaw [32] leverages hierarchical features of global and local geometry to match and align the fracture surfaces, and recovers the global pose of each piece to restore the underlying object.", + "- Neural Shape Mating [5] utilizes PointNet for feature encoding and a transformer for feature fusion to learn the correlations between assembly parts, enabling joint prediction of their poses." + ], + "bbox": [ + 500, + 145, + 893, + 387 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1.3 2BY2 Benchmark Results and Analysis", + "text_level": 1, + "bbox": [ + 500, + 406, + 826, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3 presents the quantitative performance of our method compared to all baselines. The results show that our approach outperforms the baselines across 18 fine-grained assembly tasks, with an average improvement of 0.046 in translation RMSE and 8.97 in rotation RMSE.", + "bbox": [ + 498, + 429, + 892, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, we evaluate our method on three cross-category tasks defined in Section 3.1, namely Lid Covering, Inserting and High Precision Placing, and achieve the state-of-the-art performance. Moreover, in the most comprehensive All task, we outperform the baseline by 0.123 in translation and 10.90 in rotation, demonstrating strong generalization across tasks and object shapes. In the meantime, baseline comparisons confirm the rigor and challenge of our tasks. Results on challenging tasks like Plug and Key highlight our framework's effectiveness in complex scenarios.", + "bbox": [ + 496, + 505, + 893, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We analyze that the superior performance of our designed network is due to the approach of separately predicting the poses of the two objects in a step-by-step manner. This prevents the pose errors from interfering with each other, which often occurs in other baselines when predicting both poses simultaneously. Additionally, the design of our encoder makes our network more sensitive to subtle changes in rotation and translation, resulting better performance.", + "bbox": [ + 496, + 655, + 893, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Real-World Robot Experiment", + "text_level": 1, + "bbox": [ + 500, + 786, + 772, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Real-world robot experiment setup. As shown in Figure 5, we conduct our real-world robot experiments using a UR5 robotic arm, equipped with a Robotiq 2F-85 Gripper. We select four tasks, Cup, Flower, Bread and Plug, demonstrating that our model exhibits strong generalization ability on unseen real-world objects.", + "bbox": [ + 496, + 809, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17388", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fe3a5a31411bc4697c273c9d0ebb776e192ad54a845b79757e56bef8eaeaa9da.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskJigsaw [32]Puzzlefusion++ [50]NSM [5]SE(3)-Assembly [53]Ours
RMSE(T)RMSE(R)RMSE(T)RMSE(R)RMSE(T)RMSE(R)RMSE(T)RMSE(R)RMSE(T)↓RMSE(R)↓
Lid Covering0.39833.330.40837.740.18433.450.12521.370.09016.12
Kitchenport0.47745.800.42347.230.23757.470.09317.290.06816.60
Bottle0.41134.710.38535.230.22778.580.14736.920.07627.70
Kettle0.33543.710.37238.380.21561.100.13313.150.11111.56
Coffeemachine0.52732.670.43734.640.25350.660.14224.430.07622.83
Cup0.40833.550.43933.580.26067.350.16046.460.12223.18
Inserting0.36453.580.32757.830.27569.930.19946.30.14238.03
Plug0.37256.890.34848.890.30352.260.17618.580.0949.74
Childrentoy0.26859.880.24563.210.27193.770.30280.520.24257.81
Letter0.40967.990.35772.080.31776.480.12139.240.09433.74
Bread0.22057.840.20160.920.17165.500.11151.130.09036.40
Nut0.47640.080.32347.290.27155.320.10246.680.05135.60
Coin0.40639.620.34851.400.28962.580.11128.690.10722.88
Key0.38442.850.34850.380.29063.600.08717.280.04516.32
Usb0.46367.410.34258.230.25269.900.21532.280.12828.98
Precision Placing0.37573.940.28767.810.21185.020.13457.860.11544.84
Box0.13733.720.13440.470.13072.470.07125.080.06621.53
Tissue0.29282.390.26585.180.17579.780.18373.020.11564.37
Flower0.32864.390.28359.040.24687.180.21364.890.12542.23
Teaport0.30268.330.32461.010.28856.320.08540.590.05026.11
Position0.42358.070.38957.550.25770.110.16628.170.14124.46
ALL0.36053.340.34258.230.28470.300.23352.340.11041.44
", + "bbox": [ + 76, + 88, + 893, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d63afcedd816608ab0e5ef1a9cc93c4eca7d77de0aa3eab58cda76220feee797.jpg", + "image_caption": [ + "Figure 5. Real Robot Setup. We conduct real-world robot experiments on Cup, Flower, Bread and Plug tasks." + ], + "image_footnote": [], + "bbox": [ + 78, + 481, + 472, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We place objects in the scene with random initial poses and scan them to obtain their point clouds. Using pre-trained models on selected data of 2BY2 dataset, we predict the pose of each object. A manually designed grasping pose is then applied to pick up each object, and based on the predicted poses, the robotic arm plans a trajectory to complete the assembly. We use SE(3) assembly [53] as the baseline and test our approach on 10 different initial poses. As shown in Table 4, our method significantly outperforms the baseline.", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5ceb72ad0089a4bd1c9c060a305ad6c5ea65191b187c2fb19519b792044fcaa4.jpg", + "table_caption": [ + "Table 3. Quantitative Evaluation on 2BY2 for Pairwise Object Assembly. Our method outperforms the baseline across all 18 fine-grained assembly tasks, as well as demonstrating significant improvement on three cross-category assembly tasks. It achieves a lower task average with a reduction of 0.046 in translation RMSE and 8.97 in rotation RMSE." + ], + "table_footnote": [], + "table_body": "
TaskCupFlowerBreadPlugOverall
SE(3) [53]2/104/101/102/1022.5%
Ours8/1010/106/107/1077.5%
", + "bbox": [ + 531, + 473, + 857, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Real-World Robot Experiment Success Rate Results.", + "bbox": [ + 501, + 535, + 887, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 579, + 651, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct comprehensive experiments to demonstrate the rationality of our network design and the effectiveness of each module.", + "bbox": [ + 498, + 607, + 890, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Encoder. To validate the effectiveness of two-scale Vector Neuron(VN) DGCNN, we compare it with other encoders: VN DGCNN [9], DGCNN [48], and PointNet [38].", + "bbox": [ + 498, + 654, + 893, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Two-step network design. We compare our method with an end-to-end approach, which jointly predicts the pose of $\\mathcal{P}_A$ and $\\mathcal{P}_{\\mathcal{B}}$ . Specifically, we utilize $\\mathcal{B}_A$ and the input point cloud of $\\mathcal{P}_A$ and $\\mathcal{P}_{\\mathcal{B}}$ to get the 6D pose of $P_A$ , while using the same encoder to extract $\\mathcal{E}_{\\mathcal{B}}$ and pass it through the same pose prediction head to predict the pose of $P_B$ .", + "bbox": [ + 496, + 702, + 893, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in in Table 5, we show the results of ablation studies on Lid covering, Inserting and Precision Placing, which are more comprehensive and require cross-task generalization abilities. The performance declines in both translation and rotation when we removing our two-scale VN DGCNN encoder and change it to Vector Neuron DGCNN [9], DGCNN [48], Pointnet [38]. It demonstrates", + "bbox": [ + 496, + 795, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17389", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d4f59e3949644bd077a197c54f2a47ce948e6083951ed494a49e90ed2345e1bd.jpg", + "image_caption": [ + "Figure 6. Qualitative Results Comparison. We highlight Bottle, Plug, Bread, Letter, Childrentoy, Key, and Flower tasks to demonstrate our improved translation and rotation predictions compared to baseline methods." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 890, + 506 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/190233c59c3af2ea98caa8719e2950677be593f307ecf13ba8aa4e092b73634a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskVector Neuron DGCNN [9]DGCNN [48]PointNet [38]w/o Two-stepOurs
RMSE(T)RMSE(R)RMSE(T)RMSE(R)RMSE(T)RMSE(R)RMSE(T)RMSE(R)RMSE(T) ↓RMSE(R) ↓
Lid Covering0.09818.230.24570.060.23465.320.11718.740.09016.12
Inserting0.15741.220.23462.250.23466.800.16441.110.14238.03
Precision Placing0.12148.010.24565.190.21172.470.13746.380.11544.84
ALL0.12344.670.27772.460.26475.380.13945.200.11041.44
", + "bbox": [ + 78, + 560, + 890, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation Study Results. We compare various encoders including Vector Neuron DGCNN [9], DGCNN [48], PointNet [38], and our proposed two-scale Vector Neuron DGCNN. We also compare end-to-end networks with two-step networks to demonstrate the effectiveness of each component in our network design.", + "bbox": [ + 75, + 659, + 893, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "that our encoder effectively exploits the advantage of SE(3) equivariance, enabling greater sample efficiency and more robust generalization abilities. Compared with version in Figure 4, the experiment performance declines when we change our two-step network in a joint-learning manner, proving that our two-step network design can reduce error caused by jointly predictions and thereby is more effective.", + "bbox": [ + 75, + 728, + 470, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 844, + 194, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "2BY2 is a significant step in bridging the gap between geometry-based assembly tasks and everyday object assem", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "blies. With pose and symmetry annotations for 517 object pairs across 18 fine-grained tasks, 2BY2 sets a new benchmark for 3D assembly challenges. Our two-step pairwise SE(3) pose estimation framework, which leverages equivariant features, demonstrates superior performance over existing approaches in reducing both translation and rotation errors. Robot experiments further validate the method's generalizability in practical 3D assembly scenarios. In conclusion, 2BY2 provides both a comprehensive benchmark and an effective framework, with the aim of inspiring and supporting more generalizable solution in robot manipulation.", + "bbox": [ + 496, + 728, + 893, + 895 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17390", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] 3dwarehouse. https://3dwarehouse.sketchup.com/by/su3dchallenge, 2014.3", + "[2] google-sketchup-3d-challenge. https://3dwarehouse.sketchup.com/by/su3dchallenge, 2014.3", + "[3] Benedict J Brown, Corey Toler-Franklin, Diego Nehab, Michael Burns, David Dobkin, Andreas Vlachopoulos, Christos Doumas, Szymon Rusinkiewicz, and Tim Weyrich. A system for high-volume acquisition and matching of fresco fragments: Reassembling theran wall paintings. ACM transactions on graphics (TOG), 27(3):1-9, 2008. 2", + "[4] Siddhartha Chaudhuri, Evangelos Kalogerakis, Leonidas Guibas, and Vladlen Koltun. Probabilistic reasoning for assembly-based 3d modeling. In ACM SIGGRAPH 2011 papers, pages 1-10. 2011. 2", + "[5] Yun-Chun Chen, Haoda Li, Dylan Turpin, Alec Jacobson, and Animesh Garg. Neural shape mating: Self-supervised object assembly with adversarial shape priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12724-12733, 2022. 2, 3, 6, 7", + "[6] Junfeng Cheng, Mingdong Wu, Ruiyuan Zhang, Guanqi Zhan, Chao Wu, and Hao Dong. Score-pa: Score-based 3d part assembly. In British Machine Vision Conference, 2023. 2", + "[7] Taco S Cohen, Mario Geiger, Jonas Kohler, and Max Welling. Spherical cnns. arXiv preprint arXiv:1801.10130, 2018. 2", + "[8] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\\mathrm{m} + 3\\mathrm{d}$ objects. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[9] Congyue Deng, Or Litany, Yueqi Duan, Adrien Poulenard, Andrea Tagliasacchi, and Leonidas J Guibas. Vector neurons: A general framework for so (3)-equivariant networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12200-12209, 2021. 2, 4, 7, 8", + "[10] P Kingma Diederik. Adam: A method for stochastic optimization. (No Title), 2014. 6", + "[11] Thomas Funkhouser, Michael Kazhdan, Philip Shilane, Patrick Min, William Kiefer, Ayellet Tal, Szymon Rusinkiewicz, and David Dobkin. Modeling by example. ACM transactions on graphics (TOG), 23(3):652-663, 2004. 2", + "[12] Thomas Funkhouser, Hijung Shin, Corey Toler-Franklin, Antonio García Castañeda, Benedict Brown, David Dobkin, Szymon Rusinkiewicz, and Tim Weyrich. Learning how to match fresco fragments. Journal on Computing and Cultural Heritage (JOCCH), 4(2):1-13, 2011. 2", + "[13] Chongkai Gao, Zhengrong Xue, Shuying Deng, Tianhai Liang, Siqi Yang, Lin Shao, and Huazhe Xu. Riemann: Near real-time se (3)-equivariant robot manipulation without point cloud segmentation. arXiv preprint arXiv:2403.19460, 2024. 3", + "[14] Stefan Hinterstoisser, Stefan Holzer, Cedric Cagniart, Slobodan Ilic, Kurt Konolige, Nassir Navab, and Vincent Lepetit. Multimodal templates for real-time detection of texture-less objects in heavily cluttered scenes. In 2011 International Conference on Computer Vision, pages 858-865, 2011. 2" + ], + "bbox": [ + 78, + 114, + 472, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Stefan Hinterstoisser, Vincent Lepetit, Slobodan Ilic, Stefan Holzer, Gary Bradski, Kurt Konolige, and Nassir Navab. Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In Computer Vision - ACCV 2012, pages 548-562, Berlin, Heidelberg, 2013. Springer Berlin Heidelberg. 2", + "[16] Sepidehsadat Sepid Hossieni, Mohammad Amin Shabani, Saghar Irandoust, and Yasutaka Furukawa. Puzzle: unleashing the power of diffusion models for spatial puzzle solving. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[17] Boce Hu, Xupeng Zhu, Dian Wang, Zihao Dong, Haojie Huang, Chenghao Wang, Robin Walters, and Robert Platt. Orbitgrasp: Se (3)-equivariant grasp learning. In 8th Annual Conference on Robot Learning, 2024. 3", + "[18] Haojie Huang, Dian Wang, Xupeng Zhu, Robin Walters, and Robert Platt. Edge grasp network: A graph-based se (3)-invariant approach to grasp detection. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3882-3888. IEEE, 2023.", + "[19] Haojie Huang, Haotian Liu, Dian Wang, Robin Walters, and Robert Platt. Match policy: A simple pipeline from point cloud registration to manipulation policies. arXiv preprint arXiv:2409.15517, 2024.", + "[20] Haojie Huang, Karl Schmeckpeper, Dian Wang, Ondrej Biza, Yaoyao Qian, Haotian Liu, Mingxi Jia, Robert Platt, and Robin Walters. IMAGINATION POLICY: Using generative point cloud models for learning manipulation policies. In 8th Annual Conference on Robot Learning, 2024. 2, 3", + "[21] Qi-Xing Huang, Simon Flöry, Natasha Gelfand, Michael Hofer, and Helmut Pottmann. Reassembling fractured objects by geometric matching. In ACM Siggraph 2006 papers, pages 569-578. 2006. 2", + "[22] Prakhar Jaiswal, Jinmiao Huang, and Rahul Rai. Assembly-based conceptual 3d modeling with unlabeled components using probabilistic factor graph. Computer-Aided Design, 74: 45-54, 2016. 2", + "[23] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J Davison. Rlbench: The robot learning benchmark & learning environment. IEEE Robotics and Automation Letters, 5(2):3019-3026, 2020. 2", + "[24] Benjamin Jones, Dalton Hildreth, Duowen Chen, Ilya Baran, Vladimir G Kim, and Adriana Schulz. Automate: A dataset and learning approach for automatic mating of cad assemblies. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021. 2, 3", + "[25] Yuanchen Ju, Kaizhe Hu, Guowei Zhang, Gu Zhang, Mingrun Jiang, and Huazhe Xu. Robo-abc: Affordance generalization beyond categories via semantic correspondence for robot manipulation. In European Conference on Computer Vision, pages 222–239. Springer, 2025. 2", + "[26] Evangelos Kalogerakis, Siddhartha Chaudhuri, Daphne Koller, and Vladlen Koltun. A probabilistic model for component-based shape synthesis. Acm Transactions on Graphics (TOG), 31(4):1-11, 2012. 2", + "[27] Wadim Kehl, Fabian Manhardt, Federico Tombari, Slobodan Ilic, and Nassir Navab. SSD-6D: making rgb-based" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17391", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "3d detection and 6d pose estimation great again. CoRR, abs/1711.10006, 2017. 2", + "[28] Yichen Li, Kaichun Mo, Lin Shao, Minhyuk Sung, and Leonidas Guibas. Learning 3d part assembly from a single image. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 664-682. Springer, 2020. 2", + "[29] Yulong Li, Andy Zeng, and Shuran Song. Rearrangement planning for general part assembly. In 7th Annual Conference on Robot Learning, 2023. 2", + "[30] Yuval Litvak, Armin Biess, and Aharon Bar-Hillel. Learning pose estimation for high-precision robotic assembly using simulated depth images. In 2019 International Conference on Robotics and Automation (ICRA), pages 3521–3527. IEEE, 2019. 2", + "[31] Hao-Yu Liu, Jian-Wei Guo, Hai-Yong Jiang, Yan-Chao Liu, Xiao-Peng Zhang, and Dong-Ming Yan. Puzzlenet: boundary-aware feature matching for non-overlapping 3d point clouds assembly. Journal of Computer Science and Technology, 38 (3):492-509, 2023. 2", + "[32] Jiaxin Lu, Yifan Sun, and Qixing Huang. Jigsaw: Learning to assemble multiple fractured objects. Advances in Neural Information Processing Systems, 36, 2024. 2, 6, 7", + "[33] Jianlan Luo, Eugen Solowjow, Chengtao Wen, Juan Aparicio Ojea, Alice M Agogino, Aviv Tamar, and Pieter Abbeel. Reinforcement learning on variable impedance controller for high-precision robotic assembly. In 2019 International Conference on Robotics and Automation (ICRA), pages 3080–3087. IEEE, 2019. 2", + "[34] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909-918, 2019. 3", + "[35] Yashraj Narang, Kier Storey, Iretiayo Akinola, Miles Macklin, Philipp Reist, Lukasz Wawrzyniak, Yunrong Guo, Adam Moravanszky, Gavriel State, Michelle Lu, et al. Factory: Fast contact for robotic assembly. arXiv preprint arXiv:2205.03532, 2022. 2, 3", + "[36] Abhinav Narayan, Rajendra Nagar, and Shanmuganathan Ram. Rgl-net: A recurrent graph learning framework for progressive part assembly. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 78-87, 2022. 2", + "[37] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 2", + "[38] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 7, 8", + "[39] Ismael Rodriguez, Korbinian Nottensteiner, Daniel Leidner, Michael KaBecker, Freek Stulp, and Alin Albu-Schaffer. Iteratively refined feasibility checks in robotic assembly sequence" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "planning. IEEE Robotics and Automation Letters, 4(2):1416-1423, 2019. 3", + "[40] Gianluca Scarpellini, Stefano Fiorini, Francesco Giuliani, Pietro Moreiro, and Alessio Del Bue. Diffassemble: A unified graph-diffusion model for 2d and 3d reassembly. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28098-28108, 2024. 2", + "[41] Silvia Sellán, Yun-Chun Chen, Ziyi Wu, Animesh Garg, and Alec Jacobson. Breaking bad: A dataset for geometric fracture and reassembly. Advances in Neural Information Processing Systems, 35:38885-38898, 2022. 2, 3, 6", + "[42] Hijung Shin, Christos Doumas, Thomas Funkhouser, Szymon Rusinkiewicz, Kenneth Steiglitz, Andreas Vlachopoulos, and Tim Weyrich. Analyzing and simulating fracture patterns of theran wall paintings. Journal on Computing and Cultural Heritage (JOCCH), 5(3):1-14, 2012. 2", + "[43] Chenrui Tie, Yue Chen, Ruihai Wu, Boxuan Dong, Zeyi Li, Chongkai Gao, and Hao Dong. ET-SEED: EFFICIENT TRAJECTORY-LEVEL SE(3) EQUIVARIANT DIFFUSION POLICY. In The Thirteenth International Conference on Learning Representations, 2025. 3", + "[44] Joshua Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. CoRR, abs/1703.06907, 2017. 2", + "[45] Jonathan Tremblay, Thang To, Balakumar Sundaralingam, Yu Xiang, Dieter Fox, and Stan Birchfield. Deep object pose estimation for semantic robotic grasping of household objects. ArXiv, abs/1809.10790, 2018. 2", + "[46] Weiwei Wan, Kensuke Harada, and Kazuyuki Nagata. Assembly sequence planning for motion planning. CoRR, abs/1609.03108, 2016. 3", + "[47] Ruocheng Wang, Yunzhi Zhang, Jiayuan Mao, Ran Zhang, Chin-Yi Cheng, and Jiajun Wu. Iea-manual: Seeing shape assembly step by step. Advances in Neural Information Processing Systems, 35:28428-28440, 2022. 2", + "[48] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 7, 8", + "[49] Ziming Wang and Rebecka Jornsten. Se (3)-bi-equivariant transformers for point cloud assembly. arXiv preprint arXiv:2407.09167, 2024. 2", + "[50] Zhengqing Wang, Jiacheng Chen, and Yasutaka Furukawa. Puzzlefusion++: Auto-agglomerative 3d fracture assembly by denoise and verify. arXiv preprint arXiv:2406.00259, 2024. 2, 6, 7", + "[51] Karl DD Willis, Pradeep Kumar Jayaraman, Hang Chu, Yunsheng Tian, Yifei Li, Daniele Grandi, Aditya Sanghi, Linh Tran, Joseph G Lambourne, Armando Solar-Lezama, et al. Joinable: Learning bottom-up assembly of parametric cad joints. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15849-15860, 2022. 2, 3", + "[52] Rundi Wu, Yixin Zhuang, Kai Xu, Hao Zhang, and Baoquan Chen. Pq-net: A generative part seq2seq network for 3d shapes. In Proceedings of the IEEE/CVF Conference on" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17392", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Computer Vision and Pattern Recognition, pages 829-838, 2020. 2", + "[53] Ruihai Wu, Chenrui Tie, Yushi Du, Yan Zhao, and Hao Dong. Leveraging se (3) equivariance for learning 3d geometric shape assembly. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14311-14320, 2023. 6, 7", + "[54] Fanbo Xiang, Yuzhe Qin, Kaichun Mo, Yikuan Xia, Hao Zhu, Fangchen Liu, Minghua Liu, Hanxiao Jiang, Yifu Yuan, He Wang, Li Yi, Angel X. Chang, Leonidas J. Guibas, and Hao Su. SAPIEN: A simulated part-based interactive environment. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3", + "[55] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. 2018. 2", + "[56] Yuwen Xiong, Wei-Chiu Ma, Jingkang Wang, and Raquel Urtasun. Learning compact representations for lidar completion and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1074-1083, 2023. 2", + "[57] Boshen Xu, Sipeng Zheng, and Qin Jin. Spaformer: Sequential 3d part assembly with transformers. arXiv preprint arXiv:2403.05874, 2024. 2", + "[58] Xianghao Xu, Paul Guerrero, Matthew Fisher, Siddhartha Chaudhuri, and Daniel Ritchie. Unsupervised 3d shape reconstruction by part retrieval and assembly. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8559-8567, 2023. 2", + "[59] Zhengrong Xue, Zhecheng Yuan, Jiashun Wang, Xueqian Wang, Yang Gao, and Huazhe Xu. Usek: Unsupervised se (3)-equivariant 3d keypoints for generalizable manipulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1715-1722. IEEE, 2023. 3", + "[60] Kangxue Yin, Zhiqin Chen, Siddhartha Chaudhuri, Matthew Fisher, Vladimir G Kim, and Hao Zhang. Coalesce: Component assembly by learning to synthesize connections. In 2020 International Conference on 3D Vision (3DV), pages 61-70. IEEE, 2020. 2", + "[61] Kevin Zakka, Andy Zeng, Johnny Lee, and Shuran Song. Form2fit: Learning shape priors for generalizable assembly from disassembly. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9404–9410. IEEE, 2020. 2", + "[62] Guanqi Zhan, Qingnan Fan, Kaichun Mo, Lin Shao, Baoquan Chen, Leonidas J Guibas, Hao Dong, et al. Generative 3d part assembly via dynamic graph learning. Advances in Neural Information Processing Systems, 33:6315-6326, 2020. 2", + "[63] Rufeng Zhang, Tao Kong, Weihao Wang, Xuan Han, and Mingyu You. 3d part assembly generation with instance encoded transformer. IEEE Robotics and Automation Letters, 7(4):9051-9058, 2022. 2", + "[64] Ruiyuan Zhang, Jiaxiang Liu, Zexi Li, Hao Dong, Jie Fu, and Chao Wu. Scalable geometric fracture assembly via co-creation space among assemblers. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7269–7277, 2024. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Haibo Zhao, Dian Wang, Yizhe Zhu, Xupeng Zhu, Owen Howell, Linfeng Zhao, Yaoyao Qian, Robin Walters, and Robert Platt. Hierarchical equivariant policy via frame transf. arXiv preprint arXiv:2502.05728, 2025. 3", + "[66] Junzhe Zhu, Yuanchen Ju, Junyi Zhang, Muhan Wang, Zhecheng Yuan, Kaizhe Hu, and Huazhe Xu. Densehunter: Learning 3d semantic correspondence for category-level manipulation from a single demo. International Conference on Learning Representations (ICLR) 2025, 2024. 2", + "[67] Xupeng Zhu, Dian Wang, Guanang Su, Ondrej Biza, Robin Walters, and Robert Platt. On robot grasp learning using equivariant models. Autonomous Robots, 47(8):1175-1193, 2023. 3", + "[68] Yuke Zhu, Josiah Wong, Ajay Mandlekar, Roberto Martin-Martin, Abhishek Joshi, Soroush Nasiriany, and Yifeng Zhu. robosuite: A modular simulation framework and benchmark for robot learning. arXiv preprint arXiv:2009.12293, 2020. 2" + ], + "bbox": [ + 501, + 92, + 893, + 330 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17393", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Two by Two_ Learning Multi-Task Pairwise Objects Assembly for Generalizable Robot Manipulation/57fa886e-a784-460f-94a7-de5f04b8ab96_model.json b/2025/Two by Two_ Learning Multi-Task Pairwise Objects Assembly for Generalizable Robot Manipulation/57fa886e-a784-460f-94a7-de5f04b8ab96_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1292f330e76bb892387ff69555443fc56f8c4f32 --- /dev/null +++ b/2025/Two by Two_ Learning Multi-Task Pairwise Objects Assembly for Generalizable Robot Manipulation/57fa886e-a784-460f-94a7-de5f04b8ab96_model.json @@ -0,0 +1,2466 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.158, + 0.123, + 0.815, + 0.179 + ], + "angle": 0, + "content": "Two by Two: Learning Multi-Task Pairwise Objects Assembly for Generalizable Robot Manipulation" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.206, + 0.913, + 0.261 + ], + "angle": 0, + "content": "Yu Qi \\(^{2,1*}\\) Yuanchen Ju\\(^{1,3*}\\) Tianming Wei\\(^{1,4}\\) Chi Chu\\(^{1}\\) Lawson L.S. Wong\\(^{2}\\) Huazhe Xu\\(^{1,3,5\\dagger}\\) \n\\(^{1}\\) Shanghai Qi Zhi Institute \\(^{2}\\) Northeastern University \\(^{3}\\) IIIS, Tsinghua University \n\\(^{4}\\) Shanghai Jiao Tong University \\(^{5}\\) Shanghai AI Laboratory" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.271, + 0.891, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.551, + 0.893, + 0.581 + ], + "angle": 0, + "content": "Figure 1. Overview of the 2BY2 Dataset. We propose the first large-scale daily pairwise object assembly dataset 2BY2, which contains 1,034 instances and 517 pairwise objects with pose and symmetry annotations." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.6, + 0.314, + 0.616 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.474, + 0.904 + ], + "angle": 0, + "content": "3D assembly tasks, such as furniture assembly and component fitting, play a crucial role in daily life and represent essential capabilities for future home robots. Existing benchmarks and datasets predominantly focus on assembling geometric fragments or factory parts, which fall short in addressing the complexities of everyday object interactions and assemblies. To bridge this gap, we present 2BY2, a large-scale annotated dataset for daily pairwise objects assembly, covering 18 fine-grained tasks that reflect real-life scenarios, such as plugging into sockets, arranging flowers in vases, and inserting bread into toasters. 2BY2 dataset includes 1,034 instances and 517 pairwise objects with pose and symmetry annotations, requiring approaches that align geometric shapes while accounting for functional and spatial relationships between objects. Leveraging the 2BY2 dataset, we propose a two-step SE(3) pose estimation method with equivariant features for assembly constraints. Compared to previous shape assembly methods, our approach achieves state-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.601, + 0.896, + 0.678 + ], + "angle": 0, + "content": "performance across all 18 tasks in the 2BY2 dataset. Additionally, robot experiments further validate the reliability and generalization ability of our method for complex 3D assembly tasks. More details and demonstrations can be found at https://tea-lab.github.io/TwoByTwo/." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.696, + 0.633, + 0.712 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.721, + 0.895, + 0.872 + ], + "angle": 0, + "content": "Assembly tasks are ubiquitous, such as assembling furniture, repairing household appliances, or putting together electronics. Successfully completing these tasks requires precise reasoning about the spatial relationships between pairs of objects. For robots to assist in these activities, they need to accurately estimate the 6D pose of each objects—including both their orientation and position in space. This capability is essential for domestic robots to help humans with various tasks, as it enables them to interact with their environment in a meaningful way." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.873, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Daily object pairwise assembly not only requires considering the geometric constraints and spatial relationships" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17383" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "between objects to achieve precise alignment but also needs to exhibit a certain level of generalization. Existing methods and benchmarks for solving assembly problems [31, 32, 49, 50, 63], typically focus on matching local geometric shapes, which often results in suboptimal performance in everyday assembly scenarios which require semantic and spatial alignment. This is because they are primarily trained and tested on existing assembly datasets that consist of large-scale geometric fragments, such as Breaking Bad [41] and Neural Shape Mating [5]. Compared with existing assembly tasks that focus on putting together fractures of objects, daily pairwise assembly tasks are more challenging and hold greater practical significance in human life." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.289, + 0.473, + 0.426 + ], + "angle": 0, + "content": "To bridge this gap, we introduce 2BY2, the first large-scale daily pairwise assembly dataset comprising 18 fine-grained tasks, shown as Figure 1. Compared to previous datasets and benchmarks, 2BY2 contains 1,034 instances and 517 pairwise objects with pose and symmetry annotations, covering a variety of pairwise assembly tasks reflecting everyday scenarios, which require approaches that align geometric shapes while accounting for functional and spatial relationships between objects, as shown in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.427, + 0.473, + 0.639 + ], + "angle": 0, + "content": "Building on this dataset, we propose a two-step pairwise network architecture for assembly tasks. Mimicking the human assembly process like we firstly put the vase on the table and then arrange flower in it, our approach predicts the pose of each object in a step-by-step manner to assemble them to a predefined canonical space, which refers to a standard coordinate system that aligns with the principles of the human world, with detailed definition in Section 3.2. The network leverages a custom two-scale Vector Neuron DGCNN [9] encoder with spherical convolution [7] to extract SE(3) equivariant and SO(3) invariant features from point cloud inputs. Additionally, a feature fusion module and a two-step training and evaluation strategy are used to improve pose prediction accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.473, + 0.793 + ], + "angle": 0, + "content": "We evaluate our approach on 18 tasks in 2BY2 dataset to demonstrate the effectiveness on multi-task object pairwise assembly prediction. Compared to existing baselines, our method achieves an average improvement of 0.046 in translation RMSE and 8.97 in rotation RMSE. Moreover, we validate the effectiveness of our approach on three multicategory task, namely Lid Covering, Inserting and High Precision Placing, as well as the All task, which is defined in Section 2. Besides, real-world robot experiments validate the practical applicability of our approach." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.794, + 0.393, + 0.808 + ], + "angle": 0, + "content": "Our main contributions are listed as follows:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.472, + 0.902 + ], + "angle": 0, + "content": "1. We introduce 2BY2, the first large-scale daily pairwise object assembly dataset. By providing comprehensive pose and symmetry annotations for 517 pairwise objects across 18 fine-grained tasks, 2BY2 pushes the boundaries of real-world 3D assembly challenges and establishes a benchmark for pairwise assembly tasks." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.092, + 0.895, + 0.168 + ], + "angle": 0, + "content": "2. Our two-step pairwise SE(3) pose estimation method, leveraging equivariant geometric features, demonstrates superior performance compared to existing shape assembly methods, significantly reducing translation and rotation errors and enhancing the accuracy of 6D pose estimation." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.168, + 0.895, + 0.229 + ], + "angle": 0, + "content": "3. Our approach achieves state-of-the-art performance on the benchmark, with real-world robot experiments demonstrating its capability, providing a generalizable solution for robot manipulation using pairwise object assembly." + }, + { + "type": "list", + "bbox": [ + 0.497, + 0.092, + 0.895, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.242, + 0.642, + 0.258 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.267, + 0.871, + 0.284 + ], + "angle": 0, + "content": "2.1. Object Assembly Benchmarks and Datasets" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.29, + 0.895, + 0.502 + ], + "angle": 0, + "content": "Object reassembly has led to various datasets in computer vision and robotics. In computer vision, datasets like AutoMate [24] and JoinABLe [51] focus on reassembling fragments using geometric clues, while early datasets [3, 12, 21, 42] were limited in scale. Recent efforts, such as Neural Shape Mating [5] and Breaking Bad [41], generate largescale fractured object data using parametric segmentation. In robotics, benchmarks like Factory [35], RLBench [23], and RoboSuite [68] lack diverse shapes and assembly tasks under varying initial poses. In contrast, our dataset includes over 500 diverse object pairs across 3 categories and 18 assembly tasks, providing a comprehensive benchmark for pairwise object assembly, supporting the development of generalizable methods for real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.511, + 0.688, + 0.528 + ], + "angle": 0, + "content": "2.2.3D Shape Assembly" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.535, + 0.895, + 0.747 + ], + "angle": 0, + "content": "3D shape assembly [11, 30, 33, 56, 61], also known as part assembly, involves reconstructing objects from fragments, such as shattered sculptures or disassembled furniture. Existing methods use graphical models [4, 22, 26] and neural networks [6, 25, 28, 47, 52, 57, 58, 60, 64, 66] to capture geometric and semantic relationships. Approaches such as [5, 36, 62, 63] focus on pose estimation and part assembly without relying on predefined semantic information. Few-shot learning has been applied to assembly tasks [29], while jigsaw puzzle techniques [32, 37] leverage shape completion strategies. Recent works [16, 20, 40, 50] utilize diffusion models to refine poses or point clouds for assembly. In contrast, our method introduces a two-step pairwise network for step-by-step assembly, tailored to pairwise object alignment." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.757, + 0.874, + 0.773 + ], + "angle": 0, + "content": "2.3.6D Pose Estimation for Robot Manipulation" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "6D pose estimation is crucial in robotics and computer vision for object interaction in unstructured environments [15, 45, 55]. Early handcrafted feature-based methods struggled in cluttered scenes [14, 15], while CNN-based approaches improved performance but lacked generalization [27, 55]. Domain randomization enhances robustness by varying synthetic datasets [44, 45]. In assembly tasks, 6D pose estimation aids manipulation planning with predefined" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17384" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.894, + 0.205 + ], + "angle": 0, + "content": "
Dataset#OC#OSTask NumberPairTask HierarchyEveryday ScenarioSymmetryAssemble Type
PartNet [34]2426,671-NoNoNoNoSemantic
AutoMate [24]292,5291NoNoNoNoGeometric
JoinABLe [51]68,2511NoNoNoNoSemantic
NSM dataset[5]111,2461YesNoNoNoGeometric
Breaking Bad[41]-10,4741NoNoNoNoGeometric
Factory[35]8608YesNoNoNoGeometric
2BY2 Dataset (Ours)36103418YesYesYesYesGeometric and Semantic
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.215, + 0.895, + 0.286 + ], + "angle": 0, + "content": "Table 1. Dataset Comparison. We compare 2BY2 dataset with existing datasets and benchmarks. #OC stands for the number of object categories. #OS stands for the number of object shapes. Pair denotes whether the dataset is pairwise. Task Number refers to the number of distinct assembly tasks, with the assembly of fractured pieces considered as a single task. Task Hierarchy stands for the different categories of task from coarse to fine, with ours shown in Section 3.1. Everyday Scenario means whether the assemble task has practical significance in real-world human applications. Symmetry denotes whether the dataset contains part symmetry annotation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.311, + 0.471, + 0.358 + ], + "angle": 0, + "content": "objects [39, 46]. Like [13, 17-20, 43, 59, 65, 67], our method leverages equivariant features for efficient 6D pose learning and improved generalization." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.369, + 0.218, + 0.385 + ], + "angle": 0, + "content": "3.2BY2 Dataset" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.393, + 0.296, + 0.409 + ], + "angle": 0, + "content": "3.1.2BY2 Dataset Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.471, + 0.567 + ], + "angle": 0, + "content": "We present the first large-scale 3D pairwise object assembly dataset for everyday scenarios, with detailed annotations for each object pair. The meshes in our dataset come from 3D Warehouse [1], SAPIEN PartNet-Mobility [54], Google SketchUp 3D Challenge [2], and Objaverse [8]. These meshes are manually paired, cleaned, annotated, and scaled uniformly. The 2BY2 dataset contains 517 unique pairs across three main tasks: Lid Covering, Inserting, and High Precision Placing, further subdivided into multiple subcategories, as shown in Table 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.576, + 0.243, + 0.59 + ], + "angle": 0, + "content": "3.2. Data Annotation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.471, + 0.825 + ], + "angle": 0, + "content": "To ensure high quality and reliability of our dataset, we conducted systematic cleaning and annotation of the collected meshes. First, we manually segment, integrate, and pair the meshes, classify them into Object \\(B\\) and Object \\(A\\). Object \\(B\\) is the base or the receiving component, such as the nut, the vase, the postbox. Object \\(A\\), is the fitting component, such as the bolt, the flower, the mail. This classification aligns with intuitive human assembly logic and supports our network's prediction strategy, such as positioning a nut before the bolt, as detailed in Sections 4 and 5.1. Automated scripts were used to uniformly scale meshes and align each pair to a canonical pose in world frame, defined as the object resting stably on the \\(XY\\) plane with its lowest point aligned to \\(Z = 0\\). For instance, bottles and vases are aligned as if placed on a table, and mailboxes on the ground." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.473, + 0.902 + ], + "angle": 0, + "content": "During point cloud generation, we utilized blue noise sampling method [41] to extract point clouds uniformly from each mesh surface with dimension (1024, 3). We also annotated each object category with its inherent symmetry properties, specifically considering rotational symmetry along the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.311, + 0.892, + 0.341 + ], + "angle": 0, + "content": "\\(Z\\)-axis, such as bottles, screws, and mirror symmetry along the \\(X\\)-axis, such as bread, letters." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.349, + 0.859, + 0.366 + ], + "angle": 0, + "content": "3.3. Data Division and Task Diversity Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.373, + 0.895, + 0.525 + ], + "angle": 0, + "content": "Our dataset provides diverse task coverage across categories, with each further divided into specific sub-categories, see Table 2. Objects within each category vary in shape, size, and type. To enhance generalization, the testing set includes objects with unseen geometric shapes, as shown in Figure 3. We also compute Chamfer Distance on point clouds between training and testing sets to quantify geometry differences, as shown in Figure 2. This diversity ensures generalization ability and applicability in real world scenarios and supports robust 3D matching and assembly tasks." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.536, + 0.891, + 0.695 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.705, + 0.894, + 0.762 + ], + "angle": 0, + "content": "Figure 2. Chamfer Distance Between Training and Testing Set. We normalize point clouds and compute the Chamfer Distance. For each task we calculate the distance separately between point cloud of Object \\( A \\) and Object \\( B \\) in the training set and test set." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.77, + 0.707, + 0.785 + ], + "angle": 0, + "content": "4. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The task takes two point clouds as input, namely \\(\\mathcal{P}_A\\) and \\(\\mathcal{P}_B\\), each with dimension (1024, 3). These point clouds are derived from objects \\(\\mathcal{O}_A\\) and \\(\\mathcal{O}_B\\) from predefined canonical pose, as detailed in 3.2, respectively, and is randomly augmented with SO(3) rotation and being translated to its centroid. The desired output would be two individual SE(3) pose two assemble \\(\\mathcal{O}_A\\) and \\(\\mathcal{O}_B\\) to the canonical pose." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17385" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.891, + 0.155 + ], + "angle": 0, + "content": "
TaskLid CoveringInsertingHigh Precision Placing
KitBotKetCofCupPluChiLetBreNutCoiKeyUsbBoxTisFloTeaPos
Pair Num248628261614193224202120202520604221
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.165, + 0.895, + 0.25 + ], + "angle": 0, + "content": "Table 2. 2BY2 Dataset Statistics Overview. The figure presents the number of object pairs across all task categories in the 2BY2 dataset, where each pair consists of two unique objects. The first row categorizes tasks into three major groups, while the second row provides a detailed breakdown of specific task categories. Specifically, Kit = Kitchenport, Bot = Bottle, Ket = Kettle, Cof = Coffee machine, Cup = Cup, Plu = Plug into socket, Chi = Children's toy, Let = Letter into mailbox, Bre = Bread into toaster, Nut = Bolt into nut, Coi = Coin into piggy bank, Key = Key into lock, Usb = USB cap, Box = Shoe boxing, Tis = Tissue placement on rack, Flo = Flower into vase, Tea = Teaware arrangement on tray, Pos = Positioning a cup on the coffee machine for coffee dispensing." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.273, + 0.273, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.273, + 0.468, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.475, + 0.472, + 0.56 + ], + "angle": 0, + "content": "Figure 3. Task Diversity Visualization. The image shows selected objects from four different tasks: USB, Bottle, Letter, and Plug in Socket. On the left are the objects selected on training set, and on the right is the testing set. As seen in the legend, object geometry varies in both the training and testing set, with the testing set containing novel shapes not seen in the training set." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.589, + 0.17, + 0.604 + ], + "angle": 0, + "content": "5. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.616, + 0.427, + 0.632 + ], + "angle": 0, + "content": "5.1. Two-step Pairwise Network Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.473, + 0.868 + ], + "angle": 0, + "content": "To effectively learn pairwise object assembly, we propose a two-step pairwise network architecture with two branches: Branch B \\((\\mathcal{B}_{\\mathcal{B}})\\) and Branch A \\((\\mathcal{B}_{\\mathcal{A}})\\), as shown in Figure 4. Branch B predicts the pose of \\(\\mathcal{P}_{\\mathcal{B}}\\), which is the socket, using a two-scale Vector Neuron DGCNN encoder [9] to extract SE(3) equivariant features, denoted as \\(\\mathcal{E}_{\\mathcal{B}}\\), followed by MLP-based pose prediction heads for translation and rotation. The transformed \\(\\mathcal{P}_{\\mathcal{B}}\\) and inserter object \\(\\mathcal{P}_{\\mathcal{A}}\\), which is the plug, are then passed to Branch A, which extracts SE(3) equivariant features, denoted as \\(\\mathcal{E}_{\\mathcal{A}}\\) and SO(3) invariant features \\((\\mathcal{I}_{\\mathcal{B}})\\). The features are fused through element-wise multiplication, allowing \\(\\mathcal{B}_{\\mathcal{A}}\\) to predict the pose of \\(\\mathcal{P}_{\\mathcal{A}}\\) using information from both objects. This architecture ensures geometric alignment and matching by leveraging shared feature representations while reducing feature interference." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Our two-step pairwise network is inspired by the human approach to pairwise assembly tasks. For example, when ar-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.275, + 0.895, + 0.366 + ], + "angle": 0, + "content": "ranging a vase with flowers, one intuitively first positions the vase correctly before placing the flowers inside. Similarly, inserting an envelope into a mailbox requires identifying the mailbox slot's pose first. By mimicking this sequential strategy, our model simulates human decision-making process, enabling more efficient and accurate assembly tasks." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.377, + 0.892, + 0.408 + ], + "angle": 0, + "content": "5.2. Two-scale SE(3) Equivariant and SO(3) Invariant Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.895, + 0.521 + ], + "angle": 0, + "content": "We employ a two-scale SE(3) Vector Neuron DGCNN, an enhanced variant of the original Vector Neuron DGCNN [9], as our encoder to extract SE(3) equivariant and SO(3) invariant features. This architecture leverages equivariance to improve sample efficiency of the model, while incorporating a two-scale information fusion mechanism to capture geometric features at two different scales." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.895, + 0.674 + ], + "angle": 0, + "content": "SE(3) Equivariance and SO(3) Invariance. SE(3) equivariance combines SO(3) rotation and T(3) translation equivariance: rotation equivariance ensures that a network's output rotates with the input, while translation equivariance shifts the output accordingly. SO(3) invariance means the network's output remains unchanged under any 3D rotation. By leveraging SE(3) equivariance, the model benefits from improved sample efficiency and generalization. This is particularly advantageous for assembly tasks, where objects may appear in arbitrary poses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.895, + 0.794 + ], + "angle": 0, + "content": "Vector Neuron DGCNN. The Vector Neuron Network [9] extends traditional neurons from scalars to 3D vectors, designs vector-based convolutional layers and nonlinear functions like pooling and ReLU to support SO(3) equivariant and SO(3) invariant feature extraction. VNN operates in vector space, captures richer geometric relationships and ensures more robust feature representations for downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.897, + 0.902 + ], + "angle": 0, + "content": "Two-scale Vector Neuron DGCNN. We propose a two-scale Vector Neuron DGCNN for extracting SE(3) equivariant and SO(3) invariant features \\(\\mathcal{E}_{\\mathcal{B}}\\), \\(\\mathcal{I}_{\\mathcal{B}}\\), and \\(\\mathcal{E}_{\\mathcal{A}}\\). As shown in Figure 4, the encoder comprises two branches with different \\(K\\) values, each consisting of multiple Vector Neuron convolutional layers followed by pooling. The extracted features from both branches are concatenated and further pro" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17386" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.516, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.725, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.092, + 0.882, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.171, + 0.881, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.249, + 0.882, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.327, + 0.882, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.264, + 0.516, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.246, + 0.416, + 0.343, + 0.431 + ], + "angle": 0, + "content": "Training Stage" + }, + { + "type": "image_caption", + "bbox": [ + 0.584, + 0.416, + 0.687, + 0.43 + ], + "angle": 0, + "content": "Inference Stage" + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.416, + 0.868, + 0.429 + ], + "angle": 0, + "content": "Robot Experiments" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.442, + 0.894, + 0.472 + ], + "angle": 0, + "content": "Figure 4. Our Two-Step Pairwise Network. We utilize two-scale VN DGCNN as our encoder to extract equivariant and invariant feature. We first predict the canonical pose of \\(\\mathcal{O}_B\\) and then predict the pose of \\(\\mathcal{O}_A\\) according to it." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.497, + 0.471, + 0.543 + ], + "angle": 0, + "content": "cessed through an additional Vector Neuron convolutional layer. Point clouds \\(\\mathcal{P}_{\\mathcal{B}}\\) and \\(\\mathcal{P}_{\\mathcal{A}}\\) are independently processed, forming graphs that propagate through both branches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.543, + 0.47, + 0.662 + ], + "angle": 0, + "content": "The SO(3) rotation equivariance of our encoder is ensured by the inherent equivariant properties of the Vector Neuron layers. To achieve T(3) translation equivariance, with an input point cloud \\( P = (p_{1}, p_{2}, \\dots, p_{n}) \\), \\( p_{i} \\in R^{3} \\), we compute its centroid \\( x = (\\Sigma_{i=1}^{n} p_{i}) / n \\), and get the input point cloud as \\( P' = P - x \\). In this way, our prediction is T(3) translation equivariant, i.e., \\( f \\) is our encoder and \\( \\mathcal{P} \\) is the original point cloud." + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.674, + 0.47, + 0.693 + ], + "angle": 0, + "content": "\\[\nf (\\mathcal {P} + \\mathcal {T}) = f (\\mathcal {P}) + \\mathcal {T}, \\quad T \\in R ^ {3} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.47, + 0.793 + ], + "angle": 0, + "content": "Our two-scale VN DGCNN employs dual K-nearest neighbor (KNN) values to extract features across two distinct scales, enhancing its ability to capture both local and global information. This pyramid structure enables the network to simultaneously grasp overall object shapes and fine-grained details, improving feature extraction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.334, + 0.818 + ], + "angle": 0, + "content": "5.3. Cross Object Fusion Module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We utilize point-wise multiplication, shown in Figure 4, as our cross object fusion module designed in \\( B_A \\). We fuse the feature of \\( \\mathcal{P}_B \\) and \\( \\mathcal{P}_A \\) by multiplying \\( \\mathcal{I}_B \\) and \\( \\mathcal{E}_A \\), so that each point in \\( \\mathcal{P}_A \\) will have the geometry feature of both \\( \\mathcal{P}_A \\) and \\( \\mathcal{P}_B \\). This approach integrates the geometric feature of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.497, + 0.893, + 0.543 + ], + "angle": 0, + "content": "\\(\\mathcal{P}_B\\) in each point while preserving the rotation equivariance of \\(\\mathcal{P}_A\\), i.e., \\(f\\) is an equivariant neural network, \\(R\\) is random rotation matrix," + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.555, + 0.893, + 0.574 + ], + "angle": 0, + "content": "\\[\nf \\left(R \\cdot \\left(\\mathcal {I} _ {\\mathcal {B}} * \\mathcal {E} _ {\\mathcal {A}}\\right)\\right) = R \\cdot f \\left(\\mathcal {I} _ {\\mathcal {B}} * \\mathcal {E} _ {\\mathcal {A}}\\right), \\quad R \\in R ^ {3 \\times 3} \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.586, + 0.656, + 0.601 + ], + "angle": 0, + "content": "5.4. Pose Prediction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.609, + 0.893, + 0.701 + ], + "angle": 0, + "content": "At both branches, we utilize two separate MLPs as our pose prediction head, to separately predict the translation \\( T \\in R^3 \\) and rotation \\( R \\in R^{3 \\times 3} \\). Compared to predicting translation and rotation within a single prediction head, this approach helps mitigate the issue of differing convergence speeds between the two components." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.711, + 0.795, + 0.728 + ], + "angle": 0, + "content": "5.5. Training and Evaluation Strategy" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We adopt a separate training and evaluation strategy for our network. To minimize the impact of pose prediction errors of \\(\\mathcal{P}_{\\mathcal{B}}\\) on \\(\\mathcal{P}_{\\mathcal{A}}\\), we train \\(\\mathcal{B}_{\\mathcal{A}}\\) and \\(\\mathcal{B}_{\\mathcal{B}}\\) independently. Specifically, for \\(\\mathcal{B}_{\\mathcal{A}}\\), during training, we utilize \\(\\mathcal{P}_{\\mathcal{B}}\\) under canonical pose, which is our ground truth point cloud of \\(\\mathcal{P}_{\\mathcal{B}}\\), to train our model. During testing, we first predict the pose of \\(\\mathcal{P}_{\\mathcal{B}}\\), then use the transformed \\(\\mathcal{P}_{\\mathcal{B}}\\), along with the initial \\(\\mathcal{P}_{\\mathcal{A}}\\) to predict A's pose, as shown in Figure 4. This phased, two-step training and evaluation strategy reduces errors caused by joint training of object poses, ensuring more accurate predictions." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17387" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.223, + 0.105 + ], + "angle": 0, + "content": "5.6. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.114, + 0.47, + 0.145 + ], + "angle": 0, + "content": "To train our network to robustly predict poses, we use the following equation as our loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.153, + 0.47, + 0.169 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {\\text {r o t}} \\mathcal {L} _ {\\text {r o t}} + \\lambda_ {\\text {t r a n s}} \\mathcal {L} _ {\\text {t r a n s}} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.176, + 0.47, + 0.237 + ], + "angle": 0, + "content": "Specifically, for predicted pose translation \\( T_{\\mathrm{pred}} \\in R^3 \\), rotation \\( \\mathcal{R}_{\\mathrm{pred}} \\in R^{3 \\times 3} \\) and ground truth pose translation \\( T_{\\mathrm{gt}} \\in R^3 \\) and rotation \\( \\mathcal{R}_{\\mathrm{gt}} \\in R^{3 \\times 3} \\), we use \\( \\mathcal{L}_1 \\) loss to compute our \\( \\mathcal{L}_{\\mathrm{trans}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.245, + 0.47, + 0.262 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t r a n s}} = \\mathcal {L} _ {1} \\left(T _ {\\text {p r e d}}, T _ {\\text {g t}}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.269, + 0.47, + 0.345 + ], + "angle": 0, + "content": "As for the rotation, we utilize Geodesic Distance, which measures the shortest path between two rotations on the rotation manifold. It offers a smooth and bounded angular error, ensuring stable gradients and accurately achieving precise rotation alignment." + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.351, + 0.47, + 0.392 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r o t}} = \\arccos \\left(\\frac {\\operatorname {t r} \\left(\\mathcal {R} _ {\\mathrm {g t}} \\mathcal {R} _ {\\text {p r e d}} ^ {T}\\right) - 1}{2}\\right) \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.399, + 0.21, + 0.416 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.424, + 0.47, + 0.469 + ], + "angle": 0, + "content": "In this section, we present a comprehensive evaluation and analysis of our two-step pairwise network architecture by addressing the following key questions:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.469, + 0.47, + 0.528 + ], + "angle": 0, + "content": "1. How does our network perform on 2BY2 tasks compared to existing baseline approaches, including matching-based, graph-network-based, and diffusion-based assembly methods?" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.53, + 0.47, + 0.574 + ], + "angle": 0, + "content": "2. How well does our network generalize across multiple tasks within the 2BY2 dataset? Can our network effectively handle a diverse set of tasks simultaneously?" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.575, + 0.469, + 0.59 + ], + "angle": 0, + "content": "3. Can our network generalize to real-world robot tasks?" + }, + { + "type": "list", + "bbox": [ + 0.077, + 0.469, + 0.47, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.598, + 0.36, + 0.614 + ], + "angle": 0, + "content": "6.1.2BY2 Dataset Main Experiment" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.621, + 0.266, + 0.637 + ], + "angle": 0, + "content": "6.1.1 Experiment Set Up" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.764 + ], + "angle": 0, + "content": "Tasks. We divide the 18 assembly tasks in the 2BY2 dataset into training and testing sets individually and compared the performance of our method with various baseline approaches. To further evaluate its cross-task generalization ability, we conducted additional experiments on tasks such as Lid Covering, Insertion, and High Precision Placement, as well as All task, which requires the method to handle all tasks in the entire dataset. See Table 2 for task details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.855 + ], + "angle": 0, + "content": "Evaluation metrics. Following metrics from datasets like Breaking Bad [41] and Neural Shape Mating [5], we use Root Mean Squared Error (RMSE) to evaluate both rotation and translation of the predicted SE(3) pose. Specifically, rotations are represented using Euler angles with symmetry considerations, see Section 3.2 for symmetry details." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Training parameters. We set batch size to be 4, and the initial learning rate of Adam Optimizer [10] to be 1e-4. We train models for 1000 epochs for them to fully converge." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.092, + 0.62, + 0.106 + ], + "angle": 0, + "content": "6.1.2 Baselines" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.115, + 0.894, + 0.145 + ], + "angle": 0, + "content": "We compare our method with SE-3 assembly [53], Puzzlefusion++ [50], Jigsaw [32] and Neural Shape Mating [5]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.146, + 0.894, + 0.206 + ], + "angle": 0, + "content": "- SE-3 Assembly [53] proposes a network architecture to leverage SE(3) equivariance for representations considering multi-part correlations, and predict the pose of each part jointly." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.206, + 0.894, + 0.265 + ], + "angle": 0, + "content": "- Puzzlefusion++ [50] proposes an auto-agglomerative 3D fracture assembly framework. It iteratively aligns and merges fragments using a diffusion model for 6-DoF alignment and a transformer model for verification." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.266, + 0.894, + 0.327 + ], + "angle": 0, + "content": "- Jigsaw [32] leverages hierarchical features of global and local geometry to match and align the fracture surfaces, and recovers the global pose of each piece to restore the underlying object." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.327, + 0.893, + 0.388 + ], + "angle": 0, + "content": "- Neural Shape Mating [5] utilizes PointNet for feature encoding and a transformer for feature fusion to learn the correlations between assembly parts, enabling joint prediction of their poses." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.146, + 0.894, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.407, + 0.827, + 0.422 + ], + "angle": 0, + "content": "6.1.3 2BY2 Benchmark Results and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.43, + 0.893, + 0.504 + ], + "angle": 0, + "content": "Table 3 presents the quantitative performance of our method compared to all baselines. The results show that our approach outperforms the baselines across 18 fine-grained assembly tasks, with an average improvement of 0.046 in translation RMSE and 8.97 in rotation RMSE." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.895, + 0.656 + ], + "angle": 0, + "content": "Additionally, we evaluate our method on three cross-category tasks defined in Section 3.1, namely Lid Covering, Inserting and High Precision Placing, and achieve the state-of-the-art performance. Moreover, in the most comprehensive All task, we outperform the baseline by 0.123 in translation and 10.90 in rotation, demonstrating strong generalization across tasks and object shapes. In the meantime, baseline comparisons confirm the rigor and challenge of our tasks. Results on challenging tasks like Plug and Key highlight our framework's effectiveness in complex scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.656, + 0.894, + 0.779 + ], + "angle": 0, + "content": "We analyze that the superior performance of our designed network is due to the approach of separately predicting the poses of the two objects in a step-by-step manner. This prevents the pose errors from interfering with each other, which often occurs in other baselines when predicting both poses simultaneously. Additionally, the design of our encoder makes our network more sensitive to subtle changes in rotation and translation, resulting better performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.773, + 0.804 + ], + "angle": 0, + "content": "6.2. Real-World Robot Experiment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Real-world robot experiment setup. As shown in Figure 5, we conduct our real-world robot experiments using a UR5 robotic arm, equipped with a Robotiq 2F-85 Gripper. We select four tasks, Cup, Flower, Bread and Plug, demonstrating that our model exhibits strong generalization ability on unseen real-world objects." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17388" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.089, + 0.895, + 0.399 + ], + "angle": 0, + "content": "
TaskJigsaw [32]Puzzlefusion++ [50]NSM [5]SE(3)-Assembly [53]Ours
RMSE(T)RMSE(R)RMSE(T)RMSE(R)RMSE(T)RMSE(R)RMSE(T)RMSE(R)RMSE(T)↓RMSE(R)↓
Lid Covering0.39833.330.40837.740.18433.450.12521.370.09016.12
Kitchenport0.47745.800.42347.230.23757.470.09317.290.06816.60
Bottle0.41134.710.38535.230.22778.580.14736.920.07627.70
Kettle0.33543.710.37238.380.21561.100.13313.150.11111.56
Coffeemachine0.52732.670.43734.640.25350.660.14224.430.07622.83
Cup0.40833.550.43933.580.26067.350.16046.460.12223.18
Inserting0.36453.580.32757.830.27569.930.19946.30.14238.03
Plug0.37256.890.34848.890.30352.260.17618.580.0949.74
Childrentoy0.26859.880.24563.210.27193.770.30280.520.24257.81
Letter0.40967.990.35772.080.31776.480.12139.240.09433.74
Bread0.22057.840.20160.920.17165.500.11151.130.09036.40
Nut0.47640.080.32347.290.27155.320.10246.680.05135.60
Coin0.40639.620.34851.400.28962.580.11128.690.10722.88
Key0.38442.850.34850.380.29063.600.08717.280.04516.32
Usb0.46367.410.34258.230.25269.900.21532.280.12828.98
Precision Placing0.37573.940.28767.810.21185.020.13457.860.11544.84
Box0.13733.720.13440.470.13072.470.07125.080.06621.53
Tissue0.29282.390.26585.180.17579.780.18373.020.11564.37
Flower0.32864.390.28359.040.24687.180.21364.890.12542.23
Teaport0.30268.330.32461.010.28856.320.08540.590.05026.11
Position0.42358.070.38957.550.25770.110.16628.170.14124.46
ALL0.36053.340.34258.230.28470.300.23352.340.11041.44
" + }, + { + "type": "table_caption", + "bbox": [ + 0.075, + 0.408, + 0.893, + 0.451 + ], + "angle": 0, + "content": "Table 3. Quantitative Evaluation on 2BY2 for Pairwise Object Assembly. Our method outperforms the baseline across all 18 fine-grained assembly tasks, as well as demonstrating significant improvement on three cross-category assembly tasks. It achieves a lower task average with a reduction of 0.046 in translation RMSE and 8.97 in rotation RMSE." + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.482, + 0.473, + 0.689 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.697, + 0.471, + 0.726 + ], + "angle": 0, + "content": "Figure 5. Real Robot Setup. We conduct real-world robot experiments on Cup, Flower, Bread and Plug tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We place objects in the scene with random initial poses and scan them to obtain their point clouds. Using pre-trained models on selected data of 2BY2 dataset, we predict the pose of each object. A manually designed grasping pose is then applied to pick up each object, and based on the predicted poses, the robotic arm plans a trajectory to complete the assembly. We use SE(3) assembly [53] as the baseline and test our approach on 10 different initial poses. As shown in Table 4, our method significantly outperforms the baseline." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.474, + 0.858, + 0.526 + ], + "angle": 0, + "content": "
TaskCupFlowerBreadPlugOverall
SE(3) [53]2/104/101/102/1022.5%
Ours8/1010/106/107/1077.5%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.503, + 0.536, + 0.888, + 0.55 + ], + "angle": 0, + "content": "Table 4. Real-World Robot Experiment Success Rate Results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.58, + 0.653, + 0.597 + ], + "angle": 0, + "content": "7. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.608, + 0.892, + 0.652 + ], + "angle": 0, + "content": "In this section, we conduct comprehensive experiments to demonstrate the rationality of our network design and the effectiveness of each module." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.655, + 0.894, + 0.701 + ], + "angle": 0, + "content": "Encoder. To validate the effectiveness of two-scale Vector Neuron(VN) DGCNN, we compare it with other encoders: VN DGCNN [9], DGCNN [48], and PointNet [38]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.894, + 0.794 + ], + "angle": 0, + "content": "Two-step network design. We compare our method with an end-to-end approach, which jointly predicts the pose of \\(\\mathcal{P}_A\\) and \\(\\mathcal{P}_{\\mathcal{B}}\\). Specifically, we utilize \\(\\mathcal{B}_A\\) and the input point cloud of \\(\\mathcal{P}_A\\) and \\(\\mathcal{P}_{\\mathcal{B}}\\) to get the 6D pose of \\(P_A\\), while using the same encoder to extract \\(\\mathcal{E}_{\\mathcal{B}}\\) and pass it through the same pose prediction head to predict the pose of \\(P_B\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.901 + ], + "angle": 0, + "content": "As shown in in Table 5, we show the results of ablation studies on Lid covering, Inserting and Precision Placing, which are more comprehensive and require cross-task generalization abilities. The performance declines in both translation and rotation when we removing our two-scale VN DGCNN encoder and change it to Vector Neuron DGCNN [9], DGCNN [48], Pointnet [38]. It demonstrates" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17389" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.891, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.519, + 0.895, + 0.548 + ], + "angle": 0, + "content": "Figure 6. Qualitative Results Comparison. We highlight Bottle, Plug, Bread, Letter, Childrentoy, Key, and Flower tasks to demonstrate our improved translation and rotation predictions compared to baseline methods." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.561, + 0.892, + 0.651 + ], + "angle": 0, + "content": "
TaskVector Neuron DGCNN [9]DGCNN [48]PointNet [38]w/o Two-stepOurs
RMSE(T)RMSE(R)RMSE(T)RMSE(R)RMSE(T)RMSE(R)RMSE(T)RMSE(R)RMSE(T) ↓RMSE(R) ↓
Lid Covering0.09818.230.24570.060.23465.320.11718.740.09016.12
Inserting0.15741.220.23462.250.23466.800.16441.110.14238.03
Precision Placing0.12148.010.24565.190.21172.470.13746.380.11544.84
ALL0.12344.670.27772.460.26475.380.13945.200.11041.44
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.66, + 0.895, + 0.703 + ], + "angle": 0, + "content": "Table 5. Ablation Study Results. We compare various encoders including Vector Neuron DGCNN [9], DGCNN [48], PointNet [38], and our proposed two-scale Vector Neuron DGCNN. We also compare end-to-end networks with two-step networks to demonstrate the effectiveness of each component in our network design." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.729, + 0.471, + 0.837 + ], + "angle": 0, + "content": "that our encoder effectively exploits the advantage of SE(3) equivariance, enabling greater sample efficiency and more robust generalization abilities. Compared with version in Figure 4, the experiment performance declines when we change our two-step network in a joint-learning manner, proving that our two-step network design can reduce error caused by jointly predictions and thereby is more effective." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.196, + 0.86 + ], + "angle": 0, + "content": "8. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.472, + 0.903 + ], + "angle": 0, + "content": "2BY2 is a significant step in bridging the gap between geometry-based assembly tasks and everyday object assem" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.729, + 0.895, + 0.896 + ], + "angle": 0, + "content": "blies. With pose and symmetry annotations for 517 object pairs across 18 fine-grained tasks, 2BY2 sets a new benchmark for 3D assembly challenges. Our two-step pairwise SE(3) pose estimation framework, which leverages equivariant features, demonstrates superior performance over existing approaches in reducing both translation and rotation errors. Robot experiments further validate the method's generalizability in practical 3D assembly scenarios. In conclusion, 2BY2 provides both a comprehensive benchmark and an effective framework, with the aim of inspiring and supporting more generalizable solution in robot manipulation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17390" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.472, + 0.143 + ], + "angle": 0, + "content": "[1] 3dwarehouse. https://3dwarehouse.sketchup.com/by/su3dchallenge, 2014.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.473, + 0.172 + ], + "angle": 0, + "content": "[2] google-sketchup-3d-challenge. https://3dwarehouse.sketchup.com/by/su3dchallenge, 2014.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.473, + 0.255 + ], + "angle": 0, + "content": "[3] Benedict J Brown, Corey Toler-Franklin, Diego Nehab, Michael Burns, David Dobkin, Andreas Vlachopoulos, Christos Doumas, Szymon Rusinkiewicz, and Tim Weyrich. A system for high-volume acquisition and matching of fresco fragments: Reassembling theran wall paintings. ACM transactions on graphics (TOG), 27(3):1-9, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.472, + 0.312 + ], + "angle": 0, + "content": "[4] Siddhartha Chaudhuri, Evangelos Kalogerakis, Leonidas Guibas, and Vladlen Koltun. Probabilistic reasoning for assembly-based 3d modeling. In ACM SIGGRAPH 2011 papers, pages 1-10. 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.313, + 0.472, + 0.382 + ], + "angle": 0, + "content": "[5] Yun-Chun Chen, Haoda Li, Dylan Turpin, Alec Jacobson, and Animesh Garg. Neural shape mating: Self-supervised object assembly with adversarial shape priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12724-12733, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.383, + 0.472, + 0.424 + ], + "angle": 0, + "content": "[6] Junfeng Cheng, Mingdong Wu, Ruiyuan Zhang, Guanqi Zhan, Chao Wu, and Hao Dong. Score-pa: Score-based 3d part assembly. In British Machine Vision Conference, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.425, + 0.472, + 0.452 + ], + "angle": 0, + "content": "[7] Taco S Cohen, Mario Geiger, Jonas Kohler, and Max Welling. Spherical cnns. arXiv preprint arXiv:1801.10130, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.472, + 0.522 + ], + "angle": 0, + "content": "[8] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of \\(10\\mathrm{m} + 3\\mathrm{d}\\) objects. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.524, + 0.472, + 0.592 + ], + "angle": 0, + "content": "[9] Congyue Deng, Or Litany, Yueqi Duan, Adrien Poulenard, Andrea Tagliasacchi, and Leonidas J Guibas. Vector neurons: A general framework for so (3)-equivariant networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12200-12209, 2021. 2, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.594, + 0.472, + 0.621 + ], + "angle": 0, + "content": "[10] P Kingma Diederik. Adam: A method for stochastic optimization. (No Title), 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.472, + 0.689 + ], + "angle": 0, + "content": "[11] Thomas Funkhouser, Michael Kazhdan, Philip Shilane, Patrick Min, William Kiefer, Ayellet Tal, Szymon Rusinkiewicz, and David Dobkin. Modeling by example. ACM transactions on graphics (TOG), 23(3):652-663, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.472, + 0.761 + ], + "angle": 0, + "content": "[12] Thomas Funkhouser, Hijung Shin, Corey Toler-Franklin, Antonio García Castañeda, Benedict Brown, David Dobkin, Szymon Rusinkiewicz, and Tim Weyrich. Learning how to match fresco fragments. Journal on Computing and Cultural Heritage (JOCCH), 4(2):1-13, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.472, + 0.829 + ], + "angle": 0, + "content": "[13] Chongkai Gao, Zhengrong Xue, Shuying Deng, Tianhai Liang, Siqi Yang, Lin Shao, and Huazhe Xu. Riemann: Near real-time se (3)-equivariant robot manipulation without point cloud segmentation. arXiv preprint arXiv:2403.19460, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[14] Stefan Hinterstoisser, Stefan Holzer, Cedric Cagniart, Slobodan Ilic, Kurt Konolige, Nassir Navab, and Vincent Lepetit. Multimodal templates for real-time detection of texture-less objects in heavily cluttered scenes. In 2011 International Conference on Computer Vision, pages 858-865, 2011. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.473, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.176 + ], + "angle": 0, + "content": "[15] Stefan Hinterstoisser, Vincent Lepetit, Slobodan Ilic, Stefan Holzer, Gary Bradski, Kurt Konolige, and Nassir Navab. Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In Computer Vision - ACCV 2012, pages 548-562, Berlin, Heidelberg, 2013. Springer Berlin Heidelberg. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.894, + 0.245 + ], + "angle": 0, + "content": "[16] Sepidehsadat Sepid Hossieni, Mohammad Amin Shabani, Saghar Irandoust, and Yasutaka Furukawa. Puzzle: unleashing the power of diffusion models for spatial puzzle solving. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.248, + 0.894, + 0.303 + ], + "angle": 0, + "content": "[17] Boce Hu, Xupeng Zhu, Dian Wang, Zihao Dong, Haojie Huang, Chenghao Wang, Robin Walters, and Robert Platt. Orbitgrasp: Se (3)-equivariant grasp learning. In 8th Annual Conference on Robot Learning, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.894, + 0.374 + ], + "angle": 0, + "content": "[18] Haojie Huang, Dian Wang, Xupeng Zhu, Robin Walters, and Robert Platt. Edge grasp network: A graph-based se (3)-invariant approach to grasp detection. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3882-3888. IEEE, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.893, + 0.431 + ], + "angle": 0, + "content": "[19] Haojie Huang, Haotian Liu, Dian Wang, Robin Walters, and Robert Platt. Match policy: A simple pipeline from point cloud registration to manipulation policies. arXiv preprint arXiv:2409.15517, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.894, + 0.502 + ], + "angle": 0, + "content": "[20] Haojie Huang, Karl Schmeckpeper, Dian Wang, Ondrej Biza, Yaoyao Qian, Haotian Liu, Mingxi Jia, Robert Platt, and Robin Walters. IMAGINATION POLICY: Using generative point cloud models for learning manipulation policies. In 8th Annual Conference on Robot Learning, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.894, + 0.559 + ], + "angle": 0, + "content": "[21] Qi-Xing Huang, Simon Flöry, Natasha Gelfand, Michael Hofer, and Helmut Pottmann. Reassembling fractured objects by geometric matching. In ACM Siggraph 2006 papers, pages 569-578. 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.894, + 0.615 + ], + "angle": 0, + "content": "[22] Prakhar Jaiswal, Jinmiao Huang, and Rahul Rai. Assembly-based conceptual 3d modeling with unlabeled components using probabilistic factor graph. Computer-Aided Design, 74: 45-54, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.617, + 0.894, + 0.672 + ], + "angle": 0, + "content": "[23] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J Davison. Rlbench: The robot learning benchmark & learning environment. IEEE Robotics and Automation Letters, 5(2):3019-3026, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.894, + 0.742 + ], + "angle": 0, + "content": "[24] Benjamin Jones, Dalton Hildreth, Duowen Chen, Ilya Baran, Vladimir G Kim, and Adriana Schulz. Automate: A dataset and learning approach for automatic mating of cad assemblies. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.745, + 0.894, + 0.814 + ], + "angle": 0, + "content": "[25] Yuanchen Ju, Kaizhe Hu, Guowei Zhang, Gu Zhang, Mingrun Jiang, and Huazhe Xu. Robo-abc: Affordance generalization beyond categories via semantic correspondence for robot manipulation. In European Conference on Computer Vision, pages 222–239. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.893, + 0.871 + ], + "angle": 0, + "content": "[26] Evangelos Kalogerakis, Siddhartha Chaudhuri, Daphne Koller, and Vladlen Koltun. A probabilistic model for component-based shape synthesis. Acm Transactions on Graphics (TOG), 31(4):1-11, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[27] Wadim Kehl, Fabian Manhardt, Federico Tombari, Slobodan Ilic, and Nassir Navab. SSD-6D: making rgb-based" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17391" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.471, + 0.12 + ], + "angle": 0, + "content": "3d detection and 6d pose estimation great again. CoRR, abs/1711.10006, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.192 + ], + "angle": 0, + "content": "[28] Yichen Li, Kaichun Mo, Lin Shao, Minhyuk Sung, and Leonidas Guibas. Learning 3d part assembly from a single image. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 664-682. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.471, + 0.234 + ], + "angle": 0, + "content": "[29] Yulong Li, Andy Zeng, and Shuran Song. Rearrangement planning for general part assembly. In 7th Annual Conference on Robot Learning, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.471, + 0.305 + ], + "angle": 0, + "content": "[30] Yuval Litvak, Armin Biess, and Aharon Bar-Hillel. Learning pose estimation for high-precision robotic assembly using simulated depth images. In 2019 International Conference on Robotics and Automation (ICRA), pages 3521–3527. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.471, + 0.376 + ], + "angle": 0, + "content": "[31] Hao-Yu Liu, Jian-Wei Guo, Hai-Yong Jiang, Yan-Chao Liu, Xiao-Peng Zhang, and Dong-Ming Yan. Puzzlenet: boundary-aware feature matching for non-overlapping 3d point clouds assembly. Journal of Computer Science and Technology, 38 (3):492-509, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.378, + 0.471, + 0.42 + ], + "angle": 0, + "content": "[32] Jiaxin Lu, Yifan Sun, and Qixing Huang. Jigsaw: Learning to assemble multiple fractured objects. Advances in Neural Information Processing Systems, 36, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.421, + 0.471, + 0.503 + ], + "angle": 0, + "content": "[33] Jianlan Luo, Eugen Solowjow, Chengtao Wen, Juan Aparicio Ojea, Alice M Agogino, Aviv Tamar, and Pieter Abbeel. Reinforcement learning on variable impedance controller for high-precision robotic assembly. In 2019 International Conference on Robotics and Automation (ICRA), pages 3080–3087. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.471, + 0.588 + ], + "angle": 0, + "content": "[34] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909-918, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.471, + 0.659 + ], + "angle": 0, + "content": "[35] Yashraj Narang, Kier Storey, Iretiayo Akinola, Miles Macklin, Philipp Reist, Lukasz Wawrzyniak, Yunrong Guo, Adam Moravanszky, Gavriel State, Michelle Lu, et al. Factory: Fast contact for robotic assembly. arXiv preprint arXiv:2205.03532, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.471, + 0.73 + ], + "angle": 0, + "content": "[36] Abhinav Narayan, Rajendra Nagar, and Shanmuganathan Ram. Rgl-net: A recurrent graph learning framework for progressive part assembly. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 78-87, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.471, + 0.787 + ], + "angle": 0, + "content": "[37] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.471, + 0.857 + ], + "angle": 0, + "content": "[38] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[39] Ismael Rodriguez, Korbinian Nottensteiner, Daniel Leidner, Michael KaBecker, Freek Stulp, and Alin Albu-Schaffer. Iteratively refined feasibility checks in robotic assembly sequence" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.894, + 0.12 + ], + "angle": 0, + "content": "planning. IEEE Robotics and Automation Letters, 4(2):1416-1423, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.192 + ], + "angle": 0, + "content": "[40] Gianluca Scarpellini, Stefano Fiorini, Francesco Giuliani, Pietro Moreiro, and Alessio Del Bue. Diffassemble: A unified graph-diffusion model for 2d and 3d reassembly. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28098-28108, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.893, + 0.248 + ], + "angle": 0, + "content": "[41] Silvia Sellán, Yun-Chun Chen, Ziyi Wu, Animesh Garg, and Alec Jacobson. Breaking bad: A dataset for geometric fracture and reassembly. Advances in Neural Information Processing Systems, 35:38885-38898, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.893, + 0.319 + ], + "angle": 0, + "content": "[42] Hijung Shin, Christos Doumas, Thomas Funkhouser, Szymon Rusinkiewicz, Kenneth Steiglitz, Andreas Vlachopoulos, and Tim Weyrich. Analyzing and simulating fracture patterns of theran wall paintings. Journal on Computing and Cultural Heritage (JOCCH), 5(3):1-14, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.894, + 0.39 + ], + "angle": 0, + "content": "[43] Chenrui Tie, Yue Chen, Ruihai Wu, Boxuan Dong, Zeyi Li, Chongkai Gao, and Hao Dong. ET-SEED: EFFICIENT TRAJECTORY-LEVEL SE(3) EQUIVARIANT DIFFUSION POLICY. In The Thirteenth International Conference on Learning Representations, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.893, + 0.446 + ], + "angle": 0, + "content": "[44] Joshua Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. CoRR, abs/1703.06907, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.893, + 0.503 + ], + "angle": 0, + "content": "[45] Jonathan Tremblay, Thang To, Balakumar Sundaralingam, Yu Xiang, Dieter Fox, and Stan Birchfield. Deep object pose estimation for semantic robotic grasping of household objects. ArXiv, abs/1809.10790, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.893, + 0.546 + ], + "angle": 0, + "content": "[46] Weiwei Wan, Kensuke Harada, and Kazuyuki Nagata. Assembly sequence planning for motion planning. CoRR, abs/1609.03108, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.893, + 0.603 + ], + "angle": 0, + "content": "[47] Ruocheng Wang, Yunzhi Zhang, Jiayuan Mao, Ran Zhang, Chin-Yi Cheng, and Jiajun Wu. Iea-manual: Seeing shape assembly step by step. Advances in Neural Information Processing Systems, 35:28428-28440, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.893, + 0.66 + ], + "angle": 0, + "content": "[48] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.893, + 0.703 + ], + "angle": 0, + "content": "[49] Ziming Wang and Rebecka Jornsten. Se (3)-bi-equivariant transformers for point cloud assembly. arXiv preprint arXiv:2407.09167, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.893, + 0.759 + ], + "angle": 0, + "content": "[50] Zhengqing Wang, Jiacheng Chen, and Yasutaka Furukawa. Puzzlefusion++: Auto-agglomerative 3d fracture assembly by denoise and verify. arXiv preprint arXiv:2406.00259, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.893, + 0.857 + ], + "angle": 0, + "content": "[51] Karl DD Willis, Pradeep Kumar Jayaraman, Hang Chu, Yunsheng Tian, Yifei Li, Daniele Grandi, Aditya Sanghi, Linh Tran, Joseph G Lambourne, Armando Solar-Lezama, et al. Joinable: Learning bottom-up assembly of parametric cad joints. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15849-15860, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.893, + 0.902 + ], + "angle": 0, + "content": "[52] Rundi Wu, Yixin Zhuang, Kai Xu, Hao Zhang, and Baoquan Chen. Pq-net: A generative part seq2seq network for 3d shapes. In Proceedings of the IEEE/CVF Conference on" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17392" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.119 + ], + "angle": 0, + "content": "Computer Vision and Pattern Recognition, pages 829-838, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.191 + ], + "angle": 0, + "content": "[53] Ruihai Wu, Chenrui Tie, Yushi Du, Yan Zhao, and Hao Dong. Leveraging se (3) equivariance for learning 3d geometric shape assembly. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14311-14320, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.472, + 0.275 + ], + "angle": 0, + "content": "[54] Fanbo Xiang, Yuzhe Qin, Kaichun Mo, Yikuan Xia, Hao Zhu, Fangchen Liu, Minghua Liu, Hanxiao Jiang, Yifu Yuan, He Wang, Li Yi, Angel X. Chang, Leonidas J. Guibas, and Hao Su. SAPIEN: A simulated part-based interactive environment. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.472, + 0.32 + ], + "angle": 0, + "content": "[55] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.472, + 0.389 + ], + "angle": 0, + "content": "[56] Yuwen Xiong, Wei-Chiu Ma, Jingkang Wang, and Raquel Urtasun. Learning compact representations for lidar completion and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1074-1083, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.472, + 0.433 + ], + "angle": 0, + "content": "[57] Boshen Xu, Sipeng Zheng, and Qin Jin. Spaformer: Sequential 3d part assembly with transformers. arXiv preprint arXiv:2403.05874, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.435, + 0.472, + 0.504 + ], + "angle": 0, + "content": "[58] Xianghao Xu, Paul Guerrero, Matthew Fisher, Siddhartha Chaudhuri, and Daniel Ritchie. Unsupervised 3d shape reconstruction by part retrieval and assembly. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8559-8567, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.472, + 0.575 + ], + "angle": 0, + "content": "[59] Zhengrong Xue, Zhecheng Yuan, Jiashun Wang, Xueqian Wang, Yang Gao, and Huazhe Xu. Usek: Unsupervised se (3)-equivariant 3d keypoints for generalizable manipulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1715-1722. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.472, + 0.645 + ], + "angle": 0, + "content": "[60] Kangxue Yin, Zhiqin Chen, Siddhartha Chaudhuri, Matthew Fisher, Vladimir G Kim, and Hao Zhang. Coalesce: Component assembly by learning to synthesize connections. In 2020 International Conference on 3D Vision (3DV), pages 61-70. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.472, + 0.715 + ], + "angle": 0, + "content": "[61] Kevin Zakka, Andy Zeng, Johnny Lee, and Shuran Song. Form2fit: Learning shape priors for generalizable assembly from disassembly. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9404–9410. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.472, + 0.773 + ], + "angle": 0, + "content": "[62] Guanqi Zhan, Qingnan Fan, Kaichun Mo, Lin Shao, Baoquan Chen, Leonidas J Guibas, Hao Dong, et al. Generative 3d part assembly via dynamic graph learning. Advances in Neural Information Processing Systems, 33:6315-6326, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.472, + 0.829 + ], + "angle": 0, + "content": "[63] Rufeng Zhang, Tao Kong, Weihao Wang, Xuan Han, and Mingyu You. 3d part assembly generation with instance encoded transformer. IEEE Robotics and Automation Letters, 7(4):9051-9058, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.472, + 0.899 + ], + "angle": 0, + "content": "[64] Ruiyuan Zhang, Jiaxiang Liu, Zexi Li, Hao Dong, Jie Fu, and Chao Wu. Scalable geometric fracture assembly via co-creation space among assemblers. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7269–7277, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.147 + ], + "angle": 0, + "content": "[65] Haibo Zhao, Dian Wang, Yizhe Zhu, Xupeng Zhu, Owen Howell, Linfeng Zhao, Yaoyao Qian, Robin Walters, and Robert Platt. Hierarchical equivariant policy via frame transf. arXiv preprint arXiv:2502.05728, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.219 + ], + "angle": 0, + "content": "[66] Junzhe Zhu, Yuanchen Ju, Junyi Zhang, Muhan Wang, Zhecheng Yuan, Kaizhe Hu, and Huazhe Xu. Densehunter: Learning 3d semantic correspondence for category-level manipulation from a single demo. International Conference on Learning Representations (ICLR) 2025, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.893, + 0.274 + ], + "angle": 0, + "content": "[67] Xupeng Zhu, Dian Wang, Guanang Su, Ondrej Biza, Robin Walters, and Robert Platt. On robot grasp learning using equivariant models. Autonomous Robots, 47(8):1175-1193, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.894, + 0.332 + ], + "angle": 0, + "content": "[68] Yuke Zhu, Josiah Wong, Ajay Mandlekar, Roberto Martin-Martin, Abhishek Joshi, Soroush Nasiriany, and Yifeng Zhu. robosuite: A modular simulation framework and benchmark for robot learning. arXiv preprint arXiv:2009.12293, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17393" + } + ] +] \ No newline at end of file diff --git a/2025/Two by Two_ Learning Multi-Task Pairwise Objects Assembly for Generalizable Robot Manipulation/57fa886e-a784-460f-94a7-de5f04b8ab96_origin.pdf b/2025/Two by Two_ Learning Multi-Task Pairwise Objects Assembly for Generalizable Robot Manipulation/57fa886e-a784-460f-94a7-de5f04b8ab96_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f0fc4a81e333a2cfdebcf56701c35baf4179ee32 --- /dev/null +++ b/2025/Two by Two_ Learning Multi-Task Pairwise Objects Assembly for Generalizable Robot Manipulation/57fa886e-a784-460f-94a7-de5f04b8ab96_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9dba31ec66282062bd2909844b6f0fdacc199aba1c2154361be80985231795b7 +size 4396273 diff --git a/2025/Two by Two_ Learning Multi-Task Pairwise Objects Assembly for Generalizable Robot Manipulation/full.md b/2025/Two by Two_ Learning Multi-Task Pairwise Objects Assembly for Generalizable Robot Manipulation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..aefa45126ec359173c2714320f84264c66cb0ffe --- /dev/null +++ b/2025/Two by Two_ Learning Multi-Task Pairwise Objects Assembly for Generalizable Robot Manipulation/full.md @@ -0,0 +1,330 @@ +# Two by Two: Learning Multi-Task Pairwise Objects Assembly for Generalizable Robot Manipulation + +Yu Qi $^{2,1*}$ Yuanchen Ju $^{1,3*}$ Tianming Wei $^{1,4}$ Chi Chu $^{1}$ Lawson L.S. Wong $^{2}$ Huazhe Xu $^{1,3,5\dagger}$ $^{1}$ Shanghai Qi Zhi Institute $^{2}$ Northeastern University $^{3}$ IIIS, Tsinghua University + $^{4}$ Shanghai Jiao Tong University $^{5}$ Shanghai AI Laboratory + +![](images/a5cbc2fb8e59f02077f4a25321f169f156b49bd3b50d7cf4ec0f05d3794fa9a5.jpg) +Figure 1. Overview of the 2BY2 Dataset. We propose the first large-scale daily pairwise object assembly dataset 2BY2, which contains 1,034 instances and 517 pairwise objects with pose and symmetry annotations. + +# Abstract + +3D assembly tasks, such as furniture assembly and component fitting, play a crucial role in daily life and represent essential capabilities for future home robots. Existing benchmarks and datasets predominantly focus on assembling geometric fragments or factory parts, which fall short in addressing the complexities of everyday object interactions and assemblies. To bridge this gap, we present 2BY2, a large-scale annotated dataset for daily pairwise objects assembly, covering 18 fine-grained tasks that reflect real-life scenarios, such as plugging into sockets, arranging flowers in vases, and inserting bread into toasters. 2BY2 dataset includes 1,034 instances and 517 pairwise objects with pose and symmetry annotations, requiring approaches that align geometric shapes while accounting for functional and spatial relationships between objects. Leveraging the 2BY2 dataset, we propose a two-step SE(3) pose estimation method with equivariant features for assembly constraints. Compared to previous shape assembly methods, our approach achieves state-of-the-art + +performance across all 18 tasks in the 2BY2 dataset. Additionally, robot experiments further validate the reliability and generalization ability of our method for complex 3D assembly tasks. More details and demonstrations can be found at https://tea-lab.github.io/TwoByTwo/. + +# 1. Introduction + +Assembly tasks are ubiquitous, such as assembling furniture, repairing household appliances, or putting together electronics. Successfully completing these tasks requires precise reasoning about the spatial relationships between pairs of objects. For robots to assist in these activities, they need to accurately estimate the 6D pose of each objects—including both their orientation and position in space. This capability is essential for domestic robots to help humans with various tasks, as it enables them to interact with their environment in a meaningful way. + +Daily object pairwise assembly not only requires considering the geometric constraints and spatial relationships + +between objects to achieve precise alignment but also needs to exhibit a certain level of generalization. Existing methods and benchmarks for solving assembly problems [31, 32, 49, 50, 63], typically focus on matching local geometric shapes, which often results in suboptimal performance in everyday assembly scenarios which require semantic and spatial alignment. This is because they are primarily trained and tested on existing assembly datasets that consist of large-scale geometric fragments, such as Breaking Bad [41] and Neural Shape Mating [5]. Compared with existing assembly tasks that focus on putting together fractures of objects, daily pairwise assembly tasks are more challenging and hold greater practical significance in human life. + +To bridge this gap, we introduce 2BY2, the first large-scale daily pairwise assembly dataset comprising 18 fine-grained tasks, shown as Figure 1. Compared to previous datasets and benchmarks, 2BY2 contains 1,034 instances and 517 pairwise objects with pose and symmetry annotations, covering a variety of pairwise assembly tasks reflecting everyday scenarios, which require approaches that align geometric shapes while accounting for functional and spatial relationships between objects, as shown in Table 1. + +Building on this dataset, we propose a two-step pairwise network architecture for assembly tasks. Mimicking the human assembly process like we firstly put the vase on the table and then arrange flower in it, our approach predicts the pose of each object in a step-by-step manner to assemble them to a predefined canonical space, which refers to a standard coordinate system that aligns with the principles of the human world, with detailed definition in Section 3.2. The network leverages a custom two-scale Vector Neuron DGCNN [9] encoder with spherical convolution [7] to extract SE(3) equivariant and SO(3) invariant features from point cloud inputs. Additionally, a feature fusion module and a two-step training and evaluation strategy are used to improve pose prediction accuracy. + +We evaluate our approach on 18 tasks in 2BY2 dataset to demonstrate the effectiveness on multi-task object pairwise assembly prediction. Compared to existing baselines, our method achieves an average improvement of 0.046 in translation RMSE and 8.97 in rotation RMSE. Moreover, we validate the effectiveness of our approach on three multicategory task, namely Lid Covering, Inserting and High Precision Placing, as well as the All task, which is defined in Section 2. Besides, real-world robot experiments validate the practical applicability of our approach. + +Our main contributions are listed as follows: + +1. We introduce 2BY2, the first large-scale daily pairwise object assembly dataset. By providing comprehensive pose and symmetry annotations for 517 pairwise objects across 18 fine-grained tasks, 2BY2 pushes the boundaries of real-world 3D assembly challenges and establishes a benchmark for pairwise assembly tasks. + +2. Our two-step pairwise SE(3) pose estimation method, leveraging equivariant geometric features, demonstrates superior performance compared to existing shape assembly methods, significantly reducing translation and rotation errors and enhancing the accuracy of 6D pose estimation. +3. Our approach achieves state-of-the-art performance on the benchmark, with real-world robot experiments demonstrating its capability, providing a generalizable solution for robot manipulation using pairwise object assembly. + +# 2. Related Work + +# 2.1. Object Assembly Benchmarks and Datasets + +Object reassembly has led to various datasets in computer vision and robotics. In computer vision, datasets like AutoMate [24] and JoinABLe [51] focus on reassembling fragments using geometric clues, while early datasets [3, 12, 21, 42] were limited in scale. Recent efforts, such as Neural Shape Mating [5] and Breaking Bad [41], generate largescale fractured object data using parametric segmentation. In robotics, benchmarks like Factory [35], RLBench [23], and RoboSuite [68] lack diverse shapes and assembly tasks under varying initial poses. In contrast, our dataset includes over 500 diverse object pairs across 3 categories and 18 assembly tasks, providing a comprehensive benchmark for pairwise object assembly, supporting the development of generalizable methods for real-world applications. + +# 2.2.3D Shape Assembly + +3D shape assembly [11, 30, 33, 56, 61], also known as part assembly, involves reconstructing objects from fragments, such as shattered sculptures or disassembled furniture. Existing methods use graphical models [4, 22, 26] and neural networks [6, 25, 28, 47, 52, 57, 58, 60, 64, 66] to capture geometric and semantic relationships. Approaches such as [5, 36, 62, 63] focus on pose estimation and part assembly without relying on predefined semantic information. Few-shot learning has been applied to assembly tasks [29], while jigsaw puzzle techniques [32, 37] leverage shape completion strategies. Recent works [16, 20, 40, 50] utilize diffusion models to refine poses or point clouds for assembly. In contrast, our method introduces a two-step pairwise network for step-by-step assembly, tailored to pairwise object alignment. + +# 2.3.6D Pose Estimation for Robot Manipulation + +6D pose estimation is crucial in robotics and computer vision for object interaction in unstructured environments [15, 45, 55]. Early handcrafted feature-based methods struggled in cluttered scenes [14, 15], while CNN-based approaches improved performance but lacked generalization [27, 55]. Domain randomization enhances robustness by varying synthetic datasets [44, 45]. In assembly tasks, 6D pose estimation aids manipulation planning with predefined + +
Dataset#OC#OSTask NumberPairTask HierarchyEveryday ScenarioSymmetryAssemble Type
PartNet [34]2426,671-NoNoNoNoSemantic
AutoMate [24]292,5291NoNoNoNoGeometric
JoinABLe [51]68,2511NoNoNoNoSemantic
NSM dataset[5]111,2461YesNoNoNoGeometric
Breaking Bad[41]-10,4741NoNoNoNoGeometric
Factory[35]8608YesNoNoNoGeometric
2BY2 Dataset (Ours)36103418YesYesYesYesGeometric and Semantic
+ +Table 1. Dataset Comparison. We compare 2BY2 dataset with existing datasets and benchmarks. #OC stands for the number of object categories. #OS stands for the number of object shapes. Pair denotes whether the dataset is pairwise. Task Number refers to the number of distinct assembly tasks, with the assembly of fractured pieces considered as a single task. Task Hierarchy stands for the different categories of task from coarse to fine, with ours shown in Section 3.1. Everyday Scenario means whether the assemble task has practical significance in real-world human applications. Symmetry denotes whether the dataset contains part symmetry annotation. + +objects [39, 46]. Like [13, 17-20, 43, 59, 65, 67], our method leverages equivariant features for efficient 6D pose learning and improved generalization. + +# 3.2BY2 Dataset + +# 3.1.2BY2 Dataset Overview + +We present the first large-scale 3D pairwise object assembly dataset for everyday scenarios, with detailed annotations for each object pair. The meshes in our dataset come from 3D Warehouse [1], SAPIEN PartNet-Mobility [54], Google SketchUp 3D Challenge [2], and Objaverse [8]. These meshes are manually paired, cleaned, annotated, and scaled uniformly. The 2BY2 dataset contains 517 unique pairs across three main tasks: Lid Covering, Inserting, and High Precision Placing, further subdivided into multiple subcategories, as shown in Table 2. + +# 3.2. Data Annotation + +To ensure high quality and reliability of our dataset, we conducted systematic cleaning and annotation of the collected meshes. First, we manually segment, integrate, and pair the meshes, classify them into Object $B$ and Object $A$ . Object $B$ is the base or the receiving component, such as the nut, the vase, the postbox. Object $A$ , is the fitting component, such as the bolt, the flower, the mail. This classification aligns with intuitive human assembly logic and supports our network's prediction strategy, such as positioning a nut before the bolt, as detailed in Sections 4 and 5.1. Automated scripts were used to uniformly scale meshes and align each pair to a canonical pose in world frame, defined as the object resting stably on the $XY$ plane with its lowest point aligned to $Z = 0$ . For instance, bottles and vases are aligned as if placed on a table, and mailboxes on the ground. + +During point cloud generation, we utilized blue noise sampling method [41] to extract point clouds uniformly from each mesh surface with dimension (1024, 3). We also annotated each object category with its inherent symmetry properties, specifically considering rotational symmetry along the + +$Z$ -axis, such as bottles, screws, and mirror symmetry along the $X$ -axis, such as bread, letters. + +# 3.3. Data Division and Task Diversity Analysis + +Our dataset provides diverse task coverage across categories, with each further divided into specific sub-categories, see Table 2. Objects within each category vary in shape, size, and type. To enhance generalization, the testing set includes objects with unseen geometric shapes, as shown in Figure 3. We also compute Chamfer Distance on point clouds between training and testing sets to quantify geometry differences, as shown in Figure 2. This diversity ensures generalization ability and applicability in real world scenarios and supports robust 3D matching and assembly tasks. + +![](images/fa84678daecd4f114ab87ad5c863115c41885046160521308f804e83832416b9.jpg) +Figure 2. Chamfer Distance Between Training and Testing Set. We normalize point clouds and compute the Chamfer Distance. For each task we calculate the distance separately between point cloud of Object $A$ and Object $B$ in the training set and test set. + +# 4. Problem Formulation + +The task takes two point clouds as input, namely $\mathcal{P}_A$ and $\mathcal{P}_B$ , each with dimension (1024, 3). These point clouds are derived from objects $\mathcal{O}_A$ and $\mathcal{O}_B$ from predefined canonical pose, as detailed in 3.2, respectively, and is randomly augmented with SO(3) rotation and being translated to its centroid. The desired output would be two individual SE(3) pose two assemble $\mathcal{O}_A$ and $\mathcal{O}_B$ to the canonical pose. + +
TaskLid CoveringInsertingHigh Precision Placing
KitBotKetCofCupPluChiLetBreNutCoiKeyUsbBoxTisFloTeaPos
Pair Num248628261614193224202120202520604221
+ +Table 2. 2BY2 Dataset Statistics Overview. The figure presents the number of object pairs across all task categories in the 2BY2 dataset, where each pair consists of two unique objects. The first row categorizes tasks into three major groups, while the second row provides a detailed breakdown of specific task categories. Specifically, Kit = Kitchenport, Bot = Bottle, Ket = Kettle, Cof = Coffee machine, Cup = Cup, Plu = Plug into socket, Chi = Children's toy, Let = Letter into mailbox, Bre = Bread into toaster, Nut = Bolt into nut, Coi = Coin into piggy bank, Key = Key into lock, Usb = USB cap, Box = Shoe boxing, Tis = Tissue placement on rack, Flo = Flower into vase, Tea = Teaware arrangement on tray, Pos = Positioning a cup on the coffee machine for coffee dispensing. + +![](images/0d3ba6708fc5f95c010232ecbee4aa425fc8af00308e3d5b7b04880c20e67bd3.jpg) +Figure 3. Task Diversity Visualization. The image shows selected objects from four different tasks: USB, Bottle, Letter, and Plug in Socket. On the left are the objects selected on training set, and on the right is the testing set. As seen in the legend, object geometry varies in both the training and testing set, with the testing set containing novel shapes not seen in the training set. + +![](images/aa6cd26a3d7f7e604524c8ec7094bb194075a6c98b0b8179ba5f33a465875c90.jpg) + +# 5. Method + +# 5.1. Two-step Pairwise Network Architecture + +To effectively learn pairwise object assembly, we propose a two-step pairwise network architecture with two branches: Branch B $(\mathcal{B}_{\mathcal{B}})$ and Branch A $(\mathcal{B}_{\mathcal{A}})$ , as shown in Figure 4. Branch B predicts the pose of $\mathcal{P}_{\mathcal{B}}$ , which is the socket, using a two-scale Vector Neuron DGCNN encoder [9] to extract SE(3) equivariant features, denoted as $\mathcal{E}_{\mathcal{B}}$ , followed by MLP-based pose prediction heads for translation and rotation. The transformed $\mathcal{P}_{\mathcal{B}}$ and inserter object $\mathcal{P}_{\mathcal{A}}$ , which is the plug, are then passed to Branch A, which extracts SE(3) equivariant features, denoted as $\mathcal{E}_{\mathcal{A}}$ and SO(3) invariant features $(\mathcal{I}_{\mathcal{B}})$ . The features are fused through element-wise multiplication, allowing $\mathcal{B}_{\mathcal{A}}$ to predict the pose of $\mathcal{P}_{\mathcal{A}}$ using information from both objects. This architecture ensures geometric alignment and matching by leveraging shared feature representations while reducing feature interference. + +Our two-step pairwise network is inspired by the human approach to pairwise assembly tasks. For example, when ar- + +ranging a vase with flowers, one intuitively first positions the vase correctly before placing the flowers inside. Similarly, inserting an envelope into a mailbox requires identifying the mailbox slot's pose first. By mimicking this sequential strategy, our model simulates human decision-making process, enabling more efficient and accurate assembly tasks. + +# 5.2. Two-scale SE(3) Equivariant and SO(3) Invariant Feature Extraction + +We employ a two-scale SE(3) Vector Neuron DGCNN, an enhanced variant of the original Vector Neuron DGCNN [9], as our encoder to extract SE(3) equivariant and SO(3) invariant features. This architecture leverages equivariance to improve sample efficiency of the model, while incorporating a two-scale information fusion mechanism to capture geometric features at two different scales. + +SE(3) Equivariance and SO(3) Invariance. SE(3) equivariance combines SO(3) rotation and T(3) translation equivariance: rotation equivariance ensures that a network's output rotates with the input, while translation equivariance shifts the output accordingly. SO(3) invariance means the network's output remains unchanged under any 3D rotation. By leveraging SE(3) equivariance, the model benefits from improved sample efficiency and generalization. This is particularly advantageous for assembly tasks, where objects may appear in arbitrary poses. + +Vector Neuron DGCNN. The Vector Neuron Network [9] extends traditional neurons from scalars to 3D vectors, designs vector-based convolutional layers and nonlinear functions like pooling and ReLU to support SO(3) equivariant and SO(3) invariant feature extraction. VNN operates in vector space, captures richer geometric relationships and ensures more robust feature representations for downstream tasks. + +Two-scale Vector Neuron DGCNN. We propose a two-scale Vector Neuron DGCNN for extracting SE(3) equivariant and SO(3) invariant features $\mathcal{E}_{\mathcal{B}}$ , $\mathcal{I}_{\mathcal{B}}$ , and $\mathcal{E}_{\mathcal{A}}$ . As shown in Figure 4, the encoder comprises two branches with different $K$ values, each consisting of multiple Vector Neuron convolutional layers followed by pooling. The extracted features from both branches are concatenated and further pro + +![](images/426fdc6ea95e7999c713f3f9b402dc937e672444a66d44b20ab891b96882fccb.jpg) + +![](images/ec389a409c21357f0ba92a5925b7ea26c9a8c1ce5489b63d37b43bee8c8b2ea3.jpg) +Inference Stage + +![](images/ee017c079a0e93900d5a1be86390829fd704ea4a7869d6b55c1803bc4b88ddf6.jpg) + +![](images/becc995c74a834c658d0cd186207979920259cc0fb5358cd4c2df100e69d8e30.jpg) + +![](images/15b00f593880eaf4dbb0ba131eed6e52d0012e5b76d1cd0cecc6c8de59b64b83.jpg) + +![](images/288ef862b3bdd968b1a0c42c80bfdfd0b3a1d4a8c5910493018c8f8e55f46ec0.jpg) +Robot Experiments +Figure 4. Our Two-Step Pairwise Network. We utilize two-scale VN DGCNN as our encoder to extract equivariant and invariant feature. We first predict the canonical pose of $\mathcal{O}_B$ and then predict the pose of $\mathcal{O}_A$ according to it. + +![](images/19b59556fd9937072e7e0aed90b18510e3e2fcba159b3febf5b5fda5dc08897f.jpg) +Training Stage + +cessed through an additional Vector Neuron convolutional layer. Point clouds $\mathcal{P}_{\mathcal{B}}$ and $\mathcal{P}_{\mathcal{A}}$ are independently processed, forming graphs that propagate through both branches. + +The SO(3) rotation equivariance of our encoder is ensured by the inherent equivariant properties of the Vector Neuron layers. To achieve T(3) translation equivariance, with an input point cloud $P = (p_{1}, p_{2}, \dots, p_{n})$ , $p_{i} \in R^{3}$ , we compute its centroid $x = (\Sigma_{i=1}^{n} p_{i}) / n$ , and get the input point cloud as $P' = P - x$ . In this way, our prediction is T(3) translation equivariant, i.e., $f$ is our encoder and $\mathcal{P}$ is the original point cloud. + +$$ +f (\mathcal {P} + \mathcal {T}) = f (\mathcal {P}) + \mathcal {T}, \quad T \in R ^ {3} \tag {1} +$$ + +Our two-scale VN DGCNN employs dual K-nearest neighbor (KNN) values to extract features across two distinct scales, enhancing its ability to capture both local and global information. This pyramid structure enables the network to simultaneously grasp overall object shapes and fine-grained details, improving feature extraction. + +# 5.3. Cross Object Fusion Module + +We utilize point-wise multiplication, shown in Figure 4, as our cross object fusion module designed in $B_A$ . We fuse the feature of $\mathcal{P}_B$ and $\mathcal{P}_A$ by multiplying $\mathcal{I}_B$ and $\mathcal{E}_A$ , so that each point in $\mathcal{P}_A$ will have the geometry feature of both $\mathcal{P}_A$ and $\mathcal{P}_B$ . This approach integrates the geometric feature of + +$\mathcal{P}_B$ in each point while preserving the rotation equivariance of $\mathcal{P}_A$ , i.e., $f$ is an equivariant neural network, $R$ is random rotation matrix, + +$$ +f \left(R \cdot \left(\mathcal {I} _ {\mathcal {B}} * \mathcal {E} _ {\mathcal {A}}\right)\right) = R \cdot f \left(\mathcal {I} _ {\mathcal {B}} * \mathcal {E} _ {\mathcal {A}}\right), \quad R \in R ^ {3 \times 3} \tag {2} +$$ + +# 5.4. Pose Prediction + +At both branches, we utilize two separate MLPs as our pose prediction head, to separately predict the translation $T \in R^3$ and rotation $R \in R^{3 \times 3}$ . Compared to predicting translation and rotation within a single prediction head, this approach helps mitigate the issue of differing convergence speeds between the two components. + +# 5.5. Training and Evaluation Strategy + +We adopt a separate training and evaluation strategy for our network. To minimize the impact of pose prediction errors of $\mathcal{P}_{\mathcal{B}}$ on $\mathcal{P}_{\mathcal{A}}$ , we train $\mathcal{B}_{\mathcal{A}}$ and $\mathcal{B}_{\mathcal{B}}$ independently. Specifically, for $\mathcal{B}_{\mathcal{A}}$ , during training, we utilize $\mathcal{P}_{\mathcal{B}}$ under canonical pose, which is our ground truth point cloud of $\mathcal{P}_{\mathcal{B}}$ , to train our model. During testing, we first predict the pose of $\mathcal{P}_{\mathcal{B}}$ , then use the transformed $\mathcal{P}_{\mathcal{B}}$ , along with the initial $\mathcal{P}_{\mathcal{A}}$ to predict A's pose, as shown in Figure 4. This phased, two-step training and evaluation strategy reduces errors caused by joint training of object poses, ensuring more accurate predictions. + +# 5.6. Loss Function + +To train our network to robustly predict poses, we use the following equation as our loss function: + +$$ +\mathcal {L} = \lambda_ {\text {r o t}} \mathcal {L} _ {\text {r o t}} + \lambda_ {\text {t r a n s}} \mathcal {L} _ {\text {t r a n s}} \tag {3} +$$ + +Specifically, for predicted pose translation $T_{\mathrm{pred}} \in R^3$ , rotation $\mathcal{R}_{\mathrm{pred}} \in R^{3 \times 3}$ and ground truth pose translation $T_{\mathrm{gt}} \in R^3$ and rotation $\mathcal{R}_{\mathrm{gt}} \in R^{3 \times 3}$ , we use $\mathcal{L}_1$ loss to compute our $\mathcal{L}_{\mathrm{trans}}$ : + +$$ +\mathcal {L} _ {\text {t r a n s}} = \mathcal {L} _ {1} \left(T _ {\text {p r e d}}, T _ {\text {g t}}\right) \tag {4} +$$ + +As for the rotation, we utilize Geodesic Distance, which measures the shortest path between two rotations on the rotation manifold. It offers a smooth and bounded angular error, ensuring stable gradients and accurately achieving precise rotation alignment. + +$$ +\mathcal {L} _ {\text {r o t}} = \arccos \left(\frac {\operatorname {t r} \left(\mathcal {R} _ {\mathrm {g t}} \mathcal {R} _ {\text {p r e d}} ^ {T}\right) - 1}{2}\right) \tag {5} +$$ + +# 6. Experiments + +In this section, we present a comprehensive evaluation and analysis of our two-step pairwise network architecture by addressing the following key questions: + +1. How does our network perform on 2BY2 tasks compared to existing baseline approaches, including matching-based, graph-network-based, and diffusion-based assembly methods? +2. How well does our network generalize across multiple tasks within the 2BY2 dataset? Can our network effectively handle a diverse set of tasks simultaneously? +3. Can our network generalize to real-world robot tasks? + +# 6.1.2BY2 Dataset Main Experiment + +# 6.1.1 Experiment Set Up + +Tasks. We divide the 18 assembly tasks in the 2BY2 dataset into training and testing sets individually and compared the performance of our method with various baseline approaches. To further evaluate its cross-task generalization ability, we conducted additional experiments on tasks such as Lid Covering, Insertion, and High Precision Placement, as well as All task, which requires the method to handle all tasks in the entire dataset. See Table 2 for task details. + +Evaluation metrics. Following metrics from datasets like Breaking Bad [41] and Neural Shape Mating [5], we use Root Mean Squared Error (RMSE) to evaluate both rotation and translation of the predicted SE(3) pose. Specifically, rotations are represented using Euler angles with symmetry considerations, see Section 3.2 for symmetry details. + +Training parameters. We set batch size to be 4, and the initial learning rate of Adam Optimizer [10] to be 1e-4. We train models for 1000 epochs for them to fully converge. + +# 6.1.2 Baselines + +We compare our method with SE-3 assembly [53], Puzzlefusion++ [50], Jigsaw [32] and Neural Shape Mating [5]. + +- SE-3 Assembly [53] proposes a network architecture to leverage SE(3) equivariance for representations considering multi-part correlations, and predict the pose of each part jointly. +- Puzzlefusion++ [50] proposes an auto-agglomerative 3D fracture assembly framework. It iteratively aligns and merges fragments using a diffusion model for 6-DoF alignment and a transformer model for verification. +- Jigsaw [32] leverages hierarchical features of global and local geometry to match and align the fracture surfaces, and recovers the global pose of each piece to restore the underlying object. +- Neural Shape Mating [5] utilizes PointNet for feature encoding and a transformer for feature fusion to learn the correlations between assembly parts, enabling joint prediction of their poses. + +# 6.1.3 2BY2 Benchmark Results and Analysis + +Table 3 presents the quantitative performance of our method compared to all baselines. The results show that our approach outperforms the baselines across 18 fine-grained assembly tasks, with an average improvement of 0.046 in translation RMSE and 8.97 in rotation RMSE. + +Additionally, we evaluate our method on three cross-category tasks defined in Section 3.1, namely Lid Covering, Inserting and High Precision Placing, and achieve the state-of-the-art performance. Moreover, in the most comprehensive All task, we outperform the baseline by 0.123 in translation and 10.90 in rotation, demonstrating strong generalization across tasks and object shapes. In the meantime, baseline comparisons confirm the rigor and challenge of our tasks. Results on challenging tasks like Plug and Key highlight our framework's effectiveness in complex scenarios. + +We analyze that the superior performance of our designed network is due to the approach of separately predicting the poses of the two objects in a step-by-step manner. This prevents the pose errors from interfering with each other, which often occurs in other baselines when predicting both poses simultaneously. Additionally, the design of our encoder makes our network more sensitive to subtle changes in rotation and translation, resulting better performance. + +# 6.2. Real-World Robot Experiment + +Real-world robot experiment setup. As shown in Figure 5, we conduct our real-world robot experiments using a UR5 robotic arm, equipped with a Robotiq 2F-85 Gripper. We select four tasks, Cup, Flower, Bread and Plug, demonstrating that our model exhibits strong generalization ability on unseen real-world objects. + +
TaskJigsaw [32]Puzzlefusion++ [50]NSM [5]SE(3)-Assembly [53]Ours
RMSE(T)RMSE(R)RMSE(T)RMSE(R)RMSE(T)RMSE(R)RMSE(T)RMSE(R)RMSE(T)↓RMSE(R)↓
Lid Covering0.39833.330.40837.740.18433.450.12521.370.09016.12
Kitchenport0.47745.800.42347.230.23757.470.09317.290.06816.60
Bottle0.41134.710.38535.230.22778.580.14736.920.07627.70
Kettle0.33543.710.37238.380.21561.100.13313.150.11111.56
Coffeemachine0.52732.670.43734.640.25350.660.14224.430.07622.83
Cup0.40833.550.43933.580.26067.350.16046.460.12223.18
Inserting0.36453.580.32757.830.27569.930.19946.30.14238.03
Plug0.37256.890.34848.890.30352.260.17618.580.0949.74
Childrentoy0.26859.880.24563.210.27193.770.30280.520.24257.81
Letter0.40967.990.35772.080.31776.480.12139.240.09433.74
Bread0.22057.840.20160.920.17165.500.11151.130.09036.40
Nut0.47640.080.32347.290.27155.320.10246.680.05135.60
Coin0.40639.620.34851.400.28962.580.11128.690.10722.88
Key0.38442.850.34850.380.29063.600.08717.280.04516.32
Usb0.46367.410.34258.230.25269.900.21532.280.12828.98
Precision Placing0.37573.940.28767.810.21185.020.13457.860.11544.84
Box0.13733.720.13440.470.13072.470.07125.080.06621.53
Tissue0.29282.390.26585.180.17579.780.18373.020.11564.37
Flower0.32864.390.28359.040.24687.180.21364.890.12542.23
Teaport0.30268.330.32461.010.28856.320.08540.590.05026.11
Position0.42358.070.38957.550.25770.110.16628.170.14124.46
ALL0.36053.340.34258.230.28470.300.23352.340.11041.44
+ +![](images/d63afcedd816608ab0e5ef1a9cc93c4eca7d77de0aa3eab58cda76220feee797.jpg) +Figure 5. Real Robot Setup. We conduct real-world robot experiments on Cup, Flower, Bread and Plug tasks. + +We place objects in the scene with random initial poses and scan them to obtain their point clouds. Using pre-trained models on selected data of 2BY2 dataset, we predict the pose of each object. A manually designed grasping pose is then applied to pick up each object, and based on the predicted poses, the robotic arm plans a trajectory to complete the assembly. We use SE(3) assembly [53] as the baseline and test our approach on 10 different initial poses. As shown in Table 4, our method significantly outperforms the baseline. + +Table 3. Quantitative Evaluation on 2BY2 for Pairwise Object Assembly. Our method outperforms the baseline across all 18 fine-grained assembly tasks, as well as demonstrating significant improvement on three cross-category assembly tasks. It achieves a lower task average with a reduction of 0.046 in translation RMSE and 8.97 in rotation RMSE. + +
TaskCupFlowerBreadPlugOverall
SE(3) [53]2/104/101/102/1022.5%
Ours8/1010/106/107/1077.5%
+ +Table 4. Real-World Robot Experiment Success Rate Results. + +# 7. Ablation Study + +In this section, we conduct comprehensive experiments to demonstrate the rationality of our network design and the effectiveness of each module. + +Encoder. To validate the effectiveness of two-scale Vector Neuron(VN) DGCNN, we compare it with other encoders: VN DGCNN [9], DGCNN [48], and PointNet [38]. + +Two-step network design. We compare our method with an end-to-end approach, which jointly predicts the pose of $\mathcal{P}_A$ and $\mathcal{P}_{\mathcal{B}}$ . Specifically, we utilize $\mathcal{B}_A$ and the input point cloud of $\mathcal{P}_A$ and $\mathcal{P}_{\mathcal{B}}$ to get the 6D pose of $P_A$ , while using the same encoder to extract $\mathcal{E}_{\mathcal{B}}$ and pass it through the same pose prediction head to predict the pose of $P_B$ . + +As shown in in Table 5, we show the results of ablation studies on Lid covering, Inserting and Precision Placing, which are more comprehensive and require cross-task generalization abilities. The performance declines in both translation and rotation when we removing our two-scale VN DGCNN encoder and change it to Vector Neuron DGCNN [9], DGCNN [48], Pointnet [38]. It demonstrates + +![](images/d4f59e3949644bd077a197c54f2a47ce948e6083951ed494a49e90ed2345e1bd.jpg) +Figure 6. Qualitative Results Comparison. We highlight Bottle, Plug, Bread, Letter, Childrentoy, Key, and Flower tasks to demonstrate our improved translation and rotation predictions compared to baseline methods. + +
TaskVector Neuron DGCNN [9]DGCNN [48]PointNet [38]w/o Two-stepOurs
RMSE(T)RMSE(R)RMSE(T)RMSE(R)RMSE(T)RMSE(R)RMSE(T)RMSE(R)RMSE(T) ↓RMSE(R) ↓
Lid Covering0.09818.230.24570.060.23465.320.11718.740.09016.12
Inserting0.15741.220.23462.250.23466.800.16441.110.14238.03
Precision Placing0.12148.010.24565.190.21172.470.13746.380.11544.84
ALL0.12344.670.27772.460.26475.380.13945.200.11041.44
+ +Table 5. Ablation Study Results. We compare various encoders including Vector Neuron DGCNN [9], DGCNN [48], PointNet [38], and our proposed two-scale Vector Neuron DGCNN. We also compare end-to-end networks with two-step networks to demonstrate the effectiveness of each component in our network design. + +that our encoder effectively exploits the advantage of SE(3) equivariance, enabling greater sample efficiency and more robust generalization abilities. Compared with version in Figure 4, the experiment performance declines when we change our two-step network in a joint-learning manner, proving that our two-step network design can reduce error caused by jointly predictions and thereby is more effective. + +# 8. Conclusion + +2BY2 is a significant step in bridging the gap between geometry-based assembly tasks and everyday object assem + +blies. With pose and symmetry annotations for 517 object pairs across 18 fine-grained tasks, 2BY2 sets a new benchmark for 3D assembly challenges. Our two-step pairwise SE(3) pose estimation framework, which leverages equivariant features, demonstrates superior performance over existing approaches in reducing both translation and rotation errors. Robot experiments further validate the method's generalizability in practical 3D assembly scenarios. In conclusion, 2BY2 provides both a comprehensive benchmark and an effective framework, with the aim of inspiring and supporting more generalizable solution in robot manipulation. + +# References + +[1] 3dwarehouse. https://3dwarehouse.sketchup.com/by/su3dchallenge, 2014.3 +[2] google-sketchup-3d-challenge. https://3dwarehouse.sketchup.com/by/su3dchallenge, 2014.3 +[3] Benedict J Brown, Corey Toler-Franklin, Diego Nehab, Michael Burns, David Dobkin, Andreas Vlachopoulos, Christos Doumas, Szymon Rusinkiewicz, and Tim Weyrich. A system for high-volume acquisition and matching of fresco fragments: Reassembling theran wall paintings. ACM transactions on graphics (TOG), 27(3):1-9, 2008. 2 +[4] Siddhartha Chaudhuri, Evangelos Kalogerakis, Leonidas Guibas, and Vladlen Koltun. Probabilistic reasoning for assembly-based 3d modeling. In ACM SIGGRAPH 2011 papers, pages 1-10. 2011. 2 +[5] Yun-Chun Chen, Haoda Li, Dylan Turpin, Alec Jacobson, and Animesh Garg. Neural shape mating: Self-supervised object assembly with adversarial shape priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12724-12733, 2022. 2, 3, 6, 7 +[6] Junfeng Cheng, Mingdong Wu, Ruiyuan Zhang, Guanqi Zhan, Chao Wu, and Hao Dong. Score-pa: Score-based 3d part assembly. In British Machine Vision Conference, 2023. 2 +[7] Taco S Cohen, Mario Geiger, Jonas Kohler, and Max Welling. Spherical cnns. arXiv preprint arXiv:1801.10130, 2018. 2 +[8] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects. Advances in Neural Information Processing Systems, 36, 2024. 3 +[9] Congyue Deng, Or Litany, Yueqi Duan, Adrien Poulenard, Andrea Tagliasacchi, and Leonidas J Guibas. Vector neurons: A general framework for so (3)-equivariant networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12200-12209, 2021. 2, 4, 7, 8 +[10] P Kingma Diederik. Adam: A method for stochastic optimization. (No Title), 2014. 6 +[11] Thomas Funkhouser, Michael Kazhdan, Philip Shilane, Patrick Min, William Kiefer, Ayellet Tal, Szymon Rusinkiewicz, and David Dobkin. Modeling by example. ACM transactions on graphics (TOG), 23(3):652-663, 2004. 2 +[12] Thomas Funkhouser, Hijung Shin, Corey Toler-Franklin, Antonio García Castañeda, Benedict Brown, David Dobkin, Szymon Rusinkiewicz, and Tim Weyrich. Learning how to match fresco fragments. Journal on Computing and Cultural Heritage (JOCCH), 4(2):1-13, 2011. 2 +[13] Chongkai Gao, Zhengrong Xue, Shuying Deng, Tianhai Liang, Siqi Yang, Lin Shao, and Huazhe Xu. Riemann: Near real-time se (3)-equivariant robot manipulation without point cloud segmentation. arXiv preprint arXiv:2403.19460, 2024. 3 +[14] Stefan Hinterstoisser, Stefan Holzer, Cedric Cagniart, Slobodan Ilic, Kurt Konolige, Nassir Navab, and Vincent Lepetit. Multimodal templates for real-time detection of texture-less objects in heavily cluttered scenes. In 2011 International Conference on Computer Vision, pages 858-865, 2011. 2 + +[15] Stefan Hinterstoisser, Vincent Lepetit, Slobodan Ilic, Stefan Holzer, Gary Bradski, Kurt Konolige, and Nassir Navab. Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In Computer Vision - ACCV 2012, pages 548-562, Berlin, Heidelberg, 2013. Springer Berlin Heidelberg. 2 +[16] Sepidehsadat Sepid Hossieni, Mohammad Amin Shabani, Saghar Irandoust, and Yasutaka Furukawa. Puzzle: unleashing the power of diffusion models for spatial puzzle solving. Advances in Neural Information Processing Systems, 36, 2024. 2 +[17] Boce Hu, Xupeng Zhu, Dian Wang, Zihao Dong, Haojie Huang, Chenghao Wang, Robin Walters, and Robert Platt. Orbitgrasp: Se (3)-equivariant grasp learning. In 8th Annual Conference on Robot Learning, 2024. 3 +[18] Haojie Huang, Dian Wang, Xupeng Zhu, Robin Walters, and Robert Platt. Edge grasp network: A graph-based se (3)-invariant approach to grasp detection. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3882-3888. IEEE, 2023. +[19] Haojie Huang, Haotian Liu, Dian Wang, Robin Walters, and Robert Platt. Match policy: A simple pipeline from point cloud registration to manipulation policies. arXiv preprint arXiv:2409.15517, 2024. +[20] Haojie Huang, Karl Schmeckpeper, Dian Wang, Ondrej Biza, Yaoyao Qian, Haotian Liu, Mingxi Jia, Robert Platt, and Robin Walters. IMAGINATION POLICY: Using generative point cloud models for learning manipulation policies. In 8th Annual Conference on Robot Learning, 2024. 2, 3 +[21] Qi-Xing Huang, Simon Flöry, Natasha Gelfand, Michael Hofer, and Helmut Pottmann. Reassembling fractured objects by geometric matching. In ACM Siggraph 2006 papers, pages 569-578. 2006. 2 +[22] Prakhar Jaiswal, Jinmiao Huang, and Rahul Rai. Assembly-based conceptual 3d modeling with unlabeled components using probabilistic factor graph. Computer-Aided Design, 74: 45-54, 2016. 2 +[23] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J Davison. Rlbench: The robot learning benchmark & learning environment. IEEE Robotics and Automation Letters, 5(2):3019-3026, 2020. 2 +[24] Benjamin Jones, Dalton Hildreth, Duowen Chen, Ilya Baran, Vladimir G Kim, and Adriana Schulz. Automate: A dataset and learning approach for automatic mating of cad assemblies. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021. 2, 3 +[25] Yuanchen Ju, Kaizhe Hu, Guowei Zhang, Gu Zhang, Mingrun Jiang, and Huazhe Xu. Robo-abc: Affordance generalization beyond categories via semantic correspondence for robot manipulation. In European Conference on Computer Vision, pages 222–239. Springer, 2025. 2 +[26] Evangelos Kalogerakis, Siddhartha Chaudhuri, Daphne Koller, and Vladlen Koltun. A probabilistic model for component-based shape synthesis. Acm Transactions on Graphics (TOG), 31(4):1-11, 2012. 2 +[27] Wadim Kehl, Fabian Manhardt, Federico Tombari, Slobodan Ilic, and Nassir Navab. SSD-6D: making rgb-based + +3d detection and 6d pose estimation great again. CoRR, abs/1711.10006, 2017. 2 +[28] Yichen Li, Kaichun Mo, Lin Shao, Minhyuk Sung, and Leonidas Guibas. Learning 3d part assembly from a single image. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 664-682. Springer, 2020. 2 +[29] Yulong Li, Andy Zeng, and Shuran Song. Rearrangement planning for general part assembly. In 7th Annual Conference on Robot Learning, 2023. 2 +[30] Yuval Litvak, Armin Biess, and Aharon Bar-Hillel. Learning pose estimation for high-precision robotic assembly using simulated depth images. In 2019 International Conference on Robotics and Automation (ICRA), pages 3521–3527. IEEE, 2019. 2 +[31] Hao-Yu Liu, Jian-Wei Guo, Hai-Yong Jiang, Yan-Chao Liu, Xiao-Peng Zhang, and Dong-Ming Yan. Puzzlenet: boundary-aware feature matching for non-overlapping 3d point clouds assembly. Journal of Computer Science and Technology, 38 (3):492-509, 2023. 2 +[32] Jiaxin Lu, Yifan Sun, and Qixing Huang. Jigsaw: Learning to assemble multiple fractured objects. Advances in Neural Information Processing Systems, 36, 2024. 2, 6, 7 +[33] Jianlan Luo, Eugen Solowjow, Chengtao Wen, Juan Aparicio Ojea, Alice M Agogino, Aviv Tamar, and Pieter Abbeel. Reinforcement learning on variable impedance controller for high-precision robotic assembly. In 2019 International Conference on Robotics and Automation (ICRA), pages 3080–3087. IEEE, 2019. 2 +[34] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909-918, 2019. 3 +[35] Yashraj Narang, Kier Storey, Iretiayo Akinola, Miles Macklin, Philipp Reist, Lukasz Wawrzyniak, Yunrong Guo, Adam Moravanszky, Gavriel State, Michelle Lu, et al. Factory: Fast contact for robotic assembly. arXiv preprint arXiv:2205.03532, 2022. 2, 3 +[36] Abhinav Narayan, Rajendra Nagar, and Shanmuganathan Ram. Rgl-net: A recurrent graph learning framework for progressive part assembly. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 78-87, 2022. 2 +[37] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 2 +[38] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 7, 8 +[39] Ismael Rodriguez, Korbinian Nottensteiner, Daniel Leidner, Michael KaBecker, Freek Stulp, and Alin Albu-Schaffer. Iteratively refined feasibility checks in robotic assembly sequence + +planning. IEEE Robotics and Automation Letters, 4(2):1416-1423, 2019. 3 +[40] Gianluca Scarpellini, Stefano Fiorini, Francesco Giuliani, Pietro Moreiro, and Alessio Del Bue. Diffassemble: A unified graph-diffusion model for 2d and 3d reassembly. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28098-28108, 2024. 2 +[41] Silvia Sellán, Yun-Chun Chen, Ziyi Wu, Animesh Garg, and Alec Jacobson. Breaking bad: A dataset for geometric fracture and reassembly. Advances in Neural Information Processing Systems, 35:38885-38898, 2022. 2, 3, 6 +[42] Hijung Shin, Christos Doumas, Thomas Funkhouser, Szymon Rusinkiewicz, Kenneth Steiglitz, Andreas Vlachopoulos, and Tim Weyrich. Analyzing and simulating fracture patterns of theran wall paintings. Journal on Computing and Cultural Heritage (JOCCH), 5(3):1-14, 2012. 2 +[43] Chenrui Tie, Yue Chen, Ruihai Wu, Boxuan Dong, Zeyi Li, Chongkai Gao, and Hao Dong. ET-SEED: EFFICIENT TRAJECTORY-LEVEL SE(3) EQUIVARIANT DIFFUSION POLICY. In The Thirteenth International Conference on Learning Representations, 2025. 3 +[44] Joshua Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. CoRR, abs/1703.06907, 2017. 2 +[45] Jonathan Tremblay, Thang To, Balakumar Sundaralingam, Yu Xiang, Dieter Fox, and Stan Birchfield. Deep object pose estimation for semantic robotic grasping of household objects. ArXiv, abs/1809.10790, 2018. 2 +[46] Weiwei Wan, Kensuke Harada, and Kazuyuki Nagata. Assembly sequence planning for motion planning. CoRR, abs/1609.03108, 2016. 3 +[47] Ruocheng Wang, Yunzhi Zhang, Jiayuan Mao, Ran Zhang, Chin-Yi Cheng, and Jiajun Wu. Iea-manual: Seeing shape assembly step by step. Advances in Neural Information Processing Systems, 35:28428-28440, 2022. 2 +[48] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 7, 8 +[49] Ziming Wang and Rebecka Jornsten. Se (3)-bi-equivariant transformers for point cloud assembly. arXiv preprint arXiv:2407.09167, 2024. 2 +[50] Zhengqing Wang, Jiacheng Chen, and Yasutaka Furukawa. Puzzlefusion++: Auto-agglomerative 3d fracture assembly by denoise and verify. arXiv preprint arXiv:2406.00259, 2024. 2, 6, 7 +[51] Karl DD Willis, Pradeep Kumar Jayaraman, Hang Chu, Yunsheng Tian, Yifei Li, Daniele Grandi, Aditya Sanghi, Linh Tran, Joseph G Lambourne, Armando Solar-Lezama, et al. Joinable: Learning bottom-up assembly of parametric cad joints. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15849-15860, 2022. 2, 3 +[52] Rundi Wu, Yixin Zhuang, Kai Xu, Hao Zhang, and Baoquan Chen. Pq-net: A generative part seq2seq network for 3d shapes. In Proceedings of the IEEE/CVF Conference on + +Computer Vision and Pattern Recognition, pages 829-838, 2020. 2 +[53] Ruihai Wu, Chenrui Tie, Yushi Du, Yan Zhao, and Hao Dong. Leveraging se (3) equivariance for learning 3d geometric shape assembly. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14311-14320, 2023. 6, 7 +[54] Fanbo Xiang, Yuzhe Qin, Kaichun Mo, Yikuan Xia, Hao Zhu, Fangchen Liu, Minghua Liu, Hanxiao Jiang, Yifu Yuan, He Wang, Li Yi, Angel X. Chang, Leonidas J. Guibas, and Hao Su. SAPIEN: A simulated part-based interactive environment. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3 +[55] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. 2018. 2 +[56] Yuwen Xiong, Wei-Chiu Ma, Jingkang Wang, and Raquel Urtasun. Learning compact representations for lidar completion and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1074-1083, 2023. 2 +[57] Boshen Xu, Sipeng Zheng, and Qin Jin. Spaformer: Sequential 3d part assembly with transformers. arXiv preprint arXiv:2403.05874, 2024. 2 +[58] Xianghao Xu, Paul Guerrero, Matthew Fisher, Siddhartha Chaudhuri, and Daniel Ritchie. Unsupervised 3d shape reconstruction by part retrieval and assembly. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8559-8567, 2023. 2 +[59] Zhengrong Xue, Zhecheng Yuan, Jiashun Wang, Xueqian Wang, Yang Gao, and Huazhe Xu. Usek: Unsupervised se (3)-equivariant 3d keypoints for generalizable manipulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1715-1722. IEEE, 2023. 3 +[60] Kangxue Yin, Zhiqin Chen, Siddhartha Chaudhuri, Matthew Fisher, Vladimir G Kim, and Hao Zhang. Coalesce: Component assembly by learning to synthesize connections. In 2020 International Conference on 3D Vision (3DV), pages 61-70. IEEE, 2020. 2 +[61] Kevin Zakka, Andy Zeng, Johnny Lee, and Shuran Song. Form2fit: Learning shape priors for generalizable assembly from disassembly. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9404–9410. IEEE, 2020. 2 +[62] Guanqi Zhan, Qingnan Fan, Kaichun Mo, Lin Shao, Baoquan Chen, Leonidas J Guibas, Hao Dong, et al. Generative 3d part assembly via dynamic graph learning. Advances in Neural Information Processing Systems, 33:6315-6326, 2020. 2 +[63] Rufeng Zhang, Tao Kong, Weihao Wang, Xuan Han, and Mingyu You. 3d part assembly generation with instance encoded transformer. IEEE Robotics and Automation Letters, 7(4):9051-9058, 2022. 2 +[64] Ruiyuan Zhang, Jiaxiang Liu, Zexi Li, Hao Dong, Jie Fu, and Chao Wu. Scalable geometric fracture assembly via co-creation space among assemblers. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7269–7277, 2024. 2 + +[65] Haibo Zhao, Dian Wang, Yizhe Zhu, Xupeng Zhu, Owen Howell, Linfeng Zhao, Yaoyao Qian, Robin Walters, and Robert Platt. Hierarchical equivariant policy via frame transf. arXiv preprint arXiv:2502.05728, 2025. 3 +[66] Junzhe Zhu, Yuanchen Ju, Junyi Zhang, Muhan Wang, Zhecheng Yuan, Kaizhe Hu, and Huazhe Xu. Densehunter: Learning 3d semantic correspondence for category-level manipulation from a single demo. International Conference on Learning Representations (ICLR) 2025, 2024. 2 +[67] Xupeng Zhu, Dian Wang, Guanang Su, Ondrej Biza, Robin Walters, and Robert Platt. On robot grasp learning using equivariant models. Autonomous Robots, 47(8):1175-1193, 2023. 3 +[68] Yuke Zhu, Josiah Wong, Ajay Mandlekar, Roberto Martin-Martin, Abhishek Joshi, Soroush Nasiriany, and Yifeng Zhu. robosuite: A modular simulation framework and benchmark for robot learning. arXiv preprint arXiv:2009.12293, 2020. 2 \ No newline at end of file diff --git a/2025/Two by Two_ Learning Multi-Task Pairwise Objects Assembly for Generalizable Robot Manipulation/images.zip b/2025/Two by Two_ Learning Multi-Task Pairwise Objects Assembly for Generalizable Robot Manipulation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9112dc0356ca1242e86128517da1ca07c6b1394e --- /dev/null +++ b/2025/Two by Two_ Learning Multi-Task Pairwise Objects Assembly for Generalizable Robot Manipulation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2b0aa66dad90efa8ad1ad1f0ef8f293bec8d96307a70ee19ba2550f4aca3da81 +size 723676 diff --git a/2025/Two by Two_ Learning Multi-Task Pairwise Objects Assembly for Generalizable Robot Manipulation/layout.json b/2025/Two by Two_ Learning Multi-Task Pairwise Objects Assembly for Generalizable Robot Manipulation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ae610684bd14382a6f1392494de0ffadedf17ad0 --- /dev/null +++ b/2025/Two by Two_ Learning Multi-Task Pairwise Objects Assembly for Generalizable Robot Manipulation/layout.json @@ -0,0 +1,9278 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 97, + 498, + 141 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 97, + 498, + 141 + ], + "spans": [ + { + "bbox": [ + 96, + 97, + 498, + 141 + ], + "type": "text", + "content": "Two by Two: Learning Multi-Task Pairwise Objects Assembly for Generalizable Robot Manipulation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 163, + 558, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 163, + 558, + 206 + ], + "spans": [ + { + "bbox": [ + 53, + 163, + 558, + 206 + ], + "type": "text", + "content": "Yu Qi " + }, + { + "bbox": [ + 53, + 163, + 558, + 206 + ], + "type": "inline_equation", + "content": "^{2,1*}" + }, + { + "bbox": [ + 53, + 163, + 558, + 206 + ], + "type": "text", + "content": " Yuanchen Ju" + }, + { + "bbox": [ + 53, + 163, + 558, + 206 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 53, + 163, + 558, + 206 + ], + "type": "text", + "content": " Tianming Wei" + }, + { + "bbox": [ + 53, + 163, + 558, + 206 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 53, + 163, + 558, + 206 + ], + "type": "text", + "content": " Chi Chu" + }, + { + "bbox": [ + 53, + 163, + 558, + 206 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 163, + 558, + 206 + ], + "type": "text", + "content": " Lawson L.S. Wong" + }, + { + "bbox": [ + 53, + 163, + 558, + 206 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 163, + 558, + 206 + ], + "type": "text", + "content": " Huazhe Xu" + }, + { + "bbox": [ + 53, + 163, + 558, + 206 + ], + "type": "inline_equation", + "content": "^{1,3,5\\dagger}" + }, + { + "bbox": [ + 53, + 163, + 558, + 206 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 163, + 558, + 206 + ], + "type": "text", + "content": " Shanghai Qi Zhi Institute " + }, + { + "bbox": [ + 53, + 163, + 558, + 206 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 163, + 558, + 206 + ], + "type": "text", + "content": " Northeastern University " + }, + { + "bbox": [ + 53, + 163, + 558, + 206 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 163, + 558, + 206 + ], + "type": "text", + "content": " IIIS, Tsinghua University \n" + }, + { + "bbox": [ + 53, + 163, + 558, + 206 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 53, + 163, + 558, + 206 + ], + "type": "text", + "content": " Shanghai Jiao Tong University " + }, + { + "bbox": [ + 53, + 163, + 558, + 206 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 53, + 163, + 558, + 206 + ], + "type": "text", + "content": " Shanghai AI Laboratory" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 48, + 214, + 545, + 427 + ], + "blocks": [ + { + "bbox": [ + 48, + 214, + 545, + 427 + ], + "lines": [ + { + "bbox": [ + 48, + 214, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 48, + 214, + 545, + 427 + ], + "type": "image", + "image_path": "a5cbc2fb8e59f02077f4a25321f169f156b49bd3b50d7cf4ec0f05d3794fa9a5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 436, + 546, + 460 + ], + "lines": [ + { + "bbox": [ + 46, + 436, + 546, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 546, + 460 + ], + "type": "text", + "content": "Figure 1. Overview of the 2BY2 Dataset. We propose the first large-scale daily pairwise object assembly dataset 2BY2, which contains 1,034 instances and 517 pairwise objects with pose and symmetry annotations." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 475, + 192, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 475, + 192, + 487 + ], + "spans": [ + { + "bbox": [ + 143, + 475, + 192, + 487 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 498, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 290, + 715 + ], + "type": "text", + "content": "3D assembly tasks, such as furniture assembly and component fitting, play a crucial role in daily life and represent essential capabilities for future home robots. Existing benchmarks and datasets predominantly focus on assembling geometric fragments or factory parts, which fall short in addressing the complexities of everyday object interactions and assemblies. To bridge this gap, we present 2BY2, a large-scale annotated dataset for daily pairwise objects assembly, covering 18 fine-grained tasks that reflect real-life scenarios, such as plugging into sockets, arranging flowers in vases, and inserting bread into toasters. 2BY2 dataset includes 1,034 instances and 517 pairwise objects with pose and symmetry annotations, requiring approaches that align geometric shapes while accounting for functional and spatial relationships between objects. Leveraging the 2BY2 dataset, we propose a two-step SE(3) pose estimation method with equivariant features for assembly constraints. Compared to previous shape assembly methods, our approach achieves state-of-the-art" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 475, + 548, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 548, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 548, + 536 + ], + "type": "text", + "content": "performance across all 18 tasks in the 2BY2 dataset. Additionally, robot experiments further validate the reliability and generalization ability of our method for complex 3D assembly tasks. More details and demonstrations can be found at https://tea-lab.github.io/TwoByTwo/." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 551, + 387, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 387, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 387, + 563 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 571, + 547, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 571, + 547, + 690 + ], + "spans": [ + { + "bbox": [ + 304, + 571, + 547, + 690 + ], + "type": "text", + "content": "Assembly tasks are ubiquitous, such as assembling furniture, repairing household appliances, or putting together electronics. Successfully completing these tasks requires precise reasoning about the spatial relationships between pairs of objects. For robots to assist in these activities, they need to accurately estimate the 6D pose of each objects—including both their orientation and position in space. This capability is essential for domestic robots to help humans with various tasks, as it enables them to interact with their environment in a meaningful way." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 691, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 691, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 691, + 547, + 715 + ], + "type": "text", + "content": "Daily object pairwise assembly not only requires considering the geometric constraints and spatial relationships" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17383" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "between objects to achieve precise alignment but also needs to exhibit a certain level of generalization. Existing methods and benchmarks for solving assembly problems [31, 32, 49, 50, 63], typically focus on matching local geometric shapes, which often results in suboptimal performance in everyday assembly scenarios which require semantic and spatial alignment. This is because they are primarily trained and tested on existing assembly datasets that consist of large-scale geometric fragments, such as Breaking Bad [41] and Neural Shape Mating [5]. Compared with existing assembly tasks that focus on putting together fractures of objects, daily pairwise assembly tasks are more challenging and hold greater practical significance in human life." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 289, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 289, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 289, + 337 + ], + "type": "text", + "content": "To bridge this gap, we introduce 2BY2, the first large-scale daily pairwise assembly dataset comprising 18 fine-grained tasks, shown as Figure 1. Compared to previous datasets and benchmarks, 2BY2 contains 1,034 instances and 517 pairwise objects with pose and symmetry annotations, covering a variety of pairwise assembly tasks reflecting everyday scenarios, which require approaches that align geometric shapes while accounting for functional and spatial relationships between objects, as shown in Table 1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 338, + 289, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 289, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 289, + 506 + ], + "type": "text", + "content": "Building on this dataset, we propose a two-step pairwise network architecture for assembly tasks. Mimicking the human assembly process like we firstly put the vase on the table and then arrange flower in it, our approach predicts the pose of each object in a step-by-step manner to assemble them to a predefined canonical space, which refers to a standard coordinate system that aligns with the principles of the human world, with detailed definition in Section 3.2. The network leverages a custom two-scale Vector Neuron DGCNN [9] encoder with spherical convolution [7] to extract SE(3) equivariant and SO(3) invariant features from point cloud inputs. Additionally, a feature fusion module and a two-step training and evaluation strategy are used to improve pose prediction accuracy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 506, + 289, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 289, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 289, + 628 + ], + "type": "text", + "content": "We evaluate our approach on 18 tasks in 2BY2 dataset to demonstrate the effectiveness on multi-task object pairwise assembly prediction. Compared to existing baselines, our method achieves an average improvement of 0.046 in translation RMSE and 8.97 in rotation RMSE. Moreover, we validate the effectiveness of our approach on three multicategory task, namely Lid Covering, Inserting and High Precision Placing, as well as the All task, which is defined in Section 2. Besides, real-world robot experiments validate the practical applicability of our approach." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 628, + 240, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 628, + 240, + 639 + ], + "spans": [ + { + "bbox": [ + 59, + 628, + 240, + 639 + ], + "type": "text", + "content": "Our main contributions are listed as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "1. We introduce 2BY2, the first large-scale daily pairwise object assembly dataset. By providing comprehensive pose and symmetry annotations for 517 pairwise objects across 18 fine-grained tasks, 2BY2 pushes the boundaries of real-world 3D assembly challenges and establishes a benchmark for pairwise assembly tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 181 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": "2. Our two-step pairwise SE(3) pose estimation method, leveraging equivariant geometric features, demonstrates superior performance compared to existing shape assembly methods, significantly reducing translation and rotation errors and enhancing the accuracy of 6D pose estimation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 133, + 547, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 547, + 181 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 547, + 181 + ], + "type": "text", + "content": "3. Our approach achieves state-of-the-art performance on the benchmark, with real-world robot experiments demonstrating its capability, providing a generalizable solution for robot manipulation using pairwise object assembly." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 305, + 191, + 392, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 191, + 392, + 204 + ], + "spans": [ + { + "bbox": [ + 305, + 191, + 392, + 204 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 211, + 533, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 211, + 533, + 224 + ], + "spans": [ + { + "bbox": [ + 305, + 211, + 533, + 224 + ], + "type": "text", + "content": "2.1. Object Assembly Benchmarks and Datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 229, + 547, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 229, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 229, + 547, + 397 + ], + "type": "text", + "content": "Object reassembly has led to various datasets in computer vision and robotics. In computer vision, datasets like AutoMate [24] and JoinABLe [51] focus on reassembling fragments using geometric clues, while early datasets [3, 12, 21, 42] were limited in scale. Recent efforts, such as Neural Shape Mating [5] and Breaking Bad [41], generate largescale fractured object data using parametric segmentation. In robotics, benchmarks like Factory [35], RLBench [23], and RoboSuite [68] lack diverse shapes and assembly tasks under varying initial poses. In contrast, our dataset includes over 500 diverse object pairs across 3 categories and 18 assembly tasks, providing a comprehensive benchmark for pairwise object assembly, supporting the development of generalizable methods for real-world applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 404, + 421, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 404, + 421, + 418 + ], + "spans": [ + { + "bbox": [ + 305, + 404, + 421, + 418 + ], + "type": "text", + "content": "2.2.3D Shape Assembly" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 423, + 547, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 547, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 547, + 591 + ], + "type": "text", + "content": "3D shape assembly [11, 30, 33, 56, 61], also known as part assembly, involves reconstructing objects from fragments, such as shattered sculptures or disassembled furniture. Existing methods use graphical models [4, 22, 26] and neural networks [6, 25, 28, 47, 52, 57, 58, 60, 64, 66] to capture geometric and semantic relationships. Approaches such as [5, 36, 62, 63] focus on pose estimation and part assembly without relying on predefined semantic information. Few-shot learning has been applied to assembly tasks [29], while jigsaw puzzle techniques [32, 37] leverage shape completion strategies. Recent works [16, 20, 40, 50] utilize diffusion models to refine poses or point clouds for assembly. In contrast, our method introduces a two-step pairwise network for step-by-step assembly, tailored to pairwise object alignment." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 599, + 534, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 599, + 534, + 612 + ], + "spans": [ + { + "bbox": [ + 305, + 599, + 534, + 612 + ], + "type": "text", + "content": "2.3.6D Pose Estimation for Robot Manipulation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "6D pose estimation is crucial in robotics and computer vision for object interaction in unstructured environments [15, 45, 55]. Early handcrafted feature-based methods struggled in cluttered scenes [14, 15], while CNN-based approaches improved performance but lacked generalization [27, 55]. Domain randomization enhances robustness by varying synthetic datasets [44, 45]. In assembly tasks, 6D pose estimation aids manipulation planning with predefined" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17384" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 547, + 162 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 162 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 162 + ], + "type": "table", + "html": "
Dataset#OC#OSTask NumberPairTask HierarchyEveryday ScenarioSymmetryAssemble Type
PartNet [34]2426,671-NoNoNoNoSemantic
AutoMate [24]292,5291NoNoNoNoGeometric
JoinABLe [51]68,2511NoNoNoNoSemantic
NSM dataset[5]111,2461YesNoNoNoGeometric
Breaking Bad[41]-10,4741NoNoNoNoGeometric
Factory[35]8608YesNoNoNoGeometric
2BY2 Dataset (Ours)36103418YesYesYesYesGeometric and Semantic
", + "image_path": "cbaab6546e4e6c458c1f489d1f5394cc71ecaa7d0b733b9bba15817f0a640d24.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 170, + 547, + 226 + ], + "lines": [ + { + "bbox": [ + 46, + 170, + 547, + 226 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 547, + 226 + ], + "type": "text", + "content": "Table 1. Dataset Comparison. We compare 2BY2 dataset with existing datasets and benchmarks. #OC stands for the number of object categories. #OS stands for the number of object shapes. Pair denotes whether the dataset is pairwise. Task Number refers to the number of distinct assembly tasks, with the assembly of fractured pieces considered as a single task. Task Hierarchy stands for the different categories of task from coarse to fine, with ours shown in Section 3.1. Everyday Scenario means whether the assemble task has practical significance in real-world human applications. Symmetry denotes whether the dataset contains part symmetry annotation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 246, + 288, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 246, + 288, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 288, + 283 + ], + "type": "text", + "content": "objects [39, 46]. Like [13, 17-20, 43, 59, 65, 67], our method leverages equivariant features for efficient 6D pose learning and improved generalization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 292, + 133, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 292, + 133, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 292, + 133, + 304 + ], + "type": "text", + "content": "3.2BY2 Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 311, + 181, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 311, + 181, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 311, + 181, + 323 + ], + "type": "text", + "content": "3.1.2BY2 Dataset Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 329, + 288, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 288, + 449 + ], + "type": "text", + "content": "We present the first large-scale 3D pairwise object assembly dataset for everyday scenarios, with detailed annotations for each object pair. The meshes in our dataset come from 3D Warehouse [1], SAPIEN PartNet-Mobility [54], Google SketchUp 3D Challenge [2], and Objaverse [8]. These meshes are manually paired, cleaned, annotated, and scaled uniformly. The 2BY2 dataset contains 517 unique pairs across three main tasks: Lid Covering, Inserting, and High Precision Placing, further subdivided into multiple subcategories, as shown in Table 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 456, + 148, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 456, + 148, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 148, + 467 + ], + "type": "text", + "content": "3.2. Data Annotation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 474, + 288, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 288, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 288, + 653 + ], + "type": "text", + "content": "To ensure high quality and reliability of our dataset, we conducted systematic cleaning and annotation of the collected meshes. First, we manually segment, integrate, and pair the meshes, classify them into Object " + }, + { + "bbox": [ + 46, + 474, + 288, + 653 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 474, + 288, + 653 + ], + "type": "text", + "content": " and Object " + }, + { + "bbox": [ + 46, + 474, + 288, + 653 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 474, + 288, + 653 + ], + "type": "text", + "content": ". Object " + }, + { + "bbox": [ + 46, + 474, + 288, + 653 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 474, + 288, + 653 + ], + "type": "text", + "content": " is the base or the receiving component, such as the nut, the vase, the postbox. Object " + }, + { + "bbox": [ + 46, + 474, + 288, + 653 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 474, + 288, + 653 + ], + "type": "text", + "content": ", is the fitting component, such as the bolt, the flower, the mail. This classification aligns with intuitive human assembly logic and supports our network's prediction strategy, such as positioning a nut before the bolt, as detailed in Sections 4 and 5.1. Automated scripts were used to uniformly scale meshes and align each pair to a canonical pose in world frame, defined as the object resting stably on the " + }, + { + "bbox": [ + 46, + 474, + 288, + 653 + ], + "type": "inline_equation", + "content": "XY" + }, + { + "bbox": [ + 46, + 474, + 288, + 653 + ], + "type": "text", + "content": " plane with its lowest point aligned to " + }, + { + "bbox": [ + 46, + 474, + 288, + 653 + ], + "type": "inline_equation", + "content": "Z = 0" + }, + { + "bbox": [ + 46, + 474, + 288, + 653 + ], + "type": "text", + "content": ". For instance, bottles and vases are aligned as if placed on a table, and mailboxes on the ground." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 654, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 289, + 714 + ], + "type": "text", + "content": "During point cloud generation, we utilized blue noise sampling method [41] to extract point clouds uniformly from each mesh surface with dimension (1024, 3). We also annotated each object category with its inherent symmetry properties, specifically considering rotational symmetry along the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 246, + 545, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 246, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 545, + 270 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 304, + 246, + 545, + 270 + ], + "type": "text", + "content": "-axis, such as bottles, screws, and mirror symmetry along the " + }, + { + "bbox": [ + 304, + 246, + 545, + 270 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 246, + 545, + 270 + ], + "type": "text", + "content": "-axis, such as bread, letters." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 276, + 525, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 276, + 525, + 289 + ], + "spans": [ + { + "bbox": [ + 305, + 276, + 525, + 289 + ], + "type": "text", + "content": "3.3. Data Division and Task Diversity Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 295, + 547, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 295, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 295, + 547, + 415 + ], + "type": "text", + "content": "Our dataset provides diverse task coverage across categories, with each further divided into specific sub-categories, see Table 2. Objects within each category vary in shape, size, and type. To enhance generalization, the testing set includes objects with unseen geometric shapes, as shown in Figure 3. We also compute Chamfer Distance on point clouds between training and testing sets to quantify geometry differences, as shown in Figure 2. This diversity ensures generalization ability and applicability in real world scenarios and supports robust 3D matching and assembly tasks." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 309, + 424, + 545, + 550 + ], + "blocks": [ + { + "bbox": [ + 309, + 424, + 545, + 550 + ], + "lines": [ + { + "bbox": [ + 309, + 424, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 309, + 424, + 545, + 550 + ], + "type": "image", + "image_path": "fa84678daecd4f114ab87ad5c863115c41885046160521308f804e83832416b9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 558, + 547, + 603 + ], + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 603 + ], + "type": "text", + "content": "Figure 2. Chamfer Distance Between Training and Testing Set. We normalize point clouds and compute the Chamfer Distance. For each task we calculate the distance separately between point cloud of Object " + }, + { + "bbox": [ + 304, + 558, + 547, + 603 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 558, + 547, + 603 + ], + "type": "text", + "content": " and Object " + }, + { + "bbox": [ + 304, + 558, + 547, + 603 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 558, + 547, + 603 + ], + "type": "text", + "content": " in the training set and test set." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 609, + 432, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 609, + 432, + 621 + ], + "spans": [ + { + "bbox": [ + 305, + 609, + 432, + 621 + ], + "type": "text", + "content": "4. Problem Formulation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "The task takes two point clouds as input, namely " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_A" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_B" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": ", each with dimension (1024, 3). These point clouds are derived from objects " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_A" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_B" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " from predefined canonical pose, as detailed in 3.2, respectively, and is randomly augmented with SO(3) rotation and being translated to its centroid. The desired output would be two individual SE(3) pose two assemble " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_A" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_B" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " to the canonical pose." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17385" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 545, + 122 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 122 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 122 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 122 + ], + "type": "table", + "html": "
TaskLid CoveringInsertingHigh Precision Placing
KitBotKetCofCupPluChiLetBreNutCoiKeyUsbBoxTisFloTeaPos
Pair Num248628261614193224202120202520604221
", + "image_path": "f0800a33a2e4796a3a8670c3dfe9d6285a9a4319f0afbdf2a9ded4d16a6f7add.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 130, + 547, + 198 + ], + "lines": [ + { + "bbox": [ + 46, + 130, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 130, + 547, + 198 + ], + "type": "text", + "content": "Table 2. 2BY2 Dataset Statistics Overview. The figure presents the number of object pairs across all task categories in the 2BY2 dataset, where each pair consists of two unique objects. The first row categorizes tasks into three major groups, while the second row provides a detailed breakdown of specific task categories. Specifically, Kit = Kitchenport, Bot = Bottle, Ket = Kettle, Cof = Coffee machine, Cup = Cup, Plu = Plug into socket, Chi = Children's toy, Let = Letter into mailbox, Bre = Bread into toaster, Nut = Bolt into nut, Coi = Coin into piggy bank, Key = Key into lock, Usb = USB cap, Box = Shoe boxing, Tis = Tissue placement on rack, Flo = Flower into vase, Tea = Teaware arrangement on tray, Pos = Positioning a cup on the coffee machine for coffee dispensing." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 50, + 216, + 167, + 368 + ], + "blocks": [ + { + "bbox": [ + 50, + 216, + 167, + 368 + ], + "lines": [ + { + "bbox": [ + 50, + 216, + 167, + 368 + ], + "spans": [ + { + "bbox": [ + 50, + 216, + 167, + 368 + ], + "type": "image", + "image_path": "0d3ba6708fc5f95c010232ecbee4aa425fc8af00308e3d5b7b04880c20e67bd3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 376, + 288, + 443 + ], + "lines": [ + { + "bbox": [ + 46, + 376, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 288, + 443 + ], + "type": "text", + "content": "Figure 3. Task Diversity Visualization. The image shows selected objects from four different tasks: USB, Bottle, Letter, and Plug in Socket. On the left are the objects selected on training set, and on the right is the testing set. As seen in the legend, object geometry varies in both the training and testing set, with the testing set containing novel shapes not seen in the training set." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 168, + 216, + 286, + 368 + ], + "blocks": [ + { + "bbox": [ + 168, + 216, + 286, + 368 + ], + "lines": [ + { + "bbox": [ + 168, + 216, + 286, + 368 + ], + "spans": [ + { + "bbox": [ + 168, + 216, + 286, + 368 + ], + "type": "image", + "image_path": "aa6cd26a3d7f7e604524c8ec7094bb194075a6c98b0b8179ba5f33a465875c90.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 466, + 104, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 466, + 104, + 478 + ], + "spans": [ + { + "bbox": [ + 47, + 466, + 104, + 478 + ], + "type": "text", + "content": "5. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 487, + 261, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 487, + 261, + 500 + ], + "spans": [ + { + "bbox": [ + 47, + 487, + 261, + 500 + ], + "type": "text", + "content": "5.1. Two-step Pairwise Network Architecture" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 508, + 289, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 289, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 289, + 687 + ], + "type": "text", + "content": "To effectively learn pairwise object assembly, we propose a two-step pairwise network architecture with two branches: Branch B " + }, + { + "bbox": [ + 46, + 508, + 289, + 687 + ], + "type": "inline_equation", + "content": "(\\mathcal{B}_{\\mathcal{B}})" + }, + { + "bbox": [ + 46, + 508, + 289, + 687 + ], + "type": "text", + "content": " and Branch A " + }, + { + "bbox": [ + 46, + 508, + 289, + 687 + ], + "type": "inline_equation", + "content": "(\\mathcal{B}_{\\mathcal{A}})" + }, + { + "bbox": [ + 46, + 508, + 289, + 687 + ], + "type": "text", + "content": ", as shown in Figure 4. Branch B predicts the pose of " + }, + { + "bbox": [ + 46, + 508, + 289, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{B}}" + }, + { + "bbox": [ + 46, + 508, + 289, + 687 + ], + "type": "text", + "content": ", which is the socket, using a two-scale Vector Neuron DGCNN encoder [9] to extract SE(3) equivariant features, denoted as " + }, + { + "bbox": [ + 46, + 508, + 289, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{B}}" + }, + { + "bbox": [ + 46, + 508, + 289, + 687 + ], + "type": "text", + "content": ", followed by MLP-based pose prediction heads for translation and rotation. The transformed " + }, + { + "bbox": [ + 46, + 508, + 289, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{B}}" + }, + { + "bbox": [ + 46, + 508, + 289, + 687 + ], + "type": "text", + "content": " and inserter object " + }, + { + "bbox": [ + 46, + 508, + 289, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{A}}" + }, + { + "bbox": [ + 46, + 508, + 289, + 687 + ], + "type": "text", + "content": ", which is the plug, are then passed to Branch A, which extracts SE(3) equivariant features, denoted as " + }, + { + "bbox": [ + 46, + 508, + 289, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{A}}" + }, + { + "bbox": [ + 46, + 508, + 289, + 687 + ], + "type": "text", + "content": " and SO(3) invariant features " + }, + { + "bbox": [ + 46, + 508, + 289, + 687 + ], + "type": "inline_equation", + "content": "(\\mathcal{I}_{\\mathcal{B}})" + }, + { + "bbox": [ + 46, + 508, + 289, + 687 + ], + "type": "text", + "content": ". The features are fused through element-wise multiplication, allowing " + }, + { + "bbox": [ + 46, + 508, + 289, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathcal{A}}" + }, + { + "bbox": [ + 46, + 508, + 289, + 687 + ], + "type": "text", + "content": " to predict the pose of " + }, + { + "bbox": [ + 46, + 508, + 289, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{A}}" + }, + { + "bbox": [ + 46, + 508, + 289, + 687 + ], + "type": "text", + "content": " using information from both objects. This architecture ensures geometric alignment and matching by leveraging shared feature representations while reducing feature interference." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 289, + 714 + ], + "type": "text", + "content": "Our two-step pairwise network is inspired by the human approach to pairwise assembly tasks. For example, when ar-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 217, + 547, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 547, + 289 + ], + "type": "text", + "content": "ranging a vase with flowers, one intuitively first positions the vase correctly before placing the flowers inside. Similarly, inserting an envelope into a mailbox requires identifying the mailbox slot's pose first. By mimicking this sequential strategy, our model simulates human decision-making process, enabling more efficient and accurate assembly tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 298, + 545, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 298, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 305, + 298, + 545, + 323 + ], + "type": "text", + "content": "5.2. Two-scale SE(3) Equivariant and SO(3) Invariant Feature Extraction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 329, + 547, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 547, + 412 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 547, + 412 + ], + "type": "text", + "content": "We employ a two-scale SE(3) Vector Neuron DGCNN, an enhanced variant of the original Vector Neuron DGCNN [9], as our encoder to extract SE(3) equivariant and SO(3) invariant features. This architecture leverages equivariance to improve sample efficiency of the model, while incorporating a two-scale information fusion mechanism to capture geometric features at two different scales." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 413, + 547, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 547, + 533 + ], + "type": "text", + "content": "SE(3) Equivariance and SO(3) Invariance. SE(3) equivariance combines SO(3) rotation and T(3) translation equivariance: rotation equivariance ensures that a network's output rotates with the input, while translation equivariance shifts the output accordingly. SO(3) invariance means the network's output remains unchanged under any 3D rotation. By leveraging SE(3) equivariance, the model benefits from improved sample efficiency and generalization. This is particularly advantageous for assembly tasks, where objects may appear in arbitrary poses." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 533, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 628 + ], + "type": "text", + "content": "Vector Neuron DGCNN. The Vector Neuron Network [9] extends traditional neurons from scalars to 3D vectors, designs vector-based convolutional layers and nonlinear functions like pooling and ReLU to support SO(3) equivariant and SO(3) invariant feature extraction. VNN operates in vector space, captures richer geometric relationships and ensures more robust feature representations for downstream tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 630, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 548, + 714 + ], + "type": "text", + "content": "Two-scale Vector Neuron DGCNN. We propose a two-scale Vector Neuron DGCNN for extracting SE(3) equivariant and SO(3) invariant features " + }, + { + "bbox": [ + 304, + 630, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{B}}" + }, + { + "bbox": [ + 304, + 630, + 548, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 630, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathcal{B}}" + }, + { + "bbox": [ + 304, + 630, + 548, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 630, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{A}}" + }, + { + "bbox": [ + 304, + 630, + 548, + 714 + ], + "type": "text", + "content": ". As shown in Figure 4, the encoder comprises two branches with different " + }, + { + "bbox": [ + 304, + 630, + 548, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 630, + 548, + 714 + ], + "type": "text", + "content": " values, each consisting of multiple Vector Neuron convolutional layers followed by pooling. The extracted features from both branches are concatenated and further pro" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17386" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 315, + 208 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 315, + 208 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 315, + 208 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 315, + 208 + ], + "type": "image", + "image_path": "426fdc6ea95e7999c713f3f9b402dc937e672444a66d44b20ab891b96882fccb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 443, + 319 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 443, + 319 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 443, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 443, + 319 + ], + "type": "image", + "image_path": "ec389a409c21357f0ba92a5925b7ea26c9a8c1ce5489b63d37b43bee8c8b2ea3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 357, + 329, + 420, + 340 + ], + "lines": [ + { + "bbox": [ + 357, + 329, + 420, + 340 + ], + "spans": [ + { + "bbox": [ + 357, + 329, + 420, + 340 + ], + "type": "text", + "content": "Inference Stage" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 444, + 72, + 539, + 132 + ], + "blocks": [ + { + "bbox": [ + 444, + 72, + 539, + 132 + ], + "lines": [ + { + "bbox": [ + 444, + 72, + 539, + 132 + ], + "spans": [ + { + "bbox": [ + 444, + 72, + 539, + 132 + ], + "type": "image", + "image_path": "ee017c079a0e93900d5a1be86390829fd704ea4a7869d6b55c1803bc4b88ddf6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 445, + 135, + 539, + 194 + ], + "blocks": [ + { + "bbox": [ + 445, + 135, + 539, + 194 + ], + "lines": [ + { + "bbox": [ + 445, + 135, + 539, + 194 + ], + "spans": [ + { + "bbox": [ + 445, + 135, + 539, + 194 + ], + "type": "image", + "image_path": "becc995c74a834c658d0cd186207979920259cc0fb5358cd4c2df100e69d8e30.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 445, + 197, + 539, + 256 + ], + "blocks": [ + { + "bbox": [ + 445, + 197, + 539, + 256 + ], + "lines": [ + { + "bbox": [ + 445, + 197, + 539, + 256 + ], + "spans": [ + { + "bbox": [ + 445, + 197, + 539, + 256 + ], + "type": "image", + "image_path": "15b00f593880eaf4dbb0ba131eed6e52d0012e5b76d1cd0cecc6c8de59b64b83.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 445, + 258, + 539, + 318 + ], + "blocks": [ + { + "bbox": [ + 445, + 258, + 539, + 318 + ], + "lines": [ + { + "bbox": [ + 445, + 258, + 539, + 318 + ], + "spans": [ + { + "bbox": [ + 445, + 258, + 539, + 318 + ], + "type": "image", + "image_path": "288ef862b3bdd968b1a0c42c80bfdfd0b3a1d4a8c5910493018c8f8e55f46ec0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 329, + 531, + 339 + ], + "lines": [ + { + "bbox": [ + 453, + 329, + 531, + 339 + ], + "spans": [ + { + "bbox": [ + 453, + 329, + 531, + 339 + ], + "type": "text", + "content": "Robot Experiments" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 350, + 547, + 373 + ], + "lines": [ + { + "bbox": [ + 46, + 350, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 547, + 373 + ], + "type": "text", + "content": "Figure 4. Our Two-Step Pairwise Network. We utilize two-scale VN DGCNN as our encoder to extract equivariant and invariant feature. We first predict the canonical pose of " + }, + { + "bbox": [ + 46, + 350, + 547, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_B" + }, + { + "bbox": [ + 46, + 350, + 547, + 373 + ], + "type": "text", + "content": " and then predict the pose of " + }, + { + "bbox": [ + 46, + 350, + 547, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_A" + }, + { + "bbox": [ + 46, + 350, + 547, + 373 + ], + "type": "text", + "content": " according to it." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 48, + 209, + 315, + 318 + ], + "blocks": [ + { + "bbox": [ + 48, + 209, + 315, + 318 + ], + "lines": [ + { + "bbox": [ + 48, + 209, + 315, + 318 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 315, + 318 + ], + "type": "image", + "image_path": "19b59556fd9937072e7e0aed90b18510e3e2fcba159b3febf5b5fda5dc08897f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 150, + 329, + 209, + 341 + ], + "lines": [ + { + "bbox": [ + 150, + 329, + 209, + 341 + ], + "spans": [ + { + "bbox": [ + 150, + 329, + 209, + 341 + ], + "type": "text", + "content": "Training Stage" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 393, + 288, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 393, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 288, + 430 + ], + "type": "text", + "content": "cessed through an additional Vector Neuron convolutional layer. Point clouds " + }, + { + "bbox": [ + 46, + 393, + 288, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{B}}" + }, + { + "bbox": [ + 46, + 393, + 288, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 393, + 288, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{A}}" + }, + { + "bbox": [ + 46, + 393, + 288, + 430 + ], + "type": "text", + "content": " are independently processed, forming graphs that propagate through both branches." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 430, + 287, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 287, + 524 + ], + "type": "text", + "content": "The SO(3) rotation equivariance of our encoder is ensured by the inherent equivariant properties of the Vector Neuron layers. To achieve T(3) translation equivariance, with an input point cloud " + }, + { + "bbox": [ + 46, + 430, + 287, + 524 + ], + "type": "inline_equation", + "content": "P = (p_{1}, p_{2}, \\dots, p_{n})" + }, + { + "bbox": [ + 46, + 430, + 287, + 524 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 430, + 287, + 524 + ], + "type": "inline_equation", + "content": "p_{i} \\in R^{3}" + }, + { + "bbox": [ + 46, + 430, + 287, + 524 + ], + "type": "text", + "content": ", we compute its centroid " + }, + { + "bbox": [ + 46, + 430, + 287, + 524 + ], + "type": "inline_equation", + "content": "x = (\\Sigma_{i=1}^{n} p_{i}) / n" + }, + { + "bbox": [ + 46, + 430, + 287, + 524 + ], + "type": "text", + "content": ", and get the input point cloud as " + }, + { + "bbox": [ + 46, + 430, + 287, + 524 + ], + "type": "inline_equation", + "content": "P' = P - x" + }, + { + "bbox": [ + 46, + 430, + 287, + 524 + ], + "type": "text", + "content": ". In this way, our prediction is T(3) translation equivariant, i.e., " + }, + { + "bbox": [ + 46, + 430, + 287, + 524 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 430, + 287, + 524 + ], + "type": "text", + "content": " is our encoder and " + }, + { + "bbox": [ + 46, + 430, + 287, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 430, + 287, + 524 + ], + "type": "text", + "content": " is the original point cloud." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 94, + 533, + 287, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 533, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 94, + 533, + 287, + 548 + ], + "type": "interline_equation", + "content": "f (\\mathcal {P} + \\mathcal {T}) = f (\\mathcal {P}) + \\mathcal {T}, \\quad T \\in R ^ {3} \\tag {1}", + "image_path": "9f3da2c3fc054afe7527632e0b3442fbc558ac4ebc7331460189387c587cdd22.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 556, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 628 + ], + "type": "text", + "content": "Our two-scale VN DGCNN employs dual K-nearest neighbor (KNN) values to extract features across two distinct scales, enhancing its ability to capture both local and global information. This pyramid structure enables the network to simultaneously grasp overall object shapes and fine-grained details, improving feature extraction." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 635, + 204, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 204, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 204, + 647 + ], + "type": "text", + "content": "5.3. Cross Object Fusion Module" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": "We utilize point-wise multiplication, shown in Figure 4, as our cross object fusion module designed in " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "B_A" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": ". We fuse the feature of " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_B" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_A" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": " by multiplying " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_B" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_A" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": ", so that each point in " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_A" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": " will have the geometry feature of both " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_A" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_B" + }, + { + "bbox": [ + 46, + 653, + 288, + 714 + ], + "type": "text", + "content": ". This approach integrates the geometric feature of" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 393, + 546, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 393, + 546, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 393, + 546, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_B" + }, + { + "bbox": [ + 304, + 393, + 546, + 430 + ], + "type": "text", + "content": " in each point while preserving the rotation equivariance of " + }, + { + "bbox": [ + 304, + 393, + 546, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_A" + }, + { + "bbox": [ + 304, + 393, + 546, + 430 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 304, + 393, + 546, + 430 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 393, + 546, + 430 + ], + "type": "text", + "content": " is an equivariant neural network, " + }, + { + "bbox": [ + 304, + 393, + 546, + 430 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 393, + 546, + 430 + ], + "type": "text", + "content": " is random rotation matrix," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 318, + 439, + 546, + 454 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 439, + 546, + 454 + ], + "spans": [ + { + "bbox": [ + 318, + 439, + 546, + 454 + ], + "type": "interline_equation", + "content": "f \\left(R \\cdot \\left(\\mathcal {I} _ {\\mathcal {B}} * \\mathcal {E} _ {\\mathcal {A}}\\right)\\right) = R \\cdot f \\left(\\mathcal {I} _ {\\mathcal {B}} * \\mathcal {E} _ {\\mathcal {A}}\\right), \\quad R \\in R ^ {3 \\times 3} \\tag {2}", + "image_path": "fb7d5ed2157f619caea58102969f045ceef2ff4143520d14149c3df95e5cb459.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 464, + 401, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 464, + 401, + 475 + ], + "spans": [ + { + "bbox": [ + 306, + 464, + 401, + 475 + ], + "type": "text", + "content": "5.4. Pose Prediction" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 482, + 546, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 482, + 546, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 546, + 555 + ], + "type": "text", + "content": "At both branches, we utilize two separate MLPs as our pose prediction head, to separately predict the translation " + }, + { + "bbox": [ + 304, + 482, + 546, + 555 + ], + "type": "inline_equation", + "content": "T \\in R^3" + }, + { + "bbox": [ + 304, + 482, + 546, + 555 + ], + "type": "text", + "content": " and rotation " + }, + { + "bbox": [ + 304, + 482, + 546, + 555 + ], + "type": "inline_equation", + "content": "R \\in R^{3 \\times 3}" + }, + { + "bbox": [ + 304, + 482, + 546, + 555 + ], + "type": "text", + "content": ". Compared to predicting translation and rotation within a single prediction head, this approach helps mitigate the issue of differing convergence speeds between the two components." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 563, + 486, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 563, + 486, + 576 + ], + "spans": [ + { + "bbox": [ + 305, + 563, + 486, + 576 + ], + "type": "text", + "content": "5.5. Training and Evaluation Strategy" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": "We adopt a separate training and evaluation strategy for our network. To minimize the impact of pose prediction errors of " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{B}}" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{A}}" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": ", we train " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathcal{A}}" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathcal{B}}" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": " independently. Specifically, for " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathcal{A}}" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": ", during training, we utilize " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{B}}" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": " under canonical pose, which is our ground truth point cloud of " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{B}}" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": ", to train our model. During testing, we first predict the pose of " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{B}}" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": ", then use the transformed " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{B}}" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": ", along with the initial " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{A}}" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": " to predict A's pose, as shown in Figure 4. This phased, two-step training and evaluation strategy reduces errors caused by joint training of object poses, ensuring more accurate predictions." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "17387" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 136, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 136, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 136, + 83 + ], + "type": "text", + "content": "5.6. Loss Function" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 90, + 287, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 90, + 287, + 114 + ], + "spans": [ + { + "bbox": [ + 47, + 90, + 287, + 114 + ], + "type": "text", + "content": "To train our network to robustly predict poses, we use the following equation as our loss function:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 114, + 121, + 287, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 121, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 114, + 121, + 287, + 133 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {\\text {r o t}} \\mathcal {L} _ {\\text {r o t}} + \\lambda_ {\\text {t r a n s}} \\mathcal {L} _ {\\text {t r a n s}} \\tag {3}", + "image_path": "33f9f7e80af40686e165e446d441306ae55e021720612a27ca3cb80f021292bc.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 139, + 287, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 139, + 287, + 187 + ], + "spans": [ + { + "bbox": [ + 47, + 139, + 287, + 187 + ], + "type": "text", + "content": "Specifically, for predicted pose translation " + }, + { + "bbox": [ + 47, + 139, + 287, + 187 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{pred}} \\in R^3" + }, + { + "bbox": [ + 47, + 139, + 287, + 187 + ], + "type": "text", + "content": ", rotation " + }, + { + "bbox": [ + 47, + 139, + 287, + 187 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathrm{pred}} \\in R^{3 \\times 3}" + }, + { + "bbox": [ + 47, + 139, + 287, + 187 + ], + "type": "text", + "content": " and ground truth pose translation " + }, + { + "bbox": [ + 47, + 139, + 287, + 187 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{gt}} \\in R^3" + }, + { + "bbox": [ + 47, + 139, + 287, + 187 + ], + "type": "text", + "content": " and rotation " + }, + { + "bbox": [ + 47, + 139, + 287, + 187 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathrm{gt}} \\in R^{3 \\times 3}" + }, + { + "bbox": [ + 47, + 139, + 287, + 187 + ], + "type": "text", + "content": ", we use " + }, + { + "bbox": [ + 47, + 139, + 287, + 187 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 47, + 139, + 287, + 187 + ], + "type": "text", + "content": " loss to compute our " + }, + { + "bbox": [ + 47, + 139, + 287, + 187 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{trans}}" + }, + { + "bbox": [ + 47, + 139, + 287, + 187 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 194, + 287, + 207 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 194, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 121, + 194, + 287, + 207 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t r a n s}} = \\mathcal {L} _ {1} \\left(T _ {\\text {p r e d}}, T _ {\\text {g t}}\\right) \\tag {4}", + "image_path": "0ca4ad587d83fe9aaae8996ea92924394b075cf73060a48167e43b8d7f20805c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 213, + 287, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 213, + 287, + 273 + ], + "spans": [ + { + "bbox": [ + 47, + 213, + 287, + 273 + ], + "type": "text", + "content": "As for the rotation, we utilize Geodesic Distance, which measures the shortest path between two rotations on the rotation manifold. It offers a smooth and bounded angular error, ensuring stable gradients and accurately achieving precise rotation alignment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 94, + 277, + 287, + 310 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 277, + 287, + 310 + ], + "spans": [ + { + "bbox": [ + 94, + 277, + 287, + 310 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r o t}} = \\arccos \\left(\\frac {\\operatorname {t r} \\left(\\mathcal {R} _ {\\mathrm {g t}} \\mathcal {R} _ {\\text {p r e d}} ^ {T}\\right) - 1}{2}\\right) \\tag {5}", + "image_path": "190370cc63282ab07d2e241798212b052f7c1983ac529d8de245f808d5792453.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 316, + 128, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 128, + 329 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 128, + 329 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 335, + 287, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 287, + 371 + ], + "type": "text", + "content": "In this section, we present a comprehensive evaluation and analysis of our two-step pairwise network architecture by addressing the following key questions:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 371, + 287, + 467 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 47, + 371, + 287, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 371, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 371, + 287, + 418 + ], + "type": "text", + "content": "1. How does our network perform on 2BY2 tasks compared to existing baseline approaches, including matching-based, graph-network-based, and diffusion-based assembly methods?" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 419, + 287, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 419, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 47, + 419, + 287, + 454 + ], + "type": "text", + "content": "2. How well does our network generalize across multiple tasks within the 2BY2 dataset? Can our network effectively handle a diverse set of tasks simultaneously?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 58, + 455, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 455, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 58, + 455, + 287, + 467 + ], + "type": "text", + "content": "3. Can our network generalize to real-world robot tasks?" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 473, + 220, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 473, + 220, + 486 + ], + "spans": [ + { + "bbox": [ + 47, + 473, + 220, + 486 + ], + "type": "text", + "content": "6.1.2BY2 Dataset Main Experiment" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 491, + 162, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 162, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 162, + 504 + ], + "type": "text", + "content": "6.1.1 Experiment Set Up" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 510, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 605 + ], + "type": "text", + "content": "Tasks. We divide the 18 assembly tasks in the 2BY2 dataset into training and testing sets individually and compared the performance of our method with various baseline approaches. To further evaluate its cross-task generalization ability, we conducted additional experiments on tasks such as Lid Covering, Insertion, and High Precision Placement, as well as All task, which requires the method to handle all tasks in the entire dataset. See Table 2 for task details." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 605, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 677 + ], + "type": "text", + "content": "Evaluation metrics. Following metrics from datasets like Breaking Bad [41] and Neural Shape Mating [5], we use Root Mean Squared Error (RMSE) to evaluate both rotation and translation of the predicted SE(3) pose. Specifically, rotations are represented using Euler angles with symmetry considerations, see Section 3.2 for symmetry details." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "Training parameters. We set batch size to be 4, and the initial learning rate of Adam Optimizer [10] to be 1e-4. We train models for 1000 epochs for them to fully converge." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 72, + 379, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 379, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 379, + 83 + ], + "type": "text", + "content": "6.1.2 Baselines" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 91, + 547, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 91, + 547, + 114 + ], + "spans": [ + { + "bbox": [ + 305, + 91, + 547, + 114 + ], + "type": "text", + "content": "We compare our method with SE-3 assembly [53], Puzzlefusion++ [50], Jigsaw [32] and Neural Shape Mating [5]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 115, + 547, + 307 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 306, + 115, + 547, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 115, + 547, + 163 + ], + "spans": [ + { + "bbox": [ + 306, + 115, + 547, + 163 + ], + "type": "text", + "content": "- SE-3 Assembly [53] proposes a network architecture to leverage SE(3) equivariance for representations considering multi-part correlations, and predict the pose of each part jointly." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 163, + 547, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 163, + 547, + 209 + ], + "spans": [ + { + "bbox": [ + 306, + 163, + 547, + 209 + ], + "type": "text", + "content": "- Puzzlefusion++ [50] proposes an auto-agglomerative 3D fracture assembly framework. It iteratively aligns and merges fragments using a diffusion model for 6-DoF alignment and a transformer model for verification." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 210, + 547, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 210, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 306, + 210, + 547, + 258 + ], + "type": "text", + "content": "- Jigsaw [32] leverages hierarchical features of global and local geometry to match and align the fracture surfaces, and recovers the global pose of each piece to restore the underlying object." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 258, + 546, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 258, + 546, + 307 + ], + "spans": [ + { + "bbox": [ + 306, + 258, + 546, + 307 + ], + "type": "text", + "content": "- Neural Shape Mating [5] utilizes PointNet for feature encoding and a transformer for feature fusion to learn the correlations between assembly parts, enabling joint prediction of their poses." + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 322, + 506, + 334 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 322, + 506, + 334 + ], + "spans": [ + { + "bbox": [ + 306, + 322, + 506, + 334 + ], + "type": "text", + "content": "6.1.3 2BY2 Benchmark Results and Analysis" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 340, + 546, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 340, + 546, + 399 + ], + "spans": [ + { + "bbox": [ + 305, + 340, + 546, + 399 + ], + "type": "text", + "content": "Table 3 presents the quantitative performance of our method compared to all baselines. The results show that our approach outperforms the baselines across 18 fine-grained assembly tasks, with an average improvement of 0.046 in translation RMSE and 8.97 in rotation RMSE." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 400, + 547, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 547, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 547, + 519 + ], + "type": "text", + "content": "Additionally, we evaluate our method on three cross-category tasks defined in Section 3.1, namely Lid Covering, Inserting and High Precision Placing, and achieve the state-of-the-art performance. Moreover, in the most comprehensive All task, we outperform the baseline by 0.123 in translation and 10.90 in rotation, demonstrating strong generalization across tasks and object shapes. In the meantime, baseline comparisons confirm the rigor and challenge of our tasks. Results on challenging tasks like Plug and Key highlight our framework's effectiveness in complex scenarios." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 519, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 519, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 519, + 547, + 616 + ], + "type": "text", + "content": "We analyze that the superior performance of our designed network is due to the approach of separately predicting the poses of the two objects in a step-by-step manner. This prevents the pose errors from interfering with each other, which often occurs in other baselines when predicting both poses simultaneously. Additionally, the design of our encoder makes our network more sensitive to subtle changes in rotation and translation, resulting better performance." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 623, + 473, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 473, + 636 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 473, + 636 + ], + "type": "text", + "content": "6.2. Real-World Robot Experiment" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": "Real-world robot experiment setup. As shown in Figure 5, we conduct our real-world robot experiments using a UR5 robotic arm, equipped with a Robotiq 2F-85 Gripper. We select four tasks, Cup, Flower, Bread and Plug, demonstrating that our model exhibits strong generalization ability on unseen real-world objects." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17388" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 70, + 547, + 316 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 316 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 316 + ], + "type": "table", + "html": "
TaskJigsaw [32]Puzzlefusion++ [50]NSM [5]SE(3)-Assembly [53]Ours
RMSE(T)RMSE(R)RMSE(T)RMSE(R)RMSE(T)RMSE(R)RMSE(T)RMSE(R)RMSE(T)↓RMSE(R)↓
Lid Covering0.39833.330.40837.740.18433.450.12521.370.09016.12
Kitchenport0.47745.800.42347.230.23757.470.09317.290.06816.60
Bottle0.41134.710.38535.230.22778.580.14736.920.07627.70
Kettle0.33543.710.37238.380.21561.100.13313.150.11111.56
Coffeemachine0.52732.670.43734.640.25350.660.14224.430.07622.83
Cup0.40833.550.43933.580.26067.350.16046.460.12223.18
Inserting0.36453.580.32757.830.27569.930.19946.30.14238.03
Plug0.37256.890.34848.890.30352.260.17618.580.0949.74
Childrentoy0.26859.880.24563.210.27193.770.30280.520.24257.81
Letter0.40967.990.35772.080.31776.480.12139.240.09433.74
Bread0.22057.840.20160.920.17165.500.11151.130.09036.40
Nut0.47640.080.32347.290.27155.320.10246.680.05135.60
Coin0.40639.620.34851.400.28962.580.11128.690.10722.88
Key0.38442.850.34850.380.29063.600.08717.280.04516.32
Usb0.46367.410.34258.230.25269.900.21532.280.12828.98
Precision Placing0.37573.940.28767.810.21185.020.13457.860.11544.84
Box0.13733.720.13440.470.13072.470.07125.080.06621.53
Tissue0.29282.390.26585.180.17579.780.18373.020.11564.37
Flower0.32864.390.28359.040.24687.180.21364.890.12542.23
Teaport0.30268.330.32461.010.28856.320.08540.590.05026.11
Position0.42358.070.38957.550.25770.110.16628.170.14124.46
ALL0.36053.340.34258.230.28470.300.23352.340.11041.44
", + "image_path": "fe3a5a31411bc4697c273c9d0ebb776e192ad54a845b79757e56bef8eaeaa9da.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 381, + 289, + 545 + ], + "blocks": [ + { + "bbox": [ + 48, + 381, + 289, + 545 + ], + "lines": [ + { + "bbox": [ + 48, + 381, + 289, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 381, + 289, + 545 + ], + "type": "image", + "image_path": "d63afcedd816608ab0e5ef1a9cc93c4eca7d77de0aa3eab58cda76220feee797.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 552, + 288, + 574 + ], + "lines": [ + { + "bbox": [ + 46, + 552, + 288, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 288, + 574 + ], + "type": "text", + "content": "Figure 5. Real Robot Setup. We conduct real-world robot experiments on Cup, Flower, Bread and Plug tasks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "We place objects in the scene with random initial poses and scan them to obtain their point clouds. Using pre-trained models on selected data of 2BY2 dataset, we predict the pose of each object. A manually designed grasping pose is then applied to pick up each object, and based on the predicted poses, the robotic arm plans a trajectory to complete the assembly. We use SE(3) assembly [53] as the baseline and test our approach on 10 different initial poses. As shown in Table 4, our method significantly outperforms the baseline." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 325, + 375, + 525, + 416 + ], + "blocks": [ + { + "bbox": [ + 45, + 323, + 546, + 357 + ], + "lines": [ + { + "bbox": [ + 45, + 323, + 546, + 357 + ], + "spans": [ + { + "bbox": [ + 45, + 323, + 546, + 357 + ], + "type": "text", + "content": "Table 3. Quantitative Evaluation on 2BY2 for Pairwise Object Assembly. Our method outperforms the baseline across all 18 fine-grained assembly tasks, as well as demonstrating significant improvement on three cross-category assembly tasks. It achieves a lower task average with a reduction of 0.046 in translation RMSE and 8.97 in rotation RMSE." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 375, + 525, + 416 + ], + "lines": [ + { + "bbox": [ + 325, + 375, + 525, + 416 + ], + "spans": [ + { + "bbox": [ + 325, + 375, + 525, + 416 + ], + "type": "table", + "html": "
TaskCupFlowerBreadPlugOverall
SE(3) [53]2/104/101/102/1022.5%
Ours8/1010/106/107/1077.5%
", + "image_path": "5ceb72ad0089a4bd1c9c060a305ad6c5ea65191b187c2fb19519b792044fcaa4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 307, + 424, + 543, + 435 + ], + "lines": [ + { + "bbox": [ + 307, + 424, + 543, + 435 + ], + "spans": [ + { + "bbox": [ + 307, + 424, + 543, + 435 + ], + "type": "text", + "content": "Table 4. Real-World Robot Experiment Success Rate Results." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 459, + 399, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 459, + 399, + 472 + ], + "spans": [ + { + "bbox": [ + 306, + 459, + 399, + 472 + ], + "type": "text", + "content": "7. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 481, + 545, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 481, + 545, + 516 + ], + "spans": [ + { + "bbox": [ + 305, + 481, + 545, + 516 + ], + "type": "text", + "content": "In this section, we conduct comprehensive experiments to demonstrate the rationality of our network design and the effectiveness of each module." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 518, + 547, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 518, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 305, + 518, + 547, + 555 + ], + "type": "text", + "content": "Encoder. To validate the effectiveness of two-scale Vector Neuron(VN) DGCNN, we compare it with other encoders: VN DGCNN [9], DGCNN [48], and PointNet [38]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "text", + "content": "Two-step network design. We compare our method with an end-to-end approach, which jointly predicts the pose of " + }, + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_A" + }, + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{B}}" + }, + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "text", + "content": ". Specifically, we utilize " + }, + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_A" + }, + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "text", + "content": " and the input point cloud of " + }, + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_A" + }, + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{B}}" + }, + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "text", + "content": " to get the 6D pose of " + }, + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "inline_equation", + "content": "P_A" + }, + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "text", + "content": ", while using the same encoder to extract " + }, + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{B}}" + }, + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "text", + "content": " and pass it through the same pose prediction head to predict the pose of " + }, + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "inline_equation", + "content": "P_B" + }, + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": "As shown in in Table 5, we show the results of ablation studies on Lid covering, Inserting and Precision Placing, which are more comprehensive and require cross-task generalization abilities. The performance declines in both translation and rotation when we removing our two-scale VN DGCNN encoder and change it to Vector Neuron DGCNN [9], DGCNN [48], Pointnet [38]. It demonstrates" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "17389" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 545, + 401 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 545, + 401 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 545, + 401 + ], + "type": "image", + "image_path": "d4f59e3949644bd077a197c54f2a47ce948e6083951ed494a49e90ed2345e1bd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 411, + 547, + 434 + ], + "lines": [ + { + "bbox": [ + 46, + 411, + 547, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 547, + 434 + ], + "type": "text", + "content": "Figure 6. Qualitative Results Comparison. We highlight Bottle, Plug, Bread, Letter, Childrentoy, Key, and Flower tasks to demonstrate our improved translation and rotation predictions compared to baseline methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 444, + 545, + 515 + ], + "blocks": [ + { + "bbox": [ + 48, + 444, + 545, + 515 + ], + "lines": [ + { + "bbox": [ + 48, + 444, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 545, + 515 + ], + "type": "table", + "html": "
TaskVector Neuron DGCNN [9]DGCNN [48]PointNet [38]w/o Two-stepOurs
RMSE(T)RMSE(R)RMSE(T)RMSE(R)RMSE(T)RMSE(R)RMSE(T)RMSE(R)RMSE(T) ↓RMSE(R) ↓
Lid Covering0.09818.230.24570.060.23465.320.11718.740.09016.12
Inserting0.15741.220.23462.250.23466.800.16441.110.14238.03
Precision Placing0.12148.010.24565.190.21172.470.13746.380.11544.84
ALL0.12344.670.27772.460.26475.380.13945.200.11041.44
", + "image_path": "190233c59c3af2ea98caa8719e2950677be593f307ecf13ba8aa4e092b73634a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 522, + 547, + 556 + ], + "lines": [ + { + "bbox": [ + 46, + 522, + 547, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 547, + 556 + ], + "type": "text", + "content": "Table 5. Ablation Study Results. We compare various encoders including Vector Neuron DGCNN [9], DGCNN [48], PointNet [38], and our proposed two-scale Vector Neuron DGCNN. We also compare end-to-end networks with two-step networks to demonstrate the effectiveness of each component in our network design." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 577, + 288, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 577, + 288, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 577, + 288, + 662 + ], + "type": "text", + "content": "that our encoder effectively exploits the advantage of SE(3) equivariance, enabling greater sample efficiency and more robust generalization abilities. Compared with version in Figure 4, the experiment performance declines when we change our two-step network in a joint-learning manner, proving that our two-step network design can reduce error caused by jointly predictions and thereby is more effective." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "text", + "content": "8. Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "2BY2 is a significant step in bridging the gap between geometry-based assembly tasks and everyday object assem" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 577, + 547, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 577, + 547, + 709 + ], + "spans": [ + { + "bbox": [ + 304, + 577, + 547, + 709 + ], + "type": "text", + "content": "blies. With pose and symmetry annotations for 517 object pairs across 18 fine-grained tasks, 2BY2 sets a new benchmark for 3D assembly challenges. Our two-step pairwise SE(3) pose estimation framework, which leverages equivariant features, demonstrates superior performance over existing approaches in reducing both translation and rotation errors. Robot experiments further validate the method's generalizability in practical 3D assembly scenarios. In conclusion, 2BY2 provides both a comprehensive benchmark and an effective framework, with the aim of inspiring and supporting more generalizable solution in robot manipulation." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17390" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 113 + ], + "type": "text", + "content": "[1] 3dwarehouse. https://3dwarehouse.sketchup.com/by/su3dchallenge, 2014.3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 289, + 136 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 289, + 136 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 289, + 136 + ], + "type": "text", + "content": "[2] google-sketchup-3d-challenge. https://3dwarehouse.sketchup.com/by/su3dchallenge, 2014.3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 137, + 289, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 289, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 289, + 201 + ], + "type": "text", + "content": "[3] Benedict J Brown, Corey Toler-Franklin, Diego Nehab, Michael Burns, David Dobkin, Andreas Vlachopoulos, Christos Doumas, Szymon Rusinkiewicz, and Tim Weyrich. A system for high-volume acquisition and matching of fresco fragments: Reassembling theran wall paintings. ACM transactions on graphics (TOG), 27(3):1-9, 2008. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 202, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 247 + ], + "type": "text", + "content": "[4] Siddhartha Chaudhuri, Evangelos Kalogerakis, Leonidas Guibas, and Vladlen Koltun. Probabilistic reasoning for assembly-based 3d modeling. In ACM SIGGRAPH 2011 papers, pages 1-10. 2011. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 247, + 288, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 302 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 302 + ], + "type": "text", + "content": "[5] Yun-Chun Chen, Haoda Li, Dylan Turpin, Alec Jacobson, and Animesh Garg. Neural shape mating: Self-supervised object assembly with adversarial shape priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12724-12733, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 303, + 288, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 303, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 303, + 288, + 335 + ], + "type": "text", + "content": "[6] Junfeng Cheng, Mingdong Wu, Ruiyuan Zhang, Guanqi Zhan, Chao Wu, and Hao Dong. Score-pa: Score-based 3d part assembly. In British Machine Vision Conference, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 336, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 336, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 336, + 288, + 357 + ], + "type": "text", + "content": "[7] Taco S Cohen, Mario Geiger, Jonas Kohler, and Max Welling. Spherical cnns. arXiv preprint arXiv:1801.10130, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 358, + 288, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 288, + 413 + ], + "type": "text", + "content": "[8] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of " + }, + { + "bbox": [ + 53, + 358, + 288, + 413 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{d}" + }, + { + "bbox": [ + 53, + 358, + 288, + 413 + ], + "type": "text", + "content": " objects. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 415, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 415, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 415, + 288, + 468 + ], + "type": "text", + "content": "[9] Congyue Deng, Or Litany, Yueqi Duan, Adrien Poulenard, Andrea Tagliasacchi, and Leonidas J Guibas. Vector neurons: A general framework for so (3)-equivariant networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12200-12209, 2021. 2, 4, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 470, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 470, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 470, + 288, + 491 + ], + "type": "text", + "content": "[10] P Kingma Diederik. Adam: A method for stochastic optimization. (No Title), 2014. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 492, + 288, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 288, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 288, + 545 + ], + "type": "text", + "content": "[11] Thomas Funkhouser, Michael Kazhdan, Philip Shilane, Patrick Min, William Kiefer, Ayellet Tal, Szymon Rusinkiewicz, and David Dobkin. Modeling by example. ACM transactions on graphics (TOG), 23(3):652-663, 2004. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 547, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 288, + 602 + ], + "type": "text", + "content": "[12] Thomas Funkhouser, Hijung Shin, Corey Toler-Franklin, Antonio García Castañeda, Benedict Brown, David Dobkin, Szymon Rusinkiewicz, and Tim Weyrich. Learning how to match fresco fragments. Journal on Computing and Cultural Heritage (JOCCH), 4(2):1-13, 2011. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 603, + 288, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 288, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 288, + 656 + ], + "type": "text", + "content": "[13] Chongkai Gao, Zhengrong Xue, Shuying Deng, Tianhai Liang, Siqi Yang, Lin Shao, and Huazhe Xu. Riemann: Near real-time se (3)-equivariant robot manipulation without point cloud segmentation. arXiv preprint arXiv:2403.19460, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "text", + "content": "[14] Stefan Hinterstoisser, Stefan Holzer, Cedric Cagniart, Slobodan Ilic, Kurt Konolige, Nassir Navab, and Vincent Lepetit. Multimodal templates for real-time detection of texture-less objects in heavily cluttered scenes. In 2011 International Conference on Computer Vision, pages 858-865, 2011. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 139 + ], + "type": "text", + "content": "[15] Stefan Hinterstoisser, Vincent Lepetit, Slobodan Ilic, Stefan Holzer, Gary Bradski, Kurt Konolige, and Nassir Navab. Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In Computer Vision - ACCV 2012, pages 548-562, Berlin, Heidelberg, 2013. Springer Berlin Heidelberg. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 140, + 547, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 547, + 194 + ], + "type": "text", + "content": "[16] Sepidehsadat Sepid Hossieni, Mohammad Amin Shabani, Saghar Irandoust, and Yasutaka Furukawa. Puzzle: unleashing the power of diffusion models for spatial puzzle solving. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 196, + 547, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 196, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 196, + 547, + 239 + ], + "type": "text", + "content": "[17] Boce Hu, Xupeng Zhu, Dian Wang, Zihao Dong, Haojie Huang, Chenghao Wang, Robin Walters, and Robert Platt. Orbitgrasp: Se (3)-equivariant grasp learning. In 8th Annual Conference on Robot Learning, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 241, + 547, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 547, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 547, + 296 + ], + "type": "text", + "content": "[18] Haojie Huang, Dian Wang, Xupeng Zhu, Robin Walters, and Robert Platt. Edge grasp network: A graph-based se (3)-invariant approach to grasp detection. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3882-3888. IEEE, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 297, + 546, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 546, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 546, + 341 + ], + "type": "text", + "content": "[19] Haojie Huang, Haotian Liu, Dian Wang, Robin Walters, and Robert Platt. Match policy: A simple pipeline from point cloud registration to manipulation policies. arXiv preprint arXiv:2409.15517, 2024." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 342, + 547, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 547, + 397 + ], + "type": "text", + "content": "[20] Haojie Huang, Karl Schmeckpeper, Dian Wang, Ondrej Biza, Yaoyao Qian, Haotian Liu, Mingxi Jia, Robert Platt, and Robin Walters. IMAGINATION POLICY: Using generative point cloud models for learning manipulation policies. In 8th Annual Conference on Robot Learning, 2024. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 399, + 547, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 547, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 547, + 442 + ], + "type": "text", + "content": "[21] Qi-Xing Huang, Simon Flöry, Natasha Gelfand, Michael Hofer, and Helmut Pottmann. Reassembling fractured objects by geometric matching. In ACM Siggraph 2006 papers, pages 569-578. 2006. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 444, + 547, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 547, + 487 + ], + "type": "text", + "content": "[22] Prakhar Jaiswal, Jinmiao Huang, and Rahul Rai. Assembly-based conceptual 3d modeling with unlabeled components using probabilistic factor graph. Computer-Aided Design, 74: 45-54, 2016. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 488, + 547, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 488, + 547, + 532 + ], + "spans": [ + { + "bbox": [ + 307, + 488, + 547, + 532 + ], + "type": "text", + "content": "[23] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J Davison. Rlbench: The robot learning benchmark & learning environment. IEEE Robotics and Automation Letters, 5(2):3019-3026, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 534, + 547, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 547, + 587 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 547, + 587 + ], + "type": "text", + "content": "[24] Benjamin Jones, Dalton Hildreth, Duowen Chen, Ilya Baran, Vladimir G Kim, and Adriana Schulz. Automate: A dataset and learning approach for automatic mating of cad assemblies. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 590, + 547, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 547, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 547, + 644 + ], + "type": "text", + "content": "[25] Yuanchen Ju, Kaizhe Hu, Guowei Zhang, Gu Zhang, Mingrun Jiang, and Huazhe Xu. Robo-abc: Affordance generalization beyond categories via semantic correspondence for robot manipulation. In European Conference on Computer Vision, pages 222–239. Springer, 2025. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 646, + 546, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 546, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 546, + 689 + ], + "type": "text", + "content": "[26] Evangelos Kalogerakis, Siddhartha Chaudhuri, Daphne Koller, and Vladlen Koltun. A probabilistic model for component-based shape synthesis. Acm Transactions on Graphics (TOG), 31(4):1-11, 2012. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 691, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 547, + 713 + ], + "type": "text", + "content": "[27] Wadim Kehl, Fabian Manhardt, Federico Tombari, Slobodan Ilic, and Nassir Navab. SSD-6D: making rgb-based" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17391" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 288, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 288, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 288, + 95 + ], + "type": "text", + "content": "3d detection and 6d pose estimation great again. CoRR, abs/1711.10006, 2017. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 152 + ], + "type": "text", + "content": "[28] Yichen Li, Kaichun Mo, Lin Shao, Minhyuk Sung, and Leonidas Guibas. Learning 3d part assembly from a single image. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 664-682. Springer, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 185 + ], + "type": "text", + "content": "[29] Yulong Li, Andy Zeng, and Shuran Song. Rearrangement planning for general part assembly. In 7th Annual Conference on Robot Learning, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 288, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 241 + ], + "type": "text", + "content": "[30] Yuval Litvak, Armin Biess, and Aharon Bar-Hillel. Learning pose estimation for high-precision robotic assembly using simulated depth images. In 2019 International Conference on Robotics and Automation (ICRA), pages 3521–3527. IEEE, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 243, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 288, + 297 + ], + "type": "text", + "content": "[31] Hao-Yu Liu, Jian-Wei Guo, Hai-Yong Jiang, Yan-Chao Liu, Xiao-Peng Zhang, and Dong-Ming Yan. Puzzlenet: boundary-aware feature matching for non-overlapping 3d point clouds assembly. Journal of Computer Science and Technology, 38 (3):492-509, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 299, + 288, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 299, + 288, + 332 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 288, + 332 + ], + "type": "text", + "content": "[32] Jiaxin Lu, Yifan Sun, and Qixing Huang. Jigsaw: Learning to assemble multiple fractured objects. Advances in Neural Information Processing Systems, 36, 2024. 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 333, + 288, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 333, + 288, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 333, + 288, + 398 + ], + "type": "text", + "content": "[33] Jianlan Luo, Eugen Solowjow, Chengtao Wen, Juan Aparicio Ojea, Alice M Agogino, Aviv Tamar, and Pieter Abbeel. Reinforcement learning on variable impedance controller for high-precision robotic assembly. In 2019 International Conference on Robotics and Automation (ICRA), pages 3080–3087. IEEE, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 288, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 465 + ], + "type": "text", + "content": "[34] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909-918, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 467, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 288, + 521 + ], + "type": "text", + "content": "[35] Yashraj Narang, Kier Storey, Iretiayo Akinola, Miles Macklin, Philipp Reist, Lukasz Wawrzyniak, Yunrong Guo, Adam Moravanszky, Gavriel State, Michelle Lu, et al. Factory: Fast contact for robotic assembly. arXiv preprint arXiv:2205.03532, 2022. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 523, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 288, + 578 + ], + "type": "text", + "content": "[36] Abhinav Narayan, Rajendra Nagar, and Shanmuganathan Ram. Rgl-net: A recurrent graph learning framework for progressive part assembly. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 78-87, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 288, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 288, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 288, + 623 + ], + "type": "text", + "content": "[37] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "type": "text", + "content": "[38] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "text", + "content": "[39] Ismael Rodriguez, Korbinian Nottensteiner, Daniel Leidner, Michael KaBecker, Freek Stulp, and Alin Albu-Schaffer. Iteratively refined feasibility checks in robotic assembly sequence" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 547, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 547, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 547, + 95 + ], + "type": "text", + "content": "planning. IEEE Robotics and Automation Letters, 4(2):1416-1423, 2019. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 547, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 152 + ], + "type": "text", + "content": "[40] Gianluca Scarpellini, Stefano Fiorini, Francesco Giuliani, Pietro Moreiro, and Alessio Del Bue. Diffassemble: A unified graph-diffusion model for 2d and 3d reassembly. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28098-28108, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 546, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 546, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 546, + 196 + ], + "type": "text", + "content": "[41] Silvia Sellán, Yun-Chun Chen, Ziyi Wu, Animesh Garg, and Alec Jacobson. Breaking bad: A dataset for geometric fracture and reassembly. Advances in Neural Information Processing Systems, 35:38885-38898, 2022. 2, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 197, + 546, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 546, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 546, + 252 + ], + "type": "text", + "content": "[42] Hijung Shin, Christos Doumas, Thomas Funkhouser, Szymon Rusinkiewicz, Kenneth Steiglitz, Andreas Vlachopoulos, and Tim Weyrich. Analyzing and simulating fracture patterns of theran wall paintings. Journal on Computing and Cultural Heritage (JOCCH), 5(3):1-14, 2012. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 254, + 547, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 547, + 308 + ], + "type": "text", + "content": "[43] Chenrui Tie, Yue Chen, Ruihai Wu, Boxuan Dong, Zeyi Li, Chongkai Gao, and Hao Dong. ET-SEED: EFFICIENT TRAJECTORY-LEVEL SE(3) EQUIVARIANT DIFFUSION POLICY. In The Thirteenth International Conference on Learning Representations, 2025. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 309, + 546, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 546, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 546, + 353 + ], + "type": "text", + "content": "[44] Joshua Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. CoRR, abs/1703.06907, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 354, + 546, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 546, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 546, + 398 + ], + "type": "text", + "content": "[45] Jonathan Tremblay, Thang To, Balakumar Sundaralingam, Yu Xiang, Dieter Fox, and Stan Birchfield. Deep object pose estimation for semantic robotic grasping of household objects. ArXiv, abs/1809.10790, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 399, + 546, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 546, + 432 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 546, + 432 + ], + "type": "text", + "content": "[46] Weiwei Wan, Kensuke Harada, and Kazuyuki Nagata. Assembly sequence planning for motion planning. CoRR, abs/1609.03108, 2016. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 434, + 546, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 546, + 477 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 546, + 477 + ], + "type": "text", + "content": "[47] Ruocheng Wang, Yunzhi Zhang, Jiayuan Mao, Ran Zhang, Chin-Yi Cheng, and Jiajun Wu. Iea-manual: Seeing shape assembly step by step. Advances in Neural Information Processing Systems, 35:28428-28440, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 478, + 546, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 546, + 522 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 546, + 522 + ], + "type": "text", + "content": "[48] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 523, + 546, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 546, + 556 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 546, + 556 + ], + "type": "text", + "content": "[49] Ziming Wang and Rebecka Jornsten. Se (3)-bi-equivariant transformers for point cloud assembly. arXiv preprint arXiv:2407.09167, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 557, + 546, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 546, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 546, + 601 + ], + "type": "text", + "content": "[50] Zhengqing Wang, Jiacheng Chen, and Yasutaka Furukawa. Puzzlefusion++: Auto-agglomerative 3d fracture assembly by denoise and verify. arXiv preprint arXiv:2406.00259, 2024. 2, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 602, + 546, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 546, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 546, + 678 + ], + "type": "text", + "content": "[51] Karl DD Willis, Pradeep Kumar Jayaraman, Hang Chu, Yunsheng Tian, Yifei Li, Daniele Grandi, Aditya Sanghi, Linh Tran, Joseph G Lambourne, Armando Solar-Lezama, et al. Joinable: Learning bottom-up assembly of parametric cad joints. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15849-15860, 2022. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 680, + 546, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 546, + 714 + ], + "type": "text", + "content": "[52] Rundi Wu, Yixin Zhuang, Kai Xu, Hao Zhang, and Baoquan Chen. Pq-net: A generative part seq2seq network for 3d shapes. In Proceedings of the IEEE/CVF Conference on" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17392" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "type": "text", + "content": "Computer Vision and Pattern Recognition, pages 829-838, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 151 + ], + "type": "text", + "content": "[53] Ruihai Wu, Chenrui Tie, Yushi Du, Yan Zhao, and Hao Dong. Leveraging se (3) equivariance for learning 3d geometric shape assembly. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14311-14320, 2023. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 217 + ], + "type": "text", + "content": "[54] Fanbo Xiang, Yuzhe Qin, Kaichun Mo, Yikuan Xia, Hao Zhu, Fangchen Liu, Minghua Liu, Hanxiao Jiang, Yifu Yuan, He Wang, Li Yi, Angel X. Chang, Leonidas J. Guibas, and Hao Su. SAPIEN: A simulated part-based interactive environment. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 288, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 288, + 253 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 288, + 253 + ], + "type": "text", + "content": "[55] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 288, + 308 + ], + "type": "text", + "content": "[56] Yuwen Xiong, Wei-Chiu Ma, Jingkang Wang, and Raquel Urtasun. Learning compact representations for lidar completion and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1074-1083, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 288, + 342 + ], + "type": "text", + "content": "[57] Boshen Xu, Sipeng Zheng, and Qin Jin. Spaformer: Sequential 3d part assembly with transformers. arXiv preprint arXiv:2403.05874, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 344, + 288, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 344, + 288, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 344, + 288, + 399 + ], + "type": "text", + "content": "[58] Xianghao Xu, Paul Guerrero, Matthew Fisher, Siddhartha Chaudhuri, and Daniel Ritchie. Unsupervised 3d shape reconstruction by part retrieval and assembly. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8559-8567, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 288, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 455 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 455 + ], + "type": "text", + "content": "[59] Zhengrong Xue, Zhecheng Yuan, Jiashun Wang, Xueqian Wang, Yang Gao, and Huazhe Xu. Usek: Unsupervised se (3)-equivariant 3d keypoints for generalizable manipulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1715-1722. IEEE, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 288, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 288, + 510 + ], + "type": "text", + "content": "[60] Kangxue Yin, Zhiqin Chen, Siddhartha Chaudhuri, Matthew Fisher, Vladimir G Kim, and Hao Zhang. Coalesce: Component assembly by learning to synthesize connections. In 2020 International Conference on 3D Vision (3DV), pages 61-70. IEEE, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 512, + 288, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 288, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 288, + 566 + ], + "type": "text", + "content": "[61] Kevin Zakka, Andy Zeng, Johnny Lee, and Shuran Song. Form2fit: Learning shape priors for generalizable assembly from disassembly. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9404–9410. IEEE, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "type": "text", + "content": "[62] Guanqi Zhan, Qingnan Fan, Kaichun Mo, Lin Shao, Baoquan Chen, Leonidas J Guibas, Hao Dong, et al. Generative 3d part assembly via dynamic graph learning. Advances in Neural Information Processing Systems, 33:6315-6326, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 288, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 656 + ], + "type": "text", + "content": "[63] Rufeng Zhang, Tao Kong, Weihao Wang, Xuan Han, and Mingyu You. 3d part assembly generation with instance encoded transformer. IEEE Robotics and Automation Letters, 7(4):9051-9058, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[64] Ruiyuan Zhang, Jiaxiang Liu, Zexi Li, Hao Dong, Jie Fu, and Chao Wu. Scalable geometric fracture assembly via co-creation space among assemblers. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7269–7277, 2024. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 262 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "text", + "content": "[65] Haibo Zhao, Dian Wang, Yizhe Zhu, Xupeng Zhu, Owen Howell, Linfeng Zhao, Yaoyao Qian, Robin Walters, and Robert Platt. Hierarchical equivariant policy via frame transf. arXiv preprint arXiv:2502.05728, 2025. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 547, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 173 + ], + "type": "text", + "content": "[66] Junzhe Zhu, Yuanchen Ju, Junyi Zhang, Muhan Wang, Zhecheng Yuan, Kaizhe Hu, and Huazhe Xu. Densehunter: Learning 3d semantic correspondence for category-level manipulation from a single demo. International Conference on Learning Representations (ICLR) 2025, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 546, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 546, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 546, + 217 + ], + "type": "text", + "content": "[67] Xupeng Zhu, Dian Wang, Guanang Su, Ondrej Biza, Robin Walters, and Robert Platt. On robot grasp learning using equivariant models. Autonomous Robots, 47(8):1175-1193, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 219, + 547, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 547, + 262 + ], + "type": "text", + "content": "[68] Yuke Zhu, Josiah Wong, Ajay Mandlekar, Roberto Martin-Martin, Abhishek Joshi, Soroush Nasiriany, and Yifeng Zhu. robosuite: A modular simulation framework and benchmark for robot learning. arXiv preprint arXiv:2009.12293, 2020. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17393" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Two is Better than One_ Efficient Ensemble Defense for Robust and Compact Models/f8d31d9f-8dff-45eb-9d02-43167ab5a77a_content_list.json b/2025/Two is Better than One_ Efficient Ensemble Defense for Robust and Compact Models/f8d31d9f-8dff-45eb-9d02-43167ab5a77a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..855b2aac957c331d76bbb13fb2f97ff602d4f30d --- /dev/null +++ b/2025/Two is Better than One_ Efficient Ensemble Defense for Robust and Compact Models/f8d31d9f-8dff-45eb-9d02-43167ab5a77a_content_list.json @@ -0,0 +1,1999 @@ +[ + { + "type": "text", + "text": "Two is Better than One: Efficient Ensemble Defense for Robust and Compact Models", + "text_level": 1, + "bbox": [ + 194, + 130, + 803, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yoojin Jung Byung Cheol Song", + "bbox": [ + 362, + 204, + 635, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Department of Electrical and Computer Engineering, Inha University Incheon, Republic of Korea", + "bbox": [ + 225, + 222, + 769, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yoon0505@gmail.com, bcsong@inha.ac.kr", + "bbox": [ + 323, + 258, + 666, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 309, + 326, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning-based computer vision systems adopt complex and large architectures to improve performance, yet they face challenges in deployment on resource-constrained mobile and edge devices. To address this issue, model compression techniques such as pruning, quantization, and matrix factorization have been proposed; however, these compressed models are often highly vulnerable to adversarial attacks. We introduce the Efficient Ensemble Defense (EED) technique, which diversifies the compression of a single base model based on different pruning importance scores and enhances ensemble diversity to achieve high adversarial robustness and resource efficiency. EED dynamically determines the number of necessary sub-models during the inference stage, minimizing unnecessary computations while maintaining high robustness. On the CIFAR-10 and SVHN datasets, EED demonstrated state-of-the-art robustness performance compared to existing adversarial pruning techniques, along with an inference speed improvement of up to 1.86 times. This proves that EED is a powerful defense solution in resource-constrained environments.", + "bbox": [ + 89, + 342, + 483, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 676, + 220, + 693 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Convolutional Neural Networks (CNNs) have shown remarkable performance across various computer vision tasks, such as image classification, object detection, and semantic segmentation[18, 33, 36]. However, CNNs have become increasingly wider and deeper to achieve higher performance, leading to complex architectures with vast numbers of parameters. This complexity inherently results in significant computational costs and storage demands, making deployment on resource-constrained edge devices challenging[4, 42]. To address this challenge, a number of model compression techniques, including pruning[14, 24], quantization[3], and matrix decomposition[8], have been proposed to minimize model size and computational load", + "bbox": [ + 89, + 703, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e096d87d155417cb0031f40ba099c84fc66be49dfab080bf834e915603e72821.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 311, + 666, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c94bb48dc532955dfc01c11083e4d6574dbb24eb823279cda220acde293248c2.jpg", + "image_caption": [], + "image_footnote": [ + "Important nodes or connections may be removed incorrectly due to adversarial attack" + ], + "bbox": [ + 687, + 311, + 807, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/00a49362902aad1dadfca3eaa37b44ed40e9fc49092ecbf60a09778070a0a78a.jpg", + "image_caption": [ + "Figure 1. Challenges in AP and the potential of ensemble methods: AP weakens the network by removing critical parameters, reducing robustness and increasing vulnerability to adversarial attacks. An ensemble approach mitigates pruning effects by combining pruned models, enhancing robustness through their collective strength (illustrated in green)." + ], + "image_footnote": [ + "Ensemble could help compensate" + ], + "bbox": [ + 609, + 422, + 707, + 530 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "without compromising accuracy.", + "bbox": [ + 511, + 654, + 732, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meanwhile, deep learning models are highly vulnerable to adversarial attacks that induce false predictions by disturbing the input image with small perturbations[11, 37]. This vulnerability is particularly pronounced in sparsely compressed models, which are even more sensitive to adversarial attacks than standard models[35, 44]. Consequently, common adversarial defense techniques such as adversarial training (AT) [20, 27], label smoothing [41, 51], and defensive distillation[31], while generally effective, often experience a degradation in performance when applied to compressed models. This indicates that existing compression techniques may compromise adversarial robustness even if they maintain accuracy on clean test sets. In response, adversarial pruning (AP), a model compression technique that considers adversarial robustness[34, 48, 52], has emerged;", + "bbox": [ + 509, + 674, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9696", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "however, AP still exhibits lower robustness compared to non-compressed defense models[32].", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We assume that this degradation in robustness arises from reduced generalization capability and information loss caused by pruning. Specifically, adversarial attacks may cause the model to misidentify and even remove critical nodes or connections during the pruning process. In Sec.3.3, we validate this assumption by observing importance scores, which reflect how each component contributes to the loss function through the gradients of the weights[50]. Our findings demonstrate that adversarial examples significantly impact the importance scores of individual parameters.", + "bbox": [ + 88, + 126, + 482, + 292 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Ensemble defense[7, 39] has recently gained attention in adversarial defense research for enhancing the stability and robustness of multiple base classifiers against attacks. As illustrated in Fig.1, we believe that ensemble defense methods can help mitigate generalization degradation and information loss during the pruning process. Since ensemble learning can compensate for deficiencies in individual compressed models due to removed parameters by other models and offset individual errors by combining outputs. However, traditional ensemble defenses inherently suffer from a large capacity, as they require multiple base models[15, 30]. This raises an important question: Is it possible to develop an ensemble defense algorithm that not only achieves robustness but also results in a compact compressed model?", + "bbox": [ + 91, + 297, + 483, + 508 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As an answer to this, we propose Efficient Ensemble Defense (EED). Instead of using multiple base models, we compress a single base model by employing different pruning metrics and then utilize these compressed sub-models in the ensemble. This approach reduces the overall capacity compared to traditional ensemble defense models that rely on multiple base models. Given that using smaller models in an ensemble can lead to reduced defensive efficacy[46], we enhance the variety of the compressed models by incorporating a diversity term in the ensemble selection process, thereby improving ensemble effectiveness while maintaining robustness. Additionally, by adopting inference-efficient ensemble[25], we dynamically add sub-networks and perform ensemble operations as needed during inference, improving overall efficiency.", + "bbox": [ + 91, + 513, + 483, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of our work are as follows:", + "bbox": [ + 109, + 744, + 447, + 758 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Propose EED, that overcomes the limitations of traditional AP by enhancing the ensemble diversity and compensating for individual errors, thereby achieving robustness against adversarial attacks while being compact.", + "- Reduce memory by using a single base model in ensemble, and decrease computational costs by dynamically adding models only when necessary during inference.", + "- Achieve state-of-the-art performance across various attack benchmarks on the CIFAR-10 and SVHN datasets." + ], + "bbox": [ + 88, + 763, + 482, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 89, + 663, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Adversarial Attack and Defense", + "text_level": 1, + "bbox": [ + 511, + 114, + 795, + 130 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adversarial attacks exploit the inherent vulnerabilities of deep learning models by adding specific noise to the input, thereby inducing incorrect predictions. These specially crafted inputs, designed to deceive deep learning models, are called adversarial examples. Early studies highlighted the susceptibility of deep learning models to adversarial examples[37]. Then, more potent attacks such as FGSM[11], PGD[27], and AutoAttack[6] have emerged.", + "bbox": [ + 511, + 137, + 903, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, defense methods against attacks have also been evolved. The most well-known adversarial training (AT)[1, 20, 27] enhanced robustness by training adversarial examples. Other strategies involve regularizing geometric properties such as the curvature of the loss function [41, 45], or restoring potentially attacked images to purified versions[29, 49]. Adversarial pruning (AP) and ensemble defense are also one of adversarial defenses (details are given in Sec.3.1.)", + "bbox": [ + 511, + 258, + 903, + 395 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Adversarial Pruning", + "text_level": 1, + "bbox": [ + 511, + 406, + 709, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "AP is a pruning technique (primarily weight pruning) that considers the adversarial context, aiming to reduce the complexity of neural networks while preserving robustness against adversarial attacks. HYDRA demonstrated that empirical adversarial robustness can be achieved through a lowest weight magnitude (LWM) pruning method. R-ADMM[48] and HARP[52] incorporated the alternating direction method of multipliers (ADMM) framework into LWM-based pruning to enhance robustness. In some studies [12, 22, 26], a traditional pruning method, i.e., the Lottery Ticket Hypothesis, was integrated with AT. However, they all failed to maintain generalized robustness at higher sparsity rates $(s_r)[13, 19, 38]$ , and struggled to defend against adaptive attacks[9, 16]. This suggests that conventional AP has not fully resolved the trade-off between model capacity and robustness.", + "bbox": [ + 511, + 429, + 906, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Ensemble Defense", + "text_level": 1, + "bbox": [ + 511, + 681, + 689, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Ensemble defense (ED), which applies ensemble learning to adversarial defense, has been actively studied due to its promising performance[7, 39]. The core of ED is to improve predictions through an ensemble diversity loss that considers adversarial robustness during training. ADP[30] used Shannon entropy and geometric diversity for uncertainty, while GAL[17] utilized cosine similarities between model loss gradients to formalize diversity. DVERGE[47] introduced vulnerability diversity to ensure that each model is robust to the weaknesses of others, and MOL[5] increased perturbation diversity by using expert models trained on different attacks. Thus, ED has significantly enhanced robustness by reducing the risk of overfitting to adversarial ex", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9697", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "amples and diversifying the decision boundary, though with increased model capacity and computational costs.", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 133, + 230, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Notations", + "text_level": 1, + "bbox": [ + 89, + 159, + 202, + 172 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversarial Attack Let $x \\in \\mathbb{R}^d$ be an input example, $y \\in \\{1, \\dots, C\\}$ the true label, and $f(\\cdot)$ a classifier with parameters $\\theta$ . The classifier's prediction for input $x$ is given by $f(x) = \\arg \\max_{i \\in \\{1, \\dots, C\\}} f_i(x)$ , where $f_i(x)$ denotes the predicted probability for class $i$ . An adversarial example $x'$ is crafted by adding a perturbation $\\delta$ to $x$ , aiming to maximize the model's prediction error while keeping $x'$ close to $x$ under a norm constraint. Formally, an adversarial attack is defined by:", + "bbox": [ + 89, + 180, + 483, + 316 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {\\prime} = x + \\delta = \\arg \\max _ {\\| x - x ^ {\\prime} \\| < \\epsilon} \\mathcal {L} (f (x ^ {\\prime}), y) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 325, + 482, + 351 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}$ is the loss function and $\\epsilon$ is the perturbation budget that constrains the strength of the attack under an $\\ell_p$ -norm.", + "bbox": [ + 89, + 359, + 483, + 390 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversarial Training AT is a defense method that improves model robustness by training on adversarially perturbed samples alongside clean samples. The objective for AT is usually formulated by:", + "bbox": [ + 89, + 391, + 483, + 450 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {(x, y) \\sim \\mathcal {D}} \\left[ \\max _ {\\delta : \\| \\delta \\| _ {p} \\leq \\epsilon} \\mathcal {L} \\left(f \\left(x ^ {\\prime}; \\theta\\right), y\\right) \\right] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 459, + 482, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{D}$ is the data distribution.", + "bbox": [ + 89, + 501, + 307, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversarial Pruning Given a pre-trained model $f(\\theta)$ under AT, we aim to maintain robustness while applying a pruning mask $M$ across layers to achieve compression. The general pruning is carried out by removing parameters until the network reaches a desired $s_r$ . So, pruning involves creating $M$ , formalized as follows:", + "bbox": [ + 89, + 516, + 483, + 606 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nM = \\arg \\min _ {\\| \\mathbf {m} \\| _ {0} \\leq 1 - s _ {r}} \\mathcal {L} (f (\\theta) \\otimes m, x, y). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 617, + 482, + 642 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nm = \\left\\{ \\begin{array}{l l} 0, & \\text {i f} \\theta_ {i} \\text {i s t o b e p r u n e d} \\\\ 1, & \\text {o t h e r w i s e} \\end{array} , \\quad \\forall \\theta_ {i} \\in \\theta \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 651, + 482, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For AP, we follow the strategy from HARP[52]. To enhance adversarial robustness, inner maximization generates adversarial examples, targeting a worst-case loss under the PGD attacks:", + "bbox": [ + 89, + 695, + 483, + 753 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\delta} \\mathcal {L} _ {r} (f (\\theta) \\odot M, x + \\delta , y) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 755, + 482, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_r$ defines the robust loss function, incorporating attack-specific parameters. And the global non-uniform pruning strategy jointly optimizes a compression rate $c_{r}$ and importance scores $I$ across layers. This is defined by:", + "bbox": [ + 89, + 781, + 483, + 842 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {c _ {r}, I} \\mathbb {E} _ {(x, y) \\sim D} \\left[ \\max _ {\\delta} \\mathcal {L} _ {r} (f (\\theta) \\odot M, x + \\delta , y) \\right] \\tag {6} \\\\ + \\varphi \\cdot \\mathcal {L} _ {p} (f (\\theta) \\odot M, a _ {t}) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 849, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_p$ enforces target compression $a_{t}$ across all layers $L$ to guide non-uniform pruning rates.", + "bbox": [ + 511, + 90, + 906, + 122 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Pruning Importance Score", + "text_level": 1, + "bbox": [ + 513, + 132, + 756, + 148 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The importance score's definition varies by pruning method. This section details its definition and computation, as discussed in Sec.3.1.", + "bbox": [ + 511, + 155, + 905, + 199 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neuron Importance Score (NIS) NIS [50] is a common pruning score that minimizes the weighted distance between the original and pruned final responses in a given layer. Given a neural network $f^{(n)}$ with $n$ layers, and the importance score of the final layer response $I^{(n)}$ , the importance score of the $i$ -th layer $I^{(i)}$ is calculated as follows:", + "bbox": [ + 511, + 200, + 905, + 292 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI ^ {(i)} = \\left| w ^ {(i + 1)} \\right| ^ {T} \\cdot \\left| w ^ {(i + 2)} \\right| ^ {T} \\dots \\left| w ^ {(n)} \\right| ^ {T} \\cdot I ^ {(n)} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 303, + 905, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $w^{(i)}$ refers to the weight matrix of the $i$ -th layer. In this case, the importance of a neuron in the network can be computed recursively along the network. Thus, $I^{(i)}$ can be propagated from the importance score of the $(i + 1)$ -th layer as follow:", + "bbox": [ + 511, + 337, + 905, + 411 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {N I S} ^ {(i)} = \\left| w ^ {(i + 1)} \\right| ^ {T} \\cdot I ^ {(i + 1)} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 411, + 903, + 433 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Empirical Risk Minimization (ERM) The importance score based on ERM, used in HYDRA[34] and HARP, is defined as follows, based on the weight magnitudes of a pre-trained model.", + "bbox": [ + 511, + 441, + 905, + 501 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {E R M} ^ {(i)} = \\frac {\\eta \\cdot | \\theta^ {(i)} |}{\\max (| \\theta^ {(i)} |)} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 513, + 905, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $I_{ERM}^{(i)}$ represents the importance score for the weights of the $i$ -th layer, $\\theta^{(i)}$ denotes the weight matrix of that layer, and $\\eta$ is a scaling factor determined by the layer's structure and receptive field size.", + "bbox": [ + 511, + 563, + 905, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversarial Saliency Estimation (ASE) MAD [21] computes the ASE for each parameter, removing the least salient ones. Adversarial saliency is estimated via a second-order Taylor expansion, with the saliency score $I_{ASE}$ for each parameter $w^{(i)}$ defined as:", + "bbox": [ + 511, + 626, + 905, + 700 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {A S E} = \\frac {1}{2} \\left[ \\frac {\\partial^ {2} \\mathcal {L}}{\\partial \\left(w ^ {(i)}\\right) ^ {2}} \\right] \\left| w ^ {(i)} \\right| ^ {2} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 714, + 903, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $\\frac{\\partial^2\\mathcal{L}}{\\partial(w^{(i)})^2}$ is the Hessian of the loss function with respect to $w^{(i)}$ , reflecting the impact of parameter changes on the loss, and $|w^{(i)}|^2$ reflects the magnitude and importance of the corresponding parameter.", + "bbox": [ + 511, + 756, + 905, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Batch Normalization Scaling Factor (BNSF) BNAP [43] identifies the magnitude of the Batch Normalization (BN) scaling factor $\\gamma$ as a robustness measure across channels, used as the pruning criterion. The BN layer normalizes weights $w_{ik}^{j}$ into the following effective weights:", + "bbox": [ + 511, + 825, + 905, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9698", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f29e9f634c90f9504fc5a3607953c9fcfe767ca235f95fd6361906a96313ed95.jpg", + "image_caption": [ + "Clean" + ], + "image_footnote": [ + "(a) Importance score of models with different inputs." + ], + "bbox": [ + 96, + 97, + 181, + 186 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d37ec6b339baabc2e9253d3599ee026f441f7201f270abb0495a89d8dfca76d0.jpg", + "image_caption": [ + "Attack" + ], + "image_footnote": [], + "bbox": [ + 197, + 99, + 281, + 186 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5dc2ad8b079621336b4c10967e0de735dca9a74038f9c54833465f715dce03b4.jpg", + "image_caption": [ + "AT" + ], + "image_footnote": [], + "bbox": [ + 303, + 99, + 387, + 186 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/19b4f980a11e32b48926026ccdfdde77504c4ba7fd8131964e9fb88b2655a1b3.jpg", + "image_caption": [ + "NIS" + ], + "image_footnote": [ + "(b) Importance score of a model with different scoring metrics." + ], + "bbox": [ + 408, + 99, + 491, + 186 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/119d785a023bdc73f48944695313f05eebc553d09226bafb5bb78ed28c92dbc8.jpg", + "image_caption": [ + "ERM" + ], + "image_footnote": [], + "bbox": [ + 511, + 99, + 594, + 186 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cb9442754b7db8f627c132e968293ccfd02dce31a9f95617a085aac4cbf33004.jpg", + "image_caption": [ + "ASE" + ], + "image_footnote": [], + "bbox": [ + 614, + 99, + 697, + 186 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/24a63754435d4f7cf3449fe6efd3eb9f512c4ed00224f8f29374362324fa72a2.jpg", + "image_caption": [ + "BNSF" + ], + "image_footnote": [], + "bbox": [ + 717, + 99, + 800, + 186 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bcf2920daf0c2c82ef10b0a1149d4593dd9f8c0cf51a6d2a814de4f535151ffc.jpg", + "image_caption": [ + "Figure 2. Importance score visualization for the 4-5-6-3 CNN model on the MNIST dataset: yellow/red indicate higher importance, green/blue indicate lower." + ], + "image_footnote": [], + "bbox": [ + 826, + 99, + 875, + 202 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {w} _ {i k} ^ {j} = \\frac {\\gamma_ {i} ^ {j} w _ {i k} ^ {j}}{\\left(\\sigma_ {i} ^ {j}\\right) ^ {2} + \\epsilon} \\approx \\frac {\\gamma_ {i} ^ {j} w _ {i k} ^ {j}}{\\sigma_ {i} ^ {j}} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 301, + 482, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $i$ is the layer index, $j$ is the channel index of $i$ -th layer, and $k$ is the channel index in the $(i - 1)$ -th layer. In this case, $\\epsilon$ is small enough to be neglected, and $\\gamma_i^j$ and $\\sigma_i^j$ each play a role in readjusting the weights per channel. Therefore, the computed weights can be used as the importance score.", + "bbox": [ + 89, + 348, + 483, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. A Closer Look at Adversarial Pruning via Importance Score", + "text_level": 1, + "bbox": [ + 89, + 452, + 482, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Least Importance Score [50] is a common pruning criterion that assesses each parameter's (weights, filters, channels) importance to overall network performance, pruning the least important first. This score clarifies the impact of adversarial examples on pruning. We compared importance scores between models trained on clean data, adversarial data, and adversarily-trained models using a simple CNN architecture.", + "bbox": [ + 89, + 491, + 483, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 2a, importance scores differ significantly between models trained on clean versus adversarial datasets. In the AT model, trained on both, scores do not consistently fall between clean and adversarial values but more closely resemble the clean model. Notably, nodes highly important in both cases (e.g., the second node in the second layer) may appear less important in the AT model. This suggests that training on both datasets may obscure importance, potentially leading to pruning of critical nodes or connections.", + "bbox": [ + 89, + 612, + 483, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Moreover, we evaluated the same model with different importance scoring methods mentioned earlier. The results in Fig.2b suggest that the order of importance for nodes and connections can vary depending on the scoring method. In other words, each sub-network pruned with a different importance scoring method may contain important information that is not shared by the others. This implies that sub-models sharing the same base model could act as multiple models in ED, while complementing each other.", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Efficient Ensemble Defense (EED)", + "text_level": 1, + "bbox": [ + 511, + 287, + 823, + 304 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A common hypothesis in adversarial ensemble defense is that training and combining multiple base models enhances adversarial robustness compared to a single model [7]. However, traditional ED methods often increase model capacity, which is counterproductive for reducing computational and memory costs. But is it truly impossible to achieve an efficient adversarial ensemble defense for a compact model?", + "bbox": [ + 511, + 315, + 903, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose a novel Efficient Ensemble Defense (EED) method that leverages model compression while maintaining robustness. Sec.3.3 has demonstrated the potential for sub-models, pruned based on different importance scores from a single base model, to act as multiple base models in an ensemble. By ensembling these sub-models, we can reduce model capacity compared to traditional ED. However, it is observed that the smaller the model during ensemble, the more the robustness performance decreases; thus, high diversity is required to mitigate this issue[30]. Therefore, we enhance the ensemble effect and maintain robustness through robust diversity evaluation and misclassification regularization. Furthermore, to improve inference speed—a primary goal of model compression—we dynamically add sub-models during inference and perform ensemble only as needed, thus enhancing inference efficiency.", + "bbox": [ + 511, + 436, + 905, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "EED is designed with three main considerations: 1) the diversity of pruned sub-models, 2) improvement of robustness during the ensemble process, and 3) maintaining compactness throughout the ensemble process.", + "bbox": [ + 511, + 680, + 905, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Enhancing Robust Diversity", + "text_level": 1, + "bbox": [ + 511, + 756, + 764, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our strategy begins with generating sub-models from a base model through pruning, where each sub-model undergoes an AP process based on a unique importance score. This AP adheres to the principles described in Sec.3.1, utilizing four importance scoring methods detailed in Sec.3.2. To further enhance diversity, we divide the training dataset for pruning into multiple subsets, with each subset designated to train a specific sub-model. Some subsets are shared across all", + "bbox": [ + 511, + 780, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9699", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Efficient Ensemble Defense (EED)" + ], + "code_body": "Input: Training data $D$ , base model $M$ \nOutput: Ensemble Selected \nfor each subset $D_{i}\\subset D$ do $F_{i}\\gets$ AdversarialPruning $(M,D_{i})$ Set $(D)\\leftarrow \\operatorname {Set}(D)\\cup F_i$ \nend for \nEnsSet $\\leftarrow$ FormEnsembleSet(Set(D)) \nfor each ensemble $E\\in \\mathit{ensSet}$ do $RD\\gets$ CalculateRobustDiversity(E) \nend for \nSelected $\\leftarrow$ SelectEnsemble(EnsSet, RD) \nfor each training example $(x,y)\\in D$ do $\\mathcal{L}_{\\mathrm{EED}}(x,y)\\gets \\mathcal{L}_E + \\mathcal{L}_R + \\mathcal{L}_C$ Selected $\\leftarrow$ UpdateModels(Selected, L_EED) \nend for {INFERENCE PHASE} \nfor each test example $x\\in T$ do pred $\\leftarrow$ InitializePredictions(Selected) t $\\leftarrow 1$ while not sufficient robustness do pred $\\leftarrow$ UpdateEnsemble(pred, Selected[t]) t $\\leftarrow t + 1$ qt $\\leftarrow$ CalculateStoppingCriterion(pred) if robustness criterion met then break end if end while \nend for \nOutput pred", + "bbox": [ + 107, + 109, + 480, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sub-models to retain common core knowledge, while the remaining subsets are distributed across different sub-models to ensure each to learn somewhat different aspects of the data. These subsets are randomly reshuffled whenever the importance scoring method changes. This approach amplifies model diversity, as each sub-model becomes specialized in defending against attacks targeting specific patterns or features in the data. Ultimately, we establish a model pool composed of sub-models with varying importance scoring methods or trained on different datasets.", + "bbox": [ + 89, + 580, + 483, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let this be denoted as pool of $N$ models trained on the dataset $D$ , represented as $Set(D) = \\{F_0, \\dots, F_{N-1}\\}$ . The ensemble set $EnsSet$ represents the collection of all possible ensemble teams with sizes $S$ ranging from 2 to $N$ formed from $Set(D)$ . The number of ensemble teams is given by $|EnsSet| = \\sum_{S=2}^{N} \\binom{N}{S} = 2^N - (1 + N)$ .", + "bbox": [ + 89, + 731, + 483, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To quantify the diversity among sub-models within each ensemble team, we define a Robust Diversity (RD) score based on the probability of defense failure for the model. The diversity score relies on two probabilities: $p_{one}$ and $p_{two}$ . $p_{one}$ is the probability that a randomly chosen model", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "fails to defend (viz. the attack succeeds), while $p_{two}$ is the expected probability that two randomly selected models both fail to defend:", + "bbox": [ + 511, + 90, + 906, + 135 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {o n e} = \\sum_ {i = 1} ^ {S} \\frac {S _ {i}}{S} p _ {i}, \\quad p _ {t w o} = \\sum_ {i = 1} ^ {S} \\frac {S _ {i} (i - 1)}{S (S - 1)} p _ {i} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 143, + 906, + 185 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $p_i$ represents the defense failure probability of the $i$ -th model among the total $S$ models. Based on the two probabilities, the RD score is defined by:", + "bbox": [ + 511, + 194, + 906, + 239 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR D = 1 - \\frac {p _ {\\text {t w o}}}{p _ {\\text {o n e}}} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 250, + 906, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This RD metric measures diversity. $RD = 1$ indicates maximum diversity (i.e., even if one model fails in defense, the likelihood of other models succeeding is high), while $RD = 0$ indicates no diversity (i.e., both models fail in defense). The EED framework applies RD-based ensemble selection to form the smallest ensemble from the EnsSet of sub-models with the high RD scores.", + "bbox": [ + 511, + 286, + 906, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Regularization against Misclassification", + "text_level": 1, + "bbox": [ + 511, + 401, + 854, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As in the existing ED technique [47], the classifier is trained by combining the classification loss, regularization $(Reg)$ and diversity $(Div)$ terms. Specifically, for a single training example $(x,y)$ , the ensemble loss function is defined by:", + "bbox": [ + 511, + 422, + 906, + 484 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {E} (x, y) = \\sum_ {i = 1} ^ {N} \\ell \\left(F _ {i} (x), y\\right) + \\alpha R e g (h (x)) \\tag {14} \\\\ + \\beta D i v (h ^ {1} (x), h ^ {2} (x), \\ldots , h ^ {N} (x), y) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 494, + 905, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\alpha$ and $\\beta$ are hyperparameters, and $h^i (x) = (h_1^i (x),\\ldots ,h_C^i (x)), i\\in \\mathbb{N}$ represents the output of the ith classifier. The final output is calculated using the average combiner $h(x) = \\frac{1}{N}\\sum_{i = 1}^{N}h^{i}(x)$ .", + "bbox": [ + 511, + 566, + 905, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To improve the robustness of EED, we employ an additional regularization term that focuses on the probability scores of the most misclassified class, inspired by iGAT [7]. For a given example $(x,y)$ , the proposed term is defined by:", + "bbox": [ + 511, + 628, + 905, + 689 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {R} (x, y) = - \\psi_ {0 / 1} (c (h ^ {1} (x), \\dots , h ^ {N} (x)), y) \\\\ \\cdot \\log \\left(1 - C \\max _ {i = 1} ^ {N} \\max _ {j = 1} ^ {C} h _ {i} ^ {j} (x)\\right) \\tag {15} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 698, + 903, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The error function $\\psi_{0/1}(f, y) = \\frac{1}{|D|} \\sum_{x \\in D} 1 [f_y(x) < \\max_{c=y} f_c(x)]$ returns 0 if $f$ correctly predicts the label $y$ , and 1 otherwise. This regularization term plays a crucial role in guiding the model to reduce misclassifications, especially improving its ability to handle ambiguous data points or difficult input examples. By doing so, each base classifier focuses more on the examples it misclassified, enhancing the robustness of the overall ensemble model across diverse input spaces.", + "bbox": [ + 511, + 762, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9700", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Regularization for Compactness", + "text_level": 1, + "bbox": [ + 89, + 90, + 375, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Another key objective of the EED framework is maintaining the compressed models' compactness while maximizing the ensemble merit. To achieve this, we design a compactness regularization term $\\mathcal{L}_C$ to maintain weight and activation sparsity among the sub-models, thereby improving the compactness of the ensemble. The weight regularization term controls the sparsity of the sub-model weights, ensuring that they remain sparse above a certain threshold, thus suppressing model capacity growth. Specifically, the weight sparsity regularization term $\\mathcal{L}_W$ is defined as follows:", + "bbox": [ + 89, + 112, + 483, + 263 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {W} (x, y) = \\sum_ {i = 1} ^ {N} \\| W _ {i} \\| _ {1}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 273, + 483, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $\\| W_i\\| _1$ denotes the $L_{1}$ -norm of the weights $W_{i}$ of the $i$ -th sub-model, which encourages sparsity by minimizing the sum of the absolute values of the weights.", + "bbox": [ + 89, + 325, + 483, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To control the activation sparsity of the sub-models, we use an additional sparsity regularization term $\\mathcal{L}_A$ , which encourages each sub-model to reduce unnecessary neuron activations. This term is defined as follows:", + "bbox": [ + 89, + 369, + 483, + 431 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {A} (x, y) = \\sum_ {i = 1} ^ {N} \\| h ^ {i} (x) \\| _ {0}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 439, + 482, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $\\| h^{i}(x)\\|_{0}$ represents the sum of the number of nonzero activations in the output $h^i (x)$ of the $i$ -th sub-model.", + "bbox": [ + 89, + 491, + 482, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Thus, $\\mathcal{L}_C$ can be summarized as follows:", + "bbox": [ + 109, + 521, + 383, + 536 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C} (x, y) = \\lambda_ {1} \\mathcal {L} _ {W} + \\lambda_ {2} \\mathcal {L} _ {A} \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 547, + 482, + 564 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The hyperparameter $\\lambda_{1}$ regulates weight sparsity, with higher values leading to more sparse weights and reduced model capacity. Similarly, $\\lambda_{2}$ controls activation sparsity, making neuron activations sparser, which results in a more compact model structure that saves memory and computational resources.", + "bbox": [ + 89, + 574, + 483, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, the overall loss function of EED is as follows:", + "bbox": [ + 109, + 665, + 470, + 680 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {E E D} (x, y) = \\mathcal {L} _ {E} (x, y) + \\omega \\mathcal {L} _ {R} (x, y) + \\gamma \\mathcal {L} _ {C} (x, y) \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 691, + 482, + 708 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The EED framework provides high robustness against adversarial attacks while ensuring compactness through these multiple regularization terms, making it suitable to efficient applications like real-time inference.", + "bbox": [ + 89, + 718, + 483, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Dynamic Ensemble for Inference Efficiency", + "text_level": 1, + "bbox": [ + 89, + 787, + 460, + 804 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the training phase, EED optimizes all sub-models to ensure diversity and robustness. To enhance the compactness and efficiency of EED during the inference phase, we adopt an inference-efficient ensemble method [25]. We apply this to adversarial defense by dynamically constructing an ensemble of varying sizes for each sample, termed Dynamic", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference Ensemble (DIE). Specifically, for a given sample $x$ , EED dynamically determines the optimal point to stop while adding each sub-models progressively. As a result, by monitoring the uncertainty of each sub-model on adversarial examples and only introducing additional models when necessary, we avoid unnecessary computations while maintaining high robustness.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DIE sequentially adds sub-models and evaluates the robustness of the predictions at each stage. Starting with the first sub-model $F_{1}$ , a new sub-model $F_{t}$ is added at each stage $t$ , and the intermediate ensemble prediction $\\hat{y}_{ens}^{t}$ is updated. These intermediate predictions serve as a criterion for assessing how robust the added sub-models are. The intermediate ensemble prediction based on the predictions $\\hat{y}_i$ of each sub-model $F_{i}$ is computed by:", + "bbox": [ + 511, + 196, + 905, + 318 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} _ {\\text {e n s}} ^ {t} = \\frac {1}{t} \\sum_ {i = 1} ^ {t} \\hat {y} _ {i} \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 327, + 903, + 367 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, each $\\hat{y}_{\\text{ens}}^t$ represents the combined prediction of the sub-models for the input (both clean and adversarial examples), and it plays a role in strengthening the overall robustness of the ensemble.", + "bbox": [ + 511, + 378, + 903, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At each stage $t$ , the early stopping probability $q_{t}$ is assessed. If the current ensemble performance is determined to be robust enough, the addition of sub-models is halted. The value of $q_{t}$ is computed based on the prediction uncertainty and confidence. The prediction uncertainty is computed by KL-divergence of $t$ and $(t - 1)$ -th ensemble predictions, and the prediction confidence setting follows WoC[40]. The optimal stopping point $z_{x}$ is defined as follows, based on a geometric distribution modeling with discrete Bernoulli trials:", + "bbox": [ + 511, + 438, + 905, + 589 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nz _ {x} = \\arg \\max _ {t} \\left(q _ {t} \\prod_ {i = 1} ^ {t - 1} \\left(1 - q _ {i}\\right)\\right) \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 595, + 903, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $q_{t}$ is dynamically adjusted at each stage based on robustness, and sub-models are added only when necessary, allowing for efficient use of computational resources.", + "bbox": [ + 511, + 646, + 903, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The final ensemble prediction $\\hat{y}_{ens}$ is calculated as the average of the predictions of all sub-models up to the stopping point $z_{x}$ :", + "bbox": [ + 511, + 691, + 903, + 736 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} _ {e n s} = \\frac {1}{z _ {x}} \\sum_ {i = 1} ^ {z _ {x}} \\hat {y} _ {i} \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 648, + 737, + 903, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This approach takes into account the predictions of all submodels while avoiding the inclusion of unnecessary ones, thereby enhancing efficiency. By applying different ensemble sizes for each sample, the overall computational cost is reduced while still being robust against adversarial attacks.", + "bbox": [ + 511, + 779, + 905, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DIE finds a balance between efficiency and robustness, quickly identifying the optimal combination of sub-models to ensure strong performance even in real-time inference", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9701", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/13dda296c696f0828469014759fa4714b7eaa9d4faea483274f6d3b93cbf0af0.jpg", + "table_caption": [ + "Table 1. Comparison of various AP methods and EED on CIFAR-10 and SVHN datasets when sparsity rate ${s}_{r} = {80}\\%$" + ], + "table_footnote": [], + "table_body": "
SettingCIFAR-10SVHNPrams
CleanPGDAAC&WDFSpeed upCleanPGDAAC&WDFSpeed up
Resnet 18AT[27]87.0556.1448.0257.6053.101.00x93.3756.2750.1458.8559.801.00x11.2M
R-ADMM[48]81.2548.0043.9249.1739.111.68x74.8149.7337.4052.6243.401.63x
HYDRA[34]77.3652.9243.7449.6445.911.77x91.0652.2247.6255.1346.131.74x
RST[10]61.0241.0118.3851.0226.821.86x82.3946.2936.3552.6539.271.83x
MAD[21]74.1850.3841.2754.1738.441.69x92.8451.6539.8759.8044.721.70x
Flying Bird[2]81.0751.6244.4156.0845.291.79x90.2152.0642.0157.3052.231.76x2.2M
HARP[52]83.8452.5645.3656.5747.041.81x92.6054.1645.8957.2851.241.80x
TwinRep[23]77.2652.0443.5256.6046.181.82x92.9653.8344.7356.2648.611.78x
EED(ours)86.1355.7148.1357.0351.971.86x93.1555.7450.1858.3756.051.85x
VGG 16AT[27]82.7054.4948.5254.9156.911.00x93.0657.6452.2855.3661.491.00x14.7M
R-ADMM[48]76.5147.1940.4144.8048.291.65x65.4048.3944.6448.9249.851.66x
HYDRA[34]78.8347.8544.8545.7349.411.72x90.8253.6646.3052.6353.421.72x
RST[10]77.2939.9030.1532.0242.031.83x82.3941.9445.5243.5838.811.81x
MAD[21]73.6548.8140.8847.1645.501.80x89.9449.4640.3153.0347.371.69x
Flying Bird[2]77.3649.4345.7350.3848.341.68x90.1153.2949.1852.8150.781.73x2.9M
HARP[52]80.9251.7744.0252.7952.171.73x93.2255.5344.7352.9656.161.77x
TwinRep[23]76.0450.1243.2652.4251.591.79x92.6054.6149.0651.2456.581.75x
EED(ours)81.3954.2647.4953.2753.941.79x93.1457.1951.4953.1260.271.77x
", + "bbox": [ + 91, + 114, + 923, + 366 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "settings. As a result, the EED framework can lower the average inference cost in adversarial attack scenarios while maintaining high level of robustness, making it an effective defense, even in resource-constrained environments.", + "bbox": [ + 89, + 385, + 482, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experimental Evaluations", + "text_level": 1, + "bbox": [ + 89, + 460, + 333, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Experiment Setups", + "text_level": 1, + "bbox": [ + 89, + 484, + 272, + 501 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Classifiers and Datasets We utilize two specific architectures, ResNet18 and VGG16, as classifiers and conduct experiments focusing on CIFAR-10 and SVHN. We divide the dataset into $d$ subsets for pruning into various sub-models, setting $d = 4$ to create three sub-models for each pruning metric. $25\\%$ of the dataset is shared across all sub-models, while the remaining $75\\%$ is allocated separately for each of the three sub-models. With four pruning metrics in play, we end up with a total of $N = 12$ sub-models, which together form the ensemble set, denoted as EnsSet. For ensemble learning, the entire dataset is used.", + "bbox": [ + 89, + 507, + 482, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Attacks We use PGD-20 attacker with a step size of 2/255 to generate adversarial examples for training, and Auto Attack(AA)[6], C&W[1] and Deepfool(DF)[28] for evaluation. Except for DF using $l_{2}$ norm, the rest use $l_{\\infty}$ norm, and all attacks use perturbation strength $\\epsilon = 8 / 255$ .", + "bbox": [ + 89, + 672, + 482, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training We use the Madry[27] as AT method. The batch size is set to 512, and the multi-step learning rates are \\{0.01, 0.002\\} for CIFAR-10 and \\{0.1, 0.02, 0.004\\} for SVHN. The hyperparameters are set as follows: $\\omega = 10$ , $\\lambda_{1} = 0.7$ , $\\lambda_{2} = 0.25$ , and the RD threshold is set to 0.7. $c_{r}$ is set to be more than four times sparser than the desirable $s_{r}$ , considering the size of $N$ , and $\\gamma$ is adjusted to ensure that the $s_{r}$ is met even after the ensemble. (When $s_{r} = 80\\%$ , we set $c_{r} = 95\\%$ and $\\gamma = 4$ .) Each experiment is run on an NVIDIA RTX A6000 GPU with 8 CPU cores.", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Result Comparison and Analysis", + "text_level": 1, + "bbox": [ + 511, + 383, + 802, + 401 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This section compares the proposed EED with existing APs for sparsity rate $(s_r)$ of $80\\%$ on the CIFAR-10 and SVHN datasets. Table 1 demonstrates that EED consistently provided superior adversarial robustness, with significantly higher accuracy against attacks such as PGD, AA, C&W, and DF compared to other methods. Additionally, EED maintained high performance on clean data, indicating a proper balance between adversarial robustness and clean accuracy.", + "bbox": [ + 511, + 410, + 906, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that EED demonstrates superior robustness, particularly against strong attacks such as AA and PGD, outperforming existing defense methods like HYDRA, HARP, and TwinRep. For example, in the CIFAR-10 experiment using the ResNet-18 architecture, EED achieved $55.71\\%$ accuracy against PGD and $48.13\\%$ against AA, showing an improvement of over $2\\%$ compared to other methods and achieving the highest robustness. Moreover, despite using only $20\\%$ of the model's parameters, the performance drop in AT was minimal, less than $0.5\\%$ . Also, EED showed stronger defense capabilities than AT against PGD in the ResNet-18-based SVHN experiment.", + "bbox": [ + 511, + 549, + 906, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As a result, EED accomplished remarkable performance improvements without sacrificing efficiency. Notably, it also records up to 1.86 times faster inference speeds, comparable to or exceeding existing techniques. This is crucial for practical applications where computational resources may be limited, demonstrating that EED is an ideal defense method for resource-constrained environments. Furthermore, EED showed consistent performance across different model architectures such as ResNet-18 and VGG-16, suggesting that EED can be applied to a wide range of network architectures with minimal adjustment.", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9702", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/36788b395246e5fe5c137d24f0967a931957007def8242751b45ec680fea6db3.jpg", + "image_caption": [ + "Figure 3. Robust performance by $s_r$ ." + ], + "image_footnote": [], + "bbox": [ + 98, + 95, + 313, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/253520b5df4c39fe39333c8d1f7a28a2e60ec467abcf044af2f7b96f56edbd1b.jpg", + "image_caption": [ + "(a) RD analysis on ensemble team size." + ], + "image_footnote": [], + "bbox": [ + 362, + 90, + 617, + 253 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a23a51dd1ba532a789bfdc6c567670c2a02030253dc94f10caee63acb05550ab.jpg", + "image_caption": [ + "(b) RD analysis on sub-model sparsity.", + "Figure 4. Correlation between RD and robustness in different ensemble components." + ], + "image_footnote": [], + "bbox": [ + 635, + 92, + 890, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 315, + 254, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation studies are conducted on Resnet18 and CIFAR-10. Changes along Sparsity Fig.3 shows the change in robustness performance with respect to the sparsity rate. EED demonstrated the highest performance compared to the existing top three methods, except for a very high $s_r$ (95%) against AA. EED tended to show a decrease in performance as $s_r$ increases, which is believed to be due to the reduction in ensemble diversity caused by the higher compression rate. On the other hand, at lower $s_r$ , EED exhibited higher robustness than AT due to the enhanced ensemble diversity. More detailed experimental results for different $s_r$ can be found in the supplementary materials.", + "bbox": [ + 88, + 338, + 482, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analyzing Robust Diversity Fig.4 analyzes the role of RD in EED. Fig.3a, which shows robustness according to ensemble team size, demonstrates that a high RD is necessary for small ensemble sizes to achieve high performance. Fig.3b shows robustness from the RD perspective with respect to the compression rate $(c_r)$ of the sub-models. We can find that to maintain high performance while keeping sub-models small, a high RD is also required.", + "bbox": [ + 88, + 521, + 482, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analyzing Regularization Components EED has two regularization terms, as mentioned in Sec.4.2 and Sec.4.3. Regularization against misclassification $(L_{R})$ is controlled by $\\omega$ , while regularization for compactness $(L_{C})$ is controlled by $\\gamma$ . This section analyzes the impact of each term on the performance of EED. Tab.2 shows that $L_{R}$ has a greater effect on clean performance, while $L_{C}$ has a larger impact on attack performance. In particular, when the compactness term is not used, attack performance significantly deteriorates, which is likely due to the need of reducing the ensemble size further to meet the $s_{r}$ , thereby diminishing the ensemble advantages.", + "bbox": [ + 88, + 643, + 482, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Dynamic Inference Ensemble Tab.3 analyzes the effect of DIE for different sparsity rates. At $s_r = 80\\%$ , DIE had little impact on the main performance and only showed a slight speed improvement. On the other hand, at $s_r = 50\\%$ , activating DIE resulted in performance im", + "bbox": [ + 88, + 824, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/250d644cfb6ced5c5858cc3843708853db21f9427ffb7a790ca6cd9671f0a044.jpg", + "table_caption": [ + "Table 2. Analysis of EED with regularization terms." + ], + "table_footnote": [], + "table_body": "
SettingsCleanPGDAA
CIFAR-10EED86.1355.7148.13
ω = 074.1647.4536.37
Resnet 18γ = 076.3339.1031.59
ω,γ = 073.9537.8328.94
", + "bbox": [ + 566, + 335, + 854, + 409 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7e2cba1cf25af6abe2277ad5abeddd5d026b3c02e09e4e20979685d183be9759.jpg", + "table_caption": [ + "Table 3. Correlation between dynamic inference ensemble and ${s}_{r}$ ." + ], + "table_footnote": [], + "table_body": "
SettingsCleanPGDAASpeed up
sr=80%w/ dynamic86.1355.7148.131.86x
w/o dynamic86.2055.6248.111.80x
sr=50%w/ dynamic86.1556.0952.351.64x
w/o dynamic85.9354.2850.761.27x
", + "bbox": [ + 522, + 444, + 897, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "provements of $1.81\\%$ and $1.59\\%$ for PGD and AA, respectively, while significantly improving inference speed. This suggests that DIE is more effective at lower sparsity. The reason for this performance improvement is that, with lower sparsity, the ensemble can be larger, thus enhancing the optimization effect of the dynamic ensemble. In contrast, without DIE, the likelihood of conflicts between the sub-models may increase as the ensemble grows larger.", + "bbox": [ + 511, + 542, + 905, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 678, + 633, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes the Efficient Ensemble Defense (EED) technique to maintain both compression and robustness of deep learning models. EED constructs an ensemble of sub-models using different pruning importance scores to enhance adversarial robustness while reducing model size and computational cost. By addressing information loss and limited model diversity caused by pruning with ensemble methods and regularization, EED outperforms existing AP methods in accuracy and robustness. Experimental results show superior resilience of EED against various attacks on the CIFAR-10 and SVHN, maintaining inference efficiency. EED also proves its applicability in diverse neural network architectures and resource-constrained environments.", + "bbox": [ + 509, + 703, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9703", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.RS-2022-00155915 (Artificial Intelligence Convergence Innovation Human Resources Development (Inha University)) and No. 2021-0-02068 (AI Innovation Hub)), and this work was partly supported by the National Research Foundation of Korea (NRF) grant funded by the MSIT (No. 2022R1A2C2010095).", + "bbox": [ + 89, + 113, + 483, + 239 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 267, + 186, + 282 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE symposium on security and privacy (sp), pages 39-57. IEEE, 2017. 2, 7", + "[2] Tianlong Chen, Zhenyu Zhang, Pengjun Wang, Santosh Balachandra, Haoyu Ma, Zehao Wang, and Zhangyang Wang. Sparsity winning twice: Better robust generalization from more efficient training. arXiv preprint arXiv:2202.09844, 2022. 7", + "[3] Wenlin Chen, James Wilson, Stephen Tyree, Kilian Weinberger, and Yixin Chen. Compressing neural networks with the hashing trick. In International conference on machine learning, pages 2285-2294. PMLR, 2015. 1", + "[4] Wuyang Chen, Ziyu Jiang, Zhangyang Wang, Kexin Cui, and Xiaoning Qian. Collaborative global-local networks for memory-efficient segmentation of ultra-high resolution images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8924-8933, 2019. 1", + "[5] Hao Cheng, Kaidi Xu, Zhengang Li, Pu Zhao, Chenan Wang, Xue Lin, Bhavya Kailkhura, and Ryan Goldhahn. More or less (mol): Defending against multiple perturbation attacks on deep neural networks through model ensemble and compression. In 2022 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW), pages 645-655. IEEE, 2022. 2", + "[6] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 2, 7", + "[7] Yian Deng and Tingting Mu. Understanding and improving ensemble adversarial defense. NeurIPS, 36, 2024. 2, 4, 5", + "[8] Misha Denil, Babak Shakibi, Laurent Dinh, Marc'Aurelio Ranzato, and Nando De Freitas. Predicting parameters in deep learning. Advances in neural information processing systems, 26, 2013. 1", + "[9] Guneet S Dhillon, Kamyar Azizzadenesheli, Zachary C Lipton, Jeremy Bernstein, Jean Kossaifi, Aran Khanna, and Anima Anandkumar. Stochastic activation pruning for robust adversarial defense. arXiv preprint arXiv:1803.01442, 2018. 2", + "[10] Yonggan Fu, Qixuan Yu, Yang Zhang, Shang Wu, Xu Ouyang, David Cox, and Yingyan Lin. Drawing robust scratch tickets: Subnetworks with inborn robustness are" + ], + "bbox": [ + 93, + 292, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "found within randomly initialized networks. Advances in Neural Information Processing Systems, 34:13059-13072, 2021. 7", + "[11] Ian Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In International Conference on Learning Representations, 2015. 1, 2", + "[12] Yiwen Guo, Chao Zhang, Changshui Zhang, and Yurong Chen. Sparse dnns with improved adversarial robustness. Advances in neural information processing systems, 31, 2018. 2", + "[13] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. 2", + "[14] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. Advances in neural information processing systems, 28, 2015. 1", + "[15] Warren He, James Wei, Xinyun Chen, Nicholas Carlini, and Dawn Song. Adversarial example defense: Ensembles of weak defenses are not strong. In 11th USENIX workshop on offensive technologies (WOOT 17), 2017. 2", + "[16] Tong Jian, Zifeng Wang, Yanzhi Wang, Jennifer Dy, and Stratis Ioannidis. Pruning adversarially robust neural networks without adversarial examples. In 2022 IEEE International Conference on Data Mining (ICDM), pages 993-998. IEEE, 2022. 2", + "[17] Sanjay Kariyappa and Moinuddin K Qureshi. Improving adversarial robustness of ensembles with diversity training. arXiv preprint arXiv:1901.09981, 2019. 2", + "[18] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 1", + "[19] Souvik Kundu, Mahdi Nazemi, Peter A Beerel, and Massoud Pedram. Dnr: A tunable robust pruning framework through dynamic network rewiring of dnns. In Proceedings of the 26th Asia and South Pacific Design Automation Conference, pages 344–350, 2021. 2", + "[20] Alexey Kurakin, Ian Goodfellow, and Samy Bengio. Adversarial machine learning at scale. arXiv preprint arXiv:1611.01236, 2016. 1, 2", + "[21] Byung-Kwan Lee, Junho Kim, and Yong Man Ro. Masking adversarial damage: Finding adversarial saliency for robust and sparse network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15126-15136, 2022. 3, 7", + "[22] Bai Li, Shiqi Wang, Yunhan Jia, Yantao Lu, Zhenyu Zhong, Lawrence Carin, and Suman Jana. Towards practical lottery ticket hypothesis for adversarial training. arXiv preprint arXiv:2003.05733, 2020. 2", + "[23] Chenhao Li, Qiang Qiu, Zhibin Zhang, Jiafeng Guo, and Xueqi Cheng. Learning adversarially robust sparse networks via weight reparameterization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8527-8535, 2023. 7", + "[24] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. arXiv preprint arXiv:1608.08710, 2016. 1" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9704", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Ziyue Li, Kan Ren, Yifan Yang, Xinyang Jiang, Yuqing Yang, and Dongsheng Li. Towards inference efficient deep ensemble learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8711-8719, 2023. 2, 6", + "[26] Ningyi Liao, Shufan Wang, Liyao Xiang, Nanyang Ye, Shuo Shao, and Pengzhi Chu. Achieving adversarial robustness via sparsity. Machine Learning, pages 1-27, 2022. 2", + "[27] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 1, 2, 7", + "[28] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2574–2582, 2016. 7", + "[29] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Anima Anandkumar. Diffusion models for adversarial purification. arXiv preprint arXiv:2205.07460, 2022. 2", + "[30] Tianyu Pang, Kun Xu, Chao Du, Ning Chen, and Jun Zhu. Improving adversarial robustness via promoting ensemble diversity. In International Conference on Machine Learning, pages 4970-4979. PMLR, 2019. 2, 4", + "[31] Nicolas Papernot, Patrick McDaniel, Xi Wu, Somesh Jha, and Ananthram Swami. Distillation as a defense to adversarial perturbations against deep neural networks. In 2016 IEEE symposium on security and privacy (SP), pages 582-597. IEEE, 2016. 1", + "[32] Giorgio Piras, Maura Pintor, Ambra Demontis, Battista Biggio, Giorgio Giacinto, and Fabio Roli. Adversarial pruning: A survey and benchmark of pruning methods for adversarial robustness. arXiv preprint arXiv:2409.01249, 2024. 2", + "[33] J Redmon. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2016. 1", + "[34] Vikash Sehwag, Shiqi Wang, Prateek Mittal, and Suman Jana. Hydra: Pruning adversarially robust neural networks. Advances in Neural Information Processing Systems, 33: 19655-19666, 2020. 1, 3, 7", + "[35] Ilia Shumailov, Yiren Zhao, Robert Mullins, and Ross Anderson. To compress or not to compress: Understanding the interactions between adversarial attacks and neural network compression. Proceedings of Machine Learning and Systems, 1:230-240, 2019. 1", + "[36] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. Advances in neural information processing systems, 27, 2014. 1", + "[37] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations, 2014. 1, 2", + "[38] Lukas Timpl, Rahim Entezari, Hanie Sedghi, Behnam Neyshabur, and Olga Saukh. Understanding the effect of sparsity on neural networks robustness. arXiv preprint arXiv:2206.10915, 2022. 2" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Florian Tramér, Alexey Kurakin, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. Ensemble adversarial training: Attacks and defenses. In International Conference on Learning Representations, 2018. 2", + "[40] Xiaofang Wang, Dan Kondratyuk, Eric Christiansen, Kris M Kitani, Yair Movshovitz-Attias, and Elad Eban. Wisdom of committees: An overlooked approach to faster and more accurate models. In International Conference on Learning Representations, 2021. 6", + "[41] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In International conference on learning representations, 2019. 1, 2", + "[42] Yue Wang, Jianghao Shen, Ting-Kuei Hu, Pengfei Xu, Tan Nguyen, Richard Baraniuk, Zhangyang Wang, and Yingyan Lin. Dual dynamic inference: Enabling more efficient, adaptive, and controllable deep inference. IEEE Journal of Selected Topics in Signal Processing, 14(4):623-633, 2020. 1", + "[43] Xiao Wei, Yao Zhu, and Shu-Tao Xia. Batch normalization assisted adversarial pruning: Towards lightweight, sparse and robust models. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2021. 3", + "[44] Boxi Wu, Jinghui Chen, Deng Cai, Xiaofei He, and Quanquan Gu. Do wider neural networks really help adversarial robustness? Advances in Neural Information Processing Systems, 34:7054-7067, 2021. 1", + "[45] Dongxian Wu, Shu-Tao Xia, and Yisen Wang. Adversarial weight perturbation helps robust generalization. Advances in neural information processing systems, 33:2958-2969, 2020. 2", + "[46] Yanzhao Wu, Ling Liu, Zhongwei Xie, Ka-Ho Chow, and Wenqi Wei. Boosting ensemble accuracy by revisiting ensemble diversity metrics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16469-16477, 2021. 2", + "[47] Huanrui Yang, Jingyang Zhang, Hongliang Dong, Nathan Inkawhich, Andrew Gardner, Andrew Touchet, Wesley Wilkes, Heath Berry, and Hai Li. Diverse: diversifying vulnerabilities for enhanced robust generation of ensembles. Advances in Neural Information Processing Systems, 33: 5505-5515, 2020. 2, 5", + "[48] Shaokai Ye, Kaidi Xu, Sijia Liu, Hao Cheng, Jan-Henrik Lambrechts, Huan Zhang, Aojun Zhou, Kaisheng Ma, Yanzhi Wang, and Xue Lin. Adversarial robustness vs. model compression, or both? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 111-120, 2019. 1, 2, 7", + "[49] Jongmin Yoon, Sung Ju Hwang, and Juho Lee. Adversarial purification with score-based generative models. In International Conference on Machine Learning, pages 12062-12072. PMLR, 2021. 2", + "[50] Ruichi Yu, Ang Li, Chun-Fu Chen, Jui-Hsin Lai, Vlad I Morariu, Xintong Han, Mingfei Gao, Ching-Yung Lin, and Larry S Davis. Nisp: Pruning networks using neuron importance score propagation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9194-9203, 2018. 2, 3, 4" + ], + "bbox": [ + 516, + 90, + 905, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9705", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In International conference on machine learning, pages 7472-7482. PMLR, 2019. 1", + "[52] Qi Zhao and Christian Wressnegger. Holistic adversarially robust pruning. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 3, 7" + ], + "bbox": [ + 91, + 90, + 480, + 205 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9706", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Two is Better than One_ Efficient Ensemble Defense for Robust and Compact Models/f8d31d9f-8dff-45eb-9d02-43167ab5a77a_model.json b/2025/Two is Better than One_ Efficient Ensemble Defense for Robust and Compact Models/f8d31d9f-8dff-45eb-9d02-43167ab5a77a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..01a7c176380eaa7a61c2740ef46d01a929447bf5 --- /dev/null +++ b/2025/Two is Better than One_ Efficient Ensemble Defense for Robust and Compact Models/f8d31d9f-8dff-45eb-9d02-43167ab5a77a_model.json @@ -0,0 +1,2675 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.195, + 0.131, + 0.804, + 0.177 + ], + "angle": 0, + "content": "Two is Better than One: Efficient Ensemble Defense for Robust and Compact Models" + }, + { + "type": "text", + "bbox": [ + 0.363, + 0.205, + 0.636, + 0.222 + ], + "angle": 0, + "content": "Yoojin Jung Byung Cheol Song" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.223, + 0.771, + 0.257 + ], + "angle": 0, + "content": "Department of Electrical and Computer Engineering, Inha University Incheon, Republic of Korea" + }, + { + "type": "text", + "bbox": [ + 0.324, + 0.26, + 0.668, + 0.275 + ], + "angle": 0, + "content": "yoon0505@gmail.com, bcsong@inha.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.31, + 0.327, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.343, + 0.485, + 0.645 + ], + "angle": 0, + "content": "Deep learning-based computer vision systems adopt complex and large architectures to improve performance, yet they face challenges in deployment on resource-constrained mobile and edge devices. To address this issue, model compression techniques such as pruning, quantization, and matrix factorization have been proposed; however, these compressed models are often highly vulnerable to adversarial attacks. We introduce the Efficient Ensemble Defense (EED) technique, which diversifies the compression of a single base model based on different pruning importance scores and enhances ensemble diversity to achieve high adversarial robustness and resource efficiency. EED dynamically determines the number of necessary sub-models during the inference stage, minimizing unnecessary computations while maintaining high robustness. On the CIFAR-10 and SVHN datasets, EED demonstrated state-of-the-art robustness performance compared to existing adversarial pruning techniques, along with an inference speed improvement of up to 1.86 times. This proves that EED is a powerful defense solution in resource-constrained environments." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.678, + 0.222, + 0.694 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Convolutional Neural Networks (CNNs) have shown remarkable performance across various computer vision tasks, such as image classification, object detection, and semantic segmentation[18, 33, 36]. However, CNNs have become increasingly wider and deeper to achieve higher performance, leading to complex architectures with vast numbers of parameters. This complexity inherently results in significant computational costs and storage demands, making deployment on resource-constrained edge devices challenging[4, 42]. To address this challenge, a number of model compression techniques, including pruning[14, 24], quantization[3], and matrix decomposition[8], have been proposed to minimize model size and computational load" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.312, + 0.668, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.312, + 0.808, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.8, + 0.355, + 0.877, + 0.404 + ], + "angle": 0, + "content": "Important nodes or connections may be removed incorrectly due to adversarial attack" + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.424, + 0.708, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.719, + 0.476, + 0.774, + 0.507 + ], + "angle": 0, + "content": "Ensemble could help compensate" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.539, + 0.908, + 0.624 + ], + "angle": 0, + "content": "Figure 1. Challenges in AP and the potential of ensemble methods: AP weakens the network by removing critical parameters, reducing robustness and increasing vulnerability to adversarial attacks. An ensemble approach mitigates pruning effects by combining pruned models, enhancing robustness through their collective strength (illustrated in green)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.655, + 0.733, + 0.671 + ], + "angle": 0, + "content": "without compromising accuracy." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Meanwhile, deep learning models are highly vulnerable to adversarial attacks that induce false predictions by disturbing the input image with small perturbations[11, 37]. This vulnerability is particularly pronounced in sparsely compressed models, which are even more sensitive to adversarial attacks than standard models[35, 44]. Consequently, common adversarial defense techniques such as adversarial training (AT) [20, 27], label smoothing [41, 51], and defensive distillation[31], while generally effective, often experience a degradation in performance when applied to compressed models. This indicates that existing compression techniques may compromise adversarial robustness even if they maintain accuracy on clean test sets. In response, adversarial pruning (AP), a model compression technique that considers adversarial robustness[34, 48, 52], has emerged;" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9696" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "however, AP still exhibits lower robustness compared to non-compressed defense models[32]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.127, + 0.483, + 0.293 + ], + "angle": 0, + "content": "We assume that this degradation in robustness arises from reduced generalization capability and information loss caused by pruning. Specifically, adversarial attacks may cause the model to misidentify and even remove critical nodes or connections during the pruning process. In Sec.3.3, we validate this assumption by observing importance scores, which reflect how each component contributes to the loss function through the gradients of the weights[50]. Our findings demonstrate that adversarial examples significantly impact the importance scores of individual parameters." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.298, + 0.485, + 0.509 + ], + "angle": 0, + "content": "Ensemble defense[7, 39] has recently gained attention in adversarial defense research for enhancing the stability and robustness of multiple base classifiers against attacks. As illustrated in Fig.1, we believe that ensemble defense methods can help mitigate generalization degradation and information loss during the pruning process. Since ensemble learning can compensate for deficiencies in individual compressed models due to removed parameters by other models and offset individual errors by combining outputs. However, traditional ensemble defenses inherently suffer from a large capacity, as they require multiple base models[15, 30]. This raises an important question: Is it possible to develop an ensemble defense algorithm that not only achieves robustness but also results in a compact compressed model?" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.514, + 0.485, + 0.741 + ], + "angle": 0, + "content": "As an answer to this, we propose Efficient Ensemble Defense (EED). Instead of using multiple base models, we compress a single base model by employing different pruning metrics and then utilize these compressed sub-models in the ensemble. This approach reduces the overall capacity compared to traditional ensemble defense models that rely on multiple base models. Given that using smaller models in an ensemble can lead to reduced defensive efficacy[46], we enhance the variety of the compressed models by incorporating a diversity term in the ensemble selection process, thereby improving ensemble effectiveness while maintaining robustness. Additionally, by adopting inference-efficient ensemble[25], we dynamically add sub-networks and perform ensemble operations as needed during inference, improving overall efficiency." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.745, + 0.449, + 0.76 + ], + "angle": 0, + "content": "The main contributions of our work are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.825 + ], + "angle": 0, + "content": "- Propose EED, that overcomes the limitations of traditional AP by enhancing the ensemble diversity and compensating for individual errors, thereby achieving robustness against adversarial attacks while being compact." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.826, + 0.483, + 0.871 + ], + "angle": 0, + "content": "- Reduce memory by using a single base model in ensemble, and decrease computational costs by dynamically adding models only when necessary during inference." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.871, + 0.483, + 0.9 + ], + "angle": 0, + "content": "- Achieve state-of-the-art performance across various attack benchmarks on the CIFAR-10 and SVHN datasets." + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.665, + 0.106 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.116, + 0.796, + 0.131 + ], + "angle": 0, + "content": "2.1. Adversarial Attack and Defense" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.905, + 0.259 + ], + "angle": 0, + "content": "Adversarial attacks exploit the inherent vulnerabilities of deep learning models by adding specific noise to the input, thereby inducing incorrect predictions. These specially crafted inputs, designed to deceive deep learning models, are called adversarial examples. Early studies highlighted the susceptibility of deep learning models to adversarial examples[37]. Then, more potent attacks such as FGSM[11], PGD[27], and AutoAttack[6] have emerged." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.26, + 0.905, + 0.396 + ], + "angle": 0, + "content": "On the other hand, defense methods against attacks have also been evolved. The most well-known adversarial training (AT)[1, 20, 27] enhanced robustness by training adversarial examples. Other strategies involve regularizing geometric properties such as the curvature of the loss function [41, 45], or restoring potentially attacked images to purified versions[29, 49]. Adversarial pruning (AP) and ensemble defense are also one of adversarial defenses (details are given in Sec.3.1.)" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.407, + 0.71, + 0.423 + ], + "angle": 0, + "content": "2.2. Adversarial Pruning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.907, + 0.672 + ], + "angle": 0, + "content": "AP is a pruning technique (primarily weight pruning) that considers the adversarial context, aiming to reduce the complexity of neural networks while preserving robustness against adversarial attacks. HYDRA demonstrated that empirical adversarial robustness can be achieved through a lowest weight magnitude (LWM) pruning method. R-ADMM[48] and HARP[52] incorporated the alternating direction method of multipliers (ADMM) framework into LWM-based pruning to enhance robustness. In some studies [12, 22, 26], a traditional pruning method, i.e., the Lottery Ticket Hypothesis, was integrated with AT. However, they all failed to maintain generalized robustness at higher sparsity rates \\((s_r)[13, 19, 38]\\), and struggled to defend against adaptive attacks[9, 16]. This suggests that conventional AP has not fully resolved the trade-off between model capacity and robustness." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.691, + 0.697 + ], + "angle": 0, + "content": "2.3. Ensemble Defense" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Ensemble defense (ED), which applies ensemble learning to adversarial defense, has been actively studied due to its promising performance[7, 39]. The core of ED is to improve predictions through an ensemble diversity loss that considers adversarial robustness during training. ADP[30] used Shannon entropy and geometric diversity for uncertainty, while GAL[17] utilized cosine similarities between model loss gradients to formalize diversity. DVERGE[47] introduced vulnerability diversity to ensure that each model is robust to the weaknesses of others, and MOL[5] increased perturbation diversity by using expert models trained on different attacks. Thus, ED has significantly enhanced robustness by reducing the risk of overfitting to adversarial ex" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9697" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "amples and diversifying the decision boundary, though with increased model capacity and computational costs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.135, + 0.231, + 0.151 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.16, + 0.204, + 0.173 + ], + "angle": 0, + "content": "3.1. Notations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.181, + 0.484, + 0.318 + ], + "angle": 0, + "content": "Adversarial Attack Let \\( x \\in \\mathbb{R}^d \\) be an input example, \\( y \\in \\{1, \\dots, C\\} \\) the true label, and \\( f(\\cdot) \\) a classifier with parameters \\( \\theta \\). The classifier's prediction for input \\( x \\) is given by \\( f(x) = \\arg \\max_{i \\in \\{1, \\dots, C\\}} f_i(x) \\), where \\( f_i(x) \\) denotes the predicted probability for class \\( i \\). An adversarial example \\( x' \\) is crafted by adding a perturbation \\( \\delta \\) to \\( x \\), aiming to maximize the model's prediction error while keeping \\( x' \\) close to \\( x \\) under a norm constraint. Formally, an adversarial attack is defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.327, + 0.483, + 0.352 + ], + "angle": 0, + "content": "\\[\nx ^ {\\prime} = x + \\delta = \\arg \\max _ {\\| x - x ^ {\\prime} \\| < \\epsilon} \\mathcal {L} (f (x ^ {\\prime}), y) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.361, + 0.484, + 0.391 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}\\) is the loss function and \\(\\epsilon\\) is the perturbation budget that constrains the strength of the attack under an \\(\\ell_p\\)-norm." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.392, + 0.484, + 0.451 + ], + "angle": 0, + "content": "Adversarial Training AT is a defense method that improves model robustness by training on adversarially perturbed samples alongside clean samples. The objective for AT is usually formulated by:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.46, + 0.483, + 0.495 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {(x, y) \\sim \\mathcal {D}} \\left[ \\max _ {\\delta : \\| \\delta \\| _ {p} \\leq \\epsilon} \\mathcal {L} \\left(f \\left(x ^ {\\prime}; \\theta\\right), y\\right) \\right] \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.502, + 0.308, + 0.516 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}\\) is the data distribution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.517, + 0.484, + 0.607 + ], + "angle": 0, + "content": "Adversarial Pruning Given a pre-trained model \\( f(\\theta) \\) under AT, we aim to maintain robustness while applying a pruning mask \\( M \\) across layers to achieve compression. The general pruning is carried out by removing parameters until the network reaches a desired \\( s_r \\). So, pruning involves creating \\( M \\), formalized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.618, + 0.483, + 0.643 + ], + "angle": 0, + "content": "\\[\nM = \\arg \\min _ {\\| \\mathbf {m} \\| _ {0} \\leq 1 - s _ {r}} \\mathcal {L} (f (\\theta) \\otimes m, x, y). \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.652, + 0.483, + 0.692 + ], + "angle": 0, + "content": "\\[\nm = \\left\\{ \\begin{array}{l l} 0, & \\text {i f} \\theta_ {i} \\text {i s t o b e p r u n e d} \\\\ 1, & \\text {o t h e r w i s e} \\end{array} , \\quad \\forall \\theta_ {i} \\in \\theta \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.696, + 0.484, + 0.755 + ], + "angle": 0, + "content": "For AP, we follow the strategy from HARP[52]. To enhance adversarial robustness, inner maximization generates adversarial examples, targeting a worst-case loss under the PGD attacks:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.756, + 0.483, + 0.778 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\delta} \\mathcal {L} _ {r} (f (\\theta) \\odot M, x + \\delta , y) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.782, + 0.484, + 0.843 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_r\\) defines the robust loss function, incorporating attack-specific parameters. And the global non-uniform pruning strategy jointly optimizes a compression rate \\(c_{r}\\) and importance scores \\(I\\) across layers. This is defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.851, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {c _ {r}, I} \\mathbb {E} _ {(x, y) \\sim D} \\left[ \\max _ {\\delta} \\mathcal {L} _ {r} (f (\\theta) \\odot M, x + \\delta , y) \\right] \\tag {6} \\\\ + \\varphi \\cdot \\mathcal {L} _ {p} (f (\\theta) \\odot M, a _ {t}) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.123 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_p\\) enforces target compression \\(a_{t}\\) across all layers \\(L\\) to guide non-uniform pruning rates." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.133, + 0.757, + 0.15 + ], + "angle": 0, + "content": "3.2. Pruning Importance Score" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.156, + 0.906, + 0.2 + ], + "angle": 0, + "content": "The importance score's definition varies by pruning method. This section details its definition and computation, as discussed in Sec.3.1." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.202, + 0.906, + 0.293 + ], + "angle": 0, + "content": "Neuron Importance Score (NIS) NIS [50] is a common pruning score that minimizes the weighted distance between the original and pruned final responses in a given layer. Given a neural network \\( f^{(n)} \\) with \\( n \\) layers, and the importance score of the final layer response \\( I^{(n)} \\), the importance score of the \\( i \\)-th layer \\( I^{(i)} \\) is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.304, + 0.906, + 0.327 + ], + "angle": 0, + "content": "\\[\nI ^ {(i)} = \\left| w ^ {(i + 1)} \\right| ^ {T} \\cdot \\left| w ^ {(i + 2)} \\right| ^ {T} \\dots \\left| w ^ {(n)} \\right| ^ {T} \\cdot I ^ {(n)} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.338, + 0.906, + 0.412 + ], + "angle": 0, + "content": "Here, \\( w^{(i)} \\) refers to the weight matrix of the \\( i \\)-th layer. In this case, the importance of a neuron in the network can be computed recursively along the network. Thus, \\( I^{(i)} \\) can be propagated from the importance score of the \\( (i + 1) \\)-th layer as follow:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.412, + 0.905, + 0.434 + ], + "angle": 0, + "content": "\\[\nI _ {N I S} ^ {(i)} = \\left| w ^ {(i + 1)} \\right| ^ {T} \\cdot I ^ {(i + 1)} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.442, + 0.906, + 0.502 + ], + "angle": 0, + "content": "Empirical Risk Minimization (ERM) The importance score based on ERM, used in HYDRA[34] and HARP, is defined as follows, based on the weight magnitudes of a pre-trained model." + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.514, + 0.906, + 0.55 + ], + "angle": 0, + "content": "\\[\nI _ {E R M} ^ {(i)} = \\frac {\\eta \\cdot | \\theta^ {(i)} |}{\\max (| \\theta^ {(i)} |)} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.564, + 0.906, + 0.625 + ], + "angle": 0, + "content": "Here, \\( I_{ERM}^{(i)} \\) represents the importance score for the weights of the \\( i \\)-th layer, \\( \\theta^{(i)} \\) denotes the weight matrix of that layer, and \\( \\eta \\) is a scaling factor determined by the layer's structure and receptive field size." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.627, + 0.906, + 0.701 + ], + "angle": 0, + "content": "Adversarial Saliency Estimation (ASE) MAD [21] computes the ASE for each parameter, removing the least salient ones. Adversarial saliency is estimated via a second-order Taylor expansion, with the saliency score \\( I_{ASE} \\) for each parameter \\( w^{(i)} \\) defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.715, + 0.905, + 0.75 + ], + "angle": 0, + "content": "\\[\nI _ {A S E} = \\frac {1}{2} \\left[ \\frac {\\partial^ {2} \\mathcal {L}}{\\partial \\left(w ^ {(i)}\\right) ^ {2}} \\right] \\left| w ^ {(i)} \\right| ^ {2} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.757, + 0.906, + 0.825 + ], + "angle": 0, + "content": "Here, \\(\\frac{\\partial^2\\mathcal{L}}{\\partial(w^{(i)})^2}\\) is the Hessian of the loss function with respect to \\(w^{(i)}\\), reflecting the impact of parameter changes on the loss, and \\(|w^{(i)}|^2\\) reflects the magnitude and importance of the corresponding parameter." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Batch Normalization Scaling Factor (BNSF) BNAP [43] identifies the magnitude of the Batch Normalization (BN) scaling factor \\(\\gamma\\) as a robustness measure across channels, used as the pruning criterion. The BN layer normalizes weights \\(w_{ik}^{j}\\) into the following effective weights:" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9698" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.098, + 0.182, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.199, + 0.154, + 0.209 + ], + "angle": 0, + "content": "Clean" + }, + { + "type": "image", + "bbox": [ + 0.199, + 0.101, + 0.282, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.199, + 0.258, + 0.209 + ], + "angle": 0, + "content": "Attack" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.1, + 0.388, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.199, + 0.354, + 0.209 + ], + "angle": 0, + "content": "AT" + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.1, + 0.493, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.44, + 0.199, + 0.461, + 0.209 + ], + "angle": 0, + "content": "NIS" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.1, + 0.596, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.199, + 0.567, + 0.209 + ], + "angle": 0, + "content": "ERM" + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.1, + 0.698, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.643, + 0.199, + 0.668, + 0.209 + ], + "angle": 0, + "content": "ASE" + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.1, + 0.802, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.199, + 0.776, + 0.209 + ], + "angle": 0, + "content": "BNSF" + }, + { + "type": "image", + "bbox": [ + 0.828, + 0.1, + 0.876, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.102, + 0.218, + 0.381, + 0.231 + ], + "angle": 0, + "content": "(a) Importance score of models with different inputs." + }, + { + "type": "image_footnote", + "bbox": [ + 0.489, + 0.218, + 0.82, + 0.231 + ], + "angle": 0, + "content": "(b) Importance score of a model with different scoring metrics." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.242, + 0.905, + 0.271 + ], + "angle": 0, + "content": "Figure 2. Importance score visualization for the 4-5-6-3 CNN model on the MNIST dataset: yellow/red indicate higher importance, green/blue indicate lower." + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.303, + 0.483, + 0.342 + ], + "angle": 0, + "content": "\\[\n\\hat {w} _ {i k} ^ {j} = \\frac {\\gamma_ {i} ^ {j} w _ {i k} ^ {j}}{\\left(\\sigma_ {i} ^ {j}\\right) ^ {2} + \\epsilon} \\approx \\frac {\\gamma_ {i} ^ {j} w _ {i k} ^ {j}}{\\sigma_ {i} ^ {j}} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.349, + 0.484, + 0.44 + ], + "angle": 0, + "content": "where \\(i\\) is the layer index, \\(j\\) is the channel index of \\(i\\)-th layer, and \\(k\\) is the channel index in the \\((i - 1)\\)-th layer. In this case, \\(\\epsilon\\) is small enough to be neglected, and \\(\\gamma_i^j\\) and \\(\\sigma_i^j\\) each play a role in readjusting the weights per channel. Therefore, the computed weights can be used as the importance score." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.453, + 0.483, + 0.485 + ], + "angle": 0, + "content": "3.3. A Closer Look at Adversarial Pruning via Importance Score" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.61 + ], + "angle": 0, + "content": "Least Importance Score [50] is a common pruning criterion that assesses each parameter's (weights, filters, channels) importance to overall network performance, pruning the least important first. This score clarifies the impact of adversarial examples on pruning. We compared importance scores between models trained on clean data, adversarial data, and adversarily-trained models using a simple CNN architecture." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.484, + 0.763 + ], + "angle": 0, + "content": "As shown in Fig. 2a, importance scores differ significantly between models trained on clean versus adversarial datasets. In the AT model, trained on both, scores do not consistently fall between clean and adversarial values but more closely resemble the clean model. Notably, nodes highly important in both cases (e.g., the second node in the second layer) may appear less important in the AT model. This suggests that training on both datasets may obscure importance, potentially leading to pruning of critical nodes or connections." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Moreover, we evaluated the same model with different importance scoring methods mentioned earlier. The results in Fig.2b suggest that the order of importance for nodes and connections can vary depending on the scoring method. In other words, each sub-network pruned with a different importance scoring method may contain important information that is not shared by the others. This implies that sub-models sharing the same base model could act as multiple models in ED, while complementing each other." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.288, + 0.824, + 0.305 + ], + "angle": 0, + "content": "4. Efficient Ensemble Defense (EED)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.316, + 0.905, + 0.435 + ], + "angle": 0, + "content": "A common hypothesis in adversarial ensemble defense is that training and combining multiple base models enhances adversarial robustness compared to a single model [7]. However, traditional ED methods often increase model capacity, which is counterproductive for reducing computational and memory costs. But is it truly impossible to achieve an efficient adversarial ensemble defense for a compact model?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.438, + 0.906, + 0.68 + ], + "angle": 0, + "content": "We propose a novel Efficient Ensemble Defense (EED) method that leverages model compression while maintaining robustness. Sec.3.3 has demonstrated the potential for sub-models, pruned based on different importance scores from a single base model, to act as multiple base models in an ensemble. By ensembling these sub-models, we can reduce model capacity compared to traditional ED. However, it is observed that the smaller the model during ensemble, the more the robustness performance decreases; thus, high diversity is required to mitigate this issue[30]. Therefore, we enhance the ensemble effect and maintain robustness through robust diversity evaluation and misclassification regularization. Furthermore, to improve inference speed—a primary goal of model compression—we dynamically add sub-models during inference and perform ensemble only as needed, thus enhancing inference efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.681, + 0.906, + 0.743 + ], + "angle": 0, + "content": "EED is designed with three main considerations: 1) the diversity of pruned sub-models, 2) improvement of robustness during the ensemble process, and 3) maintaining compactness throughout the ensemble process." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.766, + 0.773 + ], + "angle": 0, + "content": "4.1. Enhancing Robust Diversity" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Our strategy begins with generating sub-models from a base model through pruning, where each sub-model undergoes an AP process based on a unique importance score. This AP adheres to the principles described in Sec.3.1, utilizing four importance scoring methods detailed in Sec.3.2. To further enhance diversity, we divide the training dataset for pruning into multiple subsets, with each subset designated to train a specific sub-model. Some subsets are shared across all" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9699" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.093, + 0.091, + 0.414, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Efficient Ensemble Defense (EED)" + }, + { + "type": "algorithm", + "bbox": [ + 0.108, + 0.11, + 0.482, + 0.552 + ], + "angle": 0, + "content": "Input: Training data \\(D\\) , base model \\(M\\) \nOutput: Ensemble Selected \nfor each subset \\(D_{i}\\subset D\\) do \\(F_{i}\\gets\\) AdversarialPruning \\((M,D_{i})\\) Set \\((D)\\leftarrow \\operatorname {Set}(D)\\cup F_i\\) \nend for \nEnsSet \\(\\leftarrow\\) FormEnsembleSet(Set(D)) \nfor each ensemble \\(E\\in \\mathit{ensSet}\\) do \\(RD\\gets\\) CalculateRobustDiversity(E) \nend for \nSelected \\(\\leftarrow\\) SelectEnsemble(EnsSet, RD) \nfor each training example \\((x,y)\\in D\\) do \\(\\mathcal{L}_{\\mathrm{EED}}(x,y)\\gets \\mathcal{L}_E + \\mathcal{L}_R + \\mathcal{L}_C\\) Selected \\(\\leftarrow\\) UpdateModels(Selected, L_EED) \nend for {INFERENCE PHASE} \nfor each test example \\(x\\in T\\) do pred \\(\\leftarrow\\) InitializePredictions(Selected) t \\(\\leftarrow 1\\) while not sufficient robustness do pred \\(\\leftarrow\\) UpdateEnsemble(pred, Selected[t]) t \\(\\leftarrow t + 1\\) qt \\(\\leftarrow\\) CalculateStoppingCriterion(pred) if robustness criterion met then break end if end while \nend for \nOutput pred" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.484, + 0.731 + ], + "angle": 0, + "content": "sub-models to retain common core knowledge, while the remaining subsets are distributed across different sub-models to ensure each to learn somewhat different aspects of the data. These subsets are randomly reshuffled whenever the importance scoring method changes. This approach amplifies model diversity, as each sub-model becomes specialized in defending against attacks targeting specific patterns or features in the data. Ultimately, we establish a model pool composed of sub-models with varying importance scoring methods or trained on different datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.732, + 0.484, + 0.826 + ], + "angle": 0, + "content": "Let this be denoted as pool of \\(N\\) models trained on the dataset \\(D\\), represented as \\(Set(D) = \\{F_0, \\dots, F_{N-1}\\}\\). The ensemble set \\(EnsSet\\) represents the collection of all possible ensemble teams with sizes \\(S\\) ranging from 2 to \\(N\\) formed from \\(Set(D)\\). The number of ensemble teams is given by \\(|EnsSet| = \\sum_{S=2}^{N} \\binom{N}{S} = 2^N - (1 + N)\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To quantify the diversity among sub-models within each ensemble team, we define a Robust Diversity (RD) score based on the probability of defense failure for the model. The diversity score relies on two probabilities: \\( p_{one} \\) and \\( p_{two} \\). \\( p_{one} \\) is the probability that a randomly chosen model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.136 + ], + "angle": 0, + "content": "fails to defend (viz. the attack succeeds), while \\( p_{two} \\) is the expected probability that two randomly selected models both fail to defend:" + }, + { + "type": "equation", + "bbox": [ + 0.546, + 0.144, + 0.907, + 0.186 + ], + "angle": 0, + "content": "\\[\np _ {o n e} = \\sum_ {i = 1} ^ {S} \\frac {S _ {i}}{S} p _ {i}, \\quad p _ {t w o} = \\sum_ {i = 1} ^ {S} \\frac {S _ {i} (i - 1)}{S (S - 1)} p _ {i} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.195, + 0.907, + 0.241 + ], + "angle": 0, + "content": "Here, \\( p_i \\) represents the defense failure probability of the \\( i \\)-th model among the total \\( S \\) models. Based on the two probabilities, the RD score is defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.652, + 0.25, + 0.907, + 0.279 + ], + "angle": 0, + "content": "\\[\nR D = 1 - \\frac {p _ {\\text {t w o}}}{p _ {\\text {o n e}}} \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.287, + 0.907, + 0.394 + ], + "angle": 0, + "content": "This RD metric measures diversity. \\(RD = 1\\) indicates maximum diversity (i.e., even if one model fails in defense, the likelihood of other models succeeding is high), while \\(RD = 0\\) indicates no diversity (i.e., both models fail in defense). The EED framework applies RD-based ensemble selection to form the smallest ensemble from the EnsSet of sub-models with the high RD scores." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.402, + 0.855, + 0.418 + ], + "angle": 0, + "content": "4.2. Regularization against Misclassification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.907, + 0.485 + ], + "angle": 0, + "content": "As in the existing ED technique [47], the classifier is trained by combining the classification loss, regularization \\((Reg)\\) and diversity \\((Div)\\) terms. Specifically, for a single training example \\((x,y)\\), the ensemble loss function is defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.495, + 0.906, + 0.558 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {E} (x, y) = \\sum_ {i = 1} ^ {N} \\ell \\left(F _ {i} (x), y\\right) + \\alpha R e g (h (x)) \\tag {14} \\\\ + \\beta D i v (h ^ {1} (x), h ^ {2} (x), \\ldots , h ^ {N} (x), y) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.568, + 0.906, + 0.629 + ], + "angle": 0, + "content": "Here, \\(\\alpha\\) and \\(\\beta\\) are hyperparameters, and \\(h^i (x) = (h_1^i (x),\\ldots ,h_C^i (x)), i\\in \\mathbb{N}\\) represents the output of the ith classifier. The final output is calculated using the average combiner \\(h(x) = \\frac{1}{N}\\sum_{i = 1}^{N}h^{i}(x)\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.69 + ], + "angle": 0, + "content": "To improve the robustness of EED, we employ an additional regularization term that focuses on the probability scores of the most misclassified class, inspired by iGAT [7]. For a given example \\((x,y)\\), the proposed term is defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.699, + 0.905, + 0.754 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {R} (x, y) = - \\psi_ {0 / 1} (c (h ^ {1} (x), \\dots , h ^ {N} (x)), y) \\\\ \\cdot \\log \\left(1 - C \\max _ {i = 1} ^ {N} \\max _ {j = 1} ^ {C} h _ {i} ^ {j} (x)\\right) \\tag {15} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.763, + 0.907, + 0.902 + ], + "angle": 0, + "content": "The error function \\(\\psi_{0/1}(f, y) = \\frac{1}{|D|} \\sum_{x \\in D} 1 [f_y(x) < \\max_{c=y} f_c(x)]\\) returns 0 if \\(f\\) correctly predicts the label \\(y\\), and 1 otherwise. This regularization term plays a crucial role in guiding the model to reduce misclassifications, especially improving its ability to handle ambiguous data points or difficult input examples. By doing so, each base classifier focuses more on the examples it misclassified, enhancing the robustness of the overall ensemble model across diverse input spaces." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9700" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.376, + 0.108 + ], + "angle": 0, + "content": "4.3. Regularization for Compactness" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.265 + ], + "angle": 0, + "content": "Another key objective of the EED framework is maintaining the compressed models' compactness while maximizing the ensemble merit. To achieve this, we design a compactness regularization term \\(\\mathcal{L}_C\\) to maintain weight and activation sparsity among the sub-models, thereby improving the compactness of the ensemble. The weight regularization term controls the sparsity of the sub-model weights, ensuring that they remain sparse above a certain threshold, thus suppressing model capacity growth. Specifically, the weight sparsity regularization term \\(\\mathcal{L}_W\\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.274, + 0.484, + 0.317 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {W} (x, y) = \\sum_ {i = 1} ^ {N} \\| W _ {i} \\| _ {1}, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.326, + 0.484, + 0.371 + ], + "angle": 0, + "content": "Here, \\( \\| W_i\\| _1 \\) denotes the \\( L_{1} \\)-norm of the weights \\( W_{i} \\) of the \\( i \\)-th sub-model, which encourages sparsity by minimizing the sum of the absolute values of the weights." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.371, + 0.484, + 0.432 + ], + "angle": 0, + "content": "To control the activation sparsity of the sub-models, we use an additional sparsity regularization term \\(\\mathcal{L}_A\\), which encourages each sub-model to reduce unnecessary neuron activations. This term is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.44, + 0.483, + 0.481 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {A} (x, y) = \\sum_ {i = 1} ^ {N} \\| h ^ {i} (x) \\| _ {0}, \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.483, + 0.522 + ], + "angle": 0, + "content": "Here, \\( \\| h^{i}(x)\\|_{0} \\) represents the sum of the number of nonzero activations in the output \\( h^i (x) \\) of the \\( i \\)-th sub-model." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.522, + 0.384, + 0.537 + ], + "angle": 0, + "content": "Thus, \\(\\mathcal{L}_C\\) can be summarized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.193, + 0.549, + 0.483, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C} (x, y) = \\lambda_ {1} \\mathcal {L} _ {W} + \\lambda_ {2} \\mathcal {L} _ {A} \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.575, + 0.484, + 0.664 + ], + "angle": 0, + "content": "The hyperparameter \\(\\lambda_{1}\\) regulates weight sparsity, with higher values leading to more sparse weights and reduced model capacity. Similarly, \\(\\lambda_{2}\\) controls activation sparsity, making neuron activations sparser, which results in a more compact model structure that saves memory and computational resources." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.666, + 0.471, + 0.681 + ], + "angle": 0, + "content": "Finally, the overall loss function of EED is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.692, + 0.483, + 0.709 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {E E D} (x, y) = \\mathcal {L} _ {E} (x, y) + \\omega \\mathcal {L} _ {R} (x, y) + \\gamma \\mathcal {L} _ {C} (x, y) \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.484, + 0.78 + ], + "angle": 0, + "content": "The EED framework provides high robustness against adversarial attacks while ensuring compactness through these multiple regularization terms, making it suitable to efficient applications like real-time inference." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.788, + 0.461, + 0.805 + ], + "angle": 0, + "content": "4.4. Dynamic Ensemble for Inference Efficiency" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In the training phase, EED optimizes all sub-models to ensure diversity and robustness. To enhance the compactness and efficiency of EED during the inference phase, we adopt an inference-efficient ensemble method [25]. We apply this to adversarial defense by dynamically constructing an ensemble of varying sizes for each sample, termed Dynamic" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "Inference Ensemble (DIE). Specifically, for a given sample \\( x \\), EED dynamically determines the optimal point to stop while adding each sub-models progressively. As a result, by monitoring the uncertainty of each sub-model on adversarial examples and only introducing additional models when necessary, we avoid unnecessary computations while maintaining high robustness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.198, + 0.906, + 0.319 + ], + "angle": 0, + "content": "DIE sequentially adds sub-models and evaluates the robustness of the predictions at each stage. Starting with the first sub-model \\( F_{1} \\), a new sub-model \\( F_{t} \\) is added at each stage \\( t \\), and the intermediate ensemble prediction \\( \\hat{y}_{ens}^{t} \\) is updated. These intermediate predictions serve as a criterion for assessing how robust the added sub-models are. The intermediate ensemble prediction based on the predictions \\( \\hat{y}_i \\) of each sub-model \\( F_{i} \\) is computed by:" + }, + { + "type": "equation", + "bbox": [ + 0.653, + 0.328, + 0.905, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\hat {y} _ {\\text {e n s}} ^ {t} = \\frac {1}{t} \\sum_ {i = 1} ^ {t} \\hat {y} _ {i} \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.905, + 0.438 + ], + "angle": 0, + "content": "Here, each \\(\\hat{y}_{\\text{ens}}^t\\) represents the combined prediction of the sub-models for the input (both clean and adversarial examples), and it plays a role in strengthening the overall robustness of the ensemble." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.906, + 0.59 + ], + "angle": 0, + "content": "At each stage \\( t \\), the early stopping probability \\( q_{t} \\) is assessed. If the current ensemble performance is determined to be robust enough, the addition of sub-models is halted. The value of \\( q_{t} \\) is computed based on the prediction uncertainty and confidence. The prediction uncertainty is computed by KL-divergence of \\( t \\) and \\( (t - 1) \\)-th ensemble predictions, and the prediction confidence setting follows WoC[40]. The optimal stopping point \\( z_{x} \\) is defined as follows, based on a geometric distribution modeling with discrete Bernoulli trials:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.597, + 0.905, + 0.639 + ], + "angle": 0, + "content": "\\[\nz _ {x} = \\arg \\max _ {t} \\left(q _ {t} \\prod_ {i = 1} ^ {t - 1} \\left(1 - q _ {i}\\right)\\right) \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.647, + 0.905, + 0.692 + ], + "angle": 0, + "content": "Here, \\( q_{t} \\) is dynamically adjusted at each stage based on robustness, and sub-models are added only when necessary, allowing for efficient use of computational resources." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.693, + 0.905, + 0.737 + ], + "angle": 0, + "content": "The final ensemble prediction \\(\\hat{y}_{ens}\\) is calculated as the average of the predictions of all sub-models up to the stopping point \\(z_{x}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.65, + 0.738, + 0.905, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\hat {y} _ {e n s} = \\frac {1}{z _ {x}} \\sum_ {i = 1} ^ {z _ {x}} \\hat {y} _ {i} \\tag {22}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.855 + ], + "angle": 0, + "content": "This approach takes into account the predictions of all submodels while avoiding the inclusion of unnecessary ones, thereby enhancing efficiency. By applying different ensemble sizes for each sample, the overall computational cost is reduced while still being robust against adversarial attacks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "DIE finds a balance between efficiency and robustness, quickly identifying the optimal combination of sub-models to ensure strong performance even in real-time inference" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9701" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.149, + 0.09, + 0.848, + 0.104 + ], + "angle": 0, + "content": "Table 1. Comparison of various AP methods and EED on CIFAR-10 and SVHN datasets when sparsity rate \\( {s}_{r} = {80}\\% \\)" + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.116, + 0.924, + 0.367 + ], + "angle": 0, + "content": "
SettingCIFAR-10SVHNPrams
CleanPGDAAC&WDFSpeed upCleanPGDAAC&WDFSpeed up
Resnet 18AT[27]87.0556.1448.0257.6053.101.00x93.3756.2750.1458.8559.801.00x11.2M
R-ADMM[48]81.2548.0043.9249.1739.111.68x74.8149.7337.4052.6243.401.63x
HYDRA[34]77.3652.9243.7449.6445.911.77x91.0652.2247.6255.1346.131.74x
RST[10]61.0241.0118.3851.0226.821.86x82.3946.2936.3552.6539.271.83x
MAD[21]74.1850.3841.2754.1738.441.69x92.8451.6539.8759.8044.721.70x
Flying Bird[2]81.0751.6244.4156.0845.291.79x90.2152.0642.0157.3052.231.76x2.2M
HARP[52]83.8452.5645.3656.5747.041.81x92.6054.1645.8957.2851.241.80x
TwinRep[23]77.2652.0443.5256.6046.181.82x92.9653.8344.7356.2648.611.78x
EED(ours)86.1355.7148.1357.0351.971.86x93.1555.7450.1858.3756.051.85x
VGG 16AT[27]82.7054.4948.5254.9156.911.00x93.0657.6452.2855.3661.491.00x14.7M
R-ADMM[48]76.5147.1940.4144.8048.291.65x65.4048.3944.6448.9249.851.66x
HYDRA[34]78.8347.8544.8545.7349.411.72x90.8253.6646.3052.6353.421.72x
RST[10]77.2939.9030.1532.0242.031.83x82.3941.9445.5243.5838.811.81x
MAD[21]73.6548.8140.8847.1645.501.80x89.9449.4640.3153.0347.371.69x
Flying Bird[2]77.3649.4345.7350.3848.341.68x90.1153.2949.1852.8150.781.73x2.9M
HARP[52]80.9251.7744.0252.7952.171.73x93.2255.5344.7352.9656.161.77x
TwinRep[23]76.0450.1243.2652.4251.591.79x92.6054.6149.0651.2456.581.75x
EED(ours)81.3954.2647.4953.2753.941.79x93.1457.1951.4953.1260.271.77x
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.386, + 0.483, + 0.448 + ], + "angle": 0, + "content": "settings. As a result, the EED framework can lower the average inference cost in adversarial attack scenarios while maintaining high level of robustness, making it an effective defense, even in resource-constrained environments." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.461, + 0.334, + 0.478 + ], + "angle": 0, + "content": "5. Experimental Evaluations" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.485, + 0.274, + 0.502 + ], + "angle": 0, + "content": "5.1. Experiment Setups" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.483, + 0.674 + ], + "angle": 0, + "content": "Classifiers and Datasets We utilize two specific architectures, ResNet18 and VGG16, as classifiers and conduct experiments focusing on CIFAR-10 and SVHN. We divide the dataset into \\(d\\) subsets for pruning into various sub-models, setting \\(d = 4\\) to create three sub-models for each pruning metric. \\(25\\%\\) of the dataset is shared across all sub-models, while the remaining \\(75\\%\\) is allocated separately for each of the three sub-models. With four pruning metrics in play, we end up with a total of \\(N = 12\\) sub-models, which together form the ensemble set, denoted as EnsSet. For ensemble learning, the entire dataset is used." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Attacks We use PGD-20 attacker with a step size of 2/255 to generate adversarial examples for training, and Auto Attack(AA)[6], C&W[1] and Deepfool(DF)[28] for evaluation. Except for DF using \\( l_{2} \\) norm, the rest use \\( l_{\\infty} \\) norm, and all attacks use perturbation strength \\( \\epsilon = 8 / 255 \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Training We use the Madry[27] as AT method. The batch size is set to 512, and the multi-step learning rates are \\{0.01, 0.002\\} for CIFAR-10 and \\{0.1, 0.02, 0.004\\} for SVHN. The hyperparameters are set as follows: \\(\\omega = 10\\), \\(\\lambda_{1} = 0.7\\), \\(\\lambda_{2} = 0.25\\), and the RD threshold is set to 0.7. \\(c_{r}\\) is set to be more than four times sparser than the desirable \\(s_{r}\\), considering the size of \\(N\\), and \\(\\gamma\\) is adjusted to ensure that the \\(s_{r}\\) is met even after the ensemble. (When \\(s_{r} = 80\\%\\), we set \\(c_{r} = 95\\%\\) and \\(\\gamma = 4\\).) Each experiment is run on an NVIDIA RTX A6000 GPU with 8 CPU cores." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.385, + 0.803, + 0.402 + ], + "angle": 0, + "content": "5.2. Result Comparison and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.411, + 0.907, + 0.547 + ], + "angle": 0, + "content": "This section compares the proposed EED with existing APs for sparsity rate \\((s_r)\\) of \\(80\\%\\) on the CIFAR-10 and SVHN datasets. Table 1 demonstrates that EED consistently provided superior adversarial robustness, with significantly higher accuracy against attacks such as PGD, AA, C&W, and DF compared to other methods. Additionally, EED maintained high performance on clean data, indicating a proper balance between adversarial robustness and clean accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.907, + 0.732 + ], + "angle": 0, + "content": "Note that EED demonstrates superior robustness, particularly against strong attacks such as AA and PGD, outperforming existing defense methods like HYDRA, HARP, and TwinRep. For example, in the CIFAR-10 experiment using the ResNet-18 architecture, EED achieved \\(55.71\\%\\) accuracy against PGD and \\(48.13\\%\\) against AA, showing an improvement of over \\(2\\%\\) compared to other methods and achieving the highest robustness. Moreover, despite using only \\(20\\%\\) of the model's parameters, the performance drop in AT was minimal, less than \\(0.5\\%\\). Also, EED showed stronger defense capabilities than AT against PGD in the ResNet-18-based SVHN experiment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "As a result, EED accomplished remarkable performance improvements without sacrificing efficiency. Notably, it also records up to 1.86 times faster inference speeds, comparable to or exceeding existing techniques. This is crucial for practical applications where computational resources may be limited, demonstrating that EED is an ideal defense method for resource-constrained environments. Furthermore, EED showed consistent performance across different model architectures such as ResNet-18 and VGG-16, suggesting that EED can be applied to a wide range of network architectures with minimal adjustment." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9702" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.096, + 0.314, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.275, + 0.314, + 0.289 + ], + "angle": 0, + "content": "Figure 3. Robust performance by \\( s_r \\)." + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.091, + 0.619, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.387, + 0.258, + 0.595, + 0.27 + ], + "angle": 0, + "content": "(a) RD analysis on ensemble team size." + }, + { + "type": "image", + "bbox": [ + 0.637, + 0.093, + 0.892, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.663, + 0.258, + 0.867, + 0.271 + ], + "angle": 0, + "content": "(b) RD analysis on sub-model sparsity." + }, + { + "type": "image_caption", + "bbox": [ + 0.379, + 0.279, + 0.88, + 0.294 + ], + "angle": 0, + "content": "Figure 4. Correlation between RD and robustness in different ensemble components." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.316, + 0.255, + 0.331 + ], + "angle": 0, + "content": "5.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.339, + 0.483, + 0.52 + ], + "angle": 0, + "content": "Ablation studies are conducted on Resnet18 and CIFAR-10. Changes along Sparsity Fig.3 shows the change in robustness performance with respect to the sparsity rate. EED demonstrated the highest performance compared to the existing top three methods, except for a very high \\( s_r \\) (95%) against AA. EED tended to show a decrease in performance as \\( s_r \\) increases, which is believed to be due to the reduction in ensemble diversity caused by the higher compression rate. On the other hand, at lower \\( s_r \\), EED exhibited higher robustness than AT due to the enhanced ensemble diversity. More detailed experimental results for different \\( s_r \\) can be found in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.522, + 0.483, + 0.643 + ], + "angle": 0, + "content": "Analyzing Robust Diversity Fig.4 analyzes the role of RD in EED. Fig.3a, which shows robustness according to ensemble team size, demonstrates that a high RD is necessary for small ensemble sizes to achieve high performance. Fig.3b shows robustness from the RD perspective with respect to the compression rate \\((c_r)\\) of the sub-models. We can find that to maintain high performance while keeping sub-models small, a high RD is also required." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.644, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Analyzing Regularization Components EED has two regularization terms, as mentioned in Sec.4.2 and Sec.4.3. Regularization against misclassification \\((L_{R})\\) is controlled by \\(\\omega\\), while regularization for compactness \\((L_{C})\\) is controlled by \\(\\gamma\\). This section analyzes the impact of each term on the performance of EED. Tab.2 shows that \\(L_{R}\\) has a greater effect on clean performance, while \\(L_{C}\\) has a larger impact on attack performance. In particular, when the compactness term is not used, attack performance significantly deteriorates, which is likely due to the need of reducing the ensemble size further to meet the \\(s_{r}\\), thereby diminishing the ensemble advantages." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Effect of Dynamic Inference Ensemble Tab.3 analyzes the effect of DIE for different sparsity rates. At \\( s_r = 80\\% \\), DIE had little impact on the main performance and only showed a slight speed improvement. On the other hand, at \\( s_r = 50\\% \\), activating DIE resulted in performance im" + }, + { + "type": "table_caption", + "bbox": [ + 0.555, + 0.315, + 0.863, + 0.328 + ], + "angle": 0, + "content": "Table 2. Analysis of EED with regularization terms." + }, + { + "type": "table", + "bbox": [ + 0.568, + 0.337, + 0.855, + 0.41 + ], + "angle": 0, + "content": "
SettingsCleanPGDAA
CIFAR-10EED86.1355.7148.13
ω = 074.1647.4536.37
Resnet 18γ = 076.3339.1031.59
ω,γ = 073.9537.8328.94
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.422, + 0.904, + 0.436 + ], + "angle": 0, + "content": "Table 3. Correlation between dynamic inference ensemble and \\( {s}_{r} \\) ." + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.445, + 0.898, + 0.523 + ], + "angle": 0, + "content": "
SettingsCleanPGDAASpeed up
sr=80%w/ dynamic86.1355.7148.131.86x
w/o dynamic86.2055.6248.111.80x
sr=50%w/ dynamic86.1556.0952.351.64x
w/o dynamic85.9354.2850.761.27x
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.544, + 0.906, + 0.665 + ], + "angle": 0, + "content": "provements of \\(1.81\\%\\) and \\(1.59\\%\\) for PGD and AA, respectively, while significantly improving inference speed. This suggests that DIE is more effective at lower sparsity. The reason for this performance improvement is that, with lower sparsity, the ensemble can be larger, thus enhancing the optimization effect of the dynamic ensemble. In contrast, without DIE, the likelihood of conflicts between the sub-models may increase as the ensemble grows larger." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.679, + 0.634, + 0.695 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.704, + 0.907, + 0.9 + ], + "angle": 0, + "content": "This paper proposes the Efficient Ensemble Defense (EED) technique to maintain both compression and robustness of deep learning models. EED constructs an ensemble of sub-models using different pruning importance scores to enhance adversarial robustness while reducing model size and computational cost. By addressing information loss and limited model diversity caused by pruning with ensemble methods and regularization, EED outperforms existing AP methods in accuracy and robustness. Experimental results show superior resilience of EED against various attacks on the CIFAR-10 and SVHN, maintaining inference efficiency. EED also proves its applicability in diverse neural network architectures and resource-constrained environments." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9703" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.114, + 0.484, + 0.241 + ], + "angle": 0, + "content": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.RS-2022-00155915 (Artificial Intelligence Convergence Innovation Human Resources Development (Inha University)) and No. 2021-0-02068 (AI Innovation Hub)), and this work was partly supported by the National Research Foundation of Korea (NRF) grant funded by the MSIT (No. 2022R1A2C2010095)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.268, + 0.187, + 0.284 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.293, + 0.484, + 0.335 + ], + "angle": 0, + "content": "[1] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE symposium on security and privacy (sp), pages 39-57. IEEE, 2017. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.336, + 0.484, + 0.404 + ], + "angle": 0, + "content": "[2] Tianlong Chen, Zhenyu Zhang, Pengjun Wang, Santosh Balachandra, Haoyu Ma, Zehao Wang, and Zhangyang Wang. Sparsity winning twice: Better robust generalization from more efficient training. arXiv preprint arXiv:2202.09844, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.407, + 0.484, + 0.463 + ], + "angle": 0, + "content": "[3] Wenlin Chen, James Wilson, Stephen Tyree, Kilian Weinberger, and Yixin Chen. Compressing neural networks with the hashing trick. In International conference on machine learning, pages 2285-2294. PMLR, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.464, + 0.484, + 0.546 + ], + "angle": 0, + "content": "[4] Wuyang Chen, Ziyu Jiang, Zhangyang Wang, Kexin Cui, and Xiaoning Qian. Collaborative global-local networks for memory-efficient segmentation of ultra-high resolution images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8924-8933, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.548, + 0.484, + 0.645 + ], + "angle": 0, + "content": "[5] Hao Cheng, Kaidi Xu, Zhengang Li, Pu Zhao, Chenan Wang, Xue Lin, Bhavya Kailkhura, and Ryan Goldhahn. More or less (mol): Defending against multiple perturbation attacks on deep neural networks through model ensemble and compression. In 2022 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW), pages 645-655. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[6] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.484, + 0.731 + ], + "angle": 0, + "content": "[7] Yian Deng and Tingting Mu. Understanding and improving ensemble adversarial defense. NeurIPS, 36, 2024. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.733, + 0.484, + 0.787 + ], + "angle": 0, + "content": "[8] Misha Denil, Babak Shakibi, Laurent Dinh, Marc'Aurelio Ranzato, and Nando De Freitas. Predicting parameters in deep learning. Advances in neural information processing systems, 26, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.789, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[9] Guneet S Dhillon, Kamyar Azizzadenesheli, Zachary C Lipton, Jeremy Bernstein, Jean Kossaifi, Aran Khanna, and Anima Anandkumar. Stochastic activation pruning for robust adversarial defense. arXiv preprint arXiv:1803.01442, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[10] Yonggan Fu, Qixuan Yu, Yang Zhang, Shang Wu, Xu Ouyang, David Cox, and Yingyan Lin. Drawing robust scratch tickets: Subnetworks with inborn robustness are" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.293, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.133 + ], + "angle": 0, + "content": "found within randomly initialized networks. Advances in Neural Information Processing Systems, 34:13059-13072, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.906, + 0.188 + ], + "angle": 0, + "content": "[11] Ian Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In International Conference on Learning Representations, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.906, + 0.245 + ], + "angle": 0, + "content": "[12] Yiwen Guo, Chao Zhang, Changshui Zhang, and Yurong Chen. Sparse dnns with improved adversarial robustness. Advances in neural information processing systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.906, + 0.3 + ], + "angle": 0, + "content": "[13] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.906, + 0.356 + ], + "angle": 0, + "content": "[14] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. Advances in neural information processing systems, 28, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.906, + 0.413 + ], + "angle": 0, + "content": "[15] Warren He, James Wei, Xinyun Chen, Nicholas Carlini, and Dawn Song. Adversarial example defense: Ensembles of weak defenses are not strong. In 11th USENIX workshop on offensive technologies (WOOT 17), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.906, + 0.482 + ], + "angle": 0, + "content": "[16] Tong Jian, Zifeng Wang, Yanzhi Wang, Jennifer Dy, and Stratis Ioannidis. Pruning adversarially robust neural networks without adversarial examples. In 2022 IEEE International Conference on Data Mining (ICDM), pages 993-998. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.906, + 0.525 + ], + "angle": 0, + "content": "[17] Sanjay Kariyappa and Moinuddin K Qureshi. Improving adversarial robustness of ensembles with diversity training. arXiv preprint arXiv:1901.09981, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.906, + 0.553 + ], + "angle": 0, + "content": "[18] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.906, + 0.622 + ], + "angle": 0, + "content": "[19] Souvik Kundu, Mahdi Nazemi, Peter A Beerel, and Massoud Pedram. Dnr: A tunable robust pruning framework through dynamic network rewiring of dnns. In Proceedings of the 26th Asia and South Pacific Design Automation Conference, pages 344–350, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.906, + 0.663 + ], + "angle": 0, + "content": "[20] Alexey Kurakin, Ian Goodfellow, and Samy Bengio. Adversarial machine learning at scale. arXiv preprint arXiv:1611.01236, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.906, + 0.733 + ], + "angle": 0, + "content": "[21] Byung-Kwan Lee, Junho Kim, and Yong Man Ro. Masking adversarial damage: Finding adversarial saliency for robust and sparse network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15126-15136, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.906, + 0.789 + ], + "angle": 0, + "content": "[22] Bai Li, Shiqi Wang, Yunhan Jia, Yantao Lu, Zhenyu Zhong, Lawrence Carin, and Suman Jana. Towards practical lottery ticket hypothesis for adversarial training. arXiv preprint arXiv:2003.05733, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.907, + 0.858 + ], + "angle": 0, + "content": "[23] Chenhao Li, Qiang Qiu, Zhibin Zhang, Jiafeng Guo, and Xueqi Cheng. Learning adversarially robust sparse networks via weight reparameterization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8527-8535, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[24] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. arXiv preprint arXiv:1608.08710, 2016. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9704" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.148 + ], + "angle": 0, + "content": "[25] Ziyue Li, Kan Ren, Yifan Yang, Xinyang Jiang, Yuqing Yang, and Dongsheng Li. Towards inference efficient deep ensemble learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8711-8719, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.192 + ], + "angle": 0, + "content": "[26] Ningyi Liao, Shufan Wang, Liyao Xiang, Nanyang Ye, Shuo Shao, and Pengzhi Chu. Achieving adversarial robustness via sparsity. Machine Learning, pages 1-27, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[27] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[28] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2574–2582, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[29] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Anima Anandkumar. Diffusion models for adversarial purification. arXiv preprint arXiv:2205.07460, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.433 + ], + "angle": 0, + "content": "[30] Tianyu Pang, Kun Xu, Chao Du, Ning Chen, and Jun Zhu. Improving adversarial robustness via promoting ensemble diversity. In International Conference on Machine Learning, pages 4970-4979. PMLR, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[31] Nicolas Papernot, Patrick McDaniel, Xi Wu, Somesh Jha, and Ananthram Swami. Distillation as a defense to adversarial perturbations against deep neural networks. In 2016 IEEE symposium on security and privacy (SP), pages 582-597. IEEE, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[32] Giorgio Piras, Maura Pintor, Ambra Demontis, Battista Biggio, Giorgio Giacinto, and Fabio Roli. Adversarial pruning: A survey and benchmark of pruning methods for adversarial robustness. arXiv preprint arXiv:2409.01249, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[33] J Redmon. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[34] Vikash Sehwag, Shiqi Wang, Prateek Mittal, and Suman Jana. Hydra: Pruning adversarially robust neural networks. Advances in Neural Information Processing Systems, 33: 19655-19666, 2020. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[35] Ilia Shumailov, Yiren Zhao, Robert Mullins, and Ross Anderson. To compress or not to compress: Understanding the interactions between adversarial attacks and neural network compression. Proceedings of Machine Learning and Systems, 1:230-240, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[36] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. Advances in neural information processing systems, 27, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[37] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[38] Lukas Timpl, Rahim Entezari, Hanie Sedghi, Behnam Neyshabur, and Olga Saukh. Understanding the effect of sparsity on neural networks robustness. arXiv preprint arXiv:2206.10915, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[39] Florian Tramér, Alexey Kurakin, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. Ensemble adversarial training: Attacks and defenses. In International Conference on Learning Representations, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[40] Xiaofang Wang, Dan Kondratyuk, Eric Christiansen, Kris M Kitani, Yair Movshovitz-Attias, and Elad Eban. Wisdom of committees: An overlooked approach to faster and more accurate models. In International Conference on Learning Representations, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.906, + 0.277 + ], + "angle": 0, + "content": "[41] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In International conference on learning representations, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.906, + 0.348 + ], + "angle": 0, + "content": "[42] Yue Wang, Jianghao Shen, Ting-Kuei Hu, Pengfei Xu, Tan Nguyen, Richard Baraniuk, Zhangyang Wang, and Yingyan Lin. Dual dynamic inference: Enabling more efficient, adaptive, and controllable deep inference. IEEE Journal of Selected Topics in Signal Processing, 14(4):623-633, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.906, + 0.405 + ], + "angle": 0, + "content": "[43] Xiao Wei, Yao Zhu, and Shu-Tao Xia. Batch normalization assisted adversarial pruning: Towards lightweight, sparse and robust models. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.906, + 0.462 + ], + "angle": 0, + "content": "[44] Boxi Wu, Jinghui Chen, Deng Cai, Xiaofei He, and Quanquan Gu. Do wider neural networks really help adversarial robustness? Advances in Neural Information Processing Systems, 34:7054-7067, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.906, + 0.518 + ], + "angle": 0, + "content": "[45] Dongxian Wu, Shu-Tao Xia, and Yisen Wang. Adversarial weight perturbation helps robust generalization. Advances in neural information processing systems, 33:2958-2969, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.52, + 0.906, + 0.59 + ], + "angle": 0, + "content": "[46] Yanzhao Wu, Ling Liu, Zhongwei Xie, Ka-Ho Chow, and Wenqi Wei. Boosting ensemble accuracy by revisiting ensemble diversity metrics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16469-16477, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.906, + 0.673 + ], + "angle": 0, + "content": "[47] Huanrui Yang, Jingyang Zhang, Hongliang Dong, Nathan Inkawhich, Andrew Gardner, Andrew Touchet, Wesley Wilkes, Heath Berry, and Hai Li. Diverse: diversifying vulnerabilities for enhanced robust generation of ensembles. Advances in Neural Information Processing Systems, 33: 5505-5515, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.906, + 0.759 + ], + "angle": 0, + "content": "[48] Shaokai Ye, Kaidi Xu, Sijia Liu, Hao Cheng, Jan-Henrik Lambrechts, Huan Zhang, Aojun Zhou, Kaisheng Ma, Yanzhi Wang, and Xue Lin. Adversarial robustness vs. model compression, or both? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 111-120, 2019. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.906, + 0.815 + ], + "angle": 0, + "content": "[49] Jongmin Yoon, Sung Ju Hwang, and Juho Lee. Adversarial purification with score-based generative models. In International Conference on Machine Learning, pages 12062-12072. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[50] Ruichi Yu, Ang Li, Chun-Fu Chen, Jui-Hsin Lai, Vlad I Morariu, Xintong Han, Mingfei Gao, Ching-Yung Lin, and Larry S Davis. Nisp: Pruning networks using neuron importance score propagation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9194-9203, 2018. 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "9705" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[51] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In International conference on machine learning, pages 7472-7482. PMLR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[52] Qi Zhao and Christian Wressnegger. Holistic adversarially robust pruning. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.956 + ], + "angle": 0, + "content": "9706" + } + ] +] \ No newline at end of file diff --git a/2025/Two is Better than One_ Efficient Ensemble Defense for Robust and Compact Models/f8d31d9f-8dff-45eb-9d02-43167ab5a77a_origin.pdf b/2025/Two is Better than One_ Efficient Ensemble Defense for Robust and Compact Models/f8d31d9f-8dff-45eb-9d02-43167ab5a77a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1ad2e54efdd570e40f5c4583dc89708684990d0a --- /dev/null +++ b/2025/Two is Better than One_ Efficient Ensemble Defense for Robust and Compact Models/f8d31d9f-8dff-45eb-9d02-43167ab5a77a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05bb6522188a82ce1e91a1bc7c1f3d6e5a7a7d3e5b41f03513cf68b5679f79df +size 1426116 diff --git a/2025/Two is Better than One_ Efficient Ensemble Defense for Robust and Compact Models/full.md b/2025/Two is Better than One_ Efficient Ensemble Defense for Robust and Compact Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c94b2ac74e4c1454669b541b2daf8b2aa5c6fcb3 --- /dev/null +++ b/2025/Two is Better than One_ Efficient Ensemble Defense for Robust and Compact Models/full.md @@ -0,0 +1,431 @@ +# Two is Better than One: Efficient Ensemble Defense for Robust and Compact Models + +Yoojin Jung Byung Cheol Song + +Department of Electrical and Computer Engineering, Inha University Incheon, Republic of Korea + +yoon0505@gmail.com, bcsong@inha.ac.kr + +# Abstract + +Deep learning-based computer vision systems adopt complex and large architectures to improve performance, yet they face challenges in deployment on resource-constrained mobile and edge devices. To address this issue, model compression techniques such as pruning, quantization, and matrix factorization have been proposed; however, these compressed models are often highly vulnerable to adversarial attacks. We introduce the Efficient Ensemble Defense (EED) technique, which diversifies the compression of a single base model based on different pruning importance scores and enhances ensemble diversity to achieve high adversarial robustness and resource efficiency. EED dynamically determines the number of necessary sub-models during the inference stage, minimizing unnecessary computations while maintaining high robustness. On the CIFAR-10 and SVHN datasets, EED demonstrated state-of-the-art robustness performance compared to existing adversarial pruning techniques, along with an inference speed improvement of up to 1.86 times. This proves that EED is a powerful defense solution in resource-constrained environments. + +# 1. Introduction + +Convolutional Neural Networks (CNNs) have shown remarkable performance across various computer vision tasks, such as image classification, object detection, and semantic segmentation[18, 33, 36]. However, CNNs have become increasingly wider and deeper to achieve higher performance, leading to complex architectures with vast numbers of parameters. This complexity inherently results in significant computational costs and storage demands, making deployment on resource-constrained edge devices challenging[4, 42]. To address this challenge, a number of model compression techniques, including pruning[14, 24], quantization[3], and matrix decomposition[8], have been proposed to minimize model size and computational load + +![](images/e096d87d155417cb0031f40ba099c84fc66be49dfab080bf834e915603e72821.jpg) + +![](images/c94bb48dc532955dfc01c11083e4d6574dbb24eb823279cda220acde293248c2.jpg) +Important nodes or connections may be removed incorrectly due to adversarial attack + +Figure 1. Challenges in AP and the potential of ensemble methods: AP weakens the network by removing critical parameters, reducing robustness and increasing vulnerability to adversarial attacks. An ensemble approach mitigates pruning effects by combining pruned models, enhancing robustness through their collective strength (illustrated in green). +![](images/00a49362902aad1dadfca3eaa37b44ed40e9fc49092ecbf60a09778070a0a78a.jpg) +Ensemble could help compensate + +without compromising accuracy. + +Meanwhile, deep learning models are highly vulnerable to adversarial attacks that induce false predictions by disturbing the input image with small perturbations[11, 37]. This vulnerability is particularly pronounced in sparsely compressed models, which are even more sensitive to adversarial attacks than standard models[35, 44]. Consequently, common adversarial defense techniques such as adversarial training (AT) [20, 27], label smoothing [41, 51], and defensive distillation[31], while generally effective, often experience a degradation in performance when applied to compressed models. This indicates that existing compression techniques may compromise adversarial robustness even if they maintain accuracy on clean test sets. In response, adversarial pruning (AP), a model compression technique that considers adversarial robustness[34, 48, 52], has emerged; + +however, AP still exhibits lower robustness compared to non-compressed defense models[32]. + +We assume that this degradation in robustness arises from reduced generalization capability and information loss caused by pruning. Specifically, adversarial attacks may cause the model to misidentify and even remove critical nodes or connections during the pruning process. In Sec.3.3, we validate this assumption by observing importance scores, which reflect how each component contributes to the loss function through the gradients of the weights[50]. Our findings demonstrate that adversarial examples significantly impact the importance scores of individual parameters. + +Ensemble defense[7, 39] has recently gained attention in adversarial defense research for enhancing the stability and robustness of multiple base classifiers against attacks. As illustrated in Fig.1, we believe that ensemble defense methods can help mitigate generalization degradation and information loss during the pruning process. Since ensemble learning can compensate for deficiencies in individual compressed models due to removed parameters by other models and offset individual errors by combining outputs. However, traditional ensemble defenses inherently suffer from a large capacity, as they require multiple base models[15, 30]. This raises an important question: Is it possible to develop an ensemble defense algorithm that not only achieves robustness but also results in a compact compressed model? + +As an answer to this, we propose Efficient Ensemble Defense (EED). Instead of using multiple base models, we compress a single base model by employing different pruning metrics and then utilize these compressed sub-models in the ensemble. This approach reduces the overall capacity compared to traditional ensemble defense models that rely on multiple base models. Given that using smaller models in an ensemble can lead to reduced defensive efficacy[46], we enhance the variety of the compressed models by incorporating a diversity term in the ensemble selection process, thereby improving ensemble effectiveness while maintaining robustness. Additionally, by adopting inference-efficient ensemble[25], we dynamically add sub-networks and perform ensemble operations as needed during inference, improving overall efficiency. + +The main contributions of our work are as follows: + +- Propose EED, that overcomes the limitations of traditional AP by enhancing the ensemble diversity and compensating for individual errors, thereby achieving robustness against adversarial attacks while being compact. +- Reduce memory by using a single base model in ensemble, and decrease computational costs by dynamically adding models only when necessary during inference. +- Achieve state-of-the-art performance across various attack benchmarks on the CIFAR-10 and SVHN datasets. + +# 2. Related Works + +# 2.1. Adversarial Attack and Defense + +Adversarial attacks exploit the inherent vulnerabilities of deep learning models by adding specific noise to the input, thereby inducing incorrect predictions. These specially crafted inputs, designed to deceive deep learning models, are called adversarial examples. Early studies highlighted the susceptibility of deep learning models to adversarial examples[37]. Then, more potent attacks such as FGSM[11], PGD[27], and AutoAttack[6] have emerged. + +On the other hand, defense methods against attacks have also been evolved. The most well-known adversarial training (AT)[1, 20, 27] enhanced robustness by training adversarial examples. Other strategies involve regularizing geometric properties such as the curvature of the loss function [41, 45], or restoring potentially attacked images to purified versions[29, 49]. Adversarial pruning (AP) and ensemble defense are also one of adversarial defenses (details are given in Sec.3.1.) + +# 2.2. Adversarial Pruning + +AP is a pruning technique (primarily weight pruning) that considers the adversarial context, aiming to reduce the complexity of neural networks while preserving robustness against adversarial attacks. HYDRA demonstrated that empirical adversarial robustness can be achieved through a lowest weight magnitude (LWM) pruning method. R-ADMM[48] and HARP[52] incorporated the alternating direction method of multipliers (ADMM) framework into LWM-based pruning to enhance robustness. In some studies [12, 22, 26], a traditional pruning method, i.e., the Lottery Ticket Hypothesis, was integrated with AT. However, they all failed to maintain generalized robustness at higher sparsity rates $(s_r)[13, 19, 38]$ , and struggled to defend against adaptive attacks[9, 16]. This suggests that conventional AP has not fully resolved the trade-off between model capacity and robustness. + +# 2.3. Ensemble Defense + +Ensemble defense (ED), which applies ensemble learning to adversarial defense, has been actively studied due to its promising performance[7, 39]. The core of ED is to improve predictions through an ensemble diversity loss that considers adversarial robustness during training. ADP[30] used Shannon entropy and geometric diversity for uncertainty, while GAL[17] utilized cosine similarities between model loss gradients to formalize diversity. DVERGE[47] introduced vulnerability diversity to ensure that each model is robust to the weaknesses of others, and MOL[5] increased perturbation diversity by using expert models trained on different attacks. Thus, ED has significantly enhanced robustness by reducing the risk of overfitting to adversarial ex + +amples and diversifying the decision boundary, though with increased model capacity and computational costs. + +# 3. Preliminaries + +# 3.1. Notations + +Adversarial Attack Let $x \in \mathbb{R}^d$ be an input example, $y \in \{1, \dots, C\}$ the true label, and $f(\cdot)$ a classifier with parameters $\theta$ . The classifier's prediction for input $x$ is given by $f(x) = \arg \max_{i \in \{1, \dots, C\}} f_i(x)$ , where $f_i(x)$ denotes the predicted probability for class $i$ . An adversarial example $x'$ is crafted by adding a perturbation $\delta$ to $x$ , aiming to maximize the model's prediction error while keeping $x'$ close to $x$ under a norm constraint. Formally, an adversarial attack is defined by: + +$$ +x ^ {\prime} = x + \delta = \arg \max _ {\| x - x ^ {\prime} \| < \epsilon} \mathcal {L} (f (x ^ {\prime}), y) \tag {1} +$$ + +where $\mathcal{L}$ is the loss function and $\epsilon$ is the perturbation budget that constrains the strength of the attack under an $\ell_p$ -norm. + +Adversarial Training AT is a defense method that improves model robustness by training on adversarially perturbed samples alongside clean samples. The objective for AT is usually formulated by: + +$$ +\min _ {\theta} \mathbb {E} _ {(x, y) \sim \mathcal {D}} \left[ \max _ {\delta : \| \delta \| _ {p} \leq \epsilon} \mathcal {L} \left(f \left(x ^ {\prime}; \theta\right), y\right) \right] \tag {2} +$$ + +where $\mathcal{D}$ is the data distribution. + +Adversarial Pruning Given a pre-trained model $f(\theta)$ under AT, we aim to maintain robustness while applying a pruning mask $M$ across layers to achieve compression. The general pruning is carried out by removing parameters until the network reaches a desired $s_r$ . So, pruning involves creating $M$ , formalized as follows: + +$$ +M = \arg \min _ {\| \mathbf {m} \| _ {0} \leq 1 - s _ {r}} \mathcal {L} (f (\theta) \otimes m, x, y). \tag {3} +$$ + +$$ +m = \left\{ \begin{array}{l l} 0, & \text {i f} \theta_ {i} \text {i s t o b e p r u n e d} \\ 1, & \text {o t h e r w i s e} \end{array} , \quad \forall \theta_ {i} \in \theta \right. \tag {4} +$$ + +For AP, we follow the strategy from HARP[52]. To enhance adversarial robustness, inner maximization generates adversarial examples, targeting a worst-case loss under the PGD attacks: + +$$ +\max _ {\delta} \mathcal {L} _ {r} (f (\theta) \odot M, x + \delta , y) \tag {5} +$$ + +where $\mathcal{L}_r$ defines the robust loss function, incorporating attack-specific parameters. And the global non-uniform pruning strategy jointly optimizes a compression rate $c_{r}$ and importance scores $I$ across layers. This is defined by: + +$$ +\begin{array}{l} \min _ {c _ {r}, I} \mathbb {E} _ {(x, y) \sim D} \left[ \max _ {\delta} \mathcal {L} _ {r} (f (\theta) \odot M, x + \delta , y) \right] \tag {6} \\ + \varphi \cdot \mathcal {L} _ {p} (f (\theta) \odot M, a _ {t}) \\ \end{array} +$$ + +where $\mathcal{L}_p$ enforces target compression $a_{t}$ across all layers $L$ to guide non-uniform pruning rates. + +# 3.2. Pruning Importance Score + +The importance score's definition varies by pruning method. This section details its definition and computation, as discussed in Sec.3.1. + +Neuron Importance Score (NIS) NIS [50] is a common pruning score that minimizes the weighted distance between the original and pruned final responses in a given layer. Given a neural network $f^{(n)}$ with $n$ layers, and the importance score of the final layer response $I^{(n)}$ , the importance score of the $i$ -th layer $I^{(i)}$ is calculated as follows: + +$$ +I ^ {(i)} = \left| w ^ {(i + 1)} \right| ^ {T} \cdot \left| w ^ {(i + 2)} \right| ^ {T} \dots \left| w ^ {(n)} \right| ^ {T} \cdot I ^ {(n)} \tag {7} +$$ + +Here, $w^{(i)}$ refers to the weight matrix of the $i$ -th layer. In this case, the importance of a neuron in the network can be computed recursively along the network. Thus, $I^{(i)}$ can be propagated from the importance score of the $(i + 1)$ -th layer as follow: + +$$ +I _ {N I S} ^ {(i)} = \left| w ^ {(i + 1)} \right| ^ {T} \cdot I ^ {(i + 1)} \tag {8} +$$ + +Empirical Risk Minimization (ERM) The importance score based on ERM, used in HYDRA[34] and HARP, is defined as follows, based on the weight magnitudes of a pre-trained model. + +$$ +I _ {E R M} ^ {(i)} = \frac {\eta \cdot | \theta^ {(i)} |}{\max (| \theta^ {(i)} |)} \tag {9} +$$ + +Here, $I_{ERM}^{(i)}$ represents the importance score for the weights of the $i$ -th layer, $\theta^{(i)}$ denotes the weight matrix of that layer, and $\eta$ is a scaling factor determined by the layer's structure and receptive field size. + +Adversarial Saliency Estimation (ASE) MAD [21] computes the ASE for each parameter, removing the least salient ones. Adversarial saliency is estimated via a second-order Taylor expansion, with the saliency score $I_{ASE}$ for each parameter $w^{(i)}$ defined as: + +$$ +I _ {A S E} = \frac {1}{2} \left[ \frac {\partial^ {2} \mathcal {L}}{\partial \left(w ^ {(i)}\right) ^ {2}} \right] \left| w ^ {(i)} \right| ^ {2} \tag {10} +$$ + +Here, $\frac{\partial^2\mathcal{L}}{\partial(w^{(i)})^2}$ is the Hessian of the loss function with respect to $w^{(i)}$ , reflecting the impact of parameter changes on the loss, and $|w^{(i)}|^2$ reflects the magnitude and importance of the corresponding parameter. + +Batch Normalization Scaling Factor (BNSF) BNAP [43] identifies the magnitude of the Batch Normalization (BN) scaling factor $\gamma$ as a robustness measure across channels, used as the pruning criterion. The BN layer normalizes weights $w_{ik}^{j}$ into the following effective weights: + +Clean +![](images/f29e9f634c90f9504fc5a3607953c9fcfe767ca235f95fd6361906a96313ed95.jpg) +(a) Importance score of models with different inputs. + +![](images/d37ec6b339baabc2e9253d3599ee026f441f7201f270abb0495a89d8dfca76d0.jpg) +Attack + +![](images/5dc2ad8b079621336b4c10967e0de735dca9a74038f9c54833465f715dce03b4.jpg) +AT + +NIS +![](images/19b4f980a11e32b48926026ccdfdde77504c4ba7fd8131964e9fb88b2655a1b3.jpg) +(b) Importance score of a model with different scoring metrics. + +![](images/119d785a023bdc73f48944695313f05eebc553d09226bafb5bb78ed28c92dbc8.jpg) +ERM + +![](images/cb9442754b7db8f627c132e968293ccfd02dce31a9f95617a085aac4cbf33004.jpg) +ASE + +![](images/24a63754435d4f7cf3449fe6efd3eb9f512c4ed00224f8f29374362324fa72a2.jpg) +BNSF + +![](images/bcf2920daf0c2c82ef10b0a1149d4593dd9f8c0cf51a6d2a814de4f535151ffc.jpg) +Figure 2. Importance score visualization for the 4-5-6-3 CNN model on the MNIST dataset: yellow/red indicate higher importance, green/blue indicate lower. + +$$ +\hat {w} _ {i k} ^ {j} = \frac {\gamma_ {i} ^ {j} w _ {i k} ^ {j}}{\left(\sigma_ {i} ^ {j}\right) ^ {2} + \epsilon} \approx \frac {\gamma_ {i} ^ {j} w _ {i k} ^ {j}}{\sigma_ {i} ^ {j}} \tag {11} +$$ + +where $i$ is the layer index, $j$ is the channel index of $i$ -th layer, and $k$ is the channel index in the $(i - 1)$ -th layer. In this case, $\epsilon$ is small enough to be neglected, and $\gamma_i^j$ and $\sigma_i^j$ each play a role in readjusting the weights per channel. Therefore, the computed weights can be used as the importance score. + +# 3.3. A Closer Look at Adversarial Pruning via Importance Score + +Least Importance Score [50] is a common pruning criterion that assesses each parameter's (weights, filters, channels) importance to overall network performance, pruning the least important first. This score clarifies the impact of adversarial examples on pruning. We compared importance scores between models trained on clean data, adversarial data, and adversarily-trained models using a simple CNN architecture. + +As shown in Fig. 2a, importance scores differ significantly between models trained on clean versus adversarial datasets. In the AT model, trained on both, scores do not consistently fall between clean and adversarial values but more closely resemble the clean model. Notably, nodes highly important in both cases (e.g., the second node in the second layer) may appear less important in the AT model. This suggests that training on both datasets may obscure importance, potentially leading to pruning of critical nodes or connections. + +Moreover, we evaluated the same model with different importance scoring methods mentioned earlier. The results in Fig.2b suggest that the order of importance for nodes and connections can vary depending on the scoring method. In other words, each sub-network pruned with a different importance scoring method may contain important information that is not shared by the others. This implies that sub-models sharing the same base model could act as multiple models in ED, while complementing each other. + +# 4. Efficient Ensemble Defense (EED) + +A common hypothesis in adversarial ensemble defense is that training and combining multiple base models enhances adversarial robustness compared to a single model [7]. However, traditional ED methods often increase model capacity, which is counterproductive for reducing computational and memory costs. But is it truly impossible to achieve an efficient adversarial ensemble defense for a compact model? + +We propose a novel Efficient Ensemble Defense (EED) method that leverages model compression while maintaining robustness. Sec.3.3 has demonstrated the potential for sub-models, pruned based on different importance scores from a single base model, to act as multiple base models in an ensemble. By ensembling these sub-models, we can reduce model capacity compared to traditional ED. However, it is observed that the smaller the model during ensemble, the more the robustness performance decreases; thus, high diversity is required to mitigate this issue[30]. Therefore, we enhance the ensemble effect and maintain robustness through robust diversity evaluation and misclassification regularization. Furthermore, to improve inference speed—a primary goal of model compression—we dynamically add sub-models during inference and perform ensemble only as needed, thus enhancing inference efficiency. + +EED is designed with three main considerations: 1) the diversity of pruned sub-models, 2) improvement of robustness during the ensemble process, and 3) maintaining compactness throughout the ensemble process. + +# 4.1. Enhancing Robust Diversity + +Our strategy begins with generating sub-models from a base model through pruning, where each sub-model undergoes an AP process based on a unique importance score. This AP adheres to the principles described in Sec.3.1, utilizing four importance scoring methods detailed in Sec.3.2. To further enhance diversity, we divide the training dataset for pruning into multiple subsets, with each subset designated to train a specific sub-model. Some subsets are shared across all + +Algorithm 1 Efficient Ensemble Defense (EED) +Input: Training data $D$ , base model $M$ +Output: Ensemble Selected +for each subset $D_{i}\subset D$ do $F_{i}\gets$ AdversarialPruning $(M,D_{i})$ Set $(D)\leftarrow \operatorname {Set}(D)\cup F_i$ +end for +EnsSet $\leftarrow$ FormEnsembleSet(Set(D)) +for each ensemble $E\in \mathit{ensSet}$ do $RD\gets$ CalculateRobustDiversity(E) +end for +Selected $\leftarrow$ SelectEnsemble(EnsSet, RD) +for each training example $(x,y)\in D$ do $\mathcal{L}_{\mathrm{EED}}(x,y)\gets \mathcal{L}_E + \mathcal{L}_R + \mathcal{L}_C$ Selected $\leftarrow$ UpdateModels(Selected, L_EED) +end for {INFERENCE PHASE} +for each test example $x\in T$ do pred $\leftarrow$ InitializePredictions(Selected) t $\leftarrow 1$ while not sufficient robustness do pred $\leftarrow$ UpdateEnsemble(pred, Selected[t]) t $\leftarrow t + 1$ qt $\leftarrow$ CalculateStoppingCriterion(pred) if robustness criterion met then break end if end while +end for +Output pred + +sub-models to retain common core knowledge, while the remaining subsets are distributed across different sub-models to ensure each to learn somewhat different aspects of the data. These subsets are randomly reshuffled whenever the importance scoring method changes. This approach amplifies model diversity, as each sub-model becomes specialized in defending against attacks targeting specific patterns or features in the data. Ultimately, we establish a model pool composed of sub-models with varying importance scoring methods or trained on different datasets. + +Let this be denoted as pool of $N$ models trained on the dataset $D$ , represented as $Set(D) = \{F_0, \dots, F_{N-1}\}$ . The ensemble set $EnsSet$ represents the collection of all possible ensemble teams with sizes $S$ ranging from 2 to $N$ formed from $Set(D)$ . The number of ensemble teams is given by $|EnsSet| = \sum_{S=2}^{N} \binom{N}{S} = 2^N - (1 + N)$ . + +To quantify the diversity among sub-models within each ensemble team, we define a Robust Diversity (RD) score based on the probability of defense failure for the model. The diversity score relies on two probabilities: $p_{one}$ and $p_{two}$ . $p_{one}$ is the probability that a randomly chosen model + +fails to defend (viz. the attack succeeds), while $p_{two}$ is the expected probability that two randomly selected models both fail to defend: + +$$ +p _ {o n e} = \sum_ {i = 1} ^ {S} \frac {S _ {i}}{S} p _ {i}, \quad p _ {t w o} = \sum_ {i = 1} ^ {S} \frac {S _ {i} (i - 1)}{S (S - 1)} p _ {i} \tag {12} +$$ + +Here, $p_i$ represents the defense failure probability of the $i$ -th model among the total $S$ models. Based on the two probabilities, the RD score is defined by: + +$$ +R D = 1 - \frac {p _ {\text {t w o}}}{p _ {\text {o n e}}} \tag {13} +$$ + +This RD metric measures diversity. $RD = 1$ indicates maximum diversity (i.e., even if one model fails in defense, the likelihood of other models succeeding is high), while $RD = 0$ indicates no diversity (i.e., both models fail in defense). The EED framework applies RD-based ensemble selection to form the smallest ensemble from the EnsSet of sub-models with the high RD scores. + +# 4.2. Regularization against Misclassification + +As in the existing ED technique [47], the classifier is trained by combining the classification loss, regularization $(Reg)$ and diversity $(Div)$ terms. Specifically, for a single training example $(x,y)$ , the ensemble loss function is defined by: + +$$ +\begin{array}{l} \mathcal {L} _ {E} (x, y) = \sum_ {i = 1} ^ {N} \ell \left(F _ {i} (x), y\right) + \alpha R e g (h (x)) \tag {14} \\ + \beta D i v (h ^ {1} (x), h ^ {2} (x), \ldots , h ^ {N} (x), y) \\ \end{array} +$$ + +Here, $\alpha$ and $\beta$ are hyperparameters, and $h^i (x) = (h_1^i (x),\ldots ,h_C^i (x)), i\in \mathbb{N}$ represents the output of the ith classifier. The final output is calculated using the average combiner $h(x) = \frac{1}{N}\sum_{i = 1}^{N}h^{i}(x)$ . + +To improve the robustness of EED, we employ an additional regularization term that focuses on the probability scores of the most misclassified class, inspired by iGAT [7]. For a given example $(x,y)$ , the proposed term is defined by: + +$$ +\begin{array}{l} \mathcal {L} _ {R} (x, y) = - \psi_ {0 / 1} (c (h ^ {1} (x), \dots , h ^ {N} (x)), y) \\ \cdot \log \left(1 - C \max _ {i = 1} ^ {N} \max _ {j = 1} ^ {C} h _ {i} ^ {j} (x)\right) \tag {15} \\ \end{array} +$$ + +The error function $\psi_{0/1}(f, y) = \frac{1}{|D|} \sum_{x \in D} 1 [f_y(x) < \max_{c=y} f_c(x)]$ returns 0 if $f$ correctly predicts the label $y$ , and 1 otherwise. This regularization term plays a crucial role in guiding the model to reduce misclassifications, especially improving its ability to handle ambiguous data points or difficult input examples. By doing so, each base classifier focuses more on the examples it misclassified, enhancing the robustness of the overall ensemble model across diverse input spaces. + +# 4.3. Regularization for Compactness + +Another key objective of the EED framework is maintaining the compressed models' compactness while maximizing the ensemble merit. To achieve this, we design a compactness regularization term $\mathcal{L}_C$ to maintain weight and activation sparsity among the sub-models, thereby improving the compactness of the ensemble. The weight regularization term controls the sparsity of the sub-model weights, ensuring that they remain sparse above a certain threshold, thus suppressing model capacity growth. Specifically, the weight sparsity regularization term $\mathcal{L}_W$ is defined as follows: + +$$ +\mathcal {L} _ {W} (x, y) = \sum_ {i = 1} ^ {N} \| W _ {i} \| _ {1}, \tag {16} +$$ + +Here, $\| W_i\| _1$ denotes the $L_{1}$ -norm of the weights $W_{i}$ of the $i$ -th sub-model, which encourages sparsity by minimizing the sum of the absolute values of the weights. + +To control the activation sparsity of the sub-models, we use an additional sparsity regularization term $\mathcal{L}_A$ , which encourages each sub-model to reduce unnecessary neuron activations. This term is defined as follows: + +$$ +\mathcal {L} _ {A} (x, y) = \sum_ {i = 1} ^ {N} \| h ^ {i} (x) \| _ {0}, \tag {17} +$$ + +Here, $\| h^{i}(x)\|_{0}$ represents the sum of the number of nonzero activations in the output $h^i (x)$ of the $i$ -th sub-model. + +Thus, $\mathcal{L}_C$ can be summarized as follows: + +$$ +\mathcal {L} _ {C} (x, y) = \lambda_ {1} \mathcal {L} _ {W} + \lambda_ {2} \mathcal {L} _ {A} \tag {18} +$$ + +The hyperparameter $\lambda_{1}$ regulates weight sparsity, with higher values leading to more sparse weights and reduced model capacity. Similarly, $\lambda_{2}$ controls activation sparsity, making neuron activations sparser, which results in a more compact model structure that saves memory and computational resources. + +Finally, the overall loss function of EED is as follows: + +$$ +\mathcal {L} _ {E E D} (x, y) = \mathcal {L} _ {E} (x, y) + \omega \mathcal {L} _ {R} (x, y) + \gamma \mathcal {L} _ {C} (x, y) \tag {19} +$$ + +The EED framework provides high robustness against adversarial attacks while ensuring compactness through these multiple regularization terms, making it suitable to efficient applications like real-time inference. + +# 4.4. Dynamic Ensemble for Inference Efficiency + +In the training phase, EED optimizes all sub-models to ensure diversity and robustness. To enhance the compactness and efficiency of EED during the inference phase, we adopt an inference-efficient ensemble method [25]. We apply this to adversarial defense by dynamically constructing an ensemble of varying sizes for each sample, termed Dynamic + +Inference Ensemble (DIE). Specifically, for a given sample $x$ , EED dynamically determines the optimal point to stop while adding each sub-models progressively. As a result, by monitoring the uncertainty of each sub-model on adversarial examples and only introducing additional models when necessary, we avoid unnecessary computations while maintaining high robustness. + +DIE sequentially adds sub-models and evaluates the robustness of the predictions at each stage. Starting with the first sub-model $F_{1}$ , a new sub-model $F_{t}$ is added at each stage $t$ , and the intermediate ensemble prediction $\hat{y}_{ens}^{t}$ is updated. These intermediate predictions serve as a criterion for assessing how robust the added sub-models are. The intermediate ensemble prediction based on the predictions $\hat{y}_i$ of each sub-model $F_{i}$ is computed by: + +$$ +\hat {y} _ {\text {e n s}} ^ {t} = \frac {1}{t} \sum_ {i = 1} ^ {t} \hat {y} _ {i} \tag {20} +$$ + +Here, each $\hat{y}_{\text{ens}}^t$ represents the combined prediction of the sub-models for the input (both clean and adversarial examples), and it plays a role in strengthening the overall robustness of the ensemble. + +At each stage $t$ , the early stopping probability $q_{t}$ is assessed. If the current ensemble performance is determined to be robust enough, the addition of sub-models is halted. The value of $q_{t}$ is computed based on the prediction uncertainty and confidence. The prediction uncertainty is computed by KL-divergence of $t$ and $(t - 1)$ -th ensemble predictions, and the prediction confidence setting follows WoC[40]. The optimal stopping point $z_{x}$ is defined as follows, based on a geometric distribution modeling with discrete Bernoulli trials: + +$$ +z _ {x} = \arg \max _ {t} \left(q _ {t} \prod_ {i = 1} ^ {t - 1} \left(1 - q _ {i}\right)\right) \tag {21} +$$ + +Here, $q_{t}$ is dynamically adjusted at each stage based on robustness, and sub-models are added only when necessary, allowing for efficient use of computational resources. + +The final ensemble prediction $\hat{y}_{ens}$ is calculated as the average of the predictions of all sub-models up to the stopping point $z_{x}$ : + +$$ +\hat {y} _ {e n s} = \frac {1}{z _ {x}} \sum_ {i = 1} ^ {z _ {x}} \hat {y} _ {i} \tag {22} +$$ + +This approach takes into account the predictions of all submodels while avoiding the inclusion of unnecessary ones, thereby enhancing efficiency. By applying different ensemble sizes for each sample, the overall computational cost is reduced while still being robust against adversarial attacks. + +DIE finds a balance between efficiency and robustness, quickly identifying the optimal combination of sub-models to ensure strong performance even in real-time inference + +Table 1. Comparison of various AP methods and EED on CIFAR-10 and SVHN datasets when sparsity rate ${s}_{r} = {80}\%$ + +
SettingCIFAR-10SVHNPrams
CleanPGDAAC&WDFSpeed upCleanPGDAAC&WDFSpeed up
Resnet 18AT[27]87.0556.1448.0257.6053.101.00x93.3756.2750.1458.8559.801.00x11.2M
R-ADMM[48]81.2548.0043.9249.1739.111.68x74.8149.7337.4052.6243.401.63x
HYDRA[34]77.3652.9243.7449.6445.911.77x91.0652.2247.6255.1346.131.74x
RST[10]61.0241.0118.3851.0226.821.86x82.3946.2936.3552.6539.271.83x
MAD[21]74.1850.3841.2754.1738.441.69x92.8451.6539.8759.8044.721.70x
Flying Bird[2]81.0751.6244.4156.0845.291.79x90.2152.0642.0157.3052.231.76x2.2M
HARP[52]83.8452.5645.3656.5747.041.81x92.6054.1645.8957.2851.241.80x
TwinRep[23]77.2652.0443.5256.6046.181.82x92.9653.8344.7356.2648.611.78x
EED(ours)86.1355.7148.1357.0351.971.86x93.1555.7450.1858.3756.051.85x
VGG 16AT[27]82.7054.4948.5254.9156.911.00x93.0657.6452.2855.3661.491.00x14.7M
R-ADMM[48]76.5147.1940.4144.8048.291.65x65.4048.3944.6448.9249.851.66x
HYDRA[34]78.8347.8544.8545.7349.411.72x90.8253.6646.3052.6353.421.72x
RST[10]77.2939.9030.1532.0242.031.83x82.3941.9445.5243.5838.811.81x
MAD[21]73.6548.8140.8847.1645.501.80x89.9449.4640.3153.0347.371.69x
Flying Bird[2]77.3649.4345.7350.3848.341.68x90.1153.2949.1852.8150.781.73x2.9M
HARP[52]80.9251.7744.0252.7952.171.73x93.2255.5344.7352.9656.161.77x
TwinRep[23]76.0450.1243.2652.4251.591.79x92.6054.6149.0651.2456.581.75x
EED(ours)81.3954.2647.4953.2753.941.79x93.1457.1951.4953.1260.271.77x
+ +settings. As a result, the EED framework can lower the average inference cost in adversarial attack scenarios while maintaining high level of robustness, making it an effective defense, even in resource-constrained environments. + +# 5. Experimental Evaluations + +# 5.1. Experiment Setups + +Classifiers and Datasets We utilize two specific architectures, ResNet18 and VGG16, as classifiers and conduct experiments focusing on CIFAR-10 and SVHN. We divide the dataset into $d$ subsets for pruning into various sub-models, setting $d = 4$ to create three sub-models for each pruning metric. $25\%$ of the dataset is shared across all sub-models, while the remaining $75\%$ is allocated separately for each of the three sub-models. With four pruning metrics in play, we end up with a total of $N = 12$ sub-models, which together form the ensemble set, denoted as EnsSet. For ensemble learning, the entire dataset is used. + +Attacks We use PGD-20 attacker with a step size of 2/255 to generate adversarial examples for training, and Auto Attack(AA)[6], C&W[1] and Deepfool(DF)[28] for evaluation. Except for DF using $l_{2}$ norm, the rest use $l_{\infty}$ norm, and all attacks use perturbation strength $\epsilon = 8 / 255$ . + +Training We use the Madry[27] as AT method. The batch size is set to 512, and the multi-step learning rates are \{0.01, 0.002\} for CIFAR-10 and \{0.1, 0.02, 0.004\} for SVHN. The hyperparameters are set as follows: $\omega = 10$ , $\lambda_{1} = 0.7$ , $\lambda_{2} = 0.25$ , and the RD threshold is set to 0.7. $c_{r}$ is set to be more than four times sparser than the desirable $s_{r}$ , considering the size of $N$ , and $\gamma$ is adjusted to ensure that the $s_{r}$ is met even after the ensemble. (When $s_{r} = 80\%$ , we set $c_{r} = 95\%$ and $\gamma = 4$ .) Each experiment is run on an NVIDIA RTX A6000 GPU with 8 CPU cores. + +# 5.2. Result Comparison and Analysis + +This section compares the proposed EED with existing APs for sparsity rate $(s_r)$ of $80\%$ on the CIFAR-10 and SVHN datasets. Table 1 demonstrates that EED consistently provided superior adversarial robustness, with significantly higher accuracy against attacks such as PGD, AA, C&W, and DF compared to other methods. Additionally, EED maintained high performance on clean data, indicating a proper balance between adversarial robustness and clean accuracy. + +Note that EED demonstrates superior robustness, particularly against strong attacks such as AA and PGD, outperforming existing defense methods like HYDRA, HARP, and TwinRep. For example, in the CIFAR-10 experiment using the ResNet-18 architecture, EED achieved $55.71\%$ accuracy against PGD and $48.13\%$ against AA, showing an improvement of over $2\%$ compared to other methods and achieving the highest robustness. Moreover, despite using only $20\%$ of the model's parameters, the performance drop in AT was minimal, less than $0.5\%$ . Also, EED showed stronger defense capabilities than AT against PGD in the ResNet-18-based SVHN experiment. + +As a result, EED accomplished remarkable performance improvements without sacrificing efficiency. Notably, it also records up to 1.86 times faster inference speeds, comparable to or exceeding existing techniques. This is crucial for practical applications where computational resources may be limited, demonstrating that EED is an ideal defense method for resource-constrained environments. Furthermore, EED showed consistent performance across different model architectures such as ResNet-18 and VGG-16, suggesting that EED can be applied to a wide range of network architectures with minimal adjustment. + +![](images/36788b395246e5fe5c137d24f0967a931957007def8242751b45ec680fea6db3.jpg) +Figure 3. Robust performance by $s_r$ . + +![](images/253520b5df4c39fe39333c8d1f7a28a2e60ec467abcf044af2f7b96f56edbd1b.jpg) +(a) RD analysis on ensemble team size. + +![](images/a23a51dd1ba532a789bfdc6c567670c2a02030253dc94f10caee63acb05550ab.jpg) +(b) RD analysis on sub-model sparsity. +Figure 4. Correlation between RD and robustness in different ensemble components. + +# 5.3. Ablation Studies + +Ablation studies are conducted on Resnet18 and CIFAR-10. Changes along Sparsity Fig.3 shows the change in robustness performance with respect to the sparsity rate. EED demonstrated the highest performance compared to the existing top three methods, except for a very high $s_r$ (95%) against AA. EED tended to show a decrease in performance as $s_r$ increases, which is believed to be due to the reduction in ensemble diversity caused by the higher compression rate. On the other hand, at lower $s_r$ , EED exhibited higher robustness than AT due to the enhanced ensemble diversity. More detailed experimental results for different $s_r$ can be found in the supplementary materials. + +Analyzing Robust Diversity Fig.4 analyzes the role of RD in EED. Fig.3a, which shows robustness according to ensemble team size, demonstrates that a high RD is necessary for small ensemble sizes to achieve high performance. Fig.3b shows robustness from the RD perspective with respect to the compression rate $(c_r)$ of the sub-models. We can find that to maintain high performance while keeping sub-models small, a high RD is also required. + +Analyzing Regularization Components EED has two regularization terms, as mentioned in Sec.4.2 and Sec.4.3. Regularization against misclassification $(L_{R})$ is controlled by $\omega$ , while regularization for compactness $(L_{C})$ is controlled by $\gamma$ . This section analyzes the impact of each term on the performance of EED. Tab.2 shows that $L_{R}$ has a greater effect on clean performance, while $L_{C}$ has a larger impact on attack performance. In particular, when the compactness term is not used, attack performance significantly deteriorates, which is likely due to the need of reducing the ensemble size further to meet the $s_{r}$ , thereby diminishing the ensemble advantages. + +Effect of Dynamic Inference Ensemble Tab.3 analyzes the effect of DIE for different sparsity rates. At $s_r = 80\%$ , DIE had little impact on the main performance and only showed a slight speed improvement. On the other hand, at $s_r = 50\%$ , activating DIE resulted in performance im + +Table 2. Analysis of EED with regularization terms. + +
SettingsCleanPGDAA
CIFAR-10EED86.1355.7148.13
ω = 074.1647.4536.37
Resnet 18γ = 076.3339.1031.59
ω,γ = 073.9537.8328.94
+ +Table 3. Correlation between dynamic inference ensemble and ${s}_{r}$ . + +
SettingsCleanPGDAASpeed up
sr=80%w/ dynamic86.1355.7148.131.86x
w/o dynamic86.2055.6248.111.80x
sr=50%w/ dynamic86.1556.0952.351.64x
w/o dynamic85.9354.2850.761.27x
+ +provements of $1.81\%$ and $1.59\%$ for PGD and AA, respectively, while significantly improving inference speed. This suggests that DIE is more effective at lower sparsity. The reason for this performance improvement is that, with lower sparsity, the ensemble can be larger, thus enhancing the optimization effect of the dynamic ensemble. In contrast, without DIE, the likelihood of conflicts between the sub-models may increase as the ensemble grows larger. + +# 6. Conclusion + +This paper proposes the Efficient Ensemble Defense (EED) technique to maintain both compression and robustness of deep learning models. EED constructs an ensemble of sub-models using different pruning importance scores to enhance adversarial robustness while reducing model size and computational cost. By addressing information loss and limited model diversity caused by pruning with ensemble methods and regularization, EED outperforms existing AP methods in accuracy and robustness. Experimental results show superior resilience of EED against various attacks on the CIFAR-10 and SVHN, maintaining inference efficiency. EED also proves its applicability in diverse neural network architectures and resource-constrained environments. + +# Acknowledgment + +This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.RS-2022-00155915 (Artificial Intelligence Convergence Innovation Human Resources Development (Inha University)) and No. 2021-0-02068 (AI Innovation Hub)), and this work was partly supported by the National Research Foundation of Korea (NRF) grant funded by the MSIT (No. 2022R1A2C2010095). + +# References + +[1] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE symposium on security and privacy (sp), pages 39-57. IEEE, 2017. 2, 7 +[2] Tianlong Chen, Zhenyu Zhang, Pengjun Wang, Santosh Balachandra, Haoyu Ma, Zehao Wang, and Zhangyang Wang. Sparsity winning twice: Better robust generalization from more efficient training. arXiv preprint arXiv:2202.09844, 2022. 7 +[3] Wenlin Chen, James Wilson, Stephen Tyree, Kilian Weinberger, and Yixin Chen. Compressing neural networks with the hashing trick. In International conference on machine learning, pages 2285-2294. PMLR, 2015. 1 +[4] Wuyang Chen, Ziyu Jiang, Zhangyang Wang, Kexin Cui, and Xiaoning Qian. Collaborative global-local networks for memory-efficient segmentation of ultra-high resolution images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8924-8933, 2019. 1 +[5] Hao Cheng, Kaidi Xu, Zhengang Li, Pu Zhao, Chenan Wang, Xue Lin, Bhavya Kailkhura, and Ryan Goldhahn. More or less (mol): Defending against multiple perturbation attacks on deep neural networks through model ensemble and compression. In 2022 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW), pages 645-655. IEEE, 2022. 2 +[6] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 2, 7 +[7] Yian Deng and Tingting Mu. Understanding and improving ensemble adversarial defense. NeurIPS, 36, 2024. 2, 4, 5 +[8] Misha Denil, Babak Shakibi, Laurent Dinh, Marc'Aurelio Ranzato, and Nando De Freitas. Predicting parameters in deep learning. Advances in neural information processing systems, 26, 2013. 1 +[9] Guneet S Dhillon, Kamyar Azizzadenesheli, Zachary C Lipton, Jeremy Bernstein, Jean Kossaifi, Aran Khanna, and Anima Anandkumar. Stochastic activation pruning for robust adversarial defense. arXiv preprint arXiv:1803.01442, 2018. 2 +[10] Yonggan Fu, Qixuan Yu, Yang Zhang, Shang Wu, Xu Ouyang, David Cox, and Yingyan Lin. Drawing robust scratch tickets: Subnetworks with inborn robustness are + +found within randomly initialized networks. Advances in Neural Information Processing Systems, 34:13059-13072, 2021. 7 +[11] Ian Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In International Conference on Learning Representations, 2015. 1, 2 +[12] Yiwen Guo, Chao Zhang, Changshui Zhang, and Yurong Chen. Sparse dnns with improved adversarial robustness. Advances in neural information processing systems, 31, 2018. 2 +[13] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. 2 +[14] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. Advances in neural information processing systems, 28, 2015. 1 +[15] Warren He, James Wei, Xinyun Chen, Nicholas Carlini, and Dawn Song. Adversarial example defense: Ensembles of weak defenses are not strong. In 11th USENIX workshop on offensive technologies (WOOT 17), 2017. 2 +[16] Tong Jian, Zifeng Wang, Yanzhi Wang, Jennifer Dy, and Stratis Ioannidis. Pruning adversarially robust neural networks without adversarial examples. In 2022 IEEE International Conference on Data Mining (ICDM), pages 993-998. IEEE, 2022. 2 +[17] Sanjay Kariyappa and Moinuddin K Qureshi. Improving adversarial robustness of ensembles with diversity training. arXiv preprint arXiv:1901.09981, 2019. 2 +[18] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 1 +[19] Souvik Kundu, Mahdi Nazemi, Peter A Beerel, and Massoud Pedram. Dnr: A tunable robust pruning framework through dynamic network rewiring of dnns. In Proceedings of the 26th Asia and South Pacific Design Automation Conference, pages 344–350, 2021. 2 +[20] Alexey Kurakin, Ian Goodfellow, and Samy Bengio. Adversarial machine learning at scale. arXiv preprint arXiv:1611.01236, 2016. 1, 2 +[21] Byung-Kwan Lee, Junho Kim, and Yong Man Ro. Masking adversarial damage: Finding adversarial saliency for robust and sparse network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15126-15136, 2022. 3, 7 +[22] Bai Li, Shiqi Wang, Yunhan Jia, Yantao Lu, Zhenyu Zhong, Lawrence Carin, and Suman Jana. Towards practical lottery ticket hypothesis for adversarial training. arXiv preprint arXiv:2003.05733, 2020. 2 +[23] Chenhao Li, Qiang Qiu, Zhibin Zhang, Jiafeng Guo, and Xueqi Cheng. Learning adversarially robust sparse networks via weight reparameterization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8527-8535, 2023. 7 +[24] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. arXiv preprint arXiv:1608.08710, 2016. 1 + +[25] Ziyue Li, Kan Ren, Yifan Yang, Xinyang Jiang, Yuqing Yang, and Dongsheng Li. Towards inference efficient deep ensemble learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8711-8719, 2023. 2, 6 +[26] Ningyi Liao, Shufan Wang, Liyao Xiang, Nanyang Ye, Shuo Shao, and Pengzhi Chu. Achieving adversarial robustness via sparsity. Machine Learning, pages 1-27, 2022. 2 +[27] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 1, 2, 7 +[28] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2574–2582, 2016. 7 +[29] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Anima Anandkumar. Diffusion models for adversarial purification. arXiv preprint arXiv:2205.07460, 2022. 2 +[30] Tianyu Pang, Kun Xu, Chao Du, Ning Chen, and Jun Zhu. Improving adversarial robustness via promoting ensemble diversity. In International Conference on Machine Learning, pages 4970-4979. PMLR, 2019. 2, 4 +[31] Nicolas Papernot, Patrick McDaniel, Xi Wu, Somesh Jha, and Ananthram Swami. Distillation as a defense to adversarial perturbations against deep neural networks. In 2016 IEEE symposium on security and privacy (SP), pages 582-597. IEEE, 2016. 1 +[32] Giorgio Piras, Maura Pintor, Ambra Demontis, Battista Biggio, Giorgio Giacinto, and Fabio Roli. Adversarial pruning: A survey and benchmark of pruning methods for adversarial robustness. arXiv preprint arXiv:2409.01249, 2024. 2 +[33] J Redmon. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2016. 1 +[34] Vikash Sehwag, Shiqi Wang, Prateek Mittal, and Suman Jana. Hydra: Pruning adversarially robust neural networks. Advances in Neural Information Processing Systems, 33: 19655-19666, 2020. 1, 3, 7 +[35] Ilia Shumailov, Yiren Zhao, Robert Mullins, and Ross Anderson. To compress or not to compress: Understanding the interactions between adversarial attacks and neural network compression. Proceedings of Machine Learning and Systems, 1:230-240, 2019. 1 +[36] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. Advances in neural information processing systems, 27, 2014. 1 +[37] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations, 2014. 1, 2 +[38] Lukas Timpl, Rahim Entezari, Hanie Sedghi, Behnam Neyshabur, and Olga Saukh. Understanding the effect of sparsity on neural networks robustness. arXiv preprint arXiv:2206.10915, 2022. 2 + +[39] Florian Tramér, Alexey Kurakin, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. Ensemble adversarial training: Attacks and defenses. In International Conference on Learning Representations, 2018. 2 +[40] Xiaofang Wang, Dan Kondratyuk, Eric Christiansen, Kris M Kitani, Yair Movshovitz-Attias, and Elad Eban. Wisdom of committees: An overlooked approach to faster and more accurate models. In International Conference on Learning Representations, 2021. 6 +[41] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In International conference on learning representations, 2019. 1, 2 +[42] Yue Wang, Jianghao Shen, Ting-Kuei Hu, Pengfei Xu, Tan Nguyen, Richard Baraniuk, Zhangyang Wang, and Yingyan Lin. Dual dynamic inference: Enabling more efficient, adaptive, and controllable deep inference. IEEE Journal of Selected Topics in Signal Processing, 14(4):623-633, 2020. 1 +[43] Xiao Wei, Yao Zhu, and Shu-Tao Xia. Batch normalization assisted adversarial pruning: Towards lightweight, sparse and robust models. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2021. 3 +[44] Boxi Wu, Jinghui Chen, Deng Cai, Xiaofei He, and Quanquan Gu. Do wider neural networks really help adversarial robustness? Advances in Neural Information Processing Systems, 34:7054-7067, 2021. 1 +[45] Dongxian Wu, Shu-Tao Xia, and Yisen Wang. Adversarial weight perturbation helps robust generalization. Advances in neural information processing systems, 33:2958-2969, 2020. 2 +[46] Yanzhao Wu, Ling Liu, Zhongwei Xie, Ka-Ho Chow, and Wenqi Wei. Boosting ensemble accuracy by revisiting ensemble diversity metrics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16469-16477, 2021. 2 +[47] Huanrui Yang, Jingyang Zhang, Hongliang Dong, Nathan Inkawhich, Andrew Gardner, Andrew Touchet, Wesley Wilkes, Heath Berry, and Hai Li. Diverse: diversifying vulnerabilities for enhanced robust generation of ensembles. Advances in Neural Information Processing Systems, 33: 5505-5515, 2020. 2, 5 +[48] Shaokai Ye, Kaidi Xu, Sijia Liu, Hao Cheng, Jan-Henrik Lambrechts, Huan Zhang, Aojun Zhou, Kaisheng Ma, Yanzhi Wang, and Xue Lin. Adversarial robustness vs. model compression, or both? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 111-120, 2019. 1, 2, 7 +[49] Jongmin Yoon, Sung Ju Hwang, and Juho Lee. Adversarial purification with score-based generative models. In International Conference on Machine Learning, pages 12062-12072. PMLR, 2021. 2 +[50] Ruichi Yu, Ang Li, Chun-Fu Chen, Jui-Hsin Lai, Vlad I Morariu, Xintong Han, Mingfei Gao, Ching-Yung Lin, and Larry S Davis. Nisp: Pruning networks using neuron importance score propagation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9194-9203, 2018. 2, 3, 4 + +[51] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In International conference on machine learning, pages 7472-7482. PMLR, 2019. 1 +[52] Qi Zhao and Christian Wressnegger. Holistic adversarially robust pruning. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 3, 7 \ No newline at end of file diff --git a/2025/Two is Better than One_ Efficient Ensemble Defense for Robust and Compact Models/images.zip b/2025/Two is Better than One_ Efficient Ensemble Defense for Robust and Compact Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1656fa1230cdb471c232757c5c3bfede1a8465bf --- /dev/null +++ b/2025/Two is Better than One_ Efficient Ensemble Defense for Robust and Compact Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e91b09057d42056c3b289de6cdb673fcf935dc9741a6c620fa1887c9385fae0a +size 457317 diff --git a/2025/Two is Better than One_ Efficient Ensemble Defense for Robust and Compact Models/layout.json b/2025/Two is Better than One_ Efficient Ensemble Defense for Robust and Compact Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1623ca930339c0f56d94646ecf8a8ffe0e7d965c --- /dev/null +++ b/2025/Two is Better than One_ Efficient Ensemble Defense for Robust and Compact Models/layout.json @@ -0,0 +1,11758 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 119, + 103, + 492, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 103, + 492, + 140 + ], + "spans": [ + { + "bbox": [ + 119, + 103, + 492, + 140 + ], + "type": "text", + "content": "Two is Better than One: Efficient Ensemble Defense for Robust and Compact Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 222, + 162, + 389, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 162, + 389, + 175 + ], + "spans": [ + { + "bbox": [ + 222, + 162, + 389, + 175 + ], + "type": "text", + "content": "Yoojin Jung Byung Cheol Song" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 176, + 471, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 176, + 471, + 203 + ], + "spans": [ + { + "bbox": [ + 138, + 176, + 471, + 203 + ], + "type": "text", + "content": "Department of Electrical and Computer Engineering, Inha University Incheon, Republic of Korea" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 198, + 205, + 408, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 205, + 408, + 217 + ], + "spans": [ + { + "bbox": [ + 198, + 205, + 408, + 217 + ], + "type": "text", + "content": "yoon0505@gmail.com, bcsong@inha.ac.kr" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 245, + 200, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 245, + 200, + 257 + ], + "spans": [ + { + "bbox": [ + 152, + 245, + 200, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 271, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 296, + 510 + ], + "type": "text", + "content": "Deep learning-based computer vision systems adopt complex and large architectures to improve performance, yet they face challenges in deployment on resource-constrained mobile and edge devices. To address this issue, model compression techniques such as pruning, quantization, and matrix factorization have been proposed; however, these compressed models are often highly vulnerable to adversarial attacks. We introduce the Efficient Ensemble Defense (EED) technique, which diversifies the compression of a single base model based on different pruning importance scores and enhances ensemble diversity to achieve high adversarial robustness and resource efficiency. EED dynamically determines the number of necessary sub-models during the inference stage, minimizing unnecessary computations while maintaining high robustness. On the CIFAR-10 and SVHN datasets, EED demonstrated state-of-the-art robustness performance compared to existing adversarial pruning techniques, along with an inference speed improvement of up to 1.86 times. This proves that EED is a powerful defense solution in resource-constrained environments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 536, + 135, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 536, + 135, + 549 + ], + "spans": [ + { + "bbox": [ + 56, + 536, + 135, + 549 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": "Convolutional Neural Networks (CNNs) have shown remarkable performance across various computer vision tasks, such as image classification, object detection, and semantic segmentation[18, 33, 36]. However, CNNs have become increasingly wider and deeper to achieve higher performance, leading to complex architectures with vast numbers of parameters. This complexity inherently results in significant computational costs and storage demands, making deployment on resource-constrained edge devices challenging[4, 42]. To address this challenge, a number of model compression techniques, including pruning[14, 24], quantization[3], and matrix decomposition[8], have been proposed to minimize model size and computational load" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 326, + 247, + 408, + 331 + ], + "blocks": [ + { + "bbox": [ + 326, + 247, + 408, + 331 + ], + "lines": [ + { + "bbox": [ + 326, + 247, + 408, + 331 + ], + "spans": [ + { + "bbox": [ + 326, + 247, + 408, + 331 + ], + "type": "image", + "image_path": "e096d87d155417cb0031f40ba099c84fc66be49dfab080bf834e915603e72821.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 421, + 247, + 494, + 331 + ], + "blocks": [ + { + "bbox": [ + 421, + 247, + 494, + 331 + ], + "lines": [ + { + "bbox": [ + 421, + 247, + 494, + 331 + ], + "spans": [ + { + "bbox": [ + 421, + 247, + 494, + 331 + ], + "type": "image", + "image_path": "c94bb48dc532955dfc01c11083e4d6574dbb24eb823279cda220acde293248c2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 489, + 281, + 536, + 319 + ], + "lines": [ + { + "bbox": [ + 489, + 281, + 536, + 319 + ], + "spans": [ + { + "bbox": [ + 489, + 281, + 536, + 319 + ], + "type": "text", + "content": "Important nodes or connections may be removed incorrectly due to adversarial attack" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 373, + 335, + 433, + 420 + ], + "blocks": [ + { + "bbox": [ + 373, + 335, + 433, + 420 + ], + "lines": [ + { + "bbox": [ + 373, + 335, + 433, + 420 + ], + "spans": [ + { + "bbox": [ + 373, + 335, + 433, + 420 + ], + "type": "image", + "image_path": "00a49362902aad1dadfca3eaa37b44ed40e9fc49092ecbf60a09778070a0a78a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 376, + 473, + 401 + ], + "lines": [ + { + "bbox": [ + 440, + 376, + 473, + 401 + ], + "spans": [ + { + "bbox": [ + 440, + 376, + 473, + 401 + ], + "type": "text", + "content": "Ensemble could help compensate" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 313, + 426, + 555, + 494 + ], + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 494 + ], + "type": "text", + "content": "Figure 1. Challenges in AP and the potential of ensemble methods: AP weakens the network by removing critical parameters, reducing robustness and increasing vulnerability to adversarial attacks. An ensemble approach mitigates pruning effects by combining pruned models, enhancing robustness through their collective strength (illustrated in green)." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 518, + 448, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 448, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 448, + 531 + ], + "type": "text", + "content": "without compromising accuracy." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 534, + 556, + 715 + ], + "type": "text", + "content": "Meanwhile, deep learning models are highly vulnerable to adversarial attacks that induce false predictions by disturbing the input image with small perturbations[11, 37]. This vulnerability is particularly pronounced in sparsely compressed models, which are even more sensitive to adversarial attacks than standard models[35, 44]. Consequently, common adversarial defense techniques such as adversarial training (AT) [20, 27], label smoothing [41, 51], and defensive distillation[31], while generally effective, often experience a degradation in performance when applied to compressed models. This indicates that existing compression techniques may compromise adversarial robustness even if they maintain accuracy on clean test sets. In response, adversarial pruning (AP), a model compression technique that considers adversarial robustness[34, 48, 52], has emerged;" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9696" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "however, AP still exhibits lower robustness compared to non-compressed defense models[32]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 100, + 295, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 100, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 54, + 100, + 295, + 232 + ], + "type": "text", + "content": "We assume that this degradation in robustness arises from reduced generalization capability and information loss caused by pruning. Specifically, adversarial attacks may cause the model to misidentify and even remove critical nodes or connections during the pruning process. In Sec.3.3, we validate this assumption by observing importance scores, which reflect how each component contributes to the loss function through the gradients of the weights[50]. Our findings demonstrate that adversarial examples significantly impact the importance scores of individual parameters." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 236, + 296, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 236, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 236, + 296, + 403 + ], + "type": "text", + "content": "Ensemble defense[7, 39] has recently gained attention in adversarial defense research for enhancing the stability and robustness of multiple base classifiers against attacks. As illustrated in Fig.1, we believe that ensemble defense methods can help mitigate generalization degradation and information loss during the pruning process. Since ensemble learning can compensate for deficiencies in individual compressed models due to removed parameters by other models and offset individual errors by combining outputs. However, traditional ensemble defenses inherently suffer from a large capacity, as they require multiple base models[15, 30]. This raises an important question: Is it possible to develop an ensemble defense algorithm that not only achieves robustness but also results in a compact compressed model?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 407, + 296, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 407, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 407, + 296, + 586 + ], + "type": "text", + "content": "As an answer to this, we propose Efficient Ensemble Defense (EED). Instead of using multiple base models, we compress a single base model by employing different pruning metrics and then utilize these compressed sub-models in the ensemble. This approach reduces the overall capacity compared to traditional ensemble defense models that rely on multiple base models. Given that using smaller models in an ensemble can lead to reduced defensive efficacy[46], we enhance the variety of the compressed models by incorporating a diversity term in the ensemble selection process, thereby improving ensemble effectiveness while maintaining robustness. Additionally, by adopting inference-efficient ensemble[25], we dynamically add sub-networks and perform ensemble operations as needed during inference, improving overall efficiency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 590, + 274, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 590, + 274, + 601 + ], + "spans": [ + { + "bbox": [ + 67, + 590, + 274, + 601 + ], + "type": "text", + "content": "The main contributions of our work are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 605, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 54, + 605, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 653 + ], + "type": "text", + "content": "- Propose EED, that overcomes the limitations of traditional AP by enhancing the ensemble diversity and compensating for individual errors, thereby achieving robustness against adversarial attacks while being compact." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 654, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 654, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 654, + 295, + 689 + ], + "type": "text", + "content": "- Reduce memory by using a single base model in ensemble, and decrease computational costs by dynamically adding models only when necessary during inference." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 689, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 689, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 54, + 689, + 295, + 712 + ], + "type": "text", + "content": "- Achieve state-of-the-art performance across various attack benchmarks on the CIFAR-10 and SVHN datasets." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 71, + 406, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 406, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 406, + 83 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 91, + 487, + 103 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 487, + 103 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 487, + 103 + ], + "type": "text", + "content": "2.1. Adversarial Attack and Defense" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 109, + 553, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 553, + 205 + ], + "type": "text", + "content": "Adversarial attacks exploit the inherent vulnerabilities of deep learning models by adding specific noise to the input, thereby inducing incorrect predictions. These specially crafted inputs, designed to deceive deep learning models, are called adversarial examples. Early studies highlighted the susceptibility of deep learning models to adversarial examples[37]. Then, more potent attacks such as FGSM[11], PGD[27], and AutoAttack[6] have emerged." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 205, + 553, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 313 + ], + "type": "text", + "content": "On the other hand, defense methods against attacks have also been evolved. The most well-known adversarial training (AT)[1, 20, 27] enhanced robustness by training adversarial examples. Other strategies involve regularizing geometric properties such as the curvature of the loss function [41, 45], or restoring potentially attacked images to purified versions[29, 49]. Adversarial pruning (AP) and ensemble defense are also one of adversarial defenses (details are given in Sec.3.1.)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 322, + 434, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 434, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 434, + 335 + ], + "type": "text", + "content": "2.2. Adversarial Pruning" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 340, + 555, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 532 + ], + "type": "text", + "content": "AP is a pruning technique (primarily weight pruning) that considers the adversarial context, aiming to reduce the complexity of neural networks while preserving robustness against adversarial attacks. HYDRA demonstrated that empirical adversarial robustness can be achieved through a lowest weight magnitude (LWM) pruning method. R-ADMM[48] and HARP[52] incorporated the alternating direction method of multipliers (ADMM) framework into LWM-based pruning to enhance robustness. In some studies [12, 22, 26], a traditional pruning method, i.e., the Lottery Ticket Hypothesis, was integrated with AT. However, they all failed to maintain generalized robustness at higher sparsity rates " + }, + { + "bbox": [ + 313, + 340, + 555, + 532 + ], + "type": "inline_equation", + "content": "(s_r)[13, 19, 38]" + }, + { + "bbox": [ + 313, + 340, + 555, + 532 + ], + "type": "text", + "content": ", and struggled to defend against adaptive attacks[9, 16]. This suggests that conventional AP has not fully resolved the trade-off between model capacity and robustness." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 540, + 422, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 422, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 422, + 552 + ], + "type": "text", + "content": "2.3. Ensemble Defense" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "Ensemble defense (ED), which applies ensemble learning to adversarial defense, has been actively studied due to its promising performance[7, 39]. The core of ED is to improve predictions through an ensemble diversity loss that considers adversarial robustness during training. ADP[30] used Shannon entropy and geometric diversity for uncertainty, while GAL[17] utilized cosine similarities between model loss gradients to formalize diversity. DVERGE[47] introduced vulnerability diversity to ensure that each model is robust to the weaknesses of others, and MOL[5] increased perturbation diversity by using expert models trained on different attacks. Thus, ED has significantly enhanced robustness by reducing the risk of overfitting to adversarial ex" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9697" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "amples and diversifying the decision boundary, though with increased model capacity and computational costs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 106, + 141, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 106, + 141, + 119 + ], + "spans": [ + { + "bbox": [ + 55, + 106, + 141, + 119 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 126, + 124, + 137 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 126, + 124, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 126, + 124, + 137 + ], + "type": "text", + "content": "3.1. Notations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 143, + 296, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 143, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 143, + 296, + 251 + ], + "type": "text", + "content": "Adversarial Attack Let " + }, + { + "bbox": [ + 55, + 143, + 296, + 251 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 143, + 296, + 251 + ], + "type": "text", + "content": " be an input example, " + }, + { + "bbox": [ + 55, + 143, + 296, + 251 + ], + "type": "inline_equation", + "content": "y \\in \\{1, \\dots, C\\}" + }, + { + "bbox": [ + 55, + 143, + 296, + 251 + ], + "type": "text", + "content": " the true label, and " + }, + { + "bbox": [ + 55, + 143, + 296, + 251 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 55, + 143, + 296, + 251 + ], + "type": "text", + "content": " a classifier with parameters " + }, + { + "bbox": [ + 55, + 143, + 296, + 251 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 143, + 296, + 251 + ], + "type": "text", + "content": ". The classifier's prediction for input " + }, + { + "bbox": [ + 55, + 143, + 296, + 251 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 143, + 296, + 251 + ], + "type": "text", + "content": " is given by " + }, + { + "bbox": [ + 55, + 143, + 296, + 251 + ], + "type": "inline_equation", + "content": "f(x) = \\arg \\max_{i \\in \\{1, \\dots, C\\}} f_i(x)" + }, + { + "bbox": [ + 55, + 143, + 296, + 251 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 143, + 296, + 251 + ], + "type": "inline_equation", + "content": "f_i(x)" + }, + { + "bbox": [ + 55, + 143, + 296, + 251 + ], + "type": "text", + "content": " denotes the predicted probability for class " + }, + { + "bbox": [ + 55, + 143, + 296, + 251 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 143, + 296, + 251 + ], + "type": "text", + "content": ". An adversarial example " + }, + { + "bbox": [ + 55, + 143, + 296, + 251 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 55, + 143, + 296, + 251 + ], + "type": "text", + "content": " is crafted by adding a perturbation " + }, + { + "bbox": [ + 55, + 143, + 296, + 251 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 143, + 296, + 251 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 143, + 296, + 251 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 143, + 296, + 251 + ], + "type": "text", + "content": ", aiming to maximize the model's prediction error while keeping " + }, + { + "bbox": [ + 55, + 143, + 296, + 251 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 55, + 143, + 296, + 251 + ], + "type": "text", + "content": " close to " + }, + { + "bbox": [ + 55, + 143, + 296, + 251 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 143, + 296, + 251 + ], + "type": "text", + "content": " under a norm constraint. Formally, an adversarial attack is defined by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 94, + 258, + 295, + 278 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 258, + 295, + 278 + ], + "spans": [ + { + "bbox": [ + 94, + 258, + 295, + 278 + ], + "type": "interline_equation", + "content": "x ^ {\\prime} = x + \\delta = \\arg \\max _ {\\| x - x ^ {\\prime} \\| < \\epsilon} \\mathcal {L} (f (x ^ {\\prime}), y) \\tag {1}", + "image_path": "1a59404a4bc15c4e8acbaf493ebaad07e0689b5f6b67d456dcb980a5c16d6fe4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 285, + 296, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 296, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 296, + 309 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 285, + 296, + 309 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 285, + 296, + 309 + ], + "type": "text", + "content": " is the loss function and " + }, + { + "bbox": [ + 55, + 285, + 296, + 309 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 285, + 296, + 309 + ], + "type": "text", + "content": " is the perturbation budget that constrains the strength of the attack under an " + }, + { + "bbox": [ + 55, + 285, + 296, + 309 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 55, + 285, + 296, + 309 + ], + "type": "text", + "content": "-norm." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 310, + 296, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 296, + 357 + ], + "type": "text", + "content": "Adversarial Training AT is a defense method that improves model robustness by training on adversarially perturbed samples alongside clean samples. The objective for AT is usually formulated by:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 94, + 364, + 295, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 364, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 94, + 364, + 295, + 392 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {(x, y) \\sim \\mathcal {D}} \\left[ \\max _ {\\delta : \\| \\delta \\| _ {p} \\leq \\epsilon} \\mathcal {L} \\left(f \\left(x ^ {\\prime}; \\theta\\right), y\\right) \\right] \\tag {2}", + "image_path": "129c31f83e4cb48e3ec29010718d5a1963b6b29201013f13d4fe7a98867354ee.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 397, + 188, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 188, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 188, + 408 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 397, + 188, + 408 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 397, + 188, + 408 + ], + "type": "text", + "content": " is the data distribution." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 409, + 296, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 296, + 480 + ], + "type": "text", + "content": "Adversarial Pruning Given a pre-trained model " + }, + { + "bbox": [ + 55, + 409, + 296, + 480 + ], + "type": "inline_equation", + "content": "f(\\theta)" + }, + { + "bbox": [ + 55, + 409, + 296, + 480 + ], + "type": "text", + "content": " under AT, we aim to maintain robustness while applying a pruning mask " + }, + { + "bbox": [ + 55, + 409, + 296, + 480 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 409, + 296, + 480 + ], + "type": "text", + "content": " across layers to achieve compression. The general pruning is carried out by removing parameters until the network reaches a desired " + }, + { + "bbox": [ + 55, + 409, + 296, + 480 + ], + "type": "inline_equation", + "content": "s_r" + }, + { + "bbox": [ + 55, + 409, + 296, + 480 + ], + "type": "text", + "content": ". So, pruning involves creating " + }, + { + "bbox": [ + 55, + 409, + 296, + 480 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 409, + 296, + 480 + ], + "type": "text", + "content": ", formalized as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 92, + 489, + 295, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 489, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 92, + 489, + 295, + 509 + ], + "type": "interline_equation", + "content": "M = \\arg \\min _ {\\| \\mathbf {m} \\| _ {0} \\leq 1 - s _ {r}} \\mathcal {L} (f (\\theta) \\otimes m, x, y). \\tag {3}", + "image_path": "0221fa073e2e3cc2a80ccbe606a30e9fb64cd51aea74f1ede6ab91b595170df5.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 86, + 516, + 295, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 516, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 86, + 516, + 295, + 548 + ], + "type": "interline_equation", + "content": "m = \\left\\{ \\begin{array}{l l} 0, & \\text {i f} \\theta_ {i} \\text {i s t o b e p r u n e d} \\\\ 1, & \\text {o t h e r w i s e} \\end{array} , \\quad \\forall \\theta_ {i} \\in \\theta \\right. \\tag {4}", + "image_path": "3e33e8123de138ba5590c17c7e8e1d94a8a515906875ece5fd832547c5a6f449.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 551, + 296, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 296, + 597 + ], + "type": "text", + "content": "For AP, we follow the strategy from HARP[52]. To enhance adversarial robustness, inner maximization generates adversarial examples, targeting a worst-case loss under the PGD attacks:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 598, + 295, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 598, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 116, + 598, + 295, + 616 + ], + "type": "interline_equation", + "content": "\\max _ {\\delta} \\mathcal {L} _ {r} (f (\\theta) \\odot M, x + \\delta , y) \\tag {5}", + "image_path": "6c78faf0436e985b66b6442a43cda355b2f90ff6cfd3611bf393e1fe91d0e80d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 619, + 296, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 619, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 55, + 619, + 296, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 619, + 296, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 55, + 619, + 296, + 667 + ], + "type": "text", + "content": " defines the robust loss function, incorporating attack-specific parameters. And the global non-uniform pruning strategy jointly optimizes a compression rate " + }, + { + "bbox": [ + 55, + 619, + 296, + 667 + ], + "type": "inline_equation", + "content": "c_{r}" + }, + { + "bbox": [ + 55, + 619, + 296, + 667 + ], + "type": "text", + "content": " and importance scores " + }, + { + "bbox": [ + 55, + 619, + 296, + 667 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 619, + 296, + 667 + ], + "type": "text", + "content": " across layers. This is defined by:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 80, + 673, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 673, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 80, + 673, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {c _ {r}, I} \\mathbb {E} _ {(x, y) \\sim D} \\left[ \\max _ {\\delta} \\mathcal {L} _ {r} (f (\\theta) \\odot M, x + \\delta , y) \\right] \\tag {6} \\\\ + \\varphi \\cdot \\mathcal {L} _ {p} (f (\\theta) \\odot M, a _ {t}) \\\\ \\end{array}", + "image_path": "4aaff51108820f6205a4f622f2d6ad8c0be795e5d969b02d746bec1d7c539da5.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": " enforces target compression " + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": " across all layers " + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": " to guide non-uniform pruning rates." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 105, + 463, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 105, + 463, + 118 + ], + "spans": [ + { + "bbox": [ + 314, + 105, + 463, + 118 + ], + "type": "text", + "content": "3.2. Pruning Importance Score" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 123, + 554, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 123, + 554, + 158 + ], + "spans": [ + { + "bbox": [ + 313, + 123, + 554, + 158 + ], + "type": "text", + "content": "The importance score's definition varies by pruning method. This section details its definition and computation, as discussed in Sec.3.1." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 159, + 554, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 159, + 554, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 554, + 232 + ], + "type": "text", + "content": "Neuron Importance Score (NIS) NIS [50] is a common pruning score that minimizes the weighted distance between the original and pruned final responses in a given layer. Given a neural network " + }, + { + "bbox": [ + 313, + 159, + 554, + 232 + ], + "type": "inline_equation", + "content": "f^{(n)}" + }, + { + "bbox": [ + 313, + 159, + 554, + 232 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 159, + 554, + 232 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 159, + 554, + 232 + ], + "type": "text", + "content": " layers, and the importance score of the final layer response " + }, + { + "bbox": [ + 313, + 159, + 554, + 232 + ], + "type": "inline_equation", + "content": "I^{(n)}" + }, + { + "bbox": [ + 313, + 159, + 554, + 232 + ], + "type": "text", + "content": ", the importance score of the " + }, + { + "bbox": [ + 313, + 159, + 554, + 232 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 159, + 554, + 232 + ], + "type": "text", + "content": "-th layer " + }, + { + "bbox": [ + 313, + 159, + 554, + 232 + ], + "type": "inline_equation", + "content": "I^{(i)}" + }, + { + "bbox": [ + 313, + 159, + 554, + 232 + ], + "type": "text", + "content": " is calculated as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 341, + 240, + 554, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 240, + 554, + 258 + ], + "spans": [ + { + "bbox": [ + 341, + 240, + 554, + 258 + ], + "type": "interline_equation", + "content": "I ^ {(i)} = \\left| w ^ {(i + 1)} \\right| ^ {T} \\cdot \\left| w ^ {(i + 2)} \\right| ^ {T} \\dots \\left| w ^ {(n)} \\right| ^ {T} \\cdot I ^ {(n)} \\tag {7}", + "image_path": "05b66a155802963c6c50912e5853850b9b4b8abb89c16d04347c566e415836a7.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 267, + 554, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 267, + 554, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 554, + 326 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 267, + 554, + 326 + ], + "type": "inline_equation", + "content": "w^{(i)}" + }, + { + "bbox": [ + 313, + 267, + 554, + 326 + ], + "type": "text", + "content": " refers to the weight matrix of the " + }, + { + "bbox": [ + 313, + 267, + 554, + 326 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 267, + 554, + 326 + ], + "type": "text", + "content": "-th layer. In this case, the importance of a neuron in the network can be computed recursively along the network. Thus, " + }, + { + "bbox": [ + 313, + 267, + 554, + 326 + ], + "type": "inline_equation", + "content": "I^{(i)}" + }, + { + "bbox": [ + 313, + 267, + 554, + 326 + ], + "type": "text", + "content": " can be propagated from the importance score of the " + }, + { + "bbox": [ + 313, + 267, + 554, + 326 + ], + "type": "inline_equation", + "content": "(i + 1)" + }, + { + "bbox": [ + 313, + 267, + 554, + 326 + ], + "type": "text", + "content": "-th layer as follow:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 380, + 326, + 553, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 326, + 553, + 343 + ], + "spans": [ + { + "bbox": [ + 380, + 326, + 553, + 343 + ], + "type": "interline_equation", + "content": "I _ {N I S} ^ {(i)} = \\left| w ^ {(i + 1)} \\right| ^ {T} \\cdot I ^ {(i + 1)} \\tag {8}", + "image_path": "c48cdb2e76882614a724a7ce5bab4694154c0a1ce4a3d2151b8478220896817a.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 350, + 554, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 554, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 554, + 397 + ], + "type": "text", + "content": "Empirical Risk Minimization (ERM) The importance score based on ERM, used in HYDRA[34] and HARP, is defined as follows, based on the weight magnitudes of a pre-trained model." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 389, + 407, + 554, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 407, + 554, + 435 + ], + "spans": [ + { + "bbox": [ + 389, + 407, + 554, + 435 + ], + "type": "interline_equation", + "content": "I _ {E R M} ^ {(i)} = \\frac {\\eta \\cdot | \\theta^ {(i)} |}{\\max (| \\theta^ {(i)} |)} \\tag {9}", + "image_path": "45273e8618b3aae33b726b2f9f35db7b56d940d101077a5d29bcbed67fdb7abd.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 446, + 554, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 446, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 554, + 495 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 446, + 554, + 495 + ], + "type": "inline_equation", + "content": "I_{ERM}^{(i)}" + }, + { + "bbox": [ + 313, + 446, + 554, + 495 + ], + "type": "text", + "content": " represents the importance score for the weights of the " + }, + { + "bbox": [ + 313, + 446, + 554, + 495 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 446, + 554, + 495 + ], + "type": "text", + "content": "-th layer, " + }, + { + "bbox": [ + 313, + 446, + 554, + 495 + ], + "type": "inline_equation", + "content": "\\theta^{(i)}" + }, + { + "bbox": [ + 313, + 446, + 554, + 495 + ], + "type": "text", + "content": " denotes the weight matrix of that layer, and " + }, + { + "bbox": [ + 313, + 446, + 554, + 495 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 446, + 554, + 495 + ], + "type": "text", + "content": " is a scaling factor determined by the layer's structure and receptive field size." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 496, + 554, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 554, + 555 + ], + "type": "text", + "content": "Adversarial Saliency Estimation (ASE) MAD [21] computes the ASE for each parameter, removing the least salient ones. Adversarial saliency is estimated via a second-order Taylor expansion, with the saliency score " + }, + { + "bbox": [ + 313, + 496, + 554, + 555 + ], + "type": "inline_equation", + "content": "I_{ASE}" + }, + { + "bbox": [ + 313, + 496, + 554, + 555 + ], + "type": "text", + "content": " for each parameter " + }, + { + "bbox": [ + 313, + 496, + 554, + 555 + ], + "type": "inline_equation", + "content": "w^{(i)}" + }, + { + "bbox": [ + 313, + 496, + 554, + 555 + ], + "type": "text", + "content": " defined as:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 372, + 566, + 553, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 566, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 372, + 566, + 553, + 594 + ], + "type": "interline_equation", + "content": "I _ {A S E} = \\frac {1}{2} \\left[ \\frac {\\partial^ {2} \\mathcal {L}}{\\partial \\left(w ^ {(i)}\\right) ^ {2}} \\right] \\left| w ^ {(i)} \\right| ^ {2} \\tag {10}", + "image_path": "e6269e5f0833b900c955cf8621bda3496d8b01219d24496236d128744141e98f.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 599, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 554, + 653 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 599, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\frac{\\partial^2\\mathcal{L}}{\\partial(w^{(i)})^2}" + }, + { + "bbox": [ + 313, + 599, + 554, + 653 + ], + "type": "text", + "content": " is the Hessian of the loss function with respect to " + }, + { + "bbox": [ + 313, + 599, + 554, + 653 + ], + "type": "inline_equation", + "content": "w^{(i)}" + }, + { + "bbox": [ + 313, + 599, + 554, + 653 + ], + "type": "text", + "content": ", reflecting the impact of parameter changes on the loss, and " + }, + { + "bbox": [ + 313, + 599, + 554, + 653 + ], + "type": "inline_equation", + "content": "|w^{(i)}|^2" + }, + { + "bbox": [ + 313, + 599, + 554, + 653 + ], + "type": "text", + "content": " reflects the magnitude and importance of the corresponding parameter." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": "Batch Normalization Scaling Factor (BNSF) BNAP [43] identifies the magnitude of the Batch Normalization (BN) scaling factor " + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": " as a robustness measure across channels, used as the pruning criterion. The BN layer normalizes weights " + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "inline_equation", + "content": "w_{ik}^{j}" + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": " into the following effective weights:" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9698" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 77, + 111, + 148 + ], + "blocks": [ + { + "bbox": [ + 59, + 77, + 111, + 148 + ], + "lines": [ + { + "bbox": [ + 59, + 77, + 111, + 148 + ], + "spans": [ + { + "bbox": [ + 59, + 77, + 111, + 148 + ], + "type": "image", + "image_path": "f29e9f634c90f9504fc5a3607953c9fcfe767ca235f95fd6361906a96313ed95.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 157, + 94, + 165 + ], + "lines": [ + { + "bbox": [ + 75, + 157, + 94, + 165 + ], + "spans": [ + { + "bbox": [ + 75, + 157, + 94, + 165 + ], + "type": "text", + "content": "Clean" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 172, + 233, + 182 + ], + "lines": [ + { + "bbox": [ + 62, + 172, + 233, + 182 + ], + "spans": [ + { + "bbox": [ + 62, + 172, + 233, + 182 + ], + "type": "text", + "content": "(a) Importance score of models with different inputs." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 121, + 79, + 172, + 148 + ], + "blocks": [ + { + "bbox": [ + 121, + 79, + 172, + 148 + ], + "lines": [ + { + "bbox": [ + 121, + 79, + 172, + 148 + ], + "spans": [ + { + "bbox": [ + 121, + 79, + 172, + 148 + ], + "type": "image", + "image_path": "d37ec6b339baabc2e9253d3599ee026f441f7201f270abb0495a89d8dfca76d0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 157, + 157, + 165 + ], + "lines": [ + { + "bbox": [ + 137, + 157, + 157, + 165 + ], + "spans": [ + { + "bbox": [ + 137, + 157, + 157, + 165 + ], + "type": "text", + "content": "Attack" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 186, + 79, + 237, + 148 + ], + "blocks": [ + { + "bbox": [ + 186, + 79, + 237, + 148 + ], + "lines": [ + { + "bbox": [ + 186, + 79, + 237, + 148 + ], + "spans": [ + { + "bbox": [ + 186, + 79, + 237, + 148 + ], + "type": "image", + "image_path": "5dc2ad8b079621336b4c10967e0de735dca9a74038f9c54833465f715dce03b4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 157, + 216, + 165 + ], + "lines": [ + { + "bbox": [ + 206, + 157, + 216, + 165 + ], + "spans": [ + { + "bbox": [ + 206, + 157, + 216, + 165 + ], + "type": "text", + "content": "AT" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 250, + 79, + 301, + 148 + ], + "blocks": [ + { + "bbox": [ + 250, + 79, + 301, + 148 + ], + "lines": [ + { + "bbox": [ + 250, + 79, + 301, + 148 + ], + "spans": [ + { + "bbox": [ + 250, + 79, + 301, + 148 + ], + "type": "image", + "image_path": "19b4f980a11e32b48926026ccdfdde77504c4ba7fd8131964e9fb88b2655a1b3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 269, + 157, + 282, + 165 + ], + "lines": [ + { + "bbox": [ + 269, + 157, + 282, + 165 + ], + "spans": [ + { + "bbox": [ + 269, + 157, + 282, + 165 + ], + "type": "text", + "content": "NIS" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 299, + 172, + 501, + 182 + ], + "lines": [ + { + "bbox": [ + 299, + 172, + 501, + 182 + ], + "spans": [ + { + "bbox": [ + 299, + 172, + 501, + 182 + ], + "type": "text", + "content": "(b) Importance score of a model with different scoring metrics." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 313, + 79, + 364, + 148 + ], + "blocks": [ + { + "bbox": [ + 313, + 79, + 364, + 148 + ], + "lines": [ + { + "bbox": [ + 313, + 79, + 364, + 148 + ], + "spans": [ + { + "bbox": [ + 313, + 79, + 364, + 148 + ], + "type": "image", + "image_path": "119d785a023bdc73f48944695313f05eebc553d09226bafb5bb78ed28c92dbc8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 157, + 347, + 165 + ], + "lines": [ + { + "bbox": [ + 329, + 157, + 347, + 165 + ], + "spans": [ + { + "bbox": [ + 329, + 157, + 347, + 165 + ], + "type": "text", + "content": "ERM" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 376, + 79, + 427, + 148 + ], + "blocks": [ + { + "bbox": [ + 376, + 79, + 427, + 148 + ], + "lines": [ + { + "bbox": [ + 376, + 79, + 427, + 148 + ], + "spans": [ + { + "bbox": [ + 376, + 79, + 427, + 148 + ], + "type": "image", + "image_path": "cb9442754b7db8f627c132e968293ccfd02dce31a9f95617a085aac4cbf33004.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 393, + 157, + 408, + 165 + ], + "lines": [ + { + "bbox": [ + 393, + 157, + 408, + 165 + ], + "spans": [ + { + "bbox": [ + 393, + 157, + 408, + 165 + ], + "type": "text", + "content": "ASE" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 439, + 79, + 490, + 148 + ], + "blocks": [ + { + "bbox": [ + 439, + 79, + 490, + 148 + ], + "lines": [ + { + "bbox": [ + 439, + 79, + 490, + 148 + ], + "spans": [ + { + "bbox": [ + 439, + 79, + 490, + 148 + ], + "type": "image", + "image_path": "24a63754435d4f7cf3449fe6efd3eb9f512c4ed00224f8f29374362324fa72a2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 454, + 157, + 474, + 165 + ], + "lines": [ + { + "bbox": [ + 454, + 157, + 474, + 165 + ], + "spans": [ + { + "bbox": [ + 454, + 157, + 474, + 165 + ], + "type": "text", + "content": "BNSF" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 506, + 79, + 536, + 160 + ], + "blocks": [ + { + "bbox": [ + 506, + 79, + 536, + 160 + ], + "lines": [ + { + "bbox": [ + 506, + 79, + 536, + 160 + ], + "spans": [ + { + "bbox": [ + 506, + 79, + 536, + 160 + ], + "type": "image", + "image_path": "bcf2920daf0c2c82ef10b0a1149d4593dd9f8c0cf51a6d2a814de4f535151ffc.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 191, + 553, + 214 + ], + "lines": [ + { + "bbox": [ + 55, + 191, + 553, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 553, + 214 + ], + "type": "text", + "content": "Figure 2. Importance score visualization for the 4-5-6-3 CNN model on the MNIST dataset: yellow/red indicate higher importance, green/blue indicate lower." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 119, + 239, + 295, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 239, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 119, + 239, + 295, + 270 + ], + "type": "interline_equation", + "content": "\\hat {w} _ {i k} ^ {j} = \\frac {\\gamma_ {i} ^ {j} w _ {i k} ^ {j}}{\\left(\\sigma_ {i} ^ {j}\\right) ^ {2} + \\epsilon} \\approx \\frac {\\gamma_ {i} ^ {j} w _ {i k} ^ {j}}{\\sigma_ {i} ^ {j}} \\tag {11}", + "image_path": "13621e8c7e247977eb0ffc609fe81e83314f655be6ef5bbdf18c1371238a4a9f.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 276, + 296, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 296, + 348 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 276, + 296, + 348 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 276, + 296, + 348 + ], + "type": "text", + "content": " is the layer index, " + }, + { + "bbox": [ + 55, + 276, + 296, + 348 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 276, + 296, + 348 + ], + "type": "text", + "content": " is the channel index of " + }, + { + "bbox": [ + 55, + 276, + 296, + 348 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 276, + 296, + 348 + ], + "type": "text", + "content": "-th layer, and " + }, + { + "bbox": [ + 55, + 276, + 296, + 348 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 276, + 296, + 348 + ], + "type": "text", + "content": " is the channel index in the " + }, + { + "bbox": [ + 55, + 276, + 296, + 348 + ], + "type": "inline_equation", + "content": "(i - 1)" + }, + { + "bbox": [ + 55, + 276, + 296, + 348 + ], + "type": "text", + "content": "-th layer. In this case, " + }, + { + "bbox": [ + 55, + 276, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 276, + 296, + 348 + ], + "type": "text", + "content": " is small enough to be neglected, and " + }, + { + "bbox": [ + 55, + 276, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\gamma_i^j" + }, + { + "bbox": [ + 55, + 276, + 296, + 348 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 276, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\sigma_i^j" + }, + { + "bbox": [ + 55, + 276, + 296, + 348 + ], + "type": "text", + "content": " each play a role in readjusting the weights per channel. Therefore, the computed weights can be used as the importance score." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 358, + 295, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 295, + 384 + ], + "type": "text", + "content": "3.3. A Closer Look at Adversarial Pruning via Importance Score" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 389, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 483 + ], + "type": "text", + "content": "Least Importance Score [50] is a common pruning criterion that assesses each parameter's (weights, filters, channels) importance to overall network performance, pruning the least important first. This score clarifies the impact of adversarial examples on pruning. We compared importance scores between models trained on clean data, adversarial data, and adversarily-trained models using a simple CNN architecture." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 485, + 296, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 296, + 604 + ], + "type": "text", + "content": "As shown in Fig. 2a, importance scores differ significantly between models trained on clean versus adversarial datasets. In the AT model, trained on both, scores do not consistently fall between clean and adversarial values but more closely resemble the clean model. Notably, nodes highly important in both cases (e.g., the second node in the second layer) may appear less important in the AT model. This suggests that training on both datasets may obscure importance, potentially leading to pruning of critical nodes or connections." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "Moreover, we evaluated the same model with different importance scoring methods mentioned earlier. The results in Fig.2b suggest that the order of importance for nodes and connections can vary depending on the scoring method. In other words, each sub-network pruned with a different importance scoring method may contain important information that is not shared by the others. This implies that sub-models sharing the same base model could act as multiple models in ED, while complementing each other." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 228, + 504, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 504, + 241 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 504, + 241 + ], + "type": "text", + "content": "4. Efficient Ensemble Defense (EED)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 250, + 553, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 250, + 553, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 250, + 553, + 344 + ], + "type": "text", + "content": "A common hypothesis in adversarial ensemble defense is that training and combining multiple base models enhances adversarial robustness compared to a single model [7]. However, traditional ED methods often increase model capacity, which is counterproductive for reducing computational and memory costs. But is it truly impossible to achieve an efficient adversarial ensemble defense for a compact model?" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 346, + 554, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 554, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 554, + 538 + ], + "type": "text", + "content": "We propose a novel Efficient Ensemble Defense (EED) method that leverages model compression while maintaining robustness. Sec.3.3 has demonstrated the potential for sub-models, pruned based on different importance scores from a single base model, to act as multiple base models in an ensemble. By ensembling these sub-models, we can reduce model capacity compared to traditional ED. However, it is observed that the smaller the model during ensemble, the more the robustness performance decreases; thus, high diversity is required to mitigate this issue[30]. Therefore, we enhance the ensemble effect and maintain robustness through robust diversity evaluation and misclassification regularization. Furthermore, to improve inference speed—a primary goal of model compression—we dynamically add sub-models during inference and perform ensemble only as needed, thus enhancing inference efficiency." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 539, + 554, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 554, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 554, + 588 + ], + "type": "text", + "content": "EED is designed with three main considerations: 1) the diversity of pruned sub-models, 2) improvement of robustness during the ensemble process, and 3) maintaining compactness throughout the ensemble process." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 599, + 468, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 468, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 468, + 612 + ], + "type": "text", + "content": "4.1. Enhancing Robust Diversity" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "content": "Our strategy begins with generating sub-models from a base model through pruning, where each sub-model undergoes an AP process based on a unique importance score. This AP adheres to the principles described in Sec.3.1, utilizing four importance scoring methods detailed in Sec.3.2. To further enhance diversity, we divide the training dataset for pruning into multiple subsets, with each subset designated to train a specific sub-model. Some subsets are shared across all" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9699" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 66, + 87, + 294, + 437 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 253, + 84 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 253, + 84 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 253, + 84 + ], + "type": "text", + "content": "Algorithm 1 Efficient Ensemble Defense (EED)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "lines": [ + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "text", + "content": "Input: Training data " + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "text", + "content": " , base model " + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "text", + "content": " \nOutput: Ensemble Selected \nfor each subset " + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "inline_equation", + "content": "D_{i}\\subset D" + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "text", + "content": " do " + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "inline_equation", + "content": "F_{i}\\gets" + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "text", + "content": " AdversarialPruning " + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "inline_equation", + "content": "(M,D_{i})" + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "text", + "content": " Set " + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "inline_equation", + "content": "(D)\\leftarrow \\operatorname {Set}(D)\\cup F_i" + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "text", + "content": " \nend for \nEnsSet " + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "text", + "content": " FormEnsembleSet(Set(D)) \nfor each ensemble " + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "inline_equation", + "content": "E\\in \\mathit{ensSet}" + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "text", + "content": " do " + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "inline_equation", + "content": "RD\\gets" + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "text", + "content": " CalculateRobustDiversity(E) \nend for \nSelected " + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "text", + "content": " SelectEnsemble(EnsSet, RD) \nfor each training example " + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "inline_equation", + "content": "(x,y)\\in D" + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "text", + "content": " do " + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{EED}}(x,y)\\gets \\mathcal{L}_E + \\mathcal{L}_R + \\mathcal{L}_C" + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "text", + "content": " Selected " + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "text", + "content": " UpdateModels(Selected, L_EED) \nend for {INFERENCE PHASE} \nfor each test example " + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "inline_equation", + "content": "x\\in T" + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "text", + "content": " do pred " + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "text", + "content": " InitializePredictions(Selected) t " + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "inline_equation", + "content": "\\leftarrow 1" + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "text", + "content": " while not sufficient robustness do pred " + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "text", + "content": " UpdateEnsemble(pred, Selected[t]) t " + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "inline_equation", + "content": "\\leftarrow t + 1" + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "text", + "content": " qt " + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 66, + 87, + 294, + 437 + ], + "type": "text", + "content": " CalculateStoppingCriterion(pred) if robustness criterion met then break end if end while \nend for \nOutput pred" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 460, + 296, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 578 + ], + "type": "text", + "content": "sub-models to retain common core knowledge, while the remaining subsets are distributed across different sub-models to ensure each to learn somewhat different aspects of the data. These subsets are randomly reshuffled whenever the importance scoring method changes. This approach amplifies model diversity, as each sub-model becomes specialized in defending against attacks targeting specific patterns or features in the data. Ultimately, we establish a model pool composed of sub-models with varying importance scoring methods or trained on different datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 579, + 296, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 579, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 579, + 296, + 654 + ], + "type": "text", + "content": "Let this be denoted as pool of " + }, + { + "bbox": [ + 55, + 579, + 296, + 654 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 579, + 296, + 654 + ], + "type": "text", + "content": " models trained on the dataset " + }, + { + "bbox": [ + 55, + 579, + 296, + 654 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 579, + 296, + 654 + ], + "type": "text", + "content": ", represented as " + }, + { + "bbox": [ + 55, + 579, + 296, + 654 + ], + "type": "inline_equation", + "content": "Set(D) = \\{F_0, \\dots, F_{N-1}\\}" + }, + { + "bbox": [ + 55, + 579, + 296, + 654 + ], + "type": "text", + "content": ". The ensemble set " + }, + { + "bbox": [ + 55, + 579, + 296, + 654 + ], + "type": "inline_equation", + "content": "EnsSet" + }, + { + "bbox": [ + 55, + 579, + 296, + 654 + ], + "type": "text", + "content": " represents the collection of all possible ensemble teams with sizes " + }, + { + "bbox": [ + 55, + 579, + 296, + 654 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 579, + 296, + 654 + ], + "type": "text", + "content": " ranging from 2 to " + }, + { + "bbox": [ + 55, + 579, + 296, + 654 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 579, + 296, + 654 + ], + "type": "text", + "content": " formed from " + }, + { + "bbox": [ + 55, + 579, + 296, + 654 + ], + "type": "inline_equation", + "content": "Set(D)" + }, + { + "bbox": [ + 55, + 579, + 296, + 654 + ], + "type": "text", + "content": ". The number of ensemble teams is given by " + }, + { + "bbox": [ + 55, + 579, + 296, + 654 + ], + "type": "inline_equation", + "content": "|EnsSet| = \\sum_{S=2}^{N} \\binom{N}{S} = 2^N - (1 + N)" + }, + { + "bbox": [ + 55, + 579, + 296, + 654 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "To quantify the diversity among sub-models within each ensemble team, we define a Robust Diversity (RD) score based on the probability of defense failure for the model. The diversity score relies on two probabilities: " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "p_{one}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "p_{two}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "p_{one}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " is the probability that a randomly chosen model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 107 + ], + "type": "text", + "content": "fails to defend (viz. the attack succeeds), while " + }, + { + "bbox": [ + 313, + 72, + 555, + 107 + ], + "type": "inline_equation", + "content": "p_{two}" + }, + { + "bbox": [ + 313, + 72, + 555, + 107 + ], + "type": "text", + "content": " is the expected probability that two randomly selected models both fail to defend:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 334, + 114, + 555, + 147 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 114, + 555, + 147 + ], + "spans": [ + { + "bbox": [ + 334, + 114, + 555, + 147 + ], + "type": "interline_equation", + "content": "p _ {o n e} = \\sum_ {i = 1} ^ {S} \\frac {S _ {i}}{S} p _ {i}, \\quad p _ {t w o} = \\sum_ {i = 1} ^ {S} \\frac {S _ {i} (i - 1)}{S (S - 1)} p _ {i} \\tag {12}", + "image_path": "998f8c7149c879ec9ca5177e3aa737d304eff9688437ed6e7e34c8fd77b91ea9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 154, + 555, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 154, + 555, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 555, + 190 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 154, + 555, + 190 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 154, + 555, + 190 + ], + "type": "text", + "content": " represents the defense failure probability of the " + }, + { + "bbox": [ + 313, + 154, + 555, + 190 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 154, + 555, + 190 + ], + "type": "text", + "content": "-th model among the total " + }, + { + "bbox": [ + 313, + 154, + 555, + 190 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 154, + 555, + 190 + ], + "type": "text", + "content": " models. Based on the two probabilities, the RD score is defined by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 399, + 198, + 555, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 198, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 399, + 198, + 555, + 220 + ], + "type": "interline_equation", + "content": "R D = 1 - \\frac {p _ {\\text {t w o}}}{p _ {\\text {o n e}}} \\tag {13}", + "image_path": "2be8690344461967f787ff39772db2aa742e4232b96a3de45abc23b648b59b2d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 227, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 227, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 227, + 555, + 312 + ], + "type": "text", + "content": "This RD metric measures diversity. " + }, + { + "bbox": [ + 313, + 227, + 555, + 312 + ], + "type": "inline_equation", + "content": "RD = 1" + }, + { + "bbox": [ + 313, + 227, + 555, + 312 + ], + "type": "text", + "content": " indicates maximum diversity (i.e., even if one model fails in defense, the likelihood of other models succeeding is high), while " + }, + { + "bbox": [ + 313, + 227, + 555, + 312 + ], + "type": "inline_equation", + "content": "RD = 0" + }, + { + "bbox": [ + 313, + 227, + 555, + 312 + ], + "type": "text", + "content": " indicates no diversity (i.e., both models fail in defense). The EED framework applies RD-based ensemble selection to form the smallest ensemble from the EnsSet of sub-models with the high RD scores." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 318, + 523, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 523, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 523, + 331 + ], + "type": "text", + "content": "4.2. Regularization against Misclassification" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 335, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 555, + 384 + ], + "type": "text", + "content": "As in the existing ED technique [47], the classifier is trained by combining the classification loss, regularization " + }, + { + "bbox": [ + 313, + 335, + 555, + 384 + ], + "type": "inline_equation", + "content": "(Reg)" + }, + { + "bbox": [ + 313, + 335, + 555, + 384 + ], + "type": "text", + "content": " and diversity " + }, + { + "bbox": [ + 313, + 335, + 555, + 384 + ], + "type": "inline_equation", + "content": "(Div)" + }, + { + "bbox": [ + 313, + 335, + 555, + 384 + ], + "type": "text", + "content": " terms. Specifically, for a single training example " + }, + { + "bbox": [ + 313, + 335, + 555, + 384 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 313, + 335, + 555, + 384 + ], + "type": "text", + "content": ", the ensemble loss function is defined by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 337, + 392, + 554, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 392, + 554, + 441 + ], + "spans": [ + { + "bbox": [ + 337, + 392, + 554, + 441 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {E} (x, y) = \\sum_ {i = 1} ^ {N} \\ell \\left(F _ {i} (x), y\\right) + \\alpha R e g (h (x)) \\tag {14} \\\\ + \\beta D i v (h ^ {1} (x), h ^ {2} (x), \\ldots , h ^ {N} (x), y) \\\\ \\end{array}", + "image_path": "95f57e8c0f468ad3141a98bcf1e9f8cc80069cf5ef30292aadc815e0ae6422bf.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 449, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 498 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 449, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 449, + 554, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 449, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 449, + 554, + 498 + ], + "type": "text", + "content": " are hyperparameters, and " + }, + { + "bbox": [ + 313, + 449, + 554, + 498 + ], + "type": "inline_equation", + "content": "h^i (x) = (h_1^i (x),\\ldots ,h_C^i (x)), i\\in \\mathbb{N}" + }, + { + "bbox": [ + 313, + 449, + 554, + 498 + ], + "type": "text", + "content": " represents the output of the ith classifier. The final output is calculated using the average combiner " + }, + { + "bbox": [ + 313, + 449, + 554, + 498 + ], + "type": "inline_equation", + "content": "h(x) = \\frac{1}{N}\\sum_{i = 1}^{N}h^{i}(x)" + }, + { + "bbox": [ + 313, + 449, + 554, + 498 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 498, + 554, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 546 + ], + "type": "text", + "content": "To improve the robustness of EED, we employ an additional regularization term that focuses on the probability scores of the most misclassified class, inspired by iGAT [7]. For a given example " + }, + { + "bbox": [ + 313, + 498, + 554, + 546 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 313, + 498, + 554, + 546 + ], + "type": "text", + "content": ", the proposed term is defined by:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 334, + 553, + 553, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 553, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 334, + 553, + 553, + 597 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {R} (x, y) = - \\psi_ {0 / 1} (c (h ^ {1} (x), \\dots , h ^ {N} (x)), y) \\\\ \\cdot \\log \\left(1 - C \\max _ {i = 1} ^ {N} \\max _ {j = 1} ^ {C} h _ {i} ^ {j} (x)\\right) \\tag {15} \\\\ \\end{array}", + "image_path": "cb33b7ea5ef8dbce713baaad28cd8d60cd93234b7baa4fdfab668444d827c12d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 604, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 555, + 714 + ], + "type": "text", + "content": "The error function " + }, + { + "bbox": [ + 313, + 604, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\psi_{0/1}(f, y) = \\frac{1}{|D|} \\sum_{x \\in D} 1 [f_y(x) < \\max_{c=y} f_c(x)]" + }, + { + "bbox": [ + 313, + 604, + 555, + 714 + ], + "type": "text", + "content": " returns 0 if " + }, + { + "bbox": [ + 313, + 604, + 555, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 604, + 555, + 714 + ], + "type": "text", + "content": " correctly predicts the label " + }, + { + "bbox": [ + 313, + 604, + 555, + 714 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 604, + 555, + 714 + ], + "type": "text", + "content": ", and 1 otherwise. This regularization term plays a crucial role in guiding the model to reduce misclassifications, especially improving its ability to handle ambiguous data points or difficult input examples. By doing so, each base classifier focuses more on the examples it misclassified, enhancing the robustness of the overall ensemble model across diverse input spaces." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9700" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 230, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 230, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 230, + 85 + ], + "type": "text", + "content": "4.3. Regularization for Compactness" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 209 + ], + "type": "text", + "content": "Another key objective of the EED framework is maintaining the compressed models' compactness while maximizing the ensemble merit. To achieve this, we design a compactness regularization term " + }, + { + "bbox": [ + 55, + 89, + 296, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_C" + }, + { + "bbox": [ + 55, + 89, + 296, + 209 + ], + "type": "text", + "content": " to maintain weight and activation sparsity among the sub-models, thereby improving the compactness of the ensemble. The weight regularization term controls the sparsity of the sub-model weights, ensuring that they remain sparse above a certain threshold, thus suppressing model capacity growth. Specifically, the weight sparsity regularization term " + }, + { + "bbox": [ + 55, + 89, + 296, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_W" + }, + { + "bbox": [ + 55, + 89, + 296, + 209 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 124, + 217, + 296, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 217, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 124, + 217, + 296, + 251 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {W} (x, y) = \\sum_ {i = 1} ^ {N} \\| W _ {i} \\| _ {1}, \\tag {16}", + "image_path": "acf25a04003f664b677e7452d5e4dee28931ec0ad78b63648713ba9c2867f0b5.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 258, + 296, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 293 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 258, + 296, + 293 + ], + "type": "inline_equation", + "content": "\\| W_i\\| _1" + }, + { + "bbox": [ + 55, + 258, + 296, + 293 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 55, + 258, + 296, + 293 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 258, + 296, + 293 + ], + "type": "text", + "content": "-norm of the weights " + }, + { + "bbox": [ + 55, + 258, + 296, + 293 + ], + "type": "inline_equation", + "content": "W_{i}" + }, + { + "bbox": [ + 55, + 258, + 296, + 293 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 55, + 258, + 296, + 293 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 258, + 296, + 293 + ], + "type": "text", + "content": "-th sub-model, which encourages sparsity by minimizing the sum of the absolute values of the weights." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 293, + 296, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 296, + 342 + ], + "type": "text", + "content": "To control the activation sparsity of the sub-models, we use an additional sparsity regularization term " + }, + { + "bbox": [ + 55, + 293, + 296, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_A" + }, + { + "bbox": [ + 55, + 293, + 296, + 342 + ], + "type": "text", + "content": ", which encourages each sub-model to reduce unnecessary neuron activations. This term is defined as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 121, + 348, + 295, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 348, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 121, + 348, + 295, + 380 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {A} (x, y) = \\sum_ {i = 1} ^ {N} \\| h ^ {i} (x) \\| _ {0}, \\tag {17}", + "image_path": "6ee14c5bfe3376e7a6b8c11117c5a7368fd23fddf5a3308a394eb4cb2e41bca8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 389, + 295, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 295, + 413 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 389, + 295, + 413 + ], + "type": "inline_equation", + "content": "\\| h^{i}(x)\\|_{0}" + }, + { + "bbox": [ + 55, + 389, + 295, + 413 + ], + "type": "text", + "content": " represents the sum of the number of nonzero activations in the output " + }, + { + "bbox": [ + 55, + 389, + 295, + 413 + ], + "type": "inline_equation", + "content": "h^i (x)" + }, + { + "bbox": [ + 55, + 389, + 295, + 413 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 55, + 389, + 295, + 413 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 389, + 295, + 413 + ], + "type": "text", + "content": "-th sub-model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 413, + 235, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 413, + 235, + 425 + ], + "spans": [ + { + "bbox": [ + 67, + 413, + 235, + 425 + ], + "type": "text", + "content": "Thus, " + }, + { + "bbox": [ + 67, + 413, + 235, + 425 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_C" + }, + { + "bbox": [ + 67, + 413, + 235, + 425 + ], + "type": "text", + "content": " can be summarized as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 118, + 434, + 295, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 434, + 295, + 447 + ], + "spans": [ + { + "bbox": [ + 118, + 434, + 295, + 447 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C} (x, y) = \\lambda_ {1} \\mathcal {L} _ {W} + \\lambda_ {2} \\mathcal {L} _ {A} \\tag {18}", + "image_path": "e0e50a602d5a87d816062e1157975a4f8aac8566fa39d482161e88c5c46c60c5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 455, + 296, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 296, + 525 + ], + "type": "text", + "content": "The hyperparameter " + }, + { + "bbox": [ + 55, + 455, + 296, + 525 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 55, + 455, + 296, + 525 + ], + "type": "text", + "content": " regulates weight sparsity, with higher values leading to more sparse weights and reduced model capacity. Similarly, " + }, + { + "bbox": [ + 55, + 455, + 296, + 525 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 55, + 455, + 296, + 525 + ], + "type": "text", + "content": " controls activation sparsity, making neuron activations sparser, which results in a more compact model structure that saves memory and computational resources." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 527, + 288, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 527, + 288, + 539 + ], + "spans": [ + { + "bbox": [ + 67, + 527, + 288, + 539 + ], + "type": "text", + "content": "Finally, the overall loss function of EED is as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 548, + 295, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 548, + 295, + 561 + ], + "spans": [ + { + "bbox": [ + 61, + 548, + 295, + 561 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {E E D} (x, y) = \\mathcal {L} _ {E} (x, y) + \\omega \\mathcal {L} _ {R} (x, y) + \\gamma \\mathcal {L} _ {C} (x, y) \\tag {19}", + "image_path": "657f27a3c6a8e990732f5170eea5bc5b341d290b7de44d1248c9c17cc7531495.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 569, + 296, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 296, + 617 + ], + "type": "text", + "content": "The EED framework provides high robustness against adversarial attacks while ensuring compactness through these multiple regularization terms, making it suitable to efficient applications like real-time inference." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 624, + 282, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 282, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 282, + 637 + ], + "type": "text", + "content": "4.4. Dynamic Ensemble for Inference Efficiency" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "In the training phase, EED optimizes all sub-models to ensure diversity and robustness. To enhance the compactness and efficiency of EED during the inference phase, we adopt an inference-efficient ensemble method [25]. We apply this to adversarial defense by dynamically constructing an ensemble of varying sizes for each sample, termed Dynamic" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "Inference Ensemble (DIE). Specifically, for a given sample " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": ", EED dynamically determines the optimal point to stop while adding each sub-models progressively. As a result, by monitoring the uncertainty of each sub-model on adversarial examples and only introducing additional models when necessary, we avoid unnecessary computations while maintaining high robustness." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 156, + 554, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 554, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 554, + 252 + ], + "type": "text", + "content": "DIE sequentially adds sub-models and evaluates the robustness of the predictions at each stage. Starting with the first sub-model " + }, + { + "bbox": [ + 313, + 156, + 554, + 252 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 313, + 156, + 554, + 252 + ], + "type": "text", + "content": ", a new sub-model " + }, + { + "bbox": [ + 313, + 156, + 554, + 252 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 313, + 156, + 554, + 252 + ], + "type": "text", + "content": " is added at each stage " + }, + { + "bbox": [ + 313, + 156, + 554, + 252 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 156, + 554, + 252 + ], + "type": "text", + "content": ", and the intermediate ensemble prediction " + }, + { + "bbox": [ + 313, + 156, + 554, + 252 + ], + "type": "inline_equation", + "content": "\\hat{y}_{ens}^{t}" + }, + { + "bbox": [ + 313, + 156, + 554, + 252 + ], + "type": "text", + "content": " is updated. These intermediate predictions serve as a criterion for assessing how robust the added sub-models are. The intermediate ensemble prediction based on the predictions " + }, + { + "bbox": [ + 313, + 156, + 554, + 252 + ], + "type": "inline_equation", + "content": "\\hat{y}_i" + }, + { + "bbox": [ + 313, + 156, + 554, + 252 + ], + "type": "text", + "content": " of each sub-model " + }, + { + "bbox": [ + 313, + 156, + 554, + 252 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 313, + 156, + 554, + 252 + ], + "type": "text", + "content": " is computed by:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 399, + 259, + 553, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 259, + 553, + 291 + ], + "spans": [ + { + "bbox": [ + 399, + 259, + 553, + 291 + ], + "type": "interline_equation", + "content": "\\hat {y} _ {\\text {e n s}} ^ {t} = \\frac {1}{t} \\sum_ {i = 1} ^ {t} \\hat {y} _ {i} \\tag {20}", + "image_path": "bbd64cdf3dc07a90a30b523af2d35503e819b36d97603f32539a2d70ef03310a.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 300, + 553, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 346 + ], + "type": "text", + "content": "Here, each " + }, + { + "bbox": [ + 313, + 300, + 553, + 346 + ], + "type": "inline_equation", + "content": "\\hat{y}_{\\text{ens}}^t" + }, + { + "bbox": [ + 313, + 300, + 553, + 346 + ], + "type": "text", + "content": " represents the combined prediction of the sub-models for the input (both clean and adversarial examples), and it plays a role in strengthening the overall robustness of the ensemble." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 347, + 554, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 554, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 554, + 467 + ], + "type": "text", + "content": "At each stage " + }, + { + "bbox": [ + 313, + 347, + 554, + 467 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 347, + 554, + 467 + ], + "type": "text", + "content": ", the early stopping probability " + }, + { + "bbox": [ + 313, + 347, + 554, + 467 + ], + "type": "inline_equation", + "content": "q_{t}" + }, + { + "bbox": [ + 313, + 347, + 554, + 467 + ], + "type": "text", + "content": " is assessed. If the current ensemble performance is determined to be robust enough, the addition of sub-models is halted. The value of " + }, + { + "bbox": [ + 313, + 347, + 554, + 467 + ], + "type": "inline_equation", + "content": "q_{t}" + }, + { + "bbox": [ + 313, + 347, + 554, + 467 + ], + "type": "text", + "content": " is computed based on the prediction uncertainty and confidence. The prediction uncertainty is computed by KL-divergence of " + }, + { + "bbox": [ + 313, + 347, + 554, + 467 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 347, + 554, + 467 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 347, + 554, + 467 + ], + "type": "inline_equation", + "content": "(t - 1)" + }, + { + "bbox": [ + 313, + 347, + 554, + 467 + ], + "type": "text", + "content": "-th ensemble predictions, and the prediction confidence setting follows WoC[40]. The optimal stopping point " + }, + { + "bbox": [ + 313, + 347, + 554, + 467 + ], + "type": "inline_equation", + "content": "z_{x}" + }, + { + "bbox": [ + 313, + 347, + 554, + 467 + ], + "type": "text", + "content": " is defined as follows, based on a geometric distribution modeling with discrete Bernoulli trials:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 367, + 472, + 553, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 472, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 367, + 472, + 553, + 506 + ], + "type": "interline_equation", + "content": "z _ {x} = \\arg \\max _ {t} \\left(q _ {t} \\prod_ {i = 1} ^ {t - 1} \\left(1 - q _ {i}\\right)\\right) \\tag {21}", + "image_path": "8ac4fcd9883516f0a8e5d71cbc86bc65852c6382c7ba9fea50985f18b88724b2.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 512, + 553, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 512, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 512, + 553, + 548 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 512, + 553, + 548 + ], + "type": "inline_equation", + "content": "q_{t}" + }, + { + "bbox": [ + 313, + 512, + 553, + 548 + ], + "type": "text", + "content": " is dynamically adjusted at each stage based on robustness, and sub-models are added only when necessary, allowing for efficient use of computational resources." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 548, + 553, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 553, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 553, + 583 + ], + "type": "text", + "content": "The final ensemble prediction " + }, + { + "bbox": [ + 313, + 548, + 553, + 583 + ], + "type": "inline_equation", + "content": "\\hat{y}_{ens}" + }, + { + "bbox": [ + 313, + 548, + 553, + 583 + ], + "type": "text", + "content": " is calculated as the average of the predictions of all sub-models up to the stopping point " + }, + { + "bbox": [ + 313, + 548, + 553, + 583 + ], + "type": "inline_equation", + "content": "z_{x}" + }, + { + "bbox": [ + 313, + 548, + 553, + 583 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 397, + 584, + 553, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 584, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 397, + 584, + 553, + 613 + ], + "type": "interline_equation", + "content": "\\hat {y} _ {e n s} = \\frac {1}{z _ {x}} \\sum_ {i = 1} ^ {z _ {x}} \\hat {y} _ {i} \\tag {22}", + "image_path": "22386384a92f5cd5e039665d35173d6ab0933845ee26259a1ded11a6f3d1a53a.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 617, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 677 + ], + "type": "text", + "content": "This approach takes into account the predictions of all submodels while avoiding the inclusion of unnecessary ones, thereby enhancing efficiency. By applying different ensemble sizes for each sample, the overall computational cost is reduced while still being robust against adversarial attacks." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "DIE finds a balance between efficiency and robustness, quickly identifying the optimal combination of sub-models to ensure strong performance even in real-time inference" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "9701" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 91, + 565, + 290 + ], + "blocks": [ + { + "bbox": [ + 91, + 71, + 518, + 82 + ], + "lines": [ + { + "bbox": [ + 91, + 71, + 518, + 82 + ], + "spans": [ + { + "bbox": [ + 91, + 71, + 518, + 82 + ], + "type": "text", + "content": "Table 1. Comparison of various AP methods and EED on CIFAR-10 and SVHN datasets when sparsity rate " + }, + { + "bbox": [ + 91, + 71, + 518, + 82 + ], + "type": "inline_equation", + "content": "{s}_{r} = {80}\\%" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 91, + 565, + 290 + ], + "lines": [ + { + "bbox": [ + 56, + 91, + 565, + 290 + ], + "spans": [ + { + "bbox": [ + 56, + 91, + 565, + 290 + ], + "type": "table", + "html": "
SettingCIFAR-10SVHNPrams
CleanPGDAAC&WDFSpeed upCleanPGDAAC&WDFSpeed up
Resnet 18AT[27]87.0556.1448.0257.6053.101.00x93.3756.2750.1458.8559.801.00x11.2M
R-ADMM[48]81.2548.0043.9249.1739.111.68x74.8149.7337.4052.6243.401.63x
HYDRA[34]77.3652.9243.7449.6445.911.77x91.0652.2247.6255.1346.131.74x
RST[10]61.0241.0118.3851.0226.821.86x82.3946.2936.3552.6539.271.83x
MAD[21]74.1850.3841.2754.1738.441.69x92.8451.6539.8759.8044.721.70x
Flying Bird[2]81.0751.6244.4156.0845.291.79x90.2152.0642.0157.3052.231.76x2.2M
HARP[52]83.8452.5645.3656.5747.041.81x92.6054.1645.8957.2851.241.80x
TwinRep[23]77.2652.0443.5256.6046.181.82x92.9653.8344.7356.2648.611.78x
EED(ours)86.1355.7148.1357.0351.971.86x93.1555.7450.1858.3756.051.85x
VGG 16AT[27]82.7054.4948.5254.9156.911.00x93.0657.6452.2855.3661.491.00x14.7M
R-ADMM[48]76.5147.1940.4144.8048.291.65x65.4048.3944.6448.9249.851.66x
HYDRA[34]78.8347.8544.8545.7349.411.72x90.8253.6646.3052.6353.421.72x
RST[10]77.2939.9030.1532.0242.031.83x82.3941.9445.5243.5838.811.81x
MAD[21]73.6548.8140.8847.1645.501.80x89.9449.4640.3153.0347.371.69x
Flying Bird[2]77.3649.4345.7350.3848.341.68x90.1153.2949.1852.8150.781.73x2.9M
HARP[52]80.9251.7744.0252.7952.171.73x93.2255.5344.7352.9656.161.77x
TwinRep[23]76.0450.1243.2652.4251.591.79x92.6054.6149.0651.2456.581.75x
EED(ours)81.3954.2647.4953.2753.941.79x93.1457.1951.4953.1260.271.77x
", + "image_path": "13dda296c696f0828469014759fa4714b7eaa9d4faea483274f6d3b93cbf0af0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 305, + 295, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 305, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 305, + 295, + 354 + ], + "type": "text", + "content": "settings. As a result, the EED framework can lower the average inference cost in adversarial attack scenarios while maintaining high level of robustness, making it an effective defense, even in resource-constrained environments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 365, + 204, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 204, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 204, + 378 + ], + "type": "text", + "content": "5. Experimental Evaluations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 384, + 167, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 167, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 167, + 397 + ], + "type": "text", + "content": "5.1. Experiment Setups" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 402, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 295, + 533 + ], + "type": "text", + "content": "Classifiers and Datasets We utilize two specific architectures, ResNet18 and VGG16, as classifiers and conduct experiments focusing on CIFAR-10 and SVHN. We divide the dataset into " + }, + { + "bbox": [ + 55, + 402, + 295, + 533 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 402, + 295, + 533 + ], + "type": "text", + "content": " subsets for pruning into various sub-models, setting " + }, + { + "bbox": [ + 55, + 402, + 295, + 533 + ], + "type": "inline_equation", + "content": "d = 4" + }, + { + "bbox": [ + 55, + 402, + 295, + 533 + ], + "type": "text", + "content": " to create three sub-models for each pruning metric. " + }, + { + "bbox": [ + 55, + 402, + 295, + 533 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 55, + 402, + 295, + 533 + ], + "type": "text", + "content": " of the dataset is shared across all sub-models, while the remaining " + }, + { + "bbox": [ + 55, + 402, + 295, + 533 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 55, + 402, + 295, + 533 + ], + "type": "text", + "content": " is allocated separately for each of the three sub-models. With four pruning metrics in play, we end up with a total of " + }, + { + "bbox": [ + 55, + 402, + 295, + 533 + ], + "type": "inline_equation", + "content": "N = 12" + }, + { + "bbox": [ + 55, + 402, + 295, + 533 + ], + "type": "text", + "content": " sub-models, which together form the ensemble set, denoted as EnsSet. For ensemble learning, the entire dataset is used." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 533, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 593 + ], + "type": "text", + "content": "Attacks We use PGD-20 attacker with a step size of 2/255 to generate adversarial examples for training, and Auto Attack(AA)[6], C&W[1] and Deepfool(DF)[28] for evaluation. Except for DF using " + }, + { + "bbox": [ + 55, + 533, + 295, + 593 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 55, + 533, + 295, + 593 + ], + "type": "text", + "content": " norm, the rest use " + }, + { + "bbox": [ + 55, + 533, + 295, + 593 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 55, + 533, + 295, + 593 + ], + "type": "text", + "content": " norm, and all attacks use perturbation strength " + }, + { + "bbox": [ + 55, + 533, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\epsilon = 8 / 255" + }, + { + "bbox": [ + 55, + 533, + 295, + 593 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Training We use the Madry[27] as AT method. The batch size is set to 512, and the multi-step learning rates are \\{0.01, 0.002\\} for CIFAR-10 and \\{0.1, 0.02, 0.004\\} for SVHN. The hyperparameters are set as follows: " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\omega = 10" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = 0.7" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{2} = 0.25" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": ", and the RD threshold is set to 0.7. " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "c_{r}" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": " is set to be more than four times sparser than the desirable " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "s_{r}" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": ", considering the size of " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": " is adjusted to ensure that the " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "s_{r}" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": " is met even after the ensemble. (When " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "s_{r} = 80\\%" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": ", we set " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "c_{r} = 95\\%" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\gamma = 4" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": ".) Each experiment is run on an NVIDIA RTX A6000 GPU with 8 CPU cores." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 304, + 491, + 318 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 491, + 318 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 491, + 318 + ], + "type": "text", + "content": "5.2. Result Comparison and Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 325, + 555, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 555, + 433 + ], + "type": "text", + "content": "This section compares the proposed EED with existing APs for sparsity rate " + }, + { + "bbox": [ + 313, + 325, + 555, + 433 + ], + "type": "inline_equation", + "content": "(s_r)" + }, + { + "bbox": [ + 313, + 325, + 555, + 433 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 325, + 555, + 433 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 313, + 325, + 555, + 433 + ], + "type": "text", + "content": " on the CIFAR-10 and SVHN datasets. Table 1 demonstrates that EED consistently provided superior adversarial robustness, with significantly higher accuracy against attacks such as PGD, AA, C&W, and DF compared to other methods. Additionally, EED maintained high performance on clean data, indicating a proper balance between adversarial robustness and clean accuracy." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 435, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 555, + 579 + ], + "type": "text", + "content": "Note that EED demonstrates superior robustness, particularly against strong attacks such as AA and PGD, outperforming existing defense methods like HYDRA, HARP, and TwinRep. For example, in the CIFAR-10 experiment using the ResNet-18 architecture, EED achieved " + }, + { + "bbox": [ + 313, + 435, + 555, + 579 + ], + "type": "inline_equation", + "content": "55.71\\%" + }, + { + "bbox": [ + 313, + 435, + 555, + 579 + ], + "type": "text", + "content": " accuracy against PGD and " + }, + { + "bbox": [ + 313, + 435, + 555, + 579 + ], + "type": "inline_equation", + "content": "48.13\\%" + }, + { + "bbox": [ + 313, + 435, + 555, + 579 + ], + "type": "text", + "content": " against AA, showing an improvement of over " + }, + { + "bbox": [ + 313, + 435, + 555, + 579 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 313, + 435, + 555, + 579 + ], + "type": "text", + "content": " compared to other methods and achieving the highest robustness. Moreover, despite using only " + }, + { + "bbox": [ + 313, + 435, + 555, + 579 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 435, + 555, + 579 + ], + "type": "text", + "content": " of the model's parameters, the performance drop in AT was minimal, less than " + }, + { + "bbox": [ + 313, + 435, + 555, + 579 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 313, + 435, + 555, + 579 + ], + "type": "text", + "content": ". Also, EED showed stronger defense capabilities than AT against PGD in the ResNet-18-based SVHN experiment." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "As a result, EED accomplished remarkable performance improvements without sacrificing efficiency. Notably, it also records up to 1.86 times faster inference speeds, comparable to or exceeding existing techniques. This is crucial for practical applications where computational resources may be limited, demonstrating that EED is an ideal defense method for resource-constrained environments. Furthermore, EED showed consistent performance across different model architectures such as ResNet-18 and VGG-16, suggesting that EED can be applied to a wide range of network architectures with minimal adjustment." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9702" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 76, + 192, + 205 + ], + "blocks": [ + { + "bbox": [ + 60, + 76, + 192, + 205 + ], + "lines": [ + { + "bbox": [ + 60, + 76, + 192, + 205 + ], + "spans": [ + { + "bbox": [ + 60, + 76, + 192, + 205 + ], + "type": "image", + "image_path": "36788b395246e5fe5c137d24f0967a931957007def8242751b45ec680fea6db3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 217, + 192, + 228 + ], + "lines": [ + { + "bbox": [ + 58, + 217, + 192, + 228 + ], + "spans": [ + { + "bbox": [ + 58, + 217, + 192, + 228 + ], + "type": "text", + "content": "Figure 3. Robust performance by " + }, + { + "bbox": [ + 58, + 217, + 192, + 228 + ], + "type": "inline_equation", + "content": "s_r" + }, + { + "bbox": [ + 58, + 217, + 192, + 228 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 222, + 72, + 378, + 201 + ], + "blocks": [ + { + "bbox": [ + 222, + 72, + 378, + 201 + ], + "lines": [ + { + "bbox": [ + 222, + 72, + 378, + 201 + ], + "spans": [ + { + "bbox": [ + 222, + 72, + 378, + 201 + ], + "type": "image", + "image_path": "253520b5df4c39fe39333c8d1f7a28a2e60ec467abcf044af2f7b96f56edbd1b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 236, + 204, + 364, + 213 + ], + "lines": [ + { + "bbox": [ + 236, + 204, + 364, + 213 + ], + "spans": [ + { + "bbox": [ + 236, + 204, + 364, + 213 + ], + "type": "text", + "content": "(a) RD analysis on ensemble team size." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 389, + 73, + 545, + 201 + ], + "blocks": [ + { + "bbox": [ + 389, + 73, + 545, + 201 + ], + "lines": [ + { + "bbox": [ + 389, + 73, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 389, + 73, + 545, + 201 + ], + "type": "image", + "image_path": "a23a51dd1ba532a789bfdc6c567670c2a02030253dc94f10caee63acb05550ab.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 204, + 530, + 214 + ], + "lines": [ + { + "bbox": [ + 405, + 204, + 530, + 214 + ], + "spans": [ + { + "bbox": [ + 405, + 204, + 530, + 214 + ], + "type": "text", + "content": "(b) RD analysis on sub-model sparsity." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 231, + 220, + 538, + 232 + ], + "lines": [ + { + "bbox": [ + 231, + 220, + 538, + 232 + ], + "spans": [ + { + "bbox": [ + 231, + 220, + 538, + 232 + ], + "type": "text", + "content": "Figure 4. Correlation between RD and robustness in different ensemble components." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 250, + 156, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 250, + 156, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 156, + 262 + ], + "type": "text", + "content": "5.3. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 268, + 295, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 268, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 54, + 268, + 295, + 411 + ], + "type": "text", + "content": "Ablation studies are conducted on Resnet18 and CIFAR-10. Changes along Sparsity Fig.3 shows the change in robustness performance with respect to the sparsity rate. EED demonstrated the highest performance compared to the existing top three methods, except for a very high " + }, + { + "bbox": [ + 54, + 268, + 295, + 411 + ], + "type": "inline_equation", + "content": "s_r" + }, + { + "bbox": [ + 54, + 268, + 295, + 411 + ], + "type": "text", + "content": " (95%) against AA. EED tended to show a decrease in performance as " + }, + { + "bbox": [ + 54, + 268, + 295, + 411 + ], + "type": "inline_equation", + "content": "s_r" + }, + { + "bbox": [ + 54, + 268, + 295, + 411 + ], + "type": "text", + "content": " increases, which is believed to be due to the reduction in ensemble diversity caused by the higher compression rate. On the other hand, at lower " + }, + { + "bbox": [ + 54, + 268, + 295, + 411 + ], + "type": "inline_equation", + "content": "s_r" + }, + { + "bbox": [ + 54, + 268, + 295, + 411 + ], + "type": "text", + "content": ", EED exhibited higher robustness than AT due to the enhanced ensemble diversity. More detailed experimental results for different " + }, + { + "bbox": [ + 54, + 268, + 295, + 411 + ], + "type": "inline_equation", + "content": "s_r" + }, + { + "bbox": [ + 54, + 268, + 295, + 411 + ], + "type": "text", + "content": " can be found in the supplementary materials." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 413, + 295, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 413, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 54, + 413, + 295, + 509 + ], + "type": "text", + "content": "Analyzing Robust Diversity Fig.4 analyzes the role of RD in EED. Fig.3a, which shows robustness according to ensemble team size, demonstrates that a high RD is necessary for small ensemble sizes to achieve high performance. Fig.3b shows robustness from the RD perspective with respect to the compression rate " + }, + { + "bbox": [ + 54, + 413, + 295, + 509 + ], + "type": "inline_equation", + "content": "(c_r)" + }, + { + "bbox": [ + 54, + 413, + 295, + 509 + ], + "type": "text", + "content": " of the sub-models. We can find that to maintain high performance while keeping sub-models small, a high RD is also required." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "text", + "content": "Analyzing Regularization Components EED has two regularization terms, as mentioned in Sec.4.2 and Sec.4.3. Regularization against misclassification " + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "inline_equation", + "content": "(L_{R})" + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "text", + "content": " is controlled by " + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "text", + "content": ", while regularization for compactness " + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "inline_equation", + "content": "(L_{C})" + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "text", + "content": " is controlled by " + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "text", + "content": ". This section analyzes the impact of each term on the performance of EED. Tab.2 shows that " + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "inline_equation", + "content": "L_{R}" + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "text", + "content": " has a greater effect on clean performance, while " + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "inline_equation", + "content": "L_{C}" + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "text", + "content": " has a larger impact on attack performance. In particular, when the compactness term is not used, attack performance significantly deteriorates, which is likely due to the need of reducing the ensemble size further to meet the " + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "inline_equation", + "content": "s_{r}" + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "text", + "content": ", thereby diminishing the ensemble advantages." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 653, + 295, + 713 + ], + "type": "text", + "content": "Effect of Dynamic Inference Ensemble Tab.3 analyzes the effect of DIE for different sparsity rates. At " + }, + { + "bbox": [ + 54, + 653, + 295, + 713 + ], + "type": "inline_equation", + "content": "s_r = 80\\%" + }, + { + "bbox": [ + 54, + 653, + 295, + 713 + ], + "type": "text", + "content": ", DIE had little impact on the main performance and only showed a slight speed improvement. On the other hand, at " + }, + { + "bbox": [ + 54, + 653, + 295, + 713 + ], + "type": "inline_equation", + "content": "s_r = 50\\%" + }, + { + "bbox": [ + 54, + 653, + 295, + 713 + ], + "type": "text", + "content": ", activating DIE resulted in performance im" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 347, + 266, + 523, + 324 + ], + "blocks": [ + { + "bbox": [ + 339, + 249, + 528, + 259 + ], + "lines": [ + { + "bbox": [ + 339, + 249, + 528, + 259 + ], + "spans": [ + { + "bbox": [ + 339, + 249, + 528, + 259 + ], + "type": "text", + "content": "Table 2. Analysis of EED with regularization terms." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 347, + 266, + 523, + 324 + ], + "lines": [ + { + "bbox": [ + 347, + 266, + 523, + 324 + ], + "spans": [ + { + "bbox": [ + 347, + 266, + 523, + 324 + ], + "type": "table", + "html": "
SettingsCleanPGDAA
CIFAR-10EED86.1355.7148.13
ω = 074.1647.4536.37
Resnet 18γ = 076.3339.1031.59
ω,γ = 073.9537.8328.94
", + "image_path": "250d644cfb6ced5c5858cc3843708853db21f9427ffb7a790ca6cd9671f0a044.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 320, + 352, + 549, + 414 + ], + "blocks": [ + { + "bbox": [ + 313, + 334, + 553, + 345 + ], + "lines": [ + { + "bbox": [ + 313, + 334, + 553, + 345 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 553, + 345 + ], + "type": "text", + "content": "Table 3. Correlation between dynamic inference ensemble and " + }, + { + "bbox": [ + 313, + 334, + 553, + 345 + ], + "type": "inline_equation", + "content": "{s}_{r}" + }, + { + "bbox": [ + 313, + 334, + 553, + 345 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 352, + 549, + 414 + ], + "lines": [ + { + "bbox": [ + 320, + 352, + 549, + 414 + ], + "spans": [ + { + "bbox": [ + 320, + 352, + 549, + 414 + ], + "type": "table", + "html": "
SettingsCleanPGDAASpeed up
sr=80%w/ dynamic86.1355.7148.131.86x
w/o dynamic86.2055.6248.111.80x
sr=50%w/ dynamic86.1556.0952.351.64x
w/o dynamic85.9354.2850.761.27x
", + "image_path": "7e2cba1cf25af6abe2277ad5abeddd5d026b3c02e09e4e20979685d183be9759.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 430, + 554, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 554, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 554, + 526 + ], + "type": "text", + "content": "provements of " + }, + { + "bbox": [ + 313, + 430, + 554, + 526 + ], + "type": "inline_equation", + "content": "1.81\\%" + }, + { + "bbox": [ + 313, + 430, + 554, + 526 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 430, + 554, + 526 + ], + "type": "inline_equation", + "content": "1.59\\%" + }, + { + "bbox": [ + 313, + 430, + 554, + 526 + ], + "type": "text", + "content": " for PGD and AA, respectively, while significantly improving inference speed. This suggests that DIE is more effective at lower sparsity. The reason for this performance improvement is that, with lower sparsity, the ensemble can be larger, thus enhancing the optimization effect of the dynamic ensemble. In contrast, without DIE, the likelihood of conflicts between the sub-models may increase as the ensemble grows larger." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 557, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 557, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 557, + 555, + 712 + ], + "type": "text", + "content": "This paper proposes the Efficient Ensemble Defense (EED) technique to maintain both compression and robustness of deep learning models. EED constructs an ensemble of sub-models using different pruning importance scores to enhance adversarial robustness while reducing model size and computational cost. By addressing information loss and limited model diversity caused by pruning with ensemble methods and regularization, EED outperforms existing AP methods in accuracy and robustness. Experimental results show superior resilience of EED against various attacks on the CIFAR-10 and SVHN, maintaining inference efficiency. EED also proves its applicability in diverse neural network architectures and resource-constrained environments." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9703" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 190 + ], + "type": "text", + "content": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.RS-2022-00155915 (Artificial Intelligence Convergence Innovation Human Resources Development (Inha University)) and No. 2021-0-02068 (AI Innovation Hub)), and this work was partly supported by the National Research Foundation of Korea (NRF) grant funded by the MSIT (No. 2022R1A2C2010095)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 212, + 114, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 212, + 114, + 224 + ], + "spans": [ + { + "bbox": [ + 56, + 212, + 114, + 224 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 232, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 232, + 296, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 232, + 296, + 265 + ], + "spans": [ + { + "bbox": [ + 61, + 232, + 296, + 265 + ], + "type": "text", + "content": "[1] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In 2017 IEEE symposium on security and privacy (sp), pages 39-57. IEEE, 2017. 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 266, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 266, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 62, + 266, + 296, + 319 + ], + "type": "text", + "content": "[2] Tianlong Chen, Zhenyu Zhang, Pengjun Wang, Santosh Balachandra, Haoyu Ma, Zehao Wang, and Zhangyang Wang. Sparsity winning twice: Better robust generalization from more efficient training. arXiv preprint arXiv:2202.09844, 2022. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 322, + 296, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 322, + 296, + 366 + ], + "spans": [ + { + "bbox": [ + 62, + 322, + 296, + 366 + ], + "type": "text", + "content": "[3] Wenlin Chen, James Wilson, Stephen Tyree, Kilian Weinberger, and Yixin Chen. Compressing neural networks with the hashing trick. In International conference on machine learning, pages 2285-2294. PMLR, 2015. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 367, + 296, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 367, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 62, + 367, + 296, + 432 + ], + "type": "text", + "content": "[4] Wuyang Chen, Ziyu Jiang, Zhangyang Wang, Kexin Cui, and Xiaoning Qian. Collaborative global-local networks for memory-efficient segmentation of ultra-high resolution images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8924-8933, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 434, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 296, + 510 + ], + "type": "text", + "content": "[5] Hao Cheng, Kaidi Xu, Zhengang Li, Pu Zhao, Chenan Wang, Xue Lin, Bhavya Kailkhura, and Ryan Goldhahn. More or less (mol): Defending against multiple perturbation attacks on deep neural networks through model ensemble and compression. In 2022 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW), pages 645-655. IEEE, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 512, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 296, + 555 + ], + "type": "text", + "content": "[6] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 557, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 296, + 578 + ], + "type": "text", + "content": "[7] Yian Deng and Tingting Mu. Understanding and improving ensemble adversarial defense. NeurIPS, 36, 2024. 2, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 580, + 296, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 580, + 296, + 623 + ], + "spans": [ + { + "bbox": [ + 62, + 580, + 296, + 623 + ], + "type": "text", + "content": "[8] Misha Denil, Babak Shakibi, Laurent Dinh, Marc'Aurelio Ranzato, and Nando De Freitas. Predicting parameters in deep learning. Advances in neural information processing systems, 26, 2013. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 624, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 296, + 678 + ], + "type": "text", + "content": "[9] Guneet S Dhillon, Kamyar Azizzadenesheli, Zachary C Lipton, Jeremy Bernstein, Jean Kossaifi, Aran Khanna, and Anima Anandkumar. Stochastic activation pruning for robust adversarial defense. arXiv preprint arXiv:1803.01442, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "text", + "content": "[10] Yonggan Fu, Qixuan Yu, Yang Zhang, Shang Wu, Xu Ouyang, David Cox, and Yingyan Lin. Drawing robust scratch tickets: Subnetworks with inborn robustness are" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 105 + ], + "type": "text", + "content": "found within randomly initialized networks. Advances in Neural Information Processing Systems, 34:13059-13072, 2021. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 106, + 554, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 554, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 554, + 148 + ], + "type": "text", + "content": "[11] Ian Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In International Conference on Learning Representations, 2015. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 150, + 554, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 554, + 194 + ], + "type": "text", + "content": "[12] Yiwen Guo, Chao Zhang, Changshui Zhang, and Yurong Chen. Sparse dnns with improved adversarial robustness. Advances in neural information processing systems, 31, 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 195, + 554, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 554, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 554, + 237 + ], + "type": "text", + "content": "[13] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 239, + 554, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 554, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 554, + 281 + ], + "type": "text", + "content": "[14] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. Advances in neural information processing systems, 28, 2015. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 283, + 554, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 554, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 554, + 327 + ], + "type": "text", + "content": "[15] Warren He, James Wei, Xinyun Chen, Nicholas Carlini, and Dawn Song. Adversarial example defense: Ensembles of weak defenses are not strong. In 11th USENIX workshop on offensive technologies (WOOT 17), 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 327, + 554, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 554, + 381 + ], + "type": "text", + "content": "[16] Tong Jian, Zifeng Wang, Yanzhi Wang, Jennifer Dy, and Stratis Ioannidis. Pruning adversarially robust neural networks without adversarial examples. In 2022 IEEE International Conference on Data Mining (ICDM), pages 993-998. IEEE, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 383, + 554, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 554, + 415 + ], + "type": "text", + "content": "[17] Sanjay Kariyappa and Moinuddin K Qureshi. Improving adversarial robustness of ensembles with diversity training. arXiv preprint arXiv:1901.09981, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 416, + 554, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 554, + 437 + ], + "type": "text", + "content": "[18] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 438, + 554, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 554, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 554, + 492 + ], + "type": "text", + "content": "[19] Souvik Kundu, Mahdi Nazemi, Peter A Beerel, and Massoud Pedram. Dnr: A tunable robust pruning framework through dynamic network rewiring of dnns. In Proceedings of the 26th Asia and South Pacific Design Automation Conference, pages 344–350, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 493, + 554, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 554, + 525 + ], + "type": "text", + "content": "[20] Alexey Kurakin, Ian Goodfellow, and Samy Bengio. Adversarial machine learning at scale. arXiv preprint arXiv:1611.01236, 2016. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 526, + 554, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 554, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 554, + 580 + ], + "type": "text", + "content": "[21] Byung-Kwan Lee, Junho Kim, and Yong Man Ro. Masking adversarial damage: Finding adversarial saliency for robust and sparse network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15126-15136, 2022. 3, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 582, + 554, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 554, + 624 + ], + "type": "text", + "content": "[22] Bai Li, Shiqi Wang, Yunhan Jia, Yantao Lu, Zhenyu Zhong, Lawrence Carin, and Suman Jana. Towards practical lottery ticket hypothesis for adversarial training. arXiv preprint arXiv:2003.05733, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 625, + 555, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 555, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 555, + 679 + ], + "type": "text", + "content": "[23] Chenhao Li, Qiang Qiu, Zhibin Zhang, Jiafeng Guo, and Xueqi Cheng. Learning adversarially robust sparse networks via weight reparameterization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8527-8535, 2023. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[24] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. arXiv preprint arXiv:1608.08710, 2016. 1" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9704" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "text", + "content": "[25] Ziyue Li, Kan Ren, Yifan Yang, Xinyang Jiang, Yuqing Yang, and Dongsheng Li. Towards inference efficient deep ensemble learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8711-8719, 2023. 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 152 + ], + "type": "text", + "content": "[26] Ningyi Liao, Shufan Wang, Liyao Xiang, Nanyang Ye, Shuo Shao, and Pengzhi Chu. Achieving adversarial robustness via sparsity. Machine Learning, pages 1-27, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "text", + "content": "[27] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 1, 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "text", + "content": "[28] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2574–2582, 2016. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "type": "text", + "content": "[29] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Anima Anandkumar. Diffusion models for adversarial purification. arXiv preprint arXiv:2205.07460, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "type": "text", + "content": "[30] Tianyu Pang, Kun Xu, Chao Du, Ning Chen, and Jun Zhu. Improving adversarial robustness via promoting ensemble diversity. In International Conference on Machine Learning, pages 4970-4979. PMLR, 2019. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 397 + ], + "type": "text", + "content": "[31] Nicolas Papernot, Patrick McDaniel, Xi Wu, Somesh Jha, and Ananthram Swami. Distillation as a defense to adversarial perturbations against deep neural networks. In 2016 IEEE symposium on security and privacy (SP), pages 582-597. IEEE, 2016. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 443 + ], + "type": "text", + "content": "[32] Giorgio Piras, Maura Pintor, Ambra Demontis, Battista Biggio, Giorgio Giacinto, and Fabio Roli. Adversarial pruning: A survey and benchmark of pruning methods for adversarial robustness. arXiv preprint arXiv:2409.01249, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 444, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 294, + 477 + ], + "type": "text", + "content": "[33] J Redmon. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2016. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "type": "text", + "content": "[34] Vikash Sehwag, Shiqi Wang, Prateek Mittal, and Suman Jana. Hydra: Pruning adversarially robust neural networks. Advances in Neural Information Processing Systems, 33: 19655-19666, 2020. 1, 3, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "text", + "content": "[35] Ilia Shumailov, Yiren Zhao, Robert Mullins, and Ross Anderson. To compress or not to compress: Understanding the interactions between adversarial attacks and neural network compression. Proceedings of Machine Learning and Systems, 1:230-240, 2019. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "type": "text", + "content": "[36] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. Advances in neural information processing systems, 27, 2014. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 668 + ], + "type": "text", + "content": "[37] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In International Conference on Learning Representations, 2014. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "text", + "content": "[38] Lukas Timpl, Rahim Entezari, Hanie Sedghi, Behnam Neyshabur, and Olga Saukh. Understanding the effect of sparsity on neural networks robustness. arXiv preprint arXiv:2206.10915, 2022. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 117 + ], + "type": "text", + "content": "[39] Florian Tramér, Alexey Kurakin, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. Ensemble adversarial training: Attacks and defenses. In International Conference on Learning Representations, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 173 + ], + "type": "text", + "content": "[40] Xiaofang Wang, Dan Kondratyuk, Eric Christiansen, Kris M Kitani, Yair Movshovitz-Attias, and Elad Eban. Wisdom of committees: An overlooked approach to faster and more accurate models. In International Conference on Learning Representations, 2021. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 554, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 554, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 554, + 219 + ], + "type": "text", + "content": "[41] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In International conference on learning representations, 2019. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 554, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 554, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 554, + 275 + ], + "type": "text", + "content": "[42] Yue Wang, Jianghao Shen, Ting-Kuei Hu, Pengfei Xu, Tan Nguyen, Richard Baraniuk, Zhangyang Wang, and Yingyan Lin. Dual dynamic inference: Enabling more efficient, adaptive, and controllable deep inference. IEEE Journal of Selected Topics in Signal Processing, 14(4):623-633, 2020. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 276, + 554, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 554, + 320 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 554, + 320 + ], + "type": "text", + "content": "[43] Xiao Wei, Yao Zhu, and Shu-Tao Xia. Batch normalization assisted adversarial pruning: Towards lightweight, sparse and robust models. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 321, + 554, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 554, + 365 + ], + "type": "text", + "content": "[44] Boxi Wu, Jinghui Chen, Deng Cai, Xiaofei He, and Quanquan Gu. Do wider neural networks really help adversarial robustness? Advances in Neural Information Processing Systems, 34:7054-7067, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 366, + 554, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 554, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 554, + 410 + ], + "type": "text", + "content": "[45] Dongxian Wu, Shu-Tao Xia, and Yisen Wang. Adversarial weight perturbation helps robust generalization. Advances in neural information processing systems, 33:2958-2969, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 411, + 554, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 554, + 467 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 554, + 467 + ], + "type": "text", + "content": "[46] Yanzhao Wu, Ling Liu, Zhongwei Xie, Ka-Ho Chow, and Wenqi Wei. Boosting ensemble accuracy by revisiting ensemble diversity metrics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16469-16477, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 468, + 554, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 554, + 533 + ], + "type": "text", + "content": "[47] Huanrui Yang, Jingyang Zhang, Hongliang Dong, Nathan Inkawhich, Andrew Gardner, Andrew Touchet, Wesley Wilkes, Heath Berry, and Hai Li. Diverse: diversifying vulnerabilities for enhanced robust generation of ensembles. Advances in Neural Information Processing Systems, 33: 5505-5515, 2020. 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 535, + 554, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 554, + 601 + ], + "type": "text", + "content": "[48] Shaokai Ye, Kaidi Xu, Sijia Liu, Hao Cheng, Jan-Henrik Lambrechts, Huan Zhang, Aojun Zhou, Kaisheng Ma, Yanzhi Wang, and Xue Lin. Adversarial robustness vs. model compression, or both? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 111-120, 2019. 1, 2, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 602, + 554, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 554, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 554, + 645 + ], + "type": "text", + "content": "[49] Jongmin Yoon, Sung Ju Hwang, and Juho Lee. Adversarial purification with score-based generative models. In International Conference on Machine Learning, pages 12062-12072. PMLR, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 712 + ], + "type": "text", + "content": "[50] Ruichi Yu, Ang Li, Chun-Fu Chen, Jui-Hsin Lai, Vlad I Morariu, Xintong Han, Mingfei Gao, Ching-Yung Lin, and Larry S Davis. Nisp: Pruning networks using neuron importance score propagation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9194-9203, 2018. 2, 3, 4" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9705" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 163 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "text", + "content": "[51] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In International conference on machine learning, pages 7472-7482. PMLR, 2019. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 163 + ], + "type": "text", + "content": "[52] Qi Zhao and Christian Wressnegger. Holistic adversarially robust pruning. In The Eleventh International Conference on Learning Representations, 2023. 1, 2, 3, 7" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "9706" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Type-R_ Automatically Retouching Typos for Text-to-Image Generation/6f176b19-a787-4b85-9ebe-081a3f4759be_content_list.json b/2025/Type-R_ Automatically Retouching Typos for Text-to-Image Generation/6f176b19-a787-4b85-9ebe-081a3f4759be_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b54ffafa625f08f8d5dcd723f64852787dea9210 --- /dev/null +++ b/2025/Type-R_ Automatically Retouching Typos for Text-to-Image Generation/6f176b19-a787-4b85-9ebe-081a3f4759be_content_list.json @@ -0,0 +1,1889 @@ +[ + { + "type": "text", + "text": "Type-R: Automatically Retouching Typos for Text-to-Image Generation", + "text_level": 1, + "bbox": [ + 138, + 130, + 859, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wataru Shimoda1 Naoto Inoue1 Daichi Haraguchi1 Hayato Mitani2 Seiichi Uchida2 Kota Yamaguchi1 CyberAgent Kyushu University", + "bbox": [ + 277, + 179, + 718, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Create a poster. Display the text 'CVPR' at the top of the image. NASHVILLE TENNESSEE is below the text. The date information 2025 JUNE is required. This image contains the picture of Music City Center at the lower right and cityscapes at the left.", + "bbox": [ + 96, + 309, + 285, + 410 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prompt", + "bbox": [ + 165, + 436, + 218, + 450 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b80130e76e0085375c97dfe722e0e026a5e06be172cea49f3f6bb098bd71b2e1.jpg", + "image_caption": [ + "Generated image" + ], + "image_footnote": [], + "bbox": [ + 290, + 276, + 491, + 424 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b697a7f71940d60fef9741ae7df78653404ac874a2839a94e77b2c18b2478669.jpg", + "image_caption": [ + "Type-R" + ], + "image_footnote": [], + "bbox": [ + 493, + 268, + 694, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cb2b0080be1dcbfc7beec7ced6256970329e2146cf16f5cf5bd458b2d0f5f293.jpg", + "image_caption": [ + "Figure 1. Type-R automatically corrects typographic errors in an image generated from any text-to-image model without sacrificing their visual quality. We highlight prompt words corresponding to the words in the image with the same color." + ], + "image_footnote": [], + "bbox": [ + 581, + 426, + 651, + 463 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e28eda8aaedb55f708f4fd1f1fbb7ac4a17febe48bb40cb3ab1a00ad382dbd98.jpg", + "image_caption": [ + "Typo-corrected image" + ], + "image_footnote": [], + "bbox": [ + 696, + 276, + 897, + 424 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 521, + 326, + 537 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While recent text-to-image models can generate photorealistic images from text prompts that reflect detailed instructions, they still face significant challenges in accurately rendering words in the image. In this paper, we propose to retouch erroneous text renderings in the post-processing pipeline. Our approach, called Type-R, identifies typographical errors in the generated image, erases the erroneous text, regenerates text boxes for missing words, and finally corrects typos in the rendered words. Through extensive experiments, we show that Type-R, in combination with the latest text-to-image models such as Stable Diffusion or Flux, achieves the highest text rendering accuracy while maintaining image quality and also outperforms text-focused generation baselines in terms of balancing text accuracy and image quality.", + "bbox": [ + 86, + 551, + 485, + 781 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 806, + 220, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The rapid advancement of text-to-image generation models triggers a surge of application interest in various creative domains. Graphic design is one of the domains from which", + "bbox": [ + 89, + 832, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "text-to-image models potentially benefit, with the goal of helping generate appealing posters, advertisements, book covers, webpages, or signs, such that these display media convey the intended message to the viewers. However, one of the major obstacles is text rendering on the image [25]. Popular state-of-the-art text-to-image models such as Stable Diffusion [6, 30] or Flux [3] still face notable challenges in accurately rendering words or characters (Fig. 1 second column). Common issues include missing words, unintended characters, or word-level errors that insert another similar-looking word. While it might be possible to tune the text-to-image models with increasing amounts of training data, the resulting image is not guaranteed to be typo-free.", + "bbox": [ + 511, + 523, + 906, + 719 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, instead of modifying the image generator, we propose automatically retouching the erroneous text renderings of the already generated image. Our approach, named Type-R, detects typographic errors in a generated image and automatically fixes erroneous content. Figure 1 shows an example of the input prompt, the generated image by Flux.1-dev [3], detected errors by Type-R, and the final retouched image. Type-R successfully corrects the typographic errors without artifacts compared to the initial generation. Our approach is fully automatic, independent of the base text-to-image generator, and easily pluggable to any existing text-to-image model.", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Code and project page are available.", + "bbox": [ + 107, + 886, + 308, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2745", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Through extensive experiments, we study the limitations in the current text-to-image models, including both the popular text-to-image generators [3, 30] and typophy-focussed models [9, 10], and show that Type-R significantly improves the text rendering accuracy of text-to-image models while maintaining graphic design quality. We also study several design choices, e.g., how changing the base text-to-image generator or the OCR model affects the resulting visual quality and text accuracy, and show that Type-R in combination with Flux [3] achieves the best trade-off between text rendering accuracy and quality scores by GPT evaluation on Mario-Eval benchmark [9].", + "bbox": [ + 89, + 90, + 480, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our contributions in the following", + "bbox": [ + 109, + 273, + 436, + 287 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose Type-R, which automatically retouches typographic errors in generated images by text-to-image models while maintaining the design quality.", + "2. We empirically show that the combination of the state-of-the-art text-to-image generator and Type-R achieves the best quality-accuracy trade-off and outperforms the approach specifically designed for typography." + ], + "bbox": [ + 89, + 287, + 482, + 395 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 409, + 232, + 424 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Text-to-Image Generation", + "text_level": 1, + "bbox": [ + 89, + 434, + 326, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-image generation is the task of creating images from text prompts. With the development of large image-capture datasets, there has been significant progress in generating photorealistic images from diffusion-based models [1, 2, 30, 34, 36]. However, rendering accurate text is still one of the remaining problems. Liu et al. [25] point out the challenge in accurately rendering characters in text-to-image generation models, where it is common to observe missing glyphs, redundant glyphs, or wrong characters. The text accuracy problem persists in the recent text-to-image models such as Stable Diffusion 3 [11] or Flux [3].", + "bbox": [ + 89, + 458, + 482, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, instead of modifying or tuning the text-to-image models, we utilize feedback from external models to improve text rendering. Several works report the successful use of external models in improving image generation. RPG [45] introduces a technique for region-specific prompt guidance for noise control. Wu et al. [44] apply object detection to generated images and refine detection results using Large Language Models. In our task setup, we take advantage of pre-trained Vision Language Models (VLMs), OCR models, and text-inpainting models to retouch typographic errors.", + "bbox": [ + 89, + 625, + 482, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Visual Text Generation", + "text_level": 1, + "bbox": [ + 89, + 801, + 303, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several text-to-image models [2, 26, 27, 36, 52] have reported that better text encoder significantly improves text rendering capabilities, where the encoder can learn glyph-level information from the training corpus. Adding OCR losses also improve performances [20, 39, 53]. While im", + "bbox": [ + 89, + 825, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "proving the capability of the base text-to-image generator might eventually reduce typographic errors, we consider a post-processing approach independent of the base model.", + "bbox": [ + 511, + 90, + 903, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another line of work to improve text rendering is via conditional inputs to the text-to-image models. Some literature attempts at conditioning generation on text positions and glyph masks to improve the text rendering accuracies [9, 10, 28, 29, 39, 47, 51, 54]. Particularly, TextDiffusers [9, 10] generate images with two steps: layout planning, which generates text position information from prompts, followed by image generation conditioned on the first layout information. TextDiffuser [9] and TextDiffuser 2 [9] differ in how they generate layout information, where the former generates character-by-character masks and the latter generates layout in language. One major limitation of conditional generation is the training requirements; i.e., building a sufficient training corpus for the conditioning module is often difficult. Our experiments suggest that TextDiffusers suffer from inadequate image quality compared to the state-of-the-art image generators, likely due to the bias in its training data.", + "bbox": [ + 511, + 140, + 906, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We take the approach of retouching afterward using external pre-trained models to take full advantage of the best text-to-image models without fine-tuning. A few works incorporate external models for text rendering. Very recently, Lakhanpal et al. reports Refining T2I [16] that applies OCR to the generated images by TextDiffuser and optimizes the layouts to improve the legibility of texts, while the same limitation to TextDiffusers still applies. Li et al. [21] propose to generate background images by removing generated scene text. Unlike prior works, we propose a method for generating images that adheres as much as possible to the initial generation results, making full use of OCR models, text removal models, and VLMs to improve text rendering.", + "bbox": [ + 511, + 417, + 908, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Text Editing and Synthesis", + "text_level": 1, + "bbox": [ + 511, + 635, + 754, + 651 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another relevant task is raster text editing, which directly manipulates texts in the image. While GAN-based methods [31, 35, 43, 46] have been popular, recent visual text generation methods [14, 18, 39-41, 50, 55] are capable of editing texts in the image directly. While text editing and our task setup seem similar, text editing is only a part of the typography retouching task, which involves further tasks such as automatic error detection or missing word insertion.", + "bbox": [ + 511, + 662, + 906, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 511, + 810, + 620, + 828 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our Type-R pipeline consists of four stages, which are all automatic: 1) error detection, 2) text erasing, 3) layout regeneration, and 4) typo correction. We show the pipeline overview in Fig. 2.", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2746", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/da7b0999f7424eef3557508a0b85fa2761b2bec67f0e774281a6743e1c8e0c9d.jpg", + "image_caption": [ + "Generated image" + ], + "image_footnote": [ + "Create an advertisement for a hamburger with a pop background. Display the texts '10% OFF and HAMBURGER' with a tasty burger, fried potatoes, and a Coke. The text 'Combo Meals' is required. Prompt" + ], + "bbox": [ + 96, + 87, + 248, + 205 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/df2d3b673ee05efb2be9e89e91b57815dd84fbce33aafb825389787ecc58d188.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 87, + 408, + 205 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/db645485cc70c71938434ce1f57df5d4c07561865d3b5a7078aa1920d22c3355.jpg", + "image_caption": [ + "Generation", + "Error detection", + "Figure 2. Illustration of the Type-R pipeline. Type-R automatically detects errors, erases unintended texts, inserts missing words, and corrects spelling errors in the image." + ], + "image_footnote": [], + "bbox": [ + 256, + 210, + 408, + 284 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7b27e99f440aaa70e4c90a60f597cdcb1561a5c8849cdc197f3042a276bd3bad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 87, + 566, + 205 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2b630661ec278ff3b568e862f49b517f601db572ee655793d1a6fc5740dbaaae.jpg", + "image_caption": [ + "Text erasing" + ], + "image_footnote": [], + "bbox": [ + 413, + 210, + 565, + 284 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b5d5466011a573b8bf20e6f4edd341467c9aba83d770805ad11488fedbee07c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 570, + 87, + 723, + 205 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8d12660020e55b3e3718902cc82dc65c113d80da4c0beb9ff56b71ad008b262b.jpg", + "image_caption": [ + "Layout regeneration" + ], + "image_footnote": [], + "bbox": [ + 571, + 210, + 723, + 284 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/77e45fe2dce800364594956fd18bfe0fe22385c32e059f43e80919285ac57ad6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 87, + 883, + 205 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/99438419cdb13ddb5a10afa1aa71fdb5fe492803bdd645c9b680d00fc3cec239.jpg", + "image_caption": [ + "3 Typo correction" + ], + "image_footnote": [], + "bbox": [ + 730, + 210, + 880, + 284 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Error Detection", + "text_level": 1, + "bbox": [ + 89, + 367, + 250, + 382 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We denote an image $I$ generated by an off-the-shelf text-to-image model $G$ with a prompt, which requests to show $N$ words $W = \\{w_{i}\\}_{i = 1}^{N}$ in $I$ . To ensure how much this expectation is satisfied, we first detect words in the image $I$ . Specifically, a scene text detection model predicts polygonal word regions $\\hat{B}$ . We remove polygons with a height less than a threshold $\\theta$ . Then, a scene text recognition model parses the words $\\hat{W} = \\{\\hat{w}_i\\}_{i = 1}^{\\hat{N}}$ in the regions $\\hat{B}$ . As noted in Section 1, the model $G$ is often not perfect, and $W \\neq \\hat{W}$ ; therefore, we need to identify errors in $\\hat{W}$ by considering $W$ is a ground-truth.", + "bbox": [ + 89, + 390, + 483, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Roughly speaking, this identification task is a matching problem between $W$ and $\\hat{W}$ . More specifically, we consider the matching between the word set $\\mathcal{W}$ and $\\hat{\\mathcal{W}}$ of equal size $\\mathcal{N}$ by optimal transport:", + "bbox": [ + 89, + 560, + 483, + 619 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {O T} (\\mathcal {W}, \\hat {\\mathcal {W}}) \\equiv \\max _ {\\pi} \\sum_ {i = 1} ^ {\\mathcal {N}} \\operatorname {L e v} \\left(w _ {i}, \\hat {w} _ {\\pi (i)}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 631, + 480, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $w_{i}\\in \\mathcal{W},\\hat{w}_{i}\\in \\hat{\\mathcal{W}},\\pi$ is a one-by-one matching, and $\\mathrm{Levenshtein}$ distance to measure the text similarity. This matching works only when the two sets are of equal size, but typographic errors result in different sizes because of missing or unintended words. Here, we equalize two word sets $W$ and $\\hat{W}$ by padding tokens. Let us define $\\mathcal{N} = \\max (N,\\hat{N})$ . We express the padded word sets by:", + "bbox": [ + 89, + 686, + 483, + 792 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {W} = W \\cup \\left\\{p _ {i} \\right\\} _ {i = N + 1} ^ {\\mathcal {N}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 818, + 480, + 837 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {W}} = \\hat {W} \\cup \\left\\{\\hat {p} _ {i} \\right\\} _ {i = \\hat {N} + 1} ^ {\\mathcal {N}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 839, + 480, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $p_i$ and $\\hat{p}_i$ are the padding tokens. Here, a word matching a padding token indicates a typographic error; i.e.,", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a paired word with $p$ is an unintended word, and a paired word with $\\hat{p}$ is a missing word. We assign a constant distance to any word matching to the padding tokens. Any pairs having non-zero distance indicate spelling errors.", + "bbox": [ + 511, + 368, + 903, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The following sections describe our approach to minimizing the above optimal transportation cost. We align the number of words by erasing redundant texts when $N < \\hat{N}$ (Sec. 3.2) and insert text boxes by layout regeneration when $N > \\hat{N}$ (Sec. 3.3). Finally, we fix spelling errors in typo correction (Sec. 3.4).", + "bbox": [ + 511, + 429, + 905, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Text Erasing", + "text_level": 1, + "bbox": [ + 513, + 527, + 648, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Type-R erases unintended texts using an off-the-shelf inpainting method [37] if the number of detected words exceeds the number of the words specified by a prompt, i.e., $N < \\hat{N}$ , or the generated image includes filtered texts. This step reduces the number of words in a generated image to equal or less than the number of specified words by the prompt. The inpainting model removes specified paint regions. In our setup, we specify regions that we want to erase in $\\hat{B}$ and composite the inpainted results to the original image. Since text detection may not always fully cover text areas, we slightly enlarge the text region masks to ensure complete removal.", + "bbox": [ + 511, + 551, + 906, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Layout Regeneration", + "text_level": 1, + "bbox": [ + 511, + 741, + 712, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Layout regeneration fills in missing text regions so that the resulting $\\hat{N}$ equals the number of words specified by a prompt $N$ when $\\hat{N} < N$ . This is equivalent to obtaining position information from the image for missing words of the prompt, and we address this task using a VLM, specifically GPT-4o [4]. Inspired by works on layout generation by zero/few-shot learning of LLMs [12, 23], we ask GPT to plan layouts for each missing word from the image and layout information of valid text boxes in JSON format.", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2747", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For simplicity, we represent the layout information as a bounding box with four coordinates. When text detection outputs are in the polygonal form, we convert them to boxes by finding the smallest four-sided bounding box that encloses each polygon.", + "bbox": [ + 89, + 90, + 482, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Tipo Correction", + "text_level": 1, + "bbox": [ + 89, + 179, + 253, + 195 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Typo correction tries to resolve all word pairs with non-zero edit distance. To address word-level spelling errors, we employ a pre-trained text editing model [39] to automatically retouch the word region. Using an inpainting approach, the text editing model generates raster images containing specified text in the selected areas. Similar to text-to-image models, text editing models may not always render the specified text accurately but with a much lower computation cost than large text-to-image models. We repeatedly apply the text editing model until we fix the spelling errors or we reach the predefined number of attempts. Note that we only composite the areas of corrected words in the edited image during the iteration. We show the pseudo-codes in Algorithm 1.", + "bbox": [ + 89, + 203, + 483, + 400 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 A pseudo-code of typo correction." + ], + "code_body": "1: function TyPO CORRECTION(Image I, Words W, Detected regions $\\hat{B}$ , Maximum iterations $t_\\mathrm{max})$ \n2: for $t = 1$ to $t_\\mathrm{max}$ do \n3: $I_{\\mathrm{edit}}\\gets$ Text-Editor(I,W, $\\hat{B}$ ) \n4: $\\hat{W}\\gets$ Text-Recognition $(I_{\\mathrm{edit}},\\hat{B})$ \n5: $\\mathcal{O}\\gets \\mathrm{Diff}(W,\\hat{W})\\qquad \\triangleright$ Find spelling errors \n6: $I\\gets$ Composite $(I,I_{\\mathrm{edit}},\\hat{B},\\mathcal{O})$ \n7: if $\\mathcal{O}\\neq \\varnothing$ then \n8: $W,\\hat{B}\\gets$ Update(W, $\\hat{B},\\mathcal{O})$ \n9: else \n10: break \n11: end if \n12: end for \n13: return I \n14: end function", + "bbox": [ + 93, + 435, + 483, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 699, + 215, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 89, + 724, + 186, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We evaluate the performance of visual text generation on MARIO-Eval benchmark [9]. This benchmark contains 5,414 prompts, of which 5,000 have image data. We use all 5,414 prompts in the main comparison, i.e., in Tab. 1. We randomly pick 500 of the 5,414 prompts as the subset for analysis in Sec. 4.6 because our experiments involve paid APIs. For the selection of hyperparameters in filtering, we use 500 prompts randomly selected from the MARIO-LAION dataset [9] as the validation set. Note that the validation does not overlap with the benchmark.", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Metrics", + "text_level": 1, + "bbox": [ + 513, + 90, + 609, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We evaluate the quality of generated images using a GPT-based evaluation similar to TextDiffuser-2 [10]. In our evaluation, we adopt a rating-based evaluation approach [15] instead of the voting evaluation because GPT-based voting is sensitive to the option order and the number of options. In our settings, GPT assigns scores from 1 to 10 for both graphic design quality and content alignment between generated images and prompts. For text rendering accuracy, we measure OCR accuracy commonly used for assessing text rendering performance used in text-to-image generation [9, 10]. We also measure Fréchet Inception Distance (FID) [13] to quantify distribution differences between generated images and the full 5000 images in the MARIO-Eval benchmark, and CLIP score [32] to compute the similarity between each image and its prompt-based on the CLIP.", + "bbox": [ + 511, + 116, + 906, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 361, + 730, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use multiple external models in our Type-R pipeline without fine-tuning. We use Deepsolo [48] for text region detection, Paddle [5] for scene text recognition, LaMa [38] for text erasing, GPT-4o [4] for layout regeneration, and AnyText [39] for text editing. In addition, we use GPT to augment prompts following [42]. When augmenting prompts with GPT, we specify what categories of images to generate and how to draw the text in quotation marks to make it stand out. For all GPT utilization in this paper, i.e., augmentation, layout planning, and evaluation, we use the same version of OpenAI API 2023-05-15.", + "bbox": [ + 511, + 387, + 906, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When the response of the API is corrupted, we retry requests up to five times and give up. For typo correction iteration, we set the maximum number of correction attempts $t_{\\mathrm{max}}$ to 10. We chose the hyperparameter $\\theta$ for filtering to $4\\%$ of the overall image height through grid search.", + "bbox": [ + 511, + 556, + 905, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.4. Baselines", + "text_level": 1, + "bbox": [ + 513, + 651, + 620, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since Type-R is a post-processing module, we combine Type-R with the two base text-to-image models to build a baseline. In addition, we chose a few different baselines tailored for text rendering.", + "bbox": [ + 511, + 676, + 905, + 737 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Type-R w/ SD3 combines Stable Diffusion 3 [11] and our approach.", + "bbox": [ + 511, + 739, + 905, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Type-R w/ Flux combines Flux [3] and our approach.", + "bbox": [ + 511, + 773, + 870, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "TextDiffuser [9] is a text-to-image generation method based on layout planning, which generates character-wise masks and generates images using the masks as conditions.", + "bbox": [ + 511, + 791, + 905, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "TextDiffuser-2 [10] is a state-of-the-art text-to-image generation method based on layout planning. Unlike TextDiffuser, TextDiffuser-2 plans layouts via LLM and generates images by embedding the layout information into prompts.", + "bbox": [ + 511, + 839, + 903, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2748", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/994473edf7b9c9b48ce64157a4252de501bb31254371053c3944a85baa6e3932.jpg", + "image_caption": [ + "Figure 3. Comparisons of generated images by Flux, Flux w/ Type-R, and TextDiffuser models. The left column shows prompts, and the right columns present generated images by each method." + ], + "image_footnote": [], + "bbox": [ + 130, + 94, + 861, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.5. Results", + "text_level": 1, + "bbox": [ + 89, + 523, + 184, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Type-R vs. TextDiffusers Table 1 summarizes evaluation metrics of TextDiffuser models and Type-R. The result suggests that Type-R significantly outperforms the TextDiffuser models in GPT evaluation scores, and also Type-R achieves higher content alignment. The higher GPT scores indicate better aesthetic quality, as also suggested in the qualitative results in Figure 3. While the TextDiffuser models tend to generate images with simple backgrounds and texts (i.e., low quality in terms of graphic designs), Type-R successfully generates images with contextually rich backgrounds and seamlessly integrated text through precise and corrective adjustments.", + "bbox": [ + 89, + 547, + 482, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In OCR evaluations, SD3 with Type-R yields lower scores than the TextDiffuser models, whereas Flux with Type-R achieves higher accuracy. This result suggests that the enhancement of Type-R is influenced by the capability of the base text-to-image model.", + "bbox": [ + 89, + 729, + 482, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "User Study We also conducted a user study in which participants rated images using the same questions and scoring criteria as the GPT evaluations to validate the reliability of the GPT evaluation. We prepared 100 prompts and generated 100 corresponding images, which resulted in scores", + "bbox": [ + 89, + 824, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/32f502007b201536385bdd6ebdd2819612acee1b9305bde501d35efd2cace022.jpg", + "table_caption": [ + "Table 1. Evaluation metrics for Type-R and TextDiffuser baselines. \"Graphic\" and \"Match\" indicate graphic design quality and content alignment." + ], + "table_footnote": [], + "table_body": "
MethodGPT↑OCR ↑FID↓CLIP ↑
GraphicMatch
TD [9]4.677.1655.442.034.4
TD-2 [10]4.977.1756.233.834.7
Type-R
SD3[11]7.308.4648.645.033.0
Flux[3]7.678.5562.043.133.1
", + "bbox": [ + 511, + 575, + 911, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "from 5 participants for a total of 500 ratings per method.", + "bbox": [ + 511, + 717, + 885, + 732 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2 presents the user study results. Both SD3 and Flux with Type-R have higher scores than the TextDiffuser models, consistent with the GPT evaluation. The differences in scores are small among all baselines, although we observe that all scores of Type-R are statistically significant compared to TextDiffuser baselines. The p-value of the paired t-test with Type-R w/ Flux is $1.55e - 11$ for TextDiffuser and $4.14e - 18$ for TextDiffuser 2 in text-image matching evaluation. The small difference might be due to a lack of confidence in rating design or the ability to perceive finer defects in graphic design.", + "bbox": [ + 509, + 734, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2749", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ba8dca761a6201232d625dc4b954dfdbaf04e491598972a3681426f3e1df69fa.jpg", + "table_caption": [ + "Table 2. Userstudy for reliability of GPT evaluation." + ], + "table_footnote": [], + "table_body": "
MethodHuman ↑
GraphicMatch
TextDiffuser [9]5.24 ±2.185.08 ±2.51
TextDiffuser-2 [10]5.30 ±1.605.06 ±2.39
Type-R
SD3[11]5.94 ±1.865.99 ±2.34
Flux[3]6.23 ±1.816.04 ±2.34
", + "bbox": [ + 133, + 114, + 444, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/42a168055db44ba50a930569c8f5ba22ac5c5a342d1b3f91eb43fe74b513f216.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 168, + 238, + 274, + 318 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b2a4f9edad73f22541b842e6e3fb99d780aa16e556cba528984318ccb6759f48.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 238, + 401, + 318 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2617594a8f8c1b70525e15f7ff3a8f3fe5dbf0feb1e0fd9254017515ddaa0191.jpg", + "image_caption": [ + "B&W" + ], + "image_footnote": [], + "bbox": [ + 166, + 320, + 274, + 410 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/59c4635c3f14832267ec908a5db311575123405d6357b0e75b409e77b3faec1d.jpg", + "image_caption": [ + "w/ white box", + "Figure 4. Examples of naive baselines." + ], + "image_footnote": [], + "bbox": [ + 295, + 319, + 401, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FID and CLIP scores Although TextDiffuser yields higher FID and CLIP scores, we argue that these scores may not accurately capture human perception of image quality. For FID, the reference dataset is limited to 5,000 images that are not always consistent in quality, which likely causes the FID scores to overfit the narrow reference corpus, making it unsuitable for evaluating graphic design quality.", + "bbox": [ + 88, + 482, + 482, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TextDiffuser also yields higher CLIP scores. We suspect this result is due to the text-spotting bias of the CLIP model, which tends to \"parrot\" the visual text embedding in images [24]. That is, the CLIP score tends to prefer words over visual objects in the image, leading to a higher score for the word-only image without a meaningful background visual, which TextDiffusers tend to generate.", + "bbox": [ + 88, + 589, + 482, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.6. Additional Study", + "text_level": 1, + "bbox": [ + 89, + 707, + 256, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quality-accuracy Trade-off Here, we compare Type-R with several baselines in terms of the trade-off between image quality and OCR accuracy. In this study, we introduce the following backend text-to-image models.", + "bbox": [ + 89, + 731, + 482, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Type-R w/Dall-E 3 combines Dall-E 3 [1] with Type-R.", + "bbox": [ + 89, + 792, + 467, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Type-R w/ SD 3.5 combines Stable Diffusion 3.5 [6] with Type-R.", + "bbox": [ + 89, + 809, + 480, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare the raw output (Raw) from the backend text-to-image models and the output with Type-R. We also consider the following naive baselines (Fig. 4). We heuristically arrange text boxes for all the naive baselines, ignoring", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cdcc5f6e3ac771a26980494add07b8371df32e6c58528a0d5d93c4ad56a0b36b.jpg", + "image_caption": [ + "Figure 5. Plot of the relation between OCR accuracy and graphic design quality by GPT. Raw represents the results of text-to-image generation with prompt augmentation." + ], + "image_footnote": [], + "bbox": [ + 517, + 93, + 893, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the image content.", + "bbox": [ + 511, + 489, + 638, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Simple text erases all texts in the generated image from the backend text-to-image models and uses it as a background.", + "bbox": [ + 511, + 503, + 905, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Simple text w/ white box also removes all texts and white out the text box regions.", + "bbox": [ + 511, + 534, + 903, + 564 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "B&W always renders text on a white background.", + "bbox": [ + 511, + 565, + 844, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 5 plots the trade-off between GPT score and OCR accuracy, which shows that naive baselines can achieve high OCR accuracy for much lower GPT score. On the other hand, Flux and Dall-E 3 with Type-R show much higher graphic design quality with a slightly degrading OCR accuracy, almost reaching the accuracy of the simple text rendering. Type-R exhibits robustness across multiple backend text-to-image models, as demonstrated in the plot where Type-R effectively improves OCR accuracy for any backend models while maintaining high graphic design quality.", + "bbox": [ + 511, + 580, + 905, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Results Figure 6 shows a few generated images by Type-R demonstrating how each external model improves text rendering accuracy. In the first row, the text-editing model accurately modifies the text using a fancy font that aligns with the surrounding context. The second row presents an instance of Type-R successfully removing unnecessary texts. The third row provides an example of removing small text and adding a missing text through layout planning, where the inserted new text implicitly utilizes surrounding font styles to achieve a seamless and natural", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2750", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/097d5dc41a14f77cf80af4a6599ef89c9578f88919fc24c3b72a9733fcdfcb58.jpg", + "image_caption": [ + "Prompt: 'Nobliaux' et sorcires" + ], + "image_footnote": [], + "bbox": [ + 94, + 101, + 220, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/03527dabf3d3dba81a3cb9fd617edb84ed345a4e6e14db1db2aeaf8d1b46db52.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 101, + 349, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/02f0db7562851db71bdda99c7e96ab3afad03c2f9651eb7287b9761c7f11c645.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 101, + 477, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6a0f62b2de2cdc51d5a35a7b5e58471cff8416acd30f45021839cd34dde9dc1f.jpg", + "image_caption": [ + "Prompt: Image of MINI 'GPS' MAGNETIC CAR TRACKER", + "Prompt: Title 'Elise Chain 'Reaction' Large Paperback" + ], + "image_footnote": [], + "bbox": [ + 94, + 215, + 223, + 315 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f359ce52b84a2a0958efc3e8ea14646469578017679795e8617532dd2654d9a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 215, + 349, + 315 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4981f8ad047343329bd9a243fdc5c2a678be5cfeb3c1cf2e5963eafbe696588c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 215, + 477, + 315 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/49c016edf16e509923513e0cdaff0eede1b5fbeb649ec02c1f24c1e403dc4fa7.jpg", + "image_caption": [ + "Text-to-Image", + "Figure 6. Examples of generated images through Type-R." + ], + "image_footnote": [], + "bbox": [ + 96, + 333, + 223, + 433 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/88e31cc8199f472b5ba913e30283d1d113c6ec20a30ebc7f3ce3f9d9c579f722.jpg", + "image_caption": [ + "Text erasing $^+$ Layout regeneration" + ], + "image_footnote": [], + "bbox": [ + 233, + 333, + 349, + 433 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2b3d2cd5972aead784b7eb070b03c3099aeacb04bd185fc1f48dc689194b06a7.jpg", + "image_caption": [ + "Typo Corrected image" + ], + "image_footnote": [], + "bbox": [ + 361, + 333, + 475, + 433 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "appearance.", + "bbox": [ + 89, + 513, + 171, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since Type-R is independent of the backend text-to-image models, we can take advantage of the state-of-the-art models. Figure 7 presents examples with very long prompts, showing that the latest text-to-image model, Flux, accurately captures detailed information from these prompts, whereas TextDiffuser-2 generates corrupted images when handling such prompts, as CLIP is unable to process tokens exceeding 77.", + "bbox": [ + 88, + 527, + 482, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Longer Iterations The final typo correction step iteratively refines misspelled words. We show the improvement over iterations in Fig. 8 for several backend models. Our typo correction approach consistently improves OCR accuracy as we repeat the iterations. Our approach achieves $80\\%$ of the total improvement within four iterations independent of the backend model.", + "bbox": [ + 89, + 669, + 482, + 775 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.7. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 786, + 243, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Type-R Pipeline Modules Tab. 3 summarizes the ablation results for the modules in Type-R pipeline: augmentation for prompt, layout correction (text erasing and layout regeneration), and typo correction. We observe that each module effectively improves OCR accuracy with minimal negative impact on graphic scores by GPT. Layout correc", + "bbox": [ + 89, + 809, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7c8ffddcf5ea391de4fbe82f3a89910822e2b53feb12875a4c1ab9b0213f800d.jpg", + "table_caption": [ + "Table 3. Ablation study for Type-R pipeline modules." + ], + "table_footnote": [], + "table_body": "
T2IAugCorrectionGraphic ↑OCR ↑
LayoutTypo
SD3 [11]✓6.778.6
7.5512.2
✓✓7.4227.4
✓✓7.5320.4
✓✓✓7.2750.0
Flux [3]✓6.7914.4
7.8028.0
✓✓7.8353.8
✓✓7.7329.4
✓✓✓7.6563.0
", + "bbox": [ + 514, + 114, + 903, + 300 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9734dcb2ac198f20b5f92fecdb0aeddad6c0595b5716c7cdae02c0d7fc0fcf23.jpg", + "table_caption": [ + "Table 4. Comparing backend OCR models." + ], + "table_footnote": [], + "table_body": "
DetectionRecognitionOCR ↑
Paddle [5]Paddle [5]57.4
CRAFT [8]Paddle [5]47.2
Hi-SAM [49]Paddle [5]60.8
Mask TextSpotter v3 [22]Paddle [5]61.0
DeepSolo [48]Paddle [5]63.0
DeepSolo [48]Baek et al. [7]62.0
DeepSolo [48]TrOCR [19]60.0
DeepSolo [48]CLIP4STR-B* [33]60.8
", + "bbox": [ + 517, + 340, + 903, + 483 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion significantly improves Flux's OCR accuracy and further boosts OCR performance in combination with typo correction. The comparison between Flux and SD3 suggests Flux exhibits lower rates of glyph corruption.", + "bbox": [ + 511, + 510, + 905, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "OCR Models We summarize the effect of replacing the backend OCR models in Tab. 4. Both detection and recognition models impact OCR accuracy, with detection model performance changes having a more substantial effect than those of the recognition model. We observe that CRAFT [8] frequently merges words separated by spaces into a unit, leading to a notable performance drop, as Type-R assumes that spaces are excluded from texts and does not account for this issue. DeepSolo [48] achieves the highest performance.", + "bbox": [ + 511, + 597, + 905, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For text recognition, improvements in OCR accuracy do not always align with scene text recognition performance. While CLIP4STR-B* [33] is one of the state-of-the-art scene text recognition models, the OCR accuracy of CLIP4STR-B* is lower than the OCR accuracy of Baek et al. [7] and Paddle [5]. Paddle achieves the highest performance, likely because AnyText [39], trained with a loss function that incorporates the Paddle model, tends to generate words that are more easily recognized by Paddle. This suggests that compatibility between the text editing and recognition models is essential for effective correction.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2751", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/96023c0f29d2b4c6c231d855e275d63ac1e3c2c0fe51e7cd4e87f7e74c249053.jpg", + "image_caption": [], + "image_footnote": [ + "Create a sleek and dynamic MLB logo poster for the Colorado rockies: Use a bold color palette featuring deep purples, black, and metallic silver. Incorporate iconic mountains and baseball imagery, with prominent use of the text Colorado rockedies: Ensure the text is stylishly integrated and clearly legible. Keep the design clean and simple, focusing on strong, sharp lines that convey professional atheticism and team spirit.", + "A beautifully designed poster showcasing Brazilian cultural elements, suitable for a gallery. The composition features vibrant colors such as yellow, green, and blue, representing the Brazilian flag. In the center, there is a cheerful smiley face with a glossy finish. Below the smiley face, in bold and elegant typography, the word 'Obrigado' is prominently displayed. Traditional Brazilian patterns and motifs subtly decorate the background, adding a festive and welcoming atmosphere. The overall design is simple yet striking, perfect for a gallery wrap canvas that conveys gratitude with a touch of Brazilian charm." + ], + "bbox": [ + 312, + 88, + 604, + 202 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3ad0a4f8f8c705f8fa7b6e5b421ab8398761fb553c2704e7af18fdea4392a275.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 88, + 895, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bf9545ff7b8359d3b1e1cabf533767cd1c2c77a334f1d612c1073c58e1e126c7.jpg", + "image_caption": [ + "Flux" + ], + "image_footnote": [], + "bbox": [ + 328, + 203, + 449, + 311 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/73d8fd7669ab051f04d2aa858cb42b69067fbb7701ca1952ed74ebd97714cc79.jpg", + "image_caption": [ + "Type-R w/ Flux" + ], + "image_footnote": [], + "bbox": [ + 472, + 205, + 598, + 311 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5643d23e340a61efa2829c5f47a161b838326851e0a4002830615ce39303bc0a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 606, + 200, + 702, + 313 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/108f755b3de14df4ce443cfae66a02a9c04b2c42e488083e1caae1a0a4def7e1.jpg", + "image_caption": [ + "TextDiffuser-2", + "Figure 7. Generated images from long prompts by Flux, Flux w/ Type-R and TextDiffuser-2." + ], + "image_footnote": [], + "bbox": [ + 705, + 200, + 897, + 314 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4ff027d1d2535b97d7ef0aa6403950d3cacbe4eb174784bc1dd3c9393a411b12.jpg", + "image_caption": [ + "Figure 8. Relationship in the typo correction iteration and OCR* accuracy. Note that OCR* accuracy is computed by OCR models in Type-R which is distinct from the OCR model for evaluation." + ], + "image_footnote": [], + "bbox": [ + 104, + 380, + 475, + 513 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/733f1d0e2a9fc9359dae15bae4357cccb8217260fce82a5f7b87431fd1398fa3.jpg", + "table_caption": [ + "Table 5. Comparing backend text erasers." + ], + "table_footnote": [], + "table_body": "
MethodErase allOCR ↑
LaMa [37]✓62.4
63.0
Garnet [17]✓53.4
53.2
", + "bbox": [ + 176, + 611, + 398, + 693 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Text Erasing Methods We present the OCR accuracy when replacing the backend text erasing model in Tab. 5. In this study, we compare LaMa [37], a generic inpainting model, and Garnet, a text-specific inpainting model. Since inpainting models tend to be influenced by surrounding text regions, we optionally erase all text regions to prevent filling in letter-like artifacts and only focus on the background. Table 5 shows that LaMa performs well, and erasing all texts slightly improves the final OCR accuracy. We observe that Garnet tends to leave outlines of texts when erasing text, and the outlines seem to be recognized as words in the OCR evaluation.", + "bbox": [ + 89, + 719, + 483, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitation", + "text_level": 1, + "bbox": [ + 513, + 378, + 627, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Type-R depends on the performance of several external models and the base text-to-image models. We observe that OCR models sometimes fail to correctly parse all texts generated by text-to-image models. We also observe that AnyText [39] tends to generate unstable outlines of texts and degrading brightness of pixels from original images, which tends to cause incompatibility in composition. Text-to-image models sometimes fail to render with detailed layout specifications and seem to suffer from uppercase and lowercase distinctions. Although our method is applicable in any combination, we have to improve these backend models to resolve the failure cases.", + "bbox": [ + 511, + 404, + 906, + 584 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "There are other minor remaining problems. Type-R decomposes texts into words by the space. There are texts whose reading order is meaningful, but Type-R does not guarantee keeping the reading order of words in the generated image. Type-R has a relatively high computational requirement due to the stacking of multiple backend models in the pipeline, though it might be able to use the same backend in multiple stages to mitigate the issue; e.g., the test editing model can also erase texts.", + "bbox": [ + 511, + 585, + 908, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 738, + 633, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose an automatic typography correction method named Type-R, which automatically detects and fixes typographic errors in the image generated by text-to-image models. Through extensive experiments, we show that Type-R effectively corrects typos while maintaining the quality of the generated images and outperforms baselines in terms of the quality-accuracy trade-off. Our approach does not require training or fine-tuning and is readily plug-gable to any existing text-to-image model.", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2752", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] DALL-E 3. https://openai.com/dall-e-3. Accessed: 2024-03-15. 2, 6", + "[2] DeepFloyd IF. https://github.com/deep-floyd/IF. Accessed: 2024-11-14. 2", + "[3] FLUX.1 [dev]. https://blackforestlabs.ai/announcing-black-forest-labs/. Accessed: 2024-11-14. 1, 2, 4, 5, 6, 7", + "[4] GPT-4o. https://openai.com/index/gpt-4o-system-card/. Accessed: 2024-11-14. 3, 4", + "[5] PaddlePaddle. https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.7/doc/doc_ch/PP-OCRv4_introduction.md.4,7", + "[6] Stable Diffusion 3.5. https://huggingface.co/stabilityai/stable-diffusion-3.5-large. Accessed: 2024-11-8. 1, 6", + "[7] Jeonghun Baek, Geewook Kim, Junyeop Lee, Sungrae Park, Dongyoon Han, Sangwoo Yun, Seong Joon Oh, and Hwalsuk Lee. What is wrong with scene text recognition model comparisons? dataset and model analysis. In ICCV, 2019. 7", + "[8] Youngmin Baek, Bado Lee, Dongyoon Han, Sangdoo Yun, and Hwalsuk Lee. Character region awareness for text detection. In CVPR, 2019. 7", + "[9] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser: Diffusion models as text painters. In NeurIPS, 2023. 2, 4, 5, 6", + "[10] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser-2: Unleashing the power of language models for text rendering. In ECCV, 2024. 2, 4, 5, 6", + "[11] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 2, 4, 5, 6, 7", + "[12] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. In NeurIPS, 2023. 3", + "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. arXiv preprint arXiv:1706.08500, 2017. 4", + "[14] Jiabao Ji, Guanhua Zhang, Zhaowen Wang, Bairu Hou, Zhifei Zhang, Brian Price, and Shiyu Chang. Improving diffusion models for scene text editing with dual encoders. arXiv preprint arXiv:2304.05568, 2023. 2", + "[15] Peidong Jia, Chenxuan Li, Zeyu Liu, Yichao Shen, Xingru Chen, Yuhui Yuan, Yinglin Zheng, Dong Chen, Ji Li, Xiaodong Xie, et al. Cole: A hierarchical generation framework for graphic design. arXiv preprint arXiv:2311.16974, 2023. 4", + "[16] Sanyam Lakhanpal, Shivang Chopra, Vinija Jain, Aman Chadha, and Man Luo. Refining text-to-image generation:" + ], + "bbox": [ + 93, + 114, + 480, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Towards accurate training-free glyph-enhanced image generation. arXiv preprint arXiv:2403.16422, 2024. 2", + "[17] Hyeonsu Lee and Chankyu Choi. The surprisingly straightforward scene text removal method with gated attention and region of interest generation: A comprehensive prominent model analysis. In ECCV, 2022. 8", + "[18] Chao Li, Chen Jiang, Xiaolong Liu, Jun Zhao, and Guoxin Wang. Joytype: A robust design for multilingual visual text creation. arXiv preprint arXiv:2409.17524, 2024. 2", + "[19] Minghao Li, Tengchao Lv, Jingye Chen, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, and Furu Wei. Trocr: Transformer-based optical character recognition with pre-trained models. In AAAI, 2023. 7", + "[20] Wenbo Li, Guohao Li, Zhibin Lan, Xue Xu, Wanru Zhuang, Jiachen Liu, Xinyan Xiao, and Jinsong Su. Empowering backbone models for visual text generation with input granularity control and glyph-aware training. arXiv preprint arXiv:2410.04439, 2024. 2", + "[21] Zhenhang Li, Yan Shu, Weichao Zeng, Dongbao Yang, and Yu Zhou. First creating backgrounds then rendering texts: A new paradigm for visual text blending. In ECAI, 2024. 2", + "[22] Minghui Liao, Guan Pang, Jing Huang, Tal Hassner, and Xiang Bai. Mask TextSpotter v3: Segmentation proposal network for robust scene text spotting. In ECCV, 2020. 7", + "[23] Jiawei Lin, Jiaqi Guo, Shizhao Sun, Zijiang Yang, JianGuang Lou, and Dongmei Zhang. Layoutprompter: Awaken the design ability of large language models. In NeurIPS, 2023. 3", + "[24] Yiqi Lin, Conghui He, Alex Jinpeng Wang, Bin Wang, Wei-jia Li, and Mike Zheng Shou. Parrot captions teach clip to spot text. In ECCV, 2024. 6", + "[25] Rosanne Liu, Dan Garrette, Chitwan Sahara, William Chan, Adam Roberts, Sharan Narang, Irina Blok, RJ Mical, Mohammad Norouzi, and Noah Constant. Character-aware models improve visual text rendering. In ACL, 2023. 1, 2", + "[26] Zeyu Liu, Weicong Liang, Zhanhao Liang, Chong Luo, Ji Li, Gao Huang, and Yuhui Yuan. Glyph-byt5: A customized text encoder for accurate visual text rendering. In ECCV, 2024. 2", + "[27] Zeyu Liu, Weicong Liang, Yiming Zhao, Bohan Chen, Ji Li, and Yuhui Yuan. Glyph-byt5-v2: A strong aesthetic baseline for accurate multilingual visual text rendering. arXiv preprint arXiv:2406.10208, 2024. 2", + "[28] Jian Ma, Mingjun Zhao, Chen Chen, Ruichen Wang, Di Niu, Haonan Lu, and Xiaodong Lin. Glyphdraw: Seamlessly rendering text with intricate spatial structures in text-to-image generation. arXiv preprint arXiv:2303.17870, 2023. 2", + "[29] KhayTze Peong, Seiichi Uchida, and Daichi Haraguchi. Typographic text generation with off-the-shelf diffusion model. In ICDAR, 2024. 2", + "[30] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2023. 1, 2", + "[31] Yadong Qu, Qingfeng Tan, Hongtao Xie, Jianjun Xu, Yuxin Wang, and Yongdong Zhang. Exploring stroke-level modifications for scene text editing. In AAAI, 2023. 2" + ], + "bbox": [ + 517, + 93, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2753", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 4", + "[33] Miao Rang, Zhenni Bi, Chuanjian Liu, Yunhe Wang, and Kai Han. An empirical study of scaling law forOCR. In CVPR, 2024. 7", + "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2", + "[35] Prasun Roy, Saumik Bhattacharya, Subhankar Ghosh, and Umapada Pal. STEFANN: scene text editor using font adaptive neural network. In CVPR, 2020. 2", + "[36] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2", + "[37] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In WACV, 2022. 3, 8", + "[38] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 4", + "[39] Yuxiang Tuo, Wangmeng Xiang, Jun-Yan He, Yifeng Geng, and Xuansong Xie. Anytext: Multilingual visual text generation and editing. In ICLR, 2023. 2, 4, 7, 8", + "[40] Aoqiang Wang, Jian Wang, Zhenyu Yan, Wenxiang Shang, Ran Lin, and Zhao Zhang. Textmaster: Universal controllable text edit. arXiv preprint arXiv:2410.09879, 2024.", + "[41] Yibin Wang, Weizhong Zhang, Jianwei Zheng, and Cheng Jin. High fidelity scene text synthesis. arXiv preprint arXiv:2405.14701, 2024. 2", + "[42] Jules White, Quchen Fu, Sam Hays, Michael Sandborn, Carlos Olea, Henry Gilbert, Ashraf Elnashar, Jesse Spencer-Smith, and Douglas C Schmidt. A prompt pattern catalog to enhance prompt engineering with chatgpt. arXiv preprint arXiv:2302.11382, 2023. 4", + "[43] Liang Wu, Chengquan Zhang, Jiaming Liu, Junyu Han, Jingtuo Liu, Errui Ding, and Xiang Bai. Editing text in the wild. In ACM MM, 2018. 2", + "[44] Tsung-Han Wu, Long Lian, Joseph E Gonzalez, Boyi Li, and Trevor Darrell. Self-correcting llm-controlled diffusion models. In CVPR, 2024. 2", + "[45] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and CUI Bin. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. In ICML, 2024. 2", + "[46] Qiangpeng Yang, Jun Huang, and Wei Lin. Swaptext: Image based texts transfer in scenes. In CVPR, pages 14700-14709, 2020. 2" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Yukang Yang, Dongnan Gui, Yuhui Yuan, Weicong Liang, Haisong Ding, Han Hu, and Kai Chen. Glyphcontrol: Glyph conditional control for visual text generation. In NeurIPS, 2023. 2", + "[48] Maoyuan Ye, Jing Zhang, Shanshan Zhao, Juhua Liu, Tongliang Liu, Bo Du, and Dacheng Tao. DeepSolo: Let transformer decoder with explicit points solo for text spotting. In CVPR, 2023. 4, 7", + "[49] Maoyuan Ye, Jing Zhang, Juhua Liu, Chenyu Liu, Baocai Yin, Cong Liu, Bo Du, and Dacheng Tao. Hi-sam: Marrying segment anything model for hierarchical text segmentation. arXiv preprint arXiv:2401.17904, 2024. 7", + "[50] Weichao Zeng, Yan Shu, Zhenhang Li, Dongbao Yang, and Yu Zhou. TextCtrl: Diffusion-based scene text editing with prior guidance control. In NeurIPS, 2024. 2", + "[51] Lingjun Zhang, Xinyuan Chen, Yaohui Wang, Yue Lu, and Yu Qiao. Brush your text: Synthesize any scene text on images via diffusion model. In AAAI, 2024. 2", + "[52] Qilong Zhangli, Jindong Jiang, Di Liu, Licheng Yu, Xiaoliang Dai, Ankit Ramchandani, Guan Pang, Dimitris N Metaxas, and Praveen Krishnan. Layout-agnostic scene text image synthesis with diffusion models. In CVPR, 2024. 2", + "[53] Yiming Zhao and Zhouhui Lian. Udifftext: A unified framework for high-quality text synthesis in arbitrary images via character-aware diffusion models. In ECCV, 2024. 2", + "[54] Zhen Zhao, Jingqun Tang, Binghong Wu, Chunhui Lin, Shu Wei, Hao Liu, Xin Tan, Zhizhong Zhang, Can Huang, and Yuan Xie. Harmonizing visual text comprehension and generation. In NeurIPS, 2024. 2", + "[55] Yuanzhi Zhu, Jiawei Liu, Feiyu Gao, Wenyu Liu, Xinggang Wang, Peng Wang, Fei Huang, Cong Yao, and Zhibo Yang. Visual text generation in the wild. In ECCV, 2024. 2" + ], + "bbox": [ + 514, + 90, + 903, + 544 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2754", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Type-R_ Automatically Retouching Typos for Text-to-Image Generation/6f176b19-a787-4b85-9ebe-081a3f4759be_model.json b/2025/Type-R_ Automatically Retouching Typos for Text-to-Image Generation/6f176b19-a787-4b85-9ebe-081a3f4759be_model.json new file mode 100644 index 0000000000000000000000000000000000000000..802d98d3953aba0c83cce9779d646f7cc34177ea --- /dev/null +++ b/2025/Type-R_ Automatically Retouching Typos for Text-to-Image Generation/6f176b19-a787-4b85-9ebe-081a3f4759be_model.json @@ -0,0 +1,2706 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.139, + 0.131, + 0.86, + 0.154 + ], + "angle": 0, + "content": "Type-R: Automatically Retouching Typos for Text-to-Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.278, + 0.18, + 0.72, + 0.236 + ], + "angle": 0, + "content": "Wataru Shimoda1 Naoto Inoue1 Daichi Haraguchi1 Hayato Mitani2 Seiichi Uchida2 Kota Yamaguchi1 CyberAgent Kyushu University" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.31, + 0.287, + 0.411 + ], + "angle": 0, + "content": "Create a poster. Display the text 'CVPR' at the top of the image. NASHVILLE TENNESSEE is below the text. The date information 2025 JUNE is required. This image contains the picture of Music City Center at the lower right and cityscapes at the left." + }, + { + "type": "text", + "bbox": [ + 0.166, + 0.438, + 0.22, + 0.451 + ], + "angle": 0, + "content": "Prompt" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.277, + 0.493, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.356, + 0.431, + 0.428, + 0.457 + ], + "angle": 0, + "content": "Generated image" + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.27, + 0.696, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.438, + 0.571, + 0.451 + ], + "angle": 0, + "content": "Type-R" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.427, + 0.653, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.277, + 0.898, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.431, + 0.849, + 0.457 + ], + "angle": 0, + "content": "Typo-corrected image" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.481, + 0.908, + 0.51 + ], + "angle": 0, + "content": "Figure 1. Type-R automatically corrects typographic errors in an image generated from any text-to-image model without sacrificing their visual quality. We highlight prompt words corresponding to the words in the image with the same color." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.522, + 0.327, + 0.538 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.553, + 0.486, + 0.782 + ], + "angle": 0, + "content": "While recent text-to-image models can generate photorealistic images from text prompts that reflect detailed instructions, they still face significant challenges in accurately rendering words in the image. In this paper, we propose to retouch erroneous text renderings in the post-processing pipeline. Our approach, called Type-R, identifies typographical errors in the generated image, erases the erroneous text, regenerates text boxes for missing words, and finally corrects typos in the rendered words. Through extensive experiments, we show that Type-R, in combination with the latest text-to-image models such as Stable Diffusion or Flux, achieves the highest text rendering accuracy while maintaining image quality and also outperforms text-focused generation baselines in terms of balancing text accuracy and image quality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.808, + 0.222, + 0.823 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.833, + 0.485, + 0.88 + ], + "angle": 0, + "content": "The rapid advancement of text-to-image generation models triggers a surge of application interest in various creative domains. Graphic design is one of the domains from which" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.72 + ], + "angle": 0, + "content": "text-to-image models potentially benefit, with the goal of helping generate appealing posters, advertisements, book covers, webpages, or signs, such that these display media convey the intended message to the viewers. However, one of the major obstacles is text rendering on the image [25]. Popular state-of-the-art text-to-image models such as Stable Diffusion [6, 30] or Flux [3] still face notable challenges in accurately rendering words or characters (Fig. 1 second column). Common issues include missing words, unintended characters, or word-level errors that insert another similar-looking word. While it might be possible to tune the text-to-image models with increasing amounts of training data, the resulting image is not guaranteed to be typo-free." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this work, instead of modifying the image generator, we propose automatically retouching the erroneous text renderings of the already generated image. Our approach, named Type-R, detects typographic errors in a generated image and automatically fixes erroneous content. Figure 1 shows an example of the input prompt, the generated image by Flux.1-dev [3], detected errors by Type-R, and the final retouched image. Type-R successfully corrects the typographic errors without artifacts compared to the initial generation. Our approach is fully automatic, independent of the base text-to-image generator, and easily pluggable to any existing text-to-image model." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.309, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)Code and project page are available." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2745" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.272 + ], + "angle": 0, + "content": "Through extensive experiments, we study the limitations in the current text-to-image models, including both the popular text-to-image generators [3, 30] and typophy-focussed models [9, 10], and show that Type-R significantly improves the text rendering accuracy of text-to-image models while maintaining graphic design quality. We also study several design choices, e.g., how changing the base text-to-image generator or the OCR model affects the resulting visual quality and text accuracy, and show that Type-R in combination with Flux [3] achieves the best trade-off between text rendering accuracy and quality scores by GPT evaluation on Mario-Eval benchmark [9]." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.274, + 0.437, + 0.288 + ], + "angle": 0, + "content": "We summarize our contributions in the following" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.289, + 0.483, + 0.335 + ], + "angle": 0, + "content": "1. We propose Type-R, which automatically retouches typographic errors in generated images by text-to-image models while maintaining the design quality." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.335, + 0.483, + 0.396 + ], + "angle": 0, + "content": "2. We empirically show that the combination of the state-of-the-art text-to-image generator and Type-R achieves the best quality-accuracy trade-off and outperforms the approach specifically designed for typography." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.289, + 0.483, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.41, + 0.233, + 0.425 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.435, + 0.327, + 0.451 + ], + "angle": 0, + "content": "2.1. Text-to-Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.483, + 0.625 + ], + "angle": 0, + "content": "Text-to-image generation is the task of creating images from text prompts. With the development of large image-capture datasets, there has been significant progress in generating photorealistic images from diffusion-based models [1, 2, 30, 34, 36]. However, rendering accurate text is still one of the remaining problems. Liu et al. [25] point out the challenge in accurately rendering characters in text-to-image generation models, where it is common to observe missing glyphs, redundant glyphs, or wrong characters. The text accuracy problem persists in the recent text-to-image models such as Stable Diffusion 3 [11] or Flux [3]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.626, + 0.483, + 0.791 + ], + "angle": 0, + "content": "In this work, instead of modifying or tuning the text-to-image models, we utilize feedback from external models to improve text rendering. Several works report the successful use of external models in improving image generation. RPG [45] introduces a technique for region-specific prompt guidance for noise control. Wu et al. [44] apply object detection to generated images and refine detection results using Large Language Models. In our task setup, we take advantage of pre-trained Vision Language Models (VLMs), OCR models, and text-inpainting models to retouch typographic errors." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.802, + 0.305, + 0.817 + ], + "angle": 0, + "content": "2.2. Visual Text Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Several text-to-image models [2, 26, 27, 36, 52] have reported that better text encoder significantly improves text rendering capabilities, where the encoder can learn glyph-level information from the training corpus. Adding OCR losses also improve performances [20, 39, 53]. While im" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.138 + ], + "angle": 0, + "content": "proving the capability of the base text-to-image generator might eventually reduce typographic errors, we consider a post-processing approach independent of the base model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.141, + 0.907, + 0.413 + ], + "angle": 0, + "content": "Another line of work to improve text rendering is via conditional inputs to the text-to-image models. Some literature attempts at conditioning generation on text positions and glyph masks to improve the text rendering accuracies [9, 10, 28, 29, 39, 47, 51, 54]. Particularly, TextDiffusers [9, 10] generate images with two steps: layout planning, which generates text position information from prompts, followed by image generation conditioned on the first layout information. TextDiffuser [9] and TextDiffuser 2 [9] differ in how they generate layout information, where the former generates character-by-character masks and the latter generates layout in language. One major limitation of conditional generation is the training requirements; i.e., building a sufficient training corpus for the conditioning module is often difficult. Our experiments suggest that TextDiffusers suffer from inadequate image quality compared to the state-of-the-art image generators, likely due to the bias in its training data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.909, + 0.614 + ], + "angle": 0, + "content": "We take the approach of retouching afterward using external pre-trained models to take full advantage of the best text-to-image models without fine-tuning. A few works incorporate external models for text rendering. Very recently, Lakhanpal et al. reports Refining T2I [16] that applies OCR to the generated images by TextDiffuser and optimizes the layouts to improve the legibility of texts, while the same limitation to TextDiffusers still applies. Li et al. [21] propose to generate background images by removing generated scene text. Unlike prior works, we propose a method for generating images that adheres as much as possible to the initial generation results, making full use of OCR models, text removal models, and VLMs to improve text rendering." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.636, + 0.755, + 0.652 + ], + "angle": 0, + "content": "2.3. Text Editing and Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.664, + 0.907, + 0.785 + ], + "angle": 0, + "content": "Another relevant task is raster text editing, which directly manipulates texts in the image. While GAN-based methods [31, 35, 43, 46] have been popular, recent visual text generation methods [14, 18, 39-41, 50, 55] are capable of editing texts in the image directly. While text editing and our task setup seem similar, text editing is only a part of the typography retouching task, which involves further tasks such as automatic error detection or missing word insertion." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.811, + 0.622, + 0.829 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Our Type-R pipeline consists of four stages, which are all automatic: 1) error detection, 2) text erasing, 3) layout regeneration, and 4) typo correction. We show the pipeline overview in Fig. 2." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2746" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.088, + 0.249, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.207, + 0.194, + 0.217 + ], + "angle": 0, + "content": "Generated image" + }, + { + "type": "image_footnote", + "bbox": [ + 0.1, + 0.225, + 0.244, + 0.278 + ], + "angle": 0, + "content": "Create an advertisement for a hamburger with a pop background. Display the texts '10% OFF and HAMBURGER' with a tasty burger, fried potatoes, and a Coke. The text 'Combo Meals' is required. Prompt" + }, + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.287, + 0.213, + 0.298 + ], + "angle": 0, + "content": "Generation" + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.088, + 0.409, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.211, + 0.409, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.282, + 0.287, + 0.386, + 0.298 + ], + "angle": 0, + "content": "Error detection" + }, + { + "type": "image", + "bbox": [ + 0.413, + 0.088, + 0.567, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.211, + 0.566, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.449, + 0.288, + 0.533, + 0.301 + ], + "angle": 0, + "content": "Text erasing" + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.088, + 0.725, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.211, + 0.724, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.288, + 0.718, + 0.301 + ], + "angle": 0, + "content": "Layout regeneration" + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.088, + 0.885, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.732, + 0.211, + 0.882, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.286, + 0.869, + 0.301 + ], + "angle": 0, + "content": "3 Typo correction" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.313, + 0.907, + 0.342 + ], + "angle": 0, + "content": "Figure 2. Illustration of the Type-R pipeline. Type-R automatically detects errors, erases unintended texts, inserts missing words, and corrects spelling errors in the image." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.368, + 0.25, + 0.383 + ], + "angle": 0, + "content": "3.1. Error Detection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.391, + 0.484, + 0.56 + ], + "angle": 0, + "content": "We denote an image \\(I\\) generated by an off-the-shelf text-to-image model \\(G\\) with a prompt, which requests to show \\(N\\) words \\(W = \\{w_{i}\\}_{i = 1}^{N}\\) in \\(I\\). To ensure how much this expectation is satisfied, we first detect words in the image \\(I\\). Specifically, a scene text detection model predicts polygonal word regions \\(\\hat{B}\\). We remove polygons with a height less than a threshold \\(\\theta\\). Then, a scene text recognition model parses the words \\(\\hat{W} = \\{\\hat{w}_i\\}_{i = 1}^{\\hat{N}}\\) in the regions \\(\\hat{B}\\). As noted in Section 1, the model \\(G\\) is often not perfect, and \\(W \\neq \\hat{W}\\); therefore, we need to identify errors in \\(\\hat{W}\\) by considering \\(W\\) is a ground-truth." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.561, + 0.484, + 0.621 + ], + "angle": 0, + "content": "Roughly speaking, this identification task is a matching problem between \\( W \\) and \\( \\hat{W} \\). More specifically, we consider the matching between the word set \\( \\mathcal{W} \\) and \\( \\hat{\\mathcal{W}} \\) of equal size \\( \\mathcal{N} \\) by optimal transport:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.632, + 0.482, + 0.674 + ], + "angle": 0, + "content": "\\[\n\\mathrm {O T} (\\mathcal {W}, \\hat {\\mathcal {W}}) \\equiv \\max _ {\\pi} \\sum_ {i = 1} ^ {\\mathcal {N}} \\operatorname {L e v} \\left(w _ {i}, \\hat {w} _ {\\pi (i)}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.484, + 0.794 + ], + "angle": 0, + "content": "where \\(w_{i}\\in \\mathcal{W},\\hat{w}_{i}\\in \\hat{\\mathcal{W}},\\pi\\) is a one-by-one matching, and \\(\\mathrm{Levenshtein}\\) distance to measure the text similarity. This matching works only when the two sets are of equal size, but typographic errors result in different sizes because of missing or unintended words. Here, we equalize two word sets \\(W\\) and \\(\\hat{W}\\) by padding tokens. Let us define \\(\\mathcal{N} = \\max (N,\\hat{N})\\). We express the padded word sets by:" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.819, + 0.482, + 0.838 + ], + "angle": 0, + "content": "\\[\n\\mathcal {W} = W \\cup \\left\\{p _ {i} \\right\\} _ {i = N + 1} ^ {\\mathcal {N}}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.84, + 0.482, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {W}} = \\hat {W} \\cup \\left\\{\\hat {p} _ {i} \\right\\} _ {i = \\hat {N} + 1} ^ {\\mathcal {N}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\( p_i \\) and \\( \\hat{p}_i \\) are the padding tokens. Here, a word matching a padding token indicates a typographic error; i.e.," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.369, + 0.905, + 0.429 + ], + "angle": 0, + "content": "a paired word with \\( p \\) is an unintended word, and a paired word with \\( \\hat{p} \\) is a missing word. We assign a constant distance to any word matching to the padding tokens. Any pairs having non-zero distance indicate spelling errors." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.43, + 0.906, + 0.52 + ], + "angle": 0, + "content": "The following sections describe our approach to minimizing the above optimal transportation cost. We align the number of words by erasing redundant texts when \\( N < \\hat{N} \\) (Sec. 3.2) and insert text boxes by layout regeneration when \\( N > \\hat{N} \\) (Sec. 3.3). Finally, we fix spelling errors in typo correction (Sec. 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.528, + 0.649, + 0.545 + ], + "angle": 0, + "content": "3.2. Text Erasing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.907, + 0.734 + ], + "angle": 0, + "content": "Type-R erases unintended texts using an off-the-shelf inpainting method [37] if the number of detected words exceeds the number of the words specified by a prompt, i.e., \\( N < \\hat{N} \\), or the generated image includes filtered texts. This step reduces the number of words in a generated image to equal or less than the number of specified words by the prompt. The inpainting model removes specified paint regions. In our setup, we specify regions that we want to erase in \\( \\hat{B} \\) and composite the inpainted results to the original image. Since text detection may not always fully cover text areas, we slightly enlarge the text region masks to ensure complete removal." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.742, + 0.714, + 0.759 + ], + "angle": 0, + "content": "3.3. Layout Regeneration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Layout regeneration fills in missing text regions so that the resulting \\(\\hat{N}\\) equals the number of words specified by a prompt \\(N\\) when \\(\\hat{N} < N\\). This is equivalent to obtaining position information from the image for missing words of the prompt, and we address this task using a VLM, specifically GPT-4o [4]. Inspired by works on layout generation by zero/few-shot learning of LLMs [12, 23], we ask GPT to plan layouts for each missing word from the image and layout information of valid text boxes in JSON format." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2747" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.168 + ], + "angle": 0, + "content": "For simplicity, we represent the layout information as a bounding box with four coordinates. When text detection outputs are in the polygonal form, we convert them to boxes by finding the smallest four-sided bounding box that encloses each polygon." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.18, + 0.254, + 0.196 + ], + "angle": 0, + "content": "3.4. Tipo Correction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.204, + 0.484, + 0.401 + ], + "angle": 0, + "content": "Typo correction tries to resolve all word pairs with non-zero edit distance. To address word-level spelling errors, we employ a pre-trained text editing model [39] to automatically retouch the word region. Using an inpainting approach, the text editing model generates raster images containing specified text in the selected areas. Similar to text-to-image models, text editing models may not always render the specified text accurately but with a much lower computation cost than large text-to-image models. We repeatedly apply the text editing model until we fix the spelling errors or we reach the predefined number of attempts. Note that we only composite the areas of corrected words in the edited image during the iteration. We show the pseudo-codes in Algorithm 1." + }, + { + "type": "code_caption", + "bbox": [ + 0.094, + 0.418, + 0.408, + 0.433 + ], + "angle": 0, + "content": "Algorithm 1 A pseudo-code of typo correction." + }, + { + "type": "algorithm", + "bbox": [ + 0.094, + 0.436, + 0.484, + 0.665 + ], + "angle": 0, + "content": "1: function TyPO CORRECTION(Image I, Words W, Detected regions \\(\\hat{B}\\) , Maximum iterations \\(t_\\mathrm{max})\\) \n2: for \\(t = 1\\) to \\(t_\\mathrm{max}\\) do \n3: \\(I_{\\mathrm{edit}}\\gets\\) Text-Editor(I,W,\\(\\hat{B}\\)) \n4: \\(\\hat{W}\\gets\\) Text-Recognition \\((I_{\\mathrm{edit}},\\hat{B})\\) \n5: \\(\\mathcal{O}\\gets \\mathrm{Diff}(W,\\hat{W})\\qquad \\triangleright\\) Find spelling errors \n6: \\(I\\gets\\) Composite \\((I,I_{\\mathrm{edit}},\\hat{B},\\mathcal{O})\\) \n7: if \\(\\mathcal{O}\\neq \\varnothing\\) then \n8: \\(W,\\hat{B}\\gets\\) Update(W,\\(\\hat{B},\\mathcal{O})\\) \n9: else \n10: break \n11: end if \n12: end for \n13: return I \n14: end function" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.7, + 0.216, + 0.717 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.726, + 0.187, + 0.74 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We evaluate the performance of visual text generation on MARIO-Eval benchmark [9]. This benchmark contains 5,414 prompts, of which 5,000 have image data. We use all 5,414 prompts in the main comparison, i.e., in Tab. 1. We randomly pick 500 of the 5,414 prompts as the subset for analysis in Sec. 4.6 because our experiments involve paid APIs. For the selection of hyperparameters in filtering, we use 500 prompts randomly selected from the MARIO-LAION dataset [9] as the validation set. Note that the validation does not overlap with the benchmark." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.61, + 0.106 + ], + "angle": 0, + "content": "4.2. Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.117, + 0.907, + 0.345 + ], + "angle": 0, + "content": "We evaluate the quality of generated images using a GPT-based evaluation similar to TextDiffuser-2 [10]. In our evaluation, we adopt a rating-based evaluation approach [15] instead of the voting evaluation because GPT-based voting is sensitive to the option order and the number of options. In our settings, GPT assigns scores from 1 to 10 for both graphic design quality and content alignment between generated images and prompts. For text rendering accuracy, we measure OCR accuracy commonly used for assessing text rendering performance used in text-to-image generation [9, 10]. We also measure Fréchet Inception Distance (FID) [13] to quantify distribution differences between generated images and the full 5000 images in the MARIO-Eval benchmark, and CLIP score [32] to compute the similarity between each image and its prompt-based on the CLIP." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.362, + 0.731, + 0.379 + ], + "angle": 0, + "content": "4.3. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.907, + 0.554 + ], + "angle": 0, + "content": "We use multiple external models in our Type-R pipeline without fine-tuning. We use Deepsolo [48] for text region detection, Paddle [5] for scene text recognition, LaMa [38] for text erasing, GPT-4o [4] for layout regeneration, and AnyText [39] for text editing. In addition, we use GPT to augment prompts following [42]. When augmenting prompts with GPT, we specify what categories of images to generate and how to draw the text in quotation marks to make it stand out. For all GPT utilization in this paper, i.e., augmentation, layout planning, and evaluation, we use the same version of OpenAI API 2023-05-15." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.558, + 0.906, + 0.634 + ], + "angle": 0, + "content": "When the response of the API is corrupted, we retry requests up to five times and give up. For typo correction iteration, we set the maximum number of correction attempts \\( t_{\\mathrm{max}} \\) to 10. We chose the hyperparameter \\( \\theta \\) for filtering to \\( 4\\% \\) of the overall image height through grid search." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.652, + 0.622, + 0.667 + ], + "angle": 0, + "content": "4.4. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.678, + 0.906, + 0.738 + ], + "angle": 0, + "content": "Since Type-R is a post-processing module, we combine Type-R with the two base text-to-image models to build a baseline. In addition, we chose a few different baselines tailored for text rendering." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.741, + 0.906, + 0.772 + ], + "angle": 0, + "content": "Type-R w/ SD3 combines Stable Diffusion 3 [11] and our approach." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.774, + 0.871, + 0.79 + ], + "angle": 0, + "content": "Type-R w/ Flux combines Flux [3] and our approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.792, + 0.906, + 0.838 + ], + "angle": 0, + "content": "TextDiffuser [9] is a text-to-image generation method based on layout planning, which generates character-wise masks and generates images using the masks as conditions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.902 + ], + "angle": 0, + "content": "TextDiffuser-2 [10] is a state-of-the-art text-to-image generation method based on layout planning. Unlike TextDiffuser, TextDiffuser-2 plans layouts via LLM and generates images by embedding the layout information into prompts." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2748" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.131, + 0.095, + 0.862, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.471, + 0.908, + 0.501 + ], + "angle": 0, + "content": "Figure 3. Comparisons of generated images by Flux, Flux w/ Type-R, and TextDiffuser models. The left column shows prompts, and the right columns present generated images by each method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.525, + 0.185, + 0.54 + ], + "angle": 0, + "content": "4.5. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.549, + 0.483, + 0.729 + ], + "angle": 0, + "content": "Type-R vs. TextDiffusers Table 1 summarizes evaluation metrics of TextDiffuser models and Type-R. The result suggests that Type-R significantly outperforms the TextDiffuser models in GPT evaluation scores, and also Type-R achieves higher content alignment. The higher GPT scores indicate better aesthetic quality, as also suggested in the qualitative results in Figure 3. While the TextDiffuser models tend to generate images with simple backgrounds and texts (i.e., low quality in terms of graphic designs), Type-R successfully generates images with contextually rich backgrounds and seamlessly integrated text through precise and corrective adjustments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.731, + 0.483, + 0.807 + ], + "angle": 0, + "content": "In OCR evaluations, SD3 with Type-R yields lower scores than the TextDiffuser models, whereas Flux with Type-R achieves higher accuracy. This result suggests that the enhancement of Type-R is influenced by the capability of the base text-to-image model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.903 + ], + "angle": 0, + "content": "User Study We also conducted a user study in which participants rated images using the same questions and scoring criteria as the GPT evaluations to validate the reliability of the GPT evaluation. We prepared 100 prompts and generated 100 corresponding images, which resulted in scores" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.524, + 0.907, + 0.568 + ], + "angle": 0, + "content": "Table 1. Evaluation metrics for Type-R and TextDiffuser baselines. \"Graphic\" and \"Match\" indicate graphic design quality and content alignment." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.577, + 0.912, + 0.688 + ], + "angle": 0, + "content": "
MethodGPT↑OCR ↑FID↓CLIP ↑
GraphicMatch
TD [9]4.677.1655.442.034.4
TD-2 [10]4.977.1756.233.834.7
Type-R
SD3[11]7.308.4648.645.033.0
Flux[3]7.678.5562.043.133.1
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.718, + 0.887, + 0.733 + ], + "angle": 0, + "content": "from 5 participants for a total of 500 ratings per method." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Table 2 presents the user study results. Both SD3 and Flux with Type-R have higher scores than the TextDiffuser models, consistent with the GPT evaluation. The differences in scores are small among all baselines, although we observe that all scores of Type-R are statistically significant compared to TextDiffuser baselines. The p-value of the paired t-test with Type-R w/ Flux is \\(1.55e - 11\\) for TextDiffuser and \\(4.14e - 18\\) for TextDiffuser 2 in text-image matching evaluation. The small difference might be due to a lack of confidence in rating design or the ability to perceive finer defects in graphic design." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2749" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.132, + 0.089, + 0.443, + 0.104 + ], + "angle": 0, + "content": "Table 2. Userstudy for reliability of GPT evaluation." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.115, + 0.446, + 0.226 + ], + "angle": 0, + "content": "
MethodHuman ↑
GraphicMatch
TextDiffuser [9]5.24 ±2.185.08 ±2.51
TextDiffuser-2 [10]5.30 ±1.605.06 ±2.39
Type-R
SD3[11]5.94 ±1.865.99 ±2.34
Flux[3]6.23 ±1.816.04 ±2.34
" + }, + { + "type": "image", + "bbox": [ + 0.169, + 0.239, + 0.276, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.239, + 0.403, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.168, + 0.321, + 0.276, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.32, + 0.403, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.205, + 0.413, + 0.234, + 0.422 + ], + "angle": 0, + "content": "B&W" + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.413, + 0.387, + 0.424 + ], + "angle": 0, + "content": "w/ white box" + }, + { + "type": "image_caption", + "bbox": [ + 0.17, + 0.441, + 0.402, + 0.455 + ], + "angle": 0, + "content": "Figure 4. Examples of naive baselines." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.483, + 0.483, + 0.59 + ], + "angle": 0, + "content": "FID and CLIP scores Although TextDiffuser yields higher FID and CLIP scores, we argue that these scores may not accurately capture human perception of image quality. For FID, the reference dataset is limited to 5,000 images that are not always consistent in quality, which likely causes the FID scores to overfit the narrow reference corpus, making it unsuitable for evaluating graphic design quality." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.59, + 0.483, + 0.696 + ], + "angle": 0, + "content": "TextDiffuser also yields higher CLIP scores. We suspect this result is due to the text-spotting bias of the CLIP model, which tends to \"parrot\" the visual text embedding in images [24]. That is, the CLIP score tends to prefer words over visual objects in the image, leading to a higher score for the word-only image without a meaningful background visual, which TextDiffusers tend to generate." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.708, + 0.258, + 0.725 + ], + "angle": 0, + "content": "4.6. Additional Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.732, + 0.483, + 0.793 + ], + "angle": 0, + "content": "Quality-accuracy Trade-off Here, we compare Type-R with several baselines in terms of the trade-off between image quality and OCR accuracy. In this study, we introduce the following backend text-to-image models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.794, + 0.468, + 0.809 + ], + "angle": 0, + "content": "Type-R w/Dall-E 3 combines Dall-E 3 [1] with Type-R." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.482, + 0.84 + ], + "angle": 0, + "content": "Type-R w/ SD 3.5 combines Stable Diffusion 3.5 [6] with Type-R." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We compare the raw output (Raw) from the backend text-to-image models and the output with Type-R. We also consider the following naive baselines (Fig. 4). We heuristically arrange text boxes for all the naive baselines, ignoring" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.094, + 0.895, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.421, + 0.907, + 0.464 + ], + "angle": 0, + "content": "Figure 5. Plot of the relation between OCR accuracy and graphic design quality by GPT. Raw represents the results of text-to-image generation with prompt augmentation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.491, + 0.64, + 0.505 + ], + "angle": 0, + "content": "the image content." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.505, + 0.906, + 0.535 + ], + "angle": 0, + "content": "Simple text erases all texts in the generated image from the backend text-to-image models and uses it as a background." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.905, + 0.565 + ], + "angle": 0, + "content": "Simple text w/ white box also removes all texts and white out the text box regions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.566, + 0.845, + 0.581 + ], + "angle": 0, + "content": "B&W always renders text on a white background." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.906, + 0.732 + ], + "angle": 0, + "content": "Fig. 5 plots the trade-off between GPT score and OCR accuracy, which shows that naive baselines can achieve high OCR accuracy for much lower GPT score. On the other hand, Flux and Dall-E 3 with Type-R show much higher graphic design quality with a slightly degrading OCR accuracy, almost reaching the accuracy of the simple text rendering. Type-R exhibits robustness across multiple backend text-to-image models, as demonstrated in the plot where Type-R effectively improves OCR accuracy for any backend models while maintaining high graphic design quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Qualitative Results Figure 6 shows a few generated images by Type-R demonstrating how each external model improves text rendering accuracy. In the first row, the text-editing model accurately modifies the text using a fancy font that aligns with the surrounding context. The second row presents an instance of Type-R successfully removing unnecessary texts. The third row provides an example of removing small text and adding a missing text through layout planning, where the inserted new text implicitly utilizes surrounding font styles to achieve a seamless and natural" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2750" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.21, + 0.09, + 0.356, + 0.101 + ], + "angle": 0, + "content": "Prompt: 'Nobliaux' et sorcires" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.102, + 0.222, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.102, + 0.351, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.102, + 0.478, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.204, + 0.437, + 0.216 + ], + "angle": 0, + "content": "Prompt: Image of MINI 'GPS' MAGNETIC CAR TRACKER" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.217, + 0.225, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.217, + 0.351, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.217, + 0.478, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.319, + 0.41, + 0.332 + ], + "angle": 0, + "content": "Prompt: Title 'Elise Chain 'Reaction' Large Paperback" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.334, + 0.225, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.436, + 0.201, + 0.447 + ], + "angle": 0, + "content": "Text-to-Image" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.334, + 0.351, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.436, + 0.348, + 0.458 + ], + "angle": 0, + "content": "Text erasing \\(^+\\) Layout regeneration" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.334, + 0.477, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.364, + 0.436, + 0.45, + 0.457 + ], + "angle": 0, + "content": "Typo Corrected image" + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.472, + 0.457, + 0.487 + ], + "angle": 0, + "content": "Figure 6. Examples of generated images through Type-R." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.173, + 0.528 + ], + "angle": 0, + "content": "appearance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.529, + 0.483, + 0.651 + ], + "angle": 0, + "content": "Since Type-R is independent of the backend text-to-image models, we can take advantage of the state-of-the-art models. Figure 7 presents examples with very long prompts, showing that the latest text-to-image model, Flux, accurately captures detailed information from these prompts, whereas TextDiffuser-2 generates corrupted images when handling such prompts, as CLIP is unable to process tokens exceeding 77." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.67, + 0.483, + 0.776 + ], + "angle": 0, + "content": "Longer Iterations The final typo correction step iteratively refines misspelled words. We show the improvement over iterations in Fig. 8 for several backend models. Our typo correction approach consistently improves OCR accuracy as we repeat the iterations. Our approach achieves \\(80\\%\\) of the total improvement within four iterations independent of the backend model." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.787, + 0.245, + 0.803 + ], + "angle": 0, + "content": "4.7. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Type-R Pipeline Modules Tab. 3 summarizes the ablation results for the modules in Type-R pipeline: augmentation for prompt, layout correction (text erasing and layout regeneration), and typo correction. We observe that each module effectively improves OCR accuracy with minimal negative impact on graphic scores by GPT. Layout correc" + }, + { + "type": "table_caption", + "bbox": [ + 0.55, + 0.09, + 0.869, + 0.104 + ], + "angle": 0, + "content": "Table 3. Ablation study for Type-R pipeline modules." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.115, + 0.905, + 0.301 + ], + "angle": 0, + "content": "
T2IAugCorrectionGraphic ↑OCR ↑
LayoutTypo
SD3 [11]✓6.778.6
7.5512.2
✓✓7.4227.4
✓✓7.5320.4
✓✓✓7.2750.0
Flux [3]✓6.7914.4
7.8028.0
✓✓7.8353.8
✓✓7.7329.4
✓✓✓7.6563.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.58, + 0.317, + 0.838, + 0.332 + ], + "angle": 0, + "content": "Table 4. Comparing backend OCR models." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.342, + 0.905, + 0.484 + ], + "angle": 0, + "content": "
DetectionRecognitionOCR ↑
Paddle [5]Paddle [5]57.4
CRAFT [8]Paddle [5]47.2
Hi-SAM [49]Paddle [5]60.8
Mask TextSpotter v3 [22]Paddle [5]61.0
DeepSolo [48]Paddle [5]63.0
DeepSolo [48]Baek et al. [7]62.0
DeepSolo [48]TrOCR [19]60.0
DeepSolo [48]CLIP4STR-B* [33]60.8
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.511, + 0.906, + 0.573 + ], + "angle": 0, + "content": "tion significantly improves Flux's OCR accuracy and further boosts OCR performance in combination with typo correction. The comparison between Flux and SD3 suggests Flux exhibits lower rates of glyph corruption." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.906, + 0.734 + ], + "angle": 0, + "content": "OCR Models We summarize the effect of replacing the backend OCR models in Tab. 4. Both detection and recognition models impact OCR accuracy, with detection model performance changes having a more substantial effect than those of the recognition model. We observe that CRAFT [8] frequently merges words separated by spaces into a unit, leading to a notable performance drop, as Type-R assumes that spaces are excluded from texts and does not account for this issue. DeepSolo [48] achieves the highest performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "For text recognition, improvements in OCR accuracy do not always align with scene text recognition performance. While CLIP4STR-B* [33] is one of the state-of-the-art scene text recognition models, the OCR accuracy of CLIP4STR-B* is lower than the OCR accuracy of Baek et al. [7] and Paddle [5]. Paddle achieves the highest performance, likely because AnyText [39], trained with a loss function that incorporates the Paddle model, tends to generate words that are more easily recognized by Paddle. This suggests that compatibility between the text editing and recognition models is essential for effective correction." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2751" + } + ], + [ + { + "type": "image_footnote", + "bbox": [ + 0.102, + 0.103, + 0.297, + 0.185 + ], + "angle": 0, + "content": "Create a sleek and dynamic MLB logo poster for the Colorado rockies: Use a bold color palette featuring deep purples, black, and metallic silver. Incorporate iconic mountains and baseball imagery, with prominent use of the text Colorado rockedies: Ensure the text is stylishly integrated and clearly legible. Keep the design clean and simple, focusing on strong, sharp lines that convey professional atheticism and team spirit." + }, + { + "type": "image", + "bbox": [ + 0.313, + 0.089, + 0.605, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.608, + 0.089, + 0.896, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.094, + 0.204, + 0.315, + 0.308 + ], + "angle": 0, + "content": "A beautifully designed poster showcasing Brazilian cultural elements, suitable for a gallery. The composition features vibrant colors such as yellow, green, and blue, representing the Brazilian flag. In the center, there is a cheerful smiley face with a glossy finish. Below the smiley face, in bold and elegant typography, the word 'Obrigado' is prominently displayed. Traditional Brazilian patterns and motifs subtly decorate the background, adding a festive and welcoming atmosphere. The overall design is simple yet striking, perfect for a gallery wrap canvas that conveys gratitude with a touch of Brazilian charm." + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.204, + 0.45, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.315, + 0.403, + 0.326 + ], + "angle": 0, + "content": "Flux" + }, + { + "type": "image", + "bbox": [ + 0.473, + 0.206, + 0.599, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.483, + 0.314, + 0.585, + 0.327 + ], + "angle": 0, + "content": "Type-R w/ Flux" + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.202, + 0.704, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.202, + 0.898, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.315, + 0.803, + 0.326 + ], + "angle": 0, + "content": "TextDiffuser-2" + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.34, + 0.772, + 0.354 + ], + "angle": 0, + "content": "Figure 7. Generated images from long prompts by Flux, Flux w/ Type-R and TextDiffuser-2." + }, + { + "type": "image", + "bbox": [ + 0.105, + 0.381, + 0.476, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.529, + 0.483, + 0.571 + ], + "angle": 0, + "content": "Figure 8. Relationship in the typo correction iteration and OCR* accuracy. Note that OCR* accuracy is computed by OCR models in Type-R which is distinct from the OCR model for evaluation." + }, + { + "type": "table_caption", + "bbox": [ + 0.163, + 0.588, + 0.411, + 0.602 + ], + "angle": 0, + "content": "Table 5. Comparing backend text erasers." + }, + { + "type": "table", + "bbox": [ + 0.177, + 0.612, + 0.399, + 0.694 + ], + "angle": 0, + "content": "
MethodErase allOCR ↑
LaMa [37]✓62.4
63.0
Garnet [17]✓53.4
53.2
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.9 + ], + "angle": 0, + "content": "Text Erasing Methods We present the OCR accuracy when replacing the backend text erasing model in Tab. 5. In this study, we compare LaMa [37], a generic inpainting model, and Garnet, a text-specific inpainting model. Since inpainting models tend to be influenced by surrounding text regions, we optionally erase all text regions to prevent filling in letter-like artifacts and only focus on the background. Table 5 shows that LaMa performs well, and erasing all texts slightly improves the final OCR accuracy. We observe that Garnet tends to leave outlines of texts when erasing text, and the outlines seem to be recognized as words in the OCR evaluation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.38, + 0.628, + 0.395 + ], + "angle": 0, + "content": "5. Limitation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.405, + 0.907, + 0.585 + ], + "angle": 0, + "content": "Type-R depends on the performance of several external models and the base text-to-image models. We observe that OCR models sometimes fail to correctly parse all texts generated by text-to-image models. We also observe that AnyText [39] tends to generate unstable outlines of texts and degrading brightness of pixels from original images, which tends to cause incompatibility in composition. Text-to-image models sometimes fail to render with detailed layout specifications and seem to suffer from uppercase and lowercase distinctions. Although our method is applicable in any combination, we have to improve these backend models to resolve the failure cases." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.587, + 0.909, + 0.724 + ], + "angle": 0, + "content": "There are other minor remaining problems. Type-R decomposes texts into words by the space. There are texts whose reading order is meaningful, but Type-R does not guarantee keeping the reading order of words in the generated image. Type-R has a relatively high computational requirement due to the stacking of multiple backend models in the pipeline, though it might be able to use the same backend in multiple stages to mitigate the issue; e.g., the test editing model can also erase texts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.739, + 0.634, + 0.755 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this paper, we propose an automatic typography correction method named Type-R, which automatically detects and fixes typographic errors in the image generated by text-to-image models. Through extensive experiments, we show that Type-R effectively corrects typos while maintaining the quality of the generated images and outperforms baselines in terms of the quality-accuracy trade-off. Our approach does not require training or fine-tuning and is readily plug-gable to any existing text-to-image model." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2752" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.142 + ], + "angle": 0, + "content": "[1] DALL-E 3. https://openai.com/dall-e-3. Accessed: 2024-03-15. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.145, + 0.482, + 0.171 + ], + "angle": 0, + "content": "[2] DeepFloyd IF. https://github.com/deep-floyd/IF. Accessed: 2024-11-14. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.174, + 0.482, + 0.214 + ], + "angle": 0, + "content": "[3] FLUX.1 [dev]. https://blackforestlabs.ai/announcing-black-forest-labs/. Accessed: 2024-11-14. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.217, + 0.482, + 0.244 + ], + "angle": 0, + "content": "[4] GPT-4o. https://openai.com/index/gpt-4o-system-card/. Accessed: 2024-11-14. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.247, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[5] PaddlePaddle. https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.7/doc/doc_ch/PP-OCRv4_introduction.md.4,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.304, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[6] Stable Diffusion 3.5. https://huggingface.co/stabilityai/stable-diffusion-3.5-large. Accessed: 2024-11-8. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.347, + 0.482, + 0.402 + ], + "angle": 0, + "content": "[7] Jeonghun Baek, Geewook Kim, Junyeop Lee, Sungrae Park, Dongyoon Han, Sangwoo Yun, Seong Joon Oh, and Hwalsuk Lee. What is wrong with scene text recognition model comparisons? dataset and model analysis. In ICCV, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.404, + 0.482, + 0.444 + ], + "angle": 0, + "content": "[8] Youngmin Baek, Bado Lee, Dongyoon Han, Sangdoo Yun, and Hwalsuk Lee. Character region awareness for text detection. In CVPR, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.447, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[9] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser: Diffusion models as text painters. In NeurIPS, 2023. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.49, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[10] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser-2: Unleashing the power of language models for text rendering. In ECCV, 2024. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.548, + 0.482, + 0.615 + ], + "angle": 0, + "content": "[11] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.618, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[12] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. In NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.688, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. arXiv preprint arXiv:1706.08500, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[14] Jiabao Ji, Guanhua Zhang, Zhaowen Wang, Bairu Hou, Zhifei Zhang, Brian Price, and Shiyu Chang. Improving diffusion models for scene text editing with dual encoders. arXiv preprint arXiv:2304.05568, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[15] Peidong Jia, Chenxuan Li, Zeyu Liu, Yichao Shen, Xingru Chen, Yuhui Yuan, Yinglin Zheng, Dong Chen, Ji Li, Xiaodong Xie, et al. Cole: A hierarchical generation framework for graphic design. arXiv preprint arXiv:2311.16974, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[16] Sanyam Lakhanpal, Shivang Chopra, Vinija Jain, Aman Chadha, and Man Luo. Refining text-to-image generation:" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Towards accurate training-free glyph-enhanced image generation. arXiv preprint arXiv:2403.16422, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[17] Hyeonsu Lee and Chankyu Choi. The surprisingly straightforward scene text removal method with gated attention and region of interest generation: A comprehensive prominent model analysis. In ECCV, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[18] Chao Li, Chen Jiang, Xiaolong Liu, Jun Zhao, and Guoxin Wang. Joytype: A robust design for multilingual visual text creation. arXiv preprint arXiv:2409.17524, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.222, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[19] Minghao Li, Tengchao Lv, Jingye Chen, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, and Furu Wei. Trocr: Transformer-based optical character recognition with pre-trained models. In AAAI, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.279, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[20] Wenbo Li, Guohao Li, Zhibin Lan, Xue Xu, Wanru Zhuang, Jiachen Liu, Xinyan Xiao, and Jinsong Su. Empowering backbone models for visual text generation with input granularity control and glyph-aware training. arXiv preprint arXiv:2410.04439, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[21] Zhenhang Li, Yan Shu, Weichao Zeng, Dongbao Yang, and Yu Zhou. First creating backgrounds then rendering texts: A new paradigm for visual text blending. In ECAI, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.392, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[22] Minghui Liao, Guan Pang, Jing Huang, Tal Hassner, and Xiang Bai. Mask TextSpotter v3: Segmentation proposal network for robust scene text spotting. In ECCV, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.435, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[23] Jiawei Lin, Jiaqi Guo, Shizhao Sun, Zijiang Yang, JianGuang Lou, and Dongmei Zhang. Layoutprompter: Awaken the design ability of large language models. In NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.491, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[24] Yiqi Lin, Conghui He, Alex Jinpeng Wang, Bin Wang, Wei-jia Li, and Mike Zheng Shou. Parrot captions teach clip to spot text. In ECCV, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.534, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[25] Rosanne Liu, Dan Garrette, Chitwan Sahara, William Chan, Adam Roberts, Sharan Narang, Irina Blok, RJ Mical, Mohammad Norouzi, and Noah Constant. Character-aware models improve visual text rendering. In ACL, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.905, + 0.643 + ], + "angle": 0, + "content": "[26] Zeyu Liu, Weicong Liang, Zhanhao Liang, Chong Luo, Ji Li, Gao Huang, and Yuhui Yuan. Glyph-byt5: A customized text encoder for accurate visual text rendering. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[27] Zeyu Liu, Weicong Liang, Yiming Zhao, Bohan Chen, Ji Li, and Yuhui Yuan. Glyph-byt5-v2: A strong aesthetic baseline for accurate multilingual visual text rendering. arXiv preprint arXiv:2406.10208, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[28] Jian Ma, Mingjun Zhao, Chen Chen, Ruichen Wang, Di Niu, Haonan Lu, and Xiaodong Lin. Glyphdraw: Seamlessly rendering text with intricate spatial structures in text-to-image generation. arXiv preprint arXiv:2303.17870, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.761, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[29] KhayTze Peong, Seiichi Uchida, and Daichi Haraguchi. Typographic text generation with off-the-shelf diffusion model. In ICDAR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[30] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[31] Yadong Qu, Qingfeng Tan, Hongtao Xie, Jianjun Xu, Yuxin Wang, and Yongdong Zhang. Exploring stroke-level modifications for scene text editing. In AAAI, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2753" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[33] Miao Rang, Zhenni Bi, Chuanjian Liu, Yunhe Wang, and Kai Han. An empirical study of scaling law forOCR. In CVPR, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.482, + 0.249 + ], + "angle": 0, + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.252, + 0.482, + 0.293 + ], + "angle": 0, + "content": "[35] Prasun Roy, Saumik Bhattacharya, Subhankar Ghosh, and Umapada Pal. STEFANN: scene text editor using font adaptive neural network. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.296, + 0.482, + 0.365 + ], + "angle": 0, + "content": "[36] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.368, + 0.482, + 0.436 + ], + "angle": 0, + "content": "[37] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In WACV, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.439, + 0.482, + 0.508 + ], + "angle": 0, + "content": "[38] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[39] Yuxiang Tuo, Wangmeng Xiang, Jun-Yan He, Yifeng Geng, and Xuansong Xie. Anytext: Multilingual visual text generation and editing. In ICLR, 2023. 2, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.555, + 0.482, + 0.596 + ], + "angle": 0, + "content": "[40] Aoqiang Wang, Jian Wang, Zhenyu Yan, Wenxiang Shang, Ran Lin, and Zhao Zhang. Textmaster: Universal controllable text edit. arXiv preprint arXiv:2410.09879, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.599, + 0.482, + 0.639 + ], + "angle": 0, + "content": "[41] Yibin Wang, Weizhong Zhang, Jianwei Zheng, and Cheng Jin. High fidelity scene text synthesis. arXiv preprint arXiv:2405.14701, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.642, + 0.482, + 0.711 + ], + "angle": 0, + "content": "[42] Jules White, Quchen Fu, Sam Hays, Michael Sandborn, Carlos Olea, Henry Gilbert, Ashraf Elnashar, Jesse Spencer-Smith, and Douglas C Schmidt. A prompt pattern catalog to enhance prompt engineering with chatgpt. arXiv preprint arXiv:2302.11382, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.714, + 0.482, + 0.755 + ], + "angle": 0, + "content": "[43] Liang Wu, Chengquan Zhang, Jiaming Liu, Junyu Han, Jingtuo Liu, Errui Ding, and Xiang Bai. Editing text in the wild. In ACM MM, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.758, + 0.482, + 0.798 + ], + "angle": 0, + "content": "[44] Tsung-Han Wu, Long Lian, Joseph E Gonzalez, Boyi Li, and Trevor Darrell. Self-correcting llm-controlled diffusion models. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[45] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and CUI Bin. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. In ICML, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[46] Qiangpeng Yang, Jun Huang, and Wei Lin. Swaptext: Image based texts transfer in scenes. In CVPR, pages 14700-14709, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[47] Yukang Yang, Dongnan Gui, Yuhui Yuan, Weicong Liang, Haisong Ding, Han Hu, and Kai Chen. Glyphcontrol: Glyph conditional control for visual text generation. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.149, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[48] Maoyuan Ye, Jing Zhang, Shanshan Zhao, Juhua Liu, Tongliang Liu, Bo Du, and Dacheng Tao. DeepSolo: Let transformer decoder with explicit points solo for text spotting. In CVPR, 2023. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[49] Maoyuan Ye, Jing Zhang, Juhua Liu, Chenyu Liu, Baocai Yin, Cong Liu, Bo Du, and Dacheng Tao. Hi-sam: Marrying segment anything model for hierarchical text segmentation. arXiv preprint arXiv:2401.17904, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.263, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[50] Weichao Zeng, Yan Shu, Zhenhang Li, Dongbao Yang, and Yu Zhou. TextCtrl: Diffusion-based scene text editing with prior guidance control. In NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.305, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[51] Lingjun Zhang, Xinyuan Chen, Yaohui Wang, Yue Lu, and Yu Qiao. Brush your text: Synthesize any scene text on images via diffusion model. In AAAI, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.349, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[52] Qilong Zhangli, Jindong Jiang, Di Liu, Licheng Yu, Xiaoliang Dai, Ankit Ramchandani, Guan Pang, Dimitris N Metaxas, and Praveen Krishnan. Layout-agnostic scene text image synthesis with diffusion models. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.405, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[53] Yiming Zhao and Zhouhui Lian. Udifftext: A unified framework for high-quality text synthesis in arbitrary images via character-aware diffusion models. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.448, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[54] Zhen Zhao, Jingqun Tang, Binghong Wu, Chunhui Lin, Shu Wei, Hao Liu, Xin Tan, Zhizhong Zhang, Can Huang, and Yuan Xie. Harmonizing visual text comprehension and generation. In NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.504, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[55] Yuanzhi Zhu, Jiawei Liu, Feiyu Gao, Wenyu Liu, Xinggang Wang, Peng Wang, Fei Huang, Cong Yao, and Zhibo Yang. Visual text generation in the wild. In ECCV, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2754" + } + ] +] \ No newline at end of file diff --git a/2025/Type-R_ Automatically Retouching Typos for Text-to-Image Generation/6f176b19-a787-4b85-9ebe-081a3f4759be_origin.pdf b/2025/Type-R_ Automatically Retouching Typos for Text-to-Image Generation/6f176b19-a787-4b85-9ebe-081a3f4759be_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..db88026101e8345775ae2f7885a9bd27cb9d50f5 --- /dev/null +++ b/2025/Type-R_ Automatically Retouching Typos for Text-to-Image Generation/6f176b19-a787-4b85-9ebe-081a3f4759be_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b7f4d13db4b0bce730ed184693f3a83e798d56907759509d15f53c1d6ec3c7e1 +size 3830374 diff --git a/2025/Type-R_ Automatically Retouching Typos for Text-to-Image Generation/full.md b/2025/Type-R_ Automatically Retouching Typos for Text-to-Image Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e91736009b6991dace67f15742f36838d8152132 --- /dev/null +++ b/2025/Type-R_ Automatically Retouching Typos for Text-to-Image Generation/full.md @@ -0,0 +1,385 @@ +# Type-R: Automatically Retouching Typos for Text-to-Image Generation + +Wataru Shimoda1 Naoto Inoue1 Daichi Haraguchi1 Hayato Mitani2 Seiichi Uchida2 Kota Yamaguchi1 CyberAgent Kyushu University + +Create a poster. Display the text 'CVPR' at the top of the image. NASHVILLE TENNESSEE is below the text. The date information 2025 JUNE is required. This image contains the picture of Music City Center at the lower right and cityscapes at the left. + +Prompt + +![](images/b80130e76e0085375c97dfe722e0e026a5e06be172cea49f3f6bb098bd71b2e1.jpg) +Generated image + +![](images/b697a7f71940d60fef9741ae7df78653404ac874a2839a94e77b2c18b2478669.jpg) +Type-R + +![](images/cb2b0080be1dcbfc7beec7ced6256970329e2146cf16f5cf5bd458b2d0f5f293.jpg) +Figure 1. Type-R automatically corrects typographic errors in an image generated from any text-to-image model without sacrificing their visual quality. We highlight prompt words corresponding to the words in the image with the same color. + +![](images/e28eda8aaedb55f708f4fd1f1fbb7ac4a17febe48bb40cb3ab1a00ad382dbd98.jpg) +Typo-corrected image + +# Abstract + +While recent text-to-image models can generate photorealistic images from text prompts that reflect detailed instructions, they still face significant challenges in accurately rendering words in the image. In this paper, we propose to retouch erroneous text renderings in the post-processing pipeline. Our approach, called Type-R, identifies typographical errors in the generated image, erases the erroneous text, regenerates text boxes for missing words, and finally corrects typos in the rendered words. Through extensive experiments, we show that Type-R, in combination with the latest text-to-image models such as Stable Diffusion or Flux, achieves the highest text rendering accuracy while maintaining image quality and also outperforms text-focused generation baselines in terms of balancing text accuracy and image quality. + +# 1. Introduction + +The rapid advancement of text-to-image generation models triggers a surge of application interest in various creative domains. Graphic design is one of the domains from which + +text-to-image models potentially benefit, with the goal of helping generate appealing posters, advertisements, book covers, webpages, or signs, such that these display media convey the intended message to the viewers. However, one of the major obstacles is text rendering on the image [25]. Popular state-of-the-art text-to-image models such as Stable Diffusion [6, 30] or Flux [3] still face notable challenges in accurately rendering words or characters (Fig. 1 second column). Common issues include missing words, unintended characters, or word-level errors that insert another similar-looking word. While it might be possible to tune the text-to-image models with increasing amounts of training data, the resulting image is not guaranteed to be typo-free. + +In this work, instead of modifying the image generator, we propose automatically retouching the erroneous text renderings of the already generated image. Our approach, named Type-R, detects typographic errors in a generated image and automatically fixes erroneous content. Figure 1 shows an example of the input prompt, the generated image by Flux.1-dev [3], detected errors by Type-R, and the final retouched image. Type-R successfully corrects the typographic errors without artifacts compared to the initial generation. Our approach is fully automatic, independent of the base text-to-image generator, and easily pluggable to any existing text-to-image model. + +Through extensive experiments, we study the limitations in the current text-to-image models, including both the popular text-to-image generators [3, 30] and typophy-focussed models [9, 10], and show that Type-R significantly improves the text rendering accuracy of text-to-image models while maintaining graphic design quality. We also study several design choices, e.g., how changing the base text-to-image generator or the OCR model affects the resulting visual quality and text accuracy, and show that Type-R in combination with Flux [3] achieves the best trade-off between text rendering accuracy and quality scores by GPT evaluation on Mario-Eval benchmark [9]. + +We summarize our contributions in the following + +1. We propose Type-R, which automatically retouches typographic errors in generated images by text-to-image models while maintaining the design quality. +2. We empirically show that the combination of the state-of-the-art text-to-image generator and Type-R achieves the best quality-accuracy trade-off and outperforms the approach specifically designed for typography. + +# 2. Related Work + +# 2.1. Text-to-Image Generation + +Text-to-image generation is the task of creating images from text prompts. With the development of large image-capture datasets, there has been significant progress in generating photorealistic images from diffusion-based models [1, 2, 30, 34, 36]. However, rendering accurate text is still one of the remaining problems. Liu et al. [25] point out the challenge in accurately rendering characters in text-to-image generation models, where it is common to observe missing glyphs, redundant glyphs, or wrong characters. The text accuracy problem persists in the recent text-to-image models such as Stable Diffusion 3 [11] or Flux [3]. + +In this work, instead of modifying or tuning the text-to-image models, we utilize feedback from external models to improve text rendering. Several works report the successful use of external models in improving image generation. RPG [45] introduces a technique for region-specific prompt guidance for noise control. Wu et al. [44] apply object detection to generated images and refine detection results using Large Language Models. In our task setup, we take advantage of pre-trained Vision Language Models (VLMs), OCR models, and text-inpainting models to retouch typographic errors. + +# 2.2. Visual Text Generation + +Several text-to-image models [2, 26, 27, 36, 52] have reported that better text encoder significantly improves text rendering capabilities, where the encoder can learn glyph-level information from the training corpus. Adding OCR losses also improve performances [20, 39, 53]. While im + +proving the capability of the base text-to-image generator might eventually reduce typographic errors, we consider a post-processing approach independent of the base model. + +Another line of work to improve text rendering is via conditional inputs to the text-to-image models. Some literature attempts at conditioning generation on text positions and glyph masks to improve the text rendering accuracies [9, 10, 28, 29, 39, 47, 51, 54]. Particularly, TextDiffusers [9, 10] generate images with two steps: layout planning, which generates text position information from prompts, followed by image generation conditioned on the first layout information. TextDiffuser [9] and TextDiffuser 2 [9] differ in how they generate layout information, where the former generates character-by-character masks and the latter generates layout in language. One major limitation of conditional generation is the training requirements; i.e., building a sufficient training corpus for the conditioning module is often difficult. Our experiments suggest that TextDiffusers suffer from inadequate image quality compared to the state-of-the-art image generators, likely due to the bias in its training data. + +We take the approach of retouching afterward using external pre-trained models to take full advantage of the best text-to-image models without fine-tuning. A few works incorporate external models for text rendering. Very recently, Lakhanpal et al. reports Refining T2I [16] that applies OCR to the generated images by TextDiffuser and optimizes the layouts to improve the legibility of texts, while the same limitation to TextDiffusers still applies. Li et al. [21] propose to generate background images by removing generated scene text. Unlike prior works, we propose a method for generating images that adheres as much as possible to the initial generation results, making full use of OCR models, text removal models, and VLMs to improve text rendering. + +# 2.3. Text Editing and Synthesis + +Another relevant task is raster text editing, which directly manipulates texts in the image. While GAN-based methods [31, 35, 43, 46] have been popular, recent visual text generation methods [14, 18, 39-41, 50, 55] are capable of editing texts in the image directly. While text editing and our task setup seem similar, text editing is only a part of the typography retouching task, which involves further tasks such as automatic error detection or missing word insertion. + +# 3. Approach + +Our Type-R pipeline consists of four stages, which are all automatic: 1) error detection, 2) text erasing, 3) layout regeneration, and 4) typo correction. We show the pipeline overview in Fig. 2. + +Generated image +![](images/da7b0999f7424eef3557508a0b85fa2761b2bec67f0e774281a6743e1c8e0c9d.jpg) +Create an advertisement for a hamburger with a pop background. Display the texts '10% OFF and HAMBURGER' with a tasty burger, fried potatoes, and a Coke. The text 'Combo Meals' is required. Prompt + +![](images/df2d3b673ee05efb2be9e89e91b57815dd84fbce33aafb825389787ecc58d188.jpg) + +![](images/db645485cc70c71938434ce1f57df5d4c07561865d3b5a7078aa1920d22c3355.jpg) +Generation +Error detection +Figure 2. Illustration of the Type-R pipeline. Type-R automatically detects errors, erases unintended texts, inserts missing words, and corrects spelling errors in the image. + +![](images/7b27e99f440aaa70e4c90a60f597cdcb1561a5c8849cdc197f3042a276bd3bad.jpg) + +![](images/2b630661ec278ff3b568e862f49b517f601db572ee655793d1a6fc5740dbaaae.jpg) +Text erasing + +![](images/b5d5466011a573b8bf20e6f4edd341467c9aba83d770805ad11488fedbee07c6.jpg) + +![](images/8d12660020e55b3e3718902cc82dc65c113d80da4c0beb9ff56b71ad008b262b.jpg) +Layout regeneration + +![](images/77e45fe2dce800364594956fd18bfe0fe22385c32e059f43e80919285ac57ad6.jpg) + +![](images/99438419cdb13ddb5a10afa1aa71fdb5fe492803bdd645c9b680d00fc3cec239.jpg) +3 Typo correction + +# 3.1. Error Detection + +We denote an image $I$ generated by an off-the-shelf text-to-image model $G$ with a prompt, which requests to show $N$ words $W = \{w_{i}\}_{i = 1}^{N}$ in $I$ . To ensure how much this expectation is satisfied, we first detect words in the image $I$ . Specifically, a scene text detection model predicts polygonal word regions $\hat{B}$ . We remove polygons with a height less than a threshold $\theta$ . Then, a scene text recognition model parses the words $\hat{W} = \{\hat{w}_i\}_{i = 1}^{\hat{N}}$ in the regions $\hat{B}$ . As noted in Section 1, the model $G$ is often not perfect, and $W \neq \hat{W}$ ; therefore, we need to identify errors in $\hat{W}$ by considering $W$ is a ground-truth. + +Roughly speaking, this identification task is a matching problem between $W$ and $\hat{W}$ . More specifically, we consider the matching between the word set $\mathcal{W}$ and $\hat{\mathcal{W}}$ of equal size $\mathcal{N}$ by optimal transport: + +$$ +\mathrm {O T} (\mathcal {W}, \hat {\mathcal {W}}) \equiv \max _ {\pi} \sum_ {i = 1} ^ {\mathcal {N}} \operatorname {L e v} \left(w _ {i}, \hat {w} _ {\pi (i)}\right), \tag {1} +$$ + +where $w_{i}\in \mathcal{W},\hat{w}_{i}\in \hat{\mathcal{W}},\pi$ is a one-by-one matching, and $\mathrm{Levenshtein}$ distance to measure the text similarity. This matching works only when the two sets are of equal size, but typographic errors result in different sizes because of missing or unintended words. Here, we equalize two word sets $W$ and $\hat{W}$ by padding tokens. Let us define $\mathcal{N} = \max (N,\hat{N})$ . We express the padded word sets by: + +$$ +\mathcal {W} = W \cup \left\{p _ {i} \right\} _ {i = N + 1} ^ {\mathcal {N}}, \tag {2} +$$ + +$$ +\hat {\mathcal {W}} = \hat {W} \cup \left\{\hat {p} _ {i} \right\} _ {i = \hat {N} + 1} ^ {\mathcal {N}}, \tag {3} +$$ + +where $p_i$ and $\hat{p}_i$ are the padding tokens. Here, a word matching a padding token indicates a typographic error; i.e., + +a paired word with $p$ is an unintended word, and a paired word with $\hat{p}$ is a missing word. We assign a constant distance to any word matching to the padding tokens. Any pairs having non-zero distance indicate spelling errors. + +The following sections describe our approach to minimizing the above optimal transportation cost. We align the number of words by erasing redundant texts when $N < \hat{N}$ (Sec. 3.2) and insert text boxes by layout regeneration when $N > \hat{N}$ (Sec. 3.3). Finally, we fix spelling errors in typo correction (Sec. 3.4). + +# 3.2. Text Erasing + +Type-R erases unintended texts using an off-the-shelf inpainting method [37] if the number of detected words exceeds the number of the words specified by a prompt, i.e., $N < \hat{N}$ , or the generated image includes filtered texts. This step reduces the number of words in a generated image to equal or less than the number of specified words by the prompt. The inpainting model removes specified paint regions. In our setup, we specify regions that we want to erase in $\hat{B}$ and composite the inpainted results to the original image. Since text detection may not always fully cover text areas, we slightly enlarge the text region masks to ensure complete removal. + +# 3.3. Layout Regeneration + +Layout regeneration fills in missing text regions so that the resulting $\hat{N}$ equals the number of words specified by a prompt $N$ when $\hat{N} < N$ . This is equivalent to obtaining position information from the image for missing words of the prompt, and we address this task using a VLM, specifically GPT-4o [4]. Inspired by works on layout generation by zero/few-shot learning of LLMs [12, 23], we ask GPT to plan layouts for each missing word from the image and layout information of valid text boxes in JSON format. + +For simplicity, we represent the layout information as a bounding box with four coordinates. When text detection outputs are in the polygonal form, we convert them to boxes by finding the smallest four-sided bounding box that encloses each polygon. + +# 3.4. Tipo Correction + +Typo correction tries to resolve all word pairs with non-zero edit distance. To address word-level spelling errors, we employ a pre-trained text editing model [39] to automatically retouch the word region. Using an inpainting approach, the text editing model generates raster images containing specified text in the selected areas. Similar to text-to-image models, text editing models may not always render the specified text accurately but with a much lower computation cost than large text-to-image models. We repeatedly apply the text editing model until we fix the spelling errors or we reach the predefined number of attempts. Note that we only composite the areas of corrected words in the edited image during the iteration. We show the pseudo-codes in Algorithm 1. + +Algorithm 1 A pseudo-code of typo correction. +1: function TyPO CORRECTION(Image I, Words W, Detected regions $\hat{B}$ , Maximum iterations $t_\mathrm{max})$ +2: for $t = 1$ to $t_\mathrm{max}$ do +3: $I_{\mathrm{edit}}\gets$ Text-Editor(I,W, $\hat{B}$ ) +4: $\hat{W}\gets$ Text-Recognition $(I_{\mathrm{edit}},\hat{B})$ +5: $\mathcal{O}\gets \mathrm{Diff}(W,\hat{W})\qquad \triangleright$ Find spelling errors +6: $I\gets$ Composite $(I,I_{\mathrm{edit}},\hat{B},\mathcal{O})$ +7: if $\mathcal{O}\neq \varnothing$ then +8: $W,\hat{B}\gets$ Update(W, $\hat{B},\mathcal{O})$ +9: else +10: break +11: end if +12: end for +13: return I +14: end function + +# 4. Experiment + +# 4.1. Dataset + +We evaluate the performance of visual text generation on MARIO-Eval benchmark [9]. This benchmark contains 5,414 prompts, of which 5,000 have image data. We use all 5,414 prompts in the main comparison, i.e., in Tab. 1. We randomly pick 500 of the 5,414 prompts as the subset for analysis in Sec. 4.6 because our experiments involve paid APIs. For the selection of hyperparameters in filtering, we use 500 prompts randomly selected from the MARIO-LAION dataset [9] as the validation set. Note that the validation does not overlap with the benchmark. + +# 4.2. Metrics + +We evaluate the quality of generated images using a GPT-based evaluation similar to TextDiffuser-2 [10]. In our evaluation, we adopt a rating-based evaluation approach [15] instead of the voting evaluation because GPT-based voting is sensitive to the option order and the number of options. In our settings, GPT assigns scores from 1 to 10 for both graphic design quality and content alignment between generated images and prompts. For text rendering accuracy, we measure OCR accuracy commonly used for assessing text rendering performance used in text-to-image generation [9, 10]. We also measure Fréchet Inception Distance (FID) [13] to quantify distribution differences between generated images and the full 5000 images in the MARIO-Eval benchmark, and CLIP score [32] to compute the similarity between each image and its prompt-based on the CLIP. + +# 4.3. Implementation Details + +We use multiple external models in our Type-R pipeline without fine-tuning. We use Deepsolo [48] for text region detection, Paddle [5] for scene text recognition, LaMa [38] for text erasing, GPT-4o [4] for layout regeneration, and AnyText [39] for text editing. In addition, we use GPT to augment prompts following [42]. When augmenting prompts with GPT, we specify what categories of images to generate and how to draw the text in quotation marks to make it stand out. For all GPT utilization in this paper, i.e., augmentation, layout planning, and evaluation, we use the same version of OpenAI API 2023-05-15. + +When the response of the API is corrupted, we retry requests up to five times and give up. For typo correction iteration, we set the maximum number of correction attempts $t_{\mathrm{max}}$ to 10. We chose the hyperparameter $\theta$ for filtering to $4\%$ of the overall image height through grid search. + +# 4.4. Baselines + +Since Type-R is a post-processing module, we combine Type-R with the two base text-to-image models to build a baseline. In addition, we chose a few different baselines tailored for text rendering. + +Type-R w/ SD3 combines Stable Diffusion 3 [11] and our approach. + +Type-R w/ Flux combines Flux [3] and our approach. + +TextDiffuser [9] is a text-to-image generation method based on layout planning, which generates character-wise masks and generates images using the masks as conditions. + +TextDiffuser-2 [10] is a state-of-the-art text-to-image generation method based on layout planning. Unlike TextDiffuser, TextDiffuser-2 plans layouts via LLM and generates images by embedding the layout information into prompts. + +![](images/994473edf7b9c9b48ce64157a4252de501bb31254371053c3944a85baa6e3932.jpg) +Figure 3. Comparisons of generated images by Flux, Flux w/ Type-R, and TextDiffuser models. The left column shows prompts, and the right columns present generated images by each method. + +# 4.5. Results + +Type-R vs. TextDiffusers Table 1 summarizes evaluation metrics of TextDiffuser models and Type-R. The result suggests that Type-R significantly outperforms the TextDiffuser models in GPT evaluation scores, and also Type-R achieves higher content alignment. The higher GPT scores indicate better aesthetic quality, as also suggested in the qualitative results in Figure 3. While the TextDiffuser models tend to generate images with simple backgrounds and texts (i.e., low quality in terms of graphic designs), Type-R successfully generates images with contextually rich backgrounds and seamlessly integrated text through precise and corrective adjustments. + +In OCR evaluations, SD3 with Type-R yields lower scores than the TextDiffuser models, whereas Flux with Type-R achieves higher accuracy. This result suggests that the enhancement of Type-R is influenced by the capability of the base text-to-image model. + +User Study We also conducted a user study in which participants rated images using the same questions and scoring criteria as the GPT evaluations to validate the reliability of the GPT evaluation. We prepared 100 prompts and generated 100 corresponding images, which resulted in scores + +Table 1. Evaluation metrics for Type-R and TextDiffuser baselines. "Graphic" and "Match" indicate graphic design quality and content alignment. + +
MethodGPT↑OCR ↑FID↓CLIP ↑
GraphicMatch
TD [9]4.677.1655.442.034.4
TD-2 [10]4.977.1756.233.834.7
Type-R
SD3[11]7.308.4648.645.033.0
Flux[3]7.678.5562.043.133.1
+ +from 5 participants for a total of 500 ratings per method. + +Table 2 presents the user study results. Both SD3 and Flux with Type-R have higher scores than the TextDiffuser models, consistent with the GPT evaluation. The differences in scores are small among all baselines, although we observe that all scores of Type-R are statistically significant compared to TextDiffuser baselines. The p-value of the paired t-test with Type-R w/ Flux is $1.55e - 11$ for TextDiffuser and $4.14e - 18$ for TextDiffuser 2 in text-image matching evaluation. The small difference might be due to a lack of confidence in rating design or the ability to perceive finer defects in graphic design. + +Table 2. Userstudy for reliability of GPT evaluation. + +
MethodHuman ↑
GraphicMatch
TextDiffuser [9]5.24 ±2.185.08 ±2.51
TextDiffuser-2 [10]5.30 ±1.605.06 ±2.39
Type-R
SD3[11]5.94 ±1.865.99 ±2.34
Flux[3]6.23 ±1.816.04 ±2.34
+ +![](images/42a168055db44ba50a930569c8f5ba22ac5c5a342d1b3f91eb43fe74b513f216.jpg) + +![](images/b2a4f9edad73f22541b842e6e3fb99d780aa16e556cba528984318ccb6759f48.jpg) + +![](images/2617594a8f8c1b70525e15f7ff3a8f3fe5dbf0feb1e0fd9254017515ddaa0191.jpg) +B&W + +![](images/59c4635c3f14832267ec908a5db311575123405d6357b0e75b409e77b3faec1d.jpg) +w/ white box +Figure 4. Examples of naive baselines. + +FID and CLIP scores Although TextDiffuser yields higher FID and CLIP scores, we argue that these scores may not accurately capture human perception of image quality. For FID, the reference dataset is limited to 5,000 images that are not always consistent in quality, which likely causes the FID scores to overfit the narrow reference corpus, making it unsuitable for evaluating graphic design quality. + +TextDiffuser also yields higher CLIP scores. We suspect this result is due to the text-spotting bias of the CLIP model, which tends to "parrot" the visual text embedding in images [24]. That is, the CLIP score tends to prefer words over visual objects in the image, leading to a higher score for the word-only image without a meaningful background visual, which TextDiffusers tend to generate. + +# 4.6. Additional Study + +Quality-accuracy Trade-off Here, we compare Type-R with several baselines in terms of the trade-off between image quality and OCR accuracy. In this study, we introduce the following backend text-to-image models. + +Type-R w/Dall-E 3 combines Dall-E 3 [1] with Type-R. + +Type-R w/ SD 3.5 combines Stable Diffusion 3.5 [6] with Type-R. + +We compare the raw output (Raw) from the backend text-to-image models and the output with Type-R. We also consider the following naive baselines (Fig. 4). We heuristically arrange text boxes for all the naive baselines, ignoring + +![](images/cdcc5f6e3ac771a26980494add07b8371df32e6c58528a0d5d93c4ad56a0b36b.jpg) +Figure 5. Plot of the relation between OCR accuracy and graphic design quality by GPT. Raw represents the results of text-to-image generation with prompt augmentation. + +the image content. + +Simple text erases all texts in the generated image from the backend text-to-image models and uses it as a background. + +Simple text w/ white box also removes all texts and white out the text box regions. + +B&W always renders text on a white background. + +Fig. 5 plots the trade-off between GPT score and OCR accuracy, which shows that naive baselines can achieve high OCR accuracy for much lower GPT score. On the other hand, Flux and Dall-E 3 with Type-R show much higher graphic design quality with a slightly degrading OCR accuracy, almost reaching the accuracy of the simple text rendering. Type-R exhibits robustness across multiple backend text-to-image models, as demonstrated in the plot where Type-R effectively improves OCR accuracy for any backend models while maintaining high graphic design quality. + +Qualitative Results Figure 6 shows a few generated images by Type-R demonstrating how each external model improves text rendering accuracy. In the first row, the text-editing model accurately modifies the text using a fancy font that aligns with the surrounding context. The second row presents an instance of Type-R successfully removing unnecessary texts. The third row provides an example of removing small text and adding a missing text through layout planning, where the inserted new text implicitly utilizes surrounding font styles to achieve a seamless and natural + +![](images/097d5dc41a14f77cf80af4a6599ef89c9578f88919fc24c3b72a9733fcdfcb58.jpg) +Prompt: 'Nobliaux' et sorcires + +![](images/03527dabf3d3dba81a3cb9fd617edb84ed345a4e6e14db1db2aeaf8d1b46db52.jpg) + +![](images/02f0db7562851db71bdda99c7e96ab3afad03c2f9651eb7287b9761c7f11c645.jpg) + +![](images/6a0f62b2de2cdc51d5a35a7b5e58471cff8416acd30f45021839cd34dde9dc1f.jpg) +Prompt: Image of MINI 'GPS' MAGNETIC CAR TRACKER +Prompt: Title 'Elise Chain 'Reaction' Large Paperback + +![](images/f359ce52b84a2a0958efc3e8ea14646469578017679795e8617532dd2654d9a3.jpg) + +![](images/4981f8ad047343329bd9a243fdc5c2a678be5cfeb3c1cf2e5963eafbe696588c.jpg) + +![](images/49c016edf16e509923513e0cdaff0eede1b5fbeb649ec02c1f24c1e403dc4fa7.jpg) +Text-to-Image +Figure 6. Examples of generated images through Type-R. + +![](images/88e31cc8199f472b5ba913e30283d1d113c6ec20a30ebc7f3ce3f9d9c579f722.jpg) +Text erasing $^+$ Layout regeneration + +![](images/2b3d2cd5972aead784b7eb070b03c3099aeacb04bd185fc1f48dc689194b06a7.jpg) +Typo Corrected image + +appearance. + +Since Type-R is independent of the backend text-to-image models, we can take advantage of the state-of-the-art models. Figure 7 presents examples with very long prompts, showing that the latest text-to-image model, Flux, accurately captures detailed information from these prompts, whereas TextDiffuser-2 generates corrupted images when handling such prompts, as CLIP is unable to process tokens exceeding 77. + +Longer Iterations The final typo correction step iteratively refines misspelled words. We show the improvement over iterations in Fig. 8 for several backend models. Our typo correction approach consistently improves OCR accuracy as we repeat the iterations. Our approach achieves $80\%$ of the total improvement within four iterations independent of the backend model. + +# 4.7. Ablation Study + +Type-R Pipeline Modules Tab. 3 summarizes the ablation results for the modules in Type-R pipeline: augmentation for prompt, layout correction (text erasing and layout regeneration), and typo correction. We observe that each module effectively improves OCR accuracy with minimal negative impact on graphic scores by GPT. Layout correc + +Table 3. Ablation study for Type-R pipeline modules. + +
T2IAugCorrectionGraphic ↑OCR ↑
LayoutTypo
SD3 [11]✓6.778.6
7.5512.2
✓✓7.4227.4
✓✓7.5320.4
✓✓✓7.2750.0
Flux [3]✓6.7914.4
7.8028.0
✓✓7.8353.8
✓✓7.7329.4
✓✓✓7.6563.0
+ +Table 4. Comparing backend OCR models. + +
DetectionRecognitionOCR ↑
Paddle [5]Paddle [5]57.4
CRAFT [8]Paddle [5]47.2
Hi-SAM [49]Paddle [5]60.8
Mask TextSpotter v3 [22]Paddle [5]61.0
DeepSolo [48]Paddle [5]63.0
DeepSolo [48]Baek et al. [7]62.0
DeepSolo [48]TrOCR [19]60.0
DeepSolo [48]CLIP4STR-B* [33]60.8
+ +tion significantly improves Flux's OCR accuracy and further boosts OCR performance in combination with typo correction. The comparison between Flux and SD3 suggests Flux exhibits lower rates of glyph corruption. + +OCR Models We summarize the effect of replacing the backend OCR models in Tab. 4. Both detection and recognition models impact OCR accuracy, with detection model performance changes having a more substantial effect than those of the recognition model. We observe that CRAFT [8] frequently merges words separated by spaces into a unit, leading to a notable performance drop, as Type-R assumes that spaces are excluded from texts and does not account for this issue. DeepSolo [48] achieves the highest performance. + +For text recognition, improvements in OCR accuracy do not always align with scene text recognition performance. While CLIP4STR-B* [33] is one of the state-of-the-art scene text recognition models, the OCR accuracy of CLIP4STR-B* is lower than the OCR accuracy of Baek et al. [7] and Paddle [5]. Paddle achieves the highest performance, likely because AnyText [39], trained with a loss function that incorporates the Paddle model, tends to generate words that are more easily recognized by Paddle. This suggests that compatibility between the text editing and recognition models is essential for effective correction. + +![](images/96023c0f29d2b4c6c231d855e275d63ac1e3c2c0fe51e7cd4e87f7e74c249053.jpg) +Create a sleek and dynamic MLB logo poster for the Colorado rockies: Use a bold color palette featuring deep purples, black, and metallic silver. Incorporate iconic mountains and baseball imagery, with prominent use of the text Colorado rockedies: Ensure the text is stylishly integrated and clearly legible. Keep the design clean and simple, focusing on strong, sharp lines that convey professional atheticism and team spirit. +A beautifully designed poster showcasing Brazilian cultural elements, suitable for a gallery. The composition features vibrant colors such as yellow, green, and blue, representing the Brazilian flag. In the center, there is a cheerful smiley face with a glossy finish. Below the smiley face, in bold and elegant typography, the word 'Obrigado' is prominently displayed. Traditional Brazilian patterns and motifs subtly decorate the background, adding a festive and welcoming atmosphere. The overall design is simple yet striking, perfect for a gallery wrap canvas that conveys gratitude with a touch of Brazilian charm. + +![](images/3ad0a4f8f8c705f8fa7b6e5b421ab8398761fb553c2704e7af18fdea4392a275.jpg) + +![](images/bf9545ff7b8359d3b1e1cabf533767cd1c2c77a334f1d612c1073c58e1e126c7.jpg) +Flux + +![](images/73d8fd7669ab051f04d2aa858cb42b69067fbb7701ca1952ed74ebd97714cc79.jpg) +Type-R w/ Flux + +![](images/5643d23e340a61efa2829c5f47a161b838326851e0a4002830615ce39303bc0a.jpg) + +![](images/108f755b3de14df4ce443cfae66a02a9c04b2c42e488083e1caae1a0a4def7e1.jpg) +TextDiffuser-2 +Figure 7. Generated images from long prompts by Flux, Flux w/ Type-R and TextDiffuser-2. + +![](images/4ff027d1d2535b97d7ef0aa6403950d3cacbe4eb174784bc1dd3c9393a411b12.jpg) +Figure 8. Relationship in the typo correction iteration and OCR* accuracy. Note that OCR* accuracy is computed by OCR models in Type-R which is distinct from the OCR model for evaluation. + +Table 5. Comparing backend text erasers. + +
MethodErase allOCR ↑
LaMa [37]✓62.4
63.0
Garnet [17]✓53.4
53.2
+ +Text Erasing Methods We present the OCR accuracy when replacing the backend text erasing model in Tab. 5. In this study, we compare LaMa [37], a generic inpainting model, and Garnet, a text-specific inpainting model. Since inpainting models tend to be influenced by surrounding text regions, we optionally erase all text regions to prevent filling in letter-like artifacts and only focus on the background. Table 5 shows that LaMa performs well, and erasing all texts slightly improves the final OCR accuracy. We observe that Garnet tends to leave outlines of texts when erasing text, and the outlines seem to be recognized as words in the OCR evaluation. + +# 5. Limitation + +Type-R depends on the performance of several external models and the base text-to-image models. We observe that OCR models sometimes fail to correctly parse all texts generated by text-to-image models. We also observe that AnyText [39] tends to generate unstable outlines of texts and degrading brightness of pixels from original images, which tends to cause incompatibility in composition. Text-to-image models sometimes fail to render with detailed layout specifications and seem to suffer from uppercase and lowercase distinctions. Although our method is applicable in any combination, we have to improve these backend models to resolve the failure cases. + +There are other minor remaining problems. Type-R decomposes texts into words by the space. There are texts whose reading order is meaningful, but Type-R does not guarantee keeping the reading order of words in the generated image. Type-R has a relatively high computational requirement due to the stacking of multiple backend models in the pipeline, though it might be able to use the same backend in multiple stages to mitigate the issue; e.g., the test editing model can also erase texts. + +# 6. Conclusion + +In this paper, we propose an automatic typography correction method named Type-R, which automatically detects and fixes typographic errors in the image generated by text-to-image models. Through extensive experiments, we show that Type-R effectively corrects typos while maintaining the quality of the generated images and outperforms baselines in terms of the quality-accuracy trade-off. Our approach does not require training or fine-tuning and is readily plug-gable to any existing text-to-image model. + +# References + +[1] DALL-E 3. https://openai.com/dall-e-3. Accessed: 2024-03-15. 2, 6 +[2] DeepFloyd IF. https://github.com/deep-floyd/IF. Accessed: 2024-11-14. 2 +[3] FLUX.1 [dev]. https://blackforestlabs.ai/announcing-black-forest-labs/. Accessed: 2024-11-14. 1, 2, 4, 5, 6, 7 +[4] GPT-4o. https://openai.com/index/gpt-4o-system-card/. Accessed: 2024-11-14. 3, 4 +[5] PaddlePaddle. https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.7/doc/doc_ch/PP-OCRv4_introduction.md.4,7 +[6] Stable Diffusion 3.5. https://huggingface.co/stabilityai/stable-diffusion-3.5-large. Accessed: 2024-11-8. 1, 6 +[7] Jeonghun Baek, Geewook Kim, Junyeop Lee, Sungrae Park, Dongyoon Han, Sangwoo Yun, Seong Joon Oh, and Hwalsuk Lee. What is wrong with scene text recognition model comparisons? dataset and model analysis. In ICCV, 2019. 7 +[8] Youngmin Baek, Bado Lee, Dongyoon Han, Sangdoo Yun, and Hwalsuk Lee. Character region awareness for text detection. In CVPR, 2019. 7 +[9] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser: Diffusion models as text painters. In NeurIPS, 2023. 2, 4, 5, 6 +[10] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser-2: Unleashing the power of language models for text rendering. In ECCV, 2024. 2, 4, 5, 6 +[11] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 2, 4, 5, 6, 7 +[12] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. In NeurIPS, 2023. 3 +[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. arXiv preprint arXiv:1706.08500, 2017. 4 +[14] Jiabao Ji, Guanhua Zhang, Zhaowen Wang, Bairu Hou, Zhifei Zhang, Brian Price, and Shiyu Chang. Improving diffusion models for scene text editing with dual encoders. arXiv preprint arXiv:2304.05568, 2023. 2 +[15] Peidong Jia, Chenxuan Li, Zeyu Liu, Yichao Shen, Xingru Chen, Yuhui Yuan, Yinglin Zheng, Dong Chen, Ji Li, Xiaodong Xie, et al. Cole: A hierarchical generation framework for graphic design. arXiv preprint arXiv:2311.16974, 2023. 4 +[16] Sanyam Lakhanpal, Shivang Chopra, Vinija Jain, Aman Chadha, and Man Luo. Refining text-to-image generation: + +Towards accurate training-free glyph-enhanced image generation. arXiv preprint arXiv:2403.16422, 2024. 2 +[17] Hyeonsu Lee and Chankyu Choi. The surprisingly straightforward scene text removal method with gated attention and region of interest generation: A comprehensive prominent model analysis. In ECCV, 2022. 8 +[18] Chao Li, Chen Jiang, Xiaolong Liu, Jun Zhao, and Guoxin Wang. Joytype: A robust design for multilingual visual text creation. arXiv preprint arXiv:2409.17524, 2024. 2 +[19] Minghao Li, Tengchao Lv, Jingye Chen, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, and Furu Wei. Trocr: Transformer-based optical character recognition with pre-trained models. In AAAI, 2023. 7 +[20] Wenbo Li, Guohao Li, Zhibin Lan, Xue Xu, Wanru Zhuang, Jiachen Liu, Xinyan Xiao, and Jinsong Su. Empowering backbone models for visual text generation with input granularity control and glyph-aware training. arXiv preprint arXiv:2410.04439, 2024. 2 +[21] Zhenhang Li, Yan Shu, Weichao Zeng, Dongbao Yang, and Yu Zhou. First creating backgrounds then rendering texts: A new paradigm for visual text blending. In ECAI, 2024. 2 +[22] Minghui Liao, Guan Pang, Jing Huang, Tal Hassner, and Xiang Bai. Mask TextSpotter v3: Segmentation proposal network for robust scene text spotting. In ECCV, 2020. 7 +[23] Jiawei Lin, Jiaqi Guo, Shizhao Sun, Zijiang Yang, JianGuang Lou, and Dongmei Zhang. Layoutprompter: Awaken the design ability of large language models. In NeurIPS, 2023. 3 +[24] Yiqi Lin, Conghui He, Alex Jinpeng Wang, Bin Wang, Wei-jia Li, and Mike Zheng Shou. Parrot captions teach clip to spot text. In ECCV, 2024. 6 +[25] Rosanne Liu, Dan Garrette, Chitwan Sahara, William Chan, Adam Roberts, Sharan Narang, Irina Blok, RJ Mical, Mohammad Norouzi, and Noah Constant. Character-aware models improve visual text rendering. In ACL, 2023. 1, 2 +[26] Zeyu Liu, Weicong Liang, Zhanhao Liang, Chong Luo, Ji Li, Gao Huang, and Yuhui Yuan. Glyph-byt5: A customized text encoder for accurate visual text rendering. In ECCV, 2024. 2 +[27] Zeyu Liu, Weicong Liang, Yiming Zhao, Bohan Chen, Ji Li, and Yuhui Yuan. Glyph-byt5-v2: A strong aesthetic baseline for accurate multilingual visual text rendering. arXiv preprint arXiv:2406.10208, 2024. 2 +[28] Jian Ma, Mingjun Zhao, Chen Chen, Ruichen Wang, Di Niu, Haonan Lu, and Xiaodong Lin. Glyphdraw: Seamlessly rendering text with intricate spatial structures in text-to-image generation. arXiv preprint arXiv:2303.17870, 2023. 2 +[29] KhayTze Peong, Seiichi Uchida, and Daichi Haraguchi. Typographic text generation with off-the-shelf diffusion model. In ICDAR, 2024. 2 +[30] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2023. 1, 2 +[31] Yadong Qu, Qingfeng Tan, Hongtao Xie, Jianjun Xu, Yuxin Wang, and Yongdong Zhang. Exploring stroke-level modifications for scene text editing. In AAAI, 2023. 2 + +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 4 +[33] Miao Rang, Zhenni Bi, Chuanjian Liu, Yunhe Wang, and Kai Han. An empirical study of scaling law forOCR. In CVPR, 2024. 7 +[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2 +[35] Prasun Roy, Saumik Bhattacharya, Subhankar Ghosh, and Umapada Pal. STEFANN: scene text editor using font adaptive neural network. In CVPR, 2020. 2 +[36] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2 +[37] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In WACV, 2022. 3, 8 +[38] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 4 +[39] Yuxiang Tuo, Wangmeng Xiang, Jun-Yan He, Yifeng Geng, and Xuansong Xie. Anytext: Multilingual visual text generation and editing. In ICLR, 2023. 2, 4, 7, 8 +[40] Aoqiang Wang, Jian Wang, Zhenyu Yan, Wenxiang Shang, Ran Lin, and Zhao Zhang. Textmaster: Universal controllable text edit. arXiv preprint arXiv:2410.09879, 2024. +[41] Yibin Wang, Weizhong Zhang, Jianwei Zheng, and Cheng Jin. High fidelity scene text synthesis. arXiv preprint arXiv:2405.14701, 2024. 2 +[42] Jules White, Quchen Fu, Sam Hays, Michael Sandborn, Carlos Olea, Henry Gilbert, Ashraf Elnashar, Jesse Spencer-Smith, and Douglas C Schmidt. A prompt pattern catalog to enhance prompt engineering with chatgpt. arXiv preprint arXiv:2302.11382, 2023. 4 +[43] Liang Wu, Chengquan Zhang, Jiaming Liu, Junyu Han, Jingtuo Liu, Errui Ding, and Xiang Bai. Editing text in the wild. In ACM MM, 2018. 2 +[44] Tsung-Han Wu, Long Lian, Joseph E Gonzalez, Boyi Li, and Trevor Darrell. Self-correcting llm-controlled diffusion models. In CVPR, 2024. 2 +[45] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and CUI Bin. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. In ICML, 2024. 2 +[46] Qiangpeng Yang, Jun Huang, and Wei Lin. Swaptext: Image based texts transfer in scenes. In CVPR, pages 14700-14709, 2020. 2 + +[47] Yukang Yang, Dongnan Gui, Yuhui Yuan, Weicong Liang, Haisong Ding, Han Hu, and Kai Chen. Glyphcontrol: Glyph conditional control for visual text generation. In NeurIPS, 2023. 2 +[48] Maoyuan Ye, Jing Zhang, Shanshan Zhao, Juhua Liu, Tongliang Liu, Bo Du, and Dacheng Tao. DeepSolo: Let transformer decoder with explicit points solo for text spotting. In CVPR, 2023. 4, 7 +[49] Maoyuan Ye, Jing Zhang, Juhua Liu, Chenyu Liu, Baocai Yin, Cong Liu, Bo Du, and Dacheng Tao. Hi-sam: Marrying segment anything model for hierarchical text segmentation. arXiv preprint arXiv:2401.17904, 2024. 7 +[50] Weichao Zeng, Yan Shu, Zhenhang Li, Dongbao Yang, and Yu Zhou. TextCtrl: Diffusion-based scene text editing with prior guidance control. In NeurIPS, 2024. 2 +[51] Lingjun Zhang, Xinyuan Chen, Yaohui Wang, Yue Lu, and Yu Qiao. Brush your text: Synthesize any scene text on images via diffusion model. In AAAI, 2024. 2 +[52] Qilong Zhangli, Jindong Jiang, Di Liu, Licheng Yu, Xiaoliang Dai, Ankit Ramchandani, Guan Pang, Dimitris N Metaxas, and Praveen Krishnan. Layout-agnostic scene text image synthesis with diffusion models. In CVPR, 2024. 2 +[53] Yiming Zhao and Zhouhui Lian. Udifftext: A unified framework for high-quality text synthesis in arbitrary images via character-aware diffusion models. In ECCV, 2024. 2 +[54] Zhen Zhao, Jingqun Tang, Binghong Wu, Chunhui Lin, Shu Wei, Hao Liu, Xin Tan, Zhizhong Zhang, Can Huang, and Yuan Xie. Harmonizing visual text comprehension and generation. In NeurIPS, 2024. 2 +[55] Yuanzhi Zhu, Jiawei Liu, Feiyu Gao, Wenyu Liu, Xinggang Wang, Peng Wang, Fei Huang, Cong Yao, and Zhibo Yang. Visual text generation in the wild. In ECCV, 2024. 2 \ No newline at end of file diff --git a/2025/Type-R_ Automatically Retouching Typos for Text-to-Image Generation/images.zip b/2025/Type-R_ Automatically Retouching Typos for Text-to-Image Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..79a3ed01833b79b45983806b499fb0eab8bc760f --- /dev/null +++ b/2025/Type-R_ Automatically Retouching Typos for Text-to-Image Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:677efb321768aed44963ebf7c8c2b3652375c6f78183bfed1680691f166c9c91 +size 801725 diff --git a/2025/Type-R_ Automatically Retouching Typos for Text-to-Image Generation/layout.json b/2025/Type-R_ Automatically Retouching Typos for Text-to-Image Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..22038de7e6328c8396452717366dae090dcdead4 --- /dev/null +++ b/2025/Type-R_ Automatically Retouching Typos for Text-to-Image Generation/layout.json @@ -0,0 +1,9704 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 103, + 526, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 103, + 526, + 121 + ], + "spans": [ + { + "bbox": [ + 85, + 103, + 526, + 121 + ], + "type": "text", + "content": "Type-R: Automatically Retouching Typos for Text-to-Image Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 170, + 142, + 440, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 142, + 440, + 186 + ], + "spans": [ + { + "bbox": [ + 170, + 142, + 440, + 186 + ], + "type": "text", + "content": "Wataru Shimoda1 Naoto Inoue1 Daichi Haraguchi1 Hayato Mitani2 Seiichi Uchida2 Kota Yamaguchi1 CyberAgent Kyushu University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 245, + 175, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 245, + 175, + 325 + ], + "spans": [ + { + "bbox": [ + 59, + 245, + 175, + 325 + ], + "type": "text", + "content": "Create a poster. Display the text 'CVPR' at the top of the image. NASHVILLE TENNESSEE is below the text. The date information 2025 JUNE is required. This image contains the picture of Music City Center at the lower right and cityscapes at the left." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 101, + 346, + 134, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 346, + 134, + 357 + ], + "spans": [ + { + "bbox": [ + 101, + 346, + 134, + 357 + ], + "type": "text", + "content": "Prompt" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 178, + 219, + 301, + 336 + ], + "blocks": [ + { + "bbox": [ + 178, + 219, + 301, + 336 + ], + "lines": [ + { + "bbox": [ + 178, + 219, + 301, + 336 + ], + "spans": [ + { + "bbox": [ + 178, + 219, + 301, + 336 + ], + "type": "image", + "image_path": "b80130e76e0085375c97dfe722e0e026a5e06be172cea49f3f6bb098bd71b2e1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 341, + 261, + 361 + ], + "lines": [ + { + "bbox": [ + 217, + 341, + 261, + 361 + ], + "spans": [ + { + "bbox": [ + 217, + 341, + 261, + 361 + ], + "type": "text", + "content": "Generated image" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 302, + 213, + 425, + 335 + ], + "blocks": [ + { + "bbox": [ + 302, + 213, + 425, + 335 + ], + "lines": [ + { + "bbox": [ + 302, + 213, + 425, + 335 + ], + "spans": [ + { + "bbox": [ + 302, + 213, + 425, + 335 + ], + "type": "image", + "image_path": "b697a7f71940d60fef9741ae7df78653404ac874a2839a94e77b2c18b2478669.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 319, + 346, + 349, + 357 + ], + "lines": [ + { + "bbox": [ + 319, + 346, + 349, + 357 + ], + "spans": [ + { + "bbox": [ + 319, + 346, + 349, + 357 + ], + "type": "text", + "content": "Type-R" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 356, + 338, + 399, + 367 + ], + "blocks": [ + { + "bbox": [ + 356, + 338, + 399, + 367 + ], + "lines": [ + { + "bbox": [ + 356, + 338, + 399, + 367 + ], + "spans": [ + { + "bbox": [ + 356, + 338, + 399, + 367 + ], + "type": "image", + "image_path": "cb2b0080be1dcbfc7beec7ced6256970329e2146cf16f5cf5bd458b2d0f5f293.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 380, + 555, + 403 + ], + "lines": [ + { + "bbox": [ + 55, + 380, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 555, + 403 + ], + "type": "text", + "content": "Figure 1. Type-R automatically corrects typographic errors in an image generated from any text-to-image model without sacrificing their visual quality. We highlight prompt words corresponding to the words in the image with the same color." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 426, + 219, + 549, + 336 + ], + "blocks": [ + { + "bbox": [ + 426, + 219, + 549, + 336 + ], + "lines": [ + { + "bbox": [ + 426, + 219, + 549, + 336 + ], + "spans": [ + { + "bbox": [ + 426, + 219, + 549, + 336 + ], + "type": "image", + "image_path": "e28eda8aaedb55f708f4fd1f1fbb7ac4a17febe48bb40cb3ab1a00ad382dbd98.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 341, + 519, + 361 + ], + "lines": [ + { + "bbox": [ + 456, + 341, + 519, + 361 + ], + "spans": [ + { + "bbox": [ + 456, + 341, + 519, + 361 + ], + "type": "text", + "content": "Typo-corrected image" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 151, + 413, + 200, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 413, + 200, + 426 + ], + "spans": [ + { + "bbox": [ + 151, + 413, + 200, + 426 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 437, + 297, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 437, + 297, + 619 + ], + "spans": [ + { + "bbox": [ + 53, + 437, + 297, + 619 + ], + "type": "text", + "content": "While recent text-to-image models can generate photorealistic images from text prompts that reflect detailed instructions, they still face significant challenges in accurately rendering words in the image. In this paper, we propose to retouch erroneous text renderings in the post-processing pipeline. Our approach, called Type-R, identifies typographical errors in the generated image, erases the erroneous text, regenerates text boxes for missing words, and finally corrects typos in the rendered words. Through extensive experiments, we show that Type-R, in combination with the latest text-to-image models such as Stable Diffusion or Flux, achieves the highest text rendering accuracy while maintaining image quality and also outperforms text-focused generation baselines in terms of balancing text accuracy and image quality." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 639, + 135, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 135, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 135, + 651 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "type": "text", + "content": "The rapid advancement of text-to-image generation models triggers a surge of application interest in various creative domains. Graphic design is one of the domains from which" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 415, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 570 + ], + "type": "text", + "content": "text-to-image models potentially benefit, with the goal of helping generate appealing posters, advertisements, book covers, webpages, or signs, such that these display media convey the intended message to the viewers. However, one of the major obstacles is text rendering on the image [25]. Popular state-of-the-art text-to-image models such as Stable Diffusion [6, 30] or Flux [3] still face notable challenges in accurately rendering words or characters (Fig. 1 second column). Common issues include missing words, unintended characters, or word-level errors that insert another similar-looking word. While it might be possible to tune the text-to-image models with increasing amounts of training data, the resulting image is not guaranteed to be typo-free." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "In this work, instead of modifying the image generator, we propose automatically retouching the erroneous text renderings of the already generated image. Our approach, named Type-R, detects typographic errors in a generated image and automatically fixes erroneous content. Figure 1 shows an example of the input prompt, the generated image by Flux.1-dev [3], detected errors by Type-R, and the final retouched image. Type-R successfully corrects the typographic errors without artifacts compared to the initial generation. Our approach is fully automatic, independent of the base text-to-image generator, and easily pluggable to any existing text-to-image model." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 189, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 189, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 189, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 702, + 189, + 713 + ], + "type": "text", + "content": "Code and project page are available." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2745" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": "Through extensive experiments, we study the limitations in the current text-to-image models, including both the popular text-to-image generators [3, 30] and typophy-focussed models [9, 10], and show that Type-R significantly improves the text rendering accuracy of text-to-image models while maintaining graphic design quality. We also study several design choices, e.g., how changing the base text-to-image generator or the OCR model affects the resulting visual quality and text accuracy, and show that Type-R in combination with Flux [3] achieves the best trade-off between text rendering accuracy and quality scores by GPT evaluation on Mario-Eval benchmark [9]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 217, + 267, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 217, + 267, + 228 + ], + "spans": [ + { + "bbox": [ + 67, + 217, + 267, + 228 + ], + "type": "text", + "content": "We summarize our contributions in the following" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 228, + 295, + 313 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 56, + 228, + 295, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 295, + 265 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 295, + 265 + ], + "type": "text", + "content": "1. We propose Type-R, which automatically retouches typographic errors in generated images by text-to-image models while maintaining the design quality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 265, + 295, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 295, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 295, + 313 + ], + "type": "text", + "content": "2. We empirically show that the combination of the state-of-the-art text-to-image generator and Type-R achieves the best quality-accuracy trade-off and outperforms the approach specifically designed for typography." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 324, + 142, + 336 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 142, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 142, + 336 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 344, + 200, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 344, + 200, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 200, + 357 + ], + "type": "text", + "content": "2.1. Text-to-Image Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 363, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 295, + 495 + ], + "type": "text", + "content": "Text-to-image generation is the task of creating images from text prompts. With the development of large image-capture datasets, there has been significant progress in generating photorealistic images from diffusion-based models [1, 2, 30, 34, 36]. However, rendering accurate text is still one of the remaining problems. Liu et al. [25] point out the challenge in accurately rendering characters in text-to-image generation models, where it is common to observe missing glyphs, redundant glyphs, or wrong characters. The text accuracy problem persists in the recent text-to-image models such as Stable Diffusion 3 [11] or Flux [3]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 495, + 295, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 626 + ], + "type": "text", + "content": "In this work, instead of modifying or tuning the text-to-image models, we utilize feedback from external models to improve text rendering. Several works report the successful use of external models in improving image generation. RPG [45] introduces a technique for region-specific prompt guidance for noise control. Wu et al. [44] apply object detection to generated images and refine detection results using Large Language Models. In our task setup, we take advantage of pre-trained Vision Language Models (VLMs), OCR models, and text-inpainting models to retouch typographic errors." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 635, + 186, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 186, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 186, + 647 + ], + "type": "text", + "content": "2.2. Visual Text Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": "Several text-to-image models [2, 26, 27, 36, 52] have reported that better text encoder significantly improves text rendering capabilities, where the encoder can learn glyph-level information from the training corpus. Adding OCR losses also improve performances [20, 39, 53]. While im" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "content": "proving the capability of the base text-to-image generator might eventually reduce typographic errors, we consider a post-processing approach independent of the base model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 111, + 555, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 111, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 111, + 555, + 327 + ], + "type": "text", + "content": "Another line of work to improve text rendering is via conditional inputs to the text-to-image models. Some literature attempts at conditioning generation on text positions and glyph masks to improve the text rendering accuracies [9, 10, 28, 29, 39, 47, 51, 54]. Particularly, TextDiffusers [9, 10] generate images with two steps: layout planning, which generates text position information from prompts, followed by image generation conditioned on the first layout information. TextDiffuser [9] and TextDiffuser 2 [9] differ in how they generate layout information, where the former generates character-by-character masks and the latter generates layout in language. One major limitation of conditional generation is the training requirements; i.e., building a sufficient training corpus for the conditioning module is often difficult. Our experiments suggest that TextDiffusers suffer from inadequate image quality compared to the state-of-the-art image generators, likely due to the bias in its training data." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 331, + 556, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 556, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 556, + 486 + ], + "type": "text", + "content": "We take the approach of retouching afterward using external pre-trained models to take full advantage of the best text-to-image models without fine-tuning. A few works incorporate external models for text rendering. Very recently, Lakhanpal et al. reports Refining T2I [16] that applies OCR to the generated images by TextDiffuser and optimizes the layouts to improve the legibility of texts, while the same limitation to TextDiffusers still applies. Li et al. [21] propose to generate background images by removing generated scene text. Unlike prior works, we propose a method for generating images that adheres as much as possible to the initial generation results, making full use of OCR models, text removal models, and VLMs to improve text rendering." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 503, + 462, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 462, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 462, + 516 + ], + "type": "text", + "content": "2.3. Text Editing and Synthesis" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 525, + 555, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 555, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 555, + 621 + ], + "type": "text", + "content": "Another relevant task is raster text editing, which directly manipulates texts in the image. While GAN-based methods [31, 35, 43, 46] have been popular, recent visual text generation methods [14, 18, 39-41, 50, 55] are capable of editing texts in the image directly. While text editing and our task setup seem similar, text editing is only a part of the typography retouching task, which involves further tasks such as automatic error detection or missing word insertion." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 642, + 380, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 380, + 656 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 380, + 656 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "Our Type-R pipeline consists of four stages, which are all automatic: 1) error detection, 2) text erasing, 3) layout regeneration, and 4) typo correction. We show the pipeline overview in Fig. 2." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2746" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 69, + 152, + 163 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 152, + 163 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 152, + 163 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 152, + 163 + ], + "type": "image", + "image_path": "da7b0999f7424eef3557508a0b85fa2761b2bec67f0e774281a6743e1c8e0c9d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 163, + 118, + 171 + ], + "lines": [ + { + "bbox": [ + 64, + 163, + 118, + 171 + ], + "spans": [ + { + "bbox": [ + 64, + 163, + 118, + 171 + ], + "type": "text", + "content": "Generated image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 178, + 149, + 220 + ], + "lines": [ + { + "bbox": [ + 61, + 178, + 149, + 220 + ], + "spans": [ + { + "bbox": [ + 61, + 178, + 149, + 220 + ], + "type": "text", + "content": "Create an advertisement for a hamburger with a pop background. Display the texts '10% OFF and HAMBURGER' with a tasty burger, fried potatoes, and a Coke. The text 'Combo Meals' is required. Prompt" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 156, + 69, + 250, + 163 + ], + "blocks": [ + { + "bbox": [ + 156, + 69, + 250, + 163 + ], + "lines": [ + { + "bbox": [ + 156, + 69, + 250, + 163 + ], + "spans": [ + { + "bbox": [ + 156, + 69, + 250, + 163 + ], + "type": "image", + "image_path": "df2d3b673ee05efb2be9e89e91b57815dd84fbce33aafb825389787ecc58d188.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 157, + 167, + 250, + 225 + ], + "blocks": [ + { + "bbox": [ + 83, + 227, + 130, + 236 + ], + "lines": [ + { + "bbox": [ + 83, + 227, + 130, + 236 + ], + "spans": [ + { + "bbox": [ + 83, + 227, + 130, + 236 + ], + "type": "text", + "content": "Generation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 157, + 167, + 250, + 225 + ], + "lines": [ + { + "bbox": [ + 157, + 167, + 250, + 225 + ], + "spans": [ + { + "bbox": [ + 157, + 167, + 250, + 225 + ], + "type": "image", + "image_path": "db645485cc70c71938434ce1f57df5d4c07561865d3b5a7078aa1920d22c3355.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 172, + 227, + 236, + 236 + ], + "lines": [ + { + "bbox": [ + 172, + 227, + 236, + 236 + ], + "spans": [ + { + "bbox": [ + 172, + 227, + 236, + 236 + ], + "type": "text", + "content": "Error detection" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 247, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 247, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 555, + 270 + ], + "type": "text", + "content": "Figure 2. Illustration of the Type-R pipeline. Type-R automatically detects errors, erases unintended texts, inserts missing words, and corrects spelling errors in the image." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 252, + 69, + 347, + 163 + ], + "blocks": [ + { + "bbox": [ + 252, + 69, + 347, + 163 + ], + "lines": [ + { + "bbox": [ + 252, + 69, + 347, + 163 + ], + "spans": [ + { + "bbox": [ + 252, + 69, + 347, + 163 + ], + "type": "image", + "image_path": "7b27e99f440aaa70e4c90a60f597cdcb1561a5c8849cdc197f3042a276bd3bad.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 253, + 167, + 346, + 225 + ], + "blocks": [ + { + "bbox": [ + 253, + 167, + 346, + 225 + ], + "lines": [ + { + "bbox": [ + 253, + 167, + 346, + 225 + ], + "spans": [ + { + "bbox": [ + 253, + 167, + 346, + 225 + ], + "type": "image", + "image_path": "2b630661ec278ff3b568e862f49b517f601db572ee655793d1a6fc5740dbaaae.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 274, + 228, + 326, + 238 + ], + "lines": [ + { + "bbox": [ + 274, + 228, + 326, + 238 + ], + "spans": [ + { + "bbox": [ + 274, + 228, + 326, + 238 + ], + "type": "text", + "content": "Text erasing" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 349, + 69, + 443, + 163 + ], + "blocks": [ + { + "bbox": [ + 349, + 69, + 443, + 163 + ], + "lines": [ + { + "bbox": [ + 349, + 69, + 443, + 163 + ], + "spans": [ + { + "bbox": [ + 349, + 69, + 443, + 163 + ], + "type": "image", + "image_path": "b5d5466011a573b8bf20e6f4edd341467c9aba83d770805ad11488fedbee07c6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 350, + 167, + 443, + 225 + ], + "blocks": [ + { + "bbox": [ + 350, + 167, + 443, + 225 + ], + "lines": [ + { + "bbox": [ + 350, + 167, + 443, + 225 + ], + "spans": [ + { + "bbox": [ + 350, + 167, + 443, + 225 + ], + "type": "image", + "image_path": "8d12660020e55b3e3718902cc82dc65c113d80da4c0beb9ff56b71ad008b262b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 228, + 439, + 238 + ], + "lines": [ + { + "bbox": [ + 356, + 228, + 439, + 238 + ], + "spans": [ + { + "bbox": [ + 356, + 228, + 439, + 238 + ], + "type": "text", + "content": "Layout regeneration" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 446, + 69, + 541, + 163 + ], + "blocks": [ + { + "bbox": [ + 446, + 69, + 541, + 163 + ], + "lines": [ + { + "bbox": [ + 446, + 69, + 541, + 163 + ], + "spans": [ + { + "bbox": [ + 446, + 69, + 541, + 163 + ], + "type": "image", + "image_path": "77e45fe2dce800364594956fd18bfe0fe22385c32e059f43e80919285ac57ad6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 447, + 167, + 539, + 225 + ], + "blocks": [ + { + "bbox": [ + 447, + 167, + 539, + 225 + ], + "lines": [ + { + "bbox": [ + 447, + 167, + 539, + 225 + ], + "spans": [ + { + "bbox": [ + 447, + 167, + 539, + 225 + ], + "type": "image", + "image_path": "99438419cdb13ddb5a10afa1aa71fdb5fe492803bdd645c9b680d00fc3cec239.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 454, + 226, + 531, + 238 + ], + "lines": [ + { + "bbox": [ + 454, + 226, + 531, + 238 + ], + "spans": [ + { + "bbox": [ + 454, + 226, + 531, + 238 + ], + "type": "text", + "content": "3 Typo correction" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 291, + 153, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 153, + 303 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 153, + 303 + ], + "type": "text", + "content": "3.1. Error Detection" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "text", + "content": "We denote an image " + }, + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "text", + "content": " generated by an off-the-shelf text-to-image model " + }, + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "text", + "content": " with a prompt, which requests to show " + }, + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "text", + "content": " words " + }, + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "inline_equation", + "content": "W = \\{w_{i}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "text", + "content": ". To ensure how much this expectation is satisfied, we first detect words in the image " + }, + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "text", + "content": ". Specifically, a scene text detection model predicts polygonal word regions " + }, + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\hat{B}" + }, + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "text", + "content": ". We remove polygons with a height less than a threshold " + }, + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "text", + "content": ". Then, a scene text recognition model parses the words " + }, + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\hat{W} = \\{\\hat{w}_i\\}_{i = 1}^{\\hat{N}}" + }, + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "text", + "content": " in the regions " + }, + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\hat{B}" + }, + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "text", + "content": ". As noted in Section 1, the model " + }, + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "text", + "content": " is often not perfect, and " + }, + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "inline_equation", + "content": "W \\neq \\hat{W}" + }, + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "text", + "content": "; therefore, we need to identify errors in " + }, + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\hat{W}" + }, + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "text", + "content": " by considering " + }, + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 309, + 296, + 443 + ], + "type": "text", + "content": " is a ground-truth." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 444, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 296, + 491 + ], + "type": "text", + "content": "Roughly speaking, this identification task is a matching problem between " + }, + { + "bbox": [ + 55, + 444, + 296, + 491 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 444, + 296, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 444, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\hat{W}" + }, + { + "bbox": [ + 55, + 444, + 296, + 491 + ], + "type": "text", + "content": ". More specifically, we consider the matching between the word set " + }, + { + "bbox": [ + 55, + 444, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 55, + 444, + 296, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 444, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{W}}" + }, + { + "bbox": [ + 55, + 444, + 296, + 491 + ], + "type": "text", + "content": " of equal size " + }, + { + "bbox": [ + 55, + 444, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 55, + 444, + 296, + 491 + ], + "type": "text", + "content": " by optimal transport:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 94, + 500, + 294, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 500, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 94, + 500, + 294, + 533 + ], + "type": "interline_equation", + "content": "\\mathrm {O T} (\\mathcal {W}, \\hat {\\mathcal {W}}) \\equiv \\max _ {\\pi} \\sum_ {i = 1} ^ {\\mathcal {N}} \\operatorname {L e v} \\left(w _ {i}, \\hat {w} _ {\\pi (i)}\\right), \\tag {1}", + "image_path": "b65a890324c45ad9924a993bda38bbca2ca647d4f23ffe4f323ab13afa65020f.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 544, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 296, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 544, + 296, + 628 + ], + "type": "inline_equation", + "content": "w_{i}\\in \\mathcal{W},\\hat{w}_{i}\\in \\hat{\\mathcal{W}},\\pi" + }, + { + "bbox": [ + 55, + 544, + 296, + 628 + ], + "type": "text", + "content": " is a one-by-one matching, and " + }, + { + "bbox": [ + 55, + 544, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{Levenshtein}" + }, + { + "bbox": [ + 55, + 544, + 296, + 628 + ], + "type": "text", + "content": " distance to measure the text similarity. This matching works only when the two sets are of equal size, but typographic errors result in different sizes because of missing or unintended words. Here, we equalize two word sets " + }, + { + "bbox": [ + 55, + 544, + 296, + 628 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 544, + 296, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 544, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\hat{W}" + }, + { + "bbox": [ + 55, + 544, + 296, + 628 + ], + "type": "text", + "content": " by padding tokens. Let us define " + }, + { + "bbox": [ + 55, + 544, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{N} = \\max (N,\\hat{N})" + }, + { + "bbox": [ + 55, + 544, + 296, + 628 + ], + "type": "text", + "content": ". We express the padded word sets by:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 127, + 648, + 294, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 648, + 294, + 663 + ], + "spans": [ + { + "bbox": [ + 127, + 648, + 294, + 663 + ], + "type": "interline_equation", + "content": "\\mathcal {W} = W \\cup \\left\\{p _ {i} \\right\\} _ {i = N + 1} ^ {\\mathcal {N}}, \\tag {2}", + "image_path": "24759dd8d482c5373e7a327659061cb1b36719514ef2bc5e9243f860534f8d38.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 127, + 665, + 294, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 665, + 294, + 682 + ], + "spans": [ + { + "bbox": [ + 127, + 665, + 294, + 682 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {W}} = \\hat {W} \\cup \\left\\{\\hat {p} _ {i} \\right\\} _ {i = \\hat {N} + 1} ^ {\\mathcal {N}}, \\tag {3}", + "image_path": "7bfeff7aae9ab6807050f6fbae6847cc65543ac9aad04385809b46d7127a86da.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{p}_i" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " are the padding tokens. Here, a word matching a padding token indicates a typographic error; i.e.," + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 292, + 553, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 553, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 553, + 339 + ], + "type": "text", + "content": "a paired word with " + }, + { + "bbox": [ + 313, + 292, + 553, + 339 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 292, + 553, + 339 + ], + "type": "text", + "content": " is an unintended word, and a paired word with " + }, + { + "bbox": [ + 313, + 292, + 553, + 339 + ], + "type": "inline_equation", + "content": "\\hat{p}" + }, + { + "bbox": [ + 313, + 292, + 553, + 339 + ], + "type": "text", + "content": " is a missing word. We assign a constant distance to any word matching to the padding tokens. Any pairs having non-zero distance indicate spelling errors." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 340, + 554, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 554, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 554, + 411 + ], + "type": "text", + "content": "The following sections describe our approach to minimizing the above optimal transportation cost. We align the number of words by erasing redundant texts when " + }, + { + "bbox": [ + 313, + 340, + 554, + 411 + ], + "type": "inline_equation", + "content": "N < \\hat{N}" + }, + { + "bbox": [ + 313, + 340, + 554, + 411 + ], + "type": "text", + "content": " (Sec. 3.2) and insert text boxes by layout regeneration when " + }, + { + "bbox": [ + 313, + 340, + 554, + 411 + ], + "type": "inline_equation", + "content": "N > \\hat{N}" + }, + { + "bbox": [ + 313, + 340, + 554, + 411 + ], + "type": "text", + "content": " (Sec. 3.3). Finally, we fix spelling errors in typo correction (Sec. 3.4)." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 314, + 418, + 397, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 418, + 397, + 431 + ], + "spans": [ + { + "bbox": [ + 314, + 418, + 397, + 431 + ], + "type": "text", + "content": "3.2. Text Erasing" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 437, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 581 + ], + "type": "text", + "content": "Type-R erases unintended texts using an off-the-shelf inpainting method [37] if the number of detected words exceeds the number of the words specified by a prompt, i.e., " + }, + { + "bbox": [ + 313, + 437, + 555, + 581 + ], + "type": "inline_equation", + "content": "N < \\hat{N}" + }, + { + "bbox": [ + 313, + 437, + 555, + 581 + ], + "type": "text", + "content": ", or the generated image includes filtered texts. This step reduces the number of words in a generated image to equal or less than the number of specified words by the prompt. The inpainting model removes specified paint regions. In our setup, we specify regions that we want to erase in " + }, + { + "bbox": [ + 313, + 437, + 555, + 581 + ], + "type": "inline_equation", + "content": "\\hat{B}" + }, + { + "bbox": [ + 313, + 437, + 555, + 581 + ], + "type": "text", + "content": " and composite the inpainted results to the original image. Since text detection may not always fully cover text areas, we slightly enlarge the text region masks to ensure complete removal." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 587, + 436, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 436, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 436, + 601 + ], + "type": "text", + "content": "3.3. Layout Regeneration" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "Layout regeneration fills in missing text regions so that the resulting " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\hat{N}" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " equals the number of words specified by a prompt " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\hat{N} < N" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": ". This is equivalent to obtaining position information from the image for missing words of the prompt, and we address this task using a VLM, specifically GPT-4o [4]. Inspired by works on layout generation by zero/few-shot learning of LLMs [12, 23], we ask GPT to plan layouts for each missing word from the image and layout information of valid text boxes in JSON format." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2747" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "content": "For simplicity, we represent the layout information as a bounding box with four coordinates. When text detection outputs are in the polygonal form, we convert them to boxes by finding the smallest four-sided bounding box that encloses each polygon." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 142, + 155, + 155 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 142, + 155, + 155 + ], + "spans": [ + { + "bbox": [ + 55, + 142, + 155, + 155 + ], + "type": "text", + "content": "3.4. Tipo Correction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 161, + 296, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 161, + 296, + 317 + ], + "spans": [ + { + "bbox": [ + 55, + 161, + 296, + 317 + ], + "type": "text", + "content": "Typo correction tries to resolve all word pairs with non-zero edit distance. To address word-level spelling errors, we employ a pre-trained text editing model [39] to automatically retouch the word region. Using an inpainting approach, the text editing model generates raster images containing specified text in the selected areas. Similar to text-to-image models, text editing models may not always render the specified text accurately but with a much lower computation cost than large text-to-image models. We repeatedly apply the text editing model until we fix the spelling errors or we reach the predefined number of attempts. Note that we only composite the areas of corrected words in the edited image during the iteration. We show the pseudo-codes in Algorithm 1." + } + ] + } + ], + "index": 2 + }, + { + "type": "code", + "bbox": [ + 57, + 345, + 296, + 526 + ], + "blocks": [ + { + "bbox": [ + 57, + 331, + 249, + 342 + ], + "lines": [ + { + "bbox": [ + 57, + 331, + 249, + 342 + ], + "spans": [ + { + "bbox": [ + 57, + 331, + 249, + 342 + ], + "type": "text", + "content": "Algorithm 1 A pseudo-code of typo correction." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "lines": [ + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "spans": [ + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "type": "text", + "content": "1: function TyPO CORRECTION(Image I, Words W, Detected regions " + }, + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "type": "inline_equation", + "content": "\\hat{B}" + }, + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "type": "text", + "content": " , Maximum iterations " + }, + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "type": "inline_equation", + "content": "t_\\mathrm{max})" + }, + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "type": "text", + "content": " \n2: for " + }, + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "type": "inline_equation", + "content": "t_\\mathrm{max}" + }, + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "type": "text", + "content": " do \n3: " + }, + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{edit}}\\gets" + }, + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "type": "text", + "content": " Text-Editor(I,W," + }, + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "type": "inline_equation", + "content": "\\hat{B}" + }, + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "type": "text", + "content": ") \n4: " + }, + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "type": "inline_equation", + "content": "\\hat{W}\\gets" + }, + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "type": "text", + "content": " Text-Recognition " + }, + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "type": "inline_equation", + "content": "(I_{\\mathrm{edit}},\\hat{B})" + }, + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "type": "text", + "content": " \n5: " + }, + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "type": "inline_equation", + "content": "\\mathcal{O}\\gets \\mathrm{Diff}(W,\\hat{W})\\qquad \\triangleright" + }, + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "type": "text", + "content": " Find spelling errors \n6: " + }, + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "type": "inline_equation", + "content": "I\\gets" + }, + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "type": "text", + "content": " Composite " + }, + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "type": "inline_equation", + "content": "(I,I_{\\mathrm{edit}},\\hat{B},\\mathcal{O})" + }, + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "type": "text", + "content": " \n7: if " + }, + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "type": "inline_equation", + "content": "\\mathcal{O}\\neq \\varnothing" + }, + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "type": "text", + "content": " then \n8: " + }, + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "type": "inline_equation", + "content": "W,\\hat{B}\\gets" + }, + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "type": "text", + "content": " Update(W," + }, + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "type": "inline_equation", + "content": "\\hat{B},\\mathcal{O})" + }, + { + "bbox": [ + 57, + 345, + 296, + 526 + ], + "type": "text", + "content": " \n9: else \n10: break \n11: end if \n12: end for \n13: return I \n14: end function" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "code_body" + } + ], + "index": 4, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 554, + 132, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 132, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 132, + 567 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 574, + 114, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 114, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 114, + 586 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "We evaluate the performance of visual text generation on MARIO-Eval benchmark [9]. This benchmark contains 5,414 prompts, of which 5,000 have image data. We use all 5,414 prompts in the main comparison, i.e., in Tab. 1. We randomly pick 500 of the 5,414 prompts as the subset for analysis in Sec. 4.6 because our experiments involve paid APIs. For the selection of hyperparameters in filtering, we use 500 prompts randomly selected from the MARIO-LAION dataset [9] as the validation set. Note that the validation does not overlap with the benchmark." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 72, + 373, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 373, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 373, + 83 + ], + "type": "text", + "content": "4.2. Metrics" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 92, + 555, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 92, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 92, + 555, + 273 + ], + "type": "text", + "content": "We evaluate the quality of generated images using a GPT-based evaluation similar to TextDiffuser-2 [10]. In our evaluation, we adopt a rating-based evaluation approach [15] instead of the voting evaluation because GPT-based voting is sensitive to the option order and the number of options. In our settings, GPT assigns scores from 1 to 10 for both graphic design quality and content alignment between generated images and prompts. For text rendering accuracy, we measure OCR accuracy commonly used for assessing text rendering performance used in text-to-image generation [9, 10]. We also measure Fréchet Inception Distance (FID) [13] to quantify distribution differences between generated images and the full 5000 images in the MARIO-Eval benchmark, and CLIP score [32] to compute the similarity between each image and its prompt-based on the CLIP." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 286, + 447, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 286, + 447, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 286, + 447, + 300 + ], + "type": "text", + "content": "4.3. Implementation Details" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 307, + 555, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 555, + 438 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 555, + 438 + ], + "type": "text", + "content": "We use multiple external models in our Type-R pipeline without fine-tuning. We use Deepsolo [48] for text region detection, Paddle [5] for scene text recognition, LaMa [38] for text erasing, GPT-4o [4] for layout regeneration, and AnyText [39] for text editing. In addition, we use GPT to augment prompts following [42]. When augmenting prompts with GPT, we specify what categories of images to generate and how to draw the text in quotation marks to make it stand out. For all GPT utilization in this paper, i.e., augmentation, layout planning, and evaluation, we use the same version of OpenAI API 2023-05-15." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 441, + 554, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 554, + 502 + ], + "type": "text", + "content": "When the response of the API is corrupted, we retry requests up to five times and give up. For typo correction iteration, we set the maximum number of correction attempts " + }, + { + "bbox": [ + 313, + 441, + 554, + 502 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{max}}" + }, + { + "bbox": [ + 313, + 441, + 554, + 502 + ], + "type": "text", + "content": " to 10. We chose the hyperparameter " + }, + { + "bbox": [ + 313, + 441, + 554, + 502 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 441, + 554, + 502 + ], + "type": "text", + "content": " for filtering to " + }, + { + "bbox": [ + 313, + 441, + 554, + 502 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 313, + 441, + 554, + 502 + ], + "type": "text", + "content": " of the overall image height through grid search." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 516, + 380, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 516, + 380, + 528 + ], + "spans": [ + { + "bbox": [ + 314, + 516, + 380, + 528 + ], + "type": "text", + "content": "4.4. Baselines" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 536, + 554, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 554, + 584 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 554, + 584 + ], + "type": "text", + "content": "Since Type-R is a post-processing module, we combine Type-R with the two base text-to-image models to build a baseline. In addition, we chose a few different baselines tailored for text rendering." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 586, + 554, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 554, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 554, + 611 + ], + "type": "text", + "content": "Type-R w/ SD3 combines Stable Diffusion 3 [11] and our approach." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 613, + 533, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 533, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 533, + 625 + ], + "type": "text", + "content": "Type-R w/ Flux combines Flux [3] and our approach." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 627, + 554, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 627, + 554, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 627, + 554, + 663 + ], + "type": "text", + "content": "TextDiffuser [9] is a text-to-image generation method based on layout planning, which generates character-wise masks and generates images using the masks as conditions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "content": "TextDiffuser-2 [10] is a state-of-the-art text-to-image generation method based on layout planning. Unlike TextDiffuser, TextDiffuser-2 plans layouts via LLM and generates images by embedding the layout information into prompts." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2748" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 75, + 527, + 358 + ], + "blocks": [ + { + "bbox": [ + 80, + 75, + 527, + 358 + ], + "lines": [ + { + "bbox": [ + 80, + 75, + 527, + 358 + ], + "spans": [ + { + "bbox": [ + 80, + 75, + 527, + 358 + ], + "type": "image", + "image_path": "994473edf7b9c9b48ce64157a4252de501bb31254371053c3944a85baa6e3932.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 373, + 555, + 396 + ], + "lines": [ + { + "bbox": [ + 54, + 373, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 555, + 396 + ], + "type": "text", + "content": "Figure 3. Comparisons of generated images by Flux, Flux w/ Type-R, and TextDiffuser models. The left column shows prompts, and the right columns present generated images by each method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 415, + 113, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 113, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 113, + 427 + ], + "type": "text", + "content": "4.5. Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 434, + 295, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 577 + ], + "type": "text", + "content": "Type-R vs. TextDiffusers Table 1 summarizes evaluation metrics of TextDiffuser models and Type-R. The result suggests that Type-R significantly outperforms the TextDiffuser models in GPT evaluation scores, and also Type-R achieves higher content alignment. The higher GPT scores indicate better aesthetic quality, as also suggested in the qualitative results in Figure 3. While the TextDiffuser models tend to generate images with simple backgrounds and texts (i.e., low quality in terms of graphic designs), Type-R successfully generates images with contextually rich backgrounds and seamlessly integrated text through precise and corrective adjustments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 578, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 295, + 639 + ], + "type": "text", + "content": "In OCR evaluations, SD3 with Type-R yields lower scores than the TextDiffuser models, whereas Flux with Type-R achieves higher accuracy. This result suggests that the enhancement of Type-R is influenced by the capability of the base text-to-image model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "text", + "content": "User Study We also conducted a user study in which participants rated images using the same questions and scoring criteria as the GPT evaluations to validate the reliability of the GPT evaluation. We prepared 100 prompts and generated 100 corresponding images, which resulted in scores" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 313, + 456, + 558, + 544 + ], + "blocks": [ + { + "bbox": [ + 313, + 415, + 555, + 449 + ], + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 449 + ], + "type": "text", + "content": "Table 1. Evaluation metrics for Type-R and TextDiffuser baselines. \"Graphic\" and \"Match\" indicate graphic design quality and content alignment." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 456, + 558, + 544 + ], + "lines": [ + { + "bbox": [ + 313, + 456, + 558, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 558, + 544 + ], + "type": "table", + "html": "
MethodGPT↑OCR ↑FID↓CLIP ↑
GraphicMatch
TD [9]4.677.1655.442.034.4
TD-2 [10]4.977.1756.233.834.7
Type-R
SD3[11]7.308.4648.645.033.0
Flux[3]7.678.5562.043.133.1
", + "image_path": "32f502007b201536385bdd6ebdd2819612acee1b9305bde501d35efd2cace022.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 568, + 542, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 542, + 580 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 542, + 580 + ], + "type": "text", + "content": "from 5 participants for a total of 500 ratings per method." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 582, + 555, + 715 + ], + "type": "text", + "content": "Table 2 presents the user study results. Both SD3 and Flux with Type-R have higher scores than the TextDiffuser models, consistent with the GPT evaluation. The differences in scores are small among all baselines, although we observe that all scores of Type-R are statistically significant compared to TextDiffuser baselines. The p-value of the paired t-test with Type-R w/ Flux is " + }, + { + "bbox": [ + 312, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "1.55e - 11" + }, + { + "bbox": [ + 312, + 582, + 555, + 715 + ], + "type": "text", + "content": " for TextDiffuser and " + }, + { + "bbox": [ + 312, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "4.14e - 18" + }, + { + "bbox": [ + 312, + 582, + 555, + 715 + ], + "type": "text", + "content": " for TextDiffuser 2 in text-image matching evaluation. The small difference might be due to a lack of confidence in rating design or the ability to perceive finer defects in graphic design." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2749" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 91, + 272, + 178 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 271, + 82 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 271, + 82 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 271, + 82 + ], + "type": "text", + "content": "Table 2. Userstudy for reliability of GPT evaluation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 91, + 272, + 178 + ], + "lines": [ + { + "bbox": [ + 82, + 91, + 272, + 178 + ], + "spans": [ + { + "bbox": [ + 82, + 91, + 272, + 178 + ], + "type": "table", + "html": "
MethodHuman ↑
GraphicMatch
TextDiffuser [9]5.24 ±2.185.08 ±2.51
TextDiffuser-2 [10]5.30 ±1.605.06 ±2.39
Type-R
SD3[11]5.94 ±1.865.99 ±2.34
Flux[3]6.23 ±1.816.04 ±2.34
", + "image_path": "ba8dca761a6201232d625dc4b954dfdbaf04e491598972a3681426f3e1df69fa.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 103, + 189, + 168, + 252 + ], + "blocks": [ + { + "bbox": [ + 103, + 189, + 168, + 252 + ], + "lines": [ + { + "bbox": [ + 103, + 189, + 168, + 252 + ], + "spans": [ + { + "bbox": [ + 103, + 189, + 168, + 252 + ], + "type": "image", + "image_path": "42a168055db44ba50a930569c8f5ba22ac5c5a342d1b3f91eb43fe74b513f216.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 181, + 189, + 246, + 252 + ], + "blocks": [ + { + "bbox": [ + 181, + 189, + 246, + 252 + ], + "lines": [ + { + "bbox": [ + 181, + 189, + 246, + 252 + ], + "spans": [ + { + "bbox": [ + 181, + 189, + 246, + 252 + ], + "type": "image", + "image_path": "b2a4f9edad73f22541b842e6e3fb99d780aa16e556cba528984318ccb6759f48.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 102, + 254, + 168, + 325 + ], + "blocks": [ + { + "bbox": [ + 102, + 254, + 168, + 325 + ], + "lines": [ + { + "bbox": [ + 102, + 254, + 168, + 325 + ], + "spans": [ + { + "bbox": [ + 102, + 254, + 168, + 325 + ], + "type": "image", + "image_path": "2617594a8f8c1b70525e15f7ff3a8f3fe5dbf0feb1e0fd9254017515ddaa0191.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 125, + 327, + 143, + 334 + ], + "lines": [ + { + "bbox": [ + 125, + 327, + 143, + 334 + ], + "spans": [ + { + "bbox": [ + 125, + 327, + 143, + 334 + ], + "type": "text", + "content": "B&W" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 181, + 253, + 246, + 326 + ], + "blocks": [ + { + "bbox": [ + 181, + 253, + 246, + 326 + ], + "lines": [ + { + "bbox": [ + 181, + 253, + 246, + 326 + ], + "spans": [ + { + "bbox": [ + 181, + 253, + 246, + 326 + ], + "type": "image", + "image_path": "59c4635c3f14832267ec908a5db311575123405d6357b0e75b409e77b3faec1d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 327, + 236, + 335 + ], + "lines": [ + { + "bbox": [ + 189, + 327, + 236, + 335 + ], + "spans": [ + { + "bbox": [ + 189, + 327, + 236, + 335 + ], + "type": "text", + "content": "w/ white box" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 104, + 349, + 246, + 360 + ], + "lines": [ + { + "bbox": [ + 104, + 349, + 246, + 360 + ], + "spans": [ + { + "bbox": [ + 104, + 349, + 246, + 360 + ], + "type": "text", + "content": "Figure 4. Examples of naive baselines." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 382, + 295, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 382, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 54, + 382, + 295, + 467 + ], + "type": "text", + "content": "FID and CLIP scores Although TextDiffuser yields higher FID and CLIP scores, we argue that these scores may not accurately capture human perception of image quality. For FID, the reference dataset is limited to 5,000 images that are not always consistent in quality, which likely causes the FID scores to overfit the narrow reference corpus, making it unsuitable for evaluating graphic design quality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 467, + 295, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 467, + 295, + 551 + ], + "spans": [ + { + "bbox": [ + 54, + 467, + 295, + 551 + ], + "type": "text", + "content": "TextDiffuser also yields higher CLIP scores. We suspect this result is due to the text-spotting bias of the CLIP model, which tends to \"parrot\" the visual text embedding in images [24]. That is, the CLIP score tends to prefer words over visual objects in the image, leading to a higher score for the word-only image without a meaningful background visual, which TextDiffusers tend to generate." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 560, + 157, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 157, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 157, + 574 + ], + "type": "text", + "content": "4.6. Additional Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 579, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 579, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 579, + 295, + 628 + ], + "type": "text", + "content": "Quality-accuracy Trade-off Here, we compare Type-R with several baselines in terms of the trade-off between image quality and OCR accuracy. In this study, we introduce the following backend text-to-image models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 628, + 286, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 286, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 286, + 640 + ], + "type": "text", + "content": "Type-R w/Dall-E 3 combines Dall-E 3 [1] with Type-R." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 641, + 294, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 294, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 294, + 665 + ], + "type": "text", + "content": "Type-R w/ SD 3.5 combines Stable Diffusion 3.5 [6] with Type-R." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "We compare the raw output (Raw) from the backend text-to-image models and the output with Type-R. We also consider the following naive baselines (Fig. 4). We heuristically arrange text boxes for all the naive baselines, ignoring" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 317, + 74, + 547, + 323 + ], + "blocks": [ + { + "bbox": [ + 317, + 74, + 547, + 323 + ], + "lines": [ + { + "bbox": [ + 317, + 74, + 547, + 323 + ], + "spans": [ + { + "bbox": [ + 317, + 74, + 547, + 323 + ], + "type": "image", + "image_path": "cdcc5f6e3ac771a26980494add07b8371df32e6c58528a0d5d93c4ad56a0b36b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 333, + 555, + 367 + ], + "lines": [ + { + "bbox": [ + 313, + 333, + 555, + 367 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 555, + 367 + ], + "type": "text", + "content": "Figure 5. Plot of the relation between OCR accuracy and graphic design quality by GPT. Raw represents the results of text-to-image generation with prompt augmentation." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 388, + 391, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 391, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 391, + 399 + ], + "type": "text", + "content": "the image content." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 399, + 554, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 554, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 554, + 423 + ], + "type": "text", + "content": "Simple text erases all texts in the generated image from the backend text-to-image models and uses it as a background." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 423, + 553, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 553, + 447 + ], + "type": "text", + "content": "Simple text w/ white box also removes all texts and white out the text box regions." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 448, + 517, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 517, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 517, + 460 + ], + "type": "text", + "content": "B&W always renders text on a white background." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 460, + 554, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 579 + ], + "type": "text", + "content": "Fig. 5 plots the trade-off between GPT score and OCR accuracy, which shows that naive baselines can achieve high OCR accuracy for much lower GPT score. On the other hand, Flux and Dall-E 3 with Type-R show much higher graphic design quality with a slightly degrading OCR accuracy, almost reaching the accuracy of the simple text rendering. Type-R exhibits robustness across multiple backend text-to-image models, as demonstrated in the plot where Type-R effectively improves OCR accuracy for any backend models while maintaining high graphic design quality." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "Qualitative Results Figure 6 shows a few generated images by Type-R demonstrating how each external model improves text rendering accuracy. In the first row, the text-editing model accurately modifies the text using a fancy font that aligns with the surrounding context. The second row presents an instance of Type-R successfully removing unnecessary texts. The third row provides an example of removing small text and adding a missing text through layout planning, where the inserted new text implicitly utilizes surrounding font styles to achieve a seamless and natural" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2750" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 80, + 135, + 158 + ], + "blocks": [ + { + "bbox": [ + 128, + 71, + 217, + 79 + ], + "lines": [ + { + "bbox": [ + 128, + 71, + 217, + 79 + ], + "spans": [ + { + "bbox": [ + 128, + 71, + 217, + 79 + ], + "type": "text", + "content": "Prompt: 'Nobliaux' et sorcires" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 80, + 135, + 158 + ], + "lines": [ + { + "bbox": [ + 58, + 80, + 135, + 158 + ], + "spans": [ + { + "bbox": [ + 58, + 80, + 135, + 158 + ], + "type": "image", + "image_path": "097d5dc41a14f77cf80af4a6599ef89c9578f88919fc24c3b72a9733fcdfcb58.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 136, + 80, + 214, + 158 + ], + "blocks": [ + { + "bbox": [ + 136, + 80, + 214, + 158 + ], + "lines": [ + { + "bbox": [ + 136, + 80, + 214, + 158 + ], + "spans": [ + { + "bbox": [ + 136, + 80, + 214, + 158 + ], + "type": "image", + "image_path": "03527dabf3d3dba81a3cb9fd617edb84ed345a4e6e14db1db2aeaf8d1b46db52.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 215, + 80, + 292, + 158 + ], + "blocks": [ + { + "bbox": [ + 215, + 80, + 292, + 158 + ], + "lines": [ + { + "bbox": [ + 215, + 80, + 292, + 158 + ], + "spans": [ + { + "bbox": [ + 215, + 80, + 292, + 158 + ], + "type": "image", + "image_path": "02f0db7562851db71bdda99c7e96ab3afad03c2f9651eb7287b9761c7f11c645.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 171, + 137, + 250 + ], + "blocks": [ + { + "bbox": [ + 83, + 161, + 267, + 171 + ], + "lines": [ + { + "bbox": [ + 83, + 161, + 267, + 171 + ], + "spans": [ + { + "bbox": [ + 83, + 161, + 267, + 171 + ], + "type": "text", + "content": "Prompt: Image of MINI 'GPS' MAGNETIC CAR TRACKER" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 171, + 137, + 250 + ], + "lines": [ + { + "bbox": [ + 58, + 171, + 137, + 250 + ], + "spans": [ + { + "bbox": [ + 58, + 171, + 137, + 250 + ], + "type": "image", + "image_path": "6a0f62b2de2cdc51d5a35a7b5e58471cff8416acd30f45021839cd34dde9dc1f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 252, + 250, + 262 + ], + "lines": [ + { + "bbox": [ + 86, + 252, + 250, + 262 + ], + "spans": [ + { + "bbox": [ + 86, + 252, + 250, + 262 + ], + "type": "text", + "content": "Prompt: Title 'Elise Chain 'Reaction' Large Paperback" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 137, + 171, + 214, + 250 + ], + "blocks": [ + { + "bbox": [ + 137, + 171, + 214, + 250 + ], + "lines": [ + { + "bbox": [ + 137, + 171, + 214, + 250 + ], + "spans": [ + { + "bbox": [ + 137, + 171, + 214, + 250 + ], + "type": "image", + "image_path": "f359ce52b84a2a0958efc3e8ea14646469578017679795e8617532dd2654d9a3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 215, + 171, + 292, + 250 + ], + "blocks": [ + { + "bbox": [ + 215, + 171, + 292, + 250 + ], + "lines": [ + { + "bbox": [ + 215, + 171, + 292, + 250 + ], + "spans": [ + { + "bbox": [ + 215, + 171, + 292, + 250 + ], + "type": "image", + "image_path": "4981f8ad047343329bd9a243fdc5c2a678be5cfeb3c1cf2e5963eafbe696588c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 59, + 264, + 137, + 343 + ], + "blocks": [ + { + "bbox": [ + 59, + 264, + 137, + 343 + ], + "lines": [ + { + "bbox": [ + 59, + 264, + 137, + 343 + ], + "spans": [ + { + "bbox": [ + 59, + 264, + 137, + 343 + ], + "type": "image", + "image_path": "49c016edf16e509923513e0cdaff0eede1b5fbeb649ec02c1f24c1e403dc4fa7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 345, + 123, + 354 + ], + "lines": [ + { + "bbox": [ + 72, + 345, + 123, + 354 + ], + "spans": [ + { + "bbox": [ + 72, + 345, + 123, + 354 + ], + "type": "text", + "content": "Text-to-Image" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 373, + 279, + 385 + ], + "lines": [ + { + "bbox": [ + 70, + 373, + 279, + 385 + ], + "spans": [ + { + "bbox": [ + 70, + 373, + 279, + 385 + ], + "type": "text", + "content": "Figure 6. Examples of generated images through Type-R." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 143, + 264, + 214, + 343 + ], + "blocks": [ + { + "bbox": [ + 143, + 264, + 214, + 343 + ], + "lines": [ + { + "bbox": [ + 143, + 264, + 214, + 343 + ], + "spans": [ + { + "bbox": [ + 143, + 264, + 214, + 343 + ], + "type": "image", + "image_path": "88e31cc8199f472b5ba913e30283d1d113c6ec20a30ebc7f3ce3f9d9c579f722.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 345, + 212, + 362 + ], + "lines": [ + { + "bbox": [ + 143, + 345, + 212, + 362 + ], + "spans": [ + { + "bbox": [ + 143, + 345, + 212, + 362 + ], + "type": "text", + "content": "Text erasing " + }, + { + "bbox": [ + 143, + 345, + 212, + 362 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 143, + 345, + 212, + 362 + ], + "type": "text", + "content": " Layout regeneration" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 221, + 264, + 291, + 343 + ], + "blocks": [ + { + "bbox": [ + 221, + 264, + 291, + 343 + ], + "lines": [ + { + "bbox": [ + 221, + 264, + 291, + 343 + ], + "spans": [ + { + "bbox": [ + 221, + 264, + 291, + 343 + ], + "type": "image", + "image_path": "2b3d2cd5972aead784b7eb070b03c3099aeacb04bd185fc1f48dc689194b06a7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 222, + 345, + 275, + 361 + ], + "lines": [ + { + "bbox": [ + 222, + 345, + 275, + 361 + ], + "spans": [ + { + "bbox": [ + 222, + 345, + 275, + 361 + ], + "type": "text", + "content": "Typo Corrected image" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 407, + 105, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 105, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 105, + 418 + ], + "type": "text", + "content": "appearance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 54, + 418, + 295, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 418, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 54, + 418, + 295, + 515 + ], + "type": "text", + "content": "Since Type-R is independent of the backend text-to-image models, we can take advantage of the state-of-the-art models. Figure 7 presents examples with very long prompts, showing that the latest text-to-image model, Flux, accurately captures detailed information from these prompts, whereas TextDiffuser-2 generates corrupted images when handling such prompts, as CLIP is unable to process tokens exceeding 77." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 530, + 295, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 295, + 614 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 295, + 614 + ], + "type": "text", + "content": "Longer Iterations The final typo correction step iteratively refines misspelled words. We show the improvement over iterations in Fig. 8 for several backend models. Our typo correction approach consistently improves OCR accuracy as we repeat the iterations. Our approach achieves " + }, + { + "bbox": [ + 55, + 530, + 295, + 614 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 55, + 530, + 295, + 614 + ], + "type": "text", + "content": " of the total improvement within four iterations independent of the backend model." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 623, + 149, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 149, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 149, + 635 + ], + "type": "text", + "content": "4.7. Ablation Study" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": "Type-R Pipeline Modules Tab. 3 summarizes the ablation results for the modules in Type-R pipeline: augmentation for prompt, layout correction (text erasing and layout regeneration), and typo correction. We observe that each module effectively improves OCR accuracy with minimal negative impact on graphic scores by GPT. Layout correc" + } + ] + } + ], + "index": 20 + }, + { + "type": "table", + "bbox": [ + 315, + 91, + 553, + 238 + ], + "blocks": [ + { + "bbox": [ + 336, + 71, + 531, + 82 + ], + "lines": [ + { + "bbox": [ + 336, + 71, + 531, + 82 + ], + "spans": [ + { + "bbox": [ + 336, + 71, + 531, + 82 + ], + "type": "text", + "content": "Table 3. Ablation study for Type-R pipeline modules." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 91, + 553, + 238 + ], + "lines": [ + { + "bbox": [ + 315, + 91, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 315, + 91, + 553, + 238 + ], + "type": "table", + "html": "
T2IAugCorrectionGraphic ↑OCR ↑
LayoutTypo
SD3 [11]✓6.778.6
7.5512.2
✓✓7.4227.4
✓✓7.5320.4
✓✓✓7.2750.0
Flux [3]✓6.7914.4
7.8028.0
✓✓7.8353.8
✓✓7.7329.4
✓✓✓7.6563.0
", + "image_path": "7c8ffddcf5ea391de4fbe82f3a89910822e2b53feb12875a4c1ab9b0213f800d.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_body" + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 317, + 270, + 553, + 383 + ], + "blocks": [ + { + "bbox": [ + 354, + 251, + 512, + 262 + ], + "lines": [ + { + "bbox": [ + 354, + 251, + 512, + 262 + ], + "spans": [ + { + "bbox": [ + 354, + 251, + 512, + 262 + ], + "type": "text", + "content": "Table 4. Comparing backend OCR models." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 270, + 553, + 383 + ], + "lines": [ + { + "bbox": [ + 317, + 270, + 553, + 383 + ], + "spans": [ + { + "bbox": [ + 317, + 270, + 553, + 383 + ], + "type": "table", + "html": "
DetectionRecognitionOCR ↑
Paddle [5]Paddle [5]57.4
CRAFT [8]Paddle [5]47.2
Hi-SAM [49]Paddle [5]60.8
Mask TextSpotter v3 [22]Paddle [5]61.0
DeepSolo [48]Paddle [5]63.0
DeepSolo [48]Baek et al. [7]62.0
DeepSolo [48]TrOCR [19]60.0
DeepSolo [48]CLIP4STR-B* [33]60.8
", + "image_path": "9734dcb2ac198f20b5f92fecdb0aeddad6c0595b5716c7cdae02c0d7fc0fcf23.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 404, + 554, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 554, + 453 + ], + "type": "text", + "content": "tion significantly improves Flux's OCR accuracy and further boosts OCR performance in combination with typo correction. The comparison between Flux and SD3 suggests Flux exhibits lower rates of glyph corruption." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 473, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 554, + 581 + ], + "type": "text", + "content": "OCR Models We summarize the effect of replacing the backend OCR models in Tab. 4. Both detection and recognition models impact OCR accuracy, with detection model performance changes having a more substantial effect than those of the recognition model. We observe that CRAFT [8] frequently merges words separated by spaces into a unit, leading to a notable performance drop, as Type-R assumes that spaces are excluded from texts and does not account for this issue. DeepSolo [48] achieves the highest performance." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "For text recognition, improvements in OCR accuracy do not always align with scene text recognition performance. While CLIP4STR-B* [33] is one of the state-of-the-art scene text recognition models, the OCR accuracy of CLIP4STR-B* is lower than the OCR accuracy of Baek et al. [7] and Paddle [5]. Paddle achieves the highest performance, likely because AnyText [39], trained with a loss function that incorporates the Paddle model, tends to generate words that are more easily recognized by Paddle. This suggests that compatibility between the text editing and recognition models is essential for effective correction." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2751" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 191, + 70, + 370, + 160 + ], + "blocks": [ + { + "bbox": [ + 62, + 81, + 181, + 146 + ], + "lines": [ + { + "bbox": [ + 62, + 81, + 181, + 146 + ], + "spans": [ + { + "bbox": [ + 62, + 81, + 181, + 146 + ], + "type": "text", + "content": "Create a sleek and dynamic MLB logo poster for the Colorado rockies: Use a bold color palette featuring deep purples, black, and metallic silver. Incorporate iconic mountains and baseball imagery, with prominent use of the text Colorado rockedies: Ensure the text is stylishly integrated and clearly legible. Keep the design clean and simple, focusing on strong, sharp lines that convey professional atheticism and team spirit." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 191, + 70, + 370, + 160 + ], + "lines": [ + { + "bbox": [ + 191, + 70, + 370, + 160 + ], + "spans": [ + { + "bbox": [ + 191, + 70, + 370, + 160 + ], + "type": "image", + "image_path": "96023c0f29d2b4c6c231d855e275d63ac1e3c2c0fe51e7cd4e87f7e74c249053.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 161, + 192, + 243 + ], + "lines": [ + { + "bbox": [ + 57, + 161, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 57, + 161, + 192, + 243 + ], + "type": "text", + "content": "A beautifully designed poster showcasing Brazilian cultural elements, suitable for a gallery. The composition features vibrant colors such as yellow, green, and blue, representing the Brazilian flag. In the center, there is a cheerful smiley face with a glossy finish. Below the smiley face, in bold and elegant typography, the word 'Obrigado' is prominently displayed. Traditional Brazilian patterns and motifs subtly decorate the background, adding a festive and welcoming atmosphere. The overall design is simple yet striking, perfect for a gallery wrap canvas that conveys gratitude with a touch of Brazilian charm." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 372, + 70, + 548, + 159 + ], + "blocks": [ + { + "bbox": [ + 372, + 70, + 548, + 159 + ], + "lines": [ + { + "bbox": [ + 372, + 70, + 548, + 159 + ], + "spans": [ + { + "bbox": [ + 372, + 70, + 548, + 159 + ], + "type": "image", + "image_path": "3ad0a4f8f8c705f8fa7b6e5b421ab8398761fb553c2704e7af18fdea4392a275.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 201, + 161, + 275, + 247 + ], + "blocks": [ + { + "bbox": [ + 201, + 161, + 275, + 247 + ], + "lines": [ + { + "bbox": [ + 201, + 161, + 275, + 247 + ], + "spans": [ + { + "bbox": [ + 201, + 161, + 275, + 247 + ], + "type": "image", + "image_path": "bf9545ff7b8359d3b1e1cabf533767cd1c2c77a334f1d612c1073c58e1e126c7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 249, + 246, + 258 + ], + "lines": [ + { + "bbox": [ + 227, + 249, + 246, + 258 + ], + "spans": [ + { + "bbox": [ + 227, + 249, + 246, + 258 + ], + "type": "text", + "content": "Flux" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 289, + 163, + 366, + 247 + ], + "blocks": [ + { + "bbox": [ + 289, + 163, + 366, + 247 + ], + "lines": [ + { + "bbox": [ + 289, + 163, + 366, + 247 + ], + "spans": [ + { + "bbox": [ + 289, + 163, + 366, + 247 + ], + "type": "image", + "image_path": "73d8fd7669ab051f04d2aa858cb42b69067fbb7701ca1952ed74ebd97714cc79.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 295, + 248, + 358, + 258 + ], + "lines": [ + { + "bbox": [ + 295, + 248, + 358, + 258 + ], + "spans": [ + { + "bbox": [ + 295, + 248, + 358, + 258 + ], + "type": "text", + "content": "Type-R w/ Flux" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 371, + 159, + 430, + 248 + ], + "blocks": [ + { + "bbox": [ + 371, + 159, + 430, + 248 + ], + "lines": [ + { + "bbox": [ + 371, + 159, + 430, + 248 + ], + "spans": [ + { + "bbox": [ + 371, + 159, + 430, + 248 + ], + "type": "image", + "image_path": "5643d23e340a61efa2829c5f47a161b838326851e0a4002830615ce39303bc0a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 432, + 159, + 549, + 249 + ], + "blocks": [ + { + "bbox": [ + 432, + 159, + 549, + 249 + ], + "lines": [ + { + "bbox": [ + 432, + 159, + 549, + 249 + ], + "spans": [ + { + "bbox": [ + 432, + 159, + 549, + 249 + ], + "type": "image", + "image_path": "108f755b3de14df4ce443cfae66a02a9c04b2c42e488083e1caae1a0a4def7e1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 249, + 491, + 258 + ], + "lines": [ + { + "bbox": [ + 430, + 249, + 491, + 258 + ], + "spans": [ + { + "bbox": [ + 430, + 249, + 491, + 258 + ], + "type": "text", + "content": "TextDiffuser-2" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 136, + 269, + 472, + 280 + ], + "lines": [ + { + "bbox": [ + 136, + 269, + 472, + 280 + ], + "spans": [ + { + "bbox": [ + 136, + 269, + 472, + 280 + ], + "type": "text", + "content": "Figure 7. Generated images from long prompts by Flux, Flux w/ Type-R and TextDiffuser-2." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 64, + 301, + 291, + 407 + ], + "blocks": [ + { + "bbox": [ + 64, + 301, + 291, + 407 + ], + "lines": [ + { + "bbox": [ + 64, + 301, + 291, + 407 + ], + "spans": [ + { + "bbox": [ + 64, + 301, + 291, + 407 + ], + "type": "image", + "image_path": "4ff027d1d2535b97d7ef0aa6403950d3cacbe4eb174784bc1dd3c9393a411b12.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 418, + 295, + 452 + ], + "lines": [ + { + "bbox": [ + 55, + 418, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 295, + 452 + ], + "type": "text", + "content": "Figure 8. Relationship in the typo correction iteration and OCR* accuracy. Note that OCR* accuracy is computed by OCR models in Type-R which is distinct from the OCR model for evaluation." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 108, + 484, + 244, + 549 + ], + "blocks": [ + { + "bbox": [ + 99, + 465, + 251, + 476 + ], + "lines": [ + { + "bbox": [ + 99, + 465, + 251, + 476 + ], + "spans": [ + { + "bbox": [ + 99, + 465, + 251, + 476 + ], + "type": "text", + "content": "Table 5. Comparing backend text erasers." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 108, + 484, + 244, + 549 + ], + "lines": [ + { + "bbox": [ + 108, + 484, + 244, + 549 + ], + "spans": [ + { + "bbox": [ + 108, + 484, + 244, + 549 + ], + "type": "table", + "html": "
MethodErase allOCR ↑
LaMa [37]✓62.4
63.0
Garnet [17]✓53.4
53.2
", + "image_path": "733f1d0e2a9fc9359dae15bae4357cccb8217260fce82a5f7b87431fd1398fa3.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 570, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 712 + ], + "type": "text", + "content": "Text Erasing Methods We present the OCR accuracy when replacing the backend text erasing model in Tab. 5. In this study, we compare LaMa [37], a generic inpainting model, and Garnet, a text-specific inpainting model. Since inpainting models tend to be influenced by surrounding text regions, we optionally erase all text regions to prevent filling in letter-like artifacts and only focus on the background. Table 5 shows that LaMa performs well, and erasing all texts slightly improves the final OCR accuracy. We observe that Garnet tends to leave outlines of texts when erasing text, and the outlines seem to be recognized as words in the OCR evaluation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 300, + 384, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 300, + 384, + 312 + ], + "spans": [ + { + "bbox": [ + 314, + 300, + 384, + 312 + ], + "type": "text", + "content": "5. Limitation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 320, + 555, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 555, + 463 + ], + "type": "text", + "content": "Type-R depends on the performance of several external models and the base text-to-image models. We observe that OCR models sometimes fail to correctly parse all texts generated by text-to-image models. We also observe that AnyText [39] tends to generate unstable outlines of texts and degrading brightness of pixels from original images, which tends to cause incompatibility in composition. Text-to-image models sometimes fail to render with detailed layout specifications and seem to suffer from uppercase and lowercase distinctions. Although our method is applicable in any combination, we have to improve these backend models to resolve the failure cases." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 464, + 556, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 556, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 556, + 573 + ], + "type": "text", + "content": "There are other minor remaining problems. Type-R decomposes texts into words by the space. There are texts whose reading order is meaningful, but Type-R does not guarantee keeping the reading order of words in the generated image. Type-R has a relatively high computational requirement due to the stacking of multiple backend models in the pipeline, though it might be able to use the same backend in multiple stages to mitigate the issue; e.g., the test editing model can also erase texts." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 585, + 388, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 585, + 388, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 585, + 388, + 597 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "In this paper, we propose an automatic typography correction method named Type-R, which automatically detects and fixes typographic errors in the image generated by text-to-image models. Through extensive experiments, we show that Type-R effectively corrects typos while maintaining the quality of the generated images and outperforms baselines in terms of the quality-accuracy trade-off. Our approach does not require training or fine-tuning and is readily plug-gable to any existing text-to-image model." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2752" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 112 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 112 + ], + "type": "text", + "content": "[1] DALL-E 3. https://openai.com/dall-e-3. Accessed: 2024-03-15. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 114, + 294, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 114, + 294, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 294, + 135 + ], + "type": "text", + "content": "[2] DeepFloyd IF. https://github.com/deep-floyd/IF. Accessed: 2024-11-14. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 137, + 294, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 137, + 294, + 169 + ], + "spans": [ + { + "bbox": [ + 62, + 137, + 294, + 169 + ], + "type": "text", + "content": "[3] FLUX.1 [dev]. https://blackforestlabs.ai/announcing-black-forest-labs/. Accessed: 2024-11-14. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 171, + 294, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 294, + 193 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 294, + 193 + ], + "type": "text", + "content": "[4] GPT-4o. https://openai.com/index/gpt-4o-system-card/. Accessed: 2024-11-14. 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 195, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 195, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 62, + 195, + 294, + 237 + ], + "type": "text", + "content": "[5] PaddlePaddle. https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.7/doc/doc_ch/PP-OCRv4_introduction.md.4,7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 240, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 240, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 62, + 240, + 294, + 272 + ], + "type": "text", + "content": "[6] Stable Diffusion 3.5. https://huggingface.co/stabilityai/stable-diffusion-3.5-large. Accessed: 2024-11-8. 1, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 274, + 294, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 294, + 318 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 294, + 318 + ], + "type": "text", + "content": "[7] Jeonghun Baek, Geewook Kim, Junyeop Lee, Sungrae Park, Dongyoon Han, Sangwoo Yun, Seong Joon Oh, and Hwalsuk Lee. What is wrong with scene text recognition model comparisons? dataset and model analysis. In ICCV, 2019. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 319, + 294, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 294, + 351 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 294, + 351 + ], + "type": "text", + "content": "[8] Youngmin Baek, Bado Lee, Dongyoon Han, Sangdoo Yun, and Hwalsuk Lee. Character region awareness for text detection. In CVPR, 2019. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 354, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 354, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 62, + 354, + 294, + 386 + ], + "type": "text", + "content": "[9] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser: Diffusion models as text painters. In NeurIPS, 2023. 2, 4, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 388, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 388, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 57, + 388, + 294, + 430 + ], + "type": "text", + "content": "[10] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser-2: Unleashing the power of language models for text rendering. In ECCV, 2024. 2, 4, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 434, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 434, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 57, + 434, + 294, + 487 + ], + "type": "text", + "content": "[11] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 489, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 489, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 57, + 489, + 294, + 542 + ], + "type": "text", + "content": "[12] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Eric Wang, and William Yang Wang. Layoutgpt: Compositional visual planning and generation with large language models. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 544, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 294, + 588 + ], + "type": "text", + "content": "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. arXiv preprint arXiv:1706.08500, 2017. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 590, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 294, + 634 + ], + "type": "text", + "content": "[14] Jiabao Ji, Guanhua Zhang, Zhaowen Wang, Bairu Hou, Zhifei Zhang, Brian Price, and Shiyu Chang. Improving diffusion models for scene text editing with dual encoders. arXiv preprint arXiv:2304.05568, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 294, + 689 + ], + "type": "text", + "content": "[15] Peidong Jia, Chenxuan Li, Zeyu Liu, Yichao Shen, Xingru Chen, Yuhui Yuan, Yinglin Zheng, Dong Chen, Ji Li, Xiaodong Xie, et al. Cole: A hierarchical generation framework for graphic design. arXiv preprint arXiv:2311.16974, 2023. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "text", + "content": "[16] Sanyam Lakhanpal, Shivang Chopra, Vinija Jain, Aman Chadha, and Man Luo. Refining text-to-image generation:" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "Towards accurate training-free glyph-enhanced image generation. arXiv preprint arXiv:2403.16422, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 140 + ], + "type": "text", + "content": "[17] Hyeonsu Lee and Chankyu Choi. The surprisingly straightforward scene text removal method with gated attention and region of interest generation: A comprehensive prominent model analysis. In ECCV, 2022. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 141, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 173 + ], + "type": "text", + "content": "[18] Chao Li, Chen Jiang, Xiaolong Liu, Jun Zhao, and Guoxin Wang. Joytype: A robust design for multilingual visual text creation. arXiv preprint arXiv:2409.17524, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 175, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 218 + ], + "type": "text", + "content": "[19] Minghao Li, Tengchao Lv, Jingye Chen, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, and Furu Wei. Trocr: Transformer-based optical character recognition with pre-trained models. In AAAI, 2023. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 220, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 220, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 220, + 553, + 274 + ], + "type": "text", + "content": "[20] Wenbo Li, Guohao Li, Zhibin Lan, Xue Xu, Wanru Zhuang, Jiachen Liu, Xinyan Xiao, and Jinsong Su. Empowering backbone models for visual text generation with input granularity control and glyph-aware training. arXiv preprint arXiv:2410.04439, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 276, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 553, + 308 + ], + "type": "text", + "content": "[21] Zhenhang Li, Yan Shu, Weichao Zeng, Dongbao Yang, and Yu Zhou. First creating backgrounds then rendering texts: A new paradigm for visual text blending. In ECAI, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 310, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 310, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 317, + 310, + 553, + 342 + ], + "type": "text", + "content": "[22] Minghui Liao, Guan Pang, Jing Huang, Tal Hassner, and Xiang Bai. Mask TextSpotter v3: Segmentation proposal network for robust scene text spotting. In ECCV, 2020. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 344, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 344, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 317, + 344, + 553, + 386 + ], + "type": "text", + "content": "[23] Jiawei Lin, Jiaqi Guo, Shizhao Sun, Zijiang Yang, JianGuang Lou, and Dongmei Zhang. Layoutprompter: Awaken the design ability of large language models. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 388, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 553, + 421 + ], + "type": "text", + "content": "[24] Yiqi Lin, Conghui He, Alex Jinpeng Wang, Bin Wang, Wei-jia Li, and Mike Zheng Shou. Parrot captions teach clip to spot text. In ECCV, 2024. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 422, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 465 + ], + "type": "text", + "content": "[25] Rosanne Liu, Dan Garrette, Chitwan Sahara, William Chan, Adam Roberts, Sharan Narang, Irina Blok, RJ Mical, Mohammad Norouzi, and Noah Constant. Character-aware models improve visual text rendering. In ACL, 2023. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 468, + 553, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 509 + ], + "type": "text", + "content": "[26] Zeyu Liu, Weicong Liang, Zhanhao Liang, Chong Luo, Ji Li, Gao Huang, and Yuhui Yuan. Glyph-byt5: A customized text encoder for accurate visual text rendering. In ECCV, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 512, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 553, + 555 + ], + "type": "text", + "content": "[27] Zeyu Liu, Weicong Liang, Yiming Zhao, Bohan Chen, Ji Li, and Yuhui Yuan. Glyph-byt5-v2: A strong aesthetic baseline for accurate multilingual visual text rendering. arXiv preprint arXiv:2406.10208, 2024. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 557, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 553, + 600 + ], + "type": "text", + "content": "[28] Jian Ma, Mingjun Zhao, Chen Chen, Ruichen Wang, Di Niu, Haonan Lu, and Xiaodong Lin. Glyphdraw: Seamlessly rendering text with intricate spatial structures in text-to-image generation. arXiv preprint arXiv:2303.17870, 2023. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 602, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 602, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 602, + 553, + 633 + ], + "type": "text", + "content": "[29] KhayTze Peong, Seiichi Uchida, and Daichi Haraguchi. Typographic text generation with off-the-shelf diffusion model. In ICDAR, 2024. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 635, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 679 + ], + "type": "text", + "content": "[30] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2023. 1, 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "text", + "content": "[31] Yadong Qu, Qingfeng Tan, Hongtao Xie, Jianjun Xu, Yuxin Wang, and Yongdong Zhang. Exploring stroke-level modifications for scene text editing. In AAAI, 2023. 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2753" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 130, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 294, + 162 + ], + "type": "text", + "content": "[33] Miao Rang, Zhenni Bi, Chuanjian Liu, Yunhe Wang, and Kai Han. An empirical study of scaling law forOCR. In CVPR, 2024. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 294, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 294, + 197 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 294, + 197 + ], + "type": "text", + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 199, + 294, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 199, + 294, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 199, + 294, + 232 + ], + "type": "text", + "content": "[35] Prasun Roy, Saumik Bhattacharya, Subhankar Ghosh, and Umapada Pal. STEFANN: scene text editor using font adaptive neural network. In CVPR, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 234, + 294, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 234, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 56, + 234, + 294, + 289 + ], + "type": "text", + "content": "[36] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 291, + 294, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 291, + 294, + 345 + ], + "spans": [ + { + "bbox": [ + 56, + 291, + 294, + 345 + ], + "type": "text", + "content": "[37] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In WACV, 2022. 3, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 347, + 294, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 347, + 294, + 402 + ], + "spans": [ + { + "bbox": [ + 56, + 347, + 294, + 402 + ], + "type": "text", + "content": "[38] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 437 + ], + "type": "text", + "content": "[39] Yuxiang Tuo, Wangmeng Xiang, Jun-Yan He, Yifeng Geng, and Xuansong Xie. Anytext: Multilingual visual text generation and editing. In ICLR, 2023. 2, 4, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 439, + 294, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 439, + 294, + 472 + ], + "spans": [ + { + "bbox": [ + 56, + 439, + 294, + 472 + ], + "type": "text", + "content": "[40] Aoqiang Wang, Jian Wang, Zhenyu Yan, Wenxiang Shang, Ran Lin, and Zhao Zhang. Textmaster: Universal controllable text edit. arXiv preprint arXiv:2410.09879, 2024." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 474, + 294, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 474, + 294, + 506 + ], + "spans": [ + { + "bbox": [ + 56, + 474, + 294, + 506 + ], + "type": "text", + "content": "[41] Yibin Wang, Weizhong Zhang, Jianwei Zheng, and Cheng Jin. High fidelity scene text synthesis. arXiv preprint arXiv:2405.14701, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 508, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 508, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 508, + 294, + 563 + ], + "type": "text", + "content": "[42] Jules White, Quchen Fu, Sam Hays, Michael Sandborn, Carlos Olea, Henry Gilbert, Ashraf Elnashar, Jesse Spencer-Smith, and Douglas C Schmidt. A prompt pattern catalog to enhance prompt engineering with chatgpt. arXiv preprint arXiv:2302.11382, 2023. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 565, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 565, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 565, + 294, + 597 + ], + "type": "text", + "content": "[43] Liang Wu, Chengquan Zhang, Jiaming Liu, Junyu Han, Jingtuo Liu, Errui Ding, and Xiang Bai. Editing text in the wild. In ACM MM, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 600, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 600, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 600, + 294, + 632 + ], + "type": "text", + "content": "[44] Tsung-Han Wu, Long Lian, Joseph E Gonzalez, Boyi Li, and Trevor Darrell. Self-correcting llm-controlled diffusion models. In CVPR, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 677 + ], + "type": "text", + "content": "[45] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and CUI Bin. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. In ICML, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "text", + "content": "[46] Qiangpeng Yang, Jun Huang, and Wei Lin. Swaptext: Image based texts transfer in scenes. In CVPR, pages 14700-14709, 2020. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 431 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "text", + "content": "[47] Yukang Yang, Dongnan Gui, Yuhui Yuan, Weicong Liang, Haisong Ding, Han Hu, and Kai Chen. Glyphcontrol: Glyph conditional control for visual text generation. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 553, + 162 + ], + "type": "text", + "content": "[48] Maoyuan Ye, Jing Zhang, Shanshan Zhao, Juhua Liu, Tongliang Liu, Bo Du, and Dacheng Tao. DeepSolo: Let transformer decoder with explicit points solo for text spotting. In CVPR, 2023. 4, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 315, + 163, + 553, + 206 + ], + "type": "text", + "content": "[49] Maoyuan Ye, Jing Zhang, Juhua Liu, Chenyu Liu, Baocai Yin, Cong Liu, Bo Du, and Dacheng Tao. Hi-sam: Marrying segment anything model for hierarchical text segmentation. arXiv preprint arXiv:2401.17904, 2024. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 208, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 208, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 315, + 208, + 553, + 240 + ], + "type": "text", + "content": "[50] Weichao Zeng, Yan Shu, Zhenhang Li, Dongbao Yang, and Yu Zhou. TextCtrl: Diffusion-based scene text editing with prior guidance control. In NeurIPS, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 241, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 241, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 315, + 241, + 553, + 274 + ], + "type": "text", + "content": "[51] Lingjun Zhang, Xinyuan Chen, Yaohui Wang, Yue Lu, and Yu Qiao. Brush your text: Synthesize any scene text on images via diffusion model. In AAAI, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 315, + 276, + 553, + 319 + ], + "type": "text", + "content": "[52] Qilong Zhangli, Jindong Jiang, Di Liu, Licheng Yu, Xiaoliang Dai, Ankit Ramchandani, Guan Pang, Dimitris N Metaxas, and Praveen Krishnan. Layout-agnostic scene text image synthesis with diffusion models. In CVPR, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 320, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 320, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 315, + 320, + 553, + 352 + ], + "type": "text", + "content": "[53] Yiming Zhao and Zhouhui Lian. Udifftext: A unified framework for high-quality text synthesis in arbitrary images via character-aware diffusion models. In ECCV, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 354, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 354, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 315, + 354, + 553, + 397 + ], + "type": "text", + "content": "[54] Zhen Zhao, Jingqun Tang, Binghong Wu, Chunhui Lin, Shu Wei, Hao Liu, Xin Tan, Zhizhong Zhang, Can Huang, and Yuan Xie. Harmonizing visual text comprehension and generation. In NeurIPS, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 399, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 399, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 315, + 399, + 553, + 431 + ], + "type": "text", + "content": "[55] Yuanzhi Zhu, Jiawei Liu, Feiyu Gao, Wenyu Liu, Xinggang Wang, Peng Wang, Fei Huang, Cong Yao, and Zhibo Yang. Visual text generation in the wild. In ECCV, 2024. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2754" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/U-Know-DiffPAN_ An Uncertainty-aware Knowledge Distillation Diffusion Framework with Details Enhancement for PAN-Sharpening/ca20f1f2-4ee9-4b83-80ca-6e2c71410bad_content_list.json b/2025/U-Know-DiffPAN_ An Uncertainty-aware Knowledge Distillation Diffusion Framework with Details Enhancement for PAN-Sharpening/ca20f1f2-4ee9-4b83-80ca-6e2c71410bad_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bfbdc136d7981e5d6c76ec479478808623ed7645 --- /dev/null +++ b/2025/U-Know-DiffPAN_ An Uncertainty-aware Knowledge Distillation Diffusion Framework with Details Enhancement for PAN-Sharpening/ca20f1f2-4ee9-4b83-80ca-6e2c71410bad_content_list.json @@ -0,0 +1,2029 @@ +[ + { + "type": "text", + "text": "U-Know-DiffPAN: An Uncertainty-aware Knowledge Distillation Diffusion Framework with Details Enhancement for PAN-Sharpening", + "text_level": 1, + "bbox": [ + 122, + 128, + 875, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sungpyo Kim", + "bbox": [ + 137, + 203, + 250, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 KAIST", + "bbox": [ + 161, + 224, + 228, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ksp04204@kaist.ac.kr", + "bbox": [ + 114, + 247, + 274, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jeonghyeok Do", + "bbox": [ + 344, + 203, + 470, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ KAIST", + "bbox": [ + 374, + 224, + 441, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ehwjdgur0913@kaist.ac.kr", + "bbox": [ + 312, + 247, + 503, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jaehyup Lee†", + "bbox": [ + 568, + 203, + 676, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ KNU", + "bbox": [ + 593, + 224, + 645, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jaehyuplee@knu.ac.kr", + "bbox": [ + 542, + 247, + 700, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Munchurl Kim", + "bbox": [ + 750, + 203, + 875, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1KAIST", + "bbox": [ + 776, + 224, + 844, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mkimee@kaist.ac.kr", + "bbox": [ + 738, + 247, + 880, + 258 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ab6af6b1aaff5f0f456ea32e538cf3b54298a70bd8c9b8f9355d561f7b00c523.jpg", + "image_caption": [ + "Figure 1. Visual comparison of PAN-sharpening results on the full-resolution WV3 dataset. The rightmost image shows the output of our proposed U-Know-DiffPAN framework, specifically the result produced by FSA-S (frequency selective attention student network). Notably, the proposed framework generates more detailed and robust results, particularly in high-uncertainty regions, outperforming the state-of-the-art model CANConv [14] and recent diffusion-based methods PanDiff [31] and TMDiff [56]. As highlighted in the red box, our approach successfully restores challenging highly-uncertain regions, such as cars, where other models fall short." + ], + "image_footnote": [], + "bbox": [ + 94, + 295, + 184, + 434 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/34b1b64b9497bb181f9865af225cf03a9e90f436c317c1aff15eb4414724022a.jpg", + "image_caption": [ + "PanDiff (TGRS 2023)" + ], + "image_footnote": [], + "bbox": [ + 184, + 295, + 362, + 433 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f53a2e345610728e2c67de6e7daf4d679ee27cbf2338295882bc27ab88d679f7.jpg", + "image_caption": [ + "TMDiff (TGRS 2024)" + ], + "image_footnote": [], + "bbox": [ + 364, + 295, + 542, + 433 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1a235cba609c2ce645eeafd4d9897b9d6e1f8f2644e71fce7dcdabad8081ee1d.jpg", + "image_caption": [ + "CANConv (CVPR 2024)" + ], + "image_footnote": [], + "bbox": [ + 542, + 295, + 722, + 433 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bca9a5628c4c47d655d5d2cacca7df7e1bf0a4a5036c8fcd975ce184b0c92d02.jpg", + "image_caption": [ + "U-Know-DiffPAN (Ours)" + ], + "image_footnote": [], + "bbox": [ + 723, + 295, + 903, + 433 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 529, + 326, + 545 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Conventional methods for PAN-sharpening often struggle to restore fine details due to limitations in leveraging high-frequency information. Moreover, diffusion-based approaches lack sufficient conditioning to fully utilize Panchromatic (PAN) images and low-resolution multispectral (LRMS) inputs effectively. To address these challenges, we propose an uncertainty-aware knowledge distillation diffusion framework with details enhancement for PAN-sharpening, called U-Know-DiffPAN. The U-Know-DiffPAN incorporates uncertainty-aware knowledge distillation for effective transfer of feature details from our teacher model to a student one. The teacher model in our U-Know-DiffPAN captures frequency details through frequency selective attention, facilitating accurate reverse process learning. By conditioning the encoder on compact vector representations of PAN and LRMS and the decoder on Wavelet transforms, we enable rich frequency utilization. So, the high-capacity teacher model distills frequency-rich features into a lightweight student model aided by an un", + "bbox": [ + 88, + 561, + 483, + 849 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "certainty map. From this, the teacher model can guide the student model to focus on difficult image regions for PAN-sharpening via the usage of the uncertainty map. Extensive experiments on diverse datasets demonstrate the robustness and superior performance of our U-Know-DiffPAN over very recent state-of-the-art PAN-sharpening methods. The project page is available at https://kaist-viclab.github.io/U-Know-DiffPAN-site/.", + "bbox": [ + 511, + 531, + 906, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 680, + 645, + 695 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Satellite imagery plays a vital role in a wide range of applications, including environmental monitoring, surveillance, and mapping [2, 10, 15, 29, 67]. High-resolution multispectral (HRMS) images that combine high spatial and spectral fidelity are commonly generated using PAN-sharpening techniques, fusing low-resolution multispectral (LRMS) images with high-resolution (HR) panchromatic (PAN) images. The primary objective of PAN-sharpening is to retain the spatial detail of PAN images while preserving the spectral information of LRMS images. Recently, deep learning (DL)-based approaches have gained attraction in PAN-sharpening tasks, leveraging deep networks to perform end-to-end image regression. However, DL models", + "bbox": [ + 509, + 704, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Korea Advanced Institute of Science and Technology (KAIST).", + "bbox": [ + 107, + 862, + 450, + 876 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{2}$ Kyungpook National University (KNU).", + "bbox": [ + 109, + 876, + 328, + 888 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Co-corresponding authors (equal contribution).", + "bbox": [ + 109, + 888, + 364, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23069", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "are typically limited by the local receptive fields of convolutional layers, which primarily focus on local patterns and fail to capture long-range dependencies. This often leads to overly smooth outputs with reduced high-frequency details.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome this limitation, generative models such as generative adversarial networks (GANs) [18] have been applied to PAN-sharpening, aiming to produce more realistic results. Despite their success, GANs [28, 30] pose additional challenges, including training instability, mode collapse, and artifact generation, which can degrade image quality. Recently, diffusion models [21, 38, 42, 55] have demonstrated strong performance across a variety of image generation and restoration tasks. In PAN-sharpening, diffusion-based methods [31, 56] have shown promise by using PAN and LRMS images as conditioning inputs. However, they are limited by simple conditioning strategies that do not fully utilize the complementary information provided by the PAN and LRMS inputs, thus leading to suboptimal restoration quality. Additionally, they often incur high computational costs and longer inference times compared to other DL-based methods, limiting their practical applicability.", + "bbox": [ + 89, + 154, + 482, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose U-Know-DiffPAN demonstrated in Fig. 2, a novel PAN-sharpening framework that incorporates an uncertainty-aware knowledge distillation (U-Know) strategy. Our framework is designed to fully leverage the frequency information in PAN and LRMS inputs while reducing computational costs. Our teacher model, is built to maximize information extraction from PAN and LRMS images by utilizing (i) a Feed Forward Attention (FFA) block in the encoder to capture compact vector representations of PAN and LRMS, and (ii) a High Quality Frequency Enhancement (HQFE) block in the decoder, composed of Fourier Transform Channel Attention (FTCA) and Stationary Wavelet Transform Cross Attention (SWTCA) modules. The FTCA enhances crucial frequency components through Fourier transform-based [52] attention, while the SwTCA incorporates high-frequency information from PAN and low-frequency information from LRMS into the decoder.", + "bbox": [ + 89, + 429, + 482, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our U-Know strategy leverages uncertainty-aware knowledge distillation by transferring the features from the teacher model to a light-weight student model. The teacher model denoises HRMS images along with an uncertainty map, identifying spatial regions where the prediction confidence is low. Then based of the U-Know strategy, the teacher model distills to the student model not only the frequency-aware features but also the uncertainty map to address spatial weaknesses in the teacher's predictions. This approach leads to improved performance and reduced computational cost, making our U-Know-DiffPAN more effective and efficient for PAN-sharpening. Our contributions are summarized as follows:", + "bbox": [ + 89, + 704, + 482, + 898 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We firstly incorporate the uncertainty-aware knowledge distillation into diffusion based PAN-sharpening with the guidance of pixel-wise restoration uncertainty and frequency-selective attention for details enhancement;", + "- Our teacher model leverages compact vector representations to enhance encoding efficiency and employs frequency selective attention in the decoder to refine frequency components for PAN-sharpening.", + "- We conduct extensive experiments on multiple datasets [11], including WV3, QB, and GF2, demonstrating state-of-the-art performance and surpassing very recent methods in PAN-sharpening with large margins." + ], + "bbox": [ + 513, + 90, + 903, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 290, + 661, + 306 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. PAN-Sharpening", + "text_level": 1, + "bbox": [ + 513, + 316, + 679, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "PAN-sharpening is a technique to fuse high-resolution (HR) panchromatic (PAN) images with low-resolution multispectral (LRMS) images to produce HR multispectral (HRMS) images. Traditional PAN-sharpening techniques are broadly categorized into three approaches: component substitution (CS) [8, 41], multi-resolution analysis (MRA) [1, 34], and variational optimization (VO) [3]. Modern PAN-sharpening methods primarily consist of CNN-based [14, 25, 47, 58], transformer-based [25, 68], and GAN-based [28, 30] models. LAGConv [25] introduces local-context adaptive convolutional kernels with a global harmonic bias mechanism. S2DBPN [64] uses a spatial-spectral dual back-projection network for effective PAN and LRMS fusion. DCPNet [68] establishes a transformer-based dual-task parallel framework to optimize reconstruction. CANConv [14] integrates a non-local adaptive convolution module, enhancing spatial adaptability and non-local self-similarity. Besides the above methods, there are other DL-based methods for PAN-sharpening [9, 13, 17, 22, 23, 27]. While deep learning (DL)-based methods can capture complex mappings, they encounter challenges. CNN-based approaches tend to produce overly smooth results due to local receptive field of convolution layers. GAN-based models, though capable of generating realistic textures, may suffer from instability and introduce unrealistic artifacts.", + "bbox": [ + 511, + 340, + 903, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion-based approaches. Diffusion models (DMs) [21, 42] offer a promising alternative to conventional DL-based methods for image restoration [48, 55], iteratively refining noisy images through progressive denoising. Unlike GANs [18, 26], which can be prone to instability, DMs use iterative denoising steps with flexible conditioning, making them robust for tasks such as image restoration and PAN-sharpening. Recent diffusion-based PAN-sharpening models [31, 56] have shown significant advancements. PanDiff [31], however, employs a relatively simplistic conditioning approach that limits its ability to fully utilize the information from the input data, sometimes resulting in subopti", + "bbox": [ + 511, + 719, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23070", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mal restoration quality. In contrast, TMDiff [56] employs a complex conditioning mechanism, requiring an additional pretrained CLIP [36] text encoder, which adds unnecessary complexity without substantial performance gains. Our proposed framework, U-Know-DiffPAN, handles these limitations by integrating an advanced conditioning strategy that maximizes the use of input information. Specifically, our U-Know-DiffPAN conditions the encoder on compact vector representations of PAN and LRMS, while the decoder is conditioned on the stationary wavelet transform [32] of PAN and LRMS. This approach allows our U-Know-DiffPAN to effectively capture both frequency details and spatial fidelity, achieving superior restoration quality compared to existing diffusion-based methods.", + "bbox": [ + 89, + 90, + 483, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Knowledge Distillation", + "text_level": 1, + "bbox": [ + 89, + 311, + 302, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Knowledge Distillation (KD) [20] is a well-known method for transferring knowledge from a large, model (teacher) to a smaller model (student), enabling deployment in resource-constrained settings while retaining performance. KD has demonstrated effectiveness across a range of computer vision tasks, including classification [7, 57, 65], object detection [24, 49, 60, 70], and image restoration [43, 59, 66]. However, previous KD methods, often fall short in tasks such as super-resolution [66], where capturing complicated detail is essential. This limitation arises because conventional KD student learns all regions equally, without focusing on areas that require concentrated learning. To address this, we propose uncertainty-aware knowledge distillation (U-Know), which introduces an uncertainty map into the distillation process. Our U-Know strategy informs the student of uncertain regions, areas the teacher model finds challenging to restore, enabling the student model to refine crucial details, thereby enhancing fidelity in both spatial and spectral domains.", + "bbox": [ + 89, + 333, + 483, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 89, + 633, + 191, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview of U-Know-DiffPAN", + "text_level": 1, + "bbox": [ + 89, + 657, + 361, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2 shows the overview of our U-Know-DiffPAN framework. The proposed U-Know-DiffPAN framework comprises a teacher network with frequency-selective attention (FSA-T) and a student network (FSA-S), designed within a knowledge distillation paradigm. PAN image $\\mathbf{I}_{\\mathrm{PAN}}$ and interpolated LRMS image $\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{LR}}$ are inputted in concatenation to generate HRMS image $\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{HR}}$ . Both FSA-T and FSA-S are designed to predict the residuals as:", + "bbox": [ + 89, + 680, + 483, + 801 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {0} = \\mathbf {I} _ {\\mathrm {M S}} ^ {\\mathrm {H R}} - \\mathbf {I} _ {\\mathrm {M S}} ^ {\\mathrm {L R}}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 810, + 482, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our framework operates in two main stages: (i) pre-training the FSA-T within a diffusion process to produce an initial prediction $\\widehat{\\mathbf{X}}_0$ alongside an uncertainty map $\\hat{\\theta}$ that identifies spatially weak regions; and (ii) training the FSA-S by", + "bbox": [ + 89, + 840, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Step1. Teacher (FSA-T) pre-training", + "bbox": [ + 519, + 92, + 666, + 102 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/adf6fec8ea1d84b4cf8fefc26a393936306bd5d67eed577bfb76b3431c3d3979.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 104, + 630, + 170 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/12ab21e8c633dacbb7aa7e1862747fdac4bf95ef862d5c6b2ce0a8b1afa16e0f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 104, + 874, + 171 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Step2. Student (FSA-S) training", + "bbox": [ + 519, + 172, + 648, + 184 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1d50faf5ecd3ae8177e643efa50e922af91bfa781f53f1c509cbeeb87276ac20.jpg", + "image_caption": [ + "Figure 2. Overview of our uncertainty-aware knowledge-distillation diffusion framework with details enhancement, called U-Know-DiffPAN." + ], + "image_footnote": [], + "bbox": [ + 519, + 186, + 632, + 246 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/59e338e4cde93290f2a51cd51b2a9ac6fc542f39081f58d6042213d0473a6363.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 186, + 901, + 246 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "leveraging this uncertainty map to guide the FSA-S in refining these regions through the KD. In Sec. 3.2, we describe the FSA-T's diffusion-based training process, which incorporates frequency information from PAN and LRMS inputs. Sec. 3.3 details the FSA-T architecture, specifically designed to enhance frequency components. Finally, in Sec. 3.4, we explain the FSA-S's uncertainty-aware knowledge distillation framework, which utilizes the uncertainty map to achieve precise and reliable PAN-sharpening.", + "bbox": [ + 511, + 310, + 905, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Diffusion Process of Teacher Network (FSA-T)", + "text_level": 1, + "bbox": [ + 511, + 454, + 905, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models operate by defining a forward process that gradually adds noise to data until it becomes indistinguishable from random noise, and a reverse process that learns to denoise the data, reconstructing the original information. We leverage this framework into a frequency-selective attention teacher network (FSA-T) as our teacher network, to generate frequency enhanced HRMS images with their pixel-wise uncertainty information.", + "bbox": [ + 511, + 476, + 905, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Forward process. In the forward diffusion process, noise is progressively added to the residual image, $\\mathbf{X}_0$ , until it converges to a Gaussian distribution, $\\mathbf{X}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ , after $T$ timesteps. At each timestep $t \\in \\{1, \\dots, T\\}$ , a noisy version of the residual is obtained as follows [21]:", + "bbox": [ + 511, + 598, + 905, + 674 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {X} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 681, + 903, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ is Gaussian noise, and $\\bar{\\alpha}_t = \\prod_{s=1}^t (1 - \\beta_s)$ controls the noise level at step $t$ .", + "bbox": [ + 511, + 707, + 903, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reverse process. Following the DDPM [21], the reverse process aims to recover $\\mathbf{X}_0$ from $\\mathbf{X}_T$ by predicting a denoised image at each step. To perform denoising at each timestep, the denoising network (FSA-T), denoted as $\\Psi$ , is designed to predict the original $\\mathbf{X}_0$ rather than the added noise $\\epsilon$ , as we found this empirically to be beneficial for PAN-sharpening. Additionally, $\\Psi$ simultaneously produces an uncertainty map $\\hat{\\theta}$ for each prediction, providing pixelwise confidence information as:", + "bbox": [ + 511, + 739, + 905, + 873 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\widehat {\\mathbf {X}} _ {0} \\mid \\widehat {\\boldsymbol {\\theta}} \\right] = \\Psi \\left(\\left[ \\mathbf {X} _ {t} \\mid \\mathbf {I} _ {\\text {P A N}} \\mid \\mathbf {I} _ {\\mathrm {M S}} ^ {\\mathrm {L R}} \\right]; \\mathbf {v}, \\mathcal {S} - \\text {C o n d}; t\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 878, + 903, + 905 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23071", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a2a0e0bcf4075dbb7ba1c9fcdfa5dc1651fe7f943632525bd546219e47e12585.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 89, + 256, + 267 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/65a9812a2f6b025a690644bfb9952031ec187f6f7a90b6433f6aa4d6542c8409.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 259, + 89, + 759, + 268 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5143c06fceb3b4937c9e978678ee64ec052fc45b53632a817a8f7e0c79468fcc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 93, + 901, + 266 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/751d94a4da0449bbcb18709c4d3f90f85e7447266244962bf47db8460c990d8d.jpg", + "image_caption": [ + "Figure 3. Architecture of our proposed teacher model with frequency-selective attention, denoted as FSA-T, for PAN-sharpening. The FSA-T is designed to fully utilize frequency information for details enhancement from PAN and LRMS inputs (more details in Supplemental.)" + ], + "image_footnote": [], + "bbox": [ + 94, + 270, + 256, + 345 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/efa7c75e788eea5ee898f43fe4618f5b69483a314bdabbcaaeb018db2c433391.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 259, + 273, + 480, + 344 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7dadd3b623aa519e83469df927456b6ccf6079903df4a319907b20ccd46a8687.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 271, + 679, + 345 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f565785ec6e442234fbbb45b8791270223e45b9d510b80aa370ad2837cc31b2d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 691, + 271, + 901, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\left[\\cdot \\mid \\cdot \\right]$ indicates channel-wise concatenation, $\\widehat{\\mathbf{X}}_0$ is a predicted residual, $\\hat{\\theta}$ is a predicted uncertainty map, $\\mathbf{v}$ and $S$ -Cond represent a learning compact representation (vector) and Stationary Wavelet Transform (SWT)-based conditioning terms, respectively. This formulation allows the FSA-T to iteratively refine the noisy input, guided by the conditioned information ( $\\mathbf{v}$ , $S$ -Cond), toward an accurate reconstruction of the $\\mathbf{X}_0$ . $\\hat{\\theta}$ captures pixel-wise prediction confidence, which guides the FSA-S to more rigorously learn the regions with high uncertainty. To ensure positivity, we apply the SoftPlus [69] activation to $\\hat{\\theta}$ .", + "bbox": [ + 88, + 393, + 482, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Uncertainty-driven diffusion loss. Inspired by [33, 39], we employ the following uncertainty-driven diffusion loss function, $\\mathcal{L}_{\\mathrm{U - Diff}}$ , to optimize the teacher network $\\Psi$ :", + "bbox": [ + 89, + 561, + 483, + 608 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {U - D i f f}} = \\left\\| \\frac {1}{2 \\hat {\\boldsymbol {\\theta}}} \\odot \\left| \\widehat {\\mathbf {X}} _ {0} - \\mathbf {X} _ {0} \\right| + \\frac {1}{2} \\log \\hat {\\boldsymbol {\\theta}} \\right\\| _ {1}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 614, + 482, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\theta}$ serves as the estimated variance term and is regarded as the uncertainty map in our framework. This objective function encourages the model to minimize prediction errors in low-confidence regions, thus yielding a more accurate and spatially robust output.", + "bbox": [ + 89, + 657, + 483, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Details of Teacher Network Architecture", + "text_level": 1, + "bbox": [ + 89, + 742, + 437, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our teacher network, FSA-T, is designed to enhance frequency details for PAN-sharpening tasks by fully utilizing complementary information from PAN and LRMS inputs. The FSA-T adopts an encoder-decoder (U-Net [37]) structure, where each stage is specifically tailored to capture and refine critical spatial and frequency features. The encoder of FSA-T employs Feed Forward Attention (FFA) blocks to extract compact and informative representations from the PAN and LRMS inputs. This attention mechanism allows", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the encoder to focus on essential features, preparing the data for effective frequency enhancement in subsequent layers.", + "bbox": [ + 511, + 393, + 903, + 424 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The decoder of FSA-T integrates High-Quality Frequency Enhancement (HQFE) blocks at each decoding level to further refine frequency components. Each HQFE block comprises two key modules: Fourier Transform Channel Attention (FTCA) that emphasizes important frequency components via Fourier Transform based attention, and Stationary Wavelet Transform Cross Attention (SWTCA) that merges high-frequency information from PAN and low-frequency content from LRMS, both of which are decomposed by SwT [32]. By incorporating these specialize components, FSA-T effectively produces $\\widehat{\\mathbf{X}}_0$ with enhanced spatial fidelity, establishing an effective uncertainty-aware knowledge distillation (U-Know) strategy.", + "bbox": [ + 511, + 425, + 905, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Feed Forward Attention (FFA). The FFA block in Fig. 3 is designed to leverage compact information [55] from the PAN and LRMS input to establish a rough structural representation during encoding. A compact vector representation $\\mathbf{v} \\in \\mathbb{R}^{1 \\times D}$ is obtained by a vector extractor $(\\mathcal{E})$ as:", + "bbox": [ + 511, + 623, + 905, + 699 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} = \\mathcal {E} \\left(\\mathbf {I} _ {\\mathrm {P A N}}, \\mathbf {I} _ {\\mathrm {M S}} ^ {\\mathrm {L R}}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 710, + 903, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathbf{v}$ is then used as a dynamic modulation [26, 50] parameter within the FFA, enabling $\\mathcal{E}$ to incorporate $\\mathbf{v}$ and establish an approximate structure during restoration as follows:", + "bbox": [ + 511, + 742, + 903, + 787 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n[ \\boldsymbol {\\beta} _ {n} \\mid \\gamma_ {n} ] = \\operatorname {L i n e a r} (\\mathbf {v}),\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 800, + 723, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {n} ^ {\\prime} = \\gamma_ {n} \\odot \\mathrm {G N} \\left(\\mathbf {f} _ {n}\\right) + \\beta_ {n}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 819, + 903, + 835 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {n} \\leftarrow \\operatorname {C o n v} \\left(\\mathbf {f} _ {n} ^ {\\prime}\\right) \\odot \\operatorname {G E L U} \\left(\\operatorname {C o n v} \\left(\\mathbf {f} _ {n} ^ {\\prime}\\right)\\right) + \\mathbf {f} _ {n},\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 838, + 849, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{f}_n\\in \\mathbb{R}^{C\\times h\\times w}$ denotes input feature for the $n$ th encoder, Linear indicates Linear operation, $\\odot$ denotes", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23072", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "element-wise multiplication, GN indicates group normalization [54], GELU indicates GELU [19] activation and Conv consists of a sequential $1 \\times 1$ convolution and a $3 \\times 3$ depth-wise convolution layer (Details of $\\mathcal{E}$ is in the Supplemental.).", + "bbox": [ + 89, + 90, + 480, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fourier Transform Channel Attention (FTCA). The FTCA block in Fig. 3 performs self-attention in frequency domain by leveraging the 2D Fast Fourier Transform (FFT), denoted as $\\mathcal{F}$ . $\\mathcal{F}$ allows for efficient identification and emphasis of critical frequency components. For the $n$ -th decoder block in Fig. 3, the input feature $\\mathbf{f}_n \\in \\mathbb{R}^{C \\times h \\times w}$ of $C$ channels is transformed into query $(\\mathbf{Q})$ , key $(\\mathbf{K})$ , and value $(\\mathbf{V})$ representations. The 2D FFT $\\mathcal{F}$ is then applied independently to each channel of $\\mathbf{Q}$ , $\\mathbf{K}$ , and $\\mathbf{V}$ , mapping them transform from time domain to frequency domain as:", + "bbox": [ + 89, + 167, + 482, + 318 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n[ \\mathbf {Q} \\mid \\mathbf {K} \\mid \\mathbf {V} ] = \\mathcal {F} (\\operatorname {C o n v} \\left(\\mathbf {f} _ {n}\\right)), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 330, + 482, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{Q},\\mathbf{K}$ and $\\mathbf{V}$ reside in $\\mathbb{C}^{C\\times h\\times (\\frac{w}{2} +1)}$ due to the Hermitian symmetry property [6]. Note that each frequency domain representation of $\\mathbf{Q},\\mathbf{K}$ , and $\\mathbf{V}$ includes real and imaginary parts. So, we decompose $\\mathbf{Q},\\mathbf{K}$ , and $\\mathbf{V}$ into real and imaginary parts as:", + "bbox": [ + 89, + 359, + 483, + 436 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\mathbf {Q} _ {\\mathbb {R}} \\mid \\mathbf {K} _ {\\mathbb {R}} \\mid \\mathbf {V} _ {\\mathbb {R}} \\right] + j \\left[ \\mathbf {Q} _ {\\mathbb {I}} \\mid \\mathbf {K} _ {\\mathbb {I}} \\mid \\mathbf {V} _ {\\mathbb {I}} \\right] = \\left[ \\mathbf {Q} \\mid \\mathbf {K} \\mid \\mathbf {V} \\right], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 448, + 482, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $j = \\sqrt{-1}$ , and $\\mathbf{Q}_{\\mathbb{R}}$ , $\\mathbf{K}_{\\mathbb{R}}$ , $\\mathbf{V}_{\\mathbb{R}}$ , $\\mathbf{Q}_{\\mathbb{I}}$ , $\\mathbf{K}_{\\mathbb{I}}$ , $\\mathbf{V}_{\\mathbb{I}}$ are in $\\mathbb{R}^{C\\times h\\times (\\frac{w}{2} + 1)}$ . Self-attention operations are then performed separately on the real and imaginary parts as follows:", + "bbox": [ + 89, + 477, + 483, + 523 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {l} = \\operatorname {S o f t M a x} \\left(\\mathbf {Q} _ {l} \\mathbf {K} _ {l} ^ {\\top} / \\sqrt {C}\\right) \\mathbf {V} _ {l}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 535, + 482, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $l \\in \\{\\mathbb{R},\\mathbb{I}\\}$ indicates the real or imaginary feature, respectively. Finally, the inverse Fourier Transform $\\mathcal{F}^{-1}$ is applied to $\\mathbf{f}_{\\mathbb{R}}$ and $\\mathbf{f}_{\\mathbb{I}}$ to map them back to the time domain:", + "bbox": [ + 89, + 568, + 483, + 614 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {n} \\leftarrow \\mathcal {F} ^ {- 1} \\left(\\mathbf {f} _ {\\mathbb {R}} + j \\mathbf {f} _ {\\mathbb {I}}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 626, + 482, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This process ensures that refined frequency information be selectively captured and utilized.", + "bbox": [ + 89, + 656, + 482, + 685 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stationary Wavelet Transform Cross Attention (SWTCA). The SwTCA block is designed to leverage Wavelet components of $\\mathbf{I}_{\\mathrm{PAN}}$ and $\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{LR}}$ motivated by [5, 35, 40, 71]. As shown in Fig. 3, it enhances the features outputted by the FTCA block, utilizing the Stationary Wavelet Transform (SWT) [32], denoted as $S$ . The SwTCA block selectively incorporates relevant frequency components from $\\mathbf{I}_{\\mathrm{PAN}}$ and $\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{LR}}$ through a two-stage cross-attention process. The SwT-based conditioning terms ( $S$ -Cond) are constructed by decomposing $\\mathbf{I}_{\\mathrm{PAN}}$ and $\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{LR}}$ using $S$ into four components:", + "bbox": [ + 89, + 686, + 483, + 853 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{r l} & {\\left[ \\mathbf {L} _ {\\mathrm {P A N}} \\mid \\mathbf {H} _ {\\mathrm {P A N}} \\mid \\mathbf {V} _ {\\mathrm {P A N}} \\mid \\mathbf {D} _ {\\mathrm {P A N}} \\right] = \\mathcal {S} (\\mathbf {I} _ {\\mathrm {P A N}}),} \\\\ & {\\left[ \\mathbf {L} _ {\\mathrm {M S}} ^ {\\mathrm {L R}} \\mid \\mathbf {H} _ {\\mathrm {M S}} ^ {\\mathrm {L R}} \\mid \\mathbf {V} _ {\\mathrm {M S}} ^ {\\mathrm {L R}} \\mid \\mathbf {D} _ {\\mathrm {M S}} ^ {\\mathrm {L R}} \\right] = \\mathcal {S} (\\mathbf {I} _ {\\mathrm {M S}} ^ {\\mathrm {L R}}),} \\end{array} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 864, + 482, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{L}$ represent the low-frequency approximation components, while $\\mathbf{H},\\mathbf{V},\\mathbf{D}$ denote the high-frequency horizontal, vertical, and diagonal details, respectively. Since $\\mathbf{I}_{\\mathrm{PAN}}$ provides high spatial resolution but lacks spectral information, only its high-frequency components $\\mathbf{H}_{\\mathrm{PAN}},$ $\\mathbf{V}_{\\mathrm{PAN}}$ ,and $\\mathbf{D}_{\\mathrm{PAN}}$ are used. Conversely, $\\mathbf{L}_{\\mathrm{MS}}^{\\mathrm{LR}}$ from $\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{LR}}$ contains richer spectral information. These components are concatenated channel-wise to form $S$ -Cond as:", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} - \\text {C o n d} = \\left[ \\mathbf {L} _ {\\mathrm {M S}} ^ {\\mathrm {L R}} \\mid \\mathbf {H} _ {\\mathrm {P A N}} \\mid \\mathbf {V} _ {\\mathrm {P A N}} \\mid \\mathbf {D} _ {\\mathrm {P A N}} \\right]. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 220, + 903, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To inject $S$ -Cond into the feature $\\mathbf{f}_n$ obtained from FTCA at the $n$ -th decoder block, we employ a two-stage cross-attention process. In the first stage, $\\mathbf{Q}_1$ and $\\mathbf{K}_1$ are derived from $S$ -Cond, directing attention toward critical frequency components, while $\\mathbf{V}_1$ is derived from $\\mathbf{f}_n$ . The resulting features are reshaped to $C \\times hw$ for channel attention [63], producing an intermediate feature $\\mathbf{f}_n^1 \\in \\mathbb{R}^{C \\times h \\times w}$ as:", + "bbox": [ + 511, + 246, + 906, + 353 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\mathbf {Q} _ {1} \\mid \\mathbf {K} _ {1} \\right] = \\operatorname {C o n v} (S - \\text {C o n d}), \\mathbf {V} _ {1} = \\operatorname {C o n v} (\\mathbf {f} _ {n}),\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 373, + 852, + 390 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {n} ^ {1} = \\operatorname {S o f t M a x} \\left(\\mathbf {Q} _ {1} \\mathbf {K} _ {1} ^ {\\top} / \\sqrt {C}\\right) \\mathbf {V} _ {1}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 390, + 903, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{Q}_1, \\mathbf{K}_1, \\mathbf{V}_1$ are in $\\mathbb{R}^{C \\times hw}$ . In the second stage, $\\mathbf{Q}_2$ is derived from the intermediate feature $\\mathbf{f}_n^1$ , while $\\mathbf{K}_2$ and $\\mathbf{V}_2$ come from $S$ -Cond, enabling the model to refine the correlation between the reconstructed feature $\\mathbf{f}_n$ and significant frequency components in $S$ -Cond. This stage yields the final feature $\\mathbf{f}_n \\in \\mathbb{R}^{C \\times h \\times w}$ after reshaping:", + "bbox": [ + 511, + 428, + 905, + 518 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {2} = \\operatorname {C o n v} \\left(\\mathbf {f} _ {n} ^ {1}\\right), \\left[ \\mathbf {K} _ {2} \\mid \\mathbf {V} _ {2} \\right] = \\operatorname {C o n v} (S - C o n d),\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 540, + 852, + 556 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {n} \\leftarrow \\operatorname {S o f t M a x} \\left(\\mathbf {Q} _ {2} \\mathbf {K} _ {2} ^ {\\top} / \\sqrt {C}\\right) \\mathbf {V} _ {2}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 556, + 903, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Through these two stages of cross-attention, the final feature $\\mathbf{f}_n$ effectively incorporates essential frequency components from both PAN and MS inputs, ensuring enhanced frequency representation and high-quality HRMS output.", + "bbox": [ + 511, + 593, + 905, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Uncertainty-Aware Knowledge Distillation", + "text_level": 1, + "bbox": [ + 511, + 662, + 875, + 678 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address the substantial computational demands of the teacher (FSA-T) $\\Psi$ , we propose an novel uncertainty-aware knowledge distillation (U-Know) strategy.", + "bbox": [ + 511, + 684, + 905, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Student network (FSA-S) $\\psi$ . The student model FSA-S is designed with the same number of encoder and decoder blocks as FSA-T. Each block consists solely of Res-blocks, and unlike the teacher, the student operates without any additional conditional inputs. As shown in Fig. 2, the lightweight student model $\\psi$ predicts the denoised image $\\widetilde{\\mathbf{X}}_0$ from the input $\\mathbf{X}_t$ at timestep $t$ as:", + "bbox": [ + 511, + 729, + 905, + 834 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {\\mathbf {X}} _ {0} = \\psi \\left(\\left[ \\mathbf {X} _ {t} \\mid \\mathbf {I} _ {\\text {P A N}} \\mid \\mathbf {I} _ {\\mathrm {M S}} ^ {\\mathrm {L R}} \\right]; t\\right). \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 840, + 903, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "U-Know loss function. FSA-T $\\Psi$ generates an uncertainty map $\\hat{\\theta}$ along with the output $\\bar{\\mathbf{X}}_0$ during training.", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23073", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "These two outputs are frozen and used in the U-Know loss function, $\\mathcal{L}_{\\mathrm{U - Know}}$ , to train the FSA-S $\\psi$ . The total training objective combines a hard loss $(\\mathcal{L}_{\\mathrm{hard}})$ , a soft loss $(\\mathcal{L}_{\\mathrm{soft}})$ , and a feature loss $(\\mathcal{L}_{\\mathrm{feat}})$ to distill frequency-rich features and enhance regions where $\\Psi$ exhibits spatial weaknesses points. The total loss is defined as:", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {U - K n o w}} = \\mathcal {L} _ {\\text {h a r d}} + \\lambda_ {s} \\mathcal {L} _ {\\text {s o f t}} + \\lambda_ {f} \\mathcal {L} _ {\\text {f e a t}}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 193, + 482, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda_{s}$ and $\\lambda_{f}$ control the contributions of $\\mathcal{L}_{\\mathrm{soft}}$ and $\\mathcal{L}_{\\mathrm{feat}}$ respectively. $\\mathcal{L}_{\\mathrm{hard}}$ measures the difference between $\\psi$ output $\\widetilde{\\mathbf{X}}_0$ and the ground truth $\\mathbf{X}_0$ , weighted by $\\Psi$ 's uncertainty map $\\hat{\\theta}$ to emphasize challenging regions:", + "bbox": [ + 89, + 220, + 482, + 281 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {h a r d}} = \\left\\| (\\tau + \\hat {\\boldsymbol {\\theta}}) \\odot \\left| \\widetilde {\\mathbf {X}} _ {0} - \\mathbf {X} _ {0} \\right| \\right\\| _ {1}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 290, + 482, + 319 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\tau$ is a hyper-parameter. $\\mathcal{L}_{\\mathrm{soft}}$ computes the difference between $\\psi$ 's output $\\widetilde{\\mathbf{X}}_0$ and $\\Psi$ 's output $\\widehat{\\mathbf{X}}_0$ , weighted by the $\\hat{\\theta}$ to help $\\psi$ better approximate the $\\Psi$ 's knowledge:", + "bbox": [ + 89, + 327, + 483, + 372 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s o f t}} = \\left\\| (\\tau - \\hat {\\boldsymbol {\\theta}}) \\odot \\left| \\widetilde {\\mathbf {X}} _ {0} - \\widehat {\\mathbf {X}} _ {0} \\right| \\right\\| _ {1}. \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 381, + 482, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the regions where $\\hat{\\pmb{\\theta}}$ values are high, indicating that $\\Psi$ struggles with the task, we apply a weight of $(\\tau + \\hat{\\pmb{\\theta}})$ and use $\\mathbf{X}_0$ as the hard ground truth to guide $\\psi$ toward the true target. Conversely, in the areas where $\\hat{\\pmb{\\theta}}$ values are low, indicating that $\\Psi$ performs well, we apply a weight of $(\\tau - \\hat{\\pmb{\\theta}})$ and use $\\Psi$ 's output $\\widehat{\\mathbf{X}}_0$ as the soft ground truth, allowing $\\psi$ to approximate $\\Psi$ 's knowledge in these regions. These dual loss formulations enable $\\Psi$ to focus on challenging areas while effectively leveraging the $\\Psi$ 's strengths. $\\mathcal{L}_{\\mathrm{feat}}$ ensures alignment between intermediate features of $\\Psi$ and $\\psi$ as [4]:", + "bbox": [ + 89, + 421, + 483, + 575 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\text {f e a t}} ^ {i} = \\sqrt {\\left\\| \\mathbf {f} _ {i} - \\mathbf {s} _ {i} \\right\\| _ {1} ^ {2} + \\gamma}, \\quad \\mathcal {L} _ {\\text {f e a t}} = \\sum_ {i} \\alpha_ {i} \\ell_ {\\text {f e a t}} ^ {i}, \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 585, + 482, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\ell_{\\mathrm{feat}}^{i}$ is a feature loss for the $i$ -th feature, $\\alpha_{i}$ is a weight for the $i$ -th feature, and $\\mathbf{f}_i$ and $\\mathbf{s}_i$ denote the $i$ -th intermediate feature maps of $\\Psi$ and $\\psi$ respectively. The small constant $\\gamma$ ensures smooth convergence during optimization.", + "bbox": [ + 89, + 623, + 483, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 696, + 215, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets, Metrics and Training Details", + "text_level": 1, + "bbox": [ + 89, + 720, + 421, + 738 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a43a7358a918bb182e443964aee5cc75d390562015300b9a6be4e8f5b3f60841.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SatelliteWorldView-3QuickBirdGaoFen-2
Number of Band844
Spatial Resolution (m)PAN0.30.60.8
LRMS1.22.43.2
Radiometric Resolution (bit)111110
Number of (Train / Test) Images9,714 / 2017,139 / 2019,809 / 20
Patch SizePAN64×64×164×64×164×64×1
LRMS16×16×816×16×416×16×4
", + "bbox": [ + 99, + 755, + 480, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Detailed information of Worldview-3, QuickBird, and GaoFen-2 datasets.", + "bbox": [ + 89, + 864, + 482, + 891 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 summarizes the characteristics of three datasets from PanCollection [11] for our experiments: WorldView-3 (WV3), QuickBird (QB), and GaoFen-2 (GF2). All experiments were conducted with a batch size of 32 for 300K iterations per dataset on an NVIDIA GeForce RTX 3090 GPU. The initial learning rate was set to $10^{-4}$ with a decay factor of 0.5 applied every 10K iterations, using the AdamW optimizer [12]. For the diffusion process of our FSA-T and FSA-S, the total diffusion time step $T$ was set to 500, with the DDIM [42] sampling strategy used to generate HRMS images in 25 steps. To evaluate the model performance, we employed widely accepted metrics that capture various aspects of image quality and similarity. For the reduced-resolution datasets of WV3, QB, and GF2, we used six standard metrics: PSNR [51], SSIM [51], SAM [62], ERGAS [46], SCC [16], and Q4/Q8 [44]. For the full-resolution datasets of WV3, QB, and GF2, we used HQNR [45] with $D_{\\lambda}$ , and $D_{S}$ to comprehensively assess the results.", + "bbox": [ + 511, + 90, + 906, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 378, + 714, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To ensure a thorough and equitable evaluation, we used the official codes for each compared method whenever available. For methods without official codes, we re-implemented them based on the papers [31, 56] to maintain fairness in comparison.", + "bbox": [ + 511, + 402, + 906, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative analysis. As shown in Table 2 and Table 3, our FSA-T achieves superior performance on the WV3, QB, and GF2 datasets compared to all other methods. The GF2 dataset contains relatively easier (less complex) images, so the teacher model can fully leverage its high-capacity features. In contrast, for the more challenging (more complex) WV3 and QB datasets, our FSA-S surpasses FSA-T across several metrics. This improvement represents the effectiveness of our uncertainty-aware knowledge distillation (U-Know) strategy, which enables the student model (FSA-S) to address the spatial weaknesses of the teacher model by focusing on uncertain regions. From this, the FSA-S can benefit more generalization from the knowledge distillation with uncertainty maps, leading to better performances than the FSA-T for the test datasets. Consequently, our FSA-S demonstrates an enhanced adaptability and robustness on complex datasets, where precise handling of spatial details (high frequency components) is essential.", + "bbox": [ + 511, + 479, + 908, + 751 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative analysis. Fig. 4 shows the output HRMS images $\\hat{\\mathbf{I}}_{\\mathrm{MS}}^{\\mathrm{HR}}$ in the first rows and the error maps (MAE images) between $\\hat{\\mathbf{I}}_{\\mathrm{MS}}^{\\mathrm{HR}}$ and ground truth $\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{HR}}$ . As shown, our framework generates the results $\\hat{\\mathbf{I}}_{\\mathrm{MS}}^{\\mathrm{HR}}$ closer to $\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{HR}}$ in the reduced resolution GF2 data set. Furthermore, Fig. 1 demonstrates that even in the full-resolution comparison where ground truth is unavailable, our framework enables to generate more detailed and stable results compared to all other state-of-the-art approaches.", + "bbox": [ + 511, + 752, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23074", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5aebfec8f96cd6fd21a2b67324d157abdfbe4af1af70fa2111c16cd78b7c3b1e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
GF2 DatasetReduced-ResolutionFull-Resolution
TypesMethodsPublicationsPSNR↑SSIM↑SAM↓ERGAS↓SCC↑Q4↑Dλ↓Ds↓HQNR↑
Non-DMsPanNet [58]ICCV 201739.1970.9591.0501.0380.9750.9630.0200.0520.929
MSDCNN [61]JSTARS 201840.7300.9710.9460.8620.9830.9720.0260.0790.898
FusionNet [53]ICCV 202139.8660.9660.9710.9600.9800.9670.0340.1050.865
LAGConv [25]AAAI 202241.1470.9740.8860.8160.9850.9740.0300.0780.895
S2DBPN [64]TGRS 202342.6860.9800.7720.6860.9900.9810.0200.0460.935
DCPNet [68]TGRS 202442.3120.9790.8060.7240.9880.9800.0240.0240.953
CANConv [14]CVPR 202443.1660.9820.7220.6530.9910.9830.0190.0630.919
Diffusion ModelsPanDiff [31]TGRS 202342.8270.9800.7670.6740.9900.9810.0200.0450.936
TMDiff [56]TGRS 202441.8960.9770.7640.7540.9880.9790.0290.0300.942
FSA-SOurs44.5850.9860.6240.5480.9930.9870.0180.0370.944
FSA-TOurs44.7570.9880.6030.5370.9940.9880.0170.0290.953
", + "bbox": [ + 99, + 88, + 901, + 268 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9ffbd1959f59ea5589663053d82b40ff031fc0ae5076fb74b43214f03e57bfc7.jpg", + "table_caption": [ + "Table 2. Comparison of different models on the GaoFen-2 (GF2) dataset. Blue indicates the second-best performance, while red highlights the best-performing model. Standard deviations of the metrics across test samples are provided in the Supplemental." + ], + "table_footnote": [], + "table_body": "
TypesMethodsWV3 Dataset (Reduced-Resolution)QB Dataset (Reduced-Resolution)
PSNR↑SSIM↑SAM↓ERGAS↓SCC↑Q8↑PSNR↑SSIM↑SAM↓ERGAS↓SCC↑Q4↑
Non-DMsPanNet [58]36.1480.9663.4022.5380.9790.91335.5630.9395.2734.8560.9660.911
MSDCNN [61]36.3290.9673.3002.4890.9790.91437.0400.9544.8284.0740.9770.925
FusionNet [53]36.5690.9683.1882.4280.9810.91636.8210.9524.8924.1830.9750.923
LAGConv [25]36.7320.9703.1532.3800.9810.91637.5650.9584.6823.8450.9800.930
S2DBPN [64]37.2160.9723.0192.2450.9850.91737.3140.9564.8493.9560.9800.928
DCPNet [68]37.0090.9723.0832.3010.9840.91538.0790.9634.4203.6180.9830.935
CANConv [14]37.4410.9732.9272.1630.9850.91837.7950.9604.5543.7400.9820.935
Diffusion ModelsPanDiff [31]37.0290.9713.0582.2760.9840.91337.8420.9594.6113.7230.9820.935
TMDiff [56]37.4770.9732.8852.1510.9860.91537.6420.9584.6273.8040.9810.930
FSA-S37.9300.9762.7972.0460.9880.92238.3610.9644.3373.5000.9840.938
FSA-T37.8940.9762.8012.0550.9870.92138.3430.9644.3493.5020.9850.938
", + "bbox": [ + 99, + 313, + 901, + 478 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/89a59f39da9ed831a7d68e9e447c5fd4358c7edd9c0f3eb790a8d2b40f8e9a1e.jpg", + "image_caption": [ + "Figure 4. Visual comparison of PAN-sharpening results on the reduced-resolution GF2 dataset. The first row shows RGB of outputs $\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{HR}}$ and the second row displays Error Map, the difference between output $\\hat{\\mathbf{I}}_{\\mathrm{MS}}^{\\mathrm{HR}}$ and ground truth $\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{HR}}$ . Both FSA-T and FSA-S achieve more detailed results compared to state-of-the-art models." + ], + "image_footnote": [], + "bbox": [ + 93, + 523, + 903, + 719 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/29ec0caa76aea8560757a3939ec0b5df7c4b80de311f6b007f9a87d4370271d6.jpg", + "table_caption": [ + "Table 3. Performance comparison of different models on WV3 and QB datasets. Red and blue highlights the best- and 2nd-best performing models. The results for the full-resolution datasets and detailed results with standard deviations are provided in the Supplemental." + ], + "table_footnote": [], + "table_body": "
MethodsParams. (M)FLOPs (T)Time (s)Memory (GB)
PanDiff [31]12.5560.47119.5223.260
TMDiff [56]153.9395.51767.46110.483
FSA-T25.4921.40225.4955.910
FSA-S9.1150.34612.2872.136
", + "bbox": [ + 125, + 782, + 454, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Computational complexity comparison for diffusion-based models. Best values are highlighted in red.", + "bbox": [ + 89, + 851, + 482, + 878 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Computational complexity analysis. Table 4 compares the computational cost of diffusion-based models, detailing parameters (Params.), floating-point operations (FLOPs), inference time (Time) and memory usage. This demonstrates the efficiency of our FSA-S over other models. In particular, the TMDiff has significantly higher numbers of parameters and FLOPs. The FSA-T, while computationally", + "bbox": [ + 511, + 785, + 906, + 893 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23075", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "intensive than the FSA-S, still provides a balanced trade-off between efficiency and performance.", + "bbox": [ + 89, + 90, + 485, + 122 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 132, + 256, + 148 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/419ae6fa61c9328953033815efb5e94c57c0cb2454f4265898c0e5e772a63e4f.jpg", + "image_caption": [ + "Figure 5. Visualization of uncertainty map $\\hat{\\theta}$ , Error Map, and ground truth $\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{HR}}$ of reduced WV3." + ], + "image_footnote": [], + "bbox": [ + 94, + 166, + 483, + 339 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of uncertainty map $\\hat{\\theta}$ of $\\Psi$ . The uncertainty map $\\hat{\\theta}$ learned through Eq. 4 shows that the areas with low uncertainty appear dark, while the areas with high uncertainty appear bright. Fig. 5 visualizes the uncertainty maps $\\hat{\\theta}$ , error maps and ground truth $\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{HR}}$ for the reduced WV3 dataset. As shown in Fig. 5, it can be noted that the high-uncertainty regions mainly correspond to high-frequency areas such as challenging object edges. The learned uncertainty maps tend to resemble the error maps.", + "bbox": [ + 89, + 388, + 483, + 526 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/310414b3c0dcfb7ae2eb96b3b390262b0468a2793b736d27327cacc9fcde5519.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
EncoderDecoderGF2 Dataset (Reduced-Resolution)
FFAHQFESAM↓ERGAS↓SCC↑Q4↑
0.654 ± 0.1120.661 ± 0.0760.992 ± 0.0010.986 ± 0.007
✓0.654 ± 0.1120.636 ± 0.0770.993 ± 0.0020.986 ± 0.007
✓0.617 ± 0.1170.556 ± 0.1040.993 ± 0.0020.987 ± 0.007
✓✓0.603 ± 0.1020.537 ± 0.0770.994 ± 0.0010.988 ± 0.006
", + "bbox": [ + 96, + 537, + 480, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of FFA, HQFE blocks in FSA-T. Table 5 presents the effectiveness of FFA and HQFE. As noted, there are a significant improvement across all metrics. The performance gains are much greater than when FFA is only applied to the encoder or HQFE only to the decoder, as simultaneous condition injection creates a synergistic effect with both FFA and HQFE because FFA sets up initial styles or \"sketches\" for $\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{HR}}$ images, followed by the decoder that restores frequency details via the HQFE. This step-by-step learning greatly improves overall performance.", + "bbox": [ + 89, + 654, + 482, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of the U-Know loss. To verify the effectiveness of our uncertainty-aware knowledge distillation loss $\\mathcal{L}_{\\mathrm{U - Know}}$ (U-Know loss), we trained the student model (FSA-S) using three losses: a widely used $L_{1}$ , a conventional knowledge distillation loss $\\mathcal{L}_{\\mathrm{KD}}$ which is same from our method $\\mathcal{L}_{\\mathrm{U - Know}}$ but without considering the uncertainty", + "bbox": [ + 89, + 810, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c84454788afe5523ce9e9764149407bb6b2099507dc493b37208901f8f2f6415.jpg", + "table_caption": [ + "Table 5. Comparison of Results with and without FFA and HQFE Blocks in FSA-T. Best values are highlighted in red." + ], + "table_footnote": [], + "table_body": "
Loss func.GF2 Dataset (Full-Resolution)
Dλ↓Ds↓HQNR ↑
L10.026 ± 0.0140.040 ± 0.0170.935 ± 0.020
LKD0.025 ± 0.0150.038 ± 0.0160.938 ± 0.021
LU-Know0.018 ± 0.0110.037 ± 0.0070.944 ± 0.012
", + "bbox": [ + 550, + 88, + 877, + 152 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Impact of our U-Know loss function design. Best values are highlighted in red.", + "bbox": [ + 511, + 162, + 906, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "map, and our $\\mathcal{L}_{\\mathrm{U - Know}}$ . Table 6 shows the performance comparisons for the three losses. As shown, when the FSA-S was trained with $\\mathcal{L}_{\\mathrm{U - Know}}$ , it achieved the best performance across all three metrics. This highlights the effectiveness of $\\mathcal{L}_{\\mathrm{U - Know}}$ that enables the FSA-S to better leverage the teacher (FSA-T)'s prior knowledge including the uncertainty maps $\\hat{\\theta}$ , output features $\\mathbf{f}_i$ , and output HRMS $\\hat{\\mathbf{I}}_{\\mathrm{MS}}^{\\mathrm{HR}}$ for improved learning.", + "bbox": [ + 511, + 209, + 906, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Limitations", + "text_level": 1, + "bbox": [ + 511, + 342, + 640, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inference time limitations. Despite U-Know-DiffPAN's superior performance across multiple metrics, one notable limitation lies in its inference speed, compared to nondiffusion-based models. Diffusion models unavoidably require multiple sampling steps during inference, making this process inevitably slower. This iterative approach, while being beneficial for image generation of high quality, limits the model's efficiency in real-time applications. Future work could focus on optimizing the inference process by pruning sampling steps, while maintaining performance.", + "bbox": [ + 511, + 364, + 906, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 527, + 633, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced U-Know-DiffPAN, a novel diffusion-based PAN-sharpening framework that effectively leverages frequency information from PAN and LRMS inputs to generate high-quality HRMS images. Our Uncertainty-aware Knowledge Distillation (U-Know) strategy enables efficient feature transfer by distilling both frequency-rich features and uncertainty maps from the teacher model, enhancing the restoration performance for spatially weak regions while reducing computational costs for the student model (FSA-S). Extensive experiments on multiple datasets, including WV3, QB, and GF2, demonstrate that our U-Know-DiffPAN outperforms very recent state-of-the-art methods in PAN-sharpening, establishing a new benchmark for performance and efficiency in satellite image processing.", + "bbox": [ + 511, + 553, + 908, + 767 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgment", + "text_level": 1, + "bbox": [ + 511, + 777, + 684, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by National Research Foundation of Korea (NRF) grant funded by the Korean Government [Ministry of Science and ICT (Information and Communications Technology)] (Project Number: RS-2024-00338513, Project Title: AI-based Computer Vision Study for Satellite Image Processing and Analysis, $100\\%$ ).", + "bbox": [ + 511, + 801, + 908, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23076", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Bruno Aiazzi, Luciano Alparone, Stefano Baronti, Andrea Garzelli, and Massimo Selva. Mtf-tailored multiscale fusion of high-resolution ms and pan imagery. Photogrammetric Engineering & Remote Sensing, 72(5):591-596, 2006. 2", + "[2] Israa Amro, Javier Mateos, Miguel Vega, Rafael Molina, and Aggelos K Katsaggelos. A survey of classical methods and new trends in pansharpening of multispectral images. EURASIP Journal on Advances in Signal Processing, 2011:1-22, 2011. 1", + "[3] Coloma Ballester, Vicent Caselles, Laura Igual, Joan Verdera, and Bernard Rouge. A variational model for $\\mathfrak{p}+$ xs image fusion. International Journal of Computer Vision, 69:43-58, 2006. 2", + "[4] Jonathan T Barron. A general and adaptive robust loss function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4331-4339, 2019. 6", + "[5] Zihan Cao, Shiqi Cao, Liang-Jian Deng, Xiao Wu, Junming Hou, and Gemine Vivone. Diffusion model with disentangled modulations for sharpening multispectral and hyperspectral images. Information Fusion, 104:102158, 2024. 5", + "[6] Enrico Celeghini, Manuel Gadella, and Mariano A del Olmo. Hermite functions and fourier series. Symmetry, 13(5):853, 2021. 5", + "[7] Wei-Chun Chen, Chia-Che Chang, and Che-Rung Lee. Knowledge distillation with feature maps for image classification. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part III 14, pages 200–215. Springer, 2019. 3", + "[8] Jaewan Choi, Kiyun Yu, and Yongil Kim. A new adaptive component-substitution-based satellite image fusion by using partial replacement. IEEE transactions on geoscience and remote sensing, 49(1):295-309, 2010. 2", + "[9] Matteo Ciotola, Sergio Vitale, Antonio Mazza, Giovanni Poggi, and Giuseppe Scarpa. Pansharpening by convolutional neural networks in the full resolution framework. IEEE Transactions on Geoscience and Remote Sensing, 60: 1-17, 2022. 2", + "[10] Matteo Ciotola, Giuseppe Guarino, Gemine Vivone, Giovanni Poggi, Jocelyn Chanussot, Antonio Plaza, and Giuseppe Scarpa. Hyperspectral pansharpening: Critical review, tools and future perspectives. arXiv preprint arXiv:2407.01355, 2024. 1", + "[11] Liang-Jian Deng, Gemine Vivone, Mercedes E Paoletti, Giuseppe Scarpa, Jiang He, Yongjun Zhang, Jocelyn Chanussot, and Antonio Plaza. Machine learning in pan-sharpening: A benchmark, from shallow to deep networks. IEEE Geoscience and Remote Sensing Magazine, 10(3): 279-315, 2022. 2, 6", + "[12] P Kingma Diederik. Adam: A method for stochastic optimization. (No Title), 2014. 6", + "[13] Kento Doi and Akira Iwasaki. Sscnet: Spectral-spatial consistency optimization of cnn for pansharpening. In IGARSS 2019-2019 IEEE International Geoscience and Remote Sensing Symposium, pages 3141-3144. IEEE, 2019. 2" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Yule Duan, Xiao Wu, Haoyu Deng, and Liang-Jian Deng. Content-adaptive non-local convolution for remote sensing pansharpening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27738-27747, 2024. 1, 2, 7", + "[15] Joan Duran, Antoni Buades, Bartomeu Coll, Catalina Sbert, and Gwendoline Blanchet. A survey of pansharpening methods with a new band-decoupled variational model. ISPRS Journal of Photogrammetry and Remote Sensing, 125:78-105, 2017. 1", + "[16] Andrea Garzelli and Filippo Nencini. Hypercomplex quality assessment of multi/hyperspectral images. IEEE Geoscience and Remote Sensing Letters, 6(4):662-665, 2009. 6", + "[17] Anaïs Gastineau, Jean-François Aujol, Yannick Berthoumieu, and Christian Germain. A residual dense generative adversarial network for pansharpening with geometrical constraints. In 2020 IEEE International Conference on Image Processing (ICIP), pages 493-497. IEEE, 2020. 2", + "[18] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2", + "[19] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5", + "[20] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3", + "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3", + "[22] Jianwen Hu, Pei Hu, Xudong Kang, Hui Zhang, and Shaosheng Fan. Pan-sharpening via multiscale dynamic convolutional neural network. IEEE Transactions on Geoscience and Remote Sensing, 59(3):2231-2244, 2020. 2", + "[23] Wei Huang, Liang Xiao, Zhihui Wei, Hongyi Liu, and Songze Tang. A new pan-sharpening method with deep neural networks. IEEE Geoscience and Remote Sensing Letters, 12(5):1037-1041, 2015. 2", + "[24] Zeyi Huang, Yang Zou, BVK Kumar, and Dong Huang. Comprehensive attention self-distillation for weakly-supervised object detection. Advances in neural information processing systems, 33:16797-16807, 2020. 3", + "[25] Zi-Rong Jin, Tian-Jing Zhang, Tai-Xiang Jiang, Gemine Vivone, and Liang-Jian Deng. Lagconv: Local-context adaptive convolution kernels with global harmonic bias for pan-sharpening. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1113-1121, 2022. 2, 7", + "[26] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2, 4", + "[27] Jaehyup Lee, Soomin Seo, and Munchurl Kim. Sipsa-net: Shift-invariant pan sharpening with moving object alignment for satellite imagery. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10166-10174, 2021. 2", + "[28] Xiaobo Liu, Xiang Liu, Haoran Dai, Xudong Kang, Antonio Plaza, and Wenjie Zu. Mun-gan: A multiscale unsupervised network for remote sensing image pansharpening." + ], + "bbox": [ + 516, + 92, + 905, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23077", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE Transactions on Geoscience and Remote Sensing, 61: 1-18, 2023. 2", + "[29] Laetitia Loncan, Luis B De Almeida, José M Bioucas-Dias, Xavier Briottet, Jocelyn Chanussot, Nicolas Dobigeon, Sophie Fabre, Wenzhi Liao, Giorgio A Licciardi, Miguel Simoes, et al. Hyperspectral pansharpening: A review. IEEE Geoscience and remote sensing magazine, 3(3):27-46, 2015. 1", + "[30] Jiayi Ma, Wei Yu, Chen Chen, Pengwei Liang, Xiaojie Guo, and Junjun Jiang. Pan-gan: An unsupervised pan-sharpening method for remote sensing image fusion. Information Fusion, 62:110-120, 2020. 2", + "[31] Qingyan Meng, Wenxu Shi, Sijia Li, and Linlin Zhang. Pandiff: A novel pansharpening method based on denoising diffusion probabilistic model. IEEE Transactions on Geoscience and Remote Sensing, 61:1-17, 2023. 1, 2, 6, 7", + "[32] Guy P Nason and Bernard W Silverman. The stationary wavelet transform and some statistical applications. In Wavelets and statistics, pages 281-299. Springer, 1995. 3, 4, 5", + "[33] Qian Ning, Weisheng Dong, Xin Li, Jinjian Wu, and Guangming Shi. Uncertainty-driven loss for single image superresolution. Advances in Neural Information Processing Systems, 34:16398-16409, 2021. 4", + "[34] Xavier Otazu, María González-Audímaca, Octavi Fors, and Jorge Núñez. Introduction of sensor spectral response into image fusion methods. application to wavelet-based methods. IEEE Transactions on Geoscience and Remote Sensing, 43(10):2376-2385, 2005. 2", + "[35] Hao Phung, Quan Dao, and Anh Tran. Wavelet diffusion models are fast and scalable image generators. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10199-10208, 2023. 5", + "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3", + "[37] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4", + "[38] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE transactions on pattern analysis and machine intelligence, 45(4):4713-4726, 2022. 2", + "[39] Maximilian Seitzer, Arash Tavakoli, Dimitrije Antic, and Georg Martius. On the pitfalls of heteroscedastic uncertainty estimation with probabilistic neural networks. arXiv preprint arXiv:2203.09168, 2022. 4", + "[40] Shuyao Shang, Zhengyang Shan, Guangxing Liu, LunQian Wang, XingHua Wang, Zekai Zhang, and Jinglin Zhang. Resdiff: Combining cnn and diffusion model for image" + ], + "bbox": [ + 91, + 90, + 482, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "super-resolution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8975-8983, 2024. 5", + "[41] Mark J Shensa et al. The discrete wavelet transform: wedding the a trous and mallat algorithms. IEEE Transactions on signal processing, 40(10):2464-2482, 1992. 2", + "[42] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 6", + "[43] Baoli Sun, Xinchen Ye, Baopu Li, Haojie Li, Zhihui Wang, and Rui Xu. Learning scene structure guidance via cross-task knowledge transfer for single depth super-resolution. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 7792-7801, 2021. 3", + "[44] Gemine Vivone, Luciano Alparone, Jocelyn Chanussot, Mauro Dalla Mura, Andrea Garzelli, Giorgio A Licciardi, Rocco Restaino, and Lucien Wald. A critical comparison among pansharpening algorithms. IEEE Transactions on Geoscience and Remote Sensing, 53(5):2565-2586, 2014. 6", + "[45] Gemine Vivone, Mauro Dalla Mura, Andrea Garzelli, Rocco Restaino, Giuseppe Scarpa, Magnus O Ulfarsson, Luciano Alparone, and Jocelyn Chanussot. A new benchmark based on recent advances in multispectral pansharpening: Revisiting pansharpening with classical and emerging pansharpening methods. IEEE Geoscience and Remote Sensing Magazine, 9(1):53-81, 2020. 6", + "[46] Lucien Wald. Quality of high resolution synthesised images: Is there a simple criterion? In Third conference\" Fusion of Earth data: merging point measurements, raster maps and remotely sensed images\", pages 99-103. SEE/URISCA, 2000. 6", + "[47] Lucien Wald, Thierry Ranchin, and Marc Mangolini. Fusion of satellite images of different spatial resolutions: Assessing the quality of resulting images. Photogrammetric engineering and remote sensing, 63(6):691-699, 1997. 2", + "[48] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, pages 1-21, 2024. 2", + "[49] Tao Wang, Li Yuan, Xiaopeng Zhang, and Jiashi Feng. Distilling object detectors with fine-grained feature imitation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4933-4942, 2019. 3", + "[50] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 606-615, 2018. 4", + "[51] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6", + "[52] Stephen Weinstein and Paul Ebert. Data transmission by frequency-division multiplexing using the discrete fourier transform. IEEE transactions on Communication Technology, 19(5):628-634, 1971. 2", + "[53] Xiao Wu, Ting-Zhu Huang, Liang-Jian Deng, and Tian-Jing Zhang. Dynamic cross feature fusion for remote sensing pan-sharpening. In Proceedings of the IEEE/CVF International" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23078", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision, pages 14687-14696, 2021. 7", + "[54] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 5", + "[55] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13095-13105, 2023. 2, 4", + "[56] Yinghui Xing, Litao Qu, Shizhou Zhang, Jiapeng Feng, Xiuwei Zhang, and Yanning Zhang. Empower generalizability for pansharpening through text-modulated diffusion model. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1, 2, 3, 6, 7", + "[57] Kunran Xu, Lai Rui, Yishi Li, and Lin Gu. Feature normalized knowledge distillation for image classification. In European conference on computer vision, pages 664-680. Springer, 2020. 3", + "[58] Junfeng Yang, Xueyang Fu, Yuwen Hu, Yue Huang, Xinghao Ding, and John Paisley. Pannet: A deep network architecture for pan-sharpening. In Proceedings of the IEEE international conference on computer vision, pages 5449-5457, 2017. 2, 7", + "[59] Zhendong Yang, Zhe Li, Ailing Zeng, Zexian Li, Chun Yuan, and Yu Li. Vitkd: Feature-based knowledge distillation for vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1379-1388, 2024. 3", + "[60] Lewei Yao, Renjie Pi, Hang Xu, Wei Zhang, Zhenguo Li, and Tong Zhang. G-detkd: Towards general distillation framework for object detectors via contrastive and semantic-guided feature imitation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3591-3600, 2021. 3", + "[61] Qiangqiang Yuan, Yancong Wei, Xiangchao Meng, Huanfeng Shen, and Liangpei Zhang. A multiscale and multidepth convolutional neural network for remote sensing imagery pan-sharpening. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 11(3):978-989, 2018. 7", + "[62] Roberta H Yuhas, Alexander FH Goetz, and Joe W Boardman. Discrimination among semi-arid landscape endmembers using the spectral angle mapper (sam) algorithm. In JPL, Summaries of the Third Annual JPL Airborne Geoscience Workshop. Volume 1: AVIRIS Workshop, 1992. 6", + "[63] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 5", + "[64] Kai Zhang, Anfei Wang, Feng Zhang, Wenbo Wan, Jiande Sun, and Lorenzo Bruzzone. Spatial-spectral dual backprojection network for pansharpening. IEEE Transactions on Geoscience and Remote Sensing, 2023. 2, 7", + "[65] Yiman Zhang, Hanting Chen, Xinghao Chen, Yiping Deng, Chunjing Xu, and Yunhe Wang. Data-free knowledge dis" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tillation for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7852-7861, 2021. 3", + "[66] Yun Zhang, Wei Li, Simiao Li, Hanting Chen, Zhijun Tu, Wenjia Wang, Bingyi Jing, Shaohui Lin, and Jie Hu. Data upcycling knowledge distillation for image super-resolution. arXiv preprint arXiv:2309.14162, 2023. 3", + "[67] Yaqun Zhang, Zhichang Guo, Dazhi Zhang, and Boying Wu. An anisotropic variational pansharpening model with adaptive coefficients. Inverse Problems and Imaging, pages 0–0, 2024. 1", + "[68] Yafei Zhang, Xuji Yang, Huafeng Li, Minghong Xie, and Zhengtao Yu. Dcpnet: A dual-task collaborative promotion network for pansharpening. IEEE Transactions on Geoscience and Remote Sensing, 62:1-16, 2024. 2, 7", + "[69] Hao Zheng, Zhanlei Yang, Wenju Liu, Jizhong Liang, and Yanpeng Li. Improving deep neural networks using softplus units. In 2015 International joint conference on neural networks (IJCNN), pages 1-4. IEEE, 2015. 4", + "[70] Wu Zheng, Weiliang Tang, Li Jiang, and Chi-Wing Fu. Sessd: Self-ensembling single-stage object detector from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14494–14503, 2021. 3", + "[71] Jie Zhou, Daniel L Civco, and John A Silander. A wavelet transform method to merge landsat tm and spot panchromatic data. International journal of remote sensing, 19(4):743-757, 1998. 5" + ], + "bbox": [ + 516, + 92, + 903, + 486 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23079", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/U-Know-DiffPAN_ An Uncertainty-aware Knowledge Distillation Diffusion Framework with Details Enhancement for PAN-Sharpening/ca20f1f2-4ee9-4b83-80ca-6e2c71410bad_model.json b/2025/U-Know-DiffPAN_ An Uncertainty-aware Knowledge Distillation Diffusion Framework with Details Enhancement for PAN-Sharpening/ca20f1f2-4ee9-4b83-80ca-6e2c71410bad_model.json new file mode 100644 index 0000000000000000000000000000000000000000..15794a7b67d612f28d7bc88fc2ac3075596bcc77 --- /dev/null +++ b/2025/U-Know-DiffPAN_ An Uncertainty-aware Knowledge Distillation Diffusion Framework with Details Enhancement for PAN-Sharpening/ca20f1f2-4ee9-4b83-80ca-6e2c71410bad_model.json @@ -0,0 +1,2840 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.123, + 0.13, + 0.877, + 0.177 + ], + "angle": 0, + "content": "U-Know-DiffPAN: An Uncertainty-aware Knowledge Distillation Diffusion Framework with Details Enhancement for PAN-Sharpening" + }, + { + "type": "text", + "bbox": [ + 0.138, + 0.204, + 0.251, + 0.222 + ], + "angle": 0, + "content": "Sungpyo Kim" + }, + { + "type": "text", + "bbox": [ + 0.162, + 0.225, + 0.229, + 0.242 + ], + "angle": 0, + "content": "1 KAIST" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.248, + 0.275, + 0.26 + ], + "angle": 0, + "content": "ksp04204@kaist.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.345, + 0.204, + 0.471, + 0.222 + ], + "angle": 0, + "content": "Jeonghyeok Do" + }, + { + "type": "text", + "bbox": [ + 0.375, + 0.225, + 0.442, + 0.242 + ], + "angle": 0, + "content": "\\(^{1}\\)KAIST" + }, + { + "type": "text", + "bbox": [ + 0.313, + 0.248, + 0.504, + 0.26 + ], + "angle": 0, + "content": "ehwjdgur0913@kaist.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.569, + 0.204, + 0.678, + 0.222 + ], + "angle": 0, + "content": "Jaehyup Lee†" + }, + { + "type": "text", + "bbox": [ + 0.594, + 0.225, + 0.646, + 0.242 + ], + "angle": 0, + "content": "\\(^{2}\\)KNU" + }, + { + "type": "text", + "bbox": [ + 0.543, + 0.248, + 0.701, + 0.26 + ], + "angle": 0, + "content": "jaehyuplee@knu.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.751, + 0.204, + 0.877, + 0.222 + ], + "angle": 0, + "content": "Munchurl Kim" + }, + { + "type": "text", + "bbox": [ + 0.777, + 0.225, + 0.845, + 0.242 + ], + "angle": 0, + "content": "1KAIST" + }, + { + "type": "text", + "bbox": [ + 0.739, + 0.248, + 0.882, + 0.26 + ], + "angle": 0, + "content": "mkimee@kaist.ac.kr" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.282, + 0.336, + 0.295 + ], + "angle": 0, + "content": "PanDiff (TGRS 2023)" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.296, + 0.185, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.186, + 0.296, + 0.364, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.282, + 0.515, + 0.295 + ], + "angle": 0, + "content": "TMDiff (TGRS 2024)" + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.296, + 0.544, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.566, + 0.282, + 0.704, + 0.295 + ], + "angle": 0, + "content": "CANConv (CVPR 2024)" + }, + { + "type": "image", + "bbox": [ + 0.544, + 0.296, + 0.723, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.282, + 0.888, + 0.295 + ], + "angle": 0, + "content": "U-Know-DiffPAN (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.296, + 0.904, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.44, + 0.907, + 0.51 + ], + "angle": 0, + "content": "Figure 1. Visual comparison of PAN-sharpening results on the full-resolution WV3 dataset. The rightmost image shows the output of our proposed U-Know-DiffPAN framework, specifically the result produced by FSA-S (frequency selective attention student network). Notably, the proposed framework generates more detailed and robust results, particularly in high-uncertainty regions, outperforming the state-of-the-art model CANConv [14] and recent diffusion-based methods PanDiff [31] and TMDiff [56]. As highlighted in the red box, our approach successfully restores challenging highly-uncertain regions, such as cars, where other models fall short." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.53, + 0.327, + 0.546 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.562, + 0.485, + 0.851 + ], + "angle": 0, + "content": "Conventional methods for PAN-sharpening often struggle to restore fine details due to limitations in leveraging high-frequency information. Moreover, diffusion-based approaches lack sufficient conditioning to fully utilize Panchromatic (PAN) images and low-resolution multispectral (LRMS) inputs effectively. To address these challenges, we propose an uncertainty-aware knowledge distillation diffusion framework with details enhancement for PAN-sharpening, called U-Know-DiffPAN. The U-Know-DiffPAN incorporates uncertainty-aware knowledge distillation for effective transfer of feature details from our teacher model to a student one. The teacher model in our U-Know-DiffPAN captures frequency details through frequency selective attention, facilitating accurate reverse process learning. By conditioning the encoder on compact vector representations of PAN and LRMS and the decoder on Wavelet transforms, we enable rich frequency utilization. So, the high-capacity teacher model distills frequency-rich features into a lightweight student model aided by an un" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.907, + 0.653 + ], + "angle": 0, + "content": "certainty map. From this, the teacher model can guide the student model to focus on difficult image regions for PAN-sharpening via the usage of the uncertainty map. Extensive experiments on diverse datasets demonstrate the robustness and superior performance of our U-Know-DiffPAN over very recent state-of-the-art PAN-sharpening methods. The project page is available at https://kaist-viclab.github.io/U-Know-DiffPAN-site/." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.681, + 0.646, + 0.696 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.705, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Satellite imagery plays a vital role in a wide range of applications, including environmental monitoring, surveillance, and mapping [2, 10, 15, 29, 67]. High-resolution multispectral (HRMS) images that combine high spatial and spectral fidelity are commonly generated using PAN-sharpening techniques, fusing low-resolution multispectral (LRMS) images with high-resolution (HR) panchromatic (PAN) images. The primary objective of PAN-sharpening is to retain the spatial detail of PAN images while preserving the spectral information of LRMS images. Recently, deep learning (DL)-based approaches have gained attraction in PAN-sharpening tasks, leveraging deep networks to perform end-to-end image regression. However, DL models" + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.863, + 0.452, + 0.877 + ], + "angle": 0, + "content": "\\(^{1}\\) Korea Advanced Institute of Science and Technology (KAIST)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.877, + 0.33, + 0.89 + ], + "angle": 0, + "content": "\\(^{2}\\)Kyungpook National University (KNU)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.89, + 0.365, + 0.902 + ], + "angle": 0, + "content": "† Co-corresponding authors (equal contribution)." + }, + { + "type": "list", + "bbox": [ + 0.108, + 0.863, + 0.452, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23069" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "are typically limited by the local receptive fields of convolutional layers, which primarily focus on local patterns and fail to capture long-range dependencies. This often leads to overly smooth outputs with reduced high-frequency details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.155, + 0.483, + 0.426 + ], + "angle": 0, + "content": "To overcome this limitation, generative models such as generative adversarial networks (GANs) [18] have been applied to PAN-sharpening, aiming to produce more realistic results. Despite their success, GANs [28, 30] pose additional challenges, including training instability, mode collapse, and artifact generation, which can degrade image quality. Recently, diffusion models [21, 38, 42, 55] have demonstrated strong performance across a variety of image generation and restoration tasks. In PAN-sharpening, diffusion-based methods [31, 56] have shown promise by using PAN and LRMS images as conditioning inputs. However, they are limited by simple conditioning strategies that do not fully utilize the complementary information provided by the PAN and LRMS inputs, thus leading to suboptimal restoration quality. Additionally, they often incur high computational costs and longer inference times compared to other DL-based methods, limiting their practical applicability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.483, + 0.701 + ], + "angle": 0, + "content": "To address these challenges, we propose U-Know-DiffPAN demonstrated in Fig. 2, a novel PAN-sharpening framework that incorporates an uncertainty-aware knowledge distillation (U-Know) strategy. Our framework is designed to fully leverage the frequency information in PAN and LRMS inputs while reducing computational costs. Our teacher model, is built to maximize information extraction from PAN and LRMS images by utilizing (i) a Feed Forward Attention (FFA) block in the encoder to capture compact vector representations of PAN and LRMS, and (ii) a High Quality Frequency Enhancement (HQFE) block in the decoder, composed of Fourier Transform Channel Attention (FTCA) and Stationary Wavelet Transform Cross Attention (SWTCA) modules. The FTCA enhances crucial frequency components through Fourier transform-based [52] attention, while the SwTCA incorporates high-frequency information from PAN and low-frequency information from LRMS into the decoder." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.9 + ], + "angle": 0, + "content": "Our U-Know strategy leverages uncertainty-aware knowledge distillation by transferring the features from the teacher model to a light-weight student model. The teacher model denoises HRMS images along with an uncertainty map, identifying spatial regions where the prediction confidence is low. Then based of the U-Know strategy, the teacher model distills to the student model not only the frequency-aware features but also the uncertainty map to address spatial weaknesses in the teacher's predictions. This approach leads to improved performance and reduced computational cost, making our U-Know-DiffPAN more effective and efficient for PAN-sharpening. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.151 + ], + "angle": 0, + "content": "- We firstly incorporate the uncertainty-aware knowledge distillation into diffusion based PAN-sharpening with the guidance of pixel-wise restoration uncertainty and frequency-selective attention for details enhancement;" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.153, + 0.905, + 0.212 + ], + "angle": 0, + "content": "- Our teacher model leverages compact vector representations to enhance encoding efficiency and employs frequency selective attention in the decoder to refine frequency components for PAN-sharpening." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.213, + 0.905, + 0.273 + ], + "angle": 0, + "content": "- We conduct extensive experiments on multiple datasets [11], including WV3, QB, and GF2, demonstrating state-of-the-art performance and surpassing very recent methods in PAN-sharpening with large margins." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.291, + 0.663, + 0.307 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.318, + 0.681, + 0.334 + ], + "angle": 0, + "content": "2.1. PAN-Sharpening" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.341, + 0.905, + 0.717 + ], + "angle": 0, + "content": "PAN-sharpening is a technique to fuse high-resolution (HR) panchromatic (PAN) images with low-resolution multispectral (LRMS) images to produce HR multispectral (HRMS) images. Traditional PAN-sharpening techniques are broadly categorized into three approaches: component substitution (CS) [8, 41], multi-resolution analysis (MRA) [1, 34], and variational optimization (VO) [3]. Modern PAN-sharpening methods primarily consist of CNN-based [14, 25, 47, 58], transformer-based [25, 68], and GAN-based [28, 30] models. LAGConv [25] introduces local-context adaptive convolutional kernels with a global harmonic bias mechanism. S2DBPN [64] uses a spatial-spectral dual back-projection network for effective PAN and LRMS fusion. DCPNet [68] establishes a transformer-based dual-task parallel framework to optimize reconstruction. CANConv [14] integrates a non-local adaptive convolution module, enhancing spatial adaptability and non-local self-similarity. Besides the above methods, there are other DL-based methods for PAN-sharpening [9, 13, 17, 22, 23, 27]. While deep learning (DL)-based methods can capture complex mappings, they encounter challenges. CNN-based approaches tend to produce overly smooth results due to local receptive field of convolution layers. GAN-based models, though capable of generating realistic textures, may suffer from instability and introduce unrealistic artifacts." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.72, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Diffusion-based approaches. Diffusion models (DMs) [21, 42] offer a promising alternative to conventional DL-based methods for image restoration [48, 55], iteratively refining noisy images through progressive denoising. Unlike GANs [18, 26], which can be prone to instability, DMs use iterative denoising steps with flexible conditioning, making them robust for tasks such as image restoration and PAN-sharpening. Recent diffusion-based PAN-sharpening models [31, 56] have shown significant advancements. PanDiff [31], however, employs a relatively simplistic conditioning approach that limits its ability to fully utilize the information from the input data, sometimes resulting in subopti" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23070" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.304 + ], + "angle": 0, + "content": "mal restoration quality. In contrast, TMDiff [56] employs a complex conditioning mechanism, requiring an additional pretrained CLIP [36] text encoder, which adds unnecessary complexity without substantial performance gains. Our proposed framework, U-Know-DiffPAN, handles these limitations by integrating an advanced conditioning strategy that maximizes the use of input information. Specifically, our U-Know-DiffPAN conditions the encoder on compact vector representations of PAN and LRMS, while the decoder is conditioned on the stationary wavelet transform [32] of PAN and LRMS. This approach allows our U-Know-DiffPAN to effectively capture both frequency details and spatial fidelity, achieving superior restoration quality compared to existing diffusion-based methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.312, + 0.303, + 0.328 + ], + "angle": 0, + "content": "2.2. Knowledge Distillation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.334, + 0.485, + 0.622 + ], + "angle": 0, + "content": "Knowledge Distillation (KD) [20] is a well-known method for transferring knowledge from a large, model (teacher) to a smaller model (student), enabling deployment in resource-constrained settings while retaining performance. KD has demonstrated effectiveness across a range of computer vision tasks, including classification [7, 57, 65], object detection [24, 49, 60, 70], and image restoration [43, 59, 66]. However, previous KD methods, often fall short in tasks such as super-resolution [66], where capturing complicated detail is essential. This limitation arises because conventional KD student learns all regions equally, without focusing on areas that require concentrated learning. To address this, we propose uncertainty-aware knowledge distillation (U-Know), which introduces an uncertainty map into the distillation process. Our U-Know strategy informs the student of uncertain regions, areas the teacher model finds challenging to restore, enabling the student model to refine crucial details, thereby enhancing fidelity in both spatial and spectral domains." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.635, + 0.192, + 0.65 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.659, + 0.362, + 0.673 + ], + "angle": 0, + "content": "3.1. Overview of U-Know-DiffPAN" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.681, + 0.484, + 0.803 + ], + "angle": 0, + "content": "Fig. 2 shows the overview of our U-Know-DiffPAN framework. The proposed U-Know-DiffPAN framework comprises a teacher network with frequency-selective attention (FSA-T) and a student network (FSA-S), designed within a knowledge distillation paradigm. PAN image \\(\\mathbf{I}_{\\mathrm{PAN}}\\) and interpolated LRMS image \\(\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{LR}}\\) are inputted in concatenation to generate HRMS image \\(\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{HR}}\\). Both FSA-T and FSA-S are designed to predict the residuals as:" + }, + { + "type": "equation", + "bbox": [ + 0.226, + 0.811, + 0.483, + 0.831 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {0} = \\mathbf {I} _ {\\mathrm {M S}} ^ {\\mathrm {H R}} - \\mathbf {I} _ {\\mathrm {M S}} ^ {\\mathrm {L R}}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Our framework operates in two main stages: (i) pre-training the FSA-T within a diffusion process to produce an initial prediction \\(\\widehat{\\mathbf{X}}_0\\) alongside an uncertainty map \\(\\hat{\\theta}\\) that identifies spatially weak regions; and (ii) training the FSA-S by" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.093, + 0.668, + 0.103 + ], + "angle": 0, + "content": "Step1. Teacher (FSA-T) pre-training" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.106, + 0.631, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.106, + 0.875, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.174, + 0.65, + 0.185 + ], + "angle": 0, + "content": "Step2. Student (FSA-S) training" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.187, + 0.633, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.187, + 0.902, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.258, + 0.906, + 0.3 + ], + "angle": 0, + "content": "Figure 2. Overview of our uncertainty-aware knowledge-distillation diffusion framework with details enhancement, called U-Know-DiffPAN." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.906, + 0.448 + ], + "angle": 0, + "content": "leveraging this uncertainty map to guide the FSA-S in refining these regions through the KD. In Sec. 3.2, we describe the FSA-T's diffusion-based training process, which incorporates frequency information from PAN and LRMS inputs. Sec. 3.3 details the FSA-T architecture, specifically designed to enhance frequency components. Finally, in Sec. 3.4, we explain the FSA-S's uncertainty-aware knowledge distillation framework, which utilizes the uncertainty map to achieve precise and reliable PAN-sharpening." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.455, + 0.906, + 0.47 + ], + "angle": 0, + "content": "3.2. Diffusion Process of Teacher Network (FSA-T)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.906, + 0.599 + ], + "angle": 0, + "content": "Diffusion models operate by defining a forward process that gradually adds noise to data until it becomes indistinguishable from random noise, and a reverse process that learns to denoise the data, reconstructing the original information. We leverage this framework into a frequency-selective attention teacher network (FSA-T) as our teacher network, to generate frequency enhanced HRMS images with their pixel-wise uncertainty information." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.599, + 0.906, + 0.675 + ], + "angle": 0, + "content": "Forward process. In the forward diffusion process, noise is progressively added to the residual image, \\(\\mathbf{X}_0\\), until it converges to a Gaussian distribution, \\(\\mathbf{X}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\), after \\(T\\) timesteps. At each timestep \\(t \\in \\{1, \\dots, T\\}\\), a noisy version of the residual is obtained as follows [21]:" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.683, + 0.905, + 0.7 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {X} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.708, + 0.904, + 0.739 + ], + "angle": 0, + "content": "where \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) is Gaussian noise, and \\(\\bar{\\alpha}_t = \\prod_{s=1}^t (1 - \\beta_s)\\) controls the noise level at step \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.74, + 0.906, + 0.874 + ], + "angle": 0, + "content": "Reverse process. Following the DDPM [21], the reverse process aims to recover \\(\\mathbf{X}_0\\) from \\(\\mathbf{X}_T\\) by predicting a denoised image at each step. To perform denoising at each timestep, the denoising network (FSA-T), denoted as \\(\\Psi\\), is designed to predict the original \\(\\mathbf{X}_0\\) rather than the added noise \\(\\epsilon\\), as we found this empirically to be beneficial for PAN-sharpening. Additionally, \\(\\Psi\\) simultaneously produces an uncertainty map \\(\\hat{\\theta}\\) for each prediction, providing pixelwise confidence information as:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.88, + 0.905, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\left[ \\widehat {\\mathbf {X}} _ {0} \\mid \\widehat {\\boldsymbol {\\theta}} \\right] = \\Psi \\left(\\left[ \\mathbf {X} _ {t} \\mid \\mathbf {I} _ {\\text {P A N}} \\mid \\mathbf {I} _ {\\mathrm {M S}} ^ {\\mathrm {L R}} \\right]; \\mathbf {v}, \\mathcal {S} - \\text {C o n d}; t\\right), \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23071" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.258, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.09, + 0.761, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.094, + 0.903, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.271, + 0.258, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.274, + 0.482, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.272, + 0.681, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.272, + 0.902, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.355, + 0.905, + 0.385 + ], + "angle": 0, + "content": "Figure 3. Architecture of our proposed teacher model with frequency-selective attention, denoted as FSA-T, for PAN-sharpening. The FSA-T is designed to fully utilize frequency information for details enhancement from PAN and LRMS inputs (more details in Supplemental.)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.394, + 0.483, + 0.562 + ], + "angle": 0, + "content": "where \\(\\left[\\cdot \\mid \\cdot \\right]\\) indicates channel-wise concatenation, \\(\\widehat{\\mathbf{X}}_0\\) is a predicted residual, \\(\\hat{\\theta}\\) is a predicted uncertainty map, \\(\\mathbf{v}\\) and \\(S\\)-Cond represent a learning compact representation (vector) and Stationary Wavelet Transform (SWT)-based conditioning terms, respectively. This formulation allows the FSA-T to iteratively refine the noisy input, guided by the conditioned information (\\(\\mathbf{v}\\), \\(S\\)-Cond), toward an accurate reconstruction of the \\(\\mathbf{X}_0\\). \\(\\hat{\\theta}\\) captures pixel-wise prediction confidence, which guides the FSA-S to more rigorously learn the regions with high uncertainty. To ensure positivity, we apply the SoftPlus [69] activation to \\(\\hat{\\theta}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.563, + 0.484, + 0.609 + ], + "angle": 0, + "content": "Uncertainty-driven diffusion loss. Inspired by [33, 39], we employ the following uncertainty-driven diffusion loss function, \\(\\mathcal{L}_{\\mathrm{U - Diff}}\\), to optimize the teacher network \\(\\Psi\\):" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.615, + 0.483, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {U - D i f f}} = \\left\\| \\frac {1}{2 \\hat {\\boldsymbol {\\theta}}} \\odot \\left| \\widehat {\\mathbf {X}} _ {0} - \\mathbf {X} _ {0} \\right| + \\frac {1}{2} \\log \\hat {\\boldsymbol {\\theta}} \\right\\| _ {1}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.484, + 0.737 + ], + "angle": 0, + "content": "where \\(\\hat{\\theta}\\) serves as the estimated variance term and is regarded as the uncertainty map in our framework. This objective function encourages the model to minimize prediction errors in low-confidence regions, thus yielding a more accurate and spatially robust output." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.439, + 0.759 + ], + "angle": 0, + "content": "3.3. Details of Teacher Network Architecture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our teacher network, FSA-T, is designed to enhance frequency details for PAN-sharpening tasks by fully utilizing complementary information from PAN and LRMS inputs. The FSA-T adopts an encoder-decoder (U-Net [37]) structure, where each stage is specifically tailored to capture and refine critical spatial and frequency features. The encoder of FSA-T employs Feed Forward Attention (FFA) blocks to extract compact and informative representations from the PAN and LRMS inputs. This attention mechanism allows" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.905, + 0.425 + ], + "angle": 0, + "content": "the encoder to focus on essential features, preparing the data for effective frequency enhancement in subsequent layers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.906, + 0.623 + ], + "angle": 0, + "content": "The decoder of FSA-T integrates High-Quality Frequency Enhancement (HQFE) blocks at each decoding level to further refine frequency components. Each HQFE block comprises two key modules: Fourier Transform Channel Attention (FTCA) that emphasizes important frequency components via Fourier Transform based attention, and Stationary Wavelet Transform Cross Attention (SWTCA) that merges high-frequency information from PAN and low-frequency content from LRMS, both of which are decomposed by SwT [32]. By incorporating these specialize components, FSA-T effectively produces \\(\\widehat{\\mathbf{X}}_0\\) with enhanced spatial fidelity, establishing an effective uncertainty-aware knowledge distillation (U-Know) strategy." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.624, + 0.906, + 0.7 + ], + "angle": 0, + "content": "Feed Forward Attention (FFA). The FFA block in Fig. 3 is designed to leverage compact information [55] from the PAN and LRMS input to establish a rough structural representation during encoding. A compact vector representation \\(\\mathbf{v} \\in \\mathbb{R}^{1 \\times D}\\) is obtained by a vector extractor \\((\\mathcal{E})\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.642, + 0.711, + 0.905, + 0.731 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} = \\mathcal {E} \\left(\\mathbf {I} _ {\\mathrm {P A N}}, \\mathbf {I} _ {\\mathrm {M S}} ^ {\\mathrm {L R}}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.743, + 0.905, + 0.789 + ], + "angle": 0, + "content": "\\(\\mathbf{v}\\) is then used as a dynamic modulation [26, 50] parameter within the FFA, enabling \\(\\mathcal{E}\\) to incorporate \\(\\mathbf{v}\\) and establish an approximate structure during restoration as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.801, + 0.725, + 0.818 + ], + "angle": 0, + "content": "\\[\n[ \\boldsymbol {\\beta} _ {n} \\mid \\gamma_ {n} ] = \\operatorname {L i n e a r} (\\mathbf {v}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.82, + 0.905, + 0.837 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {n} ^ {\\prime} = \\gamma_ {n} \\odot \\mathrm {G N} \\left(\\mathbf {f} _ {n}\\right) + \\beta_ {n}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.839, + 0.85, + 0.856 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {n} \\leftarrow \\operatorname {C o n v} \\left(\\mathbf {f} _ {n} ^ {\\prime}\\right) \\odot \\operatorname {G E L U} \\left(\\operatorname {C o n v} \\left(\\mathbf {f} _ {n} ^ {\\prime}\\right)\\right) + \\mathbf {f} _ {n},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.87, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{f}_n\\in \\mathbb{R}^{C\\times h\\times w}\\) denotes input feature for the \\(n\\) th encoder, Linear indicates Linear operation, \\(\\odot\\) denotes" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23072" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.166 + ], + "angle": 0, + "content": "element-wise multiplication, GN indicates group normalization [54], GELU indicates GELU [19] activation and Conv consists of a sequential \\( 1 \\times 1 \\) convolution and a \\( 3 \\times 3 \\) depth-wise convolution layer (Details of \\( \\mathcal{E} \\) is in the Supplemental.)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.483, + 0.319 + ], + "angle": 0, + "content": "Fourier Transform Channel Attention (FTCA). The FTCA block in Fig. 3 performs self-attention in frequency domain by leveraging the 2D Fast Fourier Transform (FFT), denoted as \\(\\mathcal{F}\\). \\(\\mathcal{F}\\) allows for efficient identification and emphasis of critical frequency components. For the \\(n\\)-th decoder block in Fig. 3, the input feature \\(\\mathbf{f}_n \\in \\mathbb{R}^{C \\times h \\times w}\\) of \\(C\\) channels is transformed into query \\((\\mathbf{Q})\\), key \\((\\mathbf{K})\\), and value \\((\\mathbf{V})\\) representations. The 2D FFT \\(\\mathcal{F}\\) is then applied independently to each channel of \\(\\mathbf{Q}\\), \\(\\mathbf{K}\\), and \\(\\mathbf{V}\\), mapping them transform from time domain to frequency domain as:" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.332, + 0.483, + 0.35 + ], + "angle": 0, + "content": "\\[\n[ \\mathbf {Q} \\mid \\mathbf {K} \\mid \\mathbf {V} ] = \\mathcal {F} (\\operatorname {C o n v} \\left(\\mathbf {f} _ {n}\\right)), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.36, + 0.484, + 0.437 + ], + "angle": 0, + "content": "where \\(\\mathbf{Q},\\mathbf{K}\\) and \\(\\mathbf{V}\\) reside in \\(\\mathbb{C}^{C\\times h\\times (\\frac{w}{2} +1)}\\) due to the Hermitian symmetry property [6]. Note that each frequency domain representation of \\(\\mathbf{Q},\\mathbf{K}\\), and \\(\\mathbf{V}\\) includes real and imaginary parts. So, we decompose \\(\\mathbf{Q},\\mathbf{K}\\), and \\(\\mathbf{V}\\) into real and imaginary parts as:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.449, + 0.483, + 0.467 + ], + "angle": 0, + "content": "\\[\n\\left[ \\mathbf {Q} _ {\\mathbb {R}} \\mid \\mathbf {K} _ {\\mathbb {R}} \\mid \\mathbf {V} _ {\\mathbb {R}} \\right] + j \\left[ \\mathbf {Q} _ {\\mathbb {I}} \\mid \\mathbf {K} _ {\\mathbb {I}} \\mid \\mathbf {V} _ {\\mathbb {I}} \\right] = \\left[ \\mathbf {Q} \\mid \\mathbf {K} \\mid \\mathbf {V} \\right], \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.484, + 0.525 + ], + "angle": 0, + "content": "where \\(j = \\sqrt{-1}\\), and \\(\\mathbf{Q}_{\\mathbb{R}}\\), \\(\\mathbf{K}_{\\mathbb{R}}\\), \\(\\mathbf{V}_{\\mathbb{R}}\\), \\(\\mathbf{Q}_{\\mathbb{I}}\\), \\(\\mathbf{K}_{\\mathbb{I}}\\), \\(\\mathbf{V}_{\\mathbb{I}}\\) are in \\(\\mathbb{R}^{C\\times h\\times (\\frac{w}{2} + 1)}\\). Self-attention operations are then performed separately on the real and imaginary parts as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.536, + 0.483, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {l} = \\operatorname {S o f t M a x} \\left(\\mathbf {Q} _ {l} \\mathbf {K} _ {l} ^ {\\top} / \\sqrt {C}\\right) \\mathbf {V} _ {l}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.484, + 0.615 + ], + "angle": 0, + "content": "where \\(l \\in \\{\\mathbb{R},\\mathbb{I}\\}\\) indicates the real or imaginary feature, respectively. Finally, the inverse Fourier Transform \\(\\mathcal{F}^{-1}\\) is applied to \\(\\mathbf{f}_{\\mathbb{R}}\\) and \\(\\mathbf{f}_{\\mathbb{I}}\\) to map them back to the time domain:" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.627, + 0.483, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {n} \\leftarrow \\mathcal {F} ^ {- 1} \\left(\\mathbf {f} _ {\\mathbb {R}} + j \\mathbf {f} _ {\\mathbb {I}}\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.483, + 0.686 + ], + "angle": 0, + "content": "This process ensures that refined frequency information be selectively captured and utilized." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.484, + 0.854 + ], + "angle": 0, + "content": "Stationary Wavelet Transform Cross Attention (SWTCA). The SwTCA block is designed to leverage Wavelet components of \\(\\mathbf{I}_{\\mathrm{PAN}}\\) and \\(\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{LR}}\\) motivated by [5, 35, 40, 71]. As shown in Fig. 3, it enhances the features outputted by the FTCA block, utilizing the Stationary Wavelet Transform (SWT) [32], denoted as \\(S\\). The SwTCA block selectively incorporates relevant frequency components from \\(\\mathbf{I}_{\\mathrm{PAN}}\\) and \\(\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{LR}}\\) through a two-stage cross-attention process. The SwT-based conditioning terms (\\(S\\)-Cond) are constructed by decomposing \\(\\mathbf{I}_{\\mathrm{PAN}}\\) and \\(\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{LR}}\\) using \\(S\\) into four components:" + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.866, + 0.483, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{r l} & {\\left[ \\mathbf {L} _ {\\mathrm {P A N}} \\mid \\mathbf {H} _ {\\mathrm {P A N}} \\mid \\mathbf {V} _ {\\mathrm {P A N}} \\mid \\mathbf {D} _ {\\mathrm {P A N}} \\right] = \\mathcal {S} (\\mathbf {I} _ {\\mathrm {P A N}}),} \\\\ & {\\left[ \\mathbf {L} _ {\\mathrm {M S}} ^ {\\mathrm {L R}} \\mid \\mathbf {H} _ {\\mathrm {M S}} ^ {\\mathrm {L R}} \\mid \\mathbf {V} _ {\\mathrm {M S}} ^ {\\mathrm {L R}} \\mid \\mathbf {D} _ {\\mathrm {M S}} ^ {\\mathrm {L R}} \\right] = \\mathcal {S} (\\mathbf {I} _ {\\mathrm {M S}} ^ {\\mathrm {L R}}),} \\end{array} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "where \\(\\mathbf{L}\\) represent the low-frequency approximation components, while \\(\\mathbf{H},\\mathbf{V},\\mathbf{D}\\) denote the high-frequency horizontal, vertical, and diagonal details, respectively. Since \\(\\mathbf{I}_{\\mathrm{PAN}}\\) provides high spatial resolution but lacks spectral information, only its high-frequency components \\(\\mathbf{H}_{\\mathrm{PAN}},\\) \\(\\mathbf{V}_{\\mathrm{PAN}}\\) ,and \\(\\mathbf{D}_{\\mathrm{PAN}}\\) are used. Conversely, \\(\\mathbf{L}_{\\mathrm{MS}}^{\\mathrm{LR}}\\) from \\(\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{LR}}\\) contains richer spectral information. These components are concatenated channel-wise to form \\(S\\) -Cond as:" + }, + { + "type": "equation", + "bbox": [ + 0.569, + 0.221, + 0.905, + 0.24 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} - \\text {C o n d} = \\left[ \\mathbf {L} _ {\\mathrm {M S}} ^ {\\mathrm {L R}} \\mid \\mathbf {H} _ {\\mathrm {P A N}} \\mid \\mathbf {V} _ {\\mathrm {P A N}} \\mid \\mathbf {D} _ {\\mathrm {P A N}} \\right]. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.247, + 0.907, + 0.354 + ], + "angle": 0, + "content": "To inject \\(S\\)-Cond into the feature \\(\\mathbf{f}_n\\) obtained from FTCA at the \\(n\\)-th decoder block, we employ a two-stage cross-attention process. In the first stage, \\(\\mathbf{Q}_1\\) and \\(\\mathbf{K}_1\\) are derived from \\(S\\)-Cond, directing attention toward critical frequency components, while \\(\\mathbf{V}_1\\) is derived from \\(\\mathbf{f}_n\\). The resulting features are reshaped to \\(C \\times hw\\) for channel attention [63], producing an intermediate feature \\(\\mathbf{f}_n^1 \\in \\mathbb{R}^{C \\times h \\times w}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.374, + 0.854, + 0.391 + ], + "angle": 0, + "content": "\\[\n\\left[ \\mathbf {Q} _ {1} \\mid \\mathbf {K} _ {1} \\right] = \\operatorname {C o n v} (S - \\text {C o n d}), \\mathbf {V} _ {1} = \\operatorname {C o n v} (\\mathbf {f} _ {n}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.391, + 0.905, + 0.42 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {n} ^ {1} = \\operatorname {S o f t M a x} \\left(\\mathbf {Q} _ {1} \\mathbf {K} _ {1} ^ {\\top} / \\sqrt {C}\\right) \\mathbf {V} _ {1}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.906, + 0.52 + ], + "angle": 0, + "content": "where \\(\\mathbf{Q}_1, \\mathbf{K}_1, \\mathbf{V}_1\\) are in \\(\\mathbb{R}^{C \\times hw}\\). In the second stage, \\(\\mathbf{Q}_2\\) is derived from the intermediate feature \\(\\mathbf{f}_n^1\\), while \\(\\mathbf{K}_2\\) and \\(\\mathbf{V}_2\\) come from \\(S\\)-Cond, enabling the model to refine the correlation between the reconstructed feature \\(\\mathbf{f}_n\\) and significant frequency components in \\(S\\)-Cond. This stage yields the final feature \\(\\mathbf{f}_n \\in \\mathbb{R}^{C \\times h \\times w}\\) after reshaping:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.541, + 0.853, + 0.558 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {2} = \\operatorname {C o n v} \\left(\\mathbf {f} _ {n} ^ {1}\\right), \\left[ \\mathbf {K} _ {2} \\mid \\mathbf {V} _ {2} \\right] = \\operatorname {C o n v} (S - C o n d),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.558, + 0.905, + 0.586 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {n} \\leftarrow \\operatorname {S o f t M a x} \\left(\\mathbf {Q} _ {2} \\mathbf {K} _ {2} ^ {\\top} / \\sqrt {C}\\right) \\mathbf {V} _ {2}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.906, + 0.655 + ], + "angle": 0, + "content": "Through these two stages of cross-attention, the final feature \\(\\mathbf{f}_n\\) effectively incorporates essential frequency components from both PAN and MS inputs, ensuring enhanced frequency representation and high-quality HRMS output." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.663, + 0.877, + 0.679 + ], + "angle": 0, + "content": "3.4. Uncertainty-Aware Knowledge Distillation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.685, + 0.906, + 0.729 + ], + "angle": 0, + "content": "To address the substantial computational demands of the teacher (FSA-T) \\(\\Psi\\), we propose an novel uncertainty-aware knowledge distillation (U-Know) strategy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.73, + 0.906, + 0.835 + ], + "angle": 0, + "content": "Student network (FSA-S) \\(\\psi\\). The student model FSA-S is designed with the same number of encoder and decoder blocks as FSA-T. Each block consists solely of Res-blocks, and unlike the teacher, the student operates without any additional conditional inputs. As shown in Fig. 2, the lightweight student model \\(\\psi\\) predicts the denoised image \\(\\widetilde{\\mathbf{X}}_0\\) from the input \\(\\mathbf{X}_t\\) at timestep \\(t\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.842, + 0.905, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\widetilde {\\mathbf {X}} _ {0} = \\psi \\left(\\left[ \\mathbf {X} _ {t} \\mid \\mathbf {I} _ {\\text {P A N}} \\mid \\mathbf {I} _ {\\mathrm {M S}} ^ {\\mathrm {L R}} \\right]; t\\right). \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "U-Know loss function. FSA-T \\(\\Psi\\) generates an uncertainty map \\(\\hat{\\theta}\\) along with the output \\(\\bar{\\mathbf{X}}_0\\) during training." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23073" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "These two outputs are frozen and used in the U-Know loss function, \\(\\mathcal{L}_{\\mathrm{U - Know}}\\), to train the FSA-S \\(\\psi\\). The total training objective combines a hard loss \\((\\mathcal{L}_{\\mathrm{hard}})\\), a soft loss \\((\\mathcal{L}_{\\mathrm{soft}})\\), and a feature loss \\((\\mathcal{L}_{\\mathrm{feat}})\\) to distill frequency-rich features and enhance regions where \\(\\Psi\\) exhibits spatial weaknesses points. The total loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.194, + 0.483, + 0.211 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {U - K n o w}} = \\mathcal {L} _ {\\text {h a r d}} + \\lambda_ {s} \\mathcal {L} _ {\\text {s o f t}} + \\lambda_ {f} \\mathcal {L} _ {\\text {f e a t}}, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.222, + 0.483, + 0.282 + ], + "angle": 0, + "content": "where \\(\\lambda_{s}\\) and \\(\\lambda_{f}\\) control the contributions of \\(\\mathcal{L}_{\\mathrm{soft}}\\) and \\(\\mathcal{L}_{\\mathrm{feat}}\\) respectively. \\(\\mathcal{L}_{\\mathrm{hard}}\\) measures the difference between \\(\\psi\\) output \\(\\widetilde{\\mathbf{X}}_0\\) and the ground truth \\(\\mathbf{X}_0\\), weighted by \\(\\Psi\\)'s uncertainty map \\(\\hat{\\theta}\\) to emphasize challenging regions:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.291, + 0.483, + 0.32 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {h a r d}} = \\left\\| (\\tau + \\hat {\\boldsymbol {\\theta}}) \\odot \\left| \\widetilde {\\mathbf {X}} _ {0} - \\mathbf {X} _ {0} \\right| \\right\\| _ {1}, \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.328, + 0.484, + 0.373 + ], + "angle": 0, + "content": "where \\(\\tau\\) is a hyper-parameter. \\(\\mathcal{L}_{\\mathrm{soft}}\\) computes the difference between \\(\\psi\\) 's output \\(\\widetilde{\\mathbf{X}}_0\\) and \\(\\Psi\\) 's output \\(\\widehat{\\mathbf{X}}_0\\), weighted by the \\(\\hat{\\theta}\\) to help \\(\\psi\\) better approximate the \\(\\Psi\\) 's knowledge:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.382, + 0.483, + 0.41 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s o f t}} = \\left\\| (\\tau - \\hat {\\boldsymbol {\\theta}}) \\odot \\left| \\widetilde {\\mathbf {X}} _ {0} - \\widehat {\\mathbf {X}} _ {0} \\right| \\right\\| _ {1}. \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.422, + 0.484, + 0.576 + ], + "angle": 0, + "content": "In the regions where \\(\\hat{\\pmb{\\theta}}\\) values are high, indicating that \\(\\Psi\\) struggles with the task, we apply a weight of \\((\\tau + \\hat{\\pmb{\\theta}})\\) and use \\(\\mathbf{X}_0\\) as the hard ground truth to guide \\(\\psi\\) toward the true target. Conversely, in the areas where \\(\\hat{\\pmb{\\theta}}\\) values are low, indicating that \\(\\Psi\\) performs well, we apply a weight of \\((\\tau - \\hat{\\pmb{\\theta}})\\) and use \\(\\Psi\\)'s output \\(\\widehat{\\mathbf{X}}_0\\) as the soft ground truth, allowing \\(\\psi\\) to approximate \\(\\Psi\\)'s knowledge in these regions. These dual loss formulations enable \\(\\Psi\\) to focus on challenging areas while effectively leveraging the \\(\\Psi\\)'s strengths. \\(\\mathcal{L}_{\\mathrm{feat}}\\) ensures alignment between intermediate features of \\(\\Psi\\) and \\(\\psi\\) as [4]:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.586, + 0.483, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\text {f e a t}} ^ {i} = \\sqrt {\\left\\| \\mathbf {f} _ {i} - \\mathbf {s} _ {i} \\right\\| _ {1} ^ {2} + \\gamma}, \\quad \\mathcal {L} _ {\\text {f e a t}} = \\sum_ {i} \\alpha_ {i} \\ell_ {\\text {f e a t}} ^ {i}, \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.624, + 0.484, + 0.684 + ], + "angle": 0, + "content": "where \\(\\ell_{\\mathrm{feat}}^{i}\\) is a feature loss for the \\(i\\)-th feature, \\(\\alpha_{i}\\) is a weight for the \\(i\\)-th feature, and \\(\\mathbf{f}_i\\) and \\(\\mathbf{s}_i\\) denote the \\(i\\)-th intermediate feature maps of \\(\\Psi\\) and \\(\\psi\\) respectively. The small constant \\(\\gamma\\) ensures smooth convergence during optimization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.217, + 0.714 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.722, + 0.422, + 0.739 + ], + "angle": 0, + "content": "4.1. Datasets, Metrics and Training Details" + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.756, + 0.482, + 0.855 + ], + "angle": 0, + "content": "
SatelliteWorldView-3QuickBirdGaoFen-2
Number of Band844
Spatial Resolution (m)PAN0.30.60.8
LRMS1.22.43.2
Radiometric Resolution (bit)111110
Number of (Train / Test) Images9,714 / 2017,139 / 2019,809 / 20
Patch SizePAN64×64×164×64×164×64×1
LRMS16×16×816×16×416×16×4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.865, + 0.483, + 0.892 + ], + "angle": 0, + "content": "Table 1. Detailed information of Worldview-3, QuickBird, and GaoFen-2 datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.365 + ], + "angle": 0, + "content": "Table 1 summarizes the characteristics of three datasets from PanCollection [11] for our experiments: WorldView-3 (WV3), QuickBird (QB), and GaoFen-2 (GF2). All experiments were conducted with a batch size of 32 for 300K iterations per dataset on an NVIDIA GeForce RTX 3090 GPU. The initial learning rate was set to \\(10^{-4}\\) with a decay factor of 0.5 applied every 10K iterations, using the AdamW optimizer [12]. For the diffusion process of our FSA-T and FSA-S, the total diffusion time step \\(T\\) was set to 500, with the DDIM [42] sampling strategy used to generate HRMS images in 25 steps. To evaluate the model performance, we employed widely accepted metrics that capture various aspects of image quality and similarity. For the reduced-resolution datasets of WV3, QB, and GF2, we used six standard metrics: PSNR [51], SSIM [51], SAM [62], ERGAS [46], SCC [16], and Q4/Q8 [44]. For the full-resolution datasets of WV3, QB, and GF2, we used HQNR [45] with \\(D_{\\lambda}\\), and \\(D_{S}\\) to comprehensively assess the results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.379, + 0.715, + 0.395 + ], + "angle": 0, + "content": "4.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.907, + 0.478 + ], + "angle": 0, + "content": "To ensure a thorough and equitable evaluation, we used the official codes for each compared method whenever available. For methods without official codes, we re-implemented them based on the papers [31, 56] to maintain fairness in comparison." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.48, + 0.909, + 0.752 + ], + "angle": 0, + "content": "Quantitative analysis. As shown in Table 2 and Table 3, our FSA-T achieves superior performance on the WV3, QB, and GF2 datasets compared to all other methods. The GF2 dataset contains relatively easier (less complex) images, so the teacher model can fully leverage its high-capacity features. In contrast, for the more challenging (more complex) WV3 and QB datasets, our FSA-S surpasses FSA-T across several metrics. This improvement represents the effectiveness of our uncertainty-aware knowledge distillation (U-Know) strategy, which enables the student model (FSA-S) to address the spatial weaknesses of the teacher model by focusing on uncertain regions. From this, the FSA-S can benefit more generalization from the knowledge distillation with uncertainty maps, leading to better performances than the FSA-T for the test datasets. Consequently, our FSA-S demonstrates an enhanced adaptability and robustness on complex datasets, where precise handling of spatial details (high frequency components) is essential." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.753, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Qualitative analysis. Fig. 4 shows the output HRMS images \\(\\hat{\\mathbf{I}}_{\\mathrm{MS}}^{\\mathrm{HR}}\\) in the first rows and the error maps (MAE images) between \\(\\hat{\\mathbf{I}}_{\\mathrm{MS}}^{\\mathrm{HR}}\\) and ground truth \\(\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{HR}}\\). As shown, our framework generates the results \\(\\hat{\\mathbf{I}}_{\\mathrm{MS}}^{\\mathrm{HR}}\\) closer to \\(\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{HR}}\\) in the reduced resolution GF2 data set. Furthermore, Fig. 1 demonstrates that even in the full-resolution comparison where ground truth is unavailable, our framework enables to generate more detailed and stable results compared to all other state-of-the-art approaches." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23074" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.101, + 0.089, + 0.903, + 0.27 + ], + "angle": 0, + "content": "
GF2 DatasetReduced-ResolutionFull-Resolution
TypesMethodsPublicationsPSNR↑SSIM↑SAM↓ERGAS↓SCC↑Q4↑Dλ↓Ds↓HQNR↑
Non-DMsPanNet [58]ICCV 201739.1970.9591.0501.0380.9750.9630.0200.0520.929
MSDCNN [61]JSTARS 201840.7300.9710.9460.8620.9830.9720.0260.0790.898
FusionNet [53]ICCV 202139.8660.9660.9710.9600.9800.9670.0340.1050.865
LAGConv [25]AAAI 202241.1470.9740.8860.8160.9850.9740.0300.0780.895
S2DBPN [64]TGRS 202342.6860.9800.7720.6860.9900.9810.0200.0460.935
DCPNet [68]TGRS 202442.3120.9790.8060.7240.9880.9800.0240.0240.953
CANConv [14]CVPR 202443.1660.9820.7220.6530.9910.9830.0190.0630.919
Diffusion ModelsPanDiff [31]TGRS 202342.8270.9800.7670.6740.9900.9810.0200.0450.936
TMDiff [56]TGRS 202441.8960.9770.7640.7540.9880.9790.0290.0300.942
FSA-SOurs44.5850.9860.6240.5480.9930.9870.0180.0370.944
FSA-TOurs44.7570.9880.6030.5370.9940.9880.0170.0290.953
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.273, + 0.905, + 0.302 + ], + "angle": 0, + "content": "Table 2. Comparison of different models on the GaoFen-2 (GF2) dataset. Blue indicates the second-best performance, while red highlights the best-performing model. Standard deviations of the metrics across test samples are provided in the Supplemental." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.314, + 0.903, + 0.479 + ], + "angle": 0, + "content": "
TypesMethodsWV3 Dataset (Reduced-Resolution)QB Dataset (Reduced-Resolution)
PSNR↑SSIM↑SAM↓ERGAS↓SCC↑Q8↑PSNR↑SSIM↑SAM↓ERGAS↓SCC↑Q4↑
Non-DMsPanNet [58]36.1480.9663.4022.5380.9790.91335.5630.9395.2734.8560.9660.911
MSDCNN [61]36.3290.9673.3002.4890.9790.91437.0400.9544.8284.0740.9770.925
FusionNet [53]36.5690.9683.1882.4280.9810.91636.8210.9524.8924.1830.9750.923
LAGConv [25]36.7320.9703.1532.3800.9810.91637.5650.9584.6823.8450.9800.930
S2DBPN [64]37.2160.9723.0192.2450.9850.91737.3140.9564.8493.9560.9800.928
DCPNet [68]37.0090.9723.0832.3010.9840.91538.0790.9634.4203.6180.9830.935
CANConv [14]37.4410.9732.9272.1630.9850.91837.7950.9604.5543.7400.9820.935
Diffusion ModelsPanDiff [31]37.0290.9713.0582.2760.9840.91337.8420.9594.6113.7230.9820.935
TMDiff [56]37.4770.9732.8852.1510.9860.91537.6420.9584.6273.8040.9810.930
FSA-S37.9300.9762.7972.0460.9880.92238.3610.9644.3373.5000.9840.938
FSA-T37.8940.9762.8012.0550.9870.92138.3430.9644.3493.5020.9850.938
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.482, + 0.907, + 0.511 + ], + "angle": 0, + "content": "Table 3. Performance comparison of different models on WV3 and QB datasets. Red and blue highlights the best- and 2nd-best performing models. The results for the full-resolution datasets and detailed results with standard deviations are provided in the Supplemental." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.525, + 0.905, + 0.72 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.722, + 0.907, + 0.767 + ], + "angle": 0, + "content": "Figure 4. Visual comparison of PAN-sharpening results on the reduced-resolution GF2 dataset. The first row shows RGB of outputs \\(\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{HR}}\\) and the second row displays Error Map, the difference between output \\(\\hat{\\mathbf{I}}_{\\mathrm{MS}}^{\\mathrm{HR}}\\) and ground truth \\(\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{HR}}\\). Both FSA-T and FSA-S achieve more detailed results compared to state-of-the-art models." + }, + { + "type": "table", + "bbox": [ + 0.127, + 0.783, + 0.455, + 0.842 + ], + "angle": 0, + "content": "
MethodsParams. (M)FLOPs (T)Time (s)Memory (GB)
PanDiff [31]12.5560.47119.5223.260
TMDiff [56]153.9395.51767.46110.483
FSA-T25.4921.40225.4955.910
FSA-S9.1150.34612.2872.136
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.852, + 0.483, + 0.88 + ], + "angle": 0, + "content": "Table 4. Computational complexity comparison for diffusion-based models. Best values are highlighted in red." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.786, + 0.907, + 0.894 + ], + "angle": 0, + "content": "Computational complexity analysis. Table 4 compares the computational cost of diffusion-based models, detailing parameters (Params.), floating-point operations (FLOPs), inference time (Time) and memory usage. This demonstrates the efficiency of our FSA-S over other models. In particular, the TMDiff has significantly higher numbers of parameters and FLOPs. The FSA-T, while computationally" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23075" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.123 + ], + "angle": 0, + "content": "intensive than the FSA-S, still provides a balanced trade-off between efficiency and performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.133, + 0.258, + 0.149 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.167, + 0.485, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.348, + 0.485, + 0.379 + ], + "angle": 0, + "content": "Figure 5. Visualization of uncertainty map \\(\\hat{\\theta}\\), Error Map, and ground truth \\(\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{HR}}\\) of reduced WV3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.389, + 0.484, + 0.527 + ], + "angle": 0, + "content": "Visualization of uncertainty map \\(\\hat{\\theta}\\) of \\(\\Psi\\). The uncertainty map \\(\\hat{\\theta}\\) learned through Eq. 4 shows that the areas with low uncertainty appear dark, while the areas with high uncertainty appear bright. Fig. 5 visualizes the uncertainty maps \\(\\hat{\\theta}\\), error maps and ground truth \\(\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{HR}}\\) for the reduced WV3 dataset. As shown in Fig. 5, it can be noted that the high-uncertainty regions mainly correspond to high-frequency areas such as challenging object edges. The learned uncertainty maps tend to resemble the error maps." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.539, + 0.482, + 0.6 + ], + "angle": 0, + "content": "
EncoderDecoderGF2 Dataset (Reduced-Resolution)
FFAHQFESAM↓ERGAS↓SCC↑Q4↑
0.654 ± 0.1120.661 ± 0.0760.992 ± 0.0010.986 ± 0.007
✓0.654 ± 0.1120.636 ± 0.0770.993 ± 0.0020.986 ± 0.007
✓0.617 ± 0.1170.556 ± 0.1040.993 ± 0.0020.987 ± 0.007
✓✓0.603 ± 0.1020.537 ± 0.0770.994 ± 0.0010.988 ± 0.006
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.609, + 0.484, + 0.641 + ], + "angle": 0, + "content": "Table 5. Comparison of Results with and without FFA and HQFE Blocks in FSA-T. Best values are highlighted in red." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Effect of FFA, HQFE blocks in FSA-T. Table 5 presents the effectiveness of FFA and HQFE. As noted, there are a significant improvement across all metrics. The performance gains are much greater than when FFA is only applied to the encoder or HQFE only to the decoder, as simultaneous condition injection creates a synergistic effect with both FFA and HQFE because FFA sets up initial styles or \"sketches\" for \\(\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{HR}}\\) images, followed by the decoder that restores frequency details via the HQFE. This step-by-step learning greatly improves overall performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.904 + ], + "angle": 0, + "content": "Impact of the U-Know loss. To verify the effectiveness of our uncertainty-aware knowledge distillation loss \\(\\mathcal{L}_{\\mathrm{U - Know}}\\) (U-Know loss), we trained the student model (FSA-S) using three losses: a widely used \\(L_{1}\\), a conventional knowledge distillation loss \\(\\mathcal{L}_{\\mathrm{KD}}\\) which is same from our method \\(\\mathcal{L}_{\\mathrm{U - Know}}\\) but without considering the uncertainty" + }, + { + "type": "table", + "bbox": [ + 0.551, + 0.089, + 0.878, + 0.153 + ], + "angle": 0, + "content": "
Loss func.GF2 Dataset (Full-Resolution)
Dλ↓Ds↓HQNR ↑
L10.026 ± 0.0140.040 ± 0.0170.935 ± 0.020
LKD0.025 ± 0.0150.038 ± 0.0160.938 ± 0.021
LU-Know0.018 ± 0.0110.037 ± 0.0070.944 ± 0.012
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.163, + 0.907, + 0.192 + ], + "angle": 0, + "content": "Table 6. Impact of our U-Know loss function design. Best values are highlighted in red." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.21, + 0.907, + 0.336 + ], + "angle": 0, + "content": "map, and our \\(\\mathcal{L}_{\\mathrm{U - Know}}\\). Table 6 shows the performance comparisons for the three losses. As shown, when the FSA-S was trained with \\(\\mathcal{L}_{\\mathrm{U - Know}}\\), it achieved the best performance across all three metrics. This highlights the effectiveness of \\(\\mathcal{L}_{\\mathrm{U - Know}}\\) that enables the FSA-S to better leverage the teacher (FSA-T)'s prior knowledge including the uncertainty maps \\(\\hat{\\theta}\\), output features \\(\\mathbf{f}_i\\), and output HRMS \\(\\hat{\\mathbf{I}}_{\\mathrm{MS}}^{\\mathrm{HR}}\\) for improved learning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.343, + 0.642, + 0.358 + ], + "angle": 0, + "content": "4.4. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.365, + 0.907, + 0.518 + ], + "angle": 0, + "content": "Inference time limitations. Despite U-Know-DiffPAN's superior performance across multiple metrics, one notable limitation lies in its inference speed, compared to nondiffusion-based models. Diffusion models unavoidably require multiple sampling steps during inference, making this process inevitably slower. This iterative approach, while being beneficial for image generation of high quality, limits the model's efficiency in real-time applications. Future work could focus on optimizing the inference process by pruning sampling steps, while maintaining performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.529, + 0.634, + 0.545 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.909, + 0.768 + ], + "angle": 0, + "content": "We introduced U-Know-DiffPAN, a novel diffusion-based PAN-sharpening framework that effectively leverages frequency information from PAN and LRMS inputs to generate high-quality HRMS images. Our Uncertainty-aware Knowledge Distillation (U-Know) strategy enables efficient feature transfer by distilling both frequency-rich features and uncertainty maps from the teacher model, enhancing the restoration performance for spatially weak regions while reducing computational costs for the student model (FSA-S). Extensive experiments on multiple datasets, including WV3, QB, and GF2, demonstrate that our U-Know-DiffPAN outperforms very recent state-of-the-art methods in PAN-sharpening, establishing a new benchmark for performance and efficiency in satellite image processing." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.779, + 0.686, + 0.796 + ], + "angle": 0, + "content": "6. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.803, + 0.909, + 0.889 + ], + "angle": 0, + "content": "This work was supported by National Research Foundation of Korea (NRF) grant funded by the Korean Government [Ministry of Science and ICT (Information and Communications Technology)] (Project Number: RS-2024-00338513, Project Title: AI-based Computer Vision Study for Satellite Image Processing and Analysis, \\(100\\%\\))." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23076" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Bruno Aiazzi, Luciano Alparone, Stefano Baronti, Andrea Garzelli, and Massimo Selva. Mtf-tailored multiscale fusion of high-resolution ms and pan imagery. Photogrammetric Engineering & Remote Sensing, 72(5):591-596, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.24 + ], + "angle": 0, + "content": "[2] Israa Amro, Javier Mateos, Miguel Vega, Rafael Molina, and Aggelos K Katsaggelos. A survey of classical methods and new trends in pansharpening of multispectral images. EURASIP Journal on Advances in Signal Processing, 2011:1-22, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.483, + 0.297 + ], + "angle": 0, + "content": "[3] Coloma Ballester, Vicent Caselles, Laura Igual, Joan Verdera, and Bernard Rouge. A variational model for \\(\\mathfrak{p}+\\) xs image fusion. International Journal of Computer Vision, 69:43-58, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.483, + 0.352 + ], + "angle": 0, + "content": "[4] Jonathan T Barron. A general and adaptive robust loss function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4331-4339, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.483, + 0.409 + ], + "angle": 0, + "content": "[5] Zihan Cao, Shiqi Cao, Liang-Jian Deng, Xiao Wu, Junming Hou, and Gemine Vivone. Diffusion model with disentangled modulations for sharpening multispectral and hyperspectral images. Information Fusion, 104:102158, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.41, + 0.482, + 0.45 + ], + "angle": 0, + "content": "[6] Enrico Celeghini, Manuel Gadella, and Mariano A del Olmo. Hermite functions and fourier series. Symmetry, 13(5):853, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.453, + 0.483, + 0.535 + ], + "angle": 0, + "content": "[7] Wei-Chun Chen, Chia-Che Chang, and Che-Rung Lee. Knowledge distillation with feature maps for image classification. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part III 14, pages 200–215. Springer, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.537, + 0.483, + 0.592 + ], + "angle": 0, + "content": "[8] Jaewan Choi, Kiyun Yu, and Yongil Kim. A new adaptive component-substitution-based satellite image fusion by using partial replacement. IEEE transactions on geoscience and remote sensing, 49(1):295-309, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.483, + 0.661 + ], + "angle": 0, + "content": "[9] Matteo Ciotola, Sergio Vitale, Antonio Mazza, Giovanni Poggi, and Giuseppe Scarpa. Pansharpening by convolutional neural networks in the full resolution framework. IEEE Transactions on Geoscience and Remote Sensing, 60: 1-17, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.483, + 0.732 + ], + "angle": 0, + "content": "[10] Matteo Ciotola, Giuseppe Guarino, Gemine Vivone, Giovanni Poggi, Jocelyn Chanussot, Antonio Plaza, and Giuseppe Scarpa. Hyperspectral pansharpening: Critical review, tools and future perspectives. arXiv preprint arXiv:2407.01355, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[11] Liang-Jian Deng, Gemine Vivone, Mercedes E Paoletti, Giuseppe Scarpa, Jiang He, Yongjun Zhang, Jocelyn Chanussot, and Antonio Plaza. Machine learning in pan-sharpening: A benchmark, from shallow to deep networks. IEEE Geoscience and Remote Sensing Magazine, 10(3): 279-315, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[12] P Kingma Diederik. Adam: A method for stochastic optimization. (No Title), 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Kento Doi and Akira Iwasaki. Sscnet: Spectral-spatial consistency optimization of cnn for pansharpening. In IGARSS 2019-2019 IEEE International Geoscience and Remote Sensing Symposium, pages 3141-3144. IEEE, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[14] Yule Duan, Xiao Wu, Haoyu Deng, and Liang-Jian Deng. Content-adaptive non-local convolution for remote sensing pansharpening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27738-27747, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.906, + 0.232 + ], + "angle": 0, + "content": "[15] Joan Duran, Antoni Buades, Bartomeu Coll, Catalina Sbert, and Gwendoline Blanchet. A survey of pansharpening methods with a new band-decoupled variational model. ISPRS Journal of Photogrammetry and Remote Sensing, 125:78-105, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[16] Andrea Garzelli and Filippo Nencini. Hypercomplex quality assessment of multi/hyperspectral images. IEEE Geoscience and Remote Sensing Letters, 6(4):662-665, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.275, + 0.905, + 0.344 + ], + "angle": 0, + "content": "[17] Anaïs Gastineau, Jean-François Aujol, Yannick Berthoumieu, and Christian Germain. A residual dense generative adversarial network for pansharpening with geometrical constraints. In 2020 IEEE International Conference on Image Processing (ICIP), pages 493-497. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.344, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[18] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.4, + 0.905, + 0.427 + ], + "angle": 0, + "content": "[19] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[20] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.456, + 0.905, + 0.497 + ], + "angle": 0, + "content": "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.498, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[22] Jianwen Hu, Pei Hu, Xudong Kang, Hui Zhang, and Shaosheng Fan. Pan-sharpening via multiscale dynamic convolutional neural network. IEEE Transactions on Geoscience and Remote Sensing, 59(3):2231-2244, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.554, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[23] Wei Huang, Liang Xiao, Zhihui Wei, Hongyi Liu, and Songze Tang. A new pan-sharpening method with deep neural networks. IEEE Geoscience and Remote Sensing Letters, 12(5):1037-1041, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.664 + ], + "angle": 0, + "content": "[24] Zeyi Huang, Yang Zou, BVK Kumar, and Dong Huang. Comprehensive attention self-distillation for weakly-supervised object detection. Advances in neural information processing systems, 33:16797-16807, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[25] Zi-Rong Jin, Tian-Jing Zhang, Tai-Xiang Jiang, Gemine Vivone, and Liang-Jian Deng. Lagconv: Local-context adaptive convolution kernels with global harmonic bias for pan-sharpening. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1113-1121, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[26] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[27] Jaehyup Lee, Soomin Seo, and Munchurl Kim. Sipsa-net: Shift-invariant pan sharpening with moving object alignment for satellite imagery. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10166-10174, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[28] Xiaobo Liu, Xiang Liu, Haoran Dai, Xudong Kang, Antonio Plaza, and Wenjie Zu. Mun-gan: A multiscale unsupervised network for remote sensing image pansharpening." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23077" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "IEEE Transactions on Geoscience and Remote Sensing, 61: 1-18, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[29] Laetitia Loncan, Luis B De Almeida, José M Bioucas-Dias, Xavier Briottet, Jocelyn Chanussot, Nicolas Dobigeon, Sophie Fabre, Wenzhi Liao, Giorgio A Licciardi, Miguel Simoes, et al. Hyperspectral pansharpening: A review. IEEE Geoscience and remote sensing magazine, 3(3):27-46, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[30] Jiayi Ma, Wei Yu, Chen Chen, Pengwei Liang, Xiaojie Guo, and Junjun Jiang. Pan-gan: An unsupervised pan-sharpening method for remote sensing image fusion. Information Fusion, 62:110-120, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[31] Qingyan Meng, Wenxu Shi, Sijia Li, and Linlin Zhang. Pandiff: A novel pansharpening method based on denoising diffusion probabilistic model. IEEE Transactions on Geoscience and Remote Sensing, 61:1-17, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[32] Guy P Nason and Bernard W Silverman. The stationary wavelet transform and some statistical applications. In Wavelets and statistics, pages 281-299. Springer, 1995. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.433 + ], + "angle": 0, + "content": "[33] Qian Ning, Weisheng Dong, Xin Li, Jinjian Wu, and Guangming Shi. Uncertainty-driven loss for single image superresolution. Advances in Neural Information Processing Systems, 34:16398-16409, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[34] Xavier Otazu, María González-Audímaca, Octavi Fors, and Jorge Núñez. Introduction of sensor spectral response into image fusion methods. application to wavelet-based methods. IEEE Transactions on Geoscience and Remote Sensing, 43(10):2376-2385, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.561 + ], + "angle": 0, + "content": "[35] Hao Phung, Quan Dao, and Anh Tran. Wavelet diffusion models are fast and scalable image generators. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10199-10208, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[37] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[38] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE transactions on pattern analysis and machine intelligence, 45(4):4713-4726, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[39] Maximilian Seitzer, Arash Tavakoli, Dimitrije Antic, and Georg Martius. On the pitfalls of heteroscedastic uncertainty estimation with probabilistic neural networks. arXiv preprint arXiv:2203.09168, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[40] Shuyao Shang, Zhengyang Shan, Guangxing Liu, LunQian Wang, XingHua Wang, Zekai Zhang, and Jinglin Zhang. Resdiff: Combining cnn and diffusion model for image" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "super-resolution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8975-8983, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[41] Mark J Shensa et al. The discrete wavelet transform: wedding the a trous and mallat algorithms. IEEE Transactions on signal processing, 40(10):2464-2482, 1992. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[42] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[43] Baoli Sun, Xinchen Ye, Baopu Li, Haojie Li, Zhihui Wang, and Rui Xu. Learning scene structure guidance via cross-task knowledge transfer for single depth super-resolution. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 7792-7801, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[44] Gemine Vivone, Luciano Alparone, Jocelyn Chanussot, Mauro Dalla Mura, Andrea Garzelli, Giorgio A Licciardi, Rocco Restaino, and Lucien Wald. A critical comparison among pansharpening algorithms. IEEE Transactions on Geoscience and Remote Sensing, 53(5):2565-2586, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[45] Gemine Vivone, Mauro Dalla Mura, Andrea Garzelli, Rocco Restaino, Giuseppe Scarpa, Magnus O Ulfarsson, Luciano Alparone, and Jocelyn Chanussot. A new benchmark based on recent advances in multispectral pansharpening: Revisiting pansharpening with classical and emerging pansharpening methods. IEEE Geoscience and Remote Sensing Magazine, 9(1):53-81, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.509 + ], + "angle": 0, + "content": "[46] Lucien Wald. Quality of high resolution synthesised images: Is there a simple criterion? In Third conference\" Fusion of Earth data: merging point measurements, raster maps and remotely sensed images\", pages 99-103. SEE/URISCA, 2000. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.567 + ], + "angle": 0, + "content": "[47] Lucien Wald, Thierry Ranchin, and Marc Mangolini. Fusion of satellite images of different spatial resolutions: Assessing the quality of resulting images. Photogrammetric engineering and remote sensing, 63(6):691-699, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.907, + 0.622 + ], + "angle": 0, + "content": "[48] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, pages 1-21, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.678 + ], + "angle": 0, + "content": "[49] Tao Wang, Li Yuan, Xiaopeng Zhang, and Jiashi Feng. Distilling object detectors with fine-grained feature imitation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4933-4942, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[50] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 606-615, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[51] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[52] Stephen Weinstein and Paul Ebert. Data transmission by frequency-division multiplexing using the discrete fourier transform. IEEE transactions on Communication Technology, 19(5):628-634, 1971. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[53] Xiao Wu, Ting-Zhu Huang, Liang-Jian Deng, and Tian-Jing Zhang. Dynamic cross feature fusion for remote sensing pan-sharpening. In Proceedings of the IEEE/CVF International" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23078" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "Conference on Computer Vision, pages 14687-14696, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.162 + ], + "angle": 0, + "content": "[54] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[55] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13095-13105, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[56] Yinghui Xing, Litao Qu, Shizhou Zhang, Jiapeng Feng, Xiuwei Zhang, and Yanning Zhang. Empower generalizability for pansharpening through text-modulated diffusion model. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.303, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[57] Kunran Xu, Lai Rui, Yishi Li, and Lin Gu. Feature normalized knowledge distillation for image classification. In European conference on computer vision, pages 664-680. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.482, + 0.424 + ], + "angle": 0, + "content": "[58] Junfeng Yang, Xueyang Fu, Yuwen Hu, Yue Huang, Xinghao Ding, and John Paisley. Pannet: A deep network architecture for pan-sharpening. In Proceedings of the IEEE international conference on computer vision, pages 5449-5457, 2017. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[59] Zhendong Yang, Zhe Li, Ailing Zeng, Zexian Li, Chun Yuan, and Yu Li. Vitkd: Feature-based knowledge distillation for vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1379-1388, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[60] Lewei Yao, Renjie Pi, Hang Xu, Wei Zhang, Zhenguo Li, and Tong Zhang. G-detkd: Towards general distillation framework for object detectors via contrastive and semantic-guided feature imitation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3591-3600, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.662 + ], + "angle": 0, + "content": "[61] Qiangqiang Yuan, Yancong Wei, Xiangchao Meng, Huanfeng Shen, and Liangpei Zhang. A multiscale and multidepth convolutional neural network for remote sensing imagery pan-sharpening. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 11(3):978-989, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[62] Roberta H Yuhas, Alexander FH Goetz, and Joe W Boardman. Discrimination among semi-arid landscape endmembers using the spectral angle mapper (sam) algorithm. In JPL, Summaries of the Third Annual JPL Airborne Geoscience Workshop. Volume 1: AVIRIS Workshop, 1992. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[63] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.873 + ], + "angle": 0, + "content": "[64] Kai Zhang, Anfei Wang, Feng Zhang, Wenbo Wan, Jiande Sun, and Lorenzo Bruzzone. Spatial-spectral dual backprojection network for pansharpening. IEEE Transactions on Geoscience and Remote Sensing, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[65] Yiman Zhang, Hanting Chen, Xinghao Chen, Yiping Deng, Chunjing Xu, and Yunhe Wang. Data-free knowledge dis" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "tillation for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7852-7861, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[66] Yun Zhang, Wei Li, Simiao Li, Hanting Chen, Zhijun Tu, Wenjia Wang, Bingyi Jing, Shaohui Lin, and Jie Hu. Data upcycling knowledge distillation for image super-resolution. arXiv preprint arXiv:2309.14162, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[67] Yaqun Zhang, Zhichang Guo, Dazhi Zhang, and Boying Wu. An anisotropic variational pansharpening model with adaptive coefficients. Inverse Problems and Imaging, pages 0–0, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[68] Yafei Zhang, Xuji Yang, Huafeng Li, Minghong Xie, and Zhengtao Yu. Dcpnet: A dual-task collaborative promotion network for pansharpening. IEEE Transactions on Geoscience and Remote Sensing, 62:1-16, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[69] Hao Zheng, Zhanlei Yang, Wenju Liu, Jizhong Liang, and Yanpeng Li. Improving deep neural networks using softplus units. In 2015 International joint conference on neural networks (IJCNN), pages 1-4. IEEE, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[70] Wu Zheng, Weiliang Tang, Li Jiang, and Chi-Wing Fu. Sessd: Self-ensembling single-stage object detector from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14494–14503, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[71] Jie Zhou, Daniel L Civco, and John A Silander. A wavelet transform method to merge landsat tm and spot panchromatic data. International journal of remote sensing, 19(4):743-757, 1998. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23079" + } + ] +] \ No newline at end of file diff --git a/2025/U-Know-DiffPAN_ An Uncertainty-aware Knowledge Distillation Diffusion Framework with Details Enhancement for PAN-Sharpening/ca20f1f2-4ee9-4b83-80ca-6e2c71410bad_origin.pdf b/2025/U-Know-DiffPAN_ An Uncertainty-aware Knowledge Distillation Diffusion Framework with Details Enhancement for PAN-Sharpening/ca20f1f2-4ee9-4b83-80ca-6e2c71410bad_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0ce42ca556466250722935812d25ab66c7e8c3a3 --- /dev/null +++ b/2025/U-Know-DiffPAN_ An Uncertainty-aware Knowledge Distillation Diffusion Framework with Details Enhancement for PAN-Sharpening/ca20f1f2-4ee9-4b83-80ca-6e2c71410bad_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e7aa1c94da18b1a4e5748e9ae495131601c9af4160f6132905e4149f5437007 +size 2512660 diff --git a/2025/U-Know-DiffPAN_ An Uncertainty-aware Knowledge Distillation Diffusion Framework with Details Enhancement for PAN-Sharpening/full.md b/2025/U-Know-DiffPAN_ An Uncertainty-aware Knowledge Distillation Diffusion Framework with Details Enhancement for PAN-Sharpening/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c220efb51e0bec4d0b727eb83d2c381b541fe8b1 --- /dev/null +++ b/2025/U-Know-DiffPAN_ An Uncertainty-aware Knowledge Distillation Diffusion Framework with Details Enhancement for PAN-Sharpening/full.md @@ -0,0 +1,429 @@ +# U-Know-DiffPAN: An Uncertainty-aware Knowledge Distillation Diffusion Framework with Details Enhancement for PAN-Sharpening + +Sungpyo Kim + +1 KAIST + +ksp04204@kaist.ac.kr + +Jeonghyeok Do + +$^{1}$ KAIST + +ehwjdgur0913@kaist.ac.kr + +Jaehyup Lee† + +$^{2}$ KNU + +jaehyuplee@knu.ac.kr + +Munchurl Kim + +1KAIST + +mkimee@kaist.ac.kr + +![](images/ab6af6b1aaff5f0f456ea32e538cf3b54298a70bd8c9b8f9355d561f7b00c523.jpg) +Figure 1. Visual comparison of PAN-sharpening results on the full-resolution WV3 dataset. The rightmost image shows the output of our proposed U-Know-DiffPAN framework, specifically the result produced by FSA-S (frequency selective attention student network). Notably, the proposed framework generates more detailed and robust results, particularly in high-uncertainty regions, outperforming the state-of-the-art model CANConv [14] and recent diffusion-based methods PanDiff [31] and TMDiff [56]. As highlighted in the red box, our approach successfully restores challenging highly-uncertain regions, such as cars, where other models fall short. + +![](images/34b1b64b9497bb181f9865af225cf03a9e90f436c317c1aff15eb4414724022a.jpg) +PanDiff (TGRS 2023) + +![](images/f53a2e345610728e2c67de6e7daf4d679ee27cbf2338295882bc27ab88d679f7.jpg) +TMDiff (TGRS 2024) + +![](images/1a235cba609c2ce645eeafd4d9897b9d6e1f8f2644e71fce7dcdabad8081ee1d.jpg) +CANConv (CVPR 2024) + +![](images/bca9a5628c4c47d655d5d2cacca7df7e1bf0a4a5036c8fcd975ce184b0c92d02.jpg) +U-Know-DiffPAN (Ours) + +# Abstract + +Conventional methods for PAN-sharpening often struggle to restore fine details due to limitations in leveraging high-frequency information. Moreover, diffusion-based approaches lack sufficient conditioning to fully utilize Panchromatic (PAN) images and low-resolution multispectral (LRMS) inputs effectively. To address these challenges, we propose an uncertainty-aware knowledge distillation diffusion framework with details enhancement for PAN-sharpening, called U-Know-DiffPAN. The U-Know-DiffPAN incorporates uncertainty-aware knowledge distillation for effective transfer of feature details from our teacher model to a student one. The teacher model in our U-Know-DiffPAN captures frequency details through frequency selective attention, facilitating accurate reverse process learning. By conditioning the encoder on compact vector representations of PAN and LRMS and the decoder on Wavelet transforms, we enable rich frequency utilization. So, the high-capacity teacher model distills frequency-rich features into a lightweight student model aided by an un + +certainty map. From this, the teacher model can guide the student model to focus on difficult image regions for PAN-sharpening via the usage of the uncertainty map. Extensive experiments on diverse datasets demonstrate the robustness and superior performance of our U-Know-DiffPAN over very recent state-of-the-art PAN-sharpening methods. The project page is available at https://kaist-viclab.github.io/U-Know-DiffPAN-site/. + +# 1. Introduction + +Satellite imagery plays a vital role in a wide range of applications, including environmental monitoring, surveillance, and mapping [2, 10, 15, 29, 67]. High-resolution multispectral (HRMS) images that combine high spatial and spectral fidelity are commonly generated using PAN-sharpening techniques, fusing low-resolution multispectral (LRMS) images with high-resolution (HR) panchromatic (PAN) images. The primary objective of PAN-sharpening is to retain the spatial detail of PAN images while preserving the spectral information of LRMS images. Recently, deep learning (DL)-based approaches have gained attraction in PAN-sharpening tasks, leveraging deep networks to perform end-to-end image regression. However, DL models + +are typically limited by the local receptive fields of convolutional layers, which primarily focus on local patterns and fail to capture long-range dependencies. This often leads to overly smooth outputs with reduced high-frequency details. + +To overcome this limitation, generative models such as generative adversarial networks (GANs) [18] have been applied to PAN-sharpening, aiming to produce more realistic results. Despite their success, GANs [28, 30] pose additional challenges, including training instability, mode collapse, and artifact generation, which can degrade image quality. Recently, diffusion models [21, 38, 42, 55] have demonstrated strong performance across a variety of image generation and restoration tasks. In PAN-sharpening, diffusion-based methods [31, 56] have shown promise by using PAN and LRMS images as conditioning inputs. However, they are limited by simple conditioning strategies that do not fully utilize the complementary information provided by the PAN and LRMS inputs, thus leading to suboptimal restoration quality. Additionally, they often incur high computational costs and longer inference times compared to other DL-based methods, limiting their practical applicability. + +To address these challenges, we propose U-Know-DiffPAN demonstrated in Fig. 2, a novel PAN-sharpening framework that incorporates an uncertainty-aware knowledge distillation (U-Know) strategy. Our framework is designed to fully leverage the frequency information in PAN and LRMS inputs while reducing computational costs. Our teacher model, is built to maximize information extraction from PAN and LRMS images by utilizing (i) a Feed Forward Attention (FFA) block in the encoder to capture compact vector representations of PAN and LRMS, and (ii) a High Quality Frequency Enhancement (HQFE) block in the decoder, composed of Fourier Transform Channel Attention (FTCA) and Stationary Wavelet Transform Cross Attention (SWTCA) modules. The FTCA enhances crucial frequency components through Fourier transform-based [52] attention, while the SwTCA incorporates high-frequency information from PAN and low-frequency information from LRMS into the decoder. + +Our U-Know strategy leverages uncertainty-aware knowledge distillation by transferring the features from the teacher model to a light-weight student model. The teacher model denoises HRMS images along with an uncertainty map, identifying spatial regions where the prediction confidence is low. Then based of the U-Know strategy, the teacher model distills to the student model not only the frequency-aware features but also the uncertainty map to address spatial weaknesses in the teacher's predictions. This approach leads to improved performance and reduced computational cost, making our U-Know-DiffPAN more effective and efficient for PAN-sharpening. Our contributions are summarized as follows: + +- We firstly incorporate the uncertainty-aware knowledge distillation into diffusion based PAN-sharpening with the guidance of pixel-wise restoration uncertainty and frequency-selective attention for details enhancement; +- Our teacher model leverages compact vector representations to enhance encoding efficiency and employs frequency selective attention in the decoder to refine frequency components for PAN-sharpening. +- We conduct extensive experiments on multiple datasets [11], including WV3, QB, and GF2, demonstrating state-of-the-art performance and surpassing very recent methods in PAN-sharpening with large margins. + +# 2. Related Works + +# 2.1. PAN-Sharpening + +PAN-sharpening is a technique to fuse high-resolution (HR) panchromatic (PAN) images with low-resolution multispectral (LRMS) images to produce HR multispectral (HRMS) images. Traditional PAN-sharpening techniques are broadly categorized into three approaches: component substitution (CS) [8, 41], multi-resolution analysis (MRA) [1, 34], and variational optimization (VO) [3]. Modern PAN-sharpening methods primarily consist of CNN-based [14, 25, 47, 58], transformer-based [25, 68], and GAN-based [28, 30] models. LAGConv [25] introduces local-context adaptive convolutional kernels with a global harmonic bias mechanism. S2DBPN [64] uses a spatial-spectral dual back-projection network for effective PAN and LRMS fusion. DCPNet [68] establishes a transformer-based dual-task parallel framework to optimize reconstruction. CANConv [14] integrates a non-local adaptive convolution module, enhancing spatial adaptability and non-local self-similarity. Besides the above methods, there are other DL-based methods for PAN-sharpening [9, 13, 17, 22, 23, 27]. While deep learning (DL)-based methods can capture complex mappings, they encounter challenges. CNN-based approaches tend to produce overly smooth results due to local receptive field of convolution layers. GAN-based models, though capable of generating realistic textures, may suffer from instability and introduce unrealistic artifacts. + +Diffusion-based approaches. Diffusion models (DMs) [21, 42] offer a promising alternative to conventional DL-based methods for image restoration [48, 55], iteratively refining noisy images through progressive denoising. Unlike GANs [18, 26], which can be prone to instability, DMs use iterative denoising steps with flexible conditioning, making them robust for tasks such as image restoration and PAN-sharpening. Recent diffusion-based PAN-sharpening models [31, 56] have shown significant advancements. PanDiff [31], however, employs a relatively simplistic conditioning approach that limits its ability to fully utilize the information from the input data, sometimes resulting in subopti + +mal restoration quality. In contrast, TMDiff [56] employs a complex conditioning mechanism, requiring an additional pretrained CLIP [36] text encoder, which adds unnecessary complexity without substantial performance gains. Our proposed framework, U-Know-DiffPAN, handles these limitations by integrating an advanced conditioning strategy that maximizes the use of input information. Specifically, our U-Know-DiffPAN conditions the encoder on compact vector representations of PAN and LRMS, while the decoder is conditioned on the stationary wavelet transform [32] of PAN and LRMS. This approach allows our U-Know-DiffPAN to effectively capture both frequency details and spatial fidelity, achieving superior restoration quality compared to existing diffusion-based methods. + +# 2.2. Knowledge Distillation + +Knowledge Distillation (KD) [20] is a well-known method for transferring knowledge from a large, model (teacher) to a smaller model (student), enabling deployment in resource-constrained settings while retaining performance. KD has demonstrated effectiveness across a range of computer vision tasks, including classification [7, 57, 65], object detection [24, 49, 60, 70], and image restoration [43, 59, 66]. However, previous KD methods, often fall short in tasks such as super-resolution [66], where capturing complicated detail is essential. This limitation arises because conventional KD student learns all regions equally, without focusing on areas that require concentrated learning. To address this, we propose uncertainty-aware knowledge distillation (U-Know), which introduces an uncertainty map into the distillation process. Our U-Know strategy informs the student of uncertain regions, areas the teacher model finds challenging to restore, enabling the student model to refine crucial details, thereby enhancing fidelity in both spatial and spectral domains. + +# 3. Methods + +# 3.1. Overview of U-Know-DiffPAN + +Fig. 2 shows the overview of our U-Know-DiffPAN framework. The proposed U-Know-DiffPAN framework comprises a teacher network with frequency-selective attention (FSA-T) and a student network (FSA-S), designed within a knowledge distillation paradigm. PAN image $\mathbf{I}_{\mathrm{PAN}}$ and interpolated LRMS image $\mathbf{I}_{\mathrm{MS}}^{\mathrm{LR}}$ are inputted in concatenation to generate HRMS image $\mathbf{I}_{\mathrm{MS}}^{\mathrm{HR}}$ . Both FSA-T and FSA-S are designed to predict the residuals as: + +$$ +\mathbf {X} _ {0} = \mathbf {I} _ {\mathrm {M S}} ^ {\mathrm {H R}} - \mathbf {I} _ {\mathrm {M S}} ^ {\mathrm {L R}}. \tag {1} +$$ + +Our framework operates in two main stages: (i) pre-training the FSA-T within a diffusion process to produce an initial prediction $\widehat{\mathbf{X}}_0$ alongside an uncertainty map $\hat{\theta}$ that identifies spatially weak regions; and (ii) training the FSA-S by + +Step1. Teacher (FSA-T) pre-training + +![](images/adf6fec8ea1d84b4cf8fefc26a393936306bd5d67eed577bfb76b3431c3d3979.jpg) + +![](images/12ab21e8c633dacbb7aa7e1862747fdac4bf95ef862d5c6b2ce0a8b1afa16e0f.jpg) + +Step2. Student (FSA-S) training + +![](images/1d50faf5ecd3ae8177e643efa50e922af91bfa781f53f1c509cbeeb87276ac20.jpg) +Figure 2. Overview of our uncertainty-aware knowledge-distillation diffusion framework with details enhancement, called U-Know-DiffPAN. + +![](images/59e338e4cde93290f2a51cd51b2a9ac6fc542f39081f58d6042213d0473a6363.jpg) + +leveraging this uncertainty map to guide the FSA-S in refining these regions through the KD. In Sec. 3.2, we describe the FSA-T's diffusion-based training process, which incorporates frequency information from PAN and LRMS inputs. Sec. 3.3 details the FSA-T architecture, specifically designed to enhance frequency components. Finally, in Sec. 3.4, we explain the FSA-S's uncertainty-aware knowledge distillation framework, which utilizes the uncertainty map to achieve precise and reliable PAN-sharpening. + +# 3.2. Diffusion Process of Teacher Network (FSA-T) + +Diffusion models operate by defining a forward process that gradually adds noise to data until it becomes indistinguishable from random noise, and a reverse process that learns to denoise the data, reconstructing the original information. We leverage this framework into a frequency-selective attention teacher network (FSA-T) as our teacher network, to generate frequency enhanced HRMS images with their pixel-wise uncertainty information. + +Forward process. In the forward diffusion process, noise is progressively added to the residual image, $\mathbf{X}_0$ , until it converges to a Gaussian distribution, $\mathbf{X}_T \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ , after $T$ timesteps. At each timestep $t \in \{1, \dots, T\}$ , a noisy version of the residual is obtained as follows [21]: + +$$ +\mathbf {X} _ {t} = \sqrt {\bar {\alpha} _ {t}} \mathbf {X} _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , \tag {2} +$$ + +where $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ is Gaussian noise, and $\bar{\alpha}_t = \prod_{s=1}^t (1 - \beta_s)$ controls the noise level at step $t$ . + +Reverse process. Following the DDPM [21], the reverse process aims to recover $\mathbf{X}_0$ from $\mathbf{X}_T$ by predicting a denoised image at each step. To perform denoising at each timestep, the denoising network (FSA-T), denoted as $\Psi$ , is designed to predict the original $\mathbf{X}_0$ rather than the added noise $\epsilon$ , as we found this empirically to be beneficial for PAN-sharpening. Additionally, $\Psi$ simultaneously produces an uncertainty map $\hat{\theta}$ for each prediction, providing pixelwise confidence information as: + +$$ +\left[ \widehat {\mathbf {X}} _ {0} \mid \widehat {\boldsymbol {\theta}} \right] = \Psi \left(\left[ \mathbf {X} _ {t} \mid \mathbf {I} _ {\text {P A N}} \mid \mathbf {I} _ {\mathrm {M S}} ^ {\mathrm {L R}} \right]; \mathbf {v}, \mathcal {S} - \text {C o n d}; t\right), \tag {3} +$$ + +![](images/a2a0e0bcf4075dbb7ba1c9fcdfa5dc1651fe7f943632525bd546219e47e12585.jpg) + +![](images/65a9812a2f6b025a690644bfb9952031ec187f6f7a90b6433f6aa4d6542c8409.jpg) + +![](images/5143c06fceb3b4937c9e978678ee64ec052fc45b53632a817a8f7e0c79468fcc.jpg) + +![](images/751d94a4da0449bbcb18709c4d3f90f85e7447266244962bf47db8460c990d8d.jpg) +Figure 3. Architecture of our proposed teacher model with frequency-selective attention, denoted as FSA-T, for PAN-sharpening. The FSA-T is designed to fully utilize frequency information for details enhancement from PAN and LRMS inputs (more details in Supplemental.) + +![](images/efa7c75e788eea5ee898f43fe4618f5b69483a314bdabbcaaeb018db2c433391.jpg) + +![](images/7dadd3b623aa519e83469df927456b6ccf6079903df4a319907b20ccd46a8687.jpg) + +![](images/f565785ec6e442234fbbb45b8791270223e45b9d510b80aa370ad2837cc31b2d.jpg) + +where $\left[\cdot \mid \cdot \right]$ indicates channel-wise concatenation, $\widehat{\mathbf{X}}_0$ is a predicted residual, $\hat{\theta}$ is a predicted uncertainty map, $\mathbf{v}$ and $S$ -Cond represent a learning compact representation (vector) and Stationary Wavelet Transform (SWT)-based conditioning terms, respectively. This formulation allows the FSA-T to iteratively refine the noisy input, guided by the conditioned information ( $\mathbf{v}$ , $S$ -Cond), toward an accurate reconstruction of the $\mathbf{X}_0$ . $\hat{\theta}$ captures pixel-wise prediction confidence, which guides the FSA-S to more rigorously learn the regions with high uncertainty. To ensure positivity, we apply the SoftPlus [69] activation to $\hat{\theta}$ . + +Uncertainty-driven diffusion loss. Inspired by [33, 39], we employ the following uncertainty-driven diffusion loss function, $\mathcal{L}_{\mathrm{U - Diff}}$ , to optimize the teacher network $\Psi$ : + +$$ +\mathcal {L} _ {\mathrm {U - D i f f}} = \left\| \frac {1}{2 \hat {\boldsymbol {\theta}}} \odot \left| \widehat {\mathbf {X}} _ {0} - \mathbf {X} _ {0} \right| + \frac {1}{2} \log \hat {\boldsymbol {\theta}} \right\| _ {1}, \tag {4} +$$ + +where $\hat{\theta}$ serves as the estimated variance term and is regarded as the uncertainty map in our framework. This objective function encourages the model to minimize prediction errors in low-confidence regions, thus yielding a more accurate and spatially robust output. + +# 3.3. Details of Teacher Network Architecture + +Our teacher network, FSA-T, is designed to enhance frequency details for PAN-sharpening tasks by fully utilizing complementary information from PAN and LRMS inputs. The FSA-T adopts an encoder-decoder (U-Net [37]) structure, where each stage is specifically tailored to capture and refine critical spatial and frequency features. The encoder of FSA-T employs Feed Forward Attention (FFA) blocks to extract compact and informative representations from the PAN and LRMS inputs. This attention mechanism allows + +the encoder to focus on essential features, preparing the data for effective frequency enhancement in subsequent layers. + +The decoder of FSA-T integrates High-Quality Frequency Enhancement (HQFE) blocks at each decoding level to further refine frequency components. Each HQFE block comprises two key modules: Fourier Transform Channel Attention (FTCA) that emphasizes important frequency components via Fourier Transform based attention, and Stationary Wavelet Transform Cross Attention (SWTCA) that merges high-frequency information from PAN and low-frequency content from LRMS, both of which are decomposed by SwT [32]. By incorporating these specialize components, FSA-T effectively produces $\widehat{\mathbf{X}}_0$ with enhanced spatial fidelity, establishing an effective uncertainty-aware knowledge distillation (U-Know) strategy. + +Feed Forward Attention (FFA). The FFA block in Fig. 3 is designed to leverage compact information [55] from the PAN and LRMS input to establish a rough structural representation during encoding. A compact vector representation $\mathbf{v} \in \mathbb{R}^{1 \times D}$ is obtained by a vector extractor $(\mathcal{E})$ as: + +$$ +\mathbf {v} = \mathcal {E} \left(\mathbf {I} _ {\mathrm {P A N}}, \mathbf {I} _ {\mathrm {M S}} ^ {\mathrm {L R}}\right). \tag {5} +$$ + +$\mathbf{v}$ is then used as a dynamic modulation [26, 50] parameter within the FFA, enabling $\mathcal{E}$ to incorporate $\mathbf{v}$ and establish an approximate structure during restoration as follows: + +$$ +[ \boldsymbol {\beta} _ {n} \mid \gamma_ {n} ] = \operatorname {L i n e a r} (\mathbf {v}), +$$ + +$$ +\mathbf {f} _ {n} ^ {\prime} = \gamma_ {n} \odot \mathrm {G N} \left(\mathbf {f} _ {n}\right) + \beta_ {n}, \tag {6} +$$ + +$$ +\mathbf {f} _ {n} \leftarrow \operatorname {C o n v} \left(\mathbf {f} _ {n} ^ {\prime}\right) \odot \operatorname {G E L U} \left(\operatorname {C o n v} \left(\mathbf {f} _ {n} ^ {\prime}\right)\right) + \mathbf {f} _ {n}, +$$ + +where $\mathbf{f}_n\in \mathbb{R}^{C\times h\times w}$ denotes input feature for the $n$ th encoder, Linear indicates Linear operation, $\odot$ denotes + +element-wise multiplication, GN indicates group normalization [54], GELU indicates GELU [19] activation and Conv consists of a sequential $1 \times 1$ convolution and a $3 \times 3$ depth-wise convolution layer (Details of $\mathcal{E}$ is in the Supplemental.). + +Fourier Transform Channel Attention (FTCA). The FTCA block in Fig. 3 performs self-attention in frequency domain by leveraging the 2D Fast Fourier Transform (FFT), denoted as $\mathcal{F}$ . $\mathcal{F}$ allows for efficient identification and emphasis of critical frequency components. For the $n$ -th decoder block in Fig. 3, the input feature $\mathbf{f}_n \in \mathbb{R}^{C \times h \times w}$ of $C$ channels is transformed into query $(\mathbf{Q})$ , key $(\mathbf{K})$ , and value $(\mathbf{V})$ representations. The 2D FFT $\mathcal{F}$ is then applied independently to each channel of $\mathbf{Q}$ , $\mathbf{K}$ , and $\mathbf{V}$ , mapping them transform from time domain to frequency domain as: + +$$ +[ \mathbf {Q} \mid \mathbf {K} \mid \mathbf {V} ] = \mathcal {F} (\operatorname {C o n v} \left(\mathbf {f} _ {n}\right)), \tag {7} +$$ + +where $\mathbf{Q},\mathbf{K}$ and $\mathbf{V}$ reside in $\mathbb{C}^{C\times h\times (\frac{w}{2} +1)}$ due to the Hermitian symmetry property [6]. Note that each frequency domain representation of $\mathbf{Q},\mathbf{K}$ , and $\mathbf{V}$ includes real and imaginary parts. So, we decompose $\mathbf{Q},\mathbf{K}$ , and $\mathbf{V}$ into real and imaginary parts as: + +$$ +\left[ \mathbf {Q} _ {\mathbb {R}} \mid \mathbf {K} _ {\mathbb {R}} \mid \mathbf {V} _ {\mathbb {R}} \right] + j \left[ \mathbf {Q} _ {\mathbb {I}} \mid \mathbf {K} _ {\mathbb {I}} \mid \mathbf {V} _ {\mathbb {I}} \right] = \left[ \mathbf {Q} \mid \mathbf {K} \mid \mathbf {V} \right], \tag {8} +$$ + +where $j = \sqrt{-1}$ , and $\mathbf{Q}_{\mathbb{R}}$ , $\mathbf{K}_{\mathbb{R}}$ , $\mathbf{V}_{\mathbb{R}}$ , $\mathbf{Q}_{\mathbb{I}}$ , $\mathbf{K}_{\mathbb{I}}$ , $\mathbf{V}_{\mathbb{I}}$ are in $\mathbb{R}^{C\times h\times (\frac{w}{2} + 1)}$ . Self-attention operations are then performed separately on the real and imaginary parts as follows: + +$$ +\mathbf {f} _ {l} = \operatorname {S o f t M a x} \left(\mathbf {Q} _ {l} \mathbf {K} _ {l} ^ {\top} / \sqrt {C}\right) \mathbf {V} _ {l}, \tag {9} +$$ + +where $l \in \{\mathbb{R},\mathbb{I}\}$ indicates the real or imaginary feature, respectively. Finally, the inverse Fourier Transform $\mathcal{F}^{-1}$ is applied to $\mathbf{f}_{\mathbb{R}}$ and $\mathbf{f}_{\mathbb{I}}$ to map them back to the time domain: + +$$ +\mathbf {f} _ {n} \leftarrow \mathcal {F} ^ {- 1} \left(\mathbf {f} _ {\mathbb {R}} + j \mathbf {f} _ {\mathbb {I}}\right). \tag {10} +$$ + +This process ensures that refined frequency information be selectively captured and utilized. + +Stationary Wavelet Transform Cross Attention (SWTCA). The SwTCA block is designed to leverage Wavelet components of $\mathbf{I}_{\mathrm{PAN}}$ and $\mathbf{I}_{\mathrm{MS}}^{\mathrm{LR}}$ motivated by [5, 35, 40, 71]. As shown in Fig. 3, it enhances the features outputted by the FTCA block, utilizing the Stationary Wavelet Transform (SWT) [32], denoted as $S$ . The SwTCA block selectively incorporates relevant frequency components from $\mathbf{I}_{\mathrm{PAN}}$ and $\mathbf{I}_{\mathrm{MS}}^{\mathrm{LR}}$ through a two-stage cross-attention process. The SwT-based conditioning terms ( $S$ -Cond) are constructed by decomposing $\mathbf{I}_{\mathrm{PAN}}$ and $\mathbf{I}_{\mathrm{MS}}^{\mathrm{LR}}$ using $S$ into four components: + +$$ +\begin{array}{r l} & {\left[ \mathbf {L} _ {\mathrm {P A N}} \mid \mathbf {H} _ {\mathrm {P A N}} \mid \mathbf {V} _ {\mathrm {P A N}} \mid \mathbf {D} _ {\mathrm {P A N}} \right] = \mathcal {S} (\mathbf {I} _ {\mathrm {P A N}}),} \\ & {\left[ \mathbf {L} _ {\mathrm {M S}} ^ {\mathrm {L R}} \mid \mathbf {H} _ {\mathrm {M S}} ^ {\mathrm {L R}} \mid \mathbf {V} _ {\mathrm {M S}} ^ {\mathrm {L R}} \mid \mathbf {D} _ {\mathrm {M S}} ^ {\mathrm {L R}} \right] = \mathcal {S} (\mathbf {I} _ {\mathrm {M S}} ^ {\mathrm {L R}}),} \end{array} \tag {11} +$$ + +where $\mathbf{L}$ represent the low-frequency approximation components, while $\mathbf{H},\mathbf{V},\mathbf{D}$ denote the high-frequency horizontal, vertical, and diagonal details, respectively. Since $\mathbf{I}_{\mathrm{PAN}}$ provides high spatial resolution but lacks spectral information, only its high-frequency components $\mathbf{H}_{\mathrm{PAN}},$ $\mathbf{V}_{\mathrm{PAN}}$ ,and $\mathbf{D}_{\mathrm{PAN}}$ are used. Conversely, $\mathbf{L}_{\mathrm{MS}}^{\mathrm{LR}}$ from $\mathbf{I}_{\mathrm{MS}}^{\mathrm{LR}}$ contains richer spectral information. These components are concatenated channel-wise to form $S$ -Cond as: + +$$ +\mathcal {S} - \text {C o n d} = \left[ \mathbf {L} _ {\mathrm {M S}} ^ {\mathrm {L R}} \mid \mathbf {H} _ {\mathrm {P A N}} \mid \mathbf {V} _ {\mathrm {P A N}} \mid \mathbf {D} _ {\mathrm {P A N}} \right]. \tag {12} +$$ + +To inject $S$ -Cond into the feature $\mathbf{f}_n$ obtained from FTCA at the $n$ -th decoder block, we employ a two-stage cross-attention process. In the first stage, $\mathbf{Q}_1$ and $\mathbf{K}_1$ are derived from $S$ -Cond, directing attention toward critical frequency components, while $\mathbf{V}_1$ is derived from $\mathbf{f}_n$ . The resulting features are reshaped to $C \times hw$ for channel attention [63], producing an intermediate feature $\mathbf{f}_n^1 \in \mathbb{R}^{C \times h \times w}$ as: + +$$ +\left[ \mathbf {Q} _ {1} \mid \mathbf {K} _ {1} \right] = \operatorname {C o n v} (S - \text {C o n d}), \mathbf {V} _ {1} = \operatorname {C o n v} (\mathbf {f} _ {n}), +$$ + +$$ +\mathbf {f} _ {n} ^ {1} = \operatorname {S o f t M a x} \left(\mathbf {Q} _ {1} \mathbf {K} _ {1} ^ {\top} / \sqrt {C}\right) \mathbf {V} _ {1}, \tag {13} +$$ + +where $\mathbf{Q}_1, \mathbf{K}_1, \mathbf{V}_1$ are in $\mathbb{R}^{C \times hw}$ . In the second stage, $\mathbf{Q}_2$ is derived from the intermediate feature $\mathbf{f}_n^1$ , while $\mathbf{K}_2$ and $\mathbf{V}_2$ come from $S$ -Cond, enabling the model to refine the correlation between the reconstructed feature $\mathbf{f}_n$ and significant frequency components in $S$ -Cond. This stage yields the final feature $\mathbf{f}_n \in \mathbb{R}^{C \times h \times w}$ after reshaping: + +$$ +\mathbf {Q} _ {2} = \operatorname {C o n v} \left(\mathbf {f} _ {n} ^ {1}\right), \left[ \mathbf {K} _ {2} \mid \mathbf {V} _ {2} \right] = \operatorname {C o n v} (S - C o n d), +$$ + +$$ +\mathbf {f} _ {n} \leftarrow \operatorname {S o f t M a x} \left(\mathbf {Q} _ {2} \mathbf {K} _ {2} ^ {\top} / \sqrt {C}\right) \mathbf {V} _ {2}. \tag {14} +$$ + +Through these two stages of cross-attention, the final feature $\mathbf{f}_n$ effectively incorporates essential frequency components from both PAN and MS inputs, ensuring enhanced frequency representation and high-quality HRMS output. + +# 3.4. Uncertainty-Aware Knowledge Distillation + +To address the substantial computational demands of the teacher (FSA-T) $\Psi$ , we propose an novel uncertainty-aware knowledge distillation (U-Know) strategy. + +Student network (FSA-S) $\psi$ . The student model FSA-S is designed with the same number of encoder and decoder blocks as FSA-T. Each block consists solely of Res-blocks, and unlike the teacher, the student operates without any additional conditional inputs. As shown in Fig. 2, the lightweight student model $\psi$ predicts the denoised image $\widetilde{\mathbf{X}}_0$ from the input $\mathbf{X}_t$ at timestep $t$ as: + +$$ +\widetilde {\mathbf {X}} _ {0} = \psi \left(\left[ \mathbf {X} _ {t} \mid \mathbf {I} _ {\text {P A N}} \mid \mathbf {I} _ {\mathrm {M S}} ^ {\mathrm {L R}} \right]; t\right). \tag {15} +$$ + +U-Know loss function. FSA-T $\Psi$ generates an uncertainty map $\hat{\theta}$ along with the output $\bar{\mathbf{X}}_0$ during training. + +These two outputs are frozen and used in the U-Know loss function, $\mathcal{L}_{\mathrm{U - Know}}$ , to train the FSA-S $\psi$ . The total training objective combines a hard loss $(\mathcal{L}_{\mathrm{hard}})$ , a soft loss $(\mathcal{L}_{\mathrm{soft}})$ , and a feature loss $(\mathcal{L}_{\mathrm{feat}})$ to distill frequency-rich features and enhance regions where $\Psi$ exhibits spatial weaknesses points. The total loss is defined as: + +$$ +\mathcal {L} _ {\mathrm {U - K n o w}} = \mathcal {L} _ {\text {h a r d}} + \lambda_ {s} \mathcal {L} _ {\text {s o f t}} + \lambda_ {f} \mathcal {L} _ {\text {f e a t}}, \tag {16} +$$ + +where $\lambda_{s}$ and $\lambda_{f}$ control the contributions of $\mathcal{L}_{\mathrm{soft}}$ and $\mathcal{L}_{\mathrm{feat}}$ respectively. $\mathcal{L}_{\mathrm{hard}}$ measures the difference between $\psi$ output $\widetilde{\mathbf{X}}_0$ and the ground truth $\mathbf{X}_0$ , weighted by $\Psi$ 's uncertainty map $\hat{\theta}$ to emphasize challenging regions: + +$$ +\mathcal {L} _ {\text {h a r d}} = \left\| (\tau + \hat {\boldsymbol {\theta}}) \odot \left| \widetilde {\mathbf {X}} _ {0} - \mathbf {X} _ {0} \right| \right\| _ {1}, \tag {17} +$$ + +where $\tau$ is a hyper-parameter. $\mathcal{L}_{\mathrm{soft}}$ computes the difference between $\psi$ 's output $\widetilde{\mathbf{X}}_0$ and $\Psi$ 's output $\widehat{\mathbf{X}}_0$ , weighted by the $\hat{\theta}$ to help $\psi$ better approximate the $\Psi$ 's knowledge: + +$$ +\mathcal {L} _ {\text {s o f t}} = \left\| (\tau - \hat {\boldsymbol {\theta}}) \odot \left| \widetilde {\mathbf {X}} _ {0} - \widehat {\mathbf {X}} _ {0} \right| \right\| _ {1}. \tag {18} +$$ + +In the regions where $\hat{\pmb{\theta}}$ values are high, indicating that $\Psi$ struggles with the task, we apply a weight of $(\tau + \hat{\pmb{\theta}})$ and use $\mathbf{X}_0$ as the hard ground truth to guide $\psi$ toward the true target. Conversely, in the areas where $\hat{\pmb{\theta}}$ values are low, indicating that $\Psi$ performs well, we apply a weight of $(\tau - \hat{\pmb{\theta}})$ and use $\Psi$ 's output $\widehat{\mathbf{X}}_0$ as the soft ground truth, allowing $\psi$ to approximate $\Psi$ 's knowledge in these regions. These dual loss formulations enable $\Psi$ to focus on challenging areas while effectively leveraging the $\Psi$ 's strengths. $\mathcal{L}_{\mathrm{feat}}$ ensures alignment between intermediate features of $\Psi$ and $\psi$ as [4]: + +$$ +\ell_ {\text {f e a t}} ^ {i} = \sqrt {\left\| \mathbf {f} _ {i} - \mathbf {s} _ {i} \right\| _ {1} ^ {2} + \gamma}, \quad \mathcal {L} _ {\text {f e a t}} = \sum_ {i} \alpha_ {i} \ell_ {\text {f e a t}} ^ {i}, \tag {19} +$$ + +where $\ell_{\mathrm{feat}}^{i}$ is a feature loss for the $i$ -th feature, $\alpha_{i}$ is a weight for the $i$ -th feature, and $\mathbf{f}_i$ and $\mathbf{s}_i$ denote the $i$ -th intermediate feature maps of $\Psi$ and $\psi$ respectively. The small constant $\gamma$ ensures smooth convergence during optimization. + +# 4. Experiment + +# 4.1. Datasets, Metrics and Training Details + +
SatelliteWorldView-3QuickBirdGaoFen-2
Number of Band844
Spatial Resolution (m)PAN0.30.60.8
LRMS1.22.43.2
Radiometric Resolution (bit)111110
Number of (Train / Test) Images9,714 / 2017,139 / 2019,809 / 20
Patch SizePAN64×64×164×64×164×64×1
LRMS16×16×816×16×416×16×4
+ +Table 1. Detailed information of Worldview-3, QuickBird, and GaoFen-2 datasets. + +Table 1 summarizes the characteristics of three datasets from PanCollection [11] for our experiments: WorldView-3 (WV3), QuickBird (QB), and GaoFen-2 (GF2). All experiments were conducted with a batch size of 32 for 300K iterations per dataset on an NVIDIA GeForce RTX 3090 GPU. The initial learning rate was set to $10^{-4}$ with a decay factor of 0.5 applied every 10K iterations, using the AdamW optimizer [12]. For the diffusion process of our FSA-T and FSA-S, the total diffusion time step $T$ was set to 500, with the DDIM [42] sampling strategy used to generate HRMS images in 25 steps. To evaluate the model performance, we employed widely accepted metrics that capture various aspects of image quality and similarity. For the reduced-resolution datasets of WV3, QB, and GF2, we used six standard metrics: PSNR [51], SSIM [51], SAM [62], ERGAS [46], SCC [16], and Q4/Q8 [44]. For the full-resolution datasets of WV3, QB, and GF2, we used HQNR [45] with $D_{\lambda}$ , and $D_{S}$ to comprehensively assess the results. + +# 4.2. Experimental Results + +To ensure a thorough and equitable evaluation, we used the official codes for each compared method whenever available. For methods without official codes, we re-implemented them based on the papers [31, 56] to maintain fairness in comparison. + +Quantitative analysis. As shown in Table 2 and Table 3, our FSA-T achieves superior performance on the WV3, QB, and GF2 datasets compared to all other methods. The GF2 dataset contains relatively easier (less complex) images, so the teacher model can fully leverage its high-capacity features. In contrast, for the more challenging (more complex) WV3 and QB datasets, our FSA-S surpasses FSA-T across several metrics. This improvement represents the effectiveness of our uncertainty-aware knowledge distillation (U-Know) strategy, which enables the student model (FSA-S) to address the spatial weaknesses of the teacher model by focusing on uncertain regions. From this, the FSA-S can benefit more generalization from the knowledge distillation with uncertainty maps, leading to better performances than the FSA-T for the test datasets. Consequently, our FSA-S demonstrates an enhanced adaptability and robustness on complex datasets, where precise handling of spatial details (high frequency components) is essential. + +Qualitative analysis. Fig. 4 shows the output HRMS images $\hat{\mathbf{I}}_{\mathrm{MS}}^{\mathrm{HR}}$ in the first rows and the error maps (MAE images) between $\hat{\mathbf{I}}_{\mathrm{MS}}^{\mathrm{HR}}$ and ground truth $\mathbf{I}_{\mathrm{MS}}^{\mathrm{HR}}$ . As shown, our framework generates the results $\hat{\mathbf{I}}_{\mathrm{MS}}^{\mathrm{HR}}$ closer to $\mathbf{I}_{\mathrm{MS}}^{\mathrm{HR}}$ in the reduced resolution GF2 data set. Furthermore, Fig. 1 demonstrates that even in the full-resolution comparison where ground truth is unavailable, our framework enables to generate more detailed and stable results compared to all other state-of-the-art approaches. + +
GF2 DatasetReduced-ResolutionFull-Resolution
TypesMethodsPublicationsPSNR↑SSIM↑SAM↓ERGAS↓SCC↑Q4↑Dλ↓Ds↓HQNR↑
Non-DMsPanNet [58]ICCV 201739.1970.9591.0501.0380.9750.9630.0200.0520.929
MSDCNN [61]JSTARS 201840.7300.9710.9460.8620.9830.9720.0260.0790.898
FusionNet [53]ICCV 202139.8660.9660.9710.9600.9800.9670.0340.1050.865
LAGConv [25]AAAI 202241.1470.9740.8860.8160.9850.9740.0300.0780.895
S2DBPN [64]TGRS 202342.6860.9800.7720.6860.9900.9810.0200.0460.935
DCPNet [68]TGRS 202442.3120.9790.8060.7240.9880.9800.0240.0240.953
CANConv [14]CVPR 202443.1660.9820.7220.6530.9910.9830.0190.0630.919
Diffusion ModelsPanDiff [31]TGRS 202342.8270.9800.7670.6740.9900.9810.0200.0450.936
TMDiff [56]TGRS 202441.8960.9770.7640.7540.9880.9790.0290.0300.942
FSA-SOurs44.5850.9860.6240.5480.9930.9870.0180.0370.944
FSA-TOurs44.7570.9880.6030.5370.9940.9880.0170.0290.953
+ +Table 2. Comparison of different models on the GaoFen-2 (GF2) dataset. Blue indicates the second-best performance, while red highlights the best-performing model. Standard deviations of the metrics across test samples are provided in the Supplemental. + +
TypesMethodsWV3 Dataset (Reduced-Resolution)QB Dataset (Reduced-Resolution)
PSNR↑SSIM↑SAM↓ERGAS↓SCC↑Q8↑PSNR↑SSIM↑SAM↓ERGAS↓SCC↑Q4↑
Non-DMsPanNet [58]36.1480.9663.4022.5380.9790.91335.5630.9395.2734.8560.9660.911
MSDCNN [61]36.3290.9673.3002.4890.9790.91437.0400.9544.8284.0740.9770.925
FusionNet [53]36.5690.9683.1882.4280.9810.91636.8210.9524.8924.1830.9750.923
LAGConv [25]36.7320.9703.1532.3800.9810.91637.5650.9584.6823.8450.9800.930
S2DBPN [64]37.2160.9723.0192.2450.9850.91737.3140.9564.8493.9560.9800.928
DCPNet [68]37.0090.9723.0832.3010.9840.91538.0790.9634.4203.6180.9830.935
CANConv [14]37.4410.9732.9272.1630.9850.91837.7950.9604.5543.7400.9820.935
Diffusion ModelsPanDiff [31]37.0290.9713.0582.2760.9840.91337.8420.9594.6113.7230.9820.935
TMDiff [56]37.4770.9732.8852.1510.9860.91537.6420.9584.6273.8040.9810.930
FSA-S37.9300.9762.7972.0460.9880.92238.3610.9644.3373.5000.9840.938
FSA-T37.8940.9762.8012.0550.9870.92138.3430.9644.3493.5020.9850.938
+ +![](images/89a59f39da9ed831a7d68e9e447c5fd4358c7edd9c0f3eb790a8d2b40f8e9a1e.jpg) +Figure 4. Visual comparison of PAN-sharpening results on the reduced-resolution GF2 dataset. The first row shows RGB of outputs $\mathbf{I}_{\mathrm{MS}}^{\mathrm{HR}}$ and the second row displays Error Map, the difference between output $\hat{\mathbf{I}}_{\mathrm{MS}}^{\mathrm{HR}}$ and ground truth $\mathbf{I}_{\mathrm{MS}}^{\mathrm{HR}}$ . Both FSA-T and FSA-S achieve more detailed results compared to state-of-the-art models. + +Table 3. Performance comparison of different models on WV3 and QB datasets. Red and blue highlights the best- and 2nd-best performing models. The results for the full-resolution datasets and detailed results with standard deviations are provided in the Supplemental. + +
MethodsParams. (M)FLOPs (T)Time (s)Memory (GB)
PanDiff [31]12.5560.47119.5223.260
TMDiff [56]153.9395.51767.46110.483
FSA-T25.4921.40225.4955.910
FSA-S9.1150.34612.2872.136
+ +Table 4. Computational complexity comparison for diffusion-based models. Best values are highlighted in red. + +Computational complexity analysis. Table 4 compares the computational cost of diffusion-based models, detailing parameters (Params.), floating-point operations (FLOPs), inference time (Time) and memory usage. This demonstrates the efficiency of our FSA-S over other models. In particular, the TMDiff has significantly higher numbers of parameters and FLOPs. The FSA-T, while computationally + +intensive than the FSA-S, still provides a balanced trade-off between efficiency and performance. + +# 4.3. Ablation Studies + +![](images/419ae6fa61c9328953033815efb5e94c57c0cb2454f4265898c0e5e772a63e4f.jpg) +Figure 5. Visualization of uncertainty map $\hat{\theta}$ , Error Map, and ground truth $\mathbf{I}_{\mathrm{MS}}^{\mathrm{HR}}$ of reduced WV3. + +Visualization of uncertainty map $\hat{\theta}$ of $\Psi$ . The uncertainty map $\hat{\theta}$ learned through Eq. 4 shows that the areas with low uncertainty appear dark, while the areas with high uncertainty appear bright. Fig. 5 visualizes the uncertainty maps $\hat{\theta}$ , error maps and ground truth $\mathbf{I}_{\mathrm{MS}}^{\mathrm{HR}}$ for the reduced WV3 dataset. As shown in Fig. 5, it can be noted that the high-uncertainty regions mainly correspond to high-frequency areas such as challenging object edges. The learned uncertainty maps tend to resemble the error maps. + +
EncoderDecoderGF2 Dataset (Reduced-Resolution)
FFAHQFESAM↓ERGAS↓SCC↑Q4↑
0.654 ± 0.1120.661 ± 0.0760.992 ± 0.0010.986 ± 0.007
✓0.654 ± 0.1120.636 ± 0.0770.993 ± 0.0020.986 ± 0.007
✓0.617 ± 0.1170.556 ± 0.1040.993 ± 0.0020.987 ± 0.007
✓✓0.603 ± 0.1020.537 ± 0.0770.994 ± 0.0010.988 ± 0.006
+ +Effect of FFA, HQFE blocks in FSA-T. Table 5 presents the effectiveness of FFA and HQFE. As noted, there are a significant improvement across all metrics. The performance gains are much greater than when FFA is only applied to the encoder or HQFE only to the decoder, as simultaneous condition injection creates a synergistic effect with both FFA and HQFE because FFA sets up initial styles or "sketches" for $\mathbf{I}_{\mathrm{MS}}^{\mathrm{HR}}$ images, followed by the decoder that restores frequency details via the HQFE. This step-by-step learning greatly improves overall performance. + +Impact of the U-Know loss. To verify the effectiveness of our uncertainty-aware knowledge distillation loss $\mathcal{L}_{\mathrm{U - Know}}$ (U-Know loss), we trained the student model (FSA-S) using three losses: a widely used $L_{1}$ , a conventional knowledge distillation loss $\mathcal{L}_{\mathrm{KD}}$ which is same from our method $\mathcal{L}_{\mathrm{U - Know}}$ but without considering the uncertainty + +Table 5. Comparison of Results with and without FFA and HQFE Blocks in FSA-T. Best values are highlighted in red. + +
Loss func.GF2 Dataset (Full-Resolution)
Dλ↓Ds↓HQNR ↑
L10.026 ± 0.0140.040 ± 0.0170.935 ± 0.020
LKD0.025 ± 0.0150.038 ± 0.0160.938 ± 0.021
LU-Know0.018 ± 0.0110.037 ± 0.0070.944 ± 0.012
+ +Table 6. Impact of our U-Know loss function design. Best values are highlighted in red. + +map, and our $\mathcal{L}_{\mathrm{U - Know}}$ . Table 6 shows the performance comparisons for the three losses. As shown, when the FSA-S was trained with $\mathcal{L}_{\mathrm{U - Know}}$ , it achieved the best performance across all three metrics. This highlights the effectiveness of $\mathcal{L}_{\mathrm{U - Know}}$ that enables the FSA-S to better leverage the teacher (FSA-T)'s prior knowledge including the uncertainty maps $\hat{\theta}$ , output features $\mathbf{f}_i$ , and output HRMS $\hat{\mathbf{I}}_{\mathrm{MS}}^{\mathrm{HR}}$ for improved learning. + +# 4.4. Limitations + +Inference time limitations. Despite U-Know-DiffPAN's superior performance across multiple metrics, one notable limitation lies in its inference speed, compared to nondiffusion-based models. Diffusion models unavoidably require multiple sampling steps during inference, making this process inevitably slower. This iterative approach, while being beneficial for image generation of high quality, limits the model's efficiency in real-time applications. Future work could focus on optimizing the inference process by pruning sampling steps, while maintaining performance. + +# 5. Conclusion + +We introduced U-Know-DiffPAN, a novel diffusion-based PAN-sharpening framework that effectively leverages frequency information from PAN and LRMS inputs to generate high-quality HRMS images. Our Uncertainty-aware Knowledge Distillation (U-Know) strategy enables efficient feature transfer by distilling both frequency-rich features and uncertainty maps from the teacher model, enhancing the restoration performance for spatially weak regions while reducing computational costs for the student model (FSA-S). Extensive experiments on multiple datasets, including WV3, QB, and GF2, demonstrate that our U-Know-DiffPAN outperforms very recent state-of-the-art methods in PAN-sharpening, establishing a new benchmark for performance and efficiency in satellite image processing. + +# 6. Acknowledgment + +This work was supported by National Research Foundation of Korea (NRF) grant funded by the Korean Government [Ministry of Science and ICT (Information and Communications Technology)] (Project Number: RS-2024-00338513, Project Title: AI-based Computer Vision Study for Satellite Image Processing and Analysis, $100\%$ ). + +# References + +[1] Bruno Aiazzi, Luciano Alparone, Stefano Baronti, Andrea Garzelli, and Massimo Selva. Mtf-tailored multiscale fusion of high-resolution ms and pan imagery. Photogrammetric Engineering & Remote Sensing, 72(5):591-596, 2006. 2 +[2] Israa Amro, Javier Mateos, Miguel Vega, Rafael Molina, and Aggelos K Katsaggelos. A survey of classical methods and new trends in pansharpening of multispectral images. EURASIP Journal on Advances in Signal Processing, 2011:1-22, 2011. 1 +[3] Coloma Ballester, Vicent Caselles, Laura Igual, Joan Verdera, and Bernard Rouge. A variational model for $\mathfrak{p}+$ xs image fusion. International Journal of Computer Vision, 69:43-58, 2006. 2 +[4] Jonathan T Barron. A general and adaptive robust loss function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4331-4339, 2019. 6 +[5] Zihan Cao, Shiqi Cao, Liang-Jian Deng, Xiao Wu, Junming Hou, and Gemine Vivone. Diffusion model with disentangled modulations for sharpening multispectral and hyperspectral images. Information Fusion, 104:102158, 2024. 5 +[6] Enrico Celeghini, Manuel Gadella, and Mariano A del Olmo. Hermite functions and fourier series. Symmetry, 13(5):853, 2021. 5 +[7] Wei-Chun Chen, Chia-Che Chang, and Che-Rung Lee. Knowledge distillation with feature maps for image classification. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part III 14, pages 200–215. Springer, 2019. 3 +[8] Jaewan Choi, Kiyun Yu, and Yongil Kim. A new adaptive component-substitution-based satellite image fusion by using partial replacement. IEEE transactions on geoscience and remote sensing, 49(1):295-309, 2010. 2 +[9] Matteo Ciotola, Sergio Vitale, Antonio Mazza, Giovanni Poggi, and Giuseppe Scarpa. Pansharpening by convolutional neural networks in the full resolution framework. IEEE Transactions on Geoscience and Remote Sensing, 60: 1-17, 2022. 2 +[10] Matteo Ciotola, Giuseppe Guarino, Gemine Vivone, Giovanni Poggi, Jocelyn Chanussot, Antonio Plaza, and Giuseppe Scarpa. Hyperspectral pansharpening: Critical review, tools and future perspectives. arXiv preprint arXiv:2407.01355, 2024. 1 +[11] Liang-Jian Deng, Gemine Vivone, Mercedes E Paoletti, Giuseppe Scarpa, Jiang He, Yongjun Zhang, Jocelyn Chanussot, and Antonio Plaza. Machine learning in pan-sharpening: A benchmark, from shallow to deep networks. IEEE Geoscience and Remote Sensing Magazine, 10(3): 279-315, 2022. 2, 6 +[12] P Kingma Diederik. Adam: A method for stochastic optimization. (No Title), 2014. 6 +[13] Kento Doi and Akira Iwasaki. Sscnet: Spectral-spatial consistency optimization of cnn for pansharpening. In IGARSS 2019-2019 IEEE International Geoscience and Remote Sensing Symposium, pages 3141-3144. IEEE, 2019. 2 + +[14] Yule Duan, Xiao Wu, Haoyu Deng, and Liang-Jian Deng. Content-adaptive non-local convolution for remote sensing pansharpening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27738-27747, 2024. 1, 2, 7 +[15] Joan Duran, Antoni Buades, Bartomeu Coll, Catalina Sbert, and Gwendoline Blanchet. A survey of pansharpening methods with a new band-decoupled variational model. ISPRS Journal of Photogrammetry and Remote Sensing, 125:78-105, 2017. 1 +[16] Andrea Garzelli and Filippo Nencini. Hypercomplex quality assessment of multi/hyperspectral images. IEEE Geoscience and Remote Sensing Letters, 6(4):662-665, 2009. 6 +[17] Anaïs Gastineau, Jean-François Aujol, Yannick Berthoumieu, and Christian Germain. A residual dense generative adversarial network for pansharpening with geometrical constraints. In 2020 IEEE International Conference on Image Processing (ICIP), pages 493-497. IEEE, 2020. 2 +[18] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2 +[19] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5 +[20] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3 +[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3 +[22] Jianwen Hu, Pei Hu, Xudong Kang, Hui Zhang, and Shaosheng Fan. Pan-sharpening via multiscale dynamic convolutional neural network. IEEE Transactions on Geoscience and Remote Sensing, 59(3):2231-2244, 2020. 2 +[23] Wei Huang, Liang Xiao, Zhihui Wei, Hongyi Liu, and Songze Tang. A new pan-sharpening method with deep neural networks. IEEE Geoscience and Remote Sensing Letters, 12(5):1037-1041, 2015. 2 +[24] Zeyi Huang, Yang Zou, BVK Kumar, and Dong Huang. Comprehensive attention self-distillation for weakly-supervised object detection. Advances in neural information processing systems, 33:16797-16807, 2020. 3 +[25] Zi-Rong Jin, Tian-Jing Zhang, Tai-Xiang Jiang, Gemine Vivone, and Liang-Jian Deng. Lagconv: Local-context adaptive convolution kernels with global harmonic bias for pan-sharpening. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1113-1121, 2022. 2, 7 +[26] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2, 4 +[27] Jaehyup Lee, Soomin Seo, and Munchurl Kim. Sipsa-net: Shift-invariant pan sharpening with moving object alignment for satellite imagery. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10166-10174, 2021. 2 +[28] Xiaobo Liu, Xiang Liu, Haoran Dai, Xudong Kang, Antonio Plaza, and Wenjie Zu. Mun-gan: A multiscale unsupervised network for remote sensing image pansharpening. + +IEEE Transactions on Geoscience and Remote Sensing, 61: 1-18, 2023. 2 +[29] Laetitia Loncan, Luis B De Almeida, José M Bioucas-Dias, Xavier Briottet, Jocelyn Chanussot, Nicolas Dobigeon, Sophie Fabre, Wenzhi Liao, Giorgio A Licciardi, Miguel Simoes, et al. Hyperspectral pansharpening: A review. IEEE Geoscience and remote sensing magazine, 3(3):27-46, 2015. 1 +[30] Jiayi Ma, Wei Yu, Chen Chen, Pengwei Liang, Xiaojie Guo, and Junjun Jiang. Pan-gan: An unsupervised pan-sharpening method for remote sensing image fusion. Information Fusion, 62:110-120, 2020. 2 +[31] Qingyan Meng, Wenxu Shi, Sijia Li, and Linlin Zhang. Pandiff: A novel pansharpening method based on denoising diffusion probabilistic model. IEEE Transactions on Geoscience and Remote Sensing, 61:1-17, 2023. 1, 2, 6, 7 +[32] Guy P Nason and Bernard W Silverman. The stationary wavelet transform and some statistical applications. In Wavelets and statistics, pages 281-299. Springer, 1995. 3, 4, 5 +[33] Qian Ning, Weisheng Dong, Xin Li, Jinjian Wu, and Guangming Shi. Uncertainty-driven loss for single image superresolution. Advances in Neural Information Processing Systems, 34:16398-16409, 2021. 4 +[34] Xavier Otazu, María González-Audímaca, Octavi Fors, and Jorge Núñez. Introduction of sensor spectral response into image fusion methods. application to wavelet-based methods. IEEE Transactions on Geoscience and Remote Sensing, 43(10):2376-2385, 2005. 2 +[35] Hao Phung, Quan Dao, and Anh Tran. Wavelet diffusion models are fast and scalable image generators. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10199-10208, 2023. 5 +[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3 +[37] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4 +[38] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE transactions on pattern analysis and machine intelligence, 45(4):4713-4726, 2022. 2 +[39] Maximilian Seitzer, Arash Tavakoli, Dimitrije Antic, and Georg Martius. On the pitfalls of heteroscedastic uncertainty estimation with probabilistic neural networks. arXiv preprint arXiv:2203.09168, 2022. 4 +[40] Shuyao Shang, Zhengyang Shan, Guangxing Liu, LunQian Wang, XingHua Wang, Zekai Zhang, and Jinglin Zhang. Resdiff: Combining cnn and diffusion model for image + +super-resolution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8975-8983, 2024. 5 +[41] Mark J Shensa et al. The discrete wavelet transform: wedding the a trous and mallat algorithms. IEEE Transactions on signal processing, 40(10):2464-2482, 1992. 2 +[42] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 6 +[43] Baoli Sun, Xinchen Ye, Baopu Li, Haojie Li, Zhihui Wang, and Rui Xu. Learning scene structure guidance via cross-task knowledge transfer for single depth super-resolution. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 7792-7801, 2021. 3 +[44] Gemine Vivone, Luciano Alparone, Jocelyn Chanussot, Mauro Dalla Mura, Andrea Garzelli, Giorgio A Licciardi, Rocco Restaino, and Lucien Wald. A critical comparison among pansharpening algorithms. IEEE Transactions on Geoscience and Remote Sensing, 53(5):2565-2586, 2014. 6 +[45] Gemine Vivone, Mauro Dalla Mura, Andrea Garzelli, Rocco Restaino, Giuseppe Scarpa, Magnus O Ulfarsson, Luciano Alparone, and Jocelyn Chanussot. A new benchmark based on recent advances in multispectral pansharpening: Revisiting pansharpening with classical and emerging pansharpening methods. IEEE Geoscience and Remote Sensing Magazine, 9(1):53-81, 2020. 6 +[46] Lucien Wald. Quality of high resolution synthesised images: Is there a simple criterion? In Third conference" Fusion of Earth data: merging point measurements, raster maps and remotely sensed images", pages 99-103. SEE/URISCA, 2000. 6 +[47] Lucien Wald, Thierry Ranchin, and Marc Mangolini. Fusion of satellite images of different spatial resolutions: Assessing the quality of resulting images. Photogrammetric engineering and remote sensing, 63(6):691-699, 1997. 2 +[48] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, pages 1-21, 2024. 2 +[49] Tao Wang, Li Yuan, Xiaopeng Zhang, and Jiashi Feng. Distilling object detectors with fine-grained feature imitation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4933-4942, 2019. 3 +[50] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 606-615, 2018. 4 +[51] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6 +[52] Stephen Weinstein and Paul Ebert. Data transmission by frequency-division multiplexing using the discrete fourier transform. IEEE transactions on Communication Technology, 19(5):628-634, 1971. 2 +[53] Xiao Wu, Ting-Zhu Huang, Liang-Jian Deng, and Tian-Jing Zhang. Dynamic cross feature fusion for remote sensing pan-sharpening. In Proceedings of the IEEE/CVF International + +Conference on Computer Vision, pages 14687-14696, 2021. 7 +[54] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 5 +[55] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13095-13105, 2023. 2, 4 +[56] Yinghui Xing, Litao Qu, Shizhou Zhang, Jiapeng Feng, Xiuwei Zhang, and Yanning Zhang. Empower generalizability for pansharpening through text-modulated diffusion model. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1, 2, 3, 6, 7 +[57] Kunran Xu, Lai Rui, Yishi Li, and Lin Gu. Feature normalized knowledge distillation for image classification. In European conference on computer vision, pages 664-680. Springer, 2020. 3 +[58] Junfeng Yang, Xueyang Fu, Yuwen Hu, Yue Huang, Xinghao Ding, and John Paisley. Pannet: A deep network architecture for pan-sharpening. In Proceedings of the IEEE international conference on computer vision, pages 5449-5457, 2017. 2, 7 +[59] Zhendong Yang, Zhe Li, Ailing Zeng, Zexian Li, Chun Yuan, and Yu Li. Vitkd: Feature-based knowledge distillation for vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1379-1388, 2024. 3 +[60] Lewei Yao, Renjie Pi, Hang Xu, Wei Zhang, Zhenguo Li, and Tong Zhang. G-detkd: Towards general distillation framework for object detectors via contrastive and semantic-guided feature imitation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3591-3600, 2021. 3 +[61] Qiangqiang Yuan, Yancong Wei, Xiangchao Meng, Huanfeng Shen, and Liangpei Zhang. A multiscale and multidepth convolutional neural network for remote sensing imagery pan-sharpening. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 11(3):978-989, 2018. 7 +[62] Roberta H Yuhas, Alexander FH Goetz, and Joe W Boardman. Discrimination among semi-arid landscape endmembers using the spectral angle mapper (sam) algorithm. In JPL, Summaries of the Third Annual JPL Airborne Geoscience Workshop. Volume 1: AVIRIS Workshop, 1992. 6 +[63] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 5 +[64] Kai Zhang, Anfei Wang, Feng Zhang, Wenbo Wan, Jiande Sun, and Lorenzo Bruzzone. Spatial-spectral dual backprojection network for pansharpening. IEEE Transactions on Geoscience and Remote Sensing, 2023. 2, 7 +[65] Yiman Zhang, Hanting Chen, Xinghao Chen, Yiping Deng, Chunjing Xu, and Yunhe Wang. Data-free knowledge dis + +tillation for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7852-7861, 2021. 3 +[66] Yun Zhang, Wei Li, Simiao Li, Hanting Chen, Zhijun Tu, Wenjia Wang, Bingyi Jing, Shaohui Lin, and Jie Hu. Data upcycling knowledge distillation for image super-resolution. arXiv preprint arXiv:2309.14162, 2023. 3 +[67] Yaqun Zhang, Zhichang Guo, Dazhi Zhang, and Boying Wu. An anisotropic variational pansharpening model with adaptive coefficients. Inverse Problems and Imaging, pages 0–0, 2024. 1 +[68] Yafei Zhang, Xuji Yang, Huafeng Li, Minghong Xie, and Zhengtao Yu. Dcpnet: A dual-task collaborative promotion network for pansharpening. IEEE Transactions on Geoscience and Remote Sensing, 62:1-16, 2024. 2, 7 +[69] Hao Zheng, Zhanlei Yang, Wenju Liu, Jizhong Liang, and Yanpeng Li. Improving deep neural networks using softplus units. In 2015 International joint conference on neural networks (IJCNN), pages 1-4. IEEE, 2015. 4 +[70] Wu Zheng, Weiliang Tang, Li Jiang, and Chi-Wing Fu. Sessd: Self-ensembling single-stage object detector from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14494–14503, 2021. 3 +[71] Jie Zhou, Daniel L Civco, and John A Silander. A wavelet transform method to merge landsat tm and spot panchromatic data. International journal of remote sensing, 19(4):743-757, 1998. 5 \ No newline at end of file diff --git a/2025/U-Know-DiffPAN_ An Uncertainty-aware Knowledge Distillation Diffusion Framework with Details Enhancement for PAN-Sharpening/images.zip b/2025/U-Know-DiffPAN_ An Uncertainty-aware Knowledge Distillation Diffusion Framework with Details Enhancement for PAN-Sharpening/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..99ff15455da7298f3c83a76259975099b3c68c2d --- /dev/null +++ b/2025/U-Know-DiffPAN_ An Uncertainty-aware Knowledge Distillation Diffusion Framework with Details Enhancement for PAN-Sharpening/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9368cd3f7a913957447b9dfee7fef5d63d483fcb6c1696e53629ca8c019e1883 +size 828996 diff --git a/2025/U-Know-DiffPAN_ An Uncertainty-aware Knowledge Distillation Diffusion Framework with Details Enhancement for PAN-Sharpening/layout.json b/2025/U-Know-DiffPAN_ An Uncertainty-aware Knowledge Distillation Diffusion Framework with Details Enhancement for PAN-Sharpening/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f07b30f1c94cea45915ec6fe05dc16278415f352 --- /dev/null +++ b/2025/U-Know-DiffPAN_ An Uncertainty-aware Knowledge Distillation Diffusion Framework with Details Enhancement for PAN-Sharpening/layout.json @@ -0,0 +1,12886 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 102, + 536, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 102, + 536, + 140 + ], + "spans": [ + { + "bbox": [ + 75, + 102, + 536, + 140 + ], + "type": "text", + "content": "U-Know-DiffPAN: An Uncertainty-aware Knowledge Distillation Diffusion Framework with Details Enhancement for PAN-Sharpening" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 84, + 161, + 153, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 161, + 153, + 175 + ], + "spans": [ + { + "bbox": [ + 84, + 161, + 153, + 175 + ], + "type": "text", + "content": "Sungpyo Kim" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 178, + 140, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 178, + 140, + 191 + ], + "spans": [ + { + "bbox": [ + 99, + 178, + 140, + 191 + ], + "type": "text", + "content": "1 KAIST" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 196, + 168, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 196, + 168, + 205 + ], + "spans": [ + { + "bbox": [ + 70, + 196, + 168, + 205 + ], + "type": "text", + "content": "ksp04204@kaist.ac.kr" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 211, + 161, + 288, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 161, + 288, + 175 + ], + "spans": [ + { + "bbox": [ + 211, + 161, + 288, + 175 + ], + "type": "text", + "content": "Jeonghyeok Do" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 229, + 178, + 270, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 178, + 270, + 191 + ], + "spans": [ + { + "bbox": [ + 229, + 178, + 270, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 229, + 178, + 270, + 191 + ], + "type": "text", + "content": "KAIST" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 191, + 196, + 308, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 191, + 196, + 308, + 205 + ], + "spans": [ + { + "bbox": [ + 191, + 196, + 308, + 205 + ], + "type": "text", + "content": "ehwjdgur0913@kaist.ac.kr" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 348, + 161, + 414, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 161, + 414, + 175 + ], + "spans": [ + { + "bbox": [ + 348, + 161, + 414, + 175 + ], + "type": "text", + "content": "Jaehyup Lee†" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 363, + 178, + 395, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 178, + 395, + 191 + ], + "spans": [ + { + "bbox": [ + 363, + 178, + 395, + 191 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 363, + 178, + 395, + 191 + ], + "type": "text", + "content": "KNU" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 332, + 196, + 429, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 196, + 429, + 205 + ], + "spans": [ + { + "bbox": [ + 332, + 196, + 429, + 205 + ], + "type": "text", + "content": "jaehyuplee@knu.ac.kr" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 459, + 161, + 536, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 459, + 161, + 536, + 175 + ], + "spans": [ + { + "bbox": [ + 459, + 161, + 536, + 175 + ], + "type": "text", + "content": "Munchurl Kim" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 475, + 178, + 517, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 475, + 178, + 517, + 191 + ], + "spans": [ + { + "bbox": [ + 475, + 178, + 517, + 191 + ], + "type": "text", + "content": "1KAIST" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 452, + 196, + 539, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 452, + 196, + 539, + 205 + ], + "spans": [ + { + "bbox": [ + 452, + 196, + 539, + 205 + ], + "type": "text", + "content": "mkimee@kaist.ac.kr" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 58, + 234, + 113, + 344 + ], + "blocks": [ + { + "bbox": [ + 58, + 234, + 113, + 344 + ], + "lines": [ + { + "bbox": [ + 58, + 234, + 113, + 344 + ], + "spans": [ + { + "bbox": [ + 58, + 234, + 113, + 344 + ], + "type": "image", + "image_path": "ab6af6b1aaff5f0f456ea32e538cf3b54298a70bd8c9b8f9355d561f7b00c523.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 348, + 555, + 403 + ], + "lines": [ + { + "bbox": [ + 55, + 348, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 555, + 403 + ], + "type": "text", + "content": "Figure 1. Visual comparison of PAN-sharpening results on the full-resolution WV3 dataset. The rightmost image shows the output of our proposed U-Know-DiffPAN framework, specifically the result produced by FSA-S (frequency selective attention student network). Notably, the proposed framework generates more detailed and robust results, particularly in high-uncertainty regions, outperforming the state-of-the-art model CANConv [14] and recent diffusion-based methods PanDiff [31] and TMDiff [56]. As highlighted in the red box, our approach successfully restores challenging highly-uncertain regions, such as cars, where other models fall short." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 113, + 234, + 222, + 343 + ], + "blocks": [ + { + "bbox": [ + 131, + 223, + 205, + 233 + ], + "lines": [ + { + "bbox": [ + 131, + 223, + 205, + 233 + ], + "spans": [ + { + "bbox": [ + 131, + 223, + 205, + 233 + ], + "type": "text", + "content": "PanDiff (TGRS 2023)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 113, + 234, + 222, + 343 + ], + "lines": [ + { + "bbox": [ + 113, + 234, + 222, + 343 + ], + "spans": [ + { + "bbox": [ + 113, + 234, + 222, + 343 + ], + "type": "image", + "image_path": "34b1b64b9497bb181f9865af225cf03a9e90f436c317c1aff15eb4414724022a.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 223, + 234, + 332, + 343 + ], + "blocks": [ + { + "bbox": [ + 241, + 223, + 315, + 233 + ], + "lines": [ + { + "bbox": [ + 241, + 223, + 315, + 233 + ], + "spans": [ + { + "bbox": [ + 241, + 223, + 315, + 233 + ], + "type": "text", + "content": "TMDiff (TGRS 2024)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 223, + 234, + 332, + 343 + ], + "lines": [ + { + "bbox": [ + 223, + 234, + 332, + 343 + ], + "spans": [ + { + "bbox": [ + 223, + 234, + 332, + 343 + ], + "type": "image", + "image_path": "f53a2e345610728e2c67de6e7daf4d679ee27cbf2338295882bc27ab88d679f7.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 332, + 234, + 442, + 343 + ], + "blocks": [ + { + "bbox": [ + 346, + 223, + 430, + 233 + ], + "lines": [ + { + "bbox": [ + 346, + 223, + 430, + 233 + ], + "spans": [ + { + "bbox": [ + 346, + 223, + 430, + 233 + ], + "type": "text", + "content": "CANConv (CVPR 2024)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 332, + 234, + 442, + 343 + ], + "lines": [ + { + "bbox": [ + 332, + 234, + 442, + 343 + ], + "spans": [ + { + "bbox": [ + 332, + 234, + 442, + 343 + ], + "type": "image", + "image_path": "1a235cba609c2ce645eeafd4d9897b9d6e1f8f2644e71fce7dcdabad8081ee1d.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 443, + 234, + 553, + 343 + ], + "blocks": [ + { + "bbox": [ + 453, + 223, + 543, + 233 + ], + "lines": [ + { + "bbox": [ + 453, + 223, + 543, + 233 + ], + "spans": [ + { + "bbox": [ + 453, + 223, + 543, + 233 + ], + "type": "text", + "content": "U-Know-DiffPAN (Ours)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 443, + 234, + 553, + 343 + ], + "lines": [ + { + "bbox": [ + 443, + 234, + 553, + 343 + ], + "spans": [ + { + "bbox": [ + 443, + 234, + 553, + 343 + ], + "type": "image", + "image_path": "bca9a5628c4c47d655d5d2cacca7df7e1bf0a4a5036c8fcd975ce184b0c92d02.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 151, + 419, + 200, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 419, + 200, + 432 + ], + "spans": [ + { + "bbox": [ + 151, + 419, + 200, + 432 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 54, + 445, + 296, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 445, + 296, + 673 + ], + "spans": [ + { + "bbox": [ + 54, + 445, + 296, + 673 + ], + "type": "text", + "content": "Conventional methods for PAN-sharpening often struggle to restore fine details due to limitations in leveraging high-frequency information. Moreover, diffusion-based approaches lack sufficient conditioning to fully utilize Panchromatic (PAN) images and low-resolution multispectral (LRMS) inputs effectively. To address these challenges, we propose an uncertainty-aware knowledge distillation diffusion framework with details enhancement for PAN-sharpening, called U-Know-DiffPAN. The U-Know-DiffPAN incorporates uncertainty-aware knowledge distillation for effective transfer of feature details from our teacher model to a student one. The teacher model in our U-Know-DiffPAN captures frequency details through frequency selective attention, facilitating accurate reverse process learning. By conditioning the encoder on compact vector representations of PAN and LRMS and the decoder on Wavelet transforms, we enable rich frequency utilization. So, the high-capacity teacher model distills frequency-rich features into a lightweight student model aided by an un" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 421, + 555, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 555, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 555, + 517 + ], + "type": "text", + "content": "certainty map. From this, the teacher model can guide the student model to focus on difficult image regions for PAN-sharpening via the usage of the uncertainty map. Extensive experiments on diverse datasets demonstrate the robustness and superior performance of our U-Know-DiffPAN over very recent state-of-the-art PAN-sharpening methods. The project page is available at https://kaist-viclab.github.io/U-Know-DiffPAN-site/." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 314, + 539, + 395, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 539, + 395, + 551 + ], + "spans": [ + { + "bbox": [ + 314, + 539, + 395, + 551 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 558, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 558, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 558, + 555, + 715 + ], + "type": "text", + "content": "Satellite imagery plays a vital role in a wide range of applications, including environmental monitoring, surveillance, and mapping [2, 10, 15, 29, 67]. High-resolution multispectral (HRMS) images that combine high spatial and spectral fidelity are commonly generated using PAN-sharpening techniques, fusing low-resolution multispectral (LRMS) images with high-resolution (HR) panchromatic (PAN) images. The primary objective of PAN-sharpening is to retain the spatial detail of PAN images while preserving the spectral information of LRMS images. Recently, deep learning (DL)-based approaches have gained attraction in PAN-sharpening tasks, leveraging deep networks to perform end-to-end image regression. However, DL models" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 683, + 276, + 694 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 683, + 276, + 694 + ], + "spans": [ + { + "bbox": [ + 66, + 683, + 276, + 694 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 683, + 276, + 694 + ], + "type": "text", + "content": " Korea Advanced Institute of Science and Technology (KAIST)." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 67, + 694, + 201, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 694, + 201, + 704 + ], + "spans": [ + { + "bbox": [ + 67, + 694, + 201, + 704 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 694, + 201, + 704 + ], + "type": "text", + "content": "Kyungpook National University (KNU)." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 67, + 704, + 223, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 704, + 223, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 704, + 223, + 714 + ], + "type": "text", + "content": "† Co-corresponding authors (equal contribution)." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23069" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "are typically limited by the local receptive fields of convolutional layers, which primarily focus on local patterns and fail to capture long-range dependencies. This often leads to overly smooth outputs with reduced high-frequency details." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 122, + 295, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 295, + 337 + ], + "type": "text", + "content": "To overcome this limitation, generative models such as generative adversarial networks (GANs) [18] have been applied to PAN-sharpening, aiming to produce more realistic results. Despite their success, GANs [28, 30] pose additional challenges, including training instability, mode collapse, and artifact generation, which can degrade image quality. Recently, diffusion models [21, 38, 42, 55] have demonstrated strong performance across a variety of image generation and restoration tasks. In PAN-sharpening, diffusion-based methods [31, 56] have shown promise by using PAN and LRMS images as conditioning inputs. However, they are limited by simple conditioning strategies that do not fully utilize the complementary information provided by the PAN and LRMS inputs, thus leading to suboptimal restoration quality. Additionally, they often incur high computational costs and longer inference times compared to other DL-based methods, limiting their practical applicability." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 340, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 295, + 555 + ], + "type": "text", + "content": "To address these challenges, we propose U-Know-DiffPAN demonstrated in Fig. 2, a novel PAN-sharpening framework that incorporates an uncertainty-aware knowledge distillation (U-Know) strategy. Our framework is designed to fully leverage the frequency information in PAN and LRMS inputs while reducing computational costs. Our teacher model, is built to maximize information extraction from PAN and LRMS images by utilizing (i) a Feed Forward Attention (FFA) block in the encoder to capture compact vector representations of PAN and LRMS, and (ii) a High Quality Frequency Enhancement (HQFE) block in the decoder, composed of Fourier Transform Channel Attention (FTCA) and Stationary Wavelet Transform Cross Attention (SWTCA) modules. The FTCA enhances crucial frequency components through Fourier transform-based [52] attention, while the SwTCA incorporates high-frequency information from PAN and low-frequency information from LRMS into the decoder." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 558, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 712 + ], + "type": "text", + "content": "Our U-Know strategy leverages uncertainty-aware knowledge distillation by transferring the features from the teacher model to a light-weight student model. The teacher model denoises HRMS images along with an uncertainty map, identifying spatial regions where the prediction confidence is low. Then based of the U-Know strategy, the teacher model distills to the student model not only the frequency-aware features but also the uncertainty map to address spatial weaknesses in the teacher's predictions. This approach leads to improved performance and reduced computational cost, making our U-Know-DiffPAN more effective and efficient for PAN-sharpening. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 72, + 553, + 216 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 314, + 72, + 553, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 553, + 119 + ], + "type": "text", + "content": "- We firstly incorporate the uncertainty-aware knowledge distillation into diffusion based PAN-sharpening with the guidance of pixel-wise restoration uncertainty and frequency-selective attention for details enhancement;" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 121, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 121, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 314, + 121, + 553, + 167 + ], + "type": "text", + "content": "- Our teacher model leverages compact vector representations to enhance encoding efficiency and employs frequency selective attention in the decoder to refine frequency components for PAN-sharpening." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 168, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 168, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 314, + 168, + 553, + 216 + ], + "type": "text", + "content": "- We conduct extensive experiments on multiple datasets [11], including WV3, QB, and GF2, demonstrating state-of-the-art performance and surpassing very recent methods in PAN-sharpening with large margins." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 230, + 405, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 230, + 405, + 243 + ], + "spans": [ + { + "bbox": [ + 314, + 230, + 405, + 243 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 251, + 416, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 251, + 416, + 264 + ], + "spans": [ + { + "bbox": [ + 314, + 251, + 416, + 264 + ], + "type": "text", + "content": "2.1. PAN-Sharpening" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 270, + 553, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 553, + 567 + ], + "type": "text", + "content": "PAN-sharpening is a technique to fuse high-resolution (HR) panchromatic (PAN) images with low-resolution multispectral (LRMS) images to produce HR multispectral (HRMS) images. Traditional PAN-sharpening techniques are broadly categorized into three approaches: component substitution (CS) [8, 41], multi-resolution analysis (MRA) [1, 34], and variational optimization (VO) [3]. Modern PAN-sharpening methods primarily consist of CNN-based [14, 25, 47, 58], transformer-based [25, 68], and GAN-based [28, 30] models. LAGConv [25] introduces local-context adaptive convolutional kernels with a global harmonic bias mechanism. S2DBPN [64] uses a spatial-spectral dual back-projection network for effective PAN and LRMS fusion. DCPNet [68] establishes a transformer-based dual-task parallel framework to optimize reconstruction. CANConv [14] integrates a non-local adaptive convolution module, enhancing spatial adaptability and non-local self-similarity. Besides the above methods, there are other DL-based methods for PAN-sharpening [9, 13, 17, 22, 23, 27]. While deep learning (DL)-based methods can capture complex mappings, they encounter challenges. CNN-based approaches tend to produce overly smooth results due to local receptive field of convolution layers. GAN-based models, though capable of generating realistic textures, may suffer from instability and introduce unrealistic artifacts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "content": "Diffusion-based approaches. Diffusion models (DMs) [21, 42] offer a promising alternative to conventional DL-based methods for image restoration [48, 55], iteratively refining noisy images through progressive denoising. Unlike GANs [18, 26], which can be prone to instability, DMs use iterative denoising steps with flexible conditioning, making them robust for tasks such as image restoration and PAN-sharpening. Recent diffusion-based PAN-sharpening models [31, 56] have shown significant advancements. PanDiff [31], however, employs a relatively simplistic conditioning approach that limits its ability to fully utilize the information from the input data, sometimes resulting in subopti" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23070" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": "mal restoration quality. In contrast, TMDiff [56] employs a complex conditioning mechanism, requiring an additional pretrained CLIP [36] text encoder, which adds unnecessary complexity without substantial performance gains. Our proposed framework, U-Know-DiffPAN, handles these limitations by integrating an advanced conditioning strategy that maximizes the use of input information. Specifically, our U-Know-DiffPAN conditions the encoder on compact vector representations of PAN and LRMS, while the decoder is conditioned on the stationary wavelet transform [32] of PAN and LRMS. This approach allows our U-Know-DiffPAN to effectively capture both frequency details and spatial fidelity, achieving superior restoration quality compared to existing diffusion-based methods." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 247, + 185, + 259 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 185, + 259 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 185, + 259 + ], + "type": "text", + "content": "2.2. Knowledge Distillation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 264, + 296, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 296, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 296, + 492 + ], + "type": "text", + "content": "Knowledge Distillation (KD) [20] is a well-known method for transferring knowledge from a large, model (teacher) to a smaller model (student), enabling deployment in resource-constrained settings while retaining performance. KD has demonstrated effectiveness across a range of computer vision tasks, including classification [7, 57, 65], object detection [24, 49, 60, 70], and image restoration [43, 59, 66]. However, previous KD methods, often fall short in tasks such as super-resolution [66], where capturing complicated detail is essential. This limitation arises because conventional KD student learns all regions equally, without focusing on areas that require concentrated learning. To address this, we propose uncertainty-aware knowledge distillation (U-Know), which introduces an uncertainty map into the distillation process. Our U-Know strategy informs the student of uncertain regions, areas the teacher model finds challenging to restore, enabling the student model to refine crucial details, thereby enhancing fidelity in both spatial and spectral domains." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 502, + 117, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 117, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 117, + 514 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 521, + 221, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 221, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 221, + 533 + ], + "type": "text", + "content": "3.1. Overview of U-Know-DiffPAN" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "content": "Fig. 2 shows the overview of our U-Know-DiffPAN framework. The proposed U-Know-DiffPAN framework comprises a teacher network with frequency-selective attention (FSA-T) and a student network (FSA-S), designed within a knowledge distillation paradigm. PAN image " + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{PAN}}" + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "content": " and interpolated LRMS image " + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{LR}}" + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "content": " are inputted in concatenation to generate HRMS image " + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{HR}}" + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "content": ". Both FSA-T and FSA-S are designed to predict the residuals as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 642, + 295, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 642, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 138, + 642, + 295, + 658 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {0} = \\mathbf {I} _ {\\mathrm {M S}} ^ {\\mathrm {H R}} - \\mathbf {I} _ {\\mathrm {M S}} ^ {\\mathrm {L R}}. \\tag {1}", + "image_path": "8fc6bb8efaf6a97fe8e5b737385501fbc5e1b228019cf3247b70d2d5e0c4850d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "content": "Our framework operates in two main stages: (i) pre-training the FSA-T within a diffusion process to produce an initial prediction " + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{X}}_0" + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "content": " alongside an uncertainty map " + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}" + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "content": " that identifies spatially weak regions; and (ii) training the FSA-S by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 318, + 73, + 408, + 81 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 73, + 408, + 81 + ], + "spans": [ + { + "bbox": [ + 318, + 73, + 408, + 81 + ], + "type": "text", + "content": "Step1. Teacher (FSA-T) pre-training" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 319, + 83, + 386, + 135 + ], + "blocks": [ + { + "bbox": [ + 319, + 83, + 386, + 135 + ], + "lines": [ + { + "bbox": [ + 319, + 83, + 386, + 135 + ], + "spans": [ + { + "bbox": [ + 319, + 83, + 386, + 135 + ], + "type": "image", + "image_path": "adf6fec8ea1d84b4cf8fefc26a393936306bd5d67eed577bfb76b3431c3d3979.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 392, + 83, + 535, + 136 + ], + "blocks": [ + { + "bbox": [ + 392, + 83, + 535, + 136 + ], + "lines": [ + { + "bbox": [ + 392, + 83, + 535, + 136 + ], + "spans": [ + { + "bbox": [ + 392, + 83, + 535, + 136 + ], + "type": "image", + "image_path": "12ab21e8c633dacbb7aa7e1862747fdac4bf95ef862d5c6b2ce0a8b1afa16e0f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 318, + 137, + 397, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 137, + 397, + 146 + ], + "spans": [ + { + "bbox": [ + 318, + 137, + 397, + 146 + ], + "type": "text", + "content": "Step2. Student (FSA-S) training" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 318, + 148, + 387, + 195 + ], + "blocks": [ + { + "bbox": [ + 318, + 148, + 387, + 195 + ], + "lines": [ + { + "bbox": [ + 318, + 148, + 387, + 195 + ], + "spans": [ + { + "bbox": [ + 318, + 148, + 387, + 195 + ], + "type": "image", + "image_path": "1d50faf5ecd3ae8177e643efa50e922af91bfa781f53f1c509cbeeb87276ac20.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 204, + 554, + 237 + ], + "lines": [ + { + "bbox": [ + 313, + 204, + 554, + 237 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 554, + 237 + ], + "type": "text", + "content": "Figure 2. Overview of our uncertainty-aware knowledge-distillation diffusion framework with details enhancement, called U-Know-DiffPAN." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 392, + 148, + 552, + 195 + ], + "blocks": [ + { + "bbox": [ + 392, + 148, + 552, + 195 + ], + "lines": [ + { + "bbox": [ + 392, + 148, + 552, + 195 + ], + "spans": [ + { + "bbox": [ + 392, + 148, + 552, + 195 + ], + "type": "image", + "image_path": "59e338e4cde93290f2a51cd51b2a9ac6fc542f39081f58d6042213d0473a6363.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 246, + 554, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 554, + 354 + ], + "type": "text", + "content": "leveraging this uncertainty map to guide the FSA-S in refining these regions through the KD. In Sec. 3.2, we describe the FSA-T's diffusion-based training process, which incorporates frequency information from PAN and LRMS inputs. Sec. 3.3 details the FSA-T architecture, specifically designed to enhance frequency components. Finally, in Sec. 3.4, we explain the FSA-S's uncertainty-aware knowledge distillation framework, which utilizes the uncertainty map to achieve precise and reliable PAN-sharpening." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 360, + 554, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 554, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 554, + 372 + ], + "type": "text", + "content": "3.2. Diffusion Process of Teacher Network (FSA-T)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 377, + 554, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 554, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 554, + 474 + ], + "type": "text", + "content": "Diffusion models operate by defining a forward process that gradually adds noise to data until it becomes indistinguishable from random noise, and a reverse process that learns to denoise the data, reconstructing the original information. We leverage this framework into a frequency-selective attention teacher network (FSA-T) as our teacher network, to generate frequency enhanced HRMS images with their pixel-wise uncertainty information." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 474, + 554, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 534 + ], + "type": "text", + "content": "Forward process. In the forward diffusion process, noise is progressively added to the residual image, " + }, + { + "bbox": [ + 313, + 474, + 554, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_0" + }, + { + "bbox": [ + 313, + 474, + 554, + 534 + ], + "type": "text", + "content": ", until it converges to a Gaussian distribution, " + }, + { + "bbox": [ + 313, + 474, + 554, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 313, + 474, + 554, + 534 + ], + "type": "text", + "content": ", after " + }, + { + "bbox": [ + 313, + 474, + 554, + 534 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 474, + 554, + 534 + ], + "type": "text", + "content": " timesteps. At each timestep " + }, + { + "bbox": [ + 313, + 474, + 554, + 534 + ], + "type": "inline_equation", + "content": "t \\in \\{1, \\dots, T\\}" + }, + { + "bbox": [ + 313, + 474, + 554, + 534 + ], + "type": "text", + "content": ", a noisy version of the residual is obtained as follows [21]:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 375, + 540, + 553, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 540, + 553, + 554 + ], + "spans": [ + { + "bbox": [ + 375, + 540, + 553, + 554 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {X} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\tag {2}", + "image_path": "b624402d0a2f14e2124a02aa903725d7a6352e8642b63ef2f838aab2da14a7c8.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 560, + 553, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 553, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 553, + 585 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 560, + 553, + 585 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 313, + 560, + 553, + 585 + ], + "type": "text", + "content": " is Gaussian noise, and " + }, + { + "bbox": [ + 313, + 560, + 553, + 585 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t = \\prod_{s=1}^t (1 - \\beta_s)" + }, + { + "bbox": [ + 313, + 560, + 553, + 585 + ], + "type": "text", + "content": " controls the noise level at step " + }, + { + "bbox": [ + 313, + 560, + 553, + 585 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 560, + 553, + 585 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 586, + 554, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 554, + 692 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 554, + 692 + ], + "type": "text", + "content": "Reverse process. Following the DDPM [21], the reverse process aims to recover " + }, + { + "bbox": [ + 313, + 586, + 554, + 692 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_0" + }, + { + "bbox": [ + 313, + 586, + 554, + 692 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 586, + 554, + 692 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_T" + }, + { + "bbox": [ + 313, + 586, + 554, + 692 + ], + "type": "text", + "content": " by predicting a denoised image at each step. To perform denoising at each timestep, the denoising network (FSA-T), denoted as " + }, + { + "bbox": [ + 313, + 586, + 554, + 692 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 313, + 586, + 554, + 692 + ], + "type": "text", + "content": ", is designed to predict the original " + }, + { + "bbox": [ + 313, + 586, + 554, + 692 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_0" + }, + { + "bbox": [ + 313, + 586, + 554, + 692 + ], + "type": "text", + "content": " rather than the added noise " + }, + { + "bbox": [ + 313, + 586, + 554, + 692 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 586, + 554, + 692 + ], + "type": "text", + "content": ", as we found this empirically to be beneficial for PAN-sharpening. Additionally, " + }, + { + "bbox": [ + 313, + 586, + 554, + 692 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 313, + 586, + 554, + 692 + ], + "type": "text", + "content": " simultaneously produces an uncertainty map " + }, + { + "bbox": [ + 313, + 586, + 554, + 692 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}" + }, + { + "bbox": [ + 313, + 586, + 554, + 692 + ], + "type": "text", + "content": " for each prediction, providing pixelwise confidence information as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 324, + 696, + 553, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 696, + 553, + 717 + ], + "spans": [ + { + "bbox": [ + 324, + 696, + 553, + 717 + ], + "type": "interline_equation", + "content": "\\left[ \\widehat {\\mathbf {X}} _ {0} \\mid \\widehat {\\boldsymbol {\\theta}} \\right] = \\Psi \\left(\\left[ \\mathbf {X} _ {t} \\mid \\mathbf {I} _ {\\text {P A N}} \\mid \\mathbf {I} _ {\\mathrm {M S}} ^ {\\mathrm {L R}} \\right]; \\mathbf {v}, \\mathcal {S} - \\text {C o n d}; t\\right), \\tag {3}", + "image_path": "1bf4b8814d112ab28b5f5ee45525ec279b625cd733e95c8cf71df4619549ccac.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23071" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 157, + 212 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 157, + 212 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 157, + 212 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 157, + 212 + ], + "type": "image", + "image_path": "a2a0e0bcf4075dbb7ba1c9fcdfa5dc1651fe7f943632525bd546219e47e12585.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 159, + 71, + 465, + 213 + ], + "blocks": [ + { + "bbox": [ + 159, + 71, + 465, + 213 + ], + "lines": [ + { + "bbox": [ + 159, + 71, + 465, + 213 + ], + "spans": [ + { + "bbox": [ + 159, + 71, + 465, + 213 + ], + "type": "image", + "image_path": "65a9812a2f6b025a690644bfb9952031ec187f6f7a90b6433f6aa4d6542c8409.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 466, + 74, + 552, + 211 + ], + "blocks": [ + { + "bbox": [ + 466, + 74, + 552, + 211 + ], + "lines": [ + { + "bbox": [ + 466, + 74, + 552, + 211 + ], + "spans": [ + { + "bbox": [ + 466, + 74, + 552, + 211 + ], + "type": "image", + "image_path": "5143c06fceb3b4937c9e978678ee64ec052fc45b53632a817a8f7e0c79468fcc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 214, + 157, + 274 + ], + "blocks": [ + { + "bbox": [ + 58, + 214, + 157, + 274 + ], + "lines": [ + { + "bbox": [ + 58, + 214, + 157, + 274 + ], + "spans": [ + { + "bbox": [ + 58, + 214, + 157, + 274 + ], + "type": "image", + "image_path": "751d94a4da0449bbcb18709c4d3f90f85e7447266244962bf47db8460c990d8d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 281, + 553, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 281, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 553, + 304 + ], + "type": "text", + "content": "Figure 3. Architecture of our proposed teacher model with frequency-selective attention, denoted as FSA-T, for PAN-sharpening. The FSA-T is designed to fully utilize frequency information for details enhancement from PAN and LRMS inputs (more details in Supplemental.)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 159, + 217, + 294, + 273 + ], + "blocks": [ + { + "bbox": [ + 159, + 217, + 294, + 273 + ], + "lines": [ + { + "bbox": [ + 159, + 217, + 294, + 273 + ], + "spans": [ + { + "bbox": [ + 159, + 217, + 294, + 273 + ], + "type": "image", + "image_path": "efa7c75e788eea5ee898f43fe4618f5b69483a314bdabbcaaeb018db2c433391.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 299, + 215, + 416, + 274 + ], + "blocks": [ + { + "bbox": [ + 299, + 215, + 416, + 274 + ], + "lines": [ + { + "bbox": [ + 299, + 215, + 416, + 274 + ], + "spans": [ + { + "bbox": [ + 299, + 215, + 416, + 274 + ], + "type": "image", + "image_path": "7dadd3b623aa519e83469df927456b6ccf6079903df4a319907b20ccd46a8687.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 423, + 215, + 552, + 273 + ], + "blocks": [ + { + "bbox": [ + 423, + 215, + 552, + 273 + ], + "lines": [ + { + "bbox": [ + 423, + 215, + 552, + 273 + ], + "spans": [ + { + "bbox": [ + 423, + 215, + 552, + 273 + ], + "type": "image", + "image_path": "f565785ec6e442234fbbb45b8791270223e45b9d510b80aa370ad2837cc31b2d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 312, + 295, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 312, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 54, + 312, + 295, + 445 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 312, + 295, + 445 + ], + "type": "inline_equation", + "content": "\\left[\\cdot \\mid \\cdot \\right]" + }, + { + "bbox": [ + 54, + 312, + 295, + 445 + ], + "type": "text", + "content": " indicates channel-wise concatenation, " + }, + { + "bbox": [ + 54, + 312, + 295, + 445 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{X}}_0" + }, + { + "bbox": [ + 54, + 312, + 295, + 445 + ], + "type": "text", + "content": " is a predicted residual, " + }, + { + "bbox": [ + 54, + 312, + 295, + 445 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}" + }, + { + "bbox": [ + 54, + 312, + 295, + 445 + ], + "type": "text", + "content": " is a predicted uncertainty map, " + }, + { + "bbox": [ + 54, + 312, + 295, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 54, + 312, + 295, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 312, + 295, + 445 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 54, + 312, + 295, + 445 + ], + "type": "text", + "content": "-Cond represent a learning compact representation (vector) and Stationary Wavelet Transform (SWT)-based conditioning terms, respectively. This formulation allows the FSA-T to iteratively refine the noisy input, guided by the conditioned information (" + }, + { + "bbox": [ + 54, + 312, + 295, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 54, + 312, + 295, + 445 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 312, + 295, + 445 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 54, + 312, + 295, + 445 + ], + "type": "text", + "content": "-Cond), toward an accurate reconstruction of the " + }, + { + "bbox": [ + 54, + 312, + 295, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_0" + }, + { + "bbox": [ + 54, + 312, + 295, + 445 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 54, + 312, + 295, + 445 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}" + }, + { + "bbox": [ + 54, + 312, + 295, + 445 + ], + "type": "text", + "content": " captures pixel-wise prediction confidence, which guides the FSA-S to more rigorously learn the regions with high uncertainty. To ensure positivity, we apply the SoftPlus [69] activation to " + }, + { + "bbox": [ + 54, + 312, + 295, + 445 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}" + }, + { + "bbox": [ + 54, + 312, + 295, + 445 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 445, + 296, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 296, + 482 + ], + "type": "text", + "content": "Uncertainty-driven diffusion loss. Inspired by [33, 39], we employ the following uncertainty-driven diffusion loss function, " + }, + { + "bbox": [ + 55, + 445, + 296, + 482 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{U - Diff}}" + }, + { + "bbox": [ + 55, + 445, + 296, + 482 + ], + "type": "text", + "content": ", to optimize the teacher network " + }, + { + "bbox": [ + 55, + 445, + 296, + 482 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 55, + 445, + 296, + 482 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 87, + 487, + 295, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 487, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 87, + 487, + 295, + 515 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {U - D i f f}} = \\left\\| \\frac {1}{2 \\hat {\\boldsymbol {\\theta}}} \\odot \\left| \\widehat {\\mathbf {X}} _ {0} - \\mathbf {X} _ {0} \\right| + \\frac {1}{2} \\log \\hat {\\boldsymbol {\\theta}} \\right\\| _ {1}, \\tag {4}", + "image_path": "d55f8502768a5ac5c8a77707aa88d6cc327721cc6d34f79ce1d8a77281bff0bb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 521, + 296, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 583 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 521, + 296, + 583 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}" + }, + { + "bbox": [ + 55, + 521, + 296, + 583 + ], + "type": "text", + "content": " serves as the estimated variance term and is regarded as the uncertainty map in our framework. This objective function encourages the model to minimize prediction errors in low-confidence regions, thus yielding a more accurate and spatially robust output." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 588, + 268, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 268, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 268, + 601 + ], + "type": "text", + "content": "3.3. Details of Teacher Network Architecture" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "Our teacher network, FSA-T, is designed to enhance frequency details for PAN-sharpening tasks by fully utilizing complementary information from PAN and LRMS inputs. The FSA-T adopts an encoder-decoder (U-Net [37]) structure, where each stage is specifically tailored to capture and refine critical spatial and frequency features. The encoder of FSA-T employs Feed Forward Attention (FFA) blocks to extract compact and informative representations from the PAN and LRMS inputs. This attention mechanism allows" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 312, + 553, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 336 + ], + "type": "text", + "content": "the encoder to focus on essential features, preparing the data for effective frequency enhancement in subsequent layers." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 337, + 554, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 554, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 554, + 493 + ], + "type": "text", + "content": "The decoder of FSA-T integrates High-Quality Frequency Enhancement (HQFE) blocks at each decoding level to further refine frequency components. Each HQFE block comprises two key modules: Fourier Transform Channel Attention (FTCA) that emphasizes important frequency components via Fourier Transform based attention, and Stationary Wavelet Transform Cross Attention (SWTCA) that merges high-frequency information from PAN and low-frequency content from LRMS, both of which are decomposed by SwT [32]. By incorporating these specialize components, FSA-T effectively produces " + }, + { + "bbox": [ + 313, + 337, + 554, + 493 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{X}}_0" + }, + { + "bbox": [ + 313, + 337, + 554, + 493 + ], + "type": "text", + "content": " with enhanced spatial fidelity, establishing an effective uncertainty-aware knowledge distillation (U-Know) strategy." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 494, + 554, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 554, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 554, + 554 + ], + "type": "text", + "content": "Feed Forward Attention (FFA). The FFA block in Fig. 3 is designed to leverage compact information [55] from the PAN and LRMS input to establish a rough structural representation during encoding. A compact vector representation " + }, + { + "bbox": [ + 313, + 494, + 554, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{v} \\in \\mathbb{R}^{1 \\times D}" + }, + { + "bbox": [ + 313, + 494, + 554, + 554 + ], + "type": "text", + "content": " is obtained by a vector extractor " + }, + { + "bbox": [ + 313, + 494, + 554, + 554 + ], + "type": "inline_equation", + "content": "(\\mathcal{E})" + }, + { + "bbox": [ + 313, + 494, + 554, + 554 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 392, + 563, + 553, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 563, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 392, + 563, + 553, + 578 + ], + "type": "interline_equation", + "content": "\\mathbf {v} = \\mathcal {E} \\left(\\mathbf {I} _ {\\mathrm {P A N}}, \\mathbf {I} _ {\\mathrm {M S}} ^ {\\mathrm {L R}}\\right). \\tag {5}", + "image_path": "001364dc60ad8e53723b1f0dbfd6ee77de0a0bd3c8323461d5cf1a4ac8c6d518.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 588, + 553, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 553, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 313, + 588, + 553, + 624 + ], + "type": "text", + "content": " is then used as a dynamic modulation [26, 50] parameter within the FFA, enabling " + }, + { + "bbox": [ + 313, + 588, + 553, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 588, + 553, + 624 + ], + "type": "text", + "content": " to incorporate " + }, + { + "bbox": [ + 313, + 588, + 553, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 313, + 588, + 553, + 624 + ], + "type": "text", + "content": " and establish an approximate structure during restoration as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 347, + 634, + 443, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 634, + 443, + 647 + ], + "spans": [ + { + "bbox": [ + 347, + 634, + 443, + 647 + ], + "type": "interline_equation", + "content": "[ \\boldsymbol {\\beta} _ {n} \\mid \\gamma_ {n} ] = \\operatorname {L i n e a r} (\\mathbf {v}),", + "image_path": "922020674285fb4ce06caad69921296b3035434b8de3421311ca57f9c003f2da.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 347, + 649, + 553, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 649, + 553, + 662 + ], + "spans": [ + { + "bbox": [ + 347, + 649, + 553, + 662 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {n} ^ {\\prime} = \\gamma_ {n} \\odot \\mathrm {G N} \\left(\\mathbf {f} _ {n}\\right) + \\beta_ {n}, \\tag {6}", + "image_path": "8a23fa8cff1564dea3739d43008dc6d657758ef9f0d0aa40c37b0db4f4643394.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 347, + 664, + 520, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 664, + 520, + 677 + ], + "spans": [ + { + "bbox": [ + 347, + 664, + 520, + 677 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {n} \\leftarrow \\operatorname {C o n v} \\left(\\mathbf {f} _ {n} ^ {\\prime}\\right) \\odot \\operatorname {G E L U} \\left(\\operatorname {C o n v} \\left(\\mathbf {f} _ {n} ^ {\\prime}\\right)\\right) + \\mathbf {f} _ {n},", + "image_path": "0be3fc8d76ba68d336978fac40bafa348ae16e67da2fa1a93aff97119a0e5804.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_n\\in \\mathbb{R}^{C\\times h\\times w}" + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": " denotes input feature for the " + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": " th encoder, Linear indicates Linear operation, " + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": " denotes" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23072" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "content": "element-wise multiplication, GN indicates group normalization [54], GELU indicates GELU [19] activation and Conv consists of a sequential " + }, + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "content": " convolution and a " + }, + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "content": " depth-wise convolution layer (Details of " + }, + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "content": " is in the Supplemental.)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "type": "text", + "content": "Fourier Transform Channel Attention (FTCA). The FTCA block in Fig. 3 performs self-attention in frequency domain by leveraging the 2D Fast Fourier Transform (FFT), denoted as " + }, + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "type": "text", + "content": " allows for efficient identification and emphasis of critical frequency components. For the " + }, + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "type": "text", + "content": "-th decoder block in Fig. 3, the input feature " + }, + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_n \\in \\mathbb{R}^{C \\times h \\times w}" + }, + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "type": "text", + "content": " channels is transformed into query " + }, + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "type": "inline_equation", + "content": "(\\mathbf{Q})" + }, + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "type": "text", + "content": ", key " + }, + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "type": "inline_equation", + "content": "(\\mathbf{K})" + }, + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "type": "text", + "content": ", and value " + }, + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "type": "inline_equation", + "content": "(\\mathbf{V})" + }, + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "type": "text", + "content": " representations. The 2D FFT " + }, + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "type": "text", + "content": " is then applied independently to each channel of " + }, + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 55, + 133, + 295, + 252 + ], + "type": "text", + "content": ", mapping them transform from time domain to frequency domain as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 112, + 262, + 295, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 262, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 112, + 262, + 295, + 277 + ], + "type": "interline_equation", + "content": "[ \\mathbf {Q} \\mid \\mathbf {K} \\mid \\mathbf {V} ] = \\mathcal {F} (\\operatorname {C o n v} \\left(\\mathbf {f} _ {n}\\right)), \\tag {7}", + "image_path": "e412a427a945d8e4f373298132c1188ecb05201cb7d4164fb44259234ef84375.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{Q},\\mathbf{K}" + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "text", + "content": " reside in " + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "inline_equation", + "content": "\\mathbb{C}^{C\\times h\\times (\\frac{w}{2} +1)}" + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "text", + "content": " due to the Hermitian symmetry property [6]. Note that each frequency domain representation of " + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{Q},\\mathbf{K}" + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "text", + "content": " includes real and imaginary parts. So, we decompose " + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{Q},\\mathbf{K}" + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 55, + 285, + 296, + 346 + ], + "type": "text", + "content": " into real and imaginary parts as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 355, + 295, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 355, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 64, + 355, + 295, + 369 + ], + "type": "interline_equation", + "content": "\\left[ \\mathbf {Q} _ {\\mathbb {R}} \\mid \\mathbf {K} _ {\\mathbb {R}} \\mid \\mathbf {V} _ {\\mathbb {R}} \\right] + j \\left[ \\mathbf {Q} _ {\\mathbb {I}} \\mid \\mathbf {K} _ {\\mathbb {I}} \\mid \\mathbf {V} _ {\\mathbb {I}} \\right] = \\left[ \\mathbf {Q} \\mid \\mathbf {K} \\mid \\mathbf {V} \\right], \\tag {8}", + "image_path": "d0d31a5b02b90c47324f7b83deb0b80595b2912b940911f2e3c3778f0ad7344f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 378, + 296, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 296, + 415 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 378, + 296, + 415 + ], + "type": "inline_equation", + "content": "j = \\sqrt{-1}" + }, + { + "bbox": [ + 55, + 378, + 296, + 415 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 378, + 296, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathbb{R}}" + }, + { + "bbox": [ + 55, + 378, + 296, + 415 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 378, + 296, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{\\mathbb{R}}" + }, + { + "bbox": [ + 55, + 378, + 296, + 415 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 378, + 296, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{\\mathbb{R}}" + }, + { + "bbox": [ + 55, + 378, + 296, + 415 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 378, + 296, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathbb{I}}" + }, + { + "bbox": [ + 55, + 378, + 296, + 415 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 378, + 296, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{\\mathbb{I}}" + }, + { + "bbox": [ + 55, + 378, + 296, + 415 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 378, + 296, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{\\mathbb{I}}" + }, + { + "bbox": [ + 55, + 378, + 296, + 415 + ], + "type": "text", + "content": " are in " + }, + { + "bbox": [ + 55, + 378, + 296, + 415 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C\\times h\\times (\\frac{w}{2} + 1)}" + }, + { + "bbox": [ + 55, + 378, + 296, + 415 + ], + "type": "text", + "content": ". Self-attention operations are then performed separately on the real and imaginary parts as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 424, + 295, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 424, + 295, + 444 + ], + "spans": [ + { + "bbox": [ + 107, + 424, + 295, + 444 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {l} = \\operatorname {S o f t M a x} \\left(\\mathbf {Q} _ {l} \\mathbf {K} _ {l} ^ {\\top} / \\sqrt {C}\\right) \\mathbf {V} _ {l}, \\tag {9}", + "image_path": "afa76a876e098541b6bbdfbee0e4ebca9689ee2a33d8c8e6d3cfb14d8861dc42.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 450, + 296, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 296, + 487 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 450, + 296, + 487 + ], + "type": "inline_equation", + "content": "l \\in \\{\\mathbb{R},\\mathbb{I}\\}" + }, + { + "bbox": [ + 55, + 450, + 296, + 487 + ], + "type": "text", + "content": " indicates the real or imaginary feature, respectively. Finally, the inverse Fourier Transform " + }, + { + "bbox": [ + 55, + 450, + 296, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{-1}" + }, + { + "bbox": [ + 55, + 450, + 296, + 487 + ], + "type": "text", + "content": " is applied to " + }, + { + "bbox": [ + 55, + 450, + 296, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathbb{R}}" + }, + { + "bbox": [ + 55, + 450, + 296, + 487 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 450, + 296, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathbb{I}}" + }, + { + "bbox": [ + 55, + 450, + 296, + 487 + ], + "type": "text", + "content": " to map them back to the time domain:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 129, + 496, + 295, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 496, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 129, + 496, + 295, + 510 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {n} \\leftarrow \\mathcal {F} ^ {- 1} \\left(\\mathbf {f} _ {\\mathbb {R}} + j \\mathbf {f} _ {\\mathbb {I}}\\right). \\tag {10}", + "image_path": "6a53e06c8e2e77ab80d2a6abe9d65b0a48f98ee5e3da5d3d5afe120c91117f9a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 520, + 295, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 543 + ], + "type": "text", + "content": "This process ensures that refined frequency information be selectively captured and utilized." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 544, + 296, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 296, + 676 + ], + "type": "text", + "content": "Stationary Wavelet Transform Cross Attention (SWTCA). The SwTCA block is designed to leverage Wavelet components of " + }, + { + "bbox": [ + 55, + 544, + 296, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{PAN}}" + }, + { + "bbox": [ + 55, + 544, + 296, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 544, + 296, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{LR}}" + }, + { + "bbox": [ + 55, + 544, + 296, + 676 + ], + "type": "text", + "content": " motivated by [5, 35, 40, 71]. As shown in Fig. 3, it enhances the features outputted by the FTCA block, utilizing the Stationary Wavelet Transform (SWT) [32], denoted as " + }, + { + "bbox": [ + 55, + 544, + 296, + 676 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 544, + 296, + 676 + ], + "type": "text", + "content": ". The SwTCA block selectively incorporates relevant frequency components from " + }, + { + "bbox": [ + 55, + 544, + 296, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{PAN}}" + }, + { + "bbox": [ + 55, + 544, + 296, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 544, + 296, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{LR}}" + }, + { + "bbox": [ + 55, + 544, + 296, + 676 + ], + "type": "text", + "content": " through a two-stage cross-attention process. The SwT-based conditioning terms (" + }, + { + "bbox": [ + 55, + 544, + 296, + 676 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 544, + 296, + 676 + ], + "type": "text", + "content": "-Cond) are constructed by decomposing " + }, + { + "bbox": [ + 55, + 544, + 296, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{PAN}}" + }, + { + "bbox": [ + 55, + 544, + 296, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 544, + 296, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{LR}}" + }, + { + "bbox": [ + 55, + 544, + 296, + 676 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 55, + 544, + 296, + 676 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 544, + 296, + 676 + ], + "type": "text", + "content": " into four components:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 89, + 685, + 295, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 685, + 295, + 717 + ], + "spans": [ + { + "bbox": [ + 89, + 685, + 295, + 717 + ], + "type": "interline_equation", + "content": "\\begin{array}{r l} & {\\left[ \\mathbf {L} _ {\\mathrm {P A N}} \\mid \\mathbf {H} _ {\\mathrm {P A N}} \\mid \\mathbf {V} _ {\\mathrm {P A N}} \\mid \\mathbf {D} _ {\\mathrm {P A N}} \\right] = \\mathcal {S} (\\mathbf {I} _ {\\mathrm {P A N}}),} \\\\ & {\\left[ \\mathbf {L} _ {\\mathrm {M S}} ^ {\\mathrm {L R}} \\mid \\mathbf {H} _ {\\mathrm {M S}} ^ {\\mathrm {L R}} \\mid \\mathbf {V} _ {\\mathrm {M S}} ^ {\\mathrm {L R}} \\mid \\mathbf {D} _ {\\mathrm {M S}} ^ {\\mathrm {L R}} \\right] = \\mathcal {S} (\\mathbf {I} _ {\\mathrm {M S}} ^ {\\mathrm {L R}}),} \\end{array} \\tag {11}", + "image_path": "47167f66fa97ad3a9bd2ef7a2dc54ed2410cc53fc56363c01a18b3e01f2131de.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " represent the low-frequency approximation components, while " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{H},\\mathbf{V},\\mathbf{D}" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " denote the high-frequency horizontal, vertical, and diagonal details, respectively. Since " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{PAN}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " provides high spatial resolution but lacks spectral information, only its high-frequency components " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{\\mathrm{PAN}}," + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{\\mathrm{PAN}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " ,and " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{PAN}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " are used. Conversely, " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{MS}}^{\\mathrm{LR}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{LR}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " contains richer spectral information. These components are concatenated channel-wise to form " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " -Cond as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 348, + 175, + 553, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 175, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 348, + 175, + 553, + 190 + ], + "type": "interline_equation", + "content": "\\mathcal {S} - \\text {C o n d} = \\left[ \\mathbf {L} _ {\\mathrm {M S}} ^ {\\mathrm {L R}} \\mid \\mathbf {H} _ {\\mathrm {P A N}} \\mid \\mathbf {V} _ {\\mathrm {P A N}} \\mid \\mathbf {D} _ {\\mathrm {P A N}} \\right]. \\tag {12}", + "image_path": "9021715f6dbee5d6c6cbda4e70d6f38c722564968a8c1ead75565f3203cb5d00.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 195, + 555, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 195, + 555, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 195, + 555, + 280 + ], + "type": "text", + "content": "To inject " + }, + { + "bbox": [ + 313, + 195, + 555, + 280 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 195, + 555, + 280 + ], + "type": "text", + "content": "-Cond into the feature " + }, + { + "bbox": [ + 313, + 195, + 555, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_n" + }, + { + "bbox": [ + 313, + 195, + 555, + 280 + ], + "type": "text", + "content": " obtained from FTCA at the " + }, + { + "bbox": [ + 313, + 195, + 555, + 280 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 195, + 555, + 280 + ], + "type": "text", + "content": "-th decoder block, we employ a two-stage cross-attention process. In the first stage, " + }, + { + "bbox": [ + 313, + 195, + 555, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_1" + }, + { + "bbox": [ + 313, + 195, + 555, + 280 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 195, + 555, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_1" + }, + { + "bbox": [ + 313, + 195, + 555, + 280 + ], + "type": "text", + "content": " are derived from " + }, + { + "bbox": [ + 313, + 195, + 555, + 280 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 195, + 555, + 280 + ], + "type": "text", + "content": "-Cond, directing attention toward critical frequency components, while " + }, + { + "bbox": [ + 313, + 195, + 555, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_1" + }, + { + "bbox": [ + 313, + 195, + 555, + 280 + ], + "type": "text", + "content": " is derived from " + }, + { + "bbox": [ + 313, + 195, + 555, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_n" + }, + { + "bbox": [ + 313, + 195, + 555, + 280 + ], + "type": "text", + "content": ". The resulting features are reshaped to " + }, + { + "bbox": [ + 313, + 195, + 555, + 280 + ], + "type": "inline_equation", + "content": "C \\times hw" + }, + { + "bbox": [ + 313, + 195, + 555, + 280 + ], + "type": "text", + "content": " for channel attention [63], producing an intermediate feature " + }, + { + "bbox": [ + 313, + 195, + 555, + 280 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_n^1 \\in \\mathbb{R}^{C \\times h \\times w}" + }, + { + "bbox": [ + 313, + 195, + 555, + 280 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 332, + 296, + 522, + 309 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 296, + 522, + 309 + ], + "spans": [ + { + "bbox": [ + 332, + 296, + 522, + 309 + ], + "type": "interline_equation", + "content": "\\left[ \\mathbf {Q} _ {1} \\mid \\mathbf {K} _ {1} \\right] = \\operatorname {C o n v} (S - \\text {C o n d}), \\mathbf {V} _ {1} = \\operatorname {C o n v} (\\mathbf {f} _ {n}),", + "image_path": "fa1d3fe993f7933c095b0ef0ab92a703d65823d21dfdbee4422cae3c32eeada2.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 329, + 309, + 553, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 309, + 553, + 332 + ], + "spans": [ + { + "bbox": [ + 329, + 309, + 553, + 332 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {n} ^ {1} = \\operatorname {S o f t M a x} \\left(\\mathbf {Q} _ {1} \\mathbf {K} _ {1} ^ {\\top} / \\sqrt {C}\\right) \\mathbf {V} _ {1}, \\tag {13}", + "image_path": "23b686ff55f71a4446904006d278512dc50a82901a788667fa6535f293a8e5c8.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_1, \\mathbf{K}_1, \\mathbf{V}_1" + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "text", + "content": " are in " + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C \\times hw}" + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "text", + "content": ". In the second stage, " + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_2" + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "text", + "content": " is derived from the intermediate feature " + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_n^1" + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_2" + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_2" + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "text", + "content": " come from " + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "text", + "content": "-Cond, enabling the model to refine the correlation between the reconstructed feature " + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_n" + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "text", + "content": " and significant frequency components in " + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "text", + "content": "-Cond. This stage yields the final feature " + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_n \\in \\mathbb{R}^{C \\times h \\times w}" + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "text", + "content": " after reshaping:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 329, + 428, + 522, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 428, + 522, + 441 + ], + "spans": [ + { + "bbox": [ + 329, + 428, + 522, + 441 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {2} = \\operatorname {C o n v} \\left(\\mathbf {f} _ {n} ^ {1}\\right), \\left[ \\mathbf {K} _ {2} \\mid \\mathbf {V} _ {2} \\right] = \\operatorname {C o n v} (S - C o n d),", + "image_path": "23ef4359e465147126a235653e687b6077ac3112949cc127323df1bea787936f.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 331, + 441, + 553, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 441, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 331, + 441, + 553, + 464 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {n} \\leftarrow \\operatorname {S o f t M a x} \\left(\\mathbf {Q} _ {2} \\mathbf {K} _ {2} ^ {\\top} / \\sqrt {C}\\right) \\mathbf {V} _ {2}. \\tag {14}", + "image_path": "6d8f20de5bd900a3aa48b8b6564a60a25f051606ff4ac3a2de114793e94b08e4.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 470, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 554, + 518 + ], + "type": "text", + "content": "Through these two stages of cross-attention, the final feature " + }, + { + "bbox": [ + 313, + 470, + 554, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_n" + }, + { + "bbox": [ + 313, + 470, + 554, + 518 + ], + "type": "text", + "content": " effectively incorporates essential frequency components from both PAN and MS inputs, ensuring enhanced frequency representation and high-quality HRMS output." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 525, + 536, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 536, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 536, + 537 + ], + "type": "text", + "content": "3.4. Uncertainty-Aware Knowledge Distillation" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 542, + 554, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 554, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 554, + 577 + ], + "type": "text", + "content": "To address the substantial computational demands of the teacher (FSA-T) " + }, + { + "bbox": [ + 313, + 542, + 554, + 577 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 313, + 542, + 554, + 577 + ], + "type": "text", + "content": ", we propose an novel uncertainty-aware knowledge distillation (U-Know) strategy." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 578, + 554, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 554, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 554, + 661 + ], + "type": "text", + "content": "Student network (FSA-S) " + }, + { + "bbox": [ + 313, + 578, + 554, + 661 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 578, + 554, + 661 + ], + "type": "text", + "content": ". The student model FSA-S is designed with the same number of encoder and decoder blocks as FSA-T. Each block consists solely of Res-blocks, and unlike the teacher, the student operates without any additional conditional inputs. As shown in Fig. 2, the lightweight student model " + }, + { + "bbox": [ + 313, + 578, + 554, + 661 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 578, + 554, + 661 + ], + "type": "text", + "content": " predicts the denoised image " + }, + { + "bbox": [ + 313, + 578, + 554, + 661 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{X}}_0" + }, + { + "bbox": [ + 313, + 578, + 554, + 661 + ], + "type": "text", + "content": " from the input " + }, + { + "bbox": [ + 313, + 578, + 554, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t" + }, + { + "bbox": [ + 313, + 578, + 554, + 661 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 313, + 578, + 554, + 661 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 578, + 554, + 661 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 367, + 666, + 553, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 666, + 553, + 683 + ], + "spans": [ + { + "bbox": [ + 367, + 666, + 553, + 683 + ], + "type": "interline_equation", + "content": "\\widetilde {\\mathbf {X}} _ {0} = \\psi \\left(\\left[ \\mathbf {X} _ {t} \\mid \\mathbf {I} _ {\\text {P A N}} \\mid \\mathbf {I} _ {\\mathrm {M S}} ^ {\\mathrm {L R}} \\right]; t\\right). \\tag {15}", + "image_path": "72206a03eec4d528b3e5f31a298698fd15e95810c2524e91189e7416189aa2ce.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "U-Know loss function. FSA-T " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " generates an uncertainty map " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " along with the output " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{X}}_0" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " during training." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23073" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "These two outputs are frozen and used in the U-Know loss function, " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{U - Know}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": ", to train the FSA-S " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": ". The total training objective combines a hard loss " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{hard}})" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": ", a soft loss " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{soft}})" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": ", and a feature loss " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{feat}})" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " to distill frequency-rich features and enhance regions where " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " exhibits spatial weaknesses points. The total loss is defined as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 99, + 153, + 295, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 153, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 99, + 153, + 295, + 167 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {U - K n o w}} = \\mathcal {L} _ {\\text {h a r d}} + \\lambda_ {s} \\mathcal {L} _ {\\text {s o f t}} + \\lambda_ {f} \\mathcal {L} _ {\\text {f e a t}}, \\tag {16}", + "image_path": "d5b552fe6dab96983cd50266df4535176e9792ef5976fd7e09612d8ce73e6bbb.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 175, + 295, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 295, + 223 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 295, + 223 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 175, + 295, + 223 + ], + "type": "inline_equation", + "content": "\\lambda_{s}" + }, + { + "bbox": [ + 55, + 175, + 295, + 223 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 175, + 295, + 223 + ], + "type": "inline_equation", + "content": "\\lambda_{f}" + }, + { + "bbox": [ + 55, + 175, + 295, + 223 + ], + "type": "text", + "content": " control the contributions of " + }, + { + "bbox": [ + 55, + 175, + 295, + 223 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{soft}}" + }, + { + "bbox": [ + 55, + 175, + 295, + 223 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 175, + 295, + 223 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{feat}}" + }, + { + "bbox": [ + 55, + 175, + 295, + 223 + ], + "type": "text", + "content": " respectively. " + }, + { + "bbox": [ + 55, + 175, + 295, + 223 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{hard}}" + }, + { + "bbox": [ + 55, + 175, + 295, + 223 + ], + "type": "text", + "content": " measures the difference between " + }, + { + "bbox": [ + 55, + 175, + 295, + 223 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 55, + 175, + 295, + 223 + ], + "type": "text", + "content": " output " + }, + { + "bbox": [ + 55, + 175, + 295, + 223 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{X}}_0" + }, + { + "bbox": [ + 55, + 175, + 295, + 223 + ], + "type": "text", + "content": " and the ground truth " + }, + { + "bbox": [ + 55, + 175, + 295, + 223 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_0" + }, + { + "bbox": [ + 55, + 175, + 295, + 223 + ], + "type": "text", + "content": ", weighted by " + }, + { + "bbox": [ + 55, + 175, + 295, + 223 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 55, + 175, + 295, + 223 + ], + "type": "text", + "content": "'s uncertainty map " + }, + { + "bbox": [ + 55, + 175, + 295, + 223 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}" + }, + { + "bbox": [ + 55, + 175, + 295, + 223 + ], + "type": "text", + "content": " to emphasize challenging regions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 100, + 230, + 295, + 253 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 230, + 295, + 253 + ], + "spans": [ + { + "bbox": [ + 100, + 230, + 295, + 253 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {h a r d}} = \\left\\| (\\tau + \\hat {\\boldsymbol {\\theta}}) \\odot \\left| \\widetilde {\\mathbf {X}} _ {0} - \\mathbf {X} _ {0} \\right| \\right\\| _ {1}, \\tag {17}", + "image_path": "2aeab3a570b457a78c30cbeb16788aff5b13c394e57c490e4ecf9bdbef2a23bf.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 259, + 296, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 259, + 296, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 296, + 295 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 259, + 296, + 295 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 259, + 296, + 295 + ], + "type": "text", + "content": " is a hyper-parameter. " + }, + { + "bbox": [ + 55, + 259, + 296, + 295 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{soft}}" + }, + { + "bbox": [ + 55, + 259, + 296, + 295 + ], + "type": "text", + "content": " computes the difference between " + }, + { + "bbox": [ + 55, + 259, + 296, + 295 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 55, + 259, + 296, + 295 + ], + "type": "text", + "content": " 's output " + }, + { + "bbox": [ + 55, + 259, + 296, + 295 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{X}}_0" + }, + { + "bbox": [ + 55, + 259, + 296, + 295 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 259, + 296, + 295 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 55, + 259, + 296, + 295 + ], + "type": "text", + "content": " 's output " + }, + { + "bbox": [ + 55, + 259, + 296, + 295 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{X}}_0" + }, + { + "bbox": [ + 55, + 259, + 296, + 295 + ], + "type": "text", + "content": ", weighted by the " + }, + { + "bbox": [ + 55, + 259, + 296, + 295 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}" + }, + { + "bbox": [ + 55, + 259, + 296, + 295 + ], + "type": "text", + "content": " to help " + }, + { + "bbox": [ + 55, + 259, + 296, + 295 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 55, + 259, + 296, + 295 + ], + "type": "text", + "content": " better approximate the " + }, + { + "bbox": [ + 55, + 259, + 296, + 295 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 55, + 259, + 296, + 295 + ], + "type": "text", + "content": " 's knowledge:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 101, + 302, + 295, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 302, + 295, + 324 + ], + "spans": [ + { + "bbox": [ + 101, + 302, + 295, + 324 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s o f t}} = \\left\\| (\\tau - \\hat {\\boldsymbol {\\theta}}) \\odot \\left| \\widetilde {\\mathbf {X}} _ {0} - \\widehat {\\mathbf {X}} _ {0} \\right| \\right\\| _ {1}. \\tag {18}", + "image_path": "5ad04330bd1badf1e7566ae1e547d107dc741d66bbf92f6a377be0d25327a62b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "text", + "content": "In the regions where " + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{\\theta}}" + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "text", + "content": " values are high, indicating that " + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "text", + "content": " struggles with the task, we apply a weight of " + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "inline_equation", + "content": "(\\tau + \\hat{\\pmb{\\theta}})" + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "text", + "content": " and use " + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_0" + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "text", + "content": " as the hard ground truth to guide " + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "text", + "content": " toward the true target. Conversely, in the areas where " + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{\\theta}}" + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "text", + "content": " values are low, indicating that " + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "text", + "content": " performs well, we apply a weight of " + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "inline_equation", + "content": "(\\tau - \\hat{\\pmb{\\theta}})" + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "text", + "content": " and use " + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "text", + "content": "'s output " + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{X}}_0" + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "text", + "content": " as the soft ground truth, allowing " + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "text", + "content": " to approximate " + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "text", + "content": "'s knowledge in these regions. These dual loss formulations enable " + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "text", + "content": " to focus on challenging areas while effectively leveraging the " + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "text", + "content": "'s strengths. " + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{feat}}" + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "text", + "content": " ensures alignment between intermediate features of " + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 55, + 334, + 296, + 456 + ], + "type": "text", + "content": " as [4]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 464, + 295, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 464, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 70, + 464, + 295, + 485 + ], + "type": "interline_equation", + "content": "\\ell_ {\\text {f e a t}} ^ {i} = \\sqrt {\\left\\| \\mathbf {f} _ {i} - \\mathbf {s} _ {i} \\right\\| _ {1} ^ {2} + \\gamma}, \\quad \\mathcal {L} _ {\\text {f e a t}} = \\sum_ {i} \\alpha_ {i} \\ell_ {\\text {f e a t}} ^ {i}, \\tag {19}", + "image_path": "8702ad0c58dd49c1b62ef69472e5b7df2ecb6476e0085d957dbec1cfc1fdef0f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 494, + 296, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 296, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 296, + 541 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 494, + 296, + 541 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{feat}}^{i}" + }, + { + "bbox": [ + 55, + 494, + 296, + 541 + ], + "type": "text", + "content": " is a feature loss for the " + }, + { + "bbox": [ + 55, + 494, + 296, + 541 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 494, + 296, + 541 + ], + "type": "text", + "content": "-th feature, " + }, + { + "bbox": [ + 55, + 494, + 296, + 541 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 55, + 494, + 296, + 541 + ], + "type": "text", + "content": " is a weight for the " + }, + { + "bbox": [ + 55, + 494, + 296, + 541 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 494, + 296, + 541 + ], + "type": "text", + "content": "-th feature, and " + }, + { + "bbox": [ + 55, + 494, + 296, + 541 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 55, + 494, + 296, + 541 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 494, + 296, + 541 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i" + }, + { + "bbox": [ + 55, + 494, + 296, + 541 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 55, + 494, + 296, + 541 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 494, + 296, + 541 + ], + "type": "text", + "content": "-th intermediate feature maps of " + }, + { + "bbox": [ + 55, + 494, + 296, + 541 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 55, + 494, + 296, + 541 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 494, + 296, + 541 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 55, + 494, + 296, + 541 + ], + "type": "text", + "content": " respectively. The small constant " + }, + { + "bbox": [ + 55, + 494, + 296, + 541 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 494, + 296, + 541 + ], + "type": "text", + "content": " ensures smooth convergence during optimization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 552, + 132, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 132, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 132, + 565 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 571, + 258, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 258, + 585 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 258, + 585 + ], + "type": "text", + "content": "4.1. Datasets, Metrics and Training Details" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 61, + 598, + 294, + 677 + ], + "blocks": [ + { + "bbox": [ + 61, + 598, + 294, + 677 + ], + "lines": [ + { + "bbox": [ + 61, + 598, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 61, + 598, + 294, + 677 + ], + "type": "table", + "html": "
SatelliteWorldView-3QuickBirdGaoFen-2
Number of Band844
Spatial Resolution (m)PAN0.30.60.8
LRMS1.22.43.2
Radiometric Resolution (bit)111110
Number of (Train / Test) Images9,714 / 2017,139 / 2019,809 / 20
Patch SizePAN64×64×164×64×164×64×1
LRMS16×16×816×16×416×16×4
", + "image_path": "a43a7358a918bb182e443964aee5cc75d390562015300b9a6be4e8f5b3f60841.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 685, + 295, + 706 + ], + "lines": [ + { + "bbox": [ + 55, + 685, + 295, + 706 + ], + "spans": [ + { + "bbox": [ + 55, + 685, + 295, + 706 + ], + "type": "text", + "content": "Table 1. Detailed information of Worldview-3, QuickBird, and GaoFen-2 datasets." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 72, + 555, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 289 + ], + "type": "text", + "content": "Table 1 summarizes the characteristics of three datasets from PanCollection [11] for our experiments: WorldView-3 (WV3), QuickBird (QB), and GaoFen-2 (GF2). All experiments were conducted with a batch size of 32 for 300K iterations per dataset on an NVIDIA GeForce RTX 3090 GPU. The initial learning rate was set to " + }, + { + "bbox": [ + 313, + 72, + 555, + 289 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 313, + 72, + 555, + 289 + ], + "type": "text", + "content": " with a decay factor of 0.5 applied every 10K iterations, using the AdamW optimizer [12]. For the diffusion process of our FSA-T and FSA-S, the total diffusion time step " + }, + { + "bbox": [ + 313, + 72, + 555, + 289 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 72, + 555, + 289 + ], + "type": "text", + "content": " was set to 500, with the DDIM [42] sampling strategy used to generate HRMS images in 25 steps. To evaluate the model performance, we employed widely accepted metrics that capture various aspects of image quality and similarity. For the reduced-resolution datasets of WV3, QB, and GF2, we used six standard metrics: PSNR [51], SSIM [51], SAM [62], ERGAS [46], SCC [16], and Q4/Q8 [44]. For the full-resolution datasets of WV3, QB, and GF2, we used HQNR [45] with " + }, + { + "bbox": [ + 313, + 72, + 555, + 289 + ], + "type": "inline_equation", + "content": "D_{\\lambda}" + }, + { + "bbox": [ + 313, + 72, + 555, + 289 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 72, + 555, + 289 + ], + "type": "inline_equation", + "content": "D_{S}" + }, + { + "bbox": [ + 313, + 72, + 555, + 289 + ], + "type": "text", + "content": " to comprehensively assess the results." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 300, + 437, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 437, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 437, + 312 + ], + "type": "text", + "content": "4.2. Experimental Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 319, + 555, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 378 + ], + "type": "text", + "content": "To ensure a thorough and equitable evaluation, we used the official codes for each compared method whenever available. For methods without official codes, we re-implemented them based on the papers [31, 56] to maintain fairness in comparison." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 380, + 556, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 556, + 595 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 556, + 595 + ], + "type": "text", + "content": "Quantitative analysis. As shown in Table 2 and Table 3, our FSA-T achieves superior performance on the WV3, QB, and GF2 datasets compared to all other methods. The GF2 dataset contains relatively easier (less complex) images, so the teacher model can fully leverage its high-capacity features. In contrast, for the more challenging (more complex) WV3 and QB datasets, our FSA-S surpasses FSA-T across several metrics. This improvement represents the effectiveness of our uncertainty-aware knowledge distillation (U-Know) strategy, which enables the student model (FSA-S) to address the spatial weaknesses of the teacher model by focusing on uncertain regions. From this, the FSA-S can benefit more generalization from the knowledge distillation with uncertainty maps, leading to better performances than the FSA-T for the test datasets. Consequently, our FSA-S demonstrates an enhanced adaptability and robustness on complex datasets, where precise handling of spatial details (high frequency components) is essential." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 596, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 596, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 596, + 556, + 714 + ], + "type": "text", + "content": "Qualitative analysis. Fig. 4 shows the output HRMS images " + }, + { + "bbox": [ + 313, + 596, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_{\\mathrm{MS}}^{\\mathrm{HR}}" + }, + { + "bbox": [ + 313, + 596, + 556, + 714 + ], + "type": "text", + "content": " in the first rows and the error maps (MAE images) between " + }, + { + "bbox": [ + 313, + 596, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_{\\mathrm{MS}}^{\\mathrm{HR}}" + }, + { + "bbox": [ + 313, + 596, + 556, + 714 + ], + "type": "text", + "content": " and ground truth " + }, + { + "bbox": [ + 313, + 596, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{HR}}" + }, + { + "bbox": [ + 313, + 596, + 556, + 714 + ], + "type": "text", + "content": ". As shown, our framework generates the results " + }, + { + "bbox": [ + 313, + 596, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_{\\mathrm{MS}}^{\\mathrm{HR}}" + }, + { + "bbox": [ + 313, + 596, + 556, + 714 + ], + "type": "text", + "content": " closer to " + }, + { + "bbox": [ + 313, + 596, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{HR}}" + }, + { + "bbox": [ + 313, + 596, + 556, + 714 + ], + "type": "text", + "content": " in the reduced resolution GF2 data set. Furthermore, Fig. 1 demonstrates that even in the full-resolution comparison where ground truth is unavailable, our framework enables to generate more detailed and stable results compared to all other state-of-the-art approaches." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23074" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 552, + 213 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 552, + 213 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 552, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 552, + 213 + ], + "type": "table", + "html": "
GF2 DatasetReduced-ResolutionFull-Resolution
TypesMethodsPublicationsPSNR↑SSIM↑SAM↓ERGAS↓SCC↑Q4↑Dλ↓Ds↓HQNR↑
Non-DMsPanNet [58]ICCV 201739.1970.9591.0501.0380.9750.9630.0200.0520.929
MSDCNN [61]JSTARS 201840.7300.9710.9460.8620.9830.9720.0260.0790.898
FusionNet [53]ICCV 202139.8660.9660.9710.9600.9800.9670.0340.1050.865
LAGConv [25]AAAI 202241.1470.9740.8860.8160.9850.9740.0300.0780.895
S2DBPN [64]TGRS 202342.6860.9800.7720.6860.9900.9810.0200.0460.935
DCPNet [68]TGRS 202442.3120.9790.8060.7240.9880.9800.0240.0240.953
CANConv [14]CVPR 202443.1660.9820.7220.6530.9910.9830.0190.0630.919
Diffusion ModelsPanDiff [31]TGRS 202342.8270.9800.7670.6740.9900.9810.0200.0450.936
TMDiff [56]TGRS 202441.8960.9770.7640.7540.9880.9790.0290.0300.942
FSA-SOurs44.5850.9860.6240.5480.9930.9870.0180.0370.944
FSA-TOurs44.7570.9880.6030.5370.9940.9880.0170.0290.953
", + "image_path": "5aebfec8f96cd6fd21a2b67324d157abdfbe4af1af70fa2111c16cd78b7c3b1e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 61, + 248, + 552, + 379 + ], + "blocks": [ + { + "bbox": [ + 55, + 216, + 553, + 239 + ], + "lines": [ + { + "bbox": [ + 55, + 216, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 553, + 239 + ], + "type": "text", + "content": "Table 2. Comparison of different models on the GaoFen-2 (GF2) dataset. Blue indicates the second-best performance, while red highlights the best-performing model. Standard deviations of the metrics across test samples are provided in the Supplemental." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 248, + 552, + 379 + ], + "lines": [ + { + "bbox": [ + 61, + 248, + 552, + 379 + ], + "spans": [ + { + "bbox": [ + 61, + 248, + 552, + 379 + ], + "type": "table", + "html": "
TypesMethodsWV3 Dataset (Reduced-Resolution)QB Dataset (Reduced-Resolution)
PSNR↑SSIM↑SAM↓ERGAS↓SCC↑Q8↑PSNR↑SSIM↑SAM↓ERGAS↓SCC↑Q4↑
Non-DMsPanNet [58]36.1480.9663.4022.5380.9790.91335.5630.9395.2734.8560.9660.911
MSDCNN [61]36.3290.9673.3002.4890.9790.91437.0400.9544.8284.0740.9770.925
FusionNet [53]36.5690.9683.1882.4280.9810.91636.8210.9524.8924.1830.9750.923
LAGConv [25]36.7320.9703.1532.3800.9810.91637.5650.9584.6823.8450.9800.930
S2DBPN [64]37.2160.9723.0192.2450.9850.91737.3140.9564.8493.9560.9800.928
DCPNet [68]37.0090.9723.0832.3010.9840.91538.0790.9634.4203.6180.9830.935
CANConv [14]37.4410.9732.9272.1630.9850.91837.7950.9604.5543.7400.9820.935
Diffusion ModelsPanDiff [31]37.0290.9713.0582.2760.9840.91337.8420.9594.6113.7230.9820.935
TMDiff [56]37.4770.9732.8852.1510.9860.91537.6420.9584.6273.8040.9810.930
FSA-S37.9300.9762.7972.0460.9880.92238.3610.9644.3373.5000.9840.938
FSA-T37.8940.9762.8012.0550.9870.92138.3430.9644.3493.5020.9850.938
", + "image_path": "9ffbd1959f59ea5589663053d82b40ff031fc0ae5076fb74b43214f03e57bfc7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 57, + 415, + 553, + 570 + ], + "blocks": [ + { + "bbox": [ + 57, + 415, + 553, + 570 + ], + "lines": [ + { + "bbox": [ + 57, + 415, + 553, + 570 + ], + "spans": [ + { + "bbox": [ + 57, + 415, + 553, + 570 + ], + "type": "image", + "image_path": "89a59f39da9ed831a7d68e9e447c5fd4358c7edd9c0f3eb790a8d2b40f8e9a1e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 571, + 555, + 607 + ], + "lines": [ + { + "bbox": [ + 55, + 571, + 555, + 607 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 555, + 607 + ], + "type": "text", + "content": "Figure 4. Visual comparison of PAN-sharpening results on the reduced-resolution GF2 dataset. The first row shows RGB of outputs " + }, + { + "bbox": [ + 55, + 571, + 555, + 607 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{HR}}" + }, + { + "bbox": [ + 55, + 571, + 555, + 607 + ], + "type": "text", + "content": " and the second row displays Error Map, the difference between output " + }, + { + "bbox": [ + 55, + 571, + 555, + 607 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_{\\mathrm{MS}}^{\\mathrm{HR}}" + }, + { + "bbox": [ + 55, + 571, + 555, + 607 + ], + "type": "text", + "content": " and ground truth " + }, + { + "bbox": [ + 55, + 571, + 555, + 607 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{HR}}" + }, + { + "bbox": [ + 55, + 571, + 555, + 607 + ], + "type": "text", + "content": ". Both FSA-T and FSA-S achieve more detailed results compared to state-of-the-art models." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 77, + 620, + 278, + 666 + ], + "blocks": [ + { + "bbox": [ + 55, + 381, + 555, + 404 + ], + "lines": [ + { + "bbox": [ + 55, + 381, + 555, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 555, + 404 + ], + "type": "text", + "content": "Table 3. Performance comparison of different models on WV3 and QB datasets. Red and blue highlights the best- and 2nd-best performing models. The results for the full-resolution datasets and detailed results with standard deviations are provided in the Supplemental." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 77, + 620, + 278, + 666 + ], + "lines": [ + { + "bbox": [ + 77, + 620, + 278, + 666 + ], + "spans": [ + { + "bbox": [ + 77, + 620, + 278, + 666 + ], + "type": "table", + "html": "
MethodsParams. (M)FLOPs (T)Time (s)Memory (GB)
PanDiff [31]12.5560.47119.5223.260
TMDiff [56]153.9395.51767.46110.483
FSA-T25.4921.40225.4955.910
FSA-S9.1150.34612.2872.136
", + "image_path": "29ec0caa76aea8560757a3939ec0b5df7c4b80de311f6b007f9a87d4370271d6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 674, + 295, + 696 + ], + "lines": [ + { + "bbox": [ + 55, + 674, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 674, + 295, + 696 + ], + "type": "text", + "content": "Table 4. Computational complexity comparison for diffusion-based models. Best values are highlighted in red." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 622, + 555, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 622, + 555, + 708 + ], + "spans": [ + { + "bbox": [ + 313, + 622, + 555, + 708 + ], + "type": "text", + "content": "Computational complexity analysis. Table 4 compares the computational cost of diffusion-based models, detailing parameters (Params.), floating-point operations (FLOPs), inference time (Time) and memory usage. This demonstrates the efficiency of our FSA-S over other models. In particular, the TMDiff has significantly higher numbers of parameters and FLOPs. The FSA-T, while computationally" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23075" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "intensive than the FSA-S, still provides a balanced trade-off between efficiency and performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 105, + 157, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 105, + 157, + 118 + ], + "spans": [ + { + "bbox": [ + 55, + 105, + 157, + 118 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 132, + 296, + 269 + ], + "blocks": [ + { + "bbox": [ + 58, + 132, + 296, + 269 + ], + "lines": [ + { + "bbox": [ + 58, + 132, + 296, + 269 + ], + "spans": [ + { + "bbox": [ + 58, + 132, + 296, + 269 + ], + "type": "image", + "image_path": "419ae6fa61c9328953033815efb5e94c57c0cb2454f4265898c0e5e772a63e4f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 275, + 296, + 300 + ], + "lines": [ + { + "bbox": [ + 55, + 275, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 296, + 300 + ], + "type": "text", + "content": "Figure 5. Visualization of uncertainty map " + }, + { + "bbox": [ + 55, + 275, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}" + }, + { + "bbox": [ + 55, + 275, + 296, + 300 + ], + "type": "text", + "content": ", Error Map, and ground truth " + }, + { + "bbox": [ + 55, + 275, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{HR}}" + }, + { + "bbox": [ + 55, + 275, + 296, + 300 + ], + "type": "text", + "content": " of reduced WV3." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 308, + 296, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 296, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 296, + 417 + ], + "type": "text", + "content": "Visualization of uncertainty map " + }, + { + "bbox": [ + 55, + 308, + 296, + 417 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}" + }, + { + "bbox": [ + 55, + 308, + 296, + 417 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 308, + 296, + 417 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 55, + 308, + 296, + 417 + ], + "type": "text", + "content": ". The uncertainty map " + }, + { + "bbox": [ + 55, + 308, + 296, + 417 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}" + }, + { + "bbox": [ + 55, + 308, + 296, + 417 + ], + "type": "text", + "content": " learned through Eq. 4 shows that the areas with low uncertainty appear dark, while the areas with high uncertainty appear bright. Fig. 5 visualizes the uncertainty maps " + }, + { + "bbox": [ + 55, + 308, + 296, + 417 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}" + }, + { + "bbox": [ + 55, + 308, + 296, + 417 + ], + "type": "text", + "content": ", error maps and ground truth " + }, + { + "bbox": [ + 55, + 308, + 296, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{HR}}" + }, + { + "bbox": [ + 55, + 308, + 296, + 417 + ], + "type": "text", + "content": " for the reduced WV3 dataset. As shown in Fig. 5, it can be noted that the high-uncertainty regions mainly correspond to high-frequency areas such as challenging object edges. The learned uncertainty maps tend to resemble the error maps." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 59, + 426, + 294, + 475 + ], + "blocks": [ + { + "bbox": [ + 59, + 426, + 294, + 475 + ], + "lines": [ + { + "bbox": [ + 59, + 426, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 59, + 426, + 294, + 475 + ], + "type": "table", + "html": "
EncoderDecoderGF2 Dataset (Reduced-Resolution)
FFAHQFESAM↓ERGAS↓SCC↑Q4↑
0.654 ± 0.1120.661 ± 0.0760.992 ± 0.0010.986 ± 0.007
✓0.654 ± 0.1120.636 ± 0.0770.993 ± 0.0020.986 ± 0.007
✓0.617 ± 0.1170.556 ± 0.1040.993 ± 0.0020.987 ± 0.007
✓✓0.603 ± 0.1020.537 ± 0.0770.994 ± 0.0010.988 ± 0.006
", + "image_path": "310414b3c0dcfb7ae2eb96b3b390262b0468a2793b736d27327cacc9fcde5519.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 518, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 295, + 640 + ], + "type": "text", + "content": "Effect of FFA, HQFE blocks in FSA-T. Table 5 presents the effectiveness of FFA and HQFE. As noted, there are a significant improvement across all metrics. The performance gains are much greater than when FFA is only applied to the encoder or HQFE only to the decoder, as simultaneous condition injection creates a synergistic effect with both FFA and HQFE because FFA sets up initial styles or \"sketches\" for " + }, + { + "bbox": [ + 55, + 518, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{MS}}^{\\mathrm{HR}}" + }, + { + "bbox": [ + 55, + 518, + 295, + 640 + ], + "type": "text", + "content": " images, followed by the decoder that restores frequency details via the HQFE. This step-by-step learning greatly improves overall performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": "Impact of the U-Know loss. To verify the effectiveness of our uncertainty-aware knowledge distillation loss " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{U - Know}}" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": " (U-Know loss), we trained the student model (FSA-S) using three losses: a widely used " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": ", a conventional knowledge distillation loss " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{KD}}" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": " which is same from our method " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{U - Know}}" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": " but without considering the uncertainty" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 337, + 70, + 537, + 121 + ], + "blocks": [ + { + "bbox": [ + 55, + 482, + 296, + 507 + ], + "lines": [ + { + "bbox": [ + 55, + 482, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 296, + 507 + ], + "type": "text", + "content": "Table 5. Comparison of Results with and without FFA and HQFE Blocks in FSA-T. Best values are highlighted in red." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 337, + 70, + 537, + 121 + ], + "lines": [ + { + "bbox": [ + 337, + 70, + 537, + 121 + ], + "spans": [ + { + "bbox": [ + 337, + 70, + 537, + 121 + ], + "type": "table", + "html": "
Loss func.GF2 Dataset (Full-Resolution)
Dλ↓Ds↓HQNR ↑
L10.026 ± 0.0140.040 ± 0.0170.935 ± 0.020
LKD0.025 ± 0.0150.038 ± 0.0160.938 ± 0.021
LU-Know0.018 ± 0.0110.037 ± 0.0070.944 ± 0.012
", + "image_path": "c84454788afe5523ce9e9764149407bb6b2099507dc493b37208901f8f2f6415.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 129, + 555, + 152 + ], + "lines": [ + { + "bbox": [ + 313, + 129, + 555, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 129, + 555, + 152 + ], + "type": "text", + "content": "Table 6. Impact of our U-Know loss function design. Best values are highlighted in red." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 166, + 555, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 166, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 166, + 555, + 266 + ], + "type": "text", + "content": "map, and our " + }, + { + "bbox": [ + 313, + 166, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{U - Know}}" + }, + { + "bbox": [ + 313, + 166, + 555, + 266 + ], + "type": "text", + "content": ". Table 6 shows the performance comparisons for the three losses. As shown, when the FSA-S was trained with " + }, + { + "bbox": [ + 313, + 166, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{U - Know}}" + }, + { + "bbox": [ + 313, + 166, + 555, + 266 + ], + "type": "text", + "content": ", it achieved the best performance across all three metrics. This highlights the effectiveness of " + }, + { + "bbox": [ + 313, + 166, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{U - Know}}" + }, + { + "bbox": [ + 313, + 166, + 555, + 266 + ], + "type": "text", + "content": " that enables the FSA-S to better leverage the teacher (FSA-T)'s prior knowledge including the uncertainty maps " + }, + { + "bbox": [ + 313, + 166, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}" + }, + { + "bbox": [ + 313, + 166, + 555, + 266 + ], + "type": "text", + "content": ", output features " + }, + { + "bbox": [ + 313, + 166, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 313, + 166, + 555, + 266 + ], + "type": "text", + "content": ", and output HRMS " + }, + { + "bbox": [ + 313, + 166, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_{\\mathrm{MS}}^{\\mathrm{HR}}" + }, + { + "bbox": [ + 313, + 166, + 555, + 266 + ], + "type": "text", + "content": " for improved learning." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 271, + 392, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 392, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 392, + 283 + ], + "type": "text", + "content": "4.4. Limitations" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 289, + 555, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 410 + ], + "type": "text", + "content": "Inference time limitations. Despite U-Know-DiffPAN's superior performance across multiple metrics, one notable limitation lies in its inference speed, compared to nondiffusion-based models. Diffusion models unavoidably require multiple sampling steps during inference, making this process inevitably slower. This iterative approach, while being beneficial for image generation of high quality, limits the model's efficiency in real-time applications. Future work could focus on optimizing the inference process by pruning sampling steps, while maintaining performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 418, + 388, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 388, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 388, + 431 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 438, + 556, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 556, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 556, + 608 + ], + "type": "text", + "content": "We introduced U-Know-DiffPAN, a novel diffusion-based PAN-sharpening framework that effectively leverages frequency information from PAN and LRMS inputs to generate high-quality HRMS images. Our Uncertainty-aware Knowledge Distillation (U-Know) strategy enables efficient feature transfer by distilling both frequency-rich features and uncertainty maps from the teacher model, enhancing the restoration performance for spatially weak regions while reducing computational costs for the student model (FSA-S). Extensive experiments on multiple datasets, including WV3, QB, and GF2, demonstrate that our U-Know-DiffPAN outperforms very recent state-of-the-art methods in PAN-sharpening, establishing a new benchmark for performance and efficiency in satellite image processing." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 616, + 419, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 419, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 419, + 630 + ], + "type": "text", + "content": "6. Acknowledgment" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 635, + 556, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 556, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 556, + 704 + ], + "type": "text", + "content": "This work was supported by National Research Foundation of Korea (NRF) grant funded by the Korean Government [Ministry of Science and ICT (Information and Communications Technology)] (Project Number: RS-2024-00338513, Project Title: AI-based Computer Vision Study for Satellite Image Processing and Analysis, " + }, + { + "bbox": [ + 313, + 635, + 556, + 704 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 313, + 635, + 556, + 704 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23076" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Bruno Aiazzi, Luciano Alparone, Stefano Baronti, Andrea Garzelli, and Massimo Selva. Mtf-tailored multiscale fusion of high-resolution ms and pan imagery. Photogrammetric Engineering & Remote Sensing, 72(5):591-596, 2006. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "type": "text", + "content": "[2] Israa Amro, Javier Mateos, Miguel Vega, Rafael Molina, and Aggelos K Katsaggelos. A survey of classical methods and new trends in pansharpening of multispectral images. EURASIP Journal on Advances in Signal Processing, 2011:1-22, 2011. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 295, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 295, + 235 + ], + "type": "text", + "content": "[3] Coloma Ballester, Vicent Caselles, Laura Igual, Joan Verdera, and Bernard Rouge. A variational model for " + }, + { + "bbox": [ + 62, + 191, + 295, + 235 + ], + "type": "inline_equation", + "content": "\\mathfrak{p}+" + }, + { + "bbox": [ + 62, + 191, + 295, + 235 + ], + "type": "text", + "content": " xs image fusion. International Journal of Computer Vision, 69:43-58, 2006. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 295, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 295, + 278 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 295, + 278 + ], + "type": "text", + "content": "[4] Jonathan T Barron. A general and adaptive robust loss function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4331-4339, 2019. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 280, + 295, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 295, + 323 + ], + "type": "text", + "content": "[5] Zihan Cao, Shiqi Cao, Liang-Jian Deng, Xiao Wu, Junming Hou, and Gemine Vivone. Diffusion model with disentangled modulations for sharpening multispectral and hyperspectral images. Information Fusion, 104:102158, 2024. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 324, + 294, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 324, + 294, + 356 + ], + "spans": [ + { + "bbox": [ + 62, + 324, + 294, + 356 + ], + "type": "text", + "content": "[6] Enrico Celeghini, Manuel Gadella, and Mariano A del Olmo. Hermite functions and fourier series. Symmetry, 13(5):853, 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 358, + 295, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 358, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 358, + 295, + 423 + ], + "type": "text", + "content": "[7] Wei-Chun Chen, Chia-Che Chang, and Che-Rung Lee. Knowledge distillation with feature maps for image classification. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part III 14, pages 200–215. Springer, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 425, + 295, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 425, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 425, + 295, + 468 + ], + "type": "text", + "content": "[8] Jaewan Choi, Kiyun Yu, and Yongil Kim. A new adaptive component-substitution-based satellite image fusion by using partial replacement. IEEE transactions on geoscience and remote sensing, 49(1):295-309, 2010. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 469, + 295, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 295, + 523 + ], + "type": "text", + "content": "[9] Matteo Ciotola, Sergio Vitale, Antonio Mazza, Giovanni Poggi, and Giuseppe Scarpa. Pansharpening by convolutional neural networks in the full resolution framework. IEEE Transactions on Geoscience and Remote Sensing, 60: 1-17, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 525, + 295, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 295, + 579 + ], + "type": "text", + "content": "[10] Matteo Ciotola, Giuseppe Guarino, Gemine Vivone, Giovanni Poggi, Jocelyn Chanussot, Antonio Plaza, and Giuseppe Scarpa. Hyperspectral pansharpening: Critical review, tools and future perspectives. arXiv preprint arXiv:2407.01355, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 580, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 295, + 645 + ], + "type": "text", + "content": "[11] Liang-Jian Deng, Gemine Vivone, Mercedes E Paoletti, Giuseppe Scarpa, Jiang He, Yongjun Zhang, Jocelyn Chanussot, and Antonio Plaza. Machine learning in pan-sharpening: A benchmark, from shallow to deep networks. IEEE Geoscience and Remote Sensing Magazine, 10(3): 279-315, 2022. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 668 + ], + "type": "text", + "content": "[12] P Kingma Diederik. Adam: A method for stochastic optimization. (No Title), 2014. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[13] Kento Doi and Akira Iwasaki. Sscnet: Spectral-spatial consistency optimization of cnn for pansharpening. In IGARSS 2019-2019 IEEE International Geoscience and Remote Sensing Symposium, pages 3141-3144. IEEE, 2019. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "text", + "content": "[14] Yule Duan, Xiao Wu, Haoyu Deng, and Liang-Jian Deng. Content-adaptive non-local convolution for remote sensing pansharpening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27738-27747, 2024. 1, 2, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 129, + 554, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 554, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 554, + 183 + ], + "type": "text", + "content": "[15] Joan Duran, Antoni Buades, Bartomeu Coll, Catalina Sbert, and Gwendoline Blanchet. A survey of pansharpening methods with a new band-decoupled variational model. ISPRS Journal of Photogrammetry and Remote Sensing, 125:78-105, 2017. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 184, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 553, + 217 + ], + "type": "text", + "content": "[16] Andrea Garzelli and Filippo Nencini. Hypercomplex quality assessment of multi/hyperspectral images. IEEE Geoscience and Remote Sensing Letters, 6(4):662-665, 2009. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 217, + 553, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 217, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 317, + 217, + 553, + 272 + ], + "type": "text", + "content": "[17] Anaïs Gastineau, Jean-François Aujol, Yannick Berthoumieu, and Christian Germain. A residual dense generative adversarial network for pansharpening with geometrical constraints. In 2020 IEEE International Conference on Image Processing (ICIP), pages 493-497. IEEE, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 272, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 272, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 317, + 272, + 553, + 316 + ], + "type": "text", + "content": "[18] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 316, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 316, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 317, + 316, + 553, + 338 + ], + "type": "text", + "content": "[19] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 338, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 553, + 360 + ], + "type": "text", + "content": "[20] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 361, + 553, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 553, + 393 + ], + "type": "text", + "content": "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 394, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 394, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 317, + 394, + 553, + 437 + ], + "type": "text", + "content": "[22] Jianwen Hu, Pei Hu, Xudong Kang, Hui Zhang, and Shaosheng Fan. Pan-sharpening via multiscale dynamic convolutional neural network. IEEE Transactions on Geoscience and Remote Sensing, 59(3):2231-2244, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 438, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 438, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 317, + 438, + 553, + 480 + ], + "type": "text", + "content": "[23] Wei Huang, Liang Xiao, Zhihui Wei, Hongyi Liu, and Songze Tang. A new pan-sharpening method with deep neural networks. IEEE Geoscience and Remote Sensing Letters, 12(5):1037-1041, 2015. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 482, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 525 + ], + "type": "text", + "content": "[24] Zeyi Huang, Yang Zou, BVK Kumar, and Dong Huang. Comprehensive attention self-distillation for weakly-supervised object detection. Advances in neural information processing systems, 33:16797-16807, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 526, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 581 + ], + "type": "text", + "content": "[25] Zi-Rong Jin, Tian-Jing Zhang, Tai-Xiang Jiang, Gemine Vivone, and Liang-Jian Deng. Lagconv: Local-context adaptive convolution kernels with global harmonic bias for pan-sharpening. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1113-1121, 2022. 2, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 582, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 553, + 624 + ], + "type": "text", + "content": "[26] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 625, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 679 + ], + "type": "text", + "content": "[27] Jaehyup Lee, Soomin Seo, and Munchurl Kim. Sipsa-net: Shift-invariant pan sharpening with moving object alignment for satellite imagery. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10166-10174, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 681, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 714 + ], + "type": "text", + "content": "[28] Xiaobo Liu, Xiang Liu, Haoran Dai, Xudong Kang, Antonio Plaza, and Wenjie Zu. Mun-gan: A multiscale unsupervised network for remote sensing image pansharpening." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23077" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 94 + ], + "type": "text", + "content": "IEEE Transactions on Geoscience and Remote Sensing, 61: 1-18, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "text", + "content": "[29] Laetitia Loncan, Luis B De Almeida, José M Bioucas-Dias, Xavier Briottet, Jocelyn Chanussot, Nicolas Dobigeon, Sophie Fabre, Wenzhi Liao, Giorgio A Licciardi, Miguel Simoes, et al. Hyperspectral pansharpening: A review. IEEE Geoscience and remote sensing magazine, 3(3):27-46, 2015. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 207 + ], + "type": "text", + "content": "[30] Jiayi Ma, Wei Yu, Chen Chen, Pengwei Liang, Xiaojie Guo, and Junjun Jiang. Pan-gan: An unsupervised pan-sharpening method for remote sensing image fusion. Information Fusion, 62:110-120, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "text", + "content": "[31] Qingyan Meng, Wenxu Shi, Sijia Li, and Linlin Zhang. Pandiff: A novel pansharpening method based on denoising diffusion probabilistic model. IEEE Transactions on Geoscience and Remote Sensing, 61:1-17, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "text", + "content": "[32] Guy P Nason and Bernard W Silverman. The stationary wavelet transform and some statistical applications. In Wavelets and statistics, pages 281-299. Springer, 1995. 3, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 342 + ], + "type": "text", + "content": "[33] Qian Ning, Weisheng Dong, Xin Li, Jinjian Wu, and Guangming Shi. Uncertainty-driven loss for single image superresolution. Advances in Neural Information Processing Systems, 34:16398-16409, 2021. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 344, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 294, + 398 + ], + "type": "text", + "content": "[34] Xavier Otazu, María González-Audímaca, Octavi Fors, and Jorge Núñez. Introduction of sensor spectral response into image fusion methods. application to wavelet-based methods. IEEE Transactions on Geoscience and Remote Sensing, 43(10):2376-2385, 2005. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 444 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 444 + ], + "type": "text", + "content": "[35] Hao Phung, Quan Dao, and Anh Tran. Wavelet diffusion models are fast and scalable image generators. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10199-10208, 2023. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 510 + ], + "type": "text", + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 578 + ], + "type": "text", + "content": "[37] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "text", + "content": "[38] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE transactions on pattern analysis and machine intelligence, 45(4):4713-4726, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "type": "text", + "content": "[39] Maximilian Seitzer, Arash Tavakoli, Dimitrije Antic, and Georg Martius. On the pitfalls of heteroscedastic uncertainty estimation with probabilistic neural networks. arXiv preprint arXiv:2203.09168, 2022. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "type": "text", + "content": "[40] Shuyao Shang, Zhengyang Shan, Guangxing Liu, LunQian Wang, XingHua Wang, Zekai Zhang, and Jinglin Zhang. Resdiff: Combining cnn and diffusion model for image" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "super-resolution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8975-8983, 2024. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 95, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 128 + ], + "type": "text", + "content": "[41] Mark J Shensa et al. The discrete wavelet transform: wedding the a trous and mallat algorithms. IEEE Transactions on signal processing, 40(10):2464-2482, 1992. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 161 + ], + "type": "text", + "content": "[42] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 162, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 217 + ], + "type": "text", + "content": "[43] Baoli Sun, Xinchen Ye, Baopu Li, Haojie Li, Zhihui Wang, and Rui Xu. Learning scene structure guidance via cross-task knowledge transfer for single depth super-resolution. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 7792-7801, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "type": "text", + "content": "[44] Gemine Vivone, Luciano Alparone, Jocelyn Chanussot, Mauro Dalla Mura, Andrea Garzelli, Giorgio A Licciardi, Rocco Restaino, and Lucien Wald. A critical comparison among pansharpening algorithms. IEEE Transactions on Geoscience and Remote Sensing, 53(5):2565-2586, 2014. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 272, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 348 + ], + "type": "text", + "content": "[45] Gemine Vivone, Mauro Dalla Mura, Andrea Garzelli, Rocco Restaino, Giuseppe Scarpa, Magnus O Ulfarsson, Luciano Alparone, and Jocelyn Chanussot. A new benchmark based on recent advances in multispectral pansharpening: Revisiting pansharpening with classical and emerging pansharpening methods. IEEE Geoscience and Remote Sensing Magazine, 9(1):53-81, 2020. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "text", + "content": "[46] Lucien Wald. Quality of high resolution synthesised images: Is there a simple criterion? In Third conference\" Fusion of Earth data: merging point measurements, raster maps and remotely sensed images\", pages 99-103. SEE/URISCA, 2000. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 404, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 449 + ], + "type": "text", + "content": "[47] Lucien Wald, Thierry Ranchin, and Marc Mangolini. Fusion of satellite images of different spatial resolutions: Assessing the quality of resulting images. Photogrammetric engineering and remote sensing, 63(6):691-699, 1997. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 449, + 555, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 555, + 492 + ], + "type": "text", + "content": "[48] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, pages 1-21, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "type": "text", + "content": "[49] Tao Wang, Li Yuan, Xiaopeng Zhang, and Jiashi Feng. Distilling object detectors with fine-grained feature imitation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4933-4942, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 537, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 591 + ], + "type": "text", + "content": "[50] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 606-615, 2018. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "type": "text", + "content": "[51] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 636, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 680 + ], + "type": "text", + "content": "[52] Stephen Weinstein and Paul Ebert. Data transmission by frequency-division multiplexing using the discrete fourier transform. IEEE transactions on Communication Technology, 19(5):628-634, 1971. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 681, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 714 + ], + "type": "text", + "content": "[53] Xiao Wu, Ting-Zhu Huang, Liang-Jian Deng, and Tian-Jing Zhang. Dynamic cross feature fusion for remote sensing pan-sharpening. In Proceedings of the IEEE/CVF International" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23078" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 94 + ], + "type": "text", + "content": "Conference on Computer Vision, pages 14687-14696, 2021. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 128 + ], + "type": "text", + "content": "[54] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "text", + "content": "[55] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13095-13105, 2023. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 239 + ], + "type": "text", + "content": "[56] Yinghui Xing, Litao Qu, Shizhou Zhang, Jiapeng Feng, Xiuwei Zhang, and Yanning Zhang. Empower generalizability for pansharpening through text-modulated diffusion model. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "text", + "content": "[57] Kunran Xu, Lai Rui, Yishi Li, and Lin Gu. Feature normalized knowledge distillation for image classification. In European conference on computer vision, pages 664-680. Springer, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 284, + 294, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 294, + 335 + ], + "type": "text", + "content": "[58] Junfeng Yang, Xueyang Fu, Yuwen Hu, Yue Huang, Xinghao Ding, and John Paisley. Pannet: A deep network architecture for pan-sharpening. In Proceedings of the IEEE international conference on computer vision, pages 5449-5457, 2017. 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "type": "text", + "content": "[59] Zhendong Yang, Zhe Li, Ailing Zeng, Zexian Li, Chun Yuan, and Yu Li. Vitkd: Feature-based knowledge distillation for vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1379-1388, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 394, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 294, + 458 + ], + "type": "text", + "content": "[60] Lewei Yao, Renjie Pi, Hang Xu, Wei Zhang, Zhenguo Li, and Tong Zhang. G-detkd: Towards general distillation framework for object detectors via contrastive and semantic-guided feature imitation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3591-3600, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 294, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 524 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 524 + ], + "type": "text", + "content": "[61] Qiangqiang Yuan, Yancong Wei, Xiangchao Meng, Huanfeng Shen, and Liangpei Zhang. A multiscale and multidepth convolutional neural network for remote sensing imagery pan-sharpening. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 11(3):978-989, 2018. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 526, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 581 + ], + "type": "text", + "content": "[62] Roberta H Yuhas, Alexander FH Goetz, and Joe W Boardman. Discrimination among semi-arid landscape endmembers using the spectral angle mapper (sam) algorithm. In JPL, Summaries of the Third Annual JPL Airborne Geoscience Workshop. Volume 1: AVIRIS Workshop, 1992. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 582, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 645 + ], + "type": "text", + "content": "[63] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 691 + ], + "type": "text", + "content": "[64] Kai Zhang, Anfei Wang, Feng Zhang, Wenbo Wan, Jiande Sun, and Lorenzo Bruzzone. Spatial-spectral dual backprojection network for pansharpening. IEEE Transactions on Geoscience and Remote Sensing, 2023. 2, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[65] Yiman Zhang, Hanting Chen, Xinghao Chen, Yiping Deng, Chunjing Xu, and Yunhe Wang. Data-free knowledge dis" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 385 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "tillation for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7852-7861, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[66] Yun Zhang, Wei Li, Simiao Li, Hanting Chen, Zhijun Tu, Wenjia Wang, Bingyi Jing, Shaohui Lin, and Jie Hu. Data upcycling knowledge distillation for image super-resolution. arXiv preprint arXiv:2309.14162, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[67] Yaqun Zhang, Zhichang Guo, Dazhi Zhang, and Boying Wu. An anisotropic variational pansharpening model with adaptive coefficients. Inverse Problems and Imaging, pages 0–0, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "type": "text", + "content": "[68] Yafei Zhang, Xuji Yang, Huafeng Li, Minghong Xie, and Zhengtao Yu. Dcpnet: A dual-task collaborative promotion network for pansharpening. IEEE Transactions on Geoscience and Remote Sensing, 62:1-16, 2024. 2, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "text", + "content": "[69] Hao Zheng, Zhanlei Yang, Wenju Liu, Jizhong Liang, and Yanpeng Li. Improving deep neural networks using softplus units. In 2015 International joint conference on neural networks (IJCNN), pages 1-4. IEEE, 2015. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "type": "text", + "content": "[70] Wu Zheng, Weiliang Tang, Li Jiang, and Chi-Wing Fu. Sessd: Self-ensembling single-stage object detector from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14494–14503, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "type": "text", + "content": "[71] Jie Zhou, Daniel L Civco, and John A Silander. A wavelet transform method to merge landsat tm and spot panchromatic data. International journal of remote sensing, 19(4):743-757, 1998. 5" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23079" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/UA-Pose_ Uncertainty-Aware 6D Object Pose Estimation and Online Object Completion with Partial References/0d3743ff-3af5-445c-ae4c-ba5a33b9d7b4_content_list.json b/2025/UA-Pose_ Uncertainty-Aware 6D Object Pose Estimation and Online Object Completion with Partial References/0d3743ff-3af5-445c-ae4c-ba5a33b9d7b4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2c63323047427558d881019e34f6f2543aaf2b12 --- /dev/null +++ b/2025/UA-Pose_ Uncertainty-Aware 6D Object Pose Estimation and Online Object Completion with Partial References/0d3743ff-3af5-445c-ae4c-ba5a33b9d7b4_content_list.json @@ -0,0 +1,1566 @@ +[ + { + "type": "text", + "text": "UA-Pose: Uncertainty-Aware 6D Object Pose Estimation and Online Object Completion with Partial References", + "text_level": 1, + "bbox": [ + 104, + 128, + 867, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ming-Feng Li $^{1}$ Xin Yang $^{4}$ Fu-En Wang $^{4}$ Hritam Basak $^{2}$ Yuyin Sun $^{4}$ Shreekant Gayaka $^{4}$ Min Sun $^{3,4}$ Cheng-Hao Kuo $^{4}$", + "bbox": [ + 220, + 202, + 759, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Carnegie Mellon University $^{2}$ Stony Brook University $^{3}$ National Tsing Hua University $^{4}$ Amazon", + "bbox": [ + 122, + 256, + 844, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 327, + 313, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "6D object pose estimation has shown strong generalizability to novel objects. However, existing methods often require either a complete, well-reconstructed 3D model or numerous reference images that fully cover the object. Estimating 6D poses from partial references, which capture only fragments of an object's appearance and geometry, remains challenging. To address this, we propose UA-Pose, an uncertainty-aware approach for 6D object pose estimation and online object completion specifically designed for partial references. We assume access to either (1) a limited set of RGBD images with known poses or (2) a single 2D image. For the first case, we initialize a partial object 3D model based on the provided images and poses, while for the second, we use image-to-3D techniques to generate an initial object 3D model. Our method integrates uncertainty into the incomplete 3D model, distinguishing between seen and unseen regions. This uncertainty enables confidence assessment in pose estimation and guides an uncertainty-aware sampling strategy for online object completion, enhancing robustness in pose estimation accuracy and improving object completeness. We evaluate our method on the YCB-Video, YCBIInEOAT, and HO3D datasets, including RGBD sequences of YCB objects manipulated by robots and human hands. Experimental results demonstrate significant performance improvements over existing methods, particularly when object observations are incomplete or partially captured. Project page: https://minfenli.github.io/UA-Pose/", + "bbox": [ + 75, + 359, + 473, + 784 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 814, + 209, + 829 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "6D object pose estimation, which determines the rigid 6-degree-of-freedom transformation between an object and a camera, is essential for various real-world applications such as robotic manipulation [17, 38, 39, 49] and augmented re", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/37db0d36fb83e091b613238948cfc2a2ea7b152f0be3e6b070408c9a83247ef3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 328, + 864, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a205788b8f026eca1d12a5b0f168034c341c780a54780064154ebaa72266780b.jpg", + "image_caption": [ + "Figure 1. Overview of UA-Pose. (a) We initialize an incomplete object 3D model $\\mathcal{M}$ (Sec. 3.2), which is a hybrid mesh that combines texture, geometry, and uncertainty. (b) Given a sequence of RGBD test images, our method estimates the 6D pose of the object in each test image (Sec. 3.3) based on the incomplete model $\\mathcal{M}$ . (c) While more test images are captured, we store them with estimated poses in a memory pool, which will be used for online object completion (Sec. 3.4) to refine the incomplete model $\\mathcal{M}$ ." + ], + "image_footnote": [], + "bbox": [ + 527, + 479, + 859, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ality [22, 29]. Recent research introduced model-based and model-free approaches that can generalize to arbitrary novel objects. These methods push the boundaries of pose estimation for real-world applications and can be divided into two scenarios based on the object information available beforehand. Model-based methods [1, 2, 4, 11, 13, 16, 23, 24, 28] require textured 3D CAD models of the object to estimate 6D object poses in 2D images by establishing 2D-3D correspondences. In contrast, model-free methods leverage a large set of posed reference RGB images [19], RGBD images [10, 41], or video sequences [7, 31] to estimate object poses without the need for CAD models.", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1180", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To bridge model-based and model-free methods, FoundationPose [41] proposes a unified framework that supports both model-based and model-free setups, using 3D CAD models or posed RGBD images (referred to as references). While effective in generalizing to novel objects, FoundationPose and similar model-free methods [7, 10, 31] still depend on high-quality 3D models or sufficient posed RGBD images that capture the object from sufficient viewpoints. However, assuming the availability of high-quality 3D models is often impractical due to the vast number of objects that remain uncaptured or unmodeled in real-world environments. Likewise, obtaining adequate posed RGBD images is challenging, as real-world settings frequently involve limited viewpoints, occlusions from nearby objects, and complexities in estimating camera poses in dynamic scenes.", + "bbox": [ + 75, + 90, + 472, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we present UA-Pose, an uncertainty-aware approach for 6D pose estimation and online object completion as shown in Fig. 1. Our approach is designed to work with partial reference inputs that can include valuable meta-information for robotic tasks, such as grasping patterns, detection markers, or affordance labels. We consider two scenarios: (1) When a limited set of reference RGBD images with known poses is available, we initialize an incomplete 3D model using these images and poses; (2) When only a single unposed RGB image is available, we apply image-to-3D techniques to generate an initial object 3D model for pose estimation. By marking unseen regions as uncertain, we introduce a hybrid representation (Sec. 3.2) that integrates texture, geometry, and uncertainty. The uncertainty is utilized during our proposed uncertainty-aware pose estimation (Sec. 3.3) to provide confidence for estimated poses. Additionally, we propose an uncertainty-aware sampling strategy to select informative images for online object completion, iteratively refining the object 3D model during testing (Sec. 3.4).", + "bbox": [ + 75, + 320, + 472, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate our approach on the YCB-Video, YCBI-nEOAT, and HO3D datasets. The results demonstrate significant improvements over existing methods and highlight the applicability of our uncertainty-aware approach in real-world scenarios where fully captured 3D models are often unavailable. To summarize, our contributions are:", + "bbox": [ + 76, + 626, + 468, + 717 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose an uncertainty-aware approach for 6D object pose estimation (Sec. 3.3) and online object completion (Sec. 3.4), which improves both pose accuracy and object completeness when only partial references are available.", + "- We develop a hybrid representation (Sec. 3.2) incorporating uncertainty. The uncertainty enables confidence assessment in pose estimation and supports an uncertainty-aware sampling strategy for online object completion.", + "- We demonstrate pose estimation from a single RGB image by leveraging single-image-to-3D methods, which generate an initial model for pose estimation and provide augmented data for online object completion (Sec. 3.5)." + ], + "bbox": [ + 76, + 719, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Our method supports different partial object references as input, which could be associated to meta-information, making it feasible for real-world applications.", + "bbox": [ + 500, + 90, + 890, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 151, + 640, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Model-based Object Pose Estimation", + "text_level": 1, + "bbox": [ + 500, + 176, + 818, + 193 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional model-based pose estimation methods can be divided into two categories: instance-level [8, 9, 12, 25] and category-level [3, 15, 32, 34, 46, 47]. Both types of methods assume that a textured CAD model of the specific object or similar object categories is available during training. However, these approaches are limited to objects seen during training, significantly restricting their applicability in real-world scenarios where many objects remain unseen or unmodeled. To address this limitation, zero-shot methods have been proposed [2, 11, 13, 16, 28], which estimate the pose of novel objects by providing their CAD models at test time. These methods use object CAD models to either render 2D images for feature matching and 2D-3D correspondence estimation or leverage point clouds from CAD models to estimate 3D-3D correspondences. While these approaches improve generalization to unseen objects, they still require access to high-quality 3D models at test time, which is often impractical in real-world settings.", + "bbox": [ + 496, + 200, + 893, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Model-free Object Pose estimation", + "text_level": 1, + "bbox": [ + 500, + 484, + 800, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model-free methods eliminate the need for an explicit textured CAD model. Gen6D [19] introduces a pipeline to estimate object poses from RGB images with known poses. However, due to limited training data, Gen6D faces challenges with out-of-distribution test cases and images affected by occlusions. OnePose [31] and its extension OnePose++ [7] use structure-from-motion techniques to reconstruct 3D point clouds from sequential RGB images and employ pre-trained 2D-3D matching networks to estimate the pose of test images. FS6D [10] employs a transformer-based architecture to extract features from multi-view RGBD images for pose estimation. More recently, FoundationPose [41], trained on a large synthetic dataset, presents a unified framework that supports both model-based and model-free setups. In its model-free setting, FoundationPose requires posed reference images that capture the object from sufficient viewpoints to reconstruct 3D models for robust pose estimation. In general, these model-free pose estimation methods rely on either sequential RGB images with known poses or densely posed RGBD images. However, assuming pre-captured posed RGBD images for unseen objects is impractical for two key reasons: first, real-world applications often involve numerous unseen objects, making it infeasible to pre-capture and collect images for each object; and second, even if sufficient images of the novel object are available, accurately estimating cam", + "bbox": [ + 496, + 508, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1181", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "era poses in dynamic environments is challenging.", + "bbox": [ + 76, + 90, + 410, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Object Pose Tracking and Reconstruction", + "text_level": 1, + "bbox": [ + 76, + 114, + 431, + 130 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Assuming test images for pose estimation are sequential, 6D object pose tracking methods [36, 40] leverage temporal cues to achieve efficient, smooth, and accurate pose predictions across consecutive frames by estimating relative pose transformations from each frame to the initial frame. For instance, BundleSDF [40] is a 6D pose tracking method that estimates poses from sequential RGBD images while simultaneously training a signed distance field (SDF) [45] representation to maintain consistent pose estimates over time. However, pose tracking methods are constrained to tracking object poses within sequential images from a single video, without support for using reference images for initial pose estimation. In contrast, our method is designed to estimate object poses based on given object references, which may include metadata beneficial for robotic applications, such as grasping patterns, barcodes, or affordance labels.", + "bbox": [ + 75, + 137, + 472, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Single-Image-to-3D and Pose Estimation", + "text_level": 1, + "bbox": [ + 76, + 387, + 423, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, single-image-to-3D methods have been developed to generate 3D models from a single image [18, 20, 21, 26, 27, 33, 43]. These methods offer potential for pose estimation when textured CAD models are unavailable. For instance, GigaPose [23] explored using generated models for pose estimation; however, relying on generated 3D models that are not closely similar to real objects can lead to imprecise pose estimates. To address this, we apply a single-image-to-3D approach [43] to synthesize a generated model for early pose estimation. As real RGBD images are gradually collected during testing, they are used to create a refined object model, while the generated model from single-image-to-3D methods is leveraged to produce augmented data to improve object geometry in the uncaptured parts. Incorporating these augmented data into the SDF training yields more complete object models, ultimately improving the robustness and accuracy of pose estimation.", + "bbox": [ + 75, + 409, + 472, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 679, + 168, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An overview of our method is illustrated in Fig. 1. We begin by initializing an incomplete object 3D model $\\mathcal{M}$ using partial references, which can be either: (1) a few reference RGBD images with known poses or (2) a single unposed RGB image of the object. In this work, we propose a hybrid object representation (Sec. 3.2) that integrates texture, geometry, and uncertainty, serving as the incomplete model $\\mathcal{M}$ for uncertainty-aware pose estimation (Sec. 3.3).", + "bbox": [ + 75, + 703, + 468, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a sequence of $k$ RGB test images, $\\mathcal{I} = \\{I_0,I_1,\\dots ,I_{k - 1}\\}$ , corresponding depth maps, and an initial segmentation mask $m_0$ for the object of interest in the first image $I_0$ , our method estimates the 6D pose of the object $\\xi_{i}$ in each subsequent test image $I_{i}$ using", + "bbox": [ + 76, + 825, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the proposed uncertainty-aware pose estimation approach (Sec. 3.3), based on the incomplete 3D model $\\mathcal{M}$ . During testing, newly captured test RGBD images and their estimated poses will be appended to a memory pool. These images will be used for online object completion (Sec. 3.4) to refine the incomplete object 3D model $\\mathcal{M}$ . Notably, when the input reference is a single unposed RGB image, we employ a single-image-to-3D approach [43] to initialize a generated model $\\hat{\\mathcal{M}}$ to assist in pose estimation and object completion. More details are described in Sec. 3.5.", + "bbox": [ + 496, + 90, + 890, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Background: Model-free Pose Estimation", + "text_level": 1, + "bbox": [ + 498, + 253, + 854, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given RGBD reference images of a novel object and their camera poses, the task of model-free pose estimation is to determine the 6D object poses (i.e., the rigid 6D transformation from the object to the camera) of that novel object in test RGBD images. FoundationPose [41] is a foundational model for 6D pose estimation trained on a large-scale dataset. In the model-free setup, it uses sufficient RGBD reference images (e.g., 16 RGBD images) that fully cover the object to train a Signed Distance Field (SDF) and then extracts a mesh by marching cubes as the 3D object model. Given the 3D object model to render object images, it introduces a two-stage pipeline with two specialized networks: a pose refinement network, which refines pose hypotheses by estimating the relative pose between the test image and rendered object images, and a pose selection network, which identifies the best pose by comparing the test image with rendered images based on these refined poses. However, RGBD reference images that fully represent the object's appearance and geometry are usually unavailable in real-world scenarios. As a result, model-free methods may struggle to provide accurate pose estimates when only incomplete object models or sparse observations are available.", + "bbox": [ + 496, + 276, + 890, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Hybrid Object Representation", + "text_level": 1, + "bbox": [ + 498, + 619, + 769, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the limitations, we introduce a hybrid object representation that integrates the object's texture, geometry, and uncertainty, as shown in Fig. 2. This uncertainty marks the seen and unseen regions of an incomplete 3D object model and provides confidence in pose estimations, helping to filter out less reliable results. Given a set of RGBD images, object masks, and their camera poses, we train an object-centric neural Signed Distance Field (SDF) [45] to learn the partial object's appearance and geometry.", + "bbox": [ + 496, + 643, + 890, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We then extract a 3D mesh using the marching cubes algorithm to explicitly model uncertainty. Inspired by Visual Hull [14], we define object uncertainty based on visibility inferred from 2D object masks. By projecting these masks onto the mesh $\\mathcal{M}$ , we label regions of $\\mathcal{M}$ seen in the reference images as \"certain\" and those unseen as \"uncertain\", as shown in Fig. 2. This approach allows pose estimation to account for both the seen and unseen parts of the object.", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1182", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8f86fd6dc3858ab38ac2f5fa4a989cda6c02879314baf8645078b91afd2704f4.jpg", + "image_caption": [ + "Figure 2. Hybrid Object Representation Modeling. We propose a hybrid object representation (Sec. 3.2) that integrates the object's texture, geometry, and uncertainty. (a) First, a neural SDF is trained and extracted as a mesh representing the object's appearance and geometry (Sec. 3.2.1). (b) Then, we check the visibility of each mesh vertex from the viewpoint of each reference image to create the uncertainty map (Sec. 3.2.2), which reflects the seen and unseen regions of an incomplete 3D object model." + ], + "image_footnote": [], + "bbox": [ + 81, + 94, + 462, + 200 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Neural SDF Training and Mesh Extraction", + "text_level": 1, + "bbox": [ + 76, + 321, + 429, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We follow SDF training methods similar to [40, 41] to train a neural SDF for object modeling. The SDF defines the object surface as the set of 3D points:", + "bbox": [ + 75, + 344, + 468, + 390 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS = \\{x \\in \\mathbb {R} ^ {3} \\mid \\Omega (x) = 0 \\}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 397, + 468, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Omega : \\mathbb{R}^3 \\to \\mathbb{R}$ is the SDF geometry function that maps each 3D point $x$ to a signed distance value $s$ and $s = 0$ indicates the object's surface. Given RGBD images, object masks, and their camera poses, we use a neural SDF to reconstruct the object's appearance and geometry, represented by two networks [45] as shown in Fig. 2. The geometric network $\\Omega : x \\mapsto s$ takes a 3D point $x \\in \\mathbb{R}^3$ as input and outputs a signed distance value $s \\in \\mathbb{R}$ . The appearance network $\\Phi : (f_{\\Omega(x)}, d) \\mapsto c$ takes an intermediate feature vector $f_{\\Omega(x)}$ from the geometric network along with a view direction $d \\in \\mathbb{R}^3$ and outputs the color $c \\in \\mathbb{R}^3$ for the point.", + "bbox": [ + 75, + 421, + 468, + 587 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "**Rendering.** Given the object pose $\\xi$ , we render an image by casting rays through each pixel. Along each ray $r$ , 3D points are sampled at multiple positions as follows:", + "bbox": [ + 75, + 588, + 468, + 633 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {i} (r) = o (r) + t _ {i} d (r), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 641, + 468, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $o(r)$ denotes the ray origin and $d(r)$ denotes the ray direction from the camera to the object determined by the object pose $\\xi$ . The parameter $t_i \\in \\mathbb{R}_+$ determines the point position along the ray. Following the approach in [40], the color $c$ of a ray $r$ is computed by integrating sampled points near the object surface through volumetric rendering:", + "bbox": [ + 75, + 664, + 468, + 755 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nc (r) = \\int_ {z (r) - \\lambda} ^ {z (r) + 0. 5 \\lambda} w \\left(x _ {i}\\right) \\Phi \\left(f _ {\\Omega \\left(x _ {i}\\right)}, d \\left(x _ {i}\\right)\\right) d t, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 762, + 468, + 800 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw \\left(x _ {i}\\right) = \\frac {1}{1 + e ^ {- \\alpha \\Omega \\left(x _ {i}\\right)}} \\frac {1}{1 + e ^ {\\alpha \\Omega \\left(x _ {i}\\right)}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 806, + 468, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w(x_{i})$ is a bell-shaped probability density function [35] based on the distance of each point from the implicit object surface and $\\alpha$ adjusts the sharpness of the density distribution. $z(r)$ represents the ray's depth from the", + "bbox": [ + 75, + 840, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "depth image and $\\lambda$ is the truncation distance. During volumetric rendering, $z(r)$ and $\\lambda$ are used to exclude 3D points far from the object surface, enhancing rendering efficiency and preventing the SDF from overfitting to empty regions.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During training, we supervise this rendering by comparing it to the reference RGB images with the color loss:", + "bbox": [ + 496, + 152, + 890, + 181 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = \\frac {1}{| \\mathcal {R} |} \\sum_ {r \\in \\mathcal {R}} \\| c (r) - \\bar {c} (r) \\| _ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 191, + 890, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\bar{c}(r)$ is the ground-truth color at the pixel through which the ray $r$ passes. For geometry learning, we leverage depth information from the provided depth maps to supervise the neural (SDF) through two losses [40]: the near-surface loss $\\mathcal{L}_s$ and the empty space loss $\\mathcal{L}_e$ . $\\mathcal{L}_s$ encourage the SDF surface to align closely with depth values and $\\mathcal{L}_e$ enforces that points along the ray between the camera and the surface remain empty. Additionally, we apply an eikonal regularization loss $\\mathcal{L}_{eik}$ [5] to the near-surface SDF to enforce smoothness. The total loss is represented as:", + "bbox": [ + 496, + 241, + 892, + 392 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = w _ {c} \\mathcal {L} _ {c} + w _ {e} \\mathcal {L} _ {e} + w _ {s} \\mathcal {L} _ {s} + w _ {e i k} \\mathcal {L} _ {e i k}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 406, + 890, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w_{c}, w_{e}, w_{s}$ , and $w_{eik}$ are weighting factors.", + "bbox": [ + 496, + 435, + 831, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To explicitly model uncertainty, we convert the implicit surface defined by the neural SDF into an explicit 3D mesh $\\mathcal{E} = (V,C,F)$ with the vertices $V$ , vertex colors $C$ , and faces $F$ . To achieve this, we extract a mesh from the trained neural SDF representation using the marching cubes algorithm, which guarantees a closed-surface mesh including vertices from any viewing angle. This allows us to model the visibility of the object from any viewing angle even if some parts are not visible in the reference images.", + "bbox": [ + 496, + 450, + 890, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each vertex $v_{i} \\in V$ is assigned a color $c_{i} \\in C$ , which is computed using the appearance function $\\varPhi$ , which maps the intermediate feature vector $f_{\\Omega(v_{i})}$ from the geometry network and a view direction $d \\in \\mathbb{R}^3$ to the RGB color:", + "bbox": [ + 496, + 585, + 890, + 646 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nc _ {i} = \\Phi \\left(f _ {\\Omega \\left(v _ {i}\\right)}, d\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 659, + 890, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Uncertainty Modeling", + "text_level": 1, + "bbox": [ + 500, + 694, + 707, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To model the uncertainty map that represents whether a vertex is visible from the viewpoints of input references, for each vertex $v_{i} \\in V$ , we check its visibility from the viewpoint of each reference image $\\bar{I}_j \\in \\tilde{\\mathcal{I}}$ through the mesh rasterization process, which rasterizes 3D vertices onto 2D image pixels. We define an uncertainty score $u(v_{i})$ based on whether $v_{i}$ is visible from any of the viewpoints in the reference images $\\bar{I}_j$ . This uncertainty score $u(v_{i}) \\in \\{0,1\\}$ labels vertices on the mesh as \"certain\" if they are seen in any reference image $(u(v_{i}) = 0)$ and as \"uncertain\" if they remain unseen $(u(v_{i}) = 1)$ . The resulting uncertainty map $\\mathcal{U} = \\{u(v_i) \\mid v_i \\in V\\}$ is incorporated into our hybrid object", + "bbox": [ + 496, + 719, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1183", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4ef30e617b60be67a5e5b93f5e2b6e395d349da375deb031384565aac2fc26c9.jpg", + "image_caption": [ + "Figure 3. Pipeline for Uncertainty-Aware Pose Estimation. Through uncertainty modeling (Sec. 3.2.2), we propose an uncertainty-aware pipeline to assess the confidence of estimated poses (Sec. 3.3). For each estimated pose, we calculate its seen IoU, which measures the overlap between the seen regions of the 3D model and the 2D object mask for the test image. When seen IoU is high, the pose is deemed confident, and both the pose and its RGBD image are added to the memory pool for object completion. Conversely, if the estimated pose has a low seen IoU, it is considered unreliable. In such cases, we employ an uncertainty-aware image sampling strategy to select images from the memory pool for online object completion (Sec. 3.4), thus enhancing pose estimation by refining the object model during testing." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 885, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "representation $\\mathcal{M} = (\\mathcal{E},\\mathcal{U})$ , allowing the pose estimation process to distinguish between seen and unseen regions.", + "bbox": [ + 75, + 314, + 467, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Guidance from Uncertainty Map. Given an object pose $\\xi \\in \\mathbb{R}^{3\\times 4}$ , the model $\\mathcal{M}$ can be used to render an RGB image $I_{\\xi}^{\\mathrm{rend}}\\in \\mathbb{R}^{H\\times W\\times 3}$ , a depth map $D_{\\xi}^{\\mathrm{rend}}\\in \\mathbb{R}^{H\\times W}$ , a rendered object mask $m_{\\xi}^{\\mathrm{rend}}\\in \\{0,1\\}^{H\\times W}$ , and an uncertainty map $U_{\\xi}^{\\mathrm{rend}}\\in \\{0,1\\}^{H\\times W}$ at a resolution of $H\\times W$ . To evaluate the validity and confidence of an estimated pose, we introduce two metrics: uncertainty rate and seen IoU. The uncertainty rate measures the proportion of the uncertain pixels over the total pixels within the rendered object mask:", + "bbox": [ + 75, + 345, + 468, + 484 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\sum \\left(U _ {\\xi} ^ {\\text {r e n d}} \\odot m _ {\\xi} ^ {\\text {r e n d}}\\right)}{\\sum \\left(m _ {\\xi} ^ {\\text {r e n d}}\\right)}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 494, + 468, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\odot$ denotes element-wise logical multiplication. The seen IoU measures the overlap between seen (certain) regions of the rendered object mask and the object mask in the test image, which is defined as:", + "bbox": [ + 75, + 537, + 468, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {I o U} \\left(\\neg U _ {\\xi} ^ {\\text {r e n d}} \\odot m _ {\\xi} ^ {\\text {r e n d}}, m ^ {\\text {t e s t}}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 608, + 468, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $m^{\\mathrm{test}}$ is the object mask from the test image, $\\neg U_{\\xi}^{\\mathrm{rend}}\\odot$ $m_{\\xi}^{\\mathrm{rend}}$ identifies \"certain\" pixels within the rendered object mask, and IoU represents the Intersection over Union function. Since pose estimation is more robust when based on seen regions of the 3D model, a high seen IoU indicates high confidence in the pose estimation, whereas a low seen IoU suggests that the estimation is not reliable.", + "bbox": [ + 75, + 638, + 468, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Uncentainty-aware Pose Estimation", + "text_level": 1, + "bbox": [ + 76, + 756, + 387, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After modeling the uncertainty, the hybrid object representation could be used in our uncertainty-aware pose estimation pipeline, as shown in Fig. 3. In this work, we leverage the pose refinement and the pose selection modules of [41] and assume that the test images are sequential. To obtain per-frame 2D object masks in test images for calculating the uncertainty rate and seen IoU, given first frame object mask $m_0$ , the object masks of the following", + "bbox": [ + 75, + 780, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "frames $\\{m_1, m_2, \\ldots, m_{k-1}\\}$ is determined by off-the-shelf segmentation methods [44].", + "bbox": [ + 498, + 314, + 890, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pose Initialization. Given an RGBD video, object masks for each frame, and an incomplete object model $\\mathcal{M}$ , we begin by estimating the object pose for the first frame by generating multiple pose hypotheses. For each hypothesis, the translation is initialized using the 3D point projected from the center of the 2D object mask, determined by the median depth value within this region. For rotation, we uniformly sample $N_{v}$ viewpoints from an icosphere centered on the object and orient the camera to face the object's center. Subsequently, $N_{in}$ in-plane rotations are sampled and applied to each viewpoint, yielding $N_{v} \\cdot N_{in}$ hypothesized object poses. Each object pose can be represented as $[\\pmb{R}|\\pmb{t}] \\in \\mathbb{S}\\mathbb{E}(3)$ , where $\\pmb{R} \\in \\mathbb{S}\\mathbb{O}(3)$ is the rotation and $\\pmb{t} \\in \\mathbb{R}^3$ is the translation.", + "bbox": [ + 496, + 348, + 892, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pose Refinement. After generating $N_v \\cdot N_{in}$ pose hypotheses, we apply the refinement module [41] to improve pose accuracy. Specifically, given a rendered RGBD image of the object from each pose hypothesis and the test RGBD image of the target object, the network outputs a pose update $\\Delta R \\in \\mathbb{SO}(3)$ and $\\Delta t \\in \\mathbb{R}^3$ to refine the pose hypothesis, aligning it more closely with the observed object pose in the test image. Each refined pose is represented as $[R^{+}|t^{+}] \\in \\mathbb{SE}(3)$ , where $t^{+} = t + \\Delta t$ and $R^{+} = \\Delta R \\otimes R$ . This refinement process could be iteratively repeated. In this work, each pose is refined five times for pose estimation on the first frame and twice for subsequent frames.", + "bbox": [ + 496, + 564, + 893, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pose Selection. Given a set of refined pose hypotheses, we apply our proposed metrics (Sec. 3.2.2) to filter out unreliable poses: the uncertainty rate, which measures the proportion of uncertain pixels within the rendered object mask, and the seen IoU, which quantifies the overlap between previously seen (certain) regions of the object and the object mask in the test image. A pose hypothesis is discarded when its uncertainty rate is above the threshold $T_{u}$ or its seen IoU is below the threshold $T_{s}$ , ensuring that only reliable hypotheses that adequately cover seen object regions are con-", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1184", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sidered. At last, we use the pose selection module [41] to identify the most accurate pose as the final estimate.", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pose Tracking. Given the sequential nature of RGBD video frames, object poses in adjacent frames are typically similar. For each frame following the initial one, we take the estimated pose from the previous frame as the sole pose hypothesis and refine it using the pose refinement module to generate the estimated pose for the current frame. This iterative process continues across the video sequence, leveraging temporal cues for stable and smooth pose estimation.", + "bbox": [ + 75, + 121, + 467, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Memory Pool. To enhance pose estimation by completing the object model during testing, we select test RGBD images with their estimated poses to incrementally refine the object representation. For efficiency, we introduce a memory pool $\\mathcal{P} = \\{\\xi_0,\\xi_1,\\xi_2,\\dots ,\\xi_{|\\mathcal{P}| - 1}\\}$ that stores the most informative $|\\mathcal{P}|$ viewpoints to maintain a compact yet diverse set of test image frames for object completion, where $\\xi_{i}$ is the estimated pose corresponding to a test image $I_{i}$ . The first frame $\\xi_{i}$ is automatically added to $\\mathcal{P}$ , establishing the canonical coordinate system for the novel object. Subsequent frames are added when their viewpoints contribute to the multi-view diversity in the pool. Specifically, before adding a new frame, we calculate the rotational geodesic distance between the new frame and the last frame added to the pool. A frame is included in $\\mathcal{P}$ only if this geodesic distance exceeds a threshold $T_{geo}$ , ensuring that each frame offers a unique viewpoint to enhance model completeness while keeping $\\mathcal{P}$ compact. Note that not all estimated poses are reliable, adding inaccurate results to the memory pool may cause noisy SDF training during object completion. To mitigate this, we introduce an image-filtering strategy to filter out low-confidence estimations. As illustrated in Fig. 3, when a pose is estimated, its confidence is evaluated using our proposed metric, seen IoU, which measures the overlap between the seen regions of the 3D model and the 2D object mask in the test image. Since a high seen IoU indicates greater confidence in an estimated pose, frames are added to the memory pool only if their seen IoU exceeds a threshold $T_{conf}$ , ensuring that low-confidence poses are excluded.", + "bbox": [ + 75, + 244, + 470, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Uncentainty-aware Object Completion", + "text_level": 1, + "bbox": [ + 76, + 694, + 410, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig. 3, online object completion is triggered when the seen IoU of the current test frame falls below a threshold $T_{complete}$ , indicating that the current pose estimation may no longer be reliable. As new frames are continuously added, the memory pool $\\mathcal{P}$ can grow excessively large, which may reduce SDF training efficiency. To manage this, we perform uncertainty-aware sampling to select the $K$ most informative frames for object completion.", + "bbox": [ + 75, + 718, + 467, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Uncertainty-aware Sampling. Early in the video sequence, when $|\\mathcal{P}| \\leq K$ , all frames in the pool are included without selection. However, once the pool size exceeds $K$ we apply an uncertainty-aware image selection strategy to", + "bbox": [ + 75, + 839, + 467, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5f5e32d554693063e3a24cf45252f3a8a34cb877e17513714866070a0db819e4.jpg", + "image_caption": [ + "Figure 4. Pipeline for leveraging generated models. In Sec. 3.5, the initial model generated by image-to-3D techniques is used to estimate poses for the first few frames. As more frames are captured, a refined object model is trained to replace the initial generated model, resulting in a more complete representation. To further leverage the generated 3D model, we render RGBD images from the image-to-3D generated model as augmented data for SDF training. This supervision enhances the object geometry in unseen regions, providing more complete information for pose estimation." + ], + "image_footnote": [], + "bbox": [ + 506, + 92, + 781, + 243 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8d659f63bc6c501eddc73a3d3b7a0a4baf595c4464b985a4a8ad6214c95aea94.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 795, + 92, + 880, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "maximize the coverage of unseen object regions. This process selects a subset, $\\mathcal{P}' \\subset \\mathcal{P}$ , that reveals the most unseen areas of the object. The selection process begins by including the initial and most recent frames in the pool to establish the starting and ending viewpoints. Next, an iterative strategy is employed: at each iteration, we select the pose that reveals the largest unseen region. This continues until $K$ poses are selected.", + "bbox": [ + 498, + 380, + 890, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Leveraging Image-to-3D Generation", + "text_level": 1, + "bbox": [ + 500, + 513, + 815, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given a single unposed RGB image, we employ image-to-3D techniques to generate an initial object 3D model $\\hat{\\mathcal{M}}$ for pose estimation. The viewpoint associated with the initial RGB image is labeled as \"certain,\" while other areas that are inferred by the image-to-3D method are marked \"uncertain\". The generated model is then rescaled to align with the object size in the test images. Specifically, we use the depth map and object mask from the first frame of the test image to approximate a rough model size. We then sample slightly larger and smaller models, generate pose hypotheses, and employ the pose selection module [41] to select the optimal pose hypothesis and corresponding model size.", + "bbox": [ + 498, + 537, + 890, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Refer to Fig. 4, the initial generated model is used to estimate poses for the first few test frames until the rotational geodesic distance between the initial pose and a new estimated pose exceeds a threshold $T_{gen}$ . As more frames are captured, a refined object model is trained to replace the initial generated model, resulting in a more complete representation. To further leverage the generated 3D model, we render RGBD images from the image-to-3D generated model as augmented data for SDF training. This supervision enhances the object geometry in unseen regions, providing more complete information for pose estimation. See the supplementary materials for additional details.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1185", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 89, + 200, + 107 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 114, + 179, + 128 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "YCB-Video [42]: The YCB-Video dataset contains RGBD video sequences of everyday objects from the YCB objects, widely used for benchmarking 6D object pose estimation.", + "bbox": [ + 75, + 137, + 467, + 183 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "YCBCInEOAT [37]: YCBCInEOAT extends the YCB-Video dataset with RGBD sequences of YCB objects manipulated by a dual-arm robot. This dataset focuses on real-world scenarios involving robot-object interactions.", + "bbox": [ + 75, + 183, + 467, + 243 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "HO3D [6]: The HO3D (Hand-Object 3D) dataset contains RGBD sequences of human hands interacting with objects, capturing complex hand-object interactions and occlusions.", + "bbox": [ + 75, + 243, + 467, + 289 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Experimental setup", + "text_level": 1, + "bbox": [ + 76, + 299, + 264, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We evaluate our method against state-of-the-art RGBD model-free 6D pose estimation methods [10, 30, 41] under two scenarios where partial object references are given as (1) A limited set of reference RGBD images (2 views) that include known poses. (2) A single unposed RGB object image. In scenario (2), we leverage image-to-3D techniques [43] to generate an initial object 3D model and iteratively estimate the object pose during testing.", + "bbox": [ + 75, + 321, + 467, + 443 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics. We consider ADD and ADD-S metrics for 6D pose estimation, which measure the accuracy of estimated poses. The Area Under the Curve (AUC) is calculated and reported for both the ADD and ADD-S metrics, following the protocols in [40-42]. Furthermore, to demonstrate the effectiveness of our method for online object completion, we report the Chamfer Distance (CD), which measures the average distance between points of the reconstructed model and the ground-truth model.", + "bbox": [ + 75, + 443, + 467, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison Results on YCB-Video", + "text_level": 1, + "bbox": [ + 76, + 589, + 383, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first evaluate UA-Pose on the YCB-Video [42] dataset, a widely used benchmark for 6D pose estimation, and compare our approach with state-of-the-art model-free methods [10, 30, 41] with a limited set of reference images.", + "bbox": [ + 75, + 613, + 467, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As presented in Tab. 1, our method significantly outperforms FoundationPose [41] with only 2 reference images in all metrics. Moreover, both our method and FoundationPose outperforms LoFTR [30] and FS6D-DPM [10] with 16 reference images with a large margin. This shows that FoundationPose is the strongest baseline for our further analysis. When only a single unposed RGB image of the object is available, we apply a single-image-to-3D approach [43] to generate an initial 3D model for pose estimation. Experimental results, shown in Tab. 1, indicate that directly using the generated model with FoundationPose may not yield optimal results, as these models may not accurately represent real objects. In contrast, UA-Pose uses the generated model only for initial pose estimation and for rendering RGBD images as augmented data to aid object", + "bbox": [ + 75, + 674, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8d81e4efa51b9c976587e3f90356bf82c2673c9348d731d79ed9123193866ab4.jpg", + "table_caption": [ + "Table 1. Quantitative results on the YCB-Video dataset. \"Comp.\" indicates whether the method includes online completion, and \"-\" denotes that the method does not reconstruct object shapes. Results for [10, 30] are adopted from [41]. Experiments for \"single RGB + Image-to-3D\" are conducted on a subset of the YCB-Video dataset. See supplementary materials for details." + ], + "table_footnote": [], + "table_body": "
InputComp.Mean
ADDADD-SCD
LoFTR [30]16 referencesNo26.252.5-
FS6D-DPM [10]16 referencesNo42.188.4-
FoundationPose [41]2 referencesNo87.494.30.57
Ours2 referencesYes92.896.50.53
FoundationPose [41]single RGB + Image-to-3DNo88.993.60.67
Ourssingle RGB + Image-to-3DYes93.296.90.62
", + "bbox": [ + 503, + 172, + 890, + 258 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f843c93225d3329a87054396655807da502fe3fb685821cf6f33a15e1539db30.jpg", + "table_caption": [ + "Table 2. Quantitative results on the YCBIInEOAT dataset. \"Comp.\" indicates that the method performs online completion." + ], + "table_footnote": [], + "table_body": "
InputComp.Mean
ADDADD-SCD
FoundationPose [41]2 referencesNo68.5284.800.60
Ours2 referencesYes89.9994.350.57
FoundationPose [41]single RGB + Image-to-3DNo83.3990.660.61
Ourssingle RGB + Image-to-3DYes89.2493.990.60
", + "bbox": [ + 503, + 294, + 890, + 358 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "completion, as described in Sec. 3.5. This approach reconstructs a model more closely aligned with the real object, leading to improved pose estimation accuracy.", + "bbox": [ + 498, + 375, + 890, + 420 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Comparison Results on YCBIInEOAT", + "text_level": 1, + "bbox": [ + 498, + 431, + 821, + 448 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluated UA-Pose on the YCBIInEOAT [37] dataset to assess its performance in challenging robotic scenarios in real-world where objects are manipulated by robotic arms.", + "bbox": [ + 496, + 455, + 890, + 501 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 2, when 2 reference images are provided, UA-Pose demonstrates superior performance compared to FoundationPose, highlighting the effectiveness of our uncertainty-aware mechanisms and online object completion for robust pose estimation in robotic scenarios. As illustrated in Fig. 5, simply using the incomplete model generated from the partial references often leads to inaccurate estimations. In contrast, our method achieves robust pose estimation with uncertainty-aware mechanisms and online object completion. When only a single unposed RGB image of the object is available, UA-Pose also consistently surpasses FoundationPose across all metrics. This demonstrates the flexibility of UA-Pose to utilize diverse partial references, which could carry useful meta-information.", + "bbox": [ + 496, + 501, + 890, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Comparison Results on HO3D", + "text_level": 1, + "bbox": [ + 498, + 724, + 769, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluated UA-Pose on the HO3D dataset to assess its effectiveness in complex hand-object interaction scenarios where objects are frequently rotated and occluded.", + "bbox": [ + 496, + 750, + 890, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Tab. 3, when 2 reference images are provided, FoundationPose fails to make accurate estimations when objects in the test video exhibit frequent and strong rotations. In contrast, UA-Pose accurately estimates object poses under the challenging hand-object interaction scenarios. As for utilizing the generated models from the single-image-to-3D approach for pose estimation, UA-Pose", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1186", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3884863dd735546e24505d63eabfee59bbea2529a50d1f2d0d1d39439655d070.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison on the YCBIInEOAT. Object image (left) and depth map (right) pairs are rendered based on the estimated object poses using the incomplete model. Without uncertainty-aware mechanisms and Online Object Completion (referred to as \"w/o OOC\"), the incomplete model often leads to inaccurate estimations. In contrast, our method (referred to as \"w/ OOC\") achieves robust pose estimation by incorporating uncertainty-aware mechanisms and online object completion." + ], + "image_footnote": [], + "bbox": [ + 84, + 84, + 460, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "achieves more accurate estimations than FoundationPose. Note that even if the generated models in Tab. 3 show lower Chamfer Distance (CD) than refined models obtained from object completion, directly using generated models for FoundationPose results in considerably lower ADD and ADD-S. This indicates that even though generated models have good geometry (low CD), geometry alone cannot accurately represent real objects for pose estimation because the textures of generated models are not good enough in unseen areas that are not covered by the initial partial references.", + "bbox": [ + 75, + 354, + 468, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 513, + 228, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Key Components. Tab. 4 presents an ablation study for key components of UA-Pose on the YCBCInEOAT dataset. All experiments start with the initialized model based on two reference images and are evaluated using ADD, ADD-S, and CD. Additionally, we report $N_{rebuild}$ , the total number of times object completion is applied across test sequences to refine the incomplete model $\\mathcal{M}$ . The \"w/o object completion\" setup disables online object completion, illustrating its importance for achieving accurate pose estimation. In \"w/o uncertainty-aware object completion\", object completion is applied every time a new test image is added to the memory pool, instead of completing the model only when the seen IoU threshold is below $T_{complete}$ . This leads to poorer performance and a higher $N_{rebuild}$ count, increasing unnecessary computation for object completion. The \"w/o image-filtering strategy\" experiment omits the image-filtering strategy (Sec. 3.3) for excluding unreliable poses, allowing low-confidence poses to be added to the memory pool and resulting in decreased pose accuracy. In \"w/o uncertainty-aware sampling\", we use the geodesic distance-based sampling strategy from [40] instead of our proposed uncertainty-aware sampling strategy (Sec. 3.4) to sample images for object completion. The results demonstrate our proposed sampling method is better at selecting infor", + "bbox": [ + 75, + 537, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/74715dc9812bc615b1b1f6afff271d5e5ff11af116d5549436f40cd1f82701e2.jpg", + "table_caption": [ + "Table 3. Quantitative results on the HO3D dataset. \"Comp.\" indicates that the method performs online completion." + ], + "table_footnote": [], + "table_body": "
InputComp.Mean
ADDADD-SCD
FoundationPose [41]2 referencesNo45.6761.570.78
Ours2 referencesYes91.5195.560.69
FoundationPose [41]single RGB + Image-to-3DNo72.0687.870.76
Ourssingle RGB + Image-to-3DYes83.2393.590.88
", + "bbox": [ + 503, + 117, + 890, + 181 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7da8fc97de9de723977d9faea0f05d637018848002c53e38c3187257edce8473.jpg", + "table_caption": [ + "Table 4. Ablation study. We demonstrate the importance of UA-Pose's key components on the YCBCInEOAT dataset." + ], + "table_footnote": [], + "table_body": "
MeanNrebuild
ADDADDSCD
w/o object completion68.5284.800.60-
w/o uncertainty-aware object completion85.8392.390.72181
w/o image-filtering strategy87.7093.370.7187
w/o uncertainty-aware sampling87.8893.390.6859
Ours (full)89.9994.350.5758
", + "bbox": [ + 511, + 215, + 875, + 292 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8cd9f33de3907fda7da0499871f699c7674e57cec332f1e036f8788fd5eebfbd.jpg", + "table_caption": [ + "Table 5. Additional study. We compare our method with FoundationPose [41] and pose tracking methods [36, 40] on the YCBI-nEOAT dataset. \"Comp.\" means the method performs online completion, and \"\\*\" denotes using cropped ground-truth meshes [40], which exclude invisible parts in the test sequence, to calculate CD." + ], + "table_footnote": [], + "table_body": "
InputComp.Mean
ADDADD-SCD
BundleTrack [36]first-frame RGBDNo87.3492.532.81*
BundleSDF [40]first-frame RGBDYes86.9593.771.16*
FoundationPose [41]first-frame RGBDNo74.3484.581.12*
Oursfirst-frame RGBDYes88.3893.820.75*
", + "bbox": [ + 511, + 367, + 875, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mative images for accurate pose estimation. Finally, the \"Ours (full)\" setup that incorporates all our key components achieves the best performance across ADD, ADD-S, CD, and the fewest $N_{\\text{rebuild}}$ for object completion.", + "bbox": [ + 496, + 439, + 890, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additional Study. When no external reference images are available, we demonstrate that we could initialize the object model for pose estimation using the first frame of the test RGBD image sequence. Besides the strongest model-free pose estimation baseline [41], we also compare our method with strong model-free pose tracking approaches [36, 40, 48] on the YCBIInEOAT dataset. Refer to Tab. 5, our method achieves the highest performance across ADD, ADD-S, and CD metrics. Note that unlike pose-tracking methods which are limited to estimating relative poses from sequential frames within the same RGBD video, our method can incorporate object reference images as input for useful meta-information.", + "bbox": [ + 496, + 500, + 892, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 709, + 617, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced UA-Pose, an uncertainty-aware approach for 6D object pose estimation that addresses the limitations of previous model-free methods in the scenarios in which input object references are incomplete or partially captured. Incorporating mechanisms to handle uncertainty and leverage online object completion to complete object models, our method demonstrates substantial performance improvements on the YCB-Video, YCBIInEOAT, and HO3D datasets. We demonstrate UA-Pose is a robust solution for practical real-world applications where full 3D object models and extensive image references are usually unavailable.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1187", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Philipp Ausserlechner, David Haberger, Stefan Thalhammer, Jean-Baptiste Weibel, and Markus Vincze. Zs6d: Zero-shot 6d object pose estimation using vision transformers. ICRA, 2024. 1", + "[2] Andrea Caraffa, Davide Boscaini, Amir Hamza, and Fabio Poiesi. Freeze: Training-free zero-shot 6d pose estimation with geometric and vision foundation models. ECCV, 2024. 1, 2", + "[3] Dengsheng Chen, Jun Li, Zheng Wang, and Kai Xu. Learning canonical shape space for category-level 6D object pose and size estimation. In CVPR, 2020. 2", + "[4] Jianqiu Chen, Zikun Zhou, Mingshan Sun, Rui Zhao, Liwei Wu, Tianpeng Bao, and Zhenyu He. Zeropose: Cadprompted zero-shot object 6d pose estimation in cluttered scenes. TCSVT, 2024. 1", + "[5] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, 2020. 4", + "[6] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In CVPR, 2020. 7", + "[7] Xingyi He, Jiaming Sun, Yuang Wang, Di Huang, Hujun Bao, and Xiaowei Zhou. OnePose++: Keypoint-free one-shot object pose estimation without CAD models. NeurIPS, 2022. 1, 2", + "[8] Yisheng He, Wei Sun, Haibin Huang, Jianran Liu, Haoqiang Fan, and Jian Sun. PVN3D: A deep point-wise 3D keypoints voting network for 6DoF pose estimation. In CVPR, 2020. 2", + "[9] Yisheng He, Haibin Huang, Haoqiang Fan, Qifeng Chen, and Jian Sun. FFB6D: A full flow bidirectional fusion network for 6D pose estimation. In CVPR, 2021. 2", + "[10] Yisheng He, Yao Wang, Haoqiang Fan, Jian Sun, and Qifeng Chen. Fs6d: Few-shot 6d pose estimation of novel objects. In CVPR, 2022. 1, 2, 7", + "[11] Junwen Huang, Hao Yu, Kuan-Ting Yu, Nassir Navab, Slobodan Ilic, and Benjamin Busam. Matchu: Matching unseen objects for 6d pose estimation from rgb-d images. In CVPR, 2024. 1, 2", + "[12] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. CosyPose: Consistent multi-view multi-object 6D pose estimation. In ECCV, 2020. 2", + "[13] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. MegaPose: 6D pose estimation of novel objects via render & compare. In CoRL, 2022. 1, 2", + "[14] Aldo Laurentini. The visual hull concept for silhouette-based image understanding. IEEE TPAMI, 1994. 3", + "[15] Taeyeop Lee, Jonathan Tremblay, Valts Blukis, Bowen Wen, Byeong-Uk Lee, Inkyu Shin, Stan Birchfield, In So Kweon, and Kuk-Jin Yoon. TTA-COPE: Test-time adaptation for category-level object pose estimation. In CVPR, 2023. 2", + "[16] Jiehong Lin, Lihua Liu, Dekun Lu, and Kui Jia. Sam-6d: Segment anything model meets zero-shot 6d object pose estimation. In CVPR, 2024. 1, 2" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Jian Liu, Wei Sun, Chongpei Liu, Xing Zhang, and Qiang Fu. Robotic continuous grasping system by shape transformer-guided multi-object category-level 6d pose estimation. IEEE Transactions on Industrial Informatics, 2023. 1", + "[18] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In ICCV, 2023. 3", + "[19] Yuan Liu, Yilin Wen, Sida Peng, Cheng Lin, Xiaoxiao Long, Taku Komura, and Wenping Wang. Gen6D: Generalizable model-free 6-DoF object pose estimation from RGB images. ECCV, 2022. 1, 2", + "[20] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Learning to generate multiview-consistent images from a single-view image. ICLR, 2024. 3", + "[21] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, and Wenping Wang. Wonder3d: Single image to 3d using cross-domain diffusion. CVPR, 2024. 3", + "[22] Eric Marchand, Hideaki Uchiyama, and Fabien Spindler. Pose estimation for augmented reality: A hands-on survey. TVCG, 2015. 1", + "[23] Van Nguyen Nguyen, Thibault Groueix, Mathieu Salzmann, and Vincent Lepetit. Gigapose: Fast and robust novel object pose estimation via one correspondence. CVPR, 2024. 1, 3", + "[24] Evin Pinar Ornek, Yann Labbe, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, and Tomas Hodan. Found- pose: Unseen object pose estimation with foundation features. ECCV, 2024. 1", + "[25] Kiru Park, Timothy Patten, and Markus Vincze. Pix2Pose: Pixel-wise coordinate regression of objects for 6D pose estimation. In ICCV, 2019. 2", + "[26] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, and Bernard Ghanem. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. ICLR, 2024. 3", + "[27] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 3", + "[28] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. OSOP: A multi-stage one shot object pose estimation framework. In CVPR, 2022. 1, 2", + "[29] Yongzhi Su, Jason Rambach, Nareg Minaskan, Paul Lesur, Alain Pagani, and Didier Stricker. Deep multi-state object pose estimation for augmented reality assembly. In ISMAR-Adjunct, 2019. 1", + "[30] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In CVPR, 2021. 7", + "[31] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. OnePose: One-shot object pose estimation without CAD models. In CVPR, 2022. 1, 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1188", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Meng Tian, Marcelo H Ang, and Gim Hee Lee. Shape prior deformation for categorical 6D object pose and size estimation. In ECCV, 2020. 2", + "[33] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In ECCV, 2025. 3", + "[34] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6D object pose and size estimation. In CVPR, 2019. 2", + "[35] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 4", + "[36] Bowen Wen and Kostas Bekris. Bundletrack: 6d pose tracking for novel objects without instance or category-level 3d models. In IROS, 2021. 3, 8", + "[37] Bowen Wen, Chaitanya Mitash, Baozhang Ren, and Kostas E Bekris. se (3)-tracknet: Data-driven 6d pose tracking by calibrating image residuals in synthetic domains. In IROS, 2020. 7", + "[38] Bowen Wen, Wenzhao Lian, Kostas Bekris, and Stefan Schaal. CatGrasp: Learning category-level task-relevant grasping in clutter from simulation. In ICRA, 2022. 1", + "[39] Bowen Wen, Wenzhao Lian, Kostas Bekris, and Stefan Schaal. You only demonstrate once: Category-level manipulation from single visual demonstration. RSS, 2022. 1", + "[40] Bowen Wen, Jonathan Tremblay, Valts Blukis, Stephen Tyree, Thomas Müller, Alex Evans, Dieter Fox, Jan Kautz, and Stan Birchfield. Bundlesdf: Neural 6-dof tracking and 3d reconstruction of unknown objects. In CVPR, 2023. 3, 4, 7, 8", + "[41] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. Foundationpose: Unified 6d pose estimation and tracking of novel objects. CVPR, 2024. 1, 2, 3, 4, 5, 6, 7, 8", + "[42] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. RSS, 2018. 7", + "[43] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 3, 7", + "[44] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 5", + "[45] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. NeurIPS, 2020. 3, 4", + "[46] Ruida Zhang, Yan Di, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. SSP-Pose: Symmetry-aware shape prior deformation for direct category-level object pose estimation. In IROS, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Linfang Zheng, Chen Wang, Yinghan Sun, Esha Dasgupta, Hua Chen, Ales Leonardis, Wei Zhang, and Hyung Jin Chang. HS-Pose: Hybrid scope feature extraction for category-level object pose estimation. In CVPR, 2023. 2", + "[48] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In CVPR, 2022. 8", + "[49] Chungang Zhuang, Shaofei Li, and Han Ding. Instance segmentation based 6d pose estimation of industrial objects using point clouds for robotic bin-picking. Robotics and Computer-Integrated Manufacturing, 2023. 1" + ], + "bbox": [ + 501, + 90, + 890, + 261 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1189", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/UA-Pose_ Uncertainty-Aware 6D Object Pose Estimation and Online Object Completion with Partial References/0d3743ff-3af5-445c-ae4c-ba5a33b9d7b4_model.json b/2025/UA-Pose_ Uncertainty-Aware 6D Object Pose Estimation and Online Object Completion with Partial References/0d3743ff-3af5-445c-ae4c-ba5a33b9d7b4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..09bdd6584450fce816a0782813277cefd72f52dd --- /dev/null +++ b/2025/UA-Pose_ Uncertainty-Aware 6D Object Pose Estimation and Online Object Completion with Partial References/0d3743ff-3af5-445c-ae4c-ba5a33b9d7b4_model.json @@ -0,0 +1,2123 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.105, + 0.13, + 0.868, + 0.177 + ], + "angle": 0, + "content": "UA-Pose: Uncertainty-Aware 6D Object Pose Estimation and Online Object Completion with Partial References" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.203, + 0.761, + 0.24 + ], + "angle": 0, + "content": "Ming-Feng Li\\(^{1}\\) Xin Yang\\(^{4}\\) Fu-En Wang\\(^{4}\\) Hritam Basak\\(^{2}\\) Yuyin Sun\\(^{4}\\) Shreekant Gayaka\\(^{4}\\) Min Sun\\(^{3,4}\\) Cheng-Hao Kuo\\(^{4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.257, + 0.846, + 0.293 + ], + "angle": 0, + "content": "\\(^{1}\\)Carnegie Mellon University \\(^{2}\\)Stony Brook University \\(^{3}\\)National Tsing Hua University \\(^{4}\\)Amazon" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.328, + 0.314, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.361, + 0.474, + 0.785 + ], + "angle": 0, + "content": "6D object pose estimation has shown strong generalizability to novel objects. However, existing methods often require either a complete, well-reconstructed 3D model or numerous reference images that fully cover the object. Estimating 6D poses from partial references, which capture only fragments of an object's appearance and geometry, remains challenging. To address this, we propose UA-Pose, an uncertainty-aware approach for 6D object pose estimation and online object completion specifically designed for partial references. We assume access to either (1) a limited set of RGBD images with known poses or (2) a single 2D image. For the first case, we initialize a partial object 3D model based on the provided images and poses, while for the second, we use image-to-3D techniques to generate an initial object 3D model. Our method integrates uncertainty into the incomplete 3D model, distinguishing between seen and unseen regions. This uncertainty enables confidence assessment in pose estimation and guides an uncertainty-aware sampling strategy for online object completion, enhancing robustness in pose estimation accuracy and improving object completeness. We evaluate our method on the YCB-Video, YCBIInEOAT, and HO3D datasets, including RGBD sequences of YCB objects manipulated by robots and human hands. Experimental results demonstrate significant performance improvements over existing methods, particularly when object observations are incomplete or partially captured. Project page: https://minfenli.github.io/UA-Pose/" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.815, + 0.21, + 0.83 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "6D object pose estimation, which determines the rigid 6-degree-of-freedom transformation between an object and a camera, is essential for various real-world applications such as robotic manipulation [17, 38, 39, 49] and augmented re" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.329, + 0.865, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.48, + 0.861, + 0.603 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.604, + 0.895, + 0.716 + ], + "angle": 0, + "content": "Figure 1. Overview of UA-Pose. (a) We initialize an incomplete object 3D model \\(\\mathcal{M}\\) (Sec. 3.2), which is a hybrid mesh that combines texture, geometry, and uncertainty. (b) Given a sequence of RGBD test images, our method estimates the 6D pose of the object in each test image (Sec. 3.3) based on the incomplete model \\(\\mathcal{M}\\). (c) While more test images are captured, we store them with estimated poses in a memory pool, which will be used for online object completion (Sec. 3.4) to refine the incomplete model \\(\\mathcal{M}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "ality [22, 29]. Recent research introduced model-based and model-free approaches that can generalize to arbitrary novel objects. These methods push the boundaries of pose estimation for real-world applications and can be divided into two scenarios based on the object information available beforehand. Model-based methods [1, 2, 4, 11, 13, 16, 23, 24, 28] require textured 3D CAD models of the object to estimate 6D object poses in 2D images by establishing 2D-3D correspondences. In contrast, model-free methods leverage a large set of posed reference RGB images [19], RGBD images [10, 41], or video sequences [7, 31] to estimate object poses without the need for CAD models." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1180" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.319 + ], + "angle": 0, + "content": "To bridge model-based and model-free methods, FoundationPose [41] proposes a unified framework that supports both model-based and model-free setups, using 3D CAD models or posed RGBD images (referred to as references). While effective in generalizing to novel objects, FoundationPose and similar model-free methods [7, 10, 31] still depend on high-quality 3D models or sufficient posed RGBD images that capture the object from sufficient viewpoints. However, assuming the availability of high-quality 3D models is often impractical due to the vast number of objects that remain uncaptured or unmodeled in real-world environments. Likewise, obtaining adequate posed RGBD images is challenging, as real-world settings frequently involve limited viewpoints, occlusions from nearby objects, and complexities in estimating camera poses in dynamic scenes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.321, + 0.473, + 0.625 + ], + "angle": 0, + "content": "To address these challenges, we present UA-Pose, an uncertainty-aware approach for 6D pose estimation and online object completion as shown in Fig. 1. Our approach is designed to work with partial reference inputs that can include valuable meta-information for robotic tasks, such as grasping patterns, detection markers, or affordance labels. We consider two scenarios: (1) When a limited set of reference RGBD images with known poses is available, we initialize an incomplete 3D model using these images and poses; (2) When only a single unposed RGB image is available, we apply image-to-3D techniques to generate an initial object 3D model for pose estimation. By marking unseen regions as uncertain, we introduce a hybrid representation (Sec. 3.2) that integrates texture, geometry, and uncertainty. The uncertainty is utilized during our proposed uncertainty-aware pose estimation (Sec. 3.3) to provide confidence for estimated poses. Additionally, we propose an uncertainty-aware sampling strategy to select informative images for online object completion, iteratively refining the object 3D model during testing (Sec. 3.4)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.627, + 0.47, + 0.718 + ], + "angle": 0, + "content": "We evaluate our approach on the YCB-Video, YCBI-nEOAT, and HO3D datasets. The results demonstrate significant improvements over existing methods and highlight the applicability of our uncertainty-aware approach in real-world scenarios where fully captured 3D models are often unavailable. To summarize, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.72, + 0.47, + 0.781 + ], + "angle": 0, + "content": "- We propose an uncertainty-aware approach for 6D object pose estimation (Sec. 3.3) and online object completion (Sec. 3.4), which improves both pose accuracy and object completeness when only partial references are available." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.781, + 0.47, + 0.841 + ], + "angle": 0, + "content": "- We develop a hybrid representation (Sec. 3.2) incorporating uncertainty. The uncertainty enables confidence assessment in pose estimation and supports an uncertainty-aware sampling strategy for online object completion." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.841, + 0.47, + 0.902 + ], + "angle": 0, + "content": "- We demonstrate pose estimation from a single RGB image by leveraging single-image-to-3D methods, which generate an initial model for pose estimation and provide augmented data for online object completion (Sec. 3.5)." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.72, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "- Our method supports different partial object references as input, which could be associated to meta-information, making it feasible for real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.152, + 0.642, + 0.168 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.177, + 0.819, + 0.194 + ], + "angle": 0, + "content": "2.1. Model-based Object Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.201, + 0.895, + 0.475 + ], + "angle": 0, + "content": "Traditional model-based pose estimation methods can be divided into two categories: instance-level [8, 9, 12, 25] and category-level [3, 15, 32, 34, 46, 47]. Both types of methods assume that a textured CAD model of the specific object or similar object categories is available during training. However, these approaches are limited to objects seen during training, significantly restricting their applicability in real-world scenarios where many objects remain unseen or unmodeled. To address this limitation, zero-shot methods have been proposed [2, 11, 13, 16, 28], which estimate the pose of novel objects by providing their CAD models at test time. These methods use object CAD models to either render 2D images for feature matching and 2D-3D correspondence estimation or leverage point clouds from CAD models to estimate 3D-3D correspondences. While these approaches improve generalization to unseen objects, they still require access to high-quality 3D models at test time, which is often impractical in real-world settings." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.485, + 0.802, + 0.501 + ], + "angle": 0, + "content": "2.2. Model-free Object Pose estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Model-free methods eliminate the need for an explicit textured CAD model. Gen6D [19] introduces a pipeline to estimate object poses from RGB images with known poses. However, due to limited training data, Gen6D faces challenges with out-of-distribution test cases and images affected by occlusions. OnePose [31] and its extension OnePose++ [7] use structure-from-motion techniques to reconstruct 3D point clouds from sequential RGB images and employ pre-trained 2D-3D matching networks to estimate the pose of test images. FS6D [10] employs a transformer-based architecture to extract features from multi-view RGBD images for pose estimation. More recently, FoundationPose [41], trained on a large synthetic dataset, presents a unified framework that supports both model-based and model-free setups. In its model-free setting, FoundationPose requires posed reference images that capture the object from sufficient viewpoints to reconstruct 3D models for robust pose estimation. In general, these model-free pose estimation methods rely on either sequential RGB images with known poses or densely posed RGBD images. However, assuming pre-captured posed RGBD images for unseen objects is impractical for two key reasons: first, real-world applications often involve numerous unseen objects, making it infeasible to pre-capture and collect images for each object; and second, even if sufficient images of the novel object are available, accurately estimating cam" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1181" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.411, + 0.108 + ], + "angle": 0, + "content": "era poses in dynamic environments is challenging." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.115, + 0.433, + 0.131 + ], + "angle": 0, + "content": "2.3. Object Pose Tracking and Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.473, + 0.381 + ], + "angle": 0, + "content": "Assuming test images for pose estimation are sequential, 6D object pose tracking methods [36, 40] leverage temporal cues to achieve efficient, smooth, and accurate pose predictions across consecutive frames by estimating relative pose transformations from each frame to the initial frame. For instance, BundleSDF [40] is a 6D pose tracking method that estimates poses from sequential RGBD images while simultaneously training a signed distance field (SDF) [45] representation to maintain consistent pose estimates over time. However, pose tracking methods are constrained to tracking object poses within sequential images from a single video, without support for using reference images for initial pose estimation. In contrast, our method is designed to estimate object poses based on given object references, which may include metadata beneficial for robotic applications, such as grasping patterns, barcodes, or affordance labels." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.388, + 0.424, + 0.403 + ], + "angle": 0, + "content": "2.4. Single-Image-to-3D and Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.473, + 0.669 + ], + "angle": 0, + "content": "Recently, single-image-to-3D methods have been developed to generate 3D models from a single image [18, 20, 21, 26, 27, 33, 43]. These methods offer potential for pose estimation when textured CAD models are unavailable. For instance, GigaPose [23] explored using generated models for pose estimation; however, relying on generated 3D models that are not closely similar to real objects can lead to imprecise pose estimates. To address this, we apply a single-image-to-3D approach [43] to synthesize a generated model for early pose estimation. As real RGBD images are gradually collected during testing, they are used to create a refined object model, while the generated model from single-image-to-3D methods is leveraged to produce augmented data to improve object geometry in the uncaptured parts. Incorporating these augmented data into the SDF training yields more complete object models, ultimately improving the robustness and accuracy of pose estimation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.68, + 0.169, + 0.695 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.825 + ], + "angle": 0, + "content": "An overview of our method is illustrated in Fig. 1. We begin by initializing an incomplete object 3D model \\(\\mathcal{M}\\) using partial references, which can be either: (1) a few reference RGBD images with known poses or (2) a single unposed RGB image of the object. In this work, we propose a hybrid object representation (Sec. 3.2) that integrates texture, geometry, and uncertainty, serving as the incomplete model \\(\\mathcal{M}\\) for uncertainty-aware pose estimation (Sec. 3.3)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Given a sequence of \\(k\\) RGB test images, \\(\\mathcal{I} = \\{I_0,I_1,\\dots ,I_{k - 1}\\}\\), corresponding depth maps, and an initial segmentation mask \\(m_0\\) for the object of interest in the first image \\(I_0\\), our method estimates the 6D pose of the object \\(\\xi_{i}\\) in each subsequent test image \\(I_{i}\\) using" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.243 + ], + "angle": 0, + "content": "the proposed uncertainty-aware pose estimation approach (Sec. 3.3), based on the incomplete 3D model \\(\\mathcal{M}\\). During testing, newly captured test RGBD images and their estimated poses will be appended to a memory pool. These images will be used for online object completion (Sec. 3.4) to refine the incomplete object 3D model \\(\\mathcal{M}\\). Notably, when the input reference is a single unposed RGB image, we employ a single-image-to-3D approach [43] to initialize a generated model \\(\\hat{\\mathcal{M}}\\) to assist in pose estimation and object completion. More details are described in Sec. 3.5." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.254, + 0.855, + 0.269 + ], + "angle": 0, + "content": "3.1. Background: Model-free Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.277, + 0.892, + 0.61 + ], + "angle": 0, + "content": "Given RGBD reference images of a novel object and their camera poses, the task of model-free pose estimation is to determine the 6D object poses (i.e., the rigid 6D transformation from the object to the camera) of that novel object in test RGBD images. FoundationPose [41] is a foundational model for 6D pose estimation trained on a large-scale dataset. In the model-free setup, it uses sufficient RGBD reference images (e.g., 16 RGBD images) that fully cover the object to train a Signed Distance Field (SDF) and then extracts a mesh by marching cubes as the 3D object model. Given the 3D object model to render object images, it introduces a two-stage pipeline with two specialized networks: a pose refinement network, which refines pose hypotheses by estimating the relative pose between the test image and rendered object images, and a pose selection network, which identifies the best pose by comparing the test image with rendered images based on these refined poses. However, RGBD reference images that fully represent the object's appearance and geometry are usually unavailable in real-world scenarios. As a result, model-free methods may struggle to provide accurate pose estimates when only incomplete object models or sparse observations are available." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.62, + 0.771, + 0.637 + ], + "angle": 0, + "content": "3.2. Hybrid Object Representation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.78 + ], + "angle": 0, + "content": "To address the limitations, we introduce a hybrid object representation that integrates the object's texture, geometry, and uncertainty, as shown in Fig. 2. This uncertainty marks the seen and unseen regions of an incomplete 3D object model and provides confidence in pose estimations, helping to filter out less reliable results. Given a set of RGBD images, object masks, and their camera poses, we train an object-centric neural Signed Distance Field (SDF) [45] to learn the partial object's appearance and geometry." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We then extract a 3D mesh using the marching cubes algorithm to explicitly model uncertainty. Inspired by Visual Hull [14], we define object uncertainty based on visibility inferred from 2D object masks. By projecting these masks onto the mesh \\(\\mathcal{M}\\), we label regions of \\(\\mathcal{M}\\) seen in the reference images as \"certain\" and those unseen as \"uncertain\", as shown in Fig. 2. This approach allows pose estimation to account for both the seen and unseen parts of the object." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1182" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.095, + 0.464, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.206, + 0.47, + 0.318 + ], + "angle": 0, + "content": "Figure 2. Hybrid Object Representation Modeling. We propose a hybrid object representation (Sec. 3.2) that integrates the object's texture, geometry, and uncertainty. (a) First, a neural SDF is trained and extracted as a mesh representing the object's appearance and geometry (Sec. 3.2.1). (b) Then, we check the visibility of each mesh vertex from the viewpoint of each reference image to create the uncertainty map (Sec. 3.2.2), which reflects the seen and unseen regions of an incomplete 3D object model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.322, + 0.43, + 0.337 + ], + "angle": 0, + "content": "3.2.1 Neural SDF Training and Mesh Extraction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.345, + 0.47, + 0.391 + ], + "angle": 0, + "content": "We follow SDF training methods similar to [40, 41] to train a neural SDF for object modeling. The SDF defines the object surface as the set of 3D points:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.398, + 0.47, + 0.416 + ], + "angle": 0, + "content": "\\[\nS = \\{x \\in \\mathbb {R} ^ {3} \\mid \\Omega (x) = 0 \\}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.47, + 0.588 + ], + "angle": 0, + "content": "where \\(\\Omega : \\mathbb{R}^3 \\to \\mathbb{R}\\) is the SDF geometry function that maps each 3D point \\(x\\) to a signed distance value \\(s\\) and \\(s = 0\\) indicates the object's surface. Given RGBD images, object masks, and their camera poses, we use a neural SDF to reconstruct the object's appearance and geometry, represented by two networks [45] as shown in Fig. 2. The geometric network \\(\\Omega : x \\mapsto s\\) takes a 3D point \\(x \\in \\mathbb{R}^3\\) as input and outputs a signed distance value \\(s \\in \\mathbb{R}\\). The appearance network \\(\\Phi : (f_{\\Omega(x)}, d) \\mapsto c\\) takes an intermediate feature vector \\(f_{\\Omega(x)}\\) from the geometric network along with a view direction \\(d \\in \\mathbb{R}^3\\) and outputs the color \\(c \\in \\mathbb{R}^3\\) for the point." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.589, + 0.469, + 0.634 + ], + "angle": 0, + "content": "**Rendering.** Given the object pose \\(\\xi\\), we render an image by casting rays through each pixel. Along each ray \\(r\\), 3D points are sampled at multiple positions as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.642, + 0.469, + 0.659 + ], + "angle": 0, + "content": "\\[\nx _ {i} (r) = o (r) + t _ {i} d (r), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.665, + 0.47, + 0.756 + ], + "angle": 0, + "content": "where \\(o(r)\\) denotes the ray origin and \\(d(r)\\) denotes the ray direction from the camera to the object determined by the object pose \\(\\xi\\). The parameter \\(t_i \\in \\mathbb{R}_+\\) determines the point position along the ray. Following the approach in [40], the color \\(c\\) of a ray \\(r\\) is computed by integrating sampled points near the object surface through volumetric rendering:" + }, + { + "type": "equation", + "bbox": [ + 0.123, + 0.763, + 0.47, + 0.801 + ], + "angle": 0, + "content": "\\[\nc (r) = \\int_ {z (r) - \\lambda} ^ {z (r) + 0. 5 \\lambda} w \\left(x _ {i}\\right) \\Phi \\left(f _ {\\Omega \\left(x _ {i}\\right)}, d \\left(x _ {i}\\right)\\right) d t, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.807, + 0.469, + 0.838 + ], + "angle": 0, + "content": "\\[\nw \\left(x _ {i}\\right) = \\frac {1}{1 + e ^ {- \\alpha \\Omega \\left(x _ {i}\\right)}} \\frac {1}{1 + e ^ {\\alpha \\Omega \\left(x _ {i}\\right)}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\( w(x_{i}) \\) is a bell-shaped probability density function [35] based on the distance of each point from the implicit object surface and \\( \\alpha \\) adjusts the sharpness of the density distribution. \\( z(r) \\) represents the ray's depth from the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "depth image and \\(\\lambda\\) is the truncation distance. During volumetric rendering, \\(z(r)\\) and \\(\\lambda\\) are used to exclude 3D points far from the object surface, enhancing rendering efficiency and preventing the SDF from overfitting to empty regions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.892, + 0.183 + ], + "angle": 0, + "content": "During training, we supervise this rendering by comparing it to the reference RGB images with the color loss:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.193, + 0.892, + 0.23 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = \\frac {1}{| \\mathcal {R} |} \\sum_ {r \\in \\mathcal {R}} \\| c (r) - \\bar {c} (r) \\| _ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.242, + 0.893, + 0.393 + ], + "angle": 0, + "content": "where \\(\\bar{c}(r)\\) is the ground-truth color at the pixel through which the ray \\(r\\) passes. For geometry learning, we leverage depth information from the provided depth maps to supervise the neural (SDF) through two losses [40]: the near-surface loss \\(\\mathcal{L}_s\\) and the empty space loss \\(\\mathcal{L}_e\\). \\(\\mathcal{L}_s\\) encourage the SDF surface to align closely with depth values and \\(\\mathcal{L}_e\\) enforces that points along the ray between the camera and the surface remain empty. Additionally, we apply an eikonal regularization loss \\(\\mathcal{L}_{eik}\\) [5] to the near-surface SDF to enforce smoothness. The total loss is represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.407, + 0.891, + 0.423 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = w _ {c} \\mathcal {L} _ {c} + w _ {e} \\mathcal {L} _ {e} + w _ {s} \\mathcal {L} _ {s} + w _ {e i k} \\mathcal {L} _ {e i k}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.436, + 0.833, + 0.451 + ], + "angle": 0, + "content": "where \\(w_{c}, w_{e}, w_{s}\\), and \\(w_{eik}\\) are weighting factors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.451, + 0.892, + 0.586 + ], + "angle": 0, + "content": "To explicitly model uncertainty, we convert the implicit surface defined by the neural SDF into an explicit 3D mesh \\(\\mathcal{E} = (V,C,F)\\) with the vertices \\(V\\), vertex colors \\(C\\), and faces \\(F\\). To achieve this, we extract a mesh from the trained neural SDF representation using the marching cubes algorithm, which guarantees a closed-surface mesh including vertices from any viewing angle. This allows us to model the visibility of the object from any viewing angle even if some parts are not visible in the reference images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.587, + 0.892, + 0.647 + ], + "angle": 0, + "content": "Each vertex \\( v_{i} \\in V \\) is assigned a color \\( c_{i} \\in C \\), which is computed using the appearance function \\( \\varPhi \\), which maps the intermediate feature vector \\( f_{\\Omega(v_{i})} \\) from the geometry network and a view direction \\( d \\in \\mathbb{R}^3 \\) to the RGB color:" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.66, + 0.891, + 0.677 + ], + "angle": 0, + "content": "\\[\nc _ {i} = \\Phi \\left(f _ {\\Omega \\left(v _ {i}\\right)}, d\\right). \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.708, + 0.71 + ], + "angle": 0, + "content": "3.2.2 Uncertainty Modeling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.902 + ], + "angle": 0, + "content": "To model the uncertainty map that represents whether a vertex is visible from the viewpoints of input references, for each vertex \\( v_{i} \\in V \\), we check its visibility from the viewpoint of each reference image \\( \\bar{I}_j \\in \\tilde{\\mathcal{I}} \\) through the mesh rasterization process, which rasterizes 3D vertices onto 2D image pixels. We define an uncertainty score \\( u(v_{i}) \\) based on whether \\( v_{i} \\) is visible from any of the viewpoints in the reference images \\( \\bar{I}_j \\). This uncertainty score \\( u(v_{i}) \\in \\{0,1\\} \\) labels vertices on the mesh as \"certain\" if they are seen in any reference image \\( (u(v_{i}) = 0) \\) and as \"uncertain\" if they remain unseen \\( (u(v_{i}) = 1) \\). The resulting uncertainty map \\( \\mathcal{U} = \\{u(v_i) \\mid v_i \\in V\\} \\) is incorporated into our hybrid object" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1183" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.09, + 0.887, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.221, + 0.894, + 0.306 + ], + "angle": 0, + "content": "Figure 3. Pipeline for Uncertainty-Aware Pose Estimation. Through uncertainty modeling (Sec. 3.2.2), we propose an uncertainty-aware pipeline to assess the confidence of estimated poses (Sec. 3.3). For each estimated pose, we calculate its seen IoU, which measures the overlap between the seen regions of the 3D model and the 2D object mask for the test image. When seen IoU is high, the pose is deemed confident, and both the pose and its RGBD image are added to the memory pool for object completion. Conversely, if the estimated pose has a low seen IoU, it is considered unreliable. In such cases, we employ an uncertainty-aware image sampling strategy to select images from the memory pool for online object completion (Sec. 3.4), thus enhancing pose estimation by refining the object model during testing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.315, + 0.468, + 0.345 + ], + "angle": 0, + "content": "representation \\(\\mathcal{M} = (\\mathcal{E},\\mathcal{U})\\), allowing the pose estimation process to distinguish between seen and unseen regions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.346, + 0.47, + 0.485 + ], + "angle": 0, + "content": "Guidance from Uncertainty Map. Given an object pose \\(\\xi \\in \\mathbb{R}^{3\\times 4}\\), the model \\(\\mathcal{M}\\) can be used to render an RGB image \\(I_{\\xi}^{\\mathrm{rend}}\\in \\mathbb{R}^{H\\times W\\times 3}\\), a depth map \\(D_{\\xi}^{\\mathrm{rend}}\\in \\mathbb{R}^{H\\times W}\\), a rendered object mask \\(m_{\\xi}^{\\mathrm{rend}}\\in \\{0,1\\}^{H\\times W}\\), and an uncertainty map \\(U_{\\xi}^{\\mathrm{rend}}\\in \\{0,1\\}^{H\\times W}\\) at a resolution of \\(H\\times W\\). To evaluate the validity and confidence of an estimated pose, we introduce two metrics: uncertainty rate and seen IoU. The uncertainty rate measures the proportion of the uncertain pixels over the total pixels within the rendered object mask:" + }, + { + "type": "equation", + "bbox": [ + 0.207, + 0.495, + 0.47, + 0.534 + ], + "angle": 0, + "content": "\\[\n\\frac {\\sum \\left(U _ {\\xi} ^ {\\text {r e n d}} \\odot m _ {\\xi} ^ {\\text {r e n d}}\\right)}{\\sum \\left(m _ {\\xi} ^ {\\text {r e n d}}\\right)}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.47, + 0.6 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes element-wise logical multiplication. The seen IoU measures the overlap between seen (certain) regions of the rendered object mask and the object mask in the test image, which is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.609, + 0.47, + 0.63 + ], + "angle": 0, + "content": "\\[\n\\mathrm {I o U} \\left(\\neg U _ {\\xi} ^ {\\text {r e n d}} \\odot m _ {\\xi} ^ {\\text {r e n d}}, m ^ {\\text {t e s t}}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.47, + 0.749 + ], + "angle": 0, + "content": "where \\(m^{\\mathrm{test}}\\) is the object mask from the test image, \\(\\neg U_{\\xi}^{\\mathrm{rend}}\\odot\\) \\(m_{\\xi}^{\\mathrm{rend}}\\) identifies \"certain\" pixels within the rendered object mask, and IoU represents the Intersection over Union function. Since pose estimation is more robust when based on seen regions of the 3D model, a high seen IoU indicates high confidence in the pose estimation, whereas a low seen IoU suggests that the estimation is not reliable." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.388, + 0.773 + ], + "angle": 0, + "content": "3.3. Uncentainty-aware Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.902 + ], + "angle": 0, + "content": "After modeling the uncertainty, the hybrid object representation could be used in our uncertainty-aware pose estimation pipeline, as shown in Fig. 3. In this work, we leverage the pose refinement and the pose selection modules of [41] and assume that the test images are sequential. To obtain per-frame 2D object masks in test images for calculating the uncertainty rate and seen IoU, given first frame object mask \\( m_0 \\), the object masks of the following" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.315, + 0.892, + 0.345 + ], + "angle": 0, + "content": "frames \\(\\{m_1, m_2, \\ldots, m_{k-1}\\}\\) is determined by off-the-shelf segmentation methods [44]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.349, + 0.893, + 0.56 + ], + "angle": 0, + "content": "Pose Initialization. Given an RGBD video, object masks for each frame, and an incomplete object model \\(\\mathcal{M}\\), we begin by estimating the object pose for the first frame by generating multiple pose hypotheses. For each hypothesis, the translation is initialized using the 3D point projected from the center of the 2D object mask, determined by the median depth value within this region. For rotation, we uniformly sample \\(N_{v}\\) viewpoints from an icosphere centered on the object and orient the camera to face the object's center. Subsequently, \\(N_{in}\\) in-plane rotations are sampled and applied to each viewpoint, yielding \\(N_{v} \\cdot N_{in}\\) hypothesized object poses. Each object pose can be represented as \\([\\pmb{R}|\\pmb{t}] \\in \\mathbb{S}\\mathbb{E}(3)\\), where \\(\\pmb{R} \\in \\mathbb{S}\\mathbb{O}(3)\\) is the rotation and \\(\\pmb{t} \\in \\mathbb{R}^3\\) is the translation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.894, + 0.746 + ], + "angle": 0, + "content": "Pose Refinement. After generating \\(N_v \\cdot N_{in}\\) pose hypotheses, we apply the refinement module [41] to improve pose accuracy. Specifically, given a rendered RGBD image of the object from each pose hypothesis and the test RGBD image of the target object, the network outputs a pose update \\(\\Delta R \\in \\mathbb{SO}(3)\\) and \\(\\Delta t \\in \\mathbb{R}^3\\) to refine the pose hypothesis, aligning it more closely with the observed object pose in the test image. Each refined pose is represented as \\([R^{+}|t^{+}] \\in \\mathbb{SE}(3)\\), where \\(t^{+} = t + \\Delta t\\) and \\(R^{+} = \\Delta R \\otimes R\\). This refinement process could be iteratively repeated. In this work, each pose is refined five times for pose estimation on the first frame and twice for subsequent frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Pose Selection. Given a set of refined pose hypotheses, we apply our proposed metrics (Sec. 3.2.2) to filter out unreliable poses: the uncertainty rate, which measures the proportion of uncertain pixels within the rendered object mask, and the seen IoU, which quantifies the overlap between previously seen (certain) regions of the object and the object mask in the test image. A pose hypothesis is discarded when its uncertainty rate is above the threshold \\( T_{u} \\) or its seen IoU is below the threshold \\( T_{s} \\), ensuring that only reliable hypotheses that adequately cover seen object regions are con-" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1184" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "sidered. At last, we use the pose selection module [41] to identify the most accurate pose as the final estimate." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.468, + 0.244 + ], + "angle": 0, + "content": "Pose Tracking. Given the sequential nature of RGBD video frames, object poses in adjacent frames are typically similar. For each frame following the initial one, we take the estimated pose from the previous frame as the sole pose hypothesis and refine it using the pose refinement module to generate the estimated pose for the current frame. This iterative process continues across the video sequence, leveraging temporal cues for stable and smooth pose estimation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.245, + 0.471, + 0.684 + ], + "angle": 0, + "content": "Memory Pool. To enhance pose estimation by completing the object model during testing, we select test RGBD images with their estimated poses to incrementally refine the object representation. For efficiency, we introduce a memory pool \\(\\mathcal{P} = \\{\\xi_0,\\xi_1,\\xi_2,\\dots ,\\xi_{|\\mathcal{P}| - 1}\\}\\) that stores the most informative \\(|\\mathcal{P}|\\) viewpoints to maintain a compact yet diverse set of test image frames for object completion, where \\(\\xi_{i}\\) is the estimated pose corresponding to a test image \\(I_{i}\\). The first frame \\(\\xi_{i}\\) is automatically added to \\(\\mathcal{P}\\), establishing the canonical coordinate system for the novel object. Subsequent frames are added when their viewpoints contribute to the multi-view diversity in the pool. Specifically, before adding a new frame, we calculate the rotational geodesic distance between the new frame and the last frame added to the pool. A frame is included in \\(\\mathcal{P}\\) only if this geodesic distance exceeds a threshold \\(T_{geo}\\), ensuring that each frame offers a unique viewpoint to enhance model completeness while keeping \\(\\mathcal{P}\\) compact. Note that not all estimated poses are reliable, adding inaccurate results to the memory pool may cause noisy SDF training during object completion. To mitigate this, we introduce an image-filtering strategy to filter out low-confidence estimations. As illustrated in Fig. 3, when a pose is estimated, its confidence is evaluated using our proposed metric, seen IoU, which measures the overlap between the seen regions of the 3D model and the 2D object mask in the test image. Since a high seen IoU indicates greater confidence in an estimated pose, frames are added to the memory pool only if their seen IoU exceeds a threshold \\(T_{conf}\\), ensuring that low-confidence poses are excluded." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.695, + 0.411, + 0.712 + ], + "angle": 0, + "content": "3.4. Uncentainty-aware Object Completion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.468, + 0.84 + ], + "angle": 0, + "content": "As shown in Fig. 3, online object completion is triggered when the seen IoU of the current test frame falls below a threshold \\( T_{complete} \\), indicating that the current pose estimation may no longer be reliable. As new frames are continuously added, the memory pool \\( \\mathcal{P} \\) can grow excessively large, which may reduce SDF training efficiency. To manage this, we perform uncertainty-aware sampling to select the \\( K \\) most informative frames for object completion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.902 + ], + "angle": 0, + "content": "Uncertainty-aware Sampling. Early in the video sequence, when \\( |\\mathcal{P}| \\leq K \\), all frames in the pool are included without selection. However, once the pool size exceeds \\( K \\) we apply an uncertainty-aware image selection strategy to" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.093, + 0.782, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.797, + 0.093, + 0.882, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.244, + 0.892, + 0.368 + ], + "angle": 0, + "content": "Figure 4. Pipeline for leveraging generated models. In Sec. 3.5, the initial model generated by image-to-3D techniques is used to estimate poses for the first few frames. As more frames are captured, a refined object model is trained to replace the initial generated model, resulting in a more complete representation. To further leverage the generated 3D model, we render RGBD images from the image-to-3D generated model as augmented data for SDF training. This supervision enhances the object geometry in unseen regions, providing more complete information for pose estimation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.381, + 0.892, + 0.502 + ], + "angle": 0, + "content": "maximize the coverage of unseen object regions. This process selects a subset, \\(\\mathcal{P}' \\subset \\mathcal{P}\\), that reveals the most unseen areas of the object. The selection process begins by including the initial and most recent frames in the pool to establish the starting and ending viewpoints. Next, an iterative strategy is employed: at each iteration, we select the pose that reveals the largest unseen region. This continues until \\(K\\) poses are selected." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.514, + 0.816, + 0.53 + ], + "angle": 0, + "content": "3.5. Leveraging Image-to-3D Generation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.538, + 0.892, + 0.719 + ], + "angle": 0, + "content": "Given a single unposed RGB image, we employ image-to-3D techniques to generate an initial object 3D model \\(\\hat{\\mathcal{M}}\\) for pose estimation. The viewpoint associated with the initial RGB image is labeled as \"certain,\" while other areas that are inferred by the image-to-3D method are marked \"uncertain\". The generated model is then rescaled to align with the object size in the test images. Specifically, we use the depth map and object mask from the first frame of the test image to approximate a rough model size. We then sample slightly larger and smaller models, generate pose hypotheses, and employ the pose selection module [41] to select the optimal pose hypothesis and corresponding model size." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Refer to Fig. 4, the initial generated model is used to estimate poses for the first few test frames until the rotational geodesic distance between the initial pose and a new estimated pose exceeds a threshold \\( T_{gen} \\). As more frames are captured, a refined object model is trained to replace the initial generated model, resulting in a more complete representation. To further leverage the generated 3D model, we render RGBD images from the image-to-3D generated model as augmented data for SDF training. This supervision enhances the object geometry in unseen regions, providing more complete information for pose estimation. See the supplementary materials for additional details." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1185" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.09, + 0.201, + 0.108 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.115, + 0.18, + 0.13 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.468, + 0.184 + ], + "angle": 0, + "content": "YCB-Video [42]: The YCB-Video dataset contains RGBD video sequences of everyday objects from the YCB objects, widely used for benchmarking 6D object pose estimation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.184, + 0.468, + 0.244 + ], + "angle": 0, + "content": "YCBCInEOAT [37]: YCBCInEOAT extends the YCB-Video dataset with RGBD sequences of YCB objects manipulated by a dual-arm robot. This dataset focuses on real-world scenarios involving robot-object interactions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.468, + 0.29 + ], + "angle": 0, + "content": "HO3D [6]: The HO3D (Hand-Object 3D) dataset contains RGBD sequences of human hands interacting with objects, capturing complex hand-object interactions and occlusions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.3, + 0.265, + 0.317 + ], + "angle": 0, + "content": "4.2. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.468, + 0.444 + ], + "angle": 0, + "content": "Baselines. We evaluate our method against state-of-the-art RGBD model-free 6D pose estimation methods [10, 30, 41] under two scenarios where partial object references are given as (1) A limited set of reference RGBD images (2 views) that include known poses. (2) A single unposed RGB object image. In scenario (2), we leverage image-to-3D techniques [43] to generate an initial object 3D model and iteratively estimate the object pose during testing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.444, + 0.468, + 0.581 + ], + "angle": 0, + "content": "Metrics. We consider ADD and ADD-S metrics for 6D pose estimation, which measure the accuracy of estimated poses. The Area Under the Curve (AUC) is calculated and reported for both the ADD and ADD-S metrics, following the protocols in [40-42]. Furthermore, to demonstrate the effectiveness of our method for online object completion, we report the Chamfer Distance (CD), which measures the average distance between points of the reconstructed model and the ground-truth model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.59, + 0.385, + 0.607 + ], + "angle": 0, + "content": "4.3. Comparison Results on YCB-Video" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.468, + 0.674 + ], + "angle": 0, + "content": "We first evaluate UA-Pose on the YCB-Video [42] dataset, a widely used benchmark for 6D pose estimation, and compare our approach with state-of-the-art model-free methods [10, 30, 41] with a limited set of reference images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.468, + 0.901 + ], + "angle": 0, + "content": "As presented in Tab. 1, our method significantly outperforms FoundationPose [41] with only 2 reference images in all metrics. Moreover, both our method and FoundationPose outperforms LoFTR [30] and FS6D-DPM [10] with 16 reference images with a large margin. This shows that FoundationPose is the strongest baseline for our further analysis. When only a single unposed RGB image of the object is available, we apply a single-image-to-3D approach [43] to generate an initial 3D model for pose estimation. Experimental results, shown in Tab. 1, indicate that directly using the generated model with FoundationPose may not yield optimal results, as these models may not accurately represent real objects. In contrast, UA-Pose uses the generated model only for initial pose estimation and for rendering RGBD images as augmented data to aid object" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.173 + ], + "angle": 0, + "content": "Table 1. Quantitative results on the YCB-Video dataset. \"Comp.\" indicates whether the method includes online completion, and \"-\" denotes that the method does not reconstruct object shapes. Results for [10, 30] are adopted from [41]. Experiments for \"single RGB + Image-to-3D\" are conducted on a subset of the YCB-Video dataset. See supplementary materials for details." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.173, + 0.892, + 0.259 + ], + "angle": 0, + "content": "
InputComp.Mean
ADDADD-SCD
LoFTR [30]16 referencesNo26.252.5-
FS6D-DPM [10]16 referencesNo42.188.4-
FoundationPose [41]2 referencesNo87.494.30.57
Ours2 referencesYes92.896.50.53
FoundationPose [41]single RGB + Image-to-3DNo88.993.60.67
Ourssingle RGB + Image-to-3DYes93.296.90.62
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.267, + 0.892, + 0.294 + ], + "angle": 0, + "content": "Table 2. Quantitative results on the YCBIInEOAT dataset. \"Comp.\" indicates that the method performs online completion." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.295, + 0.892, + 0.359 + ], + "angle": 0, + "content": "
InputComp.Mean
ADDADD-SCD
FoundationPose [41]2 referencesNo68.5284.800.60
Ours2 referencesYes89.9994.350.57
FoundationPose [41]single RGB + Image-to-3DNo83.3990.660.61
Ourssingle RGB + Image-to-3DYes89.2493.990.60
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.375, + 0.892, + 0.421 + ], + "angle": 0, + "content": "completion, as described in Sec. 3.5. This approach reconstructs a model more closely aligned with the real object, leading to improved pose estimation accuracy." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.432, + 0.822, + 0.449 + ], + "angle": 0, + "content": "4.4. Comparison Results on YCBIInEOAT" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.456, + 0.892, + 0.502 + ], + "angle": 0, + "content": "We evaluated UA-Pose on the YCBIInEOAT [37] dataset to assess its performance in challenging robotic scenarios in real-world where objects are manipulated by robotic arms." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.502, + 0.892, + 0.713 + ], + "angle": 0, + "content": "In Table 2, when 2 reference images are provided, UA-Pose demonstrates superior performance compared to FoundationPose, highlighting the effectiveness of our uncertainty-aware mechanisms and online object completion for robust pose estimation in robotic scenarios. As illustrated in Fig. 5, simply using the incomplete model generated from the partial references often leads to inaccurate estimations. In contrast, our method achieves robust pose estimation with uncertainty-aware mechanisms and online object completion. When only a single unposed RGB image of the object is available, UA-Pose also consistently surpasses FoundationPose across all metrics. This demonstrates the flexibility of UA-Pose to utilize diverse partial references, which could carry useful meta-information." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.725, + 0.771, + 0.741 + ], + "angle": 0, + "content": "4.5. Comparison Results on HO3D" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.795 + ], + "angle": 0, + "content": "We evaluated UA-Pose on the HO3D dataset to assess its effectiveness in complex hand-object interaction scenarios where objects are frequently rotated and occluded." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As shown in Tab. 3, when 2 reference images are provided, FoundationPose fails to make accurate estimations when objects in the test video exhibit frequent and strong rotations. In contrast, UA-Pose accurately estimates object poses under the challenging hand-object interaction scenarios. As for utilizing the generated models from the single-image-to-3D approach for pose estimation, UA-Pose" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1186" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.085, + 0.462, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.232, + 0.47, + 0.344 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison on the YCBIInEOAT. Object image (left) and depth map (right) pairs are rendered based on the estimated object poses using the incomplete model. Without uncertainty-aware mechanisms and Online Object Completion (referred to as \"w/o OOC\"), the incomplete model often leads to inaccurate estimations. In contrast, our method (referred to as \"w/ OOC\") achieves robust pose estimation by incorporating uncertainty-aware mechanisms and online object completion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.356, + 0.47, + 0.506 + ], + "angle": 0, + "content": "achieves more accurate estimations than FoundationPose. Note that even if the generated models in Tab. 3 show lower Chamfer Distance (CD) than refined models obtained from object completion, directly using generated models for FoundationPose results in considerably lower ADD and ADD-S. This indicates that even though generated models have good geometry (low CD), geometry alone cannot accurately represent real objects for pose estimation because the textures of generated models are not good enough in unseen areas that are not covered by the initial partial references." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.515, + 0.23, + 0.531 + ], + "angle": 0, + "content": "4.6. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Key Components. Tab. 4 presents an ablation study for key components of UA-Pose on the YCBCInEOAT dataset. All experiments start with the initialized model based on two reference images and are evaluated using ADD, ADD-S, and CD. Additionally, we report \\( N_{rebuild} \\), the total number of times object completion is applied across test sequences to refine the incomplete model \\( \\mathcal{M} \\). The \"w/o object completion\" setup disables online object completion, illustrating its importance for achieving accurate pose estimation. In \"w/o uncertainty-aware object completion\", object completion is applied every time a new test image is added to the memory pool, instead of completing the model only when the seen IoU threshold is below \\( T_{complete} \\). This leads to poorer performance and a higher \\( N_{rebuild} \\) count, increasing unnecessary computation for object completion. The \"w/o image-filtering strategy\" experiment omits the image-filtering strategy (Sec. 3.3) for excluding unreliable poses, allowing low-confidence poses to be added to the memory pool and resulting in decreased pose accuracy. In \"w/o uncertainty-aware sampling\", we use the geodesic distance-based sampling strategy from [40] instead of our proposed uncertainty-aware sampling strategy (Sec. 3.4) to sample images for object completion. The results demonstrate our proposed sampling method is better at selecting infor" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.117 + ], + "angle": 0, + "content": "Table 3. Quantitative results on the HO3D dataset. \"Comp.\" indicates that the method performs online completion." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.118, + 0.892, + 0.182 + ], + "angle": 0, + "content": "
InputComp.Mean
ADDADD-SCD
FoundationPose [41]2 referencesNo45.6761.570.78
Ours2 referencesYes91.5195.560.69
FoundationPose [41]single RGB + Image-to-3DNo72.0687.870.76
Ourssingle RGB + Image-to-3DYes83.2393.590.88
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.188, + 0.892, + 0.215 + ], + "angle": 0, + "content": "Table 4. Ablation study. We demonstrate the importance of UA-Pose's key components on the YCBCInEOAT dataset." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.217, + 0.877, + 0.293 + ], + "angle": 0, + "content": "
MeanNrebuild
ADDADDSCD
w/o object completion68.5284.800.60-
w/o uncertainty-aware object completion85.8392.390.72181
w/o image-filtering strategy87.7093.370.7187
w/o uncertainty-aware sampling87.8893.390.6859
Ours (full)89.9994.350.5758
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.298, + 0.892, + 0.366 + ], + "angle": 0, + "content": "Table 5. Additional study. We compare our method with FoundationPose [41] and pose tracking methods [36, 40] on the YCBI-nEOAT dataset. \"Comp.\" means the method performs online completion, and \"\\*\" denotes using cropped ground-truth meshes [40], which exclude invisible parts in the test sequence, to calculate CD." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.368, + 0.877, + 0.432 + ], + "angle": 0, + "content": "
InputComp.Mean
ADDADD-SCD
BundleTrack [36]first-frame RGBDNo87.3492.532.81*
BundleSDF [40]first-frame RGBDYes86.9593.771.16*
FoundationPose [41]first-frame RGBDNo74.3484.581.12*
Oursfirst-frame RGBDYes88.3893.820.75*
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.892, + 0.5 + ], + "angle": 0, + "content": "mative images for accurate pose estimation. Finally, the \"Ours (full)\" setup that incorporates all our key components achieves the best performance across ADD, ADD-S, CD, and the fewest \\( N_{\\text{rebuild}} \\) for object completion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.893, + 0.697 + ], + "angle": 0, + "content": "Additional Study. When no external reference images are available, we demonstrate that we could initialize the object model for pose estimation using the first frame of the test RGBD image sequence. Besides the strongest model-free pose estimation baseline [41], we also compare our method with strong model-free pose tracking approaches [36, 40, 48] on the YCBIInEOAT dataset. Refer to Tab. 5, our method achieves the highest performance across ADD, ADD-S, and CD metrics. Note that unlike pose-tracking methods which are limited to estimating relative poses from sequential frames within the same RGBD video, our method can incorporate object reference images as input for useful meta-information." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.619, + 0.725 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We introduced UA-Pose, an uncertainty-aware approach for 6D object pose estimation that addresses the limitations of previous model-free methods in the scenarios in which input object references are incomplete or partially captured. Incorporating mechanisms to handle uncertainty and leverage online object completion to complete object models, our method demonstrates substantial performance improvements on the YCB-Video, YCBIInEOAT, and HO3D datasets. We demonstrate UA-Pose is a robust solution for practical real-world applications where full 3D object models and extensive image references are usually unavailable." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1187" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Philipp Ausserlechner, David Haberger, Stefan Thalhammer, Jean-Baptiste Weibel, and Markus Vincze. Zs6d: Zero-shot 6d object pose estimation using vision transformers. ICRA, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.468, + 0.226 + ], + "angle": 0, + "content": "[2] Andrea Caraffa, Davide Boscaini, Amir Hamza, and Fabio Poiesi. Freeze: Training-free zero-shot 6d pose estimation with geometric and vision foundation models. ECCV, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.229, + 0.468, + 0.27 + ], + "angle": 0, + "content": "[3] Dengsheng Chen, Jun Li, Zheng Wang, and Kai Xu. Learning canonical shape space for category-level 6D object pose and size estimation. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.272, + 0.468, + 0.327 + ], + "angle": 0, + "content": "[4] Jianqiu Chen, Zikun Zhou, Mingshan Sun, Rui Zhao, Liwei Wu, Tianpeng Bao, and Zhenyu He. Zeropose: Cadprompted zero-shot object 6d pose estimation in cluttered scenes. TCSVT, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.329, + 0.468, + 0.37 + ], + "angle": 0, + "content": "[5] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.468, + 0.413 + ], + "angle": 0, + "content": "[6] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In CVPR, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.416, + 0.468, + 0.47 + ], + "angle": 0, + "content": "[7] Xingyi He, Jiaming Sun, Yuang Wang, Di Huang, Hujun Bao, and Xiaowei Zhou. OnePose++: Keypoint-free one-shot object pose estimation without CAD models. NeurIPS, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.473, + 0.468, + 0.514 + ], + "angle": 0, + "content": "[8] Yisheng He, Wei Sun, Haibin Huang, Jianran Liu, Haoqiang Fan, and Jian Sun. PVN3D: A deep point-wise 3D keypoints voting network for 6DoF pose estimation. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.516, + 0.468, + 0.557 + ], + "angle": 0, + "content": "[9] Yisheng He, Haibin Huang, Haoqiang Fan, Qifeng Chen, and Jian Sun. FFB6D: A full flow bidirectional fusion network for 6D pose estimation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.559, + 0.468, + 0.599 + ], + "angle": 0, + "content": "[10] Yisheng He, Yao Wang, Haoqiang Fan, Jian Sun, and Qifeng Chen. Fs6d: Few-shot 6d pose estimation of novel objects. In CVPR, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.602, + 0.468, + 0.657 + ], + "angle": 0, + "content": "[11] Junwen Huang, Hao Yu, Kuan-Ting Yu, Nassir Navab, Slobodan Ilic, and Benjamin Busam. Matchu: Matching unseen objects for 6d pose estimation from rgb-d images. In CVPR, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.468, + 0.701 + ], + "angle": 0, + "content": "[12] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. CosyPose: Consistent multi-view multi-object 6D pose estimation. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.468, + 0.77 + ], + "angle": 0, + "content": "[13] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. MegaPose: 6D pose estimation of novel objects via render & compare. In CoRL, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.468, + 0.801 + ], + "angle": 0, + "content": "[14] Aldo Laurentini. The visual hull concept for silhouette-based image understanding. IEEE TPAMI, 1994. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.468, + 0.857 + ], + "angle": 0, + "content": "[15] Taeyeop Lee, Jonathan Tremblay, Valts Blukis, Bowen Wen, Byeong-Uk Lee, Inkyu Shin, Stan Birchfield, In So Kweon, and Kuk-Jin Yoon. TTA-COPE: Test-time adaptation for category-level object pose estimation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[16] Jiehong Lin, Lihua Liu, Dekun Lu, and Kui Jia. Sam-6d: Segment anything model meets zero-shot 6d object pose estimation. In CVPR, 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[17] Jian Liu, Wei Sun, Chongpei Liu, Xing Zhang, and Qiang Fu. Robotic continuous grasping system by shape transformer-guided multi-object category-level 6d pose estimation. IEEE Transactions on Industrial Informatics, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[18] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.244 + ], + "angle": 0, + "content": "[19] Yuan Liu, Yilin Wen, Sida Peng, Cheng Lin, Xiaoxiao Long, Taku Komura, and Wenping Wang. Gen6D: Generalizable model-free 6-DoF object pose estimation from RGB images. ECCV, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[20] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Learning to generate multiview-consistent images from a single-view image. ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.302, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[21] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, and Wenping Wang. Wonder3d: Single image to 3d using cross-domain diffusion. CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[22] Eric Marchand, Hideaki Uchiyama, and Fabien Spindler. Pose estimation for augmented reality: A hands-on survey. TVCG, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[23] Van Nguyen Nguyen, Thibault Groueix, Mathieu Salzmann, and Vincent Lepetit. Gigapose: Fast and robust novel object pose estimation via one correspondence. CVPR, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.457, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[24] Evin Pinar Ornek, Yann Labbe, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, and Tomas Hodan. Found- pose: Unseen object pose estimation with foundation features. ECCV, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.512, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[25] Kiru Park, Timothy Patten, and Markus Vincze. Pix2Pose: Pixel-wise coordinate regression of objects for 6D pose estimation. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.554, + 0.892, + 0.634 + ], + "angle": 0, + "content": "[26] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, and Bernard Ghanem. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.892, + 0.703 + ], + "angle": 0, + "content": "[27] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.707, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[28] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. OSOP: A multi-stage one shot object pose estimation framework. In CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.749, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[29] Yongzhi Su, Jason Rambach, Nareg Minaskan, Paul Lesur, Alain Pagani, and Didier Stricker. Deep multi-state object pose estimation for augmented reality assembly. In ISMAR-Adjunct, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[30] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In CVPR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[31] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. OnePose: One-shot object pose estimation without CAD models. In CVPR, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1188" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[32] Meng Tian, Marcelo H Ang, and Gim Hee Lee. Shape prior deformation for categorical 6D object pose and size estimation. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[33] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In ECCV, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.469, + 0.263 + ], + "angle": 0, + "content": "[34] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6D object pose and size estimation. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.266, + 0.469, + 0.32 + ], + "angle": 0, + "content": "[35] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.324, + 0.469, + 0.364 + ], + "angle": 0, + "content": "[36] Bowen Wen and Kostas Bekris. Bundletrack: 6d pose tracking for novel objects without instance or category-level 3d models. In IROS, 2021. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.368, + 0.469, + 0.421 + ], + "angle": 0, + "content": "[37] Bowen Wen, Chaitanya Mitash, Baozhang Ren, and Kostas E Bekris. se (3)-tracknet: Data-driven 6d pose tracking by calibrating image residuals in synthetic domains. In IROS, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.425, + 0.469, + 0.467 + ], + "angle": 0, + "content": "[38] Bowen Wen, Wenzhao Lian, Kostas Bekris, and Stefan Schaal. CatGrasp: Learning category-level task-relevant grasping in clutter from simulation. In ICRA, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.511 + ], + "angle": 0, + "content": "[39] Bowen Wen, Wenzhao Lian, Kostas Bekris, and Stefan Schaal. You only demonstrate once: Category-level manipulation from single visual demonstration. RSS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.513, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[40] Bowen Wen, Jonathan Tremblay, Valts Blukis, Stephen Tyree, Thomas Müller, Alex Evans, Dieter Fox, Jan Kautz, and Stan Birchfield. Bundlesdf: Neural 6-dof tracking and 3d reconstruction of unknown objects. In CVPR, 2023. 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.584, + 0.469, + 0.626 + ], + "angle": 0, + "content": "[41] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. Foundationpose: Unified 6d pose estimation and tracking of novel objects. CVPR, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.629, + 0.469, + 0.67 + ], + "angle": 0, + "content": "[42] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. RSS, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.672, + 0.469, + 0.74 + ], + "angle": 0, + "content": "[43] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[44] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[45] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. NeurIPS, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[46] Ruida Zhang, Yan Di, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. SSP-Pose: Symmetry-aware shape prior deformation for direct category-level object pose estimation. In IROS, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[47] Linfang Zheng, Chen Wang, Yinghan Sun, Esha Dasgupta, Hua Chen, Ales Leonardis, Wei Zhang, and Hyung Jin Chang. HS-Pose: Hybrid scope feature extraction for category-level object pose estimation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[48] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In CVPR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.206, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[49] Chungang Zhuang, Shaofei Li, and Han Ding. Instance segmentation based 6d pose estimation of industrial objects using point clouds for robotic bin-picking. Robotics and Computer-Integrated Manufacturing, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1189" + } + ] +] \ No newline at end of file diff --git a/2025/UA-Pose_ Uncertainty-Aware 6D Object Pose Estimation and Online Object Completion with Partial References/0d3743ff-3af5-445c-ae4c-ba5a33b9d7b4_origin.pdf b/2025/UA-Pose_ Uncertainty-Aware 6D Object Pose Estimation and Online Object Completion with Partial References/0d3743ff-3af5-445c-ae4c-ba5a33b9d7b4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7d91fc907cfdc43b956cd2d3b9aa5fb88e4cb1ec --- /dev/null +++ b/2025/UA-Pose_ Uncertainty-Aware 6D Object Pose Estimation and Online Object Completion with Partial References/0d3743ff-3af5-445c-ae4c-ba5a33b9d7b4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3cae97d5dc24850ccfff862c2fac9fe02b9c47be1302df40022b963623003f7 +size 5425497 diff --git a/2025/UA-Pose_ Uncertainty-Aware 6D Object Pose Estimation and Online Object Completion with Partial References/full.md b/2025/UA-Pose_ Uncertainty-Aware 6D Object Pose Estimation and Online Object Completion with Partial References/full.md new file mode 100644 index 0000000000000000000000000000000000000000..32b60abff227742de764b96552d7cf1a44331840 --- /dev/null +++ b/2025/UA-Pose_ Uncertainty-Aware 6D Object Pose Estimation and Online Object Completion with Partial References/full.md @@ -0,0 +1,313 @@ +# UA-Pose: Uncertainty-Aware 6D Object Pose Estimation and Online Object Completion with Partial References + +Ming-Feng Li $^{1}$ Xin Yang $^{4}$ Fu-En Wang $^{4}$ Hritam Basak $^{2}$ Yuyin Sun $^{4}$ Shreekant Gayaka $^{4}$ Min Sun $^{3,4}$ Cheng-Hao Kuo $^{4}$ + +$^{1}$ Carnegie Mellon University $^{2}$ Stony Brook University $^{3}$ National Tsing Hua University $^{4}$ Amazon + +# Abstract + +6D object pose estimation has shown strong generalizability to novel objects. However, existing methods often require either a complete, well-reconstructed 3D model or numerous reference images that fully cover the object. Estimating 6D poses from partial references, which capture only fragments of an object's appearance and geometry, remains challenging. To address this, we propose UA-Pose, an uncertainty-aware approach for 6D object pose estimation and online object completion specifically designed for partial references. We assume access to either (1) a limited set of RGBD images with known poses or (2) a single 2D image. For the first case, we initialize a partial object 3D model based on the provided images and poses, while for the second, we use image-to-3D techniques to generate an initial object 3D model. Our method integrates uncertainty into the incomplete 3D model, distinguishing between seen and unseen regions. This uncertainty enables confidence assessment in pose estimation and guides an uncertainty-aware sampling strategy for online object completion, enhancing robustness in pose estimation accuracy and improving object completeness. We evaluate our method on the YCB-Video, YCBIInEOAT, and HO3D datasets, including RGBD sequences of YCB objects manipulated by robots and human hands. Experimental results demonstrate significant performance improvements over existing methods, particularly when object observations are incomplete or partially captured. Project page: https://minfenli.github.io/UA-Pose/ + +# 1. Introduction + +6D object pose estimation, which determines the rigid 6-degree-of-freedom transformation between an object and a camera, is essential for various real-world applications such as robotic manipulation [17, 38, 39, 49] and augmented re + +![](images/37db0d36fb83e091b613238948cfc2a2ea7b152f0be3e6b070408c9a83247ef3.jpg) + +![](images/a205788b8f026eca1d12a5b0f168034c341c780a54780064154ebaa72266780b.jpg) +Figure 1. Overview of UA-Pose. (a) We initialize an incomplete object 3D model $\mathcal{M}$ (Sec. 3.2), which is a hybrid mesh that combines texture, geometry, and uncertainty. (b) Given a sequence of RGBD test images, our method estimates the 6D pose of the object in each test image (Sec. 3.3) based on the incomplete model $\mathcal{M}$ . (c) While more test images are captured, we store them with estimated poses in a memory pool, which will be used for online object completion (Sec. 3.4) to refine the incomplete model $\mathcal{M}$ . + +ality [22, 29]. Recent research introduced model-based and model-free approaches that can generalize to arbitrary novel objects. These methods push the boundaries of pose estimation for real-world applications and can be divided into two scenarios based on the object information available beforehand. Model-based methods [1, 2, 4, 11, 13, 16, 23, 24, 28] require textured 3D CAD models of the object to estimate 6D object poses in 2D images by establishing 2D-3D correspondences. In contrast, model-free methods leverage a large set of posed reference RGB images [19], RGBD images [10, 41], or video sequences [7, 31] to estimate object poses without the need for CAD models. + +To bridge model-based and model-free methods, FoundationPose [41] proposes a unified framework that supports both model-based and model-free setups, using 3D CAD models or posed RGBD images (referred to as references). While effective in generalizing to novel objects, FoundationPose and similar model-free methods [7, 10, 31] still depend on high-quality 3D models or sufficient posed RGBD images that capture the object from sufficient viewpoints. However, assuming the availability of high-quality 3D models is often impractical due to the vast number of objects that remain uncaptured or unmodeled in real-world environments. Likewise, obtaining adequate posed RGBD images is challenging, as real-world settings frequently involve limited viewpoints, occlusions from nearby objects, and complexities in estimating camera poses in dynamic scenes. + +To address these challenges, we present UA-Pose, an uncertainty-aware approach for 6D pose estimation and online object completion as shown in Fig. 1. Our approach is designed to work with partial reference inputs that can include valuable meta-information for robotic tasks, such as grasping patterns, detection markers, or affordance labels. We consider two scenarios: (1) When a limited set of reference RGBD images with known poses is available, we initialize an incomplete 3D model using these images and poses; (2) When only a single unposed RGB image is available, we apply image-to-3D techniques to generate an initial object 3D model for pose estimation. By marking unseen regions as uncertain, we introduce a hybrid representation (Sec. 3.2) that integrates texture, geometry, and uncertainty. The uncertainty is utilized during our proposed uncertainty-aware pose estimation (Sec. 3.3) to provide confidence for estimated poses. Additionally, we propose an uncertainty-aware sampling strategy to select informative images for online object completion, iteratively refining the object 3D model during testing (Sec. 3.4). + +We evaluate our approach on the YCB-Video, YCBI-nEOAT, and HO3D datasets. The results demonstrate significant improvements over existing methods and highlight the applicability of our uncertainty-aware approach in real-world scenarios where fully captured 3D models are often unavailable. To summarize, our contributions are: + +- We propose an uncertainty-aware approach for 6D object pose estimation (Sec. 3.3) and online object completion (Sec. 3.4), which improves both pose accuracy and object completeness when only partial references are available. +- We develop a hybrid representation (Sec. 3.2) incorporating uncertainty. The uncertainty enables confidence assessment in pose estimation and supports an uncertainty-aware sampling strategy for online object completion. +- We demonstrate pose estimation from a single RGB image by leveraging single-image-to-3D methods, which generate an initial model for pose estimation and provide augmented data for online object completion (Sec. 3.5). + +- Our method supports different partial object references as input, which could be associated to meta-information, making it feasible for real-world applications. + +# 2. Related Work + +# 2.1. Model-based Object Pose Estimation + +Traditional model-based pose estimation methods can be divided into two categories: instance-level [8, 9, 12, 25] and category-level [3, 15, 32, 34, 46, 47]. Both types of methods assume that a textured CAD model of the specific object or similar object categories is available during training. However, these approaches are limited to objects seen during training, significantly restricting their applicability in real-world scenarios where many objects remain unseen or unmodeled. To address this limitation, zero-shot methods have been proposed [2, 11, 13, 16, 28], which estimate the pose of novel objects by providing their CAD models at test time. These methods use object CAD models to either render 2D images for feature matching and 2D-3D correspondence estimation or leverage point clouds from CAD models to estimate 3D-3D correspondences. While these approaches improve generalization to unseen objects, they still require access to high-quality 3D models at test time, which is often impractical in real-world settings. + +# 2.2. Model-free Object Pose estimation + +Model-free methods eliminate the need for an explicit textured CAD model. Gen6D [19] introduces a pipeline to estimate object poses from RGB images with known poses. However, due to limited training data, Gen6D faces challenges with out-of-distribution test cases and images affected by occlusions. OnePose [31] and its extension OnePose++ [7] use structure-from-motion techniques to reconstruct 3D point clouds from sequential RGB images and employ pre-trained 2D-3D matching networks to estimate the pose of test images. FS6D [10] employs a transformer-based architecture to extract features from multi-view RGBD images for pose estimation. More recently, FoundationPose [41], trained on a large synthetic dataset, presents a unified framework that supports both model-based and model-free setups. In its model-free setting, FoundationPose requires posed reference images that capture the object from sufficient viewpoints to reconstruct 3D models for robust pose estimation. In general, these model-free pose estimation methods rely on either sequential RGB images with known poses or densely posed RGBD images. However, assuming pre-captured posed RGBD images for unseen objects is impractical for two key reasons: first, real-world applications often involve numerous unseen objects, making it infeasible to pre-capture and collect images for each object; and second, even if sufficient images of the novel object are available, accurately estimating cam + +era poses in dynamic environments is challenging. + +# 2.3. Object Pose Tracking and Reconstruction + +Assuming test images for pose estimation are sequential, 6D object pose tracking methods [36, 40] leverage temporal cues to achieve efficient, smooth, and accurate pose predictions across consecutive frames by estimating relative pose transformations from each frame to the initial frame. For instance, BundleSDF [40] is a 6D pose tracking method that estimates poses from sequential RGBD images while simultaneously training a signed distance field (SDF) [45] representation to maintain consistent pose estimates over time. However, pose tracking methods are constrained to tracking object poses within sequential images from a single video, without support for using reference images for initial pose estimation. In contrast, our method is designed to estimate object poses based on given object references, which may include metadata beneficial for robotic applications, such as grasping patterns, barcodes, or affordance labels. + +# 2.4. Single-Image-to-3D and Pose Estimation + +Recently, single-image-to-3D methods have been developed to generate 3D models from a single image [18, 20, 21, 26, 27, 33, 43]. These methods offer potential for pose estimation when textured CAD models are unavailable. For instance, GigaPose [23] explored using generated models for pose estimation; however, relying on generated 3D models that are not closely similar to real objects can lead to imprecise pose estimates. To address this, we apply a single-image-to-3D approach [43] to synthesize a generated model for early pose estimation. As real RGBD images are gradually collected during testing, they are used to create a refined object model, while the generated model from single-image-to-3D methods is leveraged to produce augmented data to improve object geometry in the uncaptured parts. Incorporating these augmented data into the SDF training yields more complete object models, ultimately improving the robustness and accuracy of pose estimation. + +# 3. Method + +An overview of our method is illustrated in Fig. 1. We begin by initializing an incomplete object 3D model $\mathcal{M}$ using partial references, which can be either: (1) a few reference RGBD images with known poses or (2) a single unposed RGB image of the object. In this work, we propose a hybrid object representation (Sec. 3.2) that integrates texture, geometry, and uncertainty, serving as the incomplete model $\mathcal{M}$ for uncertainty-aware pose estimation (Sec. 3.3). + +Given a sequence of $k$ RGB test images, $\mathcal{I} = \{I_0,I_1,\dots ,I_{k - 1}\}$ , corresponding depth maps, and an initial segmentation mask $m_0$ for the object of interest in the first image $I_0$ , our method estimates the 6D pose of the object $\xi_{i}$ in each subsequent test image $I_{i}$ using + +the proposed uncertainty-aware pose estimation approach (Sec. 3.3), based on the incomplete 3D model $\mathcal{M}$ . During testing, newly captured test RGBD images and their estimated poses will be appended to a memory pool. These images will be used for online object completion (Sec. 3.4) to refine the incomplete object 3D model $\mathcal{M}$ . Notably, when the input reference is a single unposed RGB image, we employ a single-image-to-3D approach [43] to initialize a generated model $\hat{\mathcal{M}}$ to assist in pose estimation and object completion. More details are described in Sec. 3.5. + +# 3.1. Background: Model-free Pose Estimation + +Given RGBD reference images of a novel object and their camera poses, the task of model-free pose estimation is to determine the 6D object poses (i.e., the rigid 6D transformation from the object to the camera) of that novel object in test RGBD images. FoundationPose [41] is a foundational model for 6D pose estimation trained on a large-scale dataset. In the model-free setup, it uses sufficient RGBD reference images (e.g., 16 RGBD images) that fully cover the object to train a Signed Distance Field (SDF) and then extracts a mesh by marching cubes as the 3D object model. Given the 3D object model to render object images, it introduces a two-stage pipeline with two specialized networks: a pose refinement network, which refines pose hypotheses by estimating the relative pose between the test image and rendered object images, and a pose selection network, which identifies the best pose by comparing the test image with rendered images based on these refined poses. However, RGBD reference images that fully represent the object's appearance and geometry are usually unavailable in real-world scenarios. As a result, model-free methods may struggle to provide accurate pose estimates when only incomplete object models or sparse observations are available. + +# 3.2. Hybrid Object Representation + +To address the limitations, we introduce a hybrid object representation that integrates the object's texture, geometry, and uncertainty, as shown in Fig. 2. This uncertainty marks the seen and unseen regions of an incomplete 3D object model and provides confidence in pose estimations, helping to filter out less reliable results. Given a set of RGBD images, object masks, and their camera poses, we train an object-centric neural Signed Distance Field (SDF) [45] to learn the partial object's appearance and geometry. + +We then extract a 3D mesh using the marching cubes algorithm to explicitly model uncertainty. Inspired by Visual Hull [14], we define object uncertainty based on visibility inferred from 2D object masks. By projecting these masks onto the mesh $\mathcal{M}$ , we label regions of $\mathcal{M}$ seen in the reference images as "certain" and those unseen as "uncertain", as shown in Fig. 2. This approach allows pose estimation to account for both the seen and unseen parts of the object. + +![](images/8f86fd6dc3858ab38ac2f5fa4a989cda6c02879314baf8645078b91afd2704f4.jpg) +Figure 2. Hybrid Object Representation Modeling. We propose a hybrid object representation (Sec. 3.2) that integrates the object's texture, geometry, and uncertainty. (a) First, a neural SDF is trained and extracted as a mesh representing the object's appearance and geometry (Sec. 3.2.1). (b) Then, we check the visibility of each mesh vertex from the viewpoint of each reference image to create the uncertainty map (Sec. 3.2.2), which reflects the seen and unseen regions of an incomplete 3D object model. + +# 3.2.1 Neural SDF Training and Mesh Extraction + +We follow SDF training methods similar to [40, 41] to train a neural SDF for object modeling. The SDF defines the object surface as the set of 3D points: + +$$ +S = \{x \in \mathbb {R} ^ {3} \mid \Omega (x) = 0 \}. \tag {1} +$$ + +where $\Omega : \mathbb{R}^3 \to \mathbb{R}$ is the SDF geometry function that maps each 3D point $x$ to a signed distance value $s$ and $s = 0$ indicates the object's surface. Given RGBD images, object masks, and their camera poses, we use a neural SDF to reconstruct the object's appearance and geometry, represented by two networks [45] as shown in Fig. 2. The geometric network $\Omega : x \mapsto s$ takes a 3D point $x \in \mathbb{R}^3$ as input and outputs a signed distance value $s \in \mathbb{R}$ . The appearance network $\Phi : (f_{\Omega(x)}, d) \mapsto c$ takes an intermediate feature vector $f_{\Omega(x)}$ from the geometric network along with a view direction $d \in \mathbb{R}^3$ and outputs the color $c \in \mathbb{R}^3$ for the point. + +**Rendering.** Given the object pose $\xi$ , we render an image by casting rays through each pixel. Along each ray $r$ , 3D points are sampled at multiple positions as follows: + +$$ +x _ {i} (r) = o (r) + t _ {i} d (r), \tag {2} +$$ + +where $o(r)$ denotes the ray origin and $d(r)$ denotes the ray direction from the camera to the object determined by the object pose $\xi$ . The parameter $t_i \in \mathbb{R}_+$ determines the point position along the ray. Following the approach in [40], the color $c$ of a ray $r$ is computed by integrating sampled points near the object surface through volumetric rendering: + +$$ +c (r) = \int_ {z (r) - \lambda} ^ {z (r) + 0. 5 \lambda} w \left(x _ {i}\right) \Phi \left(f _ {\Omega \left(x _ {i}\right)}, d \left(x _ {i}\right)\right) d t, \tag {3} +$$ + +$$ +w \left(x _ {i}\right) = \frac {1}{1 + e ^ {- \alpha \Omega \left(x _ {i}\right)}} \frac {1}{1 + e ^ {\alpha \Omega \left(x _ {i}\right)}}, \tag {4} +$$ + +where $w(x_{i})$ is a bell-shaped probability density function [35] based on the distance of each point from the implicit object surface and $\alpha$ adjusts the sharpness of the density distribution. $z(r)$ represents the ray's depth from the + +depth image and $\lambda$ is the truncation distance. During volumetric rendering, $z(r)$ and $\lambda$ are used to exclude 3D points far from the object surface, enhancing rendering efficiency and preventing the SDF from overfitting to empty regions. + +During training, we supervise this rendering by comparing it to the reference RGB images with the color loss: + +$$ +\mathcal {L} _ {c} = \frac {1}{| \mathcal {R} |} \sum_ {r \in \mathcal {R}} \| c (r) - \bar {c} (r) \| _ {2}, \tag {5} +$$ + +where $\bar{c}(r)$ is the ground-truth color at the pixel through which the ray $r$ passes. For geometry learning, we leverage depth information from the provided depth maps to supervise the neural (SDF) through two losses [40]: the near-surface loss $\mathcal{L}_s$ and the empty space loss $\mathcal{L}_e$ . $\mathcal{L}_s$ encourage the SDF surface to align closely with depth values and $\mathcal{L}_e$ enforces that points along the ray between the camera and the surface remain empty. Additionally, we apply an eikonal regularization loss $\mathcal{L}_{eik}$ [5] to the near-surface SDF to enforce smoothness. The total loss is represented as: + +$$ +\mathcal {L} = w _ {c} \mathcal {L} _ {c} + w _ {e} \mathcal {L} _ {e} + w _ {s} \mathcal {L} _ {s} + w _ {e i k} \mathcal {L} _ {e i k}, \tag {6} +$$ + +where $w_{c}, w_{e}, w_{s}$ , and $w_{eik}$ are weighting factors. + +To explicitly model uncertainty, we convert the implicit surface defined by the neural SDF into an explicit 3D mesh $\mathcal{E} = (V,C,F)$ with the vertices $V$ , vertex colors $C$ , and faces $F$ . To achieve this, we extract a mesh from the trained neural SDF representation using the marching cubes algorithm, which guarantees a closed-surface mesh including vertices from any viewing angle. This allows us to model the visibility of the object from any viewing angle even if some parts are not visible in the reference images. + +Each vertex $v_{i} \in V$ is assigned a color $c_{i} \in C$ , which is computed using the appearance function $\varPhi$ , which maps the intermediate feature vector $f_{\Omega(v_{i})}$ from the geometry network and a view direction $d \in \mathbb{R}^3$ to the RGB color: + +$$ +c _ {i} = \Phi \left(f _ {\Omega \left(v _ {i}\right)}, d\right). \tag {7} +$$ + +# 3.2.2 Uncertainty Modeling + +To model the uncertainty map that represents whether a vertex is visible from the viewpoints of input references, for each vertex $v_{i} \in V$ , we check its visibility from the viewpoint of each reference image $\bar{I}_j \in \tilde{\mathcal{I}}$ through the mesh rasterization process, which rasterizes 3D vertices onto 2D image pixels. We define an uncertainty score $u(v_{i})$ based on whether $v_{i}$ is visible from any of the viewpoints in the reference images $\bar{I}_j$ . This uncertainty score $u(v_{i}) \in \{0,1\}$ labels vertices on the mesh as "certain" if they are seen in any reference image $(u(v_{i}) = 0)$ and as "uncertain" if they remain unseen $(u(v_{i}) = 1)$ . The resulting uncertainty map $\mathcal{U} = \{u(v_i) \mid v_i \in V\}$ is incorporated into our hybrid object + +![](images/4ef30e617b60be67a5e5b93f5e2b6e395d349da375deb031384565aac2fc26c9.jpg) +Figure 3. Pipeline for Uncertainty-Aware Pose Estimation. Through uncertainty modeling (Sec. 3.2.2), we propose an uncertainty-aware pipeline to assess the confidence of estimated poses (Sec. 3.3). For each estimated pose, we calculate its seen IoU, which measures the overlap between the seen regions of the 3D model and the 2D object mask for the test image. When seen IoU is high, the pose is deemed confident, and both the pose and its RGBD image are added to the memory pool for object completion. Conversely, if the estimated pose has a low seen IoU, it is considered unreliable. In such cases, we employ an uncertainty-aware image sampling strategy to select images from the memory pool for online object completion (Sec. 3.4), thus enhancing pose estimation by refining the object model during testing. + +representation $\mathcal{M} = (\mathcal{E},\mathcal{U})$ , allowing the pose estimation process to distinguish between seen and unseen regions. + +Guidance from Uncertainty Map. Given an object pose $\xi \in \mathbb{R}^{3\times 4}$ , the model $\mathcal{M}$ can be used to render an RGB image $I_{\xi}^{\mathrm{rend}}\in \mathbb{R}^{H\times W\times 3}$ , a depth map $D_{\xi}^{\mathrm{rend}}\in \mathbb{R}^{H\times W}$ , a rendered object mask $m_{\xi}^{\mathrm{rend}}\in \{0,1\}^{H\times W}$ , and an uncertainty map $U_{\xi}^{\mathrm{rend}}\in \{0,1\}^{H\times W}$ at a resolution of $H\times W$ . To evaluate the validity and confidence of an estimated pose, we introduce two metrics: uncertainty rate and seen IoU. The uncertainty rate measures the proportion of the uncertain pixels over the total pixels within the rendered object mask: + +$$ +\frac {\sum \left(U _ {\xi} ^ {\text {r e n d}} \odot m _ {\xi} ^ {\text {r e n d}}\right)}{\sum \left(m _ {\xi} ^ {\text {r e n d}}\right)}, \tag {8} +$$ + +where $\odot$ denotes element-wise logical multiplication. The seen IoU measures the overlap between seen (certain) regions of the rendered object mask and the object mask in the test image, which is defined as: + +$$ +\mathrm {I o U} \left(\neg U _ {\xi} ^ {\text {r e n d}} \odot m _ {\xi} ^ {\text {r e n d}}, m ^ {\text {t e s t}}\right), \tag {9} +$$ + +where $m^{\mathrm{test}}$ is the object mask from the test image, $\neg U_{\xi}^{\mathrm{rend}}\odot$ $m_{\xi}^{\mathrm{rend}}$ identifies "certain" pixels within the rendered object mask, and IoU represents the Intersection over Union function. Since pose estimation is more robust when based on seen regions of the 3D model, a high seen IoU indicates high confidence in the pose estimation, whereas a low seen IoU suggests that the estimation is not reliable. + +# 3.3. Uncentainty-aware Pose Estimation + +After modeling the uncertainty, the hybrid object representation could be used in our uncertainty-aware pose estimation pipeline, as shown in Fig. 3. In this work, we leverage the pose refinement and the pose selection modules of [41] and assume that the test images are sequential. To obtain per-frame 2D object masks in test images for calculating the uncertainty rate and seen IoU, given first frame object mask $m_0$ , the object masks of the following + +frames $\{m_1, m_2, \ldots, m_{k-1}\}$ is determined by off-the-shelf segmentation methods [44]. + +Pose Initialization. Given an RGBD video, object masks for each frame, and an incomplete object model $\mathcal{M}$ , we begin by estimating the object pose for the first frame by generating multiple pose hypotheses. For each hypothesis, the translation is initialized using the 3D point projected from the center of the 2D object mask, determined by the median depth value within this region. For rotation, we uniformly sample $N_{v}$ viewpoints from an icosphere centered on the object and orient the camera to face the object's center. Subsequently, $N_{in}$ in-plane rotations are sampled and applied to each viewpoint, yielding $N_{v} \cdot N_{in}$ hypothesized object poses. Each object pose can be represented as $[\pmb{R}|\pmb{t}] \in \mathbb{S}\mathbb{E}(3)$ , where $\pmb{R} \in \mathbb{S}\mathbb{O}(3)$ is the rotation and $\pmb{t} \in \mathbb{R}^3$ is the translation. + +Pose Refinement. After generating $N_v \cdot N_{in}$ pose hypotheses, we apply the refinement module [41] to improve pose accuracy. Specifically, given a rendered RGBD image of the object from each pose hypothesis and the test RGBD image of the target object, the network outputs a pose update $\Delta R \in \mathbb{SO}(3)$ and $\Delta t \in \mathbb{R}^3$ to refine the pose hypothesis, aligning it more closely with the observed object pose in the test image. Each refined pose is represented as $[R^{+}|t^{+}] \in \mathbb{SE}(3)$ , where $t^{+} = t + \Delta t$ and $R^{+} = \Delta R \otimes R$ . This refinement process could be iteratively repeated. In this work, each pose is refined five times for pose estimation on the first frame and twice for subsequent frames. + +Pose Selection. Given a set of refined pose hypotheses, we apply our proposed metrics (Sec. 3.2.2) to filter out unreliable poses: the uncertainty rate, which measures the proportion of uncertain pixels within the rendered object mask, and the seen IoU, which quantifies the overlap between previously seen (certain) regions of the object and the object mask in the test image. A pose hypothesis is discarded when its uncertainty rate is above the threshold $T_{u}$ or its seen IoU is below the threshold $T_{s}$ , ensuring that only reliable hypotheses that adequately cover seen object regions are con- + +sidered. At last, we use the pose selection module [41] to identify the most accurate pose as the final estimate. + +Pose Tracking. Given the sequential nature of RGBD video frames, object poses in adjacent frames are typically similar. For each frame following the initial one, we take the estimated pose from the previous frame as the sole pose hypothesis and refine it using the pose refinement module to generate the estimated pose for the current frame. This iterative process continues across the video sequence, leveraging temporal cues for stable and smooth pose estimation. + +Memory Pool. To enhance pose estimation by completing the object model during testing, we select test RGBD images with their estimated poses to incrementally refine the object representation. For efficiency, we introduce a memory pool $\mathcal{P} = \{\xi_0,\xi_1,\xi_2,\dots ,\xi_{|\mathcal{P}| - 1}\}$ that stores the most informative $|\mathcal{P}|$ viewpoints to maintain a compact yet diverse set of test image frames for object completion, where $\xi_{i}$ is the estimated pose corresponding to a test image $I_{i}$ . The first frame $\xi_{i}$ is automatically added to $\mathcal{P}$ , establishing the canonical coordinate system for the novel object. Subsequent frames are added when their viewpoints contribute to the multi-view diversity in the pool. Specifically, before adding a new frame, we calculate the rotational geodesic distance between the new frame and the last frame added to the pool. A frame is included in $\mathcal{P}$ only if this geodesic distance exceeds a threshold $T_{geo}$ , ensuring that each frame offers a unique viewpoint to enhance model completeness while keeping $\mathcal{P}$ compact. Note that not all estimated poses are reliable, adding inaccurate results to the memory pool may cause noisy SDF training during object completion. To mitigate this, we introduce an image-filtering strategy to filter out low-confidence estimations. As illustrated in Fig. 3, when a pose is estimated, its confidence is evaluated using our proposed metric, seen IoU, which measures the overlap between the seen regions of the 3D model and the 2D object mask in the test image. Since a high seen IoU indicates greater confidence in an estimated pose, frames are added to the memory pool only if their seen IoU exceeds a threshold $T_{conf}$ , ensuring that low-confidence poses are excluded. + +# 3.4. Uncentainty-aware Object Completion + +As shown in Fig. 3, online object completion is triggered when the seen IoU of the current test frame falls below a threshold $T_{complete}$ , indicating that the current pose estimation may no longer be reliable. As new frames are continuously added, the memory pool $\mathcal{P}$ can grow excessively large, which may reduce SDF training efficiency. To manage this, we perform uncertainty-aware sampling to select the $K$ most informative frames for object completion. + +Uncertainty-aware Sampling. Early in the video sequence, when $|\mathcal{P}| \leq K$ , all frames in the pool are included without selection. However, once the pool size exceeds $K$ we apply an uncertainty-aware image selection strategy to + +![](images/5f5e32d554693063e3a24cf45252f3a8a34cb877e17513714866070a0db819e4.jpg) +Figure 4. Pipeline for leveraging generated models. In Sec. 3.5, the initial model generated by image-to-3D techniques is used to estimate poses for the first few frames. As more frames are captured, a refined object model is trained to replace the initial generated model, resulting in a more complete representation. To further leverage the generated 3D model, we render RGBD images from the image-to-3D generated model as augmented data for SDF training. This supervision enhances the object geometry in unseen regions, providing more complete information for pose estimation. + +![](images/8d659f63bc6c501eddc73a3d3b7a0a4baf595c4464b985a4a8ad6214c95aea94.jpg) + +maximize the coverage of unseen object regions. This process selects a subset, $\mathcal{P}' \subset \mathcal{P}$ , that reveals the most unseen areas of the object. The selection process begins by including the initial and most recent frames in the pool to establish the starting and ending viewpoints. Next, an iterative strategy is employed: at each iteration, we select the pose that reveals the largest unseen region. This continues until $K$ poses are selected. + +# 3.5. Leveraging Image-to-3D Generation + +Given a single unposed RGB image, we employ image-to-3D techniques to generate an initial object 3D model $\hat{\mathcal{M}}$ for pose estimation. The viewpoint associated with the initial RGB image is labeled as "certain," while other areas that are inferred by the image-to-3D method are marked "uncertain". The generated model is then rescaled to align with the object size in the test images. Specifically, we use the depth map and object mask from the first frame of the test image to approximate a rough model size. We then sample slightly larger and smaller models, generate pose hypotheses, and employ the pose selection module [41] to select the optimal pose hypothesis and corresponding model size. + +Refer to Fig. 4, the initial generated model is used to estimate poses for the first few test frames until the rotational geodesic distance between the initial pose and a new estimated pose exceeds a threshold $T_{gen}$ . As more frames are captured, a refined object model is trained to replace the initial generated model, resulting in a more complete representation. To further leverage the generated 3D model, we render RGBD images from the image-to-3D generated model as augmented data for SDF training. This supervision enhances the object geometry in unseen regions, providing more complete information for pose estimation. See the supplementary materials for additional details. + +# 4. Experiment + +# 4.1. Datasets + +YCB-Video [42]: The YCB-Video dataset contains RGBD video sequences of everyday objects from the YCB objects, widely used for benchmarking 6D object pose estimation. + +YCBCInEOAT [37]: YCBCInEOAT extends the YCB-Video dataset with RGBD sequences of YCB objects manipulated by a dual-arm robot. This dataset focuses on real-world scenarios involving robot-object interactions. + +HO3D [6]: The HO3D (Hand-Object 3D) dataset contains RGBD sequences of human hands interacting with objects, capturing complex hand-object interactions and occlusions. + +# 4.2. Experimental setup + +Baselines. We evaluate our method against state-of-the-art RGBD model-free 6D pose estimation methods [10, 30, 41] under two scenarios where partial object references are given as (1) A limited set of reference RGBD images (2 views) that include known poses. (2) A single unposed RGB object image. In scenario (2), we leverage image-to-3D techniques [43] to generate an initial object 3D model and iteratively estimate the object pose during testing. + +Metrics. We consider ADD and ADD-S metrics for 6D pose estimation, which measure the accuracy of estimated poses. The Area Under the Curve (AUC) is calculated and reported for both the ADD and ADD-S metrics, following the protocols in [40-42]. Furthermore, to demonstrate the effectiveness of our method for online object completion, we report the Chamfer Distance (CD), which measures the average distance between points of the reconstructed model and the ground-truth model. + +# 4.3. Comparison Results on YCB-Video + +We first evaluate UA-Pose on the YCB-Video [42] dataset, a widely used benchmark for 6D pose estimation, and compare our approach with state-of-the-art model-free methods [10, 30, 41] with a limited set of reference images. + +As presented in Tab. 1, our method significantly outperforms FoundationPose [41] with only 2 reference images in all metrics. Moreover, both our method and FoundationPose outperforms LoFTR [30] and FS6D-DPM [10] with 16 reference images with a large margin. This shows that FoundationPose is the strongest baseline for our further analysis. When only a single unposed RGB image of the object is available, we apply a single-image-to-3D approach [43] to generate an initial 3D model for pose estimation. Experimental results, shown in Tab. 1, indicate that directly using the generated model with FoundationPose may not yield optimal results, as these models may not accurately represent real objects. In contrast, UA-Pose uses the generated model only for initial pose estimation and for rendering RGBD images as augmented data to aid object + +Table 1. Quantitative results on the YCB-Video dataset. "Comp." indicates whether the method includes online completion, and "-" denotes that the method does not reconstruct object shapes. Results for [10, 30] are adopted from [41]. Experiments for "single RGB + Image-to-3D" are conducted on a subset of the YCB-Video dataset. See supplementary materials for details. + +
InputComp.Mean
ADDADD-SCD
LoFTR [30]16 referencesNo26.252.5-
FS6D-DPM [10]16 referencesNo42.188.4-
FoundationPose [41]2 referencesNo87.494.30.57
Ours2 referencesYes92.896.50.53
FoundationPose [41]single RGB + Image-to-3DNo88.993.60.67
Ourssingle RGB + Image-to-3DYes93.296.90.62
+ +Table 2. Quantitative results on the YCBIInEOAT dataset. "Comp." indicates that the method performs online completion. + +
InputComp.Mean
ADDADD-SCD
FoundationPose [41]2 referencesNo68.5284.800.60
Ours2 referencesYes89.9994.350.57
FoundationPose [41]single RGB + Image-to-3DNo83.3990.660.61
Ourssingle RGB + Image-to-3DYes89.2493.990.60
+ +completion, as described in Sec. 3.5. This approach reconstructs a model more closely aligned with the real object, leading to improved pose estimation accuracy. + +# 4.4. Comparison Results on YCBIInEOAT + +We evaluated UA-Pose on the YCBIInEOAT [37] dataset to assess its performance in challenging robotic scenarios in real-world where objects are manipulated by robotic arms. + +In Table 2, when 2 reference images are provided, UA-Pose demonstrates superior performance compared to FoundationPose, highlighting the effectiveness of our uncertainty-aware mechanisms and online object completion for robust pose estimation in robotic scenarios. As illustrated in Fig. 5, simply using the incomplete model generated from the partial references often leads to inaccurate estimations. In contrast, our method achieves robust pose estimation with uncertainty-aware mechanisms and online object completion. When only a single unposed RGB image of the object is available, UA-Pose also consistently surpasses FoundationPose across all metrics. This demonstrates the flexibility of UA-Pose to utilize diverse partial references, which could carry useful meta-information. + +# 4.5. Comparison Results on HO3D + +We evaluated UA-Pose on the HO3D dataset to assess its effectiveness in complex hand-object interaction scenarios where objects are frequently rotated and occluded. + +As shown in Tab. 3, when 2 reference images are provided, FoundationPose fails to make accurate estimations when objects in the test video exhibit frequent and strong rotations. In contrast, UA-Pose accurately estimates object poses under the challenging hand-object interaction scenarios. As for utilizing the generated models from the single-image-to-3D approach for pose estimation, UA-Pose + +![](images/3884863dd735546e24505d63eabfee59bbea2529a50d1f2d0d1d39439655d070.jpg) +Figure 5. Qualitative comparison on the YCBIInEOAT. Object image (left) and depth map (right) pairs are rendered based on the estimated object poses using the incomplete model. Without uncertainty-aware mechanisms and Online Object Completion (referred to as "w/o OOC"), the incomplete model often leads to inaccurate estimations. In contrast, our method (referred to as "w/ OOC") achieves robust pose estimation by incorporating uncertainty-aware mechanisms and online object completion. + +achieves more accurate estimations than FoundationPose. Note that even if the generated models in Tab. 3 show lower Chamfer Distance (CD) than refined models obtained from object completion, directly using generated models for FoundationPose results in considerably lower ADD and ADD-S. This indicates that even though generated models have good geometry (low CD), geometry alone cannot accurately represent real objects for pose estimation because the textures of generated models are not good enough in unseen areas that are not covered by the initial partial references. + +# 4.6. Ablation Study + +Key Components. Tab. 4 presents an ablation study for key components of UA-Pose on the YCBCInEOAT dataset. All experiments start with the initialized model based on two reference images and are evaluated using ADD, ADD-S, and CD. Additionally, we report $N_{rebuild}$ , the total number of times object completion is applied across test sequences to refine the incomplete model $\mathcal{M}$ . The "w/o object completion" setup disables online object completion, illustrating its importance for achieving accurate pose estimation. In "w/o uncertainty-aware object completion", object completion is applied every time a new test image is added to the memory pool, instead of completing the model only when the seen IoU threshold is below $T_{complete}$ . This leads to poorer performance and a higher $N_{rebuild}$ count, increasing unnecessary computation for object completion. The "w/o image-filtering strategy" experiment omits the image-filtering strategy (Sec. 3.3) for excluding unreliable poses, allowing low-confidence poses to be added to the memory pool and resulting in decreased pose accuracy. In "w/o uncertainty-aware sampling", we use the geodesic distance-based sampling strategy from [40] instead of our proposed uncertainty-aware sampling strategy (Sec. 3.4) to sample images for object completion. The results demonstrate our proposed sampling method is better at selecting infor + +Table 3. Quantitative results on the HO3D dataset. "Comp." indicates that the method performs online completion. + +
InputComp.Mean
ADDADD-SCD
FoundationPose [41]2 referencesNo45.6761.570.78
Ours2 referencesYes91.5195.560.69
FoundationPose [41]single RGB + Image-to-3DNo72.0687.870.76
Ourssingle RGB + Image-to-3DYes83.2393.590.88
+ +Table 4. Ablation study. We demonstrate the importance of UA-Pose's key components on the YCBCInEOAT dataset. + +
MeanNrebuild
ADDADDSCD
w/o object completion68.5284.800.60-
w/o uncertainty-aware object completion85.8392.390.72181
w/o image-filtering strategy87.7093.370.7187
w/o uncertainty-aware sampling87.8893.390.6859
Ours (full)89.9994.350.5758
+ +Table 5. Additional study. We compare our method with FoundationPose [41] and pose tracking methods [36, 40] on the YCBI-nEOAT dataset. "Comp." means the method performs online completion, and "\*" denotes using cropped ground-truth meshes [40], which exclude invisible parts in the test sequence, to calculate CD. + +
InputComp.Mean
ADDADD-SCD
BundleTrack [36]first-frame RGBDNo87.3492.532.81*
BundleSDF [40]first-frame RGBDYes86.9593.771.16*
FoundationPose [41]first-frame RGBDNo74.3484.581.12*
Oursfirst-frame RGBDYes88.3893.820.75*
+ +mative images for accurate pose estimation. Finally, the "Ours (full)" setup that incorporates all our key components achieves the best performance across ADD, ADD-S, CD, and the fewest $N_{\text{rebuild}}$ for object completion. + +Additional Study. When no external reference images are available, we demonstrate that we could initialize the object model for pose estimation using the first frame of the test RGBD image sequence. Besides the strongest model-free pose estimation baseline [41], we also compare our method with strong model-free pose tracking approaches [36, 40, 48] on the YCBIInEOAT dataset. Refer to Tab. 5, our method achieves the highest performance across ADD, ADD-S, and CD metrics. Note that unlike pose-tracking methods which are limited to estimating relative poses from sequential frames within the same RGBD video, our method can incorporate object reference images as input for useful meta-information. + +# 5. Conclusion + +We introduced UA-Pose, an uncertainty-aware approach for 6D object pose estimation that addresses the limitations of previous model-free methods in the scenarios in which input object references are incomplete or partially captured. Incorporating mechanisms to handle uncertainty and leverage online object completion to complete object models, our method demonstrates substantial performance improvements on the YCB-Video, YCBIInEOAT, and HO3D datasets. We demonstrate UA-Pose is a robust solution for practical real-world applications where full 3D object models and extensive image references are usually unavailable. + +# References + +[1] Philipp Ausserlechner, David Haberger, Stefan Thalhammer, Jean-Baptiste Weibel, and Markus Vincze. Zs6d: Zero-shot 6d object pose estimation using vision transformers. ICRA, 2024. 1 +[2] Andrea Caraffa, Davide Boscaini, Amir Hamza, and Fabio Poiesi. Freeze: Training-free zero-shot 6d pose estimation with geometric and vision foundation models. ECCV, 2024. 1, 2 +[3] Dengsheng Chen, Jun Li, Zheng Wang, and Kai Xu. Learning canonical shape space for category-level 6D object pose and size estimation. In CVPR, 2020. 2 +[4] Jianqiu Chen, Zikun Zhou, Mingshan Sun, Rui Zhao, Liwei Wu, Tianpeng Bao, and Zhenyu He. Zeropose: Cadprompted zero-shot object 6d pose estimation in cluttered scenes. TCSVT, 2024. 1 +[5] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, 2020. 4 +[6] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In CVPR, 2020. 7 +[7] Xingyi He, Jiaming Sun, Yuang Wang, Di Huang, Hujun Bao, and Xiaowei Zhou. OnePose++: Keypoint-free one-shot object pose estimation without CAD models. NeurIPS, 2022. 1, 2 +[8] Yisheng He, Wei Sun, Haibin Huang, Jianran Liu, Haoqiang Fan, and Jian Sun. PVN3D: A deep point-wise 3D keypoints voting network for 6DoF pose estimation. In CVPR, 2020. 2 +[9] Yisheng He, Haibin Huang, Haoqiang Fan, Qifeng Chen, and Jian Sun. FFB6D: A full flow bidirectional fusion network for 6D pose estimation. In CVPR, 2021. 2 +[10] Yisheng He, Yao Wang, Haoqiang Fan, Jian Sun, and Qifeng Chen. Fs6d: Few-shot 6d pose estimation of novel objects. In CVPR, 2022. 1, 2, 7 +[11] Junwen Huang, Hao Yu, Kuan-Ting Yu, Nassir Navab, Slobodan Ilic, and Benjamin Busam. Matchu: Matching unseen objects for 6d pose estimation from rgb-d images. In CVPR, 2024. 1, 2 +[12] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. CosyPose: Consistent multi-view multi-object 6D pose estimation. In ECCV, 2020. 2 +[13] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. MegaPose: 6D pose estimation of novel objects via render & compare. In CoRL, 2022. 1, 2 +[14] Aldo Laurentini. The visual hull concept for silhouette-based image understanding. IEEE TPAMI, 1994. 3 +[15] Taeyeop Lee, Jonathan Tremblay, Valts Blukis, Bowen Wen, Byeong-Uk Lee, Inkyu Shin, Stan Birchfield, In So Kweon, and Kuk-Jin Yoon. TTA-COPE: Test-time adaptation for category-level object pose estimation. In CVPR, 2023. 2 +[16] Jiehong Lin, Lihua Liu, Dekun Lu, and Kui Jia. Sam-6d: Segment anything model meets zero-shot 6d object pose estimation. In CVPR, 2024. 1, 2 + +[17] Jian Liu, Wei Sun, Chongpei Liu, Xing Zhang, and Qiang Fu. Robotic continuous grasping system by shape transformer-guided multi-object category-level 6d pose estimation. IEEE Transactions on Industrial Informatics, 2023. 1 +[18] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In ICCV, 2023. 3 +[19] Yuan Liu, Yilin Wen, Sida Peng, Cheng Lin, Xiaoxiao Long, Taku Komura, and Wenping Wang. Gen6D: Generalizable model-free 6-DoF object pose estimation from RGB images. ECCV, 2022. 1, 2 +[20] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Learning to generate multiview-consistent images from a single-view image. ICLR, 2024. 3 +[21] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, and Wenping Wang. Wonder3d: Single image to 3d using cross-domain diffusion. CVPR, 2024. 3 +[22] Eric Marchand, Hideaki Uchiyama, and Fabien Spindler. Pose estimation for augmented reality: A hands-on survey. TVCG, 2015. 1 +[23] Van Nguyen Nguyen, Thibault Groueix, Mathieu Salzmann, and Vincent Lepetit. Gigapose: Fast and robust novel object pose estimation via one correspondence. CVPR, 2024. 1, 3 +[24] Evin Pinar Ornek, Yann Labbe, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, and Tomas Hodan. Found- pose: Unseen object pose estimation with foundation features. ECCV, 2024. 1 +[25] Kiru Park, Timothy Patten, and Markus Vincze. Pix2Pose: Pixel-wise coordinate regression of objects for 6D pose estimation. In ICCV, 2019. 2 +[26] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, and Bernard Ghanem. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. ICLR, 2024. 3 +[27] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 3 +[28] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. OSOP: A multi-stage one shot object pose estimation framework. In CVPR, 2022. 1, 2 +[29] Yongzhi Su, Jason Rambach, Nareg Minaskan, Paul Lesur, Alain Pagani, and Didier Stricker. Deep multi-state object pose estimation for augmented reality assembly. In ISMAR-Adjunct, 2019. 1 +[30] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In CVPR, 2021. 7 +[31] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. OnePose: One-shot object pose estimation without CAD models. In CVPR, 2022. 1, 2 + +[32] Meng Tian, Marcelo H Ang, and Gim Hee Lee. Shape prior deformation for categorical 6D object pose and size estimation. In ECCV, 2020. 2 +[33] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In ECCV, 2025. 3 +[34] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6D object pose and size estimation. In CVPR, 2019. 2 +[35] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 4 +[36] Bowen Wen and Kostas Bekris. Bundletrack: 6d pose tracking for novel objects without instance or category-level 3d models. In IROS, 2021. 3, 8 +[37] Bowen Wen, Chaitanya Mitash, Baozhang Ren, and Kostas E Bekris. se (3)-tracknet: Data-driven 6d pose tracking by calibrating image residuals in synthetic domains. In IROS, 2020. 7 +[38] Bowen Wen, Wenzhao Lian, Kostas Bekris, and Stefan Schaal. CatGrasp: Learning category-level task-relevant grasping in clutter from simulation. In ICRA, 2022. 1 +[39] Bowen Wen, Wenzhao Lian, Kostas Bekris, and Stefan Schaal. You only demonstrate once: Category-level manipulation from single visual demonstration. RSS, 2022. 1 +[40] Bowen Wen, Jonathan Tremblay, Valts Blukis, Stephen Tyree, Thomas Müller, Alex Evans, Dieter Fox, Jan Kautz, and Stan Birchfield. Bundlesdf: Neural 6-dof tracking and 3d reconstruction of unknown objects. In CVPR, 2023. 3, 4, 7, 8 +[41] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. Foundationpose: Unified 6d pose estimation and tracking of novel objects. CVPR, 2024. 1, 2, 3, 4, 5, 6, 7, 8 +[42] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. RSS, 2018. 7 +[43] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 3, 7 +[44] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 5 +[45] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. NeurIPS, 2020. 3, 4 +[46] Ruida Zhang, Yan Di, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. SSP-Pose: Symmetry-aware shape prior deformation for direct category-level object pose estimation. In IROS, 2022. 2 + +[47] Linfang Zheng, Chen Wang, Yinghan Sun, Esha Dasgupta, Hua Chen, Ales Leonardis, Wei Zhang, and Hyung Jin Chang. HS-Pose: Hybrid scope feature extraction for category-level object pose estimation. In CVPR, 2023. 2 +[48] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In CVPR, 2022. 8 +[49] Chungang Zhuang, Shaofei Li, and Han Ding. Instance segmentation based 6d pose estimation of industrial objects using point clouds for robotic bin-picking. Robotics and Computer-Integrated Manufacturing, 2023. 1 \ No newline at end of file diff --git a/2025/UA-Pose_ Uncertainty-Aware 6D Object Pose Estimation and Online Object Completion with Partial References/images.zip b/2025/UA-Pose_ Uncertainty-Aware 6D Object Pose Estimation and Online Object Completion with Partial References/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d4c5fa2e2b4874f0334eac59e1040588d76d60f1 --- /dev/null +++ b/2025/UA-Pose_ Uncertainty-Aware 6D Object Pose Estimation and Online Object Completion with Partial References/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd95e6da805f3daaf98651304c00dc17fd7a264d6951336de2ff625ccfae5e49 +size 365084 diff --git a/2025/UA-Pose_ Uncertainty-Aware 6D Object Pose Estimation and Online Object Completion with Partial References/layout.json b/2025/UA-Pose_ Uncertainty-Aware 6D Object Pose Estimation and Online Object Completion with Partial References/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b23b2fa8d6cc5d66ad4951ffcbf53a5e0d51cecf --- /dev/null +++ b/2025/UA-Pose_ Uncertainty-Aware 6D Object Pose Estimation and Online Object Completion with Partial References/layout.json @@ -0,0 +1,9112 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 102, + 531, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 102, + 531, + 140 + ], + "spans": [ + { + "bbox": [ + 64, + 102, + 531, + 140 + ], + "type": "text", + "content": "UA-Pose: Uncertainty-Aware 6D Object Pose Estimation and Online Object Completion with Partial References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 160, + 465, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 160, + 465, + 190 + ], + "spans": [ + { + "bbox": [ + 135, + 160, + 465, + 190 + ], + "type": "text", + "content": "Ming-Feng Li" + }, + { + "bbox": [ + 135, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 160, + 465, + 190 + ], + "type": "text", + "content": " Xin Yang" + }, + { + "bbox": [ + 135, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 135, + 160, + 465, + 190 + ], + "type": "text", + "content": " Fu-En Wang" + }, + { + "bbox": [ + 135, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 135, + 160, + 465, + 190 + ], + "type": "text", + "content": " Hritam Basak" + }, + { + "bbox": [ + 135, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 160, + 465, + 190 + ], + "type": "text", + "content": " Yuyin Sun" + }, + { + "bbox": [ + 135, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 135, + 160, + 465, + 190 + ], + "type": "text", + "content": " Shreekant Gayaka" + }, + { + "bbox": [ + 135, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 135, + 160, + 465, + 190 + ], + "type": "text", + "content": " Min Sun" + }, + { + "bbox": [ + 135, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 135, + 160, + 465, + 190 + ], + "type": "text", + "content": " Cheng-Hao Kuo" + }, + { + "bbox": [ + 135, + 160, + 465, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 203, + 517, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 203, + 517, + 232 + ], + "spans": [ + { + "bbox": [ + 75, + 203, + 517, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 203, + 517, + 232 + ], + "type": "text", + "content": "Carnegie Mellon University " + }, + { + "bbox": [ + 75, + 203, + 517, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 203, + 517, + 232 + ], + "type": "text", + "content": "Stony Brook University " + }, + { + "bbox": [ + 75, + 203, + 517, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 75, + 203, + 517, + 232 + ], + "type": "text", + "content": "National Tsing Hua University " + }, + { + "bbox": [ + 75, + 203, + 517, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 75, + 203, + 517, + 232 + ], + "type": "text", + "content": "Amazon" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 259, + 192, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 259, + 192, + 272 + ], + "spans": [ + { + "bbox": [ + 143, + 259, + 192, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 285, + 290, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 290, + 621 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 290, + 621 + ], + "type": "text", + "content": "6D object pose estimation has shown strong generalizability to novel objects. However, existing methods often require either a complete, well-reconstructed 3D model or numerous reference images that fully cover the object. Estimating 6D poses from partial references, which capture only fragments of an object's appearance and geometry, remains challenging. To address this, we propose UA-Pose, an uncertainty-aware approach for 6D object pose estimation and online object completion specifically designed for partial references. We assume access to either (1) a limited set of RGBD images with known poses or (2) a single 2D image. For the first case, we initialize a partial object 3D model based on the provided images and poses, while for the second, we use image-to-3D techniques to generate an initial object 3D model. Our method integrates uncertainty into the incomplete 3D model, distinguishing between seen and unseen regions. This uncertainty enables confidence assessment in pose estimation and guides an uncertainty-aware sampling strategy for online object completion, enhancing robustness in pose estimation accuracy and improving object completeness. We evaluate our method on the YCB-Video, YCBIInEOAT, and HO3D datasets, including RGBD sequences of YCB objects manipulated by robots and human hands. Experimental results demonstrate significant performance improvements over existing methods, particularly when object observations are incomplete or partially captured. Project page: https://minfenli.github.io/UA-Pose/" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 645, + 128, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 128, + 657 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 128, + 657 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "6D object pose estimation, which determines the rigid 6-degree-of-freedom transformation between an object and a camera, is essential for various real-world applications such as robotic manipulation [17, 38, 39, 49] and augmented re" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 323, + 260, + 529, + 376 + ], + "blocks": [ + { + "bbox": [ + 323, + 260, + 529, + 376 + ], + "lines": [ + { + "bbox": [ + 323, + 260, + 529, + 376 + ], + "spans": [ + { + "bbox": [ + 323, + 260, + 529, + 376 + ], + "type": "image", + "image_path": "37db0d36fb83e091b613238948cfc2a2ea7b152f0be3e6b070408c9a83247ef3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 323, + 380, + 526, + 477 + ], + "blocks": [ + { + "bbox": [ + 323, + 380, + 526, + 477 + ], + "lines": [ + { + "bbox": [ + 323, + 380, + 526, + 477 + ], + "spans": [ + { + "bbox": [ + 323, + 380, + 526, + 477 + ], + "type": "image", + "image_path": "a205788b8f026eca1d12a5b0f168034c341c780a54780064154ebaa72266780b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 478, + 547, + 567 + ], + "lines": [ + { + "bbox": [ + 304, + 478, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 478, + 547, + 567 + ], + "type": "text", + "content": "Figure 1. Overview of UA-Pose. (a) We initialize an incomplete object 3D model " + }, + { + "bbox": [ + 304, + 478, + 547, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 478, + 547, + 567 + ], + "type": "text", + "content": " (Sec. 3.2), which is a hybrid mesh that combines texture, geometry, and uncertainty. (b) Given a sequence of RGBD test images, our method estimates the 6D pose of the object in each test image (Sec. 3.3) based on the incomplete model " + }, + { + "bbox": [ + 304, + 478, + 547, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 478, + 547, + 567 + ], + "type": "text", + "content": ". (c) While more test images are captured, we store them with estimated poses in a memory pool, which will be used for online object completion (Sec. 3.4) to refine the incomplete model " + }, + { + "bbox": [ + 304, + 478, + 547, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 478, + 547, + 567 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "ality [22, 29]. Recent research introduced model-based and model-free approaches that can generalize to arbitrary novel objects. These methods push the boundaries of pose estimation for real-world applications and can be divided into two scenarios based on the object information available beforehand. Model-based methods [1, 2, 4, 11, 13, 16, 23, 24, 28] require textured 3D CAD models of the object to estimate 6D object poses in 2D images by establishing 2D-3D correspondences. In contrast, model-free methods leverage a large set of posed reference RGB images [19], RGBD images [10, 41], or video sequences [7, 31] to estimate object poses without the need for CAD models." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1180" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 252 + ], + "type": "text", + "content": "To bridge model-based and model-free methods, FoundationPose [41] proposes a unified framework that supports both model-based and model-free setups, using 3D CAD models or posed RGBD images (referred to as references). While effective in generalizing to novel objects, FoundationPose and similar model-free methods [7, 10, 31] still depend on high-quality 3D models or sufficient posed RGBD images that capture the object from sufficient viewpoints. However, assuming the availability of high-quality 3D models is often impractical due to the vast number of objects that remain uncaptured or unmodeled in real-world environments. Likewise, obtaining adequate posed RGBD images is challenging, as real-world settings frequently involve limited viewpoints, occlusions from nearby objects, and complexities in estimating camera poses in dynamic scenes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 254, + 289, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 254, + 289, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 254, + 289, + 495 + ], + "type": "text", + "content": "To address these challenges, we present UA-Pose, an uncertainty-aware approach for 6D pose estimation and online object completion as shown in Fig. 1. Our approach is designed to work with partial reference inputs that can include valuable meta-information for robotic tasks, such as grasping patterns, detection markers, or affordance labels. We consider two scenarios: (1) When a limited set of reference RGBD images with known poses is available, we initialize an incomplete 3D model using these images and poses; (2) When only a single unposed RGB image is available, we apply image-to-3D techniques to generate an initial object 3D model for pose estimation. By marking unseen regions as uncertain, we introduce a hybrid representation (Sec. 3.2) that integrates texture, geometry, and uncertainty. The uncertainty is utilized during our proposed uncertainty-aware pose estimation (Sec. 3.3) to provide confidence for estimated poses. Additionally, we propose an uncertainty-aware sampling strategy to select informative images for online object completion, iteratively refining the object 3D model during testing (Sec. 3.4)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 496, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 496, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 496, + 287, + 568 + ], + "type": "text", + "content": "We evaluate our approach on the YCB-Video, YCBI-nEOAT, and HO3D datasets. The results demonstrate significant improvements over existing methods and highlight the applicability of our uncertainty-aware approach in real-world scenarios where fully captured 3D models are often unavailable. To summarize, our contributions are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 570, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 570, + 287, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 287, + 618 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 287, + 618 + ], + "type": "text", + "content": "- We propose an uncertainty-aware approach for 6D object pose estimation (Sec. 3.3) and online object completion (Sec. 3.4), which improves both pose accuracy and object completeness when only partial references are available." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 618, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 618, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 47, + 618, + 287, + 666 + ], + "type": "text", + "content": "- We develop a hybrid representation (Sec. 3.2) incorporating uncertainty. The uncertainty enables confidence assessment in pose estimation and supports an uncertainty-aware sampling strategy for online object completion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 666, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 666, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 666, + 287, + 714 + ], + "type": "text", + "content": "- We demonstrate pose estimation from a single RGB image by leveraging single-image-to-3D methods, which generate an initial model for pose estimation and provide augmented data for online object completion (Sec. 3.5)." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 109 + ], + "type": "text", + "content": "- Our method supports different partial object references as input, which could be associated to meta-information, making it feasible for real-world applications." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 120, + 392, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 120, + 392, + 133 + ], + "spans": [ + { + "bbox": [ + 306, + 120, + 392, + 133 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 140, + 501, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 140, + 501, + 153 + ], + "spans": [ + { + "bbox": [ + 306, + 140, + 501, + 153 + ], + "type": "text", + "content": "2.1. Model-based Object Pose Estimation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 159, + 547, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 159, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 547, + 376 + ], + "type": "text", + "content": "Traditional model-based pose estimation methods can be divided into two categories: instance-level [8, 9, 12, 25] and category-level [3, 15, 32, 34, 46, 47]. Both types of methods assume that a textured CAD model of the specific object or similar object categories is available during training. However, these approaches are limited to objects seen during training, significantly restricting their applicability in real-world scenarios where many objects remain unseen or unmodeled. To address this limitation, zero-shot methods have been proposed [2, 11, 13, 16, 28], which estimate the pose of novel objects by providing their CAD models at test time. These methods use object CAD models to either render 2D images for feature matching and 2D-3D correspondence estimation or leverage point clouds from CAD models to estimate 3D-3D correspondences. While these approaches improve generalization to unseen objects, they still require access to high-quality 3D models at test time, which is often impractical in real-world settings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 384, + 490, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 384, + 490, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 384, + 490, + 396 + ], + "type": "text", + "content": "2.2. Model-free Object Pose estimation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 403, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 547, + 715 + ], + "type": "text", + "content": "Model-free methods eliminate the need for an explicit textured CAD model. Gen6D [19] introduces a pipeline to estimate object poses from RGB images with known poses. However, due to limited training data, Gen6D faces challenges with out-of-distribution test cases and images affected by occlusions. OnePose [31] and its extension OnePose++ [7] use structure-from-motion techniques to reconstruct 3D point clouds from sequential RGB images and employ pre-trained 2D-3D matching networks to estimate the pose of test images. FS6D [10] employs a transformer-based architecture to extract features from multi-view RGBD images for pose estimation. More recently, FoundationPose [41], trained on a large synthetic dataset, presents a unified framework that supports both model-based and model-free setups. In its model-free setting, FoundationPose requires posed reference images that capture the object from sufficient viewpoints to reconstruct 3D models for robust pose estimation. In general, these model-free pose estimation methods rely on either sequential RGB images with known poses or densely posed RGBD images. However, assuming pre-captured posed RGBD images for unseen objects is impractical for two key reasons: first, real-world applications often involve numerous unseen objects, making it infeasible to pre-capture and collect images for each object; and second, even if sufficient images of the novel object are available, accurately estimating cam" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "1181" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 251, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 251, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 251, + 85 + ], + "type": "text", + "content": "era poses in dynamic environments is challenging." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 264, + 103 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 264, + 103 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 264, + 103 + ], + "type": "text", + "content": "2.3. Object Pose Tracking and Reconstruction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 109, + 289, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 289, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 289, + 301 + ], + "type": "text", + "content": "Assuming test images for pose estimation are sequential, 6D object pose tracking methods [36, 40] leverage temporal cues to achieve efficient, smooth, and accurate pose predictions across consecutive frames by estimating relative pose transformations from each frame to the initial frame. For instance, BundleSDF [40] is a 6D pose tracking method that estimates poses from sequential RGBD images while simultaneously training a signed distance field (SDF) [45] representation to maintain consistent pose estimates over time. However, pose tracking methods are constrained to tracking object poses within sequential images from a single video, without support for using reference images for initial pose estimation. In contrast, our method is designed to estimate object poses based on given object references, which may include metadata beneficial for robotic applications, such as grasping patterns, barcodes, or affordance labels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 307, + 259, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 307, + 259, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 307, + 259, + 319 + ], + "type": "text", + "content": "2.4. Single-Image-to-3D and Pose Estimation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 324, + 289, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 289, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 289, + 529 + ], + "type": "text", + "content": "Recently, single-image-to-3D methods have been developed to generate 3D models from a single image [18, 20, 21, 26, 27, 33, 43]. These methods offer potential for pose estimation when textured CAD models are unavailable. For instance, GigaPose [23] explored using generated models for pose estimation; however, relying on generated 3D models that are not closely similar to real objects can lead to imprecise pose estimates. To address this, we apply a single-image-to-3D approach [43] to synthesize a generated model for early pose estimation. As real RGBD images are gradually collected during testing, they are used to create a refined object model, while the generated model from single-image-to-3D methods is leveraged to produce augmented data to improve object geometry in the uncaptured parts. Incorporating these augmented data into the SDF training yields more complete object models, ultimately improving the robustness and accuracy of pose estimation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 538, + 103, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 103, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 103, + 550 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "type": "text", + "content": "An overview of our method is illustrated in Fig. 1. We begin by initializing an incomplete object 3D model " + }, + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "type": "text", + "content": " using partial references, which can be either: (1) a few reference RGBD images with known poses or (2) a single unposed RGB image of the object. In this work, we propose a hybrid object representation (Sec. 3.2) that integrates texture, geometry, and uncertainty, serving as the incomplete model " + }, + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 557, + 287, + 653 + ], + "type": "text", + "content": " for uncertainty-aware pose estimation (Sec. 3.3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 654, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 288, + 714 + ], + "type": "text", + "content": "Given a sequence of " + }, + { + "bbox": [ + 47, + 654, + 288, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 654, + 288, + 714 + ], + "type": "text", + "content": " RGB test images, " + }, + { + "bbox": [ + 47, + 654, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\{I_0,I_1,\\dots ,I_{k - 1}\\}" + }, + { + "bbox": [ + 47, + 654, + 288, + 714 + ], + "type": "text", + "content": ", corresponding depth maps, and an initial segmentation mask " + }, + { + "bbox": [ + 47, + 654, + 288, + 714 + ], + "type": "inline_equation", + "content": "m_0" + }, + { + "bbox": [ + 47, + 654, + 288, + 714 + ], + "type": "text", + "content": " for the object of interest in the first image " + }, + { + "bbox": [ + 47, + 654, + 288, + 714 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 47, + 654, + 288, + 714 + ], + "type": "text", + "content": ", our method estimates the 6D pose of the object " + }, + { + "bbox": [ + 47, + 654, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\xi_{i}" + }, + { + "bbox": [ + 47, + 654, + 288, + 714 + ], + "type": "text", + "content": " in each subsequent test image " + }, + { + "bbox": [ + 47, + 654, + 288, + 714 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 47, + 654, + 288, + 714 + ], + "type": "text", + "content": " using" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": "the proposed uncertainty-aware pose estimation approach (Sec. 3.3), based on the incomplete 3D model " + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": ". During testing, newly captured test RGBD images and their estimated poses will be appended to a memory pool. These images will be used for online object completion (Sec. 3.4) to refine the incomplete object 3D model " + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": ". Notably, when the input reference is a single unposed RGB image, we employ a single-image-to-3D approach [43] to initialize a generated model " + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": " to assist in pose estimation and object completion. More details are described in Sec. 3.5." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 201, + 523, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 201, + 523, + 213 + ], + "spans": [ + { + "bbox": [ + 305, + 201, + 523, + 213 + ], + "type": "text", + "content": "3.1. Background: Model-free Pose Estimation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 219, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 545, + 483 + ], + "type": "text", + "content": "Given RGBD reference images of a novel object and their camera poses, the task of model-free pose estimation is to determine the 6D object poses (i.e., the rigid 6D transformation from the object to the camera) of that novel object in test RGBD images. FoundationPose [41] is a foundational model for 6D pose estimation trained on a large-scale dataset. In the model-free setup, it uses sufficient RGBD reference images (e.g., 16 RGBD images) that fully cover the object to train a Signed Distance Field (SDF) and then extracts a mesh by marching cubes as the 3D object model. Given the 3D object model to render object images, it introduces a two-stage pipeline with two specialized networks: a pose refinement network, which refines pose hypotheses by estimating the relative pose between the test image and rendered object images, and a pose selection network, which identifies the best pose by comparing the test image with rendered images based on these refined poses. However, RGBD reference images that fully represent the object's appearance and geometry are usually unavailable in real-world scenarios. As a result, model-free methods may struggle to provide accurate pose estimates when only incomplete object models or sparse observations are available." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 491, + 471, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 491, + 471, + 504 + ], + "spans": [ + { + "bbox": [ + 305, + 491, + 471, + 504 + ], + "type": "text", + "content": "3.2. Hybrid Object Representation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 510, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 617 + ], + "type": "text", + "content": "To address the limitations, we introduce a hybrid object representation that integrates the object's texture, geometry, and uncertainty, as shown in Fig. 2. This uncertainty marks the seen and unseen regions of an incomplete 3D object model and provides confidence in pose estimations, helping to filter out less reliable results. Given a set of RGBD images, object masks, and their camera poses, we train an object-centric neural Signed Distance Field (SDF) [45] to learn the partial object's appearance and geometry." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "We then extract a 3D mesh using the marching cubes algorithm to explicitly model uncertainty. Inspired by Visual Hull [14], we define object uncertainty based on visibility inferred from 2D object masks. By projecting these masks onto the mesh " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": ", we label regions of " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": " seen in the reference images as \"certain\" and those unseen as \"uncertain\", as shown in Fig. 2. This approach allows pose estimation to account for both the seen and unseen parts of the object." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1182" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 75, + 283, + 159 + ], + "blocks": [ + { + "bbox": [ + 50, + 75, + 283, + 159 + ], + "lines": [ + { + "bbox": [ + 50, + 75, + 283, + 159 + ], + "spans": [ + { + "bbox": [ + 50, + 75, + 283, + 159 + ], + "type": "image", + "image_path": "8f86fd6dc3858ab38ac2f5fa4a989cda6c02879314baf8645078b91afd2704f4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 163, + 287, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 163, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 163, + 287, + 251 + ], + "type": "text", + "content": "Figure 2. Hybrid Object Representation Modeling. We propose a hybrid object representation (Sec. 3.2) that integrates the object's texture, geometry, and uncertainty. (a) First, a neural SDF is trained and extracted as a mesh representing the object's appearance and geometry (Sec. 3.2.1). (b) Then, we check the visibility of each mesh vertex from the viewpoint of each reference image to create the uncertainty map (Sec. 3.2.2), which reflects the seen and unseen regions of an incomplete 3D object model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 255, + 263, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 255, + 263, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 255, + 263, + 266 + ], + "type": "text", + "content": "3.2.1 Neural SDF Training and Mesh Extraction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 273, + 287, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 273, + 287, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 287, + 309 + ], + "type": "text", + "content": "We follow SDF training methods similar to [40, 41] to train a neural SDF for object modeling. The SDF defines the object surface as the set of 3D points:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 315, + 287, + 329 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 315, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 110, + 315, + 287, + 329 + ], + "type": "interline_equation", + "content": "S = \\{x \\in \\mathbb {R} ^ {3} \\mid \\Omega (x) = 0 \\}. \\tag {1}", + "image_path": "2a3db72453ee47dbe02bfc475e333b4298a2bdf43a65f89befe9782d502eec66.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 334, + 287, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 287, + 465 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 334, + 287, + 465 + ], + "type": "inline_equation", + "content": "\\Omega : \\mathbb{R}^3 \\to \\mathbb{R}" + }, + { + "bbox": [ + 46, + 334, + 287, + 465 + ], + "type": "text", + "content": " is the SDF geometry function that maps each 3D point " + }, + { + "bbox": [ + 46, + 334, + 287, + 465 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 334, + 287, + 465 + ], + "type": "text", + "content": " to a signed distance value " + }, + { + "bbox": [ + 46, + 334, + 287, + 465 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 334, + 287, + 465 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 334, + 287, + 465 + ], + "type": "inline_equation", + "content": "s = 0" + }, + { + "bbox": [ + 46, + 334, + 287, + 465 + ], + "type": "text", + "content": " indicates the object's surface. Given RGBD images, object masks, and their camera poses, we use a neural SDF to reconstruct the object's appearance and geometry, represented by two networks [45] as shown in Fig. 2. The geometric network " + }, + { + "bbox": [ + 46, + 334, + 287, + 465 + ], + "type": "inline_equation", + "content": "\\Omega : x \\mapsto s" + }, + { + "bbox": [ + 46, + 334, + 287, + 465 + ], + "type": "text", + "content": " takes a 3D point " + }, + { + "bbox": [ + 46, + 334, + 287, + 465 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 334, + 287, + 465 + ], + "type": "text", + "content": " as input and outputs a signed distance value " + }, + { + "bbox": [ + 46, + 334, + 287, + 465 + ], + "type": "inline_equation", + "content": "s \\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 334, + 287, + 465 + ], + "type": "text", + "content": ". The appearance network " + }, + { + "bbox": [ + 46, + 334, + 287, + 465 + ], + "type": "inline_equation", + "content": "\\Phi : (f_{\\Omega(x)}, d) \\mapsto c" + }, + { + "bbox": [ + 46, + 334, + 287, + 465 + ], + "type": "text", + "content": " takes an intermediate feature vector " + }, + { + "bbox": [ + 46, + 334, + 287, + 465 + ], + "type": "inline_equation", + "content": "f_{\\Omega(x)}" + }, + { + "bbox": [ + 46, + 334, + 287, + 465 + ], + "type": "text", + "content": " from the geometric network along with a view direction " + }, + { + "bbox": [ + 46, + 334, + 287, + 465 + ], + "type": "inline_equation", + "content": "d \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 334, + 287, + 465 + ], + "type": "text", + "content": " and outputs the color " + }, + { + "bbox": [ + 46, + 334, + 287, + 465 + ], + "type": "inline_equation", + "content": "c \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 334, + 287, + 465 + ], + "type": "text", + "content": " for the point." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 466, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 466, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 466, + 287, + 502 + ], + "type": "text", + "content": "**Rendering.** Given the object pose " + }, + { + "bbox": [ + 46, + 466, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 46, + 466, + 287, + 502 + ], + "type": "text", + "content": ", we render an image by casting rays through each pixel. Along each ray " + }, + { + "bbox": [ + 46, + 466, + 287, + 502 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 466, + 287, + 502 + ], + "type": "text", + "content": ", 3D points are sampled at multiple positions as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 119, + 508, + 287, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 508, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 119, + 508, + 287, + 521 + ], + "type": "interline_equation", + "content": "x _ {i} (r) = o (r) + t _ {i} d (r), \\tag {2}", + "image_path": "928c1192977d1ca63a3da603a535be7c55c55bf016e22479d3b66828b64de6d6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 526, + 287, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 526, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 526, + 287, + 598 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 526, + 287, + 598 + ], + "type": "inline_equation", + "content": "o(r)" + }, + { + "bbox": [ + 46, + 526, + 287, + 598 + ], + "type": "text", + "content": " denotes the ray origin and " + }, + { + "bbox": [ + 46, + 526, + 287, + 598 + ], + "type": "inline_equation", + "content": "d(r)" + }, + { + "bbox": [ + 46, + 526, + 287, + 598 + ], + "type": "text", + "content": " denotes the ray direction from the camera to the object determined by the object pose " + }, + { + "bbox": [ + 46, + 526, + 287, + 598 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 46, + 526, + 287, + 598 + ], + "type": "text", + "content": ". The parameter " + }, + { + "bbox": [ + 46, + 526, + 287, + 598 + ], + "type": "inline_equation", + "content": "t_i \\in \\mathbb{R}_+" + }, + { + "bbox": [ + 46, + 526, + 287, + 598 + ], + "type": "text", + "content": " determines the point position along the ray. Following the approach in [40], the color " + }, + { + "bbox": [ + 46, + 526, + 287, + 598 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 526, + 287, + 598 + ], + "type": "text", + "content": " of a ray " + }, + { + "bbox": [ + 46, + 526, + 287, + 598 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 526, + 287, + 598 + ], + "type": "text", + "content": " is computed by integrating sampled points near the object surface through volumetric rendering:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 75, + 604, + 287, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 604, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 75, + 604, + 287, + 634 + ], + "type": "interline_equation", + "content": "c (r) = \\int_ {z (r) - \\lambda} ^ {z (r) + 0. 5 \\lambda} w \\left(x _ {i}\\right) \\Phi \\left(f _ {\\Omega \\left(x _ {i}\\right)}, d \\left(x _ {i}\\right)\\right) d t, \\tag {3}", + "image_path": "1ab64af7a4e3ac4a50cf357b79a1d6172dc0d17cc5e244c0ff7cd47c09ece825.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 91, + 639, + 287, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 639, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 91, + 639, + 287, + 663 + ], + "type": "interline_equation", + "content": "w \\left(x _ {i}\\right) = \\frac {1}{1 + e ^ {- \\alpha \\Omega \\left(x _ {i}\\right)}} \\frac {1}{1 + e ^ {\\alpha \\Omega \\left(x _ {i}\\right)}}, \\tag {4}", + "image_path": "0cd4e8a9e4c61a248248fec525c723031be31d19bab797144184770d9beb74c0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "inline_equation", + "content": "w(x_{i})" + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "content": " is a bell-shaped probability density function [35] based on the distance of each point from the implicit object surface and " + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "content": " adjusts the sharpness of the density distribution. " + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "inline_equation", + "content": "z(r)" + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "content": " represents the ray's depth from the" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "depth image and " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " is the truncation distance. During volumetric rendering, " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "z(r)" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " are used to exclude 3D points far from the object surface, enhancing rendering efficiency and preventing the SDF from overfitting to empty regions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 121, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 545, + 144 + ], + "type": "text", + "content": "During training, we supervise this rendering by comparing it to the reference RGB images with the color loss:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 361, + 152, + 545, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 152, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 361, + 152, + 545, + 182 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = \\frac {1}{| \\mathcal {R} |} \\sum_ {r \\in \\mathcal {R}} \\| c (r) - \\bar {c} (r) \\| _ {2}, \\tag {5}", + "image_path": "1115697c1b5ebf67b4302fc7e105cb4e139ae57dd7015bd1f4b545cc1916ee14.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 191, + 546, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 191, + 546, + 311 + ], + "spans": [ + { + "bbox": [ + 304, + 191, + 546, + 311 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 191, + 546, + 311 + ], + "type": "inline_equation", + "content": "\\bar{c}(r)" + }, + { + "bbox": [ + 304, + 191, + 546, + 311 + ], + "type": "text", + "content": " is the ground-truth color at the pixel through which the ray " + }, + { + "bbox": [ + 304, + 191, + 546, + 311 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 191, + 546, + 311 + ], + "type": "text", + "content": " passes. For geometry learning, we leverage depth information from the provided depth maps to supervise the neural (SDF) through two losses [40]: the near-surface loss " + }, + { + "bbox": [ + 304, + 191, + 546, + 311 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 304, + 191, + 546, + 311 + ], + "type": "text", + "content": " and the empty space loss " + }, + { + "bbox": [ + 304, + 191, + 546, + 311 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_e" + }, + { + "bbox": [ + 304, + 191, + 546, + 311 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 191, + 546, + 311 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 304, + 191, + 546, + 311 + ], + "type": "text", + "content": " encourage the SDF surface to align closely with depth values and " + }, + { + "bbox": [ + 304, + 191, + 546, + 311 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_e" + }, + { + "bbox": [ + 304, + 191, + 546, + 311 + ], + "type": "text", + "content": " enforces that points along the ray between the camera and the surface remain empty. Additionally, we apply an eikonal regularization loss " + }, + { + "bbox": [ + 304, + 191, + 546, + 311 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{eik}" + }, + { + "bbox": [ + 304, + 191, + 546, + 311 + ], + "type": "text", + "content": " [5] to the near-surface SDF to enforce smoothness. The total loss is represented as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 347, + 322, + 545, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 322, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 347, + 322, + 545, + 335 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = w _ {c} \\mathcal {L} _ {c} + w _ {e} \\mathcal {L} _ {e} + w _ {s} \\mathcal {L} _ {s} + w _ {e i k} \\mathcal {L} _ {e i k}, \\tag {6}", + "image_path": "96b6610cc0e994f9dad07420379eff5878182a1f8650f1c0517e2af8ca554064.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 345, + 509, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 345, + 509, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 345, + 509, + 357 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 345, + 509, + 357 + ], + "type": "inline_equation", + "content": "w_{c}, w_{e}, w_{s}" + }, + { + "bbox": [ + 304, + 345, + 509, + 357 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 345, + 509, + 357 + ], + "type": "inline_equation", + "content": "w_{eik}" + }, + { + "bbox": [ + 304, + 345, + 509, + 357 + ], + "type": "text", + "content": " are weighting factors." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 357, + 545, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 545, + 464 + ], + "type": "text", + "content": "To explicitly model uncertainty, we convert the implicit surface defined by the neural SDF into an explicit 3D mesh " + }, + { + "bbox": [ + 304, + 357, + 545, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = (V,C,F)" + }, + { + "bbox": [ + 304, + 357, + 545, + 464 + ], + "type": "text", + "content": " with the vertices " + }, + { + "bbox": [ + 304, + 357, + 545, + 464 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 357, + 545, + 464 + ], + "type": "text", + "content": ", vertex colors " + }, + { + "bbox": [ + 304, + 357, + 545, + 464 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 357, + 545, + 464 + ], + "type": "text", + "content": ", and faces " + }, + { + "bbox": [ + 304, + 357, + 545, + 464 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 357, + 545, + 464 + ], + "type": "text", + "content": ". To achieve this, we extract a mesh from the trained neural SDF representation using the marching cubes algorithm, which guarantees a closed-surface mesh including vertices from any viewing angle. This allows us to model the visibility of the object from any viewing angle even if some parts are not visible in the reference images." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 464, + 545, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 545, + 512 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 545, + 512 + ], + "type": "text", + "content": "Each vertex " + }, + { + "bbox": [ + 304, + 464, + 545, + 512 + ], + "type": "inline_equation", + "content": "v_{i} \\in V" + }, + { + "bbox": [ + 304, + 464, + 545, + 512 + ], + "type": "text", + "content": " is assigned a color " + }, + { + "bbox": [ + 304, + 464, + 545, + 512 + ], + "type": "inline_equation", + "content": "c_{i} \\in C" + }, + { + "bbox": [ + 304, + 464, + 545, + 512 + ], + "type": "text", + "content": ", which is computed using the appearance function " + }, + { + "bbox": [ + 304, + 464, + 545, + 512 + ], + "type": "inline_equation", + "content": "\\varPhi" + }, + { + "bbox": [ + 304, + 464, + 545, + 512 + ], + "type": "text", + "content": ", which maps the intermediate feature vector " + }, + { + "bbox": [ + 304, + 464, + 545, + 512 + ], + "type": "inline_equation", + "content": "f_{\\Omega(v_{i})}" + }, + { + "bbox": [ + 304, + 464, + 545, + 512 + ], + "type": "text", + "content": " from the geometry network and a view direction " + }, + { + "bbox": [ + 304, + 464, + 545, + 512 + ], + "type": "inline_equation", + "content": "d \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 464, + 545, + 512 + ], + "type": "text", + "content": " to the RGB color:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 388, + 522, + 545, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 522, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 388, + 522, + 545, + 536 + ], + "type": "interline_equation", + "content": "c _ {i} = \\Phi \\left(f _ {\\Omega \\left(v _ {i}\\right)}, d\\right). \\tag {7}", + "image_path": "c7c6632690675aada0762d6aa023d85eb487f181536b6fd72fa8a6fa94605e72.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 550, + 433, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 433, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 433, + 562 + ], + "type": "text", + "content": "3.2.2 Uncertainty Modeling" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 570, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 714 + ], + "type": "text", + "content": "To model the uncertainty map that represents whether a vertex is visible from the viewpoints of input references, for each vertex " + }, + { + "bbox": [ + 304, + 570, + 545, + 714 + ], + "type": "inline_equation", + "content": "v_{i} \\in V" + }, + { + "bbox": [ + 304, + 570, + 545, + 714 + ], + "type": "text", + "content": ", we check its visibility from the viewpoint of each reference image " + }, + { + "bbox": [ + 304, + 570, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\bar{I}_j \\in \\tilde{\\mathcal{I}}" + }, + { + "bbox": [ + 304, + 570, + 545, + 714 + ], + "type": "text", + "content": " through the mesh rasterization process, which rasterizes 3D vertices onto 2D image pixels. We define an uncertainty score " + }, + { + "bbox": [ + 304, + 570, + 545, + 714 + ], + "type": "inline_equation", + "content": "u(v_{i})" + }, + { + "bbox": [ + 304, + 570, + 545, + 714 + ], + "type": "text", + "content": " based on whether " + }, + { + "bbox": [ + 304, + 570, + 545, + 714 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 304, + 570, + 545, + 714 + ], + "type": "text", + "content": " is visible from any of the viewpoints in the reference images " + }, + { + "bbox": [ + 304, + 570, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\bar{I}_j" + }, + { + "bbox": [ + 304, + 570, + 545, + 714 + ], + "type": "text", + "content": ". This uncertainty score " + }, + { + "bbox": [ + 304, + 570, + 545, + 714 + ], + "type": "inline_equation", + "content": "u(v_{i}) \\in \\{0,1\\}" + }, + { + "bbox": [ + 304, + 570, + 545, + 714 + ], + "type": "text", + "content": " labels vertices on the mesh as \"certain\" if they are seen in any reference image " + }, + { + "bbox": [ + 304, + 570, + 545, + 714 + ], + "type": "inline_equation", + "content": "(u(v_{i}) = 0)" + }, + { + "bbox": [ + 304, + 570, + 545, + 714 + ], + "type": "text", + "content": " and as \"uncertain\" if they remain unseen " + }, + { + "bbox": [ + 304, + 570, + 545, + 714 + ], + "type": "inline_equation", + "content": "(u(v_{i}) = 1)" + }, + { + "bbox": [ + 304, + 570, + 545, + 714 + ], + "type": "text", + "content": ". The resulting uncertainty map " + }, + { + "bbox": [ + 304, + 570, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{U} = \\{u(v_i) \\mid v_i \\in V\\}" + }, + { + "bbox": [ + 304, + 570, + 545, + 714 + ], + "type": "text", + "content": " is incorporated into our hybrid object" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1183" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 542, + 173 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 542, + 173 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 542, + 173 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 542, + 173 + ], + "type": "image", + "image_path": "4ef30e617b60be67a5e5b93f5e2b6e395d349da375deb031384565aac2fc26c9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 175, + 547, + 242 + ], + "lines": [ + { + "bbox": [ + 46, + 175, + 547, + 242 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 547, + 242 + ], + "type": "text", + "content": "Figure 3. Pipeline for Uncertainty-Aware Pose Estimation. Through uncertainty modeling (Sec. 3.2.2), we propose an uncertainty-aware pipeline to assess the confidence of estimated poses (Sec. 3.3). For each estimated pose, we calculate its seen IoU, which measures the overlap between the seen regions of the 3D model and the 2D object mask for the test image. When seen IoU is high, the pose is deemed confident, and both the pose and its RGBD image are added to the memory pool for object completion. Conversely, if the estimated pose has a low seen IoU, it is considered unreliable. In such cases, we employ an uncertainty-aware image sampling strategy to select images from the memory pool for online object completion (Sec. 3.4), thus enhancing pose estimation by refining the object model during testing." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 249, + 286, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 249, + 286, + 273 + ], + "spans": [ + { + "bbox": [ + 46, + 249, + 286, + 273 + ], + "type": "text", + "content": "representation " + }, + { + "bbox": [ + 46, + 249, + 286, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = (\\mathcal{E},\\mathcal{U})" + }, + { + "bbox": [ + 46, + 249, + 286, + 273 + ], + "type": "text", + "content": ", allowing the pose estimation process to distinguish between seen and unseen regions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 274, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 274, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 287, + 384 + ], + "type": "text", + "content": "Guidance from Uncertainty Map. Given an object pose " + }, + { + "bbox": [ + 46, + 274, + 287, + 384 + ], + "type": "inline_equation", + "content": "\\xi \\in \\mathbb{R}^{3\\times 4}" + }, + { + "bbox": [ + 46, + 274, + 287, + 384 + ], + "type": "text", + "content": ", the model " + }, + { + "bbox": [ + 46, + 274, + 287, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 274, + 287, + 384 + ], + "type": "text", + "content": " can be used to render an RGB image " + }, + { + "bbox": [ + 46, + 274, + 287, + 384 + ], + "type": "inline_equation", + "content": "I_{\\xi}^{\\mathrm{rend}}\\in \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 46, + 274, + 287, + 384 + ], + "type": "text", + "content": ", a depth map " + }, + { + "bbox": [ + 46, + 274, + 287, + 384 + ], + "type": "inline_equation", + "content": "D_{\\xi}^{\\mathrm{rend}}\\in \\mathbb{R}^{H\\times W}" + }, + { + "bbox": [ + 46, + 274, + 287, + 384 + ], + "type": "text", + "content": ", a rendered object mask " + }, + { + "bbox": [ + 46, + 274, + 287, + 384 + ], + "type": "inline_equation", + "content": "m_{\\xi}^{\\mathrm{rend}}\\in \\{0,1\\}^{H\\times W}" + }, + { + "bbox": [ + 46, + 274, + 287, + 384 + ], + "type": "text", + "content": ", and an uncertainty map " + }, + { + "bbox": [ + 46, + 274, + 287, + 384 + ], + "type": "inline_equation", + "content": "U_{\\xi}^{\\mathrm{rend}}\\in \\{0,1\\}^{H\\times W}" + }, + { + "bbox": [ + 46, + 274, + 287, + 384 + ], + "type": "text", + "content": " at a resolution of " + }, + { + "bbox": [ + 46, + 274, + 287, + 384 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 46, + 274, + 287, + 384 + ], + "type": "text", + "content": ". To evaluate the validity and confidence of an estimated pose, we introduce two metrics: uncertainty rate and seen IoU. The uncertainty rate measures the proportion of the uncertain pixels over the total pixels within the rendered object mask:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 392, + 287, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 392, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 126, + 392, + 287, + 422 + ], + "type": "interline_equation", + "content": "\\frac {\\sum \\left(U _ {\\xi} ^ {\\text {r e n d}} \\odot m _ {\\xi} ^ {\\text {r e n d}}\\right)}{\\sum \\left(m _ {\\xi} ^ {\\text {r e n d}}\\right)}, \\tag {8}", + "image_path": "e0ac61c965d9924597ba291995875f6027b5e43231bda6dd6c98b87828ec07b6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 426, + 287, + 475 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 426, + 287, + 475 + ], + "type": "text", + "content": " denotes element-wise logical multiplication. The seen IoU measures the overlap between seen (certain) regions of the rendered object mask and the object mask in the test image, which is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 110, + 482, + 287, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 482, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 110, + 482, + 287, + 498 + ], + "type": "interline_equation", + "content": "\\mathrm {I o U} \\left(\\neg U _ {\\xi} ^ {\\text {r e n d}} \\odot m _ {\\xi} ^ {\\text {r e n d}}, m ^ {\\text {t e s t}}\\right), \\tag {9}", + "image_path": "491182f8c6caec77898fc80a1259f3ebfd2bd56059555b69fd5444121ccbd472.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 506, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 506, + 287, + 593 + ], + "type": "inline_equation", + "content": "m^{\\mathrm{test}}" + }, + { + "bbox": [ + 46, + 506, + 287, + 593 + ], + "type": "text", + "content": " is the object mask from the test image, " + }, + { + "bbox": [ + 46, + 506, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\neg U_{\\xi}^{\\mathrm{rend}}\\odot" + }, + { + "bbox": [ + 46, + 506, + 287, + 593 + ], + "type": "inline_equation", + "content": "m_{\\xi}^{\\mathrm{rend}}" + }, + { + "bbox": [ + 46, + 506, + 287, + 593 + ], + "type": "text", + "content": " identifies \"certain\" pixels within the rendered object mask, and IoU represents the Intersection over Union function. Since pose estimation is more robust when based on seen regions of the 3D model, a high seen IoU indicates high confidence in the pose estimation, whereas a low seen IoU suggests that the estimation is not reliable." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 599, + 237, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 237, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 237, + 612 + ], + "type": "text", + "content": "3.3. Uncentainty-aware Pose Estimation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": "After modeling the uncertainty, the hybrid object representation could be used in our uncertainty-aware pose estimation pipeline, as shown in Fig. 3. In this work, we leverage the pose refinement and the pose selection modules of [41] and assume that the test images are sequential. To obtain per-frame 2D object masks in test images for calculating the uncertainty rate and seen IoU, given first frame object mask " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "m_0" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": ", the object masks of the following" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 249, + 545, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 249, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 305, + 249, + 545, + 273 + ], + "type": "text", + "content": "frames " + }, + { + "bbox": [ + 305, + 249, + 545, + 273 + ], + "type": "inline_equation", + "content": "\\{m_1, m_2, \\ldots, m_{k-1}\\}" + }, + { + "bbox": [ + 305, + 249, + 545, + 273 + ], + "type": "text", + "content": " is determined by off-the-shelf segmentation methods [44]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 276, + 546, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 276, + 546, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 546, + 443 + ], + "type": "text", + "content": "Pose Initialization. Given an RGBD video, object masks for each frame, and an incomplete object model " + }, + { + "bbox": [ + 304, + 276, + 546, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 276, + 546, + 443 + ], + "type": "text", + "content": ", we begin by estimating the object pose for the first frame by generating multiple pose hypotheses. For each hypothesis, the translation is initialized using the 3D point projected from the center of the 2D object mask, determined by the median depth value within this region. For rotation, we uniformly sample " + }, + { + "bbox": [ + 304, + 276, + 546, + 443 + ], + "type": "inline_equation", + "content": "N_{v}" + }, + { + "bbox": [ + 304, + 276, + 546, + 443 + ], + "type": "text", + "content": " viewpoints from an icosphere centered on the object and orient the camera to face the object's center. Subsequently, " + }, + { + "bbox": [ + 304, + 276, + 546, + 443 + ], + "type": "inline_equation", + "content": "N_{in}" + }, + { + "bbox": [ + 304, + 276, + 546, + 443 + ], + "type": "text", + "content": " in-plane rotations are sampled and applied to each viewpoint, yielding " + }, + { + "bbox": [ + 304, + 276, + 546, + 443 + ], + "type": "inline_equation", + "content": "N_{v} \\cdot N_{in}" + }, + { + "bbox": [ + 304, + 276, + 546, + 443 + ], + "type": "text", + "content": " hypothesized object poses. Each object pose can be represented as " + }, + { + "bbox": [ + 304, + 276, + 546, + 443 + ], + "type": "inline_equation", + "content": "[\\pmb{R}|\\pmb{t}] \\in \\mathbb{S}\\mathbb{E}(3)" + }, + { + "bbox": [ + 304, + 276, + 546, + 443 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 276, + 546, + 443 + ], + "type": "inline_equation", + "content": "\\pmb{R} \\in \\mathbb{S}\\mathbb{O}(3)" + }, + { + "bbox": [ + 304, + 276, + 546, + 443 + ], + "type": "text", + "content": " is the rotation and " + }, + { + "bbox": [ + 304, + 276, + 546, + 443 + ], + "type": "inline_equation", + "content": "\\pmb{t} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 276, + 546, + 443 + ], + "type": "text", + "content": " is the translation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 447, + 547, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 547, + 590 + ], + "type": "text", + "content": "Pose Refinement. After generating " + }, + { + "bbox": [ + 304, + 447, + 547, + 590 + ], + "type": "inline_equation", + "content": "N_v \\cdot N_{in}" + }, + { + "bbox": [ + 304, + 447, + 547, + 590 + ], + "type": "text", + "content": " pose hypotheses, we apply the refinement module [41] to improve pose accuracy. Specifically, given a rendered RGBD image of the object from each pose hypothesis and the test RGBD image of the target object, the network outputs a pose update " + }, + { + "bbox": [ + 304, + 447, + 547, + 590 + ], + "type": "inline_equation", + "content": "\\Delta R \\in \\mathbb{SO}(3)" + }, + { + "bbox": [ + 304, + 447, + 547, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 447, + 547, + 590 + ], + "type": "inline_equation", + "content": "\\Delta t \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 447, + 547, + 590 + ], + "type": "text", + "content": " to refine the pose hypothesis, aligning it more closely with the observed object pose in the test image. Each refined pose is represented as " + }, + { + "bbox": [ + 304, + 447, + 547, + 590 + ], + "type": "inline_equation", + "content": "[R^{+}|t^{+}] \\in \\mathbb{SE}(3)" + }, + { + "bbox": [ + 304, + 447, + 547, + 590 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 447, + 547, + 590 + ], + "type": "inline_equation", + "content": "t^{+} = t + \\Delta t" + }, + { + "bbox": [ + 304, + 447, + 547, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 447, + 547, + 590 + ], + "type": "inline_equation", + "content": "R^{+} = \\Delta R \\otimes R" + }, + { + "bbox": [ + 304, + 447, + 547, + 590 + ], + "type": "text", + "content": ". This refinement process could be iteratively repeated. In this work, each pose is refined five times for pose estimation on the first frame and twice for subsequent frames." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "Pose Selection. Given a set of refined pose hypotheses, we apply our proposed metrics (Sec. 3.2.2) to filter out unreliable poses: the uncertainty rate, which measures the proportion of uncertain pixels within the rendered object mask, and the seen IoU, which quantifies the overlap between previously seen (certain) regions of the object and the object mask in the test image. A pose hypothesis is discarded when its uncertainty rate is above the threshold " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "T_{u}" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": " or its seen IoU is below the threshold " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "T_{s}" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": ", ensuring that only reliable hypotheses that adequately cover seen object regions are con-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1184" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "sidered. At last, we use the pose selection module [41] to identify the most accurate pose as the final estimate." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 286, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 286, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 286, + 193 + ], + "type": "text", + "content": "Pose Tracking. Given the sequential nature of RGBD video frames, object poses in adjacent frames are typically similar. For each frame following the initial one, we take the estimated pose from the previous frame as the sole pose hypothesis and refine it using the pose refinement module to generate the estimated pose for the current frame. This iterative process continues across the video sequence, leveraging temporal cues for stable and smooth pose estimation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 194, + 288, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 288, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 288, + 541 + ], + "type": "text", + "content": "Memory Pool. To enhance pose estimation by completing the object model during testing, we select test RGBD images with their estimated poses to incrementally refine the object representation. For efficiency, we introduce a memory pool " + }, + { + "bbox": [ + 46, + 194, + 288, + 541 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{\\xi_0,\\xi_1,\\xi_2,\\dots ,\\xi_{|\\mathcal{P}| - 1}\\}" + }, + { + "bbox": [ + 46, + 194, + 288, + 541 + ], + "type": "text", + "content": " that stores the most informative " + }, + { + "bbox": [ + 46, + 194, + 288, + 541 + ], + "type": "inline_equation", + "content": "|\\mathcal{P}|" + }, + { + "bbox": [ + 46, + 194, + 288, + 541 + ], + "type": "text", + "content": " viewpoints to maintain a compact yet diverse set of test image frames for object completion, where " + }, + { + "bbox": [ + 46, + 194, + 288, + 541 + ], + "type": "inline_equation", + "content": "\\xi_{i}" + }, + { + "bbox": [ + 46, + 194, + 288, + 541 + ], + "type": "text", + "content": " is the estimated pose corresponding to a test image " + }, + { + "bbox": [ + 46, + 194, + 288, + 541 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 46, + 194, + 288, + 541 + ], + "type": "text", + "content": ". The first frame " + }, + { + "bbox": [ + 46, + 194, + 288, + 541 + ], + "type": "inline_equation", + "content": "\\xi_{i}" + }, + { + "bbox": [ + 46, + 194, + 288, + 541 + ], + "type": "text", + "content": " is automatically added to " + }, + { + "bbox": [ + 46, + 194, + 288, + 541 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 194, + 288, + 541 + ], + "type": "text", + "content": ", establishing the canonical coordinate system for the novel object. Subsequent frames are added when their viewpoints contribute to the multi-view diversity in the pool. Specifically, before adding a new frame, we calculate the rotational geodesic distance between the new frame and the last frame added to the pool. A frame is included in " + }, + { + "bbox": [ + 46, + 194, + 288, + 541 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 194, + 288, + 541 + ], + "type": "text", + "content": " only if this geodesic distance exceeds a threshold " + }, + { + "bbox": [ + 46, + 194, + 288, + 541 + ], + "type": "inline_equation", + "content": "T_{geo}" + }, + { + "bbox": [ + 46, + 194, + 288, + 541 + ], + "type": "text", + "content": ", ensuring that each frame offers a unique viewpoint to enhance model completeness while keeping " + }, + { + "bbox": [ + 46, + 194, + 288, + 541 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 194, + 288, + 541 + ], + "type": "text", + "content": " compact. Note that not all estimated poses are reliable, adding inaccurate results to the memory pool may cause noisy SDF training during object completion. To mitigate this, we introduce an image-filtering strategy to filter out low-confidence estimations. As illustrated in Fig. 3, when a pose is estimated, its confidence is evaluated using our proposed metric, seen IoU, which measures the overlap between the seen regions of the 3D model and the 2D object mask in the test image. Since a high seen IoU indicates greater confidence in an estimated pose, frames are added to the memory pool only if their seen IoU exceeds a threshold " + }, + { + "bbox": [ + 46, + 194, + 288, + 541 + ], + "type": "inline_equation", + "content": "T_{conf}" + }, + { + "bbox": [ + 46, + 194, + 288, + 541 + ], + "type": "text", + "content": ", ensuring that low-confidence poses are excluded." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 550, + 251, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 251, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 251, + 563 + ], + "type": "text", + "content": "3.4. Uncentainty-aware Object Completion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 569, + 286, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 286, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 286, + 665 + ], + "type": "text", + "content": "As shown in Fig. 3, online object completion is triggered when the seen IoU of the current test frame falls below a threshold " + }, + { + "bbox": [ + 46, + 569, + 286, + 665 + ], + "type": "inline_equation", + "content": "T_{complete}" + }, + { + "bbox": [ + 46, + 569, + 286, + 665 + ], + "type": "text", + "content": ", indicating that the current pose estimation may no longer be reliable. As new frames are continuously added, the memory pool " + }, + { + "bbox": [ + 46, + 569, + 286, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 569, + 286, + 665 + ], + "type": "text", + "content": " can grow excessively large, which may reduce SDF training efficiency. To manage this, we perform uncertainty-aware sampling to select the " + }, + { + "bbox": [ + 46, + 569, + 286, + 665 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 569, + 286, + 665 + ], + "type": "text", + "content": " most informative frames for object completion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "content": "Uncertainty-aware Sampling. Early in the video sequence, when " + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "inline_equation", + "content": "|\\mathcal{P}| \\leq K" + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "content": ", all frames in the pool are included without selection. However, once the pool size exceeds " + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "content": " we apply an uncertainty-aware image selection strategy to" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 73, + 478, + 193 + ], + "blocks": [ + { + "bbox": [ + 310, + 73, + 478, + 193 + ], + "lines": [ + { + "bbox": [ + 310, + 73, + 478, + 193 + ], + "spans": [ + { + "bbox": [ + 310, + 73, + 478, + 193 + ], + "type": "image", + "image_path": "5f5e32d554693063e3a24cf45252f3a8a34cb877e17513714866070a0db819e4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 193, + 545, + 291 + ], + "lines": [ + { + "bbox": [ + 305, + 193, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 305, + 193, + 545, + 291 + ], + "type": "text", + "content": "Figure 4. Pipeline for leveraging generated models. In Sec. 3.5, the initial model generated by image-to-3D techniques is used to estimate poses for the first few frames. As more frames are captured, a refined object model is trained to replace the initial generated model, resulting in a more complete representation. To further leverage the generated 3D model, we render RGBD images from the image-to-3D generated model as augmented data for SDF training. This supervision enhances the object geometry in unseen regions, providing more complete information for pose estimation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 487, + 73, + 539, + 193 + ], + "blocks": [ + { + "bbox": [ + 487, + 73, + 539, + 193 + ], + "lines": [ + { + "bbox": [ + 487, + 73, + 539, + 193 + ], + "spans": [ + { + "bbox": [ + 487, + 73, + 539, + 193 + ], + "type": "image", + "image_path": "8d659f63bc6c501eddc73a3d3b7a0a4baf595c4464b985a4a8ad6214c95aea94.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 301, + 545, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 301, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 305, + 301, + 545, + 397 + ], + "type": "text", + "content": "maximize the coverage of unseen object regions. This process selects a subset, " + }, + { + "bbox": [ + 305, + 301, + 545, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{P}' \\subset \\mathcal{P}" + }, + { + "bbox": [ + 305, + 301, + 545, + 397 + ], + "type": "text", + "content": ", that reveals the most unseen areas of the object. The selection process begins by including the initial and most recent frames in the pool to establish the starting and ending viewpoints. Next, an iterative strategy is employed: at each iteration, we select the pose that reveals the largest unseen region. This continues until " + }, + { + "bbox": [ + 305, + 301, + 545, + 397 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 301, + 545, + 397 + ], + "type": "text", + "content": " poses are selected." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 407, + 499, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 407, + 499, + 419 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 499, + 419 + ], + "type": "text", + "content": "3.5. Leveraging Image-to-3D Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 426, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 426, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 305, + 426, + 545, + 569 + ], + "type": "text", + "content": "Given a single unposed RGB image, we employ image-to-3D techniques to generate an initial object 3D model " + }, + { + "bbox": [ + 305, + 426, + 545, + 569 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}}" + }, + { + "bbox": [ + 305, + 426, + 545, + 569 + ], + "type": "text", + "content": " for pose estimation. The viewpoint associated with the initial RGB image is labeled as \"certain,\" while other areas that are inferred by the image-to-3D method are marked \"uncertain\". The generated model is then rescaled to align with the object size in the test images. Specifically, we use the depth map and object mask from the first frame of the test image to approximate a rough model size. We then sample slightly larger and smaller models, generate pose hypotheses, and employ the pose selection module [41] to select the optimal pose hypothesis and corresponding model size." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Refer to Fig. 4, the initial generated model is used to estimate poses for the first few test frames until the rotational geodesic distance between the initial pose and a new estimated pose exceeds a threshold " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "T_{gen}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ". As more frames are captured, a refined object model is trained to replace the initial generated model, resulting in a more complete representation. To further leverage the generated 3D model, we render RGBD images from the image-to-3D generated model as augmented data for SDF training. This supervision enhances the object geometry in unseen regions, providing more complete information for pose estimation. See the supplementary materials for additional details." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1185" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 123, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 123, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 123, + 85 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 110, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 110, + 102 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 110, + 102 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 109, + 286, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 286, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 286, + 145 + ], + "type": "text", + "content": "YCB-Video [42]: The YCB-Video dataset contains RGBD video sequences of everyday objects from the YCB objects, widely used for benchmarking 6D object pose estimation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 145, + 286, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 145, + 286, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 145, + 286, + 193 + ], + "type": "text", + "content": "YCBCInEOAT [37]: YCBCInEOAT extends the YCB-Video dataset with RGBD sequences of YCB objects manipulated by a dual-arm robot. This dataset focuses on real-world scenarios involving robot-object interactions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 193, + 286, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 286, + 229 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 286, + 229 + ], + "type": "text", + "content": "HO3D [6]: The HO3D (Hand-Object 3D) dataset contains RGBD sequences of human hands interacting with objects, capturing complex hand-object interactions and occlusions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 237, + 162, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 237, + 162, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 237, + 162, + 251 + ], + "type": "text", + "content": "4.2. Experimental setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 255, + 286, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 286, + 351 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 286, + 351 + ], + "type": "text", + "content": "Baselines. We evaluate our method against state-of-the-art RGBD model-free 6D pose estimation methods [10, 30, 41] under two scenarios where partial object references are given as (1) A limited set of reference RGBD images (2 views) that include known poses. (2) A single unposed RGB object image. In scenario (2), we leverage image-to-3D techniques [43] to generate an initial object 3D model and iteratively estimate the object pose during testing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 351, + 286, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 351, + 286, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 286, + 460 + ], + "type": "text", + "content": "Metrics. We consider ADD and ADD-S metrics for 6D pose estimation, which measure the accuracy of estimated poses. The Area Under the Curve (AUC) is calculated and reported for both the ADD and ADD-S metrics, following the protocols in [40-42]. Furthermore, to demonstrate the effectiveness of our method for online object completion, we report the Chamfer Distance (CD), which measures the average distance between points of the reconstructed model and the ground-truth model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 467, + 235, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 467, + 235, + 480 + ], + "spans": [ + { + "bbox": [ + 47, + 467, + 235, + 480 + ], + "type": "text", + "content": "4.3. Comparison Results on YCB-Video" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 486, + 286, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 286, + 533 + ], + "type": "text", + "content": "We first evaluate UA-Pose on the YCB-Video [42] dataset, a widely used benchmark for 6D pose estimation, and compare our approach with state-of-the-art model-free methods [10, 30, 41] with a limited set of reference images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 534, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 286, + 713 + ], + "type": "text", + "content": "As presented in Tab. 1, our method significantly outperforms FoundationPose [41] with only 2 reference images in all metrics. Moreover, both our method and FoundationPose outperforms LoFTR [30] and FS6D-DPM [10] with 16 reference images with a large margin. This shows that FoundationPose is the strongest baseline for our further analysis. When only a single unposed RGB image of the object is available, we apply a single-image-to-3D approach [43] to generate an initial 3D model for pose estimation. Experimental results, shown in Tab. 1, indicate that directly using the generated model with FoundationPose may not yield optimal results, as these models may not accurately represent real objects. In contrast, UA-Pose uses the generated model only for initial pose estimation and for rendering RGBD images as augmented data to aid object" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 308, + 137, + 545, + 205 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 137 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 137 + ], + "type": "text", + "content": "Table 1. Quantitative results on the YCB-Video dataset. \"Comp.\" indicates whether the method includes online completion, and \"-\" denotes that the method does not reconstruct object shapes. Results for [10, 30] are adopted from [41]. Experiments for \"single RGB + Image-to-3D\" are conducted on a subset of the YCB-Video dataset. See supplementary materials for details." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 137, + 545, + 205 + ], + "lines": [ + { + "bbox": [ + 308, + 137, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 137, + 545, + 205 + ], + "type": "table", + "html": "
InputComp.Mean
ADDADD-SCD
LoFTR [30]16 referencesNo26.252.5-
FS6D-DPM [10]16 referencesNo42.188.4-
FoundationPose [41]2 referencesNo87.494.30.57
Ours2 referencesYes92.896.50.53
FoundationPose [41]single RGB + Image-to-3DNo88.993.60.67
Ourssingle RGB + Image-to-3DYes93.296.90.62
", + "image_path": "8d81e4efa51b9c976587e3f90356bf82c2673c9348d731d79ed9123193866ab4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 308, + 233, + 545, + 284 + ], + "blocks": [ + { + "bbox": [ + 305, + 211, + 545, + 232 + ], + "lines": [ + { + "bbox": [ + 305, + 211, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 305, + 211, + 545, + 232 + ], + "type": "text", + "content": "Table 2. Quantitative results on the YCBIInEOAT dataset. \"Comp.\" indicates that the method performs online completion." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 233, + 545, + 284 + ], + "lines": [ + { + "bbox": [ + 308, + 233, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 308, + 233, + 545, + 284 + ], + "type": "table", + "html": "
InputComp.Mean
ADDADD-SCD
FoundationPose [41]2 referencesNo68.5284.800.60
Ours2 referencesYes89.9994.350.57
FoundationPose [41]single RGB + Image-to-3DNo83.3990.660.61
Ourssingle RGB + Image-to-3DYes89.2493.990.60
", + "image_path": "f843c93225d3329a87054396655807da502fe3fb685821cf6f33a15e1539db30.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 297, + 545, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 297, + 545, + 333 + ], + "spans": [ + { + "bbox": [ + 305, + 297, + 545, + 333 + ], + "type": "text", + "content": "completion, as described in Sec. 3.5. This approach reconstructs a model more closely aligned with the real object, leading to improved pose estimation accuracy." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 342, + 503, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 342, + 503, + 355 + ], + "spans": [ + { + "bbox": [ + 305, + 342, + 503, + 355 + ], + "type": "text", + "content": "4.4. Comparison Results on YCBIInEOAT" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 361, + 545, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 397 + ], + "type": "text", + "content": "We evaluated UA-Pose on the YCBIInEOAT [37] dataset to assess its performance in challenging robotic scenarios in real-world where objects are manipulated by robotic arms." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 397, + 545, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 397, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 545, + 564 + ], + "type": "text", + "content": "In Table 2, when 2 reference images are provided, UA-Pose demonstrates superior performance compared to FoundationPose, highlighting the effectiveness of our uncertainty-aware mechanisms and online object completion for robust pose estimation in robotic scenarios. As illustrated in Fig. 5, simply using the incomplete model generated from the partial references often leads to inaccurate estimations. In contrast, our method achieves robust pose estimation with uncertainty-aware mechanisms and online object completion. When only a single unposed RGB image of the object is available, UA-Pose also consistently surpasses FoundationPose across all metrics. This demonstrates the flexibility of UA-Pose to utilize diverse partial references, which could carry useful meta-information." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 574, + 471, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 574, + 471, + 586 + ], + "spans": [ + { + "bbox": [ + 305, + 574, + 471, + 586 + ], + "type": "text", + "content": "4.5. Comparison Results on HO3D" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 594, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 629 + ], + "type": "text", + "content": "We evaluated UA-Pose on the HO3D dataset to assess its effectiveness in complex hand-object interaction scenarios where objects are frequently rotated and occluded." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "As shown in Tab. 3, when 2 reference images are provided, FoundationPose fails to make accurate estimations when objects in the test video exhibit frequent and strong rotations. In contrast, UA-Pose accurately estimates object poses under the challenging hand-object interaction scenarios. As for utilizing the generated models from the single-image-to-3D approach for pose estimation, UA-Pose" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1186" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 67, + 282, + 182 + ], + "blocks": [ + { + "bbox": [ + 52, + 67, + 282, + 182 + ], + "lines": [ + { + "bbox": [ + 52, + 67, + 282, + 182 + ], + "spans": [ + { + "bbox": [ + 52, + 67, + 282, + 182 + ], + "type": "image", + "image_path": "3884863dd735546e24505d63eabfee59bbea2529a50d1f2d0d1d39439655d070.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 183, + 287, + 272 + ], + "lines": [ + { + "bbox": [ + 46, + 183, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 287, + 272 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison on the YCBIInEOAT. Object image (left) and depth map (right) pairs are rendered based on the estimated object poses using the incomplete model. Without uncertainty-aware mechanisms and Online Object Completion (referred to as \"w/o OOC\"), the incomplete model often leads to inaccurate estimations. In contrast, our method (referred to as \"w/ OOC\") achieves robust pose estimation by incorporating uncertainty-aware mechanisms and online object completion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 281, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 400 + ], + "type": "text", + "content": "achieves more accurate estimations than FoundationPose. Note that even if the generated models in Tab. 3 show lower Chamfer Distance (CD) than refined models obtained from object completion, directly using generated models for FoundationPose results in considerably lower ADD and ADD-S. This indicates that even though generated models have good geometry (low CD), geometry alone cannot accurately represent real objects for pose estimation because the textures of generated models are not good enough in unseen areas that are not covered by the initial partial references." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 407, + 140, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 140, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 140, + 420 + ], + "type": "text", + "content": "4.6. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 426, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 288, + 715 + ], + "type": "text", + "content": "Key Components. Tab. 4 presents an ablation study for key components of UA-Pose on the YCBCInEOAT dataset. All experiments start with the initialized model based on two reference images and are evaluated using ADD, ADD-S, and CD. Additionally, we report " + }, + { + "bbox": [ + 46, + 426, + 288, + 715 + ], + "type": "inline_equation", + "content": "N_{rebuild}" + }, + { + "bbox": [ + 46, + 426, + 288, + 715 + ], + "type": "text", + "content": ", the total number of times object completion is applied across test sequences to refine the incomplete model " + }, + { + "bbox": [ + 46, + 426, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 426, + 288, + 715 + ], + "type": "text", + "content": ". The \"w/o object completion\" setup disables online object completion, illustrating its importance for achieving accurate pose estimation. In \"w/o uncertainty-aware object completion\", object completion is applied every time a new test image is added to the memory pool, instead of completing the model only when the seen IoU threshold is below " + }, + { + "bbox": [ + 46, + 426, + 288, + 715 + ], + "type": "inline_equation", + "content": "T_{complete}" + }, + { + "bbox": [ + 46, + 426, + 288, + 715 + ], + "type": "text", + "content": ". This leads to poorer performance and a higher " + }, + { + "bbox": [ + 46, + 426, + 288, + 715 + ], + "type": "inline_equation", + "content": "N_{rebuild}" + }, + { + "bbox": [ + 46, + 426, + 288, + 715 + ], + "type": "text", + "content": " count, increasing unnecessary computation for object completion. The \"w/o image-filtering strategy\" experiment omits the image-filtering strategy (Sec. 3.3) for excluding unreliable poses, allowing low-confidence poses to be added to the memory pool and resulting in decreased pose accuracy. In \"w/o uncertainty-aware sampling\", we use the geodesic distance-based sampling strategy from [40] instead of our proposed uncertainty-aware sampling strategy (Sec. 3.4) to sample images for object completion. The results demonstrate our proposed sampling method is better at selecting infor" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 308, + 93, + 545, + 144 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 92 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 92 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 92 + ], + "type": "text", + "content": "Table 3. Quantitative results on the HO3D dataset. \"Comp.\" indicates that the method performs online completion." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 93, + 545, + 144 + ], + "lines": [ + { + "bbox": [ + 308, + 93, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 308, + 93, + 545, + 144 + ], + "type": "table", + "html": "
InputComp.Mean
ADDADD-SCD
FoundationPose [41]2 referencesNo45.6761.570.78
Ours2 referencesYes91.5195.560.69
FoundationPose [41]single RGB + Image-to-3DNo72.0687.870.76
Ourssingle RGB + Image-to-3DYes83.2393.590.88
", + "image_path": "74715dc9812bc615b1b1f6afff271d5e5ff11af116d5549436f40cd1f82701e2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 313, + 171, + 536, + 232 + ], + "blocks": [ + { + "bbox": [ + 305, + 148, + 545, + 170 + ], + "lines": [ + { + "bbox": [ + 305, + 148, + 545, + 170 + ], + "spans": [ + { + "bbox": [ + 305, + 148, + 545, + 170 + ], + "type": "text", + "content": "Table 4. Ablation study. We demonstrate the importance of UA-Pose's key components on the YCBCInEOAT dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 171, + 536, + 232 + ], + "lines": [ + { + "bbox": [ + 313, + 171, + 536, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 536, + 232 + ], + "type": "table", + "html": "
MeanNrebuild
ADDADDSCD
w/o object completion68.5284.800.60-
w/o uncertainty-aware object completion85.8392.390.72181
w/o image-filtering strategy87.7093.370.7187
w/o uncertainty-aware sampling87.8893.390.6859
Ours (full)89.9994.350.5758
", + "image_path": "7da8fc97de9de723977d9faea0f05d637018848002c53e38c3187257edce8473.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 313, + 291, + 536, + 342 + ], + "blocks": [ + { + "bbox": [ + 305, + 236, + 545, + 289 + ], + "lines": [ + { + "bbox": [ + 305, + 236, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 305, + 236, + 545, + 289 + ], + "type": "text", + "content": "Table 5. Additional study. We compare our method with FoundationPose [41] and pose tracking methods [36, 40] on the YCBI-nEOAT dataset. \"Comp.\" means the method performs online completion, and \"\\*\" denotes using cropped ground-truth meshes [40], which exclude invisible parts in the test sequence, to calculate CD." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 291, + 536, + 342 + ], + "lines": [ + { + "bbox": [ + 313, + 291, + 536, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 536, + 342 + ], + "type": "table", + "html": "
InputComp.Mean
ADDADD-SCD
BundleTrack [36]first-frame RGBDNo87.3492.532.81*
BundleSDF [40]first-frame RGBDYes86.9593.771.16*
FoundationPose [41]first-frame RGBDNo74.3484.581.12*
Oursfirst-frame RGBDYes88.3893.820.75*
", + "image_path": "8cd9f33de3907fda7da0499871f699c7674e57cec332f1e036f8788fd5eebfbd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 348, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 545, + 396 + ], + "type": "text", + "content": "mative images for accurate pose estimation. Finally, the \"Ours (full)\" setup that incorporates all our key components achieves the best performance across ADD, ADD-S, CD, and the fewest " + }, + { + "bbox": [ + 304, + 348, + 545, + 396 + ], + "type": "inline_equation", + "content": "N_{\\text{rebuild}}" + }, + { + "bbox": [ + 304, + 348, + 545, + 396 + ], + "type": "text", + "content": " for object completion." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 396, + 546, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 546, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 546, + 552 + ], + "type": "text", + "content": "Additional Study. When no external reference images are available, we demonstrate that we could initialize the object model for pose estimation using the first frame of the test RGBD image sequence. Besides the strongest model-free pose estimation baseline [41], we also compare our method with strong model-free pose tracking approaches [36, 40, 48] on the YCBIInEOAT dataset. Refer to Tab. 5, our method achieves the highest performance across ADD, ADD-S, and CD metrics. Note that unlike pose-tracking methods which are limited to estimating relative poses from sequential frames within the same RGBD video, our method can incorporate object reference images as input for useful meta-information." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 562, + 378, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 378, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 378, + 574 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "We introduced UA-Pose, an uncertainty-aware approach for 6D object pose estimation that addresses the limitations of previous model-free methods in the scenarios in which input object references are incomplete or partially captured. Incorporating mechanisms to handle uncertainty and leverage online object completion to complete object models, our method demonstrates substantial performance improvements on the YCB-Video, YCBIInEOAT, and HO3D datasets. We demonstrate UA-Pose is a robust solution for practical real-world applications where full 3D object models and extensive image references are usually unavailable." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1187" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Philipp Ausserlechner, David Haberger, Stefan Thalhammer, Jean-Baptiste Weibel, and Markus Vincze. Zs6d: Zero-shot 6d object pose estimation using vision transformers. ICRA, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 286, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 286, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 286, + 178 + ], + "type": "text", + "content": "[2] Andrea Caraffa, Davide Boscaini, Amir Hamza, and Fabio Poiesi. Freeze: Training-free zero-shot 6d pose estimation with geometric and vision foundation models. ECCV, 2024. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 181, + 286, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 181, + 286, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 181, + 286, + 213 + ], + "type": "text", + "content": "[3] Dengsheng Chen, Jun Li, Zheng Wang, and Kai Xu. Learning canonical shape space for category-level 6D object pose and size estimation. In CVPR, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 215, + 286, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 215, + 286, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 215, + 286, + 258 + ], + "type": "text", + "content": "[4] Jianqiu Chen, Zikun Zhou, Mingshan Sun, Rui Zhao, Liwei Wu, Tianpeng Bao, and Zhenyu He. Zeropose: Cadprompted zero-shot object 6d pose estimation in cluttered scenes. TCSVT, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 260, + 286, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 260, + 286, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 260, + 286, + 293 + ], + "type": "text", + "content": "[5] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, 2020. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 295, + 286, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 286, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 286, + 327 + ], + "type": "text", + "content": "[6] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In CVPR, 2020. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 329, + 286, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 329, + 286, + 372 + ], + "spans": [ + { + "bbox": [ + 53, + 329, + 286, + 372 + ], + "type": "text", + "content": "[7] Xingyi He, Jiaming Sun, Yuang Wang, Di Huang, Hujun Bao, and Xiaowei Zhou. OnePose++: Keypoint-free one-shot object pose estimation without CAD models. NeurIPS, 2022. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 374, + 286, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 374, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 374, + 286, + 407 + ], + "type": "text", + "content": "[8] Yisheng He, Wei Sun, Haibin Huang, Jianran Liu, Haoqiang Fan, and Jian Sun. PVN3D: A deep point-wise 3D keypoints voting network for 6DoF pose estimation. In CVPR, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 408, + 286, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 408, + 286, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 408, + 286, + 441 + ], + "type": "text", + "content": "[9] Yisheng He, Haibin Huang, Haoqiang Fan, Qifeng Chen, and Jian Sun. FFB6D: A full flow bidirectional fusion network for 6D pose estimation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 442, + 286, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 442, + 286, + 474 + ], + "spans": [ + { + "bbox": [ + 48, + 442, + 286, + 474 + ], + "type": "text", + "content": "[10] Yisheng He, Yao Wang, Haoqiang Fan, Jian Sun, and Qifeng Chen. Fs6d: Few-shot 6d pose estimation of novel objects. In CVPR, 2022. 1, 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 476, + 286, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 476, + 286, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 476, + 286, + 520 + ], + "type": "text", + "content": "[11] Junwen Huang, Hao Yu, Kuan-Ting Yu, Nassir Navab, Slobodan Ilic, and Benjamin Busam. Matchu: Matching unseen objects for 6d pose estimation from rgb-d images. In CVPR, 2024. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 522, + 286, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 286, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 286, + 555 + ], + "type": "text", + "content": "[12] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. CosyPose: Consistent multi-view multi-object 6D pose estimation. In ECCV, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 556, + 286, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 286, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 286, + 609 + ], + "type": "text", + "content": "[13] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. MegaPose: 6D pose estimation of novel objects via render & compare. In CoRL, 2022. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 612, + 286, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 286, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 286, + 634 + ], + "type": "text", + "content": "[14] Aldo Laurentini. The visual hull concept for silhouette-based image understanding. IEEE TPAMI, 1994. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 635, + 286, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 286, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 286, + 678 + ], + "type": "text", + "content": "[15] Taeyeop Lee, Jonathan Tremblay, Valts Blukis, Bowen Wen, Byeong-Uk Lee, Inkyu Shin, Stan Birchfield, In So Kweon, and Kuk-Jin Yoon. TTA-COPE: Test-time adaptation for category-level object pose estimation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "text", + "content": "[16] Jiehong Lin, Lihua Liu, Dekun Lu, and Kui Jia. Sam-6d: Segment anything model meets zero-shot 6d object pose estimation. In CVPR, 2024. 1, 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[17] Jian Liu, Wei Sun, Chongpei Liu, Xing Zhang, and Qiang Fu. Robotic continuous grasping system by shape transformer-guided multi-object category-level 6d pose estimation. IEEE Transactions on Industrial Informatics, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[18] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In ICCV, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 152, + 545, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 193 + ], + "type": "text", + "content": "[19] Yuan Liu, Yilin Wen, Sida Peng, Cheng Lin, Xiaoxiao Long, Taku Komura, and Wenping Wang. Gen6D: Generalizable model-free 6-DoF object pose estimation from RGB images. ECCV, 2022. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 195, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 545, + 237 + ], + "type": "text", + "content": "[20] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Learning to generate multiview-consistent images from a single-view image. ICLR, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "type": "text", + "content": "[21] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, and Wenping Wang. Wonder3d: Single image to 3d using cross-domain diffusion. CVPR, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 294, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 326 + ], + "type": "text", + "content": "[22] Eric Marchand, Hideaki Uchiyama, and Fabien Spindler. Pose estimation for augmented reality: A hands-on survey. TVCG, 2015. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 327, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 360 + ], + "type": "text", + "content": "[23] Van Nguyen Nguyen, Thibault Groueix, Mathieu Salzmann, and Vincent Lepetit. Gigapose: Fast and robust novel object pose estimation via one correspondence. CVPR, 2024. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 361, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 545, + 403 + ], + "type": "text", + "content": "[24] Evin Pinar Ornek, Yann Labbe, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, and Tomas Hodan. Found- pose: Unseen object pose estimation with foundation features. ECCV, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 405, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 405, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 307, + 405, + 545, + 436 + ], + "type": "text", + "content": "[25] Kiru Park, Timothy Patten, and Markus Vincze. Pix2Pose: Pixel-wise coordinate regression of objects for 6D pose estimation. In ICCV, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 438, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 438, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 307, + 438, + 545, + 502 + ], + "type": "text", + "content": "[26] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, and Bernard Ghanem. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. ICLR, 2024. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 504, + 545, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 545, + 556 + ], + "type": "text", + "content": "[27] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 559, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 545, + 591 + ], + "type": "text", + "content": "[28] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. OSOP: A multi-stage one shot object pose estimation framework. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 593, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 593, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 593, + 545, + 635 + ], + "type": "text", + "content": "[29] Yongzhi Su, Jason Rambach, Nareg Minaskan, Paul Lesur, Alain Pagani, and Didier Stricker. Deep multi-state object pose estimation for augmented reality assembly. In ISMAR-Adjunct, 2019. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 636, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 545, + 668 + ], + "type": "text", + "content": "[30] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In CVPR, 2021. 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[31] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. OnePose: One-shot object pose estimation without CAD models. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1188" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[32] Meng Tian, Marcelo H Ang, and Gim Hee Lee. Shape prior deformation for categorical 6D object pose and size estimation. In ECCV, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[33] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In ECCV, 2025. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 164, + 287, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 287, + 208 + ], + "type": "text", + "content": "[34] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6D object pose and size estimation. In CVPR, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 210, + 287, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 210, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 48, + 210, + 287, + 253 + ], + "type": "text", + "content": "[35] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 256, + 287, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 256, + 287, + 288 + ], + "spans": [ + { + "bbox": [ + 48, + 256, + 287, + 288 + ], + "type": "text", + "content": "[36] Bowen Wen and Kostas Bekris. Bundletrack: 6d pose tracking for novel objects without instance or category-level 3d models. In IROS, 2021. 3, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 291, + 287, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 291, + 287, + 333 + ], + "spans": [ + { + "bbox": [ + 48, + 291, + 287, + 333 + ], + "type": "text", + "content": "[37] Bowen Wen, Chaitanya Mitash, Baozhang Ren, and Kostas E Bekris. se (3)-tracknet: Data-driven 6d pose tracking by calibrating image residuals in synthetic domains. In IROS, 2020. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 336, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 336, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 336, + 287, + 369 + ], + "type": "text", + "content": "[38] Bowen Wen, Wenzhao Lian, Kostas Bekris, and Stefan Schaal. CatGrasp: Learning category-level task-relevant grasping in clutter from simulation. In ICRA, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 371, + 287, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 404 + ], + "type": "text", + "content": "[39] Bowen Wen, Wenzhao Lian, Kostas Bekris, and Stefan Schaal. You only demonstrate once: Category-level manipulation from single visual demonstration. RSS, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 406, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 406, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 406, + 287, + 459 + ], + "type": "text", + "content": "[40] Bowen Wen, Jonathan Tremblay, Valts Blukis, Stephen Tyree, Thomas Müller, Alex Evans, Dieter Fox, Jan Kautz, and Stan Birchfield. Bundlesdf: Neural 6-dof tracking and 3d reconstruction of unknown objects. In CVPR, 2023. 3, 4, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 462, + 287, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 462, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 48, + 462, + 287, + 495 + ], + "type": "text", + "content": "[41] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. Foundationpose: Unified 6d pose estimation and tracking of novel objects. CVPR, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 498, + 287, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 287, + 530 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 287, + 530 + ], + "type": "text", + "content": "[42] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. RSS, 2018. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 532, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 532, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 532, + 287, + 586 + ], + "type": "text", + "content": "[43] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 3, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "type": "text", + "content": "[44] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[45] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. NeurIPS, 2020. 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[46] Ruida Zhang, Yan Di, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. SSP-Pose: Symmetry-aware shape prior deformation for direct category-level object pose estimation. In IROS, 2022. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 207 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "text", + "content": "[47] Linfang Zheng, Chen Wang, Yinghan Sun, Esha Dasgupta, Hua Chen, Ales Leonardis, Wei Zhang, and Hyung Jin Chang. HS-Pose: Hybrid scope feature extraction for category-level object pose estimation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[48] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In CVPR, 2022. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 163, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 207 + ], + "type": "text", + "content": "[49] Chungang Zhuang, Shaofei Li, and Han Ding. Instance segmentation based 6d pose estimation of industrial objects using point clouds for robotic bin-picking. Robotics and Computer-Integrated Manufacturing, 2023. 1" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1189" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/UCM-VeID V2_ A Richer Dataset and A Pre-training Method for UAV Cross-Modality Vehicle Re-Identification/9b7d21db-8ee1-4718-8ddf-98653daa1c9e_content_list.json b/2025/UCM-VeID V2_ A Richer Dataset and A Pre-training Method for UAV Cross-Modality Vehicle Re-Identification/9b7d21db-8ee1-4718-8ddf-98653daa1c9e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6d05faa762a7860c67aca1a3ee93c89109a85bdc --- /dev/null +++ b/2025/UCM-VeID V2_ A Richer Dataset and A Pre-training Method for UAV Cross-Modality Vehicle Re-Identification/9b7d21db-8ee1-4718-8ddf-98653daa1c9e_content_list.json @@ -0,0 +1,1778 @@ +[ + { + "type": "text", + "text": "UCM-VeID V2: A Richer Dataset and A Pre-training Method for UAV Cross-Modality Vehicle Re-Identification", + "text_level": 1, + "bbox": [ + 145, + 128, + 852, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xingyue Liu Jiahao Qi Chen Chen KangCheng Bin Ping Zhong* National University of Defense Technology, China", + "bbox": [ + 205, + 203, + 795, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{liuxingyue18, qijiahao1996, chenchen21c, binkc21, zhongping}@nudt.edu.cn", + "bbox": [ + 173, + 241, + 816, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cross-Modality Re-Identification (VT-ReID) aims to achieve around-the-clock target matching, benefiting from the strengths of both RGB and infrared (IR) modalities. However, the field is hindered by limited datasets, particularly for vehicle VT-ReID, and by challenges such as modality bias training (MBT), stemming from biased pre-training on ImageNet. To tackle the above issues, this paper introduces an dataset benchmark, named UCM-VeID V2, for vehicle VT-ReID, and proposes a new self-supervised pre-training method, Cross-Modality Patch-Mixed Self-supervised Learning (PMSL). UCM-VeID V2 dataset features a significant increase in data volume, along with enhancements in multiple aspects. PMSL addresses MBT by learning modality-invariant features through Patch-Mixed Image Reconstruction (PMIR) and Modality Discrimination Adversarial Learning (MDAL), and enhances discriminability with Modality-Augmented Contrasting Cluster (MACC). Comprehensive experiments are carried out to validate the proposed method.", + "bbox": [ + 89, + 321, + 483, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 638, + 220, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cross-modality vehicle re-identification (ReID), also referred to as vehicle visible-infrared (VI) ReID [22], involves retrieving a matching vehicle image from a gallery set captured in the visible (RGB) or infrared (IR) spectrum, given a query vehicle image in the alternate modality. This process, where images are obtained from cameras of differing modalities, has garnered significant attention in applications such as video surveillance and public security [1, 28], primarily due to its advantages in enabling continuous, intelligent monitoring across day and night conditions.", + "bbox": [ + 89, + 664, + 482, + 815 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite its importance, research on vehicle VT-ReID remains relatively underdeveloped. A primary challenge lies in the lack of essential resources, including a comprehensive large-scale dataset and widely accessible benchmark", + "bbox": [ + 89, + 816, + 483, + 877 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/650e42b7c9288e3d2e5a94cc8eb2f7d7be8c4f5764bcc15f9eddcb9f62c1411f.jpg", + "image_caption": [ + "(a) Multiple Platforms" + ], + "image_footnote": [], + "bbox": [ + 516, + 290, + 905, + 358 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f24f985c52a5b5e52306a3bd1abb9b7ef9fe325624d2186bb5eb1743a4496851.jpg", + "image_caption": [ + "(b) Wider time spans" + ], + "image_footnote": [], + "bbox": [ + 516, + 369, + 903, + 436 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cd42070e455eb5c1d8d9e73efb3768b4a8c733f11e6a8072d8ebe4c2c6ec027f.jpg", + "image_caption": [ + "(c) Richer weather conditions" + ], + "image_footnote": [], + "bbox": [ + 516, + 448, + 903, + 515 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e881e6a447ae79ac1ed301ff496a789a0cfd9fef23e009eec007a435d70c4f42.jpg", + "image_caption": [ + "(d) Diverse Scenes" + ], + "image_footnote": [], + "bbox": [ + 516, + 527, + 903, + 594 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/79c7482d3305bee0d43233ed7746884924d63901f0a88dd194056ead19dde2ea.jpg", + "image_caption": [ + "(e) Samples", + "Figure 1. Increments and samples of UCM-VeID V2. (a)(b)(c)(d) are varies platforms, time periods, weather conditions and locations for data acquisition, respectively. (e) shows the 4 IDs with different modalities, view-points and scales samples." + ], + "image_footnote": [], + "bbox": [ + 516, + 606, + 903, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "results. Existing remote sensing vehicle ReID datasets, such as VRAI [31], UAV-VeID [29], VRU [24], and VeRi-UAV [19], primarily focus on the single RGB modality, lacking the infrared data necessary for cross-modality analysis. For this, [22] has introduced the first UAV RGB-IR vehicle ReID dataset, UCM-VeID, to support cross-modality vehicle ReID research. However, this dataset is constrained by its single data acquisition platform and limited number of", + "bbox": [ + 511, + 779, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 107, + 886, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22286", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "identities, reducing its suitability for generalizing to diverse real-world scenarios. Consequently, developing large-scale and more diverse vehicle VT-ReID datasets is essential to advance research and better meet the requirements of practical applications.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another key challenge in vehicle VT-ReID development is the need for further research on VT-ReID methods, particularly to address the modality bias training (MBT) problem, which remains under-explored. As introduced by [16], MBT refers to the modality imbalance that arises when VT-ReID models are initialized with pre-trained ImageNet weights [6]. The modality imbalance stems from training on RGB-exclusive ImageNet data, which skews feature learning towards the RGB modality and reduces the IR discrimination capability. Moreover, the MBT issue is amplified in large-scale multi-modal datasets: as shown in Fig. 2, models initialized with ImageNet pre-trained weights exhibit a performance drop on larger datasets, particularly in terms of discriminative ability. The potential underlying cause of the above phenomenon is that ImageNet pretrained weights perform well when handling infrared data with lower modality diversity. However, they lack sufficient generalization ability when faced with larger datasets that exhibit higher modality diversity, which exacerbates modality bias and degrades model performance.", + "bbox": [ + 91, + 169, + 482, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, this paper introduces a more comprehensive dataset for vehicle VT-ReID and a novel pre-training paradigm called Cross-modality Patch-mixed Self-supervised Learning (CPSL) to mitigate the MBT problem. The proposed dataset, UCM-VeID V2, includes 113,929 RGB images and 51,162 IR images of 3,407 vehicles captured by five cameras. Compared to UCM-VeID [22], it is significantly expanded across multiple dimensions, including additional cameras and acquisition platforms, extended capture times, a broader range of weather conditions, and more diverse shooting scenarios, as illustrated in Fig. 1. For the MBT challenge, inspired by [13], CPSL aligns the RGB and IR domains by incorporating a Patch-Mixed Image Reconstruction (PMIR) self-supervised task. To further eliminate the modality bias, CPSL designs a Modality Discrimination Adversarial Learning (MDAL) strategy to exact modality invariant features through an adversarial training manner. In addition, considering discrimination insufficiency, we introduce the Modality-Augmented Contrasting Learning (MACL) task, where patch-mixed images are used as augmented positive examples for contrastive learning.", + "bbox": [ + 91, + 473, + 482, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are as follows:", + "bbox": [ + 109, + 808, + 447, + 821 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the UCM-VeID V2 dataset, which significantly enhances the original UCM-VeID with greater scale and diversity across multiple dimensions.", + "- We propose a novel self-supervised paradigm to pre-train an encoder, replacing the ImageNet-trained ResNet50, to" + ], + "bbox": [ + 91, + 825, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0263be20b404b1e2de1e5a9efece11042bdd6f5df8cc7c52a25946019ee56a66.jpg", + "image_caption": [ + "(a) Training based on UCM-VeID" + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 696, + 195 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/74a9675ea7a5b6f8f2451ad9f3d8324cc16bb1c0c18b28049cbc42f7e0711420.jpg", + "image_caption": [ + "(b) Training based on UCM-VeID V2", + "Figure 2. The feature distribution of pre-trained embeddings is visualized. Panels (a) and (b) show the embeddings in 2D space for models trained on the UCM-VeID and UCM-VeID V2 datasets, respectively, with dots and crosses representing different identities in RGB and IR modalities. The experiment uses the AGW baseline [37]. As indicated by the red circles in (a), the same-ID samples are more compact. In contrast, the same-ID samples are more spread out in (b), which ultimately degrades Re-ID performance." + ], + "image_footnote": [], + "bbox": [ + 696, + 88, + 867, + 195 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/67328e746653e6e617c3de88a8a6e3da2f8881c52a006470abc67f82f6a0a519.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 874, + 89, + 901, + 193 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "address the MBT issue.", + "bbox": [ + 529, + 354, + 686, + 369 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 383, + 663, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Vehicle Re-ID Datasets", + "text_level": 1, + "bbox": [ + 513, + 409, + 725, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vehicle Re-ID research has flourished over the past few years, significantly benefited by the availability of vehicle Re-ID datasets. The most widely-employed datasets are primarily captured via fixed surveillance cameras, such as VehicleID [20], VeRI-776 [21], VERI-Wild [23] and Vehicle1M [10]. Compared to fixed surveillance, UAV-based cameras offer superior mobility, flexibility, and convenience, leading to the development of key UAV-based vehicle Re-ID datasets such as VRAI [31], UAV-VeID [29], VRU [24], and VeRi-UAV [19]. Furthermore, Liu et al. [22] have advanced the first RGB-IR vehicle Re-ID dataset, named UCM-VeID, to support the research of cross-modality vehicle Re-ID. However, the UCM-VeID dataset exhibits limitations in both size and diversity, highlighting an urgent need for the further expansion and enrichment of the dataset, ensuring it remains a potent benchmark for advancing the frontiers of vehicle Re-ID research in the face of evolving challenges and complexities.", + "bbox": [ + 511, + 430, + 906, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Modality Bias Training", + "text_level": 1, + "bbox": [ + 511, + 712, + 728, + 728 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The MBT problem is essentially a problem of unbalanced training. Imbalanced model learning has received extensive attention in multi-modal learning [15, 32], and great progresses have been made in VT-ReID [16, 30]. To our best knowledge, there are two distinct approaches to tackle the MBT issue. The first approach is to control the training process, such as learning rate adjusting [15], gradient modulation [27], and ID-exclusive sampling strategy [16]. The second approach is to use modal balanced weights rather than ImageNet pre-trained ResNet50 avoiding MBT problems from the beginning. Wan et al. [30] devised a self-", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22287", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/f88c127df84ed430d7ef49e12042e9c4a74b3bbafdab337289289aa866e0112c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetVenueIdentitiesImagesModalitiesCamerasPlatformCamera ViewTarget StateColorTypeOrientation
VRAIICCV201913022137613RGB2UAVFlexibleMove97—
UAV-VeIDIJCV2021460141917RGB2UAVFlexibleMove——4
UCM-VeIDTMM202475329928RGB/IR2UAVFlexibleMove&Static1298
UCM-VeID V2ours3407165091RGB/IR5UAVFlexibleMove&Static1298
", + "bbox": [ + 94, + 88, + 898, + 170 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Comparisons among other vehicle Re-ID datasets and UCM-VeID V2", + "bbox": [ + 264, + 179, + 732, + 193 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "supervised pre-training method to obtain a set of new unbiased weights, which achieves comparable performance to that of ImageNet-trained ResNet50. The second approach addresses the MBT problem from the outset by using pretrained weights that are modally balanced, instead of relying on subsequent adjustments. In contrast to the first approach, the second method more fundamentally resolves the modality bias, offering stronger generalization and more stable performance. Consequently, we follow the second approach to carry out the MBT problem.", + "bbox": [ + 89, + 220, + 483, + 371 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Self-Supervised Learning", + "text_level": 1, + "bbox": [ + 89, + 382, + 323, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-supervised learning, as a subset of unsupervised learning, constructs pretext tasks to extract valuable insights from large-scale unlabeled data, thereby learning representations beneficial for downstream tasks. Common methods involve contextual pretext tasks to learn discriminative features, such as Word2Vec [25], Jigsaw [26], Rotation [8], Colorization [38], and MAE [14]. Another prominent approach, contrastive learning, builds discriminative representations by comparing instances within the same feature space, as seen in methods like MoCo [12] and BYOL [9], both of which perform well on downstream tasks. Nevertheless, above methods are not specifically tailored for the Re-ID task, particularly for VT-ReID task. In this paper, we leverage the consistency of different modalities to construct a patch-mixed image reconstruction self-supervised task aimed at extracting shared features, thereby mitigating the modality discrepancy.", + "bbox": [ + 89, + 404, + 483, + 661 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Benchmark: UCM-VeID V2", + "text_level": 1, + "bbox": [ + 89, + 675, + 352, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To further enhance the UAV cross-modal vehicle re-identification dataset, we have significantly expanded the first edition, UCM-VeID [22], by an order of magnitude in terms of UAV platforms, identities and so on. This section details the expanded dataset called UCM-VeID V2.", + "bbox": [ + 89, + 700, + 483, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Dataset Description", + "text_level": 1, + "bbox": [ + 89, + 787, + 277, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(1) Dataset Partition. UCM-VeID V2 comprises a total of 3407 identities with 165091 samples. The dataset is split into a training set and a test set in an $8:2$ ratio, resulting in a training set with 2,693 identities and 131,451 samples, while the test set contains 674 identities and 33,640 samples. Furthermore, the query set and gallery set are derived", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "from the test set. Specifically, one sample is drawn from each ID within the test set to constitute the query set, while the remaining samples collectively form the gallery set. It is worth noting that the extraction process is random, and crucially, every sample in the query set has a corresponding sample with the same ID in the gallery set.", + "bbox": [ + 511, + 220, + 906, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(2) Dataset Statistics. To maximize the utility of the UCM-VeID V2 dataset, we perform a detailed statistical analysis of its key attributes, providing insights into the data patterns relevant to Re-ID tasks.", + "bbox": [ + 511, + 311, + 906, + 371 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- ID distribution of 5 cameras. Data is collected using 3 drone platforms equipped with 5 different cameras. $D300_{z}$ camera captures fewer IDs compared to the other four cameras, as Fig. 3(a) shown. The $D300_{z}$ camera captures high-resolution images, primarily focused on nighttime data, thereby enhancing the dataset's diversity.", + "- Distribution of ID samples. Samples are collected from different camera perspectives, ensuring diversity and ample representation for each ID. As shown in Fig. 3(b), most IDs have 20-40 samples, providing comprehensive coverage of target's appearance across different conditions.", + "- Distribution of Resolution. As shown in Fig. 3(c), resolutions vary significantly across cameras. For example, $D300_{i}r$ samples cluster around 10,000-15,000 pixels, while DM3P samples are predominantly 45,000-50,000 pixels, highlighting a resolution disparity that presents a core challenge in UAV-based VT-ReID.", + "- Energy Distribution of IR Samples. The dataset spans all seasons and covers a 24-hour period, capturing both dynamic and static vehicles, leading to diverse infrared energy levels. Fig. 3(d) reveals that most infrared energy values are between 80-100, corresponding to moderate daylight. Extreme conditions, like intense heat or cold, are less common but contribute to the dataset's breadth.", + "(3) Dataset Attributes. Attributes annotation is an integral part of building reliable, high-performance data sets. In this dataset, we annotate the below attributes." + ], + "bbox": [ + 511, + 371, + 905, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Vehicle Type. Vehicles are classified into categories such as car, SUV, minibus, MPV, bus, mini truck, truck, and pickup. As shown in Fig. 3(e), cars and SUVs make up the majority, while trucks and pickups are less frequent. This distribution reflects the urban focus of our data collection, where cars and SUVs are more prevalent.", + "bbox": [ + 514, + 795, + 905, + 885 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Vehicle color. We categorize vehicle colors into thirteen", + "bbox": [ + 514, + 885, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22288", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a19c874acef395028b97b1d4c6f296b53e6c41657a75ee410e33daba391ad4d0.jpg", + "image_caption": [ + "ID Distribution of Different Cameras", + "(a)" + ], + "image_footnote": [], + "bbox": [ + 99, + 109, + 292, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fec54423b6a3d274c3dbb2103f4febe2a2c904b7d30e96f8a2bf167fd9ef8d10.jpg", + "image_caption": [ + "Distribution of ID Samples", + "(b)" + ], + "image_footnote": [], + "bbox": [ + 303, + 107, + 491, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b7c36c07a56e51a86fe6656e404885347f0ceffa1bcb92af4e584b1f369f37eb.jpg", + "image_caption": [ + "Distribution of Resolution", + "(c)" + ], + "image_footnote": [], + "bbox": [ + 496, + 101, + 699, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/64f782825eaf9bef9199b5e3372266e79ed829407f64abf0e56b910c3e094097.jpg", + "image_caption": [ + "Energy Distribution of IR Samples", + "(d)" + ], + "image_footnote": [], + "bbox": [ + 700, + 104, + 895, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/50986752ee54355497badb93d78ca4013f9b8eab9a452f6f2128475caf0a78b9.jpg", + "image_caption": [ + "Distribution of Vehicle Type", + "(e)" + ], + "image_footnote": [], + "bbox": [ + 94, + 255, + 290, + 364 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/27e34b5eb58b0121bbaf46bb74036621e70a7cbdfb626ad9e6de631beade163d.jpg", + "image_caption": [ + "Distribution of Vehicle Color", + "(f)" + ], + "image_footnote": [], + "bbox": [ + 292, + 255, + 488, + 369 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/97e128641af1f8169fae5af1768c0fc7e4b829b54a37555caeef39f3e8b71957.jpg", + "image_caption": [ + "Distribution of Orientation Class", + "(g)" + ], + "image_footnote": [], + "bbox": [ + 496, + 255, + 699, + 371 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ff207fa11f345721ae98eed1447981517ea1d712e0df69e5228509cf987ab27a.jpg", + "image_caption": [ + "Orientation Classification Criteria", + "(h)", + "Figure 3. The statistics and annotation attributes of UCM-VeID V2 dataset. Notation that, the 'k' in (c) means the unit of 1000 pixels." + ], + "image_footnote": [], + "bbox": [ + 720, + 251, + 879, + 373 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "classes, shown in Fig. 3(f), ordered by proportion as follows: white, black, gray, red, dark gray, cyan, brown, blue, golden, yellow, others, green, and orange. Despite some imbalance, this distribution mirrors real-world vehicle color trends.", + "bbox": [ + 104, + 439, + 482, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Vehicle Orientation. Based on [22] and Fig. 3(h), we define eight orientation categories. Fig. 3(g) shows that orientations 2, 4, 6, and 8 are most common, a result of the UAV's flexible viewing angles. This orientation imbalance highlights the dataset's realism and the challenges in UAV-based Re-ID tasks.", + "bbox": [ + 91, + 515, + 482, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Comparison with Other Datasets", + "text_level": 1, + "bbox": [ + 89, + 619, + 379, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast to other vehicle Re-ID datasets, as detailed in Tab. 1, our dataset has several unique characteristics:", + "bbox": [ + 89, + 642, + 482, + 672 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Multi modal information. Compared with RGB-based datasets, such as VRAI [31], UAV-VeID [29], UCMVeID [22] and UCM-VeID V2 integrate both RGB and IR modalities, facilitating cross-modality vehicle Re-ID tasks, which holds promise as a viable solution to achieve round-the-clock vehicle Re-Identification.", + "2. Flexible viewpoint, scale and orientations. The UCMVeID V2 dataset, captured via a drone-mounted camera, offers varied perspectives, scales, and orientations that differ from fixed-camera datasets like RGBN300 [18] and MSVR310 [40]. This diversity substantially enriches the dataset, making it a valuable resource for robust vehicle Re-ID research.", + "3. Overall enhancement. Compared to UCM-VeID, the UCM-VeID V2 dataset offers a substantial increase in" + ], + "bbox": [ + 89, + 674, + 482, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "data volume and significant improvements in several dimensions, including additional cameras and platforms, extended capture times, a broader variety of weather conditions, and a wider range of shooting scenarios, as shown in Fig. 1. These enhancements make it a richer and more comprehensive dataset, supporting diverse and robust data-driven research and applications.", + "bbox": [ + 531, + 439, + 906, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Proposed pre-training method", + "text_level": 1, + "bbox": [ + 511, + 561, + 790, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Overall Model Structure.", + "text_level": 1, + "bbox": [ + 511, + 585, + 743, + 601 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proposed pre-training VT-ReID method is illustrated in Fig. 4. Firstly, style learning and image generation techniques are used to create pairs of RGB and IR images. Then, we segment these images into patches and randomly swap patches between different modalities to create two new patch-mixed images. Finally, the patch-mixed images are fed into the encoder to complete Patch-Mixed Image Reconstruction (PMIR), Modality Discrimination Adversarial Learning (MDAL) and Modality-Augmented Contrasting Learning (MACL) tasks, where PMIR and MDAL are adopted to learn modality-invariant features, and MACL is utilized to enhance ID discriminability.", + "bbox": [ + 511, + 609, + 906, + 790 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Patch-Mixed Unified Embedding", + "text_level": 1, + "bbox": [ + 511, + 801, + 802, + 818 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross-modality data generation. Given the limited size of Re-ID dataset, we first train a CycleGAN [41] to generate cross-modality versions of ImageNet. Specifically, we use CycleGAN to learn style mapping between RGB and IR images from the UCM-VeID V2 dataset with a gener", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22289", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/aada07c53c2ccd2652fbc21c54a0bfdb8be043485aaae892ad35a6544e049979.jpg", + "image_caption": [ + "Figure 4. Pre-training paradigm, which comprises preprocessing and three self-supervised tasks. Preprocessing involves obtaining paired RGB and IR images through style learning and image generation, followed by patches segmentation and swapping, PMIR and MDAL then focus on learning modality-invariant features, while MACL aims to enhance the ID discriminability." + ], + "image_footnote": [], + "bbox": [ + 101, + 94, + 903, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ator $G$ . To avoid data leakage, we use ImageNet dataset for pre-training. For an input RGB image $\\pmb{x}_i^{\\mathrm{RGB}}$ in the ImageNet, we generate the corresponding IR image as $\\pmb{x}_i^{\\mathrm{IR}} = G(\\pmb{x}_i^{\\mathrm{RGB}})$ . The resulting cross-modality ImageNet dataset is then $\\{(\\pmb{x}_i^{\\mathrm{RGB}},\\pmb{x}_i^{\\mathrm{IR}})\\}_{i = 1}^N$ , where $N$ is the total number.", + "bbox": [ + 89, + 367, + 482, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Random patch mixture. Our method aims to teach the backbone to fully leverage the cross-modality information for modality bias elimination. Therefore, for the data input aspect, we mixture the RGB and IR image pair in a patch swapping manner. Mathematically, given a RGB-IR image pair $(\\pmb{x}_i^{\\mathrm{RGB}},\\pmb{x}_i^{\\mathrm{IR}})$ , we first establish a patch mask $M$ with the same size as $\\pmb{x}_i^{\\mathrm{RGB}}(\\pmb{x}_i^{\\mathrm{IR}})$ that consists of $K\\times K$ regular non-overlapping patches. With a predefined mask ratio $\\eta$ , $\\lfloor \\eta \\times K\\times K\\rfloor$ patches in $M$ are randomly selected following a uniform distribution and set as all-zeros matrices. In this paper, we predefined a mask ratio $\\eta = 0.75$ . A higher $\\eta$ leads to more patch exchanges, increases the difficulty of the reconstruction task, and enables the model to learn more shared features. Then, the process of the proposed random path mixture can be formulated as:", + "bbox": [ + 89, + 443, + 483, + 670 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {\\boldsymbol {x}} _ {i} ^ {\\mathrm {R G B}} = \\boldsymbol {x} _ {i} ^ {\\mathrm {R G B}} \\odot \\boldsymbol {M} + \\boldsymbol {x} _ {i} ^ {\\mathrm {I R}} \\odot (1 - \\boldsymbol {M}) \\\\ \\hat {\\boldsymbol {x}} _ {i} ^ {\\mathrm {I R}} = \\boldsymbol {x} _ {i} ^ {\\mathrm {I R}} \\odot \\boldsymbol {M} + \\boldsymbol {x} _ {i} ^ {\\mathrm {R G B}} \\odot (1 - \\boldsymbol {M}), \\end{array} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 675, + 480, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, based on the Eq. (1), the cross-modality patch-mixed unified embeddings are finally achieved as:", + "bbox": [ + 89, + 719, + 482, + 750 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {f} _ {i} ^ {\\mathrm {R G B}}, \\boldsymbol {f} _ {i} ^ {\\mathrm {I R}} = \\boldsymbol {F} \\left(\\hat {\\boldsymbol {x}} _ {i} ^ {\\mathrm {R G B}}\\right), \\boldsymbol {F} \\left(\\hat {\\boldsymbol {x}} _ {i} ^ {\\mathrm {I R}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 755, + 482, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $F(\\cdot)$ is the feature extractor and ResNet50 [11] is employed in this paper. $f_{i}^{\\mathrm{RGB}}$ and $f_{i}^{\\mathrm{IR}}$ are the latent embeddings of mixed images.", + "bbox": [ + 89, + 780, + 483, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Self-supervised Multi-task Learning.", + "text_level": 1, + "bbox": [ + 89, + 833, + 408, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the proposed unified embedding module, we can acquire the latent features containing cross-modality information. However, these features can not be directly employed", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to the Re-ID task due to modality-bias information and insufficient discriminability [30, 34, 37]. To address above these issues, we devise three self-supervised tasks.", + "bbox": [ + 511, + 367, + 903, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Patch-Mixed Image Reconstruction. In this task, we utilize the mixed features, $f_{i}^{\\mathrm{RGB}}$ and $f_{i}^{\\mathrm{IR}}$ , to reconstruct the input images, $x_{i}^{\\mathrm{RGB}}$ and $x_{i}^{\\mathrm{IR}}$ , respectively. This approach is motivated by the inherent advantage of image reconstruction as a naturally unsupervised learning task, eliminating the need for additional labels and making it suitable for training large models. Moreover, for RGB image reconstruction, $f_{i}^{\\mathrm{RGB}}$ contains minimal RGB information but abundant IR information, compelling the network to focus on modality-shared features. The limited RGB information acts as style information that supports the reconstruction of the RGB image. Similarly, for IR image reconstruction, $x_{i}^{\\mathrm{IR}}$ using $f_{i}^{\\mathrm{IR}}$ , where the feature contains mostly RGB information with only a small amount of IR data, thus also ensuring the model learns to emphasize cross-modality features effectively. According to above analysis, the process of patch-mixed image reconstruction can be formulated as:", + "bbox": [ + 511, + 412, + 906, + 670 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{r l} \\bar {\\boldsymbol {x}} _ {i} ^ {\\mathrm {R G B}} & = \\boldsymbol {H} _ {\\mathrm {R G B}} \\left(\\boldsymbol {f} _ {i} ^ {\\mathrm {R G B}}\\right) \\\\ \\bar {\\boldsymbol {x}} _ {i} ^ {\\mathrm {I R}} & = \\boldsymbol {H} _ {\\mathrm {I R}} \\left(\\boldsymbol {f} _ {i} ^ {\\mathrm {I R}}\\right), \\end{array} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 680, + 903, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\bar{\\pmb{x}}_i^{\\mathrm{RGB}}$ and $\\bar{\\pmb{x}}_i^{\\mathrm{IR}}$ are the reconstructed RGB and IR images. $H_{\\mathrm{RGB}}(\\cdot)$ and $H_{\\mathrm{IR}}(\\cdot)$ are networks for RGB and IR image reconstruction respectively. Finally, a pixel-wise image reconstruction loss with $l_{2}$ -norm is employed to guide the reconstruction process as:", + "bbox": [ + 511, + 729, + 906, + 808 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r e c}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\| \\bar {\\boldsymbol {x}} _ {i} ^ {\\mathrm {R G B}} - \\boldsymbol {x} _ {i} ^ {\\mathrm {R G B}} \\| _ {2} ^ {2} + \\| \\bar {\\boldsymbol {x}} _ {i} ^ {\\mathrm {I R}} - \\boldsymbol {x} _ {i} ^ {\\mathrm {I R}} \\| _ {2} ^ {2}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 818, + 903, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Modality Discrimination Adversarial Learning. In this task, we introduce adversarial training to encourage", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22290", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the learned features to be modality-invariant. Specifically, we employ a discriminator $D(\\cdot)$ to distinguish between the mixed features $f_{i}^{\\mathrm{RGB}}$ and $f_{i}^{\\mathrm{IR}}$ . Initially, we fix the feature extractor $F$ and train the discriminator $D$ to correctly classify the mixed features based on their corresponding modality labels. This step ensures that $D$ can accurately distinguish between features from different modalities. The objective for the discriminator $D$ can be formulated as:", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D}} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left[ y _ {i} \\log \\boldsymbol {D} (\\boldsymbol {f} _ {i}) + (1 - y _ {i}) \\log (1 - \\boldsymbol {D} (\\boldsymbol {f} _ {i})) \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 223, + 482, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $y_{i}$ is the modality label, $f_{i}$ represents either $f_{i}^{\\mathrm{RGB}}$ or $f_{i}^{\\mathrm{IR}}$ , and $K$ is the batch size. In the second step, we fix the discriminator $D$ and update the feature extractor $F$ to generate mixed features that the discriminator cannot distinguish. This step forces the feature extractor to produce modality-invariant features, effectively minimizing the modality bias in the learned features. The objective for the feature extractor $F$ during this phase is to maximize the classification error of $D$ as:", + "bbox": [ + 89, + 252, + 483, + 388 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {F}} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left[ \\left(1 - y _ {i}\\right) \\log \\boldsymbol {D} (\\boldsymbol {f} _ {i}) + y _ {i} \\log \\left(1 - \\boldsymbol {D} (\\boldsymbol {f} _ {i})\\right) \\right]. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 400, + 482, + 419 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The adversarial training process iterates between these two steps until the model converges, achieving a balance where the discriminator $D$ cannot reliably distinguish the modality of the features generated by $F$ .", + "bbox": [ + 89, + 430, + 482, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Modality-Augmented Contrasting Cluster. In this task, we focus on enhancing the discriminative power of the learned features by employing a contrastive learning approach. Unlike traditional methods that rely on data augmentation to generate negative samples, our approach does not introduce additional data augmentation but instead utilizes the inherent modality-specific features for contrastive learning. We treat $f_{i}^{\\mathrm{RGB}}$ and $f_{i}^{\\mathrm{IR}}$ as positive pairs within the contrastive learning framework. The contrastive loss function aims to maximize the similarity between the mixed features of different modalities, which can be formulated as:", + "bbox": [ + 89, + 491, + 483, + 656 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n t r a s t}} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {f} _ {i} ^ {\\mathrm {R G B}} , \\boldsymbol {f} _ {i} ^ {\\mathrm {I R}}\\right)\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {f} _ {i} ^ {\\mathrm {R G B}}, \\boldsymbol {f} _ {j} ^ {\\mathrm {I R}}\\right)\\right)}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 667, + 482, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathrm{sim}(\\cdot ,\\cdot)$ denotes a similarity measure (such as cosine similarity) between features. The loss function encourages the feature extractor to learn representations that pull together positive pairs $(f_{i}^{\\mathrm{RGB}}$ and $f_{i}^{\\mathrm{IR}})$ while pushing apart other features. By iteratively optimizing this contrastive loss, the feature extractor $\\pmb{F}$ is guided to produce more discriminative features, thereby improving the overall performance of the cross-modal image re-identification task.", + "bbox": [ + 89, + 720, + 482, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In summary, the overall loss function for the entire training process can be expressed as:", + "bbox": [ + 89, + 842, + 482, + 872 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l l}} = \\mathcal {L} _ {\\text {r e c}} + \\left(\\mathcal {L} _ {\\mathrm {D}} + \\mathcal {L} _ {\\mathrm {F}}\\right) + \\mathcal {L} _ {\\text {c o n t r a s t}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 885, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 90, + 645, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Implementation and Evaluation Protocols", + "text_level": 1, + "bbox": [ + 511, + 114, + 870, + 131 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pre-training Stage. For pre-training, we implement CPSL using PyTorch on an NVIDIA RTX A100. Optimization is conducted with an SGD optimizer (momentum of 0.9, weight decay of 0.0005), an initial learning rate of 0.1 with linear warm-up, and a cosine decay schedule without restarts. Data augmentation includes random grayscale, horizontal flipping, Gaussian filtering, and interpolation sampling. Images are resized to $256 \\times 256$ , and the mask size $N$ is set to 4.", + "bbox": [ + 511, + 137, + 903, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fine-Tuning Stage. For fine-tuning, we implement VT-ReID models using PyTorch on an NVIDIA GeForce RTX 3090 GPU. Images are resized to $180 \\times 256$ . Optimization uses an SGD optimizer (momentum of 0.9, weight decay of 0.0005) with an initial learning rate of 0.1 and linear warm-up over 80 epochs. The model is initialized with CPSL checkpoints while other hyper-parameters and data augmentations remain unchanged, as this setup is not specifically tuned for CPSL. Further optimization may improve results, which we leave for future work.", + "bbox": [ + 511, + 273, + 905, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Protocols. Following [34], we use the cumulative matching characteristic (CMC) curve and mean average precision (mAP) for evaluation. CMC assesses rank-k accuracy, with rank-1, rank-10, and rank-20 scores reported in our experiments. mAP reflects the overall retrieval performance across all query samples.", + "bbox": [ + 511, + 426, + 905, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparison with pre-trained weights", + "text_level": 1, + "bbox": [ + 511, + 527, + 836, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate the effectiveness of our proposed pre-training method, we employ three different backbone initialization strategies during fine-tuning: random initialization (Random Init.), initialization with ImageNet pre-trained weights (ImageNet-1K), and initialization using the proposed pretraining weights (CPSL). To enhance experimental rigor, we perform supervised fine-tuning with four state-of-the-art VT-ReID methods: AGW [37], DDAG [36], MMD-ReID [17], and DEEN [39].", + "bbox": [ + 511, + 551, + 905, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "From Tab. 2, we can observe that the VT-ReID models trained from scratch generally under-performance, with all methods failing to converge. The primary cause of the gradient explosion is attributed to the triplet loss, suggesting that randomly initialized models struggle to establish semantically meaningful Euclidean feature spaces.", + "bbox": [ + 511, + 686, + 905, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method CPSL achieves superior results across all testing and query modes. In contrast to ImageNet-1K pretraining, we do not rely on additional training data or incorporate any supervised training processes. The superior performance is primarily attributed to the design of our self-supervised tasks, which account for modality bias and eliminate cross-modality differences.", + "bbox": [ + 511, + 779, + 905, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It can also be observed that CPSL consistently improves", + "bbox": [ + 531, + 885, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22291", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/90035aec1a424133b4b2479dc66f5859239216b91ddbffa0e96675f88424aa99.jpg", + "table_caption": [], + "table_footnote": [ + "* The bold entries represent the best performance in table." + ], + "table_body": "
MethodVenuePre-TrainSingle-shotMulti-shot
IR2RGBRGB2IRIR2RGBRGB2IR
rank1rank10mAPrank1rank10mAPrank1rank10mAPrank1rank10mAP
AGW [37]TPAMI 2021Random Init.4.3518.759.384.8919.79.297.2227.264.826.6726.175.53
ImageNet-1K15.7545.9825.7416.1747.5826.5420.158.1216.2120.4558.1618.3
CPSL23.4753.0633.6925.3658.5436.5731.9167.5424.1332.6468.0724.01
DDAG [36]ECCV 2020Random Init.5.2222.8111.485.0622.6911.228.5532.725.847.0630.056.25
ImageNet-1K20.4648.7830.1219.4649.1529.2127.5864.7520.2826.6965.5422.87
CPSL22.2350.8631.9923.4551.3332.9431.9667.9421.731.1666.5824.51
MMD-ReID [17]BMVC 2021Random Init.4.3415.279.864.9218.539.935.9919.3710.785.6920.3410.42
ImageNet-1K18.9348.6428.9919.2749.1429.0822.0455.8719.6422.9659.2731.41
CPSL21.9851.8532.3922.5253.3831.9928.3462.0922.129.0462.8521.8
DEEN [39]CVPR 2023Random Init.7.2931.7415.246.1127.0613.297.8837.447.487.3932.817.34
ImageNet-1K18.2653.0130.3220.4656.2332.0923.8364.1422.0622.6566.6121.43
CPSL22.5759.8531.0923.759.5633.3525.6267.5223.4424.5368.6323.97
", + "bbox": [ + 91, + 88, + 924, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "performance across various VT-ReID methods, which indicates that the weights generated by CPSL exhibit strong robustness and adaptability. This suggests that the CPSL method is effective in enhancing the generalization capabilities of the models, allowing them to perform well under different conditions and with diverse Re-ID techniques.", + "bbox": [ + 89, + 415, + 483, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Comparisons With Other SSL Methods", + "text_level": 1, + "bbox": [ + 89, + 515, + 429, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare CPSL with multiple ResNet50 pre-trained models on the UCM-VeID V2. The methods involved can be categorized into four groups: (1) method based on VT-ReID fine-tuning (MMGL [30]), (2) method based pretext task design (Jigsaw [26], Rotation [8]), (3) method based self-supervised learning (SSL) pre-training (InstDisc [35], SimCLR [3], SwAV [2], BYOL [9], SimSiam [4], MoCoV2 [5], DenseCL [33], Relative-Loc[7]).", + "bbox": [ + 89, + 537, + 483, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Tab. 3, MMGL performs suboptimally. This is partly because MMGL is tailored for pedestrian targets, using a top-down human segmentation approach that struggles with vehicles' variable orientations. Additionally, the released MMGL weights were fine-tuned on pedestrian VT-ReID datasets, embedding pedestrian-specific features that do not generalize well to vehicle targets.", + "bbox": [ + 89, + 659, + 482, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SSL methods share a similar goal with our approach by aiming to endow the encoder with discriminative capabilities. However, methods such as SwAV [2] and BYOL [9] perform poorly, as indicated by the stagnant triplet loss during fine-tuning, suggesting these methods fail to establish a meaningful Euclidean space from the outset. Although some SSL methods outperform those based on pretext task design, they still underperform compared to CPSL. This is because they focus solely on target discriminability through", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a44db2b428e177baf2971c10ae46443b6cf59d0e5518781d1417330cb4108b53.jpg", + "table_caption": [ + "Table 2. Comparisons between the proposed CPSL and some state-of-the-art methods on UCM-VeID V2 datasets. The rank-k (%) and mAP (%) are demonstrated." + ], + "table_footnote": [], + "table_body": "
MethodRank1Rank10Rank20mAP
method based on VT-ReID fine-tuning:
MMGL [30]21.8256.5668.0316.14
method based pretext task design:
Jigsaw [26]21.2456.7767.2515.66
RotationNet [8]26.6263.0372.7719.32
method based on SSL pre-training:
InstDisc [35]24.0962.0773.1620.04
SimCLR [3]12.4348.5762.1511.17
SwAV [2]6.5728.8940.225.17
BYOL [9]6.8925.5136.875.17
SimSiam [4]27.2465.7473.1322.36
MoCo V2 [5]28.8967.4674.9822.44
DenseCL [33]27.8964.9975.1222.41
Relative-Loc[7]18.9155.7968.0615.53
ours32.6468.0776.3224.01
", + "bbox": [ + 517, + 412, + 901, + 659 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Comparison with other SSL methods. The rank-k (%) and mAP (%) are demonstrated.", + "bbox": [ + 511, + 669, + 906, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "contrastive learning without addressing modality differences. In contrast, our PMIR and MDAL modules effectively mitigate cross-modality differences, enabling shared feature learning.", + "bbox": [ + 511, + 727, + 906, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 801, + 666, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To assess the effectiveness of each component in our method, we conducted ablation studies on the proposed pretraining method, removing specific modules while keeping all other experimental settings constant. Our method includes three key modules: PMIR (R), MDAL (D), and", + "bbox": [ + 511, + 825, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22292", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e4f90f269178d1b0c81e37b87b9bdd8f4c427c03e03919d6c462775e11ed28b0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSingle-shotMulti-shot
rank1rank10mAPrank1rank10mAP
Random Init.4.8919.79.296.6726.175.53
ImageNet-1K16.1747.5826.5420.4558.1618.3
R7.6426.9312.219.3342.7210.59
R+D14.6349.5722.4618.8854.9518.32
R+C22.4955.4830.3428.9765.1621.96
R+D+C25.3658.5436.5732.6468.0724.01
", + "bbox": [ + 91, + 88, + 483, + 218 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MACL (C). Here, R and D target modality alignment, while C enhances discriminative representation. The ablation study results are shown in Tab. 4.", + "bbox": [ + 88, + 282, + 482, + 329 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of PMIR. PMIR aims to create a shared latent representation space for RGB and IR modalities via reconstruction tasks. As shown in Tab. 4, compare to the results of Random Init., RANK1 and mAP improve by $2.57\\%$ and $2.92\\%$ under the single-shot query mode, indicating PMIR's effectiveness in reducing modality discrepancies.", + "bbox": [ + 89, + 330, + 483, + 420 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of MDAL. MDAL uses modality adversarial learning to further align modalities. Tab. 4 shows a RANK1 and mAP increase of $6.99\\%$ and $10.25\\%$ under single-shot query mode, demonstrating that MDAL helps guide the encoder toward shared feature learning.", + "bbox": [ + 89, + 421, + 483, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of MACL. MACL employs modality contrastive learning to enhance feature clustering and discriminative power. As evidenced in Tab. 4, RANK1 and mAP improve by $7.86\\%$ and $7.88\\%$ in single-shot query mode, confirming MACL's success in learning discriminative features for VT-ReID.", + "bbox": [ + 89, + 496, + 483, + 585 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/19e55203ceae4a77f105f06993cda464c4fdf56b19efbc01668663f03c8ca43b.jpg", + "table_caption": [ + "Table 4. The quantitative results of ablation studies. The rank-k $(\\%)$ and mAP $(\\%)$ are demonstrated." + ], + "table_footnote": [], + "table_body": "
Pn2 × 23 × 34 × 45 × 5
rank124.6627.3232.6429.05
mAP16.3820.9124.0122.44
", + "bbox": [ + 91, + 597, + 483, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Different partition stripe numbers ${P}_{n}\\left( {N \\times N}\\right)$ . The rank-k (%) and mAP (%) are demonstrated.", + "bbox": [ + 89, + 667, + 482, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Parameter Analysis", + "text_level": 1, + "bbox": [ + 89, + 726, + 279, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The partition stripe number $P_{n}$ directly determines the size of the image blocks to be swapped and is an important hyperparameter of our method. To verify the impact of this hyperparameter, we conducted experiments with different values for the partition stripe number, with the relevant results presented in Tab. 5. From the table, it can be observed that as the partition stripe number increases, the model's performance initially improves before declining. This is because, as the partition stripe number increases, the semantic content within each image block transitions from being more", + "bbox": [ + 88, + 750, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6f3023b35357547ed5b2e3343fcb0c70f381171787f24a24e2e9786da093930e.jpg", + "image_caption": [ + "(a) ImageNet Pre-Training" + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 694, + 193 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/238a5c5513c100a4ca072473311bcbd5f265e17604c36a57f048aa1c3fcbfe4a.jpg", + "image_caption": [ + "(b) PMSL Pre-Training", + "Figure 5. Visualization of the pre-learned features. Different colors represent distinct identities, while circle and cross symbols denote RGB and IR modalities, respectively. The experiment is conducted on UCM-VeID V2 dataset with the AGW baseline [37]." + ], + "image_footnote": [], + "bbox": [ + 694, + 88, + 867, + 193 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/23e415c255f9555e89e81104ef8eb30a508277ab37c7089ab7c19fcc34c15183.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 874, + 89, + 901, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "discrete to becoming sparser. This indicates that in practical applications, appropriate adjustment of this parameter is necessary to achieve optimal performance.", + "bbox": [ + 511, + 301, + 906, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.6. Visualization of the feature distribution", + "text_level": 1, + "bbox": [ + 511, + 358, + 852, + 373 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To better understand the proposed pre-training paradigm, we visualize the t-SNE distributions of features pre-learned using ImageNet-1K and the CPSL checkpoints on the UCM-VeID V2 dataset, utilizing AGW [37]. As shown in Fig. 5, we randomly select 15 identities to visualize all corresponding images.", + "bbox": [ + 511, + 380, + 905, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 5 demonstrates that, the modal gap between PMSL features is smaller than that between ImageNet features (highlighted by the red circle). Specifically, PMSL features show less variation across modalities, indicating better consistency and robustness when processing RGB and IR images. In contrast, ImageNet-pretrained features exhibit significant modal variation, which can negatively impact performance in VT-ReID tasks. Thus, PMSL not only eliminates cross-modality discrepancy but also improves the model's discriminative ability for downstream tasks, providing stronger support for cross-modal vehicle re-identification.", + "bbox": [ + 511, + 470, + 905, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 667, + 633, + 683 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel dataset that expands existing ones in terms of scale, acquisition platforms, collection scenarios, time, and climate conditions. To address the MBT problem, we break it into two sub-problems: modality discrepancy and insufficient discriminability. We propose an end-to-end cross-modal pre-training framework to address these issues. To reduce modality discrepancies, the framework incorporates self-supervised tasks based on image block swapping and modality adversarial learning. For enhanced discriminability, we design an unsupervised clustering-guided multi-modal contrastive learning approach. Extensive experiments on the new dataset validate the effectiveness of our method.", + "bbox": [ + 511, + 691, + 905, + 871 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by the National Natural Science Foundation of China (62201586).", + "bbox": [ + 511, + 872, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22293", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 119, + 187, + 136 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yan Bai, Yihang Lou, Feng Gao, Shiqi Wang, Yuwei Wu, and Ling-Yu Duan. Group-sensitive triplet embedding for vehicle reidentification. IEEE Trans. Multim., 20(9):2385-2399, 2018. 1", + "[2] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In Proc. Adv. Neural Inf. Process. Syst., pages 9912-9924, 2020. 7", + "[3] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In Proc. Int. Conf. Mach. Learn., pages 1597-1607, 2020. 7", + "[4] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 15750-15758, 2021. 7", + "[5] Xinlei Chen, Haoqi Fan, Ross B. Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. CoRR, abs/2003.04297, 2020. 7", + "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proc. IEEE Conf. Comput. Vis. Pattern Recognition., pages 248-255, 2009. 2", + "[7] Carl Doersch, Abhinav Gupta, and Alexei A. Efros. Unsupervised visual representation learning by context prediction. In Proc. IEEE Int. Conf. Comput. Vis., pages 1422-1430. IEEE Computer Society, 2015. 7", + "[8] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In Proc. Int. Conf. Learn. Represent., 2018. 3, 7", + "[9] Jean-Bastien Grill, Florian Strub, Florent Alché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Ávila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent - A new approach to self-supervised learning. In Proc. Adv. Neural Inf. Process. Syst., pages 21271-21284, 2020. 3, 7", + "[10] Haiyun Guo, Chaoyang Zhao, Zhiwei Liu, Jinqiao Wang, and Hanqing Lu. Learning coarse-to-fine structured feature embedding for vehicle re-identification. In Proc. AAAI Conf. Artif. Intell., pages 6853–6860, 2018. 2", + "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 770-778, 2016. 5", + "[12] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 9726-9735, 2020. 3", + "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 15979-15988, 2022. 2" + ], + "bbox": [ + 93, + 146, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 15979-15988, 2022. 3", + "[15] Yu Huang, Junyang Lin, Chang Zhou, Hongxia Yang, and Longbo Huang. Modality competition: What makes joint training of multi-modal network fail in deep learning? (provably). In Proc. Int. Conf. Mach. Learn., pages 9226-9259, 2022. 2", + "[16] Yan Huang, Qiang Wu, Jingsong Xu, Yi Zhong, Peng Zhang, and Zhaoxiang Zhang. Alleviating modality bias training for infrared-visible person re-identification. IEEE Trans. Multim., 24:1570-1582, 2022. 2", + "[17] Chaitra Jambigi, Ruchit Rawal, and Anirban Chakraborty. MMD-ReID: A simple but effective solution for visible-thermal person reid. In Proc. BMVC, pages 12026–12035, 2021. 6, 7", + "[18] H. Li, C. Li, X. Zhu, A. Zheng, and B. Luo. Multi-spectral vehicle re-identification: A challenge. In Proc. AAAI Conf. Artif. Intell., pages 11345–11353, 2020. 4", + "[19] Chunsheng Liu, Ye Song, Faliang Chang, Shuang Li, Ruimin Ke, and Yinhai Wang. Posture calibration based cross-view and hard-sensitive metric learning for UAV-based vehicle re-identification. IEEE Trans. Inf. Forensics Secur., 23:19246-19257, 2022. 1, 2", + "[20] Hongye Liu, Yonghong Tian, Yaowei Wang, Lu Pang, and Tiejun Huang. Deep relative distance learning: Tell the difference between similar vehicles. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 2167-2175, 2016. 2", + "[21] Xinchen Liu, Wu Liu, Tao Mei, and Huadong Ma. A deep learning-based approach to progressive vehicle re-identification for urban surveillance. In Proc. Eur. Conf. Comput. Vis., pages 869–884, 2016. 2", + "[22] Xingyue Liu, Jiahao Qi, Chen Chen, Kangcheng Bin, and Ping Zhong. Beyond sharing weights in decoupling feature learning network for UAV rgb-infrared vehicle re-identification. IEEE Trans. Multim., 2024. 1, 2, 3, 4", + "[23] Yihang Lou, Yan Bai, Jun Liu, Shiqi Wang, and Lingyu Duan. VERI-Wild: A large dataset and a new method for vehicle re-identification in the wild. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3235-3243, 2019. 2", + "[24] Mingming Lu, Yongchuan Xu, and Haifeng Li. Vehicle re-identification based on UAV viewpoint: Dataset and method. Remote. Sens., 14(18):4603, 2022. 1, 2", + "[25] Tomás Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In Proc. Int. Conf. Learn. Represent., 2013. 3", + "[26] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In Proc. Eur. Conf. Comput. Vis., pages 69-84, 2016. 3, 7", + "[27] Xiaokang Peng, Yake Wei, Andong Deng, Dong Wang, and Di Hu. Balanced multimodal learning via on-the-fly gradient modulation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 8228-8237, 2022. 2", + "[28] Shangzhi Teng, Shiliang Zhang, Qingming Huang, and Nicu Sebe. Multi-view spatial attention embedding for vehicle re-identification. IEEE Trans. Circuits Syst. Video Technol., 31 (2):816-827, 2021. 1" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22294", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Shangzhi Teng, Shiliang Zhang, Qingming Huang, and Nicu Sebe. Viewpoint and scale consistency reinforcement for UAV vehicle re-identification. Int. J. Comput. Vis., 129(3): 719-735, 2021. 1, 2, 4", + "[30] Lin Wan, Qianyan Jing, Zongyuan Sun, Chuang Zhang, Zhihang Li, and Yehansen Chen. Self-supervised modality-aware multiple granularity pre-training for rgb-infrared person re-identification. IEEE Trans. Inf. Forensics Secur., 18: 3044–3057, 2023. 2, 5, 7", + "[31] Peng Wang, Bingliang Jiao, Lu Yang, Yifei Yang, Shizhou Zhang, Wei Wei, and Yanning Zhang. Vehicle re-identification in aerial imagery: Dataset and approach. In Proc. IEEE Int. Conf. Comput. Vis., pages 460–469, 2019. 1, 2, 4", + "[32] Weiyao Wang, Du Tran, and Matt Feiszli. What makes training multi-modal classification networks hard? In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 12692-12702, 2020. 2", + "[33] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3024-3033. Computer Vision Foundation / IEEE, 2021. 7", + "[34] Ancong Wu, Wei-Shi Zheng, Hong-Xing Yu, Shaogang Gong, and Jianhuang Lai. RGB-infrared cross-modality person re-identification. In Proc. IEEE Int. Conf. Comput. Vis., pages 5390-5399, 2017. 5, 6", + "[35] Zhirong Wu, Yuanjun Xiong, Stella X. Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3733-3742, 2018. 7", + "[36] Mang Ye, Jianbing Shen, David J. Crandall, Ling Shao, and Jiebo Luo. Dynamic dual-attentive aggregation learning for visible-infrared person re-identification. Proc. Eur. Conf. Comput. Vision, 12362:229–247, 2020. 6, 7", + "[37] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven C. H. Hoi. Deep learning for person re-identification: A survey and outlook. IEEE Trans. Pattern Anal. Mach. Intell., 44:2872–2893, 2020. 2, 5, 6, 7, 8", + "[38] Richard Zhang, Phillip Isola, and Alexei A. Efros. Colorful image colorization. In Proc. Eur. Conf. Comput. Vis., pages 649-666, 2016. 3", + "[39] Yukang Zhang and Hanzi Wang. Diverse embedding expansion network and low-light cross-modality benchmark for visible-infrared person re-identification. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 2153–2162, 2023. 6, 7", + "[40] Aihua Zheng, Xianpeng Zhu, Chenglong Li, Jin Tang, and Jixin Ma. Cross-directional consistency network with adaptive layer normalization for multi-spectral vehicle re-identification and a high-quality benchmark. Inf. Fusion, 100:101901, 2023. 4", + "[41] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proc. IEEE Int. Conf. Comput. Vis., pages 2242-2251, 2017. 4" + ], + "bbox": [ + 91, + 90, + 480, + 882 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22295", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/UCM-VeID V2_ A Richer Dataset and A Pre-training Method for UAV Cross-Modality Vehicle Re-Identification/9b7d21db-8ee1-4718-8ddf-98653daa1c9e_model.json b/2025/UCM-VeID V2_ A Richer Dataset and A Pre-training Method for UAV Cross-Modality Vehicle Re-Identification/9b7d21db-8ee1-4718-8ddf-98653daa1c9e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7600ce41d502d17b19f6574d54a1c5ddc6e00636 --- /dev/null +++ b/2025/UCM-VeID V2_ A Richer Dataset and A Pre-training Method for UAV Cross-Modality Vehicle Re-Identification/9b7d21db-8ee1-4718-8ddf-98653daa1c9e_model.json @@ -0,0 +1,2519 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.146, + 0.13, + 0.853, + 0.175 + ], + "angle": 0, + "content": "UCM-VeID V2: A Richer Dataset and A Pre-training Method for UAV Cross-Modality Vehicle Re-Identification" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.204, + 0.796, + 0.24 + ], + "angle": 0, + "content": "Xingyue Liu Jiahao Qi Chen Chen KangCheng Bin Ping Zhong* National University of Defense Technology, China" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.242, + 0.817, + 0.258 + ], + "angle": 0, + "content": "{liuxingyue18, qijiahao1996, chenchen21c, binkc21, zhongping}@nudt.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.323, + 0.485, + 0.611 + ], + "angle": 0, + "content": "Cross-Modality Re-Identification (VT-ReID) aims to achieve around-the-clock target matching, benefiting from the strengths of both RGB and infrared (IR) modalities. However, the field is hindered by limited datasets, particularly for vehicle VT-ReID, and by challenges such as modality bias training (MBT), stemming from biased pre-training on ImageNet. To tackle the above issues, this paper introduces an dataset benchmark, named UCM-VeID V2, for vehicle VT-ReID, and proposes a new self-supervised pre-training method, Cross-Modality Patch-Mixed Self-supervised Learning (PMSL). UCM-VeID V2 dataset features a significant increase in data volume, along with enhancements in multiple aspects. PMSL addresses MBT by learning modality-invariant features through Patch-Mixed Image Reconstruction (PMIR) and Modality Discrimination Adversarial Learning (MDAL), and enhances discriminability with Modality-Augmented Contrasting Cluster (MACC). Comprehensive experiments are carried out to validate the proposed method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.64, + 0.222, + 0.656 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.665, + 0.483, + 0.816 + ], + "angle": 0, + "content": "Cross-modality vehicle re-identification (ReID), also referred to as vehicle visible-infrared (VI) ReID [22], involves retrieving a matching vehicle image from a gallery set captured in the visible (RGB) or infrared (IR) spectrum, given a query vehicle image in the alternate modality. This process, where images are obtained from cameras of differing modalities, has garnered significant attention in applications such as video surveillance and public security [1, 28], primarily due to its advantages in enabling continuous, intelligent monitoring across day and night conditions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.817, + 0.484, + 0.878 + ], + "angle": 0, + "content": "Despite its importance, research on vehicle VT-ReID remains relatively underdeveloped. A primary challenge lies in the lack of essential resources, including a comprehensive large-scale dataset and widely accessible benchmark" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.291, + 0.906, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.661, + 0.358, + 0.762, + 0.368 + ], + "angle": 0, + "content": "(a) Multiple Platforms" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.37, + 0.905, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.437, + 0.758, + 0.447 + ], + "angle": 0, + "content": "(b) Wider time spans" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.516, + 0.776, + 0.526 + ], + "angle": 0, + "content": "(c) Richer weather conditions" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.528, + 0.905, + 0.595 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.595, + 0.753, + 0.604 + ], + "angle": 0, + "content": "(d) Diverse Scenes" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.607, + 0.905, + 0.674 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.673, + 0.737, + 0.684 + ], + "angle": 0, + "content": "(e) Samples" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.696, + 0.907, + 0.753 + ], + "angle": 0, + "content": "Figure 1. Increments and samples of UCM-VeID V2. (a)(b)(c)(d) are varies platforms, time periods, weather conditions and locations for data acquisition, respectively. (e) shows the 4 IDs with different modalities, view-points and scales samples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "results. Existing remote sensing vehicle ReID datasets, such as VRAI [31], UAV-VeID [29], VRU [24], and VeRi-UAV [19], primarily focus on the single RGB modality, lacking the infrared data necessary for cross-modality analysis. For this, [22] has introduced the first UAV RGB-IR vehicle ReID dataset, UCM-VeID, to support cross-modality vehicle ReID research. However, this dataset is constrained by its single data acquisition platform and limited number of" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.235, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "22286" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "identities, reducing its suitability for generalizing to diverse real-world scenarios. Consequently, developing large-scale and more diverse vehicle VT-ReID datasets is essential to advance research and better meet the requirements of practical applications." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.17, + 0.483, + 0.472 + ], + "angle": 0, + "content": "Another key challenge in vehicle VT-ReID development is the need for further research on VT-ReID methods, particularly to address the modality bias training (MBT) problem, which remains under-explored. As introduced by [16], MBT refers to the modality imbalance that arises when VT-ReID models are initialized with pre-trained ImageNet weights [6]. The modality imbalance stems from training on RGB-exclusive ImageNet data, which skews feature learning towards the RGB modality and reduces the IR discrimination capability. Moreover, the MBT issue is amplified in large-scale multi-modal datasets: as shown in Fig. 2, models initialized with ImageNet pre-trained weights exhibit a performance drop on larger datasets, particularly in terms of discriminative ability. The potential underlying cause of the above phenomenon is that ImageNet pretrained weights perform well when handling infrared data with lower modality diversity. However, they lack sufficient generalization ability when faced with larger datasets that exhibit higher modality diversity, which exacerbates modality bias and degrades model performance." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.474, + 0.483, + 0.806 + ], + "angle": 0, + "content": "To address these challenges, this paper introduces a more comprehensive dataset for vehicle VT-ReID and a novel pre-training paradigm called Cross-modality Patch-mixed Self-supervised Learning (CPSL) to mitigate the MBT problem. The proposed dataset, UCM-VeID V2, includes 113,929 RGB images and 51,162 IR images of 3,407 vehicles captured by five cameras. Compared to UCM-VeID [22], it is significantly expanded across multiple dimensions, including additional cameras and acquisition platforms, extended capture times, a broader range of weather conditions, and more diverse shooting scenarios, as illustrated in Fig. 1. For the MBT challenge, inspired by [13], CPSL aligns the RGB and IR domains by incorporating a Patch-Mixed Image Reconstruction (PMIR) self-supervised task. To further eliminate the modality bias, CPSL designs a Modality Discrimination Adversarial Learning (MDAL) strategy to exact modality invariant features through an adversarial training manner. In addition, considering discrimination insufficiency, we introduce the Modality-Augmented Contrasting Learning (MACL) task, where patch-mixed images are used as augmented positive examples for contrastive learning." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.809, + 0.449, + 0.823 + ], + "angle": 0, + "content": "In summary, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.826, + 0.483, + 0.87 + ], + "angle": 0, + "content": "- We introduce the UCM-VeID V2 dataset, which significantly enhances the original UCM-VeID with greater scale and diversity across multiple dimensions." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.872, + 0.484, + 0.902 + ], + "angle": 0, + "content": "- We propose a novel self-supervised paradigm to pre-train an encoder, replacing the ImageNet-trained ResNet50, to" + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.697, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.197, + 0.68, + 0.207 + ], + "angle": 0, + "content": "(a) Training based on UCM-VeID" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.089, + 0.868, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.711, + 0.197, + 0.862, + 0.207 + ], + "angle": 0, + "content": "(b) Training based on UCM-VeID V2" + }, + { + "type": "image", + "bbox": [ + 0.875, + 0.09, + 0.903, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.218, + 0.905, + 0.329 + ], + "angle": 0, + "content": "Figure 2. The feature distribution of pre-trained embeddings is visualized. Panels (a) and (b) show the embeddings in 2D space for models trained on the UCM-VeID and UCM-VeID V2 datasets, respectively, with dots and crosses representing different identities in RGB and IR modalities. The experiment uses the AGW baseline [37]. As indicated by the red circles in (a), the same-ID samples are more compact. In contrast, the same-ID samples are more spread out in (b), which ultimately degrades Re-ID performance." + }, + { + "type": "text", + "bbox": [ + 0.53, + 0.356, + 0.687, + 0.37 + ], + "angle": 0, + "content": "address the MBT issue." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.384, + 0.664, + 0.4 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.41, + 0.726, + 0.424 + ], + "angle": 0, + "content": "2.1. Vehicle Re-ID Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.907, + 0.704 + ], + "angle": 0, + "content": "Vehicle Re-ID research has flourished over the past few years, significantly benefited by the availability of vehicle Re-ID datasets. The most widely-employed datasets are primarily captured via fixed surveillance cameras, such as VehicleID [20], VeRI-776 [21], VERI-Wild [23] and Vehicle1M [10]. Compared to fixed surveillance, UAV-based cameras offer superior mobility, flexibility, and convenience, leading to the development of key UAV-based vehicle Re-ID datasets such as VRAI [31], UAV-VeID [29], VRU [24], and VeRi-UAV [19]. Furthermore, Liu et al. [22] have advanced the first RGB-IR vehicle Re-ID dataset, named UCM-VeID, to support the research of cross-modality vehicle Re-ID. However, the UCM-VeID dataset exhibits limitations in both size and diversity, highlighting an urgent need for the further expansion and enrichment of the dataset, ensuring it remains a potent benchmark for advancing the frontiers of vehicle Re-ID research in the face of evolving challenges and complexities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.713, + 0.729, + 0.729 + ], + "angle": 0, + "content": "2.2. Modality Bias Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The MBT problem is essentially a problem of unbalanced training. Imbalanced model learning has received extensive attention in multi-modal learning [15, 32], and great progresses have been made in VT-ReID [16, 30]. To our best knowledge, there are two distinct approaches to tackle the MBT issue. The first approach is to control the training process, such as learning rate adjusting [15], gradient modulation [27], and ID-exclusive sampling strategy [16]. The second approach is to use modal balanced weights rather than ImageNet pre-trained ResNet50 avoiding MBT problems from the beginning. Wan et al. [30] devised a self-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22287" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.9, + 0.171 + ], + "angle": 0, + "content": "
DatasetVenueIdentitiesImagesModalitiesCamerasPlatformCamera ViewTarget StateColorTypeOrientation
VRAIICCV201913022137613RGB2UAVFlexibleMove97—
UAV-VeIDIJCV2021460141917RGB2UAVFlexibleMove——4
UCM-VeIDTMM202475329928RGB/IR2UAVFlexibleMove&Static1298
UCM-VeID V2ours3407165091RGB/IR5UAVFlexibleMove&Static1298
" + }, + { + "type": "table_caption", + "bbox": [ + 0.265, + 0.18, + 0.733, + 0.194 + ], + "angle": 0, + "content": "Table 1. Comparisons among other vehicle Re-ID datasets and UCM-VeID V2" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.221, + 0.484, + 0.372 + ], + "angle": 0, + "content": "supervised pre-training method to obtain a set of new unbiased weights, which achieves comparable performance to that of ImageNet-trained ResNet50. The second approach addresses the MBT problem from the outset by using pretrained weights that are modally balanced, instead of relying on subsequent adjustments. In contrast to the first approach, the second method more fundamentally resolves the modality bias, offering stronger generalization and more stable performance. Consequently, we follow the second approach to carry out the MBT problem." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.383, + 0.325, + 0.399 + ], + "angle": 0, + "content": "2.3. Self-Supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.405, + 0.484, + 0.662 + ], + "angle": 0, + "content": "Self-supervised learning, as a subset of unsupervised learning, constructs pretext tasks to extract valuable insights from large-scale unlabeled data, thereby learning representations beneficial for downstream tasks. Common methods involve contextual pretext tasks to learn discriminative features, such as Word2Vec [25], Jigsaw [26], Rotation [8], Colorization [38], and MAE [14]. Another prominent approach, contrastive learning, builds discriminative representations by comparing instances within the same feature space, as seen in methods like MoCo [12] and BYOL [9], both of which perform well on downstream tasks. Nevertheless, above methods are not specifically tailored for the Re-ID task, particularly for VT-ReID task. In this paper, we leverage the consistency of different modalities to construct a patch-mixed image reconstruction self-supervised task aimed at extracting shared features, thereby mitigating the modality discrepancy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.676, + 0.353, + 0.693 + ], + "angle": 0, + "content": "3. Benchmark: UCM-VeID V2" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.702, + 0.484, + 0.779 + ], + "angle": 0, + "content": "To further enhance the UAV cross-modal vehicle re-identification dataset, we have significantly expanded the first edition, UCM-VeID [22], by an order of magnitude in terms of UAV platforms, identities and so on. This section details the expanded dataset called UCM-VeID V2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.279, + 0.804 + ], + "angle": 0, + "content": "3.1. Dataset Description" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "(1) Dataset Partition. UCM-VeID V2 comprises a total of 3407 identities with 165091 samples. The dataset is split into a training set and a test set in an \\(8:2\\) ratio, resulting in a training set with 2,693 identities and 131,451 samples, while the test set contains 674 identities and 33,640 samples. Furthermore, the query set and gallery set are derived" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.221, + 0.907, + 0.312 + ], + "angle": 0, + "content": "from the test set. Specifically, one sample is drawn from each ID within the test set to constitute the query set, while the remaining samples collectively form the gallery set. It is worth noting that the extraction process is random, and crucially, every sample in the query set has a corresponding sample with the same ID in the gallery set." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.312, + 0.907, + 0.372 + ], + "angle": 0, + "content": "(2) Dataset Statistics. To maximize the utility of the UCM-VeID V2 dataset, we perform a detailed statistical analysis of its key attributes, providing insights into the data patterns relevant to Re-ID tasks." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.372, + 0.906, + 0.463 + ], + "angle": 0, + "content": "- ID distribution of 5 cameras. Data is collected using 3 drone platforms equipped with 5 different cameras. \\(D300_{z}\\) camera captures fewer IDs compared to the other four cameras, as Fig. 3(a) shown. The \\(D300_{z}\\) camera captures high-resolution images, primarily focused on nighttime data, thereby enhancing the dataset's diversity." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.463, + 0.906, + 0.551 + ], + "angle": 0, + "content": "- Distribution of ID samples. Samples are collected from different camera perspectives, ensuring diversity and ample representation for each ID. As shown in Fig. 3(b), most IDs have 20-40 samples, providing comprehensive coverage of target's appearance across different conditions." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.554, + 0.906, + 0.644 + ], + "angle": 0, + "content": "- Distribution of Resolution. As shown in Fig. 3(c), resolutions vary significantly across cameras. For example, \\(D300_{i}r\\) samples cluster around 10,000-15,000 pixels, while DM3P samples are predominantly 45,000-50,000 pixels, highlighting a resolution disparity that presents a core challenge in UAV-based VT-ReID." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.644, + 0.906, + 0.749 + ], + "angle": 0, + "content": "- Energy Distribution of IR Samples. The dataset spans all seasons and covers a 24-hour period, capturing both dynamic and static vehicles, leading to diverse infrared energy levels. Fig. 3(d) reveals that most infrared energy values are between 80-100, corresponding to moderate daylight. Extreme conditions, like intense heat or cold, are less common but contribute to the dataset's breadth." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.75, + 0.905, + 0.794 + ], + "angle": 0, + "content": "(3) Dataset Attributes. Attributes annotation is an integral part of building reliable, high-performance data sets. In this dataset, we annotate the below attributes." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.372, + 0.906, + 0.794 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.796, + 0.906, + 0.886 + ], + "angle": 0, + "content": "- Vehicle Type. Vehicles are classified into categories such as car, SUV, minibus, MPV, bus, mini truck, truck, and pickup. As shown in Fig. 3(e), cars and SUVs make up the majority, while trucks and pickups are less frequent. This distribution reflects the urban focus of our data collection, where cars and SUVs are more prevalent." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "- Vehicle color. We categorize vehicle colors into thirteen" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22288" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.092, + 0.284, + 0.102 + ], + "angle": 0, + "content": "ID Distribution of Different Cameras" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.111, + 0.294, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.192, + 0.225, + 0.21, + 0.237 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.342, + 0.092, + 0.461, + 0.102 + ], + "angle": 0, + "content": "Distribution of ID Samples" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.108, + 0.493, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.394, + 0.225, + 0.412, + 0.237 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.09, + 0.657, + 0.099 + ], + "angle": 0, + "content": "Distribution of Resolution" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.102, + 0.7, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.225, + 0.618, + 0.237 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.722, + 0.09, + 0.874, + 0.101 + ], + "angle": 0, + "content": "Energy Distribution of IR Samples" + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.105, + 0.897, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.796, + 0.225, + 0.815, + 0.237 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.242, + 0.262, + 0.254 + ], + "angle": 0, + "content": "Distribution of Vehicle Type" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.256, + 0.291, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.192, + 0.377, + 0.211, + 0.389 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image_caption", + "bbox": [ + 0.335, + 0.242, + 0.464, + 0.253 + ], + "angle": 0, + "content": "Distribution of Vehicle Color" + }, + { + "type": "image", + "bbox": [ + 0.294, + 0.256, + 0.49, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.393, + 0.376, + 0.411, + 0.389 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.243, + 0.678, + 0.252 + ], + "angle": 0, + "content": "Distribution of Orientation Class" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.256, + 0.7, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.376, + 0.619, + 0.389 + ], + "angle": 0, + "content": "(g)" + }, + { + "type": "image_caption", + "bbox": [ + 0.723, + 0.24, + 0.877, + 0.25 + ], + "angle": 0, + "content": "Orientation Classification Criteria" + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.252, + 0.88, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.796, + 0.376, + 0.816, + 0.389 + ], + "angle": 0, + "content": "(h)" + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.399, + 0.892, + 0.414 + ], + "angle": 0, + "content": "Figure 3. The statistics and annotation attributes of UCM-VeID V2 dataset. Notation that, the 'k' in (c) means the unit of 1000 pixels." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.44, + 0.483, + 0.514 + ], + "angle": 0, + "content": "classes, shown in Fig. 3(f), ordered by proportion as follows: white, black, gray, red, dark gray, cyan, brown, blue, golden, yellow, others, green, and orange. Despite some imbalance, this distribution mirrors real-world vehicle color trends." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.516, + 0.483, + 0.605 + ], + "angle": 0, + "content": "- Vehicle Orientation. Based on [22] and Fig. 3(h), we define eight orientation categories. Fig. 3(g) shows that orientations 2, 4, 6, and 8 are most common, a result of the UAV's flexible viewing angles. This orientation imbalance highlights the dataset's realism and the challenges in UAV-based Re-ID tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.62, + 0.38, + 0.636 + ], + "angle": 0, + "content": "3.2. Comparison with Other Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.673 + ], + "angle": 0, + "content": "In contrast to other vehicle Re-ID datasets, as detailed in Tab. 1, our dataset has several unique characteristics:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.675, + 0.483, + 0.763 + ], + "angle": 0, + "content": "1. Multi modal information. Compared with RGB-based datasets, such as VRAI [31], UAV-VeID [29], UCMVeID [22] and UCM-VeID V2 integrate both RGB and IR modalities, facilitating cross-modality vehicle Re-ID tasks, which holds promise as a viable solution to achieve round-the-clock vehicle Re-Identification." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.765, + 0.483, + 0.869 + ], + "angle": 0, + "content": "2. Flexible viewpoint, scale and orientations. The UCMVeID V2 dataset, captured via a drone-mounted camera, offers varied perspectives, scales, and orientations that differ from fixed-camera datasets like RGBN300 [18] and MSVR310 [40]. This diversity substantially enriches the dataset, making it a valuable resource for robust vehicle Re-ID research." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.9 + ], + "angle": 0, + "content": "3. Overall enhancement. Compared to UCM-VeID, the UCM-VeID V2 dataset offers a substantial increase in" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.675, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.44, + 0.907, + 0.547 + ], + "angle": 0, + "content": "data volume and significant improvements in several dimensions, including additional cameras and platforms, extended capture times, a broader variety of weather conditions, and a wider range of shooting scenarios, as shown in Fig. 1. These enhancements make it a richer and more comprehensive dataset, supporting diverse and robust data-driven research and applications." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.562, + 0.791, + 0.58 + ], + "angle": 0, + "content": "4. Proposed pre-training method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.587, + 0.744, + 0.602 + ], + "angle": 0, + "content": "4.1. Overall Model Structure." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.907, + 0.791 + ], + "angle": 0, + "content": "The proposed pre-training VT-ReID method is illustrated in Fig. 4. Firstly, style learning and image generation techniques are used to create pairs of RGB and IR images. Then, we segment these images into patches and randomly swap patches between different modalities to create two new patch-mixed images. Finally, the patch-mixed images are fed into the encoder to complete Patch-Mixed Image Reconstruction (PMIR), Modality Discrimination Adversarial Learning (MDAL) and Modality-Augmented Contrasting Learning (MACL) tasks, where PMIR and MDAL are adopted to learn modality-invariant features, and MACL is utilized to enhance ID discriminability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.803, + 0.819 + ], + "angle": 0, + "content": "4.2. Patch-Mixed Unified Embedding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Cross-modality data generation. Given the limited size of Re-ID dataset, we first train a CycleGAN [41] to generate cross-modality versions of ImageNet. Specifically, we use CycleGAN to learn style mapping between RGB and IR images from the UCM-VeID V2 dataset with a gener" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22289" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.095, + 0.905, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.3, + 0.908, + 0.342 + ], + "angle": 0, + "content": "Figure 4. Pre-training paradigm, which comprises preprocessing and three self-supervised tasks. Preprocessing involves obtaining paired RGB and IR images through style learning and image generation, followed by patches segmentation and swapping, PMIR and MDAL then focus on learning modality-invariant features, while MACL aims to enhance the ID discriminability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.483, + 0.444 + ], + "angle": 0, + "content": "ator \\(G\\). To avoid data leakage, we use ImageNet dataset for pre-training. For an input RGB image \\(\\pmb{x}_i^{\\mathrm{RGB}}\\) in the ImageNet, we generate the corresponding IR image as \\(\\pmb{x}_i^{\\mathrm{IR}} = G(\\pmb{x}_i^{\\mathrm{RGB}})\\). The resulting cross-modality ImageNet dataset is then \\(\\{(\\pmb{x}_i^{\\mathrm{RGB}},\\pmb{x}_i^{\\mathrm{IR}})\\}_{i = 1}^N\\), where \\(N\\) is the total number." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.444, + 0.484, + 0.671 + ], + "angle": 0, + "content": "Random patch mixture. Our method aims to teach the backbone to fully leverage the cross-modality information for modality bias elimination. Therefore, for the data input aspect, we mixture the RGB and IR image pair in a patch swapping manner. Mathematically, given a RGB-IR image pair \\((\\pmb{x}_i^{\\mathrm{RGB}},\\pmb{x}_i^{\\mathrm{IR}})\\), we first establish a patch mask \\(M\\) with the same size as \\(\\pmb{x}_i^{\\mathrm{RGB}}(\\pmb{x}_i^{\\mathrm{IR}})\\) that consists of \\(K\\times K\\) regular non-overlapping patches. With a predefined mask ratio \\(\\eta\\), \\(\\lfloor \\eta \\times K\\times K\\rfloor\\) patches in \\(M\\) are randomly selected following a uniform distribution and set as all-zeros matrices. In this paper, we predefined a mask ratio \\(\\eta = 0.75\\). A higher \\(\\eta\\) leads to more patch exchanges, increases the difficulty of the reconstruction task, and enables the model to learn more shared features. Then, the process of the proposed random path mixture can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.676, + 0.482, + 0.716 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {\\boldsymbol {x}} _ {i} ^ {\\mathrm {R G B}} = \\boldsymbol {x} _ {i} ^ {\\mathrm {R G B}} \\odot \\boldsymbol {M} + \\boldsymbol {x} _ {i} ^ {\\mathrm {I R}} \\odot (1 - \\boldsymbol {M}) \\\\ \\hat {\\boldsymbol {x}} _ {i} ^ {\\mathrm {I R}} = \\boldsymbol {x} _ {i} ^ {\\mathrm {I R}} \\odot \\boldsymbol {M} + \\boldsymbol {x} _ {i} ^ {\\mathrm {R G B}} \\odot (1 - \\boldsymbol {M}), \\end{array} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.751 + ], + "angle": 0, + "content": "Then, based on the Eq. (1), the cross-modality patch-mixed unified embeddings are finally achieved as:" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.756, + 0.483, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {f} _ {i} ^ {\\mathrm {R G B}}, \\boldsymbol {f} _ {i} ^ {\\mathrm {I R}} = \\boldsymbol {F} \\left(\\hat {\\boldsymbol {x}} _ {i} ^ {\\mathrm {R G B}}\\right), \\boldsymbol {F} \\left(\\hat {\\boldsymbol {x}} _ {i} ^ {\\mathrm {I R}}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.827 + ], + "angle": 0, + "content": "where \\( F(\\cdot) \\) is the feature extractor and ResNet50 [11] is employed in this paper. \\( f_{i}^{\\mathrm{RGB}} \\) and \\( f_{i}^{\\mathrm{IR}} \\) are the latent embeddings of mixed images." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.834, + 0.409, + 0.851 + ], + "angle": 0, + "content": "4.3. Self-supervised Multi-task Learning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "With the proposed unified embedding module, we can acquire the latent features containing cross-modality information. However, these features can not be directly employed" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.368, + 0.905, + 0.413 + ], + "angle": 0, + "content": "to the Re-ID task due to modality-bias information and insufficient discriminability [30, 34, 37]. To address above these issues, we devise three self-supervised tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.414, + 0.907, + 0.671 + ], + "angle": 0, + "content": "Patch-Mixed Image Reconstruction. In this task, we utilize the mixed features, \\( f_{i}^{\\mathrm{RGB}} \\) and \\( f_{i}^{\\mathrm{IR}} \\), to reconstruct the input images, \\( x_{i}^{\\mathrm{RGB}} \\) and \\( x_{i}^{\\mathrm{IR}} \\), respectively. This approach is motivated by the inherent advantage of image reconstruction as a naturally unsupervised learning task, eliminating the need for additional labels and making it suitable for training large models. Moreover, for RGB image reconstruction, \\( f_{i}^{\\mathrm{RGB}} \\) contains minimal RGB information but abundant IR information, compelling the network to focus on modality-shared features. The limited RGB information acts as style information that supports the reconstruction of the RGB image. Similarly, for IR image reconstruction, \\( x_{i}^{\\mathrm{IR}} \\) using \\( f_{i}^{\\mathrm{IR}} \\), where the feature contains mostly RGB information with only a small amount of IR data, thus also ensuring the model learns to emphasize cross-modality features effectively. According to above analysis, the process of patch-mixed image reconstruction can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.681, + 0.905, + 0.719 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{r l} \\bar {\\boldsymbol {x}} _ {i} ^ {\\mathrm {R G B}} & = \\boldsymbol {H} _ {\\mathrm {R G B}} \\left(\\boldsymbol {f} _ {i} ^ {\\mathrm {R G B}}\\right) \\\\ \\bar {\\boldsymbol {x}} _ {i} ^ {\\mathrm {I R}} & = \\boldsymbol {H} _ {\\mathrm {I R}} \\left(\\boldsymbol {f} _ {i} ^ {\\mathrm {I R}}\\right), \\end{array} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.731, + 0.907, + 0.809 + ], + "angle": 0, + "content": "where \\(\\bar{\\pmb{x}}_i^{\\mathrm{RGB}}\\) and \\(\\bar{\\pmb{x}}_i^{\\mathrm{IR}}\\) are the reconstructed RGB and IR images. \\(H_{\\mathrm{RGB}}(\\cdot)\\) and \\(H_{\\mathrm{IR}}(\\cdot)\\) are networks for RGB and IR image reconstruction respectively. Finally, a pixel-wise image reconstruction loss with \\(l_{2}\\)-norm is employed to guide the reconstruction process as:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.819, + 0.905, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r e c}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\| \\bar {\\boldsymbol {x}} _ {i} ^ {\\mathrm {R G B}} - \\boldsymbol {x} _ {i} ^ {\\mathrm {R G B}} \\| _ {2} ^ {2} + \\| \\bar {\\boldsymbol {x}} _ {i} ^ {\\mathrm {I R}} - \\boldsymbol {x} _ {i} ^ {\\mathrm {I R}} \\| _ {2} ^ {2}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Modality Discrimination Adversarial Learning. In this task, we introduce adversarial training to encourage" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22290" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "the learned features to be modality-invariant. Specifically, we employ a discriminator \\( D(\\cdot) \\) to distinguish between the mixed features \\( f_{i}^{\\mathrm{RGB}} \\) and \\( f_{i}^{\\mathrm{IR}} \\). Initially, we fix the feature extractor \\( F \\) and train the discriminator \\( D \\) to correctly classify the mixed features based on their corresponding modality labels. This step ensures that \\( D \\) can accurately distinguish between features from different modalities. The objective for the discriminator \\( D \\) can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.224, + 0.483, + 0.243 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D}} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left[ y _ {i} \\log \\boldsymbol {D} (\\boldsymbol {f} _ {i}) + (1 - y _ {i}) \\log (1 - \\boldsymbol {D} (\\boldsymbol {f} _ {i})) \\right], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.253, + 0.484, + 0.389 + ], + "angle": 0, + "content": "where \\( y_{i} \\) is the modality label, \\( f_{i} \\) represents either \\( f_{i}^{\\mathrm{RGB}} \\) or \\( f_{i}^{\\mathrm{IR}} \\), and \\( K \\) is the batch size. In the second step, we fix the discriminator \\( D \\) and update the feature extractor \\( F \\) to generate mixed features that the discriminator cannot distinguish. This step forces the feature extractor to produce modality-invariant features, effectively minimizing the modality bias in the learned features. The objective for the feature extractor \\( F \\) during this phase is to maximize the classification error of \\( D \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.401, + 0.483, + 0.42 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {F}} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left[ \\left(1 - y _ {i}\\right) \\log \\boldsymbol {D} (\\boldsymbol {f} _ {i}) + y _ {i} \\log \\left(1 - \\boldsymbol {D} (\\boldsymbol {f} _ {i})\\right) \\right]. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.431, + 0.483, + 0.492 + ], + "angle": 0, + "content": "The adversarial training process iterates between these two steps until the model converges, achieving a balance where the discriminator \\( D \\) cannot reliably distinguish the modality of the features generated by \\( F \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.657 + ], + "angle": 0, + "content": "Modality-Augmented Contrasting Cluster. In this task, we focus on enhancing the discriminative power of the learned features by employing a contrastive learning approach. Unlike traditional methods that rely on data augmentation to generate negative samples, our approach does not introduce additional data augmentation but instead utilizes the inherent modality-specific features for contrastive learning. We treat \\( f_{i}^{\\mathrm{RGB}} \\) and \\( f_{i}^{\\mathrm{IR}} \\) as positive pairs within the contrastive learning framework. The contrastive loss function aims to maximize the similarity between the mixed features of different modalities, which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.669, + 0.483, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n t r a s t}} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {f} _ {i} ^ {\\mathrm {R G B}} , \\boldsymbol {f} _ {i} ^ {\\mathrm {I R}}\\right)\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {f} _ {i} ^ {\\mathrm {R G B}}, \\boldsymbol {f} _ {j} ^ {\\mathrm {I R}}\\right)\\right)}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.722, + 0.483, + 0.842 + ], + "angle": 0, + "content": "where \\(\\mathrm{sim}(\\cdot ,\\cdot)\\) denotes a similarity measure (such as cosine similarity) between features. The loss function encourages the feature extractor to learn representations that pull together positive pairs \\((f_{i}^{\\mathrm{RGB}}\\) and \\(f_{i}^{\\mathrm{IR}})\\) while pushing apart other features. By iteratively optimizing this contrastive loss, the feature extractor \\(\\pmb{F}\\) is guided to produce more discriminative features, thereby improving the overall performance of the cross-modal image re-identification task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.843, + 0.483, + 0.873 + ], + "angle": 0, + "content": "In summary, the overall loss function for the entire training process can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l l}} = \\mathcal {L} _ {\\text {r e c}} + \\left(\\mathcal {L} _ {\\mathrm {D}} + \\mathcal {L} _ {\\mathrm {F}}\\right) + \\mathcal {L} _ {\\text {c o n t r a s t}}. \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.646, + 0.107 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.116, + 0.871, + 0.132 + ], + "angle": 0, + "content": "5.1. Implementation and Evaluation Protocols" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.905, + 0.273 + ], + "angle": 0, + "content": "Pre-training Stage. For pre-training, we implement CPSL using PyTorch on an NVIDIA RTX A100. Optimization is conducted with an SGD optimizer (momentum of 0.9, weight decay of 0.0005), an initial learning rate of 0.1 with linear warm-up, and a cosine decay schedule without restarts. Data augmentation includes random grayscale, horizontal flipping, Gaussian filtering, and interpolation sampling. Images are resized to \\(256 \\times 256\\), and the mask size \\(N\\) is set to 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.275, + 0.906, + 0.426 + ], + "angle": 0, + "content": "Fine-Tuning Stage. For fine-tuning, we implement VT-ReID models using PyTorch on an NVIDIA GeForce RTX 3090 GPU. Images are resized to \\(180 \\times 256\\). Optimization uses an SGD optimizer (momentum of 0.9, weight decay of 0.0005) with an initial learning rate of 0.1 and linear warm-up over 80 epochs. The model is initialized with CPSL checkpoints while other hyper-parameters and data augmentations remain unchanged, as this setup is not specifically tuned for CPSL. Further optimization may improve results, which we leave for future work." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.906, + 0.518 + ], + "angle": 0, + "content": "Evaluation Protocols. Following [34], we use the cumulative matching characteristic (CMC) curve and mean average precision (mAP) for evaluation. CMC assesses rank-k accuracy, with rank-1, rank-10, and rank-20 scores reported in our experiments. mAP reflects the overall retrieval performance across all query samples." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.529, + 0.838, + 0.545 + ], + "angle": 0, + "content": "5.2. Comparison with pre-trained weights" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.906, + 0.687 + ], + "angle": 0, + "content": "To validate the effectiveness of our proposed pre-training method, we employ three different backbone initialization strategies during fine-tuning: random initialization (Random Init.), initialization with ImageNet pre-trained weights (ImageNet-1K), and initialization using the proposed pretraining weights (CPSL). To enhance experimental rigor, we perform supervised fine-tuning with four state-of-the-art VT-ReID methods: AGW [37], DDAG [36], MMD-ReID [17], and DEEN [39]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.906, + 0.779 + ], + "angle": 0, + "content": "From Tab. 2, we can observe that the VT-ReID models trained from scratch generally under-performance, with all methods failing to converge. The primary cause of the gradient explosion is attributed to the triplet loss, suggesting that randomly initialized models struggle to establish semantically meaningful Euclidean feature spaces." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.884 + ], + "angle": 0, + "content": "Our method CPSL achieves superior results across all testing and query modes. In contrast to ImageNet-1K pretraining, we do not rely on additional training data or incorporate any supervised training processes. The superior performance is primarily attributed to the design of our self-supervised tasks, which account for modality bias and eliminate cross-modality differences." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "It can also be observed that CPSL consistently improves" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22291" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.925, + 0.336 + ], + "angle": 0, + "content": "
MethodVenuePre-TrainSingle-shotMulti-shot
IR2RGBRGB2IRIR2RGBRGB2IR
rank1rank10mAPrank1rank10mAPrank1rank10mAPrank1rank10mAP
AGW [37]TPAMI 2021Random Init.4.3518.759.384.8919.79.297.2227.264.826.6726.175.53
ImageNet-1K15.7545.9825.7416.1747.5826.5420.158.1216.2120.4558.1618.3
CPSL23.4753.0633.6925.3658.5436.5731.9167.5424.1332.6468.0724.01
DDAG [36]ECCV 2020Random Init.5.2222.8111.485.0622.6911.228.5532.725.847.0630.056.25
ImageNet-1K20.4648.7830.1219.4649.1529.2127.5864.7520.2826.6965.5422.87
CPSL22.2350.8631.9923.4551.3332.9431.9667.9421.731.1666.5824.51
MMD-ReID [17]BMVC 2021Random Init.4.3415.279.864.9218.539.935.9919.3710.785.6920.3410.42
ImageNet-1K18.9348.6428.9919.2749.1429.0822.0455.8719.6422.9659.2731.41
CPSL21.9851.8532.3922.5253.3831.9928.3462.0922.129.0462.8521.8
DEEN [39]CVPR 2023Random Init.7.2931.7415.246.1127.0613.297.8837.447.487.3932.817.34
ImageNet-1K18.2653.0130.3220.4656.2332.0923.8364.1422.0622.6566.6121.43
CPSL22.5759.8531.0923.759.5633.3525.6267.5223.4424.5368.6323.97
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.103, + 0.338, + 0.413, + 0.351 + ], + "angle": 0, + "content": "* The bold entries represent the best performance in table." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.361, + 0.908, + 0.39 + ], + "angle": 0, + "content": "Table 2. Comparisons between the proposed CPSL and some state-of-the-art methods on UCM-VeID V2 datasets. The rank-k (%) and mAP (%) are demonstrated." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.416, + 0.484, + 0.507 + ], + "angle": 0, + "content": "performance across various VT-ReID methods, which indicates that the weights generated by CPSL exhibit strong robustness and adaptability. This suggests that the CPSL method is effective in enhancing the generalization capabilities of the models, allowing them to perform well under different conditions and with diverse Re-ID techniques." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.516, + 0.431, + 0.533 + ], + "angle": 0, + "content": "5.3. Comparisons With Other SSL Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.484, + 0.659 + ], + "angle": 0, + "content": "We compare CPSL with multiple ResNet50 pre-trained models on the UCM-VeID V2. The methods involved can be categorized into four groups: (1) method based on VT-ReID fine-tuning (MMGL [30]), (2) method based pretext task design (Jigsaw [26], Rotation [8]), (3) method based self-supervised learning (SSL) pre-training (InstDisc [35], SimCLR [3], SwAV [2], BYOL [9], SimSiam [4], MoCoV2 [5], DenseCL [33], Relative-Loc[7])." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.765 + ], + "angle": 0, + "content": "As shown in Tab. 3, MMGL performs suboptimally. This is partly because MMGL is tailored for pedestrian targets, using a top-down human segmentation approach that struggles with vehicles' variable orientations. Additionally, the released MMGL weights were fine-tuned on pedestrian VT-ReID datasets, embedding pedestrian-specific features that do not generalize well to vehicle targets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "SSL methods share a similar goal with our approach by aiming to endow the encoder with discriminative capabilities. However, methods such as SwAV [2] and BYOL [9] perform poorly, as indicated by the stagnant triplet loss during fine-tuning, suggesting these methods fail to establish a meaningful Euclidean space from the outset. Although some SSL methods outperform those based on pretext task design, they still underperform compared to CPSL. This is because they focus solely on target discriminability through" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.413, + 0.903, + 0.66 + ], + "angle": 0, + "content": "
MethodRank1Rank10Rank20mAP
method based on VT-ReID fine-tuning:
MMGL [30]21.8256.5668.0316.14
method based pretext task design:
Jigsaw [26]21.2456.7767.2515.66
RotationNet [8]26.6263.0372.7719.32
method based on SSL pre-training:
InstDisc [35]24.0962.0773.1620.04
SimCLR [3]12.4348.5762.1511.17
SwAV [2]6.5728.8940.225.17
BYOL [9]6.8925.5136.875.17
SimSiam [4]27.2465.7473.1322.36
MoCo V2 [5]28.8967.4674.9822.44
DenseCL [33]27.8964.9975.1222.41
Relative-Loc[7]18.9155.7968.0615.53
ours32.6468.0776.3224.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.67, + 0.907, + 0.699 + ], + "angle": 0, + "content": "Table 3. Comparison with other SSL methods. The rank-k (%) and mAP (%) are demonstrated." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.728, + 0.907, + 0.79 + ], + "angle": 0, + "content": "contrastive learning without addressing modality differences. In contrast, our PMIR and MDAL modules effectively mitigate cross-modality differences, enabling shared feature learning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.667, + 0.818 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To assess the effectiveness of each component in our method, we conducted ablation studies on the proposed pretraining method, removing specific modules while keeping all other experimental settings constant. Our method includes three key modules: PMIR (R), MDAL (D), and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22292" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.485, + 0.219 + ], + "angle": 0, + "content": "
MethodSingle-shotMulti-shot
rank1rank10mAPrank1rank10mAP
Random Init.4.8919.79.296.6726.175.53
ImageNet-1K16.1747.5826.5420.4558.1618.3
R7.6426.9312.219.3342.7210.59
R+D14.6349.5722.4618.8854.9518.32
R+C22.4955.4830.3428.9765.1621.96
R+D+C25.3658.5436.5732.6468.0724.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.229, + 0.484, + 0.259 + ], + "angle": 0, + "content": "Table 4. The quantitative results of ablation studies. The rank-k \\((\\%)\\) and mAP \\((\\%)\\) are demonstrated." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.284, + 0.483, + 0.33 + ], + "angle": 0, + "content": "MACL (C). Here, R and D target modality alignment, while C enhances discriminative representation. The ablation study results are shown in Tab. 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.331, + 0.484, + 0.421 + ], + "angle": 0, + "content": "Effectiveness of PMIR. PMIR aims to create a shared latent representation space for RGB and IR modalities via reconstruction tasks. As shown in Tab. 4, compare to the results of Random Init., RANK1 and mAP improve by \\(2.57\\%\\) and \\(2.92\\%\\) under the single-shot query mode, indicating PMIR's effectiveness in reducing modality discrepancies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.422, + 0.484, + 0.496 + ], + "angle": 0, + "content": "Effectiveness of MDAL. MDAL uses modality adversarial learning to further align modalities. Tab. 4 shows a RANK1 and mAP increase of \\(6.99\\%\\) and \\(10.25\\%\\) under single-shot query mode, demonstrating that MDAL helps guide the encoder toward shared feature learning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.497, + 0.484, + 0.587 + ], + "angle": 0, + "content": "Effectiveness of MACL. MACL employs modality contrastive learning to enhance feature clustering and discriminative power. As evidenced in Tab. 4, RANK1 and mAP improve by \\(7.86\\%\\) and \\(7.88\\%\\) in single-shot query mode, confirming MACL's success in learning discriminative features for VT-ReID." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.598, + 0.485, + 0.66 + ], + "angle": 0, + "content": "
Pn2 × 23 × 34 × 45 × 5
rank124.6627.3232.6429.05
mAP16.3820.9124.0122.44
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.669, + 0.483, + 0.698 + ], + "angle": 0, + "content": "Table 5. Different partition stripe numbers \\( {P}_{n}\\left( {N \\times N}\\right) \\) . The rank-k (%) and mAP (%) are demonstrated." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.281, + 0.744 + ], + "angle": 0, + "content": "5.5. Parameter Analysis" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The partition stripe number \\( P_{n} \\) directly determines the size of the image blocks to be swapped and is an important hyperparameter of our method. To verify the impact of this hyperparameter, we conducted experiments with different values for the partition stripe number, with the relevant results presented in Tab. 5. From the table, it can be observed that as the partition stripe number increases, the model's performance initially improves before declining. This is because, as the partition stripe number increases, the semantic content within each image block transitions from being more" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.695, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.196, + 0.663, + 0.206 + ], + "angle": 0, + "content": "(a) ImageNet Pre-Training" + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.089, + 0.868, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.734, + 0.196, + 0.831, + 0.206 + ], + "angle": 0, + "content": "(b) PMSL Pre-Training" + }, + { + "type": "image", + "bbox": [ + 0.875, + 0.09, + 0.903, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.22, + 0.905, + 0.276 + ], + "angle": 0, + "content": "Figure 5. Visualization of the pre-learned features. Different colors represent distinct identities, while circle and cross symbols denote RGB and IR modalities, respectively. The experiment is conducted on UCM-VeID V2 dataset with the AGW baseline [37]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.303, + 0.907, + 0.35 + ], + "angle": 0, + "content": "discrete to becoming sparser. This indicates that in practical applications, appropriate adjustment of this parameter is necessary to achieve optimal performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.359, + 0.854, + 0.374 + ], + "angle": 0, + "content": "5.6. Visualization of the feature distribution" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.906, + 0.472 + ], + "angle": 0, + "content": "To better understand the proposed pre-training paradigm, we visualize the t-SNE distributions of features pre-learned using ImageNet-1K and the CPSL checkpoints on the UCM-VeID V2 dataset, utilizing AGW [37]. As shown in Fig. 5, we randomly select 15 identities to visualize all corresponding images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.472, + 0.906, + 0.653 + ], + "angle": 0, + "content": "Fig. 5 demonstrates that, the modal gap between PMSL features is smaller than that between ImageNet features (highlighted by the red circle). Specifically, PMSL features show less variation across modalities, indicating better consistency and robustness when processing RGB and IR images. In contrast, ImageNet-pretrained features exhibit significant modal variation, which can negatively impact performance in VT-ReID tasks. Thus, PMSL not only eliminates cross-modality discrepancy but also improves the model's discriminative ability for downstream tasks, providing stronger support for cross-modal vehicle re-identification." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.668, + 0.634, + 0.684 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.692, + 0.906, + 0.872 + ], + "angle": 0, + "content": "In this paper, we introduce a novel dataset that expands existing ones in terms of scale, acquisition platforms, collection scenarios, time, and climate conditions. To address the MBT problem, we break it into two sub-problems: modality discrepancy and insufficient discriminability. We propose an end-to-end cross-modal pre-training framework to address these issues. To reduce modality discrepancies, the framework incorporates self-supervised tasks based on image block swapping and modality adversarial learning. For enhanced discriminability, we design an unsupervised clustering-guided multi-modal contrastive learning approach. Extensive experiments on the new dataset validate the effectiveness of our method." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.873, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by the National Natural Science Foundation of China (62201586)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "22293" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.121, + 0.188, + 0.137 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.147, + 0.483, + 0.202 + ], + "angle": 0, + "content": "[1] Yan Bai, Yihang Lou, Feng Gao, Shiqi Wang, Yuwei Wu, and Ling-Yu Duan. Group-sensitive triplet embedding for vehicle reidentification. IEEE Trans. Multim., 20(9):2385-2399, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.204, + 0.483, + 0.272 + ], + "angle": 0, + "content": "[2] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In Proc. Adv. Neural Inf. Process. Syst., pages 9912-9924, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.275, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[3] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In Proc. Int. Conf. Mach. Learn., pages 1597-1607, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.333, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[4] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 15750-15758, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.376, + 0.482, + 0.416 + ], + "angle": 0, + "content": "[5] Xinlei Chen, Haoqi Fan, Ross B. Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. CoRR, abs/2003.04297, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.419, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proc. IEEE Conf. Comput. Vis. Pattern Recognition., pages 248-255, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.476, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[7] Carl Doersch, Abhinav Gupta, and Alexei A. Efros. Unsupervised visual representation learning by context prediction. In Proc. IEEE Int. Conf. Comput. Vis., pages 1422-1430. IEEE Computer Society, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.533, + 0.482, + 0.574 + ], + "angle": 0, + "content": "[8] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In Proc. Int. Conf. Learn. Represent., 2018. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.577, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[9] Jean-Bastien Grill, Florian Strub, Florent Alché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Ávila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent - A new approach to self-supervised learning. In Proc. Adv. Neural Inf. Process. Syst., pages 21271-21284, 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.675, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[10] Haiyun Guo, Chaoyang Zhao, Zhiwei Liu, Jinqiao Wang, and Hanqing Lu. Learning coarse-to-fine structured feature embedding for vehicle re-identification. In Proc. AAAI Conf. Artif. Intell., pages 6853–6860, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.732, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[12] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 9726-9735, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 15979-15988, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.147, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[14] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 15979-15988, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.216 + ], + "angle": 0, + "content": "[15] Yu Huang, Junyang Lin, Chang Zhou, Hongxia Yang, and Longbo Huang. Modality competition: What makes joint training of multi-modal network fail in deep learning? (provably). In Proc. Int. Conf. Mach. Learn., pages 9226-9259, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[16] Yan Huang, Qiang Wu, Jingsong Xu, Yi Zhong, Peng Zhang, and Zhaoxiang Zhang. Alleviating modality bias training for infrared-visible person re-identification. IEEE Trans. Multim., 24:1570-1582, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.904, + 0.327 + ], + "angle": 0, + "content": "[17] Chaitra Jambigi, Ruchit Rawal, and Anirban Chakraborty. MMD-ReID: A simple but effective solution for visible-thermal person reid. In Proc. BMVC, pages 12026–12035, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[18] H. Li, C. Li, X. Zhu, A. Zheng, and B. Luo. Multi-spectral vehicle re-identification: A challenge. In Proc. AAAI Conf. Artif. Intell., pages 11345–11353, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.372, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[19] Chunsheng Liu, Ye Song, Faliang Chang, Shuang Li, Ruimin Ke, and Yinhai Wang. Posture calibration based cross-view and hard-sensitive metric learning for UAV-based vehicle re-identification. IEEE Trans. Inf. Forensics Secur., 23:19246-19257, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[20] Hongye Liu, Yonghong Tian, Yaowei Wang, Lu Pang, and Tiejun Huang. Deep relative distance learning: Tell the difference between similar vehicles. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 2167-2175, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[21] Xinchen Liu, Wu Liu, Tao Mei, and Huadong Ma. A deep learning-based approach to progressive vehicle re-identification for urban surveillance. In Proc. Eur. Conf. Comput. Vis., pages 869–884, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[22] Xingyue Liu, Jiahao Qi, Chen Chen, Kangcheng Bin, and Ping Zhong. Beyond sharing weights in decoupling feature learning network for UAV rgb-infrared vehicle re-identification. IEEE Trans. Multim., 2024. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[23] Yihang Lou, Yan Bai, Jun Liu, Shiqi Wang, and Lingyu Duan. VERI-Wild: A large dataset and a new method for vehicle re-identification in the wild. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3235-3243, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.904, + 0.704 + ], + "angle": 0, + "content": "[24] Mingming Lu, Yongchuan Xu, and Haifeng Li. Vehicle re-identification based on UAV viewpoint: Dataset and method. Remote. Sens., 14(18):4603, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[25] Tomás Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In Proc. Int. Conf. Learn. Represent., 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[26] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In Proc. Eur. Conf. Comput. Vis., pages 69-84, 2016. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[27] Xiaokang Peng, Yake Wei, Andong Deng, Dong Wang, and Di Hu. Balanced multimodal learning via on-the-fly gradient modulation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 8228-8237, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[28] Shangzhi Teng, Shiliang Zhang, Qingming Huang, and Nicu Sebe. Multi-view spatial attention embedding for vehicle re-identification. IEEE Trans. Circuits Syst. Video Technol., 31 (2):816-827, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22294" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[29] Shangzhi Teng, Shiliang Zhang, Qingming Huang, and Nicu Sebe. Viewpoint and scale consistency reinforcement for UAV vehicle re-identification. Int. J. Comput. Vis., 129(3): 719-735, 2021. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[30] Lin Wan, Qianyan Jing, Zongyuan Sun, Chuang Zhang, Zhihang Li, and Yehansen Chen. Self-supervised modality-aware multiple granularity pre-training for rgb-infrared person re-identification. IEEE Trans. Inf. Forensics Secur., 18: 3044–3057, 2023. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[31] Peng Wang, Bingliang Jiao, Lu Yang, Yifei Yang, Shizhou Zhang, Wei Wei, and Yanning Zhang. Vehicle re-identification in aerial imagery: Dataset and approach. In Proc. IEEE Int. Conf. Comput. Vis., pages 460–469, 2019. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.29, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[32] Weiyao Wang, Du Tran, and Matt Feiszli. What makes training multi-modal classification networks hard? In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 12692-12702, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.346, + 0.482, + 0.415 + ], + "angle": 0, + "content": "[33] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3024-3033. Computer Vision Foundation / IEEE, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[34] Ancong Wu, Wei-Shi Zheng, Hong-Xing Yu, Shaogang Gong, and Jianhuang Lai. RGB-infrared cross-modality person re-identification. In Proc. IEEE Int. Conf. Comput. Vis., pages 5390-5399, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.475, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[35] Zhirong Wu, Yuanjun Xiong, Stella X. Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3733-3742, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.531, + 0.482, + 0.586 + ], + "angle": 0, + "content": "[36] Mang Ye, Jianbing Shen, David J. Crandall, Ling Shao, and Jiebo Luo. Dynamic dual-attentive aggregation learning for visible-infrared person re-identification. Proc. Eur. Conf. Comput. Vision, 12362:229–247, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.588, + 0.482, + 0.642 + ], + "angle": 0, + "content": "[37] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven C. H. Hoi. Deep learning for person re-identification: A survey and outlook. IEEE Trans. Pattern Anal. Mach. Intell., 44:2872–2893, 2020. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.644, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[38] Richard Zhang, Phillip Isola, and Alexei A. Efros. Colorful image colorization. In Proc. Eur. Conf. Comput. Vis., pages 649-666, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.687, + 0.482, + 0.754 + ], + "angle": 0, + "content": "[39] Yukang Zhang and Hanzi Wang. Diverse embedding expansion network and low-light cross-modality benchmark for visible-infrared person re-identification. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 2153–2162, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.757, + 0.482, + 0.826 + ], + "angle": 0, + "content": "[40] Aihua Zheng, Xianpeng Zhu, Chenglong Li, Jin Tang, and Jixin Ma. Cross-directional consistency network with adaptive layer normalization for multi-spectral vehicle re-identification and a high-quality benchmark. Inf. Fusion, 100:101901, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.828, + 0.482, + 0.883 + ], + "angle": 0, + "content": "[41] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proc. IEEE Int. Conf. Comput. Vis., pages 2242-2251, 2017. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.883 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22295" + } + ] +] \ No newline at end of file diff --git a/2025/UCM-VeID V2_ A Richer Dataset and A Pre-training Method for UAV Cross-Modality Vehicle Re-Identification/9b7d21db-8ee1-4718-8ddf-98653daa1c9e_origin.pdf b/2025/UCM-VeID V2_ A Richer Dataset and A Pre-training Method for UAV Cross-Modality Vehicle Re-Identification/9b7d21db-8ee1-4718-8ddf-98653daa1c9e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..64f8f751f06a0875a955abcac43776b1a46cdefd --- /dev/null +++ b/2025/UCM-VeID V2_ A Richer Dataset and A Pre-training Method for UAV Cross-Modality Vehicle Re-Identification/9b7d21db-8ee1-4718-8ddf-98653daa1c9e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b681dc2936fc520e17b90b0ca4790ccc57b7a953f91e15896d4dac03e9c47c52 +size 4259868 diff --git a/2025/UCM-VeID V2_ A Richer Dataset and A Pre-training Method for UAV Cross-Modality Vehicle Re-Identification/full.md b/2025/UCM-VeID V2_ A Richer Dataset and A Pre-training Method for UAV Cross-Modality Vehicle Re-Identification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d5e39ee2437c217d8212a825a2def8735a3bcfa0 --- /dev/null +++ b/2025/UCM-VeID V2_ A Richer Dataset and A Pre-training Method for UAV Cross-Modality Vehicle Re-Identification/full.md @@ -0,0 +1,356 @@ +# UCM-VeID V2: A Richer Dataset and A Pre-training Method for UAV Cross-Modality Vehicle Re-Identification + +Xingyue Liu Jiahao Qi Chen Chen KangCheng Bin Ping Zhong* National University of Defense Technology, China + +{liuxingyue18, qijiahao1996, chenchen21c, binkc21, zhongping}@nudt.edu.cn + +# Abstract + +Cross-Modality Re-Identification (VT-ReID) aims to achieve around-the-clock target matching, benefiting from the strengths of both RGB and infrared (IR) modalities. However, the field is hindered by limited datasets, particularly for vehicle VT-ReID, and by challenges such as modality bias training (MBT), stemming from biased pre-training on ImageNet. To tackle the above issues, this paper introduces an dataset benchmark, named UCM-VeID V2, for vehicle VT-ReID, and proposes a new self-supervised pre-training method, Cross-Modality Patch-Mixed Self-supervised Learning (PMSL). UCM-VeID V2 dataset features a significant increase in data volume, along with enhancements in multiple aspects. PMSL addresses MBT by learning modality-invariant features through Patch-Mixed Image Reconstruction (PMIR) and Modality Discrimination Adversarial Learning (MDAL), and enhances discriminability with Modality-Augmented Contrasting Cluster (MACC). Comprehensive experiments are carried out to validate the proposed method. + +# 1. Introduction + +Cross-modality vehicle re-identification (ReID), also referred to as vehicle visible-infrared (VI) ReID [22], involves retrieving a matching vehicle image from a gallery set captured in the visible (RGB) or infrared (IR) spectrum, given a query vehicle image in the alternate modality. This process, where images are obtained from cameras of differing modalities, has garnered significant attention in applications such as video surveillance and public security [1, 28], primarily due to its advantages in enabling continuous, intelligent monitoring across day and night conditions. + +Despite its importance, research on vehicle VT-ReID remains relatively underdeveloped. A primary challenge lies in the lack of essential resources, including a comprehensive large-scale dataset and widely accessible benchmark + +![](images/650e42b7c9288e3d2e5a94cc8eb2f7d7be8c4f5764bcc15f9eddcb9f62c1411f.jpg) +(a) Multiple Platforms + +![](images/f24f985c52a5b5e52306a3bd1abb9b7ef9fe325624d2186bb5eb1743a4496851.jpg) +(b) Wider time spans + +![](images/cd42070e455eb5c1d8d9e73efb3768b4a8c733f11e6a8072d8ebe4c2c6ec027f.jpg) +(c) Richer weather conditions + +![](images/e881e6a447ae79ac1ed301ff496a789a0cfd9fef23e009eec007a435d70c4f42.jpg) +(d) Diverse Scenes + +![](images/79c7482d3305bee0d43233ed7746884924d63901f0a88dd194056ead19dde2ea.jpg) +(e) Samples +Figure 1. Increments and samples of UCM-VeID V2. (a)(b)(c)(d) are varies platforms, time periods, weather conditions and locations for data acquisition, respectively. (e) shows the 4 IDs with different modalities, view-points and scales samples. + +results. Existing remote sensing vehicle ReID datasets, such as VRAI [31], UAV-VeID [29], VRU [24], and VeRi-UAV [19], primarily focus on the single RGB modality, lacking the infrared data necessary for cross-modality analysis. For this, [22] has introduced the first UAV RGB-IR vehicle ReID dataset, UCM-VeID, to support cross-modality vehicle ReID research. However, this dataset is constrained by its single data acquisition platform and limited number of + +identities, reducing its suitability for generalizing to diverse real-world scenarios. Consequently, developing large-scale and more diverse vehicle VT-ReID datasets is essential to advance research and better meet the requirements of practical applications. + +Another key challenge in vehicle VT-ReID development is the need for further research on VT-ReID methods, particularly to address the modality bias training (MBT) problem, which remains under-explored. As introduced by [16], MBT refers to the modality imbalance that arises when VT-ReID models are initialized with pre-trained ImageNet weights [6]. The modality imbalance stems from training on RGB-exclusive ImageNet data, which skews feature learning towards the RGB modality and reduces the IR discrimination capability. Moreover, the MBT issue is amplified in large-scale multi-modal datasets: as shown in Fig. 2, models initialized with ImageNet pre-trained weights exhibit a performance drop on larger datasets, particularly in terms of discriminative ability. The potential underlying cause of the above phenomenon is that ImageNet pretrained weights perform well when handling infrared data with lower modality diversity. However, they lack sufficient generalization ability when faced with larger datasets that exhibit higher modality diversity, which exacerbates modality bias and degrades model performance. + +To address these challenges, this paper introduces a more comprehensive dataset for vehicle VT-ReID and a novel pre-training paradigm called Cross-modality Patch-mixed Self-supervised Learning (CPSL) to mitigate the MBT problem. The proposed dataset, UCM-VeID V2, includes 113,929 RGB images and 51,162 IR images of 3,407 vehicles captured by five cameras. Compared to UCM-VeID [22], it is significantly expanded across multiple dimensions, including additional cameras and acquisition platforms, extended capture times, a broader range of weather conditions, and more diverse shooting scenarios, as illustrated in Fig. 1. For the MBT challenge, inspired by [13], CPSL aligns the RGB and IR domains by incorporating a Patch-Mixed Image Reconstruction (PMIR) self-supervised task. To further eliminate the modality bias, CPSL designs a Modality Discrimination Adversarial Learning (MDAL) strategy to exact modality invariant features through an adversarial training manner. In addition, considering discrimination insufficiency, we introduce the Modality-Augmented Contrasting Learning (MACL) task, where patch-mixed images are used as augmented positive examples for contrastive learning. + +In summary, our main contributions are as follows: + +- We introduce the UCM-VeID V2 dataset, which significantly enhances the original UCM-VeID with greater scale and diversity across multiple dimensions. +- We propose a novel self-supervised paradigm to pre-train an encoder, replacing the ImageNet-trained ResNet50, to + +![](images/0263be20b404b1e2de1e5a9efece11042bdd6f5df8cc7c52a25946019ee56a66.jpg) +(a) Training based on UCM-VeID + +![](images/74a9675ea7a5b6f8f2451ad9f3d8324cc16bb1c0c18b28049cbc42f7e0711420.jpg) +(b) Training based on UCM-VeID V2 +Figure 2. The feature distribution of pre-trained embeddings is visualized. Panels (a) and (b) show the embeddings in 2D space for models trained on the UCM-VeID and UCM-VeID V2 datasets, respectively, with dots and crosses representing different identities in RGB and IR modalities. The experiment uses the AGW baseline [37]. As indicated by the red circles in (a), the same-ID samples are more compact. In contrast, the same-ID samples are more spread out in (b), which ultimately degrades Re-ID performance. + +![](images/67328e746653e6e617c3de88a8a6e3da2f8881c52a006470abc67f82f6a0a519.jpg) + +address the MBT issue. + +# 2. Related Works + +# 2.1. Vehicle Re-ID Datasets + +Vehicle Re-ID research has flourished over the past few years, significantly benefited by the availability of vehicle Re-ID datasets. The most widely-employed datasets are primarily captured via fixed surveillance cameras, such as VehicleID [20], VeRI-776 [21], VERI-Wild [23] and Vehicle1M [10]. Compared to fixed surveillance, UAV-based cameras offer superior mobility, flexibility, and convenience, leading to the development of key UAV-based vehicle Re-ID datasets such as VRAI [31], UAV-VeID [29], VRU [24], and VeRi-UAV [19]. Furthermore, Liu et al. [22] have advanced the first RGB-IR vehicle Re-ID dataset, named UCM-VeID, to support the research of cross-modality vehicle Re-ID. However, the UCM-VeID dataset exhibits limitations in both size and diversity, highlighting an urgent need for the further expansion and enrichment of the dataset, ensuring it remains a potent benchmark for advancing the frontiers of vehicle Re-ID research in the face of evolving challenges and complexities. + +# 2.2. Modality Bias Training + +The MBT problem is essentially a problem of unbalanced training. Imbalanced model learning has received extensive attention in multi-modal learning [15, 32], and great progresses have been made in VT-ReID [16, 30]. To our best knowledge, there are two distinct approaches to tackle the MBT issue. The first approach is to control the training process, such as learning rate adjusting [15], gradient modulation [27], and ID-exclusive sampling strategy [16]. The second approach is to use modal balanced weights rather than ImageNet pre-trained ResNet50 avoiding MBT problems from the beginning. Wan et al. [30] devised a self- + +
DatasetVenueIdentitiesImagesModalitiesCamerasPlatformCamera ViewTarget StateColorTypeOrientation
VRAIICCV201913022137613RGB2UAVFlexibleMove97—
UAV-VeIDIJCV2021460141917RGB2UAVFlexibleMove——4
UCM-VeIDTMM202475329928RGB/IR2UAVFlexibleMove&Static1298
UCM-VeID V2ours3407165091RGB/IR5UAVFlexibleMove&Static1298
+ +Table 1. Comparisons among other vehicle Re-ID datasets and UCM-VeID V2 + +supervised pre-training method to obtain a set of new unbiased weights, which achieves comparable performance to that of ImageNet-trained ResNet50. The second approach addresses the MBT problem from the outset by using pretrained weights that are modally balanced, instead of relying on subsequent adjustments. In contrast to the first approach, the second method more fundamentally resolves the modality bias, offering stronger generalization and more stable performance. Consequently, we follow the second approach to carry out the MBT problem. + +# 2.3. Self-Supervised Learning + +Self-supervised learning, as a subset of unsupervised learning, constructs pretext tasks to extract valuable insights from large-scale unlabeled data, thereby learning representations beneficial for downstream tasks. Common methods involve contextual pretext tasks to learn discriminative features, such as Word2Vec [25], Jigsaw [26], Rotation [8], Colorization [38], and MAE [14]. Another prominent approach, contrastive learning, builds discriminative representations by comparing instances within the same feature space, as seen in methods like MoCo [12] and BYOL [9], both of which perform well on downstream tasks. Nevertheless, above methods are not specifically tailored for the Re-ID task, particularly for VT-ReID task. In this paper, we leverage the consistency of different modalities to construct a patch-mixed image reconstruction self-supervised task aimed at extracting shared features, thereby mitigating the modality discrepancy. + +# 3. Benchmark: UCM-VeID V2 + +To further enhance the UAV cross-modal vehicle re-identification dataset, we have significantly expanded the first edition, UCM-VeID [22], by an order of magnitude in terms of UAV platforms, identities and so on. This section details the expanded dataset called UCM-VeID V2. + +# 3.1. Dataset Description + +(1) Dataset Partition. UCM-VeID V2 comprises a total of 3407 identities with 165091 samples. The dataset is split into a training set and a test set in an $8:2$ ratio, resulting in a training set with 2,693 identities and 131,451 samples, while the test set contains 674 identities and 33,640 samples. Furthermore, the query set and gallery set are derived + +from the test set. Specifically, one sample is drawn from each ID within the test set to constitute the query set, while the remaining samples collectively form the gallery set. It is worth noting that the extraction process is random, and crucially, every sample in the query set has a corresponding sample with the same ID in the gallery set. + +(2) Dataset Statistics. To maximize the utility of the UCM-VeID V2 dataset, we perform a detailed statistical analysis of its key attributes, providing insights into the data patterns relevant to Re-ID tasks. + +- ID distribution of 5 cameras. Data is collected using 3 drone platforms equipped with 5 different cameras. $D300_{z}$ camera captures fewer IDs compared to the other four cameras, as Fig. 3(a) shown. The $D300_{z}$ camera captures high-resolution images, primarily focused on nighttime data, thereby enhancing the dataset's diversity. +- Distribution of ID samples. Samples are collected from different camera perspectives, ensuring diversity and ample representation for each ID. As shown in Fig. 3(b), most IDs have 20-40 samples, providing comprehensive coverage of target's appearance across different conditions. +- Distribution of Resolution. As shown in Fig. 3(c), resolutions vary significantly across cameras. For example, $D300_{i}r$ samples cluster around 10,000-15,000 pixels, while DM3P samples are predominantly 45,000-50,000 pixels, highlighting a resolution disparity that presents a core challenge in UAV-based VT-ReID. +- Energy Distribution of IR Samples. The dataset spans all seasons and covers a 24-hour period, capturing both dynamic and static vehicles, leading to diverse infrared energy levels. Fig. 3(d) reveals that most infrared energy values are between 80-100, corresponding to moderate daylight. Extreme conditions, like intense heat or cold, are less common but contribute to the dataset's breadth. +(3) Dataset Attributes. Attributes annotation is an integral part of building reliable, high-performance data sets. In this dataset, we annotate the below attributes. + +- Vehicle Type. Vehicles are classified into categories such as car, SUV, minibus, MPV, bus, mini truck, truck, and pickup. As shown in Fig. 3(e), cars and SUVs make up the majority, while trucks and pickups are less frequent. This distribution reflects the urban focus of our data collection, where cars and SUVs are more prevalent. + +- Vehicle color. We categorize vehicle colors into thirteen + +![](images/a19c874acef395028b97b1d4c6f296b53e6c41657a75ee410e33daba391ad4d0.jpg) +ID Distribution of Different Cameras +(a) + +![](images/fec54423b6a3d274c3dbb2103f4febe2a2c904b7d30e96f8a2bf167fd9ef8d10.jpg) +Distribution of ID Samples +(b) + +![](images/b7c36c07a56e51a86fe6656e404885347f0ceffa1bcb92af4e584b1f369f37eb.jpg) +Distribution of Resolution +(c) + +![](images/64f782825eaf9bef9199b5e3372266e79ed829407f64abf0e56b910c3e094097.jpg) +Energy Distribution of IR Samples +(d) + +![](images/50986752ee54355497badb93d78ca4013f9b8eab9a452f6f2128475caf0a78b9.jpg) +Distribution of Vehicle Type +(e) + +![](images/27e34b5eb58b0121bbaf46bb74036621e70a7cbdfb626ad9e6de631beade163d.jpg) +Distribution of Vehicle Color +(f) + +![](images/97e128641af1f8169fae5af1768c0fc7e4b829b54a37555caeef39f3e8b71957.jpg) +Distribution of Orientation Class +(g) + +![](images/ff207fa11f345721ae98eed1447981517ea1d712e0df69e5228509cf987ab27a.jpg) +Orientation Classification Criteria +(h) +Figure 3. The statistics and annotation attributes of UCM-VeID V2 dataset. Notation that, the 'k' in (c) means the unit of 1000 pixels. + +classes, shown in Fig. 3(f), ordered by proportion as follows: white, black, gray, red, dark gray, cyan, brown, blue, golden, yellow, others, green, and orange. Despite some imbalance, this distribution mirrors real-world vehicle color trends. + +- Vehicle Orientation. Based on [22] and Fig. 3(h), we define eight orientation categories. Fig. 3(g) shows that orientations 2, 4, 6, and 8 are most common, a result of the UAV's flexible viewing angles. This orientation imbalance highlights the dataset's realism and the challenges in UAV-based Re-ID tasks. + +# 3.2. Comparison with Other Datasets + +In contrast to other vehicle Re-ID datasets, as detailed in Tab. 1, our dataset has several unique characteristics: + +1. Multi modal information. Compared with RGB-based datasets, such as VRAI [31], UAV-VeID [29], UCMVeID [22] and UCM-VeID V2 integrate both RGB and IR modalities, facilitating cross-modality vehicle Re-ID tasks, which holds promise as a viable solution to achieve round-the-clock vehicle Re-Identification. +2. Flexible viewpoint, scale and orientations. The UCMVeID V2 dataset, captured via a drone-mounted camera, offers varied perspectives, scales, and orientations that differ from fixed-camera datasets like RGBN300 [18] and MSVR310 [40]. This diversity substantially enriches the dataset, making it a valuable resource for robust vehicle Re-ID research. +3. Overall enhancement. Compared to UCM-VeID, the UCM-VeID V2 dataset offers a substantial increase in + +data volume and significant improvements in several dimensions, including additional cameras and platforms, extended capture times, a broader variety of weather conditions, and a wider range of shooting scenarios, as shown in Fig. 1. These enhancements make it a richer and more comprehensive dataset, supporting diverse and robust data-driven research and applications. + +# 4. Proposed pre-training method + +# 4.1. Overall Model Structure. + +The proposed pre-training VT-ReID method is illustrated in Fig. 4. Firstly, style learning and image generation techniques are used to create pairs of RGB and IR images. Then, we segment these images into patches and randomly swap patches between different modalities to create two new patch-mixed images. Finally, the patch-mixed images are fed into the encoder to complete Patch-Mixed Image Reconstruction (PMIR), Modality Discrimination Adversarial Learning (MDAL) and Modality-Augmented Contrasting Learning (MACL) tasks, where PMIR and MDAL are adopted to learn modality-invariant features, and MACL is utilized to enhance ID discriminability. + +# 4.2. Patch-Mixed Unified Embedding + +Cross-modality data generation. Given the limited size of Re-ID dataset, we first train a CycleGAN [41] to generate cross-modality versions of ImageNet. Specifically, we use CycleGAN to learn style mapping between RGB and IR images from the UCM-VeID V2 dataset with a gener + +![](images/aada07c53c2ccd2652fbc21c54a0bfdb8be043485aaae892ad35a6544e049979.jpg) +Figure 4. Pre-training paradigm, which comprises preprocessing and three self-supervised tasks. Preprocessing involves obtaining paired RGB and IR images through style learning and image generation, followed by patches segmentation and swapping, PMIR and MDAL then focus on learning modality-invariant features, while MACL aims to enhance the ID discriminability. + +ator $G$ . To avoid data leakage, we use ImageNet dataset for pre-training. For an input RGB image $\pmb{x}_i^{\mathrm{RGB}}$ in the ImageNet, we generate the corresponding IR image as $\pmb{x}_i^{\mathrm{IR}} = G(\pmb{x}_i^{\mathrm{RGB}})$ . The resulting cross-modality ImageNet dataset is then $\{(\pmb{x}_i^{\mathrm{RGB}},\pmb{x}_i^{\mathrm{IR}})\}_{i = 1}^N$ , where $N$ is the total number. + +Random patch mixture. Our method aims to teach the backbone to fully leverage the cross-modality information for modality bias elimination. Therefore, for the data input aspect, we mixture the RGB and IR image pair in a patch swapping manner. Mathematically, given a RGB-IR image pair $(\pmb{x}_i^{\mathrm{RGB}},\pmb{x}_i^{\mathrm{IR}})$ , we first establish a patch mask $M$ with the same size as $\pmb{x}_i^{\mathrm{RGB}}(\pmb{x}_i^{\mathrm{IR}})$ that consists of $K\times K$ regular non-overlapping patches. With a predefined mask ratio $\eta$ , $\lfloor \eta \times K\times K\rfloor$ patches in $M$ are randomly selected following a uniform distribution and set as all-zeros matrices. In this paper, we predefined a mask ratio $\eta = 0.75$ . A higher $\eta$ leads to more patch exchanges, increases the difficulty of the reconstruction task, and enables the model to learn more shared features. Then, the process of the proposed random path mixture can be formulated as: + +$$ +\begin{array}{l} \hat {\boldsymbol {x}} _ {i} ^ {\mathrm {R G B}} = \boldsymbol {x} _ {i} ^ {\mathrm {R G B}} \odot \boldsymbol {M} + \boldsymbol {x} _ {i} ^ {\mathrm {I R}} \odot (1 - \boldsymbol {M}) \\ \hat {\boldsymbol {x}} _ {i} ^ {\mathrm {I R}} = \boldsymbol {x} _ {i} ^ {\mathrm {I R}} \odot \boldsymbol {M} + \boldsymbol {x} _ {i} ^ {\mathrm {R G B}} \odot (1 - \boldsymbol {M}), \end{array} \tag {1} +$$ + +Then, based on the Eq. (1), the cross-modality patch-mixed unified embeddings are finally achieved as: + +$$ +\boldsymbol {f} _ {i} ^ {\mathrm {R G B}}, \boldsymbol {f} _ {i} ^ {\mathrm {I R}} = \boldsymbol {F} \left(\hat {\boldsymbol {x}} _ {i} ^ {\mathrm {R G B}}\right), \boldsymbol {F} \left(\hat {\boldsymbol {x}} _ {i} ^ {\mathrm {I R}}\right), \tag {2} +$$ + +where $F(\cdot)$ is the feature extractor and ResNet50 [11] is employed in this paper. $f_{i}^{\mathrm{RGB}}$ and $f_{i}^{\mathrm{IR}}$ are the latent embeddings of mixed images. + +# 4.3. Self-supervised Multi-task Learning. + +With the proposed unified embedding module, we can acquire the latent features containing cross-modality information. However, these features can not be directly employed + +to the Re-ID task due to modality-bias information and insufficient discriminability [30, 34, 37]. To address above these issues, we devise three self-supervised tasks. + +Patch-Mixed Image Reconstruction. In this task, we utilize the mixed features, $f_{i}^{\mathrm{RGB}}$ and $f_{i}^{\mathrm{IR}}$ , to reconstruct the input images, $x_{i}^{\mathrm{RGB}}$ and $x_{i}^{\mathrm{IR}}$ , respectively. This approach is motivated by the inherent advantage of image reconstruction as a naturally unsupervised learning task, eliminating the need for additional labels and making it suitable for training large models. Moreover, for RGB image reconstruction, $f_{i}^{\mathrm{RGB}}$ contains minimal RGB information but abundant IR information, compelling the network to focus on modality-shared features. The limited RGB information acts as style information that supports the reconstruction of the RGB image. Similarly, for IR image reconstruction, $x_{i}^{\mathrm{IR}}$ using $f_{i}^{\mathrm{IR}}$ , where the feature contains mostly RGB information with only a small amount of IR data, thus also ensuring the model learns to emphasize cross-modality features effectively. According to above analysis, the process of patch-mixed image reconstruction can be formulated as: + +$$ +\begin{array}{r l} \bar {\boldsymbol {x}} _ {i} ^ {\mathrm {R G B}} & = \boldsymbol {H} _ {\mathrm {R G B}} \left(\boldsymbol {f} _ {i} ^ {\mathrm {R G B}}\right) \\ \bar {\boldsymbol {x}} _ {i} ^ {\mathrm {I R}} & = \boldsymbol {H} _ {\mathrm {I R}} \left(\boldsymbol {f} _ {i} ^ {\mathrm {I R}}\right), \end{array} \tag {3} +$$ + +where $\bar{\pmb{x}}_i^{\mathrm{RGB}}$ and $\bar{\pmb{x}}_i^{\mathrm{IR}}$ are the reconstructed RGB and IR images. $H_{\mathrm{RGB}}(\cdot)$ and $H_{\mathrm{IR}}(\cdot)$ are networks for RGB and IR image reconstruction respectively. Finally, a pixel-wise image reconstruction loss with $l_{2}$ -norm is employed to guide the reconstruction process as: + +$$ +\mathcal {L} _ {\mathrm {r e c}} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left(\| \bar {\boldsymbol {x}} _ {i} ^ {\mathrm {R G B}} - \boldsymbol {x} _ {i} ^ {\mathrm {R G B}} \| _ {2} ^ {2} + \| \bar {\boldsymbol {x}} _ {i} ^ {\mathrm {I R}} - \boldsymbol {x} _ {i} ^ {\mathrm {I R}} \| _ {2} ^ {2}\right). \tag {4} +$$ + +Modality Discrimination Adversarial Learning. In this task, we introduce adversarial training to encourage + +the learned features to be modality-invariant. Specifically, we employ a discriminator $D(\cdot)$ to distinguish between the mixed features $f_{i}^{\mathrm{RGB}}$ and $f_{i}^{\mathrm{IR}}$ . Initially, we fix the feature extractor $F$ and train the discriminator $D$ to correctly classify the mixed features based on their corresponding modality labels. This step ensures that $D$ can accurately distinguish between features from different modalities. The objective for the discriminator $D$ can be formulated as: + +$$ +\mathcal {L} _ {\mathrm {D}} = - \frac {1}{N} \sum_ {i = 1} ^ {N} \left[ y _ {i} \log \boldsymbol {D} (\boldsymbol {f} _ {i}) + (1 - y _ {i}) \log (1 - \boldsymbol {D} (\boldsymbol {f} _ {i})) \right], \tag {5} +$$ + +where $y_{i}$ is the modality label, $f_{i}$ represents either $f_{i}^{\mathrm{RGB}}$ or $f_{i}^{\mathrm{IR}}$ , and $K$ is the batch size. In the second step, we fix the discriminator $D$ and update the feature extractor $F$ to generate mixed features that the discriminator cannot distinguish. This step forces the feature extractor to produce modality-invariant features, effectively minimizing the modality bias in the learned features. The objective for the feature extractor $F$ during this phase is to maximize the classification error of $D$ as: + +$$ +\mathcal {L} _ {\mathrm {F}} = - \frac {1}{N} \sum_ {i = 1} ^ {N} \left[ \left(1 - y _ {i}\right) \log \boldsymbol {D} (\boldsymbol {f} _ {i}) + y _ {i} \log \left(1 - \boldsymbol {D} (\boldsymbol {f} _ {i})\right) \right]. \tag {6} +$$ + +The adversarial training process iterates between these two steps until the model converges, achieving a balance where the discriminator $D$ cannot reliably distinguish the modality of the features generated by $F$ . + +Modality-Augmented Contrasting Cluster. In this task, we focus on enhancing the discriminative power of the learned features by employing a contrastive learning approach. Unlike traditional methods that rely on data augmentation to generate negative samples, our approach does not introduce additional data augmentation but instead utilizes the inherent modality-specific features for contrastive learning. We treat $f_{i}^{\mathrm{RGB}}$ and $f_{i}^{\mathrm{IR}}$ as positive pairs within the contrastive learning framework. The contrastive loss function aims to maximize the similarity between the mixed features of different modalities, which can be formulated as: + +$$ +\mathcal {L} _ {\text {c o n t r a s t}} = - \frac {1}{N} \sum_ {i = 1} ^ {N} \log \frac {\exp \left(\operatorname {s i m} \left(\boldsymbol {f} _ {i} ^ {\mathrm {R G B}} , \boldsymbol {f} _ {i} ^ {\mathrm {I R}}\right)\right)}{\sum_ {j = 1} ^ {N} \exp \left(\operatorname {s i m} \left(\boldsymbol {f} _ {i} ^ {\mathrm {R G B}}, \boldsymbol {f} _ {j} ^ {\mathrm {I R}}\right)\right)}, \tag {7} +$$ + +where $\mathrm{sim}(\cdot ,\cdot)$ denotes a similarity measure (such as cosine similarity) between features. The loss function encourages the feature extractor to learn representations that pull together positive pairs $(f_{i}^{\mathrm{RGB}}$ and $f_{i}^{\mathrm{IR}})$ while pushing apart other features. By iteratively optimizing this contrastive loss, the feature extractor $\pmb{F}$ is guided to produce more discriminative features, thereby improving the overall performance of the cross-modal image re-identification task. + +In summary, the overall loss function for the entire training process can be expressed as: + +$$ +\mathcal {L} _ {\text {a l l}} = \mathcal {L} _ {\text {r e c}} + \left(\mathcal {L} _ {\mathrm {D}} + \mathcal {L} _ {\mathrm {F}}\right) + \mathcal {L} _ {\text {c o n t r a s t}}. \tag {8} +$$ + +# 5. Experiments + +# 5.1. Implementation and Evaluation Protocols + +Pre-training Stage. For pre-training, we implement CPSL using PyTorch on an NVIDIA RTX A100. Optimization is conducted with an SGD optimizer (momentum of 0.9, weight decay of 0.0005), an initial learning rate of 0.1 with linear warm-up, and a cosine decay schedule without restarts. Data augmentation includes random grayscale, horizontal flipping, Gaussian filtering, and interpolation sampling. Images are resized to $256 \times 256$ , and the mask size $N$ is set to 4. + +Fine-Tuning Stage. For fine-tuning, we implement VT-ReID models using PyTorch on an NVIDIA GeForce RTX 3090 GPU. Images are resized to $180 \times 256$ . Optimization uses an SGD optimizer (momentum of 0.9, weight decay of 0.0005) with an initial learning rate of 0.1 and linear warm-up over 80 epochs. The model is initialized with CPSL checkpoints while other hyper-parameters and data augmentations remain unchanged, as this setup is not specifically tuned for CPSL. Further optimization may improve results, which we leave for future work. + +Evaluation Protocols. Following [34], we use the cumulative matching characteristic (CMC) curve and mean average precision (mAP) for evaluation. CMC assesses rank-k accuracy, with rank-1, rank-10, and rank-20 scores reported in our experiments. mAP reflects the overall retrieval performance across all query samples. + +# 5.2. Comparison with pre-trained weights + +To validate the effectiveness of our proposed pre-training method, we employ three different backbone initialization strategies during fine-tuning: random initialization (Random Init.), initialization with ImageNet pre-trained weights (ImageNet-1K), and initialization using the proposed pretraining weights (CPSL). To enhance experimental rigor, we perform supervised fine-tuning with four state-of-the-art VT-ReID methods: AGW [37], DDAG [36], MMD-ReID [17], and DEEN [39]. + +From Tab. 2, we can observe that the VT-ReID models trained from scratch generally under-performance, with all methods failing to converge. The primary cause of the gradient explosion is attributed to the triplet loss, suggesting that randomly initialized models struggle to establish semantically meaningful Euclidean feature spaces. + +Our method CPSL achieves superior results across all testing and query modes. In contrast to ImageNet-1K pretraining, we do not rely on additional training data or incorporate any supervised training processes. The superior performance is primarily attributed to the design of our self-supervised tasks, which account for modality bias and eliminate cross-modality differences. + +It can also be observed that CPSL consistently improves + +
MethodVenuePre-TrainSingle-shotMulti-shot
IR2RGBRGB2IRIR2RGBRGB2IR
rank1rank10mAPrank1rank10mAPrank1rank10mAPrank1rank10mAP
AGW [37]TPAMI 2021Random Init.4.3518.759.384.8919.79.297.2227.264.826.6726.175.53
ImageNet-1K15.7545.9825.7416.1747.5826.5420.158.1216.2120.4558.1618.3
CPSL23.4753.0633.6925.3658.5436.5731.9167.5424.1332.6468.0724.01
DDAG [36]ECCV 2020Random Init.5.2222.8111.485.0622.6911.228.5532.725.847.0630.056.25
ImageNet-1K20.4648.7830.1219.4649.1529.2127.5864.7520.2826.6965.5422.87
CPSL22.2350.8631.9923.4551.3332.9431.9667.9421.731.1666.5824.51
MMD-ReID [17]BMVC 2021Random Init.4.3415.279.864.9218.539.935.9919.3710.785.6920.3410.42
ImageNet-1K18.9348.6428.9919.2749.1429.0822.0455.8719.6422.9659.2731.41
CPSL21.9851.8532.3922.5253.3831.9928.3462.0922.129.0462.8521.8
DEEN [39]CVPR 2023Random Init.7.2931.7415.246.1127.0613.297.8837.447.487.3932.817.34
ImageNet-1K18.2653.0130.3220.4656.2332.0923.8364.1422.0622.6566.6121.43
CPSL22.5759.8531.0923.759.5633.3525.6267.5223.4424.5368.6323.97
+ +* The bold entries represent the best performance in table. + +performance across various VT-ReID methods, which indicates that the weights generated by CPSL exhibit strong robustness and adaptability. This suggests that the CPSL method is effective in enhancing the generalization capabilities of the models, allowing them to perform well under different conditions and with diverse Re-ID techniques. + +# 5.3. Comparisons With Other SSL Methods + +We compare CPSL with multiple ResNet50 pre-trained models on the UCM-VeID V2. The methods involved can be categorized into four groups: (1) method based on VT-ReID fine-tuning (MMGL [30]), (2) method based pretext task design (Jigsaw [26], Rotation [8]), (3) method based self-supervised learning (SSL) pre-training (InstDisc [35], SimCLR [3], SwAV [2], BYOL [9], SimSiam [4], MoCoV2 [5], DenseCL [33], Relative-Loc[7]). + +As shown in Tab. 3, MMGL performs suboptimally. This is partly because MMGL is tailored for pedestrian targets, using a top-down human segmentation approach that struggles with vehicles' variable orientations. Additionally, the released MMGL weights were fine-tuned on pedestrian VT-ReID datasets, embedding pedestrian-specific features that do not generalize well to vehicle targets. + +SSL methods share a similar goal with our approach by aiming to endow the encoder with discriminative capabilities. However, methods such as SwAV [2] and BYOL [9] perform poorly, as indicated by the stagnant triplet loss during fine-tuning, suggesting these methods fail to establish a meaningful Euclidean space from the outset. Although some SSL methods outperform those based on pretext task design, they still underperform compared to CPSL. This is because they focus solely on target discriminability through + +Table 2. Comparisons between the proposed CPSL and some state-of-the-art methods on UCM-VeID V2 datasets. The rank-k (%) and mAP (%) are demonstrated. + +
MethodRank1Rank10Rank20mAP
method based on VT-ReID fine-tuning:
MMGL [30]21.8256.5668.0316.14
method based pretext task design:
Jigsaw [26]21.2456.7767.2515.66
RotationNet [8]26.6263.0372.7719.32
method based on SSL pre-training:
InstDisc [35]24.0962.0773.1620.04
SimCLR [3]12.4348.5762.1511.17
SwAV [2]6.5728.8940.225.17
BYOL [9]6.8925.5136.875.17
SimSiam [4]27.2465.7473.1322.36
MoCo V2 [5]28.8967.4674.9822.44
DenseCL [33]27.8964.9975.1222.41
Relative-Loc[7]18.9155.7968.0615.53
ours32.6468.0776.3224.01
+ +Table 3. Comparison with other SSL methods. The rank-k (%) and mAP (%) are demonstrated. + +contrastive learning without addressing modality differences. In contrast, our PMIR and MDAL modules effectively mitigate cross-modality differences, enabling shared feature learning. + +# 5.4. Ablation Study + +To assess the effectiveness of each component in our method, we conducted ablation studies on the proposed pretraining method, removing specific modules while keeping all other experimental settings constant. Our method includes three key modules: PMIR (R), MDAL (D), and + +
MethodSingle-shotMulti-shot
rank1rank10mAPrank1rank10mAP
Random Init.4.8919.79.296.6726.175.53
ImageNet-1K16.1747.5826.5420.4558.1618.3
R7.6426.9312.219.3342.7210.59
R+D14.6349.5722.4618.8854.9518.32
R+C22.4955.4830.3428.9765.1621.96
R+D+C25.3658.5436.5732.6468.0724.01
+ +MACL (C). Here, R and D target modality alignment, while C enhances discriminative representation. The ablation study results are shown in Tab. 4. + +Effectiveness of PMIR. PMIR aims to create a shared latent representation space for RGB and IR modalities via reconstruction tasks. As shown in Tab. 4, compare to the results of Random Init., RANK1 and mAP improve by $2.57\%$ and $2.92\%$ under the single-shot query mode, indicating PMIR's effectiveness in reducing modality discrepancies. + +Effectiveness of MDAL. MDAL uses modality adversarial learning to further align modalities. Tab. 4 shows a RANK1 and mAP increase of $6.99\%$ and $10.25\%$ under single-shot query mode, demonstrating that MDAL helps guide the encoder toward shared feature learning. + +Effectiveness of MACL. MACL employs modality contrastive learning to enhance feature clustering and discriminative power. As evidenced in Tab. 4, RANK1 and mAP improve by $7.86\%$ and $7.88\%$ in single-shot query mode, confirming MACL's success in learning discriminative features for VT-ReID. + +Table 4. The quantitative results of ablation studies. The rank-k $(\%)$ and mAP $(\%)$ are demonstrated. + +
Pn2 × 23 × 34 × 45 × 5
rank124.6627.3232.6429.05
mAP16.3820.9124.0122.44
+ +Table 5. Different partition stripe numbers ${P}_{n}\left( {N \times N}\right)$ . The rank-k (%) and mAP (%) are demonstrated. + +# 5.5. Parameter Analysis + +The partition stripe number $P_{n}$ directly determines the size of the image blocks to be swapped and is an important hyperparameter of our method. To verify the impact of this hyperparameter, we conducted experiments with different values for the partition stripe number, with the relevant results presented in Tab. 5. From the table, it can be observed that as the partition stripe number increases, the model's performance initially improves before declining. This is because, as the partition stripe number increases, the semantic content within each image block transitions from being more + +![](images/6f3023b35357547ed5b2e3343fcb0c70f381171787f24a24e2e9786da093930e.jpg) +(a) ImageNet Pre-Training + +![](images/238a5c5513c100a4ca072473311bcbd5f265e17604c36a57f048aa1c3fcbfe4a.jpg) +(b) PMSL Pre-Training +Figure 5. Visualization of the pre-learned features. Different colors represent distinct identities, while circle and cross symbols denote RGB and IR modalities, respectively. The experiment is conducted on UCM-VeID V2 dataset with the AGW baseline [37]. + +![](images/23e415c255f9555e89e81104ef8eb30a508277ab37c7089ab7c19fcc34c15183.jpg) + +discrete to becoming sparser. This indicates that in practical applications, appropriate adjustment of this parameter is necessary to achieve optimal performance. + +# 5.6. Visualization of the feature distribution + +To better understand the proposed pre-training paradigm, we visualize the t-SNE distributions of features pre-learned using ImageNet-1K and the CPSL checkpoints on the UCM-VeID V2 dataset, utilizing AGW [37]. As shown in Fig. 5, we randomly select 15 identities to visualize all corresponding images. + +Fig. 5 demonstrates that, the modal gap between PMSL features is smaller than that between ImageNet features (highlighted by the red circle). Specifically, PMSL features show less variation across modalities, indicating better consistency and robustness when processing RGB and IR images. In contrast, ImageNet-pretrained features exhibit significant modal variation, which can negatively impact performance in VT-ReID tasks. Thus, PMSL not only eliminates cross-modality discrepancy but also improves the model's discriminative ability for downstream tasks, providing stronger support for cross-modal vehicle re-identification. + +# 6. Conclusion + +In this paper, we introduce a novel dataset that expands existing ones in terms of scale, acquisition platforms, collection scenarios, time, and climate conditions. To address the MBT problem, we break it into two sub-problems: modality discrepancy and insufficient discriminability. We propose an end-to-end cross-modal pre-training framework to address these issues. To reduce modality discrepancies, the framework incorporates self-supervised tasks based on image block swapping and modality adversarial learning. For enhanced discriminability, we design an unsupervised clustering-guided multi-modal contrastive learning approach. Extensive experiments on the new dataset validate the effectiveness of our method. + +Acknowledgements. This work was supported by the National Natural Science Foundation of China (62201586). + +# References + +[1] Yan Bai, Yihang Lou, Feng Gao, Shiqi Wang, Yuwei Wu, and Ling-Yu Duan. Group-sensitive triplet embedding for vehicle reidentification. IEEE Trans. Multim., 20(9):2385-2399, 2018. 1 +[2] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In Proc. Adv. Neural Inf. Process. Syst., pages 9912-9924, 2020. 7 +[3] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In Proc. Int. Conf. Mach. Learn., pages 1597-1607, 2020. 7 +[4] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 15750-15758, 2021. 7 +[5] Xinlei Chen, Haoqi Fan, Ross B. Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. CoRR, abs/2003.04297, 2020. 7 +[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proc. IEEE Conf. Comput. Vis. Pattern Recognition., pages 248-255, 2009. 2 +[7] Carl Doersch, Abhinav Gupta, and Alexei A. Efros. Unsupervised visual representation learning by context prediction. In Proc. IEEE Int. Conf. Comput. Vis., pages 1422-1430. IEEE Computer Society, 2015. 7 +[8] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In Proc. Int. Conf. Learn. Represent., 2018. 3, 7 +[9] Jean-Bastien Grill, Florian Strub, Florent Alché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Ávila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent - A new approach to self-supervised learning. In Proc. Adv. Neural Inf. Process. Syst., pages 21271-21284, 2020. 3, 7 +[10] Haiyun Guo, Chaoyang Zhao, Zhiwei Liu, Jinqiao Wang, and Hanqing Lu. Learning coarse-to-fine structured feature embedding for vehicle re-identification. In Proc. AAAI Conf. Artif. Intell., pages 6853–6860, 2018. 2 +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 770-778, 2016. 5 +[12] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 9726-9735, 2020. 3 +[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 15979-15988, 2022. 2 + +[14] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 15979-15988, 2022. 3 +[15] Yu Huang, Junyang Lin, Chang Zhou, Hongxia Yang, and Longbo Huang. Modality competition: What makes joint training of multi-modal network fail in deep learning? (provably). In Proc. Int. Conf. Mach. Learn., pages 9226-9259, 2022. 2 +[16] Yan Huang, Qiang Wu, Jingsong Xu, Yi Zhong, Peng Zhang, and Zhaoxiang Zhang. Alleviating modality bias training for infrared-visible person re-identification. IEEE Trans. Multim., 24:1570-1582, 2022. 2 +[17] Chaitra Jambigi, Ruchit Rawal, and Anirban Chakraborty. MMD-ReID: A simple but effective solution for visible-thermal person reid. In Proc. BMVC, pages 12026–12035, 2021. 6, 7 +[18] H. Li, C. Li, X. Zhu, A. Zheng, and B. Luo. Multi-spectral vehicle re-identification: A challenge. In Proc. AAAI Conf. Artif. Intell., pages 11345–11353, 2020. 4 +[19] Chunsheng Liu, Ye Song, Faliang Chang, Shuang Li, Ruimin Ke, and Yinhai Wang. Posture calibration based cross-view and hard-sensitive metric learning for UAV-based vehicle re-identification. IEEE Trans. Inf. Forensics Secur., 23:19246-19257, 2022. 1, 2 +[20] Hongye Liu, Yonghong Tian, Yaowei Wang, Lu Pang, and Tiejun Huang. Deep relative distance learning: Tell the difference between similar vehicles. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 2167-2175, 2016. 2 +[21] Xinchen Liu, Wu Liu, Tao Mei, and Huadong Ma. A deep learning-based approach to progressive vehicle re-identification for urban surveillance. In Proc. Eur. Conf. Comput. Vis., pages 869–884, 2016. 2 +[22] Xingyue Liu, Jiahao Qi, Chen Chen, Kangcheng Bin, and Ping Zhong. Beyond sharing weights in decoupling feature learning network for UAV rgb-infrared vehicle re-identification. IEEE Trans. Multim., 2024. 1, 2, 3, 4 +[23] Yihang Lou, Yan Bai, Jun Liu, Shiqi Wang, and Lingyu Duan. VERI-Wild: A large dataset and a new method for vehicle re-identification in the wild. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3235-3243, 2019. 2 +[24] Mingming Lu, Yongchuan Xu, and Haifeng Li. Vehicle re-identification based on UAV viewpoint: Dataset and method. Remote. Sens., 14(18):4603, 2022. 1, 2 +[25] Tomás Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In Proc. Int. Conf. Learn. Represent., 2013. 3 +[26] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In Proc. Eur. Conf. Comput. Vis., pages 69-84, 2016. 3, 7 +[27] Xiaokang Peng, Yake Wei, Andong Deng, Dong Wang, and Di Hu. Balanced multimodal learning via on-the-fly gradient modulation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 8228-8237, 2022. 2 +[28] Shangzhi Teng, Shiliang Zhang, Qingming Huang, and Nicu Sebe. Multi-view spatial attention embedding for vehicle re-identification. IEEE Trans. Circuits Syst. Video Technol., 31 (2):816-827, 2021. 1 + +[29] Shangzhi Teng, Shiliang Zhang, Qingming Huang, and Nicu Sebe. Viewpoint and scale consistency reinforcement for UAV vehicle re-identification. Int. J. Comput. Vis., 129(3): 719-735, 2021. 1, 2, 4 +[30] Lin Wan, Qianyan Jing, Zongyuan Sun, Chuang Zhang, Zhihang Li, and Yehansen Chen. Self-supervised modality-aware multiple granularity pre-training for rgb-infrared person re-identification. IEEE Trans. Inf. Forensics Secur., 18: 3044–3057, 2023. 2, 5, 7 +[31] Peng Wang, Bingliang Jiao, Lu Yang, Yifei Yang, Shizhou Zhang, Wei Wei, and Yanning Zhang. Vehicle re-identification in aerial imagery: Dataset and approach. In Proc. IEEE Int. Conf. Comput. Vis., pages 460–469, 2019. 1, 2, 4 +[32] Weiyao Wang, Du Tran, and Matt Feiszli. What makes training multi-modal classification networks hard? In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 12692-12702, 2020. 2 +[33] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3024-3033. Computer Vision Foundation / IEEE, 2021. 7 +[34] Ancong Wu, Wei-Shi Zheng, Hong-Xing Yu, Shaogang Gong, and Jianhuang Lai. RGB-infrared cross-modality person re-identification. In Proc. IEEE Int. Conf. Comput. Vis., pages 5390-5399, 2017. 5, 6 +[35] Zhirong Wu, Yuanjun Xiong, Stella X. Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3733-3742, 2018. 7 +[36] Mang Ye, Jianbing Shen, David J. Crandall, Ling Shao, and Jiebo Luo. Dynamic dual-attentive aggregation learning for visible-infrared person re-identification. Proc. Eur. Conf. Comput. Vision, 12362:229–247, 2020. 6, 7 +[37] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven C. H. Hoi. Deep learning for person re-identification: A survey and outlook. IEEE Trans. Pattern Anal. Mach. Intell., 44:2872–2893, 2020. 2, 5, 6, 7, 8 +[38] Richard Zhang, Phillip Isola, and Alexei A. Efros. Colorful image colorization. In Proc. Eur. Conf. Comput. Vis., pages 649-666, 2016. 3 +[39] Yukang Zhang and Hanzi Wang. Diverse embedding expansion network and low-light cross-modality benchmark for visible-infrared person re-identification. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 2153–2162, 2023. 6, 7 +[40] Aihua Zheng, Xianpeng Zhu, Chenglong Li, Jin Tang, and Jixin Ma. Cross-directional consistency network with adaptive layer normalization for multi-spectral vehicle re-identification and a high-quality benchmark. Inf. Fusion, 100:101901, 2023. 4 +[41] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proc. IEEE Int. Conf. Comput. Vis., pages 2242-2251, 2017. 4 \ No newline at end of file diff --git a/2025/UCM-VeID V2_ A Richer Dataset and A Pre-training Method for UAV Cross-Modality Vehicle Re-Identification/images.zip b/2025/UCM-VeID V2_ A Richer Dataset and A Pre-training Method for UAV Cross-Modality Vehicle Re-Identification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d821ff4393bad429b22594f486fa7bea2ae916e5 --- /dev/null +++ b/2025/UCM-VeID V2_ A Richer Dataset and A Pre-training Method for UAV Cross-Modality Vehicle Re-Identification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9eac77c63e15d692742f1c10e587d0131f47df574cb70f9106ec9a9feb645393 +size 650421 diff --git a/2025/UCM-VeID V2_ A Richer Dataset and A Pre-training Method for UAV Cross-Modality Vehicle Re-Identification/layout.json b/2025/UCM-VeID V2_ A Richer Dataset and A Pre-training Method for UAV Cross-Modality Vehicle Re-Identification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5ba74890e5464f9b59c1de8a4fddefe735e9957a --- /dev/null +++ b/2025/UCM-VeID V2_ A Richer Dataset and A Pre-training Method for UAV Cross-Modality Vehicle Re-Identification/layout.json @@ -0,0 +1,9206 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 102, + 522, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 102, + 522, + 138 + ], + "spans": [ + { + "bbox": [ + 89, + 102, + 522, + 138 + ], + "type": "text", + "content": "UCM-VeID V2: A Richer Dataset and A Pre-training Method for UAV Cross-Modality Vehicle Re-Identification" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 161, + 487, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 161, + 487, + 190 + ], + "spans": [ + { + "bbox": [ + 126, + 161, + 487, + 190 + ], + "type": "text", + "content": "Xingyue Liu Jiahao Qi Chen Chen KangCheng Bin Ping Zhong* National University of Defense Technology, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 191, + 500, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 191, + 500, + 204 + ], + "spans": [ + { + "bbox": [ + 106, + 191, + 500, + 204 + ], + "type": "text", + "content": "{liuxingyue18, qijiahao1996, chenchen21c, binkc21, zhongping}@nudt.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 255, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 483 + ], + "type": "text", + "content": "Cross-Modality Re-Identification (VT-ReID) aims to achieve around-the-clock target matching, benefiting from the strengths of both RGB and infrared (IR) modalities. However, the field is hindered by limited datasets, particularly for vehicle VT-ReID, and by challenges such as modality bias training (MBT), stemming from biased pre-training on ImageNet. To tackle the above issues, this paper introduces an dataset benchmark, named UCM-VeID V2, for vehicle VT-ReID, and proposes a new self-supervised pre-training method, Cross-Modality Patch-Mixed Self-supervised Learning (PMSL). UCM-VeID V2 dataset features a significant increase in data volume, along with enhancements in multiple aspects. PMSL addresses MBT by learning modality-invariant features through Patch-Mixed Image Reconstruction (PMIR) and Modality Discrimination Adversarial Learning (MDAL), and enhances discriminability with Modality-Augmented Contrasting Cluster (MACC). Comprehensive experiments are carried out to validate the proposed method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 506, + 135, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 135, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 135, + 519 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 526, + 295, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 295, + 646 + ], + "type": "text", + "content": "Cross-modality vehicle re-identification (ReID), also referred to as vehicle visible-infrared (VI) ReID [22], involves retrieving a matching vehicle image from a gallery set captured in the visible (RGB) or infrared (IR) spectrum, given a query vehicle image in the alternate modality. This process, where images are obtained from cameras of differing modalities, has garnered significant attention in applications such as video surveillance and public security [1, 28], primarily due to its advantages in enabling continuous, intelligent monitoring across day and night conditions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 647, + 296, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 296, + 695 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 296, + 695 + ], + "type": "text", + "content": "Despite its importance, research on vehicle VT-ReID remains relatively underdeveloped. A primary challenge lies in the lack of essential resources, including a comprehensive large-scale dataset and widely accessible benchmark" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 230, + 554, + 284 + ], + "blocks": [ + { + "bbox": [ + 316, + 230, + 554, + 284 + ], + "lines": [ + { + "bbox": [ + 316, + 230, + 554, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 554, + 284 + ], + "type": "image", + "image_path": "650e42b7c9288e3d2e5a94cc8eb2f7d7be8c4f5764bcc15f9eddcb9f62c1411f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 404, + 283, + 466, + 291 + ], + "lines": [ + { + "bbox": [ + 404, + 283, + 466, + 291 + ], + "spans": [ + { + "bbox": [ + 404, + 283, + 466, + 291 + ], + "type": "text", + "content": "(a) Multiple Platforms" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 293, + 553, + 346 + ], + "blocks": [ + { + "bbox": [ + 316, + 293, + 553, + 346 + ], + "lines": [ + { + "bbox": [ + 316, + 293, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 316, + 293, + 553, + 346 + ], + "type": "image", + "image_path": "f24f985c52a5b5e52306a3bd1abb9b7ef9fe325624d2186bb5eb1743a4496851.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 346, + 463, + 354 + ], + "lines": [ + { + "bbox": [ + 406, + 346, + 463, + 354 + ], + "spans": [ + { + "bbox": [ + 406, + 346, + 463, + 354 + ], + "type": "text", + "content": "(b) Wider time spans" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 316, + 355, + 553, + 408 + ], + "blocks": [ + { + "bbox": [ + 316, + 355, + 553, + 408 + ], + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 408 + ], + "type": "image", + "image_path": "cd42070e455eb5c1d8d9e73efb3768b4a8c733f11e6a8072d8ebe4c2c6ec027f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 408, + 474, + 416 + ], + "lines": [ + { + "bbox": [ + 394, + 408, + 474, + 416 + ], + "spans": [ + { + "bbox": [ + 394, + 408, + 474, + 416 + ], + "type": "text", + "content": "(c) Richer weather conditions" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 316, + 418, + 553, + 471 + ], + "blocks": [ + { + "bbox": [ + 316, + 418, + 553, + 471 + ], + "lines": [ + { + "bbox": [ + 316, + 418, + 553, + 471 + ], + "spans": [ + { + "bbox": [ + 316, + 418, + 553, + 471 + ], + "type": "image", + "image_path": "e881e6a447ae79ac1ed301ff496a789a0cfd9fef23e009eec007a435d70c4f42.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 471, + 460, + 478 + ], + "lines": [ + { + "bbox": [ + 409, + 471, + 460, + 478 + ], + "spans": [ + { + "bbox": [ + 409, + 471, + 460, + 478 + ], + "type": "text", + "content": "(d) Diverse Scenes" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 316, + 480, + 553, + 533 + ], + "blocks": [ + { + "bbox": [ + 316, + 480, + 553, + 533 + ], + "lines": [ + { + "bbox": [ + 316, + 480, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 480, + 553, + 533 + ], + "type": "image", + "image_path": "79c7482d3305bee0d43233ed7746884924d63901f0a88dd194056ead19dde2ea.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 533, + 451, + 541 + ], + "lines": [ + { + "bbox": [ + 419, + 533, + 451, + 541 + ], + "spans": [ + { + "bbox": [ + 419, + 533, + 451, + 541 + ], + "type": "text", + "content": "(e) Samples" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 551, + 555, + 596 + ], + "lines": [ + { + "bbox": [ + 313, + 551, + 555, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 555, + 596 + ], + "type": "text", + "content": "Figure 1. Increments and samples of UCM-VeID V2. (a)(b)(c)(d) are varies platforms, time periods, weather conditions and locations for data acquisition, respectively. (e) shows the 4 IDs with different modalities, view-points and scales samples." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "results. Existing remote sensing vehicle ReID datasets, such as VRAI [31], UAV-VeID [29], VRU [24], and VeRi-UAV [19], primarily focus on the single RGB modality, lacking the infrared data necessary for cross-modality analysis. For this, [22] has introduced the first UAV RGB-IR vehicle ReID dataset, UCM-VeID, to support cross-modality vehicle ReID research. However, this dataset is constrained by its single data acquisition platform and limited number of" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 143, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22286" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "identities, reducing its suitability for generalizing to diverse real-world scenarios. Consequently, developing large-scale and more diverse vehicle VT-ReID datasets is essential to advance research and better meet the requirements of practical applications." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 134, + 295, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 134, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 134, + 295, + 373 + ], + "type": "text", + "content": "Another key challenge in vehicle VT-ReID development is the need for further research on VT-ReID methods, particularly to address the modality bias training (MBT) problem, which remains under-explored. As introduced by [16], MBT refers to the modality imbalance that arises when VT-ReID models are initialized with pre-trained ImageNet weights [6]. The modality imbalance stems from training on RGB-exclusive ImageNet data, which skews feature learning towards the RGB modality and reduces the IR discrimination capability. Moreover, the MBT issue is amplified in large-scale multi-modal datasets: as shown in Fig. 2, models initialized with ImageNet pre-trained weights exhibit a performance drop on larger datasets, particularly in terms of discriminative ability. The potential underlying cause of the above phenomenon is that ImageNet pretrained weights perform well when handling infrared data with lower modality diversity. However, they lack sufficient generalization ability when faced with larger datasets that exhibit higher modality diversity, which exacerbates modality bias and degrades model performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 375, + 295, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 375, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 56, + 375, + 295, + 638 + ], + "type": "text", + "content": "To address these challenges, this paper introduces a more comprehensive dataset for vehicle VT-ReID and a novel pre-training paradigm called Cross-modality Patch-mixed Self-supervised Learning (CPSL) to mitigate the MBT problem. The proposed dataset, UCM-VeID V2, includes 113,929 RGB images and 51,162 IR images of 3,407 vehicles captured by five cameras. Compared to UCM-VeID [22], it is significantly expanded across multiple dimensions, including additional cameras and acquisition platforms, extended capture times, a broader range of weather conditions, and more diverse shooting scenarios, as illustrated in Fig. 1. For the MBT challenge, inspired by [13], CPSL aligns the RGB and IR domains by incorporating a Patch-Mixed Image Reconstruction (PMIR) self-supervised task. To further eliminate the modality bias, CPSL designs a Modality Discrimination Adversarial Learning (MDAL) strategy to exact modality invariant features through an adversarial training manner. In addition, considering discrimination insufficiency, we introduce the Modality-Augmented Contrasting Learning (MACL) task, where patch-mixed images are used as augmented positive examples for contrastive learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 640, + 274, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 640, + 274, + 651 + ], + "spans": [ + { + "bbox": [ + 67, + 640, + 274, + 651 + ], + "type": "text", + "content": "In summary, our main contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 654, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 654, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 654, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 654, + 295, + 689 + ], + "type": "text", + "content": "- We introduce the UCM-VeID V2 dataset, which significantly enhances the original UCM-VeID with greater scale and diversity across multiple dimensions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 690, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 296, + 714 + ], + "type": "text", + "content": "- We propose a novel self-supervised paradigm to pre-train an encoder, replacing the ImageNet-trained ResNet50, to" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 426, + 155 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 426, + 155 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 426, + 155 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 426, + 155 + ], + "type": "image", + "image_path": "0263be20b404b1e2de1e5a9efece11042bdd6f5df8cc7c52a25946019ee56a66.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 156, + 416, + 163 + ], + "lines": [ + { + "bbox": [ + 332, + 156, + 416, + 163 + ], + "spans": [ + { + "bbox": [ + 332, + 156, + 416, + 163 + ], + "type": "text", + "content": "(a) Training based on UCM-VeID" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 426, + 70, + 531, + 155 + ], + "blocks": [ + { + "bbox": [ + 426, + 70, + 531, + 155 + ], + "lines": [ + { + "bbox": [ + 426, + 70, + 531, + 155 + ], + "spans": [ + { + "bbox": [ + 426, + 70, + 531, + 155 + ], + "type": "image", + "image_path": "74a9675ea7a5b6f8f2451ad9f3d8324cc16bb1c0c18b28049cbc42f7e0711420.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 156, + 527, + 163 + ], + "lines": [ + { + "bbox": [ + 435, + 156, + 527, + 163 + ], + "spans": [ + { + "bbox": [ + 435, + 156, + 527, + 163 + ], + "type": "text", + "content": "(b) Training based on UCM-VeID V2" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 172, + 553, + 260 + ], + "lines": [ + { + "bbox": [ + 313, + 172, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 313, + 172, + 553, + 260 + ], + "type": "text", + "content": "Figure 2. The feature distribution of pre-trained embeddings is visualized. Panels (a) and (b) show the embeddings in 2D space for models trained on the UCM-VeID and UCM-VeID V2 datasets, respectively, with dots and crosses representing different identities in RGB and IR modalities. The experiment uses the AGW baseline [37]. As indicated by the red circles in (a), the same-ID samples are more compact. In contrast, the same-ID samples are more spread out in (b), which ultimately degrades Re-ID performance." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 535, + 71, + 552, + 153 + ], + "blocks": [ + { + "bbox": [ + 535, + 71, + 552, + 153 + ], + "lines": [ + { + "bbox": [ + 535, + 71, + 552, + 153 + ], + "spans": [ + { + "bbox": [ + 535, + 71, + 552, + 153 + ], + "type": "image", + "image_path": "67328e746653e6e617c3de88a8a6e3da2f8881c52a006470abc67f82f6a0a519.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 324, + 281, + 420, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 281, + 420, + 293 + ], + "spans": [ + { + "bbox": [ + 324, + 281, + 420, + 293 + ], + "type": "text", + "content": "address the MBT issue." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 304, + 406, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 304, + 406, + 316 + ], + "spans": [ + { + "bbox": [ + 314, + 304, + 406, + 316 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 324, + 444, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 324, + 444, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 324, + 444, + 335 + ], + "type": "text", + "content": "2.1. Vehicle Re-ID Datasets" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 341, + 555, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 555, + 557 + ], + "type": "text", + "content": "Vehicle Re-ID research has flourished over the past few years, significantly benefited by the availability of vehicle Re-ID datasets. The most widely-employed datasets are primarily captured via fixed surveillance cameras, such as VehicleID [20], VeRI-776 [21], VERI-Wild [23] and Vehicle1M [10]. Compared to fixed surveillance, UAV-based cameras offer superior mobility, flexibility, and convenience, leading to the development of key UAV-based vehicle Re-ID datasets such as VRAI [31], UAV-VeID [29], VRU [24], and VeRi-UAV [19]. Furthermore, Liu et al. [22] have advanced the first RGB-IR vehicle Re-ID dataset, named UCM-VeID, to support the research of cross-modality vehicle Re-ID. However, the UCM-VeID dataset exhibits limitations in both size and diversity, highlighting an urgent need for the further expansion and enrichment of the dataset, ensuring it remains a potent benchmark for advancing the frontiers of vehicle Re-ID research in the face of evolving challenges and complexities." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 564, + 446, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 446, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 446, + 577 + ], + "type": "text", + "content": "2.2. Modality Bias Training" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "The MBT problem is essentially a problem of unbalanced training. Imbalanced model learning has received extensive attention in multi-modal learning [15, 32], and great progresses have been made in VT-ReID [16, 30]. To our best knowledge, there are two distinct approaches to tackle the MBT issue. The first approach is to control the training process, such as learning rate adjusting [15], gradient modulation [27], and ID-exclusive sampling strategy [16]. The second approach is to use modal balanced weights rather than ImageNet pre-trained ResNet50 avoiding MBT problems from the beginning. Wan et al. [30] devised a self-" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22287" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 550, + 135 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 550, + 135 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 550, + 135 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 550, + 135 + ], + "type": "table", + "html": "
DatasetVenueIdentitiesImagesModalitiesCamerasPlatformCamera ViewTarget StateColorTypeOrientation
VRAIICCV201913022137613RGB2UAVFlexibleMove97—
UAV-VeIDIJCV2021460141917RGB2UAVFlexibleMove——4
UCM-VeIDTMM202475329928RGB/IR2UAVFlexibleMove&Static1298
UCM-VeID V2ours3407165091RGB/IR5UAVFlexibleMove&Static1298
", + "image_path": "f88c127df84ed430d7ef49e12042e9c4a74b3bbafdab337289289aa866e0112c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 162, + 142, + 448, + 153 + ], + "lines": [ + { + "bbox": [ + 162, + 142, + 448, + 153 + ], + "spans": [ + { + "bbox": [ + 162, + 142, + 448, + 153 + ], + "type": "text", + "content": "Table 1. Comparisons among other vehicle Re-ID datasets and UCM-VeID V2" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 175, + 296, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 296, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 296, + 294 + ], + "type": "text", + "content": "supervised pre-training method to obtain a set of new unbiased weights, which achieves comparable performance to that of ImageNet-trained ResNet50. The second approach addresses the MBT problem from the outset by using pretrained weights that are modally balanced, instead of relying on subsequent adjustments. In contrast to the first approach, the second method more fundamentally resolves the modality bias, offering stronger generalization and more stable performance. Consequently, we follow the second approach to carry out the MBT problem." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 303, + 198, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 198, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 198, + 316 + ], + "type": "text", + "content": "2.3. Self-Supervised Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 320, + 296, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 320, + 296, + 524 + ], + "spans": [ + { + "bbox": [ + 55, + 320, + 296, + 524 + ], + "type": "text", + "content": "Self-supervised learning, as a subset of unsupervised learning, constructs pretext tasks to extract valuable insights from large-scale unlabeled data, thereby learning representations beneficial for downstream tasks. Common methods involve contextual pretext tasks to learn discriminative features, such as Word2Vec [25], Jigsaw [26], Rotation [8], Colorization [38], and MAE [14]. Another prominent approach, contrastive learning, builds discriminative representations by comparing instances within the same feature space, as seen in methods like MoCo [12] and BYOL [9], both of which perform well on downstream tasks. Nevertheless, above methods are not specifically tailored for the Re-ID task, particularly for VT-ReID task. In this paper, we leverage the consistency of different modalities to construct a patch-mixed image reconstruction self-supervised task aimed at extracting shared features, thereby mitigating the modality discrepancy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 535, + 216, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 535, + 216, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 535, + 216, + 548 + ], + "type": "text", + "content": "3. Benchmark: UCM-VeID V2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 555, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 296, + 616 + ], + "type": "text", + "content": "To further enhance the UAV cross-modal vehicle re-identification dataset, we have significantly expanded the first edition, UCM-VeID [22], by an order of magnitude in terms of UAV platforms, identities and so on. This section details the expanded dataset called UCM-VeID V2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 624, + 170, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 170, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 170, + 636 + ], + "type": "text", + "content": "3.1. Dataset Description" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "(1) Dataset Partition. UCM-VeID V2 comprises a total of 3407 identities with 165091 samples. The dataset is split into a training set and a test set in an " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "8:2" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " ratio, resulting in a training set with 2,693 identities and 131,451 samples, while the test set contains 674 identities and 33,640 samples. Furthermore, the query set and gallery set are derived" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 175, + 555, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 555, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 555, + 247 + ], + "type": "text", + "content": "from the test set. Specifically, one sample is drawn from each ID within the test set to constitute the query set, while the remaining samples collectively form the gallery set. It is worth noting that the extraction process is random, and crucially, every sample in the query set has a corresponding sample with the same ID in the gallery set." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 247, + 555, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 555, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 555, + 294 + ], + "type": "text", + "content": "(2) Dataset Statistics. To maximize the utility of the UCM-VeID V2 dataset, we perform a detailed statistical analysis of its key attributes, providing insights into the data patterns relevant to Re-ID tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 294, + 554, + 628 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 314, + 294, + 554, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 294, + 554, + 366 + ], + "spans": [ + { + "bbox": [ + 314, + 294, + 554, + 366 + ], + "type": "text", + "content": "- ID distribution of 5 cameras. Data is collected using 3 drone platforms equipped with 5 different cameras. " + }, + { + "bbox": [ + 314, + 294, + 554, + 366 + ], + "type": "inline_equation", + "content": "D300_{z}" + }, + { + "bbox": [ + 314, + 294, + 554, + 366 + ], + "type": "text", + "content": " camera captures fewer IDs compared to the other four cameras, as Fig. 3(a) shown. The " + }, + { + "bbox": [ + 314, + 294, + 554, + 366 + ], + "type": "inline_equation", + "content": "D300_{z}" + }, + { + "bbox": [ + 314, + 294, + 554, + 366 + ], + "type": "text", + "content": " camera captures high-resolution images, primarily focused on nighttime data, thereby enhancing the dataset's diversity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 366, + 554, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 366, + 554, + 436 + ], + "spans": [ + { + "bbox": [ + 315, + 366, + 554, + 436 + ], + "type": "text", + "content": "- Distribution of ID samples. Samples are collected from different camera perspectives, ensuring diversity and ample representation for each ID. As shown in Fig. 3(b), most IDs have 20-40 samples, providing comprehensive coverage of target's appearance across different conditions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 438, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 438, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 315, + 438, + 554, + 510 + ], + "type": "text", + "content": "- Distribution of Resolution. As shown in Fig. 3(c), resolutions vary significantly across cameras. For example, " + }, + { + "bbox": [ + 315, + 438, + 554, + 510 + ], + "type": "inline_equation", + "content": "D300_{i}r" + }, + { + "bbox": [ + 315, + 438, + 554, + 510 + ], + "type": "text", + "content": " samples cluster around 10,000-15,000 pixels, while DM3P samples are predominantly 45,000-50,000 pixels, highlighting a resolution disparity that presents a core challenge in UAV-based VT-ReID." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 510, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 510, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 315, + 510, + 554, + 593 + ], + "type": "text", + "content": "- Energy Distribution of IR Samples. The dataset spans all seasons and covers a 24-hour period, capturing both dynamic and static vehicles, leading to diverse infrared energy levels. Fig. 3(d) reveals that most infrared energy values are between 80-100, corresponding to moderate daylight. Extreme conditions, like intense heat or cold, are less common but contribute to the dataset's breadth." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 553, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 628 + ], + "type": "text", + "content": "(3) Dataset Attributes. Attributes annotation is an integral part of building reliable, high-performance data sets. In this dataset, we annotate the below attributes." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 315, + 630, + 554, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 630, + 554, + 701 + ], + "spans": [ + { + "bbox": [ + 315, + 630, + 554, + 701 + ], + "type": "text", + "content": "- Vehicle Type. Vehicles are classified into categories such as car, SUV, minibus, MPV, bus, mini truck, truck, and pickup. As shown in Fig. 3(e), cars and SUVs make up the majority, while trucks and pickups are less frequent. This distribution reflects the urban focus of our data collection, where cars and SUVs are more prevalent." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 701, + 553, + 713 + ], + "type": "text", + "content": "- Vehicle color. We categorize vehicle colors into thirteen" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22288" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 87, + 179, + 175 + ], + "blocks": [ + { + "bbox": [ + 72, + 72, + 173, + 80 + ], + "lines": [ + { + "bbox": [ + 72, + 72, + 173, + 80 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 173, + 80 + ], + "type": "text", + "content": "ID Distribution of Different Cameras" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 87, + 179, + 175 + ], + "lines": [ + { + "bbox": [ + 61, + 87, + 179, + 175 + ], + "spans": [ + { + "bbox": [ + 61, + 87, + 179, + 175 + ], + "type": "image", + "image_path": "a19c874acef395028b97b1d4c6f296b53e6c41657a75ee410e33daba391ad4d0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 117, + 178, + 128, + 187 + ], + "lines": [ + { + "bbox": [ + 117, + 178, + 128, + 187 + ], + "spans": [ + { + "bbox": [ + 117, + 178, + 128, + 187 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 186, + 85, + 301, + 175 + ], + "blocks": [ + { + "bbox": [ + 209, + 72, + 282, + 80 + ], + "lines": [ + { + "bbox": [ + 209, + 72, + 282, + 80 + ], + "spans": [ + { + "bbox": [ + 209, + 72, + 282, + 80 + ], + "type": "text", + "content": "Distribution of ID Samples" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 186, + 85, + 301, + 175 + ], + "lines": [ + { + "bbox": [ + 186, + 85, + 301, + 175 + ], + "spans": [ + { + "bbox": [ + 186, + 85, + 301, + 175 + ], + "type": "image", + "image_path": "fec54423b6a3d274c3dbb2103f4febe2a2c904b7d30e96f8a2bf167fd9ef8d10.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 178, + 252, + 187 + ], + "lines": [ + { + "bbox": [ + 241, + 178, + 252, + 187 + ], + "spans": [ + { + "bbox": [ + 241, + 178, + 252, + 187 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 304, + 80, + 428, + 175 + ], + "blocks": [ + { + "bbox": [ + 332, + 71, + 402, + 78 + ], + "lines": [ + { + "bbox": [ + 332, + 71, + 402, + 78 + ], + "spans": [ + { + "bbox": [ + 332, + 71, + 402, + 78 + ], + "type": "text", + "content": "Distribution of Resolution" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 80, + 428, + 175 + ], + "lines": [ + { + "bbox": [ + 304, + 80, + 428, + 175 + ], + "spans": [ + { + "bbox": [ + 304, + 80, + 428, + 175 + ], + "type": "image", + "image_path": "b7c36c07a56e51a86fe6656e404885347f0ceffa1bcb92af4e584b1f369f37eb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 366, + 178, + 378, + 187 + ], + "lines": [ + { + "bbox": [ + 366, + 178, + 378, + 187 + ], + "spans": [ + { + "bbox": [ + 366, + 178, + 378, + 187 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 429, + 83, + 548, + 175 + ], + "blocks": [ + { + "bbox": [ + 441, + 71, + 534, + 79 + ], + "lines": [ + { + "bbox": [ + 441, + 71, + 534, + 79 + ], + "spans": [ + { + "bbox": [ + 441, + 71, + 534, + 79 + ], + "type": "text", + "content": "Energy Distribution of IR Samples" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 429, + 83, + 548, + 175 + ], + "lines": [ + { + "bbox": [ + 429, + 83, + 548, + 175 + ], + "spans": [ + { + "bbox": [ + 429, + 83, + 548, + 175 + ], + "type": "image", + "image_path": "64f782825eaf9bef9199b5e3372266e79ed829407f64abf0e56b910c3e094097.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 487, + 178, + 498, + 187 + ], + "lines": [ + { + "bbox": [ + 487, + 178, + 498, + 187 + ], + "spans": [ + { + "bbox": [ + 487, + 178, + 498, + 187 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 58, + 202, + 178, + 289 + ], + "blocks": [ + { + "bbox": [ + 83, + 191, + 160, + 201 + ], + "lines": [ + { + "bbox": [ + 83, + 191, + 160, + 201 + ], + "spans": [ + { + "bbox": [ + 83, + 191, + 160, + 201 + ], + "type": "text", + "content": "Distribution of Vehicle Type" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 202, + 178, + 289 + ], + "lines": [ + { + "bbox": [ + 58, + 202, + 178, + 289 + ], + "spans": [ + { + "bbox": [ + 58, + 202, + 178, + 289 + ], + "type": "image", + "image_path": "50986752ee54355497badb93d78ca4013f9b8eab9a452f6f2128475caf0a78b9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 117, + 298, + 129, + 308 + ], + "lines": [ + { + "bbox": [ + 117, + 298, + 129, + 308 + ], + "spans": [ + { + "bbox": [ + 117, + 298, + 129, + 308 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 179, + 202, + 299, + 293 + ], + "blocks": [ + { + "bbox": [ + 205, + 191, + 283, + 200 + ], + "lines": [ + { + "bbox": [ + 205, + 191, + 283, + 200 + ], + "spans": [ + { + "bbox": [ + 205, + 191, + 283, + 200 + ], + "type": "text", + "content": "Distribution of Vehicle Color" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 179, + 202, + 299, + 293 + ], + "lines": [ + { + "bbox": [ + 179, + 202, + 299, + 293 + ], + "spans": [ + { + "bbox": [ + 179, + 202, + 299, + 293 + ], + "type": "image", + "image_path": "27e34b5eb58b0121bbaf46bb74036621e70a7cbdfb626ad9e6de631beade163d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 240, + 297, + 251, + 308 + ], + "lines": [ + { + "bbox": [ + 240, + 297, + 251, + 308 + ], + "spans": [ + { + "bbox": [ + 240, + 297, + 251, + 308 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 304, + 202, + 428, + 294 + ], + "blocks": [ + { + "bbox": [ + 324, + 192, + 414, + 199 + ], + "lines": [ + { + "bbox": [ + 324, + 192, + 414, + 199 + ], + "spans": [ + { + "bbox": [ + 324, + 192, + 414, + 199 + ], + "type": "text", + "content": "Distribution of Orientation Class" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 202, + 428, + 294 + ], + "lines": [ + { + "bbox": [ + 304, + 202, + 428, + 294 + ], + "spans": [ + { + "bbox": [ + 304, + 202, + 428, + 294 + ], + "type": "image", + "image_path": "97e128641af1f8169fae5af1768c0fc7e4b829b54a37555caeef39f3e8b71957.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 366, + 297, + 378, + 308 + ], + "lines": [ + { + "bbox": [ + 366, + 297, + 378, + 308 + ], + "spans": [ + { + "bbox": [ + 366, + 297, + 378, + 308 + ], + "type": "text", + "content": "(g)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 441, + 199, + 538, + 296 + ], + "blocks": [ + { + "bbox": [ + 442, + 190, + 536, + 198 + ], + "lines": [ + { + "bbox": [ + 442, + 190, + 536, + 198 + ], + "spans": [ + { + "bbox": [ + 442, + 190, + 536, + 198 + ], + "type": "text", + "content": "Orientation Classification Criteria" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 441, + 199, + 538, + 296 + ], + "lines": [ + { + "bbox": [ + 441, + 199, + 538, + 296 + ], + "spans": [ + { + "bbox": [ + 441, + 199, + 538, + 296 + ], + "type": "image", + "image_path": "ff207fa11f345721ae98eed1447981517ea1d712e0df69e5228509cf987ab27a.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 487, + 297, + 499, + 308 + ], + "lines": [ + { + "bbox": [ + 487, + 297, + 499, + 308 + ], + "spans": [ + { + "bbox": [ + 487, + 297, + 499, + 308 + ], + "type": "text", + "content": "(h)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 316, + 545, + 327 + ], + "lines": [ + { + "bbox": [ + 63, + 316, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 63, + 316, + 545, + 327 + ], + "type": "text", + "content": "Figure 3. The statistics and annotation attributes of UCM-VeID V2 dataset. Notation that, the 'k' in (c) means the unit of 1000 pixels." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 64, + 348, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 348, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 64, + 348, + 295, + 407 + ], + "type": "text", + "content": "classes, shown in Fig. 3(f), ordered by proportion as follows: white, black, gray, red, dark gray, cyan, brown, blue, golden, yellow, others, green, and orange. Despite some imbalance, this distribution mirrors real-world vehicle color trends." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 56, + 408, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 408, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 56, + 408, + 295, + 479 + ], + "type": "text", + "content": "- Vehicle Orientation. Based on [22] and Fig. 3(h), we define eight orientation categories. Fig. 3(g) shows that orientations 2, 4, 6, and 8 are most common, a result of the UAV's flexible viewing angles. This orientation imbalance highlights the dataset's realism and the challenges in UAV-based Re-ID tasks." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 491, + 232, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 232, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 232, + 503 + ], + "type": "text", + "content": "3.2. Comparison with Other Datasets" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 509, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 533 + ], + "type": "text", + "content": "In contrast to other vehicle Re-ID datasets, as detailed in Tab. 1, our dataset has several unique characteristics:" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 534, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 56, + 534, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 295, + 604 + ], + "type": "text", + "content": "1. Multi modal information. Compared with RGB-based datasets, such as VRAI [31], UAV-VeID [29], UCMVeID [22] and UCM-VeID V2 integrate both RGB and IR modalities, facilitating cross-modality vehicle Re-ID tasks, which holds promise as a viable solution to achieve round-the-clock vehicle Re-Identification." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 688 + ], + "type": "text", + "content": "2. Flexible viewpoint, scale and orientations. The UCMVeID V2 dataset, captured via a drone-mounted camera, offers varied perspectives, scales, and orientations that differ from fixed-camera datasets like RGBN300 [18] and MSVR310 [40]. This diversity substantially enriches the dataset, making it a valuable resource for robust vehicle Re-ID research." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "text", + "content": "3. Overall enhancement. Compared to UCM-VeID, the UCM-VeID V2 dataset offers a substantial increase in" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 325, + 348, + 555, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 348, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 325, + 348, + 555, + 433 + ], + "type": "text", + "content": "data volume and significant improvements in several dimensions, including additional cameras and platforms, extended capture times, a broader variety of weather conditions, and a wider range of shooting scenarios, as shown in Fig. 1. These enhancements make it a richer and more comprehensive dataset, supporting diverse and robust data-driven research and applications." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 445, + 484, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 484, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 484, + 459 + ], + "type": "text", + "content": "4. Proposed pre-training method" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 464, + 455, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 455, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 455, + 476 + ], + "type": "text", + "content": "4.1. Overall Model Structure." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 483, + 555, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 626 + ], + "type": "text", + "content": "The proposed pre-training VT-ReID method is illustrated in Fig. 4. Firstly, style learning and image generation techniques are used to create pairs of RGB and IR images. Then, we segment these images into patches and randomly swap patches between different modalities to create two new patch-mixed images. Finally, the patch-mixed images are fed into the encoder to complete Patch-Mixed Image Reconstruction (PMIR), Modality Discrimination Adversarial Learning (MDAL) and Modality-Augmented Contrasting Learning (MACL) tasks, where PMIR and MDAL are adopted to learn modality-invariant features, and MACL is utilized to enhance ID discriminability." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 635, + 491, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 491, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 491, + 648 + ], + "type": "text", + "content": "4.2. Patch-Mixed Unified Embedding" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Cross-modality data generation. Given the limited size of Re-ID dataset, we first train a CycleGAN [41] to generate cross-modality versions of ImageNet. Specifically, we use CycleGAN to learn style mapping between RGB and IR images from the UCM-VeID V2 dataset with a gener" + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22289" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 75, + 553, + 223 + ], + "blocks": [ + { + "bbox": [ + 62, + 75, + 553, + 223 + ], + "lines": [ + { + "bbox": [ + 62, + 75, + 553, + 223 + ], + "spans": [ + { + "bbox": [ + 62, + 75, + 553, + 223 + ], + "type": "image", + "image_path": "aada07c53c2ccd2652fbc21c54a0bfdb8be043485aaae892ad35a6544e049979.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 237, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 237, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 555, + 270 + ], + "type": "text", + "content": "Figure 4. Pre-training paradigm, which comprises preprocessing and three self-supervised tasks. Preprocessing involves obtaining paired RGB and IR images through style learning and image generation, followed by patches segmentation and swapping, PMIR and MDAL then focus on learning modality-invariant features, while MACL aims to enhance the ID discriminability." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 291, + 295, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 295, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 295, + 351 + ], + "type": "text", + "content": "ator " + }, + { + "bbox": [ + 55, + 291, + 295, + 351 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 291, + 295, + 351 + ], + "type": "text", + "content": ". To avoid data leakage, we use ImageNet dataset for pre-training. For an input RGB image " + }, + { + "bbox": [ + 55, + 291, + 295, + 351 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^{\\mathrm{RGB}}" + }, + { + "bbox": [ + 55, + 291, + 295, + 351 + ], + "type": "text", + "content": " in the ImageNet, we generate the corresponding IR image as " + }, + { + "bbox": [ + 55, + 291, + 295, + 351 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^{\\mathrm{IR}} = G(\\pmb{x}_i^{\\mathrm{RGB}})" + }, + { + "bbox": [ + 55, + 291, + 295, + 351 + ], + "type": "text", + "content": ". The resulting cross-modality ImageNet dataset is then " + }, + { + "bbox": [ + 55, + 291, + 295, + 351 + ], + "type": "inline_equation", + "content": "\\{(\\pmb{x}_i^{\\mathrm{RGB}},\\pmb{x}_i^{\\mathrm{IR}})\\}_{i = 1}^N" + }, + { + "bbox": [ + 55, + 291, + 295, + 351 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 291, + 295, + 351 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 291, + 295, + 351 + ], + "type": "text", + "content": " is the total number." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 351, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 296, + 531 + ], + "type": "text", + "content": "Random patch mixture. Our method aims to teach the backbone to fully leverage the cross-modality information for modality bias elimination. Therefore, for the data input aspect, we mixture the RGB and IR image pair in a patch swapping manner. Mathematically, given a RGB-IR image pair " + }, + { + "bbox": [ + 55, + 351, + 296, + 531 + ], + "type": "inline_equation", + "content": "(\\pmb{x}_i^{\\mathrm{RGB}},\\pmb{x}_i^{\\mathrm{IR}})" + }, + { + "bbox": [ + 55, + 351, + 296, + 531 + ], + "type": "text", + "content": ", we first establish a patch mask " + }, + { + "bbox": [ + 55, + 351, + 296, + 531 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 351, + 296, + 531 + ], + "type": "text", + "content": " with the same size as " + }, + { + "bbox": [ + 55, + 351, + 296, + 531 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^{\\mathrm{RGB}}(\\pmb{x}_i^{\\mathrm{IR}})" + }, + { + "bbox": [ + 55, + 351, + 296, + 531 + ], + "type": "text", + "content": " that consists of " + }, + { + "bbox": [ + 55, + 351, + 296, + 531 + ], + "type": "inline_equation", + "content": "K\\times K" + }, + { + "bbox": [ + 55, + 351, + 296, + 531 + ], + "type": "text", + "content": " regular non-overlapping patches. With a predefined mask ratio " + }, + { + "bbox": [ + 55, + 351, + 296, + 531 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 351, + 296, + 531 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 351, + 296, + 531 + ], + "type": "inline_equation", + "content": "\\lfloor \\eta \\times K\\times K\\rfloor" + }, + { + "bbox": [ + 55, + 351, + 296, + 531 + ], + "type": "text", + "content": " patches in " + }, + { + "bbox": [ + 55, + 351, + 296, + 531 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 351, + 296, + 531 + ], + "type": "text", + "content": " are randomly selected following a uniform distribution and set as all-zeros matrices. In this paper, we predefined a mask ratio " + }, + { + "bbox": [ + 55, + 351, + 296, + 531 + ], + "type": "inline_equation", + "content": "\\eta = 0.75" + }, + { + "bbox": [ + 55, + 351, + 296, + 531 + ], + "type": "text", + "content": ". A higher " + }, + { + "bbox": [ + 55, + 351, + 296, + 531 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 351, + 296, + 531 + ], + "type": "text", + "content": " leads to more patch exchanges, increases the difficulty of the reconstruction task, and enables the model to learn more shared features. Then, the process of the proposed random path mixture can be formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 535, + 294, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 535, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 93, + 535, + 294, + 567 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {\\boldsymbol {x}} _ {i} ^ {\\mathrm {R G B}} = \\boldsymbol {x} _ {i} ^ {\\mathrm {R G B}} \\odot \\boldsymbol {M} + \\boldsymbol {x} _ {i} ^ {\\mathrm {I R}} \\odot (1 - \\boldsymbol {M}) \\\\ \\hat {\\boldsymbol {x}} _ {i} ^ {\\mathrm {I R}} = \\boldsymbol {x} _ {i} ^ {\\mathrm {I R}} \\odot \\boldsymbol {M} + \\boldsymbol {x} _ {i} ^ {\\mathrm {R G B}} \\odot (1 - \\boldsymbol {M}), \\end{array} \\tag {1}", + "image_path": "2abe34b76581ec19bce78a2ecd0d9c2fbd0a3a2c2fe1180fa1f030c33734be44.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 594 + ], + "type": "text", + "content": "Then, based on the Eq. (1), the cross-modality patch-mixed unified embeddings are finally achieved as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 109, + 598, + 295, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 598, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 109, + 598, + 295, + 613 + ], + "type": "interline_equation", + "content": "\\boldsymbol {f} _ {i} ^ {\\mathrm {R G B}}, \\boldsymbol {f} _ {i} ^ {\\mathrm {I R}} = \\boldsymbol {F} \\left(\\hat {\\boldsymbol {x}} _ {i} ^ {\\mathrm {R G B}}\\right), \\boldsymbol {F} \\left(\\hat {\\boldsymbol {x}} _ {i} ^ {\\mathrm {I R}}\\right), \\tag {2}", + "image_path": "45006dd10ed41fa4c00cf3da1b0b41600493154c8122c02b8a5a60d341d78658.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 618, + 296, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 618, + 296, + 654 + ], + "type": "inline_equation", + "content": "F(\\cdot)" + }, + { + "bbox": [ + 55, + 618, + 296, + 654 + ], + "type": "text", + "content": " is the feature extractor and ResNet50 [11] is employed in this paper. " + }, + { + "bbox": [ + 55, + 618, + 296, + 654 + ], + "type": "inline_equation", + "content": "f_{i}^{\\mathrm{RGB}}" + }, + { + "bbox": [ + 55, + 618, + 296, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 618, + 296, + 654 + ], + "type": "inline_equation", + "content": "f_{i}^{\\mathrm{IR}}" + }, + { + "bbox": [ + 55, + 618, + 296, + 654 + ], + "type": "text", + "content": " are the latent embeddings of mixed images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 660, + 250, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 250, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 250, + 673 + ], + "type": "text", + "content": "4.3. Self-supervised Multi-task Learning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "With the proposed unified embedding module, we can acquire the latent features containing cross-modality information. However, these features can not be directly employed" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 291, + 553, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 553, + 327 + ], + "type": "text", + "content": "to the Re-ID task due to modality-bias information and insufficient discriminability [30, 34, 37]. To address above these issues, we devise three self-supervised tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 327, + 555, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 531 + ], + "type": "text", + "content": "Patch-Mixed Image Reconstruction. In this task, we utilize the mixed features, " + }, + { + "bbox": [ + 313, + 327, + 555, + 531 + ], + "type": "inline_equation", + "content": "f_{i}^{\\mathrm{RGB}}" + }, + { + "bbox": [ + 313, + 327, + 555, + 531 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 327, + 555, + 531 + ], + "type": "inline_equation", + "content": "f_{i}^{\\mathrm{IR}}" + }, + { + "bbox": [ + 313, + 327, + 555, + 531 + ], + "type": "text", + "content": ", to reconstruct the input images, " + }, + { + "bbox": [ + 313, + 327, + 555, + 531 + ], + "type": "inline_equation", + "content": "x_{i}^{\\mathrm{RGB}}" + }, + { + "bbox": [ + 313, + 327, + 555, + 531 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 327, + 555, + 531 + ], + "type": "inline_equation", + "content": "x_{i}^{\\mathrm{IR}}" + }, + { + "bbox": [ + 313, + 327, + 555, + 531 + ], + "type": "text", + "content": ", respectively. This approach is motivated by the inherent advantage of image reconstruction as a naturally unsupervised learning task, eliminating the need for additional labels and making it suitable for training large models. Moreover, for RGB image reconstruction, " + }, + { + "bbox": [ + 313, + 327, + 555, + 531 + ], + "type": "inline_equation", + "content": "f_{i}^{\\mathrm{RGB}}" + }, + { + "bbox": [ + 313, + 327, + 555, + 531 + ], + "type": "text", + "content": " contains minimal RGB information but abundant IR information, compelling the network to focus on modality-shared features. The limited RGB information acts as style information that supports the reconstruction of the RGB image. Similarly, for IR image reconstruction, " + }, + { + "bbox": [ + 313, + 327, + 555, + 531 + ], + "type": "inline_equation", + "content": "x_{i}^{\\mathrm{IR}}" + }, + { + "bbox": [ + 313, + 327, + 555, + 531 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 313, + 327, + 555, + 531 + ], + "type": "inline_equation", + "content": "f_{i}^{\\mathrm{IR}}" + }, + { + "bbox": [ + 313, + 327, + 555, + 531 + ], + "type": "text", + "content": ", where the feature contains mostly RGB information with only a small amount of IR data, thus also ensuring the model learns to emphasize cross-modality features effectively. According to above analysis, the process of patch-mixed image reconstruction can be formulated as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 388, + 539, + 553, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 539, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 388, + 539, + 553, + 569 + ], + "type": "interline_equation", + "content": "\\begin{array}{r l} \\bar {\\boldsymbol {x}} _ {i} ^ {\\mathrm {R G B}} & = \\boldsymbol {H} _ {\\mathrm {R G B}} \\left(\\boldsymbol {f} _ {i} ^ {\\mathrm {R G B}}\\right) \\\\ \\bar {\\boldsymbol {x}} _ {i} ^ {\\mathrm {I R}} & = \\boldsymbol {H} _ {\\mathrm {I R}} \\left(\\boldsymbol {f} _ {i} ^ {\\mathrm {I R}}\\right), \\end{array} \\tag {3}", + "image_path": "4824f1f6ca2498150c163e3d2b93a25710e6eee97a7f9f774567a00649740783.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 578, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 555, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 578, + 555, + 640 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{x}}_i^{\\mathrm{RGB}}" + }, + { + "bbox": [ + 313, + 578, + 555, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 578, + 555, + 640 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{x}}_i^{\\mathrm{IR}}" + }, + { + "bbox": [ + 313, + 578, + 555, + 640 + ], + "type": "text", + "content": " are the reconstructed RGB and IR images. " + }, + { + "bbox": [ + 313, + 578, + 555, + 640 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{RGB}}(\\cdot)" + }, + { + "bbox": [ + 313, + 578, + 555, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 578, + 555, + 640 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{IR}}(\\cdot)" + }, + { + "bbox": [ + 313, + 578, + 555, + 640 + ], + "type": "text", + "content": " are networks for RGB and IR image reconstruction respectively. Finally, a pixel-wise image reconstruction loss with " + }, + { + "bbox": [ + 313, + 578, + 555, + 640 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 313, + 578, + 555, + 640 + ], + "type": "text", + "content": "-norm is employed to guide the reconstruction process as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 322, + 648, + 553, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 648, + 553, + 681 + ], + "spans": [ + { + "bbox": [ + 322, + 648, + 553, + 681 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r e c}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\| \\bar {\\boldsymbol {x}} _ {i} ^ {\\mathrm {R G B}} - \\boldsymbol {x} _ {i} ^ {\\mathrm {R G B}} \\| _ {2} ^ {2} + \\| \\bar {\\boldsymbol {x}} _ {i} ^ {\\mathrm {I R}} - \\boldsymbol {x} _ {i} ^ {\\mathrm {I R}} \\| _ {2} ^ {2}\\right). \\tag {4}", + "image_path": "cbf81c06ed3dbdf01f99983ae398ba7e6222bcbaecfbb0df953395350070496f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Modality Discrimination Adversarial Learning. In this task, we introduce adversarial training to encourage" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22290" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "the learned features to be modality-invariant. Specifically, we employ a discriminator " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "D(\\cdot)" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " to distinguish between the mixed features " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "f_{i}^{\\mathrm{RGB}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "f_{i}^{\\mathrm{IR}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": ". Initially, we fix the feature extractor " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " and train the discriminator " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " to correctly classify the mixed features based on their corresponding modality labels. This step ensures that " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " can accurately distinguish between features from different modalities. The objective for the discriminator " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " can be formulated as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 177, + 295, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 177, + 295, + 192 + ], + "spans": [ + { + "bbox": [ + 61, + 177, + 295, + 192 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D}} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left[ y _ {i} \\log \\boldsymbol {D} (\\boldsymbol {f} _ {i}) + (1 - y _ {i}) \\log (1 - \\boldsymbol {D} (\\boldsymbol {f} _ {i})) \\right], \\tag {5}", + "image_path": "b2b7ca1bdbae1441e4b4c0a4f6535b7f18a58594419625d5756d27395436081d.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 200, + 296, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 200, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 200, + 296, + 308 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 200, + 296, + 308 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 55, + 200, + 296, + 308 + ], + "type": "text", + "content": " is the modality label, " + }, + { + "bbox": [ + 55, + 200, + 296, + 308 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 55, + 200, + 296, + 308 + ], + "type": "text", + "content": " represents either " + }, + { + "bbox": [ + 55, + 200, + 296, + 308 + ], + "type": "inline_equation", + "content": "f_{i}^{\\mathrm{RGB}}" + }, + { + "bbox": [ + 55, + 200, + 296, + 308 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 200, + 296, + 308 + ], + "type": "inline_equation", + "content": "f_{i}^{\\mathrm{IR}}" + }, + { + "bbox": [ + 55, + 200, + 296, + 308 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 200, + 296, + 308 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 200, + 296, + 308 + ], + "type": "text", + "content": " is the batch size. In the second step, we fix the discriminator " + }, + { + "bbox": [ + 55, + 200, + 296, + 308 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 200, + 296, + 308 + ], + "type": "text", + "content": " and update the feature extractor " + }, + { + "bbox": [ + 55, + 200, + 296, + 308 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 200, + 296, + 308 + ], + "type": "text", + "content": " to generate mixed features that the discriminator cannot distinguish. This step forces the feature extractor to produce modality-invariant features, effectively minimizing the modality bias in the learned features. The objective for the feature extractor " + }, + { + "bbox": [ + 55, + 200, + 296, + 308 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 200, + 296, + 308 + ], + "type": "text", + "content": " during this phase is to maximize the classification error of " + }, + { + "bbox": [ + 55, + 200, + 296, + 308 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 200, + 296, + 308 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 317, + 295, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 317, + 295, + 332 + ], + "spans": [ + { + "bbox": [ + 61, + 317, + 295, + 332 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {F}} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left[ \\left(1 - y _ {i}\\right) \\log \\boldsymbol {D} (\\boldsymbol {f} _ {i}) + y _ {i} \\log \\left(1 - \\boldsymbol {D} (\\boldsymbol {f} _ {i})\\right) \\right]. \\tag {6}", + "image_path": "494365ab3c87ec233416c9bf225a93a1f69715da2bb069dc1019c4037441a7cc.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 341, + 295, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 295, + 389 + ], + "type": "text", + "content": "The adversarial training process iterates between these two steps until the model converges, achieving a balance where the discriminator " + }, + { + "bbox": [ + 55, + 341, + 295, + 389 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 341, + 295, + 389 + ], + "type": "text", + "content": " cannot reliably distinguish the modality of the features generated by " + }, + { + "bbox": [ + 55, + 341, + 295, + 389 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 341, + 295, + 389 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 389, + 296, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 520 + ], + "type": "text", + "content": "Modality-Augmented Contrasting Cluster. In this task, we focus on enhancing the discriminative power of the learned features by employing a contrastive learning approach. Unlike traditional methods that rely on data augmentation to generate negative samples, our approach does not introduce additional data augmentation but instead utilizes the inherent modality-specific features for contrastive learning. We treat " + }, + { + "bbox": [ + 55, + 389, + 296, + 520 + ], + "type": "inline_equation", + "content": "f_{i}^{\\mathrm{RGB}}" + }, + { + "bbox": [ + 55, + 389, + 296, + 520 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 389, + 296, + 520 + ], + "type": "inline_equation", + "content": "f_{i}^{\\mathrm{IR}}" + }, + { + "bbox": [ + 55, + 389, + 296, + 520 + ], + "type": "text", + "content": " as positive pairs within the contrastive learning framework. The contrastive loss function aims to maximize the similarity between the mixed features of different modalities, which can be formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 529, + 295, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 529, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 62, + 529, + 295, + 563 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n t r a s t}} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {f} _ {i} ^ {\\mathrm {R G B}} , \\boldsymbol {f} _ {i} ^ {\\mathrm {I R}}\\right)\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\operatorname {s i m} \\left(\\boldsymbol {f} _ {i} ^ {\\mathrm {R G B}}, \\boldsymbol {f} _ {j} ^ {\\mathrm {I R}}\\right)\\right)}, \\tag {7}", + "image_path": "b360db5c121af02bd84207df72c7457d6d1bf6a48527a5d4d6a5e0bbec99fb25.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 571, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 295, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 571, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 55, + 571, + 295, + 666 + ], + "type": "text", + "content": " denotes a similarity measure (such as cosine similarity) between features. The loss function encourages the feature extractor to learn representations that pull together positive pairs " + }, + { + "bbox": [ + 55, + 571, + 295, + 666 + ], + "type": "inline_equation", + "content": "(f_{i}^{\\mathrm{RGB}}" + }, + { + "bbox": [ + 55, + 571, + 295, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 571, + 295, + 666 + ], + "type": "inline_equation", + "content": "f_{i}^{\\mathrm{IR}})" + }, + { + "bbox": [ + 55, + 571, + 295, + 666 + ], + "type": "text", + "content": " while pushing apart other features. By iteratively optimizing this contrastive loss, the feature extractor " + }, + { + "bbox": [ + 55, + 571, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\pmb{F}" + }, + { + "bbox": [ + 55, + 571, + 295, + 666 + ], + "type": "text", + "content": " is guided to produce more discriminative features, thereby improving the overall performance of the cross-modal image re-identification task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 667, + 295, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 667, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 55, + 667, + 295, + 691 + ], + "type": "text", + "content": "In summary, the overall loss function for the entire training process can be expressed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 102, + 701, + 295, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 102, + 701, + 295, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l l}} = \\mathcal {L} _ {\\text {r e c}} + \\left(\\mathcal {L} _ {\\mathrm {D}} + \\mathcal {L} _ {\\mathrm {F}}\\right) + \\mathcal {L} _ {\\text {c o n t r a s t}}. \\tag {8}", + "image_path": "31d47818f743b8f69134c6ec9332e3d127620d31d0d8469c89e4a12c36085697.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 72, + 395, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 395, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 395, + 84 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 91, + 533, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 533, + 104 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 533, + 104 + ], + "type": "text", + "content": "5.1. Implementation and Evaluation Protocols" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 109, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 553, + 216 + ], + "type": "text", + "content": "Pre-training Stage. For pre-training, we implement CPSL using PyTorch on an NVIDIA RTX A100. Optimization is conducted with an SGD optimizer (momentum of 0.9, weight decay of 0.0005), an initial learning rate of 0.1 with linear warm-up, and a cosine decay schedule without restarts. Data augmentation includes random grayscale, horizontal flipping, Gaussian filtering, and interpolation sampling. Images are resized to " + }, + { + "bbox": [ + 313, + 109, + 553, + 216 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 109, + 553, + 216 + ], + "type": "text", + "content": ", and the mask size " + }, + { + "bbox": [ + 313, + 109, + 553, + 216 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 109, + 553, + 216 + ], + "type": "text", + "content": " is set to 4." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 217, + 554, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 554, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 554, + 337 + ], + "type": "text", + "content": "Fine-Tuning Stage. For fine-tuning, we implement VT-ReID models using PyTorch on an NVIDIA GeForce RTX 3090 GPU. Images are resized to " + }, + { + "bbox": [ + 313, + 217, + 554, + 337 + ], + "type": "inline_equation", + "content": "180 \\times 256" + }, + { + "bbox": [ + 313, + 217, + 554, + 337 + ], + "type": "text", + "content": ". Optimization uses an SGD optimizer (momentum of 0.9, weight decay of 0.0005) with an initial learning rate of 0.1 and linear warm-up over 80 epochs. The model is initialized with CPSL checkpoints while other hyper-parameters and data augmentations remain unchanged, as this setup is not specifically tuned for CPSL. Further optimization may improve results, which we leave for future work." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 338, + 554, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 554, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 554, + 410 + ], + "type": "text", + "content": "Evaluation Protocols. Following [34], we use the cumulative matching characteristic (CMC) curve and mean average precision (mAP) for evaluation. CMC assesses rank-k accuracy, with rank-1, rank-10, and rank-20 scores reported in our experiments. mAP reflects the overall retrieval performance across all query samples." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 418, + 512, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 512, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 512, + 431 + ], + "type": "text", + "content": "5.2. Comparison with pre-trained weights" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 437, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 544 + ], + "type": "text", + "content": "To validate the effectiveness of our proposed pre-training method, we employ three different backbone initialization strategies during fine-tuning: random initialization (Random Init.), initialization with ImageNet pre-trained weights (ImageNet-1K), and initialization using the proposed pretraining weights (CPSL). To enhance experimental rigor, we perform supervised fine-tuning with four state-of-the-art VT-ReID methods: AGW [37], DDAG [36], MMD-ReID [17], and DEEN [39]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 544, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 616 + ], + "type": "text", + "content": "From Tab. 2, we can observe that the VT-ReID models trained from scratch generally under-performance, with all methods failing to converge. The primary cause of the gradient explosion is attributed to the triplet loss, suggesting that randomly initialized models struggle to establish semantically meaningful Euclidean feature spaces." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 617, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 700 + ], + "type": "text", + "content": "Our method CPSL achieves superior results across all testing and query modes. In contrast to ImageNet-1K pretraining, we do not rely on additional training data or incorporate any supervised training processes. The superior performance is primarily attributed to the design of our self-supervised tasks, which account for modality bias and eliminate cross-modality differences." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "type": "text", + "content": "It can also be observed that CPSL consistently improves" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22291" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 566, + 266 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 566, + 266 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 566, + 266 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 566, + 266 + ], + "type": "table", + "html": "
MethodVenuePre-TrainSingle-shotMulti-shot
IR2RGBRGB2IRIR2RGBRGB2IR
rank1rank10mAPrank1rank10mAPrank1rank10mAPrank1rank10mAP
AGW [37]TPAMI 2021Random Init.4.3518.759.384.8919.79.297.2227.264.826.6726.175.53
ImageNet-1K15.7545.9825.7416.1747.5826.5420.158.1216.2120.4558.1618.3
CPSL23.4753.0633.6925.3658.5436.5731.9167.5424.1332.6468.0724.01
DDAG [36]ECCV 2020Random Init.5.2222.8111.485.0622.6911.228.5532.725.847.0630.056.25
ImageNet-1K20.4648.7830.1219.4649.1529.2127.5864.7520.2826.6965.5422.87
CPSL22.2350.8631.9923.4551.3332.9431.9667.9421.731.1666.5824.51
MMD-ReID [17]BMVC 2021Random Init.4.3415.279.864.9218.539.935.9919.3710.785.6920.3410.42
ImageNet-1K18.9348.6428.9919.2749.1429.0822.0455.8719.6422.9659.2731.41
CPSL21.9851.8532.3922.5253.3831.9928.3462.0922.129.0462.8521.8
DEEN [39]CVPR 2023Random Init.7.2931.7415.246.1127.0613.297.8837.447.487.3932.817.34
ImageNet-1K18.2653.0130.3220.4656.2332.0923.8364.1422.0622.6566.6121.43
CPSL22.5759.8531.0923.759.5633.3525.6267.5223.4424.5368.6323.97
", + "image_path": "90035aec1a424133b4b2479dc66f5859239216b91ddbffa0e96675f88424aa99.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 63, + 267, + 252, + 277 + ], + "lines": [ + { + "bbox": [ + 63, + 267, + 252, + 277 + ], + "spans": [ + { + "bbox": [ + 63, + 267, + 252, + 277 + ], + "type": "text", + "content": "* The bold entries represent the best performance in table." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 329, + 296, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 296, + 401 + ], + "type": "text", + "content": "performance across various VT-ReID methods, which indicates that the weights generated by CPSL exhibit strong robustness and adaptability. This suggests that the CPSL method is effective in enhancing the generalization capabilities of the models, allowing them to perform well under different conditions and with diverse Re-ID techniques." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 408, + 263, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 263, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 263, + 422 + ], + "type": "text", + "content": "5.3. Comparisons With Other SSL Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 426, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 521 + ], + "type": "text", + "content": "We compare CPSL with multiple ResNet50 pre-trained models on the UCM-VeID V2. The methods involved can be categorized into four groups: (1) method based on VT-ReID fine-tuning (MMGL [30]), (2) method based pretext task design (Jigsaw [26], Rotation [8]), (3) method based self-supervised learning (SSL) pre-training (InstDisc [35], SimCLR [3], SwAV [2], BYOL [9], SimSiam [4], MoCoV2 [5], DenseCL [33], Relative-Loc[7])." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 605 + ], + "type": "text", + "content": "As shown in Tab. 3, MMGL performs suboptimally. This is partly because MMGL is tailored for pedestrian targets, using a top-down human segmentation approach that struggles with vehicles' variable orientations. Additionally, the released MMGL weights were fine-tuned on pedestrian VT-ReID datasets, embedding pedestrian-specific features that do not generalize well to vehicle targets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "SSL methods share a similar goal with our approach by aiming to endow the encoder with discriminative capabilities. However, methods such as SwAV [2] and BYOL [9] perform poorly, as indicated by the stagnant triplet loss during fine-tuning, suggesting these methods fail to establish a meaningful Euclidean space from the outset. Although some SSL methods outperform those based on pretext task design, they still underperform compared to CPSL. This is because they focus solely on target discriminability through" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 327, + 552, + 522 + ], + "blocks": [ + { + "bbox": [ + 55, + 285, + 555, + 308 + ], + "lines": [ + { + "bbox": [ + 55, + 285, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 555, + 308 + ], + "type": "text", + "content": "Table 2. Comparisons between the proposed CPSL and some state-of-the-art methods on UCM-VeID V2 datasets. The rank-k (%) and mAP (%) are demonstrated." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 327, + 552, + 522 + ], + "lines": [ + { + "bbox": [ + 317, + 327, + 552, + 522 + ], + "spans": [ + { + "bbox": [ + 317, + 327, + 552, + 522 + ], + "type": "table", + "html": "
MethodRank1Rank10Rank20mAP
method based on VT-ReID fine-tuning:
MMGL [30]21.8256.5668.0316.14
method based pretext task design:
Jigsaw [26]21.2456.7767.2515.66
RotationNet [8]26.6263.0372.7719.32
method based on SSL pre-training:
InstDisc [35]24.0962.0773.1620.04
SimCLR [3]12.4348.5762.1511.17
SwAV [2]6.5728.8940.225.17
BYOL [9]6.8925.5136.875.17
SimSiam [4]27.2465.7473.1322.36
MoCo V2 [5]28.8967.4674.9822.44
DenseCL [33]27.8964.9975.1222.41
Relative-Loc[7]18.9155.7968.0615.53
ours32.6468.0776.3224.01
", + "image_path": "a44db2b428e177baf2971c10ae46443b6cf59d0e5518781d1417330cb4108b53.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 530, + 555, + 553 + ], + "lines": [ + { + "bbox": [ + 313, + 530, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 555, + 553 + ], + "type": "text", + "content": "Table 3. Comparison with other SSL methods. The rank-k (%) and mAP (%) are demonstrated." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 576, + 555, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 555, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 555, + 625 + ], + "type": "text", + "content": "contrastive learning without addressing modality differences. In contrast, our PMIR and MDAL modules effectively mitigate cross-modality differences, enabling shared feature learning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 635, + 408, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 408, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 408, + 647 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "To assess the effectiveness of each component in our method, we conducted ablation studies on the proposed pretraining method, removing specific modules while keeping all other experimental settings constant. Our method includes three key modules: PMIR (R), MDAL (D), and" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22292" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 296, + 173 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 296, + 173 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 296, + 173 + ], + "type": "table", + "html": "
MethodSingle-shotMulti-shot
rank1rank10mAPrank1rank10mAP
Random Init.4.8919.79.296.6726.175.53
ImageNet-1K16.1747.5826.5420.4558.1618.3
R7.6426.9312.219.3342.7210.59
R+D14.6349.5722.4618.8854.9518.32
R+C22.4955.4830.3428.9765.1621.96
R+D+C25.3658.5436.5732.6468.0724.01
", + "image_path": "e4f90f269178d1b0c81e37b87b9bdd8f4c427c03e03919d6c462775e11ed28b0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 224, + 295, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 224, + 295, + 261 + ], + "spans": [ + { + "bbox": [ + 54, + 224, + 295, + 261 + ], + "type": "text", + "content": "MACL (C). Here, R and D target modality alignment, while C enhances discriminative representation. The ablation study results are shown in Tab. 4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 262, + 296, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 296, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 296, + 333 + ], + "type": "text", + "content": "Effectiveness of PMIR. PMIR aims to create a shared latent representation space for RGB and IR modalities via reconstruction tasks. As shown in Tab. 4, compare to the results of Random Init., RANK1 and mAP improve by " + }, + { + "bbox": [ + 55, + 262, + 296, + 333 + ], + "type": "inline_equation", + "content": "2.57\\%" + }, + { + "bbox": [ + 55, + 262, + 296, + 333 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 262, + 296, + 333 + ], + "type": "inline_equation", + "content": "2.92\\%" + }, + { + "bbox": [ + 55, + 262, + 296, + 333 + ], + "type": "text", + "content": " under the single-shot query mode, indicating PMIR's effectiveness in reducing modality discrepancies." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 334, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 296, + 392 + ], + "type": "text", + "content": "Effectiveness of MDAL. MDAL uses modality adversarial learning to further align modalities. Tab. 4 shows a RANK1 and mAP increase of " + }, + { + "bbox": [ + 55, + 334, + 296, + 392 + ], + "type": "inline_equation", + "content": "6.99\\%" + }, + { + "bbox": [ + 55, + 334, + 296, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 334, + 296, + 392 + ], + "type": "inline_equation", + "content": "10.25\\%" + }, + { + "bbox": [ + 55, + 334, + 296, + 392 + ], + "type": "text", + "content": " under single-shot query mode, demonstrating that MDAL helps guide the encoder toward shared feature learning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 393, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 393, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 296, + 464 + ], + "type": "text", + "content": "Effectiveness of MACL. MACL employs modality contrastive learning to enhance feature clustering and discriminative power. As evidenced in Tab. 4, RANK1 and mAP improve by " + }, + { + "bbox": [ + 55, + 393, + 296, + 464 + ], + "type": "inline_equation", + "content": "7.86\\%" + }, + { + "bbox": [ + 55, + 393, + 296, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 393, + 296, + 464 + ], + "type": "inline_equation", + "content": "7.88\\%" + }, + { + "bbox": [ + 55, + 393, + 296, + 464 + ], + "type": "text", + "content": " in single-shot query mode, confirming MACL's success in learning discriminative features for VT-ReID." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 56, + 473, + 296, + 522 + ], + "blocks": [ + { + "bbox": [ + 55, + 181, + 296, + 205 + ], + "lines": [ + { + "bbox": [ + 55, + 181, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 181, + 296, + 205 + ], + "type": "text", + "content": "Table 4. The quantitative results of ablation studies. The rank-k " + }, + { + "bbox": [ + 55, + 181, + 296, + 205 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 181, + 296, + 205 + ], + "type": "text", + "content": " and mAP " + }, + { + "bbox": [ + 55, + 181, + 296, + 205 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 181, + 296, + 205 + ], + "type": "text", + "content": " are demonstrated." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 473, + 296, + 522 + ], + "lines": [ + { + "bbox": [ + 56, + 473, + 296, + 522 + ], + "spans": [ + { + "bbox": [ + 56, + 473, + 296, + 522 + ], + "type": "table", + "html": "
Pn2 × 23 × 34 × 45 × 5
rank124.6627.3232.6429.05
mAP16.3820.9124.0122.44
", + "image_path": "19e55203ceae4a77f105f06993cda464c4fdf56b19efbc01668663f03c8ca43b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 529, + 295, + 552 + ], + "lines": [ + { + "bbox": [ + 55, + 529, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 295, + 552 + ], + "type": "text", + "content": "Table 5. Different partition stripe numbers " + }, + { + "bbox": [ + 55, + 529, + 295, + 552 + ], + "type": "inline_equation", + "content": "{P}_{n}\\left( {N \\times N}\\right)" + }, + { + "bbox": [ + 55, + 529, + 295, + 552 + ], + "type": "text", + "content": " . The rank-k (%) and mAP (%) are demonstrated." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 575, + 171, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 171, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 171, + 589 + ], + "type": "text", + "content": "5.5. Parameter Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "content": "The partition stripe number " + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "P_{n}" + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "content": " directly determines the size of the image blocks to be swapped and is an important hyperparameter of our method. To verify the impact of this hyperparameter, we conducted experiments with different values for the partition stripe number, with the relevant results presented in Tab. 5. From the table, it can be observed that as the partition stripe number increases, the model's performance initially improves before declining. This is because, as the partition stripe number increases, the semantic content within each image block transitions from being more" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 425, + 153 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 425, + 153 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 425, + 153 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 425, + 153 + ], + "type": "image", + "image_path": "6f3023b35357547ed5b2e3343fcb0c70f381171787f24a24e2e9786da093930e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 155, + 405, + 163 + ], + "lines": [ + { + "bbox": [ + 339, + 155, + 405, + 163 + ], + "spans": [ + { + "bbox": [ + 339, + 155, + 405, + 163 + ], + "type": "text", + "content": "(a) ImageNet Pre-Training" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 425, + 70, + 531, + 153 + ], + "blocks": [ + { + "bbox": [ + 425, + 70, + 531, + 153 + ], + "lines": [ + { + "bbox": [ + 425, + 70, + 531, + 153 + ], + "spans": [ + { + "bbox": [ + 425, + 70, + 531, + 153 + ], + "type": "image", + "image_path": "238a5c5513c100a4ca072473311bcbd5f265e17604c36a57f048aa1c3fcbfe4a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 155, + 508, + 163 + ], + "lines": [ + { + "bbox": [ + 449, + 155, + 508, + 163 + ], + "spans": [ + { + "bbox": [ + 449, + 155, + 508, + 163 + ], + "type": "text", + "content": "(b) PMSL Pre-Training" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 174, + 553, + 218 + ], + "lines": [ + { + "bbox": [ + 313, + 174, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 553, + 218 + ], + "type": "text", + "content": "Figure 5. Visualization of the pre-learned features. Different colors represent distinct identities, while circle and cross symbols denote RGB and IR modalities, respectively. The experiment is conducted on UCM-VeID V2 dataset with the AGW baseline [37]." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 535, + 71, + 552, + 152 + ], + "blocks": [ + { + "bbox": [ + 535, + 71, + 552, + 152 + ], + "lines": [ + { + "bbox": [ + 535, + 71, + 552, + 152 + ], + "spans": [ + { + "bbox": [ + 535, + 71, + 552, + 152 + ], + "type": "image", + "image_path": "23e415c255f9555e89e81104ef8eb30a508277ab37c7089ab7c19fcc34c15183.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 239, + 555, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 555, + 277 + ], + "type": "text", + "content": "discrete to becoming sparser. This indicates that in practical applications, appropriate adjustment of this parameter is necessary to achieve optimal performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 284, + 522, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 522, + 296 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 522, + 296 + ], + "type": "text", + "content": "5.6. Visualization of the feature distribution" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 301, + 554, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 554, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 554, + 373 + ], + "type": "text", + "content": "To better understand the proposed pre-training paradigm, we visualize the t-SNE distributions of features pre-learned using ImageNet-1K and the CPSL checkpoints on the UCM-VeID V2 dataset, utilizing AGW [37]. As shown in Fig. 5, we randomly select 15 identities to visualize all corresponding images." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 373, + 554, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 554, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 554, + 517 + ], + "type": "text", + "content": "Fig. 5 demonstrates that, the modal gap between PMSL features is smaller than that between ImageNet features (highlighted by the red circle). Specifically, PMSL features show less variation across modalities, indicating better consistency and robustness when processing RGB and IR images. In contrast, ImageNet-pretrained features exhibit significant modal variation, which can negatively impact performance in VT-ReID tasks. Thus, PMSL not only eliminates cross-modality discrepancy but also improves the model's discriminative ability for downstream tasks, providing stronger support for cross-modal vehicle re-identification." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 529, + 388, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 388, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 388, + 541 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 548, + 554, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 554, + 690 + ], + "type": "text", + "content": "In this paper, we introduce a novel dataset that expands existing ones in terms of scale, acquisition platforms, collection scenarios, time, and climate conditions. To address the MBT problem, we break it into two sub-problems: modality discrepancy and insufficient discriminability. We propose an end-to-end cross-modal pre-training framework to address these issues. To reduce modality discrepancies, the framework incorporates self-supervised tasks based on image block swapping and modality adversarial learning. For enhanced discriminability, we design an unsupervised clustering-guided multi-modal contrastive learning approach. Extensive experiments on the new dataset validate the effectiveness of our method." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 691, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 691, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 691, + 554, + 713 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by the National Natural Science Foundation of China (62201586)." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22293" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 95, + 115, + 108 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 115, + 108 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 115, + 108 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 116, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 116, + 295, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 116, + 295, + 159 + ], + "spans": [ + { + "bbox": [ + 61, + 116, + 295, + 159 + ], + "type": "text", + "content": "[1] Yan Bai, Yihang Lou, Feng Gao, Shiqi Wang, Yuwei Wu, and Ling-Yu Duan. Group-sensitive triplet embedding for vehicle reidentification. IEEE Trans. Multim., 20(9):2385-2399, 2018. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 161, + 295, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 161, + 295, + 215 + ], + "spans": [ + { + "bbox": [ + 61, + 161, + 295, + 215 + ], + "type": "text", + "content": "[2] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In Proc. Adv. Neural Inf. Process. Syst., pages 9912-9924, 2020. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 217, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 217, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 62, + 217, + 294, + 261 + ], + "type": "text", + "content": "[3] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In Proc. Int. Conf. Mach. Learn., pages 1597-1607, 2020. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 263, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 263, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 62, + 263, + 294, + 296 + ], + "type": "text", + "content": "[4] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 15750-15758, 2021. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 297, + 294, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 297, + 294, + 329 + ], + "spans": [ + { + "bbox": [ + 62, + 297, + 294, + 329 + ], + "type": "text", + "content": "[5] Xinlei Chen, Haoqi Fan, Ross B. Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. CoRR, abs/2003.04297, 2020. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 331, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 294, + 374 + ], + "type": "text", + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proc. IEEE Conf. Comput. Vis. Pattern Recognition., pages 248-255, 2009. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 376, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 294, + 420 + ], + "type": "text", + "content": "[7] Carl Doersch, Abhinav Gupta, and Alexei A. Efros. Unsupervised visual representation learning by context prediction. In Proc. IEEE Int. Conf. Comput. Vis., pages 1422-1430. IEEE Computer Society, 2015. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 422, + 294, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 294, + 454 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 294, + 454 + ], + "type": "text", + "content": "[8] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In Proc. Int. Conf. Learn. Represent., 2018. 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 456, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 456, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 456, + 294, + 533 + ], + "type": "text", + "content": "[9] Jean-Bastien Grill, Florian Strub, Florent Alché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Ávila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent - A new approach to self-supervised learning. In Proc. Adv. Neural Inf. Process. Syst., pages 21271-21284, 2020. 3, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 534, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 534, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 57, + 534, + 294, + 577 + ], + "type": "text", + "content": "[10] Haiyun Guo, Chaoyang Zhao, Zhiwei Liu, Jinqiao Wang, and Hanqing Lu. Learning coarse-to-fine structured feature embedding for vehicle re-identification. In Proc. AAAI Conf. Artif. Intell., pages 6853–6860, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 579, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 579, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 579, + 294, + 622 + ], + "type": "text", + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 770-778, 2016. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 667 + ], + "type": "text", + "content": "[12] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised visual representation learning. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 9726-9735, 2020. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "text", + "content": "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 15979-15988, 2022. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[14] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. Masked autoencoders are scalable vision learners. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 15979-15988, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "type": "text", + "content": "[15] Yu Huang, Junyang Lin, Chang Zhou, Hongxia Yang, and Longbo Huang. Modality competition: What makes joint training of multi-modal network fail in deep learning? (provably). In Proc. Int. Conf. Mach. Learn., pages 9226-9259, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "type": "text", + "content": "[16] Yan Huang, Qiang Wu, Jingsong Xu, Yi Zhong, Peng Zhang, and Zhaoxiang Zhang. Alleviating modality bias training for infrared-visible person re-identification. IEEE Trans. Multim., 24:1570-1582, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 553, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 258 + ], + "type": "text", + "content": "[17] Chaitra Jambigi, Ruchit Rawal, and Anirban Chakraborty. MMD-ReID: A simple but effective solution for visible-thermal person reid. In Proc. BMVC, pages 12026–12035, 2021. 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 261, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 293 + ], + "type": "text", + "content": "[18] H. Li, C. Li, X. Zhu, A. Zheng, and B. Luo. Multi-spectral vehicle re-identification: A challenge. In Proc. AAAI Conf. Artif. Intell., pages 11345–11353, 2020. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 294, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 553, + 348 + ], + "type": "text", + "content": "[19] Chunsheng Liu, Ye Song, Faliang Chang, Shuang Li, Ruimin Ke, and Yinhai Wang. Posture calibration based cross-view and hard-sensitive metric learning for UAV-based vehicle re-identification. IEEE Trans. Inf. Forensics Secur., 23:19246-19257, 2022. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "text", + "content": "[20] Hongye Liu, Yonghong Tian, Yaowei Wang, Lu Pang, and Tiejun Huang. Deep relative distance learning: Tell the difference between similar vehicles. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 2167-2175, 2016. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "type": "text", + "content": "[21] Xinchen Liu, Wu Liu, Tao Mei, and Huadong Ma. A deep learning-based approach to progressive vehicle re-identification for urban surveillance. In Proc. Eur. Conf. Comput. Vis., pages 869–884, 2016. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "type": "text", + "content": "[22] Xingyue Liu, Jiahao Qi, Chen Chen, Kangcheng Bin, and Ping Zhong. Beyond sharing weights in decoupling feature learning network for UAV rgb-infrared vehicle re-identification. IEEE Trans. Multim., 2024. 1, 2, 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 482, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 525 + ], + "type": "text", + "content": "[23] Yihang Lou, Yan Bai, Jun Liu, Shiqi Wang, and Lingyu Duan. VERI-Wild: A large dataset and a new method for vehicle re-identification in the wild. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3235-3243, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 526, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 557 + ], + "type": "text", + "content": "[24] Mingming Lu, Yongchuan Xu, and Haifeng Li. Vehicle re-identification based on UAV viewpoint: Dataset and method. Remote. Sens., 14(18):4603, 2022. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 559, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 592 + ], + "type": "text", + "content": "[25] Tomás Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. In Proc. Int. Conf. Learn. Represent., 2013. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 593, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 553, + 624 + ], + "type": "text", + "content": "[26] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In Proc. Eur. Conf. Comput. Vis., pages 69-84, 2016. 3, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 625, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 669 + ], + "type": "text", + "content": "[27] Xiaokang Peng, Yake Wei, Andong Deng, Dong Wang, and Di Hu. Balanced multimodal learning via on-the-fly gradient modulation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 8228-8237, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[28] Shangzhi Teng, Shiliang Zhang, Qingming Huang, and Nicu Sebe. Multi-view spatial attention embedding for vehicle re-identification. IEEE Trans. Circuits Syst. Video Technol., 31 (2):816-827, 2021. 1" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22294" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 699 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[29] Shangzhi Teng, Shiliang Zhang, Qingming Huang, and Nicu Sebe. Viewpoint and scale consistency reinforcement for UAV vehicle re-identification. Int. J. Comput. Vis., 129(3): 719-735, 2021. 1, 2, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "text", + "content": "[30] Lin Wan, Qianyan Jing, Zongyuan Sun, Chuang Zhang, Zhihang Li, and Yehansen Chen. Self-supervised modality-aware multiple granularity pre-training for rgb-infrared person re-identification. IEEE Trans. Inf. Forensics Secur., 18: 3044–3057, 2023. 2, 5, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "text", + "content": "[31] Peng Wang, Bingliang Jiao, Lu Yang, Yifei Yang, Shizhou Zhang, Wei Wei, and Yanning Zhang. Vehicle re-identification in aerial imagery: Dataset and approach. In Proc. IEEE Int. Conf. Comput. Vis., pages 460–469, 2019. 1, 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 229, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 229, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 294, + 272 + ], + "type": "text", + "content": "[32] Weiyao Wang, Du Tran, and Matt Feiszli. What makes training multi-modal classification networks hard? In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 12692-12702, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 294, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 328 + ], + "type": "text", + "content": "[33] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3024-3033. Computer Vision Foundation / IEEE, 2021. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 374 + ], + "type": "text", + "content": "[34] Ancong Wu, Wei-Shi Zheng, Hong-Xing Yu, Shaogang Gong, and Jianhuang Lai. RGB-infrared cross-modality person re-identification. In Proc. IEEE Int. Conf. Comput. Vis., pages 5390-5399, 2017. 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 419 + ], + "type": "text", + "content": "[35] Zhirong Wu, Yuanjun Xiong, Stella X. Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3733-3742, 2018. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 420, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 420, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 420, + 294, + 464 + ], + "type": "text", + "content": "[36] Mang Ye, Jianbing Shen, David J. Crandall, Ling Shao, and Jiebo Luo. Dynamic dual-attentive aggregation learning for visible-infrared person re-identification. Proc. Eur. Conf. Comput. Vision, 12362:229–247, 2020. 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 465, + 294, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 465, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 465, + 294, + 508 + ], + "type": "text", + "content": "[37] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven C. H. Hoi. Deep learning for person re-identification: A survey and outlook. IEEE Trans. Pattern Anal. Mach. Intell., 44:2872–2893, 2020. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 510, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 294, + 542 + ], + "type": "text", + "content": "[38] Richard Zhang, Phillip Isola, and Alexei A. Efros. Colorful image colorization. In Proc. Eur. Conf. Comput. Vis., pages 649-666, 2016. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 544, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 597 + ], + "type": "text", + "content": "[39] Yukang Zhang and Hanzi Wang. Diverse embedding expansion network and low-light cross-modality benchmark for visible-infrared person re-identification. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 2153–2162, 2023. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 599, + 294, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 599, + 294, + 654 + ], + "spans": [ + { + "bbox": [ + 56, + 599, + 294, + 654 + ], + "type": "text", + "content": "[40] Aihua Zheng, Xianpeng Zhu, Chenglong Li, Jin Tang, and Jixin Ma. Cross-directional consistency network with adaptive layer normalization for multi-spectral vehicle re-identification and a high-quality benchmark. Inf. Fusion, 100:101901, 2023. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 655, + 294, + 699 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 655, + 294, + 699 + ], + "spans": [ + { + "bbox": [ + 56, + 655, + 294, + 699 + ], + "type": "text", + "content": "[41] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proc. IEEE Int. Conf. Comput. Vis., pages 2242-2251, 2017. 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22295" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/UCOD-DPL_ Unsupervised Camouflaged Object Detection via Dynamic Pseudo-label Learning/dec1b26d-4dd9-4c06-a461-c03d0e31f356_content_list.json b/2025/UCOD-DPL_ Unsupervised Camouflaged Object Detection via Dynamic Pseudo-label Learning/dec1b26d-4dd9-4c06-a461-c03d0e31f356_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ac956173083e4a33dda4ec68fbb70f4c37e22d6e --- /dev/null +++ b/2025/UCOD-DPL_ Unsupervised Camouflaged Object Detection via Dynamic Pseudo-label Learning/dec1b26d-4dd9-4c06-a461-c03d0e31f356_content_list.json @@ -0,0 +1,1925 @@ +[ + { + "type": "text", + "text": "UCOD-DPL: Unsupervised Camouflaged Object Detection via Dynamic Pseudo-label Learning", + "text_level": 1, + "bbox": [ + 204, + 130, + 794, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weiqi Yan $^{1}$ , Lvhai Chen $^{1}$ , Huaijia Ko $^{1}$ , Shengchuan Zhang $^{1*}$ , Yan Zhang $^{1}$ , Liujuan Cao $^{1}$", + "bbox": [ + 207, + 204, + 787, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Key Laboratory of Multimedia Trusted Perception and Efficient Computing,", + "bbox": [ + 243, + 222, + 753, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ministry of Education of China, Xiamen University, 361005, P.R. China.", + "bbox": [ + 259, + 239, + 736, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "weiqi_yan@outlook.com lvhaichen2002@gmail.com zsc_2016@xmu.edu.cn", + "bbox": [ + 210, + 258, + 779, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unsupervised Camouflaged Object Detection (UCOD) has gained attention since it doesn't need to rely on extensive pixel-level labels. Existing UCOD methods typically generate pseudo-labels using fixed strategies and train $1 \\times 1$ convolutional layers as a simple decoder, leading to low performance compared to fully-supervised methods. We emphasize two drawbacks in these approaches: 1). The model is prone to fitting incorrect knowledge due to the pseudolabel containing substantial noise. 2). The simple decoder fails to capture and learn the semantic features of camouflaged objects, especially for small-sized objects, due to the low-resolution pseudo-labels and severe confusion between foreground and background pixels. To this end, we propose a UCOD method with a teacher-student framework via Dynamic Pseudo-label Learning called UCOD-DPL, which contains an Adaptive Pseudo-label Module (APM), a Dual-Branch Adversarial (DBA) decoder, and a Look-Twice mechanism. The APM module adaptively combines pseudo-labels generated by fixed strategies and the teacher model to prevent the model from overfitting incorrect knowledge while preserving the ability for self-correction; the DBA decoder takes adversarial learning of different segmentation objectives, guides the model to overcome the foreground-background confusion of camouflaged objects, and the Look-Twice mechanism mimics the human tendency to zoom in on camouflaged objects and performs secondary refinement on small-sized objects. Extensive experiments show that our method demonstrates outstanding performance, even surpassing some existing fully supervised methods. The code is available now1.", + "bbox": [ + 86, + 352, + 483, + 806 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5ac8efa5b68b91a90ea10acb677ac58490f6f67c51ff90cef9129cf4c0b41b87.jpg", + "image_caption": [ + "Figure 1. Comparison between our method and previous pseudolabel-based methods." + ], + "image_footnote": [], + "bbox": [ + 519, + 309, + 906, + 512 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 584, + 643, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"Camouflage\" originates from the natural behavior of objects blending into their surroundings using similar textures and colors to evade \"predators\"[4]. Camouflaged Object Detection (COD) is a challenging semantic segmentation task aimed at learning to segment objects that are visually hidden in their backgrounds, which has significant applications in several important fields [11, 50].", + "bbox": [ + 509, + 611, + 906, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although fully-supervised methods have achieved significant progress in the COD tasks [19, 38, 49, 59, 60], they heavily rely on large-scale human annotations for training. However, it is difficult to get pixel-level human labels for camouflaged objects due to their complex visual attributes [27]. Therefore, this reliance on annotations has sparked increasing interest in the task of unsupervised camouflage object detection (UCOD) among researchers. However, existing UCOD methods typically use pseudo-labels generated by fixed strategies to train a single $1 \\times 1$ convolutional layer [44, 63]. By investing in existing methods, we identified two main issues: 1). The pseudo-label con", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 107, + 875, + 233, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://github.com/Heartfirey/UCOD-DPL", + "bbox": [ + 109, + 887, + 418, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "30365", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c037aa66092f0a1f0d582e430edd9af8efac42e7e807c6839a8cba512a77e4a9.jpg", + "image_caption": [ + "Small-sized" + ], + "image_footnote": [], + "bbox": [ + 127, + 106, + 212, + 162 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0896fe4269343f73288e1be4ac1ae579ca458ea16653dff23238022e09984fdb.jpg", + "image_caption": [ + "Huge-sized" + ], + "image_footnote": [], + "bbox": [ + 218, + 106, + 302, + 162 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/abdac31dd9a5aea9a1280a5e1ae313a3bc47d84d78d00663bfee90522fccf1d0.jpg", + "image_caption": [ + "Multiple" + ], + "image_footnote": [], + "bbox": [ + 308, + 106, + 390, + 162 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/560531f721e64c89cd6d43697af21c7e59b8e59b058534ebda9cdd426ec4e159.jpg", + "image_caption": [ + "Uncertainty" + ], + "image_footnote": [], + "bbox": [ + 397, + 104, + 480, + 162 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ef753cf78e411b7724b52ed8462e5e3657c85f7a4a772376ef1b1957147ae3f7.jpg", + "image_caption": [ + "CocosDg" + ], + "image_footnote": [], + "bbox": [ + 96, + 223, + 120, + 279 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/37c5dc4ac4c08502e20ef8140b38c7803a50f816266c7ae11c483dfb75e4d2e9.jpg", + "image_caption": [ + "1" + ], + "image_footnote": [], + "bbox": [ + 127, + 164, + 212, + 220 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/549641dd98f921a8c5679018e0f06c1935125b49424aeee67aee7eb3b68cb56b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 127, + 222, + 212, + 279 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e5e2045dd2258de7408624fdca3dbac1ff29b150781030371efd01eba7bd0783.jpg", + "image_caption": [ + "Figure 2. Some examples of low-quality pseudo-labels generated using fixed strategies. Based on features extracted by DINOv2, we construct pseudo-labels using the Background-Seed from FOUND [44] and the MaskCut in UCOD-DA [44], in challenging scenarios (e.g. small-sized objects, huge-sized objects, multiple instances, and uncertain objects) for visualization." + ], + "image_footnote": [], + "bbox": [ + 127, + 280, + 212, + 337 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c28da8f540c7727fad1f22bf991d391ad860722a319d80e4aabd6c591d301dbc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 164, + 302, + 220 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e5b5861e533732c198bbd4e4bccf4f4aba2d9164fa40bf4fd0d03a1f5b2b3cda.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 220, + 302, + 279 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b126565ffeb8f8d829db44757da0ba6b8fda433bc0f85aa3a2fa20970e948cc0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 280, + 302, + 337 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/de55faa6b59c87878981166000c394e2fc556aa8ae7a3cb5c9697577b62548f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 308, + 164, + 390, + 220 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dfc2be28101e075b0e2865dc8e42e2cd9e30732770969d56bacbacf0b6e53e46.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 308, + 220, + 390, + 279 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/301cc198cda7d14e37dbf9e905d9c708da16277b655a67962eae80afb20006d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 308, + 280, + 390, + 337 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c22fe784f1af5024970c4e127c8a598aeb840a1a88013bed875d5ee63a0fde84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 397, + 164, + 480, + 220 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c647237a4427a22322048b2573886a66ea50a195eb5f28afeeaa8c7b08d7e86e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 397, + 220, + 480, + 279 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/26683d13d030418266cb11fa7b2ac1ead541e51029cfca245385d7da1ef905cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 397, + 280, + 480, + 337 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "structed by fixed strategies contains substantial noise, which makes the model prone to fitting incorrect knowledge. 2). These pseudo-labels have low resolution and severe confusion between foreground and background pixels, and simple $1 \\times 1$ convolution fails to capture and learn the semantic features of camouflaged objects, especially for small-sized objects. Consequently, their segmentation results fall far behind those of fully supervised methods, which limits their application in real-world scenarios.", + "bbox": [ + 89, + 460, + 482, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a brand-new model called UCOD-DPL. As shown in Fig. 1, we introduce a teacher-student framework consisting of an Adaptive Pseudo-label Mixing (APM) module, a Dual-Branch Adversarial (DBA) decoder, and a Look-Twice refinement strategy. As illustrated in Fig. 2, pseudo-labels constructed by fixed strategies are generally low-quality and contain substantial noise. To prevent the model from fitting uncorrectable erroneous knowledge in the pseudo-labels, we design the APM module with a discriminator and a dynamic scoring mechanism to mix the pseudo-labels from the fixed strategy and the teacher branch adaptively. To help the model overcome foreground-background confusion, we propose a DBA decoder inspired by predator-prey dynamics. This decoder uses different segmentation targets (foreground and background) along with an orthogonal loss of features to deeply mine and learn distinctive features of the foreground and background. Additionally, inspired by [37, 38], we mimic the zooming-in behavior humans use when examining camouflaged images by performing secondary refinement on", + "bbox": [ + 89, + 599, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "small-sized camouflaged objects to enhance the model's segmentation accuracy when dealing with small targets.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the experiment, we train our model on the commonly used COD training set and evaluate it on multiple test sets. The results demonstrate that our method achieves outstanding performance, even comparable to some fully supervised COD methods when using DINOv2 as the backbone.", + "bbox": [ + 511, + 123, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows:", + "bbox": [ + 532, + 199, + 870, + 213 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose a novel UCOD model called UCOD-DPL, and extensive experiments demonstrate that our method exhibits superior performance, reaching the performance level of some fully supervised algorithms.", + "2. We introduce a teacher-student framework with an APM module. By using the APM, we ensure that the model can learn knowledge from pseudo-labels while preserving the ability to identify and correct the erroneous.", + "3. To address the issues of low-resolution pseudo-labels and foreground-background confusion, we design a DBA decoder that uses a foreground-background adversarial task, enabling the model to learn the camouflaged features of the foreground adaptively.", + "4. To enhance the model's ability to segment small-sized objects, we design a Look-Twice refinement mechanism. This mechanism mimics human behavior by performing a secondary segmentation on smaller camouflaged objects, thereby refining the segmentation results." + ], + "bbox": [ + 513, + 215, + 903, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 503, + 661, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Unsupervised Semantic Segmentation", + "text_level": 1, + "bbox": [ + 513, + 530, + 838, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unsupervised Semantic Segmentation (USS) aims to train a model without any human annotations to produce the semantic segmentation mask. Before the surge of deep learning methods, researchers relied on numerous handcrafted methods [7, 16, 57, 68] that used one or more priors related to foreground regions in the image to generate segmentation masks, such as center[26], contrast[24], and boundary[54]. In deep learning-based methods, some approaches utilize generative models [17] to produce a segmentation mask, which will be used to synthesize a realistic image by copying the corresponding region into a background which is either synthesized [3] or taken from a different image [1]. Another category of methods leverages pseudo-labels constructed by handcrafted fixed strategies to train the detector. [62] adopts diverse noisy pseudo-labels via distinct unsupervised handcrafted methods to train the detector to predict a saliency map free from noise in labels. And [34, 44] proposed to refine the pseudo-labels produced by fixed-strategy by training the model via a self-supervised iterative refinement manner. These methods adopt a self-supervision manner that highly relies on pseudo-labels generated directly from one or multiple fixed strategies. If the pseudo-labels contain substantial noise, these methods remain prone to", + "bbox": [ + 511, + 553, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "30366", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "fitting incorrect knowledge, which is particularly evident when dealing with camouflaged objects.", + "bbox": [ + 89, + 90, + 485, + 122 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Camouflaged Object Detection", + "text_level": 1, + "bbox": [ + 89, + 133, + 362, + 148 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Camouflaged Object Detection (COD) is a complex task aimed at segmenting objects that are intentionally designed to merge seamlessly with their environment. This task has an extensive history in its domain. Existing fully-supervised COD methods [10, 37, 38, 69] have achieved excellent performance by training on large-scale annotated datasets. These methods utilize the rich information (e.g. boundary [25, 45, 66], texture [25, 42, 67] and other information like frequency domain and depth [28, 52, 56, 65]) provided by pixel-level annotations and design a series of sophisticated strategies to extract unique visual characteristics of camouflaged objects. However, the extensive pixel-level annotations are hardly gained, which has drawn the attention of researchers to Unsupervised Camouflaged Object Detection (UCOD) and USS methods [63]. These methods typically rely on pseudo-labels generated through fixed strategies to supervise model learning for segmentation. However, due to limitations in algorithmic complexity and robustness, these fixed strategies struggle to work effectively at high resolutions and easily introduce significant noise. Consequently, these methods are prone to learning incorrect noise from the pseudo-labels, leading to suboptimal segmentation performance and limiting the application in the real world of UCOD models.", + "bbox": [ + 89, + 156, + 483, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Self-supervised Learning", + "text_level": 1, + "bbox": [ + 89, + 530, + 320, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-Supervised Learning (SSL) has revolutionized the way models leverage unlabeled data. Among current SSL approaches, some methods adopt contrastive learning to learn image feature representations. MoCo [20] utilizes a dynamic memory bank to maintain a large pool of negative samples; SimCLR [6] employs various data augmentations and a projection head to maximize similarities between differently augmented views of the same instance; and BYOL [18] constructs a teacher-student framework with the teacher model's parameters are updated by Exponential Moving Average (EMA) of the student model. Following BYOL, DINO [5] applied two interactive encoders sharing the same structure but with different parameter sets and update strategies. Inspired by self-supervised pretraining techniques in natural language processing [40, 41, 46], other SSL methods design pretext tasks (e.g. colorization, rotation prediction, and patch reconstruction) on unlabeled images [5, 6, 14, 15, 20, 29, 33, 35]. BeiT [2] adopts a Transformer-based architecture to reconstruct masked image patches, MAE [21] implements a masked autoencoder that efficiently reconstructs the missing parts of an image. In this paper, we employ DINOv1 [5] and DINOv2 [36] to extract robust and generalized image features.", + "bbox": [ + 89, + 551, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 513, + 89, + 648, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overall Framework", + "text_level": 1, + "bbox": [ + 513, + 114, + 702, + 128 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The structure of our proposed method is shown in Fig. 3, which includes a teacher-student framework with an Adaptive Pseudo-label Mixing (APM) module, a Dual-Branch Adversarial (DBA) decoder, and a Look-Twice strategy. Both the teacher and student models are configured as DBA decoders.", + "bbox": [ + 511, + 136, + 906, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Firstly, the $i$ -th unlabeled input image passes through the self-supervised pretrained backbone to obtain high-level semantic image feature $F_{i}$ . We then employ the background seed method to generate a low-resolution fixed-strategy pseudo-label $\\hat{P}_i^{fs}$ , which utilizes pixel-level similarity comparison and was proposed in [44]. Simultaneously, the image feature $F_{i}$ is fed into a teacher-student model. The teacher model generates a higher-resolution teacher's pseudo label $\\hat{P}_i^t$ , and the student model produces the foreground and background predicted mask $\\hat{Y}_i^{FG}, \\hat{Y}_i^{BG}$ . Then, $\\hat{P}_i^{fs}$ and $\\hat{P}_i^t$ are fed into the APM module to construct the mixed dynamic pseudo-label $P_{i}$ . Finally, we use the mixed dynamic pseudo-label $P_{i}$ to supervise the final foreground and background predicted masks $\\hat{Y}_i^{FG}, \\hat{Y}_i^{BG}$ . We further apply a Look-Twice strategy to perform a secondary refinement on small-sized objects to improve the model's segmentation performance on small-sized samples.", + "bbox": [ + 511, + 227, + 906, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Adaptive Pseudo-Label Merging Module", + "text_level": 1, + "bbox": [ + 513, + 494, + 862, + 510 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing unsupervised semantic segmentation methods typically adopt a fixed-strategy (e.g. similarity comparison [32, 43, 53]) or utilize a dual-branch model (e.g. fixed-strategy or learnable branch [44, 63]) structure for self-supervised learning to generate segmentation masks. However, we believe the performance of these methods is constrained by the strategy's design or the pseudo-labels' quality, preventing the model from achieving higher accuracy.", + "bbox": [ + 511, + 516, + 905, + 637 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce a discriminator $\\mathcal{D}$ , which inputs a foreground segmentation mask (i.e. fixed-strategy pseudo-label $\\hat{P}_i^{fs}$ or student's foreground predicted mask $\\hat{Y}_i^{FG}$ ) and outputs the probability that the mask originates from the fixed-strategy branch:", + "bbox": [ + 511, + 638, + 905, + 712 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} _ {i} ^ {p 1} = \\mathcal {D} \\left(\\hat {P} _ {i} ^ {f s}\\right), \\hat {y} _ {i} ^ {p 2} = \\mathcal {D} \\left(\\hat {Y} _ {i} ^ {F G}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 719, + 903, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{y}}_i^{p1},\\hat{\\mathbf{y}}_i^{p2}$ denote the predicted probability that the input mask belongs to the fixed-strategy branch. We then introduce a scoring function $S(\\cdot)$ with a temporal constraint:", + "bbox": [ + 511, + 747, + 906, + 794 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} \\left(\\hat {\\mathrm {y}} _ {i} ^ {p 1}, \\hat {\\mathrm {y}} _ {i} ^ {p 2}\\right) = \\operatorname {C L I P} \\left(\\frac {t}{T} + \\frac {1}{2} \\left(1 + \\cos \\left(\\pi \\times \\left| \\hat {\\mathrm {y}} _ {i} ^ {p 1} - \\hat {\\mathrm {y}} _ {i} ^ {p 2} \\right|\\right)\\right) _ {0} ^ {1}, \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 806, + 916, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $t, T$ denote the current and total training epochs, $\\mathrm{CLIP}(x)_0^1$ denotes truncating $x$ so that it lies within the interval [0, 1]. In the early stage of training, the teacher model", + "bbox": [ + 511, + 854, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "30367", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/500a56bfabcc9538ae36538d3af179e5afd19ea4959a3c6175fd76829ca5b5fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 88, + 903, + 292 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f58d65e25dfd0c36d887656828ddc15dc7851f6fec0e0893f7195b056fc37be1.jpg", + "image_caption": [ + "Figure 3. The main framework of our proposed method. The model contains a teacher-student framework, an Adaptive Psdudo-label Mixing (APM) module, a Dual-Branch Adversarial (DBA) decoder, and a Look-Twice strategy." + ], + "image_footnote": [], + "bbox": [ + 99, + 297, + 905, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "has not yet developed sufficient localization and basic semantic segmentation capabilities for camouflaged objects, making it unable to supervise the student model effectively. At this point, there is a significant difference between fixed-strategy pseudo-label $\\hat{P}_i^{fs}$ and the student's foreground predicted mask $\\hat{Y}_i^{FG}$ . To help the model learn the fundamental foreground localization and segmentation tasks, we use a higher proportion of $\\hat{P}_i^{fs}$ to supervise model training. Later in the training process, as the $\\hat{Y}_i^{FG}$ gradually approaches $\\hat{P}_i^{fs}$ , we increase the proportion of the teacher's pseudo-label $\\hat{P}_i^t$ as it can make more stable predictions. This adjustment helps prevent the model from overfitting to noise and incorrect knowledge in the $\\hat{P}_i^{fs}$ . We use the obtained score as the mixing weight $W_i^t$ to combine the fixed-strategy pseudo-label $\\hat{P}_i^{fs}$ and teacher's predictions $\\hat{P}_i^t$ :", + "bbox": [ + 89, + 523, + 483, + 760 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {W} _ {i} ^ {t} = \\mathcal {S} \\left(\\hat {\\mathrm {y}} _ {i} ^ {p 1}, \\hat {\\mathrm {y}} _ {i} ^ {p 2}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 768, + 482, + 796 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {i} = \\mathrm {W} _ {i} ^ {t} \\hat {P} _ {i} ^ {t} + (1 - \\mathrm {W} _ {i} ^ {t}) \\hat {P} _ {i} ^ {f s},\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 792, + 387, + 811 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P_{i}$ denotes the mixed dynamic pseudo-label. For the training of the discriminator, we set a binary classification task by using $\\hat{P}_i^{fs}$ and $\\hat{Y}_i^{FG}$ as the input and assign the discriminator label $\\mathrm{y}_i^{p1} = 1$ for the fixed-strategy pseudo-label $\\hat{P}_i^{fs}$ and $\\mathrm{y}_i^{p2} = 0$ for the student model branch $\\hat{Y}_i^{FG}$ . Then,", + "bbox": [ + 89, + 821, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we employ a cross-entropy supervision loss $\\mathcal{L}_{\\mathrm{BCE}}$ to train the discriminator:", + "bbox": [ + 511, + 523, + 906, + 551 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d i s} = \\mathcal {L} _ {\\mathrm {B C E}} \\left(\\hat {\\mathrm {y}} _ {i} ^ {p j}, \\mathrm {y} _ {i} ^ {p j}\\right), j \\in \\{1, 2 \\}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 559, + 906, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During the training, we treat the student model as a kind of generator and alternately train the decoder and the discriminator in a GAN-like manner.", + "bbox": [ + 511, + 583, + 905, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Dual-Branch Adversarial Decoder", + "text_level": 1, + "bbox": [ + 511, + 636, + 815, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In previous methods [44, 63], predictions are typically generated by a simple $1 \\times 1$ convolution, which limits the ability to capture camouflage features thoroughly. We propose a Dual-Branch Adversarial decoder to enhance accuracy and robustness in this task. The DBA decoder consists of two parallel branches: one is responsible for segmenting the foreground predicted mask $\\hat{Y}_i^{FG}$ , while the other is for segmenting the background predicted mask $\\hat{Y}_i^{BG}$ .", + "bbox": [ + 511, + 657, + 906, + 780 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, for $i$ -th image feature $F_{i}$ extracted from the backbone model, a convolutional layer is used to double its channel, then we decouple it into foreground and background features $F_{i}^{FG}, F_{i}^{BG}$ with the same shape of $F_{i}$ :", + "bbox": [ + 511, + 780, + 906, + 842 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {i} ^ {F G}, F _ {i} ^ {B G} = \\operatorname {S p l i t} \\left(\\operatorname {C o n v} \\left(F _ {i}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 845, + 903, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\text{Split}(\\cdot)$ denotes splitting a tensor along the channel dimension equally. We then employ two learnable em", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "30368", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "beddings $\\mathrm{E}_{FG},\\mathrm{E}_{BG}$ to store foreground and backgroundrelated knowledge. These embeddings are used to calculate foreground and background attention queries $Q_{i}^{FG^{\\prime}}$ $Q_{i}^{BG^{\\prime}}$ ..", + "bbox": [ + 89, + 90, + 483, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {i} ^ {F G ^ {\\prime}} = F _ {i} ^ {F G} \\times \\mathrm {E} _ {F G}, Q _ {i} ^ {B G ^ {\\prime}} = F _ {i} ^ {B G} \\times \\mathrm {E} _ {B G}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 148, + 482, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we apply multiplicative attention with a residual connection to weigh the features, enhancing important details. The resulting features are then passed through a $1 \\times 1$ convolution to output the foreground predicted mask $\\hat{Y}_i^{FG}$ and background mask $\\hat{Y}_i^{BG}$ :", + "bbox": [ + 89, + 172, + 483, + 252 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {Y} _ {i} ^ {F G} = \\operatorname {C o n v} ^ {\\prime} \\left(\\operatorname {S i g m o i d} \\left(Q _ {i} ^ {F G ^ {\\prime}} \\times F _ {i} ^ {F G}\\right) + F _ {i} ^ {F G}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 260, + 482, + 287 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {Y} _ {i} ^ {B G} = \\operatorname {C o n v} ^ {\\prime} (\\operatorname {S i g m o i d} \\left(Q _ {i} ^ {B G ^ {\\prime}} \\times F _ {i} ^ {B G}\\right) + F _ {i} ^ {B G}).\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 282, + 444, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the student model, we use $Y_{i}^{FG}, \\hat{Y}_{i}^{BG}$ as the student's foreground and background prediction masks. For the teacher model, we further generate the teacher's pseudolabel $P_{i}$ using the combination of the foreground and background prediction masks.", + "bbox": [ + 89, + 314, + 483, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by [55], we apply an orthogonal loss $\\mathcal{L}_{\\perp}$ for $Q_{i}^{FG^{\\prime}}\\in \\mathbb{R}^{(n\\times m)\\times c},Q_{i}^{BG^{\\prime}}\\in \\mathbb{R}^{(n\\times m)\\times c}$ as a regularisation constraint to encourage them to focus on distinct features:", + "bbox": [ + 89, + 391, + 483, + 436 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {M} = Q _ {i} ^ {F G ^ {\\prime}} \\left(Q _ {i} ^ {B G ^ {\\prime}}\\right) ^ {\\top},\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 446, + 361, + 465 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\perp} = \\frac {1}{(n \\times m) ^ {2}} \\sum_ {j = 1} ^ {n \\times m} \\sum_ {k = 1} ^ {n \\times m} (1 - \\delta_ {j k}) (\\mathbb {M} [ j, k ]) ^ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 470, + 482, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $n, m, c$ indicate the height, width, and number of channels. When $j = k$ , $\\delta_{jk} = 1$ ; otherwise $\\delta_{jk} = 0$ . For the predictions $\\hat{Y}_i^{FG}, \\hat{Y}_i^{BG}$ , we use a binary cross-entropy loss $\\mathcal{L}_{\\mathrm{BCE}}$ for supervision, where the background prediction is inverted before applying the supervision:", + "bbox": [ + 89, + 521, + 483, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s e g} = \\mathcal {L} _ {\\mathrm {B C E}} (\\hat {Y} _ {i} ^ {F G}, P _ {i}) + \\mathcal {L} _ {\\mathrm {B C E}} \\left((1 - \\hat {Y} _ {i} ^ {B G}), P _ {i}\\right). (9)\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 608, + 483, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Look-Twice Refinement", + "text_level": 1, + "bbox": [ + 89, + 643, + 310, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The segmentation quality of COD is often influenced by object size. Due to reduced spatial resolution, the model struggles to produce fine segmentation for small-sized camouflaged objects. Inspired by previous works [37, 38], we attempt to mimic human behavior when observing small-sized objects: first coarsely locating the camouflaged object in the image, then zooming in to carefully retrieve its details. To this end, we designed the Look-Twice mechanism. First, the input image is processed by the model to obtain a coarse segmentation result $\\hat{Y}_i^{FG}$ . We define the set of all foreground objects as $\\mathrm{C}^{FG}$ , which can be calculated as:", + "bbox": [ + 89, + 664, + 483, + 832 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {C} ^ {\\mathrm {F G}} = \\text {C o n n e c t C o m p o n e n t} \\left(\\hat {Y} _ {i} ^ {F G}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 840, + 482, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where ConnectComponent(\\cdot) denotes the connected components labelling algorithm. Next, for each connect", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "component $c_k^{FG} \\in \\mathbb{C}^{FG}$ , we calculate its foreground area ratio $r_k^{FG}$ :", + "bbox": [ + 511, + 89, + 905, + 122 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr _ {k} ^ {F G} = \\frac {\\sum_ {(x , y) \\in \\mathrm {c} _ {k} ^ {F G}} \\hat {Y} _ {i} ^ {F G} (x , y)}{H \\times W}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 128, + 903, + 164 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $H, W$ denote the height and width of the foreground area of the output mask $\\tilde{Y}_i^{FG}$ . We consider all regions where $r_i^{FG} < \\tau$ as small-sized object areas. We set $\\tau = 0.15$ when using DINOv2 as the backbone. The hyperparameter ablation study on $\\tau$ will be provided in the experiments section.", + "bbox": [ + 511, + 167, + 905, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each cropped small-sized object region, we aim to retain a certain proportion of the background to provide the model with sufficient context to segment the foreground from the background. Therefore, for each connected component region $c_k^{FG}$ , we calculate the local foreground area ratio $s_k^{FG}$ and the global area ratio $s_k^{BG}$ :", + "bbox": [ + 511, + 258, + 905, + 349 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {k} ^ {F G} = \\frac {\\sum_ {(x , y) \\in c _ {k} ^ {F G}} \\hat {Y} _ {i} ^ {F G} (x , y)}{H _ {k} ^ {F G} \\times W _ {k} ^ {F G}}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 354, + 903, + 398 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {k} ^ {B G} = \\frac {H _ {k} ^ {F G} \\times W _ {k} ^ {F G}}{H \\times W},\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 396, + 759, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $H_{k}^{FG},W_{k}^{FG}$ denote the height and width of the bounding box of the $k$ -th connected component $\\mathrm{c}_k^{FG}$ ..", + "bbox": [ + 511, + 433, + 905, + 465 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {s c a l e} _ {k} ^ {F G} = 1 - \\left(s _ {k} ^ {F G} / s _ {k} ^ {B G}\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 470, + 903, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we calculate the expansion ratio using Eq. (13) and enlarge the bounding box of the connected component region by this ratio. The expanded bounding box is then used to crop the related area of the input image, and the cropped patches are resized to the original input size. During the training process, these patches are treated as augmented data. At test time, we re-infer these patches, scale the resulting foreground predicted mask back to the original size, and paste it into the coarse mask as the refined result. However, we found that this mechanism still has some weakness: for cases where objects are occluded and the annotations are fragmented, this method may mistakenly amplify a part of the object as a small object. This can serve as a point of our future work.", + "bbox": [ + 511, + 494, + 905, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Total Loss", + "text_level": 1, + "bbox": [ + 511, + 713, + 629, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the total loss $\\mathcal{L}_{tot}$ can be defined as follows:", + "bbox": [ + 511, + 734, + 859, + 751 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t o t} = \\mathcal {L} _ {s e g} + \\mathcal {L} _ {\\perp} + \\mathcal {L} _ {d i s}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 757, + 903, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{seg}$ is the segmentation loss defined by Eq. (9), $\\mathcal{L}_{\\perp}$ is the orthogonal loss defined by Eq. (8), and $\\mathcal{L}_{dis}$ is the discriminator loss defined by Eq. (4). When training on the teacher-student framework, we freeze the parameters of the discriminator and use $\\mathcal{L}_{seg}$ to supervise the model training. Conversely, when training the discriminator, we freeze the parameters of the other parts of the model and train the discriminator only using $\\mathcal{L}_{dis}$ .", + "bbox": [ + 511, + 779, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "30369", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d49f7e8cb0c745161f0aa8f4316084b0dd6dda57b92b08cae9cd77a40f5df15a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsCHAMELEON (87)CAMO-Test (250)COD10K-Test (2,026)NC4K (4,121)
Sm↑FωβFmβEmφM↓Sm↑FωβFmβEmφM↓Sm↑FωβFmβEmφM↓Sm↑FωβFmβEmφM↓
Fully-Supervised Methods
SINet20[10].872.806.827.946.034.751.606.675.771.100.771.551.634.806.051.808.723.769.871.058
C2FNet21[47].888.828.844.946.032.796.719.762.864.080.813.686.723.900.036.838.762.794.904.049
MGL-R21[61].893.812.834.941.030.775.673.726.842.088.814.666.710.890.035.833.739.782.893.053
UGTR21[58].887.794.819.940.031.784.684.735.851.086.817.666.711.890.036.839.746.787.899.052
BGNet22[48].901.851.860.954.027.812.749.789.870.073.831.722.753.901.033.851.788.820.907.044
ZoomNet22[37].902.845.864.958.023.820.752.794.878.066.838.729.766.888.029.853.784.818.896.043
SINetv22[12].888.816.835.961.030.820.743.782.882.070.815.680.718.887.037.847.770.805.903.048
HitNet23[22].921.897.900.972.019.849.809.831.906.055.871.806.823.935.023.875.834.853.926.037
FSPNet23[23].908.851.867.965.023.856.799.830.899.050.851.735.769.895.026.879.816.843.915.035
BiRefNet24[64].929.911.922.968.016.932.914.922.974.015.913.874.888.960.014.914.894.909.953.023
Semi-Supervised Methods
CamoTeacher24(1%)[27].652.472.558.714.093.621.456.545.669.136.699.517.788.797.062.718.599.779.814.090
CamoTeacher24(5%)[27].729.587.656.785.070.669.523.601.711.122.745.583.827.840.050.777.677.834.859.071
CamoTeacher24(10%)[27].756.617.684.813.065.701.560.742.795.112.759.594.836.854.049.791.687.842.868.068
SCOD-ND24(10%)[13].850.773-.928.036.789.732-.859.077.819.725-.891.033.838.787-.903.046
Unsupervised Methods
BigGW21[51].547.244.294.527.257.565.299.349.528.282.528.185.246.497.261.608.319.391.565.246
TokenCut22[53].654.496.536.740.132.633.498.543.706.163.658.469.502.735.103.725.615.649.802.101
TokenCut22 w/B.S.[53].655.351.393.582.169.639.383.434.595.195.666.334.399.609.127.735.478.547.683.133
SpectralSeg22[32].575.410.440.628.220.579.450.481.648.235.575.360.388.595.193.669.535.562.719.159
SelfMask22[43].619.436.481.675.176.617.483.536.698.176.637.431.469.679.131.716.593.634.777.114
SelfMask22 w/U.B.[43].629.447.491.683.169.627.495.547.708.182.645.440.478.687.125.723.601.642.784.110
FOUND22-DINOv1[44].684.542.590.810.095.685.584.633.782.129.670.482.520.751.085.741.637.674.824.084
*FOUND22-DINOv2[44].829.757.781.911.040.770.704.740.849.090.767.641.668.847.045.816.756.783.893.052
UCOS-DA23-DINOv1[63].715.591.629.802.095.701.606.646.784.127.689.513.546.740.086.755.656.689.819.085
*UCOS-DA23-DINOv2[63].750.639.666.808.091.702.604.633.751.148.655.467.495.687.120.731.617.644.785.103
OursDINOv1.734.625.680.854.072.706.621.689.801.108.727.577.627.822.059.761.680.737.851.074
OursDINOv2.864.825.838.931.031.793.747.779.862.077.834.763.779.916.031.850.818.835.923.043
", + "bbox": [ + 94, + 88, + 906, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison of our methods with recent methods. We compared our proposed methods with competing unsupervised, semi-supervised, and full-supervised methods. Bold indicates the best result in group settings, and underline indicates the second-best result. * denotes the version that reimplemented by us.", + "bbox": [ + 89, + 473, + 906, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 541, + 215, + 558 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 566, + 297, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Trainsets. To fairly compare with the existing works, following [10, 30], we used a combination of 1000 images from the CAMO-Train and 3040 images from COD10K-Train as the training dataset of our experiments. Following the unsupervised learning settings, we do not use any ground-truth label during the training process.", + "bbox": [ + 89, + 592, + 482, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Testing Sets. We evaluate the performance of our methods on four mainstream COD benchmark testing sets: CHAMELEON with 76 test images, CAMO with 250 test images, COD10K with 2026 test images, and NC4K with 4121 test images.", + "bbox": [ + 89, + 685, + 482, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Protocol. For a fair and comprehensive evaluation, we employ the S-measure $(S_m)$ [8], mean and weighted F-measure $(\\mathcal{F}_\\beta^m,\\mathcal{F}_\\beta^\\omega)$ [31], mean E-measure $(\\mathcal{E}_{\\epsilon}^{m})$ [9], mean absolute error $(\\mathcal{M})$ [39] as the main metrics.", + "bbox": [ + 89, + 762, + 482, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Following [44], we employ the simple background seed methods as our fixed strategy branch to generate the coarse pseudo-label. And following [44, 63], we use the strong self-supervised pretrained backbone DINO [5, 36] as our encoder. The teacher model", + "bbox": [ + 89, + 824, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "is updated via EMA with a momentum $\\eta$ of 0.99. We train the model for 25 epochs. The batch size is set to 32 for each GPU during training. All experiments are implemented with PyTorch 2.1 and run on a machine with Intel(R) Xeon(R) Silver 4214R CPU @ 2.40GHz, 512GiB RAM, and 1 NVIDIA Titan A100-40G GPUs. All experiments use the same random seed. More implementation details will be provided in the supplementary materials.", + "bbox": [ + 511, + 542, + 906, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 511, + 678, + 653, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Analysis. We show visualizations of a series of camouflaged object segmentation masks predicted by our method and related methods in some challenging scenarios. As shown in Fig. 4, we noticed that our method achieves higher segmentation accuracy than existing approaches, with clearer edges and a more detailed representation of the camouflaged object. Note that DINOv2 is employed for all methods that require a backbone to extract features.", + "bbox": [ + 511, + 702, + 906, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Analysis. In Tab. 1, we compared our proposed method's performance with competing USS and UCOD models on four COD test datasets. To further highlight our model's competitiveness, we also compared ex", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "30370", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f46a89772bbc1654c1a1fab1dd8c61f22cc6d71261dde5982a42629c3e683b42.jpg", + "image_caption": [ + "Figure 4. Visual comparison of our method with other existing methods in challenging scenarios. Our method has clearer and more precise segmentation boundaries and correctly recognizes depth-artifacted objects." + ], + "image_footnote": [], + "bbox": [ + 96, + 93, + 897, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "isting fully-supervised and semi-supervised methods. The results show that our model outperformed all existing USS and UCOD methods across all metrics and datasets, thus achieving State-Of-The-Art (SOTA) performance. Additionally, based on DINOv2 backbone, our model has surpassed several semi-supervised and fully-supervised methods across all four datasets, demonstrating its superior performance, effectiveness, and robustness.", + "bbox": [ + 89, + 518, + 483, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 651, + 243, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To verify the effectiveness of our methods, we employ $DI-NOv2$ as the backbone and conduct comprehensive ablation studies on the COD10K-Test dataset with 2,026 images.", + "bbox": [ + 89, + 672, + 482, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Module Ablation. We first conduct ablation experiments on the teacher-student framework (Tea-Stu) and our proposed methods: Adaptive Pseudo-label Mixing (APM) module, Dual-Branch Adversarial (DBA) decoder, and Look-Twice strategy. To thoroughly demonstrate their effectiveness, we replaced the decoder with a single $1 \\times 1$ convolutional layer following [44, 63]. For all ablations on the APM module, we used the $1:1$ constant mixing strategy to replace it. The ablation results are shown in Tab. 2. As a result, we find that when only the teacher-student framework is used, the simple $1 \\times 1$ convolution fails to learn camouflage-related knowledge adequately. When training", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/94999939b255f78e815bb0a5b4e292037407b3b1df3fb382a67c50aca4468b8e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingsCOD10K (2026)
Tea-StuAPMDBALook-TwiceSm↑Fωβ↑Fmβ↑Emφ↑M↓
✓.626.448.464.658.112
✓✓.633.452.472.668.129
✓.607.409.435.631.129
✓✓.603.404.431.625.133
✓✓.724.582.604.800.054
✓✓✓.788.679.700.879.043
✓✓✓.784.675.692.869.036
✓✓✓✓.834.763.779.916.031
", + "bbox": [ + 517, + 516, + 908, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation study to evaluate the proposed modules. We retrained our model with different settings on the same learning rate and epochs, then benchmarked on COD10K-testset to validate the effectiveness.", + "bbox": [ + 511, + 679, + 906, + 736 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with a DBA decoder only, the model lacks the guidance of the teacher-student framework, thus easily learning the incorrect knowledge from the low-quality pseudo-labels constructed by fixed-strategy. The proposed Look-Twice mechanism is effective only when the network can produce sufficiently high-quality predictions. If the network has poor localization and segmentation capabilities, Look-Twice will degrade performance.", + "bbox": [ + 511, + 763, + 906, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mixing Strategy Ablation. We continue conducting ab-", + "bbox": [ + 511, + 885, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "30371", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/73d05e0996f2ad857bd64cd64c8f30c38a5cf0a425a0bf0f9638b0787ec19781.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Mixing StrategiesCOD10K (2026)
Sm↑Fωβ↑Fmβ↑Eφ↑M↓
Proportional(1:1) Mixing.707.555.572.763.082
Linear Decay Mixing.809.716.723.888.038
APM.834.763.779.916.031
", + "bbox": [ + 94, + 88, + 482, + 170 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4f3057181878e30e094f1cc195b8e5535dee4ae504bb963561870aa8c1dbdb6c.jpg", + "image_caption": [ + "Figure 5. Hyper-parameter ablation of small-sized object ratio $\\tau$ . The model performs optimally and shows stable results when the value of $\\tau$ is set around 0.15." + ], + "image_footnote": [], + "bbox": [ + 99, + 242, + 468, + 380 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9c12f184efb53acdc1c26ebd60956ece2c2cfa7e4f1f1ba11d1df4e04dad3d6b.jpg", + "table_caption": [ + "Table 3. Ablation to the pseudo-label mixing strategy. We retrained our model with different pseudo-label mixing strategies and benchmarked on the COD10K-Test dataset." + ], + "table_footnote": [], + "table_body": "
Fixed Generation StrategiesCOD10K (2026)
Sm↑Fωβ↑Fmβ↑Emφ↑M↓
Null(Pure Black/White).459.002.001.250.091
Random Perlin Noise.667.489.543.788.072
Background Seed.834.763.779.916.031
", + "bbox": [ + 94, + 454, + 482, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Fixed Pseudo-label Generation Strategy Ablation. We retrained our model with a replaced fixed strategy and benchmarked it on the COD10K-Test dataset.", + "bbox": [ + 89, + 542, + 482, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "lation studies on the mixing strategies. We compared the proportional(1:1), linear decay, and APM mixing strategies. As shown in Tab. 3, the APM module provides a method for fusing pseudo-labels that is more aligned with the model's training process and achieving optimal performance.", + "bbox": [ + 89, + 613, + 482, + 688 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hyper-parameter Ablation. We conducted ablation study on the threshold $\\tau$ for the small-sized object ratio mentioned in Sec. 3.4. As shown in Fig. 5, when $\\tau \\approx 0.15$ , the model achieves optimal and stable performance. Therefore, we set $\\tau$ to 0.15 in the final model.", + "bbox": [ + 89, + 688, + 482, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Study on Foreground-sizes. We divided the test set by the proportion of test set prospects at $2\\%$ intervals, then benchmarked the performance between our method and some previous SOTA methods[44, 63]. As shown in Fig. 6, our method has significant advantages when dealing with smaller-sized camouflaged targets.", + "bbox": [ + 89, + 763, + 482, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fixed Pseudo-label Generation Strategy Ablation. We further investigated the effect of the pseudo-labels constructed by the fixed-strategy. For each input image, we", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9283571ec3bb47002208c1ee0b09a24fd8100561a05124ff3767814af27a63e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 88, + 712, + 188 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e6f62818efa23ab878b50496c92983054a0701f4478add2bbf493b9980bc2bd1.jpg", + "image_caption": [ + "Figure 6. Performance comparison for different foreground sizes on COD10K-Test dataset." + ], + "image_footnote": [], + "bbox": [ + 514, + 189, + 712, + 290 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/64f6a5107d93963b83679ae70d0618ca8c9412b0cf8707566ca24fb47f217f9c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 88, + 905, + 189 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d8e6e54f6f1a3eb10a5753a404a0f77665ca280763698d09042fd0792333e539.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 189, + 905, + 290 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "generated a random shape of Perlin noise on a plain black background. We used this \"fabricated pseudo-label\" as well as the pure black and white masks to replace the fixed-strategy pseudo-label during the training process. The results are shown in Tab. 4. When using random noise, the noise may overlap with the foreground region, allowing the network to explore some knowledge related to the camouflaged foreground. However, when using a solid color mask, the model fails to learn the basic definition of the segmentation task to produce valid predictions.", + "bbox": [ + 511, + 358, + 906, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 513, + 526, + 638, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Considering the existing UCOD methods suffer from main issues: 1). The low-quality pseudo-labels constructed by fixed strategies contain substantial noise, making the model prone to fitting incorrect knowledge. 2). The low-resolution pseudo-labels cause severe confusion between the foreground and background. Simple $1 \\times 1$ convolution fails to capture and learn the semantic features of camouflaged objects, especially for small-sized samples. To solve these problems, we propose UCOD-DPL in this paper. We introduce a teacher-student framework with an Adaptive Pseudolabel Module, a Dual-Branch Adversarial decoder, and a Look-Twice mechanism. The APM dynamically fuses pseudo-labels from the fixed strategy and teacher model, effectively preventing the model from overfitting the noise and erroneous knowledge. The DBA decoder mines foreground and background knowledge relevant to camouflaged objects while resisting incorrect information from pseudolabels. The Look-Twice refinement mechanism enhances camouflaged object learning and refinement of small objects. We conducted extensive benchmarking on multiple COD test datasets, and the results show that our method outperforms all previous approaches and even reaches the performance level of some fully supervised methods.", + "bbox": [ + 511, + 553, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "30372", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Science Fund for Distinguished Young Scholars (No.62025603), the National Natural Science Foundation of China (No. U21B2037, No. U22B2051, No. U23A20383, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No. 2021J06003, No. 2022J06001).", + "bbox": [ + 89, + 113, + 485, + 227 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 252, + 187, + 268 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Relja Arandjelović and Andrew Zisserman. Object discovery with a copy-pasting gan. arXiv preprint arXiv:1905.11369, 2019. 2", + "[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.3", + "[3] Adam Bielski and Paolo Favaro. Emergence of object segmentation in perturbed generative models. Advances in Neural Information Processing Systems, 32, 2019. 2", + "[4] TIM Caro. The adaptive significance of coloration in mammals. BioScience, 55(2):125-136, 2005. 1", + "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In The IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 3, 6", + "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 3", + "[7] Ming-Ming Cheng, Niloy J Mitra, Xiaolei Huang, Philip HS Torr, and Shi-Min Hu. Global contrast based salient region detection. IEEE transactions on pattern analysis and machine intelligence, 37(3):569-582, 2014. 2", + "[8] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In The IEEE/CVF International Conference on Computer Vision, pages 4548-4557, 2017. 6", + "[9] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, Ming-Ming Cheng, and Ali Borji. Enhanced-alignment measure for binary foreground map evaluation. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, pages 698-704. AAAI Press, 2018. 6", + "[10] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 2777-2787, 2020. 3, 6", + "[11] Deng-Ping Fan, Ge-Peng Ji, Tao Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, and Ling Shao. Pranet: Parallel reverse attention network for polyp segmentation. In International conference on medical image computing and computer-assisted intervention, pages 263-273. Springer, 2020. 1", + "[12] Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. Concealed object detection. IEEE Transactions on" + ], + "bbox": [ + 93, + 277, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Pattern Analysis and Machine Intelligence, 44(10):6024-6042, 2022. 6", + "[13] Yuanbin Fu, Jie Ying, Houlei Lv, and Xiaojie Guo. Semi-supervised camouflaged object detection from noisy data. In ACM International Conference on Multimedia, 2024. 6", + "[14] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations, 2018. 3", + "[15] Spyros Gidaris, Andrei Bursuc, Gilles Puy, Nikos Komodakis, Matthieu Cord, and Patrick Pérez. Obow: Online bag-of-visual-words generation for self-supervised learning. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 6830–6840, 2021. 3", + "[16] Stas Goferman, Lihi Zelnik-Manor, and Ayellet Tal. Context-aware saliency detection. IEEE transactions on pattern analysis and machine intelligence, 34(10):1915-1926, 2011. 2", + "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2", + "[18] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 3", + "[19] Chunming He, Kai Li, Yachao Zhang, Longxiang Tang, Yu-lun Zhang, Zhenhua Guo, and Xiu Li. Camouflaged object detection with feature decomposition and edge reconstruction. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 22046-22055, 2023. 1", + "[20] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 9729-9738, 2020. 3", + "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 16000-16009, 2022. 3", + "[22] Xiaobin Hu, Shuo Wang, Xuebin Qin, Hang Dai, Wenqi Ren, Donghao Luo, Ying Tai, and Ling Shao. High-resolution iterative feedback network for camouflaged object detection. In Association for the Advancement of Artificial Intelligence, pages 881-889, 2023. 6", + "[23] Zhou Huang, Hang Dai, Tian-Zhu Xiang, Shuo Wang, Huai-Xin Chen, Jie Qin, and Huan Xiong. Feature shrinkage pyramid for camouflaged object detection with transformers. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 5557-5566, 2023. 6", + "[24] Laurent Itti, Christof Koch, and Ernst Niebur. A model of saliency-based visual attention for rapid scene analysis. IEEE Transactions on pattern analysis and machine intelligence, 20(11):1254-1259, 1998. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "30373", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Ge-Peng Ji, Deng-Ping Fan, Yu-Cheng Chou, Dengxin Dai, Alexander Liniger, and Luc Van Gool. Deep gradient learning for efficient camouflaged object detection. Machine Intelligence Research, 20:92-108, 2023. 3", + "[26] Tilke Judd, Krista Ehinger, Frédo Durand, and Antonio Torralba. Learning to predict where humans look. In The IEEE/CVF International Conference on Computer Vision, pages 2106-2113. IEEE, 2009. 2", + "[27] Xunfa Lai, Zhiyu Yang, Jie Hu, Shengchuan Zhang, Liujuan Cao, Guannan Jiang, Zhiyu Wang, Songan Zhang, and Rongrong Ji. Camoteacher: Dual-rotation consistency learning for semi-supervised camouflaged object detection. In European Conference on Computer Vision, 2024. 1, 6", + "[28] Jiaying Lin, Xin Tan, Ke Xu, Lizhuang Ma, and Rynson W. H. Lau. Frequency-aware camouflaged object detection. ACM Trans. Multimedia Comput. Commun. Appl., 19 (2), 2023. 3", + "[29] Jianghang Lin, Yunhang Shen, Bingquan Wang, Shaohui Lin, Ke Li, and Liujuan Cao. Weakly supervised open-vocabulary object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3404-3412, 2024. 3", + "[30] Naisong Luo, Yuwen Pan, Rui Sun, Tianzhu Zhang, Zhiwei Xiong, and Feng Wu. Camouflaged instance segmentation via explicit de-camouflaging. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 17918-17927, 2023. 6", + "[31] Ran Margolin, Lihi Zelnik-Manor, and Ayellet Tal. How to evaluate foreground maps? In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 248-255, 2014. 6", + "[32] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 8364-8375, 2022. 3, 6", + "[33] Peng Mi, Jianghang Lin, Yiyi Zhou, Yunhang Shen, Gen Luo, Xiaoshuai Sun, Liujuan Cao, Rongrong Fu, Qiang Xu, and Rongrong Ji. Active teacher for semi-supervised object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14482-14491, 2022. 3", + "[34] Tam Nguyen, Maximilian Dax, Chaithanya Kumar Mummadi, Nhung Ngo, Thi Hoai Phuong Nguyen, Zhongyu Lou, and Thomas Brox. Deepusps: Deep robust unsupervised saliency prediction via self-supervision. Advances in Neural Information Processing Systems, 32, 2019. 2", + "[35] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 3", + "[36] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou," + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 3, 6", + "[37] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, 2022. 2, 3, 5, 6", + "[38] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoomnext: A unified collaborative pyramid network for camouflaged object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024, 1, 2, 3, 5", + "[39] Federico Perazzi, Philipp Krahenbuhl, Yael Pritch, and Alexander Hornung. Saliency filters: Contrast based filtering for salient region detection. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 733-740. IEEE, 2012. 6", + "[40] Alec Radford. Improving language understanding by generative pre-training. OpenAI blog, 2018. 3", + "[41] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3", + "[42] Jingjing Ren, Xiaowei Hu, Lei Zhu, Xuemiao Xu, Yangyang Xu, Weiming Wang, Zijun Deng, and Pheng-Ann Heng. Deep texture-aware features for camouflaged object detection. IEEE Transactions on Circuits and Systems for Video Technology, 33(3):1157-1167, 2023. 3", + "[43] Gyungin Shin, Samuel Albanie, and Weidi Xie. Unsupervised salient object detection with spectral cluster voting. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 3971-3980, 2022. 3, 6", + "[44] Oriane Simeoni, Chloe Sekkat, Gilles Puy, Antonín Vobecký, Éloi Zablocki, and Patrick Pérez. Unsupervised object localization: Observing the background to discover objects. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 3176-3186, 2023. 1, 2, 3, 4, 6, 7, 8", + "[45] Dongyue Sun, Shiyao Jiang, and Lin Qi. Edge-aware mirror network for camouflaged object detection, 2023. 3", + "[46] Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. Bert4rec: Sequential recommendation with bidirectional encoder representations from transformer. In Proceedings of the 28th ACM international conference on information and knowledge management, pages 1441-1450, 2019. 3", + "[47] Yujia Sun, Geng Chen, Tao Zhou, Yi Zhang, and Nian Liu. Context-aware cross-level fusion network for camouflaged object detection. In The International Joint Conference on Artificial Intelligence, pages 1025–1031, 2021. 6", + "[48] Yujia Sun, Shuo Wang, Chenglizhao Chen, and Tian-Zhu Xiang. Boundary-guided camouflaged object detection. In The International Joint Conference on Artificial Intelligence, pages 1335-1341, 2022. 6", + "[49] Yanguang Sun, Hanyu Xuan, Jian Yang, and Lei Luo. Glconet: Learning multisource perception representation for camouflaged object detection. IEEE Transactions on Neural Networks and Learning Systems, 2024. 1" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "30374", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Domen Tabernik, Samo Šela, Jure Skvarč, and Danijel Skočaj. Segmentation-based deep-learning approach for surface-defect detection. Journal of Intelligent Manufacturing, 31(3):759-776, 2020. 1", + "[51] Andrey Voynov, Stanislav Morozov, and Artem Babenko. Object segmentation without labels with large-scale generative models. In International Conference on Machine Learning, pages 10596-10606. PMLR, 2021. 6", + "[52] Qingwei Wang, Jinyu Yang, Xiaosheng Yu, Fangyi Wang, Peng Chen, and Feng Zheng. Depth-aided camouflaged object detection. In Proceedings of the 31st ACM International Conference on Multimedia, page 3297–3306, 2023. 3", + "[53] Yangtao Wang, Xi Shen, Shell Xu Hu, Yuan Yuan, James L Crowley, and Dominique Vaufreydaz. Self-supervised transformers for unsupervised object discovery using normalized cut. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 14543-14553, 2022. 3, 6", + "[54] Yichen Wei, Fang Wen, Wangjiang Zhu, and Jian Sun. Geodesic saliency using background priors. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part III 12, pages 29-42. Springer, 2012. 2", + "[55] Among Wu, Rui Liu, Yahong Han, Linchao Zhu, and Yi Yang. Vector-decomposed disentanglement for domain-invariant object detection. The IEEE/CVF International Conference on Computer Vision, 2021. 5", + "[56] Mochu Xiang, Jing Zhang, Yunqiu Lv, Aixuan Li, Yiran Zhong, and Yuchao Dai. Exploring depth contribution for camouflaged object detection, 2022. 3", + "[57] Qiong Yan, Li Xu, Jianping Shi, and Jiaya Jia. Hierarchical saliency detection. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 1155-1162, 2013. 2", + "[58] Fan Yang, Qiang Zhai, Xin Li, Rui Huang, Ao Luo, Hong Cheng, and Deng-Ping Fan. Uncertainty-guided transformer reasoning for camouflaged object detection. In The IEEE/CVF International Conference on Computer Vision, pages 4146-4155, 2021. 6", + "[59] Siyuan Yao, Hao Sun, Tian-Zhu Xiang, Xiao Wang, and Xiaochun Cao. Hierarchical graph interaction transformer with dynamic token clustering for camouflaged object detection. IEEE Transactions on Image Processing, 2024. 1", + "[60] Bowen Yin, Xuying Zhang, Deng-Ping Fan, Shaohui Jiao, Ming-Ming Cheng, Luc Van Gool, and Qibin Hou. Camoformer: Masked separable attention for camouflaged object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1", + "[61] Qiang Zhai, Xin Li, Fan Yang, Chenglizhao Chen, Hong Cheng, and Deng-Ping Fan. Mutual graph learning for camouflaged object detection. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 12997-13007, 2021. 6", + "[62] Jing Zhang, Tong Zhang, Yuchao Dai, Mehrtash Harandi, and Richard Hartley. Deep unsupervised saliency detection: A multiple noisy labeling perspective. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 9029-9038, 2018. 2", + "[63] Yi Zhang and Chengyi Wu. Unsupervised camouflaged object segmentation as domain adaptation. In The IEEE/CVF" + ], + "bbox": [ + 91, + 90, + 480, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "International Conference on Computer Vision, pages 4334-4344, 2023. 1, 3, 4, 6, 7, 8", + "[64] Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, and Nicu Sebe. Bilateral reference for high-resolution dichotomous image segmentation. CAAI Artificial Intelligence Research, 3:9150038, 2024. 6", + "[65] Yijie Zhong, Bo Li, Lv Tang, Senyun Kuang, Shuang Wu, and Shouhong Ding. Detecting camouflaged object in frequency domain. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 4494-4503, 2022. 3", + "[66] Hongwei Zhu, Peng Li, Haoran Xie, Xu Yan, Dong Liang, Dapeng Chen, Mingqiang Wei, and Jing Qin. I can find you! boundary-guided separated attention network for camouflaged object detection. In Association for the Advancement of Artificial Intelligence, 2022. 3", + "[67] Jinchao Zhu, Xiaoyu Zhang, Shuo Zhang, and Junnan Liu. Inferring camouflaged objects by texture-aware interactive guidance network. In Association for the Advancement of Artificial Intelligence, 2021. 3", + "[68] Wangjiang Zhu, Shuang Liang, Yichen Wei, and Jian Sun. Saliency optimization from robust background detection. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 2814-2821, 2014. 2", + "[69] Mingchen Zhuge, Xiankai Lu, Yiyou Guo, Zhihua Cai, and Shuhan Chen. Cubenet: X-shape connection for camouflaged object detection. Pattern Recognition, 127:108644, 2022. 3" + ], + "bbox": [ + 516, + 92, + 906, + 472 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "30375", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/UCOD-DPL_ Unsupervised Camouflaged Object Detection via Dynamic Pseudo-label Learning/dec1b26d-4dd9-4c06-a461-c03d0e31f356_model.json b/2025/UCOD-DPL_ Unsupervised Camouflaged Object Detection via Dynamic Pseudo-label Learning/dec1b26d-4dd9-4c06-a461-c03d0e31f356_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7fd7e37be390b69c8984a9acdcbdb26c1a417e3b --- /dev/null +++ b/2025/UCOD-DPL_ Unsupervised Camouflaged Object Detection via Dynamic Pseudo-label Learning/dec1b26d-4dd9-4c06-a461-c03d0e31f356_model.json @@ -0,0 +1,2719 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.205, + 0.131, + 0.795, + 0.177 + ], + "angle": 0, + "content": "UCOD-DPL: Unsupervised Camouflaged Object Detection via Dynamic Pseudo-label Learning" + }, + { + "type": "text", + "bbox": [ + 0.208, + 0.205, + 0.788, + 0.222 + ], + "angle": 0, + "content": "Weiqi Yan\\(^{1}\\), Lvhai Chen\\(^{1}\\), Huaijia Ko\\(^{1}\\), Shengchuan Zhang\\(^{1*}\\), Yan Zhang\\(^{1}\\), Liujuan Cao\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.244, + 0.223, + 0.754, + 0.239 + ], + "angle": 0, + "content": "1 Key Laboratory of Multimedia Trusted Perception and Efficient Computing," + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.241, + 0.737, + 0.256 + ], + "angle": 0, + "content": "Ministry of Education of China, Xiamen University, 361005, P.R. China." + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.26, + 0.78, + 0.274 + ], + "angle": 0, + "content": "weiqi_yan@outlook.com lvhaichen2002@gmail.com zsc_2016@xmu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.353, + 0.485, + 0.807 + ], + "angle": 0, + "content": "Unsupervised Camouflaged Object Detection (UCOD) has gained attention since it doesn't need to rely on extensive pixel-level labels. Existing UCOD methods typically generate pseudo-labels using fixed strategies and train \\(1 \\times 1\\) convolutional layers as a simple decoder, leading to low performance compared to fully-supervised methods. We emphasize two drawbacks in these approaches: 1). The model is prone to fitting incorrect knowledge due to the pseudolabel containing substantial noise. 2). The simple decoder fails to capture and learn the semantic features of camouflaged objects, especially for small-sized objects, due to the low-resolution pseudo-labels and severe confusion between foreground and background pixels. To this end, we propose a UCOD method with a teacher-student framework via Dynamic Pseudo-label Learning called UCOD-DPL, which contains an Adaptive Pseudo-label Module (APM), a Dual-Branch Adversarial (DBA) decoder, and a Look-Twice mechanism. The APM module adaptively combines pseudo-labels generated by fixed strategies and the teacher model to prevent the model from overfitting incorrect knowledge while preserving the ability for self-correction; the DBA decoder takes adversarial learning of different segmentation objectives, guides the model to overcome the foreground-background confusion of camouflaged objects, and the Look-Twice mechanism mimics the human tendency to zoom in on camouflaged objects and performs secondary refinement on small-sized objects. Extensive experiments show that our method demonstrates outstanding performance, even surpassing some existing fully supervised methods. The code is available now1." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.31, + 0.907, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.529, + 0.907, + 0.556 + ], + "angle": 0, + "content": "Figure 1. Comparison between our method and previous pseudolabel-based methods." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.585, + 0.645, + 0.601 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.612, + 0.907, + 0.718 + ], + "angle": 0, + "content": "\"Camouflage\" originates from the natural behavior of objects blending into their surroundings using similar textures and colors to evade \"predators\"[4]. Camouflaged Object Detection (COD) is a challenging semantic segmentation task aimed at learning to segment objects that are visually hidden in their backgrounds, which has significant applications in several important fields [11, 50]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Although fully-supervised methods have achieved significant progress in the COD tasks [19, 38, 49, 59, 60], they heavily rely on large-scale human annotations for training. However, it is difficult to get pixel-level human labels for camouflaged objects due to their complex visual attributes [27]. Therefore, this reliance on annotations has sparked increasing interest in the task of unsupervised camouflage object detection (UCOD) among researchers. However, existing UCOD methods typically use pseudo-labels generated by fixed strategies to train a single \\(1 \\times 1\\) convolutional layer [44, 63]. By investing in existing methods, we identified two main issues: 1). The pseudo-label con" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.235, + 0.888 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.419, + 0.9 + ], + "angle": 0, + "content": "1https://github.com/Heartfirey/UCOD-DPL" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.419, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30365" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.124, + 0.119, + 0.148 + ], + "angle": 0, + "content": "1" + }, + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.092, + 0.212, + 0.102 + ], + "angle": 0, + "content": "Small-sized" + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.107, + 0.213, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.231, + 0.092, + 0.288, + 0.103 + ], + "angle": 0, + "content": "Huge-sized" + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.107, + 0.303, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.329, + 0.092, + 0.371, + 0.103 + ], + "angle": 0, + "content": "Multiple" + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.107, + 0.392, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.41, + 0.092, + 0.471, + 0.103 + ], + "angle": 0, + "content": "Uncertainty" + }, + { + "type": "image", + "bbox": [ + 0.398, + 0.106, + 0.481, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.224, + 0.121, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.291, + 0.121, + 0.328 + ], + "angle": 0, + "content": "CocosDg" + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.165, + 0.213, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.223, + 0.213, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.281, + 0.213, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.165, + 0.303, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.222, + 0.303, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.281, + 0.303, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.165, + 0.392, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.222, + 0.391, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.281, + 0.391, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.398, + 0.165, + 0.481, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.398, + 0.222, + 0.481, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.398, + 0.281, + 0.481, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.349, + 0.483, + 0.432 + ], + "angle": 0, + "content": "Figure 2. Some examples of low-quality pseudo-labels generated using fixed strategies. Based on features extracted by DINOv2, we construct pseudo-labels using the Background-Seed from FOUND [44] and the MaskCut in UCOD-DA [44], in challenging scenarios (e.g. small-sized objects, huge-sized objects, multiple instances, and uncertain objects) for visualization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.483, + 0.597 + ], + "angle": 0, + "content": "structed by fixed strategies contains substantial noise, which makes the model prone to fitting incorrect knowledge. 2). These pseudo-labels have low resolution and severe confusion between foreground and background pixels, and simple \\(1 \\times 1\\) convolution fails to capture and learn the semantic features of camouflaged objects, especially for small-sized objects. Consequently, their segmentation results fall far behind those of fully supervised methods, which limits their application in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a brand-new model called UCOD-DPL. As shown in Fig. 1, we introduce a teacher-student framework consisting of an Adaptive Pseudo-label Mixing (APM) module, a Dual-Branch Adversarial (DBA) decoder, and a Look-Twice refinement strategy. As illustrated in Fig. 2, pseudo-labels constructed by fixed strategies are generally low-quality and contain substantial noise. To prevent the model from fitting uncorrectable erroneous knowledge in the pseudo-labels, we design the APM module with a discriminator and a dynamic scoring mechanism to mix the pseudo-labels from the fixed strategy and the teacher branch adaptively. To help the model overcome foreground-background confusion, we propose a DBA decoder inspired by predator-prey dynamics. This decoder uses different segmentation targets (foreground and background) along with an orthogonal loss of features to deeply mine and learn distinctive features of the foreground and background. Additionally, inspired by [37, 38], we mimic the zooming-in behavior humans use when examining camouflaged images by performing secondary refinement on" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "small-sized camouflaged objects to enhance the model's segmentation accuracy when dealing with small targets." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.124, + 0.905, + 0.198 + ], + "angle": 0, + "content": "In the experiment, we train our model on the commonly used COD training set and evaluate it on multiple test sets. The results demonstrate that our method achieves outstanding performance, even comparable to some fully supervised COD methods when using DINOv2 as the backbone." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.2, + 0.872, + 0.214 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.216, + 0.905, + 0.277 + ], + "angle": 0, + "content": "1. We propose a novel UCOD model called UCOD-DPL, and extensive experiments demonstrate that our method exhibits superior performance, reaching the performance level of some fully supervised algorithms." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.277, + 0.905, + 0.336 + ], + "angle": 0, + "content": "2. We introduce a teacher-student framework with an APM module. By using the APM, we ensure that the model can learn knowledge from pseudo-labels while preserving the ability to identify and correct the erroneous." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.337, + 0.905, + 0.412 + ], + "angle": 0, + "content": "3. To address the issues of low-resolution pseudo-labels and foreground-background confusion, we design a DBA decoder that uses a foreground-background adversarial task, enabling the model to learn the camouflaged features of the foreground adaptively." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.413, + 0.905, + 0.488 + ], + "angle": 0, + "content": "4. To enhance the model's ability to segment small-sized objects, we design a Look-Twice refinement mechanism. This mechanism mimics human behavior by performing a secondary segmentation on smaller camouflaged objects, thereby refining the segmentation results." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.216, + 0.905, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.505, + 0.663, + 0.52 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.531, + 0.839, + 0.547 + ], + "angle": 0, + "content": "2.1. Unsupervised Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Unsupervised Semantic Segmentation (USS) aims to train a model without any human annotations to produce the semantic segmentation mask. Before the surge of deep learning methods, researchers relied on numerous handcrafted methods [7, 16, 57, 68] that used one or more priors related to foreground regions in the image to generate segmentation masks, such as center[26], contrast[24], and boundary[54]. In deep learning-based methods, some approaches utilize generative models [17] to produce a segmentation mask, which will be used to synthesize a realistic image by copying the corresponding region into a background which is either synthesized [3] or taken from a different image [1]. Another category of methods leverages pseudo-labels constructed by handcrafted fixed strategies to train the detector. [62] adopts diverse noisy pseudo-labels via distinct unsupervised handcrafted methods to train the detector to predict a saliency map free from noise in labels. And [34, 44] proposed to refine the pseudo-labels produced by fixed-strategy by training the model via a self-supervised iterative refinement manner. These methods adopt a self-supervision manner that highly relies on pseudo-labels generated directly from one or multiple fixed strategies. If the pseudo-labels contain substantial noise, these methods remain prone to" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30366" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.123 + ], + "angle": 0, + "content": "fitting incorrect knowledge, which is particularly evident when dealing with camouflaged objects." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.134, + 0.364, + 0.15 + ], + "angle": 0, + "content": "2.2. Camouflaged Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.157, + 0.485, + 0.519 + ], + "angle": 0, + "content": "Camouflaged Object Detection (COD) is a complex task aimed at segmenting objects that are intentionally designed to merge seamlessly with their environment. This task has an extensive history in its domain. Existing fully-supervised COD methods [10, 37, 38, 69] have achieved excellent performance by training on large-scale annotated datasets. These methods utilize the rich information (e.g. boundary [25, 45, 66], texture [25, 42, 67] and other information like frequency domain and depth [28, 52, 56, 65]) provided by pixel-level annotations and design a series of sophisticated strategies to extract unique visual characteristics of camouflaged objects. However, the extensive pixel-level annotations are hardly gained, which has drawn the attention of researchers to Unsupervised Camouflaged Object Detection (UCOD) and USS methods [63]. These methods typically rely on pseudo-labels generated through fixed strategies to supervise model learning for segmentation. However, due to limitations in algorithmic complexity and robustness, these fixed strategies struggle to work effectively at high resolutions and easily introduce significant noise. Consequently, these methods are prone to learning incorrect noise from the pseudo-labels, leading to suboptimal segmentation performance and limiting the application in the real world of UCOD models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.531, + 0.321, + 0.547 + ], + "angle": 0, + "content": "2.3. Self-supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Self-Supervised Learning (SSL) has revolutionized the way models leverage unlabeled data. Among current SSL approaches, some methods adopt contrastive learning to learn image feature representations. MoCo [20] utilizes a dynamic memory bank to maintain a large pool of negative samples; SimCLR [6] employs various data augmentations and a projection head to maximize similarities between differently augmented views of the same instance; and BYOL [18] constructs a teacher-student framework with the teacher model's parameters are updated by Exponential Moving Average (EMA) of the student model. Following BYOL, DINO [5] applied two interactive encoders sharing the same structure but with different parameter sets and update strategies. Inspired by self-supervised pretraining techniques in natural language processing [40, 41, 46], other SSL methods design pretext tasks (e.g. colorization, rotation prediction, and patch reconstruction) on unlabeled images [5, 6, 14, 15, 20, 29, 33, 35]. BeiT [2] adopts a Transformer-based architecture to reconstruct masked image patches, MAE [21] implements a masked autoencoder that efficiently reconstructs the missing parts of an image. In this paper, we employ DINOv1 [5] and DINOv2 [36] to extract robust and generalized image features." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.65, + 0.108 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.115, + 0.704, + 0.13 + ], + "angle": 0, + "content": "3.1. Overall Framework" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.137, + 0.907, + 0.226 + ], + "angle": 0, + "content": "The structure of our proposed method is shown in Fig. 3, which includes a teacher-student framework with an Adaptive Pseudo-label Mixing (APM) module, a Dual-Branch Adversarial (DBA) decoder, and a Look-Twice strategy. Both the teacher and student models are configured as DBA decoders." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.228, + 0.908, + 0.488 + ], + "angle": 0, + "content": "Firstly, the \\(i\\)-th unlabeled input image passes through the self-supervised pretrained backbone to obtain high-level semantic image feature \\(F_{i}\\). We then employ the background seed method to generate a low-resolution fixed-strategy pseudo-label \\(\\hat{P}_i^{fs}\\), which utilizes pixel-level similarity comparison and was proposed in [44]. Simultaneously, the image feature \\(F_{i}\\) is fed into a teacher-student model. The teacher model generates a higher-resolution teacher's pseudo label \\(\\hat{P}_i^t\\), and the student model produces the foreground and background predicted mask \\(\\hat{Y}_i^{FG}, \\hat{Y}_i^{BG}\\). Then, \\(\\hat{P}_i^{fs}\\) and \\(\\hat{P}_i^t\\) are fed into the APM module to construct the mixed dynamic pseudo-label \\(P_{i}\\). Finally, we use the mixed dynamic pseudo-label \\(P_{i}\\) to supervise the final foreground and background predicted masks \\(\\hat{Y}_i^{FG}, \\hat{Y}_i^{BG}\\). We further apply a Look-Twice strategy to perform a secondary refinement on small-sized objects to improve the model's segmentation performance on small-sized samples." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.495, + 0.863, + 0.511 + ], + "angle": 0, + "content": "3.2. Adaptive Pseudo-Label Merging Module" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.517, + 0.906, + 0.638 + ], + "angle": 0, + "content": "Existing unsupervised semantic segmentation methods typically adopt a fixed-strategy (e.g. similarity comparison [32, 43, 53]) or utilize a dual-branch model (e.g. fixed-strategy or learnable branch [44, 63]) structure for self-supervised learning to generate segmentation masks. However, we believe the performance of these methods is constrained by the strategy's design or the pseudo-labels' quality, preventing the model from achieving higher accuracy." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.639, + 0.906, + 0.713 + ], + "angle": 0, + "content": "We introduce a discriminator \\(\\mathcal{D}\\), which inputs a foreground segmentation mask (i.e. fixed-strategy pseudo-label \\(\\hat{P}_i^{fs}\\) or student's foreground predicted mask \\(\\hat{Y}_i^{FG}\\)) and outputs the probability that the mask originates from the fixed-strategy branch:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.72, + 0.905, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\hat {y} _ {i} ^ {p 1} = \\mathcal {D} \\left(\\hat {P} _ {i} ^ {f s}\\right), \\hat {y} _ {i} ^ {p 2} = \\mathcal {D} \\left(\\hat {Y} _ {i} ^ {F G}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.748, + 0.907, + 0.795 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{y}}_i^{p1},\\hat{\\mathbf{y}}_i^{p2}\\) denote the predicted probability that the input mask belongs to the fixed-strategy branch. We then introduce a scoring function \\(S(\\cdot)\\) with a temporal constraint:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.807, + 0.918, + 0.855 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} \\left(\\hat {\\mathrm {y}} _ {i} ^ {p 1}, \\hat {\\mathrm {y}} _ {i} ^ {p 2}\\right) = \\operatorname {C L I P} \\left(\\frac {t}{T} + \\frac {1}{2} \\left(1 + \\cos \\left(\\pi \\times \\left| \\hat {\\mathrm {y}} _ {i} ^ {p 1} - \\hat {\\mathrm {y}} _ {i} ^ {p 2} \\right|\\right)\\right) _ {0} ^ {1}, \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\( t, T \\) denote the current and total training epochs, \\( \\mathrm{CLIP}(x)_0^1 \\) denotes truncating \\( x \\) so that it lies within the interval [0, 1]. In the early stage of training, the teacher model" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30367" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.904, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.298, + 0.906, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.47, + 0.908, + 0.498 + ], + "angle": 0, + "content": "Figure 3. The main framework of our proposed method. The model contains a teacher-student framework, an Adaptive Psdudo-label Mixing (APM) module, a Dual-Branch Adversarial (DBA) decoder, and a Look-Twice strategy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.485, + 0.761 + ], + "angle": 0, + "content": "has not yet developed sufficient localization and basic semantic segmentation capabilities for camouflaged objects, making it unable to supervise the student model effectively. At this point, there is a significant difference between fixed-strategy pseudo-label \\(\\hat{P}_i^{fs}\\) and the student's foreground predicted mask \\(\\hat{Y}_i^{FG}\\). To help the model learn the fundamental foreground localization and segmentation tasks, we use a higher proportion of \\(\\hat{P}_i^{fs}\\) to supervise model training. Later in the training process, as the \\(\\hat{Y}_i^{FG}\\) gradually approaches \\(\\hat{P}_i^{fs}\\), we increase the proportion of the teacher's pseudo-label \\(\\hat{P}_i^t\\) as it can make more stable predictions. This adjustment helps prevent the model from overfitting to noise and incorrect knowledge in the \\(\\hat{P}_i^{fs}\\). We use the obtained score as the mixing weight \\(W_i^t\\) to combine the fixed-strategy pseudo-label \\(\\hat{P}_i^{fs}\\) and teacher's predictions \\(\\hat{P}_i^t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.77, + 0.483, + 0.797 + ], + "angle": 0, + "content": "\\[\n\\mathrm {W} _ {i} ^ {t} = \\mathcal {S} \\left(\\hat {\\mathrm {y}} _ {i} ^ {p 1}, \\hat {\\mathrm {y}} _ {i} ^ {p 2}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.794, + 0.388, + 0.812 + ], + "angle": 0, + "content": "\\[\nP _ {i} = \\mathrm {W} _ {i} ^ {t} \\hat {P} _ {i} ^ {t} + (1 - \\mathrm {W} _ {i} ^ {t}) \\hat {P} _ {i} ^ {f s},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.822, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(P_{i}\\) denotes the mixed dynamic pseudo-label. For the training of the discriminator, we set a binary classification task by using \\(\\hat{P}_i^{fs}\\) and \\(\\hat{Y}_i^{FG}\\) as the input and assign the discriminator label \\(\\mathrm{y}_i^{p1} = 1\\) for the fixed-strategy pseudo-label \\(\\hat{P}_i^{fs}\\) and \\(\\mathrm{y}_i^{p2} = 0\\) for the student model branch \\(\\hat{Y}_i^{FG}\\). Then," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.525, + 0.907, + 0.553 + ], + "angle": 0, + "content": "we employ a cross-entropy supervision loss \\(\\mathcal{L}_{\\mathrm{BCE}}\\) to train the discriminator:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.56, + 0.907, + 0.579 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d i s} = \\mathcal {L} _ {\\mathrm {B C E}} \\left(\\hat {\\mathrm {y}} _ {i} ^ {p j}, \\mathrm {y} _ {i} ^ {p j}\\right), j \\in \\{1, 2 \\}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.629 + ], + "angle": 0, + "content": "During the training, we treat the student model as a kind of generator and alternately train the decoder and the discriminator in a GAN-like manner." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.637, + 0.816, + 0.652 + ], + "angle": 0, + "content": "3.3. Dual-Branch Adversarial Decoder" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.781 + ], + "angle": 0, + "content": "In previous methods [44, 63], predictions are typically generated by a simple \\(1 \\times 1\\) convolution, which limits the ability to capture camouflage features thoroughly. We propose a Dual-Branch Adversarial decoder to enhance accuracy and robustness in this task. The DBA decoder consists of two parallel branches: one is responsible for segmenting the foreground predicted mask \\(\\hat{Y}_i^{FG}\\), while the other is for segmenting the background predicted mask \\(\\hat{Y}_i^{BG}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.907, + 0.843 + ], + "angle": 0, + "content": "First, for \\(i\\)-th image feature \\(F_{i}\\) extracted from the backbone model, a convolutional layer is used to double its channel, then we decouple it into foreground and background features \\(F_{i}^{FG}, F_{i}^{BG}\\) with the same shape of \\(F_{i}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.846, + 0.905, + 0.865 + ], + "angle": 0, + "content": "\\[\nF _ {i} ^ {F G}, F _ {i} ^ {B G} = \\operatorname {S p l i t} \\left(\\operatorname {C o n v} \\left(F _ {i}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "where \\( \\text{Split}(\\cdot) \\) denotes splitting a tensor along the channel dimension equally. We then employ two learnable em" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30368" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.138 + ], + "angle": 0, + "content": "beddings \\(\\mathrm{E}_{FG},\\mathrm{E}_{BG}\\) to store foreground and backgroundrelated knowledge. These embeddings are used to calculate foreground and background attention queries \\(Q_{i}^{FG^{\\prime}}\\) \\(Q_{i}^{BG^{\\prime}}\\) .." + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.149, + 0.483, + 0.169 + ], + "angle": 0, + "content": "\\[\nQ _ {i} ^ {F G ^ {\\prime}} = F _ {i} ^ {F G} \\times \\mathrm {E} _ {F G}, Q _ {i} ^ {B G ^ {\\prime}} = F _ {i} ^ {B G} \\times \\mathrm {E} _ {B G}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.174, + 0.484, + 0.253 + ], + "angle": 0, + "content": "Finally, we apply multiplicative attention with a residual connection to weigh the features, enhancing important details. The resulting features are then passed through a \\(1 \\times 1\\) convolution to output the foreground predicted mask \\(\\hat{Y}_i^{FG}\\) and background mask \\(\\hat{Y}_i^{BG}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.261, + 0.483, + 0.289 + ], + "angle": 0, + "content": "\\[\n\\hat {Y} _ {i} ^ {F G} = \\operatorname {C o n v} ^ {\\prime} \\left(\\operatorname {S i g m o i d} \\left(Q _ {i} ^ {F G ^ {\\prime}} \\times F _ {i} ^ {F G}\\right) + F _ {i} ^ {F G}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.284, + 0.445, + 0.304 + ], + "angle": 0, + "content": "\\[\n\\hat {Y} _ {i} ^ {B G} = \\operatorname {C o n v} ^ {\\prime} (\\operatorname {S i g m o i d} \\left(Q _ {i} ^ {B G ^ {\\prime}} \\times F _ {i} ^ {B G}\\right) + F _ {i} ^ {B G}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.315, + 0.484, + 0.391 + ], + "angle": 0, + "content": "For the student model, we use \\( Y_{i}^{FG}, \\hat{Y}_{i}^{BG} \\) as the student's foreground and background prediction masks. For the teacher model, we further generate the teacher's pseudolabel \\( P_{i} \\) using the combination of the foreground and background prediction masks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.392, + 0.484, + 0.437 + ], + "angle": 0, + "content": "Inspired by [55], we apply an orthogonal loss \\(\\mathcal{L}_{\\perp}\\) for \\(Q_{i}^{FG^{\\prime}}\\in \\mathbb{R}^{(n\\times m)\\times c},Q_{i}^{BG^{\\prime}}\\in \\mathbb{R}^{(n\\times m)\\times c}\\) as a regularisation constraint to encourage them to focus on distinct features:" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.447, + 0.362, + 0.467 + ], + "angle": 0, + "content": "\\[\n\\mathbb {M} = Q _ {i} ^ {F G ^ {\\prime}} \\left(Q _ {i} ^ {B G ^ {\\prime}}\\right) ^ {\\top},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.471, + 0.483, + 0.512 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\perp} = \\frac {1}{(n \\times m) ^ {2}} \\sum_ {j = 1} ^ {n \\times m} \\sum_ {k = 1} ^ {n \\times m} (1 - \\delta_ {j k}) (\\mathbb {M} [ j, k ]) ^ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.484, + 0.6 + ], + "angle": 0, + "content": "where \\(n, m, c\\) indicate the height, width, and number of channels. When \\(j = k\\), \\(\\delta_{jk} = 1\\); otherwise \\(\\delta_{jk} = 0\\). For the predictions \\(\\hat{Y}_i^{FG}, \\hat{Y}_i^{BG}\\), we use a binary cross-entropy loss \\(\\mathcal{L}_{\\mathrm{BCE}}\\) for supervision, where the background prediction is inverted before applying the supervision:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.609, + 0.484, + 0.636 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s e g} = \\mathcal {L} _ {\\mathrm {B C E}} (\\hat {Y} _ {i} ^ {F G}, P _ {i}) + \\mathcal {L} _ {\\mathrm {B C E}} \\left((1 - \\hat {Y} _ {i} ^ {B G}), P _ {i}\\right). (9)\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.644, + 0.312, + 0.658 + ], + "angle": 0, + "content": "3.4. Look-Twice Refinement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.665, + 0.484, + 0.833 + ], + "angle": 0, + "content": "The segmentation quality of COD is often influenced by object size. Due to reduced spatial resolution, the model struggles to produce fine segmentation for small-sized camouflaged objects. Inspired by previous works [37, 38], we attempt to mimic human behavior when observing small-sized objects: first coarsely locating the camouflaged object in the image, then zooming in to carefully retrieve its details. To this end, we designed the Look-Twice mechanism. First, the input image is processed by the model to obtain a coarse segmentation result \\(\\hat{Y}_i^{FG}\\). We define the set of all foreground objects as \\(\\mathrm{C}^{FG}\\), which can be calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.841, + 0.483, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\mathrm {C} ^ {\\mathrm {F G}} = \\text {C o n n e c t C o m p o n e n t} \\left(\\hat {Y} _ {i} ^ {F G}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where ConnectComponent(\\cdot) denotes the connected components labelling algorithm. Next, for each connect" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.09, + 0.906, + 0.123 + ], + "angle": 0, + "content": "component \\(c_k^{FG} \\in \\mathbb{C}^{FG}\\), we calculate its foreground area ratio \\(r_k^{FG}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.129, + 0.905, + 0.165 + ], + "angle": 0, + "content": "\\[\nr _ {k} ^ {F G} = \\frac {\\sum_ {(x , y) \\in \\mathrm {c} _ {k} ^ {F G}} \\hat {Y} _ {i} ^ {F G} (x , y)}{H \\times W}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.169, + 0.906, + 0.258 + ], + "angle": 0, + "content": "where \\(H, W\\) denote the height and width of the foreground area of the output mask \\(\\tilde{Y}_i^{FG}\\). We consider all regions where \\(r_i^{FG} < \\tau\\) as small-sized object areas. We set \\(\\tau = 0.15\\) when using DINOv2 as the backbone. The hyperparameter ablation study on \\(\\tau\\) will be provided in the experiments section." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.259, + 0.906, + 0.351 + ], + "angle": 0, + "content": "For each cropped small-sized object region, we aim to retain a certain proportion of the background to provide the model with sufficient context to segment the foreground from the background. Therefore, for each connected component region \\( c_k^{FG} \\), we calculate the local foreground area ratio \\( s_k^{FG} \\) and the global area ratio \\( s_k^{BG} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.356, + 0.905, + 0.399 + ], + "angle": 0, + "content": "\\[\ns _ {k} ^ {F G} = \\frac {\\sum_ {(x , y) \\in c _ {k} ^ {F G}} \\hat {Y} _ {i} ^ {F G} (x , y)}{H _ {k} ^ {F G} \\times W _ {k} ^ {F G}}, \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.397, + 0.761, + 0.43 + ], + "angle": 0, + "content": "\\[\ns _ {k} ^ {B G} = \\frac {H _ {k} ^ {F G} \\times W _ {k} ^ {F G}}{H \\times W},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.434, + 0.906, + 0.467 + ], + "angle": 0, + "content": "where \\(H_{k}^{FG},W_{k}^{FG}\\) denote the height and width of the bounding box of the \\(k\\) -th connected component \\(\\mathrm{c}_k^{FG}\\) .." + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.471, + 0.905, + 0.491 + ], + "angle": 0, + "content": "\\[\n\\operatorname {s c a l e} _ {k} ^ {F G} = 1 - \\left(s _ {k} ^ {F G} / s _ {k} ^ {B G}\\right). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.496, + 0.906, + 0.707 + ], + "angle": 0, + "content": "Finally, we calculate the expansion ratio using Eq. (13) and enlarge the bounding box of the connected component region by this ratio. The expanded bounding box is then used to crop the related area of the input image, and the cropped patches are resized to the original input size. During the training process, these patches are treated as augmented data. At test time, we re-infer these patches, scale the resulting foreground predicted mask back to the original size, and paste it into the coarse mask as the refined result. However, we found that this mechanism still has some weakness: for cases where objects are occluded and the annotations are fragmented, this method may mistakenly amplify a part of the object as a small object. This can serve as a point of our future work." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.714, + 0.63, + 0.729 + ], + "angle": 0, + "content": "3.5. Total Loss" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.736, + 0.86, + 0.752 + ], + "angle": 0, + "content": "Finally, the total loss \\(\\mathcal{L}_{tot}\\) can be defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.758, + 0.905, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t o t} = \\mathcal {L} _ {s e g} + \\mathcal {L} _ {\\perp} + \\mathcal {L} _ {d i s}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{seg}\\) is the segmentation loss defined by Eq. (9), \\(\\mathcal{L}_{\\perp}\\) is the orthogonal loss defined by Eq. (8), and \\(\\mathcal{L}_{dis}\\) is the discriminator loss defined by Eq. (4). When training on the teacher-student framework, we freeze the parameters of the discriminator and use \\(\\mathcal{L}_{seg}\\) to supervise the model training. Conversely, when training the discriminator, we freeze the parameters of the other parts of the model and train the discriminator only using \\(\\mathcal{L}_{dis}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30369" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.907, + 0.468 + ], + "angle": 0, + "content": "
MethodsCHAMELEON (87)CAMO-Test (250)COD10K-Test (2,026)NC4K (4,121)
Sm↑FωβFmβEmφM↓Sm↑FωβFmβEmφM↓Sm↑FωβFmβEmφM↓Sm↑FωβFmβEmφM↓
Fully-Supervised Methods
SINet20[10].872.806.827.946.034.751.606.675.771.100.771.551.634.806.051.808.723.769.871.058
C2FNet21[47].888.828.844.946.032.796.719.762.864.080.813.686.723.900.036.838.762.794.904.049
MGL-R21[61].893.812.834.941.030.775.673.726.842.088.814.666.710.890.035.833.739.782.893.053
UGTR21[58].887.794.819.940.031.784.684.735.851.086.817.666.711.890.036.839.746.787.899.052
BGNet22[48].901.851.860.954.027.812.749.789.870.073.831.722.753.901.033.851.788.820.907.044
ZoomNet22[37].902.845.864.958.023.820.752.794.878.066.838.729.766.888.029.853.784.818.896.043
SINetv22[12].888.816.835.961.030.820.743.782.882.070.815.680.718.887.037.847.770.805.903.048
HitNet23[22].921.897.900.972.019.849.809.831.906.055.871.806.823.935.023.875.834.853.926.037
FSPNet23[23].908.851.867.965.023.856.799.830.899.050.851.735.769.895.026.879.816.843.915.035
BiRefNet24[64].929.911.922.968.016.932.914.922.974.015.913.874.888.960.014.914.894.909.953.023
Semi-Supervised Methods
CamoTeacher24(1%)[27].652.472.558.714.093.621.456.545.669.136.699.517.788.797.062.718.599.779.814.090
CamoTeacher24(5%)[27].729.587.656.785.070.669.523.601.711.122.745.583.827.840.050.777.677.834.859.071
CamoTeacher24(10%)[27].756.617.684.813.065.701.560.742.795.112.759.594.836.854.049.791.687.842.868.068
SCOD-ND24(10%)[13].850.773-.928.036.789.732-.859.077.819.725-.891.033.838.787-.903.046
Unsupervised Methods
BigGW21[51].547.244.294.527.257.565.299.349.528.282.528.185.246.497.261.608.319.391.565.246
TokenCut22[53].654.496.536.740.132.633.498.543.706.163.658.469.502.735.103.725.615.649.802.101
TokenCut22 w/B.S.[53].655.351.393.582.169.639.383.434.595.195.666.334.399.609.127.735.478.547.683.133
SpectralSeg22[32].575.410.440.628.220.579.450.481.648.235.575.360.388.595.193.669.535.562.719.159
SelfMask22[43].619.436.481.675.176.617.483.536.698.176.637.431.469.679.131.716.593.634.777.114
SelfMask22 w/U.B.[43].629.447.491.683.169.627.495.547.708.182.645.440.478.687.125.723.601.642.784.110
FOUND22-DINOv1[44].684.542.590.810.095.685.584.633.782.129.670.482.520.751.085.741.637.674.824.084
*FOUND22-DINOv2[44].829.757.781.911.040.770.704.740.849.090.767.641.668.847.045.816.756.783.893.052
UCOS-DA23-DINOv1[63].715.591.629.802.095.701.606.646.784.127.689.513.546.740.086.755.656.689.819.085
*UCOS-DA23-DINOv2[63].750.639.666.808.091.702.604.633.751.148.655.467.495.687.120.731.617.644.785.103
OursDINOv1.734.625.680.854.072.706.621.689.801.108.727.577.627.822.059.761.680.737.851.074
OursDINOv2.864.825.838.931.031.793.747.779.862.077.834.763.779.916.031.850.818.835.923.043
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.474, + 0.907, + 0.517 + ], + "angle": 0, + "content": "Table 1. Comparison of our methods with recent methods. We compared our proposed methods with competing unsupervised, semi-supervised, and full-supervised methods. Bold indicates the best result in group settings, and underline indicates the second-best result. * denotes the version that reimplemented by us." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.542, + 0.216, + 0.559 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.568, + 0.299, + 0.586 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.593, + 0.483, + 0.684 + ], + "angle": 0, + "content": "Trainsets. To fairly compare with the existing works, following [10, 30], we used a combination of 1000 images from the CAMO-Train and 3040 images from COD10K-Train as the training dataset of our experiments. Following the unsupervised learning settings, we do not use any ground-truth label during the training process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.483, + 0.761 + ], + "angle": 0, + "content": "Testing Sets. We evaluate the performance of our methods on four mainstream COD benchmark testing sets: CHAMELEON with 76 test images, CAMO with 250 test images, COD10K with 2026 test images, and NC4K with 4121 test images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.763, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Evaluation Protocol. For a fair and comprehensive evaluation, we employ the S-measure \\((S_m)\\) [8], mean and weighted F-measure \\((\\mathcal{F}_\\beta^m,\\mathcal{F}_\\beta^\\omega)\\) [31], mean E-measure \\((\\mathcal{E}_{\\epsilon}^{m})\\) [9], mean absolute error \\((\\mathcal{M})\\) [39] as the main metrics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. Following [44], we employ the simple background seed methods as our fixed strategy branch to generate the coarse pseudo-label. And following [44, 63], we use the strong self-supervised pretrained backbone DINO [5, 36] as our encoder. The teacher model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.543, + 0.907, + 0.665 + ], + "angle": 0, + "content": "is updated via EMA with a momentum \\(\\eta\\) of 0.99. We train the model for 25 epochs. The batch size is set to 32 for each GPU during training. All experiments are implemented with PyTorch 2.1 and run on a machine with Intel(R) Xeon(R) Silver 4214R CPU @ 2.40GHz, 512GiB RAM, and 1 NVIDIA Titan A100-40G GPUs. All experiments use the same random seed. More implementation details will be provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.679, + 0.655, + 0.695 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.907, + 0.838 + ], + "angle": 0, + "content": "Qualitative Analysis. We show visualizations of a series of camouflaged object segmentation masks predicted by our method and related methods in some challenging scenarios. As shown in Fig. 4, we noticed that our method achieves higher segmentation accuracy than existing approaches, with clearer edges and a more detailed representation of the camouflaged object. Note that DINOv2 is employed for all methods that require a backbone to extract features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Quantitative Analysis. In Tab. 1, we compared our proposed method's performance with competing USS and UCOD models on four COD test datasets. To further highlight our model's competitiveness, we also compared ex" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "30370" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.094, + 0.898, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.466, + 0.908, + 0.496 + ], + "angle": 0, + "content": "Figure 4. Visual comparison of our method with other existing methods in challenging scenarios. Our method has clearer and more precise segmentation boundaries and correctly recognizes depth-artifacted objects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.484, + 0.643 + ], + "angle": 0, + "content": "isting fully-supervised and semi-supervised methods. The results show that our model outperformed all existing USS and UCOD methods across all metrics and datasets, thus achieving State-Of-The-Art (SOTA) performance. Additionally, based on DINOv2 backbone, our model has surpassed several semi-supervised and fully-supervised methods across all four datasets, demonstrating its superior performance, effectiveness, and robustness." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.652, + 0.245, + 0.669 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.719 + ], + "angle": 0, + "content": "To verify the effectiveness of our methods, we employ \\( DI-NOv2 \\) as the backbone and conduct comprehensive ablation studies on the COD10K-Test dataset with 2,026 images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Module Ablation. We first conduct ablation experiments on the teacher-student framework (Tea-Stu) and our proposed methods: Adaptive Pseudo-label Mixing (APM) module, Dual-Branch Adversarial (DBA) decoder, and Look-Twice strategy. To thoroughly demonstrate their effectiveness, we replaced the decoder with a single \\(1 \\times 1\\) convolutional layer following [44, 63]. For all ablations on the APM module, we used the \\(1:1\\) constant mixing strategy to replace it. The ablation results are shown in Tab. 2. As a result, we find that when only the teacher-student framework is used, the simple \\(1 \\times 1\\) convolution fails to learn camouflage-related knowledge adequately. When training" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.517, + 0.909, + 0.671 + ], + "angle": 0, + "content": "
SettingsCOD10K (2026)
Tea-StuAPMDBALook-TwiceSm↑Fωβ↑Fmβ↑Emφ↑M↓
✓.626.448.464.658.112
✓✓.633.452.472.668.129
✓.607.409.435.631.129
✓✓.603.404.431.625.133
✓✓.724.582.604.800.054
✓✓✓.788.679.700.879.043
✓✓✓.784.675.692.869.036
✓✓✓✓.834.763.779.916.031
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.68, + 0.908, + 0.737 + ], + "angle": 0, + "content": "Table 2. Ablation study to evaluate the proposed modules. We retrained our model with different settings on the same learning rate and epochs, then benchmarked on COD10K-testset to validate the effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.907, + 0.884 + ], + "angle": 0, + "content": "with a DBA decoder only, the model lacks the guidance of the teacher-student framework, thus easily learning the incorrect knowledge from the low-quality pseudo-labels constructed by fixed-strategy. The proposed Look-Twice mechanism is effective only when the network can produce sufficiently high-quality predictions. If the network has poor localization and segmentation capabilities, Look-Twice will degrade performance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Mixing Strategy Ablation. We continue conducting ab-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "30371" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.483, + 0.171 + ], + "angle": 0, + "content": "
Mixing StrategiesCOD10K (2026)
Sm↑Fωβ↑Fmβ↑Eφ↑M↓
Proportional(1:1) Mixing.707.555.572.763.082
Linear Decay Mixing.809.716.723.888.038
APM.834.763.779.916.031
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.18, + 0.483, + 0.222 + ], + "angle": 0, + "content": "Table 3. Ablation to the pseudo-label mixing strategy. We retrained our model with different pseudo-label mixing strategies and benchmarked on the COD10K-Test dataset." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.243, + 0.47, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.398, + 0.483, + 0.441 + ], + "angle": 0, + "content": "Figure 5. Hyper-parameter ablation of small-sized object ratio \\(\\tau\\). The model performs optimally and shows stable results when the value of \\(\\tau\\) is set around 0.15." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.455, + 0.483, + 0.535 + ], + "angle": 0, + "content": "
Fixed Generation StrategiesCOD10K (2026)
Sm↑Fωβ↑Fmβ↑Emφ↑M↓
Null(Pure Black/White).459.002.001.250.091
Random Perlin Noise.667.489.543.788.072
Background Seed.834.763.779.916.031
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.544, + 0.483, + 0.587 + ], + "angle": 0, + "content": "Table 4. Fixed Pseudo-label Generation Strategy Ablation. We retrained our model with a replaced fixed strategy and benchmarked it on the COD10K-Test dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.689 + ], + "angle": 0, + "content": "lation studies on the mixing strategies. We compared the proportional(1:1), linear decay, and APM mixing strategies. As shown in Tab. 3, the APM module provides a method for fusing pseudo-labels that is more aligned with the model's training process and achieving optimal performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.764 + ], + "angle": 0, + "content": "Hyper-parameter Ablation. We conducted ablation study on the threshold \\(\\tau\\) for the small-sized object ratio mentioned in Sec. 3.4. As shown in Fig. 5, when \\(\\tau \\approx 0.15\\), the model achieves optimal and stable performance. Therefore, we set \\(\\tau\\) to 0.15 in the final model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Study on Foreground-sizes. We divided the test set by the proportion of test set prospects at \\(2\\%\\) intervals, then benchmarked the performance between our method and some previous SOTA methods[44, 63]. As shown in Fig. 6, our method has significant advantages when dealing with smaller-sized camouflaged targets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Fixed Pseudo-label Generation Strategy Ablation. We further investigated the effect of the pseudo-labels constructed by the fixed-strategy. For each input image, we" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.714, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.19, + 0.714, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.089, + 0.906, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.19, + 0.906, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.302, + 0.906, + 0.33 + ], + "angle": 0, + "content": "Figure 6. Performance comparison for different foreground sizes on COD10K-Test dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.359, + 0.907, + 0.51 + ], + "angle": 0, + "content": "generated a random shape of Perlin noise on a plain black background. We used this \"fabricated pseudo-label\" as well as the pure black and white masks to replace the fixed-strategy pseudo-label during the training process. The results are shown in Tab. 4. When using random noise, the noise may overlap with the foreground region, allowing the network to explore some knowledge related to the camouflaged foreground. However, when using a solid color mask, the model fails to learn the basic definition of the segmentation task to produce valid predictions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.527, + 0.64, + 0.543 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Considering the existing UCOD methods suffer from main issues: 1). The low-quality pseudo-labels constructed by fixed strategies contain substantial noise, making the model prone to fitting incorrect knowledge. 2). The low-resolution pseudo-labels cause severe confusion between the foreground and background. Simple \\(1 \\times 1\\) convolution fails to capture and learn the semantic features of camouflaged objects, especially for small-sized samples. To solve these problems, we propose UCOD-DPL in this paper. We introduce a teacher-student framework with an Adaptive Pseudolabel Module, a Dual-Branch Adversarial decoder, and a Look-Twice mechanism. The APM dynamically fuses pseudo-labels from the fixed strategy and teacher model, effectively preventing the model from overfitting the noise and erroneous knowledge. The DBA decoder mines foreground and background knowledge relevant to camouflaged objects while resisting incorrect information from pseudolabels. The Look-Twice refinement mechanism enhances camouflaged object learning and refinement of small objects. We conducted extensive benchmarking on multiple COD test datasets, and the results show that our method outperforms all previous approaches and even reaches the performance level of some fully supervised methods." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30372" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.228 + ], + "angle": 0, + "content": "This work was supported by the National Science Fund for Distinguished Young Scholars (No.62025603), the National Natural Science Foundation of China (No. U21B2037, No. U22B2051, No. U23A20383, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No. 2021J06003, No. 2022J06001)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.253, + 0.188, + 0.269 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.278, + 0.484, + 0.319 + ], + "angle": 0, + "content": "[1] Relja Arandjelović and Andrew Zisserman. Object discovery with a copy-pasting gan. arXiv preprint arXiv:1905.11369, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.322, + 0.484, + 0.362 + ], + "angle": 0, + "content": "[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.365, + 0.483, + 0.406 + ], + "angle": 0, + "content": "[3] Adam Bielski and Paolo Favaro. Emergence of object segmentation in perturbed generative models. Advances in Neural Information Processing Systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.408, + 0.482, + 0.435 + ], + "angle": 0, + "content": "[4] TIM Caro. The adaptive significance of coloration in mammals. BioScience, 55(2):125-136, 2005. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.436, + 0.483, + 0.505 + ], + "angle": 0, + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In The IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.507, + 0.483, + 0.562 + ], + "angle": 0, + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.564, + 0.482, + 0.618 + ], + "angle": 0, + "content": "[7] Ming-Ming Cheng, Niloy J Mitra, Xiaolei Huang, Philip HS Torr, and Shi-Min Hu. Global contrast based salient region detection. IEEE transactions on pattern analysis and machine intelligence, 37(3):569-582, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.62, + 0.482, + 0.675 + ], + "angle": 0, + "content": "[8] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In The IEEE/CVF International Conference on Computer Vision, pages 4548-4557, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.677, + 0.482, + 0.745 + ], + "angle": 0, + "content": "[9] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, Ming-Ming Cheng, and Ali Borji. Enhanced-alignment measure for binary foreground map evaluation. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, pages 698-704. AAAI Press, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[10] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 2777-2787, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.873 + ], + "angle": 0, + "content": "[11] Deng-Ping Fan, Ge-Peng Ji, Tao Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, and Ling Shao. Pranet: Parallel reverse attention network for polyp segmentation. In International conference on medical image computing and computer-assisted intervention, pages 263-273. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. Concealed object detection. IEEE Transactions on" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.278, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Pattern Analysis and Machine Intelligence, 44(10):6024-6042, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.124, + 0.905, + 0.165 + ], + "angle": 0, + "content": "[13] Yuanbin Fu, Jie Ying, Houlei Lv, and Xiaojie Guo. Semi-supervised camouflaged object detection from noisy data. In ACM International Conference on Multimedia, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.168, + 0.905, + 0.222 + ], + "angle": 0, + "content": "[14] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.226, + 0.905, + 0.295 + ], + "angle": 0, + "content": "[15] Spyros Gidaris, Andrei Bursuc, Gilles Puy, Nikos Komodakis, Matthieu Cord, and Patrick Pérez. Obow: Online bag-of-visual-words generation for self-supervised learning. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 6830–6840, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.298, + 0.905, + 0.352 + ], + "angle": 0, + "content": "[16] Stas Goferman, Lihi Zelnik-Manor, and Ayellet Tal. Context-aware saliency detection. IEEE transactions on pattern analysis and machine intelligence, 34(10):1915-1926, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.356, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.414, + 0.905, + 0.497 + ], + "angle": 0, + "content": "[18] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.5, + 0.905, + 0.569 + ], + "angle": 0, + "content": "[19] Chunming He, Kai Li, Yachao Zhang, Longxiang Tang, Yu-lun Zhang, Zhenhua Guo, and Xiu Li. Camouflaged object detection with feature decomposition and edge reconstruction. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 22046-22055, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.572, + 0.905, + 0.627 + ], + "angle": 0, + "content": "[20] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 9729-9738, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.63, + 0.905, + 0.697 + ], + "angle": 0, + "content": "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 16000-16009, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.702, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[22] Xiaobin Hu, Shuo Wang, Xuebin Qin, Hang Dai, Wenqi Ren, Donghao Luo, Ying Tai, and Ling Shao. High-resolution iterative feedback network for camouflaged object detection. In Association for the Advancement of Artificial Intelligence, pages 881-889, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.774, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[23] Zhou Huang, Hang Dai, Tian-Zhu Xiang, Shuo Wang, Huai-Xin Chen, Jie Qin, and Huan Xiong. Feature shrinkage pyramid for camouflaged object detection with transformers. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 5557-5566, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Laurent Itti, Christof Koch, and Ernst Niebur. A model of saliency-based visual attention for rapid scene analysis. IEEE Transactions on pattern analysis and machine intelligence, 20(11):1254-1259, 1998. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30373" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[25] Ge-Peng Ji, Deng-Ping Fan, Yu-Cheng Chou, Dengxin Dai, Alexander Liniger, and Luc Van Gool. Deep gradient learning for efficient camouflaged object detection. Machine Intelligence Research, 20:92-108, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[26] Tilke Judd, Krista Ehinger, Frédo Durand, and Antonio Torralba. Learning to predict where humans look. In The IEEE/CVF International Conference on Computer Vision, pages 2106-2113. IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[27] Xunfa Lai, Zhiyu Yang, Jie Hu, Shengchuan Zhang, Liujuan Cao, Guannan Jiang, Zhiyu Wang, Songan Zhang, and Rongrong Ji. Camoteacher: Dual-rotation consistency learning for semi-supervised camouflaged object detection. In European Conference on Computer Vision, 2024. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[28] Jiaying Lin, Xin Tan, Ke Xu, Lizhuang Ma, and Rynson W. H. Lau. Frequency-aware camouflaged object detection. ACM Trans. Multimedia Comput. Commun. Appl., 19 (2), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.335, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[29] Jianghang Lin, Yunhang Shen, Bingquan Wang, Shaohui Lin, Ke Li, and Liujuan Cao. Weakly supervised open-vocabulary object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3404-3412, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.474 + ], + "angle": 0, + "content": "[30] Naisong Luo, Yuwen Pan, Rui Sun, Tianzhu Zhang, Zhiwei Xiong, and Feng Wu. Camouflaged instance segmentation via explicit de-camouflaging. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 17918-17927, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[31] Ran Margolin, Lihi Zelnik-Manor, and Ayellet Tal. How to evaluate foreground maps? In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 248-255, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[32] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 8364-8375, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.605, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[33] Peng Mi, Jianghang Lin, Yiyi Zhou, Yunhang Shen, Gen Luo, Xiaoshuai Sun, Liujuan Cao, Rongrong Fu, Qiang Xu, and Rongrong Ji. Active teacher for semi-supervised object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14482-14491, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[34] Tam Nguyen, Maximilian Dax, Chaithanya Kumar Mummadi, Nhung Ngo, Thi Hoai Phuong Nguyen, Zhongyu Lou, and Thomas Brox. Deepusps: Deep robust unsupervised saliency prediction via self-supervision. Advances in Neural Information Processing Systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[35] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[36] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[37] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[38] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoomnext: A unified collaborative pyramid network for camouflaged object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024, 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[39] Federico Perazzi, Philipp Krahenbuhl, Yael Pritch, and Alexander Hornung. Saliency filters: Contrast based filtering for salient region detection. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 733-740. IEEE, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[40] Alec Radford. Improving language understanding by generative pre-training. OpenAI blog, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[41] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[42] Jingjing Ren, Xiaowei Hu, Lei Zhu, Xuemiao Xu, Yangyang Xu, Weiming Wang, Zijun Deng, and Pheng-Ann Heng. Deep texture-aware features for camouflaged object detection. IEEE Transactions on Circuits and Systems for Video Technology, 33(3):1157-1167, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[43] Gyungin Shin, Samuel Albanie, and Weidi Xie. Unsupervised salient object detection with spectral cluster voting. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 3971-3980, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[44] Oriane Simeoni, Chloe Sekkat, Gilles Puy, Antonín Vobecký, Éloi Zablocki, and Patrick Pérez. Unsupervised object localization: Observing the background to discover objects. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 3176-3186, 2023. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[45] Dongyue Sun, Shiyao Jiang, and Lin Qi. Edge-aware mirror network for camouflaged object detection, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.732 + ], + "angle": 0, + "content": "[46] Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. Bert4rec: Sequential recommendation with bidirectional encoder representations from transformer. In Proceedings of the 28th ACM international conference on information and knowledge management, pages 1441-1450, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[47] Yujia Sun, Geng Chen, Tao Zhou, Yi Zhang, and Nian Liu. Context-aware cross-level fusion network for camouflaged object detection. In The International Joint Conference on Artificial Intelligence, pages 1025–1031, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[48] Yujia Sun, Shuo Wang, Chenglizhao Chen, and Tian-Zhu Xiang. Boundary-guided camouflaged object detection. In The International Joint Conference on Artificial Intelligence, pages 1335-1341, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Yanguang Sun, Hanyu Xuan, Jian Yang, and Lei Luo. Glconet: Learning multisource perception representation for camouflaged object detection. IEEE Transactions on Neural Networks and Learning Systems, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "30374" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[50] Domen Tabernik, Samo Šela, Jure Skvarč, and Danijel Skočaj. Segmentation-based deep-learning approach for surface-defect detection. Journal of Intelligent Manufacturing, 31(3):759-776, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[51] Andrey Voynov, Stanislav Morozov, and Artem Babenko. Object segmentation without labels with large-scale generative models. In International Conference on Machine Learning, pages 10596-10606. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[52] Qingwei Wang, Jinyu Yang, Xiaosheng Yu, Fangyi Wang, Peng Chen, and Feng Zheng. Depth-aided camouflaged object detection. In Proceedings of the 31st ACM International Conference on Multimedia, page 3297–3306, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.329 + ], + "angle": 0, + "content": "[53] Yangtao Wang, Xi Shen, Shell Xu Hu, Yuan Yuan, James L Crowley, and Dominique Vaufreydaz. Self-supervised transformers for unsupervised object discovery using normalized cut. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 14543-14553, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[54] Yichen Wei, Fang Wen, Wangjiang Zhu, and Jian Sun. Geodesic saliency using background priors. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part III 12, pages 29-42. Springer, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.401, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[55] Among Wu, Rui Liu, Yahong Han, Linchao Zhu, and Yi Yang. Vector-decomposed disentanglement for domain-invariant object detection. The IEEE/CVF International Conference on Computer Vision, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[56] Mochu Xiang, Jing Zhang, Yunqiu Lv, Aixuan Li, Yiran Zhong, and Yuchao Dai. Exploring depth contribution for camouflaged object detection, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.539 + ], + "angle": 0, + "content": "[57] Qiong Yan, Li Xu, Jianping Shi, and Jiaya Jia. Hierarchical saliency detection. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 1155-1162, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.609 + ], + "angle": 0, + "content": "[58] Fan Yang, Qiang Zhai, Xin Li, Rui Huang, Ao Luo, Hong Cheng, and Deng-Ping Fan. Uncertainty-guided transformer reasoning for camouflaged object detection. In The IEEE/CVF International Conference on Computer Vision, pages 4146-4155, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.61, + 0.482, + 0.664 + ], + "angle": 0, + "content": "[59] Siyuan Yao, Hao Sun, Tian-Zhu Xiang, Xiao Wang, and Xiaochun Cao. Hierarchical graph interaction transformer with dynamic token clustering for camouflaged object detection. IEEE Transactions on Image Processing, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[60] Bowen Yin, Xuying Zhang, Deng-Ping Fan, Shaohui Jiao, Ming-Ming Cheng, Luc Van Gool, and Qibin Hou. Camoformer: Masked separable attention for camouflaged object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[61] Qiang Zhai, Xin Li, Fan Yang, Chenglizhao Chen, Hong Cheng, and Deng-Ping Fan. Mutual graph learning for camouflaged object detection. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 12997-13007, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[62] Jing Zhang, Tong Zhang, Yuchao Dai, Mehrtash Harandi, and Richard Hartley. Deep unsupervised saliency detection: A multiple noisy labeling perspective. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 9029-9038, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[63] Yi Zhang and Chengyi Wu. Unsupervised camouflaged object segmentation as domain adaptation. In The IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "International Conference on Computer Vision, pages 4334-4344, 2023. 1, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[64] Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, and Nicu Sebe. Bilateral reference for high-resolution dichotomous image segmentation. CAAI Artificial Intelligence Research, 3:9150038, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[65] Yijie Zhong, Bo Li, Lv Tang, Senyun Kuang, Shuang Wu, and Shouhong Ding. Detecting camouflaged object in frequency domain. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 4494-4503, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[66] Hongwei Zhu, Peng Li, Haoran Xie, Xu Yan, Dong Liang, Dapeng Chen, Mingqiang Wei, and Jing Qin. I can find you! boundary-guided separated attention network for camouflaged object detection. In Association for the Advancement of Artificial Intelligence, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.907, + 0.361 + ], + "angle": 0, + "content": "[67] Jinchao Zhu, Xiaoyu Zhang, Shuo Zhang, and Junnan Liu. Inferring camouflaged objects by texture-aware interactive guidance network. In Association for the Advancement of Artificial Intelligence, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[68] Wangjiang Zhu, Shuang Liang, Yichen Wei, and Jian Sun. Saliency optimization from robust background detection. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 2814-2821, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[69] Mingchen Zhuge, Xiankai Lu, Yiyou Guo, Zhihua Cai, and Shuhan Chen. Cubenet: X-shape connection for camouflaged object detection. Pattern Recognition, 127:108644, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "30375" + } + ] +] \ No newline at end of file diff --git a/2025/UCOD-DPL_ Unsupervised Camouflaged Object Detection via Dynamic Pseudo-label Learning/dec1b26d-4dd9-4c06-a461-c03d0e31f356_origin.pdf b/2025/UCOD-DPL_ Unsupervised Camouflaged Object Detection via Dynamic Pseudo-label Learning/dec1b26d-4dd9-4c06-a461-c03d0e31f356_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3d33545f0146bc66b6e44fda39182153efb6f4ce --- /dev/null +++ b/2025/UCOD-DPL_ Unsupervised Camouflaged Object Detection via Dynamic Pseudo-label Learning/dec1b26d-4dd9-4c06-a461-c03d0e31f356_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6aa79b8eeffe0ca0fdaaa401c8ba7c200cbee0cf2585313d55bd093d8e2fe347 +size 3442790 diff --git a/2025/UCOD-DPL_ Unsupervised Camouflaged Object Detection via Dynamic Pseudo-label Learning/full.md b/2025/UCOD-DPL_ Unsupervised Camouflaged Object Detection via Dynamic Pseudo-label Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4d7dc37569b4c476f1d1a276302bef64414b812c --- /dev/null +++ b/2025/UCOD-DPL_ Unsupervised Camouflaged Object Detection via Dynamic Pseudo-label Learning/full.md @@ -0,0 +1,398 @@ +# UCOD-DPL: Unsupervised Camouflaged Object Detection via Dynamic Pseudo-label Learning + +Weiqi Yan $^{1}$ , Lvhai Chen $^{1}$ , Huaijia Ko $^{1}$ , Shengchuan Zhang $^{1*}$ , Yan Zhang $^{1}$ , Liujuan Cao $^{1}$ + +1 Key Laboratory of Multimedia Trusted Perception and Efficient Computing, + +Ministry of Education of China, Xiamen University, 361005, P.R. China. + +weiqi_yan@outlook.com lvhaichen2002@gmail.com zsc_2016@xmu.edu.cn + +# Abstract + +Unsupervised Camouflaged Object Detection (UCOD) has gained attention since it doesn't need to rely on extensive pixel-level labels. Existing UCOD methods typically generate pseudo-labels using fixed strategies and train $1 \times 1$ convolutional layers as a simple decoder, leading to low performance compared to fully-supervised methods. We emphasize two drawbacks in these approaches: 1). The model is prone to fitting incorrect knowledge due to the pseudolabel containing substantial noise. 2). The simple decoder fails to capture and learn the semantic features of camouflaged objects, especially for small-sized objects, due to the low-resolution pseudo-labels and severe confusion between foreground and background pixels. To this end, we propose a UCOD method with a teacher-student framework via Dynamic Pseudo-label Learning called UCOD-DPL, which contains an Adaptive Pseudo-label Module (APM), a Dual-Branch Adversarial (DBA) decoder, and a Look-Twice mechanism. The APM module adaptively combines pseudo-labels generated by fixed strategies and the teacher model to prevent the model from overfitting incorrect knowledge while preserving the ability for self-correction; the DBA decoder takes adversarial learning of different segmentation objectives, guides the model to overcome the foreground-background confusion of camouflaged objects, and the Look-Twice mechanism mimics the human tendency to zoom in on camouflaged objects and performs secondary refinement on small-sized objects. Extensive experiments show that our method demonstrates outstanding performance, even surpassing some existing fully supervised methods. The code is available now1. + +![](images/5ac8efa5b68b91a90ea10acb677ac58490f6f67c51ff90cef9129cf4c0b41b87.jpg) +Figure 1. Comparison between our method and previous pseudolabel-based methods. + +# 1. Introduction + +"Camouflage" originates from the natural behavior of objects blending into their surroundings using similar textures and colors to evade "predators"[4]. Camouflaged Object Detection (COD) is a challenging semantic segmentation task aimed at learning to segment objects that are visually hidden in their backgrounds, which has significant applications in several important fields [11, 50]. + +Although fully-supervised methods have achieved significant progress in the COD tasks [19, 38, 49, 59, 60], they heavily rely on large-scale human annotations for training. However, it is difficult to get pixel-level human labels for camouflaged objects due to their complex visual attributes [27]. Therefore, this reliance on annotations has sparked increasing interest in the task of unsupervised camouflage object detection (UCOD) among researchers. However, existing UCOD methods typically use pseudo-labels generated by fixed strategies to train a single $1 \times 1$ convolutional layer [44, 63]. By investing in existing methods, we identified two main issues: 1). The pseudo-label con + +![](images/c037aa66092f0a1f0d582e430edd9af8efac42e7e807c6839a8cba512a77e4a9.jpg) +Small-sized + +![](images/0896fe4269343f73288e1be4ac1ae579ca458ea16653dff23238022e09984fdb.jpg) +Huge-sized + +![](images/abdac31dd9a5aea9a1280a5e1ae313a3bc47d84d78d00663bfee90522fccf1d0.jpg) +Multiple + +![](images/560531f721e64c89cd6d43697af21c7e59b8e59b058534ebda9cdd426ec4e159.jpg) +Uncertainty + +![](images/ef753cf78e411b7724b52ed8462e5e3657c85f7a4a772376ef1b1957147ae3f7.jpg) +CocosDg + +![](images/37c5dc4ac4c08502e20ef8140b38c7803a50f816266c7ae11c483dfb75e4d2e9.jpg) +1 + +![](images/549641dd98f921a8c5679018e0f06c1935125b49424aeee67aee7eb3b68cb56b.jpg) + +![](images/e5e2045dd2258de7408624fdca3dbac1ff29b150781030371efd01eba7bd0783.jpg) +Figure 2. Some examples of low-quality pseudo-labels generated using fixed strategies. Based on features extracted by DINOv2, we construct pseudo-labels using the Background-Seed from FOUND [44] and the MaskCut in UCOD-DA [44], in challenging scenarios (e.g. small-sized objects, huge-sized objects, multiple instances, and uncertain objects) for visualization. + +![](images/c28da8f540c7727fad1f22bf991d391ad860722a319d80e4aabd6c591d301dbc.jpg) + +![](images/e5b5861e533732c198bbd4e4bccf4f4aba2d9164fa40bf4fd0d03a1f5b2b3cda.jpg) + +![](images/b126565ffeb8f8d829db44757da0ba6b8fda433bc0f85aa3a2fa20970e948cc0.jpg) + +![](images/de55faa6b59c87878981166000c394e2fc556aa8ae7a3cb5c9697577b62548f5.jpg) + +![](images/dfc2be28101e075b0e2865dc8e42e2cd9e30732770969d56bacbacf0b6e53e46.jpg) + +![](images/301cc198cda7d14e37dbf9e905d9c708da16277b655a67962eae80afb20006d6.jpg) + +![](images/c22fe784f1af5024970c4e127c8a598aeb840a1a88013bed875d5ee63a0fde84.jpg) + +![](images/c647237a4427a22322048b2573886a66ea50a195eb5f28afeeaa8c7b08d7e86e.jpg) + +![](images/26683d13d030418266cb11fa7b2ac1ead541e51029cfca245385d7da1ef905cc.jpg) + +structed by fixed strategies contains substantial noise, which makes the model prone to fitting incorrect knowledge. 2). These pseudo-labels have low resolution and severe confusion between foreground and background pixels, and simple $1 \times 1$ convolution fails to capture and learn the semantic features of camouflaged objects, especially for small-sized objects. Consequently, their segmentation results fall far behind those of fully supervised methods, which limits their application in real-world scenarios. + +In this paper, we propose a brand-new model called UCOD-DPL. As shown in Fig. 1, we introduce a teacher-student framework consisting of an Adaptive Pseudo-label Mixing (APM) module, a Dual-Branch Adversarial (DBA) decoder, and a Look-Twice refinement strategy. As illustrated in Fig. 2, pseudo-labels constructed by fixed strategies are generally low-quality and contain substantial noise. To prevent the model from fitting uncorrectable erroneous knowledge in the pseudo-labels, we design the APM module with a discriminator and a dynamic scoring mechanism to mix the pseudo-labels from the fixed strategy and the teacher branch adaptively. To help the model overcome foreground-background confusion, we propose a DBA decoder inspired by predator-prey dynamics. This decoder uses different segmentation targets (foreground and background) along with an orthogonal loss of features to deeply mine and learn distinctive features of the foreground and background. Additionally, inspired by [37, 38], we mimic the zooming-in behavior humans use when examining camouflaged images by performing secondary refinement on + +small-sized camouflaged objects to enhance the model's segmentation accuracy when dealing with small targets. + +In the experiment, we train our model on the commonly used COD training set and evaluate it on multiple test sets. The results demonstrate that our method achieves outstanding performance, even comparable to some fully supervised COD methods when using DINOv2 as the backbone. + +Our main contributions are summarized as follows: + +1. We propose a novel UCOD model called UCOD-DPL, and extensive experiments demonstrate that our method exhibits superior performance, reaching the performance level of some fully supervised algorithms. +2. We introduce a teacher-student framework with an APM module. By using the APM, we ensure that the model can learn knowledge from pseudo-labels while preserving the ability to identify and correct the erroneous. +3. To address the issues of low-resolution pseudo-labels and foreground-background confusion, we design a DBA decoder that uses a foreground-background adversarial task, enabling the model to learn the camouflaged features of the foreground adaptively. +4. To enhance the model's ability to segment small-sized objects, we design a Look-Twice refinement mechanism. This mechanism mimics human behavior by performing a secondary segmentation on smaller camouflaged objects, thereby refining the segmentation results. + +# 2. Related Works + +# 2.1. Unsupervised Semantic Segmentation + +Unsupervised Semantic Segmentation (USS) aims to train a model without any human annotations to produce the semantic segmentation mask. Before the surge of deep learning methods, researchers relied on numerous handcrafted methods [7, 16, 57, 68] that used one or more priors related to foreground regions in the image to generate segmentation masks, such as center[26], contrast[24], and boundary[54]. In deep learning-based methods, some approaches utilize generative models [17] to produce a segmentation mask, which will be used to synthesize a realistic image by copying the corresponding region into a background which is either synthesized [3] or taken from a different image [1]. Another category of methods leverages pseudo-labels constructed by handcrafted fixed strategies to train the detector. [62] adopts diverse noisy pseudo-labels via distinct unsupervised handcrafted methods to train the detector to predict a saliency map free from noise in labels. And [34, 44] proposed to refine the pseudo-labels produced by fixed-strategy by training the model via a self-supervised iterative refinement manner. These methods adopt a self-supervision manner that highly relies on pseudo-labels generated directly from one or multiple fixed strategies. If the pseudo-labels contain substantial noise, these methods remain prone to + +fitting incorrect knowledge, which is particularly evident when dealing with camouflaged objects. + +# 2.2. Camouflaged Object Detection + +Camouflaged Object Detection (COD) is a complex task aimed at segmenting objects that are intentionally designed to merge seamlessly with their environment. This task has an extensive history in its domain. Existing fully-supervised COD methods [10, 37, 38, 69] have achieved excellent performance by training on large-scale annotated datasets. These methods utilize the rich information (e.g. boundary [25, 45, 66], texture [25, 42, 67] and other information like frequency domain and depth [28, 52, 56, 65]) provided by pixel-level annotations and design a series of sophisticated strategies to extract unique visual characteristics of camouflaged objects. However, the extensive pixel-level annotations are hardly gained, which has drawn the attention of researchers to Unsupervised Camouflaged Object Detection (UCOD) and USS methods [63]. These methods typically rely on pseudo-labels generated through fixed strategies to supervise model learning for segmentation. However, due to limitations in algorithmic complexity and robustness, these fixed strategies struggle to work effectively at high resolutions and easily introduce significant noise. Consequently, these methods are prone to learning incorrect noise from the pseudo-labels, leading to suboptimal segmentation performance and limiting the application in the real world of UCOD models. + +# 2.3. Self-supervised Learning + +Self-Supervised Learning (SSL) has revolutionized the way models leverage unlabeled data. Among current SSL approaches, some methods adopt contrastive learning to learn image feature representations. MoCo [20] utilizes a dynamic memory bank to maintain a large pool of negative samples; SimCLR [6] employs various data augmentations and a projection head to maximize similarities between differently augmented views of the same instance; and BYOL [18] constructs a teacher-student framework with the teacher model's parameters are updated by Exponential Moving Average (EMA) of the student model. Following BYOL, DINO [5] applied two interactive encoders sharing the same structure but with different parameter sets and update strategies. Inspired by self-supervised pretraining techniques in natural language processing [40, 41, 46], other SSL methods design pretext tasks (e.g. colorization, rotation prediction, and patch reconstruction) on unlabeled images [5, 6, 14, 15, 20, 29, 33, 35]. BeiT [2] adopts a Transformer-based architecture to reconstruct masked image patches, MAE [21] implements a masked autoencoder that efficiently reconstructs the missing parts of an image. In this paper, we employ DINOv1 [5] and DINOv2 [36] to extract robust and generalized image features. + +# 3. Methodology + +# 3.1. Overall Framework + +The structure of our proposed method is shown in Fig. 3, which includes a teacher-student framework with an Adaptive Pseudo-label Mixing (APM) module, a Dual-Branch Adversarial (DBA) decoder, and a Look-Twice strategy. Both the teacher and student models are configured as DBA decoders. + +Firstly, the $i$ -th unlabeled input image passes through the self-supervised pretrained backbone to obtain high-level semantic image feature $F_{i}$ . We then employ the background seed method to generate a low-resolution fixed-strategy pseudo-label $\hat{P}_i^{fs}$ , which utilizes pixel-level similarity comparison and was proposed in [44]. Simultaneously, the image feature $F_{i}$ is fed into a teacher-student model. The teacher model generates a higher-resolution teacher's pseudo label $\hat{P}_i^t$ , and the student model produces the foreground and background predicted mask $\hat{Y}_i^{FG}, \hat{Y}_i^{BG}$ . Then, $\hat{P}_i^{fs}$ and $\hat{P}_i^t$ are fed into the APM module to construct the mixed dynamic pseudo-label $P_{i}$ . Finally, we use the mixed dynamic pseudo-label $P_{i}$ to supervise the final foreground and background predicted masks $\hat{Y}_i^{FG}, \hat{Y}_i^{BG}$ . We further apply a Look-Twice strategy to perform a secondary refinement on small-sized objects to improve the model's segmentation performance on small-sized samples. + +# 3.2. Adaptive Pseudo-Label Merging Module + +Existing unsupervised semantic segmentation methods typically adopt a fixed-strategy (e.g. similarity comparison [32, 43, 53]) or utilize a dual-branch model (e.g. fixed-strategy or learnable branch [44, 63]) structure for self-supervised learning to generate segmentation masks. However, we believe the performance of these methods is constrained by the strategy's design or the pseudo-labels' quality, preventing the model from achieving higher accuracy. + +We introduce a discriminator $\mathcal{D}$ , which inputs a foreground segmentation mask (i.e. fixed-strategy pseudo-label $\hat{P}_i^{fs}$ or student's foreground predicted mask $\hat{Y}_i^{FG}$ ) and outputs the probability that the mask originates from the fixed-strategy branch: + +$$ +\hat {y} _ {i} ^ {p 1} = \mathcal {D} \left(\hat {P} _ {i} ^ {f s}\right), \hat {y} _ {i} ^ {p 2} = \mathcal {D} \left(\hat {Y} _ {i} ^ {F G}\right), \tag {1} +$$ + +where $\hat{\mathbf{y}}_i^{p1},\hat{\mathbf{y}}_i^{p2}$ denote the predicted probability that the input mask belongs to the fixed-strategy branch. We then introduce a scoring function $S(\cdot)$ with a temporal constraint: + +$$ +\mathcal {S} \left(\hat {\mathrm {y}} _ {i} ^ {p 1}, \hat {\mathrm {y}} _ {i} ^ {p 2}\right) = \operatorname {C L I P} \left(\frac {t}{T} + \frac {1}{2} \left(1 + \cos \left(\pi \times \left| \hat {\mathrm {y}} _ {i} ^ {p 1} - \hat {\mathrm {y}} _ {i} ^ {p 2} \right|\right)\right) _ {0} ^ {1}, \right. \tag {2} +$$ + +where $t, T$ denote the current and total training epochs, $\mathrm{CLIP}(x)_0^1$ denotes truncating $x$ so that it lies within the interval [0, 1]. In the early stage of training, the teacher model + +![](images/500a56bfabcc9538ae36538d3af179e5afd19ea4959a3c6175fd76829ca5b5fe.jpg) + +![](images/f58d65e25dfd0c36d887656828ddc15dc7851f6fec0e0893f7195b056fc37be1.jpg) +Figure 3. The main framework of our proposed method. The model contains a teacher-student framework, an Adaptive Psdudo-label Mixing (APM) module, a Dual-Branch Adversarial (DBA) decoder, and a Look-Twice strategy. + +has not yet developed sufficient localization and basic semantic segmentation capabilities for camouflaged objects, making it unable to supervise the student model effectively. At this point, there is a significant difference between fixed-strategy pseudo-label $\hat{P}_i^{fs}$ and the student's foreground predicted mask $\hat{Y}_i^{FG}$ . To help the model learn the fundamental foreground localization and segmentation tasks, we use a higher proportion of $\hat{P}_i^{fs}$ to supervise model training. Later in the training process, as the $\hat{Y}_i^{FG}$ gradually approaches $\hat{P}_i^{fs}$ , we increase the proportion of the teacher's pseudo-label $\hat{P}_i^t$ as it can make more stable predictions. This adjustment helps prevent the model from overfitting to noise and incorrect knowledge in the $\hat{P}_i^{fs}$ . We use the obtained score as the mixing weight $W_i^t$ to combine the fixed-strategy pseudo-label $\hat{P}_i^{fs}$ and teacher's predictions $\hat{P}_i^t$ : + +$$ +\mathrm {W} _ {i} ^ {t} = \mathcal {S} \left(\hat {\mathrm {y}} _ {i} ^ {p 1}, \hat {\mathrm {y}} _ {i} ^ {p 2}\right), \tag {3} +$$ + +$$ +P _ {i} = \mathrm {W} _ {i} ^ {t} \hat {P} _ {i} ^ {t} + (1 - \mathrm {W} _ {i} ^ {t}) \hat {P} _ {i} ^ {f s}, +$$ + +where $P_{i}$ denotes the mixed dynamic pseudo-label. For the training of the discriminator, we set a binary classification task by using $\hat{P}_i^{fs}$ and $\hat{Y}_i^{FG}$ as the input and assign the discriminator label $\mathrm{y}_i^{p1} = 1$ for the fixed-strategy pseudo-label $\hat{P}_i^{fs}$ and $\mathrm{y}_i^{p2} = 0$ for the student model branch $\hat{Y}_i^{FG}$ . Then, + +we employ a cross-entropy supervision loss $\mathcal{L}_{\mathrm{BCE}}$ to train the discriminator: + +$$ +\mathcal {L} _ {d i s} = \mathcal {L} _ {\mathrm {B C E}} \left(\hat {\mathrm {y}} _ {i} ^ {p j}, \mathrm {y} _ {i} ^ {p j}\right), j \in \{1, 2 \}. \tag {4} +$$ + +During the training, we treat the student model as a kind of generator and alternately train the decoder and the discriminator in a GAN-like manner. + +# 3.3. Dual-Branch Adversarial Decoder + +In previous methods [44, 63], predictions are typically generated by a simple $1 \times 1$ convolution, which limits the ability to capture camouflage features thoroughly. We propose a Dual-Branch Adversarial decoder to enhance accuracy and robustness in this task. The DBA decoder consists of two parallel branches: one is responsible for segmenting the foreground predicted mask $\hat{Y}_i^{FG}$ , while the other is for segmenting the background predicted mask $\hat{Y}_i^{BG}$ . + +First, for $i$ -th image feature $F_{i}$ extracted from the backbone model, a convolutional layer is used to double its channel, then we decouple it into foreground and background features $F_{i}^{FG}, F_{i}^{BG}$ with the same shape of $F_{i}$ : + +$$ +F _ {i} ^ {F G}, F _ {i} ^ {B G} = \operatorname {S p l i t} \left(\operatorname {C o n v} \left(F _ {i}\right)\right), \tag {5} +$$ + +where $\text{Split}(\cdot)$ denotes splitting a tensor along the channel dimension equally. We then employ two learnable em + +beddings $\mathrm{E}_{FG},\mathrm{E}_{BG}$ to store foreground and backgroundrelated knowledge. These embeddings are used to calculate foreground and background attention queries $Q_{i}^{FG^{\prime}}$ $Q_{i}^{BG^{\prime}}$ .. + +$$ +Q _ {i} ^ {F G ^ {\prime}} = F _ {i} ^ {F G} \times \mathrm {E} _ {F G}, Q _ {i} ^ {B G ^ {\prime}} = F _ {i} ^ {B G} \times \mathrm {E} _ {B G}. \tag {6} +$$ + +Finally, we apply multiplicative attention with a residual connection to weigh the features, enhancing important details. The resulting features are then passed through a $1 \times 1$ convolution to output the foreground predicted mask $\hat{Y}_i^{FG}$ and background mask $\hat{Y}_i^{BG}$ : + +$$ +\hat {Y} _ {i} ^ {F G} = \operatorname {C o n v} ^ {\prime} \left(\operatorname {S i g m o i d} \left(Q _ {i} ^ {F G ^ {\prime}} \times F _ {i} ^ {F G}\right) + F _ {i} ^ {F G}\right), \tag {7} +$$ + +$$ +\hat {Y} _ {i} ^ {B G} = \operatorname {C o n v} ^ {\prime} (\operatorname {S i g m o i d} \left(Q _ {i} ^ {B G ^ {\prime}} \times F _ {i} ^ {B G}\right) + F _ {i} ^ {B G}). +$$ + +For the student model, we use $Y_{i}^{FG}, \hat{Y}_{i}^{BG}$ as the student's foreground and background prediction masks. For the teacher model, we further generate the teacher's pseudolabel $P_{i}$ using the combination of the foreground and background prediction masks. + +Inspired by [55], we apply an orthogonal loss $\mathcal{L}_{\perp}$ for $Q_{i}^{FG^{\prime}}\in \mathbb{R}^{(n\times m)\times c},Q_{i}^{BG^{\prime}}\in \mathbb{R}^{(n\times m)\times c}$ as a regularisation constraint to encourage them to focus on distinct features: + +$$ +\mathbb {M} = Q _ {i} ^ {F G ^ {\prime}} \left(Q _ {i} ^ {B G ^ {\prime}}\right) ^ {\top}, +$$ + +$$ +\mathcal {L} _ {\perp} = \frac {1}{(n \times m) ^ {2}} \sum_ {j = 1} ^ {n \times m} \sum_ {k = 1} ^ {n \times m} (1 - \delta_ {j k}) (\mathbb {M} [ j, k ]) ^ {2}, \tag {8} +$$ + +where $n, m, c$ indicate the height, width, and number of channels. When $j = k$ , $\delta_{jk} = 1$ ; otherwise $\delta_{jk} = 0$ . For the predictions $\hat{Y}_i^{FG}, \hat{Y}_i^{BG}$ , we use a binary cross-entropy loss $\mathcal{L}_{\mathrm{BCE}}$ for supervision, where the background prediction is inverted before applying the supervision: + +$$ +\mathcal {L} _ {s e g} = \mathcal {L} _ {\mathrm {B C E}} (\hat {Y} _ {i} ^ {F G}, P _ {i}) + \mathcal {L} _ {\mathrm {B C E}} \left((1 - \hat {Y} _ {i} ^ {B G}), P _ {i}\right). (9) +$$ + +# 3.4. Look-Twice Refinement + +The segmentation quality of COD is often influenced by object size. Due to reduced spatial resolution, the model struggles to produce fine segmentation for small-sized camouflaged objects. Inspired by previous works [37, 38], we attempt to mimic human behavior when observing small-sized objects: first coarsely locating the camouflaged object in the image, then zooming in to carefully retrieve its details. To this end, we designed the Look-Twice mechanism. First, the input image is processed by the model to obtain a coarse segmentation result $\hat{Y}_i^{FG}$ . We define the set of all foreground objects as $\mathrm{C}^{FG}$ , which can be calculated as: + +$$ +\mathrm {C} ^ {\mathrm {F G}} = \text {C o n n e c t C o m p o n e n t} \left(\hat {Y} _ {i} ^ {F G}\right), \tag {10} +$$ + +where ConnectComponent(\cdot) denotes the connected components labelling algorithm. Next, for each connect + +component $c_k^{FG} \in \mathbb{C}^{FG}$ , we calculate its foreground area ratio $r_k^{FG}$ : + +$$ +r _ {k} ^ {F G} = \frac {\sum_ {(x , y) \in \mathrm {c} _ {k} ^ {F G}} \hat {Y} _ {i} ^ {F G} (x , y)}{H \times W}, \tag {11} +$$ + +where $H, W$ denote the height and width of the foreground area of the output mask $\tilde{Y}_i^{FG}$ . We consider all regions where $r_i^{FG} < \tau$ as small-sized object areas. We set $\tau = 0.15$ when using DINOv2 as the backbone. The hyperparameter ablation study on $\tau$ will be provided in the experiments section. + +For each cropped small-sized object region, we aim to retain a certain proportion of the background to provide the model with sufficient context to segment the foreground from the background. Therefore, for each connected component region $c_k^{FG}$ , we calculate the local foreground area ratio $s_k^{FG}$ and the global area ratio $s_k^{BG}$ : + +$$ +s _ {k} ^ {F G} = \frac {\sum_ {(x , y) \in c _ {k} ^ {F G}} \hat {Y} _ {i} ^ {F G} (x , y)}{H _ {k} ^ {F G} \times W _ {k} ^ {F G}}, \tag {12} +$$ + +$$ +s _ {k} ^ {B G} = \frac {H _ {k} ^ {F G} \times W _ {k} ^ {F G}}{H \times W}, +$$ + +where $H_{k}^{FG},W_{k}^{FG}$ denote the height and width of the bounding box of the $k$ -th connected component $\mathrm{c}_k^{FG}$ .. + +$$ +\operatorname {s c a l e} _ {k} ^ {F G} = 1 - \left(s _ {k} ^ {F G} / s _ {k} ^ {B G}\right). \tag {13} +$$ + +Finally, we calculate the expansion ratio using Eq. (13) and enlarge the bounding box of the connected component region by this ratio. The expanded bounding box is then used to crop the related area of the input image, and the cropped patches are resized to the original input size. During the training process, these patches are treated as augmented data. At test time, we re-infer these patches, scale the resulting foreground predicted mask back to the original size, and paste it into the coarse mask as the refined result. However, we found that this mechanism still has some weakness: for cases where objects are occluded and the annotations are fragmented, this method may mistakenly amplify a part of the object as a small object. This can serve as a point of our future work. + +# 3.5. Total Loss + +Finally, the total loss $\mathcal{L}_{tot}$ can be defined as follows: + +$$ +\mathcal {L} _ {t o t} = \mathcal {L} _ {s e g} + \mathcal {L} _ {\perp} + \mathcal {L} _ {d i s}, \tag {14} +$$ + +where $\mathcal{L}_{seg}$ is the segmentation loss defined by Eq. (9), $\mathcal{L}_{\perp}$ is the orthogonal loss defined by Eq. (8), and $\mathcal{L}_{dis}$ is the discriminator loss defined by Eq. (4). When training on the teacher-student framework, we freeze the parameters of the discriminator and use $\mathcal{L}_{seg}$ to supervise the model training. Conversely, when training the discriminator, we freeze the parameters of the other parts of the model and train the discriminator only using $\mathcal{L}_{dis}$ . + +
MethodsCHAMELEON (87)CAMO-Test (250)COD10K-Test (2,026)NC4K (4,121)
Sm↑FωβFmβEmφM↓Sm↑FωβFmβEmφM↓Sm↑FωβFmβEmφM↓Sm↑FωβFmβEmφM↓
Fully-Supervised Methods
SINet20[10].872.806.827.946.034.751.606.675.771.100.771.551.634.806.051.808.723.769.871.058
C2FNet21[47].888.828.844.946.032.796.719.762.864.080.813.686.723.900.036.838.762.794.904.049
MGL-R21[61].893.812.834.941.030.775.673.726.842.088.814.666.710.890.035.833.739.782.893.053
UGTR21[58].887.794.819.940.031.784.684.735.851.086.817.666.711.890.036.839.746.787.899.052
BGNet22[48].901.851.860.954.027.812.749.789.870.073.831.722.753.901.033.851.788.820.907.044
ZoomNet22[37].902.845.864.958.023.820.752.794.878.066.838.729.766.888.029.853.784.818.896.043
SINetv22[12].888.816.835.961.030.820.743.782.882.070.815.680.718.887.037.847.770.805.903.048
HitNet23[22].921.897.900.972.019.849.809.831.906.055.871.806.823.935.023.875.834.853.926.037
FSPNet23[23].908.851.867.965.023.856.799.830.899.050.851.735.769.895.026.879.816.843.915.035
BiRefNet24[64].929.911.922.968.016.932.914.922.974.015.913.874.888.960.014.914.894.909.953.023
Semi-Supervised Methods
CamoTeacher24(1%)[27].652.472.558.714.093.621.456.545.669.136.699.517.788.797.062.718.599.779.814.090
CamoTeacher24(5%)[27].729.587.656.785.070.669.523.601.711.122.745.583.827.840.050.777.677.834.859.071
CamoTeacher24(10%)[27].756.617.684.813.065.701.560.742.795.112.759.594.836.854.049.791.687.842.868.068
SCOD-ND24(10%)[13].850.773-.928.036.789.732-.859.077.819.725-.891.033.838.787-.903.046
Unsupervised Methods
BigGW21[51].547.244.294.527.257.565.299.349.528.282.528.185.246.497.261.608.319.391.565.246
TokenCut22[53].654.496.536.740.132.633.498.543.706.163.658.469.502.735.103.725.615.649.802.101
TokenCut22 w/B.S.[53].655.351.393.582.169.639.383.434.595.195.666.334.399.609.127.735.478.547.683.133
SpectralSeg22[32].575.410.440.628.220.579.450.481.648.235.575.360.388.595.193.669.535.562.719.159
SelfMask22[43].619.436.481.675.176.617.483.536.698.176.637.431.469.679.131.716.593.634.777.114
SelfMask22 w/U.B.[43].629.447.491.683.169.627.495.547.708.182.645.440.478.687.125.723.601.642.784.110
FOUND22-DINOv1[44].684.542.590.810.095.685.584.633.782.129.670.482.520.751.085.741.637.674.824.084
*FOUND22-DINOv2[44].829.757.781.911.040.770.704.740.849.090.767.641.668.847.045.816.756.783.893.052
UCOS-DA23-DINOv1[63].715.591.629.802.095.701.606.646.784.127.689.513.546.740.086.755.656.689.819.085
*UCOS-DA23-DINOv2[63].750.639.666.808.091.702.604.633.751.148.655.467.495.687.120.731.617.644.785.103
OursDINOv1.734.625.680.854.072.706.621.689.801.108.727.577.627.822.059.761.680.737.851.074
OursDINOv2.864.825.838.931.031.793.747.779.862.077.834.763.779.916.031.850.818.835.923.043
+ +Table 1. Comparison of our methods with recent methods. We compared our proposed methods with competing unsupervised, semi-supervised, and full-supervised methods. Bold indicates the best result in group settings, and underline indicates the second-best result. * denotes the version that reimplemented by us. + +# 4. Experiment + +# 4.1. Experimental Settings + +Trainsets. To fairly compare with the existing works, following [10, 30], we used a combination of 1000 images from the CAMO-Train and 3040 images from COD10K-Train as the training dataset of our experiments. Following the unsupervised learning settings, we do not use any ground-truth label during the training process. + +Testing Sets. We evaluate the performance of our methods on four mainstream COD benchmark testing sets: CHAMELEON with 76 test images, CAMO with 250 test images, COD10K with 2026 test images, and NC4K with 4121 test images. + +Evaluation Protocol. For a fair and comprehensive evaluation, we employ the S-measure $(S_m)$ [8], mean and weighted F-measure $(\mathcal{F}_\beta^m,\mathcal{F}_\beta^\omega)$ [31], mean E-measure $(\mathcal{E}_{\epsilon}^{m})$ [9], mean absolute error $(\mathcal{M})$ [39] as the main metrics. + +Implementation Details. Following [44], we employ the simple background seed methods as our fixed strategy branch to generate the coarse pseudo-label. And following [44, 63], we use the strong self-supervised pretrained backbone DINO [5, 36] as our encoder. The teacher model + +is updated via EMA with a momentum $\eta$ of 0.99. We train the model for 25 epochs. The batch size is set to 32 for each GPU during training. All experiments are implemented with PyTorch 2.1 and run on a machine with Intel(R) Xeon(R) Silver 4214R CPU @ 2.40GHz, 512GiB RAM, and 1 NVIDIA Titan A100-40G GPUs. All experiments use the same random seed. More implementation details will be provided in the supplementary materials. + +# 4.2. Main Results + +Qualitative Analysis. We show visualizations of a series of camouflaged object segmentation masks predicted by our method and related methods in some challenging scenarios. As shown in Fig. 4, we noticed that our method achieves higher segmentation accuracy than existing approaches, with clearer edges and a more detailed representation of the camouflaged object. Note that DINOv2 is employed for all methods that require a backbone to extract features. + +Quantitative Analysis. In Tab. 1, we compared our proposed method's performance with competing USS and UCOD models on four COD test datasets. To further highlight our model's competitiveness, we also compared ex + +![](images/f46a89772bbc1654c1a1fab1dd8c61f22cc6d71261dde5982a42629c3e683b42.jpg) +Figure 4. Visual comparison of our method with other existing methods in challenging scenarios. Our method has clearer and more precise segmentation boundaries and correctly recognizes depth-artifacted objects. + +isting fully-supervised and semi-supervised methods. The results show that our model outperformed all existing USS and UCOD methods across all metrics and datasets, thus achieving State-Of-The-Art (SOTA) performance. Additionally, based on DINOv2 backbone, our model has surpassed several semi-supervised and fully-supervised methods across all four datasets, demonstrating its superior performance, effectiveness, and robustness. + +# 4.3. Ablation Study + +To verify the effectiveness of our methods, we employ $DI-NOv2$ as the backbone and conduct comprehensive ablation studies on the COD10K-Test dataset with 2,026 images. + +Module Ablation. We first conduct ablation experiments on the teacher-student framework (Tea-Stu) and our proposed methods: Adaptive Pseudo-label Mixing (APM) module, Dual-Branch Adversarial (DBA) decoder, and Look-Twice strategy. To thoroughly demonstrate their effectiveness, we replaced the decoder with a single $1 \times 1$ convolutional layer following [44, 63]. For all ablations on the APM module, we used the $1:1$ constant mixing strategy to replace it. The ablation results are shown in Tab. 2. As a result, we find that when only the teacher-student framework is used, the simple $1 \times 1$ convolution fails to learn camouflage-related knowledge adequately. When training + +
SettingsCOD10K (2026)
Tea-StuAPMDBALook-TwiceSm↑Fωβ↑Fmβ↑Emφ↑M↓
✓.626.448.464.658.112
✓✓.633.452.472.668.129
✓.607.409.435.631.129
✓✓.603.404.431.625.133
✓✓.724.582.604.800.054
✓✓✓.788.679.700.879.043
✓✓✓.784.675.692.869.036
✓✓✓✓.834.763.779.916.031
+ +Table 2. Ablation study to evaluate the proposed modules. We retrained our model with different settings on the same learning rate and epochs, then benchmarked on COD10K-testset to validate the effectiveness. + +with a DBA decoder only, the model lacks the guidance of the teacher-student framework, thus easily learning the incorrect knowledge from the low-quality pseudo-labels constructed by fixed-strategy. The proposed Look-Twice mechanism is effective only when the network can produce sufficiently high-quality predictions. If the network has poor localization and segmentation capabilities, Look-Twice will degrade performance. + +Mixing Strategy Ablation. We continue conducting ab- + +
Mixing StrategiesCOD10K (2026)
Sm↑Fωβ↑Fmβ↑Eφ↑M↓
Proportional(1:1) Mixing.707.555.572.763.082
Linear Decay Mixing.809.716.723.888.038
APM.834.763.779.916.031
+ +![](images/4f3057181878e30e094f1cc195b8e5535dee4ae504bb963561870aa8c1dbdb6c.jpg) +Figure 5. Hyper-parameter ablation of small-sized object ratio $\tau$ . The model performs optimally and shows stable results when the value of $\tau$ is set around 0.15. + +Table 3. Ablation to the pseudo-label mixing strategy. We retrained our model with different pseudo-label mixing strategies and benchmarked on the COD10K-Test dataset. + +
Fixed Generation StrategiesCOD10K (2026)
Sm↑Fωβ↑Fmβ↑Emφ↑M↓
Null(Pure Black/White).459.002.001.250.091
Random Perlin Noise.667.489.543.788.072
Background Seed.834.763.779.916.031
+ +Table 4. Fixed Pseudo-label Generation Strategy Ablation. We retrained our model with a replaced fixed strategy and benchmarked it on the COD10K-Test dataset. + +lation studies on the mixing strategies. We compared the proportional(1:1), linear decay, and APM mixing strategies. As shown in Tab. 3, the APM module provides a method for fusing pseudo-labels that is more aligned with the model's training process and achieving optimal performance. + +Hyper-parameter Ablation. We conducted ablation study on the threshold $\tau$ for the small-sized object ratio mentioned in Sec. 3.4. As shown in Fig. 5, when $\tau \approx 0.15$ , the model achieves optimal and stable performance. Therefore, we set $\tau$ to 0.15 in the final model. + +Study on Foreground-sizes. We divided the test set by the proportion of test set prospects at $2\%$ intervals, then benchmarked the performance between our method and some previous SOTA methods[44, 63]. As shown in Fig. 6, our method has significant advantages when dealing with smaller-sized camouflaged targets. + +Fixed Pseudo-label Generation Strategy Ablation. We further investigated the effect of the pseudo-labels constructed by the fixed-strategy. For each input image, we + +![](images/9283571ec3bb47002208c1ee0b09a24fd8100561a05124ff3767814af27a63e6.jpg) + +![](images/e6f62818efa23ab878b50496c92983054a0701f4478add2bbf493b9980bc2bd1.jpg) +Figure 6. Performance comparison for different foreground sizes on COD10K-Test dataset. + +![](images/64f6a5107d93963b83679ae70d0618ca8c9412b0cf8707566ca24fb47f217f9c.jpg) + +![](images/d8e6e54f6f1a3eb10a5753a404a0f77665ca280763698d09042fd0792333e539.jpg) + +generated a random shape of Perlin noise on a plain black background. We used this "fabricated pseudo-label" as well as the pure black and white masks to replace the fixed-strategy pseudo-label during the training process. The results are shown in Tab. 4. When using random noise, the noise may overlap with the foreground region, allowing the network to explore some knowledge related to the camouflaged foreground. However, when using a solid color mask, the model fails to learn the basic definition of the segmentation task to produce valid predictions. + +# 5. Conclusions + +Considering the existing UCOD methods suffer from main issues: 1). The low-quality pseudo-labels constructed by fixed strategies contain substantial noise, making the model prone to fitting incorrect knowledge. 2). The low-resolution pseudo-labels cause severe confusion between the foreground and background. Simple $1 \times 1$ convolution fails to capture and learn the semantic features of camouflaged objects, especially for small-sized samples. To solve these problems, we propose UCOD-DPL in this paper. We introduce a teacher-student framework with an Adaptive Pseudolabel Module, a Dual-Branch Adversarial decoder, and a Look-Twice mechanism. The APM dynamically fuses pseudo-labels from the fixed strategy and teacher model, effectively preventing the model from overfitting the noise and erroneous knowledge. The DBA decoder mines foreground and background knowledge relevant to camouflaged objects while resisting incorrect information from pseudolabels. The Look-Twice refinement mechanism enhances camouflaged object learning and refinement of small objects. We conducted extensive benchmarking on multiple COD test datasets, and the results show that our method outperforms all previous approaches and even reaches the performance level of some fully supervised methods. + +# Acknowledgements + +This work was supported by the National Science Fund for Distinguished Young Scholars (No.62025603), the National Natural Science Foundation of China (No. U21B2037, No. U22B2051, No. U23A20383, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No. 2021J06003, No. 2022J06001). + +# References + +[1] Relja Arandjelović and Andrew Zisserman. Object discovery with a copy-pasting gan. arXiv preprint arXiv:1905.11369, 2019. 2 +[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.3 +[3] Adam Bielski and Paolo Favaro. Emergence of object segmentation in perturbed generative models. Advances in Neural Information Processing Systems, 32, 2019. 2 +[4] TIM Caro. The adaptive significance of coloration in mammals. BioScience, 55(2):125-136, 2005. 1 +[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In The IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 3, 6 +[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 3 +[7] Ming-Ming Cheng, Niloy J Mitra, Xiaolei Huang, Philip HS Torr, and Shi-Min Hu. Global contrast based salient region detection. IEEE transactions on pattern analysis and machine intelligence, 37(3):569-582, 2014. 2 +[8] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In The IEEE/CVF International Conference on Computer Vision, pages 4548-4557, 2017. 6 +[9] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, Ming-Ming Cheng, and Ali Borji. Enhanced-alignment measure for binary foreground map evaluation. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, pages 698-704. AAAI Press, 2018. 6 +[10] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 2777-2787, 2020. 3, 6 +[11] Deng-Ping Fan, Ge-Peng Ji, Tao Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, and Ling Shao. Pranet: Parallel reverse attention network for polyp segmentation. In International conference on medical image computing and computer-assisted intervention, pages 263-273. Springer, 2020. 1 +[12] Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. Concealed object detection. IEEE Transactions on + +Pattern Analysis and Machine Intelligence, 44(10):6024-6042, 2022. 6 +[13] Yuanbin Fu, Jie Ying, Houlei Lv, and Xiaojie Guo. Semi-supervised camouflaged object detection from noisy data. In ACM International Conference on Multimedia, 2024. 6 +[14] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations, 2018. 3 +[15] Spyros Gidaris, Andrei Bursuc, Gilles Puy, Nikos Komodakis, Matthieu Cord, and Patrick Pérez. Obow: Online bag-of-visual-words generation for self-supervised learning. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 6830–6840, 2021. 3 +[16] Stas Goferman, Lihi Zelnik-Manor, and Ayellet Tal. Context-aware saliency detection. IEEE transactions on pattern analysis and machine intelligence, 34(10):1915-1926, 2011. 2 +[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2 +[18] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 3 +[19] Chunming He, Kai Li, Yachao Zhang, Longxiang Tang, Yu-lun Zhang, Zhenhua Guo, and Xiu Li. Camouflaged object detection with feature decomposition and edge reconstruction. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 22046-22055, 2023. 1 +[20] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 9729-9738, 2020. 3 +[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 16000-16009, 2022. 3 +[22] Xiaobin Hu, Shuo Wang, Xuebin Qin, Hang Dai, Wenqi Ren, Donghao Luo, Ying Tai, and Ling Shao. High-resolution iterative feedback network for camouflaged object detection. In Association for the Advancement of Artificial Intelligence, pages 881-889, 2023. 6 +[23] Zhou Huang, Hang Dai, Tian-Zhu Xiang, Shuo Wang, Huai-Xin Chen, Jie Qin, and Huan Xiong. Feature shrinkage pyramid for camouflaged object detection with transformers. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 5557-5566, 2023. 6 +[24] Laurent Itti, Christof Koch, and Ernst Niebur. A model of saliency-based visual attention for rapid scene analysis. IEEE Transactions on pattern analysis and machine intelligence, 20(11):1254-1259, 1998. 2 + +[25] Ge-Peng Ji, Deng-Ping Fan, Yu-Cheng Chou, Dengxin Dai, Alexander Liniger, and Luc Van Gool. Deep gradient learning for efficient camouflaged object detection. Machine Intelligence Research, 20:92-108, 2023. 3 +[26] Tilke Judd, Krista Ehinger, Frédo Durand, and Antonio Torralba. Learning to predict where humans look. In The IEEE/CVF International Conference on Computer Vision, pages 2106-2113. IEEE, 2009. 2 +[27] Xunfa Lai, Zhiyu Yang, Jie Hu, Shengchuan Zhang, Liujuan Cao, Guannan Jiang, Zhiyu Wang, Songan Zhang, and Rongrong Ji. Camoteacher: Dual-rotation consistency learning for semi-supervised camouflaged object detection. In European Conference on Computer Vision, 2024. 1, 6 +[28] Jiaying Lin, Xin Tan, Ke Xu, Lizhuang Ma, and Rynson W. H. Lau. Frequency-aware camouflaged object detection. ACM Trans. Multimedia Comput. Commun. Appl., 19 (2), 2023. 3 +[29] Jianghang Lin, Yunhang Shen, Bingquan Wang, Shaohui Lin, Ke Li, and Liujuan Cao. Weakly supervised open-vocabulary object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3404-3412, 2024. 3 +[30] Naisong Luo, Yuwen Pan, Rui Sun, Tianzhu Zhang, Zhiwei Xiong, and Feng Wu. Camouflaged instance segmentation via explicit de-camouflaging. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 17918-17927, 2023. 6 +[31] Ran Margolin, Lihi Zelnik-Manor, and Ayellet Tal. How to evaluate foreground maps? In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 248-255, 2014. 6 +[32] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 8364-8375, 2022. 3, 6 +[33] Peng Mi, Jianghang Lin, Yiyi Zhou, Yunhang Shen, Gen Luo, Xiaoshuai Sun, Liujuan Cao, Rongrong Fu, Qiang Xu, and Rongrong Ji. Active teacher for semi-supervised object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14482-14491, 2022. 3 +[34] Tam Nguyen, Maximilian Dax, Chaithanya Kumar Mummadi, Nhung Ngo, Thi Hoai Phuong Nguyen, Zhongyu Lou, and Thomas Brox. Deepusps: Deep robust unsupervised saliency prediction via self-supervision. Advances in Neural Information Processing Systems, 32, 2019. 2 +[35] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 3 +[36] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, + +Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 3, 6 +[37] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, 2022. 2, 3, 5, 6 +[38] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoomnext: A unified collaborative pyramid network for camouflaged object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024, 1, 2, 3, 5 +[39] Federico Perazzi, Philipp Krahenbuhl, Yael Pritch, and Alexander Hornung. Saliency filters: Contrast based filtering for salient region detection. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 733-740. IEEE, 2012. 6 +[40] Alec Radford. Improving language understanding by generative pre-training. OpenAI blog, 2018. 3 +[41] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3 +[42] Jingjing Ren, Xiaowei Hu, Lei Zhu, Xuemiao Xu, Yangyang Xu, Weiming Wang, Zijun Deng, and Pheng-Ann Heng. Deep texture-aware features for camouflaged object detection. IEEE Transactions on Circuits and Systems for Video Technology, 33(3):1157-1167, 2023. 3 +[43] Gyungin Shin, Samuel Albanie, and Weidi Xie. Unsupervised salient object detection with spectral cluster voting. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 3971-3980, 2022. 3, 6 +[44] Oriane Simeoni, Chloe Sekkat, Gilles Puy, Antonín Vobecký, Éloi Zablocki, and Patrick Pérez. Unsupervised object localization: Observing the background to discover objects. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 3176-3186, 2023. 1, 2, 3, 4, 6, 7, 8 +[45] Dongyue Sun, Shiyao Jiang, and Lin Qi. Edge-aware mirror network for camouflaged object detection, 2023. 3 +[46] Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. Bert4rec: Sequential recommendation with bidirectional encoder representations from transformer. In Proceedings of the 28th ACM international conference on information and knowledge management, pages 1441-1450, 2019. 3 +[47] Yujia Sun, Geng Chen, Tao Zhou, Yi Zhang, and Nian Liu. Context-aware cross-level fusion network for camouflaged object detection. In The International Joint Conference on Artificial Intelligence, pages 1025–1031, 2021. 6 +[48] Yujia Sun, Shuo Wang, Chenglizhao Chen, and Tian-Zhu Xiang. Boundary-guided camouflaged object detection. In The International Joint Conference on Artificial Intelligence, pages 1335-1341, 2022. 6 +[49] Yanguang Sun, Hanyu Xuan, Jian Yang, and Lei Luo. Glconet: Learning multisource perception representation for camouflaged object detection. IEEE Transactions on Neural Networks and Learning Systems, 2024. 1 + +[50] Domen Tabernik, Samo Šela, Jure Skvarč, and Danijel Skočaj. Segmentation-based deep-learning approach for surface-defect detection. Journal of Intelligent Manufacturing, 31(3):759-776, 2020. 1 +[51] Andrey Voynov, Stanislav Morozov, and Artem Babenko. Object segmentation without labels with large-scale generative models. In International Conference on Machine Learning, pages 10596-10606. PMLR, 2021. 6 +[52] Qingwei Wang, Jinyu Yang, Xiaosheng Yu, Fangyi Wang, Peng Chen, and Feng Zheng. Depth-aided camouflaged object detection. In Proceedings of the 31st ACM International Conference on Multimedia, page 3297–3306, 2023. 3 +[53] Yangtao Wang, Xi Shen, Shell Xu Hu, Yuan Yuan, James L Crowley, and Dominique Vaufreydaz. Self-supervised transformers for unsupervised object discovery using normalized cut. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 14543-14553, 2022. 3, 6 +[54] Yichen Wei, Fang Wen, Wangjiang Zhu, and Jian Sun. Geodesic saliency using background priors. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part III 12, pages 29-42. Springer, 2012. 2 +[55] Among Wu, Rui Liu, Yahong Han, Linchao Zhu, and Yi Yang. Vector-decomposed disentanglement for domain-invariant object detection. The IEEE/CVF International Conference on Computer Vision, 2021. 5 +[56] Mochu Xiang, Jing Zhang, Yunqiu Lv, Aixuan Li, Yiran Zhong, and Yuchao Dai. Exploring depth contribution for camouflaged object detection, 2022. 3 +[57] Qiong Yan, Li Xu, Jianping Shi, and Jiaya Jia. Hierarchical saliency detection. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 1155-1162, 2013. 2 +[58] Fan Yang, Qiang Zhai, Xin Li, Rui Huang, Ao Luo, Hong Cheng, and Deng-Ping Fan. Uncertainty-guided transformer reasoning for camouflaged object detection. In The IEEE/CVF International Conference on Computer Vision, pages 4146-4155, 2021. 6 +[59] Siyuan Yao, Hao Sun, Tian-Zhu Xiang, Xiao Wang, and Xiaochun Cao. Hierarchical graph interaction transformer with dynamic token clustering for camouflaged object detection. IEEE Transactions on Image Processing, 2024. 1 +[60] Bowen Yin, Xuying Zhang, Deng-Ping Fan, Shaohui Jiao, Ming-Ming Cheng, Luc Van Gool, and Qibin Hou. Camoformer: Masked separable attention for camouflaged object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1 +[61] Qiang Zhai, Xin Li, Fan Yang, Chenglizhao Chen, Hong Cheng, and Deng-Ping Fan. Mutual graph learning for camouflaged object detection. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 12997-13007, 2021. 6 +[62] Jing Zhang, Tong Zhang, Yuchao Dai, Mehrtash Harandi, and Richard Hartley. Deep unsupervised saliency detection: A multiple noisy labeling perspective. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 9029-9038, 2018. 2 +[63] Yi Zhang and Chengyi Wu. Unsupervised camouflaged object segmentation as domain adaptation. In The IEEE/CVF + +International Conference on Computer Vision, pages 4334-4344, 2023. 1, 3, 4, 6, 7, 8 +[64] Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, and Nicu Sebe. Bilateral reference for high-resolution dichotomous image segmentation. CAAI Artificial Intelligence Research, 3:9150038, 2024. 6 +[65] Yijie Zhong, Bo Li, Lv Tang, Senyun Kuang, Shuang Wu, and Shouhong Ding. Detecting camouflaged object in frequency domain. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 4494-4503, 2022. 3 +[66] Hongwei Zhu, Peng Li, Haoran Xie, Xu Yan, Dong Liang, Dapeng Chen, Mingqiang Wei, and Jing Qin. I can find you! boundary-guided separated attention network for camouflaged object detection. In Association for the Advancement of Artificial Intelligence, 2022. 3 +[67] Jinchao Zhu, Xiaoyu Zhang, Shuo Zhang, and Junnan Liu. Inferring camouflaged objects by texture-aware interactive guidance network. In Association for the Advancement of Artificial Intelligence, 2021. 3 +[68] Wangjiang Zhu, Shuang Liang, Yichen Wei, and Jian Sun. Saliency optimization from robust background detection. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 2814-2821, 2014. 2 +[69] Mingchen Zhuge, Xiankai Lu, Yiyou Guo, Zhihua Cai, and Shuhan Chen. Cubenet: X-shape connection for camouflaged object detection. Pattern Recognition, 127:108644, 2022. 3 \ No newline at end of file diff --git a/2025/UCOD-DPL_ Unsupervised Camouflaged Object Detection via Dynamic Pseudo-label Learning/images.zip b/2025/UCOD-DPL_ Unsupervised Camouflaged Object Detection via Dynamic Pseudo-label Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..437d1c13ddefd85dcf8b69a820c294b0b8a0a601 --- /dev/null +++ b/2025/UCOD-DPL_ Unsupervised Camouflaged Object Detection via Dynamic Pseudo-label Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ee3677afdeeb463d6b1a6fdabdc33b3ecc15ad9c84b473f7729e0b41cd0ddbc +size 961567 diff --git a/2025/UCOD-DPL_ Unsupervised Camouflaged Object Detection via Dynamic Pseudo-label Learning/layout.json b/2025/UCOD-DPL_ Unsupervised Camouflaged Object Detection via Dynamic Pseudo-label Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c2dedef4d72b5f40d189a88ae0d72d18db8d227a --- /dev/null +++ b/2025/UCOD-DPL_ Unsupervised Camouflaged Object Detection via Dynamic Pseudo-label Learning/layout.json @@ -0,0 +1,10548 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 125, + 103, + 486, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 103, + 486, + 140 + ], + "spans": [ + { + "bbox": [ + 125, + 103, + 486, + 140 + ], + "type": "text", + "content": "UCOD-DPL: Unsupervised Camouflaged Object Detection via Dynamic Pseudo-label Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 162, + 482, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 162, + 482, + 175 + ], + "spans": [ + { + "bbox": [ + 127, + 162, + 482, + 175 + ], + "type": "text", + "content": "Weiqi Yan" + }, + { + "bbox": [ + 127, + 162, + 482, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 162, + 482, + 175 + ], + "type": "text", + "content": ", Lvhai Chen" + }, + { + "bbox": [ + 127, + 162, + 482, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 162, + 482, + 175 + ], + "type": "text", + "content": ", Huaijia Ko" + }, + { + "bbox": [ + 127, + 162, + 482, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 162, + 482, + 175 + ], + "type": "text", + "content": ", Shengchuan Zhang" + }, + { + "bbox": [ + 127, + 162, + 482, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 127, + 162, + 482, + 175 + ], + "type": "text", + "content": ", Yan Zhang" + }, + { + "bbox": [ + 127, + 162, + 482, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 162, + 482, + 175 + ], + "type": "text", + "content": ", Liujuan Cao" + }, + { + "bbox": [ + 127, + 162, + 482, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 149, + 176, + 461, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 176, + 461, + 189 + ], + "spans": [ + { + "bbox": [ + 149, + 176, + 461, + 189 + ], + "type": "text", + "content": "1 Key Laboratory of Multimedia Trusted Perception and Efficient Computing," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 190, + 451, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 190, + 451, + 202 + ], + "spans": [ + { + "bbox": [ + 159, + 190, + 451, + 202 + ], + "type": "text", + "content": "Ministry of Education of China, Xiamen University, 361005, P.R. China." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 205, + 477, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 205, + 477, + 217 + ], + "spans": [ + { + "bbox": [ + 129, + 205, + 477, + 217 + ], + "type": "text", + "content": "weiqi_yan@outlook.com lvhaichen2002@gmail.com zsc_2016@xmu.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 279, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 279, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 53, + 279, + 296, + 639 + ], + "type": "text", + "content": "Unsupervised Camouflaged Object Detection (UCOD) has gained attention since it doesn't need to rely on extensive pixel-level labels. Existing UCOD methods typically generate pseudo-labels using fixed strategies and train " + }, + { + "bbox": [ + 53, + 279, + 296, + 639 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 53, + 279, + 296, + 639 + ], + "type": "text", + "content": " convolutional layers as a simple decoder, leading to low performance compared to fully-supervised methods. We emphasize two drawbacks in these approaches: 1). The model is prone to fitting incorrect knowledge due to the pseudolabel containing substantial noise. 2). The simple decoder fails to capture and learn the semantic features of camouflaged objects, especially for small-sized objects, due to the low-resolution pseudo-labels and severe confusion between foreground and background pixels. To this end, we propose a UCOD method with a teacher-student framework via Dynamic Pseudo-label Learning called UCOD-DPL, which contains an Adaptive Pseudo-label Module (APM), a Dual-Branch Adversarial (DBA) decoder, and a Look-Twice mechanism. The APM module adaptively combines pseudo-labels generated by fixed strategies and the teacher model to prevent the model from overfitting incorrect knowledge while preserving the ability for self-correction; the DBA decoder takes adversarial learning of different segmentation objectives, guides the model to overcome the foreground-background confusion of camouflaged objects, and the Look-Twice mechanism mimics the human tendency to zoom in on camouflaged objects and performs secondary refinement on small-sized objects. Extensive experiments show that our method demonstrates outstanding performance, even surpassing some existing fully supervised methods. The code is available now1." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 245, + 555, + 406 + ], + "blocks": [ + { + "bbox": [ + 318, + 245, + 555, + 406 + ], + "lines": [ + { + "bbox": [ + 318, + 245, + 555, + 406 + ], + "spans": [ + { + "bbox": [ + 318, + 245, + 555, + 406 + ], + "type": "image", + "image_path": "5ac8efa5b68b91a90ea10acb677ac58490f6f67c51ff90cef9129cf4c0b41b87.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 418, + 555, + 440 + ], + "lines": [ + { + "bbox": [ + 313, + 418, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 555, + 440 + ], + "type": "text", + "content": "Figure 1. Comparison between our method and previous pseudolabel-based methods." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 463, + 394, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 463, + 394, + 475 + ], + "spans": [ + { + "bbox": [ + 315, + 463, + 394, + 475 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 484, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 484, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 312, + 484, + 555, + 568 + ], + "type": "text", + "content": "\"Camouflage\" originates from the natural behavior of objects blending into their surroundings using similar textures and colors to evade \"predators\"[4]. Camouflaged Object Detection (COD) is a challenging semantic segmentation task aimed at learning to segment objects that are visually hidden in their backgrounds, which has significant applications in several important fields [11, 50]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "Although fully-supervised methods have achieved significant progress in the COD tasks [19, 38, 49, 59, 60], they heavily rely on large-scale human annotations for training. However, it is difficult to get pixel-level human labels for camouflaged objects due to their complex visual attributes [27]. Therefore, this reliance on annotations has sparked increasing interest in the task of unsupervised camouflage object detection (UCOD) among researchers. However, existing UCOD methods typically use pseudo-labels generated by fixed strategies to train a single " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": " convolutional layer [44, 63]. By investing in existing methods, we identified two main issues: 1). The pseudo-label con" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 143, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 143, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 143, + 703 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 67, + 703, + 256, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 256, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 256, + 712 + ], + "type": "text", + "content": "1https://github.com/Heartfirey/UCOD-DPL" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30365" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 84, + 130, + 129 + ], + "blocks": [ + { + "bbox": [ + 83, + 72, + 129, + 80 + ], + "lines": [ + { + "bbox": [ + 83, + 72, + 129, + 80 + ], + "spans": [ + { + "bbox": [ + 83, + 72, + 129, + 80 + ], + "type": "text", + "content": "Small-sized" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 78, + 84, + 130, + 129 + ], + "lines": [ + { + "bbox": [ + 78, + 84, + 130, + 129 + ], + "spans": [ + { + "bbox": [ + 78, + 84, + 130, + 129 + ], + "type": "image", + "image_path": "c037aa66092f0a1f0d582e430edd9af8efac42e7e807c6839a8cba512a77e4a9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 134, + 84, + 185, + 129 + ], + "blocks": [ + { + "bbox": [ + 141, + 72, + 176, + 81 + ], + "lines": [ + { + "bbox": [ + 141, + 72, + 176, + 81 + ], + "spans": [ + { + "bbox": [ + 141, + 72, + 176, + 81 + ], + "type": "text", + "content": "Huge-sized" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 134, + 84, + 185, + 129 + ], + "lines": [ + { + "bbox": [ + 134, + 84, + 185, + 129 + ], + "spans": [ + { + "bbox": [ + 134, + 84, + 185, + 129 + ], + "type": "image", + "image_path": "0896fe4269343f73288e1be4ac1ae579ca458ea16653dff23238022e09984fdb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 189, + 84, + 239, + 129 + ], + "blocks": [ + { + "bbox": [ + 201, + 72, + 227, + 81 + ], + "lines": [ + { + "bbox": [ + 201, + 72, + 227, + 81 + ], + "spans": [ + { + "bbox": [ + 201, + 72, + 227, + 81 + ], + "type": "text", + "content": "Multiple" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 189, + 84, + 239, + 129 + ], + "lines": [ + { + "bbox": [ + 189, + 84, + 239, + 129 + ], + "spans": [ + { + "bbox": [ + 189, + 84, + 239, + 129 + ], + "type": "image", + "image_path": "abdac31dd9a5aea9a1280a5e1ae313a3bc47d84d78d00663bfee90522fccf1d0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 243, + 83, + 294, + 129 + ], + "blocks": [ + { + "bbox": [ + 250, + 72, + 288, + 81 + ], + "lines": [ + { + "bbox": [ + 250, + 72, + 288, + 81 + ], + "spans": [ + { + "bbox": [ + 250, + 72, + 288, + 81 + ], + "type": "text", + "content": "Uncertainty" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 243, + 83, + 294, + 129 + ], + "lines": [ + { + "bbox": [ + 243, + 83, + 294, + 129 + ], + "spans": [ + { + "bbox": [ + 243, + 83, + 294, + 129 + ], + "type": "image", + "image_path": "560531f721e64c89cd6d43697af21c7e59b8e59b058534ebda9cdd426ec4e159.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 59, + 177, + 74, + 221 + ], + "blocks": [ + { + "bbox": [ + 59, + 177, + 74, + 221 + ], + "lines": [ + { + "bbox": [ + 59, + 177, + 74, + 221 + ], + "spans": [ + { + "bbox": [ + 59, + 177, + 74, + 221 + ], + "type": "image", + "image_path": "ef753cf78e411b7724b52ed8462e5e3657c85f7a4a772376ef1b1957147ae3f7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 230, + 74, + 259 + ], + "lines": [ + { + "bbox": [ + 59, + 230, + 74, + 259 + ], + "spans": [ + { + "bbox": [ + 59, + 230, + 74, + 259 + ], + "type": "text", + "content": "CocosDg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 78, + 130, + 130, + 175 + ], + "blocks": [ + { + "bbox": [ + 61, + 98, + 72, + 117 + ], + "lines": [ + { + "bbox": [ + 61, + 98, + 72, + 117 + ], + "spans": [ + { + "bbox": [ + 61, + 98, + 72, + 117 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 78, + 130, + 130, + 175 + ], + "lines": [ + { + "bbox": [ + 78, + 130, + 130, + 175 + ], + "spans": [ + { + "bbox": [ + 78, + 130, + 130, + 175 + ], + "type": "image", + "image_path": "37c5dc4ac4c08502e20ef8140b38c7803a50f816266c7ae11c483dfb75e4d2e9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 78, + 176, + 130, + 221 + ], + "blocks": [ + { + "bbox": [ + 78, + 176, + 130, + 221 + ], + "lines": [ + { + "bbox": [ + 78, + 176, + 130, + 221 + ], + "spans": [ + { + "bbox": [ + 78, + 176, + 130, + 221 + ], + "type": "image", + "image_path": "549641dd98f921a8c5679018e0f06c1935125b49424aeee67aee7eb3b68cb56b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 78, + 222, + 130, + 267 + ], + "blocks": [ + { + "bbox": [ + 78, + 222, + 130, + 267 + ], + "lines": [ + { + "bbox": [ + 78, + 222, + 130, + 267 + ], + "spans": [ + { + "bbox": [ + 78, + 222, + 130, + 267 + ], + "type": "image", + "image_path": "e5e2045dd2258de7408624fdca3dbac1ff29b150781030371efd01eba7bd0783.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 276, + 295, + 342 + ], + "lines": [ + { + "bbox": [ + 55, + 276, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 295, + 342 + ], + "type": "text", + "content": "Figure 2. Some examples of low-quality pseudo-labels generated using fixed strategies. Based on features extracted by DINOv2, we construct pseudo-labels using the Background-Seed from FOUND [44] and the MaskCut in UCOD-DA [44], in challenging scenarios (e.g. small-sized objects, huge-sized objects, multiple instances, and uncertain objects) for visualization." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 134, + 130, + 185, + 175 + ], + "blocks": [ + { + "bbox": [ + 134, + 130, + 185, + 175 + ], + "lines": [ + { + "bbox": [ + 134, + 130, + 185, + 175 + ], + "spans": [ + { + "bbox": [ + 134, + 130, + 185, + 175 + ], + "type": "image", + "image_path": "c28da8f540c7727fad1f22bf991d391ad860722a319d80e4aabd6c591d301dbc.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 134, + 175, + 185, + 221 + ], + "blocks": [ + { + "bbox": [ + 134, + 175, + 185, + 221 + ], + "lines": [ + { + "bbox": [ + 134, + 175, + 185, + 221 + ], + "spans": [ + { + "bbox": [ + 134, + 175, + 185, + 221 + ], + "type": "image", + "image_path": "e5b5861e533732c198bbd4e4bccf4f4aba2d9164fa40bf4fd0d03a1f5b2b3cda.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 134, + 222, + 185, + 267 + ], + "blocks": [ + { + "bbox": [ + 134, + 222, + 185, + 267 + ], + "lines": [ + { + "bbox": [ + 134, + 222, + 185, + 267 + ], + "spans": [ + { + "bbox": [ + 134, + 222, + 185, + 267 + ], + "type": "image", + "image_path": "b126565ffeb8f8d829db44757da0ba6b8fda433bc0f85aa3a2fa20970e948cc0.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 189, + 130, + 239, + 175 + ], + "blocks": [ + { + "bbox": [ + 189, + 130, + 239, + 175 + ], + "lines": [ + { + "bbox": [ + 189, + 130, + 239, + 175 + ], + "spans": [ + { + "bbox": [ + 189, + 130, + 239, + 175 + ], + "type": "image", + "image_path": "de55faa6b59c87878981166000c394e2fc556aa8ae7a3cb5c9697577b62548f5.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 189, + 175, + 239, + 221 + ], + "blocks": [ + { + "bbox": [ + 189, + 175, + 239, + 221 + ], + "lines": [ + { + "bbox": [ + 189, + 175, + 239, + 221 + ], + "spans": [ + { + "bbox": [ + 189, + 175, + 239, + 221 + ], + "type": "image", + "image_path": "dfc2be28101e075b0e2865dc8e42e2cd9e30732770969d56bacbacf0b6e53e46.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 189, + 222, + 239, + 267 + ], + "blocks": [ + { + "bbox": [ + 189, + 222, + 239, + 267 + ], + "lines": [ + { + "bbox": [ + 189, + 222, + 239, + 267 + ], + "spans": [ + { + "bbox": [ + 189, + 222, + 239, + 267 + ], + "type": "image", + "image_path": "301cc198cda7d14e37dbf9e905d9c708da16277b655a67962eae80afb20006d6.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 243, + 130, + 294, + 175 + ], + "blocks": [ + { + "bbox": [ + 243, + 130, + 294, + 175 + ], + "lines": [ + { + "bbox": [ + 243, + 130, + 294, + 175 + ], + "spans": [ + { + "bbox": [ + 243, + 130, + 294, + 175 + ], + "type": "image", + "image_path": "c22fe784f1af5024970c4e127c8a598aeb840a1a88013bed875d5ee63a0fde84.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 243, + 175, + 294, + 221 + ], + "blocks": [ + { + "bbox": [ + 243, + 175, + 294, + 221 + ], + "lines": [ + { + "bbox": [ + 243, + 175, + 294, + 221 + ], + "spans": [ + { + "bbox": [ + 243, + 175, + 294, + 221 + ], + "type": "image", + "image_path": "c647237a4427a22322048b2573886a66ea50a195eb5f28afeeaa8c7b08d7e86e.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 243, + 222, + 294, + 267 + ], + "blocks": [ + { + "bbox": [ + 243, + 222, + 294, + 267 + ], + "lines": [ + { + "bbox": [ + 243, + 222, + 294, + 267 + ], + "spans": [ + { + "bbox": [ + 243, + 222, + 294, + 267 + ], + "type": "image", + "image_path": "26683d13d030418266cb11fa7b2ac1ead541e51029cfca245385d7da1ef905cc.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 365, + 295, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 472 + ], + "type": "text", + "content": "structed by fixed strategies contains substantial noise, which makes the model prone to fitting incorrect knowledge. 2). These pseudo-labels have low resolution and severe confusion between foreground and background pixels, and simple " + }, + { + "bbox": [ + 55, + 365, + 295, + 472 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 55, + 365, + 295, + 472 + ], + "type": "text", + "content": " convolution fails to capture and learn the semantic features of camouflaged objects, especially for small-sized objects. Consequently, their segmentation results fall far behind those of fully supervised methods, which limits their application in real-world scenarios." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 475, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 713 + ], + "type": "text", + "content": "In this paper, we propose a brand-new model called UCOD-DPL. As shown in Fig. 1, we introduce a teacher-student framework consisting of an Adaptive Pseudo-label Mixing (APM) module, a Dual-Branch Adversarial (DBA) decoder, and a Look-Twice refinement strategy. As illustrated in Fig. 2, pseudo-labels constructed by fixed strategies are generally low-quality and contain substantial noise. To prevent the model from fitting uncorrectable erroneous knowledge in the pseudo-labels, we design the APM module with a discriminator and a dynamic scoring mechanism to mix the pseudo-labels from the fixed strategy and the teacher branch adaptively. To help the model overcome foreground-background confusion, we propose a DBA decoder inspired by predator-prey dynamics. This decoder uses different segmentation targets (foreground and background) along with an orthogonal loss of features to deeply mine and learn distinctive features of the foreground and background. Additionally, inspired by [37, 38], we mimic the zooming-in behavior humans use when examining camouflaged images by performing secondary refinement on" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "small-sized camouflaged objects to enhance the model's segmentation accuracy when dealing with small targets." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 98, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 98, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 98, + 553, + 156 + ], + "type": "text", + "content": "In the experiment, we train our model on the commonly used COD training set and evaluate it on multiple test sets. The results demonstrate that our method achieves outstanding performance, even comparable to some fully supervised COD methods when using DINOv2 as the backbone." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 326, + 158, + 533, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 158, + 533, + 169 + ], + "spans": [ + { + "bbox": [ + 326, + 158, + 533, + 169 + ], + "type": "text", + "content": "Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 314, + 171, + 553, + 386 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 314, + 171, + 553, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 171, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 314, + 171, + 553, + 219 + ], + "type": "text", + "content": "1. We propose a novel UCOD model called UCOD-DPL, and extensive experiments demonstrate that our method exhibits superior performance, reaching the performance level of some fully supervised algorithms." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 314, + 219, + 553, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 219, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 314, + 219, + 553, + 266 + ], + "type": "text", + "content": "2. We introduce a teacher-student framework with an APM module. By using the APM, we ensure that the model can learn knowledge from pseudo-labels while preserving the ability to identify and correct the erroneous." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 314, + 266, + 553, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 266, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 314, + 266, + 553, + 326 + ], + "type": "text", + "content": "3. To address the issues of low-resolution pseudo-labels and foreground-background confusion, we design a DBA decoder that uses a foreground-background adversarial task, enabling the model to learn the camouflaged features of the foreground adaptively." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 314, + 327, + 553, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 327, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 314, + 327, + 553, + 386 + ], + "type": "text", + "content": "4. To enhance the model's ability to segment small-sized objects, we design a Look-Twice refinement mechanism. This mechanism mimics human behavior by performing a secondary segmentation on smaller camouflaged objects, thereby refining the segmentation results." + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 399, + 405, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 399, + 405, + 411 + ], + "spans": [ + { + "bbox": [ + 314, + 399, + 405, + 411 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 314, + 420, + 513, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 420, + 513, + 433 + ], + "spans": [ + { + "bbox": [ + 314, + 420, + 513, + 433 + ], + "type": "text", + "content": "2.1. Unsupervised Semantic Segmentation" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 438, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 555, + 713 + ], + "type": "text", + "content": "Unsupervised Semantic Segmentation (USS) aims to train a model without any human annotations to produce the semantic segmentation mask. Before the surge of deep learning methods, researchers relied on numerous handcrafted methods [7, 16, 57, 68] that used one or more priors related to foreground regions in the image to generate segmentation masks, such as center[26], contrast[24], and boundary[54]. In deep learning-based methods, some approaches utilize generative models [17] to produce a segmentation mask, which will be used to synthesize a realistic image by copying the corresponding region into a background which is either synthesized [3] or taken from a different image [1]. Another category of methods leverages pseudo-labels constructed by handcrafted fixed strategies to train the detector. [62] adopts diverse noisy pseudo-labels via distinct unsupervised handcrafted methods to train the detector to predict a saliency map free from noise in labels. And [34, 44] proposed to refine the pseudo-labels produced by fixed-strategy by training the model via a self-supervised iterative refinement manner. These methods adopt a self-supervision manner that highly relies on pseudo-labels generated directly from one or multiple fixed strategies. If the pseudo-labels contain substantial noise, these methods remain prone to" + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30366" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "fitting incorrect knowledge, which is particularly evident when dealing with camouflaged objects." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 106, + 222, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 106, + 222, + 118 + ], + "spans": [ + { + "bbox": [ + 55, + 106, + 222, + 118 + ], + "type": "text", + "content": "2.2. Camouflaged Object Detection" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 124, + 296, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 124, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 124, + 296, + 411 + ], + "type": "text", + "content": "Camouflaged Object Detection (COD) is a complex task aimed at segmenting objects that are intentionally designed to merge seamlessly with their environment. This task has an extensive history in its domain. Existing fully-supervised COD methods [10, 37, 38, 69] have achieved excellent performance by training on large-scale annotated datasets. These methods utilize the rich information (e.g. boundary [25, 45, 66], texture [25, 42, 67] and other information like frequency domain and depth [28, 52, 56, 65]) provided by pixel-level annotations and design a series of sophisticated strategies to extract unique visual characteristics of camouflaged objects. However, the extensive pixel-level annotations are hardly gained, which has drawn the attention of researchers to Unsupervised Camouflaged Object Detection (UCOD) and USS methods [63]. These methods typically rely on pseudo-labels generated through fixed strategies to supervise model learning for segmentation. However, due to limitations in algorithmic complexity and robustness, these fixed strategies struggle to work effectively at high resolutions and easily introduce significant noise. Consequently, these methods are prone to learning incorrect noise from the pseudo-labels, leading to suboptimal segmentation performance and limiting the application in the real world of UCOD models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 420, + 196, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 196, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 196, + 433 + ], + "type": "text", + "content": "2.3. Self-supervised Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 437, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 714 + ], + "type": "text", + "content": "Self-Supervised Learning (SSL) has revolutionized the way models leverage unlabeled data. Among current SSL approaches, some methods adopt contrastive learning to learn image feature representations. MoCo [20] utilizes a dynamic memory bank to maintain a large pool of negative samples; SimCLR [6] employs various data augmentations and a projection head to maximize similarities between differently augmented views of the same instance; and BYOL [18] constructs a teacher-student framework with the teacher model's parameters are updated by Exponential Moving Average (EMA) of the student model. Following BYOL, DINO [5] applied two interactive encoders sharing the same structure but with different parameter sets and update strategies. Inspired by self-supervised pretraining techniques in natural language processing [40, 41, 46], other SSL methods design pretext tasks (e.g. colorization, rotation prediction, and patch reconstruction) on unlabeled images [5, 6, 14, 15, 20, 29, 33, 35]. BeiT [2] adopts a Transformer-based architecture to reconstruct masked image patches, MAE [21] implements a masked autoencoder that efficiently reconstructs the missing parts of an image. In this paper, we employ DINOv1 [5] and DINOv2 [36] to extract robust and generalized image features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 71, + 397, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 397, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 397, + 85 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 91, + 430, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 430, + 102 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 430, + 102 + ], + "type": "text", + "content": "3.1. Overall Framework" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 108, + 555, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 555, + 178 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 555, + 178 + ], + "type": "text", + "content": "The structure of our proposed method is shown in Fig. 3, which includes a teacher-student framework with an Adaptive Pseudo-label Mixing (APM) module, a Dual-Branch Adversarial (DBA) decoder, and a Look-Twice strategy. Both the teacher and student models are configured as DBA decoders." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 180, + 555, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 555, + 386 + ], + "type": "text", + "content": "Firstly, the " + }, + { + "bbox": [ + 313, + 180, + 555, + 386 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 180, + 555, + 386 + ], + "type": "text", + "content": "-th unlabeled input image passes through the self-supervised pretrained backbone to obtain high-level semantic image feature " + }, + { + "bbox": [ + 313, + 180, + 555, + 386 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 313, + 180, + 555, + 386 + ], + "type": "text", + "content": ". We then employ the background seed method to generate a low-resolution fixed-strategy pseudo-label " + }, + { + "bbox": [ + 313, + 180, + 555, + 386 + ], + "type": "inline_equation", + "content": "\\hat{P}_i^{fs}" + }, + { + "bbox": [ + 313, + 180, + 555, + 386 + ], + "type": "text", + "content": ", which utilizes pixel-level similarity comparison and was proposed in [44]. Simultaneously, the image feature " + }, + { + "bbox": [ + 313, + 180, + 555, + 386 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 313, + 180, + 555, + 386 + ], + "type": "text", + "content": " is fed into a teacher-student model. The teacher model generates a higher-resolution teacher's pseudo label " + }, + { + "bbox": [ + 313, + 180, + 555, + 386 + ], + "type": "inline_equation", + "content": "\\hat{P}_i^t" + }, + { + "bbox": [ + 313, + 180, + 555, + 386 + ], + "type": "text", + "content": ", and the student model produces the foreground and background predicted mask " + }, + { + "bbox": [ + 313, + 180, + 555, + 386 + ], + "type": "inline_equation", + "content": "\\hat{Y}_i^{FG}, \\hat{Y}_i^{BG}" + }, + { + "bbox": [ + 313, + 180, + 555, + 386 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 313, + 180, + 555, + 386 + ], + "type": "inline_equation", + "content": "\\hat{P}_i^{fs}" + }, + { + "bbox": [ + 313, + 180, + 555, + 386 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 180, + 555, + 386 + ], + "type": "inline_equation", + "content": "\\hat{P}_i^t" + }, + { + "bbox": [ + 313, + 180, + 555, + 386 + ], + "type": "text", + "content": " are fed into the APM module to construct the mixed dynamic pseudo-label " + }, + { + "bbox": [ + 313, + 180, + 555, + 386 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 313, + 180, + 555, + 386 + ], + "type": "text", + "content": ". Finally, we use the mixed dynamic pseudo-label " + }, + { + "bbox": [ + 313, + 180, + 555, + 386 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 313, + 180, + 555, + 386 + ], + "type": "text", + "content": " to supervise the final foreground and background predicted masks " + }, + { + "bbox": [ + 313, + 180, + 555, + 386 + ], + "type": "inline_equation", + "content": "\\hat{Y}_i^{FG}, \\hat{Y}_i^{BG}" + }, + { + "bbox": [ + 313, + 180, + 555, + 386 + ], + "type": "text", + "content": ". We further apply a Look-Twice strategy to perform a secondary refinement on small-sized objects to improve the model's segmentation performance on small-sized samples." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 392, + 528, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 392, + 528, + 404 + ], + "spans": [ + { + "bbox": [ + 314, + 392, + 528, + 404 + ], + "type": "text", + "content": "3.2. Adaptive Pseudo-Label Merging Module" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 409, + 554, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 554, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 554, + 505 + ], + "type": "text", + "content": "Existing unsupervised semantic segmentation methods typically adopt a fixed-strategy (e.g. similarity comparison [32, 43, 53]) or utilize a dual-branch model (e.g. fixed-strategy or learnable branch [44, 63]) structure for self-supervised learning to generate segmentation masks. However, we believe the performance of these methods is constrained by the strategy's design or the pseudo-labels' quality, preventing the model from achieving higher accuracy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 506, + 554, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 554, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 554, + 564 + ], + "type": "text", + "content": "We introduce a discriminator " + }, + { + "bbox": [ + 313, + 506, + 554, + 564 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 506, + 554, + 564 + ], + "type": "text", + "content": ", which inputs a foreground segmentation mask (i.e. fixed-strategy pseudo-label " + }, + { + "bbox": [ + 313, + 506, + 554, + 564 + ], + "type": "inline_equation", + "content": "\\hat{P}_i^{fs}" + }, + { + "bbox": [ + 313, + 506, + 554, + 564 + ], + "type": "text", + "content": " or student's foreground predicted mask " + }, + { + "bbox": [ + 313, + 506, + 554, + 564 + ], + "type": "inline_equation", + "content": "\\hat{Y}_i^{FG}" + }, + { + "bbox": [ + 313, + 506, + 554, + 564 + ], + "type": "text", + "content": ") and outputs the probability that the mask originates from the fixed-strategy branch:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 365, + 570, + 553, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 570, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 365, + 570, + 553, + 586 + ], + "type": "interline_equation", + "content": "\\hat {y} _ {i} ^ {p 1} = \\mathcal {D} \\left(\\hat {P} _ {i} ^ {f s}\\right), \\hat {y} _ {i} ^ {p 2} = \\mathcal {D} \\left(\\hat {Y} _ {i} ^ {F G}\\right), \\tag {1}", + "image_path": "3ec339c1407b2905cad450c9f3b16b8502f3712ffff4514b1b2ce210d9d43f34.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 592, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 555, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 592, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_i^{p1},\\hat{\\mathbf{y}}_i^{p2}" + }, + { + "bbox": [ + 313, + 592, + 555, + 629 + ], + "type": "text", + "content": " denote the predicted probability that the input mask belongs to the fixed-strategy branch. We then introduce a scoring function " + }, + { + "bbox": [ + 313, + 592, + 555, + 629 + ], + "type": "inline_equation", + "content": "S(\\cdot)" + }, + { + "bbox": [ + 313, + 592, + 555, + 629 + ], + "type": "text", + "content": " with a temporal constraint:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 639, + 561, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 561, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 561, + 677 + ], + "type": "interline_equation", + "content": "\\mathcal {S} \\left(\\hat {\\mathrm {y}} _ {i} ^ {p 1}, \\hat {\\mathrm {y}} _ {i} ^ {p 2}\\right) = \\operatorname {C L I P} \\left(\\frac {t}{T} + \\frac {1}{2} \\left(1 + \\cos \\left(\\pi \\times \\left| \\hat {\\mathrm {y}} _ {i} ^ {p 1} - \\hat {\\mathrm {y}} _ {i} ^ {p 2} \\right|\\right)\\right) _ {0} ^ {1}, \\right. \\tag {2}", + "image_path": "c7c36ee798f5a160bafb15108ac9202ea72377515bd263475f149c873ae984d9.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "t, T" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " denote the current and total training epochs, " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}(x)_0^1" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " denotes truncating " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " so that it lies within the interval [0, 1]. In the early stage of training, the teacher model" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30367" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 553, + 232 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 232 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 232 + ], + "type": "image", + "image_path": "500a56bfabcc9538ae36538d3af179e5afd19ea4959a3c6175fd76829ca5b5fe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 236, + 554, + 358 + ], + "blocks": [ + { + "bbox": [ + 61, + 236, + 554, + 358 + ], + "lines": [ + { + "bbox": [ + 61, + 236, + 554, + 358 + ], + "spans": [ + { + "bbox": [ + 61, + 236, + 554, + 358 + ], + "type": "image", + "image_path": "f58d65e25dfd0c36d887656828ddc15dc7851f6fec0e0893f7195b056fc37be1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 372, + 555, + 394 + ], + "lines": [ + { + "bbox": [ + 55, + 372, + 555, + 394 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 555, + 394 + ], + "type": "text", + "content": "Figure 3. The main framework of our proposed method. The model contains a teacher-student framework, an Adaptive Psdudo-label Mixing (APM) module, a Dual-Branch Adversarial (DBA) decoder, and a Look-Twice strategy." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 415, + 296, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 602 + ], + "type": "text", + "content": "has not yet developed sufficient localization and basic semantic segmentation capabilities for camouflaged objects, making it unable to supervise the student model effectively. At this point, there is a significant difference between fixed-strategy pseudo-label " + }, + { + "bbox": [ + 55, + 415, + 296, + 602 + ], + "type": "inline_equation", + "content": "\\hat{P}_i^{fs}" + }, + { + "bbox": [ + 55, + 415, + 296, + 602 + ], + "type": "text", + "content": " and the student's foreground predicted mask " + }, + { + "bbox": [ + 55, + 415, + 296, + 602 + ], + "type": "inline_equation", + "content": "\\hat{Y}_i^{FG}" + }, + { + "bbox": [ + 55, + 415, + 296, + 602 + ], + "type": "text", + "content": ". To help the model learn the fundamental foreground localization and segmentation tasks, we use a higher proportion of " + }, + { + "bbox": [ + 55, + 415, + 296, + 602 + ], + "type": "inline_equation", + "content": "\\hat{P}_i^{fs}" + }, + { + "bbox": [ + 55, + 415, + 296, + 602 + ], + "type": "text", + "content": " to supervise model training. Later in the training process, as the " + }, + { + "bbox": [ + 55, + 415, + 296, + 602 + ], + "type": "inline_equation", + "content": "\\hat{Y}_i^{FG}" + }, + { + "bbox": [ + 55, + 415, + 296, + 602 + ], + "type": "text", + "content": " gradually approaches " + }, + { + "bbox": [ + 55, + 415, + 296, + 602 + ], + "type": "inline_equation", + "content": "\\hat{P}_i^{fs}" + }, + { + "bbox": [ + 55, + 415, + 296, + 602 + ], + "type": "text", + "content": ", we increase the proportion of the teacher's pseudo-label " + }, + { + "bbox": [ + 55, + 415, + 296, + 602 + ], + "type": "inline_equation", + "content": "\\hat{P}_i^t" + }, + { + "bbox": [ + 55, + 415, + 296, + 602 + ], + "type": "text", + "content": " as it can make more stable predictions. This adjustment helps prevent the model from overfitting to noise and incorrect knowledge in the " + }, + { + "bbox": [ + 55, + 415, + 296, + 602 + ], + "type": "inline_equation", + "content": "\\hat{P}_i^{fs}" + }, + { + "bbox": [ + 55, + 415, + 296, + 602 + ], + "type": "text", + "content": ". We use the obtained score as the mixing weight " + }, + { + "bbox": [ + 55, + 415, + 296, + 602 + ], + "type": "inline_equation", + "content": "W_i^t" + }, + { + "bbox": [ + 55, + 415, + 296, + 602 + ], + "type": "text", + "content": " to combine the fixed-strategy pseudo-label " + }, + { + "bbox": [ + 55, + 415, + 296, + 602 + ], + "type": "inline_equation", + "content": "\\hat{P}_i^{fs}" + }, + { + "bbox": [ + 55, + 415, + 296, + 602 + ], + "type": "text", + "content": " and teacher's predictions " + }, + { + "bbox": [ + 55, + 415, + 296, + 602 + ], + "type": "inline_equation", + "content": "\\hat{P}_i^t" + }, + { + "bbox": [ + 55, + 415, + 296, + 602 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 609, + 295, + 631 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 609, + 295, + 631 + ], + "spans": [ + { + "bbox": [ + 113, + 609, + 295, + 631 + ], + "type": "interline_equation", + "content": "\\mathrm {W} _ {i} ^ {t} = \\mathcal {S} \\left(\\hat {\\mathrm {y}} _ {i} ^ {p 1}, \\hat {\\mathrm {y}} _ {i} ^ {p 2}\\right), \\tag {3}", + "image_path": "cc066310db5ebed9b75d0a2502367b77d31d730064df39d7b2b6035f37ca6ab3.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 628, + 237, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 628, + 237, + 643 + ], + "spans": [ + { + "bbox": [ + 115, + 628, + 237, + 643 + ], + "type": "interline_equation", + "content": "P _ {i} = \\mathrm {W} _ {i} ^ {t} \\hat {P} _ {i} ^ {t} + (1 - \\mathrm {W} _ {i} ^ {t}) \\hat {P} _ {i} ^ {f s},", + "image_path": "25f588b6e17214b3a3ab5b0d5336c5c2f1c2b5f271f32d283872d9f2bf913643.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "text", + "content": " denotes the mixed dynamic pseudo-label. For the training of the discriminator, we set a binary classification task by using " + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{P}_i^{fs}" + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{Y}_i^{FG}" + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "text", + "content": " as the input and assign the discriminator label " + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{y}_i^{p1} = 1" + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "text", + "content": " for the fixed-strategy pseudo-label " + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{P}_i^{fs}" + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{y}_i^{p2} = 0" + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "text", + "content": " for the student model branch " + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{Y}_i^{FG}" + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "text", + "content": ". Then," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 415, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 437 + ], + "type": "text", + "content": "we employ a cross-entropy supervision loss " + }, + { + "bbox": [ + 313, + 415, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BCE}}" + }, + { + "bbox": [ + 313, + 415, + 555, + 437 + ], + "type": "text", + "content": " to train the discriminator:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 361, + 443, + 555, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 443, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 361, + 443, + 555, + 458 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d i s} = \\mathcal {L} _ {\\mathrm {B C E}} \\left(\\hat {\\mathrm {y}} _ {i} ^ {p j}, \\mathrm {y} _ {i} ^ {p j}\\right), j \\in \\{1, 2 \\}. \\tag {4}", + "image_path": "35cfccbeb8d3e92f72c28daf56ad2b5d7ba5f6a7a1ffc19e08b840fc170033f2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 462, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 498 + ], + "type": "text", + "content": "During the training, we treat the student model as a kind of generator and alternately train the decoder and the discriminator in a GAN-like manner." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 504, + 499, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 499, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 499, + 516 + ], + "type": "text", + "content": "3.3. Dual-Branch Adversarial Decoder" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 521, + 555, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 618 + ], + "type": "text", + "content": "In previous methods [44, 63], predictions are typically generated by a simple " + }, + { + "bbox": [ + 313, + 521, + 555, + 618 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 313, + 521, + 555, + 618 + ], + "type": "text", + "content": " convolution, which limits the ability to capture camouflage features thoroughly. We propose a Dual-Branch Adversarial decoder to enhance accuracy and robustness in this task. The DBA decoder consists of two parallel branches: one is responsible for segmenting the foreground predicted mask " + }, + { + "bbox": [ + 313, + 521, + 555, + 618 + ], + "type": "inline_equation", + "content": "\\hat{Y}_i^{FG}" + }, + { + "bbox": [ + 313, + 521, + 555, + 618 + ], + "type": "text", + "content": ", while the other is for segmenting the background predicted mask " + }, + { + "bbox": [ + 313, + 521, + 555, + 618 + ], + "type": "inline_equation", + "content": "\\hat{Y}_i^{BG}" + }, + { + "bbox": [ + 313, + 521, + 555, + 618 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 618, + 555, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 555, + 667 + ], + "type": "text", + "content": "First, for " + }, + { + "bbox": [ + 313, + 618, + 555, + 667 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 618, + 555, + 667 + ], + "type": "text", + "content": "-th image feature " + }, + { + "bbox": [ + 313, + 618, + 555, + 667 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 313, + 618, + 555, + 667 + ], + "type": "text", + "content": " extracted from the backbone model, a convolutional layer is used to double its channel, then we decouple it into foreground and background features " + }, + { + "bbox": [ + 313, + 618, + 555, + 667 + ], + "type": "inline_equation", + "content": "F_{i}^{FG}, F_{i}^{BG}" + }, + { + "bbox": [ + 313, + 618, + 555, + 667 + ], + "type": "text", + "content": " with the same shape of " + }, + { + "bbox": [ + 313, + 618, + 555, + 667 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 313, + 618, + 555, + 667 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 367, + 670, + 553, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 670, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 367, + 670, + 553, + 685 + ], + "type": "interline_equation", + "content": "F _ {i} ^ {F G}, F _ {i} ^ {B G} = \\operatorname {S p l i t} \\left(\\operatorname {C o n v} \\left(F _ {i}\\right)\\right), \\tag {5}", + "image_path": "3393f7e98b3cdc4276ea7404dc0b2894695f3ba5601f51b95d4ded7044c4137b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\text{Split}(\\cdot)" + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": " denotes splitting a tensor along the channel dimension equally. We then employ two learnable em" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30368" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": "beddings " + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{FG},\\mathrm{E}_{BG}" + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": " to store foreground and backgroundrelated knowledge. These embeddings are used to calculate foreground and background attention queries " + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "inline_equation", + "content": "Q_{i}^{FG^{\\prime}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "inline_equation", + "content": "Q_{i}^{BG^{\\prime}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 79, + 118, + 295, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 118, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 79, + 118, + 295, + 133 + ], + "type": "interline_equation", + "content": "Q _ {i} ^ {F G ^ {\\prime}} = F _ {i} ^ {F G} \\times \\mathrm {E} _ {F G}, Q _ {i} ^ {B G ^ {\\prime}} = F _ {i} ^ {B G} \\times \\mathrm {E} _ {B G}. \\tag {6}", + "image_path": "a510f8864b7ca4fe7ff7a160f8ed8269c2bceae57c2599af8623700e2c22cb09.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 137, + 296, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 137, + 296, + 200 + ], + "spans": [ + { + "bbox": [ + 55, + 137, + 296, + 200 + ], + "type": "text", + "content": "Finally, we apply multiplicative attention with a residual connection to weigh the features, enhancing important details. The resulting features are then passed through a " + }, + { + "bbox": [ + 55, + 137, + 296, + 200 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 55, + 137, + 296, + 200 + ], + "type": "text", + "content": " convolution to output the foreground predicted mask " + }, + { + "bbox": [ + 55, + 137, + 296, + 200 + ], + "type": "inline_equation", + "content": "\\hat{Y}_i^{FG}" + }, + { + "bbox": [ + 55, + 137, + 296, + 200 + ], + "type": "text", + "content": " and background mask " + }, + { + "bbox": [ + 55, + 137, + 296, + 200 + ], + "type": "inline_equation", + "content": "\\hat{Y}_i^{BG}" + }, + { + "bbox": [ + 55, + 137, + 296, + 200 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 206, + 295, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 206, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 67, + 206, + 295, + 228 + ], + "type": "interline_equation", + "content": "\\hat {Y} _ {i} ^ {F G} = \\operatorname {C o n v} ^ {\\prime} \\left(\\operatorname {S i g m o i d} \\left(Q _ {i} ^ {F G ^ {\\prime}} \\times F _ {i} ^ {F G}\\right) + F _ {i} ^ {F G}\\right), \\tag {7}", + "image_path": "69b5e7993ec651d04c3cf91dee71b5c6abec2b5a47f538585bee573f3527d1d8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 224, + 272, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 224, + 272, + 240 + ], + "spans": [ + { + "bbox": [ + 67, + 224, + 272, + 240 + ], + "type": "interline_equation", + "content": "\\hat {Y} _ {i} ^ {B G} = \\operatorname {C o n v} ^ {\\prime} (\\operatorname {S i g m o i d} \\left(Q _ {i} ^ {B G ^ {\\prime}} \\times F _ {i} ^ {B G}\\right) + F _ {i} ^ {B G}).", + "image_path": "2bffccb8f8db1660ff52d8285ceed5ee023ac55e58a1c4c5a21219da6708b9c2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 249, + 296, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 249, + 296, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 296, + 309 + ], + "type": "text", + "content": "For the student model, we use " + }, + { + "bbox": [ + 55, + 249, + 296, + 309 + ], + "type": "inline_equation", + "content": "Y_{i}^{FG}, \\hat{Y}_{i}^{BG}" + }, + { + "bbox": [ + 55, + 249, + 296, + 309 + ], + "type": "text", + "content": " as the student's foreground and background prediction masks. For the teacher model, we further generate the teacher's pseudolabel " + }, + { + "bbox": [ + 55, + 249, + 296, + 309 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 55, + 249, + 296, + 309 + ], + "type": "text", + "content": " using the combination of the foreground and background prediction masks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 310, + 296, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 296, + 346 + ], + "type": "text", + "content": "Inspired by [55], we apply an orthogonal loss " + }, + { + "bbox": [ + 55, + 310, + 296, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\perp}" + }, + { + "bbox": [ + 55, + 310, + 296, + 346 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 310, + 296, + 346 + ], + "type": "inline_equation", + "content": "Q_{i}^{FG^{\\prime}}\\in \\mathbb{R}^{(n\\times m)\\times c},Q_{i}^{BG^{\\prime}}\\in \\mathbb{R}^{(n\\times m)\\times c}" + }, + { + "bbox": [ + 55, + 310, + 296, + 346 + ], + "type": "text", + "content": " as a regularisation constraint to encourage them to focus on distinct features:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 354, + 221, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 354, + 221, + 369 + ], + "spans": [ + { + "bbox": [ + 132, + 354, + 221, + 369 + ], + "type": "interline_equation", + "content": "\\mathbb {M} = Q _ {i} ^ {F G ^ {\\prime}} \\left(Q _ {i} ^ {B G ^ {\\prime}}\\right) ^ {\\top},", + "image_path": "241edbaeb23fdce3355a8b5afe8e4cad26a6249fb8fddca039880e638b8aa6f5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 73, + 373, + 295, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 373, + 295, + 405 + ], + "spans": [ + { + "bbox": [ + 73, + 373, + 295, + 405 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\perp} = \\frac {1}{(n \\times m) ^ {2}} \\sum_ {j = 1} ^ {n \\times m} \\sum_ {k = 1} ^ {n \\times m} (1 - \\delta_ {j k}) (\\mathbb {M} [ j, k ]) ^ {2}, \\tag {8}", + "image_path": "7ab136ab8de706ffa57c7f003cac75a188a4d3588f35cc31589594a08e6525c2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 413, + 296, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 413, + 296, + 475 + ], + "type": "inline_equation", + "content": "n, m, c" + }, + { + "bbox": [ + 55, + 413, + 296, + 475 + ], + "type": "text", + "content": " indicate the height, width, and number of channels. When " + }, + { + "bbox": [ + 55, + 413, + 296, + 475 + ], + "type": "inline_equation", + "content": "j = k" + }, + { + "bbox": [ + 55, + 413, + 296, + 475 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 413, + 296, + 475 + ], + "type": "inline_equation", + "content": "\\delta_{jk} = 1" + }, + { + "bbox": [ + 55, + 413, + 296, + 475 + ], + "type": "text", + "content": "; otherwise " + }, + { + "bbox": [ + 55, + 413, + 296, + 475 + ], + "type": "inline_equation", + "content": "\\delta_{jk} = 0" + }, + { + "bbox": [ + 55, + 413, + 296, + 475 + ], + "type": "text", + "content": ". For the predictions " + }, + { + "bbox": [ + 55, + 413, + 296, + 475 + ], + "type": "inline_equation", + "content": "\\hat{Y}_i^{FG}, \\hat{Y}_i^{BG}" + }, + { + "bbox": [ + 55, + 413, + 296, + 475 + ], + "type": "text", + "content": ", we use a binary cross-entropy loss " + }, + { + "bbox": [ + 55, + 413, + 296, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BCE}}" + }, + { + "bbox": [ + 55, + 413, + 296, + 475 + ], + "type": "text", + "content": " for supervision, where the background prediction is inverted before applying the supervision:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 482, + 296, + 503 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 482, + 296, + 503 + ], + "spans": [ + { + "bbox": [ + 61, + 482, + 296, + 503 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s e g} = \\mathcal {L} _ {\\mathrm {B C E}} (\\hat {Y} _ {i} ^ {F G}, P _ {i}) + \\mathcal {L} _ {\\mathrm {B C E}} \\left((1 - \\hat {Y} _ {i} ^ {B G}), P _ {i}\\right). (9)", + "image_path": "f2f35ece9a5a08f59c1bcfb64049e46a1ceb9b7c8ba1d93acd29b8b252771504.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 510, + 190, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 190, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 190, + 521 + ], + "type": "text", + "content": "3.4. Look-Twice Refinement" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 526, + 296, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 296, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 296, + 659 + ], + "type": "text", + "content": "The segmentation quality of COD is often influenced by object size. Due to reduced spatial resolution, the model struggles to produce fine segmentation for small-sized camouflaged objects. Inspired by previous works [37, 38], we attempt to mimic human behavior when observing small-sized objects: first coarsely locating the camouflaged object in the image, then zooming in to carefully retrieve its details. To this end, we designed the Look-Twice mechanism. First, the input image is processed by the model to obtain a coarse segmentation result " + }, + { + "bbox": [ + 55, + 526, + 296, + 659 + ], + "type": "inline_equation", + "content": "\\hat{Y}_i^{FG}" + }, + { + "bbox": [ + 55, + 526, + 296, + 659 + ], + "type": "text", + "content": ". We define the set of all foreground objects as " + }, + { + "bbox": [ + 55, + 526, + 296, + 659 + ], + "type": "inline_equation", + "content": "\\mathrm{C}^{FG}" + }, + { + "bbox": [ + 55, + 526, + 296, + 659 + ], + "type": "text", + "content": ", which can be calculated as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 92, + 666, + 295, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 666, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 92, + 666, + 295, + 681 + ], + "type": "interline_equation", + "content": "\\mathrm {C} ^ {\\mathrm {F G}} = \\text {C o n n e c t C o m p o n e n t} \\left(\\hat {Y} _ {i} ^ {F G}\\right), \\tag {10}", + "image_path": "3372360487c859f4cc2bd2b3003ef269c6c586152f1abea53c0f22c564da1be3.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where ConnectComponent(\\cdot) denotes the connected components labelling algorithm. Next, for each connect" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 71, + 554, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 71, + 554, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 554, + 97 + ], + "type": "text", + "content": "component " + }, + { + "bbox": [ + 313, + 71, + 554, + 97 + ], + "type": "inline_equation", + "content": "c_k^{FG} \\in \\mathbb{C}^{FG}" + }, + { + "bbox": [ + 313, + 71, + 554, + 97 + ], + "type": "text", + "content": ", we calculate its foreground area ratio " + }, + { + "bbox": [ + 313, + 71, + 554, + 97 + ], + "type": "inline_equation", + "content": "r_k^{FG}" + }, + { + "bbox": [ + 313, + 71, + 554, + 97 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 367, + 102, + 553, + 130 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 102, + 553, + 130 + ], + "spans": [ + { + "bbox": [ + 367, + 102, + 553, + 130 + ], + "type": "interline_equation", + "content": "r _ {k} ^ {F G} = \\frac {\\sum_ {(x , y) \\in \\mathrm {c} _ {k} ^ {F G}} \\hat {Y} _ {i} ^ {F G} (x , y)}{H \\times W}, \\tag {11}", + "image_path": "7ea64e96d0a3dfd2a3202e6cf4dadc35c71570bc2388d424726dfe7e0cc46c1e.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 133, + 554, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 554, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 554, + 204 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 133, + 554, + 204 + ], + "type": "inline_equation", + "content": "H, W" + }, + { + "bbox": [ + 313, + 133, + 554, + 204 + ], + "type": "text", + "content": " denote the height and width of the foreground area of the output mask " + }, + { + "bbox": [ + 313, + 133, + 554, + 204 + ], + "type": "inline_equation", + "content": "\\tilde{Y}_i^{FG}" + }, + { + "bbox": [ + 313, + 133, + 554, + 204 + ], + "type": "text", + "content": ". We consider all regions where " + }, + { + "bbox": [ + 313, + 133, + 554, + 204 + ], + "type": "inline_equation", + "content": "r_i^{FG} < \\tau" + }, + { + "bbox": [ + 313, + 133, + 554, + 204 + ], + "type": "text", + "content": " as small-sized object areas. We set " + }, + { + "bbox": [ + 313, + 133, + 554, + 204 + ], + "type": "inline_equation", + "content": "\\tau = 0.15" + }, + { + "bbox": [ + 313, + 133, + 554, + 204 + ], + "type": "text", + "content": " when using DINOv2 as the backbone. The hyperparameter ablation study on " + }, + { + "bbox": [ + 313, + 133, + 554, + 204 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 133, + 554, + 204 + ], + "type": "text", + "content": " will be provided in the experiments section." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 205, + 554, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 554, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 554, + 277 + ], + "type": "text", + "content": "For each cropped small-sized object region, we aim to retain a certain proportion of the background to provide the model with sufficient context to segment the foreground from the background. Therefore, for each connected component region " + }, + { + "bbox": [ + 313, + 205, + 554, + 277 + ], + "type": "inline_equation", + "content": "c_k^{FG}" + }, + { + "bbox": [ + 313, + 205, + 554, + 277 + ], + "type": "text", + "content": ", we calculate the local foreground area ratio " + }, + { + "bbox": [ + 313, + 205, + 554, + 277 + ], + "type": "inline_equation", + "content": "s_k^{FG}" + }, + { + "bbox": [ + 313, + 205, + 554, + 277 + ], + "type": "text", + "content": " and the global area ratio " + }, + { + "bbox": [ + 313, + 205, + 554, + 277 + ], + "type": "inline_equation", + "content": "s_k^{BG}" + }, + { + "bbox": [ + 313, + 205, + 554, + 277 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 367, + 281, + 553, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 281, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 367, + 281, + 553, + 316 + ], + "type": "interline_equation", + "content": "s _ {k} ^ {F G} = \\frac {\\sum_ {(x , y) \\in c _ {k} ^ {F G}} \\hat {Y} _ {i} ^ {F G} (x , y)}{H _ {k} ^ {F G} \\times W _ {k} ^ {F G}}, \\tag {12}", + "image_path": "8846f1ce583462d1bd0244fc14f9ffa3af5e75a8b564ed8188a8293c975947af.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 369, + 314, + 465, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 314, + 465, + 340 + ], + "spans": [ + { + "bbox": [ + 369, + 314, + 465, + 340 + ], + "type": "interline_equation", + "content": "s _ {k} ^ {B G} = \\frac {H _ {k} ^ {F G} \\times W _ {k} ^ {F G}}{H \\times W},", + "image_path": "6562e0da7a38fb493e73fca1b9af3eefc42674bf2c96be97af6e3ec7ea3f4992.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 343, + 554, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 554, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 554, + 369 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 343, + 554, + 369 + ], + "type": "inline_equation", + "content": "H_{k}^{FG},W_{k}^{FG}" + }, + { + "bbox": [ + 313, + 343, + 554, + 369 + ], + "type": "text", + "content": " denote the height and width of the bounding box of the " + }, + { + "bbox": [ + 313, + 343, + 554, + 369 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 343, + 554, + 369 + ], + "type": "text", + "content": " -th connected component " + }, + { + "bbox": [ + 313, + 343, + 554, + 369 + ], + "type": "inline_equation", + "content": "\\mathrm{c}_k^{FG}" + }, + { + "bbox": [ + 313, + 343, + 554, + 369 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 375, + 373, + 553, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 373, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 375, + 373, + 553, + 388 + ], + "type": "interline_equation", + "content": "\\operatorname {s c a l e} _ {k} ^ {F G} = 1 - \\left(s _ {k} ^ {F G} / s _ {k} ^ {B G}\\right). \\tag {13}", + "image_path": "987acb93159230ccba99ce70eb2124822555af1284e6dc7fd868f1365f73e28b.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 392, + 554, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 554, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 554, + 559 + ], + "type": "text", + "content": "Finally, we calculate the expansion ratio using Eq. (13) and enlarge the bounding box of the connected component region by this ratio. The expanded bounding box is then used to crop the related area of the input image, and the cropped patches are resized to the original input size. During the training process, these patches are treated as augmented data. At test time, we re-infer these patches, scale the resulting foreground predicted mask back to the original size, and paste it into the coarse mask as the refined result. However, we found that this mechanism still has some weakness: for cases where objects are occluded and the annotations are fragmented, this method may mistakenly amplify a part of the object as a small object. This can serve as a point of our future work." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 565, + 385, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 565, + 385, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 565, + 385, + 577 + ], + "type": "text", + "content": "3.5. Total Loss" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 582, + 526, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 526, + 595 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 526, + 595 + ], + "type": "text", + "content": "Finally, the total loss " + }, + { + "bbox": [ + 313, + 582, + 526, + 595 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{tot}" + }, + { + "bbox": [ + 313, + 582, + 526, + 595 + ], + "type": "text", + "content": " can be defined as follows:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 378, + 600, + 553, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 600, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 378, + 600, + 553, + 613 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t o t} = \\mathcal {L} _ {s e g} + \\mathcal {L} _ {\\perp} + \\mathcal {L} _ {d i s}, \\tag {14}", + "image_path": "9c2af2f6f033244598e35e5d76ab0cac8c2bf2186989486c380ac7445a114a90.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 617, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 617, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{seg}" + }, + { + "bbox": [ + 313, + 617, + 554, + 714 + ], + "type": "text", + "content": " is the segmentation loss defined by Eq. (9), " + }, + { + "bbox": [ + 313, + 617, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\perp}" + }, + { + "bbox": [ + 313, + 617, + 554, + 714 + ], + "type": "text", + "content": " is the orthogonal loss defined by Eq. (8), and " + }, + { + "bbox": [ + 313, + 617, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dis}" + }, + { + "bbox": [ + 313, + 617, + 554, + 714 + ], + "type": "text", + "content": " is the discriminator loss defined by Eq. (4). When training on the teacher-student framework, we freeze the parameters of the discriminator and use " + }, + { + "bbox": [ + 313, + 617, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{seg}" + }, + { + "bbox": [ + 313, + 617, + 554, + 714 + ], + "type": "text", + "content": " to supervise the model training. Conversely, when training the discriminator, we freeze the parameters of the other parts of the model and train the discriminator only using " + }, + { + "bbox": [ + 313, + 617, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dis}" + }, + { + "bbox": [ + 313, + 617, + 554, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30369" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 555, + 370 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 370 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 370 + ], + "type": "table", + "html": "
MethodsCHAMELEON (87)CAMO-Test (250)COD10K-Test (2,026)NC4K (4,121)
Sm↑FωβFmβEmφM↓Sm↑FωβFmβEmφM↓Sm↑FωβFmβEmφM↓Sm↑FωβFmβEmφM↓
Fully-Supervised Methods
SINet20[10].872.806.827.946.034.751.606.675.771.100.771.551.634.806.051.808.723.769.871.058
C2FNet21[47].888.828.844.946.032.796.719.762.864.080.813.686.723.900.036.838.762.794.904.049
MGL-R21[61].893.812.834.941.030.775.673.726.842.088.814.666.710.890.035.833.739.782.893.053
UGTR21[58].887.794.819.940.031.784.684.735.851.086.817.666.711.890.036.839.746.787.899.052
BGNet22[48].901.851.860.954.027.812.749.789.870.073.831.722.753.901.033.851.788.820.907.044
ZoomNet22[37].902.845.864.958.023.820.752.794.878.066.838.729.766.888.029.853.784.818.896.043
SINetv22[12].888.816.835.961.030.820.743.782.882.070.815.680.718.887.037.847.770.805.903.048
HitNet23[22].921.897.900.972.019.849.809.831.906.055.871.806.823.935.023.875.834.853.926.037
FSPNet23[23].908.851.867.965.023.856.799.830.899.050.851.735.769.895.026.879.816.843.915.035
BiRefNet24[64].929.911.922.968.016.932.914.922.974.015.913.874.888.960.014.914.894.909.953.023
Semi-Supervised Methods
CamoTeacher24(1%)[27].652.472.558.714.093.621.456.545.669.136.699.517.788.797.062.718.599.779.814.090
CamoTeacher24(5%)[27].729.587.656.785.070.669.523.601.711.122.745.583.827.840.050.777.677.834.859.071
CamoTeacher24(10%)[27].756.617.684.813.065.701.560.742.795.112.759.594.836.854.049.791.687.842.868.068
SCOD-ND24(10%)[13].850.773-.928.036.789.732-.859.077.819.725-.891.033.838.787-.903.046
Unsupervised Methods
BigGW21[51].547.244.294.527.257.565.299.349.528.282.528.185.246.497.261.608.319.391.565.246
TokenCut22[53].654.496.536.740.132.633.498.543.706.163.658.469.502.735.103.725.615.649.802.101
TokenCut22 w/B.S.[53].655.351.393.582.169.639.383.434.595.195.666.334.399.609.127.735.478.547.683.133
SpectralSeg22[32].575.410.440.628.220.579.450.481.648.235.575.360.388.595.193.669.535.562.719.159
SelfMask22[43].619.436.481.675.176.617.483.536.698.176.637.431.469.679.131.716.593.634.777.114
SelfMask22 w/U.B.[43].629.447.491.683.169.627.495.547.708.182.645.440.478.687.125.723.601.642.784.110
FOUND22-DINOv1[44].684.542.590.810.095.685.584.633.782.129.670.482.520.751.085.741.637.674.824.084
*FOUND22-DINOv2[44].829.757.781.911.040.770.704.740.849.090.767.641.668.847.045.816.756.783.893.052
UCOS-DA23-DINOv1[63].715.591.629.802.095.701.606.646.784.127.689.513.546.740.086.755.656.689.819.085
*UCOS-DA23-DINOv2[63].750.639.666.808.091.702.604.633.751.148.655.467.495.687.120.731.617.644.785.103
OursDINOv1.734.625.680.854.072.706.621.689.801.108.727.577.627.822.059.761.680.737.851.074
OursDINOv2.864.825.838.931.031.793.747.779.862.077.834.763.779.916.031.850.818.835.923.043
", + "image_path": "d49f7e8cb0c745161f0aa8f4316084b0dd6dda57b92b08cae9cd77a40f5df15a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 375, + 555, + 409 + ], + "lines": [ + { + "bbox": [ + 55, + 375, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 555, + 409 + ], + "type": "text", + "content": "Table 1. Comparison of our methods with recent methods. We compared our proposed methods with competing unsupervised, semi-supervised, and full-supervised methods. Bold indicates the best result in group settings, and underline indicates the second-best result. * denotes the version that reimplemented by us." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 429, + 132, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 429, + 132, + 442 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 132, + 442 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 449, + 182, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 182, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 182, + 464 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 469, + 295, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 295, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 295, + 541 + ], + "type": "text", + "content": "Trainsets. To fairly compare with the existing works, following [10, 30], we used a combination of 1000 images from the CAMO-Train and 3040 images from COD10K-Train as the training dataset of our experiments. Following the unsupervised learning settings, we do not use any ground-truth label during the training process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 543, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 295, + 602 + ], + "type": "text", + "content": "Testing Sets. We evaluate the performance of our methods on four mainstream COD benchmark testing sets: CHAMELEON with 76 test images, CAMO with 250 test images, COD10K with 2026 test images, and NC4K with 4121 test images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 604, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 295, + 652 + ], + "type": "text", + "content": "Evaluation Protocol. For a fair and comprehensive evaluation, we employ the S-measure " + }, + { + "bbox": [ + 55, + 604, + 295, + 652 + ], + "type": "inline_equation", + "content": "(S_m)" + }, + { + "bbox": [ + 55, + 604, + 295, + 652 + ], + "type": "text", + "content": " [8], mean and weighted F-measure " + }, + { + "bbox": [ + 55, + 604, + 295, + 652 + ], + "type": "inline_equation", + "content": "(\\mathcal{F}_\\beta^m,\\mathcal{F}_\\beta^\\omega)" + }, + { + "bbox": [ + 55, + 604, + 295, + 652 + ], + "type": "text", + "content": " [31], mean E-measure " + }, + { + "bbox": [ + 55, + 604, + 295, + 652 + ], + "type": "inline_equation", + "content": "(\\mathcal{E}_{\\epsilon}^{m})" + }, + { + "bbox": [ + 55, + 604, + 295, + 652 + ], + "type": "text", + "content": " [9], mean absolute error " + }, + { + "bbox": [ + 55, + 604, + 295, + 652 + ], + "type": "inline_equation", + "content": "(\\mathcal{M})" + }, + { + "bbox": [ + 55, + 604, + 295, + 652 + ], + "type": "text", + "content": " [39] as the main metrics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "Implementation Details. Following [44], we employ the simple background seed methods as our fixed strategy branch to generate the coarse pseudo-label. And following [44, 63], we use the strong self-supervised pretrained backbone DINO [5, 36] as our encoder. The teacher model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 430, + 555, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 555, + 526 + ], + "type": "text", + "content": "is updated via EMA with a momentum " + }, + { + "bbox": [ + 313, + 430, + 555, + 526 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 430, + 555, + 526 + ], + "type": "text", + "content": " of 0.99. We train the model for 25 epochs. The batch size is set to 32 for each GPU during training. All experiments are implemented with PyTorch 2.1 and run on a machine with Intel(R) Xeon(R) Silver 4214R CPU @ 2.40GHz, 512GiB RAM, and 1 NVIDIA Titan A100-40G GPUs. All experiments use the same random seed. More implementation details will be provided in the supplementary materials." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 537, + 400, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 400, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 400, + 550 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 556, + 555, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 555, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 555, + 663 + ], + "type": "text", + "content": "Qualitative Analysis. We show visualizations of a series of camouflaged object segmentation masks predicted by our method and related methods in some challenging scenarios. As shown in Fig. 4, we noticed that our method achieves higher segmentation accuracy than existing approaches, with clearer edges and a more detailed representation of the camouflaged object. Note that DINOv2 is employed for all methods that require a backbone to extract features." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "Quantitative Analysis. In Tab. 1, we compared our proposed method's performance with competing USS and UCOD models on four COD test datasets. To further highlight our model's competitiveness, we also compared ex" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "30370" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 74, + 549, + 360 + ], + "blocks": [ + { + "bbox": [ + 59, + 74, + 549, + 360 + ], + "lines": [ + { + "bbox": [ + 59, + 74, + 549, + 360 + ], + "spans": [ + { + "bbox": [ + 59, + 74, + 549, + 360 + ], + "type": "image", + "image_path": "f46a89772bbc1654c1a1fab1dd8c61f22cc6d71261dde5982a42629c3e683b42.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 369, + 555, + 392 + ], + "lines": [ + { + "bbox": [ + 55, + 369, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 555, + 392 + ], + "type": "text", + "content": "Figure 4. Visual comparison of our method with other existing methods in challenging scenarios. Our method has clearer and more precise segmentation boundaries and correctly recognizes depth-artifacted objects." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 411, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 296, + 509 + ], + "type": "text", + "content": "isting fully-supervised and semi-supervised methods. The results show that our model outperformed all existing USS and UCOD methods across all metrics and datasets, thus achieving State-Of-The-Art (SOTA) performance. Additionally, based on DINOv2 backbone, our model has surpassed several semi-supervised and fully-supervised methods across all four datasets, demonstrating its superior performance, effectiveness, and robustness." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 516, + 149, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 149, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 149, + 529 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 533, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 569 + ], + "type": "text", + "content": "To verify the effectiveness of our methods, we employ " + }, + { + "bbox": [ + 55, + 533, + 295, + 569 + ], + "type": "inline_equation", + "content": "DI-NOv2" + }, + { + "bbox": [ + 55, + 533, + 295, + 569 + ], + "type": "text", + "content": " as the backbone and conduct comprehensive ablation studies on the COD10K-Test dataset with 2,026 images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "Module Ablation. We first conduct ablation experiments on the teacher-student framework (Tea-Stu) and our proposed methods: Adaptive Pseudo-label Mixing (APM) module, Dual-Branch Adversarial (DBA) decoder, and Look-Twice strategy. To thoroughly demonstrate their effectiveness, we replaced the decoder with a single " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " convolutional layer following [44, 63]. For all ablations on the APM module, we used the " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "1:1" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " constant mixing strategy to replace it. The ablation results are shown in Tab. 2. As a result, we find that when only the teacher-student framework is used, the simple " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " convolution fails to learn camouflage-related knowledge adequately. When training" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 409, + 556, + 531 + ], + "blocks": [ + { + "bbox": [ + 317, + 409, + 556, + 531 + ], + "lines": [ + { + "bbox": [ + 317, + 409, + 556, + 531 + ], + "spans": [ + { + "bbox": [ + 317, + 409, + 556, + 531 + ], + "type": "table", + "html": "
SettingsCOD10K (2026)
Tea-StuAPMDBALook-TwiceSm↑Fωβ↑Fmβ↑Emφ↑M↓
✓.626.448.464.658.112
✓✓.633.452.472.668.129
✓.607.409.435.631.129
✓✓.603.404.431.625.133
✓✓.724.582.604.800.054
✓✓✓.788.679.700.879.043
✓✓✓.784.675.692.869.036
✓✓✓✓.834.763.779.916.031
", + "image_path": "94999939b255f78e815bb0a5b4e292037407b3b1df3fb382a67c50aca4468b8e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 538, + 555, + 583 + ], + "lines": [ + { + "bbox": [ + 313, + 538, + 555, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 555, + 583 + ], + "type": "text", + "content": "Table 2. Ablation study to evaluate the proposed modules. We retrained our model with different settings on the same learning rate and epochs, then benchmarked on COD10K-testset to validate the effectiveness." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 605, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 700 + ], + "type": "text", + "content": "with a DBA decoder only, the model lacks the guidance of the teacher-student framework, thus easily learning the incorrect knowledge from the low-quality pseudo-labels constructed by fixed-strategy. The proposed Look-Twice mechanism is effective only when the network can produce sufficiently high-quality predictions. If the network has poor localization and segmentation capabilities, Look-Twice will degrade performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "type": "text", + "content": "Mixing Strategy Ablation. We continue conducting ab-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30371" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 295, + 135 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 295, + 135 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 295, + 135 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 295, + 135 + ], + "type": "table", + "html": "
Mixing StrategiesCOD10K (2026)
Sm↑Fωβ↑Fmβ↑Eφ↑M↓
Proportional(1:1) Mixing.707.555.572.763.082
Linear Decay Mixing.809.716.723.888.038
APM.834.763.779.916.031
", + "image_path": "73d05e0996f2ad857bd64cd64c8f30c38a5cf0a425a0bf0f9638b0787ec19781.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 192, + 287, + 301 + ], + "blocks": [ + { + "bbox": [ + 61, + 192, + 287, + 301 + ], + "lines": [ + { + "bbox": [ + 61, + 192, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 61, + 192, + 287, + 301 + ], + "type": "image", + "image_path": "4f3057181878e30e094f1cc195b8e5535dee4ae504bb963561870aa8c1dbdb6c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 315, + 295, + 349 + ], + "lines": [ + { + "bbox": [ + 55, + 315, + 295, + 349 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 295, + 349 + ], + "type": "text", + "content": "Figure 5. Hyper-parameter ablation of small-sized object ratio " + }, + { + "bbox": [ + 55, + 315, + 295, + 349 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 315, + 295, + 349 + ], + "type": "text", + "content": ". The model performs optimally and shows stable results when the value of " + }, + { + "bbox": [ + 55, + 315, + 295, + 349 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 315, + 295, + 349 + ], + "type": "text", + "content": " is set around 0.15." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 360, + 295, + 423 + ], + "blocks": [ + { + "bbox": [ + 55, + 142, + 295, + 175 + ], + "lines": [ + { + "bbox": [ + 55, + 142, + 295, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 142, + 295, + 175 + ], + "type": "text", + "content": "Table 3. Ablation to the pseudo-label mixing strategy. We retrained our model with different pseudo-label mixing strategies and benchmarked on the COD10K-Test dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 360, + 295, + 423 + ], + "lines": [ + { + "bbox": [ + 58, + 360, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 58, + 360, + 295, + 423 + ], + "type": "table", + "html": "
Fixed Generation StrategiesCOD10K (2026)
Sm↑Fωβ↑Fmβ↑Emφ↑M↓
Null(Pure Black/White).459.002.001.250.091
Random Perlin Noise.667.489.543.788.072
Background Seed.834.763.779.916.031
", + "image_path": "9c12f184efb53acdc1c26ebd60956ece2c2cfa7e4f1f1ba11d1df4e04dad3d6b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 430, + 295, + 464 + ], + "lines": [ + { + "bbox": [ + 55, + 430, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 295, + 464 + ], + "type": "text", + "content": "Table 4. Fixed Pseudo-label Generation Strategy Ablation. We retrained our model with a replaced fixed strategy and benchmarked it on the COD10K-Test dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 486, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 545 + ], + "type": "text", + "content": "lation studies on the mixing strategies. We compared the proportional(1:1), linear decay, and APM mixing strategies. As shown in Tab. 3, the APM module provides a method for fusing pseudo-labels that is more aligned with the model's training process and achieving optimal performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "type": "text", + "content": "Hyper-parameter Ablation. We conducted ablation study on the threshold " + }, + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "type": "text", + "content": " for the small-sized object ratio mentioned in Sec. 3.4. As shown in Fig. 5, when " + }, + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\tau \\approx 0.15" + }, + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "type": "text", + "content": ", the model achieves optimal and stable performance. Therefore, we set " + }, + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "type": "text", + "content": " to 0.15 in the final model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": "Study on Foreground-sizes. We divided the test set by the proportion of test set prospects at " + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": " intervals, then benchmarked the performance between our method and some previous SOTA methods[44, 63]. As shown in Fig. 6, our method has significant advantages when dealing with smaller-sized camouflaged targets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Fixed Pseudo-label Generation Strategy Ablation. We further investigated the effect of the pseudo-labels constructed by the fixed-strategy. For each input image, we" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 436, + 149 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 436, + 149 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 436, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 436, + 149 + ], + "type": "image", + "image_path": "9283571ec3bb47002208c1ee0b09a24fd8100561a05124ff3767814af27a63e6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 315, + 150, + 436, + 230 + ], + "blocks": [ + { + "bbox": [ + 315, + 150, + 436, + 230 + ], + "lines": [ + { + "bbox": [ + 315, + 150, + 436, + 230 + ], + "spans": [ + { + "bbox": [ + 315, + 150, + 436, + 230 + ], + "type": "image", + "image_path": "e6f62818efa23ab878b50496c92983054a0701f4478add2bbf493b9980bc2bd1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 239, + 554, + 261 + ], + "lines": [ + { + "bbox": [ + 313, + 239, + 554, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 554, + 261 + ], + "type": "text", + "content": "Figure 6. Performance comparison for different foreground sizes on COD10K-Test dataset." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 439, + 70, + 554, + 150 + ], + "blocks": [ + { + "bbox": [ + 439, + 70, + 554, + 150 + ], + "lines": [ + { + "bbox": [ + 439, + 70, + 554, + 150 + ], + "spans": [ + { + "bbox": [ + 439, + 70, + 554, + 150 + ], + "type": "image", + "image_path": "64f6a5107d93963b83679ae70d0618ca8c9412b0cf8707566ca24fb47f217f9c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 439, + 150, + 554, + 230 + ], + "blocks": [ + { + "bbox": [ + 439, + 150, + 554, + 230 + ], + "lines": [ + { + "bbox": [ + 439, + 150, + 554, + 230 + ], + "spans": [ + { + "bbox": [ + 439, + 150, + 554, + 230 + ], + "type": "image", + "image_path": "d8e6e54f6f1a3eb10a5753a404a0f77665ca280763698d09042fd0792333e539.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 284, + 555, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 555, + 403 + ], + "type": "text", + "content": "generated a random shape of Perlin noise on a plain black background. We used this \"fabricated pseudo-label\" as well as the pure black and white masks to replace the fixed-strategy pseudo-label during the training process. The results are shown in Tab. 4. When using random noise, the noise may overlap with the foreground region, allowing the network to explore some knowledge related to the camouflaged foreground. However, when using a solid color mask, the model fails to learn the basic definition of the segmentation task to produce valid predictions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 417, + 391, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 417, + 391, + 430 + ], + "spans": [ + { + "bbox": [ + 314, + 417, + 391, + 430 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "text", + "content": "Considering the existing UCOD methods suffer from main issues: 1). The low-quality pseudo-labels constructed by fixed strategies contain substantial noise, making the model prone to fitting incorrect knowledge. 2). The low-resolution pseudo-labels cause severe confusion between the foreground and background. Simple " + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "text", + "content": " convolution fails to capture and learn the semantic features of camouflaged objects, especially for small-sized samples. To solve these problems, we propose UCOD-DPL in this paper. We introduce a teacher-student framework with an Adaptive Pseudolabel Module, a Dual-Branch Adversarial decoder, and a Look-Twice mechanism. The APM dynamically fuses pseudo-labels from the fixed strategy and teacher model, effectively preventing the model from overfitting the noise and erroneous knowledge. The DBA decoder mines foreground and background knowledge relevant to camouflaged objects while resisting incorrect information from pseudolabels. The Look-Twice refinement mechanism enhances camouflaged object learning and refinement of small objects. We conducted extensive benchmarking on multiple COD test datasets, and the results show that our method outperforms all previous approaches and even reaches the performance level of some fully supervised methods." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30372" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 180 + ], + "type": "text", + "content": "This work was supported by the National Science Fund for Distinguished Young Scholars (No.62025603), the National Natural Science Foundation of China (No. U21B2037, No. U22B2051, No. U23A20383, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No. 2021J06003, No. 2022J06001)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 200, + 115, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 200, + 115, + 213 + ], + "spans": [ + { + "bbox": [ + 56, + 200, + 115, + 213 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 220, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 220, + 296, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 220, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 61, + 220, + 296, + 252 + ], + "type": "text", + "content": "[1] Relja Arandjelović and Andrew Zisserman. Object discovery with a copy-pasting gan. arXiv preprint arXiv:1905.11369, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 255, + 296, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 255, + 296, + 286 + ], + "spans": [ + { + "bbox": [ + 61, + 255, + 296, + 286 + ], + "type": "text", + "content": "[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 289, + 295, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 289, + 295, + 321 + ], + "spans": [ + { + "bbox": [ + 62, + 289, + 295, + 321 + ], + "type": "text", + "content": "[3] Adam Bielski and Paolo Favaro. Emergence of object segmentation in perturbed generative models. Advances in Neural Information Processing Systems, 32, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 323, + 294, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 323, + 294, + 344 + ], + "spans": [ + { + "bbox": [ + 62, + 323, + 294, + 344 + ], + "type": "text", + "content": "[4] TIM Caro. The adaptive significance of coloration in mammals. BioScience, 55(2):125-136, 2005. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 345, + 295, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 345, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 62, + 345, + 295, + 399 + ], + "type": "text", + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In The IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 401, + 295, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 401, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 401, + 295, + 445 + ], + "type": "text", + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 446, + 294, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 446, + 294, + 489 + ], + "spans": [ + { + "bbox": [ + 62, + 446, + 294, + 489 + ], + "type": "text", + "content": "[7] Ming-Ming Cheng, Niloy J Mitra, Xiaolei Huang, Philip HS Torr, and Shi-Min Hu. Global contrast based salient region detection. IEEE transactions on pattern analysis and machine intelligence, 37(3):569-582, 2014. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 491, + 294, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 294, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 294, + 534 + ], + "type": "text", + "content": "[8] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In The IEEE/CVF International Conference on Computer Vision, pages 4548-4557, 2017. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 536, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 294, + 590 + ], + "type": "text", + "content": "[9] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, Ming-Ming Cheng, and Ali Borji. Enhanced-alignment measure for binary foreground map evaluation. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, pages 698-704. AAAI Press, 2018. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "text", + "content": "[10] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 2777-2787, 2020. 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 636, + 295, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 691 + ], + "type": "text", + "content": "[11] Deng-Ping Fan, Ge-Peng Ji, Tao Zhou, Geng Chen, Huazhu Fu, Jianbing Shen, and Ling Shao. Pranet: Parallel reverse attention network for polyp segmentation. In International conference on medical image computing and computer-assisted intervention, pages 263-273. Springer, 2020. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[12] Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. Concealed object detection. IEEE Transactions on" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "Pattern Analysis and Machine Intelligence, 44(10):6024-6042, 2022. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 98, + 553, + 130 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 98, + 553, + 130 + ], + "spans": [ + { + "bbox": [ + 316, + 98, + 553, + 130 + ], + "type": "text", + "content": "[13] Yuanbin Fu, Jie Ying, Houlei Lv, and Xiaojie Guo. Semi-supervised camouflaged object detection from noisy data. In ACM International Conference on Multimedia, 2024. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 133, + 553, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 133, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 316, + 133, + 553, + 175 + ], + "type": "text", + "content": "[14] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations, 2018. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 178, + 553, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 178, + 553, + 233 + ], + "spans": [ + { + "bbox": [ + 317, + 178, + 553, + 233 + ], + "type": "text", + "content": "[15] Spyros Gidaris, Andrei Bursuc, Gilles Puy, Nikos Komodakis, Matthieu Cord, and Patrick Pérez. Obow: Online bag-of-visual-words generation for self-supervised learning. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 6830–6840, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 236, + 553, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 236, + 553, + 278 + ], + "spans": [ + { + "bbox": [ + 317, + 236, + 553, + 278 + ], + "type": "text", + "content": "[16] Stas Goferman, Lihi Zelnik-Manor, and Ayellet Tal. Context-aware saliency detection. IEEE transactions on pattern analysis and machine intelligence, 34(10):1915-1926, 2011. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 281, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 281, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 281, + 553, + 326 + ], + "type": "text", + "content": "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 327, + 553, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 327, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 317, + 327, + 553, + 393 + ], + "type": "text", + "content": "[18] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 396, + 553, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 396, + 553, + 450 + ], + "spans": [ + { + "bbox": [ + 317, + 396, + 553, + 450 + ], + "type": "text", + "content": "[19] Chunming He, Kai Li, Yachao Zhang, Longxiang Tang, Yu-lun Zhang, Zhenhua Guo, and Xiu Li. Camouflaged object detection with feature decomposition and edge reconstruction. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 22046-22055, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 453, + 553, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 453, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 317, + 453, + 553, + 496 + ], + "type": "text", + "content": "[20] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 9729-9738, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 498, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 498, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 317, + 498, + 553, + 552 + ], + "type": "text", + "content": "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 16000-16009, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 555, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 555, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 317, + 555, + 553, + 610 + ], + "type": "text", + "content": "[22] Xiaobin Hu, Shuo Wang, Xuebin Qin, Hang Dai, Wenqi Ren, Donghao Luo, Ying Tai, and Ling Shao. High-resolution iterative feedback network for camouflaged object detection. In Association for the Advancement of Artificial Intelligence, pages 881-889, 2023. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 613, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 667 + ], + "type": "text", + "content": "[23] Zhou Huang, Hang Dai, Tian-Zhu Xiang, Shuo Wang, Huai-Xin Chen, Jie Qin, and Huan Xiong. Feature shrinkage pyramid for camouflaged object detection with transformers. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 5557-5566, 2023. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[24] Laurent Itti, Christof Koch, and Ernst Niebur. A model of saliency-based visual attention for rapid scene analysis. IEEE Transactions on pattern analysis and machine intelligence, 20(11):1254-1259, 1998. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30373" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[25] Ge-Peng Ji, Deng-Ping Fan, Yu-Cheng Chou, Dengxin Dai, Alexander Liniger, and Luc Van Gool. Deep gradient learning for efficient camouflaged object detection. Machine Intelligence Research, 20:92-108, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[26] Tilke Judd, Krista Ehinger, Frédo Durand, and Antonio Torralba. Learning to predict where humans look. In The IEEE/CVF International Conference on Computer Vision, pages 2106-2113. IEEE, 2009. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 218 + ], + "type": "text", + "content": "[27] Xunfa Lai, Zhiyu Yang, Jie Hu, Shengchuan Zhang, Liujuan Cao, Guannan Jiang, Zhiyu Wang, Songan Zhang, and Rongrong Ji. Camoteacher: Dual-rotation consistency learning for semi-supervised camouflaged object detection. In European Conference on Computer Vision, 2024. 1, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "text", + "content": "[28] Jiaying Lin, Xin Tan, Ke Xu, Lizhuang Ma, and Rynson W. H. Lau. Frequency-aware camouflaged object detection. ACM Trans. Multimedia Comput. Commun. Appl., 19 (2), 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 265, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 294, + 319 + ], + "type": "text", + "content": "[29] Jianghang Lin, Yunhang Shen, Bingquan Wang, Shaohui Lin, Ke Li, and Liujuan Cao. Weakly supervised open-vocabulary object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3404-3412, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 294, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 375 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 375 + ], + "type": "text", + "content": "[30] Naisong Luo, Yuwen Pan, Rui Sun, Tianzhu Zhang, Zhiwei Xiong, and Feng Wu. Camouflaged instance segmentation via explicit de-camouflaging. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 17918-17927, 2023. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 377, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 420 + ], + "type": "text", + "content": "[31] Ran Margolin, Lihi Zelnik-Manor, and Ayellet Tal. How to evaluate foreground maps? In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 248-255, 2014. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 477 + ], + "type": "text", + "content": "[32] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 8364-8375, 2022. 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 479, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 544 + ], + "type": "text", + "content": "[33] Peng Mi, Jianghang Lin, Yiyi Zhou, Yunhang Shen, Gen Luo, Xiaoshuai Sun, Liujuan Cao, Rongrong Fu, Qiang Xu, and Rongrong Ji. Active teacher for semi-supervised object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14482-14491, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 545, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 601 + ], + "type": "text", + "content": "[34] Tam Nguyen, Maximilian Dax, Chaithanya Kumar Mummadi, Nhung Ngo, Thi Hoai Phuong Nguyen, Zhongyu Lou, and Thomas Brox. Deepusps: Deep robust unsupervised saliency prediction via self-supervision. Advances in Neural Information Processing Systems, 32, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 602, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 644 + ], + "type": "text", + "content": "[35] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 713 + ], + "type": "text", + "content": "[36] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou," + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "type": "text", + "content": "Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 3, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "type": "text", + "content": "[37] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 162, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 215 + ], + "type": "text", + "content": "[38] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoomnext: A unified collaborative pyramid network for camouflaged object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024, 1, 2, 3, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 217, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 270 + ], + "type": "text", + "content": "[39] Federico Perazzi, Philipp Krahenbuhl, Yael Pritch, and Alexander Hornung. Saliency filters: Contrast based filtering for salient region detection. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 733-740. IEEE, 2012. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 272, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 293 + ], + "type": "text", + "content": "[40] Alec Radford. Improving language understanding by generative pre-training. OpenAI blog, 2018. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 294, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 327 + ], + "type": "text", + "content": "[41] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "text", + "content": "[42] Jingjing Ren, Xiaowei Hu, Lei Zhu, Xuemiao Xu, Yangyang Xu, Weiming Wang, Zijun Deng, and Pheng-Ann Heng. Deep texture-aware features for camouflaged object detection. IEEE Transactions on Circuits and Systems for Video Technology, 33(3):1157-1167, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "text", + "content": "[43] Gyungin Shin, Samuel Albanie, and Weidi Xie. Unsupervised salient object detection with spectral cluster voting. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 3971-3980, 2022. 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 426, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 491 + ], + "type": "text", + "content": "[44] Oriane Simeoni, Chloe Sekkat, Gilles Puy, Antonín Vobecký, Éloi Zablocki, and Patrick Pérez. Unsupervised object localization: Observing the background to discover objects. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 3176-3186, 2023. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 493, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 514 + ], + "type": "text", + "content": "[45] Dongyue Sun, Shiyao Jiang, and Lin Qi. Edge-aware mirror network for camouflaged object detection, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 515, + 553, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 579 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 579 + ], + "type": "text", + "content": "[46] Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. Bert4rec: Sequential recommendation with bidirectional encoder representations from transformer. In Proceedings of the 28th ACM international conference on information and knowledge management, pages 1441-1450, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 581, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 553, + 624 + ], + "type": "text", + "content": "[47] Yujia Sun, Geng Chen, Tao Zhou, Yi Zhang, and Nian Liu. Context-aware cross-level fusion network for camouflaged object detection. In The International Joint Conference on Artificial Intelligence, pages 1025–1031, 2021. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "text", + "content": "[48] Yujia Sun, Shuo Wang, Chenglizhao Chen, and Tian-Zhu Xiang. Boundary-guided camouflaged object detection. In The International Joint Conference on Artificial Intelligence, pages 1335-1341, 2022. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[49] Yanguang Sun, Hanyu Xuan, Jian Yang, and Lei Luo. Glconet: Learning multisource perception representation for camouflaged object detection. IEEE Transactions on Neural Networks and Learning Systems, 2024. 1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "30374" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[50] Domen Tabernik, Samo Šela, Jure Skvarč, and Danijel Skočaj. Segmentation-based deep-learning approach for surface-defect detection. Journal of Intelligent Manufacturing, 31(3):759-776, 2020. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "text", + "content": "[51] Andrey Voynov, Stanislav Morozov, and Artem Babenko. Object segmentation without labels with large-scale generative models. In International Conference on Machine Learning, pages 10596-10606. PMLR, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "type": "text", + "content": "[52] Qingwei Wang, Jinyu Yang, Xiaosheng Yu, Fangyi Wang, Peng Chen, and Feng Zheng. Depth-aided camouflaged object detection. In Proceedings of the 31st ACM International Conference on Multimedia, page 3297–3306, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 260 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 260 + ], + "type": "text", + "content": "[53] Yangtao Wang, Xi Shen, Shell Xu Hu, Yuan Yuan, James L Crowley, and Dominique Vaufreydaz. Self-supervised transformers for unsupervised object discovery using normalized cut. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 14543-14553, 2022. 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "type": "text", + "content": "[54] Yichen Wei, Fang Wen, Wangjiang Zhu, and Jian Sun. Geodesic saliency using background priors. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part III 12, pages 29-42. Springer, 2012. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 317, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 294, + 360 + ], + "type": "text", + "content": "[55] Among Wu, Rui Liu, Yahong Han, Linchao Zhu, and Yi Yang. Vector-decomposed disentanglement for domain-invariant object detection. The IEEE/CVF International Conference on Computer Vision, 2021. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 392 + ], + "type": "text", + "content": "[56] Mochu Xiang, Jing Zhang, Yunqiu Lv, Aixuan Li, Yiran Zhong, and Yuchao Dai. Exploring depth contribution for camouflaged object detection, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "type": "text", + "content": "[57] Qiong Yan, Li Xu, Jianping Shi, and Jiaya Jia. Hierarchical saliency detection. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 1155-1162, 2013. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 426, + 294, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 482 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 482 + ], + "type": "text", + "content": "[58] Fan Yang, Qiang Zhai, Xin Li, Rui Huang, Ao Luo, Hong Cheng, and Deng-Ping Fan. Uncertainty-guided transformer reasoning for camouflaged object detection. In The IEEE/CVF International Conference on Computer Vision, pages 4146-4155, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 483, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 483, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 483, + 294, + 525 + ], + "type": "text", + "content": "[59] Siyuan Yao, Hao Sun, Tian-Zhu Xiang, Xiao Wang, and Xiaochun Cao. Hierarchical graph interaction transformer with dynamic token clustering for camouflaged object detection. IEEE Transactions on Image Processing, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 526, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 581 + ], + "type": "text", + "content": "[60] Bowen Yin, Xuying Zhang, Deng-Ping Fan, Shaohui Jiao, Ming-Ming Cheng, Luc Van Gool, and Qibin Hou. Camoformer: Masked separable attention for camouflaged object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "type": "text", + "content": "[61] Qiang Zhai, Xin Li, Fan Yang, Chenglizhao Chen, Hong Cheng, and Deng-Ping Fan. Mutual graph learning for camouflaged object detection. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 12997-13007, 2021. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "type": "text", + "content": "[62] Jing Zhang, Tong Zhang, Yuchao Dai, Mehrtash Harandi, and Richard Hartley. Deep unsupervised saliency detection: A multiple noisy labeling perspective. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 9029-9038, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "text", + "content": "[63] Yi Zhang and Chengyi Wu. Unsupervised camouflaged object segmentation as domain adaptation. In The IEEE/CVF" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 374 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "International Conference on Computer Vision, pages 4334-4344, 2023. 1, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[64] Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, and Nicu Sebe. Bilateral reference for high-resolution dichotomous image segmentation. CAAI Artificial Intelligence Research, 3:9150038, 2024. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "text", + "content": "[65] Yijie Zhong, Bo Li, Lv Tang, Senyun Kuang, Shuang Wu, and Shouhong Ding. Detecting camouflaged object in frequency domain. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 4494-4503, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 186, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 240 + ], + "type": "text", + "content": "[66] Hongwei Zhu, Peng Li, Haoran Xie, Xu Yan, Dong Liang, Dapeng Chen, Mingqiang Wei, and Jing Qin. I can find you! boundary-guided separated attention network for camouflaged object detection. In Association for the Advancement of Artificial Intelligence, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 241, + 555, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 555, + 285 + ], + "type": "text", + "content": "[67] Jinchao Zhu, Xiaoyu Zhang, Shuo Zhang, and Junnan Liu. Inferring camouflaged objects by texture-aware interactive guidance network. In Association for the Advancement of Artificial Intelligence, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "text", + "content": "[68] Wangjiang Zhu, Shuang Liang, Yichen Wei, and Jian Sun. Saliency optimization from robust background detection. In The IEEE/CVF Computer Vision and Pattern Recognition Conference, pages 2814-2821, 2014. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "text", + "content": "[69] Mingchen Zhuge, Xiankai Lu, Yiyou Guo, Zhihua Cai, and Shuhan Chen. Cubenet: X-shape connection for camouflaged object detection. Pattern Recognition, 127:108644, 2022. 3" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "30375" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/UHD-processer_ Unified UHD Image Restoration with Progressive Frequency Learning and Degradation-aware Prompts/2e2768b6-2887-4bd5-935c-d0b880ebe319_content_list.json b/2025/UHD-processer_ Unified UHD Image Restoration with Progressive Frequency Learning and Degradation-aware Prompts/2e2768b6-2887-4bd5-935c-d0b880ebe319_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c9dc62639b7952d8860512ec7414a245de389fce --- /dev/null +++ b/2025/UHD-processer_ Unified UHD Image Restoration with Progressive Frequency Learning and Degradation-aware Prompts/2e2768b6-2887-4bd5-935c-d0b880ebe319_content_list.json @@ -0,0 +1,2070 @@ +[ + { + "type": "text", + "text": "UHD-processor: Unified UHD Image Restoration with Progressive Frequency Learning and Degradation-aware Prompts", + "text_level": 1, + "bbox": [ + 166, + 130, + 833, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yidi Liu, DongLi, Xueyang Fu†, Xin Lu, Jie Huang, Zheng-jun Zha \nSchool of Information Science and Technology and MoE Key Laboratory of \nBrain-inspired Intelligent Perception and Cognition, \nUniversity of Science and Technology of China, Hefei, 230026, China", + "bbox": [ + 197, + 203, + 800, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "liuyidi2023@mail.ustc.edu.cn, xyfu@ustc.edu.cn † Corresponding Author", + "bbox": [ + 210, + 277, + 781, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 328, + 326, + 344 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce UHD-Processor, a unified and robust framework for all-in-one image restoration, which is particularly resource-efficient for Ultra-High-Definition (UHD) images. To address the limitations of traditional all-in-one methods that rely on complex restoration backbones, our strategy employs a frequency domain decoupling progressive learning technique, motivated by curriculum learning, to incrementally learn restoration mappings from low to high frequencies. This approach incorporates specialized sub-network modules to effectively tackle different frequency bands in a divide-and-conquer manner, significantly enhancing the learning capability of simpler networks. Moreover, to accommodate the high-resolution characteristics of UHD images, we developed a variational autoencoder (VAE)-based framework that reduces computational complexity by modeling a concise latent space. It integrates task-specific degradation awareness in the encoder and frequency selection in the decoder, enhancing task comprehension and generalization. Our unified model is able to handle various degradations such as denoising, deblurring, dehazing, lowlighting, etc. Experimental evaluations extensively showcase the effectiveness of our dual-strategy approach, significantly improving UHD image restoration and achieving cutting-edge performance across diverse conditions. The code will be available at https://github.com/1yd2022/UHD-processor", + "bbox": [ + 89, + 361, + 485, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 782, + 223, + 799 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ultra-high-definition (UHD) images are increasingly utilized across various fields, prompting substantial research in UHD image restoration for specific degradation types, such as low-light enhancement, deblurring, dehazing, and deraining [2, 15, 29, 36]. However, real-world applications like object detection [21], video segmentation [16], and au", + "bbox": [ + 89, + 809, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/72d4c86909474046daf42a55f096b0941a7b360703e19113aa0530755009f44b.jpg", + "image_caption": [ + "(a) Comparison of PSNR with SOTA All-in-One Methods" + ], + "image_footnote": [], + "bbox": [ + 535, + 335, + 751, + 497 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f468997ca2074c2028c01d8cca464a3a0eba431f2f695ded2e63a6e228823f9e.jpg", + "image_caption": [ + "(b) Observation 1" + ], + "image_footnote": [], + "bbox": [ + 759, + 339, + 895, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b8fadce7ba92206ae2a8ccd616838cb66522413d76ac01bc0b22f67c0f24756b.jpg", + "image_caption": [ + "(c) Observation 2" + ], + "image_footnote": [], + "bbox": [ + 761, + 428, + 895, + 500 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/046f1dc832e655b802a32dcf6eebc7f7feb05c97b3c5411637ead85369738392.jpg", + "image_caption": [ + "(d) Observation 3: The Variation of Gap Between Degradations Across Frequency Bands" + ], + "image_footnote": [], + "bbox": [ + 532, + 510, + 767, + 608 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c6efaf951ec6ec0d222cdb2f14ea555fc4b1229a514857e39b200d2a1edddfca.jpg", + "image_caption": [ + "Figure 1. (a) Comparison of PSNR between our method and the previous state-of-the-art (SOTA) all-in-one method across six degradation settings in UHD images. (b) Observe the performance gains when integrating the same promptIR prompt branch into Restormer models with varying parameter sizes. (c) Comparison of low-to-high frequency training order on UHD and regular images, showing greater effectiveness for UHD tasks. (d) Quantitative evaluation of restoration tasks across low, high, and full-frequency bands." + ], + "image_footnote": [], + "bbox": [ + 767, + 511, + 898, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tonomous driving [3] face complex degradation patterns that vary across time and space, making reliance on single-function models impractical. This calls for a generalized model that can address multiple degradation types to support robust low-level vision tasks.", + "bbox": [ + 511, + 747, + 906, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In response to the above requirements, all-in-one(AIO) approaches aim to use a single model to manage various degradations, but they generally target standard-resolution images and encounter computational difficulties when scaled to UHD. This gap highlights the pressing need for an efficient", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23121", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "all-in-one image restoration framework tailored to UHD images. For standard-resolution tasks [14, 20], successful approaches often include two key features, i.e., degradation-aware branches and strong network backbones. The former is implemented through techniques like prompts [10] or mixtures of experts (MoE) [47] to enhance adaptability to different degradation types, while the later contains advanced backbones like Restormer [40] to further strengthens learning and generalization.", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, due to the high-resolution nature of UHD images, directly applying these strategies results in a significant computational burden, making full-resolution inference infeasible on consumer-grade GPUs. As shown in Figure 1(b), PromptIR extends the Restormer backbone with a prompt branch. We adjust their parameter counts by varying Restormer's module numbers, keeping the prompt branch fixed. Reducing parameters significantly lowers performance for both, meanwhile, the benefits introduced by the prompt branch also gradually diminish.", + "bbox": [ + 89, + 229, + 480, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These results indicate that while the prompt branch enhances all-in-one task performance, a strong backbone is crucial to fully leverage these improvements. Moreover, compared to Restormer, the additional degradation-aware prompt branch introduced by PromptIR adds 6.87M parameters, which is clearly burdensome for UHD tasks.", + "bbox": [ + 89, + 381, + 480, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the above observations, this work aims to tackle two challenges faced by traditional AIO methods in UHD scenarios: reliance on complex restoration backbones and redundancy in degradation-aware branch parameters.", + "bbox": [ + 89, + 474, + 480, + 535 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle the first challenge, at the model architecture level, we propose a VAE-based framework that transforms restoration tasks from pixel space into a compact latent space, thereby reducing the computational cost of UHD image processing. This alleviates the network's learning burden, enabling a simpler model to restore effectively. At the learning strategy level, inspired by curriculum learning [34], we develop a Frequency Domain Decoupling Progressive Learning Strategy. As shown in Figure 1(d), for UHD AIO tasks, higher frequency bands present greater restoration challenges. Consequently, we decompose the task into manageable frequency components, enabling sub-networks to progressively learn restoration mappings from low to high frequencies. This divide-and-conquer approach significantly enhances the network's adaptability and efficiency.", + "bbox": [ + 89, + 536, + 482, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the second challenge, we introduce efficient prompt strategies for both the VAE encoder and decoder. For encoding, we propose a low-rank degradation-aware prompt to indicate degradation type and severity. This is achieved by creating distinct low-rank parameter spaces per degradation type and learning scale-invariant convolution kernels, which adapt to UHD resolution differences between training and testing. For efficiency, we reparameterize these convolution branches during inference. To maintain efficiency,", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "we reparameterize these convolution branches at inference. In decoding, we observe that degradation types exhibit distinct distributions across frequency bands. Accordingly, we introduce a frequency band selection prompt, allowing the decoder to adaptively choose between low-frequency features from the decoder and mid- or high-frequency features from the encoder based on the specific degradation type.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By integrating these components, we propose UHD-Processor, an efficient all-in-one image restoration framework tailored for ultra-high-definition (UHD) images. As shown in Figure 1(a), our proposed UHD-Processor achieves state-of-the-art performance while reducing parameters by $95.1\\%$ and FLOPs by $97.4\\%$ compared to promptIR [20]. Our contributions are summarized as follows:", + "bbox": [ + 511, + 198, + 903, + 301 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We introduce a VAE-based image restoration framework that operates within a compact latent space, significantly enhancing UHD image restoration efficiency.", + "2. We design a frequency domain decoupling progressive learning strategy, which enables effective restoration across frequency bands from low to high frequencies through a divide-and-conquer approach, to improve adaptability to various optimization objectives.", + "3. We develop a prompt learning framework with low-rank degradation-aware prompts for encoding and frequency selection prompts for decoding, improving the network's adaptability to various degradation types." + ], + "bbox": [ + 514, + 306, + 903, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 502, + 653, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Ultra-High-Definition Restoration", + "text_level": 1, + "bbox": [ + 513, + 527, + 810, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As the demand for UHD image processing grows in imaging systems, various approaches have been proposed to restore clarity in UHD images through diverse network architectures. These include bilateral learning for image dehazing [45], multi-scale separable-patch integration networks for video deblurring [24], and Fourier embedding networks for low-light image enhancement [15]. Recently, generalized architectures for UHD image restoration have evolved to address multiple restoration tasks through unified frameworks. For instance, [29] introduces a correlation matching mechanism between high-resolution and low-resolution branches to enhance the utilization of native resolution information, while [31] incorporates gradient and normal priors into the model design. However the approach about UHD all in one is not yet fully explored.", + "bbox": [ + 511, + 550, + 903, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. All-in-One Image Restoration", + "text_level": 1, + "bbox": [ + 513, + 787, + 777, + 803 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "All-in-one image restoration, which aims to recover clean images from various types of degradation using a single unified model, has gained significant attention for its efficiency and practical applications in low-level vision tasks [14, 20, 33, 38, 41]. Early methods, such as AirNet [14], utilized contrastive learning to extract degra", + "bbox": [ + 511, + 810, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23122", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dation representations, while IDR [41]introduced a two-stage meta-learning approach to address diverse degradation types. More recently, prompt-based approaches like PromptIR [8]and PIP [17]further enhanced restoration flexibility through vision prompts. Additionally, large-scale pre-trained models, as demonstrated by DiffUIR [44] and DA-CLIP [18], have been adapted for all-in-one restoration with strong performance. Please refer to the supplementary material for more detailed related work.", + "bbox": [ + 89, + 90, + 485, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 242, + 181, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 89, + 268, + 200, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The framework of our method is shown in Figure 2. Overall, our approach follows a two-stage paradigm. In the first stage, we train a VAE on clean images with an image reconstruction task, as shown in Figure 2(a). In the second stage, as shown in Figure 2(b), we freeze the VAE weights and train the remaining components for image restoration tasks. To mitigate high-frequency loss during the VAE encoding process, we pass high-frequency components from the wavelet transformation between the encoder adapter and the decoder adapter. Additionally, prompts are injected through the adapters: in the encoding phase, we apply degradation-aware low-rank prompts (Figure 2(d)) to adapt the VAE to different degradations, encoding them into a unified compact latent space; in the decoding phase, we use frequency selection prompts (Figure 2(c)) for each type of degradation, enabling efficient adaptive fusion of low-frequency decoder information with high-frequency encoder information.", + "bbox": [ + 89, + 291, + 483, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The latent restoration network is divided into $M$ subnetworks, each corresponding to one of the $M$ decomposed frequency bands of the input. During Frequency Domain Decoupling Progressive Learning (Figure 2(e)), we progressively increase from the lowest to the highest frequency bands, incrementally adding sub-networks until all frequency bands are restored. Attributing the strong performance to the robust latent space encoding and frequency-progressive learning strategy, we achieve excellent results with a simple latent space restoration network. In this work, we adopt the Cubic-Mixer [45] for its efficiency, results of other architectures and the interaction detail between the adapter and VAE, please refer to the supplementary materials.", + "bbox": [ + 89, + 549, + 483, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Frequency Domain Decoupling Progressive Learning Strategy", + "text_level": 1, + "bbox": [ + 89, + 757, + 482, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For all-in-one tasks, which aim to handle multiple types of degradation with a single model, we can represent this as $F(\\{\\tau^t\\}; \\theta)_{t \\in [1, N]}$ , where $N$ represents the number of tasks, $\\tau^t$ denotes the input for task $t$ , and $\\theta$ represents the model parameters. Each task $\\tau^t$ corresponds to a distinct optimization objective; for instance, the low-light enhancement task focuses on adjusting the global brightness of the image, while", + "bbox": [ + 89, + 794, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the deraining task aims to remove local rain streaks. Consequently, it is challenging for a network trained from scratch to identify a suitable gradient direction for all tasks during initial training, which often results in slow convergence or convergence to suboptimal solutions.", + "bbox": [ + 511, + 90, + 906, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by the concept of curriculum learning [34][19], a learning process that progresses from easy to difficult tasks is more beneficial for optimizing complex tasks. MIO-IR [13] employs a sequential learning strategy to incrementally add tasks, thereby mitigating the interference caused by joint training. However, while extending from single-task to multitask learning is a form of curriculum learning, the relationships among tasks remain ambiguous, making it challenging to establish a generalized strategy for determining the optimal task sequence. Therefore, we revisit how to design a more general curriculum learning strategy for UHD all-in-one tasks.", + "bbox": [ + 511, + 167, + 908, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1. Observation", + "text_level": 1, + "bbox": [ + 513, + 361, + 645, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "First, we assess which processes are easier or harder for a neural network to learn during training, starting from random initialization to being capable of handling multiple degradations effectively. Inspired by the frequency principle of neural networks [23][25][37][30], we approach this problem from the perspective of the frequency domain.", + "bbox": [ + 511, + 381, + 906, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For a restoration task $\\tau^t$ , we decompose it into $M$ frequency bands, with the $f$ -th frequency band denoted as $\\tau_f^t$ . The multi-degradation restoration process is represented as $F(\\tau_f^t;\\theta)_{t\\in [1,N],f\\in [1,M]}$ . In sequential learning, tasks are learned progressively from 1 to $N$ , where each stage involves restoring all frequency bands, represented as $F(\\{\\tau_f^t\\}_{f = 1}^M;\\theta)_{t = 1\\to N}$ . The main drawback of this approach is the lack of strong inter-stage connections, and the relationships between all degradation types are only modeled in the final stage, which hinders the network's ability to comprehensively learn the interplay between different degradations.", + "bbox": [ + 511, + 473, + 908, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address this, we shift from task-based to frequency band-based progressive learning, represented as $F(\\{\\tau_f^t\\}_{t=1}^N; \\theta)_{f=1 \\to M}$ . Each frequency band captures global image information and maintains a physical relationship with other bands. Learning by frequency bands is smoother and more natural, as knowledge from earlier bands supports learning subsequent bands.", + "bbox": [ + 511, + 642, + 908, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To analyze the differences between degradations under low- and high-frequency conditions, we conducted a toy experiment. We selected four types of degradations: low-light and haze, which predominantly affect the low-frequency bands of an image, and denoising and deblurring, which mainly impact the high-frequency bands. We applied a low-pass filter to decompose images into low- and high-frequency components, then trained separate models for each. To quantify the gap between different tasks, we introduced the metric $\\hat{z}_{ij}^{s}$ from [43], which is used to measure the discrepancy", + "bbox": [ + 511, + 750, + 910, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23123", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f04361381cdecd0c74ec0b94828d0c107be49b16b57d7f8a45fec4282c0f08c3.jpg", + "image_caption": [ + "Figure 2. The overall framework of our proposed UHD-Processor, which includes: (a) Pre-training a VAE on clean images for an image reconstruction task. (b) During the image restoration task, we keep the VAE frozen and efficiently fine-tune it using an adapter, where latentIR represents the latent space image restoration network. (c) Degradation-Specific Frequency Selection Prompt (DFSP) for the decoder. (d) Degradation-Aware Low Rank Prompt (DALR) for the encoder, and (e) an illustration of the Frequency Domain Decoupling Progressive Learning Strategy." + ], + "image_footnote": [], + "bbox": [ + 98, + 89, + 338, + 467 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ff5635cbce2670a008074d46ed9c56c733d8478887a37414bb97637eeb3da48e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 343, + 89, + 901, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "between pre-training and fine-tuning tasks (higher values indicate smaller differences). For a detailed discussion of this metric, please refer to the supplementary material.", + "bbox": [ + 88, + 575, + 482, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Experimental results in Figure 1(d) indicate that, for both low and high frequencies, the value of $\\hat{z}_{ij}^{s}$ is consistently higher compared to the entire frequency spectrum. This underscores the critical role of frequency-wise decomposition in minimizing task-specific discrepancies within individual frequency bands. Moreover, $\\hat{z}_{ij}^{s}$ is significantly higher in the low-frequency band than in the high-frequency band, suggesting that a progressive learning sequence from low to high frequencies is better suited for establishing an effective curriculum learning framework.", + "bbox": [ + 88, + 625, + 482, + 775 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further investigate the differences in conclusions between standard and UHD resolutions, we conducted experiments on UHD data for four types of degradations and their corresponding downsampled versions using an all-in-one training setup. As shown in Figure 1(c), By adopting a training sequence progressing from low to high frequencies, we observed improvements for both resolutions compared to direct training. However, the enhancement was more pro", + "bbox": [ + 88, + 779, + 485, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "nounced in UHD scenarios, which can be attributed to the richer high-frequency texture information present in UHD images.", + "bbox": [ + 511, + 575, + 906, + 621 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Frequency Domain Decoupling Progressive Learning Strategy", + "text_level": 1, + "bbox": [ + 511, + 633, + 906, + 665 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, we propose the Frequency Decoupling Progressive Learning strategy, which leverages multiple wavelet transformation (MWT) to decompose the input image $X$ and ground truth $\\Upsilon$ into multiple frequency sub-bands $\\{X_f\\}_{f=1}^M$ and $\\{Y_f\\}_{f=1}^M$ .", + "bbox": [ + 511, + 670, + 908, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During training, the learning process progressively advances from low to high frequencies, with each frequency band $X_{f}$ managed by a dedicated restoration sub-network $F_{f}$ . As training progresses and higher frequency bands are introduced, additional restoration sub-networks are incrementally incorporated, forming a divide-and-conquer approach to effectively handle each frequency band. The process can be formulated as Algorithm 1. This pseudocode omits the detailed steps of multiple wavelet transformation (MWT) and its corresponding inverse process (MIWT). For", + "bbox": [ + 511, + 750, + 910, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23124", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a more comprehensive description of the algorithm, please refer to the supplementary materials.", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 Frequency Domain Decoupling Progressive Learning", + "bbox": [ + 91, + 133, + 482, + 165 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Input: $\\{X_f\\}_{f=1}^M, \\{Y_f\\}_{f=1}^M \\gets \\mathrm{MWT}(X), \\mathrm{MWT}(Y)$", + "2: Init: Initialize $\\{\\theta_f\\}_{f=1}^M$ for sub-networks, set stage thresholds $\\{T_k\\}_{k=1}^M$ , and define total_iter", + "3: current_iter $\\leftarrow 0, k \\leftarrow 1$", + "4: while current_iter $\\leq$ total_iter do", + "5: $\\{\\hat{X}_f\\mid f = 1,\\dots ,k\\} \\gets F_f(X_f;\\theta_f)$", + "6: $\\dot{X}\\gets \\mathrm{MIWT}(\\{\\dot{X}_f\\}_{f = 1}^k,\\{Y_f\\}_{f = k + 1}^M)$", + "7: $\\mathcal{L}\\gets \\operatorname {Loss}(\\hat{X},Y)$", + "8: $\\{\\theta_f\\}_{f = 1}^k\\gets \\{\\theta_f\\}_{f = 1}^k -\\alpha \\nabla_{\\{\\theta_f\\}_{f = 1}^k}\\mathcal{L}(\\tilde{X},Y)$", + "9: current_iter $\\leftarrow$ current_iter + 1", + "10: if current_iter $>T_{k}$ and $k < M$ then", + "11: $k\\gets k + 1$", + "12: end if", + "13: end while" + ], + "bbox": [ + 99, + 169, + 480, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Efficient Adaptive Prompt Learning", + "text_level": 1, + "bbox": [ + 89, + 416, + 403, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on our VAE-based restoration framework, we design two different prompt strategies for the distinct behaviors of the VAE encoding and decoding processes. For the encoding process, since the VAE is pre-trained on clean images, directly encoding various degraded images introduces a significant domain gap. Therefore, we introduce a low-rank degradation-aware prompt to the encoder, enabling it to encode different degraded images into the pre-trained compact and regularized latent space, thereby reducing the difficulty of the subsequent restoration mapping.", + "bbox": [ + 88, + 438, + 483, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the decoding stage, despite the successful mitigation of degradation effects within the latent space, challenges persist with the loss of high-frequency details when decoding back to pixel space. To address this issue, we incorporate an adapter to inject high-frequency information from the corresponding stages of the encoder into the decoder. However, since different degradations impact specific frequency bands differently, we employ a frequency-selective prompt in the decoding process. This approach facilitates an adaptive fusion of high-frequency details from the encoder with the low-frequency information from the decoder, tailored to the specific frequency bands affected by the degradation.", + "bbox": [ + 89, + 589, + 483, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, to allow the network to effectively recognize the type of degradation, we pre-train a degradation classifier. Since this classification task is not particularly challenging, a simple classification network is sufficient to achieve high accuracy.", + "bbox": [ + 89, + 770, + 483, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.1. Degradation-Aware Low Rank Prompt", + "text_level": 1, + "bbox": [ + 89, + 851, + 408, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The current methods for prompt injection in all-in-one tasks primarily follow the design of PromptIR [20]. In this frame", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "work, for the input feature $F_{i}\\in \\mathbb{R}^{H\\times W\\times C}$ , a learnable prompt tensor $P_{c}\\in \\mathbb{R}^{N\\times \\hat{H}\\times \\hat{W}\\times C}$ is first constructed, where $N$ denotes the number of prompt components, and $\\hat{H}$ and $\\hat{W}$ represent the spatial dimensions, while the channel count $C$ aligns with that of the input. Subsequently, $P_{c}$ is interpolated to match the size of $F_{i}$ , enabling further interaction to generate the prompts.", + "bbox": [ + 511, + 89, + 906, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While this design has shown promising performance in tasks at standard resolutions, it faces two significant challenges in UHD scenarios. First, the entire learnable tensor $P_{c}$ comprises a substantial number of parameters due to the dimensions $\\hat{H} \\times \\hat{W}$ . Furthermore, the prompt is homogeneous across all degradation types, which means its complexity cannot be adapted based on the difficulty of each specific degradation. This lack of customization makes the approach inefficient for UHD image restoration, where the difficulty of different degradation categories can vary significantly.", + "bbox": [ + 511, + 199, + 906, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Second, since the dimensions $\\hat{H}$ and $\\hat{W}$ of $P_{c}$ must be predetermined and interpolated to ensure adaptability to feature resolution, traditional all-in-one tasks exhibit minimal differences between training and inference resolutions. This allows the trained $P_{c}$ to generalize well during inference. However, in UHD tasks, training typically uses a patch size of $512 \\times 512$ for efficiency, while inference requires processing full images at $3840 \\times 2160$ . This significant resolution gap limits the generalization of $P_{c}$ during inference.", + "bbox": [ + 511, + 349, + 906, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address the issue of parameter efficiency, inspired by the low-rank adaptation method [9], we perform a low-rank decomposition on the spatial dimensions of the original parameter matrix $P_{c}\\in \\mathbb{R}^{N\\times \\hat{H}\\times \\hat{W}\\times C}$ , represented as:", + "bbox": [ + 511, + 487, + 906, + 550 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {c l} = \\operatorname {c a t} \\left[ B _ {n} A _ {n} \\right] _ {n \\in [ 1 \\dots N ]}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 563, + 906, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $B_{n}\\in \\mathbb{R}^{1\\times \\hat{H}\\times r\\times C}$ and $A_{n}\\in \\mathbb{R}^{1\\times r\\times \\hat{W}\\times C}$ represent the matrices resulting from the low-rank decomposition, and $\\operatorname {cat}[\\cdot ]$ represents concatenation along the first dimension, with $r\\ll \\min (\\hat{H},\\hat{W})$ . After low-rank decomposition, the original parameter matrix is effectively compressed.", + "bbox": [ + 511, + 594, + 905, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Furthermore, for different types of degradations, we can specify different ranks $r$ . More complex degradations can use a larger $r$ to enhance the parameter space, while simpler degradations can use a smaller $r$ to improve parameter efficiency. The use of different ranks $r$ introduces unique designs for various degradations, while the multiplication of the decomposed matrices preserves consistent dimensions, allowing efficient parallel processing through concatenation.", + "bbox": [ + 511, + 672, + 906, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address the issue of resolution adaptability, after obtaining the low-rank prompt tensor $P_{cl}$ , we do not directly interpolate it to match the input features. Instead, considering the scale invariance of convolution operations, we utilize $P_{cl}$ as a parameter space to learn multiple convolution operators. By using convolution as the core operator for prompt interactions, this approach is more effective in adapting to", + "bbox": [ + 511, + 795, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23125", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the differences between inference and training resolutions compared to simple interpolation.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, the prompt generation process during the encoding phase is as follows. First, the input image is passed through a classifier to obtain $N + 1$ classification scores, where $N$ represents the number of degradation types, and 1 denotes the clean image. We then further process these classification scores to derive $W_{1}$ and $W_{2}$ :", + "bbox": [ + 89, + 121, + 483, + 212 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nW _ {1} = \\operatorname {s o f t m a x} (\\sum (N), 1), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 217, + 483, + 244 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nW _ {2} = \\operatorname {s o f t m a x} (N),\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 244, + 331, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where, $W_{1}$ represents the binary classification score between degraded and clean images, serving to characterize the degree of degradation, while $W_{2}$ corresponds to the $N$ -way classification scores for the $N$ different degradation types, indicating the specific type of degradation present.", + "bbox": [ + 89, + 265, + 483, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We then construct two simple pointwise convolutions to provide degradation degree prompts(DDP), as illustrated below:", + "bbox": [ + 89, + 340, + 483, + 383 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {i 1} ^ {\\prime} = \\operatorname {c a t} \\left[ P _ {\\text {d e g}} \\left(F _ {i}\\right), P _ {\\text {c l e a n}} \\left(F _ {i}\\right) \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 388, + 483, + 412 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {i 1} = F _ {i 1} ^ {\\prime} \\odot W _ {1},\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 409, + 295, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where, $P_{deg}(\\cdot)$ and $P_{clean}(\\cdot)$ denote the pointwise convolutions used to provide degradation degree prompts, while $\\odot$ represents the weighted summation of their outputs.", + "bbox": [ + 89, + 431, + 483, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After obtaining the degradation-aware feature $F_{i1}$ , we proceed to provide prompts for degradation types(DTP). First, we define the rank of the parameter space for each degradation type as $R_{N} = [r_{1}, r_{2}, \\dots, r_{N}]$ , which allows us to derive the parameter space $P_{cl}$ based on Equation (1). Next, we utilize a simple MLP to generate a set of deep-wise convolution kernels $K_{s} = [(k_{1}:d_{1}), (k_{2}:d_{2}), \\dots, (k_{m}:d_{m})]$ , characterized by varying kernel sizes and dilation rates.", + "bbox": [ + 89, + 479, + 483, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The feature $F_{i1}$ is then processed in parallel through all multi-scale convolution group (MCKG), followed by a weighted summation with $W_{2}$ . Due to the impact of parallel multi-branch convolutional kernels on inference efficiency [6], we employ the structural reparameterization method from [7]. In the inference phase, all convolutional kernels are reparameterized and merged into a single large convolutional kernel. This process can be expressed as follows:", + "bbox": [ + 89, + 599, + 483, + 733 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nK _ {s} = \\mathrm {M L P} _ {i \\in [ 1: m ]} (P _ {c l}),\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 738, + 303, + 755 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {i 2} ^ {\\prime} = \\left\\{ \\begin{array}{l l} \\sum \\left(K _ {s} \\left(F _ {i 1}\\right)\\right) & \\text {i f i n t r a i n i n g p h a s e}, \\\\ \\operatorname {R e p} \\left[ K _ {s} \\right] \\left(F _ {i 1}\\right) & \\text {i f i n t e s t i n g p h a s e}, \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 758, + 483, + 797 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {i 2} = W _ {2} \\odot F _ {i 2} ^ {\\prime},\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 801, + 254, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where, $\\mathrm{Rep}[\\cdot ]$ refers to the reparameterization operation. The resultant $F_{i2}$ serves as the low-rank degradation-aware prompt, which is integrated into the encoder's adapter. This mechanism facilitates the encoder's ability to encode diverse degradations into a cohesive and compact latent space.", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3.2. Degradation-Specific Frequency Selection Prompt", + "text_level": 1, + "bbox": [ + 511, + 90, + 906, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given that different degradations affect distinct frequency bands [4] [11], we design prompts in the decoding phase to facilitate adaptive selection of frequency components. Specifically, this involves integrating medium-frequency information $\\{F_E^{HL}, F_E^{LH}\\}$ and high-frequency information $F_E^{HH}$ from the encoder, along with low-frequency information $F_D^{LL}$ from the decoder.", + "bbox": [ + 511, + 109, + 906, + 215 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We begin by applying the Fast Fourier Transform (FFT) to each frequency band feature in the adapter branch, represented as $F_{w} = \\{F_{D}^{LL}, F_{E}^{HL}, F_{E}^{LH}, F_{E}^{HH}\\} \\in \\mathbb{R}^{B \\times 4 \\times \\frac{\\hat{H}}{2} \\times \\frac{\\hat{W}}{2} \\times C}$ . Pointwise convolution is applied to each band to efficiently capture global features and expand the channels by a factor of $N$ , resulting in $F_{w1} \\in \\mathbb{R}^{B \\times 4 \\times N \\times \\frac{\\hat{H}}{2} \\times \\frac{\\hat{W}}{2} \\times C}$ after reshaping. A weighted summation with $W_{2}$ is then performed along the $N$ dimension, followed by the Inverse FFT (IFFT) to obtain $F_{w1}'$ :", + "bbox": [ + 511, + 215, + 906, + 358 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} F _ {w 1} = \\operatorname {c a t} \\left[ \\operatorname {P c o n v} _ {\\mathrm {i} \\in [ \\text {l o w}, \\text {m i d}, \\text {h i g h} ]} \\left(\\mathcal {F} \\left(F _ {w}\\right)\\right) \\right], \\\\ F ^ {\\prime} = \\mathcal {F} ^ {- 1} \\left(W, W, F _ {w}\\right) \\end{array} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 364, + 905, + 391 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {w 1} ^ {\\prime} = \\mathcal {F} ^ {- 1} \\left(W _ {2} \\odot F _ {w 1}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 386, + 736, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where, $\\mathrm{Pconv}_{\\mathrm{i}\\in [\\mathrm{low},\\mathrm{mid},\\mathrm{high}]}$ denotes the pointwise convolutions used to process the low, mid, and high frequency bands, while $\\mathcal{F}(\\cdot)$ and $\\mathcal{F}^{-1}(\\cdot)$ represent the spatial dimensional Fourier transform and its corresponding inverse transform, respectively.", + "bbox": [ + 511, + 410, + 906, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Subsequently, pixel shuffle is applied for frequency mixing, followed by a sigmoid activation to obtain the filter kernel $F_{w2} \\in \\mathbb{R}^{B \\times \\hat{H} \\times \\hat{W} \\times C}$ . For the input $F_{d} \\in \\mathbb{R}^{B \\times \\hat{H} \\times \\hat{W} \\times C}$ from the decoder, we perform FFT to shift to the frequency domain and apply $F_{w2}$ for frequency domain filtering. Finally, the IFFT is conducted to obtain the frequency-selective prompt $F_{w2}'$ :", + "bbox": [ + 511, + 484, + 908, + 594 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {w 2} = \\operatorname {s i g m o i d} \\left(\\operatorname {p i x e l s h u f f u l e} \\left(F _ {w 1} ^ {\\prime}\\right)\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 599, + 836, + 618 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {w 2} ^ {\\prime} = \\mathcal {F} ^ {- 1} \\left(F _ {w 2} \\otimes \\mathcal {F} \\left(F _ {w 2}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 616, + 905, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\otimes$ denotes element-wise multiplication. $F_{w2}^{\\prime}$ serves as a frequency-selective prompt injected through the adapter in the decoder, enabling the decoding process to perform a more customized multi-band feature fusion tailored to the frequency characteristics of the degradation.", + "bbox": [ + 511, + 645, + 905, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 733, + 645, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 757, + 702, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We have developed a comprehensive benchmark for evaluating UHD all-in-one restoration performance, based on datasets from UHD-LL [15], UHD-blur [5], UHD-haze [46], UHD-rain [2], and UHD-haze [31]. Additionally, we constructed a denoising dataset, UHD-noise, using 4K images from [42] as the background. For additional implementation details and dataset descriptions, please refer to the supplementary materials.", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23126", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/989b7689d88e816a6cb0f407cdca5e687ce76acd91ed16c549a77525ac55ef51.jpg", + "image_caption": [ + "Image deblurring" + ], + "image_footnote": [], + "bbox": [ + 96, + 92, + 248, + 169 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/659e7328f41190ef9aa1846f105e5a4ea7f170fc415b6bbf6cb98e12d9a2c993.jpg", + "image_caption": [ + "Histformer" + ], + "image_footnote": [], + "bbox": [ + 253, + 93, + 328, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ac1237d54e30069b3408419500440123ba0fadb2dcc8bbd97262a59bc0fc6022.jpg", + "image_caption": [ + "PromptIR" + ], + "image_footnote": [], + "bbox": [ + 334, + 93, + 410, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4ef8809299c46ac5b60fda594b974ed56d679d4567f35ac0de24b3c15049cca7.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 415, + 93, + 491, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e025b82451bec369a8468f046e4eed96acae49e459d8d1a00fc99cb83972372a.jpg", + "image_caption": [ + "Gridformer" + ], + "image_footnote": [], + "bbox": [ + 254, + 136, + 328, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3820b6dc4862582190abe227faa866e5fd3e6853288b7399e93abd783d5c8264.jpg", + "image_caption": [ + "HAIR" + ], + "image_footnote": [], + "bbox": [ + 334, + 136, + 410, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4e4826a8146fa398f18be36ff90e29f8f63d20e8aa8cf18f43b2215ddf5d85e6.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 415, + 136, + 491, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/57287db71fd0c582a97ad0048c81dd9b229d8a7d75762b03bc5c760525202c5b.jpg", + "image_caption": [ + "Low light enhancement", + "Figure 3. A comparison of visual results for four types of degradation removal with other state-of-the-art (SOTA) all-in-one methods." + ], + "image_footnote": [], + "bbox": [ + 96, + 180, + 248, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4038e26fadfb1aeea679d300e3b75c0d8919f2a9fb71de580f1d61fcf21ea990.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 253, + 181, + 328, + 214 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0414d68c91ff50d09bb4c5fd1bafab0831d3e5d3585462d3d7a56d1efc12f4ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 181, + 410, + 214 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/86b37eaf6ee2b186d50449a933fac29c005824c8c73f25511d0df54b2b43ef43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 415, + 181, + 491, + 214 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d98c9d0f5e6a442d25a6b67bdb9239efbb2454864c9efde931857f9eacf69248.jpg", + "image_caption": [ + "Histformer", + "Gridformer" + ], + "image_footnote": [], + "bbox": [ + 253, + 224, + 328, + 257 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a5d6ac59df501d7b9aad69041876c4697605528f6504e6c44b9da7ef325b3ffa.jpg", + "image_caption": [ + "PromptIR", + "HAIR" + ], + "image_footnote": [], + "bbox": [ + 334, + 224, + 410, + 257 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c0c7fa20af61602243b7608b5d6ffe6bf4d1e5df5f80e8ff3cafc5cb5b4f4f3f.jpg", + "image_caption": [ + "GT", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 415, + 224, + 491, + 257 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fd7940648b1c40934a6c388da3a088df134a594435e62365d11e4c933bcb7b4e.jpg", + "image_caption": [ + "Image denoising" + ], + "image_footnote": [], + "bbox": [ + 504, + 93, + 656, + 169 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2e54fba6b683998791ccb45d7addac44694a3e6c6ebd6d06b0358e9a05479d39.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 93, + 736, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f0f9cd3a714accb1c416174a91a640a43a3184d482e0e71453c25f96b6041a0f.jpg", + "image_caption": [ + "PromptIR" + ], + "image_footnote": [], + "bbox": [ + 741, + 93, + 818, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/699f9c65003e99886c92bc71e0236fe76039d87e11c0a74e1dd74acbeb6d923b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 823, + 93, + 900, + 126 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3577901648ab33fd82bed731f716ca9e461ebf3ee5e8a1d2c8a9e998ccae6f3d.jpg", + "image_caption": [ + "Histformer" + ], + "image_footnote": [], + "bbox": [ + 661, + 137, + 736, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/84d0423fbc73d628b6fea574dfef7dfd36de2a734837364044078a42a10ecbe8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 741, + 137, + 818, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/df0c1d6c3ec1be1c93c778315ded93ff3c9ae8e0d43509c9eca8b59b32e1ff3e.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 821, + 137, + 900, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8a7b2b617fe2b52c4301a8a8e2b001bef24bf141f3307bf83387682928cf8de0.jpg", + "image_caption": [ + "Image dehazing" + ], + "image_footnote": [], + "bbox": [ + 504, + 181, + 658, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fac6ddff917d8339bf314872342a0d65790262bc3aae1d839f2c826da2662236.jpg", + "image_caption": [ + "Gridformer" + ], + "image_footnote": [], + "bbox": [ + 661, + 181, + 736, + 214 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c10c595537485a98cc87cbb5c7de71e9faf68bff6e8c678b26c201cd408d8d10.jpg", + "image_caption": [ + "HAIR" + ], + "image_footnote": [], + "bbox": [ + 741, + 181, + 818, + 214 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fa7002e0da7863c87e766d3c03e621cf9cccca59437bb48417a59b7786ce4a5e.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 823, + 181, + 900, + 214 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5177262119abe129d47c65b526ac3485076793234eb3c1d6e53ba4c75600d472.jpg", + "image_caption": [ + "Histformer", + "Gridformer" + ], + "image_footnote": [], + "bbox": [ + 661, + 224, + 738, + 256 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4c554c03c9650b99b6d0c47fd1dbd752e37b7bbc0cac734bbc79c460a0a21b16.jpg", + "image_caption": [ + "PromptIR", + "HAIR" + ], + "image_footnote": [], + "bbox": [ + 741, + 224, + 818, + 256 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/42ccda51ace6b43f5c5d22951878bc907a035daf5ffc7610b8884b0bbda677e2.jpg", + "image_caption": [ + "GT", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 823, + 224, + 900, + 256 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f005521bdf462bb1697f5180eb9e3436680b3628632694f700b1e941ba60109a.jpg", + "table_caption": [ + "Table 1. Comparison to state-of-the-art on four degradations. PSNR (dB, ↑), SSIM (↑), and LPIPS (↓), and FS represents full-size 4K image inference. FLOPs are computed for an input size of $256 \\times 256$ . Best and second best performances are highlighted." + ], + "table_footnote": [], + "table_body": "
MethodFSFLOPsParams.Low LightDeblurringDehazingDenoisingAverage
UHD-LLUHD-blurUHD-haze\\( UHDN_{\\sigma=15} \\)\\( UHDN_{\\sigma=25} \\)\\( UHDN_{\\sigma=50} \\)
AIRNet [14]✘301G9M19.24.80921.89.75718.37.81221.33.88720.78.78418.79.47520.07.754.2843
IDR [41]✘88G15.3M23.12.91024.67.79319.12.76827.48.91225.86.87224.57.65424.14.822.2684
PromptIR [20]✘158G33M23.44.90225.77.78219.97.72728.43.92426.74.89823.72.58424.68.803.2571
CAPTNet [8]✘25G24.3M23.96.92026.11.79819.46.86825.58.86523.24.88421.98.50823.39.809.3466
NDR-Restore [38]✘196G36.9M23.84.89424.25.80220.08.89225.62.91224.37.89722.94.66923.52.846.3126
Gridformer [33]✘367G34M23.12.89825.82.78319.24.86936.04.93731.72.89826.24.62327.03.836.3754
DiffUIR-L [44]✘10G36.2M21.56.81223.85.74318.28.86436.84.93832.42.89726.08.64826.51.818.2564
Histoformer [27]✘91G16.6M23.22.90825.62.78219.78.90326.88.84525.64.87423.13.65924.04.829.3524
adalR [4]✘147G28.7M23.57.91626.35.80118.44.90132.84.92130.48.90126.48.67226.36.857.3429
HAIR [1]✘41G29M25.75.92225.78.79820.00.89435.54.91630.84.87826.26.65727.36.847.2822
Ours✓4G1.6M27.11.92526.48.80320.94.92338.94.97533.99.90327.95.67729.23.868.2541
", + "bbox": [ + 133, + 335, + 866, + 484 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f16ccaa3fe15b3b5c35fd426480ad6e1489606b8a3c55107fc0d0485653eee99.jpg", + "table_caption": [ + "Table 2. Comparison to state-of-the-art on six degradations. PSNR (dB, ↑), SSIM (↑), LPIPS (↓) and FS represents full-size 4K image inference. FLOPs are computed for an input size of $256 \\times 256$ . Best and second best performances are highlighted." + ], + "table_footnote": [], + "table_body": "
MethodFSFLOPsParams.Low LightDeblurringDehazingDenoisingDerainingDesnowingAverage
UHD-LLUHD-blurUHD-hazeUHDN\\(_{\\sigma=50}\\)UHD-rainUHD-snow
AIRNet [14]✘301G9M22.68.88723.52.87618.24.84622.38.87626.35.87627.38.92423.43.874.1861
IDR [41]✘88G15.3M24.33.91525.64.78818.68.87929.64.90628.82.90630.48.94526.27.890.1912
PromptIR [20]✘158G33M23.3.91126.48.80520.14.90124.88.83528.89.89730.78.96625.74.886.2155
CAPTNet [8]✘25G24.3M24.97.92126.32.79620.32.90321.64.56929.34.90832.21.97425.80.845.2861
NDR-Restore [38]✘196G36.9M25.12.88525.64.79119.21.89631.44.91529.24.89728.41.94826.51.889.3108
Gridformer [33]✘367G34M23.92.89825.68.78218.87.88932.86.91529.37.90428.24.94226.49.895.2321
DiffUIR-L [44]✘10G36.2M22.64.90225.08.78518.62.88933.25.92827.89.88627.36.94525.81.889.1844
Histoformer [27]✘91G16.6M25.73.91526.55.79618.73.89733.05.92427.96.88427.56.97126.59.898.1855
adalR [4]✘147G28.7M23.84.91826.86.80319.34.91032.46.92328.18.90127.72.95326.40.901.2492
HAIR [1]✘41G29M25.22.89724.77.79918.75.88332.50.91528.76.89327.89.96826.31.892.2607
Ours✓4G1.6M26.91.92426.95.80721.81.93133.73.93429.90.91532.73.97928.67.915.1839
", + "bbox": [ + 133, + 534, + 866, + 683 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cb9a2adf570c11b2699299dbb89f0c33716f56e4aa4a4731950ab48d5527522b.jpg", + "table_caption": [ + "Table 3. Comparison to state-of-the-art for single degradations. PSNR (dB, ↑), SSIM (↑) and LPIPS (↓). Best and second best performances are highlighted.", + "(a) Low light", + "(b) Deblurring", + "(c) Dehazing" + ], + "table_footnote": [], + "table_body": "
MethodFSUHD-LLLIPIPSMethodFSUHD-blurLIPIPSMethodFSUHD-hazeLIPIPS
Restormer [40]✘21.54.843.3608Restormer [40]✘25.21.752.3695Restormer [40]✘12.72.693.4560
LLformer [32]✘24.06.858.3516Uformer [35]✘25.26.751.3851Uformer [35]✘19.83.921.4220
UHDFour [15]✓26.22.900.2390Stripformer [28]✘25.05.750.3740DehazeFormer [26]✘15.37.737.3998
UHDFformer [29]✓27.11.927.2240FFTformer [12]✘25.41.757.3708TaylorFormer [22]✘20.99.919.3124
LMAR [39]✓26.27.919.2248UHDformer [29]✓28.82.844.2350UHDformer [29]✓22.58.942.1188
AirNet [14]✘24.61.827.3841AirNet✘23.48.824.3942AirNet✘18.44.848.4018
PromptIR [20]✘24.81.912.3116PromptIR✘23.24.829.3754PromptIR✘20.21.927.3642
Histoformer [27]✘25.46.894.3213Histoformer✘27.12.848.2986Histoformer✘21.72.948.2942
HAIR [1]✘26.08.901.2964HAIR✘26.74.847.3867HAIR✘20.38.912.2412
Ours✓27.22.929.2042Ours✓28.91.851.2129Ours✓23.24.953.1086
", + "bbox": [ + 150, + 736, + 846, + 878 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23127", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 89, + 90, + 459, + 107 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2.1. All-in-One: Four Degradations", + "text_level": 1, + "bbox": [ + 89, + 112, + 354, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare our UHD-Processor with specialized all-in-one restoration methods, trained on four degradation types: dehazing, denoising, deblurring, and low-light enhancement. As shown in Tab. 1, our method achieves state-of-the-art (SOTA) performance on both individual datasets and overall metrics. Specifically, the average PSNR improves by 1.93 dB over the second-best method, HAIR, while our method requires only $5.5\\%$ of the parameters and $9.7\\%$ of the FLOPS of Histoformer. Visual results, shown in Figure 3, further demonstrate that our approach effectively achieves the best degradation removal.", + "bbox": [ + 88, + 131, + 485, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2.2. All-in-One: Six Degradations", + "text_level": 1, + "bbox": [ + 89, + 305, + 341, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further extended the experiment to include six degradation types by introducing image desnow and image derain, to validate the effectiveness of our method in a more complex all-in-one setup. As shown in Tab. 2, our method continues to achieve SOTA performance on both individual datasets and average metrics. Specifically, the average PSNR improves by $1.7\\mathrm{dB}$ over the second-best method, Histoformer, while the number of parameters and FLOPS are only $9.6\\%$ and $4.3\\%$ respectively. More visual comparison results can be found in the supplementary materials.", + "bbox": [ + 88, + 324, + 485, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2.3. Single-Degradation Results", + "text_level": 1, + "bbox": [ + 89, + 483, + 326, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further evaluate the performance of our method by training instances under a single degradation protocol, comparing them with both task-specific methods and all-in-one approaches. The results, shown in Tab. 3, demonstrate that our method still achieves the best performance, validating the effectiveness of our approach.", + "bbox": [ + 88, + 502, + 483, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 603, + 243, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.1. Frequency-Progressive Learning Strategy", + "text_level": 1, + "bbox": [ + 89, + 625, + 426, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Tab. 4, we validate the effectiveness of the proposed Frequency Domain Progressive Learning (FDPL). First, compared to directly training on mixed data, using FDPL results in a significant improvement of $1.99\\mathrm{dB}$ PSNR. Next, we evaluate the sequential learning (SL) strategy proposed in [13], which advocates for learning local degradations before global ones. We test two learning orders: SL1 (noise $\\rightarrow$ blur $\\rightarrow$ low-light $\\rightarrow$ haze) and SL2 (blur $\\rightarrow$ noise $\\rightarrow$ haze $\\rightarrow$ low-light). The experimental results show that not all local-to-global learning orders lead to performance gains. In fact, SL1 even caused some performance degradation, while SL2 yielded better results. This suggests that sequential learning strategies are sensitive to the order of tasks, and the optimal sequence may not only depend on the tasks themselves but also on the specific data set, scene, or degradation severity. Therefore, it is challenging to define a universally effective task order. In contrast, our", + "bbox": [ + 89, + 643, + 485, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/09fe57beb67b5dcbdb70d5a7abfb0cad12b6c48de5f750dc46af525b68c79cd3.jpg", + "table_caption": [ + "Table 4. Ablation Study of the Learning Strategy." + ], + "table_footnote": [], + "table_body": "
MethodMixw/SL1w/SL2w/FDPL (Ours)
PSNR27.2427.0828.1229.23
SSIM0.8530.8460.8720.874
", + "bbox": [ + 576, + 104, + 841, + 150 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d934e023cd5d9bbd1619d83c30fb43ec15460961ee574e5ecb1cf38ae2a6a4b0.jpg", + "table_caption": [ + "Table 5. Ablation Study of Efficient Adaptive Prompt Learning." + ], + "table_footnote": [], + "table_body": "
Methodw/o DALPw/o DDPw/o DTPw/o DSRw/o MCKGw/o DFSPOurs
PSNR27.8628.4928.3428.8428.7228.2129.23
SSIM0.8620.8660.8580.8640.8620.8650.874
", + "bbox": [ + 514, + 183, + 908, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "FDPL strategy is task-agnostic and offers a more generalizable learning framework. Experimental results confirm that FDPL leads to more significant performance improvements. Further discussions on sequential learning can be found in the supplementary materials.", + "bbox": [ + 511, + 260, + 906, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.2. Efficient Adaptive Prompt Learning", + "text_level": 1, + "bbox": [ + 511, + 352, + 810, + 368 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The ablation study of the proposed prompt learning method, as shown in Tab. 5, demonstrates that both the DALP used in the encoding process and the DFSP used in the decoding process lead to significant performance improvements. We further analyze the components of DALP, and the results show that both the Degradation Degree Prompt (DDP) and the Degradation Type Prompt (DTP) provide more granular prompt information, improving performance consistently. The Degradation-Specific Rank (DSR) component introduces task-specific low-rank decompositions, allowing for the selection of an appropriate rank for each task, which enhances parameter efficiency while improving performance. Moreover, the Multi-Scale Convolution Kernel Group (MCKG) introduces a multi-scale convolutional branch, which strengthens the representation capacity of the prompt generation process. During inference, the multi-branch convolutions are reparameterized for efficient execution. For more detailed ablation experiments and efficiency analyses, please refer to the supplementary materials.", + "bbox": [ + 511, + 373, + 908, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 676, + 633, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we proposed UHD-Processor, an efficient and effective framework for UHD all-in-one image restoration. By leveraging Frequency Domain Decoupling Progressive Learning Strategy and an efficient prompt-integrated VAE framework, our method addresses the unique challenges of UHD restoration tasks, including denoising, deblurring, deraining, dehazing, and low-light enhancement. Experimental results demonstrated that our approach significantly outperforms existing methods, achieving state-of-the-art results with reduced computational costs. The combination of progressive frequency learning and targeted prompt-based adaptation provides a strong foundation for future advancements in UHD image restoration.", + "bbox": [ + 511, + 703, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23128", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China (NSFC) under Grants 62225207, 62436008, 62422609 and 62276243.", + "bbox": [ + 89, + 114, + 485, + 160 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 172, + 187, + 188 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jin Cao, Yi Cao, Li Pang, Deyu Meng, and Xiangyong Cao. Hair: Hypernetworks-based all-in-one image restoration, 2024. 7", + "[2] Hongming Chen, Xiang Chen, Chen Wu, Zhuoran Zheng, Jinshan Pan, and Xianping Fu. Towards ultra-high-definition image deraining: A benchmark and an efficient method, 2024. 1, 6", + "[3] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1", + "[4] Yuning Cui, Syed Waqas Zamir, Salman Khan, Alois Knoll, Mubarak Shah, and Fahad Shahbaz Khan. Adair: Adaptive all-in-one image restoration via frequency mining and modulation, 2024. 6, 7", + "[5] Senyou Deng, Wenqi Ren, Yanyang Yan, Tao Wang, Fenglong Song, and Xiaochun Cao. Multi-scale separable network for ultra-high-definition video deblurring. In the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14030-14039, 2021. 6", + "[6] Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13733-13742, 2021. 6", + "[7] Xiaohan Ding, Yiyuan Zhang, Yixiao Ge, Sijie Zhao, Lin Song, Xiangyu Yue, and Ying Shan. Unireplknet: A universal perception large-kernel convnet for audio video point cloud time-series and image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5513-5524, 2024. 6", + "[8] Hu Gao, Jing Yang, Ning Wang, Jingfan Yang, Ying Zhang, and Depeng Dang. Prompt-based all-in-one image restoration using cnns and transformer. arXiv preprint arXiv:2309.03063, 2023. 3, 7", + "[9] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 5", + "[10] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 2", + "[11] Xingyu Jiang, Xiuhui Zhang, Ning Gao, and Yue Deng. When fast fourier transform meets transformer for image restoration. 6", + "[12] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In 2023 IEEE/CVF" + ], + "bbox": [ + 93, + 198, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition (CVPR), pages 5886-5895, 2023. 7", + "[13] Xiangtao Kong, Chao Dong, and Lei Zhang. Towards effective multiple-in-one image restoration: A sequential and prompt learning strategy. arXiv preprint arXiv:2401.03379, 2024. 3, 8", + "[14] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-In-One Image Restoration for Unknown Corruption. In IEEE Conference on Computer Vision and Pattern Recognition, New Orleans, LA, 2022. 2, 7", + "[15] Chongyi Li, Chun-Le Guo, Man Zhou, Zhexin Liang, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Embedding fourier for ultra-high-definition low-light image enhancement. In ICLR, 2023. 1, 2, 6, 7", + "[16] Xiangtai Li, Haobo Yuan, Wei Li, Henghui Ding, Size Wu, Wenwei Zhang, Yining Li, Kai Chen, and Chen Change Loy. Omg-seg: Is one model good enough for all segmentation? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27948-27959, 2024. 1", + "[17] Zilong Li, Yiming Lei, Chenglong Ma, Junping Zhang, and Hongming Shan. Prompt-in-prompt learning for universal image restoration, 2023. 3", + "[18] Ziwei Luo, Fredrik K. Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B. Schön. Controlling vision-language models for multi-task image restoration. In The Twelfth International Conference on Learning Representations, 2024. 3", + "[19] Anastasia Pentina, Viktoriia Sharmanska, and Christoph H Lampert. Curriculum learning of multiple tasks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5492-5500, 2015. 3", + "[20] Vaishnav Potlapalli, Syed Waqas Zamir, Salman Khan, and Fahad Khan. Prompt: Prompting for all-in-one image restoration. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2, 5, 7", + "[21] Yifan Pu, Weicong Liang, Yiduo Hao, Yuhui Yuan, Yukang Yang, Chao Zhang, Han Hu, and Gao Huang. Rank-detr for high quality object detection. Advances in Neural Information Processing Systems, 36, 2024. 1", + "[22] Yuwei Qiu, Kaihao Zhang, Chenxi Wang, Wenhan Luo, Hongdong Li, and Zhi Jin. Mb-taylorformer: Multi-branch efficient transformer expanded by taylor formula for image dehazing, 2023. 7", + "[23] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International conference on machine learning, pages 5301-5310. PMLR, 2019. 3", + "[24] Wenqi Ren, Senyou Deng, Kaihao Zhang, Fenglong Song, Xiaochun Cao, and Ming-Hsuan Yang. Fast ultra high-definition video deblurring via multi-scale separable network. International Journal of Computer Vision, 132(5):1817-1834, 2024. 2", + "[25] Basri Ronen, David Jacobs, Yoni Kasten, and Shira Kritchman. The convergence rate of neural networks for learned functions of different frequencies. Advances in Neural Information Processing Systems, 32, 2019. 3" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23129", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Yuda Song, Zhuqing He, Hui Qian, and Xin Du. Vision transformers for single image dehazing. IEEE Transactions on Image Processing, 32:1927-1941, 2023. 7", + "[27] Shangquan Sun, Wenqi Ren, Xinwei Gao, Rui Wang, and Xiaochun Cao. Restoring images in adverse weather conditions via histogram transformer. In European Conference on Computer Vision, pages 111-129. Springer, 2025. 7", + "[28] Fu-Jen Tsai, Yan-Tsung Peng, Yen-Yu Lin, Chung-Chi Tsai, and Chia-Wen Lin. Stripformer: Strip transformer for fast image deblurring. In European conference on computer vision, pages 146-162. Springer, 2022. 7", + "[29] Cong Wang, Jinshan Pan, Wei Wang, Gang Fu, Siyuan Liang, Mengzhu Wang, Xiao-Ming Wu, and Jun Liu. Correlation matching transformation transformers for uhd image restoration. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5336-5344, 2024. 1, 2, 7", + "[30] Haohan Wang, Xindi Wu, Zeyi Huang, and Eric P Xing. High-frequency component helps explain the generalization of convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8684-8694, 2020. 3", + "[31] Liyan Wang, Cong Wang, Jinshan Pan, Xiaofeng Liu, Weixiang Zhou, Xiaoran Sun, Wei Wang, and Zhixun Su. Ultra-high-definition image restoration: New benchmarks and a dual interaction prior-driven solution, 2024. 2, 6", + "[32] Tao Wang, Kaihao Zhang, Tianrun Shen, Wenhan Luo, Bjorn Stenger, and Tong Lu. Ultra-high-definition low-light image enhancement: A benchmark and transformer-based method. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2654-2662, 2023. 7", + "[33] Tao Wang, Kaihao Zhang, Ziqian Shao, Wenhan Luo, Bjorn Stenger, Tong Lu, Tae-Kyun Kim, Wei Liu, and Hongdong Li. Gridformer: Residual dense transformer with grid structure for image restoration in adverse weather conditions. International Journal of Computer Vision, pages 1-23, 2024. 2, 7", + "[34] Xin Wang, Yudong Chen, and Wenwu Zhu. A survey on curriculum learning. IEEE transactions on pattern analysis and machine intelligence, 44(9):4555-4576, 2021. 2, 3", + "[35] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17683-17693, 2022. 7", + "[36] Boxue Xiao, Zhuoran Zheng, Xiang Chen, Chen Lv, Yunliang Zhuang, and Tao Wang. Single uhd image dehazing via interpretable pyramid network, 2022. 1", + "[37] Zhi-Qin John Xu, Yaoyu Zhang, and Yanyang Xiao. Training behavior of deep neural network in frequency domain. In Neural Information Processing: 26th International Conference, ICONIP 2019, Sydney, NSW, Australia, December 12–15, 2019, Proceedings, Part I 26, pages 264–274. Springer, 2019. 3", + "[38] Mingde Yao, Ruikang Xu, Yuanshen Guan, Jie Huang, and Zhiwei Xiong. Neural degradation representation learning for all-in-one image restoration. IEEE Transactions on Image Processing, 2024. 2, 7" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Wei Yu, Jie Huang, Bing Li, Kaiwen Zheng, Qi Zhu, Man Zhou, and Feng Zhao. Empowering resampling operation for ultra-high-definition image enhancement with model-aware guidance. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25722-25731, 2024. 7", + "[40] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 2, 7", + "[41] Jinghao Zhang, Jie Huang, Mingde Yao, Zizheng Yang, Hu Yu, Man Zhou, and Feng Zhao. Ingredient-oriented multidegradation learning for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5825-5835, 2023. 2, 3, 7", + "[42] Kaihao Zhang, Dongxu Li, Wenhan Luo, Wenqi Ren, Bjorn Stenger, Wei Liu, Hongdong Li, and Yang Ming-Hsuan. Benchmarking ultra-high-definition image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 6", + "[43] Wenlong Zhang, Xiaohui Li, SHI Guangyuan, Xiangyu Chen, Yu Qiao, Xiaoyun Zhang, Xiao-Ming Wu, and Chao Dong. Real-world image super-resolution as multi-task learning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 3", + "[44] Dian Zheng, Xiao-Ming Wu, Shuzhou Yang, Jian Zhang, Jian-Fang Hu, and Wei-Shi Zheng. Selective hourglass mapping for universal image restoration based on diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25445-25455, 2024. 3, 7", + "[45] Zhuoran Zheng and Xiuyi Jia. Uhd image deblurring via multi-scale cubic-mixer. arXiv preprint arXiv:2206.03678, 2022. 2, 3", + "[46] Zhuoran Zheng, Wenqi Ren, Xiaochun Cao, Xiaobin Hu, Tao Wang, Fenglong Song, and Xiuyi Jia. Ultra-high-definition image dehazing via multi-guided bilateral learning. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16180-16189, 2021. 6", + "[47] Yanqi Zhou, Tao Lei, Hanxiao Liu, Nan Du, Yanping Huang, Vincent Zhao, Andrew M Dai, Quoc V Le, James Laudon, et al. Mixture-of-experts with expert choice routing. Advances in Neural Information Processing Systems, 35:7103-7114, 2022. 2" + ], + "bbox": [ + 516, + 90, + 906, + 710 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23130", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/UHD-processer_ Unified UHD Image Restoration with Progressive Frequency Learning and Degradation-aware Prompts/2e2768b6-2887-4bd5-935c-d0b880ebe319_model.json b/2025/UHD-processer_ Unified UHD Image Restoration with Progressive Frequency Learning and Degradation-aware Prompts/2e2768b6-2887-4bd5-935c-d0b880ebe319_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7415865db715e9c2493ec28bf4a9669bf62b7b12 --- /dev/null +++ b/2025/UHD-processer_ Unified UHD Image Restoration with Progressive Frequency Learning and Degradation-aware Prompts/2e2768b6-2887-4bd5-935c-d0b880ebe319_model.json @@ -0,0 +1,2992 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.167, + 0.131, + 0.834, + 0.178 + ], + "angle": 0, + "content": "UHD-processor: Unified UHD Image Restoration with Progressive Frequency Learning and Degradation-aware Prompts" + }, + { + "type": "text", + "bbox": [ + 0.198, + 0.204, + 0.802, + 0.278 + ], + "angle": 0, + "content": "Yidi Liu, DongLi, Xueyang Fu†, Xin Lu, Jie Huang, Zheng-jun Zha \nSchool of Information Science and Technology and MoE Key Laboratory of \nBrain-inspired Intelligent Perception and Cognition, \nUniversity of Science and Technology of China, Hefei, 230026, China" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.279, + 0.782, + 0.295 + ], + "angle": 0, + "content": "liuyidi2023@mail.ustc.edu.cn, xyfu@ustc.edu.cn † Corresponding Author" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.329, + 0.328, + 0.345 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.362, + 0.486, + 0.754 + ], + "angle": 0, + "content": "We introduce UHD-Processor, a unified and robust framework for all-in-one image restoration, which is particularly resource-efficient for Ultra-High-Definition (UHD) images. To address the limitations of traditional all-in-one methods that rely on complex restoration backbones, our strategy employs a frequency domain decoupling progressive learning technique, motivated by curriculum learning, to incrementally learn restoration mappings from low to high frequencies. This approach incorporates specialized sub-network modules to effectively tackle different frequency bands in a divide-and-conquer manner, significantly enhancing the learning capability of simpler networks. Moreover, to accommodate the high-resolution characteristics of UHD images, we developed a variational autoencoder (VAE)-based framework that reduces computational complexity by modeling a concise latent space. It integrates task-specific degradation awareness in the encoder and frequency selection in the decoder, enhancing task comprehension and generalization. Our unified model is able to handle various degradations such as denoising, deblurring, dehazing, lowlighting, etc. Experimental evaluations extensively showcase the effectiveness of our dual-strategy approach, significantly improving UHD image restoration and achieving cutting-edge performance across diverse conditions. The code will be available at https://github.com/1yd2022/UHD-processor" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.784, + 0.224, + 0.8 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Ultra-high-definition (UHD) images are increasingly utilized across various fields, prompting substantial research in UHD image restoration for specific degradation types, such as low-light enhancement, deblurring, dehazing, and deraining [2, 15, 29, 36]. However, real-world applications like object detection [21], video segmentation [16], and au" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.336, + 0.753, + 0.498 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.501, + 0.757, + 0.509 + ], + "angle": 0, + "content": "(a) Comparison of PSNR with SOTA All-in-One Methods" + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.34, + 0.897, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.793, + 0.419, + 0.858, + 0.427 + ], + "angle": 0, + "content": "(b) Observation 1" + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.429, + 0.896, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.793, + 0.501, + 0.859, + 0.509 + ], + "angle": 0, + "content": "(c) Observation 2" + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.511, + 0.768, + 0.609 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.566, + 0.61, + 0.876, + 0.62 + ], + "angle": 0, + "content": "(d) Observation 3: The Variation of Gap Between Degradations Across Frequency Bands" + }, + { + "type": "image", + "bbox": [ + 0.769, + 0.512, + 0.9, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.622, + 0.908, + 0.734 + ], + "angle": 0, + "content": "Figure 1. (a) Comparison of PSNR between our method and the previous state-of-the-art (SOTA) all-in-one method across six degradation settings in UHD images. (b) Observe the performance gains when integrating the same promptIR prompt branch into Restormer models with varying parameter sizes. (c) Comparison of low-to-high frequency training order on UHD and regular images, showing greater effectiveness for UHD tasks. (d) Quantitative evaluation of restoration tasks across low, high, and full-frequency bands." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.908, + 0.824 + ], + "angle": 0, + "content": "tonomous driving [3] face complex degradation patterns that vary across time and space, making reliance on single-function models impractical. This calls for a generalized model that can address multiple degradation types to support robust low-level vision tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In response to the above requirements, all-in-one(AIO) approaches aim to use a single model to manage various degradations, but they generally target standard-resolution images and encounter computational difficulties when scaled to UHD. This gap highlights the pressing need for an efficient" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "23121" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "all-in-one image restoration framework tailored to UHD images. For standard-resolution tasks [14, 20], successful approaches often include two key features, i.e., degradation-aware branches and strong network backbones. The former is implemented through techniques like prompts [10] or mixtures of experts (MoE) [47] to enhance adaptability to different degradation types, while the later contains advanced backbones like Restormer [40] to further strengthens learning and generalization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.23, + 0.482, + 0.381 + ], + "angle": 0, + "content": "However, due to the high-resolution nature of UHD images, directly applying these strategies results in a significant computational burden, making full-resolution inference infeasible on consumer-grade GPUs. As shown in Figure 1(b), PromptIR extends the Restormer backbone with a prompt branch. We adjust their parameter counts by varying Restormer's module numbers, keeping the prompt branch fixed. Reducing parameters significantly lowers performance for both, meanwhile, the benefits introduced by the prompt branch also gradually diminish." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.382, + 0.482, + 0.473 + ], + "angle": 0, + "content": "These results indicate that while the prompt branch enhances all-in-one task performance, a strong backbone is crucial to fully leverage these improvements. Moreover, compared to Restormer, the additional degradation-aware prompt branch introduced by PromptIR adds 6.87M parameters, which is clearly burdensome for UHD tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.475, + 0.482, + 0.536 + ], + "angle": 0, + "content": "Based on the above observations, this work aims to tackle two challenges faced by traditional AIO methods in UHD scenarios: reliance on complex restoration backbones and redundancy in degradation-aware branch parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.483, + 0.764 + ], + "angle": 0, + "content": "To tackle the first challenge, at the model architecture level, we propose a VAE-based framework that transforms restoration tasks from pixel space into a compact latent space, thereby reducing the computational cost of UHD image processing. This alleviates the network's learning burden, enabling a simpler model to restore effectively. At the learning strategy level, inspired by curriculum learning [34], we develop a Frequency Domain Decoupling Progressive Learning Strategy. As shown in Figure 1(d), for UHD AIO tasks, higher frequency bands present greater restoration challenges. Consequently, we decompose the task into manageable frequency components, enabling sub-networks to progressively learn restoration mappings from low to high frequencies. This divide-and-conquer approach significantly enhances the network's adaptability and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To address the second challenge, we introduce efficient prompt strategies for both the VAE encoder and decoder. For encoding, we propose a low-rank degradation-aware prompt to indicate degradation type and severity. This is achieved by creating distinct low-rank parameter spaces per degradation type and learning scale-invariant convolution kernels, which adapt to UHD resolution differences between training and testing. For efficiency, we reparameterize these convolution branches during inference. To maintain efficiency," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "we reparameterize these convolution branches at inference. In decoding, we observe that degradation types exhibit distinct distributions across frequency bands. Accordingly, we introduce a frequency band selection prompt, allowing the decoder to adaptively choose between low-frequency features from the decoder and mid- or high-frequency features from the encoder based on the specific degradation type." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.199, + 0.905, + 0.303 + ], + "angle": 0, + "content": "By integrating these components, we propose UHD-Processor, an efficient all-in-one image restoration framework tailored for ultra-high-definition (UHD) images. As shown in Figure 1(a), our proposed UHD-Processor achieves state-of-the-art performance while reducing parameters by \\(95.1\\%\\) and FLOPs by \\(97.4\\%\\) compared to promptIR [20]. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.307, + 0.904, + 0.351 + ], + "angle": 0, + "content": "1. We introduce a VAE-based image restoration framework that operates within a compact latent space, significantly enhancing UHD image restoration efficiency." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.353, + 0.905, + 0.427 + ], + "angle": 0, + "content": "2. We design a frequency domain decoupling progressive learning strategy, which enables effective restoration across frequency bands from low to high frequencies through a divide-and-conquer approach, to improve adaptability to various optimization objectives." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.428, + 0.904, + 0.487 + ], + "angle": 0, + "content": "3. We develop a prompt learning framework with low-rank degradation-aware prompts for encoding and frequency selection prompts for decoding, improving the network's adaptability to various degradation types." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.307, + 0.905, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.503, + 0.655, + 0.518 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.528, + 0.811, + 0.544 + ], + "angle": 0, + "content": "2.1. Ultra-High-Definition Restoration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.905, + 0.779 + ], + "angle": 0, + "content": "As the demand for UHD image processing grows in imaging systems, various approaches have been proposed to restore clarity in UHD images through diverse network architectures. These include bilateral learning for image dehazing [45], multi-scale separable-patch integration networks for video deblurring [24], and Fourier embedding networks for low-light image enhancement [15]. Recently, generalized architectures for UHD image restoration have evolved to address multiple restoration tasks through unified frameworks. For instance, [29] introduces a correlation matching mechanism between high-resolution and low-resolution branches to enhance the utilization of native resolution information, while [31] incorporates gradient and normal priors into the model design. However the approach about UHD all in one is not yet fully explored." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.788, + 0.779, + 0.804 + ], + "angle": 0, + "content": "2.2. All-in-One Image Restoration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.901 + ], + "angle": 0, + "content": "All-in-one image restoration, which aims to recover clean images from various types of degradation using a single unified model, has gained significant attention for its efficiency and practical applications in low-level vision tasks [14, 20, 33, 38, 41]. Early methods, such as AirNet [14], utilized contrastive learning to extract degra" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23122" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.226 + ], + "angle": 0, + "content": "dation representations, while IDR [41]introduced a two-stage meta-learning approach to address diverse degradation types. More recently, prompt-based approaches like PromptIR [8]and PIP [17]further enhanced restoration flexibility through vision prompts. Additionally, large-scale pre-trained models, as demonstrated by DiffUIR [44] and DA-CLIP [18], have been adapted for all-in-one restoration with strong performance. Please refer to the supplementary material for more detailed related work." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.243, + 0.183, + 0.259 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.269, + 0.201, + 0.284 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.292, + 0.485, + 0.548 + ], + "angle": 0, + "content": "The framework of our method is shown in Figure 2. Overall, our approach follows a two-stage paradigm. In the first stage, we train a VAE on clean images with an image reconstruction task, as shown in Figure 2(a). In the second stage, as shown in Figure 2(b), we freeze the VAE weights and train the remaining components for image restoration tasks. To mitigate high-frequency loss during the VAE encoding process, we pass high-frequency components from the wavelet transformation between the encoder adapter and the decoder adapter. Additionally, prompts are injected through the adapters: in the encoding phase, we apply degradation-aware low-rank prompts (Figure 2(d)) to adapt the VAE to different degradations, encoding them into a unified compact latent space; in the decoding phase, we use frequency selection prompts (Figure 2(c)) for each type of degradation, enabling efficient adaptive fusion of low-frequency decoder information with high-frequency encoder information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.55, + 0.484, + 0.746 + ], + "angle": 0, + "content": "The latent restoration network is divided into \\(M\\) subnetworks, each corresponding to one of the \\(M\\) decomposed frequency bands of the input. During Frequency Domain Decoupling Progressive Learning (Figure 2(e)), we progressively increase from the lowest to the highest frequency bands, incrementally adding sub-networks until all frequency bands are restored. Attributing the strong performance to the robust latent space encoding and frequency-progressive learning strategy, we achieve excellent results with a simple latent space restoration network. In this work, we adopt the Cubic-Mixer [45] for its efficiency, results of other architectures and the interaction detail between the adapter and VAE, please refer to the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.758, + 0.483, + 0.79 + ], + "angle": 0, + "content": "3.2. Frequency Domain Decoupling Progressive Learning Strategy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.902 + ], + "angle": 0, + "content": "For all-in-one tasks, which aim to handle multiple types of degradation with a single model, we can represent this as \\( F(\\{\\tau^t\\}; \\theta)_{t \\in [1, N]} \\), where \\( N \\) represents the number of tasks, \\( \\tau^t \\) denotes the input for task \\( t \\), and \\( \\theta \\) represents the model parameters. Each task \\( \\tau^t \\) corresponds to a distinct optimization objective; for instance, the low-light enhancement task focuses on adjusting the global brightness of the image, while" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.168 + ], + "angle": 0, + "content": "the deraining task aims to remove local rain streaks. Consequently, it is challenging for a network trained from scratch to identify a suitable gradient direction for all tasks during initial training, which often results in slow convergence or convergence to suboptimal solutions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.169, + 0.909, + 0.349 + ], + "angle": 0, + "content": "Inspired by the concept of curriculum learning [34][19], a learning process that progresses from easy to difficult tasks is more beneficial for optimizing complex tasks. MIO-IR [13] employs a sequential learning strategy to incrementally add tasks, thereby mitigating the interference caused by joint training. However, while extending from single-task to multitask learning is a form of curriculum learning, the relationships among tasks remain ambiguous, making it challenging to establish a generalized strategy for determining the optimal task sequence. Therefore, we revisit how to design a more general curriculum learning strategy for UHD all-in-one tasks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.362, + 0.646, + 0.375 + ], + "angle": 0, + "content": "3.2.1. Observation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.382, + 0.908, + 0.472 + ], + "angle": 0, + "content": "First, we assess which processes are easier or harder for a neural network to learn during training, starting from random initialization to being capable of handling multiple degradations effectively. Inspired by the frequency principle of neural networks [23][25][37][30], we approach this problem from the perspective of the frequency domain." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.474, + 0.909, + 0.642 + ], + "angle": 0, + "content": "For a restoration task \\(\\tau^t\\), we decompose it into \\(M\\) frequency bands, with the \\(f\\)-th frequency band denoted as \\(\\tau_f^t\\). The multi-degradation restoration process is represented as \\(F(\\tau_f^t;\\theta)_{t\\in [1,N],f\\in [1,M]}\\). In sequential learning, tasks are learned progressively from 1 to \\(N\\), where each stage involves restoring all frequency bands, represented as \\(F(\\{\\tau_f^t\\}_{f = 1}^M;\\theta)_{t = 1\\to N}\\). The main drawback of this approach is the lack of strong inter-stage connections, and the relationships between all degradation types are only modeled in the final stage, which hinders the network's ability to comprehensively learn the interplay between different degradations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.643, + 0.909, + 0.749 + ], + "angle": 0, + "content": "To address this, we shift from task-based to frequency band-based progressive learning, represented as \\( F(\\{\\tau_f^t\\}_{t=1}^N; \\theta)_{f=1 \\to M} \\). Each frequency band captures global image information and maintains a physical relationship with other bands. Learning by frequency bands is smoother and more natural, as knowledge from earlier bands supports learning subsequent bands." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.911, + 0.904 + ], + "angle": 0, + "content": "To analyze the differences between degradations under low- and high-frequency conditions, we conducted a toy experiment. We selected four types of degradations: low-light and haze, which predominantly affect the low-frequency bands of an image, and denoising and deblurring, which mainly impact the high-frequency bands. We applied a low-pass filter to decompose images into low- and high-frequency components, then trained separate models for each. To quantify the gap between different tasks, we introduced the metric \\(\\hat{z}_{ij}^{s}\\) from [43], which is used to measure the discrepancy" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23123" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.339, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.344, + 0.09, + 0.902, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.48, + 0.908, + 0.552 + ], + "angle": 0, + "content": "Figure 2. The overall framework of our proposed UHD-Processor, which includes: (a) Pre-training a VAE on clean images for an image reconstruction task. (b) During the image restoration task, we keep the VAE frozen and efficiently fine-tune it using an adapter, where latentIR represents the latent space image restoration network. (c) Degradation-Specific Frequency Selection Prompt (DFSP) for the decoder. (d) Degradation-Aware Low Rank Prompt (DALR) for the encoder, and (e) an illustration of the Frequency Domain Decoupling Progressive Learning Strategy." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.577, + 0.483, + 0.623 + ], + "angle": 0, + "content": "between pre-training and fine-tuning tasks (higher values indicate smaller differences). For a detailed discussion of this metric, please refer to the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.626, + 0.483, + 0.776 + ], + "angle": 0, + "content": "Experimental results in Figure 1(d) indicate that, for both low and high frequencies, the value of \\(\\hat{z}_{ij}^{s}\\) is consistently higher compared to the entire frequency spectrum. This underscores the critical role of frequency-wise decomposition in minimizing task-specific discrepancies within individual frequency bands. Moreover, \\(\\hat{z}_{ij}^{s}\\) is significantly higher in the low-frequency band than in the high-frequency band, suggesting that a progressive learning sequence from low to high frequencies is better suited for establishing an effective curriculum learning framework." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.486, + 0.902 + ], + "angle": 0, + "content": "To further investigate the differences in conclusions between standard and UHD resolutions, we conducted experiments on UHD data for four types of degradations and their corresponding downsampled versions using an all-in-one training setup. As shown in Figure 1(c), By adopting a training sequence progressing from low to high frequencies, we observed improvements for both resolutions compared to direct training. However, the enhancement was more pro" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.577, + 0.907, + 0.622 + ], + "angle": 0, + "content": "nounced in UHD scenarios, which can be attributed to the richer high-frequency texture information present in UHD images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.635, + 0.907, + 0.666 + ], + "angle": 0, + "content": "3.2.2. Frequency Domain Decoupling Progressive Learning Strategy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.909, + 0.751 + ], + "angle": 0, + "content": "Therefore, we propose the Frequency Decoupling Progressive Learning strategy, which leverages multiple wavelet transformation (MWT) to decompose the input image \\(X\\) and ground truth \\(\\Upsilon\\) into multiple frequency sub-bands \\(\\{X_f\\}_{f=1}^M\\) and \\(\\{Y_f\\}_{f=1}^M\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.751, + 0.911, + 0.902 + ], + "angle": 0, + "content": "During training, the learning process progressively advances from low to high frequencies, with each frequency band \\( X_{f} \\) managed by a dedicated restoration sub-network \\( F_{f} \\). As training progresses and higher frequency bands are introduced, additional restoration sub-networks are incrementally incorporated, forming a divide-and-conquer approach to effectively handle each frequency band. The process can be formulated as Algorithm 1. This pseudocode omits the detailed steps of multiple wavelet transformation (MWT) and its corresponding inverse process (MIWT). For" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23124" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "a more comprehensive description of the algorithm, please refer to the supplementary materials." + }, + { + "type": "code_caption", + "bbox": [ + 0.093, + 0.134, + 0.483, + 0.166 + ], + "angle": 0, + "content": "Algorithm 1 Frequency Domain Decoupling Progressive Learning" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.17, + 0.468, + 0.188 + ], + "angle": 0, + "content": "1: Input: \\(\\{X_f\\}_{f=1}^M, \\{Y_f\\}_{f=1}^M \\gets \\mathrm{MWT}(X), \\mathrm{MWT}(Y)\\)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.189, + 0.482, + 0.221 + ], + "angle": 0, + "content": "2: Init: Initialize \\(\\{\\theta_f\\}_{f=1}^M\\) for sub-networks, set stage thresholds \\(\\{T_k\\}_{k=1}^M\\), and define total_iter" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.222, + 0.297, + 0.235 + ], + "angle": 0, + "content": "3: current_iter \\(\\leftarrow 0, k \\leftarrow 1\\)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.236, + 0.367, + 0.249 + ], + "angle": 0, + "content": "4: while current_iter \\(\\leq\\) total_iter do" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.251, + 0.38, + 0.267 + ], + "angle": 0, + "content": "5: \\(\\{\\hat{X}_f\\mid f = 1,\\dots ,k\\} \\gets F_f(X_f;\\theta_f)\\)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.268, + 0.383, + 0.284 + ], + "angle": 0, + "content": "6: \\(\\dot{X}\\gets \\mathrm{MIWT}(\\{\\dot{X}_f\\}_{f = 1}^k,\\{Y_f\\}_{f = k + 1}^M)\\)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.286, + 0.254, + 0.3 + ], + "angle": 0, + "content": "7: \\(\\mathcal{L}\\gets \\operatorname {Loss}(\\hat{X},Y)\\)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.301, + 0.43, + 0.319 + ], + "angle": 0, + "content": "8: \\(\\{\\theta_f\\}_{f = 1}^k\\gets \\{\\theta_f\\}_{f = 1}^k -\\alpha \\nabla_{\\{\\theta_f\\}_{f = 1}^k}\\mathcal{L}(\\tilde{X},Y)\\)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.32, + 0.373, + 0.332 + ], + "angle": 0, + "content": "9: current_iter \\(\\leftarrow\\) current_iter + 1" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.334, + 0.401, + 0.348 + ], + "angle": 0, + "content": "10: if current_iter \\(>T_{k}\\) and \\(k < M\\) then" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.35, + 0.227, + 0.361 + ], + "angle": 0, + "content": "11: \\(k\\gets k + 1\\)" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.365, + 0.18, + 0.376 + ], + "angle": 0, + "content": "12: end if" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.38, + 0.19, + 0.391 + ], + "angle": 0, + "content": "13: end while" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.17, + 0.482, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.417, + 0.405, + 0.433 + ], + "angle": 0, + "content": "3.3. Efficient Adaptive Prompt Learning" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.439, + 0.484, + 0.59 + ], + "angle": 0, + "content": "Based on our VAE-based restoration framework, we design two different prompt strategies for the distinct behaviors of the VAE encoding and decoding processes. For the encoding process, since the VAE is pre-trained on clean images, directly encoding various degraded images introduces a significant domain gap. Therefore, we introduce a low-rank degradation-aware prompt to the encoder, enabling it to encode different degraded images into the pre-trained compact and regularized latent space, thereby reducing the difficulty of the subsequent restoration mapping." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.59, + 0.484, + 0.77 + ], + "angle": 0, + "content": "During the decoding stage, despite the successful mitigation of degradation effects within the latent space, challenges persist with the loss of high-frequency details when decoding back to pixel space. To address this issue, we incorporate an adapter to inject high-frequency information from the corresponding stages of the encoder into the decoder. However, since different degradations impact specific frequency bands differently, we employ a frequency-selective prompt in the decoding process. This approach facilitates an adaptive fusion of high-frequency details from the encoder with the low-frequency information from the decoder, tailored to the specific frequency bands affected by the degradation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.771, + 0.484, + 0.847 + ], + "angle": 0, + "content": "Additionally, to allow the network to effectively recognize the type of degradation, we pre-train a degradation classifier. Since this classification task is not particularly challenging, a simple classification network is sufficient to achieve high accuracy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.852, + 0.409, + 0.868 + ], + "angle": 0, + "content": "3.3.1. Degradation-Aware Low Rank Prompt" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The current methods for prompt injection in all-in-one tasks primarily follow the design of PromptIR [20]. In this frame" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.09, + 0.907, + 0.199 + ], + "angle": 0, + "content": "work, for the input feature \\(F_{i}\\in \\mathbb{R}^{H\\times W\\times C}\\), a learnable prompt tensor \\(P_{c}\\in \\mathbb{R}^{N\\times \\hat{H}\\times \\hat{W}\\times C}\\) is first constructed, where \\(N\\) denotes the number of prompt components, and \\(\\hat{H}\\) and \\(\\hat{W}\\) represent the spatial dimensions, while the channel count \\(C\\) aligns with that of the input. Subsequently, \\(P_{c}\\) is interpolated to match the size of \\(F_{i}\\), enabling further interaction to generate the prompts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.2, + 0.907, + 0.351 + ], + "angle": 0, + "content": "While this design has shown promising performance in tasks at standard resolutions, it faces two significant challenges in UHD scenarios. First, the entire learnable tensor \\( P_{c} \\) comprises a substantial number of parameters due to the dimensions \\( \\hat{H} \\times \\hat{W} \\). Furthermore, the prompt is homogeneous across all degradation types, which means its complexity cannot be adapted based on the difficulty of each specific degradation. This lack of customization makes the approach inefficient for UHD image restoration, where the difficulty of different degradation categories can vary significantly." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.351, + 0.907, + 0.488 + ], + "angle": 0, + "content": "Second, since the dimensions \\(\\hat{H}\\) and \\(\\hat{W}\\) of \\(P_{c}\\) must be predetermined and interpolated to ensure adaptability to feature resolution, traditional all-in-one tasks exhibit minimal differences between training and inference resolutions. This allows the trained \\(P_{c}\\) to generalize well during inference. However, in UHD tasks, training typically uses a patch size of \\(512 \\times 512\\) for efficiency, while inference requires processing full images at \\(3840 \\times 2160\\). This significant resolution gap limits the generalization of \\(P_{c}\\) during inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.907, + 0.551 + ], + "angle": 0, + "content": "To address the issue of parameter efficiency, inspired by the low-rank adaptation method [9], we perform a low-rank decomposition on the spatial dimensions of the original parameter matrix \\(P_{c}\\in \\mathbb{R}^{N\\times \\hat{H}\\times \\hat{W}\\times C}\\), represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.564, + 0.907, + 0.583 + ], + "angle": 0, + "content": "\\[\nP _ {c l} = \\operatorname {c a t} \\left[ B _ {n} A _ {n} \\right] _ {n \\in [ 1 \\dots N ]}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.906, + 0.673 + ], + "angle": 0, + "content": "where \\(B_{n}\\in \\mathbb{R}^{1\\times \\hat{H}\\times r\\times C}\\) and \\(A_{n}\\in \\mathbb{R}^{1\\times r\\times \\hat{W}\\times C}\\) represent the matrices resulting from the low-rank decomposition, and \\(\\operatorname {cat}[\\cdot ]\\) represents concatenation along the first dimension, with \\(r\\ll \\min (\\hat{H},\\hat{W})\\) . After low-rank decomposition, the original parameter matrix is effectively compressed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.907, + 0.795 + ], + "angle": 0, + "content": "Furthermore, for different types of degradations, we can specify different ranks \\( r \\). More complex degradations can use a larger \\( r \\) to enhance the parameter space, while simpler degradations can use a smaller \\( r \\) to improve parameter efficiency. The use of different ranks \\( r \\) introduces unique designs for various degradations, while the multiplication of the decomposed matrices preserves consistent dimensions, allowing efficient parallel processing through concatenation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To address the issue of resolution adaptability, after obtaining the low-rank prompt tensor \\( P_{cl} \\), we do not directly interpolate it to match the input features. Instead, considering the scale invariance of convolution operations, we utilize \\( P_{cl} \\) as a parameter space to learn multiple convolution operators. By using convolution as the core operator for prompt interactions, this approach is more effective in adapting to" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23125" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "the differences between inference and training resolutions compared to simple interpolation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.213 + ], + "angle": 0, + "content": "Specifically, the prompt generation process during the encoding phase is as follows. First, the input image is passed through a classifier to obtain \\( N + 1 \\) classification scores, where \\( N \\) represents the number of degradation types, and 1 denotes the clean image. We then further process these classification scores to derive \\( W_{1} \\) and \\( W_{2} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.218, + 0.484, + 0.245 + ], + "angle": 0, + "content": "\\[\nW _ {1} = \\operatorname {s o f t m a x} (\\sum (N), 1), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.245, + 0.332, + 0.26 + ], + "angle": 0, + "content": "\\[\nW _ {2} = \\operatorname {s o f t m a x} (N),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.266, + 0.484, + 0.341 + ], + "angle": 0, + "content": "where, \\( W_{1} \\) represents the binary classification score between degraded and clean images, serving to characterize the degree of degradation, while \\( W_{2} \\) corresponds to the \\( N \\)-way classification scores for the \\( N \\) different degradation types, indicating the specific type of degradation present." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.484, + 0.385 + ], + "angle": 0, + "content": "We then construct two simple pointwise convolutions to provide degradation degree prompts(DDP), as illustrated below:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.39, + 0.484, + 0.414 + ], + "angle": 0, + "content": "\\[\nF _ {i 1} ^ {\\prime} = \\operatorname {c a t} \\left[ P _ {\\text {d e g}} \\left(F _ {i}\\right), P _ {\\text {c l e a n}} \\left(F _ {i}\\right) \\right], \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.41, + 0.296, + 0.427 + ], + "angle": 0, + "content": "\\[\nF _ {i 1} = F _ {i 1} ^ {\\prime} \\odot W _ {1},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.484, + 0.478 + ], + "angle": 0, + "content": "where, \\( P_{deg}(\\cdot) \\) and \\( P_{clean}(\\cdot) \\) denote the pointwise convolutions used to provide degradation degree prompts, while \\( \\odot \\) represents the weighted summation of their outputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.48, + 0.485, + 0.599 + ], + "angle": 0, + "content": "After obtaining the degradation-aware feature \\(F_{i1}\\), we proceed to provide prompts for degradation types(DTP). First, we define the rank of the parameter space for each degradation type as \\(R_{N} = [r_{1}, r_{2}, \\dots, r_{N}]\\), which allows us to derive the parameter space \\(P_{cl}\\) based on Equation (1). Next, we utilize a simple MLP to generate a set of deep-wise convolution kernels \\(K_{s} = [(k_{1}:d_{1}), (k_{2}:d_{2}), \\dots, (k_{m}:d_{m})]\\), characterized by varying kernel sizes and dilation rates." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.484, + 0.734 + ], + "angle": 0, + "content": "The feature \\( F_{i1} \\) is then processed in parallel through all multi-scale convolution group (MCKG), followed by a weighted summation with \\( W_{2} \\). Due to the impact of parallel multi-branch convolutional kernels on inference efficiency [6], we employ the structural reparameterization method from [7]. In the inference phase, all convolutional kernels are reparameterized and merged into a single large convolutional kernel. This process can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.739, + 0.304, + 0.756 + ], + "angle": 0, + "content": "\\[\nK _ {s} = \\mathrm {M L P} _ {i \\in [ 1: m ]} (P _ {c l}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.759, + 0.484, + 0.799 + ], + "angle": 0, + "content": "\\[\nF _ {i 2} ^ {\\prime} = \\left\\{ \\begin{array}{l l} \\sum \\left(K _ {s} \\left(F _ {i 1}\\right)\\right) & \\text {i f i n t r a i n i n g p h a s e}, \\\\ \\operatorname {R e p} \\left[ K _ {s} \\right] \\left(F _ {i 1}\\right) & \\text {i f i n t e s t i n g p h a s e}, \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.802, + 0.255, + 0.82 + ], + "angle": 0, + "content": "\\[\nF _ {i 2} = W _ {2} \\odot F _ {i 2} ^ {\\prime},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where, \\(\\mathrm{Rep}[\\cdot ]\\) refers to the reparameterization operation. The resultant \\(F_{i2}\\) serves as the low-rank degradation-aware prompt, which is integrated into the encoder's adapter. This mechanism facilitates the encoder's ability to encode diverse degradations into a cohesive and compact latent space." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.107 + ], + "angle": 0, + "content": "3.3.2. Degradation-Specific Frequency Selection Prompt" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.11, + 0.907, + 0.217 + ], + "angle": 0, + "content": "Given that different degradations affect distinct frequency bands [4] [11], we design prompts in the decoding phase to facilitate adaptive selection of frequency components. Specifically, this involves integrating medium-frequency information \\(\\{F_E^{HL}, F_E^{LH}\\}\\) and high-frequency information \\(F_E^{HH}\\) from the encoder, along with low-frequency information \\(F_D^{LL}\\) from the decoder." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.217, + 0.907, + 0.359 + ], + "angle": 0, + "content": "We begin by applying the Fast Fourier Transform (FFT) to each frequency band feature in the adapter branch, represented as \\( F_{w} = \\{F_{D}^{LL}, F_{E}^{HL}, F_{E}^{LH}, F_{E}^{HH}\\} \\in \\mathbb{R}^{B \\times 4 \\times \\frac{\\hat{H}}{2} \\times \\frac{\\hat{W}}{2} \\times C} \\). Pointwise convolution is applied to each band to efficiently capture global features and expand the channels by a factor of \\( N \\), resulting in \\( F_{w1} \\in \\mathbb{R}^{B \\times 4 \\times N \\times \\frac{\\hat{H}}{2} \\times \\frac{\\hat{W}}{2} \\times C} \\) after reshaping. A weighted summation with \\( W_{2} \\) is then performed along the \\( N \\) dimension, followed by the Inverse FFT (IFFT) to obtain \\( F_{w1}' \\):" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.366, + 0.906, + 0.392 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} F _ {w 1} = \\operatorname {c a t} \\left[ \\operatorname {P c o n v} _ {\\mathrm {i} \\in [ \\text {l o w}, \\text {m i d}, \\text {h i g h} ]} \\left(\\mathcal {F} \\left(F _ {w}\\right)\\right) \\right], \\\\ F ^ {\\prime} = \\mathcal {F} ^ {- 1} \\left(W, W, F _ {w}\\right) \\end{array} \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.387, + 0.738, + 0.404 + ], + "angle": 0, + "content": "\\[\nF _ {w 1} ^ {\\prime} = \\mathcal {F} ^ {- 1} \\left(W _ {2} \\odot F _ {w 1}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.411, + 0.907, + 0.485 + ], + "angle": 0, + "content": "where, \\(\\mathrm{Pconv}_{\\mathrm{i}\\in [\\mathrm{low},\\mathrm{mid},\\mathrm{high}]}\\) denotes the pointwise convolutions used to process the low, mid, and high frequency bands, while \\(\\mathcal{F}(\\cdot)\\) and \\(\\mathcal{F}^{-1}(\\cdot)\\) represent the spatial dimensional Fourier transform and its corresponding inverse transform, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.909, + 0.595 + ], + "angle": 0, + "content": "Subsequently, pixel shuffle is applied for frequency mixing, followed by a sigmoid activation to obtain the filter kernel \\( F_{w2} \\in \\mathbb{R}^{B \\times \\hat{H} \\times \\hat{W} \\times C} \\). For the input \\( F_{d} \\in \\mathbb{R}^{B \\times \\hat{H} \\times \\hat{W} \\times C} \\) from the decoder, we perform FFT to shift to the frequency domain and apply \\( F_{w2} \\) for frequency domain filtering. Finally, the IFFT is conducted to obtain the frequency-selective prompt \\( F_{w2}' \\):" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.601, + 0.838, + 0.619 + ], + "angle": 0, + "content": "\\[\nF _ {w 2} = \\operatorname {s i g m o i d} \\left(\\operatorname {p i x e l s h u f f u l e} \\left(F _ {w 1} ^ {\\prime}\\right)\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.617, + 0.906, + 0.639 + ], + "angle": 0, + "content": "\\[\nF _ {w 2} ^ {\\prime} = \\mathcal {F} ^ {- 1} \\left(F _ {w 2} \\otimes \\mathcal {F} \\left(F _ {w 2}\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.646, + 0.906, + 0.722 + ], + "angle": 0, + "content": "where \\(\\otimes\\) denotes element-wise multiplication. \\(F_{w2}^{\\prime}\\) serves as a frequency-selective prompt injected through the adapter in the decoder, enabling the decoding process to perform a more customized multi-band feature fusion tailored to the frequency characteristics of the degradation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.734, + 0.646, + 0.75 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.704, + 0.774 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We have developed a comprehensive benchmark for evaluating UHD all-in-one restoration performance, based on datasets from UHD-LL [15], UHD-blur [5], UHD-haze [46], UHD-rain [2], and UHD-haze [31]. Additionally, we constructed a denoising dataset, UHD-noise, using 4K images from [42] as the background. For additional implementation details and dataset descriptions, please refer to the supplementary materials." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23126" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.093, + 0.249, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.171, + 0.207, + 0.18 + ], + "angle": 0, + "content": "Image deblurring" + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.094, + 0.33, + 0.127 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.265, + 0.128, + 0.312, + 0.136 + ], + "angle": 0, + "content": "Histformer" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.094, + 0.411, + 0.127 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.353, + 0.128, + 0.395, + 0.136 + ], + "angle": 0, + "content": "PromptIR" + }, + { + "type": "image", + "bbox": [ + 0.416, + 0.094, + 0.493, + 0.127 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.444, + 0.128, + 0.46, + 0.136 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.137, + 0.33, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.171, + 0.315, + 0.18 + ], + "angle": 0, + "content": "Gridformer" + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.137, + 0.411, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.171, + 0.387, + 0.18 + ], + "angle": 0, + "content": "HAIR" + }, + { + "type": "image", + "bbox": [ + 0.416, + 0.137, + 0.493, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.444, + 0.171, + 0.467, + 0.18 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.181, + 0.249, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.261, + 0.225, + 0.269 + ], + "angle": 0, + "content": "Low light enhancement" + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.182, + 0.33, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.271, + 0.218, + 0.318, + 0.225 + ], + "angle": 0, + "content": "Histformer" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.182, + 0.411, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.353, + 0.218, + 0.395, + 0.226 + ], + "angle": 0, + "content": "PromptIR" + }, + { + "type": "image", + "bbox": [ + 0.416, + 0.182, + 0.493, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.444, + 0.218, + 0.46, + 0.225 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.226, + 0.33, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.26, + 0.315, + 0.268 + ], + "angle": 0, + "content": "Gridformer" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.226, + 0.411, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.36, + 0.26, + 0.386, + 0.268 + ], + "angle": 0, + "content": "HAIR" + }, + { + "type": "image", + "bbox": [ + 0.416, + 0.226, + 0.493, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.444, + 0.26, + 0.467, + 0.268 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.094, + 0.657, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.172, + 0.613, + 0.18 + ], + "angle": 0, + "content": "Image denoising" + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.094, + 0.738, + 0.127 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.129, + 0.726, + 0.137 + ], + "angle": 0, + "content": "Histformer" + }, + { + "type": "image", + "bbox": [ + 0.743, + 0.094, + 0.819, + 0.127 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.763, + 0.128, + 0.805, + 0.137 + ], + "angle": 0, + "content": "PromptIR" + }, + { + "type": "image", + "bbox": [ + 0.824, + 0.094, + 0.901, + 0.127 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.854, + 0.129, + 0.869, + 0.137 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.663, + 0.138, + 0.738, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.172, + 0.726, + 0.18 + ], + "angle": 0, + "content": "Gridformer" + }, + { + "type": "image", + "bbox": [ + 0.742, + 0.138, + 0.82, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.172, + 0.791, + 0.18 + ], + "angle": 0, + "content": "HAIR" + }, + { + "type": "image", + "bbox": [ + 0.823, + 0.138, + 0.901, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.855, + 0.172, + 0.875, + 0.18 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.182, + 0.659, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.261, + 0.611, + 0.27 + ], + "angle": 0, + "content": "Image dehazing" + }, + { + "type": "image", + "bbox": [ + 0.663, + 0.182, + 0.738, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.218, + 0.726, + 0.226 + ], + "angle": 0, + "content": "Histformer" + }, + { + "type": "image", + "bbox": [ + 0.742, + 0.182, + 0.82, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.764, + 0.217, + 0.805, + 0.226 + ], + "angle": 0, + "content": "PromptIR" + }, + { + "type": "image", + "bbox": [ + 0.824, + 0.182, + 0.901, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.854, + 0.218, + 0.869, + 0.225 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.663, + 0.226, + 0.74, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.678, + 0.26, + 0.726, + 0.269 + ], + "angle": 0, + "content": "Gridformer" + }, + { + "type": "image", + "bbox": [ + 0.743, + 0.226, + 0.82, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.26, + 0.791, + 0.268 + ], + "angle": 0, + "content": "HAIR" + }, + { + "type": "image", + "bbox": [ + 0.824, + 0.226, + 0.901, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.854, + 0.261, + 0.875, + 0.269 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.272, + 0.89, + 0.285 + ], + "angle": 0, + "content": "Figure 3. A comparison of visual results for four types of degradation removal with other state-of-the-art (SOTA) all-in-one methods." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.298, + 0.907, + 0.327 + ], + "angle": 0, + "content": "Table 1. Comparison to state-of-the-art on four degradations. PSNR (dB, ↑), SSIM (↑), and LPIPS (↓), and FS represents full-size 4K image inference. FLOPs are computed for an input size of \\(256 \\times 256\\). Best and second best performances are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.337, + 0.867, + 0.485 + ], + "angle": 0, + "content": "
MethodFSFLOPsParams.Low LightDeblurringDehazingDenoisingAverage
UHD-LLUHD-blurUHD-haze\\( UHDN_{\\sigma=15} \\)\\( UHDN_{\\sigma=25} \\)\\( UHDN_{\\sigma=50} \\)
AIRNet [14]✘301G9M19.24.80921.89.75718.37.81221.33.88720.78.78418.79.47520.07.754.2843
IDR [41]✘88G15.3M23.12.91024.67.79319.12.76827.48.91225.86.87224.57.65424.14.822.2684
PromptIR [20]✘158G33M23.44.90225.77.78219.97.72728.43.92426.74.89823.72.58424.68.803.2571
CAPTNet [8]✘25G24.3M23.96.92026.11.79819.46.86825.58.86523.24.88421.98.50823.39.809.3466
NDR-Restore [38]✘196G36.9M23.84.89424.25.80220.08.89225.62.91224.37.89722.94.66923.52.846.3126
Gridformer [33]✘367G34M23.12.89825.82.78319.24.86936.04.93731.72.89826.24.62327.03.836.3754
DiffUIR-L [44]✘10G36.2M21.56.81223.85.74318.28.86436.84.93832.42.89726.08.64826.51.818.2564
Histoformer [27]✘91G16.6M23.22.90825.62.78219.78.90326.88.84525.64.87423.13.65924.04.829.3524
adalR [4]✘147G28.7M23.57.91626.35.80118.44.90132.84.92130.48.90126.48.67226.36.857.3429
HAIR [1]✘41G29M25.75.92225.78.79820.00.89435.54.91630.84.87826.26.65727.36.847.2822
Ours✓4G1.6M27.11.92526.48.80320.94.92338.94.97533.99.90327.95.67729.23.868.2541
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.497, + 0.907, + 0.525 + ], + "angle": 0, + "content": "Table 2. Comparison to state-of-the-art on six degradations. PSNR (dB, ↑), SSIM (↑), LPIPS (↓) and FS represents full-size 4K image inference. FLOPs are computed for an input size of \\(256 \\times 256\\). Best and second best performances are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.535, + 0.867, + 0.684 + ], + "angle": 0, + "content": "
MethodFSFLOPsParams.Low LightDeblurringDehazingDenoisingDerainingDesnowingAverage
UHD-LLUHD-blurUHD-hazeUHDN\\(_{\\sigma=50}\\)UHD-rainUHD-snow
AIRNet [14]✘301G9M22.68.88723.52.87618.24.84622.38.87626.35.87627.38.92423.43.874.1861
IDR [41]✘88G15.3M24.33.91525.64.78818.68.87929.64.90628.82.90630.48.94526.27.890.1912
PromptIR [20]✘158G33M23.3.91126.48.80520.14.90124.88.83528.89.89730.78.96625.74.886.2155
CAPTNet [8]✘25G24.3M24.97.92126.32.79620.32.90321.64.56929.34.90832.21.97425.80.845.2861
NDR-Restore [38]✘196G36.9M25.12.88525.64.79119.21.89631.44.91529.24.89728.41.94826.51.889.3108
Gridformer [33]✘367G34M23.92.89825.68.78218.87.88932.86.91529.37.90428.24.94226.49.895.2321
DiffUIR-L [44]✘10G36.2M22.64.90225.08.78518.62.88933.25.92827.89.88627.36.94525.81.889.1844
Histoformer [27]✘91G16.6M25.73.91526.55.79618.73.89733.05.92427.96.88427.56.97126.59.898.1855
adalR [4]✘147G28.7M23.84.91826.86.80319.34.91032.46.92328.18.90127.72.95326.40.901.2492
HAIR [1]✘41G29M25.22.89724.77.79918.75.88332.50.91528.76.89327.89.96826.31.892.2607
Ours✓4G1.6M26.91.92426.95.80721.81.93133.73.93429.90.91532.73.97928.67.915.1839
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.696, + 0.907, + 0.723 + ], + "angle": 0, + "content": "Table 3. Comparison to state-of-the-art for single degradations. PSNR (dB, ↑), SSIM (↑) and LPIPS (↓). Best and second best performances are highlighted." + }, + { + "type": "table_caption", + "bbox": [ + 0.231, + 0.724, + 0.301, + 0.735 + ], + "angle": 0, + "content": "(a) Low light" + }, + { + "type": "table_caption", + "bbox": [ + 0.46, + 0.724, + 0.539, + 0.735 + ], + "angle": 0, + "content": "(b) Deblurring" + }, + { + "type": "table_caption", + "bbox": [ + 0.697, + 0.724, + 0.767, + 0.735 + ], + "angle": 0, + "content": "(c) Dehazing" + }, + { + "type": "table", + "bbox": [ + 0.151, + 0.737, + 0.848, + 0.879 + ], + "angle": 0, + "content": "
MethodFSUHD-LLLIPIPSMethodFSUHD-blurLIPIPSMethodFSUHD-hazeLIPIPS
Restormer [40]✘21.54.843.3608Restormer [40]✘25.21.752.3695Restormer [40]✘12.72.693.4560
LLformer [32]✘24.06.858.3516Uformer [35]✘25.26.751.3851Uformer [35]✘19.83.921.4220
UHDFour [15]✓26.22.900.2390Stripformer [28]✘25.05.750.3740DehazeFormer [26]✘15.37.737.3998
UHDFformer [29]✓27.11.927.2240FFTformer [12]✘25.41.757.3708TaylorFormer [22]✘20.99.919.3124
LMAR [39]✓26.27.919.2248UHDformer [29]✓28.82.844.2350UHDformer [29]✓22.58.942.1188
AirNet [14]✘24.61.827.3841AirNet✘23.48.824.3942AirNet✘18.44.848.4018
PromptIR [20]✘24.81.912.3116PromptIR✘23.24.829.3754PromptIR✘20.21.927.3642
Histoformer [27]✘25.46.894.3213Histoformer✘27.12.848.2986Histoformer✘21.72.948.2942
HAIR [1]✘26.08.901.2964HAIR✘26.74.847.3867HAIR✘20.38.912.2412
Ours✓27.22.929.2042Ours✓28.91.851.2129Ours✓23.24.953.1086
" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23127" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.46, + 0.108 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-Art Methods" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.113, + 0.355, + 0.129 + ], + "angle": 0, + "content": "4.2.1. All-in-One: Four Degradations" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.132, + 0.486, + 0.299 + ], + "angle": 0, + "content": "We compare our UHD-Processor with specialized all-in-one restoration methods, trained on four degradation types: dehazing, denoising, deblurring, and low-light enhancement. As shown in Tab. 1, our method achieves state-of-the-art (SOTA) performance on both individual datasets and overall metrics. Specifically, the average PSNR improves by 1.93 dB over the second-best method, HAIR, while our method requires only \\(5.5\\%\\) of the parameters and \\(9.7\\%\\) of the FLOPS of Histoformer. Visual results, shown in Figure 3, further demonstrate that our approach effectively achieves the best degradation removal." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.306, + 0.342, + 0.322 + ], + "angle": 0, + "content": "4.2.2. All-in-One: Six Degradations" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.325, + 0.486, + 0.476 + ], + "angle": 0, + "content": "We further extended the experiment to include six degradation types by introducing image desnow and image derain, to validate the effectiveness of our method in a more complex all-in-one setup. As shown in Tab. 2, our method continues to achieve SOTA performance on both individual datasets and average metrics. Specifically, the average PSNR improves by \\(1.7\\mathrm{dB}\\) over the second-best method, Histoformer, while the number of parameters and FLOPS are only \\(9.6\\%\\) and \\(4.3\\%\\) respectively. More visual comparison results can be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.484, + 0.328, + 0.5 + ], + "angle": 0, + "content": "4.2.3. Single-Degradation Results" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.503, + 0.485, + 0.595 + ], + "angle": 0, + "content": "We further evaluate the performance of our method by training instances under a single degradation protocol, comparing them with both task-specific methods and all-in-one approaches. The results, shown in Tab. 3, demonstrate that our method still achieves the best performance, validating the effectiveness of our approach." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.604, + 0.245, + 0.619 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.625, + 0.427, + 0.642 + ], + "angle": 0, + "content": "4.3.1. Frequency-Progressive Learning Strategy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.486, + 0.901 + ], + "angle": 0, + "content": "As shown in Tab. 4, we validate the effectiveness of the proposed Frequency Domain Progressive Learning (FDPL). First, compared to directly training on mixed data, using FDPL results in a significant improvement of \\(1.99\\mathrm{dB}\\) PSNR. Next, we evaluate the sequential learning (SL) strategy proposed in [13], which advocates for learning local degradations before global ones. We test two learning orders: SL1 (noise \\(\\rightarrow\\) blur \\(\\rightarrow\\) low-light \\(\\rightarrow\\) haze) and SL2 (blur \\(\\rightarrow\\) noise \\(\\rightarrow\\) haze \\(\\rightarrow\\) low-light). The experimental results show that not all local-to-global learning orders lead to performance gains. In fact, SL1 even caused some performance degradation, while SL2 yielded better results. This suggests that sequential learning strategies are sensitive to the order of tasks, and the optimal sequence may not only depend on the tasks themselves but also on the specific data set, scene, or degradation severity. Therefore, it is challenging to define a universally effective task order. In contrast, our" + }, + { + "type": "table_caption", + "bbox": [ + 0.563, + 0.09, + 0.857, + 0.105 + ], + "angle": 0, + "content": "Table 4. Ablation Study of the Learning Strategy." + }, + { + "type": "table", + "bbox": [ + 0.578, + 0.106, + 0.842, + 0.151 + ], + "angle": 0, + "content": "
MethodMixw/SL1w/SL2w/FDPL (Ours)
PSNR27.2427.0828.1229.23
SSIM0.8530.8460.8720.874
" + }, + { + "type": "table_caption", + "bbox": [ + 0.519, + 0.17, + 0.898, + 0.183 + ], + "angle": 0, + "content": "Table 5. Ablation Study of Efficient Adaptive Prompt Learning." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.184, + 0.91, + 0.23 + ], + "angle": 0, + "content": "
Methodw/o DALPw/o DDPw/o DTPw/o DSRw/o MCKGw/o DFSPOurs
PSNR27.8628.4928.3428.8428.7228.2129.23
SSIM0.8620.8660.8580.8640.8620.8650.874
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.261, + 0.908, + 0.337 + ], + "angle": 0, + "content": "FDPL strategy is task-agnostic and offers a more generalizable learning framework. Experimental results confirm that FDPL leads to more significant performance improvements. Further discussions on sequential learning can be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.353, + 0.812, + 0.369 + ], + "angle": 0, + "content": "4.3.2. Efficient Adaptive Prompt Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.374, + 0.909, + 0.662 + ], + "angle": 0, + "content": "The ablation study of the proposed prompt learning method, as shown in Tab. 5, demonstrates that both the DALP used in the encoding process and the DFSP used in the decoding process lead to significant performance improvements. We further analyze the components of DALP, and the results show that both the Degradation Degree Prompt (DDP) and the Degradation Type Prompt (DTP) provide more granular prompt information, improving performance consistently. The Degradation-Specific Rank (DSR) component introduces task-specific low-rank decompositions, allowing for the selection of an appropriate rank for each task, which enhances parameter efficiency while improving performance. Moreover, the Multi-Scale Convolution Kernel Group (MCKG) introduces a multi-scale convolutional branch, which strengthens the representation capacity of the prompt generation process. During inference, the multi-branch convolutions are reparameterized for efficient execution. For more detailed ablation experiments and efficiency analyses, please refer to the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.677, + 0.634, + 0.693 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.91, + 0.903 + ], + "angle": 0, + "content": "In this work, we proposed UHD-Processor, an efficient and effective framework for UHD all-in-one image restoration. By leveraging Frequency Domain Decoupling Progressive Learning Strategy and an efficient prompt-integrated VAE framework, our method addresses the unique challenges of UHD restoration tasks, including denoising, deblurring, deraining, dehazing, and low-light enhancement. Experimental results demonstrated that our approach significantly outperforms existing methods, achieving state-of-the-art results with reduced computational costs. The combination of progressive frequency learning and targeted prompt-based adaptation provides a strong foundation for future advancements in UHD image restoration." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23128" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.486, + 0.161 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (NSFC) under Grants 62225207, 62436008, 62422609 and 62276243." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.174, + 0.188, + 0.189 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.199, + 0.484, + 0.239 + ], + "angle": 0, + "content": "[1] Jin Cao, Yi Cao, Li Pang, Deyu Meng, and Xiangyong Cao. Hair: Hypernetworks-based all-in-one image restoration, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.242, + 0.484, + 0.295 + ], + "angle": 0, + "content": "[2] Hongming Chen, Xiang Chen, Chen Wu, Zhuoran Zheng, Jinshan Pan, and Xianping Fu. Towards ultra-high-definition image deraining: A benchmark and an efficient method, 2024. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.484, + 0.354 + ], + "angle": 0, + "content": "[3] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.355, + 0.484, + 0.409 + ], + "angle": 0, + "content": "[4] Yuning Cui, Syed Waqas Zamir, Salman Khan, Alois Knoll, Mubarak Shah, and Fahad Shahbaz Khan. Adair: Adaptive all-in-one image restoration via frequency mining and modulation, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.411, + 0.484, + 0.478 + ], + "angle": 0, + "content": "[5] Senyou Deng, Wenqi Ren, Yanyang Yan, Tao Wang, Fenglong Song, and Xiaochun Cao. Multi-scale separable network for ultra-high-definition video deblurring. In the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14030-14039, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.484, + 0.548 + ], + "angle": 0, + "content": "[6] Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13733-13742, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.484, + 0.633 + ], + "angle": 0, + "content": "[7] Xiaohan Ding, Yiyuan Zhang, Yixiao Ge, Sijie Zhao, Lin Song, Xiangyu Yue, and Ying Shan. Unireplknet: A universal perception large-kernel convnet for audio video point cloud time-series and image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5513-5524, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.484, + 0.688 + ], + "angle": 0, + "content": "[8] Hu Gao, Jing Yang, Ning Wang, Jingfan Yang, Ying Zhang, and Depeng Dang. Prompt-based all-in-one image restoration using cnns and transformer. arXiv preprint arXiv:2309.03063, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.691, + 0.484, + 0.758 + ], + "angle": 0, + "content": "[9] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.484, + 0.816 + ], + "angle": 0, + "content": "[10] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[11] Xingyu Jiang, Xiuhui Zhang, Ning Gao, and Yue Deng. When fast fourier transform meets transformer for image restoration. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In 2023 IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.199, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition (CVPR), pages 5886-5895, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.124, + 0.908, + 0.178 + ], + "angle": 0, + "content": "[13] Xiangtao Kong, Chao Dong, and Lei Zhang. Towards effective multiple-in-one image restoration: A sequential and prompt learning strategy. arXiv preprint arXiv:2401.03379, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.181, + 0.907, + 0.236 + ], + "angle": 0, + "content": "[14] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-In-One Image Restoration for Unknown Corruption. In IEEE Conference on Computer Vision and Pattern Recognition, New Orleans, LA, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.239, + 0.907, + 0.293 + ], + "angle": 0, + "content": "[15] Chongyi Li, Chun-Le Guo, Man Zhou, Zhexin Liang, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Embedding fourier for ultra-high-definition low-light image enhancement. In ICLR, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.296, + 0.907, + 0.366 + ], + "angle": 0, + "content": "[16] Xiangtai Li, Haobo Yuan, Wei Li, Henghui Ding, Size Wu, Wenwei Zhang, Yining Li, Kai Chen, and Chen Change Loy. Omg-seg: Is one model good enough for all segmentation? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27948-27959, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.369, + 0.906, + 0.41 + ], + "angle": 0, + "content": "[17] Zilong Li, Yiming Lei, Chenglong Ma, Junping Zhang, and Hongming Shan. Prompt-in-prompt learning for universal image restoration, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.906, + 0.467 + ], + "angle": 0, + "content": "[18] Ziwei Luo, Fredrik K. Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B. Schön. Controlling vision-language models for multi-task image restoration. In The Twelfth International Conference on Learning Representations, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.471, + 0.907, + 0.526 + ], + "angle": 0, + "content": "[19] Anastasia Pentina, Viktoriia Sharmanska, and Christoph H Lampert. Curriculum learning of multiple tasks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5492-5500, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.529, + 0.907, + 0.583 + ], + "angle": 0, + "content": "[20] Vaishnav Potlapalli, Syed Waqas Zamir, Salman Khan, and Fahad Khan. Prompt: Prompting for all-in-one image restoration. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.586, + 0.906, + 0.641 + ], + "angle": 0, + "content": "[21] Yifan Pu, Weicong Liang, Yiduo Hao, Yuhui Yuan, Yukang Yang, Chao Zhang, Han Hu, and Gao Huang. Rank-detr for high quality object detection. Advances in Neural Information Processing Systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.644, + 0.907, + 0.698 + ], + "angle": 0, + "content": "[22] Yuwei Qiu, Kaihao Zhang, Chenxi Wang, Wenhan Luo, Hongdong Li, and Zhi Jin. Mb-taylorformer: Multi-branch efficient transformer expanded by taylor formula for image dehazing, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.702, + 0.908, + 0.77 + ], + "angle": 0, + "content": "[23] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International conference on machine learning, pages 5301-5310. PMLR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.774, + 0.908, + 0.841 + ], + "angle": 0, + "content": "[24] Wenqi Ren, Senyou Deng, Kaihao Zhang, Fenglong Song, Xiaochun Cao, and Ming-Hsuan Yang. Fast ultra high-definition video deblurring via multi-scale separable network. International Journal of Computer Vision, 132(5):1817-1834, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[25] Basri Ronen, David Jacobs, Yoni Kasten, and Shira Kritchman. The convergence rate of neural networks for learned functions of different frequencies. Advances in Neural Information Processing Systems, 32, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "23129" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[26] Yuda Song, Zhuqing He, Hui Qian, and Xin Du. Vision transformers for single image dehazing. IEEE Transactions on Image Processing, 32:1927-1941, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.192 + ], + "angle": 0, + "content": "[27] Shangquan Sun, Wenqi Ren, Xinwei Gao, Rui Wang, and Xiaochun Cao. Restoring images in adverse weather conditions via histogram transformer. In European Conference on Computer Vision, pages 111-129. Springer, 2025. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.194, + 0.483, + 0.248 + ], + "angle": 0, + "content": "[28] Fu-Jen Tsai, Yan-Tsung Peng, Yen-Yu Lin, Chung-Chi Tsai, and Chia-Wen Lin. Stripformer: Strip transformer for fast image deblurring. In European conference on computer vision, pages 146-162. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.483, + 0.319 + ], + "angle": 0, + "content": "[29] Cong Wang, Jinshan Pan, Wei Wang, Gang Fu, Siyuan Liang, Mengzhu Wang, Xiao-Ming Wu, and Jun Liu. Correlation matching transformation transformers for uhd image restoration. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5336-5344, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.483, + 0.389 + ], + "angle": 0, + "content": "[30] Haohan Wang, Xindi Wu, Zeyi Huang, and Eric P Xing. High-frequency component helps explain the generalization of convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8684-8694, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.483, + 0.447 + ], + "angle": 0, + "content": "[31] Liyan Wang, Cong Wang, Jinshan Pan, Xiaofeng Liu, Weixiang Zhou, Xiaoran Sun, Wei Wang, and Zhixun Su. Ultra-high-definition image restoration: New benchmarks and a dual interaction prior-driven solution, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.483, + 0.518 + ], + "angle": 0, + "content": "[32] Tao Wang, Kaihao Zhang, Tianrun Shen, Wenhan Luo, Bjorn Stenger, and Tong Lu. Ultra-high-definition low-light image enhancement: A benchmark and transformer-based method. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2654-2662, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.483, + 0.601 + ], + "angle": 0, + "content": "[33] Tao Wang, Kaihao Zhang, Ziqian Shao, Wenhan Luo, Bjorn Stenger, Tong Lu, Tae-Kyun Kim, Wei Liu, and Hongdong Li. Gridformer: Residual dense transformer with grid structure for image restoration in adverse weather conditions. International Journal of Computer Vision, pages 1-23, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.483, + 0.646 + ], + "angle": 0, + "content": "[34] Xin Wang, Yudong Chen, and Wenwu Zhu. A survey on curriculum learning. IEEE transactions on pattern analysis and machine intelligence, 44(9):4555-4576, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.483, + 0.716 + ], + "angle": 0, + "content": "[35] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17683-17693, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[36] Boxue Xiao, Zhuoran Zheng, Xiang Chen, Chen Lv, Yunliang Zhuang, and Tao Wang. Single uhd image dehazing via interpretable pyramid network, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[37] Zhi-Qin John Xu, Yaoyu Zhang, and Yanyang Xiao. Training behavior of deep neural network in frequency domain. In Neural Information Processing: 26th International Conference, ICONIP 2019, Sydney, NSW, Australia, December 12–15, 2019, Proceedings, Part I 26, pages 264–274. Springer, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[38] Mingde Yao, Ruikang Xu, Yuanshen Guan, Jie Huang, and Zhiwei Xiong. Neural degradation representation learning for all-in-one image restoration. IEEE Transactions on Image Processing, 2024. 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.174 + ], + "angle": 0, + "content": "[39] Wei Yu, Jie Huang, Bing Li, Kaiwen Zheng, Qi Zhu, Man Zhou, and Feng Zhao. Empowering resampling operation for ultra-high-definition image enhancement with model-aware guidance. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25722-25731, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.907, + 0.232 + ], + "angle": 0, + "content": "[40] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.907, + 0.303 + ], + "angle": 0, + "content": "[41] Jinghao Zhang, Jie Huang, Mingde Yao, Zizheng Yang, Hu Yu, Man Zhou, and Feng Zhao. Ingredient-oriented multidegradation learning for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5825-5835, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.907, + 0.373 + ], + "angle": 0, + "content": "[42] Kaihao Zhang, Dongxu Li, Wenhan Luo, Wenqi Ren, Bjorn Stenger, Wei Liu, Hongdong Li, and Yang Ming-Hsuan. Benchmarking ultra-high-definition image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.907, + 0.443 + ], + "angle": 0, + "content": "[43] Wenlong Zhang, Xiaohui Li, SHI Guangyuan, Xiangyu Chen, Yu Qiao, Xiaoyun Zhang, Xiao-Ming Wu, and Chao Dong. Real-world image super-resolution as multi-task learning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.445, + 0.907, + 0.528 + ], + "angle": 0, + "content": "[44] Dian Zheng, Xiao-Ming Wu, Shuzhou Yang, Jian Zhang, Jian-Fang Hu, and Wei-Shi Zheng. Selective hourglass mapping for universal image restoration based on diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25445-25455, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.53, + 0.907, + 0.571 + ], + "angle": 0, + "content": "[45] Zhuoran Zheng and Xiuyi Jia. Uhd image deblurring via multi-scale cubic-mixer. arXiv preprint arXiv:2206.03678, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.572, + 0.907, + 0.642 + ], + "angle": 0, + "content": "[46] Zhuoran Zheng, Wenqi Ren, Xiaochun Cao, Xiaobin Hu, Tao Wang, Fenglong Song, and Xiuyi Jia. Ultra-high-definition image dehazing via multi-guided bilateral learning. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16180-16189, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.642, + 0.907, + 0.711 + ], + "angle": 0, + "content": "[47] Yanqi Zhou, Tao Lei, Hanxiao Liu, Nan Du, Yanping Huang, Vincent Zhao, Andrew M Dai, Quoc V Le, James Laudon, et al. Mixture-of-experts with expert choice routing. Advances in Neural Information Processing Systems, 35:7103-7114, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.711 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23130" + } + ] +] \ No newline at end of file diff --git a/2025/UHD-processer_ Unified UHD Image Restoration with Progressive Frequency Learning and Degradation-aware Prompts/2e2768b6-2887-4bd5-935c-d0b880ebe319_origin.pdf b/2025/UHD-processer_ Unified UHD Image Restoration with Progressive Frequency Learning and Degradation-aware Prompts/2e2768b6-2887-4bd5-935c-d0b880ebe319_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..35a891264b574813de0e9201f0e2c060543a9fb3 --- /dev/null +++ b/2025/UHD-processer_ Unified UHD Image Restoration with Progressive Frequency Learning and Degradation-aware Prompts/2e2768b6-2887-4bd5-935c-d0b880ebe319_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b0662c9ccd18ab79fd400d80b0572b798dd470b46c79b19b60cdfdd1023a7a7e +size 1238665 diff --git a/2025/UHD-processer_ Unified UHD Image Restoration with Progressive Frequency Learning and Degradation-aware Prompts/full.md b/2025/UHD-processer_ Unified UHD Image Restoration with Progressive Frequency Learning and Degradation-aware Prompts/full.md new file mode 100644 index 0000000000000000000000000000000000000000..88f1fcfd89e2fd403f60ce0d887dca3b99491dfa --- /dev/null +++ b/2025/UHD-processer_ Unified UHD Image Restoration with Progressive Frequency Learning and Degradation-aware Prompts/full.md @@ -0,0 +1,434 @@ +# UHD-processor: Unified UHD Image Restoration with Progressive Frequency Learning and Degradation-aware Prompts + +Yidi Liu, DongLi, Xueyang Fu†, Xin Lu, Jie Huang, Zheng-jun Zha +School of Information Science and Technology and MoE Key Laboratory of +Brain-inspired Intelligent Perception and Cognition, +University of Science and Technology of China, Hefei, 230026, China + +liuyidi2023@mail.ustc.edu.cn, xyfu@ustc.edu.cn † Corresponding Author + +# Abstract + +We introduce UHD-Processor, a unified and robust framework for all-in-one image restoration, which is particularly resource-efficient for Ultra-High-Definition (UHD) images. To address the limitations of traditional all-in-one methods that rely on complex restoration backbones, our strategy employs a frequency domain decoupling progressive learning technique, motivated by curriculum learning, to incrementally learn restoration mappings from low to high frequencies. This approach incorporates specialized sub-network modules to effectively tackle different frequency bands in a divide-and-conquer manner, significantly enhancing the learning capability of simpler networks. Moreover, to accommodate the high-resolution characteristics of UHD images, we developed a variational autoencoder (VAE)-based framework that reduces computational complexity by modeling a concise latent space. It integrates task-specific degradation awareness in the encoder and frequency selection in the decoder, enhancing task comprehension and generalization. Our unified model is able to handle various degradations such as denoising, deblurring, dehazing, lowlighting, etc. Experimental evaluations extensively showcase the effectiveness of our dual-strategy approach, significantly improving UHD image restoration and achieving cutting-edge performance across diverse conditions. The code will be available at https://github.com/1yd2022/UHD-processor + +# 1. Introduction + +Ultra-high-definition (UHD) images are increasingly utilized across various fields, prompting substantial research in UHD image restoration for specific degradation types, such as low-light enhancement, deblurring, dehazing, and deraining [2, 15, 29, 36]. However, real-world applications like object detection [21], video segmentation [16], and au + +![](images/72d4c86909474046daf42a55f096b0941a7b360703e19113aa0530755009f44b.jpg) +(a) Comparison of PSNR with SOTA All-in-One Methods + +![](images/f468997ca2074c2028c01d8cca464a3a0eba431f2f695ded2e63a6e228823f9e.jpg) +(b) Observation 1 + +![](images/b8fadce7ba92206ae2a8ccd616838cb66522413d76ac01bc0b22f67c0f24756b.jpg) +(c) Observation 2 + +![](images/046f1dc832e655b802a32dcf6eebc7f7feb05c97b3c5411637ead85369738392.jpg) +(d) Observation 3: The Variation of Gap Between Degradations Across Frequency Bands + +![](images/c6efaf951ec6ec0d222cdb2f14ea555fc4b1229a514857e39b200d2a1edddfca.jpg) +Figure 1. (a) Comparison of PSNR between our method and the previous state-of-the-art (SOTA) all-in-one method across six degradation settings in UHD images. (b) Observe the performance gains when integrating the same promptIR prompt branch into Restormer models with varying parameter sizes. (c) Comparison of low-to-high frequency training order on UHD and regular images, showing greater effectiveness for UHD tasks. (d) Quantitative evaluation of restoration tasks across low, high, and full-frequency bands. + +tonomous driving [3] face complex degradation patterns that vary across time and space, making reliance on single-function models impractical. This calls for a generalized model that can address multiple degradation types to support robust low-level vision tasks. + +In response to the above requirements, all-in-one(AIO) approaches aim to use a single model to manage various degradations, but they generally target standard-resolution images and encounter computational difficulties when scaled to UHD. This gap highlights the pressing need for an efficient + +all-in-one image restoration framework tailored to UHD images. For standard-resolution tasks [14, 20], successful approaches often include two key features, i.e., degradation-aware branches and strong network backbones. The former is implemented through techniques like prompts [10] or mixtures of experts (MoE) [47] to enhance adaptability to different degradation types, while the later contains advanced backbones like Restormer [40] to further strengthens learning and generalization. + +However, due to the high-resolution nature of UHD images, directly applying these strategies results in a significant computational burden, making full-resolution inference infeasible on consumer-grade GPUs. As shown in Figure 1(b), PromptIR extends the Restormer backbone with a prompt branch. We adjust their parameter counts by varying Restormer's module numbers, keeping the prompt branch fixed. Reducing parameters significantly lowers performance for both, meanwhile, the benefits introduced by the prompt branch also gradually diminish. + +These results indicate that while the prompt branch enhances all-in-one task performance, a strong backbone is crucial to fully leverage these improvements. Moreover, compared to Restormer, the additional degradation-aware prompt branch introduced by PromptIR adds 6.87M parameters, which is clearly burdensome for UHD tasks. + +Based on the above observations, this work aims to tackle two challenges faced by traditional AIO methods in UHD scenarios: reliance on complex restoration backbones and redundancy in degradation-aware branch parameters. + +To tackle the first challenge, at the model architecture level, we propose a VAE-based framework that transforms restoration tasks from pixel space into a compact latent space, thereby reducing the computational cost of UHD image processing. This alleviates the network's learning burden, enabling a simpler model to restore effectively. At the learning strategy level, inspired by curriculum learning [34], we develop a Frequency Domain Decoupling Progressive Learning Strategy. As shown in Figure 1(d), for UHD AIO tasks, higher frequency bands present greater restoration challenges. Consequently, we decompose the task into manageable frequency components, enabling sub-networks to progressively learn restoration mappings from low to high frequencies. This divide-and-conquer approach significantly enhances the network's adaptability and efficiency. + +To address the second challenge, we introduce efficient prompt strategies for both the VAE encoder and decoder. For encoding, we propose a low-rank degradation-aware prompt to indicate degradation type and severity. This is achieved by creating distinct low-rank parameter spaces per degradation type and learning scale-invariant convolution kernels, which adapt to UHD resolution differences between training and testing. For efficiency, we reparameterize these convolution branches during inference. To maintain efficiency, + +we reparameterize these convolution branches at inference. In decoding, we observe that degradation types exhibit distinct distributions across frequency bands. Accordingly, we introduce a frequency band selection prompt, allowing the decoder to adaptively choose between low-frequency features from the decoder and mid- or high-frequency features from the encoder based on the specific degradation type. + +By integrating these components, we propose UHD-Processor, an efficient all-in-one image restoration framework tailored for ultra-high-definition (UHD) images. As shown in Figure 1(a), our proposed UHD-Processor achieves state-of-the-art performance while reducing parameters by $95.1\%$ and FLOPs by $97.4\%$ compared to promptIR [20]. Our contributions are summarized as follows: + +1. We introduce a VAE-based image restoration framework that operates within a compact latent space, significantly enhancing UHD image restoration efficiency. +2. We design a frequency domain decoupling progressive learning strategy, which enables effective restoration across frequency bands from low to high frequencies through a divide-and-conquer approach, to improve adaptability to various optimization objectives. +3. We develop a prompt learning framework with low-rank degradation-aware prompts for encoding and frequency selection prompts for decoding, improving the network's adaptability to various degradation types. + +# 2. Related Work + +# 2.1. Ultra-High-Definition Restoration + +As the demand for UHD image processing grows in imaging systems, various approaches have been proposed to restore clarity in UHD images through diverse network architectures. These include bilateral learning for image dehazing [45], multi-scale separable-patch integration networks for video deblurring [24], and Fourier embedding networks for low-light image enhancement [15]. Recently, generalized architectures for UHD image restoration have evolved to address multiple restoration tasks through unified frameworks. For instance, [29] introduces a correlation matching mechanism between high-resolution and low-resolution branches to enhance the utilization of native resolution information, while [31] incorporates gradient and normal priors into the model design. However the approach about UHD all in one is not yet fully explored. + +# 2.2. All-in-One Image Restoration + +All-in-one image restoration, which aims to recover clean images from various types of degradation using a single unified model, has gained significant attention for its efficiency and practical applications in low-level vision tasks [14, 20, 33, 38, 41]. Early methods, such as AirNet [14], utilized contrastive learning to extract degra + +dation representations, while IDR [41]introduced a two-stage meta-learning approach to address diverse degradation types. More recently, prompt-based approaches like PromptIR [8]and PIP [17]further enhanced restoration flexibility through vision prompts. Additionally, large-scale pre-trained models, as demonstrated by DiffUIR [44] and DA-CLIP [18], have been adapted for all-in-one restoration with strong performance. Please refer to the supplementary material for more detailed related work. + +# 3. Method + +# 3.1. Overview + +The framework of our method is shown in Figure 2. Overall, our approach follows a two-stage paradigm. In the first stage, we train a VAE on clean images with an image reconstruction task, as shown in Figure 2(a). In the second stage, as shown in Figure 2(b), we freeze the VAE weights and train the remaining components for image restoration tasks. To mitigate high-frequency loss during the VAE encoding process, we pass high-frequency components from the wavelet transformation between the encoder adapter and the decoder adapter. Additionally, prompts are injected through the adapters: in the encoding phase, we apply degradation-aware low-rank prompts (Figure 2(d)) to adapt the VAE to different degradations, encoding them into a unified compact latent space; in the decoding phase, we use frequency selection prompts (Figure 2(c)) for each type of degradation, enabling efficient adaptive fusion of low-frequency decoder information with high-frequency encoder information. + +The latent restoration network is divided into $M$ subnetworks, each corresponding to one of the $M$ decomposed frequency bands of the input. During Frequency Domain Decoupling Progressive Learning (Figure 2(e)), we progressively increase from the lowest to the highest frequency bands, incrementally adding sub-networks until all frequency bands are restored. Attributing the strong performance to the robust latent space encoding and frequency-progressive learning strategy, we achieve excellent results with a simple latent space restoration network. In this work, we adopt the Cubic-Mixer [45] for its efficiency, results of other architectures and the interaction detail between the adapter and VAE, please refer to the supplementary materials. + +# 3.2. Frequency Domain Decoupling Progressive Learning Strategy + +For all-in-one tasks, which aim to handle multiple types of degradation with a single model, we can represent this as $F(\{\tau^t\}; \theta)_{t \in [1, N]}$ , where $N$ represents the number of tasks, $\tau^t$ denotes the input for task $t$ , and $\theta$ represents the model parameters. Each task $\tau^t$ corresponds to a distinct optimization objective; for instance, the low-light enhancement task focuses on adjusting the global brightness of the image, while + +the deraining task aims to remove local rain streaks. Consequently, it is challenging for a network trained from scratch to identify a suitable gradient direction for all tasks during initial training, which often results in slow convergence or convergence to suboptimal solutions. + +Inspired by the concept of curriculum learning [34][19], a learning process that progresses from easy to difficult tasks is more beneficial for optimizing complex tasks. MIO-IR [13] employs a sequential learning strategy to incrementally add tasks, thereby mitigating the interference caused by joint training. However, while extending from single-task to multitask learning is a form of curriculum learning, the relationships among tasks remain ambiguous, making it challenging to establish a generalized strategy for determining the optimal task sequence. Therefore, we revisit how to design a more general curriculum learning strategy for UHD all-in-one tasks. + +# 3.2.1. Observation + +First, we assess which processes are easier or harder for a neural network to learn during training, starting from random initialization to being capable of handling multiple degradations effectively. Inspired by the frequency principle of neural networks [23][25][37][30], we approach this problem from the perspective of the frequency domain. + +For a restoration task $\tau^t$ , we decompose it into $M$ frequency bands, with the $f$ -th frequency band denoted as $\tau_f^t$ . The multi-degradation restoration process is represented as $F(\tau_f^t;\theta)_{t\in [1,N],f\in [1,M]}$ . In sequential learning, tasks are learned progressively from 1 to $N$ , where each stage involves restoring all frequency bands, represented as $F(\{\tau_f^t\}_{f = 1}^M;\theta)_{t = 1\to N}$ . The main drawback of this approach is the lack of strong inter-stage connections, and the relationships between all degradation types are only modeled in the final stage, which hinders the network's ability to comprehensively learn the interplay between different degradations. + +To address this, we shift from task-based to frequency band-based progressive learning, represented as $F(\{\tau_f^t\}_{t=1}^N; \theta)_{f=1 \to M}$ . Each frequency band captures global image information and maintains a physical relationship with other bands. Learning by frequency bands is smoother and more natural, as knowledge from earlier bands supports learning subsequent bands. + +To analyze the differences between degradations under low- and high-frequency conditions, we conducted a toy experiment. We selected four types of degradations: low-light and haze, which predominantly affect the low-frequency bands of an image, and denoising and deblurring, which mainly impact the high-frequency bands. We applied a low-pass filter to decompose images into low- and high-frequency components, then trained separate models for each. To quantify the gap between different tasks, we introduced the metric $\hat{z}_{ij}^{s}$ from [43], which is used to measure the discrepancy + +![](images/f04361381cdecd0c74ec0b94828d0c107be49b16b57d7f8a45fec4282c0f08c3.jpg) +Figure 2. The overall framework of our proposed UHD-Processor, which includes: (a) Pre-training a VAE on clean images for an image reconstruction task. (b) During the image restoration task, we keep the VAE frozen and efficiently fine-tune it using an adapter, where latentIR represents the latent space image restoration network. (c) Degradation-Specific Frequency Selection Prompt (DFSP) for the decoder. (d) Degradation-Aware Low Rank Prompt (DALR) for the encoder, and (e) an illustration of the Frequency Domain Decoupling Progressive Learning Strategy. + +![](images/ff5635cbce2670a008074d46ed9c56c733d8478887a37414bb97637eeb3da48e.jpg) + +between pre-training and fine-tuning tasks (higher values indicate smaller differences). For a detailed discussion of this metric, please refer to the supplementary material. + +Experimental results in Figure 1(d) indicate that, for both low and high frequencies, the value of $\hat{z}_{ij}^{s}$ is consistently higher compared to the entire frequency spectrum. This underscores the critical role of frequency-wise decomposition in minimizing task-specific discrepancies within individual frequency bands. Moreover, $\hat{z}_{ij}^{s}$ is significantly higher in the low-frequency band than in the high-frequency band, suggesting that a progressive learning sequence from low to high frequencies is better suited for establishing an effective curriculum learning framework. + +To further investigate the differences in conclusions between standard and UHD resolutions, we conducted experiments on UHD data for four types of degradations and their corresponding downsampled versions using an all-in-one training setup. As shown in Figure 1(c), By adopting a training sequence progressing from low to high frequencies, we observed improvements for both resolutions compared to direct training. However, the enhancement was more pro + +nounced in UHD scenarios, which can be attributed to the richer high-frequency texture information present in UHD images. + +# 3.2.2. Frequency Domain Decoupling Progressive Learning Strategy + +Therefore, we propose the Frequency Decoupling Progressive Learning strategy, which leverages multiple wavelet transformation (MWT) to decompose the input image $X$ and ground truth $\Upsilon$ into multiple frequency sub-bands $\{X_f\}_{f=1}^M$ and $\{Y_f\}_{f=1}^M$ . + +During training, the learning process progressively advances from low to high frequencies, with each frequency band $X_{f}$ managed by a dedicated restoration sub-network $F_{f}$ . As training progresses and higher frequency bands are introduced, additional restoration sub-networks are incrementally incorporated, forming a divide-and-conquer approach to effectively handle each frequency band. The process can be formulated as Algorithm 1. This pseudocode omits the detailed steps of multiple wavelet transformation (MWT) and its corresponding inverse process (MIWT). For + +a more comprehensive description of the algorithm, please refer to the supplementary materials. + +Algorithm 1 Frequency Domain Decoupling Progressive Learning + +1: Input: $\{X_f\}_{f=1}^M, \{Y_f\}_{f=1}^M \gets \mathrm{MWT}(X), \mathrm{MWT}(Y)$ +2: Init: Initialize $\{\theta_f\}_{f=1}^M$ for sub-networks, set stage thresholds $\{T_k\}_{k=1}^M$ , and define total_iter +3: current_iter $\leftarrow 0, k \leftarrow 1$ +4: while current_iter $\leq$ total_iter do +5: $\{\hat{X}_f\mid f = 1,\dots ,k\} \gets F_f(X_f;\theta_f)$ +6: $\dot{X}\gets \mathrm{MIWT}(\{\dot{X}_f\}_{f = 1}^k,\{Y_f\}_{f = k + 1}^M)$ +7: $\mathcal{L}\gets \operatorname {Loss}(\hat{X},Y)$ +8: $\{\theta_f\}_{f = 1}^k\gets \{\theta_f\}_{f = 1}^k -\alpha \nabla_{\{\theta_f\}_{f = 1}^k}\mathcal{L}(\tilde{X},Y)$ +9: current_iter $\leftarrow$ current_iter + 1 +10: if current_iter $>T_{k}$ and $k < M$ then +11: $k\gets k + 1$ +12: end if +13: end while + +# 3.3. Efficient Adaptive Prompt Learning + +Based on our VAE-based restoration framework, we design two different prompt strategies for the distinct behaviors of the VAE encoding and decoding processes. For the encoding process, since the VAE is pre-trained on clean images, directly encoding various degraded images introduces a significant domain gap. Therefore, we introduce a low-rank degradation-aware prompt to the encoder, enabling it to encode different degraded images into the pre-trained compact and regularized latent space, thereby reducing the difficulty of the subsequent restoration mapping. + +During the decoding stage, despite the successful mitigation of degradation effects within the latent space, challenges persist with the loss of high-frequency details when decoding back to pixel space. To address this issue, we incorporate an adapter to inject high-frequency information from the corresponding stages of the encoder into the decoder. However, since different degradations impact specific frequency bands differently, we employ a frequency-selective prompt in the decoding process. This approach facilitates an adaptive fusion of high-frequency details from the encoder with the low-frequency information from the decoder, tailored to the specific frequency bands affected by the degradation. + +Additionally, to allow the network to effectively recognize the type of degradation, we pre-train a degradation classifier. Since this classification task is not particularly challenging, a simple classification network is sufficient to achieve high accuracy. + +# 3.3.1. Degradation-Aware Low Rank Prompt + +The current methods for prompt injection in all-in-one tasks primarily follow the design of PromptIR [20]. In this frame + +work, for the input feature $F_{i}\in \mathbb{R}^{H\times W\times C}$ , a learnable prompt tensor $P_{c}\in \mathbb{R}^{N\times \hat{H}\times \hat{W}\times C}$ is first constructed, where $N$ denotes the number of prompt components, and $\hat{H}$ and $\hat{W}$ represent the spatial dimensions, while the channel count $C$ aligns with that of the input. Subsequently, $P_{c}$ is interpolated to match the size of $F_{i}$ , enabling further interaction to generate the prompts. + +While this design has shown promising performance in tasks at standard resolutions, it faces two significant challenges in UHD scenarios. First, the entire learnable tensor $P_{c}$ comprises a substantial number of parameters due to the dimensions $\hat{H} \times \hat{W}$ . Furthermore, the prompt is homogeneous across all degradation types, which means its complexity cannot be adapted based on the difficulty of each specific degradation. This lack of customization makes the approach inefficient for UHD image restoration, where the difficulty of different degradation categories can vary significantly. + +Second, since the dimensions $\hat{H}$ and $\hat{W}$ of $P_{c}$ must be predetermined and interpolated to ensure adaptability to feature resolution, traditional all-in-one tasks exhibit minimal differences between training and inference resolutions. This allows the trained $P_{c}$ to generalize well during inference. However, in UHD tasks, training typically uses a patch size of $512 \times 512$ for efficiency, while inference requires processing full images at $3840 \times 2160$ . This significant resolution gap limits the generalization of $P_{c}$ during inference. + +To address the issue of parameter efficiency, inspired by the low-rank adaptation method [9], we perform a low-rank decomposition on the spatial dimensions of the original parameter matrix $P_{c}\in \mathbb{R}^{N\times \hat{H}\times \hat{W}\times C}$ , represented as: + +$$ +P _ {c l} = \operatorname {c a t} \left[ B _ {n} A _ {n} \right] _ {n \in [ 1 \dots N ]}, \tag {1} +$$ + +where $B_{n}\in \mathbb{R}^{1\times \hat{H}\times r\times C}$ and $A_{n}\in \mathbb{R}^{1\times r\times \hat{W}\times C}$ represent the matrices resulting from the low-rank decomposition, and $\operatorname {cat}[\cdot ]$ represents concatenation along the first dimension, with $r\ll \min (\hat{H},\hat{W})$ . After low-rank decomposition, the original parameter matrix is effectively compressed. + +Furthermore, for different types of degradations, we can specify different ranks $r$ . More complex degradations can use a larger $r$ to enhance the parameter space, while simpler degradations can use a smaller $r$ to improve parameter efficiency. The use of different ranks $r$ introduces unique designs for various degradations, while the multiplication of the decomposed matrices preserves consistent dimensions, allowing efficient parallel processing through concatenation. + +To address the issue of resolution adaptability, after obtaining the low-rank prompt tensor $P_{cl}$ , we do not directly interpolate it to match the input features. Instead, considering the scale invariance of convolution operations, we utilize $P_{cl}$ as a parameter space to learn multiple convolution operators. By using convolution as the core operator for prompt interactions, this approach is more effective in adapting to + +the differences between inference and training resolutions compared to simple interpolation. + +Specifically, the prompt generation process during the encoding phase is as follows. First, the input image is passed through a classifier to obtain $N + 1$ classification scores, where $N$ represents the number of degradation types, and 1 denotes the clean image. We then further process these classification scores to derive $W_{1}$ and $W_{2}$ : + +$$ +W _ {1} = \operatorname {s o f t m a x} (\sum (N), 1), \tag {2} +$$ + +$$ +W _ {2} = \operatorname {s o f t m a x} (N), +$$ + +where, $W_{1}$ represents the binary classification score between degraded and clean images, serving to characterize the degree of degradation, while $W_{2}$ corresponds to the $N$ -way classification scores for the $N$ different degradation types, indicating the specific type of degradation present. + +We then construct two simple pointwise convolutions to provide degradation degree prompts(DDP), as illustrated below: + +$$ +F _ {i 1} ^ {\prime} = \operatorname {c a t} \left[ P _ {\text {d e g}} \left(F _ {i}\right), P _ {\text {c l e a n}} \left(F _ {i}\right) \right], \tag {3} +$$ + +$$ +F _ {i 1} = F _ {i 1} ^ {\prime} \odot W _ {1}, +$$ + +where, $P_{deg}(\cdot)$ and $P_{clean}(\cdot)$ denote the pointwise convolutions used to provide degradation degree prompts, while $\odot$ represents the weighted summation of their outputs. + +After obtaining the degradation-aware feature $F_{i1}$ , we proceed to provide prompts for degradation types(DTP). First, we define the rank of the parameter space for each degradation type as $R_{N} = [r_{1}, r_{2}, \dots, r_{N}]$ , which allows us to derive the parameter space $P_{cl}$ based on Equation (1). Next, we utilize a simple MLP to generate a set of deep-wise convolution kernels $K_{s} = [(k_{1}:d_{1}), (k_{2}:d_{2}), \dots, (k_{m}:d_{m})]$ , characterized by varying kernel sizes and dilation rates. + +The feature $F_{i1}$ is then processed in parallel through all multi-scale convolution group (MCKG), followed by a weighted summation with $W_{2}$ . Due to the impact of parallel multi-branch convolutional kernels on inference efficiency [6], we employ the structural reparameterization method from [7]. In the inference phase, all convolutional kernels are reparameterized and merged into a single large convolutional kernel. This process can be expressed as follows: + +$$ +K _ {s} = \mathrm {M L P} _ {i \in [ 1: m ]} (P _ {c l}), +$$ + +$$ +F _ {i 2} ^ {\prime} = \left\{ \begin{array}{l l} \sum \left(K _ {s} \left(F _ {i 1}\right)\right) & \text {i f i n t r a i n i n g p h a s e}, \\ \operatorname {R e p} \left[ K _ {s} \right] \left(F _ {i 1}\right) & \text {i f i n t e s t i n g p h a s e}, \end{array} \right. \tag {4} +$$ + +$$ +F _ {i 2} = W _ {2} \odot F _ {i 2} ^ {\prime}, +$$ + +where, $\mathrm{Rep}[\cdot ]$ refers to the reparameterization operation. The resultant $F_{i2}$ serves as the low-rank degradation-aware prompt, which is integrated into the encoder's adapter. This mechanism facilitates the encoder's ability to encode diverse degradations into a cohesive and compact latent space. + +# 3.3.2. Degradation-Specific Frequency Selection Prompt + +Given that different degradations affect distinct frequency bands [4] [11], we design prompts in the decoding phase to facilitate adaptive selection of frequency components. Specifically, this involves integrating medium-frequency information $\{F_E^{HL}, F_E^{LH}\}$ and high-frequency information $F_E^{HH}$ from the encoder, along with low-frequency information $F_D^{LL}$ from the decoder. + +We begin by applying the Fast Fourier Transform (FFT) to each frequency band feature in the adapter branch, represented as $F_{w} = \{F_{D}^{LL}, F_{E}^{HL}, F_{E}^{LH}, F_{E}^{HH}\} \in \mathbb{R}^{B \times 4 \times \frac{\hat{H}}{2} \times \frac{\hat{W}}{2} \times C}$ . Pointwise convolution is applied to each band to efficiently capture global features and expand the channels by a factor of $N$ , resulting in $F_{w1} \in \mathbb{R}^{B \times 4 \times N \times \frac{\hat{H}}{2} \times \frac{\hat{W}}{2} \times C}$ after reshaping. A weighted summation with $W_{2}$ is then performed along the $N$ dimension, followed by the Inverse FFT (IFFT) to obtain $F_{w1}'$ : + +$$ +\begin{array}{l} F _ {w 1} = \operatorname {c a t} \left[ \operatorname {P c o n v} _ {\mathrm {i} \in [ \text {l o w}, \text {m i d}, \text {h i g h} ]} \left(\mathcal {F} \left(F _ {w}\right)\right) \right], \\ F ^ {\prime} = \mathcal {F} ^ {- 1} \left(W, W, F _ {w}\right) \end{array} \tag {5} +$$ + +$$ +F _ {w 1} ^ {\prime} = \mathcal {F} ^ {- 1} \left(W _ {2} \odot F _ {w 1}\right), +$$ + +where, $\mathrm{Pconv}_{\mathrm{i}\in [\mathrm{low},\mathrm{mid},\mathrm{high}]}$ denotes the pointwise convolutions used to process the low, mid, and high frequency bands, while $\mathcal{F}(\cdot)$ and $\mathcal{F}^{-1}(\cdot)$ represent the spatial dimensional Fourier transform and its corresponding inverse transform, respectively. + +Subsequently, pixel shuffle is applied for frequency mixing, followed by a sigmoid activation to obtain the filter kernel $F_{w2} \in \mathbb{R}^{B \times \hat{H} \times \hat{W} \times C}$ . For the input $F_{d} \in \mathbb{R}^{B \times \hat{H} \times \hat{W} \times C}$ from the decoder, we perform FFT to shift to the frequency domain and apply $F_{w2}$ for frequency domain filtering. Finally, the IFFT is conducted to obtain the frequency-selective prompt $F_{w2}'$ : + +$$ +F _ {w 2} = \operatorname {s i g m o i d} \left(\operatorname {p i x e l s h u f f u l e} \left(F _ {w 1} ^ {\prime}\right)\right), +$$ + +$$ +F _ {w 2} ^ {\prime} = \mathcal {F} ^ {- 1} \left(F _ {w 2} \otimes \mathcal {F} \left(F _ {w 2}\right)\right), \tag {6} +$$ + +where $\otimes$ denotes element-wise multiplication. $F_{w2}^{\prime}$ serves as a frequency-selective prompt injected through the adapter in the decoder, enabling the decoding process to perform a more customized multi-band feature fusion tailored to the frequency characteristics of the degradation. + +# 4. Experiments + +# 4.1. Experimental Setup + +We have developed a comprehensive benchmark for evaluating UHD all-in-one restoration performance, based on datasets from UHD-LL [15], UHD-blur [5], UHD-haze [46], UHD-rain [2], and UHD-haze [31]. Additionally, we constructed a denoising dataset, UHD-noise, using 4K images from [42] as the background. For additional implementation details and dataset descriptions, please refer to the supplementary materials. + +![](images/989b7689d88e816a6cb0f407cdca5e687ce76acd91ed16c549a77525ac55ef51.jpg) +Image deblurring + +![](images/659e7328f41190ef9aa1846f105e5a4ea7f170fc415b6bbf6cb98e12d9a2c993.jpg) +Histformer + +![](images/ac1237d54e30069b3408419500440123ba0fadb2dcc8bbd97262a59bc0fc6022.jpg) +PromptIR + +![](images/4ef8809299c46ac5b60fda594b974ed56d679d4567f35ac0de24b3c15049cca7.jpg) +GT + +![](images/e025b82451bec369a8468f046e4eed96acae49e459d8d1a00fc99cb83972372a.jpg) +Gridformer + +![](images/3820b6dc4862582190abe227faa866e5fd3e6853288b7399e93abd783d5c8264.jpg) +HAIR + +![](images/4e4826a8146fa398f18be36ff90e29f8f63d20e8aa8cf18f43b2215ddf5d85e6.jpg) +Ours + +![](images/57287db71fd0c582a97ad0048c81dd9b229d8a7d75762b03bc5c760525202c5b.jpg) +Low light enhancement +Figure 3. A comparison of visual results for four types of degradation removal with other state-of-the-art (SOTA) all-in-one methods. + +![](images/4038e26fadfb1aeea679d300e3b75c0d8919f2a9fb71de580f1d61fcf21ea990.jpg) + +![](images/0414d68c91ff50d09bb4c5fd1bafab0831d3e5d3585462d3d7a56d1efc12f4ff.jpg) + +![](images/86b37eaf6ee2b186d50449a933fac29c005824c8c73f25511d0df54b2b43ef43.jpg) + +![](images/d98c9d0f5e6a442d25a6b67bdb9239efbb2454864c9efde931857f9eacf69248.jpg) +Histformer +Gridformer + +![](images/a5d6ac59df501d7b9aad69041876c4697605528f6504e6c44b9da7ef325b3ffa.jpg) +PromptIR +HAIR + +![](images/c0c7fa20af61602243b7608b5d6ffe6bf4d1e5df5f80e8ff3cafc5cb5b4f4f3f.jpg) +GT +Ours + +![](images/fd7940648b1c40934a6c388da3a088df134a594435e62365d11e4c933bcb7b4e.jpg) +Image denoising + +![](images/2e54fba6b683998791ccb45d7addac44694a3e6c6ebd6d06b0358e9a05479d39.jpg) + +![](images/f0f9cd3a714accb1c416174a91a640a43a3184d482e0e71453c25f96b6041a0f.jpg) +PromptIR + +![](images/699f9c65003e99886c92bc71e0236fe76039d87e11c0a74e1dd74acbeb6d923b.jpg) + +![](images/3577901648ab33fd82bed731f716ca9e461ebf3ee5e8a1d2c8a9e998ccae6f3d.jpg) +Histformer + +![](images/84d0423fbc73d628b6fea574dfef7dfd36de2a734837364044078a42a10ecbe8.jpg) + +![](images/df0c1d6c3ec1be1c93c778315ded93ff3c9ae8e0d43509c9eca8b59b32e1ff3e.jpg) +GT + +![](images/8a7b2b617fe2b52c4301a8a8e2b001bef24bf141f3307bf83387682928cf8de0.jpg) +Image dehazing + +![](images/fac6ddff917d8339bf314872342a0d65790262bc3aae1d839f2c826da2662236.jpg) +Gridformer + +![](images/c10c595537485a98cc87cbb5c7de71e9faf68bff6e8c678b26c201cd408d8d10.jpg) +HAIR + +![](images/fa7002e0da7863c87e766d3c03e621cf9cccca59437bb48417a59b7786ce4a5e.jpg) +Ours + +![](images/5177262119abe129d47c65b526ac3485076793234eb3c1d6e53ba4c75600d472.jpg) +Histformer +Gridformer + +![](images/4c554c03c9650b99b6d0c47fd1dbd752e37b7bbc0cac734bbc79c460a0a21b16.jpg) +PromptIR +HAIR + +![](images/42ccda51ace6b43f5c5d22951878bc907a035daf5ffc7610b8884b0bbda677e2.jpg) +GT +Ours + +Table 1. Comparison to state-of-the-art on four degradations. PSNR (dB, ↑), SSIM (↑), and LPIPS (↓), and FS represents full-size 4K image inference. FLOPs are computed for an input size of $256 \times 256$ . Best and second best performances are highlighted. + +
MethodFSFLOPsParams.Low LightDeblurringDehazingDenoisingAverage
UHD-LLUHD-blurUHD-haze\( UHDN_{\sigma=15} \)\( UHDN_{\sigma=25} \)\( UHDN_{\sigma=50} \)
AIRNet [14]✘301G9M19.24.80921.89.75718.37.81221.33.88720.78.78418.79.47520.07.754.2843
IDR [41]✘88G15.3M23.12.91024.67.79319.12.76827.48.91225.86.87224.57.65424.14.822.2684
PromptIR [20]✘158G33M23.44.90225.77.78219.97.72728.43.92426.74.89823.72.58424.68.803.2571
CAPTNet [8]✘25G24.3M23.96.92026.11.79819.46.86825.58.86523.24.88421.98.50823.39.809.3466
NDR-Restore [38]✘196G36.9M23.84.89424.25.80220.08.89225.62.91224.37.89722.94.66923.52.846.3126
Gridformer [33]✘367G34M23.12.89825.82.78319.24.86936.04.93731.72.89826.24.62327.03.836.3754
DiffUIR-L [44]✘10G36.2M21.56.81223.85.74318.28.86436.84.93832.42.89726.08.64826.51.818.2564
Histoformer [27]✘91G16.6M23.22.90825.62.78219.78.90326.88.84525.64.87423.13.65924.04.829.3524
adalR [4]✘147G28.7M23.57.91626.35.80118.44.90132.84.92130.48.90126.48.67226.36.857.3429
HAIR [1]✘41G29M25.75.92225.78.79820.00.89435.54.91630.84.87826.26.65727.36.847.2822
Ours✓4G1.6M27.11.92526.48.80320.94.92338.94.97533.99.90327.95.67729.23.868.2541
+ +Table 2. Comparison to state-of-the-art on six degradations. PSNR (dB, ↑), SSIM (↑), LPIPS (↓) and FS represents full-size 4K image inference. FLOPs are computed for an input size of $256 \times 256$ . Best and second best performances are highlighted. + +
MethodFSFLOPsParams.Low LightDeblurringDehazingDenoisingDerainingDesnowingAverage
UHD-LLUHD-blurUHD-hazeUHDN\(_{\sigma=50}\)UHD-rainUHD-snow
AIRNet [14]✘301G9M22.68.88723.52.87618.24.84622.38.87626.35.87627.38.92423.43.874.1861
IDR [41]✘88G15.3M24.33.91525.64.78818.68.87929.64.90628.82.90630.48.94526.27.890.1912
PromptIR [20]✘158G33M23.3.91126.48.80520.14.90124.88.83528.89.89730.78.96625.74.886.2155
CAPTNet [8]✘25G24.3M24.97.92126.32.79620.32.90321.64.56929.34.90832.21.97425.80.845.2861
NDR-Restore [38]✘196G36.9M25.12.88525.64.79119.21.89631.44.91529.24.89728.41.94826.51.889.3108
Gridformer [33]✘367G34M23.92.89825.68.78218.87.88932.86.91529.37.90428.24.94226.49.895.2321
DiffUIR-L [44]✘10G36.2M22.64.90225.08.78518.62.88933.25.92827.89.88627.36.94525.81.889.1844
Histoformer [27]✘91G16.6M25.73.91526.55.79618.73.89733.05.92427.96.88427.56.97126.59.898.1855
adalR [4]✘147G28.7M23.84.91826.86.80319.34.91032.46.92328.18.90127.72.95326.40.901.2492
HAIR [1]✘41G29M25.22.89724.77.79918.75.88332.50.91528.76.89327.89.96826.31.892.2607
Ours✓4G1.6M26.91.92426.95.80721.81.93133.73.93429.90.91532.73.97928.67.915.1839
+ +Table 3. Comparison to state-of-the-art for single degradations. PSNR (dB, ↑), SSIM (↑) and LPIPS (↓). Best and second best performances are highlighted. +(a) Low light +(b) Deblurring +(c) Dehazing + +
MethodFSUHD-LLLIPIPSMethodFSUHD-blurLIPIPSMethodFSUHD-hazeLIPIPS
Restormer [40]✘21.54.843.3608Restormer [40]✘25.21.752.3695Restormer [40]✘12.72.693.4560
LLformer [32]✘24.06.858.3516Uformer [35]✘25.26.751.3851Uformer [35]✘19.83.921.4220
UHDFour [15]✓26.22.900.2390Stripformer [28]✘25.05.750.3740DehazeFormer [26]✘15.37.737.3998
UHDFformer [29]✓27.11.927.2240FFTformer [12]✘25.41.757.3708TaylorFormer [22]✘20.99.919.3124
LMAR [39]✓26.27.919.2248UHDformer [29]✓28.82.844.2350UHDformer [29]✓22.58.942.1188
AirNet [14]✘24.61.827.3841AirNet✘23.48.824.3942AirNet✘18.44.848.4018
PromptIR [20]✘24.81.912.3116PromptIR✘23.24.829.3754PromptIR✘20.21.927.3642
Histoformer [27]✘25.46.894.3213Histoformer✘27.12.848.2986Histoformer✘21.72.948.2942
HAIR [1]✘26.08.901.2964HAIR✘26.74.847.3867HAIR✘20.38.912.2412
Ours✓27.22.929.2042Ours✓28.91.851.2129Ours✓23.24.953.1086
+ +# 4.2. Comparison with State-of-the-Art Methods + +# 4.2.1. All-in-One: Four Degradations + +We compare our UHD-Processor with specialized all-in-one restoration methods, trained on four degradation types: dehazing, denoising, deblurring, and low-light enhancement. As shown in Tab. 1, our method achieves state-of-the-art (SOTA) performance on both individual datasets and overall metrics. Specifically, the average PSNR improves by 1.93 dB over the second-best method, HAIR, while our method requires only $5.5\%$ of the parameters and $9.7\%$ of the FLOPS of Histoformer. Visual results, shown in Figure 3, further demonstrate that our approach effectively achieves the best degradation removal. + +# 4.2.2. All-in-One: Six Degradations + +We further extended the experiment to include six degradation types by introducing image desnow and image derain, to validate the effectiveness of our method in a more complex all-in-one setup. As shown in Tab. 2, our method continues to achieve SOTA performance on both individual datasets and average metrics. Specifically, the average PSNR improves by $1.7\mathrm{dB}$ over the second-best method, Histoformer, while the number of parameters and FLOPS are only $9.6\%$ and $4.3\%$ respectively. More visual comparison results can be found in the supplementary materials. + +# 4.2.3. Single-Degradation Results + +We further evaluate the performance of our method by training instances under a single degradation protocol, comparing them with both task-specific methods and all-in-one approaches. The results, shown in Tab. 3, demonstrate that our method still achieves the best performance, validating the effectiveness of our approach. + +# 4.3. Ablation Study + +# 4.3.1. Frequency-Progressive Learning Strategy + +As shown in Tab. 4, we validate the effectiveness of the proposed Frequency Domain Progressive Learning (FDPL). First, compared to directly training on mixed data, using FDPL results in a significant improvement of $1.99\mathrm{dB}$ PSNR. Next, we evaluate the sequential learning (SL) strategy proposed in [13], which advocates for learning local degradations before global ones. We test two learning orders: SL1 (noise $\rightarrow$ blur $\rightarrow$ low-light $\rightarrow$ haze) and SL2 (blur $\rightarrow$ noise $\rightarrow$ haze $\rightarrow$ low-light). The experimental results show that not all local-to-global learning orders lead to performance gains. In fact, SL1 even caused some performance degradation, while SL2 yielded better results. This suggests that sequential learning strategies are sensitive to the order of tasks, and the optimal sequence may not only depend on the tasks themselves but also on the specific data set, scene, or degradation severity. Therefore, it is challenging to define a universally effective task order. In contrast, our + +Table 4. Ablation Study of the Learning Strategy. + +
MethodMixw/SL1w/SL2w/FDPL (Ours)
PSNR27.2427.0828.1229.23
SSIM0.8530.8460.8720.874
+ +Table 5. Ablation Study of Efficient Adaptive Prompt Learning. + +
Methodw/o DALPw/o DDPw/o DTPw/o DSRw/o MCKGw/o DFSPOurs
PSNR27.8628.4928.3428.8428.7228.2129.23
SSIM0.8620.8660.8580.8640.8620.8650.874
+ +FDPL strategy is task-agnostic and offers a more generalizable learning framework. Experimental results confirm that FDPL leads to more significant performance improvements. Further discussions on sequential learning can be found in the supplementary materials. + +# 4.3.2. Efficient Adaptive Prompt Learning + +The ablation study of the proposed prompt learning method, as shown in Tab. 5, demonstrates that both the DALP used in the encoding process and the DFSP used in the decoding process lead to significant performance improvements. We further analyze the components of DALP, and the results show that both the Degradation Degree Prompt (DDP) and the Degradation Type Prompt (DTP) provide more granular prompt information, improving performance consistently. The Degradation-Specific Rank (DSR) component introduces task-specific low-rank decompositions, allowing for the selection of an appropriate rank for each task, which enhances parameter efficiency while improving performance. Moreover, the Multi-Scale Convolution Kernel Group (MCKG) introduces a multi-scale convolutional branch, which strengthens the representation capacity of the prompt generation process. During inference, the multi-branch convolutions are reparameterized for efficient execution. For more detailed ablation experiments and efficiency analyses, please refer to the supplementary materials. + +# 5. Conclusion + +In this work, we proposed UHD-Processor, an efficient and effective framework for UHD all-in-one image restoration. By leveraging Frequency Domain Decoupling Progressive Learning Strategy and an efficient prompt-integrated VAE framework, our method addresses the unique challenges of UHD restoration tasks, including denoising, deblurring, deraining, dehazing, and low-light enhancement. Experimental results demonstrated that our approach significantly outperforms existing methods, achieving state-of-the-art results with reduced computational costs. The combination of progressive frequency learning and targeted prompt-based adaptation provides a strong foundation for future advancements in UHD image restoration. + +# Acknowledgement + +This work was supported by the National Natural Science Foundation of China (NSFC) under Grants 62225207, 62436008, 62422609 and 62276243. + +# References + +[1] Jin Cao, Yi Cao, Li Pang, Deyu Meng, and Xiangyong Cao. Hair: Hypernetworks-based all-in-one image restoration, 2024. 7 +[2] Hongming Chen, Xiang Chen, Chen Wu, Zhuoran Zheng, Jinshan Pan, and Xianping Fu. Towards ultra-high-definition image deraining: A benchmark and an efficient method, 2024. 1, 6 +[3] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1 +[4] Yuning Cui, Syed Waqas Zamir, Salman Khan, Alois Knoll, Mubarak Shah, and Fahad Shahbaz Khan. Adair: Adaptive all-in-one image restoration via frequency mining and modulation, 2024. 6, 7 +[5] Senyou Deng, Wenqi Ren, Yanyang Yan, Tao Wang, Fenglong Song, and Xiaochun Cao. Multi-scale separable network for ultra-high-definition video deblurring. In the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14030-14039, 2021. 6 +[6] Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13733-13742, 2021. 6 +[7] Xiaohan Ding, Yiyuan Zhang, Yixiao Ge, Sijie Zhao, Lin Song, Xiangyu Yue, and Ying Shan. Unireplknet: A universal perception large-kernel convnet for audio video point cloud time-series and image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5513-5524, 2024. 6 +[8] Hu Gao, Jing Yang, Ning Wang, Jingfan Yang, Ying Zhang, and Depeng Dang. Prompt-based all-in-one image restoration using cnns and transformer. arXiv preprint arXiv:2309.03063, 2023. 3, 7 +[9] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 5 +[10] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 2 +[11] Xingyu Jiang, Xiuhui Zhang, Ning Gao, and Yue Deng. When fast fourier transform meets transformer for image restoration. 6 +[12] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In 2023 IEEE/CVF + +Conference on Computer Vision and Pattern Recognition (CVPR), pages 5886-5895, 2023. 7 +[13] Xiangtao Kong, Chao Dong, and Lei Zhang. Towards effective multiple-in-one image restoration: A sequential and prompt learning strategy. arXiv preprint arXiv:2401.03379, 2024. 3, 8 +[14] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-In-One Image Restoration for Unknown Corruption. In IEEE Conference on Computer Vision and Pattern Recognition, New Orleans, LA, 2022. 2, 7 +[15] Chongyi Li, Chun-Le Guo, Man Zhou, Zhexin Liang, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Embedding fourier for ultra-high-definition low-light image enhancement. In ICLR, 2023. 1, 2, 6, 7 +[16] Xiangtai Li, Haobo Yuan, Wei Li, Henghui Ding, Size Wu, Wenwei Zhang, Yining Li, Kai Chen, and Chen Change Loy. Omg-seg: Is one model good enough for all segmentation? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27948-27959, 2024. 1 +[17] Zilong Li, Yiming Lei, Chenglong Ma, Junping Zhang, and Hongming Shan. Prompt-in-prompt learning for universal image restoration, 2023. 3 +[18] Ziwei Luo, Fredrik K. Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B. Schön. Controlling vision-language models for multi-task image restoration. In The Twelfth International Conference on Learning Representations, 2024. 3 +[19] Anastasia Pentina, Viktoriia Sharmanska, and Christoph H Lampert. Curriculum learning of multiple tasks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5492-5500, 2015. 3 +[20] Vaishnav Potlapalli, Syed Waqas Zamir, Salman Khan, and Fahad Khan. Prompt: Prompting for all-in-one image restoration. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2, 5, 7 +[21] Yifan Pu, Weicong Liang, Yiduo Hao, Yuhui Yuan, Yukang Yang, Chao Zhang, Han Hu, and Gao Huang. Rank-detr for high quality object detection. Advances in Neural Information Processing Systems, 36, 2024. 1 +[22] Yuwei Qiu, Kaihao Zhang, Chenxi Wang, Wenhan Luo, Hongdong Li, and Zhi Jin. Mb-taylorformer: Multi-branch efficient transformer expanded by taylor formula for image dehazing, 2023. 7 +[23] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International conference on machine learning, pages 5301-5310. PMLR, 2019. 3 +[24] Wenqi Ren, Senyou Deng, Kaihao Zhang, Fenglong Song, Xiaochun Cao, and Ming-Hsuan Yang. Fast ultra high-definition video deblurring via multi-scale separable network. International Journal of Computer Vision, 132(5):1817-1834, 2024. 2 +[25] Basri Ronen, David Jacobs, Yoni Kasten, and Shira Kritchman. The convergence rate of neural networks for learned functions of different frequencies. Advances in Neural Information Processing Systems, 32, 2019. 3 + +[26] Yuda Song, Zhuqing He, Hui Qian, and Xin Du. Vision transformers for single image dehazing. IEEE Transactions on Image Processing, 32:1927-1941, 2023. 7 +[27] Shangquan Sun, Wenqi Ren, Xinwei Gao, Rui Wang, and Xiaochun Cao. Restoring images in adverse weather conditions via histogram transformer. In European Conference on Computer Vision, pages 111-129. Springer, 2025. 7 +[28] Fu-Jen Tsai, Yan-Tsung Peng, Yen-Yu Lin, Chung-Chi Tsai, and Chia-Wen Lin. Stripformer: Strip transformer for fast image deblurring. In European conference on computer vision, pages 146-162. Springer, 2022. 7 +[29] Cong Wang, Jinshan Pan, Wei Wang, Gang Fu, Siyuan Liang, Mengzhu Wang, Xiao-Ming Wu, and Jun Liu. Correlation matching transformation transformers for uhd image restoration. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5336-5344, 2024. 1, 2, 7 +[30] Haohan Wang, Xindi Wu, Zeyi Huang, and Eric P Xing. High-frequency component helps explain the generalization of convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8684-8694, 2020. 3 +[31] Liyan Wang, Cong Wang, Jinshan Pan, Xiaofeng Liu, Weixiang Zhou, Xiaoran Sun, Wei Wang, and Zhixun Su. Ultra-high-definition image restoration: New benchmarks and a dual interaction prior-driven solution, 2024. 2, 6 +[32] Tao Wang, Kaihao Zhang, Tianrun Shen, Wenhan Luo, Bjorn Stenger, and Tong Lu. Ultra-high-definition low-light image enhancement: A benchmark and transformer-based method. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2654-2662, 2023. 7 +[33] Tao Wang, Kaihao Zhang, Ziqian Shao, Wenhan Luo, Bjorn Stenger, Tong Lu, Tae-Kyun Kim, Wei Liu, and Hongdong Li. Gridformer: Residual dense transformer with grid structure for image restoration in adverse weather conditions. International Journal of Computer Vision, pages 1-23, 2024. 2, 7 +[34] Xin Wang, Yudong Chen, and Wenwu Zhu. A survey on curriculum learning. IEEE transactions on pattern analysis and machine intelligence, 44(9):4555-4576, 2021. 2, 3 +[35] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17683-17693, 2022. 7 +[36] Boxue Xiao, Zhuoran Zheng, Xiang Chen, Chen Lv, Yunliang Zhuang, and Tao Wang. Single uhd image dehazing via interpretable pyramid network, 2022. 1 +[37] Zhi-Qin John Xu, Yaoyu Zhang, and Yanyang Xiao. Training behavior of deep neural network in frequency domain. In Neural Information Processing: 26th International Conference, ICONIP 2019, Sydney, NSW, Australia, December 12–15, 2019, Proceedings, Part I 26, pages 264–274. Springer, 2019. 3 +[38] Mingde Yao, Ruikang Xu, Yuanshen Guan, Jie Huang, and Zhiwei Xiong. Neural degradation representation learning for all-in-one image restoration. IEEE Transactions on Image Processing, 2024. 2, 7 + +[39] Wei Yu, Jie Huang, Bing Li, Kaiwen Zheng, Qi Zhu, Man Zhou, and Feng Zhao. Empowering resampling operation for ultra-high-definition image enhancement with model-aware guidance. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25722-25731, 2024. 7 +[40] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 2, 7 +[41] Jinghao Zhang, Jie Huang, Mingde Yao, Zizheng Yang, Hu Yu, Man Zhou, and Feng Zhao. Ingredient-oriented multidegradation learning for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5825-5835, 2023. 2, 3, 7 +[42] Kaihao Zhang, Dongxu Li, Wenhan Luo, Wenqi Ren, Bjorn Stenger, Wei Liu, Hongdong Li, and Yang Ming-Hsuan. Benchmarking ultra-high-definition image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 6 +[43] Wenlong Zhang, Xiaohui Li, SHI Guangyuan, Xiangyu Chen, Yu Qiao, Xiaoyun Zhang, Xiao-Ming Wu, and Chao Dong. Real-world image super-resolution as multi-task learning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 3 +[44] Dian Zheng, Xiao-Ming Wu, Shuzhou Yang, Jian Zhang, Jian-Fang Hu, and Wei-Shi Zheng. Selective hourglass mapping for universal image restoration based on diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25445-25455, 2024. 3, 7 +[45] Zhuoran Zheng and Xiuyi Jia. Uhd image deblurring via multi-scale cubic-mixer. arXiv preprint arXiv:2206.03678, 2022. 2, 3 +[46] Zhuoran Zheng, Wenqi Ren, Xiaochun Cao, Xiaobin Hu, Tao Wang, Fenglong Song, and Xiuyi Jia. Ultra-high-definition image dehazing via multi-guided bilateral learning. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16180-16189, 2021. 6 +[47] Yanqi Zhou, Tao Lei, Hanxiao Liu, Nan Du, Yanping Huang, Vincent Zhao, Andrew M Dai, Quoc V Le, James Laudon, et al. Mixture-of-experts with expert choice routing. Advances in Neural Information Processing Systems, 35:7103-7114, 2022. 2 \ No newline at end of file diff --git a/2025/UHD-processer_ Unified UHD Image Restoration with Progressive Frequency Learning and Degradation-aware Prompts/images.zip b/2025/UHD-processer_ Unified UHD Image Restoration with Progressive Frequency Learning and Degradation-aware Prompts/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..62609ce03188ccf4d2d1a94adf5000a5ad477408 --- /dev/null +++ b/2025/UHD-processer_ Unified UHD Image Restoration with Progressive Frequency Learning and Degradation-aware Prompts/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:411cd71f3a48bfec209e3671540abd06528d01c29168e7720e9c0bd40b4833e6 +size 644590 diff --git a/2025/UHD-processer_ Unified UHD Image Restoration with Progressive Frequency Learning and Degradation-aware Prompts/layout.json b/2025/UHD-processer_ Unified UHD Image Restoration with Progressive Frequency Learning and Degradation-aware Prompts/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c77b279e6fed93a6f6cc7aebc94302a3e862043e --- /dev/null +++ b/2025/UHD-processer_ Unified UHD Image Restoration with Progressive Frequency Learning and Degradation-aware Prompts/layout.json @@ -0,0 +1,11719 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 510, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 510, + 140 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 510, + 140 + ], + "type": "text", + "content": "UHD-processor: Unified UHD Image Restoration with Progressive Frequency Learning and Degradation-aware Prompts" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 121, + 161, + 490, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 161, + 490, + 220 + ], + "spans": [ + { + "bbox": [ + 121, + 161, + 490, + 220 + ], + "type": "text", + "content": "Yidi Liu, DongLi, Xueyang Fu†, Xin Lu, Jie Huang, Zheng-jun Zha \nSchool of Information Science and Technology and MoE Key Laboratory of \nBrain-inspired Intelligent Perception and Cognition, \nUniversity of Science and Technology of China, Hefei, 230026, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 129, + 220, + 478, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 220, + 478, + 233 + ], + "spans": [ + { + "bbox": [ + 129, + 220, + 478, + 233 + ], + "type": "text", + "content": "liuyidi2023@mail.ustc.edu.cn, xyfu@ustc.edu.cn † Corresponding Author" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 260, + 200, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 260, + 200, + 273 + ], + "spans": [ + { + "bbox": [ + 151, + 260, + 200, + 273 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 286, + 297, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 297, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 297, + 597 + ], + "type": "text", + "content": "We introduce UHD-Processor, a unified and robust framework for all-in-one image restoration, which is particularly resource-efficient for Ultra-High-Definition (UHD) images. To address the limitations of traditional all-in-one methods that rely on complex restoration backbones, our strategy employs a frequency domain decoupling progressive learning technique, motivated by curriculum learning, to incrementally learn restoration mappings from low to high frequencies. This approach incorporates specialized sub-network modules to effectively tackle different frequency bands in a divide-and-conquer manner, significantly enhancing the learning capability of simpler networks. Moreover, to accommodate the high-resolution characteristics of UHD images, we developed a variational autoencoder (VAE)-based framework that reduces computational complexity by modeling a concise latent space. It integrates task-specific degradation awareness in the encoder and frequency selection in the decoder, enhancing task comprehension and generalization. Our unified model is able to handle various degradations such as denoising, deblurring, dehazing, lowlighting, etc. Experimental evaluations extensively showcase the effectiveness of our dual-strategy approach, significantly improving UHD image restoration and achieving cutting-edge performance across diverse conditions. The code will be available at https://github.com/1yd2022/UHD-processor" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 620, + 137, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 620, + 137, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 620, + 137, + 633 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "type": "text", + "content": "Ultra-high-definition (UHD) images are increasingly utilized across various fields, prompting substantial research in UHD image restoration for specific degradation types, such as low-light enhancement, deblurring, dehazing, and deraining [2, 15, 29, 36]. However, real-world applications like object detection [21], video segmentation [16], and au" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 328, + 266, + 460, + 394 + ], + "blocks": [ + { + "bbox": [ + 328, + 266, + 460, + 394 + ], + "lines": [ + { + "bbox": [ + 328, + 266, + 460, + 394 + ], + "spans": [ + { + "bbox": [ + 328, + 266, + 460, + 394 + ], + "type": "image", + "image_path": "72d4c86909474046daf42a55f096b0941a7b360703e19113aa0530755009f44b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 396, + 463, + 403 + ], + "lines": [ + { + "bbox": [ + 340, + 396, + 463, + 403 + ], + "spans": [ + { + "bbox": [ + 340, + 396, + 463, + 403 + ], + "type": "text", + "content": "(a) Comparison of PSNR with SOTA All-in-One Methods" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 465, + 269, + 548, + 331 + ], + "blocks": [ + { + "bbox": [ + 465, + 269, + 548, + 331 + ], + "lines": [ + { + "bbox": [ + 465, + 269, + 548, + 331 + ], + "spans": [ + { + "bbox": [ + 465, + 269, + 548, + 331 + ], + "type": "image", + "image_path": "f468997ca2074c2028c01d8cca464a3a0eba431f2f695ded2e63a6e228823f9e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 485, + 331, + 525, + 338 + ], + "lines": [ + { + "bbox": [ + 485, + 331, + 525, + 338 + ], + "spans": [ + { + "bbox": [ + 485, + 331, + 525, + 338 + ], + "type": "text", + "content": "(b) Observation 1" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 466, + 339, + 548, + 396 + ], + "blocks": [ + { + "bbox": [ + 466, + 339, + 548, + 396 + ], + "lines": [ + { + "bbox": [ + 466, + 339, + 548, + 396 + ], + "spans": [ + { + "bbox": [ + 466, + 339, + 548, + 396 + ], + "type": "image", + "image_path": "b8fadce7ba92206ae2a8ccd616838cb66522413d76ac01bc0b22f67c0f24756b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 485, + 396, + 525, + 403 + ], + "lines": [ + { + "bbox": [ + 485, + 396, + 525, + 403 + ], + "spans": [ + { + "bbox": [ + 485, + 396, + 525, + 403 + ], + "type": "text", + "content": "(c) Observation 2" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 326, + 404, + 470, + 482 + ], + "blocks": [ + { + "bbox": [ + 326, + 404, + 470, + 482 + ], + "lines": [ + { + "bbox": [ + 326, + 404, + 470, + 482 + ], + "spans": [ + { + "bbox": [ + 326, + 404, + 470, + 482 + ], + "type": "image", + "image_path": "046f1dc832e655b802a32dcf6eebc7f7feb05c97b3c5411637ead85369738392.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 346, + 483, + 536, + 491 + ], + "lines": [ + { + "bbox": [ + 346, + 483, + 536, + 491 + ], + "spans": [ + { + "bbox": [ + 346, + 483, + 536, + 491 + ], + "type": "text", + "content": "(d) Observation 3: The Variation of Gap Between Degradations Across Frequency Bands" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 470, + 405, + 550, + 479 + ], + "blocks": [ + { + "bbox": [ + 470, + 405, + 550, + 479 + ], + "lines": [ + { + "bbox": [ + 470, + 405, + 550, + 479 + ], + "spans": [ + { + "bbox": [ + 470, + 405, + 550, + 479 + ], + "type": "image", + "image_path": "c6efaf951ec6ec0d222cdb2f14ea555fc4b1229a514857e39b200d2a1edddfca.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 492, + 555, + 581 + ], + "lines": [ + { + "bbox": [ + 313, + 492, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 555, + 581 + ], + "type": "text", + "content": "Figure 1. (a) Comparison of PSNR between our method and the previous state-of-the-art (SOTA) all-in-one method across six degradation settings in UHD images. (b) Observe the performance gains when integrating the same promptIR prompt branch into Restormer models with varying parameter sizes. (c) Comparison of low-to-high frequency training order on UHD and regular images, showing greater effectiveness for UHD tasks. (d) Quantitative evaluation of restoration tasks across low, high, and full-frequency bands." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 592, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 555, + 652 + ], + "type": "text", + "content": "tonomous driving [3] face complex degradation patterns that vary across time and space, making reliance on single-function models impractical. This calls for a generalized model that can address multiple degradation types to support robust low-level vision tasks." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "In response to the above requirements, all-in-one(AIO) approaches aim to use a single model to manage various degradations, but they generally target standard-resolution images and encounter computational difficulties when scaled to UHD. This gap highlights the pressing need for an efficient" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "23121" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "all-in-one image restoration framework tailored to UHD images. For standard-resolution tasks [14, 20], successful approaches often include two key features, i.e., degradation-aware branches and strong network backbones. The former is implemented through techniques like prompts [10] or mixtures of experts (MoE) [47] to enhance adaptability to different degradation types, while the later contains advanced backbones like Restormer [40] to further strengthens learning and generalization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 182, + 294, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 294, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 294, + 301 + ], + "type": "text", + "content": "However, due to the high-resolution nature of UHD images, directly applying these strategies results in a significant computational burden, making full-resolution inference infeasible on consumer-grade GPUs. As shown in Figure 1(b), PromptIR extends the Restormer backbone with a prompt branch. We adjust their parameter counts by varying Restormer's module numbers, keeping the prompt branch fixed. Reducing parameters significantly lowers performance for both, meanwhile, the benefits introduced by the prompt branch also gradually diminish." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 302, + 294, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 302, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 294, + 374 + ], + "type": "text", + "content": "These results indicate that while the prompt branch enhances all-in-one task performance, a strong backbone is crucial to fully leverage these improvements. Moreover, compared to Restormer, the additional degradation-aware prompt branch introduced by PromptIR adds 6.87M parameters, which is clearly burdensome for UHD tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 376, + 294, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 294, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 294, + 424 + ], + "type": "text", + "content": "Based on the above observations, this work aims to tackle two challenges faced by traditional AIO methods in UHD scenarios: reliance on complex restoration backbones and redundancy in degradation-aware branch parameters." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 425, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 295, + 605 + ], + "type": "text", + "content": "To tackle the first challenge, at the model architecture level, we propose a VAE-based framework that transforms restoration tasks from pixel space into a compact latent space, thereby reducing the computational cost of UHD image processing. This alleviates the network's learning burden, enabling a simpler model to restore effectively. At the learning strategy level, inspired by curriculum learning [34], we develop a Frequency Domain Decoupling Progressive Learning Strategy. As shown in Figure 1(d), for UHD AIO tasks, higher frequency bands present greater restoration challenges. Consequently, we decompose the task into manageable frequency components, enabling sub-networks to progressively learn restoration mappings from low to high frequencies. This divide-and-conquer approach significantly enhances the network's adaptability and efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "To address the second challenge, we introduce efficient prompt strategies for both the VAE encoder and decoder. For encoding, we propose a low-rank degradation-aware prompt to indicate degradation type and severity. This is achieved by creating distinct low-rank parameter spaces per degradation type and learning scale-invariant convolution kernels, which adapt to UHD resolution differences between training and testing. For efficiency, we reparameterize these convolution branches during inference. To maintain efficiency," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "we reparameterize these convolution branches at inference. In decoding, we observe that degradation types exhibit distinct distributions across frequency bands. Accordingly, we introduce a frequency band selection prompt, allowing the decoder to adaptively choose between low-frequency features from the decoder and mid- or high-frequency features from the encoder based on the specific degradation type." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 157, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 553, + 239 + ], + "type": "text", + "content": "By integrating these components, we propose UHD-Processor, an efficient all-in-one image restoration framework tailored for ultra-high-definition (UHD) images. As shown in Figure 1(a), our proposed UHD-Processor achieves state-of-the-art performance while reducing parameters by " + }, + { + "bbox": [ + 313, + 157, + 553, + 239 + ], + "type": "inline_equation", + "content": "95.1\\%" + }, + { + "bbox": [ + 313, + 157, + 553, + 239 + ], + "type": "text", + "content": " and FLOPs by " + }, + { + "bbox": [ + 313, + 157, + 553, + 239 + ], + "type": "inline_equation", + "content": "97.4\\%" + }, + { + "bbox": [ + 313, + 157, + 553, + 239 + ], + "type": "text", + "content": " compared to promptIR [20]. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 243, + 553, + 385 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 315, + 243, + 553, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 243, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 315, + 243, + 553, + 277 + ], + "type": "text", + "content": "1. We introduce a VAE-based image restoration framework that operates within a compact latent space, significantly enhancing UHD image restoration efficiency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 279, + 553, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 279, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 315, + 279, + 553, + 338 + ], + "type": "text", + "content": "2. We design a frequency domain decoupling progressive learning strategy, which enables effective restoration across frequency bands from low to high frequencies through a divide-and-conquer approach, to improve adaptability to various optimization objectives." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 338, + 553, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 338, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 315, + 338, + 553, + 385 + ], + "type": "text", + "content": "3. We develop a prompt learning framework with low-rank degradation-aware prompts for encoding and frequency selection prompts for decoding, improving the network's adaptability to various degradation types." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 398, + 400, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 398, + 400, + 410 + ], + "spans": [ + { + "bbox": [ + 314, + 398, + 400, + 410 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 418, + 496, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 418, + 496, + 430 + ], + "spans": [ + { + "bbox": [ + 314, + 418, + 496, + 430 + ], + "type": "text", + "content": "2.1. Ultra-High-Definition Restoration" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 436, + 553, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 553, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 553, + 616 + ], + "type": "text", + "content": "As the demand for UHD image processing grows in imaging systems, various approaches have been proposed to restore clarity in UHD images through diverse network architectures. These include bilateral learning for image dehazing [45], multi-scale separable-patch integration networks for video deblurring [24], and Fourier embedding networks for low-light image enhancement [15]. Recently, generalized architectures for UHD image restoration have evolved to address multiple restoration tasks through unified frameworks. For instance, [29] introduces a correlation matching mechanism between high-resolution and low-resolution branches to enhance the utilization of native resolution information, while [31] incorporates gradient and normal priors into the model design. However the approach about UHD all in one is not yet fully explored." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 624, + 476, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 624, + 476, + 636 + ], + "spans": [ + { + "bbox": [ + 314, + 624, + 476, + 636 + ], + "type": "text", + "content": "2.2. All-in-One Image Restoration" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "content": "All-in-one image restoration, which aims to recover clean images from various types of degradation using a single unified model, has gained significant attention for its efficiency and practical applications in low-level vision tasks [14, 20, 33, 38, 41]. Early methods, such as AirNet [14], utilized contrastive learning to extract degra" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23122" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 178 + ], + "type": "text", + "content": "dation representations, while IDR [41]introduced a two-stage meta-learning approach to address diverse degradation types. More recently, prompt-based approaches like PromptIR [8]and PIP [17]further enhanced restoration flexibility through vision prompts. Additionally, large-scale pre-trained models, as demonstrated by DiffUIR [44] and DA-CLIP [18], have been adapted for all-in-one restoration with strong performance. Please refer to the supplementary material for more detailed related work." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 192, + 111, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 111, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 111, + 205 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 213, + 123, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 123, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 123, + 224 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 231, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 231, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 296, + 434 + ], + "type": "text", + "content": "The framework of our method is shown in Figure 2. Overall, our approach follows a two-stage paradigm. In the first stage, we train a VAE on clean images with an image reconstruction task, as shown in Figure 2(a). In the second stage, as shown in Figure 2(b), we freeze the VAE weights and train the remaining components for image restoration tasks. To mitigate high-frequency loss during the VAE encoding process, we pass high-frequency components from the wavelet transformation between the encoder adapter and the decoder adapter. Additionally, prompts are injected through the adapters: in the encoding phase, we apply degradation-aware low-rank prompts (Figure 2(d)) to adapt the VAE to different degradations, encoding them into a unified compact latent space; in the decoding phase, we use frequency selection prompts (Figure 2(c)) for each type of degradation, enabling efficient adaptive fusion of low-frequency decoder information with high-frequency encoder information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 435, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 296, + 590 + ], + "type": "text", + "content": "The latent restoration network is divided into " + }, + { + "bbox": [ + 55, + 435, + 296, + 590 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 435, + 296, + 590 + ], + "type": "text", + "content": " subnetworks, each corresponding to one of the " + }, + { + "bbox": [ + 55, + 435, + 296, + 590 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 435, + 296, + 590 + ], + "type": "text", + "content": " decomposed frequency bands of the input. During Frequency Domain Decoupling Progressive Learning (Figure 2(e)), we progressively increase from the lowest to the highest frequency bands, incrementally adding sub-networks until all frequency bands are restored. Attributing the strong performance to the robust latent space encoding and frequency-progressive learning strategy, we achieve excellent results with a simple latent space restoration network. In this work, we adopt the Cubic-Mixer [45] for its efficiency, results of other architectures and the interaction detail between the adapter and VAE, please refer to the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 600, + 295, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 295, + 625 + ], + "type": "text", + "content": "3.2. Frequency Domain Decoupling Progressive Learning Strategy" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": "For all-in-one tasks, which aim to handle multiple types of degradation with a single model, we can represent this as " + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "inline_equation", + "content": "F(\\{\\tau^t\\}; \\theta)_{t \\in [1, N]}" + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": " represents the number of tasks, " + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\tau^t" + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": " denotes the input for task " + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": " represents the model parameters. Each task " + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\tau^t" + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": " corresponds to a distinct optimization objective; for instance, the low-light enhancement task focuses on adjusting the global brightness of the image, while" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "content": "the deraining task aims to remove local rain streaks. Consequently, it is challenging for a network trained from scratch to identify a suitable gradient direction for all tasks during initial training, which often results in slow convergence or convergence to suboptimal solutions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 133, + 556, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 556, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 556, + 276 + ], + "type": "text", + "content": "Inspired by the concept of curriculum learning [34][19], a learning process that progresses from easy to difficult tasks is more beneficial for optimizing complex tasks. MIO-IR [13] employs a sequential learning strategy to incrementally add tasks, thereby mitigating the interference caused by joint training. However, while extending from single-task to multitask learning is a form of curriculum learning, the relationships among tasks remain ambiguous, making it challenging to establish a generalized strategy for determining the optimal task sequence. Therefore, we revisit how to design a more general curriculum learning strategy for UHD all-in-one tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 286, + 395, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 286, + 395, + 297 + ], + "spans": [ + { + "bbox": [ + 314, + 286, + 395, + 297 + ], + "type": "text", + "content": "3.2.1. Observation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 302, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 555, + 373 + ], + "type": "text", + "content": "First, we assess which processes are easier or harder for a neural network to learn during training, starting from random initialization to being capable of handling multiple degradations effectively. Inspired by the frequency principle of neural networks [23][25][37][30], we approach this problem from the perspective of the frequency domain." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 375, + 556, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 556, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 556, + 508 + ], + "type": "text", + "content": "For a restoration task " + }, + { + "bbox": [ + 313, + 375, + 556, + 508 + ], + "type": "inline_equation", + "content": "\\tau^t" + }, + { + "bbox": [ + 313, + 375, + 556, + 508 + ], + "type": "text", + "content": ", we decompose it into " + }, + { + "bbox": [ + 313, + 375, + 556, + 508 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 375, + 556, + 508 + ], + "type": "text", + "content": " frequency bands, with the " + }, + { + "bbox": [ + 313, + 375, + 556, + 508 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 375, + 556, + 508 + ], + "type": "text", + "content": "-th frequency band denoted as " + }, + { + "bbox": [ + 313, + 375, + 556, + 508 + ], + "type": "inline_equation", + "content": "\\tau_f^t" + }, + { + "bbox": [ + 313, + 375, + 556, + 508 + ], + "type": "text", + "content": ". The multi-degradation restoration process is represented as " + }, + { + "bbox": [ + 313, + 375, + 556, + 508 + ], + "type": "inline_equation", + "content": "F(\\tau_f^t;\\theta)_{t\\in [1,N],f\\in [1,M]}" + }, + { + "bbox": [ + 313, + 375, + 556, + 508 + ], + "type": "text", + "content": ". In sequential learning, tasks are learned progressively from 1 to " + }, + { + "bbox": [ + 313, + 375, + 556, + 508 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 375, + 556, + 508 + ], + "type": "text", + "content": ", where each stage involves restoring all frequency bands, represented as " + }, + { + "bbox": [ + 313, + 375, + 556, + 508 + ], + "type": "inline_equation", + "content": "F(\\{\\tau_f^t\\}_{f = 1}^M;\\theta)_{t = 1\\to N}" + }, + { + "bbox": [ + 313, + 375, + 556, + 508 + ], + "type": "text", + "content": ". The main drawback of this approach is the lack of strong inter-stage connections, and the relationships between all degradation types are only modeled in the final stage, which hinders the network's ability to comprehensively learn the interplay between different degradations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 509, + 556, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 556, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 556, + 593 + ], + "type": "text", + "content": "To address this, we shift from task-based to frequency band-based progressive learning, represented as " + }, + { + "bbox": [ + 313, + 509, + 556, + 593 + ], + "type": "inline_equation", + "content": "F(\\{\\tau_f^t\\}_{t=1}^N; \\theta)_{f=1 \\to M}" + }, + { + "bbox": [ + 313, + 509, + 556, + 593 + ], + "type": "text", + "content": ". Each frequency band captures global image information and maintains a physical relationship with other bands. Learning by frequency bands is smoother and more natural, as knowledge from earlier bands supports learning subsequent bands." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": "To analyze the differences between degradations under low- and high-frequency conditions, we conducted a toy experiment. We selected four types of degradations: low-light and haze, which predominantly affect the low-frequency bands of an image, and denoising and deblurring, which mainly impact the high-frequency bands. We applied a low-pass filter to decompose images into low- and high-frequency components, then trained separate models for each. To quantify the gap between different tasks, we introduced the metric " + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\hat{z}_{ij}^{s}" + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": " from [43], which is used to measure the discrepancy" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23123" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 207, + 370 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 207, + 370 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 207, + 370 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 207, + 370 + ], + "type": "image", + "image_path": "f04361381cdecd0c74ec0b94828d0c107be49b16b57d7f8a45fec4282c0f08c3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 380, + 555, + 437 + ], + "lines": [ + { + "bbox": [ + 54, + 380, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 54, + 380, + 555, + 437 + ], + "type": "text", + "content": "Figure 2. The overall framework of our proposed UHD-Processor, which includes: (a) Pre-training a VAE on clean images for an image reconstruction task. (b) During the image restoration task, we keep the VAE frozen and efficiently fine-tune it using an adapter, where latentIR represents the latent space image restoration network. (c) Degradation-Specific Frequency Selection Prompt (DFSP) for the decoder. (d) Degradation-Aware Low Rank Prompt (DALR) for the encoder, and (e) an illustration of the Frequency Domain Decoupling Progressive Learning Strategy." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 210, + 71, + 552, + 371 + ], + "blocks": [ + { + "bbox": [ + 210, + 71, + 552, + 371 + ], + "lines": [ + { + "bbox": [ + 210, + 71, + 552, + 371 + ], + "spans": [ + { + "bbox": [ + 210, + 71, + 552, + 371 + ], + "type": "image", + "image_path": "ff5635cbce2670a008074d46ed9c56c733d8478887a37414bb97637eeb3da48e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 456, + 295, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 456, + 295, + 493 + ], + "spans": [ + { + "bbox": [ + 54, + 456, + 295, + 493 + ], + "type": "text", + "content": "between pre-training and fine-tuning tasks (higher values indicate smaller differences). For a detailed discussion of this metric, please refer to the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 495, + 295, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 495, + 295, + 614 + ], + "spans": [ + { + "bbox": [ + 54, + 495, + 295, + 614 + ], + "type": "text", + "content": "Experimental results in Figure 1(d) indicate that, for both low and high frequencies, the value of " + }, + { + "bbox": [ + 54, + 495, + 295, + 614 + ], + "type": "inline_equation", + "content": "\\hat{z}_{ij}^{s}" + }, + { + "bbox": [ + 54, + 495, + 295, + 614 + ], + "type": "text", + "content": " is consistently higher compared to the entire frequency spectrum. This underscores the critical role of frequency-wise decomposition in minimizing task-specific discrepancies within individual frequency bands. Moreover, " + }, + { + "bbox": [ + 54, + 495, + 295, + 614 + ], + "type": "inline_equation", + "content": "\\hat{z}_{ij}^{s}" + }, + { + "bbox": [ + 54, + 495, + 295, + 614 + ], + "type": "text", + "content": " is significantly higher in the low-frequency band than in the high-frequency band, suggesting that a progressive learning sequence from low to high frequencies is better suited for establishing an effective curriculum learning framework." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 617, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 297, + 714 + ], + "type": "text", + "content": "To further investigate the differences in conclusions between standard and UHD resolutions, we conducted experiments on UHD data for four types of degradations and their corresponding downsampled versions using an all-in-one training setup. As shown in Figure 1(c), By adopting a training sequence progressing from low to high frequencies, we observed improvements for both resolutions compared to direct training. However, the enhancement was more pro" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 456, + 555, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 555, + 492 + ], + "type": "text", + "content": "nounced in UHD scenarios, which can be attributed to the richer high-frequency texture information present in UHD images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 502, + 555, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 527 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 527 + ], + "type": "text", + "content": "3.2.2. Frequency Domain Decoupling Progressive Learning Strategy" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 531, + 556, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 556, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 556, + 594 + ], + "type": "text", + "content": "Therefore, we propose the Frequency Decoupling Progressive Learning strategy, which leverages multiple wavelet transformation (MWT) to decompose the input image " + }, + { + "bbox": [ + 313, + 531, + 556, + 594 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 531, + 556, + 594 + ], + "type": "text", + "content": " and ground truth " + }, + { + "bbox": [ + 313, + 531, + 556, + 594 + ], + "type": "inline_equation", + "content": "\\Upsilon" + }, + { + "bbox": [ + 313, + 531, + 556, + 594 + ], + "type": "text", + "content": " into multiple frequency sub-bands " + }, + { + "bbox": [ + 313, + 531, + 556, + 594 + ], + "type": "inline_equation", + "content": "\\{X_f\\}_{f=1}^M" + }, + { + "bbox": [ + 313, + 531, + 556, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 531, + 556, + 594 + ], + "type": "inline_equation", + "content": "\\{Y_f\\}_{f=1}^M" + }, + { + "bbox": [ + 313, + 531, + 556, + 594 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 594, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 557, + 714 + ], + "type": "text", + "content": "During training, the learning process progressively advances from low to high frequencies, with each frequency band " + }, + { + "bbox": [ + 313, + 594, + 557, + 714 + ], + "type": "inline_equation", + "content": "X_{f}" + }, + { + "bbox": [ + 313, + 594, + 557, + 714 + ], + "type": "text", + "content": " managed by a dedicated restoration sub-network " + }, + { + "bbox": [ + 313, + 594, + 557, + 714 + ], + "type": "inline_equation", + "content": "F_{f}" + }, + { + "bbox": [ + 313, + 594, + 557, + 714 + ], + "type": "text", + "content": ". As training progresses and higher frequency bands are introduced, additional restoration sub-networks are incrementally incorporated, forming a divide-and-conquer approach to effectively handle each frequency band. The process can be formulated as Algorithm 1. This pseudocode omits the detailed steps of multiple wavelet transformation (MWT) and its corresponding inverse process (MIWT). For" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23124" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "a more comprehensive description of the algorithm, please refer to the supplementary materials." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 131 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 131 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 131 + ], + "type": "text", + "content": "Algorithm 1 Frequency Domain Decoupling Progressive Learning" + } + ] + } + ], + "index": 1, + "type": "text" + }, + { + "bbox": [ + 61, + 134, + 294, + 309 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 63, + 134, + 286, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 134, + 286, + 148 + ], + "spans": [ + { + "bbox": [ + 63, + 134, + 286, + 148 + ], + "type": "text", + "content": "1: Input: " + }, + { + "bbox": [ + 63, + 134, + 286, + 148 + ], + "type": "inline_equation", + "content": "\\{X_f\\}_{f=1}^M, \\{Y_f\\}_{f=1}^M \\gets \\mathrm{MWT}(X), \\mathrm{MWT}(Y)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 149, + 294, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 149, + 294, + 175 + ], + "spans": [ + { + "bbox": [ + 63, + 149, + 294, + 175 + ], + "type": "text", + "content": "2: Init: Initialize " + }, + { + "bbox": [ + 63, + 149, + 294, + 175 + ], + "type": "inline_equation", + "content": "\\{\\theta_f\\}_{f=1}^M" + }, + { + "bbox": [ + 63, + 149, + 294, + 175 + ], + "type": "text", + "content": " for sub-networks, set stage thresholds " + }, + { + "bbox": [ + 63, + 149, + 294, + 175 + ], + "type": "inline_equation", + "content": "\\{T_k\\}_{k=1}^M" + }, + { + "bbox": [ + 63, + 149, + 294, + 175 + ], + "type": "text", + "content": ", and define total_iter" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 175, + 181, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 175, + 181, + 186 + ], + "spans": [ + { + "bbox": [ + 63, + 175, + 181, + 186 + ], + "type": "text", + "content": "3: current_iter " + }, + { + "bbox": [ + 63, + 175, + 181, + 186 + ], + "type": "inline_equation", + "content": "\\leftarrow 0, k \\leftarrow 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 186, + 224, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 186, + 224, + 197 + ], + "spans": [ + { + "bbox": [ + 63, + 186, + 224, + 197 + ], + "type": "text", + "content": "4: while current_iter " + }, + { + "bbox": [ + 63, + 186, + 224, + 197 + ], + "type": "inline_equation", + "content": "\\leq" + }, + { + "bbox": [ + 63, + 186, + 224, + 197 + ], + "type": "text", + "content": " total_iter do" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 198, + 232, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 198, + 232, + 211 + ], + "spans": [ + { + "bbox": [ + 63, + 198, + 232, + 211 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 63, + 198, + 232, + 211 + ], + "type": "inline_equation", + "content": "\\{\\hat{X}_f\\mid f = 1,\\dots ,k\\} \\gets F_f(X_f;\\theta_f)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 63, + 212, + 234, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 212, + 234, + 224 + ], + "spans": [ + { + "bbox": [ + 63, + 212, + 234, + 224 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 63, + 212, + 234, + 224 + ], + "type": "inline_equation", + "content": "\\dot{X}\\gets \\mathrm{MIWT}(\\{\\dot{X}_f\\}_{f = 1}^k,\\{Y_f\\}_{f = k + 1}^M)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 63, + 226, + 155, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 226, + 155, + 237 + ], + "spans": [ + { + "bbox": [ + 63, + 226, + 155, + 237 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 63, + 226, + 155, + 237 + ], + "type": "inline_equation", + "content": "\\mathcal{L}\\gets \\operatorname {Loss}(\\hat{X},Y)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 63, + 238, + 263, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 238, + 263, + 252 + ], + "spans": [ + { + "bbox": [ + 63, + 238, + 263, + 252 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 63, + 238, + 263, + 252 + ], + "type": "inline_equation", + "content": "\\{\\theta_f\\}_{f = 1}^k\\gets \\{\\theta_f\\}_{f = 1}^k -\\alpha \\nabla_{\\{\\theta_f\\}_{f = 1}^k}\\mathcal{L}(\\tilde{X},Y)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 253, + 228, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 253, + 228, + 262 + ], + "spans": [ + { + "bbox": [ + 63, + 253, + 228, + 262 + ], + "type": "text", + "content": "9: current_iter " + }, + { + "bbox": [ + 63, + 253, + 228, + 262 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 63, + 253, + 228, + 262 + ], + "type": "text", + "content": " current_iter + 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 264, + 245, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 264, + 245, + 275 + ], + "spans": [ + { + "bbox": [ + 61, + 264, + 245, + 275 + ], + "type": "text", + "content": "10: if current_iter " + }, + { + "bbox": [ + 61, + 264, + 245, + 275 + ], + "type": "inline_equation", + "content": ">T_{k}" + }, + { + "bbox": [ + 61, + 264, + 245, + 275 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 61, + 264, + 245, + 275 + ], + "type": "inline_equation", + "content": "k < M" + }, + { + "bbox": [ + 61, + 264, + 245, + 275 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 277, + 138, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 277, + 138, + 285 + ], + "spans": [ + { + "bbox": [ + 61, + 277, + 138, + 285 + ], + "type": "text", + "content": "11: " + }, + { + "bbox": [ + 61, + 277, + 138, + 285 + ], + "type": "inline_equation", + "content": "k\\gets k + 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 61, + 289, + 110, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 289, + 110, + 297 + ], + "spans": [ + { + "bbox": [ + 61, + 289, + 110, + 297 + ], + "type": "text", + "content": "12: end if" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 61, + 300, + 116, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 300, + 116, + 309 + ], + "spans": [ + { + "bbox": [ + 61, + 300, + 116, + 309 + ], + "type": "text", + "content": "13: end while" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 330, + 247, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 247, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 247, + 342 + ], + "type": "text", + "content": "3.3. Efficient Adaptive Prompt Learning" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 54, + 347, + 296, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 347, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 54, + 347, + 296, + 467 + ], + "type": "text", + "content": "Based on our VAE-based restoration framework, we design two different prompt strategies for the distinct behaviors of the VAE encoding and decoding processes. For the encoding process, since the VAE is pre-trained on clean images, directly encoding various degraded images introduces a significant domain gap. Therefore, we introduce a low-rank degradation-aware prompt to the encoder, enabling it to encode different degraded images into the pre-trained compact and regularized latent space, thereby reducing the difficulty of the subsequent restoration mapping." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 467, + 296, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 296, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 296, + 609 + ], + "type": "text", + "content": "During the decoding stage, despite the successful mitigation of degradation effects within the latent space, challenges persist with the loss of high-frequency details when decoding back to pixel space. To address this issue, we incorporate an adapter to inject high-frequency information from the corresponding stages of the encoder into the decoder. However, since different degradations impact specific frequency bands differently, we employ a frequency-selective prompt in the decoding process. This approach facilitates an adaptive fusion of high-frequency details from the encoder with the low-frequency information from the decoder, tailored to the specific frequency bands affected by the degradation." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 610, + 296, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 296, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 296, + 670 + ], + "type": "text", + "content": "Additionally, to allow the network to effectively recognize the type of degradation, we pre-train a degradation classifier. Since this classification task is not particularly challenging, a simple classification network is sufficient to achieve high accuracy." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 674, + 250, + 687 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 674, + 250, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 674, + 250, + 687 + ], + "type": "text", + "content": "3.3.1. Degradation-Aware Low Rank Prompt" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "The current methods for prompt injection in all-in-one tasks primarily follow the design of PromptIR [20]. In this frame" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 71, + 555, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 157 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 157 + ], + "type": "text", + "content": "work, for the input feature " + }, + { + "bbox": [ + 313, + 71, + 555, + 157 + ], + "type": "inline_equation", + "content": "F_{i}\\in \\mathbb{R}^{H\\times W\\times C}" + }, + { + "bbox": [ + 313, + 71, + 555, + 157 + ], + "type": "text", + "content": ", a learnable prompt tensor " + }, + { + "bbox": [ + 313, + 71, + 555, + 157 + ], + "type": "inline_equation", + "content": "P_{c}\\in \\mathbb{R}^{N\\times \\hat{H}\\times \\hat{W}\\times C}" + }, + { + "bbox": [ + 313, + 71, + 555, + 157 + ], + "type": "text", + "content": " is first constructed, where " + }, + { + "bbox": [ + 313, + 71, + 555, + 157 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 71, + 555, + 157 + ], + "type": "text", + "content": " denotes the number of prompt components, and " + }, + { + "bbox": [ + 313, + 71, + 555, + 157 + ], + "type": "inline_equation", + "content": "\\hat{H}" + }, + { + "bbox": [ + 313, + 71, + 555, + 157 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 71, + 555, + 157 + ], + "type": "inline_equation", + "content": "\\hat{W}" + }, + { + "bbox": [ + 313, + 71, + 555, + 157 + ], + "type": "text", + "content": " represent the spatial dimensions, while the channel count " + }, + { + "bbox": [ + 313, + 71, + 555, + 157 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 71, + 555, + 157 + ], + "type": "text", + "content": " aligns with that of the input. Subsequently, " + }, + { + "bbox": [ + 313, + 71, + 555, + 157 + ], + "type": "inline_equation", + "content": "P_{c}" + }, + { + "bbox": [ + 313, + 71, + 555, + 157 + ], + "type": "text", + "content": " is interpolated to match the size of " + }, + { + "bbox": [ + 313, + 71, + 555, + 157 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 313, + 71, + 555, + 157 + ], + "type": "text", + "content": ", enabling further interaction to generate the prompts." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 158, + 555, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 158, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 555, + 277 + ], + "type": "text", + "content": "While this design has shown promising performance in tasks at standard resolutions, it faces two significant challenges in UHD scenarios. First, the entire learnable tensor " + }, + { + "bbox": [ + 313, + 158, + 555, + 277 + ], + "type": "inline_equation", + "content": "P_{c}" + }, + { + "bbox": [ + 313, + 158, + 555, + 277 + ], + "type": "text", + "content": " comprises a substantial number of parameters due to the dimensions " + }, + { + "bbox": [ + 313, + 158, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\hat{H} \\times \\hat{W}" + }, + { + "bbox": [ + 313, + 158, + 555, + 277 + ], + "type": "text", + "content": ". Furthermore, the prompt is homogeneous across all degradation types, which means its complexity cannot be adapted based on the difficulty of each specific degradation. This lack of customization makes the approach inefficient for UHD image restoration, where the difficulty of different degradation categories can vary significantly." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 277, + 555, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 386 + ], + "type": "text", + "content": "Second, since the dimensions " + }, + { + "bbox": [ + 313, + 277, + 555, + 386 + ], + "type": "inline_equation", + "content": "\\hat{H}" + }, + { + "bbox": [ + 313, + 277, + 555, + 386 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 277, + 555, + 386 + ], + "type": "inline_equation", + "content": "\\hat{W}" + }, + { + "bbox": [ + 313, + 277, + 555, + 386 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 277, + 555, + 386 + ], + "type": "inline_equation", + "content": "P_{c}" + }, + { + "bbox": [ + 313, + 277, + 555, + 386 + ], + "type": "text", + "content": " must be predetermined and interpolated to ensure adaptability to feature resolution, traditional all-in-one tasks exhibit minimal differences between training and inference resolutions. This allows the trained " + }, + { + "bbox": [ + 313, + 277, + 555, + 386 + ], + "type": "inline_equation", + "content": "P_{c}" + }, + { + "bbox": [ + 313, + 277, + 555, + 386 + ], + "type": "text", + "content": " to generalize well during inference. However, in UHD tasks, training typically uses a patch size of " + }, + { + "bbox": [ + 313, + 277, + 555, + 386 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 277, + 555, + 386 + ], + "type": "text", + "content": " for efficiency, while inference requires processing full images at " + }, + { + "bbox": [ + 313, + 277, + 555, + 386 + ], + "type": "inline_equation", + "content": "3840 \\times 2160" + }, + { + "bbox": [ + 313, + 277, + 555, + 386 + ], + "type": "text", + "content": ". This significant resolution gap limits the generalization of " + }, + { + "bbox": [ + 313, + 277, + 555, + 386 + ], + "type": "inline_equation", + "content": "P_{c}" + }, + { + "bbox": [ + 313, + 277, + 555, + 386 + ], + "type": "text", + "content": " during inference." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 386, + 555, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 555, + 436 + ], + "type": "text", + "content": "To address the issue of parameter efficiency, inspired by the low-rank adaptation method [9], we perform a low-rank decomposition on the spatial dimensions of the original parameter matrix " + }, + { + "bbox": [ + 313, + 386, + 555, + 436 + ], + "type": "inline_equation", + "content": "P_{c}\\in \\mathbb{R}^{N\\times \\hat{H}\\times \\hat{W}\\times C}" + }, + { + "bbox": [ + 313, + 386, + 555, + 436 + ], + "type": "text", + "content": ", represented as:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 379, + 446, + 555, + 461 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 446, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 379, + 446, + 555, + 461 + ], + "type": "interline_equation", + "content": "P _ {c l} = \\operatorname {c a t} \\left[ B _ {n} A _ {n} \\right] _ {n \\in [ 1 \\dots N ]}, \\tag {1}", + "image_path": "cb007a6498cab740c020d8ef54f281c35bf0643de0fda2365d433377abe91c48.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 471, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 554, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 471, + 554, + 533 + ], + "type": "inline_equation", + "content": "B_{n}\\in \\mathbb{R}^{1\\times \\hat{H}\\times r\\times C}" + }, + { + "bbox": [ + 313, + 471, + 554, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 471, + 554, + 533 + ], + "type": "inline_equation", + "content": "A_{n}\\in \\mathbb{R}^{1\\times r\\times \\hat{W}\\times C}" + }, + { + "bbox": [ + 313, + 471, + 554, + 533 + ], + "type": "text", + "content": " represent the matrices resulting from the low-rank decomposition, and " + }, + { + "bbox": [ + 313, + 471, + 554, + 533 + ], + "type": "inline_equation", + "content": "\\operatorname {cat}[\\cdot ]" + }, + { + "bbox": [ + 313, + 471, + 554, + 533 + ], + "type": "text", + "content": " represents concatenation along the first dimension, with " + }, + { + "bbox": [ + 313, + 471, + 554, + 533 + ], + "type": "inline_equation", + "content": "r\\ll \\min (\\hat{H},\\hat{W})" + }, + { + "bbox": [ + 313, + 471, + 554, + 533 + ], + "type": "text", + "content": " . After low-rank decomposition, the original parameter matrix is effectively compressed." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 533, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 629 + ], + "type": "text", + "content": "Furthermore, for different types of degradations, we can specify different ranks " + }, + { + "bbox": [ + 313, + 533, + 555, + 629 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 533, + 555, + 629 + ], + "type": "text", + "content": ". More complex degradations can use a larger " + }, + { + "bbox": [ + 313, + 533, + 555, + 629 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 533, + 555, + 629 + ], + "type": "text", + "content": " to enhance the parameter space, while simpler degradations can use a smaller " + }, + { + "bbox": [ + 313, + 533, + 555, + 629 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 533, + 555, + 629 + ], + "type": "text", + "content": " to improve parameter efficiency. The use of different ranks " + }, + { + "bbox": [ + 313, + 533, + 555, + 629 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 533, + 555, + 629 + ], + "type": "text", + "content": " introduces unique designs for various degradations, while the multiplication of the decomposed matrices preserves consistent dimensions, allowing efficient parallel processing through concatenation." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "content": "To address the issue of resolution adaptability, after obtaining the low-rank prompt tensor " + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "inline_equation", + "content": "P_{cl}" + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "content": ", we do not directly interpolate it to match the input features. Instead, considering the scale invariance of convolution operations, we utilize " + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "inline_equation", + "content": "P_{cl}" + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "content": " as a parameter space to learn multiple convolution operators. By using convolution as the core operator for prompt interactions, this approach is more effective in adapting to" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23125" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "the differences between inference and training resolutions compared to simple interpolation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "text", + "content": "Specifically, the prompt generation process during the encoding phase is as follows. First, the input image is passed through a classifier to obtain " + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "inline_equation", + "content": "N + 1" + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "text", + "content": " classification scores, where " + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "text", + "content": " represents the number of degradation types, and 1 denotes the clean image. We then further process these classification scores to derive " + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "inline_equation", + "content": "W_{1}" + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "inline_equation", + "content": "W_{2}" + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 117, + 172, + 296, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 172, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 117, + 172, + 296, + 194 + ], + "type": "interline_equation", + "content": "W _ {1} = \\operatorname {s o f t m a x} (\\sum (N), 1), \\tag {2}", + "image_path": "cfa8a46f03a1f3f56548030a5387586082433eaafe08e23961c12cb5c049c42c.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 119, + 194, + 203, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 194, + 203, + 205 + ], + "spans": [ + { + "bbox": [ + 119, + 194, + 203, + 205 + ], + "type": "interline_equation", + "content": "W _ {2} = \\operatorname {s o f t m a x} (N),", + "image_path": "9cab8ba8921bf95f0067f703528c70fde40d13826f8bb5987f91fd882794d669.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 210, + 296, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 210, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 210, + 296, + 270 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 55, + 210, + 296, + 270 + ], + "type": "inline_equation", + "content": "W_{1}" + }, + { + "bbox": [ + 55, + 210, + 296, + 270 + ], + "type": "text", + "content": " represents the binary classification score between degraded and clean images, serving to characterize the degree of degradation, while " + }, + { + "bbox": [ + 55, + 210, + 296, + 270 + ], + "type": "inline_equation", + "content": "W_{2}" + }, + { + "bbox": [ + 55, + 210, + 296, + 270 + ], + "type": "text", + "content": " corresponds to the " + }, + { + "bbox": [ + 55, + 210, + 296, + 270 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 210, + 296, + 270 + ], + "type": "text", + "content": "-way classification scores for the " + }, + { + "bbox": [ + 55, + 210, + 296, + 270 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 210, + 296, + 270 + ], + "type": "text", + "content": " different degradation types, indicating the specific type of degradation present." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 270, + 296, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 304 + ], + "type": "text", + "content": "We then construct two simple pointwise convolutions to provide degradation degree prompts(DDP), as illustrated below:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 308, + 296, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 308, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 107, + 308, + 296, + 327 + ], + "type": "interline_equation", + "content": "F _ {i 1} ^ {\\prime} = \\operatorname {c a t} \\left[ P _ {\\text {d e g}} \\left(F _ {i}\\right), P _ {\\text {c l e a n}} \\left(F _ {i}\\right) \\right], \\tag {3}", + "image_path": "fec00910afa1eac278c4504f9573fa993368ece44c3397b30a6653e5cd77ef76.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 108, + 324, + 181, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 324, + 181, + 338 + ], + "spans": [ + { + "bbox": [ + 108, + 324, + 181, + 338 + ], + "type": "interline_equation", + "content": "F _ {i 1} = F _ {i 1} ^ {\\prime} \\odot W _ {1},", + "image_path": "90a759a3acca67c270ef305126de622eda3de341c65e25148525fa94eca64650.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 342, + 296, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 378 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 55, + 342, + 296, + 378 + ], + "type": "inline_equation", + "content": "P_{deg}(\\cdot)" + }, + { + "bbox": [ + 55, + 342, + 296, + 378 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 342, + 296, + 378 + ], + "type": "inline_equation", + "content": "P_{clean}(\\cdot)" + }, + { + "bbox": [ + 55, + 342, + 296, + 378 + ], + "type": "text", + "content": " denote the pointwise convolutions used to provide degradation degree prompts, while " + }, + { + "bbox": [ + 55, + 342, + 296, + 378 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 342, + 296, + 378 + ], + "type": "text", + "content": " represents the weighted summation of their outputs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 380, + 296, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 296, + 474 + ], + "type": "text", + "content": "After obtaining the degradation-aware feature " + }, + { + "bbox": [ + 55, + 380, + 296, + 474 + ], + "type": "inline_equation", + "content": "F_{i1}" + }, + { + "bbox": [ + 55, + 380, + 296, + 474 + ], + "type": "text", + "content": ", we proceed to provide prompts for degradation types(DTP). First, we define the rank of the parameter space for each degradation type as " + }, + { + "bbox": [ + 55, + 380, + 296, + 474 + ], + "type": "inline_equation", + "content": "R_{N} = [r_{1}, r_{2}, \\dots, r_{N}]" + }, + { + "bbox": [ + 55, + 380, + 296, + 474 + ], + "type": "text", + "content": ", which allows us to derive the parameter space " + }, + { + "bbox": [ + 55, + 380, + 296, + 474 + ], + "type": "inline_equation", + "content": "P_{cl}" + }, + { + "bbox": [ + 55, + 380, + 296, + 474 + ], + "type": "text", + "content": " based on Equation (1). Next, we utilize a simple MLP to generate a set of deep-wise convolution kernels " + }, + { + "bbox": [ + 55, + 380, + 296, + 474 + ], + "type": "inline_equation", + "content": "K_{s} = [(k_{1}:d_{1}), (k_{2}:d_{2}), \\dots, (k_{m}:d_{m})]" + }, + { + "bbox": [ + 55, + 380, + 296, + 474 + ], + "type": "text", + "content": ", characterized by varying kernel sizes and dilation rates." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 475, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 296, + 581 + ], + "type": "text", + "content": "The feature " + }, + { + "bbox": [ + 55, + 475, + 296, + 581 + ], + "type": "inline_equation", + "content": "F_{i1}" + }, + { + "bbox": [ + 55, + 475, + 296, + 581 + ], + "type": "text", + "content": " is then processed in parallel through all multi-scale convolution group (MCKG), followed by a weighted summation with " + }, + { + "bbox": [ + 55, + 475, + 296, + 581 + ], + "type": "inline_equation", + "content": "W_{2}" + }, + { + "bbox": [ + 55, + 475, + 296, + 581 + ], + "type": "text", + "content": ". Due to the impact of parallel multi-branch convolutional kernels on inference efficiency [6], we employ the structural reparameterization method from [7]. In the inference phase, all convolutional kernels are reparameterized and merged into a single large convolutional kernel. This process can be expressed as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 84, + 585, + 186, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 585, + 186, + 598 + ], + "spans": [ + { + "bbox": [ + 84, + 585, + 186, + 598 + ], + "type": "interline_equation", + "content": "K _ {s} = \\mathrm {M L P} _ {i \\in [ 1: m ]} (P _ {c l}),", + "image_path": "acfcdad0705ed9e60f54f821ef7e0715a6726ffb91a56e2e8ee1bc4343ed00c9.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 84, + 601, + 296, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 601, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 84, + 601, + 296, + 632 + ], + "type": "interline_equation", + "content": "F _ {i 2} ^ {\\prime} = \\left\\{ \\begin{array}{l l} \\sum \\left(K _ {s} \\left(F _ {i 1}\\right)\\right) & \\text {i f i n t r a i n i n g p h a s e}, \\\\ \\operatorname {R e p} \\left[ K _ {s} \\right] \\left(F _ {i 1}\\right) & \\text {i f i n t e s t i n g p h a s e}, \\end{array} \\right. \\tag {4}", + "image_path": "4c18b85f7f9b8c7ab8275f6de69f3578d3a977773301108dd085be3fcb201a90.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 84, + 635, + 156, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 635, + 156, + 649 + ], + "spans": [ + { + "bbox": [ + 84, + 635, + 156, + 649 + ], + "type": "interline_equation", + "content": "F _ {i 2} = W _ {2} \\odot F _ {i 2} ^ {\\prime},", + "image_path": "9cb0d23e0b73cf4006f5f0e48408597b84bb5f35be1b9edd99fe27441ff2e726.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{Rep}[\\cdot ]" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " refers to the reparameterization operation. The resultant " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "F_{i2}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " serves as the low-rank degradation-aware prompt, which is integrated into the encoder's adapter. This mechanism facilitates the encoder's ability to encode diverse degradations into a cohesive and compact latent space." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 555, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 84 + ], + "type": "text", + "content": "3.3.2. Degradation-Specific Frequency Selection Prompt" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 87, + 555, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 87, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 313, + 87, + 555, + 171 + ], + "type": "text", + "content": "Given that different degradations affect distinct frequency bands [4] [11], we design prompts in the decoding phase to facilitate adaptive selection of frequency components. Specifically, this involves integrating medium-frequency information " + }, + { + "bbox": [ + 313, + 87, + 555, + 171 + ], + "type": "inline_equation", + "content": "\\{F_E^{HL}, F_E^{LH}\\}" + }, + { + "bbox": [ + 313, + 87, + 555, + 171 + ], + "type": "text", + "content": " and high-frequency information " + }, + { + "bbox": [ + 313, + 87, + 555, + 171 + ], + "type": "inline_equation", + "content": "F_E^{HH}" + }, + { + "bbox": [ + 313, + 87, + 555, + 171 + ], + "type": "text", + "content": " from the encoder, along with low-frequency information " + }, + { + "bbox": [ + 313, + 87, + 555, + 171 + ], + "type": "inline_equation", + "content": "F_D^{LL}" + }, + { + "bbox": [ + 313, + 87, + 555, + 171 + ], + "type": "text", + "content": " from the decoder." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 171, + 555, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 555, + 284 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 555, + 284 + ], + "type": "text", + "content": "We begin by applying the Fast Fourier Transform (FFT) to each frequency band feature in the adapter branch, represented as " + }, + { + "bbox": [ + 313, + 171, + 555, + 284 + ], + "type": "inline_equation", + "content": "F_{w} = \\{F_{D}^{LL}, F_{E}^{HL}, F_{E}^{LH}, F_{E}^{HH}\\} \\in \\mathbb{R}^{B \\times 4 \\times \\frac{\\hat{H}}{2} \\times \\frac{\\hat{W}}{2} \\times C}" + }, + { + "bbox": [ + 313, + 171, + 555, + 284 + ], + "type": "text", + "content": ". Pointwise convolution is applied to each band to efficiently capture global features and expand the channels by a factor of " + }, + { + "bbox": [ + 313, + 171, + 555, + 284 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 171, + 555, + 284 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 313, + 171, + 555, + 284 + ], + "type": "inline_equation", + "content": "F_{w1} \\in \\mathbb{R}^{B \\times 4 \\times N \\times \\frac{\\hat{H}}{2} \\times \\frac{\\hat{W}}{2} \\times C}" + }, + { + "bbox": [ + 313, + 171, + 555, + 284 + ], + "type": "text", + "content": " after reshaping. A weighted summation with " + }, + { + "bbox": [ + 313, + 171, + 555, + 284 + ], + "type": "inline_equation", + "content": "W_{2}" + }, + { + "bbox": [ + 313, + 171, + 555, + 284 + ], + "type": "text", + "content": " is then performed along the " + }, + { + "bbox": [ + 313, + 171, + 555, + 284 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 171, + 555, + 284 + ], + "type": "text", + "content": " dimension, followed by the Inverse FFT (IFFT) to obtain " + }, + { + "bbox": [ + 313, + 171, + 555, + 284 + ], + "type": "inline_equation", + "content": "F_{w1}'" + }, + { + "bbox": [ + 313, + 171, + 555, + 284 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 347, + 289, + 554, + 310 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 289, + 554, + 310 + ], + "spans": [ + { + "bbox": [ + 347, + 289, + 554, + 310 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} F _ {w 1} = \\operatorname {c a t} \\left[ \\operatorname {P c o n v} _ {\\mathrm {i} \\in [ \\text {l o w}, \\text {m i d}, \\text {h i g h} ]} \\left(\\mathcal {F} \\left(F _ {w}\\right)\\right) \\right], \\\\ F ^ {\\prime} = \\mathcal {F} ^ {- 1} \\left(W, W, F _ {w}\\right) \\end{array} \\tag {5}", + "image_path": "587c29b2f4a7fc7d24f2e99abd69b75d4eecd574c8f038d8112ba76924d7343c.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 347, + 306, + 451, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 306, + 451, + 319 + ], + "spans": [ + { + "bbox": [ + 347, + 306, + 451, + 319 + ], + "type": "interline_equation", + "content": "F _ {w 1} ^ {\\prime} = \\mathcal {F} ^ {- 1} \\left(W _ {2} \\odot F _ {w 1}\\right),", + "image_path": "abdc74b032c3d86003ac6fed1d44460ced0b769e5225ef50c626f7dc94ca8efe.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 325, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 555, + 384 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 313, + 325, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\mathrm{Pconv}_{\\mathrm{i}\\in [\\mathrm{low},\\mathrm{mid},\\mathrm{high}]}" + }, + { + "bbox": [ + 313, + 325, + 555, + 384 + ], + "type": "text", + "content": " denotes the pointwise convolutions used to process the low, mid, and high frequency bands, while " + }, + { + "bbox": [ + 313, + 325, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot)" + }, + { + "bbox": [ + 313, + 325, + 555, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 325, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{-1}(\\cdot)" + }, + { + "bbox": [ + 313, + 325, + 555, + 384 + ], + "type": "text", + "content": " represent the spatial dimensional Fourier transform and its corresponding inverse transform, respectively." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 384, + 556, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 556, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 556, + 471 + ], + "type": "text", + "content": "Subsequently, pixel shuffle is applied for frequency mixing, followed by a sigmoid activation to obtain the filter kernel " + }, + { + "bbox": [ + 313, + 384, + 556, + 471 + ], + "type": "inline_equation", + "content": "F_{w2} \\in \\mathbb{R}^{B \\times \\hat{H} \\times \\hat{W} \\times C}" + }, + { + "bbox": [ + 313, + 384, + 556, + 471 + ], + "type": "text", + "content": ". For the input " + }, + { + "bbox": [ + 313, + 384, + 556, + 471 + ], + "type": "inline_equation", + "content": "F_{d} \\in \\mathbb{R}^{B \\times \\hat{H} \\times \\hat{W} \\times C}" + }, + { + "bbox": [ + 313, + 384, + 556, + 471 + ], + "type": "text", + "content": " from the decoder, we perform FFT to shift to the frequency domain and apply " + }, + { + "bbox": [ + 313, + 384, + 556, + 471 + ], + "type": "inline_equation", + "content": "F_{w2}" + }, + { + "bbox": [ + 313, + 384, + 556, + 471 + ], + "type": "text", + "content": " for frequency domain filtering. Finally, the IFFT is conducted to obtain the frequency-selective prompt " + }, + { + "bbox": [ + 313, + 384, + 556, + 471 + ], + "type": "inline_equation", + "content": "F_{w2}'" + }, + { + "bbox": [ + 313, + 384, + 556, + 471 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 356, + 475, + 512, + 490 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 475, + 512, + 490 + ], + "spans": [ + { + "bbox": [ + 356, + 475, + 512, + 490 + ], + "type": "interline_equation", + "content": "F _ {w 2} = \\operatorname {s i g m o i d} \\left(\\operatorname {p i x e l s h u f f u l e} \\left(F _ {w 1} ^ {\\prime}\\right)\\right),", + "image_path": "a7462b2b4a9a236e870c7f9f6fb477eb83279158c6c544f1922efe70331aea24.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 358, + 488, + 554, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 488, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 358, + 488, + 554, + 506 + ], + "type": "interline_equation", + "content": "F _ {w 2} ^ {\\prime} = \\mathcal {F} ^ {- 1} \\left(F _ {w 2} \\otimes \\mathcal {F} \\left(F _ {w 2}\\right)\\right), \\tag {6}", + "image_path": "4d017eeb45f2323448cef0d4c558c9e2217cba733720ae7417ff7ecf9af95e6f.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 511, + 554, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 554, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 554, + 571 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 511, + 554, + 571 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 313, + 511, + 554, + 571 + ], + "type": "text", + "content": " denotes element-wise multiplication. " + }, + { + "bbox": [ + 313, + 511, + 554, + 571 + ], + "type": "inline_equation", + "content": "F_{w2}^{\\prime}" + }, + { + "bbox": [ + 313, + 511, + 554, + 571 + ], + "type": "text", + "content": " serves as a frequency-selective prompt injected through the adapter in the decoder, enabling the decoding process to perform a more customized multi-band feature fusion tailored to the frequency characteristics of the degradation." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 581, + 395, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 395, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 395, + 594 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 600, + 430, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 430, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 430, + 613 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "We have developed a comprehensive benchmark for evaluating UHD all-in-one restoration performance, based on datasets from UHD-LL [15], UHD-blur [5], UHD-haze [46], UHD-rain [2], and UHD-haze [31]. Additionally, we constructed a denoising dataset, UHD-noise, using 4K images from [42] as the background. For additional implementation details and dataset descriptions, please refer to the supplementary materials." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23126" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 73, + 152, + 134 + ], + "blocks": [ + { + "bbox": [ + 59, + 73, + 152, + 134 + ], + "lines": [ + { + "bbox": [ + 59, + 73, + 152, + 134 + ], + "spans": [ + { + "bbox": [ + 59, + 73, + 152, + 134 + ], + "type": "image", + "image_path": "989b7689d88e816a6cb0f407cdca5e687ce76acd91ed16c549a77525ac55ef51.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 81, + 135, + 126, + 142 + ], + "lines": [ + { + "bbox": [ + 81, + 135, + 126, + 142 + ], + "spans": [ + { + "bbox": [ + 81, + 135, + 126, + 142 + ], + "type": "text", + "content": "Image deblurring" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 155, + 74, + 201, + 100 + ], + "blocks": [ + { + "bbox": [ + 155, + 74, + 201, + 100 + ], + "lines": [ + { + "bbox": [ + 155, + 74, + 201, + 100 + ], + "spans": [ + { + "bbox": [ + 155, + 74, + 201, + 100 + ], + "type": "image", + "image_path": "659e7328f41190ef9aa1846f105e5a4ea7f170fc415b6bbf6cb98e12d9a2c993.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 101, + 190, + 107 + ], + "lines": [ + { + "bbox": [ + 162, + 101, + 190, + 107 + ], + "spans": [ + { + "bbox": [ + 162, + 101, + 190, + 107 + ], + "type": "text", + "content": "Histformer" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 205, + 74, + 251, + 100 + ], + "blocks": [ + { + "bbox": [ + 205, + 74, + 251, + 100 + ], + "lines": [ + { + "bbox": [ + 205, + 74, + 251, + 100 + ], + "spans": [ + { + "bbox": [ + 205, + 74, + 251, + 100 + ], + "type": "image", + "image_path": "ac1237d54e30069b3408419500440123ba0fadb2dcc8bbd97262a59bc0fc6022.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 101, + 241, + 107 + ], + "lines": [ + { + "bbox": [ + 216, + 101, + 241, + 107 + ], + "spans": [ + { + "bbox": [ + 216, + 101, + 241, + 107 + ], + "type": "text", + "content": "PromptIR" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 254, + 74, + 301, + 100 + ], + "blocks": [ + { + "bbox": [ + 254, + 74, + 301, + 100 + ], + "lines": [ + { + "bbox": [ + 254, + 74, + 301, + 100 + ], + "spans": [ + { + "bbox": [ + 254, + 74, + 301, + 100 + ], + "type": "image", + "image_path": "4ef8809299c46ac5b60fda594b974ed56d679d4567f35ac0de24b3c15049cca7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 101, + 281, + 107 + ], + "lines": [ + { + "bbox": [ + 271, + 101, + 281, + 107 + ], + "spans": [ + { + "bbox": [ + 271, + 101, + 281, + 107 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 156, + 108, + 201, + 133 + ], + "blocks": [ + { + "bbox": [ + 156, + 108, + 201, + 133 + ], + "lines": [ + { + "bbox": [ + 156, + 108, + 201, + 133 + ], + "spans": [ + { + "bbox": [ + 156, + 108, + 201, + 133 + ], + "type": "image", + "image_path": "e025b82451bec369a8468f046e4eed96acae49e459d8d1a00fc99cb83972372a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 135, + 192, + 142 + ], + "lines": [ + { + "bbox": [ + 162, + 135, + 192, + 142 + ], + "spans": [ + { + "bbox": [ + 162, + 135, + 192, + 142 + ], + "type": "text", + "content": "Gridformer" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 205, + 108, + 251, + 133 + ], + "blocks": [ + { + "bbox": [ + 205, + 108, + 251, + 133 + ], + "lines": [ + { + "bbox": [ + 205, + 108, + 251, + 133 + ], + "spans": [ + { + "bbox": [ + 205, + 108, + 251, + 133 + ], + "type": "image", + "image_path": "3820b6dc4862582190abe227faa866e5fd3e6853288b7399e93abd783d5c8264.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 135, + 236, + 142 + ], + "lines": [ + { + "bbox": [ + 220, + 135, + 236, + 142 + ], + "spans": [ + { + "bbox": [ + 220, + 135, + 236, + 142 + ], + "type": "text", + "content": "HAIR" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 254, + 108, + 301, + 133 + ], + "blocks": [ + { + "bbox": [ + 254, + 108, + 301, + 133 + ], + "lines": [ + { + "bbox": [ + 254, + 108, + 301, + 133 + ], + "spans": [ + { + "bbox": [ + 254, + 108, + 301, + 133 + ], + "type": "image", + "image_path": "4e4826a8146fa398f18be36ff90e29f8f63d20e8aa8cf18f43b2215ddf5d85e6.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 135, + 285, + 142 + ], + "lines": [ + { + "bbox": [ + 271, + 135, + 285, + 142 + ], + "spans": [ + { + "bbox": [ + 271, + 135, + 285, + 142 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 59, + 143, + 152, + 205 + ], + "blocks": [ + { + "bbox": [ + 59, + 143, + 152, + 205 + ], + "lines": [ + { + "bbox": [ + 59, + 143, + 152, + 205 + ], + "spans": [ + { + "bbox": [ + 59, + 143, + 152, + 205 + ], + "type": "image", + "image_path": "57287db71fd0c582a97ad0048c81dd9b229d8a7d75762b03bc5c760525202c5b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 206, + 137, + 213 + ], + "lines": [ + { + "bbox": [ + 77, + 206, + 137, + 213 + ], + "spans": [ + { + "bbox": [ + 77, + 206, + 137, + 213 + ], + "type": "text", + "content": "Low light enhancement" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 64, + 215, + 544, + 225 + ], + "lines": [ + { + "bbox": [ + 64, + 215, + 544, + 225 + ], + "spans": [ + { + "bbox": [ + 64, + 215, + 544, + 225 + ], + "type": "text", + "content": "Figure 3. A comparison of visual results for four types of degradation removal with other state-of-the-art (SOTA) all-in-one methods." + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 155, + 144, + 201, + 170 + ], + "blocks": [ + { + "bbox": [ + 155, + 144, + 201, + 170 + ], + "lines": [ + { + "bbox": [ + 155, + 144, + 201, + 170 + ], + "spans": [ + { + "bbox": [ + 155, + 144, + 201, + 170 + ], + "type": "image", + "image_path": "4038e26fadfb1aeea679d300e3b75c0d8919f2a9fb71de580f1d61fcf21ea990.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 205, + 144, + 251, + 170 + ], + "blocks": [ + { + "bbox": [ + 205, + 144, + 251, + 170 + ], + "lines": [ + { + "bbox": [ + 205, + 144, + 251, + 170 + ], + "spans": [ + { + "bbox": [ + 205, + 144, + 251, + 170 + ], + "type": "image", + "image_path": "0414d68c91ff50d09bb4c5fd1bafab0831d3e5d3585462d3d7a56d1efc12f4ff.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 254, + 144, + 301, + 170 + ], + "blocks": [ + { + "bbox": [ + 254, + 144, + 301, + 170 + ], + "lines": [ + { + "bbox": [ + 254, + 144, + 301, + 170 + ], + "spans": [ + { + "bbox": [ + 254, + 144, + 301, + 170 + ], + "type": "image", + "image_path": "86b37eaf6ee2b186d50449a933fac29c005824c8c73f25511d0df54b2b43ef43.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 155, + 178, + 201, + 204 + ], + "blocks": [ + { + "bbox": [ + 165, + 172, + 194, + 178 + ], + "lines": [ + { + "bbox": [ + 165, + 172, + 194, + 178 + ], + "spans": [ + { + "bbox": [ + 165, + 172, + 194, + 178 + ], + "type": "text", + "content": "Histformer" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 155, + 178, + 201, + 204 + ], + "lines": [ + { + "bbox": [ + 155, + 178, + 201, + 204 + ], + "spans": [ + { + "bbox": [ + 155, + 178, + 201, + 204 + ], + "type": "image", + "image_path": "d98c9d0f5e6a442d25a6b67bdb9239efbb2454864c9efde931857f9eacf69248.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 205, + 192, + 212 + ], + "lines": [ + { + "bbox": [ + 162, + 205, + 192, + 212 + ], + "spans": [ + { + "bbox": [ + 162, + 205, + 192, + 212 + ], + "type": "text", + "content": "Gridformer" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 205, + 178, + 251, + 204 + ], + "blocks": [ + { + "bbox": [ + 216, + 172, + 241, + 178 + ], + "lines": [ + { + "bbox": [ + 216, + 172, + 241, + 178 + ], + "spans": [ + { + "bbox": [ + 216, + 172, + 241, + 178 + ], + "type": "text", + "content": "PromptIR" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 205, + 178, + 251, + 204 + ], + "lines": [ + { + "bbox": [ + 205, + 178, + 251, + 204 + ], + "spans": [ + { + "bbox": [ + 205, + 178, + 251, + 204 + ], + "type": "image", + "image_path": "a5d6ac59df501d7b9aad69041876c4697605528f6504e6c44b9da7ef325b3ffa.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 205, + 236, + 212 + ], + "lines": [ + { + "bbox": [ + 220, + 205, + 236, + 212 + ], + "spans": [ + { + "bbox": [ + 220, + 205, + 236, + 212 + ], + "type": "text", + "content": "HAIR" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 254, + 178, + 301, + 204 + ], + "blocks": [ + { + "bbox": [ + 271, + 172, + 281, + 178 + ], + "lines": [ + { + "bbox": [ + 271, + 172, + 281, + 178 + ], + "spans": [ + { + "bbox": [ + 271, + 172, + 281, + 178 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 254, + 178, + 301, + 204 + ], + "lines": [ + { + "bbox": [ + 254, + 178, + 301, + 204 + ], + "spans": [ + { + "bbox": [ + 254, + 178, + 301, + 204 + ], + "type": "image", + "image_path": "c0c7fa20af61602243b7608b5d6ffe6bf4d1e5df5f80e8ff3cafc5cb5b4f4f3f.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 205, + 285, + 212 + ], + "lines": [ + { + "bbox": [ + 271, + 205, + 285, + 212 + ], + "spans": [ + { + "bbox": [ + 271, + 205, + 285, + 212 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 309, + 74, + 402, + 134 + ], + "blocks": [ + { + "bbox": [ + 309, + 74, + 402, + 134 + ], + "lines": [ + { + "bbox": [ + 309, + 74, + 402, + 134 + ], + "spans": [ + { + "bbox": [ + 309, + 74, + 402, + 134 + ], + "type": "image", + "image_path": "fd7940648b1c40934a6c388da3a088df134a594435e62365d11e4c933bcb7b4e.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 136, + 375, + 142 + ], + "lines": [ + { + "bbox": [ + 331, + 136, + 375, + 142 + ], + "spans": [ + { + "bbox": [ + 331, + 136, + 375, + 142 + ], + "type": "text", + "content": "Image denoising" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 405, + 74, + 451, + 100 + ], + "blocks": [ + { + "bbox": [ + 405, + 74, + 451, + 100 + ], + "lines": [ + { + "bbox": [ + 405, + 74, + 451, + 100 + ], + "spans": [ + { + "bbox": [ + 405, + 74, + 451, + 100 + ], + "type": "image", + "image_path": "2e54fba6b683998791ccb45d7addac44694a3e6c6ebd6d06b0358e9a05479d39.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 454, + 74, + 501, + 100 + ], + "blocks": [ + { + "bbox": [ + 454, + 74, + 501, + 100 + ], + "lines": [ + { + "bbox": [ + 454, + 74, + 501, + 100 + ], + "spans": [ + { + "bbox": [ + 454, + 74, + 501, + 100 + ], + "type": "image", + "image_path": "f0f9cd3a714accb1c416174a91a640a43a3184d482e0e71453c25f96b6041a0f.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 101, + 492, + 108 + ], + "lines": [ + { + "bbox": [ + 466, + 101, + 492, + 108 + ], + "spans": [ + { + "bbox": [ + 466, + 101, + 492, + 108 + ], + "type": "text", + "content": "PromptIR" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 504, + 74, + 551, + 100 + ], + "blocks": [ + { + "bbox": [ + 504, + 74, + 551, + 100 + ], + "lines": [ + { + "bbox": [ + 504, + 74, + 551, + 100 + ], + "spans": [ + { + "bbox": [ + 504, + 74, + 551, + 100 + ], + "type": "image", + "image_path": "699f9c65003e99886c92bc71e0236fe76039d87e11c0a74e1dd74acbeb6d923b.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 405, + 109, + 451, + 133 + ], + "blocks": [ + { + "bbox": [ + 414, + 102, + 444, + 108 + ], + "lines": [ + { + "bbox": [ + 414, + 102, + 444, + 108 + ], + "spans": [ + { + "bbox": [ + 414, + 102, + 444, + 108 + ], + "type": "text", + "content": "Histformer" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 405, + 109, + 451, + 133 + ], + "lines": [ + { + "bbox": [ + 405, + 109, + 451, + 133 + ], + "spans": [ + { + "bbox": [ + 405, + 109, + 451, + 133 + ], + "type": "image", + "image_path": "3577901648ab33fd82bed731f716ca9e461ebf3ee5e8a1d2c8a9e998ccae6f3d.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 454, + 109, + 501, + 133 + ], + "blocks": [ + { + "bbox": [ + 454, + 109, + 501, + 133 + ], + "lines": [ + { + "bbox": [ + 454, + 109, + 501, + 133 + ], + "spans": [ + { + "bbox": [ + 454, + 109, + 501, + 133 + ], + "type": "image", + "image_path": "84d0423fbc73d628b6fea574dfef7dfd36de2a734837364044078a42a10ecbe8.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 503, + 109, + 551, + 133 + ], + "blocks": [ + { + "bbox": [ + 522, + 102, + 531, + 108 + ], + "lines": [ + { + "bbox": [ + 522, + 102, + 531, + 108 + ], + "spans": [ + { + "bbox": [ + 522, + 102, + 531, + 108 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 503, + 109, + 551, + 133 + ], + "lines": [ + { + "bbox": [ + 503, + 109, + 551, + 133 + ], + "spans": [ + { + "bbox": [ + 503, + 109, + 551, + 133 + ], + "type": "image", + "image_path": "df0c1d6c3ec1be1c93c778315ded93ff3c9ae8e0d43509c9eca8b59b32e1ff3e.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 309, + 144, + 403, + 205 + ], + "blocks": [ + { + "bbox": [ + 309, + 144, + 403, + 205 + ], + "lines": [ + { + "bbox": [ + 309, + 144, + 403, + 205 + ], + "spans": [ + { + "bbox": [ + 309, + 144, + 403, + 205 + ], + "type": "image", + "image_path": "8a7b2b617fe2b52c4301a8a8e2b001bef24bf141f3307bf83387682928cf8de0.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 206, + 373, + 213 + ], + "lines": [ + { + "bbox": [ + 331, + 206, + 373, + 213 + ], + "spans": [ + { + "bbox": [ + 331, + 206, + 373, + 213 + ], + "type": "text", + "content": "Image dehazing" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 405, + 144, + 451, + 170 + ], + "blocks": [ + { + "bbox": [ + 414, + 136, + 444, + 142 + ], + "lines": [ + { + "bbox": [ + 414, + 136, + 444, + 142 + ], + "spans": [ + { + "bbox": [ + 414, + 136, + 444, + 142 + ], + "type": "text", + "content": "Gridformer" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 405, + 144, + 451, + 170 + ], + "lines": [ + { + "bbox": [ + 405, + 144, + 451, + 170 + ], + "spans": [ + { + "bbox": [ + 405, + 144, + 451, + 170 + ], + "type": "image", + "image_path": "fac6ddff917d8339bf314872342a0d65790262bc3aae1d839f2c826da2662236.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 454, + 144, + 501, + 170 + ], + "blocks": [ + { + "bbox": [ + 468, + 136, + 484, + 142 + ], + "lines": [ + { + "bbox": [ + 468, + 136, + 484, + 142 + ], + "spans": [ + { + "bbox": [ + 468, + 136, + 484, + 142 + ], + "type": "text", + "content": "HAIR" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 454, + 144, + 501, + 170 + ], + "lines": [ + { + "bbox": [ + 454, + 144, + 501, + 170 + ], + "spans": [ + { + "bbox": [ + 454, + 144, + 501, + 170 + ], + "type": "image", + "image_path": "c10c595537485a98cc87cbb5c7de71e9faf68bff6e8c678b26c201cd408d8d10.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 504, + 144, + 551, + 170 + ], + "blocks": [ + { + "bbox": [ + 523, + 136, + 535, + 142 + ], + "lines": [ + { + "bbox": [ + 523, + 136, + 535, + 142 + ], + "spans": [ + { + "bbox": [ + 523, + 136, + 535, + 142 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 504, + 144, + 551, + 170 + ], + "lines": [ + { + "bbox": [ + 504, + 144, + 551, + 170 + ], + "spans": [ + { + "bbox": [ + 504, + 144, + 551, + 170 + ], + "type": "image", + "image_path": "fa7002e0da7863c87e766d3c03e621cf9cccca59437bb48417a59b7786ce4a5e.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 405, + 178, + 452, + 203 + ], + "blocks": [ + { + "bbox": [ + 414, + 172, + 444, + 178 + ], + "lines": [ + { + "bbox": [ + 414, + 172, + 444, + 178 + ], + "spans": [ + { + "bbox": [ + 414, + 172, + 444, + 178 + ], + "type": "text", + "content": "Histformer" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 405, + 178, + 452, + 203 + ], + "lines": [ + { + "bbox": [ + 405, + 178, + 452, + 203 + ], + "spans": [ + { + "bbox": [ + 405, + 178, + 452, + 203 + ], + "type": "image", + "image_path": "5177262119abe129d47c65b526ac3485076793234eb3c1d6e53ba4c75600d472.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 205, + 444, + 213 + ], + "lines": [ + { + "bbox": [ + 414, + 205, + 444, + 213 + ], + "spans": [ + { + "bbox": [ + 414, + 205, + 444, + 213 + ], + "type": "text", + "content": "Gridformer" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 454, + 178, + 501, + 203 + ], + "blocks": [ + { + "bbox": [ + 467, + 171, + 492, + 178 + ], + "lines": [ + { + "bbox": [ + 467, + 171, + 492, + 178 + ], + "spans": [ + { + "bbox": [ + 467, + 171, + 492, + 178 + ], + "type": "text", + "content": "PromptIR" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 454, + 178, + 501, + 203 + ], + "lines": [ + { + "bbox": [ + 454, + 178, + 501, + 203 + ], + "spans": [ + { + "bbox": [ + 454, + 178, + 501, + 203 + ], + "type": "image", + "image_path": "4c554c03c9650b99b6d0c47fd1dbd752e37b7bbc0cac734bbc79c460a0a21b16.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 205, + 484, + 212 + ], + "lines": [ + { + "bbox": [ + 468, + 205, + 484, + 212 + ], + "spans": [ + { + "bbox": [ + 468, + 205, + 484, + 212 + ], + "type": "text", + "content": "HAIR" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 504, + 178, + 551, + 203 + ], + "blocks": [ + { + "bbox": [ + 522, + 172, + 531, + 178 + ], + "lines": [ + { + "bbox": [ + 522, + 172, + 531, + 178 + ], + "spans": [ + { + "bbox": [ + 522, + 172, + 531, + 178 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 504, + 178, + 551, + 203 + ], + "lines": [ + { + "bbox": [ + 504, + 178, + 551, + 203 + ], + "spans": [ + { + "bbox": [ + 504, + 178, + 551, + 203 + ], + "type": "image", + "image_path": "42ccda51ace6b43f5c5d22951878bc907a035daf5ffc7610b8884b0bbda677e2.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 522, + 206, + 535, + 213 + ], + "lines": [ + { + "bbox": [ + 522, + 206, + 535, + 213 + ], + "spans": [ + { + "bbox": [ + 522, + 206, + 535, + 213 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + } + ], + "index": 54 + }, + { + "type": "table", + "bbox": [ + 82, + 266, + 530, + 384 + ], + "blocks": [ + { + "bbox": [ + 55, + 236, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 236, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 555, + 258 + ], + "type": "text", + "content": "Table 1. Comparison to state-of-the-art on four degradations. PSNR (dB, ↑), SSIM (↑), and LPIPS (↓), and FS represents full-size 4K image inference. FLOPs are computed for an input size of " + }, + { + "bbox": [ + 55, + 236, + 555, + 258 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 236, + 555, + 258 + ], + "type": "text", + "content": ". Best and second best performances are highlighted." + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 266, + 530, + 384 + ], + "lines": [ + { + "bbox": [ + 82, + 266, + 530, + 384 + ], + "spans": [ + { + "bbox": [ + 82, + 266, + 530, + 384 + ], + "type": "table", + "html": "
MethodFSFLOPsParams.Low LightDeblurringDehazingDenoisingAverage
UHD-LLUHD-blurUHD-haze\\( UHDN_{\\sigma=15} \\)\\( UHDN_{\\sigma=25} \\)\\( UHDN_{\\sigma=50} \\)
AIRNet [14]✘301G9M19.24.80921.89.75718.37.81221.33.88720.78.78418.79.47520.07.754.2843
IDR [41]✘88G15.3M23.12.91024.67.79319.12.76827.48.91225.86.87224.57.65424.14.822.2684
PromptIR [20]✘158G33M23.44.90225.77.78219.97.72728.43.92426.74.89823.72.58424.68.803.2571
CAPTNet [8]✘25G24.3M23.96.92026.11.79819.46.86825.58.86523.24.88421.98.50823.39.809.3466
NDR-Restore [38]✘196G36.9M23.84.89424.25.80220.08.89225.62.91224.37.89722.94.66923.52.846.3126
Gridformer [33]✘367G34M23.12.89825.82.78319.24.86936.04.93731.72.89826.24.62327.03.836.3754
DiffUIR-L [44]✘10G36.2M21.56.81223.85.74318.28.86436.84.93832.42.89726.08.64826.51.818.2564
Histoformer [27]✘91G16.6M23.22.90825.62.78219.78.90326.88.84525.64.87423.13.65924.04.829.3524
adalR [4]✘147G28.7M23.57.91626.35.80118.44.90132.84.92130.48.90126.48.67226.36.857.3429
HAIR [1]✘41G29M25.75.92225.78.79820.00.89435.54.91630.84.87826.26.65727.36.847.2822
Ours✓4G1.6M27.11.92526.48.80320.94.92338.94.97533.99.90327.95.67729.23.868.2541
", + "image_path": "f005521bdf462bb1697f5180eb9e3436680b3628632694f700b1e941ba60109a.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "table_body" + } + ], + "index": 58 + }, + { + "type": "table", + "bbox": [ + 82, + 423, + 530, + 541 + ], + "blocks": [ + { + "bbox": [ + 55, + 393, + 555, + 415 + ], + "lines": [ + { + "bbox": [ + 55, + 393, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 555, + 415 + ], + "type": "text", + "content": "Table 2. Comparison to state-of-the-art on six degradations. PSNR (dB, ↑), SSIM (↑), LPIPS (↓) and FS represents full-size 4K image inference. FLOPs are computed for an input size of " + }, + { + "bbox": [ + 55, + 393, + 555, + 415 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 393, + 555, + 415 + ], + "type": "text", + "content": ". Best and second best performances are highlighted." + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 423, + 530, + 541 + ], + "lines": [ + { + "bbox": [ + 82, + 423, + 530, + 541 + ], + "spans": [ + { + "bbox": [ + 82, + 423, + 530, + 541 + ], + "type": "table", + "html": "
MethodFSFLOPsParams.Low LightDeblurringDehazingDenoisingDerainingDesnowingAverage
UHD-LLUHD-blurUHD-hazeUHDN\\(_{\\sigma=50}\\)UHD-rainUHD-snow
AIRNet [14]✘301G9M22.68.88723.52.87618.24.84622.38.87626.35.87627.38.92423.43.874.1861
IDR [41]✘88G15.3M24.33.91525.64.78818.68.87929.64.90628.82.90630.48.94526.27.890.1912
PromptIR [20]✘158G33M23.3.91126.48.80520.14.90124.88.83528.89.89730.78.96625.74.886.2155
CAPTNet [8]✘25G24.3M24.97.92126.32.79620.32.90321.64.56929.34.90832.21.97425.80.845.2861
NDR-Restore [38]✘196G36.9M25.12.88525.64.79119.21.89631.44.91529.24.89728.41.94826.51.889.3108
Gridformer [33]✘367G34M23.92.89825.68.78218.87.88932.86.91529.37.90428.24.94226.49.895.2321
DiffUIR-L [44]✘10G36.2M22.64.90225.08.78518.62.88933.25.92827.89.88627.36.94525.81.889.1844
Histoformer [27]✘91G16.6M25.73.91526.55.79618.73.89733.05.92427.96.88427.56.97126.59.898.1855
adalR [4]✘147G28.7M23.84.91826.86.80319.34.91032.46.92328.18.90127.72.95326.40.901.2492
HAIR [1]✘41G29M25.22.89724.77.79918.75.88332.50.91528.76.89327.89.96826.31.892.2607
Ours✓4G1.6M26.91.92426.95.80721.81.93133.73.93429.90.91532.73.97928.67.915.1839
", + "image_path": "f16ccaa3fe15b3b5c35fd426480ad6e1489606b8a3c55107fc0d0485653eee99.jpg" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "table_body" + } + ], + "index": 60 + }, + { + "type": "table", + "bbox": [ + 92, + 583, + 518, + 696 + ], + "blocks": [ + { + "bbox": [ + 55, + 551, + 555, + 572 + ], + "lines": [ + { + "bbox": [ + 55, + 551, + 555, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 555, + 572 + ], + "type": "text", + "content": "Table 3. Comparison to state-of-the-art for single degradations. PSNR (dB, ↑), SSIM (↑) and LPIPS (↓). Best and second best performances are highlighted." + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 141, + 573, + 184, + 582 + ], + "lines": [ + { + "bbox": [ + 141, + 573, + 184, + 582 + ], + "spans": [ + { + "bbox": [ + 141, + 573, + 184, + 582 + ], + "type": "text", + "content": "(a) Low light" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 281, + 573, + 329, + 582 + ], + "lines": [ + { + "bbox": [ + 281, + 573, + 329, + 582 + ], + "spans": [ + { + "bbox": [ + 281, + 573, + 329, + 582 + ], + "type": "text", + "content": "(b) Deblurring" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 426, + 573, + 469, + 582 + ], + "lines": [ + { + "bbox": [ + 426, + 573, + 469, + 582 + ], + "spans": [ + { + "bbox": [ + 426, + 573, + 469, + 582 + ], + "type": "text", + "content": "(c) Dehazing" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 92, + 583, + 518, + 696 + ], + "lines": [ + { + "bbox": [ + 92, + 583, + 518, + 696 + ], + "spans": [ + { + "bbox": [ + 92, + 583, + 518, + 696 + ], + "type": "table", + "html": "
MethodFSUHD-LLLIPIPSMethodFSUHD-blurLIPIPSMethodFSUHD-hazeLIPIPS
Restormer [40]✘21.54.843.3608Restormer [40]✘25.21.752.3695Restormer [40]✘12.72.693.4560
LLformer [32]✘24.06.858.3516Uformer [35]✘25.26.751.3851Uformer [35]✘19.83.921.4220
UHDFour [15]✓26.22.900.2390Stripformer [28]✘25.05.750.3740DehazeFormer [26]✘15.37.737.3998
UHDFformer [29]✓27.11.927.2240FFTformer [12]✘25.41.757.3708TaylorFormer [22]✘20.99.919.3124
LMAR [39]✓26.27.919.2248UHDformer [29]✓28.82.844.2350UHDformer [29]✓22.58.942.1188
AirNet [14]✘24.61.827.3841AirNet✘23.48.824.3942AirNet✘18.44.848.4018
PromptIR [20]✘24.81.912.3116PromptIR✘23.24.829.3754PromptIR✘20.21.927.3642
Histoformer [27]✘25.46.894.3213Histoformer✘27.12.848.2986Histoformer✘21.72.948.2942
HAIR [1]✘26.08.901.2964HAIR✘26.74.847.3867HAIR✘20.38.912.2412
Ours✓27.22.929.2042Ours✓28.91.851.2129Ours✓23.24.953.1086
", + "image_path": "cb9a2adf570c11b2699299dbb89f0c33716f56e4aa4a4731950ab48d5527522b.jpg" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "table_body" + } + ], + "index": 65 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23127" + } + ] + } + ], + "index": 66 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 281, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 281, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 281, + 85 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 217, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 217, + 102 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 217, + 102 + ], + "type": "text", + "content": "4.2.1. All-in-One: Four Degradations" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 104, + 297, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 104, + 297, + 236 + ], + "spans": [ + { + "bbox": [ + 54, + 104, + 297, + 236 + ], + "type": "text", + "content": "We compare our UHD-Processor with specialized all-in-one restoration methods, trained on four degradation types: dehazing, denoising, deblurring, and low-light enhancement. As shown in Tab. 1, our method achieves state-of-the-art (SOTA) performance on both individual datasets and overall metrics. Specifically, the average PSNR improves by 1.93 dB over the second-best method, HAIR, while our method requires only " + }, + { + "bbox": [ + 54, + 104, + 297, + 236 + ], + "type": "inline_equation", + "content": "5.5\\%" + }, + { + "bbox": [ + 54, + 104, + 297, + 236 + ], + "type": "text", + "content": " of the parameters and " + }, + { + "bbox": [ + 54, + 104, + 297, + 236 + ], + "type": "inline_equation", + "content": "9.7\\%" + }, + { + "bbox": [ + 54, + 104, + 297, + 236 + ], + "type": "text", + "content": " of the FLOPS of Histoformer. Visual results, shown in Figure 3, further demonstrate that our approach effectively achieves the best degradation removal." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 242, + 209, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 209, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 209, + 255 + ], + "type": "text", + "content": "4.2.2. All-in-One: Six Degradations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 257, + 297, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 257, + 297, + 376 + ], + "spans": [ + { + "bbox": [ + 54, + 257, + 297, + 376 + ], + "type": "text", + "content": "We further extended the experiment to include six degradation types by introducing image desnow and image derain, to validate the effectiveness of our method in a more complex all-in-one setup. As shown in Tab. 2, our method continues to achieve SOTA performance on both individual datasets and average metrics. Specifically, the average PSNR improves by " + }, + { + "bbox": [ + 54, + 257, + 297, + 376 + ], + "type": "inline_equation", + "content": "1.7\\mathrm{dB}" + }, + { + "bbox": [ + 54, + 257, + 297, + 376 + ], + "type": "text", + "content": " over the second-best method, Histoformer, while the number of parameters and FLOPS are only " + }, + { + "bbox": [ + 54, + 257, + 297, + 376 + ], + "type": "inline_equation", + "content": "9.6\\%" + }, + { + "bbox": [ + 54, + 257, + 297, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 257, + 297, + 376 + ], + "type": "inline_equation", + "content": "4.3\\%" + }, + { + "bbox": [ + 54, + 257, + 297, + 376 + ], + "type": "text", + "content": " respectively. More visual comparison results can be found in the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 383, + 200, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 200, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 200, + 396 + ], + "type": "text", + "content": "4.2.3. Single-Degradation Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 398, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 398, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 54, + 398, + 296, + 471 + ], + "type": "text", + "content": "We further evaluate the performance of our method by training instances under a single degradation protocol, comparing them with both task-specific methods and all-in-one approaches. The results, shown in Tab. 3, demonstrate that our method still achieves the best performance, validating the effectiveness of our approach." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 478, + 149, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 149, + 490 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 149, + 490 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 495, + 261, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 261, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 261, + 508 + ], + "type": "text", + "content": "4.3.1. Frequency-Progressive Learning Strategy" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 510, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 297, + 713 + ], + "type": "text", + "content": "As shown in Tab. 4, we validate the effectiveness of the proposed Frequency Domain Progressive Learning (FDPL). First, compared to directly training on mixed data, using FDPL results in a significant improvement of " + }, + { + "bbox": [ + 55, + 510, + 297, + 713 + ], + "type": "inline_equation", + "content": "1.99\\mathrm{dB}" + }, + { + "bbox": [ + 55, + 510, + 297, + 713 + ], + "type": "text", + "content": " PSNR. Next, we evaluate the sequential learning (SL) strategy proposed in [13], which advocates for learning local degradations before global ones. We test two learning orders: SL1 (noise " + }, + { + "bbox": [ + 55, + 510, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 510, + 297, + 713 + ], + "type": "text", + "content": " blur " + }, + { + "bbox": [ + 55, + 510, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 510, + 297, + 713 + ], + "type": "text", + "content": " low-light " + }, + { + "bbox": [ + 55, + 510, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 510, + 297, + 713 + ], + "type": "text", + "content": " haze) and SL2 (blur " + }, + { + "bbox": [ + 55, + 510, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 510, + 297, + 713 + ], + "type": "text", + "content": " noise " + }, + { + "bbox": [ + 55, + 510, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 510, + 297, + 713 + ], + "type": "text", + "content": " haze " + }, + { + "bbox": [ + 55, + 510, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 510, + 297, + 713 + ], + "type": "text", + "content": " low-light). The experimental results show that not all local-to-global learning orders lead to performance gains. In fact, SL1 even caused some performance degradation, while SL2 yielded better results. This suggests that sequential learning strategies are sensitive to the order of tasks, and the optimal sequence may not only depend on the tasks themselves but also on the specific data set, scene, or degradation severity. Therefore, it is challenging to define a universally effective task order. In contrast, our" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 353, + 83, + 515, + 119 + ], + "blocks": [ + { + "bbox": [ + 344, + 71, + 524, + 83 + ], + "lines": [ + { + "bbox": [ + 344, + 71, + 524, + 83 + ], + "spans": [ + { + "bbox": [ + 344, + 71, + 524, + 83 + ], + "type": "text", + "content": "Table 4. Ablation Study of the Learning Strategy." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 353, + 83, + 515, + 119 + ], + "lines": [ + { + "bbox": [ + 353, + 83, + 515, + 119 + ], + "spans": [ + { + "bbox": [ + 353, + 83, + 515, + 119 + ], + "type": "table", + "html": "
MethodMixw/SL1w/SL2w/FDPL (Ours)
PSNR27.2427.0828.1229.23
SSIM0.8530.8460.8720.874
", + "image_path": "09fe57beb67b5dcbdb70d5a7abfb0cad12b6c48de5f750dc46af525b68c79cd3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 315, + 145, + 556, + 182 + ], + "blocks": [ + { + "bbox": [ + 317, + 134, + 549, + 144 + ], + "lines": [ + { + "bbox": [ + 317, + 134, + 549, + 144 + ], + "spans": [ + { + "bbox": [ + 317, + 134, + 549, + 144 + ], + "type": "text", + "content": "Table 5. Ablation Study of Efficient Adaptive Prompt Learning." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 145, + 556, + 182 + ], + "lines": [ + { + "bbox": [ + 315, + 145, + 556, + 182 + ], + "spans": [ + { + "bbox": [ + 315, + 145, + 556, + 182 + ], + "type": "table", + "html": "
Methodw/o DALPw/o DDPw/o DTPw/o DSRw/o MCKGw/o DFSPOurs
PSNR27.8628.4928.3428.8428.7228.2129.23
SSIM0.8620.8660.8580.8640.8620.8650.874
", + "image_path": "d934e023cd5d9bbd1619d83c30fb43ec15460961ee574e5ecb1cf38ae2a6a4b0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 206, + 555, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 206, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 555, + 266 + ], + "type": "text", + "content": "FDPL strategy is task-agnostic and offers a more generalizable learning framework. Experimental results confirm that FDPL leads to more significant performance improvements. Further discussions on sequential learning can be found in the supplementary materials." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 279, + 496, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 496, + 292 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 496, + 292 + ], + "type": "text", + "content": "4.3.2. Efficient Adaptive Prompt Learning" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 296, + 556, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 296, + 556, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 556, + 524 + ], + "type": "text", + "content": "The ablation study of the proposed prompt learning method, as shown in Tab. 5, demonstrates that both the DALP used in the encoding process and the DFSP used in the decoding process lead to significant performance improvements. We further analyze the components of DALP, and the results show that both the Degradation Degree Prompt (DDP) and the Degradation Type Prompt (DTP) provide more granular prompt information, improving performance consistently. The Degradation-Specific Rank (DSR) component introduces task-specific low-rank decompositions, allowing for the selection of an appropriate rank for each task, which enhances parameter efficiency while improving performance. Moreover, the Multi-Scale Convolution Kernel Group (MCKG) introduces a multi-scale convolutional branch, which strengthens the representation capacity of the prompt generation process. During inference, the multi-branch convolutions are reparameterized for efficient execution. For more detailed ablation experiments and efficiency analyses, please refer to the supplementary materials." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 536, + 388, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 388, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 388, + 548 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "content": "In this work, we proposed UHD-Processor, an efficient and effective framework for UHD all-in-one image restoration. By leveraging Frequency Domain Decoupling Progressive Learning Strategy and an efficient prompt-integrated VAE framework, our method addresses the unique challenges of UHD restoration tasks, including denoising, deblurring, deraining, dehazing, and low-light enhancement. Experimental results demonstrated that our approach significantly outperforms existing methods, achieving state-of-the-art results with reduced computational costs. The combination of progressive frequency learning and targeted prompt-based adaptation provides a strong foundation for future advancements in UHD image restoration." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23128" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 127 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (NSFC) under Grants 62225207, 62436008, 62422609 and 62276243." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 157, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 157, + 296, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 157, + 296, + 189 + ], + "spans": [ + { + "bbox": [ + 61, + 157, + 296, + 189 + ], + "type": "text", + "content": "[1] Jin Cao, Yi Cao, Li Pang, Deyu Meng, and Xiangyong Cao. Hair: Hypernetworks-based all-in-one image restoration, 2024. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 191, + 296, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 191, + 296, + 233 + ], + "spans": [ + { + "bbox": [ + 61, + 191, + 296, + 233 + ], + "type": "text", + "content": "[2] Hongming Chen, Xiang Chen, Chen Wu, Zhuoran Zheng, Jinshan Pan, and Xianping Fu. Towards ultra-high-definition image deraining: A benchmark and an efficient method, 2024. 1, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 236, + 296, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 296, + 280 + ], + "type": "text", + "content": "[3] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 281, + 296, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 281, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 281, + 296, + 323 + ], + "type": "text", + "content": "[4] Yuning Cui, Syed Waqas Zamir, Salman Khan, Alois Knoll, Mubarak Shah, and Fahad Shahbaz Khan. Adair: Adaptive all-in-one image restoration via frequency mining and modulation, 2024. 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 325, + 296, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 325, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 325, + 296, + 378 + ], + "type": "text", + "content": "[5] Senyou Deng, Wenqi Ren, Yanyang Yan, Tao Wang, Fenglong Song, and Xiaochun Cao. Multi-scale separable network for ultra-high-definition video deblurring. In the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14030-14039, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 380, + 296, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 296, + 434 + ], + "type": "text", + "content": "[6] Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13733-13742, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 436, + 296, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 296, + 501 + ], + "type": "text", + "content": "[7] Xiaohan Ding, Yiyuan Zhang, Yixiao Ge, Sijie Zhao, Lin Song, Xiangyu Yue, and Ying Shan. Unireplknet: A universal perception large-kernel convnet for audio video point cloud time-series and image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5513-5524, 2024. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 502, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 296, + 544 + ], + "type": "text", + "content": "[8] Hu Gao, Jing Yang, Ning Wang, Jingfan Yang, Ying Zhang, and Depeng Dang. Prompt-based all-in-one image restoration using cnns and transformer. arXiv preprint arXiv:2309.03063, 2023. 3, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 547, + 296, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 547, + 296, + 600 + ], + "spans": [ + { + "bbox": [ + 62, + 547, + 296, + 600 + ], + "type": "text", + "content": "[9] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 602, + 296, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 296, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 296, + 646 + ], + "type": "text", + "content": "[10] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 678 + ], + "type": "text", + "content": "[11] Xingyu Jiang, Xiuhui Zhang, Ning Gao, and Yue Deng. When fast fourier transform meets transformer for image restoration. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "text", + "content": "[12] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In 2023 IEEE/CVF" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition (CVPR), pages 5886-5895, 2023. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 98, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 98, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 98, + 555, + 140 + ], + "type": "text", + "content": "[13] Xiangtao Kong, Chao Dong, and Lei Zhang. Towards effective multiple-in-one image restoration: A sequential and prompt learning strategy. arXiv preprint arXiv:2401.03379, 2024. 3, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 143, + 555, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 143, + 555, + 186 + ], + "spans": [ + { + "bbox": [ + 316, + 143, + 555, + 186 + ], + "type": "text", + "content": "[14] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-In-One Image Restoration for Unknown Corruption. In IEEE Conference on Computer Vision and Pattern Recognition, New Orleans, LA, 2022. 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 189, + 555, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 189, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 317, + 189, + 555, + 232 + ], + "type": "text", + "content": "[15] Chongyi Li, Chun-Le Guo, Man Zhou, Zhexin Liang, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Embedding fourier for ultra-high-definition low-light image enhancement. In ICLR, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 234, + 555, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 234, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 317, + 234, + 555, + 289 + ], + "type": "text", + "content": "[16] Xiangtai Li, Haobo Yuan, Wei Li, Henghui Ding, Size Wu, Wenwei Zhang, Yining Li, Kai Chen, and Chen Change Loy. Omg-seg: Is one model good enough for all segmentation? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27948-27959, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 292, + 554, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 292, + 554, + 324 + ], + "spans": [ + { + "bbox": [ + 316, + 292, + 554, + 324 + ], + "type": "text", + "content": "[17] Zilong Li, Yiming Lei, Chenglong Ma, Junping Zhang, and Hongming Shan. Prompt-in-prompt learning for universal image restoration, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 327, + 554, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 554, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 554, + 369 + ], + "type": "text", + "content": "[18] Ziwei Luo, Fredrik K. Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B. Schön. Controlling vision-language models for multi-task image restoration. In The Twelfth International Conference on Learning Representations, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 373, + 555, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 373, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 316, + 373, + 555, + 416 + ], + "type": "text", + "content": "[19] Anastasia Pentina, Viktoriia Sharmanska, and Christoph H Lampert. Curriculum learning of multiple tasks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5492-5500, 2015. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 418, + 555, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 418, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 316, + 418, + 555, + 461 + ], + "type": "text", + "content": "[20] Vaishnav Potlapalli, Syed Waqas Zamir, Salman Khan, and Fahad Khan. Prompt: Prompting for all-in-one image restoration. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2, 5, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 464, + 554, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 464, + 554, + 507 + ], + "spans": [ + { + "bbox": [ + 317, + 464, + 554, + 507 + ], + "type": "text", + "content": "[21] Yifan Pu, Weicong Liang, Yiduo Hao, Yuhui Yuan, Yukang Yang, Chao Zhang, Han Hu, and Gao Huang. Rank-detr for high quality object detection. Advances in Neural Information Processing Systems, 36, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 510, + 555, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 510, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 317, + 510, + 555, + 552 + ], + "type": "text", + "content": "[22] Yuwei Qiu, Kaihao Zhang, Chenxi Wang, Wenhan Luo, Hongdong Li, and Zhi Jin. Mb-taylorformer: Multi-branch efficient transformer expanded by taylor formula for image dehazing, 2023. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 555, + 555, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 555, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 317, + 555, + 555, + 609 + ], + "type": "text", + "content": "[23] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International conference on machine learning, pages 5301-5310. PMLR, 2019. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 613, + 555, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 555, + 666 + ], + "type": "text", + "content": "[24] Wenqi Ren, Senyou Deng, Kaihao Zhang, Fenglong Song, Xiaochun Cao, and Ming-Hsuan Yang. Fast ultra high-definition video deblurring via multi-scale separable network. International Journal of Computer Vision, 132(5):1817-1834, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "text", + "content": "[25] Basri Ronen, David Jacobs, Yoni Kasten, and Shira Kritchman. The convergence rate of neural networks for learned functions of different frequencies. Advances in Neural Information Processing Systems, 32, 2019. 3" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23129" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[26] Yuda Song, Zhuqing He, Hui Qian, and Xin Du. Vision transformers for single image dehazing. IEEE Transactions on Image Processing, 32:1927-1941, 2023. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 152 + ], + "type": "text", + "content": "[27] Shangquan Sun, Wenqi Ren, Xinwei Gao, Rui Wang, and Xiaochun Cao. Restoring images in adverse weather conditions via histogram transformer. In European Conference on Computer Vision, pages 111-129. Springer, 2025. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 153, + 295, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 153, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 153, + 295, + 196 + ], + "type": "text", + "content": "[28] Fu-Jen Tsai, Yan-Tsung Peng, Yen-Yu Lin, Chung-Chi Tsai, and Chia-Wen Lin. Stripformer: Strip transformer for fast image deblurring. In European conference on computer vision, pages 146-162. Springer, 2022. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 295, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 252 + ], + "type": "text", + "content": "[29] Cong Wang, Jinshan Pan, Wei Wang, Gang Fu, Siyuan Liang, Mengzhu Wang, Xiao-Ming Wu, and Jun Liu. Correlation matching transformation transformers for uhd image restoration. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5336-5344, 2024. 1, 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 295, + 308 + ], + "type": "text", + "content": "[30] Haohan Wang, Xindi Wu, Zeyi Huang, and Eric P Xing. High-frequency component helps explain the generalization of convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8684-8694, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 295, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 295, + 354 + ], + "type": "text", + "content": "[31] Liyan Wang, Cong Wang, Jinshan Pan, Xiaofeng Liu, Weixiang Zhou, Xiaoran Sun, Wei Wang, and Zhixun Su. Ultra-high-definition image restoration: New benchmarks and a dual interaction prior-driven solution, 2024. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 355, + 295, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 295, + 410 + ], + "type": "text", + "content": "[32] Tao Wang, Kaihao Zhang, Tianrun Shen, Wenhan Luo, Bjorn Stenger, and Tong Lu. Ultra-high-definition low-light image enhancement: A benchmark and transformer-based method. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2654-2662, 2023. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 295, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 295, + 475 + ], + "type": "text", + "content": "[33] Tao Wang, Kaihao Zhang, Ziqian Shao, Wenhan Luo, Bjorn Stenger, Tong Lu, Tae-Kyun Kim, Wei Liu, and Hongdong Li. Gridformer: Residual dense transformer with grid structure for image restoration in adverse weather conditions. International Journal of Computer Vision, pages 1-23, 2024. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 295, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 295, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 295, + 511 + ], + "type": "text", + "content": "[34] Xin Wang, Yudong Chen, and Wenwu Zhu. A survey on curriculum learning. IEEE transactions on pattern analysis and machine intelligence, 44(9):4555-4576, 2021. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 295, + 567 + ], + "type": "text", + "content": "[35] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17683-17693, 2022. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 295, + 602 + ], + "type": "text", + "content": "[36] Boxue Xiao, Zhuoran Zheng, Xiang Chen, Chen Lv, Yunliang Zhuang, and Tao Wang. Single uhd image dehazing via interpretable pyramid network, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 602, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 295, + 667 + ], + "type": "text", + "content": "[37] Zhi-Qin John Xu, Yaoyu Zhang, and Yanyang Xiao. Training behavior of deep neural network in frequency domain. In Neural Information Processing: 26th International Conference, ICONIP 2019, Sydney, NSW, Australia, December 12–15, 2019, Proceedings, Part I 26, pages 264–274. Springer, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "type": "text", + "content": "[38] Mingde Yao, Ruikang Xu, Yuanshen Guan, Jie Huang, and Zhiwei Xiong. Neural degradation representation learning for all-in-one image restoration. IEEE Transactions on Image Processing, 2024. 2, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 563 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 137 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 137 + ], + "type": "text", + "content": "[39] Wei Yu, Jie Huang, Bing Li, Kaiwen Zheng, Qi Zhu, Man Zhou, and Feng Zhao. Empowering resampling operation for ultra-high-definition image enhancement with model-aware guidance. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25722-25731, 2024. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 555, + 183 + ], + "type": "text", + "content": "[40] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 2, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 185, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 555, + 239 + ], + "type": "text", + "content": "[41] Jinghao Zhang, Jie Huang, Mingde Yao, Zizheng Yang, Hu Yu, Man Zhou, and Feng Zhao. Ingredient-oriented multidegradation learning for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5825-5835, 2023. 2, 3, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 240, + 555, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 555, + 295 + ], + "type": "text", + "content": "[42] Kaihao Zhang, Dongxu Li, Wenhan Luo, Wenqi Ren, Bjorn Stenger, Wei Liu, Hongdong Li, and Yang Ming-Hsuan. Benchmarking ultra-high-definition image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 297, + 555, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 555, + 350 + ], + "type": "text", + "content": "[43] Wenlong Zhang, Xiaohui Li, SHI Guangyuan, Xiangyu Chen, Yu Qiao, Xiaoyun Zhang, Xiao-Ming Wu, and Chao Dong. Real-world image super-resolution as multi-task learning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 352, + 555, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 352, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 352, + 555, + 418 + ], + "type": "text", + "content": "[44] Dian Zheng, Xiao-Ming Wu, Shuzhou Yang, Jian Zhang, Jian-Fang Hu, and Wei-Shi Zheng. Selective hourglass mapping for universal image restoration based on diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25445-25455, 2024. 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 419, + 555, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 419, + 555, + 452 + ], + "spans": [ + { + "bbox": [ + 316, + 419, + 555, + 452 + ], + "type": "text", + "content": "[45] Zhuoran Zheng and Xiuyi Jia. Uhd image deblurring via multi-scale cubic-mixer. arXiv preprint arXiv:2206.03678, 2022. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 453, + 555, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 555, + 508 + ], + "type": "text", + "content": "[46] Zhuoran Zheng, Wenqi Ren, Xiaochun Cao, Xiaobin Hu, Tao Wang, Fenglong Song, and Xiuyi Jia. Ultra-high-definition image dehazing via multi-guided bilateral learning. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16180-16189, 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 508, + 555, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 508, + 555, + 563 + ], + "spans": [ + { + "bbox": [ + 316, + 508, + 555, + 563 + ], + "type": "text", + "content": "[47] Yanqi Zhou, Tao Lei, Hanxiao Liu, Nan Du, Yanping Huang, Vincent Zhao, Andrew M Dai, Quoc V Le, James Laudon, et al. Mixture-of-experts with expert choice routing. Advances in Neural Information Processing Systems, 35:7103-7114, 2022. 2" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23130" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/UIBDiffusion_ Universal Imperceptible Backdoor Attack for Diffusion Models/7f0574b2-e282-4c3e-a072-3d64d026983a_content_list.json b/2025/UIBDiffusion_ Universal Imperceptible Backdoor Attack for Diffusion Models/7f0574b2-e282-4c3e-a072-3d64d026983a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..21e54369943595c6da5ce90df35b997c3452fd32 --- /dev/null +++ b/2025/UIBDiffusion_ Universal Imperceptible Backdoor Attack for Diffusion Models/7f0574b2-e282-4c3e-a072-3d64d026983a_content_list.json @@ -0,0 +1,1591 @@ +[ + { + "type": "text", + "text": "UIBDiffusion: Universal Imperceptible Backdoor Attack for Diffusion Models", + "text_level": 1, + "bbox": [ + 106, + 128, + 890, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuning Han $^{1,*}$ Bingyin Zhao $^{2,*}$ Rui Chu $^{3}$ Feng Luo $^{4}$ Biplab Sikdar $^{2}$ Yingjie Lao $^{3,\\dagger}$", + "bbox": [ + 161, + 179, + 862, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Columbia University $^{2}$ National University of Singapore $^{3}$ Tufts University $^{4}$ Clemson University", + "bbox": [ + 99, + 214, + 913, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent studies show that diffusion models (DMs) are vulnerable to backdoor attacks. Existing backdoor attacks impose unconcealed triggers (e.g., a gray box and eyeglasses) that contain evident patterns, rendering remarkable attack effects yet easy detection upon human inspection and defensive algorithms. While it is possible to improve stealthiness by reducing the strength of the backdoor, doing so can significantly compromise its generality and effectiveness. In this paper, we propose UIBDiffusion, the universal imperceptible backdoor attack for diffusion models, which allows us to achieve superior attack and generation performance while evading state-of-the-art defenses. We propose a novel trigger generation approach based on universal adversarial perturbations (UAPs) and reveal that such perturbations, which are initially devised for fooling pre-trained discriminative models, can be adapted as potent imperceptible backdoor triggers for DMs. We evaluate UIBDiffusion on multiple types of DMs with different kinds of samplers across various datasets and targets. Experimental results demonstrate that UIBDiffusion brings three advantages: 1) Universality, the imperceptible trigger is universal (i.e., image and model agnostic) where a single trigger is effective to any images and all diffusion models with different samplers; 2) Utility, it achieves comparable generation quality (e.g., FID) and even better attack success rate (i.e., ASR) at low poison rates compared to the prior works; and 3) Undetectability, UIBDiffusion is plausible to human perception and can bypass Elijah and TERD, the SOTA defenses against backdoors for DMs. Code is available at https://github.com/TheLaoLab/UIBDiffusion.", + "bbox": [ + 91, + 300, + 483, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 781, + 220, + 796 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models (DMs) [1-3] have emerged as the state-of-the-art generative paradigm in recent years and achieved unprecedented success in image [4, 5], video [6-8] text [9-11] and audio [12-17] synthesis. The success is highly", + "bbox": [ + 89, + 805, + 482, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "attributed to training the foundation model on large-scale datasets [5]. Such a practice places DMs at risk of being compromised by data poisoning attacks [18-22], where adversaries contaminate the clean training dataset with a small fraction of malicious samples and undermine the model performance. This imposes a critical security concern on the practical deployment of DMs.", + "bbox": [ + 511, + 270, + 903, + 377 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Backdoor attacks are emerging threats to DMs. Recent studies [23-25] show that DMs are vulnerable to such attacks, which implant backdoor into victim models by training them on poisoned datasets and manipulate their behavior through inputs stamped with predefined triggers. Existing backdoor attacks for DMs usually employ unconcealed triggers such as a gray box, an image of Hello Kitty, and eyeglasses in the forward and reversed (i.e., sampling) diffusion process, as shown in Fig. 1. While such trigger designs achieve superior attack performance and preserve the decent generation ability of victim models, they are easy to detect via human inspection and advanced defensive algorithms as they have obvious patterns that can be precisely remodeled via reverse engineering [26, 27]. Motivated by this observation, we aim to design a general, powerful yet stealthy trigger. However, even for the traditional backdoor attacks against discriminative models (e.g., image classifiers) [28-30], it is challenging to acquire such a trigger since it either requires complex and carefully designed generators to generate image-specific or model-specific triggers [31-38] or suffers from a relatively low attack success rate (ASR) [39, 40], or fails to evade modern defenses [41-46]. This raises an interesting question in backdoor attacks against DMs: Is there a trigger simultaneously possessing generality, effectiveness, and stealthiness for backdoor attacks against DMs?", + "bbox": [ + 511, + 381, + 906, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Driving by the question, in this work, we propose UIB-Diffusion, the universal imperceptible backdoor triggers for DMs that bring the following advantages: 1) Universality: the trigger is image-agnostic and model-agnostic so that it can be applied to arbitrary images and DMs; 2) Utility: the trigger can achieve high ASR while maintaining the generation capability of DMs; 3) Undetectability: the trigger is plausible to humans and capable of circumventing", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 875, + 220, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19186", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/489eae18e67f17ebad7b0366a582d838a4a71684ff6d96bc63b80848149dd359.jpg", + "image_caption": [ + "Figure 1. Illustrations of the forward diffusion process (top block with the yellow background) and backward diffusion process (bottom block with the blue background) of a clean diffusion model (blue dash line), VillanDiffusion [25] (red dash line) and UIBDiffusion (Ours, green dash line). The UIBDiffusion trigger $(\\tau)$ is plausible to humans in all phases from data poisoning to forward diffusion and backward diffusion while the glasses trigger $(\\mathbf{g})$ in prior works is perceptible. UIBDiffusion trigger is highly effective since it introduces a similar distribution shift as the glasses trigger, which secures the attack performance during the backward diffusion process. However, it is hard to detect as the trigger does not possess specific patterns and cannot be inverted by existing defensive algorithms. We empirically verify the effectiveness in Section 4." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 864, + 497 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "state-of-the-art defenses. Inspired by the adversarial perturbation [47] originally designed for attacking pre-trained image classifiers, we reveal that universal adversarial perturbations [48] (UAPs), a variant of adversarial perturbations, naturally hold the aforementioned properties and can be adapted as the UIBDiffusion triggers. First, UAPs are image-agnostic and model-agnostic. Second, we demonstrate that they are effective triggers because they introduce a similar distribution shift as prior works, yet they are stealthy because they do not possess clear patterns and are hard to estimate by existing trigger inversion algorithms. Moreover, they are trivial and imperceptible noises that humans can barely identify, as shown in Fig. 1. Our contributions are summarized as follows:", + "bbox": [ + 89, + 621, + 485, + 832 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, UIBDiffusion is the first imperceptible backdoor attack against diffusion models.", + "- We discover that adversarial perturbations designed for discriminative models can be adapted as backdoor trig-" + ], + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "gers for generative models. We reveal why such triggers can achieve high ASR and evade SOTA defenses.", + "bbox": [ + 526, + 621, + 903, + 650 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a practical trigger generation approach adapted from UAP, which is specifically designed to enhance universality, attack effectiveness and accessibility of triggers.", + "- We experimentally demonstrate that UIBDiffusion is highly effective towards multiple diffusion models with different samplers and achieves comparable generation quality on benign input compared to clean DMs and superior attack success rate and input stealthiness compared to prior works.", + "- While existing works face a trade-off between utility and undetectability, we empirically show that UIBDiffusion triggers are highly resistant to the SOTA trigger inversion-based defenses, effectively bypassing Elijah [26] and TERD [27], the state-of-the-art defenses against backdoors on DMs." + ], + "bbox": [ + 511, + 651, + 908, + 891 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19187", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 91, + 89, + 240, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 114, + 259, + 130 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models are latent variable models composed of two processes. The forward diffusion process imposes Gaussian noise on input data in each time step along the Markov chain and progressively diffuses the data distribution to an isotropic Gaussian distribution. The backward diffusion process is a reversed Markov chain that gradually denoises the sampled noise from the Gaussian distribution to the input data distribution. Mainstream diffusion models include DDPM [1], LDM [3] and NCSN [49-51], etc. Despite the excellent performance, DMs suffer from slow sampling during the backward process. There exists a line of works that leverage different techniques such as generalized non-Markovian chain and Ordinary Differential Equation (ODE) to improve the samplers and accelerate the generation process: DDIM [2], DEIS [52], UniPC [53], Heun [54], DPM-Solver [55] and DPM-Solver++ [7]. We comprehensively evaluate UIBDiffusion on these models and samplers to showcase the utility of our trigger.", + "bbox": [ + 89, + 138, + 485, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Backdoor Attacks and Defenses on Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 422, + 483, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Attacks. Backdoor attacks [28-30, 56], originally explored in the context of discriminative models such as image classifiers, induce models to make wrong predictions on triggered images. Although having been extensively studied for discriminative models, backdoor attacks are under-explored in diffusion models. BadDiffusion [23], TrojDiff [24] and VillanDiffusion [25] are pioneering works in this field. Unlike traditional backdoor attacks, those targeting DMs aim to force models to synthesize a targeted image upon malicious inputs (i.e., noises with triggers) while generating normal images upon benign inputs (i.e., isotropic Gaussian noises). BadDiffusion shows how to backdoor DMs with unconcealed triggers such as a gray box and a stop sign. TrojDiff proposes to use a whole image (e.g., Hello Kitty) as the trigger and demonstrates that backdoor can be generalized to DDIM and different adversarial goals. However, both attacks are limited to certain DMs (e.g., DDPM) and samplers (e.g., ancestral and DDIM). VillanDiffusion resolves the limitations and offers a unified framework for all kinds of DMs and samplers. Other works show backdoor can be injected to text-to-image DMs and activated by text prompts [57-59]. However, these attacks draw less attention from defenders as they can eliminate the backdoor by cleansing the text encoder. Thus, our attack focuses on the primary defensive objective of the defender: pixel level (i.e., image) trigger generation.", + "bbox": [ + 89, + 462, + 483, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Defenses. Backdoor attacks against DMs are shown to be immune to defenses designed for traditional image classifier backdoors [23, 25, 44, 60, 61]. The reason is that DMs", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "take sampled noises as inputs while image classifiers take natural images as inputs [27]. Elijah [26] is the first defensive algorithm specifically proposed against DM backdoors. It presents a trigger inversion scheme that can approximate the triggers and detect and remove the injected backdoor accordingly. TERD [27] employs a similar trigger estimation strategy and improves the trigger inversion process via a unified loss and trigger refinement. Based on the reconstructed trigger, it subsequently proposes to measure the KL divergence between benign and reversed distribution to detect the backdoor injected in models. Elijah and TERD achieve remarkable detection performance on existing DM backdoors. Thus, we evaluate UIBDiffusion against them and show that UIBDiffusion can bypass both defenses.", + "bbox": [ + 511, + 90, + 906, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Adversarial Perturbation", + "text_level": 1, + "bbox": [ + 511, + 313, + 745, + 328 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The concept of adversarial perturbation is proposed in [47], demonstrating that deep neural networks are vulnerable to small additive noise. These perturbations are typically subtle and imperceptible, hence are widely used to craft adversarial examples [62-64] to fool image classifiers. Later, the work in [48] devised universal adversarial perturbation (UAP), a model-agnostic and image-agnostic variant of adversarial perturbation. We find that UAP is a particularly competitive triggers for backdoor attacks against DMs due to their imperceptibility, generalization property and accessibility.", + "bbox": [ + 511, + 335, + 908, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. UIBDiffusion", + "text_level": 1, + "bbox": [ + 511, + 516, + 651, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Threat Model", + "text_level": 1, + "bbox": [ + 511, + 542, + 656, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We adopt the same threat model and attack scenario presented in the existing backdoor attacks on DMs [23-25], where adversaries publish the backdoored models on websites such as HuggingFace, and users can acquire the pretrained models via third-party sources. Adversaries are expected to be able to inject the backdoor and verify the attack and generation performance before releasing the models. Users are supposed to get access to the backdoor models and the subset of clean training data and test the model utility on their side. Note that in practice, UIBDiffusion applies to a more stringent threat model where users can access the entire training dataset (i.e., contaminated dataset with poisoned samples) and apply any defenses before model deployment as the trigger is imperceptible and undetectable.", + "bbox": [ + 511, + 565, + 908, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. UIBDiffusion Trigger Generation", + "text_level": 1, + "bbox": [ + 511, + 787, + 807, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recall that our goal is to design a general, effective, and stealthy trigger (denoted as $\\tau$ ) for backdoor attacks against DMs. Due to the imperceptibility, generalization, and effectiveness properties of UAPs, they can be adapted as the UIBDiffusion trigger. We first briefly introduce the mechanism and generation process of UAPs. For a given image", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19188", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "classifier $\\mathcal{C}$ and a data distribution $q$ , UAP algorithms aim to find a perturbation vector $v$ such that $\\mathcal{C}(x + v) \\neq \\mathcal{C}(x)$ for most $x \\sim q$ (i.e., image agnostic). The original UAP algorithm [48] achieves this goal by measuring the minimal perturbation that pushes the perturbed input to the decision boundary, which can be mathematically expressed as:", + "bbox": [ + 89, + 90, + 483, + 183 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta v _ {i} = \\underset {d} {\\arg \\min } | | d | | _ {2} \\quad \\mathbf {s}. \\mathbf {t}. \\quad \\mathcal {C} (x _ {i} + v + d) \\neq \\mathcal {C} (x _ {i}), (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 196, + 483, + 223 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $i$ is the image index of a subset of the dataset selected for the computation. The final perturbation is updated as $\\mathcal{P}_{\\xi}(v + \\Delta v_i)$ , where $\\mathcal{P}_{\\xi}(\\cdot)$ is a projection function and $\\xi$ is the perturbation budget. Note that in the context of DMs backdoor, the perturbation $v$ is actually the imperceptible trigger $\\tau$ . We use $\\tau$ for consistent notation from now on to avoid confusion.", + "bbox": [ + 89, + 234, + 483, + 339 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, the original UAP algorithm using DeepFool [65] sometimes yields less effectiveness, and we experimentally observe it renders a relatively lower ASR as the backdoor trigger for DMs. Thus, we propose a novel trigger generation approach inspired by the GUAP [66], an enhanced version of UAP that is more effective and robust. Our trigger generation process adopts the advantage of both additive (i.e., $l_{\\infty}$ -bounded) and non-additive (i.e., spatial transformation) perturbations, where the additive perturbation is adopted as the UIBDiffusion trigger $\\tau$ and coordinately optimized with the non-additive perturbation. We denote the non-additive noise as $f$ . We employ a generator $\\mathcal{G}_{\\gamma}$ (parameterized by $\\gamma$ ), which is akin to the generator in generative adversarial nets (GAN) [67] that take a latent noise $z$ as input, and generate $f$ and trigger $\\tau$ , as shown in lines 3, 4 and 5 in Algorithm 1. In contrast to the GAN that optimizes the generator under the guide of a discriminator, we update our trigger generator under the guide of an image classifier $\\mathcal{C}$ (e.g., VGG and ResNet, etc.), which indicates whether the adversarial perturbation is effective or not, as expressed in Equation 2, where $\\mathcal{H}(\\cdot)$ is the cross entropy loss and $\\otimes$ represents for the spatial transformation operation. Meanwhile, the non-additive perturbation is updated according to the loss function expressed in Equation 3, where $S(\\cdot)$ is summed max pooling values of $x$ using a kernel of size 4 during the spatial transformation and $n$ is the number of images in the selected subset of the dataset. By incorporating such a design, we can progressively optimize the generator and hence improve the effectiveness of the generated UIBDiffusion trigger, as described in lines 5, 7 and 8 in Algorithm 1, respectively. We then optimize the trigger generation process from latent variables, with the objective of maximizing the fooling rate against the image classifier. The completed trigger generation procedures are presented in Algorithm 1 and the concrete architecture of $\\mathcal{G}_{\\gamma}$ and flow illustration are provided in the Appendix as they are similar to GAN.", + "bbox": [ + 89, + 340, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 UIBDiffusion Trigger Generation", + "text_level": 1, + "bbox": [ + 514, + 90, + 826, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Require: Image set $X$ , classifier $\\mathcal{C}$ , latent noise $z$ , generator $\\mathcal{G}_{\\gamma}$ , training epoch $M$ , number of images $n$ , non-additive and additive perturbation budget $\\Gamma$ and $\\xi$", + "bbox": [ + 514, + 109, + 906, + 155 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: for $m = 1 \\dots M$ do", + "2: for $i = 1 \\dots n$ do", + "3: $x_{i}\\sim X,z\\sim \\mathcal{N}(0,1)$", + "4: $\\mathcal{G}_{\\gamma_i}^{m - 1}(z)\\to f_i^{m - 1}$", + "5: $\\mathcal{G}_{\\gamma_i}^{m - 1}(z)\\to \\tau_i^{m - 1}$", + "6: $f = f_{i}^{m}\\gets \\frac{\\Gamma}{\\max_{i}\\sqrt{\\frac{1}{n}\\sum_{i}^{n}(\\mathcal{S}(x_{i}),f_{i}^{m - 1})}}$", + "7: $\\tau = \\tau_{i}^{m}\\gets \\frac{\\xi}{||\\tau_{i}^{m - 1}||_{\\infty}}$", + "8: $\\gamma_i^m\\gets \\gamma_i^{m - 1} - \\nabla_{\\gamma_i^{m - 1}}\\mathcal{L}_{\\mathcal{G}}\\left(\\mathcal{C}(x_i\\otimes f + \\tau),\\mathcal{C}(x_i)\\right)$", + "9: end for", + "10: end for", + "11: return Trigger $\\tau$" + ], + "bbox": [ + 519, + 156, + 903, + 342 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/533c428f255e159c4b25a11fc9a0f5305c85b43c18c4e5b2d7083c7b7d2cce04.jpg", + "image_caption": [ + "Figure 2. Illustration of trigger mechanisms. UIBDiffusion triggers introduce a similar distribution shift as triggers in prior works, which secure a high attack performance. However, since our triggers do not have a specific pattern, they are difficult to remodel by trigger inversion algorithms." + ], + "image_footnote": [], + "bbox": [ + 514, + 359, + 903, + 452 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {G}} = - \\mathcal {H} (\\mathcal {C} (x \\otimes f + \\tau), \\mathcal {C} (x)), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 563, + 906, + 580 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f} = \\max _ {i} \\sqrt {\\frac {1}{n} \\sum_ {i} ^ {n} (\\mathcal {S} (x _ {i} , f))} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 603, + 905, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The effectiveness of UIBDiffusion against backdoor attacks on generative models, such as DMs, stems from key insights into backdoor dynamics. As revealed in [23-25], backdoor for DMs needs to introduce an evident distribution shift and keep such a shift during the reversed Markov chain to secure a high ASR. We empirically find that the UIBDiffusion triggers also lead to a similar distribution shift when imposed on benign sampled noise, as shown in Fig. 2. However, because they are adapted from UAPs, which exhibit minimal, hard-to-detect patterns with subtle perturbations, we hypothesize that these triggers are difficult for trigger inversion algorithms to estimate accurately. Consequently, unlike the triggers with specific patterns, they are managed to circumvent the SOTA defenses and remain highly stealthy. We provide comprehensive and extensive evaluation results in Section 4 to support our conjecture.", + "bbox": [ + 511, + 657, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19189", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Inject Backdoor in Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 90, + 406, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prior attacks leveraged similar backdoor injection mechanisms with slight differences in sampler modifications [23-25]. UIBDiffusion employs the scheme presented in VillanDiffusion [25] since it provides a unified framework to implant backdoors in various types of diffusion models and samplers and achieves outstanding performance. We briefly introduce the injection process in this section.", + "bbox": [ + 89, + 112, + 483, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Clean diffusion process. Clean diffusion consists of forward and backward processes. The forward diffusion process diffuses images $x$ from data distribution $q(x_0)$ to the Gaussian distribution $\\mathcal{N}(0, \\mathbf{I})$ by progressively adding noise in each time step along a Markov chain. The process can be mathematically expressed as: $q(x_t | x_0) = \\mathcal{N}(\\alpha(t)x_0, \\beta^2(t)\\mathbf{I})$ with (1) $q(x_{T_{\\max}}) \\approx \\mathcal{N}(0, \\mathbf{I})$ , (2) $q(x_{T_{\\min}}) \\approx q(x_0)$ and $x_t, t \\in [T_{\\min}, T_{\\max}]$ , where $T_{\\min}, T_{\\max}$ are the first and last timestep, and $\\alpha(t), \\beta(t)$ are parameterized variables introduced in the original diffusion model work [1].", + "bbox": [ + 89, + 218, + 483, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Backward diffusion process reverse the Markov chain and generate images from a sampled Gaussian noise by gradually removing noises in each timestep and can be expressed as: $q(x_{t}|x_{t - 1}) = \\mathcal{N}(k_{t}x_{t - 1},w_{t}^{2}\\mathbf{I})$ , and $x_{t}(x_{0},\\epsilon) = \\hat{\\alpha} (t)x_{0} + \\hat{\\beta} (t)\\epsilon$ , where $\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})$ and $k_{t},w_{t},\\hat{\\alpha} (t),\\hat{\\beta} (t)$ are the same parameterized variables for the reverse process.", + "bbox": [ + 89, + 383, + 483, + 476 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Backdoor and trigger injection. We employ the Villan-Diffusion framework for backdoor injection and modify the trigger injection process in data poisoning. VillanDiffusion proposes a unified backdoor injection framework by introducing the loss function shown in Equation 4, where $\\mathcal{L}_c$ and $\\mathcal{L}_p$ are used to optimize the original generation and backdoor attack tasks, respectively. $r$ is the poisoned data, $y$ is the backdoor target, $\\zeta$ is a parameter to control backdoor against different samplers (e.g., ODE and SDE), and $\\eta_c, \\eta_p$ are weights to balance $\\mathcal{L}_c$ and $\\mathcal{L}_p$ .", + "bbox": [ + 89, + 477, + 483, + 628 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\theta} \\left(\\eta_ {c}, \\eta_ {p}, x, t, \\epsilon , r, y, \\xi\\right) = \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 638, + 482, + 661 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\eta_ {c} \\mathcal {L} _ {c} (x, t, \\epsilon) + \\eta_ {p} \\mathcal {L} _ {p} (x, t, \\epsilon , r, y, \\zeta).\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 659, + 405, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We then explain the difference in the trigger injection process between VillanDiffusion and UIBDiffusion. Villan-Diffusion injects the backdoor trigger following the method in Equation 5, where $\\mathbf{M}$ is a mask and $g$ is the trigger with specific patterns (e.g., gray box and a pair of glasses), and $\\odot$ is the element-wise multiplication.", + "bbox": [ + 89, + 679, + 483, + 770 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr (x, g) = \\mathbf {M} \\odot g + (1 - \\mathbf {M}) \\odot x. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 782, + 482, + 799 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To better align with the adversarial perturbation, we inject our trigger using the approach in Equation 6, where $\\tau$ is the imperceptible trigger synthesized by our proposed trigger generator and $\\varepsilon$ is a variable that controls the strength of the added trigger. In contrast to triggers with recognizable patterns, the UIBDiffusion triggers are generated by", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "a trainable generator with subtle perturbations and hard-to-detect patterns, which makes them effective and stealthy. The backdoor and trigger injection process is summarized in Algorithm 2.", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr (x, \\tau) = x + \\varepsilon \\odot \\tau \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 169, + 906, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 2 UIBDiffusion Backdoor And Trigger Injection", + "bbox": [ + 514, + 210, + 905, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Require: Backdoor Image Trigger $\\tau$ , Benign Dataset $X$ , Backdoor Target $y$ , Training parameters $\\theta$ , Sampler Randomness $\\zeta$ , Strength $\\varepsilon$", + "bbox": [ + 514, + 243, + 905, + 287 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Backdoor DMs with Generated Trigger:", + "2: while not converge do", + "3: $\\{x,\\eta_c,\\eta_p\\} \\sim X$", + "4: $t\\sim \\mathrm{Uniform}(\\{1,\\ldots ,T\\})$", + "5: $\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})$", + "6: Trigger injection: $r(x,\\tau) = x + \\varepsilon \\odot \\tau$", + "7: update $\\theta :\\nabla_{\\theta}\\mathcal{L}_{\\theta}^{I}(\\eta_{c},\\eta_{p},x,t,\\epsilon ,r,y,\\zeta)$", + "8: end while" + ], + "bbox": [ + 524, + 289, + 823, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 443, + 645, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 468, + 720, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and models. We follow the same practice in existing backdoor attacks against DMs [23-25] and use CIFAR-10 $(32 \\times 32)$ [68] and CelebA-HQ $(256 \\times 256)$ [69] datasets for fair comparison. We evaluate UIB-Diffusion on three DMs: DDPM, LDM and NCSN, with eleven different samplers. We adopt the same training recipe from [25] and conduct our attack on pre-trained models (i.e., google/ddpm-cifar10-32 for CIFAR-10 and CompVis/ldm-celebahq-256 for CelebA-HQ) downloaded from Google HuggingFace organization and HuggingFace, respectively. Training hyper-parameters are provided in the Appendix.", + "bbox": [ + 511, + 491, + 906, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Attack configurations and baselines. We conduct our experiments on NVIDIA L40S GPUs and employ Villan-Diffusion [25] as the backdoor injection framework. We provide two generated triggers guided by VGG [70] and ResNet [71] for CIFAR-10 and one proof-of-concept trigger for CelebA-HQ. We adopt \"hat\" and \"cat\" as targets for CIFAR-10 and \"cat\" as the target for CelebA-HQ, respectively. BadDiffusion [23] and VillanDiffusion [25] are considered as the baseline attacks for utility comparison and Elijah [26] and TERD [27] as the defenses for undetectability evaluation.", + "bbox": [ + 511, + 672, + 905, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. We comprehensively evaluate the attack and generation performance of UIBDiffusion using four standard metrics. We exploit FID [72] score to examine the quality of generated clean images. Images with", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19190", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/46501edc54f55a090cc86554d8b2d984367ddef3be71c2002cd226e698a3e6a7.jpg", + "image_caption": [ + "Figure 3. FID and MSE comparison of BadDiffusion, VillanDiffusion, and UIBDiffusion over different poison rates." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 272 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/df0779aecaa78320b9628415772caa2d151413100c8b85d2a94a91f27b8dc861.jpg", + "image_caption": [ + "Table 1. ASR comparison between VillanDiffusion (top row) and UIBDiffusion (bottom row). UIBDiffusion achieves superior performance than the prior work at a low poison rate (e.g., $5\\%$ )." + ], + "image_footnote": [], + "bbox": [ + 94, + 309, + 903, + 448 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "better quality tend to have lower FID scores. As for the attack effect, we employ Attack Success Rate (ASR), Mean Square Error (MSE) and Structural Similarity Index Measure (SSIM) for assessment. ASR is defined as the percentage of the target image generated from sampled noise stamped with backdoor triggers. Higher ASR indicates better utility. MSE is a metric to measure the pixel-level distance between the authentic and synthetic target images. A successful attack is expected to achieve low MSE. SSIM is a similar metric used to measure the similarity between two images (i.e., the real and generated target images). Superior attack performance will yield SSIM that approaches the value of 1.", + "bbox": [ + 88, + 500, + 485, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results on Different Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 710, + 416, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first evaluate UIBDiffusion on different DMs. We present the results of DDPM in the main paper and organize the results of LDM and NCSN in the Appendix. As shown in Fig. 3, Table 1 and Table 2, UIBDiffusion achieves comparable FID as the clean diffusion model (i.e., poison rate $= 0\\%$ ). It can be seen that with the poison rates increased, all backdoor attacks yield a lower MSE, demonstrating a positive correlation between poison rates and attack effect. However, the generation ability of DMs will be accordingly affected when the poison rate is high. Notably, UIBDiffusion can reach a $100\\%$ ASR with only $5\\%$ poison rate", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "while VillanDiffusion gets $56\\%$ ASR at this rate. In practice, attackers can achieve the expected attack effect without undermining the original generation ability of DMs, which further improves the stealthiness of the backdoor attack.", + "bbox": [ + 511, + 500, + 906, + 559 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7d15ea2b50eab05a764f32458652ec8f69427a85991b4669b90e10bcb1c235e1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Poison Rate0%1%2%3%4%5%10%≥ 20%
BadDiffusion [23]3.80E-46.22E-31.48E-21.10E-20.370.990.990.99
VillanDiffusion [25]3.80E-43.92E-41.25E-21.69E-20.600.990.990.99
UIBDiffusion (Ours)3.80E-40.470.960.990.990.990.990.99
", + "bbox": [ + 516, + 563, + 906, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. SSIM comparison of BadDiffusion, VillanDiffusion and UIBDiffusion. UIBDiffusion achieves higher SSIM at lower poison rates of $1\\% \\sim 4\\%$ .", + "bbox": [ + 511, + 623, + 906, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Results on Different Samplers", + "text_level": 1, + "bbox": [ + 511, + 667, + 782, + 685 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We then show the effectiveness of UIBDiffusion on different samplers. We conduct the experiments using DDPM model with 11 types of samplers. As shown in Fig. 4, UIB-Diffusion achieves a consistent performance of $100\\%$ ASR across all samplers and yield a robust target image generation effect (i.e., MSE and SSIM) at various poison rates. We also observe that the ODE solvers outperform their SDE solver counterparts.", + "bbox": [ + 511, + 690, + 906, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Resilience against SOTA Defenses", + "text_level": 1, + "bbox": [ + 511, + 818, + 808, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Elijah [26] and TERD [27] are two SOTA defenses against backdoor attacks on DMs. They employ a similar idea of trigger inversion to estimate and reverse engineer the triggers. They then apply backdoor detection and removal", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19191", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/df442f1a67b91a3928b46354a75ad9a1c06e4ea1d12159dc574bb2464df0da50.jpg", + "image_caption": [ + "Figure 4. UIBDiffusion performance against eleven different samplers across various poison rates. UIBDiffusion yields consistently high ASR, SSIM, and MSE and even better FID on ODE-based samplers compared to SDE-based samplers." + ], + "image_footnote": [], + "bbox": [ + 98, + 90, + 287, + 218 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/281a17324825adcdc6951bb4c4cdcbca827950368da8c059cddc41de22291528.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 303, + 90, + 491, + 217 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5df83fc19ec4015c3af41dc08e630cfa33169215d59626b0f29e151746456e4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 90, + 696, + 218 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d00e50ed3acef7d9fdadfdf20b13d25e74f38f7617c5db20d66e8787b35df162.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 90, + 901, + 217 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "based on the reconstructed triggers. We show that UIBDiffusion is highly resilient to both attacks since our triggers are hard to remodel. Fig. 5 demonstrates the performance of UIBDiffusion against Elijah compared to BadDifuFSion and VillanDiffusion. It can be seen that UIBDiffusion achieves $100\\%$ ASR and keeps the same MSE and SSIM before and after the backdoor detection and removal while ASR and SSIM of BadDiffusion and VillanDiffusion drop to $0\\%$ after detection, indicating superior resilience against the SOTA defense. We then present the robustness of UIBDiffusion across various poison rates. Higher poison rates usually yield higher ASR but are easier to detect. However, as shown in Table 3, UIBDiffusion managed to escape detection at all the evaluated poison rates while maintaining $100\\%$ ASR and high performance of SSIM and FID. It is worth noting that a $\\Delta SSIM = 0$ indicates that UIBDiffusion successfully bypasses the defense, achieving $100\\%$ targeted image generation. We attribute the fluctuations in FID to Elijah's use of an ineffective inverted trigger during backdoor removal.", + "bbox": [ + 88, + 275, + 485, + 575 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/096c157e2328ae1ee9f58e07d6a34f640f18e0323911c5a21a38fd53b51a623f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BadDiffusion * [23]UIBDiffusion (Ours)
Poison RateDetectedASRΔSSIMΔFIDDetectedASRΔSSIMΔFID
5%100%0%-1.0-0.070%100%0-0.98
10%100%0%-1.0+0.180%100%0+0.15
20%100%0%-1.0-0.030%100%0+5.72
30%100%0%-1.0+0.150%100%0+12.59
50%100%0%-1.0+0.200%100%0+11.56
70%100%0%-1.0-0.110%100%0+14.85
90%100%0%-1.0-0.070%100%0+5.09
", + "bbox": [ + 94, + 579, + 478, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the quality of the inverted trigger by visualizing the inverted triggers and testing if they could activate the backdoor as the actual triggers when imposed on benign sampled noise. Results are shown in Fig. 6. Sampled noise with inverted triggers should generate the target image if the defense succeeds and generate normal images otherwise. It can be seen that for triggers that have specific patterns (e.g., a stop sign and a gray box), Elijah can successfully estimate the authentic triggers and the inverted triggers are able to activate backdoors. On the other hand, the inverted trigger of UIBDiffusion fails to activate the backdoor injected in the", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DM, showing that the defense is ineffective against UIBDiffusion. We argue this is because the UIBDiffusion trigger is adapted from UAP, which is small and unstructured noise that is difficult to approximate based on the current trigger inversion loss function in Elijah and TERD.", + "bbox": [ + 511, + 275, + 906, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "UIBDiffusion performs a similar undetectability against TERD as shown in Table 4. It achieves a $100\\%$ True Positive Rate (TPR) and $0\\%$ False Positive Rate (FPR), where TPR and FPR represent the percentage of the benign or backdoor sampled noise that is correctly detected. On the other hand, triggers with obvious patterns fail to bypass the trigger inversion algorithm and are all detected by TERD. Although TRED offers a more precise trigger inversion approach using a unified loss and trigger refinement, it is still ineffective against our proposed UIBDiffusion trigger.", + "bbox": [ + 511, + 351, + 908, + 503 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4dff114813bd2cb0f6e14502d540766713e47215fe4c33bf7eaf2d85697f44f8.jpg", + "table_caption": [ + "Table 3. Elijah against BadDiffusion and UIBDiffusion with different poison rates. “*” denotes results replicated from [26]." + ], + "table_footnote": [], + "table_body": "
Trigger-TargetTPR (↑)FPR (↓)
Ours-HAT100%0%
BOX-HAT*100%100%
BOX-SHOE*100%100%
BOX-CORNER*100%100%
STOP-HAT*100%100%
STOP-SHOE*100%100%
STOP-CORNER*100%100%
", + "bbox": [ + 517, + 508, + 901, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Resilience of UIBDiffusion against TERD. “*” denotes results replicated from [27].", + "bbox": [ + 511, + 623, + 905, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Comparison of Trigger Generation Approach", + "text_level": 1, + "bbox": [ + 511, + 666, + 897, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the necessity and effectiveness of our proposed trigger generation approach by comparing it to the original UAPs and GUAPs. We employ the generated UIBDiffusion trigger, the original UAP from [48] and the GUAP from [66] to poison the benign dataset and backdoor the DMs using these three triggers, respectively. As can be seen from Table 5, the UIBDiffusion trigger outperforms the original UAP and GUAP triggers at all four metrics by large margins, achieving much better attack and generation performance. The rationale behind this is that our trigger introduces an evident distribution shift. In contrast, the original UAP trigger only slightly affects the sampled noise distribution, and the spatial transformation of GUAP is image-specific and imposes a negative impact on the backdoor im", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19192", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ec68786bb9c549cfe46620a356bf3bb5c13a8a48ed2ca3ae36ba1c131a24cddf.jpg", + "image_caption": [ + "Figure 5. Performance comparison of BadDiffusion, VillanDiffusion and our UIBDiffusion before and after the Elijah defense. ASR: the higher the better; FID: the lower the better; MSE: the lower the better; SSIM: the higher the better." + ], + "image_footnote": [], + "bbox": [ + 96, + 98, + 294, + 236 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c5dd4453540f91a2e28b13f96e930829895ef883715b870c9dd5847c6c6b972b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 300, + 98, + 493, + 236 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b593e16ad5586ee094b3c9db4eb849a18edf39bb0d92eca2cab7efffb02fbb65.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 98, + 697, + 236 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/50fb0826eea200ac7edd7ff0ff4b5353e69a0a1416dae3ab5fcd90787e2bdaee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 98, + 898, + 236 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1e03395fafd941eeb48ee2340d21b80e3352e419d0f23b30ef9a84eb96827cca.jpg", + "image_caption": [ + "Figure 6. Visualization of trigger inversion. Inverted triggers of gray box and stop sign can activate backdoor, indicating a successful estimation of the actual triggers. On the other hand, inverted triggers of UIBDiffusion fail to activate the injected backdoor." + ], + "image_footnote": [], + "bbox": [ + 96, + 300, + 480, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "plant. This shows the superiority and necessity of our trigger design choice.", + "bbox": [ + 89, + 588, + 482, + 619 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/655fb51a76a2ac3fba1cda7cf756285686cd0289cb142fd82b756ed44921c047.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FID (↓)MSE (↓)SSIM (↑)ASR (↑)
Original UAP20.889.22E-20.619043.75%
Original GUAP41.877.86E-20.19320.00%
UIBDiffusion19.775.30E-40.9967100%
", + "bbox": [ + 94, + 631, + 478, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Trigger Robustness", + "text_level": 1, + "bbox": [ + 89, + 753, + 276, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Subtle adversarial perturbations are susceptible to smoothing defenses [73] and could be mitigated by random noise. Thus, we perform a robust evaluation of the UIBDiffusion triggers and present the results in Table 6. UIBDiffusion shows strong robustness against random noise. We argue that the resilience is due to our trigger generation algorithm enhancing the adversarial perturbations via co-optimizing the additive and non-additive noise.", + "bbox": [ + 89, + 780, + 483, + 898 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c5253a3abe523eaa439c1677661472aaccbe2dd490042147d26e4466700703e8.jpg", + "table_caption": [ + "Table 5. Trigger performance comparison across the original UAP, GUAP and UIBDiffusion." + ], + "table_footnote": [], + "table_body": "
Random Noise1%2%5%10%20%
ASR99%99%100%100%100%
FID19.4819.5219.8019.3719.49
MSE2.36E-31.13E-38.03E-43.02E-45.30E-4
SSIM0.98650.99290.99440.99690.9957
", + "bbox": [ + 517, + 297, + 905, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Quantitative Evaluation of Imperceptibility", + "text_level": 1, + "bbox": [ + 511, + 397, + 880, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also quantitatively evaluate the imperceptibility of our trigger design using distance metrics of $l_{\\infty}$ and $l_{2}$ norms. We compute the distance between a poisoned image (stamping with our trigger and triggers used in prior works) and a benign image and summarize the results in Table 7. Our imperceptible triggers have much lower $l_{\\infty}$ and $l_{2}$ distances compared to triggers with specific patterns, which validates the stealthiness at the image level (i.e., the pixel space) of the UIBDiffusion trigger design.", + "bbox": [ + 511, + 419, + 906, + 556 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c61efb6226093eb7d41314aa14fc0fb82ec77cf3206c8c1cfe6ca7c87ea82f97.jpg", + "table_caption": [ + "Table 6. Robustness of UIBDiffusion against random noise." + ], + "table_footnote": [], + "table_body": "
Cifar 10CelebA-HQ-256
l∞l2l∞l2
UIBDiffusion Trigger0.041.800.0512.74
Prior Work0.476.210.8019.84
", + "bbox": [ + 517, + 561, + 905, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Quantitative comparison of triggers imperceptibility using distance metrics of $l_{\\infty}$ and $l_{2}$ norms.", + "bbox": [ + 511, + 633, + 905, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 680, + 633, + 695 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we present UIBDiffusion, the first imperceptible backdoor attack against diffusion models. We propose a novel trigger generation method based on universal adversarial perturbations and demonstrate that such perturbations for attacking discriminative models can be adapted as general, effective, and stealthy triggers for backdoor attacks against diffusion models. We empirically reveal the underlying rationale why UIBDiffusion can achieve a high attack success rate while evading the trigger inversion defensive algorithms. We hope our work sheds light on the potential risks that diffusion models face and motivates future research to develop more comprehensive defenses.", + "bbox": [ + 511, + 705, + 906, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19193", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported in part by the National Science Foundation (NSF) SaTC-2426299 and SaTC-2413046.", + "bbox": [ + 91, + 113, + 480, + 142 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 170, + 187, + 186 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3, 5", + "[2] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In 9th International Conference on Learning Representations, ICLR 2021, 2021. 3", + "[3] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, pages 10674-10685. IEEE, 2022. 1, 3", + "[4] William Peebles and Saining Xie. Scalable diffusion models with transformers. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, pages 4172-4182, 2023. 1", + "[5] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, ICML 2024. OpenReview.net, 2024. 1", + "[6] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 1", + "[7] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 3", + "[8] Yuming Jiang, Tianxing Wu, Shuai Yang, Chenyang Si, Dahua Lin, Yu Qiao, Chen Change Loy, and Ziwei Liu. Videobooth: Diffusion-based video generation with image prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6689-6700, 2024. 1", + "[9] Xiang Li, John Thickstun, Ishaan Gulrajani, Percy S Liang, and Tatsunori B Hashimoto. Diffusion-lm improves controllable text generation. Advances in Neural Information Processing Systems, 35:4328-4343, 2022. 1", + "[10] Xingqian Xu, Zhangyang Wang, Gong Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text, images and variations all in one diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7754-7765, 2023.", + "[11] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1", + "[12] Myeonghun Jeong, Hyeongju Kim, Sung Jun Cheon, Byoung Jin Choi, and Nam Soo Kim. Diff-nts: A denois" + ], + "bbox": [ + 93, + 195, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing diffusion model for text-to-speech. arXiv preprint arXiv:2104.01409, 2021. 1", + "[13] Rongjie Huang, Max WY Lam, Jun Wang, Dan Su, Dong Yu, Yi Ren, and Zhou Zhao. Fastdiff: A fast conditional diffusion model for high-quality speech synthesis. arXiv preprint arXiv:2204.09934, 2022.", + "[14] Heeseung Kim, Sungwon Kim, and Sungroh Yoon. Guidedts: A diffusion model for text-to-speech via classifier guidance. In International Conference on Machine Learning, pages 11119-11133. PMLR, 2022.", + "[15] Vadim Popov, Ivan Vovk, Vladimir Gogoryan, Tasnama Sadekova, and Mikhail Kudinov. Grad-tts: A diffusion probabilistic model for text-to-speech. In International Conference on Machine Learning, pages 8599-8608. PMLR, 2021.", + "[16] Yuancheng Wang, Zeqian Ju, Xu Tan, Lei He, Zhizheng Wu, Jiang Bian, et al. Audit: Audio editing by following instructions with latent diffusion models. Advances in Neural Information Processing Systems, 36:71340-71357, 2023.", + "[17] Shuai Shen, Wenliang Zhao, Zibin Meng, Wanhua Li, Zheng Zhu, Jie Zhou, and Jiwen Lu. Difftalk: Crafting diffusion models for generalized audio-driven portraits animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1982-1991, 2023. 1", + "[18] Bo Li, Yining Wang, Aarti Singh, and Yevgeniy Vorobeychik. Data poisoning attacks on factorization-based collaborative filtering. Advances in neural information processing systems, 29, 2016. 1", + "[19] Mengchen Zhao, Bo An, Yaodong Yu, Sulin Liu, and Sinno Pan. Data poisoning attacks on multi-task relationship learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018.", + "[20] Scott Alfeld, Xiaojin Zhu, and Paul Barford. Data poisoning attacks against autoregressive models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 30, 2016.", + "[21] Bingyin Zhao and Yingjie Lao. Towards class-oriented poisoning attacks against neural networks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3741-3750, 2022.", + "[22] Bingyin Zhao and Yingjie Lao. Clpa: Clean-label poisoning availability attacks using generative adversarial nets. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 9162-9170, 2022. 1", + "[23] Sheng-Yen Chou, Pin-Yu Chen, and Tsung-Yi Ho. How to backdoor diffusion models? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4015-4024, 2023. 1, 3, 4, 5, 6, 7", + "[24] Weixin Chen, Dawn Song, and Bo Li. Trojdiff: Trojan attacks on diffusion models with diverse targets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4035-4044, 2023. 3", + "[25] Sheng-Yen Chou, Pin-Yu Chen, and Tsung-Yi Ho. Villan-diffusion: A unified backdoor attack framework for diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 4, 5, 6", + "[26] Shengwei An, Sheng-Yen Chou, Kaiyuan Zhang, Qiuling Xu, Guanhong Tao, Guangyu Shen, Siyuan Cheng, Shiqing" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19194", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ma, Pin-Yu Chen, Tsung-Yi Ho, et al. Elijah: Eliminating backdoors injected in diffusion models via distribution shift. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 10847-10855, 2024. 1, 2, 3, 5, 6, 7", + "[27] Yichuan Mo, Hui Huang, Mingjie Li, Ang Li, and Yisen Wang. TERD: A unified framework for safeguarding diffusion models against backdoors. In *Forty-first International Conference on Machine Learning*, ICML 2024, 2024. 1, 2, 3, 5, 6, 7", + "[28] Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying vulnerabilities in the machine learning model supply chain. arXiv preprint arXiv:1708.06733, 2017. 1, 3", + "[29] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. CoRR, abs/1712.05526, 2017.", + "[30] Yingqi Liu, Shiqing Ma, Yousra Aafer, Wen-Chuan Lee, Juan Zhai, Weihang Wang, and Xiangyu Zhang. Trojaning attack on neural networks. In 25th Annual Network and Distributed System Security Symposium, NDSS. The Internet Society, 2018. 1, 3", + "[31] Yuanshun Yao, Huiying Li, Haitao Zheng, and Ben Y. Zhao. Latent backdoor attacks on deep neural networks. In Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security, CCS 2019, pages 2041-2055. ACM, 2019. 1", + "[32] Yunfei Liu, Xingjun Ma, James Bailey, and Feng Lu. Reflection backdoor: A natural backdoor attack on deep neural networks. In Computer Vision - ECCV 2020, Proceedings, Part X, volume 12355 of Lecture Notes in Computer Science, pages 182-199. Springer, 2020.", + "[33] Yuezun Li, Yiming Li, Baoyuan Wu, Longkang Li, Ran He, and Siwei Lyu. Invisible backdoor attack with sample-specific triggers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16463-16472, 2021.", + "[34] Tuan Anh Nguyen and Anh Tuan Tran. Wanet - imperceptible warping-based backdoor attack. In 9th International Conference on Learning Representations, ICLR, 2021.", + "[35] Siyuan Cheng, Yingqi Liu, Shiqing Ma, and Xiangyu Zhang. Deep feature space trojan attack of neural networks by controlled detoxification. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, pages 1148–1156. AAAI Press, 2021.", + "[36] Khoa Doan, Yingjie Lao, Weijie Zhao, and Ping Li. Lira: Learnable, imperceptible and robust backdoor attacks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11966-11976, 2021.", + "[37] Khoa Doan, Yingjie Lao, and Ping Li. Backdoor attack with imperceptible input and latent modification. Advances in Neural Information Processing Systems, 34:18944-18957, 2021.", + "[38] Khoa D Doan, Yingjie Lao, and Ping Li. Marksman backdoor: Backdoor attacks with arbitrary target class. Advances in Neural Information Processing Systems, 35:38260-38273, 2022. 1" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Aniruddha Saha, Akshayvarun Subramanya, and Hamed Pirsiavash. Hidden trigger backdoor attacks. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, pages 11957-11965. AAAI Press, 2020. 1", + "[40] Hossein Souri, Liam Fowl, Rama Chellappa, Micah Goldblum, and Tom Goldstein. Sleeper agent: Scalable hidden trigger backdoors for neural networks trained from scratch. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, 2022. 1", + "[41] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Anti-backdoor learning: Training clean models on poisoned data. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 14900-14912, 2021. 1", + "[42] Qi Zhao and Christian Wressnegger. Adversarily robust anti-backdoor learning. In Proceedings of the 2024 Workshop on Artificial Intelligence and Security, AISec 2024, pages 77-88. ACM, 2024.", + "[43] Bolun Wang, Yuanshun Yao, Shawn Shan, Huiying Li, Bimal Viswanath, Haitao Zheng, and Ben Y. Zhao. Neural cleanse: Identifying and mitigating backdoor attacks in neural networks. In 2019 IEEE Symposium on Security and Privacy, SP 2019, pages 707-723. IEEE, 2019.", + "[44] Dongxian Wu and Yisen Wang. Adversarial neuron pruning purifies backdoored deep models. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, pages 16913-16925, 2021. 3", + "[45] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Neural attention distillation: Erasing backdoor triggers from deep neural networks. In 9th International Conference on Learning Representations, ICLR 2021, 2021.", + "[46] Bingyin Zhao and Yingjie Lao. Ultraclean: A simple framework to train robust neural networks against backdoor attacks. arXiv preprint arXiv:2312.10657, 2023. 1", + "[47] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In 2nd International Conference on Learning Representations, ICLR 2014, 2014. 2, 3", + "[48] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Universal adversarial perturbations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1765-1773, 2017. 2, 3, 4, 7", + "[49] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, pages 11895-11907, 2019. 3", + "[50] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In 9th International Conference on Learning Representations, ICLR 2021. OpenReview.net, 2021." + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19195", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, 2020. 3", + "[52] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. In The Eleventh International Conference on Learning Representations, ICLR 2023, 2023. 3", + "[53] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, 2023. 3", + "[54] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, 2022. 3", + "[55] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, 2022. 3", + "[56] Joseph Clements and Yingjie Lao. Hardware trojan attacks on neural networks. arXiv preprint arXiv:1806.05768, 2018. 3", + "[57] Shengfang Zhai, Yinpeng Dong, Qingni Shen, Shi Pu, Yuejian Fang, and Hang Su. Text-to-image diffusion models can be easily backdoored through multimodal data poisoning. In Proceedings of the 31st ACM International Conference on Multimedia, MM 2023, pages 1577-1587. ACM, 2023. 3", + "[58] Yihao Huang, Qing Guo, and Felix Juefei-Xu. Zero-day backdoor attack against text-to-image diffusion models via personalization. CoRR, abs/2305.10701, 2023.", + "[59] Lukas Struppek, Dominik Hintersdorf, and Kristian Kersting. Rickrolling the artist: Injecting backdoors into text encoders for text-to-image synthesis. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, pages 4561-4573. IEEE, 2023. 3", + "[60] Bryant Chen, Wilka Carvalho, Nathalie Baracaldo, Heiko Ludwig, Benjamin Edwards, Taesung Lee, Ian M. Moly, and Biplav Srivastava. Detecting backdoor attacks on deep neural networks by activation clustering. In Workshop on Artificial Intelligence Safety 2019 co-located with the Thirty-Third AAAI Conference on Artificial Intelligence 2019 (AAAI-19), volume 2301 of CEUR Workshop Proceedings. CEUR-WS.org, 2019. 3", + "[61] Yi Zeng, Won Park, Z. Morley Mao, and Ruoxi Jia. Rethinking the backdoor attacks' triggers: A frequency perspective. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, pages 16453-16461. IEEE, 2021. 3", + "[62] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In 3rd International Conference on Learning Representations, ICLR 2015, 2015. 3", + "[63] Alexey Kurakin, Ian J. Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In 5th International" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Learning Representations, ICLR 2017, Workshop Track Proceedings, 2017.", + "[64] Chaowei Xiao, Bo Li, Jun-Yan Zhu, Warren He, Mingyan Liu, and Dawn Song. Generating adversarial examples with adversarial networks. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJ-CAI 2018, pages 3905–3911. ijcai.org, 2018. 3", + "[65] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2574–2582, 2016. 4", + "[66] Yanghao Zhang, Wenjie Ruan, Fu Wang, and Xiaowei Huang. Generalizing universal adversarial attacks beyond additive perturbations. In 2020 IEEE International Conference on Data Mining (ICDM), pages 1412-1417. IEEE, 2020. 4, 7", + "[67] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 4", + "[68] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical Report, 2009. 5", + "[69] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In 2015 IEEE International Conference on Computer Vision, ICCV 2015, pages 3730-3738. IEEE Computer Society, 2015. 5", + "[70] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In Yoshua Bengio and Yann LeCun, editors, 3rd International Conference on Learning Representations, ICLR 2015, 2015. 5", + "[71] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5", + "[72] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, pages 6626-6637, 2017. 5", + "[73] Hadi Salman, Mingjie Sun, Greg Yang, Ashish Kapoor, and J Zico Kolter. Denoised smoothing: A provable defense for pretrained classifiers. Advances in Neural Information Processing Systems, 33:21945-21957, 2020. 8" + ], + "bbox": [ + 516, + 92, + 903, + 753 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19196", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/UIBDiffusion_ Universal Imperceptible Backdoor Attack for Diffusion Models/7f0574b2-e282-4c3e-a072-3d64d026983a_model.json b/2025/UIBDiffusion_ Universal Imperceptible Backdoor Attack for Diffusion Models/7f0574b2-e282-4c3e-a072-3d64d026983a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0e661b0723d864a6f45bcb1e8ced964177ebdca7 --- /dev/null +++ b/2025/UIBDiffusion_ Universal Imperceptible Backdoor Attack for Diffusion Models/7f0574b2-e282-4c3e-a072-3d64d026983a_model.json @@ -0,0 +1,2609 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.13, + 0.892, + 0.152 + ], + "angle": 0, + "content": "UIBDiffusion: Universal Imperceptible Backdoor Attack for Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.162, + 0.18, + 0.864, + 0.2 + ], + "angle": 0, + "content": "Yuning Han\\(^{1,*}\\) Bingyin Zhao\\(^{2,*}\\) Rui Chu\\(^{3}\\) Feng Luo\\(^{4}\\) Biplab Sikdar\\(^{2}\\) Yingjie Lao\\(^{3,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.215, + 0.914, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)Columbia University \\(^{2}\\)National University of Singapore \\(^{3}\\)Tufts University \\(^{4}\\)Clemson University" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.301, + 0.485, + 0.753 + ], + "angle": 0, + "content": "Recent studies show that diffusion models (DMs) are vulnerable to backdoor attacks. Existing backdoor attacks impose unconcealed triggers (e.g., a gray box and eyeglasses) that contain evident patterns, rendering remarkable attack effects yet easy detection upon human inspection and defensive algorithms. While it is possible to improve stealthiness by reducing the strength of the backdoor, doing so can significantly compromise its generality and effectiveness. In this paper, we propose UIBDiffusion, the universal imperceptible backdoor attack for diffusion models, which allows us to achieve superior attack and generation performance while evading state-of-the-art defenses. We propose a novel trigger generation approach based on universal adversarial perturbations (UAPs) and reveal that such perturbations, which are initially devised for fooling pre-trained discriminative models, can be adapted as potent imperceptible backdoor triggers for DMs. We evaluate UIBDiffusion on multiple types of DMs with different kinds of samplers across various datasets and targets. Experimental results demonstrate that UIBDiffusion brings three advantages: 1) Universality, the imperceptible trigger is universal (i.e., image and model agnostic) where a single trigger is effective to any images and all diffusion models with different samplers; 2) Utility, it achieves comparable generation quality (e.g., FID) and even better attack success rate (i.e., ASR) at low poison rates compared to the prior works; and 3) Undetectability, UIBDiffusion is plausible to human perception and can bypass Elijah and TERD, the SOTA defenses against backdoors for DMs. Code is available at https://github.com/TheLaoLab/UIBDiffusion." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.782, + 0.222, + 0.797 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.806, + 0.483, + 0.868 + ], + "angle": 0, + "content": "Diffusion models (DMs) [1-3] have emerged as the state-of-the-art generative paradigm in recent years and achieved unprecedented success in image [4, 5], video [6-8] text [9-11] and audio [12-17] synthesis. The success is highly" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.271, + 0.905, + 0.378 + ], + "angle": 0, + "content": "attributed to training the foundation model on large-scale datasets [5]. Such a practice places DMs at risk of being compromised by data poisoning attacks [18-22], where adversaries contaminate the clean training dataset with a small fraction of malicious samples and undermine the model performance. This imposes a critical security concern on the practical deployment of DMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.382, + 0.907, + 0.775 + ], + "angle": 0, + "content": "Backdoor attacks are emerging threats to DMs. Recent studies [23-25] show that DMs are vulnerable to such attacks, which implant backdoor into victim models by training them on poisoned datasets and manipulate their behavior through inputs stamped with predefined triggers. Existing backdoor attacks for DMs usually employ unconcealed triggers such as a gray box, an image of Hello Kitty, and eyeglasses in the forward and reversed (i.e., sampling) diffusion process, as shown in Fig. 1. While such trigger designs achieve superior attack performance and preserve the decent generation ability of victim models, they are easy to detect via human inspection and advanced defensive algorithms as they have obvious patterns that can be precisely remodeled via reverse engineering [26, 27]. Motivated by this observation, we aim to design a general, powerful yet stealthy trigger. However, even for the traditional backdoor attacks against discriminative models (e.g., image classifiers) [28-30], it is challenging to acquire such a trigger since it either requires complex and carefully designed generators to generate image-specific or model-specific triggers [31-38] or suffers from a relatively low attack success rate (ASR) [39, 40], or fails to evade modern defenses [41-46]. This raises an interesting question in backdoor attacks against DMs: Is there a trigger simultaneously possessing generality, effectiveness, and stealthiness for backdoor attacks against DMs?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Driving by the question, in this work, we propose UIB-Diffusion, the universal imperceptible backdoor triggers for DMs that bring the following advantages: 1) Universality: the trigger is image-agnostic and model-agnostic so that it can be applied to arbitrary images and DMs; 2) Utility: the trigger can achieve high ASR while maintaining the generation capability of DMs; 3) Undetectability: the trigger is plausible to humans and capable of circumventing" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.221, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19186" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.865, + 0.498 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.508, + 0.908, + 0.604 + ], + "angle": 0, + "content": "Figure 1. Illustrations of the forward diffusion process (top block with the yellow background) and backward diffusion process (bottom block with the blue background) of a clean diffusion model (blue dash line), VillanDiffusion [25] (red dash line) and UIBDiffusion (Ours, green dash line). The UIBDiffusion trigger \\((\\tau)\\) is plausible to humans in all phases from data poisoning to forward diffusion and backward diffusion while the glasses trigger \\((\\mathbf{g})\\) in prior works is perceptible. UIBDiffusion trigger is highly effective since it introduces a similar distribution shift as the glasses trigger, which secures the attack performance during the backward diffusion process. However, it is hard to detect as the trigger does not possess specific patterns and cannot be inverted by existing defensive algorithms. We empirically verify the effectiveness in Section 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.622, + 0.486, + 0.833 + ], + "angle": 0, + "content": "state-of-the-art defenses. Inspired by the adversarial perturbation [47] originally designed for attacking pre-trained image classifiers, we reveal that universal adversarial perturbations [48] (UAPs), a variant of adversarial perturbations, naturally hold the aforementioned properties and can be adapted as the UIBDiffusion triggers. First, UAPs are image-agnostic and model-agnostic. Second, we demonstrate that they are effective triggers because they introduce a similar distribution shift as prior works, yet they are stealthy because they do not possess clear patterns and are hard to estimate by existing trigger inversion algorithms. Moreover, they are trivial and imperceptible noises that humans can barely identify, as shown in Fig. 1. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.84, + 0.483, + 0.871 + ], + "angle": 0, + "content": "- To the best of our knowledge, UIBDiffusion is the first imperceptible backdoor attack against diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "- We discover that adversarial perturbations designed for discriminative models can be adapted as backdoor trig-" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.622, + 0.905, + 0.651 + ], + "angle": 0, + "content": "gers for generative models. We reveal why such triggers can achieve high ASR and evade SOTA defenses." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.652, + 0.907, + 0.711 + ], + "angle": 0, + "content": "- We propose a practical trigger generation approach adapted from UAP, which is specifically designed to enhance universality, attack effectiveness and accessibility of triggers." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.713, + 0.907, + 0.802 + ], + "angle": 0, + "content": "- We experimentally demonstrate that UIBDiffusion is highly effective towards multiple diffusion models with different samplers and achieves comparable generation quality on benign input compared to clean DMs and superior attack success rate and input stealthiness compared to prior works." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.803, + 0.909, + 0.892 + ], + "angle": 0, + "content": "- While existing works face a trade-off between utility and undetectability, we empirically show that UIBDiffusion triggers are highly resistant to the SOTA trigger inversion-based defenses, effectively bypassing Elijah [26] and TERD [27], the state-of-the-art defenses against backdoors on DMs." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.652, + 0.909, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19187" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.241, + 0.106 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.116, + 0.26, + 0.131 + ], + "angle": 0, + "content": "2.1. Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.139, + 0.486, + 0.413 + ], + "angle": 0, + "content": "Diffusion models are latent variable models composed of two processes. The forward diffusion process imposes Gaussian noise on input data in each time step along the Markov chain and progressively diffuses the data distribution to an isotropic Gaussian distribution. The backward diffusion process is a reversed Markov chain that gradually denoises the sampled noise from the Gaussian distribution to the input data distribution. Mainstream diffusion models include DDPM [1], LDM [3] and NCSN [49-51], etc. Despite the excellent performance, DMs suffer from slow sampling during the backward process. There exists a line of works that leverage different techniques such as generalized non-Markovian chain and Ordinary Differential Equation (ODE) to improve the samplers and accelerate the generation process: DDIM [2], DEIS [52], UniPC [53], Heun [54], DPM-Solver [55] and DPM-Solver++ [7]. We comprehensively evaluate UIBDiffusion on these models and samplers to showcase the utility of our trigger." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.424, + 0.484, + 0.455 + ], + "angle": 0, + "content": "2.2. Backdoor Attacks and Defenses on Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.485, + 0.855 + ], + "angle": 0, + "content": "Attacks. Backdoor attacks [28-30, 56], originally explored in the context of discriminative models such as image classifiers, induce models to make wrong predictions on triggered images. Although having been extensively studied for discriminative models, backdoor attacks are under-explored in diffusion models. BadDiffusion [23], TrojDiff [24] and VillanDiffusion [25] are pioneering works in this field. Unlike traditional backdoor attacks, those targeting DMs aim to force models to synthesize a targeted image upon malicious inputs (i.e., noises with triggers) while generating normal images upon benign inputs (i.e., isotropic Gaussian noises). BadDiffusion shows how to backdoor DMs with unconcealed triggers such as a gray box and a stop sign. TrojDiff proposes to use a whole image (e.g., Hello Kitty) as the trigger and demonstrates that backdoor can be generalized to DDIM and different adversarial goals. However, both attacks are limited to certain DMs (e.g., DDPM) and samplers (e.g., ancestral and DDIM). VillanDiffusion resolves the limitations and offers a unified framework for all kinds of DMs and samplers. Other works show backdoor can be injected to text-to-image DMs and activated by text prompts [57-59]. However, these attacks draw less attention from defenders as they can eliminate the backdoor by cleansing the text encoder. Thus, our attack focuses on the primary defensive objective of the defender: pixel level (i.e., image) trigger generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Defenses. Backdoor attacks against DMs are shown to be immune to defenses designed for traditional image classifier backdoors [23, 25, 44, 60, 61]. The reason is that DMs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.304 + ], + "angle": 0, + "content": "take sampled noises as inputs while image classifiers take natural images as inputs [27]. Elijah [26] is the first defensive algorithm specifically proposed against DM backdoors. It presents a trigger inversion scheme that can approximate the triggers and detect and remove the injected backdoor accordingly. TERD [27] employs a similar trigger estimation strategy and improves the trigger inversion process via a unified loss and trigger refinement. Based on the reconstructed trigger, it subsequently proposes to measure the KL divergence between benign and reversed distribution to detect the backdoor injected in models. Elijah and TERD achieve remarkable detection performance on existing DM backdoors. Thus, we evaluate UIBDiffusion against them and show that UIBDiffusion can bypass both defenses." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.314, + 0.746, + 0.329 + ], + "angle": 0, + "content": "2.3. Adversarial Perturbation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.336, + 0.909, + 0.503 + ], + "angle": 0, + "content": "The concept of adversarial perturbation is proposed in [47], demonstrating that deep neural networks are vulnerable to small additive noise. These perturbations are typically subtle and imperceptible, hence are widely used to craft adversarial examples [62-64] to fool image classifiers. Later, the work in [48] devised universal adversarial perturbation (UAP), a model-agnostic and image-agnostic variant of adversarial perturbation. We find that UAP is a particularly competitive triggers for backdoor attacks against DMs due to their imperceptibility, generalization property and accessibility." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.517, + 0.652, + 0.533 + ], + "angle": 0, + "content": "3. UIBDiffusion" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.543, + 0.658, + 0.558 + ], + "angle": 0, + "content": "3.1. Threat Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.909, + 0.779 + ], + "angle": 0, + "content": "We adopt the same threat model and attack scenario presented in the existing backdoor attacks on DMs [23-25], where adversaries publish the backdoored models on websites such as HuggingFace, and users can acquire the pretrained models via third-party sources. Adversaries are expected to be able to inject the backdoor and verify the attack and generation performance before releasing the models. Users are supposed to get access to the backdoor models and the subset of clean training data and test the model utility on their side. Note that in practice, UIBDiffusion applies to a more stringent threat model where users can access the entire training dataset (i.e., contaminated dataset with poisoned samples) and apply any defenses before model deployment as the trigger is imperceptible and undetectable." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.808, + 0.804 + ], + "angle": 0, + "content": "3.2. UIBDiffusion Trigger Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Recall that our goal is to design a general, effective, and stealthy trigger (denoted as \\(\\tau\\)) for backdoor attacks against DMs. Due to the imperceptibility, generalization, and effectiveness properties of UAPs, they can be adapted as the UIBDiffusion trigger. We first briefly introduce the mechanism and generation process of UAPs. For a given image" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19188" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.184 + ], + "angle": 0, + "content": "classifier \\(\\mathcal{C}\\) and a data distribution \\(q\\), UAP algorithms aim to find a perturbation vector \\(v\\) such that \\(\\mathcal{C}(x + v) \\neq \\mathcal{C}(x)\\) for most \\(x \\sim q\\) (i.e., image agnostic). The original UAP algorithm [48] achieves this goal by measuring the minimal perturbation that pushes the perturbed input to the decision boundary, which can be mathematically expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.197, + 0.484, + 0.224 + ], + "angle": 0, + "content": "\\[\n\\Delta v _ {i} = \\underset {d} {\\arg \\min } | | d | | _ {2} \\quad \\mathbf {s}. \\mathbf {t}. \\quad \\mathcal {C} (x _ {i} + v + d) \\neq \\mathcal {C} (x _ {i}), (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.235, + 0.484, + 0.34 + ], + "angle": 0, + "content": "where \\(i\\) is the image index of a subset of the dataset selected for the computation. The final perturbation is updated as \\(\\mathcal{P}_{\\xi}(v + \\Delta v_i)\\), where \\(\\mathcal{P}_{\\xi}(\\cdot)\\) is a projection function and \\(\\xi\\) is the perturbation budget. Note that in the context of DMs backdoor, the perturbation \\(v\\) is actually the imperceptible trigger \\(\\tau\\). We use \\(\\tau\\) for consistent notation from now on to avoid confusion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.485, + 0.904 + ], + "angle": 0, + "content": "However, the original UAP algorithm using DeepFool [65] sometimes yields less effectiveness, and we experimentally observe it renders a relatively lower ASR as the backdoor trigger for DMs. Thus, we propose a novel trigger generation approach inspired by the GUAP [66], an enhanced version of UAP that is more effective and robust. Our trigger generation process adopts the advantage of both additive (i.e., \\( l_{\\infty} \\)-bounded) and non-additive (i.e., spatial transformation) perturbations, where the additive perturbation is adopted as the UIBDiffusion trigger \\( \\tau \\) and coordinately optimized with the non-additive perturbation. We denote the non-additive noise as \\( f \\). We employ a generator \\( \\mathcal{G}_{\\gamma} \\) (parameterized by \\( \\gamma \\)), which is akin to the generator in generative adversarial nets (GAN) [67] that take a latent noise \\( z \\) as input, and generate \\( f \\) and trigger \\( \\tau \\), as shown in lines 3, 4 and 5 in Algorithm 1. In contrast to the GAN that optimizes the generator under the guide of a discriminator, we update our trigger generator under the guide of an image classifier \\( \\mathcal{C} \\) (e.g., VGG and ResNet, etc.), which indicates whether the adversarial perturbation is effective or not, as expressed in Equation 2, where \\( \\mathcal{H}(\\cdot) \\) is the cross entropy loss and \\( \\otimes \\) represents for the spatial transformation operation. Meanwhile, the non-additive perturbation is updated according to the loss function expressed in Equation 3, where \\( S(\\cdot) \\) is summed max pooling values of \\( x \\) using a kernel of size 4 during the spatial transformation and \\( n \\) is the number of images in the selected subset of the dataset. By incorporating such a design, we can progressively optimize the generator and hence improve the effectiveness of the generated UIBDiffusion trigger, as described in lines 5, 7 and 8 in Algorithm 1, respectively. We then optimize the trigger generation process from latent variables, with the objective of maximizing the fooling rate against the image classifier. The completed trigger generation procedures are presented in Algorithm 1 and the concrete architecture of \\( \\mathcal{G}_{\\gamma} \\) and flow illustration are provided in the Appendix as they are similar to GAN." + }, + { + "type": "title", + "bbox": [ + 0.516, + 0.092, + 0.827, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 UIBDiffusion Trigger Generation" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.111, + 0.907, + 0.156 + ], + "angle": 0, + "content": "Require: Image set \\(X\\), classifier \\(\\mathcal{C}\\), latent noise \\(z\\), generator \\(\\mathcal{G}_{\\gamma}\\), training epoch \\(M\\), number of images \\(n\\), non-additive and additive perturbation budget \\(\\Gamma\\) and \\(\\xi\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.157, + 0.681, + 0.17 + ], + "angle": 0, + "content": "1: for \\( m = 1 \\dots M \\) do" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.172, + 0.68, + 0.184 + ], + "angle": 0, + "content": "2: for \\( i = 1 \\dots n \\) do" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.187, + 0.72, + 0.201 + ], + "angle": 0, + "content": "3: \\(x_{i}\\sim X,z\\sim \\mathcal{N}(0,1)\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.202, + 0.703, + 0.217 + ], + "angle": 0, + "content": "4: \\(\\mathcal{G}_{\\gamma_i}^{m - 1}(z)\\to f_i^{m - 1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.218, + 0.701, + 0.235 + ], + "angle": 0, + "content": "5: \\(\\mathcal{G}_{\\gamma_i}^{m - 1}(z)\\to \\tau_i^{m - 1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.235, + 0.82, + 0.259 + ], + "angle": 0, + "content": "6: \\(f = f_{i}^{m}\\gets \\frac{\\Gamma}{\\max_{i}\\sqrt{\\frac{1}{n}\\sum_{i}^{n}(\\mathcal{S}(x_{i}),f_{i}^{m - 1})}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.259, + 0.716, + 0.28 + ], + "angle": 0, + "content": "7: \\(\\tau = \\tau_{i}^{m}\\gets \\frac{\\xi}{||\\tau_{i}^{m - 1}||_{\\infty}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.28, + 0.904, + 0.298 + ], + "angle": 0, + "content": "8: \\(\\gamma_i^m\\gets \\gamma_i^{m - 1} - \\nabla_{\\gamma_i^{m - 1}}\\mathcal{L}_{\\mathcal{G}}\\left(\\mathcal{C}(x_i\\otimes f + \\tau),\\mathcal{C}(x_i)\\right)\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.299, + 0.613, + 0.31 + ], + "angle": 0, + "content": "9: end for" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.313, + 0.597, + 0.325 + ], + "angle": 0, + "content": "10: end for" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.328, + 0.657, + 0.343 + ], + "angle": 0, + "content": "11: return Trigger \\(\\tau\\)" + }, + { + "type": "list", + "bbox": [ + 0.52, + 0.157, + 0.904, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.361, + 0.904, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.462, + 0.907, + 0.532 + ], + "angle": 0, + "content": "Figure 2. Illustration of trigger mechanisms. UIBDiffusion triggers introduce a similar distribution shift as triggers in prior works, which secure a high attack performance. However, since our triggers do not have a specific pattern, they are difficult to remodel by trigger inversion algorithms." + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.564, + 0.907, + 0.581 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {G}} = - \\mathcal {H} (\\mathcal {C} (x \\otimes f + \\tau), \\mathcal {C} (x)), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.604, + 0.906, + 0.652 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f} = \\max _ {i} \\sqrt {\\frac {1}{n} \\sum_ {i} ^ {n} (\\mathcal {S} (x _ {i} , f))} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.903 + ], + "angle": 0, + "content": "The effectiveness of UIBDiffusion against backdoor attacks on generative models, such as DMs, stems from key insights into backdoor dynamics. As revealed in [23-25], backdoor for DMs needs to introduce an evident distribution shift and keep such a shift during the reversed Markov chain to secure a high ASR. We empirically find that the UIBDiffusion triggers also lead to a similar distribution shift when imposed on benign sampled noise, as shown in Fig. 2. However, because they are adapted from UAPs, which exhibit minimal, hard-to-detect patterns with subtle perturbations, we hypothesize that these triggers are difficult for trigger inversion algorithms to estimate accurately. Consequently, unlike the triggers with specific patterns, they are managed to circumvent the SOTA defenses and remain highly stealthy. We provide comprehensive and extensive evaluation results in Section 4 to support our conjecture." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19189" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.407, + 0.108 + ], + "angle": 0, + "content": "3.3. Inject Backdoor in Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.219 + ], + "angle": 0, + "content": "Prior attacks leveraged similar backdoor injection mechanisms with slight differences in sampler modifications [23-25]. UIBDiffusion employs the scheme presented in VillanDiffusion [25] since it provides a unified framework to implant backdoors in various types of diffusion models and samplers and achieves outstanding performance. We briefly introduce the injection process in this section." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.219, + 0.484, + 0.384 + ], + "angle": 0, + "content": "Clean diffusion process. Clean diffusion consists of forward and backward processes. The forward diffusion process diffuses images \\( x \\) from data distribution \\( q(x_0) \\) to the Gaussian distribution \\( \\mathcal{N}(0, \\mathbf{I}) \\) by progressively adding noise in each time step along a Markov chain. The process can be mathematically expressed as: \\( q(x_t | x_0) = \\mathcal{N}(\\alpha(t)x_0, \\beta^2(t)\\mathbf{I}) \\) with (1) \\( q(x_{T_{\\max}}) \\approx \\mathcal{N}(0, \\mathbf{I}) \\), (2) \\( q(x_{T_{\\min}}) \\approx q(x_0) \\) and \\( x_t, t \\in [T_{\\min}, T_{\\max}] \\), where \\( T_{\\min}, T_{\\max} \\) are the first and last timestep, and \\( \\alpha(t), \\beta(t) \\) are parameterized variables introduced in the original diffusion model work [1]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.385, + 0.484, + 0.477 + ], + "angle": 0, + "content": "Backward diffusion process reverse the Markov chain and generate images from a sampled Gaussian noise by gradually removing noises in each timestep and can be expressed as: \\( q(x_{t}|x_{t - 1}) = \\mathcal{N}(k_{t}x_{t - 1},w_{t}^{2}\\mathbf{I}) \\), and \\( x_{t}(x_{0},\\epsilon) = \\hat{\\alpha} (t)x_{0} + \\hat{\\beta} (t)\\epsilon \\), where \\( \\epsilon \\sim \\mathcal{N}(0,\\mathbf{I}) \\) and \\( k_{t},w_{t},\\hat{\\alpha} (t),\\hat{\\beta} (t) \\) are the same parameterized variables for the reverse process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.484, + 0.629 + ], + "angle": 0, + "content": "Backdoor and trigger injection. We employ the Villan-Diffusion framework for backdoor injection and modify the trigger injection process in data poisoning. VillanDiffusion proposes a unified backdoor injection framework by introducing the loss function shown in Equation 4, where \\(\\mathcal{L}_c\\) and \\(\\mathcal{L}_p\\) are used to optimize the original generation and backdoor attack tasks, respectively. \\(r\\) is the poisoned data, \\(y\\) is the backdoor target, \\(\\zeta\\) is a parameter to control backdoor against different samplers (e.g., ODE and SDE), and \\(\\eta_c, \\eta_p\\) are weights to balance \\(\\mathcal{L}_c\\) and \\(\\mathcal{L}_p\\)." + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.64, + 0.483, + 0.662 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\theta} \\left(\\eta_ {c}, \\eta_ {p}, x, t, \\epsilon , r, y, \\xi\\right) = \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.66, + 0.406, + 0.675 + ], + "angle": 0, + "content": "\\[\n\\eta_ {c} \\mathcal {L} _ {c} (x, t, \\epsilon) + \\eta_ {p} \\mathcal {L} _ {p} (x, t, \\epsilon , r, y, \\zeta).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.68, + 0.484, + 0.771 + ], + "angle": 0, + "content": "We then explain the difference in the trigger injection process between VillanDiffusion and UIBDiffusion. Villan-Diffusion injects the backdoor trigger following the method in Equation 5, where \\(\\mathbf{M}\\) is a mask and \\(g\\) is the trigger with specific patterns (e.g., gray box and a pair of glasses), and \\(\\odot\\) is the element-wise multiplication." + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.783, + 0.483, + 0.8 + ], + "angle": 0, + "content": "\\[\nr (x, g) = \\mathbf {M} \\odot g + (1 - \\mathbf {M}) \\odot x. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To better align with the adversarial perturbation, we inject our trigger using the approach in Equation 6, where \\(\\tau\\) is the imperceptible trigger synthesized by our proposed trigger generator and \\(\\varepsilon\\) is a variable that controls the strength of the added trigger. In contrast to triggers with recognizable patterns, the UIBDiffusion triggers are generated by" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "a trainable generator with subtle perturbations and hard-to-detect patterns, which makes them effective and stealthy. The backdoor and trigger injection process is summarized in Algorithm 2." + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.17, + 0.907, + 0.187 + ], + "angle": 0, + "content": "\\[\nr (x, \\tau) = x + \\varepsilon \\odot \\tau \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.212, + 0.906, + 0.24 + ], + "angle": 0, + "content": "Algorithm 2 UIBDiffusion Backdoor And Trigger Injection" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.244, + 0.906, + 0.288 + ], + "angle": 0, + "content": "Require: Backdoor Image Trigger \\(\\tau\\), Benign Dataset \\(X\\), Backdoor Target \\(y\\), Training parameters \\(\\theta\\), Sampler Randomness \\(\\zeta\\), Strength \\(\\varepsilon\\)" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.29, + 0.825, + 0.304 + ], + "angle": 0, + "content": "1: Backdoor DMs with Generated Trigger:" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.305, + 0.693, + 0.318 + ], + "angle": 0, + "content": "2: while not converge do" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.319, + 0.668, + 0.335 + ], + "angle": 0, + "content": "3: \\(\\{x,\\eta_c,\\eta_p\\} \\sim X\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.335, + 0.732, + 0.35 + ], + "angle": 0, + "content": "4: \\(t\\sim \\mathrm{Uniform}(\\{1,\\ldots ,T\\})\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.35, + 0.641, + 0.365 + ], + "angle": 0, + "content": "5: \\(\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.366, + 0.815, + 0.38 + ], + "angle": 0, + "content": "6: Trigger injection: \\(r(x,\\tau) = x + \\varepsilon \\odot \\tau\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.38, + 0.81, + 0.396 + ], + "angle": 0, + "content": "7: update \\(\\theta :\\nabla_{\\theta}\\mathcal{L}_{\\theta}^{I}(\\eta_{c},\\eta_{p},x,t,\\epsilon ,r,y,\\zeta)\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.396, + 0.613, + 0.408 + ], + "angle": 0, + "content": "8: end while" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.29, + 0.825, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.444, + 0.646, + 0.461 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.469, + 0.722, + 0.485 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.907, + 0.674 + ], + "angle": 0, + "content": "Datasets and models. We follow the same practice in existing backdoor attacks against DMs [23-25] and use CIFAR-10 \\((32 \\times 32)\\) [68] and CelebA-HQ \\((256 \\times 256)\\) [69] datasets for fair comparison. We evaluate UIB-Diffusion on three DMs: DDPM, LDM and NCSN, with eleven different samplers. We adopt the same training recipe from [25] and conduct our attack on pre-trained models (i.e., google/ddpm-cifar10-32 for CIFAR-10 and CompVis/ldm-celebahq-256 for CelebA-HQ) downloaded from Google HuggingFace organization and HuggingFace, respectively. Training hyper-parameters are provided in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Attack configurations and baselines. We conduct our experiments on NVIDIA L40S GPUs and employ Villan-Diffusion [25] as the backdoor injection framework. We provide two generated triggers guided by VGG [70] and ResNet [71] for CIFAR-10 and one proof-of-concept trigger for CelebA-HQ. We adopt \"hat\" and \"cat\" as targets for CIFAR-10 and \"cat\" as the target for CelebA-HQ, respectively. BadDiffusion [23] and VillanDiffusion [25] are considered as the baseline attacks for utility comparison and Elijah [26] and TERD [27] as the defenses for undetectability evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Evaluation metrics. We comprehensively evaluate the attack and generation performance of UIBDiffusion using four standard metrics. We exploit FID [72] score to examine the quality of generated clean images. Images with" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19190" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.904, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.282, + 0.843, + 0.298 + ], + "angle": 0, + "content": "Figure 3. FID and MSE comparison of BadDiffusion, VillanDiffusion, and UIBDiffusion over different poison rates." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.31, + 0.905, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.457, + 0.908, + 0.486 + ], + "angle": 0, + "content": "Table 1. ASR comparison between VillanDiffusion (top row) and UIBDiffusion (bottom row). UIBDiffusion achieves superior performance than the prior work at a low poison rate (e.g., \\(5\\%\\))." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.5, + 0.486, + 0.696 + ], + "angle": 0, + "content": "better quality tend to have lower FID scores. As for the attack effect, we employ Attack Success Rate (ASR), Mean Square Error (MSE) and Structural Similarity Index Measure (SSIM) for assessment. ASR is defined as the percentage of the target image generated from sampled noise stamped with backdoor triggers. Higher ASR indicates better utility. MSE is a metric to measure the pixel-level distance between the authentic and synthetic target images. A successful attack is expected to achieve low MSE. SSIM is a similar metric used to measure the similarity between two images (i.e., the real and generated target images). Superior attack performance will yield SSIM that approaches the value of 1." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.711, + 0.418, + 0.725 + ], + "angle": 0, + "content": "4.2. Results on Different Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.485, + 0.902 + ], + "angle": 0, + "content": "We first evaluate UIBDiffusion on different DMs. We present the results of DDPM in the main paper and organize the results of LDM and NCSN in the Appendix. As shown in Fig. 3, Table 1 and Table 2, UIBDiffusion achieves comparable FID as the clean diffusion model (i.e., poison rate \\(= 0\\%\\)). It can be seen that with the poison rates increased, all backdoor attacks yield a lower MSE, demonstrating a positive correlation between poison rates and attack effect. However, the generation ability of DMs will be accordingly affected when the poison rate is high. Notably, UIBDiffusion can reach a \\(100\\%\\) ASR with only \\(5\\%\\) poison rate" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.5, + 0.907, + 0.56 + ], + "angle": 0, + "content": "while VillanDiffusion gets \\(56\\%\\) ASR at this rate. In practice, attackers can achieve the expected attack effect without undermining the original generation ability of DMs, which further improves the stealthiness of the backdoor attack." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.564, + 0.907, + 0.616 + ], + "angle": 0, + "content": "
Poison Rate0%1%2%3%4%5%10%≥ 20%
BadDiffusion [23]3.80E-46.22E-31.48E-21.10E-20.370.990.990.99
VillanDiffusion [25]3.80E-43.92E-41.25E-21.69E-20.600.990.990.99
UIBDiffusion (Ours)3.80E-40.470.960.990.990.990.990.99
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.624, + 0.907, + 0.665 + ], + "angle": 0, + "content": "Table 2. SSIM comparison of BadDiffusion, VillanDiffusion and UIBDiffusion. UIBDiffusion achieves higher SSIM at lower poison rates of \\(1\\% \\sim 4\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.669, + 0.783, + 0.686 + ], + "angle": 0, + "content": "4.3. Results on Different Samplers" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.691, + 0.907, + 0.812 + ], + "angle": 0, + "content": "We then show the effectiveness of UIBDiffusion on different samplers. We conduct the experiments using DDPM model with 11 types of samplers. As shown in Fig. 4, UIB-Diffusion achieves a consistent performance of \\(100\\%\\) ASR across all samplers and yield a robust target image generation effect (i.e., MSE and SSIM) at various poison rates. We also observe that the ODE solvers outperform their SDE solver counterparts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.81, + 0.835 + ], + "angle": 0, + "content": "4.4. Resilience against SOTA Defenses" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Elijah [26] and TERD [27] are two SOTA defenses against backdoor attacks on DMs. They employ a similar idea of trigger inversion to estimate and reverse engineer the triggers. They then apply backdoor detection and removal" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19191" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.092, + 0.288, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.092, + 0.493, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.092, + 0.697, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.092, + 0.902, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.233, + 0.907, + 0.262 + ], + "angle": 0, + "content": "Figure 4. UIBDiffusion performance against eleven different samplers across various poison rates. UIBDiffusion yields consistently high ASR, SSIM, and MSE and even better FID on ODE-based samplers compared to SDE-based samplers." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.276, + 0.486, + 0.576 + ], + "angle": 0, + "content": "based on the reconstructed triggers. We show that UIBDiffusion is highly resilient to both attacks since our triggers are hard to remodel. Fig. 5 demonstrates the performance of UIBDiffusion against Elijah compared to BadDifuFSion and VillanDiffusion. It can be seen that UIBDiffusion achieves \\(100\\%\\) ASR and keeps the same MSE and SSIM before and after the backdoor detection and removal while ASR and SSIM of BadDiffusion and VillanDiffusion drop to \\(0\\%\\) after detection, indicating superior resilience against the SOTA defense. We then present the robustness of UIBDiffusion across various poison rates. Higher poison rates usually yield higher ASR but are easier to detect. However, as shown in Table 3, UIBDiffusion managed to escape detection at all the evaluated poison rates while maintaining \\(100\\%\\) ASR and high performance of SSIM and FID. It is worth noting that a \\(\\Delta SSIM = 0\\) indicates that UIBDiffusion successfully bypasses the defense, achieving \\(100\\%\\) targeted image generation. We attribute the fluctuations in FID to Elijah's use of an ineffective inverted trigger during backdoor removal." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.58, + 0.48, + 0.691 + ], + "angle": 0, + "content": "
BadDiffusion * [23]UIBDiffusion (Ours)
Poison RateDetectedASRΔSSIMΔFIDDetectedASRΔSSIMΔFID
5%100%0%-1.0-0.070%100%0-0.98
10%100%0%-1.0+0.180%100%0+0.15
20%100%0%-1.0-0.030%100%0+5.72
30%100%0%-1.0+0.150%100%0+12.59
50%100%0%-1.0+0.200%100%0+11.56
70%100%0%-1.0-0.110%100%0+14.85
90%100%0%-1.0-0.070%100%0+5.09
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.701, + 0.483, + 0.731 + ], + "angle": 0, + "content": "Table 3. Elijah against BadDiffusion and UIBDiffusion with different poison rates. “*” denotes results replicated from [26]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We evaluate the quality of the inverted trigger by visualizing the inverted triggers and testing if they could activate the backdoor as the actual triggers when imposed on benign sampled noise. Results are shown in Fig. 6. Sampled noise with inverted triggers should generate the target image if the defense succeeds and generate normal images otherwise. It can be seen that for triggers that have specific patterns (e.g., a stop sign and a gray box), Elijah can successfully estimate the authentic triggers and the inverted triggers are able to activate backdoors. On the other hand, the inverted trigger of UIBDiffusion fails to activate the backdoor injected in the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.276, + 0.907, + 0.351 + ], + "angle": 0, + "content": "DM, showing that the defense is ineffective against UIBDiffusion. We argue this is because the UIBDiffusion trigger is adapted from UAP, which is small and unstructured noise that is difficult to approximate based on the current trigger inversion loss function in Elijah and TERD." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.352, + 0.909, + 0.504 + ], + "angle": 0, + "content": "UIBDiffusion performs a similar undetectability against TERD as shown in Table 4. It achieves a \\(100\\%\\) True Positive Rate (TPR) and \\(0\\%\\) False Positive Rate (FPR), where TPR and FPR represent the percentage of the benign or backdoor sampled noise that is correctly detected. On the other hand, triggers with obvious patterns fail to bypass the trigger inversion algorithm and are all detected by TERD. Although TRED offers a more precise trigger inversion approach using a unified loss and trigger refinement, it is still ineffective against our proposed UIBDiffusion trigger." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.509, + 0.903, + 0.615 + ], + "angle": 0, + "content": "
Trigger-TargetTPR (↑)FPR (↓)
Ours-HAT100%0%
BOX-HAT*100%100%
BOX-SHOE*100%100%
BOX-CORNER*100%100%
STOP-HAT*100%100%
STOP-SHOE*100%100%
STOP-CORNER*100%100%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.624, + 0.906, + 0.654 + ], + "angle": 0, + "content": "Table 4. Resilience of UIBDiffusion against TERD. “*” denotes results replicated from [27]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.667, + 0.898, + 0.684 + ], + "angle": 0, + "content": "4.5. Comparison of Trigger Generation Approach" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We evaluate the necessity and effectiveness of our proposed trigger generation approach by comparing it to the original UAPs and GUAPs. We employ the generated UIBDiffusion trigger, the original UAP from [48] and the GUAP from [66] to poison the benign dataset and backdoor the DMs using these three triggers, respectively. As can be seen from Table 5, the UIBDiffusion trigger outperforms the original UAP and GUAP triggers at all four metrics by large margins, achieving much better attack and generation performance. The rationale behind this is that our trigger introduces an evident distribution shift. In contrast, the original UAP trigger only slightly affects the sampled noise distribution, and the spatial transformation of GUAP is image-specific and imposes a negative impact on the backdoor im" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19192" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.099, + 0.295, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.099, + 0.495, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.099, + 0.698, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.099, + 0.9, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.258, + 0.908, + 0.287 + ], + "angle": 0, + "content": "Figure 5. Performance comparison of BadDiffusion, VillanDiffusion and our UIBDiffusion before and after the Elijah defense. ASR: the higher the better; FID: the lower the better; MSE: the lower the better; SSIM: the higher the better." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.301, + 0.482, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.517, + 0.484, + 0.575 + ], + "angle": 0, + "content": "Figure 6. Visualization of trigger inversion. Inverted triggers of gray box and stop sign can activate backdoor, indicating a successful estimation of the actual triggers. On the other hand, inverted triggers of UIBDiffusion fail to activate the injected backdoor." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.589, + 0.483, + 0.62 + ], + "angle": 0, + "content": "plant. This shows the superiority and necessity of our trigger design choice." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.632, + 0.48, + 0.687 + ], + "angle": 0, + "content": "
FID (↓)MSE (↓)SSIM (↑)ASR (↑)
Original UAP20.889.22E-20.619043.75%
Original GUAP41.877.86E-20.19320.00%
UIBDiffusion19.775.30E-40.9967100%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.697, + 0.483, + 0.725 + ], + "angle": 0, + "content": "Table 5. Trigger performance comparison across the original UAP, GUAP and UIBDiffusion." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.754, + 0.277, + 0.77 + ], + "angle": 0, + "content": "4.6. Trigger Robustness" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.9 + ], + "angle": 0, + "content": "Subtle adversarial perturbations are susceptible to smoothing defenses [73] and could be mitigated by random noise. Thus, we perform a robust evaluation of the UIBDiffusion triggers and present the results in Table 6. UIBDiffusion shows strong robustness against random noise. We argue that the resilience is due to our trigger generation algorithm enhancing the adversarial perturbations via co-optimizing the additive and non-additive noise." + }, + { + "type": "table", + "bbox": [ + 0.519, + 0.298, + 0.906, + 0.363 + ], + "angle": 0, + "content": "
Random Noise1%2%5%10%20%
ASR99%99%100%100%100%
FID19.4819.5219.8019.3719.49
MSE2.36E-31.13E-38.03E-43.02E-45.30E-4
SSIM0.98650.99290.99440.99690.9957
" + }, + { + "type": "table_caption", + "bbox": [ + 0.532, + 0.373, + 0.887, + 0.387 + ], + "angle": 0, + "content": "Table 6. Robustness of UIBDiffusion against random noise." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.398, + 0.882, + 0.415 + ], + "angle": 0, + "content": "4.7. Quantitative Evaluation of Imperceptibility" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.42, + 0.907, + 0.557 + ], + "angle": 0, + "content": "We also quantitatively evaluate the imperceptibility of our trigger design using distance metrics of \\( l_{\\infty} \\) and \\( l_{2} \\) norms. We compute the distance between a poisoned image (stamping with our trigger and triggers used in prior works) and a benign image and summarize the results in Table 7. Our imperceptible triggers have much lower \\( l_{\\infty} \\) and \\( l_{2} \\) distances compared to triggers with specific patterns, which validates the stealthiness at the image level (i.e., the pixel space) of the UIBDiffusion trigger design." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.562, + 0.906, + 0.625 + ], + "angle": 0, + "content": "
Cifar 10CelebA-HQ-256
l∞l2l∞l2
UIBDiffusion Trigger0.041.800.0512.74
Prior Work0.476.210.8019.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.635, + 0.906, + 0.664 + ], + "angle": 0, + "content": "Table 7. Quantitative comparison of triggers imperceptibility using distance metrics of \\( l_{\\infty} \\) and \\( l_{2} \\) norms." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.681, + 0.634, + 0.696 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.706, + 0.907, + 0.888 + ], + "angle": 0, + "content": "In this work, we present UIBDiffusion, the first imperceptible backdoor attack against diffusion models. We propose a novel trigger generation method based on universal adversarial perturbations and demonstrate that such perturbations for attacking discriminative models can be adapted as general, effective, and stealthy triggers for backdoor attacks against diffusion models. We empirically reveal the underlying rationale why UIBDiffusion can achieve a high attack success rate while evading the trigger inversion defensive algorithms. We hope our work sheds light on the potential risks that diffusion models face and motivates future research to develop more comprehensive defenses." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19193" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.114, + 0.482, + 0.143 + ], + "angle": 0, + "content": "This work is supported in part by the National Science Foundation (NSF) SaTC-2426299 and SaTC-2413046." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.171, + 0.188, + 0.187 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.196, + 0.483, + 0.237 + ], + "angle": 0, + "content": "[1] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.239, + 0.484, + 0.28 + ], + "angle": 0, + "content": "[2] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In 9th International Conference on Learning Representations, ICLR 2021, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.282, + 0.483, + 0.35 + ], + "angle": 0, + "content": "[3] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, pages 10674-10685. IEEE, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.352, + 0.483, + 0.393 + ], + "angle": 0, + "content": "[4] William Peebles and Saining Xie. Scalable diffusion models with transformers. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, pages 4172-4182, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.395, + 0.482, + 0.491 + ], + "angle": 0, + "content": "[5] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, ICML 2024. OpenReview.net, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.493, + 0.483, + 0.547 + ], + "angle": 0, + "content": "[6] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.549, + 0.483, + 0.604 + ], + "angle": 0, + "content": "[7] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[8] Yuming Jiang, Tianxing Wu, Shuai Yang, Chenyang Si, Dahua Lin, Yu Qiao, Chen Change Loy, and Ziwei Liu. Videobooth: Diffusion-based video generation with image prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6689-6700, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.69, + 0.483, + 0.745 + ], + "angle": 0, + "content": "[9] Xiang Li, John Thickstun, Ishaan Gulrajani, Percy S Liang, and Tatsunori B Hashimoto. Diffusion-lm improves controllable text generation. Advances in Neural Information Processing Systems, 35:4328-4343, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[10] Xingqian Xu, Zhangyang Wang, Gong Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text, images and variations all in one diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7754-7765, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[11] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Myeonghun Jeong, Hyeongju Kim, Sung Jun Cheon, Byoung Jin Choi, and Nam Soo Kim. Diff-nts: A denois" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.196, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.906, + 0.12 + ], + "angle": 0, + "content": "ing diffusion model for text-to-speech. arXiv preprint arXiv:2104.01409, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.906, + 0.176 + ], + "angle": 0, + "content": "[13] Rongjie Huang, Max WY Lam, Jun Wang, Dan Su, Dong Yu, Yi Ren, and Zhou Zhao. Fastdiff: A fast conditional diffusion model for high-quality speech synthesis. arXiv preprint arXiv:2204.09934, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[14] Heeseung Kim, Sungwon Kim, and Sungroh Yoon. Guidedts: A diffusion model for text-to-speech via classifier guidance. In International Conference on Machine Learning, pages 11119-11133. PMLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[15] Vadim Popov, Ivan Vovk, Vladimir Gogoryan, Tasnama Sadekova, and Mikhail Kudinov. Grad-tts: A diffusion probabilistic model for text-to-speech. In International Conference on Machine Learning, pages 8599-8608. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[16] Yuancheng Wang, Zeqian Ju, Xu Tan, Lei He, Zhizheng Wu, Jiang Bian, et al. Audit: Audio editing by following instructions with latent diffusion models. Advances in Neural Information Processing Systems, 36:71340-71357, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[17] Shuai Shen, Wenliang Zhao, Zibin Meng, Wanhua Li, Zheng Zhu, Jie Zhou, and Jiwen Lu. Difftalk: Crafting diffusion models for generalized audio-driven portraits animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1982-1991, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.42, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[18] Bo Li, Yining Wang, Aarti Singh, and Yevgeniy Vorobeychik. Data poisoning attacks on factorization-based collaborative filtering. Advances in neural information processing systems, 29, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.476, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[19] Mengchen Zhao, Bo An, Yaodong Yu, Sulin Liu, and Sinno Pan. Data poisoning attacks on multi-task relationship learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.533, + 0.907, + 0.587 + ], + "angle": 0, + "content": "[20] Scott Alfeld, Xiaojin Zhu, and Paul Barford. Data poisoning attacks against autoregressive models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 30, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.59, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[21] Bingyin Zhao and Yingjie Lao. Towards class-oriented poisoning attacks against neural networks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3741-3750, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[22] Bingyin Zhao and Yingjie Lao. Clpa: Clean-label poisoning availability attacks using generative adversarial nets. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 9162-9170, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[23] Sheng-Yen Chou, Pin-Yu Chen, and Tsung-Yi Ho. How to backdoor diffusion models? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4015-4024, 2023. 1, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.76, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[24] Weixin Chen, Dawn Song, and Bo Li. Trojdiff: Trojan attacks on diffusion models with diverse targets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4035-4044, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[25] Sheng-Yen Chou, Pin-Yu Chen, and Tsung-Yi Ho. Villan-diffusion: A unified backdoor attack framework for diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Shengwei An, Sheng-Yen Chou, Kaiyuan Zhang, Qiuling Xu, Guanhong Tao, Guangyu Shen, Siyuan Cheng, Shiqing" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19194" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.16 + ], + "angle": 0, + "content": "Ma, Pin-Yu Chen, Tsung-Yi Ho, et al. Elijah: Eliminating backdoors injected in diffusion models via distribution shift. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 10847-10855, 2024. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.484, + 0.232 + ], + "angle": 0, + "content": "[27] Yichuan Mo, Hui Huang, Mingjie Li, Ang Li, and Yisen Wang. TERD: A unified framework for safeguarding diffusion models against backdoors. In *Forty-first International Conference on Machine Learning*, ICML 2024, 2024. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.483, + 0.286 + ], + "angle": 0, + "content": "[28] Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying vulnerabilities in the machine learning model supply chain. arXiv preprint arXiv:1708.06733, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.483, + 0.329 + ], + "angle": 0, + "content": "[29] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. CoRR, abs/1712.05526, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.483, + 0.399 + ], + "angle": 0, + "content": "[30] Yingqi Liu, Shiqing Ma, Yousra Aafer, Wen-Chuan Lee, Juan Zhai, Weihang Wang, and Xiangyu Zhang. Trojaning attack on neural networks. In 25th Annual Network and Distributed System Security Symposium, NDSS. The Internet Society, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.483, + 0.468 + ], + "angle": 0, + "content": "[31] Yuanshun Yao, Huiying Li, Haitao Zheng, and Ben Y. Zhao. Latent backdoor attacks on deep neural networks. In Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security, CCS 2019, pages 2041-2055. ACM, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.483, + 0.539 + ], + "angle": 0, + "content": "[32] Yunfei Liu, Xingjun Ma, James Bailey, and Feng Lu. Reflection backdoor: A natural backdoor attack on deep neural networks. In Computer Vision - ECCV 2020, Proceedings, Part X, volume 12355 of Lecture Notes in Computer Science, pages 182-199. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.483, + 0.607 + ], + "angle": 0, + "content": "[33] Yuezun Li, Yiming Li, Baoyuan Wu, Longkang Li, Ran He, and Siwei Lyu. Invisible backdoor attack with sample-specific triggers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16463-16472, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.483, + 0.65 + ], + "angle": 0, + "content": "[34] Tuan Anh Nguyen and Anh Tuan Tran. Wanet - imperceptible warping-based backdoor attack. In 9th International Conference on Learning Representations, ICLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.483, + 0.733 + ], + "angle": 0, + "content": "[35] Siyuan Cheng, Yingqi Liu, Shiqing Ma, and Xiangyu Zhang. Deep feature space trojan attack of neural networks by controlled detoxification. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, pages 1148–1156. AAAI Press, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[36] Khoa Doan, Yingjie Lao, Weijie Zhao, and Ping Li. Lira: Learnable, imperceptible and robust backdoor attacks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11966-11976, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[37] Khoa Doan, Yingjie Lao, and Ping Li. Backdoor attack with imperceptible input and latent modification. Advances in Neural Information Processing Systems, 34:18944-18957, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[38] Khoa D Doan, Yingjie Lao, and Ping Li. Marksman backdoor: Backdoor attacks with arbitrary target class. Advances in Neural Information Processing Systems, 35:38260-38273, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[39] Aniruddha Saha, Akshayvarun Subramanya, and Hamed Pirsiavash. Hidden trigger backdoor attacks. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, pages 11957-11965. AAAI Press, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.906, + 0.233 + ], + "angle": 0, + "content": "[40] Hossein Souri, Liam Fowl, Rama Chellappa, Micah Goldblum, and Tom Goldstein. Sleeper agent: Scalable hidden trigger backdoors for neural networks trained from scratch. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.906, + 0.318 + ], + "angle": 0, + "content": "[41] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Anti-backdoor learning: Training clean models on poisoned data. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 14900-14912, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.906, + 0.375 + ], + "angle": 0, + "content": "[42] Qi Zhao and Christian Wressnegger. Adversarily robust anti-backdoor learning. In Proceedings of the 2024 Workshop on Artificial Intelligence and Security, AISec 2024, pages 77-88. ACM, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.906, + 0.446 + ], + "angle": 0, + "content": "[43] Bolun Wang, Yuanshun Yao, Shawn Shan, Huiying Li, Bimal Viswanath, Haitao Zheng, and Ben Y. Zhao. Neural cleanse: Identifying and mitigating backdoor attacks in neural networks. In 2019 IEEE Symposium on Security and Privacy, SP 2019, pages 707-723. IEEE, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.906, + 0.516 + ], + "angle": 0, + "content": "[44] Dongxian Wu and Yisen Wang. Adversarial neuron pruning purifies backdoored deep models. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, pages 16913-16925, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.906, + 0.573 + ], + "angle": 0, + "content": "[45] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Neural attention distillation: Erasing backdoor triggers from deep neural networks. In 9th International Conference on Learning Representations, ICLR 2021, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.906, + 0.617 + ], + "angle": 0, + "content": "[46] Bingyin Zhao and Yingjie Lao. Ultraclean: A simple framework to train robust neural networks against backdoor attacks. arXiv preprint arXiv:2312.10657, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.906, + 0.687 + ], + "angle": 0, + "content": "[47] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In 2nd International Conference on Learning Representations, ICLR 2014, 2014. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.906, + 0.759 + ], + "angle": 0, + "content": "[48] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Universal adversarial perturbations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1765-1773, 2017. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.906, + 0.829 + ], + "angle": 0, + "content": "[49] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, pages 11895-11907, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[50] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In 9th International Conference on Learning Representations, ICLR 2021. OpenReview.net, 2021." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19195" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[51] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.203 + ], + "angle": 0, + "content": "[52] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. In The Eleventh International Conference on Learning Representations, ICLR 2023, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[53] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[54] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.425 + ], + "angle": 0, + "content": "[55] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.482, + 0.466 + ], + "angle": 0, + "content": "[56] Joseph Clements and Yingjie Lao. Hardware trojan attacks on neural networks. arXiv preprint arXiv:1806.05768, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.537 + ], + "angle": 0, + "content": "[57] Shengfang Zhai, Yinpeng Dong, Qingni Shen, Shi Pu, Yuejian Fang, and Hang Su. Text-to-image diffusion models can be easily backdoored through multimodal data poisoning. In Proceedings of the 31st ACM International Conference on Multimedia, MM 2023, pages 1577-1587. ACM, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[58] Yihao Huang, Qing Guo, and Felix Juefei-Xu. Zero-day backdoor attack against text-to-image diffusion models via personalization. CoRR, abs/2305.10701, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[59] Lukas Struppek, Dominik Hintersdorf, and Kristian Kersting. Rickrolling the artist: Injecting backdoors into text encoders for text-to-image synthesis. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, pages 4561-4573. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[60] Bryant Chen, Wilka Carvalho, Nathalie Baracaldo, Heiko Ludwig, Benjamin Edwards, Taesung Lee, Ian M. Moly, and Biplav Srivastava. Detecting backdoor attacks on deep neural networks by activation clustering. In Workshop on Artificial Intelligence Safety 2019 co-located with the Thirty-Third AAAI Conference on Artificial Intelligence 2019 (AAAI-19), volume 2301 of CEUR Workshop Proceedings. CEUR-WS.org, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[61] Yi Zeng, Won Park, Z. Morley Mao, and Ruoxi Jia. Rethinking the backdoor attacks' triggers: A frequency perspective. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, pages 16453-16461. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[62] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In 3rd International Conference on Learning Representations, ICLR 2015, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[63] Alexey Kurakin, Ian J. Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In 5th International" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Conference on Learning Representations, ICLR 2017, Workshop Track Proceedings, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[64] Chaowei Xiao, Bo Li, Jun-Yan Zhu, Warren He, Mingyan Liu, and Dawn Song. Generating adversarial examples with adversarial networks. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJ-CAI 2018, pages 3905–3911. ijcai.org, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[65] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2574–2582, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[66] Yanghao Zhang, Wenjie Ruan, Fu Wang, and Xiaowei Huang. Generalizing universal adversarial attacks beyond additive perturbations. In 2020 IEEE International Conference on Data Mining (ICDM), pages 1412-1417. IEEE, 2020. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.332, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[67] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.429 + ], + "angle": 0, + "content": "[68] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical Report, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.432, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[69] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In 2015 IEEE International Conference on Computer Vision, ICCV 2015, pages 3730-3738. IEEE Computer Society, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.905, + 0.557 + ], + "angle": 0, + "content": "[70] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In Yoshua Bengio and Yann LeCun, editors, 3rd International Conference on Learning Representations, ICLR 2015, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.559, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[71] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.616, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[72] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, pages 6626-6637, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.7, + 0.905, + 0.755 + ], + "angle": 0, + "content": "[73] Hadi Salman, Mingjie Sun, Greg Yang, Ashish Kapoor, and J Zico Kolter. Denoised smoothing: A provable defense for pretrained classifiers. Advances in Neural Information Processing Systems, 33:21945-21957, 2020. 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19196" + } + ] +] \ No newline at end of file diff --git a/2025/UIBDiffusion_ Universal Imperceptible Backdoor Attack for Diffusion Models/7f0574b2-e282-4c3e-a072-3d64d026983a_origin.pdf b/2025/UIBDiffusion_ Universal Imperceptible Backdoor Attack for Diffusion Models/7f0574b2-e282-4c3e-a072-3d64d026983a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4d706e7554d78cc2e6f1f15edee06411b7f008f2 --- /dev/null +++ b/2025/UIBDiffusion_ Universal Imperceptible Backdoor Attack for Diffusion Models/7f0574b2-e282-4c3e-a072-3d64d026983a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a45f39b7fb1d643ebd231da1e209a5d94a9ea39778f9da8e5f8d96abe6db57c +size 4027379 diff --git a/2025/UIBDiffusion_ Universal Imperceptible Backdoor Attack for Diffusion Models/full.md b/2025/UIBDiffusion_ Universal Imperceptible Backdoor Attack for Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f843a2b0f21dbbd4ffab5003323cde21e3b7e4de --- /dev/null +++ b/2025/UIBDiffusion_ Universal Imperceptible Backdoor Attack for Diffusion Models/full.md @@ -0,0 +1,336 @@ +# UIBDiffusion: Universal Imperceptible Backdoor Attack for Diffusion Models + +Yuning Han $^{1,*}$ Bingyin Zhao $^{2,*}$ Rui Chu $^{3}$ Feng Luo $^{4}$ Biplab Sikdar $^{2}$ Yingjie Lao $^{3,\dagger}$ + +$^{1}$ Columbia University $^{2}$ National University of Singapore $^{3}$ Tufts University $^{4}$ Clemson University + +# Abstract + +Recent studies show that diffusion models (DMs) are vulnerable to backdoor attacks. Existing backdoor attacks impose unconcealed triggers (e.g., a gray box and eyeglasses) that contain evident patterns, rendering remarkable attack effects yet easy detection upon human inspection and defensive algorithms. While it is possible to improve stealthiness by reducing the strength of the backdoor, doing so can significantly compromise its generality and effectiveness. In this paper, we propose UIBDiffusion, the universal imperceptible backdoor attack for diffusion models, which allows us to achieve superior attack and generation performance while evading state-of-the-art defenses. We propose a novel trigger generation approach based on universal adversarial perturbations (UAPs) and reveal that such perturbations, which are initially devised for fooling pre-trained discriminative models, can be adapted as potent imperceptible backdoor triggers for DMs. We evaluate UIBDiffusion on multiple types of DMs with different kinds of samplers across various datasets and targets. Experimental results demonstrate that UIBDiffusion brings three advantages: 1) Universality, the imperceptible trigger is universal (i.e., image and model agnostic) where a single trigger is effective to any images and all diffusion models with different samplers; 2) Utility, it achieves comparable generation quality (e.g., FID) and even better attack success rate (i.e., ASR) at low poison rates compared to the prior works; and 3) Undetectability, UIBDiffusion is plausible to human perception and can bypass Elijah and TERD, the SOTA defenses against backdoors for DMs. Code is available at https://github.com/TheLaoLab/UIBDiffusion. + +# 1. Introduction + +Diffusion models (DMs) [1-3] have emerged as the state-of-the-art generative paradigm in recent years and achieved unprecedented success in image [4, 5], video [6-8] text [9-11] and audio [12-17] synthesis. The success is highly + +attributed to training the foundation model on large-scale datasets [5]. Such a practice places DMs at risk of being compromised by data poisoning attacks [18-22], where adversaries contaminate the clean training dataset with a small fraction of malicious samples and undermine the model performance. This imposes a critical security concern on the practical deployment of DMs. + +Backdoor attacks are emerging threats to DMs. Recent studies [23-25] show that DMs are vulnerable to such attacks, which implant backdoor into victim models by training them on poisoned datasets and manipulate their behavior through inputs stamped with predefined triggers. Existing backdoor attacks for DMs usually employ unconcealed triggers such as a gray box, an image of Hello Kitty, and eyeglasses in the forward and reversed (i.e., sampling) diffusion process, as shown in Fig. 1. While such trigger designs achieve superior attack performance and preserve the decent generation ability of victim models, they are easy to detect via human inspection and advanced defensive algorithms as they have obvious patterns that can be precisely remodeled via reverse engineering [26, 27]. Motivated by this observation, we aim to design a general, powerful yet stealthy trigger. However, even for the traditional backdoor attacks against discriminative models (e.g., image classifiers) [28-30], it is challenging to acquire such a trigger since it either requires complex and carefully designed generators to generate image-specific or model-specific triggers [31-38] or suffers from a relatively low attack success rate (ASR) [39, 40], or fails to evade modern defenses [41-46]. This raises an interesting question in backdoor attacks against DMs: Is there a trigger simultaneously possessing generality, effectiveness, and stealthiness for backdoor attacks against DMs? + +Driving by the question, in this work, we propose UIB-Diffusion, the universal imperceptible backdoor triggers for DMs that bring the following advantages: 1) Universality: the trigger is image-agnostic and model-agnostic so that it can be applied to arbitrary images and DMs; 2) Utility: the trigger can achieve high ASR while maintaining the generation capability of DMs; 3) Undetectability: the trigger is plausible to humans and capable of circumventing + +![](images/489eae18e67f17ebad7b0366a582d838a4a71684ff6d96bc63b80848149dd359.jpg) +Figure 1. Illustrations of the forward diffusion process (top block with the yellow background) and backward diffusion process (bottom block with the blue background) of a clean diffusion model (blue dash line), VillanDiffusion [25] (red dash line) and UIBDiffusion (Ours, green dash line). The UIBDiffusion trigger $(\tau)$ is plausible to humans in all phases from data poisoning to forward diffusion and backward diffusion while the glasses trigger $(\mathbf{g})$ in prior works is perceptible. UIBDiffusion trigger is highly effective since it introduces a similar distribution shift as the glasses trigger, which secures the attack performance during the backward diffusion process. However, it is hard to detect as the trigger does not possess specific patterns and cannot be inverted by existing defensive algorithms. We empirically verify the effectiveness in Section 4. + +state-of-the-art defenses. Inspired by the adversarial perturbation [47] originally designed for attacking pre-trained image classifiers, we reveal that universal adversarial perturbations [48] (UAPs), a variant of adversarial perturbations, naturally hold the aforementioned properties and can be adapted as the UIBDiffusion triggers. First, UAPs are image-agnostic and model-agnostic. Second, we demonstrate that they are effective triggers because they introduce a similar distribution shift as prior works, yet they are stealthy because they do not possess clear patterns and are hard to estimate by existing trigger inversion algorithms. Moreover, they are trivial and imperceptible noises that humans can barely identify, as shown in Fig. 1. Our contributions are summarized as follows: + +- To the best of our knowledge, UIBDiffusion is the first imperceptible backdoor attack against diffusion models. +- We discover that adversarial perturbations designed for discriminative models can be adapted as backdoor trig- + +gers for generative models. We reveal why such triggers can achieve high ASR and evade SOTA defenses. + +- We propose a practical trigger generation approach adapted from UAP, which is specifically designed to enhance universality, attack effectiveness and accessibility of triggers. +- We experimentally demonstrate that UIBDiffusion is highly effective towards multiple diffusion models with different samplers and achieves comparable generation quality on benign input compared to clean DMs and superior attack success rate and input stealthiness compared to prior works. +- While existing works face a trade-off between utility and undetectability, we empirically show that UIBDiffusion triggers are highly resistant to the SOTA trigger inversion-based defenses, effectively bypassing Elijah [26] and TERD [27], the state-of-the-art defenses against backdoors on DMs. + +# 2. Related Works + +# 2.1. Diffusion Models + +Diffusion models are latent variable models composed of two processes. The forward diffusion process imposes Gaussian noise on input data in each time step along the Markov chain and progressively diffuses the data distribution to an isotropic Gaussian distribution. The backward diffusion process is a reversed Markov chain that gradually denoises the sampled noise from the Gaussian distribution to the input data distribution. Mainstream diffusion models include DDPM [1], LDM [3] and NCSN [49-51], etc. Despite the excellent performance, DMs suffer from slow sampling during the backward process. There exists a line of works that leverage different techniques such as generalized non-Markovian chain and Ordinary Differential Equation (ODE) to improve the samplers and accelerate the generation process: DDIM [2], DEIS [52], UniPC [53], Heun [54], DPM-Solver [55] and DPM-Solver++ [7]. We comprehensively evaluate UIBDiffusion on these models and samplers to showcase the utility of our trigger. + +# 2.2. Backdoor Attacks and Defenses on Diffusion Models + +Attacks. Backdoor attacks [28-30, 56], originally explored in the context of discriminative models such as image classifiers, induce models to make wrong predictions on triggered images. Although having been extensively studied for discriminative models, backdoor attacks are under-explored in diffusion models. BadDiffusion [23], TrojDiff [24] and VillanDiffusion [25] are pioneering works in this field. Unlike traditional backdoor attacks, those targeting DMs aim to force models to synthesize a targeted image upon malicious inputs (i.e., noises with triggers) while generating normal images upon benign inputs (i.e., isotropic Gaussian noises). BadDiffusion shows how to backdoor DMs with unconcealed triggers such as a gray box and a stop sign. TrojDiff proposes to use a whole image (e.g., Hello Kitty) as the trigger and demonstrates that backdoor can be generalized to DDIM and different adversarial goals. However, both attacks are limited to certain DMs (e.g., DDPM) and samplers (e.g., ancestral and DDIM). VillanDiffusion resolves the limitations and offers a unified framework for all kinds of DMs and samplers. Other works show backdoor can be injected to text-to-image DMs and activated by text prompts [57-59]. However, these attacks draw less attention from defenders as they can eliminate the backdoor by cleansing the text encoder. Thus, our attack focuses on the primary defensive objective of the defender: pixel level (i.e., image) trigger generation. + +Defenses. Backdoor attacks against DMs are shown to be immune to defenses designed for traditional image classifier backdoors [23, 25, 44, 60, 61]. The reason is that DMs + +take sampled noises as inputs while image classifiers take natural images as inputs [27]. Elijah [26] is the first defensive algorithm specifically proposed against DM backdoors. It presents a trigger inversion scheme that can approximate the triggers and detect and remove the injected backdoor accordingly. TERD [27] employs a similar trigger estimation strategy and improves the trigger inversion process via a unified loss and trigger refinement. Based on the reconstructed trigger, it subsequently proposes to measure the KL divergence between benign and reversed distribution to detect the backdoor injected in models. Elijah and TERD achieve remarkable detection performance on existing DM backdoors. Thus, we evaluate UIBDiffusion against them and show that UIBDiffusion can bypass both defenses. + +# 2.3. Adversarial Perturbation + +The concept of adversarial perturbation is proposed in [47], demonstrating that deep neural networks are vulnerable to small additive noise. These perturbations are typically subtle and imperceptible, hence are widely used to craft adversarial examples [62-64] to fool image classifiers. Later, the work in [48] devised universal adversarial perturbation (UAP), a model-agnostic and image-agnostic variant of adversarial perturbation. We find that UAP is a particularly competitive triggers for backdoor attacks against DMs due to their imperceptibility, generalization property and accessibility. + +# 3. UIBDiffusion + +# 3.1. Threat Model + +We adopt the same threat model and attack scenario presented in the existing backdoor attacks on DMs [23-25], where adversaries publish the backdoored models on websites such as HuggingFace, and users can acquire the pretrained models via third-party sources. Adversaries are expected to be able to inject the backdoor and verify the attack and generation performance before releasing the models. Users are supposed to get access to the backdoor models and the subset of clean training data and test the model utility on their side. Note that in practice, UIBDiffusion applies to a more stringent threat model where users can access the entire training dataset (i.e., contaminated dataset with poisoned samples) and apply any defenses before model deployment as the trigger is imperceptible and undetectable. + +# 3.2. UIBDiffusion Trigger Generation + +Recall that our goal is to design a general, effective, and stealthy trigger (denoted as $\tau$ ) for backdoor attacks against DMs. Due to the imperceptibility, generalization, and effectiveness properties of UAPs, they can be adapted as the UIBDiffusion trigger. We first briefly introduce the mechanism and generation process of UAPs. For a given image + +classifier $\mathcal{C}$ and a data distribution $q$ , UAP algorithms aim to find a perturbation vector $v$ such that $\mathcal{C}(x + v) \neq \mathcal{C}(x)$ for most $x \sim q$ (i.e., image agnostic). The original UAP algorithm [48] achieves this goal by measuring the minimal perturbation that pushes the perturbed input to the decision boundary, which can be mathematically expressed as: + +$$ +\Delta v _ {i} = \underset {d} {\arg \min } | | d | | _ {2} \quad \mathbf {s}. \mathbf {t}. \quad \mathcal {C} (x _ {i} + v + d) \neq \mathcal {C} (x _ {i}), (1) +$$ + +where $i$ is the image index of a subset of the dataset selected for the computation. The final perturbation is updated as $\mathcal{P}_{\xi}(v + \Delta v_i)$ , where $\mathcal{P}_{\xi}(\cdot)$ is a projection function and $\xi$ is the perturbation budget. Note that in the context of DMs backdoor, the perturbation $v$ is actually the imperceptible trigger $\tau$ . We use $\tau$ for consistent notation from now on to avoid confusion. + +However, the original UAP algorithm using DeepFool [65] sometimes yields less effectiveness, and we experimentally observe it renders a relatively lower ASR as the backdoor trigger for DMs. Thus, we propose a novel trigger generation approach inspired by the GUAP [66], an enhanced version of UAP that is more effective and robust. Our trigger generation process adopts the advantage of both additive (i.e., $l_{\infty}$ -bounded) and non-additive (i.e., spatial transformation) perturbations, where the additive perturbation is adopted as the UIBDiffusion trigger $\tau$ and coordinately optimized with the non-additive perturbation. We denote the non-additive noise as $f$ . We employ a generator $\mathcal{G}_{\gamma}$ (parameterized by $\gamma$ ), which is akin to the generator in generative adversarial nets (GAN) [67] that take a latent noise $z$ as input, and generate $f$ and trigger $\tau$ , as shown in lines 3, 4 and 5 in Algorithm 1. In contrast to the GAN that optimizes the generator under the guide of a discriminator, we update our trigger generator under the guide of an image classifier $\mathcal{C}$ (e.g., VGG and ResNet, etc.), which indicates whether the adversarial perturbation is effective or not, as expressed in Equation 2, where $\mathcal{H}(\cdot)$ is the cross entropy loss and $\otimes$ represents for the spatial transformation operation. Meanwhile, the non-additive perturbation is updated according to the loss function expressed in Equation 3, where $S(\cdot)$ is summed max pooling values of $x$ using a kernel of size 4 during the spatial transformation and $n$ is the number of images in the selected subset of the dataset. By incorporating such a design, we can progressively optimize the generator and hence improve the effectiveness of the generated UIBDiffusion trigger, as described in lines 5, 7 and 8 in Algorithm 1, respectively. We then optimize the trigger generation process from latent variables, with the objective of maximizing the fooling rate against the image classifier. The completed trigger generation procedures are presented in Algorithm 1 and the concrete architecture of $\mathcal{G}_{\gamma}$ and flow illustration are provided in the Appendix as they are similar to GAN. + +# Algorithm 1 UIBDiffusion Trigger Generation + +Require: Image set $X$ , classifier $\mathcal{C}$ , latent noise $z$ , generator $\mathcal{G}_{\gamma}$ , training epoch $M$ , number of images $n$ , non-additive and additive perturbation budget $\Gamma$ and $\xi$ + +1: for $m = 1 \dots M$ do +2: for $i = 1 \dots n$ do +3: $x_{i}\sim X,z\sim \mathcal{N}(0,1)$ +4: $\mathcal{G}_{\gamma_i}^{m - 1}(z)\to f_i^{m - 1}$ +5: $\mathcal{G}_{\gamma_i}^{m - 1}(z)\to \tau_i^{m - 1}$ +6: $f = f_{i}^{m}\gets \frac{\Gamma}{\max_{i}\sqrt{\frac{1}{n}\sum_{i}^{n}(\mathcal{S}(x_{i}),f_{i}^{m - 1})}}$ +7: $\tau = \tau_{i}^{m}\gets \frac{\xi}{||\tau_{i}^{m - 1}||_{\infty}}$ +8: $\gamma_i^m\gets \gamma_i^{m - 1} - \nabla_{\gamma_i^{m - 1}}\mathcal{L}_{\mathcal{G}}\left(\mathcal{C}(x_i\otimes f + \tau),\mathcal{C}(x_i)\right)$ +9: end for +10: end for +11: return Trigger $\tau$ + +![](images/533c428f255e159c4b25a11fc9a0f5305c85b43c18c4e5b2d7083c7b7d2cce04.jpg) +Figure 2. Illustration of trigger mechanisms. UIBDiffusion triggers introduce a similar distribution shift as triggers in prior works, which secure a high attack performance. However, since our triggers do not have a specific pattern, they are difficult to remodel by trigger inversion algorithms. + +$$ +\mathcal {L} _ {\mathcal {G}} = - \mathcal {H} (\mathcal {C} (x \otimes f + \tau), \mathcal {C} (x)), \tag {2} +$$ + +$$ +\mathcal {L} _ {f} = \max _ {i} \sqrt {\frac {1}{n} \sum_ {i} ^ {n} (\mathcal {S} (x _ {i} , f))} \tag {3} +$$ + +The effectiveness of UIBDiffusion against backdoor attacks on generative models, such as DMs, stems from key insights into backdoor dynamics. As revealed in [23-25], backdoor for DMs needs to introduce an evident distribution shift and keep such a shift during the reversed Markov chain to secure a high ASR. We empirically find that the UIBDiffusion triggers also lead to a similar distribution shift when imposed on benign sampled noise, as shown in Fig. 2. However, because they are adapted from UAPs, which exhibit minimal, hard-to-detect patterns with subtle perturbations, we hypothesize that these triggers are difficult for trigger inversion algorithms to estimate accurately. Consequently, unlike the triggers with specific patterns, they are managed to circumvent the SOTA defenses and remain highly stealthy. We provide comprehensive and extensive evaluation results in Section 4 to support our conjecture. + +# 3.3. Inject Backdoor in Diffusion Models + +Prior attacks leveraged similar backdoor injection mechanisms with slight differences in sampler modifications [23-25]. UIBDiffusion employs the scheme presented in VillanDiffusion [25] since it provides a unified framework to implant backdoors in various types of diffusion models and samplers and achieves outstanding performance. We briefly introduce the injection process in this section. + +Clean diffusion process. Clean diffusion consists of forward and backward processes. The forward diffusion process diffuses images $x$ from data distribution $q(x_0)$ to the Gaussian distribution $\mathcal{N}(0, \mathbf{I})$ by progressively adding noise in each time step along a Markov chain. The process can be mathematically expressed as: $q(x_t | x_0) = \mathcal{N}(\alpha(t)x_0, \beta^2(t)\mathbf{I})$ with (1) $q(x_{T_{\max}}) \approx \mathcal{N}(0, \mathbf{I})$ , (2) $q(x_{T_{\min}}) \approx q(x_0)$ and $x_t, t \in [T_{\min}, T_{\max}]$ , where $T_{\min}, T_{\max}$ are the first and last timestep, and $\alpha(t), \beta(t)$ are parameterized variables introduced in the original diffusion model work [1]. + +Backward diffusion process reverse the Markov chain and generate images from a sampled Gaussian noise by gradually removing noises in each timestep and can be expressed as: $q(x_{t}|x_{t - 1}) = \mathcal{N}(k_{t}x_{t - 1},w_{t}^{2}\mathbf{I})$ , and $x_{t}(x_{0},\epsilon) = \hat{\alpha} (t)x_{0} + \hat{\beta} (t)\epsilon$ , where $\epsilon \sim \mathcal{N}(0,\mathbf{I})$ and $k_{t},w_{t},\hat{\alpha} (t),\hat{\beta} (t)$ are the same parameterized variables for the reverse process. + +Backdoor and trigger injection. We employ the Villan-Diffusion framework for backdoor injection and modify the trigger injection process in data poisoning. VillanDiffusion proposes a unified backdoor injection framework by introducing the loss function shown in Equation 4, where $\mathcal{L}_c$ and $\mathcal{L}_p$ are used to optimize the original generation and backdoor attack tasks, respectively. $r$ is the poisoned data, $y$ is the backdoor target, $\zeta$ is a parameter to control backdoor against different samplers (e.g., ODE and SDE), and $\eta_c, \eta_p$ are weights to balance $\mathcal{L}_c$ and $\mathcal{L}_p$ . + +$$ +\mathcal {L} _ {\theta} \left(\eta_ {c}, \eta_ {p}, x, t, \epsilon , r, y, \xi\right) = \tag {4} +$$ + +$$ +\eta_ {c} \mathcal {L} _ {c} (x, t, \epsilon) + \eta_ {p} \mathcal {L} _ {p} (x, t, \epsilon , r, y, \zeta). +$$ + +We then explain the difference in the trigger injection process between VillanDiffusion and UIBDiffusion. Villan-Diffusion injects the backdoor trigger following the method in Equation 5, where $\mathbf{M}$ is a mask and $g$ is the trigger with specific patterns (e.g., gray box and a pair of glasses), and $\odot$ is the element-wise multiplication. + +$$ +r (x, g) = \mathbf {M} \odot g + (1 - \mathbf {M}) \odot x. \tag {5} +$$ + +To better align with the adversarial perturbation, we inject our trigger using the approach in Equation 6, where $\tau$ is the imperceptible trigger synthesized by our proposed trigger generator and $\varepsilon$ is a variable that controls the strength of the added trigger. In contrast to triggers with recognizable patterns, the UIBDiffusion triggers are generated by + +a trainable generator with subtle perturbations and hard-to-detect patterns, which makes them effective and stealthy. The backdoor and trigger injection process is summarized in Algorithm 2. + +$$ +r (x, \tau) = x + \varepsilon \odot \tau \tag {6} +$$ + +Algorithm 2 UIBDiffusion Backdoor And Trigger Injection + +Require: Backdoor Image Trigger $\tau$ , Benign Dataset $X$ , Backdoor Target $y$ , Training parameters $\theta$ , Sampler Randomness $\zeta$ , Strength $\varepsilon$ + +1: Backdoor DMs with Generated Trigger: +2: while not converge do +3: $\{x,\eta_c,\eta_p\} \sim X$ +4: $t\sim \mathrm{Uniform}(\{1,\ldots ,T\})$ +5: $\epsilon \sim \mathcal{N}(0,\mathbf{I})$ +6: Trigger injection: $r(x,\tau) = x + \varepsilon \odot \tau$ +7: update $\theta :\nabla_{\theta}\mathcal{L}_{\theta}^{I}(\eta_{c},\eta_{p},x,t,\epsilon ,r,y,\zeta)$ +8: end while + +# 4. Experiments + +# 4.1. Experimental Settings + +Datasets and models. We follow the same practice in existing backdoor attacks against DMs [23-25] and use CIFAR-10 $(32 \times 32)$ [68] and CelebA-HQ $(256 \times 256)$ [69] datasets for fair comparison. We evaluate UIB-Diffusion on three DMs: DDPM, LDM and NCSN, with eleven different samplers. We adopt the same training recipe from [25] and conduct our attack on pre-trained models (i.e., google/ddpm-cifar10-32 for CIFAR-10 and CompVis/ldm-celebahq-256 for CelebA-HQ) downloaded from Google HuggingFace organization and HuggingFace, respectively. Training hyper-parameters are provided in the Appendix. + +Attack configurations and baselines. We conduct our experiments on NVIDIA L40S GPUs and employ Villan-Diffusion [25] as the backdoor injection framework. We provide two generated triggers guided by VGG [70] and ResNet [71] for CIFAR-10 and one proof-of-concept trigger for CelebA-HQ. We adopt "hat" and "cat" as targets for CIFAR-10 and "cat" as the target for CelebA-HQ, respectively. BadDiffusion [23] and VillanDiffusion [25] are considered as the baseline attacks for utility comparison and Elijah [26] and TERD [27] as the defenses for undetectability evaluation. + +Evaluation metrics. We comprehensively evaluate the attack and generation performance of UIBDiffusion using four standard metrics. We exploit FID [72] score to examine the quality of generated clean images. Images with + +![](images/46501edc54f55a090cc86554d8b2d984367ddef3be71c2002cd226e698a3e6a7.jpg) +Figure 3. FID and MSE comparison of BadDiffusion, VillanDiffusion, and UIBDiffusion over different poison rates. + +![](images/df0779aecaa78320b9628415772caa2d151413100c8b85d2a94a91f27b8dc861.jpg) +Table 1. ASR comparison between VillanDiffusion (top row) and UIBDiffusion (bottom row). UIBDiffusion achieves superior performance than the prior work at a low poison rate (e.g., $5\%$ ). + +better quality tend to have lower FID scores. As for the attack effect, we employ Attack Success Rate (ASR), Mean Square Error (MSE) and Structural Similarity Index Measure (SSIM) for assessment. ASR is defined as the percentage of the target image generated from sampled noise stamped with backdoor triggers. Higher ASR indicates better utility. MSE is a metric to measure the pixel-level distance between the authentic and synthetic target images. A successful attack is expected to achieve low MSE. SSIM is a similar metric used to measure the similarity between two images (i.e., the real and generated target images). Superior attack performance will yield SSIM that approaches the value of 1. + +# 4.2. Results on Different Diffusion Models + +We first evaluate UIBDiffusion on different DMs. We present the results of DDPM in the main paper and organize the results of LDM and NCSN in the Appendix. As shown in Fig. 3, Table 1 and Table 2, UIBDiffusion achieves comparable FID as the clean diffusion model (i.e., poison rate $= 0\%$ ). It can be seen that with the poison rates increased, all backdoor attacks yield a lower MSE, demonstrating a positive correlation between poison rates and attack effect. However, the generation ability of DMs will be accordingly affected when the poison rate is high. Notably, UIBDiffusion can reach a $100\%$ ASR with only $5\%$ poison rate + +while VillanDiffusion gets $56\%$ ASR at this rate. In practice, attackers can achieve the expected attack effect without undermining the original generation ability of DMs, which further improves the stealthiness of the backdoor attack. + +
Poison Rate0%1%2%3%4%5%10%≥ 20%
BadDiffusion [23]3.80E-46.22E-31.48E-21.10E-20.370.990.990.99
VillanDiffusion [25]3.80E-43.92E-41.25E-21.69E-20.600.990.990.99
UIBDiffusion (Ours)3.80E-40.470.960.990.990.990.990.99
+ +Table 2. SSIM comparison of BadDiffusion, VillanDiffusion and UIBDiffusion. UIBDiffusion achieves higher SSIM at lower poison rates of $1\% \sim 4\%$ . + +# 4.3. Results on Different Samplers + +We then show the effectiveness of UIBDiffusion on different samplers. We conduct the experiments using DDPM model with 11 types of samplers. As shown in Fig. 4, UIB-Diffusion achieves a consistent performance of $100\%$ ASR across all samplers and yield a robust target image generation effect (i.e., MSE and SSIM) at various poison rates. We also observe that the ODE solvers outperform their SDE solver counterparts. + +# 4.4. Resilience against SOTA Defenses + +Elijah [26] and TERD [27] are two SOTA defenses against backdoor attacks on DMs. They employ a similar idea of trigger inversion to estimate and reverse engineer the triggers. They then apply backdoor detection and removal + +![](images/df442f1a67b91a3928b46354a75ad9a1c06e4ea1d12159dc574bb2464df0da50.jpg) +Figure 4. UIBDiffusion performance against eleven different samplers across various poison rates. UIBDiffusion yields consistently high ASR, SSIM, and MSE and even better FID on ODE-based samplers compared to SDE-based samplers. + +![](images/281a17324825adcdc6951bb4c4cdcbca827950368da8c059cddc41de22291528.jpg) + +![](images/5df83fc19ec4015c3af41dc08e630cfa33169215d59626b0f29e151746456e4e.jpg) + +![](images/d00e50ed3acef7d9fdadfdf20b13d25e74f38f7617c5db20d66e8787b35df162.jpg) + +based on the reconstructed triggers. We show that UIBDiffusion is highly resilient to both attacks since our triggers are hard to remodel. Fig. 5 demonstrates the performance of UIBDiffusion against Elijah compared to BadDifuFSion and VillanDiffusion. It can be seen that UIBDiffusion achieves $100\%$ ASR and keeps the same MSE and SSIM before and after the backdoor detection and removal while ASR and SSIM of BadDiffusion and VillanDiffusion drop to $0\%$ after detection, indicating superior resilience against the SOTA defense. We then present the robustness of UIBDiffusion across various poison rates. Higher poison rates usually yield higher ASR but are easier to detect. However, as shown in Table 3, UIBDiffusion managed to escape detection at all the evaluated poison rates while maintaining $100\%$ ASR and high performance of SSIM and FID. It is worth noting that a $\Delta SSIM = 0$ indicates that UIBDiffusion successfully bypasses the defense, achieving $100\%$ targeted image generation. We attribute the fluctuations in FID to Elijah's use of an ineffective inverted trigger during backdoor removal. + +
BadDiffusion * [23]UIBDiffusion (Ours)
Poison RateDetectedASRΔSSIMΔFIDDetectedASRΔSSIMΔFID
5%100%0%-1.0-0.070%100%0-0.98
10%100%0%-1.0+0.180%100%0+0.15
20%100%0%-1.0-0.030%100%0+5.72
30%100%0%-1.0+0.150%100%0+12.59
50%100%0%-1.0+0.200%100%0+11.56
70%100%0%-1.0-0.110%100%0+14.85
90%100%0%-1.0-0.070%100%0+5.09
+ +We evaluate the quality of the inverted trigger by visualizing the inverted triggers and testing if they could activate the backdoor as the actual triggers when imposed on benign sampled noise. Results are shown in Fig. 6. Sampled noise with inverted triggers should generate the target image if the defense succeeds and generate normal images otherwise. It can be seen that for triggers that have specific patterns (e.g., a stop sign and a gray box), Elijah can successfully estimate the authentic triggers and the inverted triggers are able to activate backdoors. On the other hand, the inverted trigger of UIBDiffusion fails to activate the backdoor injected in the + +DM, showing that the defense is ineffective against UIBDiffusion. We argue this is because the UIBDiffusion trigger is adapted from UAP, which is small and unstructured noise that is difficult to approximate based on the current trigger inversion loss function in Elijah and TERD. + +UIBDiffusion performs a similar undetectability against TERD as shown in Table 4. It achieves a $100\%$ True Positive Rate (TPR) and $0\%$ False Positive Rate (FPR), where TPR and FPR represent the percentage of the benign or backdoor sampled noise that is correctly detected. On the other hand, triggers with obvious patterns fail to bypass the trigger inversion algorithm and are all detected by TERD. Although TRED offers a more precise trigger inversion approach using a unified loss and trigger refinement, it is still ineffective against our proposed UIBDiffusion trigger. + +Table 3. Elijah against BadDiffusion and UIBDiffusion with different poison rates. “*” denotes results replicated from [26]. + +
Trigger-TargetTPR (↑)FPR (↓)
Ours-HAT100%0%
BOX-HAT*100%100%
BOX-SHOE*100%100%
BOX-CORNER*100%100%
STOP-HAT*100%100%
STOP-SHOE*100%100%
STOP-CORNER*100%100%
+ +Table 4. Resilience of UIBDiffusion against TERD. “*” denotes results replicated from [27]. + +# 4.5. Comparison of Trigger Generation Approach + +We evaluate the necessity and effectiveness of our proposed trigger generation approach by comparing it to the original UAPs and GUAPs. We employ the generated UIBDiffusion trigger, the original UAP from [48] and the GUAP from [66] to poison the benign dataset and backdoor the DMs using these three triggers, respectively. As can be seen from Table 5, the UIBDiffusion trigger outperforms the original UAP and GUAP triggers at all four metrics by large margins, achieving much better attack and generation performance. The rationale behind this is that our trigger introduces an evident distribution shift. In contrast, the original UAP trigger only slightly affects the sampled noise distribution, and the spatial transformation of GUAP is image-specific and imposes a negative impact on the backdoor im + +![](images/ec68786bb9c549cfe46620a356bf3bb5c13a8a48ed2ca3ae36ba1c131a24cddf.jpg) +Figure 5. Performance comparison of BadDiffusion, VillanDiffusion and our UIBDiffusion before and after the Elijah defense. ASR: the higher the better; FID: the lower the better; MSE: the lower the better; SSIM: the higher the better. + +![](images/c5dd4453540f91a2e28b13f96e930829895ef883715b870c9dd5847c6c6b972b.jpg) + +![](images/b593e16ad5586ee094b3c9db4eb849a18edf39bb0d92eca2cab7efffb02fbb65.jpg) + +![](images/50fb0826eea200ac7edd7ff0ff4b5353e69a0a1416dae3ab5fcd90787e2bdaee.jpg) + +![](images/1e03395fafd941eeb48ee2340d21b80e3352e419d0f23b30ef9a84eb96827cca.jpg) +Figure 6. Visualization of trigger inversion. Inverted triggers of gray box and stop sign can activate backdoor, indicating a successful estimation of the actual triggers. On the other hand, inverted triggers of UIBDiffusion fail to activate the injected backdoor. + +plant. This shows the superiority and necessity of our trigger design choice. + +
FID (↓)MSE (↓)SSIM (↑)ASR (↑)
Original UAP20.889.22E-20.619043.75%
Original GUAP41.877.86E-20.19320.00%
UIBDiffusion19.775.30E-40.9967100%
+ +# 4.6. Trigger Robustness + +Subtle adversarial perturbations are susceptible to smoothing defenses [73] and could be mitigated by random noise. Thus, we perform a robust evaluation of the UIBDiffusion triggers and present the results in Table 6. UIBDiffusion shows strong robustness against random noise. We argue that the resilience is due to our trigger generation algorithm enhancing the adversarial perturbations via co-optimizing the additive and non-additive noise. + +Table 5. Trigger performance comparison across the original UAP, GUAP and UIBDiffusion. + +
Random Noise1%2%5%10%20%
ASR99%99%100%100%100%
FID19.4819.5219.8019.3719.49
MSE2.36E-31.13E-38.03E-43.02E-45.30E-4
SSIM0.98650.99290.99440.99690.9957
+ +# 4.7. Quantitative Evaluation of Imperceptibility + +We also quantitatively evaluate the imperceptibility of our trigger design using distance metrics of $l_{\infty}$ and $l_{2}$ norms. We compute the distance between a poisoned image (stamping with our trigger and triggers used in prior works) and a benign image and summarize the results in Table 7. Our imperceptible triggers have much lower $l_{\infty}$ and $l_{2}$ distances compared to triggers with specific patterns, which validates the stealthiness at the image level (i.e., the pixel space) of the UIBDiffusion trigger design. + +Table 6. Robustness of UIBDiffusion against random noise. + +
Cifar 10CelebA-HQ-256
l∞l2l∞l2
UIBDiffusion Trigger0.041.800.0512.74
Prior Work0.476.210.8019.84
+ +Table 7. Quantitative comparison of triggers imperceptibility using distance metrics of $l_{\infty}$ and $l_{2}$ norms. + +# 5. Conclusion + +In this work, we present UIBDiffusion, the first imperceptible backdoor attack against diffusion models. We propose a novel trigger generation method based on universal adversarial perturbations and demonstrate that such perturbations for attacking discriminative models can be adapted as general, effective, and stealthy triggers for backdoor attacks against diffusion models. We empirically reveal the underlying rationale why UIBDiffusion can achieve a high attack success rate while evading the trigger inversion defensive algorithms. We hope our work sheds light on the potential risks that diffusion models face and motivates future research to develop more comprehensive defenses. + +# Acknowledgment + +This work is supported in part by the National Science Foundation (NSF) SaTC-2426299 and SaTC-2413046. + +# References + +[1] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3, 5 +[2] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In 9th International Conference on Learning Representations, ICLR 2021, 2021. 3 +[3] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, pages 10674-10685. IEEE, 2022. 1, 3 +[4] William Peebles and Saining Xie. Scalable diffusion models with transformers. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, pages 4172-4182, 2023. 1 +[5] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, ICML 2024. OpenReview.net, 2024. 1 +[6] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 1 +[7] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 3 +[8] Yuming Jiang, Tianxing Wu, Shuai Yang, Chenyang Si, Dahua Lin, Yu Qiao, Chen Change Loy, and Ziwei Liu. Videobooth: Diffusion-based video generation with image prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6689-6700, 2024. 1 +[9] Xiang Li, John Thickstun, Ishaan Gulrajani, Percy S Liang, and Tatsunori B Hashimoto. Diffusion-lm improves controllable text generation. Advances in Neural Information Processing Systems, 35:4328-4343, 2022. 1 +[10] Xingqian Xu, Zhangyang Wang, Gong Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text, images and variations all in one diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7754-7765, 2023. +[11] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1 +[12] Myeonghun Jeong, Hyeongju Kim, Sung Jun Cheon, Byoung Jin Choi, and Nam Soo Kim. Diff-nts: A denois + +ing diffusion model for text-to-speech. arXiv preprint arXiv:2104.01409, 2021. 1 +[13] Rongjie Huang, Max WY Lam, Jun Wang, Dan Su, Dong Yu, Yi Ren, and Zhou Zhao. Fastdiff: A fast conditional diffusion model for high-quality speech synthesis. arXiv preprint arXiv:2204.09934, 2022. +[14] Heeseung Kim, Sungwon Kim, and Sungroh Yoon. Guidedts: A diffusion model for text-to-speech via classifier guidance. In International Conference on Machine Learning, pages 11119-11133. PMLR, 2022. +[15] Vadim Popov, Ivan Vovk, Vladimir Gogoryan, Tasnama Sadekova, and Mikhail Kudinov. Grad-tts: A diffusion probabilistic model for text-to-speech. In International Conference on Machine Learning, pages 8599-8608. PMLR, 2021. +[16] Yuancheng Wang, Zeqian Ju, Xu Tan, Lei He, Zhizheng Wu, Jiang Bian, et al. Audit: Audio editing by following instructions with latent diffusion models. Advances in Neural Information Processing Systems, 36:71340-71357, 2023. +[17] Shuai Shen, Wenliang Zhao, Zibin Meng, Wanhua Li, Zheng Zhu, Jie Zhou, and Jiwen Lu. Difftalk: Crafting diffusion models for generalized audio-driven portraits animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1982-1991, 2023. 1 +[18] Bo Li, Yining Wang, Aarti Singh, and Yevgeniy Vorobeychik. Data poisoning attacks on factorization-based collaborative filtering. Advances in neural information processing systems, 29, 2016. 1 +[19] Mengchen Zhao, Bo An, Yaodong Yu, Sulin Liu, and Sinno Pan. Data poisoning attacks on multi-task relationship learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. +[20] Scott Alfeld, Xiaojin Zhu, and Paul Barford. Data poisoning attacks against autoregressive models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 30, 2016. +[21] Bingyin Zhao and Yingjie Lao. Towards class-oriented poisoning attacks against neural networks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3741-3750, 2022. +[22] Bingyin Zhao and Yingjie Lao. Clpa: Clean-label poisoning availability attacks using generative adversarial nets. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 9162-9170, 2022. 1 +[23] Sheng-Yen Chou, Pin-Yu Chen, and Tsung-Yi Ho. How to backdoor diffusion models? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4015-4024, 2023. 1, 3, 4, 5, 6, 7 +[24] Weixin Chen, Dawn Song, and Bo Li. Trojdiff: Trojan attacks on diffusion models with diverse targets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4035-4044, 2023. 3 +[25] Sheng-Yen Chou, Pin-Yu Chen, and Tsung-Yi Ho. Villan-diffusion: A unified backdoor attack framework for diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 4, 5, 6 +[26] Shengwei An, Sheng-Yen Chou, Kaiyuan Zhang, Qiuling Xu, Guanhong Tao, Guangyu Shen, Siyuan Cheng, Shiqing + +Ma, Pin-Yu Chen, Tsung-Yi Ho, et al. Elijah: Eliminating backdoors injected in diffusion models via distribution shift. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 10847-10855, 2024. 1, 2, 3, 5, 6, 7 +[27] Yichuan Mo, Hui Huang, Mingjie Li, Ang Li, and Yisen Wang. TERD: A unified framework for safeguarding diffusion models against backdoors. In *Forty-first International Conference on Machine Learning*, ICML 2024, 2024. 1, 2, 3, 5, 6, 7 +[28] Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying vulnerabilities in the machine learning model supply chain. arXiv preprint arXiv:1708.06733, 2017. 1, 3 +[29] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. CoRR, abs/1712.05526, 2017. +[30] Yingqi Liu, Shiqing Ma, Yousra Aafer, Wen-Chuan Lee, Juan Zhai, Weihang Wang, and Xiangyu Zhang. Trojaning attack on neural networks. In 25th Annual Network and Distributed System Security Symposium, NDSS. The Internet Society, 2018. 1, 3 +[31] Yuanshun Yao, Huiying Li, Haitao Zheng, and Ben Y. Zhao. Latent backdoor attacks on deep neural networks. In Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security, CCS 2019, pages 2041-2055. ACM, 2019. 1 +[32] Yunfei Liu, Xingjun Ma, James Bailey, and Feng Lu. Reflection backdoor: A natural backdoor attack on deep neural networks. In Computer Vision - ECCV 2020, Proceedings, Part X, volume 12355 of Lecture Notes in Computer Science, pages 182-199. Springer, 2020. +[33] Yuezun Li, Yiming Li, Baoyuan Wu, Longkang Li, Ran He, and Siwei Lyu. Invisible backdoor attack with sample-specific triggers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16463-16472, 2021. +[34] Tuan Anh Nguyen and Anh Tuan Tran. Wanet - imperceptible warping-based backdoor attack. In 9th International Conference on Learning Representations, ICLR, 2021. +[35] Siyuan Cheng, Yingqi Liu, Shiqing Ma, and Xiangyu Zhang. Deep feature space trojan attack of neural networks by controlled detoxification. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, pages 1148–1156. AAAI Press, 2021. +[36] Khoa Doan, Yingjie Lao, Weijie Zhao, and Ping Li. Lira: Learnable, imperceptible and robust backdoor attacks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11966-11976, 2021. +[37] Khoa Doan, Yingjie Lao, and Ping Li. Backdoor attack with imperceptible input and latent modification. Advances in Neural Information Processing Systems, 34:18944-18957, 2021. +[38] Khoa D Doan, Yingjie Lao, and Ping Li. Marksman backdoor: Backdoor attacks with arbitrary target class. Advances in Neural Information Processing Systems, 35:38260-38273, 2022. 1 + +[39] Aniruddha Saha, Akshayvarun Subramanya, and Hamed Pirsiavash. Hidden trigger backdoor attacks. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, pages 11957-11965. AAAI Press, 2020. 1 +[40] Hossein Souri, Liam Fowl, Rama Chellappa, Micah Goldblum, and Tom Goldstein. Sleeper agent: Scalable hidden trigger backdoors for neural networks trained from scratch. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, 2022. 1 +[41] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Anti-backdoor learning: Training clean models on poisoned data. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 14900-14912, 2021. 1 +[42] Qi Zhao and Christian Wressnegger. Adversarily robust anti-backdoor learning. In Proceedings of the 2024 Workshop on Artificial Intelligence and Security, AISec 2024, pages 77-88. ACM, 2024. +[43] Bolun Wang, Yuanshun Yao, Shawn Shan, Huiying Li, Bimal Viswanath, Haitao Zheng, and Ben Y. Zhao. Neural cleanse: Identifying and mitigating backdoor attacks in neural networks. In 2019 IEEE Symposium on Security and Privacy, SP 2019, pages 707-723. IEEE, 2019. +[44] Dongxian Wu and Yisen Wang. Adversarial neuron pruning purifies backdoored deep models. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, pages 16913-16925, 2021. 3 +[45] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Neural attention distillation: Erasing backdoor triggers from deep neural networks. In 9th International Conference on Learning Representations, ICLR 2021, 2021. +[46] Bingyin Zhao and Yingjie Lao. Ultraclean: A simple framework to train robust neural networks against backdoor attacks. arXiv preprint arXiv:2312.10657, 2023. 1 +[47] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In 2nd International Conference on Learning Representations, ICLR 2014, 2014. 2, 3 +[48] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Universal adversarial perturbations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1765-1773, 2017. 2, 3, 4, 7 +[49] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, pages 11895-11907, 2019. 3 +[50] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In 9th International Conference on Learning Representations, ICLR 2021. OpenReview.net, 2021. + +[51] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, 2020. 3 +[52] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. In The Eleventh International Conference on Learning Representations, ICLR 2023, 2023. 3 +[53] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, 2023. 3 +[54] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, 2022. 3 +[55] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, 2022. 3 +[56] Joseph Clements and Yingjie Lao. Hardware trojan attacks on neural networks. arXiv preprint arXiv:1806.05768, 2018. 3 +[57] Shengfang Zhai, Yinpeng Dong, Qingni Shen, Shi Pu, Yuejian Fang, and Hang Su. Text-to-image diffusion models can be easily backdoored through multimodal data poisoning. In Proceedings of the 31st ACM International Conference on Multimedia, MM 2023, pages 1577-1587. ACM, 2023. 3 +[58] Yihao Huang, Qing Guo, and Felix Juefei-Xu. Zero-day backdoor attack against text-to-image diffusion models via personalization. CoRR, abs/2305.10701, 2023. +[59] Lukas Struppek, Dominik Hintersdorf, and Kristian Kersting. Rickrolling the artist: Injecting backdoors into text encoders for text-to-image synthesis. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, pages 4561-4573. IEEE, 2023. 3 +[60] Bryant Chen, Wilka Carvalho, Nathalie Baracaldo, Heiko Ludwig, Benjamin Edwards, Taesung Lee, Ian M. Moly, and Biplav Srivastava. Detecting backdoor attacks on deep neural networks by activation clustering. In Workshop on Artificial Intelligence Safety 2019 co-located with the Thirty-Third AAAI Conference on Artificial Intelligence 2019 (AAAI-19), volume 2301 of CEUR Workshop Proceedings. CEUR-WS.org, 2019. 3 +[61] Yi Zeng, Won Park, Z. Morley Mao, and Ruoxi Jia. Rethinking the backdoor attacks' triggers: A frequency perspective. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, pages 16453-16461. IEEE, 2021. 3 +[62] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In 3rd International Conference on Learning Representations, ICLR 2015, 2015. 3 +[63] Alexey Kurakin, Ian J. Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In 5th International + +Conference on Learning Representations, ICLR 2017, Workshop Track Proceedings, 2017. +[64] Chaowei Xiao, Bo Li, Jun-Yan Zhu, Warren He, Mingyan Liu, and Dawn Song. Generating adversarial examples with adversarial networks. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJ-CAI 2018, pages 3905–3911. ijcai.org, 2018. 3 +[65] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2574–2582, 2016. 4 +[66] Yanghao Zhang, Wenjie Ruan, Fu Wang, and Xiaowei Huang. Generalizing universal adversarial attacks beyond additive perturbations. In 2020 IEEE International Conference on Data Mining (ICDM), pages 1412-1417. IEEE, 2020. 4, 7 +[67] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 4 +[68] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical Report, 2009. 5 +[69] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In 2015 IEEE International Conference on Computer Vision, ICCV 2015, pages 3730-3738. IEEE Computer Society, 2015. 5 +[70] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In Yoshua Bengio and Yann LeCun, editors, 3rd International Conference on Learning Representations, ICLR 2015, 2015. 5 +[71] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5 +[72] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, pages 6626-6637, 2017. 5 +[73] Hadi Salman, Mingjie Sun, Greg Yang, Ashish Kapoor, and J Zico Kolter. Denoised smoothing: A provable defense for pretrained classifiers. Advances in Neural Information Processing Systems, 33:21945-21957, 2020. 8 \ No newline at end of file diff --git a/2025/UIBDiffusion_ Universal Imperceptible Backdoor Attack for Diffusion Models/images.zip b/2025/UIBDiffusion_ Universal Imperceptible Backdoor Attack for Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bfe6860d5a65fd10415273c165cbc6af09c70426 --- /dev/null +++ b/2025/UIBDiffusion_ Universal Imperceptible Backdoor Attack for Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d4bd0c037762d7fe90a7592b126f05defc036438b1df2c8c7f56eca9e2a545bd +size 672453 diff --git a/2025/UIBDiffusion_ Universal Imperceptible Backdoor Attack for Diffusion Models/layout.json b/2025/UIBDiffusion_ Universal Imperceptible Backdoor Attack for Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..64388260d196a47743f82ba07cf5627983d80af1 --- /dev/null +++ b/2025/UIBDiffusion_ Universal Imperceptible Backdoor Attack for Diffusion Models/layout.json @@ -0,0 +1,10093 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 102, + 545, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 102, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 65, + 102, + 545, + 120 + ], + "type": "text", + "content": "UIBDiffusion: Universal Imperceptible Backdoor Attack for Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 142, + 528, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 142, + 528, + 158 + ], + "spans": [ + { + "bbox": [ + 99, + 142, + 528, + 158 + ], + "type": "text", + "content": "Yuning Han" + }, + { + "bbox": [ + 99, + 142, + 528, + 158 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 99, + 142, + 528, + 158 + ], + "type": "text", + "content": " Bingyin Zhao" + }, + { + "bbox": [ + 99, + 142, + 528, + 158 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 99, + 142, + 528, + 158 + ], + "type": "text", + "content": " Rui Chu" + }, + { + "bbox": [ + 99, + 142, + 528, + 158 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 99, + 142, + 528, + 158 + ], + "type": "text", + "content": " Feng Luo" + }, + { + "bbox": [ + 99, + 142, + 528, + 158 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 99, + 142, + 528, + 158 + ], + "type": "text", + "content": " Biplab Sikdar" + }, + { + "bbox": [ + 99, + 142, + 528, + 158 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 99, + 142, + 528, + 158 + ], + "type": "text", + "content": " Yingjie Lao" + }, + { + "bbox": [ + 99, + 142, + 528, + 158 + ], + "type": "inline_equation", + "content": "^{3,\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 170, + 559, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 170, + 559, + 186 + ], + "spans": [ + { + "bbox": [ + 61, + 170, + 559, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 170, + 559, + 186 + ], + "type": "text", + "content": "Columbia University " + }, + { + "bbox": [ + 61, + 170, + 559, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 170, + 559, + 186 + ], + "type": "text", + "content": "National University of Singapore " + }, + { + "bbox": [ + 61, + 170, + 559, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 170, + 559, + 186 + ], + "type": "text", + "content": "Tufts University " + }, + { + "bbox": [ + 61, + 170, + 559, + 186 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 170, + 559, + 186 + ], + "type": "text", + "content": "Clemson University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 238, + 296, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 238, + 296, + 596 + ], + "spans": [ + { + "bbox": [ + 56, + 238, + 296, + 596 + ], + "type": "text", + "content": "Recent studies show that diffusion models (DMs) are vulnerable to backdoor attacks. Existing backdoor attacks impose unconcealed triggers (e.g., a gray box and eyeglasses) that contain evident patterns, rendering remarkable attack effects yet easy detection upon human inspection and defensive algorithms. While it is possible to improve stealthiness by reducing the strength of the backdoor, doing so can significantly compromise its generality and effectiveness. In this paper, we propose UIBDiffusion, the universal imperceptible backdoor attack for diffusion models, which allows us to achieve superior attack and generation performance while evading state-of-the-art defenses. We propose a novel trigger generation approach based on universal adversarial perturbations (UAPs) and reveal that such perturbations, which are initially devised for fooling pre-trained discriminative models, can be adapted as potent imperceptible backdoor triggers for DMs. We evaluate UIBDiffusion on multiple types of DMs with different kinds of samplers across various datasets and targets. Experimental results demonstrate that UIBDiffusion brings three advantages: 1) Universality, the imperceptible trigger is universal (i.e., image and model agnostic) where a single trigger is effective to any images and all diffusion models with different samplers; 2) Utility, it achieves comparable generation quality (e.g., FID) and even better attack success rate (i.e., ASR) at low poison rates compared to the prior works; and 3) Undetectability, UIBDiffusion is plausible to human perception and can bypass Elijah and TERD, the SOTA defenses against backdoors for DMs. Code is available at https://github.com/TheLaoLab/UIBDiffusion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 619, + 135, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 619, + 135, + 631 + ], + "spans": [ + { + "bbox": [ + 56, + 619, + 135, + 631 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 638, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 638, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 638, + 295, + 687 + ], + "type": "text", + "content": "Diffusion models (DMs) [1-3] have emerged as the state-of-the-art generative paradigm in recent years and achieved unprecedented success in image [4, 5], video [6-8] text [9-11] and audio [12-17] synthesis. The success is highly" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 214, + 553, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 214, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 553, + 299 + ], + "type": "text", + "content": "attributed to training the foundation model on large-scale datasets [5]. Such a practice places DMs at risk of being compromised by data poisoning attacks [18-22], where adversaries contaminate the clean training dataset with a small fraction of malicious samples and undermine the model performance. This imposes a critical security concern on the practical deployment of DMs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 302, + 555, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 555, + 613 + ], + "type": "text", + "content": "Backdoor attacks are emerging threats to DMs. Recent studies [23-25] show that DMs are vulnerable to such attacks, which implant backdoor into victim models by training them on poisoned datasets and manipulate their behavior through inputs stamped with predefined triggers. Existing backdoor attacks for DMs usually employ unconcealed triggers such as a gray box, an image of Hello Kitty, and eyeglasses in the forward and reversed (i.e., sampling) diffusion process, as shown in Fig. 1. While such trigger designs achieve superior attack performance and preserve the decent generation ability of victim models, they are easy to detect via human inspection and advanced defensive algorithms as they have obvious patterns that can be precisely remodeled via reverse engineering [26, 27]. Motivated by this observation, we aim to design a general, powerful yet stealthy trigger. However, even for the traditional backdoor attacks against discriminative models (e.g., image classifiers) [28-30], it is challenging to acquire such a trigger since it either requires complex and carefully designed generators to generate image-specific or model-specific triggers [31-38] or suffers from a relatively low attack success rate (ASR) [39, 40], or fails to evade modern defenses [41-46]. This raises an interesting question in backdoor attacks against DMs: Is there a trigger simultaneously possessing generality, effectiveness, and stealthiness for backdoor attacks against DMs?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Driving by the question, in this work, we propose UIB-Diffusion, the universal imperceptible backdoor triggers for DMs that bring the following advantages: 1) Universality: the trigger is image-agnostic and model-agnostic so that it can be applied to arbitrary images and DMs; 2) Utility: the trigger can achieve high ASR while maintaining the generation capability of DMs; 3) Undetectability: the trigger is plausible to humans and capable of circumventing" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19186" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 529, + 394 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 529, + 394 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 529, + 394 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 529, + 394 + ], + "type": "image", + "image_path": "489eae18e67f17ebad7b0366a582d838a4a71684ff6d96bc63b80848149dd359.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 402, + 555, + 478 + ], + "lines": [ + { + "bbox": [ + 54, + 402, + 555, + 478 + ], + "spans": [ + { + "bbox": [ + 54, + 402, + 555, + 478 + ], + "type": "text", + "content": "Figure 1. Illustrations of the forward diffusion process (top block with the yellow background) and backward diffusion process (bottom block with the blue background) of a clean diffusion model (blue dash line), VillanDiffusion [25] (red dash line) and UIBDiffusion (Ours, green dash line). The UIBDiffusion trigger " + }, + { + "bbox": [ + 54, + 402, + 555, + 478 + ], + "type": "inline_equation", + "content": "(\\tau)" + }, + { + "bbox": [ + 54, + 402, + 555, + 478 + ], + "type": "text", + "content": " is plausible to humans in all phases from data poisoning to forward diffusion and backward diffusion while the glasses trigger " + }, + { + "bbox": [ + 54, + 402, + 555, + 478 + ], + "type": "inline_equation", + "content": "(\\mathbf{g})" + }, + { + "bbox": [ + 54, + 402, + 555, + 478 + ], + "type": "text", + "content": " in prior works is perceptible. UIBDiffusion trigger is highly effective since it introduces a similar distribution shift as the glasses trigger, which secures the attack performance during the backward diffusion process. However, it is hard to detect as the trigger does not possess specific patterns and cannot be inverted by existing defensive algorithms. We empirically verify the effectiveness in Section 4." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 492, + 297, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 297, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 297, + 659 + ], + "type": "text", + "content": "state-of-the-art defenses. Inspired by the adversarial perturbation [47] originally designed for attacking pre-trained image classifiers, we reveal that universal adversarial perturbations [48] (UAPs), a variant of adversarial perturbations, naturally hold the aforementioned properties and can be adapted as the UIBDiffusion triggers. First, UAPs are image-agnostic and model-agnostic. Second, we demonstrate that they are effective triggers because they introduce a similar distribution shift as prior works, yet they are stealthy because they do not possess clear patterns and are hard to estimate by existing trigger inversion algorithms. Moreover, they are trivial and imperceptible noises that humans can barely identify, as shown in Fig. 1. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "content": "- To the best of our knowledge, UIBDiffusion is the first imperceptible backdoor attack against diffusion models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "- We discover that adversarial perturbations designed for discriminative models can be adapted as backdoor trig-" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 492, + 553, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 492, + 553, + 515 + ], + "spans": [ + { + "bbox": [ + 322, + 492, + 553, + 515 + ], + "type": "text", + "content": "gers for generative models. We reveal why such triggers can achieve high ASR and evade SOTA defenses." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 516, + 556, + 706 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 516, + 555, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 516, + 555, + 563 + ], + "spans": [ + { + "bbox": [ + 314, + 516, + 555, + 563 + ], + "type": "text", + "content": "- We propose a practical trigger generation approach adapted from UAP, which is specifically designed to enhance universality, attack effectiveness and accessibility of triggers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 564, + 555, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 564, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 314, + 564, + 555, + 635 + ], + "type": "text", + "content": "- We experimentally demonstrate that UIBDiffusion is highly effective towards multiple diffusion models with different samplers and achieves comparable generation quality on benign input compared to clean DMs and superior attack success rate and input stealthiness compared to prior works." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 635, + 556, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 556, + 706 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 556, + 706 + ], + "type": "text", + "content": "- While existing works face a trade-off between utility and undetectability, we empirically show that UIBDiffusion triggers are highly resistant to the SOTA trigger inversion-based defenses, effectively bypassing Elijah [26] and TERD [27], the state-of-the-art defenses against backdoors on DMs." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19187" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 147, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 147, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 147, + 83 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 159, + 103 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 159, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 159, + 103 + ], + "type": "text", + "content": "2.1. Diffusion Models" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 110, + 297, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 110, + 297, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 110, + 297, + 327 + ], + "type": "text", + "content": "Diffusion models are latent variable models composed of two processes. The forward diffusion process imposes Gaussian noise on input data in each time step along the Markov chain and progressively diffuses the data distribution to an isotropic Gaussian distribution. The backward diffusion process is a reversed Markov chain that gradually denoises the sampled noise from the Gaussian distribution to the input data distribution. Mainstream diffusion models include DDPM [1], LDM [3] and NCSN [49-51], etc. Despite the excellent performance, DMs suffer from slow sampling during the backward process. There exists a line of works that leverage different techniques such as generalized non-Markovian chain and Ordinary Differential Equation (ODE) to improve the samplers and accelerate the generation process: DDIM [2], DEIS [52], UniPC [53], Heun [54], DPM-Solver [55] and DPM-Solver++ [7]. We comprehensively evaluate UIBDiffusion on these models and samplers to showcase the utility of our trigger." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 335, + 296, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 296, + 360 + ], + "type": "text", + "content": "2.2. Backdoor Attacks and Defenses on Diffusion Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 366, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 296, + 677 + ], + "type": "text", + "content": "Attacks. Backdoor attacks [28-30, 56], originally explored in the context of discriminative models such as image classifiers, induce models to make wrong predictions on triggered images. Although having been extensively studied for discriminative models, backdoor attacks are under-explored in diffusion models. BadDiffusion [23], TrojDiff [24] and VillanDiffusion [25] are pioneering works in this field. Unlike traditional backdoor attacks, those targeting DMs aim to force models to synthesize a targeted image upon malicious inputs (i.e., noises with triggers) while generating normal images upon benign inputs (i.e., isotropic Gaussian noises). BadDiffusion shows how to backdoor DMs with unconcealed triggers such as a gray box and a stop sign. TrojDiff proposes to use a whole image (e.g., Hello Kitty) as the trigger and demonstrates that backdoor can be generalized to DDIM and different adversarial goals. However, both attacks are limited to certain DMs (e.g., DDPM) and samplers (e.g., ancestral and DDIM). VillanDiffusion resolves the limitations and offers a unified framework for all kinds of DMs and samplers. Other works show backdoor can be injected to text-to-image DMs and activated by text prompts [57-59]. However, these attacks draw less attention from defenders as they can eliminate the backdoor by cleansing the text encoder. Thus, our attack focuses on the primary defensive objective of the defender: pixel level (i.e., image) trigger generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Defenses. Backdoor attacks against DMs are shown to be immune to defenses designed for traditional image classifier backdoors [23, 25, 44, 60, 61]. The reason is that DMs" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": "take sampled noises as inputs while image classifiers take natural images as inputs [27]. Elijah [26] is the first defensive algorithm specifically proposed against DM backdoors. It presents a trigger inversion scheme that can approximate the triggers and detect and remove the injected backdoor accordingly. TERD [27] employs a similar trigger estimation strategy and improves the trigger inversion process via a unified loss and trigger refinement. Based on the reconstructed trigger, it subsequently proposes to measure the KL divergence between benign and reversed distribution to detect the backdoor injected in models. Elijah and TERD achieve remarkable detection performance on existing DM backdoors. Thus, we evaluate UIBDiffusion against them and show that UIBDiffusion can bypass both defenses." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 248, + 456, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 456, + 260 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 456, + 260 + ], + "type": "text", + "content": "2.3. Adversarial Perturbation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 266, + 556, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 556, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 556, + 398 + ], + "type": "text", + "content": "The concept of adversarial perturbation is proposed in [47], demonstrating that deep neural networks are vulnerable to small additive noise. These perturbations are typically subtle and imperceptible, hence are widely used to craft adversarial examples [62-64] to fool image classifiers. Later, the work in [48] devised universal adversarial perturbation (UAP), a model-agnostic and image-agnostic variant of adversarial perturbation. We find that UAP is a particularly competitive triggers for backdoor attacks against DMs due to their imperceptibility, generalization property and accessibility." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 409, + 399, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 399, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 399, + 422 + ], + "type": "text", + "content": "3. UIBDiffusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 430, + 402, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 402, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 402, + 441 + ], + "type": "text", + "content": "3.1. Threat Model" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 448, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 556, + 616 + ], + "type": "text", + "content": "We adopt the same threat model and attack scenario presented in the existing backdoor attacks on DMs [23-25], where adversaries publish the backdoored models on websites such as HuggingFace, and users can acquire the pretrained models via third-party sources. Adversaries are expected to be able to inject the backdoor and verify the attack and generation performance before releasing the models. Users are supposed to get access to the backdoor models and the subset of clean training data and test the model utility on their side. Note that in practice, UIBDiffusion applies to a more stringent threat model where users can access the entire training dataset (i.e., contaminated dataset with poisoned samples) and apply any defenses before model deployment as the trigger is imperceptible and undetectable." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 624, + 494, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 494, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 494, + 636 + ], + "type": "text", + "content": "3.2. UIBDiffusion Trigger Generation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "Recall that our goal is to design a general, effective, and stealthy trigger (denoted as " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": ") for backdoor attacks against DMs. Due to the imperceptibility, generalization, and effectiveness properties of UAPs, they can be adapted as the UIBDiffusion trigger. We first briefly introduce the mechanism and generation process of UAPs. For a given image" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19188" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": "classifier " + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": " and a data distribution " + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": ", UAP algorithms aim to find a perturbation vector " + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(x + v) \\neq \\mathcal{C}(x)" + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": " for most " + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "inline_equation", + "content": "x \\sim q" + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": " (i.e., image agnostic). The original UAP algorithm [48] achieves this goal by measuring the minimal perturbation that pushes the perturbed input to the decision boundary, which can be mathematically expressed as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 156, + 296, + 177 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 156, + 296, + 177 + ], + "spans": [ + { + "bbox": [ + 61, + 156, + 296, + 177 + ], + "type": "interline_equation", + "content": "\\Delta v _ {i} = \\underset {d} {\\arg \\min } | | d | | _ {2} \\quad \\mathbf {s}. \\mathbf {t}. \\quad \\mathcal {C} (x _ {i} + v + d) \\neq \\mathcal {C} (x _ {i}), (1)", + "image_path": "eef7782f286023b2498613e153feccd69471fceedf1631adfef2e954e8e0caaf.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 186, + 296, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 186, + 296, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 296, + 269 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 186, + 296, + 269 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 186, + 296, + 269 + ], + "type": "text", + "content": " is the image index of a subset of the dataset selected for the computation. The final perturbation is updated as " + }, + { + "bbox": [ + 55, + 186, + 296, + 269 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\xi}(v + \\Delta v_i)" + }, + { + "bbox": [ + 55, + 186, + 296, + 269 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 186, + 296, + 269 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\xi}(\\cdot)" + }, + { + "bbox": [ + 55, + 186, + 296, + 269 + ], + "type": "text", + "content": " is a projection function and " + }, + { + "bbox": [ + 55, + 186, + 296, + 269 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 55, + 186, + 296, + 269 + ], + "type": "text", + "content": " is the perturbation budget. Note that in the context of DMs backdoor, the perturbation " + }, + { + "bbox": [ + 55, + 186, + 296, + 269 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 186, + 296, + 269 + ], + "type": "text", + "content": " is actually the imperceptible trigger " + }, + { + "bbox": [ + 55, + 186, + 296, + 269 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 186, + 296, + 269 + ], + "type": "text", + "content": ". We use " + }, + { + "bbox": [ + 55, + 186, + 296, + 269 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 186, + 296, + 269 + ], + "type": "text", + "content": " for consistent notation from now on to avoid confusion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "text", + "content": "However, the original UAP algorithm using DeepFool [65] sometimes yields less effectiveness, and we experimentally observe it renders a relatively lower ASR as the backdoor trigger for DMs. Thus, we propose a novel trigger generation approach inspired by the GUAP [66], an enhanced version of UAP that is more effective and robust. Our trigger generation process adopts the advantage of both additive (i.e., " + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "text", + "content": "-bounded) and non-additive (i.e., spatial transformation) perturbations, where the additive perturbation is adopted as the UIBDiffusion trigger " + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "text", + "content": " and coordinately optimized with the non-additive perturbation. We denote the non-additive noise as " + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "text", + "content": ". We employ a generator " + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\gamma}" + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "text", + "content": " (parameterized by " + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "text", + "content": "), which is akin to the generator in generative adversarial nets (GAN) [67] that take a latent noise " + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "text", + "content": " as input, and generate " + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "text", + "content": " and trigger " + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "text", + "content": ", as shown in lines 3, 4 and 5 in Algorithm 1. In contrast to the GAN that optimizes the generator under the guide of a discriminator, we update our trigger generator under the guide of an image classifier " + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "text", + "content": " (e.g., VGG and ResNet, etc.), which indicates whether the adversarial perturbation is effective or not, as expressed in Equation 2, where " + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\cdot)" + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "text", + "content": " is the cross entropy loss and " + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "text", + "content": " represents for the spatial transformation operation. Meanwhile, the non-additive perturbation is updated according to the loss function expressed in Equation 3, where " + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "inline_equation", + "content": "S(\\cdot)" + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "text", + "content": " is summed max pooling values of " + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "text", + "content": " using a kernel of size 4 during the spatial transformation and " + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "text", + "content": " is the number of images in the selected subset of the dataset. By incorporating such a design, we can progressively optimize the generator and hence improve the effectiveness of the generated UIBDiffusion trigger, as described in lines 5, 7 and 8 in Algorithm 1, respectively. We then optimize the trigger generation process from latent variables, with the objective of maximizing the fooling rate against the image classifier. The completed trigger generation procedures are presented in Algorithm 1 and the concrete architecture of " + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\gamma}" + }, + { + "bbox": [ + 55, + 270, + 296, + 715 + ], + "type": "text", + "content": " and flow illustration are provided in the Appendix as they are similar to GAN." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 315, + 72, + 506, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 506, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 506, + 84 + ], + "type": "text", + "content": "Algorithm 1 UIBDiffusion Trigger Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "spans": [ + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "text", + "content": "Require: Image set " + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "text", + "content": ", classifier " + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "text", + "content": ", latent noise " + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "text", + "content": ", generator " + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\gamma}" + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "text", + "content": ", training epoch " + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "text", + "content": ", number of images " + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "text", + "content": ", non-additive and additive perturbation budget " + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 315, + 87, + 555, + 123 + ], + "type": "inline_equation", + "content": "\\xi" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 318, + 124, + 553, + 271 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 321, + 124, + 416, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 124, + 416, + 134 + ], + "spans": [ + { + "bbox": [ + 321, + 124, + 416, + 134 + ], + "type": "text", + "content": "1: for " + }, + { + "bbox": [ + 321, + 124, + 416, + 134 + ], + "type": "inline_equation", + "content": "m = 1 \\dots M" + }, + { + "bbox": [ + 321, + 124, + 416, + 134 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 321, + 136, + 416, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 136, + 416, + 145 + ], + "spans": [ + { + "bbox": [ + 321, + 136, + 416, + 145 + ], + "type": "text", + "content": "2: for " + }, + { + "bbox": [ + 321, + 136, + 416, + 145 + ], + "type": "inline_equation", + "content": "i = 1 \\dots n" + }, + { + "bbox": [ + 321, + 136, + 416, + 145 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 321, + 148, + 440, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 148, + 440, + 159 + ], + "spans": [ + { + "bbox": [ + 321, + 148, + 440, + 159 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 321, + 148, + 440, + 159 + ], + "type": "inline_equation", + "content": "x_{i}\\sim X,z\\sim \\mathcal{N}(0,1)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 321, + 159, + 430, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 159, + 430, + 171 + ], + "spans": [ + { + "bbox": [ + 321, + 159, + 430, + 171 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 321, + 159, + 430, + 171 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\gamma_i}^{m - 1}(z)\\to f_i^{m - 1}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 321, + 172, + 429, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 172, + 429, + 186 + ], + "spans": [ + { + "bbox": [ + 321, + 172, + 429, + 186 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 321, + 172, + 429, + 186 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\gamma_i}^{m - 1}(z)\\to \\tau_i^{m - 1}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 321, + 186, + 501, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 186, + 501, + 205 + ], + "spans": [ + { + "bbox": [ + 321, + 186, + 501, + 205 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 321, + 186, + 501, + 205 + ], + "type": "inline_equation", + "content": "f = f_{i}^{m}\\gets \\frac{\\Gamma}{\\max_{i}\\sqrt{\\frac{1}{n}\\sum_{i}^{n}(\\mathcal{S}(x_{i}),f_{i}^{m - 1})}}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 321, + 205, + 438, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 205, + 438, + 221 + ], + "spans": [ + { + "bbox": [ + 321, + 205, + 438, + 221 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 321, + 205, + 438, + 221 + ], + "type": "inline_equation", + "content": "\\tau = \\tau_{i}^{m}\\gets \\frac{\\xi}{||\\tau_{i}^{m - 1}||_{\\infty}}" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 221, + 553, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 221, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 321, + 221, + 553, + 236 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 321, + 221, + 553, + 236 + ], + "type": "inline_equation", + "content": "\\gamma_i^m\\gets \\gamma_i^{m - 1} - \\nabla_{\\gamma_i^{m - 1}}\\mathcal{L}_{\\mathcal{G}}\\left(\\mathcal{C}(x_i\\otimes f + \\tau),\\mathcal{C}(x_i)\\right)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 236, + 375, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 236, + 375, + 245 + ], + "spans": [ + { + "bbox": [ + 321, + 236, + 375, + 245 + ], + "type": "text", + "content": "9: end for" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 318, + 247, + 365, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 247, + 365, + 257 + ], + "spans": [ + { + "bbox": [ + 318, + 247, + 365, + 257 + ], + "type": "text", + "content": "10: end for" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 318, + 259, + 402, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 259, + 402, + 271 + ], + "spans": [ + { + "bbox": [ + 318, + 259, + 402, + 271 + ], + "type": "text", + "content": "11: return Trigger " + }, + { + "bbox": [ + 318, + 259, + 402, + 271 + ], + "type": "inline_equation", + "content": "\\tau" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 315, + 285, + 553, + 358 + ], + "blocks": [ + { + "bbox": [ + 315, + 285, + 553, + 358 + ], + "lines": [ + { + "bbox": [ + 315, + 285, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 315, + 285, + 553, + 358 + ], + "type": "image", + "image_path": "533c428f255e159c4b25a11fc9a0f5305c85b43c18c4e5b2d7083c7b7d2cce04.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 365, + 555, + 421 + ], + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 421 + ], + "type": "text", + "content": "Figure 2. Illustration of trigger mechanisms. UIBDiffusion triggers introduce a similar distribution shift as triggers in prior works, which secure a high attack performance. However, since our triggers do not have a specific pattern, they are difficult to remodel by trigger inversion algorithms." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 367, + 446, + 555, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 446, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 367, + 446, + 555, + 460 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {G}} = - \\mathcal {H} (\\mathcal {C} (x \\otimes f + \\tau), \\mathcal {C} (x)), \\tag {2}", + "image_path": "d029682e936471d37fee50defa20530022b64a981e602ae040eade1e0b71ce37.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 372, + 478, + 554, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 478, + 554, + 516 + ], + "spans": [ + { + "bbox": [ + 372, + 478, + 554, + 516 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f} = \\max _ {i} \\sqrt {\\frac {1}{n} \\sum_ {i} ^ {n} (\\mathcal {S} (x _ {i} , f))} \\tag {3}", + "image_path": "7f6ae55d587fb2c85fcc130d7abe092f8d48c8aa0b5df3113369acab7470e770.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "content": "The effectiveness of UIBDiffusion against backdoor attacks on generative models, such as DMs, stems from key insights into backdoor dynamics. As revealed in [23-25], backdoor for DMs needs to introduce an evident distribution shift and keep such a shift during the reversed Markov chain to secure a high ASR. We empirically find that the UIBDiffusion triggers also lead to a similar distribution shift when imposed on benign sampled noise, as shown in Fig. 2. However, because they are adapted from UAPs, which exhibit minimal, hard-to-detect patterns with subtle perturbations, we hypothesize that these triggers are difficult for trigger inversion algorithms to estimate accurately. Consequently, unlike the triggers with specific patterns, they are managed to circumvent the SOTA defenses and remain highly stealthy. We provide comprehensive and extensive evaluation results in Section 4 to support our conjecture." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19189" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 249, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 249, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 249, + 85 + ], + "type": "text", + "content": "3.3. Inject Backdoor in Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 173 + ], + "type": "text", + "content": "Prior attacks leveraged similar backdoor injection mechanisms with slight differences in sampler modifications [23-25]. UIBDiffusion employs the scheme presented in VillanDiffusion [25] since it provides a unified framework to implant backdoors in various types of diffusion models and samplers and achieves outstanding performance. We briefly introduce the injection process in this section." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 173, + 296, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 173, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 173, + 296, + 304 + ], + "type": "text", + "content": "Clean diffusion process. Clean diffusion consists of forward and backward processes. The forward diffusion process diffuses images " + }, + { + "bbox": [ + 55, + 173, + 296, + 304 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 173, + 296, + 304 + ], + "type": "text", + "content": " from data distribution " + }, + { + "bbox": [ + 55, + 173, + 296, + 304 + ], + "type": "inline_equation", + "content": "q(x_0)" + }, + { + "bbox": [ + 55, + 173, + 296, + 304 + ], + "type": "text", + "content": " to the Gaussian distribution " + }, + { + "bbox": [ + 55, + 173, + 296, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 55, + 173, + 296, + 304 + ], + "type": "text", + "content": " by progressively adding noise in each time step along a Markov chain. The process can be mathematically expressed as: " + }, + { + "bbox": [ + 55, + 173, + 296, + 304 + ], + "type": "inline_equation", + "content": "q(x_t | x_0) = \\mathcal{N}(\\alpha(t)x_0, \\beta^2(t)\\mathbf{I})" + }, + { + "bbox": [ + 55, + 173, + 296, + 304 + ], + "type": "text", + "content": " with (1) " + }, + { + "bbox": [ + 55, + 173, + 296, + 304 + ], + "type": "inline_equation", + "content": "q(x_{T_{\\max}}) \\approx \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 55, + 173, + 296, + 304 + ], + "type": "text", + "content": ", (2) " + }, + { + "bbox": [ + 55, + 173, + 296, + 304 + ], + "type": "inline_equation", + "content": "q(x_{T_{\\min}}) \\approx q(x_0)" + }, + { + "bbox": [ + 55, + 173, + 296, + 304 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 173, + 296, + 304 + ], + "type": "inline_equation", + "content": "x_t, t \\in [T_{\\min}, T_{\\max}]" + }, + { + "bbox": [ + 55, + 173, + 296, + 304 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 173, + 296, + 304 + ], + "type": "inline_equation", + "content": "T_{\\min}, T_{\\max}" + }, + { + "bbox": [ + 55, + 173, + 296, + 304 + ], + "type": "text", + "content": " are the first and last timestep, and " + }, + { + "bbox": [ + 55, + 173, + 296, + 304 + ], + "type": "inline_equation", + "content": "\\alpha(t), \\beta(t)" + }, + { + "bbox": [ + 55, + 173, + 296, + 304 + ], + "type": "text", + "content": " are parameterized variables introduced in the original diffusion model work [1]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 304, + 296, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 296, + 377 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 296, + 377 + ], + "type": "text", + "content": "Backward diffusion process reverse the Markov chain and generate images from a sampled Gaussian noise by gradually removing noises in each timestep and can be expressed as: " + }, + { + "bbox": [ + 55, + 304, + 296, + 377 + ], + "type": "inline_equation", + "content": "q(x_{t}|x_{t - 1}) = \\mathcal{N}(k_{t}x_{t - 1},w_{t}^{2}\\mathbf{I})" + }, + { + "bbox": [ + 55, + 304, + 296, + 377 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 304, + 296, + 377 + ], + "type": "inline_equation", + "content": "x_{t}(x_{0},\\epsilon) = \\hat{\\alpha} (t)x_{0} + \\hat{\\beta} (t)\\epsilon" + }, + { + "bbox": [ + 55, + 304, + 296, + 377 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 304, + 296, + 377 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 55, + 304, + 296, + 377 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 304, + 296, + 377 + ], + "type": "inline_equation", + "content": "k_{t},w_{t},\\hat{\\alpha} (t),\\hat{\\beta} (t)" + }, + { + "bbox": [ + 55, + 304, + 296, + 377 + ], + "type": "text", + "content": " are the same parameterized variables for the reverse process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 378, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 296, + 498 + ], + "type": "text", + "content": "Backdoor and trigger injection. We employ the Villan-Diffusion framework for backdoor injection and modify the trigger injection process in data poisoning. VillanDiffusion proposes a unified backdoor injection framework by introducing the loss function shown in Equation 4, where " + }, + { + "bbox": [ + 55, + 378, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 55, + 378, + 296, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 378, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 55, + 378, + 296, + 498 + ], + "type": "text", + "content": " are used to optimize the original generation and backdoor attack tasks, respectively. " + }, + { + "bbox": [ + 55, + 378, + 296, + 498 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 378, + 296, + 498 + ], + "type": "text", + "content": " is the poisoned data, " + }, + { + "bbox": [ + 55, + 378, + 296, + 498 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 378, + 296, + 498 + ], + "type": "text", + "content": " is the backdoor target, " + }, + { + "bbox": [ + 55, + 378, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 55, + 378, + 296, + 498 + ], + "type": "text", + "content": " is a parameter to control backdoor against different samplers (e.g., ODE and SDE), and " + }, + { + "bbox": [ + 55, + 378, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\eta_c, \\eta_p" + }, + { + "bbox": [ + 55, + 378, + 296, + 498 + ], + "type": "text", + "content": " are weights to balance " + }, + { + "bbox": [ + 55, + 378, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 55, + 378, + 296, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 378, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 55, + 378, + 296, + 498 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 101, + 506, + 295, + 524 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 506, + 295, + 524 + ], + "spans": [ + { + "bbox": [ + 101, + 506, + 295, + 524 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\theta} \\left(\\eta_ {c}, \\eta_ {p}, x, t, \\epsilon , r, y, \\xi\\right) = \\tag {4}", + "image_path": "b0dd953622305533d6045cbc86937b19ed42c3961f93dba330b97599fe94e502.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 102, + 522, + 248, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 522, + 248, + 534 + ], + "spans": [ + { + "bbox": [ + 102, + 522, + 248, + 534 + ], + "type": "interline_equation", + "content": "\\eta_ {c} \\mathcal {L} _ {c} (x, t, \\epsilon) + \\eta_ {p} \\mathcal {L} _ {p} (x, t, \\epsilon , r, y, \\zeta).", + "image_path": "627efaa916fdf15fa3382ba4bae4470dc1e68f44fe960ae4bf185f2353fa4601.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 538, + 296, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 296, + 610 + ], + "type": "text", + "content": "We then explain the difference in the trigger injection process between VillanDiffusion and UIBDiffusion. Villan-Diffusion injects the backdoor trigger following the method in Equation 5, where " + }, + { + "bbox": [ + 55, + 538, + 296, + 610 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 55, + 538, + 296, + 610 + ], + "type": "text", + "content": " is a mask and " + }, + { + "bbox": [ + 55, + 538, + 296, + 610 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 538, + 296, + 610 + ], + "type": "text", + "content": " is the trigger with specific patterns (e.g., gray box and a pair of glasses), and " + }, + { + "bbox": [ + 55, + 538, + 296, + 610 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 538, + 296, + 610 + ], + "type": "text", + "content": " is the element-wise multiplication." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 620, + 295, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 620, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 295, + 633 + ], + "type": "interline_equation", + "content": "r (x, g) = \\mathbf {M} \\odot g + (1 - \\mathbf {M}) \\odot x. \\tag {5}", + "image_path": "ac9d6a48ed31ae363aa83746b103817f8b56841718f027eaf7d66f48b88bedde.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "To better align with the adversarial perturbation, we inject our trigger using the approach in Equation 6, where " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " is the imperceptible trigger synthesized by our proposed trigger generator and " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " is a variable that controls the strength of the added trigger. In contrast to triggers with recognizable patterns, the UIBDiffusion triggers are generated by" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "a trainable generator with subtle perturbations and hard-to-detect patterns, which makes them effective and stealthy. The backdoor and trigger injection process is summarized in Algorithm 2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 392, + 134, + 555, + 148 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 134, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 392, + 134, + 555, + 148 + ], + "type": "interline_equation", + "content": "r (x, \\tau) = x + \\varepsilon \\odot \\tau \\tag {6}", + "image_path": "6bc0f665c486a8b9f9c08902e5089c9f4debd2a9d8864e3f97d42d92cfb67bac.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 167, + 554, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 167, + 554, + 190 + ], + "spans": [ + { + "bbox": [ + 315, + 167, + 554, + 190 + ], + "type": "text", + "content": "Algorithm 2 UIBDiffusion Backdoor And Trigger Injection" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 193, + 554, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 193, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 315, + 193, + 554, + 228 + ], + "type": "text", + "content": "Require: Backdoor Image Trigger " + }, + { + "bbox": [ + 315, + 193, + 554, + 228 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 315, + 193, + 554, + 228 + ], + "type": "text", + "content": ", Benign Dataset " + }, + { + "bbox": [ + 315, + 193, + 554, + 228 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 315, + 193, + 554, + 228 + ], + "type": "text", + "content": ", Backdoor Target " + }, + { + "bbox": [ + 315, + 193, + 554, + 228 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 315, + 193, + 554, + 228 + ], + "type": "text", + "content": ", Training parameters " + }, + { + "bbox": [ + 315, + 193, + 554, + 228 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 315, + 193, + 554, + 228 + ], + "type": "text", + "content": ", Sampler Randomness " + }, + { + "bbox": [ + 315, + 193, + 554, + 228 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 315, + 193, + 554, + 228 + ], + "type": "text", + "content": ", Strength " + }, + { + "bbox": [ + 315, + 193, + 554, + 228 + ], + "type": "inline_equation", + "content": "\\varepsilon" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 229, + 504, + 323 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 321, + 229, + 504, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 229, + 504, + 240 + ], + "spans": [ + { + "bbox": [ + 321, + 229, + 504, + 240 + ], + "type": "text", + "content": "1: Backdoor DMs with Generated Trigger:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 241, + 424, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 241, + 424, + 251 + ], + "spans": [ + { + "bbox": [ + 321, + 241, + 424, + 251 + ], + "type": "text", + "content": "2: while not converge do" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 252, + 408, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 252, + 408, + 265 + ], + "spans": [ + { + "bbox": [ + 321, + 252, + 408, + 265 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 321, + 252, + 408, + 265 + ], + "type": "inline_equation", + "content": "\\{x,\\eta_c,\\eta_p\\} \\sim X" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 265, + 447, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 265, + 447, + 277 + ], + "spans": [ + { + "bbox": [ + 321, + 265, + 447, + 277 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 321, + 265, + 447, + 277 + ], + "type": "inline_equation", + "content": "t\\sim \\mathrm{Uniform}(\\{1,\\ldots ,T\\})" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 277, + 392, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 277, + 392, + 289 + ], + "spans": [ + { + "bbox": [ + 321, + 277, + 392, + 289 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 321, + 277, + 392, + 289 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 289, + 498, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 289, + 498, + 300 + ], + "spans": [ + { + "bbox": [ + 321, + 289, + 498, + 300 + ], + "type": "text", + "content": "6: Trigger injection: " + }, + { + "bbox": [ + 321, + 289, + 498, + 300 + ], + "type": "inline_equation", + "content": "r(x,\\tau) = x + \\varepsilon \\odot \\tau" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 300, + 495, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 300, + 495, + 313 + ], + "spans": [ + { + "bbox": [ + 321, + 300, + 495, + 313 + ], + "type": "text", + "content": "7: update " + }, + { + "bbox": [ + 321, + 300, + 495, + 313 + ], + "type": "inline_equation", + "content": "\\theta :\\nabla_{\\theta}\\mathcal{L}_{\\theta}^{I}(\\eta_{c},\\eta_{p},x,t,\\epsilon ,r,y,\\zeta)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 313, + 375, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 313, + 375, + 323 + ], + "spans": [ + { + "bbox": [ + 321, + 313, + 375, + 323 + ], + "type": "text", + "content": "8: end while" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 351, + 395, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 351, + 395, + 365 + ], + "spans": [ + { + "bbox": [ + 314, + 351, + 395, + 365 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 371, + 441, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 441, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 441, + 384 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 389, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 533 + ], + "type": "text", + "content": "Datasets and models. We follow the same practice in existing backdoor attacks against DMs [23-25] and use CIFAR-10 " + }, + { + "bbox": [ + 313, + 389, + 555, + 533 + ], + "type": "inline_equation", + "content": "(32 \\times 32)" + }, + { + "bbox": [ + 313, + 389, + 555, + 533 + ], + "type": "text", + "content": " [68] and CelebA-HQ " + }, + { + "bbox": [ + 313, + 389, + 555, + 533 + ], + "type": "inline_equation", + "content": "(256 \\times 256)" + }, + { + "bbox": [ + 313, + 389, + 555, + 533 + ], + "type": "text", + "content": " [69] datasets for fair comparison. We evaluate UIB-Diffusion on three DMs: DDPM, LDM and NCSN, with eleven different samplers. We adopt the same training recipe from [25] and conduct our attack on pre-trained models (i.e., google/ddpm-cifar10-32 for CIFAR-10 and CompVis/ldm-celebahq-256 for CelebA-HQ) downloaded from Google HuggingFace organization and HuggingFace, respectively. Training hyper-parameters are provided in the Appendix." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 533, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 554, + 665 + ], + "type": "text", + "content": "Attack configurations and baselines. We conduct our experiments on NVIDIA L40S GPUs and employ Villan-Diffusion [25] as the backdoor injection framework. We provide two generated triggers guided by VGG [70] and ResNet [71] for CIFAR-10 and one proof-of-concept trigger for CelebA-HQ. We adopt \"hat\" and \"cat\" as targets for CIFAR-10 and \"cat\" as the target for CelebA-HQ, respectively. BadDiffusion [23] and VillanDiffusion [25] are considered as the baseline attacks for utility comparison and Elijah [26] and TERD [27] as the defenses for undetectability evaluation." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "Evaluation metrics. We comprehensively evaluate the attack and generation performance of UIBDiffusion using four standard metrics. We exploit FID [72] score to examine the quality of generated clean images. Images with" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19190" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 216 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 216 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 216 + ], + "type": "image", + "image_path": "46501edc54f55a090cc86554d8b2d984367ddef3be71c2002cd226e698a3e6a7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 223, + 515, + 236 + ], + "lines": [ + { + "bbox": [ + 93, + 223, + 515, + 236 + ], + "spans": [ + { + "bbox": [ + 93, + 223, + 515, + 236 + ], + "type": "text", + "content": "Figure 3. FID and MSE comparison of BadDiffusion, VillanDiffusion, and UIBDiffusion over different poison rates." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 245, + 553, + 355 + ], + "blocks": [ + { + "bbox": [ + 58, + 245, + 553, + 355 + ], + "lines": [ + { + "bbox": [ + 58, + 245, + 553, + 355 + ], + "spans": [ + { + "bbox": [ + 58, + 245, + 553, + 355 + ], + "type": "image", + "image_path": "df0779aecaa78320b9628415772caa2d151413100c8b85d2a94a91f27b8dc861.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 361, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 55, + 361, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 555, + 384 + ], + "type": "text", + "content": "Table 1. ASR comparison between VillanDiffusion (top row) and UIBDiffusion (bottom row). UIBDiffusion achieves superior performance than the prior work at a low poison rate (e.g., " + }, + { + "bbox": [ + 55, + 361, + 555, + 384 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 361, + 555, + 384 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 396, + 297, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 396, + 297, + 551 + ], + "spans": [ + { + "bbox": [ + 54, + 396, + 297, + 551 + ], + "type": "text", + "content": "better quality tend to have lower FID scores. As for the attack effect, we employ Attack Success Rate (ASR), Mean Square Error (MSE) and Structural Similarity Index Measure (SSIM) for assessment. ASR is defined as the percentage of the target image generated from sampled noise stamped with backdoor triggers. Higher ASR indicates better utility. MSE is a metric to measure the pixel-level distance between the authentic and synthetic target images. A successful attack is expected to achieve low MSE. SSIM is a similar metric used to measure the similarity between two images (i.e., the real and generated target images). Superior attack performance will yield SSIM that approaches the value of 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 563, + 255, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 255, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 255, + 574 + ], + "type": "text", + "content": "4.2. Results on Different Diffusion Models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "We first evaluate UIBDiffusion on different DMs. We present the results of DDPM in the main paper and organize the results of LDM and NCSN in the Appendix. As shown in Fig. 3, Table 1 and Table 2, UIBDiffusion achieves comparable FID as the clean diffusion model (i.e., poison rate " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "= 0\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "). It can be seen that with the poison rates increased, all backdoor attacks yield a lower MSE, demonstrating a positive correlation between poison rates and attack effect. However, the generation ability of DMs will be accordingly affected when the poison rate is high. Notably, UIBDiffusion can reach a " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " ASR with only " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " poison rate" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 396, + 555, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 443 + ], + "type": "text", + "content": "while VillanDiffusion gets " + }, + { + "bbox": [ + 313, + 396, + 555, + 443 + ], + "type": "inline_equation", + "content": "56\\%" + }, + { + "bbox": [ + 313, + 396, + 555, + 443 + ], + "type": "text", + "content": " ASR at this rate. In practice, attackers can achieve the expected attack effect without undermining the original generation ability of DMs, which further improves the stealthiness of the backdoor attack." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 446, + 555, + 487 + ], + "blocks": [ + { + "bbox": [ + 316, + 446, + 555, + 487 + ], + "lines": [ + { + "bbox": [ + 316, + 446, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 446, + 555, + 487 + ], + "type": "table", + "html": "
Poison Rate0%1%2%3%4%5%10%≥ 20%
BadDiffusion [23]3.80E-46.22E-31.48E-21.10E-20.370.990.990.99
VillanDiffusion [25]3.80E-43.92E-41.25E-21.69E-20.600.990.990.99
UIBDiffusion (Ours)3.80E-40.470.960.990.990.990.990.99
", + "image_path": "7d15ea2b50eab05a764f32458652ec8f69427a85991b4669b90e10bcb1c235e1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 494, + 555, + 526 + ], + "lines": [ + { + "bbox": [ + 313, + 494, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 555, + 526 + ], + "type": "text", + "content": "Table 2. SSIM comparison of BadDiffusion, VillanDiffusion and UIBDiffusion. UIBDiffusion achieves higher SSIM at lower poison rates of " + }, + { + "bbox": [ + 313, + 494, + 555, + 526 + ], + "type": "inline_equation", + "content": "1\\% \\sim 4\\%" + }, + { + "bbox": [ + 313, + 494, + 555, + 526 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 529, + 479, + 543 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 479, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 479, + 543 + ], + "type": "text", + "content": "4.3. Results on Different Samplers" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "type": "text", + "content": "We then show the effectiveness of UIBDiffusion on different samplers. We conduct the experiments using DDPM model with 11 types of samplers. As shown in Fig. 4, UIB-Diffusion achieves a consistent performance of " + }, + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 313, + 547, + 555, + 643 + ], + "type": "text", + "content": " ASR across all samplers and yield a robust target image generation effect (i.e., MSE and SSIM) at various poison rates. We also observe that the ODE solvers outperform their SDE solver counterparts." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 648, + 495, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 495, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 495, + 661 + ], + "type": "text", + "content": "4.4. Resilience against SOTA Defenses" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Elijah [26] and TERD [27] are two SOTA defenses against backdoor attacks on DMs. They employ a similar idea of trigger inversion to estimate and reverse engineer the triggers. They then apply backdoor detection and removal" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19191" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 72, + 176, + 173 + ], + "blocks": [ + { + "bbox": [ + 60, + 72, + 176, + 173 + ], + "lines": [ + { + "bbox": [ + 60, + 72, + 176, + 173 + ], + "spans": [ + { + "bbox": [ + 60, + 72, + 176, + 173 + ], + "type": "image", + "image_path": "df442f1a67b91a3928b46354a75ad9a1c06e4ea1d12159dc574bb2464df0da50.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 184, + 555, + 207 + ], + "lines": [ + { + "bbox": [ + 55, + 184, + 555, + 207 + ], + "spans": [ + { + "bbox": [ + 55, + 184, + 555, + 207 + ], + "type": "text", + "content": "Figure 4. UIBDiffusion performance against eleven different samplers across various poison rates. UIBDiffusion yields consistently high ASR, SSIM, and MSE and even better FID on ODE-based samplers compared to SDE-based samplers." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 186, + 72, + 301, + 172 + ], + "blocks": [ + { + "bbox": [ + 186, + 72, + 301, + 172 + ], + "lines": [ + { + "bbox": [ + 186, + 72, + 301, + 172 + ], + "spans": [ + { + "bbox": [ + 186, + 72, + 301, + 172 + ], + "type": "image", + "image_path": "281a17324825adcdc6951bb4c4cdcbca827950368da8c059cddc41de22291528.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 426, + 173 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 426, + 173 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 426, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 426, + 173 + ], + "type": "image", + "image_path": "5df83fc19ec4015c3af41dc08e630cfa33169215d59626b0f29e151746456e4e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 434, + 72, + 552, + 172 + ], + "blocks": [ + { + "bbox": [ + 434, + 72, + 552, + 172 + ], + "lines": [ + { + "bbox": [ + 434, + 72, + 552, + 172 + ], + "spans": [ + { + "bbox": [ + 434, + 72, + 552, + 172 + ], + "type": "image", + "image_path": "d00e50ed3acef7d9fdadfdf20b13d25e74f38f7617c5db20d66e8787b35df162.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 218, + 297, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 218, + 297, + 456 + ], + "spans": [ + { + "bbox": [ + 54, + 218, + 297, + 456 + ], + "type": "text", + "content": "based on the reconstructed triggers. We show that UIBDiffusion is highly resilient to both attacks since our triggers are hard to remodel. Fig. 5 demonstrates the performance of UIBDiffusion against Elijah compared to BadDifuFSion and VillanDiffusion. It can be seen that UIBDiffusion achieves " + }, + { + "bbox": [ + 54, + 218, + 297, + 456 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 54, + 218, + 297, + 456 + ], + "type": "text", + "content": " ASR and keeps the same MSE and SSIM before and after the backdoor detection and removal while ASR and SSIM of BadDiffusion and VillanDiffusion drop to " + }, + { + "bbox": [ + 54, + 218, + 297, + 456 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 54, + 218, + 297, + 456 + ], + "type": "text", + "content": " after detection, indicating superior resilience against the SOTA defense. We then present the robustness of UIBDiffusion across various poison rates. Higher poison rates usually yield higher ASR but are easier to detect. However, as shown in Table 3, UIBDiffusion managed to escape detection at all the evaluated poison rates while maintaining " + }, + { + "bbox": [ + 54, + 218, + 297, + 456 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 54, + 218, + 297, + 456 + ], + "type": "text", + "content": " ASR and high performance of SSIM and FID. It is worth noting that a " + }, + { + "bbox": [ + 54, + 218, + 297, + 456 + ], + "type": "inline_equation", + "content": "\\Delta SSIM = 0" + }, + { + "bbox": [ + 54, + 218, + 297, + 456 + ], + "type": "text", + "content": " indicates that UIBDiffusion successfully bypasses the defense, achieving " + }, + { + "bbox": [ + 54, + 218, + 297, + 456 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 54, + 218, + 297, + 456 + ], + "type": "text", + "content": " targeted image generation. We attribute the fluctuations in FID to Elijah's use of an ineffective inverted trigger during backdoor removal." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 58, + 459, + 293, + 547 + ], + "blocks": [ + { + "bbox": [ + 58, + 459, + 293, + 547 + ], + "lines": [ + { + "bbox": [ + 58, + 459, + 293, + 547 + ], + "spans": [ + { + "bbox": [ + 58, + 459, + 293, + 547 + ], + "type": "table", + "html": "
BadDiffusion * [23]UIBDiffusion (Ours)
Poison RateDetectedASRΔSSIMΔFIDDetectedASRΔSSIMΔFID
5%100%0%-1.0-0.070%100%0-0.98
10%100%0%-1.0+0.180%100%0+0.15
20%100%0%-1.0-0.030%100%0+5.72
30%100%0%-1.0+0.150%100%0+12.59
50%100%0%-1.0+0.200%100%0+11.56
70%100%0%-1.0-0.110%100%0+14.85
90%100%0%-1.0-0.070%100%0+5.09
", + "image_path": "096c157e2328ae1ee9f58e07d6a34f640f18e0323911c5a21a38fd53b51a623f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "We evaluate the quality of the inverted trigger by visualizing the inverted triggers and testing if they could activate the backdoor as the actual triggers when imposed on benign sampled noise. Results are shown in Fig. 6. Sampled noise with inverted triggers should generate the target image if the defense succeeds and generate normal images otherwise. It can be seen that for triggers that have specific patterns (e.g., a stop sign and a gray box), Elijah can successfully estimate the authentic triggers and the inverted triggers are able to activate backdoors. On the other hand, the inverted trigger of UIBDiffusion fails to activate the backdoor injected in the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 218, + 555, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 555, + 277 + ], + "type": "text", + "content": "DM, showing that the defense is ineffective against UIBDiffusion. We argue this is because the UIBDiffusion trigger is adapted from UAP, which is small and unstructured noise that is difficult to approximate based on the current trigger inversion loss function in Elijah and TERD." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 278, + 556, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 556, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 556, + 399 + ], + "type": "text", + "content": "UIBDiffusion performs a similar undetectability against TERD as shown in Table 4. It achieves a " + }, + { + "bbox": [ + 313, + 278, + 556, + 399 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 313, + 278, + 556, + 399 + ], + "type": "text", + "content": " True Positive Rate (TPR) and " + }, + { + "bbox": [ + 313, + 278, + 556, + 399 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 313, + 278, + 556, + 399 + ], + "type": "text", + "content": " False Positive Rate (FPR), where TPR and FPR represent the percentage of the benign or backdoor sampled noise that is correctly detected. On the other hand, triggers with obvious patterns fail to bypass the trigger inversion algorithm and are all detected by TERD. Although TRED offers a more precise trigger inversion approach using a unified loss and trigger refinement, it is still ineffective against our proposed UIBDiffusion trigger." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 317, + 403, + 552, + 487 + ], + "blocks": [ + { + "bbox": [ + 55, + 555, + 295, + 578 + ], + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 578 + ], + "type": "text", + "content": "Table 3. Elijah against BadDiffusion and UIBDiffusion with different poison rates. “*” denotes results replicated from [26]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 403, + 552, + 487 + ], + "lines": [ + { + "bbox": [ + 317, + 403, + 552, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 403, + 552, + 487 + ], + "type": "table", + "html": "
Trigger-TargetTPR (↑)FPR (↓)
Ours-HAT100%0%
BOX-HAT*100%100%
BOX-SHOE*100%100%
BOX-CORNER*100%100%
STOP-HAT*100%100%
STOP-SHOE*100%100%
STOP-CORNER*100%100%
", + "image_path": "4dff114813bd2cb0f6e14502d540766713e47215fe4c33bf7eaf2d85697f44f8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 494, + 554, + 517 + ], + "lines": [ + { + "bbox": [ + 313, + 494, + 554, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 554, + 517 + ], + "type": "text", + "content": "Table 4. Resilience of UIBDiffusion against TERD. “*” denotes results replicated from [27]." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 528, + 549, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 549, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 549, + 541 + ], + "type": "text", + "content": "4.5. Comparison of Trigger Generation Approach" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "We evaluate the necessity and effectiveness of our proposed trigger generation approach by comparing it to the original UAPs and GUAPs. We employ the generated UIBDiffusion trigger, the original UAP from [48] and the GUAP from [66] to poison the benign dataset and backdoor the DMs using these three triggers, respectively. As can be seen from Table 5, the UIBDiffusion trigger outperforms the original UAP and GUAP triggers at all four metrics by large margins, achieving much better attack and generation performance. The rationale behind this is that our trigger introduces an evident distribution shift. In contrast, the original UAP trigger only slightly affects the sampled noise distribution, and the spatial transformation of GUAP is image-specific and imposes a negative impact on the backdoor im" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19192" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 78, + 180, + 187 + ], + "blocks": [ + { + "bbox": [ + 59, + 78, + 180, + 187 + ], + "lines": [ + { + "bbox": [ + 59, + 78, + 180, + 187 + ], + "spans": [ + { + "bbox": [ + 59, + 78, + 180, + 187 + ], + "type": "image", + "image_path": "ec68786bb9c549cfe46620a356bf3bb5c13a8a48ed2ca3ae36ba1c131a24cddf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 204, + 555, + 227 + ], + "lines": [ + { + "bbox": [ + 55, + 204, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 555, + 227 + ], + "type": "text", + "content": "Figure 5. Performance comparison of BadDiffusion, VillanDiffusion and our UIBDiffusion before and after the Elijah defense. ASR: the higher the better; FID: the lower the better; MSE: the lower the better; SSIM: the higher the better." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 184, + 78, + 302, + 187 + ], + "blocks": [ + { + "bbox": [ + 184, + 78, + 302, + 187 + ], + "lines": [ + { + "bbox": [ + 184, + 78, + 302, + 187 + ], + "spans": [ + { + "bbox": [ + 184, + 78, + 302, + 187 + ], + "type": "image", + "image_path": "c5dd4453540f91a2e28b13f96e930829895ef883715b870c9dd5847c6c6b972b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 303, + 78, + 427, + 187 + ], + "blocks": [ + { + "bbox": [ + 303, + 78, + 427, + 187 + ], + "lines": [ + { + "bbox": [ + 303, + 78, + 427, + 187 + ], + "spans": [ + { + "bbox": [ + 303, + 78, + 427, + 187 + ], + "type": "image", + "image_path": "b593e16ad5586ee094b3c9db4eb849a18edf39bb0d92eca2cab7efffb02fbb65.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 430, + 78, + 550, + 187 + ], + "blocks": [ + { + "bbox": [ + 430, + 78, + 550, + 187 + ], + "lines": [ + { + "bbox": [ + 430, + 78, + 550, + 187 + ], + "spans": [ + { + "bbox": [ + 430, + 78, + 550, + 187 + ], + "type": "image", + "image_path": "50fb0826eea200ac7edd7ff0ff4b5353e69a0a1416dae3ab5fcd90787e2bdaee.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 59, + 238, + 294, + 398 + ], + "blocks": [ + { + "bbox": [ + 59, + 238, + 294, + 398 + ], + "lines": [ + { + "bbox": [ + 59, + 238, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 59, + 238, + 294, + 398 + ], + "type": "image", + "image_path": "1e03395fafd941eeb48ee2340d21b80e3352e419d0f23b30ef9a84eb96827cca.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 409, + 296, + 455 + ], + "lines": [ + { + "bbox": [ + 55, + 409, + 296, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 296, + 455 + ], + "type": "text", + "content": "Figure 6. Visualization of trigger inversion. Inverted triggers of gray box and stop sign can activate backdoor, indicating a successful estimation of the actual triggers. On the other hand, inverted triggers of UIBDiffusion fail to activate the injected backdoor." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 466, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 295, + 491 + ], + "type": "text", + "content": "plant. This shows the superiority and necessity of our trigger design choice." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 58, + 500, + 293, + 544 + ], + "blocks": [ + { + "bbox": [ + 58, + 500, + 293, + 544 + ], + "lines": [ + { + "bbox": [ + 58, + 500, + 293, + 544 + ], + "spans": [ + { + "bbox": [ + 58, + 500, + 293, + 544 + ], + "type": "table", + "html": "
FID (↓)MSE (↓)SSIM (↑)ASR (↑)
Original UAP20.889.22E-20.619043.75%
Original GUAP41.877.86E-20.19320.00%
UIBDiffusion19.775.30E-40.9967100%
", + "image_path": "655fb51a76a2ac3fba1cda7cf756285686cd0289cb142fd82b756ed44921c047.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 597, + 169, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 169, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 169, + 609 + ], + "type": "text", + "content": "4.6. Trigger Robustness" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 618, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 712 + ], + "type": "text", + "content": "Subtle adversarial perturbations are susceptible to smoothing defenses [73] and could be mitigated by random noise. Thus, we perform a robust evaluation of the UIBDiffusion triggers and present the results in Table 6. UIBDiffusion shows strong robustness against random noise. We argue that the resilience is due to our trigger generation algorithm enhancing the adversarial perturbations via co-optimizing the additive and non-additive noise." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 317, + 236, + 554, + 287 + ], + "blocks": [ + { + "bbox": [ + 55, + 552, + 295, + 574 + ], + "lines": [ + { + "bbox": [ + 55, + 552, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 295, + 574 + ], + "type": "text", + "content": "Table 5. Trigger performance comparison across the original UAP, GUAP and UIBDiffusion." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 236, + 554, + 287 + ], + "lines": [ + { + "bbox": [ + 317, + 236, + 554, + 287 + ], + "spans": [ + { + "bbox": [ + 317, + 236, + 554, + 287 + ], + "type": "table", + "html": "
Random Noise1%2%5%10%20%
ASR99%99%100%100%100%
FID19.4819.5219.8019.3719.49
MSE2.36E-31.13E-38.03E-43.02E-45.30E-4
SSIM0.98650.99290.99440.99690.9957
", + "image_path": "c5253a3abe523eaa439c1677661472aaccbe2dd490042147d26e4466700703e8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 315, + 539, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 539, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 539, + 328 + ], + "type": "text", + "content": "4.7. Quantitative Evaluation of Imperceptibility" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 332, + 555, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 555, + 441 + ], + "type": "text", + "content": "We also quantitatively evaluate the imperceptibility of our trigger design using distance metrics of " + }, + { + "bbox": [ + 313, + 332, + 555, + 441 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 313, + 332, + 555, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 332, + 555, + 441 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 313, + 332, + 555, + 441 + ], + "type": "text", + "content": " norms. We compute the distance between a poisoned image (stamping with our trigger and triggers used in prior works) and a benign image and summarize the results in Table 7. Our imperceptible triggers have much lower " + }, + { + "bbox": [ + 313, + 332, + 555, + 441 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 313, + 332, + 555, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 332, + 555, + 441 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 313, + 332, + 555, + 441 + ], + "type": "text", + "content": " distances compared to triggers with specific patterns, which validates the stealthiness at the image level (i.e., the pixel space) of the UIBDiffusion trigger design." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 317, + 445, + 554, + 495 + ], + "blocks": [ + { + "bbox": [ + 325, + 295, + 542, + 306 + ], + "lines": [ + { + "bbox": [ + 325, + 295, + 542, + 306 + ], + "spans": [ + { + "bbox": [ + 325, + 295, + 542, + 306 + ], + "type": "text", + "content": "Table 6. Robustness of UIBDiffusion against random noise." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 445, + 554, + 495 + ], + "lines": [ + { + "bbox": [ + 317, + 445, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 554, + 495 + ], + "type": "table", + "html": "
Cifar 10CelebA-HQ-256
l∞l2l∞l2
UIBDiffusion Trigger0.041.800.0512.74
Prior Work0.476.210.8019.84
", + "image_path": "c61efb6226093eb7d41314aa14fc0fb82ec77cf3206c8c1cfe6ca7c87ea82f97.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 502, + 554, + 525 + ], + "lines": [ + { + "bbox": [ + 313, + 502, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 554, + 525 + ], + "type": "text", + "content": "Table 7. Quantitative comparison of triggers imperceptibility using distance metrics of " + }, + { + "bbox": [ + 313, + 502, + 554, + 525 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 313, + 502, + 554, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 502, + 554, + 525 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 313, + 502, + 554, + 525 + ], + "type": "text", + "content": " norms." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 539, + 388, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 388, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 388, + 551 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 559, + 555, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 555, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 555, + 703 + ], + "type": "text", + "content": "In this work, we present UIBDiffusion, the first imperceptible backdoor attack against diffusion models. We propose a novel trigger generation method based on universal adversarial perturbations and demonstrate that such perturbations for attacking discriminative models can be adapted as general, effective, and stealthy triggers for backdoor attacks against diffusion models. We empirically reveal the underlying rationale why UIBDiffusion can achieve a high attack success rate while evading the trigger inversion defensive algorithms. We hope our work sheds light on the potential risks that diffusion models face and motivates future research to develop more comprehensive defenses." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19193" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 90, + 294, + 113 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 90, + 294, + 113 + ], + "spans": [ + { + "bbox": [ + 56, + 90, + 294, + 113 + ], + "type": "text", + "content": "This work is supported in part by the National Science Foundation (NSF) SaTC-2426299 and SaTC-2413046." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 135, + 115, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 135, + 115, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 135, + 115, + 148 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 155, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 155, + 295, + 187 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 155, + 295, + 187 + ], + "spans": [ + { + "bbox": [ + 61, + 155, + 295, + 187 + ], + "type": "text", + "content": "[1] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 189, + 296, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 189, + 296, + 221 + ], + "spans": [ + { + "bbox": [ + 62, + 189, + 296, + 221 + ], + "type": "text", + "content": "[2] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In 9th International Conference on Learning Representations, ICLR 2021, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 223, + 295, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 223, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 62, + 223, + 295, + 277 + ], + "type": "text", + "content": "[3] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, pages 10674-10685. IEEE, 2022. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 278, + 295, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 278, + 295, + 311 + ], + "spans": [ + { + "bbox": [ + 62, + 278, + 295, + 311 + ], + "type": "text", + "content": "[4] William Peebles and Saining Xie. Scalable diffusion models with transformers. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, pages 4172-4182, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 312, + 294, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 312, + 294, + 388 + ], + "spans": [ + { + "bbox": [ + 62, + 312, + 294, + 388 + ], + "type": "text", + "content": "[5] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, ICML 2024. OpenReview.net, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 390, + 295, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 390, + 295, + 433 + ], + "spans": [ + { + "bbox": [ + 62, + 390, + 295, + 433 + ], + "type": "text", + "content": "[6] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 434, + 295, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 295, + 478 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 295, + 478 + ], + "type": "text", + "content": "[7] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 480, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 295, + 544 + ], + "type": "text", + "content": "[8] Yuming Jiang, Tianxing Wu, Shuai Yang, Chenyang Si, Dahua Lin, Yu Qiao, Chen Change Loy, and Ziwei Liu. Videobooth: Diffusion-based video generation with image prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6689-6700, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 546, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 546, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 546, + 295, + 590 + ], + "type": "text", + "content": "[9] Xiang Li, John Thickstun, Ishaan Gulrajani, Percy S Liang, and Tatsunori B Hashimoto. Diffusion-lm improves controllable text generation. Advances in Neural Information Processing Systems, 35:4328-4343, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 591, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 645 + ], + "type": "text", + "content": "[10] Xingqian Xu, Zhangyang Wang, Gong Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text, images and variations all in one diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7754-7765, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "text", + "content": "[11] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[12] Myeonghun Jeong, Hyeongju Kim, Sung Jun Cheon, Byoung Jin Choi, and Nam Soo Kim. Diff-nts: A denois" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "type": "text", + "content": "ing diffusion model for text-to-speech. arXiv preprint arXiv:2104.01409, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 554, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 554, + 139 + ], + "type": "text", + "content": "[13] Rongjie Huang, Max WY Lam, Jun Wang, Dan Su, Dong Yu, Yi Ren, and Zhou Zhao. Fastdiff: A fast conditional diffusion model for high-quality speech synthesis. arXiv preprint arXiv:2204.09934, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "text", + "content": "[14] Heeseung Kim, Sungwon Kim, and Sungroh Yoon. Guidedts: A diffusion model for text-to-speech via classifier guidance. In International Conference on Machine Learning, pages 11119-11133. PMLR, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 186, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 230 + ], + "type": "text", + "content": "[15] Vadim Popov, Ivan Vovk, Vladimir Gogoryan, Tasnama Sadekova, and Mikhail Kudinov. Grad-tts: A diffusion probabilistic model for text-to-speech. In International Conference on Machine Learning, pages 8599-8608. PMLR, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "type": "text", + "content": "[16] Yuancheng Wang, Zeqian Ju, Xu Tan, Lei He, Zhizheng Wu, Jiang Bian, et al. Audit: Audio editing by following instructions with latent diffusion models. Advances in Neural Information Processing Systems, 36:71340-71357, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 276, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 553, + 331 + ], + "type": "text", + "content": "[17] Shuai Shen, Wenliang Zhao, Zibin Meng, Wanhua Li, Zheng Zhu, Jie Zhou, and Jiwen Lu. Difftalk: Crafting diffusion models for generalized audio-driven portraits animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1982-1991, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 332, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 332, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 317, + 332, + 553, + 374 + ], + "type": "text", + "content": "[18] Bo Li, Yining Wang, Aarti Singh, and Yevgeniy Vorobeychik. Data poisoning attacks on factorization-based collaborative filtering. Advances in neural information processing systems, 29, 2016. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 376, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 376, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 317, + 376, + 553, + 420 + ], + "type": "text", + "content": "[19] Mengchen Zhao, Bo An, Yaodong Yu, Sulin Liu, and Sinno Pan. Data poisoning attacks on multi-task relationship learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 422, + 555, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 555, + 464 + ], + "type": "text", + "content": "[20] Scott Alfeld, Xiaojin Zhu, and Paul Barford. Data poisoning attacks against autoregressive models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 30, 2016." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 467, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 467, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 467, + 553, + 510 + ], + "type": "text", + "content": "[21] Bingyin Zhao and Yingjie Lao. Towards class-oriented poisoning attacks against neural networks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3741-3750, 2022." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 512, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 553, + 555 + ], + "type": "text", + "content": "[22] Bingyin Zhao and Yingjie Lao. Clpa: Clean-label poisoning availability attacks using generative adversarial nets. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 9162-9170, 2022. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 557, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 553, + 600 + ], + "type": "text", + "content": "[23] Sheng-Yen Chou, Pin-Yu Chen, and Tsung-Yi Ho. How to backdoor diffusion models? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4015-4024, 2023. 1, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 601, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 553, + 645 + ], + "type": "text", + "content": "[24] Weixin Chen, Dawn Song, and Bo Li. Trojdiff: Trojan attacks on diffusion models with diverse targets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4035-4044, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "text", + "content": "[25] Sheng-Yen Chou, Pin-Yu Chen, and Tsung-Yi Ho. Villan-diffusion: A unified backdoor attack framework for diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[26] Shengwei An, Sheng-Yen Chou, Kaiyuan Zhang, Qiuling Xu, Guanhong Tao, Guangyu Shen, Siyuan Cheng, Shiqing" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19194" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 126 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 126 + ], + "type": "text", + "content": "Ma, Pin-Yu Chen, Tsung-Yi Ho, et al. Elijah: Eliminating backdoors injected in diffusion models via distribution shift. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 10847-10855, 2024. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 296, + 183 + ], + "type": "text", + "content": "[27] Yichuan Mo, Hui Huang, Mingjie Li, Ang Li, and Yisen Wang. TERD: A unified framework for safeguarding diffusion models against backdoors. In *Forty-first International Conference on Machine Learning*, ICML 2024, 2024. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 295, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 295, + 226 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 295, + 226 + ], + "type": "text", + "content": "[28] Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying vulnerabilities in the machine learning model supply chain. arXiv preprint arXiv:1708.06733, 2017. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 295, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 295, + 260 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 295, + 260 + ], + "type": "text", + "content": "[29] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. CoRR, abs/1712.05526, 2017." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 295, + 316 + ], + "type": "text", + "content": "[30] Yingqi Liu, Shiqing Ma, Yousra Aafer, Wen-Chuan Lee, Juan Zhai, Weihang Wang, and Xiangyu Zhang. Trojaning attack on neural networks. In 25th Annual Network and Distributed System Security Symposium, NDSS. The Internet Society, 2018. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 295, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 295, + 370 + ], + "type": "text", + "content": "[31] Yuanshun Yao, Huiying Li, Haitao Zheng, and Ben Y. Zhao. Latent backdoor attacks on deep neural networks. In Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security, CCS 2019, pages 2041-2055. ACM, 2019. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 295, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 295, + 426 + ], + "type": "text", + "content": "[32] Yunfei Liu, Xingjun Ma, James Bailey, and Feng Lu. Reflection backdoor: A natural backdoor attack on deep neural networks. In Computer Vision - ECCV 2020, Proceedings, Part X, volume 12355 of Lecture Notes in Computer Science, pages 182-199. Springer, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 295, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 295, + 480 + ], + "type": "text", + "content": "[33] Yuezun Li, Yiming Li, Baoyuan Wu, Longkang Li, Ran He, and Siwei Lyu. Invisible backdoor attack with sample-specific triggers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16463-16472, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 481, + 295, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 295, + 514 + ], + "type": "text", + "content": "[34] Tuan Anh Nguyen and Anh Tuan Tran. Wanet - imperceptible warping-based backdoor attack. In 9th International Conference on Learning Representations, ICLR, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 514, + 295, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 295, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 295, + 580 + ], + "type": "text", + "content": "[35] Siyuan Cheng, Yingqi Liu, Shiqing Ma, and Xiangyu Zhang. Deep feature space trojan attack of neural networks by controlled detoxification. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, pages 1148–1156. AAAI Press, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 581, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 295, + 624 + ], + "type": "text", + "content": "[36] Khoa Doan, Yingjie Lao, Weijie Zhao, and Ping Li. Lira: Learnable, imperceptible and robust backdoor attacks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11966-11976, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 295, + 667 + ], + "type": "text", + "content": "[37] Khoa Doan, Yingjie Lao, and Ping Li. Backdoor attack with imperceptible input and latent modification. Advances in Neural Information Processing Systems, 34:18944-18957, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 669, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 295, + 712 + ], + "type": "text", + "content": "[38] Khoa D Doan, Yingjie Lao, and Ping Li. Marksman backdoor: Backdoor attacks with arbitrary target class. Advances in Neural Information Processing Systems, 35:38260-38273, 2022. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "text", + "content": "[39] Aniruddha Saha, Akshayvarun Subramanya, and Hamed Pirsiavash. Hidden trigger backdoor attacks. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, pages 11957-11965. AAAI Press, 2020. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 554, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 184 + ], + "type": "text", + "content": "[40] Hossein Souri, Liam Fowl, Rama Chellappa, Micah Goldblum, and Tom Goldstein. Sleeper agent: Scalable hidden trigger backdoors for neural networks trained from scratch. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 185, + 554, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 554, + 251 + ], + "type": "text", + "content": "[41] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Anti-backdoor learning: Training clean models on poisoned data. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 14900-14912, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 252, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 554, + 297 + ], + "type": "text", + "content": "[42] Qi Zhao and Christian Wressnegger. Adversarily robust anti-backdoor learning. In Proceedings of the 2024 Workshop on Artificial Intelligence and Security, AISec 2024, pages 77-88. ACM, 2024." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 297, + 554, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 554, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 554, + 353 + ], + "type": "text", + "content": "[43] Bolun Wang, Yuanshun Yao, Shawn Shan, Huiying Li, Bimal Viswanath, Haitao Zheng, and Ben Y. Zhao. Neural cleanse: Identifying and mitigating backdoor attacks in neural networks. In 2019 IEEE Symposium on Security and Privacy, SP 2019, pages 707-723. IEEE, 2019." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 354, + 554, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 554, + 408 + ], + "type": "text", + "content": "[44] Dongxian Wu and Yisen Wang. Adversarial neuron pruning purifies backdoored deep models. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, pages 16913-16925, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 410, + 554, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 554, + 453 + ], + "type": "text", + "content": "[45] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Neural attention distillation: Erasing backdoor triggers from deep neural networks. In 9th International Conference on Learning Representations, ICLR 2021, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 455, + 554, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 554, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 554, + 488 + ], + "type": "text", + "content": "[46] Bingyin Zhao and Yingjie Lao. Ultraclean: A simple framework to train robust neural networks against backdoor attacks. arXiv preprint arXiv:2312.10657, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 489, + 554, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 554, + 544 + ], + "type": "text", + "content": "[47] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In 2nd International Conference on Learning Representations, ICLR 2014, 2014. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 545, + 554, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 554, + 601 + ], + "type": "text", + "content": "[48] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Universal adversarial perturbations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1765-1773, 2017. 2, 3, 4, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 601, + 554, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 554, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 554, + 656 + ], + "type": "text", + "content": "[49] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, pages 11895-11907, 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "type": "text", + "content": "[50] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In 9th International Conference on Learning Representations, ICLR 2021. OpenReview.net, 2021." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19195" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[51] Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 160 + ], + "type": "text", + "content": "[52] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. In The Eleventh International Conference on Learning Representations, ICLR 2023, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 216 + ], + "type": "text", + "content": "[53] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "type": "text", + "content": "[54] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 294, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 336 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 336 + ], + "type": "text", + "content": "[55] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 369 + ], + "type": "text", + "content": "[56] Joseph Clements and Yingjie Lao. Hardware trojan attacks on neural networks. arXiv preprint arXiv:1806.05768, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 425 + ], + "type": "text", + "content": "[57] Shengfang Zhai, Yinpeng Dong, Qingni Shen, Shi Pu, Yuejian Fang, and Hang Su. Text-to-image diffusion models can be easily backdoored through multimodal data poisoning. In Proceedings of the 31st ACM International Conference on Multimedia, MM 2023, pages 1577-1587. ACM, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 459 + ], + "type": "text", + "content": "[58] Yihao Huang, Qing Guo, and Felix Juefei-Xu. Zero-day backdoor attack against text-to-image diffusion models via personalization. CoRR, abs/2305.10701, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "text", + "content": "[59] Lukas Struppek, Dominik Hintersdorf, and Kristian Kersting. Rickrolling the artist: Injecting backdoors into text encoders for text-to-image synthesis. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, pages 4561-4573. IEEE, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 515, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 602 + ], + "type": "text", + "content": "[60] Bryant Chen, Wilka Carvalho, Nathalie Baracaldo, Heiko Ludwig, Benjamin Edwards, Taesung Lee, Ian M. Moly, and Biplav Srivastava. Detecting backdoor attacks on deep neural networks by activation clustering. In Workshop on Artificial Intelligence Safety 2019 co-located with the Thirty-Third AAAI Conference on Artificial Intelligence 2019 (AAAI-19), volume 2301 of CEUR Workshop Proceedings. CEUR-WS.org, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 603, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 646 + ], + "type": "text", + "content": "[61] Yi Zeng, Won Park, Z. Morley Mao, and Ruoxi Jia. Rethinking the backdoor attacks' triggers: A frequency perspective. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, pages 16453-16461. IEEE, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[62] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In 3rd International Conference on Learning Representations, ICLR 2015, 2015. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[63] Alexey Kurakin, Ian J. Goodfellow, and Samy Bengio. Adversarial examples in the physical world. In 5th International" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 597 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "Conference on Learning Representations, ICLR 2017, Workshop Track Proceedings, 2017." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[64] Chaowei Xiao, Bo Li, Jun-Yan Zhu, Warren He, Mingyan Liu, and Dawn Song. Generating adversarial examples with adversarial networks. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJ-CAI 2018, pages 3905–3911. ijcai.org, 2018. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "type": "text", + "content": "[65] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2574–2582, 2016. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 207, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 553, + 261 + ], + "type": "text", + "content": "[66] Yanghao Zhang, Wenjie Ruan, Fu Wang, and Xiaowei Huang. Generalizing universal adversarial attacks beyond additive perturbations. In 2020 IEEE International Conference on Data Mining (ICDM), pages 1412-1417. IEEE, 2020. 4, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 262, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 307 + ], + "type": "text", + "content": "[67] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 308, + 553, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 339 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 339 + ], + "type": "text", + "content": "[68] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical Report, 2009. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "type": "text", + "content": "[69] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In 2015 IEEE International Conference on Computer Vision, ICCV 2015, pages 3730-3738. IEEE Computer Society, 2015. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 387, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 553, + 441 + ], + "type": "text", + "content": "[70] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In Yoshua Bengio and Yann LeCun, editors, 3rd International Conference on Learning Representations, ICLR 2015, 2015. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 442, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 442, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 442, + 553, + 487 + ], + "type": "text", + "content": "[71] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 487, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 553, + 553 + ], + "type": "text", + "content": "[72] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, pages 6626-6637, 2017. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 554, + 553, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 554, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 316, + 554, + 553, + 597 + ], + "type": "text", + "content": "[73] Hadi Salman, Mingjie Sun, Greg Yang, Ashish Kapoor, and J Zico Kolter. Denoised smoothing: A provable defense for pretrained classifiers. Advances in Neural Information Processing Systems, 33:21945-21957, 2020. 8" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19196" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/UMFN_ Unified Multi-Domain Face Normalization for Joint Cross-domain Prototype Learning and Heterogeneous Face Recognition/3e3f9780-5e1d-4d68-aee1-68dcb88c5368_content_list.json b/2025/UMFN_ Unified Multi-Domain Face Normalization for Joint Cross-domain Prototype Learning and Heterogeneous Face Recognition/3e3f9780-5e1d-4d68-aee1-68dcb88c5368_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..95ddb907315ce710a6bf4d9ebd4d2a3a01abc4a8 --- /dev/null +++ b/2025/UMFN_ Unified Multi-Domain Face Normalization for Joint Cross-domain Prototype Learning and Heterogeneous Face Recognition/3e3f9780-5e1d-4d68-aee1-68dcb88c5368_content_list.json @@ -0,0 +1,1482 @@ +[ + { + "type": "text", + "text": "UMFN: Unified Multi-Domain Face Normalization for Joint Cross-domain Prototype Learning and Heterogeneous Face Recognition", + "text_level": 1, + "bbox": [ + 124, + 128, + 875, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meng Pang1, Wenjun Zhang1, Nanrun Zhou2*, Shengbo Chen1*, Hong Rao1 \n1Nanchang University \n2Shanghai University of Engineering Science", + "bbox": [ + 197, + 202, + 795, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{mengpang@, zhangwenjun@, raohong@}ncu.edu.cn, nrzhou@sues.edu.cn, ccb02kingdom@gmail.com", + "bbox": [ + 120, + 258, + 870, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 309, + 326, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Face normalization aims to enhance the robustness and effectiveness of face recognition systems by mitigating intra-personal variations in expressions, poses, occlusions, illuminations, and domains. Existing methods face limitations in handling multiple variations and adapting to cross-domain scenarios. To address these challenges, we propose a novel Unified Multi-Domain Face Normalization Network (UMFN) model, which can process face images with various types of facial variations from different domains, and reconstruct frontal, neutral-expression facial prototypes in the target domain. As an unsupervised domain adaptation model, UMFN facilitates concurrent training on multiple datasets across domains and demonstrates strong prototype reconstruction capabilities. Notably, UMFN serves as a joint prototype and feature learning framework, enabling the simultaneous extraction of domain-agnostic identity features through a decoupling mapping network and a feature domain classifier for adversarial training. Moreover, we design an efficient Heterogeneous Face Recognition (HFR) network that fuses domain-agnostic and identity-discriminative features for HFR, and introduce contrastive learning to enhance identity recognition accuracy. Empirical studies on diverse cross-domain face datasets validate the effectiveness of our proposed method.", + "bbox": [ + 89, + 340, + 483, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 747, + 220, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Face normalization, as a core technology in the field of computer vision, is dedicated to eliminating interfering variation factors in facial images, such as pose, expression, occlusion, and illumination, and reconstructing a frontal, unoccluded, evenly illuminated facial prototype with a neutral expression [34]. This technology has shown broad application prospects in numerous areas, including criminal investiga", + "bbox": [ + 89, + 771, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8985e2ee59487c050ba845e114c60d9a26ee5e63722dfb990b40c34bb2ba5d60.jpg", + "image_caption": [ + "Figure 1. Illustration of a desired face normalization method capable of handling multiple cross-domain face normalization tasks and addressing various types of facial variations." + ], + "image_footnote": [], + "bbox": [ + 527, + 309, + 898, + 563 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion , authentication, and human-computer interaction [39]. However, existing face normalization methods face challenges in coping with increasingly complex and diversified application scenarios, highlighting the urgent need for more advanced and versatile solutions to flexibly adapt to various changes in different environments and conditions.", + "bbox": [ + 511, + 625, + 906, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The limitations of current face normalization methods lie mainly in their constraints in processing single contaminant elements and their single-domain applicability. Most methods [3, 10] focus on modeling and processing a single element, such as pose, expression, illumination, or occlusion, and struggle with multiple coexisting factors. Additionally, some unsupervised methods [41, 46] may inadvertently lose crucial identity information, leading to identity inconsistency between the generated prototype and the input face. Furthermore, existing models [32, 40] are typically domain-specific, lacking universality across multi-domain scenarios. However, practical applications often involve images", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors", + "bbox": [ + 117, + 887, + 250, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29299", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "from various domains, such as visible light, near-infrared, and sketch images used in criminal investigations. Therefore, there is an urgent need for a multi-domain face normalization method capable of handling various facial variations, as illustrated in Fig. 1.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose a Unified Multidomain Face Normalization Network (UMFN) model. This model is capable of handling facial images from different domains and containing multiple contaminant elements (such as pose, expression, occlusion, etc.), and can reconstruct a frontal, neutral facial prototype in the target domain while simultaneously extracting identity features unrelated to domain information. It is worth noting that, despite the significant achievements of diffusion models in the field of image generation, their limitations in interpretable representation learning make them difficult to meet our needs for joint prototype and feature learning [42]. Therefore, this paper adopts a fusion framework of Autoencoder and Generative Adversarial Network (GAN), as shown in Fig. 2.", + "bbox": [ + 88, + 169, + 482, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The UMFN model introduces a feature decoupling network and domain classifier for adversarial training, decoupling domain-agnostic features. By leveraging concatenated masks, it regulates clean prototype generation within specified domains or reconstructs original images, ensuring identity consistency. Additionally, a multi-scale InstanceNorm Patch-based discriminator [19] enhances image details and textures, improving visual quality. As an unsupervised domain adaptation model, UMFN enables concurrent training of diverse, unlabeled datasets within a single network, demonstrating flexibility and practicality in cross-domain prototype reconstruction. Furthermore, we design an efficient heterogeneous face recognition (HFR) network in conjunction with UMFN's encoder to extract identity-discriminative features for the generated prototype. These features are fused with domain-agnostic features of input for HFR. During HFR network training, a contrastive learning mechanism minimizes the feature distance between generated and authentic prototypes, enhancing the network's ability to accurately recognize identities.", + "bbox": [ + 91, + 383, + 482, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of the paper are summarized below:", + "bbox": [ + 109, + 686, + 467, + 700 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a unified face normalization model, dubbed UMFN, which addresses the limitations of existing methods in handling multiple variation factors and cross-domain face images. This model is capable of simultaneously processing various types of facial variations and is applicable to face images from different domains.", + "- We design an efficient HFR network that fuses domain-agnostic features from input sample with identity-discriminative features from generated prototype for HFR. Furthermore, a contrastive learning mechanism is introduced to bolster the network's capability in accurately recognizing identities.", + "- As an unsupervised domain adaptation model, UMFN en" + ], + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ables simultaneous training of multiple datasets with different domains and without identity labels in a single network. It is able to perform prototype reconstruction tasks across multiple domains, enhancing model flexibility and practicality, adapting to diverse scenarios, and reducing reliance on labeled data.", + "bbox": [ + 526, + 90, + 903, + 180 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Empirical study on multiple cross-domain face datasets with facial variations validates the superiority of UMFN in face normalization tasks in both single-modal and cross-modal scenarios, as well as in HFR task.", + "bbox": [ + 513, + 181, + 903, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 256, + 661, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Face normalization, an emerging and highly esteemed research field, has shown substantial practical value in areas like criminal identification and facial evidence collection. Mainstream face normalization methods primarily address facial variation removal in single-modal scenarios, aiming to eliminate factors such as diverse lighting conditions, exaggerated facial expressions, large head poses, or partially occlusions, thereby restoring the standard facial prototype. Existing single-modal face normalization algorithms can be broadly classified into two categories [33]: The first constructs clustering models based on auxiliary sample information to estimate facial prototypes of contaminated samples. The second leverages generative AI models, such as GANs [8] and Diffusion Models [4], to learn the mapping between contaminated and standard facial samples.", + "bbox": [ + 511, + 281, + 903, + 507 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the first category, Gao et al. [7] combined unlabeled facial samples with a labeled training set, using a Gaussian Mixture Model for clustering to estimate prototypes. Pang et al. [31] introduced a more robust semi-supervised low-rank representation model for clustering and prototype estimation, updated with labeled retrieval samples. Although effective in prototype reconstruction, these methods require prior access to the facial samples, making them unsuitable for real-time scenarios. In the second category, benefiting from generative AI models' image generation and mapping capabilities, researchers have successively proposed a series of generative learning methods, aiming to eliminate expression variations [3, 37], illumination variations [10, 41], pose variations [15, 27, 46], and partial occlusions [21, 24, 40] in input facial samples, respectively, and generate facial prototype images with consistent identities. However, these methods are typically limited to specific types of facial variations and cannot effectively address combinations of multiple types. Pang et al. [32] attempts to solve prototype reconstruction for multiple types of contamination by designing a Variation Disentangling GAN (VD-GAN), but has limitations in preserving identity characteristics and handling only single-domain contamination.", + "bbox": [ + 511, + 508, + 903, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, a few of cross-domain face synthesis methods [1, 23, 26, 28, 36, 49] have been proposed, including Cycle-consistency GAN (CycleGAN) [49], Parallel Multi", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29300", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/caa5ac8b6ec3f66c293a79d26b20b97aff27d1caa2ed286978203f688d1e02e4.jpg", + "image_caption": [ + "Figure 2. Overview of the architecture of UMFN. $x_{c} \\in X_{c}$ denotes an input image in domain $c$ , $\\hat{x}_c$ represents the reconstructed image of $x_{c}$ , $\\hat{x}_{\\bar{c}}^{rp}$ denotes the generated prototype of $x_{c}$ in domain $\\bar{c}$ , and $x_{c}^{rp}$ represents the real prototype in domain $c$ ." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 903, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "stage GANs (PMSGAN) [1], Brownian Bridge Diffusion Models (BBDM) [23], and NIR-FER Stochastic Differential Equations (NFER-SDE) [28]. These methods focus on transforming domain styles while preserving facial details, but they cannot remove facial variations from in-put faces. In light of the above, this paper aims to design a multidomain face normalization scheme applicable to universal facial variation removal and explore effective integration and optimization of domain transformation and prototype learning within a unified framework.", + "bbox": [ + 88, + 361, + 483, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 89, + 526, + 191, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This paper proposes a universal face normalization framework, UMFN, whose architecture is shown in Fig. 2. UMFN is designed to uniformly process face images with various types of facial variations from different image domains, ultimately generating standard facial prototypes in the specified domain. Correspondingly, UMFN comprises the following two core components: 1) addressing the issue of input images originating from multiple domains and generating images in the specified domain; 2) uniformly handling multiple types of facial variations while generating prototypes that preserve the identity (ID). These two components will be detailed in Sec. 3.1 and Sec. 3.2", + "bbox": [ + 88, + 551, + 483, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Feature Decoupling and Cross-Domain Generation", + "text_level": 1, + "bbox": [ + 89, + 742, + 482, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To tackle the challenge that input face images may come from different domains, we have designed a feature decoupling mapping network. This network is able to separate domain information from the latent space, resulting in domain-invariant representations. Such decoupled, domain-agnostic identity features can be applied to HFR task.", + "bbox": [ + 89, + 779, + 482, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feature Decoupling: We design a feature decoupling mapping network that initially extracts primary feature $z =$", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$G_{enc}(x_c)$ from the input image $x_c$ using a pre-trained encoder $G_{enc}$ , and then derives its latent feature representation $p = \\text{map}(z)$ through the mapping network map. Specifically, the domain feature classifier $D_c$ receives the latent representations $p$ as input and outputs the domain prediction code (i.e., the image domain to which $p$ belongs). The goal of $D_c$ is to accurately predict the domain of the input latent representation $p$ , while the mapping network map aims to confuse $D_c$ 's correct domain prediction. Through adversarial training, the interplay between the two achieves domain information decoupling. Accordingly, the objective function for training $D_c$ is presented as follows:", + "bbox": [ + 511, + 362, + 906, + 544 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {D _ {c}} = E \\left[ \\log P \\left(l _ {p} ^ {\\text {p r e}} = l _ {x _ {c}} ^ {\\text {r e a l}} \\mid p\\right) \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 555, + 906, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $l_{p}^{pre}$ is the domain label predicted by $D_{c}$ , $l_{x_{c}}^{real}$ is the real domain label of $x_{c}$ . On the other side, the objective function for training the mapping network map is as", + "bbox": [ + 511, + 585, + 905, + 633 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {m a p} = - E \\left[ \\log P \\left(l _ {p} ^ {p r e} = l _ {x _ {c}} ^ {r e a l} \\mid p\\right) \\right]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 643, + 906, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cross-domain Generation: We introduce the control mask and combine it with reconstruction learning to achieve cross-domain image generation. The core objective of reconstruction learning is to guide the decoder $G_{dec}$ to learn how to decode domain-agnostic latent representation $p = \\text{map}(G_{enc}(x_c))$ conditioned on the original attribute mask $v_c$ , thereby reconstructing the input image $x_c$ . This process ensures that the decoder not only learns to understand and apply the semantics of the mask but also fully utilizes the identity feature information in the latent representation $p$ . As such, in subsequent steps, images from specified domains can be generated by manipulating the mask. The objective function for reconstruction learning is as follows:", + "bbox": [ + 511, + 675, + 906, + 871 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {r e c} = E _ {x _ {c} \\sim p _ {d a t a}} \\left[ \\| x _ {c} - \\hat {x} _ {c} \\| _ {1} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 883, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "29301", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{x}_c = G_{dec}(p, v_c)$ is the reconstructed image of $x_c$ , $v_c$ is the original attribute mask of the input image $x_c$ , indicating whether the $x_c$ image is polluted and its domain information. Specifically, $v_c$ is composed of a four-digit one-hot vector, where the first digit indicates whether the image is contaminated, and the last three digits represent the domain it belongs to. In Eq. (3), $l_1$ norm is employed instead of the $l_2$ norm to suppress the issue of blurred generated images.", + "bbox": [ + 89, + 90, + 483, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Prototype generation", + "text_level": 1, + "bbox": [ + 89, + 220, + 290, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Different from prior single-contamination normalization methods that overly focus on facial variation categories and how to accurately model them, our method returns to the essence of the prototype reconstruction task, aiming to directly generate identity-preserved standard facial prototypes by incorporating identity information from the contaminated facial samples, potentially bypassing the challenges of modeling facial variations in a clever manner.", + "bbox": [ + 89, + 241, + 483, + 361 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notably, we introduce a prototype adversarial loss, differing from the adversarial loss in traditional GAN-based image translation methods that aim to approximate the distribution of original face data with diverse variations. Our loss aims to approximate a condensed distribution of standard facial prototypes, while also reducing overfitting to these variations. Furthermore, we have designed a multiscale Markov discriminator $D = [D_w, D_f]$ to refine the details of generated faces and stabilize training. Specifically, we concatenate the extracted domain-agnostic identity feature $p = \\text{map}(G_{\\text{enc}}(x_c))$ with a control mask $v_{\\bar{c}}$ , where the first bit in $v_{\\bar{c}}$ controls whether to perform face normalization: 1 for yes, and the last three bits control the target domain for prototype generation. Next, this concatenated feature is input into the decoder to generate a pixel-level standardized facial prototype in the specified target domain. The prototype adversarial objectives for training the generator and the multi-scale discriminator are as follows:", + "bbox": [ + 89, + 362, + 483, + 633 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {D} ^ {\\text {a d v}} = D _ {w} \\left(\\hat {x} _ {\\bar {c}} ^ {r p}\\right) + D _ {f} \\left(\\operatorname {c r o p} \\left(\\hat {x} _ {\\bar {c}} ^ {r p}\\right)\\right) \\tag {4} \\\\ - D _ {w} \\left(x _ {c} ^ {r p}\\right) - D _ {f} \\left(\\operatorname {c r o p} \\left(x _ {c} ^ {r p}\\right)\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 638, + 483, + 676 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {G _ {d e c}} ^ {a d v} = - \\left(D _ {w} \\left(\\hat {x} _ {\\bar {c}} ^ {r p}\\right) + D _ {f} \\left(\\operatorname {c r o p} \\left(\\hat {x} _ {\\bar {c}} ^ {r p}\\right)\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 681, + 482, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $D_w$ and $D_f$ represent the global path discriminator and local path discriminator, respectively. $\\hat{x}_{\\bar{c}}^{rp} = G_{dec}(map(G_{enc}(x_c), v_{\\bar{c}})$ denotes the specified domain prototype generated by inputting $x_c$ and the control mask $v_{\\bar{c}}$ while $x_c^{rp}$ represents the uncontaminated faces in the training set. The function \"crop\" is used to cut out the facial features from the input face images. Considering that the Wasserstein distance is more stable and can avoid training instability and mode collapse in GANs, we adopt the WGAN-GP [9] loss in our model to replace the traditional cross-entropy loss.", + "bbox": [ + 89, + 704, + 483, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Furthermore, to ensure that the generated prototype $\\hat{x}_{\\bar{c}}^{rp}$ can effectively preserve the identity information from the", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "input $x_{c}$ , we have designed an identity feature similarity loss function. This is achieved by feeding the generated prototype back into the encoder and mapping network to obtain the latent feature map $(G_{enc}(\\hat{x_c}^{rp}))$ , and then calculating its similarity to the latent feature $p$ of the input $x_{c}$ , thereby retaining identity. The loss function is as follows:", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {i d} = \\| p - \\operatorname {m a p} \\left(G _ {e n c} \\left(\\hat {x} _ {\\bar {c}} ^ {r p}\\right)\\right) \\| _ {2} ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 196, + 903, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\| \\cdot \\|_2^2$ denotes $l_2$ -norm. On the other hand, UMFN transfers the identity feature retention ability from the reconstruction learning task to the face prototype generation task. Since these two tasks share the same input and output domains, the transferability gap [44] between them is minimal. Therefore, the identity feature retention ability learned from the reconstruction learning task can be easily transferred to the prototype reconstruction task. Therefore, the overall loss function of the decoder $G_{dec}$ is as follows:", + "bbox": [ + 511, + 222, + 906, + 358 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {G _ {d e c}} ^ {\\text {a d v}} = \\lambda_ {1} L _ {r e c} + \\lambda_ {2} L _ {i d}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 369, + 903, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ and $\\lambda_{2}$ are balancing hyperparameters.", + "bbox": [ + 511, + 401, + 834, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Heterogeneous Face Recognition", + "text_level": 1, + "bbox": [ + 511, + 426, + 799, + 443 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At this stage, we meticulously construct a HFR network, whose core components encompass an encoder $F_{\\theta}$ (its network architecture aligns with $G_{enc}$ in UMFN) and a fully connected layer $l_{\\theta}$ . The specific structure is shown in Fig. 3. The HFR network consists of two core operations:", + "bbox": [ + 511, + 449, + 905, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Feature Fusion: During training, we concatenate the domain-agnostic features (extracted by UMFN's $G_{enc}$ and mapping network map from the source domain input image) with the identity-discriminative features (extracted by encoder $F_{\\theta}$ from the real target domain prototype). The fused features are then fed into the softmax layer for classification. In this process, the fully connected layer $l_{\\theta}$ doubles the feature dimensions, enhancing classification robustness through multi-dimensional identity information integration. The classification loss for this step is presented as", + "bbox": [ + 511, + 526, + 905, + 676 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {c l s}} = \\operatorname {s o f t m a x} \\left(l _ {\\theta} \\left[ F _ {\\theta} \\left(x _ {v, i} ^ {r p}\\right) \\oplus \\operatorname {m a p} \\left(G _ {e n c} \\left(x _ {n, i}\\right)\\right) \\right], i\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 689, + 903, + 707 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x_{n,i}$ represents source domain image and $x_{v,i}^{rp}$ represents real target domain prototype, $F_{\\theta}(x_{v,i}^{rp})$ denotes the representation of the input image $x_{v,i}^{rp}$ extracted by $F_{\\theta}$ , while map $(G_{enc}(x_{n,i}))$ denotes the domain-agnostic feature of the input image $x_{n,i}$ , the symbol $\\oplus$ indicates the concatenation of these two features. During testing, we extract the identity-discriminative feature of the generated target domain prototype from the input test source domain image using the trained $F_{\\theta}$ , and fuse it with the domain-agnostic feature of the input test source domain image for HFR.", + "bbox": [ + 511, + 719, + 905, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Contrastive Learning Mechanism: During the cross-domain generation, generated target domain prototypes may", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29302", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3f40e077f6da44a76ea1795eb727d4d76c2469a3049e7eb57e36c47fc100f6c7.jpg", + "image_caption": [ + "Figure 3. Overview of the HFR network and training process." + ], + "image_footnote": [], + "bbox": [ + 112, + 89, + 467, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "suffer identity feature loss or noise introduction. Direct reliance on such data for identity classification training can lead to overfitting to missing or noisy features, impairing classification performance. To mitigate this, we innovatively employ a contrastive learning mechanism that aligns generated prototypes with real ones in the feature space, while reducing the impact of noise and missing features. This ensures that the encoder $F_{\\theta}$ extracts features closely resembling real prototypes. Specifically, we randomly sample two pairs of real heterogeneous images, $(x_{v,1}^{rp}, x_{n,1})$ and $(x_{v,2}^{rp}, x_{n,2})$ , where $x_{n,1}$ and $x_{n,2}$ are NIR images with contamination. We then use UMFN to perform cross-domain restoration on $x_{n,1}$ and $x_{n,2}$ , generating the corresponding target VIS domain prototypes $\\hat{x}_{v,1}^{rp}$ and $\\hat{x}_{v,2}^{rp}$ . Next, we form positive pairs $(x_{v,1}^{rp}, \\hat{x}_{v,1}^{rp})$ and $(x_{v,2}^{rp}, \\hat{x}_{v,2}^{rp})$ , and negative pairs $(x_{v,1}^{rp}, \\hat{x}_{v,2}^{rp})$ and $(x_{v,2}^{rp}, \\hat{x}_{v,1}^{rp})$ . The corresponding contrastive loss is as follows:", + "bbox": [ + 89, + 439, + 483, + 696 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {c o n t} = \\sum_ {j \\neq k} \\left(1 - \\langle f _ {j}, \\tilde {f} _ {j} \\rangle\\right) + \\left(1 - \\langle f _ {k}, \\tilde {f} _ {k} \\rangle\\right) \\\\ + \\max \\left(0, \\left\\langle f _ {j}, \\tilde {f} _ {k} \\right\\rangle - m\\right) + \\max \\left(0, \\left\\langle f _ {k}, \\tilde {f} _ {j} \\right\\rangle - m\\right), \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 708, + 480, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f_{j} = l_{\\theta}[F_{\\theta}(x_{v,j}^{rp}) \\oplus map(G_{enc}(x_{n,j}))]$ , $\\langle, \\rangle$ represents the calculation of the cosine similarity between two features, and $m$ is a margin value set at 0.5. In summary, the final loss function of the HFR network is expressed as", + "bbox": [ + 89, + 776, + 482, + 838 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {h f r} = L _ {c l s} + \\alpha L _ {\\text {c o n t}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 853, + 482, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ is a trade-off parameter.", + "bbox": [ + 89, + 885, + 308, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 90, + 645, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets Descriptions", + "text_level": 1, + "bbox": [ + 511, + 114, + 714, + 131 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LAMP-HQ [45] is a newly created, challenging large-scale NIR-VIS dataset comprising over 56,000 NIR and 16,000 VIS face images from 573 individuals, featuring diverse expressions, poses, lighting conditions, occlusions, and complex backgrounds. BUAA NIR-VIS [16] includes 150 identities, each with 9 NIR and 9 VIS images, covering various head poses and expressions. CUFSF [6] consists of 597 grayscale images paired with 597 real sketch images. CASIA NIR-VIS 2.0 [25] encompasses 725 identities, with 5-50 NIR and 1-22 VIS images per identity. IJB-A [22] and LFW [17] are widely-used benchmarks for face recognition and verification, featuring diverse poses, occlusions, and expressions in the visible domain.", + "bbox": [ + 511, + 137, + 906, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 345, + 702, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the experimental section, we primarily conduct four experiments: (1) Homogeneous Multi-domain Face Normalization Experiment, (2) Heterogeneous Multi-domain Face Normalization Experiment, (3) Heterogeneous Face Recognition Experiment. (4) Ablation Study. The specific configurations of these experiments are as follows:", + "bbox": [ + 511, + 369, + 905, + 459 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) Homogeneous Multi-domain Face Normalization Experiment: In this experiment, we utilize three types of domain data from the LAMP-HQ dataset to train our model, ensuring that all data originate from the same set of identities (IDs). The experiment primarily evaluates the model's performance in facial normalization across multi-domain data within the LAMP-HQ dataset. Specifically, we employ FP data from the LAMP-HQ dataset, where all NIR and VIS images from 100 IDs are used as the training set, while the remaining IDs are designated as the test set. As the LAMP-HQ dataset only contains NIR and VIS domain data, we use software to convert NIR photographs into sketch images, thereby simulating a scenario with three types of domain data. Furthermore, we transfer the model to other datasets, namely LFW and IJB-A, for testing. We assess the model's performance from both qualitative and quantitative perspectives and compare it with the latest single-modal face normalization techniques.", + "2) Heterogeneous Multi-domain Face Normalization Experiment: We use one type of domain data from three different datasets to demonstrate that the model can effectively work without paired data training, verifying its generalization ability across datasets. Specifically, we use VIS images of 250 IDs from the LAMP-HQ dataset as VIS domain data, all NIR images of 130 IDs from the BUAA dataset as NIR domain data, and Sketch images of 400 IDs from the CUFSF dataset as Sketch domain data for training.", + "3) Heterogeneous Face Recognition Experiment: We conduct heterogeneous face recognition experiments on the" + ], + "bbox": [ + 511, + 460, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29303", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d61d0faa86dee5a206156c1629087a8a2efe2d838a0898ee3d0dc513be6fd665.jpg", + "image_caption": [ + "Figure 4. In-domain face normalization results on the LAMP-HQ dataset." + ], + "image_footnote": [], + "bbox": [ + 99, + 94, + 903, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LAMP-HQ and CASIA NIR-VIS 2.0 datasets, following the standard evaluation protocol in [43].", + "bbox": [ + 89, + 277, + 482, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4) Ablation Study Experiment: We investigate the contribution of each component within the UMFN framework and explore the roles of the contrastive learning mechanism and feature fusion operation in the HFR network.", + "bbox": [ + 89, + 309, + 482, + 368 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Implementation Details and Parameter Setting", + "text_level": 1, + "bbox": [ + 89, + 377, + 482, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We employ a Resnet-50 pre-trained on the MS1M dataset and fine-tuned on VGGFace2 [2] as the encoder $G_{enc}$ to extract 2048-dimensional identity features, keeping $G_{enc}$ fixed during training and testing. The mapping network consists of multiple linear layers, inputting a 2048-dimensional vector and outputting a 2048-dimensional vector. The feature domain classifier $D_{c}$ is composed of linear layers and outputs the predicted domain category, while the decoder $G_{dec}$ outputs images of size $224 \\times 224$ . The discriminators $D_{W}$ and $D_{f}$ are both patchGAN discriminators with a receptive field of 70, using instance normalization as the normalization layer, and the crop size is $150 \\times 150$ .", + "bbox": [ + 88, + 398, + 482, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The hyperparameters $\\lambda_{1}$ and $\\lambda_{2}$ in Eq. (7) are set at $\\lambda_{1} = 10 - \\frac{cur_{epoch}}{sum_{epoch}} \\times 9$ and $\\lambda_{2} = 10$ , where $cur_{epoch}$ is the current number of training epochs and $sum_{epoch}$ is the total number of training epochs, which is 100. $\\alpha$ in Eq. (10) is empirically set at 0.001. During optimization, we use the Adam optimizer, with the learning rate set at 0.0002 and momentum set at 0.5.", + "bbox": [ + 89, + 580, + 482, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Homogeneous Multi-domain Normalization", + "text_level": 1, + "bbox": [ + 89, + 696, + 460, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we evaluate the performance of UMFN in homogeneous multi-domain face normalization using the LAMP-HQ dataset. The in-domain normalization results of UMFN are presented in Fig. 4. Fig. 4(a) showcases the results of inputting VIS domain faces corrupted by different facial variations such as pose, lighting, occlusion, and expression, and outputting normalized VIS domain faces. Fig. 4(b) displays the outcomes of inputting corrupted faces from the NIR and Sketch domains and outputting normalized faces in their respective domains. The bottom row serves as a reference for the true prototypes. From Fig. 4, it can be observed that UMFN is capable of simultaneously", + "bbox": [ + 88, + 719, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "normalizing various types of corrupted faces from multiple different domains within their respective domains, effectively preserving the identity of the input faces while obtaining facial prototypes. Additionally, we conduct a comparison between UMFN and other popular single-domain face normalization methods, including PIM [47], TP-GAN [18], DR-GAN [38], Hassner [11], and FNM [35], using contaminated face images from the LFW dataset as input. The face normalization results of each method are presented in Fig. 5. It can be observed that the target prototype obtained through UMFN normalization excels in both image quality and identity preservation.", + "bbox": [ + 511, + 277, + 903, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Moreover, following the protocols in [48] and [20], we quantitatively evaluate UMFN on the LFW and IJB-A datasets. By inputting restored faces into the ResNet-18 [12] network for face verification, we aim to assess whether UMFN-restored faces enhance VR performance. Additionally, we report the Symmetry Index (SI)[30] to provide a comprehensive comparison of the quality of prototypes generated by UMFN, R&R [48], and CFR [20]. Tab. 1 presents the verification performance and SI of UMFN across both datasets. The results indicate that UMFN normalization significantly boosts the verification performance of the ResNet model: on IJB-A, when combined with ResNet-18, it improves VR by $5.35\\%$ , with accuracy gains also observed on LFW. The experiments demonstrate that: (1) removing pose, expression, occlusion, and illumination variations from input faces allows the model to better focus on key identity features, and (2) UMFN effectively retains critical identity information while eliminating these facial variations. In addition, UMFN outperforms R&R and CFR in terms of SI, generating highly symmetrical faces that correct facial distortions.", + "bbox": [ + 511, + 460, + 905, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, we illustrate the cross-domain normalization results of the UMFN model in Fig. 6. Specifically, the inputs consist of contaminated facial images from domain A, while the outputs represent the generated prototypes in domains B and C. The bottom row displays the reference prototypes of the inputs in the VIS domain. It can be observed that the UMFN model successfully addresses the challenges of domain transfer and prototype learning,", + "bbox": [ + 511, + 779, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29304", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c958e0b0aee430106d624c974f01f1ca0a6a16dec2cd0a84f1f803690604af89.jpg", + "image_caption": [ + "Figure 5. Comparison with other single-domain face normalization methods on LFW dataset." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 480, + 186 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e7ac4ca62da999ce48fea1eeae2d63e1f0426690a6604f7f741fe97acbbb72b9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDatasets
LFWIJB-A
VR↑SI↓VR↑SI↓
Res18 [12]98.85-90.57±1.2-
Res18+R&R [48]98.9520.9391.98±0.720.85
Res18+CFR [20]99.2328.3193.36±0.624.76
Res18+UMFN(ours)99.5714.6195.92±0.715.81
", + "bbox": [ + 94, + 232, + 478, + 348 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Verification performance and Symmetry Index (SI) on LFW and IJB-A.", + "bbox": [ + 89, + 362, + 482, + 390 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "effectively preserving the identity characteristics of the inputs when generating uncontaminated facial prototypes for the target domains. The inspiring results demonstrate that UMFN is capable of performing effective domain-specific face normalization on facial images sourced from different domains and contaminated by various types of facial variations, all while efficiently retaining the key identity information of the inputs, utilizing a single network architecture.", + "bbox": [ + 89, + 400, + 483, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Heterogeneous Multi-domain Normalization", + "text_level": 1, + "bbox": [ + 89, + 531, + 465, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we evaluate the performance of UMFN in the heterogeneous multi-domain face normalization experiment. It is particularly noteworthy that there is currently no other face normalization method capable of adequately addressing this complex issue. Fig. 7 showcase the normalization results of images from the LAMP-HQ, BUAA, and CUFSF datasets across different domains (VIS, NIR, Sketch). We observe that regardless of the dataset (or domain) the input comes from, UMFN successfully generates prototypes that are highly matched to the target domain styles provided by another dataset. It effectively eliminates interfering factors such as pose, expression, and occlusion in the input face images while accurately preserving the input identity information. It is worth noting that due to the minimal facial variations in most images of the CUFSF dataset, which generally exhibit simple facial expressions such as smiles, the changes after normalization are primarily manifested in the mouth area. Furthermore, the challenging nature of cross-domain normalization tasks involving sketch images, which contain the least amount of facial detail information, leads to slight local distortions in the generated VIS and NIR prototypes. Despite these challenges, the results demonstrate UMFN's robust multi-", + "bbox": [ + 89, + 553, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f1aa9232f686686956327a233a07e4997ed972262da688b36245f5e293920d3a.jpg", + "image_caption": [ + "Figure 6. Cross-domain face normalization results on LAMP-HQ dataset. The second and third rows depict face normalization results in different domains from the inputs shown in the first row. The bottom row shows the real VIS prototypes for reference." + ], + "image_footnote": [], + "bbox": [ + 535, + 90, + 887, + 281 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "domain and multi-type face normalization capabilities, particularly when trained on unpaired datasets from multiple domains, where its performance remains impressive.", + "bbox": [ + 511, + 366, + 906, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Heterogeneous Face Recognition", + "text_level": 1, + "bbox": [ + 511, + 421, + 802, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this experiment, we employ UMFN to restore cross-domain prototypes from NIR data in the probe set, generating VIS prototypes. Subsequently, we utilize the trained HFR network to obtain fused features (i.e., concatenation of domain-agnostic and identity-discriminative features) for both these generated prototypes and the authentic prototypes in the gallery set. Finally, by computing the cosine similarity between the fused features of the generated prototypes and those of the gallery set, we obtain a similarity matrix, upon which we calculate the rank-1 recognition accuracy and VR at $\\mathrm{FAR} = 0.001$ .", + "bbox": [ + 511, + 444, + 906, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 2 presents a comparison of our method with other advanced HFR approaches, including PACH [5], AD-CANs [13], MMTN [29], NLPF [43], and LPL [14], on LAMP-HQ and CASIA NIR-VIS 2.0 datasets. It is observed that our method achieves the best performance on both datasets in terms of rank-1 accuracy and VR. This exceptional performance stems from: (1) Incorporating a contrastive learning mechanism that aligns generated prototypes with real ones in the feature space, reducing noise and missing feature impacts. This allows the HFR encoder $F_{\\theta}$ to extract identity-discriminative features from generated prototypes that closely resembling real ones. (2) The fusion of domain-agnostic and identity-discriminative features notably enhances the identity distinguishability of the combined features.", + "bbox": [ + 511, + 611, + 908, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.7. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 848, + 666, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To assess the effectiveness of UMFN and the contributions of its components, we train five variants by respectively re", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29305", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dc85ed7baf14bac2b617006ebe1886ea12340f895ff82741bd8704f8bb721c5c.jpg", + "image_caption": [ + "(a) Results on LAMP-HQ dataset", + "Figure 7. The normalization results of face images from the LAMP-HQ, BUAA, and CUFSF datasets across different domains." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 352, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cb414ba2e0d25491404cee4d865adef3f99eb6809f4d3d19b895708a06a974ec.jpg", + "image_caption": [ + "(b) Results on BUAA dataset" + ], + "image_footnote": [], + "bbox": [ + 372, + 93, + 625, + 276 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d76b988d75d360a3a8524248721da92d7fb3716fc24c53bee718b7616d785c62.jpg", + "image_caption": [ + "(c) Results on CUFSF dataset" + ], + "image_footnote": [], + "bbox": [ + 650, + 92, + 905, + 276 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/adca369babe2e45d1d31b3af6ecd1a7ad54cbf1a2a0f7de3f59bce00397a13a7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDatasets
LAMP-HQCASIANIR-VIS 2.0
Rank-1@FAR=0.001Rank-1@FAR=0.001
PACH [5] (2020)96.978.798.998.3
ADCANs [13] (2021)97.896.199.198.5
MMTN [29] (2022)97.779.299.498.7
NLPF [43] (2024)98.898.399.799.5
LPL [14] (2024)98.798.599.697.5
UMFN (ours)99.198.799.999.7
", + "bbox": [ + 94, + 325, + 478, + 448 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4a8bfd578af0da5e2ed26834dca44ef172a5c7767f2493a7c194a52d0e9cb09f.jpg", + "table_caption": [ + "Table 2. Rank-1 accuracy and verification rate on the LAMP-HQ NIR-VIS dataset and the CASIA NIR-VIS 2.0 dataset." + ], + "table_footnote": [], + "table_body": "
Feature fusionContrastiveDatasets
LAMP-HQCASIANIR-VIS 2.0
Rank-1@FAR=0.001Rank-1@FAR=0.001
91.287.695.795.1
✓94.790.597.596.7
✓97.395.398.798.2
✓✓99.198.799.999.7
", + "bbox": [ + 94, + 500, + 480, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation experiments of the contrastive learning mechanism and feature fusion in the HFR network.", + "bbox": [ + 89, + 601, + 482, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "moving the identity feature similarity loss $L_{id}$ , reconstruction loss $L_{rec}$ , feature decoupling loss $L_{map}$ in the mapping network, global path discriminator $D_w$ , and the local path discriminator $D_f$ . The face normalization results for these variants are presented in Fig. 8. Observations indicate that: (1) Without $L_{id}$ , generated prototypes fail to preserve identity characteristics; (2) The absence of $L_{rec}$ causes control masks to fail, preventing the generation of specified domain prototypes; (3) Lack of $L_{map}$ results in the retention of original domain information in cross-domain prototypes; (4) Lack of $D_w$ results in incomplete faces with poorly generated visual effects; (5) Removing $D_f$ leads to a loss of local details in generated prototypes.", + "bbox": [ + 88, + 642, + 482, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Furthermore, Tab. 3 quantitatively analyzes the impact of the contrastive learning mechanism and feature concatenation on HFR network training. Results show that contrastive learning significantly enhances the network's uti", + "bbox": [ + 88, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/442b66f559b7f2faf407945e0ba9f61f92a6503875cf5027d5bf582fc3230232.jpg", + "image_caption": [ + "Figure 8. Face normalization results of UMFN's variants." + ], + "image_footnote": [], + "bbox": [ + 537, + 328, + 887, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "lization of generated prototypes, allowing it to extract key features similar to real prototypes while reducing the effects of noise and feature loss. Additionally, fusion of domain-invariant and identity-discriminative features improves classification performance.", + "bbox": [ + 511, + 628, + 906, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 722, + 633, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper has proposed the new UMFN model. It effectively addresses the limitations in handling diverse facial variations and cross-domain adaptability by reconstructing frontal, neutral-expression facial prototypes, and enabling unsupervised domain adaptation. Through joint prototype and feature learning, along with a well-designed HFR network, our method significantly enhances identity recognition accuracy across various domains. Experiments have validated UMFN's superiority in single-modal and cross-modal face normalization tasks, as well as in HFR task.", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29306", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the Natural Science Foundation of China (62466036, 62162041), by the Natural Science Foundation of Jiangxi Province (20232BAB212025), by the Natural Science Foundation of Henan Province (242300421402), by the High-level and Urgently Needed Overseas Talent Programs of Jiangxi Province (20232BCJ25024, 20232BCJ25026), by the Science&Technology Planning Project of Shanghai (23010501800), and by the Key Research and Development Program of Jiangxi Province (20243BBG71035).", + "bbox": [ + 89, + 114, + 483, + 255 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 284, + 187, + 300 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Bing Cao, Nannan Wang, Jie Li, Qinghua Hu, and Xinbo Gao. Face photo-sketch synthesis via full-scale identity supervision. Pattern Recognition, 124:108446, 2022. 2, 3", + "[2] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age. In FG, pages 67-74, 2018. 6", + "[3] William Carver and Ifeoma Nwogu. Facial expression neutralization with stoicnet. In WACV, pages 201-208, 2021. 1, 2", + "[4] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):10850-10869, 2023. 2", + "[5] Boyan Duan, Chaoyou Fu, Yi Li, Xingguang Song, and Ran He. Cross-spectral face hallucination via disentangling independent factors. In CVPR, pages 7930-7938, 2020. 7, 8", + "[6] Hamed Kiani Galoogahi and Terence Sim. Face sketch recognition by local radon binary pattern: Lrbp. In ICIP, pages 1837-1840, 2012. 5", + "[7] Yuan Gao, Jiayi Ma, and Alan L Yuille. Semi-supervised sparse representation based classification for face recognition with insufficient labeled samples. IEEE Transactions on Image Processing, 26(5):2545-2560, 2017. 2", + "[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NeurIPS, 27, 2014. 2", + "[9] Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein gans. NeurIPS, 30, 2017. 4", + "[10] Dequan Guo, Lingrui Zhu, Shenggui Ling, Tianxiang Li, Gexiang Zhang, Qiang Yang, Ping Wang, Shiqi Jiang, Sidong Wu, and Junbao Liu. Face illumination normalization based on generative adversarial network. Natural Computing, 22(1):105-117, 2023. 1, 2", + "[11] Tal Hassner, Shai Harel, Eran Paz, and Roee Enbar. Effective face frontalization in unconstrained images. In CVPR, pages 4295-4304, 2015. 6", + "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6, 7" + ], + "bbox": [ + 93, + 309, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Weipeng Hu and Haifeng Hu. Adversarial disentanglement spectrum variations and cross-modality attention networks for nir-vis face recognition. IEEE Transactions on Multimedia, 23:145-160, 2020. 7, 8", + "[14] Weipeng Hu, Yiming Yang, and Haifeng Hu. Pseudo label association and prototype-based invariant learning for semi-supervised nir-vis face recognition. IEEE Transactions on Image Processing, 33:1448-1463, 2024. 7, 8", + "[15] Yibo Hu, Xiang Wu, Bing Yu, Ran He, and Zhenan Sun. Pose-guided photorealistic face rotation. In CVPR, pages 8398-8406, 2018. 2", + "[16] D Huang, Jia Sun, and Y Wang. The buaa-visir face database instructions. School Comput. Sci. Eng., Beihang Univ., Beijing, China, Tech. Rep. IRIP-TR-12-FR-001, 3(3): 8, 2012. 5", + "[17] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 5", + "[18] Rui Huang, Shu Zhang, Tianyu Li, and Ran He. Beyond face rotation: Global and local perception gan for photorealistic and identity preserving frontal view synthesis. In ICCV, pages 2439-2448, 2017. 6", + "[19] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In CVPR, pages 1125-1134, 2017. 2", + "[20] Yeong-Joon Ju, Gun-Hee Lee, Jung-Ho Hong, and Seong-Whan Lee. Complete face recovery gan: Unsupervised joint face rotation and de-occlusion from a single-view image. In WACV, pages 3711–3721, 2022. 6, 7", + "[21] Yeong-Joon Ju, Gun-Hee Lee, Jung-Ho Hong, and Seong-Whan Lee. Complete face recovery gan: Unsupervised joint face rotation and de-occlusion from a single-view image. In WACV, pages 3711–3721, 2022. 2", + "[22] Brendan F Klare, Ben Klein, Emma Taborsky, Austin Blanton, Jordan Cheney, Kristen Allen, Patrick Grother, Alan Mah, and Anil K Jain. Pushing the frontiers of unconstrained face detection and recognition: Iarpa janus benchmark a. In CVPR, pages 1931-1939, 2015. 5", + "[23] Bo Li, Kaitao Xue, Bin Liu, and Yu-Kun Lai. Bbddm: Image-to-image translation with brownian bridge diffusion models. In CVPR, pages 1952-1961, 2023. 2, 3", + "[24] Chenyu Li, Shiming Ge, Daichi Zhang, and Jia Li. Look through masks: Towards masked face recognition with deocclusion distillation. In ACM MM, pages 3016-3024, 2020. 2", + "[25] Stan Li, Dong Yi, Zhen Lei, and Shengcai Liao. The casia mir-vis 2.0 face database. In CVPRW, pages 348-353, 2013. 5", + "[26] Changcheng Liang, Mingrui Zhu, Nannan Wang, Heng Yang, and Xinbo Gao. Pmsgan: Parallel multistage gans for face image translation. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2", + "[27] Xiao Luan, Jiezhong Zheng, and Weisheng Li. Learning unsupervised face normalization through frontal view reconstruction. IEEE Transactions on Circuits and Systems for Video Technology, 32(8):5201-5212, 2021. 2" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29307", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Bingjun Luo, Zewen Wang, Jinpeng Wang, Junjie Zhu, Xibin Zhao, and Yue Gao. Multi-energy guided image translation with stochastic differential equations for near-infrared facial expression recognition. In AAAI, pages 565-573, 2024. 2, 3", + "[29] Mandi Luo, Haoxue Wu, Huaibo Huang, Weizan He, and Ran He. Memory-modulated transformer network for heterogeneous face recognition. IEEE Transactions on Information Forensics and Security, 17:2095-2109, 2022. 7, 8", + "[30] Joel Ruben Antony Moniz, Christopher Beckham, Simon Rajotte, Sina Honari, and Chris Pal. Unsupervised depth estimation, 3d face rotation and replacement. Advances in neural information processing systems, 31, 2018. 6", + "[31] Meng Pang, Yiu-Ming Cheung, Qiquan Shi, and Mengke Li. Iterative dynamic generic learning for face recognition from a contaminated single-sample per person. IEEE Transactions on Neural Networks and Learning Systems, 32(4): 1560-1574, 2020. 2", + "[32] Meng Pang, Binghui Wang, Yiu-ming Cheung, Yiran Chen, and Bihan Wen. Vd-gan: A unified framework for joint prototype and representation learning from contaminated single sample per person. IEEE Transactions on Information Forensics and Security, 16:2246-2259, 2021. 1, 2", + "[33] Meng Pang, Binghui Wang, Mang Ye, Yiu-ming Cheung, Yi ran Chen, and Bihan Wen. Disp+ v: A unified framework for disentangling prototype and variation from single sample per person. IEEE Transactions on Neural Networks and Learning Systems, 34(2):867-881, 2021. 2", + "[34] Meng Pang, Binghui Wang, Siyu Huang, Yiu-Ming Cheung, and Bihan Wen. A unified framework for bidirectional prototype learning from contaminated faces across heterogeneous domains. IEEE Transactions on Information Forensics and Security, 17:1544-1557, 2022. 1", + "[35] Yichen Qian, Weihong Deng, and Jiani Hu. Unsupervised face normalization with extreme pose and expression in the wild. In CVPR, pages 9851-9858, 2019. 6", + "[36] Yue Que, Li Xiong, Weiguo Wan, Xue Xia, and Zhiwei Liu. Denoising diffusion probabilistic model for face sketch-to-photo synthesis. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2", + "[37] Lingxiao Song, Zhihe Lu, Ran He, Zhenan Sun, and Tieniu Tan. Geometry guided adversarial facial expression synthesis. In ACM MM, pages 627-635, 2018. 2", + "[38] Luan Tran, Xi Yin, and Xiaoming Liu. Disentangled representation learning gan for pose-invariant face recognition. In CVPR, pages 1415-1424, 2017. 6", + "[39] Jingwei Xin, Zikai Wei, Nannan Wang, Jie Li, Xiaoyu Wang, and Xinbo Gao. Learning a high fidelity identity representation for face frontalization. IEEE Transactions on Circuits and Systems for Video Technology, 33(11):6952-6964, 2023. 1", + "[40] Jianjin Xu, Saman Motamed, Praneetha Vaddamanu, Chen Henry Wu, Christian Haene, Jean-Charles Bazin, and Fernando De la Torre. Personalized face inpainting with diffusion models by parallel visual attention. In WACV, pages 5432–5442, 2024. 1, 2", + "[41] Weihong Xu, Xiaohua Xie, and Jianhuang Lai. Relightgan: Instance-level generative adversarial network for face illumni" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "nation transfer. IEEE Transactions on Image Processing, 30: 3450-3460, 2021. 1, 2", + "[42] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4): 1-39, 2023. 2", + "[43] Yiming Yang, Weipeng Hu, and Haifeng Hu. Neutral face learning and progressive fusion synthesis network for nir-vis face recognition. IEEE Transactions on Circuits and Systems for Video Technology, 33(10):5750-5763, 2024. 6, 7, 8", + "[44] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? NeurIPS, 27, 2014. 4", + "[45] Aijing Yu, Haoxue Wu, Huaibo Huang, Zhen Lei, and Ran He. Lamp-hq: A large-scale multi-pose high-quality database and benchmark for nir-vis face recognition. International Journal of Computer Vision, 129(5):1467-1483, 2021. 5", + "[46] Ni Zhang, Nian Liu, Junwei Han, Kaiyuan Wan, and Ling Shao. Face de-occlusion with deep cascade guidance learning. IEEE Transactions on Multimedia, 25:3217-3229, 2022. 1, 2", + "[47] Jian Zhao, Yu Cheng, Yan Xu, Lin Xiong, Jianshu Li, Fang Zhao, Karlekar Jayashree, Sugiri Pranata, Shengmei Shen, Junliang Xing, et al. Towards pose invariant face recognition in the wild. In CVPR, pages 2207-2216, 2018. 6", + "[48] Hang Zhou, Jihao Liu, Ziwei Liu, Yu Liu, and Xiaogang Wang. Rotate-and-render: Unsupervised photorealistic face rotation from single-view images. In CVPR, pages 5911-5920, 2020. 6, 7", + "[49] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, pages 2223-2232, 2017. 2" + ], + "bbox": [ + 516, + 92, + 903, + 584 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29308", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/UMFN_ Unified Multi-Domain Face Normalization for Joint Cross-domain Prototype Learning and Heterogeneous Face Recognition/3e3f9780-5e1d-4d68-aee1-68dcb88c5368_model.json b/2025/UMFN_ Unified Multi-Domain Face Normalization for Joint Cross-domain Prototype Learning and Heterogeneous Face Recognition/3e3f9780-5e1d-4d68-aee1-68dcb88c5368_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0e199473c212314f111dbb90f876de71a9363852 --- /dev/null +++ b/2025/UMFN_ Unified Multi-Domain Face Normalization for Joint Cross-domain Prototype Learning and Heterogeneous Face Recognition/3e3f9780-5e1d-4d68-aee1-68dcb88c5368_model.json @@ -0,0 +1,2101 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.125, + 0.13, + 0.877, + 0.177 + ], + "angle": 0, + "content": "UMFN: Unified Multi-Domain Face Normalization for Joint Cross-domain Prototype Learning and Heterogeneous Face Recognition" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.203, + 0.797, + 0.257 + ], + "angle": 0, + "content": "Meng Pang1, Wenjun Zhang1, Nanrun Zhou2*, Shengbo Chen1*, Hong Rao1 \n1Nanchang University \n2Shanghai University of Engineering Science" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.26, + 0.871, + 0.275 + ], + "angle": 0, + "content": "{mengpang@, zhangwenjun@, raohong@}ncu.edu.cn, nrzhou@sues.edu.cn, ccb02kingdom@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.31, + 0.327, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.485, + 0.705 + ], + "angle": 0, + "content": "Face normalization aims to enhance the robustness and effectiveness of face recognition systems by mitigating intra-personal variations in expressions, poses, occlusions, illuminations, and domains. Existing methods face limitations in handling multiple variations and adapting to cross-domain scenarios. To address these challenges, we propose a novel Unified Multi-Domain Face Normalization Network (UMFN) model, which can process face images with various types of facial variations from different domains, and reconstruct frontal, neutral-expression facial prototypes in the target domain. As an unsupervised domain adaptation model, UMFN facilitates concurrent training on multiple datasets across domains and demonstrates strong prototype reconstruction capabilities. Notably, UMFN serves as a joint prototype and feature learning framework, enabling the simultaneous extraction of domain-agnostic identity features through a decoupling mapping network and a feature domain classifier for adversarial training. Moreover, we design an efficient Heterogeneous Face Recognition (HFR) network that fuses domain-agnostic and identity-discriminative features for HFR, and introduce contrastive learning to enhance identity recognition accuracy. Empirical studies on diverse cross-domain face datasets validate the effectiveness of our proposed method." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.748, + 0.222, + 0.763 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.483, + 0.879 + ], + "angle": 0, + "content": "Face normalization, as a core technology in the field of computer vision, is dedicated to eliminating interfering variation factors in facial images, such as pose, expression, occlusion, and illumination, and reconstructing a frontal, unoccluded, evenly illuminated facial prototype with a neutral expression [34]. This technology has shown broad application prospects in numerous areas, including criminal investiga" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.31, + 0.9, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.57, + 0.907, + 0.613 + ], + "angle": 0, + "content": "Figure 1. Illustration of a desired face normalization method capable of handling multiple cross-domain face normalization tasks and addressing various types of facial variations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.907, + 0.716 + ], + "angle": 0, + "content": "tion , authentication, and human-computer interaction [39]. However, existing face normalization methods face challenges in coping with increasingly complex and diversified application scenarios, highlighting the urgent need for more advanced and versatile solutions to flexibly adapt to various changes in different environments and conditions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The limitations of current face normalization methods lie mainly in their constraints in processing single contaminant elements and their single-domain applicability. Most methods [3, 10] focus on modeling and processing a single element, such as pose, expression, illumination, or occlusion, and struggle with multiple coexisting factors. Additionally, some unsupervised methods [41, 46] may inadvertently lose crucial identity information, leading to identity inconsistency between the generated prototype and the input face. Furthermore, existing models [32, 40] are typically domain-specific, lacking universality across multi-domain scenarios. However, practical applications often involve images" + }, + { + "type": "page_footnote", + "bbox": [ + 0.119, + 0.888, + 0.251, + 0.901 + ], + "angle": 0, + "content": "*Corresponding authors" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29299" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "from various domains, such as visible light, near-infrared, and sketch images used in criminal investigations. Therefore, there is an urgent need for a multi-domain face normalization method capable of handling various facial variations, as illustrated in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.17, + 0.483, + 0.382 + ], + "angle": 0, + "content": "To address these challenges, we propose a Unified Multidomain Face Normalization Network (UMFN) model. This model is capable of handling facial images from different domains and containing multiple contaminant elements (such as pose, expression, occlusion, etc.), and can reconstruct a frontal, neutral facial prototype in the target domain while simultaneously extracting identity features unrelated to domain information. It is worth noting that, despite the significant achievements of diffusion models in the field of image generation, their limitations in interpretable representation learning make them difficult to meet our needs for joint prototype and feature learning [42]. Therefore, this paper adopts a fusion framework of Autoencoder and Generative Adversarial Network (GAN), as shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.384, + 0.483, + 0.685 + ], + "angle": 0, + "content": "The UMFN model introduces a feature decoupling network and domain classifier for adversarial training, decoupling domain-agnostic features. By leveraging concatenated masks, it regulates clean prototype generation within specified domains or reconstructs original images, ensuring identity consistency. Additionally, a multi-scale InstanceNorm Patch-based discriminator [19] enhances image details and textures, improving visual quality. As an unsupervised domain adaptation model, UMFN enables concurrent training of diverse, unlabeled datasets within a single network, demonstrating flexibility and practicality in cross-domain prototype reconstruction. Furthermore, we design an efficient heterogeneous face recognition (HFR) network in conjunction with UMFN's encoder to extract identity-discriminative features for the generated prototype. These features are fused with domain-agnostic features of input for HFR. During HFR network training, a contrastive learning mechanism minimizes the feature distance between generated and authentic prototypes, enhancing the network's ability to accurately recognize identities." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.687, + 0.468, + 0.702 + ], + "angle": 0, + "content": "The contributions of the paper are summarized below:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.705, + 0.483, + 0.795 + ], + "angle": 0, + "content": "- We propose a unified face normalization model, dubbed UMFN, which addresses the limitations of existing methods in handling multiple variation factors and cross-domain face images. This model is capable of simultaneously processing various types of facial variations and is applicable to face images from different domains." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.796, + 0.483, + 0.886 + ], + "angle": 0, + "content": "- We design an efficient HFR network that fuses domain-agnostic features from input sample with identity-discriminative features from generated prototype for HFR. Furthermore, a contrastive learning mechanism is introduced to bolster the network's capability in accurately recognizing identities." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- As an unsupervised domain adaptation model, UMFN en" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.181 + ], + "angle": 0, + "content": "ables simultaneous training of multiple datasets with different domains and without identity labels in a single network. It is able to perform prototype reconstruction tasks across multiple domains, enhancing model flexibility and practicality, adapting to diverse scenarios, and reducing reliance on labeled data." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.183, + 0.905, + 0.242 + ], + "angle": 0, + "content": "- Empirical study on multiple cross-domain face datasets with facial variations validates the superiority of UMFN in face normalization tasks in both single-modal and cross-modal scenarios, as well as in HFR task." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.257, + 0.663, + 0.273 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.282, + 0.905, + 0.508 + ], + "angle": 0, + "content": "Face normalization, an emerging and highly esteemed research field, has shown substantial practical value in areas like criminal identification and facial evidence collection. Mainstream face normalization methods primarily address facial variation removal in single-modal scenarios, aiming to eliminate factors such as diverse lighting conditions, exaggerated facial expressions, large head poses, or partially occlusions, thereby restoring the standard facial prototype. Existing single-modal face normalization algorithms can be broadly classified into two categories [33]: The first constructs clustering models based on auxiliary sample information to estimate facial prototypes of contaminated samples. The second leverages generative AI models, such as GANs [8] and Diffusion Models [4], to learn the mapping between contaminated and standard facial samples." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.509, + 0.905, + 0.854 + ], + "angle": 0, + "content": "In the first category, Gao et al. [7] combined unlabeled facial samples with a labeled training set, using a Gaussian Mixture Model for clustering to estimate prototypes. Pang et al. [31] introduced a more robust semi-supervised low-rank representation model for clustering and prototype estimation, updated with labeled retrieval samples. Although effective in prototype reconstruction, these methods require prior access to the facial samples, making them unsuitable for real-time scenarios. In the second category, benefiting from generative AI models' image generation and mapping capabilities, researchers have successively proposed a series of generative learning methods, aiming to eliminate expression variations [3, 37], illumination variations [10, 41], pose variations [15, 27, 46], and partial occlusions [21, 24, 40] in input facial samples, respectively, and generate facial prototype images with consistent identities. However, these methods are typically limited to specific types of facial variations and cannot effectively address combinations of multiple types. Pang et al. [32] attempts to solve prototype reconstruction for multiple types of contamination by designing a Variation Disentangling GAN (VD-GAN), but has limitations in preserving identity characteristics and handling only single-domain contamination." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Recently, a few of cross-domain face synthesis methods [1, 23, 26, 28, 36, 49] have been proposed, including Cycle-consistency GAN (CycleGAN) [49], Parallel Multi" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29300" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.904, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.315, + 0.908, + 0.345 + ], + "angle": 0, + "content": "Figure 2. Overview of the architecture of UMFN. \\( x_{c} \\in X_{c} \\) denotes an input image in domain \\( c \\), \\( \\hat{x}_c \\) represents the reconstructed image of \\( x_{c} \\), \\( \\hat{x}_{\\bar{c}}^{rp} \\) denotes the generated prototype of \\( x_{c} \\) in domain \\( \\bar{c} \\), and \\( x_{c}^{rp} \\) represents the real prototype in domain \\( c \\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.362, + 0.484, + 0.513 + ], + "angle": 0, + "content": "stage GANs (PMSGAN) [1], Brownian Bridge Diffusion Models (BBDM) [23], and NIR-FER Stochastic Differential Equations (NFER-SDE) [28]. These methods focus on transforming domain styles while preserving facial details, but they cannot remove facial variations from in-put faces. In light of the above, this paper aims to design a multidomain face normalization scheme applicable to universal facial variation removal and explore effective integration and optimization of domain transformation and prototype learning within a unified framework." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.527, + 0.192, + 0.543 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.553, + 0.484, + 0.734 + ], + "angle": 0, + "content": "This paper proposes a universal face normalization framework, UMFN, whose architecture is shown in Fig. 2. UMFN is designed to uniformly process face images with various types of facial variations from different image domains, ultimately generating standard facial prototypes in the specified domain. Correspondingly, UMFN comprises the following two core components: 1) addressing the issue of input images originating from multiple domains and generating images in the specified domain; 2) uniformly handling multiple types of facial variations while generating prototypes that preserve the identity (ID). These two components will be detailed in Sec. 3.1 and Sec. 3.2" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.743, + 0.483, + 0.773 + ], + "angle": 0, + "content": "3.1. Feature Decoupling and Cross-Domain Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.87 + ], + "angle": 0, + "content": "To tackle the challenge that input face images may come from different domains, we have designed a feature decoupling mapping network. This network is able to separate domain information from the latent space, resulting in domain-invariant representations. Such decoupled, domain-agnostic identity features can be applied to HFR task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Feature Decoupling: We design a feature decoupling mapping network that initially extracts primary feature \\( z = \\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.363, + 0.907, + 0.545 + ], + "angle": 0, + "content": "\\(G_{enc}(x_c)\\) from the input image \\(x_c\\) using a pre-trained encoder \\(G_{enc}\\), and then derives its latent feature representation \\(p = \\text{map}(z)\\) through the mapping network map. Specifically, the domain feature classifier \\(D_c\\) receives the latent representations \\(p\\) as input and outputs the domain prediction code (i.e., the image domain to which \\(p\\) belongs). The goal of \\(D_c\\) is to accurately predict the domain of the input latent representation \\(p\\), while the mapping network map aims to confuse \\(D_c\\)'s correct domain prediction. Through adversarial training, the interplay between the two achieves domain information decoupling. Accordingly, the objective function for training \\(D_c\\) is presented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.556, + 0.907, + 0.575 + ], + "angle": 0, + "content": "\\[\nL _ {D _ {c}} = E \\left[ \\log P \\left(l _ {p} ^ {\\text {p r e}} = l _ {x _ {c}} ^ {\\text {r e a l}} \\mid p\\right) \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.587, + 0.906, + 0.634 + ], + "angle": 0, + "content": "where \\( l_{p}^{pre} \\) is the domain label predicted by \\( D_{c} \\), \\( l_{x_{c}}^{real} \\) is the real domain label of \\( x_{c} \\). On the other side, the objective function for training the mapping network map is as" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.645, + 0.907, + 0.665 + ], + "angle": 0, + "content": "\\[\nL _ {m a p} = - E \\left[ \\log P \\left(l _ {p} ^ {p r e} = l _ {x _ {c}} ^ {r e a l} \\mid p\\right) \\right]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.676, + 0.907, + 0.872 + ], + "angle": 0, + "content": "Cross-domain Generation: We introduce the control mask and combine it with reconstruction learning to achieve cross-domain image generation. The core objective of reconstruction learning is to guide the decoder \\( G_{dec} \\) to learn how to decode domain-agnostic latent representation \\( p = \\text{map}(G_{enc}(x_c)) \\) conditioned on the original attribute mask \\( v_c \\), thereby reconstructing the input image \\( x_c \\). This process ensures that the decoder not only learns to understand and apply the semantics of the mask but also fully utilizes the identity feature information in the latent representation \\( p \\). As such, in subsequent steps, images from specified domains can be generated by manipulating the mask. The objective function for reconstruction learning is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.885, + 0.907, + 0.903 + ], + "angle": 0, + "content": "\\[\nL _ {r e c} = E _ {x _ {c} \\sim p _ {d a t a}} \\left[ \\| x _ {c} - \\hat {x} _ {c} \\| _ {1} \\right], \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29301" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.213 + ], + "angle": 0, + "content": "where \\(\\hat{x}_c = G_{dec}(p, v_c)\\) is the reconstructed image of \\(x_c\\), \\(v_c\\) is the original attribute mask of the input image \\(x_c\\), indicating whether the \\(x_c\\) image is polluted and its domain information. Specifically, \\(v_c\\) is composed of a four-digit one-hot vector, where the first digit indicates whether the image is contaminated, and the last three digits represent the domain it belongs to. In Eq. (3), \\(l_1\\) norm is employed instead of the \\(l_2\\) norm to suppress the issue of blurred generated images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.221, + 0.291, + 0.236 + ], + "angle": 0, + "content": "3.2. Prototype generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.242, + 0.484, + 0.362 + ], + "angle": 0, + "content": "Different from prior single-contamination normalization methods that overly focus on facial variation categories and how to accurately model them, our method returns to the essence of the prototype reconstruction task, aiming to directly generate identity-preserved standard facial prototypes by incorporating identity information from the contaminated facial samples, potentially bypassing the challenges of modeling facial variations in a clever manner." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.363, + 0.484, + 0.635 + ], + "angle": 0, + "content": "Notably, we introduce a prototype adversarial loss, differing from the adversarial loss in traditional GAN-based image translation methods that aim to approximate the distribution of original face data with diverse variations. Our loss aims to approximate a condensed distribution of standard facial prototypes, while also reducing overfitting to these variations. Furthermore, we have designed a multiscale Markov discriminator \\( D = [D_w, D_f] \\) to refine the details of generated faces and stabilize training. Specifically, we concatenate the extracted domain-agnostic identity feature \\( p = \\text{map}(G_{\\text{enc}}(x_c)) \\) with a control mask \\( v_{\\bar{c}} \\), where the first bit in \\( v_{\\bar{c}} \\) controls whether to perform face normalization: 1 for yes, and the last three bits control the target domain for prototype generation. Next, this concatenated feature is input into the decoder to generate a pixel-level standardized facial prototype in the specified target domain. The prototype adversarial objectives for training the generator and the multi-scale discriminator are as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.639, + 0.484, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {D} ^ {\\text {a d v}} = D _ {w} \\left(\\hat {x} _ {\\bar {c}} ^ {r p}\\right) + D _ {f} \\left(\\operatorname {c r o p} \\left(\\hat {x} _ {\\bar {c}} ^ {r p}\\right)\\right) \\tag {4} \\\\ - D _ {w} \\left(x _ {c} ^ {r p}\\right) - D _ {f} \\left(\\operatorname {c r o p} \\left(x _ {c} ^ {r p}\\right)\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.683, + 0.483, + 0.702 + ], + "angle": 0, + "content": "\\[\nL _ {G _ {d e c}} ^ {a d v} = - \\left(D _ {w} \\left(\\hat {x} _ {\\bar {c}} ^ {r p}\\right) + D _ {f} \\left(\\operatorname {c r o p} \\left(\\hat {x} _ {\\bar {c}} ^ {r p}\\right)\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.87 + ], + "angle": 0, + "content": "where \\(D_w\\) and \\(D_f\\) represent the global path discriminator and local path discriminator, respectively. \\(\\hat{x}_{\\bar{c}}^{rp} = G_{dec}(map(G_{enc}(x_c), v_{\\bar{c}})\\) denotes the specified domain prototype generated by inputting \\(x_c\\) and the control mask \\(v_{\\bar{c}}\\) while \\(x_c^{rp}\\) represents the uncontaminated faces in the training set. The function \"crop\" is used to cut out the facial features from the input face images. Considering that the Wasserstein distance is more stable and can avoid training instability and mode collapse in GANs, we adopt the WGAN-GP [9] loss in our model to replace the traditional cross-entropy loss." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Furthermore, to ensure that the generated prototype \\(\\hat{x}_{\\bar{c}}^{rp}\\) can effectively preserve the identity information from the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "input \\( x_{c} \\), we have designed an identity feature similarity loss function. This is achieved by feeding the generated prototype back into the encoder and mapping network to obtain the latent feature map \\( (G_{enc}(\\hat{x_c}^{rp})) \\), and then calculating its similarity to the latent feature \\( p \\) of the input \\( x_{c} \\), thereby retaining identity. The loss function is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.197, + 0.905, + 0.216 + ], + "angle": 0, + "content": "\\[\nL _ {i d} = \\| p - \\operatorname {m a p} \\left(G _ {e n c} \\left(\\hat {x} _ {\\bar {c}} ^ {r p}\\right)\\right) \\| _ {2} ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.223, + 0.907, + 0.359 + ], + "angle": 0, + "content": "where \\( \\| \\cdot \\|_2^2 \\) denotes \\( l_2 \\)-norm. On the other hand, UMFN transfers the identity feature retention ability from the reconstruction learning task to the face prototype generation task. Since these two tasks share the same input and output domains, the transferability gap [44] between them is minimal. Therefore, the identity feature retention ability learned from the reconstruction learning task can be easily transferred to the prototype reconstruction task. Therefore, the overall loss function of the decoder \\( G_{dec} \\) is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.371, + 0.905, + 0.39 + ], + "angle": 0, + "content": "\\[\nL _ {G _ {d e c}} ^ {\\text {a d v}} = \\lambda_ {1} L _ {r e c} + \\lambda_ {2} L _ {i d}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.402, + 0.836, + 0.417 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are balancing hyperparameters." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.428, + 0.8, + 0.444 + ], + "angle": 0, + "content": "3.3. Heterogeneous Face Recognition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.906, + 0.526 + ], + "angle": 0, + "content": "At this stage, we meticulously construct a HFR network, whose core components encompass an encoder \\(F_{\\theta}\\) (its network architecture aligns with \\(G_{enc}\\) in UMFN) and a fully connected layer \\(l_{\\theta}\\). The specific structure is shown in Fig. 3. The HFR network consists of two core operations:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.527, + 0.906, + 0.678 + ], + "angle": 0, + "content": "Feature Fusion: During training, we concatenate the domain-agnostic features (extracted by UMFN's \\( G_{enc} \\) and mapping network map from the source domain input image) with the identity-discriminative features (extracted by encoder \\( F_{\\theta} \\) from the real target domain prototype). The fused features are then fed into the softmax layer for classification. In this process, the fully connected layer \\( l_{\\theta} \\) doubles the feature dimensions, enhancing classification robustness through multi-dimensional identity information integration. The classification loss for this step is presented as" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.69, + 0.905, + 0.708 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {c l s}} = \\operatorname {s o f t m a x} \\left(l _ {\\theta} \\left[ F _ {\\theta} \\left(x _ {v, i} ^ {r p}\\right) \\oplus \\operatorname {m a p} \\left(G _ {e n c} \\left(x _ {n, i}\\right)\\right) \\right], i\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.87 + ], + "angle": 0, + "content": "where \\(x_{n,i}\\) represents source domain image and \\(x_{v,i}^{rp}\\) represents real target domain prototype, \\(F_{\\theta}(x_{v,i}^{rp})\\) denotes the representation of the input image \\(x_{v,i}^{rp}\\) extracted by \\(F_{\\theta}\\), while map \\((G_{enc}(x_{n,i}))\\) denotes the domain-agnostic feature of the input image \\(x_{n,i}\\), the symbol \\(\\oplus\\) indicates the concatenation of these two features. During testing, we extract the identity-discriminative feature of the generated target domain prototype from the input test source domain image using the trained \\(F_{\\theta}\\), and fuse it with the domain-agnostic feature of the input test source domain image for HFR." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Contrastive Learning Mechanism: During the cross-domain generation, generated target domain prototypes may" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "29302" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.09, + 0.468, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.406, + 0.471, + 0.421 + ], + "angle": 0, + "content": "Figure 3. Overview of the HFR network and training process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.484, + 0.697 + ], + "angle": 0, + "content": "suffer identity feature loss or noise introduction. Direct reliance on such data for identity classification training can lead to overfitting to missing or noisy features, impairing classification performance. To mitigate this, we innovatively employ a contrastive learning mechanism that aligns generated prototypes with real ones in the feature space, while reducing the impact of noise and missing features. This ensures that the encoder \\( F_{\\theta} \\) extracts features closely resembling real prototypes. Specifically, we randomly sample two pairs of real heterogeneous images, \\( (x_{v,1}^{rp}, x_{n,1}) \\) and \\( (x_{v,2}^{rp}, x_{n,2}) \\), where \\( x_{n,1} \\) and \\( x_{n,2} \\) are NIR images with contamination. We then use UMFN to perform cross-domain restoration on \\( x_{n,1} \\) and \\( x_{n,2} \\), generating the corresponding target VIS domain prototypes \\( \\hat{x}_{v,1}^{rp} \\) and \\( \\hat{x}_{v,2}^{rp} \\). Next, we form positive pairs \\( (x_{v,1}^{rp}, \\hat{x}_{v,1}^{rp}) \\) and \\( (x_{v,2}^{rp}, \\hat{x}_{v,2}^{rp}) \\), and negative pairs \\( (x_{v,1}^{rp}, \\hat{x}_{v,2}^{rp}) \\) and \\( (x_{v,2}^{rp}, \\hat{x}_{v,1}^{rp}) \\). The corresponding contrastive loss is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.709, + 0.482, + 0.777 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {c o n t} = \\sum_ {j \\neq k} \\left(1 - \\langle f _ {j}, \\tilde {f} _ {j} \\rangle\\right) + \\left(1 - \\langle f _ {k}, \\tilde {f} _ {k} \\rangle\\right) \\\\ + \\max \\left(0, \\left\\langle f _ {j}, \\tilde {f} _ {k} \\right\\rangle - m\\right) + \\max \\left(0, \\left\\langle f _ {k}, \\tilde {f} _ {j} \\right\\rangle - m\\right), \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.777, + 0.483, + 0.839 + ], + "angle": 0, + "content": "where \\( f_{j} = l_{\\theta}[F_{\\theta}(x_{v,j}^{rp}) \\oplus map(G_{enc}(x_{n,j}))] \\), \\( \\langle, \\rangle \\) represents the calculation of the cosine similarity between two features, and \\( m \\) is a margin value set at 0.5. In summary, the final loss function of the HFR network is expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.207, + 0.854, + 0.483, + 0.871 + ], + "angle": 0, + "content": "\\[\nL _ {h f r} = L _ {c l s} + \\alpha L _ {\\text {c o n t}}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.31, + 0.901 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a trade-off parameter." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.646, + 0.107 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.116, + 0.715, + 0.132 + ], + "angle": 0, + "content": "4.1. Datasets Descriptions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.907, + 0.336 + ], + "angle": 0, + "content": "LAMP-HQ [45] is a newly created, challenging large-scale NIR-VIS dataset comprising over 56,000 NIR and 16,000 VIS face images from 573 individuals, featuring diverse expressions, poses, lighting conditions, occlusions, and complex backgrounds. BUAA NIR-VIS [16] includes 150 identities, each with 9 NIR and 9 VIS images, covering various head poses and expressions. CUFSF [6] consists of 597 grayscale images paired with 597 real sketch images. CASIA NIR-VIS 2.0 [25] encompasses 725 identities, with 5-50 NIR and 1-22 VIS images per identity. IJB-A [22] and LFW [17] are widely-used benchmarks for face recognition and verification, featuring diverse poses, occlusions, and expressions in the visible domain." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.347, + 0.704, + 0.364 + ], + "angle": 0, + "content": "4.2. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.37, + 0.906, + 0.46 + ], + "angle": 0, + "content": "In the experimental section, we primarily conduct four experiments: (1) Homogeneous Multi-domain Face Normalization Experiment, (2) Heterogeneous Multi-domain Face Normalization Experiment, (3) Heterogeneous Face Recognition Experiment. (4) Ablation Study. The specific configurations of these experiments are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.907, + 0.734 + ], + "angle": 0, + "content": "1) Homogeneous Multi-domain Face Normalization Experiment: In this experiment, we utilize three types of domain data from the LAMP-HQ dataset to train our model, ensuring that all data originate from the same set of identities (IDs). The experiment primarily evaluates the model's performance in facial normalization across multi-domain data within the LAMP-HQ dataset. Specifically, we employ FP data from the LAMP-HQ dataset, where all NIR and VIS images from 100 IDs are used as the training set, while the remaining IDs are designated as the test set. As the LAMP-HQ dataset only contains NIR and VIS domain data, we use software to convert NIR photographs into sketch images, thereby simulating a scenario with three types of domain data. Furthermore, we transfer the model to other datasets, namely LFW and IJB-A, for testing. We assess the model's performance from both qualitative and quantitative perspectives and compare it with the latest single-modal face normalization techniques." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.906, + 0.87 + ], + "angle": 0, + "content": "2) Heterogeneous Multi-domain Face Normalization Experiment: We use one type of domain data from three different datasets to demonstrate that the model can effectively work without paired data training, verifying its generalization ability across datasets. Specifically, we use VIS images of 250 IDs from the LAMP-HQ dataset as VIS domain data, all NIR images of 130 IDs from the BUAA dataset as NIR domain data, and Sketch images of 400 IDs from the CUFSF dataset as Sketch domain data for training." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "3) Heterogeneous Face Recognition Experiment: We conduct heterogeneous face recognition experiments on the" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.462, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29303" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.095, + 0.905, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.248, + 0.716, + 0.262 + ], + "angle": 0, + "content": "Figure 4. In-domain face normalization results on the LAMP-HQ dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.278, + 0.483, + 0.308 + ], + "angle": 0, + "content": "LAMP-HQ and CASIA NIR-VIS 2.0 datasets, following the standard evaluation protocol in [43]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.31, + 0.483, + 0.369 + ], + "angle": 0, + "content": "4) Ablation Study Experiment: We investigate the contribution of each component within the UMFN framework and explore the roles of the contrastive learning mechanism and feature fusion operation in the HFR network." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.378, + 0.483, + 0.394 + ], + "angle": 0, + "content": "4.3. Implementation Details and Parameter Setting" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.4, + 0.483, + 0.581 + ], + "angle": 0, + "content": "We employ a Resnet-50 pre-trained on the MS1M dataset and fine-tuned on VGGFace2 [2] as the encoder \\( G_{enc} \\) to extract 2048-dimensional identity features, keeping \\( G_{enc} \\) fixed during training and testing. The mapping network consists of multiple linear layers, inputting a 2048-dimensional vector and outputting a 2048-dimensional vector. The feature domain classifier \\( D_{c} \\) is composed of linear layers and outputs the predicted domain category, while the decoder \\( G_{dec} \\) outputs images of size \\( 224 \\times 224 \\). The discriminators \\( D_{W} \\) and \\( D_{f} \\) are both patchGAN discriminators with a receptive field of 70, using instance normalization as the normalization layer, and the crop size is \\( 150 \\times 150 \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.688 + ], + "angle": 0, + "content": "The hyperparameters \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) in Eq. (7) are set at \\(\\lambda_{1} = 10 - \\frac{cur_{epoch}}{sum_{epoch}} \\times 9\\) and \\(\\lambda_{2} = 10\\), where \\(cur_{epoch}\\) is the current number of training epochs and \\(sum_{epoch}\\) is the total number of training epochs, which is 100. \\(\\alpha\\) in Eq. (10) is empirically set at 0.001. During optimization, we use the Adam optimizer, with the learning rate set at 0.0002 and momentum set at 0.5." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.698, + 0.462, + 0.714 + ], + "angle": 0, + "content": "4.4. Homogeneous Multi-domain Normalization" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this section, we evaluate the performance of UMFN in homogeneous multi-domain face normalization using the LAMP-HQ dataset. The in-domain normalization results of UMFN are presented in Fig. 4. Fig. 4(a) showcases the results of inputting VIS domain faces corrupted by different facial variations such as pose, lighting, occlusion, and expression, and outputting normalized VIS domain faces. Fig. 4(b) displays the outcomes of inputting corrupted faces from the NIR and Sketch domains and outputting normalized faces in their respective domains. The bottom row serves as a reference for the true prototypes. From Fig. 4, it can be observed that UMFN is capable of simultaneously" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.279, + 0.905, + 0.461 + ], + "angle": 0, + "content": "normalizing various types of corrupted faces from multiple different domains within their respective domains, effectively preserving the identity of the input faces while obtaining facial prototypes. Additionally, we conduct a comparison between UMFN and other popular single-domain face normalization methods, including PIM [47], TP-GAN [18], DR-GAN [38], Hassner [11], and FNM [35], using contaminated face images from the LFW dataset as input. The face normalization results of each method are presented in Fig. 5. It can be observed that the target prototype obtained through UMFN normalization excels in both image quality and identity preservation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.906, + 0.779 + ], + "angle": 0, + "content": "Moreover, following the protocols in [48] and [20], we quantitatively evaluate UMFN on the LFW and IJB-A datasets. By inputting restored faces into the ResNet-18 [12] network for face verification, we aim to assess whether UMFN-restored faces enhance VR performance. Additionally, we report the Symmetry Index (SI)[30] to provide a comprehensive comparison of the quality of prototypes generated by UMFN, R&R [48], and CFR [20]. Tab. 1 presents the verification performance and SI of UMFN across both datasets. The results indicate that UMFN normalization significantly boosts the verification performance of the ResNet model: on IJB-A, when combined with ResNet-18, it improves VR by \\(5.35\\%\\), with accuracy gains also observed on LFW. The experiments demonstrate that: (1) removing pose, expression, occlusion, and illumination variations from input faces allows the model to better focus on key identity features, and (2) UMFN effectively retains critical identity information while eliminating these facial variations. In addition, UMFN outperforms R&R and CFR in terms of SI, generating highly symmetrical faces that correct facial distortions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Furthermore, we illustrate the cross-domain normalization results of the UMFN model in Fig. 6. Specifically, the inputs consist of contaminated facial images from domain A, while the outputs represent the generated prototypes in domains B and C. The bottom row displays the reference prototypes of the inputs in the VIS domain. It can be observed that the UMFN model successfully addresses the challenges of domain transfer and prototype learning," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29304" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.481, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.194, + 0.483, + 0.222 + ], + "angle": 0, + "content": "Figure 5. Comparison with other single-domain face normalization methods on LFW dataset." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.233, + 0.48, + 0.349 + ], + "angle": 0, + "content": "
MethodDatasets
LFWIJB-A
VR↑SI↓VR↑SI↓
Res18 [12]98.85-90.57±1.2-
Res18+R&R [48]98.9520.9391.98±0.720.85
Res18+CFR [20]99.2328.3193.36±0.624.76
Res18+UMFN(ours)99.5714.6195.92±0.715.81
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.363, + 0.483, + 0.391 + ], + "angle": 0, + "content": "Table 1. Verification performance and Symmetry Index (SI) on LFW and IJB-A." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.401, + 0.484, + 0.522 + ], + "angle": 0, + "content": "effectively preserving the identity characteristics of the inputs when generating uncontaminated facial prototypes for the target domains. The inspiring results demonstrate that UMFN is capable of performing effective domain-specific face normalization on facial images sourced from different domains and contaminated by various types of facial variations, all while efficiently retaining the key identity information of the inputs, utilizing a single network architecture." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.532, + 0.467, + 0.547 + ], + "angle": 0, + "content": "4.5. Heterogeneous Multi-domain Normalization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this section, we evaluate the performance of UMFN in the heterogeneous multi-domain face normalization experiment. It is particularly noteworthy that there is currently no other face normalization method capable of adequately addressing this complex issue. Fig. 7 showcase the normalization results of images from the LAMP-HQ, BUAA, and CUFSF datasets across different domains (VIS, NIR, Sketch). We observe that regardless of the dataset (or domain) the input comes from, UMFN successfully generates prototypes that are highly matched to the target domain styles provided by another dataset. It effectively eliminates interfering factors such as pose, expression, and occlusion in the input face images while accurately preserving the input identity information. It is worth noting that due to the minimal facial variations in most images of the CUFSF dataset, which generally exhibit simple facial expressions such as smiles, the changes after normalization are primarily manifested in the mouth area. Furthermore, the challenging nature of cross-domain normalization tasks involving sketch images, which contain the least amount of facial detail information, leads to slight local distortions in the generated VIS and NIR prototypes. Despite these challenges, the results demonstrate UMFN's robust multi-" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.092, + 0.888, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.29, + 0.907, + 0.348 + ], + "angle": 0, + "content": "Figure 6. Cross-domain face normalization results on LAMP-HQ dataset. The second and third rows depict face normalization results in different domains from the inputs shown in the first row. The bottom row shows the real VIS prototypes for reference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.367, + 0.907, + 0.412 + ], + "angle": 0, + "content": "domain and multi-type face normalization capabilities, particularly when trained on unpaired datasets from multiple domains, where its performance remains impressive." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.422, + 0.803, + 0.438 + ], + "angle": 0, + "content": "4.6. Heterogeneous Face Recognition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.445, + 0.907, + 0.61 + ], + "angle": 0, + "content": "In this experiment, we employ UMFN to restore cross-domain prototypes from NIR data in the probe set, generating VIS prototypes. Subsequently, we utilize the trained HFR network to obtain fused features (i.e., concatenation of domain-agnostic and identity-discriminative features) for both these generated prototypes and the authentic prototypes in the gallery set. Finally, by computing the cosine similarity between the fused features of the generated prototypes and those of the gallery set, we obtain a similarity matrix, upon which we calculate the rank-1 recognition accuracy and VR at \\(\\mathrm{FAR} = 0.001\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.909, + 0.837 + ], + "angle": 0, + "content": "Tab. 2 presents a comparison of our method with other advanced HFR approaches, including PACH [5], AD-CANs [13], MMTN [29], NLPF [43], and LPL [14], on LAMP-HQ and CASIA NIR-VIS 2.0 datasets. It is observed that our method achieves the best performance on both datasets in terms of rank-1 accuracy and VR. This exceptional performance stems from: (1) Incorporating a contrastive learning mechanism that aligns generated prototypes with real ones in the feature space, reducing noise and missing feature impacts. This allows the HFR encoder \\( F_{\\theta} \\) to extract identity-discriminative features from generated prototypes that closely resembling real ones. (2) The fusion of domain-agnostic and identity-discriminative features notably enhances the identity distinguishability of the combined features." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.667, + 0.865 + ], + "angle": 0, + "content": "4.7. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To assess the effectiveness of UMFN and the contributions of its components, we train five variants by respectively re" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "29305" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.092, + 0.353, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.278, + 0.323, + 0.289 + ], + "angle": 0, + "content": "(a) Results on LAMP-HQ dataset" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.094, + 0.627, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.427, + 0.278, + 0.59, + 0.289 + ], + "angle": 0, + "content": "(b) Results on BUAA dataset" + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.093, + 0.906, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.705, + 0.278, + 0.875, + 0.289 + ], + "angle": 0, + "content": "(c) Results on CUFSF dataset" + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.296, + 0.873, + 0.311 + ], + "angle": 0, + "content": "Figure 7. The normalization results of face images from the LAMP-HQ, BUAA, and CUFSF datasets across different domains." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.327, + 0.48, + 0.449 + ], + "angle": 0, + "content": "
MethodDatasets
LAMP-HQCASIANIR-VIS 2.0
Rank-1@FAR=0.001Rank-1@FAR=0.001
PACH [5] (2020)96.978.798.998.3
ADCANs [13] (2021)97.896.199.198.5
MMTN [29] (2022)97.779.299.498.7
NLPF [43] (2024)98.898.399.799.5
LPL [14] (2024)98.798.599.697.5
UMFN (ours)99.198.799.999.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.462, + 0.483, + 0.49 + ], + "angle": 0, + "content": "Table 2. Rank-1 accuracy and verification rate on the LAMP-HQ NIR-VIS dataset and the CASIA NIR-VIS 2.0 dataset." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.501, + 0.481, + 0.59 + ], + "angle": 0, + "content": "
Feature fusionContrastiveDatasets
LAMP-HQCASIANIR-VIS 2.0
Rank-1@FAR=0.001Rank-1@FAR=0.001
91.287.695.795.1
✓94.790.597.596.7
✓97.395.398.798.2
✓✓99.198.799.999.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.602, + 0.483, + 0.631 + ], + "angle": 0, + "content": "Table 3. Ablation experiments of the contrastive learning mechanism and feature fusion in the HFR network." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.643, + 0.483, + 0.839 + ], + "angle": 0, + "content": "moving the identity feature similarity loss \\( L_{id} \\), reconstruction loss \\( L_{rec} \\), feature decoupling loss \\( L_{map} \\) in the mapping network, global path discriminator \\( D_w \\), and the local path discriminator \\( D_f \\). The face normalization results for these variants are presented in Fig. 8. Observations indicate that: (1) Without \\( L_{id} \\), generated prototypes fail to preserve identity characteristics; (2) The absence of \\( L_{rec} \\) causes control masks to fail, preventing the generation of specified domain prototypes; (3) Lack of \\( L_{map} \\) results in the retention of original domain information in cross-domain prototypes; (4) Lack of \\( D_w \\) results in incomplete faces with poorly generated visual effects; (5) Removing \\( D_f \\) leads to a loss of local details in generated prototypes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Furthermore, Tab. 3 quantitatively analyzes the impact of the contrastive learning mechanism and feature concatenation on HFR network training. Results show that contrastive learning significantly enhances the network's uti" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.329, + 0.888, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.604, + 0.881, + 0.62 + ], + "angle": 0, + "content": "Figure 8. Face normalization results of UMFN's variants." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.907, + 0.707 + ], + "angle": 0, + "content": "lization of generated prototypes, allowing it to extract key features similar to real prototypes while reducing the effects of noise and feature loss. Additionally, fusion of domain-invariant and identity-discriminative features improves classification performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.723, + 0.634, + 0.74 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This paper has proposed the new UMFN model. It effectively addresses the limitations in handling diverse facial variations and cross-domain adaptability by reconstructing frontal, neutral-expression facial prototypes, and enabling unsupervised domain adaptation. Through joint prototype and feature learning, along with a well-designed HFR network, our method significantly enhances identity recognition accuracy across various domains. Experiments have validated UMFN's superiority in single-modal and cross-modal face normalization tasks, as well as in HFR task." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29306" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.256 + ], + "angle": 0, + "content": "This work was supported in part by the Natural Science Foundation of China (62466036, 62162041), by the Natural Science Foundation of Jiangxi Province (20232BAB212025), by the Natural Science Foundation of Henan Province (242300421402), by the High-level and Urgently Needed Overseas Talent Programs of Jiangxi Province (20232BCJ25024, 20232BCJ25026), by the Science&Technology Planning Project of Shanghai (23010501800), and by the Key Research and Development Program of Jiangxi Province (20243BBG71035)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.285, + 0.188, + 0.301 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.31, + 0.484, + 0.351 + ], + "angle": 0, + "content": "[1] Bing Cao, Nannan Wang, Jie Li, Qinghua Hu, and Xinbo Gao. Face photo-sketch synthesis via full-scale identity supervision. Pattern Recognition, 124:108446, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.354, + 0.484, + 0.395 + ], + "angle": 0, + "content": "[2] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age. In FG, pages 67-74, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.398, + 0.484, + 0.437 + ], + "angle": 0, + "content": "[3] William Carver and Ifeoma Nwogu. Facial expression neutralization with stoicnet. In WACV, pages 201-208, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.442, + 0.484, + 0.496 + ], + "angle": 0, + "content": "[4] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):10850-10869, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.499, + 0.484, + 0.54 + ], + "angle": 0, + "content": "[5] Boyan Duan, Chaoyou Fu, Yi Li, Xingguang Song, and Ran He. Cross-spectral face hallucination via disentangling independent factors. In CVPR, pages 7930-7938, 2020. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.543, + 0.484, + 0.583 + ], + "angle": 0, + "content": "[6] Hamed Kiani Galoogahi and Terence Sim. Face sketch recognition by local radon binary pattern: Lrbp. In ICIP, pages 1837-1840, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.586, + 0.484, + 0.641 + ], + "angle": 0, + "content": "[7] Yuan Gao, Jiayi Ma, and Alan L Yuille. Semi-supervised sparse representation based classification for face recognition with insufficient labeled samples. IEEE Transactions on Image Processing, 26(5):2545-2560, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.644, + 0.484, + 0.697 + ], + "angle": 0, + "content": "[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NeurIPS, 27, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.701, + 0.484, + 0.742 + ], + "angle": 0, + "content": "[9] Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein gans. NeurIPS, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.745, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[10] Dequan Guo, Lingrui Zhu, Shenggui Ling, Tianxiang Li, Gexiang Zhang, Qiang Yang, Ping Wang, Shiqi Jiang, Sidong Wu, and Junbao Liu. Face illumination normalization based on generative adversarial network. Natural Computing, 22(1):105-117, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.816, + 0.483, + 0.856 + ], + "angle": 0, + "content": "[11] Tal Hassner, Shai Harel, Eran Paz, and Roee Enbar. Effective face frontalization in unconstrained images. In CVPR, pages 4295-4304, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.31, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[13] Weipeng Hu and Haifeng Hu. Adversarial disentanglement spectrum variations and cross-modality attention networks for nir-vis face recognition. IEEE Transactions on Multimedia, 23:145-160, 2020. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[14] Weipeng Hu, Yiming Yang, and Haifeng Hu. Pseudo label association and prototype-based invariant learning for semi-supervised nir-vis face recognition. IEEE Transactions on Image Processing, 33:1448-1463, 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[15] Yibo Hu, Xiang Wu, Bing Yu, Ran He, and Zhenan Sun. Pose-guided photorealistic face rotation. In CVPR, pages 8398-8406, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[16] D Huang, Jia Sun, and Y Wang. The buaa-visir face database instructions. School Comput. Sci. Eng., Beihang Univ., Beijing, China, Tech. Rep. IRIP-TR-12-FR-001, 3(3): 8, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[17] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[18] Rui Huang, Shu Zhang, Tianyu Li, and Ran He. Beyond face rotation: Global and local perception gan for photorealistic and identity preserving frontal view synthesis. In ICCV, pages 2439-2448, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[19] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In CVPR, pages 1125-1134, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[20] Yeong-Joon Ju, Gun-Hee Lee, Jung-Ho Hong, and Seong-Whan Lee. Complete face recovery gan: Unsupervised joint face rotation and de-occlusion from a single-view image. In WACV, pages 3711–3721, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[21] Yeong-Joon Ju, Gun-Hee Lee, Jung-Ho Hong, and Seong-Whan Lee. Complete face recovery gan: Unsupervised joint face rotation and de-occlusion from a single-view image. In WACV, pages 3711–3721, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[22] Brendan F Klare, Ben Klein, Emma Taborsky, Austin Blanton, Jordan Cheney, Kristen Allen, Patrick Grother, Alan Mah, and Anil K Jain. Pushing the frontiers of unconstrained face detection and recognition: Iarpa janus benchmark a. In CVPR, pages 1931-1939, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.692 + ], + "angle": 0, + "content": "[23] Bo Li, Kaitao Xue, Bin Liu, and Yu-Kun Lai. Bbddm: Image-to-image translation with brownian bridge diffusion models. In CVPR, pages 1952-1961, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.693, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[24] Chenyu Li, Shiming Ge, Daichi Zhang, and Jia Li. Look through masks: Towards masked face recognition with deocclusion distillation. In ACM MM, pages 3016-3024, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[25] Stan Li, Dong Yi, Zhen Lei, and Shengcai Liao. The casia mir-vis 2.0 face database. In CVPRW, pages 348-353, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[26] Changcheng Liang, Mingrui Zhu, Nannan Wang, Heng Yang, and Xinbo Gao. Pmsgan: Parallel multistage gans for face image translation. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Xiao Luan, Jiezhong Zheng, and Weisheng Li. Learning unsupervised face normalization through frontal view reconstruction. IEEE Transactions on Circuits and Systems for Video Technology, 32(8):5201-5212, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29307" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.148 + ], + "angle": 0, + "content": "[28] Bingjun Luo, Zewen Wang, Jinpeng Wang, Junjie Zhu, Xibin Zhao, and Yue Gao. Multi-energy guided image translation with stochastic differential equations for near-infrared facial expression recognition. In AAAI, pages 565-573, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[29] Mandi Luo, Haoxue Wu, Huaibo Huang, Weizan He, and Ran He. Memory-modulated transformer network for heterogeneous face recognition. IEEE Transactions on Information Forensics and Security, 17:2095-2109, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[30] Joel Ruben Antony Moniz, Christopher Beckham, Simon Rajotte, Sina Honari, and Chris Pal. Unsupervised depth estimation, 3d face rotation and replacement. Advances in neural information processing systems, 31, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[31] Meng Pang, Yiu-Ming Cheung, Qiquan Shi, and Mengke Li. Iterative dynamic generic learning for face recognition from a contaminated single-sample per person. IEEE Transactions on Neural Networks and Learning Systems, 32(4): 1560-1574, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[32] Meng Pang, Binghui Wang, Yiu-ming Cheung, Yiran Chen, and Bihan Wen. Vd-gan: A unified framework for joint prototype and representation learning from contaminated single sample per person. IEEE Transactions on Information Forensics and Security, 16:2246-2259, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[33] Meng Pang, Binghui Wang, Mang Ye, Yiu-ming Cheung, Yi ran Chen, and Bihan Wen. Disp+ v: A unified framework for disentangling prototype and variation from single sample per person. IEEE Transactions on Neural Networks and Learning Systems, 34(2):867-881, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.475, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[34] Meng Pang, Binghui Wang, Siyu Huang, Yiu-Ming Cheung, and Bihan Wen. A unified framework for bidirectional prototype learning from contaminated faces across heterogeneous domains. IEEE Transactions on Information Forensics and Security, 17:1544-1557, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[35] Yichen Qian, Weihong Deng, and Jiani Hu. Unsupervised face normalization with extreme pose and expression in the wild. In CVPR, pages 9851-9858, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[36] Yue Que, Li Xiong, Weiguo Wan, Xue Xia, and Zhiwei Liu. Denoising diffusion probabilistic model for face sketch-to-photo synthesis. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[37] Lingxiao Song, Zhihe Lu, Ran He, Zhenan Sun, and Tieniu Tan. Geometry guided adversarial facial expression synthesis. In ACM MM, pages 627-635, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[38] Luan Tran, Xi Yin, and Xiaoming Liu. Disentangled representation learning gan for pose-invariant face recognition. In CVPR, pages 1415-1424, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[39] Jingwei Xin, Zikai Wei, Nannan Wang, Jie Li, Xiaoyu Wang, and Xinbo Gao. Learning a high fidelity identity representation for face frontalization. IEEE Transactions on Circuits and Systems for Video Technology, 33(11):6952-6964, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[40] Jianjin Xu, Saman Motamed, Praneetha Vaddamanu, Chen Henry Wu, Christian Haene, Jean-Charles Bazin, and Fernando De la Torre. Personalized face inpainting with diffusion models by parallel visual attention. In WACV, pages 5432–5442, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[41] Weihong Xu, Xiaohua Xie, and Jianhuang Lai. Relightgan: Instance-level generative adversarial network for face illumni" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "nation transfer. IEEE Transactions on Image Processing, 30: 3450-3460, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[42] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4): 1-39, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[43] Yiming Yang, Weipeng Hu, and Haifeng Hu. Neutral face learning and progressive fusion synthesis network for nir-vis face recognition. IEEE Transactions on Circuits and Systems for Video Technology, 33(10):5750-5763, 2024. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[44] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? NeurIPS, 27, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[45] Aijing Yu, Haoxue Wu, Huaibo Huang, Zhen Lei, and Ran He. Lamp-hq: A large-scale multi-pose high-quality database and benchmark for nir-vis face recognition. International Journal of Computer Vision, 129(5):1467-1483, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[46] Ni Zhang, Nian Liu, Junwei Han, Kaiyuan Wan, and Ling Shao. Face de-occlusion with deep cascade guidance learning. IEEE Transactions on Multimedia, 25:3217-3229, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[47] Jian Zhao, Yu Cheng, Yan Xu, Lin Xiong, Jianshu Li, Fang Zhao, Karlekar Jayashree, Sugiri Pranata, Shengmei Shen, Junliang Xing, et al. Towards pose invariant face recognition in the wild. In CVPR, pages 2207-2216, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.475, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[48] Hang Zhou, Jihao Liu, Ziwei Liu, Yu Liu, and Xiaogang Wang. Rotate-and-render: Unsupervised photorealistic face rotation from single-view images. In CVPR, pages 5911-5920, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.531, + 0.905, + 0.585 + ], + "angle": 0, + "content": "[49] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, pages 2223-2232, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29308" + } + ] +] \ No newline at end of file diff --git a/2025/UMFN_ Unified Multi-Domain Face Normalization for Joint Cross-domain Prototype Learning and Heterogeneous Face Recognition/3e3f9780-5e1d-4d68-aee1-68dcb88c5368_origin.pdf b/2025/UMFN_ Unified Multi-Domain Face Normalization for Joint Cross-domain Prototype Learning and Heterogeneous Face Recognition/3e3f9780-5e1d-4d68-aee1-68dcb88c5368_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2ed69963c05a47d3d95c969f3ee1479a2af6e46d --- /dev/null +++ b/2025/UMFN_ Unified Multi-Domain Face Normalization for Joint Cross-domain Prototype Learning and Heterogeneous Face Recognition/3e3f9780-5e1d-4d68-aee1-68dcb88c5368_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76c20a1e139269ec23c9ceb0f266dbc3f32846a5ae6abeee3a615bf6f4781287 +size 8275020 diff --git a/2025/UMFN_ Unified Multi-Domain Face Normalization for Joint Cross-domain Prototype Learning and Heterogeneous Face Recognition/full.md b/2025/UMFN_ Unified Multi-Domain Face Normalization for Joint Cross-domain Prototype Learning and Heterogeneous Face Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..25f00d3d7665a812d0745f32450565c071fb3af5 --- /dev/null +++ b/2025/UMFN_ Unified Multi-Domain Face Normalization for Joint Cross-domain Prototype Learning and Heterogeneous Face Recognition/full.md @@ -0,0 +1,300 @@ +# UMFN: Unified Multi-Domain Face Normalization for Joint Cross-domain Prototype Learning and Heterogeneous Face Recognition + +Meng Pang1, Wenjun Zhang1, Nanrun Zhou2*, Shengbo Chen1*, Hong Rao1 +1Nanchang University +2Shanghai University of Engineering Science + +{mengpang@, zhangwenjun@, raohong@}ncu.edu.cn, nrzhou@sues.edu.cn, ccb02kingdom@gmail.com + +# Abstract + +Face normalization aims to enhance the robustness and effectiveness of face recognition systems by mitigating intra-personal variations in expressions, poses, occlusions, illuminations, and domains. Existing methods face limitations in handling multiple variations and adapting to cross-domain scenarios. To address these challenges, we propose a novel Unified Multi-Domain Face Normalization Network (UMFN) model, which can process face images with various types of facial variations from different domains, and reconstruct frontal, neutral-expression facial prototypes in the target domain. As an unsupervised domain adaptation model, UMFN facilitates concurrent training on multiple datasets across domains and demonstrates strong prototype reconstruction capabilities. Notably, UMFN serves as a joint prototype and feature learning framework, enabling the simultaneous extraction of domain-agnostic identity features through a decoupling mapping network and a feature domain classifier for adversarial training. Moreover, we design an efficient Heterogeneous Face Recognition (HFR) network that fuses domain-agnostic and identity-discriminative features for HFR, and introduce contrastive learning to enhance identity recognition accuracy. Empirical studies on diverse cross-domain face datasets validate the effectiveness of our proposed method. + +# 1. Introduction + +Face normalization, as a core technology in the field of computer vision, is dedicated to eliminating interfering variation factors in facial images, such as pose, expression, occlusion, and illumination, and reconstructing a frontal, unoccluded, evenly illuminated facial prototype with a neutral expression [34]. This technology has shown broad application prospects in numerous areas, including criminal investiga + +![](images/8985e2ee59487c050ba845e114c60d9a26ee5e63722dfb990b40c34bb2ba5d60.jpg) +Figure 1. Illustration of a desired face normalization method capable of handling multiple cross-domain face normalization tasks and addressing various types of facial variations. + +tion , authentication, and human-computer interaction [39]. However, existing face normalization methods face challenges in coping with increasingly complex and diversified application scenarios, highlighting the urgent need for more advanced and versatile solutions to flexibly adapt to various changes in different environments and conditions. + +The limitations of current face normalization methods lie mainly in their constraints in processing single contaminant elements and their single-domain applicability. Most methods [3, 10] focus on modeling and processing a single element, such as pose, expression, illumination, or occlusion, and struggle with multiple coexisting factors. Additionally, some unsupervised methods [41, 46] may inadvertently lose crucial identity information, leading to identity inconsistency between the generated prototype and the input face. Furthermore, existing models [32, 40] are typically domain-specific, lacking universality across multi-domain scenarios. However, practical applications often involve images + +from various domains, such as visible light, near-infrared, and sketch images used in criminal investigations. Therefore, there is an urgent need for a multi-domain face normalization method capable of handling various facial variations, as illustrated in Fig. 1. + +To address these challenges, we propose a Unified Multidomain Face Normalization Network (UMFN) model. This model is capable of handling facial images from different domains and containing multiple contaminant elements (such as pose, expression, occlusion, etc.), and can reconstruct a frontal, neutral facial prototype in the target domain while simultaneously extracting identity features unrelated to domain information. It is worth noting that, despite the significant achievements of diffusion models in the field of image generation, their limitations in interpretable representation learning make them difficult to meet our needs for joint prototype and feature learning [42]. Therefore, this paper adopts a fusion framework of Autoencoder and Generative Adversarial Network (GAN), as shown in Fig. 2. + +The UMFN model introduces a feature decoupling network and domain classifier for adversarial training, decoupling domain-agnostic features. By leveraging concatenated masks, it regulates clean prototype generation within specified domains or reconstructs original images, ensuring identity consistency. Additionally, a multi-scale InstanceNorm Patch-based discriminator [19] enhances image details and textures, improving visual quality. As an unsupervised domain adaptation model, UMFN enables concurrent training of diverse, unlabeled datasets within a single network, demonstrating flexibility and practicality in cross-domain prototype reconstruction. Furthermore, we design an efficient heterogeneous face recognition (HFR) network in conjunction with UMFN's encoder to extract identity-discriminative features for the generated prototype. These features are fused with domain-agnostic features of input for HFR. During HFR network training, a contrastive learning mechanism minimizes the feature distance between generated and authentic prototypes, enhancing the network's ability to accurately recognize identities. + +The contributions of the paper are summarized below: + +- We propose a unified face normalization model, dubbed UMFN, which addresses the limitations of existing methods in handling multiple variation factors and cross-domain face images. This model is capable of simultaneously processing various types of facial variations and is applicable to face images from different domains. +- We design an efficient HFR network that fuses domain-agnostic features from input sample with identity-discriminative features from generated prototype for HFR. Furthermore, a contrastive learning mechanism is introduced to bolster the network's capability in accurately recognizing identities. +- As an unsupervised domain adaptation model, UMFN en + +ables simultaneous training of multiple datasets with different domains and without identity labels in a single network. It is able to perform prototype reconstruction tasks across multiple domains, enhancing model flexibility and practicality, adapting to diverse scenarios, and reducing reliance on labeled data. + +- Empirical study on multiple cross-domain face datasets with facial variations validates the superiority of UMFN in face normalization tasks in both single-modal and cross-modal scenarios, as well as in HFR task. + +# 2. Related Works + +Face normalization, an emerging and highly esteemed research field, has shown substantial practical value in areas like criminal identification and facial evidence collection. Mainstream face normalization methods primarily address facial variation removal in single-modal scenarios, aiming to eliminate factors such as diverse lighting conditions, exaggerated facial expressions, large head poses, or partially occlusions, thereby restoring the standard facial prototype. Existing single-modal face normalization algorithms can be broadly classified into two categories [33]: The first constructs clustering models based on auxiliary sample information to estimate facial prototypes of contaminated samples. The second leverages generative AI models, such as GANs [8] and Diffusion Models [4], to learn the mapping between contaminated and standard facial samples. + +In the first category, Gao et al. [7] combined unlabeled facial samples with a labeled training set, using a Gaussian Mixture Model for clustering to estimate prototypes. Pang et al. [31] introduced a more robust semi-supervised low-rank representation model for clustering and prototype estimation, updated with labeled retrieval samples. Although effective in prototype reconstruction, these methods require prior access to the facial samples, making them unsuitable for real-time scenarios. In the second category, benefiting from generative AI models' image generation and mapping capabilities, researchers have successively proposed a series of generative learning methods, aiming to eliminate expression variations [3, 37], illumination variations [10, 41], pose variations [15, 27, 46], and partial occlusions [21, 24, 40] in input facial samples, respectively, and generate facial prototype images with consistent identities. However, these methods are typically limited to specific types of facial variations and cannot effectively address combinations of multiple types. Pang et al. [32] attempts to solve prototype reconstruction for multiple types of contamination by designing a Variation Disentangling GAN (VD-GAN), but has limitations in preserving identity characteristics and handling only single-domain contamination. + +Recently, a few of cross-domain face synthesis methods [1, 23, 26, 28, 36, 49] have been proposed, including Cycle-consistency GAN (CycleGAN) [49], Parallel Multi + +![](images/caa5ac8b6ec3f66c293a79d26b20b97aff27d1caa2ed286978203f688d1e02e4.jpg) +Figure 2. Overview of the architecture of UMFN. $x_{c} \in X_{c}$ denotes an input image in domain $c$ , $\hat{x}_c$ represents the reconstructed image of $x_{c}$ , $\hat{x}_{\bar{c}}^{rp}$ denotes the generated prototype of $x_{c}$ in domain $\bar{c}$ , and $x_{c}^{rp}$ represents the real prototype in domain $c$ . + +stage GANs (PMSGAN) [1], Brownian Bridge Diffusion Models (BBDM) [23], and NIR-FER Stochastic Differential Equations (NFER-SDE) [28]. These methods focus on transforming domain styles while preserving facial details, but they cannot remove facial variations from in-put faces. In light of the above, this paper aims to design a multidomain face normalization scheme applicable to universal facial variation removal and explore effective integration and optimization of domain transformation and prototype learning within a unified framework. + +# 3. Methods + +This paper proposes a universal face normalization framework, UMFN, whose architecture is shown in Fig. 2. UMFN is designed to uniformly process face images with various types of facial variations from different image domains, ultimately generating standard facial prototypes in the specified domain. Correspondingly, UMFN comprises the following two core components: 1) addressing the issue of input images originating from multiple domains and generating images in the specified domain; 2) uniformly handling multiple types of facial variations while generating prototypes that preserve the identity (ID). These two components will be detailed in Sec. 3.1 and Sec. 3.2 + +# 3.1. Feature Decoupling and Cross-Domain Generation + +To tackle the challenge that input face images may come from different domains, we have designed a feature decoupling mapping network. This network is able to separate domain information from the latent space, resulting in domain-invariant representations. Such decoupled, domain-agnostic identity features can be applied to HFR task. + +Feature Decoupling: We design a feature decoupling mapping network that initially extracts primary feature $z =$ + +$G_{enc}(x_c)$ from the input image $x_c$ using a pre-trained encoder $G_{enc}$ , and then derives its latent feature representation $p = \text{map}(z)$ through the mapping network map. Specifically, the domain feature classifier $D_c$ receives the latent representations $p$ as input and outputs the domain prediction code (i.e., the image domain to which $p$ belongs). The goal of $D_c$ is to accurately predict the domain of the input latent representation $p$ , while the mapping network map aims to confuse $D_c$ 's correct domain prediction. Through adversarial training, the interplay between the two achieves domain information decoupling. Accordingly, the objective function for training $D_c$ is presented as follows: + +$$ +L _ {D _ {c}} = E \left[ \log P \left(l _ {p} ^ {\text {p r e}} = l _ {x _ {c}} ^ {\text {r e a l}} \mid p\right) \right], \tag {1} +$$ + +where $l_{p}^{pre}$ is the domain label predicted by $D_{c}$ , $l_{x_{c}}^{real}$ is the real domain label of $x_{c}$ . On the other side, the objective function for training the mapping network map is as + +$$ +L _ {m a p} = - E \left[ \log P \left(l _ {p} ^ {p r e} = l _ {x _ {c}} ^ {r e a l} \mid p\right) \right]. \tag {2} +$$ + +Cross-domain Generation: We introduce the control mask and combine it with reconstruction learning to achieve cross-domain image generation. The core objective of reconstruction learning is to guide the decoder $G_{dec}$ to learn how to decode domain-agnostic latent representation $p = \text{map}(G_{enc}(x_c))$ conditioned on the original attribute mask $v_c$ , thereby reconstructing the input image $x_c$ . This process ensures that the decoder not only learns to understand and apply the semantics of the mask but also fully utilizes the identity feature information in the latent representation $p$ . As such, in subsequent steps, images from specified domains can be generated by manipulating the mask. The objective function for reconstruction learning is as follows: + +$$ +L _ {r e c} = E _ {x _ {c} \sim p _ {d a t a}} \left[ \| x _ {c} - \hat {x} _ {c} \| _ {1} \right], \tag {3} +$$ + +where $\hat{x}_c = G_{dec}(p, v_c)$ is the reconstructed image of $x_c$ , $v_c$ is the original attribute mask of the input image $x_c$ , indicating whether the $x_c$ image is polluted and its domain information. Specifically, $v_c$ is composed of a four-digit one-hot vector, where the first digit indicates whether the image is contaminated, and the last three digits represent the domain it belongs to. In Eq. (3), $l_1$ norm is employed instead of the $l_2$ norm to suppress the issue of blurred generated images. + +# 3.2. Prototype generation + +Different from prior single-contamination normalization methods that overly focus on facial variation categories and how to accurately model them, our method returns to the essence of the prototype reconstruction task, aiming to directly generate identity-preserved standard facial prototypes by incorporating identity information from the contaminated facial samples, potentially bypassing the challenges of modeling facial variations in a clever manner. + +Notably, we introduce a prototype adversarial loss, differing from the adversarial loss in traditional GAN-based image translation methods that aim to approximate the distribution of original face data with diverse variations. Our loss aims to approximate a condensed distribution of standard facial prototypes, while also reducing overfitting to these variations. Furthermore, we have designed a multiscale Markov discriminator $D = [D_w, D_f]$ to refine the details of generated faces and stabilize training. Specifically, we concatenate the extracted domain-agnostic identity feature $p = \text{map}(G_{\text{enc}}(x_c))$ with a control mask $v_{\bar{c}}$ , where the first bit in $v_{\bar{c}}$ controls whether to perform face normalization: 1 for yes, and the last three bits control the target domain for prototype generation. Next, this concatenated feature is input into the decoder to generate a pixel-level standardized facial prototype in the specified target domain. The prototype adversarial objectives for training the generator and the multi-scale discriminator are as follows: + +$$ +\begin{array}{l} L _ {D} ^ {\text {a d v}} = D _ {w} \left(\hat {x} _ {\bar {c}} ^ {r p}\right) + D _ {f} \left(\operatorname {c r o p} \left(\hat {x} _ {\bar {c}} ^ {r p}\right)\right) \tag {4} \\ - D _ {w} \left(x _ {c} ^ {r p}\right) - D _ {f} \left(\operatorname {c r o p} \left(x _ {c} ^ {r p}\right)\right), \\ \end{array} +$$ + +$$ +L _ {G _ {d e c}} ^ {a d v} = - \left(D _ {w} \left(\hat {x} _ {\bar {c}} ^ {r p}\right) + D _ {f} \left(\operatorname {c r o p} \left(\hat {x} _ {\bar {c}} ^ {r p}\right)\right)\right), \tag {5} +$$ + +where $D_w$ and $D_f$ represent the global path discriminator and local path discriminator, respectively. $\hat{x}_{\bar{c}}^{rp} = G_{dec}(map(G_{enc}(x_c), v_{\bar{c}})$ denotes the specified domain prototype generated by inputting $x_c$ and the control mask $v_{\bar{c}}$ while $x_c^{rp}$ represents the uncontaminated faces in the training set. The function "crop" is used to cut out the facial features from the input face images. Considering that the Wasserstein distance is more stable and can avoid training instability and mode collapse in GANs, we adopt the WGAN-GP [9] loss in our model to replace the traditional cross-entropy loss. + +Furthermore, to ensure that the generated prototype $\hat{x}_{\bar{c}}^{rp}$ can effectively preserve the identity information from the + +input $x_{c}$ , we have designed an identity feature similarity loss function. This is achieved by feeding the generated prototype back into the encoder and mapping network to obtain the latent feature map $(G_{enc}(\hat{x_c}^{rp}))$ , and then calculating its similarity to the latent feature $p$ of the input $x_{c}$ , thereby retaining identity. The loss function is as follows: + +$$ +L _ {i d} = \| p - \operatorname {m a p} \left(G _ {e n c} \left(\hat {x} _ {\bar {c}} ^ {r p}\right)\right) \| _ {2} ^ {2}, \tag {6} +$$ + +where $\| \cdot \|_2^2$ denotes $l_2$ -norm. On the other hand, UMFN transfers the identity feature retention ability from the reconstruction learning task to the face prototype generation task. Since these two tasks share the same input and output domains, the transferability gap [44] between them is minimal. Therefore, the identity feature retention ability learned from the reconstruction learning task can be easily transferred to the prototype reconstruction task. Therefore, the overall loss function of the decoder $G_{dec}$ is as follows: + +$$ +L _ {G _ {d e c}} ^ {\text {a d v}} = \lambda_ {1} L _ {r e c} + \lambda_ {2} L _ {i d}. \tag {7} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ are balancing hyperparameters. + +# 3.3. Heterogeneous Face Recognition + +At this stage, we meticulously construct a HFR network, whose core components encompass an encoder $F_{\theta}$ (its network architecture aligns with $G_{enc}$ in UMFN) and a fully connected layer $l_{\theta}$ . The specific structure is shown in Fig. 3. The HFR network consists of two core operations: + +Feature Fusion: During training, we concatenate the domain-agnostic features (extracted by UMFN's $G_{enc}$ and mapping network map from the source domain input image) with the identity-discriminative features (extracted by encoder $F_{\theta}$ from the real target domain prototype). The fused features are then fed into the softmax layer for classification. In this process, the fully connected layer $l_{\theta}$ doubles the feature dimensions, enhancing classification robustness through multi-dimensional identity information integration. The classification loss for this step is presented as + +$$ +L _ {\mathrm {c l s}} = \operatorname {s o f t m a x} \left(l _ {\theta} \left[ F _ {\theta} \left(x _ {v, i} ^ {r p}\right) \oplus \operatorname {m a p} \left(G _ {e n c} \left(x _ {n, i}\right)\right) \right], i\right), \tag {8} +$$ + +where $x_{n,i}$ represents source domain image and $x_{v,i}^{rp}$ represents real target domain prototype, $F_{\theta}(x_{v,i}^{rp})$ denotes the representation of the input image $x_{v,i}^{rp}$ extracted by $F_{\theta}$ , while map $(G_{enc}(x_{n,i}))$ denotes the domain-agnostic feature of the input image $x_{n,i}$ , the symbol $\oplus$ indicates the concatenation of these two features. During testing, we extract the identity-discriminative feature of the generated target domain prototype from the input test source domain image using the trained $F_{\theta}$ , and fuse it with the domain-agnostic feature of the input test source domain image for HFR. + +Contrastive Learning Mechanism: During the cross-domain generation, generated target domain prototypes may + +![](images/3f40e077f6da44a76ea1795eb727d4d76c2469a3049e7eb57e36c47fc100f6c7.jpg) +Figure 3. Overview of the HFR network and training process. + +suffer identity feature loss or noise introduction. Direct reliance on such data for identity classification training can lead to overfitting to missing or noisy features, impairing classification performance. To mitigate this, we innovatively employ a contrastive learning mechanism that aligns generated prototypes with real ones in the feature space, while reducing the impact of noise and missing features. This ensures that the encoder $F_{\theta}$ extracts features closely resembling real prototypes. Specifically, we randomly sample two pairs of real heterogeneous images, $(x_{v,1}^{rp}, x_{n,1})$ and $(x_{v,2}^{rp}, x_{n,2})$ , where $x_{n,1}$ and $x_{n,2}$ are NIR images with contamination. We then use UMFN to perform cross-domain restoration on $x_{n,1}$ and $x_{n,2}$ , generating the corresponding target VIS domain prototypes $\hat{x}_{v,1}^{rp}$ and $\hat{x}_{v,2}^{rp}$ . Next, we form positive pairs $(x_{v,1}^{rp}, \hat{x}_{v,1}^{rp})$ and $(x_{v,2}^{rp}, \hat{x}_{v,2}^{rp})$ , and negative pairs $(x_{v,1}^{rp}, \hat{x}_{v,2}^{rp})$ and $(x_{v,2}^{rp}, \hat{x}_{v,1}^{rp})$ . The corresponding contrastive loss is as follows: + +$$ +\begin{array}{l} L _ {c o n t} = \sum_ {j \neq k} \left(1 - \langle f _ {j}, \tilde {f} _ {j} \rangle\right) + \left(1 - \langle f _ {k}, \tilde {f} _ {k} \rangle\right) \\ + \max \left(0, \left\langle f _ {j}, \tilde {f} _ {k} \right\rangle - m\right) + \max \left(0, \left\langle f _ {k}, \tilde {f} _ {j} \right\rangle - m\right), \tag {9} \\ \end{array} +$$ + +where $f_{j} = l_{\theta}[F_{\theta}(x_{v,j}^{rp}) \oplus map(G_{enc}(x_{n,j}))]$ , $\langle, \rangle$ represents the calculation of the cosine similarity between two features, and $m$ is a margin value set at 0.5. In summary, the final loss function of the HFR network is expressed as + +$$ +L _ {h f r} = L _ {c l s} + \alpha L _ {\text {c o n t}}. \tag {10} +$$ + +where $\alpha$ is a trade-off parameter. + +# 4. Experiments + +# 4.1. Datasets Descriptions + +LAMP-HQ [45] is a newly created, challenging large-scale NIR-VIS dataset comprising over 56,000 NIR and 16,000 VIS face images from 573 individuals, featuring diverse expressions, poses, lighting conditions, occlusions, and complex backgrounds. BUAA NIR-VIS [16] includes 150 identities, each with 9 NIR and 9 VIS images, covering various head poses and expressions. CUFSF [6] consists of 597 grayscale images paired with 597 real sketch images. CASIA NIR-VIS 2.0 [25] encompasses 725 identities, with 5-50 NIR and 1-22 VIS images per identity. IJB-A [22] and LFW [17] are widely-used benchmarks for face recognition and verification, featuring diverse poses, occlusions, and expressions in the visible domain. + +# 4.2. Experimental Setup + +In the experimental section, we primarily conduct four experiments: (1) Homogeneous Multi-domain Face Normalization Experiment, (2) Heterogeneous Multi-domain Face Normalization Experiment, (3) Heterogeneous Face Recognition Experiment. (4) Ablation Study. The specific configurations of these experiments are as follows: + +1) Homogeneous Multi-domain Face Normalization Experiment: In this experiment, we utilize three types of domain data from the LAMP-HQ dataset to train our model, ensuring that all data originate from the same set of identities (IDs). The experiment primarily evaluates the model's performance in facial normalization across multi-domain data within the LAMP-HQ dataset. Specifically, we employ FP data from the LAMP-HQ dataset, where all NIR and VIS images from 100 IDs are used as the training set, while the remaining IDs are designated as the test set. As the LAMP-HQ dataset only contains NIR and VIS domain data, we use software to convert NIR photographs into sketch images, thereby simulating a scenario with three types of domain data. Furthermore, we transfer the model to other datasets, namely LFW and IJB-A, for testing. We assess the model's performance from both qualitative and quantitative perspectives and compare it with the latest single-modal face normalization techniques. +2) Heterogeneous Multi-domain Face Normalization Experiment: We use one type of domain data from three different datasets to demonstrate that the model can effectively work without paired data training, verifying its generalization ability across datasets. Specifically, we use VIS images of 250 IDs from the LAMP-HQ dataset as VIS domain data, all NIR images of 130 IDs from the BUAA dataset as NIR domain data, and Sketch images of 400 IDs from the CUFSF dataset as Sketch domain data for training. +3) Heterogeneous Face Recognition Experiment: We conduct heterogeneous face recognition experiments on the + +![](images/d61d0faa86dee5a206156c1629087a8a2efe2d838a0898ee3d0dc513be6fd665.jpg) +Figure 4. In-domain face normalization results on the LAMP-HQ dataset. + +LAMP-HQ and CASIA NIR-VIS 2.0 datasets, following the standard evaluation protocol in [43]. + +4) Ablation Study Experiment: We investigate the contribution of each component within the UMFN framework and explore the roles of the contrastive learning mechanism and feature fusion operation in the HFR network. + +# 4.3. Implementation Details and Parameter Setting + +We employ a Resnet-50 pre-trained on the MS1M dataset and fine-tuned on VGGFace2 [2] as the encoder $G_{enc}$ to extract 2048-dimensional identity features, keeping $G_{enc}$ fixed during training and testing. The mapping network consists of multiple linear layers, inputting a 2048-dimensional vector and outputting a 2048-dimensional vector. The feature domain classifier $D_{c}$ is composed of linear layers and outputs the predicted domain category, while the decoder $G_{dec}$ outputs images of size $224 \times 224$ . The discriminators $D_{W}$ and $D_{f}$ are both patchGAN discriminators with a receptive field of 70, using instance normalization as the normalization layer, and the crop size is $150 \times 150$ . + +The hyperparameters $\lambda_{1}$ and $\lambda_{2}$ in Eq. (7) are set at $\lambda_{1} = 10 - \frac{cur_{epoch}}{sum_{epoch}} \times 9$ and $\lambda_{2} = 10$ , where $cur_{epoch}$ is the current number of training epochs and $sum_{epoch}$ is the total number of training epochs, which is 100. $\alpha$ in Eq. (10) is empirically set at 0.001. During optimization, we use the Adam optimizer, with the learning rate set at 0.0002 and momentum set at 0.5. + +# 4.4. Homogeneous Multi-domain Normalization + +In this section, we evaluate the performance of UMFN in homogeneous multi-domain face normalization using the LAMP-HQ dataset. The in-domain normalization results of UMFN are presented in Fig. 4. Fig. 4(a) showcases the results of inputting VIS domain faces corrupted by different facial variations such as pose, lighting, occlusion, and expression, and outputting normalized VIS domain faces. Fig. 4(b) displays the outcomes of inputting corrupted faces from the NIR and Sketch domains and outputting normalized faces in their respective domains. The bottom row serves as a reference for the true prototypes. From Fig. 4, it can be observed that UMFN is capable of simultaneously + +normalizing various types of corrupted faces from multiple different domains within their respective domains, effectively preserving the identity of the input faces while obtaining facial prototypes. Additionally, we conduct a comparison between UMFN and other popular single-domain face normalization methods, including PIM [47], TP-GAN [18], DR-GAN [38], Hassner [11], and FNM [35], using contaminated face images from the LFW dataset as input. The face normalization results of each method are presented in Fig. 5. It can be observed that the target prototype obtained through UMFN normalization excels in both image quality and identity preservation. + +Moreover, following the protocols in [48] and [20], we quantitatively evaluate UMFN on the LFW and IJB-A datasets. By inputting restored faces into the ResNet-18 [12] network for face verification, we aim to assess whether UMFN-restored faces enhance VR performance. Additionally, we report the Symmetry Index (SI)[30] to provide a comprehensive comparison of the quality of prototypes generated by UMFN, R&R [48], and CFR [20]. Tab. 1 presents the verification performance and SI of UMFN across both datasets. The results indicate that UMFN normalization significantly boosts the verification performance of the ResNet model: on IJB-A, when combined with ResNet-18, it improves VR by $5.35\%$ , with accuracy gains also observed on LFW. The experiments demonstrate that: (1) removing pose, expression, occlusion, and illumination variations from input faces allows the model to better focus on key identity features, and (2) UMFN effectively retains critical identity information while eliminating these facial variations. In addition, UMFN outperforms R&R and CFR in terms of SI, generating highly symmetrical faces that correct facial distortions. + +Furthermore, we illustrate the cross-domain normalization results of the UMFN model in Fig. 6. Specifically, the inputs consist of contaminated facial images from domain A, while the outputs represent the generated prototypes in domains B and C. The bottom row displays the reference prototypes of the inputs in the VIS domain. It can be observed that the UMFN model successfully addresses the challenges of domain transfer and prototype learning, + +![](images/c958e0b0aee430106d624c974f01f1ca0a6a16dec2cd0a84f1f803690604af89.jpg) +Figure 5. Comparison with other single-domain face normalization methods on LFW dataset. + +
MethodDatasets
LFWIJB-A
VR↑SI↓VR↑SI↓
Res18 [12]98.85-90.57±1.2-
Res18+R&R [48]98.9520.9391.98±0.720.85
Res18+CFR [20]99.2328.3193.36±0.624.76
Res18+UMFN(ours)99.5714.6195.92±0.715.81
+ +Table 1. Verification performance and Symmetry Index (SI) on LFW and IJB-A. + +effectively preserving the identity characteristics of the inputs when generating uncontaminated facial prototypes for the target domains. The inspiring results demonstrate that UMFN is capable of performing effective domain-specific face normalization on facial images sourced from different domains and contaminated by various types of facial variations, all while efficiently retaining the key identity information of the inputs, utilizing a single network architecture. + +# 4.5. Heterogeneous Multi-domain Normalization + +In this section, we evaluate the performance of UMFN in the heterogeneous multi-domain face normalization experiment. It is particularly noteworthy that there is currently no other face normalization method capable of adequately addressing this complex issue. Fig. 7 showcase the normalization results of images from the LAMP-HQ, BUAA, and CUFSF datasets across different domains (VIS, NIR, Sketch). We observe that regardless of the dataset (or domain) the input comes from, UMFN successfully generates prototypes that are highly matched to the target domain styles provided by another dataset. It effectively eliminates interfering factors such as pose, expression, and occlusion in the input face images while accurately preserving the input identity information. It is worth noting that due to the minimal facial variations in most images of the CUFSF dataset, which generally exhibit simple facial expressions such as smiles, the changes after normalization are primarily manifested in the mouth area. Furthermore, the challenging nature of cross-domain normalization tasks involving sketch images, which contain the least amount of facial detail information, leads to slight local distortions in the generated VIS and NIR prototypes. Despite these challenges, the results demonstrate UMFN's robust multi- + +![](images/f1aa9232f686686956327a233a07e4997ed972262da688b36245f5e293920d3a.jpg) +Figure 6. Cross-domain face normalization results on LAMP-HQ dataset. The second and third rows depict face normalization results in different domains from the inputs shown in the first row. The bottom row shows the real VIS prototypes for reference. + +domain and multi-type face normalization capabilities, particularly when trained on unpaired datasets from multiple domains, where its performance remains impressive. + +# 4.6. Heterogeneous Face Recognition + +In this experiment, we employ UMFN to restore cross-domain prototypes from NIR data in the probe set, generating VIS prototypes. Subsequently, we utilize the trained HFR network to obtain fused features (i.e., concatenation of domain-agnostic and identity-discriminative features) for both these generated prototypes and the authentic prototypes in the gallery set. Finally, by computing the cosine similarity between the fused features of the generated prototypes and those of the gallery set, we obtain a similarity matrix, upon which we calculate the rank-1 recognition accuracy and VR at $\mathrm{FAR} = 0.001$ . + +Tab. 2 presents a comparison of our method with other advanced HFR approaches, including PACH [5], AD-CANs [13], MMTN [29], NLPF [43], and LPL [14], on LAMP-HQ and CASIA NIR-VIS 2.0 datasets. It is observed that our method achieves the best performance on both datasets in terms of rank-1 accuracy and VR. This exceptional performance stems from: (1) Incorporating a contrastive learning mechanism that aligns generated prototypes with real ones in the feature space, reducing noise and missing feature impacts. This allows the HFR encoder $F_{\theta}$ to extract identity-discriminative features from generated prototypes that closely resembling real ones. (2) The fusion of domain-agnostic and identity-discriminative features notably enhances the identity distinguishability of the combined features. + +# 4.7. Ablation Study + +To assess the effectiveness of UMFN and the contributions of its components, we train five variants by respectively re + +![](images/dc85ed7baf14bac2b617006ebe1886ea12340f895ff82741bd8704f8bb721c5c.jpg) +(a) Results on LAMP-HQ dataset +Figure 7. The normalization results of face images from the LAMP-HQ, BUAA, and CUFSF datasets across different domains. + +![](images/cb414ba2e0d25491404cee4d865adef3f99eb6809f4d3d19b895708a06a974ec.jpg) +(b) Results on BUAA dataset + +![](images/d76b988d75d360a3a8524248721da92d7fb3716fc24c53bee718b7616d785c62.jpg) +(c) Results on CUFSF dataset + +
MethodDatasets
LAMP-HQCASIANIR-VIS 2.0
Rank-1@FAR=0.001Rank-1@FAR=0.001
PACH [5] (2020)96.978.798.998.3
ADCANs [13] (2021)97.896.199.198.5
MMTN [29] (2022)97.779.299.498.7
NLPF [43] (2024)98.898.399.799.5
LPL [14] (2024)98.798.599.697.5
UMFN (ours)99.198.799.999.7
+ +Table 2. Rank-1 accuracy and verification rate on the LAMP-HQ NIR-VIS dataset and the CASIA NIR-VIS 2.0 dataset. + +
Feature fusionContrastiveDatasets
LAMP-HQCASIANIR-VIS 2.0
Rank-1@FAR=0.001Rank-1@FAR=0.001
91.287.695.795.1
✓94.790.597.596.7
✓97.395.398.798.2
✓✓99.198.799.999.7
+ +Table 3. Ablation experiments of the contrastive learning mechanism and feature fusion in the HFR network. + +moving the identity feature similarity loss $L_{id}$ , reconstruction loss $L_{rec}$ , feature decoupling loss $L_{map}$ in the mapping network, global path discriminator $D_w$ , and the local path discriminator $D_f$ . The face normalization results for these variants are presented in Fig. 8. Observations indicate that: (1) Without $L_{id}$ , generated prototypes fail to preserve identity characteristics; (2) The absence of $L_{rec}$ causes control masks to fail, preventing the generation of specified domain prototypes; (3) Lack of $L_{map}$ results in the retention of original domain information in cross-domain prototypes; (4) Lack of $D_w$ results in incomplete faces with poorly generated visual effects; (5) Removing $D_f$ leads to a loss of local details in generated prototypes. + +Furthermore, Tab. 3 quantitatively analyzes the impact of the contrastive learning mechanism and feature concatenation on HFR network training. Results show that contrastive learning significantly enhances the network's uti + +![](images/442b66f559b7f2faf407945e0ba9f61f92a6503875cf5027d5bf582fc3230232.jpg) +Figure 8. Face normalization results of UMFN's variants. + +lization of generated prototypes, allowing it to extract key features similar to real prototypes while reducing the effects of noise and feature loss. Additionally, fusion of domain-invariant and identity-discriminative features improves classification performance. + +# 5. Conclusion + +This paper has proposed the new UMFN model. It effectively addresses the limitations in handling diverse facial variations and cross-domain adaptability by reconstructing frontal, neutral-expression facial prototypes, and enabling unsupervised domain adaptation. Through joint prototype and feature learning, along with a well-designed HFR network, our method significantly enhances identity recognition accuracy across various domains. Experiments have validated UMFN's superiority in single-modal and cross-modal face normalization tasks, as well as in HFR task. + +# Acknowledgement + +This work was supported in part by the Natural Science Foundation of China (62466036, 62162041), by the Natural Science Foundation of Jiangxi Province (20232BAB212025), by the Natural Science Foundation of Henan Province (242300421402), by the High-level and Urgently Needed Overseas Talent Programs of Jiangxi Province (20232BCJ25024, 20232BCJ25026), by the Science&Technology Planning Project of Shanghai (23010501800), and by the Key Research and Development Program of Jiangxi Province (20243BBG71035). + +# References + +[1] Bing Cao, Nannan Wang, Jie Li, Qinghua Hu, and Xinbo Gao. Face photo-sketch synthesis via full-scale identity supervision. Pattern Recognition, 124:108446, 2022. 2, 3 +[2] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age. In FG, pages 67-74, 2018. 6 +[3] William Carver and Ifeoma Nwogu. Facial expression neutralization with stoicnet. In WACV, pages 201-208, 2021. 1, 2 +[4] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):10850-10869, 2023. 2 +[5] Boyan Duan, Chaoyou Fu, Yi Li, Xingguang Song, and Ran He. Cross-spectral face hallucination via disentangling independent factors. In CVPR, pages 7930-7938, 2020. 7, 8 +[6] Hamed Kiani Galoogahi and Terence Sim. Face sketch recognition by local radon binary pattern: Lrbp. In ICIP, pages 1837-1840, 2012. 5 +[7] Yuan Gao, Jiayi Ma, and Alan L Yuille. Semi-supervised sparse representation based classification for face recognition with insufficient labeled samples. IEEE Transactions on Image Processing, 26(5):2545-2560, 2017. 2 +[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NeurIPS, 27, 2014. 2 +[9] Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein gans. NeurIPS, 30, 2017. 4 +[10] Dequan Guo, Lingrui Zhu, Shenggui Ling, Tianxiang Li, Gexiang Zhang, Qiang Yang, Ping Wang, Shiqi Jiang, Sidong Wu, and Junbao Liu. Face illumination normalization based on generative adversarial network. Natural Computing, 22(1):105-117, 2023. 1, 2 +[11] Tal Hassner, Shai Harel, Eran Paz, and Roee Enbar. Effective face frontalization in unconstrained images. In CVPR, pages 4295-4304, 2015. 6 +[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6, 7 + +[13] Weipeng Hu and Haifeng Hu. Adversarial disentanglement spectrum variations and cross-modality attention networks for nir-vis face recognition. IEEE Transactions on Multimedia, 23:145-160, 2020. 7, 8 +[14] Weipeng Hu, Yiming Yang, and Haifeng Hu. Pseudo label association and prototype-based invariant learning for semi-supervised nir-vis face recognition. IEEE Transactions on Image Processing, 33:1448-1463, 2024. 7, 8 +[15] Yibo Hu, Xiang Wu, Bing Yu, Ran He, and Zhenan Sun. Pose-guided photorealistic face rotation. In CVPR, pages 8398-8406, 2018. 2 +[16] D Huang, Jia Sun, and Y Wang. The buaa-visir face database instructions. School Comput. Sci. Eng., Beihang Univ., Beijing, China, Tech. Rep. IRIP-TR-12-FR-001, 3(3): 8, 2012. 5 +[17] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 5 +[18] Rui Huang, Shu Zhang, Tianyu Li, and Ran He. Beyond face rotation: Global and local perception gan for photorealistic and identity preserving frontal view synthesis. In ICCV, pages 2439-2448, 2017. 6 +[19] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In CVPR, pages 1125-1134, 2017. 2 +[20] Yeong-Joon Ju, Gun-Hee Lee, Jung-Ho Hong, and Seong-Whan Lee. Complete face recovery gan: Unsupervised joint face rotation and de-occlusion from a single-view image. In WACV, pages 3711–3721, 2022. 6, 7 +[21] Yeong-Joon Ju, Gun-Hee Lee, Jung-Ho Hong, and Seong-Whan Lee. Complete face recovery gan: Unsupervised joint face rotation and de-occlusion from a single-view image. In WACV, pages 3711–3721, 2022. 2 +[22] Brendan F Klare, Ben Klein, Emma Taborsky, Austin Blanton, Jordan Cheney, Kristen Allen, Patrick Grother, Alan Mah, and Anil K Jain. Pushing the frontiers of unconstrained face detection and recognition: Iarpa janus benchmark a. In CVPR, pages 1931-1939, 2015. 5 +[23] Bo Li, Kaitao Xue, Bin Liu, and Yu-Kun Lai. Bbddm: Image-to-image translation with brownian bridge diffusion models. In CVPR, pages 1952-1961, 2023. 2, 3 +[24] Chenyu Li, Shiming Ge, Daichi Zhang, and Jia Li. Look through masks: Towards masked face recognition with deocclusion distillation. In ACM MM, pages 3016-3024, 2020. 2 +[25] Stan Li, Dong Yi, Zhen Lei, and Shengcai Liao. The casia mir-vis 2.0 face database. In CVPRW, pages 348-353, 2013. 5 +[26] Changcheng Liang, Mingrui Zhu, Nannan Wang, Heng Yang, and Xinbo Gao. Pmsgan: Parallel multistage gans for face image translation. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2 +[27] Xiao Luan, Jiezhong Zheng, and Weisheng Li. Learning unsupervised face normalization through frontal view reconstruction. IEEE Transactions on Circuits and Systems for Video Technology, 32(8):5201-5212, 2021. 2 + +[28] Bingjun Luo, Zewen Wang, Jinpeng Wang, Junjie Zhu, Xibin Zhao, and Yue Gao. Multi-energy guided image translation with stochastic differential equations for near-infrared facial expression recognition. In AAAI, pages 565-573, 2024. 2, 3 +[29] Mandi Luo, Haoxue Wu, Huaibo Huang, Weizan He, and Ran He. Memory-modulated transformer network for heterogeneous face recognition. IEEE Transactions on Information Forensics and Security, 17:2095-2109, 2022. 7, 8 +[30] Joel Ruben Antony Moniz, Christopher Beckham, Simon Rajotte, Sina Honari, and Chris Pal. Unsupervised depth estimation, 3d face rotation and replacement. Advances in neural information processing systems, 31, 2018. 6 +[31] Meng Pang, Yiu-Ming Cheung, Qiquan Shi, and Mengke Li. Iterative dynamic generic learning for face recognition from a contaminated single-sample per person. IEEE Transactions on Neural Networks and Learning Systems, 32(4): 1560-1574, 2020. 2 +[32] Meng Pang, Binghui Wang, Yiu-ming Cheung, Yiran Chen, and Bihan Wen. Vd-gan: A unified framework for joint prototype and representation learning from contaminated single sample per person. IEEE Transactions on Information Forensics and Security, 16:2246-2259, 2021. 1, 2 +[33] Meng Pang, Binghui Wang, Mang Ye, Yiu-ming Cheung, Yi ran Chen, and Bihan Wen. Disp+ v: A unified framework for disentangling prototype and variation from single sample per person. IEEE Transactions on Neural Networks and Learning Systems, 34(2):867-881, 2021. 2 +[34] Meng Pang, Binghui Wang, Siyu Huang, Yiu-Ming Cheung, and Bihan Wen. A unified framework for bidirectional prototype learning from contaminated faces across heterogeneous domains. IEEE Transactions on Information Forensics and Security, 17:1544-1557, 2022. 1 +[35] Yichen Qian, Weihong Deng, and Jiani Hu. Unsupervised face normalization with extreme pose and expression in the wild. In CVPR, pages 9851-9858, 2019. 6 +[36] Yue Que, Li Xiong, Weiguo Wan, Xue Xia, and Zhiwei Liu. Denoising diffusion probabilistic model for face sketch-to-photo synthesis. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2 +[37] Lingxiao Song, Zhihe Lu, Ran He, Zhenan Sun, and Tieniu Tan. Geometry guided adversarial facial expression synthesis. In ACM MM, pages 627-635, 2018. 2 +[38] Luan Tran, Xi Yin, and Xiaoming Liu. Disentangled representation learning gan for pose-invariant face recognition. In CVPR, pages 1415-1424, 2017. 6 +[39] Jingwei Xin, Zikai Wei, Nannan Wang, Jie Li, Xiaoyu Wang, and Xinbo Gao. Learning a high fidelity identity representation for face frontalization. IEEE Transactions on Circuits and Systems for Video Technology, 33(11):6952-6964, 2023. 1 +[40] Jianjin Xu, Saman Motamed, Praneetha Vaddamanu, Chen Henry Wu, Christian Haene, Jean-Charles Bazin, and Fernando De la Torre. Personalized face inpainting with diffusion models by parallel visual attention. In WACV, pages 5432–5442, 2024. 1, 2 +[41] Weihong Xu, Xiaohua Xie, and Jianhuang Lai. Relightgan: Instance-level generative adversarial network for face illumni + +nation transfer. IEEE Transactions on Image Processing, 30: 3450-3460, 2021. 1, 2 +[42] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4): 1-39, 2023. 2 +[43] Yiming Yang, Weipeng Hu, and Haifeng Hu. Neutral face learning and progressive fusion synthesis network for nir-vis face recognition. IEEE Transactions on Circuits and Systems for Video Technology, 33(10):5750-5763, 2024. 6, 7, 8 +[44] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? NeurIPS, 27, 2014. 4 +[45] Aijing Yu, Haoxue Wu, Huaibo Huang, Zhen Lei, and Ran He. Lamp-hq: A large-scale multi-pose high-quality database and benchmark for nir-vis face recognition. International Journal of Computer Vision, 129(5):1467-1483, 2021. 5 +[46] Ni Zhang, Nian Liu, Junwei Han, Kaiyuan Wan, and Ling Shao. Face de-occlusion with deep cascade guidance learning. IEEE Transactions on Multimedia, 25:3217-3229, 2022. 1, 2 +[47] Jian Zhao, Yu Cheng, Yan Xu, Lin Xiong, Jianshu Li, Fang Zhao, Karlekar Jayashree, Sugiri Pranata, Shengmei Shen, Junliang Xing, et al. Towards pose invariant face recognition in the wild. In CVPR, pages 2207-2216, 2018. 6 +[48] Hang Zhou, Jihao Liu, Ziwei Liu, Yu Liu, and Xiaogang Wang. Rotate-and-render: Unsupervised photorealistic face rotation from single-view images. In CVPR, pages 5911-5920, 2020. 6, 7 +[49] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, pages 2223-2232, 2017. 2 \ No newline at end of file diff --git a/2025/UMFN_ Unified Multi-Domain Face Normalization for Joint Cross-domain Prototype Learning and Heterogeneous Face Recognition/images.zip b/2025/UMFN_ Unified Multi-Domain Face Normalization for Joint Cross-domain Prototype Learning and Heterogeneous Face Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b878ab4d039f9b3924337f6b1a774d143d7ff3b4 --- /dev/null +++ b/2025/UMFN_ Unified Multi-Domain Face Normalization for Joint Cross-domain Prototype Learning and Heterogeneous Face Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2cc1fe5e105bbcc68b88aa2e7b3f640343c8a1ecbb8036c0dc961df639e731be +size 596697 diff --git a/2025/UMFN_ Unified Multi-Domain Face Normalization for Joint Cross-domain Prototype Learning and Heterogeneous Face Recognition/layout.json b/2025/UMFN_ Unified Multi-Domain Face Normalization for Joint Cross-domain Prototype Learning and Heterogeneous Face Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..db1d202bd228e6e0766efbd38593070b4cffcbd2 --- /dev/null +++ b/2025/UMFN_ Unified Multi-Domain Face Normalization for Joint Cross-domain Prototype Learning and Heterogeneous Face Recognition/layout.json @@ -0,0 +1,8748 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 76, + 102, + 536, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 102, + 536, + 140 + ], + "spans": [ + { + "bbox": [ + 76, + 102, + 536, + 140 + ], + "type": "text", + "content": "UMFN: Unified Multi-Domain Face Normalization for Joint Cross-domain Prototype Learning and Heterogeneous Face Recognition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 121, + 160, + 487, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 160, + 487, + 203 + ], + "spans": [ + { + "bbox": [ + 121, + 160, + 487, + 203 + ], + "type": "text", + "content": "Meng Pang1, Wenjun Zhang1, Nanrun Zhou2*, Shengbo Chen1*, Hong Rao1 \n1Nanchang University \n2Shanghai University of Engineering Science" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 74, + 205, + 533, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 205, + 533, + 217 + ], + "spans": [ + { + "bbox": [ + 74, + 205, + 533, + 217 + ], + "type": "text", + "content": "{mengpang@, zhangwenjun@, raohong@}ncu.edu.cn, nrzhou@sues.edu.cn, ccb02kingdom@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 245, + 200, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 245, + 200, + 257 + ], + "spans": [ + { + "bbox": [ + 152, + 245, + 200, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 270, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 558 + ], + "type": "text", + "content": "Face normalization aims to enhance the robustness and effectiveness of face recognition systems by mitigating intra-personal variations in expressions, poses, occlusions, illuminations, and domains. Existing methods face limitations in handling multiple variations and adapting to cross-domain scenarios. To address these challenges, we propose a novel Unified Multi-Domain Face Normalization Network (UMFN) model, which can process face images with various types of facial variations from different domains, and reconstruct frontal, neutral-expression facial prototypes in the target domain. As an unsupervised domain adaptation model, UMFN facilitates concurrent training on multiple datasets across domains and demonstrates strong prototype reconstruction capabilities. Notably, UMFN serves as a joint prototype and feature learning framework, enabling the simultaneous extraction of domain-agnostic identity features through a decoupling mapping network and a feature domain classifier for adversarial training. Moreover, we design an efficient Heterogeneous Face Recognition (HFR) network that fuses domain-agnostic and identity-discriminative features for HFR, and introduce contrastive learning to enhance identity recognition accuracy. Empirical studies on diverse cross-domain face datasets validate the effectiveness of our proposed method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 592, + 135, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 135, + 604 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 135, + 604 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 611, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 295, + 696 + ], + "type": "text", + "content": "Face normalization, as a core technology in the field of computer vision, is dedicated to eliminating interfering variation factors in facial images, such as pose, expression, occlusion, and illumination, and reconstructing a frontal, unoccluded, evenly illuminated facial prototype with a neutral expression [34]. This technology has shown broad application prospects in numerous areas, including criminal investiga" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 323, + 245, + 550, + 446 + ], + "blocks": [ + { + "bbox": [ + 323, + 245, + 550, + 446 + ], + "lines": [ + { + "bbox": [ + 323, + 245, + 550, + 446 + ], + "spans": [ + { + "bbox": [ + 323, + 245, + 550, + 446 + ], + "type": "image", + "image_path": "8985e2ee59487c050ba845e114c60d9a26ee5e63722dfb990b40c34bb2ba5d60.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 451, + 555, + 485 + ], + "lines": [ + { + "bbox": [ + 313, + 451, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 555, + 485 + ], + "type": "text", + "content": "Figure 1. Illustration of a desired face normalization method capable of handling multiple cross-domain face normalization tasks and addressing various types of facial variations." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 495, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 567 + ], + "type": "text", + "content": "tion , authentication, and human-computer interaction [39]. However, existing face normalization methods face challenges in coping with increasingly complex and diversified application scenarios, highlighting the urgent need for more advanced and versatile solutions to flexibly adapt to various changes in different environments and conditions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "The limitations of current face normalization methods lie mainly in their constraints in processing single contaminant elements and their single-domain applicability. Most methods [3, 10] focus on modeling and processing a single element, such as pose, expression, illumination, or occlusion, and struggle with multiple coexisting factors. Additionally, some unsupervised methods [41, 46] may inadvertently lose crucial identity information, leading to identity inconsistency between the generated prototype and the input face. Furthermore, existing models [32, 40] are typically domain-specific, lacking universality across multi-domain scenarios. However, practical applications often involve images" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 72, + 703, + 153, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 703, + 153, + 713 + ], + "spans": [ + { + "bbox": [ + 72, + 703, + 153, + 713 + ], + "type": "text", + "content": "*Corresponding authors" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29299" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "from various domains, such as visible light, near-infrared, and sketch images used in criminal investigations. Therefore, there is an urgent need for a multi-domain face normalization method capable of handling various facial variations, as illustrated in Fig. 1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 134, + 295, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 134, + 295, + 302 + ], + "spans": [ + { + "bbox": [ + 54, + 134, + 295, + 302 + ], + "type": "text", + "content": "To address these challenges, we propose a Unified Multidomain Face Normalization Network (UMFN) model. This model is capable of handling facial images from different domains and containing multiple contaminant elements (such as pose, expression, occlusion, etc.), and can reconstruct a frontal, neutral facial prototype in the target domain while simultaneously extracting identity features unrelated to domain information. It is worth noting that, despite the significant achievements of diffusion models in the field of image generation, their limitations in interpretable representation learning make them difficult to meet our needs for joint prototype and feature learning [42]. Therefore, this paper adopts a fusion framework of Autoencoder and Generative Adversarial Network (GAN), as shown in Fig. 2." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 304, + 295, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 295, + 542 + ], + "type": "text", + "content": "The UMFN model introduces a feature decoupling network and domain classifier for adversarial training, decoupling domain-agnostic features. By leveraging concatenated masks, it regulates clean prototype generation within specified domains or reconstructs original images, ensuring identity consistency. Additionally, a multi-scale InstanceNorm Patch-based discriminator [19] enhances image details and textures, improving visual quality. As an unsupervised domain adaptation model, UMFN enables concurrent training of diverse, unlabeled datasets within a single network, demonstrating flexibility and practicality in cross-domain prototype reconstruction. Furthermore, we design an efficient heterogeneous face recognition (HFR) network in conjunction with UMFN's encoder to extract identity-discriminative features for the generated prototype. These features are fused with domain-agnostic features of input for HFR. During HFR network training, a contrastive learning mechanism minimizes the feature distance between generated and authentic prototypes, enhancing the network's ability to accurately recognize identities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 544, + 286, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 544, + 286, + 555 + ], + "spans": [ + { + "bbox": [ + 67, + 544, + 286, + 555 + ], + "type": "text", + "content": "The contributions of the paper are summarized below:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 558, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 629 + ], + "type": "text", + "content": "- We propose a unified face normalization model, dubbed UMFN, which addresses the limitations of existing methods in handling multiple variation factors and cross-domain face images. This model is capable of simultaneously processing various types of facial variations and is applicable to face images from different domains." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 630, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 701 + ], + "type": "text", + "content": "- We design an efficient HFR network that fuses domain-agnostic features from input sample with identity-discriminative features from generated prototype for HFR. Furthermore, a contrastive learning mechanism is introduced to bolster the network's capability in accurately recognizing identities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "- As an unsupervised domain adaptation model, UMFN en" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 553, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 143 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 143 + ], + "type": "text", + "content": "ables simultaneous training of multiple datasets with different domains and without identity labels in a single network. It is able to perform prototype reconstruction tasks across multiple domains, enhancing model flexibility and practicality, adapting to diverse scenarios, and reducing reliance on labeled data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 144, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 144, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 314, + 144, + 553, + 191 + ], + "type": "text", + "content": "- Empirical study on multiple cross-domain face datasets with facial variations validates the superiority of UMFN in face normalization tasks in both single-modal and cross-modal scenarios, as well as in HFR task." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 203, + 405, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 203, + 405, + 216 + ], + "spans": [ + { + "bbox": [ + 314, + 203, + 405, + 216 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 223, + 553, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 553, + 402 + ], + "type": "text", + "content": "Face normalization, an emerging and highly esteemed research field, has shown substantial practical value in areas like criminal identification and facial evidence collection. Mainstream face normalization methods primarily address facial variation removal in single-modal scenarios, aiming to eliminate factors such as diverse lighting conditions, exaggerated facial expressions, large head poses, or partially occlusions, thereby restoring the standard facial prototype. Existing single-modal face normalization algorithms can be broadly classified into two categories [33]: The first constructs clustering models based on auxiliary sample information to estimate facial prototypes of contaminated samples. The second leverages generative AI models, such as GANs [8] and Diffusion Models [4], to learn the mapping between contaminated and standard facial samples." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 403, + 553, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 553, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 553, + 676 + ], + "type": "text", + "content": "In the first category, Gao et al. [7] combined unlabeled facial samples with a labeled training set, using a Gaussian Mixture Model for clustering to estimate prototypes. Pang et al. [31] introduced a more robust semi-supervised low-rank representation model for clustering and prototype estimation, updated with labeled retrieval samples. Although effective in prototype reconstruction, these methods require prior access to the facial samples, making them unsuitable for real-time scenarios. In the second category, benefiting from generative AI models' image generation and mapping capabilities, researchers have successively proposed a series of generative learning methods, aiming to eliminate expression variations [3, 37], illumination variations [10, 41], pose variations [15, 27, 46], and partial occlusions [21, 24, 40] in input facial samples, respectively, and generate facial prototype images with consistent identities. However, these methods are typically limited to specific types of facial variations and cannot effectively address combinations of multiple types. Pang et al. [32] attempts to solve prototype reconstruction for multiple types of contamination by designing a Variation Disentangling GAN (VD-GAN), but has limitations in preserving identity characteristics and handling only single-domain contamination." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Recently, a few of cross-domain face synthesis methods [1, 23, 26, 28, 36, 49] have been proposed, including Cycle-consistency GAN (CycleGAN) [49], Parallel Multi" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29300" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 553, + 240 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 553, + 240 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 553, + 240 + ], + "type": "image", + "image_path": "caa5ac8b6ec3f66c293a79d26b20b97aff27d1caa2ed286978203f688d1e02e4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 249, + 555, + 273 + ], + "lines": [ + { + "bbox": [ + 55, + 249, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 555, + 273 + ], + "type": "text", + "content": "Figure 2. Overview of the architecture of UMFN. " + }, + { + "bbox": [ + 55, + 249, + 555, + 273 + ], + "type": "inline_equation", + "content": "x_{c} \\in X_{c}" + }, + { + "bbox": [ + 55, + 249, + 555, + 273 + ], + "type": "text", + "content": " denotes an input image in domain " + }, + { + "bbox": [ + 55, + 249, + 555, + 273 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 249, + 555, + 273 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 249, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\hat{x}_c" + }, + { + "bbox": [ + 55, + 249, + 555, + 273 + ], + "type": "text", + "content": " represents the reconstructed image of " + }, + { + "bbox": [ + 55, + 249, + 555, + 273 + ], + "type": "inline_equation", + "content": "x_{c}" + }, + { + "bbox": [ + 55, + 249, + 555, + 273 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 249, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\hat{x}_{\\bar{c}}^{rp}" + }, + { + "bbox": [ + 55, + 249, + 555, + 273 + ], + "type": "text", + "content": " denotes the generated prototype of " + }, + { + "bbox": [ + 55, + 249, + 555, + 273 + ], + "type": "inline_equation", + "content": "x_{c}" + }, + { + "bbox": [ + 55, + 249, + 555, + 273 + ], + "type": "text", + "content": " in domain " + }, + { + "bbox": [ + 55, + 249, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\bar{c}" + }, + { + "bbox": [ + 55, + 249, + 555, + 273 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 249, + 555, + 273 + ], + "type": "inline_equation", + "content": "x_{c}^{rp}" + }, + { + "bbox": [ + 55, + 249, + 555, + 273 + ], + "type": "text", + "content": " represents the real prototype in domain " + }, + { + "bbox": [ + 55, + 249, + 555, + 273 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 249, + 555, + 273 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 286, + 296, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 286, + 296, + 406 + ], + "spans": [ + { + "bbox": [ + 54, + 286, + 296, + 406 + ], + "type": "text", + "content": "stage GANs (PMSGAN) [1], Brownian Bridge Diffusion Models (BBDM) [23], and NIR-FER Stochastic Differential Equations (NFER-SDE) [28]. These methods focus on transforming domain styles while preserving facial details, but they cannot remove facial variations from in-put faces. In light of the above, this paper aims to design a multidomain face normalization scheme applicable to universal facial variation removal and explore effective integration and optimization of domain transformation and prototype learning within a unified framework." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 417, + 117, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 417, + 117, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 117, + 430 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 437, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 296, + 581 + ], + "type": "text", + "content": "This paper proposes a universal face normalization framework, UMFN, whose architecture is shown in Fig. 2. UMFN is designed to uniformly process face images with various types of facial variations from different image domains, ultimately generating standard facial prototypes in the specified domain. Correspondingly, UMFN comprises the following two core components: 1) addressing the issue of input images originating from multiple domains and generating images in the specified domain; 2) uniformly handling multiple types of facial variations while generating prototypes that preserve the identity (ID). These two components will be detailed in Sec. 3.1 and Sec. 3.2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 588, + 295, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 295, + 612 + ], + "type": "text", + "content": "3.1. Feature Decoupling and Cross-Domain Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 689 + ], + "type": "text", + "content": "To tackle the challenge that input face images may come from different domains, we have designed a feature decoupling mapping network. This network is able to separate domain information from the latent space, resulting in domain-invariant representations. Such decoupled, domain-agnostic identity features can be applied to HFR task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Feature Decoupling: We design a feature decoupling mapping network that initially extracts primary feature " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "z =" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "inline_equation", + "content": "G_{enc}(x_c)" + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "text", + "content": " from the input image " + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "inline_equation", + "content": "x_c" + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "text", + "content": " using a pre-trained encoder " + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "inline_equation", + "content": "G_{enc}" + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "text", + "content": ", and then derives its latent feature representation " + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "inline_equation", + "content": "p = \\text{map}(z)" + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "text", + "content": " through the mapping network map. Specifically, the domain feature classifier " + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "inline_equation", + "content": "D_c" + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "text", + "content": " receives the latent representations " + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "text", + "content": " as input and outputs the domain prediction code (i.e., the image domain to which " + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "text", + "content": " belongs). The goal of " + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "inline_equation", + "content": "D_c" + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "text", + "content": " is to accurately predict the domain of the input latent representation " + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "text", + "content": ", while the mapping network map aims to confuse " + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "inline_equation", + "content": "D_c" + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "text", + "content": "'s correct domain prediction. Through adversarial training, the interplay between the two achieves domain information decoupling. Accordingly, the objective function for training " + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "inline_equation", + "content": "D_c" + }, + { + "bbox": [ + 313, + 287, + 555, + 431 + ], + "type": "text", + "content": " is presented as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 361, + 440, + 555, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 440, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 361, + 440, + 555, + 455 + ], + "type": "interline_equation", + "content": "L _ {D _ {c}} = E \\left[ \\log P \\left(l _ {p} ^ {\\text {p r e}} = l _ {x _ {c}} ^ {\\text {r e a l}} \\mid p\\right) \\right], \\tag {1}", + "image_path": "f8a1e3f41019ec2330e28b4dc50faab8a9311aa0c162ee972435802536324900.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 464, + 554, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 554, + 502 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 464, + 554, + 502 + ], + "type": "inline_equation", + "content": "l_{p}^{pre}" + }, + { + "bbox": [ + 313, + 464, + 554, + 502 + ], + "type": "text", + "content": " is the domain label predicted by " + }, + { + "bbox": [ + 313, + 464, + 554, + 502 + ], + "type": "inline_equation", + "content": "D_{c}" + }, + { + "bbox": [ + 313, + 464, + 554, + 502 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 464, + 554, + 502 + ], + "type": "inline_equation", + "content": "l_{x_{c}}^{real}" + }, + { + "bbox": [ + 313, + 464, + 554, + 502 + ], + "type": "text", + "content": " is the real domain label of " + }, + { + "bbox": [ + 313, + 464, + 554, + 502 + ], + "type": "inline_equation", + "content": "x_{c}" + }, + { + "bbox": [ + 313, + 464, + 554, + 502 + ], + "type": "text", + "content": ". On the other side, the objective function for training the mapping network map is as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 354, + 510, + 555, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 510, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 354, + 510, + 555, + 526 + ], + "type": "interline_equation", + "content": "L _ {m a p} = - E \\left[ \\log P \\left(l _ {p} ^ {p r e} = l _ {x _ {c}} ^ {r e a l} \\mid p\\right) \\right]. \\tag {2}", + "image_path": "63cff19897821f81611a9a6b5e22b18371dfe4a84088f432e2bba910175620f3.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 535, + 555, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 535, + 555, + 690 + ], + "spans": [ + { + "bbox": [ + 313, + 535, + 555, + 690 + ], + "type": "text", + "content": "Cross-domain Generation: We introduce the control mask and combine it with reconstruction learning to achieve cross-domain image generation. The core objective of reconstruction learning is to guide the decoder " + }, + { + "bbox": [ + 313, + 535, + 555, + 690 + ], + "type": "inline_equation", + "content": "G_{dec}" + }, + { + "bbox": [ + 313, + 535, + 555, + 690 + ], + "type": "text", + "content": " to learn how to decode domain-agnostic latent representation " + }, + { + "bbox": [ + 313, + 535, + 555, + 690 + ], + "type": "inline_equation", + "content": "p = \\text{map}(G_{enc}(x_c))" + }, + { + "bbox": [ + 313, + 535, + 555, + 690 + ], + "type": "text", + "content": " conditioned on the original attribute mask " + }, + { + "bbox": [ + 313, + 535, + 555, + 690 + ], + "type": "inline_equation", + "content": "v_c" + }, + { + "bbox": [ + 313, + 535, + 555, + 690 + ], + "type": "text", + "content": ", thereby reconstructing the input image " + }, + { + "bbox": [ + 313, + 535, + 555, + 690 + ], + "type": "inline_equation", + "content": "x_c" + }, + { + "bbox": [ + 313, + 535, + 555, + 690 + ], + "type": "text", + "content": ". This process ensures that the decoder not only learns to understand and apply the semantics of the mask but also fully utilizes the identity feature information in the latent representation " + }, + { + "bbox": [ + 313, + 535, + 555, + 690 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 535, + 555, + 690 + ], + "type": "text", + "content": ". As such, in subsequent steps, images from specified domains can be generated by manipulating the mask. The objective function for reconstruction learning is as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 367, + 700, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 700, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 367, + 700, + 555, + 715 + ], + "type": "interline_equation", + "content": "L _ {r e c} = E _ {x _ {c} \\sim p _ {d a t a}} \\left[ \\| x _ {c} - \\hat {x} _ {c} \\| _ {1} \\right], \\tag {3}", + "image_path": "5431cb6a3ed8468772fd6ca9a07ca2936a89fed6dc62668ced7cff474f0f8389.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29301" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\hat{x}_c = G_{dec}(p, v_c)" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " is the reconstructed image of " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "x_c" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "v_c" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " is the original attribute mask of the input image " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "x_c" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": ", indicating whether the " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "x_c" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " image is polluted and its domain information. Specifically, " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "v_c" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " is composed of a four-digit one-hot vector, where the first digit indicates whether the image is contaminated, and the last three digits represent the domain it belongs to. In Eq. (3), " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "l_1" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " norm is employed instead of the " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "l_2" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " norm to suppress the issue of blurred generated images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 175, + 178, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 178, + 186 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 178, + 186 + ], + "type": "text", + "content": "3.2. Prototype generation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 191, + 296, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 191, + 296, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 296, + 286 + ], + "type": "text", + "content": "Different from prior single-contamination normalization methods that overly focus on facial variation categories and how to accurately model them, our method returns to the essence of the prototype reconstruction task, aiming to directly generate identity-preserved standard facial prototypes by incorporating identity information from the contaminated facial samples, potentially bypassing the challenges of modeling facial variations in a clever manner." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 287, + 296, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 287, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 296, + 502 + ], + "type": "text", + "content": "Notably, we introduce a prototype adversarial loss, differing from the adversarial loss in traditional GAN-based image translation methods that aim to approximate the distribution of original face data with diverse variations. Our loss aims to approximate a condensed distribution of standard facial prototypes, while also reducing overfitting to these variations. Furthermore, we have designed a multiscale Markov discriminator " + }, + { + "bbox": [ + 55, + 287, + 296, + 502 + ], + "type": "inline_equation", + "content": "D = [D_w, D_f]" + }, + { + "bbox": [ + 55, + 287, + 296, + 502 + ], + "type": "text", + "content": " to refine the details of generated faces and stabilize training. Specifically, we concatenate the extracted domain-agnostic identity feature " + }, + { + "bbox": [ + 55, + 287, + 296, + 502 + ], + "type": "inline_equation", + "content": "p = \\text{map}(G_{\\text{enc}}(x_c))" + }, + { + "bbox": [ + 55, + 287, + 296, + 502 + ], + "type": "text", + "content": " with a control mask " + }, + { + "bbox": [ + 55, + 287, + 296, + 502 + ], + "type": "inline_equation", + "content": "v_{\\bar{c}}" + }, + { + "bbox": [ + 55, + 287, + 296, + 502 + ], + "type": "text", + "content": ", where the first bit in " + }, + { + "bbox": [ + 55, + 287, + 296, + 502 + ], + "type": "inline_equation", + "content": "v_{\\bar{c}}" + }, + { + "bbox": [ + 55, + 287, + 296, + 502 + ], + "type": "text", + "content": " controls whether to perform face normalization: 1 for yes, and the last three bits control the target domain for prototype generation. Next, this concatenated feature is input into the decoder to generate a pixel-level standardized facial prototype in the specified target domain. The prototype adversarial objectives for training the generator and the multi-scale discriminator are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 92, + 506, + 296, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 506, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 92, + 506, + 296, + 536 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {D} ^ {\\text {a d v}} = D _ {w} \\left(\\hat {x} _ {\\bar {c}} ^ {r p}\\right) + D _ {f} \\left(\\operatorname {c r o p} \\left(\\hat {x} _ {\\bar {c}} ^ {r p}\\right)\\right) \\tag {4} \\\\ - D _ {w} \\left(x _ {c} ^ {r p}\\right) - D _ {f} \\left(\\operatorname {c r o p} \\left(x _ {c} ^ {r p}\\right)\\right), \\\\ \\end{array}", + "image_path": "a85ea755726bf8494a65ea61548b2b63f86150de67bde209ce437b1510880763.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 540, + 295, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 540, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 86, + 540, + 295, + 555 + ], + "type": "interline_equation", + "content": "L _ {G _ {d e c}} ^ {a d v} = - \\left(D _ {w} \\left(\\hat {x} _ {\\bar {c}} ^ {r p}\\right) + D _ {f} \\left(\\operatorname {c r o p} \\left(\\hat {x} _ {\\bar {c}} ^ {r p}\\right)\\right)\\right), \\tag {5}", + "image_path": "77fb7b6dfda12bb9b72257df363404e353394a2c9322e3149d1076c76fc71a2d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 558, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 558, + 296, + 689 + ], + "type": "inline_equation", + "content": "D_w" + }, + { + "bbox": [ + 55, + 558, + 296, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 558, + 296, + 689 + ], + "type": "inline_equation", + "content": "D_f" + }, + { + "bbox": [ + 55, + 558, + 296, + 689 + ], + "type": "text", + "content": " represent the global path discriminator and local path discriminator, respectively. " + }, + { + "bbox": [ + 55, + 558, + 296, + 689 + ], + "type": "inline_equation", + "content": "\\hat{x}_{\\bar{c}}^{rp} = G_{dec}(map(G_{enc}(x_c), v_{\\bar{c}})" + }, + { + "bbox": [ + 55, + 558, + 296, + 689 + ], + "type": "text", + "content": " denotes the specified domain prototype generated by inputting " + }, + { + "bbox": [ + 55, + 558, + 296, + 689 + ], + "type": "inline_equation", + "content": "x_c" + }, + { + "bbox": [ + 55, + 558, + 296, + 689 + ], + "type": "text", + "content": " and the control mask " + }, + { + "bbox": [ + 55, + 558, + 296, + 689 + ], + "type": "inline_equation", + "content": "v_{\\bar{c}}" + }, + { + "bbox": [ + 55, + 558, + 296, + 689 + ], + "type": "text", + "content": " while " + }, + { + "bbox": [ + 55, + 558, + 296, + 689 + ], + "type": "inline_equation", + "content": "x_c^{rp}" + }, + { + "bbox": [ + 55, + 558, + 296, + 689 + ], + "type": "text", + "content": " represents the uncontaminated faces in the training set. The function \"crop\" is used to cut out the facial features from the input face images. Considering that the Wasserstein distance is more stable and can avoid training instability and mode collapse in GANs, we adopt the WGAN-GP [9] loss in our model to replace the traditional cross-entropy loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Furthermore, to ensure that the generated prototype " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\hat{x}_{\\bar{c}}^{rp}" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " can effectively preserve the identity information from the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "input " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "x_{c}" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": ", we have designed an identity feature similarity loss function. This is achieved by feeding the generated prototype back into the encoder and mapping network to obtain the latent feature map " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "(G_{enc}(\\hat{x_c}^{rp}))" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": ", and then calculating its similarity to the latent feature " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " of the input " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "x_{c}" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": ", thereby retaining identity. The loss function is as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 364, + 156, + 553, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 156, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 364, + 156, + 553, + 171 + ], + "type": "interline_equation", + "content": "L _ {i d} = \\| p - \\operatorname {m a p} \\left(G _ {e n c} \\left(\\hat {x} _ {\\bar {c}} ^ {r p}\\right)\\right) \\| _ {2} ^ {2}, \\tag {6}", + "image_path": "dcbb026c86ac3538d5b771f448e04d36418d150b7ab24c99313f36c1623859f8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 176, + 555, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 176, + 555, + 284 + ], + "spans": [ + { + "bbox": [ + 313, + 176, + 555, + 284 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 176, + 555, + 284 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_2^2" + }, + { + "bbox": [ + 313, + 176, + 555, + 284 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 313, + 176, + 555, + 284 + ], + "type": "inline_equation", + "content": "l_2" + }, + { + "bbox": [ + 313, + 176, + 555, + 284 + ], + "type": "text", + "content": "-norm. On the other hand, UMFN transfers the identity feature retention ability from the reconstruction learning task to the face prototype generation task. Since these two tasks share the same input and output domains, the transferability gap [44] between them is minimal. Therefore, the identity feature retention ability learned from the reconstruction learning task can be easily transferred to the prototype reconstruction task. Therefore, the overall loss function of the decoder " + }, + { + "bbox": [ + 313, + 176, + 555, + 284 + ], + "type": "inline_equation", + "content": "G_{dec}" + }, + { + "bbox": [ + 313, + 176, + 555, + 284 + ], + "type": "text", + "content": " is as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 380, + 293, + 553, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 293, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 380, + 293, + 553, + 308 + ], + "type": "interline_equation", + "content": "L _ {G _ {d e c}} ^ {\\text {a d v}} = \\lambda_ {1} L _ {r e c} + \\lambda_ {2} L _ {i d}. \\tag {7}", + "image_path": "48956f9867ead2fd0fa6d23b24a4ce6d5836f939cd4470133cb443b9c191c44f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 318, + 511, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 511, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 511, + 330 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 318, + 511, + 330 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 313, + 318, + 511, + 330 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 318, + 511, + 330 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 313, + 318, + 511, + 330 + ], + "type": "text", + "content": " are balancing hyperparameters." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 338, + 489, + 351 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 489, + 351 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 489, + 351 + ], + "type": "text", + "content": "3.3. Heterogeneous Face Recognition" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 356, + 554, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 554, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 554, + 416 + ], + "type": "text", + "content": "At this stage, we meticulously construct a HFR network, whose core components encompass an encoder " + }, + { + "bbox": [ + 313, + 356, + 554, + 416 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 313, + 356, + 554, + 416 + ], + "type": "text", + "content": " (its network architecture aligns with " + }, + { + "bbox": [ + 313, + 356, + 554, + 416 + ], + "type": "inline_equation", + "content": "G_{enc}" + }, + { + "bbox": [ + 313, + 356, + 554, + 416 + ], + "type": "text", + "content": " in UMFN) and a fully connected layer " + }, + { + "bbox": [ + 313, + 356, + 554, + 416 + ], + "type": "inline_equation", + "content": "l_{\\theta}" + }, + { + "bbox": [ + 313, + 356, + 554, + 416 + ], + "type": "text", + "content": ". The specific structure is shown in Fig. 3. The HFR network consists of two core operations:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "text", + "content": "Feature Fusion: During training, we concatenate the domain-agnostic features (extracted by UMFN's " + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "inline_equation", + "content": "G_{enc}" + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "text", + "content": " and mapping network map from the source domain input image) with the identity-discriminative features (extracted by encoder " + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "text", + "content": " from the real target domain prototype). The fused features are then fed into the softmax layer for classification. In this process, the fully connected layer " + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "inline_equation", + "content": "l_{\\theta}" + }, + { + "bbox": [ + 313, + 417, + 554, + 536 + ], + "type": "text", + "content": " doubles the feature dimensions, enhancing classification robustness through multi-dimensional identity information integration. The classification loss for this step is presented as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 546, + 553, + 560 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 546, + 553, + 560 + ], + "spans": [ + { + "bbox": [ + 321, + 546, + 553, + 560 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {c l s}} = \\operatorname {s o f t m a x} \\left(l _ {\\theta} \\left[ F _ {\\theta} \\left(x _ {v, i} ^ {r p}\\right) \\oplus \\operatorname {m a p} \\left(G _ {e n c} \\left(x _ {n, i}\\right)\\right) \\right], i\\right), \\tag {8}", + "image_path": "882ca5fa35345783aae81a0acfa275faff56b767c06c4454b339d4d129305c67.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "inline_equation", + "content": "x_{n,i}" + }, + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "text", + "content": " represents source domain image and " + }, + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "inline_equation", + "content": "x_{v,i}^{rp}" + }, + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "text", + "content": " represents real target domain prototype, " + }, + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "inline_equation", + "content": "F_{\\theta}(x_{v,i}^{rp})" + }, + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "text", + "content": " denotes the representation of the input image " + }, + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "inline_equation", + "content": "x_{v,i}^{rp}" + }, + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "text", + "content": " extracted by " + }, + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "text", + "content": ", while map " + }, + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "inline_equation", + "content": "(G_{enc}(x_{n,i}))" + }, + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "text", + "content": " denotes the domain-agnostic feature of the input image " + }, + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "inline_equation", + "content": "x_{n,i}" + }, + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "text", + "content": ", the symbol " + }, + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "text", + "content": " indicates the concatenation of these two features. During testing, we extract the identity-discriminative feature of the generated target domain prototype from the input test source domain image using the trained " + }, + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "text", + "content": ", and fuse it with the domain-agnostic feature of the input test source domain image for HFR." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "Contrastive Learning Mechanism: During the cross-domain generation, generated target domain prototypes may" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29302" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 71, + 286, + 315 + ], + "blocks": [ + { + "bbox": [ + 69, + 71, + 286, + 315 + ], + "lines": [ + { + "bbox": [ + 69, + 71, + 286, + 315 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 286, + 315 + ], + "type": "image", + "image_path": "3f40e077f6da44a76ea1795eb727d4d76c2469a3049e7eb57e36c47fc100f6c7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 321, + 288, + 333 + ], + "lines": [ + { + "bbox": [ + 63, + 321, + 288, + 333 + ], + "spans": [ + { + "bbox": [ + 63, + 321, + 288, + 333 + ], + "type": "text", + "content": "Figure 3. Overview of the HFR network and training process." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 348, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 296, + 552 + ], + "type": "text", + "content": "suffer identity feature loss or noise introduction. Direct reliance on such data for identity classification training can lead to overfitting to missing or noisy features, impairing classification performance. To mitigate this, we innovatively employ a contrastive learning mechanism that aligns generated prototypes with real ones in the feature space, while reducing the impact of noise and missing features. This ensures that the encoder " + }, + { + "bbox": [ + 55, + 348, + 296, + 552 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 55, + 348, + 296, + 552 + ], + "type": "text", + "content": " extracts features closely resembling real prototypes. Specifically, we randomly sample two pairs of real heterogeneous images, " + }, + { + "bbox": [ + 55, + 348, + 296, + 552 + ], + "type": "inline_equation", + "content": "(x_{v,1}^{rp}, x_{n,1})" + }, + { + "bbox": [ + 55, + 348, + 296, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 348, + 296, + 552 + ], + "type": "inline_equation", + "content": "(x_{v,2}^{rp}, x_{n,2})" + }, + { + "bbox": [ + 55, + 348, + 296, + 552 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 348, + 296, + 552 + ], + "type": "inline_equation", + "content": "x_{n,1}" + }, + { + "bbox": [ + 55, + 348, + 296, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 348, + 296, + 552 + ], + "type": "inline_equation", + "content": "x_{n,2}" + }, + { + "bbox": [ + 55, + 348, + 296, + 552 + ], + "type": "text", + "content": " are NIR images with contamination. We then use UMFN to perform cross-domain restoration on " + }, + { + "bbox": [ + 55, + 348, + 296, + 552 + ], + "type": "inline_equation", + "content": "x_{n,1}" + }, + { + "bbox": [ + 55, + 348, + 296, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 348, + 296, + 552 + ], + "type": "inline_equation", + "content": "x_{n,2}" + }, + { + "bbox": [ + 55, + 348, + 296, + 552 + ], + "type": "text", + "content": ", generating the corresponding target VIS domain prototypes " + }, + { + "bbox": [ + 55, + 348, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\hat{x}_{v,1}^{rp}" + }, + { + "bbox": [ + 55, + 348, + 296, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 348, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\hat{x}_{v,2}^{rp}" + }, + { + "bbox": [ + 55, + 348, + 296, + 552 + ], + "type": "text", + "content": ". Next, we form positive pairs " + }, + { + "bbox": [ + 55, + 348, + 296, + 552 + ], + "type": "inline_equation", + "content": "(x_{v,1}^{rp}, \\hat{x}_{v,1}^{rp})" + }, + { + "bbox": [ + 55, + 348, + 296, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 348, + 296, + 552 + ], + "type": "inline_equation", + "content": "(x_{v,2}^{rp}, \\hat{x}_{v,2}^{rp})" + }, + { + "bbox": [ + 55, + 348, + 296, + 552 + ], + "type": "text", + "content": ", and negative pairs " + }, + { + "bbox": [ + 55, + 348, + 296, + 552 + ], + "type": "inline_equation", + "content": "(x_{v,1}^{rp}, \\hat{x}_{v,2}^{rp})" + }, + { + "bbox": [ + 55, + 348, + 296, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 348, + 296, + 552 + ], + "type": "inline_equation", + "content": "(x_{v,2}^{rp}, \\hat{x}_{v,1}^{rp})" + }, + { + "bbox": [ + 55, + 348, + 296, + 552 + ], + "type": "text", + "content": ". The corresponding contrastive loss is as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 561, + 294, + 615 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 561, + 294, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 561, + 294, + 615 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {c o n t} = \\sum_ {j \\neq k} \\left(1 - \\langle f _ {j}, \\tilde {f} _ {j} \\rangle\\right) + \\left(1 - \\langle f _ {k}, \\tilde {f} _ {k} \\rangle\\right) \\\\ + \\max \\left(0, \\left\\langle f _ {j}, \\tilde {f} _ {k} \\right\\rangle - m\\right) + \\max \\left(0, \\left\\langle f _ {k}, \\tilde {f} _ {j} \\right\\rangle - m\\right), \\tag {9} \\\\ \\end{array}", + "image_path": "e39c11216c9032b64103a6e8653f2fe02825edb82f4e3a9551e2e9c0d1c03447.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 615, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 615, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 615, + 295, + 664 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 615, + 295, + 664 + ], + "type": "inline_equation", + "content": "f_{j} = l_{\\theta}[F_{\\theta}(x_{v,j}^{rp}) \\oplus map(G_{enc}(x_{n,j}))]" + }, + { + "bbox": [ + 55, + 615, + 295, + 664 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 615, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\langle, \\rangle" + }, + { + "bbox": [ + 55, + 615, + 295, + 664 + ], + "type": "text", + "content": " represents the calculation of the cosine similarity between two features, and " + }, + { + "bbox": [ + 55, + 615, + 295, + 664 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 615, + 295, + 664 + ], + "type": "text", + "content": " is a margin value set at 0.5. In summary, the final loss function of the HFR network is expressed as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 126, + 676, + 295, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 676, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 126, + 676, + 295, + 689 + ], + "type": "interline_equation", + "content": "L _ {h f r} = L _ {c l s} + \\alpha L _ {\\text {c o n t}}. \\tag {10}", + "image_path": "f94b809f0cb8579e4e29d5334626939d2831f699110252bcede91dff3949ebbd.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 701, + 189, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 189, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 189, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 701, + 189, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 701, + 189, + 713 + ], + "type": "text", + "content": " is a trade-off parameter." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 72, + 395, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 395, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 395, + 84 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 91, + 437, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 437, + 104 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 437, + 104 + ], + "type": "text", + "content": "4.1. Datasets Descriptions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 109, + 555, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 555, + 266 + ], + "type": "text", + "content": "LAMP-HQ [45] is a newly created, challenging large-scale NIR-VIS dataset comprising over 56,000 NIR and 16,000 VIS face images from 573 individuals, featuring diverse expressions, poses, lighting conditions, occlusions, and complex backgrounds. BUAA NIR-VIS [16] includes 150 identities, each with 9 NIR and 9 VIS images, covering various head poses and expressions. CUFSF [6] consists of 597 grayscale images paired with 597 real sketch images. CASIA NIR-VIS 2.0 [25] encompasses 725 identities, with 5-50 NIR and 1-22 VIS images per identity. IJB-A [22] and LFW [17] are widely-used benchmarks for face recognition and verification, featuring diverse poses, occlusions, and expressions in the visible domain." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 274, + 430, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 430, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 430, + 288 + ], + "type": "text", + "content": "4.2. Experimental Setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 293, + 554, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 554, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 554, + 364 + ], + "type": "text", + "content": "In the experimental section, we primarily conduct four experiments: (1) Homogeneous Multi-domain Face Normalization Experiment, (2) Heterogeneous Multi-domain Face Normalization Experiment, (3) Heterogeneous Face Recognition Experiment. (4) Ablation Study. The specific configurations of these experiments are as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 365, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 313, + 365, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 581 + ], + "type": "text", + "content": "1) Homogeneous Multi-domain Face Normalization Experiment: In this experiment, we utilize three types of domain data from the LAMP-HQ dataset to train our model, ensuring that all data originate from the same set of identities (IDs). The experiment primarily evaluates the model's performance in facial normalization across multi-domain data within the LAMP-HQ dataset. Specifically, we employ FP data from the LAMP-HQ dataset, where all NIR and VIS images from 100 IDs are used as the training set, while the remaining IDs are designated as the test set. As the LAMP-HQ dataset only contains NIR and VIS domain data, we use software to convert NIR photographs into sketch images, thereby simulating a scenario with three types of domain data. Furthermore, we transfer the model to other datasets, namely LFW and IJB-A, for testing. We assess the model's performance from both qualitative and quantitative perspectives and compare it with the latest single-modal face normalization techniques." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 581, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 554, + 689 + ], + "type": "text", + "content": "2) Heterogeneous Multi-domain Face Normalization Experiment: We use one type of domain data from three different datasets to demonstrate that the model can effectively work without paired data training, verifying its generalization ability across datasets. Specifically, we use VIS images of 250 IDs from the LAMP-HQ dataset as VIS domain data, all NIR images of 130 IDs from the BUAA dataset as NIR domain data, and Sketch images of 400 IDs from the CUFSF dataset as Sketch domain data for training." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "3) Heterogeneous Face Recognition Experiment: We conduct heterogeneous face recognition experiments on the" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29303" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 75, + 553, + 192 + ], + "blocks": [ + { + "bbox": [ + 61, + 75, + 553, + 192 + ], + "lines": [ + { + "bbox": [ + 61, + 75, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 61, + 75, + 553, + 192 + ], + "type": "image", + "image_path": "d61d0faa86dee5a206156c1629087a8a2efe2d838a0898ee3d0dc513be6fd665.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 196, + 438, + 207 + ], + "lines": [ + { + "bbox": [ + 170, + 196, + 438, + 207 + ], + "spans": [ + { + "bbox": [ + 170, + 196, + 438, + 207 + ], + "type": "text", + "content": "Figure 4. In-domain face normalization results on the LAMP-HQ dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 220, + 295, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 220, + 295, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 295, + 243 + ], + "type": "text", + "content": "LAMP-HQ and CASIA NIR-VIS 2.0 datasets, following the standard evaluation protocol in [43]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 245, + 295, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 245, + 295, + 292 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 295, + 292 + ], + "type": "text", + "content": "4) Ablation Study Experiment: We investigate the contribution of each component within the UMFN framework and explore the roles of the contrastive learning mechanism and feature fusion operation in the HFR network." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 299, + 295, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 295, + 312 + ], + "type": "text", + "content": "4.3. Implementation Details and Parameter Setting" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 316, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 316, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 54, + 316, + 295, + 460 + ], + "type": "text", + "content": "We employ a Resnet-50 pre-trained on the MS1M dataset and fine-tuned on VGGFace2 [2] as the encoder " + }, + { + "bbox": [ + 54, + 316, + 295, + 460 + ], + "type": "inline_equation", + "content": "G_{enc}" + }, + { + "bbox": [ + 54, + 316, + 295, + 460 + ], + "type": "text", + "content": " to extract 2048-dimensional identity features, keeping " + }, + { + "bbox": [ + 54, + 316, + 295, + 460 + ], + "type": "inline_equation", + "content": "G_{enc}" + }, + { + "bbox": [ + 54, + 316, + 295, + 460 + ], + "type": "text", + "content": " fixed during training and testing. The mapping network consists of multiple linear layers, inputting a 2048-dimensional vector and outputting a 2048-dimensional vector. The feature domain classifier " + }, + { + "bbox": [ + 54, + 316, + 295, + 460 + ], + "type": "inline_equation", + "content": "D_{c}" + }, + { + "bbox": [ + 54, + 316, + 295, + 460 + ], + "type": "text", + "content": " is composed of linear layers and outputs the predicted domain category, while the decoder " + }, + { + "bbox": [ + 54, + 316, + 295, + 460 + ], + "type": "inline_equation", + "content": "G_{dec}" + }, + { + "bbox": [ + 54, + 316, + 295, + 460 + ], + "type": "text", + "content": " outputs images of size " + }, + { + "bbox": [ + 54, + 316, + 295, + 460 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 54, + 316, + 295, + 460 + ], + "type": "text", + "content": ". The discriminators " + }, + { + "bbox": [ + 54, + 316, + 295, + 460 + ], + "type": "inline_equation", + "content": "D_{W}" + }, + { + "bbox": [ + 54, + 316, + 295, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 316, + 295, + 460 + ], + "type": "inline_equation", + "content": "D_{f}" + }, + { + "bbox": [ + 54, + 316, + 295, + 460 + ], + "type": "text", + "content": " are both patchGAN discriminators with a receptive field of 70, using instance normalization as the normalization layer, and the crop size is " + }, + { + "bbox": [ + 54, + 316, + 295, + 460 + ], + "type": "inline_equation", + "content": "150 \\times 150" + }, + { + "bbox": [ + 54, + 316, + 295, + 460 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "text", + "content": "The hyperparameters " + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "text", + "content": " in Eq. (7) are set at " + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = 10 - \\frac{cur_{epoch}}{sum_{epoch}} \\times 9" + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\lambda_{2} = 10" + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "inline_equation", + "content": "cur_{epoch}" + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "text", + "content": " is the current number of training epochs and " + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "inline_equation", + "content": "sum_{epoch}" + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "text", + "content": " is the total number of training epochs, which is 100. " + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "text", + "content": " in Eq. (10) is empirically set at 0.001. During optimization, we use the Adam optimizer, with the learning rate set at 0.0002 and momentum set at 0.5." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 552, + 282, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 282, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 282, + 565 + ], + "type": "text", + "content": "4.4. Homogeneous Multi-domain Normalization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "text", + "content": "In this section, we evaluate the performance of UMFN in homogeneous multi-domain face normalization using the LAMP-HQ dataset. The in-domain normalization results of UMFN are presented in Fig. 4. Fig. 4(a) showcases the results of inputting VIS domain faces corrupted by different facial variations such as pose, lighting, occlusion, and expression, and outputting normalized VIS domain faces. Fig. 4(b) displays the outcomes of inputting corrupted faces from the NIR and Sketch domains and outputting normalized faces in their respective domains. The bottom row serves as a reference for the true prototypes. From Fig. 4, it can be observed that UMFN is capable of simultaneously" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 220, + 553, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 553, + 365 + ], + "type": "text", + "content": "normalizing various types of corrupted faces from multiple different domains within their respective domains, effectively preserving the identity of the input faces while obtaining facial prototypes. Additionally, we conduct a comparison between UMFN and other popular single-domain face normalization methods, including PIM [47], TP-GAN [18], DR-GAN [38], Hassner [11], and FNM [35], using contaminated face images from the LFW dataset as input. The face normalization results of each method are presented in Fig. 5. It can be observed that the target prototype obtained through UMFN normalization excels in both image quality and identity preservation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 365, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 554, + 616 + ], + "type": "text", + "content": "Moreover, following the protocols in [48] and [20], we quantitatively evaluate UMFN on the LFW and IJB-A datasets. By inputting restored faces into the ResNet-18 [12] network for face verification, we aim to assess whether UMFN-restored faces enhance VR performance. Additionally, we report the Symmetry Index (SI)[30] to provide a comprehensive comparison of the quality of prototypes generated by UMFN, R&R [48], and CFR [20]. Tab. 1 presents the verification performance and SI of UMFN across both datasets. The results indicate that UMFN normalization significantly boosts the verification performance of the ResNet model: on IJB-A, when combined with ResNet-18, it improves VR by " + }, + { + "bbox": [ + 313, + 365, + 554, + 616 + ], + "type": "inline_equation", + "content": "5.35\\%" + }, + { + "bbox": [ + 313, + 365, + 554, + 616 + ], + "type": "text", + "content": ", with accuracy gains also observed on LFW. The experiments demonstrate that: (1) removing pose, expression, occlusion, and illumination variations from input faces allows the model to better focus on key identity features, and (2) UMFN effectively retains critical identity information while eliminating these facial variations. In addition, UMFN outperforms R&R and CFR in terms of SI, generating highly symmetrical faces that correct facial distortions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": "Furthermore, we illustrate the cross-domain normalization results of the UMFN model in Fig. 6. Specifically, the inputs consist of contaminated facial images from domain A, while the outputs represent the generated prototypes in domains B and C. The bottom row displays the reference prototypes of the inputs in the VIS domain. It can be observed that the UMFN model successfully addresses the challenges of domain transfer and prototype learning," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29304" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 294, + 148 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 294, + 148 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 294, + 148 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 294, + 148 + ], + "type": "image", + "image_path": "c958e0b0aee430106d624c974f01f1ca0a6a16dec2cd0a84f1f803690604af89.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 153, + 295, + 175 + ], + "lines": [ + { + "bbox": [ + 55, + 153, + 295, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 153, + 295, + 175 + ], + "type": "text", + "content": "Figure 5. Comparison with other single-domain face normalization methods on LFW dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 184, + 293, + 276 + ], + "blocks": [ + { + "bbox": [ + 58, + 184, + 293, + 276 + ], + "lines": [ + { + "bbox": [ + 58, + 184, + 293, + 276 + ], + "spans": [ + { + "bbox": [ + 58, + 184, + 293, + 276 + ], + "type": "table", + "html": "
MethodDatasets
LFWIJB-A
VR↑SI↓VR↑SI↓
Res18 [12]98.85-90.57±1.2-
Res18+R&R [48]98.9520.9391.98±0.720.85
Res18+CFR [20]99.2328.3193.36±0.624.76
Res18+UMFN(ours)99.5714.6195.92±0.715.81
", + "image_path": "e7ac4ca62da999ce48fea1eeae2d63e1f0426690a6604f7f741fe97acbbb72b9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 287, + 295, + 309 + ], + "lines": [ + { + "bbox": [ + 55, + 287, + 295, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 295, + 309 + ], + "type": "text", + "content": "Table 1. Verification performance and Symmetry Index (SI) on LFW and IJB-A." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 317, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 296, + 413 + ], + "type": "text", + "content": "effectively preserving the identity characteristics of the inputs when generating uncontaminated facial prototypes for the target domains. The inspiring results demonstrate that UMFN is capable of performing effective domain-specific face normalization on facial images sourced from different domains and contaminated by various types of facial variations, all while efficiently retaining the key identity information of the inputs, utilizing a single network architecture." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 421, + 285, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 285, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 285, + 433 + ], + "type": "text", + "content": "4.5. Heterogeneous Multi-domain Normalization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 438, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 713 + ], + "type": "text", + "content": "In this section, we evaluate the performance of UMFN in the heterogeneous multi-domain face normalization experiment. It is particularly noteworthy that there is currently no other face normalization method capable of adequately addressing this complex issue. Fig. 7 showcase the normalization results of images from the LAMP-HQ, BUAA, and CUFSF datasets across different domains (VIS, NIR, Sketch). We observe that regardless of the dataset (or domain) the input comes from, UMFN successfully generates prototypes that are highly matched to the target domain styles provided by another dataset. It effectively eliminates interfering factors such as pose, expression, and occlusion in the input face images while accurately preserving the input identity information. It is worth noting that due to the minimal facial variations in most images of the CUFSF dataset, which generally exhibit simple facial expressions such as smiles, the changes after normalization are primarily manifested in the mouth area. Furthermore, the challenging nature of cross-domain normalization tasks involving sketch images, which contain the least amount of facial detail information, leads to slight local distortions in the generated VIS and NIR prototypes. Despite these challenges, the results demonstrate UMFN's robust multi-" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 328, + 72, + 543, + 223 + ], + "blocks": [ + { + "bbox": [ + 328, + 72, + 543, + 223 + ], + "lines": [ + { + "bbox": [ + 328, + 72, + 543, + 223 + ], + "spans": [ + { + "bbox": [ + 328, + 72, + 543, + 223 + ], + "type": "image", + "image_path": "f1aa9232f686686956327a233a07e4997ed972262da688b36245f5e293920d3a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 229, + 555, + 275 + ], + "lines": [ + { + "bbox": [ + 313, + 229, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 555, + 275 + ], + "type": "text", + "content": "Figure 6. Cross-domain face normalization results on LAMP-HQ dataset. The second and third rows depict face normalization results in different domains from the inputs shown in the first row. The bottom row shows the real VIS prototypes for reference." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 290, + 555, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 555, + 326 + ], + "type": "text", + "content": "domain and multi-type face normalization capabilities, particularly when trained on unpaired datasets from multiple domains, where its performance remains impressive." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 334, + 491, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 491, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 491, + 346 + ], + "type": "text", + "content": "4.6. Heterogeneous Face Recognition" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 352, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 555, + 483 + ], + "type": "text", + "content": "In this experiment, we employ UMFN to restore cross-domain prototypes from NIR data in the probe set, generating VIS prototypes. Subsequently, we utilize the trained HFR network to obtain fused features (i.e., concatenation of domain-agnostic and identity-discriminative features) for both these generated prototypes and the authentic prototypes in the gallery set. Finally, by computing the cosine similarity between the fused features of the generated prototypes and those of the gallery set, we obtain a similarity matrix, upon which we calculate the rank-1 recognition accuracy and VR at " + }, + { + "bbox": [ + 313, + 352, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\mathrm{FAR} = 0.001" + }, + { + "bbox": [ + 313, + 352, + 555, + 483 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 484, + 556, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 556, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 556, + 662 + ], + "type": "text", + "content": "Tab. 2 presents a comparison of our method with other advanced HFR approaches, including PACH [5], AD-CANs [13], MMTN [29], NLPF [43], and LPL [14], on LAMP-HQ and CASIA NIR-VIS 2.0 datasets. It is observed that our method achieves the best performance on both datasets in terms of rank-1 accuracy and VR. This exceptional performance stems from: (1) Incorporating a contrastive learning mechanism that aligns generated prototypes with real ones in the feature space, reducing noise and missing feature impacts. This allows the HFR encoder " + }, + { + "bbox": [ + 313, + 484, + 556, + 662 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 313, + 484, + 556, + 662 + ], + "type": "text", + "content": " to extract identity-discriminative features from generated prototypes that closely resembling real ones. (2) The fusion of domain-agnostic and identity-discriminative features notably enhances the identity distinguishability of the combined features." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 672, + 408, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 408, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 408, + 685 + ], + "type": "text", + "content": "4.7. Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "To assess the effectiveness of UMFN and the contributions of its components, we train five variants by respectively re" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "29305" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 216, + 220 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 216, + 220 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 216, + 220 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 216, + 220 + ], + "type": "image", + "image_path": "dc85ed7baf14bac2b617006ebe1886ea12340f895ff82741bd8704f8bb721c5c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 220, + 197, + 228 + ], + "lines": [ + { + "bbox": [ + 83, + 220, + 197, + 228 + ], + "spans": [ + { + "bbox": [ + 83, + 220, + 197, + 228 + ], + "type": "text", + "content": "(a) Results on LAMP-HQ dataset" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 74, + 234, + 534, + 246 + ], + "lines": [ + { + "bbox": [ + 74, + 234, + 534, + 246 + ], + "spans": [ + { + "bbox": [ + 74, + 234, + 534, + 246 + ], + "type": "text", + "content": "Figure 7. The normalization results of face images from the LAMP-HQ, BUAA, and CUFSF datasets across different domains." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 228, + 74, + 383, + 219 + ], + "blocks": [ + { + "bbox": [ + 228, + 74, + 383, + 219 + ], + "lines": [ + { + "bbox": [ + 228, + 74, + 383, + 219 + ], + "spans": [ + { + "bbox": [ + 228, + 74, + 383, + 219 + ], + "type": "image", + "image_path": "cb414ba2e0d25491404cee4d865adef3f99eb6809f4d3d19b895708a06a974ec.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 220, + 361, + 228 + ], + "lines": [ + { + "bbox": [ + 261, + 220, + 361, + 228 + ], + "spans": [ + { + "bbox": [ + 261, + 220, + 361, + 228 + ], + "type": "text", + "content": "(b) Results on BUAA dataset" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 398, + 73, + 554, + 219 + ], + "blocks": [ + { + "bbox": [ + 398, + 73, + 554, + 219 + ], + "lines": [ + { + "bbox": [ + 398, + 73, + 554, + 219 + ], + "spans": [ + { + "bbox": [ + 398, + 73, + 554, + 219 + ], + "type": "image", + "image_path": "d76b988d75d360a3a8524248721da92d7fb3716fc24c53bee718b7616d785c62.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 431, + 220, + 535, + 228 + ], + "lines": [ + { + "bbox": [ + 431, + 220, + 535, + 228 + ], + "spans": [ + { + "bbox": [ + 431, + 220, + 535, + 228 + ], + "type": "text", + "content": "(c) Results on CUFSF dataset" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 58, + 258, + 293, + 355 + ], + "blocks": [ + { + "bbox": [ + 58, + 258, + 293, + 355 + ], + "lines": [ + { + "bbox": [ + 58, + 258, + 293, + 355 + ], + "spans": [ + { + "bbox": [ + 58, + 258, + 293, + 355 + ], + "type": "table", + "html": "
MethodDatasets
LAMP-HQCASIANIR-VIS 2.0
Rank-1@FAR=0.001Rank-1@FAR=0.001
PACH [5] (2020)96.978.798.998.3
ADCANs [13] (2021)97.896.199.198.5
MMTN [29] (2022)97.779.299.498.7
NLPF [43] (2024)98.898.399.799.5
LPL [14] (2024)98.798.599.697.5
UMFN (ours)99.198.799.999.7
", + "image_path": "adca369babe2e45d1d31b3af6ecd1a7ad54cbf1a2a0f7de3f59bce00397a13a7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 58, + 396, + 294, + 467 + ], + "blocks": [ + { + "bbox": [ + 55, + 365, + 295, + 388 + ], + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 388 + ], + "type": "text", + "content": "Table 2. Rank-1 accuracy and verification rate on the LAMP-HQ NIR-VIS dataset and the CASIA NIR-VIS 2.0 dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 396, + 294, + 467 + ], + "lines": [ + { + "bbox": [ + 58, + 396, + 294, + 467 + ], + "spans": [ + { + "bbox": [ + 58, + 396, + 294, + 467 + ], + "type": "table", + "html": "
Feature fusionContrastiveDatasets
LAMP-HQCASIANIR-VIS 2.0
Rank-1@FAR=0.001Rank-1@FAR=0.001
91.287.695.795.1
✓94.790.597.596.7
✓97.395.398.798.2
✓✓99.198.799.999.7
", + "image_path": "4a8bfd578af0da5e2ed26834dca44ef172a5c7767f2493a7c194a52d0e9cb09f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 476, + 295, + 499 + ], + "lines": [ + { + "bbox": [ + 55, + 476, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 476, + 295, + 499 + ], + "type": "text", + "content": "Table 3. Ablation experiments of the contrastive learning mechanism and feature fusion in the HFR network." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 509, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 509, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 54, + 509, + 295, + 664 + ], + "type": "text", + "content": "moving the identity feature similarity loss " + }, + { + "bbox": [ + 54, + 509, + 295, + 664 + ], + "type": "inline_equation", + "content": "L_{id}" + }, + { + "bbox": [ + 54, + 509, + 295, + 664 + ], + "type": "text", + "content": ", reconstruction loss " + }, + { + "bbox": [ + 54, + 509, + 295, + 664 + ], + "type": "inline_equation", + "content": "L_{rec}" + }, + { + "bbox": [ + 54, + 509, + 295, + 664 + ], + "type": "text", + "content": ", feature decoupling loss " + }, + { + "bbox": [ + 54, + 509, + 295, + 664 + ], + "type": "inline_equation", + "content": "L_{map}" + }, + { + "bbox": [ + 54, + 509, + 295, + 664 + ], + "type": "text", + "content": " in the mapping network, global path discriminator " + }, + { + "bbox": [ + 54, + 509, + 295, + 664 + ], + "type": "inline_equation", + "content": "D_w" + }, + { + "bbox": [ + 54, + 509, + 295, + 664 + ], + "type": "text", + "content": ", and the local path discriminator " + }, + { + "bbox": [ + 54, + 509, + 295, + 664 + ], + "type": "inline_equation", + "content": "D_f" + }, + { + "bbox": [ + 54, + 509, + 295, + 664 + ], + "type": "text", + "content": ". The face normalization results for these variants are presented in Fig. 8. Observations indicate that: (1) Without " + }, + { + "bbox": [ + 54, + 509, + 295, + 664 + ], + "type": "inline_equation", + "content": "L_{id}" + }, + { + "bbox": [ + 54, + 509, + 295, + 664 + ], + "type": "text", + "content": ", generated prototypes fail to preserve identity characteristics; (2) The absence of " + }, + { + "bbox": [ + 54, + 509, + 295, + 664 + ], + "type": "inline_equation", + "content": "L_{rec}" + }, + { + "bbox": [ + 54, + 509, + 295, + 664 + ], + "type": "text", + "content": " causes control masks to fail, preventing the generation of specified domain prototypes; (3) Lack of " + }, + { + "bbox": [ + 54, + 509, + 295, + 664 + ], + "type": "inline_equation", + "content": "L_{map}" + }, + { + "bbox": [ + 54, + 509, + 295, + 664 + ], + "type": "text", + "content": " results in the retention of original domain information in cross-domain prototypes; (4) Lack of " + }, + { + "bbox": [ + 54, + 509, + 295, + 664 + ], + "type": "inline_equation", + "content": "D_w" + }, + { + "bbox": [ + 54, + 509, + 295, + 664 + ], + "type": "text", + "content": " results in incomplete faces with poorly generated visual effects; (5) Removing " + }, + { + "bbox": [ + 54, + 509, + 295, + 664 + ], + "type": "inline_equation", + "content": "D_f" + }, + { + "bbox": [ + 54, + 509, + 295, + 664 + ], + "type": "text", + "content": " leads to a loss of local details in generated prototypes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "type": "text", + "content": "Furthermore, Tab. 3 quantitatively analyzes the impact of the contrastive learning mechanism and feature concatenation on HFR network training. Results show that contrastive learning significantly enhances the network's uti" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 329, + 260, + 543, + 473 + ], + "blocks": [ + { + "bbox": [ + 329, + 260, + 543, + 473 + ], + "lines": [ + { + "bbox": [ + 329, + 260, + 543, + 473 + ], + "spans": [ + { + "bbox": [ + 329, + 260, + 543, + 473 + ], + "type": "image", + "image_path": "442b66f559b7f2faf407945e0ba9f61f92a6503875cf5027d5bf582fc3230232.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 478, + 539, + 491 + ], + "lines": [ + { + "bbox": [ + 329, + 478, + 539, + 491 + ], + "spans": [ + { + "bbox": [ + 329, + 478, + 539, + 491 + ], + "type": "text", + "content": "Figure 8. Face normalization results of UMFN's variants." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 498, + 555, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 559 + ], + "type": "text", + "content": "lization of generated prototypes, allowing it to extract key features similar to real prototypes while reducing the effects of noise and feature loss. Additionally, fusion of domain-invariant and identity-discriminative features improves classification performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 572, + 388, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 572, + 388, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 572, + 388, + 586 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "This paper has proposed the new UMFN model. It effectively addresses the limitations in handling diverse facial variations and cross-domain adaptability by reconstructing frontal, neutral-expression facial prototypes, and enabling unsupervised domain adaptation. Through joint prototype and feature learning, along with a well-designed HFR network, our method significantly enhances identity recognition accuracy across various domains. Experiments have validated UMFN's superiority in single-modal and cross-modal face normalization tasks, as well as in HFR task." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29306" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 202 + ], + "type": "text", + "content": "This work was supported in part by the Natural Science Foundation of China (62466036, 62162041), by the Natural Science Foundation of Jiangxi Province (20232BAB212025), by the Natural Science Foundation of Henan Province (242300421402), by the High-level and Urgently Needed Overseas Talent Programs of Jiangxi Province (20232BCJ25024, 20232BCJ25026), by the Science&Technology Planning Project of Shanghai (23010501800), and by the Key Research and Development Program of Jiangxi Province (20243BBG71035)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 225, + 115, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 225, + 115, + 238 + ], + "spans": [ + { + "bbox": [ + 56, + 225, + 115, + 238 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 245, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 245, + 296, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 245, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 61, + 245, + 296, + 277 + ], + "type": "text", + "content": "[1] Bing Cao, Nannan Wang, Jie Li, Qinghua Hu, and Xinbo Gao. Face photo-sketch synthesis via full-scale identity supervision. Pattern Recognition, 124:108446, 2022. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 280, + 296, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 280, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 61, + 280, + 296, + 312 + ], + "type": "text", + "content": "[2] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age. In FG, pages 67-74, 2018. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 315, + 296, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 315, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 61, + 315, + 296, + 346 + ], + "type": "text", + "content": "[3] William Carver and Ifeoma Nwogu. Facial expression neutralization with stoicnet. In WACV, pages 201-208, 2021. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 350, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 350, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 350, + 296, + 392 + ], + "type": "text", + "content": "[4] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):10850-10869, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 395, + 296, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 395, + 296, + 427 + ], + "spans": [ + { + "bbox": [ + 62, + 395, + 296, + 427 + ], + "type": "text", + "content": "[5] Boyan Duan, Chaoyou Fu, Yi Li, Xingguang Song, and Ran He. Cross-spectral face hallucination via disentangling independent factors. In CVPR, pages 7930-7938, 2020. 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 430, + 296, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 430, + 296, + 461 + ], + "spans": [ + { + "bbox": [ + 62, + 430, + 296, + 461 + ], + "type": "text", + "content": "[6] Hamed Kiani Galoogahi and Terence Sim. Face sketch recognition by local radon binary pattern: Lrbp. In ICIP, pages 1837-1840, 2012. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 464, + 296, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 296, + 507 + ], + "type": "text", + "content": "[7] Yuan Gao, Jiayi Ma, and Alan L Yuille. Semi-supervised sparse representation based classification for face recognition with insufficient labeled samples. IEEE Transactions on Image Processing, 26(5):2545-2560, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 510, + 296, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 510, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 62, + 510, + 296, + 552 + ], + "type": "text", + "content": "[8] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NeurIPS, 27, 2014. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 555, + 296, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 555, + 296, + 587 + ], + "spans": [ + { + "bbox": [ + 62, + 555, + 296, + 587 + ], + "type": "text", + "content": "[9] Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein gans. NeurIPS, 30, 2017. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 590, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 295, + 644 + ], + "type": "text", + "content": "[10] Dequan Guo, Lingrui Zhu, Shenggui Ling, Tianxiang Li, Gexiang Zhang, Qiang Yang, Ping Wang, Shiqi Jiang, Sidong Wu, and Junbao Liu. Face illumination normalization based on generative adversarial network. Natural Computing, 22(1):105-117, 2023. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 646, + 295, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 646, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 57, + 646, + 295, + 677 + ], + "type": "text", + "content": "[11] Tal Hassner, Shai Harel, Eran Paz, and Roee Enbar. Effective face frontalization in unconstrained images. In CVPR, pages 4295-4304, 2015. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "text", + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6, 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[13] Weipeng Hu and Haifeng Hu. Adversarial disentanglement spectrum variations and cross-modality attention networks for nir-vis face recognition. IEEE Transactions on Multimedia, 23:145-160, 2020. 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "type": "text", + "content": "[14] Weipeng Hu, Yiming Yang, and Haifeng Hu. Pseudo label association and prototype-based invariant learning for semi-supervised nir-vis face recognition. IEEE Transactions on Image Processing, 33:1448-1463, 2024. 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 162, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 194 + ], + "type": "text", + "content": "[15] Yibo Hu, Xiang Wu, Bing Yu, Ran He, and Zhenan Sun. Pose-guided photorealistic face rotation. In CVPR, pages 8398-8406, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "type": "text", + "content": "[16] D Huang, Jia Sun, and Y Wang. The buaa-visir face database instructions. School Comput. Sci. Eng., Beihang Univ., Beijing, China, Tech. Rep. IRIP-TR-12-FR-001, 3(3): 8, 2012. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "text", + "content": "[17] Gary B Huang, Marwan Mattar, Tamara Berg, and Eric Learned-Miller. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. In Workshop on faces in 'Real-Life' Images: detection, alignment, and recognition, 2008. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 294, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 337 + ], + "type": "text", + "content": "[18] Rui Huang, Shu Zhang, Tianyu Li, and Ran He. Beyond face rotation: Global and local perception gan for photorealistic and identity preserving frontal view synthesis. In ICCV, pages 2439-2448, 2017. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "type": "text", + "content": "[19] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In CVPR, pages 1125-1134, 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "text", + "content": "[20] Yeong-Joon Ju, Gun-Hee Lee, Jung-Ho Hong, and Seong-Whan Lee. Complete face recovery gan: Unsupervised joint face rotation and de-occlusion from a single-view image. In WACV, pages 3711–3721, 2022. 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "type": "text", + "content": "[21] Yeong-Joon Ju, Gun-Hee Lee, Jung-Ho Hong, and Seong-Whan Lee. Complete face recovery gan: Unsupervised joint face rotation and de-occlusion from a single-view image. In WACV, pages 3711–3721, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "type": "text", + "content": "[22] Brendan F Klare, Ben Klein, Emma Taborsky, Austin Blanton, Jordan Cheney, Kristen Allen, Patrick Grother, Alan Mah, and Anil K Jain. Pushing the frontiers of unconstrained face detection and recognition: Iarpa janus benchmark a. In CVPR, pages 1931-1939, 2015. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 515, + 553, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 548 + ], + "type": "text", + "content": "[23] Bo Li, Kaitao Xue, Bin Liu, and Yu-Kun Lai. Bbddm: Image-to-image translation with brownian bridge diffusion models. In CVPR, pages 1952-1961, 2023. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 548, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 590 + ], + "type": "text", + "content": "[24] Chenyu Li, Shiming Ge, Daichi Zhang, and Jia Li. Look through masks: Towards masked face recognition with deocclusion distillation. In ACM MM, pages 3016-3024, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 592, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 624 + ], + "type": "text", + "content": "[25] Stan Li, Dong Yi, Zhen Lei, and Shengcai Liao. The casia mir-vis 2.0 face database. In CVPRW, pages 348-353, 2013. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "text", + "content": "[26] Changcheng Liang, Mingrui Zhu, Nannan Wang, Heng Yang, and Xinbo Gao. Pmsgan: Parallel multistage gans for face image translation. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[27] Xiao Luan, Jiezhong Zheng, and Weisheng Li. Learning unsupervised face normalization through frontal view reconstruction. IEEE Transactions on Circuits and Systems for Video Technology, 32(8):5201-5212, 2021. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29307" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "text", + "content": "[28] Bingjun Luo, Zewen Wang, Jinpeng Wang, Junjie Zhu, Xibin Zhao, and Yue Gao. Multi-energy guided image translation with stochastic differential equations for near-infrared facial expression recognition. In AAAI, pages 565-573, 2024. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[29] Mandi Luo, Haoxue Wu, Huaibo Huang, Weizan He, and Ran He. Memory-modulated transformer network for heterogeneous face recognition. IEEE Transactions on Information Forensics and Security, 17:2095-2109, 2022. 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 207 + ], + "type": "text", + "content": "[30] Joel Ruben Antony Moniz, Christopher Beckham, Simon Rajotte, Sina Honari, and Chris Pal. Unsupervised depth estimation, 3d face rotation and replacement. Advances in neural information processing systems, 31, 2018. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "type": "text", + "content": "[31] Meng Pang, Yiu-Ming Cheung, Qiquan Shi, and Mengke Li. Iterative dynamic generic learning for face recognition from a contaminated single-sample per person. IEEE Transactions on Neural Networks and Learning Systems, 32(4): 1560-1574, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 319 + ], + "type": "text", + "content": "[32] Meng Pang, Binghui Wang, Yiu-ming Cheung, Yiran Chen, and Bihan Wen. Vd-gan: A unified framework for joint prototype and representation learning from contaminated single sample per person. IEEE Transactions on Information Forensics and Security, 16:2246-2259, 2021. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "type": "text", + "content": "[33] Meng Pang, Binghui Wang, Mang Ye, Yiu-ming Cheung, Yi ran Chen, and Bihan Wen. Disp+ v: A unified framework for disentangling prototype and variation from single sample per person. IEEE Transactions on Neural Networks and Learning Systems, 34(2):867-881, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 431 + ], + "type": "text", + "content": "[34] Meng Pang, Binghui Wang, Siyu Huang, Yiu-Ming Cheung, and Bihan Wen. A unified framework for bidirectional prototype learning from contaminated faces across heterogeneous domains. IEEE Transactions on Information Forensics and Security, 17:1544-1557, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 433, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 294, + 464 + ], + "type": "text", + "content": "[35] Yichen Qian, Weihong Deng, and Jiani Hu. Unsupervised face normalization with extreme pose and expression in the wild. In CVPR, pages 9851-9858, 2019. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 466, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 294, + 510 + ], + "type": "text", + "content": "[36] Yue Que, Li Xiong, Weiguo Wan, Xue Xia, and Zhiwei Liu. Denoising diffusion probabilistic model for face sketch-to-photo synthesis. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 511, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 294, + 544 + ], + "type": "text", + "content": "[37] Lingxiao Song, Zhihe Lu, Ran He, Zhenan Sun, and Tieniu Tan. Geometry guided adversarial facial expression synthesis. In ACM MM, pages 627-635, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "type": "text", + "content": "[38] Luan Tran, Xi Yin, and Xiaoming Liu. Disentangled representation learning gan for pose-invariant face recognition. In CVPR, pages 1415-1424, 2017. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "text", + "content": "[39] Jingwei Xin, Zikai Wei, Nannan Wang, Jie Li, Xiaoyu Wang, and Xinbo Gao. Learning a high fidelity identity representation for face frontalization. IEEE Transactions on Circuits and Systems for Video Technology, 33(11):6952-6964, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[40] Jianjin Xu, Saman Motamed, Praneetha Vaddamanu, Chen Henry Wu, Christian Haene, Jean-Charles Bazin, and Fernando De la Torre. Personalized face inpainting with diffusion models by parallel visual attention. In WACV, pages 5432–5442, 2024. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[41] Weihong Xu, Xiaohua Xie, and Jianhuang Lai. Relightgan: Instance-level generative adversarial network for face illumni" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 463 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "nation transfer. IEEE Transactions on Image Processing, 30: 3450-3460, 2021. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "text", + "content": "[42] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4): 1-39, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[43] Yiming Yang, Weipeng Hu, and Haifeng Hu. Neutral face learning and progressive fusion synthesis network for nir-vis face recognition. IEEE Transactions on Circuits and Systems for Video Technology, 33(10):5750-5763, 2024. 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 228 + ], + "type": "text", + "content": "[44] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? NeurIPS, 27, 2014. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 230, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 284 + ], + "type": "text", + "content": "[45] Aijing Yu, Haoxue Wu, Huaibo Huang, Zhen Lei, and Ran He. Lamp-hq: A large-scale multi-pose high-quality database and benchmark for nir-vis face recognition. International Journal of Computer Vision, 129(5):1467-1483, 2021. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "type": "text", + "content": "[46] Ni Zhang, Nian Liu, Junwei Han, Kaiyuan Wan, and Ling Shao. Face de-occlusion with deep cascade guidance learning. IEEE Transactions on Multimedia, 25:3217-3229, 2022. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "text", + "content": "[47] Jian Zhao, Yu Cheng, Yan Xu, Lin Xiong, Jianshu Li, Fang Zhao, Karlekar Jayashree, Sugiri Pranata, Shengmei Shen, Junliang Xing, et al. Towards pose invariant face recognition in the wild. In CVPR, pages 2207-2216, 2018. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 376, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 419 + ], + "type": "text", + "content": "[48] Hang Zhou, Jihao Liu, Ziwei Liu, Yu Liu, and Xiaogang Wang. Rotate-and-render: Unsupervised photorealistic face rotation from single-view images. In CVPR, pages 5911-5920, 2020. 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 420, + 553, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 420, + 553, + 463 + ], + "spans": [ + { + "bbox": [ + 316, + 420, + 553, + 463 + ], + "type": "text", + "content": "[49] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, pages 2223-2232, 2017. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29308" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/UMotion_ Uncertainty-driven Human Motion Estimation from Inertial and Ultra-wideband Units/20c112a8-f19a-4ffc-8e5b-a5d02b15caa1_content_list.json b/2025/UMotion_ Uncertainty-driven Human Motion Estimation from Inertial and Ultra-wideband Units/20c112a8-f19a-4ffc-8e5b-a5d02b15caa1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0dc6027b9e0ea0b18aa0a651a7f1112cc5ae6110 --- /dev/null +++ b/2025/UMotion_ Uncertainty-driven Human Motion Estimation from Inertial and Ultra-wideband Units/20c112a8-f19a-4ffc-8e5b-a5d02b15caa1_content_list.json @@ -0,0 +1,1571 @@ +[ + { + "type": "text", + "text": "UMotion: Uncertainty-driven Human Motion Estimation from Inertial and Ultra-wideband Units", + "text_level": 1, + "bbox": [ + 183, + 130, + 818, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Huakun Liu Hiroki Ota Xin Wei Yutaro Hirao", + "bbox": [ + 264, + 203, + 732, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Monica Perusquía-Hernández \nHideaki Uchiyama \nKiyoshi Kiyokawa \nNara Institute of Science and Technology, Japan", + "bbox": [ + 192, + 220, + 803, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{liu.huakun.li0, ota.hiroki.oc6, wei.xin.wy0, yutaro.hirao}@is.naist.jp", + "bbox": [ + 184, + 258, + 812, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{m.perusquia, hideaki.uchiyama, kiyo}@is.naist.jp", + "bbox": [ + 281, + 277, + 712, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sparse wearable inertial measurement units (IMUs) have gained popularity for estimating 3D human motion. However, challenges such as pose ambiguity, data drift, and limited adaptability to diverse bodies persist. To address these issues, we propose UMotion, an uncertainty-driven, online fusing-all state estimation framework for 3D human shape and pose estimation, supported by six integrated, body-worn ultra-wideband (UWB) distance sensors with IMUs. UWB sensors measure inter-node distances to infer spatial relationships, aiding in resolving pose ambiguities and body shape variations when combined with anthropometric data. Unfortunately, IMUs are prone to drift, and UWB sensors are affected by body occlusions. Consequently, we develop a tightly coupled Unscented Kalman Filter (UKF) framework that fuses uncertainties from sensor data and estimated human motion based on individual body shape. The UKF iteratively refines IMU and UWB measurements by aligning them with uncertain human motion constraints in real-time, producing optimal estimates for each. Experiments on both synthetic and real-world datasets demonstrate the effectiveness of UMotion in stabilizing sensor data and the improvement over state of the art in pose accuracy. Code is available at: https://github.com/kk9six/umotion.", + "bbox": [ + 89, + 359, + 485, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 753, + 223, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Estimating 3D human motion from wearable sensors has become increasingly popular due to their portability, accessibility, and versatility. Wearable sensors, such as inertial measurement units (IMUs), enable continuous monitoring of body motion measurements across unrestricted spaces. These advances shift motion capture from controlled laboratory environments to everyday settings [13, 19, 32, 41]. This transition benefits fields such as healthcare, sports", + "bbox": [ + 89, + 779, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/09b2e00879f5d00521d89acf4a37ac88a9833eb37f7d362bf7e3cd28695e2e25.jpg", + "image_caption": [ + "Figure 1. UMotion integrates IMU-UWB data inputs and pose outputs uniformly under uncertainty, constrained by individual body structure. The online state estimation framework iteratively refines sensor data confidence and stabilizes pose estimation, reducing ambiguities and improving robustness." + ], + "image_footnote": [], + "bbox": [ + 566, + 328, + 864, + 455 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "performance, ergonomics, and emerging areas in humancomputer interaction [3, 8, 14, 31].", + "bbox": [ + 511, + 566, + 906, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One of the widely chosen wearable sensors for 3D human motion estimation is IMU. Commercial systems, such as Xsens [27], utilize 17 or more IMUs for comprehensive pose coverage. While highly accurate, these densely placed IMUs are inconvenient and intrusive. Recent studies have reduced the required number of IMUs to just six—placed on the forearms, lower legs, pelvis, and head—while still achieving promising performance through data-driven methods [1, 7, 10, 31, 32, 34, 35, 37, 39]. With fewer IMUs, pose estimation becomes under-constrained and prone to ambiguity. Recent work has attempted to disambiguate poses by incorporating temporal consistency, physics-based constraints, or additional, easy-to-integrate sensors [1, 10, 35]. However, these methods still face significant challenges, including noisy sensor data, body shape variations, and ambiguities arising from under-constrained setups with sparse sensors.", + "bbox": [ + 509, + 597, + 908, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we propose UMotion, an uncertainty-driven human motion estimation framework that combines online state estimation with an integrated system of six inertial and", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7085", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ultra-wideband (UWB) distance sensors. Prior work often infers poses from sensor data without considering the reverse influence. UMotion uniformly treats input and output under uncertainty and body-specific constraints, blending data to optimize estimates (see Fig. 1). Our approach maintains inter-sensor distances as a core system state, capturing spatial relationships between sensor nodes. These distances, combined with basic anthropometric measurements (height and weight) and IMU data, serve as inputs for our learning-based shape and pose estimators. Since sensors are inherently noisy, we propose a tightly coupled online state estimation system that associates IMU and UWB sensor measurements and pose estimates using an Unscented Kalman Filter (UKF) and an uncertainty propagation method, iteratively correcting errors and stabilizing pose estimation in real-time. Our main contributions are:", + "bbox": [ + 89, + 90, + 480, + 330 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- An ensemble learning-based human shape estimation approach using distance constraints and anthropometric measurements from six integrated IMU-UWB sensors.", + "- A learning-based method for human pose estimation from inertial and distance constraints, incorporating mesh distribution estimation.", + "- A filtering-based state estimation system that couples sensor measurements, pose estimates, and body variations in real-time to improve sensor stability and pose accuracy." + ], + "bbox": [ + 89, + 333, + 478, + 465 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 482, + 230, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pose Estimation from Sparse Inertial Sensors In contrast to the predominant use of vision-based methods, wearable sensors, specifically IMUs, offer greater freedom and flexibility. Von Marcard et al. [32] present SIP that makes inertial pose estimation practical by using only six IMUs attached to the body, combined with an offline iterative SMPL body model [17] pose optimization. Huang et al. [7] propose a real-time pose estimation method, DIP, which uses a bidirectional recurrent neural network (biRNN) [28] to learn the mapping from a sequence of six IMU measurements to SMPL body poses. To obtain sufficient training data, they synthesize IMU measurements from the AMASS dataset [18], further advancing the development of learning-based methods for sparse sensor motion capture.", + "bbox": [ + 89, + 508, + 482, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Following previous studies, TransPose [34] refines pose estimation by decomposing the end-to-end framework into a multi-stage process with intermediate joint position estimation. To disambiguate the poses, PIP [35] proposes a physics-based motion optimization, while TIP [10] incorporates a conditional Transformer model for plausible terrain generation. Training data is crucial for learning-based methods. Unlike prior studies relying on synthetic IMU data, DynalIP [39] adapts real IMU data from diverse human skeleton formats to the target SMPL model and shows superior performance when training with real sensor data. Similarly focusing on the data, PNP [37] recently addresses", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "limitations in existing IMU synthesis by incorporating noninertial effects and fictitious forces, enhancing estimation robustness through a physics-informed neural network and realistic IMU synthesis.", + "bbox": [ + 511, + 90, + 903, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast to methods focusing solely on pose estimation, our approach treats human shape and pose as equally important. By integrating shape information, we establish a tight connection between sensor data and estimated motions, forming a positive feedback loop that enhances the entire process. Additionally, most previous studies fail to fully utilize IMU accelerations due to high noise and drift issues. Within our framework, accelerations serve as control inputs for state estimation, undergoing continuous correction and acting as a crucial component in tracking spatial relationships among sensors.", + "bbox": [ + 511, + 152, + 905, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pose Estimation from Hybrid Sensors In addition to IMUs, various wearable and hybrid sensor systems have been explored to overcome inherent IMU limitations such as restricted positional awareness [9, 12, 13, 15, 16, 21–23, 25, 36]. The closest works to ours are SmartPoser [4] and Ultra Inertial Poser (UIP) [1], both of which integrate UWB and IMU sensors for pose estimation. UWB sensors complement IMUs with additional distance information between sensors while preserving the portability and flexibility of tracking devices. SmartPoser [4] focuses on wearable arm pose estimation, combining UWB measurements with IMU data using an off-the-shelf smartwatch and smartphone. UIP [1] integrates six UWB sensors with IMUs for full-body pose estimation. Inter-sensor distances, estimated using an EKF that fuses IMU and UWB data, help reduce global translation drift and minimize position jitter compared to inertial-only tracking. However, body occlusion frequently affects certain node pairs, e.g., head-knee, causing the EKF to fail in estimating distances due to rapid IMU drift and unreliable UWB measurements, resulting in unstable distance data. Our approach addresses this limitation by using output pose uncertainties and body-specific constraints as observations to refine input sensor measurements, improving the stability of both IMU and UWB data.", + "bbox": [ + 511, + 339, + 906, + 704 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 718, + 604, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 743, + 653, + 757 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SMPL body model We use the SMPL model [17] to represent human motion. SMPL decomposes the human body into pose parameters, $\\pmb{\\theta} \\in \\mathbb{R}^{23 \\times 3 + 3}$ , which define relative rotations of 23 joints and the global root joint orientation in axis-angle form, and shape parameters, $\\beta \\in \\mathbb{R}^{10}$ , capturing body shape variations across individuals. The model is defined by a linear blend-skinning (LBS) function:", + "bbox": [ + 511, + 765, + 903, + 871 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nM (\\boldsymbol {\\beta}, \\boldsymbol {\\theta}) = W \\left(\\bar {\\mathbf {T}} + B _ {s} (\\boldsymbol {\\beta}) + B _ {p} (\\boldsymbol {\\theta}), J (\\boldsymbol {\\beta}), \\boldsymbol {\\theta}, \\mathcal {W}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 883, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7086", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $B_{s}(\\boldsymbol {\\beta})$ and $B_{p}(\\boldsymbol {\\theta})$ are shape and pose blend shapes that deform the template mesh $\\bar{\\mathbf{T}}$ in the zero pose. The mesh is then reposed according to specified poses $\\theta$ , combining with joint positions $J(\\beta)$ and blend-skinning weights $\\mathcal{W}$ . Inertial Measurement Unit (IMU) A nine-axis MEMS IMU comprises an accelerometer, gyroscope, and magnetometer, measuring accelerations in the sensor local frame $F^S$ , angular velocities in $F^S$ , and magnetic field strengths relative to the Earth's magnetic field, respectively. A commonly used error model for measured accelerations of one IMU, $\\pmb{a}\\in \\mathbb{R}^{3}$ , is given by [26]:", + "bbox": [ + 89, + 90, + 480, + 257 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {a} = \\boldsymbol {a} _ {\\text {t r u e}} + \\boldsymbol {b} _ {a} + \\boldsymbol {n} _ {a}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 268, + 482, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pmb{a}_{\\mathrm{true}}$ is the true acceleration, $\\pmb{b}_a$ is the bias, modeled as a random walk process with noise $\\eta$ following a zero-mean Gaussian distribution with standard deviation $\\Sigma_{b}$ , and $\\pmb{n}_a$ is high-frequency white noise. In our work, after a frame calibration process, accelerations are represented as $\\pmb{a}^{M}$ in the SMPL body-centric frame $F^{M}$ , and orientations are represented as $\\pmb{R}^{MB}$ , indicating the rotation from bone frame $F^{B}$ to $F^{M}$ . For simplicity, we denote $\\pmb{a}$ as $\\pmb{a}^{M}$ and $\\pmb{R}$ as $\\pmb{R}^{MB}$ in the following sections. Details on the calibration process are available in supplementary materials.", + "bbox": [ + 89, + 295, + 482, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Ultra-wideband (UWB) UWB uses the time-of-flight technique to measure distances between two devices, with one acting as the transmitter and the other as the receiver. Unlike other RF techniques, such as WiFi and Bluetooth, UWB operates over a wide frequency range and transmits short pulses, achieving centimeter-level distance measurements with minimal interference [20]. Consequently, it is widely used in high-precision indoor localization systems with multiple fixed anchors and movable tags [38]. In our work, all six integrated UWB sensors are movable and alternately function as both transmitters and receivers, capturing inter-distances between each sensor pair.", + "bbox": [ + 89, + 446, + 482, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Framework Design", + "text_level": 1, + "bbox": [ + 89, + 637, + 274, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2, our method uses six IMU-UWB sensors and consists of three main modules: a shape estimator, a pose estimator, and a state estimator. The shape estimator takes anthropometric measurements and inter-distances in a T-pose as input, outputting shape parameters $\\hat{\\beta}$ of the SMPL model. The pose estimator uses filtered frame-aligned, root-normalized IMU measurements and inter-distances as input to estimate pose parameters, $\\hat{\\theta}$ , and corresponding uncertainties, $\\hat{\\Sigma}$ . Given our sensor placement, we do not observe the movement of the hands and feet. Therefore, similar to previous studies, we estimate poses for only 16 joints, including the root joint, while excluding the hand and foot joints. For simplicity, we denote this reduced set of pose parameters as $\\theta \\in \\mathbb{R}^{16\\times 3}$ , while $\\theta^{24}$ refers to the full set of pose parameters for 24 joints. Notations with a hat symbol, e.g., $\\hat{\\theta}$ , indicate the corresponding estimated values. Addi", + "bbox": [ + 89, + 657, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tionally, we do not estimate global translation, as inferring it from only body-worn IMU and UWB sensors without external references would lead to unbounded error accumulation. The state estimator integrates accelerations, UWB measurements, and the estimated poses with uncertainties to track the sensor relative positions, velocities, and acceleration biases of each node at each time step. These refined estimates are then used to refine the inputs to the pose estimator.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The estimation framework is driven by inherent uncertainties present in the system. Specifically, noisy IMU data combined with tracked inter-sensor distances serve as the primary inputs for pose estimation. The generated poses, along with their associated uncertainties that reflect sensor noise, are further constrained by the human model, which in turn refines sensor measurements and distance estimates. This feedback loop propagates uncertainties through the system, continuously updating the belief in system states based on the reliability of each data source—IMU, UWB, and pose estimator with body constraints, ultimately producing robust and accurate estimation.", + "bbox": [ + 511, + 213, + 906, + 395 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Shape and Pose Estimator", + "text_level": 1, + "bbox": [ + 511, + 411, + 751, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3.1. Shape Estimator", + "text_level": 1, + "bbox": [ + 511, + 435, + 676, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spatial inter-distances provide data that approximates the rough skeletal structure of an individual. Previous studies, such as Virtual Caliper [24] and SHAPY [2], have demonstrated a linear relationship between body shape and body measurements. Building on these findings, we adopt an ensemble learning-based shape estimator that uses experimentally selected anthropometric measurements, defined as:", + "bbox": [ + 511, + 455, + 905, + 561 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {\\beta}} = S E (H, W, \\boldsymbol {D}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 577, + 906, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Basic body measurements, height $H \\in \\mathbb{R}^1$ and weight $W \\in \\mathbb{R}^1$ , are easily accessible and provide foundational information about human body proportions. Experimentally selected inter-distances between sensors, $D \\in \\mathbb{R}^7$ , capture relative limb lengths and body structural details that height and weight alone cannot provide.", + "bbox": [ + 511, + 611, + 905, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 3, we place virtual sensors on the body mesh and conduct line-of-sight (LOS) simulation experiments with real-world tests to identify and select 7 repeatable distances out of 15 possible options. While the selection may vary based on device characteristics, the chosen distances offer a reliable foundation for adaptation. For model training, we use AutoGluon [5], an AutoML framework that automatically trains and ensembles 11 basic machine learning models, which is more stable than a single linear regression method in our validation. The estimated shape parameters $\\hat{\\beta}$ are then used to reconstruct realistic human bodies and to propagate estimated poses to spatial constraints in the state estimator.", + "bbox": [ + 511, + 704, + 906, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7087", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0d3ccd4fc14c3d62ceb39f841417717db676e66dcd35a8234acc914a351f6748.jpg", + "image_caption": [ + "Figure 2. Overview of UMotion, consisting of three main modules: the shape estimator, pose estimator, and state estimator. The shape estimator takes anthropometric measurements and inter-distances in a T-pose as input, outputting shape parameters that reconstruct a realistic body and impose strong constraints on the system. The pose estimator receives filtered IMU data and inter-distances from the state estimator to predict poses, which are fed back to refine state estimates. The entire system integrates IMU, UWB, and estimated poses within the context of individual body structure to continuously update and improve motion estimation." + ], + "image_footnote": [], + "bbox": [ + 133, + 90, + 861, + 277 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c15a1643ba69f95646107fa37e9d6d87a58a1a336efea4167f94def6aa070a21.jpg", + "image_caption": [ + "Figure 3. Visualization of selected inter-sensor distances used in the shape estimator. We place virtual sensors on the body mesh at sensor mounting points and conduct line-of-sight simulation experiments. The plots display the temporal changes in UWB measurements between various sensor pairs over time." + ], + "image_footnote": [], + "bbox": [ + 127, + 385, + 455, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2. Pose Estimator", + "text_level": 1, + "bbox": [ + 89, + 699, + 243, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our pose estimator is based on unidirectional multi-layer long short-term memory (LSTM) [6] recurrent neural network (RNN). The unidirectional LSTM preserves only past information, achieving real-time estimation without requiring future data. Given six IMU-UWB sensors, we use filtered accelerations, $\\hat{a} \\in \\mathbb{R}^{18}$ , and orientations in 9D rotation matrix form, $\\pmb{R} \\in \\mathbb{R}^{54}$ , along with 15 inter-distances $\\hat{\\pmb{d}} \\in \\mathbb{R}^{15}$ , as inputs. The estimator then outputs the poses $\\hat{\\theta}_{6D} \\in \\mathbb{R}^{96}$ for the target joints, represented in a 6D rotation matrix form [40]. Additionally, in the final layer, the network outputs the logarithm of the estimation uncertainty, $\\hat{\\Sigma} \\in \\mathbb{R}^{96}$ , capturing prediction confidence. Consequently,", + "bbox": [ + 89, + 719, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the pose estimator is defined as:", + "bbox": [ + 513, + 387, + 728, + 402 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {\\theta}} _ {6 D}, \\hat {\\boldsymbol {\\Sigma}} = P E (\\hat {\\boldsymbol {a}}, \\boldsymbol {R}, \\hat {\\boldsymbol {d}}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 412, + 906, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To handle ambiguity in the initial frame, we adopt a learning-based initialization strategy inspired by PIP [35]. A fully-connected convolutional neural network regresses the initial hidden state of the RNN based on the initial global poses, $\\hat{\\pmb{\\theta}}_0^{24}$ , and joint positions, $\\hat{J}_0 \\in \\mathbb{R}^{72}$ . This approach provides the pose estimator with a well-informed initialization, adapting to different bodies.", + "bbox": [ + 511, + 441, + 906, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use two loss functions to train the pose estimator: Mean Squared Error (MSE) and Gaussian Negative Log Likelihood (GNLL) loss. MSE is defined as:", + "bbox": [ + 511, + 551, + 906, + 597 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M S E}} \\left(\\boldsymbol {\\theta} _ {6 D}, \\hat {\\boldsymbol {\\theta}} _ {6 D}\\right) = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left\\| \\boldsymbol {\\theta} _ {6 D} - \\hat {\\boldsymbol {\\theta}} _ {6 D} \\right\\| _ {2} ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 609, + 906, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{\\theta}_{6D}$ denotes the ground truth poses and $n$ is the total number of frames. The GNLL, $\\mathcal{L}_{\\mathrm{GNLL}}(\\pmb{\\theta}_{6D},\\hat{\\pmb{\\theta}}_{6D},\\hat{\\pmb{\\Sigma}})$", + "bbox": [ + 511, + 657, + 906, + 690 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {1}{2} \\left(\\log \\left(\\max \\left(\\hat {\\boldsymbol {\\Sigma}} ^ {2}, \\varepsilon_ {\\min }\\right)\\right) + \\frac {(\\hat {\\boldsymbol {\\theta}} _ {6 D} - \\boldsymbol {\\theta} _ {6 D}) ^ {2}}{\\max \\left(\\hat {\\boldsymbol {\\Sigma}} ^ {2} , \\varepsilon_ {\\min }\\right)}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 700, + 906, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\varepsilon_{\\mathrm{min}}$ , set to $10^{-6}$ in our implementation, is used for numerical stability. First, $\\mathcal{L}_{\\mathrm{MSE}}$ helps the model converge quickly during early training. Next, $\\mathcal{L}_{\\mathrm{GNLL}}$ is applied to optimize for pose and uncertainty estimation.", + "bbox": [ + 511, + 762, + 906, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. State Estimator", + "text_level": 1, + "bbox": [ + 511, + 832, + 671, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use UKF [33] for state estimation to refine the pose estimator inputs, integrating UWB measurements and pose estimator outputs with a statistically derived IMU model.", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7088", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The UWB measurements provide direct inter-distance data, while they are affected by body occlusion issues. Meanwhile, pose estimator outputs offer constraints and inform distance measurements; however, they suffer from occasional inaccuracies due to challenging poses or sensor noise. The UKF combines these sources, balancing the strengths and limitations of each to make the optimal estimates.", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.1. State Definition", + "text_level": 1, + "bbox": [ + 89, + 207, + 246, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We define the state vector $\\pmb{x} \\in \\mathbb{R}^{15 \\times 3 + 15 \\times 3 + 6 \\times 3}$ as:", + "bbox": [ + 89, + 224, + 437, + 242 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} = \\left[ \\begin{array}{l l l l l l l l} \\boldsymbol {p} ^ {1 2} & \\dots & \\boldsymbol {p} ^ {5 6} & \\boldsymbol {v} ^ {1 2} & \\dots & \\boldsymbol {v} ^ {5 6} & \\boldsymbol {b} ^ {1} & \\dots & \\boldsymbol {b} ^ {6} \\end{array} \\right] ^ {\\mathsf {T}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 251, + 482, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{p}^{xy}$ represents the relative position between nodes $x$ and $y$ , with $x, y \\in \\{1, 2, 3, 4, 5, 6\\}$ and $x < y$ . The term $\\pmb{v}^{xy}$ denotes the relative velocity between nodes $x$ and $y$ , while $\\pmb{b}^1$ through $\\pmb{b}^6$ represent residual acceleration errors after transformation from $F^S$ to $F^M$ . These biases primarily result from orientation errors and raw accelerometer biases $\\pmb{b}_a$ . All quantities are expressed in the SMPL body-centric coordinate frame $F^M$ . For initialization, we set $\\pmb{p}^{xy}$ based on a static T-pose defined by the shape parameters $\\beta$ and the SMPL model (1). The relative velocities $\\pmb{v}^{xy}$ and biases are initialized to zero.", + "bbox": [ + 89, + 287, + 483, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.2. State Propagation", + "text_level": 1, + "bbox": [ + 89, + 463, + 263, + 478 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The control input $\\pmb{u}$ consists of the accelerations $\\pmb{a}$ of six nodes, defined as:", + "bbox": [ + 89, + 483, + 483, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {u} = \\left[ \\begin{array}{l l l} \\boldsymbol {a} ^ {1} & \\dots & \\boldsymbol {a} ^ {6} \\end{array} \\right] ^ {\\mathsf {T}}, \\boldsymbol {u} \\in \\mathbb {R} ^ {1 8}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 525, + 482, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the prediction step, the current state $\\pmb{x}$ is propagated using the control input $\\pmb{u}$ . The propagation model is derived from the strapdown inertial kinematic model with an acceleration error model (2), defined as follows:", + "bbox": [ + 89, + 559, + 483, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {v} _ {k} ^ {x y} = \\boldsymbol {v} _ {k - 1} ^ {x y} + \\left(\\boldsymbol {a} _ {k - 1} ^ {y} - \\boldsymbol {a} _ {k - 1} ^ {x}\\right) \\Delta t - \\left(\\boldsymbol {b} _ {k - 1} ^ {y} - \\boldsymbol {b} _ {k - 1} ^ {x}\\right) \\Delta t, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 632, + 482, + 669 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {p} _ {k} ^ {x y} = \\boldsymbol {p} _ {k - 1} ^ {x y} + \\boldsymbol {v} _ {k - 1} ^ {x y} \\Delta t + \\frac {1}{2} (\\boldsymbol {a} _ {k - 1} ^ {y} - \\boldsymbol {a} _ {k - 1} ^ {x}) \\Delta t ^ {2} \\\\ + \\frac {1}{2} \\left(\\boldsymbol {b} _ {k - 1} ^ {x} - \\boldsymbol {b} _ {k - 1} ^ {y}\\right) \\Delta t ^ {2}, \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 674, + 482, + 732 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {b} _ {k} ^ {x} = \\boldsymbol {b} _ {k - 1} ^ {x} + \\boldsymbol {\\eta} ^ {x}, \\boldsymbol {\\eta} ^ {x} \\sim \\mathcal {N} \\left(0, \\Sigma_ {b} ^ {2}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 734, + 482, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $k - 1$ and $k$ denote consecutive time steps, and $\\Delta t$ represents the interval between them. Together, equations (9), (10), and (11) form a state propagation model with added white noise, represented as:", + "bbox": [ + 89, + 766, + 483, + 825 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\boldsymbol {x}} _ {k} = f (\\boldsymbol {x} _ {k - 1}, \\boldsymbol {u} _ {k}) + \\mathbf {Q}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 840, + 482, + 857 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{Q}$ is the process noise covariance matrix, derived from the characteristics of the IMUs used in the system.", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.3. Measurement Update", + "text_level": 1, + "bbox": [ + 513, + 90, + 710, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The state is updated within measurement spaces using observed data $\\mathbf{z}$ . To accomplish this, we define a measurement model $h(\\mathbf{x})$ that maps the current state $\\mathbf{x}$ to measurement spaces as:", + "bbox": [ + 511, + 109, + 906, + 170 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh (\\boldsymbol {x}) = \\left[ \\begin{array}{l l l} \\| \\boldsymbol {p} ^ {x y} \\| _ {2} & \\| \\boldsymbol {v} ^ {x y} \\| _ {2} & \\boldsymbol {p} ^ {x y} \\quad \\boldsymbol {v} ^ {x y} \\end{array} \\right] ^ {\\intercal}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 178, + 906, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The UWB sensor measures the distance, $d^{xy}$ , between node $x$ and $y$ , which corresponds to the distance derived from relative positions in state, $\\| \\pmb{p}^{xy}\\|_2$ . Additionally, from consecutive distance measurements, we derive the relative velocity norm as $(d_k^{xy} - d_{k-1}^{xy}) / \\Delta t$ , corresponding to $\\| \\pmb{v}\\|_2^{xy}$ . The covariance matrix for distance measurements at time step $k$ , denoted as $\\mathbf{R}_{1,k}$ , is dynamically set based on line-of-sight conditions for each node pair, informed by $\\hat{\\theta}_k$ and UWB sensor characteristics. The covariance matrix for relative velocity measurements, $\\mathbf{R}_{2,k}$ , is computed as:", + "bbox": [ + 511, + 209, + 906, + 362 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} _ {2, k} = \\frac {\\left(\\mathbf {R} _ {1 , k} + \\mathbf {R} _ {1 , k - 1}\\right)}{\\Delta t ^ {2}}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 371, + 905, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The pose estimator outputs the joint rotations, $\\hat{\\pmb{\\theta}}$ , along with their corresponding standard deviations, $\\hat{\\Sigma}$ . We employ the unscented transform [11] to transform the pose distribution to the relative position distribution. Given $\\Theta \\in \\mathbb{R}^{96} \\sim \\mathcal{N}(\\hat{\\pmb{\\theta}}, \\hat{\\Sigma})$ , we use Van der Merwe's scaled sigma point algorithm [30] to generate $n$ sigma points, $\\mathcal{X} \\in \\mathbb{R}^{n \\times 96}$ , along with weights $\\boldsymbol{W}^m \\in \\mathbb{R}^n$ for the mean and $\\boldsymbol{W}^c \\in \\mathbb{R}^n$ for the covariance. Each sigma point in $\\mathcal{X}$ is transformed through the SMPL model (1), yielding a set of sensor-relative positions denoted as:", + "bbox": [ + 511, + 412, + 906, + 564 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Y} = \\{\\pmb {p} _ {i} ^ {x y} \\in \\mathbb {R} ^ {n \\times 1 5 \\times 3} \\mid i = 1, 2, \\dots , n \\}. \\qquad (1 5)\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 571, + 906, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We then calculate the mean and covariance of the transformed relative positions, capturing the distribution of relative positions based on pose uncertainty:", + "bbox": [ + 511, + 599, + 905, + 646 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {p}} ^ {x y} = \\sum_ {i = 1} ^ {n} \\boldsymbol {W} _ {i} ^ {m} \\mathcal {Y} _ {i}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 654, + 906, + 693 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {\\Sigma}} _ {\\hat {\\boldsymbol {p}}} ^ {2} = \\sum_ {i = 1} ^ {n} \\boldsymbol {W} _ {i} ^ {c} \\left(\\mathcal {Y} _ {i} - \\hat {\\boldsymbol {p}} ^ {x y}\\right) \\left(\\mathcal {Y} _ {i} - \\hat {\\boldsymbol {p}} ^ {x y}\\right) ^ {\\intercal}. \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 696, + 905, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The transformed relative positions, $\\hat{\\pmb{p}}^{xy}$ , along with the measurement covariance matrix $\\mathbf{R}_3 = \\hat{\\Sigma}_{\\hat{\\pmb{p}}}^2$ , are used to update $\\pmb{p}^{xy}$ in the state $\\pmb{x}$ . Similarly, we compute the relative velocity between consecutive time steps as $(\\hat{\\pmb{p}}_k^{xy} - \\hat{\\pmb{p}}_{k-1}^{xy}) / \\Delta t$ , with its covariance matrix defined as $\\mathbf{R}_{4,k} = (\\mathbf{R}_{3,k} + \\mathbf{R}_{3,k-1}) / \\Delta t^2$ , to update $\\pmb{v}^{xy}$ . The complete measurement vector $\\pmb{z}$ used for updating the state is given by:", + "bbox": [ + 511, + 744, + 906, + 854 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {z} = \\left[ d ^ {x y} \\quad \\frac {\\left(d _ {k} ^ {x y} - d _ {k - 1} ^ {x y}\\right)}{\\Delta t} \\quad \\hat {\\boldsymbol {p}} ^ {x y} \\quad \\frac {\\left(\\hat {\\boldsymbol {p}} _ {k} ^ {x y} - \\hat {\\boldsymbol {p}} _ {k - 1} ^ {x y}\\right)}{\\Delta t} \\right] ^ {\\intercal}. \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 875, + 906, + 905 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7089", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 89, + 222, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 89, + 114, + 192, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the AMASS [18], TotalCapture [29], DIP-IMU [7], and UIP [1] datasets for training and evaluation. We synthesize inertial and distance data from the AMASS dataset, following the process in previous studies [1, 7, 34]. We select specific vertices on the body mesh to represent sensor mounting positions. Accelerations are computed as finite differences of these vertex positions over time, while orientations are obtained as the global orientations of the corresponding joints. Distance measurements are synthesized by calculating Euclidean distances between selected sensor positions, adjusted for individual body shape. Training the shape estimator requires anthropometric measurements and inter-distances with corresponding shape parameters from a large number of individuals, which is challenging to collect. Therefore, we synthesize these measurements—height $H$ , weight $W$ , and inter-distances $D$ —from the 3D body mesh in a T-pose, using 479 unique body shapes (273 male and 206 female) available in the AMASS dataset. We estimate $H$ and $W$ following the methods in SHAPY [2], and we sample $D$ from the synthesized distances.", + "bbox": [ + 89, + 137, + 485, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The training data for the pose estimator comprises synthesized data from the AMASS dataset and the training set of DIP-IMU. Unlike previous studies [10, 34, 35], we exclude the TotalCapture within AMASS from the training data to prevent overlap, ensuring that synthesized distances used in testing are unseen during training. The test data includes TotalCapture, the test split of DIP-IMU, and UIP.", + "bbox": [ + 89, + 440, + 483, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Method Implementation", + "text_level": 1, + "bbox": [ + 89, + 556, + 313, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our shape estimator was implemented using the official multi-label predictor in AutoGluon 1.1 with a medium-quality preset. Our pose estimator was implemented based on PyTorch 2.3. The training process used an Adam optimizer with a learning rate of 0.0001. The pose estimator model was trained for 350 epochs with a batch size of 512 on a single NVIDIA GeForce RTX 3090 GPU, completing in approximately 35 minutes. For initial 20 epochs, we adopted $\\mathcal{L}_{MSE}$ to optimize the model, then switched to $\\mathcal{L}_{GNLL}$ for the remaining epochs. For the UKF, we experimentally set parameters $\\alpha_{UKF} = 0.2$ , $\\beta_{UKF} = 1.0$ and $\\kappa_{UKF} = -105$ to control the distribution and weighting of sample points. In the unscented transform applied to pose estimates, we used $\\alpha_{NN} = 0.09$ , $\\beta_{NN} = 1.0$ , and $\\kappa_{NN} = -93$ . We applied a factor of 10 to $\\mathbf{R}_3$ to compensate for overconfident predictions. The method achieves $60\\mathrm{Hz}$ without online LOS inference and $30\\mathrm{Hz}$ with it.", + "bbox": [ + 89, + 579, + 483, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 89, + 847, + 312, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our method against two categories of methods: (1) IMU-only methods, including DIP [7], TransPose [34],", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TIP [10], PIP [35], and PNP [37], and (2) distance augmented methods, including TIP-D, PIP-D, and UIP [1]. TIP-D and PIP-D are modified versions of TIP and PIP with distance-augmented input data, and we use their results as reported by UIP. We follow the evaluation protocol in UIP by adding ideal synthetic inter-sensor distances without noise to TotalCapture and DIP-IMU to assess the impact of distance measurements independent of noise. For evaluations on the UIP dataset, which includes both IMU and UWB data, we perform RANSAC regression to calibrate UWB measurements, as described in UIP [1], and then use calibrated measurements as input for the state estimator. We do not apply outlier filtering and use default state estimator parameters due to the absence of ground-truth data and limited information on sensor-specific characteristics in the hardware setup.", + "bbox": [ + 511, + 90, + 903, + 332 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics We use the following metrics for quantitative evaluation: 1) $SIP$ Error (in degrees): the mean global angular error of the upper arms and upper legs, focusing on joints not observed by the body-worn sensors; 2) Angular Error (in degrees): the mean global angular error across all body joints; 3) Positional Error (in centimeters): the mean global joint position error across all joints; and 4) Mesh Error (in centimeters): the mean vertex position error between the reconstructed mesh and the ground-truth mesh given the mean body shape.", + "bbox": [ + 511, + 335, + 905, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with IMU-only Methods Table 1 presents a comparison of our method against IMU-only methods on the TotalCapture and DIP-IMU datasets. On the TotalCapture dataset, our approach consistently outperforms previous methods across all metrics, improving over the SOTA PNP by $32.4\\%$ in angular error and $9.4\\%$ in mesh error. A similar trend is also observed in the results on DIP-IMU dataset. This demonstrates that inter-sensor distances serve as strong constraints for refining estimated poses.", + "bbox": [ + 511, + 489, + 905, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with Distance-augmented Methods Table 2 presents a fair comparison of our method with distance-augmented methods, as all approaches use the same input data. Additionally, all methods are trained on the AMASS dataset with the TotalCapture subset excluded, whereas the IMU-only methods in Table 1 include TotalCapture in their training data. Our method outperforms UIP, achieving a reduction in positional error of $21.5\\%$ on the TotalCapture dataset and $35.0\\%$ on the DIP-IMU dataset. On the UIP dataset, where sensor data quality is low and motions are more ambiguous than in the other two test datasets, our method achieves the lowest positional error, while UIP achieves the lowest SIP error. We achieve comparable results despite using less-filtered distance measurements, a simpler architecture, and non-optimal default parameters for the state estimator. This demonstrates the robustness of our approach in handling real-world sensor noise and the effectiveness of our online state estimation framework.", + "bbox": [ + 511, + 628, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7090", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2869f5bb18a876d74bfdb458086b086e3c5aac5e7553c91bb4b4dd9654a6e3e8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSIP Error (deg)Ang Error (deg)Pos Error (cm)Mesh Error (cm)SIP Error (deg)Ang Error (deg)Pos Error (cm)Mesh Error (cm)
TotalCaptureDIP-IMU
DIP [7]18.6217.229.4211.2217.3515.367.599.05
TransPose [34]16.5812.896.557.4217.068.866.037.17
TIP [10]13.2212.305.816.8016.909.075.636.62
PIP [35]12.9312.045.616.5115.338.785.126.02
PNP [37]10.8910.454.745.4513.718.754.975.77
UMotion10.767.064.464.9414.196.353.383.93
", + "bbox": [ + 125, + 88, + 870, + 234 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7db82d521f749891ccfb16bdfe90fd3d6ea2a8cc1d9e5bbe5275541af02fd600.jpg", + "table_caption": [ + "Table 1. Comparison with state of the art IMU-only methods on TotalCapture [29] and DIP-IMU [7]." + ], + "table_footnote": [], + "table_body": "
MethodSIP Error (deg)Pos Error (cm)SIP Error (deg)Pos Error (cm)SIP Error (deg)Pos Error (cm)
TotalCaptureDIP-IMUUIP
PIP [35]15.937.0515.986.2130.4713.62
TIP-D12.185.5115.915.2630.3413.96
PIP-D13.166.3113.795.3630.3313.27
UIP [1]11.325.4913.215.0524.1210.65
UMotion10.764.4614.193.3825.6910.33
", + "bbox": [ + 232, + 273, + 764, + 414 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e915928d92b3afaa1a6f76435238c9a8a0feed8da4abb7da25570335cd39b2a2.jpg", + "table_caption": [ + "Table 2. Comparison with distance-augmented methods on TotapCapture [29], DIP-IMU [7], and UIP [1] datasets." + ], + "table_footnote": [], + "table_body": "
MetricMean shapePredictedGT
Pos Error (cm)5.154.344.31
Mesh Error (cm)5.634.814.78
H Error (cm)5.750.39-
W Error (kg)9.450.34-
D Error (cm)3.632.332.30
C Error (cm)3.781.26-
", + "bbox": [ + 120, + 463, + 452, + 575 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ac3663c2d5090a3818893e288cb1ee094b900932e291767e1595b848e26eb642.jpg", + "table_caption": [ + "Table 3. Comparison of reconstructed body mesh errors on Total-Capture [29] using mean shape versus predicted shape parameters." + ], + "table_footnote": [], + "table_body": "
MethodSIP Error (deg)Pos Error (cm)SIP Error (deg)Pos Error (cm)
TotalCaptureDIP-IMU
J →hat9.004.3614.213.29
S →J →hat12.155.7017.354.45
PG →hat9.253.8513.632.85
S →PG →hat11.134.9615.913.91
PR →hat9.894.4213.773.22
S →PR →hat11.575.4415.744.17
Ours: S →hat10.764.4614.213.38
", + "bbox": [ + 514, + 463, + 924, + 632 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation study on pose estimator architecture.", + "bbox": [ + 545, + 642, + 870, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Module Evaluations", + "text_level": 1, + "bbox": [ + 89, + 646, + 281, + 661 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Shape Estimator Table 3 compares reconstructed body mesh on the TotalCapture dataset using the mean and predicted shape, both with the same estimated poses. With the predicted shape, positional and mesh errors are nearly as low as those obtained using the ground truth shape parameters. Height and weight estimates are accurate, with errors within $1\\mathrm{cm}$ and $1\\mathrm{kg}$ , respectively. However, the mean error in circumferences $\\pmb{C}$ of the chest, waist, hips, wrists, knees, and head remains large, as relevant measurements cannot be inferred solely from inter-sensor distances. Despite this limitation, the predicted shape still reconstructs a more realistic body than the mean shape model.", + "bbox": [ + 88, + 671, + 482, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pose Estimator We conduct an ablation study by expanding our pose estimator with intermediate layers to estimate all joint positions, $J$ , sensor global positions, $P_{G}$ , and", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sensor relative positions, $\\pmb{P}_R$ , from sensor measurements $S = (\\hat{\\pmb{a}}, R, \\hat{\\pmb{d}})$ . These intermediate estimates are then combined with $S$ to compute the final poses $\\hat{\\pmb{\\theta}}$ . This layered structure follows the framework commonly used in previous methods [1, 34]. As shown in Table 4, while regressing $\\hat{\\pmb{\\theta}}$ from ideal $J$ , $P_G$ , or $P_R$ improves accuracy, introducing an intermediate layer, that is, $S \\rightarrow \\{J, P_G, P_R\\}$ , increases errors. This error propagates through the network, ultimately degrading pose estimation performance. This suggests that integrating IMU data with distance constraints may support a simpler architecture, reducing complexity while still maintaining accuracy.", + "bbox": [ + 511, + 686, + 906, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "State Estimator The state estimator is designed to: 1) mitigate errors in inter-sensor distances, 2) filter and stabi", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7091", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4f32f9d8ff28a532371bead867e00f2bd9ccec0340428cb24f58c31aeaab7f7d.jpg", + "image_caption": [ + "Figure 4. Cumulative distribution of distance error (left) and acceleration error reduction over time (right) for various fusion settings." + ], + "image_footnote": [], + "bbox": [ + 192, + 88, + 496, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/57c0e9de5321df014c23e55ce2d008ef905b3cc2cdb9dc01018859df3d467c67.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 88, + 805, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d5210017e2471484867f7fcabae5e64f03ff1ae2f8767ec024f0311535b07567.jpg", + "image_caption": [ + "Figure 5. Joint positional error for different fusion settings." + ], + "image_footnote": [], + "bbox": [ + 130, + 276, + 442, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "imize accelerations, and 3) improve the accuracy of estimated poses. To examine it we conduct an ablation study by varying fusion measurements on the TotalCapture [29] with synthesized LOS-related noisy distances. Fig. 4 (left) shows cumulative distributions of absolute distance error for different fusion settings. The raw distances show an average error of $9.20\\mathrm{cm}$ with a standard deviation of $8.94\\mathrm{cm}$ . By fusing IMU, UWB, and pose data, this error is reduced to $2.42\\mathrm{cm}$ with a standard deviation of $2.39\\mathrm{cm}$ , demonstrating the effectiveness of data fusion for mitigating distance errors. Fig. 4 (right) shows the reduction in acceleration errors. The state estimator converges within a few seconds and adapts gradually, aligning with IMU characteristics. Fig. 5 presents joint positional errors over time for different fusion setups. Without filtering, positional error fluctuates with distance measurement quality. Fusing IMU with either UWB or pose data alone offers limited improvement because distance and pose outliers remain. The combined fusion of IMU, UWB, and pose data achieves the lowest positional error, validating that our state estimator refines pose accuracy by utilizing all available measurements.", + "bbox": [ + 89, + 455, + 483, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 89, + 786, + 302, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We developed a prototype (see supplementary material for details) that integrates BNO086 IMU and DW3000 UWB sensors to demonstrate our proposed method. Fig. 6 presents visualizations comparing pose estimates for challenging motions, illustrating the improvements of our method in disambiguating poses over IMU-only methods", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4d0ceada2480686881f32963cffdc5c2f9c6b8c3bd46cda28050e7836f55c546.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison of pose estimates from the test data collected using our developed prototype." + ], + "image_footnote": [], + "bbox": [ + 563, + 268, + 857, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and enhanced pose stability compared to UIP.", + "bbox": [ + 513, + 561, + 816, + 577 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Limitations", + "text_level": 1, + "bbox": [ + 511, + 598, + 769, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we present UMotion, an uncertainty-driven framework for 3D human shape and pose estimation that integrates six IMU and UWB sensors within a state estimation system. UMotion incorporates uncertainties from IMU, UWB, and pose estimates under individual body constraints to iteratively enhance confidence in both sensor measurements and pose accuracy. Our experiments on synthetic and real-world datasets demonstrate that UMotion outperforms existing SOTA methods in pose accuracy, and effectively stabilizes sensor measurements and pose estimation.", + "bbox": [ + 511, + 626, + 906, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "However, our shape estimator requires adaptation to specific sensor configurations and conditions, and, with limited training data, it may struggle with unique body variations. The state estimator also requires careful parameter tuning to reflect sensor characteristics accurately. Additionally, our pose estimator simplifies constraints, which may limit its performance. As future work, adding a physics-aware module could potentially enhance our method's robustness.", + "bbox": [ + 511, + 780, + 908, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7092", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rayan Armani, Changlin Qian, Jiaxi Jiang, and Christian Holz. Ultra inertial poser: Scalable motion capture and tracking from sparse inertial sensors and ultra-wideband ranging. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1, 2, 6, 7", + "[2] Vasileios Choutas, Lea Müller, Chun-Hao P Huang, Siyu Tang, Dimitrios Tzionas, and Michael J Black. Accurate 3d body shape regression using metric and semantic attributes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2718-2728, 2022. 3, 6", + "[3] Peng Dai, Yang Zhang, Tao Liu, Zhen Fan, Tianyuan Du, Zhuo Su, Xiaozheng Zheng, and Zeming Li. Hmd-poser: On-device real-time human motion tracking from scalable sparse observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 874-884, 2024. 1", + "[4] Nathan DeVrio, Vimal Mollyn, and Chris Harrison. Smartposer: Arm pose estimation with a smartphone and smartwatch using uwb and imu data. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, pages 1-11, 2023. 2", + "[5] Nick Erickson, Jonas Mueller, Alexander Shirkov, Hang Zhang, Pedro Larroy, Mu Li, and Alexander Smola. Autogluon-tabular: Robust and accurate automl for structured data. arXiv preprint arXiv:2003.06505, 2020. 3", + "[6] S Hochreiter. Long short-term memory. Neural Computation MIT-Press, 1997. 4", + "[7] Yinghao Huang, Manuel Kaufmann, Emre Aksan, Michael J Black, Otmar Hilliges, and Gerard Pons-Moll. Deep inertial poser: Learning to reconstruct human pose from sparse inertial measurements in real time. ACM Transactions on Graphics (TOG), 37(6):1-15, 2018. 1, 2, 6, 7", + "[8] Jiaxi Jiang, Paul Streli, Huajian Qiu, Andreas Fender, Larissa Laich, Patrick Snape, and Christian Holz. Avatarposer: Articulated full-body pose tracking from sparse motion sensing. In European conference on computer vision, pages 443-460. Springer, 2022. 1", + "[9] Jiaxi Jiang, Paul Streli, Manuel Meier, and Christian Holz. Egoposer: Robust real-time ego-body pose estimation in large scenes. arXiv preprint arXiv:2308.06493, 2023. 2", + "[10] Yifeng Jiang, Yuting Ye, Deepak Gopinath, Jungdam Won, Alexander W Winkler, and C Karen Liu. Transformer inertial poser: Real-time human motion reconstruction from sparse imus with simultaneous terrain generation. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 1, 2, 6, 7", + "[11] Simon J Julier, Jeffrey K Uhlmann, and Hugh F Durrant-Whyte. A new approach for filtering nonlinear systems. In Proceedings of 1995 American Control Conference-ACC'95, pages 1628-1632. IEEE, 1995. 5", + "[12] Manuel Kaufmann, Yi Zhao, Chengcheng Tang, Lingling Tao, Christopher Twigg, Jie Song, Robert Wang, and Otmar Hilliges. Em-posed: 3d human pose estimation from sparse electromagnetic trackers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11510-11520, 2021. 2" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Jiye Lee and Hanbyul Joo. Mocap everyone everywhere: Lightweight motion capture with smartwatches and a head-mounted camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1091-1100, 2024. 1, 2", + "[14] Sunmin Lee, Sebastian Starke, Yuting Ye, Jungdam Won, and Alexander Winkler. Questenvsim: Environment-aware simulated motion tracking from sparse sensors. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-9, 2023. 1", + "[15] Han Liang, Yannan He, Chengfeng Zhao, Mutian Li, Jingya Wang, Jingyi Yu, and Lan Xu. Hybridcap: Inertia-aid monocular capture of challenging human motions. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1539–1548, 2023. 2", + "[16] Huajun Liu, Xiaolin Wei, Jinxiang Chai, Inwoo Ha, and Taehyun Rhee. Realtime human motion control with a small number of inertial sensors. In Symposium on interactive 3D graphics and games, pages 133-140, 2011. 2", + "[17] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. In *Seminal Graphics Papers: Pushing the Boundaries*, Volume 2, pages 851–866. 2023. 2", + "[18] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5442-5451, 2019. 2, 6", + "[19] Vimal Mollyn, Riku Arakawa, Mayank Goel, Chris Harrison, and Karan Ahuja. Imuposer: Full-body pose estimation using imus in phones, watches, and earbuds. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, pages 1-12, 2023. 1", + "[20] Ian Oppermann, Matti Hämäläinen, and Jari Inatti. *UWB: theory and applications*. John Wiley & Sons, 2004. 3", + "[21] Shaohua Pan, Qi Ma, Xinyu Yi, Weifeng Hu, Xiong Wang, Xingkang Zhou, Jijunnan Li, and Feng Xu. Fusing monocular images and sparse imu signals for real-time human motion capture. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 2", + "[22] Gerard Pons-Moll, Andreas Baak, Thomas Helten, Meinard Müller, Hans-Peter Seidel, and Bodo Rosenhahn. Multisensor-fusion for 3d full-body human motion capture. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 663-670. IEEE, 2010.", + "[23] Jose Luis Ponton, Haoran Yun, Andreas Aristidou, Carlos Andujar, and Nuria Pelechano. Sparseposer: Real-time full-body motion reconstruction from sparse data. ACM Transactions on Graphics, 43(1):1-14, 2023. 2", + "[24] Sergi Pujades, Betty Mohler, Anne Thaler, Joachim Tesch, Naureen Mahmood, Nikolas Hesse, Heinrich H Bülthoff, and Michael J Black. The virtual caliper: rapid creation of metrically accurate avatars from 3d measurements. IEEE transactions on visualization and computer graphics, 25(5): 1887-1897, 2019. 3", + "[25] Yiming Ren, Chengfeng Zhao, Yannan He, Peishan Cong, Han Liang, Jingyi Yu, Lan Xu, and Yuexin Ma. Lidar-aid" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7093", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "inertial poser: Large-scale human motion capture by sparse inertial and lidar sensors. IEEE Transactions on Visualization and Computer Graphics, 29(5):2337-2347, 2023. 2", + "[26] Xu Ru, Nian Gu, Hang Shang, and Heng Zhang. Mems inertial sensor calibration technology: Current status and future trends. Micromachines, 13(6):879, 2022. 3", + "[27] Martin Schepers, Matteo Giuberti, Giovanni Bellusci, et al. Xsens mvn: Consistent tracking of human motion using inertial sensing. Xsens Technol, 1(8):1-8, 2018. 1", + "[28] Mike Schuster and Kuldip K Paliwal. Bidirectional recurrent neural networks. IEEE transactions on Signal Processing, 45 (11):2673-2681, 1997. 2", + "[29] Matthew Trumble, Andrew Gilbert, Charles Malleson, Adrian Hilton, and John P Collomosse. Total capture: 3d human pose estimation fusing video and inertial sensors. In BMVC, pages 1-13. London, UK, 2017. 6, 7, 8", + "[30] Rudolph Van Der Merwe. Sigma-point Kalman filters for probabilistic inference in dynamic state-space models. Oregon Health & Science University, 2004. 5", + "[31] Tom Van Wouwe, Seunghwan Lee, Antoine Falisse, Scott Delp, and C Karen Liu. Diffusion inertial poser: Human motion reconstruction from arbitrary sparse imu configurations. arXiv preprint arXiv:2308.16682, 2023. 1", + "[32] Timo Von Marcard, Bodo Rosenhahn, Michael J Black, and Gerard Pons-Moll. Sparse inertial poser: Automatic 3d human pose estimation from sparse imus. In Computer graphics forum, pages 349-360. Wiley Online Library, 2017. 1, 2", + "[33] Eric A Wan and Rudolph Van Der Merwe. The unscented kalman filter for nonlinear estimation. In Proceedings of the IEEE 2000 adaptive systems for signal processing, communications, and control symposium (Cat. No. 00EX373), pages 153-158. Ieee, 2000. 4", + "[34] Xinyu Yi, Yuxiao Zhou, and Feng Xu. Transpose: Real-time 3d human translation and pose estimation with six inertial sensors. ACM Transactions On Graphics (TOG), 40(4):1-13, 2021. 1, 2, 6, 7", + "[35] Xinyu Yi, Yuxiao Zhou, Marc Habermann, Soshi Shimada, Vladislav Golyanik, Christian Theobalt, and Feng Xu. Physical inertial poser (pip): Physics-aware real-time human motion tracking from sparse inertial sensors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13167-13178, 2022. 1, 2, 4, 6, 7", + "[36] Xinyu Yi, Yuxiao Zhou, Marc Habermann, Vladislav Golyanik, Shaohua Pan, Christian Theobalt, and Feng Xu. Ecolocate: Real-time motion capture, localization, and mapping with sparse body-mounted sensors. ACM Transactions on Graphics (TOG), 42(4):1-17, 2023. 2", + "[37] Xinyu Yi, Yuxiao Zhou, and Feng Xu. Physical non-inertial poser (pnp): Modeling non-inertial effects in sparse-inertial human motion capture. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1, 2, 6, 7", + "[38] Faheem Zafari, Athanasios Gkelias, and Kin K Leung. A survey of indoor localization systems and technologies. IEEE Communications Surveys & Tutorials, 21(3):2568-2599, 2019. 3" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Yu Zhang, Songpengcheng Xia, Lei Chu, Jiarui Yang, Qi Wu, and Ling Pei. Dynamic inertial poser (dynaip): Part-based motion dynamics learning for enhanced human pose estimation with sparse inertial sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1889-1899, 2024. 1, 2", + "[40] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5745-5753, 2019. 4", + "[41] Chengxu Zuo, Yiming Wang, Lishuang Zhan, Shihui Guo, Xinyu Yi, Feng Xu, and Yipeng Qin. Loose inertial poser: Motion capture with imu-attached loose-wear jacket. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2209–2219, 2024. 1" + ], + "bbox": [ + 516, + 90, + 906, + 316 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7094", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/UMotion_ Uncertainty-driven Human Motion Estimation from Inertial and Ultra-wideband Units/20c112a8-f19a-4ffc-8e5b-a5d02b15caa1_model.json b/2025/UMotion_ Uncertainty-driven Human Motion Estimation from Inertial and Ultra-wideband Units/20c112a8-f19a-4ffc-8e5b-a5d02b15caa1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..140872fdb1ca3bf63378e5a6259d159a16277be6 --- /dev/null +++ b/2025/UMotion_ Uncertainty-driven Human Motion Estimation from Inertial and Ultra-wideband Units/20c112a8-f19a-4ffc-8e5b-a5d02b15caa1_model.json @@ -0,0 +1,2046 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.184, + 0.131, + 0.819, + 0.175 + ], + "angle": 0, + "content": "UMotion: Uncertainty-driven Human Motion Estimation from Inertial and Ultra-wideband Units" + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.204, + 0.733, + 0.221 + ], + "angle": 0, + "content": "Huakun Liu Hiroki Ota Xin Wei Yutaro Hirao" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.222, + 0.804, + 0.258 + ], + "angle": 0, + "content": "Monica Perusquía-Hernández \nHideaki Uchiyama \nKiyoshi Kiyokawa \nNara Institute of Science and Technology, Japan" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.26, + 0.813, + 0.277 + ], + "angle": 0, + "content": "{liu.huakun.li0, ota.hiroki.oc6, wei.xin.wy0, yutaro.hirao}@is.naist.jp" + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.278, + 0.714, + 0.293 + ], + "angle": 0, + "content": "{m.perusquia, hideaki.uchiyama, kiyo}@is.naist.jp" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.36, + 0.486, + 0.725 + ], + "angle": 0, + "content": "Sparse wearable inertial measurement units (IMUs) have gained popularity for estimating 3D human motion. However, challenges such as pose ambiguity, data drift, and limited adaptability to diverse bodies persist. To address these issues, we propose UMotion, an uncertainty-driven, online fusing-all state estimation framework for 3D human shape and pose estimation, supported by six integrated, body-worn ultra-wideband (UWB) distance sensors with IMUs. UWB sensors measure inter-node distances to infer spatial relationships, aiding in resolving pose ambiguities and body shape variations when combined with anthropometric data. Unfortunately, IMUs are prone to drift, and UWB sensors are affected by body occlusions. Consequently, we develop a tightly coupled Unscented Kalman Filter (UKF) framework that fuses uncertainties from sensor data and estimated human motion based on individual body shape. The UKF iteratively refines IMU and UWB measurements by aligning them with uncertain human motion constraints in real-time, producing optimal estimates for each. Experiments on both synthetic and real-world datasets demonstrate the effectiveness of UMotion in stabilizing sensor data and the improvement over state of the art in pose accuracy. Code is available at: https://github.com/kk9six/umotion." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.754, + 0.224, + 0.77 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Estimating 3D human motion from wearable sensors has become increasingly popular due to their portability, accessibility, and versatility. Wearable sensors, such as inertial measurement units (IMUs), enable continuous monitoring of body motion measurements across unrestricted spaces. These advances shift motion capture from controlled laboratory environments to everyday settings [13, 19, 32, 41]. This transition benefits fields such as healthcare, sports" + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.329, + 0.865, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.469, + 0.908, + 0.541 + ], + "angle": 0, + "content": "Figure 1. UMotion integrates IMU-UWB data inputs and pose outputs uniformly under uncertainty, constrained by individual body structure. The online state estimation framework iteratively refines sensor data confidence and stabilizes pose estimation, reducing ambiguities and improving robustness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.597 + ], + "angle": 0, + "content": "performance, ergonomics, and emerging areas in humancomputer interaction [3, 8, 14, 31]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.598, + 0.909, + 0.854 + ], + "angle": 0, + "content": "One of the widely chosen wearable sensors for 3D human motion estimation is IMU. Commercial systems, such as Xsens [27], utilize 17 or more IMUs for comprehensive pose coverage. While highly accurate, these densely placed IMUs are inconvenient and intrusive. Recent studies have reduced the required number of IMUs to just six—placed on the forearms, lower legs, pelvis, and head—while still achieving promising performance through data-driven methods [1, 7, 10, 31, 32, 34, 35, 37, 39]. With fewer IMUs, pose estimation becomes under-constrained and prone to ambiguity. Recent work has attempted to disambiguate poses by incorporating temporal consistency, physics-based constraints, or additional, easy-to-integrate sensors [1, 10, 35]. However, these methods still face significant challenges, including noisy sensor data, body shape variations, and ambiguities arising from under-constrained setups with sparse sensors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose UMotion, an uncertainty-driven human motion estimation framework that combines online state estimation with an integrated system of six inertial and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7085" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.332 + ], + "angle": 0, + "content": "ultra-wideband (UWB) distance sensors. Prior work often infers poses from sensor data without considering the reverse influence. UMotion uniformly treats input and output under uncertainty and body-specific constraints, blending data to optimize estimates (see Fig. 1). Our approach maintains inter-sensor distances as a core system state, capturing spatial relationships between sensor nodes. These distances, combined with basic anthropometric measurements (height and weight) and IMU data, serve as inputs for our learning-based shape and pose estimators. Since sensors are inherently noisy, we propose a tightly coupled online state estimation system that associates IMU and UWB sensor measurements and pose estimates using an Unscented Kalman Filter (UKF) and an uncertainty propagation method, iteratively correcting errors and stabilizing pose estimation in real-time. Our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.334, + 0.48, + 0.375 + ], + "angle": 0, + "content": "- An ensemble learning-based human shape estimation approach using distance constraints and anthropometric measurements from six integrated IMU-UWB sensors." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.379, + 0.48, + 0.42 + ], + "angle": 0, + "content": "- A learning-based method for human pose estimation from inertial and distance constraints, incorporating mesh distribution estimation." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.424, + 0.48, + 0.466 + ], + "angle": 0, + "content": "- A filtering-based state estimation system that couples sensor measurements, pose estimates, and body variations in real-time to improve sensor stability and pose accuracy." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.334, + 0.48, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.483, + 0.232, + 0.499 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.483, + 0.719 + ], + "angle": 0, + "content": "Pose Estimation from Sparse Inertial Sensors In contrast to the predominant use of vision-based methods, wearable sensors, specifically IMUs, offer greater freedom and flexibility. Von Marcard et al. [32] present SIP that makes inertial pose estimation practical by using only six IMUs attached to the body, combined with an offline iterative SMPL body model [17] pose optimization. Huang et al. [7] propose a real-time pose estimation method, DIP, which uses a bidirectional recurrent neural network (biRNN) [28] to learn the mapping from a sequence of six IMU measurements to SMPL body poses. To obtain sufficient training data, they synthesize IMU measurements from the AMASS dataset [18], further advancing the development of learning-based methods for sparse sensor motion capture." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Following previous studies, TransPose [34] refines pose estimation by decomposing the end-to-end framework into a multi-stage process with intermediate joint position estimation. To disambiguate the poses, PIP [35] proposes a physics-based motion optimization, while TIP [10] incorporates a conditional Transformer model for plausible terrain generation. Training data is crucial for learning-based methods. Unlike prior studies relying on synthetic IMU data, DynalIP [39] adapts real IMU data from diverse human skeleton formats to the target SMPL model and shows superior performance when training with real sensor data. Similarly focusing on the data, PNP [37] recently addresses" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.151 + ], + "angle": 0, + "content": "limitations in existing IMU synthesis by incorporating noninertial effects and fictitious forces, enhancing estimation robustness through a physics-informed neural network and realistic IMU synthesis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.906, + 0.319 + ], + "angle": 0, + "content": "In contrast to methods focusing solely on pose estimation, our approach treats human shape and pose as equally important. By integrating shape information, we establish a tight connection between sensor data and estimated motions, forming a positive feedback loop that enhances the entire process. Additionally, most previous studies fail to fully utilize IMU accelerations due to high noise and drift issues. Within our framework, accelerations serve as control inputs for state estimation, undergoing continuous correction and acting as a crucial component in tracking spatial relationships among sensors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.907, + 0.705 + ], + "angle": 0, + "content": "Pose Estimation from Hybrid Sensors In addition to IMUs, various wearable and hybrid sensor systems have been explored to overcome inherent IMU limitations such as restricted positional awareness [9, 12, 13, 15, 16, 21–23, 25, 36]. The closest works to ours are SmartPoser [4] and Ultra Inertial Poser (UIP) [1], both of which integrate UWB and IMU sensors for pose estimation. UWB sensors complement IMUs with additional distance information between sensors while preserving the portability and flexibility of tracking devices. SmartPoser [4] focuses on wearable arm pose estimation, combining UWB measurements with IMU data using an off-the-shelf smartwatch and smartphone. UIP [1] integrates six UWB sensors with IMUs for full-body pose estimation. Inter-sensor distances, estimated using an EKF that fuses IMU and UWB data, help reduce global translation drift and minimize position jitter compared to inertial-only tracking. However, body occlusion frequently affects certain node pairs, e.g., head-knee, causing the EKF to fail in estimating distances due to rapid IMU drift and unreliable UWB measurements, resulting in unstable distance data. Our approach addresses this limitation by using output pose uncertainties and body-specific constraints as observations to refine input sensor measurements, improving the stability of both IMU and UWB data." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.719, + 0.605, + 0.734 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.744, + 0.655, + 0.758 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.905, + 0.872 + ], + "angle": 0, + "content": "SMPL body model We use the SMPL model [17] to represent human motion. SMPL decomposes the human body into pose parameters, \\(\\pmb{\\theta} \\in \\mathbb{R}^{23 \\times 3 + 3}\\), which define relative rotations of 23 joints and the global root joint orientation in axis-angle form, and shape parameters, \\(\\beta \\in \\mathbb{R}^{10}\\), capturing body shape variations across individuals. The model is defined by a linear blend-skinning (LBS) function:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.884, + 0.907, + 0.903 + ], + "angle": 0, + "content": "\\[\nM (\\boldsymbol {\\beta}, \\boldsymbol {\\theta}) = W \\left(\\bar {\\mathbf {T}} + B _ {s} (\\boldsymbol {\\beta}) + B _ {p} (\\boldsymbol {\\theta}), J (\\boldsymbol {\\beta}), \\boldsymbol {\\theta}, \\mathcal {W}\\right), \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7086" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.258 + ], + "angle": 0, + "content": "where \\(B_{s}(\\boldsymbol {\\beta})\\) and \\(B_{p}(\\boldsymbol {\\theta})\\) are shape and pose blend shapes that deform the template mesh \\(\\bar{\\mathbf{T}}\\) in the zero pose. The mesh is then reposed according to specified poses \\(\\theta\\), combining with joint positions \\(J(\\beta)\\) and blend-skinning weights \\(\\mathcal{W}\\). Inertial Measurement Unit (IMU) A nine-axis MEMS IMU comprises an accelerometer, gyroscope, and magnetometer, measuring accelerations in the sensor local frame \\(F^S\\), angular velocities in \\(F^S\\), and magnetic field strengths relative to the Earth's magnetic field, respectively. A commonly used error model for measured accelerations of one IMU, \\(\\pmb{a}\\in \\mathbb{R}^{3}\\), is given by [26]:" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.27, + 0.483, + 0.287 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {a} = \\boldsymbol {a} _ {\\text {t r u e}} + \\boldsymbol {b} _ {a} + \\boldsymbol {n} _ {a}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.296, + 0.483, + 0.447 + ], + "angle": 0, + "content": "where \\(\\pmb{a}_{\\mathrm{true}}\\) is the true acceleration, \\(\\pmb{b}_a\\) is the bias, modeled as a random walk process with noise \\(\\eta\\) following a zero-mean Gaussian distribution with standard deviation \\(\\Sigma_{b}\\), and \\(\\pmb{n}_a\\) is high-frequency white noise. In our work, after a frame calibration process, accelerations are represented as \\(\\pmb{a}^{M}\\) in the SMPL body-centric frame \\(F^{M}\\), and orientations are represented as \\(\\pmb{R}^{MB}\\), indicating the rotation from bone frame \\(F^{B}\\) to \\(F^{M}\\). For simplicity, we denote \\(\\pmb{a}\\) as \\(\\pmb{a}^{M}\\) and \\(\\pmb{R}\\) as \\(\\pmb{R}^{MB}\\) in the following sections. Details on the calibration process are available in supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.448, + 0.483, + 0.629 + ], + "angle": 0, + "content": "Ultra-wideband (UWB) UWB uses the time-of-flight technique to measure distances between two devices, with one acting as the transmitter and the other as the receiver. Unlike other RF techniques, such as WiFi and Bluetooth, UWB operates over a wide frequency range and transmits short pulses, achieving centimeter-level distance measurements with minimal interference [20]. Consequently, it is widely used in high-precision indoor localization systems with multiple fixed anchors and movable tags [38]. In our work, all six integrated UWB sensors are movable and alternately function as both transmitters and receivers, capturing inter-distances between each sensor pair." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.638, + 0.275, + 0.654 + ], + "angle": 0, + "content": "3.2. Framework Design" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 2, our method uses six IMU-UWB sensors and consists of three main modules: a shape estimator, a pose estimator, and a state estimator. The shape estimator takes anthropometric measurements and inter-distances in a T-pose as input, outputting shape parameters \\(\\hat{\\beta}\\) of the SMPL model. The pose estimator uses filtered frame-aligned, root-normalized IMU measurements and inter-distances as input to estimate pose parameters, \\(\\hat{\\theta}\\), and corresponding uncertainties, \\(\\hat{\\Sigma}\\). Given our sensor placement, we do not observe the movement of the hands and feet. Therefore, similar to previous studies, we estimate poses for only 16 joints, including the root joint, while excluding the hand and foot joints. For simplicity, we denote this reduced set of pose parameters as \\(\\theta \\in \\mathbb{R}^{16\\times 3}\\), while \\(\\theta^{24}\\) refers to the full set of pose parameters for 24 joints. Notations with a hat symbol, e.g., \\(\\hat{\\theta}\\), indicate the corresponding estimated values. Addi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "tionally, we do not estimate global translation, as inferring it from only body-worn IMU and UWB sensors without external references would lead to unbounded error accumulation. The state estimator integrates accelerations, UWB measurements, and the estimated poses with uncertainties to track the sensor relative positions, velocities, and acceleration biases of each node at each time step. These refined estimates are then used to refine the inputs to the pose estimator." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.214, + 0.907, + 0.396 + ], + "angle": 0, + "content": "The estimation framework is driven by inherent uncertainties present in the system. Specifically, noisy IMU data combined with tracked inter-sensor distances serve as the primary inputs for pose estimation. The generated poses, along with their associated uncertainties that reflect sensor noise, are further constrained by the human model, which in turn refines sensor measurements and distance estimates. This feedback loop propagates uncertainties through the system, continuously updating the belief in system states based on the reliability of each data source—IMU, UWB, and pose estimator with body constraints, ultimately producing robust and accurate estimation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.412, + 0.753, + 0.429 + ], + "angle": 0, + "content": "3.3. Shape and Pose Estimator" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.436, + 0.677, + 0.451 + ], + "angle": 0, + "content": "3.3.1. Shape Estimator" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.906, + 0.563 + ], + "angle": 0, + "content": "Spatial inter-distances provide data that approximates the rough skeletal structure of an individual. Previous studies, such as Virtual Caliper [24] and SHAPY [2], have demonstrated a linear relationship between body shape and body measurements. Building on these findings, we adopt an ensemble learning-based shape estimator that uses experimentally selected anthropometric measurements, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.578, + 0.907, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {\\beta}} = S E (H, W, \\boldsymbol {D}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.906, + 0.703 + ], + "angle": 0, + "content": "Basic body measurements, height \\(H \\in \\mathbb{R}^1\\) and weight \\(W \\in \\mathbb{R}^1\\), are easily accessible and provide foundational information about human body proportions. Experimentally selected inter-distances between sensors, \\(D \\in \\mathbb{R}^7\\), capture relative limb lengths and body structural details that height and weight alone cannot provide." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.9 + ], + "angle": 0, + "content": "As shown in Fig. 3, we place virtual sensors on the body mesh and conduct line-of-sight (LOS) simulation experiments with real-world tests to identify and select 7 repeatable distances out of 15 possible options. While the selection may vary based on device characteristics, the chosen distances offer a reliable foundation for adaptation. For model training, we use AutoGluon [5], an AutoML framework that automatically trains and ensembles 11 basic machine learning models, which is more stable than a single linear regression method in our validation. The estimated shape parameters \\(\\hat{\\beta}\\) are then used to reconstruct realistic human bodies and to propagate estimated poses to spatial constraints in the state estimator." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7087" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.092, + 0.862, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.291, + 0.908, + 0.364 + ], + "angle": 0, + "content": "Figure 2. Overview of UMotion, consisting of three main modules: the shape estimator, pose estimator, and state estimator. The shape estimator takes anthropometric measurements and inter-distances in a T-pose as input, outputting shape parameters that reconstruct a realistic body and impose strong constraints on the system. The pose estimator receives filtered IMU data and inter-distances from the state estimator to predict poses, which are fed back to refine state estimates. The entire system integrates IMU, UWB, and estimated poses within the context of individual body structure to continuously update and improve motion estimation." + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.386, + 0.456, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.602, + 0.486, + 0.673 + ], + "angle": 0, + "content": "Figure 3. Visualization of selected inter-sensor distances used in the shape estimator. We place virtual sensors on the body mesh at sensor mounting points and conduct line-of-sight simulation experiments. The plots display the temporal changes in UWB measurements between various sensor pairs over time." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.7, + 0.245, + 0.715 + ], + "angle": 0, + "content": "3.3.2. Pose Estimator" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Our pose estimator is based on unidirectional multi-layer long short-term memory (LSTM) [6] recurrent neural network (RNN). The unidirectional LSTM preserves only past information, achieving real-time estimation without requiring future data. Given six IMU-UWB sensors, we use filtered accelerations, \\(\\hat{a} \\in \\mathbb{R}^{18}\\), and orientations in 9D rotation matrix form, \\(\\pmb{R} \\in \\mathbb{R}^{54}\\), along with 15 inter-distances \\(\\hat{\\pmb{d}} \\in \\mathbb{R}^{15}\\), as inputs. The estimator then outputs the poses \\(\\hat{\\theta}_{6D} \\in \\mathbb{R}^{96}\\) for the target joints, represented in a 6D rotation matrix form [40]. Additionally, in the final layer, the network outputs the logarithm of the estimation uncertainty, \\(\\hat{\\Sigma} \\in \\mathbb{R}^{96}\\), capturing prediction confidence. Consequently," + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.388, + 0.729, + 0.403 + ], + "angle": 0, + "content": "the pose estimator is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.413, + 0.907, + 0.432 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {\\theta}} _ {6 D}, \\hat {\\boldsymbol {\\Sigma}} = P E (\\hat {\\boldsymbol {a}}, \\boldsymbol {R}, \\hat {\\boldsymbol {d}}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.443, + 0.907, + 0.552 + ], + "angle": 0, + "content": "To handle ambiguity in the initial frame, we adopt a learning-based initialization strategy inspired by PIP [35]. A fully-connected convolutional neural network regresses the initial hidden state of the RNN based on the initial global poses, \\(\\hat{\\pmb{\\theta}}_0^{24}\\), and joint positions, \\(\\hat{J}_0 \\in \\mathbb{R}^{72}\\). This approach provides the pose estimator with a well-informed initialization, adapting to different bodies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.598 + ], + "angle": 0, + "content": "We use two loss functions to train the pose estimator: Mean Squared Error (MSE) and Gaussian Negative Log Likelihood (GNLL) loss. MSE is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.61, + 0.907, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M S E}} \\left(\\boldsymbol {\\theta} _ {6 D}, \\hat {\\boldsymbol {\\theta}} _ {6 D}\\right) = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left\\| \\boldsymbol {\\theta} _ {6 D} - \\hat {\\boldsymbol {\\theta}} _ {6 D} \\right\\| _ {2} ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.659, + 0.908, + 0.691 + ], + "angle": 0, + "content": "where \\(\\pmb{\\theta}_{6D}\\) denotes the ground truth poses and \\(n\\) is the total number of frames. The GNLL, \\(\\mathcal{L}_{\\mathrm{GNLL}}(\\pmb{\\theta}_{6D},\\hat{\\pmb{\\theta}}_{6D},\\hat{\\pmb{\\Sigma}})\\)" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.701, + 0.907, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\frac {1}{2} \\left(\\log \\left(\\max \\left(\\hat {\\boldsymbol {\\Sigma}} ^ {2}, \\varepsilon_ {\\min }\\right)\\right) + \\frac {(\\hat {\\boldsymbol {\\theta}} _ {6 D} - \\boldsymbol {\\theta} _ {6 D}) ^ {2}}{\\max \\left(\\hat {\\boldsymbol {\\Sigma}} ^ {2} , \\varepsilon_ {\\min }\\right)}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.763, + 0.908, + 0.825 + ], + "angle": 0, + "content": "where \\(\\varepsilon_{\\mathrm{min}}\\) , set to \\(10^{-6}\\) in our implementation, is used for numerical stability. First, \\(\\mathcal{L}_{\\mathrm{MSE}}\\) helps the model converge quickly during early training. Next, \\(\\mathcal{L}_{\\mathrm{GNLL}}\\) is applied to optimize for pose and uncertainty estimation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.673, + 0.849 + ], + "angle": 0, + "content": "3.4. State Estimator" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "We use UKF [33] for state estimation to refine the pose estimator inputs, integrating UWB measurements and pose estimator outputs with a statistically derived IMU model." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7088" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "The UWB measurements provide direct inter-distance data, while they are affected by body occlusion issues. Meanwhile, pose estimator outputs offer constraints and inform distance measurements; however, they suffer from occasional inaccuracies due to challenging poses or sensor noise. The UKF combines these sources, balancing the strengths and limitations of each to make the optimal estimates." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.208, + 0.248, + 0.222 + ], + "angle": 0, + "content": "3.4.1. State Definition" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.226, + 0.438, + 0.243 + ], + "angle": 0, + "content": "We define the state vector \\(\\pmb{x} \\in \\mathbb{R}^{15 \\times 3 + 15 \\times 3 + 6 \\times 3}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.252, + 0.483, + 0.287 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} = \\left[ \\begin{array}{l l l l l l l l} \\boldsymbol {p} ^ {1 2} & \\dots & \\boldsymbol {p} ^ {5 6} & \\boldsymbol {v} ^ {1 2} & \\dots & \\boldsymbol {v} ^ {5 6} & \\boldsymbol {b} ^ {1} & \\dots & \\boldsymbol {b} ^ {6} \\end{array} \\right] ^ {\\mathsf {T}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.288, + 0.484, + 0.453 + ], + "angle": 0, + "content": "where \\( \\pmb{p}^{xy} \\) represents the relative position between nodes \\( x \\) and \\( y \\), with \\( x, y \\in \\{1, 2, 3, 4, 5, 6\\} \\) and \\( x < y \\). The term \\( \\pmb{v}^{xy} \\) denotes the relative velocity between nodes \\( x \\) and \\( y \\), while \\( \\pmb{b}^1 \\) through \\( \\pmb{b}^6 \\) represent residual acceleration errors after transformation from \\( F^S \\) to \\( F^M \\). These biases primarily result from orientation errors and raw accelerometer biases \\( \\pmb{b}_a \\). All quantities are expressed in the SMPL body-centric coordinate frame \\( F^M \\). For initialization, we set \\( \\pmb{p}^{xy} \\) based on a static T-pose defined by the shape parameters \\( \\beta \\) and the SMPL model (1). The relative velocities \\( \\pmb{v}^{xy} \\) and biases are initialized to zero." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.464, + 0.264, + 0.479 + ], + "angle": 0, + "content": "3.4.2. State Propagation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.484, + 0.484, + 0.514 + ], + "angle": 0, + "content": "The control input \\(\\pmb{u}\\) consists of the accelerations \\(\\pmb{a}\\) of six nodes, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.526, + 0.483, + 0.548 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {u} = \\left[ \\begin{array}{l l l} \\boldsymbol {a} ^ {1} & \\dots & \\boldsymbol {a} ^ {6} \\end{array} \\right] ^ {\\mathsf {T}}, \\boldsymbol {u} \\in \\mathbb {R} ^ {1 8}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.56, + 0.484, + 0.62 + ], + "angle": 0, + "content": "During the prediction step, the current state \\(\\pmb{x}\\) is propagated using the control input \\(\\pmb{u}\\). The propagation model is derived from the strapdown inertial kinematic model with an acceleration error model (2), defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.633, + 0.483, + 0.67 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {v} _ {k} ^ {x y} = \\boldsymbol {v} _ {k - 1} ^ {x y} + \\left(\\boldsymbol {a} _ {k - 1} ^ {y} - \\boldsymbol {a} _ {k - 1} ^ {x}\\right) \\Delta t - \\left(\\boldsymbol {b} _ {k - 1} ^ {y} - \\boldsymbol {b} _ {k - 1} ^ {x}\\right) \\Delta t, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.675, + 0.483, + 0.733 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {p} _ {k} ^ {x y} = \\boldsymbol {p} _ {k - 1} ^ {x y} + \\boldsymbol {v} _ {k - 1} ^ {x y} \\Delta t + \\frac {1}{2} (\\boldsymbol {a} _ {k - 1} ^ {y} - \\boldsymbol {a} _ {k - 1} ^ {x}) \\Delta t ^ {2} \\\\ + \\frac {1}{2} \\left(\\boldsymbol {b} _ {k - 1} ^ {x} - \\boldsymbol {b} _ {k - 1} ^ {y}\\right) \\Delta t ^ {2}, \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.735, + 0.483, + 0.754 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {b} _ {k} ^ {x} = \\boldsymbol {b} _ {k - 1} ^ {x} + \\boldsymbol {\\eta} ^ {x}, \\boldsymbol {\\eta} ^ {x} \\sim \\mathcal {N} \\left(0, \\Sigma_ {b} ^ {2}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.767, + 0.484, + 0.827 + ], + "angle": 0, + "content": "where \\( k - 1 \\) and \\( k \\) denote consecutive time steps, and \\( \\Delta t \\) represents the interval between them. Together, equations (9), (10), and (11) form a state propagation model with added white noise, represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.841, + 0.483, + 0.858 + ], + "angle": 0, + "content": "\\[\n\\bar {\\boldsymbol {x}} _ {k} = f (\\boldsymbol {x} _ {k - 1}, \\boldsymbol {u} _ {k}) + \\mathbf {Q}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{Q}\\) is the process noise covariance matrix, derived from the characteristics of the IMUs used in the system." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.092, + 0.712, + 0.107 + ], + "angle": 0, + "content": "3.4.3. Measurement Update" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.111, + 0.907, + 0.171 + ], + "angle": 0, + "content": "The state is updated within measurement spaces using observed data \\( \\mathbf{z} \\). To accomplish this, we define a measurement model \\( h(\\mathbf{x}) \\) that maps the current state \\( \\mathbf{x} \\) to measurement spaces as:" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.179, + 0.907, + 0.202 + ], + "angle": 0, + "content": "\\[\nh (\\boldsymbol {x}) = \\left[ \\begin{array}{l l l} \\| \\boldsymbol {p} ^ {x y} \\| _ {2} & \\| \\boldsymbol {v} ^ {x y} \\| _ {2} & \\boldsymbol {p} ^ {x y} \\quad \\boldsymbol {v} ^ {x y} \\end{array} \\right] ^ {\\intercal}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.21, + 0.907, + 0.363 + ], + "angle": 0, + "content": "The UWB sensor measures the distance, \\(d^{xy}\\), between node \\(x\\) and \\(y\\), which corresponds to the distance derived from relative positions in state, \\(\\| \\pmb{p}^{xy}\\|_2\\). Additionally, from consecutive distance measurements, we derive the relative velocity norm as \\((d_k^{xy} - d_{k-1}^{xy}) / \\Delta t\\), corresponding to \\(\\| \\pmb{v}\\|_2^{xy}\\). The covariance matrix for distance measurements at time step \\(k\\), denoted as \\(\\mathbf{R}_{1,k}\\), is dynamically set based on line-of-sight conditions for each node pair, informed by \\(\\hat{\\theta}_k\\) and UWB sensor characteristics. The covariance matrix for relative velocity measurements, \\(\\mathbf{R}_{2,k}\\), is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.372, + 0.906, + 0.403 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} _ {2, k} = \\frac {\\left(\\mathbf {R} _ {1 , k} + \\mathbf {R} _ {1 , k - 1}\\right)}{\\Delta t ^ {2}}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.413, + 0.907, + 0.565 + ], + "angle": 0, + "content": "The pose estimator outputs the joint rotations, \\(\\hat{\\pmb{\\theta}}\\), along with their corresponding standard deviations, \\(\\hat{\\Sigma}\\). We employ the unscented transform [11] to transform the pose distribution to the relative position distribution. Given \\(\\Theta \\in \\mathbb{R}^{96} \\sim \\mathcal{N}(\\hat{\\pmb{\\theta}}, \\hat{\\Sigma})\\), we use Van der Merwe's scaled sigma point algorithm [30] to generate \\(n\\) sigma points, \\(\\mathcal{X} \\in \\mathbb{R}^{n \\times 96}\\), along with weights \\(\\boldsymbol{W}^m \\in \\mathbb{R}^n\\) for the mean and \\(\\boldsymbol{W}^c \\in \\mathbb{R}^n\\) for the covariance. Each sigma point in \\(\\mathcal{X}\\) is transformed through the SMPL model (1), yielding a set of sensor-relative positions denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.573, + 0.907, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Y} = \\{\\pmb {p} _ {i} ^ {x y} \\in \\mathbb {R} ^ {n \\times 1 5 \\times 3} \\mid i = 1, 2, \\dots , n \\}. \\qquad (1 5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.601, + 0.906, + 0.647 + ], + "angle": 0, + "content": "We then calculate the mean and covariance of the transformed relative positions, capturing the distribution of relative positions based on pose uncertainty:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.655, + 0.907, + 0.694 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {p}} ^ {x y} = \\sum_ {i = 1} ^ {n} \\boldsymbol {W} _ {i} ^ {m} \\mathcal {Y} _ {i}, \\tag {16}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.698, + 0.906, + 0.736 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {\\Sigma}} _ {\\hat {\\boldsymbol {p}}} ^ {2} = \\sum_ {i = 1} ^ {n} \\boldsymbol {W} _ {i} ^ {c} \\left(\\mathcal {Y} _ {i} - \\hat {\\boldsymbol {p}} ^ {x y}\\right) \\left(\\mathcal {Y} _ {i} - \\hat {\\boldsymbol {p}} ^ {x y}\\right) ^ {\\intercal}. \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.745, + 0.907, + 0.856 + ], + "angle": 0, + "content": "The transformed relative positions, \\(\\hat{\\pmb{p}}^{xy}\\), along with the measurement covariance matrix \\(\\mathbf{R}_3 = \\hat{\\Sigma}_{\\hat{\\pmb{p}}}^2\\), are used to update \\(\\pmb{p}^{xy}\\) in the state \\(\\pmb{x}\\). Similarly, we compute the relative velocity between consecutive time steps as \\((\\hat{\\pmb{p}}_k^{xy} - \\hat{\\pmb{p}}_{k-1}^{xy}) / \\Delta t\\), with its covariance matrix defined as \\(\\mathbf{R}_{4,k} = (\\mathbf{R}_{3,k} + \\mathbf{R}_{3,k-1}) / \\Delta t^2\\), to update \\(\\pmb{v}^{xy}\\). The complete measurement vector \\(\\pmb{z}\\) used for updating the state is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.876, + 0.907, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {z} = \\left[ d ^ {x y} \\quad \\frac {\\left(d _ {k} ^ {x y} - d _ {k - 1} ^ {x y}\\right)}{\\Delta t} \\quad \\hat {\\boldsymbol {p}} ^ {x y} \\quad \\frac {\\left(\\hat {\\boldsymbol {p}} _ {k} ^ {x y} - \\hat {\\boldsymbol {p}} _ {k - 1} ^ {x y}\\right)}{\\Delta t} \\right] ^ {\\intercal}. \\tag {18}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7089" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.223, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.116, + 0.194, + 0.13 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.138, + 0.486, + 0.44 + ], + "angle": 0, + "content": "We use the AMASS [18], TotalCapture [29], DIP-IMU [7], and UIP [1] datasets for training and evaluation. We synthesize inertial and distance data from the AMASS dataset, following the process in previous studies [1, 7, 34]. We select specific vertices on the body mesh to represent sensor mounting positions. Accelerations are computed as finite differences of these vertex positions over time, while orientations are obtained as the global orientations of the corresponding joints. Distance measurements are synthesized by calculating Euclidean distances between selected sensor positions, adjusted for individual body shape. Training the shape estimator requires anthropometric measurements and inter-distances with corresponding shape parameters from a large number of individuals, which is challenging to collect. Therefore, we synthesize these measurements—height \\( H \\), weight \\( W \\), and inter-distances \\( D \\) —from the 3D body mesh in a T-pose, using 479 unique body shapes (273 male and 206 female) available in the AMASS dataset. We estimate \\( H \\) and \\( W \\) following the methods in SHAPY [2], and we sample \\( D \\) from the synthesized distances." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.441, + 0.484, + 0.547 + ], + "angle": 0, + "content": "The training data for the pose estimator comprises synthesized data from the AMASS dataset and the training set of DIP-IMU. Unlike previous studies [10, 34, 35], we exclude the TotalCapture within AMASS from the training data to prevent overlap, ensuring that synthesized distances used in testing are unseen during training. The test data includes TotalCapture, the test split of DIP-IMU, and UIP." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.558, + 0.315, + 0.573 + ], + "angle": 0, + "content": "4.2. Method Implementation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.485, + 0.837 + ], + "angle": 0, + "content": "Our shape estimator was implemented using the official multi-label predictor in AutoGluon 1.1 with a medium-quality preset. Our pose estimator was implemented based on PyTorch 2.3. The training process used an Adam optimizer with a learning rate of 0.0001. The pose estimator model was trained for 350 epochs with a batch size of 512 on a single NVIDIA GeForce RTX 3090 GPU, completing in approximately 35 minutes. For initial 20 epochs, we adopted \\(\\mathcal{L}_{MSE}\\) to optimize the model, then switched to \\(\\mathcal{L}_{GNLL}\\) for the remaining epochs. For the UKF, we experimentally set parameters \\(\\alpha_{UKF} = 0.2\\), \\(\\beta_{UKF} = 1.0\\) and \\(\\kappa_{UKF} = -105\\) to control the distribution and weighting of sample points. In the unscented transform applied to pose estimates, we used \\(\\alpha_{NN} = 0.09\\), \\(\\beta_{NN} = 1.0\\), and \\(\\kappa_{NN} = -93\\). We applied a factor of 10 to \\(\\mathbf{R}_3\\) to compensate for overconfident predictions. The method achieves \\(60\\mathrm{Hz}\\) without online LOS inference and \\(30\\mathrm{Hz}\\) with it." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.313, + 0.864 + ], + "angle": 0, + "content": "4.3. Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "We evaluate our method against two categories of methods: (1) IMU-only methods, including DIP [7], TransPose [34]," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.905, + 0.333 + ], + "angle": 0, + "content": "TIP [10], PIP [35], and PNP [37], and (2) distance augmented methods, including TIP-D, PIP-D, and UIP [1]. TIP-D and PIP-D are modified versions of TIP and PIP with distance-augmented input data, and we use their results as reported by UIP. We follow the evaluation protocol in UIP by adding ideal synthetic inter-sensor distances without noise to TotalCapture and DIP-IMU to assess the impact of distance measurements independent of noise. For evaluations on the UIP dataset, which includes both IMU and UWB data, we perform RANSAC regression to calibrate UWB measurements, as described in UIP [1], and then use calibrated measurements as input for the state estimator. We do not apply outlier filtering and use default state estimator parameters due to the absence of ground-truth data and limited information on sensor-specific characteristics in the hardware setup." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.336, + 0.906, + 0.488 + ], + "angle": 0, + "content": "Metrics We use the following metrics for quantitative evaluation: 1) \\(SIP\\) Error (in degrees): the mean global angular error of the upper arms and upper legs, focusing on joints not observed by the body-worn sensors; 2) Angular Error (in degrees): the mean global angular error across all body joints; 3) Positional Error (in centimeters): the mean global joint position error across all joints; and 4) Mesh Error (in centimeters): the mean vertex position error between the reconstructed mesh and the ground-truth mesh given the mean body shape." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.906, + 0.627 + ], + "angle": 0, + "content": "Comparison with IMU-only Methods Table 1 presents a comparison of our method against IMU-only methods on the TotalCapture and DIP-IMU datasets. On the TotalCapture dataset, our approach consistently outperforms previous methods across all metrics, improving over the SOTA PNP by \\(32.4\\%\\) in angular error and \\(9.4\\%\\) in mesh error. A similar trend is also observed in the results on DIP-IMU dataset. This demonstrates that inter-sensor distances serve as strong constraints for refining estimated poses." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Comparison with Distance-augmented Methods Table 2 presents a fair comparison of our method with distance-augmented methods, as all approaches use the same input data. Additionally, all methods are trained on the AMASS dataset with the TotalCapture subset excluded, whereas the IMU-only methods in Table 1 include TotalCapture in their training data. Our method outperforms UIP, achieving a reduction in positional error of \\(21.5\\%\\) on the TotalCapture dataset and \\(35.0\\%\\) on the DIP-IMU dataset. On the UIP dataset, where sensor data quality is low and motions are more ambiguous than in the other two test datasets, our method achieves the lowest positional error, while UIP achieves the lowest SIP error. We achieve comparable results despite using less-filtered distance measurements, a simpler architecture, and non-optimal default parameters for the state estimator. This demonstrates the robustness of our approach in handling real-world sensor noise and the effectiveness of our online state estimation framework." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7090" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.127, + 0.089, + 0.872, + 0.235 + ], + "angle": 0, + "content": "
MethodSIP Error (deg)Ang Error (deg)Pos Error (cm)Mesh Error (cm)SIP Error (deg)Ang Error (deg)Pos Error (cm)Mesh Error (cm)
TotalCaptureDIP-IMU
DIP [7]18.6217.229.4211.2217.3515.367.599.05
TransPose [34]16.5812.896.557.4217.068.866.037.17
TIP [10]13.2212.305.816.8016.909.075.636.62
PIP [35]12.9312.045.616.5115.338.785.126.02
PNP [37]10.8910.454.745.4513.718.754.975.77
UMotion10.767.064.464.9414.196.353.383.93
" + }, + { + "type": "table_caption", + "bbox": [ + 0.2, + 0.247, + 0.796, + 0.262 + ], + "angle": 0, + "content": "Table 1. Comparison with state of the art IMU-only methods on TotalCapture [29] and DIP-IMU [7]." + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.274, + 0.766, + 0.415 + ], + "angle": 0, + "content": "
MethodSIP Error (deg)Pos Error (cm)SIP Error (deg)Pos Error (cm)SIP Error (deg)Pos Error (cm)
TotalCaptureDIP-IMUUIP
PIP [35]15.937.0515.986.2130.4713.62
TIP-D12.185.5115.915.2630.3413.96
PIP-D13.166.3113.795.3630.3313.27
UIP [1]11.325.4913.215.0524.1210.65
UMotion10.764.4614.193.3825.6910.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.16, + 0.426, + 0.835, + 0.44 + ], + "angle": 0, + "content": "Table 2. Comparison with distance-augmented methods on TotapCapture [29], DIP-IMU [7], and UIP [1] datasets." + }, + { + "type": "table", + "bbox": [ + 0.122, + 0.464, + 0.453, + 0.577 + ], + "angle": 0, + "content": "
MetricMean shapePredictedGT
Pos Error (cm)5.154.344.31
Mesh Error (cm)5.634.814.78
H Error (cm)5.750.39-
W Error (kg)9.450.34-
D Error (cm)3.632.332.30
C Error (cm)3.781.26-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.589, + 0.483, + 0.617 + ], + "angle": 0, + "content": "Table 3. Comparison of reconstructed body mesh errors on Total-Capture [29] using mean shape versus predicted shape parameters." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.464, + 0.925, + 0.633 + ], + "angle": 0, + "content": "
MethodSIP Error (deg)Pos Error (cm)SIP Error (deg)Pos Error (cm)
TotalCaptureDIP-IMU
J →hat9.004.3614.213.29
S →J →hat12.155.7017.354.45
PG →hat9.253.8513.632.85
S →PG →hat11.134.9615.913.91
PR →hat9.894.4213.773.22
S →PR →hat11.575.4415.744.17
Ours: S →hat10.764.4614.213.38
" + }, + { + "type": "table_caption", + "bbox": [ + 0.546, + 0.643, + 0.871, + 0.657 + ], + "angle": 0, + "content": "Table 4. Ablation study on pose estimator architecture." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.647, + 0.282, + 0.662 + ], + "angle": 0, + "content": "4.4. Module Evaluations" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.672, + 0.483, + 0.852 + ], + "angle": 0, + "content": "Shape Estimator Table 3 compares reconstructed body mesh on the TotalCapture dataset using the mean and predicted shape, both with the same estimated poses. With the predicted shape, positional and mesh errors are nearly as low as those obtained using the ground truth shape parameters. Height and weight estimates are accurate, with errors within \\(1\\mathrm{cm}\\) and \\(1\\mathrm{kg}\\), respectively. However, the mean error in circumferences \\(\\pmb{C}\\) of the chest, waist, hips, wrists, knees, and head remains large, as relevant measurements cannot be inferred solely from inter-sensor distances. Despite this limitation, the predicted shape still reconstructs a more realistic body than the mean shape model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Pose Estimator We conduct an ablation study by expanding our pose estimator with intermediate layers to estimate all joint positions, \\(J\\), sensor global positions, \\(P_{G}\\), and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.907, + 0.869 + ], + "angle": 0, + "content": "sensor relative positions, \\( \\pmb{P}_R \\), from sensor measurements \\( S = (\\hat{\\pmb{a}}, R, \\hat{\\pmb{d}}) \\). These intermediate estimates are then combined with \\( S \\) to compute the final poses \\( \\hat{\\pmb{\\theta}} \\). This layered structure follows the framework commonly used in previous methods [1, 34]. As shown in Table 4, while regressing \\( \\hat{\\pmb{\\theta}} \\) from ideal \\( J \\), \\( P_G \\), or \\( P_R \\) improves accuracy, introducing an intermediate layer, that is, \\( S \\rightarrow \\{J, P_G, P_R\\} \\), increases errors. This error propagates through the network, ultimately degrading pose estimation performance. This suggests that integrating IMU data with distance constraints may support a simpler architecture, reducing complexity while still maintaining accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "State Estimator The state estimator is designed to: 1) mitigate errors in inter-sensor distances, 2) filter and stabi" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7091" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.193, + 0.089, + 0.497, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.089, + 0.807, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.233, + 0.89, + 0.248 + ], + "angle": 0, + "content": "Figure 4. Cumulative distribution of distance error (left) and acceleration error reduction over time (right) for various fusion settings." + }, + { + "type": "image", + "bbox": [ + 0.132, + 0.277, + 0.443, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.412, + 0.462, + 0.426 + ], + "angle": 0, + "content": "Figure 5. Joint positional error for different fusion settings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.456, + 0.485, + 0.775 + ], + "angle": 0, + "content": "imize accelerations, and 3) improve the accuracy of estimated poses. To examine it we conduct an ablation study by varying fusion measurements on the TotalCapture [29] with synthesized LOS-related noisy distances. Fig. 4 (left) shows cumulative distributions of absolute distance error for different fusion settings. The raw distances show an average error of \\(9.20\\mathrm{cm}\\) with a standard deviation of \\(8.94\\mathrm{cm}\\). By fusing IMU, UWB, and pose data, this error is reduced to \\(2.42\\mathrm{cm}\\) with a standard deviation of \\(2.39\\mathrm{cm}\\), demonstrating the effectiveness of data fusion for mitigating distance errors. Fig. 4 (right) shows the reduction in acceleration errors. The state estimator converges within a few seconds and adapts gradually, aligning with IMU characteristics. Fig. 5 presents joint positional errors over time for different fusion setups. Without filtering, positional error fluctuates with distance measurement quality. Fusing IMU with either UWB or pose data alone offers limited improvement because distance and pose outliers remain. The combined fusion of IMU, UWB, and pose data achieves the lowest positional error, validating that our state estimator refines pose accuracy by utilizing all available measurements." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.787, + 0.303, + 0.803 + ], + "angle": 0, + "content": "4.5. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We developed a prototype (see supplementary material for details) that integrates BNO086 IMU and DW3000 UWB sensors to demonstrate our proposed method. Fig. 6 presents visualizations comparing pose estimates for challenging motions, illustrating the improvements of our method in disambiguating poses over IMU-only methods" + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.27, + 0.859, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.503, + 0.907, + 0.531 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison of pose estimates from the test data collected using our developed prototype." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.562, + 0.818, + 0.578 + ], + "angle": 0, + "content": "and enhanced pose stability compared to UIP." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.599, + 0.771, + 0.615 + ], + "angle": 0, + "content": "5. Conclusion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.907, + 0.779 + ], + "angle": 0, + "content": "In this work, we present UMotion, an uncertainty-driven framework for 3D human shape and pose estimation that integrates six IMU and UWB sensors within a state estimation system. UMotion incorporates uncertainties from IMU, UWB, and pose estimates under individual body constraints to iteratively enhance confidence in both sensor measurements and pose accuracy. Our experiments on synthetic and real-world datasets demonstrate that UMotion outperforms existing SOTA methods in pose accuracy, and effectively stabilizes sensor measurements and pose estimation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.902 + ], + "angle": 0, + "content": "However, our shape estimator requires adaptation to specific sensor configurations and conditions, and, with limited training data, it may struggle with unique body variations. The state estimator also requires careful parameter tuning to reflect sensor characteristics accurately. Additionally, our pose estimator simplifies constraints, which may limit its performance. As future work, adding a physics-aware module could potentially enhance our method's robustness." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7092" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Rayan Armani, Changlin Qian, Jiaxi Jiang, and Christian Holz. Ultra inertial poser: Scalable motion capture and tracking from sparse inertial sensors and ultra-wideband ranging. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[2] Vasileios Choutas, Lea Müller, Chun-Hao P Huang, Siyu Tang, Dimitrios Tzionas, and Michael J Black. Accurate 3d body shape regression using metric and semantic attributes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2718-2728, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.269, + 0.483, + 0.352 + ], + "angle": 0, + "content": "[3] Peng Dai, Yang Zhang, Tao Liu, Zhen Fan, Tianyuan Du, Zhuo Su, Xiaozheng Zheng, and Zeming Li. Hmd-poser: On-device real-time human motion tracking from scalable sparse observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 874-884, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.483, + 0.422 + ], + "angle": 0, + "content": "[4] Nathan DeVrio, Vimal Mollyn, and Chris Harrison. Smartposer: Arm pose estimation with a smartphone and smartwatch using uwb and imu data. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, pages 1-11, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.483, + 0.478 + ], + "angle": 0, + "content": "[5] Nick Erickson, Jonas Mueller, Alexander Shirkov, Hang Zhang, Pedro Larroy, Mu Li, and Alexander Smola. Autogluon-tabular: Robust and accurate automl for structured data. arXiv preprint arXiv:2003.06505, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.48, + 0.482, + 0.507 + ], + "angle": 0, + "content": "[6] S Hochreiter. Long short-term memory. Neural Computation MIT-Press, 1997. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.483, + 0.577 + ], + "angle": 0, + "content": "[7] Yinghao Huang, Manuel Kaufmann, Emre Aksan, Michael J Black, Otmar Hilliges, and Gerard Pons-Moll. Deep inertial poser: Learning to reconstruct human pose from sparse inertial measurements in real time. ACM Transactions on Graphics (TOG), 37(6):1-15, 2018. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.483, + 0.648 + ], + "angle": 0, + "content": "[8] Jiaxi Jiang, Paul Streli, Huajian Qiu, Andreas Fender, Larissa Laich, Patrick Snape, and Christian Holz. Avatarposer: Articulated full-body pose tracking from sparse motion sensing. In European conference on computer vision, pages 443-460. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.65, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[9] Jiaxi Jiang, Paul Streli, Manuel Meier, and Christian Holz. Egoposer: Robust real-time ego-body pose estimation in large scenes. arXiv preprint arXiv:2308.06493, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.692, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[10] Yifeng Jiang, Yuting Ye, Deepak Gopinath, Jungdam Won, Alexander W Winkler, and C Karen Liu. Transformer inertial poser: Real-time human motion reconstruction from sparse imus with simultaneous terrain generation. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.483, + 0.817 + ], + "angle": 0, + "content": "[11] Simon J Julier, Jeffrey K Uhlmann, and Hugh F Durrant-Whyte. A new approach for filtering nonlinear systems. In Proceedings of 1995 American Control Conference-ACC'95, pages 1628-1632. IEEE, 1995. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[12] Manuel Kaufmann, Yi Zhao, Chengcheng Tang, Lingling Tao, Christopher Twigg, Jie Song, Robert Wang, and Otmar Hilliges. Em-posed: 3d human pose estimation from sparse electromagnetic trackers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11510-11520, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[13] Jiye Lee and Hanbyul Joo. Mocap everyone everywhere: Lightweight motion capture with smartwatches and a head-mounted camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1091-1100, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[14] Sunmin Lee, Sebastian Starke, Yuting Ye, Jungdam Won, and Alexander Winkler. Questenvsim: Environment-aware simulated motion tracking from sparse sensors. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-9, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[15] Han Liang, Yannan He, Chengfeng Zhao, Mutian Li, Jingya Wang, Jingyi Yu, and Lan Xu. Hybridcap: Inertia-aid monocular capture of challenging human motions. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1539–1548, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.292, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[16] Huajun Liu, Xiaolin Wei, Jinxiang Chai, Inwoo Ha, and Taehyun Rhee. Realtime human motion control with a small number of inertial sensors. In Symposium on interactive 3D graphics and games, pages 133-140, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[17] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. In *Seminal Graphics Papers: Pushing the Boundaries*, Volume 2, pages 851–866. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.406, + 0.907, + 0.474 + ], + "angle": 0, + "content": "[18] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5442-5451, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.476, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[19] Vimal Mollyn, Riku Arakawa, Mayank Goel, Chris Harrison, and Karan Ahuja. Imuposer: Full-body pose estimation using imus in phones, watches, and earbuds. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, pages 1-12, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.904, + 0.574 + ], + "angle": 0, + "content": "[20] Ian Oppermann, Matti Hämäläinen, and Jari Inatti. *UWB: theory and applications*. John Wiley & Sons, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.577, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[21] Shaohua Pan, Qi Ma, Xinyu Yi, Weifeng Hu, Xiong Wang, Xingkang Zhou, Jijunnan Li, and Feng Xu. Fusing monocular images and sparse imu signals for real-time human motion capture. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.648, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[22] Gerard Pons-Moll, Andreas Baak, Thomas Helten, Meinard Müller, Hans-Peter Seidel, and Bodo Rosenhahn. Multisensor-fusion for 3d full-body human motion capture. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 663-670. IEEE, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.732, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[23] Jose Luis Ponton, Haoran Yun, Andreas Aristidou, Carlos Andujar, and Nuria Pelechano. Sparseposer: Real-time full-body motion reconstruction from sparse data. ACM Transactions on Graphics, 43(1):1-14, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[24] Sergi Pujades, Betty Mohler, Anne Thaler, Joachim Tesch, Naureen Mahmood, Nikolas Hesse, Heinrich H Bülthoff, and Michael J Black. The virtual caliper: rapid creation of metrically accurate avatars from 3d measurements. IEEE transactions on visualization and computer graphics, 25(5): 1887-1897, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Yiming Ren, Chengfeng Zhao, Yannan He, Peishan Cong, Han Liang, Jingyi Yu, Lan Xu, and Yuexin Ma. Lidar-aid" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7093" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.483, + 0.135 + ], + "angle": 0, + "content": "inertial poser: Large-scale human motion capture by sparse inertial and lidar sensors. IEEE Transactions on Visualization and Computer Graphics, 29(5):2337-2347, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.137, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[26] Xu Ru, Nian Gu, Hang Shang, and Heng Zhang. Mems inertial sensor calibration technology: Current status and future trends. Micromachines, 13(6):879, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.181, + 0.482, + 0.223 + ], + "angle": 0, + "content": "[27] Martin Schepers, Matteo Giuberti, Giovanni Bellusci, et al. Xsens mvn: Consistent tracking of human motion using inertial sensing. Xsens Technol, 1(8):1-8, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.225, + 0.482, + 0.266 + ], + "angle": 0, + "content": "[28] Mike Schuster and Kuldip K Paliwal. Bidirectional recurrent neural networks. IEEE transactions on Signal Processing, 45 (11):2673-2681, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.268, + 0.482, + 0.325 + ], + "angle": 0, + "content": "[29] Matthew Trumble, Andrew Gilbert, Charles Malleson, Adrian Hilton, and John P Collomosse. Total capture: 3d human pose estimation fusing video and inertial sensors. In BMVC, pages 1-13. London, UK, 2017. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.327, + 0.482, + 0.369 + ], + "angle": 0, + "content": "[30] Rudolph Van Der Merwe. Sigma-point Kalman filters for probabilistic inference in dynamic state-space models. Oregon Health & Science University, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[31] Tom Van Wouwe, Seunghwan Lee, Antoine Falisse, Scott Delp, and C Karen Liu. Diffusion inertial poser: Human motion reconstruction from arbitrary sparse imu configurations. arXiv preprint arXiv:2308.16682, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.429, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[32] Timo Von Marcard, Bodo Rosenhahn, Michael J Black, and Gerard Pons-Moll. Sparse inertial poser: Automatic 3d human pose estimation from sparse imus. In Computer graphics forum, pages 349-360. Wiley Online Library, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.5, + 0.482, + 0.57 + ], + "angle": 0, + "content": "[33] Eric A Wan and Rudolph Van Der Merwe. The unscented kalman filter for nonlinear estimation. In Proceedings of the IEEE 2000 adaptive systems for signal processing, communications, and control symposium (Cat. No. 00EX373), pages 153-158. Ieee, 2000. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.572, + 0.482, + 0.627 + ], + "angle": 0, + "content": "[34] Xinyu Yi, Yuxiao Zhou, and Feng Xu. Transpose: Real-time 3d human translation and pose estimation with six inertial sensors. ACM Transactions On Graphics (TOG), 40(4):1-13, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.63, + 0.482, + 0.713 + ], + "angle": 0, + "content": "[35] Xinyu Yi, Yuxiao Zhou, Marc Habermann, Soshi Shimada, Vladislav Golyanik, Christian Theobalt, and Feng Xu. Physical inertial poser (pip): Physics-aware real-time human motion tracking from sparse inertial sensors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13167-13178, 2022. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.715, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[36] Xinyu Yi, Yuxiao Zhou, Marc Habermann, Vladislav Golyanik, Shaohua Pan, Christian Theobalt, and Feng Xu. Ecolocate: Real-time motion capture, localization, and mapping with sparse body-mounted sensors. ACM Transactions on Graphics (TOG), 42(4):1-17, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.787, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[37] Xinyu Yi, Yuxiao Zhou, and Feng Xu. Physical non-inertial poser (pnp): Modeling non-inertial effects in sparse-inertial human motion capture. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[38] Faheem Zafari, Athanasios Gkelias, and Kin K Leung. A survey of indoor localization systems and technologies. IEEE Communications Surveys & Tutorials, 21(3):2568-2599, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[39] Yu Zhang, Songpengcheng Xia, Lei Chu, Jiarui Yang, Qi Wu, and Ling Pei. Dynamic inertial poser (dynaip): Part-based motion dynamics learning for enhanced human pose estimation with sparse inertial sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1889-1899, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.906, + 0.245 + ], + "angle": 0, + "content": "[40] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5745-5753, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.248, + 0.906, + 0.318 + ], + "angle": 0, + "content": "[41] Chengxu Zuo, Yiming Wang, Lishuang Zhan, Shihui Guo, Xinyu Yi, Feng Xu, and Yipeng Qin. Loose inertial poser: Motion capture with imu-attached loose-wear jacket. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2209–2219, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7094" + } + ] +] \ No newline at end of file diff --git a/2025/UMotion_ Uncertainty-driven Human Motion Estimation from Inertial and Ultra-wideband Units/20c112a8-f19a-4ffc-8e5b-a5d02b15caa1_origin.pdf b/2025/UMotion_ Uncertainty-driven Human Motion Estimation from Inertial and Ultra-wideband Units/20c112a8-f19a-4ffc-8e5b-a5d02b15caa1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fd2c93ce0a9f02d068e56ef46b53f9093a9d8dcb --- /dev/null +++ b/2025/UMotion_ Uncertainty-driven Human Motion Estimation from Inertial and Ultra-wideband Units/20c112a8-f19a-4ffc-8e5b-a5d02b15caa1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4095cd1664be4aa3350c7af5b7fcb8a12822e60f9db966dfca4b1aa4f6bb9135 +size 5325137 diff --git a/2025/UMotion_ Uncertainty-driven Human Motion Estimation from Inertial and Ultra-wideband Units/full.md b/2025/UMotion_ Uncertainty-driven Human Motion Estimation from Inertial and Ultra-wideband Units/full.md new file mode 100644 index 0000000000000000000000000000000000000000..74d5bf2683a17f7cf5ad1db359e822d43452565b --- /dev/null +++ b/2025/UMotion_ Uncertainty-driven Human Motion Estimation from Inertial and Ultra-wideband Units/full.md @@ -0,0 +1,326 @@ +# UMotion: Uncertainty-driven Human Motion Estimation from Inertial and Ultra-wideband Units + +Huakun Liu Hiroki Ota Xin Wei Yutaro Hirao + +Monica Perusquía-Hernández +Hideaki Uchiyama +Kiyoshi Kiyokawa +Nara Institute of Science and Technology, Japan + +{liu.huakun.li0, ota.hiroki.oc6, wei.xin.wy0, yutaro.hirao}@is.naist.jp + +{m.perusquia, hideaki.uchiyama, kiyo}@is.naist.jp + +# Abstract + +Sparse wearable inertial measurement units (IMUs) have gained popularity for estimating 3D human motion. However, challenges such as pose ambiguity, data drift, and limited adaptability to diverse bodies persist. To address these issues, we propose UMotion, an uncertainty-driven, online fusing-all state estimation framework for 3D human shape and pose estimation, supported by six integrated, body-worn ultra-wideband (UWB) distance sensors with IMUs. UWB sensors measure inter-node distances to infer spatial relationships, aiding in resolving pose ambiguities and body shape variations when combined with anthropometric data. Unfortunately, IMUs are prone to drift, and UWB sensors are affected by body occlusions. Consequently, we develop a tightly coupled Unscented Kalman Filter (UKF) framework that fuses uncertainties from sensor data and estimated human motion based on individual body shape. The UKF iteratively refines IMU and UWB measurements by aligning them with uncertain human motion constraints in real-time, producing optimal estimates for each. Experiments on both synthetic and real-world datasets demonstrate the effectiveness of UMotion in stabilizing sensor data and the improvement over state of the art in pose accuracy. Code is available at: https://github.com/kk9six/umotion. + +# 1. Introduction + +Estimating 3D human motion from wearable sensors has become increasingly popular due to their portability, accessibility, and versatility. Wearable sensors, such as inertial measurement units (IMUs), enable continuous monitoring of body motion measurements across unrestricted spaces. These advances shift motion capture from controlled laboratory environments to everyday settings [13, 19, 32, 41]. This transition benefits fields such as healthcare, sports + +![](images/09b2e00879f5d00521d89acf4a37ac88a9833eb37f7d362bf7e3cd28695e2e25.jpg) +Figure 1. UMotion integrates IMU-UWB data inputs and pose outputs uniformly under uncertainty, constrained by individual body structure. The online state estimation framework iteratively refines sensor data confidence and stabilizes pose estimation, reducing ambiguities and improving robustness. + +performance, ergonomics, and emerging areas in humancomputer interaction [3, 8, 14, 31]. + +One of the widely chosen wearable sensors for 3D human motion estimation is IMU. Commercial systems, such as Xsens [27], utilize 17 or more IMUs for comprehensive pose coverage. While highly accurate, these densely placed IMUs are inconvenient and intrusive. Recent studies have reduced the required number of IMUs to just six—placed on the forearms, lower legs, pelvis, and head—while still achieving promising performance through data-driven methods [1, 7, 10, 31, 32, 34, 35, 37, 39]. With fewer IMUs, pose estimation becomes under-constrained and prone to ambiguity. Recent work has attempted to disambiguate poses by incorporating temporal consistency, physics-based constraints, or additional, easy-to-integrate sensors [1, 10, 35]. However, these methods still face significant challenges, including noisy sensor data, body shape variations, and ambiguities arising from under-constrained setups with sparse sensors. + +In this work, we propose UMotion, an uncertainty-driven human motion estimation framework that combines online state estimation with an integrated system of six inertial and + +ultra-wideband (UWB) distance sensors. Prior work often infers poses from sensor data without considering the reverse influence. UMotion uniformly treats input and output under uncertainty and body-specific constraints, blending data to optimize estimates (see Fig. 1). Our approach maintains inter-sensor distances as a core system state, capturing spatial relationships between sensor nodes. These distances, combined with basic anthropometric measurements (height and weight) and IMU data, serve as inputs for our learning-based shape and pose estimators. Since sensors are inherently noisy, we propose a tightly coupled online state estimation system that associates IMU and UWB sensor measurements and pose estimates using an Unscented Kalman Filter (UKF) and an uncertainty propagation method, iteratively correcting errors and stabilizing pose estimation in real-time. Our main contributions are: + +- An ensemble learning-based human shape estimation approach using distance constraints and anthropometric measurements from six integrated IMU-UWB sensors. +- A learning-based method for human pose estimation from inertial and distance constraints, incorporating mesh distribution estimation. +- A filtering-based state estimation system that couples sensor measurements, pose estimates, and body variations in real-time to improve sensor stability and pose accuracy. + +# 2. Related Work + +Pose Estimation from Sparse Inertial Sensors In contrast to the predominant use of vision-based methods, wearable sensors, specifically IMUs, offer greater freedom and flexibility. Von Marcard et al. [32] present SIP that makes inertial pose estimation practical by using only six IMUs attached to the body, combined with an offline iterative SMPL body model [17] pose optimization. Huang et al. [7] propose a real-time pose estimation method, DIP, which uses a bidirectional recurrent neural network (biRNN) [28] to learn the mapping from a sequence of six IMU measurements to SMPL body poses. To obtain sufficient training data, they synthesize IMU measurements from the AMASS dataset [18], further advancing the development of learning-based methods for sparse sensor motion capture. + +Following previous studies, TransPose [34] refines pose estimation by decomposing the end-to-end framework into a multi-stage process with intermediate joint position estimation. To disambiguate the poses, PIP [35] proposes a physics-based motion optimization, while TIP [10] incorporates a conditional Transformer model for plausible terrain generation. Training data is crucial for learning-based methods. Unlike prior studies relying on synthetic IMU data, DynalIP [39] adapts real IMU data from diverse human skeleton formats to the target SMPL model and shows superior performance when training with real sensor data. Similarly focusing on the data, PNP [37] recently addresses + +limitations in existing IMU synthesis by incorporating noninertial effects and fictitious forces, enhancing estimation robustness through a physics-informed neural network and realistic IMU synthesis. + +In contrast to methods focusing solely on pose estimation, our approach treats human shape and pose as equally important. By integrating shape information, we establish a tight connection between sensor data and estimated motions, forming a positive feedback loop that enhances the entire process. Additionally, most previous studies fail to fully utilize IMU accelerations due to high noise and drift issues. Within our framework, accelerations serve as control inputs for state estimation, undergoing continuous correction and acting as a crucial component in tracking spatial relationships among sensors. + +Pose Estimation from Hybrid Sensors In addition to IMUs, various wearable and hybrid sensor systems have been explored to overcome inherent IMU limitations such as restricted positional awareness [9, 12, 13, 15, 16, 21–23, 25, 36]. The closest works to ours are SmartPoser [4] and Ultra Inertial Poser (UIP) [1], both of which integrate UWB and IMU sensors for pose estimation. UWB sensors complement IMUs with additional distance information between sensors while preserving the portability and flexibility of tracking devices. SmartPoser [4] focuses on wearable arm pose estimation, combining UWB measurements with IMU data using an off-the-shelf smartwatch and smartphone. UIP [1] integrates six UWB sensors with IMUs for full-body pose estimation. Inter-sensor distances, estimated using an EKF that fuses IMU and UWB data, help reduce global translation drift and minimize position jitter compared to inertial-only tracking. However, body occlusion frequently affects certain node pairs, e.g., head-knee, causing the EKF to fail in estimating distances due to rapid IMU drift and unreliable UWB measurements, resulting in unstable distance data. Our approach addresses this limitation by using output pose uncertainties and body-specific constraints as observations to refine input sensor measurements, improving the stability of both IMU and UWB data. + +# 3. Method + +# 3.1. Preliminaries + +SMPL body model We use the SMPL model [17] to represent human motion. SMPL decomposes the human body into pose parameters, $\pmb{\theta} \in \mathbb{R}^{23 \times 3 + 3}$ , which define relative rotations of 23 joints and the global root joint orientation in axis-angle form, and shape parameters, $\beta \in \mathbb{R}^{10}$ , capturing body shape variations across individuals. The model is defined by a linear blend-skinning (LBS) function: + +$$ +M (\boldsymbol {\beta}, \boldsymbol {\theta}) = W \left(\bar {\mathbf {T}} + B _ {s} (\boldsymbol {\beta}) + B _ {p} (\boldsymbol {\theta}), J (\boldsymbol {\beta}), \boldsymbol {\theta}, \mathcal {W}\right), \tag {1} +$$ + +where $B_{s}(\boldsymbol {\beta})$ and $B_{p}(\boldsymbol {\theta})$ are shape and pose blend shapes that deform the template mesh $\bar{\mathbf{T}}$ in the zero pose. The mesh is then reposed according to specified poses $\theta$ , combining with joint positions $J(\beta)$ and blend-skinning weights $\mathcal{W}$ . Inertial Measurement Unit (IMU) A nine-axis MEMS IMU comprises an accelerometer, gyroscope, and magnetometer, measuring accelerations in the sensor local frame $F^S$ , angular velocities in $F^S$ , and magnetic field strengths relative to the Earth's magnetic field, respectively. A commonly used error model for measured accelerations of one IMU, $\pmb{a}\in \mathbb{R}^{3}$ , is given by [26]: + +$$ +\boldsymbol {a} = \boldsymbol {a} _ {\text {t r u e}} + \boldsymbol {b} _ {a} + \boldsymbol {n} _ {a}, \tag {2} +$$ + +where $\pmb{a}_{\mathrm{true}}$ is the true acceleration, $\pmb{b}_a$ is the bias, modeled as a random walk process with noise $\eta$ following a zero-mean Gaussian distribution with standard deviation $\Sigma_{b}$ , and $\pmb{n}_a$ is high-frequency white noise. In our work, after a frame calibration process, accelerations are represented as $\pmb{a}^{M}$ in the SMPL body-centric frame $F^{M}$ , and orientations are represented as $\pmb{R}^{MB}$ , indicating the rotation from bone frame $F^{B}$ to $F^{M}$ . For simplicity, we denote $\pmb{a}$ as $\pmb{a}^{M}$ and $\pmb{R}$ as $\pmb{R}^{MB}$ in the following sections. Details on the calibration process are available in supplementary materials. + +Ultra-wideband (UWB) UWB uses the time-of-flight technique to measure distances between two devices, with one acting as the transmitter and the other as the receiver. Unlike other RF techniques, such as WiFi and Bluetooth, UWB operates over a wide frequency range and transmits short pulses, achieving centimeter-level distance measurements with minimal interference [20]. Consequently, it is widely used in high-precision indoor localization systems with multiple fixed anchors and movable tags [38]. In our work, all six integrated UWB sensors are movable and alternately function as both transmitters and receivers, capturing inter-distances between each sensor pair. + +# 3.2. Framework Design + +As shown in Fig. 2, our method uses six IMU-UWB sensors and consists of three main modules: a shape estimator, a pose estimator, and a state estimator. The shape estimator takes anthropometric measurements and inter-distances in a T-pose as input, outputting shape parameters $\hat{\beta}$ of the SMPL model. The pose estimator uses filtered frame-aligned, root-normalized IMU measurements and inter-distances as input to estimate pose parameters, $\hat{\theta}$ , and corresponding uncertainties, $\hat{\Sigma}$ . Given our sensor placement, we do not observe the movement of the hands and feet. Therefore, similar to previous studies, we estimate poses for only 16 joints, including the root joint, while excluding the hand and foot joints. For simplicity, we denote this reduced set of pose parameters as $\theta \in \mathbb{R}^{16\times 3}$ , while $\theta^{24}$ refers to the full set of pose parameters for 24 joints. Notations with a hat symbol, e.g., $\hat{\theta}$ , indicate the corresponding estimated values. Addi + +tionally, we do not estimate global translation, as inferring it from only body-worn IMU and UWB sensors without external references would lead to unbounded error accumulation. The state estimator integrates accelerations, UWB measurements, and the estimated poses with uncertainties to track the sensor relative positions, velocities, and acceleration biases of each node at each time step. These refined estimates are then used to refine the inputs to the pose estimator. + +The estimation framework is driven by inherent uncertainties present in the system. Specifically, noisy IMU data combined with tracked inter-sensor distances serve as the primary inputs for pose estimation. The generated poses, along with their associated uncertainties that reflect sensor noise, are further constrained by the human model, which in turn refines sensor measurements and distance estimates. This feedback loop propagates uncertainties through the system, continuously updating the belief in system states based on the reliability of each data source—IMU, UWB, and pose estimator with body constraints, ultimately producing robust and accurate estimation. + +# 3.3. Shape and Pose Estimator + +# 3.3.1. Shape Estimator + +Spatial inter-distances provide data that approximates the rough skeletal structure of an individual. Previous studies, such as Virtual Caliper [24] and SHAPY [2], have demonstrated a linear relationship between body shape and body measurements. Building on these findings, we adopt an ensemble learning-based shape estimator that uses experimentally selected anthropometric measurements, defined as: + +$$ +\hat {\boldsymbol {\beta}} = S E (H, W, \boldsymbol {D}). \tag {3} +$$ + +Basic body measurements, height $H \in \mathbb{R}^1$ and weight $W \in \mathbb{R}^1$ , are easily accessible and provide foundational information about human body proportions. Experimentally selected inter-distances between sensors, $D \in \mathbb{R}^7$ , capture relative limb lengths and body structural details that height and weight alone cannot provide. + +As shown in Fig. 3, we place virtual sensors on the body mesh and conduct line-of-sight (LOS) simulation experiments with real-world tests to identify and select 7 repeatable distances out of 15 possible options. While the selection may vary based on device characteristics, the chosen distances offer a reliable foundation for adaptation. For model training, we use AutoGluon [5], an AutoML framework that automatically trains and ensembles 11 basic machine learning models, which is more stable than a single linear regression method in our validation. The estimated shape parameters $\hat{\beta}$ are then used to reconstruct realistic human bodies and to propagate estimated poses to spatial constraints in the state estimator. + +![](images/0d3ccd4fc14c3d62ceb39f841417717db676e66dcd35a8234acc914a351f6748.jpg) +Figure 2. Overview of UMotion, consisting of three main modules: the shape estimator, pose estimator, and state estimator. The shape estimator takes anthropometric measurements and inter-distances in a T-pose as input, outputting shape parameters that reconstruct a realistic body and impose strong constraints on the system. The pose estimator receives filtered IMU data and inter-distances from the state estimator to predict poses, which are fed back to refine state estimates. The entire system integrates IMU, UWB, and estimated poses within the context of individual body structure to continuously update and improve motion estimation. + +![](images/c15a1643ba69f95646107fa37e9d6d87a58a1a336efea4167f94def6aa070a21.jpg) +Figure 3. Visualization of selected inter-sensor distances used in the shape estimator. We place virtual sensors on the body mesh at sensor mounting points and conduct line-of-sight simulation experiments. The plots display the temporal changes in UWB measurements between various sensor pairs over time. + +# 3.3.2. Pose Estimator + +Our pose estimator is based on unidirectional multi-layer long short-term memory (LSTM) [6] recurrent neural network (RNN). The unidirectional LSTM preserves only past information, achieving real-time estimation without requiring future data. Given six IMU-UWB sensors, we use filtered accelerations, $\hat{a} \in \mathbb{R}^{18}$ , and orientations in 9D rotation matrix form, $\pmb{R} \in \mathbb{R}^{54}$ , along with 15 inter-distances $\hat{\pmb{d}} \in \mathbb{R}^{15}$ , as inputs. The estimator then outputs the poses $\hat{\theta}_{6D} \in \mathbb{R}^{96}$ for the target joints, represented in a 6D rotation matrix form [40]. Additionally, in the final layer, the network outputs the logarithm of the estimation uncertainty, $\hat{\Sigma} \in \mathbb{R}^{96}$ , capturing prediction confidence. Consequently, + +the pose estimator is defined as: + +$$ +\hat {\boldsymbol {\theta}} _ {6 D}, \hat {\boldsymbol {\Sigma}} = P E (\hat {\boldsymbol {a}}, \boldsymbol {R}, \hat {\boldsymbol {d}}). \tag {4} +$$ + +To handle ambiguity in the initial frame, we adopt a learning-based initialization strategy inspired by PIP [35]. A fully-connected convolutional neural network regresses the initial hidden state of the RNN based on the initial global poses, $\hat{\pmb{\theta}}_0^{24}$ , and joint positions, $\hat{J}_0 \in \mathbb{R}^{72}$ . This approach provides the pose estimator with a well-informed initialization, adapting to different bodies. + +We use two loss functions to train the pose estimator: Mean Squared Error (MSE) and Gaussian Negative Log Likelihood (GNLL) loss. MSE is defined as: + +$$ +\mathcal {L} _ {\mathrm {M S E}} \left(\boldsymbol {\theta} _ {6 D}, \hat {\boldsymbol {\theta}} _ {6 D}\right) = \frac {1}{n} \sum_ {i = 1} ^ {n} \left\| \boldsymbol {\theta} _ {6 D} - \hat {\boldsymbol {\theta}} _ {6 D} \right\| _ {2} ^ {2}, \tag {5} +$$ + +where $\pmb{\theta}_{6D}$ denotes the ground truth poses and $n$ is the total number of frames. The GNLL, $\mathcal{L}_{\mathrm{GNLL}}(\pmb{\theta}_{6D},\hat{\pmb{\theta}}_{6D},\hat{\pmb{\Sigma}})$ + +$$ +\frac {1}{2} \left(\log \left(\max \left(\hat {\boldsymbol {\Sigma}} ^ {2}, \varepsilon_ {\min }\right)\right) + \frac {(\hat {\boldsymbol {\theta}} _ {6 D} - \boldsymbol {\theta} _ {6 D}) ^ {2}}{\max \left(\hat {\boldsymbol {\Sigma}} ^ {2} , \varepsilon_ {\min }\right)}\right), \tag {6} +$$ + +where $\varepsilon_{\mathrm{min}}$ , set to $10^{-6}$ in our implementation, is used for numerical stability. First, $\mathcal{L}_{\mathrm{MSE}}$ helps the model converge quickly during early training. Next, $\mathcal{L}_{\mathrm{GNLL}}$ is applied to optimize for pose and uncertainty estimation. + +# 3.4. State Estimator + +We use UKF [33] for state estimation to refine the pose estimator inputs, integrating UWB measurements and pose estimator outputs with a statistically derived IMU model. + +The UWB measurements provide direct inter-distance data, while they are affected by body occlusion issues. Meanwhile, pose estimator outputs offer constraints and inform distance measurements; however, they suffer from occasional inaccuracies due to challenging poses or sensor noise. The UKF combines these sources, balancing the strengths and limitations of each to make the optimal estimates. + +# 3.4.1. State Definition + +We define the state vector $\pmb{x} \in \mathbb{R}^{15 \times 3 + 15 \times 3 + 6 \times 3}$ as: + +$$ +\boldsymbol {x} = \left[ \begin{array}{l l l l l l l l} \boldsymbol {p} ^ {1 2} & \dots & \boldsymbol {p} ^ {5 6} & \boldsymbol {v} ^ {1 2} & \dots & \boldsymbol {v} ^ {5 6} & \boldsymbol {b} ^ {1} & \dots & \boldsymbol {b} ^ {6} \end{array} \right] ^ {\mathsf {T}}, \tag {7} +$$ + +where $\pmb{p}^{xy}$ represents the relative position between nodes $x$ and $y$ , with $x, y \in \{1, 2, 3, 4, 5, 6\}$ and $x < y$ . The term $\pmb{v}^{xy}$ denotes the relative velocity between nodes $x$ and $y$ , while $\pmb{b}^1$ through $\pmb{b}^6$ represent residual acceleration errors after transformation from $F^S$ to $F^M$ . These biases primarily result from orientation errors and raw accelerometer biases $\pmb{b}_a$ . All quantities are expressed in the SMPL body-centric coordinate frame $F^M$ . For initialization, we set $\pmb{p}^{xy}$ based on a static T-pose defined by the shape parameters $\beta$ and the SMPL model (1). The relative velocities $\pmb{v}^{xy}$ and biases are initialized to zero. + +# 3.4.2. State Propagation + +The control input $\pmb{u}$ consists of the accelerations $\pmb{a}$ of six nodes, defined as: + +$$ +\boldsymbol {u} = \left[ \begin{array}{l l l} \boldsymbol {a} ^ {1} & \dots & \boldsymbol {a} ^ {6} \end{array} \right] ^ {\mathsf {T}}, \boldsymbol {u} \in \mathbb {R} ^ {1 8}. \tag {8} +$$ + +During the prediction step, the current state $\pmb{x}$ is propagated using the control input $\pmb{u}$ . The propagation model is derived from the strapdown inertial kinematic model with an acceleration error model (2), defined as follows: + +$$ +\boldsymbol {v} _ {k} ^ {x y} = \boldsymbol {v} _ {k - 1} ^ {x y} + \left(\boldsymbol {a} _ {k - 1} ^ {y} - \boldsymbol {a} _ {k - 1} ^ {x}\right) \Delta t - \left(\boldsymbol {b} _ {k - 1} ^ {y} - \boldsymbol {b} _ {k - 1} ^ {x}\right) \Delta t, \tag {9} +$$ + +$$ +\begin{array}{l} \boldsymbol {p} _ {k} ^ {x y} = \boldsymbol {p} _ {k - 1} ^ {x y} + \boldsymbol {v} _ {k - 1} ^ {x y} \Delta t + \frac {1}{2} (\boldsymbol {a} _ {k - 1} ^ {y} - \boldsymbol {a} _ {k - 1} ^ {x}) \Delta t ^ {2} \\ + \frac {1}{2} \left(\boldsymbol {b} _ {k - 1} ^ {x} - \boldsymbol {b} _ {k - 1} ^ {y}\right) \Delta t ^ {2}, \tag {10} \\ \end{array} +$$ + +$$ +\boldsymbol {b} _ {k} ^ {x} = \boldsymbol {b} _ {k - 1} ^ {x} + \boldsymbol {\eta} ^ {x}, \boldsymbol {\eta} ^ {x} \sim \mathcal {N} \left(0, \Sigma_ {b} ^ {2}\right), \tag {11} +$$ + +where $k - 1$ and $k$ denote consecutive time steps, and $\Delta t$ represents the interval between them. Together, equations (9), (10), and (11) form a state propagation model with added white noise, represented as: + +$$ +\bar {\boldsymbol {x}} _ {k} = f (\boldsymbol {x} _ {k - 1}, \boldsymbol {u} _ {k}) + \mathbf {Q}, \tag {12} +$$ + +where $\mathbf{Q}$ is the process noise covariance matrix, derived from the characteristics of the IMUs used in the system. + +# 3.4.3. Measurement Update + +The state is updated within measurement spaces using observed data $\mathbf{z}$ . To accomplish this, we define a measurement model $h(\mathbf{x})$ that maps the current state $\mathbf{x}$ to measurement spaces as: + +$$ +h (\boldsymbol {x}) = \left[ \begin{array}{l l l} \| \boldsymbol {p} ^ {x y} \| _ {2} & \| \boldsymbol {v} ^ {x y} \| _ {2} & \boldsymbol {p} ^ {x y} \quad \boldsymbol {v} ^ {x y} \end{array} \right] ^ {\intercal}. \tag {13} +$$ + +The UWB sensor measures the distance, $d^{xy}$ , between node $x$ and $y$ , which corresponds to the distance derived from relative positions in state, $\| \pmb{p}^{xy}\|_2$ . Additionally, from consecutive distance measurements, we derive the relative velocity norm as $(d_k^{xy} - d_{k-1}^{xy}) / \Delta t$ , corresponding to $\| \pmb{v}\|_2^{xy}$ . The covariance matrix for distance measurements at time step $k$ , denoted as $\mathbf{R}_{1,k}$ , is dynamically set based on line-of-sight conditions for each node pair, informed by $\hat{\theta}_k$ and UWB sensor characteristics. The covariance matrix for relative velocity measurements, $\mathbf{R}_{2,k}$ , is computed as: + +$$ +\mathbf {R} _ {2, k} = \frac {\left(\mathbf {R} _ {1 , k} + \mathbf {R} _ {1 , k - 1}\right)}{\Delta t ^ {2}}. \tag {14} +$$ + +The pose estimator outputs the joint rotations, $\hat{\pmb{\theta}}$ , along with their corresponding standard deviations, $\hat{\Sigma}$ . We employ the unscented transform [11] to transform the pose distribution to the relative position distribution. Given $\Theta \in \mathbb{R}^{96} \sim \mathcal{N}(\hat{\pmb{\theta}}, \hat{\Sigma})$ , we use Van der Merwe's scaled sigma point algorithm [30] to generate $n$ sigma points, $\mathcal{X} \in \mathbb{R}^{n \times 96}$ , along with weights $\boldsymbol{W}^m \in \mathbb{R}^n$ for the mean and $\boldsymbol{W}^c \in \mathbb{R}^n$ for the covariance. Each sigma point in $\mathcal{X}$ is transformed through the SMPL model (1), yielding a set of sensor-relative positions denoted as: + +$$ +\mathcal {Y} = \{\pmb {p} _ {i} ^ {x y} \in \mathbb {R} ^ {n \times 1 5 \times 3} \mid i = 1, 2, \dots , n \}. \qquad (1 5) +$$ + +We then calculate the mean and covariance of the transformed relative positions, capturing the distribution of relative positions based on pose uncertainty: + +$$ +\hat {\boldsymbol {p}} ^ {x y} = \sum_ {i = 1} ^ {n} \boldsymbol {W} _ {i} ^ {m} \mathcal {Y} _ {i}, \tag {16} +$$ + +$$ +\hat {\boldsymbol {\Sigma}} _ {\hat {\boldsymbol {p}}} ^ {2} = \sum_ {i = 1} ^ {n} \boldsymbol {W} _ {i} ^ {c} \left(\mathcal {Y} _ {i} - \hat {\boldsymbol {p}} ^ {x y}\right) \left(\mathcal {Y} _ {i} - \hat {\boldsymbol {p}} ^ {x y}\right) ^ {\intercal}. \tag {17} +$$ + +The transformed relative positions, $\hat{\pmb{p}}^{xy}$ , along with the measurement covariance matrix $\mathbf{R}_3 = \hat{\Sigma}_{\hat{\pmb{p}}}^2$ , are used to update $\pmb{p}^{xy}$ in the state $\pmb{x}$ . Similarly, we compute the relative velocity between consecutive time steps as $(\hat{\pmb{p}}_k^{xy} - \hat{\pmb{p}}_{k-1}^{xy}) / \Delta t$ , with its covariance matrix defined as $\mathbf{R}_{4,k} = (\mathbf{R}_{3,k} + \mathbf{R}_{3,k-1}) / \Delta t^2$ , to update $\pmb{v}^{xy}$ . The complete measurement vector $\pmb{z}$ used for updating the state is given by: + +$$ +\boldsymbol {z} = \left[ d ^ {x y} \quad \frac {\left(d _ {k} ^ {x y} - d _ {k - 1} ^ {x y}\right)}{\Delta t} \quad \hat {\boldsymbol {p}} ^ {x y} \quad \frac {\left(\hat {\boldsymbol {p}} _ {k} ^ {x y} - \hat {\boldsymbol {p}} _ {k - 1} ^ {x y}\right)}{\Delta t} \right] ^ {\intercal}. \tag {18} +$$ + +# 4. Experiments + +# 4.1. Datasets + +We use the AMASS [18], TotalCapture [29], DIP-IMU [7], and UIP [1] datasets for training and evaluation. We synthesize inertial and distance data from the AMASS dataset, following the process in previous studies [1, 7, 34]. We select specific vertices on the body mesh to represent sensor mounting positions. Accelerations are computed as finite differences of these vertex positions over time, while orientations are obtained as the global orientations of the corresponding joints. Distance measurements are synthesized by calculating Euclidean distances between selected sensor positions, adjusted for individual body shape. Training the shape estimator requires anthropometric measurements and inter-distances with corresponding shape parameters from a large number of individuals, which is challenging to collect. Therefore, we synthesize these measurements—height $H$ , weight $W$ , and inter-distances $D$ —from the 3D body mesh in a T-pose, using 479 unique body shapes (273 male and 206 female) available in the AMASS dataset. We estimate $H$ and $W$ following the methods in SHAPY [2], and we sample $D$ from the synthesized distances. + +The training data for the pose estimator comprises synthesized data from the AMASS dataset and the training set of DIP-IMU. Unlike previous studies [10, 34, 35], we exclude the TotalCapture within AMASS from the training data to prevent overlap, ensuring that synthesized distances used in testing are unseen during training. The test data includes TotalCapture, the test split of DIP-IMU, and UIP. + +# 4.2. Method Implementation + +Our shape estimator was implemented using the official multi-label predictor in AutoGluon 1.1 with a medium-quality preset. Our pose estimator was implemented based on PyTorch 2.3. The training process used an Adam optimizer with a learning rate of 0.0001. The pose estimator model was trained for 350 epochs with a batch size of 512 on a single NVIDIA GeForce RTX 3090 GPU, completing in approximately 35 minutes. For initial 20 epochs, we adopted $\mathcal{L}_{MSE}$ to optimize the model, then switched to $\mathcal{L}_{GNLL}$ for the remaining epochs. For the UKF, we experimentally set parameters $\alpha_{UKF} = 0.2$ , $\beta_{UKF} = 1.0$ and $\kappa_{UKF} = -105$ to control the distribution and weighting of sample points. In the unscented transform applied to pose estimates, we used $\alpha_{NN} = 0.09$ , $\beta_{NN} = 1.0$ , and $\kappa_{NN} = -93$ . We applied a factor of 10 to $\mathbf{R}_3$ to compensate for overconfident predictions. The method achieves $60\mathrm{Hz}$ without online LOS inference and $30\mathrm{Hz}$ with it. + +# 4.3. Quantitative Evaluation + +We evaluate our method against two categories of methods: (1) IMU-only methods, including DIP [7], TransPose [34], + +TIP [10], PIP [35], and PNP [37], and (2) distance augmented methods, including TIP-D, PIP-D, and UIP [1]. TIP-D and PIP-D are modified versions of TIP and PIP with distance-augmented input data, and we use their results as reported by UIP. We follow the evaluation protocol in UIP by adding ideal synthetic inter-sensor distances without noise to TotalCapture and DIP-IMU to assess the impact of distance measurements independent of noise. For evaluations on the UIP dataset, which includes both IMU and UWB data, we perform RANSAC regression to calibrate UWB measurements, as described in UIP [1], and then use calibrated measurements as input for the state estimator. We do not apply outlier filtering and use default state estimator parameters due to the absence of ground-truth data and limited information on sensor-specific characteristics in the hardware setup. + +Metrics We use the following metrics for quantitative evaluation: 1) $SIP$ Error (in degrees): the mean global angular error of the upper arms and upper legs, focusing on joints not observed by the body-worn sensors; 2) Angular Error (in degrees): the mean global angular error across all body joints; 3) Positional Error (in centimeters): the mean global joint position error across all joints; and 4) Mesh Error (in centimeters): the mean vertex position error between the reconstructed mesh and the ground-truth mesh given the mean body shape. + +Comparison with IMU-only Methods Table 1 presents a comparison of our method against IMU-only methods on the TotalCapture and DIP-IMU datasets. On the TotalCapture dataset, our approach consistently outperforms previous methods across all metrics, improving over the SOTA PNP by $32.4\%$ in angular error and $9.4\%$ in mesh error. A similar trend is also observed in the results on DIP-IMU dataset. This demonstrates that inter-sensor distances serve as strong constraints for refining estimated poses. + +Comparison with Distance-augmented Methods Table 2 presents a fair comparison of our method with distance-augmented methods, as all approaches use the same input data. Additionally, all methods are trained on the AMASS dataset with the TotalCapture subset excluded, whereas the IMU-only methods in Table 1 include TotalCapture in their training data. Our method outperforms UIP, achieving a reduction in positional error of $21.5\%$ on the TotalCapture dataset and $35.0\%$ on the DIP-IMU dataset. On the UIP dataset, where sensor data quality is low and motions are more ambiguous than in the other two test datasets, our method achieves the lowest positional error, while UIP achieves the lowest SIP error. We achieve comparable results despite using less-filtered distance measurements, a simpler architecture, and non-optimal default parameters for the state estimator. This demonstrates the robustness of our approach in handling real-world sensor noise and the effectiveness of our online state estimation framework. + +
MethodSIP Error (deg)Ang Error (deg)Pos Error (cm)Mesh Error (cm)SIP Error (deg)Ang Error (deg)Pos Error (cm)Mesh Error (cm)
TotalCaptureDIP-IMU
DIP [7]18.6217.229.4211.2217.3515.367.599.05
TransPose [34]16.5812.896.557.4217.068.866.037.17
TIP [10]13.2212.305.816.8016.909.075.636.62
PIP [35]12.9312.045.616.5115.338.785.126.02
PNP [37]10.8910.454.745.4513.718.754.975.77
UMotion10.767.064.464.9414.196.353.383.93
+ +Table 1. Comparison with state of the art IMU-only methods on TotalCapture [29] and DIP-IMU [7]. + +
MethodSIP Error (deg)Pos Error (cm)SIP Error (deg)Pos Error (cm)SIP Error (deg)Pos Error (cm)
TotalCaptureDIP-IMUUIP
PIP [35]15.937.0515.986.2130.4713.62
TIP-D12.185.5115.915.2630.3413.96
PIP-D13.166.3113.795.3630.3313.27
UIP [1]11.325.4913.215.0524.1210.65
UMotion10.764.4614.193.3825.6910.33
+ +Table 2. Comparison with distance-augmented methods on TotapCapture [29], DIP-IMU [7], and UIP [1] datasets. + +
MetricMean shapePredictedGT
Pos Error (cm)5.154.344.31
Mesh Error (cm)5.634.814.78
H Error (cm)5.750.39-
W Error (kg)9.450.34-
D Error (cm)3.632.332.30
C Error (cm)3.781.26-
+ +Table 3. Comparison of reconstructed body mesh errors on Total-Capture [29] using mean shape versus predicted shape parameters. + +
MethodSIP Error (deg)Pos Error (cm)SIP Error (deg)Pos Error (cm)
TotalCaptureDIP-IMU
J →hat9.004.3614.213.29
S →J →hat12.155.7017.354.45
PG →hat9.253.8513.632.85
S →PG →hat11.134.9615.913.91
PR →hat9.894.4213.773.22
S →PR →hat11.575.4415.744.17
Ours: S →hat10.764.4614.213.38
+ +Table 4. Ablation study on pose estimator architecture. + +# 4.4. Module Evaluations + +Shape Estimator Table 3 compares reconstructed body mesh on the TotalCapture dataset using the mean and predicted shape, both with the same estimated poses. With the predicted shape, positional and mesh errors are nearly as low as those obtained using the ground truth shape parameters. Height and weight estimates are accurate, with errors within $1\mathrm{cm}$ and $1\mathrm{kg}$ , respectively. However, the mean error in circumferences $\pmb{C}$ of the chest, waist, hips, wrists, knees, and head remains large, as relevant measurements cannot be inferred solely from inter-sensor distances. Despite this limitation, the predicted shape still reconstructs a more realistic body than the mean shape model. + +Pose Estimator We conduct an ablation study by expanding our pose estimator with intermediate layers to estimate all joint positions, $J$ , sensor global positions, $P_{G}$ , and + +sensor relative positions, $\pmb{P}_R$ , from sensor measurements $S = (\hat{\pmb{a}}, R, \hat{\pmb{d}})$ . These intermediate estimates are then combined with $S$ to compute the final poses $\hat{\pmb{\theta}}$ . This layered structure follows the framework commonly used in previous methods [1, 34]. As shown in Table 4, while regressing $\hat{\pmb{\theta}}$ from ideal $J$ , $P_G$ , or $P_R$ improves accuracy, introducing an intermediate layer, that is, $S \rightarrow \{J, P_G, P_R\}$ , increases errors. This error propagates through the network, ultimately degrading pose estimation performance. This suggests that integrating IMU data with distance constraints may support a simpler architecture, reducing complexity while still maintaining accuracy. + +State Estimator The state estimator is designed to: 1) mitigate errors in inter-sensor distances, 2) filter and stabi + +![](images/4f32f9d8ff28a532371bead867e00f2bd9ccec0340428cb24f58c31aeaab7f7d.jpg) +Figure 4. Cumulative distribution of distance error (left) and acceleration error reduction over time (right) for various fusion settings. + +![](images/57c0e9de5321df014c23e55ce2d008ef905b3cc2cdb9dc01018859df3d467c67.jpg) + +![](images/d5210017e2471484867f7fcabae5e64f03ff1ae2f8767ec024f0311535b07567.jpg) +Figure 5. Joint positional error for different fusion settings. + +imize accelerations, and 3) improve the accuracy of estimated poses. To examine it we conduct an ablation study by varying fusion measurements on the TotalCapture [29] with synthesized LOS-related noisy distances. Fig. 4 (left) shows cumulative distributions of absolute distance error for different fusion settings. The raw distances show an average error of $9.20\mathrm{cm}$ with a standard deviation of $8.94\mathrm{cm}$ . By fusing IMU, UWB, and pose data, this error is reduced to $2.42\mathrm{cm}$ with a standard deviation of $2.39\mathrm{cm}$ , demonstrating the effectiveness of data fusion for mitigating distance errors. Fig. 4 (right) shows the reduction in acceleration errors. The state estimator converges within a few seconds and adapts gradually, aligning with IMU characteristics. Fig. 5 presents joint positional errors over time for different fusion setups. Without filtering, positional error fluctuates with distance measurement quality. Fusing IMU with either UWB or pose data alone offers limited improvement because distance and pose outliers remain. The combined fusion of IMU, UWB, and pose data achieves the lowest positional error, validating that our state estimator refines pose accuracy by utilizing all available measurements. + +# 4.5. Qualitative Evaluation + +We developed a prototype (see supplementary material for details) that integrates BNO086 IMU and DW3000 UWB sensors to demonstrate our proposed method. Fig. 6 presents visualizations comparing pose estimates for challenging motions, illustrating the improvements of our method in disambiguating poses over IMU-only methods + +![](images/4d0ceada2480686881f32963cffdc5c2f9c6b8c3bd46cda28050e7836f55c546.jpg) +Figure 6. Qualitative comparison of pose estimates from the test data collected using our developed prototype. + +and enhanced pose stability compared to UIP. + +# 5. Conclusion and Limitations + +In this work, we present UMotion, an uncertainty-driven framework for 3D human shape and pose estimation that integrates six IMU and UWB sensors within a state estimation system. UMotion incorporates uncertainties from IMU, UWB, and pose estimates under individual body constraints to iteratively enhance confidence in both sensor measurements and pose accuracy. Our experiments on synthetic and real-world datasets demonstrate that UMotion outperforms existing SOTA methods in pose accuracy, and effectively stabilizes sensor measurements and pose estimation. + +However, our shape estimator requires adaptation to specific sensor configurations and conditions, and, with limited training data, it may struggle with unique body variations. The state estimator also requires careful parameter tuning to reflect sensor characteristics accurately. Additionally, our pose estimator simplifies constraints, which may limit its performance. As future work, adding a physics-aware module could potentially enhance our method's robustness. + +# References + +[1] Rayan Armani, Changlin Qian, Jiaxi Jiang, and Christian Holz. Ultra inertial poser: Scalable motion capture and tracking from sparse inertial sensors and ultra-wideband ranging. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1, 2, 6, 7 +[2] Vasileios Choutas, Lea Müller, Chun-Hao P Huang, Siyu Tang, Dimitrios Tzionas, and Michael J Black. Accurate 3d body shape regression using metric and semantic attributes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2718-2728, 2022. 3, 6 +[3] Peng Dai, Yang Zhang, Tao Liu, Zhen Fan, Tianyuan Du, Zhuo Su, Xiaozheng Zheng, and Zeming Li. Hmd-poser: On-device real-time human motion tracking from scalable sparse observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 874-884, 2024. 1 +[4] Nathan DeVrio, Vimal Mollyn, and Chris Harrison. Smartposer: Arm pose estimation with a smartphone and smartwatch using uwb and imu data. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, pages 1-11, 2023. 2 +[5] Nick Erickson, Jonas Mueller, Alexander Shirkov, Hang Zhang, Pedro Larroy, Mu Li, and Alexander Smola. Autogluon-tabular: Robust and accurate automl for structured data. arXiv preprint arXiv:2003.06505, 2020. 3 +[6] S Hochreiter. Long short-term memory. Neural Computation MIT-Press, 1997. 4 +[7] Yinghao Huang, Manuel Kaufmann, Emre Aksan, Michael J Black, Otmar Hilliges, and Gerard Pons-Moll. Deep inertial poser: Learning to reconstruct human pose from sparse inertial measurements in real time. ACM Transactions on Graphics (TOG), 37(6):1-15, 2018. 1, 2, 6, 7 +[8] Jiaxi Jiang, Paul Streli, Huajian Qiu, Andreas Fender, Larissa Laich, Patrick Snape, and Christian Holz. Avatarposer: Articulated full-body pose tracking from sparse motion sensing. In European conference on computer vision, pages 443-460. Springer, 2022. 1 +[9] Jiaxi Jiang, Paul Streli, Manuel Meier, and Christian Holz. Egoposer: Robust real-time ego-body pose estimation in large scenes. arXiv preprint arXiv:2308.06493, 2023. 2 +[10] Yifeng Jiang, Yuting Ye, Deepak Gopinath, Jungdam Won, Alexander W Winkler, and C Karen Liu. Transformer inertial poser: Real-time human motion reconstruction from sparse imus with simultaneous terrain generation. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 1, 2, 6, 7 +[11] Simon J Julier, Jeffrey K Uhlmann, and Hugh F Durrant-Whyte. A new approach for filtering nonlinear systems. In Proceedings of 1995 American Control Conference-ACC'95, pages 1628-1632. IEEE, 1995. 5 +[12] Manuel Kaufmann, Yi Zhao, Chengcheng Tang, Lingling Tao, Christopher Twigg, Jie Song, Robert Wang, and Otmar Hilliges. Em-posed: 3d human pose estimation from sparse electromagnetic trackers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11510-11520, 2021. 2 + +[13] Jiye Lee and Hanbyul Joo. Mocap everyone everywhere: Lightweight motion capture with smartwatches and a head-mounted camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1091-1100, 2024. 1, 2 +[14] Sunmin Lee, Sebastian Starke, Yuting Ye, Jungdam Won, and Alexander Winkler. Questenvsim: Environment-aware simulated motion tracking from sparse sensors. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-9, 2023. 1 +[15] Han Liang, Yannan He, Chengfeng Zhao, Mutian Li, Jingya Wang, Jingyi Yu, and Lan Xu. Hybridcap: Inertia-aid monocular capture of challenging human motions. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1539–1548, 2023. 2 +[16] Huajun Liu, Xiaolin Wei, Jinxiang Chai, Inwoo Ha, and Taehyun Rhee. Realtime human motion control with a small number of inertial sensors. In Symposium on interactive 3D graphics and games, pages 133-140, 2011. 2 +[17] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. In *Seminal Graphics Papers: Pushing the Boundaries*, Volume 2, pages 851–866. 2023. 2 +[18] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5442-5451, 2019. 2, 6 +[19] Vimal Mollyn, Riku Arakawa, Mayank Goel, Chris Harrison, and Karan Ahuja. Imuposer: Full-body pose estimation using imus in phones, watches, and earbuds. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, pages 1-12, 2023. 1 +[20] Ian Oppermann, Matti Hämäläinen, and Jari Inatti. *UWB: theory and applications*. John Wiley & Sons, 2004. 3 +[21] Shaohua Pan, Qi Ma, Xinyu Yi, Weifeng Hu, Xiong Wang, Xingkang Zhou, Jijunnan Li, and Feng Xu. Fusing monocular images and sparse imu signals for real-time human motion capture. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 2 +[22] Gerard Pons-Moll, Andreas Baak, Thomas Helten, Meinard Müller, Hans-Peter Seidel, and Bodo Rosenhahn. Multisensor-fusion for 3d full-body human motion capture. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 663-670. IEEE, 2010. +[23] Jose Luis Ponton, Haoran Yun, Andreas Aristidou, Carlos Andujar, and Nuria Pelechano. Sparseposer: Real-time full-body motion reconstruction from sparse data. ACM Transactions on Graphics, 43(1):1-14, 2023. 2 +[24] Sergi Pujades, Betty Mohler, Anne Thaler, Joachim Tesch, Naureen Mahmood, Nikolas Hesse, Heinrich H Bülthoff, and Michael J Black. The virtual caliper: rapid creation of metrically accurate avatars from 3d measurements. IEEE transactions on visualization and computer graphics, 25(5): 1887-1897, 2019. 3 +[25] Yiming Ren, Chengfeng Zhao, Yannan He, Peishan Cong, Han Liang, Jingyi Yu, Lan Xu, and Yuexin Ma. Lidar-aid + +inertial poser: Large-scale human motion capture by sparse inertial and lidar sensors. IEEE Transactions on Visualization and Computer Graphics, 29(5):2337-2347, 2023. 2 +[26] Xu Ru, Nian Gu, Hang Shang, and Heng Zhang. Mems inertial sensor calibration technology: Current status and future trends. Micromachines, 13(6):879, 2022. 3 +[27] Martin Schepers, Matteo Giuberti, Giovanni Bellusci, et al. Xsens mvn: Consistent tracking of human motion using inertial sensing. Xsens Technol, 1(8):1-8, 2018. 1 +[28] Mike Schuster and Kuldip K Paliwal. Bidirectional recurrent neural networks. IEEE transactions on Signal Processing, 45 (11):2673-2681, 1997. 2 +[29] Matthew Trumble, Andrew Gilbert, Charles Malleson, Adrian Hilton, and John P Collomosse. Total capture: 3d human pose estimation fusing video and inertial sensors. In BMVC, pages 1-13. London, UK, 2017. 6, 7, 8 +[30] Rudolph Van Der Merwe. Sigma-point Kalman filters for probabilistic inference in dynamic state-space models. Oregon Health & Science University, 2004. 5 +[31] Tom Van Wouwe, Seunghwan Lee, Antoine Falisse, Scott Delp, and C Karen Liu. Diffusion inertial poser: Human motion reconstruction from arbitrary sparse imu configurations. arXiv preprint arXiv:2308.16682, 2023. 1 +[32] Timo Von Marcard, Bodo Rosenhahn, Michael J Black, and Gerard Pons-Moll. Sparse inertial poser: Automatic 3d human pose estimation from sparse imus. In Computer graphics forum, pages 349-360. Wiley Online Library, 2017. 1, 2 +[33] Eric A Wan and Rudolph Van Der Merwe. The unscented kalman filter for nonlinear estimation. In Proceedings of the IEEE 2000 adaptive systems for signal processing, communications, and control symposium (Cat. No. 00EX373), pages 153-158. Ieee, 2000. 4 +[34] Xinyu Yi, Yuxiao Zhou, and Feng Xu. Transpose: Real-time 3d human translation and pose estimation with six inertial sensors. ACM Transactions On Graphics (TOG), 40(4):1-13, 2021. 1, 2, 6, 7 +[35] Xinyu Yi, Yuxiao Zhou, Marc Habermann, Soshi Shimada, Vladislav Golyanik, Christian Theobalt, and Feng Xu. Physical inertial poser (pip): Physics-aware real-time human motion tracking from sparse inertial sensors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13167-13178, 2022. 1, 2, 4, 6, 7 +[36] Xinyu Yi, Yuxiao Zhou, Marc Habermann, Vladislav Golyanik, Shaohua Pan, Christian Theobalt, and Feng Xu. Ecolocate: Real-time motion capture, localization, and mapping with sparse body-mounted sensors. ACM Transactions on Graphics (TOG), 42(4):1-17, 2023. 2 +[37] Xinyu Yi, Yuxiao Zhou, and Feng Xu. Physical non-inertial poser (pnp): Modeling non-inertial effects in sparse-inertial human motion capture. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1, 2, 6, 7 +[38] Faheem Zafari, Athanasios Gkelias, and Kin K Leung. A survey of indoor localization systems and technologies. IEEE Communications Surveys & Tutorials, 21(3):2568-2599, 2019. 3 + +[39] Yu Zhang, Songpengcheng Xia, Lei Chu, Jiarui Yang, Qi Wu, and Ling Pei. Dynamic inertial poser (dynaip): Part-based motion dynamics learning for enhanced human pose estimation with sparse inertial sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1889-1899, 2024. 1, 2 +[40] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5745-5753, 2019. 4 +[41] Chengxu Zuo, Yiming Wang, Lishuang Zhan, Shihui Guo, Xinyu Yi, Feng Xu, and Yipeng Qin. Loose inertial poser: Motion capture with imu-attached loose-wear jacket. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2209–2219, 2024. 1 \ No newline at end of file diff --git a/2025/UMotion_ Uncertainty-driven Human Motion Estimation from Inertial and Ultra-wideband Units/images.zip b/2025/UMotion_ Uncertainty-driven Human Motion Estimation from Inertial and Ultra-wideband Units/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f79544a14ad5f43a6196cb28f8e9195832f2729b --- /dev/null +++ b/2025/UMotion_ Uncertainty-driven Human Motion Estimation from Inertial and Ultra-wideband Units/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3eb8a955d5cc8cd84c5e454b23be320314a10e822a4eb1cf64c9b215950ec220 +size 465086 diff --git a/2025/UMotion_ Uncertainty-driven Human Motion Estimation from Inertial and Ultra-wideband Units/layout.json b/2025/UMotion_ Uncertainty-driven Human Motion Estimation from Inertial and Ultra-wideband Units/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9e86f8a03f88a1b981fa194b4f464eea125dbf06 --- /dev/null +++ b/2025/UMotion_ Uncertainty-driven Human Motion Estimation from Inertial and Ultra-wideband Units/layout.json @@ -0,0 +1,9117 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 112, + 103, + 501, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 103, + 501, + 138 + ], + "spans": [ + { + "bbox": [ + 112, + 103, + 501, + 138 + ], + "type": "text", + "content": "UMotion: Uncertainty-driven Human Motion Estimation from Inertial and Ultra-wideband Units" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 162, + 161, + 448, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 161, + 448, + 175 + ], + "spans": [ + { + "bbox": [ + 162, + 161, + 448, + 175 + ], + "type": "text", + "content": "Huakun Liu Hiroki Ota Xin Wei Yutaro Hirao" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 118, + 175, + 492, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 175, + 492, + 204 + ], + "spans": [ + { + "bbox": [ + 118, + 175, + 492, + 204 + ], + "type": "text", + "content": "Monica Perusquía-Hernández \nHideaki Uchiyama \nKiyoshi Kiyokawa \nNara Institute of Science and Technology, Japan" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 205, + 497, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 205, + 497, + 219 + ], + "spans": [ + { + "bbox": [ + 113, + 205, + 497, + 219 + ], + "type": "text", + "content": "{liu.huakun.li0, ota.hiroki.oc6, wei.xin.wy0, yutaro.hirao}@is.naist.jp" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 172, + 220, + 436, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 220, + 436, + 232 + ], + "spans": [ + { + "bbox": [ + 172, + 220, + 436, + 232 + ], + "type": "text", + "content": "{m.perusquia, hideaki.uchiyama, kiyo}@is.naist.jp" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 285, + 297, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 297, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 297, + 574 + ], + "type": "text", + "content": "Sparse wearable inertial measurement units (IMUs) have gained popularity for estimating 3D human motion. However, challenges such as pose ambiguity, data drift, and limited adaptability to diverse bodies persist. To address these issues, we propose UMotion, an uncertainty-driven, online fusing-all state estimation framework for 3D human shape and pose estimation, supported by six integrated, body-worn ultra-wideband (UWB) distance sensors with IMUs. UWB sensors measure inter-node distances to infer spatial relationships, aiding in resolving pose ambiguities and body shape variations when combined with anthropometric data. Unfortunately, IMUs are prone to drift, and UWB sensors are affected by body occlusions. Consequently, we develop a tightly coupled Unscented Kalman Filter (UKF) framework that fuses uncertainties from sensor data and estimated human motion based on individual body shape. The UKF iteratively refines IMU and UWB measurements by aligning them with uncertain human motion constraints in real-time, producing optimal estimates for each. Experiments on both synthetic and real-world datasets demonstrate the effectiveness of UMotion in stabilizing sensor data and the improvement over state of the art in pose accuracy. Code is available at: https://github.com/kk9six/umotion." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 597, + 137, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 137, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 137, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "content": "Estimating 3D human motion from wearable sensors has become increasingly popular due to their portability, accessibility, and versatility. Wearable sensors, such as inertial measurement units (IMUs), enable continuous monitoring of body motion measurements across unrestricted spaces. These advances shift motion capture from controlled laboratory environments to everyday settings [13, 19, 32, 41]. This transition benefits fields such as healthcare, sports" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 347, + 260, + 529, + 361 + ], + "blocks": [ + { + "bbox": [ + 347, + 260, + 529, + 361 + ], + "lines": [ + { + "bbox": [ + 347, + 260, + 529, + 361 + ], + "spans": [ + { + "bbox": [ + 347, + 260, + 529, + 361 + ], + "type": "image", + "image_path": "09b2e00879f5d00521d89acf4a37ac88a9833eb37f7d362bf7e3cd28695e2e25.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 371, + 555, + 428 + ], + "lines": [ + { + "bbox": [ + 313, + 371, + 555, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 555, + 428 + ], + "type": "text", + "content": "Figure 1. UMotion integrates IMU-UWB data inputs and pose outputs uniformly under uncertainty, constrained by individual body structure. The online state estimation framework iteratively refines sensor data confidence and stabilizes pose estimation, reducing ambiguities and improving robustness." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 449, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 472 + ], + "type": "text", + "content": "performance, ergonomics, and emerging areas in humancomputer interaction [3, 8, 14, 31]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 473, + 556, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 473, + 556, + 676 + ], + "spans": [ + { + "bbox": [ + 312, + 473, + 556, + 676 + ], + "type": "text", + "content": "One of the widely chosen wearable sensors for 3D human motion estimation is IMU. Commercial systems, such as Xsens [27], utilize 17 or more IMUs for comprehensive pose coverage. While highly accurate, these densely placed IMUs are inconvenient and intrusive. Recent studies have reduced the required number of IMUs to just six—placed on the forearms, lower legs, pelvis, and head—while still achieving promising performance through data-driven methods [1, 7, 10, 31, 32, 34, 35, 37, 39]. With fewer IMUs, pose estimation becomes under-constrained and prone to ambiguity. Recent work has attempted to disambiguate poses by incorporating temporal consistency, physics-based constraints, or additional, easy-to-integrate sensors [1, 10, 35]. However, these methods still face significant challenges, including noisy sensor data, body shape variations, and ambiguities arising from under-constrained setups with sparse sensors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "In this work, we propose UMotion, an uncertainty-driven human motion estimation framework that combines online state estimation with an integrated system of six inertial and" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7085" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 262 + ], + "type": "text", + "content": "ultra-wideband (UWB) distance sensors. Prior work often infers poses from sensor data without considering the reverse influence. UMotion uniformly treats input and output under uncertainty and body-specific constraints, blending data to optimize estimates (see Fig. 1). Our approach maintains inter-sensor distances as a core system state, capturing spatial relationships between sensor nodes. These distances, combined with basic anthropometric measurements (height and weight) and IMU data, serve as inputs for our learning-based shape and pose estimators. Since sensors are inherently noisy, we propose a tightly coupled online state estimation system that associates IMU and UWB sensor measurements and pose estimates using an Unscented Kalman Filter (UKF) and an uncertainty propagation method, iteratively correcting errors and stabilizing pose estimation in real-time. Our main contributions are:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 264, + 293, + 369 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 55, + 264, + 293, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 293, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 293, + 297 + ], + "type": "text", + "content": "- An ensemble learning-based human shape estimation approach using distance constraints and anthropometric measurements from six integrated IMU-UWB sensors." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 300, + 293, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 300, + 293, + 332 + ], + "spans": [ + { + "bbox": [ + 56, + 300, + 293, + 332 + ], + "type": "text", + "content": "- A learning-based method for human pose estimation from inertial and distance constraints, incorporating mesh distribution estimation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 335, + 293, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 335, + 293, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 335, + 293, + 369 + ], + "type": "text", + "content": "- A filtering-based state estimation system that couples sensor measurements, pose estimates, and body variations in real-time to improve sensor stability and pose accuracy." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 382, + 141, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 382, + 141, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 141, + 395 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 403, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 295, + 569 + ], + "type": "text", + "content": "Pose Estimation from Sparse Inertial Sensors In contrast to the predominant use of vision-based methods, wearable sensors, specifically IMUs, offer greater freedom and flexibility. Von Marcard et al. [32] present SIP that makes inertial pose estimation practical by using only six IMUs attached to the body, combined with an offline iterative SMPL body model [17] pose optimization. Huang et al. [7] propose a real-time pose estimation method, DIP, which uses a bidirectional recurrent neural network (biRNN) [28] to learn the mapping from a sequence of six IMU measurements to SMPL body poses. To obtain sufficient training data, they synthesize IMU measurements from the AMASS dataset [18], further advancing the development of learning-based methods for sparse sensor motion capture." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "Following previous studies, TransPose [34] refines pose estimation by decomposing the end-to-end framework into a multi-stage process with intermediate joint position estimation. To disambiguate the poses, PIP [35] proposes a physics-based motion optimization, while TIP [10] incorporates a conditional Transformer model for plausible terrain generation. Training data is crucial for learning-based methods. Unlike prior studies relying on synthetic IMU data, DynalIP [39] adapts real IMU data from diverse human skeleton formats to the target SMPL model and shows superior performance when training with real sensor data. Similarly focusing on the data, PNP [37] recently addresses" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "content": "limitations in existing IMU synthesis by incorporating noninertial effects and fictitious forces, enhancing estimation robustness through a physics-informed neural network and realistic IMU synthesis." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 121, + 554, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 554, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 554, + 252 + ], + "type": "text", + "content": "In contrast to methods focusing solely on pose estimation, our approach treats human shape and pose as equally important. By integrating shape information, we establish a tight connection between sensor data and estimated motions, forming a positive feedback loop that enhances the entire process. Additionally, most previous studies fail to fully utilize IMU accelerations due to high noise and drift issues. Within our framework, accelerations serve as control inputs for state estimation, undergoing continuous correction and acting as a crucial component in tracking spatial relationships among sensors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 269, + 555, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 555, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 555, + 558 + ], + "type": "text", + "content": "Pose Estimation from Hybrid Sensors In addition to IMUs, various wearable and hybrid sensor systems have been explored to overcome inherent IMU limitations such as restricted positional awareness [9, 12, 13, 15, 16, 21–23, 25, 36]. The closest works to ours are SmartPoser [4] and Ultra Inertial Poser (UIP) [1], both of which integrate UWB and IMU sensors for pose estimation. UWB sensors complement IMUs with additional distance information between sensors while preserving the portability and flexibility of tracking devices. SmartPoser [4] focuses on wearable arm pose estimation, combining UWB measurements with IMU data using an off-the-shelf smartwatch and smartphone. UIP [1] integrates six UWB sensors with IMUs for full-body pose estimation. Inter-sensor distances, estimated using an EKF that fuses IMU and UWB data, help reduce global translation drift and minimize position jitter compared to inertial-only tracking. However, body occlusion frequently affects certain node pairs, e.g., head-knee, causing the EKF to fail in estimating distances due to rapid IMU drift and unreliable UWB measurements, resulting in unstable distance data. Our approach addresses this limitation by using output pose uncertainties and body-specific constraints as observations to refine input sensor measurements, improving the stability of both IMU and UWB data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 569, + 370, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 569, + 370, + 581 + ], + "spans": [ + { + "bbox": [ + 314, + 569, + 370, + 581 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 589, + 400, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 589, + 400, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 589, + 400, + 600 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 606, + 553, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 553, + 690 + ], + "type": "text", + "content": "SMPL body model We use the SMPL model [17] to represent human motion. SMPL decomposes the human body into pose parameters, " + }, + { + "bbox": [ + 313, + 606, + 553, + 690 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta} \\in \\mathbb{R}^{23 \\times 3 + 3}" + }, + { + "bbox": [ + 313, + 606, + 553, + 690 + ], + "type": "text", + "content": ", which define relative rotations of 23 joints and the global root joint orientation in axis-angle form, and shape parameters, " + }, + { + "bbox": [ + 313, + 606, + 553, + 690 + ], + "type": "inline_equation", + "content": "\\beta \\in \\mathbb{R}^{10}" + }, + { + "bbox": [ + 313, + 606, + 553, + 690 + ], + "type": "text", + "content": ", capturing body shape variations across individuals. The model is defined by a linear blend-skinning (LBS) function:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 700, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 700, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 321, + 700, + 555, + 715 + ], + "type": "interline_equation", + "content": "M (\\boldsymbol {\\beta}, \\boldsymbol {\\theta}) = W \\left(\\bar {\\mathbf {T}} + B _ {s} (\\boldsymbol {\\beta}) + B _ {p} (\\boldsymbol {\\theta}), J (\\boldsymbol {\\beta}), \\boldsymbol {\\theta}, \\mathcal {W}\\right), \\tag {1}", + "image_path": "c76494f61480ae1b0a6cc42944462c22dc00e17826e41e4e9e736ab4696a1c42.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7086" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "inline_equation", + "content": "B_{s}(\\boldsymbol {\\beta})" + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "inline_equation", + "content": "B_{p}(\\boldsymbol {\\theta})" + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": " are shape and pose blend shapes that deform the template mesh " + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{T}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": " in the zero pose. The mesh is then reposed according to specified poses " + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": ", combining with joint positions " + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "inline_equation", + "content": "J(\\beta)" + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": " and blend-skinning weights " + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": ". Inertial Measurement Unit (IMU) A nine-axis MEMS IMU comprises an accelerometer, gyroscope, and magnetometer, measuring accelerations in the sensor local frame " + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "inline_equation", + "content": "F^S" + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": ", angular velocities in " + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "inline_equation", + "content": "F^S" + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": ", and magnetic field strengths relative to the Earth's magnetic field, respectively. A commonly used error model for measured accelerations of one IMU, " + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "inline_equation", + "content": "\\pmb{a}\\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": ", is given by [26]:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 130, + 213, + 295, + 227 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 213, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 130, + 213, + 295, + 227 + ], + "type": "interline_equation", + "content": "\\boldsymbol {a} = \\boldsymbol {a} _ {\\text {t r u e}} + \\boldsymbol {b} _ {a} + \\boldsymbol {n} _ {a}, \\tag {2}", + "image_path": "2a969d3e18e3aa46538e5e375cec45cf40001185fb5823bd6de81320243efa69.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "inline_equation", + "content": "\\pmb{a}_{\\mathrm{true}}" + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "text", + "content": " is the true acceleration, " + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "inline_equation", + "content": "\\pmb{b}_a" + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "text", + "content": " is the bias, modeled as a random walk process with noise " + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "text", + "content": " following a zero-mean Gaussian distribution with standard deviation " + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "inline_equation", + "content": "\\Sigma_{b}" + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "inline_equation", + "content": "\\pmb{n}_a" + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "text", + "content": " is high-frequency white noise. In our work, after a frame calibration process, accelerations are represented as " + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "inline_equation", + "content": "\\pmb{a}^{M}" + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "text", + "content": " in the SMPL body-centric frame " + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "inline_equation", + "content": "F^{M}" + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "text", + "content": ", and orientations are represented as " + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "inline_equation", + "content": "\\pmb{R}^{MB}" + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "text", + "content": ", indicating the rotation from bone frame " + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "inline_equation", + "content": "F^{B}" + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "inline_equation", + "content": "F^{M}" + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "text", + "content": ". For simplicity, we denote " + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "inline_equation", + "content": "\\pmb{a}" + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "inline_equation", + "content": "\\pmb{a}^{M}" + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "inline_equation", + "content": "\\pmb{R}" + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "inline_equation", + "content": "\\pmb{R}^{MB}" + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "text", + "content": " in the following sections. Details on the calibration process are available in supplementary materials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 295, + 498 + ], + "type": "text", + "content": "Ultra-wideband (UWB) UWB uses the time-of-flight technique to measure distances between two devices, with one acting as the transmitter and the other as the receiver. Unlike other RF techniques, such as WiFi and Bluetooth, UWB operates over a wide frequency range and transmits short pulses, achieving centimeter-level distance measurements with minimal interference [20]. Consequently, it is widely used in high-precision indoor localization systems with multiple fixed anchors and movable tags [38]. In our work, all six integrated UWB sensors are movable and alternately function as both transmitters and receivers, capturing inter-distances between each sensor pair." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 505, + 168, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 168, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 168, + 517 + ], + "type": "text", + "content": "3.2. Framework Design" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": "As shown in Fig. 2, our method uses six IMU-UWB sensors and consists of three main modules: a shape estimator, a pose estimator, and a state estimator. The shape estimator takes anthropometric measurements and inter-distances in a T-pose as input, outputting shape parameters " + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\beta}" + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": " of the SMPL model. The pose estimator uses filtered frame-aligned, root-normalized IMU measurements and inter-distances as input to estimate pose parameters, " + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}" + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": ", and corresponding uncertainties, " + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\Sigma}" + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": ". Given our sensor placement, we do not observe the movement of the hands and feet. Therefore, similar to previous studies, we estimate poses for only 16 joints, including the root joint, while excluding the hand and foot joints. For simplicity, we denote this reduced set of pose parameters as " + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\theta \\in \\mathbb{R}^{16\\times 3}" + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\theta^{24}" + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": " refers to the full set of pose parameters for 24 joints. Notations with a hat symbol, e.g., " + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}" + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": ", indicate the corresponding estimated values. Addi" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "tionally, we do not estimate global translation, as inferring it from only body-worn IMU and UWB sensors without external references would lead to unbounded error accumulation. The state estimator integrates accelerations, UWB measurements, and the estimated poses with uncertainties to track the sensor relative positions, velocities, and acceleration biases of each node at each time step. These refined estimates are then used to refine the inputs to the pose estimator." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 169, + 555, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 169, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 555, + 313 + ], + "type": "text", + "content": "The estimation framework is driven by inherent uncertainties present in the system. Specifically, noisy IMU data combined with tracked inter-sensor distances serve as the primary inputs for pose estimation. The generated poses, along with their associated uncertainties that reflect sensor noise, are further constrained by the human model, which in turn refines sensor measurements and distance estimates. This feedback loop propagates uncertainties through the system, continuously updating the belief in system states based on the reliability of each data source—IMU, UWB, and pose estimator with body constraints, ultimately producing robust and accurate estimation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 326, + 460, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 460, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 460, + 339 + ], + "type": "text", + "content": "3.3. Shape and Pose Estimator" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 345, + 414, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 414, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 414, + 357 + ], + "type": "text", + "content": "3.3.1. Shape Estimator" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 361, + 554, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 554, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 554, + 445 + ], + "type": "text", + "content": "Spatial inter-distances provide data that approximates the rough skeletal structure of an individual. Previous studies, such as Virtual Caliper [24] and SHAPY [2], have demonstrated a linear relationship between body shape and body measurements. Building on these findings, we adopt an ensemble learning-based shape estimator that uses experimentally selected anthropometric measurements, defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 391, + 457, + 555, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 457, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 391, + 457, + 555, + 472 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {\\beta}} = S E (H, W, \\boldsymbol {D}). \\tag {3}", + "image_path": "99e877d7986d81949a1f5b2f59ecc5cf875d6e8458e0122593e91f3919cca72f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 484, + 554, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 554, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 554, + 556 + ], + "type": "text", + "content": "Basic body measurements, height " + }, + { + "bbox": [ + 313, + 484, + 554, + 556 + ], + "type": "inline_equation", + "content": "H \\in \\mathbb{R}^1" + }, + { + "bbox": [ + 313, + 484, + 554, + 556 + ], + "type": "text", + "content": " and weight " + }, + { + "bbox": [ + 313, + 484, + 554, + 556 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}^1" + }, + { + "bbox": [ + 313, + 484, + 554, + 556 + ], + "type": "text", + "content": ", are easily accessible and provide foundational information about human body proportions. Experimentally selected inter-distances between sensors, " + }, + { + "bbox": [ + 313, + 484, + 554, + 556 + ], + "type": "inline_equation", + "content": "D \\in \\mathbb{R}^7" + }, + { + "bbox": [ + 313, + 484, + 554, + 556 + ], + "type": "text", + "content": ", capture relative limb lengths and body structural details that height and weight alone cannot provide." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 558, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 712 + ], + "type": "text", + "content": "As shown in Fig. 3, we place virtual sensors on the body mesh and conduct line-of-sight (LOS) simulation experiments with real-world tests to identify and select 7 repeatable distances out of 15 possible options. While the selection may vary based on device characteristics, the chosen distances offer a reliable foundation for adaptation. For model training, we use AutoGluon [5], an AutoML framework that automatically trains and ensembles 11 basic machine learning models, which is more stable than a single linear regression method in our validation. The estimated shape parameters " + }, + { + "bbox": [ + 313, + 558, + 555, + 712 + ], + "type": "inline_equation", + "content": "\\hat{\\beta}" + }, + { + "bbox": [ + 313, + 558, + 555, + 712 + ], + "type": "text", + "content": " are then used to reconstruct realistic human bodies and to propagate estimated poses to spatial constraints in the state estimator." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7087" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 72, + 527, + 220 + ], + "blocks": [ + { + "bbox": [ + 82, + 72, + 527, + 220 + ], + "lines": [ + { + "bbox": [ + 82, + 72, + 527, + 220 + ], + "spans": [ + { + "bbox": [ + 82, + 72, + 527, + 220 + ], + "type": "image", + "image_path": "0d3ccd4fc14c3d62ceb39f841417717db676e66dcd35a8234acc914a351f6748.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 230, + 555, + 288 + ], + "lines": [ + { + "bbox": [ + 54, + 230, + 555, + 288 + ], + "spans": [ + { + "bbox": [ + 54, + 230, + 555, + 288 + ], + "type": "text", + "content": "Figure 2. Overview of UMotion, consisting of three main modules: the shape estimator, pose estimator, and state estimator. The shape estimator takes anthropometric measurements and inter-distances in a T-pose as input, outputting shape parameters that reconstruct a realistic body and impose strong constraints on the system. The pose estimator receives filtered IMU data and inter-distances from the state estimator to predict poses, which are fed back to refine state estimates. The entire system integrates IMU, UWB, and estimated poses within the context of individual body structure to continuously update and improve motion estimation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 78, + 305, + 279, + 467 + ], + "blocks": [ + { + "bbox": [ + 78, + 305, + 279, + 467 + ], + "lines": [ + { + "bbox": [ + 78, + 305, + 279, + 467 + ], + "spans": [ + { + "bbox": [ + 78, + 305, + 279, + 467 + ], + "type": "image", + "image_path": "c15a1643ba69f95646107fa37e9d6d87a58a1a336efea4167f94def6aa070a21.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 476, + 297, + 533 + ], + "lines": [ + { + "bbox": [ + 55, + 476, + 297, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 476, + 297, + 533 + ], + "type": "text", + "content": "Figure 3. Visualization of selected inter-sensor distances used in the shape estimator. We place virtual sensors on the body mesh at sensor mounting points and conduct line-of-sight simulation experiments. The plots display the temporal changes in UWB measurements between various sensor pairs over time." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 554, + 149, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 149, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 149, + 566 + ], + "type": "text", + "content": "3.3.2. Pose Estimator" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": "Our pose estimator is based on unidirectional multi-layer long short-term memory (LSTM) [6] recurrent neural network (RNN). The unidirectional LSTM preserves only past information, achieving real-time estimation without requiring future data. Given six IMU-UWB sensors, we use filtered accelerations, " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{a} \\in \\mathbb{R}^{18}" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": ", and orientations in 9D rotation matrix form, " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{R} \\in \\mathbb{R}^{54}" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": ", along with 15 inter-distances " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{d}} \\in \\mathbb{R}^{15}" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": ", as inputs. The estimator then outputs the poses " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}_{6D} \\in \\mathbb{R}^{96}" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": " for the target joints, represented in a 6D rotation matrix form [40]. Additionally, in the final layer, the network outputs the logarithm of the estimation uncertainty, " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\Sigma} \\in \\mathbb{R}^{96}" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": ", capturing prediction confidence. Consequently," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 307, + 446, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 307, + 446, + 319 + ], + "spans": [ + { + "bbox": [ + 314, + 307, + 446, + 319 + ], + "type": "text", + "content": "the pose estimator is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 382, + 327, + 555, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 327, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 382, + 327, + 555, + 342 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {\\theta}} _ {6 D}, \\hat {\\boldsymbol {\\Sigma}} = P E (\\hat {\\boldsymbol {a}}, \\boldsymbol {R}, \\hat {\\boldsymbol {d}}). \\tag {4}", + "image_path": "58cd1f2d0a98ba8d424ca89826d31b398fbb4a1f127edf1fb85af61ac852fdf9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 350, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 437 + ], + "type": "text", + "content": "To handle ambiguity in the initial frame, we adopt a learning-based initialization strategy inspired by PIP [35]. A fully-connected convolutional neural network regresses the initial hidden state of the RNN based on the initial global poses, " + }, + { + "bbox": [ + 313, + 350, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{\\theta}}_0^{24}" + }, + { + "bbox": [ + 313, + 350, + 555, + 437 + ], + "type": "text", + "content": ", and joint positions, " + }, + { + "bbox": [ + 313, + 350, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\hat{J}_0 \\in \\mathbb{R}^{72}" + }, + { + "bbox": [ + 313, + 350, + 555, + 437 + ], + "type": "text", + "content": ". This approach provides the pose estimator with a well-informed initialization, adapting to different bodies." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 437, + 555, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 473 + ], + "type": "text", + "content": "We use two loss functions to train the pose estimator: Mean Squared Error (MSE) and Gaussian Negative Log Likelihood (GNLL) loss. MSE is defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 345, + 483, + 555, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 483, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 345, + 483, + 555, + 513 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M S E}} \\left(\\boldsymbol {\\theta} _ {6 D}, \\hat {\\boldsymbol {\\theta}} _ {6 D}\\right) = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left\\| \\boldsymbol {\\theta} _ {6 D} - \\hat {\\boldsymbol {\\theta}} _ {6 D} \\right\\| _ {2} ^ {2}, \\tag {5}", + "image_path": "eb0dfbf839fc0ee588089f3914debce39ce80ba5debe9205745d8c93ca21188a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 521, + 555, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 547 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 521, + 555, + 547 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_{6D}" + }, + { + "bbox": [ + 313, + 521, + 555, + 547 + ], + "type": "text", + "content": " denotes the ground truth poses and " + }, + { + "bbox": [ + 313, + 521, + 555, + 547 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 521, + 555, + 547 + ], + "type": "text", + "content": " is the total number of frames. The GNLL, " + }, + { + "bbox": [ + 313, + 521, + 555, + 547 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{GNLL}}(\\pmb{\\theta}_{6D},\\hat{\\pmb{\\theta}}_{6D},\\hat{\\pmb{\\Sigma}})" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 322, + 555, + 555, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 555, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 322, + 555, + 555, + 594 + ], + "type": "interline_equation", + "content": "\\frac {1}{2} \\left(\\log \\left(\\max \\left(\\hat {\\boldsymbol {\\Sigma}} ^ {2}, \\varepsilon_ {\\min }\\right)\\right) + \\frac {(\\hat {\\boldsymbol {\\theta}} _ {6 D} - \\boldsymbol {\\theta} _ {6 D}) ^ {2}}{\\max \\left(\\hat {\\boldsymbol {\\Sigma}} ^ {2} , \\varepsilon_ {\\min }\\right)}\\right), \\tag {6}", + "image_path": "bfbb5a4701c0ea76ad7e12b8157b42f49faf83ed2577aa98c182c8c21f3477e4.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 604, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 555, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 604, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\varepsilon_{\\mathrm{min}}" + }, + { + "bbox": [ + 313, + 604, + 555, + 653 + ], + "type": "text", + "content": " , set to " + }, + { + "bbox": [ + 313, + 604, + 555, + 653 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 313, + 604, + 555, + 653 + ], + "type": "text", + "content": " in our implementation, is used for numerical stability. First, " + }, + { + "bbox": [ + 313, + 604, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MSE}}" + }, + { + "bbox": [ + 313, + 604, + 555, + 653 + ], + "type": "text", + "content": " helps the model converge quickly during early training. Next, " + }, + { + "bbox": [ + 313, + 604, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{GNLL}}" + }, + { + "bbox": [ + 313, + 604, + 555, + 653 + ], + "type": "text", + "content": " is applied to optimize for pose and uncertainty estimation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 659, + 411, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 411, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 411, + 672 + ], + "type": "text", + "content": "3.4. State Estimator" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "We use UKF [33] for state estimation to refine the pose estimator inputs, integrating UWB measurements and pose estimator outputs with a statistically derived IMU model." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7088" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "The UWB measurements provide direct inter-distance data, while they are affected by body occlusion issues. Meanwhile, pose estimator outputs offer constraints and inform distance measurements; however, they suffer from occasional inaccuracies due to challenging poses or sensor noise. The UKF combines these sources, balancing the strengths and limitations of each to make the optimal estimates." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 164, + 151, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 164, + 151, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 164, + 151, + 175 + ], + "type": "text", + "content": "3.4.1. State Definition" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 178, + 268, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 178, + 268, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 268, + 192 + ], + "type": "text", + "content": "We define the state vector " + }, + { + "bbox": [ + 55, + 178, + 268, + 192 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{R}^{15 \\times 3 + 15 \\times 3 + 6 \\times 3}" + }, + { + "bbox": [ + 55, + 178, + 268, + 192 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 199, + 295, + 227 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 199, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 58, + 199, + 295, + 227 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} = \\left[ \\begin{array}{l l l l l l l l} \\boldsymbol {p} ^ {1 2} & \\dots & \\boldsymbol {p} ^ {5 6} & \\boldsymbol {v} ^ {1 2} & \\dots & \\boldsymbol {v} ^ {5 6} & \\boldsymbol {b} ^ {1} & \\dots & \\boldsymbol {b} ^ {6} \\end{array} \\right] ^ {\\mathsf {T}}, \\tag {7}", + "image_path": "8ea09df4b2838eb2dc78dcbeb1c879ae407660d1e2945785f7d360f03d2f70c7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "inline_equation", + "content": "\\pmb{p}^{xy}" + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "text", + "content": " represents the relative position between nodes " + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "inline_equation", + "content": "x, y \\in \\{1, 2, 3, 4, 5, 6\\}" + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "inline_equation", + "content": "x < y" + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "text", + "content": ". The term " + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "inline_equation", + "content": "\\pmb{v}^{xy}" + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "text", + "content": " denotes the relative velocity between nodes " + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "inline_equation", + "content": "\\pmb{b}^1" + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "inline_equation", + "content": "\\pmb{b}^6" + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "text", + "content": " represent residual acceleration errors after transformation from " + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "inline_equation", + "content": "F^S" + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "inline_equation", + "content": "F^M" + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "text", + "content": ". These biases primarily result from orientation errors and raw accelerometer biases " + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "inline_equation", + "content": "\\pmb{b}_a" + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "text", + "content": ". All quantities are expressed in the SMPL body-centric coordinate frame " + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "inline_equation", + "content": "F^M" + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "text", + "content": ". For initialization, we set " + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "inline_equation", + "content": "\\pmb{p}^{xy}" + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "text", + "content": " based on a static T-pose defined by the shape parameters " + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "text", + "content": " and the SMPL model (1). The relative velocities " + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "inline_equation", + "content": "\\pmb{v}^{xy}" + }, + { + "bbox": [ + 55, + 228, + 296, + 358 + ], + "type": "text", + "content": " and biases are initialized to zero." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 367, + 161, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 161, + 379 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 161, + 379 + ], + "type": "text", + "content": "3.4.2. State Propagation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 383, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 296, + 407 + ], + "type": "text", + "content": "The control input " + }, + { + "bbox": [ + 55, + 383, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\pmb{u}" + }, + { + "bbox": [ + 55, + 383, + 296, + 407 + ], + "type": "text", + "content": " consists of the accelerations " + }, + { + "bbox": [ + 55, + 383, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\pmb{a}" + }, + { + "bbox": [ + 55, + 383, + 296, + 407 + ], + "type": "text", + "content": " of six nodes, defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 108, + 416, + 295, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 416, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 108, + 416, + 295, + 434 + ], + "type": "interline_equation", + "content": "\\boldsymbol {u} = \\left[ \\begin{array}{l l l} \\boldsymbol {a} ^ {1} & \\dots & \\boldsymbol {a} ^ {6} \\end{array} \\right] ^ {\\mathsf {T}}, \\boldsymbol {u} \\in \\mathbb {R} ^ {1 8}. \\tag {8}", + "image_path": "8bb877a100f0fc9aa159f912b06c57992c926d825590d89bf0f0b1d90542ed56.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 443, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 296, + 491 + ], + "type": "text", + "content": "During the prediction step, the current state " + }, + { + "bbox": [ + 55, + 443, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 443, + 296, + 491 + ], + "type": "text", + "content": " is propagated using the control input " + }, + { + "bbox": [ + 55, + 443, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\pmb{u}" + }, + { + "bbox": [ + 55, + 443, + 296, + 491 + ], + "type": "text", + "content": ". The propagation model is derived from the strapdown inertial kinematic model with an acceleration error model (2), defined as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 501, + 295, + 530 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 501, + 295, + 530 + ], + "spans": [ + { + "bbox": [ + 61, + 501, + 295, + 530 + ], + "type": "interline_equation", + "content": "\\boldsymbol {v} _ {k} ^ {x y} = \\boldsymbol {v} _ {k - 1} ^ {x y} + \\left(\\boldsymbol {a} _ {k - 1} ^ {y} - \\boldsymbol {a} _ {k - 1} ^ {x}\\right) \\Delta t - \\left(\\boldsymbol {b} _ {k - 1} ^ {y} - \\boldsymbol {b} _ {k - 1} ^ {x}\\right) \\Delta t, \\tag {9}", + "image_path": "1c456d95519b1efb43763be04ea940327c37109ea7ed756626975c3e5c3430a8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 534, + 295, + 580 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 534, + 295, + 580 + ], + "spans": [ + { + "bbox": [ + 61, + 534, + 295, + 580 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {p} _ {k} ^ {x y} = \\boldsymbol {p} _ {k - 1} ^ {x y} + \\boldsymbol {v} _ {k - 1} ^ {x y} \\Delta t + \\frac {1}{2} (\\boldsymbol {a} _ {k - 1} ^ {y} - \\boldsymbol {a} _ {k - 1} ^ {x}) \\Delta t ^ {2} \\\\ + \\frac {1}{2} \\left(\\boldsymbol {b} _ {k - 1} ^ {x} - \\boldsymbol {b} _ {k - 1} ^ {y}\\right) \\Delta t ^ {2}, \\tag {10} \\\\ \\end{array}", + "image_path": "b6d86bf4715444c5a97d58af35da36e863af92940e9f5b5bc9525ed73aaef1ea.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 582, + 295, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 582, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 67, + 582, + 295, + 597 + ], + "type": "interline_equation", + "content": "\\boldsymbol {b} _ {k} ^ {x} = \\boldsymbol {b} _ {k - 1} ^ {x} + \\boldsymbol {\\eta} ^ {x}, \\boldsymbol {\\eta} ^ {x} \\sim \\mathcal {N} \\left(0, \\Sigma_ {b} ^ {2}\\right), \\tag {11}", + "image_path": "f8c4ef09df2f0e628e105cb6544c25ff8f007437ca9ee45c7e9fa49e1aafe3be.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 607, + 296, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 607, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 607, + 296, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 607, + 296, + 654 + ], + "type": "inline_equation", + "content": "k - 1" + }, + { + "bbox": [ + 55, + 607, + 296, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 607, + 296, + 654 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 607, + 296, + 654 + ], + "type": "text", + "content": " denote consecutive time steps, and " + }, + { + "bbox": [ + 55, + 607, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 55, + 607, + 296, + 654 + ], + "type": "text", + "content": " represents the interval between them. Together, equations (9), (10), and (11) form a state propagation model with added white noise, represented as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 123, + 666, + 295, + 679 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 666, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 123, + 666, + 295, + 679 + ], + "type": "interline_equation", + "content": "\\bar {\\boldsymbol {x}} _ {k} = f (\\boldsymbol {x} _ {k - 1}, \\boldsymbol {u} _ {k}) + \\mathbf {Q}, \\tag {12}", + "image_path": "07436ab73fa690696398831212e4ccf2d18c711b734d51ce0f3a3243524c8788.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is the process noise covariance matrix, derived from the characteristics of the IMUs used in the system." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 72, + 435, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 435, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 435, + 84 + ], + "type": "text", + "content": "3.4.3. Measurement Update" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 87, + 555, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 87, + 555, + 135 + ], + "spans": [ + { + "bbox": [ + 313, + 87, + 555, + 135 + ], + "type": "text", + "content": "The state is updated within measurement spaces using observed data " + }, + { + "bbox": [ + 313, + 87, + 555, + 135 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 313, + 87, + 555, + 135 + ], + "type": "text", + "content": ". To accomplish this, we define a measurement model " + }, + { + "bbox": [ + 313, + 87, + 555, + 135 + ], + "type": "inline_equation", + "content": "h(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 87, + 555, + 135 + ], + "type": "text", + "content": " that maps the current state " + }, + { + "bbox": [ + 313, + 87, + 555, + 135 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 87, + 555, + 135 + ], + "type": "text", + "content": " to measurement spaces as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 337, + 141, + 555, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 141, + 555, + 159 + ], + "spans": [ + { + "bbox": [ + 337, + 141, + 555, + 159 + ], + "type": "interline_equation", + "content": "h (\\boldsymbol {x}) = \\left[ \\begin{array}{l l l} \\| \\boldsymbol {p} ^ {x y} \\| _ {2} & \\| \\boldsymbol {v} ^ {x y} \\| _ {2} & \\boldsymbol {p} ^ {x y} \\quad \\boldsymbol {v} ^ {x y} \\end{array} \\right] ^ {\\intercal}. \\tag {13}", + "image_path": "99dc9c5f1e95dbadf1d463d6275b5e3aa46764cafd955b13cf49fbc6071b6528.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 166, + 555, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 166, + 555, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 166, + 555, + 287 + ], + "type": "text", + "content": "The UWB sensor measures the distance, " + }, + { + "bbox": [ + 313, + 166, + 555, + 287 + ], + "type": "inline_equation", + "content": "d^{xy}" + }, + { + "bbox": [ + 313, + 166, + 555, + 287 + ], + "type": "text", + "content": ", between node " + }, + { + "bbox": [ + 313, + 166, + 555, + 287 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 166, + 555, + 287 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 166, + 555, + 287 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 166, + 555, + 287 + ], + "type": "text", + "content": ", which corresponds to the distance derived from relative positions in state, " + }, + { + "bbox": [ + 313, + 166, + 555, + 287 + ], + "type": "inline_equation", + "content": "\\| \\pmb{p}^{xy}\\|_2" + }, + { + "bbox": [ + 313, + 166, + 555, + 287 + ], + "type": "text", + "content": ". Additionally, from consecutive distance measurements, we derive the relative velocity norm as " + }, + { + "bbox": [ + 313, + 166, + 555, + 287 + ], + "type": "inline_equation", + "content": "(d_k^{xy} - d_{k-1}^{xy}) / \\Delta t" + }, + { + "bbox": [ + 313, + 166, + 555, + 287 + ], + "type": "text", + "content": ", corresponding to " + }, + { + "bbox": [ + 313, + 166, + 555, + 287 + ], + "type": "inline_equation", + "content": "\\| \\pmb{v}\\|_2^{xy}" + }, + { + "bbox": [ + 313, + 166, + 555, + 287 + ], + "type": "text", + "content": ". The covariance matrix for distance measurements at time step " + }, + { + "bbox": [ + 313, + 166, + 555, + 287 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 166, + 555, + 287 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 313, + 166, + 555, + 287 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{1,k}" + }, + { + "bbox": [ + 313, + 166, + 555, + 287 + ], + "type": "text", + "content": ", is dynamically set based on line-of-sight conditions for each node pair, informed by " + }, + { + "bbox": [ + 313, + 166, + 555, + 287 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}_k" + }, + { + "bbox": [ + 313, + 166, + 555, + 287 + ], + "type": "text", + "content": " and UWB sensor characteristics. The covariance matrix for relative velocity measurements, " + }, + { + "bbox": [ + 313, + 166, + 555, + 287 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{2,k}" + }, + { + "bbox": [ + 313, + 166, + 555, + 287 + ], + "type": "text", + "content": ", is computed as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 378, + 294, + 554, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 294, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 378, + 294, + 554, + 319 + ], + "type": "interline_equation", + "content": "\\mathbf {R} _ {2, k} = \\frac {\\left(\\mathbf {R} _ {1 , k} + \\mathbf {R} _ {1 , k - 1}\\right)}{\\Delta t ^ {2}}. \\tag {14}", + "image_path": "d854536fd775801e51d97b533dd2a88fe6bfd3c1726abdc0cf6074e8ead2807a.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 327, + 555, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 447 + ], + "type": "text", + "content": "The pose estimator outputs the joint rotations, " + }, + { + "bbox": [ + 313, + 327, + 555, + 447 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{\\theta}}" + }, + { + "bbox": [ + 313, + 327, + 555, + 447 + ], + "type": "text", + "content": ", along with their corresponding standard deviations, " + }, + { + "bbox": [ + 313, + 327, + 555, + 447 + ], + "type": "inline_equation", + "content": "\\hat{\\Sigma}" + }, + { + "bbox": [ + 313, + 327, + 555, + 447 + ], + "type": "text", + "content": ". We employ the unscented transform [11] to transform the pose distribution to the relative position distribution. Given " + }, + { + "bbox": [ + 313, + 327, + 555, + 447 + ], + "type": "inline_equation", + "content": "\\Theta \\in \\mathbb{R}^{96} \\sim \\mathcal{N}(\\hat{\\pmb{\\theta}}, \\hat{\\Sigma})" + }, + { + "bbox": [ + 313, + 327, + 555, + 447 + ], + "type": "text", + "content": ", we use Van der Merwe's scaled sigma point algorithm [30] to generate " + }, + { + "bbox": [ + 313, + 327, + 555, + 447 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 327, + 555, + 447 + ], + "type": "text", + "content": " sigma points, " + }, + { + "bbox": [ + 313, + 327, + 555, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{X} \\in \\mathbb{R}^{n \\times 96}" + }, + { + "bbox": [ + 313, + 327, + 555, + 447 + ], + "type": "text", + "content": ", along with weights " + }, + { + "bbox": [ + 313, + 327, + 555, + 447 + ], + "type": "inline_equation", + "content": "\\boldsymbol{W}^m \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 313, + 327, + 555, + 447 + ], + "type": "text", + "content": " for the mean and " + }, + { + "bbox": [ + 313, + 327, + 555, + 447 + ], + "type": "inline_equation", + "content": "\\boldsymbol{W}^c \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 313, + 327, + 555, + 447 + ], + "type": "text", + "content": " for the covariance. Each sigma point in " + }, + { + "bbox": [ + 313, + 327, + 555, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 327, + 555, + 447 + ], + "type": "text", + "content": " is transformed through the SMPL model (1), yielding a set of sensor-relative positions denoted as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 349, + 453, + 555, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 453, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 349, + 453, + 555, + 469 + ], + "type": "interline_equation", + "content": "\\mathcal {Y} = \\{\\pmb {p} _ {i} ^ {x y} \\in \\mathbb {R} ^ {n \\times 1 5 \\times 3} \\mid i = 1, 2, \\dots , n \\}. \\qquad (1 5)", + "image_path": "9fcf060b086accdc1b48e60922b2c74453d9df31096b3408d8915c13ff7a4d1d.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 475, + 554, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 554, + 512 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 554, + 512 + ], + "type": "text", + "content": "We then calculate the mean and covariance of the transformed relative positions, capturing the distribution of relative positions based on pose uncertainty:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 351, + 518, + 555, + 549 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 518, + 555, + 549 + ], + "spans": [ + { + "bbox": [ + 351, + 518, + 555, + 549 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {p}} ^ {x y} = \\sum_ {i = 1} ^ {n} \\boldsymbol {W} _ {i} ^ {m} \\mathcal {Y} _ {i}, \\tag {16}", + "image_path": "2e82c0850365b18e349fa40edebab38a7af806978e3f4bef2c6e00c57523e634.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 354, + 552, + 554, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 552, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 354, + 552, + 554, + 582 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {\\Sigma}} _ {\\hat {\\boldsymbol {p}}} ^ {2} = \\sum_ {i = 1} ^ {n} \\boldsymbol {W} _ {i} ^ {c} \\left(\\mathcal {Y} _ {i} - \\hat {\\boldsymbol {p}} ^ {x y}\\right) \\left(\\mathcal {Y} _ {i} - \\hat {\\boldsymbol {p}} ^ {x y}\\right) ^ {\\intercal}. \\tag {17}", + "image_path": "501fc05f7b61a837d1d61924475c31111eeea5c63f80795b6bb27c68d4617667.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 590, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 555, + 677 + ], + "type": "text", + "content": "The transformed relative positions, " + }, + { + "bbox": [ + 313, + 590, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{p}}^{xy}" + }, + { + "bbox": [ + 313, + 590, + 555, + 677 + ], + "type": "text", + "content": ", along with the measurement covariance matrix " + }, + { + "bbox": [ + 313, + 590, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_3 = \\hat{\\Sigma}_{\\hat{\\pmb{p}}}^2" + }, + { + "bbox": [ + 313, + 590, + 555, + 677 + ], + "type": "text", + "content": ", are used to update " + }, + { + "bbox": [ + 313, + 590, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\pmb{p}^{xy}" + }, + { + "bbox": [ + 313, + 590, + 555, + 677 + ], + "type": "text", + "content": " in the state " + }, + { + "bbox": [ + 313, + 590, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 590, + 555, + 677 + ], + "type": "text", + "content": ". Similarly, we compute the relative velocity between consecutive time steps as " + }, + { + "bbox": [ + 313, + 590, + 555, + 677 + ], + "type": "inline_equation", + "content": "(\\hat{\\pmb{p}}_k^{xy} - \\hat{\\pmb{p}}_{k-1}^{xy}) / \\Delta t" + }, + { + "bbox": [ + 313, + 590, + 555, + 677 + ], + "type": "text", + "content": ", with its covariance matrix defined as " + }, + { + "bbox": [ + 313, + 590, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{4,k} = (\\mathbf{R}_{3,k} + \\mathbf{R}_{3,k-1}) / \\Delta t^2" + }, + { + "bbox": [ + 313, + 590, + 555, + 677 + ], + "type": "text", + "content": ", to update " + }, + { + "bbox": [ + 313, + 590, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\pmb{v}^{xy}" + }, + { + "bbox": [ + 313, + 590, + 555, + 677 + ], + "type": "text", + "content": ". The complete measurement vector " + }, + { + "bbox": [ + 313, + 590, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\pmb{z}" + }, + { + "bbox": [ + 313, + 590, + 555, + 677 + ], + "type": "text", + "content": " used for updating the state is given by:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 329, + 693, + 555, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 693, + 555, + 717 + ], + "spans": [ + { + "bbox": [ + 329, + 693, + 555, + 717 + ], + "type": "interline_equation", + "content": "\\boldsymbol {z} = \\left[ d ^ {x y} \\quad \\frac {\\left(d _ {k} ^ {x y} - d _ {k - 1} ^ {x y}\\right)}{\\Delta t} \\quad \\hat {\\boldsymbol {p}} ^ {x y} \\quad \\frac {\\left(\\hat {\\boldsymbol {p}} _ {k} ^ {x y} - \\hat {\\boldsymbol {p}} _ {k - 1} ^ {x y}\\right)}{\\Delta t} \\right] ^ {\\intercal}. \\tag {18}", + "image_path": "234cfd28b1a5030f958bd9800b5c01e0f459ee2fd8a38083f027e68418559be5.jpg" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7089" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 136, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 136, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 136, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 118, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 118, + 102 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 118, + 102 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 109, + 297, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 109, + 297, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 109, + 297, + 348 + ], + "type": "text", + "content": "We use the AMASS [18], TotalCapture [29], DIP-IMU [7], and UIP [1] datasets for training and evaluation. We synthesize inertial and distance data from the AMASS dataset, following the process in previous studies [1, 7, 34]. We select specific vertices on the body mesh to represent sensor mounting positions. Accelerations are computed as finite differences of these vertex positions over time, while orientations are obtained as the global orientations of the corresponding joints. Distance measurements are synthesized by calculating Euclidean distances between selected sensor positions, adjusted for individual body shape. Training the shape estimator requires anthropometric measurements and inter-distances with corresponding shape parameters from a large number of individuals, which is challenging to collect. Therefore, we synthesize these measurements—height " + }, + { + "bbox": [ + 55, + 109, + 297, + 348 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 109, + 297, + 348 + ], + "type": "text", + "content": ", weight " + }, + { + "bbox": [ + 55, + 109, + 297, + 348 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 109, + 297, + 348 + ], + "type": "text", + "content": ", and inter-distances " + }, + { + "bbox": [ + 55, + 109, + 297, + 348 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 109, + 297, + 348 + ], + "type": "text", + "content": " —from the 3D body mesh in a T-pose, using 479 unique body shapes (273 male and 206 female) available in the AMASS dataset. We estimate " + }, + { + "bbox": [ + 55, + 109, + 297, + 348 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 109, + 297, + 348 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 109, + 297, + 348 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 109, + 297, + 348 + ], + "type": "text", + "content": " following the methods in SHAPY [2], and we sample " + }, + { + "bbox": [ + 55, + 109, + 297, + 348 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 109, + 297, + 348 + ], + "type": "text", + "content": " from the synthesized distances." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 349, + 296, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 296, + 433 + ], + "type": "text", + "content": "The training data for the pose estimator comprises synthesized data from the AMASS dataset and the training set of DIP-IMU. Unlike previous studies [10, 34, 35], we exclude the TotalCapture within AMASS from the training data to prevent overlap, ensuring that synthesized distances used in testing are unseen during training. The test data includes TotalCapture, the test split of DIP-IMU, and UIP." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 441, + 192, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 192, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 192, + 453 + ], + "type": "text", + "content": "4.2. Method Implementation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 459, + 296, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 296, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 296, + 662 + ], + "type": "text", + "content": "Our shape estimator was implemented using the official multi-label predictor in AutoGluon 1.1 with a medium-quality preset. Our pose estimator was implemented based on PyTorch 2.3. The training process used an Adam optimizer with a learning rate of 0.0001. The pose estimator model was trained for 350 epochs with a batch size of 512 on a single NVIDIA GeForce RTX 3090 GPU, completing in approximately 35 minutes. For initial 20 epochs, we adopted " + }, + { + "bbox": [ + 55, + 459, + 296, + 662 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MSE}" + }, + { + "bbox": [ + 55, + 459, + 296, + 662 + ], + "type": "text", + "content": " to optimize the model, then switched to " + }, + { + "bbox": [ + 55, + 459, + 296, + 662 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{GNLL}" + }, + { + "bbox": [ + 55, + 459, + 296, + 662 + ], + "type": "text", + "content": " for the remaining epochs. For the UKF, we experimentally set parameters " + }, + { + "bbox": [ + 55, + 459, + 296, + 662 + ], + "type": "inline_equation", + "content": "\\alpha_{UKF} = 0.2" + }, + { + "bbox": [ + 55, + 459, + 296, + 662 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 459, + 296, + 662 + ], + "type": "inline_equation", + "content": "\\beta_{UKF} = 1.0" + }, + { + "bbox": [ + 55, + 459, + 296, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 459, + 296, + 662 + ], + "type": "inline_equation", + "content": "\\kappa_{UKF} = -105" + }, + { + "bbox": [ + 55, + 459, + 296, + 662 + ], + "type": "text", + "content": " to control the distribution and weighting of sample points. In the unscented transform applied to pose estimates, we used " + }, + { + "bbox": [ + 55, + 459, + 296, + 662 + ], + "type": "inline_equation", + "content": "\\alpha_{NN} = 0.09" + }, + { + "bbox": [ + 55, + 459, + 296, + 662 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 459, + 296, + 662 + ], + "type": "inline_equation", + "content": "\\beta_{NN} = 1.0" + }, + { + "bbox": [ + 55, + 459, + 296, + 662 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 459, + 296, + 662 + ], + "type": "inline_equation", + "content": "\\kappa_{NN} = -93" + }, + { + "bbox": [ + 55, + 459, + 296, + 662 + ], + "type": "text", + "content": ". We applied a factor of 10 to " + }, + { + "bbox": [ + 55, + 459, + 296, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_3" + }, + { + "bbox": [ + 55, + 459, + 296, + 662 + ], + "type": "text", + "content": " to compensate for overconfident predictions. The method achieves " + }, + { + "bbox": [ + 55, + 459, + 296, + 662 + ], + "type": "inline_equation", + "content": "60\\mathrm{Hz}" + }, + { + "bbox": [ + 55, + 459, + 296, + 662 + ], + "type": "text", + "content": " without online LOS inference and " + }, + { + "bbox": [ + 55, + 459, + 296, + 662 + ], + "type": "inline_equation", + "content": "30\\mathrm{Hz}" + }, + { + "bbox": [ + 55, + 459, + 296, + 662 + ], + "type": "text", + "content": " with it." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 671, + 191, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 191, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 191, + 684 + ], + "type": "text", + "content": "4.3. Quantitative Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "We evaluate our method against two categories of methods: (1) IMU-only methods, including DIP [7], TransPose [34]," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "content": "TIP [10], PIP [35], and PNP [37], and (2) distance augmented methods, including TIP-D, PIP-D, and UIP [1]. TIP-D and PIP-D are modified versions of TIP and PIP with distance-augmented input data, and we use their results as reported by UIP. We follow the evaluation protocol in UIP by adding ideal synthetic inter-sensor distances without noise to TotalCapture and DIP-IMU to assess the impact of distance measurements independent of noise. For evaluations on the UIP dataset, which includes both IMU and UWB data, we perform RANSAC regression to calibrate UWB measurements, as described in UIP [1], and then use calibrated measurements as input for the state estimator. We do not apply outlier filtering and use default state estimator parameters due to the absence of ground-truth data and limited information on sensor-specific characteristics in the hardware setup." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 266, + 554, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 554, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 554, + 386 + ], + "type": "text", + "content": "Metrics We use the following metrics for quantitative evaluation: 1) " + }, + { + "bbox": [ + 313, + 266, + 554, + 386 + ], + "type": "inline_equation", + "content": "SIP" + }, + { + "bbox": [ + 313, + 266, + 554, + 386 + ], + "type": "text", + "content": " Error (in degrees): the mean global angular error of the upper arms and upper legs, focusing on joints not observed by the body-worn sensors; 2) Angular Error (in degrees): the mean global angular error across all body joints; 3) Positional Error (in centimeters): the mean global joint position error across all joints; and 4) Mesh Error (in centimeters): the mean vertex position error between the reconstructed mesh and the ground-truth mesh given the mean body shape." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 388, + 554, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 496 + ], + "type": "text", + "content": "Comparison with IMU-only Methods Table 1 presents a comparison of our method against IMU-only methods on the TotalCapture and DIP-IMU datasets. On the TotalCapture dataset, our approach consistently outperforms previous methods across all metrics, improving over the SOTA PNP by " + }, + { + "bbox": [ + 313, + 388, + 554, + 496 + ], + "type": "inline_equation", + "content": "32.4\\%" + }, + { + "bbox": [ + 313, + 388, + 554, + 496 + ], + "type": "text", + "content": " in angular error and " + }, + { + "bbox": [ + 313, + 388, + 554, + 496 + ], + "type": "inline_equation", + "content": "9.4\\%" + }, + { + "bbox": [ + 313, + 388, + 554, + 496 + ], + "type": "text", + "content": " in mesh error. A similar trend is also observed in the results on DIP-IMU dataset. This demonstrates that inter-sensor distances serve as strong constraints for refining estimated poses." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 498, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 713 + ], + "type": "text", + "content": "Comparison with Distance-augmented Methods Table 2 presents a fair comparison of our method with distance-augmented methods, as all approaches use the same input data. Additionally, all methods are trained on the AMASS dataset with the TotalCapture subset excluded, whereas the IMU-only methods in Table 1 include TotalCapture in their training data. Our method outperforms UIP, achieving a reduction in positional error of " + }, + { + "bbox": [ + 313, + 498, + 554, + 713 + ], + "type": "inline_equation", + "content": "21.5\\%" + }, + { + "bbox": [ + 313, + 498, + 554, + 713 + ], + "type": "text", + "content": " on the TotalCapture dataset and " + }, + { + "bbox": [ + 313, + 498, + 554, + 713 + ], + "type": "inline_equation", + "content": "35.0\\%" + }, + { + "bbox": [ + 313, + 498, + 554, + 713 + ], + "type": "text", + "content": " on the DIP-IMU dataset. On the UIP dataset, where sensor data quality is low and motions are more ambiguous than in the other two test datasets, our method achieves the lowest positional error, while UIP achieves the lowest SIP error. We achieve comparable results despite using less-filtered distance measurements, a simpler architecture, and non-optimal default parameters for the state estimator. This demonstrates the robustness of our approach in handling real-world sensor noise and the effectiveness of our online state estimation framework." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7090" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 70, + 533, + 186 + ], + "blocks": [ + { + "bbox": [ + 77, + 70, + 533, + 186 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 533, + 186 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 533, + 186 + ], + "type": "table", + "html": "
MethodSIP Error (deg)Ang Error (deg)Pos Error (cm)Mesh Error (cm)SIP Error (deg)Ang Error (deg)Pos Error (cm)Mesh Error (cm)
TotalCaptureDIP-IMU
DIP [7]18.6217.229.4211.2217.3515.367.599.05
TransPose [34]16.5812.896.557.4217.068.866.037.17
TIP [10]13.2212.305.816.8016.909.075.636.62
PIP [35]12.9312.045.616.5115.338.785.126.02
PNP [37]10.8910.454.745.4513.718.754.975.77
UMotion10.767.064.464.9414.196.353.383.93
", + "image_path": "2869f5bb18a876d74bfdb458086b086e3c5aac5e7553c91bb4b4dd9654a6e3e8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 142, + 217, + 468, + 328 + ], + "blocks": [ + { + "bbox": [ + 122, + 195, + 487, + 207 + ], + "lines": [ + { + "bbox": [ + 122, + 195, + 487, + 207 + ], + "spans": [ + { + "bbox": [ + 122, + 195, + 487, + 207 + ], + "type": "text", + "content": "Table 1. Comparison with state of the art IMU-only methods on TotalCapture [29] and DIP-IMU [7]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 142, + 217, + 468, + 328 + ], + "lines": [ + { + "bbox": [ + 142, + 217, + 468, + 328 + ], + "spans": [ + { + "bbox": [ + 142, + 217, + 468, + 328 + ], + "type": "table", + "html": "
MethodSIP Error (deg)Pos Error (cm)SIP Error (deg)Pos Error (cm)SIP Error (deg)Pos Error (cm)
TotalCaptureDIP-IMUUIP
PIP [35]15.937.0515.986.2130.4713.62
TIP-D12.185.5115.915.2630.3413.96
PIP-D13.166.3113.795.3630.3313.27
UIP [1]11.325.4913.215.0524.1210.65
UMotion10.764.4614.193.3825.6910.33
", + "image_path": "7db82d521f749891ccfb16bdfe90fd3d6ea2a8cc1d9e5bbe5275541af02fd600.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 74, + 367, + 277, + 456 + ], + "blocks": [ + { + "bbox": [ + 97, + 337, + 511, + 348 + ], + "lines": [ + { + "bbox": [ + 97, + 337, + 511, + 348 + ], + "spans": [ + { + "bbox": [ + 97, + 337, + 511, + 348 + ], + "type": "text", + "content": "Table 2. Comparison with distance-augmented methods on TotapCapture [29], DIP-IMU [7], and UIP [1] datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 367, + 277, + 456 + ], + "lines": [ + { + "bbox": [ + 74, + 367, + 277, + 456 + ], + "spans": [ + { + "bbox": [ + 74, + 367, + 277, + 456 + ], + "type": "table", + "html": "
MetricMean shapePredictedGT
Pos Error (cm)5.154.344.31
Mesh Error (cm)5.634.814.78
H Error (cm)5.750.39-
W Error (kg)9.450.34-
D Error (cm)3.632.332.30
C Error (cm)3.781.26-
", + "image_path": "e915928d92b3afaa1a6f76435238c9a8a0feed8da4abb7da25570335cd39b2a2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 315, + 367, + 566, + 501 + ], + "blocks": [ + { + "bbox": [ + 55, + 466, + 295, + 488 + ], + "lines": [ + { + "bbox": [ + 55, + 466, + 295, + 488 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 295, + 488 + ], + "type": "text", + "content": "Table 3. Comparison of reconstructed body mesh errors on Total-Capture [29] using mean shape versus predicted shape parameters." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 367, + 566, + 501 + ], + "lines": [ + { + "bbox": [ + 315, + 367, + 566, + 501 + ], + "spans": [ + { + "bbox": [ + 315, + 367, + 566, + 501 + ], + "type": "table", + "html": "
MethodSIP Error (deg)Pos Error (cm)SIP Error (deg)Pos Error (cm)
TotalCaptureDIP-IMU
J →hat9.004.3614.213.29
S →J →hat12.155.7017.354.45
PG →hat9.253.8513.632.85
S →PG →hat11.134.9615.913.91
PR →hat9.894.4213.773.22
S →PR →hat11.575.4415.744.17
Ours: S →hat10.764.4614.213.38
", + "image_path": "ac3663c2d5090a3818893e288cb1ee094b900932e291767e1595b848e26eb642.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 334, + 509, + 533, + 520 + ], + "lines": [ + { + "bbox": [ + 334, + 509, + 533, + 520 + ], + "spans": [ + { + "bbox": [ + 334, + 509, + 533, + 520 + ], + "type": "text", + "content": "Table 4. Ablation study on pose estimator architecture." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 512, + 172, + 524 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 512, + 172, + 524 + ], + "spans": [ + { + "bbox": [ + 55, + 512, + 172, + 524 + ], + "type": "text", + "content": "4.4. Module Evaluations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 532, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 532, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 54, + 532, + 295, + 674 + ], + "type": "text", + "content": "Shape Estimator Table 3 compares reconstructed body mesh on the TotalCapture dataset using the mean and predicted shape, both with the same estimated poses. With the predicted shape, positional and mesh errors are nearly as low as those obtained using the ground truth shape parameters. Height and weight estimates are accurate, with errors within " + }, + { + "bbox": [ + 54, + 532, + 295, + 674 + ], + "type": "inline_equation", + "content": "1\\mathrm{cm}" + }, + { + "bbox": [ + 54, + 532, + 295, + 674 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 532, + 295, + 674 + ], + "type": "inline_equation", + "content": "1\\mathrm{kg}" + }, + { + "bbox": [ + 54, + 532, + 295, + 674 + ], + "type": "text", + "content": ", respectively. However, the mean error in circumferences " + }, + { + "bbox": [ + 54, + 532, + 295, + 674 + ], + "type": "inline_equation", + "content": "\\pmb{C}" + }, + { + "bbox": [ + 54, + 532, + 295, + 674 + ], + "type": "text", + "content": " of the chest, waist, hips, wrists, knees, and head remains large, as relevant measurements cannot be inferred solely from inter-sensor distances. Despite this limitation, the predicted shape still reconstructs a more realistic body than the mean shape model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Pose Estimator We conduct an ablation study by expanding our pose estimator with intermediate layers to estimate all joint positions, " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": ", sensor global positions, " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "P_{G}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": ", and" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 544, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 555, + 688 + ], + "type": "text", + "content": "sensor relative positions, " + }, + { + "bbox": [ + 313, + 544, + 555, + 688 + ], + "type": "inline_equation", + "content": "\\pmb{P}_R" + }, + { + "bbox": [ + 313, + 544, + 555, + 688 + ], + "type": "text", + "content": ", from sensor measurements " + }, + { + "bbox": [ + 313, + 544, + 555, + 688 + ], + "type": "inline_equation", + "content": "S = (\\hat{\\pmb{a}}, R, \\hat{\\pmb{d}})" + }, + { + "bbox": [ + 313, + 544, + 555, + 688 + ], + "type": "text", + "content": ". These intermediate estimates are then combined with " + }, + { + "bbox": [ + 313, + 544, + 555, + 688 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 544, + 555, + 688 + ], + "type": "text", + "content": " to compute the final poses " + }, + { + "bbox": [ + 313, + 544, + 555, + 688 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{\\theta}}" + }, + { + "bbox": [ + 313, + 544, + 555, + 688 + ], + "type": "text", + "content": ". This layered structure follows the framework commonly used in previous methods [1, 34]. As shown in Table 4, while regressing " + }, + { + "bbox": [ + 313, + 544, + 555, + 688 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{\\theta}}" + }, + { + "bbox": [ + 313, + 544, + 555, + 688 + ], + "type": "text", + "content": " from ideal " + }, + { + "bbox": [ + 313, + 544, + 555, + 688 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 313, + 544, + 555, + 688 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 544, + 555, + 688 + ], + "type": "inline_equation", + "content": "P_G" + }, + { + "bbox": [ + 313, + 544, + 555, + 688 + ], + "type": "text", + "content": ", or " + }, + { + "bbox": [ + 313, + 544, + 555, + 688 + ], + "type": "inline_equation", + "content": "P_R" + }, + { + "bbox": [ + 313, + 544, + 555, + 688 + ], + "type": "text", + "content": " improves accuracy, introducing an intermediate layer, that is, " + }, + { + "bbox": [ + 313, + 544, + 555, + 688 + ], + "type": "inline_equation", + "content": "S \\rightarrow \\{J, P_G, P_R\\}" + }, + { + "bbox": [ + 313, + 544, + 555, + 688 + ], + "type": "text", + "content": ", increases errors. This error propagates through the network, ultimately degrading pose estimation performance. This suggests that integrating IMU data with distance constraints may support a simpler architecture, reducing complexity while still maintaining accuracy." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "State Estimator The state estimator is designed to: 1) mitigate errors in inter-sensor distances, 2) filter and stabi" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7091" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 118, + 70, + 304, + 176 + ], + "blocks": [ + { + "bbox": [ + 118, + 70, + 304, + 176 + ], + "lines": [ + { + "bbox": [ + 118, + 70, + 304, + 176 + ], + "spans": [ + { + "bbox": [ + 118, + 70, + 304, + 176 + ], + "type": "image", + "image_path": "4f32f9d8ff28a532371bead867e00f2bd9ccec0340428cb24f58c31aeaab7f7d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 184, + 544, + 196 + ], + "lines": [ + { + "bbox": [ + 64, + 184, + 544, + 196 + ], + "spans": [ + { + "bbox": [ + 64, + 184, + 544, + 196 + ], + "type": "text", + "content": "Figure 4. Cumulative distribution of distance error (left) and acceleration error reduction over time (right) for various fusion settings." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 304, + 70, + 493, + 175 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 493, + 175 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 493, + 175 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 493, + 175 + ], + "type": "image", + "image_path": "57c0e9de5321df014c23e55ce2d008ef905b3cc2cdb9dc01018859df3d467c67.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 80, + 219, + 271, + 312 + ], + "blocks": [ + { + "bbox": [ + 80, + 219, + 271, + 312 + ], + "lines": [ + { + "bbox": [ + 80, + 219, + 271, + 312 + ], + "spans": [ + { + "bbox": [ + 80, + 219, + 271, + 312 + ], + "type": "image", + "image_path": "d5210017e2471484867f7fcabae5e64f03ff1ae2f8767ec024f0311535b07567.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 326, + 282, + 337 + ], + "lines": [ + { + "bbox": [ + 67, + 326, + 282, + 337 + ], + "spans": [ + { + "bbox": [ + 67, + 326, + 282, + 337 + ], + "type": "text", + "content": "Figure 5. Joint positional error for different fusion settings." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 361, + 296, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 613 + ], + "type": "text", + "content": "imize accelerations, and 3) improve the accuracy of estimated poses. To examine it we conduct an ablation study by varying fusion measurements on the TotalCapture [29] with synthesized LOS-related noisy distances. Fig. 4 (left) shows cumulative distributions of absolute distance error for different fusion settings. The raw distances show an average error of " + }, + { + "bbox": [ + 55, + 361, + 296, + 613 + ], + "type": "inline_equation", + "content": "9.20\\mathrm{cm}" + }, + { + "bbox": [ + 55, + 361, + 296, + 613 + ], + "type": "text", + "content": " with a standard deviation of " + }, + { + "bbox": [ + 55, + 361, + 296, + 613 + ], + "type": "inline_equation", + "content": "8.94\\mathrm{cm}" + }, + { + "bbox": [ + 55, + 361, + 296, + 613 + ], + "type": "text", + "content": ". By fusing IMU, UWB, and pose data, this error is reduced to " + }, + { + "bbox": [ + 55, + 361, + 296, + 613 + ], + "type": "inline_equation", + "content": "2.42\\mathrm{cm}" + }, + { + "bbox": [ + 55, + 361, + 296, + 613 + ], + "type": "text", + "content": " with a standard deviation of " + }, + { + "bbox": [ + 55, + 361, + 296, + 613 + ], + "type": "inline_equation", + "content": "2.39\\mathrm{cm}" + }, + { + "bbox": [ + 55, + 361, + 296, + 613 + ], + "type": "text", + "content": ", demonstrating the effectiveness of data fusion for mitigating distance errors. Fig. 4 (right) shows the reduction in acceleration errors. The state estimator converges within a few seconds and adapts gradually, aligning with IMU characteristics. Fig. 5 presents joint positional errors over time for different fusion setups. Without filtering, positional error fluctuates with distance measurement quality. Fusing IMU with either UWB or pose data alone offers limited improvement because distance and pose outliers remain. The combined fusion of IMU, UWB, and pose data achieves the lowest positional error, validating that our state estimator refines pose accuracy by utilizing all available measurements." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 623, + 185, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 185, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 185, + 635 + ], + "type": "text", + "content": "4.5. Qualitative Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "We developed a prototype (see supplementary material for details) that integrates BNO086 IMU and DW3000 UWB sensors to demonstrate our proposed method. Fig. 6 presents visualizations comparing pose estimates for challenging motions, illustrating the improvements of our method in disambiguating poses over IMU-only methods" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 345, + 213, + 525, + 390 + ], + "blocks": [ + { + "bbox": [ + 345, + 213, + 525, + 390 + ], + "lines": [ + { + "bbox": [ + 345, + 213, + 525, + 390 + ], + "spans": [ + { + "bbox": [ + 345, + 213, + 525, + 390 + ], + "type": "image", + "image_path": "4d0ceada2480686881f32963cffdc5c2f9c6b8c3bd46cda28050e7836f55c546.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 398, + 555, + 420 + ], + "lines": [ + { + "bbox": [ + 313, + 398, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 555, + 420 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison of pose estimates from the test data collected using our developed prototype." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 445, + 500, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 445, + 500, + 457 + ], + "spans": [ + { + "bbox": [ + 314, + 445, + 500, + 457 + ], + "type": "text", + "content": "and enhanced pose stability compared to UIP." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 474, + 471, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 471, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 471, + 487 + ], + "type": "text", + "content": "5. Conclusion and Limitations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 496, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 555, + 616 + ], + "type": "text", + "content": "In this work, we present UMotion, an uncertainty-driven framework for 3D human shape and pose estimation that integrates six IMU and UWB sensors within a state estimation system. UMotion incorporates uncertainties from IMU, UWB, and pose estimates under individual body constraints to iteratively enhance confidence in both sensor measurements and pose accuracy. Our experiments on synthetic and real-world datasets demonstrate that UMotion outperforms existing SOTA methods in pose accuracy, and effectively stabilizes sensor measurements and pose estimation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": "However, our shape estimator requires adaptation to specific sensor configurations and conditions, and, with limited training data, it may struggle with unique body variations. The state estimator also requires careful parameter tuning to reflect sensor characteristics accurately. Additionally, our pose estimator simplifies constraints, which may limit its performance. As future work, adding a physics-aware module could potentially enhance our method's robustness." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7092" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Rayan Armani, Changlin Qian, Jiaxi Jiang, and Christian Holz. Ultra inertial poser: Scalable motion capture and tracking from sparse inertial sensors and ultra-wideband ranging. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 212 + ], + "type": "text", + "content": "[2] Vasileios Choutas, Lea Müller, Chun-Hao P Huang, Siyu Tang, Dimitrios Tzionas, and Michael J Black. Accurate 3d body shape regression using metric and semantic attributes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2718-2728, 2022. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 213, + 295, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 213, + 295, + 278 + ], + "spans": [ + { + "bbox": [ + 62, + 213, + 295, + 278 + ], + "type": "text", + "content": "[3] Peng Dai, Yang Zhang, Tao Liu, Zhen Fan, Tianyuan Du, Zhuo Su, Xiaozheng Zheng, and Zeming Li. Hmd-poser: On-device real-time human motion tracking from scalable sparse observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 874-884, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 280, + 295, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 295, + 334 + ], + "type": "text", + "content": "[4] Nathan DeVrio, Vimal Mollyn, and Chris Harrison. Smartposer: Arm pose estimation with a smartphone and smartwatch using uwb and imu data. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, pages 1-11, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 335, + 295, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 295, + 378 + ], + "type": "text", + "content": "[5] Nick Erickson, Jonas Mueller, Alexander Shirkov, Hang Zhang, Pedro Larroy, Mu Li, and Alexander Smola. Autogluon-tabular: Robust and accurate automl for structured data. arXiv preprint arXiv:2003.06505, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 380, + 294, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 294, + 401 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 294, + 401 + ], + "type": "text", + "content": "[6] S Hochreiter. Long short-term memory. Neural Computation MIT-Press, 1997. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 403, + 295, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 295, + 456 + ], + "type": "text", + "content": "[7] Yinghao Huang, Manuel Kaufmann, Emre Aksan, Michael J Black, Otmar Hilliges, and Gerard Pons-Moll. Deep inertial poser: Learning to reconstruct human pose from sparse inertial measurements in real time. ACM Transactions on Graphics (TOG), 37(6):1-15, 2018. 1, 2, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 458, + 295, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 295, + 513 + ], + "type": "text", + "content": "[8] Jiaxi Jiang, Paul Streli, Huajian Qiu, Andreas Fender, Larissa Laich, Patrick Snape, and Christian Holz. Avatarposer: Articulated full-body pose tracking from sparse motion sensing. In European conference on computer vision, pages 443-460. Springer, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 514, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 294, + 546 + ], + "type": "text", + "content": "[9] Jiaxi Jiang, Paul Streli, Manuel Meier, and Christian Holz. Egoposer: Robust real-time ego-body pose estimation in large scenes. arXiv preprint arXiv:2308.06493, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 548, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 548, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 548, + 295, + 602 + ], + "type": "text", + "content": "[10] Yifeng Jiang, Yuting Ye, Deepak Gopinath, Jungdam Won, Alexander W Winkler, and C Karen Liu. Transformer inertial poser: Real-time human motion reconstruction from sparse imus with simultaneous terrain generation. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 603, + 295, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 295, + 647 + ], + "type": "text", + "content": "[11] Simon J Julier, Jeffrey K Uhlmann, and Hugh F Durrant-Whyte. A new approach for filtering nonlinear systems. In Proceedings of 1995 American Control Conference-ACC'95, pages 1628-1632. IEEE, 1995. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 712 + ], + "type": "text", + "content": "[12] Manuel Kaufmann, Yi Zhao, Chengcheng Tang, Lingling Tao, Christopher Twigg, Jie Song, Robert Wang, and Otmar Hilliges. Em-posed: 3d human pose estimation from sparse electromagnetic trackers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11510-11520, 2021. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[13] Jiye Lee and Hanbyul Joo. Mocap everyone everywhere: Lightweight motion capture with smartwatches and a head-mounted camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1091-1100, 2024. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 173 + ], + "type": "text", + "content": "[14] Sunmin Lee, Sebastian Starke, Yuting Ye, Jungdam Won, and Alexander Winkler. Questenvsim: Environment-aware simulated motion tracking from sparse sensors. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-9, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 175, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 229 + ], + "type": "text", + "content": "[15] Han Liang, Yannan He, Chengfeng Zhao, Mutian Li, Jingya Wang, Jingyi Yu, and Lan Xu. Hybridcap: Inertia-aid monocular capture of challenging human motions. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1539–1548, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 231, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 231, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 553, + 274 + ], + "type": "text", + "content": "[16] Huajun Liu, Xiaolin Wei, Jinxiang Chai, Inwoo Ha, and Taehyun Rhee. Realtime human motion control with a small number of inertial sensors. In Symposium on interactive 3D graphics and games, pages 133-140, 2011. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "type": "text", + "content": "[17] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. In *Seminal Graphics Papers: Pushing the Boundaries*, Volume 2, pages 851–866. 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 321, + 555, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 321, + 555, + 375 + ], + "spans": [ + { + "bbox": [ + 317, + 321, + 555, + 375 + ], + "type": "text", + "content": "[18] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5442-5451, 2019. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 376, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 376, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 317, + 376, + 553, + 431 + ], + "type": "text", + "content": "[19] Vimal Mollyn, Riku Arakawa, Mayank Goel, Chris Harrison, and Karan Ahuja. Imuposer: Full-body pose estimation using imus in phones, watches, and earbuds. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, pages 1-12, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 434, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 454 + ], + "type": "text", + "content": "[20] Ian Oppermann, Matti Hämäläinen, and Jari Inatti. *UWB: theory and applications*. John Wiley & Sons, 2004. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 456, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 510 + ], + "type": "text", + "content": "[21] Shaohua Pan, Qi Ma, Xinyu Yi, Weifeng Hu, Xiong Wang, Xingkang Zhou, Jijunnan Li, and Feng Xu. Fusing monocular images and sparse imu signals for real-time human motion capture. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 513, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 513, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 317, + 513, + 553, + 577 + ], + "type": "text", + "content": "[22] Gerard Pons-Moll, Andreas Baak, Thomas Helten, Meinard Müller, Hans-Peter Seidel, and Bodo Rosenhahn. Multisensor-fusion for 3d full-body human motion capture. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 663-670. IEEE, 2010." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 579, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 579, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 317, + 579, + 553, + 623 + ], + "type": "text", + "content": "[23] Jose Luis Ponton, Haoran Yun, Andreas Aristidou, Carlos Andujar, and Nuria Pelechano. Sparseposer: Real-time full-body motion reconstruction from sparse data. ACM Transactions on Graphics, 43(1):1-14, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 624, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 553, + 689 + ], + "type": "text", + "content": "[24] Sergi Pujades, Betty Mohler, Anne Thaler, Joachim Tesch, Naureen Mahmood, Nikolas Hesse, Heinrich H Bülthoff, and Michael J Black. The virtual caliper: rapid creation of metrically accurate avatars from 3d measurements. IEEE transactions on visualization and computer graphics, 25(5): 1887-1897, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[25] Yiming Ren, Chengfeng Zhao, Yannan He, Peishan Cong, Han Liang, Jingyi Yu, Lan Xu, and Yuexin Ma. Lidar-aid" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7093" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 295, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 295, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 295, + 106 + ], + "type": "text", + "content": "inertial poser: Large-scale human motion capture by sparse inertial and lidar sensors. IEEE Transactions on Visualization and Computer Graphics, 29(5):2337-2347, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 108, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 108, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 108, + 295, + 140 + ], + "type": "text", + "content": "[26] Xu Ru, Nian Gu, Hang Shang, and Heng Zhang. Mems inertial sensor calibration technology: Current status and future trends. Micromachines, 13(6):879, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 143, + 294, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 143, + 294, + 176 + ], + "spans": [ + { + "bbox": [ + 56, + 143, + 294, + 176 + ], + "type": "text", + "content": "[27] Martin Schepers, Matteo Giuberti, Giovanni Bellusci, et al. Xsens mvn: Consistent tracking of human motion using inertial sensing. Xsens Technol, 1(8):1-8, 2018. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 178, + 294, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 178, + 294, + 210 + ], + "spans": [ + { + "bbox": [ + 56, + 178, + 294, + 210 + ], + "type": "text", + "content": "[28] Mike Schuster and Kuldip K Paliwal. Bidirectional recurrent neural networks. IEEE transactions on Signal Processing, 45 (11):2673-2681, 1997. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 212, + 294, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 212, + 294, + 257 + ], + "spans": [ + { + "bbox": [ + 56, + 212, + 294, + 257 + ], + "type": "text", + "content": "[29] Matthew Trumble, Andrew Gilbert, Charles Malleson, Adrian Hilton, and John P Collomosse. Total capture: 3d human pose estimation fusing video and inertial sensors. In BMVC, pages 1-13. London, UK, 2017. 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 258, + 294, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 258, + 294, + 292 + ], + "spans": [ + { + "bbox": [ + 56, + 258, + 294, + 292 + ], + "type": "text", + "content": "[30] Rudolph Van Der Merwe. Sigma-point Kalman filters for probabilistic inference in dynamic state-space models. Oregon Health & Science University, 2004. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 293, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 294, + 337 + ], + "type": "text", + "content": "[31] Tom Van Wouwe, Seunghwan Lee, Antoine Falisse, Scott Delp, and C Karen Liu. Diffusion inertial poser: Human motion reconstruction from arbitrary sparse imu configurations. arXiv preprint arXiv:2308.16682, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 339, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 339, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 339, + 294, + 392 + ], + "type": "text", + "content": "[32] Timo Von Marcard, Bodo Rosenhahn, Michael J Black, and Gerard Pons-Moll. Sparse inertial poser: Automatic 3d human pose estimation from sparse imus. In Computer graphics forum, pages 349-360. Wiley Online Library, 2017. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 396, + 294, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 396, + 294, + 451 + ], + "spans": [ + { + "bbox": [ + 56, + 396, + 294, + 451 + ], + "type": "text", + "content": "[33] Eric A Wan and Rudolph Van Der Merwe. The unscented kalman filter for nonlinear estimation. In Proceedings of the IEEE 2000 adaptive systems for signal processing, communications, and control symposium (Cat. No. 00EX373), pages 153-158. Ieee, 2000. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 453, + 294, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 453, + 294, + 496 + ], + "spans": [ + { + "bbox": [ + 56, + 453, + 294, + 496 + ], + "type": "text", + "content": "[34] Xinyu Yi, Yuxiao Zhou, and Feng Xu. Transpose: Real-time 3d human translation and pose estimation with six inertial sensors. ACM Transactions On Graphics (TOG), 40(4):1-13, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 498, + 294, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 294, + 564 + ], + "type": "text", + "content": "[35] Xinyu Yi, Yuxiao Zhou, Marc Habermann, Soshi Shimada, Vladislav Golyanik, Christian Theobalt, and Feng Xu. Physical inertial poser (pip): Physics-aware real-time human motion tracking from sparse inertial sensors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13167-13178, 2022. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 566, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 566, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 566, + 294, + 621 + ], + "type": "text", + "content": "[36] Xinyu Yi, Yuxiao Zhou, Marc Habermann, Vladislav Golyanik, Shaohua Pan, Christian Theobalt, and Feng Xu. Ecolocate: Real-time motion capture, localization, and mapping with sparse body-mounted sensors. ACM Transactions on Graphics (TOG), 42(4):1-17, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 623, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 623, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 623, + 294, + 667 + ], + "type": "text", + "content": "[37] Xinyu Yi, Yuxiao Zhou, and Feng Xu. Physical non-inertial poser (pnp): Modeling non-inertial effects in sparse-inertial human motion capture. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "text", + "content": "[38] Faheem Zafari, Athanasios Gkelias, and Kin K Leung. A survey of indoor localization systems and technologies. IEEE Communications Surveys & Tutorials, 21(3):2568-2599, 2019. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 251 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 139 + ], + "type": "text", + "content": "[39] Yu Zhang, Songpengcheng Xia, Lei Chu, Jiarui Yang, Qi Wu, and Ling Pei. Dynamic inertial poser (dynaip): Part-based motion dynamics learning for enhanced human pose estimation with sparse inertial sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1889-1899, 2024. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 554, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 554, + 194 + ], + "type": "text", + "content": "[40] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5745-5753, 2019. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 196, + 554, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 196, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 196, + 554, + 251 + ], + "type": "text", + "content": "[41] Chengxu Zuo, Yiming Wang, Lishuang Zhan, Shihui Guo, Xinyu Yi, Feng Xu, and Yipeng Qin. Loose inertial poser: Motion capture with imu-attached loose-wear jacket. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2209–2219, 2024. 1" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7094" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/UNEM_ UNrolled Generalized EM for Transductive Few-Shot Learning/592279b8-3f9e-4b39-91e5-638d8677cb8a_content_list.json b/2025/UNEM_ UNrolled Generalized EM for Transductive Few-Shot Learning/592279b8-3f9e-4b39-91e5-638d8677cb8a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..911336311a1933a73a87ac065df848852d7b0339 --- /dev/null +++ b/2025/UNEM_ UNrolled Generalized EM for Transductive Few-Shot Learning/592279b8-3f9e-4b39-91e5-638d8677cb8a_content_list.json @@ -0,0 +1,1584 @@ +[ + { + "type": "text", + "text": "UNEM: UNrolled Generalized EM for Transductive Few-Shot Learning", + "text_level": 1, + "bbox": [ + 137, + 128, + 859, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Long Zhou $^{1,3}$ , Fereshteh Shakeri $^{2}$ , Aymen Sadraoui $^{3}$ , Mounir Kaaniche $^{3,4}$ , Jean-Christophe Pesquet $^{3}$ , Ismail Ben Ayed $^{2}$", + "bbox": [ + 238, + 191, + 754, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Politecnico di Milano", + "bbox": [ + 408, + 239, + 589, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2ÉTS Montréal", + "bbox": [ + 437, + 257, + 558, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Université Paris-Saclay, Inria, CentraleSupélec, CVN", + "bbox": [ + 282, + 275, + 712, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4Université Sorbonne Paris Nord, L2TI, UR 3043", + "bbox": [ + 302, + 294, + 692, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 356, + 326, + 372 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Transductive few-shot learning has recently triggered wide attention in computer vision. Yet, current methods introduce key hyper-parameters, which control the prediction statistics of the test batches, such as the level of class balance, affecting performances significantly. Such hyper-parameters are empirically grid-searched over validation data, and their configurations may vary substantially with the target dataset and pre-training model, making such empirical searches both sub-optimal and computationally intractable. In this work, we advocate and introduce the unrolling paradigm, also referred to as \"learning to optimize\", in the context of few-shot learning, thereby learning efficiently and effectively a set of optimized hyperparameters. Specifically, we unroll a generalization of the ubiquitous Expectation-Maximization (EM) optimizer into a neural network architecture, mapping each of its iterates to a layer and learning a set of key hyper-parameters over validation data. Our unrolling approach covers various statistical feature distributions and pre-training paradigms, including recent foundational vision-language models and standard vision-only classifiers. We report comprehensive experiments, which cover a breadth of fine-grained downstream image classification tasks, showing significant gains brought by the proposed unrolled EM algorithm over iterative variants. The achieved improvements reach up to $10\\%$ and $7.5\\%$ on vision-only and vision-language benchmarks, respectively. The source code and learned parameters are available at https://github.com/ZhouLong0/UNEM-Transductive.", + "bbox": [ + 89, + 446, + 483, + 883 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ceb9a74013acf42b64735a8a9a5a565d872d5c29145950cf8386ee1d954da2df.jpg", + "image_caption": [ + "Figure 1. Impact of the class-balance hyperparameter $\\lambda$ on the accuracy of transductive few-shot classification. The accuracy results are obtained using the EM-Dirichlet algorithm [33] applied to vision-language models (with 4-shots). The plot shows that the choice of $\\lambda$ has a strong impact on the performance, and that the optimal $\\lambda$ (indicated with the star symbol) might vary by orders of magnitudes, depending on the target downstream dataset (e.g., the ten very different fine-grained classification datasets). Further comments on the optimal $\\lambda$ values and the values chosen in [33] are provided in Section 4. The values of the learned hyper-parameters based on the proposed unrolled algorithm are illustrated and analyzed in Appendix D." + ], + "image_footnote": [], + "bbox": [ + 531, + 359, + 885, + 559 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 763, + 643, + 780 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning has transfigured computer vision, driving substantial progress in tasks such as image classification, captioning, object detection and segmentation. However, these successes often come with a high cost: the requirement for large amounts of labeled data. Additionally, the generalization of these models is seriously challenged when evaluated on new classes (concepts), unseen during pre", + "bbox": [ + 511, + 794, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9665", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "training, or when operating under distribution shifts [1, 53].", + "bbox": [ + 89, + 90, + 480, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, Few-Shot Learning (FSL) has recently attracted wide attention within the computer vision community. In the standard few-shot image classification setting, a feature extractor is first pre-trained on a set of classes, often called the base classes. Then, the model is adapted and evaluated on new classes and tasks. Each evaluation task is performed on a set of unlabeled samples (referred to as the query set), and supervised by a support set composed of few labeled samples per new class. Earlier FSL methods have relied on various concepts, including meta-learning [11, 19, 40], transfer learning [13, 38, 43], and metric learning [20, 26, 51]. However, most of these works operate within the inductive setting, where, at inference time, the prediction for each sample is made independently from the other samples in the query set.", + "bbox": [ + 89, + 109, + 482, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, significant attention has shifted to the transductive scheme, in which inference is performed jointly on a batch of query samples. Transduction leverages the statistics of the unlabeled query samples, yielding notable performance gains, and have triggered a large body of recent works in few-shot learning, with various methods and mechanisms [4, 23-25, 29, 32, 59], as detailed in Section 2. Generally, transductive methods are built upon optimizing objective functions that integrate clustering terms, either discriminative as in information maximization [4, 50] or generative as in probabilistic K-means [32, 33]. However, many among these assume a perfectly balanced class distribution within the query set, and incorporate terms in the objective function or constraints that enforce a class-balance prior. The latter could limit the applicability of these methods, whose performances were shown to drop significantly when dealing with imbalanced query sets [27, 50].", + "bbox": [ + 89, + 338, + 482, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several recent attempts tackled this limitation, to handle more realistic scenarios [24, 32, 33, 48, 50]. Yet, these methods introduce key hyper-parameters, which control the prediction statistics of the unlabeled query set, such as the level of class balance. Such hyper-parameters are empirically grid-searched over pre-defined sets of values using validation data, and their optimal configurations may vary substantially with the target dataset and pre-training model [17]. To illustrate this, we depict in Fig. 1 the accuracy as a function of a class-balance hyper-parameter for the recent method in [33]. One may observe that this hyperparameter has a crucial effect on the performance, and its optimal value might vary by orders of magnitude across the target datasets. The issue is further compounded by additional hyper-parameters, which makes intensive empirical grid searches for the hyper-parameters over validation data and pre-defined intervals of values both sub-optimal and computationally intractable. Therefore, we advocate and introduce the unrolling paradigm (also called \"learning to optimize\") in the context of few-shot learning, which enables", + "bbox": [ + 89, + 598, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to learn efficiently and effectively a set of optimized hyperparameters. Specifically, our main contributions could be summarized as follows:", + "bbox": [ + 511, + 90, + 903, + 135 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We study a generalization of the ubiquitous Expectation-Maximization (EM) algorithm, in which we make two hyper-parameters controlling prediction statistics - class balance and prediction entropy - explicit1. Our generalization encompasses several existing transductive few-shot methods as particular cases and, more importantly, enables to learn these crucial hyper-parameters through the proposed unrolling strategy.", + "2. We unroll the generalized EM optimizer into a neural network architecture, mapping each of its iterations to a layer and learning the introduced hyper-parameters over validation data. Our unrolling approach offers greater flexibility in optimizing the hyper-parameters, allowing them to vary across the network's layers. To the best of our knowledge, this study is the first to investigate unrolling in transductive few-shot learning.", + "3. We design our unrolling architecture in way that covers various statistical assumptions and pre-training paradigms: (i) Gaussian for vision-only classifiers and (ii) Dirichlet for vision-language models.", + "4. We report comprehensive experiments, which cover a breadth of fine-grained downstream image classification tasks and different pre-training models, showing substantial gains over recent state-of-the-art methods." + ], + "bbox": [ + 514, + 136, + 903, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 513, + 510, + 658, + 526 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Few-shot classification with vision-only models: Few-shot classification using vision models has been widely explored in the literature, leading to the development of various methods. The first category, inductive methods, predicts the class of each test sample (in the query set) independently from others [16, 30]. In contrast, the second category, transductive methods, which has gained increasing attention in recent years, involves jointly predicting classes for a batch of test samples in each few-shot task. Numerous research efforts in transductive methods have leveraged concepts like clustering [24, 32], label propagation [29, 59], information maximization [4, 50], optimal transport [23, 49], prototype estimation [28, 59], and variational networks [25, 44], among others. Studies have shown that transductive methods significantly outperform inductive approaches, achieving accuracy gains of up to $15\\%$ as reported in several evaluations [4, 32].", + "bbox": [ + 511, + 536, + 906, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Few-shot classification with vision-language models: Contrastive Vision-Language Pre-training (CLIP) has recently emerged as an effective model for enhancing various vision tasks through visual-text pairs. By learning transferable visual models with natural language supervision, CLIP", + "bbox": [ + 511, + 792, + 906, + 868 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1Those hyper-parameters are implicit (hidden) in the standard EM formulation.", + "bbox": [ + 513, + 875, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9666", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "demonstrates strong performance in zero-shot classification by matching the image features to text embeddings of novel classes [39]. To further enhance its classification capabilities, several studies have extended CLIP to the few-shot setting [6, 18, 55, 57, 58]. While the linear probe model [39] trains a logistic regression classifier using CLIP image features, the authors of [18] proposed a generalization of this baseline, modeling the classifier weights as learnable functions of the visual prototypes and text embeddings. Context Optimization (CoOp) [57] and its extended versions [6, 55, 58] have been developed based on the concept of prompt learning. While CoOp [57] aims to model context in prompts using continuous representations, a knowledge-guided CoOp is proposed in [55] to enhance the generalization ability by minimizing the discrepancy between the learnable prompts and the original ones. In [6], the authors learn multiple prompts, describing the characteristics of each class, through the minimization of an optimal-transport distance. Unlike prompt learning methods, which fine-tune the input text, another family of methods, referred to as adapters, aims to transform the visual or text encoders, such as [12, 56]. For instance, TIP-Adapter [56] adds a nonparametric adapter to the weight-frozen CLIP model, and updates the prior knowledge encoded in CLIP by feature retrieval. CLIP-Adapter [12] introduces a multi-layered perceptron to learn new features, and combine them with the original CLIP-encoded features via residual connections. It is worth noting that all the aforementioned methods belong to the inductive family. However, unlike vision-only models, the transductive few-shot setting is still not well investigated in the context of CLIP. To the best of our knowledge, the only transductive few-shot CLIP method is the one very recently proposed in [33], which, inspired by the Expectation Maximization algorithm, relies on the Dirichlet distribution to model the data. Moreover, as reported in [33], and unlike the behaviors observed with vision-only models, recent transductive few-shot methods do not always outperform their inductive counterparts with CLIP. This suggests that there is a need to further investigate transductive methods in the context of vision-language models, due to their aforementioned advantages with respect to inductive methods.", + "bbox": [ + 89, + 90, + 485, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Class-balance and hyperparameters setting: Hyperparameter optimization has a large impact on the performance of a given iterative algorithm or learning model, and is generally tackled through greedy or gradient-based methods [34]. In the context of transductive few-shot learning, most of the existing methods have been designed for perfectly balanced query sets (i.e., uniform class distribution), and have shown drops in performances under class-imbalanced settings. To address this flaw, various methods have been developed in the last years [24, 32, 33, 48, 50]. For instance, in [48] the categorical probability of each query", + "bbox": [ + 89, + 734, + 486, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sample is regularized to quantify the difference between the class marginal distribution and the uniform one. The works in [24, 32, 33, 50] explored various weighted terms, which are added to the objective functions, to mitigate the effect of the class-balance bias. To this end, some hyper-parameters are introduced, to weigh to contributions of such terms in the objective function. Unlike [32], where the introduced hyper-parameter has been theoretically chosen in order to compensate for the hidden class-balance bias in the used clustering objective, the hyper-parameters are often set in an empirical manner based on the validation classes of each dataset. Thus, a set of hyperparameter values (in a given range) are evaluated, and the ones yielding the highest accuracy are selected.", + "bbox": [ + 511, + 90, + 908, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed methodology", + "text_level": 1, + "bbox": [ + 511, + 319, + 728, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Generalized EM algorithm", + "text_level": 1, + "bbox": [ + 511, + 344, + 756, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preliminaries: Let us first introduce the notations to formulate our transductive few-shot inference. Let $\\{\\pmb{z}_n\\}_{1\\leq n\\leq N}$ denote the set of feature vectors extracted from a pretraining network, and which are to be classified, with $N$ the total number of samples within a given task. The whole dataset contains $K$ distinct classes, whereas the number of randomly sampled classes present in each mini-batch task might be much smaller than $K$ . This subset of sampled classes may also vary across mini-batches, and the method does not assume any prior knowledge on the specific classes that may appear in each mini-batch.", + "bbox": [ + 511, + 368, + 908, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, for a given few-shot task, let us denote by $\\mathbb{S} \\subset \\{1, \\dots, N\\}$ and $\\mathbb{Q} = \\{1, \\dots, N\\} \\setminus \\mathbb{S}$ the indices of samples within the support (labeled) and query (unlabeled) mini-batch sets, respectively. For every $n \\in \\mathbb{S}$ , $\\pmb{y}_n = (y_{n,k})_{1 \\leq k \\leq K} \\in \\{0, 1\\}^K$ are the one-hot-encoded labels, such that, for every $k \\in \\{1, \\dots, K\\}$ , $y_{n,k} = 1$ if the $n$ -th sample belongs to class $k$ , and $y_{n,k} = 0$ otherwise.", + "bbox": [ + 511, + 535, + 908, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, given the extracted feature vectors $\\mathbf{z}_n$ , we assume that the data probability distribution knowing its class $k$ is modeled by a given law, whose probability density function (pdf) is denoted by $\\mathrm{p}(\\mathbf{z}_n \\mid \\boldsymbol{\\theta}_k)$ and characterized by a vector of parameters $\\boldsymbol{\\theta}_k$ . This means that the global distribution of $\\mathbf{z}_n$ is a mixture of these pdfs.", + "bbox": [ + 511, + 642, + 908, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem formulation: Our goal is to identify the classes of the unlabeled samples in the query set by optimizing a general clustering objective function, while embedding supervision constraints from the few labeled samples within the support set. We do so by unrolling iterative block-coordinate optimizers of the objective functions over two sets of variables:", + "bbox": [ + 511, + 733, + 908, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Soft assignment vectors $\\pmb{u} = (u_n)_{1\\leq n\\leq N}\\in (\\Delta_K)^N$ where $\\Delta_K$ is the probability simplex of $\\mathbb{R}^K$ . For every $n\\in \\{1,\\dots ,N\\}$ , $\\pmb{u}_n = (u_{n,k})_{1\\leq k\\leq K}$ where, for every $k\\in \\{1,\\ldots ,K\\}$ , $u_{n,k}$ can be interpreted as the probabil", + "bbox": [ + 513, + 838, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9667", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ity that the $n$ -th sample belongs to class $k$ . These probabilities have to be determined for the query set samples.", + "bbox": [ + 102, + 90, + 482, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Feature distribution parameters $\\pmb{\\theta} = (\\pmb{\\theta}_k)_{1\\leq k\\leq K}$ .", + "bbox": [ + 89, + 121, + 429, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Consider the following general probabilistic clustering problem:", + "bbox": [ + 89, + 136, + 482, + 166 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\boldsymbol {u}, \\boldsymbol {\\theta}} {\\text {m i n i m i z e}} \\mathcal {L} (\\boldsymbol {u}, \\boldsymbol {\\theta}) + \\lambda \\Psi (\\boldsymbol {u}) + T \\Phi (\\boldsymbol {u}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 176, + 482, + 199 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "subject to $\\pmb{u}_n\\in \\Delta_K\\quad \\forall n\\in \\mathbb{Q},$", + "bbox": [ + 112, + 203, + 328, + 219 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nu _ {n, k} = y _ {n, k} \\quad \\forall n \\in \\mathbb {S}, \\forall k \\in \\{1, \\dots , K \\}.\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 222, + 459, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where weighing factor $\\lambda$ and temperature scaling $T$ are learnable optimized hyper-parameters, which we will estimate through the proposed unrolling strategy (Section 3.2), and terms $\\mathcal{L}$ , $\\Psi$ and $\\Phi$ are detailed in the following.", + "bbox": [ + 89, + 248, + 483, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- The first term in objective function (1) is the negative log-likelihood of the feature vectors:", + "bbox": [ + 89, + 309, + 482, + 338 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\boldsymbol {u}, \\boldsymbol {\\theta}) = - \\sum_ {n = 1} ^ {N} \\sum_ {k = 1} ^ {K} u _ {n, k} \\ln \\left(\\mathrm {p} \\left(\\boldsymbol {z} _ {n} \\mid \\boldsymbol {\\theta} _ {k}\\right)\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 345, + 482, + 387 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This general log-likelihood model fitting term is well known in the context of clustering methods [5, 21]. In fact, it generalizes the standard $K$ -means clustering objective to arbitrary distributions2. It is well-known that minimizing (2) has an inherent bias towards class-balanced clustering [5, 21, 32, 33].", + "bbox": [ + 102, + 393, + 483, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- The second term in (1) controls the partition complexity of the model, penalizing the number of non-empty clusters in the solution. It corresponds to the Shannon entropy of class distribution $(\\pi_k)_{1 \\leq k \\leq K}$ , defined as", + "bbox": [ + 89, + 486, + 483, + 547 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi (\\boldsymbol {u}) = - \\sum_ {k = 1} ^ {K} \\pi_ {k} \\ln \\pi_ {k}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 556, + 482, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pi_k = \\frac{1}{|\\mathbb{Q}|} \\sum_{n \\in \\mathbb{Q}} u_{n,k}$ is the proportion of query samples within class $k$ , i.e., the empirical estimate of the marginal probability of class $k$ .", + "bbox": [ + 102, + 607, + 482, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Class-balance hyper-parameter $(\\lambda)$ : The marginal entropy in (3) mitigates the class-balance bias of the log-likelihood clustering term in (2). It reaches its minimum for the extremely imbalanced solution in which all data samples are assigned to a single cluster, and its maximum for a perfectly balanced clustering. Hence, clearly, hyperparameter $\\lambda$ controls the level of class balance in the solution. As depicted in Fig. 1, this hyper-parameter has a crucial effect on the performance, and its optimal value might vary by orders of magnitude from one dataset to another. This makes exhaustive grid searches for optimal $\\lambda$ over validation sets intractable computationally, which motivates learning this hyper-parameter through our unrolling strategy, as described in Section 3.2.", + "bbox": [ + 102, + 655, + 483, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- The third term in (1) is an entropic barrier, enabling to soften assignments $u_{n,k}$ , while imposing a non-negativity constraint on each of them. It is given by", + "bbox": [ + 513, + 90, + 906, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi (\\boldsymbol {u}) = \\sum_ {n = 1} ^ {N} \\sum_ {k = 1} ^ {K} u _ {n, k} \\ln u _ {n, k}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 147, + 906, + 189 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temperature scaling hyper-parameter $(T)$ : Weighting factor $T$ in (1) controls the trade-off between the clustering term and the entropic barrier in (4), i.e., the level of softness of assignments $u_{n,k}$ . Therefore, this hyperparameter has an important effect on performances and on the marginal class probabilities appearing in the class-balance term in (3), which motivates learning it. As described in Section 3.2, our unrolling algorithm enables to cope with this extra parameter.", + "bbox": [ + 524, + 199, + 906, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Link to EM and other transductive few-shot methods: We examine solving problem (1) with an iterative block-coordinate descent algorithm (see Algorithm 1), which alternates two steps, one updating distribution parameters $\\pmb{\\theta}_k^{(\\ell)}$ and the other optimizing over class assignments $\\pmb{u}_n^{(\\ell)}$ , at each iteration $\\ell$ . For the $\\pmb{u}$ -step, and due to the nonconvexity of $\\Psi$ , we proceed with a Majorization-Minimization (MM) strategy to minimize a surrogate convex function with respect to $\\pmb{u}$ at each iteration. In addition, if $\\mathrm{p}(\\cdot|\\pmb{\\theta}_k)$ is assumed to belong to the exponential family and $\\pmb{\\theta}_k$ are its canonical parameters, the estimation w.r.t. $\\pmb{\\theta}$ , with $\\pmb{u}$ fixed, is a convex problem. We provide further details on this general iterative optimization scheme in Appendix A. It is worth noting that, when $T = 1$ and the data is modeled by the Dirichlet distribution, we recover the recent transductive few-shot learning algorithm in [33]. Interestingly, when $T = 1$ and $\\lambda = |\\mathbb{Q}|$ , we recover the well-known Expectation-Maximization (EM) algorithm for estimating the parameters of mixture of distributions; see Proposition 1 in [33]. Therefore, Algorithm 1 could be viewed as a generalized EM (GEM), in which hyper-parameters $\\lambda$ and $T$ control the class balance as well as prediction softness, and could be made learnable.", + "bbox": [ + 511, + 335, + 906, + 684 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Examples of data models: While a broad range of distribution models in the exponential family could be adopted, we focus in this paper on two popular ones.", + "bbox": [ + 511, + 685, + 905, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- The first one is the Gaussian distribution, which is commonly used in standard clustering and transductive few-shot-methods applied to vision models [32, 46, 50]. By assuming a Gaussian distribution with mean $\\pmb{\\theta}_{k}$ and identity covariance matrix, the pdf $\\mathrm{p}(\\pmb{z}_n|\\pmb{\\theta}_k)$ reads as follows", + "bbox": [ + 513, + 729, + 906, + 806 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {p} \\left(\\boldsymbol {z} _ {n} \\mid \\boldsymbol {\\theta} _ {k}\\right) \\propto \\exp \\left(- \\frac {1}{2} \\left\\| \\boldsymbol {z} _ {n} - \\boldsymbol {\\theta} _ {k} \\right\\| ^ {2}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 588, + 816, + 903, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- The second one is the Dirichlet distribution, which has recently shown good modelling performance in the context of transductive few-shot for vision-language mod", + "bbox": [ + 513, + 854, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "$^{2}\\mathrm{K}$ -means corresponds to choosing the multivariate Gaussian distribution, with identity covariance matrix, for parametric density $\\mathrm{p}(\\mathbf{z}_n|\\boldsymbol{\\theta}_k)$ .", + "bbox": [ + 89, + 875, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9668", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 GEM based few-shot classification algorithm", + "text_level": 1, + "bbox": [ + 91, + 90, + 478, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: Compute $\\mathbf{z}_n$ for the dataset samples, initialize $\\mathbf{u}_n^{(0)}$ and $\\theta_k^{(0)}$ , and fix the number of iterations $L$ , for $\\ell = 0,1,\\ldots ,L - 1$ do", + "bbox": [ + 104, + 111, + 785, + 143 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "// Update Distribution parameters for each class using a given estimation algorithm (denoted here by \"DP_est\")", + "bbox": [ + 122, + 145, + 849, + 159 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\theta} _ {k} ^ {(\\ell + 1)} = \\operatorname {D P} _ {-} \\operatorname {e s t} (\\boldsymbol {u} _ {., k} ^ {(\\ell)}, \\boldsymbol {\\theta} _ {k} ^ {(\\ell)}), \\quad \\forall k \\in \\{1, \\dots , K \\},\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 159, + 457, + 179 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "// Update class proportions", + "bbox": [ + 124, + 179, + 303, + 191 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\pi_ {k} ^ {(\\ell + 1)} = \\frac {1}{| \\mathbb {Q} |} \\sum_ {n \\in \\mathbb {Q}} u _ {n, k} ^ {(\\ell)}, \\quad \\forall k \\in \\{1, \\dots , K \\},\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 191, + 421, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "// Update assignment vectors for all query samples", + "bbox": [ + 124, + 226, + 457, + 239 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {u} _ {n} ^ {(\\ell + 1)} = \\operatorname {s o f t m a x} \\left(\\frac {1}{T} \\left(\\ln \\mathrm {p} \\left(\\boldsymbol {z} _ {n} \\mid \\boldsymbol {\\theta} _ {k} ^ {(\\ell + 1)}\\right) + \\frac {\\lambda}{| \\mathbb {Q} |} \\ln (\\pi_ {k} ^ {(\\ell + 1)})\\right) _ {k}\\right), \\quad \\forall n \\in \\mathbb {Q}.\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 239, + 640, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "end for", + "text_level": 1, + "bbox": [ + 107, + 271, + 161, + 282 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/638c5b2d4c01adbc9b9d282a46011431ac1796f80a8748d1e7a7b18db5dcad27.jpg", + "image_caption": [ + "Figure 2. An overview of the unrolled GEM algorithm for a given iteration. Each iteration $\\ell$ corresponds to a network layer $\\mathcal{L}^{(\\ell)}$ . Each layer depends on the vector of hyperparameters $(\\lambda^{(\\ell)}, T^{(\\ell)})$ ." + ], + "image_footnote": [], + "bbox": [ + 212, + 304, + 782, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "els such as CLIP [33]. For $\\mathbf{z}_n = (z_{n,i})_{1\\leq i\\leq K}$ and $\\pmb{\\theta}_{k} = (\\theta_{k,i})_{1\\leq i\\leq K}$ , the associated pdf is given by", + "bbox": [ + 102, + 497, + 482, + 530 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {p} \\left(\\boldsymbol {z} _ {n} \\mid \\boldsymbol {\\theta} _ {k}\\right) = \\frac {1}{\\mathcal {B} \\left(\\boldsymbol {\\theta} _ {k}\\right)} \\prod_ {i = 1} ^ {K} z _ {n, i} ^ {\\theta_ {k, i} - 1} \\mathbb {1} _ {\\boldsymbol {z} _ {n} \\in \\Delta_ {K}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 539, + 482, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the normalization factor $\\mathcal{B}(\\pmb{\\theta}_k)$ is", + "bbox": [ + 104, + 592, + 372, + 608 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {B} \\left(\\boldsymbol {\\theta} _ {k}\\right) = \\frac {\\prod_ {i = 1} ^ {K} \\Gamma \\left(\\theta_ {k , i}\\right)}{\\Gamma \\left(\\sum_ {i = 1} ^ {K} \\theta_ {k , i}\\right)}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 617, + 482, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and $\\Gamma$ denotes the Gamma function.", + "bbox": [ + 104, + 672, + 344, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The necessary details related to feature representations in the cases of vision-only and vision-language models, distribution parameter estimation, and the resulting GEM algorithms for Gaussian and Dirichlet laws are provided in Appendices B and C, respectively.", + "bbox": [ + 89, + 688, + 483, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. UNrolled EM architecture (UNEM)", + "text_level": 1, + "bbox": [ + 89, + 772, + 401, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overview: As discussed in Section 2, the choice of the optimal hyper-parameter, which controls the level of class balance, is a difficult task, often performed manually in the existing transductive few-shot classification methods. In addition to the class-balance parameter $\\lambda$ , Algorithm 1 involves an additional temperature parameter $T$ , which makes hyperparameter setting even more challenging. For this reason,", + "bbox": [ + 89, + 794, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we propose in this paper to resort to the unrolling (called also learning to optimize) paradigm [35], which enables to learn efficiently a set of optimized hyper-parameters. It is important to note that unrolling iterative optimization algorithms have found successful applications in diverse signal and image processing tasks [2, 7, 35]. However, to the best of our knowledge, this work is the first to leverage the unrolling paradigm for hyper-parameter optimization in a few-shot learning context. The main idea behind the learning-to-optimize paradigm is to map each iteration of a given optimization algorithm to a network layer, stack all layers together, and view the hyper-parameters to be optimized as the network's learnable parameters. More precisely, to unroll our generalized EM algorithm, the number of iterations $L$ is used as the number of layers for the neural network architecture. Thus, each iteration $\\ell \\in \\{0,1,\\dots ,L - 1\\}$ is associated with a tailored layer $\\mathcal{L}^{(\\ell)}$ , which performs the update rules defined in Algorithm 1:", + "bbox": [ + 511, + 497, + 906, + 770 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\boldsymbol {\\theta} _ {k} ^ {(\\ell + 1)}, \\pi_ {k} ^ {(\\ell + 1)}, \\boldsymbol {u} _ {n} ^ {(\\ell + 1)}\\right) = \\mathcal {L} ^ {(\\ell)} \\left(\\boldsymbol {\\theta} _ {k} ^ {(\\ell)}, \\pi_ {k} ^ {(\\ell)}, \\boldsymbol {u} _ {n} ^ {(\\ell)}; \\lambda^ {(\\ell)}, T ^ {(\\ell)}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 786, + 913, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $(\\lambda^{(\\ell)}, T^{(\\ell)})_{0 \\leq \\ell \\leq L-1}$ is the vector of hyperparameters to be learned. This leads to the UNrolled EM (UNEM) model shown in Fig. 2, which depicts the inputs and outputs of a given layer $\\mathcal{L}^{(l)}$ as well as its three main blocks for the update rules.", + "bbox": [ + 511, + 824, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9669", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/702073904ba52b0407f7950ed818c705d2325a204fe9a923703181ba12013a5a.jpg", + "image_caption": [ + "Figure 3. Overall architecture of the designed UNEM." + ], + "image_footnote": [], + "bbox": [ + 210, + 88, + 782, + 238 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Learned hyper-parameters: Instead of a restricted number of handcrafted hyper-parameters (as often used in iterative algorithms), our unrolled algorithm offers more flexibility, enabling adaptation of the hyper-parameters along the processing workflow. This means that $L$ vectors of hyper-parameters $(\\lambda^{(\\ell)}, T^{(\\ell)})_{0 \\leq \\ell \\leq L - 1}$ could be learned and applied at each layer of the network. During the training of the unrolled model, we fulfill some design constraints, as described in more details in the following.", + "bbox": [ + 88, + 281, + 482, + 416 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "More specifically, to ensure the non-negativity of hyperparameter $\\lambda^{(\\ell)}$ , we propose to express it as a Softplus function, which could be seen as a smooth approximation of the RELU activation function, yielding", + "bbox": [ + 89, + 417, + 483, + 478 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda^ {(\\ell)} = \\operatorname {S o f t p l u s} \\left(a ^ {(\\ell)}\\right) = \\log \\left(1 + \\exp \\left(a ^ {(\\ell)}\\right)\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 484, + 480, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $a^{(\\ell)}$ represents a learnable parameter for the unrolled architecture.", + "bbox": [ + 89, + 512, + 482, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Regarding temperature scaling $T^{(\\ell)}$ , we observed in our experiments that imposing the non-negativity constraint alone might be a source of instability, resulting in values very close to zero and vanishing gradient issues during the training of the unrolled architecture. To circumvent this problem, we impose a lower bound equal to 1 on the temperature scaling:", + "bbox": [ + 89, + 542, + 483, + 648 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nT ^ {(\\ell)} = 1 + \\operatorname {S o f t p l u s} (b ^ {(\\ell)}) = 1 + \\log (1 + \\exp (b ^ {(\\ell)})) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 656, + 482, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $b^{(\\ell)}$ denotes a parameter that needs to be learned during the training of the unrolled model.", + "bbox": [ + 89, + 684, + 482, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall architecture and training approach: Based on the previous considerations, the overall architecture of UNEM can be summarized as the composition of the $L$ layers $\\mathcal{L}^{(L - 1)}\\circ \\ldots \\circ \\mathcal{L}^{(0)}$ . This architecture, depicted in Figure 3, illustrates: (i) the required inputs $\\pmb{\\theta}_k^{(0)}$ and $\\pmb{u}_n^{(0)}$ as well as the feature vectors $\\pmb{z}_n$ (which are fed in all layers), (ii) the cascaded layers with their associated vector of parameters $\\pmb{w} = (\\lambda^{(\\ell)},T^{(\\ell)})_{0\\leq \\ell \\leq L - 1}$ satisfying the aforementioned constraints, and (iii) the key output $\\pmb{u}_n^{(L)}$ representing the class assignment vectors.", + "bbox": [ + 89, + 714, + 482, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The resulting neural network architecture is trained by minimizing a standard cross-entropy loss on a validation", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "set:", + "bbox": [ + 513, + 282, + 542, + 295 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {L} _ {\\mathrm {c}} (\\boldsymbol {w}) = \\sum_ {n \\in \\mathbb {Q}} \\sum_ {k = 1} ^ {K} y _ {n, k} \\log \\left(u _ {n, k} ^ {(L)}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 296, + 906, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 354, + 645, + 371 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The effectiveness of the proposed approach is validated in both vision-only and vision-language transductive few-shot learning settings. We will designate our proposed unrolled methods by UNEM-Gaussian and UNEM-Dirichlet, respectively. After describing the experimental settings, this section is structured into two main sections discussing our experiments in both evaluation scenarios.", + "bbox": [ + 511, + 380, + 906, + 486 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental settings", + "text_level": 1, + "bbox": [ + 511, + 500, + 718, + 515 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Task generation: We adopt a realistic transductive few-shot evaluation protocol, which is in line with state-of-the-art protocols [29, 32, 33, 45, 50]. While $K$ designates the total number of classes in the labeled support set $\\mathbb{S}$ , $K_{\\mathrm{eff}}$ (with $K_{\\mathrm{eff}} \\ll K$ ) denotes the number of effective classes present in the unlabeled query set $\\mathbb{Q}$ . The classes in the query set remain undisclosed during inference, while randomly selecting $|\\mathbb{Q}|$ samples. On the other hand, the support set is built by uniformly selecting $s$ images from each of the $K$ classes. In this paper, the few-shot tasks are performed with (i) 5, 10, and 20 shots for vision-only models, and (ii) 4 shots for vision-language models. Moreover, in both of our evaluation scenarios, we used $K_{\\mathrm{eff}} = 5$ and $|\\mathbb{Q}| = 75$ .", + "bbox": [ + 511, + 523, + 906, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training specifications: For fair comparison and reproducibility purposes, the CLIP's pre-trained model has been directly used in the vision-language evaluation scenario, while for the vision-only scenario, the standard pre-trained ResNet-18 and WRN28-10 backbones have been finetuned. Their training is performed on the base classes set of each vision-only dataset, using cross-entropy loss with label smoothing set to 0.9 for 90 epochs. The learning rate is set to 0.1 and then decayed by a factor of 10.", + "bbox": [ + 511, + 719, + 908, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On the other hand, the training of our unrolled architecture, composed of 10 layers (i.e., $L = 10$ ), is performed on several tasks sampled from the validation set of each dataset.", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9670", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/816412569d426a3b8f7d37e2b4664d54ce997a11df573309a78e41f0362787d8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackbonemini-ImageNet (K=20)tiered-ImageNet (K=160)
5-shot10-shot20-shot5-shot10-shot20-shot
Baseline [8]55.462.167.929.736.342.2
LR+ICI [52]55.462.168.1---
BD-CSPN [28]49.854.656.511.411.011.7
PT-MAP [17]25.727.228.45.26.06.6
LaplacianShot [60]ResNet-1857.964.268.329.635.439.1
TIM [4]66.869.970.829.328.727.8
α-TIM [50]66.771.073.943.848.351.9
α-AM [24]64.467.870.1---
PADDLE [32]62.973.579.845.461.470.6
UNEM-Gaussian66.475.680.452.365.773.2
Baseline [8]59.065.772.131.939.045.6
LR+ICI [52]58.865.772.0---
BD-CSPN [28]51.155.558.418.018.418.2
PT-MAP [17]26.528.029.35.26.06.6
LaplacianShot [60]WRN28-1061.066.871.031.437.341.5
TIM [4]72.174.976.236.139.038.5
α-TIM [50]71.575.278.345.851.455.2
α-AM [24]68.271.373.3---
PADDLE [32]62.673.079.243.959.469.9
UNEM-Gaussian71.679.283.754.166.874.7
", + "bbox": [ + 232, + 87, + 767, + 383 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/83dbb3e6233185346c6cd735c25dcf1ebda60169ad5d7d4f124d490f4fc2168d.jpg", + "table_caption": [ + "Table 1. Comparison of the proposed UNEM-Gaussian with respect to state-of-the-art methods on mini-Imagenet and tiered-Imagenet. The metric is accuracy (in percentage). Results are averaged across 1,000 tasks. Results marked with ’ - ’ were intractable to obtain." + ], + "table_footnote": [], + "table_body": "
MethodCUB (K=50)
5-shot10-shot20-shot
Baseline [8]58.668.878.2
LR+ICI [52]49.955.658.0
PT-MAP [17]12.814.014.9
LaplacianShot [60]58.866.571.0
TIM [4]68.168.969.3
α-TIM [50]74.379.383.6
α-AM [24]66.268.969.8
PADDLE [32]71.281.886.8
UNEM-Gaussian78.585.388.6
", + "bbox": [ + 145, + 444, + 431, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison of the proposed UNEM-Gaussian with respect to state-of-the-art methods on CUB. The metric is accuracy (in percentage). Results are averaged across 1,000 tasks.", + "bbox": [ + 89, + 606, + 482, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This implies that the proposed solution is highly economical in terms of parameters, typically requiring only a few tenths, compared to standard neural network architectures. Let us recall that the validation set is often used to select the hyperparameters in recent state-of-the-art transductive methods as mentioned in Section 2. Specifically, for the vision-only (resp. vision-language) models, the training of the unrolled architecture is carried out on 1,000 (resp. 100) tasks, while using an initial learning rate of 0.1 (resp. 0.5) with a decay factor of 0.5, a number of epochs equals to 80, and ADAM as optimizer. The trainable hyper-parameters are initialized to the default values set in [32] (i.e., $\\lambda = |Q|$ for the vision-only models) and [33] (i.e., $\\lambda = \\frac{K}{K_{\\mathrm{eff}}}$ for the vision-language models, with $T = 1$ ). Our architecture is", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "implemented in Pytorch (version 2.3.0) and run on NVIDIA QUADRO RTX8000 (with 48 GB of memory).", + "bbox": [ + 511, + 448, + 906, + 478 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. UNEM-Gaussian in vision-only few-shot setting", + "text_level": 1, + "bbox": [ + 511, + 484, + 906, + 502 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Let us recall that the Gaussian distribution is commonly used in standard clustering and transductive few-shot methods applied to vision-only models [32, 46, 50]. For this reason, we will focus here on the proposed UNEM-Gaussian approach.", + "bbox": [ + 511, + 508, + 905, + 582 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets: The first UNEM-Gaussian architecture is evaluated on the following standard few-shot benchmark datasets: mini-[ImageNet [51], tiered-[ImageNet [41], CUB [14]. Mini-imagenet has 100 classes split into 64 base classes, 16 validation classes and 20 test classes. The tiered-imagenet has 608 classes instead, from which we follow a standard split with 351 for base training, 97 for validation and 160 for testing. For CUB, we followed the split proposed by [8] which consists of 100 base classes, 50 validation classes and 50 test classes. For each dataset, the feature vectors $\\mathbf{z}_n$ are extracted using the fine-tuned backbones, while applying a scaling parameter $T_{z}$ to the model's output as described in Appendix B. This parameter has also been learned through the unrolling approach.", + "bbox": [ + 511, + 583, + 905, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results: The UNEM-Gaussian architecture has been compared to its original version PADDLE [32] as well as several state-of-the-art methods. Tables 1 and 2 presents the results, averaged over 1,000 tasks with 5, 10, and 20 shots. Several observations can be made. First, the proposed UNEM-Gaussian outperforms the original PADDLE algorithm [32] and the other state-of-the-art methods. The achieved gains", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9671", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dcc3a6804b601dd9a08d204180d15d465b210c08383591e6d8e56d0d8df918e7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFood101EuroSATDTDOrionPedsFlowers102Caltech101UCF101FGVC AircraftStanford CarsSUM397ImageNetAverage
Ind.Tip-Adapter [56]76.772.554.786.483.288.872.123.763.966.762.768.3
CoOp [57]76.363.252.286.281.087.767.022.261.363.459.965.5
Transd.BDSCPN [28]74.746.145.281.374.282.059.018.048.154.549.257.5
Laplacian Shot [60]76.653.052.688.485.586.867.022.260.463.856.364.8
α-TIM [50]66.146.145.387.179.183.359.420.453.453.442.757.8
PADDLE [32]71.845.950.084.782.381.963.721.356.160.652.160.9
EM-Dirichlet [33]88.750.862.692.591.390.176.124.973.580.978.473.6
UNEM-Dirichlet91.453.865.396.095.693.478.530.480.088.583.177.8
", + "bbox": [ + 143, + 89, + 854, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Comparison of the proposed UNEM-Dirichlet with respect to the state-of-the-art methods on 11 different datasets. The metric is accuracy (in percentage). Results are averaged across 1,000 tasks.", + "bbox": [ + 89, + 250, + 903, + 280 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "are much higher with fewer number of shots and reach $3.5\\%$ with mini-ImageNet, $6.9\\%$ with tiered-ImageNet, and $7.3\\%$ with CUB, while using ResNet-18 as backbone. Moreover, when WRN28-10 is used as backbone, it can be noticed that the performance of all methods has been improved, except for PADDLE, which shows results similar to those obtained with ResNet-18 pre-training. This can be explained by the fact that the class-balance parameter $\\lambda$ was set in [32] to $|\\mathbb{Q}|$ (i.e. 75), which becomes sub-optimal with WRN28-10 features. Most importantly, by learning efficiently the hyperparameters, our UNEM-Gaussian achieves higher gain with respect to its original version, yielding an accuracy gain reaching up to $10\\%$ in 5-shot scenario, when WRN28-10 is used as backbone.", + "bbox": [ + 91, + 306, + 483, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. UNEM-Dirichlet for few-shot CLIP", + "text_level": 1, + "bbox": [ + 89, + 529, + 401, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dirichlet distribution has recently demonstrated good modelling performance in the context of transductive few-shot for vision-language models such as CLIP [33]. Thus, we will focus in this second round of experiments on UNEM-Dirichlet approach.", + "bbox": [ + 89, + 551, + 482, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets: The second unrolled architecture, designated by UNEM-Dirichlet, is assessed on 11 benchmark datasets that are commonly used for CLIP scenario: Caltech101 [10], OxfordPnts [37], StanfordCars [22], Flowers102 [36], Food101 [3], FGVCAircraft [31], SUN397 [54], DTD [9], EuroSAT [15], UCF101 [47], and ImageNet [42]. These datasets cover diverse classification challenges, from object recognition (Caltech101, Food101) to fine-grained tasks (StanfordCars, FGVCAircraft) and scene understanding (SUN397, EuroSAT). For each dataset, the vision-text feature vectors $z_{n}$ are extracted using the CLIP's pre-trained model, while applying a temperature parameter $T_{z}$ as described in Appendix C. This temperature parameter has also been learned through the unrolling approach.", + "bbox": [ + 89, + 628, + 482, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results: The proposed UNEM-Dirichlet has also been compared to its original EM-Dirichlet version [33] as well as different few-shot classification methods. Table 3 depicts the accuracy results evaluated over 1,000 tasks with 4-", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "shots. Thus, it can be noticed that UNEM-Dirichlet outperforms state-of-the-art methods for most datasets. In particular, unrolling the recent EM-Dirichlet algorithm yields further improvement of about $4.2\\%$ in average. The achieved gain is more significant (reaching up to $7.5\\%$ ) with more challenging datasets having a large number of classes like FGVC Aircraft, Stanford Cars, and SUN397. This is due to the inappropriate choice of the hyperparameter $\\lambda$ for these datasets. Indeed, the selected $\\lambda$ parameter was set to $\\frac{K}{K_{\\mathrm{eff}}} |\\mathbb{Q}|$ in EM-Dirichlet approach [33]. Thus, by computing the numerical values of the hyperparameter for each of the aforementioned datasets, it can be deduced from Figure 1 that the difference between the selected value (which is 3,000 for Stanford Cars) and the optimal one (which is around 5,500 for Stanford Cars) is more important and impacts significantly the accuracy performance. This problem is well addressed by resorting to the proposed unrolled model for hyperparameters optimization.", + "bbox": [ + 511, + 306, + 906, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Appendix D includes additional results to illustrate, in both scenarios, the effects of the temperature parameter in this framework, and to show the benefits of learning variable hyperparameters across the architecture layers.", + "bbox": [ + 511, + 580, + 906, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 661, + 633, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we focus on transductive few-shot methods grounded on a generalized form of the EM algorithm. These methods include a parameter to control class balance. The proposed GEM algorithm, applicable to any mixture of distributions, incorporates also a temperature scaling parameter. The optimization process is then unrolled into a neural network architecture, enabling efficient learning of the introduced hyper-parameters. The results demonstrate the effectiveness of the proposed approach on both vision-only and vision-language models. In future work, it would be valuable to explore the potential of unrolling techniques across a broader range of computer vision tasks, especially with the recent rise of foundational vision-language models.", + "bbox": [ + 511, + 688, + 908, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9672", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Christina Baek, Yiding Jiang, Aditi Raghunathan, and Zico Kolter. Agreement-on-the-line: Predicting the performance of neural networks under distribution shift. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), pages 19274-19289, 2022. 2", + "[2] C. Bertocchi, E. Chouzenoux, M.-C. Corbineau, J.-C. Pesquet, and M. Prato. Deep unfolding of a proximal interior point method for image restoration. Inverse Problems, Special Issue on Variational Methods and Effective Algorithms for Imaging and Vision, 36(3):1-27, 2020. 5", + "[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In European Conference on Computer Vision, pages 446-461, Zurich, Switzerland, 2014. Springer. 8", + "[4] Malik Boudiaf, Ziko Imtiaz Masud, Jérôme Rony, José Dolz, Pablo Piantanida, and Ismail Ben Ayed. Transductive information maximization for few-shot learning. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), pages 2445–2457, 2020. 2, 7", + "[5] Y. Boykov, H. N. Isack, C. Olsson, and I. Ben Ayed. Volumetric bias in segmentation and reconstruction: Secrets and solutions. In IEEE/CVF International Conference on Computer Vision (ICCV), 2015. 4", + "[6] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Prompt learning with optimal transport for vision-language models. In International Conference on Learning Representations (ICLR), 2023. 3", + "[7] Tianlong Chen, Xiaohan Chen, Wuyang Chen, and Zhangyang Wang. Learning to optimize: A primer and a benchmark. Journal of Machine Learning Research (JMLR), 23:1-59, 2022. 5", + "[8] Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, and Jia-Bin Huang. A closer look at few-shot classification. In International Conference on Learning Representations (ICLR), pages 1-19, 2019. 7", + "[9] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, and A. Vedaldi. Describing textures in the wild. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 8", + "[10] Li Fei-Fei, Robert Fergus, and Pietro Perona. One-shot learning of object categories. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28(4):594-611, 2006. 8", + "[11] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International Conference on Machine Learning, 2017. 2", + "[12] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. CLIP-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 32:581-595, 2023. 3", + "[13] Bharath Hariharan and Ross B. Girshick. Low-shot visual recognition by shrinking and hallucinating features. In IEEE International Conference on Computer Vision (ICCV), pages 3018-3027, 2017. 2", + "[14] Xiangteng He and Yuxin Peng. Fine-grained visual-textual" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "representation learning. IEEE Transactions on Circuits and Systems for Video Technology, 30(2):520-531, 2020. 7", + "[15] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. EuroSAT: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 8", + "[16] Shell Xu Hu, Da Li, Jan Stuhmer, Minyoung Kim, and Timothy M Hospedales. Pushing the limits of simple pipelines for few-shot learning: External data and fine-tuning make a difference. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9068-9077, 2022. 2", + "[17] Yuqing Hu, Vincent Gripon, and Stéphane Pateux. Leveraging the feature distribution in transfer-based few-shot learning. In International Conference on Artificial Neural Networks, pages 487-499, 2021. 2, 7", + "[18] Yunshi Huang, Fereshteh Shakeri, Jose Dolz, Malik Boudiaf, Houda Bahig, and Ismail Ben Ayed. $\\mathrm{LP} + +$ : A surprisingly strong linear probe for few-shot clip. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23773-23782, 2024. 3", + "[19] Muhammad Abdullah Jamal and Guo-Jun Qi. Task agnostic meta-learning for few-shot learning. In IEEE conference on Computer Vision and Pattern Recognition (CVPR), pages 11719-11727, 2019. 2", + "[20] Leonid Karlinsky, Joseph Shtok, Sivan Harary, Eli Schwartz, Amit Aides, and Rogerio Feris. Repmet: Representative-based metric learning for classification and few-shot object detection. In IEEE conference on Computer Vision and Pattern Recognition (CVPR), pages 5197-5206, 2019. 2", + "[21] M. Kearns, Y. Mansour, and A. Ng. An information-theoretic analysis of hard and soft assignment methods for clustering. In Conference on Uncertainty in Artificial Intelligence (UAI), 1997. 4", + "[22] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 8", + "[23] Michalis Lazarou, Tania Stathaki, and Yannis Avrithis. Iterative label cleaning for transductive and semi-supervised few-shot learning. In IEEE/CVF International Conference on Computer Vision, pages 8751-8760, 2021. 2", + "[24] Michalis Lazarou, Yannis Avrithis, and Tania Stathaki. Adaptive manifold for imbalanced transductive few-shot learning. In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), page 2297-2306, 2024. 2, 3, 7", + "[25] Gao Yu Lee, Tanmoy Dam, Daniel Puiu Poenar, Vu N. Duong, and Md Meftahul Ferdaus. HELA-VFA: A hellinger distance-attention-based feature aggregation network for few-shot classification. In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2173–2183, 2024. 2", + "[26] Xiaoxu Li, Xiaochen Yang, Zhanyu Ma, and Jing-Hao Xue. Deep metric learning for few-shot image classification: A review of recent developments. Pattern Recognition, 138: 109381, 2023. 2" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9673", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Moshe Lichtenstein, Prasanna Sattigeri, Rogerio Feris, Raja Giryes, and Leonid Karlinsky. TAFSSL: Task-adaptive feature sub-space learning for few-shot classification. In European Conference on Computer Vision (ECCV), pages 22-539, 2020. 2", + "[28] Jinlu Liu, Liang Song, and Yongqiang Qin. Prototype rectification for few-shot learning. In European Conference on Computer Vision (ECCV), pages 741-756, 2020. 2, 7, 8", + "[29] Yanbin Liu, Juho Lee, Minseop Park, Saehoon Kim, Eunho Yang, Sung Ju Hwang, and Yi Yang. Learning to propagate labels: Transductive propagation network for few-shot learning. In International Conference on Learning Representations, page 34677-34688, 2019. 2, 6", + "[30] Xu Luo, Hao Wu, Ji Zhang, Lianli Gao, Jing Xu, and Jingkuan Song. A closer look at few-shot classification again. In International Conference on Machine Learning (ICML), pages 23103-23123, 2023. 2", + "[31] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. Technical report, Oxford University, 2013. 8", + "[32] Ségolène Martin, Malik Boudiaf, Emilie Chouzenoux, Jean-Christophe Pesquet, and Ismail Ben Ayed. Towards practical few-shot query sets: Transductive minimum description length inference. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), page 34677–34688, 2022. 2, 3, 4, 6, 7, 8, 5", + "[33] Ségolène Martin, Yunshi Huang, Fereshteh Shakeri, Jean-Christophe Pesquet, and Ismail Ben Ayed. Transductive zero-shot and few-shot clip. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28816-28826, 2024. 1, 2, 3, 4, 5, 6, 7, 8", + "[34] Paul Micaelli and Amos J Storkey. Gradient-based hyperparameter optimization over long horizons. In Advances in Neural Information Processing Systems, pages 10798-10809. Curran Associates, Inc., 2021. 3", + "[35] Vishal Monga, Yuelong Li, and Yonina C. Eldar. Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing. IEEE Signal Processing Magazine, 38(2):18-44, 2021. 5", + "[36] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 8", + "[37] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 8", + "[38] Hang Qi, Matthew Brown, and David G Lowe. Low-shot learning with imprinted weights. In IEEE conference on Computer Vision and Pattern Recognition (CVPR), pages 5822-5830, 2018. 2", + "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, and et al. Jack Clark. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 3" + ], + "bbox": [ + 91, + 90, + 485, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Sachin Ravi and H. Larochelle. Optimization as a model for few-shot learning. In International Conference on Learning Representations, pages 1-11, 2017. 2", + "[41] Mengye Ren, Eleni Triantafillou, Sachin Ravi, Jake Snell, Kevin Swersky, Joshua B. Tenenbaum, Hugo Larochelle, and Richard S. Zemel. Meta-learning for semi-supervised few-shot classification. In International Conference on Learning Representations (ICLR), pages 1-15, 2018. 7", + "[42] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. Int. J. Comput. Vision, 115(3):211-252, 2015. 8", + "[43] Tyler Scott, Karl Ridgeway, and Michael C. Mozer. Adapted deep embeddings: A synthesis of methods for k-shot inductive transfer learning. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), pages 76-85, 2018. 2", + "[44] Anuj Singh and Hadi Jamali-Rad. Transductive decoupled variational inference for few-shot classification. Transactions on Machine Learning Research (TMLR), 2023. 2", + "[45] Jake Snell, Kevin Swersky, and Richard S. Zemel. Prototypical networks for few-shot learning. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), pages 4080-4090, 2017. 6", + "[46] Yu Song and Changsheng Chen. MPPCANet: A feedforward learning strategy for few-shot image classification. Pattern Recognition, 113:107792, 2021. 4, 7", + "[47] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. Technical report, Center for Research in Computer Vision, University of Central Florida, 2012. arXiv:1212.0402. 8", + "[48] Ran Tao, Hao Chen, and Marios Savvides. Boosting transductive few-shot fine-tuning with margin-based uncertainty weighting and probability regularization. In IEEE conference on Computer Vision and Pattern Recognition (CVPR), pages 15752-15761, 2023. 2, 3", + "[49] Long Tian, Jingyi Feng, Xiaoqiang Chai, Wenchao Chen, Liming Wang, Xiyang Liu, and Bo Chen. Prototypes-oriented transductive few-shot learning with conditional transport. In International Conference on Computer Vision (ICCV), pages 16317-16326, 2023. 2", + "[50] Olivier Veilleux, Malik Boudiaf, Pablo Piantanida, and Ismail Ben Ayed. Realistic evaluation of transductive few-shot learning. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), pages 9290-9302, 2021. 2, 3, 4, 6, 7, 8", + "[51] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), pages 3637-3645, 2016. 2, 7", + "[52] Yikai Wang, Chengming Xu, Chen Liu, Li Zhang, and Yanwei Fu. Instance credibility inference for few-shot learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12833-12842, 2020. 7" + ], + "bbox": [ + 516, + 90, + 906, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9674", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Yimu Wang, Yihan Wu, and Hongyang Zhang. Lost domain generalization is a natural consequence of lack of training domains. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 15689-15697, 2024. 2", + "[54] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3485-3492, 2010. 8", + "[55] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6757-6767, 2023. 3", + "[56] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. TIP-adapter: Training-free adaption of clip for few-shot classification. In European Conference on Computer Vision (ECCV), pages 493-510, 2022. 3, 8", + "[57] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16816-16825, 2022. 3, 8", + "[58] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In International Conference on Computer Vision (ICCV), pages 15659-15669, 2023. 3", + "[59] Hao Zhu and Piotr Koniusz. Transductive few-shot learning with prototype-based label propagation by iterative graph refinement. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 23996-24006, 2023. 2", + "[60] Imtiaz Masud Ziko, Jose Dolz, Eric Granger, and Ismail Ben Ayed. Laplacian regularized few-shot learning. In International Conference on Machine Learning (ICML), 2020. 7, 8" + ], + "bbox": [ + 91, + 92, + 480, + 583 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9675", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/UNEM_ UNrolled Generalized EM for Transductive Few-Shot Learning/592279b8-3f9e-4b39-91e5-638d8677cb8a_model.json b/2025/UNEM_ UNrolled Generalized EM for Transductive Few-Shot Learning/592279b8-3f9e-4b39-91e5-638d8677cb8a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..034e884fddacebb3995271efb5ae485f56606533 --- /dev/null +++ b/2025/UNEM_ UNrolled Generalized EM for Transductive Few-Shot Learning/592279b8-3f9e-4b39-91e5-638d8677cb8a_model.json @@ -0,0 +1,2235 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.138, + 0.13, + 0.861, + 0.154 + ], + "angle": 0, + "content": "UNEM: UNrolled Generalized EM for Transductive Few-Shot Learning" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.193, + 0.756, + 0.23 + ], + "angle": 0, + "content": "Long Zhou\\(^{1,3}\\), Fereshteh Shakeri\\(^{2}\\), Aymen Sadraoui\\(^{3}\\), Mounir Kaaniche\\(^{3,4}\\), Jean-Christophe Pesquet\\(^{3}\\), Ismail Ben Ayed\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.409, + 0.241, + 0.59, + 0.258 + ], + "angle": 0, + "content": "1Politecnico di Milano" + }, + { + "type": "text", + "bbox": [ + 0.438, + 0.258, + 0.56, + 0.276 + ], + "angle": 0, + "content": "2ÉTS Montréal" + }, + { + "type": "text", + "bbox": [ + 0.284, + 0.276, + 0.713, + 0.294 + ], + "angle": 0, + "content": "3Université Paris-Saclay, Inria, CentraleSupélec, CVN" + }, + { + "type": "text", + "bbox": [ + 0.303, + 0.295, + 0.694, + 0.311 + ], + "angle": 0, + "content": "4Université Sorbonne Paris Nord, L2TI, UR 3043" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.357, + 0.327, + 0.373 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.485, + 0.884 + ], + "angle": 0, + "content": "Transductive few-shot learning has recently triggered wide attention in computer vision. Yet, current methods introduce key hyper-parameters, which control the prediction statistics of the test batches, such as the level of class balance, affecting performances significantly. Such hyper-parameters are empirically grid-searched over validation data, and their configurations may vary substantially with the target dataset and pre-training model, making such empirical searches both sub-optimal and computationally intractable. In this work, we advocate and introduce the unrolling paradigm, also referred to as \"learning to optimize\", in the context of few-shot learning, thereby learning efficiently and effectively a set of optimized hyperparameters. Specifically, we unroll a generalization of the ubiquitous Expectation-Maximization (EM) optimizer into a neural network architecture, mapping each of its iterates to a layer and learning a set of key hyper-parameters over validation data. Our unrolling approach covers various statistical feature distributions and pre-training paradigms, including recent foundational vision-language models and standard vision-only classifiers. We report comprehensive experiments, which cover a breadth of fine-grained downstream image classification tasks, showing significant gains brought by the proposed unrolled EM algorithm over iterative variants. The achieved improvements reach up to \\(10\\%\\) and \\(7.5\\%\\) on vision-only and vision-language benchmarks, respectively. The source code and learned parameters are available at https://github.com/ZhouLong0/UNEM-Transductive." + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.36, + 0.886, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.563, + 0.908, + 0.73 + ], + "angle": 0, + "content": "Figure 1. Impact of the class-balance hyperparameter \\(\\lambda\\) on the accuracy of transductive few-shot classification. The accuracy results are obtained using the EM-Dirichlet algorithm [33] applied to vision-language models (with 4-shots). The plot shows that the choice of \\(\\lambda\\) has a strong impact on the performance, and that the optimal \\(\\lambda\\) (indicated with the star symbol) might vary by orders of magnitudes, depending on the target downstream dataset (e.g., the ten very different fine-grained classification datasets). Further comments on the optimal \\(\\lambda\\) values and the values chosen in [33] are provided in Section 4. The values of the learned hyper-parameters based on the proposed unrolled algorithm are illustrated and analyzed in Appendix D." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.765, + 0.645, + 0.781 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Deep learning has transfigured computer vision, driving substantial progress in tasks such as image classification, captioning, object detection and segmentation. However, these successes often come with a high cost: the requirement for large amounts of labeled data. Additionally, the generalization of these models is seriously challenged when evaluated on new classes (concepts), unseen during pre" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9665" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.482, + 0.108 + ], + "angle": 0, + "content": "training, or when operating under distribution shifts [1, 53]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.11, + 0.483, + 0.337 + ], + "angle": 0, + "content": "To address these challenges, Few-Shot Learning (FSL) has recently attracted wide attention within the computer vision community. In the standard few-shot image classification setting, a feature extractor is first pre-trained on a set of classes, often called the base classes. Then, the model is adapted and evaluated on new classes and tasks. Each evaluation task is performed on a set of unlabeled samples (referred to as the query set), and supervised by a support set composed of few labeled samples per new class. Earlier FSL methods have relied on various concepts, including meta-learning [11, 19, 40], transfer learning [13, 38, 43], and metric learning [20, 26, 51]. However, most of these works operate within the inductive setting, where, at inference time, the prediction for each sample is made independently from the other samples in the query set." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.339, + 0.483, + 0.597 + ], + "angle": 0, + "content": "Recently, significant attention has shifted to the transductive scheme, in which inference is performed jointly on a batch of query samples. Transduction leverages the statistics of the unlabeled query samples, yielding notable performance gains, and have triggered a large body of recent works in few-shot learning, with various methods and mechanisms [4, 23-25, 29, 32, 59], as detailed in Section 2. Generally, transductive methods are built upon optimizing objective functions that integrate clustering terms, either discriminative as in information maximization [4, 50] or generative as in probabilistic K-means [32, 33]. However, many among these assume a perfectly balanced class distribution within the query set, and incorporate terms in the objective function or constraints that enforce a class-balance prior. The latter could limit the applicability of these methods, whose performances were shown to drop significantly when dealing with imbalanced query sets [27, 50]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Several recent attempts tackled this limitation, to handle more realistic scenarios [24, 32, 33, 48, 50]. Yet, these methods introduce key hyper-parameters, which control the prediction statistics of the unlabeled query set, such as the level of class balance. Such hyper-parameters are empirically grid-searched over pre-defined sets of values using validation data, and their optimal configurations may vary substantially with the target dataset and pre-training model [17]. To illustrate this, we depict in Fig. 1 the accuracy as a function of a class-balance hyper-parameter for the recent method in [33]. One may observe that this hyperparameter has a crucial effect on the performance, and its optimal value might vary by orders of magnitude across the target datasets. The issue is further compounded by additional hyper-parameters, which makes intensive empirical grid searches for the hyper-parameters over validation data and pre-defined intervals of values both sub-optimal and computationally intractable. Therefore, we advocate and introduce the unrolling paradigm (also called \"learning to optimize\") in the context of few-shot learning, which enables" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.136 + ], + "angle": 0, + "content": "to learn efficiently and effectively a set of optimized hyperparameters. Specifically, our main contributions could be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.137, + 0.905, + 0.258 + ], + "angle": 0, + "content": "1. We study a generalization of the ubiquitous Expectation-Maximization (EM) algorithm, in which we make two hyper-parameters controlling prediction statistics - class balance and prediction entropy - explicit1. Our generalization encompasses several existing transductive few-shot methods as particular cases and, more importantly, enables to learn these crucial hyper-parameters through the proposed unrolling strategy." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.258, + 0.905, + 0.378 + ], + "angle": 0, + "content": "2. We unroll the generalized EM optimizer into a neural network architecture, mapping each of its iterations to a layer and learning the introduced hyper-parameters over validation data. Our unrolling approach offers greater flexibility in optimizing the hyper-parameters, allowing them to vary across the network's layers. To the best of our knowledge, this study is the first to investigate unrolling in transductive few-shot learning." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.379, + 0.905, + 0.438 + ], + "angle": 0, + "content": "3. We design our unrolling architecture in way that covers various statistical assumptions and pre-training paradigms: (i) Gaussian for vision-only classifiers and (ii) Dirichlet for vision-language models." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.439, + 0.905, + 0.499 + ], + "angle": 0, + "content": "4. We report comprehensive experiments, which cover a breadth of fine-grained downstream image classification tasks and different pre-training models, showing substantial gains over recent state-of-the-art methods." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.137, + 0.905, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.511, + 0.66, + 0.527 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.907, + 0.793 + ], + "angle": 0, + "content": "Few-shot classification with vision-only models: Few-shot classification using vision models has been widely explored in the literature, leading to the development of various methods. The first category, inductive methods, predicts the class of each test sample (in the query set) independently from others [16, 30]. In contrast, the second category, transductive methods, which has gained increasing attention in recent years, involves jointly predicting classes for a batch of test samples in each few-shot task. Numerous research efforts in transductive methods have leveraged concepts like clustering [24, 32], label propagation [29, 59], information maximization [4, 50], optimal transport [23, 49], prototype estimation [28, 59], and variational networks [25, 44], among others. Studies have shown that transductive methods significantly outperform inductive approaches, achieving accuracy gains of up to \\(15\\%\\) as reported in several evaluations [4, 32]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.793, + 0.907, + 0.869 + ], + "angle": 0, + "content": "Few-shot classification with vision-language models: Contrastive Vision-Language Pre-training (CLIP) has recently emerged as an effective model for enhancing various vision tasks through visual-text pairs. By learning transferable visual models with natural language supervision, CLIP" + }, + { + "type": "page_footnote", + "bbox": [ + 0.514, + 0.875, + 0.905, + 0.9 + ], + "angle": 0, + "content": "1Those hyper-parameters are implicit (hidden) in the standard EM formulation." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9666" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.725 + ], + "angle": 0, + "content": "demonstrates strong performance in zero-shot classification by matching the image features to text embeddings of novel classes [39]. To further enhance its classification capabilities, several studies have extended CLIP to the few-shot setting [6, 18, 55, 57, 58]. While the linear probe model [39] trains a logistic regression classifier using CLIP image features, the authors of [18] proposed a generalization of this baseline, modeling the classifier weights as learnable functions of the visual prototypes and text embeddings. Context Optimization (CoOp) [57] and its extended versions [6, 55, 58] have been developed based on the concept of prompt learning. While CoOp [57] aims to model context in prompts using continuous representations, a knowledge-guided CoOp is proposed in [55] to enhance the generalization ability by minimizing the discrepancy between the learnable prompts and the original ones. In [6], the authors learn multiple prompts, describing the characteristics of each class, through the minimization of an optimal-transport distance. Unlike prompt learning methods, which fine-tune the input text, another family of methods, referred to as adapters, aims to transform the visual or text encoders, such as [12, 56]. For instance, TIP-Adapter [56] adds a nonparametric adapter to the weight-frozen CLIP model, and updates the prior knowledge encoded in CLIP by feature retrieval. CLIP-Adapter [12] introduces a multi-layered perceptron to learn new features, and combine them with the original CLIP-encoded features via residual connections. It is worth noting that all the aforementioned methods belong to the inductive family. However, unlike vision-only models, the transductive few-shot setting is still not well investigated in the context of CLIP. To the best of our knowledge, the only transductive few-shot CLIP method is the one very recently proposed in [33], which, inspired by the Expectation Maximization algorithm, relies on the Dirichlet distribution to model the data. Moreover, as reported in [33], and unlike the behaviors observed with vision-only models, recent transductive few-shot methods do not always outperform their inductive counterparts with CLIP. This suggests that there is a need to further investigate transductive methods in the context of vision-language models, due to their aforementioned advantages with respect to inductive methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.487, + 0.904 + ], + "angle": 0, + "content": "Class-balance and hyperparameters setting: Hyperparameter optimization has a large impact on the performance of a given iterative algorithm or learning model, and is generally tackled through greedy or gradient-based methods [34]. In the context of transductive few-shot learning, most of the existing methods have been designed for perfectly balanced query sets (i.e., uniform class distribution), and have shown drops in performances under class-imbalanced settings. To address this flaw, various methods have been developed in the last years [24, 32, 33, 48, 50]. For instance, in [48] the categorical probability of each query" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.304 + ], + "angle": 0, + "content": "sample is regularized to quantify the difference between the class marginal distribution and the uniform one. The works in [24, 32, 33, 50] explored various weighted terms, which are added to the objective functions, to mitigate the effect of the class-balance bias. To this end, some hyper-parameters are introduced, to weigh to contributions of such terms in the objective function. Unlike [32], where the introduced hyper-parameter has been theoretically chosen in order to compensate for the hidden class-balance bias in the used clustering objective, the hyper-parameters are often set in an empirical manner based on the validation classes of each dataset. Thus, a set of hyperparameter values (in a given range) are evaluated, and the ones yielding the highest accuracy are selected." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.32, + 0.73, + 0.337 + ], + "angle": 0, + "content": "3. Proposed methodology" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.345, + 0.757, + 0.361 + ], + "angle": 0, + "content": "3.1. Generalized EM algorithm" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.369, + 0.909, + 0.534 + ], + "angle": 0, + "content": "Preliminaries: Let us first introduce the notations to formulate our transductive few-shot inference. Let \\(\\{\\pmb{z}_n\\}_{1\\leq n\\leq N}\\) denote the set of feature vectors extracted from a pretraining network, and which are to be classified, with \\(N\\) the total number of samples within a given task. The whole dataset contains \\(K\\) distinct classes, whereas the number of randomly sampled classes present in each mini-batch task might be much smaller than \\(K\\). This subset of sampled classes may also vary across mini-batches, and the method does not assume any prior knowledge on the specific classes that may appear in each mini-batch." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.536, + 0.909, + 0.642 + ], + "angle": 0, + "content": "Moreover, for a given few-shot task, let us denote by \\(\\mathbb{S} \\subset \\{1, \\dots, N\\}\\) and \\(\\mathbb{Q} = \\{1, \\dots, N\\} \\setminus \\mathbb{S}\\) the indices of samples within the support (labeled) and query (unlabeled) mini-batch sets, respectively. For every \\(n \\in \\mathbb{S}\\), \\(\\pmb{y}_n = (y_{n,k})_{1 \\leq k \\leq K} \\in \\{0, 1\\}^K\\) are the one-hot-encoded labels, such that, for every \\(k \\in \\{1, \\dots, K\\}\\), \\(y_{n,k} = 1\\) if the \\(n\\)-th sample belongs to class \\(k\\), and \\(y_{n,k} = 0\\) otherwise." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.643, + 0.909, + 0.734 + ], + "angle": 0, + "content": "Finally, given the extracted feature vectors \\( \\mathbf{z}_n \\), we assume that the data probability distribution knowing its class \\( k \\) is modeled by a given law, whose probability density function (pdf) is denoted by \\( \\mathrm{p}(\\mathbf{z}_n \\mid \\boldsymbol{\\theta}_k) \\) and characterized by a vector of parameters \\( \\boldsymbol{\\theta}_k \\). This means that the global distribution of \\( \\mathbf{z}_n \\) is a mixture of these pdfs." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.734, + 0.909, + 0.838 + ], + "angle": 0, + "content": "Problem formulation: Our goal is to identify the classes of the unlabeled samples in the query set by optimizing a general clustering objective function, while embedding supervision constraints from the few labeled samples within the support set. We do so by unrolling iterative block-coordinate optimizers of the objective functions over two sets of variables:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.839, + 0.909, + 0.903 + ], + "angle": 0, + "content": "- Soft assignment vectors \\(\\pmb{u} = (u_n)_{1\\leq n\\leq N}\\in (\\Delta_K)^N\\) where \\(\\Delta_K\\) is the probability simplex of \\(\\mathbb{R}^K\\). For every \\(n\\in \\{1,\\dots ,N\\}\\), \\(\\pmb{u}_n = (u_{n,k})_{1\\leq k\\leq K}\\) where, for every \\(k\\in \\{1,\\ldots ,K\\}\\), \\(u_{n,k}\\) can be interpreted as the probabil" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9667" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.103, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "ity that the \\(n\\)-th sample belongs to class \\(k\\). These probabilities have to be determined for the query set samples." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.122, + 0.431, + 0.137 + ], + "angle": 0, + "content": "- Feature distribution parameters \\(\\pmb{\\theta} = (\\pmb{\\theta}_k)_{1\\leq k\\leq K}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.483, + 0.167 + ], + "angle": 0, + "content": "Consider the following general probabilistic clustering problem:" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.177, + 0.483, + 0.2 + ], + "angle": 0, + "content": "\\[\n\\underset {\\boldsymbol {u}, \\boldsymbol {\\theta}} {\\text {m i n i m i z e}} \\mathcal {L} (\\boldsymbol {u}, \\boldsymbol {\\theta}) + \\lambda \\Psi (\\boldsymbol {u}) + T \\Phi (\\boldsymbol {u}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.204, + 0.329, + 0.22 + ], + "angle": 0, + "content": "subject to \\(\\pmb{u}_n\\in \\Delta_K\\quad \\forall n\\in \\mathbb{Q},\\)" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.223, + 0.46, + 0.24 + ], + "angle": 0, + "content": "\\[\nu _ {n, k} = y _ {n, k} \\quad \\forall n \\in \\mathbb {S}, \\forall k \\in \\{1, \\dots , K \\}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.249, + 0.484, + 0.308 + ], + "angle": 0, + "content": "where weighing factor \\(\\lambda\\) and temperature scaling \\(T\\) are learnable optimized hyper-parameters, which we will estimate through the proposed unrolling strategy (Section 3.2), and terms \\(\\mathcal{L}\\), \\(\\Psi\\) and \\(\\Phi\\) are detailed in the following." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.31, + 0.483, + 0.339 + ], + "angle": 0, + "content": "- The first term in objective function (1) is the negative log-likelihood of the feature vectors:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.346, + 0.483, + 0.388 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\boldsymbol {u}, \\boldsymbol {\\theta}) = - \\sum_ {n = 1} ^ {N} \\sum_ {k = 1} ^ {K} u _ {n, k} \\ln \\left(\\mathrm {p} \\left(\\boldsymbol {z} _ {n} \\mid \\boldsymbol {\\theta} _ {k}\\right)\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.395, + 0.484, + 0.487 + ], + "angle": 0, + "content": "This general log-likelihood model fitting term is well known in the context of clustering methods [5, 21]. In fact, it generalizes the standard \\(K\\)-means clustering objective to arbitrary distributions2. It is well-known that minimizing (2) has an inherent bias towards class-balanced clustering [5, 21, 32, 33]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.487, + 0.484, + 0.549 + ], + "angle": 0, + "content": "- The second term in (1) controls the partition complexity of the model, penalizing the number of non-empty clusters in the solution. It corresponds to the Shannon entropy of class distribution \\((\\pi_k)_{1 \\leq k \\leq K}\\), defined as" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.557, + 0.483, + 0.599 + ], + "angle": 0, + "content": "\\[\n\\Psi (\\boldsymbol {u}) = - \\sum_ {k = 1} ^ {K} \\pi_ {k} \\ln \\pi_ {k}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.608, + 0.483, + 0.656 + ], + "angle": 0, + "content": "where \\(\\pi_k = \\frac{1}{|\\mathbb{Q}|} \\sum_{n \\in \\mathbb{Q}} u_{n,k}\\) is the proportion of query samples within class \\(k\\), i.e., the empirical estimate of the marginal probability of class \\(k\\)." + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.656, + 0.484, + 0.867 + ], + "angle": 0, + "content": "Class-balance hyper-parameter \\((\\lambda)\\): The marginal entropy in (3) mitigates the class-balance bias of the log-likelihood clustering term in (2). It reaches its minimum for the extremely imbalanced solution in which all data samples are assigned to a single cluster, and its maximum for a perfectly balanced clustering. Hence, clearly, hyperparameter \\(\\lambda\\) controls the level of class balance in the solution. As depicted in Fig. 1, this hyper-parameter has a crucial effect on the performance, and its optimal value might vary by orders of magnitude from one dataset to another. This makes exhaustive grid searches for optimal \\(\\lambda\\) over validation sets intractable computationally, which motivates learning this hyper-parameter through our unrolling strategy, as described in Section 3.2." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.907, + 0.138 + ], + "angle": 0, + "content": "- The third term in (1) is an entropic barrier, enabling to soften assignments \\( u_{n,k} \\), while imposing a non-negativity constraint on each of them. It is given by" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.148, + 0.907, + 0.19 + ], + "angle": 0, + "content": "\\[\n\\Phi (\\boldsymbol {u}) = \\sum_ {n = 1} ^ {N} \\sum_ {k = 1} ^ {K} u _ {n, k} \\ln u _ {n, k}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.2, + 0.907, + 0.336 + ], + "angle": 0, + "content": "Temperature scaling hyper-parameter \\((T)\\): Weighting factor \\(T\\) in (1) controls the trade-off between the clustering term and the entropic barrier in (4), i.e., the level of softness of assignments \\(u_{n,k}\\). Therefore, this hyperparameter has an important effect on performances and on the marginal class probabilities appearing in the class-balance term in (3), which motivates learning it. As described in Section 3.2, our unrolling algorithm enables to cope with this extra parameter." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.337, + 0.907, + 0.685 + ], + "angle": 0, + "content": "Link to EM and other transductive few-shot methods: We examine solving problem (1) with an iterative block-coordinate descent algorithm (see Algorithm 1), which alternates two steps, one updating distribution parameters \\(\\pmb{\\theta}_k^{(\\ell)}\\) and the other optimizing over class assignments \\(\\pmb{u}_n^{(\\ell)}\\), at each iteration \\(\\ell\\). For the \\(\\pmb{u}\\)-step, and due to the nonconvexity of \\(\\Psi\\), we proceed with a Majorization-Minimization (MM) strategy to minimize a surrogate convex function with respect to \\(\\pmb{u}\\) at each iteration. In addition, if \\(\\mathrm{p}(\\cdot|\\pmb{\\theta}_k)\\) is assumed to belong to the exponential family and \\(\\pmb{\\theta}_k\\) are its canonical parameters, the estimation w.r.t. \\(\\pmb{\\theta}\\), with \\(\\pmb{u}\\) fixed, is a convex problem. We provide further details on this general iterative optimization scheme in Appendix A. It is worth noting that, when \\(T = 1\\) and the data is modeled by the Dirichlet distribution, we recover the recent transductive few-shot learning algorithm in [33]. Interestingly, when \\(T = 1\\) and \\(\\lambda = |\\mathbb{Q}|\\), we recover the well-known Expectation-Maximization (EM) algorithm for estimating the parameters of mixture of distributions; see Proposition 1 in [33]. Therefore, Algorithm 1 could be viewed as a generalized EM (GEM), in which hyper-parameters \\(\\lambda\\) and \\(T\\) control the class balance as well as prediction softness, and could be made learnable." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.686, + 0.906, + 0.731 + ], + "angle": 0, + "content": "Examples of data models: While a broad range of distribution models in the exponential family could be adopted, we focus in this paper on two popular ones." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.731, + 0.907, + 0.808 + ], + "angle": 0, + "content": "- The first one is the Gaussian distribution, which is commonly used in standard clustering and transductive few-shot-methods applied to vision models [32, 46, 50]. By assuming a Gaussian distribution with mean \\(\\pmb{\\theta}_{k}\\) and identity covariance matrix, the pdf \\(\\mathrm{p}(\\pmb{z}_n|\\pmb{\\theta}_k)\\) reads as follows" + }, + { + "type": "equation", + "bbox": [ + 0.589, + 0.817, + 0.905, + 0.851 + ], + "angle": 0, + "content": "\\[\n\\mathrm {p} \\left(\\boldsymbol {z} _ {n} \\mid \\boldsymbol {\\theta} _ {k}\\right) \\propto \\exp \\left(- \\frac {1}{2} \\left\\| \\boldsymbol {z} _ {n} - \\boldsymbol {\\theta} _ {k} \\right\\| ^ {2}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "- The second one is the Dirichlet distribution, which has recently shown good modelling performance in the context of transductive few-shot for vision-language mod" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\(^{2}\\mathrm{K}\\)-means corresponds to choosing the multivariate Gaussian distribution, with identity covariance matrix, for parametric density \\(\\mathrm{p}(\\mathbf{z}_n|\\boldsymbol{\\theta}_k)\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9668" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.48, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 GEM based few-shot classification algorithm" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.112, + 0.787, + 0.145 + ], + "angle": 0, + "content": "Input: Compute \\( \\mathbf{z}_n \\) for the dataset samples, initialize \\( \\mathbf{u}_n^{(0)} \\) and \\( \\theta_k^{(0)} \\), and fix the number of iterations \\( L \\), for \\( \\ell = 0,1,\\ldots ,L - 1 \\) do" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.146, + 0.851, + 0.16 + ], + "angle": 0, + "content": "// Update Distribution parameters for each class using a given estimation algorithm (denoted here by \"DP_est\")" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.16, + 0.458, + 0.18 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\theta} _ {k} ^ {(\\ell + 1)} = \\operatorname {D P} _ {-} \\operatorname {e s t} (\\boldsymbol {u} _ {., k} ^ {(\\ell)}, \\boldsymbol {\\theta} _ {k} ^ {(\\ell)}), \\quad \\forall k \\in \\{1, \\dots , K \\},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.18, + 0.304, + 0.192 + ], + "angle": 0, + "content": "// Update class proportions" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.192, + 0.422, + 0.227 + ], + "angle": 0, + "content": "\\[\n\\pi_ {k} ^ {(\\ell + 1)} = \\frac {1}{| \\mathbb {Q} |} \\sum_ {n \\in \\mathbb {Q}} u _ {n, k} ^ {(\\ell)}, \\quad \\forall k \\in \\{1, \\dots , K \\},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.227, + 0.458, + 0.24 + ], + "angle": 0, + "content": "// Update assignment vectors for all query samples" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.24, + 0.642, + 0.274 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {u} _ {n} ^ {(\\ell + 1)} = \\operatorname {s o f t m a x} \\left(\\frac {1}{T} \\left(\\ln \\mathrm {p} \\left(\\boldsymbol {z} _ {n} \\mid \\boldsymbol {\\theta} _ {k} ^ {(\\ell + 1)}\\right) + \\frac {\\lambda}{| \\mathbb {Q} |} \\ln (\\pi_ {k} ^ {(\\ell + 1)})\\right) _ {k}\\right), \\quad \\forall n \\in \\mathbb {Q}.\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.272, + 0.163, + 0.284 + ], + "angle": 0, + "content": "end for" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.305, + 0.784, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.443, + 0.908, + 0.473 + ], + "angle": 0, + "content": "Figure 2. An overview of the unrolled GEM algorithm for a given iteration. Each iteration \\(\\ell\\) corresponds to a network layer \\(\\mathcal{L}^{(\\ell)}\\). Each layer depends on the vector of hyperparameters \\((\\lambda^{(\\ell)}, T^{(\\ell)})\\)." + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.498, + 0.483, + 0.531 + ], + "angle": 0, + "content": "els such as CLIP [33]. For \\( \\mathbf{z}_n = (z_{n,i})_{1\\leq i\\leq K} \\) and \\( \\pmb{\\theta}_{k} = (\\theta_{k,i})_{1\\leq i\\leq K} \\), the associated pdf is given by" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.54, + 0.483, + 0.582 + ], + "angle": 0, + "content": "\\[\n\\mathrm {p} \\left(\\boldsymbol {z} _ {n} \\mid \\boldsymbol {\\theta} _ {k}\\right) = \\frac {1}{\\mathcal {B} \\left(\\boldsymbol {\\theta} _ {k}\\right)} \\prod_ {i = 1} ^ {K} z _ {n, i} ^ {\\theta_ {k, i} - 1} \\mathbb {1} _ {\\boldsymbol {z} _ {n} \\in \\Delta_ {K}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.593, + 0.373, + 0.609 + ], + "angle": 0, + "content": "where the normalization factor \\(\\mathcal{B}(\\pmb{\\theta}_k)\\) is" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.618, + 0.483, + 0.664 + ], + "angle": 0, + "content": "\\[\n\\mathcal {B} \\left(\\boldsymbol {\\theta} _ {k}\\right) = \\frac {\\prod_ {i = 1} ^ {K} \\Gamma \\left(\\theta_ {k , i}\\right)}{\\Gamma \\left(\\sum_ {i = 1} ^ {K} \\theta_ {k , i}\\right)}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.674, + 0.345, + 0.687 + ], + "angle": 0, + "content": "and \\(\\Gamma\\) denotes the Gamma function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.765 + ], + "angle": 0, + "content": "The necessary details related to feature representations in the cases of vision-only and vision-language models, distribution parameter estimation, and the resulting GEM algorithms for Gaussian and Dirichlet laws are provided in Appendices B and C, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.402, + 0.789 + ], + "angle": 0, + "content": "3.2. UNrolled EM architecture (UNEM)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Overview: As discussed in Section 2, the choice of the optimal hyper-parameter, which controls the level of class balance, is a difficult task, often performed manually in the existing transductive few-shot classification methods. In addition to the class-balance parameter \\(\\lambda\\), Algorithm 1 involves an additional temperature parameter \\(T\\), which makes hyperparameter setting even more challenging. For this reason," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.498, + 0.907, + 0.771 + ], + "angle": 0, + "content": "we propose in this paper to resort to the unrolling (called also learning to optimize) paradigm [35], which enables to learn efficiently a set of optimized hyper-parameters. It is important to note that unrolling iterative optimization algorithms have found successful applications in diverse signal and image processing tasks [2, 7, 35]. However, to the best of our knowledge, this work is the first to leverage the unrolling paradigm for hyper-parameter optimization in a few-shot learning context. The main idea behind the learning-to-optimize paradigm is to map each iteration of a given optimization algorithm to a network layer, stack all layers together, and view the hyper-parameters to be optimized as the network's learnable parameters. More precisely, to unroll our generalized EM algorithm, the number of iterations \\( L \\) is used as the number of layers for the neural network architecture. Thus, each iteration \\( \\ell \\in \\{0,1,\\dots ,L - 1\\} \\) is associated with a tailored layer \\( \\mathcal{L}^{(\\ell)} \\), which performs the update rules defined in Algorithm 1:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.787, + 0.915, + 0.825 + ], + "angle": 0, + "content": "\\[\n\\left(\\boldsymbol {\\theta} _ {k} ^ {(\\ell + 1)}, \\pi_ {k} ^ {(\\ell + 1)}, \\boldsymbol {u} _ {n} ^ {(\\ell + 1)}\\right) = \\mathcal {L} ^ {(\\ell)} \\left(\\boldsymbol {\\theta} _ {k} ^ {(\\ell)}, \\pi_ {k} ^ {(\\ell)}, \\boldsymbol {u} _ {n} ^ {(\\ell)}; \\lambda^ {(\\ell)}, T ^ {(\\ell)}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\((\\lambda^{(\\ell)}, T^{(\\ell)})_{0 \\leq \\ell \\leq L-1}\\) is the vector of hyperparameters to be learned. This leads to the UNrolled EM (UNEM) model shown in Fig. 2, which depicts the inputs and outputs of a given layer \\(\\mathcal{L}^{(l)}\\) as well as its three main blocks for the update rules." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9669" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.212, + 0.089, + 0.784, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.242, + 0.661, + 0.255 + ], + "angle": 0, + "content": "Figure 3. Overall architecture of the designed UNEM." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.282, + 0.483, + 0.417 + ], + "angle": 0, + "content": "Learned hyper-parameters: Instead of a restricted number of handcrafted hyper-parameters (as often used in iterative algorithms), our unrolled algorithm offers more flexibility, enabling adaptation of the hyper-parameters along the processing workflow. This means that \\( L \\) vectors of hyper-parameters \\( (\\lambda^{(\\ell)}, T^{(\\ell)})_{0 \\leq \\ell \\leq L - 1} \\) could be learned and applied at each layer of the network. During the training of the unrolled model, we fulfill some design constraints, as described in more details in the following." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.418, + 0.484, + 0.479 + ], + "angle": 0, + "content": "More specifically, to ensure the non-negativity of hyperparameter \\(\\lambda^{(\\ell)}\\), we propose to express it as a Softplus function, which could be seen as a smooth approximation of the RELU activation function, yielding" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.486, + 0.482, + 0.505 + ], + "angle": 0, + "content": "\\[\n\\lambda^ {(\\ell)} = \\operatorname {S o f t p l u s} \\left(a ^ {(\\ell)}\\right) = \\log \\left(1 + \\exp \\left(a ^ {(\\ell)}\\right)\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.513, + 0.483, + 0.542 + ], + "angle": 0, + "content": "where \\(a^{(\\ell)}\\) represents a learnable parameter for the unrolled architecture." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.543, + 0.484, + 0.65 + ], + "angle": 0, + "content": "Regarding temperature scaling \\( T^{(\\ell)} \\), we observed in our experiments that imposing the non-negativity constraint alone might be a source of instability, resulting in values very close to zero and vanishing gradient issues during the training of the unrolled architecture. To circumvent this problem, we impose a lower bound equal to 1 on the temperature scaling:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.657, + 0.483, + 0.677 + ], + "angle": 0, + "content": "\\[\nT ^ {(\\ell)} = 1 + \\operatorname {S o f t p l u s} (b ^ {(\\ell)}) = 1 + \\log (1 + \\exp (b ^ {(\\ell)})) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.685, + 0.483, + 0.715 + ], + "angle": 0, + "content": "where \\( b^{(\\ell)} \\) denotes a parameter that needs to be learned during the training of the unrolled model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.715, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Overall architecture and training approach: Based on the previous considerations, the overall architecture of UNEM can be summarized as the composition of the \\(L\\) layers \\(\\mathcal{L}^{(L - 1)}\\circ \\ldots \\circ \\mathcal{L}^{(0)}\\). This architecture, depicted in Figure 3, illustrates: (i) the required inputs \\(\\pmb{\\theta}_k^{(0)}\\) and \\(\\pmb{u}_n^{(0)}\\) as well as the feature vectors \\(\\pmb{z}_n\\) (which are fed in all layers), (ii) the cascaded layers with their associated vector of parameters \\(\\pmb{w} = (\\lambda^{(\\ell)},T^{(\\ell)})_{0\\leq \\ell \\leq L - 1}\\) satisfying the aforementioned constraints, and (iii) the key output \\(\\pmb{u}_n^{(L)}\\) representing the class assignment vectors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The resulting neural network architecture is trained by minimizing a standard cross-entropy loss on a validation" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.284, + 0.543, + 0.296 + ], + "angle": 0, + "content": "set:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.297, + 0.907, + 0.34 + ], + "angle": 0, + "content": "\\[\n\\mathrm {L} _ {\\mathrm {c}} (\\boldsymbol {w}) = \\sum_ {n \\in \\mathbb {Q}} \\sum_ {k = 1} ^ {K} y _ {n, k} \\log \\left(u _ {n, k} ^ {(L)}\\right). \\tag {11}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.355, + 0.646, + 0.372 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.907, + 0.487 + ], + "angle": 0, + "content": "The effectiveness of the proposed approach is validated in both vision-only and vision-language transductive few-shot learning settings. We will designate our proposed unrolled methods by UNEM-Gaussian and UNEM-Dirichlet, respectively. After describing the experimental settings, this section is structured into two main sections discussing our experiments in both evaluation scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.5, + 0.719, + 0.516 + ], + "angle": 0, + "content": "4.1. Experimental settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.72 + ], + "angle": 0, + "content": "Task generation: We adopt a realistic transductive few-shot evaluation protocol, which is in line with state-of-the-art protocols [29, 32, 33, 45, 50]. While \\(K\\) designates the total number of classes in the labeled support set \\(\\mathbb{S}\\), \\(K_{\\mathrm{eff}}\\) (with \\(K_{\\mathrm{eff}} \\ll K\\)) denotes the number of effective classes present in the unlabeled query set \\(\\mathbb{Q}\\). The classes in the query set remain undisclosed during inference, while randomly selecting \\(|\\mathbb{Q}|\\) samples. On the other hand, the support set is built by uniformly selecting \\(s\\) images from each of the \\(K\\) classes. In this paper, the few-shot tasks are performed with (i) 5, 10, and 20 shots for vision-only models, and (ii) 4 shots for vision-language models. Moreover, in both of our evaluation scenarios, we used \\(K_{\\mathrm{eff}} = 5\\) and \\(|\\mathbb{Q}| = 75\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.855 + ], + "angle": 0, + "content": "Training specifications: For fair comparison and reproducibility purposes, the CLIP's pre-trained model has been directly used in the vision-language evaluation scenario, while for the vision-only scenario, the standard pre-trained ResNet-18 and WRN28-10 backbones have been finetuned. Their training is performed on the base classes set of each vision-only dataset, using cross-entropy loss with label smoothing set to 0.9 for 90 epochs. The learning rate is set to 0.1 and then decayed by a factor of 10." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "On the other hand, the training of our unrolled architecture, composed of 10 layers (i.e., \\( L = 10 \\)), is performed on several tasks sampled from the validation set of each dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9670" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.233, + 0.088, + 0.768, + 0.384 + ], + "angle": 0, + "content": "
MethodBackbonemini-ImageNet (K=20)tiered-ImageNet (K=160)
5-shot10-shot20-shot5-shot10-shot20-shot
Baseline [8]55.462.167.929.736.342.2
LR+ICI [52]55.462.168.1---
BD-CSPN [28]49.854.656.511.411.011.7
PT-MAP [17]25.727.228.45.26.06.6
LaplacianShot [60]ResNet-1857.964.268.329.635.439.1
TIM [4]66.869.970.829.328.727.8
α-TIM [50]66.771.073.943.848.351.9
α-AM [24]64.467.870.1---
PADDLE [32]62.973.579.845.461.470.6
UNEM-Gaussian66.475.680.452.365.773.2
Baseline [8]59.065.772.131.939.045.6
LR+ICI [52]58.865.772.0---
BD-CSPN [28]51.155.558.418.018.418.2
PT-MAP [17]26.528.029.35.26.06.6
LaplacianShot [60]WRN28-1061.066.871.031.437.341.5
TIM [4]72.174.976.236.139.038.5
α-TIM [50]71.575.278.345.851.455.2
α-AM [24]68.271.373.3---
PADDLE [32]62.673.079.243.959.469.9
UNEM-Gaussian71.679.283.754.166.874.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.394, + 0.907, + 0.422 + ], + "angle": 0, + "content": "Table 1. Comparison of the proposed UNEM-Gaussian with respect to state-of-the-art methods on mini-Imagenet and tiered-Imagenet. The metric is accuracy (in percentage). Results are averaged across 1,000 tasks. Results marked with ’ - ’ were intractable to obtain." + }, + { + "type": "table", + "bbox": [ + 0.146, + 0.445, + 0.432, + 0.598 + ], + "angle": 0, + "content": "
MethodCUB (K=50)
5-shot10-shot20-shot
Baseline [8]58.668.878.2
LR+ICI [52]49.955.658.0
PT-MAP [17]12.814.014.9
LaplacianShot [60]58.866.571.0
TIM [4]68.168.969.3
α-TIM [50]74.379.383.6
α-AM [24]66.268.969.8
PADDLE [32]71.281.886.8
UNEM-Gaussian78.585.388.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.607, + 0.483, + 0.649 + ], + "angle": 0, + "content": "Table 2. Comparison of the proposed UNEM-Gaussian with respect to state-of-the-art methods on CUB. The metric is accuracy (in percentage). Results are averaged across 1,000 tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.901 + ], + "angle": 0, + "content": "This implies that the proposed solution is highly economical in terms of parameters, typically requiring only a few tenths, compared to standard neural network architectures. Let us recall that the validation set is often used to select the hyperparameters in recent state-of-the-art transductive methods as mentioned in Section 2. Specifically, for the vision-only (resp. vision-language) models, the training of the unrolled architecture is carried out on 1,000 (resp. 100) tasks, while using an initial learning rate of 0.1 (resp. 0.5) with a decay factor of 0.5, a number of epochs equals to 80, and ADAM as optimizer. The trainable hyper-parameters are initialized to the default values set in [32] (i.e., \\(\\lambda = |Q|\\) for the vision-only models) and [33] (i.e., \\(\\lambda = \\frac{K}{K_{\\mathrm{eff}}}\\) for the vision-language models, with \\(T = 1\\)). Our architecture is" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.449, + 0.907, + 0.479 + ], + "angle": 0, + "content": "implemented in Pytorch (version 2.3.0) and run on NVIDIA QUADRO RTX8000 (with 48 GB of memory)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.486, + 0.907, + 0.503 + ], + "angle": 0, + "content": "4.2. UNEM-Gaussian in vision-only few-shot setting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.906, + 0.583 + ], + "angle": 0, + "content": "Let us recall that the Gaussian distribution is commonly used in standard clustering and transductive few-shot methods applied to vision-only models [32, 46, 50]. For this reason, we will focus here on the proposed UNEM-Gaussian approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.795 + ], + "angle": 0, + "content": "Datasets: The first UNEM-Gaussian architecture is evaluated on the following standard few-shot benchmark datasets: mini-[ImageNet [51], tiered-[ImageNet [41], CUB [14]. Mini-imagenet has 100 classes split into 64 base classes, 16 validation classes and 20 test classes. The tiered-imagenet has 608 classes instead, from which we follow a standard split with 351 for base training, 97 for validation and 160 for testing. For CUB, we followed the split proposed by [8] which consists of 100 base classes, 50 validation classes and 50 test classes. For each dataset, the feature vectors \\( \\mathbf{z}_n \\) are extracted using the fine-tuned backbones, while applying a scaling parameter \\( T_{z} \\) to the model's output as described in Appendix B. This parameter has also been learned through the unrolling approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Results: The UNEM-Gaussian architecture has been compared to its original version PADDLE [32] as well as several state-of-the-art methods. Tables 1 and 2 presents the results, averaged over 1,000 tasks with 5, 10, and 20 shots. Several observations can be made. First, the proposed UNEM-Gaussian outperforms the original PADDLE algorithm [32] and the other state-of-the-art methods. The achieved gains" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.515, + 0.956 + ], + "angle": 0, + "content": "9671" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.145, + 0.09, + 0.855, + 0.243 + ], + "angle": 0, + "content": "
MethodFood101EuroSATDTDOrionPedsFlowers102Caltech101UCF101FGVC AircraftStanford CarsSUM397ImageNetAverage
Ind.Tip-Adapter [56]76.772.554.786.483.288.872.123.763.966.762.768.3
CoOp [57]76.363.252.286.281.087.767.022.261.363.459.965.5
Transd.BDSCPN [28]74.746.145.281.374.282.059.018.048.154.549.257.5
Laplacian Shot [60]76.653.052.688.485.586.867.022.260.463.856.364.8
α-TIM [50]66.146.145.387.179.183.359.420.453.453.442.757.8
PADDLE [32]71.845.950.084.782.381.963.721.356.160.652.160.9
EM-Dirichlet [33]88.750.862.692.591.390.176.124.973.580.978.473.6
UNEM-Dirichlet91.453.865.396.095.693.478.530.480.088.583.177.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.251, + 0.905, + 0.281 + ], + "angle": 0, + "content": "Table 3. Comparison of the proposed UNEM-Dirichlet with respect to the state-of-the-art methods on 11 different datasets. The metric is accuracy (in percentage). Results are averaged across 1,000 tasks." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.308, + 0.485, + 0.519 + ], + "angle": 0, + "content": "are much higher with fewer number of shots and reach \\(3.5\\%\\) with mini-ImageNet, \\(6.9\\%\\) with tiered-ImageNet, and \\(7.3\\%\\) with CUB, while using ResNet-18 as backbone. Moreover, when WRN28-10 is used as backbone, it can be noticed that the performance of all methods has been improved, except for PADDLE, which shows results similar to those obtained with ResNet-18 pre-training. This can be explained by the fact that the class-balance parameter \\(\\lambda\\) was set in [32] to \\(|\\mathbb{Q}|\\) (i.e. 75), which becomes sub-optimal with WRN28-10 features. Most importantly, by learning efficiently the hyperparameters, our UNEM-Gaussian achieves higher gain with respect to its original version, yielding an accuracy gain reaching up to \\(10\\%\\) in 5-shot scenario, when WRN28-10 is used as backbone." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.53, + 0.402, + 0.545 + ], + "angle": 0, + "content": "4.3. UNEM-Dirichlet for few-shot CLIP" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.483, + 0.629 + ], + "angle": 0, + "content": "Dirichlet distribution has recently demonstrated good modelling performance in the context of transductive few-shot for vision-language models such as CLIP [33]. Thus, we will focus in this second round of experiments on UNEM-Dirichlet approach." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Datasets: The second unrolled architecture, designated by UNEM-Dirichlet, is assessed on 11 benchmark datasets that are commonly used for CLIP scenario: Caltech101 [10], OxfordPnts [37], StanfordCars [22], Flowers102 [36], Food101 [3], FGVCAircraft [31], SUN397 [54], DTD [9], EuroSAT [15], UCF101 [47], and ImageNet [42]. These datasets cover diverse classification challenges, from object recognition (Caltech101, Food101) to fine-grained tasks (StanfordCars, FGVCAircraft) and scene understanding (SUN397, EuroSAT). For each dataset, the vision-text feature vectors \\( z_{n} \\) are extracted using the CLIP's pre-trained model, while applying a temperature parameter \\( T_{z} \\) as described in Appendix C. This temperature parameter has also been learned through the unrolling approach." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Results: The proposed UNEM-Dirichlet has also been compared to its original EM-Dirichlet version [33] as well as different few-shot classification methods. Table 3 depicts the accuracy results evaluated over 1,000 tasks with 4-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.308, + 0.908, + 0.581 + ], + "angle": 0, + "content": "shots. Thus, it can be noticed that UNEM-Dirichlet outperforms state-of-the-art methods for most datasets. In particular, unrolling the recent EM-Dirichlet algorithm yields further improvement of about \\(4.2\\%\\) in average. The achieved gain is more significant (reaching up to \\(7.5\\%\\)) with more challenging datasets having a large number of classes like FGVC Aircraft, Stanford Cars, and SUN397. This is due to the inappropriate choice of the hyperparameter \\(\\lambda\\) for these datasets. Indeed, the selected \\(\\lambda\\) parameter was set to \\(\\frac{K}{K_{\\mathrm{eff}}} |\\mathbb{Q}|\\) in EM-Dirichlet approach [33]. Thus, by computing the numerical values of the hyperparameter for each of the aforementioned datasets, it can be deduced from Figure 1 that the difference between the selected value (which is 3,000 for Stanford Cars) and the optimal one (which is around 5,500 for Stanford Cars) is more important and impacts significantly the accuracy performance. This problem is well addressed by resorting to the proposed unrolled model for hyperparameters optimization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.907, + 0.643 + ], + "angle": 0, + "content": "Appendix D includes additional results to illustrate, in both scenarios, the effects of the temperature parameter in this framework, and to show the benefits of learning variable hyperparameters across the architecture layers." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.662, + 0.634, + 0.679 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.9 + ], + "angle": 0, + "content": "In this paper, we focus on transductive few-shot methods grounded on a generalized form of the EM algorithm. These methods include a parameter to control class balance. The proposed GEM algorithm, applicable to any mixture of distributions, incorporates also a temperature scaling parameter. The optimization process is then unrolled into a neural network architecture, enabling efficient learning of the introduced hyper-parameters. The results demonstrate the effectiveness of the proposed approach on both vision-only and vision-language models. In future work, it would be valuable to explore the potential of unrolling techniques across a broader range of computer vision tasks, especially with the recent rise of foundational vision-language models." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9672" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Christina Baek, Yiding Jiang, Aditi Raghunathan, and Zico Kolter. Agreement-on-the-line: Predicting the performance of neural networks under distribution shift. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), pages 19274-19289, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.257 + ], + "angle": 0, + "content": "[2] C. Bertocchi, E. Chouzenoux, M.-C. Corbineau, J.-C. Pesquet, and M. Prato. Deep unfolding of a proximal interior point method for image restoration. Inverse Problems, Special Issue on Variational Methods and Effective Algorithms for Imaging and Vision, 36(3):1-27, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.258, + 0.483, + 0.313 + ], + "angle": 0, + "content": "[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In European Conference on Computer Vision, pages 446-461, Zurich, Switzerland, 2014. Springer. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.315, + 0.483, + 0.384 + ], + "angle": 0, + "content": "[4] Malik Boudiaf, Ziko Imtiaz Masud, Jérôme Rony, José Dolz, Pablo Piantanida, and Ismail Ben Ayed. Transductive information maximization for few-shot learning. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), pages 2445–2457, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.385, + 0.483, + 0.442 + ], + "angle": 0, + "content": "[5] Y. Boykov, H. N. Isack, C. Olsson, and I. Ben Ayed. Volumetric bias in segmentation and reconstruction: Secrets and solutions. In IEEE/CVF International Conference on Computer Vision (ICCV), 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.444, + 0.483, + 0.499 + ], + "angle": 0, + "content": "[6] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Prompt learning with optimal transport for vision-language models. In International Conference on Learning Representations (ICLR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.501, + 0.483, + 0.555 + ], + "angle": 0, + "content": "[7] Tianlong Chen, Xiaohan Chen, Wuyang Chen, and Zhangyang Wang. Learning to optimize: A primer and a benchmark. Journal of Machine Learning Research (JMLR), 23:1-59, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.557, + 0.483, + 0.613 + ], + "angle": 0, + "content": "[8] Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, and Jia-Bin Huang. A closer look at few-shot classification. In International Conference on Learning Representations (ICLR), pages 1-19, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.615, + 0.483, + 0.669 + ], + "angle": 0, + "content": "[9] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, and A. Vedaldi. Describing textures in the wild. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.672, + 0.483, + 0.714 + ], + "angle": 0, + "content": "[10] Li Fei-Fei, Robert Fergus, and Pietro Perona. One-shot learning of object categories. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28(4):594-611, 2006. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.715, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[11] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International Conference on Machine Learning, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.759, + 0.483, + 0.827 + ], + "angle": 0, + "content": "[12] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. CLIP-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 32:581-595, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.829, + 0.483, + 0.884 + ], + "angle": 0, + "content": "[13] Bharath Hariharan and Ross B. Girshick. Low-shot visual recognition by shrinking and hallucinating features. In IEEE International Conference on Computer Vision (ICCV), pages 3018-3027, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.887, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Xiangteng He and Yuxin Peng. Fine-grained visual-textual" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.906, + 0.121 + ], + "angle": 0, + "content": "representation learning. IEEE Transactions on Circuits and Systems for Video Technology, 30(2):520-531, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.124, + 0.906, + 0.192 + ], + "angle": 0, + "content": "[15] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. EuroSAT: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.195, + 0.906, + 0.264 + ], + "angle": 0, + "content": "[16] Shell Xu Hu, Da Li, Jan Stuhmer, Minyoung Kim, and Timothy M Hospedales. Pushing the limits of simple pipelines for few-shot learning: External data and fine-tuning make a difference. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9068-9077, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.267, + 0.906, + 0.323 + ], + "angle": 0, + "content": "[17] Yuqing Hu, Vincent Gripon, and Stéphane Pateux. Leveraging the feature distribution in transfer-based few-shot learning. In International Conference on Artificial Neural Networks, pages 487-499, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.326, + 0.906, + 0.395 + ], + "angle": 0, + "content": "[18] Yunshi Huang, Fereshteh Shakeri, Jose Dolz, Malik Boudiaf, Houda Bahig, and Ismail Ben Ayed. \\(\\mathrm{LP} + +\\): A surprisingly strong linear probe for few-shot clip. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23773-23782, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.397, + 0.906, + 0.453 + ], + "angle": 0, + "content": "[19] Muhammad Abdullah Jamal and Guo-Jun Qi. Task agnostic meta-learning for few-shot learning. In IEEE conference on Computer Vision and Pattern Recognition (CVPR), pages 11719-11727, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.906, + 0.525 + ], + "angle": 0, + "content": "[20] Leonid Karlinsky, Joseph Shtok, Sivan Harary, Eli Schwartz, Amit Aides, and Rogerio Feris. Repmet: Representative-based metric learning for classification and few-shot object detection. In IEEE conference on Computer Vision and Pattern Recognition (CVPR), pages 5197-5206, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.527, + 0.906, + 0.582 + ], + "angle": 0, + "content": "[21] M. Kearns, Y. Mansour, and A. Ng. An information-theoretic analysis of hard and soft assignment methods for clustering. In Conference on Uncertainty in Artificial Intelligence (UAI), 1997. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.584, + 0.906, + 0.641 + ], + "angle": 0, + "content": "[22] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.906, + 0.699 + ], + "angle": 0, + "content": "[23] Michalis Lazarou, Tania Stathaki, and Yannis Avrithis. Iterative label cleaning for transductive and semi-supervised few-shot learning. In IEEE/CVF International Conference on Computer Vision, pages 8751-8760, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.701, + 0.906, + 0.758 + ], + "angle": 0, + "content": "[24] Michalis Lazarou, Yannis Avrithis, and Tania Stathaki. Adaptive manifold for imbalanced transductive few-shot learning. In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), page 2297-2306, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.906, + 0.841 + ], + "angle": 0, + "content": "[25] Gao Yu Lee, Tanmoy Dam, Daniel Puiu Poenar, Vu N. Duong, and Md Meftahul Ferdaus. HELA-VFA: A hellinger distance-attention-based feature aggregation network for few-shot classification. In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2173–2183, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[26] Xiaoxu Li, Xiaochen Yang, Zhanyu Ma, and Jing-Hao Xue. Deep metric learning for few-shot image classification: A review of recent developments. Pattern Recognition, 138: 109381, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9673" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[27] Moshe Lichtenstein, Prasanna Sattigeri, Rogerio Feris, Raja Giryes, and Leonid Karlinsky. TAFSSL: Task-adaptive feature sub-space learning for few-shot classification. In European Conference on Computer Vision (ECCV), pages 22-539, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.485, + 0.207 + ], + "angle": 0, + "content": "[28] Jinlu Liu, Liang Song, and Yongqiang Qin. Prototype rectification for few-shot learning. In European Conference on Computer Vision (ECCV), pages 741-756, 2020. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.276 + ], + "angle": 0, + "content": "[29] Yanbin Liu, Juho Lee, Minseop Park, Saehoon Kim, Eunho Yang, Sung Ju Hwang, and Yi Yang. Learning to propagate labels: Transductive propagation network for few-shot learning. In International Conference on Learning Representations, page 34677-34688, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.484, + 0.332 + ], + "angle": 0, + "content": "[30] Xu Luo, Hao Wu, Ji Zhang, Lianli Gao, Jing Xu, and Jingkuan Song. A closer look at few-shot classification again. In International Conference on Machine Learning (ICML), pages 23103-23123, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.484, + 0.389 + ], + "angle": 0, + "content": "[31] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. Technical report, Oxford University, 2013. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.484, + 0.475 + ], + "angle": 0, + "content": "[32] Ségolène Martin, Malik Boudiaf, Emilie Chouzenoux, Jean-Christophe Pesquet, and Ismail Ben Ayed. Towards practical few-shot query sets: Transductive minimum description length inference. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), page 34677–34688, 2022. 2, 3, 4, 6, 7, 8, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.484, + 0.546 + ], + "angle": 0, + "content": "[33] Ségolène Martin, Yunshi Huang, Fereshteh Shakeri, Jean-Christophe Pesquet, and Ismail Ben Ayed. Transductive zero-shot and few-shot clip. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28816-28826, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.484, + 0.602 + ], + "angle": 0, + "content": "[34] Paul Micaelli and Amos J Storkey. Gradient-based hyperparameter optimization over long horizons. In Advances in Neural Information Processing Systems, pages 10798-10809. Curran Associates, Inc., 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.603, + 0.484, + 0.659 + ], + "angle": 0, + "content": "[35] Vishal Monga, Yuelong Li, and Yonina C. Eldar. Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing. IEEE Signal Processing Magazine, 38(2):18-44, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.66, + 0.484, + 0.717 + ], + "angle": 0, + "content": "[36] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.484, + 0.761 + ], + "angle": 0, + "content": "[37] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.484, + 0.816 + ], + "angle": 0, + "content": "[38] Hang Qi, Matthew Brown, and David G Lowe. Low-shot learning with imprinted weights. In IEEE conference on Computer Vision and Pattern Recognition (CVPR), pages 5822-5830, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, and et al. Jack Clark. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.135 + ], + "angle": 0, + "content": "[40] Sachin Ravi and H. Larochelle. Optimization as a model for few-shot learning. In International Conference on Learning Representations, pages 1-11, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.205 + ], + "angle": 0, + "content": "[41] Mengye Ren, Eleni Triantafillou, Sachin Ravi, Jake Snell, Kevin Swersky, Joshua B. Tenenbaum, Hugo Larochelle, and Richard S. Zemel. Meta-learning for semi-supervised few-shot classification. In International Conference on Learning Representations (ICLR), pages 1-15, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.906, + 0.276 + ], + "angle": 0, + "content": "[42] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. Int. J. Comput. Vision, 115(3):211-252, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.906, + 0.348 + ], + "angle": 0, + "content": "[43] Tyler Scott, Karl Ridgeway, and Michael C. Mozer. Adapted deep embeddings: A synthesis of methods for k-shot inductive transfer learning. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), pages 76-85, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.348, + 0.906, + 0.39 + ], + "angle": 0, + "content": "[44] Anuj Singh and Hadi Jamali-Rad. Transductive decoupled variational inference for few-shot classification. Transactions on Machine Learning Research (TMLR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.906, + 0.448 + ], + "angle": 0, + "content": "[45] Jake Snell, Kevin Swersky, and Richard S. Zemel. Prototypical networks for few-shot learning. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), pages 4080-4090, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.906, + 0.49 + ], + "angle": 0, + "content": "[46] Yu Song and Changsheng Chen. MPPCANet: A feedforward learning strategy for few-shot image classification. Pattern Recognition, 113:107792, 2021. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.906, + 0.56 + ], + "angle": 0, + "content": "[47] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. Technical report, Center for Research in Computer Vision, University of Central Florida, 2012. arXiv:1212.0402. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.906, + 0.632 + ], + "angle": 0, + "content": "[48] Ran Tao, Hao Chen, and Marios Savvides. Boosting transductive few-shot fine-tuning with margin-based uncertainty weighting and probability regularization. In IEEE conference on Computer Vision and Pattern Recognition (CVPR), pages 15752-15761, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.906, + 0.702 + ], + "angle": 0, + "content": "[49] Long Tian, Jingyi Feng, Xiaoqiang Chai, Wenchao Chen, Liming Wang, Xiyang Liu, and Bo Chen. Prototypes-oriented transductive few-shot learning with conditional transport. In International Conference on Computer Vision (ICCV), pages 16317-16326, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.906, + 0.773 + ], + "angle": 0, + "content": "[50] Olivier Veilleux, Malik Boudiaf, Pablo Piantanida, and Ismail Ben Ayed. Realistic evaluation of transductive few-shot learning. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), pages 9290-9302, 2021. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.906, + 0.845 + ], + "angle": 0, + "content": "[51] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), pages 3637-3645, 2016. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[52] Yikai Wang, Chengming Xu, Chen Liu, Li Zhang, and Yanwei Fu. Instance credibility inference for few-shot learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12833-12842, 2020. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "9674" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[53] Yimu Wang, Yihan Wu, and Hongyang Zhang. Lost domain generalization is a natural consequence of lack of training domains. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 15689-15697, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[54] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3485-3492, 2010. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[55] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6757-6767, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.345 + ], + "angle": 0, + "content": "[56] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. TIP-adapter: Training-free adaption of clip for few-shot classification. In European Conference on Computer Vision (ECCV), pages 493-510, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.347, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[57] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16816-16825, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.417, + 0.482, + 0.471 + ], + "angle": 0, + "content": "[58] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In International Conference on Computer Vision (ICCV), pages 15659-15669, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.474, + 0.482, + 0.529 + ], + "angle": 0, + "content": "[59] Hao Zhu and Piotr Koniusz. Transductive few-shot learning with prototype-based label propagation by iterative graph refinement. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 23996-24006, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.531, + 0.482, + 0.584 + ], + "angle": 0, + "content": "[60] Imtiaz Masud Ziko, Jose Dolz, Eric Granger, and Ismail Ben Ayed. Laplacian regularized few-shot learning. In International Conference on Machine Learning (ICML), 2020. 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9675" + } + ] +] \ No newline at end of file diff --git a/2025/UNEM_ UNrolled Generalized EM for Transductive Few-Shot Learning/592279b8-3f9e-4b39-91e5-638d8677cb8a_origin.pdf b/2025/UNEM_ UNrolled Generalized EM for Transductive Few-Shot Learning/592279b8-3f9e-4b39-91e5-638d8677cb8a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..feec3c45c23dac8724ce9a3658739a86f36325c4 --- /dev/null +++ b/2025/UNEM_ UNrolled Generalized EM for Transductive Few-Shot Learning/592279b8-3f9e-4b39-91e5-638d8677cb8a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac3b0aeda0c3a2825852a8b0b11b6c92d5eede76df5b4fe2f2228e6bfd687505 +size 1297475 diff --git a/2025/UNEM_ UNrolled Generalized EM for Transductive Few-Shot Learning/full.md b/2025/UNEM_ UNrolled Generalized EM for Transductive Few-Shot Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1486b14a4070ede24d696d9381cc38e709fc6635 --- /dev/null +++ b/2025/UNEM_ UNrolled Generalized EM for Transductive Few-Shot Learning/full.md @@ -0,0 +1,327 @@ +# UNEM: UNrolled Generalized EM for Transductive Few-Shot Learning + +Long Zhou $^{1,3}$ , Fereshteh Shakeri $^{2}$ , Aymen Sadraoui $^{3}$ , Mounir Kaaniche $^{3,4}$ , Jean-Christophe Pesquet $^{3}$ , Ismail Ben Ayed $^{2}$ + +1Politecnico di Milano + +2ÉTS Montréal + +3Université Paris-Saclay, Inria, CentraleSupélec, CVN + +4Université Sorbonne Paris Nord, L2TI, UR 3043 + +# Abstract + +Transductive few-shot learning has recently triggered wide attention in computer vision. Yet, current methods introduce key hyper-parameters, which control the prediction statistics of the test batches, such as the level of class balance, affecting performances significantly. Such hyper-parameters are empirically grid-searched over validation data, and their configurations may vary substantially with the target dataset and pre-training model, making such empirical searches both sub-optimal and computationally intractable. In this work, we advocate and introduce the unrolling paradigm, also referred to as "learning to optimize", in the context of few-shot learning, thereby learning efficiently and effectively a set of optimized hyperparameters. Specifically, we unroll a generalization of the ubiquitous Expectation-Maximization (EM) optimizer into a neural network architecture, mapping each of its iterates to a layer and learning a set of key hyper-parameters over validation data. Our unrolling approach covers various statistical feature distributions and pre-training paradigms, including recent foundational vision-language models and standard vision-only classifiers. We report comprehensive experiments, which cover a breadth of fine-grained downstream image classification tasks, showing significant gains brought by the proposed unrolled EM algorithm over iterative variants. The achieved improvements reach up to $10\%$ and $7.5\%$ on vision-only and vision-language benchmarks, respectively. The source code and learned parameters are available at https://github.com/ZhouLong0/UNEM-Transductive. + +![](images/ceb9a74013acf42b64735a8a9a5a565d872d5c29145950cf8386ee1d954da2df.jpg) +Figure 1. Impact of the class-balance hyperparameter $\lambda$ on the accuracy of transductive few-shot classification. The accuracy results are obtained using the EM-Dirichlet algorithm [33] applied to vision-language models (with 4-shots). The plot shows that the choice of $\lambda$ has a strong impact on the performance, and that the optimal $\lambda$ (indicated with the star symbol) might vary by orders of magnitudes, depending on the target downstream dataset (e.g., the ten very different fine-grained classification datasets). Further comments on the optimal $\lambda$ values and the values chosen in [33] are provided in Section 4. The values of the learned hyper-parameters based on the proposed unrolled algorithm are illustrated and analyzed in Appendix D. + +# 1. Introduction + +Deep learning has transfigured computer vision, driving substantial progress in tasks such as image classification, captioning, object detection and segmentation. However, these successes often come with a high cost: the requirement for large amounts of labeled data. Additionally, the generalization of these models is seriously challenged when evaluated on new classes (concepts), unseen during pre + +training, or when operating under distribution shifts [1, 53]. + +To address these challenges, Few-Shot Learning (FSL) has recently attracted wide attention within the computer vision community. In the standard few-shot image classification setting, a feature extractor is first pre-trained on a set of classes, often called the base classes. Then, the model is adapted and evaluated on new classes and tasks. Each evaluation task is performed on a set of unlabeled samples (referred to as the query set), and supervised by a support set composed of few labeled samples per new class. Earlier FSL methods have relied on various concepts, including meta-learning [11, 19, 40], transfer learning [13, 38, 43], and metric learning [20, 26, 51]. However, most of these works operate within the inductive setting, where, at inference time, the prediction for each sample is made independently from the other samples in the query set. + +Recently, significant attention has shifted to the transductive scheme, in which inference is performed jointly on a batch of query samples. Transduction leverages the statistics of the unlabeled query samples, yielding notable performance gains, and have triggered a large body of recent works in few-shot learning, with various methods and mechanisms [4, 23-25, 29, 32, 59], as detailed in Section 2. Generally, transductive methods are built upon optimizing objective functions that integrate clustering terms, either discriminative as in information maximization [4, 50] or generative as in probabilistic K-means [32, 33]. However, many among these assume a perfectly balanced class distribution within the query set, and incorporate terms in the objective function or constraints that enforce a class-balance prior. The latter could limit the applicability of these methods, whose performances were shown to drop significantly when dealing with imbalanced query sets [27, 50]. + +Several recent attempts tackled this limitation, to handle more realistic scenarios [24, 32, 33, 48, 50]. Yet, these methods introduce key hyper-parameters, which control the prediction statistics of the unlabeled query set, such as the level of class balance. Such hyper-parameters are empirically grid-searched over pre-defined sets of values using validation data, and their optimal configurations may vary substantially with the target dataset and pre-training model [17]. To illustrate this, we depict in Fig. 1 the accuracy as a function of a class-balance hyper-parameter for the recent method in [33]. One may observe that this hyperparameter has a crucial effect on the performance, and its optimal value might vary by orders of magnitude across the target datasets. The issue is further compounded by additional hyper-parameters, which makes intensive empirical grid searches for the hyper-parameters over validation data and pre-defined intervals of values both sub-optimal and computationally intractable. Therefore, we advocate and introduce the unrolling paradigm (also called "learning to optimize") in the context of few-shot learning, which enables + +to learn efficiently and effectively a set of optimized hyperparameters. Specifically, our main contributions could be summarized as follows: + +1. We study a generalization of the ubiquitous Expectation-Maximization (EM) algorithm, in which we make two hyper-parameters controlling prediction statistics - class balance and prediction entropy - explicit1. Our generalization encompasses several existing transductive few-shot methods as particular cases and, more importantly, enables to learn these crucial hyper-parameters through the proposed unrolling strategy. +2. We unroll the generalized EM optimizer into a neural network architecture, mapping each of its iterations to a layer and learning the introduced hyper-parameters over validation data. Our unrolling approach offers greater flexibility in optimizing the hyper-parameters, allowing them to vary across the network's layers. To the best of our knowledge, this study is the first to investigate unrolling in transductive few-shot learning. +3. We design our unrolling architecture in way that covers various statistical assumptions and pre-training paradigms: (i) Gaussian for vision-only classifiers and (ii) Dirichlet for vision-language models. +4. We report comprehensive experiments, which cover a breadth of fine-grained downstream image classification tasks and different pre-training models, showing substantial gains over recent state-of-the-art methods. + +# 2. Related works + +Few-shot classification with vision-only models: Few-shot classification using vision models has been widely explored in the literature, leading to the development of various methods. The first category, inductive methods, predicts the class of each test sample (in the query set) independently from others [16, 30]. In contrast, the second category, transductive methods, which has gained increasing attention in recent years, involves jointly predicting classes for a batch of test samples in each few-shot task. Numerous research efforts in transductive methods have leveraged concepts like clustering [24, 32], label propagation [29, 59], information maximization [4, 50], optimal transport [23, 49], prototype estimation [28, 59], and variational networks [25, 44], among others. Studies have shown that transductive methods significantly outperform inductive approaches, achieving accuracy gains of up to $15\%$ as reported in several evaluations [4, 32]. + +Few-shot classification with vision-language models: Contrastive Vision-Language Pre-training (CLIP) has recently emerged as an effective model for enhancing various vision tasks through visual-text pairs. By learning transferable visual models with natural language supervision, CLIP + +demonstrates strong performance in zero-shot classification by matching the image features to text embeddings of novel classes [39]. To further enhance its classification capabilities, several studies have extended CLIP to the few-shot setting [6, 18, 55, 57, 58]. While the linear probe model [39] trains a logistic regression classifier using CLIP image features, the authors of [18] proposed a generalization of this baseline, modeling the classifier weights as learnable functions of the visual prototypes and text embeddings. Context Optimization (CoOp) [57] and its extended versions [6, 55, 58] have been developed based on the concept of prompt learning. While CoOp [57] aims to model context in prompts using continuous representations, a knowledge-guided CoOp is proposed in [55] to enhance the generalization ability by minimizing the discrepancy between the learnable prompts and the original ones. In [6], the authors learn multiple prompts, describing the characteristics of each class, through the minimization of an optimal-transport distance. Unlike prompt learning methods, which fine-tune the input text, another family of methods, referred to as adapters, aims to transform the visual or text encoders, such as [12, 56]. For instance, TIP-Adapter [56] adds a nonparametric adapter to the weight-frozen CLIP model, and updates the prior knowledge encoded in CLIP by feature retrieval. CLIP-Adapter [12] introduces a multi-layered perceptron to learn new features, and combine them with the original CLIP-encoded features via residual connections. It is worth noting that all the aforementioned methods belong to the inductive family. However, unlike vision-only models, the transductive few-shot setting is still not well investigated in the context of CLIP. To the best of our knowledge, the only transductive few-shot CLIP method is the one very recently proposed in [33], which, inspired by the Expectation Maximization algorithm, relies on the Dirichlet distribution to model the data. Moreover, as reported in [33], and unlike the behaviors observed with vision-only models, recent transductive few-shot methods do not always outperform their inductive counterparts with CLIP. This suggests that there is a need to further investigate transductive methods in the context of vision-language models, due to their aforementioned advantages with respect to inductive methods. + +Class-balance and hyperparameters setting: Hyperparameter optimization has a large impact on the performance of a given iterative algorithm or learning model, and is generally tackled through greedy or gradient-based methods [34]. In the context of transductive few-shot learning, most of the existing methods have been designed for perfectly balanced query sets (i.e., uniform class distribution), and have shown drops in performances under class-imbalanced settings. To address this flaw, various methods have been developed in the last years [24, 32, 33, 48, 50]. For instance, in [48] the categorical probability of each query + +sample is regularized to quantify the difference between the class marginal distribution and the uniform one. The works in [24, 32, 33, 50] explored various weighted terms, which are added to the objective functions, to mitigate the effect of the class-balance bias. To this end, some hyper-parameters are introduced, to weigh to contributions of such terms in the objective function. Unlike [32], where the introduced hyper-parameter has been theoretically chosen in order to compensate for the hidden class-balance bias in the used clustering objective, the hyper-parameters are often set in an empirical manner based on the validation classes of each dataset. Thus, a set of hyperparameter values (in a given range) are evaluated, and the ones yielding the highest accuracy are selected. + +# 3. Proposed methodology + +# 3.1. Generalized EM algorithm + +Preliminaries: Let us first introduce the notations to formulate our transductive few-shot inference. Let $\{\pmb{z}_n\}_{1\leq n\leq N}$ denote the set of feature vectors extracted from a pretraining network, and which are to be classified, with $N$ the total number of samples within a given task. The whole dataset contains $K$ distinct classes, whereas the number of randomly sampled classes present in each mini-batch task might be much smaller than $K$ . This subset of sampled classes may also vary across mini-batches, and the method does not assume any prior knowledge on the specific classes that may appear in each mini-batch. + +Moreover, for a given few-shot task, let us denote by $\mathbb{S} \subset \{1, \dots, N\}$ and $\mathbb{Q} = \{1, \dots, N\} \setminus \mathbb{S}$ the indices of samples within the support (labeled) and query (unlabeled) mini-batch sets, respectively. For every $n \in \mathbb{S}$ , $\pmb{y}_n = (y_{n,k})_{1 \leq k \leq K} \in \{0, 1\}^K$ are the one-hot-encoded labels, such that, for every $k \in \{1, \dots, K\}$ , $y_{n,k} = 1$ if the $n$ -th sample belongs to class $k$ , and $y_{n,k} = 0$ otherwise. + +Finally, given the extracted feature vectors $\mathbf{z}_n$ , we assume that the data probability distribution knowing its class $k$ is modeled by a given law, whose probability density function (pdf) is denoted by $\mathrm{p}(\mathbf{z}_n \mid \boldsymbol{\theta}_k)$ and characterized by a vector of parameters $\boldsymbol{\theta}_k$ . This means that the global distribution of $\mathbf{z}_n$ is a mixture of these pdfs. + +Problem formulation: Our goal is to identify the classes of the unlabeled samples in the query set by optimizing a general clustering objective function, while embedding supervision constraints from the few labeled samples within the support set. We do so by unrolling iterative block-coordinate optimizers of the objective functions over two sets of variables: + +- Soft assignment vectors $\pmb{u} = (u_n)_{1\leq n\leq N}\in (\Delta_K)^N$ where $\Delta_K$ is the probability simplex of $\mathbb{R}^K$ . For every $n\in \{1,\dots ,N\}$ , $\pmb{u}_n = (u_{n,k})_{1\leq k\leq K}$ where, for every $k\in \{1,\ldots ,K\}$ , $u_{n,k}$ can be interpreted as the probabil + +ity that the $n$ -th sample belongs to class $k$ . These probabilities have to be determined for the query set samples. + +- Feature distribution parameters $\pmb{\theta} = (\pmb{\theta}_k)_{1\leq k\leq K}$ . + +Consider the following general probabilistic clustering problem: + +$$ +\underset {\boldsymbol {u}, \boldsymbol {\theta}} {\text {m i n i m i z e}} \mathcal {L} (\boldsymbol {u}, \boldsymbol {\theta}) + \lambda \Psi (\boldsymbol {u}) + T \Phi (\boldsymbol {u}), \tag {1} +$$ + +subject to $\pmb{u}_n\in \Delta_K\quad \forall n\in \mathbb{Q},$ + +$$ +u _ {n, k} = y _ {n, k} \quad \forall n \in \mathbb {S}, \forall k \in \{1, \dots , K \}. +$$ + +where weighing factor $\lambda$ and temperature scaling $T$ are learnable optimized hyper-parameters, which we will estimate through the proposed unrolling strategy (Section 3.2), and terms $\mathcal{L}$ , $\Psi$ and $\Phi$ are detailed in the following. + +- The first term in objective function (1) is the negative log-likelihood of the feature vectors: + +$$ +\mathcal {L} (\boldsymbol {u}, \boldsymbol {\theta}) = - \sum_ {n = 1} ^ {N} \sum_ {k = 1} ^ {K} u _ {n, k} \ln \left(\mathrm {p} \left(\boldsymbol {z} _ {n} \mid \boldsymbol {\theta} _ {k}\right)\right) \tag {2} +$$ + +This general log-likelihood model fitting term is well known in the context of clustering methods [5, 21]. In fact, it generalizes the standard $K$ -means clustering objective to arbitrary distributions2. It is well-known that minimizing (2) has an inherent bias towards class-balanced clustering [5, 21, 32, 33]. + +- The second term in (1) controls the partition complexity of the model, penalizing the number of non-empty clusters in the solution. It corresponds to the Shannon entropy of class distribution $(\pi_k)_{1 \leq k \leq K}$ , defined as + +$$ +\Psi (\boldsymbol {u}) = - \sum_ {k = 1} ^ {K} \pi_ {k} \ln \pi_ {k}, \tag {3} +$$ + +where $\pi_k = \frac{1}{|\mathbb{Q}|} \sum_{n \in \mathbb{Q}} u_{n,k}$ is the proportion of query samples within class $k$ , i.e., the empirical estimate of the marginal probability of class $k$ . + +Class-balance hyper-parameter $(\lambda)$ : The marginal entropy in (3) mitigates the class-balance bias of the log-likelihood clustering term in (2). It reaches its minimum for the extremely imbalanced solution in which all data samples are assigned to a single cluster, and its maximum for a perfectly balanced clustering. Hence, clearly, hyperparameter $\lambda$ controls the level of class balance in the solution. As depicted in Fig. 1, this hyper-parameter has a crucial effect on the performance, and its optimal value might vary by orders of magnitude from one dataset to another. This makes exhaustive grid searches for optimal $\lambda$ over validation sets intractable computationally, which motivates learning this hyper-parameter through our unrolling strategy, as described in Section 3.2. + +- The third term in (1) is an entropic barrier, enabling to soften assignments $u_{n,k}$ , while imposing a non-negativity constraint on each of them. It is given by + +$$ +\Phi (\boldsymbol {u}) = \sum_ {n = 1} ^ {N} \sum_ {k = 1} ^ {K} u _ {n, k} \ln u _ {n, k}. \tag {4} +$$ + +Temperature scaling hyper-parameter $(T)$ : Weighting factor $T$ in (1) controls the trade-off between the clustering term and the entropic barrier in (4), i.e., the level of softness of assignments $u_{n,k}$ . Therefore, this hyperparameter has an important effect on performances and on the marginal class probabilities appearing in the class-balance term in (3), which motivates learning it. As described in Section 3.2, our unrolling algorithm enables to cope with this extra parameter. + +Link to EM and other transductive few-shot methods: We examine solving problem (1) with an iterative block-coordinate descent algorithm (see Algorithm 1), which alternates two steps, one updating distribution parameters $\pmb{\theta}_k^{(\ell)}$ and the other optimizing over class assignments $\pmb{u}_n^{(\ell)}$ , at each iteration $\ell$ . For the $\pmb{u}$ -step, and due to the nonconvexity of $\Psi$ , we proceed with a Majorization-Minimization (MM) strategy to minimize a surrogate convex function with respect to $\pmb{u}$ at each iteration. In addition, if $\mathrm{p}(\cdot|\pmb{\theta}_k)$ is assumed to belong to the exponential family and $\pmb{\theta}_k$ are its canonical parameters, the estimation w.r.t. $\pmb{\theta}$ , with $\pmb{u}$ fixed, is a convex problem. We provide further details on this general iterative optimization scheme in Appendix A. It is worth noting that, when $T = 1$ and the data is modeled by the Dirichlet distribution, we recover the recent transductive few-shot learning algorithm in [33]. Interestingly, when $T = 1$ and $\lambda = |\mathbb{Q}|$ , we recover the well-known Expectation-Maximization (EM) algorithm for estimating the parameters of mixture of distributions; see Proposition 1 in [33]. Therefore, Algorithm 1 could be viewed as a generalized EM (GEM), in which hyper-parameters $\lambda$ and $T$ control the class balance as well as prediction softness, and could be made learnable. + +Examples of data models: While a broad range of distribution models in the exponential family could be adopted, we focus in this paper on two popular ones. + +- The first one is the Gaussian distribution, which is commonly used in standard clustering and transductive few-shot-methods applied to vision models [32, 46, 50]. By assuming a Gaussian distribution with mean $\pmb{\theta}_{k}$ and identity covariance matrix, the pdf $\mathrm{p}(\pmb{z}_n|\pmb{\theta}_k)$ reads as follows + +$$ +\mathrm {p} \left(\boldsymbol {z} _ {n} \mid \boldsymbol {\theta} _ {k}\right) \propto \exp \left(- \frac {1}{2} \left\| \boldsymbol {z} _ {n} - \boldsymbol {\theta} _ {k} \right\| ^ {2}\right). \tag {5} +$$ + +- The second one is the Dirichlet distribution, which has recently shown good modelling performance in the context of transductive few-shot for vision-language mod + +# Algorithm 1 GEM based few-shot classification algorithm + +Input: Compute $\mathbf{z}_n$ for the dataset samples, initialize $\mathbf{u}_n^{(0)}$ and $\theta_k^{(0)}$ , and fix the number of iterations $L$ , for $\ell = 0,1,\ldots ,L - 1$ do + +// Update Distribution parameters for each class using a given estimation algorithm (denoted here by "DP_est") + +$$ +\boldsymbol {\theta} _ {k} ^ {(\ell + 1)} = \operatorname {D P} _ {-} \operatorname {e s t} (\boldsymbol {u} _ {., k} ^ {(\ell)}, \boldsymbol {\theta} _ {k} ^ {(\ell)}), \quad \forall k \in \{1, \dots , K \}, +$$ + +// Update class proportions + +$$ +\pi_ {k} ^ {(\ell + 1)} = \frac {1}{| \mathbb {Q} |} \sum_ {n \in \mathbb {Q}} u _ {n, k} ^ {(\ell)}, \quad \forall k \in \{1, \dots , K \}, +$$ + +// Update assignment vectors for all query samples + +$$ +\boldsymbol {u} _ {n} ^ {(\ell + 1)} = \operatorname {s o f t m a x} \left(\frac {1}{T} \left(\ln \mathrm {p} \left(\boldsymbol {z} _ {n} \mid \boldsymbol {\theta} _ {k} ^ {(\ell + 1)}\right) + \frac {\lambda}{| \mathbb {Q} |} \ln (\pi_ {k} ^ {(\ell + 1)})\right) _ {k}\right), \quad \forall n \in \mathbb {Q}. +$$ + +# end for + +![](images/638c5b2d4c01adbc9b9d282a46011431ac1796f80a8748d1e7a7b18db5dcad27.jpg) +Figure 2. An overview of the unrolled GEM algorithm for a given iteration. Each iteration $\ell$ corresponds to a network layer $\mathcal{L}^{(\ell)}$ . Each layer depends on the vector of hyperparameters $(\lambda^{(\ell)}, T^{(\ell)})$ . + +els such as CLIP [33]. For $\mathbf{z}_n = (z_{n,i})_{1\leq i\leq K}$ and $\pmb{\theta}_{k} = (\theta_{k,i})_{1\leq i\leq K}$ , the associated pdf is given by + +$$ +\mathrm {p} \left(\boldsymbol {z} _ {n} \mid \boldsymbol {\theta} _ {k}\right) = \frac {1}{\mathcal {B} \left(\boldsymbol {\theta} _ {k}\right)} \prod_ {i = 1} ^ {K} z _ {n, i} ^ {\theta_ {k, i} - 1} \mathbb {1} _ {\boldsymbol {z} _ {n} \in \Delta_ {K}}, \tag {6} +$$ + +where the normalization factor $\mathcal{B}(\pmb{\theta}_k)$ is + +$$ +\mathcal {B} \left(\boldsymbol {\theta} _ {k}\right) = \frac {\prod_ {i = 1} ^ {K} \Gamma \left(\theta_ {k , i}\right)}{\Gamma \left(\sum_ {i = 1} ^ {K} \theta_ {k , i}\right)}, \tag {7} +$$ + +and $\Gamma$ denotes the Gamma function. + +The necessary details related to feature representations in the cases of vision-only and vision-language models, distribution parameter estimation, and the resulting GEM algorithms for Gaussian and Dirichlet laws are provided in Appendices B and C, respectively. + +# 3.2. UNrolled EM architecture (UNEM) + +Overview: As discussed in Section 2, the choice of the optimal hyper-parameter, which controls the level of class balance, is a difficult task, often performed manually in the existing transductive few-shot classification methods. In addition to the class-balance parameter $\lambda$ , Algorithm 1 involves an additional temperature parameter $T$ , which makes hyperparameter setting even more challenging. For this reason, + +we propose in this paper to resort to the unrolling (called also learning to optimize) paradigm [35], which enables to learn efficiently a set of optimized hyper-parameters. It is important to note that unrolling iterative optimization algorithms have found successful applications in diverse signal and image processing tasks [2, 7, 35]. However, to the best of our knowledge, this work is the first to leverage the unrolling paradigm for hyper-parameter optimization in a few-shot learning context. The main idea behind the learning-to-optimize paradigm is to map each iteration of a given optimization algorithm to a network layer, stack all layers together, and view the hyper-parameters to be optimized as the network's learnable parameters. More precisely, to unroll our generalized EM algorithm, the number of iterations $L$ is used as the number of layers for the neural network architecture. Thus, each iteration $\ell \in \{0,1,\dots ,L - 1\}$ is associated with a tailored layer $\mathcal{L}^{(\ell)}$ , which performs the update rules defined in Algorithm 1: + +$$ +\left(\boldsymbol {\theta} _ {k} ^ {(\ell + 1)}, \pi_ {k} ^ {(\ell + 1)}, \boldsymbol {u} _ {n} ^ {(\ell + 1)}\right) = \mathcal {L} ^ {(\ell)} \left(\boldsymbol {\theta} _ {k} ^ {(\ell)}, \pi_ {k} ^ {(\ell)}, \boldsymbol {u} _ {n} ^ {(\ell)}; \lambda^ {(\ell)}, T ^ {(\ell)}\right) \tag {8} +$$ + +where $(\lambda^{(\ell)}, T^{(\ell)})_{0 \leq \ell \leq L-1}$ is the vector of hyperparameters to be learned. This leads to the UNrolled EM (UNEM) model shown in Fig. 2, which depicts the inputs and outputs of a given layer $\mathcal{L}^{(l)}$ as well as its three main blocks for the update rules. + +![](images/702073904ba52b0407f7950ed818c705d2325a204fe9a923703181ba12013a5a.jpg) +Figure 3. Overall architecture of the designed UNEM. + +Learned hyper-parameters: Instead of a restricted number of handcrafted hyper-parameters (as often used in iterative algorithms), our unrolled algorithm offers more flexibility, enabling adaptation of the hyper-parameters along the processing workflow. This means that $L$ vectors of hyper-parameters $(\lambda^{(\ell)}, T^{(\ell)})_{0 \leq \ell \leq L - 1}$ could be learned and applied at each layer of the network. During the training of the unrolled model, we fulfill some design constraints, as described in more details in the following. + +More specifically, to ensure the non-negativity of hyperparameter $\lambda^{(\ell)}$ , we propose to express it as a Softplus function, which could be seen as a smooth approximation of the RELU activation function, yielding + +$$ +\lambda^ {(\ell)} = \operatorname {S o f t p l u s} \left(a ^ {(\ell)}\right) = \log \left(1 + \exp \left(a ^ {(\ell)}\right)\right) \tag {9} +$$ + +where $a^{(\ell)}$ represents a learnable parameter for the unrolled architecture. + +Regarding temperature scaling $T^{(\ell)}$ , we observed in our experiments that imposing the non-negativity constraint alone might be a source of instability, resulting in values very close to zero and vanishing gradient issues during the training of the unrolled architecture. To circumvent this problem, we impose a lower bound equal to 1 on the temperature scaling: + +$$ +T ^ {(\ell)} = 1 + \operatorname {S o f t p l u s} (b ^ {(\ell)}) = 1 + \log (1 + \exp (b ^ {(\ell)})) \tag {10} +$$ + +where $b^{(\ell)}$ denotes a parameter that needs to be learned during the training of the unrolled model. + +Overall architecture and training approach: Based on the previous considerations, the overall architecture of UNEM can be summarized as the composition of the $L$ layers $\mathcal{L}^{(L - 1)}\circ \ldots \circ \mathcal{L}^{(0)}$ . This architecture, depicted in Figure 3, illustrates: (i) the required inputs $\pmb{\theta}_k^{(0)}$ and $\pmb{u}_n^{(0)}$ as well as the feature vectors $\pmb{z}_n$ (which are fed in all layers), (ii) the cascaded layers with their associated vector of parameters $\pmb{w} = (\lambda^{(\ell)},T^{(\ell)})_{0\leq \ell \leq L - 1}$ satisfying the aforementioned constraints, and (iii) the key output $\pmb{u}_n^{(L)}$ representing the class assignment vectors. + +The resulting neural network architecture is trained by minimizing a standard cross-entropy loss on a validation + +set: + +$$ +\mathrm {L} _ {\mathrm {c}} (\boldsymbol {w}) = \sum_ {n \in \mathbb {Q}} \sum_ {k = 1} ^ {K} y _ {n, k} \log \left(u _ {n, k} ^ {(L)}\right). \tag {11} +$$ + +# 4. Experiments + +The effectiveness of the proposed approach is validated in both vision-only and vision-language transductive few-shot learning settings. We will designate our proposed unrolled methods by UNEM-Gaussian and UNEM-Dirichlet, respectively. After describing the experimental settings, this section is structured into two main sections discussing our experiments in both evaluation scenarios. + +# 4.1. Experimental settings + +Task generation: We adopt a realistic transductive few-shot evaluation protocol, which is in line with state-of-the-art protocols [29, 32, 33, 45, 50]. While $K$ designates the total number of classes in the labeled support set $\mathbb{S}$ , $K_{\mathrm{eff}}$ (with $K_{\mathrm{eff}} \ll K$ ) denotes the number of effective classes present in the unlabeled query set $\mathbb{Q}$ . The classes in the query set remain undisclosed during inference, while randomly selecting $|\mathbb{Q}|$ samples. On the other hand, the support set is built by uniformly selecting $s$ images from each of the $K$ classes. In this paper, the few-shot tasks are performed with (i) 5, 10, and 20 shots for vision-only models, and (ii) 4 shots for vision-language models. Moreover, in both of our evaluation scenarios, we used $K_{\mathrm{eff}} = 5$ and $|\mathbb{Q}| = 75$ . + +Training specifications: For fair comparison and reproducibility purposes, the CLIP's pre-trained model has been directly used in the vision-language evaluation scenario, while for the vision-only scenario, the standard pre-trained ResNet-18 and WRN28-10 backbones have been finetuned. Their training is performed on the base classes set of each vision-only dataset, using cross-entropy loss with label smoothing set to 0.9 for 90 epochs. The learning rate is set to 0.1 and then decayed by a factor of 10. + +On the other hand, the training of our unrolled architecture, composed of 10 layers (i.e., $L = 10$ ), is performed on several tasks sampled from the validation set of each dataset. + +
MethodBackbonemini-ImageNet (K=20)tiered-ImageNet (K=160)
5-shot10-shot20-shot5-shot10-shot20-shot
Baseline [8]55.462.167.929.736.342.2
LR+ICI [52]55.462.168.1---
BD-CSPN [28]49.854.656.511.411.011.7
PT-MAP [17]25.727.228.45.26.06.6
LaplacianShot [60]ResNet-1857.964.268.329.635.439.1
TIM [4]66.869.970.829.328.727.8
α-TIM [50]66.771.073.943.848.351.9
α-AM [24]64.467.870.1---
PADDLE [32]62.973.579.845.461.470.6
UNEM-Gaussian66.475.680.452.365.773.2
Baseline [8]59.065.772.131.939.045.6
LR+ICI [52]58.865.772.0---
BD-CSPN [28]51.155.558.418.018.418.2
PT-MAP [17]26.528.029.35.26.06.6
LaplacianShot [60]WRN28-1061.066.871.031.437.341.5
TIM [4]72.174.976.236.139.038.5
α-TIM [50]71.575.278.345.851.455.2
α-AM [24]68.271.373.3---
PADDLE [32]62.673.079.243.959.469.9
UNEM-Gaussian71.679.283.754.166.874.7
+ +Table 1. Comparison of the proposed UNEM-Gaussian with respect to state-of-the-art methods on mini-Imagenet and tiered-Imagenet. The metric is accuracy (in percentage). Results are averaged across 1,000 tasks. Results marked with ’ - ’ were intractable to obtain. + +
MethodCUB (K=50)
5-shot10-shot20-shot
Baseline [8]58.668.878.2
LR+ICI [52]49.955.658.0
PT-MAP [17]12.814.014.9
LaplacianShot [60]58.866.571.0
TIM [4]68.168.969.3
α-TIM [50]74.379.383.6
α-AM [24]66.268.969.8
PADDLE [32]71.281.886.8
UNEM-Gaussian78.585.388.6
+ +Table 2. Comparison of the proposed UNEM-Gaussian with respect to state-of-the-art methods on CUB. The metric is accuracy (in percentage). Results are averaged across 1,000 tasks. + +This implies that the proposed solution is highly economical in terms of parameters, typically requiring only a few tenths, compared to standard neural network architectures. Let us recall that the validation set is often used to select the hyperparameters in recent state-of-the-art transductive methods as mentioned in Section 2. Specifically, for the vision-only (resp. vision-language) models, the training of the unrolled architecture is carried out on 1,000 (resp. 100) tasks, while using an initial learning rate of 0.1 (resp. 0.5) with a decay factor of 0.5, a number of epochs equals to 80, and ADAM as optimizer. The trainable hyper-parameters are initialized to the default values set in [32] (i.e., $\lambda = |Q|$ for the vision-only models) and [33] (i.e., $\lambda = \frac{K}{K_{\mathrm{eff}}}$ for the vision-language models, with $T = 1$ ). Our architecture is + +implemented in Pytorch (version 2.3.0) and run on NVIDIA QUADRO RTX8000 (with 48 GB of memory). + +# 4.2. UNEM-Gaussian in vision-only few-shot setting + +Let us recall that the Gaussian distribution is commonly used in standard clustering and transductive few-shot methods applied to vision-only models [32, 46, 50]. For this reason, we will focus here on the proposed UNEM-Gaussian approach. + +Datasets: The first UNEM-Gaussian architecture is evaluated on the following standard few-shot benchmark datasets: mini-[ImageNet [51], tiered-[ImageNet [41], CUB [14]. Mini-imagenet has 100 classes split into 64 base classes, 16 validation classes and 20 test classes. The tiered-imagenet has 608 classes instead, from which we follow a standard split with 351 for base training, 97 for validation and 160 for testing. For CUB, we followed the split proposed by [8] which consists of 100 base classes, 50 validation classes and 50 test classes. For each dataset, the feature vectors $\mathbf{z}_n$ are extracted using the fine-tuned backbones, while applying a scaling parameter $T_{z}$ to the model's output as described in Appendix B. This parameter has also been learned through the unrolling approach. + +Results: The UNEM-Gaussian architecture has been compared to its original version PADDLE [32] as well as several state-of-the-art methods. Tables 1 and 2 presents the results, averaged over 1,000 tasks with 5, 10, and 20 shots. Several observations can be made. First, the proposed UNEM-Gaussian outperforms the original PADDLE algorithm [32] and the other state-of-the-art methods. The achieved gains + +
MethodFood101EuroSATDTDOrionPedsFlowers102Caltech101UCF101FGVC AircraftStanford CarsSUM397ImageNetAverage
Ind.Tip-Adapter [56]76.772.554.786.483.288.872.123.763.966.762.768.3
CoOp [57]76.363.252.286.281.087.767.022.261.363.459.965.5
Transd.BDSCPN [28]74.746.145.281.374.282.059.018.048.154.549.257.5
Laplacian Shot [60]76.653.052.688.485.586.867.022.260.463.856.364.8
α-TIM [50]66.146.145.387.179.183.359.420.453.453.442.757.8
PADDLE [32]71.845.950.084.782.381.963.721.356.160.652.160.9
EM-Dirichlet [33]88.750.862.692.591.390.176.124.973.580.978.473.6
UNEM-Dirichlet91.453.865.396.095.693.478.530.480.088.583.177.8
+ +Table 3. Comparison of the proposed UNEM-Dirichlet with respect to the state-of-the-art methods on 11 different datasets. The metric is accuracy (in percentage). Results are averaged across 1,000 tasks. + +are much higher with fewer number of shots and reach $3.5\%$ with mini-ImageNet, $6.9\%$ with tiered-ImageNet, and $7.3\%$ with CUB, while using ResNet-18 as backbone. Moreover, when WRN28-10 is used as backbone, it can be noticed that the performance of all methods has been improved, except for PADDLE, which shows results similar to those obtained with ResNet-18 pre-training. This can be explained by the fact that the class-balance parameter $\lambda$ was set in [32] to $|\mathbb{Q}|$ (i.e. 75), which becomes sub-optimal with WRN28-10 features. Most importantly, by learning efficiently the hyperparameters, our UNEM-Gaussian achieves higher gain with respect to its original version, yielding an accuracy gain reaching up to $10\%$ in 5-shot scenario, when WRN28-10 is used as backbone. + +# 4.3. UNEM-Dirichlet for few-shot CLIP + +Dirichlet distribution has recently demonstrated good modelling performance in the context of transductive few-shot for vision-language models such as CLIP [33]. Thus, we will focus in this second round of experiments on UNEM-Dirichlet approach. + +Datasets: The second unrolled architecture, designated by UNEM-Dirichlet, is assessed on 11 benchmark datasets that are commonly used for CLIP scenario: Caltech101 [10], OxfordPnts [37], StanfordCars [22], Flowers102 [36], Food101 [3], FGVCAircraft [31], SUN397 [54], DTD [9], EuroSAT [15], UCF101 [47], and ImageNet [42]. These datasets cover diverse classification challenges, from object recognition (Caltech101, Food101) to fine-grained tasks (StanfordCars, FGVCAircraft) and scene understanding (SUN397, EuroSAT). For each dataset, the vision-text feature vectors $z_{n}$ are extracted using the CLIP's pre-trained model, while applying a temperature parameter $T_{z}$ as described in Appendix C. This temperature parameter has also been learned through the unrolling approach. + +Results: The proposed UNEM-Dirichlet has also been compared to its original EM-Dirichlet version [33] as well as different few-shot classification methods. Table 3 depicts the accuracy results evaluated over 1,000 tasks with 4- + +shots. Thus, it can be noticed that UNEM-Dirichlet outperforms state-of-the-art methods for most datasets. In particular, unrolling the recent EM-Dirichlet algorithm yields further improvement of about $4.2\%$ in average. The achieved gain is more significant (reaching up to $7.5\%$ ) with more challenging datasets having a large number of classes like FGVC Aircraft, Stanford Cars, and SUN397. This is due to the inappropriate choice of the hyperparameter $\lambda$ for these datasets. Indeed, the selected $\lambda$ parameter was set to $\frac{K}{K_{\mathrm{eff}}} |\mathbb{Q}|$ in EM-Dirichlet approach [33]. Thus, by computing the numerical values of the hyperparameter for each of the aforementioned datasets, it can be deduced from Figure 1 that the difference between the selected value (which is 3,000 for Stanford Cars) and the optimal one (which is around 5,500 for Stanford Cars) is more important and impacts significantly the accuracy performance. This problem is well addressed by resorting to the proposed unrolled model for hyperparameters optimization. + +Appendix D includes additional results to illustrate, in both scenarios, the effects of the temperature parameter in this framework, and to show the benefits of learning variable hyperparameters across the architecture layers. + +# 5. Conclusion + +In this paper, we focus on transductive few-shot methods grounded on a generalized form of the EM algorithm. These methods include a parameter to control class balance. The proposed GEM algorithm, applicable to any mixture of distributions, incorporates also a temperature scaling parameter. The optimization process is then unrolled into a neural network architecture, enabling efficient learning of the introduced hyper-parameters. The results demonstrate the effectiveness of the proposed approach on both vision-only and vision-language models. In future work, it would be valuable to explore the potential of unrolling techniques across a broader range of computer vision tasks, especially with the recent rise of foundational vision-language models. + +# References + +[1] Christina Baek, Yiding Jiang, Aditi Raghunathan, and Zico Kolter. Agreement-on-the-line: Predicting the performance of neural networks under distribution shift. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), pages 19274-19289, 2022. 2 +[2] C. Bertocchi, E. Chouzenoux, M.-C. Corbineau, J.-C. Pesquet, and M. Prato. Deep unfolding of a proximal interior point method for image restoration. Inverse Problems, Special Issue on Variational Methods and Effective Algorithms for Imaging and Vision, 36(3):1-27, 2020. 5 +[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In European Conference on Computer Vision, pages 446-461, Zurich, Switzerland, 2014. Springer. 8 +[4] Malik Boudiaf, Ziko Imtiaz Masud, Jérôme Rony, José Dolz, Pablo Piantanida, and Ismail Ben Ayed. Transductive information maximization for few-shot learning. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), pages 2445–2457, 2020. 2, 7 +[5] Y. Boykov, H. N. Isack, C. Olsson, and I. Ben Ayed. Volumetric bias in segmentation and reconstruction: Secrets and solutions. In IEEE/CVF International Conference on Computer Vision (ICCV), 2015. 4 +[6] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Prompt learning with optimal transport for vision-language models. In International Conference on Learning Representations (ICLR), 2023. 3 +[7] Tianlong Chen, Xiaohan Chen, Wuyang Chen, and Zhangyang Wang. Learning to optimize: A primer and a benchmark. Journal of Machine Learning Research (JMLR), 23:1-59, 2022. 5 +[8] Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, and Jia-Bin Huang. A closer look at few-shot classification. In International Conference on Learning Representations (ICLR), pages 1-19, 2019. 7 +[9] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, and A. Vedaldi. Describing textures in the wild. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 8 +[10] Li Fei-Fei, Robert Fergus, and Pietro Perona. One-shot learning of object categories. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28(4):594-611, 2006. 8 +[11] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International Conference on Machine Learning, 2017. 2 +[12] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. CLIP-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 32:581-595, 2023. 3 +[13] Bharath Hariharan and Ross B. Girshick. Low-shot visual recognition by shrinking and hallucinating features. In IEEE International Conference on Computer Vision (ICCV), pages 3018-3027, 2017. 2 +[14] Xiangteng He and Yuxin Peng. Fine-grained visual-textual + +representation learning. IEEE Transactions on Circuits and Systems for Video Technology, 30(2):520-531, 2020. 7 +[15] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. EuroSAT: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 8 +[16] Shell Xu Hu, Da Li, Jan Stuhmer, Minyoung Kim, and Timothy M Hospedales. Pushing the limits of simple pipelines for few-shot learning: External data and fine-tuning make a difference. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9068-9077, 2022. 2 +[17] Yuqing Hu, Vincent Gripon, and Stéphane Pateux. Leveraging the feature distribution in transfer-based few-shot learning. In International Conference on Artificial Neural Networks, pages 487-499, 2021. 2, 7 +[18] Yunshi Huang, Fereshteh Shakeri, Jose Dolz, Malik Boudiaf, Houda Bahig, and Ismail Ben Ayed. $\mathrm{LP} + +$ : A surprisingly strong linear probe for few-shot clip. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23773-23782, 2024. 3 +[19] Muhammad Abdullah Jamal and Guo-Jun Qi. Task agnostic meta-learning for few-shot learning. In IEEE conference on Computer Vision and Pattern Recognition (CVPR), pages 11719-11727, 2019. 2 +[20] Leonid Karlinsky, Joseph Shtok, Sivan Harary, Eli Schwartz, Amit Aides, and Rogerio Feris. Repmet: Representative-based metric learning for classification and few-shot object detection. In IEEE conference on Computer Vision and Pattern Recognition (CVPR), pages 5197-5206, 2019. 2 +[21] M. Kearns, Y. Mansour, and A. Ng. An information-theoretic analysis of hard and soft assignment methods for clustering. In Conference on Uncertainty in Artificial Intelligence (UAI), 1997. 4 +[22] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 8 +[23] Michalis Lazarou, Tania Stathaki, and Yannis Avrithis. Iterative label cleaning for transductive and semi-supervised few-shot learning. In IEEE/CVF International Conference on Computer Vision, pages 8751-8760, 2021. 2 +[24] Michalis Lazarou, Yannis Avrithis, and Tania Stathaki. Adaptive manifold for imbalanced transductive few-shot learning. In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), page 2297-2306, 2024. 2, 3, 7 +[25] Gao Yu Lee, Tanmoy Dam, Daniel Puiu Poenar, Vu N. Duong, and Md Meftahul Ferdaus. HELA-VFA: A hellinger distance-attention-based feature aggregation network for few-shot classification. In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2173–2183, 2024. 2 +[26] Xiaoxu Li, Xiaochen Yang, Zhanyu Ma, and Jing-Hao Xue. Deep metric learning for few-shot image classification: A review of recent developments. Pattern Recognition, 138: 109381, 2023. 2 + +[27] Moshe Lichtenstein, Prasanna Sattigeri, Rogerio Feris, Raja Giryes, and Leonid Karlinsky. TAFSSL: Task-adaptive feature sub-space learning for few-shot classification. In European Conference on Computer Vision (ECCV), pages 22-539, 2020. 2 +[28] Jinlu Liu, Liang Song, and Yongqiang Qin. Prototype rectification for few-shot learning. In European Conference on Computer Vision (ECCV), pages 741-756, 2020. 2, 7, 8 +[29] Yanbin Liu, Juho Lee, Minseop Park, Saehoon Kim, Eunho Yang, Sung Ju Hwang, and Yi Yang. Learning to propagate labels: Transductive propagation network for few-shot learning. In International Conference on Learning Representations, page 34677-34688, 2019. 2, 6 +[30] Xu Luo, Hao Wu, Ji Zhang, Lianli Gao, Jing Xu, and Jingkuan Song. A closer look at few-shot classification again. In International Conference on Machine Learning (ICML), pages 23103-23123, 2023. 2 +[31] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. Technical report, Oxford University, 2013. 8 +[32] Ségolène Martin, Malik Boudiaf, Emilie Chouzenoux, Jean-Christophe Pesquet, and Ismail Ben Ayed. Towards practical few-shot query sets: Transductive minimum description length inference. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), page 34677–34688, 2022. 2, 3, 4, 6, 7, 8, 5 +[33] Ségolène Martin, Yunshi Huang, Fereshteh Shakeri, Jean-Christophe Pesquet, and Ismail Ben Ayed. Transductive zero-shot and few-shot clip. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28816-28826, 2024. 1, 2, 3, 4, 5, 6, 7, 8 +[34] Paul Micaelli and Amos J Storkey. Gradient-based hyperparameter optimization over long horizons. In Advances in Neural Information Processing Systems, pages 10798-10809. Curran Associates, Inc., 2021. 3 +[35] Vishal Monga, Yuelong Li, and Yonina C. Eldar. Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing. IEEE Signal Processing Magazine, 38(2):18-44, 2021. 5 +[36] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 8 +[37] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 8 +[38] Hang Qi, Matthew Brown, and David G Lowe. Low-shot learning with imprinted weights. In IEEE conference on Computer Vision and Pattern Recognition (CVPR), pages 5822-5830, 2018. 2 +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, and et al. Jack Clark. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 3 + +[40] Sachin Ravi and H. Larochelle. Optimization as a model for few-shot learning. In International Conference on Learning Representations, pages 1-11, 2017. 2 +[41] Mengye Ren, Eleni Triantafillou, Sachin Ravi, Jake Snell, Kevin Swersky, Joshua B. Tenenbaum, Hugo Larochelle, and Richard S. Zemel. Meta-learning for semi-supervised few-shot classification. In International Conference on Learning Representations (ICLR), pages 1-15, 2018. 7 +[42] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. Int. J. Comput. Vision, 115(3):211-252, 2015. 8 +[43] Tyler Scott, Karl Ridgeway, and Michael C. Mozer. Adapted deep embeddings: A synthesis of methods for k-shot inductive transfer learning. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), pages 76-85, 2018. 2 +[44] Anuj Singh and Hadi Jamali-Rad. Transductive decoupled variational inference for few-shot classification. Transactions on Machine Learning Research (TMLR), 2023. 2 +[45] Jake Snell, Kevin Swersky, and Richard S. Zemel. Prototypical networks for few-shot learning. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), pages 4080-4090, 2017. 6 +[46] Yu Song and Changsheng Chen. MPPCANet: A feedforward learning strategy for few-shot image classification. Pattern Recognition, 113:107792, 2021. 4, 7 +[47] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. Technical report, Center for Research in Computer Vision, University of Central Florida, 2012. arXiv:1212.0402. 8 +[48] Ran Tao, Hao Chen, and Marios Savvides. Boosting transductive few-shot fine-tuning with margin-based uncertainty weighting and probability regularization. In IEEE conference on Computer Vision and Pattern Recognition (CVPR), pages 15752-15761, 2023. 2, 3 +[49] Long Tian, Jingyi Feng, Xiaoqiang Chai, Wenchao Chen, Liming Wang, Xiyang Liu, and Bo Chen. Prototypes-oriented transductive few-shot learning with conditional transport. In International Conference on Computer Vision (ICCV), pages 16317-16326, 2023. 2 +[50] Olivier Veilleux, Malik Boudiaf, Pablo Piantanida, and Ismail Ben Ayed. Realistic evaluation of transductive few-shot learning. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), pages 9290-9302, 2021. 2, 3, 4, 6, 7, 8 +[51] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), pages 3637-3645, 2016. 2, 7 +[52] Yikai Wang, Chengming Xu, Chen Liu, Li Zhang, and Yanwei Fu. Instance credibility inference for few-shot learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12833-12842, 2020. 7 + +[53] Yimu Wang, Yihan Wu, and Hongyang Zhang. Lost domain generalization is a natural consequence of lack of training domains. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 15689-15697, 2024. 2 +[54] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3485-3492, 2010. 8 +[55] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6757-6767, 2023. 3 +[56] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. TIP-adapter: Training-free adaption of clip for few-shot classification. In European Conference on Computer Vision (ECCV), pages 493-510, 2022. 3, 8 +[57] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16816-16825, 2022. 3, 8 +[58] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In International Conference on Computer Vision (ICCV), pages 15659-15669, 2023. 3 +[59] Hao Zhu and Piotr Koniusz. Transductive few-shot learning with prototype-based label propagation by iterative graph refinement. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 23996-24006, 2023. 2 +[60] Imtiaz Masud Ziko, Jose Dolz, Eric Granger, and Ismail Ben Ayed. Laplacian regularized few-shot learning. In International Conference on Machine Learning (ICML), 2020. 7, 8 \ No newline at end of file diff --git a/2025/UNEM_ UNrolled Generalized EM for Transductive Few-Shot Learning/images.zip b/2025/UNEM_ UNrolled Generalized EM for Transductive Few-Shot Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..15c814990113419fa748a07a7e914f2149bebbfa --- /dev/null +++ b/2025/UNEM_ UNrolled Generalized EM for Transductive Few-Shot Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bea3f4ef5794ae5a130eaf4c5cb121251913b4a346153298195405227aefb8f6 +size 362001 diff --git a/2025/UNEM_ UNrolled Generalized EM for Transductive Few-Shot Learning/layout.json b/2025/UNEM_ UNrolled Generalized EM for Transductive Few-Shot Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..496bed7431fbf84e8080b21569245a900cefb8fa --- /dev/null +++ b/2025/UNEM_ UNrolled Generalized EM for Transductive Few-Shot Learning/layout.json @@ -0,0 +1,9506 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 84, + 102, + 526, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 102, + 526, + 121 + ], + "spans": [ + { + "bbox": [ + 84, + 102, + 526, + 121 + ], + "type": "text", + "content": "UNEM: UNrolled Generalized EM for Transductive Few-Shot Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 152, + 462, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 152, + 462, + 182 + ], + "spans": [ + { + "bbox": [ + 146, + 152, + 462, + 182 + ], + "type": "text", + "content": "Long Zhou" + }, + { + "bbox": [ + 146, + 152, + 462, + 182 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 146, + 152, + 462, + 182 + ], + "type": "text", + "content": ", Fereshteh Shakeri" + }, + { + "bbox": [ + 146, + 152, + 462, + 182 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 152, + 462, + 182 + ], + "type": "text", + "content": ", Aymen Sadraoui" + }, + { + "bbox": [ + 146, + 152, + 462, + 182 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 146, + 152, + 462, + 182 + ], + "type": "text", + "content": ", Mounir Kaaniche" + }, + { + "bbox": [ + 146, + 152, + 462, + 182 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 146, + 152, + 462, + 182 + ], + "type": "text", + "content": ", Jean-Christophe Pesquet" + }, + { + "bbox": [ + 146, + 152, + 462, + 182 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 146, + 152, + 462, + 182 + ], + "type": "text", + "content": ", Ismail Ben Ayed" + }, + { + "bbox": [ + 146, + 152, + 462, + 182 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 250, + 190, + 361, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 190, + 361, + 204 + ], + "spans": [ + { + "bbox": [ + 250, + 190, + 361, + 204 + ], + "type": "text", + "content": "1Politecnico di Milano" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 268, + 204, + 342, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 204, + 342, + 218 + ], + "spans": [ + { + "bbox": [ + 268, + 204, + 342, + 218 + ], + "type": "text", + "content": "2ÉTS Montréal" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 173, + 218, + 436, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 218, + 436, + 232 + ], + "spans": [ + { + "bbox": [ + 173, + 218, + 436, + 232 + ], + "type": "text", + "content": "3Université Paris-Saclay, Inria, CentraleSupélec, CVN" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 185, + 233, + 424, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 185, + 233, + 424, + 246 + ], + "spans": [ + { + "bbox": [ + 185, + 233, + 424, + 246 + ], + "type": "text", + "content": "4Université Sorbonne Paris Nord, L2TI, UR 3043" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 282, + 200, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 282, + 200, + 295 + ], + "spans": [ + { + "bbox": [ + 151, + 282, + 200, + 295 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 354, + 296, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 700 + ], + "type": "text", + "content": "Transductive few-shot learning has recently triggered wide attention in computer vision. Yet, current methods introduce key hyper-parameters, which control the prediction statistics of the test batches, such as the level of class balance, affecting performances significantly. Such hyper-parameters are empirically grid-searched over validation data, and their configurations may vary substantially with the target dataset and pre-training model, making such empirical searches both sub-optimal and computationally intractable. In this work, we advocate and introduce the unrolling paradigm, also referred to as \"learning to optimize\", in the context of few-shot learning, thereby learning efficiently and effectively a set of optimized hyperparameters. Specifically, we unroll a generalization of the ubiquitous Expectation-Maximization (EM) optimizer into a neural network architecture, mapping each of its iterates to a layer and learning a set of key hyper-parameters over validation data. Our unrolling approach covers various statistical feature distributions and pre-training paradigms, including recent foundational vision-language models and standard vision-only classifiers. We report comprehensive experiments, which cover a breadth of fine-grained downstream image classification tasks, showing significant gains brought by the proposed unrolled EM algorithm over iterative variants. The achieved improvements reach up to " + }, + { + "bbox": [ + 55, + 354, + 296, + 700 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 354, + 296, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 354, + 296, + 700 + ], + "type": "inline_equation", + "content": "7.5\\%" + }, + { + "bbox": [ + 55, + 354, + 296, + 700 + ], + "type": "text", + "content": " on vision-only and vision-language benchmarks, respectively. The source code and learned parameters are available at https://github.com/ZhouLong0/UNEM-Transductive." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 325, + 285, + 542, + 443 + ], + "blocks": [ + { + "bbox": [ + 325, + 285, + 542, + 443 + ], + "lines": [ + { + "bbox": [ + 325, + 285, + 542, + 443 + ], + "spans": [ + { + "bbox": [ + 325, + 285, + 542, + 443 + ], + "type": "image", + "image_path": "ceb9a74013acf42b64735a8a9a5a565d872d5c29145950cf8386ee1d954da2df.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 445, + 555, + 578 + ], + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 578 + ], + "type": "text", + "content": "Figure 1. Impact of the class-balance hyperparameter " + }, + { + "bbox": [ + 313, + 445, + 555, + 578 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 445, + 555, + 578 + ], + "type": "text", + "content": " on the accuracy of transductive few-shot classification. The accuracy results are obtained using the EM-Dirichlet algorithm [33] applied to vision-language models (with 4-shots). The plot shows that the choice of " + }, + { + "bbox": [ + 313, + 445, + 555, + 578 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 445, + 555, + 578 + ], + "type": "text", + "content": " has a strong impact on the performance, and that the optimal " + }, + { + "bbox": [ + 313, + 445, + 555, + 578 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 445, + 555, + 578 + ], + "type": "text", + "content": " (indicated with the star symbol) might vary by orders of magnitudes, depending on the target downstream dataset (e.g., the ten very different fine-grained classification datasets). Further comments on the optimal " + }, + { + "bbox": [ + 313, + 445, + 555, + 578 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 445, + 555, + 578 + ], + "type": "text", + "content": " values and the values chosen in [33] are provided in Section 4. The values of the learned hyper-parameters based on the proposed unrolled algorithm are illustrated and analyzed in Appendix D." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 605, + 394, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 605, + 394, + 618 + ], + "spans": [ + { + "bbox": [ + 314, + 605, + 394, + 618 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "Deep learning has transfigured computer vision, driving substantial progress in tasks such as image classification, captioning, object detection and segmentation. However, these successes often come with a high cost: the requirement for large amounts of labeled data. Additionally, the generalization of these models is seriously challenged when evaluated on new classes (concepts), unseen during pre" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9665" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 85 + ], + "type": "text", + "content": "training, or when operating under distribution shifts [1, 53]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 295, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 295, + 266 + ], + "type": "text", + "content": "To address these challenges, Few-Shot Learning (FSL) has recently attracted wide attention within the computer vision community. In the standard few-shot image classification setting, a feature extractor is first pre-trained on a set of classes, often called the base classes. Then, the model is adapted and evaluated on new classes and tasks. Each evaluation task is performed on a set of unlabeled samples (referred to as the query set), and supervised by a support set composed of few labeled samples per new class. Earlier FSL methods have relied on various concepts, including meta-learning [11, 19, 40], transfer learning [13, 38, 43], and metric learning [20, 26, 51]. However, most of these works operate within the inductive setting, where, at inference time, the prediction for each sample is made independently from the other samples in the query set." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 268, + 295, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 268, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 295, + 472 + ], + "type": "text", + "content": "Recently, significant attention has shifted to the transductive scheme, in which inference is performed jointly on a batch of query samples. Transduction leverages the statistics of the unlabeled query samples, yielding notable performance gains, and have triggered a large body of recent works in few-shot learning, with various methods and mechanisms [4, 23-25, 29, 32, 59], as detailed in Section 2. Generally, transductive methods are built upon optimizing objective functions that integrate clustering terms, either discriminative as in information maximization [4, 50] or generative as in probabilistic K-means [32, 33]. However, many among these assume a perfectly balanced class distribution within the query set, and incorporate terms in the objective function or constraints that enforce a class-balance prior. The latter could limit the applicability of these methods, whose performances were shown to drop significantly when dealing with imbalanced query sets [27, 50]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "content": "Several recent attempts tackled this limitation, to handle more realistic scenarios [24, 32, 33, 48, 50]. Yet, these methods introduce key hyper-parameters, which control the prediction statistics of the unlabeled query set, such as the level of class balance. Such hyper-parameters are empirically grid-searched over pre-defined sets of values using validation data, and their optimal configurations may vary substantially with the target dataset and pre-training model [17]. To illustrate this, we depict in Fig. 1 the accuracy as a function of a class-balance hyper-parameter for the recent method in [33]. One may observe that this hyperparameter has a crucial effect on the performance, and its optimal value might vary by orders of magnitude across the target datasets. The issue is further compounded by additional hyper-parameters, which makes intensive empirical grid searches for the hyper-parameters over validation data and pre-defined intervals of values both sub-optimal and computationally intractable. Therefore, we advocate and introduce the unrolling paradigm (also called \"learning to optimize\") in the context of few-shot learning, which enables" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": "to learn efficiently and effectively a set of optimized hyperparameters. Specifically, our main contributions could be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 315, + 108, + 553, + 395 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 315, + 108, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 108, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 315, + 108, + 553, + 204 + ], + "type": "text", + "content": "1. We study a generalization of the ubiquitous Expectation-Maximization (EM) algorithm, in which we make two hyper-parameters controlling prediction statistics - class balance and prediction entropy - explicit1. Our generalization encompasses several existing transductive few-shot methods as particular cases and, more importantly, enables to learn these crucial hyper-parameters through the proposed unrolling strategy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 315, + 204, + 553, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 204, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 315, + 204, + 553, + 299 + ], + "type": "text", + "content": "2. We unroll the generalized EM optimizer into a neural network architecture, mapping each of its iterations to a layer and learning the introduced hyper-parameters over validation data. Our unrolling approach offers greater flexibility in optimizing the hyper-parameters, allowing them to vary across the network's layers. To the best of our knowledge, this study is the first to investigate unrolling in transductive few-shot learning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 300, + 553, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 300, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 315, + 300, + 553, + 346 + ], + "type": "text", + "content": "3. We design our unrolling architecture in way that covers various statistical assumptions and pre-training paradigms: (i) Gaussian for vision-only classifiers and (ii) Dirichlet for vision-language models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 347, + 553, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 347, + 553, + 395 + ], + "spans": [ + { + "bbox": [ + 315, + 347, + 553, + 395 + ], + "type": "text", + "content": "4. We report comprehensive experiments, which cover a breadth of fine-grained downstream image classification tasks and different pre-training models, showing substantial gains over recent state-of-the-art methods." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 404, + 403, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 404, + 403, + 417 + ], + "spans": [ + { + "bbox": [ + 314, + 404, + 403, + 417 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "type": "text", + "content": "Few-shot classification with vision-only models: Few-shot classification using vision models has been widely explored in the literature, leading to the development of various methods. The first category, inductive methods, predicts the class of each test sample (in the query set) independently from others [16, 30]. In contrast, the second category, transductive methods, which has gained increasing attention in recent years, involves jointly predicting classes for a batch of test samples in each few-shot task. Numerous research efforts in transductive methods have leveraged concepts like clustering [24, 32], label propagation [29, 59], information maximization [4, 50], optimal transport [23, 49], prototype estimation [28, 59], and variational networks [25, 44], among others. Studies have shown that transductive methods significantly outperform inductive approaches, achieving accuracy gains of up to " + }, + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 313, + 425, + 555, + 628 + ], + "type": "text", + "content": " as reported in several evaluations [4, 32]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 628, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 555, + 688 + ], + "type": "text", + "content": "Few-shot classification with vision-language models: Contrastive Vision-Language Pre-training (CLIP) has recently emerged as an effective model for enhancing various vision tasks through visual-text pairs. By learning transferable visual models with natural language supervision, CLIP" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 314, + 693, + 553, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 693, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 314, + 693, + 553, + 712 + ], + "type": "text", + "content": "1Those hyper-parameters are implicit (hidden) in the standard EM formulation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9666" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 574 + ], + "type": "text", + "content": "demonstrates strong performance in zero-shot classification by matching the image features to text embeddings of novel classes [39]. To further enhance its classification capabilities, several studies have extended CLIP to the few-shot setting [6, 18, 55, 57, 58]. While the linear probe model [39] trains a logistic regression classifier using CLIP image features, the authors of [18] proposed a generalization of this baseline, modeling the classifier weights as learnable functions of the visual prototypes and text embeddings. Context Optimization (CoOp) [57] and its extended versions [6, 55, 58] have been developed based on the concept of prompt learning. While CoOp [57] aims to model context in prompts using continuous representations, a knowledge-guided CoOp is proposed in [55] to enhance the generalization ability by minimizing the discrepancy between the learnable prompts and the original ones. In [6], the authors learn multiple prompts, describing the characteristics of each class, through the minimization of an optimal-transport distance. Unlike prompt learning methods, which fine-tune the input text, another family of methods, referred to as adapters, aims to transform the visual or text encoders, such as [12, 56]. For instance, TIP-Adapter [56] adds a nonparametric adapter to the weight-frozen CLIP model, and updates the prior knowledge encoded in CLIP by feature retrieval. CLIP-Adapter [12] introduces a multi-layered perceptron to learn new features, and combine them with the original CLIP-encoded features via residual connections. It is worth noting that all the aforementioned methods belong to the inductive family. However, unlike vision-only models, the transductive few-shot setting is still not well investigated in the context of CLIP. To the best of our knowledge, the only transductive few-shot CLIP method is the one very recently proposed in [33], which, inspired by the Expectation Maximization algorithm, relies on the Dirichlet distribution to model the data. Moreover, as reported in [33], and unlike the behaviors observed with vision-only models, recent transductive few-shot methods do not always outperform their inductive counterparts with CLIP. This suggests that there is a need to further investigate transductive methods in the context of vision-language models, due to their aforementioned advantages with respect to inductive methods." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 582, + 298, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 298, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 298, + 715 + ], + "type": "text", + "content": "Class-balance and hyperparameters setting: Hyperparameter optimization has a large impact on the performance of a given iterative algorithm or learning model, and is generally tackled through greedy or gradient-based methods [34]. In the context of transductive few-shot learning, most of the existing methods have been designed for perfectly balanced query sets (i.e., uniform class distribution), and have shown drops in performances under class-imbalanced settings. To address this flaw, various methods have been developed in the last years [24, 32, 33, 48, 50]. For instance, in [48] the categorical probability of each query" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 313, + 72, + 556, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 240 + ], + "type": "text", + "content": "sample is regularized to quantify the difference between the class marginal distribution and the uniform one. The works in [24, 32, 33, 50] explored various weighted terms, which are added to the objective functions, to mitigate the effect of the class-balance bias. To this end, some hyper-parameters are introduced, to weigh to contributions of such terms in the objective function. Unlike [32], where the introduced hyper-parameter has been theoretically chosen in order to compensate for the hidden class-balance bias in the used clustering objective, the hyper-parameters are often set in an empirical manner based on the validation classes of each dataset. Thus, a set of hyperparameter values (in a given range) are evaluated, and the ones yielding the highest accuracy are selected." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 253, + 446, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 446, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 446, + 266 + ], + "type": "text", + "content": "3. Proposed methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 273, + 463, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 463, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 463, + 285 + ], + "type": "text", + "content": "3.1. Generalized EM algorithm" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 292, + 556, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 556, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 556, + 422 + ], + "type": "text", + "content": "Preliminaries: Let us first introduce the notations to formulate our transductive few-shot inference. Let " + }, + { + "bbox": [ + 313, + 292, + 556, + 422 + ], + "type": "inline_equation", + "content": "\\{\\pmb{z}_n\\}_{1\\leq n\\leq N}" + }, + { + "bbox": [ + 313, + 292, + 556, + 422 + ], + "type": "text", + "content": " denote the set of feature vectors extracted from a pretraining network, and which are to be classified, with " + }, + { + "bbox": [ + 313, + 292, + 556, + 422 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 292, + 556, + 422 + ], + "type": "text", + "content": " the total number of samples within a given task. The whole dataset contains " + }, + { + "bbox": [ + 313, + 292, + 556, + 422 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 292, + 556, + 422 + ], + "type": "text", + "content": " distinct classes, whereas the number of randomly sampled classes present in each mini-batch task might be much smaller than " + }, + { + "bbox": [ + 313, + 292, + 556, + 422 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 292, + 556, + 422 + ], + "type": "text", + "content": ". This subset of sampled classes may also vary across mini-batches, and the method does not assume any prior knowledge on the specific classes that may appear in each mini-batch." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 424, + 556, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 556, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 556, + 508 + ], + "type": "text", + "content": "Moreover, for a given few-shot task, let us denote by " + }, + { + "bbox": [ + 313, + 424, + 556, + 508 + ], + "type": "inline_equation", + "content": "\\mathbb{S} \\subset \\{1, \\dots, N\\}" + }, + { + "bbox": [ + 313, + 424, + 556, + 508 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 424, + 556, + 508 + ], + "type": "inline_equation", + "content": "\\mathbb{Q} = \\{1, \\dots, N\\} \\setminus \\mathbb{S}" + }, + { + "bbox": [ + 313, + 424, + 556, + 508 + ], + "type": "text", + "content": " the indices of samples within the support (labeled) and query (unlabeled) mini-batch sets, respectively. For every " + }, + { + "bbox": [ + 313, + 424, + 556, + 508 + ], + "type": "inline_equation", + "content": "n \\in \\mathbb{S}" + }, + { + "bbox": [ + 313, + 424, + 556, + 508 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 424, + 556, + 508 + ], + "type": "inline_equation", + "content": "\\pmb{y}_n = (y_{n,k})_{1 \\leq k \\leq K} \\in \\{0, 1\\}^K" + }, + { + "bbox": [ + 313, + 424, + 556, + 508 + ], + "type": "text", + "content": " are the one-hot-encoded labels, such that, for every " + }, + { + "bbox": [ + 313, + 424, + 556, + 508 + ], + "type": "inline_equation", + "content": "k \\in \\{1, \\dots, K\\}" + }, + { + "bbox": [ + 313, + 424, + 556, + 508 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 424, + 556, + 508 + ], + "type": "inline_equation", + "content": "y_{n,k} = 1" + }, + { + "bbox": [ + 313, + 424, + 556, + 508 + ], + "type": "text", + "content": " if the " + }, + { + "bbox": [ + 313, + 424, + 556, + 508 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 424, + 556, + 508 + ], + "type": "text", + "content": "-th sample belongs to class " + }, + { + "bbox": [ + 313, + 424, + 556, + 508 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 424, + 556, + 508 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 424, + 556, + 508 + ], + "type": "inline_equation", + "content": "y_{n,k} = 0" + }, + { + "bbox": [ + 313, + 424, + 556, + 508 + ], + "type": "text", + "content": " otherwise." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 509, + 556, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 556, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 556, + 581 + ], + "type": "text", + "content": "Finally, given the extracted feature vectors " + }, + { + "bbox": [ + 313, + 509, + 556, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_n" + }, + { + "bbox": [ + 313, + 509, + 556, + 581 + ], + "type": "text", + "content": ", we assume that the data probability distribution knowing its class " + }, + { + "bbox": [ + 313, + 509, + 556, + 581 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 509, + 556, + 581 + ], + "type": "text", + "content": " is modeled by a given law, whose probability density function (pdf) is denoted by " + }, + { + "bbox": [ + 313, + 509, + 556, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{p}(\\mathbf{z}_n \\mid \\boldsymbol{\\theta}_k)" + }, + { + "bbox": [ + 313, + 509, + 556, + 581 + ], + "type": "text", + "content": " and characterized by a vector of parameters " + }, + { + "bbox": [ + 313, + 509, + 556, + 581 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\theta}_k" + }, + { + "bbox": [ + 313, + 509, + 556, + 581 + ], + "type": "text", + "content": ". This means that the global distribution of " + }, + { + "bbox": [ + 313, + 509, + 556, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_n" + }, + { + "bbox": [ + 313, + 509, + 556, + 581 + ], + "type": "text", + "content": " is a mixture of these pdfs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 581, + 556, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 556, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 556, + 663 + ], + "type": "text", + "content": "Problem formulation: Our goal is to identify the classes of the unlabeled samples in the query set by optimizing a general clustering objective function, while embedding supervision constraints from the few labeled samples within the support set. We do so by unrolling iterative block-coordinate optimizers of the objective functions over two sets of variables:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 664, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 664, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 314, + 664, + 556, + 715 + ], + "type": "text", + "content": "- Soft assignment vectors " + }, + { + "bbox": [ + 314, + 664, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{u} = (u_n)_{1\\leq n\\leq N}\\in (\\Delta_K)^N" + }, + { + "bbox": [ + 314, + 664, + 556, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 314, + 664, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\Delta_K" + }, + { + "bbox": [ + 314, + 664, + 556, + 715 + ], + "type": "text", + "content": " is the probability simplex of " + }, + { + "bbox": [ + 314, + 664, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^K" + }, + { + "bbox": [ + 314, + 664, + 556, + 715 + ], + "type": "text", + "content": ". For every " + }, + { + "bbox": [ + 314, + 664, + 556, + 715 + ], + "type": "inline_equation", + "content": "n\\in \\{1,\\dots ,N\\}" + }, + { + "bbox": [ + 314, + 664, + 556, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 314, + 664, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{u}_n = (u_{n,k})_{1\\leq k\\leq K}" + }, + { + "bbox": [ + 314, + 664, + 556, + 715 + ], + "type": "text", + "content": " where, for every " + }, + { + "bbox": [ + 314, + 664, + 556, + 715 + ], + "type": "inline_equation", + "content": "k\\in \\{1,\\ldots ,K\\}" + }, + { + "bbox": [ + 314, + 664, + 556, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 314, + 664, + 556, + 715 + ], + "type": "inline_equation", + "content": "u_{n,k}" + }, + { + "bbox": [ + 314, + 664, + 556, + 715 + ], + "type": "text", + "content": " can be interpreted as the probabil" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9667" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 295, + 95 + ], + "type": "text", + "content": "ity that the " + }, + { + "bbox": [ + 63, + 72, + 295, + 95 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 63, + 72, + 295, + 95 + ], + "type": "text", + "content": "-th sample belongs to class " + }, + { + "bbox": [ + 63, + 72, + 295, + 95 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 63, + 72, + 295, + 95 + ], + "type": "text", + "content": ". These probabilities have to be determined for the query set samples." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 263, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 263, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 263, + 108 + ], + "type": "text", + "content": "- Feature distribution parameters " + }, + { + "bbox": [ + 55, + 96, + 263, + 108 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta} = (\\pmb{\\theta}_k)_{1\\leq k\\leq K}" + }, + { + "bbox": [ + 55, + 96, + 263, + 108 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 108, + 295, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 295, + 132 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 295, + 132 + ], + "type": "text", + "content": "Consider the following general probabilistic clustering problem:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 140, + 295, + 158 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 140, + 295, + 158 + ], + "spans": [ + { + "bbox": [ + 69, + 140, + 295, + 158 + ], + "type": "interline_equation", + "content": "\\underset {\\boldsymbol {u}, \\boldsymbol {\\theta}} {\\text {m i n i m i z e}} \\mathcal {L} (\\boldsymbol {u}, \\boldsymbol {\\theta}) + \\lambda \\Psi (\\boldsymbol {u}) + T \\Phi (\\boldsymbol {u}), \\tag {1}", + "image_path": "3af1636fb19b472b67eb42503d025db08f3ee4ce7ad3bd02aeb5ac061946ce01.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 161, + 201, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 161, + 201, + 174 + ], + "spans": [ + { + "bbox": [ + 69, + 161, + 201, + 174 + ], + "type": "text", + "content": "subject to " + }, + { + "bbox": [ + 69, + 161, + 201, + 174 + ], + "type": "inline_equation", + "content": "\\pmb{u}_n\\in \\Delta_K\\quad \\forall n\\in \\mathbb{Q}," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 111, + 176, + 281, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 176, + 281, + 190 + ], + "spans": [ + { + "bbox": [ + 111, + 176, + 281, + 190 + ], + "type": "interline_equation", + "content": "u _ {n, k} = y _ {n, k} \\quad \\forall n \\in \\mathbb {S}, \\forall k \\in \\{1, \\dots , K \\}.", + "image_path": "5a6927dae294b3e0f419cb15bac4926b0896e1415560f3ceaab2d5e3f6151be0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 197, + 296, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 197, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 296, + 243 + ], + "type": "text", + "content": "where weighing factor " + }, + { + "bbox": [ + 55, + 197, + 296, + 243 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 197, + 296, + 243 + ], + "type": "text", + "content": " and temperature scaling " + }, + { + "bbox": [ + 55, + 197, + 296, + 243 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 197, + 296, + 243 + ], + "type": "text", + "content": " are learnable optimized hyper-parameters, which we will estimate through the proposed unrolling strategy (Section 3.2), and terms " + }, + { + "bbox": [ + 55, + 197, + 296, + 243 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 197, + 296, + 243 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 197, + 296, + 243 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 55, + 197, + 296, + 243 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 197, + 296, + 243 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 55, + 197, + 296, + 243 + ], + "type": "text", + "content": " are detailed in the following." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 245, + 295, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 245, + 295, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 295, + 268 + ], + "type": "text", + "content": "- The first term in objective function (1) is the negative log-likelihood of the feature vectors:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 97, + 274, + 295, + 307 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 274, + 295, + 307 + ], + "spans": [ + { + "bbox": [ + 97, + 274, + 295, + 307 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\boldsymbol {u}, \\boldsymbol {\\theta}) = - \\sum_ {n = 1} ^ {N} \\sum_ {k = 1} ^ {K} u _ {n, k} \\ln \\left(\\mathrm {p} \\left(\\boldsymbol {z} _ {n} \\mid \\boldsymbol {\\theta} _ {k}\\right)\\right) \\tag {2}", + "image_path": "518837857e376880add4d8e2f258ac46242ef30764826aad91bfaf55c13bba29.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 63, + 312, + 296, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 312, + 296, + 385 + ], + "spans": [ + { + "bbox": [ + 63, + 312, + 296, + 385 + ], + "type": "text", + "content": "This general log-likelihood model fitting term is well known in the context of clustering methods [5, 21]. In fact, it generalizes the standard " + }, + { + "bbox": [ + 63, + 312, + 296, + 385 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 63, + 312, + 296, + 385 + ], + "type": "text", + "content": "-means clustering objective to arbitrary distributions2. It is well-known that minimizing (2) has an inherent bias towards class-balanced clustering [5, 21, 32, 33]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 385, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 296, + 434 + ], + "type": "text", + "content": "- The second term in (1) controls the partition complexity of the model, penalizing the number of non-empty clusters in the solution. It corresponds to the Shannon entropy of class distribution " + }, + { + "bbox": [ + 55, + 385, + 296, + 434 + ], + "type": "inline_equation", + "content": "(\\pi_k)_{1 \\leq k \\leq K}" + }, + { + "bbox": [ + 55, + 385, + 296, + 434 + ], + "type": "text", + "content": ", defined as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 129, + 441, + 295, + 474 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 441, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 129, + 441, + 295, + 474 + ], + "type": "interline_equation", + "content": "\\Psi (\\boldsymbol {u}) = - \\sum_ {k = 1} ^ {K} \\pi_ {k} \\ln \\pi_ {k}, \\tag {3}", + "image_path": "d19be703c167d2a06e996945a92b8ec9c7fff26733d8aec6fd56fa2da197a420.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 63, + 481, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 481, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 63, + 481, + 295, + 519 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 63, + 481, + 295, + 519 + ], + "type": "inline_equation", + "content": "\\pi_k = \\frac{1}{|\\mathbb{Q}|} \\sum_{n \\in \\mathbb{Q}} u_{n,k}" + }, + { + "bbox": [ + 63, + 481, + 295, + 519 + ], + "type": "text", + "content": " is the proportion of query samples within class " + }, + { + "bbox": [ + 63, + 481, + 295, + 519 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 63, + 481, + 295, + 519 + ], + "type": "text", + "content": ", i.e., the empirical estimate of the marginal probability of class " + }, + { + "bbox": [ + 63, + 481, + 295, + 519 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 63, + 481, + 295, + 519 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 63, + 519, + 296, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 519, + 296, + 686 + ], + "spans": [ + { + "bbox": [ + 63, + 519, + 296, + 686 + ], + "type": "text", + "content": "Class-balance hyper-parameter " + }, + { + "bbox": [ + 63, + 519, + 296, + 686 + ], + "type": "inline_equation", + "content": "(\\lambda)" + }, + { + "bbox": [ + 63, + 519, + 296, + 686 + ], + "type": "text", + "content": ": The marginal entropy in (3) mitigates the class-balance bias of the log-likelihood clustering term in (2). It reaches its minimum for the extremely imbalanced solution in which all data samples are assigned to a single cluster, and its maximum for a perfectly balanced clustering. Hence, clearly, hyperparameter " + }, + { + "bbox": [ + 63, + 519, + 296, + 686 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 63, + 519, + 296, + 686 + ], + "type": "text", + "content": " controls the level of class balance in the solution. As depicted in Fig. 1, this hyper-parameter has a crucial effect on the performance, and its optimal value might vary by orders of magnitude from one dataset to another. This makes exhaustive grid searches for optimal " + }, + { + "bbox": [ + 63, + 519, + 296, + 686 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 63, + 519, + 296, + 686 + ], + "type": "text", + "content": " over validation sets intractable computationally, which motivates learning this hyper-parameter through our unrolling strategy, as described in Section 3.2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 72, + 555, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 555, + 109 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 555, + 109 + ], + "type": "text", + "content": "- The third term in (1) is an entropic barrier, enabling to soften assignments " + }, + { + "bbox": [ + 314, + 72, + 555, + 109 + ], + "type": "inline_equation", + "content": "u_{n,k}" + }, + { + "bbox": [ + 314, + 72, + 555, + 109 + ], + "type": "text", + "content": ", while imposing a non-negativity constraint on each of them. It is given by" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 377, + 117, + 555, + 150 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 117, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 377, + 117, + 555, + 150 + ], + "type": "interline_equation", + "content": "\\Phi (\\boldsymbol {u}) = \\sum_ {n = 1} ^ {N} \\sum_ {k = 1} ^ {K} u _ {n, k} \\ln u _ {n, k}. \\tag {4}", + "image_path": "9acce5ce89091c913b570428661647c1276033bf5eab05ca1139cb66c7762126.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 158, + 555, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 158, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 321, + 158, + 555, + 266 + ], + "type": "text", + "content": "Temperature scaling hyper-parameter " + }, + { + "bbox": [ + 321, + 158, + 555, + 266 + ], + "type": "inline_equation", + "content": "(T)" + }, + { + "bbox": [ + 321, + 158, + 555, + 266 + ], + "type": "text", + "content": ": Weighting factor " + }, + { + "bbox": [ + 321, + 158, + 555, + 266 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 321, + 158, + 555, + 266 + ], + "type": "text", + "content": " in (1) controls the trade-off between the clustering term and the entropic barrier in (4), i.e., the level of softness of assignments " + }, + { + "bbox": [ + 321, + 158, + 555, + 266 + ], + "type": "inline_equation", + "content": "u_{n,k}" + }, + { + "bbox": [ + 321, + 158, + 555, + 266 + ], + "type": "text", + "content": ". Therefore, this hyperparameter has an important effect on performances and on the marginal class probabilities appearing in the class-balance term in (3), which motivates learning it. As described in Section 3.2, our unrolling algorithm enables to cope with this extra parameter." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "text", + "content": "Link to EM and other transductive few-shot methods: We examine solving problem (1) with an iterative block-coordinate descent algorithm (see Algorithm 1), which alternates two steps, one updating distribution parameters " + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_k^{(\\ell)}" + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "text", + "content": " and the other optimizing over class assignments " + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\pmb{u}_n^{(\\ell)}" + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "text", + "content": ", at each iteration " + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "text", + "content": ". For the " + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\pmb{u}" + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "text", + "content": "-step, and due to the nonconvexity of " + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "text", + "content": ", we proceed with a Majorization-Minimization (MM) strategy to minimize a surrogate convex function with respect to " + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\pmb{u}" + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "text", + "content": " at each iteration. In addition, if " + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\mathrm{p}(\\cdot|\\pmb{\\theta}_k)" + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "text", + "content": " is assumed to belong to the exponential family and " + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_k" + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "text", + "content": " are its canonical parameters, the estimation w.r.t. " + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\pmb{u}" + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "text", + "content": " fixed, is a convex problem. We provide further details on this general iterative optimization scheme in Appendix A. It is worth noting that, when " + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "inline_equation", + "content": "T = 1" + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "text", + "content": " and the data is modeled by the Dirichlet distribution, we recover the recent transductive few-shot learning algorithm in [33]. Interestingly, when " + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "inline_equation", + "content": "T = 1" + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\lambda = |\\mathbb{Q}|" + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "text", + "content": ", we recover the well-known Expectation-Maximization (EM) algorithm for estimating the parameters of mixture of distributions; see Proposition 1 in [33]. Therefore, Algorithm 1 could be viewed as a generalized EM (GEM), in which hyper-parameters " + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 266, + 555, + 542 + ], + "type": "text", + "content": " control the class balance as well as prediction softness, and could be made learnable." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 543, + 554, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 554, + 578 + ], + "type": "text", + "content": "Examples of data models: While a broad range of distribution models in the exponential family could be adopted, we focus in this paper on two popular ones." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 578, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 578, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 314, + 578, + 555, + 639 + ], + "type": "text", + "content": "- The first one is the Gaussian distribution, which is commonly used in standard clustering and transductive few-shot-methods applied to vision models [32, 46, 50]. By assuming a Gaussian distribution with mean " + }, + { + "bbox": [ + 314, + 578, + 555, + 639 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_{k}" + }, + { + "bbox": [ + 314, + 578, + 555, + 639 + ], + "type": "text", + "content": " and identity covariance matrix, the pdf " + }, + { + "bbox": [ + 314, + 578, + 555, + 639 + ], + "type": "inline_equation", + "content": "\\mathrm{p}(\\pmb{z}_n|\\pmb{\\theta}_k)" + }, + { + "bbox": [ + 314, + 578, + 555, + 639 + ], + "type": "text", + "content": " reads as follows" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 360, + 647, + 553, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 647, + 553, + 673 + ], + "spans": [ + { + "bbox": [ + 360, + 647, + 553, + 673 + ], + "type": "interline_equation", + "content": "\\mathrm {p} \\left(\\boldsymbol {z} _ {n} \\mid \\boldsymbol {\\theta} _ {k}\\right) \\propto \\exp \\left(- \\frac {1}{2} \\left\\| \\boldsymbol {z} _ {n} - \\boldsymbol {\\theta} _ {k} \\right\\| ^ {2}\\right). \\tag {5}", + "image_path": "b3df29b1b6f4efe28332d750d94025e75eed83c88456349f8e34c923a4d1d43b.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 314, + 677, + 555, + 713 + ], + "type": "text", + "content": "- The second one is the Dirichlet distribution, which has recently shown good modelling performance in the context of transductive few-shot for vision-language mod" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 714 + ], + "type": "inline_equation", + "content": "^{2}\\mathrm{K}" + }, + { + "bbox": [ + 55, + 693, + 295, + 714 + ], + "type": "text", + "content": "-means corresponds to choosing the multivariate Gaussian distribution, with identity covariance matrix, for parametric density " + }, + { + "bbox": [ + 55, + 693, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{p}(\\mathbf{z}_n|\\boldsymbol{\\theta}_k)" + }, + { + "bbox": [ + 55, + 693, + 295, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9668" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 293, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 293, + 84 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 293, + 84 + ], + "type": "text", + "content": "Algorithm 1 GEM based few-shot classification algorithm" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 88, + 481, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 88, + 481, + 114 + ], + "spans": [ + { + "bbox": [ + 64, + 88, + 481, + 114 + ], + "type": "text", + "content": "Input: Compute " + }, + { + "bbox": [ + 64, + 88, + 481, + 114 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_n" + }, + { + "bbox": [ + 64, + 88, + 481, + 114 + ], + "type": "text", + "content": " for the dataset samples, initialize " + }, + { + "bbox": [ + 64, + 88, + 481, + 114 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_n^{(0)}" + }, + { + "bbox": [ + 64, + 88, + 481, + 114 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 64, + 88, + 481, + 114 + ], + "type": "inline_equation", + "content": "\\theta_k^{(0)}" + }, + { + "bbox": [ + 64, + 88, + 481, + 114 + ], + "type": "text", + "content": ", and fix the number of iterations " + }, + { + "bbox": [ + 64, + 88, + 481, + 114 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 64, + 88, + 481, + 114 + ], + "type": "text", + "content": ", for " + }, + { + "bbox": [ + 64, + 88, + 481, + 114 + ], + "type": "inline_equation", + "content": "\\ell = 0,1,\\ldots ,L - 1" + }, + { + "bbox": [ + 64, + 88, + 481, + 114 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 75, + 115, + 520, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 115, + 520, + 126 + ], + "spans": [ + { + "bbox": [ + 75, + 115, + 520, + 126 + ], + "type": "text", + "content": "// Update Distribution parameters for each class using a given estimation algorithm (denoted here by \"DP_est\")" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 76, + 126, + 280, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 126, + 280, + 142 + ], + "spans": [ + { + "bbox": [ + 76, + 126, + 280, + 142 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\theta} _ {k} ^ {(\\ell + 1)} = \\operatorname {D P} _ {-} \\operatorname {e s t} (\\boldsymbol {u} _ {., k} ^ {(\\ell)}, \\boldsymbol {\\theta} _ {k} ^ {(\\ell)}), \\quad \\forall k \\in \\{1, \\dots , K \\},", + "image_path": "b2243b30eea7022c4a314c58e73f1479f0ae03af004b7b5ec0d9af307e972629.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 76, + 142, + 186, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 142, + 186, + 152 + ], + "spans": [ + { + "bbox": [ + 76, + 142, + 186, + 152 + ], + "type": "text", + "content": "// Update class proportions" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 76, + 152, + 258, + 179 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 152, + 258, + 179 + ], + "spans": [ + { + "bbox": [ + 76, + 152, + 258, + 179 + ], + "type": "interline_equation", + "content": "\\pi_ {k} ^ {(\\ell + 1)} = \\frac {1}{| \\mathbb {Q} |} \\sum_ {n \\in \\mathbb {Q}} u _ {n, k} ^ {(\\ell)}, \\quad \\forall k \\in \\{1, \\dots , K \\},", + "image_path": "1a77b4ab23e17a0f459687b2d342f3eacb4ed2de3594e31e428201e1590e92c4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 76, + 179, + 280, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 179, + 280, + 190 + ], + "spans": [ + { + "bbox": [ + 76, + 179, + 280, + 190 + ], + "type": "text", + "content": "// Update assignment vectors for all query samples" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 190, + 392, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 190, + 392, + 217 + ], + "spans": [ + { + "bbox": [ + 76, + 190, + 392, + 217 + ], + "type": "interline_equation", + "content": "\\boldsymbol {u} _ {n} ^ {(\\ell + 1)} = \\operatorname {s o f t m a x} \\left(\\frac {1}{T} \\left(\\ln \\mathrm {p} \\left(\\boldsymbol {z} _ {n} \\mid \\boldsymbol {\\theta} _ {k} ^ {(\\ell + 1)}\\right) + \\frac {\\lambda}{| \\mathbb {Q} |} \\ln (\\pi_ {k} ^ {(\\ell + 1)})\\right) _ {k}\\right), \\quad \\forall n \\in \\mathbb {Q}.", + "image_path": "0ca653e57115ac783bd9bd12132544daf918d4aa04eef510b280f448dd0e0a8a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 215, + 99, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 215, + 99, + 224 + ], + "spans": [ + { + "bbox": [ + 66, + 215, + 99, + 224 + ], + "type": "text", + "content": "end for" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 130, + 241, + 479, + 346 + ], + "blocks": [ + { + "bbox": [ + 130, + 241, + 479, + 346 + ], + "lines": [ + { + "bbox": [ + 130, + 241, + 479, + 346 + ], + "spans": [ + { + "bbox": [ + 130, + 241, + 479, + 346 + ], + "type": "image", + "image_path": "638c5b2d4c01adbc9b9d282a46011431ac1796f80a8748d1e7a7b18db5dcad27.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 350, + 555, + 374 + ], + "lines": [ + { + "bbox": [ + 55, + 350, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 555, + 374 + ], + "type": "text", + "content": "Figure 2. An overview of the unrolled GEM algorithm for a given iteration. Each iteration " + }, + { + "bbox": [ + 55, + 350, + 555, + 374 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 55, + 350, + 555, + 374 + ], + "type": "text", + "content": " corresponds to a network layer " + }, + { + "bbox": [ + 55, + 350, + 555, + 374 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{(\\ell)}" + }, + { + "bbox": [ + 55, + 350, + 555, + 374 + ], + "type": "text", + "content": ". Each layer depends on the vector of hyperparameters " + }, + { + "bbox": [ + 55, + 350, + 555, + 374 + ], + "type": "inline_equation", + "content": "(\\lambda^{(\\ell)}, T^{(\\ell)})" + }, + { + "bbox": [ + 55, + 350, + 555, + 374 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 394, + 295, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 394, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 63, + 394, + 295, + 420 + ], + "type": "text", + "content": "els such as CLIP [33]. For " + }, + { + "bbox": [ + 63, + 394, + 295, + 420 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_n = (z_{n,i})_{1\\leq i\\leq K}" + }, + { + "bbox": [ + 63, + 394, + 295, + 420 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 63, + 394, + 295, + 420 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_{k} = (\\theta_{k,i})_{1\\leq i\\leq K}" + }, + { + "bbox": [ + 63, + 394, + 295, + 420 + ], + "type": "text", + "content": ", the associated pdf is given by" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 92, + 427, + 295, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 427, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 92, + 427, + 295, + 460 + ], + "type": "interline_equation", + "content": "\\mathrm {p} \\left(\\boldsymbol {z} _ {n} \\mid \\boldsymbol {\\theta} _ {k}\\right) = \\frac {1}{\\mathcal {B} \\left(\\boldsymbol {\\theta} _ {k}\\right)} \\prod_ {i = 1} ^ {K} z _ {n, i} ^ {\\theta_ {k, i} - 1} \\mathbb {1} _ {\\boldsymbol {z} _ {n} \\in \\Delta_ {K}}, \\tag {6}", + "image_path": "a8003e167b160719bf60aa939326f5f9bbb77524c5ac433e43209a45b4b34d33.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 64, + 469, + 228, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 469, + 228, + 482 + ], + "spans": [ + { + "bbox": [ + 64, + 469, + 228, + 482 + ], + "type": "text", + "content": "where the normalization factor " + }, + { + "bbox": [ + 64, + 469, + 228, + 482 + ], + "type": "inline_equation", + "content": "\\mathcal{B}(\\pmb{\\theta}_k)" + }, + { + "bbox": [ + 64, + 469, + 228, + 482 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 125, + 489, + 295, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 489, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 125, + 489, + 295, + 525 + ], + "type": "interline_equation", + "content": "\\mathcal {B} \\left(\\boldsymbol {\\theta} _ {k}\\right) = \\frac {\\prod_ {i = 1} ^ {K} \\Gamma \\left(\\theta_ {k , i}\\right)}{\\Gamma \\left(\\sum_ {i = 1} ^ {K} \\theta_ {k , i}\\right)}, \\tag {7}", + "image_path": "e7ac9e7401ce8dbe7985bf8b272b9fb6acc6df4ed4c12d7c23f9759204528589.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 64, + 533, + 211, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 533, + 211, + 544 + ], + "spans": [ + { + "bbox": [ + 64, + 533, + 211, + 544 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 64, + 533, + 211, + 544 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 64, + 533, + 211, + 544 + ], + "type": "text", + "content": " denotes the Gamma function." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 545, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 605 + ], + "type": "text", + "content": "The necessary details related to feature representations in the cases of vision-only and vision-language models, distribution parameter estimation, and the resulting GEM algorithms for Gaussian and Dirichlet laws are provided in Appendices B and C, respectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 612, + 246, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 246, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 246, + 624 + ], + "type": "text", + "content": "3.2. UNrolled EM architecture (UNEM)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": "Overview: As discussed in Section 2, the choice of the optimal hyper-parameter, which controls the level of class balance, is a difficult task, often performed manually in the existing transductive few-shot classification methods. In addition to the class-balance parameter " + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": ", Algorithm 1 involves an additional temperature parameter " + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": ", which makes hyperparameter setting even more challenging. For this reason," + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 394, + 555, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 555, + 610 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 555, + 610 + ], + "type": "text", + "content": "we propose in this paper to resort to the unrolling (called also learning to optimize) paradigm [35], which enables to learn efficiently a set of optimized hyper-parameters. It is important to note that unrolling iterative optimization algorithms have found successful applications in diverse signal and image processing tasks [2, 7, 35]. However, to the best of our knowledge, this work is the first to leverage the unrolling paradigm for hyper-parameter optimization in a few-shot learning context. The main idea behind the learning-to-optimize paradigm is to map each iteration of a given optimization algorithm to a network layer, stack all layers together, and view the hyper-parameters to be optimized as the network's learnable parameters. More precisely, to unroll our generalized EM algorithm, the number of iterations " + }, + { + "bbox": [ + 313, + 394, + 555, + 610 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 394, + 555, + 610 + ], + "type": "text", + "content": " is used as the number of layers for the neural network architecture. Thus, each iteration " + }, + { + "bbox": [ + 313, + 394, + 555, + 610 + ], + "type": "inline_equation", + "content": "\\ell \\in \\{0,1,\\dots ,L - 1\\}" + }, + { + "bbox": [ + 313, + 394, + 555, + 610 + ], + "type": "text", + "content": " is associated with a tailored layer " + }, + { + "bbox": [ + 313, + 394, + 555, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{(\\ell)}" + }, + { + "bbox": [ + 313, + 394, + 555, + 610 + ], + "type": "text", + "content": ", which performs the update rules defined in Algorithm 1:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 623, + 559, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 623, + 559, + 653 + ], + "spans": [ + { + "bbox": [ + 315, + 623, + 559, + 653 + ], + "type": "interline_equation", + "content": "\\left(\\boldsymbol {\\theta} _ {k} ^ {(\\ell + 1)}, \\pi_ {k} ^ {(\\ell + 1)}, \\boldsymbol {u} _ {n} ^ {(\\ell + 1)}\\right) = \\mathcal {L} ^ {(\\ell)} \\left(\\boldsymbol {\\theta} _ {k} ^ {(\\ell)}, \\pi_ {k} ^ {(\\ell)}, \\boldsymbol {u} _ {n} ^ {(\\ell)}; \\lambda^ {(\\ell)}, T ^ {(\\ell)}\\right) \\tag {8}", + "image_path": "78427b12b0d18cd72c5f1b727585e5d15c62ff3a1e2f1e5baf91131c3720c492.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "inline_equation", + "content": "(\\lambda^{(\\ell)}, T^{(\\ell)})_{0 \\leq \\ell \\leq L-1}" + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": " is the vector of hyperparameters to be learned. This leads to the UNrolled EM (UNEM) model shown in Fig. 2, which depicts the inputs and outputs of a given layer " + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{(l)}" + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": " as well as its three main blocks for the update rules." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9669" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 129, + 70, + 479, + 189 + ], + "blocks": [ + { + "bbox": [ + 129, + 70, + 479, + 189 + ], + "lines": [ + { + "bbox": [ + 129, + 70, + 479, + 189 + ], + "spans": [ + { + "bbox": [ + 129, + 70, + 479, + 189 + ], + "type": "image", + "image_path": "702073904ba52b0407f7950ed818c705d2325a204fe9a923703181ba12013a5a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 191, + 404, + 201 + ], + "lines": [ + { + "bbox": [ + 205, + 191, + 404, + 201 + ], + "spans": [ + { + "bbox": [ + 205, + 191, + 404, + 201 + ], + "type": "text", + "content": "Figure 3. Overall architecture of the designed UNEM." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 223, + 295, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 223, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 54, + 223, + 295, + 330 + ], + "type": "text", + "content": "Learned hyper-parameters: Instead of a restricted number of handcrafted hyper-parameters (as often used in iterative algorithms), our unrolled algorithm offers more flexibility, enabling adaptation of the hyper-parameters along the processing workflow. This means that " + }, + { + "bbox": [ + 54, + 223, + 295, + 330 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 54, + 223, + 295, + 330 + ], + "type": "text", + "content": " vectors of hyper-parameters " + }, + { + "bbox": [ + 54, + 223, + 295, + 330 + ], + "type": "inline_equation", + "content": "(\\lambda^{(\\ell)}, T^{(\\ell)})_{0 \\leq \\ell \\leq L - 1}" + }, + { + "bbox": [ + 54, + 223, + 295, + 330 + ], + "type": "text", + "content": " could be learned and applied at each layer of the network. During the training of the unrolled model, we fulfill some design constraints, as described in more details in the following." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 331, + 296, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 379 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 379 + ], + "type": "text", + "content": "More specifically, to ensure the non-negativity of hyperparameter " + }, + { + "bbox": [ + 55, + 331, + 296, + 379 + ], + "type": "inline_equation", + "content": "\\lambda^{(\\ell)}" + }, + { + "bbox": [ + 55, + 331, + 296, + 379 + ], + "type": "text", + "content": ", we propose to express it as a Softplus function, which could be seen as a smooth approximation of the RELU activation function, yielding" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 384, + 294, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 384, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 85, + 384, + 294, + 399 + ], + "type": "interline_equation", + "content": "\\lambda^ {(\\ell)} = \\operatorname {S o f t p l u s} \\left(a ^ {(\\ell)}\\right) = \\log \\left(1 + \\exp \\left(a ^ {(\\ell)}\\right)\\right) \\tag {9}", + "image_path": "034826c34068308884f29a4fba3ae2c3de73b3520ac1af558286f102f9d45828.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 406, + 295, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 295, + 429 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 406, + 295, + 429 + ], + "type": "inline_equation", + "content": "a^{(\\ell)}" + }, + { + "bbox": [ + 55, + 406, + 295, + 429 + ], + "type": "text", + "content": " represents a learnable parameter for the unrolled architecture." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 430, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 296, + 514 + ], + "type": "text", + "content": "Regarding temperature scaling " + }, + { + "bbox": [ + 55, + 430, + 296, + 514 + ], + "type": "inline_equation", + "content": "T^{(\\ell)}" + }, + { + "bbox": [ + 55, + 430, + 296, + 514 + ], + "type": "text", + "content": ", we observed in our experiments that imposing the non-negativity constraint alone might be a source of instability, resulting in values very close to zero and vanishing gradient issues during the training of the unrolled architecture. To circumvent this problem, we impose a lower bound equal to 1 on the temperature scaling:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 60, + 520, + 295, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 520, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 60, + 520, + 295, + 536 + ], + "type": "interline_equation", + "content": "T ^ {(\\ell)} = 1 + \\operatorname {S o f t p l u s} (b ^ {(\\ell)}) = 1 + \\log (1 + \\exp (b ^ {(\\ell)})) \\tag {10}", + "image_path": "add6b6d779495a789c0e9de86a3e29817edaadac8b24588dd740db93168ef73b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 542, + 295, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 295, + 566 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 542, + 295, + 566 + ], + "type": "inline_equation", + "content": "b^{(\\ell)}" + }, + { + "bbox": [ + 55, + 542, + 295, + 566 + ], + "type": "text", + "content": " denotes a parameter that needs to be learned during the training of the unrolled model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 566, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 295, + 689 + ], + "type": "text", + "content": "Overall architecture and training approach: Based on the previous considerations, the overall architecture of UNEM can be summarized as the composition of the " + }, + { + "bbox": [ + 55, + 566, + 295, + 689 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 566, + 295, + 689 + ], + "type": "text", + "content": " layers " + }, + { + "bbox": [ + 55, + 566, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{(L - 1)}\\circ \\ldots \\circ \\mathcal{L}^{(0)}" + }, + { + "bbox": [ + 55, + 566, + 295, + 689 + ], + "type": "text", + "content": ". This architecture, depicted in Figure 3, illustrates: (i) the required inputs " + }, + { + "bbox": [ + 55, + 566, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_k^{(0)}" + }, + { + "bbox": [ + 55, + 566, + 295, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 566, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\pmb{u}_n^{(0)}" + }, + { + "bbox": [ + 55, + 566, + 295, + 689 + ], + "type": "text", + "content": " as well as the feature vectors " + }, + { + "bbox": [ + 55, + 566, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\pmb{z}_n" + }, + { + "bbox": [ + 55, + 566, + 295, + 689 + ], + "type": "text", + "content": " (which are fed in all layers), (ii) the cascaded layers with their associated vector of parameters " + }, + { + "bbox": [ + 55, + 566, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\pmb{w} = (\\lambda^{(\\ell)},T^{(\\ell)})_{0\\leq \\ell \\leq L - 1}" + }, + { + "bbox": [ + 55, + 566, + 295, + 689 + ], + "type": "text", + "content": " satisfying the aforementioned constraints, and (iii) the key output " + }, + { + "bbox": [ + 55, + 566, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\pmb{u}_n^{(L)}" + }, + { + "bbox": [ + 55, + 566, + 295, + 689 + ], + "type": "text", + "content": " representing the class assignment vectors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "The resulting neural network architecture is trained by minimizing a standard cross-entropy loss on a validation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 224, + 332, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 224, + 332, + 234 + ], + "spans": [ + { + "bbox": [ + 314, + 224, + 332, + 234 + ], + "type": "text", + "content": "set:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 365, + 235, + 555, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 235, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 365, + 235, + 555, + 269 + ], + "type": "interline_equation", + "content": "\\mathrm {L} _ {\\mathrm {c}} (\\boldsymbol {w}) = \\sum_ {n \\in \\mathbb {Q}} \\sum_ {k = 1} ^ {K} y _ {n, k} \\log \\left(u _ {n, k} ^ {(L)}\\right). \\tag {11}", + "image_path": "019dd278dc426e1bd66880f9ebce96fd02715107028d232b60cda5284f772719.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 281, + 395, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 395, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 395, + 294 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 301, + 555, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 555, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 555, + 385 + ], + "type": "text", + "content": "The effectiveness of the proposed approach is validated in both vision-only and vision-language transductive few-shot learning settings. We will designate our proposed unrolled methods by UNEM-Gaussian and UNEM-Dirichlet, respectively. After describing the experimental settings, this section is structured into two main sections discussing our experiments in both evaluation scenarios." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 396, + 440, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 440, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 440, + 408 + ], + "type": "text", + "content": "4.1. Experimental settings" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 415, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 570 + ], + "type": "text", + "content": "Task generation: We adopt a realistic transductive few-shot evaluation protocol, which is in line with state-of-the-art protocols [29, 32, 33, 45, 50]. While " + }, + { + "bbox": [ + 313, + 415, + 555, + 570 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 415, + 555, + 570 + ], + "type": "text", + "content": " designates the total number of classes in the labeled support set " + }, + { + "bbox": [ + 313, + 415, + 555, + 570 + ], + "type": "inline_equation", + "content": "\\mathbb{S}" + }, + { + "bbox": [ + 313, + 415, + 555, + 570 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 415, + 555, + 570 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{eff}}" + }, + { + "bbox": [ + 313, + 415, + 555, + 570 + ], + "type": "text", + "content": " (with " + }, + { + "bbox": [ + 313, + 415, + 555, + 570 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{eff}} \\ll K" + }, + { + "bbox": [ + 313, + 415, + 555, + 570 + ], + "type": "text", + "content": ") denotes the number of effective classes present in the unlabeled query set " + }, + { + "bbox": [ + 313, + 415, + 555, + 570 + ], + "type": "inline_equation", + "content": "\\mathbb{Q}" + }, + { + "bbox": [ + 313, + 415, + 555, + 570 + ], + "type": "text", + "content": ". The classes in the query set remain undisclosed during inference, while randomly selecting " + }, + { + "bbox": [ + 313, + 415, + 555, + 570 + ], + "type": "inline_equation", + "content": "|\\mathbb{Q}|" + }, + { + "bbox": [ + 313, + 415, + 555, + 570 + ], + "type": "text", + "content": " samples. On the other hand, the support set is built by uniformly selecting " + }, + { + "bbox": [ + 313, + 415, + 555, + 570 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 415, + 555, + 570 + ], + "type": "text", + "content": " images from each of the " + }, + { + "bbox": [ + 313, + 415, + 555, + 570 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 415, + 555, + 570 + ], + "type": "text", + "content": " classes. In this paper, the few-shot tasks are performed with (i) 5, 10, and 20 shots for vision-only models, and (ii) 4 shots for vision-language models. Moreover, in both of our evaluation scenarios, we used " + }, + { + "bbox": [ + 313, + 415, + 555, + 570 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{eff}} = 5" + }, + { + "bbox": [ + 313, + 415, + 555, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 415, + 555, + 570 + ], + "type": "inline_equation", + "content": "|\\mathbb{Q}| = 75" + }, + { + "bbox": [ + 313, + 415, + 555, + 570 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 570, + 556, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 677 + ], + "type": "text", + "content": "Training specifications: For fair comparison and reproducibility purposes, the CLIP's pre-trained model has been directly used in the vision-language evaluation scenario, while for the vision-only scenario, the standard pre-trained ResNet-18 and WRN28-10 backbones have been finetuned. Their training is performed on the base classes set of each vision-only dataset, using cross-entropy loss with label smoothing set to 0.9 for 90 epochs. The learning rate is set to 0.1 and then decayed by a factor of 10." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "On the other hand, the training of our unrolled architecture, composed of 10 layers (i.e., " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "L = 10" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "), is performed on several tasks sampled from the validation set of each dataset." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9670" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 142, + 69, + 470, + 304 + ], + "blocks": [ + { + "bbox": [ + 142, + 69, + 470, + 304 + ], + "lines": [ + { + "bbox": [ + 142, + 69, + 470, + 304 + ], + "spans": [ + { + "bbox": [ + 142, + 69, + 470, + 304 + ], + "type": "table", + "html": "
MethodBackbonemini-ImageNet (K=20)tiered-ImageNet (K=160)
5-shot10-shot20-shot5-shot10-shot20-shot
Baseline [8]55.462.167.929.736.342.2
LR+ICI [52]55.462.168.1---
BD-CSPN [28]49.854.656.511.411.011.7
PT-MAP [17]25.727.228.45.26.06.6
LaplacianShot [60]ResNet-1857.964.268.329.635.439.1
TIM [4]66.869.970.829.328.727.8
α-TIM [50]66.771.073.943.848.351.9
α-AM [24]64.467.870.1---
PADDLE [32]62.973.579.845.461.470.6
UNEM-Gaussian66.475.680.452.365.773.2
Baseline [8]59.065.772.131.939.045.6
LR+ICI [52]58.865.772.0---
BD-CSPN [28]51.155.558.418.018.418.2
PT-MAP [17]26.528.029.35.26.06.6
LaplacianShot [60]WRN28-1061.066.871.031.437.341.5
TIM [4]72.174.976.236.139.038.5
α-TIM [50]71.575.278.345.851.455.2
α-AM [24]68.271.373.3---
PADDLE [32]62.673.079.243.959.469.9
UNEM-Gaussian71.679.283.754.166.874.7
", + "image_path": "816412569d426a3b8f7d37e2b4664d54ce997a11df573309a78e41f0362787d8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 89, + 352, + 264, + 473 + ], + "blocks": [ + { + "bbox": [ + 55, + 312, + 555, + 334 + ], + "lines": [ + { + "bbox": [ + 55, + 312, + 555, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 555, + 334 + ], + "type": "text", + "content": "Table 1. Comparison of the proposed UNEM-Gaussian with respect to state-of-the-art methods on mini-Imagenet and tiered-Imagenet. The metric is accuracy (in percentage). Results are averaged across 1,000 tasks. Results marked with ’ - ’ were intractable to obtain." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 89, + 352, + 264, + 473 + ], + "lines": [ + { + "bbox": [ + 89, + 352, + 264, + 473 + ], + "spans": [ + { + "bbox": [ + 89, + 352, + 264, + 473 + ], + "type": "table", + "html": "
MethodCUB (K=50)
5-shot10-shot20-shot
Baseline [8]58.668.878.2
LR+ICI [52]49.955.658.0
PT-MAP [17]12.814.014.9
LaplacianShot [60]58.866.571.0
TIM [4]68.168.969.3
α-TIM [50]74.379.383.6
α-AM [24]66.268.969.8
PADDLE [32]71.281.886.8
UNEM-Gaussian78.585.388.6
", + "image_path": "83dbb3e6233185346c6cd735c25dcf1ebda60169ad5d7d4f124d490f4fc2168d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 480, + 295, + 514 + ], + "lines": [ + { + "bbox": [ + 55, + 480, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 295, + 514 + ], + "type": "text", + "content": "Table 2. Comparison of the proposed UNEM-Gaussian with respect to state-of-the-art methods on CUB. The metric is accuracy (in percentage). Results are averaged across 1,000 tasks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "This implies that the proposed solution is highly economical in terms of parameters, typically requiring only a few tenths, compared to standard neural network architectures. Let us recall that the validation set is often used to select the hyperparameters in recent state-of-the-art transductive methods as mentioned in Section 2. Specifically, for the vision-only (resp. vision-language) models, the training of the unrolled architecture is carried out on 1,000 (resp. 100) tasks, while using an initial learning rate of 0.1 (resp. 0.5) with a decay factor of 0.5, a number of epochs equals to 80, and ADAM as optimizer. The trainable hyper-parameters are initialized to the default values set in [32] (i.e., " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\lambda = |Q|" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " for the vision-only models) and [33] (i.e., " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\lambda = \\frac{K}{K_{\\mathrm{eff}}}" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " for the vision-language models, with " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "T = 1" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "). Our architecture is" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 355, + 555, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 555, + 379 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 555, + 379 + ], + "type": "text", + "content": "implemented in Pytorch (version 2.3.0) and run on NVIDIA QUADRO RTX8000 (with 48 GB of memory)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 384, + 555, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 398 + ], + "type": "text", + "content": "4.2. UNEM-Gaussian in vision-only few-shot setting" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 403, + 554, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 554, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 554, + 461 + ], + "type": "text", + "content": "Let us recall that the Gaussian distribution is commonly used in standard clustering and transductive few-shot methods applied to vision-only models [32, 46, 50]. For this reason, we will focus here on the proposed UNEM-Gaussian approach." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 462, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 629 + ], + "type": "text", + "content": "Datasets: The first UNEM-Gaussian architecture is evaluated on the following standard few-shot benchmark datasets: mini-[ImageNet [51], tiered-[ImageNet [41], CUB [14]. Mini-imagenet has 100 classes split into 64 base classes, 16 validation classes and 20 test classes. The tiered-imagenet has 608 classes instead, from which we follow a standard split with 351 for base training, 97 for validation and 160 for testing. For CUB, we followed the split proposed by [8] which consists of 100 base classes, 50 validation classes and 50 test classes. For each dataset, the feature vectors " + }, + { + "bbox": [ + 313, + 462, + 554, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_n" + }, + { + "bbox": [ + 313, + 462, + 554, + 629 + ], + "type": "text", + "content": " are extracted using the fine-tuned backbones, while applying a scaling parameter " + }, + { + "bbox": [ + 313, + 462, + 554, + 629 + ], + "type": "inline_equation", + "content": "T_{z}" + }, + { + "bbox": [ + 313, + 462, + 554, + 629 + ], + "type": "text", + "content": " to the model's output as described in Appendix B. This parameter has also been learned through the unrolling approach." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "Results: The UNEM-Gaussian architecture has been compared to its original version PADDLE [32] as well as several state-of-the-art methods. Tables 1 and 2 presents the results, averaged over 1,000 tasks with 5, 10, and 20 shots. Several observations can be made. First, the proposed UNEM-Gaussian outperforms the original PADDLE algorithm [32] and the other state-of-the-art methods. The achieved gains" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9671" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 88, + 71, + 523, + 192 + ], + "blocks": [ + { + "bbox": [ + 88, + 71, + 523, + 192 + ], + "lines": [ + { + "bbox": [ + 88, + 71, + 523, + 192 + ], + "spans": [ + { + "bbox": [ + 88, + 71, + 523, + 192 + ], + "type": "table", + "html": "
MethodFood101EuroSATDTDOrionPedsFlowers102Caltech101UCF101FGVC AircraftStanford CarsSUM397ImageNetAverage
Ind.Tip-Adapter [56]76.772.554.786.483.288.872.123.763.966.762.768.3
CoOp [57]76.363.252.286.281.087.767.022.261.363.459.965.5
Transd.BDSCPN [28]74.746.145.281.374.282.059.018.048.154.549.257.5
Laplacian Shot [60]76.653.052.688.485.586.867.022.260.463.856.364.8
α-TIM [50]66.146.145.387.179.183.359.420.453.453.442.757.8
PADDLE [32]71.845.950.084.782.381.963.721.356.160.652.160.9
EM-Dirichlet [33]88.750.862.692.591.390.176.124.973.580.978.473.6
UNEM-Dirichlet91.453.865.396.095.693.478.530.480.088.583.177.8
", + "image_path": "dcc3a6804b601dd9a08d204180d15d465b210c08383591e6d8e56d0d8df918e7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 198, + 553, + 222 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 553, + 222 + ], + "type": "text", + "content": "Table 3. Comparison of the proposed UNEM-Dirichlet with respect to the state-of-the-art methods on 11 different datasets. The metric is accuracy (in percentage). Results are averaged across 1,000 tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 56, + 243, + 296, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 296, + 411 + ], + "type": "text", + "content": "are much higher with fewer number of shots and reach " + }, + { + "bbox": [ + 56, + 243, + 296, + 411 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 56, + 243, + 296, + 411 + ], + "type": "text", + "content": " with mini-ImageNet, " + }, + { + "bbox": [ + 56, + 243, + 296, + 411 + ], + "type": "inline_equation", + "content": "6.9\\%" + }, + { + "bbox": [ + 56, + 243, + 296, + 411 + ], + "type": "text", + "content": " with tiered-ImageNet, and " + }, + { + "bbox": [ + 56, + 243, + 296, + 411 + ], + "type": "inline_equation", + "content": "7.3\\%" + }, + { + "bbox": [ + 56, + 243, + 296, + 411 + ], + "type": "text", + "content": " with CUB, while using ResNet-18 as backbone. Moreover, when WRN28-10 is used as backbone, it can be noticed that the performance of all methods has been improved, except for PADDLE, which shows results similar to those obtained with ResNet-18 pre-training. This can be explained by the fact that the class-balance parameter " + }, + { + "bbox": [ + 56, + 243, + 296, + 411 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 56, + 243, + 296, + 411 + ], + "type": "text", + "content": " was set in [32] to " + }, + { + "bbox": [ + 56, + 243, + 296, + 411 + ], + "type": "inline_equation", + "content": "|\\mathbb{Q}|" + }, + { + "bbox": [ + 56, + 243, + 296, + 411 + ], + "type": "text", + "content": " (i.e. 75), which becomes sub-optimal with WRN28-10 features. Most importantly, by learning efficiently the hyperparameters, our UNEM-Gaussian achieves higher gain with respect to its original version, yielding an accuracy gain reaching up to " + }, + { + "bbox": [ + 56, + 243, + 296, + 411 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 56, + 243, + 296, + 411 + ], + "type": "text", + "content": " in 5-shot scenario, when WRN28-10 is used as backbone." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 419, + 246, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 246, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 246, + 431 + ], + "type": "text", + "content": "4.3. UNEM-Dirichlet for few-shot CLIP" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 437, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 498 + ], + "type": "text", + "content": "Dirichlet distribution has recently demonstrated good modelling performance in the context of transductive few-shot for vision-language models such as CLIP [33]. Thus, we will focus in this second round of experiments on UNEM-Dirichlet approach." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "content": "Datasets: The second unrolled architecture, designated by UNEM-Dirichlet, is assessed on 11 benchmark datasets that are commonly used for CLIP scenario: Caltech101 [10], OxfordPnts [37], StanfordCars [22], Flowers102 [36], Food101 [3], FGVCAircraft [31], SUN397 [54], DTD [9], EuroSAT [15], UCF101 [47], and ImageNet [42]. These datasets cover diverse classification challenges, from object recognition (Caltech101, Food101) to fine-grained tasks (StanfordCars, FGVCAircraft) and scene understanding (SUN397, EuroSAT). For each dataset, the vision-text feature vectors " + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "inline_equation", + "content": "z_{n}" + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "content": " are extracted using the CLIP's pre-trained model, while applying a temperature parameter " + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "inline_equation", + "content": "T_{z}" + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "content": " as described in Appendix C. This temperature parameter has also been learned through the unrolling approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "Results: The proposed UNEM-Dirichlet has also been compared to its original EM-Dirichlet version [33] as well as different few-shot classification methods. Table 3 depicts the accuracy results evaluated over 1,000 tasks with 4-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 243, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 460 + ], + "type": "text", + "content": "shots. Thus, it can be noticed that UNEM-Dirichlet outperforms state-of-the-art methods for most datasets. In particular, unrolling the recent EM-Dirichlet algorithm yields further improvement of about " + }, + { + "bbox": [ + 313, + 243, + 555, + 460 + ], + "type": "inline_equation", + "content": "4.2\\%" + }, + { + "bbox": [ + 313, + 243, + 555, + 460 + ], + "type": "text", + "content": " in average. The achieved gain is more significant (reaching up to " + }, + { + "bbox": [ + 313, + 243, + 555, + 460 + ], + "type": "inline_equation", + "content": "7.5\\%" + }, + { + "bbox": [ + 313, + 243, + 555, + 460 + ], + "type": "text", + "content": ") with more challenging datasets having a large number of classes like FGVC Aircraft, Stanford Cars, and SUN397. This is due to the inappropriate choice of the hyperparameter " + }, + { + "bbox": [ + 313, + 243, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 243, + 555, + 460 + ], + "type": "text", + "content": " for these datasets. Indeed, the selected " + }, + { + "bbox": [ + 313, + 243, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 243, + 555, + 460 + ], + "type": "text", + "content": " parameter was set to " + }, + { + "bbox": [ + 313, + 243, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\frac{K}{K_{\\mathrm{eff}}} |\\mathbb{Q}|" + }, + { + "bbox": [ + 313, + 243, + 555, + 460 + ], + "type": "text", + "content": " in EM-Dirichlet approach [33]. Thus, by computing the numerical values of the hyperparameter for each of the aforementioned datasets, it can be deduced from Figure 1 that the difference between the selected value (which is 3,000 for Stanford Cars) and the optimal one (which is around 5,500 for Stanford Cars) is more important and impacts significantly the accuracy performance. This problem is well addressed by resorting to the proposed unrolled model for hyperparameters optimization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 460, + 555, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 509 + ], + "type": "text", + "content": "Appendix D includes additional results to illustrate, in both scenarios, the effects of the temperature parameter in this framework, and to show the benefits of learning variable hyperparameters across the architecture layers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 524, + 388, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 524, + 388, + 537 + ], + "spans": [ + { + "bbox": [ + 314, + 524, + 388, + 537 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 545, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 712 + ], + "type": "text", + "content": "In this paper, we focus on transductive few-shot methods grounded on a generalized form of the EM algorithm. These methods include a parameter to control class balance. The proposed GEM algorithm, applicable to any mixture of distributions, incorporates also a temperature scaling parameter. The optimization process is then unrolled into a neural network architecture, enabling efficient learning of the introduced hyper-parameters. The results demonstrate the effectiveness of the proposed approach on both vision-only and vision-language models. In future work, it would be valuable to explore the potential of unrolling techniques across a broader range of computer vision tasks, especially with the recent rise of foundational vision-language models." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9672" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Christina Baek, Yiding Jiang, Aditi Raghunathan, and Zico Kolter. Agreement-on-the-line: Predicting the performance of neural networks under distribution shift. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), pages 19274-19289, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "text", + "content": "[2] C. Bertocchi, E. Chouzenoux, M.-C. Corbineau, J.-C. Pesquet, and M. Prato. Deep unfolding of a proximal interior point method for image restoration. Inverse Problems, Special Issue on Variational Methods and Effective Algorithms for Imaging and Vision, 36(3):1-27, 2020. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 204, + 295, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 204, + 295, + 247 + ], + "spans": [ + { + "bbox": [ + 62, + 204, + 295, + 247 + ], + "type": "text", + "content": "[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In European Conference on Computer Vision, pages 446-461, Zurich, Switzerland, 2014. Springer. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 249, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 249, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 249, + 295, + 304 + ], + "type": "text", + "content": "[4] Malik Boudiaf, Ziko Imtiaz Masud, Jérôme Rony, José Dolz, Pablo Piantanida, and Ismail Ben Ayed. Transductive information maximization for few-shot learning. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), pages 2445–2457, 2020. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 304, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 304, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 304, + 295, + 350 + ], + "type": "text", + "content": "[5] Y. Boykov, H. N. Isack, C. Olsson, and I. Ben Ayed. Volumetric bias in segmentation and reconstruction: Secrets and solutions. In IEEE/CVF International Conference on Computer Vision (ICCV), 2015. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 351, + 295, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 351, + 295, + 395 + ], + "spans": [ + { + "bbox": [ + 62, + 351, + 295, + 395 + ], + "type": "text", + "content": "[6] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Prompt learning with optimal transport for vision-language models. In International Conference on Learning Representations (ICLR), 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 396, + 295, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 396, + 295, + 439 + ], + "spans": [ + { + "bbox": [ + 62, + 396, + 295, + 439 + ], + "type": "text", + "content": "[7] Tianlong Chen, Xiaohan Chen, Wuyang Chen, and Zhangyang Wang. Learning to optimize: A primer and a benchmark. Journal of Machine Learning Research (JMLR), 23:1-59, 2022. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 441, + 295, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 441, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 441, + 295, + 485 + ], + "type": "text", + "content": "[8] Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, and Jia-Bin Huang. A closer look at few-shot classification. In International Conference on Learning Representations (ICLR), pages 1-19, 2019. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 487, + 295, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 295, + 529 + ], + "type": "text", + "content": "[9] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, and A. Vedaldi. Describing textures in the wild. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 532, + 295, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 532, + 295, + 565 + ], + "spans": [ + { + "bbox": [ + 57, + 532, + 295, + 565 + ], + "type": "text", + "content": "[10] Li Fei-Fei, Robert Fergus, and Pietro Perona. One-shot learning of object categories. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28(4):594-611, 2006. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 566, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 566, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 57, + 566, + 295, + 600 + ], + "type": "text", + "content": "[11] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International Conference on Machine Learning, 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 601, + 295, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 295, + 654 + ], + "type": "text", + "content": "[12] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. CLIP-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 32:581-595, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 656, + 295, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 656, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 57, + 656, + 295, + 700 + ], + "type": "text", + "content": "[13] Bharath Hariharan and Ross B. Girshick. Low-shot visual recognition by shrinking and hallucinating features. In IEEE International Conference on Computer Vision (ICCV), pages 3018-3027, 2017. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 702, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 702, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 702, + 295, + 713 + ], + "type": "text", + "content": "[14] Xiangteng He and Yuxin Peng. Fine-grained visual-textual" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "type": "text", + "content": "representation learning. IEEE Transactions on Circuits and Systems for Video Technology, 30(2):520-531, 2020. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 98, + 554, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 98, + 554, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 98, + 554, + 152 + ], + "type": "text", + "content": "[15] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. EuroSAT: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 154, + 554, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 154, + 554, + 209 + ], + "spans": [ + { + "bbox": [ + 316, + 154, + 554, + 209 + ], + "type": "text", + "content": "[16] Shell Xu Hu, Da Li, Jan Stuhmer, Minyoung Kim, and Timothy M Hospedales. Pushing the limits of simple pipelines for few-shot learning: External data and fine-tuning make a difference. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9068-9077, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 211, + 554, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 211, + 554, + 255 + ], + "spans": [ + { + "bbox": [ + 316, + 211, + 554, + 255 + ], + "type": "text", + "content": "[17] Yuqing Hu, Vincent Gripon, and Stéphane Pateux. Leveraging the feature distribution in transfer-based few-shot learning. In International Conference on Artificial Neural Networks, pages 487-499, 2021. 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 258, + 554, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 258, + 554, + 312 + ], + "spans": [ + { + "bbox": [ + 316, + 258, + 554, + 312 + ], + "type": "text", + "content": "[18] Yunshi Huang, Fereshteh Shakeri, Jose Dolz, Malik Boudiaf, Houda Bahig, and Ismail Ben Ayed. " + }, + { + "bbox": [ + 316, + 258, + 554, + 312 + ], + "type": "inline_equation", + "content": "\\mathrm{LP} + +" + }, + { + "bbox": [ + 316, + 258, + 554, + 312 + ], + "type": "text", + "content": ": A surprisingly strong linear probe for few-shot clip. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23773-23782, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 314, + 554, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 314, + 554, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 314, + 554, + 358 + ], + "type": "text", + "content": "[19] Muhammad Abdullah Jamal and Guo-Jun Qi. Task agnostic meta-learning for few-shot learning. In IEEE conference on Computer Vision and Pattern Recognition (CVPR), pages 11719-11727, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 360, + 554, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 554, + 415 + ], + "type": "text", + "content": "[20] Leonid Karlinsky, Joseph Shtok, Sivan Harary, Eli Schwartz, Amit Aides, and Rogerio Feris. Repmet: Representative-based metric learning for classification and few-shot object detection. In IEEE conference on Computer Vision and Pattern Recognition (CVPR), pages 5197-5206, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 417, + 554, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 417, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 316, + 417, + 554, + 460 + ], + "type": "text", + "content": "[21] M. Kearns, Y. Mansour, and A. Ng. An information-theoretic analysis of hard and soft assignment methods for clustering. In Conference on Uncertainty in Artificial Intelligence (UAI), 1997. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 462, + 554, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 462, + 554, + 507 + ], + "spans": [ + { + "bbox": [ + 316, + 462, + 554, + 507 + ], + "type": "text", + "content": "[22] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 509, + 554, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 554, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 554, + 553 + ], + "type": "text", + "content": "[23] Michalis Lazarou, Tania Stathaki, and Yannis Avrithis. Iterative label cleaning for transductive and semi-supervised few-shot learning. In IEEE/CVF International Conference on Computer Vision, pages 8751-8760, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 555, + 554, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 554, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 554, + 600 + ], + "type": "text", + "content": "[24] Michalis Lazarou, Yannis Avrithis, and Tania Stathaki. Adaptive manifold for imbalanced transductive few-shot learning. In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), page 2297-2306, 2024. 2, 3, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 601, + 554, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 554, + 666 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 554, + 666 + ], + "type": "text", + "content": "[25] Gao Yu Lee, Tanmoy Dam, Daniel Puiu Poenar, Vu N. Duong, and Md Meftahul Ferdaus. HELA-VFA: A hellinger distance-attention-based feature aggregation network for few-shot classification. In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2173–2183, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 669, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 554, + 713 + ], + "type": "text", + "content": "[26] Xiaoxu Li, Xiaochen Yang, Zhanyu Ma, and Jing-Hao Xue. Deep metric learning for few-shot image classification: A review of recent developments. Pattern Recognition, 138: 109381, 2023. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9673" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[27] Moshe Lichtenstein, Prasanna Sattigeri, Rogerio Feris, Raja Giryes, and Leonid Karlinsky. TAFSSL: Task-adaptive feature sub-space learning for few-shot classification. In European Conference on Computer Vision (ECCV), pages 22-539, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 296, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 296, + 163 + ], + "type": "text", + "content": "[28] Jinlu Liu, Liang Song, and Yongqiang Qin. Prototype rectification for few-shot learning. In European Conference on Computer Vision (ECCV), pages 741-756, 2020. 2, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 218 + ], + "type": "text", + "content": "[29] Yanbin Liu, Juho Lee, Minseop Park, Saehoon Kim, Eunho Yang, Sung Ju Hwang, and Yi Yang. Learning to propagate labels: Transductive propagation network for few-shot learning. In International Conference on Learning Representations, page 34677-34688, 2019. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 296, + 262 + ], + "type": "text", + "content": "[30] Xu Luo, Hao Wu, Ji Zhang, Lianli Gao, Jing Xu, and Jingkuan Song. A closer look at few-shot classification again. In International Conference on Machine Learning (ICML), pages 23103-23123, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 308 + ], + "type": "text", + "content": "[31] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. Technical report, Oxford University, 2013. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 296, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 296, + 376 + ], + "type": "text", + "content": "[32] Ségolène Martin, Malik Boudiaf, Emilie Chouzenoux, Jean-Christophe Pesquet, and Ismail Ben Ayed. Towards practical few-shot query sets: Transductive minimum description length inference. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), page 34677–34688, 2022. 2, 3, 4, 6, 7, 8, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 296, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 296, + 432 + ], + "type": "text", + "content": "[33] Ségolène Martin, Yunshi Huang, Fereshteh Shakeri, Jean-Christophe Pesquet, and Ismail Ben Ayed. Transductive zero-shot and few-shot clip. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28816-28826, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 433, + 296, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 296, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 296, + 476 + ], + "type": "text", + "content": "[34] Paul Micaelli and Amos J Storkey. Gradient-based hyperparameter optimization over long horizons. In Advances in Neural Information Processing Systems, pages 10798-10809. Curran Associates, Inc., 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 477, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 477, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 477, + 296, + 521 + ], + "type": "text", + "content": "[35] Vishal Monga, Yuelong Li, and Yonina C. Eldar. Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing. IEEE Signal Processing Magazine, 38(2):18-44, 2021. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 522, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 522, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 522, + 296, + 567 + ], + "type": "text", + "content": "[36] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 296, + 602 + ], + "type": "text", + "content": "[37] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 602, + 296, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 296, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 296, + 646 + ], + "type": "text", + "content": "[38] Hang Qi, Matthew Brown, and David G Lowe. Low-shot learning with imprinted weights. In IEEE conference on Computer Vision and Pattern Recognition (CVPR), pages 5822-5830, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 714 + ], + "type": "text", + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, and et al. Jack Clark. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 106 + ], + "type": "text", + "content": "[40] Sachin Ravi and H. Larochelle. Optimization as a model for few-shot learning. In International Conference on Learning Representations, pages 1-11, 2017. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "type": "text", + "content": "[41] Mengye Ren, Eleni Triantafillou, Sachin Ravi, Jake Snell, Kevin Swersky, Joshua B. Tenenbaum, Hugo Larochelle, and Richard S. Zemel. Meta-learning for semi-supervised few-shot classification. In International Conference on Learning Representations (ICLR), pages 1-15, 2018. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 554, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 554, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 554, + 218 + ], + "type": "text", + "content": "[42] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. Int. J. Comput. Vision, 115(3):211-252, 2015. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 219, + 554, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 554, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 554, + 275 + ], + "type": "text", + "content": "[43] Tyler Scott, Karl Ridgeway, and Michael C. Mozer. Adapted deep embeddings: A synthesis of methods for k-shot inductive transfer learning. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), pages 76-85, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 275, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 275, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 275, + 554, + 308 + ], + "type": "text", + "content": "[44] Anuj Singh and Hadi Jamali-Rad. Transductive decoupled variational inference for few-shot classification. Transactions on Machine Learning Research (TMLR), 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 309, + 554, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 554, + 354 + ], + "type": "text", + "content": "[45] Jake Snell, Kevin Swersky, and Richard S. Zemel. Prototypical networks for few-shot learning. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), pages 4080-4090, 2017. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 355, + 554, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 554, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 554, + 388 + ], + "type": "text", + "content": "[46] Yu Song and Changsheng Chen. MPPCANet: A feedforward learning strategy for few-shot image classification. Pattern Recognition, 113:107792, 2021. 4, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 388, + 554, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 554, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 554, + 443 + ], + "type": "text", + "content": "[47] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. Technical report, Center for Research in Computer Vision, University of Central Florida, 2012. arXiv:1212.0402. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 445, + 554, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 554, + 500 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 554, + 500 + ], + "type": "text", + "content": "[48] Ran Tao, Hao Chen, and Marios Savvides. Boosting transductive few-shot fine-tuning with margin-based uncertainty weighting and probability regularization. In IEEE conference on Computer Vision and Pattern Recognition (CVPR), pages 15752-15761, 2023. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 501, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 554, + 555 + ], + "type": "text", + "content": "[49] Long Tian, Jingyi Feng, Xiaoqiang Chai, Wenchao Chen, Liming Wang, Xiyang Liu, and Bo Chen. Prototypes-oriented transductive few-shot learning with conditional transport. In International Conference on Computer Vision (ICCV), pages 16317-16326, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 556, + 554, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 554, + 612 + ], + "type": "text", + "content": "[50] Olivier Veilleux, Malik Boudiaf, Pablo Piantanida, and Ismail Ben Ayed. Realistic evaluation of transductive few-shot learning. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), pages 9290-9302, 2021. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 613, + 554, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 554, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 554, + 669 + ], + "type": "text", + "content": "[51] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In Proceedings of the International Conference on Neural Information Processing Systems (NeurIPS), pages 3637-3645, 2016. 2, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 669, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 554, + 714 + ], + "type": "text", + "content": "[52] Yikai Wang, Chengming Xu, Chen Liu, Li Zhang, and Yanwei Fu. Instance credibility inference for few-shot learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12833-12842, 2020. 7" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "9674" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 462 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[53] Yimu Wang, Yihan Wu, and Hongyang Zhang. Lost domain generalization is a natural consequence of lack of training domains. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 15689-15697, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "text", + "content": "[54] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3485-3492, 2010. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "type": "text", + "content": "[55] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6757-6767, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 273 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 273 + ], + "type": "text", + "content": "[56] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. TIP-adapter: Training-free adaption of clip for few-shot classification. In European Conference on Computer Vision (ECCV), pages 493-510, 2022. 3, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 327 + ], + "type": "text", + "content": "[57] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16816-16825, 2022. 3, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 330, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 330, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 330, + 294, + 373 + ], + "type": "text", + "content": "[58] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In International Conference on Computer Vision (ICCV), pages 15659-15669, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 375, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 375, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 375, + 294, + 418 + ], + "type": "text", + "content": "[59] Hao Zhu and Piotr Koniusz. Transductive few-shot learning with prototype-based label propagation by iterative graph refinement. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 23996-24006, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 420, + 294, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 420, + 294, + 462 + ], + "spans": [ + { + "bbox": [ + 56, + 420, + 294, + 462 + ], + "type": "text", + "content": "[60] Imtiaz Masud Ziko, Jose Dolz, Eric Granger, and Ismail Ben Ayed. Laplacian regularized few-shot learning. In International Conference on Machine Learning (ICML), 2020. 7, 8" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9675" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/UNIALIGN_ Scaling Multimodal Alignment within One Unified Model/7e109366-fc0f-438c-9b91-2ba7f07e59e9_content_list.json b/2025/UNIALIGN_ Scaling Multimodal Alignment within One Unified Model/7e109366-fc0f-438c-9b91-2ba7f07e59e9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f1e9e1f1bc8e2f9a31cfbc1f19d7bce1d4a11348 --- /dev/null +++ b/2025/UNIALIGN_ Scaling Multimodal Alignment within One Unified Model/7e109366-fc0f-438c-9b91-2ba7f07e59e9_content_list.json @@ -0,0 +1,2362 @@ +[ + { + "type": "text", + "text": "UNIALIGN: Scaling Multimodal Alignment within One Unified Model", + "text_level": 1, + "bbox": [ + 145, + 130, + 852, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bo Zhou $^{1*}$ , Lilei Li $^{2*}$ , Yujia Wang $^{3}$ , Huafeng Liu $^{1}$ , Yazhou Yao $^{1\\dagger}$ , Wenguan Wang $^{2}$", + "bbox": [ + 151, + 179, + 841, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Nanjing University of Science and Technology, $^{2}$ Zhejiang University, $^{3}$ Zhejiang Sci-Tech University", + "bbox": [ + 191, + 200, + 803, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/NUST-Machine-Intelligence-Laboratory/UNIALIGN", + "bbox": [ + 210, + 219, + 781, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present UNIALIGN, a unified model to align an arbitrary number of modalities (e.g., image, text, audio, 3D point cloud, etc.) through one encoder and a single training phase. Existing solutions typically employ distinct encoders for each modality, resulting in increased parameters as the number of modalities grows. In contrast, UNIALIGN proposes a modality-aware adaptation of the powerful mixture-of-experts (MoE) schema and further integrates it with Low-Rank Adaptation (LoRA), efficiently scaling the encoder to accommodate inputs in diverse modalities while maintaining a fixed computational overhead. Moreover, prior work often requires separate training for each extended modality. This leads to task-specific models and further hinders the communication between modalities. To address this, we propose a soft modality binding strategy that aligns all modalities using unpaired data samples across datasets. Two additional training objectives are introduced to distill knowledge from well-aligned anchor modalities and prior multimodal models, elevating UNIALIGN into a high performance multimodal foundation model. Experiments on 11 benchmarks across 6 different modalities demonstrate that UNIALIGN could achieve comparable performance to SOTA approaches, while using merely 7.8M trainable parameters and maintaining an identical model with the same weight across all tasks.", + "bbox": [ + 88, + 300, + 483, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 705, + 220, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans accomplish tasks by processing and combining signals from different modalities [37, 53]. This gives rise to multimodal learning, which aims to build AI models that can extract and relate information across multimodal inputs [97]. Recent efforts have been made to learn a unified representation space via the alignment between modalities [18, 22, 39]. Though enabling extension to novel modalities, existing solutions typically rely on a one-versus-one paradigm where merely one extended modality (e.g., point", + "bbox": [ + 89, + 729, + 483, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b45890f026d892961b336e635b5a3c6d4df200ee34dac41aa598ba9183b70cc6.jpg", + "image_caption": [ + "Figure 1. Comparison of architectures between standard multi-modal alignment methods [18, 39, 109] and UNIALIGN: (a) Existing work adopts specialized encoders tailored to each modality; (b) UNIALIGN unifies the alignment across multiple modalities in a single model, with one training phase across all datasets." + ], + "image_footnote": [], + "bbox": [ + 517, + 268, + 905, + 511 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cloud, audio) could be aligned with the anchor modalities (e.g., text, image) at each time (Fig. 1(a)). This presents several drawbacks: i) one dedicated encoder is required for each modality, leading to exponential increases in training cost and parameter count as the number of aligned modalities grows, severely limiting scalability; ii) task-specific training for each dataset, results in multiple independent models for different tasks; and iii) the separate alignment of each extended modality hinders the model to leverage the full breadth of data across all datasets to shape the shared embedding space from a global view. This further impedes cross-modal knowledge transfer where insights gained from one modality could benefit understanding in others.", + "bbox": [ + 509, + 595, + 906, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "These observations prompt us to rethink: Is this one-versus-one alignment paradigm essential? and Can we achieve joint alignment of all modalities within one single model? In response, we propose UNIALIGN (Fig. 1(b)), which embraces a all-in-one design to realize multimodal alignment by constructing a foundation model that includes: 1 one base encoder to UNify the encoding of all anchor and", + "bbox": [ + 511, + 795, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*The first two authors contribute equally to this work.", + "bbox": [ + 107, + 875, + 392, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author: Yazhou Yao.", + "bbox": [ + 109, + 887, + 305, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29644", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "extended modalities, and $②$ one training phase to ALIGN all modalities with all samples across datasets collectively.", + "bbox": [ + 89, + 90, + 485, + 122 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Concretely, to achieve ①, recent work has consistently demonstrated that Transformers pre-trained on one modality possess exceptional generalization ability to novel modalities [13, 67]. Moreover, the mixture-of-experts (MoE) [72] schema presents a highly efficient strategy to scale model capacity under a fixed computational cost [108]. These insights motivate us to build a foundational model accommodating multiple modalities upon a single pre-trained Transformer (e.g., CLIP [67]). Here, MoE can dynamically scale and allocate model capacities by selectively activating experts to encode inputs across diverse modalities. We explore two strategies targeting at expert grouping and routing mechanism to enhance the awareness of different modalities within MoE. To preserve the knowledge learned from anchor modalities, Low-Rank Adaptation (LoRA) [27] is integrated with MoE, enabling a significant reduction in training overhead. To fulfill ②, we propose a soft modality binding strategy, leveraging massive unpaired data samples across modalities and datasets. Specifically, given unpaired data from diverse datasets with different modalities, the class labels of samples are encoded into text embeddings to create a unified representation. The semantic similarity between text embeddings can inform the alignment between modalities (e.g., audio and video), by encouraging the similarity between modality embeddings and textual label embeddings (e.g., audio $\\leftrightarrow$ textvideo and video $\\leftrightarrow$ textaudio) to be identical as the semantic similarity (i.e., $text_{\\mathrm{video}} \\leftrightarrow \\text{text}_{\\mathrm{audio}}$ ). Moreover, rooted in the principles of knowledge distillation [2, 26, 110], the alignment between extended and anchor modalities is guided by the knowledge persisting between anchor modalities, which are well-aligned in semantics compared to extended modalities. To further boost the performance, we transfer knowledge from multiple well-trained foundation models which only contain partial modalities. Given the complexity of multimodal, multi-dataset joint training, a specialized batching strategy combined with modality-aware gradient accumulation is proposed, to mitigate the potential dispersion in optimization across different modalities.", + "bbox": [ + 91, + 125, + 485, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "UNIALIGN enjoys several advantages: First, it supports flexible alignment between arbitrary number of anchor and extended modalities, while ensuring negligible increase in parameter size. Second, the alignment for all modalities is accomplished in one training session, significantly improving training efficiency. Third, unpaired data from different modalities can be used collectively for optimization. Uncommon modalities with scarce data could take full advantages of the rich resource of common modalities. Fourth, benefited from soft modality binding, extended modalities are bridged for mutual boosting. Fifth, a compact and high-performance multimodal foundational", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "model generally applicable to all tasks is delivered at minimal cost. This contrasts with existing approaches, which typically provide multiple loosely integrated modality encoders tailored to specific tasks, thereby enabling broader transferability to other domains and downstream applications [3, 8, 16, 61, 62, 95, 96, 101, 102].", + "bbox": [ + 511, + 90, + 906, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By embracing the all-in-one philosophy, UNIALIGN demonstrates remarkable efficiency in both training cost and parameter size. After a single pre-training step without further tuning, UNIALIGN achieves top-leading performance on 11 benchmarks with one identical model and 7.8M trainable parameters across 6 modalities (previous SOTA [18, 39] needing over 80M trainable parameters for each extended modality). This evidences the great potential of our unified modality encoding and aligning strategies. We hope this work could stimulate the community to rethink multimodal alignment from a universal and compact perspective.", + "bbox": [ + 511, + 181, + 908, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 361, + 655, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal Learning. Multimodal learning has attracted significant attention given its ability to address a wide range of applications by integrating information from diverse modalities such as text, images, audio, and video[49, 51, 67, 103]. Current research in this field generally falls into three categories: First, methods involving joint training of multiple modalities aim to improve performance across different downstream tasks, by leveraging shared representations[17, 78, 107]. Though incorporating multiple modalities as inputs, these work lacks explicit alignment between modalities. Second, inspired by the successes of vision-language alignment [67], some approaches extend this alignment to additional modalities[18, 38, 40, 56, 59, 91, 98, 109]. These methods typically require an independent encoder for each modalities. Though [14, 18, 39, 90] maintain an consistent architecture across tasks, they conduct separate training on individual datasets, resulting in multiple task-specific models. Additionally, these solutions often rely on massive paired data to achieve robust zero-shot performance across various tasks, ignoring the usage of rich unpaired data. Third, the integration of multimodal inputs with large language models (LLMs) has opened new avenues for multimodal learning[23, 50, 52, 88, 100]. In these approaches, all modalities are projected into the same representation space as LLMs, without alignment among extended modalities and supporting merely one anchor modality (i.e., text). This leads to suboptimal performance on certain vision tasks and limits the applications primarily to language-related fields.", + "bbox": [ + 511, + 387, + 908, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose UNIALIGN, supporting alignment between an arbitrary number of anchor and extended modalities within one model using one unified modality encoder. After a single training phase, UNIALIGN can handle a dozen of downstream tasks without any fine-tuning. UNIALIGN also achieves high training efficiency, utilizing", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29645", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "merely 7.8M training parameters for all tasks.", + "bbox": [ + 89, + 90, + 393, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-Task Learning. To maximize the utilization of resource and improve generalization across tasks, multi-task learning seeks to develop models that can handle multiple tasks simultaneously [4]. Traditional methods often require separate subnetworks for each task, which hinders scalability [83]. In recent years, the mixture-of-experts (MoE) architecture [15, 72] has gained prominence in multi-task learning [31]. MoE leverages multiple subnetworks, or \"experts\", that are selectively activated to optimize task performance while reducing computational costs [35]. This approach proves beneficial across various domains, such as language processing [31, 35] and vision tasks [41, 43, 69]. With the success of LoRA [27], there has been a surge to integrate MoE with LoRA, which shows promise in enhancing both computational efficiency and task adaptability in large-scale models [11, 21, 28, 29]. However, these methods are mainly confined to text and image domains [7, 11], and lack alignment between modalities [46]. In contrast, UNIALIGN extends MoE to a multimodal setup, devising modality-aware expert selection and routing mechanism.", + "bbox": [ + 89, + 106, + 483, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Knowledge Distillation. Knowledge distillation [26, 34, 55] is an essential technique for model compression, enabling the transfer of knowledge from large, computationally intensive models to smaller, more efficient ones. This is particularly valuable for multimodal foundation models, where the computational overload is heavy [32, 45, 79, 99]. Initially, knowledge distillation focused on extracting knowledge from individual large models, typically using the logits or intermediate feature representations to guide the training of student models [1, 25, 36, 63, 93]. Recent research has expanded this concept to integrate capabilities from multiple models, creating hybrid models that not only improve task-specific performance but also broaden the range of potential applications [68, 71, 84, 104, 105, 112].", + "bbox": [ + 89, + 407, + 483, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Compared to existing work, UNIALIGN distills knowledge from both well-aligned anchor modalities, and multiple well-trained foundation models specialized in partial modalities. This allows our method to handle a broader range of tasks that require understanding and processing information with different composition of diverse modalities.", + "bbox": [ + 89, + 619, + 483, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 723, + 181, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first introduce the unified encoding of all modalities within one single model (§3.1), then describe the universal modality alignment strategy across all datasets (§3.2), and finally present the implementation details of UNIALIGN (§3.3).", + "bbox": [ + 89, + 748, + 483, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Unified Modality Encoding", + "text_level": 1, + "bbox": [ + 89, + 833, + 334, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preliminary: Mixture-of-Experts (MoE). Let $X \\in \\mathbb{R}^{L \\times D}$ denote the tokenized input, where $L$ is the sequence length and $D$ is the token embedding size. In vanilla Transformer,", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/465c5463cfdf1aa2bb4fd9c957df010e25ee5aad93f5e5fbfa454961685ee04d.jpg", + "image_caption": [ + "Figure 2. Architecture of the unified modality encoder." + ], + "image_footnote": [], + "bbox": [ + 539, + 84, + 883, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "each encoder block comprises a multi-head self-attention (MSA) layer and a feed-forward network (FFN), as follows:", + "bbox": [ + 511, + 415, + 903, + 444 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {X} ^ {\\prime} = \\boldsymbol {X} + \\mathcal {F} ^ {\\mathrm {M S A}} (\\ln (\\boldsymbol {X})), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 449, + 903, + 465 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {X} ^ {\\prime \\prime} = \\boldsymbol {X} ^ {\\prime} + \\mathcal {F} ^ {\\mathrm {F F N}} (\\ln (\\boldsymbol {X} ^ {\\prime})), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 469, + 903, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $X''$ is the outputs and LN represents layer normalization. In the mixture-of-experts (MoE) schema, an MoE function (i.e., $\\mathcal{F}^{\\mathrm{MoE}}$ ) is adopted to replace $\\mathcal{F}^{\\mathrm{FFN}}$ :", + "bbox": [ + 511, + 497, + 903, + 542 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {X} ^ {\\prime \\prime} = \\boldsymbol {X} ^ {\\prime} + \\mathcal {F} ^ {\\mathrm {M o E}} (\\ln \\left(\\boldsymbol {X} ^ {\\prime}\\right)), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 545, + 903, + 561 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\mathrm {M o E}} \\left(\\boldsymbol {X} ^ {\\prime}\\right)) = \\sum_ {i = 1} ^ {n} \\mathcal {G} \\left(\\boldsymbol {X} ^ {\\prime}\\right) _ {i} \\mathcal {F} _ {i} ^ {\\text {E x p e r t}} \\left(\\boldsymbol {X} ^ {\\prime}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 566, + 903, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{F}^{\\mathrm{MoE}}$ jointly aggregates outputs from multiple expert networks (i.e., $\\mathcal{F}^{\\mathrm{Expert}}$ ) according to the router $\\mathcal{G}(X')_i = \\text{softmax}(\\mathcal{F}^{\\mathrm{Router}}(X'))_i$ to assign weights for different experts and render the predictions.", + "bbox": [ + 511, + 594, + 903, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multimodal Adaptation with MoE. We aim to unify the encoding of multimodal inputs within one foundation model (e.g., the image encoder of CLIP[67]). As illustrated in Fig.2, inputs of various modalities are first converted into sequential tokens. We follow [39, 42, 107] to preprocess multimodal data and implement modality-specific tokenizers, with details summarized in Appendix. To adapt the encoding of samples from various modalities, UNIALIGN explores two modality-aware MoE strategies:", + "bbox": [ + 511, + 655, + 905, + 790 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Modality-Aware Expert Grouping divides all $\\mathcal{F}^{\\mathrm{Expert}}$ in one MoE layer into multiple sets, where each set $\\mathcal{B}$ contains $n$ experts to process inputs from a specific modality $\\mathcal{M}$ :", + "bbox": [ + 511, + 790, + 903, + 835 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {B} _ {\\mathcal {M}} = \\left\\{\\mathcal {F} _ {\\mathcal {M}, 1} ^ {\\text {E x p e r t}}, \\dots , \\mathcal {F} _ {\\mathcal {M}, n} ^ {\\text {E x p e r t}} \\right\\}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 840, + 903, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Under this setup, following [46], we construct a hierarchical routing mechanism. In the first routing round, tokens are", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "29646", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "directed to a specific expert set based on the modality of inputs. Next, within a expert set, a learnable router $\\mathcal{F}_{\\mathcal{M}}^{\\text{Router}}$ for modality $\\mathcal{M}$ assigns weights to each individual expert conditioned on the input tokens. This separation lets each expert set to specialize in one modality, and allows for easy expansion to novel modalities by loading new expert sets.", + "bbox": [ + 89, + 90, + 482, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Modality-Aware Global Routing uses multiple modality-specific routers, to direct tokens across all experts in a MoE layer without grouping. Given the routing function $\\mathcal{F}_{\\mathcal{M}}^{\\text{Router}}$ for modality $M$ , the global weight $\\mathcal{G}_{\\mathcal{M}}$ is computed as:", + "bbox": [ + 89, + 181, + 482, + 243 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {\\mathcal {M}} \\left(\\boldsymbol {X} ^ {\\prime}\\right) _ {i} = \\operatorname {s o f t m a x} \\left(\\mathcal {F} _ {\\mathcal {M}} ^ {\\text {R o u t e r}} \\left(\\boldsymbol {X} ^ {\\prime}\\right)\\right) _ {i}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 251, + 482, + 270 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By allowing the experts to be accessible to each router, this strategy enables cross-modality knowledge sharing among experts. However, since the experts are not explicitly specialized for one modality, the generalization potentially at the expense of fine-grained modality specialization.", + "bbox": [ + 89, + 281, + 482, + 356 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Modality-Aware LoRA Expert. Low-Rank Adaptation (LoRA) [28] has demonstrated both effectiveness and efficiency in fine-tuning large models [9, 81]. Therefore, to maintain the knowledge in pre-trained foundation models, the original feed-forward network (i.e., $\\mathcal{F}^{\\mathrm{FFN}}$ in Eq. 2) are preserved in a fixed state instead of being replaced during the forward of MoE (i.e., Eq. 3), resulting in:", + "bbox": [ + 89, + 357, + 483, + 462 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {X} ^ {\\prime \\prime} = \\boldsymbol {X} ^ {\\prime} + \\mathcal {F} ^ {\\mathrm {F F N}} (\\ln (\\boldsymbol {X} ^ {\\prime})) + \\mathcal {F} ^ {\\mathrm {M o E}} (\\ln (\\boldsymbol {X} ^ {\\prime})). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 470, + 482, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here the experts in $\\mathcal{F}^{\\mathrm{MoE}}$ leverage $\\mathcal{F}^{\\mathrm{FFN}}$ as the base weight for LoRA adaption, with each of them employs a specific instance of LoRA (i.e., $\\mathcal{F}^{\\mathrm{LoRA}}$ ) parameterized as follows:", + "bbox": [ + 89, + 500, + 482, + 546 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\operatorname {M o E}} \\left(\\boldsymbol {X} ^ {\\prime}\\right)) = \\sum_ {i = 1} ^ {n} \\mathcal {G} \\left(\\boldsymbol {X} ^ {\\prime}\\right) _ {i} \\mathcal {F} _ {i} ^ {\\text {L o R A}} \\left(\\boldsymbol {X} ^ {\\prime}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 553, + 482, + 571 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {i} ^ {\\text {L o R A}} \\left(\\boldsymbol {X} ^ {\\prime}\\right) = \\frac {\\alpha}{r} \\cdot \\boldsymbol {B} _ {i} \\boldsymbol {A} _ {i} \\boldsymbol {X} ^ {\\prime}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 573, + 480, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where low-rank matrices $\\mathbf{A} \\in \\mathbb{R}^{d_{in} \\times r}$ and $\\mathbf{B} \\in \\mathbb{R}^{r \\times d_{out}}$ are scaled by a factor $\\alpha$ normalized by the rank $r \\ll \\min(d_{in}, d_{out})$ . In this way, we achieve seamless integration of LoRA with MoE, enabling efficient tuning and knowledge preservation of the foundation models, while supporting adaptation to inputs across various modalities.", + "bbox": [ + 89, + 609, + 482, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Universal Modality Alignment", + "text_level": 1, + "bbox": [ + 89, + 712, + 361, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Soft Modality Binding across Datasets. Prior work [18] reveals that modalities expressed in a joint embedding space could be automatically aligned through an intermediate modality. For example, the alignment of and , could lead to the alignment between with Image as the bridge. However, this approach requires massive naturally paired samples between anchor and extended modalities, which are often challenging and time-consuming to obtain. We thus propose a soft modality binding strategy, which utilizes unpaired samples across datasets with different modalities", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/408eda2d6b769a4c6ec7401c6deb1b2233c62cb3e50df4ccf3e3569909b193d1.jpg", + "image_caption": [ + "Figure 3. Our proposed soft modality binding strategy. Unpaired samples from diverse datasets with different modalities are bridged with the semantic similarity between class labels as soft guidance." + ], + "image_footnote": [], + "bbox": [ + 519, + 84, + 903, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(Fig. 3). Specifically, let $\\mathbf{y}^{\\mathcal{M}_1} = \\{\\pmb {y}_1^{\\mathcal{M}_1},\\dots ,\\pmb {y}_{N_1}^{\\mathcal{M}_1}\\}$ and $\\pmb {\\mathcal{V}}^{\\mathcal{M}_2} = \\{\\pmb {y}_1^{\\mathcal{M}_2},\\dots ,\\pmb {y}_{N_2}^{\\mathcal{M}_2}\\}$ represent text embeddings of class labels from two arbitrary datasets in modalities $\\mathcal{M}_1$ and $\\mathcal{M}_2$ , respectively. The semantic similarity of textual labels across datasets can be computed as:", + "bbox": [ + 511, + 407, + 906, + 489 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} \\left(\\boldsymbol {y} _ {i} ^ {\\mathcal {M} _ {1}}, \\boldsymbol {y} _ {j} ^ {\\mathcal {M} _ {2}}\\right) = \\frac {\\exp \\left(\\boldsymbol {y} _ {i} ^ {\\mathcal {M} _ {1}} \\cdot \\boldsymbol {y} _ {j} ^ {\\mathcal {M} _ {2}}\\right)}{\\sum_ {n = 1} ^ {N _ {2}} \\exp \\left(\\boldsymbol {y} _ {i} ^ {\\mathcal {M} _ {1}}, \\boldsymbol {y} _ {n} ^ {\\mathcal {M} _ {2}}\\right)}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 494, + 905, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarly, given samples $(\\pmb{x}^{\\mathcal{M}_1},\\pmb{y}_i^{\\mathcal{M}_1})$ from the first dataset and $(\\pmb{x}^{\\mathcal{M}_2},\\pmb{y}_j^{\\mathcal{M}_2})$ from the second dataset, the semantic similarity between $\\pmb{x}^{\\mathcal{M}_1}$ and text embedding of labels from the second dataset (i.e., $S(\\pmb{x}^{\\mathcal{M}_1},\\pmb{y}_j^{\\mathcal{M}_2}))$ , as well as $\\pmb{x}^{\\mathcal{M}_2}$ and text embeddings of labels from the first dataset (i.e., $S(\\pmb{x}^{\\mathcal{M}_2},\\pmb{y}_i^{\\mathcal{M}_1}))$ could be computed. Taking the similarity between class labels of two datasets (i.e., $S(\\pmb{y}_i^{\\mathcal{M}_1},\\pmb{y}_j^{\\mathcal{M}_2}))$ as a soft prior for alignment, we bind modalities across datasets by optimizing the following objective:", + "bbox": [ + 511, + 542, + 906, + 686 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {s o f t \\_ b i n d}} = \\mathrm {K L} \\left(\\mathcal {S} \\left(\\boldsymbol {x} ^ {\\mathcal {M} _ {1}}, \\boldsymbol {y} _ {j} ^ {\\mathcal {M} _ {2}}\\right) \\mid \\mathcal {S} \\left(\\boldsymbol {y} _ {i} ^ {\\mathcal {M} _ {1}}, \\boldsymbol {y} _ {j} ^ {\\mathcal {M} _ {2}}\\right)\\right) \\\\ + \\operatorname {K L} \\left(\\mathcal {S} \\left(\\boldsymbol {x} ^ {\\mathcal {M} _ {2}}, \\boldsymbol {y} _ {i} ^ {\\mathcal {M} _ {1}}\\right) \\mid \\mathcal {S} \\left(\\boldsymbol {y} _ {j} ^ {\\mathcal {M} _ {2}}, \\boldsymbol {y} _ {i} ^ {\\mathcal {M} _ {1}}\\right)\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 690, + 861, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where KL refers to the Kullback-Leibler divergence. Our soft modality binding strategy is different from [18], which requires instance-level textual annotations for each sample. Instead, our approach utilizes merely class labels which are instance-agnostic and challenging for alignment.", + "bbox": [ + 511, + 734, + 905, + 810 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Knowledge Distillation from Anchor Modalities. Given multiple well-aligned anchor modalities (e.g., $\\mathcal{M}_1^a$ and $\\mathcal{M}_2^a$ ) and one novel modality to be extended (e.g., $\\mathcal{M}_k^e$ ), the alignment performance between anchor and extended modalities (e.g., $\\mathcal{M}_k^e$ and $\\mathcal{M}_1^a$ ) is typically less effective than that between anchor modalities (e.g., $\\mathcal{M}_1^a$ and $\\mathcal{M}_2^a$ ). Therefore,", + "bbox": [ + 511, + 810, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29647", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e1b8aceedb68cc74451b32ccc1739996b9e076f40208947c09536b91a150e785.jpg", + "image_caption": [ + "Figure 4. Top: Knowledge distillation from anchor modalities to instruct the alignment of novel modalities. Bottom: knowledge distillation from multiple partial-modal foundation models." + ], + "image_footnote": [], + "bbox": [ + 94, + 83, + 478, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "knowledge within the alignment between anchor modalities (Fig. 4 (Top)) could be transferred to improve the alignment learning of extended modalities via knowledge distillation.", + "bbox": [ + 89, + 373, + 482, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For instance, given a 3D point cloud sample $x_{i}$ , we could obtain its tokenized input $\\pmb{x}_{i}^{\\mathcal{P}}$ in the extended modalities, along with the text embedding of class label $\\pmb{x}_{i}^{\\mathcal{T}}$ and the visual embedding of one rendered image from a specific viewpoint $\\pmb{x}_{i}^{\\mathcal{T}}$ , both in the anchor modalities. By leveraging the similarity between anchor modalities (i.e., $S(\\pmb{x}_{i}^{\\mathcal{T}},\\pmb{x}_{i}^{\\mathcal{T}})$ ) as the guidance, we could optimize the alignment between the extended modality and anchor modalities through:", + "bbox": [ + 89, + 417, + 483, + 540 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {a n c h . d i s t i l}} = \\mathrm {K L} \\left(S \\left(\\boldsymbol {x} _ {i} ^ {\\mathcal {P}}, \\boldsymbol {x} _ {i} ^ {\\mathcal {T}}\\right) \\mid S \\left(\\boldsymbol {x} _ {i} ^ {\\mathcal {I}}, \\boldsymbol {x} _ {i} ^ {\\mathcal {T}}\\right)\\right) \\tag {12} \\\\ + \\operatorname {K L} \\left(\\mathcal {S} \\left(\\boldsymbol {x} _ {i} ^ {\\mathcal {P}}, \\boldsymbol {x} _ {i} ^ {\\mathcal {I}}\\right) \\mid \\mathcal {S} \\left(\\boldsymbol {x} _ {i} ^ {\\mathcal {T}}, \\boldsymbol {x} _ {i} ^ {\\mathcal {I}}\\right)\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 544, + 482, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This knowledge distillation strategy is generally applicable to most modalities, where 3D point cloud, video, and depth data can be converted into images with class labels.", + "bbox": [ + 89, + 585, + 482, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Knowledge Distillation from Partial-Modal Foundation Models. To further boost the alignment across modalities, we employ foundation models $\\mathcal{E}_f$ skilled in partial modalities (e.g., audio $\\leftrightarrow$ text [22], point cloud $\\leftrightarrow$ image [98]) as teacher models to guide the learning of our unified modality encoder $\\mathcal{E}_u$ (i.e., Fig.4 (Bottom)) from two perspectives:", + "bbox": [ + 89, + 631, + 482, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Feature Alignment aims to ensure features from our unified modality encoder to be identical as those from partial-modal foundation models. Specifically, denoting $\\hat{F}_i^{\\mathcal{M}_k}$ and $F_{i}^{\\mathcal{M}_{k}}$ as features in modality $\\mathcal{M}_k$ output by $\\mathcal{E}_f$ and $\\mathcal{E}_u$ , we use the mean squared error (MSE) to align them as:", + "bbox": [ + 89, + 720, + 482, + 797 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {f e a t - a l i g n}} = \\left\\| \\hat {\\boldsymbol {F}} ^ {\\mathcal {M} _ {i}} - \\boldsymbol {F} ^ {\\mathcal {M} _ {i}} \\right\\| _ {2} ^ {2}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 801, + 482, + 820 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Cross-Modality Knowledge Transfer aims to benefit the learning of $\\mathcal{E}_u$ via the well-aligned representation space across modalities shaped by large-scale paired data. Specifically, given a paired sample $(x_i^{\\mathcal{M}^a}, x_i^{\\mathcal{M}^e})$ in anchor modality $\\mathcal{M}^a$ and extended modality $\\mathcal{M}^e$ respectively, the align", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ment results of our unified modality encoder are encouraged to approximate those of partial-modal foundation models:", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {m o d a l - a l i g n}} = \\mathrm {K L} \\left(\\mathcal {S} \\left(\\boldsymbol {x} _ {i} ^ {\\mathcal {M} ^ {e}}, \\boldsymbol {x} _ {i} ^ {\\mathcal {M} ^ {a}}\\right) \\mid \\mathcal {S} \\left(\\hat {\\boldsymbol {x}} _ {i} ^ {\\mathcal {M} ^ {e}}, \\hat {\\boldsymbol {x}} _ {i} ^ {\\mathcal {M} ^ {a}}\\right)\\right) \\tag {14} \\\\ + \\mathrm {K L} \\left(S \\left(\\boldsymbol {x} _ {i} ^ {\\mathcal {M} ^ {a}}, \\boldsymbol {x} _ {i} ^ {\\mathcal {M} ^ {e}}\\right) \\mid S \\left(\\hat {\\boldsymbol {x}} _ {i} ^ {\\mathcal {M} ^ {a}}, \\hat {\\boldsymbol {x}} _ {i} ^ {\\mathcal {M} ^ {e}}\\right)\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 125, + 903, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\pmb{x}}_i$ and $\\pmb{x}_i$ are outputs of $\\mathcal{E}_f$ and $\\mathcal{E}_u$ , respectively.", + "bbox": [ + 511, + 166, + 879, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In summary, the loss for knowledge distillation from partial-modal foundation models can be formulated as:", + "bbox": [ + 511, + 181, + 905, + 210 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m o d e l - d i s t i l}} = \\lambda \\mathcal {L} _ {\\text {f e a t - a l i n g}} + \\mathcal {L} _ {\\text {m o d a l - a l i n g}}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 215, + 903, + 232 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following [99], we set $\\lambda = 2000$ to scale the magnitude.", + "bbox": [ + 511, + 234, + 885, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Implementation Details", + "text_level": 1, + "bbox": [ + 513, + 257, + 730, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Network Configuration. We adopt CLIP [67] pre-trained vision encoder as the base network for unified modality encoding. The sequence length $L$ and channel dimension $D$ for tokenized inputs $X$ are set to 196 and 768, respectively. The LoRA configuration is defined with a scaling factor $\\alpha$ of 32 and a rank $r$ of 4. For knowledge distillation from partial-modal foundation models, we use ULIP [98] as the teacher model for 3D point cloud inputs, VIT-LENS[38] for depth inputs, ONE-PEACE[86] for audio and video inputs.", + "bbox": [ + 511, + 279, + 906, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Batching Strategy. The batching strategies proposed by [17, 42] sample inputs from each dataset (modality) separately, which may lead to catastrophic forgetting[42]. To address this, we develop a new batching strategy, which mixes data from all modalities within each mini-batch. Furthermore, a modality-aware gradient accumulation is proposed to perform forward pass, loss computation, and gradient calculation individually for each modality. Backpropagation is conducted only after processing all modalities.", + "bbox": [ + 511, + 416, + 905, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Objective. Following [68], we assign an equal weight to all training objectives:", + "bbox": [ + 511, + 551, + 905, + 582 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {N C E}} + \\alpha_ {1} \\mathcal {L} _ {\\text {s o f t , b i n d}} + \\alpha_ {2} \\mathcal {L} _ {\\text {a n c h , d i s t i l}} + \\alpha_ {3} \\mathcal {L} _ {\\text {m o d e l , d i s t i l}}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 584, + 903, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha_{\\{1,2,3\\}} = 1$ and $\\mathcal{L}_{\\mathrm{NCE}}$ is the modality aligning loss for paired data, as utilized in CLIP [67].", + "bbox": [ + 511, + 604, + 905, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 643, + 645, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 513, + 667, + 702, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pre-training Datasets. We perform a joint pre-training across multiple datasets with different modalities, including images from ImageNet-1K [8], 3D point clouds from ShapeNet[5], depth data from SUN RGB-D[74], audio and video from AudioSet [16] and VGGSound [6]. To address the differences in dataset size across modalities, we apply repeated augmentation[17] to small datasets (such as SUN-D) and category-balanced downsampling for large datasets (such as AudioSet) to ensure balanced performance during joint training. Once finished pre-training, we directly evaluate our model on downstream tasks without any fine-tuning.", + "bbox": [ + 511, + 688, + 906, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. For image and video data, we use a resolution of $224 \\times 224$ and apply standard augmentations. For single-view depth data, we employ the same augmentation as in", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29648", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f678c4bb8357c69dfe413ffa6d12ba3c12d21adec8dfa247282fba6014721de8.jpg", + "table_caption": [], + "table_footnote": [ + "An identical model with the same weight across all tasks.", + "Table 1. Results for image classification on ImageNet-1K (§4.2)." + ], + "table_body": "
ModelBackboneTrainable ParamsTop-1 AccuracyTop-5 Accuracy
Single Task
Swin Transformer [47]Swin-B87.7M83.596.5
ConvNeXt [48]ConvNeXt-B88.7M83.896.7
DeiT III [80]ViT-B86.9M83.896.5
Conformer [64]Conformer-B83.3M83.896.7
Hiera [70]Hiera-B+70.0M85.297.3
Multiple Tasks
UniRepLKNet [10]UniRepLKNet-S55.6M83.9-
OMNIVORE [17]Swin-B88.2M85.397.5
MetaFormer [107]ViT-B86.6M85.497.4
UNIALIGN†ViT-B7.8M85.397.2
", + "bbox": [ + 101, + 88, + 480, + 256 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e6233942872423b839c3bbb7d797939681fc53404b31a060907d254a29fa7bd7.jpg", + "table_caption": [], + "table_footnote": [ + "An identical model with the same weight across all tasks.", + "Table 2. Results for zero-shot 3d shape classification on ModelNet40 test [95] and ScanObjectNN test [82] (\\$4.3)." + ], + "table_body": "
ModelBackboneTrainable ParamsModelNet40ScanObjectNN
Top-1Top-5Top-1Top-5
Single Task
PointCLIP [106]ResNet-50-19.334.810.530.6
PointCLIPv2 [111]ViT-B-63.685.042.274.5
CLIP2Point [30]ViT-B88.3M49.581.225.559.4
RECON [66]ViT-B43.6M61.278.142.375.6
ULIP [98]ViT-B43.6M60.484.051.580.2
Multiple Tasks
VIT-LENS [38]ViT-B34.1M65.492.7--
UNIALIGN†ViT-B7.8M55.984.342.576.3
", + "bbox": [ + 101, + 297, + 480, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[38]. The augmentation of audio data also follows [38] except Mix-up [94]. We train the model with a batch size of 2048 (after gradient accumulation) for 100 epochs. An AdamW optimizer with a initial learning rate $2\\mathrm{e}^{-4}$ and a weight decay of 0.05 is employed. For knowledge distillation from partial-modal foundation models, we cache offline features to avoid repeatedly forwarding. The base modality encoder remains frozen during training, with only LoRA enhanced MoE modules and the projection head for knowledge distillation remaining trainable. Flash-attention is applied to reduce GPU memory cost.", + "bbox": [ + 89, + 491, + 483, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Reproducibility. UNIALIGN is implemented in PyTorch and trained on four NVIDIA Tesla A100 GPUs.", + "bbox": [ + 89, + 657, + 482, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results for Image Classification", + "text_level": 1, + "bbox": [ + 89, + 696, + 370, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We evaluate UNIALIGN for image classification on ImageNet-1K val [8] which contains 50,000 samples across 1,000 classes, spanning objects, animals, and more.", + "bbox": [ + 89, + 719, + 482, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metric. Following previous works [47, 48, 87], we report the count of trainable parameters, top-1 and top-5 accuracy.", + "bbox": [ + 89, + 763, + 482, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance. As shown in Table 1, UNIALIGN demonstrates impressive performance on image classification, achieving a comparable performance to modality-specific approaches (i.e., $85.3\\%$ vs. $85.2\\%$ of Hiera[70]). It is noteworthy that our method utilizes merely $9\\%$ trainable parameters and a unified modality encoder architecture which preserves the capability to process other modalities.", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4cc03c65044fc2d34ed84304b999af1b09a852eeb95b79b5def9c6c282d79ec0.jpg", + "table_caption": [], + "table_footnote": [ + "An identical model with the same weight across all tasks.", + "Table 3. Results for depth-only scene classification on NYU-D test [73] and SUN-D test [74] (\\$4.4).*: The depth maps are tokenized by the image tokenizer additionally." + ], + "table_body": "
ModelBackboneTrainable ParamsAnchor ModalityNYU-DSUN-D
Single Task
G-L-SOOR [76]D-CNN [75]--62.344.1
RecgNet [12]ResNet18 [24]11.2M-57.747.9
Depth Swin [18]Swin-B88.2M-76.469.1
Multiple Tasks
Text Paired [18]ViT-H88.1MT41.925.4
ImageBind [18]ViT-H88.3MI54.035.1
ViT-LENS [38]ViT-L50.9MI+T68.552.2
UNIALIGN†ViT-B7.8MT46.334.3
UNIALIGN†ViT-B7.8MI+T59.445.8
UNIALIGN†ViT-B7.8MI+T71.4*58.2*
", + "bbox": [ + 524, + 89, + 903, + 239 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Results for Zero-Shot 3D Shape Classification", + "text_level": 1, + "bbox": [ + 511, + 306, + 898, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We evaluate zero-shot 3D shape classification on two datasets. ModelNet40 test [95] contains 2,468 samples from 40 categories, while ScanObjectNN test [82] consists of 581 samples distributed across 15 categories.", + "bbox": [ + 511, + 330, + 903, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metric. The top-1 and top-5 accuracy are reported.", + "bbox": [ + 511, + 391, + 852, + 406 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance. As shown in Table 2, UNIALIGN achieves a top-1 accuracy of $55.9\\%$ and $42.5\\%$ on ModelNet40 val, and ScanObjectNN test, respectively. Though SOTA approaches such as ULIP [98] and ViT-LENS [38] yield higher scores, they require a modality-specific point-cloud encoder and an additional image encoder for alignment.", + "bbox": [ + 511, + 407, + 905, + 497 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Results for Depth-only Scene Classification", + "text_level": 1, + "bbox": [ + 511, + 508, + 877, + 525 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. Experiments for depth-only scene classification are conducted on SUN RGB-D test [74] which contains 4,660 samples across 19 classes, and NYU-D test [73] which contains 653 samples across 10 semantic classes.", + "bbox": [ + 511, + 532, + 905, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metric. The top-1 accuracy is adopted for evaluation.", + "bbox": [ + 511, + 593, + 867, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance. A detailed comparison of UNIALIGN against several top-leading approaches for depth-only scene classification is provided in Table 3. Building on a ViT-B backbone and aligning extended modalities with both image and text as anchor modalities, UNIALIGN achieves a top-1 accuracy of $71.4\\%$ on NYU-D test and $58.2\\%$ on SUN RGB-D test, surpassing all competitors. Notably, aligning merely to the text modality leads to a marked performance decline (i.e., $71.4\\% \\rightarrow 46.3\\%$ on NYU-D). This suggests that i) the depth modality exhibits a closer alignment with image data, and ii) UNIALIGN can effectively leverage the additional anchor modality to enhance alignment performance.", + "bbox": [ + 511, + 609, + 905, + 790 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Results for Audio Classification and Retrieval", + "text_level": 1, + "bbox": [ + 511, + 801, + 898, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate audio classification on AudioSet val [16] containing 2M samples across 527 classes, and ESC [65] including 2K data across 50 classes. Audio-text retrieval is evaluated on AudioCaps test [33] which includes 813 audio-text pairs. Moreover, following [107], we", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29649", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0c7782e2985bf55d85b3baf7d7f8db893756a4c96fb6fe00e89b93bd06978e72.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneAnchor ModalityAudioSet mAPESC Top1AudioCaps R@1R@10
Single Task
AVFIC [57]ViT-B---8.737.7
AudioCLIP [22]ViT-BI+T25.969.4--
Multiple Tasks
ImageBind-H [18]ViT-HI17.666.99.342.3
VIT-LENS [38]ViT-LI+T27.280.914.955.2
LanguageBind [109]ViT-LI+T27.791.812.253.2
UNIALIGN†ViT-BI+T23.870.511.749.3
†:An identical model with the same weight across all tasks. \nTable 4. Results for audio classification on AudioSet val [16], ESC[65], and audio-text retrieval on AudioCaps test[33] (§4.5). Due to expiration of links, UNIALIGN is trained with only 0.5M ideas from AudioSet, while methods such as ViT-Lens use 1.6M.
MethodBackbonePre-train DatasetsTrainable ParamsTop-1
Single Task
AST [19]ViT-BN/A86.9M92.6
AST [19]ViT-BAudioSet-20K86.9M96.2
SSAST [20]ViT-BAudioSet-2M89.3M97.8
Multiple Tasks
MetaFormer [107]ViT-BLAION2B1.1M78.3
MetaFormer [107]ViT-BLAION2B86.3M97.0
UniRepLKNet [10]UniRepLKNet-S-55.5M98.5
UNIALIGNViT-BJoint1.8M97.4
", + "bbox": [ + 101, + 88, + 480, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fine-tune UNIALIGN on Speech Commands V2 train[92] for speech recognition.", + "bbox": [ + 89, + 465, + 482, + 496 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metric. Following [18, 22, 38], we measure mAP for AudioSet, top-1 accuracy for ESC and Speech Commands V2, and Recall@ $k$ ( $k = \\{1,5\\}$ ) for AudioCaps.", + "bbox": [ + 89, + 498, + 482, + 544 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance for Audio classification. UNIALIGN achieves promising results on AudioSet and ESC (Table 4). Though inferior to top-leading methods, it is important to note that, due to the expiration of links, a significant number of videos from AudioSet can not be downloaded from YouTube (i.e., 1.6M used in ViT-Lens[38] vs. 0.5M used in UNIALIGN).", + "bbox": [ + 89, + 545, + 482, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance for Audio retrieval. Similarly, despite the absence of 1.1M training data, UNIALIGN can still achieve comparable performance, achieving 11.7 R@1 accuracy.", + "bbox": [ + 89, + 636, + 482, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance for Speech Recognition. As shown in Table 5, UNIALIGN demonstrates remarkable accuracy compared to existing work (e.g., $97.4\\%$ vs. $97.0\\%$ for MetaTransformer[107]), while utilizing significantly fewer trainable parameters (e.g., 1.8M vs. 86.3M). When tuning a similar number of parameters (i.e., 1.1M for Meta-Transformer, and 1.8M for UNIALIGN), our method outperforms Meta-Transformer by $19.1\\%$ (i.e., $97.4\\%$ vs. $78.3\\%$ ).", + "bbox": [ + 89, + 683, + 482, + 804 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Results for Video Classification and Retrieval", + "text_level": 1, + "bbox": [ + 89, + 816, + 475, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets. We utilize MSR-VTT val [96] which contains around 1,000 video-text pairs for zero-shot video-text retrieval, and UCF101 val [77] which contains 3,738 samples across 101 semantic classes for video classification.", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3463065951b9026352e388cb6098c84beaf94ed862132a897d39e1038d7df4c9.jpg", + "table_caption": [ + "Table 5. Results for audio classification on Speech Commands V2 val [92]. All models are fine-tuned on the training set (§4.5)." + ], + "table_footnote": [], + "table_body": "
ModelBackboneRetrieval ModalityTrainable ParamsR@1R@5
Single Task
MIL-NCE [54]ResNeXt-101V88.8M8.625.8
SupportSet [60]R3D-34V93.5M10.430.0
AVFIC [57]ViT-BA+V86.6M19.450.3
InternVideo [89]ViT-LV305M37.571.3
Multiple Tasks
ImageBind [18]ViT-HA+V89.3M36.870.0
ViT-LENS [38]ViT-LA+V127.6M37.672.6
UNIALIGN†ViT-BA+V7.8M37.070.7
", + "bbox": [ + 524, + 89, + 903, + 223 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2dd4a695cf927d3fc6d295b2ca8412e00c15334bc499fb80adffb779ba2fc289.jpg", + "table_caption": [ + "†:An identical model with the same weight across all tasks. Table 6. Results for video retrieval on MSRVTT val [96] (§4.6)." + ], + "table_footnote": [], + "table_body": "
ModelBackboneZero-ShotFine-Tuned
Single Task
ResT [44]ResNet-10158.7-
ActionCLIP [85]ViT-B58.3-
X-CLIP [58]ViT-B72.0-
Multiple Tasks
MetaFormer [107]ViT-B-46.6
ImageBind [18]ViT-H64.898.1
UniBind [50]ViT-H73.793.3
UNIALIGN†ViT-B68.594.4
", + "bbox": [ + 522, + 265, + 903, + 383 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "†:An identical model with the same weight across all tasks. Table 7. Results for video classification on UCF101 val[77] (§4.6).", + "bbox": [ + 514, + 383, + 908, + 410 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metric. Recall@ $k$ and top-1 accuracy are adopted as evaluation metrics for video-text retrieval and classification.", + "bbox": [ + 511, + 422, + 906, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance for Video-Text Retrieval. As shown in Table 6, UNIALIGN achieves a R@1 score of $37.0\\%$ and R@5 score of $70.7\\%$ , which are comparable to advanced methods while using merely 7.8M trainable parameters.", + "bbox": [ + 511, + 454, + 906, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance for Video Classification. We summarize the results for video classification in Table 7. As shown, UNIALIGN achieves an impressive accuracy of $68.5\\%$ on UCF101 val[77] without any fine-tuning, and outperforms ImageBind [18]. After fine-tuning all models on the training set of UCF101, our method achieves the second-highest performance. This may be attributed to the differences in backbone, as ImageBind utilizes ViT-H which could provide enhanced knowledge for downstream fine-tuning.", + "bbox": [ + 511, + 515, + 906, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.7. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 662, + 694, + 678 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Fig. 5, UNIALIGN facilitates accurate cross-modality retrieval. Moreover, modalities can be combined to constrain attributes of the target in another modality.", + "bbox": [ + 511, + 684, + 906, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.8. Diagnostic Experiment", + "text_level": 1, + "bbox": [ + 511, + 739, + 725, + 757 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the core designs of UNIALIGN and gain further insights, we conduct a series of ablation studies. Results are reported on SUN-D test [74] and ModelNet40 test [5].", + "bbox": [ + 511, + 763, + 906, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Modality-Aware MoE Strategy. As illustrated in §3.1, we explore two modality-aware MoE strategies: expert grouping and global routing. Results in Table 8 indicate that dividing experts into modality-specific groups yields better performance. Thus, we adopt it as the default MoE strategy. Expert Network. Table 9 presents the ablation study on", + "bbox": [ + 511, + 809, + 905, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29650", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4b991b2df21667b4f2c7d7f2e19b0e392cb3317f451eee39701e935e95a6ea8f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 130, + 92, + 148, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/83520fce3aad31423abe0e14d9da6379fcd1420d0de4e4d7ad60a12be7325322.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 166, + 93, + 230, + 131 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/456d96ca12fc305a2e1e3da1f7a95c194c781d09a07d7f272a62ae910185cc81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 93, + 300, + 130 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c6fd2eda02bf680ceb17bf6002ca3e3b541cac4060260a377ce35b63dc4afdfe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 93, + 369, + 130 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b3ccf9fc542160290081623aedc6ed43afdd32dc7bb577cb7f184e7e7bc3b67a.jpg", + "image_caption": [ + "Aircraft engine" + ], + "image_footnote": [], + "bbox": [ + 426, + 97, + 460, + 116 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/471ab080e8a627dce21e4b4181fffc0f2c98c6a7c21a2d8255829b11d2707f07.jpg", + "image_caption": [ + "Playing volleyball" + ], + "image_footnote": [], + "bbox": [ + 483, + 92, + 547, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6810b160a4190e4bdce464dc96fb2d2316f5bcb250fdad03e5b71f90fd73a280.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 94, + 647, + 108 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0ae38f0368dbcdec74429c714f72e20ad2eeb06424a5ec06f3f2418eafb48824.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 92, + 727, + 128 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f0e430a016852f7b3c47c1fbd96646c2b3401a0df95fa3e5b9389d5f708aad35.jpg", + "image_caption": [ + "Dog" + ], + "image_footnote": [], + "bbox": [ + 769, + 97, + 803, + 114 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3afd6354d70e4a60eff22c65112d101bbc8c2e07b4d5dbcdda88d0ff298d7664.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 815, + 92, + 880, + 130 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/843190b71736fa50333b9a41a8ddb607ddb61f0f3418f1997d539820dba8c77f.jpg", + "image_caption": [ + "Bathroom" + ], + "image_footnote": [], + "bbox": [ + 112, + 140, + 166, + 174 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/aa2d3368c3f33b2e83c4cd06a90783839eec841cb5b05c73ee59037f9769485b.jpg", + "image_caption": [ + "Car horn" + ], + "image_footnote": [], + "bbox": [ + 173, + 141, + 220, + 178 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2c1080163687ece604b2122bf557fb934620bfe5c7ec6559fb878aa1a19f4a59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 259, + 142, + 294, + 160 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/23c6e52cf2bf9f3e379bd47312c6ca1f70e947cfb2bf6b1cf14a8e3d502a3675.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 307, + 142, + 367, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4e5526b97456a54eb70408a15ce25f6d11d81e320a0b974b55edf620fcde61e1.jpg", + "image_caption": [ + "Playing piano" + ], + "image_footnote": [], + "bbox": [ + 426, + 141, + 457, + 161 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/61ad3d00d9d8dcb8e42597099ef7f8e9189e0657b3b13a19bb371717b537728f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 145, + 550, + 179 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d45979826f20a9f199d0319e0a3f1fc6028e9c0e717e8fc93a53d98516e50f1b.jpg", + "image_caption": [ + "Typing on computer keyboard" + ], + "image_footnote": [], + "bbox": [ + 606, + 142, + 638, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/be6eeea0e7e1d7e64f4839649e660caefb91774ef10f9f5bdb3d54f1aae247e4.jpg", + "image_caption": [ + "Computer room" + ], + "image_footnote": [], + "bbox": [ + 674, + 141, + 730, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0aab9ec8bbf4cb72f6ef7b17757ec22ea3244c74b720d4792b52481b9c52a745.jpg", + "image_caption": [ + "Dishes, pots, and pans" + ], + "image_footnote": [], + "bbox": [ + 767, + 145, + 805, + 159 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9f6cf2dfb1e362bfd37d8436320fc7507f52ed24e340401ef7d948ccdbb17fdc.jpg", + "image_caption": [ + "Kitchen" + ], + "image_footnote": [], + "bbox": [ + 826, + 141, + 880, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/740dcf664438f4b51884a15ffec731ceefe44a0d12d670a7f1c1f461c1e7a078.jpg", + "image_caption": [ + "Figure 5. Qualitative results for cross-modal retrieval. Top and middle: retrieval using one single modality. Bottom: multiple modalities can be composed to retrieve targets across different modalities (§4.7)." + ], + "image_footnote": [], + "bbox": [ + 102, + 191, + 171, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e525cb2760c631dd9ccba1d8d525b57a01c5934b89763f8d7a5f2d45451f4952.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 171, + 196, + 261, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/592482f04f90d5e62f8370ad6295c724381c506d5652fc7ebf90a19e5eefe589.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 264, + 191, + 331, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/22b61ebf4167432c93440d5caf117d5df85beac73bbff9d745ef0948c90a2f29.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 193, + 450, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/90e1f03ce63ec3ed24a28e06a5f6a45976f6127af758f597ac045b20c1796990.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 460, + 196, + 529, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/54d555bb69c8f39999ea89ce05cf8914c4f397bff25a6799064a35975a27b5f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 191, + 604, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3dcfbc1572996109a1295e77a840ba0fcb14ac3cb863464ec96ffab282711614.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 191, + 718, + 231 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fae369fdc26c804450f832f8eb6084c854083b3aa99ae8c01916775787211568.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 191, + 823, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/18afea5e8487fd595d186cbf68d74ed6a48c92152fa973e3588737c150eabf10.jpg", + "image_caption": [ + "Office" + ], + "image_footnote": [], + "bbox": [ + 826, + 191, + 880, + 219 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/93fb001fd47af195a6ccc0f2ae6655ac33632b6eba282d27723f293a1550fc07.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
#StrategiesSUN-DModelNet40
1Expert Grouping45.855.9
2Global Routing45.055.3
", + "bbox": [ + 101, + 276, + 480, + 321 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f92a8bca9046537c7713167c971733885e03844d1e90f14b9028ba73fa161802.jpg", + "table_caption": [ + "Table 8. Ablation studies on modality-aware MoE strategies." + ], + "table_footnote": [], + "table_body": "
Top-k#ExpertsInit.GPU MemSUN-DModelNet40
11rand27.2G45.054.8
11CLIP27.2G45.255.1
12rand27.4G45.155.2
12CLIP27.4G45.255.4
22rand35.9G45.755.7
22CLIP35.9G45.855.9
", + "bbox": [ + 102, + 349, + 480, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "expert network configurations. We investigate the impact of top- $k$ routing, the number of experts per modality, and weight initialization. Starting with a baseline configuration of one expert per modality and random initialization, we achieve 45.0 on SUN-D test [74] and 54.8 on ModelNet40 test [5], with 27.2G GPU memory consumption. Increasing the expert number and initializing weights from the visual encoder of CLIP both lead to improved performance. Due to the constraints of GPU memory, we use two experts per modality and top-2 routing for expert selection. Training Objectives. Next we probe the impact of training objectives i.e., soft modality binding with $\\mathcal{L}_{\\mathrm{soft\\_bind}}$ (cf., Eq. 11), as well as knowledge distillation through $\\mathcal{L}_{\\mathrm{anch\\_distil}}$ (cf., Eq. 12) and $\\mathcal{L}_{\\mathrm{model\\_distil}}$ (cf., Eq. 15). Results are summarized in Table 10, where the first row refers to the baseline that utilizes only paired data from single datasets and is trained with the standard InfoNCE loss. It can be observed that: First, our $\\mathcal{L}_{\\mathrm{soft\\_bind}}$ successfully bridges the learning of multiple modalities for mutual boosting, leads to improved performance on both datasets. Second, $\\mathcal{L}_{\\mathrm{anch\\_distil}}$ and $\\mathcal{L}_{\\mathrm{model\\_distil}}$ could further boost the performance with compelling gains. This verifies the effectiveness of our knowledge distillation strategies, and leads UNIALIGN to a high performance model with optimized efficiency.", + "bbox": [ + 89, + 492, + 482, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Modality Scaling. Finally, we examine the robustness of UNIALIGN to scale different numbers of modalities into a unified model. As shown in Table 11, modality-independent", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a85ca3388f521e0c20408c817886f93d44c18777b364824874d193917561c2d5.jpg", + "table_caption": [ + "Table 9. Ablation studies on configuration of expert network. GPU memory is reported with a batch size of 448 on one Tesla A100." + ], + "table_footnote": [], + "table_body": "
L_NCELsoft_bindLanch_distilLmodel_distilSUN-DModelNet40
✓×××40.438.6
✓✓××41.840.1
✓✓✓×42.340.9
✓✓✓✓45.855.9
", + "bbox": [ + 524, + 277, + 903, + 348 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f76d96fbb1a70a27f16d673e693ca317fc1afd95d27adbbcbc59feb75e137c03.jpg", + "table_caption": [ + "Table 10. Ablation studies on training objectives." + ], + "table_footnote": [], + "table_body": "
#Pre-Training ModalityImageNet-1KSUN-DModelNet40
1I + D + P84.734.431.6
2I + T + D + P84.939.938.1
3I + T + D + P + V85.140.338.4
4I + T + D + P + V + A85.140.438.6
5I + T + D + P + V + A *85.345.855.9
", + "bbox": [ + 524, + 378, + 903, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 11. Ablation studies on modality scaling. Results are reported using only $\\mathcal{L}_{\\mathrm{NCE}}$ as the default training loss. * denotes incorporate $\\mathcal{L}_{\\mathrm{soft\\_bind}}$ , $\\mathcal{L}_{\\mathrm{anch\\_distil}}$ , and $\\mathcal{L}_{\\mathrm{model\\_distil}}$ into training. I: Image, T: Text, D: Depth, V: Video, A: Audio, P: 3D Point Cloud.", + "bbox": [ + 511, + 463, + 905, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "training with merely $\\mathcal{L}_{\\mathrm{NCE}}$ leads to a performance degeneration as more modalities integrated (i.e., row #1-4). This suggests that a straightforward unification of multiple modalities may yield sub-optimal performance for individual ones. Nevertheless, soft modality binding and knowledge distillation (i.e., row #4 vs. #5) could mitigate this negative impact, and boost performance to a level surpassing that achieved with fewer modalities (i.e., row #1 vs. #5).", + "bbox": [ + 511, + 537, + 906, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 676, + 633, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we focus on the all-in-one paradigm for multimodal alignment, which yields a highly efficient and effective model, UNIALIGN, to unify the alignment of all anticipated modalities within one network. Inspired by MoE and knowledge distillation, we propose a modality-aware MoE to tackle multimodal inputs, and generally applicable alignment strategies to utilize unpaired data across datasets. Empirical results suggest that UNIALIGN achieves remarkable performance while being efficient in both training (7.8M trainable parameters to handle 6 modalities within a single training phase) and inference (one identical model with consistent weights across tasks). This work may potentially advance the development of compact, multimodal models.", + "bbox": [ + 511, + 703, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29651", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by the National Science and Technology Major Project (No. 2023ZD0121300), the National Natural Science Foundation of China (No. 62472222, 62302217), Natural Science Foundation of Jiangsu Province (No. BK20240080, BK20220936), \"Pioneer\" and \"Leading Goose\" R&D Program of Zhejiang (No. 2024C01161), the National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, Xi'an Jiaotong University (No. HMHAI-202403), and Bytedance Doubao Fund.", + "bbox": [ + 89, + 92, + 485, + 218 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 232, + 187, + 248 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Romero Adriana, Ballas Nicolas, K Samira Ebrahimi, Chassang Antoine, Gatta Carlo, and Bengio Yoshua. Fitnets: Hints for thin deep nets. In ICLR, 2015. 3", + "[2] Lucas Beyer, Xiaohua Zhai, Amélie Royer, Larisa Markeeva, Rohan Anil, and Alexander Kolesnikov. Knowledge distillation: A good teacher is patient and consistent. In CVPR, 2022. 2", + "[3] Xinhao Cai, Qiqia Lai, Yuwei Wang, Wenguan Wang, Zeren Sun, and Yazhou Yao. Poly kernel inception network for remote sensing detection. In CVPR, 2024. 2", + "[4] Rich Caruana. Multitask learning. Machine learning, 28: 41-75, 1997. 3", + "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5, 7, 8", + "[6] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP, 2020. 5", + "[7] Shaoxiang Chen, Zequn Jie, and Lin Ma. Llava-MoLE: Sparse Mixture of LoRA Experts for Mitigating Data Conflicts in Instruction Finetuning MLLMs. arXiv preprint arXiv:2401.16160, 2024.3", + "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2, 5, 6", + "[9] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. In NeurIPS, 2024. 4", + "[10] Xiaohan Ding, Yiyuan Zhang, Yixiao Ge, Sijie Zhao, Lin Song, Xiangyu Yue, and Ying Shan. Unireplknet: A universal perception large-kernel convnet for audio video point cloud time-series and image recognition. In CVPR, 2024. 6, 7", + "[11] Shihan Dou, Enyu Zhou, Yan Liu, Songyang Gao, Jun Zhao, Wei Shen, Yuhao Zhou, Zhiheng Xi, Xiao Wang, Xiaoran Fan, et al. Loramoe: Revolutionizing mixture of experts for maintaining world knowledge in language model alignment. arXiv preprint arXiv:2312.09979, 2023. 3", + "[12] Dapeng Du, Limin Wang, Huiling Wang, Kai Zhao, and Gangshan Wu. Translate-to-recognize networks for rgb-d scene recognition. In CVPR, 2019. 6", + "[13] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue" + ], + "bbox": [ + 101, + 256, + 480, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Cao. Eva: Exploring the limits of masked visual representation learning at scale. In CVPR, 2023. 2", + "[14] Bilal Faye, Hanane Azzag, and Mustapha Lebbah. Oneencoder: A Lightweight Framework for Progressive Alignment of Modalities. arXiv preprint arXiv:2409.11059, 2024. 2", + "[15] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1-39, 2022. 3", + "[16] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In ICASSP, 2017. 2, 5, 6, 7", + "[17] Rohit Girdhar, Mannat Singh, Nikhila Ravi, Laurens van der Maaten, Armand Joulin, and Ishan Misra. Omnivore: A Single Model for Many Visual Modalities. In CVPR, 2022. 2, 5, 6", + "[18] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In CVPR, 2023. 1, 2, 4, 6, 7", + "[19] Yuan Gong, Yu-An Chung, and James Glass. AST: Audio Spectrogram Transformer. In Interspeech, 2021. 7", + "[20] Yuan Gong, Cheng-I Lai, Yu-An Chung, and James R. Glass. Ssast: Self-Supervised Audio Spectrogram Transformer. In AAAI, 2022. 7", + "[21] Yunhao Gou, Zhili Liu, Kai Chen, Lanqing Hong, Hang Xu, Aoxue Li, Dit-Yan Yeung, James T Kwok, and Yu Zhang. Mixture of cluster-conditional lora experts for vision-language instruction tuning. arXiv preprint arXiv:2312.12379, 2023. 3", + "[22] Andrey Guzhov, Federico Raue, Jorn Hees, and Andreas Dengel. Audioclip: Extending clip to image, text and audio. In ICASSP, 2022. 1, 5, 7", + "[23] Jiaming Han, Kaixiong Gong, Yiyuan Zhang, Jiaqi Wang, Kaipeng Zhang, Dahua Lin, Yu Qiao, Peng Gao, and Xi-angyu Yue. Onellm: One Framework to Align All Modalities with Language. In CVPR, 2024. 2", + "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6", + "[25] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, Nojun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In ICCV, 2019. 3", + "[26] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2, 3", + "[27] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2021. 2, 3", + "[28] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 3, 4" + ], + "bbox": [ + 522, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29652", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Chengsong Huang, Qian Liu, Bill Yuchen Lin, Tianyu Pang, Chao Du, and Min Lin. Lorahub: Efficient cross-task generalization via dynamic lora composition. arXiv preprint arXiv:2307.13269, 2023. 3", + "[30] Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson W.H. Lau, Wanli Ouyang, and Wangmeng Zuo. Clip2point: Transfer clip to point cloud classification with image-depth pre-training. In ICCV, 2023. 6", + "[31] Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. Mixtral of experts. arXiv preprint arXiv:2401.04088, 2024. 3", + "[32] Byoungjip Kim, Sungik Choi, Dasol Hwang, Moontae Lee, and Honglak Lee. Transferring pre-trained multimodal representations with cross-modal similarity matching. In NeurIPS, 2022. 3", + "[33] Chris Dongjoo Kim, Byeongchang Kim, Hyunmin Lee, and Gunhee Kim. Audiocaps: Generating captions for audios in the wild. In NAACL-HLT, 2019. 6, 7", + "[34] Jangho Kim, SeongUk Park, and Nojun Kwak. Paraphrasing complex network: Network compression via factor transfer. In NeurIPS, 2018. 3", + "[35] Aran Komatsuzaki, Joan Puigcerver, James Lee-Thorp, Carlos Riquelme Ruiz, Basil Mustafa, Joshua Ainslie, Yi Tay, Mostafa Dehghani, and Neil Houlsby. Sparse upcycling: Training mixture-of-experts from dense checkpoints. In ICLR, 2023. 3", + "[36] Nikos Komodakis and Sergey Zagoruyko. Paying more attention to attention: improving the performance of convolutional neural networks via attention transfer. In ICLR, 2017. 3", + "[37] Konrad P Kording, Ulrik Beierholm, Wei Ji Ma, Steven Quartz, Joshua B Tenenbaum, and Ladan Shams. Causal inference in multisensory perception. PLoS one, 2(9):e943, 2007. 1", + "[38] Stan Weixian Lei, Yixiao Ge, Kun Yi, Jianfeng Zhang, Difei Gao, Dylan Sun, Yuying Ge, Ying Shan, and Mike Zheng Shou. Vit-Lens: Towards Omni-modal Representations. In CVPR, 2024. 2, 5, 6, 7", + "[39] Weixian Lei, Yixiao Ge, Kun Yi, Jianfeng Zhang, Difei Gao, Dylan Sun, Yuying Ge, Ying Shan, and Mike Zheng Shou. Vit-lens: Towards omni-modal representations. In CVPR, 2024. 1, 2, 3", + "[40] Po-han Li, Sandeep P Chinchali, and Ufuk Topcu. Csa: Data-efficient mapping of unimodal features to multimodal features. arXiv preprint arXiv:2410.07610, 2024. 2", + "[41] Yunxin Li, Shenyuan Jiang, Baotian Hu, Longyue Wang, Wanqi Zhong, Wenhan Luo, Lin Ma, and Min Zhang. UniMoE: Scaling Unified Multimodal LLMs with Mixture of Experts. arXiv, abs/2405.11273, 2024. 3", + "[42] Valerii Likhosherstov, Anurag Arnab, Krzysztof Choromanski, Mario Lucic, Yi Tay, Adrian Weller, and Mostafa Dehghani. Polyvit: Co-training vision transformers on images, videos and audio. arXiv preprint arXiv:2111.12993, 2021. 3, 5", + "[43] Bin Lin, Zhenyu Tang, Yang Ye, Jiaxi Cui, Bin Zhu, Peng Jin, Jina Huang, Junwu Zhang, Yatian Pang, Munan Ning," + ], + "bbox": [ + 99, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "et al. Moe-llava: Mixture of experts for large vision-language models. arXiv preprint arXiv:2401.15947, 2024. 3", + "[44] Chung-Ching Lin, Kevin Lin, Lijuan Wang, Zicheng Liu, and Linjie Li. Cross-modal representation learning for zero-shot action recognition. In CVPR, 2022. 7", + "[45] Haokun Lin, Haoli Bai, Zhili Liu, Lu Hou, Muyi Sun, Linqi Song, Ying Wei, and Zhenan Sun. Mope-CLIP: Structured Pruning for Efficient Vision-Language Models with Module-wise Pruning Error Metric. In CVPR, 2024. 3", + "[46] Xi Victoria Lin, Akshit Shrivastava, Liang Luo, Srinivasan Iyer, Mike Lewis, Gargi Gosh, Luke Zettlemoyer, and Armen Aghajanyan. Moma: Efficient early-fusion pretraining with mixture of modality-aware experts. arXiv preprint arXiv:2407.21770, 2024. 3", + "[47] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 6", + "[48] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022. 6", + "[49] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 2", + "[50] Yuanhuiyi Lyu, Xu Zheng, Jiazhou Zhou, and Lin Wang. Unibind: Llm-Augmented Unified and Balanced Representation Space to Bind Them All. In CVPR, 2024. 2, 7", + "[51] Jian Ma, Wenguan Wang, Yi Yang, and Feng Zheng. Ms2sl: Multimodal spoken data-driven continuous sign language production. In ACL Findings, 2024. 2", + "[52] Wenxuan Ma, Shuang Li, Jinming Zhang, Chi Harold Liu, Jingxuan Kang, Yulin Wang, and Gao Huang. Borrowing Knowledge From Pre-trained Language Model: A New Data-efficient Visual Learning Paradigm. In ICCV, 2023. 2", + "[53] M Alex Meredith and Barry E Stein. Visual, auditory, and somatosensory convergence on cells in superior colliculus results in multisensory integration. Journal of neurophysiology, 56(3):640-662, 1986. 1", + "[54] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, 2019. 7", + "[55] Seyed Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved knowledge distillation via teacher assistant. In AAAI, 2020. 3", + "[56] Luca Moschella, Valentino Maiorca, Marco Fumero, Antonio Norelli, Francesco Locatello, and Emanuele Rodola. Relative representations enable zero-shot latent space communication. In ICLR, 2023. 2", + "[57] Arsha Nagrani, Paul Hongsuck Seo, Bryan Seybold, Anja Hauth, Santiago Manen, Chen Sun, and Cordelia Schmid. Learning audio-video modalities from image captions. In ECCV, 2022. 7" + ], + "bbox": [ + 522, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29653", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[58] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. In European Conference on Computer Vision, pages 1-18. Springer, 2022. 7", + "[59] Antonio Norelli, Marco Fumero, Valentino Maiorca, Luca Moschella, Emanuele Rodola, and Francesco Locatello. Asif: Coupled data turns unimodal models to multimodal without training. In NeurIPS, 2023. 2", + "[60] Mandela Patrick, Po-Yao Huang, Yuki Asano, Florian Metze, Alexander Hauptmann, Joao Henriques, and Andrea Vedaldi. Support-set bottlenecks for video-text representation learning. arXiv preprint arXiv:2010.02824, 2020. 7", + "[61] Gensheng Pei, Fumin Shen, Yazhou Yao, Guo-Sen Xie, Zhenmin Tang, and Jinhui Tang. Hierarchical feature alignment network for unsupervised video object segmentation. In ECCV, 2022. 2", + "[62] Gensheng Pei, Tao Chen, Xiruo Jiang, Huafeng Liu, Zeren Sun, and Yazhou Yao. Videomac: Video masked autoencoders meet convnets. In CVPR, 2024. 2", + "[63] Renjing Pei, Jianzhuang Liu, Weimian Li, Bin Shao, Songcen Xu, Peng Dai, Juwei Lu, and Youliang Yan. Clipping: Distilling clip-based models with a student base for video-language retrieval. In CVPR, 2023. 3", + "[64] Zhiliang Peng, Wei Huang, Shanzhi Gu, Lingxi Xie, Yaowei Wang, Jianbin Jiao, and Qixiang Ye. Conformer: Local features coupling global representations for visual recognition. In ICCV, 2021. 6", + "[65] Karol J Piczak. Esc: Dataset for environmental sound classification. In ACM MM, 2015. 6, 7", + "[66] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with Reconstruct: Contrastive 3d Representation Learning Guided by Generative Pretraining. In ICML, 2023. 6", + "[67] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 5", + "[68] Mike Ranzinger, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. Am-RADIO: Agglomerative Vision Foundation Model Reduce All Domains Into One. In CVPR, 2024. 3, 5", + "[69] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. In NeurIPS, 2021. 3", + "[70] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. In ICML, 2023. 6", + "[71] Jinghuan Shang, Karl Schmeckpeper, Brandon B. May, Maria Vittoria Minniti, Tarik Kelestemur, David Watkins, and Laura Herlant. Theia: Distilling Diverse Vision Foundation Models for Robot Learning. In CoRL, 2024. 3", + "[72] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean." + ], + "bbox": [ + 99, + 92, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In ICLR, 2017. 2, 3", + "[73] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 6", + "[74] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In CVPR, 2015. 5, 6, 7, 8", + "[75] Xinhang Song, Luis Herranz, and Shuqiang Jiang. Depth cnns for rgb-d scene recognition: Learning from scratch better than transferring from rgb-cnns. In AAAI, 2017. 6", + "[76] Xinhang Song, Shuqiang Jiang, Bohan Wang, Chengpeng Chen, and Gongwei Chen. Image representations with spatial object-to-object relations for rgb-d scene recognition. IEEE TIP, 29:525-537, 2019. 6", + "[77] K Soomro. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 7", + "[78] Siddharth Srivastava and Gaurav Sharma. Omnivec2 - A Novel Transformer based Network for Large Scale Multimodal and Multitask Learning. In CVPR, 2024. 2", + "[79] Ximeng Sun, Pengchuan Zhang, Peizhao Zhang, Hardik Shah, Kate Saenko, and Xide Xia. Dime-fm: Distilling multimodal and efficient foundation models. In ICCV, 2023. 3", + "[80] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. In ECCV, 2022. 6", + "[81] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 4", + "[82] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 6", + "[83] Simon Vandenhende, Stamatios Georgoulis, Wouter Van Gansbeke, Marc Proesmans, Dengxin Dai, and Luc Van Gool. Multi-task learning for dense prediction tasks: A survey. IEEE TPAMI, 44(7):3614-3633, 2021. 3", + "[84] Haoxiang Wang, Pavan Kumar Anasosalu Vasu, Fartash Faghri, Raviteja Vemulapalli, Mehrdad Farajtabar, Sachin Mehta, Mohammad Rastegari, Oncel Tuzel, and Hadi Pouransari. Sam-clip: Merging vision foundation models towards semantic and spatial understanding. In CVPR, 2024. 3", + "[85] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. 7", + "[86] Peng Wang, Shijie Wang, Junyang Lin, Shuai Bai, Xiaohuan Zhou, Jingren Zhou, Xinggang Wang, and Chang Zhou. One-peace: Exploring one general representation model toward unlimited modalities. arXiv preprint arXiv:2305.11172, 2023. 5", + "[87] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu," + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "29654", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. In CVPR, 2023. 6", + "[88] Wenguan Wang, Yi Yang, and Yunhe Pan. Visual knowledge in the big model era: Retrospect and prospect. Frontiers of Information Technology & Electronic Engineering, 26(1):1-19, 2025. 2", + "[89] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 7", + "[90] Zehan Wang, Ziang Zhang, Luping Liu, Yang Zhao, Haifeng Huang, Tao Jin, and Zhou Zhao. Extending multi-modal contrastive representations. arXiv preprint arXiv:2310.08884, 2023. 2", + "[91] Zehan Wang, Yang Zhao, Haifeng Huang, Jiageng Liu, Aoxiong Yin, Li Tang, Linjun Li, Yongqi Wang, Zhang, and Zhou Zhao. Connecting multi-modal contrastive representations. In NeurIPS, 2023. 2", + "[92] Pete Warden. Speech commands: A dataset for limited-vocabulary speech recognition. arXiv preprint arXiv:1804.03209, 2018.7", + "[93] Yixuan Wei, Han Hu, Zhenda Xie, Zheng Zhang, Yue Cao, Jianmin Bao, Dong Chen, and Baining Guo. Contrastive learning rivals masked image modeling in fine-tuning via feature distillation. Tech Report, 2022. 3", + "[94] Kan Wu, Jinnian Zhang, Houwen Peng, Mengchen Liu, Bin Xiao, Jianlong Fu, and Lu Yuan. Tinyvit: Fast Pretraining Distillation for Small Vision Transformers. In ECCV, 2022. 6", + "[95] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, 2015. 2, 6", + "[96] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. 2, 7", + "[97] Peng Xu, Xiatian Zhu, and David A Clifton. Multimodal learning with transformers: A survey. IEEE TPAMI, 45 (10):12113-12132, 2023. 1", + "[98] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In CVPR, 2023. 2, 5, 6", + "[99] Chuanguang Yang, Zhulin An, Libo Huang, Junyu Bi, Xinqiang Yu, Han Yang, Boyu Diao, and Yongjun Xu. ClipKD: An Empirical Study of CLIP Model Distillation. In CVPR, 2024. 3, 5", + "[100] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models (exemplified as a video agent). In ICML, 2024. 2", + "[101] Yazhou Yao, Tao Chen, Guo-Sen Xie, Chuanyi Zhang, Fumin Shen, Qi Wu, Zhenmin Tang, and Jian Zhang. Nonsalient region object mining for weakly supervised semantic segmentation. In CVPR, 2021. 2" + ], + "bbox": [ + 93, + 92, + 482, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[102] Yazhou Yao, Zeren Sun, Chuanyi Zhang, Fumin Shen, Qi Wu, Jian Zhang, and Zhenmin Tang. Jo-src: A contrastive approach for combating noisy labels. In CVPR, 2021. 2", + "[103] Junbo Yin, Jianbing Shen, Runnan Chen, Wei Li, Ruigang Yang, Pascal Frossard, and Wenguan Wang. Is-fusion: Instance-scene collaborative fusion for multimodal 3d object detection. In ECCV, 2024. 2", + "[104] Shan You, Chang Xu, Chao Xu, and Dacheng Tao. Learning from multiple teacher networks. In KDD, 2017. 3", + "[105] Fei Yuan, Linjun Shou, Jian Pei, Wutao Lin, Ming Gong, Yan Fu, and Daxin Jiang. Reinforced multi-teacher selection for knowledge distillation. In AAAI, 2021. 3", + "[106] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xu-peng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In CVPR, 2022. 6", + "[107] Yiyuan Zhang, Kaixiong Gong, Kaipeng Zhang, Hongsheng Li, Yu Qiao, Wanli Ouyang, and Xiangyu Yue. Meta-transformer: A unified framework for multimodal learning. arXiv preprint arXiv:2307.10802, 2023. 2, 3, 6, 7", + "[108] Yanqi Zhou, Tao Lei, Hanxiao Liu, Nan Du, Yanping Huang, Vincent Zhao, Andrew M Dai, Quoc V Le, James Laudon, et al. Mixture-of-experts with expert choice routing. In NeurIPS, 2022. 2", + "[109] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, WANG HongFa, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei Li, Cai Wan Zhang, Zhifeng Li, Wei Liu, and Li Yuan. Languagebind: Extending Video-Language Pretraining to N-modality by Language-based Semantic Alignment. In ICLR, 2024. 1, 2, 7", + "[110] Xiatian Zhu, Shaogang Gong, et al. Knowledge distillation by on-the-fly native ensemble. In NeurIPS, 2018. 2", + "[111] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In ICCV, 2023. 6", + "[112] Konrad Zuchniak. Multi-teacher knowledge distillation as an effective method for compressing ensembles of neural networks. arXiv preprint arXiv:2302.07215, 2023. 3" + ], + "bbox": [ + 516, + 92, + 903, + 643 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "29655", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/UNIALIGN_ Scaling Multimodal Alignment within One Unified Model/7e109366-fc0f-438c-9b91-2ba7f07e59e9_model.json b/2025/UNIALIGN_ Scaling Multimodal Alignment within One Unified Model/7e109366-fc0f-438c-9b91-2ba7f07e59e9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5c7e83f911f9d787adb50eb38933e160b4ca0f4d --- /dev/null +++ b/2025/UNIALIGN_ Scaling Multimodal Alignment within One Unified Model/7e109366-fc0f-438c-9b91-2ba7f07e59e9_model.json @@ -0,0 +1,3689 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.146, + 0.131, + 0.853, + 0.152 + ], + "angle": 0, + "content": "UNIALIGN: Scaling Multimodal Alignment within One Unified Model" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.18, + 0.842, + 0.201 + ], + "angle": 0, + "content": "Bo Zhou\\(^{1*}\\), Lilei Li\\(^{2*}\\), Yujia Wang\\(^{3}\\), Huafeng Liu\\(^{1}\\), Yazhou Yao\\(^{1\\dagger}\\), Wenguan Wang\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.201, + 0.805, + 0.217 + ], + "angle": 0, + "content": "\\(^{1}\\) Nanjing University of Science and Technology, \\(^{2}\\) Zhejiang University, \\(^{3}\\) Zhejiang Sci-Tech University" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.22, + 0.782, + 0.234 + ], + "angle": 0, + "content": "https://github.com/NUST-Machine-Intelligence-Laboratory/UNIALIGN" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.301, + 0.485, + 0.68 + ], + "angle": 0, + "content": "We present UNIALIGN, a unified model to align an arbitrary number of modalities (e.g., image, text, audio, 3D point cloud, etc.) through one encoder and a single training phase. Existing solutions typically employ distinct encoders for each modality, resulting in increased parameters as the number of modalities grows. In contrast, UNIALIGN proposes a modality-aware adaptation of the powerful mixture-of-experts (MoE) schema and further integrates it with Low-Rank Adaptation (LoRA), efficiently scaling the encoder to accommodate inputs in diverse modalities while maintaining a fixed computational overhead. Moreover, prior work often requires separate training for each extended modality. This leads to task-specific models and further hinders the communication between modalities. To address this, we propose a soft modality binding strategy that aligns all modalities using unpaired data samples across datasets. Two additional training objectives are introduced to distill knowledge from well-aligned anchor modalities and prior multimodal models, elevating UNIALIGN into a high performance multimodal foundation model. Experiments on 11 benchmarks across 6 different modalities demonstrate that UNIALIGN could achieve comparable performance to SOTA approaches, while using merely 7.8M trainable parameters and maintaining an identical model with the same weight across all tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.706, + 0.222, + 0.722 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.731, + 0.484, + 0.868 + ], + "angle": 0, + "content": "Humans accomplish tasks by processing and combining signals from different modalities [37, 53]. This gives rise to multimodal learning, which aims to build AI models that can extract and relate information across multimodal inputs [97]. Recent efforts have been made to learn a unified representation space via the alignment between modalities [18, 22, 39]. Though enabling extension to novel modalities, existing solutions typically rely on a one-versus-one paradigm where merely one extended modality (e.g., point" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.269, + 0.906, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.514, + 0.907, + 0.584 + ], + "angle": 0, + "content": "Figure 1. Comparison of architectures between standard multi-modal alignment methods [18, 39, 109] and UNIALIGN: (a) Existing work adopts specialized encoders tailored to each modality; (b) UNIALIGN unifies the alignment across multiple modalities in a single model, with one training phase across all datasets." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.597, + 0.907, + 0.793 + ], + "angle": 0, + "content": "cloud, audio) could be aligned with the anchor modalities (e.g., text, image) at each time (Fig. 1(a)). This presents several drawbacks: i) one dedicated encoder is required for each modality, leading to exponential increases in training cost and parameter count as the number of aligned modalities grows, severely limiting scalability; ii) task-specific training for each dataset, results in multiple independent models for different tasks; and iii) the separate alignment of each extended modality hinders the model to leverage the full breadth of data across all datasets to shape the shared embedding space from a global view. This further impedes cross-modal knowledge transfer where insights gained from one modality could benefit understanding in others." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.902 + ], + "angle": 0, + "content": "These observations prompt us to rethink: Is this one-versus-one alignment paradigm essential? and Can we achieve joint alignment of all modalities within one single model? In response, we propose UNIALIGN (Fig. 1(b)), which embraces a all-in-one design to realize multimodal alignment by constructing a foundation model that includes: 1 one base encoder to UNify the encoding of all anchor and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.393, + 0.888 + ], + "angle": 0, + "content": "*The first two authors contribute equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.307, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author: Yazhou Yao." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.393, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "29644" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.123 + ], + "angle": 0, + "content": "extended modalities, and \\(②\\) one training phase to ALIGN all modalities with all samples across datasets collectively." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.125, + 0.486, + 0.716 + ], + "angle": 0, + "content": "Concretely, to achieve ①, recent work has consistently demonstrated that Transformers pre-trained on one modality possess exceptional generalization ability to novel modalities [13, 67]. Moreover, the mixture-of-experts (MoE) [72] schema presents a highly efficient strategy to scale model capacity under a fixed computational cost [108]. These insights motivate us to build a foundational model accommodating multiple modalities upon a single pre-trained Transformer (e.g., CLIP [67]). Here, MoE can dynamically scale and allocate model capacities by selectively activating experts to encode inputs across diverse modalities. We explore two strategies targeting at expert grouping and routing mechanism to enhance the awareness of different modalities within MoE. To preserve the knowledge learned from anchor modalities, Low-Rank Adaptation (LoRA) [27] is integrated with MoE, enabling a significant reduction in training overhead. To fulfill ②, we propose a soft modality binding strategy, leveraging massive unpaired data samples across modalities and datasets. Specifically, given unpaired data from diverse datasets with different modalities, the class labels of samples are encoded into text embeddings to create a unified representation. The semantic similarity between text embeddings can inform the alignment between modalities (e.g., audio and video), by encouraging the similarity between modality embeddings and textual label embeddings (e.g., audio \\(\\leftrightarrow\\) textvideo and video \\(\\leftrightarrow\\) textaudio) to be identical as the semantic similarity (i.e., \\(text_{\\mathrm{video}} \\leftrightarrow \\text{text}_{\\mathrm{audio}}\\)). Moreover, rooted in the principles of knowledge distillation [2, 26, 110], the alignment between extended and anchor modalities is guided by the knowledge persisting between anchor modalities, which are well-aligned in semantics compared to extended modalities. To further boost the performance, we transfer knowledge from multiple well-trained foundation models which only contain partial modalities. Given the complexity of multimodal, multi-dataset joint training, a specialized batching strategy combined with modality-aware gradient accumulation is proposed, to mitigate the potential dispersion in optimization across different modalities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.485, + 0.902 + ], + "angle": 0, + "content": "UNIALIGN enjoys several advantages: First, it supports flexible alignment between arbitrary number of anchor and extended modalities, while ensuring negligible increase in parameter size. Second, the alignment for all modalities is accomplished in one training session, significantly improving training efficiency. Third, unpaired data from different modalities can be used collectively for optimization. Uncommon modalities with scarce data could take full advantages of the rich resource of common modalities. Fourth, benefited from soft modality binding, extended modalities are bridged for mutual boosting. Fifth, a compact and high-performance multimodal foundational" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.182 + ], + "angle": 0, + "content": "model generally applicable to all tasks is delivered at minimal cost. This contrasts with existing approaches, which typically provide multiple loosely integrated modality encoders tailored to specific tasks, thereby enabling broader transferability to other domains and downstream applications [3, 8, 16, 61, 62, 95, 96, 101, 102]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.909, + 0.351 + ], + "angle": 0, + "content": "By embracing the all-in-one philosophy, UNIALIGN demonstrates remarkable efficiency in both training cost and parameter size. After a single pre-training step without further tuning, UNIALIGN achieves top-leading performance on 11 benchmarks with one identical model and 7.8M trainable parameters across 6 modalities (previous SOTA [18, 39] needing over 80M trainable parameters for each extended modality). This evidences the great potential of our unified modality encoding and aligning strategies. We hope this work could stimulate the community to rethink multimodal alignment from a universal and compact perspective." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.362, + 0.656, + 0.379 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.388, + 0.909, + 0.81 + ], + "angle": 0, + "content": "Multimodal Learning. Multimodal learning has attracted significant attention given its ability to address a wide range of applications by integrating information from diverse modalities such as text, images, audio, and video[49, 51, 67, 103]. Current research in this field generally falls into three categories: First, methods involving joint training of multiple modalities aim to improve performance across different downstream tasks, by leveraging shared representations[17, 78, 107]. Though incorporating multiple modalities as inputs, these work lacks explicit alignment between modalities. Second, inspired by the successes of vision-language alignment [67], some approaches extend this alignment to additional modalities[18, 38, 40, 56, 59, 91, 98, 109]. These methods typically require an independent encoder for each modalities. Though [14, 18, 39, 90] maintain an consistent architecture across tasks, they conduct separate training on individual datasets, resulting in multiple task-specific models. Additionally, these solutions often rely on massive paired data to achieve robust zero-shot performance across various tasks, ignoring the usage of rich unpaired data. Third, the integration of multimodal inputs with large language models (LLMs) has opened new avenues for multimodal learning[23, 50, 52, 88, 100]. In these approaches, all modalities are projected into the same representation space as LLMs, without alignment among extended modalities and supporting merely one anchor modality (i.e., text). This leads to suboptimal performance on certain vision tasks and limits the applications primarily to language-related fields." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this work, we propose UNIALIGN, supporting alignment between an arbitrary number of anchor and extended modalities within one model using one unified modality encoder. After a single training phase, UNIALIGN can handle a dozen of downstream tasks without any fine-tuning. UNIALIGN also achieves high training efficiency, utilizing" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "29645" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.395, + 0.106 + ], + "angle": 0, + "content": "merely 7.8M training parameters for all tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.107, + 0.484, + 0.408 + ], + "angle": 0, + "content": "Multi-Task Learning. To maximize the utilization of resource and improve generalization across tasks, multi-task learning seeks to develop models that can handle multiple tasks simultaneously [4]. Traditional methods often require separate subnetworks for each task, which hinders scalability [83]. In recent years, the mixture-of-experts (MoE) architecture [15, 72] has gained prominence in multi-task learning [31]. MoE leverages multiple subnetworks, or \"experts\", that are selectively activated to optimize task performance while reducing computational costs [35]. This approach proves beneficial across various domains, such as language processing [31, 35] and vision tasks [41, 43, 69]. With the success of LoRA [27], there has been a surge to integrate MoE with LoRA, which shows promise in enhancing both computational efficiency and task adaptability in large-scale models [11, 21, 28, 29]. However, these methods are mainly confined to text and image domains [7, 11], and lack alignment between modalities [46]. In contrast, UNIALIGN extends MoE to a multimodal setup, devising modality-aware expert selection and routing mechanism." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.409, + 0.484, + 0.62 + ], + "angle": 0, + "content": "Knowledge Distillation. Knowledge distillation [26, 34, 55] is an essential technique for model compression, enabling the transfer of knowledge from large, computationally intensive models to smaller, more efficient ones. This is particularly valuable for multimodal foundation models, where the computational overload is heavy [32, 45, 79, 99]. Initially, knowledge distillation focused on extracting knowledge from individual large models, typically using the logits or intermediate feature representations to guide the training of student models [1, 25, 36, 63, 93]. Recent research has expanded this concept to integrate capabilities from multiple models, creating hybrid models that not only improve task-specific performance but also broaden the range of potential applications [68, 71, 84, 104, 105, 112]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.62, + 0.484, + 0.71 + ], + "angle": 0, + "content": "Compared to existing work, UNIALIGN distills knowledge from both well-aligned anchor modalities, and multiple well-trained foundation models specialized in partial modalities. This allows our method to handle a broader range of tasks that require understanding and processing information with different composition of diverse modalities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.724, + 0.182, + 0.739 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.484, + 0.824 + ], + "angle": 0, + "content": "In this section, we first introduce the unified encoding of all modalities within one single model (§3.1), then describe the universal modality alignment strategy across all datasets (§3.2), and finally present the implementation details of UNIALIGN (§3.3)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.336, + 0.85 + ], + "angle": 0, + "content": "3.1. Unified Modality Encoding" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.855, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Preliminary: Mixture-of-Experts (MoE). Let \\(X \\in \\mathbb{R}^{L \\times D}\\) denote the tokenized input, where \\(L\\) is the sequence length and \\(D\\) is the token embedding size. In vanilla Transformer," + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.085, + 0.885, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.396, + 0.874, + 0.41 + ], + "angle": 0, + "content": "Figure 2. Architecture of the unified modality encoder." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.416, + 0.905, + 0.445 + ], + "angle": 0, + "content": "each encoder block comprises a multi-head self-attention (MSA) layer and a feed-forward network (FFN), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.45, + 0.905, + 0.467 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {X} ^ {\\prime} = \\boldsymbol {X} + \\mathcal {F} ^ {\\mathrm {M S A}} (\\ln (\\boldsymbol {X})), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.47, + 0.905, + 0.488 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {X} ^ {\\prime \\prime} = \\boldsymbol {X} ^ {\\prime} + \\mathcal {F} ^ {\\mathrm {F F N}} (\\ln (\\boldsymbol {X} ^ {\\prime})), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.498, + 0.905, + 0.543 + ], + "angle": 0, + "content": "where \\(X''\\) is the outputs and LN represents layer normalization. In the mixture-of-experts (MoE) schema, an MoE function (i.e., \\(\\mathcal{F}^{\\mathrm{MoE}}\\)) is adopted to replace \\(\\mathcal{F}^{\\mathrm{FFN}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.546, + 0.905, + 0.563 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {X} ^ {\\prime \\prime} = \\boldsymbol {X} ^ {\\prime} + \\mathcal {F} ^ {\\mathrm {M o E}} (\\ln \\left(\\boldsymbol {X} ^ {\\prime}\\right)), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.569, + 0.567, + 0.905, + 0.586 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\mathrm {M o E}} \\left(\\boldsymbol {X} ^ {\\prime}\\right)) = \\sum_ {i = 1} ^ {n} \\mathcal {G} \\left(\\boldsymbol {X} ^ {\\prime}\\right) _ {i} \\mathcal {F} _ {i} ^ {\\text {E x p e r t}} \\left(\\boldsymbol {X} ^ {\\prime}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.905, + 0.655 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}^{\\mathrm{MoE}}\\) jointly aggregates outputs from multiple expert networks (i.e., \\(\\mathcal{F}^{\\mathrm{Expert}}\\)) according to the router \\(\\mathcal{G}(X')_i = \\text{softmax}(\\mathcal{F}^{\\mathrm{Router}}(X'))_i\\) to assign weights for different experts and render the predictions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.906, + 0.791 + ], + "angle": 0, + "content": "Multimodal Adaptation with MoE. We aim to unify the encoding of multimodal inputs within one foundation model (e.g., the image encoder of CLIP[67]). As illustrated in Fig.2, inputs of various modalities are first converted into sequential tokens. We follow [39, 42, 107] to preprocess multimodal data and implement modality-specific tokenizers, with details summarized in Appendix. To adapt the encoding of samples from various modalities, UNIALIGN explores two modality-aware MoE strategies:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.791, + 0.905, + 0.837 + ], + "angle": 0, + "content": "- Modality-Aware Expert Grouping divides all \\(\\mathcal{F}^{\\mathrm{Expert}}\\) in one MoE layer into multiple sets, where each set \\(\\mathcal{B}\\) contains \\(n\\) experts to process inputs from a specific modality \\(\\mathcal{M}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.842, + 0.905, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\mathcal {B} _ {\\mathcal {M}} = \\left\\{\\mathcal {F} _ {\\mathcal {M}, 1} ^ {\\text {E x p e r t}}, \\dots , \\mathcal {F} _ {\\mathcal {M}, n} ^ {\\text {E x p e r t}} \\right\\}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Under this setup, following [46], we construct a hierarchical routing mechanism. In the first routing round, tokens are" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29646" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.182 + ], + "angle": 0, + "content": "directed to a specific expert set based on the modality of inputs. Next, within a expert set, a learnable router \\(\\mathcal{F}_{\\mathcal{M}}^{\\text{Router}}\\) for modality \\(\\mathcal{M}\\) assigns weights to each individual expert conditioned on the input tokens. This separation lets each expert set to specialize in one modality, and allows for easy expansion to novel modalities by loading new expert sets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.483, + 0.244 + ], + "angle": 0, + "content": "- Modality-Aware Global Routing uses multiple modality-specific routers, to direct tokens across all experts in a MoE layer without grouping. Given the routing function \\(\\mathcal{F}_{\\mathcal{M}}^{\\text{Router}}\\) for modality \\(M\\), the global weight \\(\\mathcal{G}_{\\mathcal{M}}\\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.252, + 0.483, + 0.271 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {\\mathcal {M}} \\left(\\boldsymbol {X} ^ {\\prime}\\right) _ {i} = \\operatorname {s o f t m a x} \\left(\\mathcal {F} _ {\\mathcal {M}} ^ {\\text {R o u t e r}} \\left(\\boldsymbol {X} ^ {\\prime}\\right)\\right) _ {i}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.282, + 0.483, + 0.357 + ], + "angle": 0, + "content": "By allowing the experts to be accessible to each router, this strategy enables cross-modality knowledge sharing among experts. However, since the experts are not explicitly specialized for one modality, the generalization potentially at the expense of fine-grained modality specialization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.484, + 0.463 + ], + "angle": 0, + "content": "Modality-Aware LoRA Expert. Low-Rank Adaptation (LoRA) [28] has demonstrated both effectiveness and efficiency in fine-tuning large models [9, 81]. Therefore, to maintain the knowledge in pre-trained foundation models, the original feed-forward network (i.e., \\(\\mathcal{F}^{\\mathrm{FFN}}\\) in Eq. 2) are preserved in a fixed state instead of being replaced during the forward of MoE (i.e., Eq. 3), resulting in:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.472, + 0.483, + 0.49 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {X} ^ {\\prime \\prime} = \\boldsymbol {X} ^ {\\prime} + \\mathcal {F} ^ {\\mathrm {F F N}} (\\ln (\\boldsymbol {X} ^ {\\prime})) + \\mathcal {F} ^ {\\mathrm {M o E}} (\\ln (\\boldsymbol {X} ^ {\\prime})). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.483, + 0.547 + ], + "angle": 0, + "content": "Here the experts in \\(\\mathcal{F}^{\\mathrm{MoE}}\\) leverage \\(\\mathcal{F}^{\\mathrm{FFN}}\\) as the base weight for LoRA adaption, with each of them employs a specific instance of LoRA (i.e., \\(\\mathcal{F}^{\\mathrm{LoRA}}\\)) parameterized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.554, + 0.483, + 0.572 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\operatorname {M o E}} \\left(\\boldsymbol {X} ^ {\\prime}\\right)) = \\sum_ {i = 1} ^ {n} \\mathcal {G} \\left(\\boldsymbol {X} ^ {\\prime}\\right) _ {i} \\mathcal {F} _ {i} ^ {\\text {L o R A}} \\left(\\boldsymbol {X} ^ {\\prime}\\right), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.574, + 0.482, + 0.599 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {i} ^ {\\text {L o R A}} \\left(\\boldsymbol {X} ^ {\\prime}\\right) = \\frac {\\alpha}{r} \\cdot \\boldsymbol {B} _ {i} \\boldsymbol {A} _ {i} \\boldsymbol {X} ^ {\\prime}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.483, + 0.703 + ], + "angle": 0, + "content": "where low-rank matrices \\( \\mathbf{A} \\in \\mathbb{R}^{d_{in} \\times r} \\) and \\( \\mathbf{B} \\in \\mathbb{R}^{r \\times d_{out}} \\) are scaled by a factor \\( \\alpha \\) normalized by the rank \\( r \\ll \\min(d_{in}, d_{out}) \\). In this way, we achieve seamless integration of LoRA with MoE, enabling efficient tuning and knowledge preservation of the foundation models, while supporting adaptation to inputs across various modalities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.362, + 0.729 + ], + "angle": 0, + "content": "3.2. Universal Modality Alignment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Soft Modality Binding across Datasets. Prior work [18] reveals that modalities expressed in a joint embedding space could be automatically aligned through an intermediate modality. For example, the alignment of and , could lead to the alignment between with Image as the bridge. However, this approach requires massive naturally paired samples between anchor and extended modalities, which are often challenging and time-consuming to obtain. We thus propose a soft modality binding strategy, which utilizes unpaired samples across datasets with different modalities" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.085, + 0.905, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.361, + 0.907, + 0.403 + ], + "angle": 0, + "content": "Figure 3. Our proposed soft modality binding strategy. Unpaired samples from diverse datasets with different modalities are bridged with the semantic similarity between class labels as soft guidance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.409, + 0.907, + 0.49 + ], + "angle": 0, + "content": "(Fig. 3). Specifically, let \\(\\mathbf{y}^{\\mathcal{M}_1} = \\{\\pmb {y}_1^{\\mathcal{M}_1},\\dots ,\\pmb {y}_{N_1}^{\\mathcal{M}_1}\\}\\) and \\(\\pmb {\\mathcal{V}}^{\\mathcal{M}_2} = \\{\\pmb {y}_1^{\\mathcal{M}_2},\\dots ,\\pmb {y}_{N_2}^{\\mathcal{M}_2}\\}\\) represent text embeddings of class labels from two arbitrary datasets in modalities \\(\\mathcal{M}_1\\) and \\(\\mathcal{M}_2\\), respectively. The semantic similarity of textual labels across datasets can be computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.496, + 0.906, + 0.539 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} \\left(\\boldsymbol {y} _ {i} ^ {\\mathcal {M} _ {1}}, \\boldsymbol {y} _ {j} ^ {\\mathcal {M} _ {2}}\\right) = \\frac {\\exp \\left(\\boldsymbol {y} _ {i} ^ {\\mathcal {M} _ {1}} \\cdot \\boldsymbol {y} _ {j} ^ {\\mathcal {M} _ {2}}\\right)}{\\sum_ {n = 1} ^ {N _ {2}} \\exp \\left(\\boldsymbol {y} _ {i} ^ {\\mathcal {M} _ {1}}, \\boldsymbol {y} _ {n} ^ {\\mathcal {M} _ {2}}\\right)}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.544, + 0.907, + 0.687 + ], + "angle": 0, + "content": "Similarly, given samples \\((\\pmb{x}^{\\mathcal{M}_1},\\pmb{y}_i^{\\mathcal{M}_1})\\) from the first dataset and \\((\\pmb{x}^{\\mathcal{M}_2},\\pmb{y}_j^{\\mathcal{M}_2})\\) from the second dataset, the semantic similarity between \\(\\pmb{x}^{\\mathcal{M}_1}\\) and text embedding of labels from the second dataset (i.e., \\(S(\\pmb{x}^{\\mathcal{M}_1},\\pmb{y}_j^{\\mathcal{M}_2}))\\), as well as \\(\\pmb{x}^{\\mathcal{M}_2}\\) and text embeddings of labels from the first dataset (i.e., \\(S(\\pmb{x}^{\\mathcal{M}_2},\\pmb{y}_i^{\\mathcal{M}_1}))\\) could be computed. Taking the similarity between class labels of two datasets (i.e., \\(S(\\pmb{y}_i^{\\mathcal{M}_1},\\pmb{y}_j^{\\mathcal{M}_2}))\\) as a soft prior for alignment, we bind modalities across datasets by optimizing the following objective:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.691, + 0.862, + 0.732 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {s o f t \\_ b i n d}} = \\mathrm {K L} \\left(\\mathcal {S} \\left(\\boldsymbol {x} ^ {\\mathcal {M} _ {1}}, \\boldsymbol {y} _ {j} ^ {\\mathcal {M} _ {2}}\\right) \\mid \\mathcal {S} \\left(\\boldsymbol {y} _ {i} ^ {\\mathcal {M} _ {1}}, \\boldsymbol {y} _ {j} ^ {\\mathcal {M} _ {2}}\\right)\\right) \\\\ + \\operatorname {K L} \\left(\\mathcal {S} \\left(\\boldsymbol {x} ^ {\\mathcal {M} _ {2}}, \\boldsymbol {y} _ {i} ^ {\\mathcal {M} _ {1}}\\right) \\mid \\mathcal {S} \\left(\\boldsymbol {y} _ {j} ^ {\\mathcal {M} _ {2}}, \\boldsymbol {y} _ {i} ^ {\\mathcal {M} _ {1}}\\right)\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.811 + ], + "angle": 0, + "content": "where KL refers to the Kullback-Leibler divergence. Our soft modality binding strategy is different from [18], which requires instance-level textual annotations for each sample. Instead, our approach utilizes merely class labels which are instance-agnostic and challenging for alignment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Knowledge Distillation from Anchor Modalities. Given multiple well-aligned anchor modalities (e.g., \\(\\mathcal{M}_1^a\\) and \\(\\mathcal{M}_2^a\\)) and one novel modality to be extended (e.g., \\(\\mathcal{M}_k^e\\)), the alignment performance between anchor and extended modalities (e.g., \\(\\mathcal{M}_k^e\\) and \\(\\mathcal{M}_1^a\\)) is typically less effective than that between anchor modalities (e.g., \\(\\mathcal{M}_1^a\\) and \\(\\mathcal{M}_2^a\\)). Therefore," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29647" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.084, + 0.48, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.322, + 0.483, + 0.364 + ], + "angle": 0, + "content": "Figure 4. Top: Knowledge distillation from anchor modalities to instruct the alignment of novel modalities. Bottom: knowledge distillation from multiple partial-modal foundation models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.374, + 0.483, + 0.418 + ], + "angle": 0, + "content": "knowledge within the alignment between anchor modalities (Fig. 4 (Top)) could be transferred to improve the alignment learning of extended modalities via knowledge distillation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.419, + 0.484, + 0.541 + ], + "angle": 0, + "content": "For instance, given a 3D point cloud sample \\( x_{i} \\), we could obtain its tokenized input \\( \\pmb{x}_{i}^{\\mathcal{P}} \\) in the extended modalities, along with the text embedding of class label \\( \\pmb{x}_{i}^{\\mathcal{T}} \\) and the visual embedding of one rendered image from a specific viewpoint \\( \\pmb{x}_{i}^{\\mathcal{T}} \\), both in the anchor modalities. By leveraging the similarity between anchor modalities (i.e., \\( S(\\pmb{x}_{i}^{\\mathcal{T}},\\pmb{x}_{i}^{\\mathcal{T}}) \\)) as the guidance, we could optimize the alignment between the extended modality and anchor modalities through:" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.545, + 0.483, + 0.582 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {a n c h . d i s t i l}} = \\mathrm {K L} \\left(S \\left(\\boldsymbol {x} _ {i} ^ {\\mathcal {P}}, \\boldsymbol {x} _ {i} ^ {\\mathcal {T}}\\right) \\mid S \\left(\\boldsymbol {x} _ {i} ^ {\\mathcal {I}}, \\boldsymbol {x} _ {i} ^ {\\mathcal {T}}\\right)\\right) \\tag {12} \\\\ + \\operatorname {K L} \\left(\\mathcal {S} \\left(\\boldsymbol {x} _ {i} ^ {\\mathcal {P}}, \\boldsymbol {x} _ {i} ^ {\\mathcal {I}}\\right) \\mid \\mathcal {S} \\left(\\boldsymbol {x} _ {i} ^ {\\mathcal {T}}, \\boldsymbol {x} _ {i} ^ {\\mathcal {I}}\\right)\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.586, + 0.483, + 0.631 + ], + "angle": 0, + "content": "This knowledge distillation strategy is generally applicable to most modalities, where 3D point cloud, video, and depth data can be converted into images with class labels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.632, + 0.483, + 0.721 + ], + "angle": 0, + "content": "Knowledge Distillation from Partial-Modal Foundation Models. To further boost the alignment across modalities, we employ foundation models \\(\\mathcal{E}_f\\) skilled in partial modalities (e.g., audio \\(\\leftrightarrow\\) text [22], point cloud \\(\\leftrightarrow\\) image [98]) as teacher models to guide the learning of our unified modality encoder \\(\\mathcal{E}_u\\) (i.e., Fig.4 (Bottom)) from two perspectives:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.722, + 0.483, + 0.799 + ], + "angle": 0, + "content": "- Feature Alignment aims to ensure features from our unified modality encoder to be identical as those from partial-modal foundation models. Specifically, denoting \\(\\hat{F}_i^{\\mathcal{M}_k}\\) and \\(F_{i}^{\\mathcal{M}_{k}}\\) as features in modality \\(\\mathcal{M}_k\\) output by \\(\\mathcal{E}_f\\) and \\(\\mathcal{E}_u\\), we use the mean squared error (MSE) to align them as:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.803, + 0.483, + 0.821 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {f e a t - a l i g n}} = \\left\\| \\hat {\\boldsymbol {F}} ^ {\\mathcal {M} _ {i}} - \\boldsymbol {F} ^ {\\mathcal {M} _ {i}} \\right\\| _ {2} ^ {2}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "- Cross-Modality Knowledge Transfer aims to benefit the learning of \\(\\mathcal{E}_u\\) via the well-aligned representation space across modalities shaped by large-scale paired data. Specifically, given a paired sample \\((x_i^{\\mathcal{M}^a}, x_i^{\\mathcal{M}^e})\\) in anchor modality \\(\\mathcal{M}^a\\) and extended modality \\(\\mathcal{M}^e\\) respectively, the align" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "ment results of our unified modality encoder are encouraged to approximate those of partial-modal foundation models:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.126, + 0.905, + 0.164 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {m o d a l - a l i g n}} = \\mathrm {K L} \\left(\\mathcal {S} \\left(\\boldsymbol {x} _ {i} ^ {\\mathcal {M} ^ {e}}, \\boldsymbol {x} _ {i} ^ {\\mathcal {M} ^ {a}}\\right) \\mid \\mathcal {S} \\left(\\hat {\\boldsymbol {x}} _ {i} ^ {\\mathcal {M} ^ {e}}, \\hat {\\boldsymbol {x}} _ {i} ^ {\\mathcal {M} ^ {a}}\\right)\\right) \\tag {14} \\\\ + \\mathrm {K L} \\left(S \\left(\\boldsymbol {x} _ {i} ^ {\\mathcal {M} ^ {a}}, \\boldsymbol {x} _ {i} ^ {\\mathcal {M} ^ {e}}\\right) \\mid S \\left(\\hat {\\boldsymbol {x}} _ {i} ^ {\\mathcal {M} ^ {a}}, \\hat {\\boldsymbol {x}} _ {i} ^ {\\mathcal {M} ^ {e}}\\right)\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.167, + 0.88, + 0.181 + ], + "angle": 0, + "content": "where \\(\\hat{\\pmb{x}}_i\\) and \\(\\pmb{x}_i\\) are outputs of \\(\\mathcal{E}_f\\) and \\(\\mathcal{E}_u\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.182, + 0.906, + 0.212 + ], + "angle": 0, + "content": "In summary, the loss for knowledge distillation from partial-modal foundation models can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.217, + 0.905, + 0.233 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m o d e l - d i s t i l}} = \\lambda \\mathcal {L} _ {\\text {f e a t - a l i n g}} + \\mathcal {L} _ {\\text {m o d a l - a l i n g}}. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.235, + 0.887, + 0.25 + ], + "angle": 0, + "content": "Following [99], we set \\(\\lambda = 2000\\) to scale the magnitude." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.258, + 0.731, + 0.274 + ], + "angle": 0, + "content": "3.3. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.28, + 0.907, + 0.416 + ], + "angle": 0, + "content": "Network Configuration. We adopt CLIP [67] pre-trained vision encoder as the base network for unified modality encoding. The sequence length \\( L \\) and channel dimension \\( D \\) for tokenized inputs \\( X \\) are set to 196 and 768, respectively. The LoRA configuration is defined with a scaling factor \\( \\alpha \\) of 32 and a rank \\( r \\) of 4. For knowledge distillation from partial-modal foundation models, we use ULIP [98] as the teacher model for 3D point cloud inputs, VIT-LENS[38] for depth inputs, ONE-PEACE[86] for audio and video inputs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.906, + 0.552 + ], + "angle": 0, + "content": "Batching Strategy. The batching strategies proposed by [17, 42] sample inputs from each dataset (modality) separately, which may lead to catastrophic forgetting[42]. To address this, we develop a new batching strategy, which mixes data from all modalities within each mini-batch. Furthermore, a modality-aware gradient accumulation is proposed to perform forward pass, loss computation, and gradient calculation individually for each modality. Backpropagation is conducted only after processing all modalities." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.552, + 0.906, + 0.583 + ], + "angle": 0, + "content": "Training Objective. Following [68], we assign an equal weight to all training objectives:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.585, + 0.905, + 0.601 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {N C E}} + \\alpha_ {1} \\mathcal {L} _ {\\text {s o f t , b i n d}} + \\alpha_ {2} \\mathcal {L} _ {\\text {a n c h , d i s t i l}} + \\alpha_ {3} \\mathcal {L} _ {\\text {m o d e l , d i s t i l}}, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.606, + 0.906, + 0.636 + ], + "angle": 0, + "content": "where \\(\\alpha_{\\{1,2,3\\}} = 1\\) and \\(\\mathcal{L}_{\\mathrm{NCE}}\\) is the modality aligning loss for paired data, as utilized in CLIP [67]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.645, + 0.646, + 0.663 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.668, + 0.704, + 0.684 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.856 + ], + "angle": 0, + "content": "Pre-training Datasets. We perform a joint pre-training across multiple datasets with different modalities, including images from ImageNet-1K [8], 3D point clouds from ShapeNet[5], depth data from SUN RGB-D[74], audio and video from AudioSet [16] and VGGSound [6]. To address the differences in dataset size across modalities, we apply repeated augmentation[17] to small datasets (such as SUN-D) and category-balanced downsampling for large datasets (such as AudioSet) to ensure balanced performance during joint training. Once finished pre-training, we directly evaluate our model on downstream tasks without any fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Training. For image and video data, we use a resolution of \\(224 \\times 224\\) and apply standard augmentations. For single-view depth data, we employ the same augmentation as in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29648" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.089, + 0.482, + 0.257 + ], + "angle": 0, + "content": "
ModelBackboneTrainable ParamsTop-1 AccuracyTop-5 Accuracy
Single Task
Swin Transformer [47]Swin-B87.7M83.596.5
ConvNeXt [48]ConvNeXt-B88.7M83.896.7
DeiT III [80]ViT-B86.9M83.896.5
Conformer [64]Conformer-B83.3M83.896.7
Hiera [70]Hiera-B+70.0M85.297.3
Multiple Tasks
UniRepLKNet [10]UniRepLKNet-S55.6M83.9-
OMNIVORE [17]Swin-B88.2M85.397.5
MetaFormer [107]ViT-B86.6M85.497.4
UNIALIGN†ViT-B7.8M85.397.2
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.1, + 0.258, + 0.414, + 0.27 + ], + "angle": 0, + "content": "An identical model with the same weight across all tasks." + }, + { + "type": "table_footnote", + "bbox": [ + 0.096, + 0.271, + 0.48, + 0.284 + ], + "angle": 0, + "content": "Table 1. Results for image classification on ImageNet-1K (§4.2)." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.258, + 0.48, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.298, + 0.482, + 0.435 + ], + "angle": 0, + "content": "
ModelBackboneTrainable ParamsModelNet40ScanObjectNN
Top-1Top-5Top-1Top-5
Single Task
PointCLIP [106]ResNet-50-19.334.810.530.6
PointCLIPv2 [111]ViT-B-63.685.042.274.5
CLIP2Point [30]ViT-B88.3M49.581.225.559.4
RECON [66]ViT-B43.6M61.278.142.375.6
ULIP [98]ViT-B43.6M60.484.051.580.2
Multiple Tasks
VIT-LENS [38]ViT-B34.1M65.492.7--
UNIALIGN†ViT-B7.8M55.984.342.576.3
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.1, + 0.436, + 0.414, + 0.448 + ], + "angle": 0, + "content": "An identical model with the same weight across all tasks." + }, + { + "type": "table_footnote", + "bbox": [ + 0.091, + 0.448, + 0.483, + 0.476 + ], + "angle": 0, + "content": "Table 2. Results for zero-shot 3d shape classification on ModelNet40 test [95] and ScanObjectNN test [82] (\\$4.3)." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.436, + 0.483, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.658 + ], + "angle": 0, + "content": "[38]. The augmentation of audio data also follows [38] except Mix-up [94]. We train the model with a batch size of 2048 (after gradient accumulation) for 100 epochs. An AdamW optimizer with a initial learning rate \\(2\\mathrm{e}^{-4}\\) and a weight decay of 0.05 is employed. For knowledge distillation from partial-modal foundation models, we cache offline features to avoid repeatedly forwarding. The base modality encoder remains frozen during training, with only LoRA enhanced MoE modules and the projection head for knowledge distillation remaining trainable. Flash-attention is applied to reduce GPU memory cost." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.658, + 0.483, + 0.688 + ], + "angle": 0, + "content": "Reproducibility. UNIALIGN is implemented in PyTorch and trained on four NVIDIA Tesla A100 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.371, + 0.714 + ], + "angle": 0, + "content": "4.2. Results for Image Classification" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.765 + ], + "angle": 0, + "content": "Dataset. We evaluate UNIALIGN for image classification on ImageNet-1K val [8] which contains 50,000 samples across 1,000 classes, spanning objects, animals, and more." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.795 + ], + "angle": 0, + "content": "Metric. Following previous works [47, 48, 87], we report the count of trainable parameters, top-1 and top-5 accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Performance. As shown in Table 1, UNIALIGN demonstrates impressive performance on image classification, achieving a comparable performance to modality-specific approaches (i.e., \\(85.3\\%\\) vs. \\(85.2\\%\\) of Hiera[70]). It is noteworthy that our method utilizes merely \\(9\\%\\) trainable parameters and a unified modality encoder architecture which preserves the capability to process other modalities." + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.09, + 0.905, + 0.24 + ], + "angle": 0, + "content": "
ModelBackboneTrainable ParamsAnchor ModalityNYU-DSUN-D
Single Task
G-L-SOOR [76]D-CNN [75]--62.344.1
RecgNet [12]ResNet18 [24]11.2M-57.747.9
Depth Swin [18]Swin-B88.2M-76.469.1
Multiple Tasks
Text Paired [18]ViT-H88.1MT41.925.4
ImageBind [18]ViT-H88.3MI54.035.1
ViT-LENS [38]ViT-L50.9MI+T68.552.2
UNIALIGN†ViT-B7.8MT46.334.3
UNIALIGN†ViT-B7.8MI+T59.445.8
UNIALIGN†ViT-B7.8MI+T71.4*58.2*
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.522, + 0.241, + 0.835, + 0.253 + ], + "angle": 0, + "content": "An identical model with the same weight across all tasks." + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.253, + 0.905, + 0.295 + ], + "angle": 0, + "content": "Table 3. Results for depth-only scene classification on NYU-D test [73] and SUN-D test [74] (\\$4.4).*: The depth maps are tokenized by the image tokenizer additionally." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.241, + 0.905, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.308, + 0.9, + 0.324 + ], + "angle": 0, + "content": "4.3. Results for Zero-Shot 3D Shape Classification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.331, + 0.905, + 0.391 + ], + "angle": 0, + "content": "Dataset. We evaluate zero-shot 3D shape classification on two datasets. ModelNet40 test [95] contains 2,468 samples from 40 categories, while ScanObjectNN test [82] consists of 581 samples distributed across 15 categories." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.392, + 0.854, + 0.407 + ], + "angle": 0, + "content": "Metric. The top-1 and top-5 accuracy are reported." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.408, + 0.906, + 0.498 + ], + "angle": 0, + "content": "Performance. As shown in Table 2, UNIALIGN achieves a top-1 accuracy of \\(55.9\\%\\) and \\(42.5\\%\\) on ModelNet40 val, and ScanObjectNN test, respectively. Though SOTA approaches such as ULIP [98] and ViT-LENS [38] yield higher scores, they require a modality-specific point-cloud encoder and an additional image encoder for alignment." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.51, + 0.879, + 0.526 + ], + "angle": 0, + "content": "4.4. Results for Depth-only Scene Classification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.906, + 0.593 + ], + "angle": 0, + "content": "Dataset. Experiments for depth-only scene classification are conducted on SUN RGB-D test [74] which contains 4,660 samples across 19 classes, and NYU-D test [73] which contains 653 samples across 10 semantic classes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.594, + 0.869, + 0.608 + ], + "angle": 0, + "content": "Metric. The top-1 accuracy is adopted for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.906, + 0.791 + ], + "angle": 0, + "content": "Performance. A detailed comparison of UNIALIGN against several top-leading approaches for depth-only scene classification is provided in Table 3. Building on a ViT-B backbone and aligning extended modalities with both image and text as anchor modalities, UNIALIGN achieves a top-1 accuracy of \\(71.4\\%\\) on NYU-D test and \\(58.2\\%\\) on SUN RGB-D test, surpassing all competitors. Notably, aligning merely to the text modality leads to a marked performance decline (i.e., \\(71.4\\% \\rightarrow 46.3\\%\\) on NYU-D). This suggests that i) the depth modality exhibits a closer alignment with image data, and ii) UNIALIGN can effectively leverage the additional anchor modality to enhance alignment performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.899, + 0.818 + ], + "angle": 0, + "content": "4.5. Results for Audio Classification and Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Datasets. We evaluate audio classification on AudioSet val [16] containing 2M samples across 527 classes, and ESC [65] including 2K data across 50 classes. Audio-text retrieval is evaluated on AudioCaps test [33] which includes 813 audio-text pairs. Moreover, following [107], we" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29649" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.089, + 0.482, + 0.419 + ], + "angle": 0, + "content": "
MethodBackboneAnchor ModalityAudioSet mAPESC Top1AudioCaps R@1R@10
Single Task
AVFIC [57]ViT-B---8.737.7
AudioCLIP [22]ViT-BI+T25.969.4--
Multiple Tasks
ImageBind-H [18]ViT-HI17.666.99.342.3
VIT-LENS [38]ViT-LI+T27.280.914.955.2
LanguageBind [109]ViT-LI+T27.791.812.253.2
UNIALIGN†ViT-BI+T23.870.511.749.3
†:An identical model with the same weight across all tasks. \nTable 4. Results for audio classification on AudioSet val [16], ESC[65], and audio-text retrieval on AudioCaps test[33] (§4.5). Due to expiration of links, UNIALIGN is trained with only 0.5M ideas from AudioSet, while methods such as ViT-Lens use 1.6M.
MethodBackbonePre-train DatasetsTrainable ParamsTop-1
Single Task
AST [19]ViT-BN/A86.9M92.6
AST [19]ViT-BAudioSet-20K86.9M96.2
SSAST [20]ViT-BAudioSet-2M89.3M97.8
Multiple Tasks
MetaFormer [107]ViT-BLAION2B1.1M78.3
MetaFormer [107]ViT-BLAION2B86.3M97.0
UniRepLKNet [10]UniRepLKNet-S-55.5M98.5
UNIALIGNViT-BJoint1.8M97.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.425, + 0.483, + 0.455 + ], + "angle": 0, + "content": "Table 5. Results for audio classification on Speech Commands V2 val [92]. All models are fine-tuned on the training set (§4.5)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.467, + 0.483, + 0.497 + ], + "angle": 0, + "content": "fine-tune UNIALIGN on Speech Commands V2 train[92] for speech recognition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.499, + 0.483, + 0.545 + ], + "angle": 0, + "content": "Metric. Following [18, 22, 38], we measure mAP for AudioSet, top-1 accuracy for ESC and Speech Commands V2, and Recall@ \\(k\\) (\\(k = \\{1,5\\}\\)) for AudioCaps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.546, + 0.483, + 0.636 + ], + "angle": 0, + "content": "Performance for Audio classification. UNIALIGN achieves promising results on AudioSet and ESC (Table 4). Though inferior to top-leading methods, it is important to note that, due to the expiration of links, a significant number of videos from AudioSet can not be downloaded from YouTube (i.e., 1.6M used in ViT-Lens[38] vs. 0.5M used in UNIALIGN)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.637, + 0.483, + 0.683 + ], + "angle": 0, + "content": "Performance for Audio retrieval. Similarly, despite the absence of 1.1M training data, UNIALIGN can still achieve comparable performance, achieving 11.7 R@1 accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.684, + 0.483, + 0.805 + ], + "angle": 0, + "content": "Performance for Speech Recognition. As shown in Table 5, UNIALIGN demonstrates remarkable accuracy compared to existing work (e.g., \\(97.4\\%\\) vs. \\(97.0\\%\\) for MetaTransformer[107]), while utilizing significantly fewer trainable parameters (e.g., 1.8M vs. 86.3M). When tuning a similar number of parameters (i.e., 1.1M for Meta-Transformer, and 1.8M for UNIALIGN), our method outperforms Meta-Transformer by \\(19.1\\%\\) (i.e., \\(97.4\\%\\) vs. \\(78.3\\%\\))." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.476, + 0.833 + ], + "angle": 0, + "content": "4.6. Results for Video Classification and Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Datasets. We utilize MSR-VTT val [96] which contains around 1,000 video-text pairs for zero-shot video-text retrieval, and UCF101 val [77] which contains 3,738 samples across 101 semantic classes for video classification." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.09, + 0.905, + 0.224 + ], + "angle": 0, + "content": "
ModelBackboneRetrieval ModalityTrainable ParamsR@1R@5
Single Task
MIL-NCE [54]ResNeXt-101V88.8M8.625.8
SupportSet [60]R3D-34V93.5M10.430.0
AVFIC [57]ViT-BA+V86.6M19.450.3
InternVideo [89]ViT-LV305M37.571.3
Multiple Tasks
ImageBind [18]ViT-HA+V89.3M36.870.0
ViT-LENS [38]ViT-LA+V127.6M37.672.6
UNIALIGN†ViT-BA+V7.8M37.070.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.521, + 0.224, + 0.903, + 0.251 + ], + "angle": 0, + "content": "†:An identical model with the same weight across all tasks. Table 6. Results for video retrieval on MSRVTT val [96] (§4.6)." + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.266, + 0.905, + 0.384 + ], + "angle": 0, + "content": "
ModelBackboneZero-ShotFine-Tuned
Single Task
ResT [44]ResNet-10158.7-
ActionCLIP [85]ViT-B58.3-
X-CLIP [58]ViT-B72.0-
Multiple Tasks
MetaFormer [107]ViT-B-46.6
ImageBind [18]ViT-H64.898.1
UniBind [50]ViT-H73.793.3
UNIALIGN†ViT-B68.594.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.516, + 0.385, + 0.91, + 0.411 + ], + "angle": 0, + "content": "†:An identical model with the same weight across all tasks. Table 7. Results for video classification on UCF101 val[77] (§4.6)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.424, + 0.907, + 0.453 + ], + "angle": 0, + "content": "Metric. Recall@ \\(k\\) and top-1 accuracy are adopted as evaluation metrics for video-text retrieval and classification." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.455, + 0.907, + 0.515 + ], + "angle": 0, + "content": "Performance for Video-Text Retrieval. As shown in Table 6, UNIALIGN achieves a R@1 score of \\(37.0\\%\\) and R@5 score of \\(70.7\\%\\), which are comparable to advanced methods while using merely 7.8M trainable parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.907, + 0.652 + ], + "angle": 0, + "content": "Performance for Video Classification. We summarize the results for video classification in Table 7. As shown, UNIALIGN achieves an impressive accuracy of \\(68.5\\%\\) on UCF101 val[77] without any fine-tuning, and outperforms ImageBind [18]. After fine-tuning all models on the training set of UCF101, our method achieves the second-highest performance. This may be attributed to the differences in backbone, as ImageBind utilizes ViT-H which could provide enhanced knowledge for downstream fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.663, + 0.696, + 0.679 + ], + "angle": 0, + "content": "4.7. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.685, + 0.907, + 0.731 + ], + "angle": 0, + "content": "As shown in Fig. 5, UNIALIGN facilitates accurate cross-modality retrieval. Moreover, modalities can be combined to constrain attributes of the target in another modality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.741, + 0.727, + 0.758 + ], + "angle": 0, + "content": "4.8. Diagnostic Experiment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.907, + 0.809 + ], + "angle": 0, + "content": "To evaluate the core designs of UNIALIGN and gain further insights, we conduct a series of ablation studies. Results are reported on SUN-D test [74] and ModelNet40 test [5]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Modality-Aware MoE Strategy. As illustrated in §3.1, we explore two modality-aware MoE strategies: expert grouping and global routing. Results in Table 8 indicate that dividing experts into modality-specific groups yields better performance. Thus, we adopt it as the default MoE strategy. Expert Network. Table 9 presents the ablation study on" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29650" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.131, + 0.093, + 0.15, + 0.13 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.167, + 0.094, + 0.231, + 0.132 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.094, + 0.302, + 0.131 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.094, + 0.37, + 0.131 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.428, + 0.098, + 0.462, + 0.117 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.122, + 0.474, + 0.131 + ], + "angle": 0, + "content": "Aircraft engine" + }, + { + "type": "image", + "bbox": [ + 0.485, + 0.093, + 0.549, + 0.13 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.095, + 0.648, + 0.109 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.579, + 0.121, + 0.648, + 0.131 + ], + "angle": 0, + "content": "Playing volleyball" + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.093, + 0.728, + 0.13 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.77, + 0.098, + 0.805, + 0.116 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.123, + 0.796, + 0.132 + ], + "angle": 0, + "content": "Dog" + }, + { + "type": "image", + "bbox": [ + 0.816, + 0.093, + 0.882, + 0.131 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.141, + 0.168, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.122, + 0.176, + 0.159, + 0.182 + ], + "angle": 0, + "content": "Bathroom" + }, + { + "type": "image", + "bbox": [ + 0.174, + 0.142, + 0.221, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.143, + 0.295, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.256, + 0.172, + 0.293, + 0.181 + ], + "angle": 0, + "content": "Car horn" + }, + { + "type": "image", + "bbox": [ + 0.308, + 0.143, + 0.369, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.428, + 0.142, + 0.459, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.423, + 0.171, + 0.478, + 0.18 + ], + "angle": 0, + "content": "Playing piano" + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.146, + 0.552, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.143, + 0.64, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.585, + 0.165, + 0.661, + 0.183 + ], + "angle": 0, + "content": "Typing on computer keyboard" + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.142, + 0.731, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.67, + 0.173, + 0.739, + 0.182 + ], + "angle": 0, + "content": "Computer room" + }, + { + "type": "image", + "bbox": [ + 0.769, + 0.146, + 0.806, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.766, + 0.164, + 0.815, + 0.182 + ], + "angle": 0, + "content": "Dishes, pots, and pans" + }, + { + "type": "image", + "bbox": [ + 0.827, + 0.142, + 0.882, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.838, + 0.174, + 0.87, + 0.181 + ], + "angle": 0, + "content": "Kitchen" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.192, + 0.172, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.173, + 0.197, + 0.262, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.193, + 0.332, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.402, + 0.194, + 0.451, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.462, + 0.198, + 0.531, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.193, + 0.605, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.193, + 0.72, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.734, + 0.192, + 0.825, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.828, + 0.192, + 0.882, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.84, + 0.221, + 0.87, + 0.229 + ], + "angle": 0, + "content": "Office" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.235, + 0.905, + 0.264 + ], + "angle": 0, + "content": "Figure 5. Qualitative results for cross-modal retrieval. Top and middle: retrieval using one single modality. Bottom: multiple modalities can be composed to retrieve targets across different modalities (§4.7)." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.277, + 0.482, + 0.322 + ], + "angle": 0, + "content": "
#StrategiesSUN-DModelNet40
1Expert Grouping45.855.9
2Global Routing45.055.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.106, + 0.325, + 0.467, + 0.34 + ], + "angle": 0, + "content": "Table 8. Ablation studies on modality-aware MoE strategies." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.351, + 0.481, + 0.448 + ], + "angle": 0, + "content": "
Top-k#ExpertsInit.GPU MemSUN-DModelNet40
11rand27.2G45.054.8
11CLIP27.2G45.255.1
12rand27.4G45.155.2
12CLIP27.4G45.255.4
22rand35.9G45.755.7
22CLIP35.9G45.855.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.452, + 0.483, + 0.48 + ], + "angle": 0, + "content": "Table 9. Ablation studies on configuration of expert network. GPU memory is reported with a batch size of 448 on one Tesla A100." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.483, + 0.856 + ], + "angle": 0, + "content": "expert network configurations. We investigate the impact of top-\\(k\\) routing, the number of experts per modality, and weight initialization. Starting with a baseline configuration of one expert per modality and random initialization, we achieve 45.0 on SUN-D test [74] and 54.8 on ModelNet40 test [5], with 27.2G GPU memory consumption. Increasing the expert number and initializing weights from the visual encoder of CLIP both lead to improved performance. Due to the constraints of GPU memory, we use two experts per modality and top-2 routing for expert selection. Training Objectives. Next we probe the impact of training objectives i.e., soft modality binding with \\(\\mathcal{L}_{\\mathrm{soft\\_bind}}\\) (cf., Eq. 11), as well as knowledge distillation through \\(\\mathcal{L}_{\\mathrm{anch\\_distil}}\\) (cf., Eq. 12) and \\(\\mathcal{L}_{\\mathrm{model\\_distil}}\\) (cf., Eq. 15). Results are summarized in Table 10, where the first row refers to the baseline that utilizes only paired data from single datasets and is trained with the standard InfoNCE loss. It can be observed that: First, our \\(\\mathcal{L}_{\\mathrm{soft\\_bind}}\\) successfully bridges the learning of multiple modalities for mutual boosting, leads to improved performance on both datasets. Second, \\(\\mathcal{L}_{\\mathrm{anch\\_distil}}\\) and \\(\\mathcal{L}_{\\mathrm{model\\_distil}}\\) could further boost the performance with compelling gains. This verifies the effectiveness of our knowledge distillation strategies, and leads UNIALIGN to a high performance model with optimized efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Modality Scaling. Finally, we examine the robustness of UNIALIGN to scale different numbers of modalities into a unified model. As shown in Table 11, modality-independent" + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.278, + 0.905, + 0.349 + ], + "angle": 0, + "content": "
L_NCELsoft_bindLanch_distilLmodel_distilSUN-DModelNet40
✓×××40.438.6
✓✓××41.840.1
✓✓✓×42.340.9
✓✓✓✓45.855.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.563, + 0.353, + 0.855, + 0.367 + ], + "angle": 0, + "content": "Table 10. Ablation studies on training objectives." + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.379, + 0.905, + 0.459 + ], + "angle": 0, + "content": "
#Pre-Training ModalityImageNet-1KSUN-DModelNet40
1I + D + P84.734.431.6
2I + T + D + P84.939.938.1
3I + T + D + P + V85.140.338.4
4I + T + D + P + V + A85.140.438.6
5I + T + D + P + V + A *85.345.855.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.464, + 0.906, + 0.52 + ], + "angle": 0, + "content": "Table 11. Ablation studies on modality scaling. Results are reported using only \\(\\mathcal{L}_{\\mathrm{NCE}}\\) as the default training loss. * denotes incorporate \\(\\mathcal{L}_{\\mathrm{soft\\_bind}}\\), \\(\\mathcal{L}_{\\mathrm{anch\\_distil}}\\), and \\(\\mathcal{L}_{\\mathrm{model\\_distil}}\\) into training. I: Image, T: Text, D: Depth, V: Video, A: Audio, P: 3D Point Cloud." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.658 + ], + "angle": 0, + "content": "training with merely \\(\\mathcal{L}_{\\mathrm{NCE}}\\) leads to a performance degeneration as more modalities integrated (i.e., row #1-4). This suggests that a straightforward unification of multiple modalities may yield sub-optimal performance for individual ones. Nevertheless, soft modality binding and knowledge distillation (i.e., row #4 vs. #5) could mitigate this negative impact, and boost performance to a level surpassing that achieved with fewer modalities (i.e., row #1 vs. #5)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.678, + 0.634, + 0.693 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we focus on the all-in-one paradigm for multimodal alignment, which yields a highly efficient and effective model, UNIALIGN, to unify the alignment of all anticipated modalities within one network. Inspired by MoE and knowledge distillation, we propose a modality-aware MoE to tackle multimodal inputs, and generally applicable alignment strategies to utilize unpaired data across datasets. Empirical results suggest that UNIALIGN achieves remarkable performance while being efficient in both training (7.8M trainable parameters to handle 6 modalities within a single training phase) and inference (one identical model with consistent weights across tasks). This work may potentially advance the development of compact, multimodal models." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29651" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.093, + 0.486, + 0.219 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by the National Science and Technology Major Project (No. 2023ZD0121300), the National Natural Science Foundation of China (No. 62472222, 62302217), Natural Science Foundation of Jiangsu Province (No. BK20240080, BK20220936), \"Pioneer\" and \"Leading Goose\" R&D Program of Zhejiang (No. 2024C01161), the National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, Xi'an Jiaotong University (No. HMHAI-202403), and Bytedance Doubao Fund." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.233, + 0.188, + 0.249 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.257, + 0.482, + 0.299 + ], + "angle": 0, + "content": "[1] Romero Adriana, Ballas Nicolas, K Samira Ebrahimi, Chassang Antoine, Gatta Carlo, and Bengio Yoshua. Fitnets: Hints for thin deep nets. In ICLR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.301, + 0.482, + 0.356 + ], + "angle": 0, + "content": "[2] Lucas Beyer, Xiaohua Zhai, Amélie Royer, Larisa Markeeva, Rohan Anil, and Alexander Kolesnikov. Knowledge distillation: A good teacher is patient and consistent. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.359, + 0.482, + 0.4 + ], + "angle": 0, + "content": "[3] Xinhao Cai, Qiqia Lai, Yuwei Wang, Wenguan Wang, Zeren Sun, and Yazhou Yao. Poly kernel inception network for remote sensing detection. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.402, + 0.482, + 0.428 + ], + "angle": 0, + "content": "[4] Rich Caruana. Multitask learning. Machine learning, 28: 41-75, 1997. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.431, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.502, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[6] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.545, + 0.482, + 0.599 + ], + "angle": 0, + "content": "[7] Shaoxiang Chen, Zequn Jie, and Lin Ma. Llava-MoLE: Sparse Mixture of LoRA Experts for Mitigating Data Conflicts in Instruction Finetuning MLLMs. arXiv preprint arXiv:2401.16160, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.602, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.645, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[9] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. In NeurIPS, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.688, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[10] Xiaohan Ding, Yiyuan Zhang, Yixiao Ge, Sijie Zhao, Lin Song, Xiangyu Yue, and Ying Shan. Unireplknet: A universal perception large-kernel convnet for audio video point cloud time-series and image recognition. In CVPR, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.759, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[11] Shihan Dou, Enyu Zhou, Yan Liu, Songyang Gao, Jun Zhao, Wei Shen, Yuhao Zhou, Zhiheng Xi, Xiao Wang, Xiaoran Fan, et al. Loramoe: Revolutionizing mixture of experts for maintaining world knowledge in language model alignment. arXiv preprint arXiv:2312.09979, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.83, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[12] Dapeng Du, Limin Wang, Huiling Wang, Kai Zhao, and Gangshan Wu. Translate-to-recognize networks for rgb-d scene recognition. In CVPR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[13] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue" + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.257, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Cao. Eva: Exploring the limits of masked visual representation learning at scale. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.123, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[14] Bilal Faye, Hanane Azzag, and Mustapha Lebbah. Oneencoder: A Lightweight Framework for Progressive Alignment of Modalities. arXiv preprint arXiv:2409.11059, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.181, + 0.905, + 0.235 + ], + "angle": 0, + "content": "[15] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1-39, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.238, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[16] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In ICASSP, 2017. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.309, + 0.905, + 0.364 + ], + "angle": 0, + "content": "[17] Rohit Girdhar, Mannat Singh, Nikhila Ravi, Laurens van der Maaten, Armand Joulin, and Ishan Misra. Omnivore: A Single Model for Many Visual Modalities. In CVPR, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.367, + 0.905, + 0.422 + ], + "angle": 0, + "content": "[18] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In CVPR, 2023. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.424, + 0.905, + 0.452 + ], + "angle": 0, + "content": "[19] Yuan Gong, Yu-An Chung, and James Glass. AST: Audio Spectrogram Transformer. In Interspeech, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.454, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[20] Yuan Gong, Cheng-I Lai, Yu-An Chung, and James R. Glass. Ssast: Self-Supervised Audio Spectrogram Transformer. In AAAI, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.498, + 0.905, + 0.567 + ], + "angle": 0, + "content": "[21] Yunhao Gou, Zhili Liu, Kai Chen, Lanqing Hong, Hang Xu, Aoxue Li, Dit-Yan Yeung, James T Kwok, and Yu Zhang. Mixture of cluster-conditional lora experts for vision-language instruction tuning. arXiv preprint arXiv:2312.12379, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.569, + 0.905, + 0.61 + ], + "angle": 0, + "content": "[22] Andrey Guzhov, Federico Raue, Jorn Hees, and Andreas Dengel. Audioclip: Extending clip to image, text and audio. In ICASSP, 2022. 1, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.613, + 0.905, + 0.669 + ], + "angle": 0, + "content": "[23] Jiaming Han, Kaixiong Gong, Yiyuan Zhang, Jiaqi Wang, Kaipeng Zhang, Dahua Lin, Yu Qiao, Peng Gao, and Xi-angyu Yue. Onellm: One Framework to Align All Modalities with Language. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.671, + 0.905, + 0.711 + ], + "angle": 0, + "content": "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.714, + 0.905, + 0.755 + ], + "angle": 0, + "content": "[25] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, Nojun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.758, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[26] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.788, + 0.905, + 0.842 + ], + "angle": 0, + "content": "[27] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[28] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "29652" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[29] Chengsong Huang, Qian Liu, Bill Yuchen Lin, Tianyu Pang, Chao Du, and Min Lin. Lorahub: Efficient cross-task generalization via dynamic lora composition. arXiv preprint arXiv:2307.13269, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.149, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[30] Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson W.H. Lau, Wanli Ouyang, and Wangmeng Zuo. Clip2point: Transfer clip to point cloud classification with image-depth pre-training. In ICCV, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.205, + 0.483, + 0.272 + ], + "angle": 0, + "content": "[31] Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. Mixtral of experts. arXiv preprint arXiv:2401.04088, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.274, + 0.482, + 0.327 + ], + "angle": 0, + "content": "[32] Byoungjip Kim, Sungik Choi, Dasol Hwang, Moontae Lee, and Honglak Lee. Transferring pre-trained multimodal representations with cross-modal similarity matching. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[33] Chris Dongjoo Kim, Byeongchang Kim, Hyunmin Lee, and Gunhee Kim. Audiocaps: Generating captions for audios in the wild. In NAACL-HLT, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.373, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[34] Jangho Kim, SeongUk Park, and Nojun Kwak. Paraphrasing complex network: Network compression via factor transfer. In NeurIPS, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.414, + 0.482, + 0.481 + ], + "angle": 0, + "content": "[35] Aran Komatsuzaki, Joan Puigcerver, James Lee-Thorp, Carlos Riquelme Ruiz, Basil Mustafa, Joshua Ainslie, Yi Tay, Mostafa Dehghani, and Neil Houlsby. Sparse upcycling: Training mixture-of-experts from dense checkpoints. In ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.484, + 0.482, + 0.536 + ], + "angle": 0, + "content": "[36] Nikos Komodakis and Sergey Zagoruyko. Paying more attention to attention: improving the performance of convolutional neural networks via attention transfer. In ICLR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.539, + 0.482, + 0.592 + ], + "angle": 0, + "content": "[37] Konrad P Kording, Ulrik Beierholm, Wei Ji Ma, Steven Quartz, Joshua B Tenenbaum, and Ladan Shams. Causal inference in multisensory perception. PLoS one, 2(9):e943, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.595, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[38] Stan Weixian Lei, Yixiao Ge, Kun Yi, Jianfeng Zhang, Difei Gao, Dylan Sun, Yuying Ge, Ying Shan, and Mike Zheng Shou. Vit-Lens: Towards Omni-modal Representations. In CVPR, 2024. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.651, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[39] Weixian Lei, Yixiao Ge, Kun Yi, Jianfeng Zhang, Difei Gao, Dylan Sun, Yuying Ge, Ying Shan, and Mike Zheng Shou. Vit-lens: Towards omni-modal representations. In CVPR, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.707, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[40] Po-han Li, Sandeep P Chinchali, and Ufuk Topcu. Csa: Data-efficient mapping of unimodal features to multimodal features. arXiv preprint arXiv:2410.07610, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.749, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[41] Yunxin Li, Shenyuan Jiang, Baotian Hu, Longyue Wang, Wanqi Zhong, Wenhan Luo, Lin Ma, and Min Zhang. UniMoE: Scaling Unified Multimodal LLMs with Mixture of Experts. arXiv, abs/2405.11273, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.804, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[42] Valerii Likhosherstov, Anurag Arnab, Krzysztof Choromanski, Mario Lucic, Yi Tay, Adrian Weller, and Mostafa Dehghani. Polyvit: Co-training vision transformers on images, videos and audio. arXiv preprint arXiv:2111.12993, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[43] Bin Lin, Zhenyu Tang, Yang Ye, Jiaxi Cui, Bin Zhu, Peng Jin, Jina Huang, Junwu Zhang, Yatian Pang, Munan Ning," + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "et al. Moe-llava: Mixture of experts for large vision-language models. arXiv preprint arXiv:2401.15947, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.136, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[44] Chung-Ching Lin, Kevin Lin, Lijuan Wang, Zicheng Liu, and Linjie Li. Cross-modal representation learning for zero-shot action recognition. In CVPR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.179, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[45] Haokun Lin, Haoli Bai, Zhili Liu, Lu Hou, Muyi Sun, Linqi Song, Ying Wei, and Zhenan Sun. Mope-CLIP: Structured Pruning for Efficient Vision-Language Models with Module-wise Pruning Error Metric. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.236, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[46] Xi Victoria Lin, Akshit Shrivastava, Liang Luo, Srinivasan Iyer, Mike Lewis, Gargi Gosh, Luke Zettlemoyer, and Armen Aghajanyan. Moma: Efficient early-fusion pretraining with mixture of modality-aware experts. arXiv preprint arXiv:2407.21770, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.305, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[47] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.362, + 0.904, + 0.403 + ], + "angle": 0, + "content": "[48] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.406, + 0.904, + 0.46 + ], + "angle": 0, + "content": "[49] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.462, + 0.904, + 0.503 + ], + "angle": 0, + "content": "[50] Yuanhuiyi Lyu, Xu Zheng, Jiazhou Zhou, and Lin Wang. Unibind: Llm-Augmented Unified and Balanced Representation Space to Bind Them All. In CVPR, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.505, + 0.904, + 0.546 + ], + "angle": 0, + "content": "[51] Jian Ma, Wenguan Wang, Yi Yang, and Feng Zheng. Ms2sl: Multimodal spoken data-driven continuous sign language production. In ACL Findings, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.548, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[52] Wenxuan Ma, Shuang Li, Jinming Zhang, Chi Harold Liu, Jingxuan Kang, Yulin Wang, and Gao Huang. Borrowing Knowledge From Pre-trained Language Model: A New Data-efficient Visual Learning Paradigm. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.605, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[53] M Alex Meredith and Barry E Stein. Visual, auditory, and somatosensory convergence on cells in superior colliculus results in multisensory integration. Journal of neurophysiology, 56(3):640-662, 1986. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.662, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[54] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.732, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[55] Seyed Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved knowledge distillation via teacher assistant. In AAAI, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.789, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[56] Luca Moschella, Valentino Maiorca, Marco Fumero, Antonio Norelli, Francesco Locatello, and Emanuele Rodola. Relative representations enable zero-shot latent space communication. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[57] Arsha Nagrani, Paul Hongsuck Seo, Bryan Seybold, Anja Hauth, Santiago Manen, Chen Sun, and Cordelia Schmid. Learning audio-video modalities from image captions. In ECCV, 2022. 7" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29653" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[58] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. In European Conference on Computer Vision, pages 1-18. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.162, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[59] Antonio Norelli, Marco Fumero, Valentino Maiorca, Luca Moschella, Emanuele Rodola, and Francesco Locatello. Asif: Coupled data turns unimodal models to multimodal without training. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.218, + 0.483, + 0.273 + ], + "angle": 0, + "content": "[60] Mandela Patrick, Po-Yao Huang, Yuki Asano, Florian Metze, Alexander Hauptmann, Joao Henriques, and Andrea Vedaldi. Support-set bottlenecks for video-text representation learning. arXiv preprint arXiv:2010.02824, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.274, + 0.482, + 0.327 + ], + "angle": 0, + "content": "[61] Gensheng Pei, Fumin Shen, Yazhou Yao, Guo-Sen Xie, Zhenmin Tang, and Jinhui Tang. Hierarchical feature alignment network for unsupervised video object segmentation. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[62] Gensheng Pei, Tao Chen, Xiruo Jiang, Huafeng Liu, Zeren Sun, and Yazhou Yao. Videomac: Video masked autoencoders meet convnets. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.372, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[63] Renjing Pei, Jianzhuang Liu, Weimian Li, Bin Shao, Songcen Xu, Peng Dai, Juwei Lu, and Youliang Yan. Clipping: Distilling clip-based models with a student base for video-language retrieval. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.428, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[64] Zhiliang Peng, Wei Huang, Shanzhi Gu, Lingxi Xie, Yaowei Wang, Jianbin Jiao, and Qixiang Ye. Conformer: Local features coupling global representations for visual recognition. In ICCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.484, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[65] Karol J Piczak. Esc: Dataset for environmental sound classification. In ACM MM, 2015. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.511, + 0.482, + 0.566 + ], + "angle": 0, + "content": "[66] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with Reconstruct: Contrastive 3d Representation Learning Guided by Generative Pretraining. In ICML, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.567, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[67] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.637, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[68] Mike Ranzinger, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. Am-RADIO: Agglomerative Vision Foundation Model Reduce All Domains Into One. In CVPR, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.692, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[69] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. In NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.748, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[70] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. In ICML, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[71] Jinghuan Shang, Karl Schmeckpeper, Brandon B. May, Maria Vittoria Minniti, Tarik Kelestemur, David Watkins, and Laura Herlant. Theia: Distilling Diverse Vision Foundation Models for Robot Learning. In CoRL, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[72] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean." + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In ICLR, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.122, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[73] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.166, + 0.904, + 0.208 + ], + "angle": 0, + "content": "[74] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In CVPR, 2015. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.21, + 0.905, + 0.252 + ], + "angle": 0, + "content": "[75] Xinhang Song, Luis Herranz, and Shuqiang Jiang. Depth cnns for rgb-d scene recognition: Learning from scratch better than transferring from rgb-cnns. In AAAI, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.254, + 0.905, + 0.308 + ], + "angle": 0, + "content": "[76] Xinhang Song, Shuqiang Jiang, Bohan Wang, Chengpeng Chen, and Gongwei Chen. Image representations with spatial object-to-object relations for rgb-d scene recognition. IEEE TIP, 29:525-537, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.311, + 0.905, + 0.352 + ], + "angle": 0, + "content": "[77] K Soomro. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.354, + 0.905, + 0.397 + ], + "angle": 0, + "content": "[78] Siddharth Srivastava and Gaurav Sharma. Omnivec2 - A Novel Transformer based Network for Large Scale Multimodal and Multitask Learning. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.399, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[79] Ximeng Sun, Pengchuan Zhang, Peizhao Zhang, Hardik Shah, Kate Saenko, and Xide Xia. Dime-fm: Distilling multimodal and efficient foundation models. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.456, + 0.904, + 0.484 + ], + "angle": 0, + "content": "[80] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. In ECCV, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.486, + 0.905, + 0.555 + ], + "angle": 0, + "content": "[81] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.557, + 0.905, + 0.613 + ], + "angle": 0, + "content": "[82] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.615, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[83] Simon Vandenhende, Stamatios Georgoulis, Wouter Van Gansbeke, Marc Proesmans, Dengxin Dai, and Luc Van Gool. Multi-task learning for dense prediction tasks: A survey. IEEE TPAMI, 44(7):3614-3633, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.673, + 0.905, + 0.755 + ], + "angle": 0, + "content": "[84] Haoxiang Wang, Pavan Kumar Anasosalu Vasu, Fartash Faghri, Raviteja Vemulapalli, Mehrdad Farajtabar, Sachin Mehta, Mohammad Rastegari, Oncel Tuzel, and Hadi Pouransari. Sam-clip: Merging vision foundation models towards semantic and spatial understanding. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.758, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[85] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.801, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[86] Peng Wang, Shijie Wang, Junyang Lin, Shuai Bai, Xiaohuan Zhou, Jingren Zhou, Xinggang Wang, and Chang Zhou. One-peace: Exploring one general representation model toward unlimited modalities. arXiv preprint arXiv:2305.11172, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.872, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[87] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu," + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29654" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. In CVPR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.134, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[88] Wenguan Wang, Yi Yang, and Yunhe Pan. Visual knowledge in the big model era: Retrospect and prospect. Frontiers of Information Technology & Electronic Engineering, 26(1):1-19, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.19, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[89] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.26, + 0.483, + 0.315 + ], + "angle": 0, + "content": "[90] Zehan Wang, Ziang Zhang, Luping Liu, Yang Zhao, Haifeng Huang, Tao Jin, and Zhou Zhao. Extending multi-modal contrastive representations. arXiv preprint arXiv:2310.08884, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.316, + 0.483, + 0.37 + ], + "angle": 0, + "content": "[91] Zehan Wang, Yang Zhao, Haifeng Huang, Jiageng Liu, Aoxiong Yin, Li Tang, Linjun Li, Yongqi Wang, Zhang, and Zhou Zhao. Connecting multi-modal contrastive representations. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.371, + 0.483, + 0.412 + ], + "angle": 0, + "content": "[92] Pete Warden. Speech commands: A dataset for limited-vocabulary speech recognition. arXiv preprint arXiv:1804.03209, 2018.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.413, + 0.483, + 0.468 + ], + "angle": 0, + "content": "[93] Yixuan Wei, Han Hu, Zhenda Xie, Zheng Zhang, Yue Cao, Jianmin Bao, Dong Chen, and Baining Guo. Contrastive learning rivals masked image modeling in fine-tuning via feature distillation. Tech Report, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.469, + 0.483, + 0.523 + ], + "angle": 0, + "content": "[94] Kan Wu, Jinnian Zhang, Houwen Peng, Mengchen Liu, Bin Xiao, Jianlong Fu, and Lu Yuan. Tinyvit: Fast Pretraining Distillation for Small Vision Transformers. In ECCV, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.525, + 0.483, + 0.579 + ], + "angle": 0, + "content": "[95] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, 2015. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.581, + 0.483, + 0.621 + ], + "angle": 0, + "content": "[96] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.622, + 0.483, + 0.663 + ], + "angle": 0, + "content": "[97] Peng Xu, Xiatian Zhu, and David A Clifton. Multimodal learning with transformers: A survey. IEEE TPAMI, 45 (10):12113-12132, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.665, + 0.483, + 0.733 + ], + "angle": 0, + "content": "[98] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In CVPR, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.734, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[99] Chuanguang Yang, Zhulin An, Libo Huang, Junyu Bi, Xinqiang Yu, Han Yang, Boyu Diao, and Yongjun Xu. ClipKD: An Empirical Study of CLIP Model Distillation. In CVPR, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[100] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models (exemplified as a video agent). In ICML, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[101] Yazhou Yao, Tao Chen, Guo-Sen Xie, Chuanyi Zhang, Fumin Shen, Qi Wu, Zhenmin Tang, and Jian Zhang. Nonsalient region object mining for weakly supervised semantic segmentation. In CVPR, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[102] Yazhou Yao, Zeren Sun, Chuanyi Zhang, Fumin Shen, Qi Wu, Jian Zhang, and Zhenmin Tang. Jo-src: A contrastive approach for combating noisy labels. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[103] Junbo Yin, Jianbing Shen, Runnan Chen, Wei Li, Ruigang Yang, Pascal Frossard, and Wenguan Wang. Is-fusion: Instance-scene collaborative fusion for multimodal 3d object detection. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.221 + ], + "angle": 0, + "content": "[104] Shan You, Chang Xu, Chao Xu, and Dacheng Tao. Learning from multiple teacher networks. In KDD, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[105] Fei Yuan, Linjun Shou, Jian Pei, Wutao Lin, Ming Gong, Yan Fu, and Daxin Jiang. Reinforced multi-teacher selection for knowledge distillation. In AAAI, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[106] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xu-peng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In CVPR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[107] Yiyuan Zhang, Kaixiong Gong, Kaipeng Zhang, Hongsheng Li, Yu Qiao, Wanli Ouyang, and Xiangyu Yue. Meta-transformer: A unified framework for multimodal learning. arXiv preprint arXiv:2307.10802, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[108] Yanqi Zhou, Tao Lei, Hanxiao Liu, Nan Du, Yanping Huang, Vincent Zhao, Andrew M Dai, Quoc V Le, James Laudon, et al. Mixture-of-experts with expert choice routing. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.434, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[109] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, WANG HongFa, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei Li, Cai Wan Zhang, Zhifeng Li, Wei Liu, and Li Yuan. Languagebind: Extending Video-Language Pretraining to N-modality by Language-based Semantic Alignment. In ICLR, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[110] Xiatian Zhu, Shaogang Gong, et al. Knowledge distillation by on-the-fly native ensemble. In NeurIPS, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[111] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In ICCV, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[112] Konrad Zuchniak. Multi-teacher knowledge distillation as an effective method for compressing ensembles of neural networks. arXiv preprint arXiv:2302.07215, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.644 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29655" + } + ] +] \ No newline at end of file diff --git a/2025/UNIALIGN_ Scaling Multimodal Alignment within One Unified Model/7e109366-fc0f-438c-9b91-2ba7f07e59e9_origin.pdf b/2025/UNIALIGN_ Scaling Multimodal Alignment within One Unified Model/7e109366-fc0f-438c-9b91-2ba7f07e59e9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5b4b5b2f91958a33f5c59dc522a2f4af77c1c12a --- /dev/null +++ b/2025/UNIALIGN_ Scaling Multimodal Alignment within One Unified Model/7e109366-fc0f-438c-9b91-2ba7f07e59e9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0836cceec8f441f32c41297362297ac472c6bec9116c735a24fa44ce4eb17b20 +size 1157524 diff --git a/2025/UNIALIGN_ Scaling Multimodal Alignment within One Unified Model/full.md b/2025/UNIALIGN_ Scaling Multimodal Alignment within One Unified Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..517d23c338ce87362ab9fcdd978e61b9ef9b3607 --- /dev/null +++ b/2025/UNIALIGN_ Scaling Multimodal Alignment within One Unified Model/full.md @@ -0,0 +1,515 @@ +# UNIALIGN: Scaling Multimodal Alignment within One Unified Model + +Bo Zhou $^{1*}$ , Lilei Li $^{2*}$ , Yujia Wang $^{3}$ , Huafeng Liu $^{1}$ , Yazhou Yao $^{1\dagger}$ , Wenguan Wang $^{2}$ + +$^{1}$ Nanjing University of Science and Technology, $^{2}$ Zhejiang University, $^{3}$ Zhejiang Sci-Tech University + +https://github.com/NUST-Machine-Intelligence-Laboratory/UNIALIGN + +# Abstract + +We present UNIALIGN, a unified model to align an arbitrary number of modalities (e.g., image, text, audio, 3D point cloud, etc.) through one encoder and a single training phase. Existing solutions typically employ distinct encoders for each modality, resulting in increased parameters as the number of modalities grows. In contrast, UNIALIGN proposes a modality-aware adaptation of the powerful mixture-of-experts (MoE) schema and further integrates it with Low-Rank Adaptation (LoRA), efficiently scaling the encoder to accommodate inputs in diverse modalities while maintaining a fixed computational overhead. Moreover, prior work often requires separate training for each extended modality. This leads to task-specific models and further hinders the communication between modalities. To address this, we propose a soft modality binding strategy that aligns all modalities using unpaired data samples across datasets. Two additional training objectives are introduced to distill knowledge from well-aligned anchor modalities and prior multimodal models, elevating UNIALIGN into a high performance multimodal foundation model. Experiments on 11 benchmarks across 6 different modalities demonstrate that UNIALIGN could achieve comparable performance to SOTA approaches, while using merely 7.8M trainable parameters and maintaining an identical model with the same weight across all tasks. + +# 1. Introduction + +Humans accomplish tasks by processing and combining signals from different modalities [37, 53]. This gives rise to multimodal learning, which aims to build AI models that can extract and relate information across multimodal inputs [97]. Recent efforts have been made to learn a unified representation space via the alignment between modalities [18, 22, 39]. Though enabling extension to novel modalities, existing solutions typically rely on a one-versus-one paradigm where merely one extended modality (e.g., point + +![](images/b45890f026d892961b336e635b5a3c6d4df200ee34dac41aa598ba9183b70cc6.jpg) +Figure 1. Comparison of architectures between standard multi-modal alignment methods [18, 39, 109] and UNIALIGN: (a) Existing work adopts specialized encoders tailored to each modality; (b) UNIALIGN unifies the alignment across multiple modalities in a single model, with one training phase across all datasets. + +cloud, audio) could be aligned with the anchor modalities (e.g., text, image) at each time (Fig. 1(a)). This presents several drawbacks: i) one dedicated encoder is required for each modality, leading to exponential increases in training cost and parameter count as the number of aligned modalities grows, severely limiting scalability; ii) task-specific training for each dataset, results in multiple independent models for different tasks; and iii) the separate alignment of each extended modality hinders the model to leverage the full breadth of data across all datasets to shape the shared embedding space from a global view. This further impedes cross-modal knowledge transfer where insights gained from one modality could benefit understanding in others. + +These observations prompt us to rethink: Is this one-versus-one alignment paradigm essential? and Can we achieve joint alignment of all modalities within one single model? In response, we propose UNIALIGN (Fig. 1(b)), which embraces a all-in-one design to realize multimodal alignment by constructing a foundation model that includes: 1 one base encoder to UNify the encoding of all anchor and + +extended modalities, and $②$ one training phase to ALIGN all modalities with all samples across datasets collectively. + +Concretely, to achieve ①, recent work has consistently demonstrated that Transformers pre-trained on one modality possess exceptional generalization ability to novel modalities [13, 67]. Moreover, the mixture-of-experts (MoE) [72] schema presents a highly efficient strategy to scale model capacity under a fixed computational cost [108]. These insights motivate us to build a foundational model accommodating multiple modalities upon a single pre-trained Transformer (e.g., CLIP [67]). Here, MoE can dynamically scale and allocate model capacities by selectively activating experts to encode inputs across diverse modalities. We explore two strategies targeting at expert grouping and routing mechanism to enhance the awareness of different modalities within MoE. To preserve the knowledge learned from anchor modalities, Low-Rank Adaptation (LoRA) [27] is integrated with MoE, enabling a significant reduction in training overhead. To fulfill ②, we propose a soft modality binding strategy, leveraging massive unpaired data samples across modalities and datasets. Specifically, given unpaired data from diverse datasets with different modalities, the class labels of samples are encoded into text embeddings to create a unified representation. The semantic similarity between text embeddings can inform the alignment between modalities (e.g., audio and video), by encouraging the similarity between modality embeddings and textual label embeddings (e.g., audio $\leftrightarrow$ textvideo and video $\leftrightarrow$ textaudio) to be identical as the semantic similarity (i.e., $text_{\mathrm{video}} \leftrightarrow \text{text}_{\mathrm{audio}}$ ). Moreover, rooted in the principles of knowledge distillation [2, 26, 110], the alignment between extended and anchor modalities is guided by the knowledge persisting between anchor modalities, which are well-aligned in semantics compared to extended modalities. To further boost the performance, we transfer knowledge from multiple well-trained foundation models which only contain partial modalities. Given the complexity of multimodal, multi-dataset joint training, a specialized batching strategy combined with modality-aware gradient accumulation is proposed, to mitigate the potential dispersion in optimization across different modalities. + +UNIALIGN enjoys several advantages: First, it supports flexible alignment between arbitrary number of anchor and extended modalities, while ensuring negligible increase in parameter size. Second, the alignment for all modalities is accomplished in one training session, significantly improving training efficiency. Third, unpaired data from different modalities can be used collectively for optimization. Uncommon modalities with scarce data could take full advantages of the rich resource of common modalities. Fourth, benefited from soft modality binding, extended modalities are bridged for mutual boosting. Fifth, a compact and high-performance multimodal foundational + +model generally applicable to all tasks is delivered at minimal cost. This contrasts with existing approaches, which typically provide multiple loosely integrated modality encoders tailored to specific tasks, thereby enabling broader transferability to other domains and downstream applications [3, 8, 16, 61, 62, 95, 96, 101, 102]. + +By embracing the all-in-one philosophy, UNIALIGN demonstrates remarkable efficiency in both training cost and parameter size. After a single pre-training step without further tuning, UNIALIGN achieves top-leading performance on 11 benchmarks with one identical model and 7.8M trainable parameters across 6 modalities (previous SOTA [18, 39] needing over 80M trainable parameters for each extended modality). This evidences the great potential of our unified modality encoding and aligning strategies. We hope this work could stimulate the community to rethink multimodal alignment from a universal and compact perspective. + +# 2. Related Work + +Multimodal Learning. Multimodal learning has attracted significant attention given its ability to address a wide range of applications by integrating information from diverse modalities such as text, images, audio, and video[49, 51, 67, 103]. Current research in this field generally falls into three categories: First, methods involving joint training of multiple modalities aim to improve performance across different downstream tasks, by leveraging shared representations[17, 78, 107]. Though incorporating multiple modalities as inputs, these work lacks explicit alignment between modalities. Second, inspired by the successes of vision-language alignment [67], some approaches extend this alignment to additional modalities[18, 38, 40, 56, 59, 91, 98, 109]. These methods typically require an independent encoder for each modalities. Though [14, 18, 39, 90] maintain an consistent architecture across tasks, they conduct separate training on individual datasets, resulting in multiple task-specific models. Additionally, these solutions often rely on massive paired data to achieve robust zero-shot performance across various tasks, ignoring the usage of rich unpaired data. Third, the integration of multimodal inputs with large language models (LLMs) has opened new avenues for multimodal learning[23, 50, 52, 88, 100]. In these approaches, all modalities are projected into the same representation space as LLMs, without alignment among extended modalities and supporting merely one anchor modality (i.e., text). This leads to suboptimal performance on certain vision tasks and limits the applications primarily to language-related fields. + +In this work, we propose UNIALIGN, supporting alignment between an arbitrary number of anchor and extended modalities within one model using one unified modality encoder. After a single training phase, UNIALIGN can handle a dozen of downstream tasks without any fine-tuning. UNIALIGN also achieves high training efficiency, utilizing + +merely 7.8M training parameters for all tasks. + +Multi-Task Learning. To maximize the utilization of resource and improve generalization across tasks, multi-task learning seeks to develop models that can handle multiple tasks simultaneously [4]. Traditional methods often require separate subnetworks for each task, which hinders scalability [83]. In recent years, the mixture-of-experts (MoE) architecture [15, 72] has gained prominence in multi-task learning [31]. MoE leverages multiple subnetworks, or "experts", that are selectively activated to optimize task performance while reducing computational costs [35]. This approach proves beneficial across various domains, such as language processing [31, 35] and vision tasks [41, 43, 69]. With the success of LoRA [27], there has been a surge to integrate MoE with LoRA, which shows promise in enhancing both computational efficiency and task adaptability in large-scale models [11, 21, 28, 29]. However, these methods are mainly confined to text and image domains [7, 11], and lack alignment between modalities [46]. In contrast, UNIALIGN extends MoE to a multimodal setup, devising modality-aware expert selection and routing mechanism. + +Knowledge Distillation. Knowledge distillation [26, 34, 55] is an essential technique for model compression, enabling the transfer of knowledge from large, computationally intensive models to smaller, more efficient ones. This is particularly valuable for multimodal foundation models, where the computational overload is heavy [32, 45, 79, 99]. Initially, knowledge distillation focused on extracting knowledge from individual large models, typically using the logits or intermediate feature representations to guide the training of student models [1, 25, 36, 63, 93]. Recent research has expanded this concept to integrate capabilities from multiple models, creating hybrid models that not only improve task-specific performance but also broaden the range of potential applications [68, 71, 84, 104, 105, 112]. + +Compared to existing work, UNIALIGN distills knowledge from both well-aligned anchor modalities, and multiple well-trained foundation models specialized in partial modalities. This allows our method to handle a broader range of tasks that require understanding and processing information with different composition of diverse modalities. + +# 3. Method + +In this section, we first introduce the unified encoding of all modalities within one single model (§3.1), then describe the universal modality alignment strategy across all datasets (§3.2), and finally present the implementation details of UNIALIGN (§3.3). + +# 3.1. Unified Modality Encoding + +Preliminary: Mixture-of-Experts (MoE). Let $X \in \mathbb{R}^{L \times D}$ denote the tokenized input, where $L$ is the sequence length and $D$ is the token embedding size. In vanilla Transformer, + +![](images/465c5463cfdf1aa2bb4fd9c957df010e25ee5aad93f5e5fbfa454961685ee04d.jpg) +Figure 2. Architecture of the unified modality encoder. + +each encoder block comprises a multi-head self-attention (MSA) layer and a feed-forward network (FFN), as follows: + +$$ +\boldsymbol {X} ^ {\prime} = \boldsymbol {X} + \mathcal {F} ^ {\mathrm {M S A}} (\ln (\boldsymbol {X})), \tag {1} +$$ + +$$ +\boldsymbol {X} ^ {\prime \prime} = \boldsymbol {X} ^ {\prime} + \mathcal {F} ^ {\mathrm {F F N}} (\ln (\boldsymbol {X} ^ {\prime})), \tag {2} +$$ + +where $X''$ is the outputs and LN represents layer normalization. In the mixture-of-experts (MoE) schema, an MoE function (i.e., $\mathcal{F}^{\mathrm{MoE}}$ ) is adopted to replace $\mathcal{F}^{\mathrm{FFN}}$ : + +$$ +\boldsymbol {X} ^ {\prime \prime} = \boldsymbol {X} ^ {\prime} + \mathcal {F} ^ {\mathrm {M o E}} (\ln \left(\boldsymbol {X} ^ {\prime}\right)), \tag {3} +$$ + +$$ +\mathcal {F} _ {\mathrm {M o E}} \left(\boldsymbol {X} ^ {\prime}\right)) = \sum_ {i = 1} ^ {n} \mathcal {G} \left(\boldsymbol {X} ^ {\prime}\right) _ {i} \mathcal {F} _ {i} ^ {\text {E x p e r t}} \left(\boldsymbol {X} ^ {\prime}\right), \tag {4} +$$ + +where $\mathcal{F}^{\mathrm{MoE}}$ jointly aggregates outputs from multiple expert networks (i.e., $\mathcal{F}^{\mathrm{Expert}}$ ) according to the router $\mathcal{G}(X')_i = \text{softmax}(\mathcal{F}^{\mathrm{Router}}(X'))_i$ to assign weights for different experts and render the predictions. + +Multimodal Adaptation with MoE. We aim to unify the encoding of multimodal inputs within one foundation model (e.g., the image encoder of CLIP[67]). As illustrated in Fig.2, inputs of various modalities are first converted into sequential tokens. We follow [39, 42, 107] to preprocess multimodal data and implement modality-specific tokenizers, with details summarized in Appendix. To adapt the encoding of samples from various modalities, UNIALIGN explores two modality-aware MoE strategies: + +- Modality-Aware Expert Grouping divides all $\mathcal{F}^{\mathrm{Expert}}$ in one MoE layer into multiple sets, where each set $\mathcal{B}$ contains $n$ experts to process inputs from a specific modality $\mathcal{M}$ : + +$$ +\mathcal {B} _ {\mathcal {M}} = \left\{\mathcal {F} _ {\mathcal {M}, 1} ^ {\text {E x p e r t}}, \dots , \mathcal {F} _ {\mathcal {M}, n} ^ {\text {E x p e r t}} \right\}. \tag {5} +$$ + +Under this setup, following [46], we construct a hierarchical routing mechanism. In the first routing round, tokens are + +directed to a specific expert set based on the modality of inputs. Next, within a expert set, a learnable router $\mathcal{F}_{\mathcal{M}}^{\text{Router}}$ for modality $\mathcal{M}$ assigns weights to each individual expert conditioned on the input tokens. This separation lets each expert set to specialize in one modality, and allows for easy expansion to novel modalities by loading new expert sets. + +- Modality-Aware Global Routing uses multiple modality-specific routers, to direct tokens across all experts in a MoE layer without grouping. Given the routing function $\mathcal{F}_{\mathcal{M}}^{\text{Router}}$ for modality $M$ , the global weight $\mathcal{G}_{\mathcal{M}}$ is computed as: + +$$ +\mathcal {G} _ {\mathcal {M}} \left(\boldsymbol {X} ^ {\prime}\right) _ {i} = \operatorname {s o f t m a x} \left(\mathcal {F} _ {\mathcal {M}} ^ {\text {R o u t e r}} \left(\boldsymbol {X} ^ {\prime}\right)\right) _ {i}. \tag {6} +$$ + +By allowing the experts to be accessible to each router, this strategy enables cross-modality knowledge sharing among experts. However, since the experts are not explicitly specialized for one modality, the generalization potentially at the expense of fine-grained modality specialization. + +Modality-Aware LoRA Expert. Low-Rank Adaptation (LoRA) [28] has demonstrated both effectiveness and efficiency in fine-tuning large models [9, 81]. Therefore, to maintain the knowledge in pre-trained foundation models, the original feed-forward network (i.e., $\mathcal{F}^{\mathrm{FFN}}$ in Eq. 2) are preserved in a fixed state instead of being replaced during the forward of MoE (i.e., Eq. 3), resulting in: + +$$ +\boldsymbol {X} ^ {\prime \prime} = \boldsymbol {X} ^ {\prime} + \mathcal {F} ^ {\mathrm {F F N}} (\ln (\boldsymbol {X} ^ {\prime})) + \mathcal {F} ^ {\mathrm {M o E}} (\ln (\boldsymbol {X} ^ {\prime})). \tag {7} +$$ + +Here the experts in $\mathcal{F}^{\mathrm{MoE}}$ leverage $\mathcal{F}^{\mathrm{FFN}}$ as the base weight for LoRA adaption, with each of them employs a specific instance of LoRA (i.e., $\mathcal{F}^{\mathrm{LoRA}}$ ) parameterized as follows: + +$$ +\mathcal {F} _ {\operatorname {M o E}} \left(\boldsymbol {X} ^ {\prime}\right)) = \sum_ {i = 1} ^ {n} \mathcal {G} \left(\boldsymbol {X} ^ {\prime}\right) _ {i} \mathcal {F} _ {i} ^ {\text {L o R A}} \left(\boldsymbol {X} ^ {\prime}\right), \tag {8} +$$ + +$$ +\mathcal {F} _ {i} ^ {\text {L o R A}} \left(\boldsymbol {X} ^ {\prime}\right) = \frac {\alpha}{r} \cdot \boldsymbol {B} _ {i} \boldsymbol {A} _ {i} \boldsymbol {X} ^ {\prime}, \tag {9} +$$ + +where low-rank matrices $\mathbf{A} \in \mathbb{R}^{d_{in} \times r}$ and $\mathbf{B} \in \mathbb{R}^{r \times d_{out}}$ are scaled by a factor $\alpha$ normalized by the rank $r \ll \min(d_{in}, d_{out})$ . In this way, we achieve seamless integration of LoRA with MoE, enabling efficient tuning and knowledge preservation of the foundation models, while supporting adaptation to inputs across various modalities. + +# 3.2. Universal Modality Alignment + +Soft Modality Binding across Datasets. Prior work [18] reveals that modalities expressed in a joint embedding space could be automatically aligned through an intermediate modality. For example, the alignment of and , could lead to the alignment between with Image as the bridge. However, this approach requires massive naturally paired samples between anchor and extended modalities, which are often challenging and time-consuming to obtain. We thus propose a soft modality binding strategy, which utilizes unpaired samples across datasets with different modalities + +![](images/408eda2d6b769a4c6ec7401c6deb1b2233c62cb3e50df4ccf3e3569909b193d1.jpg) +Figure 3. Our proposed soft modality binding strategy. Unpaired samples from diverse datasets with different modalities are bridged with the semantic similarity between class labels as soft guidance. + +(Fig. 3). Specifically, let $\mathbf{y}^{\mathcal{M}_1} = \{\pmb {y}_1^{\mathcal{M}_1},\dots ,\pmb {y}_{N_1}^{\mathcal{M}_1}\}$ and $\pmb {\mathcal{V}}^{\mathcal{M}_2} = \{\pmb {y}_1^{\mathcal{M}_2},\dots ,\pmb {y}_{N_2}^{\mathcal{M}_2}\}$ represent text embeddings of class labels from two arbitrary datasets in modalities $\mathcal{M}_1$ and $\mathcal{M}_2$ , respectively. The semantic similarity of textual labels across datasets can be computed as: + +$$ +\mathcal {S} \left(\boldsymbol {y} _ {i} ^ {\mathcal {M} _ {1}}, \boldsymbol {y} _ {j} ^ {\mathcal {M} _ {2}}\right) = \frac {\exp \left(\boldsymbol {y} _ {i} ^ {\mathcal {M} _ {1}} \cdot \boldsymbol {y} _ {j} ^ {\mathcal {M} _ {2}}\right)}{\sum_ {n = 1} ^ {N _ {2}} \exp \left(\boldsymbol {y} _ {i} ^ {\mathcal {M} _ {1}}, \boldsymbol {y} _ {n} ^ {\mathcal {M} _ {2}}\right)}. \tag {10} +$$ + +Similarly, given samples $(\pmb{x}^{\mathcal{M}_1},\pmb{y}_i^{\mathcal{M}_1})$ from the first dataset and $(\pmb{x}^{\mathcal{M}_2},\pmb{y}_j^{\mathcal{M}_2})$ from the second dataset, the semantic similarity between $\pmb{x}^{\mathcal{M}_1}$ and text embedding of labels from the second dataset (i.e., $S(\pmb{x}^{\mathcal{M}_1},\pmb{y}_j^{\mathcal{M}_2}))$ , as well as $\pmb{x}^{\mathcal{M}_2}$ and text embeddings of labels from the first dataset (i.e., $S(\pmb{x}^{\mathcal{M}_2},\pmb{y}_i^{\mathcal{M}_1}))$ could be computed. Taking the similarity between class labels of two datasets (i.e., $S(\pmb{y}_i^{\mathcal{M}_1},\pmb{y}_j^{\mathcal{M}_2}))$ as a soft prior for alignment, we bind modalities across datasets by optimizing the following objective: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {s o f t \_ b i n d}} = \mathrm {K L} \left(\mathcal {S} \left(\boldsymbol {x} ^ {\mathcal {M} _ {1}}, \boldsymbol {y} _ {j} ^ {\mathcal {M} _ {2}}\right) \mid \mathcal {S} \left(\boldsymbol {y} _ {i} ^ {\mathcal {M} _ {1}}, \boldsymbol {y} _ {j} ^ {\mathcal {M} _ {2}}\right)\right) \\ + \operatorname {K L} \left(\mathcal {S} \left(\boldsymbol {x} ^ {\mathcal {M} _ {2}}, \boldsymbol {y} _ {i} ^ {\mathcal {M} _ {1}}\right) \mid \mathcal {S} \left(\boldsymbol {y} _ {j} ^ {\mathcal {M} _ {2}}, \boldsymbol {y} _ {i} ^ {\mathcal {M} _ {1}}\right)\right), \\ \end{array} +$$ + +where KL refers to the Kullback-Leibler divergence. Our soft modality binding strategy is different from [18], which requires instance-level textual annotations for each sample. Instead, our approach utilizes merely class labels which are instance-agnostic and challenging for alignment. + +Knowledge Distillation from Anchor Modalities. Given multiple well-aligned anchor modalities (e.g., $\mathcal{M}_1^a$ and $\mathcal{M}_2^a$ ) and one novel modality to be extended (e.g., $\mathcal{M}_k^e$ ), the alignment performance between anchor and extended modalities (e.g., $\mathcal{M}_k^e$ and $\mathcal{M}_1^a$ ) is typically less effective than that between anchor modalities (e.g., $\mathcal{M}_1^a$ and $\mathcal{M}_2^a$ ). Therefore, + +![](images/e1b8aceedb68cc74451b32ccc1739996b9e076f40208947c09536b91a150e785.jpg) +Figure 4. Top: Knowledge distillation from anchor modalities to instruct the alignment of novel modalities. Bottom: knowledge distillation from multiple partial-modal foundation models. + +knowledge within the alignment between anchor modalities (Fig. 4 (Top)) could be transferred to improve the alignment learning of extended modalities via knowledge distillation. + +For instance, given a 3D point cloud sample $x_{i}$ , we could obtain its tokenized input $\pmb{x}_{i}^{\mathcal{P}}$ in the extended modalities, along with the text embedding of class label $\pmb{x}_{i}^{\mathcal{T}}$ and the visual embedding of one rendered image from a specific viewpoint $\pmb{x}_{i}^{\mathcal{T}}$ , both in the anchor modalities. By leveraging the similarity between anchor modalities (i.e., $S(\pmb{x}_{i}^{\mathcal{T}},\pmb{x}_{i}^{\mathcal{T}})$ ) as the guidance, we could optimize the alignment between the extended modality and anchor modalities through: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {a n c h . d i s t i l}} = \mathrm {K L} \left(S \left(\boldsymbol {x} _ {i} ^ {\mathcal {P}}, \boldsymbol {x} _ {i} ^ {\mathcal {T}}\right) \mid S \left(\boldsymbol {x} _ {i} ^ {\mathcal {I}}, \boldsymbol {x} _ {i} ^ {\mathcal {T}}\right)\right) \tag {12} \\ + \operatorname {K L} \left(\mathcal {S} \left(\boldsymbol {x} _ {i} ^ {\mathcal {P}}, \boldsymbol {x} _ {i} ^ {\mathcal {I}}\right) \mid \mathcal {S} \left(\boldsymbol {x} _ {i} ^ {\mathcal {T}}, \boldsymbol {x} _ {i} ^ {\mathcal {I}}\right)\right). \\ \end{array} +$$ + +This knowledge distillation strategy is generally applicable to most modalities, where 3D point cloud, video, and depth data can be converted into images with class labels. + +Knowledge Distillation from Partial-Modal Foundation Models. To further boost the alignment across modalities, we employ foundation models $\mathcal{E}_f$ skilled in partial modalities (e.g., audio $\leftrightarrow$ text [22], point cloud $\leftrightarrow$ image [98]) as teacher models to guide the learning of our unified modality encoder $\mathcal{E}_u$ (i.e., Fig.4 (Bottom)) from two perspectives: + +- Feature Alignment aims to ensure features from our unified modality encoder to be identical as those from partial-modal foundation models. Specifically, denoting $\hat{F}_i^{\mathcal{M}_k}$ and $F_{i}^{\mathcal{M}_{k}}$ as features in modality $\mathcal{M}_k$ output by $\mathcal{E}_f$ and $\mathcal{E}_u$ , we use the mean squared error (MSE) to align them as: + +$$ +\mathcal {L} _ {\text {f e a t - a l i g n}} = \left\| \hat {\boldsymbol {F}} ^ {\mathcal {M} _ {i}} - \boldsymbol {F} ^ {\mathcal {M} _ {i}} \right\| _ {2} ^ {2}. \tag {13} +$$ + +- Cross-Modality Knowledge Transfer aims to benefit the learning of $\mathcal{E}_u$ via the well-aligned representation space across modalities shaped by large-scale paired data. Specifically, given a paired sample $(x_i^{\mathcal{M}^a}, x_i^{\mathcal{M}^e})$ in anchor modality $\mathcal{M}^a$ and extended modality $\mathcal{M}^e$ respectively, the align + +ment results of our unified modality encoder are encouraged to approximate those of partial-modal foundation models: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {m o d a l - a l i g n}} = \mathrm {K L} \left(\mathcal {S} \left(\boldsymbol {x} _ {i} ^ {\mathcal {M} ^ {e}}, \boldsymbol {x} _ {i} ^ {\mathcal {M} ^ {a}}\right) \mid \mathcal {S} \left(\hat {\boldsymbol {x}} _ {i} ^ {\mathcal {M} ^ {e}}, \hat {\boldsymbol {x}} _ {i} ^ {\mathcal {M} ^ {a}}\right)\right) \tag {14} \\ + \mathrm {K L} \left(S \left(\boldsymbol {x} _ {i} ^ {\mathcal {M} ^ {a}}, \boldsymbol {x} _ {i} ^ {\mathcal {M} ^ {e}}\right) \mid S \left(\hat {\boldsymbol {x}} _ {i} ^ {\mathcal {M} ^ {a}}, \hat {\boldsymbol {x}} _ {i} ^ {\mathcal {M} ^ {e}}\right)\right), \\ \end{array} +$$ + +where $\hat{\pmb{x}}_i$ and $\pmb{x}_i$ are outputs of $\mathcal{E}_f$ and $\mathcal{E}_u$ , respectively. + +In summary, the loss for knowledge distillation from partial-modal foundation models can be formulated as: + +$$ +\mathcal {L} _ {\text {m o d e l - d i s t i l}} = \lambda \mathcal {L} _ {\text {f e a t - a l i n g}} + \mathcal {L} _ {\text {m o d a l - a l i n g}}. \tag {15} +$$ + +Following [99], we set $\lambda = 2000$ to scale the magnitude. + +# 3.3. Implementation Details + +Network Configuration. We adopt CLIP [67] pre-trained vision encoder as the base network for unified modality encoding. The sequence length $L$ and channel dimension $D$ for tokenized inputs $X$ are set to 196 and 768, respectively. The LoRA configuration is defined with a scaling factor $\alpha$ of 32 and a rank $r$ of 4. For knowledge distillation from partial-modal foundation models, we use ULIP [98] as the teacher model for 3D point cloud inputs, VIT-LENS[38] for depth inputs, ONE-PEACE[86] for audio and video inputs. + +Batching Strategy. The batching strategies proposed by [17, 42] sample inputs from each dataset (modality) separately, which may lead to catastrophic forgetting[42]. To address this, we develop a new batching strategy, which mixes data from all modalities within each mini-batch. Furthermore, a modality-aware gradient accumulation is proposed to perform forward pass, loss computation, and gradient calculation individually for each modality. Backpropagation is conducted only after processing all modalities. + +Training Objective. Following [68], we assign an equal weight to all training objectives: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {N C E}} + \alpha_ {1} \mathcal {L} _ {\text {s o f t , b i n d}} + \alpha_ {2} \mathcal {L} _ {\text {a n c h , d i s t i l}} + \alpha_ {3} \mathcal {L} _ {\text {m o d e l , d i s t i l}}, \tag {16} +$$ + +where $\alpha_{\{1,2,3\}} = 1$ and $\mathcal{L}_{\mathrm{NCE}}$ is the modality aligning loss for paired data, as utilized in CLIP [67]. + +# 4. Experiments + +# 4.1. Experimental Setup + +Pre-training Datasets. We perform a joint pre-training across multiple datasets with different modalities, including images from ImageNet-1K [8], 3D point clouds from ShapeNet[5], depth data from SUN RGB-D[74], audio and video from AudioSet [16] and VGGSound [6]. To address the differences in dataset size across modalities, we apply repeated augmentation[17] to small datasets (such as SUN-D) and category-balanced downsampling for large datasets (such as AudioSet) to ensure balanced performance during joint training. Once finished pre-training, we directly evaluate our model on downstream tasks without any fine-tuning. + +Training. For image and video data, we use a resolution of $224 \times 224$ and apply standard augmentations. For single-view depth data, we employ the same augmentation as in + +
ModelBackboneTrainable ParamsTop-1 AccuracyTop-5 Accuracy
Single Task
Swin Transformer [47]Swin-B87.7M83.596.5
ConvNeXt [48]ConvNeXt-B88.7M83.896.7
DeiT III [80]ViT-B86.9M83.896.5
Conformer [64]Conformer-B83.3M83.896.7
Hiera [70]Hiera-B+70.0M85.297.3
Multiple Tasks
UniRepLKNet [10]UniRepLKNet-S55.6M83.9-
OMNIVORE [17]Swin-B88.2M85.397.5
MetaFormer [107]ViT-B86.6M85.497.4
UNIALIGN†ViT-B7.8M85.397.2
+ +An identical model with the same weight across all tasks. +Table 1. Results for image classification on ImageNet-1K (§4.2). + +
ModelBackboneTrainable ParamsModelNet40ScanObjectNN
Top-1Top-5Top-1Top-5
Single Task
PointCLIP [106]ResNet-50-19.334.810.530.6
PointCLIPv2 [111]ViT-B-63.685.042.274.5
CLIP2Point [30]ViT-B88.3M49.581.225.559.4
RECON [66]ViT-B43.6M61.278.142.375.6
ULIP [98]ViT-B43.6M60.484.051.580.2
Multiple Tasks
VIT-LENS [38]ViT-B34.1M65.492.7--
UNIALIGN†ViT-B7.8M55.984.342.576.3
+ +An identical model with the same weight across all tasks. +Table 2. Results for zero-shot 3d shape classification on ModelNet40 test [95] and ScanObjectNN test [82] (\$4.3). + +[38]. The augmentation of audio data also follows [38] except Mix-up [94]. We train the model with a batch size of 2048 (after gradient accumulation) for 100 epochs. An AdamW optimizer with a initial learning rate $2\mathrm{e}^{-4}$ and a weight decay of 0.05 is employed. For knowledge distillation from partial-modal foundation models, we cache offline features to avoid repeatedly forwarding. The base modality encoder remains frozen during training, with only LoRA enhanced MoE modules and the projection head for knowledge distillation remaining trainable. Flash-attention is applied to reduce GPU memory cost. + +Reproducibility. UNIALIGN is implemented in PyTorch and trained on four NVIDIA Tesla A100 GPUs. + +# 4.2. Results for Image Classification + +Dataset. We evaluate UNIALIGN for image classification on ImageNet-1K val [8] which contains 50,000 samples across 1,000 classes, spanning objects, animals, and more. + +Metric. Following previous works [47, 48, 87], we report the count of trainable parameters, top-1 and top-5 accuracy. + +Performance. As shown in Table 1, UNIALIGN demonstrates impressive performance on image classification, achieving a comparable performance to modality-specific approaches (i.e., $85.3\%$ vs. $85.2\%$ of Hiera[70]). It is noteworthy that our method utilizes merely $9\%$ trainable parameters and a unified modality encoder architecture which preserves the capability to process other modalities. + +
ModelBackboneTrainable ParamsAnchor ModalityNYU-DSUN-D
Single Task
G-L-SOOR [76]D-CNN [75]--62.344.1
RecgNet [12]ResNet18 [24]11.2M-57.747.9
Depth Swin [18]Swin-B88.2M-76.469.1
Multiple Tasks
Text Paired [18]ViT-H88.1MT41.925.4
ImageBind [18]ViT-H88.3MI54.035.1
ViT-LENS [38]ViT-L50.9MI+T68.552.2
UNIALIGN†ViT-B7.8MT46.334.3
UNIALIGN†ViT-B7.8MI+T59.445.8
UNIALIGN†ViT-B7.8MI+T71.4*58.2*
+ +An identical model with the same weight across all tasks. +Table 3. Results for depth-only scene classification on NYU-D test [73] and SUN-D test [74] (\$4.4).*: The depth maps are tokenized by the image tokenizer additionally. + +# 4.3. Results for Zero-Shot 3D Shape Classification + +Dataset. We evaluate zero-shot 3D shape classification on two datasets. ModelNet40 test [95] contains 2,468 samples from 40 categories, while ScanObjectNN test [82] consists of 581 samples distributed across 15 categories. + +Metric. The top-1 and top-5 accuracy are reported. + +Performance. As shown in Table 2, UNIALIGN achieves a top-1 accuracy of $55.9\%$ and $42.5\%$ on ModelNet40 val, and ScanObjectNN test, respectively. Though SOTA approaches such as ULIP [98] and ViT-LENS [38] yield higher scores, they require a modality-specific point-cloud encoder and an additional image encoder for alignment. + +# 4.4. Results for Depth-only Scene Classification + +Dataset. Experiments for depth-only scene classification are conducted on SUN RGB-D test [74] which contains 4,660 samples across 19 classes, and NYU-D test [73] which contains 653 samples across 10 semantic classes. + +Metric. The top-1 accuracy is adopted for evaluation. + +Performance. A detailed comparison of UNIALIGN against several top-leading approaches for depth-only scene classification is provided in Table 3. Building on a ViT-B backbone and aligning extended modalities with both image and text as anchor modalities, UNIALIGN achieves a top-1 accuracy of $71.4\%$ on NYU-D test and $58.2\%$ on SUN RGB-D test, surpassing all competitors. Notably, aligning merely to the text modality leads to a marked performance decline (i.e., $71.4\% \rightarrow 46.3\%$ on NYU-D). This suggests that i) the depth modality exhibits a closer alignment with image data, and ii) UNIALIGN can effectively leverage the additional anchor modality to enhance alignment performance. + +# 4.5. Results for Audio Classification and Retrieval + +Datasets. We evaluate audio classification on AudioSet val [16] containing 2M samples across 527 classes, and ESC [65] including 2K data across 50 classes. Audio-text retrieval is evaluated on AudioCaps test [33] which includes 813 audio-text pairs. Moreover, following [107], we + +
MethodBackboneAnchor ModalityAudioSet mAPESC Top1AudioCaps R@1R@10
Single Task
AVFIC [57]ViT-B---8.737.7
AudioCLIP [22]ViT-BI+T25.969.4--
Multiple Tasks
ImageBind-H [18]ViT-HI17.666.99.342.3
VIT-LENS [38]ViT-LI+T27.280.914.955.2
LanguageBind [109]ViT-LI+T27.791.812.253.2
UNIALIGN†ViT-BI+T23.870.511.749.3
†:An identical model with the same weight across all tasks. +Table 4. Results for audio classification on AudioSet val [16], ESC[65], and audio-text retrieval on AudioCaps test[33] (§4.5). Due to expiration of links, UNIALIGN is trained with only 0.5M ideas from AudioSet, while methods such as ViT-Lens use 1.6M.
MethodBackbonePre-train DatasetsTrainable ParamsTop-1
Single Task
AST [19]ViT-BN/A86.9M92.6
AST [19]ViT-BAudioSet-20K86.9M96.2
SSAST [20]ViT-BAudioSet-2M89.3M97.8
Multiple Tasks
MetaFormer [107]ViT-BLAION2B1.1M78.3
MetaFormer [107]ViT-BLAION2B86.3M97.0
UniRepLKNet [10]UniRepLKNet-S-55.5M98.5
UNIALIGNViT-BJoint1.8M97.4
+ +fine-tune UNIALIGN on Speech Commands V2 train[92] for speech recognition. + +Metric. Following [18, 22, 38], we measure mAP for AudioSet, top-1 accuracy for ESC and Speech Commands V2, and Recall@ $k$ ( $k = \{1,5\}$ ) for AudioCaps. + +Performance for Audio classification. UNIALIGN achieves promising results on AudioSet and ESC (Table 4). Though inferior to top-leading methods, it is important to note that, due to the expiration of links, a significant number of videos from AudioSet can not be downloaded from YouTube (i.e., 1.6M used in ViT-Lens[38] vs. 0.5M used in UNIALIGN). + +Performance for Audio retrieval. Similarly, despite the absence of 1.1M training data, UNIALIGN can still achieve comparable performance, achieving 11.7 R@1 accuracy. + +Performance for Speech Recognition. As shown in Table 5, UNIALIGN demonstrates remarkable accuracy compared to existing work (e.g., $97.4\%$ vs. $97.0\%$ for MetaTransformer[107]), while utilizing significantly fewer trainable parameters (e.g., 1.8M vs. 86.3M). When tuning a similar number of parameters (i.e., 1.1M for Meta-Transformer, and 1.8M for UNIALIGN), our method outperforms Meta-Transformer by $19.1\%$ (i.e., $97.4\%$ vs. $78.3\%$ ). + +# 4.6. Results for Video Classification and Retrieval + +Datasets. We utilize MSR-VTT val [96] which contains around 1,000 video-text pairs for zero-shot video-text retrieval, and UCF101 val [77] which contains 3,738 samples across 101 semantic classes for video classification. + +Table 5. Results for audio classification on Speech Commands V2 val [92]. All models are fine-tuned on the training set (§4.5). + +
ModelBackboneRetrieval ModalityTrainable ParamsR@1R@5
Single Task
MIL-NCE [54]ResNeXt-101V88.8M8.625.8
SupportSet [60]R3D-34V93.5M10.430.0
AVFIC [57]ViT-BA+V86.6M19.450.3
InternVideo [89]ViT-LV305M37.571.3
Multiple Tasks
ImageBind [18]ViT-HA+V89.3M36.870.0
ViT-LENS [38]ViT-LA+V127.6M37.672.6
UNIALIGN†ViT-BA+V7.8M37.070.7
+ +†:An identical model with the same weight across all tasks. Table 6. Results for video retrieval on MSRVTT val [96] (§4.6). + +
ModelBackboneZero-ShotFine-Tuned
Single Task
ResT [44]ResNet-10158.7-
ActionCLIP [85]ViT-B58.3-
X-CLIP [58]ViT-B72.0-
Multiple Tasks
MetaFormer [107]ViT-B-46.6
ImageBind [18]ViT-H64.898.1
UniBind [50]ViT-H73.793.3
UNIALIGN†ViT-B68.594.4
+ +†:An identical model with the same weight across all tasks. Table 7. Results for video classification on UCF101 val[77] (§4.6). + +Metric. Recall@ $k$ and top-1 accuracy are adopted as evaluation metrics for video-text retrieval and classification. + +Performance for Video-Text Retrieval. As shown in Table 6, UNIALIGN achieves a R@1 score of $37.0\%$ and R@5 score of $70.7\%$ , which are comparable to advanced methods while using merely 7.8M trainable parameters. + +Performance for Video Classification. We summarize the results for video classification in Table 7. As shown, UNIALIGN achieves an impressive accuracy of $68.5\%$ on UCF101 val[77] without any fine-tuning, and outperforms ImageBind [18]. After fine-tuning all models on the training set of UCF101, our method achieves the second-highest performance. This may be attributed to the differences in backbone, as ImageBind utilizes ViT-H which could provide enhanced knowledge for downstream fine-tuning. + +# 4.7. Qualitative Results + +As shown in Fig. 5, UNIALIGN facilitates accurate cross-modality retrieval. Moreover, modalities can be combined to constrain attributes of the target in another modality. + +# 4.8. Diagnostic Experiment + +To evaluate the core designs of UNIALIGN and gain further insights, we conduct a series of ablation studies. Results are reported on SUN-D test [74] and ModelNet40 test [5]. + +Modality-Aware MoE Strategy. As illustrated in §3.1, we explore two modality-aware MoE strategies: expert grouping and global routing. Results in Table 8 indicate that dividing experts into modality-specific groups yields better performance. Thus, we adopt it as the default MoE strategy. Expert Network. Table 9 presents the ablation study on + +![](images/4b991b2df21667b4f2c7d7f2e19b0e392cb3317f451eee39701e935e95a6ea8f.jpg) + +![](images/83520fce3aad31423abe0e14d9da6379fcd1420d0de4e4d7ad60a12be7325322.jpg) + +![](images/456d96ca12fc305a2e1e3da1f7a95c194c781d09a07d7f272a62ae910185cc81.jpg) + +![](images/c6fd2eda02bf680ceb17bf6002ca3e3b541cac4060260a377ce35b63dc4afdfe.jpg) + +![](images/b3ccf9fc542160290081623aedc6ed43afdd32dc7bb577cb7f184e7e7bc3b67a.jpg) +Aircraft engine + +![](images/471ab080e8a627dce21e4b4181fffc0f2c98c6a7c21a2d8255829b11d2707f07.jpg) +Playing volleyball + +![](images/6810b160a4190e4bdce464dc96fb2d2316f5bcb250fdad03e5b71f90fd73a280.jpg) + +![](images/0ae38f0368dbcdec74429c714f72e20ad2eeb06424a5ec06f3f2418eafb48824.jpg) + +![](images/f0e430a016852f7b3c47c1fbd96646c2b3401a0df95fa3e5b9389d5f708aad35.jpg) +Dog + +![](images/3afd6354d70e4a60eff22c65112d101bbc8c2e07b4d5dbcdda88d0ff298d7664.jpg) + +![](images/843190b71736fa50333b9a41a8ddb607ddb61f0f3418f1997d539820dba8c77f.jpg) +Bathroom + +![](images/aa2d3368c3f33b2e83c4cd06a90783839eec841cb5b05c73ee59037f9769485b.jpg) +Car horn + +![](images/2c1080163687ece604b2122bf557fb934620bfe5c7ec6559fb878aa1a19f4a59.jpg) + +![](images/23c6e52cf2bf9f3e379bd47312c6ca1f70e947cfb2bf6b1cf14a8e3d502a3675.jpg) + +![](images/4e5526b97456a54eb70408a15ce25f6d11d81e320a0b974b55edf620fcde61e1.jpg) +Playing piano + +![](images/61ad3d00d9d8dcb8e42597099ef7f8e9189e0657b3b13a19bb371717b537728f.jpg) + +![](images/d45979826f20a9f199d0319e0a3f1fc6028e9c0e717e8fc93a53d98516e50f1b.jpg) +Typing on computer keyboard + +![](images/be6eeea0e7e1d7e64f4839649e660caefb91774ef10f9f5bdb3d54f1aae247e4.jpg) +Computer room + +![](images/0aab9ec8bbf4cb72f6ef7b17757ec22ea3244c74b720d4792b52481b9c52a745.jpg) +Dishes, pots, and pans + +![](images/9f6cf2dfb1e362bfd37d8436320fc7507f52ed24e340401ef7d948ccdbb17fdc.jpg) +Kitchen + +![](images/740dcf664438f4b51884a15ffec731ceefe44a0d12d670a7f1c1f461c1e7a078.jpg) +Figure 5. Qualitative results for cross-modal retrieval. Top and middle: retrieval using one single modality. Bottom: multiple modalities can be composed to retrieve targets across different modalities (§4.7). + +![](images/e525cb2760c631dd9ccba1d8d525b57a01c5934b89763f8d7a5f2d45451f4952.jpg) + +![](images/592482f04f90d5e62f8370ad6295c724381c506d5652fc7ebf90a19e5eefe589.jpg) + +![](images/22b61ebf4167432c93440d5caf117d5df85beac73bbff9d745ef0948c90a2f29.jpg) + +![](images/90e1f03ce63ec3ed24a28e06a5f6a45976f6127af758f597ac045b20c1796990.jpg) + +![](images/54d555bb69c8f39999ea89ce05cf8914c4f397bff25a6799064a35975a27b5f5.jpg) + +![](images/3dcfbc1572996109a1295e77a840ba0fcb14ac3cb863464ec96ffab282711614.jpg) + +![](images/fae369fdc26c804450f832f8eb6084c854083b3aa99ae8c01916775787211568.jpg) + +![](images/18afea5e8487fd595d186cbf68d74ed6a48c92152fa973e3588737c150eabf10.jpg) +Office + +
#StrategiesSUN-DModelNet40
1Expert Grouping45.855.9
2Global Routing45.055.3
+ +Table 8. Ablation studies on modality-aware MoE strategies. + +
Top-k#ExpertsInit.GPU MemSUN-DModelNet40
11rand27.2G45.054.8
11CLIP27.2G45.255.1
12rand27.4G45.155.2
12CLIP27.4G45.255.4
22rand35.9G45.755.7
22CLIP35.9G45.855.9
+ +expert network configurations. We investigate the impact of top- $k$ routing, the number of experts per modality, and weight initialization. Starting with a baseline configuration of one expert per modality and random initialization, we achieve 45.0 on SUN-D test [74] and 54.8 on ModelNet40 test [5], with 27.2G GPU memory consumption. Increasing the expert number and initializing weights from the visual encoder of CLIP both lead to improved performance. Due to the constraints of GPU memory, we use two experts per modality and top-2 routing for expert selection. Training Objectives. Next we probe the impact of training objectives i.e., soft modality binding with $\mathcal{L}_{\mathrm{soft\_bind}}$ (cf., Eq. 11), as well as knowledge distillation through $\mathcal{L}_{\mathrm{anch\_distil}}$ (cf., Eq. 12) and $\mathcal{L}_{\mathrm{model\_distil}}$ (cf., Eq. 15). Results are summarized in Table 10, where the first row refers to the baseline that utilizes only paired data from single datasets and is trained with the standard InfoNCE loss. It can be observed that: First, our $\mathcal{L}_{\mathrm{soft\_bind}}$ successfully bridges the learning of multiple modalities for mutual boosting, leads to improved performance on both datasets. Second, $\mathcal{L}_{\mathrm{anch\_distil}}$ and $\mathcal{L}_{\mathrm{model\_distil}}$ could further boost the performance with compelling gains. This verifies the effectiveness of our knowledge distillation strategies, and leads UNIALIGN to a high performance model with optimized efficiency. + +Modality Scaling. Finally, we examine the robustness of UNIALIGN to scale different numbers of modalities into a unified model. As shown in Table 11, modality-independent + +Table 9. Ablation studies on configuration of expert network. GPU memory is reported with a batch size of 448 on one Tesla A100. + +
L_NCELsoft_bindLanch_distilLmodel_distilSUN-DModelNet40
✓×××40.438.6
✓✓××41.840.1
✓✓✓×42.340.9
✓✓✓✓45.855.9
+ +Table 10. Ablation studies on training objectives. + +
#Pre-Training ModalityImageNet-1KSUN-DModelNet40
1I + D + P84.734.431.6
2I + T + D + P84.939.938.1
3I + T + D + P + V85.140.338.4
4I + T + D + P + V + A85.140.438.6
5I + T + D + P + V + A *85.345.855.9
+ +Table 11. Ablation studies on modality scaling. Results are reported using only $\mathcal{L}_{\mathrm{NCE}}$ as the default training loss. * denotes incorporate $\mathcal{L}_{\mathrm{soft\_bind}}$ , $\mathcal{L}_{\mathrm{anch\_distil}}$ , and $\mathcal{L}_{\mathrm{model\_distil}}$ into training. I: Image, T: Text, D: Depth, V: Video, A: Audio, P: 3D Point Cloud. + +training with merely $\mathcal{L}_{\mathrm{NCE}}$ leads to a performance degeneration as more modalities integrated (i.e., row #1-4). This suggests that a straightforward unification of multiple modalities may yield sub-optimal performance for individual ones. Nevertheless, soft modality binding and knowledge distillation (i.e., row #4 vs. #5) could mitigate this negative impact, and boost performance to a level surpassing that achieved with fewer modalities (i.e., row #1 vs. #5). + +# 5. Conclusion + +In this work, we focus on the all-in-one paradigm for multimodal alignment, which yields a highly efficient and effective model, UNIALIGN, to unify the alignment of all anticipated modalities within one network. Inspired by MoE and knowledge distillation, we propose a modality-aware MoE to tackle multimodal inputs, and generally applicable alignment strategies to utilize unpaired data across datasets. Empirical results suggest that UNIALIGN achieves remarkable performance while being efficient in both training (7.8M trainable parameters to handle 6 modalities within a single training phase) and inference (one identical model with consistent weights across tasks). This work may potentially advance the development of compact, multimodal models. + +Acknowledgement. This work was supported by the National Science and Technology Major Project (No. 2023ZD0121300), the National Natural Science Foundation of China (No. 62472222, 62302217), Natural Science Foundation of Jiangsu Province (No. BK20240080, BK20220936), "Pioneer" and "Leading Goose" R&D Program of Zhejiang (No. 2024C01161), the National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, Xi'an Jiaotong University (No. HMHAI-202403), and Bytedance Doubao Fund. + +# References + +[1] Romero Adriana, Ballas Nicolas, K Samira Ebrahimi, Chassang Antoine, Gatta Carlo, and Bengio Yoshua. Fitnets: Hints for thin deep nets. In ICLR, 2015. 3 +[2] Lucas Beyer, Xiaohua Zhai, Amélie Royer, Larisa Markeeva, Rohan Anil, and Alexander Kolesnikov. Knowledge distillation: A good teacher is patient and consistent. In CVPR, 2022. 2 +[3] Xinhao Cai, Qiqia Lai, Yuwei Wang, Wenguan Wang, Zeren Sun, and Yazhou Yao. Poly kernel inception network for remote sensing detection. In CVPR, 2024. 2 +[4] Rich Caruana. Multitask learning. Machine learning, 28: 41-75, 1997. 3 +[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5, 7, 8 +[6] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP, 2020. 5 +[7] Shaoxiang Chen, Zequn Jie, and Lin Ma. Llava-MoLE: Sparse Mixture of LoRA Experts for Mitigating Data Conflicts in Instruction Finetuning MLLMs. arXiv preprint arXiv:2401.16160, 2024.3 +[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2, 5, 6 +[9] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. In NeurIPS, 2024. 4 +[10] Xiaohan Ding, Yiyuan Zhang, Yixiao Ge, Sijie Zhao, Lin Song, Xiangyu Yue, and Ying Shan. Unireplknet: A universal perception large-kernel convnet for audio video point cloud time-series and image recognition. In CVPR, 2024. 6, 7 +[11] Shihan Dou, Enyu Zhou, Yan Liu, Songyang Gao, Jun Zhao, Wei Shen, Yuhao Zhou, Zhiheng Xi, Xiao Wang, Xiaoran Fan, et al. Loramoe: Revolutionizing mixture of experts for maintaining world knowledge in language model alignment. arXiv preprint arXiv:2312.09979, 2023. 3 +[12] Dapeng Du, Limin Wang, Huiling Wang, Kai Zhao, and Gangshan Wu. Translate-to-recognize networks for rgb-d scene recognition. In CVPR, 2019. 6 +[13] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue + +Cao. Eva: Exploring the limits of masked visual representation learning at scale. In CVPR, 2023. 2 +[14] Bilal Faye, Hanane Azzag, and Mustapha Lebbah. Oneencoder: A Lightweight Framework for Progressive Alignment of Modalities. arXiv preprint arXiv:2409.11059, 2024. 2 +[15] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1-39, 2022. 3 +[16] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In ICASSP, 2017. 2, 5, 6, 7 +[17] Rohit Girdhar, Mannat Singh, Nikhila Ravi, Laurens van der Maaten, Armand Joulin, and Ishan Misra. Omnivore: A Single Model for Many Visual Modalities. In CVPR, 2022. 2, 5, 6 +[18] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In CVPR, 2023. 1, 2, 4, 6, 7 +[19] Yuan Gong, Yu-An Chung, and James Glass. AST: Audio Spectrogram Transformer. In Interspeech, 2021. 7 +[20] Yuan Gong, Cheng-I Lai, Yu-An Chung, and James R. Glass. Ssast: Self-Supervised Audio Spectrogram Transformer. In AAAI, 2022. 7 +[21] Yunhao Gou, Zhili Liu, Kai Chen, Lanqing Hong, Hang Xu, Aoxue Li, Dit-Yan Yeung, James T Kwok, and Yu Zhang. Mixture of cluster-conditional lora experts for vision-language instruction tuning. arXiv preprint arXiv:2312.12379, 2023. 3 +[22] Andrey Guzhov, Federico Raue, Jorn Hees, and Andreas Dengel. Audioclip: Extending clip to image, text and audio. In ICASSP, 2022. 1, 5, 7 +[23] Jiaming Han, Kaixiong Gong, Yiyuan Zhang, Jiaqi Wang, Kaipeng Zhang, Dahua Lin, Yu Qiao, Peng Gao, and Xi-angyu Yue. Onellm: One Framework to Align All Modalities with Language. In CVPR, 2024. 2 +[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6 +[25] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, Nojun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In ICCV, 2019. 3 +[26] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2, 3 +[27] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2021. 2, 3 +[28] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 3, 4 + +[29] Chengsong Huang, Qian Liu, Bill Yuchen Lin, Tianyu Pang, Chao Du, and Min Lin. Lorahub: Efficient cross-task generalization via dynamic lora composition. arXiv preprint arXiv:2307.13269, 2023. 3 +[30] Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson W.H. Lau, Wanli Ouyang, and Wangmeng Zuo. Clip2point: Transfer clip to point cloud classification with image-depth pre-training. In ICCV, 2023. 6 +[31] Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. Mixtral of experts. arXiv preprint arXiv:2401.04088, 2024. 3 +[32] Byoungjip Kim, Sungik Choi, Dasol Hwang, Moontae Lee, and Honglak Lee. Transferring pre-trained multimodal representations with cross-modal similarity matching. In NeurIPS, 2022. 3 +[33] Chris Dongjoo Kim, Byeongchang Kim, Hyunmin Lee, and Gunhee Kim. Audiocaps: Generating captions for audios in the wild. In NAACL-HLT, 2019. 6, 7 +[34] Jangho Kim, SeongUk Park, and Nojun Kwak. Paraphrasing complex network: Network compression via factor transfer. In NeurIPS, 2018. 3 +[35] Aran Komatsuzaki, Joan Puigcerver, James Lee-Thorp, Carlos Riquelme Ruiz, Basil Mustafa, Joshua Ainslie, Yi Tay, Mostafa Dehghani, and Neil Houlsby. Sparse upcycling: Training mixture-of-experts from dense checkpoints. In ICLR, 2023. 3 +[36] Nikos Komodakis and Sergey Zagoruyko. Paying more attention to attention: improving the performance of convolutional neural networks via attention transfer. In ICLR, 2017. 3 +[37] Konrad P Kording, Ulrik Beierholm, Wei Ji Ma, Steven Quartz, Joshua B Tenenbaum, and Ladan Shams. Causal inference in multisensory perception. PLoS one, 2(9):e943, 2007. 1 +[38] Stan Weixian Lei, Yixiao Ge, Kun Yi, Jianfeng Zhang, Difei Gao, Dylan Sun, Yuying Ge, Ying Shan, and Mike Zheng Shou. Vit-Lens: Towards Omni-modal Representations. In CVPR, 2024. 2, 5, 6, 7 +[39] Weixian Lei, Yixiao Ge, Kun Yi, Jianfeng Zhang, Difei Gao, Dylan Sun, Yuying Ge, Ying Shan, and Mike Zheng Shou. Vit-lens: Towards omni-modal representations. In CVPR, 2024. 1, 2, 3 +[40] Po-han Li, Sandeep P Chinchali, and Ufuk Topcu. Csa: Data-efficient mapping of unimodal features to multimodal features. arXiv preprint arXiv:2410.07610, 2024. 2 +[41] Yunxin Li, Shenyuan Jiang, Baotian Hu, Longyue Wang, Wanqi Zhong, Wenhan Luo, Lin Ma, and Min Zhang. UniMoE: Scaling Unified Multimodal LLMs with Mixture of Experts. arXiv, abs/2405.11273, 2024. 3 +[42] Valerii Likhosherstov, Anurag Arnab, Krzysztof Choromanski, Mario Lucic, Yi Tay, Adrian Weller, and Mostafa Dehghani. Polyvit: Co-training vision transformers on images, videos and audio. arXiv preprint arXiv:2111.12993, 2021. 3, 5 +[43] Bin Lin, Zhenyu Tang, Yang Ye, Jiaxi Cui, Bin Zhu, Peng Jin, Jina Huang, Junwu Zhang, Yatian Pang, Munan Ning, + +et al. Moe-llava: Mixture of experts for large vision-language models. arXiv preprint arXiv:2401.15947, 2024. 3 +[44] Chung-Ching Lin, Kevin Lin, Lijuan Wang, Zicheng Liu, and Linjie Li. Cross-modal representation learning for zero-shot action recognition. In CVPR, 2022. 7 +[45] Haokun Lin, Haoli Bai, Zhili Liu, Lu Hou, Muyi Sun, Linqi Song, Ying Wei, and Zhenan Sun. Mope-CLIP: Structured Pruning for Efficient Vision-Language Models with Module-wise Pruning Error Metric. In CVPR, 2024. 3 +[46] Xi Victoria Lin, Akshit Shrivastava, Liang Luo, Srinivasan Iyer, Mike Lewis, Gargi Gosh, Luke Zettlemoyer, and Armen Aghajanyan. Moma: Efficient early-fusion pretraining with mixture of modality-aware experts. arXiv preprint arXiv:2407.21770, 2024. 3 +[47] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 6 +[48] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022. 6 +[49] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 2 +[50] Yuanhuiyi Lyu, Xu Zheng, Jiazhou Zhou, and Lin Wang. Unibind: Llm-Augmented Unified and Balanced Representation Space to Bind Them All. In CVPR, 2024. 2, 7 +[51] Jian Ma, Wenguan Wang, Yi Yang, and Feng Zheng. Ms2sl: Multimodal spoken data-driven continuous sign language production. In ACL Findings, 2024. 2 +[52] Wenxuan Ma, Shuang Li, Jinming Zhang, Chi Harold Liu, Jingxuan Kang, Yulin Wang, and Gao Huang. Borrowing Knowledge From Pre-trained Language Model: A New Data-efficient Visual Learning Paradigm. In ICCV, 2023. 2 +[53] M Alex Meredith and Barry E Stein. Visual, auditory, and somatosensory convergence on cells in superior colliculus results in multisensory integration. Journal of neurophysiology, 56(3):640-662, 1986. 1 +[54] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, 2019. 7 +[55] Seyed Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved knowledge distillation via teacher assistant. In AAAI, 2020. 3 +[56] Luca Moschella, Valentino Maiorca, Marco Fumero, Antonio Norelli, Francesco Locatello, and Emanuele Rodola. Relative representations enable zero-shot latent space communication. In ICLR, 2023. 2 +[57] Arsha Nagrani, Paul Hongsuck Seo, Bryan Seybold, Anja Hauth, Santiago Manen, Chen Sun, and Cordelia Schmid. Learning audio-video modalities from image captions. In ECCV, 2022. 7 + +[58] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. In European Conference on Computer Vision, pages 1-18. Springer, 2022. 7 +[59] Antonio Norelli, Marco Fumero, Valentino Maiorca, Luca Moschella, Emanuele Rodola, and Francesco Locatello. Asif: Coupled data turns unimodal models to multimodal without training. In NeurIPS, 2023. 2 +[60] Mandela Patrick, Po-Yao Huang, Yuki Asano, Florian Metze, Alexander Hauptmann, Joao Henriques, and Andrea Vedaldi. Support-set bottlenecks for video-text representation learning. arXiv preprint arXiv:2010.02824, 2020. 7 +[61] Gensheng Pei, Fumin Shen, Yazhou Yao, Guo-Sen Xie, Zhenmin Tang, and Jinhui Tang. Hierarchical feature alignment network for unsupervised video object segmentation. In ECCV, 2022. 2 +[62] Gensheng Pei, Tao Chen, Xiruo Jiang, Huafeng Liu, Zeren Sun, and Yazhou Yao. Videomac: Video masked autoencoders meet convnets. In CVPR, 2024. 2 +[63] Renjing Pei, Jianzhuang Liu, Weimian Li, Bin Shao, Songcen Xu, Peng Dai, Juwei Lu, and Youliang Yan. Clipping: Distilling clip-based models with a student base for video-language retrieval. In CVPR, 2023. 3 +[64] Zhiliang Peng, Wei Huang, Shanzhi Gu, Lingxi Xie, Yaowei Wang, Jianbin Jiao, and Qixiang Ye. Conformer: Local features coupling global representations for visual recognition. In ICCV, 2021. 6 +[65] Karol J Piczak. Esc: Dataset for environmental sound classification. In ACM MM, 2015. 6, 7 +[66] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with Reconstruct: Contrastive 3d Representation Learning Guided by Generative Pretraining. In ICML, 2023. 6 +[67] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 5 +[68] Mike Ranzinger, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. Am-RADIO: Agglomerative Vision Foundation Model Reduce All Domains Into One. In CVPR, 2024. 3, 5 +[69] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. In NeurIPS, 2021. 3 +[70] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. In ICML, 2023. 6 +[71] Jinghuan Shang, Karl Schmeckpeper, Brandon B. May, Maria Vittoria Minniti, Tarik Kelestemur, David Watkins, and Laura Herlant. Theia: Distilling Diverse Vision Foundation Models for Robot Learning. In CoRL, 2024. 3 +[72] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. + +Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In ICLR, 2017. 2, 3 +[73] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 6 +[74] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In CVPR, 2015. 5, 6, 7, 8 +[75] Xinhang Song, Luis Herranz, and Shuqiang Jiang. Depth cnns for rgb-d scene recognition: Learning from scratch better than transferring from rgb-cnns. In AAAI, 2017. 6 +[76] Xinhang Song, Shuqiang Jiang, Bohan Wang, Chengpeng Chen, and Gongwei Chen. Image representations with spatial object-to-object relations for rgb-d scene recognition. IEEE TIP, 29:525-537, 2019. 6 +[77] K Soomro. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 7 +[78] Siddharth Srivastava and Gaurav Sharma. Omnivec2 - A Novel Transformer based Network for Large Scale Multimodal and Multitask Learning. In CVPR, 2024. 2 +[79] Ximeng Sun, Pengchuan Zhang, Peizhao Zhang, Hardik Shah, Kate Saenko, and Xide Xia. Dime-fm: Distilling multimodal and efficient foundation models. In ICCV, 2023. 3 +[80] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. In ECCV, 2022. 6 +[81] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 4 +[82] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 6 +[83] Simon Vandenhende, Stamatios Georgoulis, Wouter Van Gansbeke, Marc Proesmans, Dengxin Dai, and Luc Van Gool. Multi-task learning for dense prediction tasks: A survey. IEEE TPAMI, 44(7):3614-3633, 2021. 3 +[84] Haoxiang Wang, Pavan Kumar Anasosalu Vasu, Fartash Faghri, Raviteja Vemulapalli, Mehrdad Farajtabar, Sachin Mehta, Mohammad Rastegari, Oncel Tuzel, and Hadi Pouransari. Sam-clip: Merging vision foundation models towards semantic and spatial understanding. In CVPR, 2024. 3 +[85] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. 7 +[86] Peng Wang, Shijie Wang, Junyang Lin, Shuai Bai, Xiaohuan Zhou, Jingren Zhou, Xinggang Wang, and Chang Zhou. One-peace: Exploring one general representation model toward unlimited modalities. arXiv preprint arXiv:2305.11172, 2023. 5 +[87] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, + +Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. In CVPR, 2023. 6 +[88] Wenguan Wang, Yi Yang, and Yunhe Pan. Visual knowledge in the big model era: Retrospect and prospect. Frontiers of Information Technology & Electronic Engineering, 26(1):1-19, 2025. 2 +[89] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 7 +[90] Zehan Wang, Ziang Zhang, Luping Liu, Yang Zhao, Haifeng Huang, Tao Jin, and Zhou Zhao. Extending multi-modal contrastive representations. arXiv preprint arXiv:2310.08884, 2023. 2 +[91] Zehan Wang, Yang Zhao, Haifeng Huang, Jiageng Liu, Aoxiong Yin, Li Tang, Linjun Li, Yongqi Wang, Zhang, and Zhou Zhao. Connecting multi-modal contrastive representations. In NeurIPS, 2023. 2 +[92] Pete Warden. Speech commands: A dataset for limited-vocabulary speech recognition. arXiv preprint arXiv:1804.03209, 2018.7 +[93] Yixuan Wei, Han Hu, Zhenda Xie, Zheng Zhang, Yue Cao, Jianmin Bao, Dong Chen, and Baining Guo. Contrastive learning rivals masked image modeling in fine-tuning via feature distillation. Tech Report, 2022. 3 +[94] Kan Wu, Jinnian Zhang, Houwen Peng, Mengchen Liu, Bin Xiao, Jianlong Fu, and Lu Yuan. Tinyvit: Fast Pretraining Distillation for Small Vision Transformers. In ECCV, 2022. 6 +[95] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, 2015. 2, 6 +[96] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. 2, 7 +[97] Peng Xu, Xiatian Zhu, and David A Clifton. Multimodal learning with transformers: A survey. IEEE TPAMI, 45 (10):12113-12132, 2023. 1 +[98] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In CVPR, 2023. 2, 5, 6 +[99] Chuanguang Yang, Zhulin An, Libo Huang, Junyu Bi, Xinqiang Yu, Han Yang, Boyu Diao, and Yongjun Xu. ClipKD: An Empirical Study of CLIP Model Distillation. In CVPR, 2024. 3, 5 +[100] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models (exemplified as a video agent). In ICML, 2024. 2 +[101] Yazhou Yao, Tao Chen, Guo-Sen Xie, Chuanyi Zhang, Fumin Shen, Qi Wu, Zhenmin Tang, and Jian Zhang. Nonsalient region object mining for weakly supervised semantic segmentation. In CVPR, 2021. 2 + +[102] Yazhou Yao, Zeren Sun, Chuanyi Zhang, Fumin Shen, Qi Wu, Jian Zhang, and Zhenmin Tang. Jo-src: A contrastive approach for combating noisy labels. In CVPR, 2021. 2 +[103] Junbo Yin, Jianbing Shen, Runnan Chen, Wei Li, Ruigang Yang, Pascal Frossard, and Wenguan Wang. Is-fusion: Instance-scene collaborative fusion for multimodal 3d object detection. In ECCV, 2024. 2 +[104] Shan You, Chang Xu, Chao Xu, and Dacheng Tao. Learning from multiple teacher networks. In KDD, 2017. 3 +[105] Fei Yuan, Linjun Shou, Jian Pei, Wutao Lin, Ming Gong, Yan Fu, and Daxin Jiang. Reinforced multi-teacher selection for knowledge distillation. In AAAI, 2021. 3 +[106] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xu-peng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In CVPR, 2022. 6 +[107] Yiyuan Zhang, Kaixiong Gong, Kaipeng Zhang, Hongsheng Li, Yu Qiao, Wanli Ouyang, and Xiangyu Yue. Meta-transformer: A unified framework for multimodal learning. arXiv preprint arXiv:2307.10802, 2023. 2, 3, 6, 7 +[108] Yanqi Zhou, Tao Lei, Hanxiao Liu, Nan Du, Yanping Huang, Vincent Zhao, Andrew M Dai, Quoc V Le, James Laudon, et al. Mixture-of-experts with expert choice routing. In NeurIPS, 2022. 2 +[109] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, WANG HongFa, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei Li, Cai Wan Zhang, Zhifeng Li, Wei Liu, and Li Yuan. Languagebind: Extending Video-Language Pretraining to N-modality by Language-based Semantic Alignment. In ICLR, 2024. 1, 2, 7 +[110] Xiatian Zhu, Shaogang Gong, et al. Knowledge distillation by on-the-fly native ensemble. In NeurIPS, 2018. 2 +[111] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In ICCV, 2023. 6 +[112] Konrad Zuchniak. Multi-teacher knowledge distillation as an effective method for compressing ensembles of neural networks. arXiv preprint arXiv:2302.07215, 2023. 3 \ No newline at end of file diff --git a/2025/UNIALIGN_ Scaling Multimodal Alignment within One Unified Model/images.zip b/2025/UNIALIGN_ Scaling Multimodal Alignment within One Unified Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5207becd981666d5206282fa32389791488bb3d5 --- /dev/null +++ b/2025/UNIALIGN_ Scaling Multimodal Alignment within One Unified Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c82e7f6c510468eaced3e5408e4e2c0f458dd826fb41c0a9f9bea3650877d055 +size 745142 diff --git a/2025/UNIALIGN_ Scaling Multimodal Alignment within One Unified Model/layout.json b/2025/UNIALIGN_ Scaling Multimodal Alignment within One Unified Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..feb6b38687c28d2307c0f32eeb5d4ed97dc9e12c --- /dev/null +++ b/2025/UNIALIGN_ Scaling Multimodal Alignment within One Unified Model/layout.json @@ -0,0 +1,13765 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 103, + 522, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 103, + 522, + 120 + ], + "spans": [ + { + "bbox": [ + 89, + 103, + 522, + 120 + ], + "type": "text", + "content": "UNIALIGN: Scaling Multimodal Alignment within One Unified Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 142, + 515, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 142, + 515, + 159 + ], + "spans": [ + { + "bbox": [ + 93, + 142, + 515, + 159 + ], + "type": "text", + "content": "Bo Zhou" + }, + { + "bbox": [ + 93, + 142, + 515, + 159 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 93, + 142, + 515, + 159 + ], + "type": "text", + "content": ", Lilei Li" + }, + { + "bbox": [ + 93, + 142, + 515, + 159 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 93, + 142, + 515, + 159 + ], + "type": "text", + "content": ", Yujia Wang" + }, + { + "bbox": [ + 93, + 142, + 515, + 159 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 93, + 142, + 515, + 159 + ], + "type": "text", + "content": ", Huafeng Liu" + }, + { + "bbox": [ + 93, + 142, + 515, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 93, + 142, + 515, + 159 + ], + "type": "text", + "content": ", Yazhou Yao" + }, + { + "bbox": [ + 93, + 142, + 515, + 159 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 93, + 142, + 515, + 159 + ], + "type": "text", + "content": ", Wenguan Wang" + }, + { + "bbox": [ + 93, + 142, + 515, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 159, + 492, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 159, + 492, + 171 + ], + "spans": [ + { + "bbox": [ + 117, + 159, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 117, + 159, + 492, + 171 + ], + "type": "text", + "content": " Nanjing University of Science and Technology, " + }, + { + "bbox": [ + 117, + 159, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 117, + 159, + 492, + 171 + ], + "type": "text", + "content": " Zhejiang University, " + }, + { + "bbox": [ + 117, + 159, + 492, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 117, + 159, + 492, + 171 + ], + "type": "text", + "content": " Zhejiang Sci-Tech University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 129, + 174, + 478, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 174, + 478, + 185 + ], + "spans": [ + { + "bbox": [ + 129, + 174, + 478, + 185 + ], + "type": "text", + "content": "https://github.com/NUST-Machine-Intelligence-Laboratory/UNIALIGN" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 238, + 296, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 238, + 296, + 538 + ], + "spans": [ + { + "bbox": [ + 54, + 238, + 296, + 538 + ], + "type": "text", + "content": "We present UNIALIGN, a unified model to align an arbitrary number of modalities (e.g., image, text, audio, 3D point cloud, etc.) through one encoder and a single training phase. Existing solutions typically employ distinct encoders for each modality, resulting in increased parameters as the number of modalities grows. In contrast, UNIALIGN proposes a modality-aware adaptation of the powerful mixture-of-experts (MoE) schema and further integrates it with Low-Rank Adaptation (LoRA), efficiently scaling the encoder to accommodate inputs in diverse modalities while maintaining a fixed computational overhead. Moreover, prior work often requires separate training for each extended modality. This leads to task-specific models and further hinders the communication between modalities. To address this, we propose a soft modality binding strategy that aligns all modalities using unpaired data samples across datasets. Two additional training objectives are introduced to distill knowledge from well-aligned anchor modalities and prior multimodal models, elevating UNIALIGN into a high performance multimodal foundation model. Experiments on 11 benchmarks across 6 different modalities demonstrate that UNIALIGN could achieve comparable performance to SOTA approaches, while using merely 7.8M trainable parameters and maintaining an identical model with the same weight across all tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 559, + 135, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 135, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 135, + 571 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 578, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 296, + 687 + ], + "type": "text", + "content": "Humans accomplish tasks by processing and combining signals from different modalities [37, 53]. This gives rise to multimodal learning, which aims to build AI models that can extract and relate information across multimodal inputs [97]. Recent efforts have been made to learn a unified representation space via the alignment between modalities [18, 22, 39]. Though enabling extension to novel modalities, existing solutions typically rely on a one-versus-one paradigm where merely one extended modality (e.g., point" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 213, + 554, + 405 + ], + "blocks": [ + { + "bbox": [ + 317, + 213, + 554, + 405 + ], + "lines": [ + { + "bbox": [ + 317, + 213, + 554, + 405 + ], + "spans": [ + { + "bbox": [ + 317, + 213, + 554, + 405 + ], + "type": "image", + "image_path": "b45890f026d892961b336e635b5a3c6d4df200ee34dac41aa598ba9183b70cc6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 407, + 555, + 462 + ], + "lines": [ + { + "bbox": [ + 313, + 407, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 555, + 462 + ], + "type": "text", + "content": "Figure 1. Comparison of architectures between standard multi-modal alignment methods [18, 39, 109] and UNIALIGN: (a) Existing work adopts specialized encoders tailored to each modality; (b) UNIALIGN unifies the alignment across multiple modalities in a single model, with one training phase across all datasets." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 472, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 472, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 312, + 472, + 555, + 628 + ], + "type": "text", + "content": "cloud, audio) could be aligned with the anchor modalities (e.g., text, image) at each time (Fig. 1(a)). This presents several drawbacks: i) one dedicated encoder is required for each modality, leading to exponential increases in training cost and parameter count as the number of aligned modalities grows, severely limiting scalability; ii) task-specific training for each dataset, results in multiple independent models for different tasks; and iii) the separate alignment of each extended modality hinders the model to leverage the full breadth of data across all datasets to shape the shared embedding space from a global view. This further impedes cross-modal knowledge transfer where insights gained from one modality could benefit understanding in others." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": "These observations prompt us to rethink: Is this one-versus-one alignment paradigm essential? and Can we achieve joint alignment of all modalities within one single model? In response, we propose UNIALIGN (Fig. 1(b)), which embraces a all-in-one design to realize multimodal alignment by constructing a foundation model that includes: 1 one base encoder to UNify the encoding of all anchor and" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 240, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 240, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 240, + 703 + ], + "type": "text", + "content": "*The first two authors contribute equally to this work." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 67, + 703, + 187, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 187, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 187, + 712 + ], + "type": "text", + "content": "† Corresponding author: Yazhou Yao." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29644" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "extended modalities, and " + }, + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "inline_equation", + "content": "②" + }, + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": " one training phase to ALIGN all modalities with all samples across datasets collectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 99, + 297, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 99, + 297, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 99, + 297, + 567 + ], + "type": "text", + "content": "Concretely, to achieve ①, recent work has consistently demonstrated that Transformers pre-trained on one modality possess exceptional generalization ability to novel modalities [13, 67]. Moreover, the mixture-of-experts (MoE) [72] schema presents a highly efficient strategy to scale model capacity under a fixed computational cost [108]. These insights motivate us to build a foundational model accommodating multiple modalities upon a single pre-trained Transformer (e.g., CLIP [67]). Here, MoE can dynamically scale and allocate model capacities by selectively activating experts to encode inputs across diverse modalities. We explore two strategies targeting at expert grouping and routing mechanism to enhance the awareness of different modalities within MoE. To preserve the knowledge learned from anchor modalities, Low-Rank Adaptation (LoRA) [27] is integrated with MoE, enabling a significant reduction in training overhead. To fulfill ②, we propose a soft modality binding strategy, leveraging massive unpaired data samples across modalities and datasets. Specifically, given unpaired data from diverse datasets with different modalities, the class labels of samples are encoded into text embeddings to create a unified representation. The semantic similarity between text embeddings can inform the alignment between modalities (e.g., audio and video), by encouraging the similarity between modality embeddings and textual label embeddings (e.g., audio " + }, + { + "bbox": [ + 56, + 99, + 297, + 567 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 56, + 99, + 297, + 567 + ], + "type": "text", + "content": " textvideo and video " + }, + { + "bbox": [ + 56, + 99, + 297, + 567 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 56, + 99, + 297, + 567 + ], + "type": "text", + "content": " textaudio) to be identical as the semantic similarity (i.e., " + }, + { + "bbox": [ + 56, + 99, + 297, + 567 + ], + "type": "inline_equation", + "content": "text_{\\mathrm{video}} \\leftrightarrow \\text{text}_{\\mathrm{audio}}" + }, + { + "bbox": [ + 56, + 99, + 297, + 567 + ], + "type": "text", + "content": "). Moreover, rooted in the principles of knowledge distillation [2, 26, 110], the alignment between extended and anchor modalities is guided by the knowledge persisting between anchor modalities, which are well-aligned in semantics compared to extended modalities. To further boost the performance, we transfer knowledge from multiple well-trained foundation models which only contain partial modalities. Given the complexity of multimodal, multi-dataset joint training, a specialized batching strategy combined with modality-aware gradient accumulation is proposed, to mitigate the potential dispersion in optimization across different modalities." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "UNIALIGN enjoys several advantages: First, it supports flexible alignment between arbitrary number of anchor and extended modalities, while ensuring negligible increase in parameter size. Second, the alignment for all modalities is accomplished in one training session, significantly improving training efficiency. Third, unpaired data from different modalities can be used collectively for optimization. Uncommon modalities with scarce data could take full advantages of the rich resource of common modalities. Fourth, benefited from soft modality binding, extended modalities are bridged for mutual boosting. Fifth, a compact and high-performance multimodal foundational" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": "model generally applicable to all tasks is delivered at minimal cost. This contrasts with existing approaches, which typically provide multiple loosely integrated modality encoders tailored to specific tasks, thereby enabling broader transferability to other domains and downstream applications [3, 8, 16, 61, 62, 95, 96, 101, 102]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 144, + 556, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 556, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 556, + 277 + ], + "type": "text", + "content": "By embracing the all-in-one philosophy, UNIALIGN demonstrates remarkable efficiency in both training cost and parameter size. After a single pre-training step without further tuning, UNIALIGN achieves top-leading performance on 11 benchmarks with one identical model and 7.8M trainable parameters across 6 modalities (previous SOTA [18, 39] needing over 80M trainable parameters for each extended modality). This evidences the great potential of our unified modality encoding and aligning strategies. We hope this work could stimulate the community to rethink multimodal alignment from a universal and compact perspective." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 286, + 401, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 286, + 401, + 300 + ], + "spans": [ + { + "bbox": [ + 314, + 286, + 401, + 300 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 307, + 556, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 556, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 556, + 641 + ], + "type": "text", + "content": "Multimodal Learning. Multimodal learning has attracted significant attention given its ability to address a wide range of applications by integrating information from diverse modalities such as text, images, audio, and video[49, 51, 67, 103]. Current research in this field generally falls into three categories: First, methods involving joint training of multiple modalities aim to improve performance across different downstream tasks, by leveraging shared representations[17, 78, 107]. Though incorporating multiple modalities as inputs, these work lacks explicit alignment between modalities. Second, inspired by the successes of vision-language alignment [67], some approaches extend this alignment to additional modalities[18, 38, 40, 56, 59, 91, 98, 109]. These methods typically require an independent encoder for each modalities. Though [14, 18, 39, 90] maintain an consistent architecture across tasks, they conduct separate training on individual datasets, resulting in multiple task-specific models. Additionally, these solutions often rely on massive paired data to achieve robust zero-shot performance across various tasks, ignoring the usage of rich unpaired data. Third, the integration of multimodal inputs with large language models (LLMs) has opened new avenues for multimodal learning[23, 50, 52, 88, 100]. In these approaches, all modalities are projected into the same representation space as LLMs, without alignment among extended modalities and supporting merely one anchor modality (i.e., text). This leads to suboptimal performance on certain vision tasks and limits the applications primarily to language-related fields." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "In this work, we propose UNIALIGN, supporting alignment between an arbitrary number of anchor and extended modalities within one model using one unified modality encoder. After a single training phase, UNIALIGN can handle a dozen of downstream tasks without any fine-tuning. UNIALIGN also achieves high training efficiency, utilizing" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "29645" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 241, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 241, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 241, + 83 + ], + "type": "text", + "content": "merely 7.8M training parameters for all tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 84, + 296, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 84, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 84, + 296, + 323 + ], + "type": "text", + "content": "Multi-Task Learning. To maximize the utilization of resource and improve generalization across tasks, multi-task learning seeks to develop models that can handle multiple tasks simultaneously [4]. Traditional methods often require separate subnetworks for each task, which hinders scalability [83]. In recent years, the mixture-of-experts (MoE) architecture [15, 72] has gained prominence in multi-task learning [31]. MoE leverages multiple subnetworks, or \"experts\", that are selectively activated to optimize task performance while reducing computational costs [35]. This approach proves beneficial across various domains, such as language processing [31, 35] and vision tasks [41, 43, 69]. With the success of LoRA [27], there has been a surge to integrate MoE with LoRA, which shows promise in enhancing both computational efficiency and task adaptability in large-scale models [11, 21, 28, 29]. However, these methods are mainly confined to text and image domains [7, 11], and lack alignment between modalities [46]. In contrast, UNIALIGN extends MoE to a multimodal setup, devising modality-aware expert selection and routing mechanism." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 323, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 491 + ], + "type": "text", + "content": "Knowledge Distillation. Knowledge distillation [26, 34, 55] is an essential technique for model compression, enabling the transfer of knowledge from large, computationally intensive models to smaller, more efficient ones. This is particularly valuable for multimodal foundation models, where the computational overload is heavy [32, 45, 79, 99]. Initially, knowledge distillation focused on extracting knowledge from individual large models, typically using the logits or intermediate feature representations to guide the training of student models [1, 25, 36, 63, 93]. Recent research has expanded this concept to integrate capabilities from multiple models, creating hybrid models that not only improve task-specific performance but also broaden the range of potential applications [68, 71, 84, 104, 105, 112]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 491, + 296, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 296, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 296, + 562 + ], + "type": "text", + "content": "Compared to existing work, UNIALIGN distills knowledge from both well-aligned anchor modalities, and multiple well-trained foundation models specialized in partial modalities. This allows our method to handle a broader range of tasks that require understanding and processing information with different composition of diverse modalities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 573, + 111, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 573, + 111, + 585 + ], + "spans": [ + { + "bbox": [ + 55, + 573, + 111, + 585 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 593, + 296, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 296, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 296, + 652 + ], + "type": "text", + "content": "In this section, we first introduce the unified encoding of all modalities within one single model (§3.1), then describe the universal modality alignment strategy across all datasets (§3.2), and finally present the implementation details of UNIALIGN (§3.3)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 660, + 205, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 205, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 205, + 673 + ], + "type": "text", + "content": "3.1. Unified Modality Encoding" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Preliminary: Mixture-of-Experts (MoE). Let " + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{L \\times D}" + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": " denote the tokenized input, where " + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": " is the sequence length and " + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": " is the token embedding size. In vanilla Transformer," + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 330, + 67, + 541, + 312 + ], + "blocks": [ + { + "bbox": [ + 330, + 67, + 541, + 312 + ], + "lines": [ + { + "bbox": [ + 330, + 67, + 541, + 312 + ], + "spans": [ + { + "bbox": [ + 330, + 67, + 541, + 312 + ], + "type": "image", + "image_path": "465c5463cfdf1aa2bb4fd9c957df010e25ee5aad93f5e5fbfa454961685ee04d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 313, + 534, + 324 + ], + "lines": [ + { + "bbox": [ + 334, + 313, + 534, + 324 + ], + "spans": [ + { + "bbox": [ + 334, + 313, + 534, + 324 + ], + "type": "text", + "content": "Figure 2. Architecture of the unified modality encoder." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 329, + 553, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 553, + 352 + ], + "type": "text", + "content": "each encoder block comprises a multi-head self-attention (MSA) layer and a feed-forward network (FFN), as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 375, + 356, + 553, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 356, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 375, + 356, + 553, + 369 + ], + "type": "interline_equation", + "content": "\\boldsymbol {X} ^ {\\prime} = \\boldsymbol {X} + \\mathcal {F} ^ {\\mathrm {M S A}} (\\ln (\\boldsymbol {X})), \\tag {1}", + "image_path": "bf3a43132173f74de878f51eef9bc1cb69b5e9e0f6a9d96488fca1563244e939.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 373, + 372, + 553, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 372, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 373, + 372, + 553, + 386 + ], + "type": "interline_equation", + "content": "\\boldsymbol {X} ^ {\\prime \\prime} = \\boldsymbol {X} ^ {\\prime} + \\mathcal {F} ^ {\\mathrm {F F N}} (\\ln (\\boldsymbol {X} ^ {\\prime})), \\tag {2}", + "image_path": "18702bbe85806f12936cc9260a4d428a41d577cff7069092a59fd98784761f26.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 394, + 553, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 553, + 430 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 394, + 553, + 430 + ], + "type": "inline_equation", + "content": "X''" + }, + { + "bbox": [ + 313, + 394, + 553, + 430 + ], + "type": "text", + "content": " is the outputs and LN represents layer normalization. In the mixture-of-experts (MoE) schema, an MoE function (i.e., " + }, + { + "bbox": [ + 313, + 394, + 553, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{\\mathrm{MoE}}" + }, + { + "bbox": [ + 313, + 394, + 553, + 430 + ], + "type": "text", + "content": ") is adopted to replace " + }, + { + "bbox": [ + 313, + 394, + 553, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{\\mathrm{FFN}}" + }, + { + "bbox": [ + 313, + 394, + 553, + 430 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 369, + 432, + 553, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 432, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 369, + 432, + 553, + 445 + ], + "type": "interline_equation", + "content": "\\boldsymbol {X} ^ {\\prime \\prime} = \\boldsymbol {X} ^ {\\prime} + \\mathcal {F} ^ {\\mathrm {M o E}} (\\ln \\left(\\boldsymbol {X} ^ {\\prime}\\right)), \\tag {3}", + "image_path": "f11fefd7c83407136ec576a9ca242e187d313afc1c767d877666857691ebac32.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 348, + 449, + 553, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 449, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 348, + 449, + 553, + 464 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\mathrm {M o E}} \\left(\\boldsymbol {X} ^ {\\prime}\\right)) = \\sum_ {i = 1} ^ {n} \\mathcal {G} \\left(\\boldsymbol {X} ^ {\\prime}\\right) _ {i} \\mathcal {F} _ {i} ^ {\\text {E x p e r t}} \\left(\\boldsymbol {X} ^ {\\prime}\\right), \\tag {4}", + "image_path": "fef3c8326cf7ffaaa0145f0ce26853987f17f4f48eb125df6d78dc845814979e.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 471, + 553, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 553, + 518 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 471, + 553, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{\\mathrm{MoE}}" + }, + { + "bbox": [ + 313, + 471, + 553, + 518 + ], + "type": "text", + "content": " jointly aggregates outputs from multiple expert networks (i.e., " + }, + { + "bbox": [ + 313, + 471, + 553, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{\\mathrm{Expert}}" + }, + { + "bbox": [ + 313, + 471, + 553, + 518 + ], + "type": "text", + "content": ") according to the router " + }, + { + "bbox": [ + 313, + 471, + 553, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(X')_i = \\text{softmax}(\\mathcal{F}^{\\mathrm{Router}}(X'))_i" + }, + { + "bbox": [ + 313, + 471, + 553, + 518 + ], + "type": "text", + "content": " to assign weights for different experts and render the predictions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 519, + 554, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 554, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 554, + 626 + ], + "type": "text", + "content": "Multimodal Adaptation with MoE. We aim to unify the encoding of multimodal inputs within one foundation model (e.g., the image encoder of CLIP[67]). As illustrated in Fig.2, inputs of various modalities are first converted into sequential tokens. We follow [39, 42, 107] to preprocess multimodal data and implement modality-specific tokenizers, with details summarized in Appendix. To adapt the encoding of samples from various modalities, UNIALIGN explores two modality-aware MoE strategies:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 626, + 553, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 626, + 553, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 626, + 553, + 662 + ], + "type": "text", + "content": "- Modality-Aware Expert Grouping divides all " + }, + { + "bbox": [ + 313, + 626, + 553, + 662 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{\\mathrm{Expert}}" + }, + { + "bbox": [ + 313, + 626, + 553, + 662 + ], + "type": "text", + "content": " in one MoE layer into multiple sets, where each set " + }, + { + "bbox": [ + 313, + 626, + 553, + 662 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 626, + 553, + 662 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 313, + 626, + 553, + 662 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 626, + 553, + 662 + ], + "type": "text", + "content": " experts to process inputs from a specific modality " + }, + { + "bbox": [ + 313, + 626, + 553, + 662 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 626, + 553, + 662 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 372, + 666, + 553, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 666, + 553, + 683 + ], + "spans": [ + { + "bbox": [ + 372, + 666, + 553, + 683 + ], + "type": "interline_equation", + "content": "\\mathcal {B} _ {\\mathcal {M}} = \\left\\{\\mathcal {F} _ {\\mathcal {M}, 1} ^ {\\text {E x p e r t}}, \\dots , \\mathcal {F} _ {\\mathcal {M}, n} ^ {\\text {E x p e r t}} \\right\\}. \\tag {5}", + "image_path": "e6567bdc8622a941134dfb03cf0aa29e090ea8f112d7c38000105f296da117e4.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Under this setup, following [46], we construct a hierarchical routing mechanism. In the first routing round, tokens are" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29646" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "text", + "content": "directed to a specific expert set based on the modality of inputs. Next, within a expert set, a learnable router " + }, + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathcal{M}}^{\\text{Router}}" + }, + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "text", + "content": " for modality " + }, + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "text", + "content": " assigns weights to each individual expert conditioned on the input tokens. This separation lets each expert set to specialize in one modality, and allows for easy expansion to novel modalities by loading new expert sets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 295, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 193 + ], + "type": "text", + "content": "- Modality-Aware Global Routing uses multiple modality-specific routers, to direct tokens across all experts in a MoE layer without grouping. Given the routing function " + }, + { + "bbox": [ + 55, + 144, + 295, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathcal{M}}^{\\text{Router}}" + }, + { + "bbox": [ + 55, + 144, + 295, + 193 + ], + "type": "text", + "content": " for modality " + }, + { + "bbox": [ + 55, + 144, + 295, + 193 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 144, + 295, + 193 + ], + "type": "text", + "content": ", the global weight " + }, + { + "bbox": [ + 55, + 144, + 295, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathcal{M}}" + }, + { + "bbox": [ + 55, + 144, + 295, + 193 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 100, + 199, + 295, + 214 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 199, + 295, + 214 + ], + "spans": [ + { + "bbox": [ + 100, + 199, + 295, + 214 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {\\mathcal {M}} \\left(\\boldsymbol {X} ^ {\\prime}\\right) _ {i} = \\operatorname {s o f t m a x} \\left(\\mathcal {F} _ {\\mathcal {M}} ^ {\\text {R o u t e r}} \\left(\\boldsymbol {X} ^ {\\prime}\\right)\\right) _ {i}. \\tag {6}", + "image_path": "928bd424a5ab34e9979411e5ffa4c189b3f1235634a6894dd80b62f38322a8b6.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 223, + 295, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 223, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 295, + 282 + ], + "type": "text", + "content": "By allowing the experts to be accessible to each router, this strategy enables cross-modality knowledge sharing among experts. However, since the experts are not explicitly specialized for one modality, the generalization potentially at the expense of fine-grained modality specialization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 283, + 296, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 296, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 296, + 366 + ], + "type": "text", + "content": "Modality-Aware LoRA Expert. Low-Rank Adaptation (LoRA) [28] has demonstrated both effectiveness and efficiency in fine-tuning large models [9, 81]. Therefore, to maintain the knowledge in pre-trained foundation models, the original feed-forward network (i.e., " + }, + { + "bbox": [ + 55, + 283, + 296, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{\\mathrm{FFN}}" + }, + { + "bbox": [ + 55, + 283, + 296, + 366 + ], + "type": "text", + "content": " in Eq. 2) are preserved in a fixed state instead of being replaced during the forward of MoE (i.e., Eq. 3), resulting in:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 373, + 295, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 373, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 70, + 373, + 295, + 388 + ], + "type": "interline_equation", + "content": "\\boldsymbol {X} ^ {\\prime \\prime} = \\boldsymbol {X} ^ {\\prime} + \\mathcal {F} ^ {\\mathrm {F F N}} (\\ln (\\boldsymbol {X} ^ {\\prime})) + \\mathcal {F} ^ {\\mathrm {M o E}} (\\ln (\\boldsymbol {X} ^ {\\prime})). \\tag {7}", + "image_path": "d71979c4dad90af7ab8c028062469853b07889d5602475ad7623187825642b37.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 396, + 295, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 433 + ], + "type": "text", + "content": "Here the experts in " + }, + { + "bbox": [ + 55, + 396, + 295, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{\\mathrm{MoE}}" + }, + { + "bbox": [ + 55, + 396, + 295, + 433 + ], + "type": "text", + "content": " leverage " + }, + { + "bbox": [ + 55, + 396, + 295, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{\\mathrm{FFN}}" + }, + { + "bbox": [ + 55, + 396, + 295, + 433 + ], + "type": "text", + "content": " as the base weight for LoRA adaption, with each of them employs a specific instance of LoRA (i.e., " + }, + { + "bbox": [ + 55, + 396, + 295, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{\\mathrm{LoRA}}" + }, + { + "bbox": [ + 55, + 396, + 295, + 433 + ], + "type": "text", + "content": ") parameterized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 438, + 295, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 438, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 89, + 438, + 295, + 453 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\operatorname {M o E}} \\left(\\boldsymbol {X} ^ {\\prime}\\right)) = \\sum_ {i = 1} ^ {n} \\mathcal {G} \\left(\\boldsymbol {X} ^ {\\prime}\\right) _ {i} \\mathcal {F} _ {i} ^ {\\text {L o R A}} \\left(\\boldsymbol {X} ^ {\\prime}\\right), \\tag {8}", + "image_path": "2b742e15a882cfa4c44270638b190c2a9418badc614ddc9e2c4d37511babd717.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 90, + 454, + 294, + 474 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 454, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 90, + 454, + 294, + 474 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {i} ^ {\\text {L o R A}} \\left(\\boldsymbol {X} ^ {\\prime}\\right) = \\frac {\\alpha}{r} \\cdot \\boldsymbol {B} _ {i} \\boldsymbol {A} _ {i} \\boldsymbol {X} ^ {\\prime}, \\tag {9}", + "image_path": "a9ccf4b23e301e8c14471c17a3302ed36731947feadb5eccaebc2825907856b0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 483, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 556 + ], + "type": "text", + "content": "where low-rank matrices " + }, + { + "bbox": [ + 55, + 483, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{d_{in} \\times r}" + }, + { + "bbox": [ + 55, + 483, + 295, + 556 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 483, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{B} \\in \\mathbb{R}^{r \\times d_{out}}" + }, + { + "bbox": [ + 55, + 483, + 295, + 556 + ], + "type": "text", + "content": " are scaled by a factor " + }, + { + "bbox": [ + 55, + 483, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 483, + 295, + 556 + ], + "type": "text", + "content": " normalized by the rank " + }, + { + "bbox": [ + 55, + 483, + 295, + 556 + ], + "type": "inline_equation", + "content": "r \\ll \\min(d_{in}, d_{out})" + }, + { + "bbox": [ + 55, + 483, + 295, + 556 + ], + "type": "text", + "content": ". In this way, we achieve seamless integration of LoRA with MoE, enabling efficient tuning and knowledge preservation of the foundation models, while supporting adaptation to inputs across various modalities." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 564, + 221, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 221, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 221, + 577 + ], + "type": "text", + "content": "3.2. Universal Modality Alignment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "Soft Modality Binding across Datasets. Prior work [18] reveals that modalities expressed in a joint embedding space could be automatically aligned through an intermediate modality. For example, the alignment of and , could lead to the alignment between with Image as the bridge. However, this approach requires massive naturally paired samples between anchor and extended modalities, which are often challenging and time-consuming to obtain. We thus propose a soft modality binding strategy, which utilizes unpaired samples across datasets with different modalities" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 318, + 67, + 553, + 284 + ], + "blocks": [ + { + "bbox": [ + 318, + 67, + 553, + 284 + ], + "lines": [ + { + "bbox": [ + 318, + 67, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 318, + 67, + 553, + 284 + ], + "type": "image", + "image_path": "408eda2d6b769a4c6ec7401c6deb1b2233c62cb3e50df4ccf3e3569909b193d1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 285, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 319 + ], + "type": "text", + "content": "Figure 3. Our proposed soft modality binding strategy. Unpaired samples from diverse datasets with different modalities are bridged with the semantic similarity between class labels as soft guidance." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 323, + 555, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 388 + ], + "type": "text", + "content": "(Fig. 3). Specifically, let " + }, + { + "bbox": [ + 313, + 323, + 555, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^{\\mathcal{M}_1} = \\{\\pmb {y}_1^{\\mathcal{M}_1},\\dots ,\\pmb {y}_{N_1}^{\\mathcal{M}_1}\\}" + }, + { + "bbox": [ + 313, + 323, + 555, + 388 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 323, + 555, + 388 + ], + "type": "inline_equation", + "content": "\\pmb {\\mathcal{V}}^{\\mathcal{M}_2} = \\{\\pmb {y}_1^{\\mathcal{M}_2},\\dots ,\\pmb {y}_{N_2}^{\\mathcal{M}_2}\\}" + }, + { + "bbox": [ + 313, + 323, + 555, + 388 + ], + "type": "text", + "content": " represent text embeddings of class labels from two arbitrary datasets in modalities " + }, + { + "bbox": [ + 313, + 323, + 555, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_1" + }, + { + "bbox": [ + 313, + 323, + 555, + 388 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 323, + 555, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_2" + }, + { + "bbox": [ + 313, + 323, + 555, + 388 + ], + "type": "text", + "content": ", respectively. The semantic similarity of textual labels across datasets can be computed as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 339, + 392, + 554, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 392, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 339, + 392, + 554, + 426 + ], + "type": "interline_equation", + "content": "\\mathcal {S} \\left(\\boldsymbol {y} _ {i} ^ {\\mathcal {M} _ {1}}, \\boldsymbol {y} _ {j} ^ {\\mathcal {M} _ {2}}\\right) = \\frac {\\exp \\left(\\boldsymbol {y} _ {i} ^ {\\mathcal {M} _ {1}} \\cdot \\boldsymbol {y} _ {j} ^ {\\mathcal {M} _ {2}}\\right)}{\\sum_ {n = 1} ^ {N _ {2}} \\exp \\left(\\boldsymbol {y} _ {i} ^ {\\mathcal {M} _ {1}}, \\boldsymbol {y} _ {n} ^ {\\mathcal {M} _ {2}}\\right)}. \\tag {10}", + "image_path": "05b3f0488107336fe081c503ab4f86864e9e0850f36283a42265cfc5396b8f45.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 430, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 555, + 544 + ], + "type": "text", + "content": "Similarly, given samples " + }, + { + "bbox": [ + 313, + 430, + 555, + 544 + ], + "type": "inline_equation", + "content": "(\\pmb{x}^{\\mathcal{M}_1},\\pmb{y}_i^{\\mathcal{M}_1})" + }, + { + "bbox": [ + 313, + 430, + 555, + 544 + ], + "type": "text", + "content": " from the first dataset and " + }, + { + "bbox": [ + 313, + 430, + 555, + 544 + ], + "type": "inline_equation", + "content": "(\\pmb{x}^{\\mathcal{M}_2},\\pmb{y}_j^{\\mathcal{M}_2})" + }, + { + "bbox": [ + 313, + 430, + 555, + 544 + ], + "type": "text", + "content": " from the second dataset, the semantic similarity between " + }, + { + "bbox": [ + 313, + 430, + 555, + 544 + ], + "type": "inline_equation", + "content": "\\pmb{x}^{\\mathcal{M}_1}" + }, + { + "bbox": [ + 313, + 430, + 555, + 544 + ], + "type": "text", + "content": " and text embedding of labels from the second dataset (i.e., " + }, + { + "bbox": [ + 313, + 430, + 555, + 544 + ], + "type": "inline_equation", + "content": "S(\\pmb{x}^{\\mathcal{M}_1},\\pmb{y}_j^{\\mathcal{M}_2}))" + }, + { + "bbox": [ + 313, + 430, + 555, + 544 + ], + "type": "text", + "content": ", as well as " + }, + { + "bbox": [ + 313, + 430, + 555, + 544 + ], + "type": "inline_equation", + "content": "\\pmb{x}^{\\mathcal{M}_2}" + }, + { + "bbox": [ + 313, + 430, + 555, + 544 + ], + "type": "text", + "content": " and text embeddings of labels from the first dataset (i.e., " + }, + { + "bbox": [ + 313, + 430, + 555, + 544 + ], + "type": "inline_equation", + "content": "S(\\pmb{x}^{\\mathcal{M}_2},\\pmb{y}_i^{\\mathcal{M}_1}))" + }, + { + "bbox": [ + 313, + 430, + 555, + 544 + ], + "type": "text", + "content": " could be computed. Taking the similarity between class labels of two datasets (i.e., " + }, + { + "bbox": [ + 313, + 430, + 555, + 544 + ], + "type": "inline_equation", + "content": "S(\\pmb{y}_i^{\\mathcal{M}_1},\\pmb{y}_j^{\\mathcal{M}_2}))" + }, + { + "bbox": [ + 313, + 430, + 555, + 544 + ], + "type": "text", + "content": " as a soft prior for alignment, we bind modalities across datasets by optimizing the following objective:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 323, + 547, + 527, + 579 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 547, + 527, + 579 + ], + "spans": [ + { + "bbox": [ + 323, + 547, + 527, + 579 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {s o f t \\_ b i n d}} = \\mathrm {K L} \\left(\\mathcal {S} \\left(\\boldsymbol {x} ^ {\\mathcal {M} _ {1}}, \\boldsymbol {y} _ {j} ^ {\\mathcal {M} _ {2}}\\right) \\mid \\mathcal {S} \\left(\\boldsymbol {y} _ {i} ^ {\\mathcal {M} _ {1}}, \\boldsymbol {y} _ {j} ^ {\\mathcal {M} _ {2}}\\right)\\right) \\\\ + \\operatorname {K L} \\left(\\mathcal {S} \\left(\\boldsymbol {x} ^ {\\mathcal {M} _ {2}}, \\boldsymbol {y} _ {i} ^ {\\mathcal {M} _ {1}}\\right) \\mid \\mathcal {S} \\left(\\boldsymbol {y} _ {j} ^ {\\mathcal {M} _ {2}}, \\boldsymbol {y} _ {i} ^ {\\mathcal {M} _ {1}}\\right)\\right), \\\\ \\end{array}", + "image_path": "3419ed6fc5326d05220dd41ed9cf1bf456f6f8694466443899df6b581238c44d.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 582, + 554, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 642 + ], + "type": "text", + "content": "where KL refers to the Kullback-Leibler divergence. Our soft modality binding strategy is different from [18], which requires instance-level textual annotations for each sample. Instead, our approach utilizes merely class labels which are instance-agnostic and challenging for alignment." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": "Knowledge Distillation from Anchor Modalities. Given multiple well-aligned anchor modalities (e.g., " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_1^a" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_2^a" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": ") and one novel modality to be extended (e.g., " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_k^e" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": "), the alignment performance between anchor and extended modalities (e.g., " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_k^e" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_1^a" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": ") is typically less effective than that between anchor modalities (e.g., " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_1^a" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_2^a" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": "). Therefore," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29647" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 66, + 293, + 252 + ], + "blocks": [ + { + "bbox": [ + 58, + 66, + 293, + 252 + ], + "lines": [ + { + "bbox": [ + 58, + 66, + 293, + 252 + ], + "spans": [ + { + "bbox": [ + 58, + 66, + 293, + 252 + ], + "type": "image", + "image_path": "e1b8aceedb68cc74451b32ccc1739996b9e076f40208947c09536b91a150e785.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 255, + 295, + 288 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 295, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 295, + 288 + ], + "type": "text", + "content": "Figure 4. Top: Knowledge distillation from anchor modalities to instruct the alignment of novel modalities. Bottom: knowledge distillation from multiple partial-modal foundation models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 296, + 295, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 296, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 295, + 331 + ], + "type": "text", + "content": "knowledge within the alignment between anchor modalities (Fig. 4 (Top)) could be transferred to improve the alignment learning of extended modalities via knowledge distillation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 331, + 296, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 428 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 428 + ], + "type": "text", + "content": "For instance, given a 3D point cloud sample " + }, + { + "bbox": [ + 55, + 331, + 296, + 428 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 55, + 331, + 296, + 428 + ], + "type": "text", + "content": ", we could obtain its tokenized input " + }, + { + "bbox": [ + 55, + 331, + 296, + 428 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{i}^{\\mathcal{P}}" + }, + { + "bbox": [ + 55, + 331, + 296, + 428 + ], + "type": "text", + "content": " in the extended modalities, along with the text embedding of class label " + }, + { + "bbox": [ + 55, + 331, + 296, + 428 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{i}^{\\mathcal{T}}" + }, + { + "bbox": [ + 55, + 331, + 296, + 428 + ], + "type": "text", + "content": " and the visual embedding of one rendered image from a specific viewpoint " + }, + { + "bbox": [ + 55, + 331, + 296, + 428 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{i}^{\\mathcal{T}}" + }, + { + "bbox": [ + 55, + 331, + 296, + 428 + ], + "type": "text", + "content": ", both in the anchor modalities. By leveraging the similarity between anchor modalities (i.e., " + }, + { + "bbox": [ + 55, + 331, + 296, + 428 + ], + "type": "inline_equation", + "content": "S(\\pmb{x}_{i}^{\\mathcal{T}},\\pmb{x}_{i}^{\\mathcal{T}})" + }, + { + "bbox": [ + 55, + 331, + 296, + 428 + ], + "type": "text", + "content": ") as the guidance, we could optimize the alignment between the extended modality and anchor modalities through:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 431, + 295, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 431, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 93, + 431, + 295, + 460 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {a n c h . d i s t i l}} = \\mathrm {K L} \\left(S \\left(\\boldsymbol {x} _ {i} ^ {\\mathcal {P}}, \\boldsymbol {x} _ {i} ^ {\\mathcal {T}}\\right) \\mid S \\left(\\boldsymbol {x} _ {i} ^ {\\mathcal {I}}, \\boldsymbol {x} _ {i} ^ {\\mathcal {T}}\\right)\\right) \\tag {12} \\\\ + \\operatorname {K L} \\left(\\mathcal {S} \\left(\\boldsymbol {x} _ {i} ^ {\\mathcal {P}}, \\boldsymbol {x} _ {i} ^ {\\mathcal {I}}\\right) \\mid \\mathcal {S} \\left(\\boldsymbol {x} _ {i} ^ {\\mathcal {T}}, \\boldsymbol {x} _ {i} ^ {\\mathcal {I}}\\right)\\right). \\\\ \\end{array}", + "image_path": "487a40416273ad9521e018fc03bd6d3e3a724918f64a7d12c9883e9b9cb8a082.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 464, + 295, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 295, + 499 + ], + "type": "text", + "content": "This knowledge distillation strategy is generally applicable to most modalities, where 3D point cloud, video, and depth data can be converted into images with class labels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 500, + 295, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 500, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 295, + 571 + ], + "type": "text", + "content": "Knowledge Distillation from Partial-Modal Foundation Models. To further boost the alignment across modalities, we employ foundation models " + }, + { + "bbox": [ + 55, + 500, + 295, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_f" + }, + { + "bbox": [ + 55, + 500, + 295, + 571 + ], + "type": "text", + "content": " skilled in partial modalities (e.g., audio " + }, + { + "bbox": [ + 55, + 500, + 295, + 571 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 55, + 500, + 295, + 571 + ], + "type": "text", + "content": " text [22], point cloud " + }, + { + "bbox": [ + 55, + 500, + 295, + 571 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 55, + 500, + 295, + 571 + ], + "type": "text", + "content": " image [98]) as teacher models to guide the learning of our unified modality encoder " + }, + { + "bbox": [ + 55, + 500, + 295, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_u" + }, + { + "bbox": [ + 55, + 500, + 295, + 571 + ], + "type": "text", + "content": " (i.e., Fig.4 (Bottom)) from two perspectives:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 571, + 295, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 295, + 632 + ], + "type": "text", + "content": "- Feature Alignment aims to ensure features from our unified modality encoder to be identical as those from partial-modal foundation models. Specifically, denoting " + }, + { + "bbox": [ + 55, + 571, + 295, + 632 + ], + "type": "inline_equation", + "content": "\\hat{F}_i^{\\mathcal{M}_k}" + }, + { + "bbox": [ + 55, + 571, + 295, + 632 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 571, + 295, + 632 + ], + "type": "inline_equation", + "content": "F_{i}^{\\mathcal{M}_{k}}" + }, + { + "bbox": [ + 55, + 571, + 295, + 632 + ], + "type": "text", + "content": " as features in modality " + }, + { + "bbox": [ + 55, + 571, + 295, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_k" + }, + { + "bbox": [ + 55, + 571, + 295, + 632 + ], + "type": "text", + "content": " output by " + }, + { + "bbox": [ + 55, + 571, + 295, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_f" + }, + { + "bbox": [ + 55, + 571, + 295, + 632 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 571, + 295, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_u" + }, + { + "bbox": [ + 55, + 571, + 295, + 632 + ], + "type": "text", + "content": ", we use the mean squared error (MSE) to align them as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 117, + 635, + 295, + 650 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 635, + 295, + 650 + ], + "spans": [ + { + "bbox": [ + 117, + 635, + 295, + 650 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {f e a t - a l i g n}} = \\left\\| \\hat {\\boldsymbol {F}} ^ {\\mathcal {M} _ {i}} - \\boldsymbol {F} ^ {\\mathcal {M} _ {i}} \\right\\| _ {2} ^ {2}. \\tag {13}", + "image_path": "7b2916afe910754275518b8f736e7e8c413cd9248b0c0c5d5b5cf30afaca56fa.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "- Cross-Modality Knowledge Transfer aims to benefit the learning of " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_u" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " via the well-aligned representation space across modalities shaped by large-scale paired data. Specifically, given a paired sample " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "(x_i^{\\mathcal{M}^a}, x_i^{\\mathcal{M}^e})" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " in anchor modality " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^a" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " and extended modality " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^e" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " respectively, the align" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "ment results of our unified modality encoder are encouraged to approximate those of partial-modal foundation models:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 328, + 99, + 553, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 99, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 328, + 99, + 553, + 129 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {m o d a l - a l i g n}} = \\mathrm {K L} \\left(\\mathcal {S} \\left(\\boldsymbol {x} _ {i} ^ {\\mathcal {M} ^ {e}}, \\boldsymbol {x} _ {i} ^ {\\mathcal {M} ^ {a}}\\right) \\mid \\mathcal {S} \\left(\\hat {\\boldsymbol {x}} _ {i} ^ {\\mathcal {M} ^ {e}}, \\hat {\\boldsymbol {x}} _ {i} ^ {\\mathcal {M} ^ {a}}\\right)\\right) \\tag {14} \\\\ + \\mathrm {K L} \\left(S \\left(\\boldsymbol {x} _ {i} ^ {\\mathcal {M} ^ {a}}, \\boldsymbol {x} _ {i} ^ {\\mathcal {M} ^ {e}}\\right) \\mid S \\left(\\hat {\\boldsymbol {x}} _ {i} ^ {\\mathcal {M} ^ {a}}, \\hat {\\boldsymbol {x}} _ {i} ^ {\\mathcal {M} ^ {e}}\\right)\\right), \\\\ \\end{array}", + "image_path": "65186a95c9cbf7eb342bedff1d54d02dbef55e7dfde0958d082ec2ad55e1685c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 132, + 538, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 132, + 538, + 143 + ], + "spans": [ + { + "bbox": [ + 313, + 132, + 538, + 143 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 132, + 538, + 143 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_i" + }, + { + "bbox": [ + 313, + 132, + 538, + 143 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 132, + 538, + 143 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i" + }, + { + "bbox": [ + 313, + 132, + 538, + 143 + ], + "type": "text", + "content": " are outputs of " + }, + { + "bbox": [ + 313, + 132, + 538, + 143 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_f" + }, + { + "bbox": [ + 313, + 132, + 538, + 143 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 132, + 538, + 143 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_u" + }, + { + "bbox": [ + 313, + 132, + 538, + 143 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 144, + 554, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 554, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 554, + 167 + ], + "type": "text", + "content": "In summary, the loss for knowledge distillation from partial-modal foundation models can be formulated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 359, + 171, + 553, + 184 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 171, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 359, + 171, + 553, + 184 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m o d e l - d i s t i l}} = \\lambda \\mathcal {L} _ {\\text {f e a t - a l i n g}} + \\mathcal {L} _ {\\text {m o d a l - a l i n g}}. \\tag {15}", + "image_path": "c326c70f1812151c773469a55c66be8b5bb8949e781adc9d286b3cca05c7598c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 186, + 542, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 542, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 542, + 198 + ], + "type": "text", + "content": "Following [99], we set " + }, + { + "bbox": [ + 313, + 186, + 542, + 198 + ], + "type": "inline_equation", + "content": "\\lambda = 2000" + }, + { + "bbox": [ + 313, + 186, + 542, + 198 + ], + "type": "text", + "content": " to scale the magnitude." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 204, + 447, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 204, + 447, + 217 + ], + "spans": [ + { + "bbox": [ + 314, + 204, + 447, + 217 + ], + "type": "text", + "content": "3.3. Implementation Details" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 221, + 555, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 555, + 329 + ], + "type": "text", + "content": "Network Configuration. We adopt CLIP [67] pre-trained vision encoder as the base network for unified modality encoding. The sequence length " + }, + { + "bbox": [ + 313, + 221, + 555, + 329 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 221, + 555, + 329 + ], + "type": "text", + "content": " and channel dimension " + }, + { + "bbox": [ + 313, + 221, + 555, + 329 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 221, + 555, + 329 + ], + "type": "text", + "content": " for tokenized inputs " + }, + { + "bbox": [ + 313, + 221, + 555, + 329 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 221, + 555, + 329 + ], + "type": "text", + "content": " are set to 196 and 768, respectively. The LoRA configuration is defined with a scaling factor " + }, + { + "bbox": [ + 313, + 221, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 221, + 555, + 329 + ], + "type": "text", + "content": " of 32 and a rank " + }, + { + "bbox": [ + 313, + 221, + 555, + 329 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 221, + 555, + 329 + ], + "type": "text", + "content": " of 4. For knowledge distillation from partial-modal foundation models, we use ULIP [98] as the teacher model for 3D point cloud inputs, VIT-LENS[38] for depth inputs, ONE-PEACE[86] for audio and video inputs." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 330, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 554, + 437 + ], + "type": "text", + "content": "Batching Strategy. The batching strategies proposed by [17, 42] sample inputs from each dataset (modality) separately, which may lead to catastrophic forgetting[42]. To address this, we develop a new batching strategy, which mixes data from all modalities within each mini-batch. Furthermore, a modality-aware gradient accumulation is proposed to perform forward pass, loss computation, and gradient calculation individually for each modality. Backpropagation is conducted only after processing all modalities." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 437, + 554, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 461 + ], + "type": "text", + "content": "Training Objective. Following [68], we assign an equal weight to all training objectives:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 463, + 553, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 463, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 317, + 463, + 553, + 475 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {N C E}} + \\alpha_ {1} \\mathcal {L} _ {\\text {s o f t , b i n d}} + \\alpha_ {2} \\mathcal {L} _ {\\text {a n c h , d i s t i l}} + \\alpha_ {3} \\mathcal {L} _ {\\text {m o d e l , d i s t i l}}, \\tag {16}", + "image_path": "44e97f9f0ab2c8d1d0363cd679ed1b49f7e1ed254c064dfd3783b05dff5cc304.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 479, + 554, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 554, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 554, + 503 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 479, + 554, + 503 + ], + "type": "inline_equation", + "content": "\\alpha_{\\{1,2,3\\}} = 1" + }, + { + "bbox": [ + 313, + 479, + 554, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 479, + 554, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{NCE}}" + }, + { + "bbox": [ + 313, + 479, + 554, + 503 + ], + "type": "text", + "content": " is the modality aligning loss for paired data, as utilized in CLIP [67]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 510, + 395, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 510, + 395, + 525 + ], + "spans": [ + { + "bbox": [ + 314, + 510, + 395, + 525 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 314, + 529, + 430, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 529, + 430, + 541 + ], + "spans": [ + { + "bbox": [ + 314, + 529, + 430, + 541 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 677 + ], + "type": "text", + "content": "Pre-training Datasets. We perform a joint pre-training across multiple datasets with different modalities, including images from ImageNet-1K [8], 3D point clouds from ShapeNet[5], depth data from SUN RGB-D[74], audio and video from AudioSet [16] and VGGSound [6]. To address the differences in dataset size across modalities, we apply repeated augmentation[17] to small datasets (such as SUN-D) and category-balanced downsampling for large datasets (such as AudioSet) to ensure balanced performance during joint training. Once finished pre-training, we directly evaluate our model on downstream tasks without any fine-tuning." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Training. For image and video data, we use a resolution of " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": " and apply standard augmentations. For single-view depth data, we employ the same augmentation as in" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29648" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 70, + 294, + 203 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 294, + 203 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 294, + 203 + ], + "type": "table", + "html": "
ModelBackboneTrainable ParamsTop-1 AccuracyTop-5 Accuracy
Single Task
Swin Transformer [47]Swin-B87.7M83.596.5
ConvNeXt [48]ConvNeXt-B88.7M83.896.7
DeiT III [80]ViT-B86.9M83.896.5
Conformer [64]Conformer-B83.3M83.896.7
Hiera [70]Hiera-B+70.0M85.297.3
Multiple Tasks
UniRepLKNet [10]UniRepLKNet-S55.6M83.9-
OMNIVORE [17]Swin-B88.2M85.397.5
MetaFormer [107]ViT-B86.6M85.497.4
UNIALIGN†ViT-B7.8M85.397.2
", + "image_path": "f678c4bb8357c69dfe413ffa6d12ba3c12d21adec8dfa247282fba6014721de8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 61, + 204, + 253, + 213 + ], + "lines": [ + { + "bbox": [ + 61, + 204, + 253, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 204, + 253, + 213 + ], + "type": "text", + "content": "An identical model with the same weight across all tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 58, + 214, + 293, + 224 + ], + "lines": [ + { + "bbox": [ + 58, + 214, + 293, + 224 + ], + "spans": [ + { + "bbox": [ + 58, + 214, + 293, + 224 + ], + "type": "text", + "content": "Table 1. Results for image classification on ImageNet-1K (§4.2)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 62, + 236, + 294, + 344 + ], + "blocks": [ + { + "bbox": [ + 62, + 236, + 294, + 344 + ], + "lines": [ + { + "bbox": [ + 62, + 236, + 294, + 344 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 294, + 344 + ], + "type": "table", + "html": "
ModelBackboneTrainable ParamsModelNet40ScanObjectNN
Top-1Top-5Top-1Top-5
Single Task
PointCLIP [106]ResNet-50-19.334.810.530.6
PointCLIPv2 [111]ViT-B-63.685.042.274.5
CLIP2Point [30]ViT-B88.3M49.581.225.559.4
RECON [66]ViT-B43.6M61.278.142.375.6
ULIP [98]ViT-B43.6M60.484.051.580.2
Multiple Tasks
VIT-LENS [38]ViT-B34.1M65.492.7--
UNIALIGN†ViT-B7.8M55.984.342.576.3
", + "image_path": "e6233942872423b839c3bbb7d797939681fc53404b31a060907d254a29fa7bd7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 61, + 345, + 253, + 354 + ], + "lines": [ + { + "bbox": [ + 61, + 345, + 253, + 354 + ], + "spans": [ + { + "bbox": [ + 61, + 345, + 253, + 354 + ], + "type": "text", + "content": "An identical model with the same weight across all tasks." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 55, + 354, + 295, + 376 + ], + "lines": [ + { + "bbox": [ + 55, + 354, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 295, + 376 + ], + "type": "text", + "content": "Table 2. Results for zero-shot 3d shape classification on ModelNet40 test [95] and ScanObjectNN test [82] (\\$4.3)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 389, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 521 + ], + "type": "text", + "content": "[38]. The augmentation of audio data also follows [38] except Mix-up [94]. We train the model with a batch size of 2048 (after gradient accumulation) for 100 epochs. An AdamW optimizer with a initial learning rate " + }, + { + "bbox": [ + 55, + 389, + 296, + 521 + ], + "type": "inline_equation", + "content": "2\\mathrm{e}^{-4}" + }, + { + "bbox": [ + 55, + 389, + 296, + 521 + ], + "type": "text", + "content": " and a weight decay of 0.05 is employed. For knowledge distillation from partial-modal foundation models, we cache offline features to avoid repeatedly forwarding. The base modality encoder remains frozen during training, with only LoRA enhanced MoE modules and the projection head for knowledge distillation remaining trainable. Flash-attention is applied to reduce GPU memory cost." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 521, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 544 + ], + "type": "text", + "content": "Reproducibility. UNIALIGN is implemented in PyTorch and trained on four NVIDIA Tesla A100 GPUs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 552, + 227, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 227, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 227, + 565 + ], + "type": "text", + "content": "4.2. Results for Image Classification" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "type": "text", + "content": "Dataset. We evaluate UNIALIGN for image classification on ImageNet-1K val [8] which contains 50,000 samples across 1,000 classes, spanning objects, animals, and more." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 605, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 629 + ], + "type": "text", + "content": "Metric. Following previous works [47, 48, 87], we report the count of trainable parameters, top-1 and top-5 accuracy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "Performance. As shown in Table 1, UNIALIGN demonstrates impressive performance on image classification, achieving a comparable performance to modality-specific approaches (i.e., " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "85.3\\%" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "85.2\\%" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " of Hiera[70]). It is noteworthy that our method utilizes merely " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " trainable parameters and a unified modality encoder architecture which preserves the capability to process other modalities." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 321, + 71, + 553, + 190 + ], + "blocks": [ + { + "bbox": [ + 321, + 71, + 553, + 190 + ], + "lines": [ + { + "bbox": [ + 321, + 71, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 321, + 71, + 553, + 190 + ], + "type": "table", + "html": "
ModelBackboneTrainable ParamsAnchor ModalityNYU-DSUN-D
Single Task
G-L-SOOR [76]D-CNN [75]--62.344.1
RecgNet [12]ResNet18 [24]11.2M-57.747.9
Depth Swin [18]Swin-B88.2M-76.469.1
Multiple Tasks
Text Paired [18]ViT-H88.1MT41.925.4
ImageBind [18]ViT-H88.3MI54.035.1
ViT-LENS [38]ViT-L50.9MI+T68.552.2
UNIALIGN†ViT-B7.8MT46.334.3
UNIALIGN†ViT-B7.8MI+T59.445.8
UNIALIGN†ViT-B7.8MI+T71.4*58.2*
", + "image_path": "4cc03c65044fc2d34ed84304b999af1b09a852eeb95b79b5def9c6c282d79ec0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 319, + 190, + 511, + 200 + ], + "lines": [ + { + "bbox": [ + 319, + 190, + 511, + 200 + ], + "spans": [ + { + "bbox": [ + 319, + 190, + 511, + 200 + ], + "type": "text", + "content": "An identical model with the same weight across all tasks." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 313, + 200, + 553, + 233 + ], + "lines": [ + { + "bbox": [ + 313, + 200, + 553, + 233 + ], + "spans": [ + { + "bbox": [ + 313, + 200, + 553, + 233 + ], + "type": "text", + "content": "Table 3. Results for depth-only scene classification on NYU-D test [73] and SUN-D test [74] (\\$4.4).*: The depth maps are tokenized by the image tokenizer additionally." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 243, + 550, + 256 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 550, + 256 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 550, + 256 + ], + "type": "text", + "content": "4.3. Results for Zero-Shot 3D Shape Classification" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 262, + 553, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 553, + 309 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 553, + 309 + ], + "type": "text", + "content": "Dataset. We evaluate zero-shot 3D shape classification on two datasets. ModelNet40 test [95] contains 2,468 samples from 40 categories, while ScanObjectNN test [82] consists of 581 samples distributed across 15 categories." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 310, + 522, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 310, + 522, + 322 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 522, + 322 + ], + "type": "text", + "content": "Metric. The top-1 and top-5 accuracy are reported." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 323, + 554, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 554, + 394 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 554, + 394 + ], + "type": "text", + "content": "Performance. As shown in Table 2, UNIALIGN achieves a top-1 accuracy of " + }, + { + "bbox": [ + 313, + 323, + 554, + 394 + ], + "type": "inline_equation", + "content": "55.9\\%" + }, + { + "bbox": [ + 313, + 323, + 554, + 394 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 323, + 554, + 394 + ], + "type": "inline_equation", + "content": "42.5\\%" + }, + { + "bbox": [ + 313, + 323, + 554, + 394 + ], + "type": "text", + "content": " on ModelNet40 val, and ScanObjectNN test, respectively. Though SOTA approaches such as ULIP [98] and ViT-LENS [38] yield higher scores, they require a modality-specific point-cloud encoder and an additional image encoder for alignment." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 403, + 537, + 416 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 537, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 537, + 416 + ], + "type": "text", + "content": "4.4. Results for Depth-only Scene Classification" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 422, + 554, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 554, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 554, + 469 + ], + "type": "text", + "content": "Dataset. Experiments for depth-only scene classification are conducted on SUN RGB-D test [74] which contains 4,660 samples across 19 classes, and NYU-D test [73] which contains 653 samples across 10 semantic classes." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 470, + 531, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 531, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 531, + 481 + ], + "type": "text", + "content": "Metric. The top-1 accuracy is adopted for evaluation." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 483, + 554, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 554, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 554, + 626 + ], + "type": "text", + "content": "Performance. A detailed comparison of UNIALIGN against several top-leading approaches for depth-only scene classification is provided in Table 3. Building on a ViT-B backbone and aligning extended modalities with both image and text as anchor modalities, UNIALIGN achieves a top-1 accuracy of " + }, + { + "bbox": [ + 313, + 483, + 554, + 626 + ], + "type": "inline_equation", + "content": "71.4\\%" + }, + { + "bbox": [ + 313, + 483, + 554, + 626 + ], + "type": "text", + "content": " on NYU-D test and " + }, + { + "bbox": [ + 313, + 483, + 554, + 626 + ], + "type": "inline_equation", + "content": "58.2\\%" + }, + { + "bbox": [ + 313, + 483, + 554, + 626 + ], + "type": "text", + "content": " on SUN RGB-D test, surpassing all competitors. Notably, aligning merely to the text modality leads to a marked performance decline (i.e., " + }, + { + "bbox": [ + 313, + 483, + 554, + 626 + ], + "type": "inline_equation", + "content": "71.4\\% \\rightarrow 46.3\\%" + }, + { + "bbox": [ + 313, + 483, + 554, + 626 + ], + "type": "text", + "content": " on NYU-D). This suggests that i) the depth modality exhibits a closer alignment with image data, and ii) UNIALIGN can effectively leverage the additional anchor modality to enhance alignment performance." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 635, + 550, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 550, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 550, + 647 + ], + "type": "text", + "content": "4.5. Results for Audio Classification and Retrieval" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Datasets. We evaluate audio classification on AudioSet val [16] containing 2M samples across 527 classes, and ESC [65] including 2K data across 50 classes. Audio-text retrieval is evaluated on AudioCaps test [33] which includes 813 audio-text pairs. Moreover, following [107], we" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29649" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 70, + 294, + 331 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 294, + 331 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 294, + 331 + ], + "type": "table", + "html": "
MethodBackboneAnchor ModalityAudioSet mAPESC Top1AudioCaps R@1R@10
Single Task
AVFIC [57]ViT-B---8.737.7
AudioCLIP [22]ViT-BI+T25.969.4--
Multiple Tasks
ImageBind-H [18]ViT-HI17.666.99.342.3
VIT-LENS [38]ViT-LI+T27.280.914.955.2
LanguageBind [109]ViT-LI+T27.791.812.253.2
UNIALIGN†ViT-BI+T23.870.511.749.3
†:An identical model with the same weight across all tasks. \nTable 4. Results for audio classification on AudioSet val [16], ESC[65], and audio-text retrieval on AudioCaps test[33] (§4.5). Due to expiration of links, UNIALIGN is trained with only 0.5M ideas from AudioSet, while methods such as ViT-Lens use 1.6M.
MethodBackbonePre-train DatasetsTrainable ParamsTop-1
Single Task
AST [19]ViT-BN/A86.9M92.6
AST [19]ViT-BAudioSet-20K86.9M96.2
SSAST [20]ViT-BAudioSet-2M89.3M97.8
Multiple Tasks
MetaFormer [107]ViT-BLAION2B1.1M78.3
MetaFormer [107]ViT-BLAION2B86.3M97.0
UniRepLKNet [10]UniRepLKNet-S-55.5M98.5
UNIALIGNViT-BJoint1.8M97.4
", + "image_path": "0c7782e2985bf55d85b3baf7d7f8db893756a4c96fb6fe00e89b93bd06978e72.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 369, + 295, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 295, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 295, + 393 + ], + "type": "text", + "content": "fine-tune UNIALIGN on Speech Commands V2 train[92] for speech recognition." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 395, + 295, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 395, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 295, + 431 + ], + "type": "text", + "content": "Metric. Following [18, 22, 38], we measure mAP for AudioSet, top-1 accuracy for ESC and Speech Commands V2, and Recall@ " + }, + { + "bbox": [ + 55, + 395, + 295, + 431 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 395, + 295, + 431 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 395, + 295, + 431 + ], + "type": "inline_equation", + "content": "k = \\{1,5\\}" + }, + { + "bbox": [ + 55, + 395, + 295, + 431 + ], + "type": "text", + "content": ") for AudioCaps." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 432, + 295, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 295, + 503 + ], + "type": "text", + "content": "Performance for Audio classification. UNIALIGN achieves promising results on AudioSet and ESC (Table 4). Though inferior to top-leading methods, it is important to note that, due to the expiration of links, a significant number of videos from AudioSet can not be downloaded from YouTube (i.e., 1.6M used in ViT-Lens[38] vs. 0.5M used in UNIALIGN)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 504, + 295, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 295, + 540 + ], + "type": "text", + "content": "Performance for Audio retrieval. Similarly, despite the absence of 1.1M training data, UNIALIGN can still achieve comparable performance, achieving 11.7 R@1 accuracy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 541, + 295, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 541, + 295, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 541, + 295, + 637 + ], + "type": "text", + "content": "Performance for Speech Recognition. As shown in Table 5, UNIALIGN demonstrates remarkable accuracy compared to existing work (e.g., " + }, + { + "bbox": [ + 55, + 541, + 295, + 637 + ], + "type": "inline_equation", + "content": "97.4\\%" + }, + { + "bbox": [ + 55, + 541, + 295, + 637 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 55, + 541, + 295, + 637 + ], + "type": "inline_equation", + "content": "97.0\\%" + }, + { + "bbox": [ + 55, + 541, + 295, + 637 + ], + "type": "text", + "content": " for MetaTransformer[107]), while utilizing significantly fewer trainable parameters (e.g., 1.8M vs. 86.3M). When tuning a similar number of parameters (i.e., 1.1M for Meta-Transformer, and 1.8M for UNIALIGN), our method outperforms Meta-Transformer by " + }, + { + "bbox": [ + 55, + 541, + 295, + 637 + ], + "type": "inline_equation", + "content": "19.1\\%" + }, + { + "bbox": [ + 55, + 541, + 295, + 637 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 55, + 541, + 295, + 637 + ], + "type": "inline_equation", + "content": "97.4\\%" + }, + { + "bbox": [ + 55, + 541, + 295, + 637 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 55, + 541, + 295, + 637 + ], + "type": "inline_equation", + "content": "78.3\\%" + }, + { + "bbox": [ + 55, + 541, + 295, + 637 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 647, + 291, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 291, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 291, + 659 + ], + "type": "text", + "content": "4.6. Results for Video Classification and Retrieval" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "Datasets. We utilize MSR-VTT val [96] which contains around 1,000 video-text pairs for zero-shot video-text retrieval, and UCF101 val [77] which contains 3,738 samples across 101 semantic classes for video classification." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 321, + 71, + 553, + 177 + ], + "blocks": [ + { + "bbox": [ + 55, + 336, + 295, + 360 + ], + "lines": [ + { + "bbox": [ + 55, + 336, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 295, + 360 + ], + "type": "text", + "content": "Table 5. Results for audio classification on Speech Commands V2 val [92]. All models are fine-tuned on the training set (§4.5)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 71, + 553, + 177 + ], + "lines": [ + { + "bbox": [ + 321, + 71, + 553, + 177 + ], + "spans": [ + { + "bbox": [ + 321, + 71, + 553, + 177 + ], + "type": "table", + "html": "
ModelBackboneRetrieval ModalityTrainable ParamsR@1R@5
Single Task
MIL-NCE [54]ResNeXt-101V88.8M8.625.8
SupportSet [60]R3D-34V93.5M10.430.0
AVFIC [57]ViT-BA+V86.6M19.450.3
InternVideo [89]ViT-LV305M37.571.3
Multiple Tasks
ImageBind [18]ViT-HA+V89.3M36.870.0
ViT-LENS [38]ViT-LA+V127.6M37.672.6
UNIALIGN†ViT-BA+V7.8M37.070.7
", + "image_path": "3463065951b9026352e388cb6098c84beaf94ed862132a897d39e1038d7df4c9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 320, + 210, + 553, + 304 + ], + "blocks": [ + { + "bbox": [ + 318, + 177, + 552, + 198 + ], + "lines": [ + { + "bbox": [ + 318, + 177, + 552, + 198 + ], + "spans": [ + { + "bbox": [ + 318, + 177, + 552, + 198 + ], + "type": "text", + "content": "†:An identical model with the same weight across all tasks. Table 6. Results for video retrieval on MSRVTT val [96] (§4.6)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 210, + 553, + 304 + ], + "lines": [ + { + "bbox": [ + 320, + 210, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 320, + 210, + 553, + 304 + ], + "type": "table", + "html": "
ModelBackboneZero-ShotFine-Tuned
Single Task
ResT [44]ResNet-10158.7-
ActionCLIP [85]ViT-B58.3-
X-CLIP [58]ViT-B72.0-
Multiple Tasks
MetaFormer [107]ViT-B-46.6
ImageBind [18]ViT-H64.898.1
UniBind [50]ViT-H73.793.3
UNIALIGN†ViT-B68.594.4
", + "image_path": "2dd4a695cf927d3fc6d295b2ca8412e00c15334bc499fb80adffb779ba2fc289.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 304, + 556, + 325 + ], + "lines": [ + { + "bbox": [ + 315, + 304, + 556, + 325 + ], + "spans": [ + { + "bbox": [ + 315, + 304, + 556, + 325 + ], + "type": "text", + "content": "†:An identical model with the same weight across all tasks. Table 7. Results for video classification on UCF101 val[77] (§4.6)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 335, + 555, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 555, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 555, + 358 + ], + "type": "text", + "content": "Metric. Recall@ " + }, + { + "bbox": [ + 313, + 335, + 555, + 358 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 335, + 555, + 358 + ], + "type": "text", + "content": " and top-1 accuracy are adopted as evaluation metrics for video-text retrieval and classification." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 360, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 555, + 407 + ], + "type": "text", + "content": "Performance for Video-Text Retrieval. As shown in Table 6, UNIALIGN achieves a R@1 score of " + }, + { + "bbox": [ + 313, + 360, + 555, + 407 + ], + "type": "inline_equation", + "content": "37.0\\%" + }, + { + "bbox": [ + 313, + 360, + 555, + 407 + ], + "type": "text", + "content": " and R@5 score of " + }, + { + "bbox": [ + 313, + 360, + 555, + 407 + ], + "type": "inline_equation", + "content": "70.7\\%" + }, + { + "bbox": [ + 313, + 360, + 555, + 407 + ], + "type": "text", + "content": ", which are comparable to advanced methods while using merely 7.8M trainable parameters." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 408, + 555, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 555, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 555, + 516 + ], + "type": "text", + "content": "Performance for Video Classification. We summarize the results for video classification in Table 7. As shown, UNIALIGN achieves an impressive accuracy of " + }, + { + "bbox": [ + 313, + 408, + 555, + 516 + ], + "type": "inline_equation", + "content": "68.5\\%" + }, + { + "bbox": [ + 313, + 408, + 555, + 516 + ], + "type": "text", + "content": " on UCF101 val[77] without any fine-tuning, and outperforms ImageBind [18]. After fine-tuning all models on the training set of UCF101, our method achieves the second-highest performance. This may be attributed to the differences in backbone, as ImageBind utilizes ViT-H which could provide enhanced knowledge for downstream fine-tuning." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 525, + 425, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 425, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 425, + 537 + ], + "type": "text", + "content": "4.7. Qualitative Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 542, + 555, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 555, + 578 + ], + "type": "text", + "content": "As shown in Fig. 5, UNIALIGN facilitates accurate cross-modality retrieval. Moreover, modalities can be combined to constrain attributes of the target in another modality." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 586, + 444, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 444, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 444, + 600 + ], + "type": "text", + "content": "4.8. Diagnostic Experiment" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 605, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 640 + ], + "type": "text", + "content": "To evaluate the core designs of UNIALIGN and gain further insights, we conduct a series of ablation studies. Results are reported on SUN-D test [74] and ModelNet40 test [5]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": "Modality-Aware MoE Strategy. As illustrated in §3.1, we explore two modality-aware MoE strategies: expert grouping and global routing. Results in Table 8 indicate that dividing experts into modality-specific groups yields better performance. Thus, we adopt it as the default MoE strategy. Expert Network. Table 9 presents the ablation study on" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29650" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 73, + 91, + 102 + ], + "blocks": [ + { + "bbox": [ + 80, + 73, + 91, + 102 + ], + "lines": [ + { + "bbox": [ + 80, + 73, + 91, + 102 + ], + "spans": [ + { + "bbox": [ + 80, + 73, + 91, + 102 + ], + "type": "image", + "image_path": "4b991b2df21667b4f2c7d7f2e19b0e392cb3317f451eee39701e935e95a6ea8f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 102, + 74, + 141, + 104 + ], + "blocks": [ + { + "bbox": [ + 102, + 74, + 141, + 104 + ], + "lines": [ + { + "bbox": [ + 102, + 74, + 141, + 104 + ], + "spans": [ + { + "bbox": [ + 102, + 74, + 141, + 104 + ], + "type": "image", + "image_path": "83520fce3aad31423abe0e14d9da6379fcd1420d0de4e4d7ad60a12be7325322.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 151, + 74, + 184, + 103 + ], + "blocks": [ + { + "bbox": [ + 151, + 74, + 184, + 103 + ], + "lines": [ + { + "bbox": [ + 151, + 74, + 184, + 103 + ], + "spans": [ + { + "bbox": [ + 151, + 74, + 184, + 103 + ], + "type": "image", + "image_path": "456d96ca12fc305a2e1e3da1f7a95c194c781d09a07d7f272a62ae910185cc81.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 187, + 74, + 226, + 103 + ], + "blocks": [ + { + "bbox": [ + 187, + 74, + 226, + 103 + ], + "lines": [ + { + "bbox": [ + 187, + 74, + 226, + 103 + ], + "spans": [ + { + "bbox": [ + 187, + 74, + 226, + 103 + ], + "type": "image", + "image_path": "c6fd2eda02bf680ceb17bf6002ca3e3b541cac4060260a377ce35b63dc4afdfe.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 261, + 77, + 282, + 92 + ], + "blocks": [ + { + "bbox": [ + 261, + 77, + 282, + 92 + ], + "lines": [ + { + "bbox": [ + 261, + 77, + 282, + 92 + ], + "spans": [ + { + "bbox": [ + 261, + 77, + 282, + 92 + ], + "type": "image", + "image_path": "b3ccf9fc542160290081623aedc6ed43afdd32dc7bb577cb7f184e7e7bc3b67a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 253, + 96, + 290, + 103 + ], + "lines": [ + { + "bbox": [ + 253, + 96, + 290, + 103 + ], + "spans": [ + { + "bbox": [ + 253, + 96, + 290, + 103 + ], + "type": "text", + "content": "Aircraft engine" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 296, + 73, + 335, + 102 + ], + "blocks": [ + { + "bbox": [ + 296, + 73, + 335, + 102 + ], + "lines": [ + { + "bbox": [ + 296, + 73, + 335, + 102 + ], + "spans": [ + { + "bbox": [ + 296, + 73, + 335, + 102 + ], + "type": "image", + "image_path": "471ab080e8a627dce21e4b4181fffc0f2c98c6a7c21a2d8255829b11d2707f07.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 95, + 396, + 103 + ], + "lines": [ + { + "bbox": [ + 354, + 95, + 396, + 103 + ], + "spans": [ + { + "bbox": [ + 354, + 95, + 396, + 103 + ], + "type": "text", + "content": "Playing volleyball" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 370, + 75, + 396, + 86 + ], + "blocks": [ + { + "bbox": [ + 370, + 75, + 396, + 86 + ], + "lines": [ + { + "bbox": [ + 370, + 75, + 396, + 86 + ], + "spans": [ + { + "bbox": [ + 370, + 75, + 396, + 86 + ], + "type": "image", + "image_path": "6810b160a4190e4bdce464dc96fb2d2316f5bcb250fdad03e5b71f90fd73a280.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 406, + 73, + 445, + 102 + ], + "blocks": [ + { + "bbox": [ + 406, + 73, + 445, + 102 + ], + "lines": [ + { + "bbox": [ + 406, + 73, + 445, + 102 + ], + "spans": [ + { + "bbox": [ + 406, + 73, + 445, + 102 + ], + "type": "image", + "image_path": "0ae38f0368dbcdec74429c714f72e20ad2eeb06424a5ec06f3f2418eafb48824.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 471, + 77, + 492, + 91 + ], + "blocks": [ + { + "bbox": [ + 471, + 77, + 492, + 91 + ], + "lines": [ + { + "bbox": [ + 471, + 77, + 492, + 91 + ], + "spans": [ + { + "bbox": [ + 471, + 77, + 492, + 91 + ], + "type": "image", + "image_path": "f0e430a016852f7b3c47c1fbd96646c2b3401a0df95fa3e5b9389d5f708aad35.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 475, + 97, + 487, + 104 + ], + "lines": [ + { + "bbox": [ + 475, + 97, + 487, + 104 + ], + "spans": [ + { + "bbox": [ + 475, + 97, + 487, + 104 + ], + "type": "text", + "content": "Dog" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 499, + 73, + 539, + 103 + ], + "blocks": [ + { + "bbox": [ + 499, + 73, + 539, + 103 + ], + "lines": [ + { + "bbox": [ + 499, + 73, + 539, + 103 + ], + "spans": [ + { + "bbox": [ + 499, + 73, + 539, + 103 + ], + "type": "image", + "image_path": "3afd6354d70e4a60eff22c65112d101bbc8c2e07b4d5dbcdda88d0ff298d7664.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 69, + 111, + 102, + 138 + ], + "blocks": [ + { + "bbox": [ + 69, + 111, + 102, + 138 + ], + "lines": [ + { + "bbox": [ + 69, + 111, + 102, + 138 + ], + "spans": [ + { + "bbox": [ + 69, + 111, + 102, + 138 + ], + "type": "image", + "image_path": "843190b71736fa50333b9a41a8ddb607ddb61f0f3418f1997d539820dba8c77f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 139, + 97, + 144 + ], + "lines": [ + { + "bbox": [ + 74, + 139, + 97, + 144 + ], + "spans": [ + { + "bbox": [ + 74, + 139, + 97, + 144 + ], + "type": "text", + "content": "Bathroom" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 106, + 112, + 135, + 141 + ], + "blocks": [ + { + "bbox": [ + 106, + 112, + 135, + 141 + ], + "lines": [ + { + "bbox": [ + 106, + 112, + 135, + 141 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 135, + 141 + ], + "type": "image", + "image_path": "aa2d3368c3f33b2e83c4cd06a90783839eec841cb5b05c73ee59037f9769485b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 136, + 179, + 143 + ], + "lines": [ + { + "bbox": [ + 156, + 136, + 179, + 143 + ], + "spans": [ + { + "bbox": [ + 156, + 136, + 179, + 143 + ], + "type": "text", + "content": "Car horn" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 159, + 113, + 180, + 127 + ], + "blocks": [ + { + "bbox": [ + 159, + 113, + 180, + 127 + ], + "lines": [ + { + "bbox": [ + 159, + 113, + 180, + 127 + ], + "spans": [ + { + "bbox": [ + 159, + 113, + 180, + 127 + ], + "type": "image", + "image_path": "2c1080163687ece604b2122bf557fb934620bfe5c7ec6559fb878aa1a19f4a59.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 188, + 113, + 225, + 140 + ], + "blocks": [ + { + "bbox": [ + 188, + 113, + 225, + 140 + ], + "lines": [ + { + "bbox": [ + 188, + 113, + 225, + 140 + ], + "spans": [ + { + "bbox": [ + 188, + 113, + 225, + 140 + ], + "type": "image", + "image_path": "23c6e52cf2bf9f3e379bd47312c6ca1f70e947cfb2bf6b1cf14a8e3d502a3675.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 261, + 112, + 280, + 128 + ], + "blocks": [ + { + "bbox": [ + 261, + 112, + 280, + 128 + ], + "lines": [ + { + "bbox": [ + 261, + 112, + 280, + 128 + ], + "spans": [ + { + "bbox": [ + 261, + 112, + 280, + 128 + ], + "type": "image", + "image_path": "4e5526b97456a54eb70408a15ce25f6d11d81e320a0b974b55edf620fcde61e1.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 135, + 292, + 142 + ], + "lines": [ + { + "bbox": [ + 258, + 135, + 292, + 142 + ], + "spans": [ + { + "bbox": [ + 258, + 135, + 292, + 142 + ], + "type": "text", + "content": "Playing piano" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 304, + 115, + 337, + 142 + ], + "blocks": [ + { + "bbox": [ + 304, + 115, + 337, + 142 + ], + "lines": [ + { + "bbox": [ + 304, + 115, + 337, + 142 + ], + "spans": [ + { + "bbox": [ + 304, + 115, + 337, + 142 + ], + "type": "image", + "image_path": "61ad3d00d9d8dcb8e42597099ef7f8e9189e0657b3b13a19bb371717b537728f.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 371, + 113, + 391, + 129 + ], + "blocks": [ + { + "bbox": [ + 371, + 113, + 391, + 129 + ], + "lines": [ + { + "bbox": [ + 371, + 113, + 391, + 129 + ], + "spans": [ + { + "bbox": [ + 371, + 113, + 391, + 129 + ], + "type": "image", + "image_path": "d45979826f20a9f199d0319e0a3f1fc6028e9c0e717e8fc93a53d98516e50f1b.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 130, + 404, + 144 + ], + "lines": [ + { + "bbox": [ + 358, + 130, + 404, + 144 + ], + "spans": [ + { + "bbox": [ + 358, + 130, + 404, + 144 + ], + "type": "text", + "content": "Typing on computer keyboard" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 413, + 112, + 447, + 137 + ], + "blocks": [ + { + "bbox": [ + 413, + 112, + 447, + 137 + ], + "lines": [ + { + "bbox": [ + 413, + 112, + 447, + 137 + ], + "spans": [ + { + "bbox": [ + 413, + 112, + 447, + 137 + ], + "type": "image", + "image_path": "be6eeea0e7e1d7e64f4839649e660caefb91774ef10f9f5bdb3d54f1aae247e4.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 137, + 452, + 144 + ], + "lines": [ + { + "bbox": [ + 410, + 137, + 452, + 144 + ], + "spans": [ + { + "bbox": [ + 410, + 137, + 452, + 144 + ], + "type": "text", + "content": "Computer room" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 470, + 115, + 493, + 126 + ], + "blocks": [ + { + "bbox": [ + 470, + 115, + 493, + 126 + ], + "lines": [ + { + "bbox": [ + 470, + 115, + 493, + 126 + ], + "spans": [ + { + "bbox": [ + 470, + 115, + 493, + 126 + ], + "type": "image", + "image_path": "0aab9ec8bbf4cb72f6ef7b17757ec22ea3244c74b720d4792b52481b9c52a745.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 129, + 498, + 144 + ], + "lines": [ + { + "bbox": [ + 468, + 129, + 498, + 144 + ], + "spans": [ + { + "bbox": [ + 468, + 129, + 498, + 144 + ], + "type": "text", + "content": "Dishes, pots, and pans" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 506, + 112, + 539, + 137 + ], + "blocks": [ + { + "bbox": [ + 506, + 112, + 539, + 137 + ], + "lines": [ + { + "bbox": [ + 506, + 112, + 539, + 137 + ], + "spans": [ + { + "bbox": [ + 506, + 112, + 539, + 137 + ], + "type": "image", + "image_path": "9f6cf2dfb1e362bfd37d8436320fc7507f52ed24e340401ef7d948ccdbb17fdc.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 512, + 137, + 532, + 143 + ], + "lines": [ + { + "bbox": [ + 512, + 137, + 532, + 143 + ], + "spans": [ + { + "bbox": [ + 512, + 137, + 532, + 143 + ], + "type": "text", + "content": "Kitchen" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 63, + 152, + 105, + 182 + ], + "blocks": [ + { + "bbox": [ + 63, + 152, + 105, + 182 + ], + "lines": [ + { + "bbox": [ + 63, + 152, + 105, + 182 + ], + "spans": [ + { + "bbox": [ + 63, + 152, + 105, + 182 + ], + "type": "image", + "image_path": "740dcf664438f4b51884a15ffec731ceefe44a0d12d670a7f1c1f461c1e7a078.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 186, + 553, + 209 + ], + "lines": [ + { + "bbox": [ + 55, + 186, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 553, + 209 + ], + "type": "text", + "content": "Figure 5. Qualitative results for cross-modal retrieval. Top and middle: retrieval using one single modality. Bottom: multiple modalities can be composed to retrieve targets across different modalities (§4.7)." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 105, + 156, + 160, + 182 + ], + "blocks": [ + { + "bbox": [ + 105, + 156, + 160, + 182 + ], + "lines": [ + { + "bbox": [ + 105, + 156, + 160, + 182 + ], + "spans": [ + { + "bbox": [ + 105, + 156, + 160, + 182 + ], + "type": "image", + "image_path": "e525cb2760c631dd9ccba1d8d525b57a01c5934b89763f8d7a5f2d45451f4952.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 162, + 152, + 203, + 182 + ], + "blocks": [ + { + "bbox": [ + 162, + 152, + 203, + 182 + ], + "lines": [ + { + "bbox": [ + 162, + 152, + 203, + 182 + ], + "spans": [ + { + "bbox": [ + 162, + 152, + 203, + 182 + ], + "type": "image", + "image_path": "592482f04f90d5e62f8370ad6295c724381c506d5652fc7ebf90a19e5eefe589.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 246, + 153, + 276, + 178 + ], + "blocks": [ + { + "bbox": [ + 246, + 153, + 276, + 178 + ], + "lines": [ + { + "bbox": [ + 246, + 153, + 276, + 178 + ], + "spans": [ + { + "bbox": [ + 246, + 153, + 276, + 178 + ], + "type": "image", + "image_path": "22b61ebf4167432c93440d5caf117d5df85beac73bbff9d745ef0948c90a2f29.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 282, + 156, + 324, + 178 + ], + "blocks": [ + { + "bbox": [ + 282, + 156, + 324, + 178 + ], + "lines": [ + { + "bbox": [ + 282, + 156, + 324, + 178 + ], + "spans": [ + { + "bbox": [ + 282, + 156, + 324, + 178 + ], + "type": "image", + "image_path": "90e1f03ce63ec3ed24a28e06a5f6a45976f6127af758f597ac045b20c1796990.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 331, + 152, + 370, + 182 + ], + "blocks": [ + { + "bbox": [ + 331, + 152, + 370, + 182 + ], + "lines": [ + { + "bbox": [ + 331, + 152, + 370, + 182 + ], + "spans": [ + { + "bbox": [ + 331, + 152, + 370, + 182 + ], + "type": "image", + "image_path": "54d555bb69c8f39999ea89ce05cf8914c4f397bff25a6799064a35975a27b5f5.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 417, + 152, + 440, + 183 + ], + "blocks": [ + { + "bbox": [ + 417, + 152, + 440, + 183 + ], + "lines": [ + { + "bbox": [ + 417, + 152, + 440, + 183 + ], + "spans": [ + { + "bbox": [ + 417, + 152, + 440, + 183 + ], + "type": "image", + "image_path": "3dcfbc1572996109a1295e77a840ba0fcb14ac3cb863464ec96ffab282711614.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 449, + 152, + 504, + 182 + ], + "blocks": [ + { + "bbox": [ + 449, + 152, + 504, + 182 + ], + "lines": [ + { + "bbox": [ + 449, + 152, + 504, + 182 + ], + "spans": [ + { + "bbox": [ + 449, + 152, + 504, + 182 + ], + "type": "image", + "image_path": "fae369fdc26c804450f832f8eb6084c854083b3aa99ae8c01916775787211568.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 506, + 152, + 539, + 174 + ], + "blocks": [ + { + "bbox": [ + 506, + 152, + 539, + 174 + ], + "lines": [ + { + "bbox": [ + 506, + 152, + 539, + 174 + ], + "spans": [ + { + "bbox": [ + 506, + 152, + 539, + 174 + ], + "type": "image", + "image_path": "18afea5e8487fd595d186cbf68d74ed6a48c92152fa973e3588737c150eabf10.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 514, + 175, + 532, + 181 + ], + "lines": [ + { + "bbox": [ + 514, + 175, + 532, + 181 + ], + "spans": [ + { + "bbox": [ + 514, + 175, + 532, + 181 + ], + "type": "text", + "content": "Office" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "table", + "bbox": [ + 62, + 219, + 294, + 255 + ], + "blocks": [ + { + "bbox": [ + 62, + 219, + 294, + 255 + ], + "lines": [ + { + "bbox": [ + 62, + 219, + 294, + 255 + ], + "spans": [ + { + "bbox": [ + 62, + 219, + 294, + 255 + ], + "type": "table", + "html": "
#StrategiesSUN-DModelNet40
1Expert Grouping45.855.9
2Global Routing45.055.3
", + "image_path": "93fb001fd47af195a6ccc0f2ae6655ac33632b6eba282d27723f293a1550fc07.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "table_body" + } + ], + "index": 41 + }, + { + "type": "table", + "bbox": [ + 63, + 277, + 294, + 354 + ], + "blocks": [ + { + "bbox": [ + 64, + 257, + 285, + 269 + ], + "lines": [ + { + "bbox": [ + 64, + 257, + 285, + 269 + ], + "spans": [ + { + "bbox": [ + 64, + 257, + 285, + 269 + ], + "type": "text", + "content": "Table 8. Ablation studies on modality-aware MoE strategies." + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 277, + 294, + 354 + ], + "lines": [ + { + "bbox": [ + 63, + 277, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 63, + 277, + 294, + 354 + ], + "type": "table", + "html": "
Top-k#ExpertsInit.GPU MemSUN-DModelNet40
11rand27.2G45.054.8
11CLIP27.2G45.255.1
12rand27.4G45.155.2
12CLIP27.4G45.255.4
22rand35.9G45.755.7
22CLIP35.9G45.855.9
", + "image_path": "f92a8bca9046537c7713167c971733885e03844d1e90f14b9028ba73fa161802.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "table_body" + } + ], + "index": 43 + }, + { + "bbox": [ + 55, + 390, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 295, + 677 + ], + "type": "text", + "content": "expert network configurations. We investigate the impact of top-" + }, + { + "bbox": [ + 55, + 390, + 295, + 677 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 390, + 295, + 677 + ], + "type": "text", + "content": " routing, the number of experts per modality, and weight initialization. Starting with a baseline configuration of one expert per modality and random initialization, we achieve 45.0 on SUN-D test [74] and 54.8 on ModelNet40 test [5], with 27.2G GPU memory consumption. Increasing the expert number and initializing weights from the visual encoder of CLIP both lead to improved performance. Due to the constraints of GPU memory, we use two experts per modality and top-2 routing for expert selection. Training Objectives. Next we probe the impact of training objectives i.e., soft modality binding with " + }, + { + "bbox": [ + 55, + 390, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{soft\\_bind}}" + }, + { + "bbox": [ + 55, + 390, + 295, + 677 + ], + "type": "text", + "content": " (cf., Eq. 11), as well as knowledge distillation through " + }, + { + "bbox": [ + 55, + 390, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{anch\\_distil}}" + }, + { + "bbox": [ + 55, + 390, + 295, + 677 + ], + "type": "text", + "content": " (cf., Eq. 12) and " + }, + { + "bbox": [ + 55, + 390, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{model\\_distil}}" + }, + { + "bbox": [ + 55, + 390, + 295, + 677 + ], + "type": "text", + "content": " (cf., Eq. 15). Results are summarized in Table 10, where the first row refers to the baseline that utilizes only paired data from single datasets and is trained with the standard InfoNCE loss. It can be observed that: First, our " + }, + { + "bbox": [ + 55, + 390, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{soft\\_bind}}" + }, + { + "bbox": [ + 55, + 390, + 295, + 677 + ], + "type": "text", + "content": " successfully bridges the learning of multiple modalities for mutual boosting, leads to improved performance on both datasets. Second, " + }, + { + "bbox": [ + 55, + 390, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{anch\\_distil}}" + }, + { + "bbox": [ + 55, + 390, + 295, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 390, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{model\\_distil}}" + }, + { + "bbox": [ + 55, + 390, + 295, + 677 + ], + "type": "text", + "content": " could further boost the performance with compelling gains. This verifies the effectiveness of our knowledge distillation strategies, and leads UNIALIGN to a high performance model with optimized efficiency." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Modality Scaling. Finally, we examine the robustness of UNIALIGN to scale different numbers of modalities into a unified model. As shown in Table 11, modality-independent" + } + ] + } + ], + "index": 46 + }, + { + "type": "table", + "bbox": [ + 321, + 220, + 553, + 276 + ], + "blocks": [ + { + "bbox": [ + 55, + 357, + 295, + 380 + ], + "lines": [ + { + "bbox": [ + 55, + 357, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 295, + 380 + ], + "type": "text", + "content": "Table 9. Ablation studies on configuration of expert network. GPU memory is reported with a batch size of 448 on one Tesla A100." + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 220, + 553, + 276 + ], + "lines": [ + { + "bbox": [ + 321, + 220, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 321, + 220, + 553, + 276 + ], + "type": "table", + "html": "
L_NCELsoft_bindLanch_distilLmodel_distilSUN-DModelNet40
✓×××40.438.6
✓✓××41.840.1
✓✓✓×42.340.9
✓✓✓✓45.855.9
", + "image_path": "a85ca3388f521e0c20408c817886f93d44c18777b364824874d193917561c2d5.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "table_body" + } + ], + "index": 47 + }, + { + "type": "table", + "bbox": [ + 321, + 300, + 553, + 363 + ], + "blocks": [ + { + "bbox": [ + 344, + 279, + 523, + 290 + ], + "lines": [ + { + "bbox": [ + 344, + 279, + 523, + 290 + ], + "spans": [ + { + "bbox": [ + 344, + 279, + 523, + 290 + ], + "type": "text", + "content": "Table 10. Ablation studies on training objectives." + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 300, + 553, + 363 + ], + "lines": [ + { + "bbox": [ + 321, + 300, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 321, + 300, + 553, + 363 + ], + "type": "table", + "html": "
#Pre-Training ModalityImageNet-1KSUN-DModelNet40
1I + D + P84.734.431.6
2I + T + D + P84.939.938.1
3I + T + D + P + V85.140.338.4
4I + T + D + P + V + A85.140.438.6
5I + T + D + P + V + A *85.345.855.9
", + "image_path": "f76d96fbb1a70a27f16d673e693ca317fc1afd95d27adbbcbc59feb75e137c03.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "table_body" + } + ], + "index": 49 + }, + { + "bbox": [ + 313, + 367, + 554, + 411 + ], + "lines": [ + { + "bbox": [ + 313, + 367, + 554, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 554, + 411 + ], + "type": "text", + "content": "Table 11. Ablation studies on modality scaling. Results are reported using only " + }, + { + "bbox": [ + 313, + 367, + 554, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{NCE}}" + }, + { + "bbox": [ + 313, + 367, + 554, + 411 + ], + "type": "text", + "content": " as the default training loss. * denotes incorporate " + }, + { + "bbox": [ + 313, + 367, + 554, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{soft\\_bind}}" + }, + { + "bbox": [ + 313, + 367, + 554, + 411 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 367, + 554, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{anch\\_distil}}" + }, + { + "bbox": [ + 313, + 367, + 554, + 411 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 367, + 554, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{model\\_distil}}" + }, + { + "bbox": [ + 313, + 367, + 554, + 411 + ], + "type": "text", + "content": " into training. I: Image, T: Text, D: Depth, V: Video, A: Audio, P: 3D Point Cloud." + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "content": "training with merely " + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{NCE}}" + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "content": " leads to a performance degeneration as more modalities integrated (i.e., row #1-4). This suggests that a straightforward unification of multiple modalities may yield sub-optimal performance for individual ones. Nevertheless, soft modality binding and knowledge distillation (i.e., row #4 vs. #5) could mitigate this negative impact, and boost performance to a level surpassing that achieved with fewer modalities (i.e., row #1 vs. #5)." + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 313, + 536, + 388, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 388, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 388, + 548 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": "In this work, we focus on the all-in-one paradigm for multimodal alignment, which yields a highly efficient and effective model, UNIALIGN, to unify the alignment of all anticipated modalities within one network. Inspired by MoE and knowledge distillation, we propose a modality-aware MoE to tackle multimodal inputs, and generally applicable alignment strategies to utilize unpaired data across datasets. Empirical results suggest that UNIALIGN achieves remarkable performance while being efficient in both training (7.8M trainable parameters to handle 6 modalities within a single training phase) and inference (one identical model with consistent weights across tasks). This work may potentially advance the development of compact, multimodal models." + } + ] + } + ], + "index": 53 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29651" + } + ] + } + ], + "index": 54 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 297, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 73, + 297, + 173 + ], + "spans": [ + { + "bbox": [ + 55, + 73, + 297, + 173 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by the National Science and Technology Major Project (No. 2023ZD0121300), the National Natural Science Foundation of China (No. 62472222, 62302217), Natural Science Foundation of Jiangsu Province (No. BK20240080, BK20220936), \"Pioneer\" and \"Leading Goose\" R&D Program of Zhejiang (No. 2024C01161), the National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, Xi'an Jiaotong University (No. HMHAI-202403), and Bytedance Doubao Fund." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 184, + 115, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 115, + 197 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 115, + 197 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 203, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 203, + 294, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 203, + 294, + 236 + ], + "spans": [ + { + "bbox": [ + 66, + 203, + 294, + 236 + ], + "type": "text", + "content": "[1] Romero Adriana, Ballas Nicolas, K Samira Ebrahimi, Chassang Antoine, Gatta Carlo, and Bengio Yoshua. Fitnets: Hints for thin deep nets. In ICLR, 2015. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 238, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 238, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 66, + 238, + 294, + 281 + ], + "type": "text", + "content": "[2] Lucas Beyer, Xiaohua Zhai, Amélie Royer, Larisa Markeeva, Rohan Anil, and Alexander Kolesnikov. Knowledge distillation: A good teacher is patient and consistent. In CVPR, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 284, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 284, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 66, + 284, + 294, + 316 + ], + "type": "text", + "content": "[3] Xinhao Cai, Qiqia Lai, Yuwei Wang, Wenguan Wang, Zeren Sun, and Yazhou Yao. Poly kernel inception network for remote sensing detection. In CVPR, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 318, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 318, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 66, + 318, + 294, + 338 + ], + "type": "text", + "content": "[4] Rich Caruana. Multitask learning. Machine learning, 28: 41-75, 1997. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 341, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 341, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 66, + 341, + 294, + 396 + ], + "type": "text", + "content": "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 397, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 397, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 66, + 397, + 294, + 430 + ], + "type": "text", + "content": "[6] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP, 2020. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 431, + 294, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 431, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 66, + 431, + 294, + 474 + ], + "type": "text", + "content": "[7] Shaoxiang Chen, Zequn Jie, and Lin Ma. Llava-MoLE: Sparse Mixture of LoRA Experts for Mitigating Data Conflicts in Instruction Finetuning MLLMs. arXiv preprint arXiv:2401.16160, 2024.3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 476, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 476, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 66, + 476, + 294, + 509 + ], + "type": "text", + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 510, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 510, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 66, + 510, + 294, + 542 + ], + "type": "text", + "content": "[9] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. In NeurIPS, 2024. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 544, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 544, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 294, + 599 + ], + "type": "text", + "content": "[10] Xiaohan Ding, Yiyuan Zhang, Yixiao Ge, Sijie Zhao, Lin Song, Xiangyu Yue, and Ying Shan. Unireplknet: A universal perception large-kernel convnet for audio video point cloud time-series and image recognition. In CVPR, 2024. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 601, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 601, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 62, + 601, + 294, + 655 + ], + "type": "text", + "content": "[11] Shihan Dou, Enyu Zhou, Yan Liu, Songyang Gao, Jun Zhao, Wei Shen, Yuhao Zhou, Zhiheng Xi, Xiao Wang, Xiaoran Fan, et al. Loramoe: Revolutionizing mixture of experts for maintaining world knowledge in language model alignment. arXiv preprint arXiv:2312.09979, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 657, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 657, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 657, + 294, + 689 + ], + "type": "text", + "content": "[12] Dapeng Du, Limin Wang, Huiling Wang, Kai Zhao, and Gangshan Wu. Translate-to-recognize networks for rgb-d scene recognition. In CVPR, 2019. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 62, + 691, + 294, + 714 + ], + "type": "text", + "content": "[13] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "text", + "content": "Cao. Eva: Exploring the limits of masked visual representation learning at scale. In CVPR, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 97, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 97, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 321, + 97, + 553, + 140 + ], + "type": "text", + "content": "[14] Bilal Faye, Hanane Azzag, and Mustapha Lebbah. Oneencoder: A Lightweight Framework for Progressive Alignment of Modalities. arXiv preprint arXiv:2409.11059, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 143, + 553, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 143, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 320, + 143, + 553, + 186 + ], + "type": "text", + "content": "[15] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1-39, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 188, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 188, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 321, + 188, + 553, + 241 + ], + "type": "text", + "content": "[16] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In ICASSP, 2017. 2, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 244, + 553, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 244, + 553, + 288 + ], + "spans": [ + { + "bbox": [ + 321, + 244, + 553, + 288 + ], + "type": "text", + "content": "[17] Rohit Girdhar, Mannat Singh, Nikhila Ravi, Laurens van der Maaten, Armand Joulin, and Ishan Misra. Omnivore: A Single Model for Many Visual Modalities. In CVPR, 2022. 2, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 290, + 553, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 290, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 321, + 290, + 553, + 334 + ], + "type": "text", + "content": "[18] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In CVPR, 2023. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 335, + 553, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 335, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 321, + 335, + 553, + 357 + ], + "type": "text", + "content": "[19] Yuan Gong, Yu-An Chung, and James Glass. AST: Audio Spectrogram Transformer. In Interspeech, 2021. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 359, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 359, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 321, + 359, + 553, + 392 + ], + "type": "text", + "content": "[20] Yuan Gong, Cheng-I Lai, Yu-An Chung, and James R. Glass. Ssast: Self-Supervised Audio Spectrogram Transformer. In AAAI, 2022. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 394, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 394, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 321, + 394, + 553, + 449 + ], + "type": "text", + "content": "[21] Yunhao Gou, Zhili Liu, Kai Chen, Lanqing Hong, Hang Xu, Aoxue Li, Dit-Yan Yeung, James T Kwok, and Yu Zhang. Mixture of cluster-conditional lora experts for vision-language instruction tuning. arXiv preprint arXiv:2312.12379, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 450, + 553, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 450, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 321, + 450, + 553, + 483 + ], + "type": "text", + "content": "[22] Andrey Guzhov, Federico Raue, Jorn Hees, and Andreas Dengel. Audioclip: Extending clip to image, text and audio. In ICASSP, 2022. 1, 5, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 485, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 485, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 321, + 485, + 553, + 529 + ], + "type": "text", + "content": "[23] Jiaming Han, Kaixiong Gong, Yiyuan Zhang, Jiaqi Wang, Kaipeng Zhang, Dahua Lin, Yu Qiao, Peng Gao, and Xi-angyu Yue. Onellm: One Framework to Align All Modalities with Language. In CVPR, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 531, + 553, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 531, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 321, + 531, + 553, + 563 + ], + "type": "text", + "content": "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 321, + 565, + 553, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 565, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 321, + 565, + 553, + 597 + ], + "type": "text", + "content": "[25] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, Nojun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In ICCV, 2019. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 321, + 600, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 600, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 321, + 600, + 553, + 622 + ], + "type": "text", + "content": "[26] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 321, + 624, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 624, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 321, + 624, + 553, + 666 + ], + "type": "text", + "content": "[27] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2021. 2, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 321, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 321, + 670, + 553, + 712 + ], + "type": "text", + "content": "[28] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 3, 4" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29652" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "text", + "content": "[29] Chengsong Huang, Qian Liu, Bill Yuchen Lin, Tianyu Pang, Chao Du, and Min Lin. Lorahub: Efficient cross-task generalization via dynamic lora composition. arXiv preprint arXiv:2307.13269, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 294, + 161 + ], + "type": "text", + "content": "[30] Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson W.H. Lau, Wanli Ouyang, and Wangmeng Zuo. Clip2point: Transfer clip to point cloud classification with image-depth pre-training. In ICCV, 2023. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 162, + 295, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 162, + 295, + 215 + ], + "spans": [ + { + "bbox": [ + 62, + 162, + 295, + 215 + ], + "type": "text", + "content": "[31] Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. Mixtral of experts. arXiv preprint arXiv:2401.04088, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 217, + 294, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 217, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 217, + 294, + 258 + ], + "type": "text", + "content": "[32] Byoungjip Kim, Sungik Choi, Dasol Hwang, Moontae Lee, and Honglak Lee. Transferring pre-trained multimodal representations with cross-modal similarity matching. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 261, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 294, + 293 + ], + "type": "text", + "content": "[33] Chris Dongjoo Kim, Byeongchang Kim, Hyunmin Lee, and Gunhee Kim. Audiocaps: Generating captions for audios in the wild. In NAACL-HLT, 2019. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 295, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 295, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 62, + 295, + 294, + 326 + ], + "type": "text", + "content": "[34] Jangho Kim, SeongUk Park, and Nojun Kwak. Paraphrasing complex network: Network compression via factor transfer. In NeurIPS, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 327, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 327, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 62, + 327, + 294, + 380 + ], + "type": "text", + "content": "[35] Aran Komatsuzaki, Joan Puigcerver, James Lee-Thorp, Carlos Riquelme Ruiz, Basil Mustafa, Joshua Ainslie, Yi Tay, Mostafa Dehghani, and Neil Houlsby. Sparse upcycling: Training mixture-of-experts from dense checkpoints. In ICLR, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 383, + 294, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 383, + 294, + 424 + ], + "spans": [ + { + "bbox": [ + 62, + 383, + 294, + 424 + ], + "type": "text", + "content": "[36] Nikos Komodakis and Sergey Zagoruyko. Paying more attention to attention: improving the performance of convolutional neural networks via attention transfer. In ICLR, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 426, + 294, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 426, + 294, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 426, + 294, + 468 + ], + "type": "text", + "content": "[37] Konrad P Kording, Ulrik Beierholm, Wei Ji Ma, Steven Quartz, Joshua B Tenenbaum, and Ladan Shams. Causal inference in multisensory perception. PLoS one, 2(9):e943, 2007. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 471, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 471, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 62, + 471, + 294, + 514 + ], + "type": "text", + "content": "[38] Stan Weixian Lei, Yixiao Ge, Kun Yi, Jianfeng Zhang, Difei Gao, Dylan Sun, Yuying Ge, Ying Shan, and Mike Zheng Shou. Vit-Lens: Towards Omni-modal Representations. In CVPR, 2024. 2, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 515, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 515, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 515, + 294, + 557 + ], + "type": "text", + "content": "[39] Weixian Lei, Yixiao Ge, Kun Yi, Jianfeng Zhang, Difei Gao, Dylan Sun, Yuying Ge, Ying Shan, and Mike Zheng Shou. Vit-lens: Towards omni-modal representations. In CVPR, 2024. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 559, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 294, + 591 + ], + "type": "text", + "content": "[40] Po-han Li, Sandeep P Chinchali, and Ufuk Topcu. Csa: Data-efficient mapping of unimodal features to multimodal features. arXiv preprint arXiv:2410.07610, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 593, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 593, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 593, + 295, + 635 + ], + "type": "text", + "content": "[41] Yunxin Li, Shenyuan Jiang, Baotian Hu, Longyue Wang, Wanqi Zhong, Wenhan Luo, Lin Ma, and Min Zhang. UniMoE: Scaling Unified Multimodal LLMs with Mixture of Experts. arXiv, abs/2405.11273, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 636, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 636, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 636, + 294, + 689 + ], + "type": "text", + "content": "[42] Valerii Likhosherstov, Anurag Arnab, Krzysztof Choromanski, Mario Lucic, Yi Tay, Adrian Weller, and Mostafa Dehghani. Polyvit: Co-training vision transformers on images, videos and audio. arXiv preprint arXiv:2111.12993, 2021. 3, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "type": "text", + "content": "[43] Bin Lin, Zhenyu Tang, Yang Ye, Jiaxi Cui, Bin Zhu, Peng Jin, Jina Huang, Junwu Zhang, Yatian Pang, Munan Ning," + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 338, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 338, + 73, + 553, + 105 + ], + "type": "text", + "content": "et al. Moe-llava: Mixture of experts for large vision-language models. arXiv preprint arXiv:2401.15947, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 321, + 107, + 553, + 140 + ], + "type": "text", + "content": "[44] Chung-Ching Lin, Kevin Lin, Lijuan Wang, Zicheng Liu, and Linjie Li. Cross-modal representation learning for zero-shot action recognition. In CVPR, 2022. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 141, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 141, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 320, + 141, + 553, + 185 + ], + "type": "text", + "content": "[45] Haokun Lin, Haoli Bai, Zhili Liu, Lu Hou, Muyi Sun, Linqi Song, Ying Wei, and Zhenan Sun. Mope-CLIP: Structured Pruning for Efficient Vision-Language Models with Module-wise Pruning Error Metric. In CVPR, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 186, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 186, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 320, + 186, + 553, + 240 + ], + "type": "text", + "content": "[46] Xi Victoria Lin, Akshit Shrivastava, Liang Luo, Srinivasan Iyer, Mike Lewis, Gargi Gosh, Luke Zettlemoyer, and Armen Aghajanyan. Moma: Efficient early-fusion pretraining with mixture of modality-aware experts. arXiv preprint arXiv:2407.21770, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 241, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 241, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 320, + 241, + 553, + 285 + ], + "type": "text", + "content": "[47] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 286, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 286, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 320, + 286, + 553, + 319 + ], + "type": "text", + "content": "[48] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 321, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 321, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 320, + 321, + 553, + 364 + ], + "type": "text", + "content": "[49] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 365, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 365, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 320, + 365, + 553, + 398 + ], + "type": "text", + "content": "[50] Yuanhuiyi Lyu, Xu Zheng, Jiazhou Zhou, and Lin Wang. Unibind: Llm-Augmented Unified and Balanced Representation Space to Bind Them All. In CVPR, 2024. 2, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 399, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 399, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 320, + 399, + 553, + 432 + ], + "type": "text", + "content": "[51] Jian Ma, Wenguan Wang, Yi Yang, and Feng Zheng. Ms2sl: Multimodal spoken data-driven continuous sign language production. In ACL Findings, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 434, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 434, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 320, + 434, + 553, + 477 + ], + "type": "text", + "content": "[52] Wenxuan Ma, Shuang Li, Jinming Zhang, Chi Harold Liu, Jingxuan Kang, Yulin Wang, and Gao Huang. Borrowing Knowledge From Pre-trained Language Model: A New Data-efficient Visual Learning Paradigm. In ICCV, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 479, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 479, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 320, + 479, + 553, + 521 + ], + "type": "text", + "content": "[53] M Alex Meredith and Barry E Stein. Visual, auditory, and somatosensory convergence on cells in superior colliculus results in multisensory integration. Journal of neurophysiology, 56(3):640-662, 1986. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 524, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 524, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 320, + 524, + 553, + 577 + ], + "type": "text", + "content": "[54] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, 2019. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 579, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 579, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 320, + 579, + 553, + 622 + ], + "type": "text", + "content": "[55] Seyed Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved knowledge distillation via teacher assistant. In AAAI, 2020. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 624, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 624, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 320, + 624, + 553, + 667 + ], + "type": "text", + "content": "[56] Luca Moschella, Valentino Maiorca, Marco Fumero, Antonio Norelli, Francesco Locatello, and Emanuele Rodola. Relative representations enable zero-shot latent space communication. In ICLR, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 320, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 320, + 670, + 553, + 712 + ], + "type": "text", + "content": "[57] Arsha Nagrani, Paul Hongsuck Seo, Bryan Seybold, Anja Hauth, Santiago Manen, Chen Sun, and Cordelia Schmid. Learning audio-video modalities from image captions. In ECCV, 2022. 7" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "29653" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 127 + ], + "type": "text", + "content": "[58] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. In European Conference on Computer Vision, pages 1-18. Springer, 2022. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 128, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 128, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 61, + 128, + 294, + 172 + ], + "type": "text", + "content": "[59] Antonio Norelli, Marco Fumero, Valentino Maiorca, Luca Moschella, Emanuele Rodola, and Francesco Locatello. Asif: Coupled data turns unimodal models to multimodal without training. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 172, + 295, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 172, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 62, + 172, + 295, + 216 + ], + "type": "text", + "content": "[60] Mandela Patrick, Po-Yao Huang, Yuki Asano, Florian Metze, Alexander Hauptmann, Joao Henriques, and Andrea Vedaldi. Support-set bottlenecks for video-text representation learning. arXiv preprint arXiv:2010.02824, 2020. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 217, + 294, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 217, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 217, + 294, + 258 + ], + "type": "text", + "content": "[61] Gensheng Pei, Fumin Shen, Yazhou Yao, Guo-Sen Xie, Zhenmin Tang, and Jinhui Tang. Hierarchical feature alignment network for unsupervised video object segmentation. In ECCV, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 261, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 294, + 293 + ], + "type": "text", + "content": "[62] Gensheng Pei, Tao Chen, Xiruo Jiang, Huafeng Liu, Zeren Sun, and Yazhou Yao. Videomac: Video masked autoencoders meet convnets. In CVPR, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 294, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 294, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 62, + 294, + 294, + 337 + ], + "type": "text", + "content": "[63] Renjing Pei, Jianzhuang Liu, Weimian Li, Bin Shao, Songcen Xu, Peng Dai, Juwei Lu, and Youliang Yan. Clipping: Distilling clip-based models with a student base for video-language retrieval. In CVPR, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 338, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 338, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 62, + 338, + 294, + 381 + ], + "type": "text", + "content": "[64] Zhiliang Peng, Wei Huang, Shanzhi Gu, Lingxi Xie, Yaowei Wang, Jianbin Jiao, and Qixiang Ye. Conformer: Local features coupling global representations for visual recognition. In ICCV, 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 383, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 383, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 62, + 383, + 294, + 403 + ], + "type": "text", + "content": "[65] Karol J Piczak. Esc: Dataset for environmental sound classification. In ACM MM, 2015. 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 404, + 294, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 404, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 62, + 404, + 294, + 448 + ], + "type": "text", + "content": "[66] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with Reconstruct: Contrastive 3d Representation Learning Guided by Generative Pretraining. In ICML, 2023. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 449, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 294, + 502 + ], + "type": "text", + "content": "[67] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 504, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 504, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 62, + 504, + 294, + 546 + ], + "type": "text", + "content": "[68] Mike Ranzinger, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. Am-RADIO: Agglomerative Vision Foundation Model Reduce All Domains Into One. In CVPR, 2024. 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 548, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 548, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 62, + 548, + 294, + 591 + ], + "type": "text", + "content": "[69] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. In NeurIPS, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 592, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 592, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 62, + 592, + 294, + 646 + ], + "type": "text", + "content": "[70] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. In ICML, 2023. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 647, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 294, + 690 + ], + "type": "text", + "content": "[71] Jinghuan Shang, Karl Schmeckpeper, Brandon B. May, Maria Vittoria Minniti, Tarik Kelestemur, David Watkins, and Laura Herlant. Theia: Distilling Diverse Vision Foundation Models for Robot Learning. In CoRL, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 691, + 294, + 713 + ], + "type": "text", + "content": "[72] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "text", + "content": "Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In ICLR, 2017. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 321, + 96, + 553, + 129 + ], + "type": "text", + "content": "[73] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 131, + 553, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 131, + 553, + 164 + ], + "spans": [ + { + "bbox": [ + 320, + 131, + 553, + 164 + ], + "type": "text", + "content": "[74] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In CVPR, 2015. 5, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 166, + 553, + 199 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 166, + 553, + 199 + ], + "spans": [ + { + "bbox": [ + 320, + 166, + 553, + 199 + ], + "type": "text", + "content": "[75] Xinhang Song, Luis Herranz, and Shuqiang Jiang. Depth cnns for rgb-d scene recognition: Learning from scratch better than transferring from rgb-cnns. In AAAI, 2017. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 201, + 553, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 201, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 320, + 201, + 553, + 243 + ], + "type": "text", + "content": "[76] Xinhang Song, Shuqiang Jiang, Bohan Wang, Chengpeng Chen, and Gongwei Chen. Image representations with spatial object-to-object relations for rgb-d scene recognition. IEEE TIP, 29:525-537, 2019. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 246, + 553, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 246, + 553, + 278 + ], + "spans": [ + { + "bbox": [ + 320, + 246, + 553, + 278 + ], + "type": "text", + "content": "[77] K Soomro. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 280, + 553, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 280, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 320, + 280, + 553, + 314 + ], + "type": "text", + "content": "[78] Siddharth Srivastava and Gaurav Sharma. Omnivec2 - A Novel Transformer based Network for Large Scale Multimodal and Multitask Learning. In CVPR, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 316, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 316, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 320, + 316, + 553, + 358 + ], + "type": "text", + "content": "[79] Ximeng Sun, Pengchuan Zhang, Peizhao Zhang, Hardik Shah, Kate Saenko, and Xide Xia. Dime-fm: Distilling multimodal and efficient foundation models. In ICCV, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 361, + 553, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 361, + 553, + 383 + ], + "spans": [ + { + "bbox": [ + 320, + 361, + 553, + 383 + ], + "type": "text", + "content": "[80] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. In ECCV, 2022. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 384, + 553, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 384, + 553, + 439 + ], + "spans": [ + { + "bbox": [ + 320, + 384, + 553, + 439 + ], + "type": "text", + "content": "[81] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 441, + 553, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 441, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 320, + 441, + 553, + 485 + ], + "type": "text", + "content": "[82] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In ICCV, 2019. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 487, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 487, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 320, + 487, + 553, + 531 + ], + "type": "text", + "content": "[83] Simon Vandenhende, Stamatios Georgoulis, Wouter Van Gansbeke, Marc Proesmans, Dengxin Dai, and Luc Van Gool. Multi-task learning for dense prediction tasks: A survey. IEEE TPAMI, 44(7):3614-3633, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 533, + 553, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 533, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 320, + 533, + 553, + 597 + ], + "type": "text", + "content": "[84] Haoxiang Wang, Pavan Kumar Anasosalu Vasu, Fartash Faghri, Raviteja Vemulapalli, Mehrdad Farajtabar, Sachin Mehta, Mohammad Rastegari, Oncel Tuzel, and Hadi Pouransari. Sam-clip: Merging vision foundation models towards semantic and spatial understanding. In CVPR, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 600, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 600, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 320, + 600, + 553, + 633 + ], + "type": "text", + "content": "[85] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 320, + 634, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 634, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 320, + 634, + 553, + 689 + ], + "type": "text", + "content": "[86] Peng Wang, Shijie Wang, Junyang Lin, Shuai Bai, Xiaohuan Zhou, Jingren Zhou, Xinggang Wang, and Chang Zhou. One-peace: Exploring one general representation model toward unlimited modalities. arXiv preprint arXiv:2305.11172, 2023. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 320, + 690, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 690, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 690, + 553, + 713 + ], + "type": "text", + "content": "[87] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu," + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "29654" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 81, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 294, + 105 + ], + "type": "text", + "content": "Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. In CVPR, 2023. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 106, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 106, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 61, + 106, + 295, + 149 + ], + "type": "text", + "content": "[88] Wenguan Wang, Yi Yang, and Yunhe Pan. Visual knowledge in the big model era: Retrospect and prospect. Frontiers of Information Technology & Electronic Engineering, 26(1):1-19, 2025. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 150, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 150, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 61, + 150, + 294, + 205 + ], + "type": "text", + "content": "[89] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 205, + 295, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 295, + 249 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 295, + 249 + ], + "type": "text", + "content": "[90] Zehan Wang, Ziang Zhang, Luping Liu, Yang Zhao, Haifeng Huang, Tao Jin, and Zhou Zhao. Extending multi-modal contrastive representations. arXiv preprint arXiv:2310.08884, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 250, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 250, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 250, + 295, + 293 + ], + "type": "text", + "content": "[91] Zehan Wang, Yang Zhao, Haifeng Huang, Jiageng Liu, Aoxiong Yin, Li Tang, Linjun Li, Yongqi Wang, Zhang, and Zhou Zhao. Connecting multi-modal contrastive representations. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 293, + 295, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 293, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 62, + 293, + 295, + 326 + ], + "type": "text", + "content": "[92] Pete Warden. Speech commands: A dataset for limited-vocabulary speech recognition. arXiv preprint arXiv:1804.03209, 2018.7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 327, + 295, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 327, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 62, + 327, + 295, + 370 + ], + "type": "text", + "content": "[93] Yixuan Wei, Han Hu, Zhenda Xie, Zheng Zhang, Yue Cao, Jianmin Bao, Dong Chen, and Baining Guo. Contrastive learning rivals masked image modeling in fine-tuning via feature distillation. Tech Report, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 371, + 295, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 371, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 62, + 371, + 295, + 414 + ], + "type": "text", + "content": "[94] Kan Wu, Jinnian Zhang, Houwen Peng, Mengchen Liu, Bin Xiao, Jianlong Fu, and Lu Yuan. Tinyvit: Fast Pretraining Distillation for Small Vision Transformers. In ECCV, 2022. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 415, + 295, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 458 + ], + "type": "text", + "content": "[95] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, 2015. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 460, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 295, + 491 + ], + "type": "text", + "content": "[96] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 492, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 492, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 492, + 295, + 525 + ], + "type": "text", + "content": "[97] Peng Xu, Xiatian Zhu, and David A Clifton. Multimodal learning with transformers: A survey. IEEE TPAMI, 45 (10):12113-12132, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 526, + 295, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 526, + 295, + 580 + ], + "spans": [ + { + "bbox": [ + 62, + 526, + 295, + 580 + ], + "type": "text", + "content": "[98] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In CVPR, 2023. 2, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 581, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 581, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 581, + 295, + 624 + ], + "type": "text", + "content": "[99] Chuanguang Yang, Zhulin An, Libo Huang, Junyu Bi, Xinqiang Yu, Han Yang, Boyu Diao, and Yongjun Xu. ClipKD: An Empirical Study of CLIP Model Distillation. In CVPR, 2024. 3, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 625, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 668 + ], + "type": "text", + "content": "[100] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models (exemplified as a video agent). In ICML, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[101] Yazhou Yao, Tao Chen, Guo-Sen Xie, Chuanyi Zhang, Fumin Shen, Qi Wu, Zhenmin Tang, and Jian Zhang. Nonsalient region object mining for weakly supervised semantic segmentation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 510 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[102] Yazhou Yao, Zeren Sun, Chuanyi Zhang, Fumin Shen, Qi Wu, Jian Zhang, and Zhenmin Tang. Jo-src: A contrastive approach for combating noisy labels. In CVPR, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "type": "text", + "content": "[103] Junbo Yin, Jianbing Shen, Runnan Chen, Wei Li, Ruigang Yang, Pascal Frossard, and Wenguan Wang. Is-fusion: Instance-scene collaborative fusion for multimodal 3d object detection. In ECCV, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 553, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 175 + ], + "type": "text", + "content": "[104] Shan You, Chang Xu, Chao Xu, and Dacheng Tao. Learning from multiple teacher networks. In KDD, 2017. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "text", + "content": "[105] Fei Yuan, Linjun Shou, Jian Pei, Wutao Lin, Ming Gong, Yan Fu, and Daxin Jiang. Reinforced multi-teacher selection for knowledge distillation. In AAAI, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 209, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 251 + ], + "type": "text", + "content": "[106] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xu-peng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In CVPR, 2022. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "text", + "content": "[107] Yiyuan Zhang, Kaixiong Gong, Kaipeng Zhang, Hongsheng Li, Yu Qiao, Wanli Ouyang, and Xiangyu Yue. Meta-transformer: A unified framework for multimodal learning. arXiv preprint arXiv:2307.10802, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "type": "text", + "content": "[108] Yanqi Zhou, Tao Lei, Hanxiao Liu, Nan Du, Yanping Huang, Vincent Zhao, Andrew M Dai, Quoc V Le, James Laudon, et al. Mixture-of-experts with expert choice routing. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 343, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 343, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 317, + 343, + 553, + 408 + ], + "type": "text", + "content": "[109] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, WANG HongFa, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei Li, Cai Wan Zhang, Zhifeng Li, Wei Liu, and Li Yuan. Languagebind: Extending Video-Language Pretraining to N-modality by Language-based Semantic Alignment. In ICLR, 2024. 1, 2, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 410, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 553, + 432 + ], + "type": "text", + "content": "[110] Xiatian Zhu, Shaogang Gong, et al. Knowledge distillation by on-the-fly native ensemble. In NeurIPS, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 434, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 476 + ], + "type": "text", + "content": "[111] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In ICCV, 2023. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "type": "text", + "content": "[112] Konrad Zuchniak. Multi-teacher knowledge distillation as an effective method for compressing ensembles of neural networks. arXiv preprint arXiv:2302.07215, 2023. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29655" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/UNIC-Adapter_ Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation/409eb6c5-d571-4968-a329-2c1ebf51e321_content_list.json b/2025/UNIC-Adapter_ Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation/409eb6c5-d571-4968-a329-2c1ebf51e321_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4770120b86dad400565198e1348df094d072085a --- /dev/null +++ b/2025/UNIC-Adapter_ Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation/409eb6c5-d571-4968-a329-2c1ebf51e321_content_list.json @@ -0,0 +1,1766 @@ +[ + { + "type": "text", + "text": "UNIC-Adapter: Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation", + "text_level": 1, + "bbox": [ + 148, + 128, + 849, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lunhao Duan\\*1 Shanshan Zhao\\*2 Wenjun Yan2 Yinglun Li2 Qing-Guo Chen2 Zhao Xu2 Weihua Luo2 Kaifu Zhang2 Mingming Gong3 Gui-Song Xia\\*14", + "bbox": [ + 161, + 202, + 834, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science, Wuhan University $^{2}$ Alibaba International Digital Commerce $^{3}$ University of Melbourne $^{4}$ School of Artificial Intelligence, Wuhan University", + "bbox": [ + 132, + 255, + 864, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, text-to-image generation models have achieved remarkable advancements, particularly with diffusion models facilitating high-quality image synthesis from textual descriptions. However, these models often struggle with achieving precise control over pixel-level layouts, object appearances, and global styles when using text prompts alone. To mitigate this issue, previous works introduce conditional images as auxiliary inputs for image generation, enhancing control but typically necessitating specialized models tailored to different types of reference inputs. In this paper, we explore a new approach to unify controllable generation within a single framework. Specifically, we propose the unified image-instruction adapter (UNIC-Adapter) built on the Multi-Modal-Diffusion Transformer architecture, to enable flexible and controllable generation across diverse conditions without the need for multiple specialized models. Our UNIC-Adapter effectively extracts multi-modal instruction information by incorporating both conditional images and task instructions, injecting this information into the image generation process through a cross-attention mechanism enhanced by Rotary Position Embedding. Experimental results across a variety of tasks, including pixel-level spatial control, subject-driven image generation, and style-image-based image synthesis, demonstrate the effectiveness of our UNIC-Adapter in unified controllable image generation.", + "bbox": [ + 89, + 358, + 483, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 766, + 220, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in text-to-image (T2I) generation models [3, 5, 12, 20, 23, 29, 38, 42, 51], primarily driven by diffusion models [16], have been witnessed in recent years. By leveraging large-scale paired text-image data [48], nu", + "bbox": [ + 89, + 791, + 482, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(a)", + "bbox": [ + 547, + 330, + 560, + 340 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A delicious afternoon tea.", + "bbox": [ + 581, + 328, + 625, + 345 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A room with the forest backdrop.", + "bbox": [ + 645, + 328, + 705, + 345 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A robot in autumn mountain and lake", + "bbox": [ + 715, + 328, + 781, + 345 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "···", + "bbox": [ + 785, + 330, + 797, + 340 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A dog with a mountain in the", + "bbox": [ + 813, + 328, + 870, + 345 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(b)", + "bbox": [ + 547, + 358, + 560, + 367 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generate image from canny edge.", + "bbox": [ + 584, + 352, + 625, + 375 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Render image with this depth map.", + "bbox": [ + 647, + 351, + 705, + 375 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Create image using style influence.", + "bbox": [ + 715, + 352, + 777, + 375 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1", + "bbox": [ + 782, + 358, + 797, + 367 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generate image from dog image", + "bbox": [ + 810, + 358, + 870, + 375 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(c)", + "bbox": [ + 547, + 402, + 560, + 411 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e13f90acba9ec6aacfb88fabb46d30e3470f85c4d61206fc2e5996f349e45db0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 378, + 638, + 434 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/67f8af2831537e9bc2cdbf08c4e98da41798927a67bf4f09deb17debb8353b64.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 378, + 710, + 434 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ec71007de18d7c771c97f474b57a70353f275ce361265f88ab2635cded58dbea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 378, + 785, + 434 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bf9742303b6ada9ef0af02f9f54e4fb61fe888b4bbcae808cc4df645c4135378.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 789, + 378, + 875, + 434 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(d)", + "bbox": [ + 547, + 479, + 560, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/56439bdfacd22dfd3e04f3aec4996bd6946b20e5b170326e6ad5723be1ef36f7.jpg", + "image_caption": [ + "Figure 1. With the UNIC-Adapter, SD3 enables flexible and controllable generation across multiple reference modalities within a single model. (a) to (d) represent text prompts, task instructions, conditional images, and generated images, respectively." + ], + "image_footnote": [], + "bbox": [ + 566, + 436, + 638, + 511 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0a25eff6fd53b9af61b0eb36efac739d6727206d4c74744092410610e3d0cd43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 436, + 710, + 511 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7c064a96448ffe8b22842ad46f2a0b7af58d5d197ab88015455f1cb5e31ae8ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 436, + 785, + 511 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2408286dc7a0cf4f9c10d82be303c4924ffee240e3f6554ba77e1632434dd055.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 789, + 436, + 875, + 511 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "merous open-source T2I models, like Stable Diffusion 1.5 (SD1.5) [42], Stable Diffusion XL (SDXL) [38], Pixart-series [5-7], Stable Diffusion 3 (SD3) [12], Hunyuan-DiT [28], and FLUX.1-dev [23] have significantly enhanced the fidelity of images generated from natural language descriptions. These models utilize a range of architectures, spanning from U-Net-based designs [43] to the more recent Diffusion-Transformer-based frameworks [35].", + "bbox": [ + 511, + 595, + 906, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite these advancements, relying solely on text prompts often falls short in defining pixel-level spatial structures and geometric details, exact object appearances, and global styles of a specific image. To address this limitation, recent research [13, 18, 33, 34, 45, 55, 58, 60] has incorporated conditional images as additional inputs alongside text prompts. These conditional images offer finer control over pixel-level [60, 62], object-level [34, 55, 58], or style-specific [13, 15, 54] features in the generated outputs. For example, ControlNet [60] enables pixel-level spatial control by integrating structural inputs, such as depth and edge maps, through a parallel encoder tailored for", + "bbox": [ + 509, + 719, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done when Lunhao Duan was a research intern at Alibaba International Digital Commerce. \n†Corresponding authors.", + "bbox": [ + 89, + 862, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7963", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "each type of control. Similarly, IP-Adapter [58] enables content-specific control by conditioning on a reference image and injecting its CLIP-based [40] embeddings via crossattention layers. However, these methods often necessitate individual model training for each control type, resulting in increased training costs and complicating the integration of different control inputs. IPAdapter-Instruct [44] extends IP-Adapter by incorporating instruction prompts to specify task types for different conditional images, enabling multi-task handling within a single model. UniControl [39] tackles various pixel-level control tasks within a unified model by introducing task-specific adapters to extract features for different visual conditions. Meanwhile, Uni-ControlNet [61] divides conditional images into two groups: local (pixel-level) and global (image-level) conditions, and designs dedicated adapters for each group. However, this approach is not fully unified, as the two adapters are trained separately.", + "bbox": [ + 89, + 90, + 480, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To further unify diverse reference inputs, including text, edge maps, content images, etc., Instruct-Imagen [17], based onImagen [46], represents them as multi-modal instructions within a single framework. OmniGen [56] first tokenizes interleaved text and images into a one-dimensional sequence and then uses a transformer model (Phi3) [1] to achieve unified image generation, such as image editing and subject-driven generation. To adapt the model for multi-modal instructions, these two methods involve multi-stage training of the entire model. To enable more efficient training for unified controllable image generation, this paper seeks to design an adapter based on a pre-trained T2I model, avoiding the need to train the entire model.", + "bbox": [ + 89, + 367, + 482, + 564 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given that DiT-based architectures [36] equipped with pure transformer blocks have become the mainstream in T2I models [5, 12, 23, 28], we choose SD3, an open-source state-of-the-art (SOTA) model, as the foundation model for studying the adapter. To achieve unified controllable image generation across various conditions, we employ the Multi-Modal-Diffusion Transformer (MM-DiT) structure [12, 23], which enables full cross-attention between text and image features to extract multi-modal instruction information. By treating text as an independent modality, the MM-DiT allocates substantial model capacity to capture and understand text-based information, thereby improving its interpretation of text prompts [12]. Leveraging the strengths of the MM-DiT, we propose a unified image-instruction adapter (UNIC-Adapter) to facilitate image generation conditioned on various types of reference images with a task instruction indicating the desired generation target. Specifically, task instruction and conditional image features are extracted by text encoders [40, 41] and a Variational Autoencoder (VAE) [12], respectively. Then, the adapter processes instruction and image features using MM-DiT blocks, which are subsequently integrated into the", + "bbox": [ + 89, + 568, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "main generation branch via a cross-attention mechanism. To further enhance spatial awareness, which is essential for pixel-level control, we incorporate Rotary Position Embedding (RoPE) [50] into the query and key features within the cross-attention. With the proposed UNIC-Adapter, SD3 can generate images under a wide array of conditions, as shown in Figure 1.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are threefold:", + "bbox": [ + 531, + 196, + 823, + 210 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We are the first to leverage the MM-DiT architecture for unified controllable generation across various types of reference images in T2I models.", + "- We introduce an effective feature injection method based on cross-attention, enhanced with RoPE encoding for better spatial control.", + "- Equipped with our UNIC-Adapter, SD3 enables unified controllable generation across 14 types of conditional images in our experiments." + ], + "bbox": [ + 511, + 213, + 903, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 364, + 653, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Text-to-Image Generation Models", + "text_level": 1, + "bbox": [ + 511, + 390, + 808, + 406 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mainstream T2I generation models [5, 12, 38, 42] are primarily based on diffusion models [16], which generate images by learning to reverse a progressive noise-adding process. In these models, text input typically serves as a conditional signal, guiding the generation process to align with the provided description. The success of transformer architectures [53], particularly in large language models [2, 52], has significantly influenced T2I model architectures, evolving them from convolution-transformer hybrids [38, 42] to fully transformer-based designs [5-7, 12, 23, 28]. Recently, the MM-DiT architecture [12, 23] has advanced this evolution by treating text and image features as distinct modalities with full-attention mechanisms, enabling each modality to attend to the other. This approach outperforms traditional cross-attention setups, where text features are used only as keys and values [12]. However, relying solely on text-based conditioning presents limitations, particularly for images that require detailed spatial layouts or specific object appearances. Our work builds upon these transformer-based diffusion models by integrating additional conditional images and task instructions, thereby enhancing the flexibility and controllability of T2I generation.", + "bbox": [ + 511, + 412, + 906, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Conditional Image-based T2I Models", + "text_level": 1, + "bbox": [ + 511, + 756, + 834, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Conditional image-based T2I models introduce auxiliary image inputs to complement textual prompts, enabling finer control over aspects such as spatial layout [26, 60, 62], object content [9, 24, 31, 49, 58], and stylistic elements [13, 14, 54]. For pixel-level spatial control, ControlNet [60] employs a duplicated U-Net encoder to process specific types of conditional images (e.g., edge, pose, and depth maps) and integrates the features into the main T2I model", + "bbox": [ + 511, + 779, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7964", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/86f0ef7ac179a0d0bcc1d0ab7b46c475a1cb21fcaca5ce2d236ff9e08f372406.jpg", + "image_caption": [ + "Figure 2. The overall architecture of our proposed UNIC-Adapter. The task instruction and conditional image features are progressively attending to each other through a series of $N$ adapter blocks. In each adapter block, the image features $Z_{\\mathrm{img}}$ from MM-DiT block in the main image generation branch serve as the query, while both task instruction features $Z_{\\mathrm{ist}}$ and conditional image features $Z_{\\mathrm{con}}$ function as keys and values. For simplicity, normalization layers and feed-forward networks are omitted in this figure." + ], + "image_footnote": [], + "bbox": [ + 155, + 88, + 844, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "through residual connections in the U-Net decoder. Another key work, IP-Adapter [58], uses the Vision Transformer (ViT) [11] encoder from CLIP [40] to extract features from conditional images. These features are then incorporated into the U-Net backbone via cross-attention layers, achieving controllable generation with contents from conditional images. IPAdapter-Instruct [44] extends this by introducing an instructional prompt that guides the interpretation of conditional images. To minimize the necessity for individual model training for each type of conditional image in ControlNet, UniControl [39] unifies nine pixel-level control tasks within a single model, leveraging task-specific feature extraction adapters and a task-aware HyperNet to modulate features based on task instructions. To achieve control with image content as well as spatial layout, UniControlNet [61] incorporates two dedicated adapters for local (pixel-level control) and global (CLIP-image embedding control) conditions, respectively. Unlike these methods, which freeze the base model and only training the newly added modules / adapters, Instruct-Imagen [17] proposes using multi-modal instructions and optimize the entire model, including the base model (Imagen [46]) and the new cross-attention layers, for image generation under different conditions. A recent work, OmniGen [56], develops a unified image generation model for various multi-modal inputs using a transformer-based architecture [1]. In this paper, we aim to enhance the generation capabilities of an existing T2I model for diverse conditions by introducing a unified image-instruction adapter, named UNIC-Adapter, without the need to train the entire model.", + "bbox": [ + 91, + 441, + 483, + 895 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 441, + 604, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first provide a brief overview of the DiT and MM-DiT employed in SOTA T2I models. Then, we describe our UNIC-Adapter, which incorporates task instructions and conditional images to enhance the model's capability to interpret various user-specified conditions.", + "bbox": [ + 511, + 465, + 906, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 511, + 549, + 643, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Transformer Block. In DiT Block-based T2I models [5], conditional text features $Z_{\\mathrm{txt}}$ extracted from text encoders [40, 41] are integrated into DiT Blocks via cross-attention layers. Specifically, in each DiT Block, the conditional text features are first mapped to key $K_{\\mathrm{txt}}$ and value $V_{\\mathrm{txt}}$ features through linear transformations $L$ , while the image feature $Z_{\\mathrm{img}}$ is mapped to a query feature $Q_{\\mathrm{img}}$ . An attention mechanism then updates the image features $Z_{\\mathrm{img}}$ by attending to $K_{\\mathrm{txt}}$ and $V_{\\mathrm{txt}}$ :", + "bbox": [ + 511, + 569, + 908, + 696 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nK _ {\\mathrm {t x t}} = L _ {\\mathrm {t x t}} ^ {k} \\left(Z _ {\\mathrm {t x t}}\\right), V _ {\\mathrm {t x t}} = L _ {\\mathrm {t x t}} ^ {v} \\left(Z _ {\\mathrm {t x t}}\\right), Q _ {\\mathrm {i m g}} = L _ {\\mathrm {i m g}} ^ {q} \\left(Z _ {\\mathrm {i m g}}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 700, + 903, + 724 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {\\mathrm {i m g}} ^ {\\prime} = \\operatorname {A t t n} \\left(Q _ {\\mathrm {i m g}}, K _ {\\mathrm {t x t}}, V _ {\\mathrm {t x t}}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 722, + 714, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where Attn indicates the multi-head attention [53]. In this setup, the same $Z_{\\mathrm{txt}}$ features are provided to DiT Blocks across different layers as static conditioning signals, which may limit the utilization of text features.", + "bbox": [ + 511, + 742, + 906, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-modal Diffusion Transformer Block. Recent T2I models [12, 23] based on MM-DiT Blocks enhance DiT Blocks by treating conditional text as an independent modality capable of dynamically interacting with image features across layers. In MM-DiT, both text and image features serve as queries, keys, and values, enabling cross-modal attention in both directions through a full-attention", + "bbox": [ + 511, + 801, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7965", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mechanism. Specifically, the MM-DiT layer performs the following operations:", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {\\mathrm {t x t}} = L _ {\\mathrm {t x t}} ^ {q} \\left(Z _ {\\mathrm {t x t}}\\right), K _ {\\mathrm {t x t}} = L _ {\\mathrm {t x t}} ^ {k} \\left(Z _ {\\mathrm {t x t}}\\right), V _ {\\mathrm {t x t}} = L _ {\\mathrm {t x t}} ^ {v} \\left(Z _ {\\mathrm {t x t}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 128, + 482, + 152 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {\\mathrm {i m g}} = L _ {\\mathrm {i m g}} ^ {q} \\left(Z _ {\\mathrm {i m g}}\\right), K _ {\\mathrm {i m g}} = L _ {\\mathrm {i m g}} ^ {k} \\left(Z _ {\\mathrm {i m g}}\\right), V _ {\\mathrm {i m g}} = L _ {\\mathrm {i m g}} ^ {v} \\left(Z _ {\\mathrm {i m g}}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 150, + 455, + 166 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These queries, keys, and values are processed in a full-attention layer, where each modality attends to both text and image features:", + "bbox": [ + 89, + 172, + 482, + 219 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {\\mathrm {t x t}} ^ {\\prime} = \\operatorname {A t t n} \\left(Q _ {\\mathrm {t x t}}, \\left[ K _ {\\mathrm {t x t}} \\| K _ {\\mathrm {i m g}} \\right], \\left[ V _ {\\mathrm {t x t}} \\| V _ {\\mathrm {i m g}} \\right]\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 229, + 482, + 255 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {\\mathrm {i m g}} ^ {\\prime} = \\operatorname {A t t n} \\left(Q _ {\\mathrm {i m g}}, \\left[ K _ {\\mathrm {t x t}} \\right\\| K _ {\\mathrm {i m g}} ], \\left[ V _ {\\mathrm {t x t}} \\right\\| V _ {\\mathrm {i m g}}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 252, + 434, + 271 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\parallel$ denotes concatenation. This enhanced interaction facilitates richer cross-modal feature extraction, as text and image features are iteratively refined across layers, improving text-guided image generation [12].", + "bbox": [ + 89, + 280, + 483, + 342 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. UNIC-Adapter", + "text_level": 1, + "bbox": [ + 89, + 349, + 243, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 2, the UNIC-Adapter introduces two additional inputs: task instructions and conditional images. This integration enables unified controllable image generation by allowing the model to interpret and apply specific features from a variety of conditional images, guided by the task instruction. Below, we provide a detailed description of its structure.", + "bbox": [ + 89, + 372, + 483, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image-instruction Feature Extraction. To enable a comprehensive understanding of task instructions and conditional images, we leverage MM-DiT blocks to enhance the interaction between these inputs. Specifically, we adopt SD3 as our base model. Following the setup in SD3, task instruction features, denoted as $Z_{\\mathrm{ist}}$ , are extracted using text encoders such as CLIP-G/14 [10], CLIP-L/15 [40], and T5 XXL [41]. Simultaneously, conditional image features, $Z_{\\mathrm{con}}$ , which represent one of several condition types (e.g., edge maps, depth maps, subject content images, or style-reference images), are obtained through the VAE. Within the MM-DiT blocks, task instruction and conditional image features undergo mutual attention updates, defined as follows:", + "bbox": [ + 89, + 477, + 483, + 670 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {\\text {i s t}} = L _ {\\text {i s t}} ^ {q} (Z _ {\\text {i s t}}), K _ {\\text {i s t}} = L _ {\\text {i s t}} ^ {k} (Z _ {\\text {i s t}}), V _ {\\text {i s t}} = L _ {\\text {i s t}} ^ {v} (Z _ {\\text {i s t}}),\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 676, + 406, + 694 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {\\text {c o n}} = L _ {\\text {c o n}} ^ {q} \\left(Z _ {\\text {c o n}}\\right), K _ {\\text {c o n}} = L _ {\\text {c o n}} ^ {k} \\left(Z _ {\\text {c o n}}\\right), V _ {\\text {c o n}} = L _ {\\text {c o n}} ^ {v} \\left(Z _ {\\text {c o n}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 696, + 482, + 723 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {\\mathrm {i s t}} ^ {\\prime} = \\operatorname {A t t n} \\left(Q _ {\\mathrm {i s t}}, \\left[ K _ {\\mathrm {i s t}} \\| K _ {\\mathrm {c o n}} \\right], \\left[ V _ {\\mathrm {i s t}} \\| V _ {\\mathrm {c o n}} \\right]\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 718, + 352, + 736 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {\\text {c o n}} ^ {\\prime} = \\operatorname {A t t n} \\left(Q _ {\\text {c o n}}, \\left[ K _ {\\text {i s t}} \\| K _ {\\text {c o n}} \\right], \\left[ V _ {\\text {i s t}} \\| V _ {\\text {c o n}} \\right]\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 739, + 359, + 757 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These updates ensure that the task instruction $Z_{\\mathrm{ist}}$ and conditional image $Z_{\\mathrm{con}}$ dynamically refine each other's features by selectively attending to relevant information.", + "bbox": [ + 89, + 763, + 482, + 810 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image-instruction Feature Injection. To improve the model's responsiveness to diverse user-defined conditions, the key and value features from both instruction and conditional image inputs are integrated simultaneously into the main image generation branch through cross-attention [17, 55, 58]. This mechanism ensures that the feature injection is", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "directed by the specific task instruction. Experiments indicate that injecting either conditional image features or task instruction features alone through cross-attention results in suboptimal performance, as demonstrated in the ablation studies of Section 4. By contrast, the simultaneous incorporation of both feature types optimally supports unified controlled image generation. Accordingly, the image features $Z_{\\mathrm{img}}^{\\prime}$ from Equation 3 are further updated as follows:", + "bbox": [ + 511, + 90, + 906, + 214 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {\\mathrm {i m g}} ^ {\\prime} = L _ {\\text {c r o s s}} ^ {q} \\left(Z _ {\\mathrm {i m g}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 231, + 903, + 258 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {\\mathrm {i m g}} ^ {\\prime \\prime} = Z _ {\\mathrm {i m g}} ^ {\\prime} + \\operatorname {A t t n} \\left(Q _ {\\mathrm {i m g}} ^ {\\prime}, \\left[ K _ {\\mathrm {i s t}} \\| K _ {\\mathrm {c o n}} \\right], \\left[ V _ {\\mathrm {i s t}} \\| V _ {\\mathrm {c o n}} \\right]\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 255, + 867, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this setup, we introduce a new linear layer, $L_{\\mathrm{cross}}^q$ , designed to enhance the model's ability to attend to image-instruction features effectively. This additional layer has been shown to improve performance, as validated in our experiments. Overall, this cross-attention operation facilitates controlled image generation, leveraging desired attributes from both the conditional image and the task instruction.", + "bbox": [ + 511, + 279, + 906, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Position Embedding. The positional information of conditional image features is crucial for pixel-level control tasks, as methods like ControlNet incorporate features through pixel-to-pixel addition. To improve pixel-level control precision, we integrate Rotary Position Embedding (RoPE) [50] into both the query and key features prior to cross-attention. RoPE provides relative positional encoding, ensuring that query and key features with closer pixel coordinates yield higher similarity scores, thereby improving spatial alignment in the generated output. It has been used in T2I [23] and text-to-video [57] models.", + "bbox": [ + 511, + 386, + 905, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we apply 1-D RoPE separately to the height and width dimensions. Given a feature vector $f \\in \\mathbb{R}^{|D|}$ at position $(h,w)$ , we split $f$ into two halves: $f_h \\in \\mathbb{R}^{|D|/2}$ for the height dimension and $f_w \\in \\mathbb{R}^{|D|/2}$ for the width. For the height component $f_h$ , the rotary matrix $R_h$ is defined as:", + "bbox": [ + 511, + 551, + 905, + 636 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR _ {h} = \\left[ \\begin{array}{c c c c c} \\cos (h \\theta_ {0}) & - \\sin (h \\theta_ {0}) & \\dots & 0 & 0 \\\\ \\sin (h \\theta_ {0}) & \\cos (h \\theta_ {0}) & \\dots & 0 & 0 \\\\ \\vdots & \\vdots & \\ddots & \\vdots & \\vdots \\\\ 0 & 0 & \\dots & \\cos (h \\theta_ {l}) & - \\sin (h \\theta_ {l}) \\\\ 0 & 0 & \\dots & \\sin (h \\theta_ {l}) & \\cos (h \\theta_ {l}) \\end{array} \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 641, + 905, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta_{i} = b^{-i / |D|}$ , with $b = 10000$ and $l = |D| / 4 - 1$ . The rotary matrix $R_{w}$ for the width component $f_{w}$ is defined analogously. The updated feature vector $f$ is then computed as:", + "bbox": [ + 511, + 728, + 905, + 787 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf = \\left[ R _ {h} f _ {h} \\| R _ {w} f _ {w} \\right]. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 789, + 903, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "RoPE is applied to both query and key features in Equation 5 and Equation 4, with the position coordinates of $K_{\\mathrm{ins}}$ set to $(0,0)$ . By embedding relative positional information in this way, RoPE enhances the model's precision in pixel-level control, while maintaining robustness and adaptability across various tasks.", + "bbox": [ + 511, + 810, + 905, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7966", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 89, + 223, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we conduct extensive experiments to evaluate the performance of our UNIC-Adapter in various conditional image generation tasks. Detailed ablation studies are also conducted to analyze the contribution of individual components within our framework. Additional experimental results are available in the Appendix.", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Tasks and Datasets", + "text_level": 1, + "bbox": [ + 89, + 215, + 272, + 231 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our evaluation covers three major tasks: pixel-level spatial control, subject-driven image generation, and style-image-based T2I generation. Below, we briefly describe the datasets used for training each task.", + "bbox": [ + 89, + 238, + 482, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pixel-level Spatial Control. For pixel-level spatial control, we employ the MultiGen-20M dataset [39], which includes 12 types of pixel-level control annotations, such as Canny, HED, Sketch, Depth, Normal, Skeleton, Bbox, Segmentation, Outpainting, Inpainting, Deblurring, and Colorization. The dataset contains 2.8 million images with 20 million control annotations, and we reserve 1,000 samples for qualitative evaluation. For segmentation map conditioning, we also use the ADE20k [63] training set, which includes approximately 20,000 samples. These annotations serve as conditional images, and we use GPT-4o [2] to generate 20 synonymous instructions for each control type, such as \"Generate an image from this edge map.\"", + "bbox": [ + 89, + 300, + 483, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subject-driven Image Generation. For subject-driven image generation, we use three datasets: OpenImages v7 [22], GRIT [37], and OpenStory++ [59]. Following previous works [31, 34], we utilize Blip-2 [25] to generate captions for images without annotations and use spaCy‡ to extract subject tags. GroundingDINO [30] is used to obtain bounding boxes for each subject, and SAM [21] is employed to generate subject segmentation maps based on these bounding boxes. After processing and filtering, we obtain 2.1 million subject-image pairs. We also use GPT4o to generate 20 instructions, such as \"Generate an image from this dog image.\"", + "bbox": [ + 89, + 496, + 483, + 678 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Style-image-based Image Generation. For style-image-based generation, we use the WikiArt [47] and Style-Booth [14] datasets. For StyleBooth, we randomly select an image from the same style subset as the conditional image, and for WikiArt, we employ CLIP-I [40] similarity to retrieve a conditional image for each training image. These two datasets comprise approximately 90,000 images. Additionally, we use GPT-4o to generate 20 instructions, such as \"Generate an image based on this style image.\"", + "bbox": [ + 89, + 678, + 483, + 814 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 824, + 307, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our UNIC-Adapter framework is based on the SD3 medium [12] model. Since instruction texts and conditional", + "bbox": [ + 89, + 845, + 482, + 876 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/fcb0e0db968be7f8643dc0be244a4a052fc00ebe13b82f8dfad90c1c9bcb5816.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBase ModelCanny (F1 Score↑)HED (SSIM↑)Seg. (mIoU↑)Depth (RMSE↓)
Single-task
T2I-Adapter [33]SD1.423.65-12.6148.40
Gligen [27]SD1.426.940.563423.7838.83
ControlNet [60]SD1.534.650.762132.5535.90
ControlNet++ [26]SD1.537.040.809743.6428.32
Multi-task
UniControl [39]SD1.530.820.796925.4439.18
Uni-ControlNet [61]SD1.527.320.691019.3940.65
OmniGen [56]Phi-335.540.823744.2328.54
OursSD338.940.836942.8931.10
", + "bbox": [ + 517, + 89, + 913, + 272 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f9f672eaa173e9be8fe73b574330d02e2378850def21b794a5a4d0237b6f0d3b.jpg", + "table_caption": [ + "Table 1. Results of different methods on four pixel-level control image generation tasks. $\\uparrow$ indicates that higher values are better, while $\\downarrow$ indicates that lower values are better. The best results for each metric are bolded, and the second-best results are underlined." + ], + "table_footnote": [], + "table_body": "
MethodBase ModelDINO↑CLIP-I↑CLIP-T↑
Fine-Tuning
Textual Inversion [32]SD1.50.5690.7800.255
DreamBooth [45]SD1.50.6680.8030.305
BLIP-Diffusion [24]SD1.50.6700.8050.302
Test Time Tuning Free
Re-Imagen [8]-0.6000.7400.270
SuTI [9]Imagen0.7410.8190.304
Kosmos-G [34]SD1.50.6940.8470.287
OmniGen [56]Phi-30.8010.8470.301
OursSD30.8160.8410.306
", + "bbox": [ + 527, + 354, + 890, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Results of different methods on the DreamBench dataset [45] for subject-driven image generation.", + "bbox": [ + 511, + 534, + 906, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "are variations of text and image modalities, the UNIC-Adapter is initialized using the base T2I model's parameters, which minimizes learning difficulty. We freeze the parameters of base SD3 medium model and the feed-forward layer after the attention layer in adapter blocks, and train all other newly introduced parameters, resulting in approximately 1.2 billion trainable parameters. We use the AdamW optimizer [19] with a learning rate of 0.0001 and a weight decay of 0.01. Training is performed on 16 H100 GPUs for 100,000 steps, with a batch size of 16 per GPU.", + "bbox": [ + 511, + 577, + 906, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Main Results", + "text_level": 1, + "bbox": [ + 511, + 739, + 651, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We present experimental results of the proposed UNIC-Adapter framework across three major tasks.", + "bbox": [ + 511, + 763, + 905, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pixel-Level Spatial Control. Following previous work [26, 56], we evaluate performance based on the similarity between control conditions extracted from generated images and the specified pixel-level control inputs. The evaluation is conducted on the test set of MultiGen-20M with canny edge, hed edge, and depth map conditions and on ADE20k test set with segmentation map condition. As", + "bbox": [ + 511, + 795, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ https://github.com/explosion/spaCy", + "bbox": [ + 107, + 886, + 303, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7967", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/44180588d9e4f998bfb3816d16d28ac5f55735739f9c5ae3e1cf0f5058cb02d6.jpg", + "image_caption": [ + "Figure 3. Visualization results of our UNIC-Adapter on twelve pixel-level control tasks from the MultiGen-20M dataset. The first and third rows show different types of conditional images, while the second and fourth rows display the corresponding generated images." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 903, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "shown in Table 1, UNIC-Adapter achieves the best performance on canny and hed edge map conditions, and demonstrates comparable performance on depth and segmentation map conditions, compared to SOTA methods like ControlNet++ [26] for single-task and OmniGen [56] for multitask scenarios. Figure 3 shows qualitative results on 12 tasks from MultiGen-20M, illustrating that UNIC-Adapter consistently produces images accurately aligned with pixel-level controls.", + "bbox": [ + 89, + 601, + 483, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Subject-Driven Image Generation. Evaluation is conducted using 30 objects and 25 prompts from the Dream-Bench dataset [45]. Following prior work [34], we generate 4 images for each prompt and measure subject fidelity using DINO [4] and CLIP-I scores and text fidelity with CLIP-T [40]. As shown in Table 2, UNIC-Adapter achieves high subject-image consistency and maintains strong alignment between the generated images and text prompts. Qualitative examples in Figure 4 demonstrate that UNIC-Adapter effectively generates images closely resembling input sub", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "jects across various contexts.", + "bbox": [ + 511, + 602, + 705, + 616 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Style-image-based Image Generation. For the same text prompt, style images from StyleBooth [14] are randomly selected as conditional images. The qualitative results in Figure 5 indicate that UNIC-Adapter effectively captures and applies stylistic elements from conditional images, generating artistically consistent outputs.", + "bbox": [ + 511, + 618, + 906, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 723, + 679, + 738 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To better understand the impact of individual components in the UNIC-Adapter framework, we conduct ablation studies by systematically modifying or removing key modules. For efficiency, the models in this part are trained on the subject-driven image generation task and four tasks (canny, hed, depth, and segmentation) from MultiGen-20M.", + "bbox": [ + 511, + 747, + 906, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Query and Key Feature Choices. We investigate the effect of different query and key feature choices in Equation 5. First, we fix the query features as $Q_{\\mathrm{txt}}$ and $Q_{\\mathrm{img}}$ and examine the impact of different key features. As shown in Ta", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7968", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/56feff493f71225acad6e603a90f25da3628d8a72f7d5d6a5bea088652d213c2.jpg", + "image_caption": [ + "Figure 4. Visualization results of our UNIC-Adapter on Dream-Bench for subject-driven generation. The first column displays the subject images, while the other three columns show the generated images based on different prompts." + ], + "image_footnote": [], + "bbox": [ + 109, + 93, + 480, + 334 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/30155aee3240742cf5fdeb2e2da5e74e5152b3d2eb5e1a215cedd9d4587321c8.jpg", + "image_caption": [ + "A cat with a mountain in the background." + ], + "image_footnote": [], + "bbox": [ + 94, + 412, + 480, + 500 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/381c90fb07ae86a9c3e66ea29d6a429b886e232bc703740197e620caf983f3eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 510, + 480, + 583 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0aa83a366708b9203717c54e7407acf0f9420c26e7f232f450bd03049deb9e65.jpg", + "image_caption": [ + "A girl with a city in the background.", + "A robot in autumn mountain and lake." + ], + "image_footnote": [], + "bbox": [ + 94, + 595, + 480, + 669 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0835dd966396b4cefedc048427e38b9ae258f4674213cb2ba47183798da879e6.jpg", + "image_caption": [ + "Figure 5. Visualization results of our UNIC-Adapter on style-image-based T2I generation. The first row shows the reference style image, and each subsequent row contains images generated from the same prompt, influenced by different style images." + ], + "image_footnote": [], + "bbox": [ + 94, + 681, + 480, + 756 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ble 3, $K_{\\mathrm{con}}$ plays a critical role in pixel-level spatial control tasks and in maintaining subject-image similarity. Relying solely on $K_{\\mathrm{ist}}$ as key feature (Exp. 3) significantly reduces performance across pixel-level control tasks and subject-", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fd90d90f86770d9a0b92957506e7569ded4c1bd5a96777057644dec4b6b99833.jpg", + "image_caption": [ + "Figure 6. Comparison of visualization results of our UNIC-Adapter with and without position embedding, and IPAdapter-Instruct* on pixel-level control generation using canny edge conditional images. UNIC-Adapter generates images that are more closely aligned with the conditional images compared to the other two methods." + ], + "image_footnote": [], + "bbox": [ + 517, + 92, + 903, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fbe71ec62ee4ae7c880b20f5ad85aa49b79d7893a7f4f02f4cf4a69c6a2b0021.jpg", + "image_caption": [ + "Figure 7. Comparison of visualization results between our UNIC-Adapter and ControlNet* on subject-driven generation. Compared to ControlNet*, UNIC-Adapter achieves better consistency between subject images and generated images." + ], + "image_footnote": [], + "bbox": [ + 524, + 426, + 903, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "image similarity. We also investigate the impact of query features and find that using $Q_{\\mathrm{txt}}$ alone as the query feature (Exp. 5) also degrades performance. In general, using $Q_{\\mathrm{img}}$ as the query and both $K_{\\mathrm{con}}$ and $K_{\\mathrm{ist}}$ as key features (Exp. 4) yields optimal performance on pixel-level control tasks and comparable performance on subject-driven generation tasks. This highlights the importance of task instruction features in unified controllable generation. We adopt this configuration for subsequent experiments.", + "bbox": [ + 511, + 642, + 906, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Importance of Position Embedding. To enhance the control precision of pixel-level tasks, we experiment with incorporating different types of position embeddings, including the absolute position embedding in SD3 and the RoPE used in our adapter, to query and key features of each layer. As shown in Table 4, RoPE yields better performance than absolute embeddings. We further examine the impact of introducing a new linear layer, $L_{\\mathrm{cross}}^q$ , for query image", + "bbox": [ + 511, + 780, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7969", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/12b6c1ef75e83078802b4453324ac5e44772a56d29ef566dff0bd930cd00a86f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Exp.NumberQueryQimgQtxtKeyKconKistCanny(F1 Score↑)HED(SSIM↑)Seg.(mIoU↑)Depth(RMSE↓)Subject-Driven
DINO↑CLIP-I↑CLIP-T↑
1✓✓✓✓29.980.784030.8434.050.7780.8270.309
2✓✓✓29.010.776729.0035.390.7690.8250.307
3✓✓✓22.380.559922.9638.820.6940.7800.319
4✓✓✓31.320.793430.9433.160.7690.8230.308
5✓✓✓19.100.373017.0646.950.6690.7820.317
6✓✓29.240.778627.8134.120.7500.8170.313
", + "bbox": [ + 135, + 88, + 859, + 224 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/42f15238d4f9cf04507df71124a26e4baf1b6439f0292a2038867dee18976b52.jpg", + "table_caption": [ + "Table 3. Results of different query and key feature choices on pixel-level control tasks and subject-driven generation task." + ], + "table_footnote": [], + "table_body": "
MethodCanny(F1 Score↑)HED(SSIM↑)Seg.(mIoU↑)Depth(RMSE↓)Subject-Driven
DINO↑CLIP-I↑CLIP-T↑
w/o PE31.320.793430.9433.160.7690.8230.308
w/ Abs PE31.510.796730.3132.980.7400.8090.314
w/ Abs PE + Lqcross36.930.834833.2232.740.7510.8110.314
w/ RoPE34.650.824832.1333.450.7610.8170.313
w/ RoPE + Lqcross37.950.842033.3232.250.7840.8290.309
w/o txt36.310.838733.7131.600.7880.8300.308
w/o txt update36.310.838733.7131.600.7720.8210.310
ControlNet*38.180.838234.8131.460.7490.8150.313
IPAdapter-Instruct*15.810.219116.1247.810.7880.8520.294
", + "bbox": [ + 145, + 265, + 849, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Results of two baseline methods and different module combinations of our UNIC-Adapter on pixel-level control tasks and subject-driven generation task.", + "bbox": [ + 89, + 465, + 903, + 493 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "features in the main branch. Table 4 shows that the addition of $L_{\\mathrm{cross}}^q$ improves results further, particularly in pixel-level control tasks and subject-image similarity. This indicates the importance of $L_{\\mathrm{cross}}^q$ in enabling query features $Q_{\\mathrm{img}}$ to retrieve specific features from conditional image and task instruction. Figure 6 shows that UNIC-Adapter without position embedding often generates images with incomplete or incorrect details from control signals.", + "bbox": [ + 88, + 512, + 480, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with Baseline Methods. We reproduce two baseline methods, ControlNet* and IPAdapter-Instruct*, in our framework. For ControlNet*, we replace the cross-attention mechanism in the image-instruction feature injection module with an addition operation (as did in ControlNet) while retaining other settings. For IPAdapter-Instruct*, we reproduce results by referring to the open-source code on our datasets. As shown in Table 4, ControlNet* achieves performance comparable to our final model on pixel-level control tasks but exhibits suboptimal subject-image similarity on the subject-driven generation task. The visualization results in Figure 7 show that ControlNet* often produces images with distorted subject content. In contrast, IPAdapter-Instruct* performs well on subject-driven generation tasks but poorly on pixel-level control tasks. Due to compressing conditional images into 16 tokens with", + "bbox": [ + 89, + 633, + 482, + 876 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "out precise spatial information, the generated images from IPAdapter-Instruct* in Figure 6 fail to accurately capture the pixel-level details of the conditioned images. In comparison, our UNIC-Adapter achieves consistently superior results across both task types, demonstrating its effectiveness and suitability for unified controllable image generation.", + "bbox": [ + 511, + 513, + 905, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 616, + 633, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present UNIC-Adapter, a unified framework for controllable T2I generation that integrates diverse conditional image inputs and task instructions within a single model. Built on the MM-DiT architecture, our approach leverages cross-attention mechanisms and RoPE to achieve precise pixel-level control, as well as high subject fidelity, across multiple generation tasks. Our extensive experiments on tasks such as pixel-level spatial control, subject-driven image generation, and style-image-based T2I generation demonstrate the effectiveness of the UNIC-Adapter.", + "bbox": [ + 511, + 642, + 906, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 511, + 806, + 681, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research is supported by the NSFC Grants under the contracts No. 62325111 and No.U22B2011.", + "bbox": [ + 511, + 832, + 903, + 862 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "$\\S$ https://github.com/unity-research/IP-Adapter-Instruct", + "bbox": [ + 107, + 886, + 397, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7970", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 2, 3", + "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. GPT-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 5", + "[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science, 2(3):8, 2023. 1", + "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 6", + "[5] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 1, 2, 3", + "[6] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- $\\backslash$ sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024.", + "[7] Junsong Chen, Yue Wu, Simian Luo, Enze Xie, Sayak Paul, Ping Luo, Hang Zhao, and Zhenguo Li. Pixart- $\\{\\backslash$ delta\\} Fast and controllable image generation with latent consistency models. arXiv preprint arXiv:2401.05252, 2024. 1, 2", + "[8] Wenhu Chen, Hexiang Hu, Chitwan Sahara, and William W Cohen. Re-Imagen: Retrieval-augmented text-to-image generator. arXiv preprint arXiv:2209.14491, 2022. 5", + "[9] Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Ruiz, Xuhui Jia, Ming-Wei Chang, and William W Cohen. Subject-driven text-to-image generation via apprenticeship learning. NeurIPS, 36, 2024. 2, 5", + "[10] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, pages 2818-2829, 2023. 4", + "[11] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3", + "[12] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Int. Conf. Mach. and Learn., 2024. 1, 2, 3, 4, 5", + "[13] Junyao Gao, Yanchen Liu, Yanan Sun, Yinhao Tang, Yanhong Zeng, Kai Chen, and Cairong Zhao. StyleShot: A snap" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "shot on any style. arXiv preprint arXiv:2407.01414, 2024. 1, 2", + "[14] Zhen Han, Chaojie Mao, Zeyinzi Jiang, Yulin Pan, and Jingfeng Zhang. StyleBooth: Image style editing with multimodal instruction. arXiv preprint arXiv:2404.12154, 2024. 2, 5, 6", + "[15] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In CVPR, pages 4775–4785, 2024. 1", + "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 1, 2", + "[17] Hexiang Hu, Kelvin CK Chan, Yu-Chuan Su, Wenhu Chen, Yandong Li, Kihyuk Sohn, Yang Zhao, Xue Ben, Boqing Gong, William Cohen, et al. Instruct-Imagen: Image generation with multi-modal instruction. In CVPR, pages 4754–4763, 2024. 2, 3, 4", + "[18] Lianghua Huang, Di Chen, Yu Liu, Yujun Shen, Deli Zhao, and Jingren Zhou.Composer: Creative and controllable image synthesis with composable conditions. arXiv preprint arXiv:2302.09778, 2023. 1", + "[19] Frank Hutter Ilya Loshchilov. Decoupled weight decay regularization. *ICLR*, 2019. 5", + "[20] Imagen-Team-Google. Imagen 3, 2024. 1", + "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 5", + "[22] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 128(7):1956-1981, 2020. 5", + "[23] Black Forest Labs. Flux: Inference repository. https://github.com/black-forest-labs/flux, 2024. Accessed: 2024-11-14. 1, 2, 3, 4", + "[24] Dongxu Li, Junnan Li, and Steven Hoi. BLIP-Diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. NeurIPS, 36, 2024. 2, 5", + "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Int. Conf. Mach. and Learn., pages 19730-19742. PMLR, 2023. 5", + "[26] Ming Li, Taojiannan Yang, Huafeng Kuang, Jie Wu, Zhaoning Wang, Xuefeng Xiao, and Chen Chen. ControlNet++: Improving conditional controls with efficient consistency feedback. In ECCV, pages 129-147. Springer, 2025. 2, 5, 6", + "[27] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. GLIGEN: Open-set grounded text-to-image generation. In CVPR, pages 22511-22521, 2023. 5", + "[28] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et al. Hunyuan-DiT: A powerful multi-resolution diffusion transformer with fine-grained" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7971", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "chinese understanding. arXiv preprint arXiv:2405.08748, 2024.1,2", + "[29] Bingchen Liu, Ehsan Akhgari, Alexander Visheratin, Aleks Kamko, Linmiao Xu, Shivam Shrirao, Joao Souza, Suhail Doshi, and Daiqing Li. Playground v3: Improving text-to-image alignment with deep-fusion large language models. arXiv preprint arXiv:2409.10695, 2024. 1", + "[30] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding DINO: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 5", + "[31] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-Diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. In ACM SIGGRAPH Conference Papers, pages 1-12, 2024. 2, 5", + "[32] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, pages 6038–6047, 2023. 5", + "[33] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, pages 4296-4304, 2024. 1, 5", + "[34] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-G: Generating images in context with multimodal large language models. arXiv preprint arXiv:2310.02992, 2023. 1, 5, 6", + "[35] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 1", + "[36] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 2", + "[37] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 5", + "[38] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 2", + "[39] Can Qin, Shu Zhang, Ning Yu, Yihao Feng, Xinyi Yang, Yingbo Zhou, Huan Wang, Juan Carlos Niebles, Caiming Xiong, Silvio Savarese, et al. UniControl: A unified diffusion model for controllable visual generation in the wild. arXiv preprint arXiv:2305.11147, 2023. 2, 3, 5", + "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Int. Conf. Mach. and Learn., pages 8748-8763. PMLR, 2021. 2, 3, 4, 5, 6", + "[41] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 2, 3, 4" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 2", + "[43] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 234–241. Springer, 2015. 1", + "[44] Ciara Rowles, Shimon Vainer, Dante De Nigris, Slava Elizarov, Konstantin Kutsy, and Simon Donne. IPAdapter-Instruct: Resolving ambiguity in image-based conditioning using instruct prompts. arXiv preprint arXiv:2408.03209, 2024. 2, 3", + "[45] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500-22510, 2023. 1, 5, 6", + "[46] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 35:36479-36494, 2022. 2, 3", + "[47] Babak Saleh and Ahmed Elgammal. Large-scale classification of fine-art paintings: Learning the right metric on the right feature. arXiv preprint arXiv:1505.00855, 2015. 5", + "[48] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. LAION-5B: An open large-scale dataset for training next generation image-text models. NeurIPS, 35:25278-25294, 2022. 1", + "[49] Kunpeng Song, Yizhe Zhu, Bingchen Liu, Qing Yan, Ahmed Elgammal, and Xiao Yang. Moma: Multimodal llm adapter for fast personalized image generation. arXiv preprint arXiv:2404.05674, 2024. 2", + "[50] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. RoFormer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 2, 4", + "[51] Kcolors Team. Kcolors: Effective training of diffusion model for photorealistic text-to-image synthesis. arXiv preprint, 2024. 1", + "[52] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2", + "[53] A Vaswani. Attention is all you need. NeurIPS, 2017. 2, 3", + "[54] Haofan Wang, Matteo Spinelli, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024. 1, 2", + "[55] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 1, 4" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7972", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. arXiv preprint arXiv:2409.11340, 2024. 2, 3, 5, 6", + "[57] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. CogVideoX: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 4", + "[58] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. IP-Adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2, 3, 4", + "[59] Zilyu Ye, Jinxiu Liu, Ruotian Peng, Jinjin Cao, Zhiyang Chen, Yiyang Zhang, Ziwei Xuan, Mingyuan Zhou, Xiaogian Shen, Mohamed Elhoseiny, et al. Openstory++: A large-scale dataset and benchmark for instance-aware open-domain visual storytelling. arXiv preprint arXiv:2408.03695, 2024. 5", + "[60] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 1, 2, 5", + "[61] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. UniControlNet: All-in-one control to text-to-image diffusion models. NeurIPS, 36, 2024. 2, 3, 5", + "[62] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In CVPR, pages 22490-22499, 2023. 1, 2", + "[63] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, pages 633-641, 2017. 5" + ], + "bbox": [ + 91, + 92, + 480, + 556 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7973", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/UNIC-Adapter_ Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation/409eb6c5-d571-4968-a329-2c1ebf51e321_model.json b/2025/UNIC-Adapter_ Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation/409eb6c5-d571-4968-a329-2c1ebf51e321_model.json new file mode 100644 index 0000000000000000000000000000000000000000..91a6464b6526ce0357e98caa27959124a66be46d --- /dev/null +++ b/2025/UNIC-Adapter_ Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation/409eb6c5-d571-4968-a329-2c1ebf51e321_model.json @@ -0,0 +1,2488 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.149, + 0.13, + 0.85, + 0.177 + ], + "angle": 0, + "content": "UNIC-Adapter: Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.162, + 0.203, + 0.835, + 0.241 + ], + "angle": 0, + "content": "Lunhao Duan\\*1 Shanshan Zhao\\*2 Wenjun Yan2 Yinglun Li2 Qing-Guo Chen2 Zhao Xu2 Weihua Luo2 Kaifu Zhang2 Mingming Gong3 Gui-Song Xia\\*14" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.256, + 0.865, + 0.294 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science, Wuhan University \\(^{2}\\)Alibaba International Digital Commerce \\(^{3}\\)University of Melbourne \\(^{4}\\)School of Artificial Intelligence, Wuhan University" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.359, + 0.485, + 0.751 + ], + "angle": 0, + "content": "Recently, text-to-image generation models have achieved remarkable advancements, particularly with diffusion models facilitating high-quality image synthesis from textual descriptions. However, these models often struggle with achieving precise control over pixel-level layouts, object appearances, and global styles when using text prompts alone. To mitigate this issue, previous works introduce conditional images as auxiliary inputs for image generation, enhancing control but typically necessitating specialized models tailored to different types of reference inputs. In this paper, we explore a new approach to unify controllable generation within a single framework. Specifically, we propose the unified image-instruction adapter (UNIC-Adapter) built on the Multi-Modal-Diffusion Transformer architecture, to enable flexible and controllable generation across diverse conditions without the need for multiple specialized models. Our UNIC-Adapter effectively extracts multi-modal instruction information by incorporating both conditional images and task instructions, injecting this information into the image generation process through a cross-attention mechanism enhanced by Rotary Position Embedding. Experimental results across a variety of tasks, including pixel-level spatial control, subject-driven image generation, and style-image-based image synthesis, demonstrate the effectiveness of our UNIC-Adapter in unified controllable image generation." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.767, + 0.222, + 0.783 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.792, + 0.483, + 0.855 + ], + "angle": 0, + "content": "Recent advancements in text-to-image (T2I) generation models [3, 5, 12, 20, 23, 29, 38, 42, 51], primarily driven by diffusion models [16], have been witnessed in recent years. By leveraging large-scale paired text-image data [48], nu" + }, + { + "type": "text", + "bbox": [ + 0.548, + 0.332, + 0.561, + 0.342 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "text", + "bbox": [ + 0.582, + 0.329, + 0.627, + 0.346 + ], + "angle": 0, + "content": "A delicious afternoon tea." + }, + { + "type": "text", + "bbox": [ + 0.646, + 0.329, + 0.706, + 0.346 + ], + "angle": 0, + "content": "A room with the forest backdrop." + }, + { + "type": "text", + "bbox": [ + 0.716, + 0.329, + 0.782, + 0.346 + ], + "angle": 0, + "content": "A robot in autumn mountain and lake" + }, + { + "type": "text", + "bbox": [ + 0.786, + 0.332, + 0.799, + 0.342 + ], + "angle": 0, + "content": "···" + }, + { + "type": "text", + "bbox": [ + 0.815, + 0.329, + 0.871, + 0.346 + ], + "angle": 0, + "content": "A dog with a mountain in the" + }, + { + "type": "text", + "bbox": [ + 0.549, + 0.359, + 0.561, + 0.368 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "text", + "bbox": [ + 0.585, + 0.353, + 0.627, + 0.376 + ], + "angle": 0, + "content": "Generate image from canny edge." + }, + { + "type": "text", + "bbox": [ + 0.648, + 0.352, + 0.706, + 0.376 + ], + "angle": 0, + "content": "Render image with this depth map." + }, + { + "type": "text", + "bbox": [ + 0.716, + 0.353, + 0.779, + 0.376 + ], + "angle": 0, + "content": "Create image using style influence." + }, + { + "type": "text", + "bbox": [ + 0.784, + 0.359, + 0.798, + 0.368 + ], + "angle": 0, + "content": "1" + }, + { + "type": "text", + "bbox": [ + 0.811, + 0.359, + 0.871, + 0.376 + ], + "angle": 0, + "content": "Generate image from dog image" + }, + { + "type": "text", + "bbox": [ + 0.549, + 0.403, + 0.561, + 0.412 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.379, + 0.639, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.379, + 0.712, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.379, + 0.787, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.379, + 0.876, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.549, + 0.481, + 0.561, + 0.49 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.437, + 0.639, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.437, + 0.712, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.437, + 0.787, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.437, + 0.876, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.522, + 0.907, + 0.579 + ], + "angle": 0, + "content": "Figure 1. With the UNIC-Adapter, SD3 enables flexible and controllable generation across multiple reference modalities within a single model. (a) to (d) represent text prompts, task instructions, conditional images, and generated images, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.907, + 0.717 + ], + "angle": 0, + "content": "merous open-source T2I models, like Stable Diffusion 1.5 (SD1.5) [42], Stable Diffusion XL (SDXL) [38], Pixart-series [5-7], Stable Diffusion 3 (SD3) [12], Hunyuan-DiT [28], and FLUX.1-dev [23] have significantly enhanced the fidelity of images generated from natural language descriptions. These models utilize a range of architectures, spanning from U-Net-based designs [43] to the more recent Diffusion-Transformer-based frameworks [35]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Despite these advancements, relying solely on text prompts often falls short in defining pixel-level spatial structures and geometric details, exact object appearances, and global styles of a specific image. To address this limitation, recent research [13, 18, 33, 34, 45, 55, 58, 60] has incorporated conditional images as additional inputs alongside text prompts. These conditional images offer finer control over pixel-level [60, 62], object-level [34, 55, 58], or style-specific [13, 15, 54] features in the generated outputs. For example, ControlNet [60] enables pixel-level spatial control by integrating structural inputs, such as depth and edge maps, through a parallel encoder tailored for" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.483, + 0.901 + ], + "angle": 0, + "content": "*This work was done when Lunhao Duan was a research intern at Alibaba International Digital Commerce. \n†Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7963" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.364 + ], + "angle": 0, + "content": "each type of control. Similarly, IP-Adapter [58] enables content-specific control by conditioning on a reference image and injecting its CLIP-based [40] embeddings via crossattention layers. However, these methods often necessitate individual model training for each control type, resulting in increased training costs and complicating the integration of different control inputs. IPAdapter-Instruct [44] extends IP-Adapter by incorporating instruction prompts to specify task types for different conditional images, enabling multi-task handling within a single model. UniControl [39] tackles various pixel-level control tasks within a unified model by introducing task-specific adapters to extract features for different visual conditions. Meanwhile, Uni-ControlNet [61] divides conditional images into two groups: local (pixel-level) and global (image-level) conditions, and designs dedicated adapters for each group. However, this approach is not fully unified, as the two adapters are trained separately." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.483, + 0.565 + ], + "angle": 0, + "content": "To further unify diverse reference inputs, including text, edge maps, content images, etc., Instruct-Imagen [17], based onImagen [46], represents them as multi-modal instructions within a single framework. OmniGen [56] first tokenizes interleaved text and images into a one-dimensional sequence and then uses a transformer model (Phi3) [1] to achieve unified image generation, such as image editing and subject-driven generation. To adapt the model for multi-modal instructions, these two methods involve multi-stage training of the entire model. To enable more efficient training for unified controllable image generation, this paper seeks to design an adapter based on a pre-trained T2I model, avoiding the need to train the entire model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Given that DiT-based architectures [36] equipped with pure transformer blocks have become the mainstream in T2I models [5, 12, 23, 28], we choose SD3, an open-source state-of-the-art (SOTA) model, as the foundation model for studying the adapter. To achieve unified controllable image generation across various conditions, we employ the Multi-Modal-Diffusion Transformer (MM-DiT) structure [12, 23], which enables full cross-attention between text and image features to extract multi-modal instruction information. By treating text as an independent modality, the MM-DiT allocates substantial model capacity to capture and understand text-based information, thereby improving its interpretation of text prompts [12]. Leveraging the strengths of the MM-DiT, we propose a unified image-instruction adapter (UNIC-Adapter) to facilitate image generation conditioned on various types of reference images with a task instruction indicating the desired generation target. Specifically, task instruction and conditional image features are extracted by text encoders [40, 41] and a Variational Autoencoder (VAE) [12], respectively. Then, the adapter processes instruction and image features using MM-DiT blocks, which are subsequently integrated into the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.197 + ], + "angle": 0, + "content": "main generation branch via a cross-attention mechanism. To further enhance spatial awareness, which is essential for pixel-level control, we incorporate Rotary Position Embedding (RoPE) [50] into the query and key features within the cross-attention. With the proposed UNIC-Adapter, SD3 can generate images under a wide array of conditions, as shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.198, + 0.825, + 0.212 + ], + "angle": 0, + "content": "In summary, our contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.214, + 0.905, + 0.258 + ], + "angle": 0, + "content": "- We are the first to leverage the MM-DiT architecture for unified controllable generation across various types of reference images in T2I models." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.259, + 0.904, + 0.303 + ], + "angle": 0, + "content": "- We introduce an effective feature injection method based on cross-attention, enhanced with RoPE encoding for better spatial control." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.304, + 0.904, + 0.349 + ], + "angle": 0, + "content": "- Equipped with our UNIC-Adapter, SD3 enables unified controllable generation across 14 types of conditional images in our experiments." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.214, + 0.905, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.365, + 0.655, + 0.381 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.391, + 0.809, + 0.407 + ], + "angle": 0, + "content": "2.1. Text-to-Image Generation Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.907, + 0.747 + ], + "angle": 0, + "content": "Mainstream T2I generation models [5, 12, 38, 42] are primarily based on diffusion models [16], which generate images by learning to reverse a progressive noise-adding process. In these models, text input typically serves as a conditional signal, guiding the generation process to align with the provided description. The success of transformer architectures [53], particularly in large language models [2, 52], has significantly influenced T2I model architectures, evolving them from convolution-transformer hybrids [38, 42] to fully transformer-based designs [5-7, 12, 23, 28]. Recently, the MM-DiT architecture [12, 23] has advanced this evolution by treating text and image features as distinct modalities with full-attention mechanisms, enabling each modality to attend to the other. This approach outperforms traditional cross-attention setups, where text features are used only as keys and values [12]. However, relying solely on text-based conditioning presents limitations, particularly for images that require detailed spatial layouts or specific object appearances. Our work builds upon these transformer-based diffusion models by integrating additional conditional images and task instructions, thereby enhancing the flexibility and controllability of T2I generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.836, + 0.773 + ], + "angle": 0, + "content": "2.2. Conditional Image-based T2I Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Conditional image-based T2I models introduce auxiliary image inputs to complement textual prompts, enabling finer control over aspects such as spatial layout [26, 60, 62], object content [9, 24, 31, 49, 58], and stylistic elements [13, 14, 54]. For pixel-level spatial control, ControlNet [60] employs a duplicated U-Net encoder to process specific types of conditional images (e.g., edge, pose, and depth maps) and integrates the features into the main T2I model" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7964" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.156, + 0.089, + 0.845, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.36, + 0.908, + 0.417 + ], + "angle": 0, + "content": "Figure 2. The overall architecture of our proposed UNIC-Adapter. The task instruction and conditional image features are progressively attending to each other through a series of \\(N\\) adapter blocks. In each adapter block, the image features \\(Z_{\\mathrm{img}}\\) from MM-DiT block in the main image generation branch serve as the query, while both task instruction features \\(Z_{\\mathrm{ist}}\\) and conditional image features \\(Z_{\\mathrm{con}}\\) function as keys and values. For simplicity, normalization layers and feed-forward networks are omitted in this figure." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.443, + 0.485, + 0.896 + ], + "angle": 0, + "content": "through residual connections in the U-Net decoder. Another key work, IP-Adapter [58], uses the Vision Transformer (ViT) [11] encoder from CLIP [40] to extract features from conditional images. These features are then incorporated into the U-Net backbone via cross-attention layers, achieving controllable generation with contents from conditional images. IPAdapter-Instruct [44] extends this by introducing an instructional prompt that guides the interpretation of conditional images. To minimize the necessity for individual model training for each type of conditional image in ControlNet, UniControl [39] unifies nine pixel-level control tasks within a single model, leveraging task-specific feature extraction adapters and a task-aware HyperNet to modulate features based on task instructions. To achieve control with image content as well as spatial layout, UniControlNet [61] incorporates two dedicated adapters for local (pixel-level control) and global (CLIP-image embedding control) conditions, respectively. Unlike these methods, which freeze the base model and only training the newly added modules / adapters, Instruct-Imagen [17] proposes using multi-modal instructions and optimize the entire model, including the base model (Imagen [46]) and the new cross-attention layers, for image generation under different conditions. A recent work, OmniGen [56], develops a unified image generation model for various multi-modal inputs using a transformer-based architecture [1]. In this paper, we aim to enhance the generation capabilities of an existing T2I model for diverse conditions by introducing a unified image-instruction adapter, named UNIC-Adapter, without the need to train the entire model." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.442, + 0.605, + 0.457 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.907, + 0.543 + ], + "angle": 0, + "content": "In this section, we first provide a brief overview of the DiT and MM-DiT employed in SOTA T2I models. Then, we describe our UNIC-Adapter, which incorporates task instructions and conditional images to enhance the model's capability to interpret various user-specified conditions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.55, + 0.644, + 0.566 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.909, + 0.697 + ], + "angle": 0, + "content": "Diffusion Transformer Block. In DiT Block-based T2I models [5], conditional text features \\( Z_{\\mathrm{txt}} \\) extracted from text encoders [40, 41] are integrated into DiT Blocks via cross-attention layers. Specifically, in each DiT Block, the conditional text features are first mapped to key \\( K_{\\mathrm{txt}} \\) and value \\( V_{\\mathrm{txt}} \\) features through linear transformations \\( L \\), while the image feature \\( Z_{\\mathrm{img}} \\) is mapped to a query feature \\( Q_{\\mathrm{img}} \\). An attention mechanism then updates the image features \\( Z_{\\mathrm{img}} \\) by attending to \\( K_{\\mathrm{txt}} \\) and \\( V_{\\mathrm{txt}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.701, + 0.905, + 0.726 + ], + "angle": 0, + "content": "\\[\nK _ {\\mathrm {t x t}} = L _ {\\mathrm {t x t}} ^ {k} \\left(Z _ {\\mathrm {t x t}}\\right), V _ {\\mathrm {t x t}} = L _ {\\mathrm {t x t}} ^ {v} \\left(Z _ {\\mathrm {t x t}}\\right), Q _ {\\mathrm {i m g}} = L _ {\\mathrm {i m g}} ^ {q} \\left(Z _ {\\mathrm {i m g}}\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.723, + 0.715, + 0.74 + ], + "angle": 0, + "content": "\\[\nZ _ {\\mathrm {i m g}} ^ {\\prime} = \\operatorname {A t t n} \\left(Q _ {\\mathrm {i m g}}, K _ {\\mathrm {t x t}}, V _ {\\mathrm {t x t}}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.743, + 0.907, + 0.803 + ], + "angle": 0, + "content": "where Attn indicates the multi-head attention [53]. In this setup, the same \\( Z_{\\mathrm{txt}} \\) features are provided to DiT Blocks across different layers as static conditioning signals, which may limit the utilization of text features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.803, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Multi-modal Diffusion Transformer Block. Recent T2I models [12, 23] based on MM-DiT Blocks enhance DiT Blocks by treating conditional text as an independent modality capable of dynamically interacting with image features across layers. In MM-DiT, both text and image features serve as queries, keys, and values, enabling cross-modal attention in both directions through a full-attention" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7965" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "mechanism. Specifically, the MM-DiT layer performs the following operations:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.129, + 0.483, + 0.153 + ], + "angle": 0, + "content": "\\[\nQ _ {\\mathrm {t x t}} = L _ {\\mathrm {t x t}} ^ {q} \\left(Z _ {\\mathrm {t x t}}\\right), K _ {\\mathrm {t x t}} = L _ {\\mathrm {t x t}} ^ {k} \\left(Z _ {\\mathrm {t x t}}\\right), V _ {\\mathrm {t x t}} = L _ {\\mathrm {t x t}} ^ {v} \\left(Z _ {\\mathrm {t x t}}\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.151, + 0.457, + 0.167 + ], + "angle": 0, + "content": "\\[\nQ _ {\\mathrm {i m g}} = L _ {\\mathrm {i m g}} ^ {q} \\left(Z _ {\\mathrm {i m g}}\\right), K _ {\\mathrm {i m g}} = L _ {\\mathrm {i m g}} ^ {k} \\left(Z _ {\\mathrm {i m g}}\\right), V _ {\\mathrm {i m g}} = L _ {\\mathrm {i m g}} ^ {v} \\left(Z _ {\\mathrm {i m g}}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.174, + 0.483, + 0.22 + ], + "angle": 0, + "content": "These queries, keys, and values are processed in a full-attention layer, where each modality attends to both text and image features:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.23, + 0.483, + 0.256 + ], + "angle": 0, + "content": "\\[\nZ _ {\\mathrm {t x t}} ^ {\\prime} = \\operatorname {A t t n} \\left(Q _ {\\mathrm {t x t}}, \\left[ K _ {\\mathrm {t x t}} \\| K _ {\\mathrm {i m g}} \\right], \\left[ V _ {\\mathrm {t x t}} \\| V _ {\\mathrm {i m g}} \\right]\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.253, + 0.435, + 0.272 + ], + "angle": 0, + "content": "\\[\nZ _ {\\mathrm {i m g}} ^ {\\prime} = \\operatorname {A t t n} \\left(Q _ {\\mathrm {i m g}}, \\left[ K _ {\\mathrm {t x t}} \\right\\| K _ {\\mathrm {i m g}} ], \\left[ V _ {\\mathrm {t x t}} \\right\\| V _ {\\mathrm {i m g}}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.281, + 0.484, + 0.343 + ], + "angle": 0, + "content": "where \\( \\parallel \\) denotes concatenation. This enhanced interaction facilitates richer cross-modal feature extraction, as text and image features are iteratively refined across layers, improving text-guided image generation [12]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.351, + 0.245, + 0.367 + ], + "angle": 0, + "content": "3.2. UNIC-Adapter" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.373, + 0.484, + 0.476 + ], + "angle": 0, + "content": "As shown in Figure 2, the UNIC-Adapter introduces two additional inputs: task instructions and conditional images. This integration enables unified controllable image generation by allowing the model to interpret and apply specific features from a variety of conditional images, guided by the task instruction. Below, we provide a detailed description of its structure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.484, + 0.671 + ], + "angle": 0, + "content": "Image-instruction Feature Extraction. To enable a comprehensive understanding of task instructions and conditional images, we leverage MM-DiT blocks to enhance the interaction between these inputs. Specifically, we adopt SD3 as our base model. Following the setup in SD3, task instruction features, denoted as \\( Z_{\\mathrm{ist}} \\), are extracted using text encoders such as CLIP-G/14 [10], CLIP-L/15 [40], and T5 XXL [41]. Simultaneously, conditional image features, \\( Z_{\\mathrm{con}} \\), which represent one of several condition types (e.g., edge maps, depth maps, subject content images, or style-reference images), are obtained through the VAE. Within the MM-DiT blocks, task instruction and conditional image features undergo mutual attention updates, defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.678, + 0.408, + 0.695 + ], + "angle": 0, + "content": "\\[\nQ _ {\\text {i s t}} = L _ {\\text {i s t}} ^ {q} (Z _ {\\text {i s t}}), K _ {\\text {i s t}} = L _ {\\text {i s t}} ^ {k} (Z _ {\\text {i s t}}), V _ {\\text {i s t}} = L _ {\\text {i s t}} ^ {v} (Z _ {\\text {i s t}}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.697, + 0.483, + 0.724 + ], + "angle": 0, + "content": "\\[\nQ _ {\\text {c o n}} = L _ {\\text {c o n}} ^ {q} \\left(Z _ {\\text {c o n}}\\right), K _ {\\text {c o n}} = L _ {\\text {c o n}} ^ {k} \\left(Z _ {\\text {c o n}}\\right), V _ {\\text {c o n}} = L _ {\\text {c o n}} ^ {v} \\left(Z _ {\\text {c o n}}\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.719, + 0.353, + 0.737 + ], + "angle": 0, + "content": "\\[\nZ _ {\\mathrm {i s t}} ^ {\\prime} = \\operatorname {A t t n} \\left(Q _ {\\mathrm {i s t}}, \\left[ K _ {\\mathrm {i s t}} \\| K _ {\\mathrm {c o n}} \\right], \\left[ V _ {\\mathrm {i s t}} \\| V _ {\\mathrm {c o n}} \\right]\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.741, + 0.361, + 0.758 + ], + "angle": 0, + "content": "\\[\nZ _ {\\text {c o n}} ^ {\\prime} = \\operatorname {A t t n} \\left(Q _ {\\text {c o n}}, \\left[ K _ {\\text {i s t}} \\| K _ {\\text {c o n}} \\right], \\left[ V _ {\\text {i s t}} \\| V _ {\\text {c o n}} \\right]\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.811 + ], + "angle": 0, + "content": "These updates ensure that the task instruction \\( Z_{\\mathrm{ist}} \\) and conditional image \\( Z_{\\mathrm{con}} \\) dynamically refine each other's features by selectively attending to relevant information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Image-instruction Feature Injection. To improve the model's responsiveness to diverse user-defined conditions, the key and value features from both instruction and conditional image inputs are integrated simultaneously into the main image generation branch through cross-attention [17, 55, 58]. This mechanism ensures that the feature injection is" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.215 + ], + "angle": 0, + "content": "directed by the specific task instruction. Experiments indicate that injecting either conditional image features or task instruction features alone through cross-attention results in suboptimal performance, as demonstrated in the ablation studies of Section 4. By contrast, the simultaneous incorporation of both feature types optimally supports unified controlled image generation. Accordingly, the image features \\( Z_{\\mathrm{img}}^{\\prime} \\) from Equation 3 are further updated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.232, + 0.905, + 0.259 + ], + "angle": 0, + "content": "\\[\nQ _ {\\mathrm {i m g}} ^ {\\prime} = L _ {\\text {c r o s s}} ^ {q} \\left(Z _ {\\mathrm {i m g}}\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.256, + 0.868, + 0.275 + ], + "angle": 0, + "content": "\\[\nZ _ {\\mathrm {i m g}} ^ {\\prime \\prime} = Z _ {\\mathrm {i m g}} ^ {\\prime} + \\operatorname {A t t n} \\left(Q _ {\\mathrm {i m g}} ^ {\\prime}, \\left[ K _ {\\mathrm {i s t}} \\| K _ {\\mathrm {c o n}} \\right], \\left[ V _ {\\mathrm {i s t}} \\| V _ {\\mathrm {c o n}} \\right]\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.28, + 0.907, + 0.386 + ], + "angle": 0, + "content": "In this setup, we introduce a new linear layer, \\( L_{\\mathrm{cross}}^q \\), designed to enhance the model's ability to attend to image-instruction features effectively. This additional layer has been shown to improve performance, as validated in our experiments. Overall, this cross-attention operation facilitates controlled image generation, leveraging desired attributes from both the conditional image and the task instruction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.906, + 0.552 + ], + "angle": 0, + "content": "Position Embedding. The positional information of conditional image features is crucial for pixel-level control tasks, as methods like ControlNet incorporate features through pixel-to-pixel addition. To improve pixel-level control precision, we integrate Rotary Position Embedding (RoPE) [50] into both the query and key features prior to cross-attention. RoPE provides relative positional encoding, ensuring that query and key features with closer pixel coordinates yield higher similarity scores, thereby improving spatial alignment in the generated output. It has been used in T2I [23] and text-to-video [57] models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.906, + 0.637 + ], + "angle": 0, + "content": "Specifically, we apply 1-D RoPE separately to the height and width dimensions. Given a feature vector \\( f \\in \\mathbb{R}^{|D|} \\) at position \\( (h,w) \\), we split \\( f \\) into two halves: \\( f_h \\in \\mathbb{R}^{|D|/2} \\) for the height dimension and \\( f_w \\in \\mathbb{R}^{|D|/2} \\) for the width. For the height component \\( f_h \\), the rotary matrix \\( R_h \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.642, + 0.906, + 0.723 + ], + "angle": 0, + "content": "\\[\nR _ {h} = \\left[ \\begin{array}{c c c c c} \\cos (h \\theta_ {0}) & - \\sin (h \\theta_ {0}) & \\dots & 0 & 0 \\\\ \\sin (h \\theta_ {0}) & \\cos (h \\theta_ {0}) & \\dots & 0 & 0 \\\\ \\vdots & \\vdots & \\ddots & \\vdots & \\vdots \\\\ 0 & 0 & \\dots & \\cos (h \\theta_ {l}) & - \\sin (h \\theta_ {l}) \\\\ 0 & 0 & \\dots & \\sin (h \\theta_ {l}) & \\cos (h \\theta_ {l}) \\end{array} \\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.729, + 0.906, + 0.788 + ], + "angle": 0, + "content": "where \\(\\theta_{i} = b^{-i / |D|}\\), with \\(b = 10000\\) and \\(l = |D| / 4 - 1\\). The rotary matrix \\(R_{w}\\) for the width component \\(f_{w}\\) is defined analogously. The updated feature vector \\(f\\) is then computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.79, + 0.905, + 0.807 + ], + "angle": 0, + "content": "\\[\nf = \\left[ R _ {h} f _ {h} \\| R _ {w} f _ {w} \\right]. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.9 + ], + "angle": 0, + "content": "RoPE is applied to both query and key features in Equation 5 and Equation 4, with the position coordinates of \\( K_{\\mathrm{ins}} \\) set to \\( (0,0) \\). By embedding relative positional information in this way, RoPE enhances the model's precision in pixel-level control, while maintaining robustness and adaptability across various tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7966" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.224, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.207 + ], + "angle": 0, + "content": "In this section, we conduct extensive experiments to evaluate the performance of our UNIC-Adapter in various conditional image generation tasks. Detailed ablation studies are also conducted to analyze the contribution of individual components within our framework. Additional experimental results are available in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.217, + 0.274, + 0.232 + ], + "angle": 0, + "content": "4.1. Tasks and Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.239, + 0.483, + 0.3 + ], + "angle": 0, + "content": "Our evaluation covers three major tasks: pixel-level spatial control, subject-driven image generation, and style-image-based T2I generation. Below, we briefly describe the datasets used for training each task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.484, + 0.496 + ], + "angle": 0, + "content": "Pixel-level Spatial Control. For pixel-level spatial control, we employ the MultiGen-20M dataset [39], which includes 12 types of pixel-level control annotations, such as Canny, HED, Sketch, Depth, Normal, Skeleton, Bbox, Segmentation, Outpainting, Inpainting, Deblurring, and Colorization. The dataset contains 2.8 million images with 20 million control annotations, and we reserve 1,000 samples for qualitative evaluation. For segmentation map conditioning, we also use the ADE20k [63] training set, which includes approximately 20,000 samples. These annotations serve as conditional images, and we use GPT-4o [2] to generate 20 synonymous instructions for each control type, such as \"Generate an image from this edge map.\"" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.497, + 0.484, + 0.679 + ], + "angle": 0, + "content": "Subject-driven Image Generation. For subject-driven image generation, we use three datasets: OpenImages v7 [22], GRIT [37], and OpenStory++ [59]. Following previous works [31, 34], we utilize Blip-2 [25] to generate captions for images without annotations and use spaCy‡ to extract subject tags. GroundingDINO [30] is used to obtain bounding boxes for each subject, and SAM [21] is employed to generate subject segmentation maps based on these bounding boxes. After processing and filtering, we obtain 2.1 million subject-image pairs. We also use GPT4o to generate 20 instructions, such as \"Generate an image from this dog image.\"" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.679, + 0.484, + 0.815 + ], + "angle": 0, + "content": "Style-image-based Image Generation. For style-image-based generation, we use the WikiArt [47] and Style-Booth [14] datasets. For StyleBooth, we randomly select an image from the same style subset as the conditional image, and for WikiArt, we employ CLIP-I [40] similarity to retrieve a conditional image for each training image. These two datasets comprise approximately 90,000 images. Additionally, we use GPT-4o to generate 20 instructions, such as \"Generate an image based on this style image.\"" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.825, + 0.308, + 0.84 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.847, + 0.483, + 0.877 + ], + "angle": 0, + "content": "Our UNIC-Adapter framework is based on the SD3 medium [12] model. Since instruction texts and conditional" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.09, + 0.915, + 0.273 + ], + "angle": 0, + "content": "
MethodBase ModelCanny (F1 Score↑)HED (SSIM↑)Seg. (mIoU↑)Depth (RMSE↓)
Single-task
T2I-Adapter [33]SD1.423.65-12.6148.40
Gligen [27]SD1.426.940.563423.7838.83
ControlNet [60]SD1.534.650.762132.5535.90
ControlNet++ [26]SD1.537.040.809743.6428.32
Multi-task
UniControl [39]SD1.530.820.796925.4439.18
Uni-ControlNet [61]SD1.527.320.691019.3940.65
OmniGen [56]Phi-335.540.823744.2328.54
OursSD338.940.836942.8931.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.283, + 0.907, + 0.34 + ], + "angle": 0, + "content": "Table 1. Results of different methods on four pixel-level control image generation tasks. \\(\\uparrow\\) indicates that higher values are better, while \\(\\downarrow\\) indicates that lower values are better. The best results for each metric are bolded, and the second-best results are underlined." + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.355, + 0.892, + 0.525 + ], + "angle": 0, + "content": "
MethodBase ModelDINO↑CLIP-I↑CLIP-T↑
Fine-Tuning
Textual Inversion [32]SD1.50.5690.7800.255
DreamBooth [45]SD1.50.6680.8030.305
BLIP-Diffusion [24]SD1.50.6700.8050.302
Test Time Tuning Free
Re-Imagen [8]-0.6000.7400.270
SuTI [9]Imagen0.7410.8190.304
Kosmos-G [34]SD1.50.6940.8470.287
OmniGen [56]Phi-30.8010.8470.301
OursSD30.8160.8410.306
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.535, + 0.907, + 0.565 + ], + "angle": 0, + "content": "Table 2. Results of different methods on the DreamBench dataset [45] for subject-driven image generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.578, + 0.907, + 0.73 + ], + "angle": 0, + "content": "are variations of text and image modalities, the UNIC-Adapter is initialized using the base T2I model's parameters, which minimizes learning difficulty. We freeze the parameters of base SD3 medium model and the feed-forward layer after the attention layer in adapter blocks, and train all other newly introduced parameters, resulting in approximately 1.2 billion trainable parameters. We use the AdamW optimizer [19] with a learning rate of 0.0001 and a weight decay of 0.01. Training is performed on 16 H100 GPUs for 100,000 steps, with a batch size of 16 per GPU." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.741, + 0.653, + 0.756 + ], + "angle": 0, + "content": "4.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.906, + 0.794 + ], + "angle": 0, + "content": "We present experimental results of the proposed UNIC-Adapter framework across three major tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Pixel-Level Spatial Control. Following previous work [26, 56], we evaluate performance based on the similarity between control conditions extracted from generated images and the specified pixel-level control inputs. The evaluation is conducted on the test set of MultiGen-20M with canny edge, hed edge, and depth map conditions and on ADE20k test set with segmentation map condition. As" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.305, + 0.901 + ], + "angle": 0, + "content": "\\(\\ddagger\\)https://github.com/explosion/spaCy" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7967" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.905, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.548, + 0.908, + 0.577 + ], + "angle": 0, + "content": "Figure 3. Visualization results of our UNIC-Adapter on twelve pixel-level control tasks from the MultiGen-20M dataset. The first and third rows show different types of conditional images, while the second and fourth rows display the corresponding generated images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.602, + 0.484, + 0.737 + ], + "angle": 0, + "content": "shown in Table 1, UNIC-Adapter achieves the best performance on canny and hed edge map conditions, and demonstrates comparable performance on depth and segmentation map conditions, compared to SOTA methods like ControlNet++ [26] for single-task and OmniGen [56] for multitask scenarios. Figure 3 shows qualitative results on 12 tasks from MultiGen-20M, illustrating that UNIC-Adapter consistently produces images accurately aligned with pixel-level controls." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Subject-Driven Image Generation. Evaluation is conducted using 30 objects and 25 prompts from the Dream-Bench dataset [45]. Following prior work [34], we generate 4 images for each prompt and measure subject fidelity using DINO [4] and CLIP-I scores and text fidelity with CLIP-T [40]. As shown in Table 2, UNIC-Adapter achieves high subject-image consistency and maintains strong alignment between the generated images and text prompts. Qualitative examples in Figure 4 demonstrate that UNIC-Adapter effectively generates images closely resembling input sub" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.603, + 0.707, + 0.617 + ], + "angle": 0, + "content": "jects across various contexts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.619, + 0.907, + 0.71 + ], + "angle": 0, + "content": "Style-image-based Image Generation. For the same text prompt, style images from StyleBooth [14] are randomly selected as conditional images. The qualitative results in Figure 5 indicate that UNIC-Adapter effectively captures and applies stylistic elements from conditional images, generating artistically consistent outputs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.724, + 0.68, + 0.739 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.907, + 0.838 + ], + "angle": 0, + "content": "To better understand the impact of individual components in the UNIC-Adapter framework, we conduct ablation studies by systematically modifying or removing key modules. For efficiency, the models in this part are trained on the subject-driven image generation task and four tasks (canny, hed, depth, and segmentation) from MultiGen-20M." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Query and Key Feature Choices. We investigate the effect of different query and key feature choices in Equation 5. First, we fix the query features as \\( Q_{\\mathrm{txt}} \\) and \\( Q_{\\mathrm{img}} \\) and examine the impact of different key features. As shown in Ta" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7968" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.11, + 0.094, + 0.482, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.345, + 0.483, + 0.402 + ], + "angle": 0, + "content": "Figure 4. Visualization results of our UNIC-Adapter on Dream-Bench for subject-driven generation. The first column displays the subject images, while the other three columns show the generated images based on different prompts." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.413, + 0.481, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.184, + 0.5, + 0.39, + 0.511 + ], + "angle": 0, + "content": "A cat with a mountain in the background." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.511, + 0.481, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.197, + 0.585, + 0.377, + 0.596 + ], + "angle": 0, + "content": "A girl with a city in the background." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.596, + 0.481, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.194, + 0.671, + 0.382, + 0.681 + ], + "angle": 0, + "content": "A robot in autumn mountain and lake." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.682, + 0.481, + 0.757 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.767, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Figure 5. Visualization results of our UNIC-Adapter on style-image-based T2I generation. The first row shows the reference style image, and each subsequent row contains images generated from the same prompt, influenced by different style images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "ble 3, \\( K_{\\mathrm{con}} \\) plays a critical role in pixel-level spatial control tasks and in maintaining subject-image similarity. Relying solely on \\( K_{\\mathrm{ist}} \\) as key feature (Exp. 3) significantly reduces performance across pixel-level control tasks and subject-" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.093, + 0.905, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.333, + 0.907, + 0.416 + ], + "angle": 0, + "content": "Figure 6. Comparison of visualization results of our UNIC-Adapter with and without position embedding, and IPAdapter-Instruct* on pixel-level control generation using canny edge conditional images. UNIC-Adapter generates images that are more closely aligned with the conditional images compared to the other two methods." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.428, + 0.904, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.567, + 0.907, + 0.623 + ], + "angle": 0, + "content": "Figure 7. Comparison of visualization results between our UNIC-Adapter and ControlNet* on subject-driven generation. Compared to ControlNet*, UNIC-Adapter achieves better consistency between subject images and generated images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.78 + ], + "angle": 0, + "content": "image similarity. We also investigate the impact of query features and find that using \\( Q_{\\mathrm{txt}} \\) alone as the query feature (Exp. 5) also degrades performance. In general, using \\( Q_{\\mathrm{img}} \\) as the query and both \\( K_{\\mathrm{con}} \\) and \\( K_{\\mathrm{ist}} \\) as key features (Exp. 4) yields optimal performance on pixel-level control tasks and comparable performance on subject-driven generation tasks. This highlights the importance of task instruction features in unified controllable generation. We adopt this configuration for subsequent experiments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Importance of Position Embedding. To enhance the control precision of pixel-level tasks, we experiment with incorporating different types of position embeddings, including the absolute position embedding in SD3 and the RoPE used in our adapter, to query and key features of each layer. As shown in Table 4, RoPE yields better performance than absolute embeddings. We further examine the impact of introducing a new linear layer, \\( L_{\\mathrm{cross}}^q \\), for query image" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7969" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.137, + 0.089, + 0.861, + 0.226 + ], + "angle": 0, + "content": "
Exp.NumberQueryQimgQtxtKeyKconKistCanny(F1 Score↑)HED(SSIM↑)Seg.(mIoU↑)Depth(RMSE↓)Subject-Driven
DINO↑CLIP-I↑CLIP-T↑
1✓✓✓✓29.980.784030.8434.050.7780.8270.309
2✓✓✓29.010.776729.0035.390.7690.8250.307
3✓✓✓22.380.559922.9638.820.6940.7800.319
4✓✓✓31.320.793430.9433.160.7690.8230.308
5✓✓✓19.100.373017.0646.950.6690.7820.317
6✓✓29.240.778627.8134.120.7500.8170.313
" + }, + { + "type": "table_caption", + "bbox": [ + 0.14, + 0.238, + 0.856, + 0.253 + ], + "angle": 0, + "content": "Table 3. Results of different query and key feature choices on pixel-level control tasks and subject-driven generation task." + }, + { + "type": "table", + "bbox": [ + 0.146, + 0.266, + 0.851, + 0.455 + ], + "angle": 0, + "content": "
MethodCanny(F1 Score↑)HED(SSIM↑)Seg.(mIoU↑)Depth(RMSE↓)Subject-Driven
DINO↑CLIP-I↑CLIP-T↑
w/o PE31.320.793430.9433.160.7690.8230.308
w/ Abs PE31.510.796730.3132.980.7400.8090.314
w/ Abs PE + Lqcross36.930.834833.2232.740.7510.8110.314
w/ RoPE34.650.824832.1333.450.7610.8170.313
w/ RoPE + Lqcross37.950.842033.3232.250.7840.8290.309
w/o txt36.310.838733.7131.600.7880.8300.308
w/o txt update36.310.838733.7131.600.7720.8210.310
ControlNet*38.180.838234.8131.460.7490.8150.313
IPAdapter-Instruct*15.810.219116.1247.810.7880.8520.294
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.466, + 0.904, + 0.494 + ], + "angle": 0, + "content": "Table 4. Results of two baseline methods and different module combinations of our UNIC-Adapter on pixel-level control tasks and subject-driven generation task." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.513, + 0.482, + 0.634 + ], + "angle": 0, + "content": "features in the main branch. Table 4 shows that the addition of \\( L_{\\mathrm{cross}}^q \\) improves results further, particularly in pixel-level control tasks and subject-image similarity. This indicates the importance of \\( L_{\\mathrm{cross}}^q \\) in enabling query features \\( Q_{\\mathrm{img}} \\) to retrieve specific features from conditional image and task instruction. Figure 6 shows that UNIC-Adapter without position embedding often generates images with incomplete or incorrect details from control signals." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.635, + 0.483, + 0.877 + ], + "angle": 0, + "content": "Comparison with Baseline Methods. We reproduce two baseline methods, ControlNet* and IPAdapter-Instruct*, in our framework. For ControlNet*, we replace the cross-attention mechanism in the image-instruction feature injection module with an addition operation (as did in ControlNet) while retaining other settings. For IPAdapter-Instruct*, we reproduce results by referring to the open-source code on our datasets. As shown in Table 4, ControlNet* achieves performance comparable to our final model on pixel-level control tasks but exhibits suboptimal subject-image similarity on the subject-driven generation task. The visualization results in Figure 7 show that ControlNet* often produces images with distorted subject content. In contrast, IPAdapter-Instruct* performs well on subject-driven generation tasks but poorly on pixel-level control tasks. Due to compressing conditional images into 16 tokens with" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.514, + 0.906, + 0.605 + ], + "angle": 0, + "content": "out precise spatial information, the generated images from IPAdapter-Instruct* in Figure 6 fail to accurately capture the pixel-level details of the conditioned images. In comparison, our UNIC-Adapter achieves consistently superior results across both task types, demonstrating its effectiveness and suitability for unified controllable image generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.617, + 0.634, + 0.634 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.795 + ], + "angle": 0, + "content": "In this paper, we present UNIC-Adapter, a unified framework for controllable T2I generation that integrates diverse conditional image inputs and task instructions within a single model. Built on the MM-DiT architecture, our approach leverages cross-attention mechanisms and RoPE to achieve precise pixel-level control, as well as high subject fidelity, across multiple generation tasks. Our extensive experiments on tasks such as pixel-level spatial control, subject-driven image generation, and style-image-based T2I generation demonstrate the effectiveness of the UNIC-Adapter." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.808, + 0.682, + 0.825 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.833, + 0.905, + 0.863 + ], + "angle": 0, + "content": "This research is supported by the NSFC Grants under the contracts No. 62325111 and No.U22B2011." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.398, + 0.902 + ], + "angle": 0, + "content": "\\(\\S\\) https://github.com/unity-research/IP-Adapter-Instruct" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7970" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.185 + ], + "angle": 0, + "content": "[1] Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.187, + 0.483, + 0.255 + ], + "angle": 0, + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. GPT-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.258, + 0.483, + 0.313 + ], + "angle": 0, + "content": "[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science, 2(3):8, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.315, + 0.483, + 0.37 + ], + "angle": 0, + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.373, + 0.483, + 0.442 + ], + "angle": 0, + "content": "[5] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.444, + 0.483, + 0.514 + ], + "angle": 0, + "content": "[6] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- \\(\\backslash\\) sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.516, + 0.483, + 0.583 + ], + "angle": 0, + "content": "[7] Junsong Chen, Yue Wu, Simian Luo, Enze Xie, Sayak Paul, Ping Luo, Hang Zhao, and Zhenguo Li. Pixart- \\(\\{\\backslash\\) delta\\} Fast and controllable image generation with latent consistency models. arXiv preprint arXiv:2401.05252, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.586, + 0.483, + 0.628 + ], + "angle": 0, + "content": "[8] Wenhu Chen, Hexiang Hu, Chitwan Sahara, and William W Cohen. Re-Imagen: Retrieval-augmented text-to-image generator. arXiv preprint arXiv:2209.14491, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.63, + 0.483, + 0.685 + ], + "angle": 0, + "content": "[9] Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Ruiz, Xuhui Jia, Ming-Wei Chang, and William W Cohen. Subject-driven text-to-image generation via apprenticeship learning. NeurIPS, 36, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.687, + 0.483, + 0.757 + ], + "angle": 0, + "content": "[10] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, pages 2818-2829, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.758, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[12] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Int. Conf. Mach. and Learn., 2024. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[13] Junyao Gao, Yanchen Liu, Yanan Sun, Yinhao Tang, Yanhong Zeng, Kai Chen, and Cairong Zhao. StyleShot: A snap" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "shot on any style. arXiv preprint arXiv:2407.01414, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[14] Zhen Han, Chaojie Mao, Zeyinzi Jiang, Yulin Pan, and Jingfeng Zhang. StyleBooth: Image style editing with multimodal instruction. arXiv preprint arXiv:2404.12154, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[15] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In CVPR, pages 4775–4785, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.258 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.26, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[17] Hexiang Hu, Kelvin CK Chan, Yu-Chuan Su, Wenhu Chen, Yandong Li, Kihyuk Sohn, Yang Zhao, Xue Ben, Boqing Gong, William Cohen, et al. Instruct-Imagen: Image generation with multi-modal instruction. In CVPR, pages 4754–4763, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.384 + ], + "angle": 0, + "content": "[18] Lianghua Huang, Di Chen, Yu Liu, Yujun Shen, Deli Zhao, and Jingren Zhou.Composer: Creative and controllable image synthesis with composable conditions. arXiv preprint arXiv:2302.09778, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.904, + 0.412 + ], + "angle": 0, + "content": "[19] Frank Hutter Ilya Loshchilov. Decoupled weight decay regularization. *ICLR*, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.797, + 0.428 + ], + "angle": 0, + "content": "[20] Imagen-Team-Google. Imagen 3, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.429, + 0.905, + 0.483 + ], + "angle": 0, + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[22] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 128(7):1956-1981, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.904, + 0.608 + ], + "angle": 0, + "content": "[23] Black Forest Labs. Flux: Inference repository. https://github.com/black-forest-labs/flux, 2024. Accessed: 2024-11-14. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.651 + ], + "angle": 0, + "content": "[24] Dongxu Li, Junnan Li, and Steven Hoi. BLIP-Diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. NeurIPS, 36, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.652, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Int. Conf. Mach. and Learn., pages 19730-19742. PMLR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[26] Ming Li, Taojiannan Yang, Huafeng Kuang, Jie Wu, Zhaoning Wang, Xuefeng Xiao, and Chen Chen. ControlNet++: Improving conditional controls with efficient consistency feedback. In ECCV, pages 129-147. Springer, 2025. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[27] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. GLIGEN: Open-set grounded text-to-image generation. In CVPR, pages 22511-22521, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[28] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et al. Hunyuan-DiT: A powerful multi-resolution diffusion transformer with fine-grained" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7971" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "chinese understanding. arXiv preprint arXiv:2405.08748, 2024.1,2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[29] Bingchen Liu, Ehsan Akhgari, Alexander Visheratin, Aleks Kamko, Linmiao Xu, Shivam Shrirao, Joao Souza, Suhail Doshi, and Daiqing Li. Playground v3: Improving text-to-image alignment with deep-fusion large language models. arXiv preprint arXiv:2409.10695, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[30] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding DINO: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[31] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-Diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. In ACM SIGGRAPH Conference Papers, pages 1-12, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[32] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, pages 6038–6047, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.433 + ], + "angle": 0, + "content": "[33] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, pages 4296-4304, 2024. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.49 + ], + "angle": 0, + "content": "[34] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-G: Generating images in context with multimodal large language models. arXiv preprint arXiv:2310.02992, 2023. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.518 + ], + "angle": 0, + "content": "[35] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.548 + ], + "angle": 0, + "content": "[36] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.549, + 0.482, + 0.605 + ], + "angle": 0, + "content": "[37] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.606, + 0.482, + 0.675 + ], + "angle": 0, + "content": "[38] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.677, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[39] Can Qin, Shu Zhang, Ning Yu, Yihao Feng, Xinyi Yang, Yingbo Zhou, Huan Wang, Juan Carlos Niebles, Caiming Xiong, Silvio Savarese, et al. UniControl: A unified diffusion model for controllable visual generation in the wild. arXiv preprint arXiv:2305.11147, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Int. Conf. Mach. and Learn., pages 8748-8763. PMLR, 2021. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[41] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[43] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 234–241. Springer, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.286 + ], + "angle": 0, + "content": "[44] Ciara Rowles, Shimon Vainer, Dante De Nigris, Slava Elizarov, Konstantin Kutsy, and Simon Donne. IPAdapter-Instruct: Resolving ambiguity in image-based conditioning using instruct prompts. arXiv preprint arXiv:2408.03209, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[45] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500-22510, 2023. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.425 + ], + "angle": 0, + "content": "[46] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 35:36479-36494, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.427, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[47] Babak Saleh and Ahmed Elgammal. Large-scale classification of fine-art paintings: Learning the right metric on the right feature. arXiv preprint arXiv:1505.00855, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[48] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. LAION-5B: An open large-scale dataset for training next generation image-text models. NeurIPS, 35:25278-25294, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.552, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[49] Kunpeng Song, Yizhe Zhu, Bingchen Liu, Qing Yan, Ahmed Elgammal, and Xiao Yang. Moma: Multimodal llm adapter for fast personalized image generation. arXiv preprint arXiv:2404.05674, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.608, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[50] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. RoFormer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.664, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[51] Kcolors Team. Kcolors: Effective training of diffusion model for photorealistic text-to-image synthesis. arXiv preprint, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.705, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[52] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.895, + 0.789 + ], + "angle": 0, + "content": "[53] A Vaswani. Attention is all you need. NeurIPS, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[54] Haofan Wang, Matteo Spinelli, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[55] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 1, 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7972" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[56] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. arXiv preprint arXiv:2409.11340, 2024. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.217 + ], + "angle": 0, + "content": "[57] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. CogVideoX: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[58] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. IP-Adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[59] Zilyu Ye, Jinxiu Liu, Ruotian Peng, Jinjin Cao, Zhiyang Chen, Yiyang Zhang, Ziwei Xuan, Mingyuan Zhou, Xiaogian Shen, Mohamed Elhoseiny, et al. Openstory++: A large-scale dataset and benchmark for instance-aware open-domain visual storytelling. arXiv preprint arXiv:2408.03695, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[60] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.404, + 0.482, + 0.457 + ], + "angle": 0, + "content": "[61] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. UniControlNet: All-in-one control to text-to-image diffusion models. NeurIPS, 36, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.46, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[62] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In CVPR, pages 22490-22499, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.482, + 0.557 + ], + "angle": 0, + "content": "[63] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, pages 633-641, 2017. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "7973" + } + ] +] \ No newline at end of file diff --git a/2025/UNIC-Adapter_ Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation/409eb6c5-d571-4968-a329-2c1ebf51e321_origin.pdf b/2025/UNIC-Adapter_ Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation/409eb6c5-d571-4968-a329-2c1ebf51e321_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..155f7a8876f051477105f2694f2d54acc952f725 --- /dev/null +++ b/2025/UNIC-Adapter_ Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation/409eb6c5-d571-4968-a329-2c1ebf51e321_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7cdb1f4c87a3fcacbe5e1c3140a03f79730de60915655f79d8e61d2e47aecce8 +size 8264905 diff --git a/2025/UNIC-Adapter_ Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation/full.md b/2025/UNIC-Adapter_ Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..35456cc5d154ec6619c4303c9cea0dbd77416910 --- /dev/null +++ b/2025/UNIC-Adapter_ Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation/full.md @@ -0,0 +1,359 @@ +# UNIC-Adapter: Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation + +Lunhao Duan\*1 Shanshan Zhao\*2 Wenjun Yan2 Yinglun Li2 Qing-Guo Chen2 Zhao Xu2 Weihua Luo2 Kaifu Zhang2 Mingming Gong3 Gui-Song Xia\*14 + +$^{1}$ School of Computer Science, Wuhan University $^{2}$ Alibaba International Digital Commerce $^{3}$ University of Melbourne $^{4}$ School of Artificial Intelligence, Wuhan University + +# Abstract + +Recently, text-to-image generation models have achieved remarkable advancements, particularly with diffusion models facilitating high-quality image synthesis from textual descriptions. However, these models often struggle with achieving precise control over pixel-level layouts, object appearances, and global styles when using text prompts alone. To mitigate this issue, previous works introduce conditional images as auxiliary inputs for image generation, enhancing control but typically necessitating specialized models tailored to different types of reference inputs. In this paper, we explore a new approach to unify controllable generation within a single framework. Specifically, we propose the unified image-instruction adapter (UNIC-Adapter) built on the Multi-Modal-Diffusion Transformer architecture, to enable flexible and controllable generation across diverse conditions without the need for multiple specialized models. Our UNIC-Adapter effectively extracts multi-modal instruction information by incorporating both conditional images and task instructions, injecting this information into the image generation process through a cross-attention mechanism enhanced by Rotary Position Embedding. Experimental results across a variety of tasks, including pixel-level spatial control, subject-driven image generation, and style-image-based image synthesis, demonstrate the effectiveness of our UNIC-Adapter in unified controllable image generation. + +# 1. Introduction + +Recent advancements in text-to-image (T2I) generation models [3, 5, 12, 20, 23, 29, 38, 42, 51], primarily driven by diffusion models [16], have been witnessed in recent years. By leveraging large-scale paired text-image data [48], nu + +(a) + +A delicious afternoon tea. + +A room with the forest backdrop. + +A robot in autumn mountain and lake + +··· + +A dog with a mountain in the + +(b) + +Generate image from canny edge. + +Render image with this depth map. + +Create image using style influence. + +1 + +Generate image from dog image + +(c) + +![](images/e13f90acba9ec6aacfb88fabb46d30e3470f85c4d61206fc2e5996f349e45db0.jpg) + +![](images/67f8af2831537e9bc2cdbf08c4e98da41798927a67bf4f09deb17debb8353b64.jpg) + +![](images/ec71007de18d7c771c97f474b57a70353f275ce361265f88ab2635cded58dbea.jpg) + +![](images/bf9742303b6ada9ef0af02f9f54e4fb61fe888b4bbcae808cc4df645c4135378.jpg) + +(d) + +![](images/56439bdfacd22dfd3e04f3aec4996bd6946b20e5b170326e6ad5723be1ef36f7.jpg) +Figure 1. With the UNIC-Adapter, SD3 enables flexible and controllable generation across multiple reference modalities within a single model. (a) to (d) represent text prompts, task instructions, conditional images, and generated images, respectively. + +![](images/0a25eff6fd53b9af61b0eb36efac739d6727206d4c74744092410610e3d0cd43.jpg) + +![](images/7c064a96448ffe8b22842ad46f2a0b7af58d5d197ab88015455f1cb5e31ae8ee.jpg) + +![](images/2408286dc7a0cf4f9c10d82be303c4924ffee240e3f6554ba77e1632434dd055.jpg) + +merous open-source T2I models, like Stable Diffusion 1.5 (SD1.5) [42], Stable Diffusion XL (SDXL) [38], Pixart-series [5-7], Stable Diffusion 3 (SD3) [12], Hunyuan-DiT [28], and FLUX.1-dev [23] have significantly enhanced the fidelity of images generated from natural language descriptions. These models utilize a range of architectures, spanning from U-Net-based designs [43] to the more recent Diffusion-Transformer-based frameworks [35]. + +Despite these advancements, relying solely on text prompts often falls short in defining pixel-level spatial structures and geometric details, exact object appearances, and global styles of a specific image. To address this limitation, recent research [13, 18, 33, 34, 45, 55, 58, 60] has incorporated conditional images as additional inputs alongside text prompts. These conditional images offer finer control over pixel-level [60, 62], object-level [34, 55, 58], or style-specific [13, 15, 54] features in the generated outputs. For example, ControlNet [60] enables pixel-level spatial control by integrating structural inputs, such as depth and edge maps, through a parallel encoder tailored for + +each type of control. Similarly, IP-Adapter [58] enables content-specific control by conditioning on a reference image and injecting its CLIP-based [40] embeddings via crossattention layers. However, these methods often necessitate individual model training for each control type, resulting in increased training costs and complicating the integration of different control inputs. IPAdapter-Instruct [44] extends IP-Adapter by incorporating instruction prompts to specify task types for different conditional images, enabling multi-task handling within a single model. UniControl [39] tackles various pixel-level control tasks within a unified model by introducing task-specific adapters to extract features for different visual conditions. Meanwhile, Uni-ControlNet [61] divides conditional images into two groups: local (pixel-level) and global (image-level) conditions, and designs dedicated adapters for each group. However, this approach is not fully unified, as the two adapters are trained separately. + +To further unify diverse reference inputs, including text, edge maps, content images, etc., Instruct-Imagen [17], based onImagen [46], represents them as multi-modal instructions within a single framework. OmniGen [56] first tokenizes interleaved text and images into a one-dimensional sequence and then uses a transformer model (Phi3) [1] to achieve unified image generation, such as image editing and subject-driven generation. To adapt the model for multi-modal instructions, these two methods involve multi-stage training of the entire model. To enable more efficient training for unified controllable image generation, this paper seeks to design an adapter based on a pre-trained T2I model, avoiding the need to train the entire model. + +Given that DiT-based architectures [36] equipped with pure transformer blocks have become the mainstream in T2I models [5, 12, 23, 28], we choose SD3, an open-source state-of-the-art (SOTA) model, as the foundation model for studying the adapter. To achieve unified controllable image generation across various conditions, we employ the Multi-Modal-Diffusion Transformer (MM-DiT) structure [12, 23], which enables full cross-attention between text and image features to extract multi-modal instruction information. By treating text as an independent modality, the MM-DiT allocates substantial model capacity to capture and understand text-based information, thereby improving its interpretation of text prompts [12]. Leveraging the strengths of the MM-DiT, we propose a unified image-instruction adapter (UNIC-Adapter) to facilitate image generation conditioned on various types of reference images with a task instruction indicating the desired generation target. Specifically, task instruction and conditional image features are extracted by text encoders [40, 41] and a Variational Autoencoder (VAE) [12], respectively. Then, the adapter processes instruction and image features using MM-DiT blocks, which are subsequently integrated into the + +main generation branch via a cross-attention mechanism. To further enhance spatial awareness, which is essential for pixel-level control, we incorporate Rotary Position Embedding (RoPE) [50] into the query and key features within the cross-attention. With the proposed UNIC-Adapter, SD3 can generate images under a wide array of conditions, as shown in Figure 1. + +In summary, our contributions are threefold: + +- We are the first to leverage the MM-DiT architecture for unified controllable generation across various types of reference images in T2I models. +- We introduce an effective feature injection method based on cross-attention, enhanced with RoPE encoding for better spatial control. +- Equipped with our UNIC-Adapter, SD3 enables unified controllable generation across 14 types of conditional images in our experiments. + +# 2. Related Work + +# 2.1. Text-to-Image Generation Models + +Mainstream T2I generation models [5, 12, 38, 42] are primarily based on diffusion models [16], which generate images by learning to reverse a progressive noise-adding process. In these models, text input typically serves as a conditional signal, guiding the generation process to align with the provided description. The success of transformer architectures [53], particularly in large language models [2, 52], has significantly influenced T2I model architectures, evolving them from convolution-transformer hybrids [38, 42] to fully transformer-based designs [5-7, 12, 23, 28]. Recently, the MM-DiT architecture [12, 23] has advanced this evolution by treating text and image features as distinct modalities with full-attention mechanisms, enabling each modality to attend to the other. This approach outperforms traditional cross-attention setups, where text features are used only as keys and values [12]. However, relying solely on text-based conditioning presents limitations, particularly for images that require detailed spatial layouts or specific object appearances. Our work builds upon these transformer-based diffusion models by integrating additional conditional images and task instructions, thereby enhancing the flexibility and controllability of T2I generation. + +# 2.2. Conditional Image-based T2I Models + +Conditional image-based T2I models introduce auxiliary image inputs to complement textual prompts, enabling finer control over aspects such as spatial layout [26, 60, 62], object content [9, 24, 31, 49, 58], and stylistic elements [13, 14, 54]. For pixel-level spatial control, ControlNet [60] employs a duplicated U-Net encoder to process specific types of conditional images (e.g., edge, pose, and depth maps) and integrates the features into the main T2I model + +![](images/86f0ef7ac179a0d0bcc1d0ab7b46c475a1cb21fcaca5ce2d236ff9e08f372406.jpg) +Figure 2. The overall architecture of our proposed UNIC-Adapter. The task instruction and conditional image features are progressively attending to each other through a series of $N$ adapter blocks. In each adapter block, the image features $Z_{\mathrm{img}}$ from MM-DiT block in the main image generation branch serve as the query, while both task instruction features $Z_{\mathrm{ist}}$ and conditional image features $Z_{\mathrm{con}}$ function as keys and values. For simplicity, normalization layers and feed-forward networks are omitted in this figure. + +through residual connections in the U-Net decoder. Another key work, IP-Adapter [58], uses the Vision Transformer (ViT) [11] encoder from CLIP [40] to extract features from conditional images. These features are then incorporated into the U-Net backbone via cross-attention layers, achieving controllable generation with contents from conditional images. IPAdapter-Instruct [44] extends this by introducing an instructional prompt that guides the interpretation of conditional images. To minimize the necessity for individual model training for each type of conditional image in ControlNet, UniControl [39] unifies nine pixel-level control tasks within a single model, leveraging task-specific feature extraction adapters and a task-aware HyperNet to modulate features based on task instructions. To achieve control with image content as well as spatial layout, UniControlNet [61] incorporates two dedicated adapters for local (pixel-level control) and global (CLIP-image embedding control) conditions, respectively. Unlike these methods, which freeze the base model and only training the newly added modules / adapters, Instruct-Imagen [17] proposes using multi-modal instructions and optimize the entire model, including the base model (Imagen [46]) and the new cross-attention layers, for image generation under different conditions. A recent work, OmniGen [56], develops a unified image generation model for various multi-modal inputs using a transformer-based architecture [1]. In this paper, we aim to enhance the generation capabilities of an existing T2I model for diverse conditions by introducing a unified image-instruction adapter, named UNIC-Adapter, without the need to train the entire model. + +# 3. Method + +In this section, we first provide a brief overview of the DiT and MM-DiT employed in SOTA T2I models. Then, we describe our UNIC-Adapter, which incorporates task instructions and conditional images to enhance the model's capability to interpret various user-specified conditions. + +# 3.1. Preliminary + +Diffusion Transformer Block. In DiT Block-based T2I models [5], conditional text features $Z_{\mathrm{txt}}$ extracted from text encoders [40, 41] are integrated into DiT Blocks via cross-attention layers. Specifically, in each DiT Block, the conditional text features are first mapped to key $K_{\mathrm{txt}}$ and value $V_{\mathrm{txt}}$ features through linear transformations $L$ , while the image feature $Z_{\mathrm{img}}$ is mapped to a query feature $Q_{\mathrm{img}}$ . An attention mechanism then updates the image features $Z_{\mathrm{img}}$ by attending to $K_{\mathrm{txt}}$ and $V_{\mathrm{txt}}$ : + +$$ +K _ {\mathrm {t x t}} = L _ {\mathrm {t x t}} ^ {k} \left(Z _ {\mathrm {t x t}}\right), V _ {\mathrm {t x t}} = L _ {\mathrm {t x t}} ^ {v} \left(Z _ {\mathrm {t x t}}\right), Q _ {\mathrm {i m g}} = L _ {\mathrm {i m g}} ^ {q} \left(Z _ {\mathrm {i m g}}\right), \tag {1} +$$ + +$$ +Z _ {\mathrm {i m g}} ^ {\prime} = \operatorname {A t t n} \left(Q _ {\mathrm {i m g}}, K _ {\mathrm {t x t}}, V _ {\mathrm {t x t}}\right), +$$ + +where Attn indicates the multi-head attention [53]. In this setup, the same $Z_{\mathrm{txt}}$ features are provided to DiT Blocks across different layers as static conditioning signals, which may limit the utilization of text features. + +Multi-modal Diffusion Transformer Block. Recent T2I models [12, 23] based on MM-DiT Blocks enhance DiT Blocks by treating conditional text as an independent modality capable of dynamically interacting with image features across layers. In MM-DiT, both text and image features serve as queries, keys, and values, enabling cross-modal attention in both directions through a full-attention + +mechanism. Specifically, the MM-DiT layer performs the following operations: + +$$ +Q _ {\mathrm {t x t}} = L _ {\mathrm {t x t}} ^ {q} \left(Z _ {\mathrm {t x t}}\right), K _ {\mathrm {t x t}} = L _ {\mathrm {t x t}} ^ {k} \left(Z _ {\mathrm {t x t}}\right), V _ {\mathrm {t x t}} = L _ {\mathrm {t x t}} ^ {v} \left(Z _ {\mathrm {t x t}}\right), \tag {2} +$$ + +$$ +Q _ {\mathrm {i m g}} = L _ {\mathrm {i m g}} ^ {q} \left(Z _ {\mathrm {i m g}}\right), K _ {\mathrm {i m g}} = L _ {\mathrm {i m g}} ^ {k} \left(Z _ {\mathrm {i m g}}\right), V _ {\mathrm {i m g}} = L _ {\mathrm {i m g}} ^ {v} \left(Z _ {\mathrm {i m g}}\right). +$$ + +These queries, keys, and values are processed in a full-attention layer, where each modality attends to both text and image features: + +$$ +Z _ {\mathrm {t x t}} ^ {\prime} = \operatorname {A t t n} \left(Q _ {\mathrm {t x t}}, \left[ K _ {\mathrm {t x t}} \| K _ {\mathrm {i m g}} \right], \left[ V _ {\mathrm {t x t}} \| V _ {\mathrm {i m g}} \right]\right), \tag {3} +$$ + +$$ +Z _ {\mathrm {i m g}} ^ {\prime} = \operatorname {A t t n} \left(Q _ {\mathrm {i m g}}, \left[ K _ {\mathrm {t x t}} \right\| K _ {\mathrm {i m g}} ], \left[ V _ {\mathrm {t x t}} \right\| V _ {\mathrm {i m g}}\right), +$$ + +where $\parallel$ denotes concatenation. This enhanced interaction facilitates richer cross-modal feature extraction, as text and image features are iteratively refined across layers, improving text-guided image generation [12]. + +# 3.2. UNIC-Adapter + +As shown in Figure 2, the UNIC-Adapter introduces two additional inputs: task instructions and conditional images. This integration enables unified controllable image generation by allowing the model to interpret and apply specific features from a variety of conditional images, guided by the task instruction. Below, we provide a detailed description of its structure. + +Image-instruction Feature Extraction. To enable a comprehensive understanding of task instructions and conditional images, we leverage MM-DiT blocks to enhance the interaction between these inputs. Specifically, we adopt SD3 as our base model. Following the setup in SD3, task instruction features, denoted as $Z_{\mathrm{ist}}$ , are extracted using text encoders such as CLIP-G/14 [10], CLIP-L/15 [40], and T5 XXL [41]. Simultaneously, conditional image features, $Z_{\mathrm{con}}$ , which represent one of several condition types (e.g., edge maps, depth maps, subject content images, or style-reference images), are obtained through the VAE. Within the MM-DiT blocks, task instruction and conditional image features undergo mutual attention updates, defined as follows: + +$$ +Q _ {\text {i s t}} = L _ {\text {i s t}} ^ {q} (Z _ {\text {i s t}}), K _ {\text {i s t}} = L _ {\text {i s t}} ^ {k} (Z _ {\text {i s t}}), V _ {\text {i s t}} = L _ {\text {i s t}} ^ {v} (Z _ {\text {i s t}}), +$$ + +$$ +Q _ {\text {c o n}} = L _ {\text {c o n}} ^ {q} \left(Z _ {\text {c o n}}\right), K _ {\text {c o n}} = L _ {\text {c o n}} ^ {k} \left(Z _ {\text {c o n}}\right), V _ {\text {c o n}} = L _ {\text {c o n}} ^ {v} \left(Z _ {\text {c o n}}\right), \tag {4} +$$ + +$$ +Z _ {\mathrm {i s t}} ^ {\prime} = \operatorname {A t t n} \left(Q _ {\mathrm {i s t}}, \left[ K _ {\mathrm {i s t}} \| K _ {\mathrm {c o n}} \right], \left[ V _ {\mathrm {i s t}} \| V _ {\mathrm {c o n}} \right]\right), +$$ + +$$ +Z _ {\text {c o n}} ^ {\prime} = \operatorname {A t t n} \left(Q _ {\text {c o n}}, \left[ K _ {\text {i s t}} \| K _ {\text {c o n}} \right], \left[ V _ {\text {i s t}} \| V _ {\text {c o n}} \right]\right). +$$ + +These updates ensure that the task instruction $Z_{\mathrm{ist}}$ and conditional image $Z_{\mathrm{con}}$ dynamically refine each other's features by selectively attending to relevant information. + +Image-instruction Feature Injection. To improve the model's responsiveness to diverse user-defined conditions, the key and value features from both instruction and conditional image inputs are integrated simultaneously into the main image generation branch through cross-attention [17, 55, 58]. This mechanism ensures that the feature injection is + +directed by the specific task instruction. Experiments indicate that injecting either conditional image features or task instruction features alone through cross-attention results in suboptimal performance, as demonstrated in the ablation studies of Section 4. By contrast, the simultaneous incorporation of both feature types optimally supports unified controlled image generation. Accordingly, the image features $Z_{\mathrm{img}}^{\prime}$ from Equation 3 are further updated as follows: + +$$ +Q _ {\mathrm {i m g}} ^ {\prime} = L _ {\text {c r o s s}} ^ {q} \left(Z _ {\mathrm {i m g}}\right), \tag {5} +$$ + +$$ +Z _ {\mathrm {i m g}} ^ {\prime \prime} = Z _ {\mathrm {i m g}} ^ {\prime} + \operatorname {A t t n} \left(Q _ {\mathrm {i m g}} ^ {\prime}, \left[ K _ {\mathrm {i s t}} \| K _ {\mathrm {c o n}} \right], \left[ V _ {\mathrm {i s t}} \| V _ {\mathrm {c o n}} \right]\right). +$$ + +In this setup, we introduce a new linear layer, $L_{\mathrm{cross}}^q$ , designed to enhance the model's ability to attend to image-instruction features effectively. This additional layer has been shown to improve performance, as validated in our experiments. Overall, this cross-attention operation facilitates controlled image generation, leveraging desired attributes from both the conditional image and the task instruction. + +Position Embedding. The positional information of conditional image features is crucial for pixel-level control tasks, as methods like ControlNet incorporate features through pixel-to-pixel addition. To improve pixel-level control precision, we integrate Rotary Position Embedding (RoPE) [50] into both the query and key features prior to cross-attention. RoPE provides relative positional encoding, ensuring that query and key features with closer pixel coordinates yield higher similarity scores, thereby improving spatial alignment in the generated output. It has been used in T2I [23] and text-to-video [57] models. + +Specifically, we apply 1-D RoPE separately to the height and width dimensions. Given a feature vector $f \in \mathbb{R}^{|D|}$ at position $(h,w)$ , we split $f$ into two halves: $f_h \in \mathbb{R}^{|D|/2}$ for the height dimension and $f_w \in \mathbb{R}^{|D|/2}$ for the width. For the height component $f_h$ , the rotary matrix $R_h$ is defined as: + +$$ +R _ {h} = \left[ \begin{array}{c c c c c} \cos (h \theta_ {0}) & - \sin (h \theta_ {0}) & \dots & 0 & 0 \\ \sin (h \theta_ {0}) & \cos (h \theta_ {0}) & \dots & 0 & 0 \\ \vdots & \vdots & \ddots & \vdots & \vdots \\ 0 & 0 & \dots & \cos (h \theta_ {l}) & - \sin (h \theta_ {l}) \\ 0 & 0 & \dots & \sin (h \theta_ {l}) & \cos (h \theta_ {l}) \end{array} \right], \tag {6} +$$ + +where $\theta_{i} = b^{-i / |D|}$ , with $b = 10000$ and $l = |D| / 4 - 1$ . The rotary matrix $R_{w}$ for the width component $f_{w}$ is defined analogously. The updated feature vector $f$ is then computed as: + +$$ +f = \left[ R _ {h} f _ {h} \| R _ {w} f _ {w} \right]. \tag {7} +$$ + +RoPE is applied to both query and key features in Equation 5 and Equation 4, with the position coordinates of $K_{\mathrm{ins}}$ set to $(0,0)$ . By embedding relative positional information in this way, RoPE enhances the model's precision in pixel-level control, while maintaining robustness and adaptability across various tasks. + +# 4. Experiments + +In this section, we conduct extensive experiments to evaluate the performance of our UNIC-Adapter in various conditional image generation tasks. Detailed ablation studies are also conducted to analyze the contribution of individual components within our framework. Additional experimental results are available in the Appendix. + +# 4.1. Tasks and Datasets + +Our evaluation covers three major tasks: pixel-level spatial control, subject-driven image generation, and style-image-based T2I generation. Below, we briefly describe the datasets used for training each task. + +Pixel-level Spatial Control. For pixel-level spatial control, we employ the MultiGen-20M dataset [39], which includes 12 types of pixel-level control annotations, such as Canny, HED, Sketch, Depth, Normal, Skeleton, Bbox, Segmentation, Outpainting, Inpainting, Deblurring, and Colorization. The dataset contains 2.8 million images with 20 million control annotations, and we reserve 1,000 samples for qualitative evaluation. For segmentation map conditioning, we also use the ADE20k [63] training set, which includes approximately 20,000 samples. These annotations serve as conditional images, and we use GPT-4o [2] to generate 20 synonymous instructions for each control type, such as "Generate an image from this edge map." + +Subject-driven Image Generation. For subject-driven image generation, we use three datasets: OpenImages v7 [22], GRIT [37], and OpenStory++ [59]. Following previous works [31, 34], we utilize Blip-2 [25] to generate captions for images without annotations and use spaCy‡ to extract subject tags. GroundingDINO [30] is used to obtain bounding boxes for each subject, and SAM [21] is employed to generate subject segmentation maps based on these bounding boxes. After processing and filtering, we obtain 2.1 million subject-image pairs. We also use GPT4o to generate 20 instructions, such as "Generate an image from this dog image." + +Style-image-based Image Generation. For style-image-based generation, we use the WikiArt [47] and Style-Booth [14] datasets. For StyleBooth, we randomly select an image from the same style subset as the conditional image, and for WikiArt, we employ CLIP-I [40] similarity to retrieve a conditional image for each training image. These two datasets comprise approximately 90,000 images. Additionally, we use GPT-4o to generate 20 instructions, such as "Generate an image based on this style image." + +# 4.2. Implementation Details + +Our UNIC-Adapter framework is based on the SD3 medium [12] model. Since instruction texts and conditional + +
MethodBase ModelCanny (F1 Score↑)HED (SSIM↑)Seg. (mIoU↑)Depth (RMSE↓)
Single-task
T2I-Adapter [33]SD1.423.65-12.6148.40
Gligen [27]SD1.426.940.563423.7838.83
ControlNet [60]SD1.534.650.762132.5535.90
ControlNet++ [26]SD1.537.040.809743.6428.32
Multi-task
UniControl [39]SD1.530.820.796925.4439.18
Uni-ControlNet [61]SD1.527.320.691019.3940.65
OmniGen [56]Phi-335.540.823744.2328.54
OursSD338.940.836942.8931.10
+ +Table 1. Results of different methods on four pixel-level control image generation tasks. $\uparrow$ indicates that higher values are better, while $\downarrow$ indicates that lower values are better. The best results for each metric are bolded, and the second-best results are underlined. + +
MethodBase ModelDINO↑CLIP-I↑CLIP-T↑
Fine-Tuning
Textual Inversion [32]SD1.50.5690.7800.255
DreamBooth [45]SD1.50.6680.8030.305
BLIP-Diffusion [24]SD1.50.6700.8050.302
Test Time Tuning Free
Re-Imagen [8]-0.6000.7400.270
SuTI [9]Imagen0.7410.8190.304
Kosmos-G [34]SD1.50.6940.8470.287
OmniGen [56]Phi-30.8010.8470.301
OursSD30.8160.8410.306
+ +Table 2. Results of different methods on the DreamBench dataset [45] for subject-driven image generation. + +are variations of text and image modalities, the UNIC-Adapter is initialized using the base T2I model's parameters, which minimizes learning difficulty. We freeze the parameters of base SD3 medium model and the feed-forward layer after the attention layer in adapter blocks, and train all other newly introduced parameters, resulting in approximately 1.2 billion trainable parameters. We use the AdamW optimizer [19] with a learning rate of 0.0001 and a weight decay of 0.01. Training is performed on 16 H100 GPUs for 100,000 steps, with a batch size of 16 per GPU. + +# 4.3. Main Results + +We present experimental results of the proposed UNIC-Adapter framework across three major tasks. + +Pixel-Level Spatial Control. Following previous work [26, 56], we evaluate performance based on the similarity between control conditions extracted from generated images and the specified pixel-level control inputs. The evaluation is conducted on the test set of MultiGen-20M with canny edge, hed edge, and depth map conditions and on ADE20k test set with segmentation map condition. As + +![](images/44180588d9e4f998bfb3816d16d28ac5f55735739f9c5ae3e1cf0f5058cb02d6.jpg) +Figure 3. Visualization results of our UNIC-Adapter on twelve pixel-level control tasks from the MultiGen-20M dataset. The first and third rows show different types of conditional images, while the second and fourth rows display the corresponding generated images. + +shown in Table 1, UNIC-Adapter achieves the best performance on canny and hed edge map conditions, and demonstrates comparable performance on depth and segmentation map conditions, compared to SOTA methods like ControlNet++ [26] for single-task and OmniGen [56] for multitask scenarios. Figure 3 shows qualitative results on 12 tasks from MultiGen-20M, illustrating that UNIC-Adapter consistently produces images accurately aligned with pixel-level controls. + +Subject-Driven Image Generation. Evaluation is conducted using 30 objects and 25 prompts from the Dream-Bench dataset [45]. Following prior work [34], we generate 4 images for each prompt and measure subject fidelity using DINO [4] and CLIP-I scores and text fidelity with CLIP-T [40]. As shown in Table 2, UNIC-Adapter achieves high subject-image consistency and maintains strong alignment between the generated images and text prompts. Qualitative examples in Figure 4 demonstrate that UNIC-Adapter effectively generates images closely resembling input sub + +jects across various contexts. + +Style-image-based Image Generation. For the same text prompt, style images from StyleBooth [14] are randomly selected as conditional images. The qualitative results in Figure 5 indicate that UNIC-Adapter effectively captures and applies stylistic elements from conditional images, generating artistically consistent outputs. + +# 4.4. Ablation Studies + +To better understand the impact of individual components in the UNIC-Adapter framework, we conduct ablation studies by systematically modifying or removing key modules. For efficiency, the models in this part are trained on the subject-driven image generation task and four tasks (canny, hed, depth, and segmentation) from MultiGen-20M. + +Query and Key Feature Choices. We investigate the effect of different query and key feature choices in Equation 5. First, we fix the query features as $Q_{\mathrm{txt}}$ and $Q_{\mathrm{img}}$ and examine the impact of different key features. As shown in Ta + +![](images/56feff493f71225acad6e603a90f25da3628d8a72f7d5d6a5bea088652d213c2.jpg) +Figure 4. Visualization results of our UNIC-Adapter on Dream-Bench for subject-driven generation. The first column displays the subject images, while the other three columns show the generated images based on different prompts. + +![](images/30155aee3240742cf5fdeb2e2da5e74e5152b3d2eb5e1a215cedd9d4587321c8.jpg) +A cat with a mountain in the background. + +![](images/381c90fb07ae86a9c3e66ea29d6a429b886e232bc703740197e620caf983f3eb.jpg) + +![](images/0aa83a366708b9203717c54e7407acf0f9420c26e7f232f450bd03049deb9e65.jpg) +A girl with a city in the background. +A robot in autumn mountain and lake. + +![](images/0835dd966396b4cefedc048427e38b9ae258f4674213cb2ba47183798da879e6.jpg) +Figure 5. Visualization results of our UNIC-Adapter on style-image-based T2I generation. The first row shows the reference style image, and each subsequent row contains images generated from the same prompt, influenced by different style images. + +ble 3, $K_{\mathrm{con}}$ plays a critical role in pixel-level spatial control tasks and in maintaining subject-image similarity. Relying solely on $K_{\mathrm{ist}}$ as key feature (Exp. 3) significantly reduces performance across pixel-level control tasks and subject- + +![](images/fd90d90f86770d9a0b92957506e7569ded4c1bd5a96777057644dec4b6b99833.jpg) +Figure 6. Comparison of visualization results of our UNIC-Adapter with and without position embedding, and IPAdapter-Instruct* on pixel-level control generation using canny edge conditional images. UNIC-Adapter generates images that are more closely aligned with the conditional images compared to the other two methods. + +![](images/fbe71ec62ee4ae7c880b20f5ad85aa49b79d7893a7f4f02f4cf4a69c6a2b0021.jpg) +Figure 7. Comparison of visualization results between our UNIC-Adapter and ControlNet* on subject-driven generation. Compared to ControlNet*, UNIC-Adapter achieves better consistency between subject images and generated images. + +image similarity. We also investigate the impact of query features and find that using $Q_{\mathrm{txt}}$ alone as the query feature (Exp. 5) also degrades performance. In general, using $Q_{\mathrm{img}}$ as the query and both $K_{\mathrm{con}}$ and $K_{\mathrm{ist}}$ as key features (Exp. 4) yields optimal performance on pixel-level control tasks and comparable performance on subject-driven generation tasks. This highlights the importance of task instruction features in unified controllable generation. We adopt this configuration for subsequent experiments. + +Importance of Position Embedding. To enhance the control precision of pixel-level tasks, we experiment with incorporating different types of position embeddings, including the absolute position embedding in SD3 and the RoPE used in our adapter, to query and key features of each layer. As shown in Table 4, RoPE yields better performance than absolute embeddings. We further examine the impact of introducing a new linear layer, $L_{\mathrm{cross}}^q$ , for query image + +
Exp.NumberQueryQimgQtxtKeyKconKistCanny(F1 Score↑)HED(SSIM↑)Seg.(mIoU↑)Depth(RMSE↓)Subject-Driven
DINO↑CLIP-I↑CLIP-T↑
1✓✓✓✓29.980.784030.8434.050.7780.8270.309
2✓✓✓29.010.776729.0035.390.7690.8250.307
3✓✓✓22.380.559922.9638.820.6940.7800.319
4✓✓✓31.320.793430.9433.160.7690.8230.308
5✓✓✓19.100.373017.0646.950.6690.7820.317
6✓✓29.240.778627.8134.120.7500.8170.313
+ +Table 3. Results of different query and key feature choices on pixel-level control tasks and subject-driven generation task. + +
MethodCanny(F1 Score↑)HED(SSIM↑)Seg.(mIoU↑)Depth(RMSE↓)Subject-Driven
DINO↑CLIP-I↑CLIP-T↑
w/o PE31.320.793430.9433.160.7690.8230.308
w/ Abs PE31.510.796730.3132.980.7400.8090.314
w/ Abs PE + Lqcross36.930.834833.2232.740.7510.8110.314
w/ RoPE34.650.824832.1333.450.7610.8170.313
w/ RoPE + Lqcross37.950.842033.3232.250.7840.8290.309
w/o txt36.310.838733.7131.600.7880.8300.308
w/o txt update36.310.838733.7131.600.7720.8210.310
ControlNet*38.180.838234.8131.460.7490.8150.313
IPAdapter-Instruct*15.810.219116.1247.810.7880.8520.294
+ +Table 4. Results of two baseline methods and different module combinations of our UNIC-Adapter on pixel-level control tasks and subject-driven generation task. + +features in the main branch. Table 4 shows that the addition of $L_{\mathrm{cross}}^q$ improves results further, particularly in pixel-level control tasks and subject-image similarity. This indicates the importance of $L_{\mathrm{cross}}^q$ in enabling query features $Q_{\mathrm{img}}$ to retrieve specific features from conditional image and task instruction. Figure 6 shows that UNIC-Adapter without position embedding often generates images with incomplete or incorrect details from control signals. + +Comparison with Baseline Methods. We reproduce two baseline methods, ControlNet* and IPAdapter-Instruct*, in our framework. For ControlNet*, we replace the cross-attention mechanism in the image-instruction feature injection module with an addition operation (as did in ControlNet) while retaining other settings. For IPAdapter-Instruct*, we reproduce results by referring to the open-source code on our datasets. As shown in Table 4, ControlNet* achieves performance comparable to our final model on pixel-level control tasks but exhibits suboptimal subject-image similarity on the subject-driven generation task. The visualization results in Figure 7 show that ControlNet* often produces images with distorted subject content. In contrast, IPAdapter-Instruct* performs well on subject-driven generation tasks but poorly on pixel-level control tasks. Due to compressing conditional images into 16 tokens with + +out precise spatial information, the generated images from IPAdapter-Instruct* in Figure 6 fail to accurately capture the pixel-level details of the conditioned images. In comparison, our UNIC-Adapter achieves consistently superior results across both task types, demonstrating its effectiveness and suitability for unified controllable image generation. + +# 5. Conclusion + +In this paper, we present UNIC-Adapter, a unified framework for controllable T2I generation that integrates diverse conditional image inputs and task instructions within a single model. Built on the MM-DiT architecture, our approach leverages cross-attention mechanisms and RoPE to achieve precise pixel-level control, as well as high subject fidelity, across multiple generation tasks. Our extensive experiments on tasks such as pixel-level spatial control, subject-driven image generation, and style-image-based T2I generation demonstrate the effectiveness of the UNIC-Adapter. + +# Acknowledgements + +This research is supported by the NSFC Grants under the contracts No. 62325111 and No.U22B2011. + +# References + +[1] Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 2, 3 +[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. GPT-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 5 +[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science, 2(3):8, 2023. 1 +[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 6 +[5] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 1, 2, 3 +[6] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- $\backslash$ sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. +[7] Junsong Chen, Yue Wu, Simian Luo, Enze Xie, Sayak Paul, Ping Luo, Hang Zhao, and Zhenguo Li. Pixart- $\{\backslash$ delta\} Fast and controllable image generation with latent consistency models. arXiv preprint arXiv:2401.05252, 2024. 1, 2 +[8] Wenhu Chen, Hexiang Hu, Chitwan Sahara, and William W Cohen. Re-Imagen: Retrieval-augmented text-to-image generator. arXiv preprint arXiv:2209.14491, 2022. 5 +[9] Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Ruiz, Xuhui Jia, Ming-Wei Chang, and William W Cohen. Subject-driven text-to-image generation via apprenticeship learning. NeurIPS, 36, 2024. 2, 5 +[10] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, pages 2818-2829, 2023. 4 +[11] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3 +[12] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Int. Conf. Mach. and Learn., 2024. 1, 2, 3, 4, 5 +[13] Junyao Gao, Yanchen Liu, Yanan Sun, Yinhao Tang, Yanhong Zeng, Kai Chen, and Cairong Zhao. StyleShot: A snap + +shot on any style. arXiv preprint arXiv:2407.01414, 2024. 1, 2 +[14] Zhen Han, Chaojie Mao, Zeyinzi Jiang, Yulin Pan, and Jingfeng Zhang. StyleBooth: Image style editing with multimodal instruction. arXiv preprint arXiv:2404.12154, 2024. 2, 5, 6 +[15] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In CVPR, pages 4775–4785, 2024. 1 +[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 1, 2 +[17] Hexiang Hu, Kelvin CK Chan, Yu-Chuan Su, Wenhu Chen, Yandong Li, Kihyuk Sohn, Yang Zhao, Xue Ben, Boqing Gong, William Cohen, et al. Instruct-Imagen: Image generation with multi-modal instruction. In CVPR, pages 4754–4763, 2024. 2, 3, 4 +[18] Lianghua Huang, Di Chen, Yu Liu, Yujun Shen, Deli Zhao, and Jingren Zhou.Composer: Creative and controllable image synthesis with composable conditions. arXiv preprint arXiv:2302.09778, 2023. 1 +[19] Frank Hutter Ilya Loshchilov. Decoupled weight decay regularization. *ICLR*, 2019. 5 +[20] Imagen-Team-Google. Imagen 3, 2024. 1 +[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 5 +[22] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 128(7):1956-1981, 2020. 5 +[23] Black Forest Labs. Flux: Inference repository. https://github.com/black-forest-labs/flux, 2024. Accessed: 2024-11-14. 1, 2, 3, 4 +[24] Dongxu Li, Junnan Li, and Steven Hoi. BLIP-Diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. NeurIPS, 36, 2024. 2, 5 +[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Int. Conf. Mach. and Learn., pages 19730-19742. PMLR, 2023. 5 +[26] Ming Li, Taojiannan Yang, Huafeng Kuang, Jie Wu, Zhaoning Wang, Xuefeng Xiao, and Chen Chen. ControlNet++: Improving conditional controls with efficient consistency feedback. In ECCV, pages 129-147. Springer, 2025. 2, 5, 6 +[27] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. GLIGEN: Open-set grounded text-to-image generation. In CVPR, pages 22511-22521, 2023. 5 +[28] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et al. Hunyuan-DiT: A powerful multi-resolution diffusion transformer with fine-grained + +chinese understanding. arXiv preprint arXiv:2405.08748, 2024.1,2 +[29] Bingchen Liu, Ehsan Akhgari, Alexander Visheratin, Aleks Kamko, Linmiao Xu, Shivam Shrirao, Joao Souza, Suhail Doshi, and Daiqing Li. Playground v3: Improving text-to-image alignment with deep-fusion large language models. arXiv preprint arXiv:2409.10695, 2024. 1 +[30] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding DINO: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 5 +[31] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-Diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. In ACM SIGGRAPH Conference Papers, pages 1-12, 2024. 2, 5 +[32] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, pages 6038–6047, 2023. 5 +[33] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, pages 4296-4304, 2024. 1, 5 +[34] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-G: Generating images in context with multimodal large language models. arXiv preprint arXiv:2310.02992, 2023. 1, 5, 6 +[35] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 1 +[36] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 2 +[37] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 5 +[38] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 2 +[39] Can Qin, Shu Zhang, Ning Yu, Yihao Feng, Xinyi Yang, Yingbo Zhou, Huan Wang, Juan Carlos Niebles, Caiming Xiong, Silvio Savarese, et al. UniControl: A unified diffusion model for controllable visual generation in the wild. arXiv preprint arXiv:2305.11147, 2023. 2, 3, 5 +[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Int. Conf. Mach. and Learn., pages 8748-8763. PMLR, 2021. 2, 3, 4, 5, 6 +[41] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 2, 3, 4 + +[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 2 +[43] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 234–241. Springer, 2015. 1 +[44] Ciara Rowles, Shimon Vainer, Dante De Nigris, Slava Elizarov, Konstantin Kutsy, and Simon Donne. IPAdapter-Instruct: Resolving ambiguity in image-based conditioning using instruct prompts. arXiv preprint arXiv:2408.03209, 2024. 2, 3 +[45] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500-22510, 2023. 1, 5, 6 +[46] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 35:36479-36494, 2022. 2, 3 +[47] Babak Saleh and Ahmed Elgammal. Large-scale classification of fine-art paintings: Learning the right metric on the right feature. arXiv preprint arXiv:1505.00855, 2015. 5 +[48] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. LAION-5B: An open large-scale dataset for training next generation image-text models. NeurIPS, 35:25278-25294, 2022. 1 +[49] Kunpeng Song, Yizhe Zhu, Bingchen Liu, Qing Yan, Ahmed Elgammal, and Xiao Yang. Moma: Multimodal llm adapter for fast personalized image generation. arXiv preprint arXiv:2404.05674, 2024. 2 +[50] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. RoFormer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 2, 4 +[51] Kcolors Team. Kcolors: Effective training of diffusion model for photorealistic text-to-image synthesis. arXiv preprint, 2024. 1 +[52] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2 +[53] A Vaswani. Attention is all you need. NeurIPS, 2017. 2, 3 +[54] Haofan Wang, Matteo Spinelli, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024. 1, 2 +[55] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 1, 4 + +[56] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. arXiv preprint arXiv:2409.11340, 2024. 2, 3, 5, 6 +[57] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. CogVideoX: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 4 +[58] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. IP-Adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2, 3, 4 +[59] Zilyu Ye, Jinxiu Liu, Ruotian Peng, Jinjin Cao, Zhiyang Chen, Yiyang Zhang, Ziwei Xuan, Mingyuan Zhou, Xiaogian Shen, Mohamed Elhoseiny, et al. Openstory++: A large-scale dataset and benchmark for instance-aware open-domain visual storytelling. arXiv preprint arXiv:2408.03695, 2024. 5 +[60] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 1, 2, 5 +[61] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. UniControlNet: All-in-one control to text-to-image diffusion models. NeurIPS, 36, 2024. 2, 3, 5 +[62] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In CVPR, pages 22490-22499, 2023. 1, 2 +[63] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, pages 633-641, 2017. 5 \ No newline at end of file diff --git a/2025/UNIC-Adapter_ Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation/images.zip b/2025/UNIC-Adapter_ Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ff2a07655b64e881b1cf3941030de49d75eaf715 --- /dev/null +++ b/2025/UNIC-Adapter_ Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aaf4ad1ee4d284b8b968c1a048c00fd03a6a563065a8e1c037371015fe937ebb +size 926516 diff --git a/2025/UNIC-Adapter_ Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation/layout.json b/2025/UNIC-Adapter_ Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4f077368ba78d55e5a415859c5e675427b45167e --- /dev/null +++ b/2025/UNIC-Adapter_ Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation/layout.json @@ -0,0 +1,8805 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 102, + 520, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 102, + 520, + 140 + ], + "spans": [ + { + "bbox": [ + 91, + 102, + 520, + 140 + ], + "type": "text", + "content": "UNIC-Adapter: Unified Image-instruction Adapter with Multi-modal Transformer for Image Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 160, + 511, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 160, + 511, + 190 + ], + "spans": [ + { + "bbox": [ + 99, + 160, + 511, + 190 + ], + "type": "text", + "content": "Lunhao Duan\\*1 Shanshan Zhao\\*2 Wenjun Yan2 Yinglun Li2 Qing-Guo Chen2 Zhao Xu2 Weihua Luo2 Kaifu Zhang2 Mingming Gong3 Gui-Song Xia\\*14" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 81, + 202, + 529, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 202, + 529, + 232 + ], + "spans": [ + { + "bbox": [ + 81, + 202, + 529, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 81, + 202, + 529, + 232 + ], + "type": "text", + "content": "School of Computer Science, Wuhan University " + }, + { + "bbox": [ + 81, + 202, + 529, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 81, + 202, + 529, + 232 + ], + "type": "text", + "content": "Alibaba International Digital Commerce " + }, + { + "bbox": [ + 81, + 202, + 529, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 81, + 202, + 529, + 232 + ], + "type": "text", + "content": "University of Melbourne " + }, + { + "bbox": [ + 81, + 202, + 529, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 81, + 202, + 529, + 232 + ], + "type": "text", + "content": "School of Artificial Intelligence, Wuhan University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 284, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 296, + 594 + ], + "type": "text", + "content": "Recently, text-to-image generation models have achieved remarkable advancements, particularly with diffusion models facilitating high-quality image synthesis from textual descriptions. However, these models often struggle with achieving precise control over pixel-level layouts, object appearances, and global styles when using text prompts alone. To mitigate this issue, previous works introduce conditional images as auxiliary inputs for image generation, enhancing control but typically necessitating specialized models tailored to different types of reference inputs. In this paper, we explore a new approach to unify controllable generation within a single framework. Specifically, we propose the unified image-instruction adapter (UNIC-Adapter) built on the Multi-Modal-Diffusion Transformer architecture, to enable flexible and controllable generation across diverse conditions without the need for multiple specialized models. Our UNIC-Adapter effectively extracts multi-modal instruction information by incorporating both conditional images and task instructions, injecting this information into the image generation process through a cross-attention mechanism enhanced by Rotary Position Embedding. Experimental results across a variety of tasks, including pixel-level spatial control, subject-driven image generation, and style-image-based image synthesis, demonstrate the effectiveness of our UNIC-Adapter in unified controllable image generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 607, + 135, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 607, + 135, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 607, + 135, + 620 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 627, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 627, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 627, + 295, + 677 + ], + "type": "text", + "content": "Recent advancements in text-to-image (T2I) generation models [3, 5, 12, 20, 23, 29, 38, 42, 51], primarily driven by diffusion models [16], have been witnessed in recent years. By leveraging large-scale paired text-image data [48], nu" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 335, + 262, + 343, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 262, + 343, + 270 + ], + "spans": [ + { + "bbox": [ + 335, + 262, + 343, + 270 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 356, + 260, + 383, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 260, + 383, + 274 + ], + "spans": [ + { + "bbox": [ + 356, + 260, + 383, + 274 + ], + "type": "text", + "content": "A delicious afternoon tea." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 395, + 260, + 432, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 260, + 432, + 274 + ], + "spans": [ + { + "bbox": [ + 395, + 260, + 432, + 274 + ], + "type": "text", + "content": "A room with the forest backdrop." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 438, + 260, + 478, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 438, + 260, + 478, + 274 + ], + "spans": [ + { + "bbox": [ + 438, + 260, + 478, + 274 + ], + "type": "text", + "content": "A robot in autumn mountain and lake" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 481, + 262, + 488, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 481, + 262, + 488, + 270 + ], + "spans": [ + { + "bbox": [ + 481, + 262, + 488, + 270 + ], + "type": "text", + "content": "···" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 498, + 260, + 533, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 498, + 260, + 533, + 274 + ], + "spans": [ + { + "bbox": [ + 498, + 260, + 533, + 274 + ], + "type": "text", + "content": "A dog with a mountain in the" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 335, + 284, + 343, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 284, + 343, + 291 + ], + "spans": [ + { + "bbox": [ + 335, + 284, + 343, + 291 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 358, + 279, + 383, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 279, + 383, + 297 + ], + "spans": [ + { + "bbox": [ + 358, + 279, + 383, + 297 + ], + "type": "text", + "content": "Generate image from canny edge." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 396, + 278, + 432, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 278, + 432, + 297 + ], + "spans": [ + { + "bbox": [ + 396, + 278, + 432, + 297 + ], + "type": "text", + "content": "Render image with this depth map." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 438, + 279, + 476, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 438, + 279, + 476, + 297 + ], + "spans": [ + { + "bbox": [ + 438, + 279, + 476, + 297 + ], + "type": "text", + "content": "Create image using style influence." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 479, + 284, + 488, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 479, + 284, + 488, + 291 + ], + "spans": [ + { + "bbox": [ + 479, + 284, + 488, + 291 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 496, + 284, + 533, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 496, + 284, + 533, + 297 + ], + "spans": [ + { + "bbox": [ + 496, + 284, + 533, + 297 + ], + "type": "text", + "content": "Generate image from dog image" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 335, + 319, + 343, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 319, + 343, + 326 + ], + "spans": [ + { + "bbox": [ + 335, + 319, + 343, + 326 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 347, + 300, + 391, + 344 + ], + "blocks": [ + { + "bbox": [ + 347, + 300, + 391, + 344 + ], + "lines": [ + { + "bbox": [ + 347, + 300, + 391, + 344 + ], + "spans": [ + { + "bbox": [ + 347, + 300, + 391, + 344 + ], + "type": "image", + "image_path": "e13f90acba9ec6aacfb88fabb46d30e3470f85c4d61206fc2e5996f349e45db0.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 392, + 300, + 435, + 344 + ], + "blocks": [ + { + "bbox": [ + 392, + 300, + 435, + 344 + ], + "lines": [ + { + "bbox": [ + 392, + 300, + 435, + 344 + ], + "spans": [ + { + "bbox": [ + 392, + 300, + 435, + 344 + ], + "type": "image", + "image_path": "67f8af2831537e9bc2cdbf08c4e98da41798927a67bf4f09deb17debb8353b64.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 436, + 300, + 481, + 344 + ], + "blocks": [ + { + "bbox": [ + 436, + 300, + 481, + 344 + ], + "lines": [ + { + "bbox": [ + 436, + 300, + 481, + 344 + ], + "spans": [ + { + "bbox": [ + 436, + 300, + 481, + 344 + ], + "type": "image", + "image_path": "ec71007de18d7c771c97f474b57a70353f275ce361265f88ab2635cded58dbea.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 483, + 300, + 536, + 344 + ], + "blocks": [ + { + "bbox": [ + 483, + 300, + 536, + 344 + ], + "lines": [ + { + "bbox": [ + 483, + 300, + 536, + 344 + ], + "spans": [ + { + "bbox": [ + 483, + 300, + 536, + 344 + ], + "type": "image", + "image_path": "bf9742303b6ada9ef0af02f9f54e4fb61fe888b4bbcae808cc4df645c4135378.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 335, + 380, + 343, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 380, + 343, + 388 + ], + "spans": [ + { + "bbox": [ + 335, + 380, + 343, + 388 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 347, + 346, + 391, + 405 + ], + "blocks": [ + { + "bbox": [ + 347, + 346, + 391, + 405 + ], + "lines": [ + { + "bbox": [ + 347, + 346, + 391, + 405 + ], + "spans": [ + { + "bbox": [ + 347, + 346, + 391, + 405 + ], + "type": "image", + "image_path": "56439bdfacd22dfd3e04f3aec4996bd6946b20e5b170326e6ad5723be1ef36f7.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 413, + 555, + 458 + ], + "lines": [ + { + "bbox": [ + 313, + 413, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 555, + 458 + ], + "type": "text", + "content": "Figure 1. With the UNIC-Adapter, SD3 enables flexible and controllable generation across multiple reference modalities within a single model. (a) to (d) represent text prompts, task instructions, conditional images, and generated images, respectively." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 391, + 346, + 435, + 405 + ], + "blocks": [ + { + "bbox": [ + 391, + 346, + 435, + 405 + ], + "lines": [ + { + "bbox": [ + 391, + 346, + 435, + 405 + ], + "spans": [ + { + "bbox": [ + 391, + 346, + 435, + 405 + ], + "type": "image", + "image_path": "0a25eff6fd53b9af61b0eb36efac739d6727206d4c74744092410610e3d0cd43.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 436, + 346, + 481, + 405 + ], + "blocks": [ + { + "bbox": [ + 436, + 346, + 481, + 405 + ], + "lines": [ + { + "bbox": [ + 436, + 346, + 481, + 405 + ], + "spans": [ + { + "bbox": [ + 436, + 346, + 481, + 405 + ], + "type": "image", + "image_path": "7c064a96448ffe8b22842ad46f2a0b7af58d5d197ab88015455f1cb5e31ae8ee.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 483, + 346, + 536, + 405 + ], + "blocks": [ + { + "bbox": [ + 483, + 346, + 536, + 405 + ], + "lines": [ + { + "bbox": [ + 483, + 346, + 536, + 405 + ], + "spans": [ + { + "bbox": [ + 483, + 346, + 536, + 405 + ], + "type": "image", + "image_path": "2408286dc7a0cf4f9c10d82be303c4924ffee240e3f6554ba77e1632434dd055.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 472, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 567 + ], + "type": "text", + "content": "merous open-source T2I models, like Stable Diffusion 1.5 (SD1.5) [42], Stable Diffusion XL (SDXL) [38], Pixart-series [5-7], Stable Diffusion 3 (SD3) [12], Hunyuan-DiT [28], and FLUX.1-dev [23] have significantly enhanced the fidelity of images generated from natural language descriptions. These models utilize a range of architectures, spanning from U-Net-based designs [43] to the more recent Diffusion-Transformer-based frameworks [35]." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 312, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 570, + 556, + 714 + ], + "type": "text", + "content": "Despite these advancements, relying solely on text prompts often falls short in defining pixel-level spatial structures and geometric details, exact object appearances, and global styles of a specific image. To address this limitation, recent research [13, 18, 33, 34, 45, 55, 58, 60] has incorporated conditional images as additional inputs alongside text prompts. These conditional images offer finer control over pixel-level [60, 62], object-level [34, 55, 58], or style-specific [13, 15, 54] features in the generated outputs. For example, ControlNet [60] enables pixel-level spatial control by integrating structural inputs, such as depth and edge maps, through a parallel encoder tailored for" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": "*This work was done when Lunhao Duan was a research intern at Alibaba International Digital Commerce. \n†Corresponding authors." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7963" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": "each type of control. Similarly, IP-Adapter [58] enables content-specific control by conditioning on a reference image and injecting its CLIP-based [40] embeddings via crossattention layers. However, these methods often necessitate individual model training for each control type, resulting in increased training costs and complicating the integration of different control inputs. IPAdapter-Instruct [44] extends IP-Adapter by incorporating instruction prompts to specify task types for different conditional images, enabling multi-task handling within a single model. UniControl [39] tackles various pixel-level control tasks within a unified model by introducing task-specific adapters to extract features for different visual conditions. Meanwhile, Uni-ControlNet [61] divides conditional images into two groups: local (pixel-level) and global (image-level) conditions, and designs dedicated adapters for each group. However, this approach is not fully unified, as the two adapters are trained separately." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 291, + 295, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 295, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 295, + 447 + ], + "type": "text", + "content": "To further unify diverse reference inputs, including text, edge maps, content images, etc., Instruct-Imagen [17], based onImagen [46], represents them as multi-modal instructions within a single framework. OmniGen [56] first tokenizes interleaved text and images into a one-dimensional sequence and then uses a transformer model (Phi3) [1] to achieve unified image generation, such as image editing and subject-driven generation. To adapt the model for multi-modal instructions, these two methods involve multi-stage training of the entire model. To enable more efficient training for unified controllable image generation, this paper seeks to design an adapter based on a pre-trained T2I model, avoiding the need to train the entire model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 450, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 714 + ], + "type": "text", + "content": "Given that DiT-based architectures [36] equipped with pure transformer blocks have become the mainstream in T2I models [5, 12, 23, 28], we choose SD3, an open-source state-of-the-art (SOTA) model, as the foundation model for studying the adapter. To achieve unified controllable image generation across various conditions, we employ the Multi-Modal-Diffusion Transformer (MM-DiT) structure [12, 23], which enables full cross-attention between text and image features to extract multi-modal instruction information. By treating text as an independent modality, the MM-DiT allocates substantial model capacity to capture and understand text-based information, thereby improving its interpretation of text prompts [12]. Leveraging the strengths of the MM-DiT, we propose a unified image-instruction adapter (UNIC-Adapter) to facilitate image generation conditioned on various types of reference images with a task instruction indicating the desired generation target. Specifically, task instruction and conditional image features are extracted by text encoders [40, 41] and a Variational Autoencoder (VAE) [12], respectively. Then, the adapter processes instruction and image features using MM-DiT blocks, which are subsequently integrated into the" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "main generation branch via a cross-attention mechanism. To further enhance spatial awareness, which is essential for pixel-level control, we incorporate Rotary Position Embedding (RoPE) [50] into the query and key features within the cross-attention. With the proposed UNIC-Adapter, SD3 can generate images under a wide array of conditions, as shown in Figure 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 325, + 156, + 504, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 156, + 504, + 167 + ], + "spans": [ + { + "bbox": [ + 325, + 156, + 504, + 167 + ], + "type": "text", + "content": "In summary, our contributions are threefold:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 169, + 553, + 276 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 169, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 169, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 314, + 169, + 553, + 204 + ], + "type": "text", + "content": "- We are the first to leverage the MM-DiT architecture for unified controllable generation across various types of reference images in T2I models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 205, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 239 + ], + "type": "text", + "content": "- We introduce an effective feature injection method based on cross-attention, enhanced with RoPE encoding for better spatial control." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 240, + 553, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 553, + 276 + ], + "type": "text", + "content": "- Equipped with our UNIC-Adapter, SD3 enables unified controllable generation across 14 types of conditional images in our experiments." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 289, + 400, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 289, + 400, + 301 + ], + "spans": [ + { + "bbox": [ + 314, + 289, + 400, + 301 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 309, + 495, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 495, + 322 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 495, + 322 + ], + "type": "text", + "content": "2.1. Text-to-Image Generation Models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 327, + 555, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 591 + ], + "type": "text", + "content": "Mainstream T2I generation models [5, 12, 38, 42] are primarily based on diffusion models [16], which generate images by learning to reverse a progressive noise-adding process. In these models, text input typically serves as a conditional signal, guiding the generation process to align with the provided description. The success of transformer architectures [53], particularly in large language models [2, 52], has significantly influenced T2I model architectures, evolving them from convolution-transformer hybrids [38, 42] to fully transformer-based designs [5-7, 12, 23, 28]. Recently, the MM-DiT architecture [12, 23] has advanced this evolution by treating text and image features as distinct modalities with full-attention mechanisms, enabling each modality to attend to the other. This approach outperforms traditional cross-attention setups, where text features are used only as keys and values [12]. However, relying solely on text-based conditioning presents limitations, particularly for images that require detailed spatial layouts or specific object appearances. Our work builds upon these transformer-based diffusion models by integrating additional conditional images and task instructions, thereby enhancing the flexibility and controllability of T2I generation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 599, + 511, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 511, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 511, + 612 + ], + "type": "text", + "content": "2.2. Conditional Image-based T2I Models" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "Conditional image-based T2I models introduce auxiliary image inputs to complement textual prompts, enabling finer control over aspects such as spatial layout [26, 60, 62], object content [9, 24, 31, 49, 58], and stylistic elements [13, 14, 54]. For pixel-level spatial control, ControlNet [60] employs a duplicated U-Net encoder to process specific types of conditional images (e.g., edge, pose, and depth maps) and integrates the features into the main T2I model" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7964" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 95, + 70, + 517, + 277 + ], + "blocks": [ + { + "bbox": [ + 95, + 70, + 517, + 277 + ], + "lines": [ + { + "bbox": [ + 95, + 70, + 517, + 277 + ], + "spans": [ + { + "bbox": [ + 95, + 70, + 517, + 277 + ], + "type": "image", + "image_path": "86f0ef7ac179a0d0bcc1d0ab7b46c475a1cb21fcaca5ce2d236ff9e08f372406.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 285, + 555, + 330 + ], + "lines": [ + { + "bbox": [ + 54, + 285, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 54, + 285, + 555, + 330 + ], + "type": "text", + "content": "Figure 2. The overall architecture of our proposed UNIC-Adapter. The task instruction and conditional image features are progressively attending to each other through a series of " + }, + { + "bbox": [ + 54, + 285, + 555, + 330 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 285, + 555, + 330 + ], + "type": "text", + "content": " adapter blocks. In each adapter block, the image features " + }, + { + "bbox": [ + 54, + 285, + 555, + 330 + ], + "type": "inline_equation", + "content": "Z_{\\mathrm{img}}" + }, + { + "bbox": [ + 54, + 285, + 555, + 330 + ], + "type": "text", + "content": " from MM-DiT block in the main image generation branch serve as the query, while both task instruction features " + }, + { + "bbox": [ + 54, + 285, + 555, + 330 + ], + "type": "inline_equation", + "content": "Z_{\\mathrm{ist}}" + }, + { + "bbox": [ + 54, + 285, + 555, + 330 + ], + "type": "text", + "content": " and conditional image features " + }, + { + "bbox": [ + 54, + 285, + 555, + 330 + ], + "type": "inline_equation", + "content": "Z_{\\mathrm{con}}" + }, + { + "bbox": [ + 54, + 285, + 555, + 330 + ], + "type": "text", + "content": " function as keys and values. For simplicity, normalization layers and feed-forward networks are omitted in this figure." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 350, + 296, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 296, + 709 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 296, + 709 + ], + "type": "text", + "content": "through residual connections in the U-Net decoder. Another key work, IP-Adapter [58], uses the Vision Transformer (ViT) [11] encoder from CLIP [40] to extract features from conditional images. These features are then incorporated into the U-Net backbone via cross-attention layers, achieving controllable generation with contents from conditional images. IPAdapter-Instruct [44] extends this by introducing an instructional prompt that guides the interpretation of conditional images. To minimize the necessity for individual model training for each type of conditional image in ControlNet, UniControl [39] unifies nine pixel-level control tasks within a single model, leveraging task-specific feature extraction adapters and a task-aware HyperNet to modulate features based on task instructions. To achieve control with image content as well as spatial layout, UniControlNet [61] incorporates two dedicated adapters for local (pixel-level control) and global (CLIP-image embedding control) conditions, respectively. Unlike these methods, which freeze the base model and only training the newly added modules / adapters, Instruct-Imagen [17] proposes using multi-modal instructions and optimize the entire model, including the base model (Imagen [46]) and the new cross-attention layers, for image generation under different conditions. A recent work, OmniGen [56], develops a unified image generation model for various multi-modal inputs using a transformer-based architecture [1]. In this paper, we aim to enhance the generation capabilities of an existing T2I model for diverse conditions by introducing a unified image-instruction adapter, named UNIC-Adapter, without the need to train the entire model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 314, + 350, + 370, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 350, + 370, + 361 + ], + "spans": [ + { + "bbox": [ + 314, + 350, + 370, + 361 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 369, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 430 + ], + "type": "text", + "content": "In this section, we first provide a brief overview of the DiT and MM-DiT employed in SOTA T2I models. Then, we describe our UNIC-Adapter, which incorporates task instructions and conditional images to enhance the model's capability to interpret various user-specified conditions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 435, + 394, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 394, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 394, + 448 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 451, + 556, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 556, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 556, + 552 + ], + "type": "text", + "content": "Diffusion Transformer Block. In DiT Block-based T2I models [5], conditional text features " + }, + { + "bbox": [ + 313, + 451, + 556, + 552 + ], + "type": "inline_equation", + "content": "Z_{\\mathrm{txt}}" + }, + { + "bbox": [ + 313, + 451, + 556, + 552 + ], + "type": "text", + "content": " extracted from text encoders [40, 41] are integrated into DiT Blocks via cross-attention layers. Specifically, in each DiT Block, the conditional text features are first mapped to key " + }, + { + "bbox": [ + 313, + 451, + 556, + 552 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{txt}}" + }, + { + "bbox": [ + 313, + 451, + 556, + 552 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 313, + 451, + 556, + 552 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{txt}}" + }, + { + "bbox": [ + 313, + 451, + 556, + 552 + ], + "type": "text", + "content": " features through linear transformations " + }, + { + "bbox": [ + 313, + 451, + 556, + 552 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 451, + 556, + 552 + ], + "type": "text", + "content": ", while the image feature " + }, + { + "bbox": [ + 313, + 451, + 556, + 552 + ], + "type": "inline_equation", + "content": "Z_{\\mathrm{img}}" + }, + { + "bbox": [ + 313, + 451, + 556, + 552 + ], + "type": "text", + "content": " is mapped to a query feature " + }, + { + "bbox": [ + 313, + 451, + 556, + 552 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{img}}" + }, + { + "bbox": [ + 313, + 451, + 556, + 552 + ], + "type": "text", + "content": ". An attention mechanism then updates the image features " + }, + { + "bbox": [ + 313, + 451, + 556, + 552 + ], + "type": "inline_equation", + "content": "Z_{\\mathrm{img}}" + }, + { + "bbox": [ + 313, + 451, + 556, + 552 + ], + "type": "text", + "content": " by attending to " + }, + { + "bbox": [ + 313, + 451, + 556, + 552 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{txt}}" + }, + { + "bbox": [ + 313, + 451, + 556, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 451, + 556, + 552 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{txt}}" + }, + { + "bbox": [ + 313, + 451, + 556, + 552 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 328, + 555, + 553, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 555, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 328, + 555, + 553, + 574 + ], + "type": "interline_equation", + "content": "K _ {\\mathrm {t x t}} = L _ {\\mathrm {t x t}} ^ {k} \\left(Z _ {\\mathrm {t x t}}\\right), V _ {\\mathrm {t x t}} = L _ {\\mathrm {t x t}} ^ {v} \\left(Z _ {\\mathrm {t x t}}\\right), Q _ {\\mathrm {i m g}} = L _ {\\mathrm {i m g}} ^ {q} \\left(Z _ {\\mathrm {i m g}}\\right), \\tag {1}", + "image_path": "debfd9bb12a7ddf3f7169033d2b3f59eab53a338cffe1c7193d7a58145ac2583.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 329, + 572, + 437, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 572, + 437, + 586 + ], + "spans": [ + { + "bbox": [ + 329, + 572, + 437, + 586 + ], + "type": "interline_equation", + "content": "Z _ {\\mathrm {i m g}} ^ {\\prime} = \\operatorname {A t t n} \\left(Q _ {\\mathrm {i m g}}, K _ {\\mathrm {t x t}}, V _ {\\mathrm {t x t}}\\right),", + "image_path": "7d1606e3b822c8a7903d44b929faac2f1532490aa2af03fceb9300546a5c7865.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 588, + 555, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 555, + 635 + ], + "type": "text", + "content": "where Attn indicates the multi-head attention [53]. In this setup, the same " + }, + { + "bbox": [ + 313, + 588, + 555, + 635 + ], + "type": "inline_equation", + "content": "Z_{\\mathrm{txt}}" + }, + { + "bbox": [ + 313, + 588, + 555, + 635 + ], + "type": "text", + "content": " features are provided to DiT Blocks across different layers as static conditioning signals, which may limit the utilization of text features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 635, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 556, + 713 + ], + "type": "text", + "content": "Multi-modal Diffusion Transformer Block. Recent T2I models [12, 23] based on MM-DiT Blocks enhance DiT Blocks by treating conditional text as an independent modality capable of dynamically interacting with image features across layers. In MM-DiT, both text and image features serve as queries, keys, and values, enabling cross-modal attention in both directions through a full-attention" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7965" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "mechanism. Specifically, the MM-DiT layer performs the following operations:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 102, + 295, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 102, + 295, + 121 + ], + "spans": [ + { + "bbox": [ + 61, + 102, + 295, + 121 + ], + "type": "interline_equation", + "content": "Q _ {\\mathrm {t x t}} = L _ {\\mathrm {t x t}} ^ {q} \\left(Z _ {\\mathrm {t x t}}\\right), K _ {\\mathrm {t x t}} = L _ {\\mathrm {t x t}} ^ {k} \\left(Z _ {\\mathrm {t x t}}\\right), V _ {\\mathrm {t x t}} = L _ {\\mathrm {t x t}} ^ {v} \\left(Z _ {\\mathrm {t x t}}\\right), \\tag {2}", + "image_path": "7e877b13f8fc7cb091fe5c530b66464816dceeba772302e0a229fc96e0d37779.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 119, + 279, + 132 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 119, + 279, + 132 + ], + "spans": [ + { + "bbox": [ + 61, + 119, + 279, + 132 + ], + "type": "interline_equation", + "content": "Q _ {\\mathrm {i m g}} = L _ {\\mathrm {i m g}} ^ {q} \\left(Z _ {\\mathrm {i m g}}\\right), K _ {\\mathrm {i m g}} = L _ {\\mathrm {i m g}} ^ {k} \\left(Z _ {\\mathrm {i m g}}\\right), V _ {\\mathrm {i m g}} = L _ {\\mathrm {i m g}} ^ {v} \\left(Z _ {\\mathrm {i m g}}\\right).", + "image_path": "b5c55dac401cc8d89bf7deeb2fc875875e14cfbe7f74b4a04e80c9e707d96904.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 137, + 295, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 137, + 295, + 174 + ], + "spans": [ + { + "bbox": [ + 55, + 137, + 295, + 174 + ], + "type": "text", + "content": "These queries, keys, and values are processed in a full-attention layer, where each modality attends to both text and image features:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 182, + 295, + 202 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 182, + 295, + 202 + ], + "spans": [ + { + "bbox": [ + 85, + 182, + 295, + 202 + ], + "type": "interline_equation", + "content": "Z _ {\\mathrm {t x t}} ^ {\\prime} = \\operatorname {A t t n} \\left(Q _ {\\mathrm {t x t}}, \\left[ K _ {\\mathrm {t x t}} \\| K _ {\\mathrm {i m g}} \\right], \\left[ V _ {\\mathrm {t x t}} \\| V _ {\\mathrm {i m g}} \\right]\\right), \\tag {3}", + "image_path": "6a435f22266457ba4b6c837dd1c6a7e54baf7269d579d3bc805258db659d67ae.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 200, + 266, + 215 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 200, + 266, + 215 + ], + "spans": [ + { + "bbox": [ + 86, + 200, + 266, + 215 + ], + "type": "interline_equation", + "content": "Z _ {\\mathrm {i m g}} ^ {\\prime} = \\operatorname {A t t n} \\left(Q _ {\\mathrm {i m g}}, \\left[ K _ {\\mathrm {t x t}} \\right\\| K _ {\\mathrm {i m g}} ], \\left[ V _ {\\mathrm {t x t}} \\right\\| V _ {\\mathrm {i m g}}\\right),", + "image_path": "0a063483be308427b41e75d18d8be5f34eb68cb9312e27c4a1b8bc9472a90fc8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 222, + 296, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 222, + 296, + 271 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 296, + 271 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 222, + 296, + 271 + ], + "type": "inline_equation", + "content": "\\parallel" + }, + { + "bbox": [ + 55, + 222, + 296, + 271 + ], + "type": "text", + "content": " denotes concatenation. This enhanced interaction facilitates richer cross-modal feature extraction, as text and image features are iteratively refined across layers, improving text-guided image generation [12]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 277, + 149, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 149, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 149, + 290 + ], + "type": "text", + "content": "3.2. UNIC-Adapter" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 295, + 296, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 296, + 376 + ], + "type": "text", + "content": "As shown in Figure 2, the UNIC-Adapter introduces two additional inputs: task instructions and conditional images. This integration enables unified controllable image generation by allowing the model to interpret and apply specific features from a variety of conditional images, guided by the task instruction. Below, we provide a detailed description of its structure." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 378, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 296, + 531 + ], + "type": "text", + "content": "Image-instruction Feature Extraction. To enable a comprehensive understanding of task instructions and conditional images, we leverage MM-DiT blocks to enhance the interaction between these inputs. Specifically, we adopt SD3 as our base model. Following the setup in SD3, task instruction features, denoted as " + }, + { + "bbox": [ + 55, + 378, + 296, + 531 + ], + "type": "inline_equation", + "content": "Z_{\\mathrm{ist}}" + }, + { + "bbox": [ + 55, + 378, + 296, + 531 + ], + "type": "text", + "content": ", are extracted using text encoders such as CLIP-G/14 [10], CLIP-L/15 [40], and T5 XXL [41]. Simultaneously, conditional image features, " + }, + { + "bbox": [ + 55, + 378, + 296, + 531 + ], + "type": "inline_equation", + "content": "Z_{\\mathrm{con}}" + }, + { + "bbox": [ + 55, + 378, + 296, + 531 + ], + "type": "text", + "content": ", which represent one of several condition types (e.g., edge maps, depth maps, subject content images, or style-reference images), are obtained through the VAE. Within the MM-DiT blocks, task instruction and conditional image features undergo mutual attention updates, defined as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 536, + 249, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 536, + 249, + 550 + ], + "spans": [ + { + "bbox": [ + 63, + 536, + 249, + 550 + ], + "type": "interline_equation", + "content": "Q _ {\\text {i s t}} = L _ {\\text {i s t}} ^ {q} (Z _ {\\text {i s t}}), K _ {\\text {i s t}} = L _ {\\text {i s t}} ^ {k} (Z _ {\\text {i s t}}), V _ {\\text {i s t}} = L _ {\\text {i s t}} ^ {v} (Z _ {\\text {i s t}}),", + "image_path": "86a465ed006da1ae27ae7dde063687f1d9a30ad9ed3da9323270c89ec25e527a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 64, + 552, + 295, + 573 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 552, + 295, + 573 + ], + "spans": [ + { + "bbox": [ + 64, + 552, + 295, + 573 + ], + "type": "interline_equation", + "content": "Q _ {\\text {c o n}} = L _ {\\text {c o n}} ^ {q} \\left(Z _ {\\text {c o n}}\\right), K _ {\\text {c o n}} = L _ {\\text {c o n}} ^ {k} \\left(Z _ {\\text {c o n}}\\right), V _ {\\text {c o n}} = L _ {\\text {c o n}} ^ {v} \\left(Z _ {\\text {c o n}}\\right), \\tag {4}", + "image_path": "ef5c2863b349be0a9737ee918872404b696a3796a5098d921dca269921a4649f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 65, + 569, + 216, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 569, + 216, + 583 + ], + "spans": [ + { + "bbox": [ + 65, + 569, + 216, + 583 + ], + "type": "interline_equation", + "content": "Z _ {\\mathrm {i s t}} ^ {\\prime} = \\operatorname {A t t n} \\left(Q _ {\\mathrm {i s t}}, \\left[ K _ {\\mathrm {i s t}} \\| K _ {\\mathrm {c o n}} \\right], \\left[ V _ {\\mathrm {i s t}} \\| V _ {\\mathrm {c o n}} \\right]\\right),", + "image_path": "9145e076fe0cac4a84b5d4bf9442507552333a0852f70e1b0cc62c45b2f52f3e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 65, + 586, + 220, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 586, + 220, + 600 + ], + "spans": [ + { + "bbox": [ + 65, + 586, + 220, + 600 + ], + "type": "interline_equation", + "content": "Z _ {\\text {c o n}} ^ {\\prime} = \\operatorname {A t t n} \\left(Q _ {\\text {c o n}}, \\left[ K _ {\\text {i s t}} \\| K _ {\\text {c o n}} \\right], \\left[ V _ {\\text {i s t}} \\| V _ {\\text {c o n}} \\right]\\right).", + "image_path": "3120e3f4bd8d0fc8489077dd2ae98e2625c8f4ff3b2c0209ddb8577f53ce727f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 605, + 295, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 642 + ], + "type": "text", + "content": "These updates ensure that the task instruction " + }, + { + "bbox": [ + 55, + 605, + 295, + 642 + ], + "type": "inline_equation", + "content": "Z_{\\mathrm{ist}}" + }, + { + "bbox": [ + 55, + 605, + 295, + 642 + ], + "type": "text", + "content": " and conditional image " + }, + { + "bbox": [ + 55, + 605, + 295, + 642 + ], + "type": "inline_equation", + "content": "Z_{\\mathrm{con}}" + }, + { + "bbox": [ + 55, + 605, + 295, + 642 + ], + "type": "text", + "content": " dynamically refine each other's features by selectively attending to relevant information." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Image-instruction Feature Injection. To improve the model's responsiveness to diverse user-defined conditions, the key and value features from both instruction and conditional image inputs are integrated simultaneously into the main image generation branch through cross-attention [17, 55, 58]. This mechanism ensures that the feature injection is" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 72, + 555, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 170 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 170 + ], + "type": "text", + "content": "directed by the specific task instruction. Experiments indicate that injecting either conditional image features or task instruction features alone through cross-attention results in suboptimal performance, as demonstrated in the ablation studies of Section 4. By contrast, the simultaneous incorporation of both feature types optimally supports unified controlled image generation. Accordingly, the image features " + }, + { + "bbox": [ + 313, + 72, + 555, + 170 + ], + "type": "inline_equation", + "content": "Z_{\\mathrm{img}}^{\\prime}" + }, + { + "bbox": [ + 313, + 72, + 555, + 170 + ], + "type": "text", + "content": " from Equation 3 are further updated as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 324, + 183, + 553, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 183, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 324, + 183, + 553, + 205 + ], + "type": "interline_equation", + "content": "Q _ {\\mathrm {i m g}} ^ {\\prime} = L _ {\\text {c r o s s}} ^ {q} \\left(Z _ {\\mathrm {i m g}}\\right), \\tag {5}", + "image_path": "3f36fe885e9502c098cc048307d0ecba56b3a227199695ec3d0d9fdcaec49d90.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 326, + 202, + 531, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 202, + 531, + 217 + ], + "spans": [ + { + "bbox": [ + 326, + 202, + 531, + 217 + ], + "type": "interline_equation", + "content": "Z _ {\\mathrm {i m g}} ^ {\\prime \\prime} = Z _ {\\mathrm {i m g}} ^ {\\prime} + \\operatorname {A t t n} \\left(Q _ {\\mathrm {i m g}} ^ {\\prime}, \\left[ K _ {\\mathrm {i s t}} \\| K _ {\\mathrm {c o n}} \\right], \\left[ V _ {\\mathrm {i s t}} \\| V _ {\\mathrm {c o n}} \\right]\\right).", + "image_path": "b3dea140081ede898488eaf1890cfa0451cef52bf76350b4afa99bccd1da9bc0.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 221, + 555, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 555, + 305 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 555, + 305 + ], + "type": "text", + "content": "In this setup, we introduce a new linear layer, " + }, + { + "bbox": [ + 313, + 221, + 555, + 305 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{cross}}^q" + }, + { + "bbox": [ + 313, + 221, + 555, + 305 + ], + "type": "text", + "content": ", designed to enhance the model's ability to attend to image-instruction features effectively. This additional layer has been shown to improve performance, as validated in our experiments. Overall, this cross-attention operation facilitates controlled image generation, leveraging desired attributes from both the conditional image and the task instruction." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 306, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 554, + 437 + ], + "type": "text", + "content": "Position Embedding. The positional information of conditional image features is crucial for pixel-level control tasks, as methods like ControlNet incorporate features through pixel-to-pixel addition. To improve pixel-level control precision, we integrate Rotary Position Embedding (RoPE) [50] into both the query and key features prior to cross-attention. RoPE provides relative positional encoding, ensuring that query and key features with closer pixel coordinates yield higher similarity scores, thereby improving spatial alignment in the generated output. It has been used in T2I [23] and text-to-video [57] models." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 437, + 554, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 504 + ], + "type": "text", + "content": "Specifically, we apply 1-D RoPE separately to the height and width dimensions. Given a feature vector " + }, + { + "bbox": [ + 313, + 437, + 554, + 504 + ], + "type": "inline_equation", + "content": "f \\in \\mathbb{R}^{|D|}" + }, + { + "bbox": [ + 313, + 437, + 554, + 504 + ], + "type": "text", + "content": " at position " + }, + { + "bbox": [ + 313, + 437, + 554, + 504 + ], + "type": "inline_equation", + "content": "(h,w)" + }, + { + "bbox": [ + 313, + 437, + 554, + 504 + ], + "type": "text", + "content": ", we split " + }, + { + "bbox": [ + 313, + 437, + 554, + 504 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 437, + 554, + 504 + ], + "type": "text", + "content": " into two halves: " + }, + { + "bbox": [ + 313, + 437, + 554, + 504 + ], + "type": "inline_equation", + "content": "f_h \\in \\mathbb{R}^{|D|/2}" + }, + { + "bbox": [ + 313, + 437, + 554, + 504 + ], + "type": "text", + "content": " for the height dimension and " + }, + { + "bbox": [ + 313, + 437, + 554, + 504 + ], + "type": "inline_equation", + "content": "f_w \\in \\mathbb{R}^{|D|/2}" + }, + { + "bbox": [ + 313, + 437, + 554, + 504 + ], + "type": "text", + "content": " for the width. For the height component " + }, + { + "bbox": [ + 313, + 437, + 554, + 504 + ], + "type": "inline_equation", + "content": "f_h" + }, + { + "bbox": [ + 313, + 437, + 554, + 504 + ], + "type": "text", + "content": ", the rotary matrix " + }, + { + "bbox": [ + 313, + 437, + 554, + 504 + ], + "type": "inline_equation", + "content": "R_h" + }, + { + "bbox": [ + 313, + 437, + 554, + 504 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 508, + 554, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 508, + 554, + 572 + ], + "spans": [ + { + "bbox": [ + 321, + 508, + 554, + 572 + ], + "type": "interline_equation", + "content": "R _ {h} = \\left[ \\begin{array}{c c c c c} \\cos (h \\theta_ {0}) & - \\sin (h \\theta_ {0}) & \\dots & 0 & 0 \\\\ \\sin (h \\theta_ {0}) & \\cos (h \\theta_ {0}) & \\dots & 0 & 0 \\\\ \\vdots & \\vdots & \\ddots & \\vdots & \\vdots \\\\ 0 & 0 & \\dots & \\cos (h \\theta_ {l}) & - \\sin (h \\theta_ {l}) \\\\ 0 & 0 & \\dots & \\sin (h \\theta_ {l}) & \\cos (h \\theta_ {l}) \\end{array} \\right], \\tag {6}", + "image_path": "43c3c72e8abd5d99845d31bb136106b4c2a63a96cc13bbe2549d8cff444cadf6.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 577, + 554, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 577, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 577, + 554, + 624 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 577, + 554, + 624 + ], + "type": "inline_equation", + "content": "\\theta_{i} = b^{-i / |D|}" + }, + { + "bbox": [ + 313, + 577, + 554, + 624 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 577, + 554, + 624 + ], + "type": "inline_equation", + "content": "b = 10000" + }, + { + "bbox": [ + 313, + 577, + 554, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 577, + 554, + 624 + ], + "type": "inline_equation", + "content": "l = |D| / 4 - 1" + }, + { + "bbox": [ + 313, + 577, + 554, + 624 + ], + "type": "text", + "content": ". The rotary matrix " + }, + { + "bbox": [ + 313, + 577, + 554, + 624 + ], + "type": "inline_equation", + "content": "R_{w}" + }, + { + "bbox": [ + 313, + 577, + 554, + 624 + ], + "type": "text", + "content": " for the width component " + }, + { + "bbox": [ + 313, + 577, + 554, + 624 + ], + "type": "inline_equation", + "content": "f_{w}" + }, + { + "bbox": [ + 313, + 577, + 554, + 624 + ], + "type": "text", + "content": " is defined analogously. The updated feature vector " + }, + { + "bbox": [ + 313, + 577, + 554, + 624 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 577, + 554, + 624 + ], + "type": "text", + "content": " is then computed as:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 392, + 625, + 553, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 625, + 553, + 639 + ], + "spans": [ + { + "bbox": [ + 392, + 625, + 553, + 639 + ], + "type": "interline_equation", + "content": "f = \\left[ R _ {h} f _ {h} \\| R _ {w} f _ {w} \\right]. \\tag {7}", + "image_path": "1b415bd1505e91013f38bf7a40a5715a3ede665478d536e891e308e3e6f766c1.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "text", + "content": "RoPE is applied to both query and key features in Equation 5 and Equation 4, with the position coordinates of " + }, + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{ins}}" + }, + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "text", + "content": " set to " + }, + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "inline_equation", + "content": "(0,0)" + }, + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "text", + "content": ". By embedding relative positional information in this way, RoPE enhances the model's precision in pixel-level control, while maintaining robustness and adaptability across various tasks." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7966" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 137, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 137, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 137, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "In this section, we conduct extensive experiments to evaluate the performance of our UNIC-Adapter in various conditional image generation tasks. Detailed ablation studies are also conducted to analyze the contribution of individual components within our framework. Additional experimental results are available in the Appendix." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 171, + 167, + 183 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 167, + 183 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 167, + 183 + ], + "type": "text", + "content": "4.1. Tasks and Datasets" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 189, + 295, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 189, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 189, + 295, + 237 + ], + "type": "text", + "content": "Our evaluation covers three major tasks: pixel-level spatial control, subject-driven image generation, and style-image-based T2I generation. Below, we briefly describe the datasets used for training each task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 238, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 296, + 392 + ], + "type": "text", + "content": "Pixel-level Spatial Control. For pixel-level spatial control, we employ the MultiGen-20M dataset [39], which includes 12 types of pixel-level control annotations, such as Canny, HED, Sketch, Depth, Normal, Skeleton, Bbox, Segmentation, Outpainting, Inpainting, Deblurring, and Colorization. The dataset contains 2.8 million images with 20 million control annotations, and we reserve 1,000 samples for qualitative evaluation. For segmentation map conditioning, we also use the ADE20k [63] training set, which includes approximately 20,000 samples. These annotations serve as conditional images, and we use GPT-4o [2] to generate 20 synonymous instructions for each control type, such as \"Generate an image from this edge map.\"" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 393, + 296, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 393, + 296, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 296, + 537 + ], + "type": "text", + "content": "Subject-driven Image Generation. For subject-driven image generation, we use three datasets: OpenImages v7 [22], GRIT [37], and OpenStory++ [59]. Following previous works [31, 34], we utilize Blip-2 [25] to generate captions for images without annotations and use spaCy‡ to extract subject tags. GroundingDINO [30] is used to obtain bounding boxes for each subject, and SAM [21] is employed to generate subject segmentation maps based on these bounding boxes. After processing and filtering, we obtain 2.1 million subject-image pairs. We also use GPT4o to generate 20 instructions, such as \"Generate an image from this dog image.\"" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 537, + 296, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 537, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 537, + 296, + 645 + ], + "type": "text", + "content": "Style-image-based Image Generation. For style-image-based generation, we use the WikiArt [47] and Style-Booth [14] datasets. For StyleBooth, we randomly select an image from the same style subset as the conditional image, and for WikiArt, we employ CLIP-I [40] similarity to retrieve a conditional image for each training image. These two datasets comprise approximately 90,000 images. Additionally, we use GPT-4o to generate 20 instructions, such as \"Generate an image based on this style image.\"" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 188, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 188, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 188, + 665 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 670, + 295, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 295, + 694 + ], + "type": "text", + "content": "Our UNIC-Adapter framework is based on the SD3 medium [12] model. Since instruction texts and conditional" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 71, + 559, + 216 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 559, + 216 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 559, + 216 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 559, + 216 + ], + "type": "table", + "html": "
MethodBase ModelCanny (F1 Score↑)HED (SSIM↑)Seg. (mIoU↑)Depth (RMSE↓)
Single-task
T2I-Adapter [33]SD1.423.65-12.6148.40
Gligen [27]SD1.426.940.563423.7838.83
ControlNet [60]SD1.534.650.762132.5535.90
ControlNet++ [26]SD1.537.040.809743.6428.32
Multi-task
UniControl [39]SD1.530.820.796925.4439.18
Uni-ControlNet [61]SD1.527.320.691019.3940.65
OmniGen [56]Phi-335.540.823744.2328.54
OursSD338.940.836942.8931.10
", + "image_path": "fcb0e0db968be7f8643dc0be244a4a052fc00ebe13b82f8dfad90c1c9bcb5816.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 323, + 281, + 545, + 415 + ], + "blocks": [ + { + "bbox": [ + 313, + 224, + 555, + 269 + ], + "lines": [ + { + "bbox": [ + 313, + 224, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 555, + 269 + ], + "type": "text", + "content": "Table 1. Results of different methods on four pixel-level control image generation tasks. " + }, + { + "bbox": [ + 313, + 224, + 555, + 269 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 313, + 224, + 555, + 269 + ], + "type": "text", + "content": " indicates that higher values are better, while " + }, + { + "bbox": [ + 313, + 224, + 555, + 269 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 313, + 224, + 555, + 269 + ], + "type": "text", + "content": " indicates that lower values are better. The best results for each metric are bolded, and the second-best results are underlined." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 281, + 545, + 415 + ], + "lines": [ + { + "bbox": [ + 323, + 281, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 323, + 281, + 545, + 415 + ], + "type": "table", + "html": "
MethodBase ModelDINO↑CLIP-I↑CLIP-T↑
Fine-Tuning
Textual Inversion [32]SD1.50.5690.7800.255
DreamBooth [45]SD1.50.6680.8030.305
BLIP-Diffusion [24]SD1.50.6700.8050.302
Test Time Tuning Free
Re-Imagen [8]-0.6000.7400.270
SuTI [9]Imagen0.7410.8190.304
Kosmos-G [34]SD1.50.6940.8470.287
OmniGen [56]Phi-30.8010.8470.301
OursSD30.8160.8410.306
", + "image_path": "f9f672eaa173e9be8fe73b574330d02e2378850def21b794a5a4d0237b6f0d3b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 423, + 555, + 447 + ], + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 447 + ], + "type": "text", + "content": "Table 2. Results of different methods on the DreamBench dataset [45] for subject-driven image generation." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 457, + 555, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 555, + 578 + ], + "type": "text", + "content": "are variations of text and image modalities, the UNIC-Adapter is initialized using the base T2I model's parameters, which minimizes learning difficulty. We freeze the parameters of base SD3 medium model and the feed-forward layer after the attention layer in adapter blocks, and train all other newly introduced parameters, resulting in approximately 1.2 billion trainable parameters. We use the AdamW optimizer [19] with a learning rate of 0.0001 and a weight decay of 0.01. Training is performed on 16 H100 GPUs for 100,000 steps, with a batch size of 16 per GPU." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 586, + 399, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 399, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 399, + 598 + ], + "type": "text", + "content": "4.3. Main Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 628 + ], + "type": "text", + "content": "We present experimental results of the proposed UNIC-Adapter framework across three major tasks." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "content": "Pixel-Level Spatial Control. Following previous work [26, 56], we evaluate performance based on the similarity between control conditions extracted from generated images and the specified pixel-level control inputs. The evaluation is conducted on the test set of MultiGen-20M with canny edge, hed edge, and depth map conditions and on ADE20k test set with segmentation map condition. As" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 186, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 186, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 186, + 713 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 66, + 702, + 186, + 713 + ], + "type": "text", + "content": "https://github.com/explosion/spaCy" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7967" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 553, + 426 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 553, + 426 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 553, + 426 + ], + "type": "image", + "image_path": "44180588d9e4f998bfb3816d16d28ac5f55735739f9c5ae3e1cf0f5058cb02d6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 434, + 555, + 456 + ], + "lines": [ + { + "bbox": [ + 55, + 434, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 555, + 456 + ], + "type": "text", + "content": "Figure 3. Visualization results of our UNIC-Adapter on twelve pixel-level control tasks from the MultiGen-20M dataset. The first and third rows show different types of conditional images, while the second and fourth rows display the corresponding generated images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 476, + 296, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 476, + 296, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 476, + 296, + 583 + ], + "type": "text", + "content": "shown in Table 1, UNIC-Adapter achieves the best performance on canny and hed edge map conditions, and demonstrates comparable performance on depth and segmentation map conditions, compared to SOTA methods like ControlNet++ [26] for single-task and OmniGen [56] for multitask scenarios. Figure 3 shows qualitative results on 12 tasks from MultiGen-20M, illustrating that UNIC-Adapter consistently produces images accurately aligned with pixel-level controls." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "Subject-Driven Image Generation. Evaluation is conducted using 30 objects and 25 prompts from the Dream-Bench dataset [45]. Following prior work [34], we generate 4 images for each prompt and measure subject fidelity using DINO [4] and CLIP-I scores and text fidelity with CLIP-T [40]. As shown in Table 2, UNIC-Adapter achieves high subject-image consistency and maintains strong alignment between the generated images and text prompts. Qualitative examples in Figure 4 demonstrate that UNIC-Adapter effectively generates images closely resembling input sub" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 477, + 432, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 477, + 432, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 432, + 488 + ], + "type": "text", + "content": "jects across various contexts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 490, + 555, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 490, + 555, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 555, + 562 + ], + "type": "text", + "content": "Style-image-based Image Generation. For the same text prompt, style images from StyleBooth [14] are randomly selected as conditional images. The qualitative results in Figure 5 indicate that UNIC-Adapter effectively captures and applies stylistic elements from conditional images, generating artistically consistent outputs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 573, + 416, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 416, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 416, + 585 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 592, + 555, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 555, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 555, + 663 + ], + "type": "text", + "content": "To better understand the impact of individual components in the UNIC-Adapter framework, we conduct ablation studies by systematically modifying or removing key modules. For efficiency, the models in this part are trained on the subject-driven image generation task and four tasks (canny, hed, depth, and segmentation) from MultiGen-20M." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "Query and Key Feature Choices. We investigate the effect of different query and key feature choices in Equation 5. First, we fix the query features as " + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{txt}}" + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{img}}" + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": " and examine the impact of different key features. As shown in Ta" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7968" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 74, + 294, + 265 + ], + "blocks": [ + { + "bbox": [ + 67, + 74, + 294, + 265 + ], + "lines": [ + { + "bbox": [ + 67, + 74, + 294, + 265 + ], + "spans": [ + { + "bbox": [ + 67, + 74, + 294, + 265 + ], + "type": "image", + "image_path": "56feff493f71225acad6e603a90f25da3628d8a72f7d5d6a5bea088652d213c2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 273, + 295, + 318 + ], + "lines": [ + { + "bbox": [ + 55, + 273, + 295, + 318 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 295, + 318 + ], + "type": "text", + "content": "Figure 4. Visualization results of our UNIC-Adapter on Dream-Bench for subject-driven generation. The first column displays the subject images, while the other three columns show the generated images based on different prompts." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 327, + 294, + 396 + ], + "blocks": [ + { + "bbox": [ + 58, + 327, + 294, + 396 + ], + "lines": [ + { + "bbox": [ + 58, + 327, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 58, + 327, + 294, + 396 + ], + "type": "image", + "image_path": "30155aee3240742cf5fdeb2e2da5e74e5152b3d2eb5e1a215cedd9d4587321c8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 112, + 396, + 238, + 404 + ], + "lines": [ + { + "bbox": [ + 112, + 396, + 238, + 404 + ], + "spans": [ + { + "bbox": [ + 112, + 396, + 238, + 404 + ], + "type": "text", + "content": "A cat with a mountain in the background." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 404, + 294, + 462 + ], + "blocks": [ + { + "bbox": [ + 58, + 404, + 294, + 462 + ], + "lines": [ + { + "bbox": [ + 58, + 404, + 294, + 462 + ], + "spans": [ + { + "bbox": [ + 58, + 404, + 294, + 462 + ], + "type": "image", + "image_path": "381c90fb07ae86a9c3e66ea29d6a429b886e232bc703740197e620caf983f3eb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 472, + 294, + 530 + ], + "blocks": [ + { + "bbox": [ + 120, + 463, + 230, + 472 + ], + "lines": [ + { + "bbox": [ + 120, + 463, + 230, + 472 + ], + "spans": [ + { + "bbox": [ + 120, + 463, + 230, + 472 + ], + "type": "text", + "content": "A girl with a city in the background." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 472, + 294, + 530 + ], + "lines": [ + { + "bbox": [ + 58, + 472, + 294, + 530 + ], + "spans": [ + { + "bbox": [ + 58, + 472, + 294, + 530 + ], + "type": "image", + "image_path": "0aa83a366708b9203717c54e7407acf0f9420c26e7f232f450bd03049deb9e65.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 118, + 531, + 233, + 539 + ], + "lines": [ + { + "bbox": [ + 118, + 531, + 233, + 539 + ], + "spans": [ + { + "bbox": [ + 118, + 531, + 233, + 539 + ], + "type": "text", + "content": "A robot in autumn mountain and lake." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 540, + 294, + 599 + ], + "blocks": [ + { + "bbox": [ + 58, + 540, + 294, + 599 + ], + "lines": [ + { + "bbox": [ + 58, + 540, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 58, + 540, + 294, + 599 + ], + "type": "image", + "image_path": "0835dd966396b4cefedc048427e38b9ae258f4674213cb2ba47183798da879e6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 607, + 295, + 652 + ], + "lines": [ + { + "bbox": [ + 55, + 607, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 607, + 295, + 652 + ], + "type": "text", + "content": "Figure 5. Visualization results of our UNIC-Adapter on style-image-based T2I generation. The first row shows the reference style image, and each subsequent row contains images generated from the same prompt, influenced by different style images." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "ble 3, " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{con}}" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " plays a critical role in pixel-level spatial control tasks and in maintaining subject-image similarity. Relying solely on " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{ist}}" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " as key feature (Exp. 3) significantly reduces performance across pixel-level control tasks and subject-" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 73, + 553, + 255 + ], + "blocks": [ + { + "bbox": [ + 317, + 73, + 553, + 255 + ], + "lines": [ + { + "bbox": [ + 317, + 73, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 317, + 73, + 553, + 255 + ], + "type": "image", + "image_path": "fd90d90f86770d9a0b92957506e7569ded4c1bd5a96777057644dec4b6b99833.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 263, + 555, + 329 + ], + "lines": [ + { + "bbox": [ + 313, + 263, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 555, + 329 + ], + "type": "text", + "content": "Figure 6. Comparison of visualization results of our UNIC-Adapter with and without position embedding, and IPAdapter-Instruct* on pixel-level control generation using canny edge conditional images. UNIC-Adapter generates images that are more closely aligned with the conditional images compared to the other two methods." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 321, + 338, + 553, + 440 + ], + "blocks": [ + { + "bbox": [ + 321, + 338, + 553, + 440 + ], + "lines": [ + { + "bbox": [ + 321, + 338, + 553, + 440 + ], + "spans": [ + { + "bbox": [ + 321, + 338, + 553, + 440 + ], + "type": "image", + "image_path": "fbe71ec62ee4ae7c880b20f5ad85aa49b79d7893a7f4f02f4cf4a69c6a2b0021.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 449, + 555, + 493 + ], + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 493 + ], + "type": "text", + "content": "Figure 7. Comparison of visualization results between our UNIC-Adapter and ControlNet* on subject-driven generation. Compared to ControlNet*, UNIC-Adapter achieves better consistency between subject images and generated images." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "text", + "content": "image similarity. We also investigate the impact of query features and find that using " + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{txt}}" + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "text", + "content": " alone as the query feature (Exp. 5) also degrades performance. In general, using " + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{img}}" + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "text", + "content": " as the query and both " + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{con}}" + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{ist}}" + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "text", + "content": " as key features (Exp. 4) yields optimal performance on pixel-level control tasks and comparable performance on subject-driven generation tasks. This highlights the importance of task instruction features in unified controllable generation. We adopt this configuration for subsequent experiments." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "content": "Importance of Position Embedding. To enhance the control precision of pixel-level tasks, we experiment with incorporating different types of position embeddings, including the absolute position embedding in SD3 and the RoPE used in our adapter, to query and key features of each layer. As shown in Table 4, RoPE yields better performance than absolute embeddings. We further examine the impact of introducing a new linear layer, " + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{cross}}^q" + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "content": ", for query image" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7969" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 70, + 526, + 178 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 526, + 178 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 526, + 178 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 526, + 178 + ], + "type": "table", + "html": "
Exp.NumberQueryQimgQtxtKeyKconKistCanny(F1 Score↑)HED(SSIM↑)Seg.(mIoU↑)Depth(RMSE↓)Subject-Driven
DINO↑CLIP-I↑CLIP-T↑
1✓✓✓✓29.980.784030.8434.050.7780.8270.309
2✓✓✓29.010.776729.0035.390.7690.8250.307
3✓✓✓22.380.559922.9638.820.6940.7800.319
4✓✓✓31.320.793430.9433.160.7690.8230.308
5✓✓✓19.100.373017.0646.950.6690.7820.317
6✓✓29.240.778627.8134.120.7500.8170.313
", + "image_path": "12b6c1ef75e83078802b4453324ac5e44772a56d29ef566dff0bd930cd00a86f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 89, + 210, + 520, + 360 + ], + "blocks": [ + { + "bbox": [ + 85, + 188, + 523, + 200 + ], + "lines": [ + { + "bbox": [ + 85, + 188, + 523, + 200 + ], + "spans": [ + { + "bbox": [ + 85, + 188, + 523, + 200 + ], + "type": "text", + "content": "Table 3. Results of different query and key feature choices on pixel-level control tasks and subject-driven generation task." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 89, + 210, + 520, + 360 + ], + "lines": [ + { + "bbox": [ + 89, + 210, + 520, + 360 + ], + "spans": [ + { + "bbox": [ + 89, + 210, + 520, + 360 + ], + "type": "table", + "html": "
MethodCanny(F1 Score↑)HED(SSIM↑)Seg.(mIoU↑)Depth(RMSE↓)Subject-Driven
DINO↑CLIP-I↑CLIP-T↑
w/o PE31.320.793430.9433.160.7690.8230.308
w/ Abs PE31.510.796730.3132.980.7400.8090.314
w/ Abs PE + Lqcross36.930.834833.2232.740.7510.8110.314
w/ RoPE34.650.824832.1333.450.7610.8170.313
w/ RoPE + Lqcross37.950.842033.3232.250.7840.8290.309
w/o txt36.310.838733.7131.600.7880.8300.308
w/o txt update36.310.838733.7131.600.7720.8210.310
ControlNet*38.180.838234.8131.460.7490.8150.313
IPAdapter-Instruct*15.810.219116.1247.810.7880.8520.294
", + "image_path": "42f15238d4f9cf04507df71124a26e4baf1b6439f0292a2038867dee18976b52.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 369, + 553, + 391 + ], + "lines": [ + { + "bbox": [ + 55, + 369, + 553, + 391 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 553, + 391 + ], + "type": "text", + "content": "Table 4. Results of two baseline methods and different module combinations of our UNIC-Adapter on pixel-level control tasks and subject-driven generation task." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 406, + 294, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 406, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 54, + 406, + 294, + 502 + ], + "type": "text", + "content": "features in the main branch. Table 4 shows that the addition of " + }, + { + "bbox": [ + 54, + 406, + 294, + 502 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{cross}}^q" + }, + { + "bbox": [ + 54, + 406, + 294, + 502 + ], + "type": "text", + "content": " improves results further, particularly in pixel-level control tasks and subject-image similarity. This indicates the importance of " + }, + { + "bbox": [ + 54, + 406, + 294, + 502 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{cross}}^q" + }, + { + "bbox": [ + 54, + 406, + 294, + 502 + ], + "type": "text", + "content": " in enabling query features " + }, + { + "bbox": [ + 54, + 406, + 294, + 502 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{img}}" + }, + { + "bbox": [ + 54, + 406, + 294, + 502 + ], + "type": "text", + "content": " to retrieve specific features from conditional image and task instruction. Figure 6 shows that UNIC-Adapter without position embedding often generates images with incomplete or incorrect details from control signals." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 502, + 295, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 295, + 694 + ], + "type": "text", + "content": "Comparison with Baseline Methods. We reproduce two baseline methods, ControlNet* and IPAdapter-Instruct*, in our framework. For ControlNet*, we replace the cross-attention mechanism in the image-instruction feature injection module with an addition operation (as did in ControlNet) while retaining other settings. For IPAdapter-Instruct*, we reproduce results by referring to the open-source code on our datasets. As shown in Table 4, ControlNet* achieves performance comparable to our final model on pixel-level control tasks but exhibits suboptimal subject-image similarity on the subject-driven generation task. The visualization results in Figure 7 show that ControlNet* often produces images with distorted subject content. In contrast, IPAdapter-Instruct* performs well on subject-driven generation tasks but poorly on pixel-level control tasks. Due to compressing conditional images into 16 tokens with" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 407, + 554, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 554, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 554, + 479 + ], + "type": "text", + "content": "out precise spatial information, the generated images from IPAdapter-Instruct* in Figure 6 fail to accurately capture the pixel-level details of the conditioned images. In comparison, our UNIC-Adapter achieves consistently superior results across both task types, demonstrating its effectiveness and suitability for unified controllable image generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 488, + 388, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 488, + 388, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 388, + 502 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 509, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 629 + ], + "type": "text", + "content": "In this paper, we present UNIC-Adapter, a unified framework for controllable T2I generation that integrates diverse conditional image inputs and task instructions within a single model. Built on the MM-DiT architecture, our approach leverages cross-attention mechanisms and RoPE to achieve precise pixel-level control, as well as high subject fidelity, across multiple generation tasks. Our extensive experiments on tasks such as pixel-level spatial control, subject-driven image generation, and style-image-based T2I generation demonstrate the effectiveness of the UNIC-Adapter." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 639, + 417, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 417, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 417, + 653 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 659, + 553, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 553, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 553, + 683 + ], + "type": "text", + "content": "This research is supported by the NSFC Grants under the contracts No. 62325111 and No.U22B2011." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 243, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 243, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 243, + 714 + ], + "type": "inline_equation", + "content": "\\S" + }, + { + "bbox": [ + 66, + 702, + 243, + 714 + ], + "type": "text", + "content": " https://github.com/unity-research/IP-Adapter-Instruct" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7970" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 146 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 146 + ], + "type": "text", + "content": "[1] Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 295, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 295, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 295, + 201 + ], + "type": "text", + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. GPT-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 204, + 295, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 204, + 295, + 247 + ], + "spans": [ + { + "bbox": [ + 61, + 204, + 295, + 247 + ], + "type": "text", + "content": "[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science, 2(3):8, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 249, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 249, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 61, + 249, + 295, + 293 + ], + "type": "text", + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 295, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 295, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 61, + 295, + 295, + 350 + ], + "type": "text", + "content": "[5] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 1, 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 351, + 295, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 351, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 61, + 351, + 295, + 407 + ], + "type": "text", + "content": "[6] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- " + }, + { + "bbox": [ + 61, + 351, + 295, + 407 + ], + "type": "inline_equation", + "content": "\\backslash" + }, + { + "bbox": [ + 61, + 351, + 295, + 407 + ], + "type": "text", + "content": " sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 408, + 295, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 408, + 295, + 461 + ], + "spans": [ + { + "bbox": [ + 62, + 408, + 295, + 461 + ], + "type": "text", + "content": "[7] Junsong Chen, Yue Wu, Simian Luo, Enze Xie, Sayak Paul, Ping Luo, Hang Zhao, and Zhenguo Li. Pixart- " + }, + { + "bbox": [ + 62, + 408, + 295, + 461 + ], + "type": "inline_equation", + "content": "\\{\\backslash" + }, + { + "bbox": [ + 62, + 408, + 295, + 461 + ], + "type": "text", + "content": " delta\\} Fast and controllable image generation with latent consistency models. arXiv preprint arXiv:2401.05252, 2024. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 464, + 295, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 295, + 497 + ], + "type": "text", + "content": "[8] Wenhu Chen, Hexiang Hu, Chitwan Sahara, and William W Cohen. Re-Imagen: Retrieval-augmented text-to-image generator. arXiv preprint arXiv:2209.14491, 2022. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 498, + 295, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 498, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 62, + 498, + 295, + 542 + ], + "type": "text", + "content": "[9] Wenhu Chen, Hexiang Hu, Yandong Li, Nataniel Ruiz, Xuhui Jia, Ming-Wei Chang, and William W Cohen. Subject-driven text-to-image generation via apprenticeship learning. NeurIPS, 36, 2024. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 544, + 295, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 295, + 599 + ], + "type": "text", + "content": "[10] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, pages 2818-2829, 2023. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 600, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 600, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 600, + 295, + 633 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "text", + "content": "[12] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Int. Conf. Mach. and Learn., 2024. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[13] Junyao Gao, Yanchen Liu, Yanan Sun, Yinhao Tang, Yanhong Zeng, Kai Chen, and Cairong Zhao. StyleShot: A snap" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "shot on any style. arXiv preprint arXiv:2407.01414, 2024. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 95, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 139 + ], + "type": "text", + "content": "[14] Zhen Han, Chaojie Mao, Zeyinzi Jiang, Yulin Pan, and Jingfeng Zhang. StyleBooth: Image style editing with multimodal instruction. arXiv preprint arXiv:2404.12154, 2024. 2, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "type": "text", + "content": "[15] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In CVPR, pages 4775–4785, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 173, + 553, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 204 + ], + "type": "text", + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 33:6840-6851, 2020. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 205, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 205, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 205, + 553, + 259 + ], + "type": "text", + "content": "[17] Hexiang Hu, Kelvin CK Chan, Yu-Chuan Su, Wenhu Chen, Yandong Li, Kihyuk Sohn, Yang Zhao, Xue Ben, Boqing Gong, William Cohen, et al. Instruct-Imagen: Image generation with multi-modal instruction. In CVPR, pages 4754–4763, 2024. 2, 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "type": "text", + "content": "[18] Lianghua Huang, Di Chen, Yu Liu, Yujun Shen, Deli Zhao, and Jingren Zhou.Composer: Creative and controllable image synthesis with composable conditions. arXiv preprint arXiv:2302.09778, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 305, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 326 + ], + "type": "text", + "content": "[19] Frank Hutter Ilya Loshchilov. Decoupled weight decay regularization. *ICLR*, 2019. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 327, + 487, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 487, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 487, + 338 + ], + "type": "text", + "content": "[20] Imagen-Team-Google. Imagen 3, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 339, + 553, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 339, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 316, + 339, + 553, + 382 + ], + "type": "text", + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 383, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 553, + 448 + ], + "type": "text", + "content": "[22] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 128(7):1956-1981, 2020. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 449, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 481 + ], + "type": "text", + "content": "[23] Black Forest Labs. Flux: Inference repository. https://github.com/black-forest-labs/flux, 2024. Accessed: 2024-11-14. 1, 2, 3, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 482, + 553, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 515 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 515 + ], + "type": "text", + "content": "[24] Dongxu Li, Junnan Li, and Steven Hoi. BLIP-Diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. NeurIPS, 36, 2024. 2, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 516, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 516, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 516, + 553, + 568 + ], + "type": "text", + "content": "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Int. Conf. Mach. and Learn., pages 19730-19742. PMLR, 2023. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 570, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 553, + 624 + ], + "type": "text", + "content": "[26] Ming Li, Taojiannan Yang, Huafeng Kuang, Jie Wu, Zhaoning Wang, Xuefeng Xiao, and Chen Chen. ControlNet++: Improving conditional controls with efficient consistency feedback. In ECCV, pages 129-147. Springer, 2025. 2, 5, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "text", + "content": "[27] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. GLIGEN: Open-set grounded text-to-image generation. In CVPR, pages 22511-22521, 2023. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[28] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et al. Hunyuan-DiT: A powerful multi-resolution diffusion transformer with fine-grained" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7971" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "chinese understanding. arXiv preprint arXiv:2405.08748, 2024.1,2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "text", + "content": "[29] Bingchen Liu, Ehsan Akhgari, Alexander Visheratin, Aleks Kamko, Linmiao Xu, Shivam Shrirao, Joao Souza, Suhail Doshi, and Daiqing Li. Playground v3: Improving text-to-image alignment with deep-fusion large language models. arXiv preprint arXiv:2409.10695, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "type": "text", + "content": "[30] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding DINO: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 252 + ], + "type": "text", + "content": "[31] Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. Subject-Diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. In ACM SIGGRAPH Conference Papers, pages 1-12, 2024. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 296 + ], + "type": "text", + "content": "[32] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, pages 6038–6047, 2023. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "type": "text", + "content": "[33] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, pages 4296-4304, 2024. 1, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 294, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 388 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 388 + ], + "type": "text", + "content": "[34] Xichen Pan, Li Dong, Shaohan Huang, Zhiliang Peng, Wenhu Chen, and Furu Wei. Kosmos-G: Generating images in context with multimodal large language models. arXiv preprint arXiv:2310.02992, 2023. 1, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 410 + ], + "type": "text", + "content": "[35] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 294, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 434 + ], + "type": "text", + "content": "[36] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195-4205, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 434, + 294, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 479 + ], + "type": "text", + "content": "[37] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 479, + 294, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 534 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 534 + ], + "type": "text", + "content": "[38] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 536, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 536, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 536, + 294, + 590 + ], + "type": "text", + "content": "[39] Can Qin, Shu Zhang, Ning Yu, Yihao Feng, Xinyi Yang, Yingbo Zhou, Huan Wang, Juan Carlos Niebles, Caiming Xiong, Silvio Savarese, et al. UniControl: A unified diffusion model for controllable visual generation in the wild. arXiv preprint arXiv:2305.11147, 2023. 2, 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 591, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 294, + 657 + ], + "type": "text", + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Int. Conf. Mach. and Learn., pages 8748-8763. PMLR, 2021. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[41] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 2, 3, 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "type": "text", + "content": "[43] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 234–241. Springer, 2015. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 173, + 553, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 226 + ], + "type": "text", + "content": "[44] Ciara Rowles, Shimon Vainer, Dante De Nigris, Slava Elizarov, Konstantin Kutsy, and Simon Donne. IPAdapter-Instruct: Resolving ambiguity in image-based conditioning using instruct prompts. arXiv preprint arXiv:2408.03209, 2024. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "text", + "content": "[45] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, pages 22500-22510, 2023. 1, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 272, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 336 + ], + "type": "text", + "content": "[46] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 35:36479-36494, 2022. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 338, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 371 + ], + "type": "text", + "content": "[47] Babak Saleh and Ahmed Elgammal. Large-scale classification of fine-art paintings: Learning the right metric on the right feature. arXiv preprint arXiv:1505.00855, 2015. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 371, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 436 + ], + "type": "text", + "content": "[48] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. LAION-5B: An open large-scale dataset for training next generation image-text models. NeurIPS, 35:25278-25294, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "type": "text", + "content": "[49] Kunpeng Song, Yizhe Zhu, Bingchen Liu, Qing Yan, Ahmed Elgammal, and Xiao Yang. Moma: Multimodal llm adapter for fast personalized image generation. arXiv preprint arXiv:2404.05674, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 481, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 481, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 481, + 553, + 525 + ], + "type": "text", + "content": "[50] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. RoFormer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 2, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 525, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 525, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 525, + 553, + 557 + ], + "type": "text", + "content": "[51] Kcolors Team. Kcolors: Effective training of diffusion model for photorealistic text-to-image synthesis. arXiv preprint, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 558, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 558, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 558, + 553, + 613 + ], + "type": "text", + "content": "[52] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 614, + 547, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 547, + 624 + ], + "type": "text", + "content": "[53] A Vaswani. Attention is all you need. NeurIPS, 2017. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "type": "text", + "content": "[54] Haofan Wang, Matteo Spinelli, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "text", + "content": "[55] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 1, 4" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7972" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 441 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[56] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. arXiv preprint arXiv:2409.11340, 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "type": "text", + "content": "[57] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. CogVideoX: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 216 + ], + "type": "text", + "content": "[58] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. IP-Adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 282 + ], + "type": "text", + "content": "[59] Zilyu Ye, Jinxiu Liu, Ruotian Peng, Jinjin Cao, Zhiyang Chen, Yiyang Zhang, Ziwei Xuan, Mingyuan Zhou, Xiaogian Shen, Mohamed Elhoseiny, et al. Openstory++: A large-scale dataset and benchmark for instance-aware open-domain visual storytelling. arXiv preprint arXiv:2408.03695, 2024. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 294, + 317 + ], + "type": "text", + "content": "[60] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 1, 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 361 + ], + "type": "text", + "content": "[61] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. UniControlNet: All-in-one control to text-to-image diffusion models. NeurIPS, 36, 2024. 2, 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 364, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 364, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 294, + 407 + ], + "type": "text", + "content": "[62] Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In CVPR, pages 22490-22499, 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 409, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 294, + 441 + ], + "type": "text", + "content": "[63] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, pages 633-641, 2017. 5" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7973" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/UNICL-SAM_ Uncertainty-Driven In-Context Segmentation with Part Prototype Discovery/fc453b95-4cba-4bb9-95fb-6c8d34e4ccaa_content_list.json b/2025/UNICL-SAM_ Uncertainty-Driven In-Context Segmentation with Part Prototype Discovery/fc453b95-4cba-4bb9-95fb-6c8d34e4ccaa_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..71729c2d2cef3868b019b237c38c030881205123 --- /dev/null +++ b/2025/UNICL-SAM_ Uncertainty-Driven In-Context Segmentation with Part Prototype Discovery/fc453b95-4cba-4bb9-95fb-6c8d34e4ccaa_content_list.json @@ -0,0 +1,1677 @@ +[ + { + "type": "text", + "text": "UNICL-SAM: Uncertainty-Driven In-Context Segmentation with Part Prototype Discovery", + "text_level": 1, + "bbox": [ + 196, + 130, + 803, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dianmo Sheng $^{1,4}$ , Dongdong Chen $^{2}$ , Zhentao Tan $^{1,4}$ , Qiankun Liu $^{3}$ , Qi Chu $^{1,4}$ , Tao Gong $^{1,4,\\dagger}$ , Bin Liu $^{1,4}$ , Jing Han $^{5}$ , Wenbin Tu $^{5}$ , Shengwei Xu $^{6}$ , Nenghai Yu $^{1,4}$ , $^{1}$ School of Cyber Science and Technology, USTC $^{2}$ Microsoft GenAI $^{3}$ USTB $^{4}$ Anhui Province Key Laboratory of Digital Security $^{5}$ BoCom Research $^{6}$ BESTI", + "bbox": [ + 143, + 202, + 872, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dmsheng@mail.ustc.edu.cn, tgong@ustc.edu.cn, xusw@besti.edu.cn", + "bbox": [ + 218, + 277, + 772, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in in-context segmentation generalists have demonstrated significant success in performing various image segmentation tasks using a limited number of labeled example images. However, real-world applications present challenges due to the variability of support examples, which often exhibit quality issues resulting from various sources and inaccurate labeling. How to extract more robust representations from these examples has always been one of the goals of in-context visual learning. In response, we propose UNICL-SAM, to better model the example distribution and extract robust representations to help in-context segmentation. We incorporate an uncertainty probabilistic module to quantify each example's reliability during both the training and testing phases. Utilizing this uncertainty estimation, we introduce an uncertainty-guided graph augmentation and feature refinement strategy, aimed at mitigating the impact of high-uncertainty regions to enhance the learning of robust representations. Subsequently, we construct prototypes for each example by aggregating part information, thereby creating reliable in-context instruction that effectively represents fine-grained local semantics. This approach serves as a valuable complement to traditional global pooling features. Experimental results demonstrate the effectiveness of the proposed framework, underscoring its potential for real-world applications.", + "bbox": [ + 89, + 358, + 485, + 736 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 762, + 220, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inspired by the great success of in-context learning [6, 32, 45] in the field of natural language processing (NLP), recent advances in vision generalist models [29, 30, 33, 43] indicate that different levels of segmentation tasks, including instance, semantic, and video segmentation, can be unified within one framework. The remarkable performance and", + "bbox": [ + 89, + 787, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "strong task generalization capabilities of these in-context segmentation models are undoubtedly promising for the research community.", + "bbox": [ + 511, + 328, + 906, + 373 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, existing in-context segmentation generalists often assume that the data in the training sets are of high quality and treat them equally within the same framework, thereby overlooking their diversity and differences. Previous works [47, 53] have also reported fluctuations in model performance based on varying examples and investigated strategies for selecting effective in-context examples. However, these studies only screened examples during the testing phase without considering optimization methods during the training phase. The above discussion motivates us to explore how to incorporate potential distributional diversity into the training phase to develop more robust in-context segmentation models.", + "bbox": [ + 511, + 380, + 908, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In light of this, we introduce UNICL-SAM, to investigate and address the modeling of example distribution differences and further feature optimization strategies within in-context segmentation. We draw inspiration from traditional Bayesian theory [3, 11, 21], which has previously addressed the assumptions related to data distribution differences and defined corresponding uncertainty measures to quantify and assess these distributional discrepancies. This framework facilitates the optimization of the model. Our intuition suggests that data uncertainty/discrepancy also exists within the context of contextual segmentation, potentially arising from varying observations in the reference examples. Thus, an uncertainty probabilistic Gaussian graph model is proposed to capture the training data distribution and estimate uncertainty for each example during the test phase. Additionally, we utilize the estimated uncertainty during training to guide the model to focus more on reliable (low-uncertainty) regions. The designed uncertainty-guided graph augmentation module enhances graph representation via learning invariant representations across different views of the graph. Further combined with an uncertainty-guided", + "bbox": [ + 511, + 583, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 114, + 886, + 245, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20201", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "feature reweighting strategy, the model yields more robust and reliable in-context indications while minimizing the influence of extraneous information. To fully harness the discriminating information delivered from in-context examples, we incorporate diverse visual instructions like previous works [33] and introduce part prototypes as a useful form of instructions to enhance in-context knowledge representation. In practical applications, the foreground regions under consideration inherently contain local semantics, which are often overlooked by simplistic global pooling methods. By effectively aggregating similar local features and enhancing the differentiation among prototypes, part prototypes can more accurately capture fine-grained semantic information. Through a well-designed framework and comprehensive experiments, our model demonstrates great performance on both clean and degraded data, effectively alleviating the impact of the aforementioned challenge. We summarize our main contribution as follows:", + "bbox": [ + 89, + 90, + 480, + 361 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose an uncertainty probabilistic module to characterize the distribution of input data. We further leverage uncertainty estimation to guide the model in learning more reliable features and refining visual instructions throughout both the training and testing phases.", + "- We introduce part prototypes as a form of visual instruction designed to effectively extract knowledge from reference examples by aggregating local semantics.", + "- Our comprehensive experiments reveal that UNICL-SAM exhibits superior performance on traditional benchmarks and shows good generalization capabilities on conducted real-world simulation robustness tests." + ], + "bbox": [ + 89, + 363, + 483, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 558, + 240, + 573 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In-Context Visual Learning. Recently, applying the in-context learning paradigm to the field of computer vision has become one of the hot spots in the research community. Bar et al. [1] pioneers in-context visual learning framework utilizing image inpainting on illustrations and infographics derived from vision-related literature, which shows competencies in fundamental CV tasks. Painter [42] employs masked image modeling on continuous pixels to facilitate in-context learning and yields highly competitive outcomes on seven tasks. Following this, SegGPT [43] offers a specialized generalist aimed at addressing diverse and unlimited segmentation tasks using a similar framework. SINE [30] building upon the SegGPT framework to tackle the task ambiguity problem. Other approaches [29, 52] investigated prompting techniques of powerful segmentation model SAM [24]. Moreover, recent work SEGIC [33] shows superior performance with a large margin compared with previous works by leveraging diverse in-context instructions.", + "bbox": [ + 89, + 583, + 482, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the aforementioned work does not take into account the potential impact of the diversity in training and", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "testing data distributions, which motivates us to develop a more robust in-context segmentation model.", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Uncertainty Estimation in computer vision. The issue of uncertainty in neural networks, which pertains to the lack of certainty estimates or under-confidence in predictions, has garnered increasing attention in recent research [7, 12, 13, 15, 20-22, 39, 48, 54]. Kendall et al. [21] delineate two primary types: epistemic uncertainty, arising from the model's limited knowledge, and aleatoric uncertainty, which is typically associated with inherent noise within the data. Nonbayesian approaches [7, 15, 44, 48] typically utilize a Gaussian distribution framework to estimate both the mean and variance of model parameters or input data, which serve as indicators of uncertainty. They effectively avoid the substantial computational overhead of the Bayesian solution. These techniques have been effectively applied to various computer vision tasks, including image classification [41, 54], semantic segmentation [2, 5, 8, 17], object detection [44, 48], person reidentification [50] and many other computer vision tasks [7, 15, 20].", + "bbox": [ + 511, + 127, + 903, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we first claim the diverse data distributions that in-context segmentation models need to navigate. We then propose an uncertainty probabilistic network and corresponding enhancement strategies for optimization.", + "bbox": [ + 511, + 398, + 903, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Robust Segmentation. Numerous studies [9, 10, 14, 37, 46, 49] have studied how to improve the performance of segmentation models in cross-domain and quality-degraded scenarios. For example, recently, RobustSAM [9] has examined the performance deterioration of the Segment Anything Model (SAM) [24] when handling low-quality images, subsequently designing a pipeline to enforce both feature and token consistency between clean and generated degraded images.", + "bbox": [ + 511, + 465, + 903, + 602 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building upon these insights, we propose uncertainty-guided graph augmentation and feature refinement techniques to obtain robust representations. The subsequent part prototype discovery stage provides discriminative local semantics, further enhancing the model's robustness.", + "bbox": [ + 511, + 603, + 903, + 678 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 691, + 604, + 707 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In-context segmentation aims to identify a specific segmentation task within the given $K$ in-context examples $ref = \\{(\\mathbf{x}_r,\\mathbf{y}_r)\\}_{i = 1}^K$ , where $\\mathbf{x}_r\\in \\mathbb{R}^{3\\times H\\times W}$ , $\\mathbf{y}_r\\in \\mathbb{R}^{1\\times H\\times W}$ are one reference image and its corresponding ground-truth label, and segment the interested area in the target image $\\mathbf{x}_t\\in \\mathbb{R}^{3\\times H\\times W}$ as a binary mask $\\mathbf{y}_t$ .", + "bbox": [ + 511, + 717, + 903, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 818, + 624, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our UNICL-SAM combines the advantages of previous work and further utilizes the proposed components to improve the robustness of the model. The overall pipeline is shown in Figure 1, which consists of four stages: fea-", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "footer", + "text": "20202", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c4225b7b9427a311790cabd61cfe181d7dce22ddabf2a0d582ec5b1faca177a6.jpg", + "image_caption": [ + "Figure 1. Overview of UNICL-SAM. With the fine-grained features, UNICL-SAM performs uncertainty quantification to model the input distribution and produce the estimated uncertainty map. The corresponding uncertainty-guided graph augmentation and feature refinement stages tend to learn robust graph representations by mitigating the impact of high uncertainty areas. Further, UNICL-SAM conducts part prototype discovery to get the part prototypes with the other three kinds of in-context instructions. Finally, the in-context instructions are fed into the query prompt generator to get the in-context prompt for guiding the mask decoder to predict segmentation results." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 903, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ture extraction, uncertainty probabilistic modeling, part prototype discovery, and final mask decoding. First, we extract and integrate multi-scale features, then compute dense correspondence between target and reference as previous works [29, 33] (Section 3.2). Then our uncertainty Gaussian graph model will provide quantified uncertainty estimation. Utilizing the uncertainty map, we further perform uncertainty-guided graph augmentation and feature refinement techniques to facilitate learning robust visual representations (Section 3.3). With the refined visual representations, part prototypes are generated as a form of in-context instruction to facilitate the extraction of more discriminative local semantics (Section 3.4). Finally, we generate incontext prompts through a prompt generator based on the rich instructions of reference and target, guiding the mask decoder model to obtain the final segmentation result (Section 3.5). We will elaborate on each stage in detail.", + "bbox": [ + 88, + 438, + 485, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Feature Extraction", + "text_level": 1, + "bbox": [ + 89, + 709, + 272, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Building upon insights from previous works [29, 33], we establish an in-context segmentation baseline by combining the advancements of existing VFMs [24, 36], thereby unleashing the power of pre-trained foundation models.", + "bbox": [ + 89, + 732, + 482, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use the frozen pre-trained image encoder to extract visual features of both the reference image $\\mathbf{x}_r$ and the target image $\\mathbf{x}_t$ , respectively. Inspired by previous works on dense prediction tasks like object detection [27] and semantic segmentation [25], we incorporate a multi-scale feature adapter to leverage information of feature maps at different scales and aggregate them to produce the final feature map", + "bbox": [ + 89, + 795, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathbf{F}_r \\in \\mathbb{R}^{C \\times H \\times W}$ and $\\mathbf{F}_t \\in \\mathbb{R}^{C \\times H \\times W}$ , where $C$ denotes the number of channels.", + "bbox": [ + 511, + 436, + 906, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then, we calculate the masked reference feature as $\\mathbf{F}_{rm} = \\mathbf{F}_r\\cdot \\mathbf{y}_r$ . The patch-level similarity $\\mathbf{Corr}\\in$ $\\mathbb{R}^{HW\\times HW}$ is calculated between $\\mathbf{F}_{rm}$ and $\\mathbf{F}_t$ as the dense correspondence. With the correspondence map, we calculate the pseudo mask $\\mathbf{y}_t^{pseudo}$ to extract the region of interest from the target feature like $\\mathbf{F}_{tm}^{pseudo} = \\mathbf{F}_t\\cdot \\mathbf{y}_t^{pseudo}$ . For details about the multi-scale feature adapter and correspondence extraction, please refer to the Appendix.", + "bbox": [ + 511, + 469, + 908, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Uncertainty Probabilistic Modeling", + "text_level": 1, + "bbox": [ + 511, + 604, + 821, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Deep neural network's output has varying degrees of uncertainty, influenced by how the training data is collected and the model's architecture, with observation noise affecting the instantiated predictions [21]. In the context of in-context segmentation, uncertainty also arises from the quality of the in-context examples. The limited number of samples in each iteration amplifies the impact of observation noise, further compounding uncertainty, which results in poor predictions. To address this, we model the data-dependent uncertainty of the input data and leverage it to more effectively optimize the network.", + "bbox": [ + 511, + 627, + 906, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Uncertainty quantification. We design our uncertainty Gaussian graph network (UGGN) to represent the graph representation using a Gaussian distribution. Given the fine-grained feature $\\mathbf{F}_m = \\{\\mathbf{F}_{rm},\\mathbf{F}_{tm}^{pseudo}\\}$ , we first construct a graph representation $\\mathcal{G} = \\{\\mathbf{V},\\mathbf{E}\\}$ , regarding each patch feature as each node in the graph $\\mathbf{V} = \\{\\mathbf{V}_i|i = 1,\\dots ,N\\}$ , $\\mathbf{V}_i\\in \\mathbf{F}_m$ . For the edge information, we uti", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20203", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/38e3801b3753d92f9d4c01b50831dab8e7667d907c57ba43ebbcf0fe964906c4.jpg", + "image_caption": [ + "Figure 2. Illustration of the uncertainty quantification and part prototype discovery stages. UGGN serves as a probabilistic graph network, with its initial layer dedicated to uncertainty quantification. Subsequent layers are designed to facilitate the generation of part prototypes." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 480, + 262 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "lize the affinity similarity as the edge weight from the $j^{th}$ node to the $k^{th}$ node is modeled by their affinity in the embedded space, which is formulated as $\\mathbf{E}(j,k) = (\\mathbf{V}_j)\\cdot (\\mathbf{V}_k)^T,\\mathbf{V}_j,\\mathbf{V}_k\\in \\mathbf{V}$ . We only retain edges with similarity greater than an established threshold $\\varepsilon$ and sparse the edge connections according to a certain top ratio $r$ as $\\mathbf{E} = r(\\{\\mathbf{E}(j,k)|\\mathbf{A}_{j,k}\\geqslant \\varepsilon \\})$ , where $\\mathbf{A}\\in \\mathbb{R}^{N\\times N}$ denotes the adjacency matrix.", + "bbox": [ + 89, + 369, + 483, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Figure 2 above, in the uncertainty quantification stage, what UGGN delivers the mean $\\mu_{\\mathcal{G}}$ and variance $\\sigma_{\\mathcal{G}}$ for each graph $\\mathcal{G}$ . Inspired by previous work [56], we first map the graph representation into a high-dimensional Gaussian space utilizing separate learnable mean and variance parameters, then update the mean and variance based on modified node aggregation. We use $\\mathbf{M}^{(l)} = [\\mu_i^{(l)},\\dots ,\\mu_N^{(l)}]\\in \\mathbb{R}^{N\\times d}$ and $\\pmb{\\Sigma}^{(l)} = [\\sigma_i^{(l)},\\dots ,\\sigma_N^{(l)}]\\in \\mathbb{R}^{N\\times d}$ to denote the matrix of means and variances for all nodes at layer $l\\in L$ of UGGN, respectively. $d$ represents the hidden dimension. For the first layer, we adopt a fully connected layer to map as follows:", + "bbox": [ + 89, + 492, + 483, + 679 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} ^ {(1)} = \\operatorname {E L U} \\left(\\mathbf {V W} _ {\\mu} ^ {(0)}\\right), \\boldsymbol {\\Sigma} ^ {(1)} = \\operatorname {R e L U} \\left(\\mathbf {V W} _ {\\sigma} ^ {(0)}\\right) \\quad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 686, + 482, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_{\\mu} \\in \\mathbb{R}^{N \\times d}$ and $\\mathbf{W}_{\\sigma} \\in \\mathbb{R}^{N \\times d}$ are learnable parameters, ELU and ReLU represent for activation functions. $h \\sim \\mathcal{N}(\\mu^{(1)}, \\sigma^{(1)})$ denote the parameterized distribution.", + "bbox": [ + 89, + 718, + 482, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, we need to sample from the estimated Gaussian distributions. However, sampling directly from the distribution is a non-differentiable operation, which will result in the model not being optimized. Referring to previous work [4], we decided to use the reparameterization trick. Specifically, we first draw a sample $\\epsilon$ from the standard Gaussian distribution $\\mathcal{N}(\\mathbf{0},\\mathbf{I})$ randomly, and then obtain the sample by computing $\\mathbf{Z} = \\mu^{(1)} + \\epsilon \\sigma^{(1)},\\mathbf{Z}\\in \\mathbb{R}^{N\\times C}$ . According to Bayesian probability theory [3, 21], we can con", + "bbox": [ + 88, + 765, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sider $\\mathbf{Z}$ as empirical samples drawn from an approximate predictive distribution. The uncertainty of the input data can be quantified by calculating channel-wise variance, expressed as $\\mathbf{U}_{data} = \\mathrm{Norm}(Var(\\mathbf{Z}))$ , $\\mathbf{U}_{data} \\in \\mathbb{R}^{N \\times 1}$ .", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Uncertainty-guided graph augmentation. With the estimated uncertainty, we can assume that a larger variance indicates a greater deviation from the expected data distribution. We utilize a softmax function to map the $\\mathbf{U}_{data}$ into a probability distribution $p_u^{node}$ , which determines the unimportance of each node in the graph. Combining with traditional graph augmentations like node dropping, and node masking, we can decide whether to drop or mask a node based on the its uncertainty.", + "bbox": [ + 511, + 152, + 905, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Furthermore, we compute the edge uncertainty using the affinity matrix of $\\mathbf{U}_{data}$ , and map to a probability distribution $p_u^{edge}$ as well. Intuitively, if the uncertainty at the edge's endpoints is higher, the uncertainty of the edge itself will be higher. This allows us to perform edge perturbation and subgraph transformation on the graph. With the augmented graph representation $\\mathcal{G}_{aug}$ and its vanilla view $\\mathcal{G}$ , we use the UGGN to map them into the same representation space, and minimize their KL divergence to mitigate the influence of high-uncertainty nodes, thereby learning a more robust graph representation:", + "bbox": [ + 511, + 287, + 906, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {K L} ^ {\\mathcal {G}} = \\sum_ {i = 1} ^ {N} K L (\\mathcal {N} \\left(\\mu_ {\\mathcal {G}}, \\sigma_ {\\mathcal {G}}\\right) | | \\mathcal {N} \\left(\\mu_ {\\mathcal {G} _ {a u g}}, \\sigma_ {\\mathcal {G} _ {a u g}}\\right)) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 465, + 906, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Uncertainty-guided feature refinement. On the other hand, the uncertainty estimation can be mapped back to spatial space as an uncertainty map $\\mathbf{U}_{map} \\in \\mathbf{R}^{H \\times W \\times 1}$ . We directly compute the dot product of reverse uncertainty map $\\mathbf{1} - \\mathbf{U}_{map}$ with the visual features $\\mathbf{F}$ and their associated labels $\\mathbf{y}$ for refinement, which will effectively alleviate the impact of high uncertainty positions, resulting in more accurate foreground representations.", + "bbox": [ + 511, + 518, + 905, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the uncertainty estimation may be less reliable in the early stages of training, we introduce an uncertainty gate to adjust the weights for uncertainty refinement dynamically. This gate is initialized with a small positive scalar to mitigate any initial unreliability.", + "bbox": [ + 511, + 640, + 905, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Part prototype Discovery", + "text_level": 1, + "bbox": [ + 511, + 726, + 743, + 743 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite we obtain a refined representation, the foreground region of example retains local semantics, resulting in varying degrees of correspondence with the target across different areas. To extract more discriminative representations, we propose the implementation of part prototype discovery.", + "bbox": [ + 511, + 750, + 905, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 2 below, we can treat the last layers of UGGN as an extension by converting a deterministic graph clustering network to a probabilistic type, i.e., a parameterized distribution for feature clustering. Then, we propose an attention mechanism for the next layers to", + "bbox": [ + 511, + 825, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20204", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "assign different weights to neighbors based on their variances since larger variances indicate more uncertainties in the latent representations. A smooth exponential function is utilized to control the effect of variances on weights $\\mathcal{A}^{(l)} = \\exp (-\\gamma \\Sigma^{(l)})$", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} ^ {(l + 1)} = \\tilde {\\mathbf {D}} ^ {- \\frac {1}{2}} \\mathbf {A} \\tilde {\\mathbf {D}} ^ {- \\frac {1}{2}} \\left(\\mathbf {M} ^ {(l)} \\odot \\mathcal {A} ^ {(l)}\\right) \\mathbf {W} _ {\\mu} ^ {(l)} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 172, + 482, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} ^ {(l + 1)} = \\tilde {\\mathbf {D}} ^ {- \\frac {1}{2}} \\mathbf {A} \\tilde {\\mathbf {D}} ^ {- \\frac {1}{2}} \\left(\\boldsymbol {\\Sigma} ^ {(l)} \\odot \\mathcal {A} ^ {(l)} \\odot \\mathcal {A} ^ {(l)}\\right) \\mathbf {W} _ {\\sigma} ^ {(l)} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 198, + 482, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{D}$ is the diagonal degree matrix calculated by $\\mathbf{D}_{i,i} = \\sum_{j}\\mathbf{A}_{i,j}$ , and $\\tilde{\\mathbf{D}} = \\mathbf{D} + \\mathbf{I}_N$ , where $\\mathbf{I}_N$ is the identity matrix.", + "bbox": [ + 89, + 226, + 482, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also sampled the updated distribution using the reparameterization trick to obtain the final node classification results. To solve the non-differentiability caused by the arg max operation to derive classification labels, we utilize the Gumbel-Softmax trick [19] to get the final one-hot like clusters $\\mathbf{C}\\mathbf{u} \\in \\mathbb{R}^{N \\times c}$ as follows:", + "bbox": [ + 89, + 257, + 483, + 345 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C l u} _ {i} = \\frac {\\exp \\left(\\left(\\log \\left(\\pi_ {i}\\right) + g _ {i}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {c} \\exp \\left(\\left(\\log \\left(\\pi_ {j}\\right) + g _ {j}\\right) / \\tau\\right)}, \\quad \\text {f o r} i = 1, \\dots , c. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 349, + 482, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $g_{i},\\dots ,g_{c}$ are i.i.d samples drawn from the Gumbel(0,1) distribution, $\\pi$ is a categorical distribution with class probabilities and $\\tau$ is the softmax temperature.", + "bbox": [ + 89, + 398, + 482, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the clusters $\\mathbf{C}\\mathbf{u}$ , we get part prototype representations $\\mathbf{P} = \\{\\mathbf{P}_i\\in \\mathbb{R}^C\\mid i = 1,\\dots ,c\\}$ by averaging $n$ node features in the same cluster:", + "bbox": [ + 89, + 443, + 483, + 487 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {i} = \\frac {1}{n} \\mathbf {C l u} _ {i} ^ {\\top} \\cdot \\mathbf {F}, \\quad \\mathbf {F} \\in \\mathbf {F} _ {m} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 489, + 482, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Mask Decoding", + "text_level": 1, + "bbox": [ + 89, + 523, + 246, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further interact with the mask decoder for mask prediction, we extracted four kinds of visual instructions for in-context prompt generation: visual features $\\mathbf{F}$ , mask average pooling features $\\mathbf{F}_p = \\mathrm{Avg\\_pool}(\\mathbf{F})$ , $\\mathbf{F} \\in \\mathbf{F}_m$ , part prototypes $\\mathbf{P}$ , and the mask label $\\mathbf{y}$ as depicted in Figure 1. We denote the $\\mathbf{I} = \\{\\mathbf{F}, \\mathbf{F}_p, \\mathbf{P}, \\mathbf{y}\\}$ as the set of in-context instructions derived from reference samples and the target image. A query-based prompt generator (QPG) is implemented to support a scalable structure and the ability to support different numbers of examples flexibly. Considering computational efficiency, we first concat them in channel dimension and project them into the latent space used by the QPG, then feed the reference instruction $\\mathbf{I}_r$ and target instruction $\\mathbf{I}_t$ to interact with learnable query tokens $\\mathbf{q}$ to get the final in-context prompt $\\mathbf{q}_{ict}$ :", + "bbox": [ + 89, + 546, + 483, + 773 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} = \\operatorname {P r o j} (\\operatorname {C o n c a t} (\\mathbf {F}, \\mathbf {F} _ {p}, \\mathbf {P}, \\mathbf {y})) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 777, + 482, + 795 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {q} _ {i c l} = Q P G \\left(\\mathbf {I} _ {r}, \\mathbf {I} _ {t}, \\mathbf {q}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 797, + 482, + 814 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we feed the target image feature $\\mathbf{F}_t$ with the incontext prompt $\\mathbf{q}_{icl}$ to the frozen mask decoder $D$ for prediction $\\mathbf{y}_{pred}$ as follows:", + "bbox": [ + 89, + 819, + 482, + 867 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} _ {p r e d} = D \\left(\\mathbf {F} _ {t}, \\mathbf {q} _ {i c l}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 873, + 482, + 891 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Loss function", + "text_level": 1, + "bbox": [ + 513, + 90, + 653, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We employ a linear combination of Binary Cross-Entropy (BCE) loss and Dice loss to supervise the mask prediction:", + "bbox": [ + 511, + 112, + 905, + 143 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {b c e} \\mathcal {L} _ {b c e} + \\lambda_ {d i c e} \\mathcal {L} _ {d i c e} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 155, + 905, + 171 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For better optimization, we also adopt some auxiliary losses. To ensure that the learned representations are indeed Gaussian distributions, we use an explicit regularization to constrain the latent representations in the first layer of UGGN as follows:", + "bbox": [ + 511, + 183, + 906, + 258 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {K L} ^ {r e g} = \\sum_ {i = 1} ^ {N} K L (\\mathcal {N} (\\mu_ {\\mathcal {G}}, \\sigma_ {\\mathcal {G}}) | | \\mathcal {N} (\\mathbf {0}, \\mathbf {I})) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 267, + 905, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the original GCN paper [23], we also impose $\\mathcal{L}_2$ regularization on parameters of the first layer of UGGN:", + "bbox": [ + 511, + 319, + 905, + 349 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {2} ^ {r e g} = \\left\\| \\mathbf {W} _ {\\mu} ^ {(0)} \\right\\| _ {2} ^ {2} + \\left\\| \\mathbf {W} _ {\\sigma} ^ {(0)} \\right\\| _ {2} ^ {2} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 359, + 903, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We find it helpful to use an auxiliary loss function in the uncertainty-guided part prototype discovery step during training to push prototypes away from each other:", + "bbox": [ + 511, + 390, + 905, + 436 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {P} = \\sum_ {\\mathbf {P} _ {i}, \\mathbf {P} _ {j} \\in \\mathbf {P}} \\max ((d - d i s t (\\mathbf {P} _ {i}, \\mathbf {P} _ {j})), 0) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 448, + 903, + 481 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Combing with the graph augmentation loss in Section 3.3, the final loss is the combination of the above loss:", + "bbox": [ + 511, + 492, + 905, + 522 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {K L} ^ {\\text {r e g}} \\mathcal {L} _ {K L} ^ {\\text {r e g}} + \\lambda_ {2} ^ {\\text {r e g}} \\mathcal {L} _ {2} ^ {\\text {r e g}} + \\lambda_ {P} \\mathcal {L} _ {P} + \\lambda_ {K L} ^ {\\mathcal {G}} \\mathcal {L} _ {K L} ^ {\\mathcal {G}} \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 547, + 905, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 594, + 645, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets.", + "text_level": 1, + "bbox": [ + 513, + 619, + 619, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train UNICL-SAM with mixed semantic segmentation datasets. Notably, our approach does not necessitate any modifications to the architecture or training pipeline when incorporating an additional dataset.", + "bbox": [ + 511, + 641, + 905, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "COCO [28] is a widely utilized dataset supporting instance segmentation, semantic segmentation, and panoptic segmentation tasks. It comprises 118K training images and 5k validation images, categorized into 80 categories.", + "bbox": [ + 511, + 702, + 905, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ADE20k [55] is a widely used benchmark for semantic segmentation, which consists of over 25K images sourced from diverse environments with 150 semantic categories.", + "bbox": [ + 511, + 763, + 905, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details.", + "text_level": 1, + "bbox": [ + 511, + 816, + 733, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We implement UNICL-SAM in PyTorch and use 8 Nvidia A6000 GPUs for our experiments. A total batch size of 64 was employed, with the model trained for 20 epochs. We merge the COCO instances for semantic segmentation and", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20205", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b81e54b9b035136037aa1a700cd713e605cf04e42b6766b8f92ec5939be792c6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsResolutionTrainable ParamsCOCO-20iFSS-1000LVIS-92i
one-shotfew-shotone-shotfew-shotone-shotfew-shot
specialist model
HSNet [34]48028M41.7*50.7*86.5*88.5*17.422.9
VAT [18]41752M42.9*49.4*90.3*90.8*18.522.7
FPTrans [51]480139M56.5*65.5*----
generalist model
Painter [42]448354M33.132.661.762.310.510.9
SegGPT [43]448354M56.167.985.689.318.625.4
PerSAM [52]-2M23.5-75.6-18.4-
Matcher [29]896-52.760.787.089.633.040.0
VRP-SAM [40]5121.6M53.9-----
ICVU [38]256309M58.0-----
SINE [30]89619M64.566.1--31.235.5
UNICL-SAM51855M77.878.784.086.334.137.4
", + "bbox": [ + 174, + 88, + 823, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "conduct the sample pools by category. We utilize DINOv2 (ViT-L/14) [36] as the default vision encoder. For the segmenter, we choose SAM (ViT-H) [24], which is pre-trained with 1B masks and 11M images, shows impressive zero-shot segmentation performance. During training, we froze the vision encoder and the segmenter, optimizing only the remaining parameters. We use AdamW [31] optimizer with a weight decay of $5e - 2$ and $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.95$ for optimization. We use a base learning rate of $1e - 4$ and a warmup of 4 epochs during training. Each image is cropped and resized to $518\\times 518$ resolution. We randomly sampled one reference pair during training. More implementation details are provided in the Appendix.", + "bbox": [ + 88, + 404, + 485, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Main Results", + "text_level": 1, + "bbox": [ + 89, + 617, + 232, + 632 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the main experiments, we compare UNICL-SAM with other specialist/generalist methods on several in-context semantic segmentation benchmarks.", + "bbox": [ + 89, + 641, + 483, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Few-shot semantic segmentation. Following previous works [29, 30, 33, 43], we evaluate UNICL-SAM on few-shot semantic segmentation benchmarks to demonstrate its great generalization performance. Additionally, we report the in-domain performance of specialist models to ensure a fair comparison. As illustrated in Table 1, UNICL-SAM shows competitive or even better results than recent powerful in-context segmentation models. For the in-domain COCO- $20^{i}$ [35] tests, UNICL-SAM outperforms previous generalist models, achieving an impressive MIoU score of 77.8 for one-shot and 78.7 for few-shot tests. For the out-of-domain FSS1000 [26] tests, UNICL-SAM also achieves competitive results that are close to specialist models, even without ever being trained on it. Furthermore, we conduct", + "bbox": [ + 89, + 689, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b909568ba547871559b4a9bad3f3adb00d1a1e5cd4121e0f94cf848fd5dcfef2.jpg", + "table_caption": [ + "Table 1. Comparison with state-of-the-art specialist and generalist models on few-shot semantic segmentation benchmarks. We report both one-shot and few-shot results for comparison. * indicates that the categories in training cover the categories in testing within the same dataset. Previous state-of-the-art results among generalists are underlined." + ], + "table_footnote": [], + "table_body": "
Multi-scale Feature AdapterPart Prototype DiscoveryUncertainty Probabilistic ModelingCOCO-20i
MIoU ↑MAE ↓
XXX71.770.039
XX✓72.990.037
X✓X72.410.038
✓X✓73.280.036
✓✓X73.780.036
X✓✓74.610.034
✓✓✓74.640.034
", + "bbox": [ + 517, + 401, + 903, + 522 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Ablation of proposed components.", + "bbox": [ + 578, + 532, + 839, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "experiments on LVIS- $92^{i}$ [29], a more challenging one-shot benchmark built upon LVIS [16]. On LVIS- $92^{i}$ , UNICL-SAM surpasses previous generalists in a one-shot setting with a MIoU score of 34.1, demonstrating superior generalization and versatility.", + "bbox": [ + 511, + 575, + 906, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 664, + 679, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct ablation experiments on each of the proposed components as follows. By default, we train the model on a small part of the COCO training set with 34k images and about $10\\%$ masks for 20 epochs. We report the Mean Intersection over Union (MIoU) and the Mean Absolute Error (MAE) scores for evaluation.", + "bbox": [ + 511, + 688, + 906, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ablations on proposed components. Table 2 illustrates the impact of the proposed components. The incorporation of our part prototype discovery technique contributes to improved segmentation performance, demonstrating its effectiveness in extracting contextual cues from examples. When combined with our uncertainty probabilistic modeling method, as indicated in line 6, the model's capabilities are further strengthened, resulting in a substantial increase", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20206", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/759064c628abc73b1364746e4faaa1c589e8b547cfcfcbca0c53c84b45290915.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Un-guided Graph AugmentationUn-guided Feature RefinementCOCO-20i
MIoU ↑MAE ↓
XX73.780.036
✓X73.820.036
X✓73.990.035
✓✓74.640.034
", + "bbox": [ + 96, + 88, + 478, + 188 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e6599f8df3e6f7acec9a91624dd39ae78929d1e05dec8d8a69101796fbec6bce.jpg", + "table_caption": [ + "Table 3. Ablation of uncertainty guided graph augmentation and feature refinement strategies." + ], + "table_footnote": [], + "table_body": "
Graph Model TypeUn-guided Graph AugmentationCOCO-20i
GCNUGGNMIoU ↑MAE ↓
✓✗✗72.120.038
✗✓✗72.410.038
✗✓✓73.820.036
", + "bbox": [ + 96, + 244, + 478, + 330 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/26434fe31582abdac61e0e5f04a49fff0a358735f2828392efc5b86f9147cb7c.jpg", + "image_caption": [ + "Figure 3. Visualization of uncertainty maps. We apply a predefined colormap to the uncertainty estimation map for visualization. Areas of high uncertainty are marked with red bboxes $\\square$ ." + ], + "image_footnote": [], + "bbox": [ + 94, + 373, + 480, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of $2.2\\mathrm{mIoU}$ . We attribute this improvement to the proposed UGGN's ability to accurately model data distribution and provide a reliable uncertainty measure, which facilitates the optimization of visual features through the two enhancement strategies. This, in turn, allows the part prototypes to yield more discriminative representations. Additionally, the integration of multi-scale adapters contributes to performance enhancement. The synergistic combination of these three components achieves optimal results, with a MIoU of 74.64 and a MAE of 0.034 on the COCO $20^{i}$ benchmark. These findings underscore the effectiveness of the proposed components. It is worth noting that in the following ablation experiments unless otherwise specified, the multi-scale feature adapter is used by default.", + "bbox": [ + 88, + 612, + 482, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablations on uncertainty probabilistic modeling. We investigate how the proposed uncertainty quantification stage and corresponding graph augmentation and feature refinement strategies affect performance. As shown in Table 3, utilizing either enhancement alone will help improve per", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/929cfafaaedae02bba3b8b7a16a12f519042814e10b1eb1156bf18dec0cb3845.jpg", + "table_caption": [ + "Table 4. Ablation of uncertainty-guided graph augmentation." + ], + "table_footnote": [], + "table_body": "
Un-guided Feature Refinement TypeCOCO-20i
MIoU ↑MAE ↓
top ratio masking43.730.117
random masking73.440.037
learnable gate74.640.034
", + "bbox": [ + 527, + 88, + 893, + 178 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dea861d5c37d90f89c72a6d0374516216759d0dca8ee87fa4f0e3da6cdeab5dd.jpg", + "table_caption": [ + "Table 5. Ablation of uncertainty-guided feature refinement type." + ], + "table_footnote": [], + "table_body": "
Number of ClustersCOCO-20i
MIoU ↑MAE ↓
573.360.037
1073.820.036
1572.500.039
2072.380.040
", + "bbox": [ + 576, + 220, + 844, + 327 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Ablation of the number of clusters of part prototype discovery.", + "bbox": [ + 511, + 337, + 903, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "formance. Further combining the two strategies can achieve better performance on the COCO $20^{i}$ benchmark.", + "bbox": [ + 511, + 396, + 903, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further explore the specific effects of the two enhancement strategies through detailed ablation experiments. Initially, we examine the impact of employing deterministic graph networks versus our uncertainty Gaussian graph network (UGGN). As depicted in Table 4, UGGN outperforms GCN by achieving an increase of 0.29 in mIoU, indicating that the network effectively models the distribution of graph representations. Building on this foundation, the introduction of uncertainty-guided graph augmentation technology further enhances model performance, resulting in an increase of 1.41 in MIoU scores.", + "bbox": [ + 511, + 429, + 906, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the uncertainty-guided feature refinement strategy, we study several different feature weighting/optimization schemes as shown in Table 5. the \"top ratio masking\" means masking nodes from areas of high uncertainty according to a preset ratio; \"random masking\" stands for masking out nodes whose associated uncertainty score is higher than a random probability map as described in [48]; \"learnable gate\" denotes our learnable gating mechanism. It can be observed that our \"learnable gate\" achieved the best result.", + "bbox": [ + 511, + 597, + 905, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We visualize the predicted uncertainty map for reference examples in the COCO test set. We can notice that there is often higher uncertainty at the foreground edge and small objects, as shown in the first row of Figure 3. Furthermore, for some annotation errors in COCO as shown in the second row, the model can give good instructions and the impact of these areas will be eliminated by our feature refinement strategy. The last line shows some naturally existing degradation and style changes. It can be seen that UGGN also exhibits ideal responses to these regions, effectively estimating the data variance.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20207", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9060a472d4834929daecad7b33688a3a305c8921a6e7b540af68f8641af61c2a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsResolutionTrainable ParamsCleanSupport ImageSupport LabelAvg. DeclineAvg. Decline Ratio
ColorBlunnessCompressionSpaceDomain shiftDeformation
SegGPT [43]448354M62.1-2.7-3.0-2.7-7.8-6.2-8.8-5.2-8.4%
Matcher [29]896-41.6-0.5-2.5-6.5-3.8-5.3-6.2-4.1-9.9%
SINE [30]89619M70.0-0.7-1.7-5.9-2.8-6.5-12.4-5.0-7.1%
UNICL-SAM51855M79.8-1.0-1.8-3.0-3.3-4.8-5.0-3.1-3.9%
", + "bbox": [ + 94, + 88, + 903, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Robustness comparison with advanced in-context segmentation generalists on constructed test sets. We present the MIoU score on clean data and the corresponding performance differences across various degraded data. Our UNICL-SAM outperforms these baselines in various degradation/domain shift scenarios with slight performance changes.", + "bbox": [ + 89, + 194, + 906, + 238 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/85636649746597e0b7ddb75e77ac458b3014e03d7bb283ecd78b5e5ffd02e2fa.jpg", + "image_caption": [ + "Figure 4. Visualization of cluster masks. We visualize the patches associated with the same cluster using a uniform color to facilitate the observation of the characteristics of the cluster mask." + ], + "image_footnote": [], + "bbox": [ + 93, + 260, + 480, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablations on part prototype discovery. We study the details of part prototype discovery stage. As shown in Table 6, we conducted an ablation experiment on the hyperparameter setting of the number of part prototype clusters. We found that the model showed the best performance when the number of clusters was 10, which was also used in the main experiments.", + "bbox": [ + 89, + 500, + 482, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also visualize the effect of part prototypes, as shown in Figure 4. The results clearly demonstrate that semantically similar patches are effectively grouped into the same cluster. Furthermore, for multiple targets within the same category, analogous local semantics are aggregated into a single cluster, thereby enhancing the modeling of foreground representations.", + "bbox": [ + 89, + 607, + 482, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Robustness Simulation Tests", + "text_level": 1, + "bbox": [ + 89, + 726, + 344, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To assess the effectiveness of our method in real-world applications, we constructed a test set featuring a variety of degraded and generated examples that may be encountered in practical scenarios, as shown in Figure 5. For detailed information on the robustness simulation test settings, please refer to the Appendix. The results in Table 7 indicate that our approach significantly outperforms existing advanced in-context segmentation models, effectively addressing the diverse challenges posed by these varied scenes. UNICLSAM excels not only in enhancing segmentation perfor", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/18ec0e824087f64185067bc0be2ad663e8e7ae3105b856ef5fd3af8ca25965d3.jpg", + "image_caption": [ + "Figure 5. A simple example of robustness simulation tests designed to assess the model's capabilities across various degradation and domain shift scenarios. Zoom in for a better view." + ], + "image_footnote": [], + "bbox": [ + 516, + 262, + 903, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mance in clean environments but also improving results in degraded conditions as well. It surpasses the previous state-of-the-art generalist average by an impressive score of 11.7 MIoU, highlighting its effectiveness across diverse conditions. Even in the most challenging robustness tests, our model's performance fluctuations remain within a narrow range, with an average decline of only 3.1 MIoU. This demonstrates its great robustness.", + "bbox": [ + 511, + 516, + 906, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 660, + 633, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we point out the data distribution uncertainty in in-context segmentation and introduce UNICL-SAM, a robust in-context segmentation framework that leverages uncertainty modeling to evaluate each input example and related graph augmentation and feature refinement strategies to help alleviate the impact of areas with high data uncertainty. We also present part prototypes as a form of in-context instruction to obtain discriminative local semantics. Leveraging the effective design, UNICL-SAM achieved superior performance on in-context segmentation benchmarks, while demonstrating exceptional robustness in conducted robustness simulation tests. We hope UNICL-SAM can promote the development of in-context segmentation generalists for real-world applications.", + "bbox": [ + 511, + 689, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20208", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgment", + "text_level": 1, + "bbox": [ + 89, + 90, + 263, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China (No.62472396, No.62121002), Anhui Provincial Science and Technology Major Project (No.2023z020006), and the advanced computing resources provided by the SuperComputing Center of USTC.", + "bbox": [ + 89, + 113, + 485, + 198 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 210, + 187, + 227 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros. Visual prompting via image inpainting. Advances in Neural Information Processing Systems, 35:25005-25017, 2022. 2", + "[2] Benjamin Bischke, Patrick Helber, Damian Borth, and Andreas Dengel. Segmentation of imbalanced classes in satellite imagery using adaptive uncertainty weighted class loss. In IGARSS 2018-2018 IEEE International Geoscience and Remote Sensing Symposium, pages 6191-6194. IEEE, 2018. 2", + "[3] Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty in neural network. In International conference on machine learning, pages 1613-1622. PMLR, 2015. 1, 4", + "[4] Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty in neural network. In International conference on machine learning, pages 1613-1622. PMLR, 2015. 4", + "[5] Patrik Olā Bressan, José Marcato Junior, José Augusto Correa Martins, Maximilian Jaderson de Melo, Diogo Nunes Gonçalves, Daniel Matte Freitas, Ana Paula Marques Ramos, Michelle Táis Garcia Furuya, Lucas Prado Osco, Jonathan de Andrade Silva, et al. Semantic segmentation with labeling uncertainty and class imbalance applied to vegetation mapping. International Journal of Applied Earth Observation and Geoinformation, 108:102690, 2022. 2", + "[6] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1", + "[7] Jie Chang, Zhonghao Lan, Changmao Cheng, and Yichen Wei. Data uncertainty learning in face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5710-5719, 2020. 2", + "[8] Bike Chen, Wei Peng, Xiaofeng Cao, and Juha Röning. Hyperbolic uncertainty aware semantic segmentation. IEEE Transactions on Intelligent Transportation Systems, 2023. 2", + "[9] Wei-Ting Chen, Yu-Jiet Vong, Sy-Yen Kuo, Sizhou Ma, and Jian Wang. Robustsam: Segment anything robustly on degraded images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4081-4091, 2024. 2", + "[10] Kazuki Endo, Masayuki Tanaka, and Masatoshi Okutomi. Semantic segmentation of degraded images using layer-wise feature adjustor. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3205-3213, 2023. 2", + "[11] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep" + ], + "bbox": [ + 93, + 236, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "learning. In international conference on machine learning, pages 1050-1059. PMLR, 2016. 1", + "[12] Jakob Gawlikowski, Cedrique Rovile Njieutcheu Tassi, Mohsin Ali, Jongseok Lee, Matthias Humt, Jianxiang Feng, Anna Kruspe, Rudolph Trielbel, Peter Jung, Ribana Roscher, et al. A survey of uncertainty in deep neural networks. Artificial Intelligence Review, 56(Suppl 1):1513-1589, 2023. 2", + "[13] Ethan Goan and Clinton Fookes. Uncertainty in real-time semantic segmentation on embedded systems. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4491-4501, 2023. 2", + "[14] Dazhou Guo, Yanting Pei, Kang Zheng, Hongkai Yu, Yuhang Lu, and Song Wang. Degraded image semantic segmentation with dense-gram networks. IEEE Transactions on Image Processing, 29:782-795, 2019. 2", + "[15] Hongji Guo, Hanjing Wang, and Qiang Ji. Uncertainty-guided probabilistic transformer for complex action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20052-20061, 2022. 2", + "[16] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 6", + "[17] Saumya Gupta, Yikai Zhang, Xiaoling Hu, Prateek Prasanna, and Chao Chen. Topology-aware uncertainty for image segmentation. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[18] Sunghwan Hong, Seokju Cho, Jisu Nam, Stephen Lin, and Seungryong Kim. Cost aggregation with 4d convolutional swim transformer for few-shot segmentation. In European Conference on Computer Vision, pages 108-126. Springer, 2022. 6", + "[19] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 5", + "[20] Kaixiang Ji, Feng Chen, Xin Guo, Yadong Xu, Jian Wang, and Jingdong Chen. Uncertainty-guided learning for improving image manipulation detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22456-22465, 2023. 2", + "[21] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 1, 2, 3, 4", + "[22] Soopil Kim, Philip Chikontwe, Sion An, and Sang Hyun Park. Uncertainty-aware semi-supervised few shot segmentation. Pattern Recognition, 137:109292, 2023. 2", + "[23] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 5", + "[24] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2, 3, 6", + "[25] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20209", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3041-3050, 2023. 3", + "[26] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few-shot segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2869–2878, 2020. 6", + "[27] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European conference on computer vision, pages 280-296. Springer, 2022. 3", + "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5", + "[29] Yang Liu, Muzhi Zhu, Hengtao Li, Hao Chen, Xinlong Wang, and Chunhua Shen. Matcher: Segment anything with one shot using all-purpose feature matching. arXiv preprint arXiv:2305.13310, 2023. 1, 2, 3, 6, 8", + "[30] Yang Liu, Chenchen Jing, Hengtao Li, Muzhi Zhu, Hao Chen, Xinlong Wang, and Chunhua Shen. A simple image segmentation framework via in-context examples. Advances in Neural Information Processing Systems, 2024. 1, 2, 6, 8", + "[31] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[32] Yao Lu, Max Bartolo, Alastair Moore, Sebastian Riedel, and Pontus Stenetorp. Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity. arXiv preprint arXiv:2104.08786, 2021. 1", + "[33] Lingchen Meng, Shiyi Lan, Hengduo Li, Jose M Alvarez, Zuxuan Wu, and Yu-Gang Jiang. Segic: Unleashing the emergent correspondence for in-context segmentation. In European Conference on Computer Vision, pages 203-220. Springer, 2024. 1, 2, 3, 6", + "[34] Juhong Min, Dahiyun Kang, and Minsu Cho. Hypercorrelation squeeze for few-shot segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6941-6952, 2021. 6", + "[35] Khoi Nguyen and Sinisa Todorovic. Feature weighting and boosting for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 622-631, 2019. 6", + "[36] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3, 6", + "[37] Xinru Shan and Chaoning Zhang. Robustness of segment anything model (sam) for autonomous driving in adverse weather conditions. arXiv preprint arXiv:2306.13290, 2023. 2", + "[38] Dianmo Sheng, Dongdong Chen, Zhentao Tan, Qiankun Liu, Qi Chu, Jianmin Bao, Tao Gong, Bin Liu, Shengwei Xu," + ], + "bbox": [ + 91, + 90, + 483, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Nenghai Yu. Towards more unified in-context visual understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13362-13372, 2024. 6", + "[39] Mahesh Subedar, Ranganath Krishnan, Paulo Lopez Meyer, Omesh Tickoo, and Jonathan Huang. Uncertainty-aware audiovisual activity recognition using deep bayesian variational inference. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6301-6310, 2019. 2", + "[40] Yanpeng Sun, Jiahui Chen, Shan Zhang, Xinyu Zhang, Qiang Chen, Gang Zhang, Errui Ding, Jingdong Wang, and Zechao Li. Vrp-sam: Sam with visual reference prompt. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23565-23574, 2024. 6", + "[41] Matias Valdenegro-Toro. Deep sub-ensembles for fast uncertainty estimation in image classification. arXiv preprint arXiv:1910.08168, 2019. 2", + "[42] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6830-6839, 2023. 2, 6", + "[43] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023. 1, 2, 6, 8", + "[44] Zhenyu Wang, Yali Li, Ye Guo, Lu Fang, and Shengjin Wang. Data-uncertainty guided multi-phase learning for semi-supervised object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4568-4577, 2021. 2", + "[45] Zhiyong Wu, Yaoxiang Wang, Jiacheng Ye, and Lingpeng Kong. Self-adaptive in-context learning: An information compression perspective for in-context example selection and ordering. arXiv preprint arXiv:2212.10375, 2022. 1", + "[46] Weihao Xia, Zhanglin Cheng, Yujiu Yang, and Jing-Hao Xue. Cooperative semantic segmentation and image restoration in adverse environmental conditions. arXiv preprint arXiv:1911.00679, 2019. 2", + "[47] Chengming Xu, Chen Liu, Yikai Wang, and Yanwei Fu. Towards global optimal visual in-context learning prompt selection. arXiv preprint arXiv:2405.15279, 2024. 1", + "[48] Fan Yang, Qiang Zhai, Xin Li, Rui Huang, Ao Luo, Hong Cheng, and Deng-Ping Fan. Uncertainty-guided transformer reasoning for camouflaged object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4146-4155, 2021. 2, 7", + "[49] Zhou Yang, Weisheng Dong, Xin Li, Jinjian Wu, Leida Li, and Guangming Shi. Self-feature distillation with uncertainty modeling for degraded image recognition. In European Conference on Computer Vision, pages 552-569. Springer, 2022. 2", + "[50] Tianyuan Yu, Da Li, Yongxin Yang, Timothy M Hospedales, and Tao Xiang. Robust person re-identification by modelling feature uncertainty. In Proceedings of the IEEE/CVF international conference on computer vision, pages 552–561, 2019. 2" + ], + "bbox": [ + 516, + 90, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20210", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Jian-Wei Zhang, Yifan Sun, Yi Yang, and Wei Chen. Featureproxy transformer for few-shot segmentation. Advances in Neural Information Processing Systems, 35:6575-6588, 2022. 6", + "[52] Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Xianzheng Ma, Hao Dong, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048, 2023. 2, 6", + "[53] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. What makes good examples for visual in-context learning? Advances in Neural Information Processing Systems, 36:17773-17794, 2023. 1", + "[54] Zhizheng Zhang, Cuiling Lan, Wenjun Zeng, Zhibo Chen, and Shih-Fu Chang. Uncertainty-aware few-shot image classification. arXiv preprint arXiv:2010.04525, 2020. 2", + "[55] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127:302-321, 2019. 5", + "[56] Dingyuan Zhu, Ziwei Zhang, Peng Cui, and Wenwu Zhu. Robust graph convolutional networks against adversarial attacks. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1399-1407, 2019. 4" + ], + "bbox": [ + 91, + 90, + 480, + 430 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20211", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/UNICL-SAM_ Uncertainty-Driven In-Context Segmentation with Part Prototype Discovery/fc453b95-4cba-4bb9-95fb-6c8d34e4ccaa_model.json b/2025/UNICL-SAM_ Uncertainty-Driven In-Context Segmentation with Part Prototype Discovery/fc453b95-4cba-4bb9-95fb-6c8d34e4ccaa_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cacfe016fba2bbde858d88e20508ad18d4de13c2 --- /dev/null +++ b/2025/UNICL-SAM_ Uncertainty-Driven In-Context Segmentation with Part Prototype Discovery/fc453b95-4cba-4bb9-95fb-6c8d34e4ccaa_model.json @@ -0,0 +1,2323 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.197, + 0.131, + 0.804, + 0.177 + ], + "angle": 0, + "content": "UNICL-SAM: Uncertainty-Driven In-Context Segmentation with Part Prototype Discovery" + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.203, + 0.874, + 0.277 + ], + "angle": 0, + "content": "Dianmo Sheng\\(^{1,4}\\), Dongdong Chen\\(^{2}\\), Zhentao Tan\\(^{1,4}\\), Qiankun Liu\\(^{3}\\), Qi Chu\\(^{1,4}\\), Tao Gong\\(^{1,4,\\dagger}\\), Bin Liu\\(^{1,4}\\), Jing Han\\(^{5}\\), Wenbin Tu\\(^{5}\\), Shengwei Xu\\(^{6}\\), Nenghai Yu\\(^{1,4}\\), \\(^{1}\\)School of Cyber Science and Technology, USTC \\(^{2}\\)Microsoft GenAI \\(^{3}\\)USTB \\(^{4}\\)Anhui Province Key Laboratory of Digital Security \\(^{5}\\)BoCom Research \\(^{6}\\)BESTI" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.278, + 0.773, + 0.292 + ], + "angle": 0, + "content": "dmsheng@mail.ustc.edu.cn, tgong@ustc.edu.cn, xusw@besti.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.359, + 0.486, + 0.737 + ], + "angle": 0, + "content": "Recent advancements in in-context segmentation generalists have demonstrated significant success in performing various image segmentation tasks using a limited number of labeled example images. However, real-world applications present challenges due to the variability of support examples, which often exhibit quality issues resulting from various sources and inaccurate labeling. How to extract more robust representations from these examples has always been one of the goals of in-context visual learning. In response, we propose UNICL-SAM, to better model the example distribution and extract robust representations to help in-context segmentation. We incorporate an uncertainty probabilistic module to quantify each example's reliability during both the training and testing phases. Utilizing this uncertainty estimation, we introduce an uncertainty-guided graph augmentation and feature refinement strategy, aimed at mitigating the impact of high-uncertainty regions to enhance the learning of robust representations. Subsequently, we construct prototypes for each example by aggregating part information, thereby creating reliable in-context instruction that effectively represents fine-grained local semantics. This approach serves as a valuable complement to traditional global pooling features. Experimental results demonstrate the effectiveness of the proposed framework, underscoring its potential for real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.763, + 0.222, + 0.78 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.789, + 0.484, + 0.88 + ], + "angle": 0, + "content": "Inspired by the great success of in-context learning [6, 32, 45] in the field of natural language processing (NLP), recent advances in vision generalist models [29, 30, 33, 43] indicate that different levels of segmentation tasks, including instance, semantic, and video segmentation, can be unified within one framework. The remarkable performance and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.329, + 0.907, + 0.374 + ], + "angle": 0, + "content": "strong task generalization capabilities of these in-context segmentation models are undoubtedly promising for the research community." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.909, + 0.577 + ], + "angle": 0, + "content": "However, existing in-context segmentation generalists often assume that the data in the training sets are of high quality and treat them equally within the same framework, thereby overlooking their diversity and differences. Previous works [47, 53] have also reported fluctuations in model performance based on varying examples and investigated strategies for selecting effective in-context examples. However, these studies only screened examples during the testing phase without considering optimization methods during the training phase. The above discussion motivates us to explore how to incorporate potential distributional diversity into the training phase to develop more robust in-context segmentation models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In light of this, we introduce UNICL-SAM, to investigate and address the modeling of example distribution differences and further feature optimization strategies within in-context segmentation. We draw inspiration from traditional Bayesian theory [3, 11, 21], which has previously addressed the assumptions related to data distribution differences and defined corresponding uncertainty measures to quantify and assess these distributional discrepancies. This framework facilitates the optimization of the model. Our intuition suggests that data uncertainty/discrepancy also exists within the context of contextual segmentation, potentially arising from varying observations in the reference examples. Thus, an uncertainty probabilistic Gaussian graph model is proposed to capture the training data distribution and estimate uncertainty for each example during the test phase. Additionally, we utilize the estimated uncertainty during training to guide the model to focus more on reliable (low-uncertainty) regions. The designed uncertainty-guided graph augmentation module enhances graph representation via learning invariant representations across different views of the graph. Further combined with an uncertainty-guided" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.246, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "20201" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.362 + ], + "angle": 0, + "content": "feature reweighting strategy, the model yields more robust and reliable in-context indications while minimizing the influence of extraneous information. To fully harness the discriminating information delivered from in-context examples, we incorporate diverse visual instructions like previous works [33] and introduce part prototypes as a useful form of instructions to enhance in-context knowledge representation. In practical applications, the foreground regions under consideration inherently contain local semantics, which are often overlooked by simplistic global pooling methods. By effectively aggregating similar local features and enhancing the differentiation among prototypes, part prototypes can more accurately capture fine-grained semantic information. Through a well-designed framework and comprehensive experiments, our model demonstrates great performance on both clean and degraded data, effectively alleviating the impact of the aforementioned challenge. We summarize our main contribution as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.364, + 0.483, + 0.439 + ], + "angle": 0, + "content": "- We propose an uncertainty probabilistic module to characterize the distribution of input data. We further leverage uncertainty estimation to guide the model in learning more reliable features and refining visual instructions throughout both the training and testing phases." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.44, + 0.483, + 0.484 + ], + "angle": 0, + "content": "- We introduce part prototypes as a form of visual instruction designed to effectively extract knowledge from reference examples by aggregating local semantics." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.484, + 0.484, + 0.543 + ], + "angle": 0, + "content": "- Our comprehensive experiments reveal that UNICL-SAM exhibits superior performance on traditional benchmarks and shows good generalization capabilities on conducted real-world simulation robustness tests." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.364, + 0.484, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.559, + 0.241, + 0.574 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.869 + ], + "angle": 0, + "content": "In-Context Visual Learning. Recently, applying the in-context learning paradigm to the field of computer vision has become one of the hot spots in the research community. Bar et al. [1] pioneers in-context visual learning framework utilizing image inpainting on illustrations and infographics derived from vision-related literature, which shows competencies in fundamental CV tasks. Painter [42] employs masked image modeling on continuous pixels to facilitate in-context learning and yields highly competitive outcomes on seven tasks. Following this, SegGPT [43] offers a specialized generalist aimed at addressing diverse and unlimited segmentation tasks using a similar framework. SINE [30] building upon the SegGPT framework to tackle the task ambiguity problem. Other approaches [29, 52] investigated prompting techniques of powerful segmentation model SAM [24]. Moreover, recent work SEGIC [33] shows superior performance with a large margin compared with previous works by leveraging diverse in-context instructions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "However, the aforementioned work does not take into account the potential impact of the diversity in training and" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "testing data distributions, which motivates us to develop a more robust in-context segmentation model." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.128, + 0.905, + 0.4 + ], + "angle": 0, + "content": "Uncertainty Estimation in computer vision. The issue of uncertainty in neural networks, which pertains to the lack of certainty estimates or under-confidence in predictions, has garnered increasing attention in recent research [7, 12, 13, 15, 20-22, 39, 48, 54]. Kendall et al. [21] delineate two primary types: epistemic uncertainty, arising from the model's limited knowledge, and aleatoric uncertainty, which is typically associated with inherent noise within the data. Nonbayesian approaches [7, 15, 44, 48] typically utilize a Gaussian distribution framework to estimate both the mean and variance of model parameters or input data, which serve as indicators of uncertainty. They effectively avoid the substantial computational overhead of the Bayesian solution. These techniques have been effectively applied to various computer vision tasks, including image classification [41, 54], semantic segmentation [2, 5, 8, 17], object detection [44, 48], person reidentification [50] and many other computer vision tasks [7, 15, 20]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.4, + 0.905, + 0.462 + ], + "angle": 0, + "content": "In this work, we first claim the diverse data distributions that in-context segmentation models need to navigate. We then propose an uncertainty probabilistic network and corresponding enhancement strategies for optimization." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.467, + 0.905, + 0.603 + ], + "angle": 0, + "content": "Robust Segmentation. Numerous studies [9, 10, 14, 37, 46, 49] have studied how to improve the performance of segmentation models in cross-domain and quality-degraded scenarios. For example, recently, RobustSAM [9] has examined the performance deterioration of the Segment Anything Model (SAM) [24] when handling low-quality images, subsequently designing a pipeline to enforce both feature and token consistency between clean and generated degraded images." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.604, + 0.905, + 0.679 + ], + "angle": 0, + "content": "Building upon these insights, we propose uncertainty-guided graph augmentation and feature refinement techniques to obtain robust representations. The subsequent part prototype discovery stage provides discriminative local semantics, further enhancing the model's robustness." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.693, + 0.605, + 0.708 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.718, + 0.905, + 0.809 + ], + "angle": 0, + "content": "In-context segmentation aims to identify a specific segmentation task within the given \\(K\\) in-context examples \\(ref = \\{(\\mathbf{x}_r,\\mathbf{y}_r)\\}_{i = 1}^K\\), where \\(\\mathbf{x}_r\\in \\mathbb{R}^{3\\times H\\times W}\\), \\(\\mathbf{y}_r\\in \\mathbb{R}^{1\\times H\\times W}\\) are one reference image and its corresponding ground-truth label, and segment the interested area in the target image \\(\\mathbf{x}_t\\in \\mathbb{R}^{3\\times H\\times W}\\) as a binary mask \\(\\mathbf{y}_t\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.625, + 0.833 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Our UNICL-SAM combines the advantages of previous work and further utilizes the proposed components to improve the robustness of the model. The overall pipeline is shown in Figure 1, which consists of four stages: fea-" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20202" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.905, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.342, + 0.908, + 0.414 + ], + "angle": 0, + "content": "Figure 1. Overview of UNICL-SAM. With the fine-grained features, UNICL-SAM performs uncertainty quantification to model the input distribution and produce the estimated uncertainty map. The corresponding uncertainty-guided graph augmentation and feature refinement stages tend to learn robust graph representations by mitigating the impact of high uncertainty areas. Further, UNICL-SAM conducts part prototype discovery to get the part prototypes with the other three kinds of in-context instructions. Finally, the in-context instructions are fed into the query prompt generator to get the in-context prompt for guiding the mask decoder to predict segmentation results." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.439, + 0.486, + 0.697 + ], + "angle": 0, + "content": "ture extraction, uncertainty probabilistic modeling, part prototype discovery, and final mask decoding. First, we extract and integrate multi-scale features, then compute dense correspondence between target and reference as previous works [29, 33] (Section 3.2). Then our uncertainty Gaussian graph model will provide quantified uncertainty estimation. Utilizing the uncertainty map, we further perform uncertainty-guided graph augmentation and feature refinement techniques to facilitate learning robust visual representations (Section 3.3). With the refined visual representations, part prototypes are generated as a form of in-context instruction to facilitate the extraction of more discriminative local semantics (Section 3.4). Finally, we generate incontext prompts through a prompt generator based on the rich instructions of reference and target, guiding the mask decoder model to obtain the final segmentation result (Section 3.5). We will elaborate on each stage in detail." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.71, + 0.273, + 0.725 + ], + "angle": 0, + "content": "3.2. Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.483, + 0.794 + ], + "angle": 0, + "content": "Building upon insights from previous works [29, 33], we establish an in-context segmentation baseline by combining the advancements of existing VFMs [24, 36], thereby unleashing the power of pre-trained foundation models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We use the frozen pre-trained image encoder to extract visual features of both the reference image \\(\\mathbf{x}_r\\) and the target image \\(\\mathbf{x}_t\\), respectively. Inspired by previous works on dense prediction tasks like object detection [27] and semantic segmentation [25], we incorporate a multi-scale feature adapter to leverage information of feature maps at different scales and aggregate them to produce the final feature map" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.437, + 0.907, + 0.468 + ], + "angle": 0, + "content": "\\(\\mathbf{F}_r \\in \\mathbb{R}^{C \\times H \\times W}\\) and \\(\\mathbf{F}_t \\in \\mathbb{R}^{C \\times H \\times W}\\), where \\(C\\) denotes the number of channels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.909, + 0.593 + ], + "angle": 0, + "content": "Then, we calculate the masked reference feature as \\(\\mathbf{F}_{rm} = \\mathbf{F}_r\\cdot \\mathbf{y}_r\\). The patch-level similarity \\(\\mathbf{Corr}\\in\\) \\(\\mathbb{R}^{HW\\times HW}\\) is calculated between \\(\\mathbf{F}_{rm}\\) and \\(\\mathbf{F}_t\\) as the dense correspondence. With the correspondence map, we calculate the pseudo mask \\(\\mathbf{y}_t^{pseudo}\\) to extract the region of interest from the target feature like \\(\\mathbf{F}_{tm}^{pseudo} = \\mathbf{F}_t\\cdot \\mathbf{y}_t^{pseudo}\\). For details about the multi-scale feature adapter and correspondence extraction, please refer to the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.606, + 0.822, + 0.622 + ], + "angle": 0, + "content": "3.3. Uncertainty Probabilistic Modeling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.907, + 0.794 + ], + "angle": 0, + "content": "Deep neural network's output has varying degrees of uncertainty, influenced by how the training data is collected and the model's architecture, with observation noise affecting the instantiated predictions [21]. In the context of in-context segmentation, uncertainty also arises from the quality of the in-context examples. The limited number of samples in each iteration amplifies the impact of observation noise, further compounding uncertainty, which results in poor predictions. To address this, we model the data-dependent uncertainty of the input data and leverage it to more effectively optimize the network." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Uncertainty quantification. We design our uncertainty Gaussian graph network (UGGN) to represent the graph representation using a Gaussian distribution. Given the fine-grained feature \\(\\mathbf{F}_m = \\{\\mathbf{F}_{rm},\\mathbf{F}_{tm}^{pseudo}\\}\\), we first construct a graph representation \\(\\mathcal{G} = \\{\\mathbf{V},\\mathbf{E}\\}\\), regarding each patch feature as each node in the graph \\(\\mathbf{V} = \\{\\mathbf{V}_i|i = 1,\\dots ,N\\}\\), \\(\\mathbf{V}_i\\in \\mathbf{F}_m\\). For the edge information, we uti" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20203" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.275, + 0.485, + 0.348 + ], + "angle": 0, + "content": "Figure 2. Illustration of the uncertainty quantification and part prototype discovery stages. UGGN serves as a probabilistic graph network, with its initial layer dedicated to uncertainty quantification. Subsequent layers are designed to facilitate the generation of part prototypes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.371, + 0.484, + 0.492 + ], + "angle": 0, + "content": "lize the affinity similarity as the edge weight from the \\(j^{th}\\) node to the \\(k^{th}\\) node is modeled by their affinity in the embedded space, which is formulated as \\(\\mathbf{E}(j,k) = (\\mathbf{V}_j)\\cdot (\\mathbf{V}_k)^T,\\mathbf{V}_j,\\mathbf{V}_k\\in \\mathbf{V}\\). We only retain edges with similarity greater than an established threshold \\(\\varepsilon\\) and sparse the edge connections according to a certain top ratio \\(r\\) as \\(\\mathbf{E} = r(\\{\\mathbf{E}(j,k)|\\mathbf{A}_{j,k}\\geqslant \\varepsilon \\})\\), where \\(\\mathbf{A}\\in \\mathbb{R}^{N\\times N}\\) denotes the adjacency matrix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.484, + 0.68 + ], + "angle": 0, + "content": "As illustrated in Figure 2 above, in the uncertainty quantification stage, what UGGN delivers the mean \\(\\mu_{\\mathcal{G}}\\) and variance \\(\\sigma_{\\mathcal{G}}\\) for each graph \\(\\mathcal{G}\\). Inspired by previous work [56], we first map the graph representation into a high-dimensional Gaussian space utilizing separate learnable mean and variance parameters, then update the mean and variance based on modified node aggregation. We use \\(\\mathbf{M}^{(l)} = [\\mu_i^{(l)},\\dots ,\\mu_N^{(l)}]\\in \\mathbb{R}^{N\\times d}\\) and \\(\\pmb{\\Sigma}^{(l)} = [\\sigma_i^{(l)},\\dots ,\\sigma_N^{(l)}]\\in \\mathbb{R}^{N\\times d}\\) to denote the matrix of means and variances for all nodes at layer \\(l\\in L\\) of UGGN, respectively. \\(d\\) represents the hidden dimension. For the first layer, we adopt a fully connected layer to map as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.688, + 0.483, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} ^ {(1)} = \\operatorname {E L U} \\left(\\mathbf {V W} _ {\\mu} ^ {(0)}\\right), \\boldsymbol {\\Sigma} ^ {(1)} = \\operatorname {R e L U} \\left(\\mathbf {V W} _ {\\sigma} ^ {(0)}\\right) \\quad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.765 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_{\\mu} \\in \\mathbb{R}^{N \\times d}\\) and \\(\\mathbf{W}_{\\sigma} \\in \\mathbb{R}^{N \\times d}\\) are learnable parameters, ELU and ReLU represent for activation functions. \\(h \\sim \\mathcal{N}(\\mu^{(1)}, \\sigma^{(1)})\\) denote the parameterized distribution." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.766, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Then, we need to sample from the estimated Gaussian distributions. However, sampling directly from the distribution is a non-differentiable operation, which will result in the model not being optimized. Referring to previous work [4], we decided to use the reparameterization trick. Specifically, we first draw a sample \\(\\epsilon\\) from the standard Gaussian distribution \\(\\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) randomly, and then obtain the sample by computing \\(\\mathbf{Z} = \\mu^{(1)} + \\epsilon \\sigma^{(1)},\\mathbf{Z}\\in \\mathbb{R}^{N\\times C}\\). According to Bayesian probability theory [3, 21], we can con" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "sider \\(\\mathbf{Z}\\) as empirical samples drawn from an approximate predictive distribution. The uncertainty of the input data can be quantified by calculating channel-wise variance, expressed as \\(\\mathbf{U}_{data} = \\mathrm{Norm}(Var(\\mathbf{Z}))\\), \\(\\mathbf{U}_{data} \\in \\mathbb{R}^{N \\times 1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.906, + 0.288 + ], + "angle": 0, + "content": "Uncertainty-guided graph augmentation. With the estimated uncertainty, we can assume that a larger variance indicates a greater deviation from the expected data distribution. We utilize a softmax function to map the \\(\\mathbf{U}_{data}\\) into a probability distribution \\(p_u^{node}\\), which determines the unimportance of each node in the graph. Combining with traditional graph augmentations like node dropping, and node masking, we can decide whether to drop or mask a node based on the its uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.907, + 0.456 + ], + "angle": 0, + "content": "Furthermore, we compute the edge uncertainty using the affinity matrix of \\(\\mathbf{U}_{data}\\), and map to a probability distribution \\(p_u^{edge}\\) as well. Intuitively, if the uncertainty at the edge's endpoints is higher, the uncertainty of the edge itself will be higher. This allows us to perform edge perturbation and subgraph transformation on the graph. With the augmented graph representation \\(\\mathcal{G}_{aug}\\) and its vanilla view \\(\\mathcal{G}\\), we use the UGGN to map them into the same representation space, and minimize their KL divergence to mitigate the influence of high-uncertainty nodes, thereby learning a more robust graph representation:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.467, + 0.907, + 0.509 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {K L} ^ {\\mathcal {G}} = \\sum_ {i = 1} ^ {N} K L (\\mathcal {N} \\left(\\mu_ {\\mathcal {G}}, \\sigma_ {\\mathcal {G}}\\right) | | \\mathcal {N} \\left(\\mu_ {\\mathcal {G} _ {a u g}}, \\sigma_ {\\mathcal {G} _ {a u g}}\\right)) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.906, + 0.64 + ], + "angle": 0, + "content": "Uncertainty-guided feature refinement. On the other hand, the uncertainty estimation can be mapped back to spatial space as an uncertainty map \\(\\mathbf{U}_{map} \\in \\mathbf{R}^{H \\times W \\times 1}\\). We directly compute the dot product of reverse uncertainty map \\(\\mathbf{1} - \\mathbf{U}_{map}\\) with the visual features \\(\\mathbf{F}\\) and their associated labels \\(\\mathbf{y}\\) for refinement, which will effectively alleviate the impact of high uncertainty positions, resulting in more accurate foreground representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.641, + 0.906, + 0.717 + ], + "angle": 0, + "content": "Since the uncertainty estimation may be less reliable in the early stages of training, we introduce an uncertainty gate to adjust the weights for uncertainty refinement dynamically. This gate is initialized with a small positive scalar to mitigate any initial unreliability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.727, + 0.744, + 0.744 + ], + "angle": 0, + "content": "3.4. Part prototype Discovery" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.826 + ], + "angle": 0, + "content": "Despite we obtain a refined representation, the foreground region of example retains local semantics, resulting in varying degrees of correspondence with the target across different areas. To extract more discriminative representations, we propose the implementation of part prototype discovery." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.902 + ], + "angle": 0, + "content": "As shown in Figure 2 below, we can treat the last layers of UGGN as an extension by converting a deterministic graph clustering network to a probabilistic type, i.e., a parameterized distribution for feature clustering. Then, we propose an attention mechanism for the next layers to" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20204" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "assign different weights to neighbors based on their variances since larger variances indicate more uncertainties in the latent representations. A smooth exponential function is utilized to control the effect of variances on weights \\(\\mathcal{A}^{(l)} = \\exp (-\\gamma \\Sigma^{(l)})\\)" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.173, + 0.483, + 0.197 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} ^ {(l + 1)} = \\tilde {\\mathbf {D}} ^ {- \\frac {1}{2}} \\mathbf {A} \\tilde {\\mathbf {D}} ^ {- \\frac {1}{2}} \\left(\\mathbf {M} ^ {(l)} \\odot \\mathcal {A} ^ {(l)}\\right) \\mathbf {W} _ {\\mu} ^ {(l)} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.199, + 0.483, + 0.221 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} ^ {(l + 1)} = \\tilde {\\mathbf {D}} ^ {- \\frac {1}{2}} \\mathbf {A} \\tilde {\\mathbf {D}} ^ {- \\frac {1}{2}} \\left(\\boldsymbol {\\Sigma} ^ {(l)} \\odot \\mathcal {A} ^ {(l)} \\odot \\mathcal {A} ^ {(l)}\\right) \\mathbf {W} _ {\\sigma} ^ {(l)} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.227, + 0.483, + 0.258 + ], + "angle": 0, + "content": "where \\(\\mathbf{D}\\) is the diagonal degree matrix calculated by \\(\\mathbf{D}_{i,i} = \\sum_{j}\\mathbf{A}_{i,j}\\), and \\(\\tilde{\\mathbf{D}} = \\mathbf{D} + \\mathbf{I}_N\\), where \\(\\mathbf{I}_N\\) is the identity matrix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.258, + 0.484, + 0.346 + ], + "angle": 0, + "content": "We also sampled the updated distribution using the reparameterization trick to obtain the final node classification results. To solve the non-differentiability caused by the arg max operation to derive classification labels, we utilize the Gumbel-Softmax trick [19] to get the final one-hot like clusters \\(\\mathbf{C}\\mathbf{u} \\in \\mathbb{R}^{N \\times c}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.351, + 0.483, + 0.398 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C l u} _ {i} = \\frac {\\exp \\left(\\left(\\log \\left(\\pi_ {i}\\right) + g _ {i}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {c} \\exp \\left(\\left(\\log \\left(\\pi_ {j}\\right) + g _ {j}\\right) / \\tau\\right)}, \\quad \\text {f o r} i = 1, \\dots , c. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.399, + 0.483, + 0.444 + ], + "angle": 0, + "content": "where \\(g_{i},\\dots ,g_{c}\\) are i.i.d samples drawn from the Gumbel(0,1) distribution, \\(\\pi\\) is a categorical distribution with class probabilities and \\(\\tau\\) is the softmax temperature." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.444, + 0.484, + 0.488 + ], + "angle": 0, + "content": "With the clusters \\(\\mathbf{C}\\mathbf{u}\\), we get part prototype representations \\(\\mathbf{P} = \\{\\mathbf{P}_i\\in \\mathbb{R}^C\\mid i = 1,\\dots ,c\\}\\) by averaging \\(n\\) node features in the same cluster:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.491, + 0.483, + 0.52 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {i} = \\frac {1}{n} \\mathbf {C l u} _ {i} ^ {\\top} \\cdot \\mathbf {F}, \\quad \\mathbf {F} \\in \\mathbf {F} _ {m} \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.524, + 0.248, + 0.54 + ], + "angle": 0, + "content": "3.5. Mask Decoding" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.547, + 0.484, + 0.775 + ], + "angle": 0, + "content": "To further interact with the mask decoder for mask prediction, we extracted four kinds of visual instructions for in-context prompt generation: visual features \\(\\mathbf{F}\\), mask average pooling features \\(\\mathbf{F}_p = \\mathrm{Avg\\_pool}(\\mathbf{F})\\), \\(\\mathbf{F} \\in \\mathbf{F}_m\\), part prototypes \\(\\mathbf{P}\\), and the mask label \\(\\mathbf{y}\\) as depicted in Figure 1. We denote the \\(\\mathbf{I} = \\{\\mathbf{F}, \\mathbf{F}_p, \\mathbf{P}, \\mathbf{y}\\}\\) as the set of in-context instructions derived from reference samples and the target image. A query-based prompt generator (QPG) is implemented to support a scalable structure and the ability to support different numbers of examples flexibly. Considering computational efficiency, we first concat them in channel dimension and project them into the latent space used by the QPG, then feed the reference instruction \\(\\mathbf{I}_r\\) and target instruction \\(\\mathbf{I}_t\\) to interact with learnable query tokens \\(\\mathbf{q}\\) to get the final in-context prompt \\(\\mathbf{q}_{ict}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.779, + 0.483, + 0.796 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} = \\operatorname {P r o j} (\\operatorname {C o n c a t} (\\mathbf {F}, \\mathbf {F} _ {p}, \\mathbf {P}, \\mathbf {y})) \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.799, + 0.483, + 0.815 + ], + "angle": 0, + "content": "\\[\n\\mathbf {q} _ {i c l} = Q P G \\left(\\mathbf {I} _ {r}, \\mathbf {I} _ {t}, \\mathbf {q}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.82, + 0.483, + 0.868 + ], + "angle": 0, + "content": "Finally, we feed the target image feature \\(\\mathbf{F}_t\\) with the incontext prompt \\(\\mathbf{q}_{icl}\\) to the frozen mask decoder \\(D\\) for prediction \\(\\mathbf{y}_{pred}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.874, + 0.483, + 0.892 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} _ {p r e d} = D \\left(\\mathbf {F} _ {t}, \\mathbf {q} _ {i c l}\\right) \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.092, + 0.655, + 0.106 + ], + "angle": 0, + "content": "3.6. Loss function" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.906, + 0.144 + ], + "angle": 0, + "content": "We employ a linear combination of Binary Cross-Entropy (BCE) loss and Dice loss to supervise the mask prediction:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.156, + 0.906, + 0.172 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {b c e} \\mathcal {L} _ {b c e} + \\lambda_ {d i c e} \\mathcal {L} _ {d i c e} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.184, + 0.907, + 0.259 + ], + "angle": 0, + "content": "For better optimization, we also adopt some auxiliary losses. To ensure that the learned representations are indeed Gaussian distributions, we use an explicit regularization to constrain the latent representations in the first layer of UGGN as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.268, + 0.906, + 0.31 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {K L} ^ {r e g} = \\sum_ {i = 1} ^ {N} K L (\\mathcal {N} (\\mu_ {\\mathcal {G}}, \\sigma_ {\\mathcal {G}}) | | \\mathcal {N} (\\mathbf {0}, \\mathbf {I})) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.32, + 0.906, + 0.351 + ], + "angle": 0, + "content": "Following the original GCN paper [23], we also impose \\(\\mathcal{L}_2\\) regularization on parameters of the first layer of UGGN:" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.361, + 0.905, + 0.381 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {2} ^ {r e g} = \\left\\| \\mathbf {W} _ {\\mu} ^ {(0)} \\right\\| _ {2} ^ {2} + \\left\\| \\mathbf {W} _ {\\sigma} ^ {(0)} \\right\\| _ {2} ^ {2} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.391, + 0.906, + 0.438 + ], + "angle": 0, + "content": "We find it helpful to use an auxiliary loss function in the uncertainty-guided part prototype discovery step during training to push prototypes away from each other:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.449, + 0.905, + 0.482 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {P} = \\sum_ {\\mathbf {P} _ {i}, \\mathbf {P} _ {j} \\in \\mathbf {P}} \\max ((d - d i s t (\\mathbf {P} _ {i}, \\mathbf {P} _ {j})), 0) \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.493, + 0.906, + 0.523 + ], + "angle": 0, + "content": "Combing with the graph augmentation loss in Section 3.3, the final loss is the combination of the above loss:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.548, + 0.906, + 0.582 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {K L} ^ {\\text {r e g}} \\mathcal {L} _ {K L} ^ {\\text {r e g}} + \\lambda_ {2} ^ {\\text {r e g}} \\mathcal {L} _ {2} ^ {\\text {r e g}} + \\lambda_ {P} \\mathcal {L} _ {P} + \\lambda_ {K L} ^ {\\mathcal {G}} \\mathcal {L} _ {K L} ^ {\\mathcal {G}} \\tag {14}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.595, + 0.646, + 0.612 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.621, + 0.62, + 0.635 + ], + "angle": 0, + "content": "4.1. Datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.906, + 0.702 + ], + "angle": 0, + "content": "We train UNICL-SAM with mixed semantic segmentation datasets. Notably, our approach does not necessitate any modifications to the architecture or training pipeline when incorporating an additional dataset." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.703, + 0.906, + 0.763 + ], + "angle": 0, + "content": "COCO [28] is a widely utilized dataset supporting instance segmentation, semantic segmentation, and panoptic segmentation tasks. It comprises 118K training images and 5k validation images, categorized into 80 categories." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.764, + 0.906, + 0.809 + ], + "angle": 0, + "content": "ADE20k [55] is a widely used benchmark for semantic segmentation, which consists of over 25K images sourced from diverse environments with 150 semantic categories." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.735, + 0.835 + ], + "angle": 0, + "content": "4.2. Implementation Details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We implement UNICL-SAM in PyTorch and use 8 Nvidia A6000 GPUs for our experiments. A total batch size of 64 was employed, with the model trained for 20 epochs. We merge the COCO instances for semantic segmentation and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20205" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.176, + 0.089, + 0.825, + 0.327 + ], + "angle": 0, + "content": "
MethodsResolutionTrainable ParamsCOCO-20iFSS-1000LVIS-92i
one-shotfew-shotone-shotfew-shotone-shotfew-shot
specialist model
HSNet [34]48028M41.7*50.7*86.5*88.5*17.422.9
VAT [18]41752M42.9*49.4*90.3*90.8*18.522.7
FPTrans [51]480139M56.5*65.5*----
generalist model
Painter [42]448354M33.132.661.762.310.510.9
SegGPT [43]448354M56.167.985.689.318.625.4
PerSAM [52]-2M23.5-75.6-18.4-
Matcher [29]896-52.760.787.089.633.040.0
VRP-SAM [40]5121.6M53.9-----
ICVU [38]256309M58.0-----
SINE [30]89619M64.566.1--31.235.5
UNICL-SAM51855M77.878.784.086.334.137.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.336, + 0.908, + 0.381 + ], + "angle": 0, + "content": "Table 1. Comparison with state-of-the-art specialist and generalist models on few-shot semantic segmentation benchmarks. We report both one-shot and few-shot results for comparison. * indicates that the categories in training cover the categories in testing within the same dataset. Previous state-of-the-art results among generalists are underlined." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.405, + 0.486, + 0.603 + ], + "angle": 0, + "content": "conduct the sample pools by category. We utilize DINOv2 (ViT-L/14) [36] as the default vision encoder. For the segmenter, we choose SAM (ViT-H) [24], which is pre-trained with 1B masks and 11M images, shows impressive zero-shot segmentation performance. During training, we froze the vision encoder and the segmenter, optimizing only the remaining parameters. We use AdamW [31] optimizer with a weight decay of \\(5e - 2\\) and \\(\\beta_{1} = 0.9\\), \\(\\beta_{2} = 0.95\\) for optimization. We use a base learning rate of \\(1e - 4\\) and a warmup of 4 epochs during training. Each image is cropped and resized to \\(518\\times 518\\) resolution. We randomly sampled one reference pair during training. More implementation details are provided in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.618, + 0.233, + 0.633 + ], + "angle": 0, + "content": "4.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.484, + 0.687 + ], + "angle": 0, + "content": "For the main experiments, we compare UNICL-SAM with other specialist/generalist methods on several in-context semantic segmentation benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Few-shot semantic segmentation. Following previous works [29, 30, 33, 43], we evaluate UNICL-SAM on few-shot semantic segmentation benchmarks to demonstrate its great generalization performance. Additionally, we report the in-domain performance of specialist models to ensure a fair comparison. As illustrated in Table 1, UNICL-SAM shows competitive or even better results than recent powerful in-context segmentation models. For the in-domain COCO- \\(20^{i}\\) [35] tests, UNICL-SAM outperforms previous generalist models, achieving an impressive MIoU score of 77.8 for one-shot and 78.7 for few-shot tests. For the out-of-domain FSS1000 [26] tests, UNICL-SAM also achieves competitive results that are close to specialist models, even without ever being trained on it. Furthermore, we conduct" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.402, + 0.905, + 0.523 + ], + "angle": 0, + "content": "
Multi-scale Feature AdapterPart Prototype DiscoveryUncertainty Probabilistic ModelingCOCO-20i
MIoU ↑MAE ↓
XXX71.770.039
XX✓72.990.037
X✓X72.410.038
✓X✓73.280.036
✓✓X73.780.036
X✓✓74.610.034
✓✓✓74.640.034
" + }, + { + "type": "table_caption", + "bbox": [ + 0.58, + 0.533, + 0.84, + 0.548 + ], + "angle": 0, + "content": "Table 2. Ablation of proposed components." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.577, + 0.907, + 0.653 + ], + "angle": 0, + "content": "experiments on LVIS- \\(92^{i}\\) [29], a more challenging one-shot benchmark built upon LVIS [16]. On LVIS- \\(92^{i}\\), UNICL-SAM surpasses previous generalists in a one-shot setting with a MIoU score of 34.1, demonstrating superior generalization and versatility." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.68, + 0.681 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.779 + ], + "angle": 0, + "content": "We conduct ablation experiments on each of the proposed components as follows. By default, we train the model on a small part of the COCO training set with 34k images and about \\(10\\%\\) masks for 20 epochs. We report the Mean Intersection over Union (MIoU) and the Mean Absolute Error (MAE) scores for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Ablations on proposed components. Table 2 illustrates the impact of the proposed components. The incorporation of our part prototype discovery technique contributes to improved segmentation performance, demonstrating its effectiveness in extracting contextual cues from examples. When combined with our uncertainty probabilistic modeling method, as indicated in line 6, the model's capabilities are further strengthened, resulting in a substantial increase" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20206" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.48, + 0.189 + ], + "angle": 0, + "content": "
Un-guided Graph AugmentationUn-guided Feature RefinementCOCO-20i
MIoU ↑MAE ↓
XX73.780.036
✓X73.820.036
X✓73.990.035
✓✓74.640.034
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.2, + 0.483, + 0.228 + ], + "angle": 0, + "content": "Table 3. Ablation of uncertainty guided graph augmentation and feature refinement strategies." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.246, + 0.48, + 0.332 + ], + "angle": 0, + "content": "
Graph Model TypeUn-guided Graph AugmentationCOCO-20i
GCNUGGNMIoU ↑MAE ↓
✓✗✗72.120.038
✗✓✗72.410.038
✗✓✓73.820.036
" + }, + { + "type": "table_caption", + "bbox": [ + 0.106, + 0.342, + 0.468, + 0.357 + ], + "angle": 0, + "content": "Table 4. Ablation of uncertainty-guided graph augmentation." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.374, + 0.481, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.541, + 0.483, + 0.584 + ], + "angle": 0, + "content": "Figure 3. Visualization of uncertainty maps. We apply a predefined colormap to the uncertainty estimation map for visualization. Areas of high uncertainty are marked with red bboxes \\(\\square\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.613, + 0.483, + 0.824 + ], + "angle": 0, + "content": "of \\(2.2\\mathrm{mIoU}\\). We attribute this improvement to the proposed UGGN's ability to accurately model data distribution and provide a reliable uncertainty measure, which facilitates the optimization of visual features through the two enhancement strategies. This, in turn, allows the part prototypes to yield more discriminative representations. Additionally, the integration of multi-scale adapters contributes to performance enhancement. The synergistic combination of these three components achieves optimal results, with a MIoU of 74.64 and a MAE of 0.034 on the COCO \\(20^{i}\\) benchmark. These findings underscore the effectiveness of the proposed components. It is worth noting that in the following ablation experiments unless otherwise specified, the multi-scale feature adapter is used by default." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Ablations on uncertainty probabilistic modeling. We investigate how the proposed uncertainty quantification stage and corresponding graph augmentation and feature refinement strategies affect performance. As shown in Table 3, utilizing either enhancement alone will help improve per" + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.089, + 0.895, + 0.179 + ], + "angle": 0, + "content": "
Un-guided Feature Refinement TypeCOCO-20i
MIoU ↑MAE ↓
top ratio masking43.730.117
random masking73.440.037
learnable gate74.640.034
" + }, + { + "type": "table_caption", + "bbox": [ + 0.518, + 0.189, + 0.9, + 0.205 + ], + "angle": 0, + "content": "Table 5. Ablation of uncertainty-guided feature refinement type." + }, + { + "type": "table", + "bbox": [ + 0.577, + 0.221, + 0.846, + 0.328 + ], + "angle": 0, + "content": "
Number of ClustersCOCO-20i
MIoU ↑MAE ↓
573.360.037
1073.820.036
1572.500.039
2072.380.040
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.338, + 0.905, + 0.366 + ], + "angle": 0, + "content": "Table 6. Ablation of the number of clusters of part prototype discovery." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.905, + 0.427 + ], + "angle": 0, + "content": "formance. Further combining the two strategies can achieve better performance on the COCO \\(20^{i}\\) benchmark." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.907, + 0.595 + ], + "angle": 0, + "content": "We further explore the specific effects of the two enhancement strategies through detailed ablation experiments. Initially, we examine the impact of employing deterministic graph networks versus our uncertainty Gaussian graph network (UGGN). As depicted in Table 4, UGGN outperforms GCN by achieving an increase of 0.29 in mIoU, indicating that the network effectively models the distribution of graph representations. Building on this foundation, the introduction of uncertainty-guided graph augmentation technology further enhances model performance, resulting in an increase of 1.41 in MIoU scores." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.906, + 0.734 + ], + "angle": 0, + "content": "For the uncertainty-guided feature refinement strategy, we study several different feature weighting/optimization schemes as shown in Table 5. the \"top ratio masking\" means masking nodes from areas of high uncertainty according to a preset ratio; \"random masking\" stands for masking out nodes whose associated uncertainty score is higher than a random probability map as described in [48]; \"learnable gate\" denotes our learnable gating mechanism. It can be observed that our \"learnable gate\" achieved the best result." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We visualize the predicted uncertainty map for reference examples in the COCO test set. We can notice that there is often higher uncertainty at the foreground edge and small objects, as shown in the first row of Figure 3. Furthermore, for some annotation errors in COCO as shown in the second row, the model can give good instructions and the impact of these areas will be eliminated by our feature refinement strategy. The last line shows some naturally existing degradation and style changes. It can be seen that UGGN also exhibits ideal responses to these regions, effectively estimating the data variance." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20207" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.905, + 0.187 + ], + "angle": 0, + "content": "
MethodsResolutionTrainable ParamsCleanSupport ImageSupport LabelAvg. DeclineAvg. Decline Ratio
ColorBlunnessCompressionSpaceDomain shiftDeformation
SegGPT [43]448354M62.1-2.7-3.0-2.7-7.8-6.2-8.8-5.2-8.4%
Matcher [29]896-41.6-0.5-2.5-6.5-3.8-5.3-6.2-4.1-9.9%
SINE [30]89619M70.0-0.7-1.7-5.9-2.8-6.5-12.4-5.0-7.1%
UNICL-SAM51855M79.8-1.0-1.8-3.0-3.3-4.8-5.0-3.1-3.9%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.195, + 0.908, + 0.239 + ], + "angle": 0, + "content": "Table 7. Robustness comparison with advanced in-context segmentation generalists on constructed test sets. We present the MIoU score on clean data and the corresponding performance differences across various degraded data. Our UNICL-SAM outperforms these baselines in various degradation/domain shift scenarios with slight performance changes." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.261, + 0.482, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.43, + 0.483, + 0.471 + ], + "angle": 0, + "content": "Figure 4. Visualization of cluster masks. We visualize the patches associated with the same cluster using a uniform color to facilitate the observation of the characteristics of the cluster mask." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.483, + 0.606 + ], + "angle": 0, + "content": "Ablations on part prototype discovery. We study the details of part prototype discovery stage. As shown in Table 6, we conducted an ablation experiment on the hyperparameter setting of the number of part prototype clusters. We found that the model showed the best performance when the number of clusters was 10, which was also used in the main experiments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.608, + 0.483, + 0.713 + ], + "angle": 0, + "content": "We also visualize the effect of part prototypes, as shown in Figure 4. The results clearly demonstrate that semantically similar patches are effectively grouped into the same cluster. Furthermore, for multiple targets within the same category, analogous local semantics are aggregated into a single cluster, thereby enhancing the modeling of foreground representations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.345, + 0.741 + ], + "angle": 0, + "content": "4.5. Robustness Simulation Tests" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To assess the effectiveness of our method in real-world applications, we constructed a test set featuring a variety of degraded and generated examples that may be encountered in practical scenarios, as shown in Figure 5. For detailed information on the robustness simulation test settings, please refer to the Appendix. The results in Table 7 indicate that our approach significantly outperforms existing advanced in-context segmentation models, effectively addressing the diverse challenges posed by these varied scenes. UNICLSAM excels not only in enhancing segmentation perfor" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.442, + 0.908, + 0.484 + ], + "angle": 0, + "content": "Figure 5. A simple example of robustness simulation tests designed to assess the model's capabilities across various degradation and domain shift scenarios. Zoom in for a better view." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.517, + 0.907, + 0.639 + ], + "angle": 0, + "content": "mance in clean environments but also improving results in degraded conditions as well. It surpasses the previous state-of-the-art generalist average by an impressive score of 11.7 MIoU, highlighting its effectiveness across diverse conditions. Even in the most challenging robustness tests, our model's performance fluctuations remain within a narrow range, with an average decline of only 3.1 MIoU. This demonstrates its great robustness." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.661, + 0.634, + 0.677 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we point out the data distribution uncertainty in in-context segmentation and introduce UNICL-SAM, a robust in-context segmentation framework that leverages uncertainty modeling to evaluate each input example and related graph augmentation and feature refinement strategies to help alleviate the impact of areas with high data uncertainty. We also present part prototypes as a form of in-context instruction to obtain discriminative local semantics. Leveraging the effective design, UNICL-SAM achieved superior performance on in-context segmentation benchmarks, while demonstrating exceptional robustness in conducted robustness simulation tests. We hope UNICL-SAM can promote the development of in-context segmentation generalists for real-world applications." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20208" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.264, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.199 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (No.62472396, No.62121002), Anhui Provincial Science and Technology Major Project (No.2023z020006), and the advanced computing resources provided by the SuperComputing Center of USTC." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.212, + 0.188, + 0.228 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.237, + 0.484, + 0.293 + ], + "angle": 0, + "content": "[1] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros. Visual prompting via image inpainting. Advances in Neural Information Processing Systems, 35:25005-25017, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.295, + 0.484, + 0.376 + ], + "angle": 0, + "content": "[2] Benjamin Bischke, Patrick Helber, Damian Borth, and Andreas Dengel. Segmentation of imbalanced classes in satellite imagery using adaptive uncertainty weighted class loss. In IGARSS 2018-2018 IEEE International Geoscience and Remote Sensing Symposium, pages 6191-6194. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.379, + 0.485, + 0.434 + ], + "angle": 0, + "content": "[3] Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty in neural network. In International conference on machine learning, pages 1613-1622. PMLR, 2015. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.435, + 0.484, + 0.49 + ], + "angle": 0, + "content": "[4] Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty in neural network. In International conference on machine learning, pages 1613-1622. PMLR, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.492, + 0.484, + 0.603 + ], + "angle": 0, + "content": "[5] Patrik Olā Bressan, José Marcato Junior, José Augusto Correa Martins, Maximilian Jaderson de Melo, Diogo Nunes Gonçalves, Daniel Matte Freitas, Ana Paula Marques Ramos, Michelle Táis Garcia Furuya, Lucas Prado Osco, Jonathan de Andrade Silva, et al. Semantic segmentation with labeling uncertainty and class imbalance applied to vegetation mapping. International Journal of Applied Earth Observation and Geoinformation, 108:102690, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.604, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[6] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.633, + 0.483, + 0.688 + ], + "angle": 0, + "content": "[7] Jie Chang, Zhonghao Lan, Changmao Cheng, and Yichen Wei. Data uncertainty learning in face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5710-5719, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.689, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[8] Bike Chen, Wei Peng, Xiaofeng Cao, and Juha Röning. Hyperbolic uncertainty aware semantic segmentation. IEEE Transactions on Intelligent Transportation Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.732, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[9] Wei-Ting Chen, Yu-Jiet Vong, Sy-Yen Kuo, Sizhou Ma, and Jian Wang. Robustsam: Segment anything robustly on degraded images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4081-4091, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[10] Kazuki Endo, Masayuki Tanaka, and Masatoshi Okutomi. Semantic segmentation of degraded images using layer-wise feature adjustor. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3205-3213, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[11] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.237, + 0.485, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "learning. In international conference on machine learning, pages 1050-1059. PMLR, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[12] Jakob Gawlikowski, Cedrique Rovile Njieutcheu Tassi, Mohsin Ali, Jongseok Lee, Matthias Humt, Jianxiang Feng, Anna Kruspe, Rudolph Trielbel, Peter Jung, Ribana Roscher, et al. A survey of uncertainty in deep neural networks. Artificial Intelligence Review, 56(Suppl 1):1513-1589, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[13] Ethan Goan and Clinton Fookes. Uncertainty in real-time semantic segmentation on embedded systems. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4491-4501, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[14] Dazhou Guo, Yanting Pei, Kang Zheng, Hongkai Yu, Yuhang Lu, and Song Wang. Degraded image semantic segmentation with dense-gram networks. IEEE Transactions on Image Processing, 29:782-795, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[15] Hongji Guo, Hanjing Wang, and Qiang Ji. Uncertainty-guided probabilistic transformer for complex action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20052-20061, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.427 + ], + "angle": 0, + "content": "[16] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[17] Saumya Gupta, Yikai Zhang, Xiaoling Hu, Prateek Prasanna, and Chao Chen. Topology-aware uncertainty for image segmentation. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[18] Sunghwan Hong, Seokju Cho, Jisu Nam, Stephen Lin, and Seungryong Kim. Cost aggregation with 4d convolutional swim transformer for few-shot segmentation. In European Conference on Computer Vision, pages 108-126. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[19] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.907, + 0.663 + ], + "angle": 0, + "content": "[20] Kaixiang Ji, Feng Chen, Xin Guo, Yadong Xu, Jian Wang, and Jingdong Chen. Uncertainty-guided learning for improving image manipulation detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22456-22465, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[21] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[22] Soopil Kim, Philip Chikontwe, Sion An, and Sang Hyun Park. Uncertainty-aware semi-supervised few shot segmentation. Pattern Recognition, 137:109292, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[23] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[24] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[25] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20209" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3041-3050, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.219 + ], + "angle": 0, + "content": "[26] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few-shot segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2869–2878, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[27] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European conference on computer vision, pages 280-296. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.483, + 0.36 + ], + "angle": 0, + "content": "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.483, + 0.416 + ], + "angle": 0, + "content": "[29] Yang Liu, Muzhi Zhu, Hengtao Li, Hao Chen, Xinlong Wang, and Chunhua Shen. Matcher: Segment anything with one shot using all-purpose feature matching. arXiv preprint arXiv:2305.13310, 2023. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.483, + 0.473 + ], + "angle": 0, + "content": "[30] Yang Liu, Chenchen Jing, Hengtao Li, Muzhi Zhu, Hao Chen, Xinlong Wang, and Chunhua Shen. A simple image segmentation framework via in-context examples. Advances in Neural Information Processing Systems, 2024. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.475, + 0.483, + 0.503 + ], + "angle": 0, + "content": "[31] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.483, + 0.56 + ], + "angle": 0, + "content": "[32] Yao Lu, Max Bartolo, Alastair Moore, Sebastian Riedel, and Pontus Stenetorp. Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity. arXiv preprint arXiv:2104.08786, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.483, + 0.63 + ], + "angle": 0, + "content": "[33] Lingchen Meng, Shiyi Lan, Hengduo Li, Jose M Alvarez, Zuxuan Wu, and Yu-Gang Jiang. Segic: Unleashing the emergent correspondence for in-context segmentation. In European Conference on Computer Vision, pages 203-220. Springer, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[34] Juhong Min, Dahiyun Kang, and Minsu Cho. Hypercorrelation squeeze for few-shot segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6941-6952, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.483, + 0.744 + ], + "angle": 0, + "content": "[35] Khoi Nguyen and Sinisa Todorovic. Feature weighting and boosting for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 622-631, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[36] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[37] Xinru Shan and Chaoning Zhang. Robustness of segment anything model (sam) for autonomous driving in adverse weather conditions. arXiv preprint arXiv:2306.13290, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[38] Dianmo Sheng, Dongdong Chen, Zhentao Tan, Qiankun Liu, Qi Chu, Jianmin Bao, Tao Gong, Bin Liu, Shengwei Xu," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.906, + 0.147 + ], + "angle": 0, + "content": "and Nenghai Yu. Towards more unified in-context visual understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13362-13372, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.906, + 0.231 + ], + "angle": 0, + "content": "[39] Mahesh Subedar, Ranganath Krishnan, Paulo Lopez Meyer, Omesh Tickoo, and Jonathan Huang. Uncertainty-aware audiovisual activity recognition using deep bayesian variational inference. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6301-6310, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.232, + 0.905, + 0.301 + ], + "angle": 0, + "content": "[40] Yanpeng Sun, Jiahui Chen, Shan Zhang, Xinyu Zhang, Qiang Chen, Gang Zhang, Errui Ding, Jingdong Wang, and Zechao Li. Vrp-sam: Sam with visual reference prompt. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23565-23574, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[41] Matias Valdenegro-Toro. Deep sub-ensembles for fast uncertainty estimation in image classification. arXiv preprint arXiv:1910.08168, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.343, + 0.907, + 0.413 + ], + "angle": 0, + "content": "[42] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6830-6839, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.467 + ], + "angle": 0, + "content": "[43] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[44] Zhenyu Wang, Yali Li, Ye Guo, Lu Fang, and Shengjin Wang. Data-uncertainty guided multi-phase learning for semi-supervised object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4568-4577, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[45] Zhiyong Wu, Yaoxiang Wang, Jiacheng Ye, and Lingpeng Kong. Self-adaptive in-context learning: An information compression perspective for in-context example selection and ordering. arXiv preprint arXiv:2212.10375, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[46] Weihao Xia, Zhanglin Cheng, Yujiu Yang, and Jing-Hao Xue. Cooperative semantic segmentation and image restoration in adverse environmental conditions. arXiv preprint arXiv:1911.00679, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.65, + 0.905, + 0.692 + ], + "angle": 0, + "content": "[47] Chengming Xu, Chen Liu, Yikai Wang, and Yanwei Fu. Towards global optimal visual in-context learning prompt selection. arXiv preprint arXiv:2405.15279, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[48] Fan Yang, Qiang Zhai, Xin Li, Rui Huang, Ao Luo, Hong Cheng, and Deng-Ping Fan. Uncertainty-guided transformer reasoning for camouflaged object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4146-4155, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[49] Zhou Yang, Weisheng Dong, Xin Li, Jinjian Wu, Leida Li, and Guangming Shi. Self-feature distillation with uncertainty modeling for degraded image recognition. In European Conference on Computer Vision, pages 552-569. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[50] Tianyuan Yu, Da Li, Yongxin Yang, Timothy M Hospedales, and Tao Xiang. Robust person re-identification by modelling feature uncertainty. In Proceedings of the IEEE/CVF international conference on computer vision, pages 552–561, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20210" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[51] Jian-Wei Zhang, Yifan Sun, Yi Yang, and Wei Chen. Featureproxy transformer for few-shot segmentation. Advances in Neural Information Processing Systems, 35:6575-6588, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[52] Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Xianzheng Ma, Hao Dong, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[53] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. What makes good examples for visual in-context learning? Advances in Neural Information Processing Systems, 36:17773-17794, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.304 + ], + "angle": 0, + "content": "[54] Zhizheng Zhang, Cuiling Lan, Wenjun Zeng, Zhibo Chen, and Shih-Fu Chang. Uncertainty-aware few-shot image classification. arXiv preprint arXiv:2010.04525, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[55] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127:302-321, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[56] Dingyuan Zhu, Ziwei Zhang, Peng Cui, and Wenwu Zhu. Robust graph convolutional networks against adversarial attacks. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1399-1407, 2019. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20211" + } + ] +] \ No newline at end of file diff --git a/2025/UNICL-SAM_ Uncertainty-Driven In-Context Segmentation with Part Prototype Discovery/fc453b95-4cba-4bb9-95fb-6c8d34e4ccaa_origin.pdf b/2025/UNICL-SAM_ Uncertainty-Driven In-Context Segmentation with Part Prototype Discovery/fc453b95-4cba-4bb9-95fb-6c8d34e4ccaa_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1e7cbf0c34d59492584cca15328032d10428627b --- /dev/null +++ b/2025/UNICL-SAM_ Uncertainty-Driven In-Context Segmentation with Part Prototype Discovery/fc453b95-4cba-4bb9-95fb-6c8d34e4ccaa_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0350279d321893544c432508d83a0cbba6ae151221253a2cd5d53ae214d41c81 +size 3777193 diff --git a/2025/UNICL-SAM_ Uncertainty-Driven In-Context Segmentation with Part Prototype Discovery/full.md b/2025/UNICL-SAM_ Uncertainty-Driven In-Context Segmentation with Part Prototype Discovery/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e23cf189e860fd96edf8bf3f557bce47e8de1df4 --- /dev/null +++ b/2025/UNICL-SAM_ Uncertainty-Driven In-Context Segmentation with Part Prototype Discovery/full.md @@ -0,0 +1,344 @@ +# UNICL-SAM: Uncertainty-Driven In-Context Segmentation with Part Prototype Discovery + +Dianmo Sheng $^{1,4}$ , Dongdong Chen $^{2}$ , Zhentao Tan $^{1,4}$ , Qiankun Liu $^{3}$ , Qi Chu $^{1,4}$ , Tao Gong $^{1,4,\dagger}$ , Bin Liu $^{1,4}$ , Jing Han $^{5}$ , Wenbin Tu $^{5}$ , Shengwei Xu $^{6}$ , Nenghai Yu $^{1,4}$ , $^{1}$ School of Cyber Science and Technology, USTC $^{2}$ Microsoft GenAI $^{3}$ USTB $^{4}$ Anhui Province Key Laboratory of Digital Security $^{5}$ BoCom Research $^{6}$ BESTI + +dmsheng@mail.ustc.edu.cn, tgong@ustc.edu.cn, xusw@besti.edu.cn + +# Abstract + +Recent advancements in in-context segmentation generalists have demonstrated significant success in performing various image segmentation tasks using a limited number of labeled example images. However, real-world applications present challenges due to the variability of support examples, which often exhibit quality issues resulting from various sources and inaccurate labeling. How to extract more robust representations from these examples has always been one of the goals of in-context visual learning. In response, we propose UNICL-SAM, to better model the example distribution and extract robust representations to help in-context segmentation. We incorporate an uncertainty probabilistic module to quantify each example's reliability during both the training and testing phases. Utilizing this uncertainty estimation, we introduce an uncertainty-guided graph augmentation and feature refinement strategy, aimed at mitigating the impact of high-uncertainty regions to enhance the learning of robust representations. Subsequently, we construct prototypes for each example by aggregating part information, thereby creating reliable in-context instruction that effectively represents fine-grained local semantics. This approach serves as a valuable complement to traditional global pooling features. Experimental results demonstrate the effectiveness of the proposed framework, underscoring its potential for real-world applications. + +# 1. Introduction + +Inspired by the great success of in-context learning [6, 32, 45] in the field of natural language processing (NLP), recent advances in vision generalist models [29, 30, 33, 43] indicate that different levels of segmentation tasks, including instance, semantic, and video segmentation, can be unified within one framework. The remarkable performance and + +strong task generalization capabilities of these in-context segmentation models are undoubtedly promising for the research community. + +However, existing in-context segmentation generalists often assume that the data in the training sets are of high quality and treat them equally within the same framework, thereby overlooking their diversity and differences. Previous works [47, 53] have also reported fluctuations in model performance based on varying examples and investigated strategies for selecting effective in-context examples. However, these studies only screened examples during the testing phase without considering optimization methods during the training phase. The above discussion motivates us to explore how to incorporate potential distributional diversity into the training phase to develop more robust in-context segmentation models. + +In light of this, we introduce UNICL-SAM, to investigate and address the modeling of example distribution differences and further feature optimization strategies within in-context segmentation. We draw inspiration from traditional Bayesian theory [3, 11, 21], which has previously addressed the assumptions related to data distribution differences and defined corresponding uncertainty measures to quantify and assess these distributional discrepancies. This framework facilitates the optimization of the model. Our intuition suggests that data uncertainty/discrepancy also exists within the context of contextual segmentation, potentially arising from varying observations in the reference examples. Thus, an uncertainty probabilistic Gaussian graph model is proposed to capture the training data distribution and estimate uncertainty for each example during the test phase. Additionally, we utilize the estimated uncertainty during training to guide the model to focus more on reliable (low-uncertainty) regions. The designed uncertainty-guided graph augmentation module enhances graph representation via learning invariant representations across different views of the graph. Further combined with an uncertainty-guided + +feature reweighting strategy, the model yields more robust and reliable in-context indications while minimizing the influence of extraneous information. To fully harness the discriminating information delivered from in-context examples, we incorporate diverse visual instructions like previous works [33] and introduce part prototypes as a useful form of instructions to enhance in-context knowledge representation. In practical applications, the foreground regions under consideration inherently contain local semantics, which are often overlooked by simplistic global pooling methods. By effectively aggregating similar local features and enhancing the differentiation among prototypes, part prototypes can more accurately capture fine-grained semantic information. Through a well-designed framework and comprehensive experiments, our model demonstrates great performance on both clean and degraded data, effectively alleviating the impact of the aforementioned challenge. We summarize our main contribution as follows: + +- We propose an uncertainty probabilistic module to characterize the distribution of input data. We further leverage uncertainty estimation to guide the model in learning more reliable features and refining visual instructions throughout both the training and testing phases. +- We introduce part prototypes as a form of visual instruction designed to effectively extract knowledge from reference examples by aggregating local semantics. +- Our comprehensive experiments reveal that UNICL-SAM exhibits superior performance on traditional benchmarks and shows good generalization capabilities on conducted real-world simulation robustness tests. + +# 2. Related Works + +In-Context Visual Learning. Recently, applying the in-context learning paradigm to the field of computer vision has become one of the hot spots in the research community. Bar et al. [1] pioneers in-context visual learning framework utilizing image inpainting on illustrations and infographics derived from vision-related literature, which shows competencies in fundamental CV tasks. Painter [42] employs masked image modeling on continuous pixels to facilitate in-context learning and yields highly competitive outcomes on seven tasks. Following this, SegGPT [43] offers a specialized generalist aimed at addressing diverse and unlimited segmentation tasks using a similar framework. SINE [30] building upon the SegGPT framework to tackle the task ambiguity problem. Other approaches [29, 52] investigated prompting techniques of powerful segmentation model SAM [24]. Moreover, recent work SEGIC [33] shows superior performance with a large margin compared with previous works by leveraging diverse in-context instructions. + +However, the aforementioned work does not take into account the potential impact of the diversity in training and + +testing data distributions, which motivates us to develop a more robust in-context segmentation model. + +Uncertainty Estimation in computer vision. The issue of uncertainty in neural networks, which pertains to the lack of certainty estimates or under-confidence in predictions, has garnered increasing attention in recent research [7, 12, 13, 15, 20-22, 39, 48, 54]. Kendall et al. [21] delineate two primary types: epistemic uncertainty, arising from the model's limited knowledge, and aleatoric uncertainty, which is typically associated with inherent noise within the data. Nonbayesian approaches [7, 15, 44, 48] typically utilize a Gaussian distribution framework to estimate both the mean and variance of model parameters or input data, which serve as indicators of uncertainty. They effectively avoid the substantial computational overhead of the Bayesian solution. These techniques have been effectively applied to various computer vision tasks, including image classification [41, 54], semantic segmentation [2, 5, 8, 17], object detection [44, 48], person reidentification [50] and many other computer vision tasks [7, 15, 20]. + +In this work, we first claim the diverse data distributions that in-context segmentation models need to navigate. We then propose an uncertainty probabilistic network and corresponding enhancement strategies for optimization. + +Robust Segmentation. Numerous studies [9, 10, 14, 37, 46, 49] have studied how to improve the performance of segmentation models in cross-domain and quality-degraded scenarios. For example, recently, RobustSAM [9] has examined the performance deterioration of the Segment Anything Model (SAM) [24] when handling low-quality images, subsequently designing a pipeline to enforce both feature and token consistency between clean and generated degraded images. + +Building upon these insights, we propose uncertainty-guided graph augmentation and feature refinement techniques to obtain robust representations. The subsequent part prototype discovery stage provides discriminative local semantics, further enhancing the model's robustness. + +# 3. Method + +In-context segmentation aims to identify a specific segmentation task within the given $K$ in-context examples $ref = \{(\mathbf{x}_r,\mathbf{y}_r)\}_{i = 1}^K$ , where $\mathbf{x}_r\in \mathbb{R}^{3\times H\times W}$ , $\mathbf{y}_r\in \mathbb{R}^{1\times H\times W}$ are one reference image and its corresponding ground-truth label, and segment the interested area in the target image $\mathbf{x}_t\in \mathbb{R}^{3\times H\times W}$ as a binary mask $\mathbf{y}_t$ . + +# 3.1. Overview + +Our UNICL-SAM combines the advantages of previous work and further utilizes the proposed components to improve the robustness of the model. The overall pipeline is shown in Figure 1, which consists of four stages: fea- + +![](images/c4225b7b9427a311790cabd61cfe181d7dce22ddabf2a0d582ec5b1faca177a6.jpg) +Figure 1. Overview of UNICL-SAM. With the fine-grained features, UNICL-SAM performs uncertainty quantification to model the input distribution and produce the estimated uncertainty map. The corresponding uncertainty-guided graph augmentation and feature refinement stages tend to learn robust graph representations by mitigating the impact of high uncertainty areas. Further, UNICL-SAM conducts part prototype discovery to get the part prototypes with the other three kinds of in-context instructions. Finally, the in-context instructions are fed into the query prompt generator to get the in-context prompt for guiding the mask decoder to predict segmentation results. + +ture extraction, uncertainty probabilistic modeling, part prototype discovery, and final mask decoding. First, we extract and integrate multi-scale features, then compute dense correspondence between target and reference as previous works [29, 33] (Section 3.2). Then our uncertainty Gaussian graph model will provide quantified uncertainty estimation. Utilizing the uncertainty map, we further perform uncertainty-guided graph augmentation and feature refinement techniques to facilitate learning robust visual representations (Section 3.3). With the refined visual representations, part prototypes are generated as a form of in-context instruction to facilitate the extraction of more discriminative local semantics (Section 3.4). Finally, we generate incontext prompts through a prompt generator based on the rich instructions of reference and target, guiding the mask decoder model to obtain the final segmentation result (Section 3.5). We will elaborate on each stage in detail. + +# 3.2. Feature Extraction + +Building upon insights from previous works [29, 33], we establish an in-context segmentation baseline by combining the advancements of existing VFMs [24, 36], thereby unleashing the power of pre-trained foundation models. + +We use the frozen pre-trained image encoder to extract visual features of both the reference image $\mathbf{x}_r$ and the target image $\mathbf{x}_t$ , respectively. Inspired by previous works on dense prediction tasks like object detection [27] and semantic segmentation [25], we incorporate a multi-scale feature adapter to leverage information of feature maps at different scales and aggregate them to produce the final feature map + +$\mathbf{F}_r \in \mathbb{R}^{C \times H \times W}$ and $\mathbf{F}_t \in \mathbb{R}^{C \times H \times W}$ , where $C$ denotes the number of channels. + +Then, we calculate the masked reference feature as $\mathbf{F}_{rm} = \mathbf{F}_r\cdot \mathbf{y}_r$ . The patch-level similarity $\mathbf{Corr}\in$ $\mathbb{R}^{HW\times HW}$ is calculated between $\mathbf{F}_{rm}$ and $\mathbf{F}_t$ as the dense correspondence. With the correspondence map, we calculate the pseudo mask $\mathbf{y}_t^{pseudo}$ to extract the region of interest from the target feature like $\mathbf{F}_{tm}^{pseudo} = \mathbf{F}_t\cdot \mathbf{y}_t^{pseudo}$ . For details about the multi-scale feature adapter and correspondence extraction, please refer to the Appendix. + +# 3.3. Uncertainty Probabilistic Modeling + +Deep neural network's output has varying degrees of uncertainty, influenced by how the training data is collected and the model's architecture, with observation noise affecting the instantiated predictions [21]. In the context of in-context segmentation, uncertainty also arises from the quality of the in-context examples. The limited number of samples in each iteration amplifies the impact of observation noise, further compounding uncertainty, which results in poor predictions. To address this, we model the data-dependent uncertainty of the input data and leverage it to more effectively optimize the network. + +Uncertainty quantification. We design our uncertainty Gaussian graph network (UGGN) to represent the graph representation using a Gaussian distribution. Given the fine-grained feature $\mathbf{F}_m = \{\mathbf{F}_{rm},\mathbf{F}_{tm}^{pseudo}\}$ , we first construct a graph representation $\mathcal{G} = \{\mathbf{V},\mathbf{E}\}$ , regarding each patch feature as each node in the graph $\mathbf{V} = \{\mathbf{V}_i|i = 1,\dots ,N\}$ , $\mathbf{V}_i\in \mathbf{F}_m$ . For the edge information, we uti + +![](images/38e3801b3753d92f9d4c01b50831dab8e7667d907c57ba43ebbcf0fe964906c4.jpg) +Figure 2. Illustration of the uncertainty quantification and part prototype discovery stages. UGGN serves as a probabilistic graph network, with its initial layer dedicated to uncertainty quantification. Subsequent layers are designed to facilitate the generation of part prototypes. + +lize the affinity similarity as the edge weight from the $j^{th}$ node to the $k^{th}$ node is modeled by their affinity in the embedded space, which is formulated as $\mathbf{E}(j,k) = (\mathbf{V}_j)\cdot (\mathbf{V}_k)^T,\mathbf{V}_j,\mathbf{V}_k\in \mathbf{V}$ . We only retain edges with similarity greater than an established threshold $\varepsilon$ and sparse the edge connections according to a certain top ratio $r$ as $\mathbf{E} = r(\{\mathbf{E}(j,k)|\mathbf{A}_{j,k}\geqslant \varepsilon \})$ , where $\mathbf{A}\in \mathbb{R}^{N\times N}$ denotes the adjacency matrix. + +As illustrated in Figure 2 above, in the uncertainty quantification stage, what UGGN delivers the mean $\mu_{\mathcal{G}}$ and variance $\sigma_{\mathcal{G}}$ for each graph $\mathcal{G}$ . Inspired by previous work [56], we first map the graph representation into a high-dimensional Gaussian space utilizing separate learnable mean and variance parameters, then update the mean and variance based on modified node aggregation. We use $\mathbf{M}^{(l)} = [\mu_i^{(l)},\dots ,\mu_N^{(l)}]\in \mathbb{R}^{N\times d}$ and $\pmb{\Sigma}^{(l)} = [\sigma_i^{(l)},\dots ,\sigma_N^{(l)}]\in \mathbb{R}^{N\times d}$ to denote the matrix of means and variances for all nodes at layer $l\in L$ of UGGN, respectively. $d$ represents the hidden dimension. For the first layer, we adopt a fully connected layer to map as follows: + +$$ +\mathbf {M} ^ {(1)} = \operatorname {E L U} \left(\mathbf {V W} _ {\mu} ^ {(0)}\right), \boldsymbol {\Sigma} ^ {(1)} = \operatorname {R e L U} \left(\mathbf {V W} _ {\sigma} ^ {(0)}\right) \quad (1) +$$ + +where $\mathbf{W}_{\mu} \in \mathbb{R}^{N \times d}$ and $\mathbf{W}_{\sigma} \in \mathbb{R}^{N \times d}$ are learnable parameters, ELU and ReLU represent for activation functions. $h \sim \mathcal{N}(\mu^{(1)}, \sigma^{(1)})$ denote the parameterized distribution. + +Then, we need to sample from the estimated Gaussian distributions. However, sampling directly from the distribution is a non-differentiable operation, which will result in the model not being optimized. Referring to previous work [4], we decided to use the reparameterization trick. Specifically, we first draw a sample $\epsilon$ from the standard Gaussian distribution $\mathcal{N}(\mathbf{0},\mathbf{I})$ randomly, and then obtain the sample by computing $\mathbf{Z} = \mu^{(1)} + \epsilon \sigma^{(1)},\mathbf{Z}\in \mathbb{R}^{N\times C}$ . According to Bayesian probability theory [3, 21], we can con + +sider $\mathbf{Z}$ as empirical samples drawn from an approximate predictive distribution. The uncertainty of the input data can be quantified by calculating channel-wise variance, expressed as $\mathbf{U}_{data} = \mathrm{Norm}(Var(\mathbf{Z}))$ , $\mathbf{U}_{data} \in \mathbb{R}^{N \times 1}$ . + +Uncertainty-guided graph augmentation. With the estimated uncertainty, we can assume that a larger variance indicates a greater deviation from the expected data distribution. We utilize a softmax function to map the $\mathbf{U}_{data}$ into a probability distribution $p_u^{node}$ , which determines the unimportance of each node in the graph. Combining with traditional graph augmentations like node dropping, and node masking, we can decide whether to drop or mask a node based on the its uncertainty. + +Furthermore, we compute the edge uncertainty using the affinity matrix of $\mathbf{U}_{data}$ , and map to a probability distribution $p_u^{edge}$ as well. Intuitively, if the uncertainty at the edge's endpoints is higher, the uncertainty of the edge itself will be higher. This allows us to perform edge perturbation and subgraph transformation on the graph. With the augmented graph representation $\mathcal{G}_{aug}$ and its vanilla view $\mathcal{G}$ , we use the UGGN to map them into the same representation space, and minimize their KL divergence to mitigate the influence of high-uncertainty nodes, thereby learning a more robust graph representation: + +$$ +\mathcal {L} _ {K L} ^ {\mathcal {G}} = \sum_ {i = 1} ^ {N} K L (\mathcal {N} \left(\mu_ {\mathcal {G}}, \sigma_ {\mathcal {G}}\right) | | \mathcal {N} \left(\mu_ {\mathcal {G} _ {a u g}}, \sigma_ {\mathcal {G} _ {a u g}}\right)) \tag {2} +$$ + +Uncertainty-guided feature refinement. On the other hand, the uncertainty estimation can be mapped back to spatial space as an uncertainty map $\mathbf{U}_{map} \in \mathbf{R}^{H \times W \times 1}$ . We directly compute the dot product of reverse uncertainty map $\mathbf{1} - \mathbf{U}_{map}$ with the visual features $\mathbf{F}$ and their associated labels $\mathbf{y}$ for refinement, which will effectively alleviate the impact of high uncertainty positions, resulting in more accurate foreground representations. + +Since the uncertainty estimation may be less reliable in the early stages of training, we introduce an uncertainty gate to adjust the weights for uncertainty refinement dynamically. This gate is initialized with a small positive scalar to mitigate any initial unreliability. + +# 3.4. Part prototype Discovery + +Despite we obtain a refined representation, the foreground region of example retains local semantics, resulting in varying degrees of correspondence with the target across different areas. To extract more discriminative representations, we propose the implementation of part prototype discovery. + +As shown in Figure 2 below, we can treat the last layers of UGGN as an extension by converting a deterministic graph clustering network to a probabilistic type, i.e., a parameterized distribution for feature clustering. Then, we propose an attention mechanism for the next layers to + +assign different weights to neighbors based on their variances since larger variances indicate more uncertainties in the latent representations. A smooth exponential function is utilized to control the effect of variances on weights $\mathcal{A}^{(l)} = \exp (-\gamma \Sigma^{(l)})$ + +$$ +\mathbf {M} ^ {(l + 1)} = \tilde {\mathbf {D}} ^ {- \frac {1}{2}} \mathbf {A} \tilde {\mathbf {D}} ^ {- \frac {1}{2}} \left(\mathbf {M} ^ {(l)} \odot \mathcal {A} ^ {(l)}\right) \mathbf {W} _ {\mu} ^ {(l)} \tag {3} +$$ + +$$ +\boldsymbol {\Sigma} ^ {(l + 1)} = \tilde {\mathbf {D}} ^ {- \frac {1}{2}} \mathbf {A} \tilde {\mathbf {D}} ^ {- \frac {1}{2}} \left(\boldsymbol {\Sigma} ^ {(l)} \odot \mathcal {A} ^ {(l)} \odot \mathcal {A} ^ {(l)}\right) \mathbf {W} _ {\sigma} ^ {(l)} \tag {4} +$$ + +where $\mathbf{D}$ is the diagonal degree matrix calculated by $\mathbf{D}_{i,i} = \sum_{j}\mathbf{A}_{i,j}$ , and $\tilde{\mathbf{D}} = \mathbf{D} + \mathbf{I}_N$ , where $\mathbf{I}_N$ is the identity matrix. + +We also sampled the updated distribution using the reparameterization trick to obtain the final node classification results. To solve the non-differentiability caused by the arg max operation to derive classification labels, we utilize the Gumbel-Softmax trick [19] to get the final one-hot like clusters $\mathbf{C}\mathbf{u} \in \mathbb{R}^{N \times c}$ as follows: + +$$ +\mathbf {C l u} _ {i} = \frac {\exp \left(\left(\log \left(\pi_ {i}\right) + g _ {i}\right) / \tau\right)}{\sum_ {j = 1} ^ {c} \exp \left(\left(\log \left(\pi_ {j}\right) + g _ {j}\right) / \tau\right)}, \quad \text {f o r} i = 1, \dots , c. \tag {5} +$$ + +where $g_{i},\dots ,g_{c}$ are i.i.d samples drawn from the Gumbel(0,1) distribution, $\pi$ is a categorical distribution with class probabilities and $\tau$ is the softmax temperature. + +With the clusters $\mathbf{C}\mathbf{u}$ , we get part prototype representations $\mathbf{P} = \{\mathbf{P}_i\in \mathbb{R}^C\mid i = 1,\dots ,c\}$ by averaging $n$ node features in the same cluster: + +$$ +\mathbf {P} _ {i} = \frac {1}{n} \mathbf {C l u} _ {i} ^ {\top} \cdot \mathbf {F}, \quad \mathbf {F} \in \mathbf {F} _ {m} \tag {6} +$$ + +# 3.5. Mask Decoding + +To further interact with the mask decoder for mask prediction, we extracted four kinds of visual instructions for in-context prompt generation: visual features $\mathbf{F}$ , mask average pooling features $\mathbf{F}_p = \mathrm{Avg\_pool}(\mathbf{F})$ , $\mathbf{F} \in \mathbf{F}_m$ , part prototypes $\mathbf{P}$ , and the mask label $\mathbf{y}$ as depicted in Figure 1. We denote the $\mathbf{I} = \{\mathbf{F}, \mathbf{F}_p, \mathbf{P}, \mathbf{y}\}$ as the set of in-context instructions derived from reference samples and the target image. A query-based prompt generator (QPG) is implemented to support a scalable structure and the ability to support different numbers of examples flexibly. Considering computational efficiency, we first concat them in channel dimension and project them into the latent space used by the QPG, then feed the reference instruction $\mathbf{I}_r$ and target instruction $\mathbf{I}_t$ to interact with learnable query tokens $\mathbf{q}$ to get the final in-context prompt $\mathbf{q}_{ict}$ : + +$$ +\mathbf {I} = \operatorname {P r o j} (\operatorname {C o n c a t} (\mathbf {F}, \mathbf {F} _ {p}, \mathbf {P}, \mathbf {y})) \tag {7} +$$ + +$$ +\mathbf {q} _ {i c l} = Q P G \left(\mathbf {I} _ {r}, \mathbf {I} _ {t}, \mathbf {q}\right) \tag {8} +$$ + +Finally, we feed the target image feature $\mathbf{F}_t$ with the incontext prompt $\mathbf{q}_{icl}$ to the frozen mask decoder $D$ for prediction $\mathbf{y}_{pred}$ as follows: + +$$ +\mathbf {y} _ {p r e d} = D \left(\mathbf {F} _ {t}, \mathbf {q} _ {i c l}\right) \tag {9} +$$ + +# 3.6. Loss function + +We employ a linear combination of Binary Cross-Entropy (BCE) loss and Dice loss to supervise the mask prediction: + +$$ +\mathcal {L} _ {\text {m a s k}} = \lambda_ {b c e} \mathcal {L} _ {b c e} + \lambda_ {d i c e} \mathcal {L} _ {d i c e} \tag {10} +$$ + +For better optimization, we also adopt some auxiliary losses. To ensure that the learned representations are indeed Gaussian distributions, we use an explicit regularization to constrain the latent representations in the first layer of UGGN as follows: + +$$ +\mathcal {L} _ {K L} ^ {r e g} = \sum_ {i = 1} ^ {N} K L (\mathcal {N} (\mu_ {\mathcal {G}}, \sigma_ {\mathcal {G}}) | | \mathcal {N} (\mathbf {0}, \mathbf {I})) \tag {11} +$$ + +Following the original GCN paper [23], we also impose $\mathcal{L}_2$ regularization on parameters of the first layer of UGGN: + +$$ +\mathcal {L} _ {2} ^ {r e g} = \left\| \mathbf {W} _ {\mu} ^ {(0)} \right\| _ {2} ^ {2} + \left\| \mathbf {W} _ {\sigma} ^ {(0)} \right\| _ {2} ^ {2} \tag {12} +$$ + +We find it helpful to use an auxiliary loss function in the uncertainty-guided part prototype discovery step during training to push prototypes away from each other: + +$$ +\mathcal {L} _ {P} = \sum_ {\mathbf {P} _ {i}, \mathbf {P} _ {j} \in \mathbf {P}} \max ((d - d i s t (\mathbf {P} _ {i}, \mathbf {P} _ {j})), 0) \tag {13} +$$ + +Combing with the graph augmentation loss in Section 3.3, the final loss is the combination of the above loss: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {m a s k}} + \lambda_ {K L} ^ {\text {r e g}} \mathcal {L} _ {K L} ^ {\text {r e g}} + \lambda_ {2} ^ {\text {r e g}} \mathcal {L} _ {2} ^ {\text {r e g}} + \lambda_ {P} \mathcal {L} _ {P} + \lambda_ {K L} ^ {\mathcal {G}} \mathcal {L} _ {K L} ^ {\mathcal {G}} \tag {14} +$$ + +# 4. Experiments + +# 4.1. Datasets. + +We train UNICL-SAM with mixed semantic segmentation datasets. Notably, our approach does not necessitate any modifications to the architecture or training pipeline when incorporating an additional dataset. + +COCO [28] is a widely utilized dataset supporting instance segmentation, semantic segmentation, and panoptic segmentation tasks. It comprises 118K training images and 5k validation images, categorized into 80 categories. + +ADE20k [55] is a widely used benchmark for semantic segmentation, which consists of over 25K images sourced from diverse environments with 150 semantic categories. + +# 4.2. Implementation Details. + +We implement UNICL-SAM in PyTorch and use 8 Nvidia A6000 GPUs for our experiments. A total batch size of 64 was employed, with the model trained for 20 epochs. We merge the COCO instances for semantic segmentation and + +
MethodsResolutionTrainable ParamsCOCO-20iFSS-1000LVIS-92i
one-shotfew-shotone-shotfew-shotone-shotfew-shot
specialist model
HSNet [34]48028M41.7*50.7*86.5*88.5*17.422.9
VAT [18]41752M42.9*49.4*90.3*90.8*18.522.7
FPTrans [51]480139M56.5*65.5*----
generalist model
Painter [42]448354M33.132.661.762.310.510.9
SegGPT [43]448354M56.167.985.689.318.625.4
PerSAM [52]-2M23.5-75.6-18.4-
Matcher [29]896-52.760.787.089.633.040.0
VRP-SAM [40]5121.6M53.9-----
ICVU [38]256309M58.0-----
SINE [30]89619M64.566.1--31.235.5
UNICL-SAM51855M77.878.784.086.334.137.4
+ +conduct the sample pools by category. We utilize DINOv2 (ViT-L/14) [36] as the default vision encoder. For the segmenter, we choose SAM (ViT-H) [24], which is pre-trained with 1B masks and 11M images, shows impressive zero-shot segmentation performance. During training, we froze the vision encoder and the segmenter, optimizing only the remaining parameters. We use AdamW [31] optimizer with a weight decay of $5e - 2$ and $\beta_{1} = 0.9$ , $\beta_{2} = 0.95$ for optimization. We use a base learning rate of $1e - 4$ and a warmup of 4 epochs during training. Each image is cropped and resized to $518\times 518$ resolution. We randomly sampled one reference pair during training. More implementation details are provided in the Appendix. + +# 4.3. Main Results + +For the main experiments, we compare UNICL-SAM with other specialist/generalist methods on several in-context semantic segmentation benchmarks. + +Few-shot semantic segmentation. Following previous works [29, 30, 33, 43], we evaluate UNICL-SAM on few-shot semantic segmentation benchmarks to demonstrate its great generalization performance. Additionally, we report the in-domain performance of specialist models to ensure a fair comparison. As illustrated in Table 1, UNICL-SAM shows competitive or even better results than recent powerful in-context segmentation models. For the in-domain COCO- $20^{i}$ [35] tests, UNICL-SAM outperforms previous generalist models, achieving an impressive MIoU score of 77.8 for one-shot and 78.7 for few-shot tests. For the out-of-domain FSS1000 [26] tests, UNICL-SAM also achieves competitive results that are close to specialist models, even without ever being trained on it. Furthermore, we conduct + +Table 1. Comparison with state-of-the-art specialist and generalist models on few-shot semantic segmentation benchmarks. We report both one-shot and few-shot results for comparison. * indicates that the categories in training cover the categories in testing within the same dataset. Previous state-of-the-art results among generalists are underlined. + +
Multi-scale Feature AdapterPart Prototype DiscoveryUncertainty Probabilistic ModelingCOCO-20i
MIoU ↑MAE ↓
XXX71.770.039
XX✓72.990.037
X✓X72.410.038
✓X✓73.280.036
✓✓X73.780.036
X✓✓74.610.034
✓✓✓74.640.034
+ +Table 2. Ablation of proposed components. + +experiments on LVIS- $92^{i}$ [29], a more challenging one-shot benchmark built upon LVIS [16]. On LVIS- $92^{i}$ , UNICL-SAM surpasses previous generalists in a one-shot setting with a MIoU score of 34.1, demonstrating superior generalization and versatility. + +# 4.4. Ablation Studies + +We conduct ablation experiments on each of the proposed components as follows. By default, we train the model on a small part of the COCO training set with 34k images and about $10\%$ masks for 20 epochs. We report the Mean Intersection over Union (MIoU) and the Mean Absolute Error (MAE) scores for evaluation. + +Ablations on proposed components. Table 2 illustrates the impact of the proposed components. The incorporation of our part prototype discovery technique contributes to improved segmentation performance, demonstrating its effectiveness in extracting contextual cues from examples. When combined with our uncertainty probabilistic modeling method, as indicated in line 6, the model's capabilities are further strengthened, resulting in a substantial increase + +
Un-guided Graph AugmentationUn-guided Feature RefinementCOCO-20i
MIoU ↑MAE ↓
XX73.780.036
✓X73.820.036
X✓73.990.035
✓✓74.640.034
+ +Table 3. Ablation of uncertainty guided graph augmentation and feature refinement strategies. + +
Graph Model TypeUn-guided Graph AugmentationCOCO-20i
GCNUGGNMIoU ↑MAE ↓
✓✗✗72.120.038
✗✓✗72.410.038
✗✓✓73.820.036
+ +![](images/26434fe31582abdac61e0e5f04a49fff0a358735f2828392efc5b86f9147cb7c.jpg) +Figure 3. Visualization of uncertainty maps. We apply a predefined colormap to the uncertainty estimation map for visualization. Areas of high uncertainty are marked with red bboxes $\square$ . + +of $2.2\mathrm{mIoU}$ . We attribute this improvement to the proposed UGGN's ability to accurately model data distribution and provide a reliable uncertainty measure, which facilitates the optimization of visual features through the two enhancement strategies. This, in turn, allows the part prototypes to yield more discriminative representations. Additionally, the integration of multi-scale adapters contributes to performance enhancement. The synergistic combination of these three components achieves optimal results, with a MIoU of 74.64 and a MAE of 0.034 on the COCO $20^{i}$ benchmark. These findings underscore the effectiveness of the proposed components. It is worth noting that in the following ablation experiments unless otherwise specified, the multi-scale feature adapter is used by default. + +Ablations on uncertainty probabilistic modeling. We investigate how the proposed uncertainty quantification stage and corresponding graph augmentation and feature refinement strategies affect performance. As shown in Table 3, utilizing either enhancement alone will help improve per + +Table 4. Ablation of uncertainty-guided graph augmentation. + +
Un-guided Feature Refinement TypeCOCO-20i
MIoU ↑MAE ↓
top ratio masking43.730.117
random masking73.440.037
learnable gate74.640.034
+ +Table 5. Ablation of uncertainty-guided feature refinement type. + +
Number of ClustersCOCO-20i
MIoU ↑MAE ↓
573.360.037
1073.820.036
1572.500.039
2072.380.040
+ +Table 6. Ablation of the number of clusters of part prototype discovery. + +formance. Further combining the two strategies can achieve better performance on the COCO $20^{i}$ benchmark. + +We further explore the specific effects of the two enhancement strategies through detailed ablation experiments. Initially, we examine the impact of employing deterministic graph networks versus our uncertainty Gaussian graph network (UGGN). As depicted in Table 4, UGGN outperforms GCN by achieving an increase of 0.29 in mIoU, indicating that the network effectively models the distribution of graph representations. Building on this foundation, the introduction of uncertainty-guided graph augmentation technology further enhances model performance, resulting in an increase of 1.41 in MIoU scores. + +For the uncertainty-guided feature refinement strategy, we study several different feature weighting/optimization schemes as shown in Table 5. the "top ratio masking" means masking nodes from areas of high uncertainty according to a preset ratio; "random masking" stands for masking out nodes whose associated uncertainty score is higher than a random probability map as described in [48]; "learnable gate" denotes our learnable gating mechanism. It can be observed that our "learnable gate" achieved the best result. + +We visualize the predicted uncertainty map for reference examples in the COCO test set. We can notice that there is often higher uncertainty at the foreground edge and small objects, as shown in the first row of Figure 3. Furthermore, for some annotation errors in COCO as shown in the second row, the model can give good instructions and the impact of these areas will be eliminated by our feature refinement strategy. The last line shows some naturally existing degradation and style changes. It can be seen that UGGN also exhibits ideal responses to these regions, effectively estimating the data variance. + +
MethodsResolutionTrainable ParamsCleanSupport ImageSupport LabelAvg. DeclineAvg. Decline Ratio
ColorBlunnessCompressionSpaceDomain shiftDeformation
SegGPT [43]448354M62.1-2.7-3.0-2.7-7.8-6.2-8.8-5.2-8.4%
Matcher [29]896-41.6-0.5-2.5-6.5-3.8-5.3-6.2-4.1-9.9%
SINE [30]89619M70.0-0.7-1.7-5.9-2.8-6.5-12.4-5.0-7.1%
UNICL-SAM51855M79.8-1.0-1.8-3.0-3.3-4.8-5.0-3.1-3.9%
+ +Table 7. Robustness comparison with advanced in-context segmentation generalists on constructed test sets. We present the MIoU score on clean data and the corresponding performance differences across various degraded data. Our UNICL-SAM outperforms these baselines in various degradation/domain shift scenarios with slight performance changes. + +![](images/85636649746597e0b7ddb75e77ac458b3014e03d7bb283ecd78b5e5ffd02e2fa.jpg) +Figure 4. Visualization of cluster masks. We visualize the patches associated with the same cluster using a uniform color to facilitate the observation of the characteristics of the cluster mask. + +Ablations on part prototype discovery. We study the details of part prototype discovery stage. As shown in Table 6, we conducted an ablation experiment on the hyperparameter setting of the number of part prototype clusters. We found that the model showed the best performance when the number of clusters was 10, which was also used in the main experiments. + +We also visualize the effect of part prototypes, as shown in Figure 4. The results clearly demonstrate that semantically similar patches are effectively grouped into the same cluster. Furthermore, for multiple targets within the same category, analogous local semantics are aggregated into a single cluster, thereby enhancing the modeling of foreground representations. + +# 4.5. Robustness Simulation Tests + +To assess the effectiveness of our method in real-world applications, we constructed a test set featuring a variety of degraded and generated examples that may be encountered in practical scenarios, as shown in Figure 5. For detailed information on the robustness simulation test settings, please refer to the Appendix. The results in Table 7 indicate that our approach significantly outperforms existing advanced in-context segmentation models, effectively addressing the diverse challenges posed by these varied scenes. UNICLSAM excels not only in enhancing segmentation perfor + +![](images/18ec0e824087f64185067bc0be2ad663e8e7ae3105b856ef5fd3af8ca25965d3.jpg) +Figure 5. A simple example of robustness simulation tests designed to assess the model's capabilities across various degradation and domain shift scenarios. Zoom in for a better view. + +mance in clean environments but also improving results in degraded conditions as well. It surpasses the previous state-of-the-art generalist average by an impressive score of 11.7 MIoU, highlighting its effectiveness across diverse conditions. Even in the most challenging robustness tests, our model's performance fluctuations remain within a narrow range, with an average decline of only 3.1 MIoU. This demonstrates its great robustness. + +# 5. Conclusion + +In this work, we point out the data distribution uncertainty in in-context segmentation and introduce UNICL-SAM, a robust in-context segmentation framework that leverages uncertainty modeling to evaluate each input example and related graph augmentation and feature refinement strategies to help alleviate the impact of areas with high data uncertainty. We also present part prototypes as a form of in-context instruction to obtain discriminative local semantics. Leveraging the effective design, UNICL-SAM achieved superior performance on in-context segmentation benchmarks, while demonstrating exceptional robustness in conducted robustness simulation tests. We hope UNICL-SAM can promote the development of in-context segmentation generalists for real-world applications. + +# 6. Acknowledgment + +This work was supported by the National Natural Science Foundation of China (No.62472396, No.62121002), Anhui Provincial Science and Technology Major Project (No.2023z020006), and the advanced computing resources provided by the SuperComputing Center of USTC. + +# References + +[1] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros. Visual prompting via image inpainting. Advances in Neural Information Processing Systems, 35:25005-25017, 2022. 2 +[2] Benjamin Bischke, Patrick Helber, Damian Borth, and Andreas Dengel. Segmentation of imbalanced classes in satellite imagery using adaptive uncertainty weighted class loss. In IGARSS 2018-2018 IEEE International Geoscience and Remote Sensing Symposium, pages 6191-6194. IEEE, 2018. 2 +[3] Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty in neural network. In International conference on machine learning, pages 1613-1622. PMLR, 2015. 1, 4 +[4] Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty in neural network. In International conference on machine learning, pages 1613-1622. PMLR, 2015. 4 +[5] Patrik Olā Bressan, José Marcato Junior, José Augusto Correa Martins, Maximilian Jaderson de Melo, Diogo Nunes Gonçalves, Daniel Matte Freitas, Ana Paula Marques Ramos, Michelle Táis Garcia Furuya, Lucas Prado Osco, Jonathan de Andrade Silva, et al. Semantic segmentation with labeling uncertainty and class imbalance applied to vegetation mapping. International Journal of Applied Earth Observation and Geoinformation, 108:102690, 2022. 2 +[6] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1 +[7] Jie Chang, Zhonghao Lan, Changmao Cheng, and Yichen Wei. Data uncertainty learning in face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5710-5719, 2020. 2 +[8] Bike Chen, Wei Peng, Xiaofeng Cao, and Juha Röning. Hyperbolic uncertainty aware semantic segmentation. IEEE Transactions on Intelligent Transportation Systems, 2023. 2 +[9] Wei-Ting Chen, Yu-Jiet Vong, Sy-Yen Kuo, Sizhou Ma, and Jian Wang. Robustsam: Segment anything robustly on degraded images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4081-4091, 2024. 2 +[10] Kazuki Endo, Masayuki Tanaka, and Masatoshi Okutomi. Semantic segmentation of degraded images using layer-wise feature adjustor. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3205-3213, 2023. 2 +[11] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep + +learning. In international conference on machine learning, pages 1050-1059. PMLR, 2016. 1 +[12] Jakob Gawlikowski, Cedrique Rovile Njieutcheu Tassi, Mohsin Ali, Jongseok Lee, Matthias Humt, Jianxiang Feng, Anna Kruspe, Rudolph Trielbel, Peter Jung, Ribana Roscher, et al. A survey of uncertainty in deep neural networks. Artificial Intelligence Review, 56(Suppl 1):1513-1589, 2023. 2 +[13] Ethan Goan and Clinton Fookes. Uncertainty in real-time semantic segmentation on embedded systems. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4491-4501, 2023. 2 +[14] Dazhou Guo, Yanting Pei, Kang Zheng, Hongkai Yu, Yuhang Lu, and Song Wang. Degraded image semantic segmentation with dense-gram networks. IEEE Transactions on Image Processing, 29:782-795, 2019. 2 +[15] Hongji Guo, Hanjing Wang, and Qiang Ji. Uncertainty-guided probabilistic transformer for complex action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20052-20061, 2022. 2 +[16] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 6 +[17] Saumya Gupta, Yikai Zhang, Xiaoling Hu, Prateek Prasanna, and Chao Chen. Topology-aware uncertainty for image segmentation. Advances in Neural Information Processing Systems, 36, 2024. 2 +[18] Sunghwan Hong, Seokju Cho, Jisu Nam, Stephen Lin, and Seungryong Kim. Cost aggregation with 4d convolutional swim transformer for few-shot segmentation. In European Conference on Computer Vision, pages 108-126. Springer, 2022. 6 +[19] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 5 +[20] Kaixiang Ji, Feng Chen, Xin Guo, Yadong Xu, Jian Wang, and Jingdong Chen. Uncertainty-guided learning for improving image manipulation detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22456-22465, 2023. 2 +[21] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 1, 2, 3, 4 +[22] Soopil Kim, Philip Chikontwe, Sion An, and Sang Hyun Park. Uncertainty-aware semi-supervised few shot segmentation. Pattern Recognition, 137:109292, 2023. 2 +[23] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 5 +[24] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2, 3, 6 +[25] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards + +a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3041-3050, 2023. 3 +[26] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few-shot segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2869–2878, 2020. 6 +[27] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European conference on computer vision, pages 280-296. Springer, 2022. 3 +[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5 +[29] Yang Liu, Muzhi Zhu, Hengtao Li, Hao Chen, Xinlong Wang, and Chunhua Shen. Matcher: Segment anything with one shot using all-purpose feature matching. arXiv preprint arXiv:2305.13310, 2023. 1, 2, 3, 6, 8 +[30] Yang Liu, Chenchen Jing, Hengtao Li, Muzhi Zhu, Hao Chen, Xinlong Wang, and Chunhua Shen. A simple image segmentation framework via in-context examples. Advances in Neural Information Processing Systems, 2024. 1, 2, 6, 8 +[31] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[32] Yao Lu, Max Bartolo, Alastair Moore, Sebastian Riedel, and Pontus Stenetorp. Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity. arXiv preprint arXiv:2104.08786, 2021. 1 +[33] Lingchen Meng, Shiyi Lan, Hengduo Li, Jose M Alvarez, Zuxuan Wu, and Yu-Gang Jiang. Segic: Unleashing the emergent correspondence for in-context segmentation. In European Conference on Computer Vision, pages 203-220. Springer, 2024. 1, 2, 3, 6 +[34] Juhong Min, Dahiyun Kang, and Minsu Cho. Hypercorrelation squeeze for few-shot segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6941-6952, 2021. 6 +[35] Khoi Nguyen and Sinisa Todorovic. Feature weighting and boosting for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 622-631, 2019. 6 +[36] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3, 6 +[37] Xinru Shan and Chaoning Zhang. Robustness of segment anything model (sam) for autonomous driving in adverse weather conditions. arXiv preprint arXiv:2306.13290, 2023. 2 +[38] Dianmo Sheng, Dongdong Chen, Zhentao Tan, Qiankun Liu, Qi Chu, Jianmin Bao, Tao Gong, Bin Liu, Shengwei Xu, + +and Nenghai Yu. Towards more unified in-context visual understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13362-13372, 2024. 6 +[39] Mahesh Subedar, Ranganath Krishnan, Paulo Lopez Meyer, Omesh Tickoo, and Jonathan Huang. Uncertainty-aware audiovisual activity recognition using deep bayesian variational inference. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6301-6310, 2019. 2 +[40] Yanpeng Sun, Jiahui Chen, Shan Zhang, Xinyu Zhang, Qiang Chen, Gang Zhang, Errui Ding, Jingdong Wang, and Zechao Li. Vrp-sam: Sam with visual reference prompt. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23565-23574, 2024. 6 +[41] Matias Valdenegro-Toro. Deep sub-ensembles for fast uncertainty estimation in image classification. arXiv preprint arXiv:1910.08168, 2019. 2 +[42] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6830-6839, 2023. 2, 6 +[43] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023. 1, 2, 6, 8 +[44] Zhenyu Wang, Yali Li, Ye Guo, Lu Fang, and Shengjin Wang. Data-uncertainty guided multi-phase learning for semi-supervised object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4568-4577, 2021. 2 +[45] Zhiyong Wu, Yaoxiang Wang, Jiacheng Ye, and Lingpeng Kong. Self-adaptive in-context learning: An information compression perspective for in-context example selection and ordering. arXiv preprint arXiv:2212.10375, 2022. 1 +[46] Weihao Xia, Zhanglin Cheng, Yujiu Yang, and Jing-Hao Xue. Cooperative semantic segmentation and image restoration in adverse environmental conditions. arXiv preprint arXiv:1911.00679, 2019. 2 +[47] Chengming Xu, Chen Liu, Yikai Wang, and Yanwei Fu. Towards global optimal visual in-context learning prompt selection. arXiv preprint arXiv:2405.15279, 2024. 1 +[48] Fan Yang, Qiang Zhai, Xin Li, Rui Huang, Ao Luo, Hong Cheng, and Deng-Ping Fan. Uncertainty-guided transformer reasoning for camouflaged object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4146-4155, 2021. 2, 7 +[49] Zhou Yang, Weisheng Dong, Xin Li, Jinjian Wu, Leida Li, and Guangming Shi. Self-feature distillation with uncertainty modeling for degraded image recognition. In European Conference on Computer Vision, pages 552-569. Springer, 2022. 2 +[50] Tianyuan Yu, Da Li, Yongxin Yang, Timothy M Hospedales, and Tao Xiang. Robust person re-identification by modelling feature uncertainty. In Proceedings of the IEEE/CVF international conference on computer vision, pages 552–561, 2019. 2 + +[51] Jian-Wei Zhang, Yifan Sun, Yi Yang, and Wei Chen. Featureproxy transformer for few-shot segmentation. Advances in Neural Information Processing Systems, 35:6575-6588, 2022. 6 +[52] Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Xianzheng Ma, Hao Dong, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048, 2023. 2, 6 +[53] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. What makes good examples for visual in-context learning? Advances in Neural Information Processing Systems, 36:17773-17794, 2023. 1 +[54] Zhizheng Zhang, Cuiling Lan, Wenjun Zeng, Zhibo Chen, and Shih-Fu Chang. Uncertainty-aware few-shot image classification. arXiv preprint arXiv:2010.04525, 2020. 2 +[55] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127:302-321, 2019. 5 +[56] Dingyuan Zhu, Ziwei Zhang, Peng Cui, and Wenwu Zhu. Robust graph convolutional networks against adversarial attacks. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1399-1407, 2019. 4 \ No newline at end of file diff --git a/2025/UNICL-SAM_ Uncertainty-Driven In-Context Segmentation with Part Prototype Discovery/images.zip b/2025/UNICL-SAM_ Uncertainty-Driven In-Context Segmentation with Part Prototype Discovery/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d197bdcc654181796d8e2feb25b8d30f2bafdde9 --- /dev/null +++ b/2025/UNICL-SAM_ Uncertainty-Driven In-Context Segmentation with Part Prototype Discovery/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d822544c5a500efd2f178d3460b258ef446b4348e50d08e0e56a74dd71f6be1 +size 534582 diff --git a/2025/UNICL-SAM_ Uncertainty-Driven In-Context Segmentation with Part Prototype Discovery/layout.json b/2025/UNICL-SAM_ Uncertainty-Driven In-Context Segmentation with Part Prototype Discovery/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..228e745a37ad850137846856efe080abe9d9fdd5 --- /dev/null +++ b/2025/UNICL-SAM_ Uncertainty-Driven In-Context Segmentation with Part Prototype Discovery/layout.json @@ -0,0 +1,9285 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 120, + 103, + 492, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 103, + 492, + 140 + ], + "spans": [ + { + "bbox": [ + 120, + 103, + 492, + 140 + ], + "type": "text", + "content": "UNICL-SAM: Uncertainty-Driven In-Context Segmentation with Part Prototype Discovery" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "spans": [ + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "text", + "content": "Dianmo Sheng" + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "text", + "content": ", Dongdong Chen" + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "text", + "content": ", Zhentao Tan" + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "text", + "content": ", Qiankun Liu" + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "text", + "content": ", Qi Chu" + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "text", + "content": ", Tao Gong" + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "inline_equation", + "content": "^{1,4,\\dagger}" + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "text", + "content": ", Bin Liu" + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "text", + "content": ", Jing Han" + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "text", + "content": ", Wenbin Tu" + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "text", + "content": ", Shengwei Xu" + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "text", + "content": ", Nenghai Yu" + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "text", + "content": "School of Cyber Science and Technology, USTC " + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "text", + "content": "Microsoft GenAI " + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "text", + "content": "USTB " + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "text", + "content": "Anhui Province Key Laboratory of Digital Security " + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "text", + "content": "BoCom Research " + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 88, + 160, + 534, + 219 + ], + "type": "text", + "content": "BESTI" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 220, + 473, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 220, + 473, + 231 + ], + "spans": [ + { + "bbox": [ + 134, + 220, + 473, + 231 + ], + "type": "text", + "content": "dmsheng@mail.ustc.edu.cn, tgong@ustc.edu.cn, xusw@besti.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 284, + 297, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 297, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 297, + 583 + ], + "type": "text", + "content": "Recent advancements in in-context segmentation generalists have demonstrated significant success in performing various image segmentation tasks using a limited number of labeled example images. However, real-world applications present challenges due to the variability of support examples, which often exhibit quality issues resulting from various sources and inaccurate labeling. How to extract more robust representations from these examples has always been one of the goals of in-context visual learning. In response, we propose UNICL-SAM, to better model the example distribution and extract robust representations to help in-context segmentation. We incorporate an uncertainty probabilistic module to quantify each example's reliability during both the training and testing phases. Utilizing this uncertainty estimation, we introduce an uncertainty-guided graph augmentation and feature refinement strategy, aimed at mitigating the impact of high-uncertainty regions to enhance the learning of robust representations. Subsequently, we construct prototypes for each example by aggregating part information, thereby creating reliable in-context instruction that effectively represents fine-grained local semantics. This approach serves as a valuable complement to traditional global pooling features. Experimental results demonstrate the effectiveness of the proposed framework, underscoring its potential for real-world applications." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 604, + 135, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 135, + 617 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 135, + 617 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 624, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 296, + 696 + ], + "type": "text", + "content": "Inspired by the great success of in-context learning [6, 32, 45] in the field of natural language processing (NLP), recent advances in vision generalist models [29, 30, 33, 43] indicate that different levels of segmentation tasks, including instance, semantic, and video segmentation, can be unified within one framework. The remarkable performance and" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 260, + 555, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 555, + 296 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 555, + 296 + ], + "type": "text", + "content": "strong task generalization capabilities of these in-context segmentation models are undoubtedly promising for the research community." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 301, + 556, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 556, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 556, + 456 + ], + "type": "text", + "content": "However, existing in-context segmentation generalists often assume that the data in the training sets are of high quality and treat them equally within the same framework, thereby overlooking their diversity and differences. Previous works [47, 53] have also reported fluctuations in model performance based on varying examples and investigated strategies for selecting effective in-context examples. However, these studies only screened examples during the testing phase without considering optimization methods during the training phase. The above discussion motivates us to explore how to incorporate potential distributional diversity into the training phase to develop more robust in-context segmentation models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 462, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 556, + 715 + ], + "type": "text", + "content": "In light of this, we introduce UNICL-SAM, to investigate and address the modeling of example distribution differences and further feature optimization strategies within in-context segmentation. We draw inspiration from traditional Bayesian theory [3, 11, 21], which has previously addressed the assumptions related to data distribution differences and defined corresponding uncertainty measures to quantify and assess these distributional discrepancies. This framework facilitates the optimization of the model. Our intuition suggests that data uncertainty/discrepancy also exists within the context of contextual segmentation, potentially arising from varying observations in the reference examples. Thus, an uncertainty probabilistic Gaussian graph model is proposed to capture the training data distribution and estimate uncertainty for each example during the test phase. Additionally, we utilize the estimated uncertainty during training to guide the model to focus more on reliable (low-uncertainty) regions. The designed uncertainty-guided graph augmentation module enhances graph representation via learning invariant representations across different views of the graph. Further combined with an uncertainty-guided" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 150, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 150, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 150, + 713 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "20201" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 286 + ], + "type": "text", + "content": "feature reweighting strategy, the model yields more robust and reliable in-context indications while minimizing the influence of extraneous information. To fully harness the discriminating information delivered from in-context examples, we incorporate diverse visual instructions like previous works [33] and introduce part prototypes as a useful form of instructions to enhance in-context knowledge representation. In practical applications, the foreground regions under consideration inherently contain local semantics, which are often overlooked by simplistic global pooling methods. By effectively aggregating similar local features and enhancing the differentiation among prototypes, part prototypes can more accurately capture fine-grained semantic information. Through a well-designed framework and comprehensive experiments, our model demonstrates great performance on both clean and degraded data, effectively alleviating the impact of the aforementioned challenge. We summarize our main contribution as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 288, + 296, + 430 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 55, + 288, + 295, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 295, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 295, + 347 + ], + "type": "text", + "content": "- We propose an uncertainty probabilistic module to characterize the distribution of input data. We further leverage uncertainty estimation to guide the model in learning more reliable features and refining visual instructions throughout both the training and testing phases." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 348, + 295, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 348, + 295, + 383 + ], + "spans": [ + { + "bbox": [ + 56, + 348, + 295, + 383 + ], + "type": "text", + "content": "- We introduce part prototypes as a form of visual instruction designed to effectively extract knowledge from reference examples by aggregating local semantics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 383, + 296, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 296, + 430 + ], + "type": "text", + "content": "- Our comprehensive experiments reveal that UNICL-SAM exhibits superior performance on traditional benchmarks and shows good generalization capabilities on conducted real-world simulation robustness tests." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 442, + 147, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 442, + 147, + 454 + ], + "spans": [ + { + "bbox": [ + 55, + 442, + 147, + 454 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 462, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 688 + ], + "type": "text", + "content": "In-Context Visual Learning. Recently, applying the in-context learning paradigm to the field of computer vision has become one of the hot spots in the research community. Bar et al. [1] pioneers in-context visual learning framework utilizing image inpainting on illustrations and infographics derived from vision-related literature, which shows competencies in fundamental CV tasks. Painter [42] employs masked image modeling on continuous pixels to facilitate in-context learning and yields highly competitive outcomes on seven tasks. Following this, SegGPT [43] offers a specialized generalist aimed at addressing diverse and unlimited segmentation tasks using a similar framework. SINE [30] building upon the SegGPT framework to tackle the task ambiguity problem. Other approaches [29, 52] investigated prompting techniques of powerful segmentation model SAM [24]. Moreover, recent work SEGIC [33] shows superior performance with a large margin compared with previous works by leveraging diverse in-context instructions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "However, the aforementioned work does not take into account the potential impact of the diversity in training and" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "testing data distributions, which motivates us to develop a more robust in-context segmentation model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 101, + 553, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 101, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 101, + 553, + 316 + ], + "type": "text", + "content": "Uncertainty Estimation in computer vision. The issue of uncertainty in neural networks, which pertains to the lack of certainty estimates or under-confidence in predictions, has garnered increasing attention in recent research [7, 12, 13, 15, 20-22, 39, 48, 54]. Kendall et al. [21] delineate two primary types: epistemic uncertainty, arising from the model's limited knowledge, and aleatoric uncertainty, which is typically associated with inherent noise within the data. Nonbayesian approaches [7, 15, 44, 48] typically utilize a Gaussian distribution framework to estimate both the mean and variance of model parameters or input data, which serve as indicators of uncertainty. They effectively avoid the substantial computational overhead of the Bayesian solution. These techniques have been effectively applied to various computer vision tasks, including image classification [41, 54], semantic segmentation [2, 5, 8, 17], object detection [44, 48], person reidentification [50] and many other computer vision tasks [7, 15, 20]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 316, + 553, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 553, + 365 + ], + "type": "text", + "content": "In this work, we first claim the diverse data distributions that in-context segmentation models need to navigate. We then propose an uncertainty probabilistic network and corresponding enhancement strategies for optimization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 369, + 553, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 553, + 477 + ], + "type": "text", + "content": "Robust Segmentation. Numerous studies [9, 10, 14, 37, 46, 49] have studied how to improve the performance of segmentation models in cross-domain and quality-degraded scenarios. For example, recently, RobustSAM [9] has examined the performance deterioration of the Segment Anything Model (SAM) [24] when handling low-quality images, subsequently designing a pipeline to enforce both feature and token consistency between clean and generated degraded images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 478, + 553, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 553, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 553, + 537 + ], + "type": "text", + "content": "Building upon these insights, we propose uncertainty-guided graph augmentation and feature refinement techniques to obtain robust representations. The subsequent part prototype discovery stage provides discriminative local semantics, further enhancing the model's robustness." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 548, + 370, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 548, + 370, + 560 + ], + "spans": [ + { + "bbox": [ + 314, + 548, + 370, + 560 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 568, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 553, + 640 + ], + "type": "text", + "content": "In-context segmentation aims to identify a specific segmentation task within the given " + }, + { + "bbox": [ + 313, + 568, + 553, + 640 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 568, + 553, + 640 + ], + "type": "text", + "content": " in-context examples " + }, + { + "bbox": [ + 313, + 568, + 553, + 640 + ], + "type": "inline_equation", + "content": "ref = \\{(\\mathbf{x}_r,\\mathbf{y}_r)\\}_{i = 1}^K" + }, + { + "bbox": [ + 313, + 568, + 553, + 640 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 568, + 553, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_r\\in \\mathbb{R}^{3\\times H\\times W}" + }, + { + "bbox": [ + 313, + 568, + 553, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 568, + 553, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_r\\in \\mathbb{R}^{1\\times H\\times W}" + }, + { + "bbox": [ + 313, + 568, + 553, + 640 + ], + "type": "text", + "content": " are one reference image and its corresponding ground-truth label, and segment the interested area in the target image " + }, + { + "bbox": [ + 313, + 568, + 553, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t\\in \\mathbb{R}^{3\\times H\\times W}" + }, + { + "bbox": [ + 313, + 568, + 553, + 640 + ], + "type": "text", + "content": " as a binary mask " + }, + { + "bbox": [ + 313, + 568, + 553, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_t" + }, + { + "bbox": [ + 313, + 568, + 553, + 640 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 648, + 382, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 382, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 382, + 659 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "Our UNICL-SAM combines the advantages of previous work and further utilizes the proposed components to improve the robustness of the model. The overall pipeline is shown in Figure 1, which consists of four stages: fea-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20202" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 553, + 262 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 262 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 262 + ], + "type": "image", + "image_path": "c4225b7b9427a311790cabd61cfe181d7dce22ddabf2a0d582ec5b1faca177a6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 270, + 555, + 327 + ], + "lines": [ + { + "bbox": [ + 54, + 270, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 555, + 327 + ], + "type": "text", + "content": "Figure 1. Overview of UNICL-SAM. With the fine-grained features, UNICL-SAM performs uncertainty quantification to model the input distribution and produce the estimated uncertainty map. The corresponding uncertainty-guided graph augmentation and feature refinement stages tend to learn robust graph representations by mitigating the impact of high uncertainty areas. Further, UNICL-SAM conducts part prototype discovery to get the part prototypes with the other three kinds of in-context instructions. Finally, the in-context instructions are fed into the query prompt generator to get the in-context prompt for guiding the mask decoder to predict segmentation results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 347, + 297, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 347, + 297, + 552 + ], + "spans": [ + { + "bbox": [ + 54, + 347, + 297, + 552 + ], + "type": "text", + "content": "ture extraction, uncertainty probabilistic modeling, part prototype discovery, and final mask decoding. First, we extract and integrate multi-scale features, then compute dense correspondence between target and reference as previous works [29, 33] (Section 3.2). Then our uncertainty Gaussian graph model will provide quantified uncertainty estimation. Utilizing the uncertainty map, we further perform uncertainty-guided graph augmentation and feature refinement techniques to facilitate learning robust visual representations (Section 3.3). With the refined visual representations, part prototypes are generated as a form of in-context instruction to facilitate the extraction of more discriminative local semantics (Section 3.4). Finally, we generate incontext prompts through a prompt generator based on the rich instructions of reference and target, guiding the mask decoder model to obtain the final segmentation result (Section 3.5). We will elaborate on each stage in detail." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 562, + 167, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 167, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 167, + 574 + ], + "type": "text", + "content": "3.2. Feature Extraction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 580, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 628 + ], + "type": "text", + "content": "Building upon insights from previous works [29, 33], we establish an in-context segmentation baseline by combining the advancements of existing VFMs [24, 36], thereby unleashing the power of pre-trained foundation models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": "We use the frozen pre-trained image encoder to extract visual features of both the reference image " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_r" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": " and the target image " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": ", respectively. Inspired by previous works on dense prediction tasks like object detection [27] and semantic segmentation [25], we incorporate a multi-scale feature adapter to leverage information of feature maps at different scales and aggregate them to produce the final feature map" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 346, + 555, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_r \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 313, + 346, + 555, + 370 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 346, + 555, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 313, + 346, + 555, + 370 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 346, + 555, + 370 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 346, + 555, + 370 + ], + "type": "text", + "content": " denotes the number of channels." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 372, + 556, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 556, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 556, + 469 + ], + "type": "text", + "content": "Then, we calculate the masked reference feature as " + }, + { + "bbox": [ + 313, + 372, + 556, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{rm} = \\mathbf{F}_r\\cdot \\mathbf{y}_r" + }, + { + "bbox": [ + 313, + 372, + 556, + 469 + ], + "type": "text", + "content": ". The patch-level similarity " + }, + { + "bbox": [ + 313, + 372, + 556, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{Corr}\\in" + }, + { + "bbox": [ + 313, + 372, + 556, + 469 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{HW\\times HW}" + }, + { + "bbox": [ + 313, + 372, + 556, + 469 + ], + "type": "text", + "content": " is calculated between " + }, + { + "bbox": [ + 313, + 372, + 556, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{rm}" + }, + { + "bbox": [ + 313, + 372, + 556, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 372, + 556, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t" + }, + { + "bbox": [ + 313, + 372, + 556, + 469 + ], + "type": "text", + "content": " as the dense correspondence. With the correspondence map, we calculate the pseudo mask " + }, + { + "bbox": [ + 313, + 372, + 556, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_t^{pseudo}" + }, + { + "bbox": [ + 313, + 372, + 556, + 469 + ], + "type": "text", + "content": " to extract the region of interest from the target feature like " + }, + { + "bbox": [ + 313, + 372, + 556, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{tm}^{pseudo} = \\mathbf{F}_t\\cdot \\mathbf{y}_t^{pseudo}" + }, + { + "bbox": [ + 313, + 372, + 556, + 469 + ], + "type": "text", + "content": ". For details about the multi-scale feature adapter and correspondence extraction, please refer to the Appendix." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 479, + 503, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 503, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 503, + 492 + ], + "type": "text", + "content": "3.3. Uncertainty Probabilistic Modeling" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 497, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 628 + ], + "type": "text", + "content": "Deep neural network's output has varying degrees of uncertainty, influenced by how the training data is collected and the model's architecture, with observation noise affecting the instantiated predictions [21]. In the context of in-context segmentation, uncertainty also arises from the quality of the in-context examples. The limited number of samples in each iteration amplifies the impact of observation noise, further compounding uncertainty, which results in poor predictions. To address this, we model the data-dependent uncertainty of the input data and leverage it to more effectively optimize the network." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Uncertainty quantification. We design our uncertainty Gaussian graph network (UGGN) to represent the graph representation using a Gaussian distribution. Given the fine-grained feature " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_m = \\{\\mathbf{F}_{rm},\\mathbf{F}_{tm}^{pseudo}\\}" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": ", we first construct a graph representation " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = \\{\\mathbf{V},\\mathbf{E}\\}" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": ", regarding each patch feature as each node in the graph " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = \\{\\mathbf{V}_i|i = 1,\\dots ,N\\}" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_i\\in \\mathbf{F}_m" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": ". For the edge information, we uti" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20203" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 294, + 208 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 208 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 208 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 208 + ], + "type": "image", + "image_path": "38e3801b3753d92f9d4c01b50831dab8e7667d907c57ba43ebbcf0fe964906c4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 217, + 296, + 275 + ], + "lines": [ + { + "bbox": [ + 55, + 217, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 296, + 275 + ], + "type": "text", + "content": "Figure 2. Illustration of the uncertainty quantification and part prototype discovery stages. UGGN serves as a probabilistic graph network, with its initial layer dedicated to uncertainty quantification. Subsequent layers are designed to facilitate the generation of part prototypes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 293, + 296, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 296, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 296, + 389 + ], + "type": "text", + "content": "lize the affinity similarity as the edge weight from the " + }, + { + "bbox": [ + 55, + 293, + 296, + 389 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 55, + 293, + 296, + 389 + ], + "type": "text", + "content": " node to the " + }, + { + "bbox": [ + 55, + 293, + 296, + 389 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 55, + 293, + 296, + 389 + ], + "type": "text", + "content": " node is modeled by their affinity in the embedded space, which is formulated as " + }, + { + "bbox": [ + 55, + 293, + 296, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{E}(j,k) = (\\mathbf{V}_j)\\cdot (\\mathbf{V}_k)^T,\\mathbf{V}_j,\\mathbf{V}_k\\in \\mathbf{V}" + }, + { + "bbox": [ + 55, + 293, + 296, + 389 + ], + "type": "text", + "content": ". We only retain edges with similarity greater than an established threshold " + }, + { + "bbox": [ + 55, + 293, + 296, + 389 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 55, + 293, + 296, + 389 + ], + "type": "text", + "content": " and sparse the edge connections according to a certain top ratio " + }, + { + "bbox": [ + 55, + 293, + 296, + 389 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 293, + 296, + 389 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 293, + 296, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{E} = r(\\{\\mathbf{E}(j,k)|\\mathbf{A}_{j,k}\\geqslant \\varepsilon \\})" + }, + { + "bbox": [ + 55, + 293, + 296, + 389 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 293, + 296, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{A}\\in \\mathbb{R}^{N\\times N}" + }, + { + "bbox": [ + 55, + 293, + 296, + 389 + ], + "type": "text", + "content": " denotes the adjacency matrix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 390, + 296, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 296, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 296, + 538 + ], + "type": "text", + "content": "As illustrated in Figure 2 above, in the uncertainty quantification stage, what UGGN delivers the mean " + }, + { + "bbox": [ + 55, + 390, + 296, + 538 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathcal{G}}" + }, + { + "bbox": [ + 55, + 390, + 296, + 538 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 55, + 390, + 296, + 538 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathcal{G}}" + }, + { + "bbox": [ + 55, + 390, + 296, + 538 + ], + "type": "text", + "content": " for each graph " + }, + { + "bbox": [ + 55, + 390, + 296, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 390, + 296, + 538 + ], + "type": "text", + "content": ". Inspired by previous work [56], we first map the graph representation into a high-dimensional Gaussian space utilizing separate learnable mean and variance parameters, then update the mean and variance based on modified node aggregation. We use " + }, + { + "bbox": [ + 55, + 390, + 296, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^{(l)} = [\\mu_i^{(l)},\\dots ,\\mu_N^{(l)}]\\in \\mathbb{R}^{N\\times d}" + }, + { + "bbox": [ + 55, + 390, + 296, + 538 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 390, + 296, + 538 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}^{(l)} = [\\sigma_i^{(l)},\\dots ,\\sigma_N^{(l)}]\\in \\mathbb{R}^{N\\times d}" + }, + { + "bbox": [ + 55, + 390, + 296, + 538 + ], + "type": "text", + "content": " to denote the matrix of means and variances for all nodes at layer " + }, + { + "bbox": [ + 55, + 390, + 296, + 538 + ], + "type": "inline_equation", + "content": "l\\in L" + }, + { + "bbox": [ + 55, + 390, + 296, + 538 + ], + "type": "text", + "content": " of UGGN, respectively. " + }, + { + "bbox": [ + 55, + 390, + 296, + 538 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 390, + 296, + 538 + ], + "type": "text", + "content": " represents the hidden dimension. For the first layer, we adopt a fully connected layer to map as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 78, + 544, + 295, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 544, + 295, + 562 + ], + "spans": [ + { + "bbox": [ + 78, + 544, + 295, + 562 + ], + "type": "interline_equation", + "content": "\\mathbf {M} ^ {(1)} = \\operatorname {E L U} \\left(\\mathbf {V W} _ {\\mu} ^ {(0)}\\right), \\boldsymbol {\\Sigma} ^ {(1)} = \\operatorname {R e L U} \\left(\\mathbf {V W} _ {\\sigma} ^ {(0)}\\right) \\quad (1)", + "image_path": "ec4e08ac34820cb9cb9fd5ba55e7fd5163562f8496901fae0d563533852af6a7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mu} \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\sigma} \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "text", + "content": " are learnable parameters, ELU and ReLU represent for activation functions. " + }, + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "inline_equation", + "content": "h \\sim \\mathcal{N}(\\mu^{(1)}, \\sigma^{(1)})" + }, + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "text", + "content": " denote the parameterized distribution." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 606, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 606, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 606, + 296, + 714 + ], + "type": "text", + "content": "Then, we need to sample from the estimated Gaussian distributions. However, sampling directly from the distribution is a non-differentiable operation, which will result in the model not being optimized. Referring to previous work [4], we decided to use the reparameterization trick. Specifically, we first draw a sample " + }, + { + "bbox": [ + 54, + 606, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 54, + 606, + 296, + 714 + ], + "type": "text", + "content": " from the standard Gaussian distribution " + }, + { + "bbox": [ + 54, + 606, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 54, + 606, + 296, + 714 + ], + "type": "text", + "content": " randomly, and then obtain the sample by computing " + }, + { + "bbox": [ + 54, + 606, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} = \\mu^{(1)} + \\epsilon \\sigma^{(1)},\\mathbf{Z}\\in \\mathbb{R}^{N\\times C}" + }, + { + "bbox": [ + 54, + 606, + 296, + 714 + ], + "type": "text", + "content": ". According to Bayesian probability theory [3, 21], we can con" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "sider " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " as empirical samples drawn from an approximate predictive distribution. The uncertainty of the input data can be quantified by calculating channel-wise variance, expressed as " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{data} = \\mathrm{Norm}(Var(\\mathbf{Z}))" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{data} \\in \\mathbb{R}^{N \\times 1}" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 121, + 554, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 554, + 228 + ], + "type": "text", + "content": "Uncertainty-guided graph augmentation. With the estimated uncertainty, we can assume that a larger variance indicates a greater deviation from the expected data distribution. We utilize a softmax function to map the " + }, + { + "bbox": [ + 313, + 121, + 554, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{data}" + }, + { + "bbox": [ + 313, + 121, + 554, + 228 + ], + "type": "text", + "content": " into a probability distribution " + }, + { + "bbox": [ + 313, + 121, + 554, + 228 + ], + "type": "inline_equation", + "content": "p_u^{node}" + }, + { + "bbox": [ + 313, + 121, + 554, + 228 + ], + "type": "text", + "content": ", which determines the unimportance of each node in the graph. Combining with traditional graph augmentations like node dropping, and node masking, we can decide whether to drop or mask a node based on the its uncertainty." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 228, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 361 + ], + "type": "text", + "content": "Furthermore, we compute the edge uncertainty using the affinity matrix of " + }, + { + "bbox": [ + 313, + 228, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{data}" + }, + { + "bbox": [ + 313, + 228, + 555, + 361 + ], + "type": "text", + "content": ", and map to a probability distribution " + }, + { + "bbox": [ + 313, + 228, + 555, + 361 + ], + "type": "inline_equation", + "content": "p_u^{edge}" + }, + { + "bbox": [ + 313, + 228, + 555, + 361 + ], + "type": "text", + "content": " as well. Intuitively, if the uncertainty at the edge's endpoints is higher, the uncertainty of the edge itself will be higher. This allows us to perform edge perturbation and subgraph transformation on the graph. With the augmented graph representation " + }, + { + "bbox": [ + 313, + 228, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{aug}" + }, + { + "bbox": [ + 313, + 228, + 555, + 361 + ], + "type": "text", + "content": " and its vanilla view " + }, + { + "bbox": [ + 313, + 228, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 228, + 555, + 361 + ], + "type": "text", + "content": ", we use the UGGN to map them into the same representation space, and minimize their KL divergence to mitigate the influence of high-uncertainty nodes, thereby learning a more robust graph representation:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 328, + 369, + 555, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 369, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 328, + 369, + 555, + 403 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {K L} ^ {\\mathcal {G}} = \\sum_ {i = 1} ^ {N} K L (\\mathcal {N} \\left(\\mu_ {\\mathcal {G}}, \\sigma_ {\\mathcal {G}}\\right) | | \\mathcal {N} \\left(\\mu_ {\\mathcal {G} _ {a u g}}, \\sigma_ {\\mathcal {G} _ {a u g}}\\right)) \\tag {2}", + "image_path": "2f79f027ecee161932448d5db790a35b9d6fee461a0cca7246a85852d8396953.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 411, + 554, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 554, + 506 + ], + "type": "text", + "content": "Uncertainty-guided feature refinement. On the other hand, the uncertainty estimation can be mapped back to spatial space as an uncertainty map " + }, + { + "bbox": [ + 313, + 411, + 554, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{map} \\in \\mathbf{R}^{H \\times W \\times 1}" + }, + { + "bbox": [ + 313, + 411, + 554, + 506 + ], + "type": "text", + "content": ". We directly compute the dot product of reverse uncertainty map " + }, + { + "bbox": [ + 313, + 411, + 554, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{1} - \\mathbf{U}_{map}" + }, + { + "bbox": [ + 313, + 411, + 554, + 506 + ], + "type": "text", + "content": " with the visual features " + }, + { + "bbox": [ + 313, + 411, + 554, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 313, + 411, + 554, + 506 + ], + "type": "text", + "content": " and their associated labels " + }, + { + "bbox": [ + 313, + 411, + 554, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 313, + 411, + 554, + 506 + ], + "type": "text", + "content": " for refinement, which will effectively alleviate the impact of high uncertainty positions, resulting in more accurate foreground representations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 507, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 554, + 567 + ], + "type": "text", + "content": "Since the uncertainty estimation may be less reliable in the early stages of training, we introduce an uncertainty gate to adjust the weights for uncertainty refinement dynamically. This gate is initialized with a small positive scalar to mitigate any initial unreliability." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 575, + 455, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 455, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 455, + 589 + ], + "type": "text", + "content": "3.4. Part prototype Discovery" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 554, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 654 + ], + "type": "text", + "content": "Despite we obtain a refined representation, the foreground region of example retains local semantics, resulting in varying degrees of correspondence with the target across different areas. To extract more discriminative representations, we propose the implementation of part prototype discovery." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": "As shown in Figure 2 below, we can treat the last layers of UGGN as an extension by converting a deterministic graph clustering network to a probabilistic type, i.e., a parameterized distribution for feature clustering. Then, we propose an attention mechanism for the next layers to" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20204" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "assign different weights to neighbors based on their variances since larger variances indicate more uncertainties in the latent representations. A smooth exponential function is utilized to control the effect of variances on weights " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^{(l)} = \\exp (-\\gamma \\Sigma^{(l)})" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 91, + 137, + 295, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 137, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 91, + 137, + 295, + 156 + ], + "type": "interline_equation", + "content": "\\mathbf {M} ^ {(l + 1)} = \\tilde {\\mathbf {D}} ^ {- \\frac {1}{2}} \\mathbf {A} \\tilde {\\mathbf {D}} ^ {- \\frac {1}{2}} \\left(\\mathbf {M} ^ {(l)} \\odot \\mathcal {A} ^ {(l)}\\right) \\mathbf {W} _ {\\mu} ^ {(l)} \\tag {3}", + "image_path": "577f3535d03b2eea7270431692d485dac2a57a04cb6691e9ca7ab75b1cd309ae.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 72, + 157, + 295, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 157, + 295, + 175 + ], + "spans": [ + { + "bbox": [ + 72, + 157, + 295, + 175 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} ^ {(l + 1)} = \\tilde {\\mathbf {D}} ^ {- \\frac {1}{2}} \\mathbf {A} \\tilde {\\mathbf {D}} ^ {- \\frac {1}{2}} \\left(\\boldsymbol {\\Sigma} ^ {(l)} \\odot \\mathcal {A} ^ {(l)} \\odot \\mathcal {A} ^ {(l)}\\right) \\mathbf {W} _ {\\sigma} ^ {(l)} \\tag {4}", + "image_path": "40d6906adfb502e8a790a0bfdafa3c4a3c7ccfc71d7cc1b425f8504055175361.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 179, + 295, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 179, + 295, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 179, + 295, + 204 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 179, + 295, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 55, + 179, + 295, + 204 + ], + "type": "text", + "content": " is the diagonal degree matrix calculated by " + }, + { + "bbox": [ + 55, + 179, + 295, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{i,i} = \\sum_{j}\\mathbf{A}_{i,j}" + }, + { + "bbox": [ + 55, + 179, + 295, + 204 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 179, + 295, + 204 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{D}} = \\mathbf{D} + \\mathbf{I}_N" + }, + { + "bbox": [ + 55, + 179, + 295, + 204 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 179, + 295, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_N" + }, + { + "bbox": [ + 55, + 179, + 295, + 204 + ], + "type": "text", + "content": " is the identity matrix." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 204, + 296, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 296, + 274 + ], + "type": "text", + "content": "We also sampled the updated distribution using the reparameterization trick to obtain the final node classification results. To solve the non-differentiability caused by the arg max operation to derive classification labels, we utilize the Gumbel-Softmax trick [19] to get the final one-hot like clusters " + }, + { + "bbox": [ + 55, + 204, + 296, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{C}\\mathbf{u} \\in \\mathbb{R}^{N \\times c}" + }, + { + "bbox": [ + 55, + 204, + 296, + 274 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 277, + 295, + 315 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 295, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 295, + 315 + ], + "type": "interline_equation", + "content": "\\mathbf {C l u} _ {i} = \\frac {\\exp \\left(\\left(\\log \\left(\\pi_ {i}\\right) + g _ {i}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {c} \\exp \\left(\\left(\\log \\left(\\pi_ {j}\\right) + g _ {j}\\right) / \\tau\\right)}, \\quad \\text {f o r} i = 1, \\dots , c. \\tag {5}", + "image_path": "ef1705652403b5132999d36f160d5bd392b236c8a45d9b3444228b42f8f0b9b3.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 316, + 295, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 351 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 316, + 295, + 351 + ], + "type": "inline_equation", + "content": "g_{i},\\dots ,g_{c}" + }, + { + "bbox": [ + 55, + 316, + 295, + 351 + ], + "type": "text", + "content": " are i.i.d samples drawn from the Gumbel(0,1) distribution, " + }, + { + "bbox": [ + 55, + 316, + 295, + 351 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 316, + 295, + 351 + ], + "type": "text", + "content": " is a categorical distribution with class probabilities and " + }, + { + "bbox": [ + 55, + 316, + 295, + 351 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 316, + 295, + 351 + ], + "type": "text", + "content": " is the softmax temperature." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 351, + 296, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 296, + 386 + ], + "type": "text", + "content": "With the clusters " + }, + { + "bbox": [ + 55, + 351, + 296, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{C}\\mathbf{u}" + }, + { + "bbox": [ + 55, + 351, + 296, + 386 + ], + "type": "text", + "content": ", we get part prototype representations " + }, + { + "bbox": [ + 55, + 351, + 296, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = \\{\\mathbf{P}_i\\in \\mathbb{R}^C\\mid i = 1,\\dots ,c\\}" + }, + { + "bbox": [ + 55, + 351, + 296, + 386 + ], + "type": "text", + "content": " by averaging " + }, + { + "bbox": [ + 55, + 351, + 296, + 386 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 351, + 296, + 386 + ], + "type": "text", + "content": " node features in the same cluster:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 388, + 295, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 388, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 116, + 388, + 295, + 411 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {i} = \\frac {1}{n} \\mathbf {C l u} _ {i} ^ {\\top} \\cdot \\mathbf {F}, \\quad \\mathbf {F} \\in \\mathbf {F} _ {m} \\tag {6}", + "image_path": "b01921660a66ef309f38c556cadd81a98844bbcc4d5b8b8fd8c14cb3b391a221.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 415, + 151, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 151, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 151, + 427 + ], + "type": "text", + "content": "3.5. Mask Decoding" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "text", + "content": "To further interact with the mask decoder for mask prediction, we extracted four kinds of visual instructions for in-context prompt generation: visual features " + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "text", + "content": ", mask average pooling features " + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_p = \\mathrm{Avg\\_pool}(\\mathbf{F})" + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{F} \\in \\mathbf{F}_m" + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "text", + "content": ", part prototypes " + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "text", + "content": ", and the mask label " + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "text", + "content": " as depicted in Figure 1. We denote the " + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{I} = \\{\\mathbf{F}, \\mathbf{F}_p, \\mathbf{P}, \\mathbf{y}\\}" + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "text", + "content": " as the set of in-context instructions derived from reference samples and the target image. A query-based prompt generator (QPG) is implemented to support a scalable structure and the ability to support different numbers of examples flexibly. Considering computational efficiency, we first concat them in channel dimension and project them into the latent space used by the QPG, then feed the reference instruction " + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_r" + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "text", + "content": " and target instruction " + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_t" + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "text", + "content": " to interact with learnable query tokens " + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "text", + "content": " to get the final in-context prompt " + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{ict}" + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 110, + 616, + 295, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 616, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 110, + 616, + 295, + 630 + ], + "type": "interline_equation", + "content": "\\mathbf {I} = \\operatorname {P r o j} (\\operatorname {C o n c a t} (\\mathbf {F}, \\mathbf {F} _ {p}, \\mathbf {P}, \\mathbf {y})) \\tag {7}", + "image_path": "26ebc28448a5f1f5c0fa1589f5fc8fd3f17ec8d4046272947a73cc4c1746161a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 129, + 632, + 295, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 632, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 129, + 632, + 295, + 645 + ], + "type": "interline_equation", + "content": "\\mathbf {q} _ {i c l} = Q P G \\left(\\mathbf {I} _ {r}, \\mathbf {I} _ {t}, \\mathbf {q}\\right) \\tag {8}", + "image_path": "d8163fbe6f4a0115c91a344a1f0695446e95bf3eb68eb78c40fe489b62d81258.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 649, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 649, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 649, + 295, + 687 + ], + "type": "text", + "content": "Finally, we feed the target image feature " + }, + { + "bbox": [ + 55, + 649, + 295, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_t" + }, + { + "bbox": [ + 55, + 649, + 295, + 687 + ], + "type": "text", + "content": " with the incontext prompt " + }, + { + "bbox": [ + 55, + 649, + 295, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{icl}" + }, + { + "bbox": [ + 55, + 649, + 295, + 687 + ], + "type": "text", + "content": " to the frozen mask decoder " + }, + { + "bbox": [ + 55, + 649, + 295, + 687 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 649, + 295, + 687 + ], + "type": "text", + "content": " for prediction " + }, + { + "bbox": [ + 55, + 649, + 295, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{pred}" + }, + { + "bbox": [ + 55, + 649, + 295, + 687 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 692, + 295, + 706 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 692, + 295, + 706 + ], + "spans": [ + { + "bbox": [ + 133, + 692, + 295, + 706 + ], + "type": "interline_equation", + "content": "\\mathbf {y} _ {p r e d} = D \\left(\\mathbf {F} _ {t}, \\mathbf {q} _ {i c l}\\right) \\tag {9}", + "image_path": "89f239d5c1d470c7b0619592abbf77e1aa9e8a385d6a47838ccc66d5bcd7744b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 72, + 400, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 400, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 400, + 83 + ], + "type": "text", + "content": "3.6. Loss function" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 89, + 554, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 554, + 114 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 554, + 114 + ], + "type": "text", + "content": "We employ a linear combination of Binary Cross-Entropy (BCE) loss and Dice loss to supervise the mask prediction:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 367, + 123, + 554, + 136 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 123, + 554, + 136 + ], + "spans": [ + { + "bbox": [ + 367, + 123, + 554, + 136 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {b c e} \\mathcal {L} _ {b c e} + \\lambda_ {d i c e} \\mathcal {L} _ {d i c e} \\tag {10}", + "image_path": "073976a1d8036932185ce0a0e2ae0247faf469816f510808954d3244ecbe3821.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 145, + 555, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 145, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 145, + 555, + 205 + ], + "type": "text", + "content": "For better optimization, we also adopt some auxiliary losses. To ensure that the learned representations are indeed Gaussian distributions, we use an explicit regularization to constrain the latent representations in the first layer of UGGN as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 353, + 212, + 554, + 245 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 212, + 554, + 245 + ], + "spans": [ + { + "bbox": [ + 353, + 212, + 554, + 245 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {K L} ^ {r e g} = \\sum_ {i = 1} ^ {N} K L (\\mathcal {N} (\\mu_ {\\mathcal {G}}, \\sigma_ {\\mathcal {G}}) | | \\mathcal {N} (\\mathbf {0}, \\mathbf {I})) \\tag {11}", + "image_path": "55ac5c16be251689541f4179ae8e3750b36163c221521565267084d9bddaa8db.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 253, + 554, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 554, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 554, + 277 + ], + "type": "text", + "content": "Following the original GCN paper [23], we also impose " + }, + { + "bbox": [ + 313, + 253, + 554, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 313, + 253, + 554, + 277 + ], + "type": "text", + "content": " regularization on parameters of the first layer of UGGN:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 375, + 285, + 553, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 285, + 553, + 301 + ], + "spans": [ + { + "bbox": [ + 375, + 285, + 553, + 301 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {2} ^ {r e g} = \\left\\| \\mathbf {W} _ {\\mu} ^ {(0)} \\right\\| _ {2} ^ {2} + \\left\\| \\mathbf {W} _ {\\sigma} ^ {(0)} \\right\\| _ {2} ^ {2} \\tag {12}", + "image_path": "3af5f7ea18bdf361e9aee3d6e07ebcec7c62d01f42887483345a4a0c94b926f0.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 309, + 554, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 554, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 554, + 346 + ], + "type": "text", + "content": "We find it helpful to use an auxiliary loss function in the uncertainty-guided part prototype discovery step during training to push prototypes away from each other:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 349, + 355, + 553, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 355, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 349, + 355, + 553, + 381 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {P} = \\sum_ {\\mathbf {P} _ {i}, \\mathbf {P} _ {j} \\in \\mathbf {P}} \\max ((d - d i s t (\\mathbf {P} _ {i}, \\mathbf {P} _ {j})), 0) \\tag {13}", + "image_path": "fc9a86beefcb006fa031ce76628dfcbe0ff531e144cc8743d9cbdeed5c296f43.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 390, + 554, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 554, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 554, + 414 + ], + "type": "text", + "content": "Combing with the graph augmentation loss in Section 3.3, the final loss is the combination of the above loss:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 314, + 434, + 554, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 434, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 314, + 434, + 554, + 460 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {K L} ^ {\\text {r e g}} \\mathcal {L} _ {K L} ^ {\\text {r e g}} + \\lambda_ {2} ^ {\\text {r e g}} \\mathcal {L} _ {2} ^ {\\text {r e g}} + \\lambda_ {P} \\mathcal {L} _ {P} + \\lambda_ {K L} ^ {\\mathcal {G}} \\mathcal {L} _ {K L} ^ {\\mathcal {G}} \\tag {14}", + "image_path": "0e79ebb9033d51cbf339bd2264970fb6b4f2d270a5f552ce5c85f03257612ecd.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 314, + 471, + 395, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 471, + 395, + 484 + ], + "spans": [ + { + "bbox": [ + 314, + 471, + 395, + 484 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 314, + 491, + 379, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 491, + 379, + 502 + ], + "spans": [ + { + "bbox": [ + 314, + 491, + 379, + 502 + ], + "type": "text", + "content": "4.1. Datasets." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 508, + 554, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 554, + 555 + ], + "type": "text", + "content": "We train UNICL-SAM with mixed semantic segmentation datasets. Notably, our approach does not necessitate any modifications to the architecture or training pipeline when incorporating an additional dataset." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 556, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 554, + 604 + ], + "type": "text", + "content": "COCO [28] is a widely utilized dataset supporting instance segmentation, semantic segmentation, and panoptic segmentation tasks. It comprises 118K training images and 5k validation images, categorized into 80 categories." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 605, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 640 + ], + "type": "text", + "content": "ADE20k [55] is a widely used benchmark for semantic segmentation, which consists of over 25K images sourced from diverse environments with 150 semantic categories." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 647, + 449, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 449, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 449, + 661 + ], + "type": "text", + "content": "4.2. Implementation Details." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "We implement UNICL-SAM in PyTorch and use 8 Nvidia A6000 GPUs for our experiments. A total batch size of 64 was employed, with the model trained for 20 epochs. We merge the COCO instances for semantic segmentation and" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20205" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 107, + 70, + 504, + 258 + ], + "blocks": [ + { + "bbox": [ + 107, + 70, + 504, + 258 + ], + "lines": [ + { + "bbox": [ + 107, + 70, + 504, + 258 + ], + "spans": [ + { + "bbox": [ + 107, + 70, + 504, + 258 + ], + "type": "table", + "html": "
MethodsResolutionTrainable ParamsCOCO-20iFSS-1000LVIS-92i
one-shotfew-shotone-shotfew-shotone-shotfew-shot
specialist model
HSNet [34]48028M41.7*50.7*86.5*88.5*17.422.9
VAT [18]41752M42.9*49.4*90.3*90.8*18.522.7
FPTrans [51]480139M56.5*65.5*----
generalist model
Painter [42]448354M33.132.661.762.310.510.9
SegGPT [43]448354M56.167.985.689.318.625.4
PerSAM [52]-2M23.5-75.6-18.4-
Matcher [29]896-52.760.787.089.633.040.0
VRP-SAM [40]5121.6M53.9-----
ICVU [38]256309M58.0-----
SINE [30]89619M64.566.1--31.235.5
UNICL-SAM51855M77.878.784.086.334.137.4
", + "image_path": "b81e54b9b035136037aa1a700cd713e605cf04e42b6766b8f92ec5939be792c6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 320, + 297, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 320, + 297, + 477 + ], + "spans": [ + { + "bbox": [ + 54, + 320, + 297, + 477 + ], + "type": "text", + "content": "conduct the sample pools by category. We utilize DINOv2 (ViT-L/14) [36] as the default vision encoder. For the segmenter, we choose SAM (ViT-H) [24], which is pre-trained with 1B masks and 11M images, shows impressive zero-shot segmentation performance. During training, we froze the vision encoder and the segmenter, optimizing only the remaining parameters. We use AdamW [31] optimizer with a weight decay of " + }, + { + "bbox": [ + 54, + 320, + 297, + 477 + ], + "type": "inline_equation", + "content": "5e - 2" + }, + { + "bbox": [ + 54, + 320, + 297, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 320, + 297, + 477 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 54, + 320, + 297, + 477 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 320, + 297, + 477 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.95" + }, + { + "bbox": [ + 54, + 320, + 297, + 477 + ], + "type": "text", + "content": " for optimization. We use a base learning rate of " + }, + { + "bbox": [ + 54, + 320, + 297, + 477 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 54, + 320, + 297, + 477 + ], + "type": "text", + "content": " and a warmup of 4 epochs during training. Each image is cropped and resized to " + }, + { + "bbox": [ + 54, + 320, + 297, + 477 + ], + "type": "inline_equation", + "content": "518\\times 518" + }, + { + "bbox": [ + 54, + 320, + 297, + 477 + ], + "type": "text", + "content": " resolution. We randomly sampled one reference pair during training. More implementation details are provided in the Appendix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 489, + 142, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 489, + 142, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 142, + 501 + ], + "type": "text", + "content": "4.3. Main Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 508, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 296, + 544 + ], + "type": "text", + "content": "For the main experiments, we compare UNICL-SAM with other specialist/generalist methods on several in-context semantic segmentation benchmarks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": "Few-shot semantic segmentation. Following previous works [29, 30, 33, 43], we evaluate UNICL-SAM on few-shot semantic segmentation benchmarks to demonstrate its great generalization performance. Additionally, we report the in-domain performance of specialist models to ensure a fair comparison. As illustrated in Table 1, UNICL-SAM shows competitive or even better results than recent powerful in-context segmentation models. For the in-domain COCO- " + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "inline_equation", + "content": "20^{i}" + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": " [35] tests, UNICL-SAM outperforms previous generalist models, achieving an impressive MIoU score of 77.8 for one-shot and 78.7 for few-shot tests. For the out-of-domain FSS1000 [26] tests, UNICL-SAM also achieves competitive results that are close to specialist models, even without ever being trained on it. Furthermore, we conduct" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 318, + 553, + 414 + ], + "blocks": [ + { + "bbox": [ + 54, + 266, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 54, + 266, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 555, + 301 + ], + "type": "text", + "content": "Table 1. Comparison with state-of-the-art specialist and generalist models on few-shot semantic segmentation benchmarks. We report both one-shot and few-shot results for comparison. * indicates that the categories in training cover the categories in testing within the same dataset. Previous state-of-the-art results among generalists are underlined." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 318, + 553, + 414 + ], + "lines": [ + { + "bbox": [ + 317, + 318, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 317, + 318, + 553, + 414 + ], + "type": "table", + "html": "
Multi-scale Feature AdapterPart Prototype DiscoveryUncertainty Probabilistic ModelingCOCO-20i
MIoU ↑MAE ↓
XXX71.770.039
XX✓72.990.037
X✓X72.410.038
✓X✓73.280.036
✓✓X73.780.036
X✓✓74.610.034
✓✓✓74.640.034
", + "image_path": "b909568ba547871559b4a9bad3f3adb00d1a1e5cd4121e0f94cf848fd5dcfef2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 354, + 422, + 514, + 434 + ], + "lines": [ + { + "bbox": [ + 354, + 422, + 514, + 434 + ], + "spans": [ + { + "bbox": [ + 354, + 422, + 514, + 434 + ], + "type": "text", + "content": "Table 2. Ablation of proposed components." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 456, + 555, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 555, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 555, + 517 + ], + "type": "text", + "content": "experiments on LVIS- " + }, + { + "bbox": [ + 313, + 456, + 555, + 517 + ], + "type": "inline_equation", + "content": "92^{i}" + }, + { + "bbox": [ + 313, + 456, + 555, + 517 + ], + "type": "text", + "content": " [29], a more challenging one-shot benchmark built upon LVIS [16]. On LVIS- " + }, + { + "bbox": [ + 313, + 456, + 555, + 517 + ], + "type": "inline_equation", + "content": "92^{i}" + }, + { + "bbox": [ + 313, + 456, + 555, + 517 + ], + "type": "text", + "content": ", UNICL-SAM surpasses previous generalists in a one-shot setting with a MIoU score of 34.1, demonstrating superior generalization and versatility." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 526, + 416, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 416, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 416, + 539 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 545, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 616 + ], + "type": "text", + "content": "We conduct ablation experiments on each of the proposed components as follows. By default, we train the model on a small part of the COCO training set with 34k images and about " + }, + { + "bbox": [ + 313, + 545, + 555, + 616 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 545, + 555, + 616 + ], + "type": "text", + "content": " masks for 20 epochs. We report the Mean Intersection over Union (MIoU) and the Mean Absolute Error (MAE) scores for evaluation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "Ablations on proposed components. Table 2 illustrates the impact of the proposed components. The incorporation of our part prototype discovery technique contributes to improved segmentation performance, demonstrating its effectiveness in extracting contextual cues from examples. When combined with our uncertainty probabilistic modeling method, as indicated in line 6, the model's capabilities are further strengthened, resulting in a substantial increase" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "20206" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 293, + 149 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 293, + 149 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 293, + 149 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 293, + 149 + ], + "type": "table", + "html": "
Un-guided Graph AugmentationUn-guided Feature RefinementCOCO-20i
MIoU ↑MAE ↓
XX73.780.036
✓X73.820.036
X✓73.990.035
✓✓74.640.034
", + "image_path": "759064c628abc73b1364746e4faaa1c589e8b547cfcfcbca0c53c84b45290915.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 194, + 293, + 262 + ], + "blocks": [ + { + "bbox": [ + 55, + 158, + 295, + 180 + ], + "lines": [ + { + "bbox": [ + 55, + 158, + 295, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 295, + 180 + ], + "type": "text", + "content": "Table 3. Ablation of uncertainty guided graph augmentation and feature refinement strategies." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 194, + 293, + 262 + ], + "lines": [ + { + "bbox": [ + 59, + 194, + 293, + 262 + ], + "spans": [ + { + "bbox": [ + 59, + 194, + 293, + 262 + ], + "type": "table", + "html": "
Graph Model TypeUn-guided Graph AugmentationCOCO-20i
GCNUGGNMIoU ↑MAE ↓
✓✗✗72.120.038
✗✓✗72.410.038
✗✓✓73.820.036
", + "image_path": "e6599f8df3e6f7acec9a91624dd39ae78929d1e05dec8d8a69101796fbec6bce.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 296, + 294, + 418 + ], + "blocks": [ + { + "bbox": [ + 58, + 296, + 294, + 418 + ], + "lines": [ + { + "bbox": [ + 58, + 296, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 58, + 296, + 294, + 418 + ], + "type": "image", + "image_path": "26434fe31582abdac61e0e5f04a49fff0a358735f2828392efc5b86f9147cb7c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 428, + 295, + 462 + ], + "lines": [ + { + "bbox": [ + 55, + 428, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 428, + 295, + 462 + ], + "type": "text", + "content": "Figure 3. Visualization of uncertainty maps. We apply a predefined colormap to the uncertainty estimation map for visualization. Areas of high uncertainty are marked with red bboxes " + }, + { + "bbox": [ + 55, + 428, + 295, + 462 + ], + "type": "inline_equation", + "content": "\\square" + }, + { + "bbox": [ + 55, + 428, + 295, + 462 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 485, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 485, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 54, + 485, + 295, + 652 + ], + "type": "text", + "content": "of " + }, + { + "bbox": [ + 54, + 485, + 295, + 652 + ], + "type": "inline_equation", + "content": "2.2\\mathrm{mIoU}" + }, + { + "bbox": [ + 54, + 485, + 295, + 652 + ], + "type": "text", + "content": ". We attribute this improvement to the proposed UGGN's ability to accurately model data distribution and provide a reliable uncertainty measure, which facilitates the optimization of visual features through the two enhancement strategies. This, in turn, allows the part prototypes to yield more discriminative representations. Additionally, the integration of multi-scale adapters contributes to performance enhancement. The synergistic combination of these three components achieves optimal results, with a MIoU of 74.64 and a MAE of 0.034 on the COCO " + }, + { + "bbox": [ + 54, + 485, + 295, + 652 + ], + "type": "inline_equation", + "content": "20^{i}" + }, + { + "bbox": [ + 54, + 485, + 295, + 652 + ], + "type": "text", + "content": " benchmark. These findings underscore the effectiveness of the proposed components. It is worth noting that in the following ablation experiments unless otherwise specified, the multi-scale feature adapter is used by default." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Ablations on uncertainty probabilistic modeling. We investigate how the proposed uncertainty quantification stage and corresponding graph augmentation and feature refinement strategies affect performance. As shown in Table 3, utilizing either enhancement alone will help improve per" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 323, + 70, + 547, + 141 + ], + "blocks": [ + { + "bbox": [ + 64, + 270, + 286, + 282 + ], + "lines": [ + { + "bbox": [ + 64, + 270, + 286, + 282 + ], + "spans": [ + { + "bbox": [ + 64, + 270, + 286, + 282 + ], + "type": "text", + "content": "Table 4. Ablation of uncertainty-guided graph augmentation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 70, + 547, + 141 + ], + "lines": [ + { + "bbox": [ + 323, + 70, + 547, + 141 + ], + "spans": [ + { + "bbox": [ + 323, + 70, + 547, + 141 + ], + "type": "table", + "html": "
Un-guided Feature Refinement TypeCOCO-20i
MIoU ↑MAE ↓
top ratio masking43.730.117
random masking73.440.037
learnable gate74.640.034
", + "image_path": "929cfafaaedae02bba3b8b7a16a12f519042814e10b1eb1156bf18dec0cb3845.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 353, + 175, + 517, + 259 + ], + "blocks": [ + { + "bbox": [ + 317, + 149, + 550, + 162 + ], + "lines": [ + { + "bbox": [ + 317, + 149, + 550, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 149, + 550, + 162 + ], + "type": "text", + "content": "Table 5. Ablation of uncertainty-guided feature refinement type." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 353, + 175, + 517, + 259 + ], + "lines": [ + { + "bbox": [ + 353, + 175, + 517, + 259 + ], + "spans": [ + { + "bbox": [ + 353, + 175, + 517, + 259 + ], + "type": "table", + "html": "
Number of ClustersCOCO-20i
MIoU ↑MAE ↓
573.360.037
1073.820.036
1572.500.039
2072.380.040
", + "image_path": "dea861d5c37d90f89c72a6d0374516216759d0dca8ee87fa4f0e3da6cdeab5dd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 267, + 553, + 289 + ], + "lines": [ + { + "bbox": [ + 313, + 267, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 553, + 289 + ], + "type": "text", + "content": "Table 6. Ablation of the number of clusters of part prototype discovery." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 314, + 553, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 553, + 338 + ], + "type": "text", + "content": "formance. Further combining the two strategies can achieve better performance on the COCO " + }, + { + "bbox": [ + 313, + 314, + 553, + 338 + ], + "type": "inline_equation", + "content": "20^{i}" + }, + { + "bbox": [ + 313, + 314, + 553, + 338 + ], + "type": "text", + "content": " benchmark." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 340, + 555, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 471 + ], + "type": "text", + "content": "We further explore the specific effects of the two enhancement strategies through detailed ablation experiments. Initially, we examine the impact of employing deterministic graph networks versus our uncertainty Gaussian graph network (UGGN). As depicted in Table 4, UGGN outperforms GCN by achieving an increase of 0.29 in mIoU, indicating that the network effectively models the distribution of graph representations. Building on this foundation, the introduction of uncertainty-guided graph augmentation technology further enhances model performance, resulting in an increase of 1.41 in MIoU scores." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 473, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 554, + 581 + ], + "type": "text", + "content": "For the uncertainty-guided feature refinement strategy, we study several different feature weighting/optimization schemes as shown in Table 5. the \"top ratio masking\" means masking nodes from areas of high uncertainty according to a preset ratio; \"random masking\" stands for masking out nodes whose associated uncertainty score is higher than a random probability map as described in [48]; \"learnable gate\" denotes our learnable gating mechanism. It can be observed that our \"learnable gate\" achieved the best result." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "We visualize the predicted uncertainty map for reference examples in the COCO test set. We can notice that there is often higher uncertainty at the foreground edge and small objects, as shown in the first row of Figure 3. Furthermore, for some annotation errors in COCO as shown in the second row, the model can give good instructions and the impact of these areas will be eliminated by our feature refinement strategy. The last line shows some naturally existing degradation and style changes. It can be seen that UGGN also exhibits ideal responses to these regions, effectively estimating the data variance." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20207" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 148 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 148 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 148 + ], + "type": "table", + "html": "
MethodsResolutionTrainable ParamsCleanSupport ImageSupport LabelAvg. DeclineAvg. Decline Ratio
ColorBlunnessCompressionSpaceDomain shiftDeformation
SegGPT [43]448354M62.1-2.7-3.0-2.7-7.8-6.2-8.8-5.2-8.4%
Matcher [29]896-41.6-0.5-2.5-6.5-3.8-5.3-6.2-4.1-9.9%
SINE [30]89619M70.0-0.7-1.7-5.9-2.8-6.5-12.4-5.0-7.1%
UNICL-SAM51855M79.8-1.0-1.8-3.0-3.3-4.8-5.0-3.1-3.9%
", + "image_path": "9060a472d4834929daecad7b33688a3a305c8921a6e7b540af68f8641af61c2a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 154, + 555, + 189 + ], + "lines": [ + { + "bbox": [ + 55, + 154, + 555, + 189 + ], + "spans": [ + { + "bbox": [ + 55, + 154, + 555, + 189 + ], + "type": "text", + "content": "Table 7. Robustness comparison with advanced in-context segmentation generalists on constructed test sets. We present the MIoU score on clean data and the corresponding performance differences across various degraded data. Our UNICL-SAM outperforms these baselines in various degradation/domain shift scenarios with slight performance changes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 57, + 206, + 294, + 331 + ], + "blocks": [ + { + "bbox": [ + 57, + 206, + 294, + 331 + ], + "lines": [ + { + "bbox": [ + 57, + 206, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 57, + 206, + 294, + 331 + ], + "type": "image", + "image_path": "85636649746597e0b7ddb75e77ac458b3014e03d7bb283ecd78b5e5ffd02e2fa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 340, + 295, + 373 + ], + "lines": [ + { + "bbox": [ + 55, + 340, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 295, + 373 + ], + "type": "text", + "content": "Figure 4. Visualization of cluster masks. We visualize the patches associated with the same cluster using a uniform color to facilitate the observation of the characteristics of the cluster mask." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 396, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 479 + ], + "type": "text", + "content": "Ablations on part prototype discovery. We study the details of part prototype discovery stage. As shown in Table 6, we conducted an ablation experiment on the hyperparameter setting of the number of part prototype clusters. We found that the model showed the best performance when the number of clusters was 10, which was also used in the main experiments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 481, + 295, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 295, + 564 + ], + "type": "text", + "content": "We also visualize the effect of part prototypes, as shown in Figure 4. The results clearly demonstrate that semantically similar patches are effectively grouped into the same cluster. Furthermore, for multiple targets within the same category, analogous local semantics are aggregated into a single cluster, thereby enhancing the modeling of foreground representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 575, + 211, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 211, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 211, + 586 + ], + "type": "text", + "content": "4.5. Robustness Simulation Tests" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "To assess the effectiveness of our method in real-world applications, we constructed a test set featuring a variety of degraded and generated examples that may be encountered in practical scenarios, as shown in Figure 5. For detailed information on the robustness simulation test settings, please refer to the Appendix. The results in Table 7 indicate that our approach significantly outperforms existing advanced in-context segmentation models, effectively addressing the diverse challenges posed by these varied scenes. UNICLSAM excels not only in enhancing segmentation perfor" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 208, + 553, + 342 + ], + "blocks": [ + { + "bbox": [ + 316, + 208, + 553, + 342 + ], + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 342 + ], + "type": "image", + "image_path": "18ec0e824087f64185067bc0be2ad663e8e7ae3105b856ef5fd3af8ca25965d3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 350, + 555, + 383 + ], + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 383 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 383 + ], + "type": "text", + "content": "Figure 5. A simple example of robustness simulation tests designed to assess the model's capabilities across various degradation and domain shift scenarios. Zoom in for a better view." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 409, + 555, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 555, + 506 + ], + "type": "text", + "content": "mance in clean environments but also improving results in degraded conditions as well. It surpasses the previous state-of-the-art generalist average by an impressive score of 11.7 MIoU, highlighting its effectiveness across diverse conditions. Even in the most challenging robustness tests, our model's performance fluctuations remain within a narrow range, with an average decline of only 3.1 MIoU. This demonstrates its great robustness." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 523, + 388, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 523, + 388, + 536 + ], + "spans": [ + { + "bbox": [ + 314, + 523, + 388, + 536 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": "In this work, we point out the data distribution uncertainty in in-context segmentation and introduce UNICL-SAM, a robust in-context segmentation framework that leverages uncertainty modeling to evaluate each input example and related graph augmentation and feature refinement strategies to help alleviate the impact of areas with high data uncertainty. We also present part prototypes as a form of in-context instruction to obtain discriminative local semantics. Leveraging the effective design, UNICL-SAM achieved superior performance on in-context segmentation benchmarks, while demonstrating exceptional robustness in conducted robustness simulation tests. We hope UNICL-SAM can promote the development of in-context segmentation generalists for real-world applications." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20208" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 161, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 161, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 161, + 85 + ], + "type": "text", + "content": "6. Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 157 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 157 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (No.62472396, No.62121002), Anhui Provincial Science and Technology Major Project (No.2023z020006), and the advanced computing resources provided by the SuperComputing Center of USTC." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 167, + 115, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 167, + 115, + 180 + ], + "spans": [ + { + "bbox": [ + 56, + 167, + 115, + 180 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 187, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 187, + 296, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 187, + 296, + 232 + ], + "spans": [ + { + "bbox": [ + 61, + 187, + 296, + 232 + ], + "type": "text", + "content": "[1] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros. Visual prompting via image inpainting. Advances in Neural Information Processing Systems, 35:25005-25017, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 233, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 233, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 61, + 233, + 296, + 297 + ], + "type": "text", + "content": "[2] Benjamin Bischke, Patrick Helber, Damian Borth, and Andreas Dengel. Segmentation of imbalanced classes in satellite imagery using adaptive uncertainty weighted class loss. In IGARSS 2018-2018 IEEE International Geoscience and Remote Sensing Symposium, pages 6191-6194. IEEE, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 300, + 296, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 300, + 296, + 343 + ], + "spans": [ + { + "bbox": [ + 61, + 300, + 296, + 343 + ], + "type": "text", + "content": "[3] Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty in neural network. In International conference on machine learning, pages 1613-1622. PMLR, 2015. 1, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 344, + 296, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 344, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 61, + 344, + 296, + 388 + ], + "type": "text", + "content": "[4] Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty in neural network. In International conference on machine learning, pages 1613-1622. PMLR, 2015. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 389, + 296, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 389, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 61, + 389, + 296, + 477 + ], + "type": "text", + "content": "[5] Patrik Olā Bressan, José Marcato Junior, José Augusto Correa Martins, Maximilian Jaderson de Melo, Diogo Nunes Gonçalves, Daniel Matte Freitas, Ana Paula Marques Ramos, Michelle Táis Garcia Furuya, Lucas Prado Osco, Jonathan de Andrade Silva, et al. Semantic segmentation with labeling uncertainty and class imbalance applied to vegetation mapping. International Journal of Applied Earth Observation and Geoinformation, 108:102690, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 478, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 478, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 62, + 478, + 294, + 499 + ], + "type": "text", + "content": "[6] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 501, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 501, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 501, + 295, + 544 + ], + "type": "text", + "content": "[7] Jie Chang, Zhonghao Lan, Changmao Cheng, and Yichen Wei. Data uncertainty learning in face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5710-5719, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 545, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 545, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 545, + 295, + 578 + ], + "type": "text", + "content": "[8] Bike Chen, Wei Peng, Xiaofeng Cao, and Juha Röning. Hyperbolic uncertainty aware semantic segmentation. IEEE Transactions on Intelligent Transportation Systems, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 579, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 579, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 62, + 579, + 295, + 633 + ], + "type": "text", + "content": "[9] Wei-Ting Chen, Yu-Jiet Vong, Sy-Yen Kuo, Sizhou Ma, and Jian Wang. Robustsam: Segment anything robustly on degraded images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4081-4091, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 296, + 689 + ], + "type": "text", + "content": "[10] Kazuki Endo, Masayuki Tanaka, and Masatoshi Okutomi. Semantic segmentation of degraded images using layer-wise feature adjustor. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3205-3213, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "text", + "content": "[11] Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "learning. In international conference on machine learning, pages 1050-1059. PMLR, 2016. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[12] Jakob Gawlikowski, Cedrique Rovile Njieutcheu Tassi, Mohsin Ali, Jongseok Lee, Matthias Humt, Jianxiang Feng, Anna Kruspe, Rudolph Trielbel, Peter Jung, Ribana Roscher, et al. A survey of uncertainty in deep neural networks. Artificial Intelligence Review, 56(Suppl 1):1513-1589, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "type": "text", + "content": "[13] Ethan Goan and Clinton Fookes. Uncertainty in real-time semantic segmentation on embedded systems. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4491-4501, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 195, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 239 + ], + "type": "text", + "content": "[14] Dazhou Guo, Yanting Pei, Kang Zheng, Hongkai Yu, Yuhang Lu, and Song Wang. Degraded image semantic segmentation with dense-gram networks. IEEE Transactions on Image Processing, 29:782-795, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "text", + "content": "[15] Hongji Guo, Hanjing Wang, and Qiang Ji. Uncertainty-guided probabilistic transformer for complex action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20052-20061, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 294, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 338 + ], + "type": "text", + "content": "[16] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "text", + "content": "[17] Saumya Gupta, Yikai Zhang, Xiaoling Hu, Prateek Prasanna, and Chao Chen. Topology-aware uncertainty for image segmentation. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 383, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 436 + ], + "type": "text", + "content": "[18] Sunghwan Hong, Seokju Cho, Jisu Nam, Stephen Lin, and Seungryong Kim. Cost aggregation with 4d convolutional swim transformer for few-shot segmentation. In European Conference on Computer Vision, pages 108-126. Springer, 2022. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 437, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 469 + ], + "type": "text", + "content": "[19] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 471, + 555, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 555, + 525 + ], + "type": "text", + "content": "[20] Kaixiang Ji, Feng Chen, Xin Guo, Yadong Xu, Jian Wang, and Jingdong Chen. Uncertainty-guided learning for improving image manipulation detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22456-22465, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 526, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 558 + ], + "type": "text", + "content": "[21] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 1, 2, 3, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "type": "text", + "content": "[22] Soopil Kim, Philip Chikontwe, Sion An, and Sang Hyun Park. Uncertainty-aware semi-supervised few shot segmentation. Pattern Recognition, 137:109292, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 593, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 624 + ], + "type": "text", + "content": "[23] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 625, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 689 + ], + "type": "text", + "content": "[24] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2, 3, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[25] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20209" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 75, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 296, + 116 + ], + "type": "text", + "content": "a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3041-3050, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "type": "text", + "content": "[26] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few-shot segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2869–2878, 2020. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 295, + 218 + ], + "type": "text", + "content": "[27] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European conference on computer vision, pages 280-296. Springer, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 295, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 295, + 285 + ], + "type": "text", + "content": "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 295, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 295, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 295, + 329 + ], + "type": "text", + "content": "[29] Yang Liu, Muzhi Zhu, Hengtao Li, Hao Chen, Xinlong Wang, and Chunhua Shen. Matcher: Segment anything with one shot using all-purpose feature matching. arXiv preprint arXiv:2305.13310, 2023. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 295, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 295, + 374 + ], + "type": "text", + "content": "[30] Yang Liu, Chenchen Jing, Hengtao Li, Muzhi Zhu, Hao Chen, Xinlong Wang, and Chunhua Shen. A simple image segmentation framework via in-context examples. Advances in Neural Information Processing Systems, 2024. 1, 2, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 295, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 295, + 398 + ], + "type": "text", + "content": "[31] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 295, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 295, + 443 + ], + "type": "text", + "content": "[32] Yao Lu, Max Bartolo, Alastair Moore, Sebastian Riedel, and Pontus Stenetorp. Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity. arXiv preprint arXiv:2104.08786, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 444, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 295, + 498 + ], + "type": "text", + "content": "[33] Lingchen Meng, Shiyi Lan, Hengduo Li, Jose M Alvarez, Zuxuan Wu, and Yu-Gang Jiang. Segic: Unleashing the emergent correspondence for in-context segmentation. In European Conference on Computer Vision, pages 203-220. Springer, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 500, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 295, + 544 + ], + "type": "text", + "content": "[34] Juhong Min, Dahiyun Kang, and Minsu Cho. Hypercorrelation squeeze for few-shot segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6941-6952, 2021. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 295, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 295, + 589 + ], + "type": "text", + "content": "[35] Khoi Nguyen and Sinisa Todorovic. Feature weighting and boosting for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 622-631, 2019. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 295, + 645 + ], + "type": "text", + "content": "[36] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "type": "text", + "content": "[37] Xinru Shan and Chaoning Zhang. Robustness of segment anything model (sam) for autonomous driving in adverse weather conditions. arXiv preprint arXiv:2306.13290, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 690, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 295, + 714 + ], + "type": "text", + "content": "[38] Dianmo Sheng, Dongdong Chen, Zhentao Tan, Qiankun Liu, Qi Chu, Jianmin Bao, Tao Gong, Bin Liu, Shengwei Xu," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 72, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 554, + 116 + ], + "type": "text", + "content": "and Nenghai Yu. Towards more unified in-context visual understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13362-13372, 2024. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 554, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 182 + ], + "type": "text", + "content": "[39] Mahesh Subedar, Ranganath Krishnan, Paulo Lopez Meyer, Omesh Tickoo, and Jonathan Huang. Uncertainty-aware audiovisual activity recognition using deep bayesian variational inference. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6301-6310, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 183, + 553, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 183, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 316, + 183, + 553, + 238 + ], + "type": "text", + "content": "[40] Yanpeng Sun, Jiahui Chen, Shan Zhang, Xinyu Zhang, Qiang Chen, Gang Zhang, Errui Ding, Jingdong Wang, and Zechao Li. Vrp-sam: Sam with visual reference prompt. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23565-23574, 2024. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 239, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 270 + ], + "type": "text", + "content": "[41] Matias Valdenegro-Toro. Deep sub-ensembles for fast uncertainty estimation in image classification. arXiv preprint arXiv:1910.08168, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 271, + 555, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 271, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 271, + 555, + 327 + ], + "type": "text", + "content": "[42] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6830-6839, 2023. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 327, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 369 + ], + "type": "text", + "content": "[43] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023. 1, 2, 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 371, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 426 + ], + "type": "text", + "content": "[44] Zhenyu Wang, Yali Li, Ye Guo, Lu Fang, and Shengjin Wang. Data-uncertainty guided multi-phase learning for semi-supervised object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4568-4577, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "type": "text", + "content": "[45] Zhiyong Wu, Yaoxiang Wang, Jiacheng Ye, and Lingpeng Kong. Self-adaptive in-context learning: An information compression perspective for in-context example selection and ordering. arXiv preprint arXiv:2212.10375, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "text", + "content": "[46] Weihao Xia, Zhanglin Cheng, Yujiu Yang, and Jing-Hao Xue. Cooperative semantic segmentation and image restoration in adverse environmental conditions. arXiv preprint arXiv:1911.00679, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 514, + 553, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 553, + 548 + ], + "type": "text", + "content": "[47] Chengming Xu, Chen Liu, Yikai Wang, and Yanwei Fu. Towards global optimal visual in-context learning prompt selection. arXiv preprint arXiv:2405.15279, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 548, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 603 + ], + "type": "text", + "content": "[48] Fan Yang, Qiang Zhai, Xin Li, Rui Huang, Ao Luo, Hong Cheng, and Deng-Ping Fan. Uncertainty-guided transformer reasoning for camouflaged object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4146-4155, 2021. 2, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 603, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 658 + ], + "type": "text", + "content": "[49] Zhou Yang, Weisheng Dong, Xin Li, Jinjian Wu, Leida Li, and Guangming Shi. Self-feature distillation with uncertainty modeling for degraded image recognition. In European Conference on Computer Vision, pages 552-569. Springer, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[50] Tianyuan Yu, Da Li, Yongxin Yang, Timothy M Hospedales, and Tao Xiang. Robust person re-identification by modelling feature uncertainty. In Proceedings of the IEEE/CVF international conference on computer vision, pages 552–561, 2019. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20210" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 341 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "text", + "content": "[51] Jian-Wei Zhang, Yifan Sun, Yi Yang, and Wei Chen. Featureproxy transformer for few-shot segmentation. Advances in Neural Information Processing Systems, 35:6575-6588, 2022. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[52] Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Xianzheng Ma, Hao Dong, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048, 2023. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 205 + ], + "type": "text", + "content": "[53] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. What makes good examples for visual in-context learning? Advances in Neural Information Processing Systems, 36:17773-17794, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 240 + ], + "type": "text", + "content": "[54] Zhizheng Zhang, Cuiling Lan, Wenjun Zeng, Zhibo Chen, and Shih-Fu Chang. Uncertainty-aware few-shot image classification. arXiv preprint arXiv:2010.04525, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "type": "text", + "content": "[55] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127:302-321, 2019. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 341 + ], + "type": "text", + "content": "[56] Dingyuan Zhu, Ziwei Zhang, Peng Cui, and Wenwu Zhu. Robust graph convolutional networks against adversarial attacks. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1399-1407, 2019. 4" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20211" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/UNOPose_ Unseen Object Pose Estimation with an Unposed RGB-D Reference Image/94d3840c-f3fe-4b2b-9388-48ccee10c9d7_content_list.json b/2025/UNOPose_ Unseen Object Pose Estimation with an Unposed RGB-D Reference Image/94d3840c-f3fe-4b2b-9388-48ccee10c9d7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4a0cd631535cc117031a6d2f29bb3c36194216f0 --- /dev/null +++ b/2025/UNOPose_ Unseen Object Pose Estimation with an Unposed RGB-D Reference Image/94d3840c-f3fe-4b2b-9388-48ccee10c9d7_content_list.json @@ -0,0 +1,1676 @@ +[ + { + "type": "text", + "text": "UNOpose: Unseen Object Pose Estimation with an Unposed RGB-D Reference Image", + "text_level": 1, + "bbox": [ + 259, + 130, + 738, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xingyu Liu $^{1, *}$ , Gu Wang $^{1, *}$ , Ruida Zhang $^{1}$ , Chenyangguang Zhang $^{1}$ , Federico Tombari $^{2,3}$ , and Xiangyang Ji $^{1}$ $^{1}$ Tsinghua University, BNRist $^{2}$ Technical University of Munich, $^{3}$ Google", + "bbox": [ + 205, + 202, + 792, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{liuxy21@mails., wanggu1@, zhangrd23@mails., zcyg22@mails., xyji@}tsinghua.edu.cn,", + "bbox": [ + 135, + 258, + 857, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tombari@in.tum.de", + "bbox": [ + 419, + 277, + 571, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unseen object pose estimation methods often rely on CAD models or multiple reference views, making the onboarding stage costly. To simplify reference acquisition, we aim to estimate the unseen object's pose through a single unposed RGB-D reference image. While previous works leverage reference images as pose anchors to limit the range of relative pose, our scenario presents significant challenges since the relative transformation could vary across the entire $SE(3)$ space. Moreover, factors like occlusion, sensor noise, and extreme geometry could result in low viewpoint overlap. To address these challenges, we present a novel approach and benchmark, termed UNOPose1, for UNseen One-reference-based object Pose estimation. Building upon a coarse-to-fine paradigm, UNOPose constructs an $SE(3)$ -invariant reference frame to standardize object representation despite pose and size variations. To alleviate small overlap across viewpoints, we recalibrate the weight of each correspondence based on its predicted likelihood of being within the overlapping region. Evaluated on our proposed benchmark based on the BOP Challenge, UNOPose demonstrates superior performance, significantly outperforming traditional and learning-based methods in the one-reference setting and remaining competitive with CAD-model-based methods. The code and dataset are available at github.com/shanice-l/UNOPose.", + "bbox": [ + 89, + 357, + 485, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 750, + 220, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Localizing an object in Euclidean space by estimating its 6DoF pose, i.e., 3DoF orientation and 3DoF position, plays a crucial role in augmented/virtual reality [59, 76], scene understanding [37, 63] and robotic manipulation [23, 26, 80]. The vast majority of works [42, 48, 68, 78, 83-86] focuses on instance-level object pose estimation, where the train", + "bbox": [ + 89, + 773, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing and testing datasets consist of an identical set of known object instances, and the CAD models of objects are often required for generating training images and labels [16, 56]. More recently, Wang et al. [87] and its successors [6, 79, 88] extended this paradigm to the category level, intending to estimate poses for novel instances within predefined categories without requiring CAD models of target objects. Both paradigms have limitations in open-world applications, where annotating and training for new objects beyond known categories would be very labor-intensive and sometimes prohibitive.", + "bbox": [ + 511, + 327, + 908, + 492 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To mitigate this problem, recent research efforts [3, 43, 51, 90] have shifted towards pose estimation of arbitrary novel objects, where the target object is unseen during training. This task presents a great challenge due to the inherent uncertainty of the novel object's canonical frame (defined in the CAD model of the object). Current approaches [43, 57, 62] often tackle this by using omnidirectional reference views to cover the target object, separating pose estimation task to viewpoint selection and relative pose estimation to known pose anchors. If CAD models are available during the test phase, reference views can be easily produced through image rendering [43, 51, 61, 90]. Alternatively, without CAD models, reference views can be obtained by capturing multiple images of the object [29, 31, 55, 77]. Both setups have shortcomings, as creating a CAD model of an object or labeling a substantial number of object poses is limited in scalability. For example, they can hardly adapt to in-the-wild scenarios where novel objects are introduced unpredictably.", + "bbox": [ + 511, + 500, + 908, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To sum up, while existing methods exhibit promising transfer capability to novel objects, they are notably constrained by their reliance on the CAD model or multiple reference views (Fig. 1 (a, b)). In this work, we aim to minimize onboarding cost by estimating novel object poses using a single unposed RGB-D reference image (Fig. 1). This setup is different from existing methods mainly in two aspects. First, we focus on estimating the relative pose between", + "bbox": [ + 511, + 779, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contributions.", + "bbox": [ + 109, + 875, + 223, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1UNO $(\\prime \\mathrm{u}:\\mathrm{no}\\overline{\\mathrm{u}} /$ ) means one in Spanish and Italian.", + "bbox": [ + 114, + 887, + 387, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22023", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ea83ed8cb74d85b90fc84596b8b7beecb05bd2252b35bf5958d3239b5ca762a9.jpg", + "image_caption": [ + "Figure 1. Illustration of unseen object pose estimation. Given a query image presenting a target object unseen during training, we aim to estimate its segmentation and 6DoF pose w.r.t. a reference frame. While previous methods [43, 57, 77, 90] often rely on the CAD model or multiple RGB(-D) images for reference, we merely use one unposed RGB-D reference image." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 903, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "two unposed viewpoints rather than the absolute pose. Note that the absolute pose of a novel object becomes ill-posed without a certain canonical frame, while relative pose estimation remains well-defined regardless of the canonical frame's definition. Secondly, some existing relative pose estimation methods [98, 105, 106] merely use RGB modality to predict the 3DoF rotation, limiting their ability to predict relative translation and generalize beyond their training datasets. In contrast, we fully leverage RGB-D modality to predict the 6DoF relative pose, meanwhile enhancing the network's generalizability to novel objects and environments.", + "bbox": [ + 88, + 412, + 483, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nevertheless, estimating the pose of unseen objects with only one reference view presents significant challenges from various aspects. First, previous methods [3, 43] select the most similar reference view as a pose anchor, distinctly reducing the search space for relative poses. However, in our scenario, where both the target and reference object poses are unknown, the relative pose can vary across the entire $SE(3)$ space. Moreover, in partial-to-partial object matching, factors such as occlusion, sensor noise, and extreme geometry can severely interfere with the matching process. To address these challenges, we propose a novel approach and benchmark for UNseen One-reference-based object Pose estimation (UNOPose). Our pipeline leverages the strong generalization capabilities of vision foundation models [40, 65] to produce an effective segmentation of the unseen object. Then, UNOPose employs a coarse-to-fine paradigm for estimating the relative pose between the reference and the query objects. To alleviate the challenge of diverse pose and size variations, we introduce the $SE(3)$ -invariant global reference frame (GRF) to standardize object representation. Subsequently, a hierarchical object encoding", + "bbox": [ + 89, + 583, + 485, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "paradigm based on the local reference frame (LRF) further captures fine-grained geometric details. For achieving reliable correspondences in partial-to-partial object matching, we harness an overlap predictor to identify and concentrate on the overlapping region.", + "bbox": [ + 511, + 412, + 906, + 488 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, we propose a new benchmark devised from the BOP challenge [34] to facilitate evaluation and future research of unseen object pose estimation with one reference. Extensive experiments on YCB-V [92], LM-O [2] and TUD-L [32] datasets demonstrate that our UNOPose surpasses all compared methods on a single reference setting. To our surprise, UNOPose with an unposed reference is even on par with some $SE(3)$ -invariant-feature-based methods relying on CAD models (Ours $70.9\\%$ vs. ZTE-PPF [1] $69.0\\%$ vs. Koenig-PPF [41] $75.1\\%$ w.r.t. $\\mathrm{AR}_{\\mathrm{BOP}}$ metric).", + "bbox": [ + 511, + 488, + 908, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 532, + 640, + 857, + 654 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, we are the first to conduct unseen object 6DoF pose estimation leveraging a single unposed RGB-D reference.", + "- Based on the BOP Challenge, we devise a new extensive benchmark tailored for unseen object segmentation and pose estimation with one reference. Additionally, we evaluate several traditional and learning-based methods on this benchmark for completeness.", + "- We introduce UNOPose, a network for learning relative transformation between reference and query objects. To achieve this, we propose the $SE(3)$ -invariant global and local reference frames, enabling standardized object representations despite variations in pose and size. Furthermore, the network can automatically adjust the confidence of each correspondence by incorporating an overlap predictor." + ], + "bbox": [ + 532, + 656, + 908, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22024", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 91, + 89, + 232, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Class-specific Pose Estimation. Class-specific object pose estimation aims at predicting 6DoF poses of either instance of a known object (instance-level) or unseen objects within a known category (category-level). Instance-level object poses are often solved by direct regression [35, 47, 56, 85] using deep neural networks, or by establishing 2D-3D [27, 48, 68] or 3D-3D correspondences [30, 73, 83] which are then leveraged by RANSAC-based PnP/Kabsch algorithms. The instance-level setting requires expensive data generation/annotation [18, 19, 81, 102] and training for every new object. To address this limitation, category-level pose estimation [87] is proposed to estimate 9DoF poses of novel objects among specific categories without CAD models. Mainstream approaches can also be categorized into direct regression [6, 9, 10, 17, 50, 54, 99–101] or correspondence-based [8, 24, 79, 87] methods. However, this setting remains constrained to a limited number of categories, given the additional challenge of aligning canonical frames, managing symmetric objects, and categorizing objects in similar categories correctly (e.g. can and bottle).", + "bbox": [ + 93, + 114, + 483, + 416 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Novel Object Pose Estimation. Estimating the pose of novel objects beyond known classes, i.e., objects at inference time are unseen during training, is a useful yet challenging task. Existing works solve this problem through image-based matching [3, 43, 60, 61, 74, 90, 104] or feature-based matching approaches [4, 29, 36, 51, 62, 66, 77]. Notably, given a target image, OVE6D [3] and MegaPose [43] retrieve the most similar viewpoint from a pre-rendered omnidirectional image database as a coarse pose estimate. A customized neural network then refines the coarse pose estimate. FoundationPose [90] further builds a pose ranking network to score each refined pose hypothesis. Feature-matching methods learn local feature descriptors to construct pixel-level or point-level correspondences. For example, OnePose [77] and OnePose++ [29] reconstruct the 3D point cloud of an unseen object to establish 2D-3D correspondences, whereas SAM-6D [51] builds 3D-3D correspondences. More recently, some methods [4, 66] directly exploit the power of foundation models for zero-shot pose estimation. Exemplarily, FoundPose [66] extracts DINOv2 visual features [65], while FreeZe [4] further utilizes generalizable geometric features [69] to conduct feature matching.", + "bbox": [ + 93, + 417, + 483, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Relative Pose Estimation. Orthogonal to previous works which rely on known instances or multiple reference views, relative object pose estimation estimates a relative transformation of the unseen object with only one reference image. One closely related topic is camera pose estimation from two views. For example, RelPose [98] and RelPose++ [49] infer a distribution of relative rotations by leveraging an energy-based formulation. Moreover, iFusion [91] optimizes the unknown relative pose by inverting the novel view synthesis diffusion model [53]. Relative camera pose estimation relies", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "on static background to establish correspondence between viewpoints. In the field of relative object pose estimation, given two RGB images, 3DAHV [105] presents a hypothesis-and-verification framework to score each relative pose hypothesis. Extended by that, DVMNet [106] introduces a hypothesis-free pipeline to compute relative poses via deep voxel matching. The above works have limited capability in inferring 3DoF relative translations. In this work, with the introduction of depth data, we can estimate a complete 6DoF pose including rotation and translation.", + "bbox": [ + 511, + 90, + 906, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point Cloud Registration. Our approach follows the paradigm of correspondence-based methods [14, 15, 45, 89], which initially extract 3D local feature descriptors to establish correspondences between source and target point clouds. Subsequently, they estimate the relative transformation using techniques like SVD, RANSAC, or Hough Voting [44]. Some feature descriptors are proposed to ensure rotational invariance. For example, FPFH [72] descriptor is computed with geometric surface properties. PPF [22] uses the distance and angles to describe the relation of point pairs. TOLDI [93] proposes a robust local reference frame using keypoint normals and neighboring projections. With the advent of deep learning, neural networks are increasingly employed to extract 3D local or global feature descriptors [12, 38, 71, 89, 97].", + "bbox": [ + 511, + 242, + 908, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. UNO Object Segmentation and Pose Estimation", + "text_level": 1, + "bbox": [ + 511, + 482, + 908, + 516 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 511, + 525, + 718, + 540 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Assuming an arbitrary unseen rigid object in a query image, our goal is to estimate the object's mask $M_q$ and its 6DoF relative pose $\\Delta \\mathbf{T} \\in SE(3)$ with a masked RGB-D reference image exhibiting the target object without major occlusion or truncation. As illustrated in Fig. 1, the input is:", + "bbox": [ + 511, + 549, + 906, + 625 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) $[I_q|D_q]\\in \\mathbb{R}^{H\\times W\\times 4}$ : The query RGB-D image;", + "2) $\\left[\\bar{I}_p|D_p\\right] \\in \\mathbb{R}^{H \\times W \\times 4}$ and $M_p \\in \\mathbb{R}^{H \\times W}$ : The reference RGB-D image and the corresponding binary mask indicating the target object;", + "3) $K_{q} \\in \\mathbb{R}^{3 \\times 3}$ and $K_{p} \\in \\mathbb{R}^{3 \\times 3}$ : Camera intrinsics of the query and reference images." + ], + "bbox": [ + 514, + 625, + 906, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Optionally, if the pose $\\mathbf{T}_p \\in SE(3)$ of the reference object in the camera frame is known, the query object pose $\\mathbf{T}_q \\in SE(3)$ can be recovered by $\\mathbf{T}_q = \\Delta \\mathbf{T} \\mathbf{T}_p$ . Note that, for practicality, the method should not rely on the absolute pose of the reference object, since the world frame can vary arbitrarily in different applications. We only use the reference object pose during the inference stage for evaluation on standard object pose datasets.", + "bbox": [ + 511, + 715, + 906, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. UNO Object Segmentation", + "text_level": 1, + "bbox": [ + 511, + 845, + 754, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Firstly, we segment the query object from a cluttered background. Thanks to the great generalization ability of vision", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22025", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3004c9497e8ca6b6f792ef25e37a1f4b4f1722e393b7e280f65fc4b943bf5754.jpg", + "image_caption": [ + "Figure 2. The network architecture of UNOpose. Given the query and reference point clouds $\\mathbf{Q}_{cam}$ and $\\mathbf{P}_{cam}$ in the camera frame, UNOpose first transforms them into the $SE(3)$ -invariant global reference frame (GRF). Then feature descriptors are extracted from sparse point sets for constructing the coarse correlation matrix. For achieving precise correspondences, the fine pose estimation module exploits structural details using positional encoding and local reference frame (LRF) encoding." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 425 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "foundation models, methods like [7, 51, 61] can effectively segment novel objects with CAD models. However, unlike previous approaches, which generate diverse descriptors from multiple rendered views, we can access only a single reference image. To address this challenge, we use the SAM model [40] to predict all possible mask proposals from the query image, and then score each mask proposal by comparing DINOv2 [65] descriptors between the query and reference images using cosine similarity, identifying the most similar mask $M_{q}$ . Please refer to the appendix for more details about UNO segmentation.", + "bbox": [ + 89, + 506, + 483, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. UNO Object Pose Estimation", + "text_level": 1, + "bbox": [ + 89, + 683, + 352, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 Overview of UNOPose", + "text_level": 1, + "bbox": [ + 89, + 707, + 302, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the predicted mask $M_{q}$ of the query image and mask $M_{p}$ of the reference image, we crop and back-project the object of interest from depth maps $D_{q}, D_{p}$ into the camera space as two point sets $\\mathbf{Q}_{cam} \\in \\mathbb{R}^{N^{Q} \\times 3}$ and $\\mathbf{P}_{cam} \\in \\mathbb{R}^{N^{P} \\times 3}$ , where $N^{Q}$ and $N^{P}$ denotes the point numbers of query and reference point clouds, respectively. Our goal is to recover the relative transformation $\\Delta \\mathbf{T} = \\{\\Delta \\mathbf{R}, \\Delta \\mathbf{t}\\}$ by minimizing the correspondence distance", + "bbox": [ + 89, + 731, + 483, + 857 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min \\sum_ {(\\mathbf {q}, \\mathbf {p}) \\in \\mathbf {C}} \\| \\Delta \\mathbf {R} \\mathbf {q} + \\Delta \\mathbf {t} - \\mathbf {p} \\| _ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 869, + 483, + 904 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{C}$ is the predicted correspondence set between $\\mathbf{Q}_{cam}$ and $\\mathbf{P}_{cam}$ . We exploit both color and geometric cues to build this correspondence and follow the broadly used coarse-to-fine paradigm in point cloud registration to solve Eq. (1). The network is demonstrated in Fig. 2.", + "bbox": [ + 511, + 506, + 906, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2 Coarse-to-fine Pose Estimation", + "text_level": 1, + "bbox": [ + 511, + 599, + 782, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Constructing a Pose-invariant Reference Frame. Given only an unposed reference image, the relative pose to predict is arbitrary in the $SE(3)$ space, which renders a significant challenge for achieving robust correspondences. Hence, we introduce a pose-invariant global reference frame (GRF) and transform $\\mathbf{Q}_{cam}$ and $\\mathbf{P}_{cam}$ into GRF as $\\mathbf{Q}_G$ and $\\mathbf{P}_G$ .", + "bbox": [ + 511, + 623, + 906, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concretely, taking $\\mathbf{Q}_{cam}$ as an example, transforming the point cloud to GRF involves a 7DoF coordinate transformation $\\{\\mathbf{R}_G\\in SO(3),\\mathbf{t}_G\\in \\mathbb{R}^3,s_G\\in \\mathbb{R}\\}$ :", + "bbox": [ + 511, + 714, + 906, + 760 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {G} = \\left\\{\\mathbf {R} _ {G} ^ {\\top} \\left(\\mathbf {q} - \\mathbf {t} _ {G}\\right) / s _ {G} \\mid \\mathbf {q} \\in \\mathbf {Q} _ {c a m} \\right\\}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 763, + 906, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The origin of GRF is located at the object center $\\mathbf{c}_Q$ for translation invariance, and the radius of the point cloud is rescaled to 1 for size invariance, computed as", + "bbox": [ + 511, + 789, + 906, + 834 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {t} _ {G} = \\mathbf {c} _ {Q}, \\quad s _ {G} = \\max _ {\\mathbf {q} \\in \\mathbf {Q} _ {c a m}} \\| \\mathbf {q} - \\mathbf {c} _ {Q} \\| _ {2}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 840, + 906, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The key ingredient is to devise the rotation $\\mathbf{R}_G = [\\mathbf{r}_{Gx}|\\mathbf{r}_{Gy}|\\mathbf{r}_{Gz}]$ , where $\\mathbf{r}_{Gx},\\mathbf{r}_{Gy},\\mathbf{r}_{Gz}$ are the three columns", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22026", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of $\\mathbf{R}_G$ , ensuring orientation invariance for the transformed point cloud. Inspired by former works [25, 75, 93], we use the normal vector of the object center $\\mathbf{n}(\\mathbf{c}_Q)$ as $\\mathbf{r}_{Gz}$ , and project all points in $\\mathbf{Q}_{cam}$ onto the tangent plane of $\\mathbf{c}_Q$ , summarizing the projected vectors to determine $\\mathbf{r}_{Gx}$ . Then, $\\mathbf{r}_{Gy}$ is obtained by the cross-product of $\\mathbf{r}_{Gx}$ and $\\mathbf{r}_{Gz}$ , i.e., $\\mathbf{r}_{Gy} = \\mathbf{r}_{Gx} \\times \\mathbf{r}_{Gz}$ . This ensures that the XY-plane evenly divides the point cloud, with the x-axis representing the principal projection direction. The GRF is $SE(3)$ -invariant as it can be uniquely determined by the spatial distribution of the point cloud, making the transformed point cloud robust to variations in pose and size.", + "bbox": [ + 89, + 90, + 483, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Concretely, we perform SVD on the covariance matrix $\\mathrm{Cov}(\\mathbf{Q}_{cam}) = \\frac{1}{N_Q}\\mathbf{Q}_{cam}^\\top \\mathbf{Q}_{cam} - \\mathbf{c}_Q\\mathbf{c}_Q^\\top$ and use the singular vector w.r.t. the minimum singular value to determine $\\mathbf{r}_{Gz}$ :", + "bbox": [ + 89, + 273, + 483, + 321 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {r} _ {G z} = \\left\\{ \\begin{array}{l l} \\mathbf {n} \\left(\\mathbf {c} _ {Q}\\right), & \\text {i f} \\mathbf {n} \\left(\\mathbf {c} _ {Q}\\right) ^ {\\top} \\sum_ {\\mathbf {q} \\in \\mathbf {Q} _ {c a m}} \\left(\\mathbf {c} _ {Q} - \\mathbf {q}\\right) > 0 \\\\ - \\mathbf {n} \\left(\\mathbf {c} _ {Q}\\right), & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 334, + 483, + 387 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Afterwards, $\\mathbf{r}_{Gx}$ is computed as", + "bbox": [ + 89, + 388, + 303, + 402 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {r} _ {G x} = \\sum_ {\\mathbf {q} \\in \\mathbf {Q} _ {c a m}} w _ {q} \\left(\\left(\\mathbf {q} - \\mathbf {c} _ {Q}\\right) - \\mathbf {r} _ {G z} ^ {\\top} \\left(\\mathbf {q} - \\mathbf {c} _ {Q}\\right) \\mathbf {r} _ {G z}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 416, + 483, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w_{q}$ is a weight w.r.t. distance between $\\mathbf{q}$ and $\\mathbf{c}_Q$ (See the appendix for details).", + "bbox": [ + 89, + 463, + 482, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous works often depend on complex networks or computationally expensive PPF features for $SE(3)$ invariance [46, 94, 96]. In contrast, our transformation is computationally efficient and can seamlessly adapt to a wide range of network architectures.", + "bbox": [ + 89, + 494, + 483, + 569 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Coarse Pose Estimation. Given point clouds $\\mathbf{Q}_G$ and $\\mathbf{P}_G$ in GRF, we sample two sparse point sets $\\mathbf{Q}_G^c\\in \\mathbb{R}^{N^c\\times 3}$ and $\\mathbf{P}_G^c\\in \\mathbb{R}^{N^c\\times 3}$ to efficiently obtain a coarse pose initialization $\\Delta \\mathbf{T}_{init}$ . Specifically, we leverage a geometry encoder [71] and a color encoder [65] to extract point cloud and RGB features separately. Features are further concatenated as $f_{Q}^{c}\\in \\mathbb{R}^{N^{c}\\times d}$ and $f_{P}^{c}\\in \\mathbb{R}^{N^{c}\\times d}$ , where $d$ is the dimension of embeddings. Following [51], we add a learnable background token for assigning non-overlapping points. The embeddings, denoted as $\\hat{f}_Q^c\\in \\mathbb{R}^{(N^c +1)\\times d}$ and $\\hat{f}_P^c\\in \\mathbb{R}^{(N^c +1)\\times d}$ , are processed by three stacked Geometric Transformer decoding modules [71].", + "bbox": [ + 89, + 571, + 483, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The outputs of the last decoder $\\hat{F}_Q^c\\in \\mathbb{R}^{(N^c +1)\\times D}$ and $\\hat{F}_P^c\\in \\mathbb{R}^{(N^c +1)\\times D}$ are point-wise features for building the correlation matrix. However, in our setting, the overlap ratio can be influenced by complicated factors like viewpoint, occlusion or depth noise. To solve this issue, the network additionally predicts overlap confidence $\\hat{O}_Q^c\\in \\mathbb{R}^{(N^c +1)\\times 1}$ and $\\hat{O}_P^c\\in \\mathbb{R}^{(N^c +1)\\times 1}$ , which indicate point-wise probabilities of being in the overlapping region. The overlap-aware correlation matrix $\\mathbf{X}^c\\in \\mathbb{R}^{(N^c +1)\\times (N^c +1)}$ can thus be com", + "bbox": [ + 89, + 757, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "puted by", + "bbox": [ + 513, + 92, + 576, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} ^ {c} = \\operatorname {s o f t m a x} [ (\\hat {O} _ {Q} ^ {c} \\odot \\hat {F} _ {Q} ^ {c}) (\\hat {O} _ {P} ^ {c} \\odot \\hat {F} _ {P} ^ {c}) ^ {\\top} ]. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 114, + 906, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here $\\odot$ indicates element-wise multiplication. Each element in $\\mathbf{X}^c$ suggests a correlation score between a point pair in $\\mathbf{Q}^c$ and $\\mathbf{P}^c$ .", + "bbox": [ + 511, + 142, + 906, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once $\\mathbf{X}^c$ is computed, we can extract all possible corresponding point pairs between $\\mathbf{Q}^c$ and $\\mathbf{P}^c$ as well as their correlation scores to solve Eq. (1). We first compute pose hypotheses by sampling $N_{H}$ triplets of point pairs according to the distribution of $\\mathbf{X}^c$ . Then each pose hypothesis $\\Delta \\mathbf{T}_h = \\{\\Delta \\mathbf{R}_h, \\Delta \\mathbf{t}_h\\}$ is scored with the reciprocal of distance $D_h$ as in [27, 51]", + "bbox": [ + 511, + 188, + 908, + 294 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} D _ {h} = \\frac {1}{N ^ {c}} \\sum_ {\\mathbf {p} ^ {c} \\in \\mathbf {P} ^ {c}} \\min _ {\\mathbf {q} ^ {c} \\in \\mathbf {Q} ^ {c}} \\| \\Delta \\mathbf {R} _ {h} ^ {\\top} \\left(\\mathbf {q} ^ {c} - \\Delta \\mathbf {t} _ {h}\\right) - \\mathbf {p} ^ {c} \\| _ {2}, \\tag {7} \\\\ \\xi_ {h} = \\frac {1}{D _ {h}}, \\quad h = 1, 2, 3, \\dots , N _ {H}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 301, + 906, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The pose hypothesis with the highest score $\\xi_{h}$ is selected as the initial pose prediction $\\Delta \\mathbf{T}_{init} = \\{\\Delta \\mathbf{R}_{init},\\Delta \\mathbf{t}_{init}\\}$ and further sent to transform the input of the next stage.", + "bbox": [ + 511, + 380, + 905, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fine Pose Estimation. After transforming $\\mathbf{Q}_{cam}$ to $\\tilde{\\mathbf{Q}}_{cam}$ with the initial pose prediction $\\Delta \\mathbf{T}_{init}$ , we perform a fine matching process between two dense point sets, i.e., $\\tilde{\\mathbf{Q}}^f \\in \\mathbb{R}^{N^f \\times 3}$ and $\\mathbf{P}^f \\in \\mathbb{R}^{N^f \\times 3}$ , to achieve a more accurate pose. The network exploits geometric details through a hierarchical encoding paradigm, comprising a positional encoding layer and a local reference frame encoding layer. For each point, we first encode its global position using a mini-PointNet [70], and then construct SE(3)-invariant local reference frame (LRF) to gather local descriptors. These two encodings complement each other, as local descriptors capture fine geometric structures within small neighborhoods, while the positional encoding layer offers global geometric context.", + "bbox": [ + 511, + 426, + 906, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Leveraging $\\tilde{\\mathbf{Q}}^f$ for example, the process of constructing LRF encoding is as follows. For each point $\\tilde{\\mathbf{q}}^m$ in $\\tilde{\\mathbf{Q}}^f$ , we build a local region set $\\tilde{\\mathcal{Q}}^m = \\{\\tilde{\\mathbf{q}}_j\\}$ where $\\| \\tilde{\\mathbf{q}}_j - \\tilde{\\mathbf{q}}^m \\|_2 \\leq r\\}_{j=1}^{N_D}$ by grouping $N_D$ neighboring points of $\\tilde{\\mathbf{q}}^m$ . The transformation pose $\\{\\mathbf{R}_L^m, \\mathbf{t}_L^m, s_L^m\\}$ for LRF is computed similarly to GRF (Eq. 3, 4, 5), except that LRF is built upon a local point set while GRF is based on the entire point cloud. By computing the transformation pose, we calculate local point descriptors as", + "bbox": [ + 511, + 625, + 908, + 762 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {Q} _ {L} ^ {m} = \\tilde {\\mathcal {Q}} _ {L} ^ {m} = \\left\\{\\left(\\mathbf {R} _ {L} ^ {m}\\right) ^ {\\top} \\left(\\tilde {\\mathbf {q}} _ {j} - \\mathbf {t} _ {L} ^ {m}\\right) / s _ {L} ^ {m} \\right\\} _ {j = 1} ^ {N _ {D}}, \\tag {8} \\\\ \\mathbf {Q} _ {L} ^ {f} = \\{\\mathcal {Q} _ {L} ^ {m} \\} _ {m = 1} ^ {N ^ {f}}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 768, + 906, + 813 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly, we transform $\\mathbf{P}^f$ to $\\mathbf{P}_L^f$ . Note that $\\mathcal{Q}_L^m = \\tilde{\\mathcal{Q}}_L^m$ since LRF is pose-invariant. Then, the LRF encoding is extracted from $\\mathbf{Q}_L^f$ and $\\mathbf{P}_L^f$ with a three-layer MLP.", + "bbox": [ + 511, + 823, + 903, + 871 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Positional encoding and LRF encoding are combined with the geometric and color features as the input of the", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22027", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d4732933d15ba4561c6067fe7b492eab1cf23cd25cc088583b2ad1a729b2b5bb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSeg. ModelModalityCAD modelLM-OTUD-LYCB-VAvgTime (s)
ZeroPose [7]SAM [40]RGB✓35.642.153.443.73.90
CNOS [61]FastSAM [107]RGB✓39.748.059.949.20.23
CNOS [61]SAM [40]RGB✓39.639.159.546.11.71
SAM-6D [51]FastSAM [107]RGB-D✓42.251.762.152.01.47
SAM-6D [51]SAM [40]RGB-D✓46.056.960.554.54.53
UNOseg (Ours)FastSAM [107]RGBX37.350.367.351.60.24
UNOseg (Ours)SAM [40]RGBX39.756.266.854.22.68
", + "bbox": [ + 168, + 88, + 826, + 214 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/032cc425a3e4ff453dc116eb41c3992a46720aae7d9f593df801c68c5d5de629.jpg", + "table_caption": [ + "Table 1. Segmentation results (mAP) on LM-O, TUD-L, and YCB-V. Results of other methods are obtained from the BOP website bop.felk.cvut.cz/leaderboards/segmentation-unseen-bop23. We denote the best score in bold and the second best score with underline." + ], + "table_footnote": [], + "table_body": "
MethodSeg. ModelModalityRef. TypeLM-OTUD-LYCB-VAvgTime (s)
ZTE-PPF [1]MRCNN [28]DModel66.390.450.269.00.74
Koenig-PPF [41]Retina/MRCNNRGB-DModel63.192.070.175.10.99
Ref. AlignCenterUNOSeg-SAMDImage27.328.847.534.52.69
FPFH+RANSAC [72]UNOSeg-SAMDImage31.031.050.037.36.38
FPFH+MAC [103]UNOSeg-SAMDImage22.522.149.631.4136.94
PPF [22]UNOSeg-SAMDImage29.714.838.327.611.79
PPF_3D_ICP [22]UNOSeg-SAMDImage44.729.166.846.914.27
FCGF+RANSAC [12]UNOSeg-SAMDImage38.959.057.651.810.96
FCGF+MAC [103]UNOSeg-SAMDImage33.948.351.044.460.53
UTOPIC [11]UNOSeg-SAMDImage13.735.410.519.94.00
GeDi [69]UNOSeg-SAMDImage42.867.360.656.948.89
FreeZe [4]UNOSeg-SAMRGB-DImage45.568.365.559.852.96
SAM-6D* [51]UNOSeg-SAMRGB-DPosed Image54.529.768.150.84.21
UNOpose (Ours)UNOSeg-FastSAMRGB-DImage55.866.581.968.11.11
UNOpose (Ours)UNOSeg-SAMRGB-DImage58.771.083.170.93.70
", + "bbox": [ + 132, + 263, + 864, + 478 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Pose estimation results on LM-O, TUD-L, and YCB-V. The mean Average Recall (%) of the BOP metric and the average time (s) per image are reported. We highlight the best scores of each setting in bold. Specifically, for adapting SAM-6D [51] to our setting, we randomly choose one of the two templates which [51] uses as a canonical reference during training.", + "bbox": [ + 89, + 488, + 906, + 531 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Geometric Transformer. Similar to initial pose prediction, by decoding these features, we obtain the fine point-wise features $\\hat{F}_Q^f,\\hat{F}_P^f$ and overlap confidencees $\\hat{O}_Q^f,\\hat{O}_P^f$ , which are then used to obtain the overlap-aware fine correlation matrix $\\mathbf{X}^{f}\\in \\mathbb{R}^{(N^{f} + 1)\\times (N^{f} + 1)}$ . The final pose $\\Delta \\mathbf{T}$ is predicted by solving Eq. (1) using $\\mathbf{X}^f$ with the weighted SVD algorithm.", + "bbox": [ + 89, + 553, + 483, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 666, + 223, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 693, + 279, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Implemented with PyTorch [67], we use the DINOv2 [13, 65] pre-trained ViT-base [20] to encode color information. GeoTransformers [71] with attention [82] and sparse-to-dense linear attention [51] are respectively employed to extract coarse and fine geometric correspondences. The network is trained with standard MegaPose synthetic dataset [43] for BOP unseen object pose estimation track [34]. It contains $\\sim 2\\mathrm{M}$ images of $\\sim 50\\mathrm{K}$ objects from GSO [21] and ShapeNet [5]. We train 3 epochs with a batch size of 32 on 4 RTX 3090 GPUs. The network is trained with Adam [39] and cosine annealing [58], using a base learning rate of $10^{-4}$ . Moreover, we lever", + "bbox": [ + 89, + 719, + 485, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "age InfoNCE [64] loss for learning correlation matrix and weighted BCE loss for supervising overlap prediction.", + "bbox": [ + 511, + 554, + 906, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and Benchmarks. We evaluate our method on BOP test split [33] of three different real-world datasets: LM-O [2], TUD-L [32], and YCB-V [92]. LM-O consists of 200 images from one clutter scene, where eight objects with occlusion are provided for testing. TUD-L comprises three moving objects with mild occlusion and diverse lighting conditions. YCB-V is very challenging due to severe occlusions and sensor noise. It comprises 900 testing images from 12 scenes consisting of 21 objects. For each query object in the test split, we randomly select a reference object beyond the test set (in the training or evaluation set). Unless specified, we constrain the maximal rotation difference between reference and target to $50^{\\circ}$ to ensure the ground-truth correspondence set is not empty. Moreover, the occlusion ratio of reference objects is required to be less than $30\\%$ .", + "bbox": [ + 511, + 590, + 908, + 818 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. For evaluating instance segmentation, we follow [33] to use the mean Average Precision (mAP) metric, which is computed across different Intersection-over-Union (IoU) thresholds (\\{0.5, 0.55, ..., 0.95\\}). Only annotated instances with at least $10\\%$ projected visible surfaces", + "bbox": [ + 511, + 825, + 910, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22028", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/98115c3f7cea2900bcbeea20f01e09c35312b241d8dd2a7157c8611f93c066ed.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CategoryVSDMSSDMSPDARBOP
Dragon47.173.269.463.2
Frog59.452.657.656.4
Can56.865.665.262.5
Avg54.463.764.160.7
", + "bbox": [ + 102, + 95, + 480, + 189 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6eb2f5d7bdcd5fd12f6a209e40b106170a7c2944df73cec60c5589e27e596756.jpg", + "table_caption": [ + "(a)" + ], + "table_footnote": [], + "table_body": "
Ref.
Ref.
", + "bbox": [ + 506, + 89, + 888, + 189 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "are considered for evaluation. For object pose evaluation, we use the symmetry-aware BOP metric [33]. It is calculated as the mean of the Average Recall of Visible Surface Discrepancy (VSD), Maximum Symmetry-Aware Surface Distance (MSSD), and Maximum Symmetry-Aware Projection Distance (MSPD) (See [33] for details).", + "bbox": [ + 89, + 250, + 483, + 340 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. UNO Object Segmentation and Pose Estimation Results", + "text_level": 1, + "bbox": [ + 89, + 348, + 483, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "UNO Object Segmentation Results. We compare our UNOSeg with some existing CAD-model-based unseen object segmentation methods [7, 51, 61] in Tab. 1. Leveraging a single image as a reference, our UNOSeg achieves comparable results with state-of-the-art methods (Ours $54.2\\%$ vs. SAM-6D $54.5\\%$ ) at a faster speed (Ours 2.68s vs. SAM-6D 4.53s). Notably, UNOSeg-FastSAM achieves the best performance on YCB-V (mAP $67.3\\%$ ). Compared to FastSAM [107], SAM [40] achieves consistently better results on three datasets. Therefore, we choose UNOSeg-SAM as the default segmentation for subsequent experiments.", + "bbox": [ + 89, + 387, + 483, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "UNO Object Pose Estimation Results. Since we are the first to conduct relative 6DoF pose estimation for unseen objects and build a brand-new benchmark, we re-implement some traditional [22, 72, 103] and learning-based methods [4, 11, 12, 51, 69] for comparison. Note that we leverage identical data pre-processing, segmentation, query and reference image pairs, and evaluation protocols for all methods for fair comparison. Moreover, we also compare UNOPose with two CAD-model-based pose estimators leveraging $SE(3)$ -invariant PPF feature [1, 41]. All results are illustrated in Tab. 2. It clearly shows that our method surpasses all image-reference-based methods by a large margin, and even achieves comparable results with CAD-model-based methods using class-specific detectors [28, 52] (Ours $70.9\\%$ vs. ZTE-PPF [1] $69.0\\%$ vs. Koenig-PPF [41] $75.1\\%$ ). \"Ref. AlignCenter\" directly uses the reference rotation as prediction, meanwhile leveraging the shift between query and reference centers as relative translation, so it can be regarded as the baseline of all methods. Among all traditional descriptors, PPF with Iterative Closest Point (ICP) refinement [22] achieves the best results (AR $46.9\\%$ ). Learning-based methods demonstrate varying generalization abilities on unseen objects. For example, the GeoTransformer-based", + "bbox": [ + 89, + 553, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b9374e63da4f81e8637ab892dedd0fff979da821e8f2cf7a4f2163be4e239063.jpg", + "table_caption": [ + "(b)", + "Table 3. One reference per category experiments on TUD-L. Blue and green contours denote GT and estimated poses, respectively." + ], + "table_footnote": [], + "table_body": "
ROWGRFLRFFPEVSDMSSDMSPD\\( \\mathrm{AR}_{\\text{BOP}} \\)
A0✘✘✘64.668.159.263.9
A1✓✘✘66.472.863.267.5
A2✘✓✘66.269.560.465.4
A3✓✓✘66.973.263.567.9
B0✘✘✓79.885.075.580.1
B1✓✘✓79.485.276.680.4
B2✘✓✓81.185.176.680.9
B3✓✓✓80.786.077.781.5
", + "bbox": [ + 519, + 247, + 898, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation of GRF and LRF on YCB-V. FPE stands for \"fine pose estimation\".", + "bbox": [ + 511, + 380, + 906, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "method UTOPIC, designed for registering object-level partial point clouds, performs suboptimally in our setting. In contrast, the adapted unseen object pose estimation methods, i.e. SAM-6D [51] and FreeZe [4], achieve satisfactory results compared to their counterparts.", + "bbox": [ + 511, + 436, + 906, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Run-time Analysis. Evaluated on a single RTX 3090 GPU, our pipeline runs at 3.70s for one image, including 2.68s for segmentation and 1.02s for pose estimation (Tab. 1 - 2).", + "bbox": [ + 511, + 512, + 906, + 558 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. One Reference for a Category", + "text_level": 1, + "bbox": [ + 511, + 569, + 777, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To minimize the efforts of acquiring references, we leverage a single reference for all objects of the same category in a dataset. Tab. 3 illustrates the qualitative and quantitative experimental results on TUD-L. It is worth noting that the target object in TUD-L is dynamic rather than static, preventing methods from leveraging background information. This setup poses challenges for relative pose estimation methods that rely on scene features, making them likely to fail in this task. Despite this, UNOPose achieves impressive performance (60.7% in terms of the $\\mathrm{AR}_{\\mathrm{BOP}}$ metric) with only a single reference per category. The visualization results in Tab. 3b further show UNOPose is capable of predicting omnidirectional relative poses. Additional experiments on reference selection are provided in the appendix.", + "bbox": [ + 511, + 593, + 908, + 805 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 816, + 679, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of GRF and LRF. As demonstrated in Sec. 3.3.2, GRF transfers arbitrary object poses to an $SE(3)$ -invariant frame, while LRF assists with extracting precise local descriptors for fine pose estimation. Tab. 4 shows that", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22029", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1d4260866584a2d3b0cbf9ae248d38f31524b6ea23bfb1b4e18c03f28c9404e2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ROWMethodVSDMSSDMSPD\\( \\mathrm{AR}_{\\text{BOP}} \\)
C0B3 + Soft Correspondence [95]77.181.772.777.2
C1B3 + Overlap Predictor82.687.479.483.1
C2A3 + Overlap Predictor68.474.664.869.2
D0C1: Seg. → CNOS [61]77.480.973.677.3
D1C1: Seg. → SAM-6D [51]78.382.174.378.2
D2C1: Seg. → MRCNN [28, 42]81.285.777.981.6
D3C1: Seg. → GT Seg.86.689.282.186.0
E0C1: Ref. Real → Rendered83.789.683.185.5
", + "bbox": [ + 94, + 88, + 478, + 212 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation of overlap predictor, segmentation, and reference type on YCB-V.", + "bbox": [ + 89, + 223, + 483, + 251 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/85f1dfbd821bff5428a02769c25d05f473cb29697d79d389fa2aaf40cca98d1d.jpg", + "image_caption": [ + "Figure 3. Ablation of initial rotation distance between query and reference objects. We categorize all testing objects into nine groups according to the initial rotation distance, and evaluate the $\\mathrm{AR}_{\\mathrm{BOP}}$ metric and overlap ratio for each group separately." + ], + "image_footnote": [], + "bbox": [ + 109, + 266, + 464, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the effectiveness of GRF and LRF is complementary in UNOpose. Specifically, GRF brings significant performance enhancement in the coarse pose estimation (Tab. 4 A1 vs. A0, A3 vs. A2), providing an accurate pose initialization for the next stage. Meanwhile, LRF is constructed to exploit fine-grained local information, primarily enhancing the fine pose estimation results (Tab. 4 B0 vs. B2, B1 vs. B3). With both GRF and LRF, our network achieves obvious performance improvement (Tab. 4 A3 vs. A0, B3 vs. B0).", + "bbox": [ + 89, + 565, + 482, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on Correspondence Loss. In partial-to-partial point cloud registration, only some of the points can find their counterparts on the target point cloud. Therefore, the introduction of overlap prediction helps to distinguish the points in the overlap region and achieves good results (Tab. 4-5 C1 vs. B3, C2 vs. A3). Moreover, we consider soft correspondence loss in [95] as an alternative, which predicts probabilities of overlap between each point pair. However, the result did not meet the expectations (Tab. 5 C0).", + "bbox": [ + 89, + 702, + 482, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on Segmentation. Tab. 1 showcases the advantage of our UNOSeg over other segmentation methods on YCB-V. Furthermore, substituting UNOSeg with CNOS [61], SAM-6D [51], or MRCNN [28, 42] results in a noticeable", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "decline in UNOpose's performance (Tab. 5 D0-D2). Moreover, substituting UNOSeg with ground-truth segmentation brings $\\approx 3\\%$ performance enhancement (Tab. 5 D3 vs. C1).", + "bbox": [ + 511, + 90, + 906, + 136 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Using Rendered References. We replaced the real reference with a pose-identical rendered version and present the results in Tab. 5. With reduced depth noise and domain gap between training and testing, UNOPose works even better using the rendered reference (Tab. 5 E0 vs. C1). However, this ablation requires the object model, which is more difficult to acquire than an unposed real image.", + "bbox": [ + 511, + 136, + 908, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on Reference Rotation Distance. To investigate the impact of the initial rotation distance between the query and reference images, we randomly select the reference object with a maximum rotation distance of $90^{\\circ}$ for each query object. Further, all testing objects are categorized into nine groups according to initial rotation distances. We separately evaluate each group and present the results in Fig. 3. Meanwhile, we calculate the overlap ratio between reference and query instances of each group and visualize the mean and standard deviation. Notably, performance significantly decreases when the rotation distance exceeds $50^{\\circ}$ . This decline is attributed to the reduced overlap between query and reference viewpoints. However, our network still shows favorable transfer ability in extreme relative pose estimation (rotation distance $(80^{\\circ}, 90^{\\circ}]$ , overlap ratio $45.4\\%$ , $\\mathrm{AR}_{\\mathrm{BOP}} 54.8\\%$ ).", + "bbox": [ + 511, + 242, + 908, + 469 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The appendix presents further implementation details, visualization results, and more experimental analysis on reference selection and extensive datasets.", + "bbox": [ + 511, + 469, + 906, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion, Limitation and Future work", + "text_level": 1, + "bbox": [ + 511, + 529, + 877, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work has introduced UNOPose, a new approach for unseen object pose estimation with one reference. The key idea is constructing the $SE(3)$ -invariant reference frame for tackling diverse pose and size variations. Moreover, we propose an overlap predictor for handling low-overlap scenarios. To evaluate the proposed method, we devise a new benchmark based on the BOP challenge and compare some state-of-the-art methods. Experimental results show UNOPose surpasses all compared reference-based methods significantly, and is competitive with some $SE(3)$ -invariant-feature-based methods relying on CAD models.", + "bbox": [ + 511, + 554, + 908, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation and Future Work. While efficient, correspondence-based methods rely on predicting robust and dense correspondence between query and reference objects, thereby limited in extreme two-view geometry applications. Future work will focus on reconstructing the unseen object from a single reference [53, 108] and concurrently estimating its object pose.", + "bbox": [ + 511, + 720, + 908, + 827 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported in part by the National Natural Science Foundation of China under Grant No. 62406169, and in part by the China Postdoctoral Science Foundation under Grant No. 2024M761673.", + "bbox": [ + 511, + 830, + 908, + 886 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22030", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anonymous. Zte-ppf. https://bop.felk(cvut.cz/method_info/250/, 2022. 2, 6, 7", + "[2] Eric Brachmann, Alexander Krull, Frank Michel, Stefan Gumhold, Jamie Shotton, and Carsten Rother. Learning 6D object pose estimation using 3D object coordinates. In ECCV, 2014. 2, 6", + "[3] Dingding Cai, Janne Heikkilä, and Esa Rahtu. Ove6d: Object viewpoint encoding for depth-based 6d object pose estimation. In CVPR, pages 6803-6813, 2022. 1, 2, 3", + "[4] Andrea Caraffa, Davide Boscaini, Amir Hamza, and Fabio Poiesi. Freeze: Training-free zero-shot 6d pose estimation with geometric and vision foundation models. In ECCV, 2024. 3, 6, 7", + "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012, 2015. 6", + "[6] Dengsheng Chen, Jun Li, Zheng Wang, and Kai Xu. Learning canonical shape space for category-level 6D object pose and size estimation. In CVPR, pages 11970-11979, 2020. 1, 3", + "[7] Jianqiu Chen, Zikun Zhou, Mingshan Sun, Rui Zhao, Liwei Wu, Tianpeng Bao, and Zhenyu He. Zeropose: Cadprompted zero-shot object 6d pose estimation in cluttered scenes. IEEE TCSVT, 2024. 4, 6, 7", + "[8] Kai Chen and Qi Dou. SGPA: Structure-guided prior adaptation for category-level 6D object pose estimation. In ICCV, pages 2773-2782, 2021. 3", + "[9] Wei Chen, Xi Jia, Hyung Jin Chang, Jinming Duan, Shen Linlin, and Ales Leonardis. FS-Net: Fast shape-based network for category-level 6D object pose estimation with decoupled rotation mechanism. In CVPR, pages 1581–1590, 2021. 3", + "[10] Yamei Chen, Yan Di, Guangyao Zhai, Fabian Manhardt, Chenyangguang Zhang, Ruida Zhang, Federico Tombari, Nassir Navab, and Benjamin Busam. Secondpose: Se (3)-consistent dual-stream feature fusion for category-level pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9959-9969, 2024. 3", + "[11] Zhilei Chen, Honghua Chen, Lina Gong, Xuefeng Yan, Jun Wang, Yanwen Guo, Jing Qin, and Mingqiang Wei. Utopic: Uncertainty-aware overlap prediction network for partial point cloud registration. In Computer Graphics Forum, pages 87-98. Wiley Online Library, 2022. 6, 7", + "[12] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In ICCV, pages 8958-8966, 2019. 3, 6, 7", + "[13] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In ICLR, 2024. 6", + "[14] Haowen Deng, Tolga Birdal, and Slobodan Ilic. Ppf-foldnet: Unsupervised learning of rotation invariant 3d local descriptors. In ECCV, pages 602-618, 2018. 3" + ], + "bbox": [ + 101, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Haowen Deng, Tolga Birdal, and Slobodan Ilic. Ppfnet: Global context aware local features for robust 3d point matching. In CVPR, pages 195-205, 2018. 3", + "[16] Maximilian Denninger, Dominik Winkelbauer, Martin Sundermeyer, Wout Boerdijk, Markus Knauer, Klaus H. Strobl, Matthias Humt, and Rudolph Triebel. Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software, 8(82):4901, 2023. 1", + "[17] Yan Di, Ruida Zhang, Zhiqiang Lou, Fabian Manhardt, Xi-angyang Ji, Nassir Navab, and Federico Tombari. Gpv-posed Category-level object pose estimation via geometry-guided point-wise voting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6781–6791, 2022. 3", + "[18] Yan Di, Chenyangguang Zhang, Ruida Zhang, Fabian Manhardt, Yongzhi Su, Jason Rambach, Didier Stricker, Xi-angyang Ji, and Federico Tombari. U-red: Unsupervised 3d shape retrieval and deformation for partial point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8884–8895, 2023. 3", + "[19] Yan Di, Chenyangguang Zhang, Chaowei Wang, Ruida Zhang, Guangyao Zhai, Yanyan Li, Bowen Fu, Xiangyang Ji, and Shan Gao. Shapematcher: Self-supervised joint shape canonicalization segmentation retrieval and deformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21017-21028, 2024. 3", + "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 6", + "[21] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In ICRA, pages 2553-2560. IEEE, 2022. 6", + "[22] Bertram Drost, Markus Ulrich, Nassir Navab, and Slobodan Ilic. Model globally, match locally: Efficient and robust 3d object recognition. In CVPR, pages 998-1005. IEEE, 2010. 3, 6, 7", + "[23] Guoguang Du, Kai Wang, Shiguo Lian, and Kaiyong Zhao. Vision-based robotic grasping from object localization, object pose estimation to grasp estimation for parallel grippers: a review. Artificial Intelligence Review, 54(3):1677-1734, 2021. 1", + "[24] Zhaoxin Fan, Zhenbo Song, Jian Xu, Zhicheng Wang, Kejian Wu, Hongyan Liu, and Jun He. Object level depth reconstruction for category level 6d object pose estimation from monocular rgb image. In ECCV, pages 220-236. Springer, 2022. 3", + "[25] Andrea Frome, Daniel Huber, Ravi Kolluri, Thomas Bulow, and Jitendra Malik. Recognizing objects in range data using regional point descriptors. In ECCV, pages 224-237. Springer, 2004. 5", + "[26] Bowen Fu, Sek Kun Leong, Yan Di, Gu Wang, Jiwen Tang, and Xiangyang Ji. Lanpose: Language-instructed 6d object pose estimation for robotic assembly. In ECCVW, 2024. 1" + ], + "bbox": [ + 522, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22031", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Rasmus Laurvig Haugaard and Anders Glent Buch. Surfemb: Dense and continuous correspondence distributions for object pose estimation with learnt surface embeddings. In CVPR, pages 6749-6758, 2022. 3, 5", + "[28] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask R-CNN. In ICCV, pages 2961-2969, 2017. 6, 7, 8", + "[29] Xingyi He, Jiaming Sun, Yuang Wang, Di Huang, Hujun Bao, and Xiaowei Zhou. Onepose++: Keypoint-free one-shot object pose estimation without cad models. NeurIPS, 35:35103-35115, 2022. 1, 3", + "[30] Yisheng He, Haibin Huang, Haoqiang Fan, Qifeng Chen, and Jian Sun. Ffb6d: A full flow bidirectional fusion network for 6d pose estimation. In CVPR, pages 3003-3013, 2021. 3", + "[31] Yisheng He, Yao Wang, Haoqiang Fan, Jian Sun, and Qifeng Chen. Fs6d: Few-shot 6d pose estimation of novel objects. In CVPR, pages 6814-6824, 2022. 1", + "[32] Tomas Hodan, Frank Michel, Eric Brachmann, Wadim Kehl, Anders GlentBuch, Dirk Kraft, Bertram Drost, Joel Vidal, Stephan Ihrke, Xenophon Zabulis, et al. BOP: Benchmark for 6D Object Pose Estimation. In ECCV, pages 19-34, 2018. 2, 6", + "[33] Tomas Hodan, Martin Sundermeyer, Bertram Drost, Yann Labbe, Eric Brachmann, Frank Michel, Carsten Rother, and Jiri Matas. BOP Challenge 2020 on 6D object localization. In ECCVW, pages 577-594, 2020. 6, 7", + "[34] Tomas Hodan, Martin Sundermeyer, Yann Labbe, Van Nguyen Nguyen, Gu Wang, Eric Brachmann, Bertram Drost, Vincent Lepetit, Carsten Rother, and Jiri Matas. Bop challenge 2023 on detection, segmentation and pose estimation of seen and unseen rigid objects. In CVPRW, 2024. 2, 6", + "[35] Yinlin Hu, Pascal Fua, Wei Wang, and Mathieu Salzmann. Single-stage 6d object pose estimation. In CVPR, pages 2930-2939, 2020. 3", + "[36] Junwen Huang, Hao Yu, Kuan-Ting Yu, Nassir Navab, Slobodan Ilic, and Benjamin Busam. Matchu: Matching unseen objects for 6d pose estimation from rgb-d images. In CVPR, 2024. 3", + "[37] Siyuan Huang, Siyuan Qi, Yinxue Xiao, Yixin Zhu, Ying Nian Wu, and Song-Chun Zhu. Cooperative holistic scene understanding: Unifying 3d object, layout, and camera pose estimation. NeurIPS, 31, 2018. 1", + "[38] Shengyu Huang, Zan Gojcic, Mikhail Usvyatsov, Andreas Wieser, and Konrad Schindler. Predator: Registration of 3d point clouds with low overlap. In CVPR, pages 4267-4276, 2021. 3", + "[39] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6", + "[40] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 2, 4, 6, 7", + "[41] Rebecca König and Bertram Drost. A hybrid approach for 6 dof pose estimation. In ECCV, pages 700-706. Springer, 2020. 2, 6, 7" + ], + "bbox": [ + 99, + 92, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. CosyPose: Consistent multi-view multi-object 6D pose estimation. In ECCV, pages 574-591, 2020. 1, 8", + "[43] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. Megapose: 6d pose estimation of novel objects via render & compare. In CoRL, pages 715-725. PMLR, 2023. 1, 2, 3, 6", + "[44] Junha Lee, Seungwook Kim, Minsu Cho, and Jaesik Park. Deep hough voting for robust global registration. In ICCV, pages 15994-16003, 2021. 3", + "[45] Jiahao Li, Changhao Zhang, Ziyao Xu, Hangning Zhou, and Chi Zhang. Iterative distance-aware similarity matrix convolution with mutual-supervised point elimination for efficient point cloud registration. In ECCV, pages 378-394. Springer, 2020. 3", + "[46] Xiaolong Li, Yijia Weng, Li Yi, Leonidas J Guibas, A Abbott, Shuran Song, and He Wang. Leveraging se (3) equivariance for self-supervised category-level object pose estimation from point clouds. NeurIPS, 34, 2021. 5", + "[47] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. DeepIM: Deep iterative matching for 6D pose estimation. IJCV, 128(3):657-678, 2020. 3", + "[48] Zhigang Li, Gu Wang, and Xiangyang Ji. CDPN: Coordinates-Based Disentangled Pose Network for Real-Time RGB-Based 6-DoF Object Pose Estimation. In ICCV, pages 7678-7687, 2019. 1, 3", + "[49] Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. In 3DV, 2024. 3", + "[50] Jiehong Lin, Zewei Wei, Zhihao Li, Songcen Xu, Kui Jia, and Yuanqing Li. DualPoseNet: Category-level 6D object pose and size estimation using dual pose network with refined learning of pose consistency. In ICCV, pages 3560-3569, 2021. 3", + "[51] Jiehong Lin, Lihua Liu, Dekun Lu, and Kui Jia. Sam-6d: Segment anything model meets zero-shot 6d object pose estimation. In CVPR, 2024. 1, 3, 4, 5, 6, 7, 8", + "[52] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 7", + "[53] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In ICCV, pages 9298–9309, 2023. 3, 8", + "[54] Xingyu Liu, Gu Wang, Yi Li, and Xiangyang Ji. CATRE: iterative point clouds alignment for category-level object pose refinement. In ECCV, 2022. 3", + "[55] Xingyu Liu, Yingyue Li, Chengxi Li, Gu Wang, Chenyangguang Zhang, Ziqin Huang, and Xiangyang Ji. Gfreedet: Exploiting gaussian splatting and foundation models for model-free unseen object detection in the bop challenge 2024. arXiv preprint arXiv:2412.01552, 2024. 1", + "[56] Xingyu Liu, Chenyangguang Zhang, Gu Wang, Ruida Zhang, and Xiangyang Ji. Rasim: A range-aware high-fidelity rgb-d data simulation pipeline for real-world applications. In ICRA, pages 17057-17064. IEEE, 2024. 1, 3" + ], + "bbox": [ + 524, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22032", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Yuan Liu, Yilin Wen, Sida Peng, Cheng Lin, Xiaoxiao Long, Taku Komura, and Wenping Wang. Gen6d: Generalizable model-free 6-dof object pose estimation from rgb images. In ECCV, 2022. 1, 2", + "[58] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In *ICLR*, 2017. 6", + "[59] Eric Marchand, Hideaki Uchiyama, and Fabien Spindler. Pose estimation for augmented reality: a hands-on survey. IEEE TVCG, 22(12):2633-2651, 2015. 1", + "[60] Van Nguyen Nguyen, Yinlin Hu, Yang Xiao, Mathieu Salzmann, and Vincent Lepetit. Templates for 3d object pose estimation revisited: Generalization to new objects and robustness to occlusions. In CVPR, pages 6771-6780, 2022. 3", + "[61] Van Nguyen Nguyen, Thibault Groueix, Georgy Ponimatkin, Vincent Lepetit, and Tomas Hodan. Cnos: A strong baseline for cad-based novel object segmentation. In ICCV, pages 2134-2140, 2023. 1, 3, 4, 6, 7, 8", + "[62] Van Nguyen Nguyen, Thibault Groueix, Mathieu Salzmann, and Vincent Lepetit. Gigapose: Fast and robust novel object pose estimation via one correspondence. In CVPR, 2024. 1, 3", + "[63] Yinyu Nie, Xiaoguang Han, Shihui Guo, Yujuan Zheng, Jian Chang, and Jian Jun Zhang. Total3DUnderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In CVPR, pages 55-64, 2020. 1", + "[64] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 6", + "[65] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. TMLR, 2024. 2, 3, 4, 5, 6", + "[66] Evin Pinar Örnek, Yann Labbé, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, and Tomas Hodan. Found- pose: Unseen object pose estimation with foundation features. In CVPR, 2024. 3", + "[67] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. PyTorch: An imperative style, high-performance deep learning library. In NeurlPS, pages 8026-8037, 2019. 6", + "[68] Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, and Hujun Bao. Pynet: Pixel-wise voting network for 6dof pose estimation. In CVPR, pages 4561-4570, 2019. 1, 3", + "[69] Fabio Poiesi and Davide Boscaini. Learning general and distinctive 3d local deep descriptors for point cloud registration. IEEE TPAMI, 45(3):3979-3985, 2022. 3, 6, 7", + "[70] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. PointNet: Deep learning on point sets for 3D classification and segmentation. CVPR, 1(2):4, 2017. 5" + ], + "bbox": [ + 99, + 92, + 485, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[71] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, Slobodan Ilic, Dewen Hu, and Kai Xu. Geotransformer: Fast and robust point cloud registration with geometric transformer. IEEE TPAMI, 2023. 3, 5, 6", + "[72] Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (fpfh) for 3d registration. In ICRA, pages 3212-3217. IEEE, 2009. 3, 6, 7", + "[73] Ivan Shugurov, Sergey Zakharov, and Slobodan Ilic. Dpodv2: Dense correspondence-based 6 dof pose estimation. IEEE TPAMI, 44(11):7417-7435, 2021. 3", + "[74] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. Osop: A multi-stage one shot object pose estimation framework. In CVPR, pages 6835-6844, 2022. 3", + "[75] Fridtjof Stein, Gérard Medioni, et al. Structural indexing: Efficient 3-d object recognition. IEEE TPAMI, 14(2):125-145, 1992. 5", + "[76] Yongzhi Su, Jason Rambach, Nareg Minaskan, Paul Lesur, Alain Pagani, and Didier Stricker. Deep multi-state object pose estimation for augmented reality assembly. In IEEE International Symposium on Mixed and Augmented Reality Adjunct (ISMAR-Adjunct), pages 222–227, 2019. 1", + "[77] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. Onepose: One-shot object pose estimation without cad models. In CVPR, pages 6825-6834, 2022. 1, 2, 3", + "[78] Jingyi Tang, Gu Wang, Zeyu Chen, Shengquan Li, Xiu Li, and Xiangyang Ji. Fafa: Frequency-aware flow-aided self-supervision for underwater object pose estimation. In ECCV, pages 352–369. Springer, 2024. 1", + "[79] Meng Tian, Marcelo H Ang Jr, and Gim Hee Lee. Shape prior deformation for categorical 6D object pose and size estimation. In ECCV, 2020. 1, 3", + "[80] Jonathan Tremblay, Thang To, Balakumar Sundaralingam, Yu Xiang, Dieter Fox, and Stan Birchfield. Deep object pose estimation for semantic robotic grasping of household objects. In CoRL, pages 306-316, 2018. 1", + "[81] Mikaela Angelina Uy, Vladimir G Kim, Minhyuk Sung, Noam Aigerman, Siddhartha Chaudhuri, and Leonidas J Guibas. Joint learning of 3d shape retrieval and deformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11713-11722, 2021. 3", + "[82] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 6", + "[83] Chen Wang, Danfei Xu, Yuke Zhu, Roberto Martin-Martin, Cewu Lu, Li Fei-Fei, and Silvio Savarese. DenseFusion: 6D object pose estimation by iterative dense fusion. In CVPR, pages 3343–3352, 2019. 1, 3", + "[84] Gu Wang, Fabian Manhardt, Jianzhun Shao, Xiangyang Ji, Nassir Navab, and Federico Tombari. Self6d: Self-supervised monocular 6d object pose estimation. In ECCV, pages 108-125. Springer, 2020.", + "[85] Gu Wang, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. GDR-Net: Geometry-guided direct regression network for monocular 6D object pose estimation. In CVPR, pages 16611–16621, 2021. 3" + ], + "bbox": [ + 522, + 92, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22033", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "ref_text", + "text": "[94] Yang You, Ruoxi Shi, Weiming Wang, and Cewu Lu. CPPf: Towards robust category-level 9d pose estimation in the wild. In CVPR, pages 6866-6875, 2022. 5", + "bbox": [ + 101, + 498, + 483, + 540 + ], + "page_idx": 11 + }, + { + "type": "ref_text", + "text": "[95] Hao Yu, Fu Li, Mahdi Saleh, Benjamin Busam, and Slobodan Ilic. Cofinet: Reliable coarse-to-fine correspondences for robust pointcloud registration. NeurIPS, 34, 2021. 8", + "bbox": [ + 101, + 542, + 483, + 583 + ], + "page_idx": 11 + }, + { + "type": "ref_text", + "text": "[96] Hao Yu, Zheng Qin, Ji Hou, Mahdi Saleh, Dongsheng Li, Benjamin Busam, and Slobodan Ilic. Rotation-invariant transformer for point cloud matching. In CVPR, pages 5384-5393, 2023. 5", + "bbox": [ + 101, + 585, + 483, + 640 + ], + "page_idx": 11 + }, + { + "type": "ref_text", + "text": "[97] Hao Yu, Ji Hou, Zheng Qin, Mahdi Saleh, Ivan Shugurov, Kai Wang, Benjamin Busam, and Slobodan Ilic. Riga: Rotation-invariant and globally-aware descriptors for point cloud registration. IEEE TPAMI, 46(5):3796-3812, 2024. 3", + "bbox": [ + 101, + 643, + 483, + 699 + ], + "page_idx": 11 + }, + { + "type": "ref_text", + "text": "[98] Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose: Predicting probabilistic relative rotation for single objects in the wild. In ECCV, pages 592-611. Springer, 2022. 2, 3", + "bbox": [ + 101, + 702, + 483, + 756 + ], + "page_idx": 11 + }, + { + "type": "ref_text", + "text": "[99] Ruida Zhang, Yan Di, Zhiqiang Lou, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. Rbp-posed: Residual bounding box projection for category-level pose estimation. In European conference on computer vision, pages 655–672. Springer, 2022. 3", + "bbox": [ + 101, + 758, + 483, + 828 + ], + "page_idx": 11 + }, + { + "type": "ref_text", + "text": "[100] Ruida Zhang, Yan Di, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. Ssp-pose: Symmetry-aware shape prior deformation for direct category-level object pose estimation. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7452–7459. IEEE, 2022.", + "bbox": [ + 94, + 830, + 482, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[86] Gu Wang, Fabian Manhardt, Xingyu Liu, Xiangyang Ji, and Federico Tombari. Occlusion-aware self-supervised monocular 6D object pose estimation. IEEE TPAMI, 2024. 1" + ], + "bbox": [ + 94, + 90, + 485, + 145 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[87] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6D object pose and size estimation. In CVPR, pages 2642-2651, 2019. 1, 3", + "[88] Jiaze Wang, Kai Chen, and Qi Dou. Category-level 6D object pose estimation via cascaded relation and recurrent reconstruction networks. In IROS, 2021. 1", + "[89] Yue Wang and Justin M Solomon. Deep closest point: Learning representations for point cloud registration. In ICCV, pages 3523-3532, 2019. 3", + "[90] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. FoundationPose: Unified 6d pose estimation and tracking of novel objects. In CVPR, 2024. 1, 2, 3", + "[91] Chin-Hsuan Wu, Yen-Chun Chen, Bolivar Solarte, Lu Yuan, and Min Sun. ifusion: Inverting diffusion for pose-free reconstruction from sparse views. arXiv preprint arXiv:2312.17250, 2023. 3", + "[92] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6D object pose estimation in cluttered scenes. RSS, 2018. 2, 6", + "[93] Jiaqi Yang, Qian Zhang, Yang Xiao, and Zhiguo Cao. Toldi: An effective and robust approach for 3d local shape description. PR, 65:175-187, 2017. 3, 5", + "[101] Ruida Zhang, Ziqin Huang, Gu Wang, Chenyangguang Zhang, Yan Di, Xingxing Zuo, Jiwen Tang, and Xiangyang Ji. Lapose: Laplacian mixture shape modeling for rgb-based category-level object pose estimation. In ECCV, 2024. 3", + "[102] Ruida Zhang, Chenyangguang Zhang, Yan Di, Fabian Manhardt, Xingyu Liu, Federico Tombari, and Xiangyang Ji. Kp-red: Exploiting semantic keypoints for joint 3d shape retrieval and deformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20540-20550, 2024. 3", + "[103] Xiyu Zhang, Jiaqi Yang, Shikun Zhang, and Yanning Zhang. 3d registration with maximal cliques. In CVPR, pages 17745–17754, 2023. 6, 7", + "[104] Chen Zhao, Yinlin Hu, and Mathieu Salzmann. Fusing local similarities for retrieval-based 3d orientation estimation of unseen objects. In ECCV, pages 106-122. Springer, 2022. 3", + "[105] Chen Zhao, Tong Zhang, and Mathieu Salzmann. 3d-aware hypothesis & verification for generalizable relative object pose estimation. In ICLR, 2023. 2, 3", + "[106] Chen Zhao, Tong Zhang, Zheng Dang, and Mathieu Salzmann. Dvmnet: Computing relative pose for unseen objects beyond hypotheses. In CVPR, 2024. 2, 3", + "[107] Xu Zhao, Wenchao Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, and Jinqiao Wang. Fast segment anything. arXiv preprint arXiv:2306.12156, 2023. 6, 7", + "[108] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splattering: Fast and generalizable single-view 3d reconstruction with transformers. In CVPR, pages 10324-10335, 2024. 8" + ], + "bbox": [ + 94, + 90, + 908, + 513 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "22034", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/UNOPose_ Unseen Object Pose Estimation with an Unposed RGB-D Reference Image/94d3840c-f3fe-4b2b-9388-48ccee10c9d7_model.json b/2025/UNOPose_ Unseen Object Pose Estimation with an Unposed RGB-D Reference Image/94d3840c-f3fe-4b2b-9388-48ccee10c9d7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8b2083aec3b244b19020a0b33e0457c7f9d6d325 --- /dev/null +++ b/2025/UNOPose_ Unseen Object Pose Estimation with an Unposed RGB-D Reference Image/94d3840c-f3fe-4b2b-9388-48ccee10c9d7_model.json @@ -0,0 +1,2776 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.26, + 0.131, + 0.74, + 0.176 + ], + "angle": 0, + "content": "UNOpose: Unseen Object Pose Estimation with an Unposed RGB-D Reference Image" + }, + { + "type": "text", + "bbox": [ + 0.206, + 0.203, + 0.794, + 0.257 + ], + "angle": 0, + "content": "Xingyu Liu\\(^{1, *}\\), Gu Wang\\(^{1, *}\\), Ruida Zhang\\(^{1}\\), Chenyangguang Zhang\\(^{1}\\), Federico Tombari\\(^{2,3}\\), and Xiangyang Ji\\(^{1}\\) \n\\(^{1}\\)Tsinghua University, BNRist \\(^{2}\\)Technical University of Munich, \\(^{3}\\)Google" + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.26, + 0.858, + 0.275 + ], + "angle": 0, + "content": "{liuxy21@mails., wanggu1@, zhangrd23@mails., zcyg22@mails., xyji@}tsinghua.edu.cn," + }, + { + "type": "text", + "bbox": [ + 0.42, + 0.278, + 0.573, + 0.291 + ], + "angle": 0, + "content": "tombari@in.tum.de" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.486, + 0.721 + ], + "angle": 0, + "content": "Unseen object pose estimation methods often rely on CAD models or multiple reference views, making the onboarding stage costly. To simplify reference acquisition, we aim to estimate the unseen object's pose through a single unposed RGB-D reference image. While previous works leverage reference images as pose anchors to limit the range of relative pose, our scenario presents significant challenges since the relative transformation could vary across the entire \\(SE(3)\\) space. Moreover, factors like occlusion, sensor noise, and extreme geometry could result in low viewpoint overlap. To address these challenges, we present a novel approach and benchmark, termed UNOPose1, for UNseen One-reference-based object Pose estimation. Building upon a coarse-to-fine paradigm, UNOPose constructs an \\(SE(3)\\)-invariant reference frame to standardize object representation despite pose and size variations. To alleviate small overlap across viewpoints, we recalibrate the weight of each correspondence based on its predicted likelihood of being within the overlapping region. Evaluated on our proposed benchmark based on the BOP Challenge, UNOPose demonstrates superior performance, significantly outperforming traditional and learning-based methods in the one-reference setting and remaining competitive with CAD-model-based methods. The code and dataset are available at github.com/shanice-l/UNOPose." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.75, + 0.222, + 0.765 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.775, + 0.485, + 0.866 + ], + "angle": 0, + "content": "Localizing an object in Euclidean space by estimating its 6DoF pose, i.e., 3DoF orientation and 3DoF position, plays a crucial role in augmented/virtual reality [59, 76], scene understanding [37, 63] and robotic manipulation [23, 26, 80]. The vast majority of works [42, 48, 68, 78, 83-86] focuses on instance-level object pose estimation, where the train" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.328, + 0.909, + 0.493 + ], + "angle": 0, + "content": "ing and testing datasets consist of an identical set of known object instances, and the CAD models of objects are often required for generating training images and labels [16, 56]. More recently, Wang et al. [87] and its successors [6, 79, 88] extended this paradigm to the category level, intending to estimate poses for novel instances within predefined categories without requiring CAD models of target objects. Both paradigms have limitations in open-world applications, where annotating and training for new objects beyond known categories would be very labor-intensive and sometimes prohibitive." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.909, + 0.774 + ], + "angle": 0, + "content": "To mitigate this problem, recent research efforts [3, 43, 51, 90] have shifted towards pose estimation of arbitrary novel objects, where the target object is unseen during training. This task presents a great challenge due to the inherent uncertainty of the novel object's canonical frame (defined in the CAD model of the object). Current approaches [43, 57, 62] often tackle this by using omnidirectional reference views to cover the target object, separating pose estimation task to viewpoint selection and relative pose estimation to known pose anchors. If CAD models are available during the test phase, reference views can be easily produced through image rendering [43, 51, 61, 90]. Alternatively, without CAD models, reference views can be obtained by capturing multiple images of the object [29, 31, 55, 77]. Both setups have shortcomings, as creating a CAD model of an object or labeling a substantial number of object poses is limited in scalability. For example, they can hardly adapt to in-the-wild scenarios where novel objects are introduced unpredictably." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.911, + 0.903 + ], + "angle": 0, + "content": "To sum up, while existing methods exhibit promising transfer capability to novel objects, they are notably constrained by their reliance on the CAD model or multiple reference views (Fig. 1 (a, b)). In this work, we aim to minimize onboarding cost by estimating novel object poses using a single unposed RGB-D reference image (Fig. 1). This setup is different from existing methods mainly in two aspects. First, we focus on estimating the relative pose between" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.875, + 0.225, + 0.888 + ], + "angle": 0, + "content": "*Equal contributions." + }, + { + "type": "page_footnote", + "bbox": [ + 0.116, + 0.888, + 0.388, + 0.901 + ], + "angle": 0, + "content": "1UNO \\((\\prime \\mathrm{u}:\\mathrm{no}\\overline{\\mathrm{u}} /\\) ) means one in Spanish and Italian." + }, + { + "type": "list", + "bbox": [ + 0.11, + 0.875, + 0.388, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22023" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.905, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.362, + 0.908, + 0.406 + ], + "angle": 0, + "content": "Figure 1. Illustration of unseen object pose estimation. Given a query image presenting a target object unseen during training, we aim to estimate its segmentation and 6DoF pose w.r.t. a reference frame. While previous methods [43, 57, 77, 90] often rely on the CAD model or multiple RGB(-D) images for reference, we merely use one unposed RGB-D reference image." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.413, + 0.484, + 0.581 + ], + "angle": 0, + "content": "two unposed viewpoints rather than the absolute pose. Note that the absolute pose of a novel object becomes ill-posed without a certain canonical frame, while relative pose estimation remains well-defined regardless of the canonical frame's definition. Secondly, some existing relative pose estimation methods [98, 105, 106] merely use RGB modality to predict the 3DoF rotation, limiting their ability to predict relative translation and generalize beyond their training datasets. In contrast, we fully leverage RGB-D modality to predict the 6DoF relative pose, meanwhile enhancing the network's generalizability to novel objects and environments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Nevertheless, estimating the pose of unseen objects with only one reference view presents significant challenges from various aspects. First, previous methods [3, 43] select the most similar reference view as a pose anchor, distinctly reducing the search space for relative poses. However, in our scenario, where both the target and reference object poses are unknown, the relative pose can vary across the entire \\(SE(3)\\) space. Moreover, in partial-to-partial object matching, factors such as occlusion, sensor noise, and extreme geometry can severely interfere with the matching process. To address these challenges, we propose a novel approach and benchmark for UNseen One-reference-based object Pose estimation (UNOPose). Our pipeline leverages the strong generalization capabilities of vision foundation models [40, 65] to produce an effective segmentation of the unseen object. Then, UNOPose employs a coarse-to-fine paradigm for estimating the relative pose between the reference and the query objects. To alleviate the challenge of diverse pose and size variations, we introduce the \\(SE(3)\\)-invariant global reference frame (GRF) to standardize object representation. Subsequently, a hierarchical object encoding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.907, + 0.489 + ], + "angle": 0, + "content": "paradigm based on the local reference frame (LRF) further captures fine-grained geometric details. For achieving reliable correspondences in partial-to-partial object matching, we harness an overlap predictor to identify and concentrate on the overlapping region." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.909, + 0.64 + ], + "angle": 0, + "content": "Moreover, we propose a new benchmark devised from the BOP challenge [34] to facilitate evaluation and future research of unseen object pose estimation with one reference. Extensive experiments on YCB-V [92], LM-O [2] and TUD-L [32] datasets demonstrate that our UNOPose surpasses all compared methods on a single reference setting. To our surprise, UNOPose with an unposed reference is even on par with some \\(SE(3)\\)-invariant-feature-based methods relying on CAD models (Ours \\(70.9\\%\\) vs. ZTE-PPF [1] \\(69.0\\%\\) vs. Koenig-PPF [41] \\(75.1\\%\\) w.r.t. \\(\\mathrm{AR}_{\\mathrm{BOP}}\\) metric)." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.641, + 0.859, + 0.655 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.657, + 0.905, + 0.701 + ], + "angle": 0, + "content": "- To the best of our knowledge, we are the first to conduct unseen object 6DoF pose estimation leveraging a single unposed RGB-D reference." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.702, + 0.907, + 0.776 + ], + "angle": 0, + "content": "- Based on the BOP Challenge, we devise a new extensive benchmark tailored for unseen object segmentation and pose estimation with one reference. Additionally, we evaluate several traditional and learning-based methods on this benchmark for completeness." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.778, + 0.909, + 0.899 + ], + "angle": 0, + "content": "- We introduce UNOPose, a network for learning relative transformation between reference and query objects. To achieve this, we propose the \\(SE(3)\\)-invariant global and local reference frames, enabling standardized object representations despite variations in pose and size. Furthermore, the network can automatically adjust the confidence of each correspondence by incorporating an overlap predictor." + }, + { + "type": "list", + "bbox": [ + 0.533, + 0.657, + 0.909, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22024" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.233, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.417 + ], + "angle": 0, + "content": "Class-specific Pose Estimation. Class-specific object pose estimation aims at predicting 6DoF poses of either instance of a known object (instance-level) or unseen objects within a known category (category-level). Instance-level object poses are often solved by direct regression [35, 47, 56, 85] using deep neural networks, or by establishing 2D-3D [27, 48, 68] or 3D-3D correspondences [30, 73, 83] which are then leveraged by RANSAC-based PnP/Kabsch algorithms. The instance-level setting requires expensive data generation/annotation [18, 19, 81, 102] and training for every new object. To address this limitation, category-level pose estimation [87] is proposed to estimate 9DoF poses of novel objects among specific categories without CAD models. Mainstream approaches can also be categorized into direct regression [6, 9, 10, 17, 50, 54, 99–101] or correspondence-based [8, 24, 79, 87] methods. However, this setting remains constrained to a limited number of categories, given the additional challenge of aligning canonical frames, managing symmetric objects, and categorizing objects in similar categories correctly (e.g. can and bottle)." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.418, + 0.484, + 0.75 + ], + "angle": 0, + "content": "Novel Object Pose Estimation. Estimating the pose of novel objects beyond known classes, i.e., objects at inference time are unseen during training, is a useful yet challenging task. Existing works solve this problem through image-based matching [3, 43, 60, 61, 74, 90, 104] or feature-based matching approaches [4, 29, 36, 51, 62, 66, 77]. Notably, given a target image, OVE6D [3] and MegaPose [43] retrieve the most similar viewpoint from a pre-rendered omnidirectional image database as a coarse pose estimate. A customized neural network then refines the coarse pose estimate. FoundationPose [90] further builds a pose ranking network to score each refined pose hypothesis. Feature-matching methods learn local feature descriptors to construct pixel-level or point-level correspondences. For example, OnePose [77] and OnePose++ [29] reconstruct the 3D point cloud of an unseen object to establish 2D-3D correspondences, whereas SAM-6D [51] builds 3D-3D correspondences. More recently, some methods [4, 66] directly exploit the power of foundation models for zero-shot pose estimation. Exemplarily, FoundPose [66] extracts DINOv2 visual features [65], while FreeZe [4] further utilizes generalizable geometric features [69] to conduct feature matching." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.751, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Relative Pose Estimation. Orthogonal to previous works which rely on known instances or multiple reference views, relative object pose estimation estimates a relative transformation of the unseen object with only one reference image. One closely related topic is camera pose estimation from two views. For example, RelPose [98] and RelPose++ [49] infer a distribution of relative rotations by leveraging an energy-based formulation. Moreover, iFusion [91] optimizes the unknown relative pose by inverting the novel view synthesis diffusion model [53]. Relative camera pose estimation relies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.242 + ], + "angle": 0, + "content": "on static background to establish correspondence between viewpoints. In the field of relative object pose estimation, given two RGB images, 3DAHV [105] presents a hypothesis-and-verification framework to score each relative pose hypothesis. Extended by that, DVMNet [106] introduces a hypothesis-free pipeline to compute relative poses via deep voxel matching. The above works have limited capability in inferring 3DoF relative translations. In this work, with the introduction of depth data, we can estimate a complete 6DoF pose including rotation and translation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.243, + 0.909, + 0.469 + ], + "angle": 0, + "content": "Point Cloud Registration. Our approach follows the paradigm of correspondence-based methods [14, 15, 45, 89], which initially extract 3D local feature descriptors to establish correspondences between source and target point clouds. Subsequently, they estimate the relative transformation using techniques like SVD, RANSAC, or Hough Voting [44]. Some feature descriptors are proposed to ensure rotational invariance. For example, FPFH [72] descriptor is computed with geometric surface properties. PPF [22] uses the distance and angles to describe the relation of point pairs. TOLDI [93] proposes a robust local reference frame using keypoint normals and neighboring projections. With the advent of deep learning, neural networks are increasingly employed to extract 3D local or global feature descriptors [12, 38, 71, 89, 97]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.483, + 0.909, + 0.517 + ], + "angle": 0, + "content": "3. UNO Object Segmentation and Pose Estimation" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.526, + 0.719, + 0.541 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.907, + 0.625 + ], + "angle": 0, + "content": "Assuming an arbitrary unseen rigid object in a query image, our goal is to estimate the object's mask \\(M_q\\) and its 6DoF relative pose \\(\\Delta \\mathbf{T} \\in SE(3)\\) with a masked RGB-D reference image exhibiting the target object without major occlusion or truncation. As illustrated in Fig. 1, the input is:" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.625, + 0.854, + 0.641 + ], + "angle": 0, + "content": "1) \\([I_q|D_q]\\in \\mathbb{R}^{H\\times W\\times 4}\\): The query RGB-D image;" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.64, + 0.907, + 0.685 + ], + "angle": 0, + "content": "2) \\(\\left[\\bar{I}_p|D_p\\right] \\in \\mathbb{R}^{H \\times W \\times 4}\\) and \\(M_p \\in \\mathbb{R}^{H \\times W}\\): The reference RGB-D image and the corresponding binary mask indicating the target object;" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.686, + 0.905, + 0.715 + ], + "angle": 0, + "content": "3) \\(K_{q} \\in \\mathbb{R}^{3 \\times 3}\\) and \\(K_{p} \\in \\mathbb{R}^{3 \\times 3}\\): Camera intrinsics of the query and reference images." + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.625, + 0.907, + 0.715 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.717, + 0.907, + 0.837 + ], + "angle": 0, + "content": "Optionally, if the pose \\(\\mathbf{T}_p \\in SE(3)\\) of the reference object in the camera frame is known, the query object pose \\(\\mathbf{T}_q \\in SE(3)\\) can be recovered by \\(\\mathbf{T}_q = \\Delta \\mathbf{T} \\mathbf{T}_p\\). Note that, for practicality, the method should not rely on the absolute pose of the reference object, since the world frame can vary arbitrarily in different applications. We only use the reference object pose during the inference stage for evaluation on standard object pose datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.847, + 0.756, + 0.864 + ], + "angle": 0, + "content": "3.2. UNO Object Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Firstly, we segment the query object from a cluttered background. Thanks to the great generalization ability of vision" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22025" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.904, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.435, + 0.908, + 0.493 + ], + "angle": 0, + "content": "Figure 2. The network architecture of UNOpose. Given the query and reference point clouds \\(\\mathbf{Q}_{cam}\\) and \\(\\mathbf{P}_{cam}\\) in the camera frame, UNOpose first transforms them into the \\(SE(3)\\)-invariant global reference frame (GRF). Then feature descriptors are extracted from sparse point sets for constructing the coarse correlation matrix. For achieving precise correspondences, the fine pose estimation module exploits structural details using positional encoding and local reference frame (LRF) encoding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.484, + 0.675 + ], + "angle": 0, + "content": "foundation models, methods like [7, 51, 61] can effectively segment novel objects with CAD models. However, unlike previous approaches, which generate diverse descriptors from multiple rendered views, we can access only a single reference image. To address this challenge, we use the SAM model [40] to predict all possible mask proposals from the query image, and then score each mask proposal by comparing DINOv2 [65] descriptors between the query and reference images using cosine similarity, identifying the most similar mask \\( M_{q} \\). Please refer to the appendix for more details about UNO segmentation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.684, + 0.353, + 0.701 + ], + "angle": 0, + "content": "3.3. UNO Object Pose Estimation" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.708, + 0.303, + 0.723 + ], + "angle": 0, + "content": "3.3.1 Overview of UNOPose" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.732, + 0.485, + 0.858 + ], + "angle": 0, + "content": "Given the predicted mask \\(M_{q}\\) of the query image and mask \\(M_{p}\\) of the reference image, we crop and back-project the object of interest from depth maps \\(D_{q}, D_{p}\\) into the camera space as two point sets \\(\\mathbf{Q}_{cam} \\in \\mathbb{R}^{N^{Q} \\times 3}\\) and \\(\\mathbf{P}_{cam} \\in \\mathbb{R}^{N^{P} \\times 3}\\), where \\(N^{Q}\\) and \\(N^{P}\\) denotes the point numbers of query and reference point clouds, respectively. Our goal is to recover the relative transformation \\(\\Delta \\mathbf{T} = \\{\\Delta \\mathbf{R}, \\Delta \\mathbf{t}\\}\\) by minimizing the correspondence distance" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.87, + 0.484, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\min \\sum_ {(\\mathbf {q}, \\mathbf {p}) \\in \\mathbf {C}} \\| \\Delta \\mathbf {R} \\mathbf {q} + \\Delta \\mathbf {t} - \\mathbf {p} \\| _ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.908, + 0.584 + ], + "angle": 0, + "content": "where \\(\\mathbf{C}\\) is the predicted correspondence set between \\(\\mathbf{Q}_{cam}\\) and \\(\\mathbf{P}_{cam}\\). We exploit both color and geometric cues to build this correspondence and follow the broadly used coarse-to-fine paradigm in point cloud registration to solve Eq. (1). The network is demonstrated in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.6, + 0.783, + 0.615 + ], + "angle": 0, + "content": "3.3.2 Coarse-to-fine Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.624, + 0.907, + 0.715 + ], + "angle": 0, + "content": "Constructing a Pose-invariant Reference Frame. Given only an unposed reference image, the relative pose to predict is arbitrary in the \\(SE(3)\\) space, which renders a significant challenge for achieving robust correspondences. Hence, we introduce a pose-invariant global reference frame (GRF) and transform \\(\\mathbf{Q}_{cam}\\) and \\(\\mathbf{P}_{cam}\\) into GRF as \\(\\mathbf{Q}_G\\) and \\(\\mathbf{P}_G\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.715, + 0.907, + 0.761 + ], + "angle": 0, + "content": "Concretely, taking \\(\\mathbf{Q}_{cam}\\) as an example, transforming the point cloud to GRF involves a 7DoF coordinate transformation \\(\\{\\mathbf{R}_G\\in SO(3),\\mathbf{t}_G\\in \\mathbb{R}^3,s_G\\in \\mathbb{R}\\}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.765, + 0.907, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {G} = \\left\\{\\mathbf {R} _ {G} ^ {\\top} \\left(\\mathbf {q} - \\mathbf {t} _ {G}\\right) / s _ {G} \\mid \\mathbf {q} \\in \\mathbf {Q} _ {c a m} \\right\\}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.79, + 0.907, + 0.835 + ], + "angle": 0, + "content": "The origin of GRF is located at the object center \\(\\mathbf{c}_Q\\) for translation invariance, and the radius of the point cloud is rescaled to 1 for size invariance, computed as" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.842, + 0.907, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\mathbf {t} _ {G} = \\mathbf {c} _ {Q}, \\quad s _ {G} = \\max _ {\\mathbf {q} \\in \\mathbf {Q} _ {c a m}} \\| \\mathbf {q} - \\mathbf {c} _ {Q} \\| _ {2}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "The key ingredient is to devise the rotation \\(\\mathbf{R}_G = [\\mathbf{r}_{Gx}|\\mathbf{r}_{Gy}|\\mathbf{r}_{Gz}]\\), where \\(\\mathbf{r}_{Gx},\\mathbf{r}_{Gy},\\mathbf{r}_{Gz}\\) are the three columns" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22026" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.272 + ], + "angle": 0, + "content": "of \\(\\mathbf{R}_G\\), ensuring orientation invariance for the transformed point cloud. Inspired by former works [25, 75, 93], we use the normal vector of the object center \\(\\mathbf{n}(\\mathbf{c}_Q)\\) as \\(\\mathbf{r}_{Gz}\\), and project all points in \\(\\mathbf{Q}_{cam}\\) onto the tangent plane of \\(\\mathbf{c}_Q\\), summarizing the projected vectors to determine \\(\\mathbf{r}_{Gx}\\). Then, \\(\\mathbf{r}_{Gy}\\) is obtained by the cross-product of \\(\\mathbf{r}_{Gx}\\) and \\(\\mathbf{r}_{Gz}\\), i.e., \\(\\mathbf{r}_{Gy} = \\mathbf{r}_{Gx} \\times \\mathbf{r}_{Gz}\\). This ensures that the XY-plane evenly divides the point cloud, with the x-axis representing the principal projection direction. The GRF is \\(SE(3)\\)-invariant as it can be uniquely determined by the spatial distribution of the point cloud, making the transformed point cloud robust to variations in pose and size." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.274, + 0.484, + 0.322 + ], + "angle": 0, + "content": "Concretely, we perform SVD on the covariance matrix \\(\\mathrm{Cov}(\\mathbf{Q}_{cam}) = \\frac{1}{N_Q}\\mathbf{Q}_{cam}^\\top \\mathbf{Q}_{cam} - \\mathbf{c}_Q\\mathbf{c}_Q^\\top\\) and use the singular vector w.r.t. the minimum singular value to determine \\(\\mathbf{r}_{Gz}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.335, + 0.484, + 0.388 + ], + "angle": 0, + "content": "\\[\n\\mathbf {r} _ {G z} = \\left\\{ \\begin{array}{l l} \\mathbf {n} \\left(\\mathbf {c} _ {Q}\\right), & \\text {i f} \\mathbf {n} \\left(\\mathbf {c} _ {Q}\\right) ^ {\\top} \\sum_ {\\mathbf {q} \\in \\mathbf {Q} _ {c a m}} \\left(\\mathbf {c} _ {Q} - \\mathbf {q}\\right) > 0 \\\\ - \\mathbf {n} \\left(\\mathbf {c} _ {Q}\\right), & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.389, + 0.305, + 0.404 + ], + "angle": 0, + "content": "Afterwards, \\(\\mathbf{r}_{Gx}\\) is computed as" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.417, + 0.484, + 0.452 + ], + "angle": 0, + "content": "\\[\n\\mathbf {r} _ {G x} = \\sum_ {\\mathbf {q} \\in \\mathbf {Q} _ {c a m}} w _ {q} \\left(\\left(\\mathbf {q} - \\mathbf {c} _ {Q}\\right) - \\mathbf {r} _ {G z} ^ {\\top} \\left(\\mathbf {q} - \\mathbf {c} _ {Q}\\right) \\mathbf {r} _ {G z}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.464, + 0.483, + 0.495 + ], + "angle": 0, + "content": "where \\(w_{q}\\) is a weight w.r.t. distance between \\(\\mathbf{q}\\) and \\(\\mathbf{c}_Q\\) (See the appendix for details)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.496, + 0.484, + 0.57 + ], + "angle": 0, + "content": "Previous works often depend on complex networks or computationally expensive PPF features for \\(SE(3)\\) invariance [46, 94, 96]. In contrast, our transformation is computationally efficient and can seamlessly adapt to a wide range of network architectures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.573, + 0.484, + 0.756 + ], + "angle": 0, + "content": "Coarse Pose Estimation. Given point clouds \\(\\mathbf{Q}_G\\) and \\(\\mathbf{P}_G\\) in GRF, we sample two sparse point sets \\(\\mathbf{Q}_G^c\\in \\mathbb{R}^{N^c\\times 3}\\) and \\(\\mathbf{P}_G^c\\in \\mathbb{R}^{N^c\\times 3}\\) to efficiently obtain a coarse pose initialization \\(\\Delta \\mathbf{T}_{init}\\). Specifically, we leverage a geometry encoder [71] and a color encoder [65] to extract point cloud and RGB features separately. Features are further concatenated as \\(f_{Q}^{c}\\in \\mathbb{R}^{N^{c}\\times d}\\) and \\(f_{P}^{c}\\in \\mathbb{R}^{N^{c}\\times d}\\), where \\(d\\) is the dimension of embeddings. Following [51], we add a learnable background token for assigning non-overlapping points. The embeddings, denoted as \\(\\hat{f}_Q^c\\in \\mathbb{R}^{(N^c +1)\\times d}\\) and \\(\\hat{f}_P^c\\in \\mathbb{R}^{(N^c +1)\\times d}\\), are processed by three stacked Geometric Transformer decoding modules [71]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.758, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The outputs of the last decoder \\(\\hat{F}_Q^c\\in \\mathbb{R}^{(N^c +1)\\times D}\\) and \\(\\hat{F}_P^c\\in \\mathbb{R}^{(N^c +1)\\times D}\\) are point-wise features for building the correlation matrix. However, in our setting, the overlap ratio can be influenced by complicated factors like viewpoint, occlusion or depth noise. To solve this issue, the network additionally predicts overlap confidence \\(\\hat{O}_Q^c\\in \\mathbb{R}^{(N^c +1)\\times 1}\\) and \\(\\hat{O}_P^c\\in \\mathbb{R}^{(N^c +1)\\times 1}\\), which indicate point-wise probabilities of being in the overlapping region. The overlap-aware correlation matrix \\(\\mathbf{X}^c\\in \\mathbb{R}^{(N^c +1)\\times (N^c +1)}\\) can thus be com" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.093, + 0.577, + 0.107 + ], + "angle": 0, + "content": "puted by" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.115, + 0.907, + 0.136 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} ^ {c} = \\operatorname {s o f t m a x} [ (\\hat {O} _ {Q} ^ {c} \\odot \\hat {F} _ {Q} ^ {c}) (\\hat {O} _ {P} ^ {c} \\odot \\hat {F} _ {P} ^ {c}) ^ {\\top} ]. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.143, + 0.907, + 0.188 + ], + "angle": 0, + "content": "Here \\(\\odot\\) indicates element-wise multiplication. Each element in \\(\\mathbf{X}^c\\) suggests a correlation score between a point pair in \\(\\mathbf{Q}^c\\) and \\(\\mathbf{P}^c\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.189, + 0.909, + 0.295 + ], + "angle": 0, + "content": "Once \\(\\mathbf{X}^c\\) is computed, we can extract all possible corresponding point pairs between \\(\\mathbf{Q}^c\\) and \\(\\mathbf{P}^c\\) as well as their correlation scores to solve Eq. (1). We first compute pose hypotheses by sampling \\(N_{H}\\) triplets of point pairs according to the distribution of \\(\\mathbf{X}^c\\). Then each pose hypothesis \\(\\Delta \\mathbf{T}_h = \\{\\Delta \\mathbf{R}_h, \\Delta \\mathbf{t}_h\\}\\) is scored with the reciprocal of distance \\(D_h\\) as in [27, 51]" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.303, + 0.907, + 0.373 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} D _ {h} = \\frac {1}{N ^ {c}} \\sum_ {\\mathbf {p} ^ {c} \\in \\mathbf {P} ^ {c}} \\min _ {\\mathbf {q} ^ {c} \\in \\mathbf {Q} ^ {c}} \\| \\Delta \\mathbf {R} _ {h} ^ {\\top} \\left(\\mathbf {q} ^ {c} - \\Delta \\mathbf {t} _ {h}\\right) - \\mathbf {p} ^ {c} \\| _ {2}, \\tag {7} \\\\ \\xi_ {h} = \\frac {1}{D _ {h}}, \\quad h = 1, 2, 3, \\dots , N _ {H}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.381, + 0.906, + 0.426 + ], + "angle": 0, + "content": "The pose hypothesis with the highest score \\(\\xi_{h}\\) is selected as the initial pose prediction \\(\\Delta \\mathbf{T}_{init} = \\{\\Delta \\mathbf{R}_{init},\\Delta \\mathbf{t}_{init}\\}\\) and further sent to transform the input of the next stage." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.427, + 0.907, + 0.625 + ], + "angle": 0, + "content": "Fine Pose Estimation. After transforming \\(\\mathbf{Q}_{cam}\\) to \\(\\tilde{\\mathbf{Q}}_{cam}\\) with the initial pose prediction \\(\\Delta \\mathbf{T}_{init}\\), we perform a fine matching process between two dense point sets, i.e., \\(\\tilde{\\mathbf{Q}}^f \\in \\mathbb{R}^{N^f \\times 3}\\) and \\(\\mathbf{P}^f \\in \\mathbb{R}^{N^f \\times 3}\\), to achieve a more accurate pose. The network exploits geometric details through a hierarchical encoding paradigm, comprising a positional encoding layer and a local reference frame encoding layer. For each point, we first encode its global position using a mini-PointNet [70], and then construct SE(3)-invariant local reference frame (LRF) to gather local descriptors. These two encodings complement each other, as local descriptors capture fine geometric structures within small neighborhoods, while the positional encoding layer offers global geometric context." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.625, + 0.909, + 0.763 + ], + "angle": 0, + "content": "Leveraging \\(\\tilde{\\mathbf{Q}}^f\\) for example, the process of constructing LRF encoding is as follows. For each point \\(\\tilde{\\mathbf{q}}^m\\) in \\(\\tilde{\\mathbf{Q}}^f\\), we build a local region set \\(\\tilde{\\mathcal{Q}}^m = \\{\\tilde{\\mathbf{q}}_j\\}\\) where \\(\\| \\tilde{\\mathbf{q}}_j - \\tilde{\\mathbf{q}}^m \\|_2 \\leq r\\}_{j=1}^{N_D}\\) by grouping \\(N_D\\) neighboring points of \\(\\tilde{\\mathbf{q}}^m\\). The transformation pose \\(\\{\\mathbf{R}_L^m, \\mathbf{t}_L^m, s_L^m\\}\\) for LRF is computed similarly to GRF (Eq. 3, 4, 5), except that LRF is built upon a local point set while GRF is based on the entire point cloud. By computing the transformation pose, we calculate local point descriptors as" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.77, + 0.907, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {Q} _ {L} ^ {m} = \\tilde {\\mathcal {Q}} _ {L} ^ {m} = \\left\\{\\left(\\mathbf {R} _ {L} ^ {m}\\right) ^ {\\top} \\left(\\tilde {\\mathbf {q}} _ {j} - \\mathbf {t} _ {L} ^ {m}\\right) / s _ {L} ^ {m} \\right\\} _ {j = 1} ^ {N _ {D}}, \\tag {8} \\\\ \\mathbf {Q} _ {L} ^ {f} = \\{\\mathcal {Q} _ {L} ^ {m} \\} _ {m = 1} ^ {N ^ {f}}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.824, + 0.905, + 0.872 + ], + "angle": 0, + "content": "Similarly, we transform \\(\\mathbf{P}^f\\) to \\(\\mathbf{P}_L^f\\). Note that \\(\\mathcal{Q}_L^m = \\tilde{\\mathcal{Q}}_L^m\\) since LRF is pose-invariant. Then, the LRF encoding is extracted from \\(\\mathbf{Q}_L^f\\) and \\(\\mathbf{P}_L^f\\) with a three-layer MLP." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Positional encoding and LRF encoding are combined with the geometric and color features as the input of the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22027" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.169, + 0.089, + 0.828, + 0.215 + ], + "angle": 0, + "content": "
MethodSeg. ModelModalityCAD modelLM-OTUD-LYCB-VAvgTime (s)
ZeroPose [7]SAM [40]RGB✓35.642.153.443.73.90
CNOS [61]FastSAM [107]RGB✓39.748.059.949.20.23
CNOS [61]SAM [40]RGB✓39.639.159.546.11.71
SAM-6D [51]FastSAM [107]RGB-D✓42.251.762.152.01.47
SAM-6D [51]SAM [40]RGB-D✓46.056.960.554.54.53
UNOseg (Ours)FastSAM [107]RGBX37.350.367.351.60.24
UNOseg (Ours)SAM [40]RGBX39.756.266.854.22.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.226, + 0.907, + 0.255 + ], + "angle": 0, + "content": "Table 1. Segmentation results (mAP) on LM-O, TUD-L, and YCB-V. Results of other methods are obtained from the BOP website bop.felk.cvut.cz/leaderboards/segmentation-unseen-bop23. We denote the best score in bold and the second best score with underline." + }, + { + "type": "table", + "bbox": [ + 0.133, + 0.264, + 0.865, + 0.479 + ], + "angle": 0, + "content": "
MethodSeg. ModelModalityRef. TypeLM-OTUD-LYCB-VAvgTime (s)
ZTE-PPF [1]MRCNN [28]DModel66.390.450.269.00.74
Koenig-PPF [41]Retina/MRCNNRGB-DModel63.192.070.175.10.99
Ref. AlignCenterUNOSeg-SAMDImage27.328.847.534.52.69
FPFH+RANSAC [72]UNOSeg-SAMDImage31.031.050.037.36.38
FPFH+MAC [103]UNOSeg-SAMDImage22.522.149.631.4136.94
PPF [22]UNOSeg-SAMDImage29.714.838.327.611.79
PPF_3D_ICP [22]UNOSeg-SAMDImage44.729.166.846.914.27
FCGF+RANSAC [12]UNOSeg-SAMDImage38.959.057.651.810.96
FCGF+MAC [103]UNOSeg-SAMDImage33.948.351.044.460.53
UTOPIC [11]UNOSeg-SAMDImage13.735.410.519.94.00
GeDi [69]UNOSeg-SAMDImage42.867.360.656.948.89
FreeZe [4]UNOSeg-SAMRGB-DImage45.568.365.559.852.96
SAM-6D* [51]UNOSeg-SAMRGB-DPosed Image54.529.768.150.84.21
UNOpose (Ours)UNOSeg-FastSAMRGB-DImage55.866.581.968.11.11
UNOpose (Ours)UNOSeg-SAMRGB-DImage58.771.083.170.93.70
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.489, + 0.907, + 0.532 + ], + "angle": 0, + "content": "Table 2. Pose estimation results on LM-O, TUD-L, and YCB-V. The mean Average Recall (%) of the BOP metric and the average time (s) per image are reported. We highlight the best scores of each setting in bold. Specifically, for adapting SAM-6D [51] to our setting, we randomly choose one of the two templates which [51] uses as a canonical reference during training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.649 + ], + "angle": 0, + "content": "Geometric Transformer. Similar to initial pose prediction, by decoding these features, we obtain the fine point-wise features \\(\\hat{F}_Q^f,\\hat{F}_P^f\\) and overlap confidencees \\(\\hat{O}_Q^f,\\hat{O}_P^f\\) , which are then used to obtain the overlap-aware fine correlation matrix \\(\\mathbf{X}^{f}\\in \\mathbb{R}^{(N^{f} + 1)\\times (N^{f} + 1)}\\) . The final pose \\(\\Delta \\mathbf{T}\\) is predicted by solving Eq. (1) using \\(\\mathbf{X}^f\\) with the weighted SVD algorithm." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.224, + 0.684 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.694, + 0.281, + 0.71 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. Implemented with PyTorch [67], we use the DINOv2 [13, 65] pre-trained ViT-base [20] to encode color information. GeoTransformers [71] with attention [82] and sparse-to-dense linear attention [51] are respectively employed to extract coarse and fine geometric correspondences. The network is trained with standard MegaPose synthetic dataset [43] for BOP unseen object pose estimation track [34]. It contains \\(\\sim 2\\mathrm{M}\\) images of \\(\\sim 50\\mathrm{K}\\) objects from GSO [21] and ShapeNet [5]. We train 3 epochs with a batch size of 32 on 4 RTX 3090 GPUs. The network is trained with Adam [39] and cosine annealing [58], using a base learning rate of \\(10^{-4}\\). Moreover, we lever" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.555, + 0.907, + 0.587 + ], + "angle": 0, + "content": "age InfoNCE [64] loss for learning correlation matrix and weighted BCE loss for supervising overlap prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.592, + 0.909, + 0.819 + ], + "angle": 0, + "content": "Datasets and Benchmarks. We evaluate our method on BOP test split [33] of three different real-world datasets: LM-O [2], TUD-L [32], and YCB-V [92]. LM-O consists of 200 images from one clutter scene, where eight objects with occlusion are provided for testing. TUD-L comprises three moving objects with mild occlusion and diverse lighting conditions. YCB-V is very challenging due to severe occlusions and sensor noise. It comprises 900 testing images from 12 scenes consisting of 21 objects. For each query object in the test split, we randomly select a reference object beyond the test set (in the training or evaluation set). Unless specified, we constrain the maximal rotation difference between reference and target to \\(50^{\\circ}\\) to ensure the ground-truth correspondence set is not empty. Moreover, the occlusion ratio of reference objects is required to be less than \\(30\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Evaluation Metrics. For evaluating instance segmentation, we follow [33] to use the mean Average Precision (mAP) metric, which is computed across different Intersection-over-Union (IoU) thresholds (\\{0.5, 0.55, ..., 0.95\\}). Only annotated instances with at least \\(10\\%\\) projected visible surfaces" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22028" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.096, + 0.481, + 0.19 + ], + "angle": 0, + "content": "
CategoryVSDMSSDMSPDARBOP
Dragon47.173.269.463.2
Frog59.452.657.656.4
Can56.865.665.262.5
Avg54.463.764.160.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.284, + 0.193, + 0.302, + 0.204 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.09, + 0.89, + 0.19 + ], + "angle": 0, + "content": "
Ref.
Ref.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.697, + 0.193, + 0.715, + 0.204 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.217, + 0.894, + 0.232 + ], + "angle": 0, + "content": "Table 3. One reference per category experiments on TUD-L. Blue and green contours denote GT and estimated poses, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.251, + 0.484, + 0.342 + ], + "angle": 0, + "content": "are considered for evaluation. For object pose evaluation, we use the symmetry-aware BOP metric [33]. It is calculated as the mean of the Average Recall of Visible Surface Discrepancy (VSD), Maximum Symmetry-Aware Surface Distance (MSSD), and Maximum Symmetry-Aware Projection Distance (MSPD) (See [33] for details)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.349, + 0.484, + 0.379 + ], + "angle": 0, + "content": "4.2. UNO Object Segmentation and Pose Estimation Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.388, + 0.484, + 0.553 + ], + "angle": 0, + "content": "UNO Object Segmentation Results. We compare our UNOSeg with some existing CAD-model-based unseen object segmentation methods [7, 51, 61] in Tab. 1. Leveraging a single image as a reference, our UNOSeg achieves comparable results with state-of-the-art methods (Ours \\(54.2\\%\\) vs. SAM-6D \\(54.5\\%\\) ) at a faster speed (Ours 2.68s vs. SAM-6D 4.53s). Notably, UNOSeg-FastSAM achieves the best performance on YCB-V (mAP \\(67.3\\%\\)). Compared to FastSAM [107], SAM [40] achieves consistently better results on three datasets. Therefore, we choose UNOSeg-SAM as the default segmentation for subsequent experiments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.486, + 0.902 + ], + "angle": 0, + "content": "UNO Object Pose Estimation Results. Since we are the first to conduct relative 6DoF pose estimation for unseen objects and build a brand-new benchmark, we re-implement some traditional [22, 72, 103] and learning-based methods [4, 11, 12, 51, 69] for comparison. Note that we leverage identical data pre-processing, segmentation, query and reference image pairs, and evaluation protocols for all methods for fair comparison. Moreover, we also compare UNOPose with two CAD-model-based pose estimators leveraging \\(SE(3)\\)-invariant PPF feature [1, 41]. All results are illustrated in Tab. 2. It clearly shows that our method surpasses all image-reference-based methods by a large margin, and even achieves comparable results with CAD-model-based methods using class-specific detectors [28, 52] (Ours \\(70.9\\%\\) vs. ZTE-PPF [1] \\(69.0\\%\\) vs. Koenig-PPF [41] \\(75.1\\%\\)). \"Ref. AlignCenter\" directly uses the reference rotation as prediction, meanwhile leveraging the shift between query and reference centers as relative translation, so it can be regarded as the baseline of all methods. Among all traditional descriptors, PPF with Iterative Closest Point (ICP) refinement [22] achieves the best results (AR \\(46.9\\%\\)). Learning-based methods demonstrate varying generalization abilities on unseen objects. For example, the GeoTransformer-based" + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.248, + 0.899, + 0.371 + ], + "angle": 0, + "content": "
ROWGRFLRFFPEVSDMSSDMSPD\\( \\mathrm{AR}_{\\text{BOP}} \\)
A0✘✘✘64.668.159.263.9
A1✓✘✘66.472.863.267.5
A2✘✓✘66.269.560.465.4
A3✓✓✘66.973.263.567.9
B0✘✘✓79.885.075.580.1
B1✓✘✓79.485.276.680.4
B2✘✓✓81.185.176.680.9
B3✓✓✓80.786.077.781.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.381, + 0.907, + 0.409 + ], + "angle": 0, + "content": "Table 4. Ablation of GRF and LRF on YCB-V. FPE stands for \"fine pose estimation\"." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.437, + 0.907, + 0.513 + ], + "angle": 0, + "content": "method UTOPIC, designed for registering object-level partial point clouds, performs suboptimally in our setting. In contrast, the adapted unseen object pose estimation methods, i.e. SAM-6D [51] and FreeZe [4], achieve satisfactory results compared to their counterparts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.907, + 0.559 + ], + "angle": 0, + "content": "Run-time Analysis. Evaluated on a single RTX 3090 GPU, our pipeline runs at 3.70s for one image, including 2.68s for segmentation and 1.02s for pose estimation (Tab. 1 - 2)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.57, + 0.779, + 0.587 + ], + "angle": 0, + "content": "4.3. One Reference for a Category" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.909, + 0.806 + ], + "angle": 0, + "content": "To minimize the efforts of acquiring references, we leverage a single reference for all objects of the same category in a dataset. Tab. 3 illustrates the qualitative and quantitative experimental results on TUD-L. It is worth noting that the target object in TUD-L is dynamic rather than static, preventing methods from leveraging background information. This setup poses challenges for relative pose estimation methods that rely on scene features, making them likely to fail in this task. Despite this, UNOPose achieves impressive performance (60.7% in terms of the \\(\\mathrm{AR}_{\\mathrm{BOP}}\\) metric) with only a single reference per category. The visualization results in Tab. 3b further show UNOPose is capable of predicting omnidirectional relative poses. Additional experiments on reference selection are provided in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.817, + 0.68, + 0.831 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of GRF and LRF. As demonstrated in Sec. 3.3.2, GRF transfers arbitrary object poses to an \\(SE(3)\\)-invariant frame, while LRF assists with extracting precise local descriptors for fine pose estimation. Tab. 4 shows that" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22029" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.48, + 0.213 + ], + "angle": 0, + "content": "
ROWMethodVSDMSSDMSPD\\( \\mathrm{AR}_{\\text{BOP}} \\)
C0B3 + Soft Correspondence [95]77.181.772.777.2
C1B3 + Overlap Predictor82.687.479.483.1
C2A3 + Overlap Predictor68.474.664.869.2
D0C1: Seg. → CNOS [61]77.480.973.677.3
D1C1: Seg. → SAM-6D [51]78.382.174.378.2
D2C1: Seg. → MRCNN [28, 42]81.285.777.981.6
D3C1: Seg. → GT Seg.86.689.282.186.0
E0C1: Ref. Real → Rendered83.789.683.185.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.224, + 0.485, + 0.252 + ], + "angle": 0, + "content": "Table 5. Ablation of overlap predictor, segmentation, and reference type on YCB-V." + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.267, + 0.465, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.48, + 0.483, + 0.539 + ], + "angle": 0, + "content": "Figure 3. Ablation of initial rotation distance between query and reference objects. We categorize all testing objects into nine groups according to the initial rotation distance, and evaluate the \\(\\mathrm{AR}_{\\mathrm{BOP}}\\) metric and overlap ratio for each group separately." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.483, + 0.702 + ], + "angle": 0, + "content": "the effectiveness of GRF and LRF is complementary in UNOpose. Specifically, GRF brings significant performance enhancement in the coarse pose estimation (Tab. 4 A1 vs. A0, A3 vs. A2), providing an accurate pose initialization for the next stage. Meanwhile, LRF is constructed to exploit fine-grained local information, primarily enhancing the fine pose estimation results (Tab. 4 B0 vs. B2, B1 vs. B3). With both GRF and LRF, our network achieves obvious performance improvement (Tab. 4 A3 vs. A0, B3 vs. B0)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Ablation on Correspondence Loss. In partial-to-partial point cloud registration, only some of the points can find their counterparts on the target point cloud. Therefore, the introduction of overlap prediction helps to distinguish the points in the overlap region and achieves good results (Tab. 4-5 C1 vs. B3, C2 vs. A3). Moreover, we consider soft correspondence loss in [95] as an alternative, which predicts probabilities of overlap between each point pair. However, the result did not meet the expectations (Tab. 5 C0)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Ablation on Segmentation. Tab. 1 showcases the advantage of our UNOSeg over other segmentation methods on YCB-V. Furthermore, substituting UNOSeg with CNOS [61], SAM-6D [51], or MRCNN [28, 42] results in a noticeable" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.137 + ], + "angle": 0, + "content": "decline in UNOpose's performance (Tab. 5 D0-D2). Moreover, substituting UNOSeg with ground-truth segmentation brings \\(\\approx 3\\%\\) performance enhancement (Tab. 5 D3 vs. C1)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.909, + 0.243 + ], + "angle": 0, + "content": "Using Rendered References. We replaced the real reference with a pose-identical rendered version and present the results in Tab. 5. With reduced depth noise and domain gap between training and testing, UNOPose works even better using the rendered reference (Tab. 5 E0 vs. C1). However, this ablation requires the object model, which is more difficult to acquire than an unposed real image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.243, + 0.909, + 0.47 + ], + "angle": 0, + "content": "Ablation on Reference Rotation Distance. To investigate the impact of the initial rotation distance between the query and reference images, we randomly select the reference object with a maximum rotation distance of \\(90^{\\circ}\\) for each query object. Further, all testing objects are categorized into nine groups according to initial rotation distances. We separately evaluate each group and present the results in Fig. 3. Meanwhile, we calculate the overlap ratio between reference and query instances of each group and visualize the mean and standard deviation. Notably, performance significantly decreases when the rotation distance exceeds \\(50^{\\circ}\\). This decline is attributed to the reduced overlap between query and reference viewpoints. However, our network still shows favorable transfer ability in extreme relative pose estimation (rotation distance \\((80^{\\circ}, 90^{\\circ}]\\), overlap ratio \\(45.4\\%\\), \\(\\mathrm{AR}_{\\mathrm{BOP}} 54.8\\%\\))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.908, + 0.515 + ], + "angle": 0, + "content": "The appendix presents further implementation details, visualization results, and more experimental analysis on reference selection and extensive datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.53, + 0.878, + 0.547 + ], + "angle": 0, + "content": "5. Conclusion, Limitation and Future work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.555, + 0.909, + 0.72 + ], + "angle": 0, + "content": "This work has introduced UNOPose, a new approach for unseen object pose estimation with one reference. The key idea is constructing the \\(SE(3)\\)-invariant reference frame for tackling diverse pose and size variations. Moreover, we propose an overlap predictor for handling low-overlap scenarios. To evaluate the proposed method, we devise a new benchmark based on the BOP challenge and compare some state-of-the-art methods. Experimental results show UNOPose surpasses all compared reference-based methods significantly, and is competitive with some \\(SE(3)\\)-invariant-feature-based methods relying on CAD models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.722, + 0.909, + 0.828 + ], + "angle": 0, + "content": "Limitation and Future Work. While efficient, correspondence-based methods rely on predicting robust and dense correspondence between query and reference objects, thereby limited in extreme two-view geometry applications. Future work will focus on reconstructing the unseen object from a single reference [53, 108] and concurrently estimating its object pose." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.831, + 0.909, + 0.887 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported in part by the National Natural Science Foundation of China under Grant No. 62406169, and in part by the China Postdoctoral Science Foundation under Grant No. 2024M761673." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22030" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.116, + 0.484, + 0.143 + ], + "angle": 0, + "content": "[1] Anonymous. Zte-ppf. https://bop.felk(cvut.cz/method_info/250/, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.145, + 0.483, + 0.2 + ], + "angle": 0, + "content": "[2] Eric Brachmann, Alexander Krull, Frank Michel, Stefan Gumhold, Jamie Shotton, and Carsten Rother. Learning 6D object pose estimation using 3D object coordinates. In ECCV, 2014. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.202, + 0.484, + 0.243 + ], + "angle": 0, + "content": "[3] Dingding Cai, Janne Heikkilä, and Esa Rahtu. Ove6d: Object viewpoint encoding for depth-based 6d object pose estimation. In CVPR, pages 6803-6813, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.246, + 0.484, + 0.3 + ], + "angle": 0, + "content": "[4] Andrea Caraffa, Davide Boscaini, Amir Hamza, and Fabio Poiesi. Freeze: Training-free zero-shot 6d pose estimation with geometric and vision foundation models. In ECCV, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.303, + 0.484, + 0.371 + ], + "angle": 0, + "content": "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.374, + 0.484, + 0.428 + ], + "angle": 0, + "content": "[6] Dengsheng Chen, Jun Li, Zheng Wang, and Kai Xu. Learning canonical shape space for category-level 6D object pose and size estimation. In CVPR, pages 11970-11979, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.431, + 0.484, + 0.486 + ], + "angle": 0, + "content": "[7] Jianqiu Chen, Zikun Zhou, Mingshan Sun, Rui Zhao, Liwei Wu, Tianpeng Bao, and Zhenyu He. Zeropose: Cadprompted zero-shot object 6d pose estimation in cluttered scenes. IEEE TCSVT, 2024. 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.489, + 0.484, + 0.53 + ], + "angle": 0, + "content": "[8] Kai Chen and Qi Dou. SGPA: Structure-guided prior adaptation for category-level 6D object pose estimation. In ICCV, pages 2773-2782, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.532, + 0.484, + 0.6 + ], + "angle": 0, + "content": "[9] Wei Chen, Xi Jia, Hyung Jin Chang, Jinming Duan, Shen Linlin, and Ales Leonardis. FS-Net: Fast shape-based network for category-level 6D object pose estimation with decoupled rotation mechanism. In CVPR, pages 1581–1590, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.603, + 0.484, + 0.698 + ], + "angle": 0, + "content": "[10] Yamei Chen, Yan Di, Guangyao Zhai, Fabian Manhardt, Chenyangguang Zhang, Ruida Zhang, Federico Tombari, Nassir Navab, and Benjamin Busam. Secondpose: Se (3)-consistent dual-stream feature fusion for category-level pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9959-9969, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.702, + 0.484, + 0.77 + ], + "angle": 0, + "content": "[11] Zhilei Chen, Honghua Chen, Lina Gong, Xuefeng Yan, Jun Wang, Yanwen Guo, Jing Qin, and Mingqiang Wei. Utopic: Uncertainty-aware overlap prediction network for partial point cloud registration. In Computer Graphics Forum, pages 87-98. Wiley Online Library, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.773, + 0.484, + 0.814 + ], + "angle": 0, + "content": "[12] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In ICCV, pages 8958-8966, 2019. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.816, + 0.484, + 0.856 + ], + "angle": 0, + "content": "[13] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In ICLR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[14] Haowen Deng, Tolga Birdal, and Slobodan Ilic. Ppf-foldnet: Unsupervised learning of rotation invariant 3d local descriptors. In ECCV, pages 602-618, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.116, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "[15] Haowen Deng, Tolga Birdal, and Slobodan Ilic. Ppfnet: Global context aware local features for robust 3d point matching. In CVPR, pages 195-205, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.136, + 0.908, + 0.204 + ], + "angle": 0, + "content": "[16] Maximilian Denninger, Dominik Winkelbauer, Martin Sundermeyer, Wout Boerdijk, Markus Knauer, Klaus H. Strobl, Matthias Humt, and Rudolph Triebel. Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software, 8(82):4901, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.205, + 0.907, + 0.287 + ], + "angle": 0, + "content": "[17] Yan Di, Ruida Zhang, Zhiqiang Lou, Fabian Manhardt, Xi-angyang Ji, Nassir Navab, and Federico Tombari. Gpv-posed Category-level object pose estimation via geometry-guided point-wise voting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6781–6791, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.288, + 0.907, + 0.37 + ], + "angle": 0, + "content": "[18] Yan Di, Chenyangguang Zhang, Ruida Zhang, Fabian Manhardt, Yongzhi Su, Jason Rambach, Didier Stricker, Xi-angyang Ji, and Federico Tombari. U-red: Unsupervised 3d shape retrieval and deformation for partial point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8884–8895, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.372, + 0.907, + 0.467 + ], + "angle": 0, + "content": "[19] Yan Di, Chenyangguang Zhang, Chaowei Wang, Ruida Zhang, Guangyao Zhai, Yanyan Li, Bowen Fu, Xiangyang Ji, and Shan Gao. Shapematcher: Self-supervised joint shape canonicalization segmentation retrieval and deformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21017-21028, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.469, + 0.907, + 0.538 + ], + "angle": 0, + "content": "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.539, + 0.907, + 0.607 + ], + "angle": 0, + "content": "[21] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In ICRA, pages 2553-2560. IEEE, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.609, + 0.907, + 0.662 + ], + "angle": 0, + "content": "[22] Bertram Drost, Markus Ulrich, Nassir Navab, and Slobodan Ilic. Model globally, match locally: Efficient and robust 3d object recognition. In CVPR, pages 998-1005. IEEE, 2010. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.665, + 0.907, + 0.732 + ], + "angle": 0, + "content": "[23] Guoguang Du, Kai Wang, Shiguo Lian, and Kaiyong Zhao. Vision-based robotic grasping from object localization, object pose estimation to grasp estimation for parallel grippers: a review. Artificial Intelligence Review, 54(3):1677-1734, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.734, + 0.907, + 0.801 + ], + "angle": 0, + "content": "[24] Zhaoxin Fan, Zhenbo Song, Jian Xu, Zhicheng Wang, Kejian Wu, Hongyan Liu, and Jun He. Object level depth reconstruction for category level 6d object pose estimation from monocular rgb image. In ECCV, pages 220-236. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.804, + 0.907, + 0.859 + ], + "angle": 0, + "content": "[25] Andrea Frome, Daniel Huber, Ravi Kolluri, Thomas Bulow, and Jitendra Malik. Recognizing objects in range data using regional point descriptors. In ECCV, pages 224-237. Springer, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.86, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[26] Bowen Fu, Sek Kun Leong, Yan Di, Gu Wang, Jiwen Tang, and Xiangyang Ji. Lanpose: Language-instructed 6d object pose estimation for robotic assembly. In ECCVW, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22031" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.148 + ], + "angle": 0, + "content": "[27] Rasmus Laurvig Haugaard and Anders Glent Buch. Surfemb: Dense and continuous correspondence distributions for object pose estimation with learnt surface embeddings. In CVPR, pages 6749-6758, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.151, + 0.485, + 0.191 + ], + "angle": 0, + "content": "[28] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask R-CNN. In ICCV, pages 2961-2969, 2017. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.194, + 0.484, + 0.248 + ], + "angle": 0, + "content": "[29] Xingyi He, Jiaming Sun, Yuang Wang, Di Huang, Hujun Bao, and Xiaowei Zhou. Onepose++: Keypoint-free one-shot object pose estimation without cad models. NeurIPS, 35:35103-35115, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.252, + 0.484, + 0.294 + ], + "angle": 0, + "content": "[30] Yisheng He, Haibin Huang, Haoqiang Fan, Qifeng Chen, and Jian Sun. Ffb6d: A full flow bidirectional fusion network for 6d pose estimation. In CVPR, pages 3003-3013, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.297, + 0.484, + 0.337 + ], + "angle": 0, + "content": "[31] Yisheng He, Yao Wang, Haoqiang Fan, Jian Sun, and Qifeng Chen. Fs6d: Few-shot 6d pose estimation of novel objects. In CVPR, pages 6814-6824, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.34, + 0.484, + 0.408 + ], + "angle": 0, + "content": "[32] Tomas Hodan, Frank Michel, Eric Brachmann, Wadim Kehl, Anders GlentBuch, Dirk Kraft, Bertram Drost, Joel Vidal, Stephan Ihrke, Xenophon Zabulis, et al. BOP: Benchmark for 6D Object Pose Estimation. In ECCV, pages 19-34, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.412, + 0.484, + 0.466 + ], + "angle": 0, + "content": "[33] Tomas Hodan, Martin Sundermeyer, Bertram Drost, Yann Labbe, Eric Brachmann, Frank Michel, Carsten Rother, and Jiri Matas. BOP Challenge 2020 on 6D object localization. In ECCVW, pages 577-594, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.47, + 0.484, + 0.55 + ], + "angle": 0, + "content": "[34] Tomas Hodan, Martin Sundermeyer, Yann Labbe, Van Nguyen Nguyen, Gu Wang, Eric Brachmann, Bertram Drost, Vincent Lepetit, Carsten Rother, and Jiri Matas. Bop challenge 2023 on detection, segmentation and pose estimation of seen and unseen rigid objects. In CVPRW, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.555, + 0.484, + 0.595 + ], + "angle": 0, + "content": "[35] Yinlin Hu, Pascal Fua, Wei Wang, and Mathieu Salzmann. Single-stage 6d object pose estimation. In CVPR, pages 2930-2939, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.599, + 0.484, + 0.652 + ], + "angle": 0, + "content": "[36] Junwen Huang, Hao Yu, Kuan-Ting Yu, Nassir Navab, Slobodan Ilic, and Benjamin Busam. Matchu: Matching unseen objects for 6d pose estimation from rgb-d images. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.656, + 0.484, + 0.711 + ], + "angle": 0, + "content": "[37] Siyuan Huang, Siyuan Qi, Yinxue Xiao, Yixin Zhu, Ying Nian Wu, and Song-Chun Zhu. Cooperative holistic scene understanding: Unifying 3d object, layout, and camera pose estimation. NeurIPS, 31, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.715, + 0.484, + 0.768 + ], + "angle": 0, + "content": "[38] Shengyu Huang, Zan Gojcic, Mikhail Usvyatsov, Andreas Wieser, and Konrad Schindler. Predator: Registration of 3d point clouds with low overlap. In CVPR, pages 4267-4276, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.772, + 0.484, + 0.799 + ], + "angle": 0, + "content": "[39] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.802, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[40] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.86, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[41] Rebecca König and Bertram Drost. A hybrid approach for 6 dof pose estimation. In ECCV, pages 700-706. Springer, 2020. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.907, + 0.135 + ], + "angle": 0, + "content": "[42] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. CosyPose: Consistent multi-view multi-object 6D pose estimation. In ECCV, pages 574-591, 2020. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.137, + 0.907, + 0.205 + ], + "angle": 0, + "content": "[43] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. Megapose: 6d pose estimation of novel objects via render & compare. In CoRL, pages 715-725. PMLR, 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.208, + 0.907, + 0.248 + ], + "angle": 0, + "content": "[44] Junha Lee, Seungwook Kim, Minsu Cho, and Jaesik Park. Deep hough voting for robust global registration. In ICCV, pages 15994-16003, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.251, + 0.907, + 0.318 + ], + "angle": 0, + "content": "[45] Jiahao Li, Changhao Zhang, Ziyao Xu, Hangning Zhou, and Chi Zhang. Iterative distance-aware similarity matrix convolution with mutual-supervised point elimination for efficient point cloud registration. In ECCV, pages 378-394. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.321, + 0.907, + 0.375 + ], + "angle": 0, + "content": "[46] Xiaolong Li, Yijia Weng, Li Yi, Leonidas J Guibas, A Abbott, Shuran Song, and He Wang. Leveraging se (3) equivariance for self-supervised category-level object pose estimation from point clouds. NeurIPS, 34, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.378, + 0.907, + 0.417 + ], + "angle": 0, + "content": "[47] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. DeepIM: Deep iterative matching for 6D pose estimation. IJCV, 128(3):657-678, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.42, + 0.907, + 0.474 + ], + "angle": 0, + "content": "[48] Zhigang Li, Gu Wang, and Xiangyang Ji. CDPN: Coordinates-Based Disentangled Pose Network for Real-Time RGB-Based 6-DoF Object Pose Estimation. In ICCV, pages 7678-7687, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.477, + 0.906, + 0.517 + ], + "angle": 0, + "content": "[49] Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. In 3DV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.519, + 0.907, + 0.587 + ], + "angle": 0, + "content": "[50] Jiehong Lin, Zewei Wei, Zhihao Li, Songcen Xu, Kui Jia, and Yuanqing Li. DualPoseNet: Category-level 6D object pose and size estimation using dual pose network with refined learning of pose consistency. In ICCV, pages 3560-3569, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.591, + 0.907, + 0.631 + ], + "angle": 0, + "content": "[51] Jiehong Lin, Lihua Liu, Dekun Lu, and Kui Jia. Sam-6d: Segment anything model meets zero-shot 6d object pose estimation. In CVPR, 2024. 1, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.633, + 0.907, + 0.673 + ], + "angle": 0, + "content": "[52] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.676, + 0.907, + 0.73 + ], + "angle": 0, + "content": "[53] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In ICCV, pages 9298–9309, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.733, + 0.907, + 0.773 + ], + "angle": 0, + "content": "[54] Xingyu Liu, Gu Wang, Yi Li, and Xiangyang Ji. CATRE: iterative point clouds alignment for category-level object pose refinement. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.776, + 0.907, + 0.844 + ], + "angle": 0, + "content": "[55] Xingyu Liu, Yingyue Li, Chengxi Li, Gu Wang, Chenyangguang Zhang, Ziqin Huang, and Xiangyang Ji. Gfreedet: Exploiting gaussian splatting and foundation models for model-free unseen object detection in the bop challenge 2024. arXiv preprint arXiv:2412.01552, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.847, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[56] Xingyu Liu, Chenyangguang Zhang, Gu Wang, Ruida Zhang, and Xiangyang Ji. Rasim: A range-aware high-fidelity rgb-d data simulation pipeline for real-world applications. In ICRA, pages 17057-17064. IEEE, 2024. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22032" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[57] Yuan Liu, Yilin Wen, Sida Peng, Cheng Lin, Xiaoxiao Long, Taku Komura, and Wenping Wang. Gen6d: Generalizable model-free 6-dof object pose estimation from rgb images. In ECCV, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.151, + 0.485, + 0.178 + ], + "angle": 0, + "content": "[58] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In *ICLR*, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.181, + 0.484, + 0.221 + ], + "angle": 0, + "content": "[59] Eric Marchand, Hideaki Uchiyama, and Fabien Spindler. Pose estimation for augmented reality: a hands-on survey. IEEE TVCG, 22(12):2633-2651, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.225, + 0.484, + 0.293 + ], + "angle": 0, + "content": "[60] Van Nguyen Nguyen, Yinlin Hu, Yang Xiao, Mathieu Salzmann, and Vincent Lepetit. Templates for 3d object pose estimation revisited: Generalization to new objects and robustness to occlusions. In CVPR, pages 6771-6780, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.297, + 0.483, + 0.352 + ], + "angle": 0, + "content": "[61] Van Nguyen Nguyen, Thibault Groueix, Georgy Ponimatkin, Vincent Lepetit, and Tomas Hodan. Cnos: A strong baseline for cad-based novel object segmentation. In ICCV, pages 2134-2140, 2023. 1, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.355, + 0.484, + 0.408 + ], + "angle": 0, + "content": "[62] Van Nguyen Nguyen, Thibault Groueix, Mathieu Salzmann, and Vincent Lepetit. Gigapose: Fast and robust novel object pose estimation via one correspondence. In CVPR, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.413, + 0.484, + 0.468 + ], + "angle": 0, + "content": "[63] Yinyu Nie, Xiaoguang Han, Shihui Guo, Yujuan Zheng, Jian Chang, and Jian Jun Zhang. Total3DUnderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In CVPR, pages 55-64, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.471, + 0.484, + 0.512 + ], + "angle": 0, + "content": "[64] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.515, + 0.484, + 0.639 + ], + "angle": 0, + "content": "[65] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. TMLR, 2024. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.642, + 0.484, + 0.696 + ], + "angle": 0, + "content": "[66] Evin Pinar Örnek, Yann Labbé, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, and Tomas Hodan. Found- pose: Unseen object pose estimation with foundation features. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.7, + 0.484, + 0.768 + ], + "angle": 0, + "content": "[67] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. PyTorch: An imperative style, high-performance deep learning library. In NeurlPS, pages 8026-8037, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.772, + 0.482, + 0.813 + ], + "angle": 0, + "content": "[68] Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, and Hujun Bao. Pynet: Pixel-wise voting network for 6dof pose estimation. In CVPR, pages 4561-4570, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.816, + 0.484, + 0.856 + ], + "angle": 0, + "content": "[69] Fabio Poiesi and Davide Boscaini. Learning general and distinctive 3d local deep descriptors for point cloud registration. IEEE TPAMI, 45(3):3979-3985, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[70] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. PointNet: Deep learning on point sets for 3D classification and segmentation. CVPR, 1(2):4, 2017. 5" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.093, + 0.908, + 0.147 + ], + "angle": 0, + "content": "[71] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, Slobodan Ilic, Dewen Hu, and Kai Xu. Geotransformer: Fast and robust point cloud registration with geometric transformer. IEEE TPAMI, 2023. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.149, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[72] Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (fpfh) for 3d registration. In ICRA, pages 3212-3217. IEEE, 2009. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.192, + 0.908, + 0.231 + ], + "angle": 0, + "content": "[73] Ivan Shugurov, Sergey Zakharov, and Slobodan Ilic. Dpodv2: Dense correspondence-based 6 dof pose estimation. IEEE TPAMI, 44(11):7417-7435, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.233, + 0.908, + 0.274 + ], + "angle": 0, + "content": "[74] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. Osop: A multi-stage one shot object pose estimation framework. In CVPR, pages 6835-6844, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.275, + 0.908, + 0.314 + ], + "angle": 0, + "content": "[75] Fridtjof Stein, Gérard Medioni, et al. Structural indexing: Efficient 3-d object recognition. IEEE TPAMI, 14(2):125-145, 1992. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.317, + 0.907, + 0.385 + ], + "angle": 0, + "content": "[76] Yongzhi Su, Jason Rambach, Nareg Minaskan, Paul Lesur, Alain Pagani, and Didier Stricker. Deep multi-state object pose estimation for augmented reality assembly. In IEEE International Symposium on Mixed and Augmented Reality Adjunct (ISMAR-Adjunct), pages 222–227, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.387, + 0.908, + 0.44 + ], + "angle": 0, + "content": "[77] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. Onepose: One-shot object pose estimation without cad models. In CVPR, pages 6825-6834, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.443, + 0.908, + 0.496 + ], + "angle": 0, + "content": "[78] Jingyi Tang, Gu Wang, Zeyu Chen, Shengquan Li, Xiu Li, and Xiangyang Ji. Fafa: Frequency-aware flow-aided self-supervision for underwater object pose estimation. In ECCV, pages 352–369. Springer, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.498, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[79] Meng Tian, Marcelo H Ang Jr, and Gim Hee Lee. Shape prior deformation for categorical 6D object pose and size estimation. In ECCV, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.54, + 0.907, + 0.594 + ], + "angle": 0, + "content": "[80] Jonathan Tremblay, Thang To, Balakumar Sundaralingam, Yu Xiang, Dieter Fox, and Stan Birchfield. Deep object pose estimation for semantic robotic grasping of household objects. In CoRL, pages 306-316, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.596, + 0.908, + 0.676 + ], + "angle": 0, + "content": "[81] Mikaela Angelina Uy, Vladimir G Kim, Minhyuk Sung, Noam Aigerman, Siddhartha Chaudhuri, and Leonidas J Guibas. Joint learning of 3d shape retrieval and deformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11713-11722, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.679, + 0.908, + 0.733 + ], + "angle": 0, + "content": "[82] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.735, + 0.907, + 0.789 + ], + "angle": 0, + "content": "[83] Chen Wang, Danfei Xu, Yuke Zhu, Roberto Martin-Martin, Cewu Lu, Li Fei-Fei, and Silvio Savarese. DenseFusion: 6D object pose estimation by iterative dense fusion. In CVPR, pages 3343–3352, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.791, + 0.908, + 0.845 + ], + "angle": 0, + "content": "[84] Gu Wang, Fabian Manhardt, Jianzhun Shao, Xiangyang Ji, Nassir Navab, and Federico Tombari. Self6d: Self-supervised monocular 6d object pose estimation. In ECCV, pages 108-125. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.847, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[85] Gu Wang, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. GDR-Net: Geometry-guided direct regression network for monocular 6D object pose estimation. In CVPR, pages 16611–16621, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22033" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.146 + ], + "angle": 0, + "content": "[86] Gu Wang, Fabian Manhardt, Xingyu Liu, Xiangyang Ji, and Federico Tombari. Occlusion-aware self-supervised monocular 6D object pose estimation. IEEE TPAMI, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.15, + 0.483, + 0.207 + ], + "angle": 0, + "content": "[87] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6D object pose and size estimation. In CVPR, pages 2642-2651, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.209, + 0.483, + 0.248 + ], + "angle": 0, + "content": "[88] Jiaze Wang, Kai Chen, and Qi Dou. Category-level 6D object pose estimation via cascaded relation and recurrent reconstruction networks. In IROS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.252, + 0.485, + 0.294 + ], + "angle": 0, + "content": "[89] Yue Wang and Justin M Solomon. Deep closest point: Learning representations for point cloud registration. In ICCV, pages 3523-3532, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.297, + 0.484, + 0.337 + ], + "angle": 0, + "content": "[90] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. FoundationPose: Unified 6d pose estimation and tracking of novel objects. In CVPR, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.34, + 0.484, + 0.395 + ], + "angle": 0, + "content": "[91] Chin-Hsuan Wu, Yen-Chun Chen, Bolivar Solarte, Lu Yuan, and Min Sun. ifusion: Inverting diffusion for pose-free reconstruction from sparse views. arXiv preprint arXiv:2312.17250, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.398, + 0.484, + 0.451 + ], + "angle": 0, + "content": "[92] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6D object pose estimation in cluttered scenes. RSS, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.455, + 0.484, + 0.496 + ], + "angle": 0, + "content": "[93] Jiaqi Yang, Qian Zhang, Yang Xiao, and Zhiguo Cao. Toldi: An effective and robust approach for 3d local shape description. PR, 65:175-187, 2017. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.499, + 0.484, + 0.541 + ], + "angle": 0, + "content": "[94] Yang You, Ruoxi Shi, Weiming Wang, and Cewu Lu. CPPf: Towards robust category-level 9d pose estimation in the wild. In CVPR, pages 6866-6875, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.543, + 0.484, + 0.584 + ], + "angle": 0, + "content": "[95] Hao Yu, Fu Li, Mahdi Saleh, Benjamin Busam, and Slobodan Ilic. Cofinet: Reliable coarse-to-fine correspondences for robust pointcloud registration. NeurIPS, 34, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.587, + 0.484, + 0.641 + ], + "angle": 0, + "content": "[96] Hao Yu, Zheng Qin, Ji Hou, Mahdi Saleh, Dongsheng Li, Benjamin Busam, and Slobodan Ilic. Rotation-invariant transformer for point cloud matching. In CVPR, pages 5384-5393, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.645, + 0.484, + 0.7 + ], + "angle": 0, + "content": "[97] Hao Yu, Ji Hou, Zheng Qin, Mahdi Saleh, Ivan Shugurov, Kai Wang, Benjamin Busam, and Slobodan Ilic. Riga: Rotation-invariant and globally-aware descriptors for point cloud registration. IEEE TPAMI, 46(5):3796-3812, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.703, + 0.484, + 0.757 + ], + "angle": 0, + "content": "[98] Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose: Predicting probabilistic relative rotation for single objects in the wild. In ECCV, pages 592-611. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.76, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[99] Ruida Zhang, Yan Di, Zhiqiang Lou, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. Rbp-posed: Residual bounding box projection for category-level pose estimation. In European conference on computer vision, pages 655–672. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[100] Ruida Zhang, Yan Di, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. Ssp-pose: Symmetry-aware shape prior deformation for direct category-level object pose estimation. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7452–7459. IEEE, 2022." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.092, + 0.486, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[101] Ruida Zhang, Ziqin Huang, Gu Wang, Chenyangguang Zhang, Yan Di, Xingxing Zuo, Jiwen Tang, and Xiangyang Ji. Lapose: Laplacian mixture shape modeling for rgb-based category-level object pose estimation. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.907, + 0.232 + ], + "angle": 0, + "content": "[102] Ruida Zhang, Chenyangguang Zhang, Yan Di, Fabian Manhardt, Xingyu Liu, Federico Tombari, and Xiangyang Ji. Kp-red: Exploiting semantic keypoints for joint 3d shape retrieval and deformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20540-20550, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.909, + 0.274 + ], + "angle": 0, + "content": "[103] Xiyu Zhang, Jiaqi Yang, Shikun Zhang, and Yanning Zhang. 3d registration with maximal cliques. In CVPR, pages 17745–17754, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.906, + 0.318 + ], + "angle": 0, + "content": "[104] Chen Zhao, Yinlin Hu, and Mathieu Salzmann. Fusing local similarities for retrieval-based 3d orientation estimation of unseen objects. In ECCV, pages 106-122. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.32, + 0.906, + 0.36 + ], + "angle": 0, + "content": "[105] Chen Zhao, Tong Zhang, and Mathieu Salzmann. 3d-aware hypothesis & verification for generalizable relative object pose estimation. In ICLR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.906, + 0.403 + ], + "angle": 0, + "content": "[106] Chen Zhao, Tong Zhang, Zheng Dang, and Mathieu Salzmann. Dvmnet: Computing relative pose for unseen objects beyond hypotheses. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.405, + 0.906, + 0.446 + ], + "angle": 0, + "content": "[107] Xu Zhao, Wenchao Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, and Jinqiao Wang. Fast segment anything. arXiv preprint arXiv:2306.12156, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.448, + 0.907, + 0.515 + ], + "angle": 0, + "content": "[108] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splattering: Fast and generalizable single-view 3d reconstruction with transformers. In CVPR, pages 10324-10335, 2024. 8" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.092, + 0.909, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22034" + } + ] +] \ No newline at end of file diff --git a/2025/UNOPose_ Unseen Object Pose Estimation with an Unposed RGB-D Reference Image/94d3840c-f3fe-4b2b-9388-48ccee10c9d7_origin.pdf b/2025/UNOPose_ Unseen Object Pose Estimation with an Unposed RGB-D Reference Image/94d3840c-f3fe-4b2b-9388-48ccee10c9d7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..28da4ca11767f22ca17513289cd230ad981a29ed --- /dev/null +++ b/2025/UNOPose_ Unseen Object Pose Estimation with an Unposed RGB-D Reference Image/94d3840c-f3fe-4b2b-9388-48ccee10c9d7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:552aa9e8d549163f9d09a95c4aa9e5810045fa76b65b5422956f521da0e2f2da +size 4784174 diff --git a/2025/UNOPose_ Unseen Object Pose Estimation with an Unposed RGB-D Reference Image/full.md b/2025/UNOPose_ Unseen Object Pose Estimation with an Unposed RGB-D Reference Image/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3cf9bf0eb354290adc72882a8d59b26bf7a41557 --- /dev/null +++ b/2025/UNOPose_ Unseen Object Pose Estimation with an Unposed RGB-D Reference Image/full.md @@ -0,0 +1,360 @@ +# UNOpose: Unseen Object Pose Estimation with an Unposed RGB-D Reference Image + +Xingyu Liu $^{1, *}$ , Gu Wang $^{1, *}$ , Ruida Zhang $^{1}$ , Chenyangguang Zhang $^{1}$ , Federico Tombari $^{2,3}$ , and Xiangyang Ji $^{1}$ $^{1}$ Tsinghua University, BNRist $^{2}$ Technical University of Munich, $^{3}$ Google + +{liuxy21@mails., wanggu1@, zhangrd23@mails., zcyg22@mails., xyji@}tsinghua.edu.cn, + +tombari@in.tum.de + +# Abstract + +Unseen object pose estimation methods often rely on CAD models or multiple reference views, making the onboarding stage costly. To simplify reference acquisition, we aim to estimate the unseen object's pose through a single unposed RGB-D reference image. While previous works leverage reference images as pose anchors to limit the range of relative pose, our scenario presents significant challenges since the relative transformation could vary across the entire $SE(3)$ space. Moreover, factors like occlusion, sensor noise, and extreme geometry could result in low viewpoint overlap. To address these challenges, we present a novel approach and benchmark, termed UNOPose1, for UNseen One-reference-based object Pose estimation. Building upon a coarse-to-fine paradigm, UNOPose constructs an $SE(3)$ -invariant reference frame to standardize object representation despite pose and size variations. To alleviate small overlap across viewpoints, we recalibrate the weight of each correspondence based on its predicted likelihood of being within the overlapping region. Evaluated on our proposed benchmark based on the BOP Challenge, UNOPose demonstrates superior performance, significantly outperforming traditional and learning-based methods in the one-reference setting and remaining competitive with CAD-model-based methods. The code and dataset are available at github.com/shanice-l/UNOPose. + +# 1. Introduction + +Localizing an object in Euclidean space by estimating its 6DoF pose, i.e., 3DoF orientation and 3DoF position, plays a crucial role in augmented/virtual reality [59, 76], scene understanding [37, 63] and robotic manipulation [23, 26, 80]. The vast majority of works [42, 48, 68, 78, 83-86] focuses on instance-level object pose estimation, where the train + +ing and testing datasets consist of an identical set of known object instances, and the CAD models of objects are often required for generating training images and labels [16, 56]. More recently, Wang et al. [87] and its successors [6, 79, 88] extended this paradigm to the category level, intending to estimate poses for novel instances within predefined categories without requiring CAD models of target objects. Both paradigms have limitations in open-world applications, where annotating and training for new objects beyond known categories would be very labor-intensive and sometimes prohibitive. + +To mitigate this problem, recent research efforts [3, 43, 51, 90] have shifted towards pose estimation of arbitrary novel objects, where the target object is unseen during training. This task presents a great challenge due to the inherent uncertainty of the novel object's canonical frame (defined in the CAD model of the object). Current approaches [43, 57, 62] often tackle this by using omnidirectional reference views to cover the target object, separating pose estimation task to viewpoint selection and relative pose estimation to known pose anchors. If CAD models are available during the test phase, reference views can be easily produced through image rendering [43, 51, 61, 90]. Alternatively, without CAD models, reference views can be obtained by capturing multiple images of the object [29, 31, 55, 77]. Both setups have shortcomings, as creating a CAD model of an object or labeling a substantial number of object poses is limited in scalability. For example, they can hardly adapt to in-the-wild scenarios where novel objects are introduced unpredictably. + +To sum up, while existing methods exhibit promising transfer capability to novel objects, they are notably constrained by their reliance on the CAD model or multiple reference views (Fig. 1 (a, b)). In this work, we aim to minimize onboarding cost by estimating novel object poses using a single unposed RGB-D reference image (Fig. 1). This setup is different from existing methods mainly in two aspects. First, we focus on estimating the relative pose between + +![](images/ea83ed8cb74d85b90fc84596b8b7beecb05bd2252b35bf5958d3239b5ca762a9.jpg) +Figure 1. Illustration of unseen object pose estimation. Given a query image presenting a target object unseen during training, we aim to estimate its segmentation and 6DoF pose w.r.t. a reference frame. While previous methods [43, 57, 77, 90] often rely on the CAD model or multiple RGB(-D) images for reference, we merely use one unposed RGB-D reference image. + +two unposed viewpoints rather than the absolute pose. Note that the absolute pose of a novel object becomes ill-posed without a certain canonical frame, while relative pose estimation remains well-defined regardless of the canonical frame's definition. Secondly, some existing relative pose estimation methods [98, 105, 106] merely use RGB modality to predict the 3DoF rotation, limiting their ability to predict relative translation and generalize beyond their training datasets. In contrast, we fully leverage RGB-D modality to predict the 6DoF relative pose, meanwhile enhancing the network's generalizability to novel objects and environments. + +Nevertheless, estimating the pose of unseen objects with only one reference view presents significant challenges from various aspects. First, previous methods [3, 43] select the most similar reference view as a pose anchor, distinctly reducing the search space for relative poses. However, in our scenario, where both the target and reference object poses are unknown, the relative pose can vary across the entire $SE(3)$ space. Moreover, in partial-to-partial object matching, factors such as occlusion, sensor noise, and extreme geometry can severely interfere with the matching process. To address these challenges, we propose a novel approach and benchmark for UNseen One-reference-based object Pose estimation (UNOPose). Our pipeline leverages the strong generalization capabilities of vision foundation models [40, 65] to produce an effective segmentation of the unseen object. Then, UNOPose employs a coarse-to-fine paradigm for estimating the relative pose between the reference and the query objects. To alleviate the challenge of diverse pose and size variations, we introduce the $SE(3)$ -invariant global reference frame (GRF) to standardize object representation. Subsequently, a hierarchical object encoding + +paradigm based on the local reference frame (LRF) further captures fine-grained geometric details. For achieving reliable correspondences in partial-to-partial object matching, we harness an overlap predictor to identify and concentrate on the overlapping region. + +Moreover, we propose a new benchmark devised from the BOP challenge [34] to facilitate evaluation and future research of unseen object pose estimation with one reference. Extensive experiments on YCB-V [92], LM-O [2] and TUD-L [32] datasets demonstrate that our UNOPose surpasses all compared methods on a single reference setting. To our surprise, UNOPose with an unposed reference is even on par with some $SE(3)$ -invariant-feature-based methods relying on CAD models (Ours $70.9\%$ vs. ZTE-PPF [1] $69.0\%$ vs. Koenig-PPF [41] $75.1\%$ w.r.t. $\mathrm{AR}_{\mathrm{BOP}}$ metric). + +Our contributions can be summarized as follows: + +- To the best of our knowledge, we are the first to conduct unseen object 6DoF pose estimation leveraging a single unposed RGB-D reference. +- Based on the BOP Challenge, we devise a new extensive benchmark tailored for unseen object segmentation and pose estimation with one reference. Additionally, we evaluate several traditional and learning-based methods on this benchmark for completeness. +- We introduce UNOPose, a network for learning relative transformation between reference and query objects. To achieve this, we propose the $SE(3)$ -invariant global and local reference frames, enabling standardized object representations despite variations in pose and size. Furthermore, the network can automatically adjust the confidence of each correspondence by incorporating an overlap predictor. + +# 2. Related Work + +Class-specific Pose Estimation. Class-specific object pose estimation aims at predicting 6DoF poses of either instance of a known object (instance-level) or unseen objects within a known category (category-level). Instance-level object poses are often solved by direct regression [35, 47, 56, 85] using deep neural networks, or by establishing 2D-3D [27, 48, 68] or 3D-3D correspondences [30, 73, 83] which are then leveraged by RANSAC-based PnP/Kabsch algorithms. The instance-level setting requires expensive data generation/annotation [18, 19, 81, 102] and training for every new object. To address this limitation, category-level pose estimation [87] is proposed to estimate 9DoF poses of novel objects among specific categories without CAD models. Mainstream approaches can also be categorized into direct regression [6, 9, 10, 17, 50, 54, 99–101] or correspondence-based [8, 24, 79, 87] methods. However, this setting remains constrained to a limited number of categories, given the additional challenge of aligning canonical frames, managing symmetric objects, and categorizing objects in similar categories correctly (e.g. can and bottle). + +Novel Object Pose Estimation. Estimating the pose of novel objects beyond known classes, i.e., objects at inference time are unseen during training, is a useful yet challenging task. Existing works solve this problem through image-based matching [3, 43, 60, 61, 74, 90, 104] or feature-based matching approaches [4, 29, 36, 51, 62, 66, 77]. Notably, given a target image, OVE6D [3] and MegaPose [43] retrieve the most similar viewpoint from a pre-rendered omnidirectional image database as a coarse pose estimate. A customized neural network then refines the coarse pose estimate. FoundationPose [90] further builds a pose ranking network to score each refined pose hypothesis. Feature-matching methods learn local feature descriptors to construct pixel-level or point-level correspondences. For example, OnePose [77] and OnePose++ [29] reconstruct the 3D point cloud of an unseen object to establish 2D-3D correspondences, whereas SAM-6D [51] builds 3D-3D correspondences. More recently, some methods [4, 66] directly exploit the power of foundation models for zero-shot pose estimation. Exemplarily, FoundPose [66] extracts DINOv2 visual features [65], while FreeZe [4] further utilizes generalizable geometric features [69] to conduct feature matching. + +Relative Pose Estimation. Orthogonal to previous works which rely on known instances or multiple reference views, relative object pose estimation estimates a relative transformation of the unseen object with only one reference image. One closely related topic is camera pose estimation from two views. For example, RelPose [98] and RelPose++ [49] infer a distribution of relative rotations by leveraging an energy-based formulation. Moreover, iFusion [91] optimizes the unknown relative pose by inverting the novel view synthesis diffusion model [53]. Relative camera pose estimation relies + +on static background to establish correspondence between viewpoints. In the field of relative object pose estimation, given two RGB images, 3DAHV [105] presents a hypothesis-and-verification framework to score each relative pose hypothesis. Extended by that, DVMNet [106] introduces a hypothesis-free pipeline to compute relative poses via deep voxel matching. The above works have limited capability in inferring 3DoF relative translations. In this work, with the introduction of depth data, we can estimate a complete 6DoF pose including rotation and translation. + +Point Cloud Registration. Our approach follows the paradigm of correspondence-based methods [14, 15, 45, 89], which initially extract 3D local feature descriptors to establish correspondences between source and target point clouds. Subsequently, they estimate the relative transformation using techniques like SVD, RANSAC, or Hough Voting [44]. Some feature descriptors are proposed to ensure rotational invariance. For example, FPFH [72] descriptor is computed with geometric surface properties. PPF [22] uses the distance and angles to describe the relation of point pairs. TOLDI [93] proposes a robust local reference frame using keypoint normals and neighboring projections. With the advent of deep learning, neural networks are increasingly employed to extract 3D local or global feature descriptors [12, 38, 71, 89, 97]. + +# 3. UNO Object Segmentation and Pose Estimation + +# 3.1. Problem Formulation + +Assuming an arbitrary unseen rigid object in a query image, our goal is to estimate the object's mask $M_q$ and its 6DoF relative pose $\Delta \mathbf{T} \in SE(3)$ with a masked RGB-D reference image exhibiting the target object without major occlusion or truncation. As illustrated in Fig. 1, the input is: + +1) $[I_q|D_q]\in \mathbb{R}^{H\times W\times 4}$ : The query RGB-D image; +2) $\left[\bar{I}_p|D_p\right] \in \mathbb{R}^{H \times W \times 4}$ and $M_p \in \mathbb{R}^{H \times W}$ : The reference RGB-D image and the corresponding binary mask indicating the target object; +3) $K_{q} \in \mathbb{R}^{3 \times 3}$ and $K_{p} \in \mathbb{R}^{3 \times 3}$ : Camera intrinsics of the query and reference images. + +Optionally, if the pose $\mathbf{T}_p \in SE(3)$ of the reference object in the camera frame is known, the query object pose $\mathbf{T}_q \in SE(3)$ can be recovered by $\mathbf{T}_q = \Delta \mathbf{T} \mathbf{T}_p$ . Note that, for practicality, the method should not rely on the absolute pose of the reference object, since the world frame can vary arbitrarily in different applications. We only use the reference object pose during the inference stage for evaluation on standard object pose datasets. + +# 3.2. UNO Object Segmentation + +Firstly, we segment the query object from a cluttered background. Thanks to the great generalization ability of vision + +![](images/3004c9497e8ca6b6f792ef25e37a1f4b4f1722e393b7e280f65fc4b943bf5754.jpg) +Figure 2. The network architecture of UNOpose. Given the query and reference point clouds $\mathbf{Q}_{cam}$ and $\mathbf{P}_{cam}$ in the camera frame, UNOpose first transforms them into the $SE(3)$ -invariant global reference frame (GRF). Then feature descriptors are extracted from sparse point sets for constructing the coarse correlation matrix. For achieving precise correspondences, the fine pose estimation module exploits structural details using positional encoding and local reference frame (LRF) encoding. + +foundation models, methods like [7, 51, 61] can effectively segment novel objects with CAD models. However, unlike previous approaches, which generate diverse descriptors from multiple rendered views, we can access only a single reference image. To address this challenge, we use the SAM model [40] to predict all possible mask proposals from the query image, and then score each mask proposal by comparing DINOv2 [65] descriptors between the query and reference images using cosine similarity, identifying the most similar mask $M_{q}$ . Please refer to the appendix for more details about UNO segmentation. + +# 3.3. UNO Object Pose Estimation + +# 3.3.1 Overview of UNOPose + +Given the predicted mask $M_{q}$ of the query image and mask $M_{p}$ of the reference image, we crop and back-project the object of interest from depth maps $D_{q}, D_{p}$ into the camera space as two point sets $\mathbf{Q}_{cam} \in \mathbb{R}^{N^{Q} \times 3}$ and $\mathbf{P}_{cam} \in \mathbb{R}^{N^{P} \times 3}$ , where $N^{Q}$ and $N^{P}$ denotes the point numbers of query and reference point clouds, respectively. Our goal is to recover the relative transformation $\Delta \mathbf{T} = \{\Delta \mathbf{R}, \Delta \mathbf{t}\}$ by minimizing the correspondence distance + +$$ +\min \sum_ {(\mathbf {q}, \mathbf {p}) \in \mathbf {C}} \| \Delta \mathbf {R} \mathbf {q} + \Delta \mathbf {t} - \mathbf {p} \| _ {2}, \tag {1} +$$ + +where $\mathbf{C}$ is the predicted correspondence set between $\mathbf{Q}_{cam}$ and $\mathbf{P}_{cam}$ . We exploit both color and geometric cues to build this correspondence and follow the broadly used coarse-to-fine paradigm in point cloud registration to solve Eq. (1). The network is demonstrated in Fig. 2. + +# 3.3.2 Coarse-to-fine Pose Estimation + +Constructing a Pose-invariant Reference Frame. Given only an unposed reference image, the relative pose to predict is arbitrary in the $SE(3)$ space, which renders a significant challenge for achieving robust correspondences. Hence, we introduce a pose-invariant global reference frame (GRF) and transform $\mathbf{Q}_{cam}$ and $\mathbf{P}_{cam}$ into GRF as $\mathbf{Q}_G$ and $\mathbf{P}_G$ . + +Concretely, taking $\mathbf{Q}_{cam}$ as an example, transforming the point cloud to GRF involves a 7DoF coordinate transformation $\{\mathbf{R}_G\in SO(3),\mathbf{t}_G\in \mathbb{R}^3,s_G\in \mathbb{R}\}$ : + +$$ +\mathbf {Q} _ {G} = \left\{\mathbf {R} _ {G} ^ {\top} \left(\mathbf {q} - \mathbf {t} _ {G}\right) / s _ {G} \mid \mathbf {q} \in \mathbf {Q} _ {c a m} \right\}. \tag {2} +$$ + +The origin of GRF is located at the object center $\mathbf{c}_Q$ for translation invariance, and the radius of the point cloud is rescaled to 1 for size invariance, computed as + +$$ +\mathbf {t} _ {G} = \mathbf {c} _ {Q}, \quad s _ {G} = \max _ {\mathbf {q} \in \mathbf {Q} _ {c a m}} \| \mathbf {q} - \mathbf {c} _ {Q} \| _ {2}. \tag {3} +$$ + +The key ingredient is to devise the rotation $\mathbf{R}_G = [\mathbf{r}_{Gx}|\mathbf{r}_{Gy}|\mathbf{r}_{Gz}]$ , where $\mathbf{r}_{Gx},\mathbf{r}_{Gy},\mathbf{r}_{Gz}$ are the three columns + +of $\mathbf{R}_G$ , ensuring orientation invariance for the transformed point cloud. Inspired by former works [25, 75, 93], we use the normal vector of the object center $\mathbf{n}(\mathbf{c}_Q)$ as $\mathbf{r}_{Gz}$ , and project all points in $\mathbf{Q}_{cam}$ onto the tangent plane of $\mathbf{c}_Q$ , summarizing the projected vectors to determine $\mathbf{r}_{Gx}$ . Then, $\mathbf{r}_{Gy}$ is obtained by the cross-product of $\mathbf{r}_{Gx}$ and $\mathbf{r}_{Gz}$ , i.e., $\mathbf{r}_{Gy} = \mathbf{r}_{Gx} \times \mathbf{r}_{Gz}$ . This ensures that the XY-plane evenly divides the point cloud, with the x-axis representing the principal projection direction. The GRF is $SE(3)$ -invariant as it can be uniquely determined by the spatial distribution of the point cloud, making the transformed point cloud robust to variations in pose and size. + +Concretely, we perform SVD on the covariance matrix $\mathrm{Cov}(\mathbf{Q}_{cam}) = \frac{1}{N_Q}\mathbf{Q}_{cam}^\top \mathbf{Q}_{cam} - \mathbf{c}_Q\mathbf{c}_Q^\top$ and use the singular vector w.r.t. the minimum singular value to determine $\mathbf{r}_{Gz}$ : + +$$ +\mathbf {r} _ {G z} = \left\{ \begin{array}{l l} \mathbf {n} \left(\mathbf {c} _ {Q}\right), & \text {i f} \mathbf {n} \left(\mathbf {c} _ {Q}\right) ^ {\top} \sum_ {\mathbf {q} \in \mathbf {Q} _ {c a m}} \left(\mathbf {c} _ {Q} - \mathbf {q}\right) > 0 \\ - \mathbf {n} \left(\mathbf {c} _ {Q}\right), & \text {o t h e r w i s e .} \end{array} \right. \tag {4} +$$ + +Afterwards, $\mathbf{r}_{Gx}$ is computed as + +$$ +\mathbf {r} _ {G x} = \sum_ {\mathbf {q} \in \mathbf {Q} _ {c a m}} w _ {q} \left(\left(\mathbf {q} - \mathbf {c} _ {Q}\right) - \mathbf {r} _ {G z} ^ {\top} \left(\mathbf {q} - \mathbf {c} _ {Q}\right) \mathbf {r} _ {G z}\right), \tag {5} +$$ + +where $w_{q}$ is a weight w.r.t. distance between $\mathbf{q}$ and $\mathbf{c}_Q$ (See the appendix for details). + +Previous works often depend on complex networks or computationally expensive PPF features for $SE(3)$ invariance [46, 94, 96]. In contrast, our transformation is computationally efficient and can seamlessly adapt to a wide range of network architectures. + +Coarse Pose Estimation. Given point clouds $\mathbf{Q}_G$ and $\mathbf{P}_G$ in GRF, we sample two sparse point sets $\mathbf{Q}_G^c\in \mathbb{R}^{N^c\times 3}$ and $\mathbf{P}_G^c\in \mathbb{R}^{N^c\times 3}$ to efficiently obtain a coarse pose initialization $\Delta \mathbf{T}_{init}$ . Specifically, we leverage a geometry encoder [71] and a color encoder [65] to extract point cloud and RGB features separately. Features are further concatenated as $f_{Q}^{c}\in \mathbb{R}^{N^{c}\times d}$ and $f_{P}^{c}\in \mathbb{R}^{N^{c}\times d}$ , where $d$ is the dimension of embeddings. Following [51], we add a learnable background token for assigning non-overlapping points. The embeddings, denoted as $\hat{f}_Q^c\in \mathbb{R}^{(N^c +1)\times d}$ and $\hat{f}_P^c\in \mathbb{R}^{(N^c +1)\times d}$ , are processed by three stacked Geometric Transformer decoding modules [71]. + +The outputs of the last decoder $\hat{F}_Q^c\in \mathbb{R}^{(N^c +1)\times D}$ and $\hat{F}_P^c\in \mathbb{R}^{(N^c +1)\times D}$ are point-wise features for building the correlation matrix. However, in our setting, the overlap ratio can be influenced by complicated factors like viewpoint, occlusion or depth noise. To solve this issue, the network additionally predicts overlap confidence $\hat{O}_Q^c\in \mathbb{R}^{(N^c +1)\times 1}$ and $\hat{O}_P^c\in \mathbb{R}^{(N^c +1)\times 1}$ , which indicate point-wise probabilities of being in the overlapping region. The overlap-aware correlation matrix $\mathbf{X}^c\in \mathbb{R}^{(N^c +1)\times (N^c +1)}$ can thus be com + +puted by + +$$ +\mathbf {X} ^ {c} = \operatorname {s o f t m a x} [ (\hat {O} _ {Q} ^ {c} \odot \hat {F} _ {Q} ^ {c}) (\hat {O} _ {P} ^ {c} \odot \hat {F} _ {P} ^ {c}) ^ {\top} ]. \tag {6} +$$ + +Here $\odot$ indicates element-wise multiplication. Each element in $\mathbf{X}^c$ suggests a correlation score between a point pair in $\mathbf{Q}^c$ and $\mathbf{P}^c$ . + +Once $\mathbf{X}^c$ is computed, we can extract all possible corresponding point pairs between $\mathbf{Q}^c$ and $\mathbf{P}^c$ as well as their correlation scores to solve Eq. (1). We first compute pose hypotheses by sampling $N_{H}$ triplets of point pairs according to the distribution of $\mathbf{X}^c$ . Then each pose hypothesis $\Delta \mathbf{T}_h = \{\Delta \mathbf{R}_h, \Delta \mathbf{t}_h\}$ is scored with the reciprocal of distance $D_h$ as in [27, 51] + +$$ +\begin{array}{l} D _ {h} = \frac {1}{N ^ {c}} \sum_ {\mathbf {p} ^ {c} \in \mathbf {P} ^ {c}} \min _ {\mathbf {q} ^ {c} \in \mathbf {Q} ^ {c}} \| \Delta \mathbf {R} _ {h} ^ {\top} \left(\mathbf {q} ^ {c} - \Delta \mathbf {t} _ {h}\right) - \mathbf {p} ^ {c} \| _ {2}, \tag {7} \\ \xi_ {h} = \frac {1}{D _ {h}}, \quad h = 1, 2, 3, \dots , N _ {H}. \\ \end{array} +$$ + +The pose hypothesis with the highest score $\xi_{h}$ is selected as the initial pose prediction $\Delta \mathbf{T}_{init} = \{\Delta \mathbf{R}_{init},\Delta \mathbf{t}_{init}\}$ and further sent to transform the input of the next stage. + +Fine Pose Estimation. After transforming $\mathbf{Q}_{cam}$ to $\tilde{\mathbf{Q}}_{cam}$ with the initial pose prediction $\Delta \mathbf{T}_{init}$ , we perform a fine matching process between two dense point sets, i.e., $\tilde{\mathbf{Q}}^f \in \mathbb{R}^{N^f \times 3}$ and $\mathbf{P}^f \in \mathbb{R}^{N^f \times 3}$ , to achieve a more accurate pose. The network exploits geometric details through a hierarchical encoding paradigm, comprising a positional encoding layer and a local reference frame encoding layer. For each point, we first encode its global position using a mini-PointNet [70], and then construct SE(3)-invariant local reference frame (LRF) to gather local descriptors. These two encodings complement each other, as local descriptors capture fine geometric structures within small neighborhoods, while the positional encoding layer offers global geometric context. + +Leveraging $\tilde{\mathbf{Q}}^f$ for example, the process of constructing LRF encoding is as follows. For each point $\tilde{\mathbf{q}}^m$ in $\tilde{\mathbf{Q}}^f$ , we build a local region set $\tilde{\mathcal{Q}}^m = \{\tilde{\mathbf{q}}_j\}$ where $\| \tilde{\mathbf{q}}_j - \tilde{\mathbf{q}}^m \|_2 \leq r\}_{j=1}^{N_D}$ by grouping $N_D$ neighboring points of $\tilde{\mathbf{q}}^m$ . The transformation pose $\{\mathbf{R}_L^m, \mathbf{t}_L^m, s_L^m\}$ for LRF is computed similarly to GRF (Eq. 3, 4, 5), except that LRF is built upon a local point set while GRF is based on the entire point cloud. By computing the transformation pose, we calculate local point descriptors as + +$$ +\begin{array}{l} \mathcal {Q} _ {L} ^ {m} = \tilde {\mathcal {Q}} _ {L} ^ {m} = \left\{\left(\mathbf {R} _ {L} ^ {m}\right) ^ {\top} \left(\tilde {\mathbf {q}} _ {j} - \mathbf {t} _ {L} ^ {m}\right) / s _ {L} ^ {m} \right\} _ {j = 1} ^ {N _ {D}}, \tag {8} \\ \mathbf {Q} _ {L} ^ {f} = \{\mathcal {Q} _ {L} ^ {m} \} _ {m = 1} ^ {N ^ {f}}. \\ \end{array} +$$ + +Similarly, we transform $\mathbf{P}^f$ to $\mathbf{P}_L^f$ . Note that $\mathcal{Q}_L^m = \tilde{\mathcal{Q}}_L^m$ since LRF is pose-invariant. Then, the LRF encoding is extracted from $\mathbf{Q}_L^f$ and $\mathbf{P}_L^f$ with a three-layer MLP. + +Positional encoding and LRF encoding are combined with the geometric and color features as the input of the + +
MethodSeg. ModelModalityCAD modelLM-OTUD-LYCB-VAvgTime (s)
ZeroPose [7]SAM [40]RGB✓35.642.153.443.73.90
CNOS [61]FastSAM [107]RGB✓39.748.059.949.20.23
CNOS [61]SAM [40]RGB✓39.639.159.546.11.71
SAM-6D [51]FastSAM [107]RGB-D✓42.251.762.152.01.47
SAM-6D [51]SAM [40]RGB-D✓46.056.960.554.54.53
UNOseg (Ours)FastSAM [107]RGBX37.350.367.351.60.24
UNOseg (Ours)SAM [40]RGBX39.756.266.854.22.68
+ +Table 1. Segmentation results (mAP) on LM-O, TUD-L, and YCB-V. Results of other methods are obtained from the BOP website bop.felk.cvut.cz/leaderboards/segmentation-unseen-bop23. We denote the best score in bold and the second best score with underline. + +
MethodSeg. ModelModalityRef. TypeLM-OTUD-LYCB-VAvgTime (s)
ZTE-PPF [1]MRCNN [28]DModel66.390.450.269.00.74
Koenig-PPF [41]Retina/MRCNNRGB-DModel63.192.070.175.10.99
Ref. AlignCenterUNOSeg-SAMDImage27.328.847.534.52.69
FPFH+RANSAC [72]UNOSeg-SAMDImage31.031.050.037.36.38
FPFH+MAC [103]UNOSeg-SAMDImage22.522.149.631.4136.94
PPF [22]UNOSeg-SAMDImage29.714.838.327.611.79
PPF_3D_ICP [22]UNOSeg-SAMDImage44.729.166.846.914.27
FCGF+RANSAC [12]UNOSeg-SAMDImage38.959.057.651.810.96
FCGF+MAC [103]UNOSeg-SAMDImage33.948.351.044.460.53
UTOPIC [11]UNOSeg-SAMDImage13.735.410.519.94.00
GeDi [69]UNOSeg-SAMDImage42.867.360.656.948.89
FreeZe [4]UNOSeg-SAMRGB-DImage45.568.365.559.852.96
SAM-6D* [51]UNOSeg-SAMRGB-DPosed Image54.529.768.150.84.21
UNOpose (Ours)UNOSeg-FastSAMRGB-DImage55.866.581.968.11.11
UNOpose (Ours)UNOSeg-SAMRGB-DImage58.771.083.170.93.70
+ +Table 2. Pose estimation results on LM-O, TUD-L, and YCB-V. The mean Average Recall (%) of the BOP metric and the average time (s) per image are reported. We highlight the best scores of each setting in bold. Specifically, for adapting SAM-6D [51] to our setting, we randomly choose one of the two templates which [51] uses as a canonical reference during training. + +Geometric Transformer. Similar to initial pose prediction, by decoding these features, we obtain the fine point-wise features $\hat{F}_Q^f,\hat{F}_P^f$ and overlap confidencees $\hat{O}_Q^f,\hat{O}_P^f$ , which are then used to obtain the overlap-aware fine correlation matrix $\mathbf{X}^{f}\in \mathbb{R}^{(N^{f} + 1)\times (N^{f} + 1)}$ . The final pose $\Delta \mathbf{T}$ is predicted by solving Eq. (1) using $\mathbf{X}^f$ with the weighted SVD algorithm. + +# 4. Experiments + +# 4.1. Experimental Setup + +Implementation Details. Implemented with PyTorch [67], we use the DINOv2 [13, 65] pre-trained ViT-base [20] to encode color information. GeoTransformers [71] with attention [82] and sparse-to-dense linear attention [51] are respectively employed to extract coarse and fine geometric correspondences. The network is trained with standard MegaPose synthetic dataset [43] for BOP unseen object pose estimation track [34]. It contains $\sim 2\mathrm{M}$ images of $\sim 50\mathrm{K}$ objects from GSO [21] and ShapeNet [5]. We train 3 epochs with a batch size of 32 on 4 RTX 3090 GPUs. The network is trained with Adam [39] and cosine annealing [58], using a base learning rate of $10^{-4}$ . Moreover, we lever + +age InfoNCE [64] loss for learning correlation matrix and weighted BCE loss for supervising overlap prediction. + +Datasets and Benchmarks. We evaluate our method on BOP test split [33] of three different real-world datasets: LM-O [2], TUD-L [32], and YCB-V [92]. LM-O consists of 200 images from one clutter scene, where eight objects with occlusion are provided for testing. TUD-L comprises three moving objects with mild occlusion and diverse lighting conditions. YCB-V is very challenging due to severe occlusions and sensor noise. It comprises 900 testing images from 12 scenes consisting of 21 objects. For each query object in the test split, we randomly select a reference object beyond the test set (in the training or evaluation set). Unless specified, we constrain the maximal rotation difference between reference and target to $50^{\circ}$ to ensure the ground-truth correspondence set is not empty. Moreover, the occlusion ratio of reference objects is required to be less than $30\%$ . + +Evaluation Metrics. For evaluating instance segmentation, we follow [33] to use the mean Average Precision (mAP) metric, which is computed across different Intersection-over-Union (IoU) thresholds (\{0.5, 0.55, ..., 0.95\}). Only annotated instances with at least $10\%$ projected visible surfaces + +
CategoryVSDMSSDMSPDARBOP
Dragon47.173.269.463.2
Frog59.452.657.656.4
Can56.865.665.262.5
Avg54.463.764.160.7
+ +(a) + +
Ref.
Ref.
+ +are considered for evaluation. For object pose evaluation, we use the symmetry-aware BOP metric [33]. It is calculated as the mean of the Average Recall of Visible Surface Discrepancy (VSD), Maximum Symmetry-Aware Surface Distance (MSSD), and Maximum Symmetry-Aware Projection Distance (MSPD) (See [33] for details). + +# 4.2. UNO Object Segmentation and Pose Estimation Results + +UNO Object Segmentation Results. We compare our UNOSeg with some existing CAD-model-based unseen object segmentation methods [7, 51, 61] in Tab. 1. Leveraging a single image as a reference, our UNOSeg achieves comparable results with state-of-the-art methods (Ours $54.2\%$ vs. SAM-6D $54.5\%$ ) at a faster speed (Ours 2.68s vs. SAM-6D 4.53s). Notably, UNOSeg-FastSAM achieves the best performance on YCB-V (mAP $67.3\%$ ). Compared to FastSAM [107], SAM [40] achieves consistently better results on three datasets. Therefore, we choose UNOSeg-SAM as the default segmentation for subsequent experiments. + +UNO Object Pose Estimation Results. Since we are the first to conduct relative 6DoF pose estimation for unseen objects and build a brand-new benchmark, we re-implement some traditional [22, 72, 103] and learning-based methods [4, 11, 12, 51, 69] for comparison. Note that we leverage identical data pre-processing, segmentation, query and reference image pairs, and evaluation protocols for all methods for fair comparison. Moreover, we also compare UNOPose with two CAD-model-based pose estimators leveraging $SE(3)$ -invariant PPF feature [1, 41]. All results are illustrated in Tab. 2. It clearly shows that our method surpasses all image-reference-based methods by a large margin, and even achieves comparable results with CAD-model-based methods using class-specific detectors [28, 52] (Ours $70.9\%$ vs. ZTE-PPF [1] $69.0\%$ vs. Koenig-PPF [41] $75.1\%$ ). "Ref. AlignCenter" directly uses the reference rotation as prediction, meanwhile leveraging the shift between query and reference centers as relative translation, so it can be regarded as the baseline of all methods. Among all traditional descriptors, PPF with Iterative Closest Point (ICP) refinement [22] achieves the best results (AR $46.9\%$ ). Learning-based methods demonstrate varying generalization abilities on unseen objects. For example, the GeoTransformer-based + +(b) +Table 3. One reference per category experiments on TUD-L. Blue and green contours denote GT and estimated poses, respectively. + +
ROWGRFLRFFPEVSDMSSDMSPD\( \mathrm{AR}_{\text{BOP}} \)
A0✘✘✘64.668.159.263.9
A1✓✘✘66.472.863.267.5
A2✘✓✘66.269.560.465.4
A3✓✓✘66.973.263.567.9
B0✘✘✓79.885.075.580.1
B1✓✘✓79.485.276.680.4
B2✘✓✓81.185.176.680.9
B3✓✓✓80.786.077.781.5
+ +Table 4. Ablation of GRF and LRF on YCB-V. FPE stands for "fine pose estimation". + +method UTOPIC, designed for registering object-level partial point clouds, performs suboptimally in our setting. In contrast, the adapted unseen object pose estimation methods, i.e. SAM-6D [51] and FreeZe [4], achieve satisfactory results compared to their counterparts. + +Run-time Analysis. Evaluated on a single RTX 3090 GPU, our pipeline runs at 3.70s for one image, including 2.68s for segmentation and 1.02s for pose estimation (Tab. 1 - 2). + +# 4.3. One Reference for a Category + +To minimize the efforts of acquiring references, we leverage a single reference for all objects of the same category in a dataset. Tab. 3 illustrates the qualitative and quantitative experimental results on TUD-L. It is worth noting that the target object in TUD-L is dynamic rather than static, preventing methods from leveraging background information. This setup poses challenges for relative pose estimation methods that rely on scene features, making them likely to fail in this task. Despite this, UNOPose achieves impressive performance (60.7% in terms of the $\mathrm{AR}_{\mathrm{BOP}}$ metric) with only a single reference per category. The visualization results in Tab. 3b further show UNOPose is capable of predicting omnidirectional relative poses. Additional experiments on reference selection are provided in the appendix. + +# 4.4. Ablation Studies + +Effectiveness of GRF and LRF. As demonstrated in Sec. 3.3.2, GRF transfers arbitrary object poses to an $SE(3)$ -invariant frame, while LRF assists with extracting precise local descriptors for fine pose estimation. Tab. 4 shows that + +
ROWMethodVSDMSSDMSPD\( \mathrm{AR}_{\text{BOP}} \)
C0B3 + Soft Correspondence [95]77.181.772.777.2
C1B3 + Overlap Predictor82.687.479.483.1
C2A3 + Overlap Predictor68.474.664.869.2
D0C1: Seg. → CNOS [61]77.480.973.677.3
D1C1: Seg. → SAM-6D [51]78.382.174.378.2
D2C1: Seg. → MRCNN [28, 42]81.285.777.981.6
D3C1: Seg. → GT Seg.86.689.282.186.0
E0C1: Ref. Real → Rendered83.789.683.185.5
+ +Table 5. Ablation of overlap predictor, segmentation, and reference type on YCB-V. + +![](images/85f1dfbd821bff5428a02769c25d05f473cb29697d79d389fa2aaf40cca98d1d.jpg) +Figure 3. Ablation of initial rotation distance between query and reference objects. We categorize all testing objects into nine groups according to the initial rotation distance, and evaluate the $\mathrm{AR}_{\mathrm{BOP}}$ metric and overlap ratio for each group separately. + +the effectiveness of GRF and LRF is complementary in UNOpose. Specifically, GRF brings significant performance enhancement in the coarse pose estimation (Tab. 4 A1 vs. A0, A3 vs. A2), providing an accurate pose initialization for the next stage. Meanwhile, LRF is constructed to exploit fine-grained local information, primarily enhancing the fine pose estimation results (Tab. 4 B0 vs. B2, B1 vs. B3). With both GRF and LRF, our network achieves obvious performance improvement (Tab. 4 A3 vs. A0, B3 vs. B0). + +Ablation on Correspondence Loss. In partial-to-partial point cloud registration, only some of the points can find their counterparts on the target point cloud. Therefore, the introduction of overlap prediction helps to distinguish the points in the overlap region and achieves good results (Tab. 4-5 C1 vs. B3, C2 vs. A3). Moreover, we consider soft correspondence loss in [95] as an alternative, which predicts probabilities of overlap between each point pair. However, the result did not meet the expectations (Tab. 5 C0). + +Ablation on Segmentation. Tab. 1 showcases the advantage of our UNOSeg over other segmentation methods on YCB-V. Furthermore, substituting UNOSeg with CNOS [61], SAM-6D [51], or MRCNN [28, 42] results in a noticeable + +decline in UNOpose's performance (Tab. 5 D0-D2). Moreover, substituting UNOSeg with ground-truth segmentation brings $\approx 3\%$ performance enhancement (Tab. 5 D3 vs. C1). + +Using Rendered References. We replaced the real reference with a pose-identical rendered version and present the results in Tab. 5. With reduced depth noise and domain gap between training and testing, UNOPose works even better using the rendered reference (Tab. 5 E0 vs. C1). However, this ablation requires the object model, which is more difficult to acquire than an unposed real image. + +Ablation on Reference Rotation Distance. To investigate the impact of the initial rotation distance between the query and reference images, we randomly select the reference object with a maximum rotation distance of $90^{\circ}$ for each query object. Further, all testing objects are categorized into nine groups according to initial rotation distances. We separately evaluate each group and present the results in Fig. 3. Meanwhile, we calculate the overlap ratio between reference and query instances of each group and visualize the mean and standard deviation. Notably, performance significantly decreases when the rotation distance exceeds $50^{\circ}$ . This decline is attributed to the reduced overlap between query and reference viewpoints. However, our network still shows favorable transfer ability in extreme relative pose estimation (rotation distance $(80^{\circ}, 90^{\circ}]$ , overlap ratio $45.4\%$ , $\mathrm{AR}_{\mathrm{BOP}} 54.8\%$ ). + +The appendix presents further implementation details, visualization results, and more experimental analysis on reference selection and extensive datasets. + +# 5. Conclusion, Limitation and Future work + +This work has introduced UNOPose, a new approach for unseen object pose estimation with one reference. The key idea is constructing the $SE(3)$ -invariant reference frame for tackling diverse pose and size variations. Moreover, we propose an overlap predictor for handling low-overlap scenarios. To evaluate the proposed method, we devise a new benchmark based on the BOP challenge and compare some state-of-the-art methods. Experimental results show UNOPose surpasses all compared reference-based methods significantly, and is competitive with some $SE(3)$ -invariant-feature-based methods relying on CAD models. + +Limitation and Future Work. While efficient, correspondence-based methods rely on predicting robust and dense correspondence between query and reference objects, thereby limited in extreme two-view geometry applications. Future work will focus on reconstructing the unseen object from a single reference [53, 108] and concurrently estimating its object pose. + +Acknowledgements. This work was supported in part by the National Natural Science Foundation of China under Grant No. 62406169, and in part by the China Postdoctoral Science Foundation under Grant No. 2024M761673. + +# References + +[1] Anonymous. Zte-ppf. https://bop.felk(cvut.cz/method_info/250/, 2022. 2, 6, 7 +[2] Eric Brachmann, Alexander Krull, Frank Michel, Stefan Gumhold, Jamie Shotton, and Carsten Rother. Learning 6D object pose estimation using 3D object coordinates. In ECCV, 2014. 2, 6 +[3] Dingding Cai, Janne Heikkilä, and Esa Rahtu. Ove6d: Object viewpoint encoding for depth-based 6d object pose estimation. In CVPR, pages 6803-6813, 2022. 1, 2, 3 +[4] Andrea Caraffa, Davide Boscaini, Amir Hamza, and Fabio Poiesi. Freeze: Training-free zero-shot 6d pose estimation with geometric and vision foundation models. In ECCV, 2024. 3, 6, 7 +[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012, 2015. 6 +[6] Dengsheng Chen, Jun Li, Zheng Wang, and Kai Xu. Learning canonical shape space for category-level 6D object pose and size estimation. In CVPR, pages 11970-11979, 2020. 1, 3 +[7] Jianqiu Chen, Zikun Zhou, Mingshan Sun, Rui Zhao, Liwei Wu, Tianpeng Bao, and Zhenyu He. Zeropose: Cadprompted zero-shot object 6d pose estimation in cluttered scenes. IEEE TCSVT, 2024. 4, 6, 7 +[8] Kai Chen and Qi Dou. SGPA: Structure-guided prior adaptation for category-level 6D object pose estimation. In ICCV, pages 2773-2782, 2021. 3 +[9] Wei Chen, Xi Jia, Hyung Jin Chang, Jinming Duan, Shen Linlin, and Ales Leonardis. FS-Net: Fast shape-based network for category-level 6D object pose estimation with decoupled rotation mechanism. In CVPR, pages 1581–1590, 2021. 3 +[10] Yamei Chen, Yan Di, Guangyao Zhai, Fabian Manhardt, Chenyangguang Zhang, Ruida Zhang, Federico Tombari, Nassir Navab, and Benjamin Busam. Secondpose: Se (3)-consistent dual-stream feature fusion for category-level pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9959-9969, 2024. 3 +[11] Zhilei Chen, Honghua Chen, Lina Gong, Xuefeng Yan, Jun Wang, Yanwen Guo, Jing Qin, and Mingqiang Wei. Utopic: Uncertainty-aware overlap prediction network for partial point cloud registration. In Computer Graphics Forum, pages 87-98. Wiley Online Library, 2022. 6, 7 +[12] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In ICCV, pages 8958-8966, 2019. 3, 6, 7 +[13] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In ICLR, 2024. 6 +[14] Haowen Deng, Tolga Birdal, and Slobodan Ilic. Ppf-foldnet: Unsupervised learning of rotation invariant 3d local descriptors. In ECCV, pages 602-618, 2018. 3 + +[15] Haowen Deng, Tolga Birdal, and Slobodan Ilic. Ppfnet: Global context aware local features for robust 3d point matching. In CVPR, pages 195-205, 2018. 3 +[16] Maximilian Denninger, Dominik Winkelbauer, Martin Sundermeyer, Wout Boerdijk, Markus Knauer, Klaus H. Strobl, Matthias Humt, and Rudolph Triebel. Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software, 8(82):4901, 2023. 1 +[17] Yan Di, Ruida Zhang, Zhiqiang Lou, Fabian Manhardt, Xi-angyang Ji, Nassir Navab, and Federico Tombari. Gpv-posed Category-level object pose estimation via geometry-guided point-wise voting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6781–6791, 2022. 3 +[18] Yan Di, Chenyangguang Zhang, Ruida Zhang, Fabian Manhardt, Yongzhi Su, Jason Rambach, Didier Stricker, Xi-angyang Ji, and Federico Tombari. U-red: Unsupervised 3d shape retrieval and deformation for partial point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8884–8895, 2023. 3 +[19] Yan Di, Chenyangguang Zhang, Chaowei Wang, Ruida Zhang, Guangyao Zhai, Yanyan Li, Bowen Fu, Xiangyang Ji, and Shan Gao. Shapematcher: Self-supervised joint shape canonicalization segmentation retrieval and deformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21017-21028, 2024. 3 +[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 6 +[21] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In ICRA, pages 2553-2560. IEEE, 2022. 6 +[22] Bertram Drost, Markus Ulrich, Nassir Navab, and Slobodan Ilic. Model globally, match locally: Efficient and robust 3d object recognition. In CVPR, pages 998-1005. IEEE, 2010. 3, 6, 7 +[23] Guoguang Du, Kai Wang, Shiguo Lian, and Kaiyong Zhao. Vision-based robotic grasping from object localization, object pose estimation to grasp estimation for parallel grippers: a review. Artificial Intelligence Review, 54(3):1677-1734, 2021. 1 +[24] Zhaoxin Fan, Zhenbo Song, Jian Xu, Zhicheng Wang, Kejian Wu, Hongyan Liu, and Jun He. Object level depth reconstruction for category level 6d object pose estimation from monocular rgb image. In ECCV, pages 220-236. Springer, 2022. 3 +[25] Andrea Frome, Daniel Huber, Ravi Kolluri, Thomas Bulow, and Jitendra Malik. Recognizing objects in range data using regional point descriptors. In ECCV, pages 224-237. Springer, 2004. 5 +[26] Bowen Fu, Sek Kun Leong, Yan Di, Gu Wang, Jiwen Tang, and Xiangyang Ji. Lanpose: Language-instructed 6d object pose estimation for robotic assembly. In ECCVW, 2024. 1 + +[27] Rasmus Laurvig Haugaard and Anders Glent Buch. Surfemb: Dense and continuous correspondence distributions for object pose estimation with learnt surface embeddings. In CVPR, pages 6749-6758, 2022. 3, 5 +[28] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask R-CNN. In ICCV, pages 2961-2969, 2017. 6, 7, 8 +[29] Xingyi He, Jiaming Sun, Yuang Wang, Di Huang, Hujun Bao, and Xiaowei Zhou. Onepose++: Keypoint-free one-shot object pose estimation without cad models. NeurIPS, 35:35103-35115, 2022. 1, 3 +[30] Yisheng He, Haibin Huang, Haoqiang Fan, Qifeng Chen, and Jian Sun. Ffb6d: A full flow bidirectional fusion network for 6d pose estimation. In CVPR, pages 3003-3013, 2021. 3 +[31] Yisheng He, Yao Wang, Haoqiang Fan, Jian Sun, and Qifeng Chen. Fs6d: Few-shot 6d pose estimation of novel objects. In CVPR, pages 6814-6824, 2022. 1 +[32] Tomas Hodan, Frank Michel, Eric Brachmann, Wadim Kehl, Anders GlentBuch, Dirk Kraft, Bertram Drost, Joel Vidal, Stephan Ihrke, Xenophon Zabulis, et al. BOP: Benchmark for 6D Object Pose Estimation. In ECCV, pages 19-34, 2018. 2, 6 +[33] Tomas Hodan, Martin Sundermeyer, Bertram Drost, Yann Labbe, Eric Brachmann, Frank Michel, Carsten Rother, and Jiri Matas. BOP Challenge 2020 on 6D object localization. In ECCVW, pages 577-594, 2020. 6, 7 +[34] Tomas Hodan, Martin Sundermeyer, Yann Labbe, Van Nguyen Nguyen, Gu Wang, Eric Brachmann, Bertram Drost, Vincent Lepetit, Carsten Rother, and Jiri Matas. Bop challenge 2023 on detection, segmentation and pose estimation of seen and unseen rigid objects. In CVPRW, 2024. 2, 6 +[35] Yinlin Hu, Pascal Fua, Wei Wang, and Mathieu Salzmann. Single-stage 6d object pose estimation. In CVPR, pages 2930-2939, 2020. 3 +[36] Junwen Huang, Hao Yu, Kuan-Ting Yu, Nassir Navab, Slobodan Ilic, and Benjamin Busam. Matchu: Matching unseen objects for 6d pose estimation from rgb-d images. In CVPR, 2024. 3 +[37] Siyuan Huang, Siyuan Qi, Yinxue Xiao, Yixin Zhu, Ying Nian Wu, and Song-Chun Zhu. Cooperative holistic scene understanding: Unifying 3d object, layout, and camera pose estimation. NeurIPS, 31, 2018. 1 +[38] Shengyu Huang, Zan Gojcic, Mikhail Usvyatsov, Andreas Wieser, and Konrad Schindler. Predator: Registration of 3d point clouds with low overlap. In CVPR, pages 4267-4276, 2021. 3 +[39] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6 +[40] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 2, 4, 6, 7 +[41] Rebecca König and Bertram Drost. A hybrid approach for 6 dof pose estimation. In ECCV, pages 700-706. Springer, 2020. 2, 6, 7 + +[42] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. CosyPose: Consistent multi-view multi-object 6D pose estimation. In ECCV, pages 574-591, 2020. 1, 8 +[43] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. Megapose: 6d pose estimation of novel objects via render & compare. In CoRL, pages 715-725. PMLR, 2023. 1, 2, 3, 6 +[44] Junha Lee, Seungwook Kim, Minsu Cho, and Jaesik Park. Deep hough voting for robust global registration. In ICCV, pages 15994-16003, 2021. 3 +[45] Jiahao Li, Changhao Zhang, Ziyao Xu, Hangning Zhou, and Chi Zhang. Iterative distance-aware similarity matrix convolution with mutual-supervised point elimination for efficient point cloud registration. In ECCV, pages 378-394. Springer, 2020. 3 +[46] Xiaolong Li, Yijia Weng, Li Yi, Leonidas J Guibas, A Abbott, Shuran Song, and He Wang. Leveraging se (3) equivariance for self-supervised category-level object pose estimation from point clouds. NeurIPS, 34, 2021. 5 +[47] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. DeepIM: Deep iterative matching for 6D pose estimation. IJCV, 128(3):657-678, 2020. 3 +[48] Zhigang Li, Gu Wang, and Xiangyang Ji. CDPN: Coordinates-Based Disentangled Pose Network for Real-Time RGB-Based 6-DoF Object Pose Estimation. In ICCV, pages 7678-7687, 2019. 1, 3 +[49] Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. In 3DV, 2024. 3 +[50] Jiehong Lin, Zewei Wei, Zhihao Li, Songcen Xu, Kui Jia, and Yuanqing Li. DualPoseNet: Category-level 6D object pose and size estimation using dual pose network with refined learning of pose consistency. In ICCV, pages 3560-3569, 2021. 3 +[51] Jiehong Lin, Lihua Liu, Dekun Lu, and Kui Jia. Sam-6d: Segment anything model meets zero-shot 6d object pose estimation. In CVPR, 2024. 1, 3, 4, 5, 6, 7, 8 +[52] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 7 +[53] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In ICCV, pages 9298–9309, 2023. 3, 8 +[54] Xingyu Liu, Gu Wang, Yi Li, and Xiangyang Ji. CATRE: iterative point clouds alignment for category-level object pose refinement. In ECCV, 2022. 3 +[55] Xingyu Liu, Yingyue Li, Chengxi Li, Gu Wang, Chenyangguang Zhang, Ziqin Huang, and Xiangyang Ji. Gfreedet: Exploiting gaussian splatting and foundation models for model-free unseen object detection in the bop challenge 2024. arXiv preprint arXiv:2412.01552, 2024. 1 +[56] Xingyu Liu, Chenyangguang Zhang, Gu Wang, Ruida Zhang, and Xiangyang Ji. Rasim: A range-aware high-fidelity rgb-d data simulation pipeline for real-world applications. In ICRA, pages 17057-17064. IEEE, 2024. 1, 3 + +[57] Yuan Liu, Yilin Wen, Sida Peng, Cheng Lin, Xiaoxiao Long, Taku Komura, and Wenping Wang. Gen6d: Generalizable model-free 6-dof object pose estimation from rgb images. In ECCV, 2022. 1, 2 +[58] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In *ICLR*, 2017. 6 +[59] Eric Marchand, Hideaki Uchiyama, and Fabien Spindler. Pose estimation for augmented reality: a hands-on survey. IEEE TVCG, 22(12):2633-2651, 2015. 1 +[60] Van Nguyen Nguyen, Yinlin Hu, Yang Xiao, Mathieu Salzmann, and Vincent Lepetit. Templates for 3d object pose estimation revisited: Generalization to new objects and robustness to occlusions. In CVPR, pages 6771-6780, 2022. 3 +[61] Van Nguyen Nguyen, Thibault Groueix, Georgy Ponimatkin, Vincent Lepetit, and Tomas Hodan. Cnos: A strong baseline for cad-based novel object segmentation. In ICCV, pages 2134-2140, 2023. 1, 3, 4, 6, 7, 8 +[62] Van Nguyen Nguyen, Thibault Groueix, Mathieu Salzmann, and Vincent Lepetit. Gigapose: Fast and robust novel object pose estimation via one correspondence. In CVPR, 2024. 1, 3 +[63] Yinyu Nie, Xiaoguang Han, Shihui Guo, Yujuan Zheng, Jian Chang, and Jian Jun Zhang. Total3DUnderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In CVPR, pages 55-64, 2020. 1 +[64] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 6 +[65] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. TMLR, 2024. 2, 3, 4, 5, 6 +[66] Evin Pinar Örnek, Yann Labbé, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, and Tomas Hodan. Found- pose: Unseen object pose estimation with foundation features. In CVPR, 2024. 3 +[67] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. PyTorch: An imperative style, high-performance deep learning library. In NeurlPS, pages 8026-8037, 2019. 6 +[68] Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, and Hujun Bao. Pynet: Pixel-wise voting network for 6dof pose estimation. In CVPR, pages 4561-4570, 2019. 1, 3 +[69] Fabio Poiesi and Davide Boscaini. Learning general and distinctive 3d local deep descriptors for point cloud registration. IEEE TPAMI, 45(3):3979-3985, 2022. 3, 6, 7 +[70] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. PointNet: Deep learning on point sets for 3D classification and segmentation. CVPR, 1(2):4, 2017. 5 + +[71] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, Slobodan Ilic, Dewen Hu, and Kai Xu. Geotransformer: Fast and robust point cloud registration with geometric transformer. IEEE TPAMI, 2023. 3, 5, 6 +[72] Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (fpfh) for 3d registration. In ICRA, pages 3212-3217. IEEE, 2009. 3, 6, 7 +[73] Ivan Shugurov, Sergey Zakharov, and Slobodan Ilic. Dpodv2: Dense correspondence-based 6 dof pose estimation. IEEE TPAMI, 44(11):7417-7435, 2021. 3 +[74] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. Osop: A multi-stage one shot object pose estimation framework. In CVPR, pages 6835-6844, 2022. 3 +[75] Fridtjof Stein, Gérard Medioni, et al. Structural indexing: Efficient 3-d object recognition. IEEE TPAMI, 14(2):125-145, 1992. 5 +[76] Yongzhi Su, Jason Rambach, Nareg Minaskan, Paul Lesur, Alain Pagani, and Didier Stricker. Deep multi-state object pose estimation for augmented reality assembly. In IEEE International Symposium on Mixed and Augmented Reality Adjunct (ISMAR-Adjunct), pages 222–227, 2019. 1 +[77] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. Onepose: One-shot object pose estimation without cad models. In CVPR, pages 6825-6834, 2022. 1, 2, 3 +[78] Jingyi Tang, Gu Wang, Zeyu Chen, Shengquan Li, Xiu Li, and Xiangyang Ji. Fafa: Frequency-aware flow-aided self-supervision for underwater object pose estimation. In ECCV, pages 352–369. Springer, 2024. 1 +[79] Meng Tian, Marcelo H Ang Jr, and Gim Hee Lee. Shape prior deformation for categorical 6D object pose and size estimation. In ECCV, 2020. 1, 3 +[80] Jonathan Tremblay, Thang To, Balakumar Sundaralingam, Yu Xiang, Dieter Fox, and Stan Birchfield. Deep object pose estimation for semantic robotic grasping of household objects. In CoRL, pages 306-316, 2018. 1 +[81] Mikaela Angelina Uy, Vladimir G Kim, Minhyuk Sung, Noam Aigerman, Siddhartha Chaudhuri, and Leonidas J Guibas. Joint learning of 3d shape retrieval and deformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11713-11722, 2021. 3 +[82] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 6 +[83] Chen Wang, Danfei Xu, Yuke Zhu, Roberto Martin-Martin, Cewu Lu, Li Fei-Fei, and Silvio Savarese. DenseFusion: 6D object pose estimation by iterative dense fusion. In CVPR, pages 3343–3352, 2019. 1, 3 +[84] Gu Wang, Fabian Manhardt, Jianzhun Shao, Xiangyang Ji, Nassir Navab, and Federico Tombari. Self6d: Self-supervised monocular 6d object pose estimation. In ECCV, pages 108-125. Springer, 2020. +[85] Gu Wang, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. GDR-Net: Geometry-guided direct regression network for monocular 6D object pose estimation. In CVPR, pages 16611–16621, 2021. 3 + +[94] Yang You, Ruoxi Shi, Weiming Wang, and Cewu Lu. CPPf: Towards robust category-level 9d pose estimation in the wild. In CVPR, pages 6866-6875, 2022. 5 + +[95] Hao Yu, Fu Li, Mahdi Saleh, Benjamin Busam, and Slobodan Ilic. Cofinet: Reliable coarse-to-fine correspondences for robust pointcloud registration. NeurIPS, 34, 2021. 8 + +[96] Hao Yu, Zheng Qin, Ji Hou, Mahdi Saleh, Dongsheng Li, Benjamin Busam, and Slobodan Ilic. Rotation-invariant transformer for point cloud matching. In CVPR, pages 5384-5393, 2023. 5 + +[97] Hao Yu, Ji Hou, Zheng Qin, Mahdi Saleh, Ivan Shugurov, Kai Wang, Benjamin Busam, and Slobodan Ilic. Riga: Rotation-invariant and globally-aware descriptors for point cloud registration. IEEE TPAMI, 46(5):3796-3812, 2024. 3 + +[98] Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose: Predicting probabilistic relative rotation for single objects in the wild. In ECCV, pages 592-611. Springer, 2022. 2, 3 + +[99] Ruida Zhang, Yan Di, Zhiqiang Lou, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. Rbp-posed: Residual bounding box projection for category-level pose estimation. In European conference on computer vision, pages 655–672. Springer, 2022. 3 + +[100] Ruida Zhang, Yan Di, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. Ssp-pose: Symmetry-aware shape prior deformation for direct category-level object pose estimation. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7452–7459. IEEE, 2022. + +[86] Gu Wang, Fabian Manhardt, Xingyu Liu, Xiangyang Ji, and Federico Tombari. Occlusion-aware self-supervised monocular 6D object pose estimation. IEEE TPAMI, 2024. 1 + +[87] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6D object pose and size estimation. In CVPR, pages 2642-2651, 2019. 1, 3 +[88] Jiaze Wang, Kai Chen, and Qi Dou. Category-level 6D object pose estimation via cascaded relation and recurrent reconstruction networks. In IROS, 2021. 1 +[89] Yue Wang and Justin M Solomon. Deep closest point: Learning representations for point cloud registration. In ICCV, pages 3523-3532, 2019. 3 +[90] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. FoundationPose: Unified 6d pose estimation and tracking of novel objects. In CVPR, 2024. 1, 2, 3 +[91] Chin-Hsuan Wu, Yen-Chun Chen, Bolivar Solarte, Lu Yuan, and Min Sun. ifusion: Inverting diffusion for pose-free reconstruction from sparse views. arXiv preprint arXiv:2312.17250, 2023. 3 +[92] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6D object pose estimation in cluttered scenes. RSS, 2018. 2, 6 +[93] Jiaqi Yang, Qian Zhang, Yang Xiao, and Zhiguo Cao. Toldi: An effective and robust approach for 3d local shape description. PR, 65:175-187, 2017. 3, 5 +[101] Ruida Zhang, Ziqin Huang, Gu Wang, Chenyangguang Zhang, Yan Di, Xingxing Zuo, Jiwen Tang, and Xiangyang Ji. Lapose: Laplacian mixture shape modeling for rgb-based category-level object pose estimation. In ECCV, 2024. 3 +[102] Ruida Zhang, Chenyangguang Zhang, Yan Di, Fabian Manhardt, Xingyu Liu, Federico Tombari, and Xiangyang Ji. Kp-red: Exploiting semantic keypoints for joint 3d shape retrieval and deformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20540-20550, 2024. 3 +[103] Xiyu Zhang, Jiaqi Yang, Shikun Zhang, and Yanning Zhang. 3d registration with maximal cliques. In CVPR, pages 17745–17754, 2023. 6, 7 +[104] Chen Zhao, Yinlin Hu, and Mathieu Salzmann. Fusing local similarities for retrieval-based 3d orientation estimation of unseen objects. In ECCV, pages 106-122. Springer, 2022. 3 +[105] Chen Zhao, Tong Zhang, and Mathieu Salzmann. 3d-aware hypothesis & verification for generalizable relative object pose estimation. In ICLR, 2023. 2, 3 +[106] Chen Zhao, Tong Zhang, Zheng Dang, and Mathieu Salzmann. Dvmnet: Computing relative pose for unseen objects beyond hypotheses. In CVPR, 2024. 2, 3 +[107] Xu Zhao, Wenchao Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, and Jinqiao Wang. Fast segment anything. arXiv preprint arXiv:2306.12156, 2023. 6, 7 +[108] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splattering: Fast and generalizable single-view 3d reconstruction with transformers. In CVPR, pages 10324-10335, 2024. 8 \ No newline at end of file diff --git a/2025/UNOPose_ Unseen Object Pose Estimation with an Unposed RGB-D Reference Image/images.zip b/2025/UNOPose_ Unseen Object Pose Estimation with an Unposed RGB-D Reference Image/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..854009264bf533ff69761805abbcf8f9b7415064 --- /dev/null +++ b/2025/UNOPose_ Unseen Object Pose Estimation with an Unposed RGB-D Reference Image/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0438716638cc3d41e476ed9f3bd50f5b67e4dd1bdd6ed7b0445f1c9b92366c5e +size 603917 diff --git a/2025/UNOPose_ Unseen Object Pose Estimation with an Unposed RGB-D Reference Image/layout.json b/2025/UNOPose_ Unseen Object Pose Estimation with an Unposed RGB-D Reference Image/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ba0e3f679fe217ef4edb59c944002fdcf04f93e1 --- /dev/null +++ b/2025/UNOPose_ Unseen Object Pose Estimation with an Unposed RGB-D Reference Image/layout.json @@ -0,0 +1,11257 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 103, + 452, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 103, + 452, + 139 + ], + "spans": [ + { + "bbox": [ + 159, + 103, + 452, + 139 + ], + "type": "text", + "content": "UNOpose: Unseen Object Pose Estimation with an Unposed RGB-D Reference Image" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 160, + 485, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 160, + 485, + 203 + ], + "spans": [ + { + "bbox": [ + 126, + 160, + 485, + 203 + ], + "type": "text", + "content": "Xingyu Liu" + }, + { + "bbox": [ + 126, + 160, + 485, + 203 + ], + "type": "inline_equation", + "content": "^{1, *}" + }, + { + "bbox": [ + 126, + 160, + 485, + 203 + ], + "type": "text", + "content": ", Gu Wang" + }, + { + "bbox": [ + 126, + 160, + 485, + 203 + ], + "type": "inline_equation", + "content": "^{1, *}" + }, + { + "bbox": [ + 126, + 160, + 485, + 203 + ], + "type": "text", + "content": ", Ruida Zhang" + }, + { + "bbox": [ + 126, + 160, + 485, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 126, + 160, + 485, + 203 + ], + "type": "text", + "content": ", Chenyangguang Zhang" + }, + { + "bbox": [ + 126, + 160, + 485, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 126, + 160, + 485, + 203 + ], + "type": "text", + "content": ", Federico Tombari" + }, + { + "bbox": [ + 126, + 160, + 485, + 203 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 126, + 160, + 485, + 203 + ], + "type": "text", + "content": ", and Xiangyang Ji" + }, + { + "bbox": [ + 126, + 160, + 485, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 126, + 160, + 485, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 126, + 160, + 485, + 203 + ], + "type": "text", + "content": "Tsinghua University, BNRist " + }, + { + "bbox": [ + 126, + 160, + 485, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 126, + 160, + 485, + 203 + ], + "type": "text", + "content": "Technical University of Munich, " + }, + { + "bbox": [ + 126, + 160, + 485, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 126, + 160, + 485, + 203 + ], + "type": "text", + "content": "Google" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 83, + 205, + 525, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 205, + 525, + 217 + ], + "spans": [ + { + "bbox": [ + 83, + 205, + 525, + 217 + ], + "type": "text", + "content": "{liuxy21@mails., wanggu1@, zhangrd23@mails., zcyg22@mails., xyji@}tsinghua.edu.cn," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 257, + 220, + 350, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 220, + 350, + 230 + ], + "spans": [ + { + "bbox": [ + 257, + 220, + 350, + 230 + ], + "type": "text", + "content": "tombari@in.tum.de" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 283, + 297, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 297, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 297, + 571 + ], + "type": "text", + "content": "Unseen object pose estimation methods often rely on CAD models or multiple reference views, making the onboarding stage costly. To simplify reference acquisition, we aim to estimate the unseen object's pose through a single unposed RGB-D reference image. While previous works leverage reference images as pose anchors to limit the range of relative pose, our scenario presents significant challenges since the relative transformation could vary across the entire " + }, + { + "bbox": [ + 55, + 283, + 297, + 571 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 55, + 283, + 297, + 571 + ], + "type": "text", + "content": " space. Moreover, factors like occlusion, sensor noise, and extreme geometry could result in low viewpoint overlap. To address these challenges, we present a novel approach and benchmark, termed UNOPose1, for UNseen One-reference-based object Pose estimation. Building upon a coarse-to-fine paradigm, UNOPose constructs an " + }, + { + "bbox": [ + 55, + 283, + 297, + 571 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 55, + 283, + 297, + 571 + ], + "type": "text", + "content": "-invariant reference frame to standardize object representation despite pose and size variations. To alleviate small overlap across viewpoints, we recalibrate the weight of each correspondence based on its predicted likelihood of being within the overlapping region. Evaluated on our proposed benchmark based on the BOP Challenge, UNOPose demonstrates superior performance, significantly outperforming traditional and learning-based methods in the one-reference setting and remaining competitive with CAD-model-based methods. The code and dataset are available at github.com/shanice-l/UNOPose." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 135, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 135, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 135, + 605 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 296, + 685 + ], + "type": "text", + "content": "Localizing an object in Euclidean space by estimating its 6DoF pose, i.e., 3DoF orientation and 3DoF position, plays a crucial role in augmented/virtual reality [59, 76], scene understanding [37, 63] and robotic manipulation [23, 26, 80]. The vast majority of works [42, 48, 68, 78, 83-86] focuses on instance-level object pose estimation, where the train" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 259, + 556, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 556, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 556, + 390 + ], + "type": "text", + "content": "ing and testing datasets consist of an identical set of known object instances, and the CAD models of objects are often required for generating training images and labels [16, 56]. More recently, Wang et al. [87] and its successors [6, 79, 88] extended this paradigm to the category level, intending to estimate poses for novel instances within predefined categories without requiring CAD models of target objects. Both paradigms have limitations in open-world applications, where annotating and training for new objects beyond known categories would be very labor-intensive and sometimes prohibitive." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 396, + 556, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 556, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 556, + 613 + ], + "type": "text", + "content": "To mitigate this problem, recent research efforts [3, 43, 51, 90] have shifted towards pose estimation of arbitrary novel objects, where the target object is unseen during training. This task presents a great challenge due to the inherent uncertainty of the novel object's canonical frame (defined in the CAD model of the object). Current approaches [43, 57, 62] often tackle this by using omnidirectional reference views to cover the target object, separating pose estimation task to viewpoint selection and relative pose estimation to known pose anchors. If CAD models are available during the test phase, reference views can be easily produced through image rendering [43, 51, 61, 90]. Alternatively, without CAD models, reference views can be obtained by capturing multiple images of the object [29, 31, 55, 77]. Both setups have shortcomings, as creating a CAD model of an object or labeling a substantial number of object poses is limited in scalability. For example, they can hardly adapt to in-the-wild scenarios where novel objects are introduced unpredictably." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "content": "To sum up, while existing methods exhibit promising transfer capability to novel objects, they are notably constrained by their reliance on the CAD model or multiple reference views (Fig. 1 (a, b)). In this work, we aim to minimize onboarding cost by estimating novel object poses using a single unposed RGB-D reference image (Fig. 1). This setup is different from existing methods mainly in two aspects. First, we focus on estimating the relative pose between" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 693, + 137, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 693, + 137, + 703 + ], + "spans": [ + { + "bbox": [ + 67, + 693, + 137, + 703 + ], + "type": "text", + "content": "*Equal contributions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 70, + 703, + 237, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 237, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 237, + 713 + ], + "type": "text", + "content": "1UNO " + }, + { + "bbox": [ + 70, + 703, + 237, + 713 + ], + "type": "inline_equation", + "content": "(\\prime \\mathrm{u}:\\mathrm{no}\\overline{\\mathrm{u}} /" + }, + { + "bbox": [ + 70, + 703, + 237, + 713 + ], + "type": "text", + "content": " ) means one in Spanish and Italian." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22023" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 553, + 277 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 277 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 277 + ], + "type": "image", + "image_path": "ea83ed8cb74d85b90fc84596b8b7beecb05bd2252b35bf5958d3239b5ca762a9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 286, + 555, + 321 + ], + "lines": [ + { + "bbox": [ + 54, + 286, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 54, + 286, + 555, + 321 + ], + "type": "text", + "content": "Figure 1. Illustration of unseen object pose estimation. Given a query image presenting a target object unseen during training, we aim to estimate its segmentation and 6DoF pose w.r.t. a reference frame. While previous methods [43, 57, 77, 90] often rely on the CAD model or multiple RGB(-D) images for reference, we merely use one unposed RGB-D reference image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 327, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 327, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 54, + 327, + 296, + 460 + ], + "type": "text", + "content": "two unposed viewpoints rather than the absolute pose. Note that the absolute pose of a novel object becomes ill-posed without a certain canonical frame, while relative pose estimation remains well-defined regardless of the canonical frame's definition. Secondly, some existing relative pose estimation methods [98, 105, 106] merely use RGB modality to predict the 3DoF rotation, limiting their ability to predict relative translation and generalize beyond their training datasets. In contrast, we fully leverage RGB-D modality to predict the 6DoF relative pose, meanwhile enhancing the network's generalizability to novel objects and environments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 462, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 297, + 715 + ], + "type": "text", + "content": "Nevertheless, estimating the pose of unseen objects with only one reference view presents significant challenges from various aspects. First, previous methods [3, 43] select the most similar reference view as a pose anchor, distinctly reducing the search space for relative poses. However, in our scenario, where both the target and reference object poses are unknown, the relative pose can vary across the entire " + }, + { + "bbox": [ + 55, + 462, + 297, + 715 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 55, + 462, + 297, + 715 + ], + "type": "text", + "content": " space. Moreover, in partial-to-partial object matching, factors such as occlusion, sensor noise, and extreme geometry can severely interfere with the matching process. To address these challenges, we propose a novel approach and benchmark for UNseen One-reference-based object Pose estimation (UNOPose). Our pipeline leverages the strong generalization capabilities of vision foundation models [40, 65] to produce an effective segmentation of the unseen object. Then, UNOPose employs a coarse-to-fine paradigm for estimating the relative pose between the reference and the query objects. To alleviate the challenge of diverse pose and size variations, we introduce the " + }, + { + "bbox": [ + 55, + 462, + 297, + 715 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 55, + 462, + 297, + 715 + ], + "type": "text", + "content": "-invariant global reference frame (GRF) to standardize object representation. Subsequently, a hierarchical object encoding" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 327, + 555, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 387 + ], + "type": "text", + "content": "paradigm based on the local reference frame (LRF) further captures fine-grained geometric details. For achieving reliable correspondences in partial-to-partial object matching, we harness an overlap predictor to identify and concentrate on the overlapping region." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 387, + 556, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 556, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 556, + 506 + ], + "type": "text", + "content": "Moreover, we propose a new benchmark devised from the BOP challenge [34] to facilitate evaluation and future research of unseen object pose estimation with one reference. Extensive experiments on YCB-V [92], LM-O [2] and TUD-L [32] datasets demonstrate that our UNOPose surpasses all compared methods on a single reference setting. To our surprise, UNOPose with an unposed reference is even on par with some " + }, + { + "bbox": [ + 313, + 387, + 556, + 506 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 313, + 387, + 556, + 506 + ], + "type": "text", + "content": "-invariant-feature-based methods relying on CAD models (Ours " + }, + { + "bbox": [ + 313, + 387, + 556, + 506 + ], + "type": "inline_equation", + "content": "70.9\\%" + }, + { + "bbox": [ + 313, + 387, + 556, + 506 + ], + "type": "text", + "content": " vs. ZTE-PPF [1] " + }, + { + "bbox": [ + 313, + 387, + 556, + 506 + ], + "type": "inline_equation", + "content": "69.0\\%" + }, + { + "bbox": [ + 313, + 387, + 556, + 506 + ], + "type": "text", + "content": " vs. Koenig-PPF [41] " + }, + { + "bbox": [ + 313, + 387, + 556, + 506 + ], + "type": "inline_equation", + "content": "75.1\\%" + }, + { + "bbox": [ + 313, + 387, + 556, + 506 + ], + "type": "text", + "content": " w.r.t. " + }, + { + "bbox": [ + 313, + 387, + 556, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{AR}_{\\mathrm{BOP}}" + }, + { + "bbox": [ + 313, + 387, + 556, + 506 + ], + "type": "text", + "content": " metric)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 326, + 507, + 525, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 507, + 525, + 518 + ], + "spans": [ + { + "bbox": [ + 326, + 507, + 525, + 518 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 326, + 520, + 556, + 712 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 326, + 520, + 553, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 520, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 326, + 520, + 553, + 555 + ], + "type": "text", + "content": "- To the best of our knowledge, we are the first to conduct unseen object 6DoF pose estimation leveraging a single unposed RGB-D reference." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 326, + 555, + 555, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 555, + 555, + 614 + ], + "spans": [ + { + "bbox": [ + 326, + 555, + 555, + 614 + ], + "type": "text", + "content": "- Based on the BOP Challenge, we devise a new extensive benchmark tailored for unseen object segmentation and pose estimation with one reference. Additionally, we evaluate several traditional and learning-based methods on this benchmark for completeness." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 326, + 616, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 616, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 326, + 616, + 556, + 712 + ], + "type": "text", + "content": "- We introduce UNOPose, a network for learning relative transformation between reference and query objects. To achieve this, we propose the " + }, + { + "bbox": [ + 326, + 616, + 556, + 712 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 326, + 616, + 556, + 712 + ], + "type": "text", + "content": "-invariant global and local reference frames, enabling standardized object representations despite variations in pose and size. Furthermore, the network can automatically adjust the confidence of each correspondence by incorporating an overlap predictor." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22024" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 91, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 57, + 91, + 296, + 330 + ], + "type": "text", + "content": "Class-specific Pose Estimation. Class-specific object pose estimation aims at predicting 6DoF poses of either instance of a known object (instance-level) or unseen objects within a known category (category-level). Instance-level object poses are often solved by direct regression [35, 47, 56, 85] using deep neural networks, or by establishing 2D-3D [27, 48, 68] or 3D-3D correspondences [30, 73, 83] which are then leveraged by RANSAC-based PnP/Kabsch algorithms. The instance-level setting requires expensive data generation/annotation [18, 19, 81, 102] and training for every new object. To address this limitation, category-level pose estimation [87] is proposed to estimate 9DoF poses of novel objects among specific categories without CAD models. Mainstream approaches can also be categorized into direct regression [6, 9, 10, 17, 50, 54, 99–101] or correspondence-based [8, 24, 79, 87] methods. However, this setting remains constrained to a limited number of categories, given the additional challenge of aligning canonical frames, managing symmetric objects, and categorizing objects in similar categories correctly (e.g. can and bottle)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 331, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 331, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 57, + 331, + 296, + 594 + ], + "type": "text", + "content": "Novel Object Pose Estimation. Estimating the pose of novel objects beyond known classes, i.e., objects at inference time are unseen during training, is a useful yet challenging task. Existing works solve this problem through image-based matching [3, 43, 60, 61, 74, 90, 104] or feature-based matching approaches [4, 29, 36, 51, 62, 66, 77]. Notably, given a target image, OVE6D [3] and MegaPose [43] retrieve the most similar viewpoint from a pre-rendered omnidirectional image database as a coarse pose estimate. A customized neural network then refines the coarse pose estimate. FoundationPose [90] further builds a pose ranking network to score each refined pose hypothesis. Feature-matching methods learn local feature descriptors to construct pixel-level or point-level correspondences. For example, OnePose [77] and OnePose++ [29] reconstruct the 3D point cloud of an unseen object to establish 2D-3D correspondences, whereas SAM-6D [51] builds 3D-3D correspondences. More recently, some methods [4, 66] directly exploit the power of foundation models for zero-shot pose estimation. Exemplarily, FoundPose [66] extracts DINOv2 visual features [65], while FreeZe [4] further utilizes generalizable geometric features [69] to conduct feature matching." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Relative Pose Estimation. Orthogonal to previous works which rely on known instances or multiple reference views, relative object pose estimation estimates a relative transformation of the unseen object with only one reference image. One closely related topic is camera pose estimation from two views. For example, RelPose [98] and RelPose++ [49] infer a distribution of relative rotations by leveraging an energy-based formulation. Moreover, iFusion [91] optimizes the unknown relative pose by inverting the novel view synthesis diffusion model [53]. Relative camera pose estimation relies" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 555, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 191 + ], + "type": "text", + "content": "on static background to establish correspondence between viewpoints. In the field of relative object pose estimation, given two RGB images, 3DAHV [105] presents a hypothesis-and-verification framework to score each relative pose hypothesis. Extended by that, DVMNet [106] introduces a hypothesis-free pipeline to compute relative poses via deep voxel matching. The above works have limited capability in inferring 3DoF relative translations. In this work, with the introduction of depth data, we can estimate a complete 6DoF pose including rotation and translation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 192, + 556, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 556, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 556, + 371 + ], + "type": "text", + "content": "Point Cloud Registration. Our approach follows the paradigm of correspondence-based methods [14, 15, 45, 89], which initially extract 3D local feature descriptors to establish correspondences between source and target point clouds. Subsequently, they estimate the relative transformation using techniques like SVD, RANSAC, or Hough Voting [44]. Some feature descriptors are proposed to ensure rotational invariance. For example, FPFH [72] descriptor is computed with geometric surface properties. PPF [22] uses the distance and angles to describe the relation of point pairs. TOLDI [93] proposes a robust local reference frame using keypoint normals and neighboring projections. With the advent of deep learning, neural networks are increasingly employed to extract 3D local or global feature descriptors [12, 38, 71, 89, 97]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 382, + 556, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 556, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 556, + 409 + ], + "type": "text", + "content": "3. UNO Object Segmentation and Pose Estimation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 416, + 440, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 440, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 440, + 428 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 435, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 555, + 495 + ], + "type": "text", + "content": "Assuming an arbitrary unseen rigid object in a query image, our goal is to estimate the object's mask " + }, + { + "bbox": [ + 313, + 435, + 555, + 495 + ], + "type": "inline_equation", + "content": "M_q" + }, + { + "bbox": [ + 313, + 435, + 555, + 495 + ], + "type": "text", + "content": " and its 6DoF relative pose " + }, + { + "bbox": [ + 313, + 435, + 555, + 495 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{T} \\in SE(3)" + }, + { + "bbox": [ + 313, + 435, + 555, + 495 + ], + "type": "text", + "content": " with a masked RGB-D reference image exhibiting the target object without major occlusion or truncation. As illustrated in Fig. 1, the input is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 495, + 555, + 566 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 315, + 495, + 522, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 495, + 522, + 507 + ], + "spans": [ + { + "bbox": [ + 315, + 495, + 522, + 507 + ], + "type": "text", + "content": "1) " + }, + { + "bbox": [ + 315, + 495, + 522, + 507 + ], + "type": "inline_equation", + "content": "[I_q|D_q]\\in \\mathbb{R}^{H\\times W\\times 4}" + }, + { + "bbox": [ + 315, + 495, + 522, + 507 + ], + "type": "text", + "content": ": The query RGB-D image;" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 506, + 555, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 506, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 315, + 506, + 555, + 542 + ], + "type": "text", + "content": "2) " + }, + { + "bbox": [ + 315, + 506, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\left[\\bar{I}_p|D_p\\right] \\in \\mathbb{R}^{H \\times W \\times 4}" + }, + { + "bbox": [ + 315, + 506, + 555, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 315, + 506, + 555, + 542 + ], + "type": "inline_equation", + "content": "M_p \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 315, + 506, + 555, + 542 + ], + "type": "text", + "content": ": The reference RGB-D image and the corresponding binary mask indicating the target object;" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 543, + 553, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 543, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 315, + 543, + 553, + 566 + ], + "type": "text", + "content": "3) " + }, + { + "bbox": [ + 315, + 543, + 553, + 566 + ], + "type": "inline_equation", + "content": "K_{q} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 315, + 543, + 553, + 566 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 315, + 543, + 553, + 566 + ], + "type": "inline_equation", + "content": "K_{p} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 315, + 543, + 553, + 566 + ], + "type": "text", + "content": ": Camera intrinsics of the query and reference images." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 567, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 555, + 662 + ], + "type": "text", + "content": "Optionally, if the pose " + }, + { + "bbox": [ + 313, + 567, + 555, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_p \\in SE(3)" + }, + { + "bbox": [ + 313, + 567, + 555, + 662 + ], + "type": "text", + "content": " of the reference object in the camera frame is known, the query object pose " + }, + { + "bbox": [ + 313, + 567, + 555, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_q \\in SE(3)" + }, + { + "bbox": [ + 313, + 567, + 555, + 662 + ], + "type": "text", + "content": " can be recovered by " + }, + { + "bbox": [ + 313, + 567, + 555, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_q = \\Delta \\mathbf{T} \\mathbf{T}_p" + }, + { + "bbox": [ + 313, + 567, + 555, + 662 + ], + "type": "text", + "content": ". Note that, for practicality, the method should not rely on the absolute pose of the reference object, since the world frame can vary arbitrarily in different applications. We only use the reference object pose during the inference stage for evaluation on standard object pose datasets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 670, + 462, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 462, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 462, + 684 + ], + "type": "text", + "content": "3.2. UNO Object Segmentation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Firstly, we segment the query object from a cluttered background. Thanks to the great generalization ability of vision" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22025" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 337 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 337 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 337 + ], + "type": "image", + "image_path": "3004c9497e8ca6b6f792ef25e37a1f4b4f1722e393b7e280f65fc4b943bf5754.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 344, + 555, + 390 + ], + "lines": [ + { + "bbox": [ + 55, + 344, + 555, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 555, + 390 + ], + "type": "text", + "content": "Figure 2. The network architecture of UNOpose. Given the query and reference point clouds " + }, + { + "bbox": [ + 55, + 344, + 555, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{cam}" + }, + { + "bbox": [ + 55, + 344, + 555, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 344, + 555, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{cam}" + }, + { + "bbox": [ + 55, + 344, + 555, + 390 + ], + "type": "text", + "content": " in the camera frame, UNOpose first transforms them into the " + }, + { + "bbox": [ + 55, + 344, + 555, + 390 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 55, + 344, + 555, + 390 + ], + "type": "text", + "content": "-invariant global reference frame (GRF). Then feature descriptors are extracted from sparse point sets for constructing the coarse correlation matrix. For achieving precise correspondences, the fine pose estimation module exploits structural details using positional encoding and local reference frame (LRF) encoding." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 401, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 296, + 534 + ], + "type": "text", + "content": "foundation models, methods like [7, 51, 61] can effectively segment novel objects with CAD models. However, unlike previous approaches, which generate diverse descriptors from multiple rendered views, we can access only a single reference image. To address this challenge, we use the SAM model [40] to predict all possible mask proposals from the query image, and then score each mask proposal by comparing DINOv2 [65] descriptors between the query and reference images using cosine similarity, identifying the most similar mask " + }, + { + "bbox": [ + 55, + 401, + 296, + 534 + ], + "type": "inline_equation", + "content": "M_{q}" + }, + { + "bbox": [ + 55, + 401, + 296, + 534 + ], + "type": "text", + "content": ". Please refer to the appendix for more details about UNO segmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 541, + 216, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 541, + 216, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 541, + 216, + 555 + ], + "type": "text", + "content": "3.3. UNO Object Pose Estimation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 560, + 185, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 185, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 185, + 572 + ], + "type": "text", + "content": "3.3.1 Overview of UNOPose" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 579, + 296, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 579, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 55, + 579, + 296, + 679 + ], + "type": "text", + "content": "Given the predicted mask " + }, + { + "bbox": [ + 55, + 579, + 296, + 679 + ], + "type": "inline_equation", + "content": "M_{q}" + }, + { + "bbox": [ + 55, + 579, + 296, + 679 + ], + "type": "text", + "content": " of the query image and mask " + }, + { + "bbox": [ + 55, + 579, + 296, + 679 + ], + "type": "inline_equation", + "content": "M_{p}" + }, + { + "bbox": [ + 55, + 579, + 296, + 679 + ], + "type": "text", + "content": " of the reference image, we crop and back-project the object of interest from depth maps " + }, + { + "bbox": [ + 55, + 579, + 296, + 679 + ], + "type": "inline_equation", + "content": "D_{q}, D_{p}" + }, + { + "bbox": [ + 55, + 579, + 296, + 679 + ], + "type": "text", + "content": " into the camera space as two point sets " + }, + { + "bbox": [ + 55, + 579, + 296, + 679 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{cam} \\in \\mathbb{R}^{N^{Q} \\times 3}" + }, + { + "bbox": [ + 55, + 579, + 296, + 679 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 579, + 296, + 679 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{cam} \\in \\mathbb{R}^{N^{P} \\times 3}" + }, + { + "bbox": [ + 55, + 579, + 296, + 679 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 579, + 296, + 679 + ], + "type": "inline_equation", + "content": "N^{Q}" + }, + { + "bbox": [ + 55, + 579, + 296, + 679 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 579, + 296, + 679 + ], + "type": "inline_equation", + "content": "N^{P}" + }, + { + "bbox": [ + 55, + 579, + 296, + 679 + ], + "type": "text", + "content": " denotes the point numbers of query and reference point clouds, respectively. Our goal is to recover the relative transformation " + }, + { + "bbox": [ + 55, + 579, + 296, + 679 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{T} = \\{\\Delta \\mathbf{R}, \\Delta \\mathbf{t}\\}" + }, + { + "bbox": [ + 55, + 579, + 296, + 679 + ], + "type": "text", + "content": " by minimizing the correspondence distance" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 689, + 296, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 689, + 296, + 716 + ], + "spans": [ + { + "bbox": [ + 104, + 689, + 296, + 716 + ], + "type": "interline_equation", + "content": "\\min \\sum_ {(\\mathbf {q}, \\mathbf {p}) \\in \\mathbf {C}} \\| \\Delta \\mathbf {R} \\mathbf {q} + \\Delta \\mathbf {t} - \\mathbf {p} \\| _ {2}, \\tag {1}", + "image_path": "0237807c32f766e69c0803bb93388230eea2af12cb38e71daf15a11039654e01.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 401, + 555, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 555, + 462 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 401, + 555, + 462 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 313, + 401, + 555, + 462 + ], + "type": "text", + "content": " is the predicted correspondence set between " + }, + { + "bbox": [ + 313, + 401, + 555, + 462 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{cam}" + }, + { + "bbox": [ + 313, + 401, + 555, + 462 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 401, + 555, + 462 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{cam}" + }, + { + "bbox": [ + 313, + 401, + 555, + 462 + ], + "type": "text", + "content": ". We exploit both color and geometric cues to build this correspondence and follow the broadly used coarse-to-fine paradigm in point cloud registration to solve Eq. (1). The network is demonstrated in Fig. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 475, + 479, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 479, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 479, + 487 + ], + "type": "text", + "content": "3.3.2 Coarse-to-fine Pose Estimation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 494, + 555, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 555, + 566 + ], + "type": "text", + "content": "Constructing a Pose-invariant Reference Frame. Given only an unposed reference image, the relative pose to predict is arbitrary in the " + }, + { + "bbox": [ + 313, + 494, + 555, + 566 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 313, + 494, + 555, + 566 + ], + "type": "text", + "content": " space, which renders a significant challenge for achieving robust correspondences. Hence, we introduce a pose-invariant global reference frame (GRF) and transform " + }, + { + "bbox": [ + 313, + 494, + 555, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{cam}" + }, + { + "bbox": [ + 313, + 494, + 555, + 566 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 494, + 555, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{cam}" + }, + { + "bbox": [ + 313, + 494, + 555, + 566 + ], + "type": "text", + "content": " into GRF as " + }, + { + "bbox": [ + 313, + 494, + 555, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_G" + }, + { + "bbox": [ + 313, + 494, + 555, + 566 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 494, + 555, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_G" + }, + { + "bbox": [ + 313, + 494, + 555, + 566 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 566, + 555, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 566, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 555, + 602 + ], + "type": "text", + "content": "Concretely, taking " + }, + { + "bbox": [ + 313, + 566, + 555, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{cam}" + }, + { + "bbox": [ + 313, + 566, + 555, + 602 + ], + "type": "text", + "content": " as an example, transforming the point cloud to GRF involves a 7DoF coordinate transformation " + }, + { + "bbox": [ + 313, + 566, + 555, + 602 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{R}_G\\in SO(3),\\mathbf{t}_G\\in \\mathbb{R}^3,s_G\\in \\mathbb{R}\\}" + }, + { + "bbox": [ + 313, + 566, + 555, + 602 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 353, + 605, + 555, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 605, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 353, + 605, + 555, + 622 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {G} = \\left\\{\\mathbf {R} _ {G} ^ {\\top} \\left(\\mathbf {q} - \\mathbf {t} _ {G}\\right) / s _ {G} \\mid \\mathbf {q} \\in \\mathbf {Q} _ {c a m} \\right\\}. \\tag {2}", + "image_path": "5e9ae25cc52db8c059030e5bfee5d9ebcf854c59964a2910b9015e9a44029d25.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 625, + 555, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 625, + 555, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 625, + 555, + 661 + ], + "type": "text", + "content": "The origin of GRF is located at the object center " + }, + { + "bbox": [ + 313, + 625, + 555, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_Q" + }, + { + "bbox": [ + 313, + 625, + 555, + 661 + ], + "type": "text", + "content": " for translation invariance, and the radius of the point cloud is rescaled to 1 for size invariance, computed as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 354, + 666, + 555, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 666, + 555, + 685 + ], + "spans": [ + { + "bbox": [ + 354, + 666, + 555, + 685 + ], + "type": "interline_equation", + "content": "\\mathbf {t} _ {G} = \\mathbf {c} _ {Q}, \\quad s _ {G} = \\max _ {\\mathbf {q} \\in \\mathbf {Q} _ {c a m}} \\| \\mathbf {q} - \\mathbf {c} _ {Q} \\| _ {2}. \\tag {3}", + "image_path": "38e5529576eb7cef8e78c0af4a7545599005ec2e5a279ae3f506f8ac10f094a9.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "The key ingredient is to devise the rotation " + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_G = [\\mathbf{r}_{Gx}|\\mathbf{r}_{Gy}|\\mathbf{r}_{Gz}]" + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{Gx},\\mathbf{r}_{Gy},\\mathbf{r}_{Gz}" + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": " are the three columns" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22026" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": "of " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_G" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ", ensuring orientation invariance for the transformed point cloud. Inspired by former works [25, 75, 93], we use the normal vector of the object center " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{n}(\\mathbf{c}_Q)" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{Gz}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ", and project all points in " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{cam}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": " onto the tangent plane of " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_Q" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ", summarizing the projected vectors to determine " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{Gx}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{Gy}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": " is obtained by the cross-product of " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{Gx}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{Gz}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{Gy} = \\mathbf{r}_{Gx} \\times \\mathbf{r}_{Gz}" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ". This ensures that the XY-plane evenly divides the point cloud, with the x-axis representing the principal projection direction. The GRF is " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": "-invariant as it can be uniquely determined by the spatial distribution of the point cloud, making the transformed point cloud robust to variations in pose and size." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 217, + 296, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 296, + 255 + ], + "type": "text", + "content": "Concretely, we perform SVD on the covariance matrix " + }, + { + "bbox": [ + 55, + 217, + 296, + 255 + ], + "type": "inline_equation", + "content": "\\mathrm{Cov}(\\mathbf{Q}_{cam}) = \\frac{1}{N_Q}\\mathbf{Q}_{cam}^\\top \\mathbf{Q}_{cam} - \\mathbf{c}_Q\\mathbf{c}_Q^\\top" + }, + { + "bbox": [ + 55, + 217, + 296, + 255 + ], + "type": "text", + "content": " and use the singular vector w.r.t. the minimum singular value to determine " + }, + { + "bbox": [ + 55, + 217, + 296, + 255 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{Gz}" + }, + { + "bbox": [ + 55, + 217, + 296, + 255 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 265, + 296, + 307 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 296, + 307 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 296, + 307 + ], + "type": "interline_equation", + "content": "\\mathbf {r} _ {G z} = \\left\\{ \\begin{array}{l l} \\mathbf {n} \\left(\\mathbf {c} _ {Q}\\right), & \\text {i f} \\mathbf {n} \\left(\\mathbf {c} _ {Q}\\right) ^ {\\top} \\sum_ {\\mathbf {q} \\in \\mathbf {Q} _ {c a m}} \\left(\\mathbf {c} _ {Q} - \\mathbf {q}\\right) > 0 \\\\ - \\mathbf {n} \\left(\\mathbf {c} _ {Q}\\right), & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {4}", + "image_path": "02dc3bd437e92e914df156e1b849323872e40b453470f0ed0d6bd08c4c18a568.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 308, + 186, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 186, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 186, + 319 + ], + "type": "text", + "content": "Afterwards, " + }, + { + "bbox": [ + 55, + 308, + 186, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{Gx}" + }, + { + "bbox": [ + 55, + 308, + 186, + 319 + ], + "type": "text", + "content": " is computed as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 330, + 296, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 330, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 64, + 330, + 296, + 357 + ], + "type": "interline_equation", + "content": "\\mathbf {r} _ {G x} = \\sum_ {\\mathbf {q} \\in \\mathbf {Q} _ {c a m}} w _ {q} \\left(\\left(\\mathbf {q} - \\mathbf {c} _ {Q}\\right) - \\mathbf {r} _ {G z} ^ {\\top} \\left(\\mathbf {q} - \\mathbf {c} _ {Q}\\right) \\mathbf {r} _ {G z}\\right), \\tag {5}", + "image_path": "60574e31bcfc67be5b5c4428f37b4b55411375e791034e1b87cbc301954b8ae3.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 367, + 295, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 295, + 392 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 367, + 295, + 392 + ], + "type": "inline_equation", + "content": "w_{q}" + }, + { + "bbox": [ + 55, + 367, + 295, + 392 + ], + "type": "text", + "content": " is a weight w.r.t. distance between " + }, + { + "bbox": [ + 55, + 367, + 295, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 55, + 367, + 295, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 367, + 295, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_Q" + }, + { + "bbox": [ + 55, + 367, + 295, + 392 + ], + "type": "text", + "content": " (See the appendix for details)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 392, + 296, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 296, + 451 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 296, + 451 + ], + "type": "text", + "content": "Previous works often depend on complex networks or computationally expensive PPF features for " + }, + { + "bbox": [ + 55, + 392, + 296, + 451 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 55, + 392, + 296, + 451 + ], + "type": "text", + "content": " invariance [46, 94, 96]. In contrast, our transformation is computationally efficient and can seamlessly adapt to a wide range of network architectures." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 453, + 296, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 296, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 296, + 598 + ], + "type": "text", + "content": "Coarse Pose Estimation. Given point clouds " + }, + { + "bbox": [ + 55, + 453, + 296, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_G" + }, + { + "bbox": [ + 55, + 453, + 296, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 453, + 296, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_G" + }, + { + "bbox": [ + 55, + 453, + 296, + 598 + ], + "type": "text", + "content": " in GRF, we sample two sparse point sets " + }, + { + "bbox": [ + 55, + 453, + 296, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_G^c\\in \\mathbb{R}^{N^c\\times 3}" + }, + { + "bbox": [ + 55, + 453, + 296, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 453, + 296, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_G^c\\in \\mathbb{R}^{N^c\\times 3}" + }, + { + "bbox": [ + 55, + 453, + 296, + 598 + ], + "type": "text", + "content": " to efficiently obtain a coarse pose initialization " + }, + { + "bbox": [ + 55, + 453, + 296, + 598 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{T}_{init}" + }, + { + "bbox": [ + 55, + 453, + 296, + 598 + ], + "type": "text", + "content": ". Specifically, we leverage a geometry encoder [71] and a color encoder [65] to extract point cloud and RGB features separately. Features are further concatenated as " + }, + { + "bbox": [ + 55, + 453, + 296, + 598 + ], + "type": "inline_equation", + "content": "f_{Q}^{c}\\in \\mathbb{R}^{N^{c}\\times d}" + }, + { + "bbox": [ + 55, + 453, + 296, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 453, + 296, + 598 + ], + "type": "inline_equation", + "content": "f_{P}^{c}\\in \\mathbb{R}^{N^{c}\\times d}" + }, + { + "bbox": [ + 55, + 453, + 296, + 598 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 453, + 296, + 598 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 453, + 296, + 598 + ], + "type": "text", + "content": " is the dimension of embeddings. Following [51], we add a learnable background token for assigning non-overlapping points. The embeddings, denoted as " + }, + { + "bbox": [ + 55, + 453, + 296, + 598 + ], + "type": "inline_equation", + "content": "\\hat{f}_Q^c\\in \\mathbb{R}^{(N^c +1)\\times d}" + }, + { + "bbox": [ + 55, + 453, + 296, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 453, + 296, + 598 + ], + "type": "inline_equation", + "content": "\\hat{f}_P^c\\in \\mathbb{R}^{(N^c +1)\\times d}" + }, + { + "bbox": [ + 55, + 453, + 296, + 598 + ], + "type": "text", + "content": ", are processed by three stacked Geometric Transformer decoding modules [71]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 600, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 296, + 713 + ], + "type": "text", + "content": "The outputs of the last decoder " + }, + { + "bbox": [ + 55, + 600, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\hat{F}_Q^c\\in \\mathbb{R}^{(N^c +1)\\times D}" + }, + { + "bbox": [ + 55, + 600, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 600, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\hat{F}_P^c\\in \\mathbb{R}^{(N^c +1)\\times D}" + }, + { + "bbox": [ + 55, + 600, + 296, + 713 + ], + "type": "text", + "content": " are point-wise features for building the correlation matrix. However, in our setting, the overlap ratio can be influenced by complicated factors like viewpoint, occlusion or depth noise. To solve this issue, the network additionally predicts overlap confidence " + }, + { + "bbox": [ + 55, + 600, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\hat{O}_Q^c\\in \\mathbb{R}^{(N^c +1)\\times 1}" + }, + { + "bbox": [ + 55, + 600, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 600, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\hat{O}_P^c\\in \\mathbb{R}^{(N^c +1)\\times 1}" + }, + { + "bbox": [ + 55, + 600, + 296, + 713 + ], + "type": "text", + "content": ", which indicate point-wise probabilities of being in the overlapping region. The overlap-aware correlation matrix " + }, + { + "bbox": [ + 55, + 600, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^c\\in \\mathbb{R}^{(N^c +1)\\times (N^c +1)}" + }, + { + "bbox": [ + 55, + 600, + 296, + 713 + ], + "type": "text", + "content": " can thus be com" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 73, + 353, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 73, + 353, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 73, + 353, + 84 + ], + "type": "text", + "content": "puted by" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 343, + 91, + 555, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 91, + 555, + 107 + ], + "spans": [ + { + "bbox": [ + 343, + 91, + 555, + 107 + ], + "type": "interline_equation", + "content": "\\mathbf {X} ^ {c} = \\operatorname {s o f t m a x} [ (\\hat {O} _ {Q} ^ {c} \\odot \\hat {F} _ {Q} ^ {c}) (\\hat {O} _ {P} ^ {c} \\odot \\hat {F} _ {P} ^ {c}) ^ {\\top} ]. \\tag {6}", + "image_path": "69fae2386ebdaf1d9421651031e07f4f06dd7f80bb05acfe37d9caefce885f4f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 113, + 555, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 113, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 313, + 113, + 555, + 148 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 313, + 113, + 555, + 148 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 113, + 555, + 148 + ], + "type": "text", + "content": " indicates element-wise multiplication. Each element in " + }, + { + "bbox": [ + 313, + 113, + 555, + 148 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^c" + }, + { + "bbox": [ + 313, + 113, + 555, + 148 + ], + "type": "text", + "content": " suggests a correlation score between a point pair in " + }, + { + "bbox": [ + 313, + 113, + 555, + 148 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^c" + }, + { + "bbox": [ + 313, + 113, + 555, + 148 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 113, + 555, + 148 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^c" + }, + { + "bbox": [ + 313, + 113, + 555, + 148 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 149, + 556, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 149, + 556, + 233 + ], + "spans": [ + { + "bbox": [ + 313, + 149, + 556, + 233 + ], + "type": "text", + "content": "Once " + }, + { + "bbox": [ + 313, + 149, + 556, + 233 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^c" + }, + { + "bbox": [ + 313, + 149, + 556, + 233 + ], + "type": "text", + "content": " is computed, we can extract all possible corresponding point pairs between " + }, + { + "bbox": [ + 313, + 149, + 556, + 233 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^c" + }, + { + "bbox": [ + 313, + 149, + 556, + 233 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 149, + 556, + 233 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^c" + }, + { + "bbox": [ + 313, + 149, + 556, + 233 + ], + "type": "text", + "content": " as well as their correlation scores to solve Eq. (1). We first compute pose hypotheses by sampling " + }, + { + "bbox": [ + 313, + 149, + 556, + 233 + ], + "type": "inline_equation", + "content": "N_{H}" + }, + { + "bbox": [ + 313, + 149, + 556, + 233 + ], + "type": "text", + "content": " triplets of point pairs according to the distribution of " + }, + { + "bbox": [ + 313, + 149, + 556, + 233 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^c" + }, + { + "bbox": [ + 313, + 149, + 556, + 233 + ], + "type": "text", + "content": ". Then each pose hypothesis " + }, + { + "bbox": [ + 313, + 149, + 556, + 233 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{T}_h = \\{\\Delta \\mathbf{R}_h, \\Delta \\mathbf{t}_h\\}" + }, + { + "bbox": [ + 313, + 149, + 556, + 233 + ], + "type": "text", + "content": " is scored with the reciprocal of distance " + }, + { + "bbox": [ + 313, + 149, + 556, + 233 + ], + "type": "inline_equation", + "content": "D_h" + }, + { + "bbox": [ + 313, + 149, + 556, + 233 + ], + "type": "text", + "content": " as in [27, 51]" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 323, + 239, + 555, + 295 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 239, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 323, + 239, + 555, + 295 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} D _ {h} = \\frac {1}{N ^ {c}} \\sum_ {\\mathbf {p} ^ {c} \\in \\mathbf {P} ^ {c}} \\min _ {\\mathbf {q} ^ {c} \\in \\mathbf {Q} ^ {c}} \\| \\Delta \\mathbf {R} _ {h} ^ {\\top} \\left(\\mathbf {q} ^ {c} - \\Delta \\mathbf {t} _ {h}\\right) - \\mathbf {p} ^ {c} \\| _ {2}, \\tag {7} \\\\ \\xi_ {h} = \\frac {1}{D _ {h}}, \\quad h = 1, 2, 3, \\dots , N _ {H}. \\\\ \\end{array}", + "image_path": "6b4a188a37e3d0e228daa48119b11db98420667821dd368cebf8e0f6cc593640.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 301, + 554, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 554, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 554, + 337 + ], + "type": "text", + "content": "The pose hypothesis with the highest score " + }, + { + "bbox": [ + 313, + 301, + 554, + 337 + ], + "type": "inline_equation", + "content": "\\xi_{h}" + }, + { + "bbox": [ + 313, + 301, + 554, + 337 + ], + "type": "text", + "content": " is selected as the initial pose prediction " + }, + { + "bbox": [ + 313, + 301, + 554, + 337 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{T}_{init} = \\{\\Delta \\mathbf{R}_{init},\\Delta \\mathbf{t}_{init}\\}" + }, + { + "bbox": [ + 313, + 301, + 554, + 337 + ], + "type": "text", + "content": " and further sent to transform the input of the next stage." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 338, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 495 + ], + "type": "text", + "content": "Fine Pose Estimation. After transforming " + }, + { + "bbox": [ + 313, + 338, + 555, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{cam}" + }, + { + "bbox": [ + 313, + 338, + 555, + 495 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 338, + 555, + 495 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{Q}}_{cam}" + }, + { + "bbox": [ + 313, + 338, + 555, + 495 + ], + "type": "text", + "content": " with the initial pose prediction " + }, + { + "bbox": [ + 313, + 338, + 555, + 495 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{T}_{init}" + }, + { + "bbox": [ + 313, + 338, + 555, + 495 + ], + "type": "text", + "content": ", we perform a fine matching process between two dense point sets, i.e., " + }, + { + "bbox": [ + 313, + 338, + 555, + 495 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{Q}}^f \\in \\mathbb{R}^{N^f \\times 3}" + }, + { + "bbox": [ + 313, + 338, + 555, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 338, + 555, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^f \\in \\mathbb{R}^{N^f \\times 3}" + }, + { + "bbox": [ + 313, + 338, + 555, + 495 + ], + "type": "text", + "content": ", to achieve a more accurate pose. The network exploits geometric details through a hierarchical encoding paradigm, comprising a positional encoding layer and a local reference frame encoding layer. For each point, we first encode its global position using a mini-PointNet [70], and then construct SE(3)-invariant local reference frame (LRF) to gather local descriptors. These two encodings complement each other, as local descriptors capture fine geometric structures within small neighborhoods, while the positional encoding layer offers global geometric context." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 495, + 556, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 556, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 556, + 604 + ], + "type": "text", + "content": "Leveraging " + }, + { + "bbox": [ + 313, + 495, + 556, + 604 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{Q}}^f" + }, + { + "bbox": [ + 313, + 495, + 556, + 604 + ], + "type": "text", + "content": " for example, the process of constructing LRF encoding is as follows. For each point " + }, + { + "bbox": [ + 313, + 495, + 556, + 604 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{q}}^m" + }, + { + "bbox": [ + 313, + 495, + 556, + 604 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 495, + 556, + 604 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{Q}}^f" + }, + { + "bbox": [ + 313, + 495, + 556, + 604 + ], + "type": "text", + "content": ", we build a local region set " + }, + { + "bbox": [ + 313, + 495, + 556, + 604 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{Q}}^m = \\{\\tilde{\\mathbf{q}}_j\\}" + }, + { + "bbox": [ + 313, + 495, + 556, + 604 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 495, + 556, + 604 + ], + "type": "inline_equation", + "content": "\\| \\tilde{\\mathbf{q}}_j - \\tilde{\\mathbf{q}}^m \\|_2 \\leq r\\}_{j=1}^{N_D}" + }, + { + "bbox": [ + 313, + 495, + 556, + 604 + ], + "type": "text", + "content": " by grouping " + }, + { + "bbox": [ + 313, + 495, + 556, + 604 + ], + "type": "inline_equation", + "content": "N_D" + }, + { + "bbox": [ + 313, + 495, + 556, + 604 + ], + "type": "text", + "content": " neighboring points of " + }, + { + "bbox": [ + 313, + 495, + 556, + 604 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{q}}^m" + }, + { + "bbox": [ + 313, + 495, + 556, + 604 + ], + "type": "text", + "content": ". The transformation pose " + }, + { + "bbox": [ + 313, + 495, + 556, + 604 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{R}_L^m, \\mathbf{t}_L^m, s_L^m\\}" + }, + { + "bbox": [ + 313, + 495, + 556, + 604 + ], + "type": "text", + "content": " for LRF is computed similarly to GRF (Eq. 3, 4, 5), except that LRF is built upon a local point set while GRF is based on the entire point cloud. By computing the transformation pose, we calculate local point descriptors as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 345, + 609, + 555, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 609, + 555, + 644 + ], + "spans": [ + { + "bbox": [ + 345, + 609, + 555, + 644 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {Q} _ {L} ^ {m} = \\tilde {\\mathcal {Q}} _ {L} ^ {m} = \\left\\{\\left(\\mathbf {R} _ {L} ^ {m}\\right) ^ {\\top} \\left(\\tilde {\\mathbf {q}} _ {j} - \\mathbf {t} _ {L} ^ {m}\\right) / s _ {L} ^ {m} \\right\\} _ {j = 1} ^ {N _ {D}}, \\tag {8} \\\\ \\mathbf {Q} _ {L} ^ {f} = \\{\\mathcal {Q} _ {L} ^ {m} \\} _ {m = 1} ^ {N ^ {f}}. \\\\ \\end{array}", + "image_path": "95a109343a7b6ecedde448a3feffb35d9e6bfb44845a051ee4ab1b6027bc7e36.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 652, + 553, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 553, + 690 + ], + "type": "text", + "content": "Similarly, we transform " + }, + { + "bbox": [ + 313, + 652, + 553, + 690 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^f" + }, + { + "bbox": [ + 313, + 652, + 553, + 690 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 652, + 553, + 690 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_L^f" + }, + { + "bbox": [ + 313, + 652, + 553, + 690 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 313, + 652, + 553, + 690 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_L^m = \\tilde{\\mathcal{Q}}_L^m" + }, + { + "bbox": [ + 313, + 652, + 553, + 690 + ], + "type": "text", + "content": " since LRF is pose-invariant. Then, the LRF encoding is extracted from " + }, + { + "bbox": [ + 313, + 652, + 553, + 690 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_L^f" + }, + { + "bbox": [ + 313, + 652, + 553, + 690 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 652, + 553, + 690 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_L^f" + }, + { + "bbox": [ + 313, + 652, + 553, + 690 + ], + "type": "text", + "content": " with a three-layer MLP." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Positional encoding and LRF encoding are combined with the geometric and color features as the input of the" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22027" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 103, + 70, + 506, + 170 + ], + "blocks": [ + { + "bbox": [ + 103, + 70, + 506, + 170 + ], + "lines": [ + { + "bbox": [ + 103, + 70, + 506, + 170 + ], + "spans": [ + { + "bbox": [ + 103, + 70, + 506, + 170 + ], + "type": "table", + "html": "
MethodSeg. ModelModalityCAD modelLM-OTUD-LYCB-VAvgTime (s)
ZeroPose [7]SAM [40]RGB✓35.642.153.443.73.90
CNOS [61]FastSAM [107]RGB✓39.748.059.949.20.23
CNOS [61]SAM [40]RGB✓39.639.159.546.11.71
SAM-6D [51]FastSAM [107]RGB-D✓42.251.762.152.01.47
SAM-6D [51]SAM [40]RGB-D✓46.056.960.554.54.53
UNOseg (Ours)FastSAM [107]RGBX37.350.367.351.60.24
UNOseg (Ours)SAM [40]RGBX39.756.266.854.22.68
", + "image_path": "d4732933d15ba4561c6067fe7b492eab1cf23cd25cc088583b2ad1a729b2b5bb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 81, + 209, + 529, + 379 + ], + "blocks": [ + { + "bbox": [ + 55, + 178, + 555, + 201 + ], + "lines": [ + { + "bbox": [ + 55, + 178, + 555, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 555, + 201 + ], + "type": "text", + "content": "Table 1. Segmentation results (mAP) on LM-O, TUD-L, and YCB-V. Results of other methods are obtained from the BOP website bop.felk.cvut.cz/leaderboards/segmentation-unseen-bop23. We denote the best score in bold and the second best score with underline." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 81, + 209, + 529, + 379 + ], + "lines": [ + { + "bbox": [ + 81, + 209, + 529, + 379 + ], + "spans": [ + { + "bbox": [ + 81, + 209, + 529, + 379 + ], + "type": "table", + "html": "
MethodSeg. ModelModalityRef. TypeLM-OTUD-LYCB-VAvgTime (s)
ZTE-PPF [1]MRCNN [28]DModel66.390.450.269.00.74
Koenig-PPF [41]Retina/MRCNNRGB-DModel63.192.070.175.10.99
Ref. AlignCenterUNOSeg-SAMDImage27.328.847.534.52.69
FPFH+RANSAC [72]UNOSeg-SAMDImage31.031.050.037.36.38
FPFH+MAC [103]UNOSeg-SAMDImage22.522.149.631.4136.94
PPF [22]UNOSeg-SAMDImage29.714.838.327.611.79
PPF_3D_ICP [22]UNOSeg-SAMDImage44.729.166.846.914.27
FCGF+RANSAC [12]UNOSeg-SAMDImage38.959.057.651.810.96
FCGF+MAC [103]UNOSeg-SAMDImage33.948.351.044.460.53
UTOPIC [11]UNOSeg-SAMDImage13.735.410.519.94.00
GeDi [69]UNOSeg-SAMDImage42.867.360.656.948.89
FreeZe [4]UNOSeg-SAMRGB-DImage45.568.365.559.852.96
SAM-6D* [51]UNOSeg-SAMRGB-DPosed Image54.529.768.150.84.21
UNOpose (Ours)UNOSeg-FastSAMRGB-DImage55.866.581.968.11.11
UNOpose (Ours)UNOSeg-SAMRGB-DImage58.771.083.170.93.70
", + "image_path": "032cc425a3e4ff453dc116eb41c3992a46720aae7d9f593df801c68c5d5de629.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 387, + 555, + 421 + ], + "lines": [ + { + "bbox": [ + 55, + 387, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 555, + 421 + ], + "type": "text", + "content": "Table 2. Pose estimation results on LM-O, TUD-L, and YCB-V. The mean Average Recall (%) of the BOP metric and the average time (s) per image are reported. We highlight the best scores of each setting in bold. Specifically, for adapting SAM-6D [51] to our setting, we randomly choose one of the two templates which [51] uses as a canonical reference during training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 438, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 514 + ], + "type": "text", + "content": "Geometric Transformer. Similar to initial pose prediction, by decoding these features, we obtain the fine point-wise features " + }, + { + "bbox": [ + 55, + 438, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\hat{F}_Q^f,\\hat{F}_P^f" + }, + { + "bbox": [ + 55, + 438, + 296, + 514 + ], + "type": "text", + "content": " and overlap confidencees " + }, + { + "bbox": [ + 55, + 438, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\hat{O}_Q^f,\\hat{O}_P^f" + }, + { + "bbox": [ + 55, + 438, + 296, + 514 + ], + "type": "text", + "content": " , which are then used to obtain the overlap-aware fine correlation matrix " + }, + { + "bbox": [ + 55, + 438, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{f}\\in \\mathbb{R}^{(N^{f} + 1)\\times (N^{f} + 1)}" + }, + { + "bbox": [ + 55, + 438, + 296, + 514 + ], + "type": "text", + "content": " . The final pose " + }, + { + "bbox": [ + 55, + 438, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{T}" + }, + { + "bbox": [ + 55, + 438, + 296, + 514 + ], + "type": "text", + "content": " is predicted by solving Eq. (1) using " + }, + { + "bbox": [ + 55, + 438, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^f" + }, + { + "bbox": [ + 55, + 438, + 296, + 514 + ], + "type": "text", + "content": " with the weighted SVD algorithm." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 528, + 137, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 137, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 137, + 541 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 549, + 171, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 171, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 171, + 562 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 297, + 713 + ], + "type": "text", + "content": "Implementation Details. Implemented with PyTorch [67], we use the DINOv2 [13, 65] pre-trained ViT-base [20] to encode color information. GeoTransformers [71] with attention [82] and sparse-to-dense linear attention [51] are respectively employed to extract coarse and fine geometric correspondences. The network is trained with standard MegaPose synthetic dataset [43] for BOP unseen object pose estimation track [34]. It contains " + }, + { + "bbox": [ + 55, + 570, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\sim 2\\mathrm{M}" + }, + { + "bbox": [ + 55, + 570, + 297, + 713 + ], + "type": "text", + "content": " images of " + }, + { + "bbox": [ + 55, + 570, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\sim 50\\mathrm{K}" + }, + { + "bbox": [ + 55, + 570, + 297, + 713 + ], + "type": "text", + "content": " objects from GSO [21] and ShapeNet [5]. We train 3 epochs with a batch size of 32 on 4 RTX 3090 GPUs. The network is trained with Adam [39] and cosine annealing [58], using a base learning rate of " + }, + { + "bbox": [ + 55, + 570, + 297, + 713 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 55, + 570, + 297, + 713 + ], + "type": "text", + "content": ". Moreover, we lever" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 439, + 555, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 555, + 464 + ], + "type": "text", + "content": "age InfoNCE [64] loss for learning correlation matrix and weighted BCE loss for supervising overlap prediction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 468, + 556, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 556, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 556, + 648 + ], + "type": "text", + "content": "Datasets and Benchmarks. We evaluate our method on BOP test split [33] of three different real-world datasets: LM-O [2], TUD-L [32], and YCB-V [92]. LM-O consists of 200 images from one clutter scene, where eight objects with occlusion are provided for testing. TUD-L comprises three moving objects with mild occlusion and diverse lighting conditions. YCB-V is very challenging due to severe occlusions and sensor noise. It comprises 900 testing images from 12 scenes consisting of 21 objects. For each query object in the test split, we randomly select a reference object beyond the test set (in the training or evaluation set). Unless specified, we constrain the maximal rotation difference between reference and target to " + }, + { + "bbox": [ + 313, + 468, + 556, + 648 + ], + "type": "inline_equation", + "content": "50^{\\circ}" + }, + { + "bbox": [ + 313, + 468, + 556, + 648 + ], + "type": "text", + "content": " to ensure the ground-truth correspondence set is not empty. Moreover, the occlusion ratio of reference objects is required to be less than " + }, + { + "bbox": [ + 313, + 468, + 556, + 648 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 468, + 556, + 648 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 654, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 557, + 715 + ], + "type": "text", + "content": "Evaluation Metrics. For evaluating instance segmentation, we follow [33] to use the mean Average Precision (mAP) metric, which is computed across different Intersection-over-Union (IoU) thresholds (\\{0.5, 0.55, ..., 0.95\\}). Only annotated instances with at least " + }, + { + "bbox": [ + 313, + 654, + 557, + 715 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 654, + 557, + 715 + ], + "type": "text", + "content": " projected visible surfaces" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22028" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 76, + 294, + 150 + ], + "blocks": [ + { + "bbox": [ + 63, + 76, + 294, + 150 + ], + "lines": [ + { + "bbox": [ + 63, + 76, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 63, + 76, + 294, + 150 + ], + "type": "table", + "html": "
CategoryVSDMSSDMSPDARBOP
Dragon47.173.269.463.2
Frog59.452.657.656.4
Can56.865.665.262.5
Avg54.463.764.160.7
", + "image_path": "98115c3f7cea2900bcbeea20f01e09c35312b241d8dd2a7157c8611f93c066ed.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 310, + 71, + 544, + 150 + ], + "blocks": [ + { + "bbox": [ + 173, + 152, + 184, + 161 + ], + "lines": [ + { + "bbox": [ + 173, + 152, + 184, + 161 + ], + "spans": [ + { + "bbox": [ + 173, + 152, + 184, + 161 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 71, + 544, + 150 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 544, + 150 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 544, + 150 + ], + "type": "table", + "html": "
Ref.
Ref.
", + "image_path": "6eb2f5d7bdcd5fd12f6a209e40b106170a7c2944df73cec60c5589e27e596756.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 198, + 296, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 296, + 270 + ], + "type": "text", + "content": "are considered for evaluation. For object pose evaluation, we use the symmetry-aware BOP metric [33]. It is calculated as the mean of the Average Recall of Visible Surface Discrepancy (VSD), Maximum Symmetry-Aware Surface Distance (MSSD), and Maximum Symmetry-Aware Projection Distance (MSPD) (See [33] for details)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 276, + 296, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 296, + 300 + ], + "type": "text", + "content": "4.2. UNO Object Segmentation and Pose Estimation Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 307, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 307, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 296, + 437 + ], + "type": "text", + "content": "UNO Object Segmentation Results. We compare our UNOSeg with some existing CAD-model-based unseen object segmentation methods [7, 51, 61] in Tab. 1. Leveraging a single image as a reference, our UNOSeg achieves comparable results with state-of-the-art methods (Ours " + }, + { + "bbox": [ + 55, + 307, + 296, + 437 + ], + "type": "inline_equation", + "content": "54.2\\%" + }, + { + "bbox": [ + 55, + 307, + 296, + 437 + ], + "type": "text", + "content": " vs. SAM-6D " + }, + { + "bbox": [ + 55, + 307, + 296, + 437 + ], + "type": "inline_equation", + "content": "54.5\\%" + }, + { + "bbox": [ + 55, + 307, + 296, + 437 + ], + "type": "text", + "content": " ) at a faster speed (Ours 2.68s vs. SAM-6D 4.53s). Notably, UNOSeg-FastSAM achieves the best performance on YCB-V (mAP " + }, + { + "bbox": [ + 55, + 307, + 296, + 437 + ], + "type": "inline_equation", + "content": "67.3\\%" + }, + { + "bbox": [ + 55, + 307, + 296, + 437 + ], + "type": "text", + "content": "). Compared to FastSAM [107], SAM [40] achieves consistently better results on three datasets. Therefore, we choose UNOSeg-SAM as the default segmentation for subsequent experiments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 438, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 297, + 714 + ], + "type": "text", + "content": "UNO Object Pose Estimation Results. Since we are the first to conduct relative 6DoF pose estimation for unseen objects and build a brand-new benchmark, we re-implement some traditional [22, 72, 103] and learning-based methods [4, 11, 12, 51, 69] for comparison. Note that we leverage identical data pre-processing, segmentation, query and reference image pairs, and evaluation protocols for all methods for fair comparison. Moreover, we also compare UNOPose with two CAD-model-based pose estimators leveraging " + }, + { + "bbox": [ + 55, + 438, + 297, + 714 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 55, + 438, + 297, + 714 + ], + "type": "text", + "content": "-invariant PPF feature [1, 41]. All results are illustrated in Tab. 2. It clearly shows that our method surpasses all image-reference-based methods by a large margin, and even achieves comparable results with CAD-model-based methods using class-specific detectors [28, 52] (Ours " + }, + { + "bbox": [ + 55, + 438, + 297, + 714 + ], + "type": "inline_equation", + "content": "70.9\\%" + }, + { + "bbox": [ + 55, + 438, + 297, + 714 + ], + "type": "text", + "content": " vs. ZTE-PPF [1] " + }, + { + "bbox": [ + 55, + 438, + 297, + 714 + ], + "type": "inline_equation", + "content": "69.0\\%" + }, + { + "bbox": [ + 55, + 438, + 297, + 714 + ], + "type": "text", + "content": " vs. Koenig-PPF [41] " + }, + { + "bbox": [ + 55, + 438, + 297, + 714 + ], + "type": "inline_equation", + "content": "75.1\\%" + }, + { + "bbox": [ + 55, + 438, + 297, + 714 + ], + "type": "text", + "content": "). \"Ref. AlignCenter\" directly uses the reference rotation as prediction, meanwhile leveraging the shift between query and reference centers as relative translation, so it can be regarded as the baseline of all methods. Among all traditional descriptors, PPF with Iterative Closest Point (ICP) refinement [22] achieves the best results (AR " + }, + { + "bbox": [ + 55, + 438, + 297, + 714 + ], + "type": "inline_equation", + "content": "46.9\\%" + }, + { + "bbox": [ + 55, + 438, + 297, + 714 + ], + "type": "text", + "content": "). Learning-based methods demonstrate varying generalization abilities on unseen objects. For example, the GeoTransformer-based" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 318, + 196, + 550, + 293 + ], + "blocks": [ + { + "bbox": [ + 426, + 152, + 437, + 161 + ], + "lines": [ + { + "bbox": [ + 426, + 152, + 437, + 161 + ], + "spans": [ + { + "bbox": [ + 426, + 152, + 437, + 161 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 171, + 547, + 183 + ], + "lines": [ + { + "bbox": [ + 61, + 171, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 61, + 171, + 547, + 183 + ], + "type": "text", + "content": "Table 3. One reference per category experiments on TUD-L. Blue and green contours denote GT and estimated poses, respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 196, + 550, + 293 + ], + "lines": [ + { + "bbox": [ + 318, + 196, + 550, + 293 + ], + "spans": [ + { + "bbox": [ + 318, + 196, + 550, + 293 + ], + "type": "table", + "html": "
ROWGRFLRFFPEVSDMSSDMSPD\\( \\mathrm{AR}_{\\text{BOP}} \\)
A0✘✘✘64.668.159.263.9
A1✓✘✘66.472.863.267.5
A2✘✓✘66.269.560.465.4
A3✓✓✘66.973.263.567.9
B0✘✘✓79.885.075.580.1
B1✓✘✓79.485.276.680.4
B2✘✓✓81.185.176.680.9
B3✓✓✓80.786.077.781.5
", + "image_path": "b9374e63da4f81e8637ab892dedd0fff979da821e8f2cf7a4f2163be4e239063.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 301, + 555, + 323 + ], + "lines": [ + { + "bbox": [ + 313, + 301, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 555, + 323 + ], + "type": "text", + "content": "Table 4. Ablation of GRF and LRF on YCB-V. FPE stands for \"fine pose estimation\"." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 346, + 555, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 406 + ], + "type": "text", + "content": "method UTOPIC, designed for registering object-level partial point clouds, performs suboptimally in our setting. In contrast, the adapted unseen object pose estimation methods, i.e. SAM-6D [51] and FreeZe [4], achieve satisfactory results compared to their counterparts." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 406, + 555, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 555, + 442 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 555, + 442 + ], + "type": "text", + "content": "Run-time Analysis. Evaluated on a single RTX 3090 GPU, our pipeline runs at 3.70s for one image, including 2.68s for segmentation and 1.02s for pose estimation (Tab. 1 - 2)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 451, + 476, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 476, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 476, + 464 + ], + "type": "text", + "content": "4.3. One Reference for a Category" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 470, + 556, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 556, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 556, + 638 + ], + "type": "text", + "content": "To minimize the efforts of acquiring references, we leverage a single reference for all objects of the same category in a dataset. Tab. 3 illustrates the qualitative and quantitative experimental results on TUD-L. It is worth noting that the target object in TUD-L is dynamic rather than static, preventing methods from leveraging background information. This setup poses challenges for relative pose estimation methods that rely on scene features, making them likely to fail in this task. Despite this, UNOPose achieves impressive performance (60.7% in terms of the " + }, + { + "bbox": [ + 313, + 470, + 556, + 638 + ], + "type": "inline_equation", + "content": "\\mathrm{AR}_{\\mathrm{BOP}}" + }, + { + "bbox": [ + 313, + 470, + 556, + 638 + ], + "type": "text", + "content": " metric) with only a single reference per category. The visualization results in Tab. 3b further show UNOPose is capable of predicting omnidirectional relative poses. Additional experiments on reference selection are provided in the appendix." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 647, + 416, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 416, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 416, + 658 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "Effectiveness of GRF and LRF. As demonstrated in Sec. 3.3.2, GRF transfers arbitrary object poses to an " + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "-invariant frame, while LRF assists with extracting precise local descriptors for fine pose estimation. Tab. 4 shows that" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22029" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 293, + 168 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 293, + 168 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 293, + 168 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 293, + 168 + ], + "type": "table", + "html": "
ROWMethodVSDMSSDMSPD\\( \\mathrm{AR}_{\\text{BOP}} \\)
C0B3 + Soft Correspondence [95]77.181.772.777.2
C1B3 + Overlap Predictor82.687.479.483.1
C2A3 + Overlap Predictor68.474.664.869.2
D0C1: Seg. → CNOS [61]77.480.973.677.3
D1C1: Seg. → SAM-6D [51]78.382.174.378.2
D2C1: Seg. → MRCNN [28, 42]81.285.777.981.6
D3C1: Seg. → GT Seg.86.689.282.186.0
E0C1: Ref. Real → Rendered83.789.683.185.5
", + "image_path": "1d4260866584a2d3b0cbf9ae248d38f31524b6ea23bfb1b4e18c03f28c9404e2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 177, + 296, + 199 + ], + "lines": [ + { + "bbox": [ + 55, + 177, + 296, + 199 + ], + "spans": [ + { + "bbox": [ + 55, + 177, + 296, + 199 + ], + "type": "text", + "content": "Table 5. Ablation of overlap predictor, segmentation, and reference type on YCB-V." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 67, + 211, + 284, + 373 + ], + "blocks": [ + { + "bbox": [ + 67, + 211, + 284, + 373 + ], + "lines": [ + { + "bbox": [ + 67, + 211, + 284, + 373 + ], + "spans": [ + { + "bbox": [ + 67, + 211, + 284, + 373 + ], + "type": "image", + "image_path": "85f1dfbd821bff5428a02769c25d05f473cb29697d79d389fa2aaf40cca98d1d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 380, + 295, + 426 + ], + "lines": [ + { + "bbox": [ + 55, + 380, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 295, + 426 + ], + "type": "text", + "content": "Figure 3. Ablation of initial rotation distance between query and reference objects. We categorize all testing objects into nine groups according to the initial rotation distance, and evaluate the " + }, + { + "bbox": [ + 55, + 380, + 295, + 426 + ], + "type": "inline_equation", + "content": "\\mathrm{AR}_{\\mathrm{BOP}}" + }, + { + "bbox": [ + 55, + 380, + 295, + 426 + ], + "type": "text", + "content": " metric and overlap ratio for each group separately." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 448, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 295, + 555 + ], + "type": "text", + "content": "the effectiveness of GRF and LRF is complementary in UNOpose. Specifically, GRF brings significant performance enhancement in the coarse pose estimation (Tab. 4 A1 vs. A0, A3 vs. A2), providing an accurate pose initialization for the next stage. Meanwhile, LRF is constructed to exploit fine-grained local information, primarily enhancing the fine pose estimation results (Tab. 4 B0 vs. B2, B1 vs. B3). With both GRF and LRF, our network achieves obvious performance improvement (Tab. 4 A3 vs. A0, B3 vs. B0)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 556, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 295, + 664 + ], + "type": "text", + "content": "Ablation on Correspondence Loss. In partial-to-partial point cloud registration, only some of the points can find their counterparts on the target point cloud. Therefore, the introduction of overlap prediction helps to distinguish the points in the overlap region and achieves good results (Tab. 4-5 C1 vs. B3, C2 vs. A3). Moreover, we consider soft correspondence loss in [95] as an alternative, which predicts probabilities of overlap between each point pair. However, the result did not meet the expectations (Tab. 5 C0)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Ablation on Segmentation. Tab. 1 showcases the advantage of our UNOSeg over other segmentation methods on YCB-V. Furthermore, substituting UNOSeg with CNOS [61], SAM-6D [51], or MRCNN [28, 42] results in a noticeable" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "content": "decline in UNOpose's performance (Tab. 5 D0-D2). Moreover, substituting UNOSeg with ground-truth segmentation brings " + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "inline_equation", + "content": "\\approx 3\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "content": " performance enhancement (Tab. 5 D3 vs. C1)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 108, + 556, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 556, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 556, + 192 + ], + "type": "text", + "content": "Using Rendered References. We replaced the real reference with a pose-identical rendered version and present the results in Tab. 5. With reduced depth noise and domain gap between training and testing, UNOPose works even better using the rendered reference (Tab. 5 E0 vs. C1). However, this ablation requires the object model, which is more difficult to acquire than an unposed real image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 192, + 556, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 556, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 556, + 372 + ], + "type": "text", + "content": "Ablation on Reference Rotation Distance. To investigate the impact of the initial rotation distance between the query and reference images, we randomly select the reference object with a maximum rotation distance of " + }, + { + "bbox": [ + 313, + 192, + 556, + 372 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 313, + 192, + 556, + 372 + ], + "type": "text", + "content": " for each query object. Further, all testing objects are categorized into nine groups according to initial rotation distances. We separately evaluate each group and present the results in Fig. 3. Meanwhile, we calculate the overlap ratio between reference and query instances of each group and visualize the mean and standard deviation. Notably, performance significantly decreases when the rotation distance exceeds " + }, + { + "bbox": [ + 313, + 192, + 556, + 372 + ], + "type": "inline_equation", + "content": "50^{\\circ}" + }, + { + "bbox": [ + 313, + 192, + 556, + 372 + ], + "type": "text", + "content": ". This decline is attributed to the reduced overlap between query and reference viewpoints. However, our network still shows favorable transfer ability in extreme relative pose estimation (rotation distance " + }, + { + "bbox": [ + 313, + 192, + 556, + 372 + ], + "type": "inline_equation", + "content": "(80^{\\circ}, 90^{\\circ}]" + }, + { + "bbox": [ + 313, + 192, + 556, + 372 + ], + "type": "text", + "content": ", overlap ratio " + }, + { + "bbox": [ + 313, + 192, + 556, + 372 + ], + "type": "inline_equation", + "content": "45.4\\%" + }, + { + "bbox": [ + 313, + 192, + 556, + 372 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 192, + 556, + 372 + ], + "type": "inline_equation", + "content": "\\mathrm{AR}_{\\mathrm{BOP}} 54.8\\%" + }, + { + "bbox": [ + 313, + 192, + 556, + 372 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 372, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 555, + 407 + ], + "type": "text", + "content": "The appendix presents further implementation details, visualization results, and more experimental analysis on reference selection and extensive datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 419, + 537, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 537, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 537, + 433 + ], + "type": "text", + "content": "5. Conclusion, Limitation and Future work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 439, + 556, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 556, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 556, + 570 + ], + "type": "text", + "content": "This work has introduced UNOPose, a new approach for unseen object pose estimation with one reference. The key idea is constructing the " + }, + { + "bbox": [ + 313, + 439, + 556, + 570 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 313, + 439, + 556, + 570 + ], + "type": "text", + "content": "-invariant reference frame for tackling diverse pose and size variations. Moreover, we propose an overlap predictor for handling low-overlap scenarios. To evaluate the proposed method, we devise a new benchmark based on the BOP challenge and compare some state-of-the-art methods. Experimental results show UNOPose surpasses all compared reference-based methods significantly, and is competitive with some " + }, + { + "bbox": [ + 313, + 439, + 556, + 570 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 313, + 439, + 556, + 570 + ], + "type": "text", + "content": "-invariant-feature-based methods relying on CAD models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 571, + 556, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 556, + 655 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 556, + 655 + ], + "type": "text", + "content": "Limitation and Future Work. While efficient, correspondence-based methods rely on predicting robust and dense correspondence between query and reference objects, thereby limited in extreme two-view geometry applications. Future work will focus on reconstructing the unseen object from a single reference [53, 108] and concurrently estimating its object pose." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 658, + 556, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 556, + 702 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 556, + 702 + ], + "type": "text", + "content": "Acknowledgements. This work was supported in part by the National Natural Science Foundation of China under Grant No. 62406169, and in part by the China Postdoctoral Science Foundation under Grant No. 2024M761673." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22030" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 91, + 296, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 91, + 296, + 113 + ], + "spans": [ + { + "bbox": [ + 66, + 91, + 296, + 113 + ], + "type": "text", + "content": "[1] Anonymous. Zte-ppf. https://bop.felk(cvut.cz/method_info/250/, 2022. 2, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 114, + 295, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 114, + 295, + 158 + ], + "spans": [ + { + "bbox": [ + 66, + 114, + 295, + 158 + ], + "type": "text", + "content": "[2] Eric Brachmann, Alexander Krull, Frank Michel, Stefan Gumhold, Jamie Shotton, and Carsten Rother. Learning 6D object pose estimation using 3D object coordinates. In ECCV, 2014. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 159, + 296, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 159, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 66, + 159, + 296, + 192 + ], + "type": "text", + "content": "[3] Dingding Cai, Janne Heikkilä, and Esa Rahtu. Ove6d: Object viewpoint encoding for depth-based 6d object pose estimation. In CVPR, pages 6803-6813, 2022. 1, 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 194, + 296, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 194, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 66, + 194, + 296, + 237 + ], + "type": "text", + "content": "[4] Andrea Caraffa, Davide Boscaini, Amir Hamza, and Fabio Poiesi. Freeze: Training-free zero-shot 6d pose estimation with geometric and vision foundation models. In ECCV, 2024. 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 239, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 239, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 66, + 239, + 296, + 293 + ], + "type": "text", + "content": "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012, 2015. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 296, + 296, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 296, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 66, + 296, + 296, + 338 + ], + "type": "text", + "content": "[6] Dengsheng Chen, Jun Li, Zheng Wang, and Kai Xu. Learning canonical shape space for category-level 6D object pose and size estimation. In CVPR, pages 11970-11979, 2020. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 341, + 296, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 341, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 66, + 341, + 296, + 384 + ], + "type": "text", + "content": "[7] Jianqiu Chen, Zikun Zhou, Mingshan Sun, Rui Zhao, Liwei Wu, Tianpeng Bao, and Zhenyu He. Zeropose: Cadprompted zero-shot object 6d pose estimation in cluttered scenes. IEEE TCSVT, 2024. 4, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 387, + 296, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 387, + 296, + 419 + ], + "spans": [ + { + "bbox": [ + 66, + 387, + 296, + 419 + ], + "type": "text", + "content": "[8] Kai Chen and Qi Dou. SGPA: Structure-guided prior adaptation for category-level 6D object pose estimation. In ICCV, pages 2773-2782, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 421, + 296, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 421, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 66, + 421, + 296, + 475 + ], + "type": "text", + "content": "[9] Wei Chen, Xi Jia, Hyung Jin Chang, Jinming Duan, Shen Linlin, and Ales Leonardis. FS-Net: Fast shape-based network for category-level 6D object pose estimation with decoupled rotation mechanism. In CVPR, pages 1581–1590, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 477, + 296, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 477, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 62, + 477, + 296, + 552 + ], + "type": "text", + "content": "[10] Yamei Chen, Yan Di, Guangyao Zhai, Fabian Manhardt, Chenyangguang Zhang, Ruida Zhang, Federico Tombari, Nassir Navab, and Benjamin Busam. Secondpose: Se (3)-consistent dual-stream feature fusion for category-level pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9959-9969, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 555, + 296, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 555, + 296, + 609 + ], + "spans": [ + { + "bbox": [ + 62, + 555, + 296, + 609 + ], + "type": "text", + "content": "[11] Zhilei Chen, Honghua Chen, Lina Gong, Xuefeng Yan, Jun Wang, Yanwen Guo, Jing Qin, and Mingqiang Wei. Utopic: Uncertainty-aware overlap prediction network for partial point cloud registration. In Computer Graphics Forum, pages 87-98. Wiley Online Library, 2022. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 612, + 296, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 612, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 62, + 612, + 296, + 644 + ], + "type": "text", + "content": "[12] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In ICCV, pages 8958-8966, 2019. 3, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 646, + 296, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 646, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 62, + 646, + 296, + 677 + ], + "type": "text", + "content": "[13] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In ICLR, 2024. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 680, + 296, + 713 + ], + "type": "text", + "content": "[14] Haowen Deng, Tolga Birdal, and Slobodan Ilic. Ppf-foldnet: Unsupervised learning of rotation invariant 3d local descriptors. In ECCV, pages 602-618, 2018. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 106 + ], + "type": "text", + "content": "[15] Haowen Deng, Tolga Birdal, and Slobodan Ilic. Ppfnet: Global context aware local features for robust 3d point matching. In CVPR, pages 195-205, 2018. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 107, + 555, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 107, + 555, + 161 + ], + "spans": [ + { + "bbox": [ + 320, + 107, + 555, + 161 + ], + "type": "text", + "content": "[16] Maximilian Denninger, Dominik Winkelbauer, Martin Sundermeyer, Wout Boerdijk, Markus Knauer, Klaus H. Strobl, Matthias Humt, and Rudolph Triebel. Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software, 8(82):4901, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 162, + 555, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 162, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 320, + 162, + 555, + 227 + ], + "type": "text", + "content": "[17] Yan Di, Ruida Zhang, Zhiqiang Lou, Fabian Manhardt, Xi-angyang Ji, Nassir Navab, and Federico Tombari. Gpv-posed Category-level object pose estimation via geometry-guided point-wise voting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6781–6791, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 228, + 555, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 228, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 320, + 228, + 555, + 293 + ], + "type": "text", + "content": "[18] Yan Di, Chenyangguang Zhang, Ruida Zhang, Fabian Manhardt, Yongzhi Su, Jason Rambach, Didier Stricker, Xi-angyang Ji, and Federico Tombari. U-red: Unsupervised 3d shape retrieval and deformation for partial point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8884–8895, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 294, + 555, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 294, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 320, + 294, + 555, + 369 + ], + "type": "text", + "content": "[19] Yan Di, Chenyangguang Zhang, Chaowei Wang, Ruida Zhang, Guangyao Zhai, Yanyan Li, Bowen Fu, Xiangyang Ji, and Shan Gao. Shapematcher: Self-supervised joint shape canonicalization segmentation retrieval and deformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21017-21028, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 371, + 555, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 371, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 320, + 371, + 555, + 426 + ], + "type": "text", + "content": "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 426, + 555, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 426, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 320, + 426, + 555, + 480 + ], + "type": "text", + "content": "[21] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In ICRA, pages 2553-2560. IEEE, 2022. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 482, + 555, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 482, + 555, + 524 + ], + "spans": [ + { + "bbox": [ + 320, + 482, + 555, + 524 + ], + "type": "text", + "content": "[22] Bertram Drost, Markus Ulrich, Nassir Navab, and Slobodan Ilic. Model globally, match locally: Efficient and robust 3d object recognition. In CVPR, pages 998-1005. IEEE, 2010. 3, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 526, + 555, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 526, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 320, + 526, + 555, + 579 + ], + "type": "text", + "content": "[23] Guoguang Du, Kai Wang, Shiguo Lian, and Kaiyong Zhao. Vision-based robotic grasping from object localization, object pose estimation to grasp estimation for parallel grippers: a review. Artificial Intelligence Review, 54(3):1677-1734, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 581, + 555, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 581, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 320, + 581, + 555, + 634 + ], + "type": "text", + "content": "[24] Zhaoxin Fan, Zhenbo Song, Jian Xu, Zhicheng Wang, Kejian Wu, Hongyan Liu, and Jun He. Object level depth reconstruction for category level 6d object pose estimation from monocular rgb image. In ECCV, pages 220-236. Springer, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 636, + 555, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 636, + 555, + 680 + ], + "spans": [ + { + "bbox": [ + 320, + 636, + 555, + 680 + ], + "type": "text", + "content": "[25] Andrea Frome, Daniel Huber, Ravi Kolluri, Thomas Bulow, and Jitendra Malik. Recognizing objects in range data using regional point descriptors. In ECCV, pages 224-237. Springer, 2004. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 681, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 681, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 681, + 555, + 713 + ], + "type": "text", + "content": "[26] Bowen Fu, Sek Kun Leong, Yan Di, Gu Wang, Jiwen Tang, and Xiangyang Ji. Lanpose: Language-instructed 6d object pose estimation for robotic assembly. In ECCVW, 2024. 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22031" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 297, + 117 + ], + "type": "text", + "content": "[27] Rasmus Laurvig Haugaard and Anders Glent Buch. Surfemb: Dense and continuous correspondence distributions for object pose estimation with learnt surface embeddings. In CVPR, pages 6749-6758, 2022. 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 119, + 296, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 119, + 296, + 151 + ], + "spans": [ + { + "bbox": [ + 61, + 119, + 296, + 151 + ], + "type": "text", + "content": "[28] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask R-CNN. In ICCV, pages 2961-2969, 2017. 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 153, + 296, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 153, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 62, + 153, + 296, + 196 + ], + "type": "text", + "content": "[29] Xingyi He, Jiaming Sun, Yuang Wang, Di Huang, Hujun Bao, and Xiaowei Zhou. Onepose++: Keypoint-free one-shot object pose estimation without cad models. NeurIPS, 35:35103-35115, 2022. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 199, + 296, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 199, + 296, + 232 + ], + "spans": [ + { + "bbox": [ + 62, + 199, + 296, + 232 + ], + "type": "text", + "content": "[30] Yisheng He, Haibin Huang, Haoqiang Fan, Qifeng Chen, and Jian Sun. Ffb6d: A full flow bidirectional fusion network for 6d pose estimation. In CVPR, pages 3003-3013, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 235, + 296, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 235, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 62, + 235, + 296, + 266 + ], + "type": "text", + "content": "[31] Yisheng He, Yao Wang, Haoqiang Fan, Jian Sun, and Qifeng Chen. Fs6d: Few-shot 6d pose estimation of novel objects. In CVPR, pages 6814-6824, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 269, + 296, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 269, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 269, + 296, + 323 + ], + "type": "text", + "content": "[32] Tomas Hodan, Frank Michel, Eric Brachmann, Wadim Kehl, Anders GlentBuch, Dirk Kraft, Bertram Drost, Joel Vidal, Stephan Ihrke, Xenophon Zabulis, et al. BOP: Benchmark for 6D Object Pose Estimation. In ECCV, pages 19-34, 2018. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 326, + 296, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 326, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 326, + 296, + 369 + ], + "type": "text", + "content": "[33] Tomas Hodan, Martin Sundermeyer, Bertram Drost, Yann Labbe, Eric Brachmann, Frank Michel, Carsten Rother, and Jiri Matas. BOP Challenge 2020 on 6D object localization. In ECCVW, pages 577-594, 2020. 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 372, + 296, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 372, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 372, + 296, + 435 + ], + "type": "text", + "content": "[34] Tomas Hodan, Martin Sundermeyer, Yann Labbe, Van Nguyen Nguyen, Gu Wang, Eric Brachmann, Bertram Drost, Vincent Lepetit, Carsten Rother, and Jiri Matas. Bop challenge 2023 on detection, segmentation and pose estimation of seen and unseen rigid objects. In CVPRW, 2024. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 439, + 296, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 439, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 62, + 439, + 296, + 471 + ], + "type": "text", + "content": "[35] Yinlin Hu, Pascal Fua, Wei Wang, and Mathieu Salzmann. Single-stage 6d object pose estimation. In CVPR, pages 2930-2939, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 474, + 296, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 474, + 296, + 516 + ], + "spans": [ + { + "bbox": [ + 62, + 474, + 296, + 516 + ], + "type": "text", + "content": "[36] Junwen Huang, Hao Yu, Kuan-Ting Yu, Nassir Navab, Slobodan Ilic, and Benjamin Busam. Matchu: Matching unseen objects for 6d pose estimation from rgb-d images. In CVPR, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 519, + 296, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 519, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 62, + 519, + 296, + 563 + ], + "type": "text", + "content": "[37] Siyuan Huang, Siyuan Qi, Yinxue Xiao, Yixin Zhu, Ying Nian Wu, and Song-Chun Zhu. Cooperative holistic scene understanding: Unifying 3d object, layout, and camera pose estimation. NeurIPS, 31, 2018. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 566, + 296, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 566, + 296, + 608 + ], + "spans": [ + { + "bbox": [ + 62, + 566, + 296, + 608 + ], + "type": "text", + "content": "[38] Shengyu Huang, Zan Gojcic, Mikhail Usvyatsov, Andreas Wieser, and Konrad Schindler. Predator: Registration of 3d point clouds with low overlap. In CVPR, pages 4267-4276, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 611, + 296, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 611, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 62, + 611, + 296, + 632 + ], + "type": "text", + "content": "[39] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 635, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 296, + 678 + ], + "type": "text", + "content": "[40] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 2, 4, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 681, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 681, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 681, + 296, + 712 + ], + "type": "text", + "content": "[41] Rebecca König and Bertram Drost. A hybrid approach for 6 dof pose estimation. In ECCV, pages 700-706. Springer, 2020. 2, 6, 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 321, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 555, + 106 + ], + "type": "text", + "content": "[42] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. CosyPose: Consistent multi-view multi-object 6D pose estimation. In ECCV, pages 574-591, 2020. 1, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 108, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 108, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 321, + 108, + 555, + 162 + ], + "type": "text", + "content": "[43] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. Megapose: 6d pose estimation of novel objects via render & compare. In CoRL, pages 715-725. PMLR, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 164, + 555, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 164, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 321, + 164, + 555, + 196 + ], + "type": "text", + "content": "[44] Junha Lee, Seungwook Kim, Minsu Cho, and Jaesik Park. Deep hough voting for robust global registration. In ICCV, pages 15994-16003, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 198, + 555, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 198, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 321, + 198, + 555, + 251 + ], + "type": "text", + "content": "[45] Jiahao Li, Changhao Zhang, Ziyao Xu, Hangning Zhou, and Chi Zhang. Iterative distance-aware similarity matrix convolution with mutual-supervised point elimination for efficient point cloud registration. In ECCV, pages 378-394. Springer, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 254, + 555, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 254, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 321, + 254, + 555, + 297 + ], + "type": "text", + "content": "[46] Xiaolong Li, Yijia Weng, Li Yi, Leonidas J Guibas, A Abbott, Shuran Song, and He Wang. Leveraging se (3) equivariance for self-supervised category-level object pose estimation from point clouds. NeurIPS, 34, 2021. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 299, + 555, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 299, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 321, + 299, + 555, + 330 + ], + "type": "text", + "content": "[47] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. DeepIM: Deep iterative matching for 6D pose estimation. IJCV, 128(3):657-678, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 332, + 555, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 332, + 555, + 375 + ], + "spans": [ + { + "bbox": [ + 321, + 332, + 555, + 375 + ], + "type": "text", + "content": "[48] Zhigang Li, Gu Wang, and Xiangyang Ji. CDPN: Coordinates-Based Disentangled Pose Network for Real-Time RGB-Based 6-DoF Object Pose Estimation. In ICCV, pages 7678-7687, 2019. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 377, + 554, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 377, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 321, + 377, + 554, + 409 + ], + "type": "text", + "content": "[49] Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. In 3DV, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 411, + 555, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 411, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 321, + 411, + 555, + 464 + ], + "type": "text", + "content": "[50] Jiehong Lin, Zewei Wei, Zhihao Li, Songcen Xu, Kui Jia, and Yuanqing Li. DualPoseNet: Category-level 6D object pose and size estimation using dual pose network with refined learning of pose consistency. In ICCV, pages 3560-3569, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 468, + 555, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 468, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 321, + 468, + 555, + 499 + ], + "type": "text", + "content": "[51] Jiehong Lin, Lihua Liu, Dekun Lu, and Kui Jia. Sam-6d: Segment anything model meets zero-shot 6d object pose estimation. In CVPR, 2024. 1, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 501, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 501, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 321, + 501, + 555, + 533 + ], + "type": "text", + "content": "[52] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 535, + 555, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 535, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 321, + 535, + 555, + 578 + ], + "type": "text", + "content": "[53] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In ICCV, pages 9298–9309, 2023. 3, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 321, + 580, + 555, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 580, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 321, + 580, + 555, + 612 + ], + "type": "text", + "content": "[54] Xingyu Liu, Gu Wang, Yi Li, and Xiangyang Ji. CATRE: iterative point clouds alignment for category-level object pose refinement. In ECCV, 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 321, + 614, + 555, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 614, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 321, + 614, + 555, + 668 + ], + "type": "text", + "content": "[55] Xingyu Liu, Yingyue Li, Chengxi Li, Gu Wang, Chenyangguang Zhang, Ziqin Huang, and Xiangyang Ji. Gfreedet: Exploiting gaussian splatting and foundation models for model-free unseen object detection in the bop challenge 2024. arXiv preprint arXiv:2412.01552, 2024. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 321, + 670, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 670, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 321, + 670, + 555, + 712 + ], + "type": "text", + "content": "[56] Xingyu Liu, Chenyangguang Zhang, Gu Wang, Ruida Zhang, and Xiangyang Ji. Rasim: A range-aware high-fidelity rgb-d data simulation pipeline for real-world applications. In ICRA, pages 17057-17064. IEEE, 2024. 1, 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22032" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 297, + 116 + ], + "type": "text", + "content": "[57] Yuan Liu, Yilin Wen, Sida Peng, Cheng Lin, Xiaoxiao Long, Taku Komura, and Wenping Wang. Gen6d: Generalizable model-free 6-dof object pose estimation from rgb images. In ECCV, 2022. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 119, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 119, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 61, + 119, + 296, + 140 + ], + "type": "text", + "content": "[58] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In *ICLR*, 2017. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 143, + 296, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 143, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 62, + 143, + 296, + 175 + ], + "type": "text", + "content": "[59] Eric Marchand, Hideaki Uchiyama, and Fabien Spindler. Pose estimation for augmented reality: a hands-on survey. IEEE TVCG, 22(12):2633-2651, 2015. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 178, + 296, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 178, + 296, + 232 + ], + "spans": [ + { + "bbox": [ + 62, + 178, + 296, + 232 + ], + "type": "text", + "content": "[60] Van Nguyen Nguyen, Yinlin Hu, Yang Xiao, Mathieu Salzmann, and Vincent Lepetit. Templates for 3d object pose estimation revisited: Generalization to new objects and robustness to occlusions. In CVPR, pages 6771-6780, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 235, + 295, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 235, + 295, + 278 + ], + "spans": [ + { + "bbox": [ + 62, + 235, + 295, + 278 + ], + "type": "text", + "content": "[61] Van Nguyen Nguyen, Thibault Groueix, Georgy Ponimatkin, Vincent Lepetit, and Tomas Hodan. Cnos: A strong baseline for cad-based novel object segmentation. In ICCV, pages 2134-2140, 2023. 1, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 281, + 296, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 281, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 281, + 296, + 323 + ], + "type": "text", + "content": "[62] Van Nguyen Nguyen, Thibault Groueix, Mathieu Salzmann, and Vincent Lepetit. Gigapose: Fast and robust novel object pose estimation via one correspondence. In CVPR, 2024. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 327, + 296, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 327, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 62, + 327, + 296, + 370 + ], + "type": "text", + "content": "[63] Yinyu Nie, Xiaoguang Han, Shihui Guo, Yujuan Zheng, Jian Chang, and Jian Jun Zhang. Total3DUnderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In CVPR, pages 55-64, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 373, + 296, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 373, + 296, + 405 + ], + "spans": [ + { + "bbox": [ + 62, + 373, + 296, + 405 + ], + "type": "text", + "content": "[64] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 407, + 296, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 407, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 62, + 407, + 296, + 506 + ], + "type": "text", + "content": "[65] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. TMLR, 2024. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 508, + 296, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 508, + 296, + 551 + ], + "spans": [ + { + "bbox": [ + 62, + 508, + 296, + 551 + ], + "type": "text", + "content": "[66] Evin Pinar Örnek, Yann Labbé, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, and Tomas Hodan. Found- pose: Unseen object pose estimation with foundation features. In CVPR, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 554, + 296, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 554, + 296, + 608 + ], + "spans": [ + { + "bbox": [ + 62, + 554, + 296, + 608 + ], + "type": "text", + "content": "[67] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. PyTorch: An imperative style, high-performance deep learning library. In NeurlPS, pages 8026-8037, 2019. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 611, + 294, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 611, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 62, + 611, + 294, + 643 + ], + "type": "text", + "content": "[68] Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, and Hujun Bao. Pynet: Pixel-wise voting network for 6dof pose estimation. In CVPR, pages 4561-4570, 2019. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 646, + 296, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 646, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 62, + 646, + 296, + 677 + ], + "type": "text", + "content": "[69] Fabio Poiesi and Davide Boscaini. Learning general and distinctive 3d local deep descriptors for point cloud registration. IEEE TPAMI, 45(3):3979-3985, 2022. 3, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 681, + 296, + 713 + ], + "type": "text", + "content": "[70] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. PointNet: Deep learning on point sets for 3D classification and segmentation. CVPR, 1(2):4, 2017. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "type": "text", + "content": "[71] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, Slobodan Ilic, Dewen Hu, and Kai Xu. Geotransformer: Fast and robust point cloud registration with geometric transformer. IEEE TPAMI, 2023. 3, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 118, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 555, + 150 + ], + "type": "text", + "content": "[72] Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (fpfh) for 3d registration. In ICRA, pages 3212-3217. IEEE, 2009. 3, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 152, + 555, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 152, + 555, + 182 + ], + "spans": [ + { + "bbox": [ + 320, + 152, + 555, + 182 + ], + "type": "text", + "content": "[73] Ivan Shugurov, Sergey Zakharov, and Slobodan Ilic. Dpodv2: Dense correspondence-based 6 dof pose estimation. IEEE TPAMI, 44(11):7417-7435, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 184, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 184, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 320, + 184, + 555, + 217 + ], + "type": "text", + "content": "[74] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. Osop: A multi-stage one shot object pose estimation framework. In CVPR, pages 6835-6844, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 217, + 555, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 217, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 320, + 217, + 555, + 248 + ], + "type": "text", + "content": "[75] Fridtjof Stein, Gérard Medioni, et al. Structural indexing: Efficient 3-d object recognition. IEEE TPAMI, 14(2):125-145, 1992. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 251, + 555, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 251, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 320, + 251, + 555, + 304 + ], + "type": "text", + "content": "[76] Yongzhi Su, Jason Rambach, Nareg Minaskan, Paul Lesur, Alain Pagani, and Didier Stricker. Deep multi-state object pose estimation for augmented reality assembly. In IEEE International Symposium on Mixed and Augmented Reality Adjunct (ISMAR-Adjunct), pages 222–227, 2019. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 306, + 555, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 306, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 320, + 306, + 555, + 348 + ], + "type": "text", + "content": "[77] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. Onepose: One-shot object pose estimation without cad models. In CVPR, pages 6825-6834, 2022. 1, 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 350, + 555, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 350, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 320, + 350, + 555, + 392 + ], + "type": "text", + "content": "[78] Jingyi Tang, Gu Wang, Zeyu Chen, Shengquan Li, Xiu Li, and Xiangyang Ji. Fafa: Frequency-aware flow-aided self-supervision for underwater object pose estimation. In ECCV, pages 352–369. Springer, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 394, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 394, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 320, + 394, + 553, + 425 + ], + "type": "text", + "content": "[79] Meng Tian, Marcelo H Ang Jr, and Gim Hee Lee. Shape prior deformation for categorical 6D object pose and size estimation. In ECCV, 2020. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 427, + 555, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 427, + 555, + 470 + ], + "spans": [ + { + "bbox": [ + 320, + 427, + 555, + 470 + ], + "type": "text", + "content": "[80] Jonathan Tremblay, Thang To, Balakumar Sundaralingam, Yu Xiang, Dieter Fox, and Stan Birchfield. Deep object pose estimation for semantic robotic grasping of household objects. In CoRL, pages 306-316, 2018. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 472, + 555, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 472, + 555, + 535 + ], + "spans": [ + { + "bbox": [ + 320, + 472, + 555, + 535 + ], + "type": "text", + "content": "[81] Mikaela Angelina Uy, Vladimir G Kim, Minhyuk Sung, Noam Aigerman, Siddhartha Chaudhuri, and Leonidas J Guibas. Joint learning of 3d shape retrieval and deformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11713-11722, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 537, + 555, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 537, + 555, + 580 + ], + "spans": [ + { + "bbox": [ + 320, + 537, + 555, + 580 + ], + "type": "text", + "content": "[82] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 582, + 555, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 582, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 320, + 582, + 555, + 624 + ], + "type": "text", + "content": "[83] Chen Wang, Danfei Xu, Yuke Zhu, Roberto Martin-Martin, Cewu Lu, Li Fei-Fei, and Silvio Savarese. DenseFusion: 6D object pose estimation by iterative dense fusion. In CVPR, pages 3343–3352, 2019. 1, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 626, + 555, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 626, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 320, + 626, + 555, + 669 + ], + "type": "text", + "content": "[84] Gu Wang, Fabian Manhardt, Jianzhun Shao, Xiangyang Ji, Nassir Navab, and Federico Tombari. Self6d: Self-supervised monocular 6d object pose estimation. In ECCV, pages 108-125. Springer, 2020." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 670, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 670, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 670, + 555, + 713 + ], + "type": "text", + "content": "[85] Gu Wang, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. GDR-Net: Geometry-guided direct regression network for monocular 6D object pose estimation. In CVPR, pages 16611–16621, 2021. 3" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22033" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 62, + 395, + 296, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 395, + 296, + 428 + ], + "spans": [ + { + "bbox": [ + 62, + 395, + 296, + 428 + ], + "type": "text", + "content": "[94] Yang You, Ruoxi Shi, Weiming Wang, and Cewu Lu. CPPf: Towards robust category-level 9d pose estimation in the wild. In CVPR, pages 6866-6875, 2022. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 430, + 296, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 430, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 62, + 430, + 296, + 462 + ], + "type": "text", + "content": "[95] Hao Yu, Fu Li, Mahdi Saleh, Benjamin Busam, and Slobodan Ilic. Cofinet: Reliable coarse-to-fine correspondences for robust pointcloud registration. NeurIPS, 34, 2021. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 464, + 296, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 296, + 507 + ], + "type": "text", + "content": "[96] Hao Yu, Zheng Qin, Ji Hou, Mahdi Saleh, Dongsheng Li, Benjamin Busam, and Slobodan Ilic. Rotation-invariant transformer for point cloud matching. In CVPR, pages 5384-5393, 2023. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 510, + 296, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 510, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 62, + 510, + 296, + 554 + ], + "type": "text", + "content": "[97] Hao Yu, Ji Hou, Zheng Qin, Mahdi Saleh, Ivan Shugurov, Kai Wang, Benjamin Busam, and Slobodan Ilic. Riga: Rotation-invariant and globally-aware descriptors for point cloud registration. IEEE TPAMI, 46(5):3796-3812, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 556, + 296, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 556, + 296, + 599 + ], + "spans": [ + { + "bbox": [ + 62, + 556, + 296, + 599 + ], + "type": "text", + "content": "[98] Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose: Predicting probabilistic relative rotation for single objects in the wild. In ECCV, pages 592-611. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 601, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 601, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 601, + 296, + 656 + ], + "type": "text", + "content": "[99] Ruida Zhang, Yan Di, Zhiqiang Lou, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. Rbp-posed: Residual bounding box projection for category-level pose estimation. In European conference on computer vision, pages 655–672. Springer, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 58, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 658, + 295, + 712 + ], + "type": "text", + "content": "[100] Ruida Zhang, Yan Di, Fabian Manhardt, Federico Tombari, and Xiangyang Ji. Ssp-pose: Symmetry-aware shape prior deformation for direct category-level object pose estimation. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7452–7459. IEEE, 2022." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 58, + 72, + 297, + 115 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 72, + 297, + 115 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 297, + 115 + ], + "type": "text", + "content": "[86] Gu Wang, Fabian Manhardt, Xingyu Liu, Xiangyang Ji, and Federico Tombari. Occlusion-aware self-supervised monocular 6D object pose estimation. IEEE TPAMI, 2024. 1" + } + ] + } + ], + "index": 0 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 58, + 72, + 556, + 407 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 61, + 118, + 295, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 295, + 163 + ], + "type": "text", + "content": "[87] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6D object pose and size estimation. In CVPR, pages 2642-2651, 2019. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 165, + 295, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 165, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 62, + 165, + 295, + 196 + ], + "type": "text", + "content": "[88] Jiaze Wang, Kai Chen, and Qi Dou. Category-level 6D object pose estimation via cascaded relation and recurrent reconstruction networks. In IROS, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 199, + 296, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 199, + 296, + 232 + ], + "spans": [ + { + "bbox": [ + 62, + 199, + 296, + 232 + ], + "type": "text", + "content": "[89] Yue Wang and Justin M Solomon. Deep closest point: Learning representations for point cloud registration. In ICCV, pages 3523-3532, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 235, + 296, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 235, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 62, + 235, + 296, + 266 + ], + "type": "text", + "content": "[90] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. FoundationPose: Unified 6d pose estimation and tracking of novel objects. In CVPR, 2024. 1, 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 269, + 296, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 269, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 269, + 296, + 312 + ], + "type": "text", + "content": "[91] Chin-Hsuan Wu, Yen-Chun Chen, Bolivar Solarte, Lu Yuan, and Min Sun. ifusion: Inverting diffusion for pose-free reconstruction from sparse views. arXiv preprint arXiv:2312.17250, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 315, + 296, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 315, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 315, + 296, + 357 + ], + "type": "text", + "content": "[92] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6D object pose estimation in cluttered scenes. RSS, 2018. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 360, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 360, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 360, + 296, + 392 + ], + "type": "text", + "content": "[93] Jiaqi Yang, Qian Zhang, Yang Xiao, and Zhiguo Cao. Toldi: An effective and robust approach for 3d local shape description. PR, 65:175-187, 2017. 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[101] Ruida Zhang, Ziqin Huang, Gu Wang, Chenyangguang Zhang, Yan Di, Xingxing Zuo, Jiwen Tang, and Xiangyang Ji. Lapose: Laplacian mixture shape modeling for rgb-based category-level object pose estimation. In ECCV, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 555, + 183 + ], + "type": "text", + "content": "[102] Ruida Zhang, Chenyangguang Zhang, Yan Di, Fabian Manhardt, Xingyu Liu, Federico Tombari, and Xiangyang Ji. Kp-red: Exploiting semantic keypoints for joint 3d shape retrieval and deformation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20540-20550, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 185, + 556, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 556, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 556, + 217 + ], + "type": "text", + "content": "[103] Xiyu Zhang, Jiaqi Yang, Shikun Zhang, and Yanning Zhang. 3d registration with maximal cliques. In CVPR, pages 17745–17754, 2023. 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 219, + 554, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 554, + 251 + ], + "type": "text", + "content": "[104] Chen Zhao, Yinlin Hu, and Mathieu Salzmann. Fusing local similarities for retrieval-based 3d orientation estimation of unseen objects. In ECCV, pages 106-122. Springer, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 253, + 554, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 253, + 554, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 253, + 554, + 285 + ], + "type": "text", + "content": "[105] Chen Zhao, Tong Zhang, and Mathieu Salzmann. 3d-aware hypothesis & verification for generalizable relative object pose estimation. In ICLR, 2023. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 286, + 554, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 554, + 319 + ], + "type": "text", + "content": "[106] Chen Zhao, Tong Zhang, Zheng Dang, and Mathieu Salzmann. Dvmnet: Computing relative pose for unseen objects beyond hypotheses. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 320, + 554, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 320, + 554, + 353 + ], + "spans": [ + { + "bbox": [ + 317, + 320, + 554, + 353 + ], + "type": "text", + "content": "[107] Xu Zhao, Wenchao Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, and Jinqiao Wang. Fast segment anything. arXiv preprint arXiv:2306.12156, 2023. 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 354, + 555, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 555, + 407 + ], + "type": "text", + "content": "[108] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splattering: Fast and generalizable single-view 3d reconstruction with transformers. In CVPR, pages 10324-10335, 2024. 8" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22034" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/UPME_ An Unsupervised Peer Review Framework for Multimodal Large Language Model Evaluation/42fa34eb-b512-4300-b1c0-40e45c00118f_content_list.json b/2025/UPME_ An Unsupervised Peer Review Framework for Multimodal Large Language Model Evaluation/42fa34eb-b512-4300-b1c0-40e45c00118f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e293322383a19338dd2a1782b2bf7a7fa4117f8e --- /dev/null +++ b/2025/UPME_ An Unsupervised Peer Review Framework for Multimodal Large Language Model Evaluation/42fa34eb-b512-4300-b1c0-40e45c00118f_content_list.json @@ -0,0 +1,1906 @@ +[ + { + "type": "text", + "text": "UPME: An Unsupervised Peer Review Framework for Multimodal Large Language Model Evaluation", + "text_level": 1, + "bbox": [ + 223, + 130, + 776, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qihui Zhang $^{1,3*}$ Munan Ning $^{1*}$ Zheyuan Liu $^{1,3}$ Yue Huang $^{4}$ Shuo Yang $^{1}$ Yanbo Wang $^{1}$ Jiayi Ye $^{1}$ Xiao Chen $^{5}$ Yibing Song $^{2,3}$ Li Yuan $^{1\\dagger}$", + "bbox": [ + 176, + 200, + 818, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Electrical and Computer Engineering, Peking University $^{2}$ Hupan Lab $^{3}$ DAMO Academy, Alibaba Group $^{4}$ University of Notre Dame $^{5}$ Tsinghua University", + "bbox": [ + 150, + 253, + 844, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimodal Large Language Models (MLLMs) have emerged to tackle the challenges of Visual Question Answering (VQA), sparking a new research focus on conducting objective evaluations of these models. Existing evaluation methods face limitations due to the significant human workload required to design Q&A pairs for visual images, which inherently restricts the scale and scope of evaluations. Although automated MLLM-as-judge approaches attempt to reduce the human workload through automatic evaluations, they often introduce biases. To address these problems, we propose an Unsupervised Peer review MLLM Evaluation framework. It utilizes only image data, allowing models to automatically generate questions and conduct peer review assessments of answers from other models, effectively alleviating the reliance on human workload. Additionally, we introduce the vision-language scoring system to mitigate the bias issues, which focuses on three aspects: (i) response correctness; (ii) visual understanding and reasoning; and (iii) image-text correlation. Experimental results demonstrate that UPME achieves a Pearson correlation of 0.944 with human evaluations on the MMstar dataset and 0.814 on the ScienceQA dataset, indicating that our framework closely aligns with human-designed benchmarks and inherent human preferences.", + "bbox": [ + 89, + 358, + 483, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 738, + 220, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, Large Language Models (LLMs) have made significant strides in reasoning and application capabilities [3, 32, 47], enabling diverse applications such as text and code generation [1, 19, 40, 48]. Notably, due to the exceptional natural language understanding capabilities of LLMs, they are increasingly employed for automated evaluations, a pro", + "bbox": [ + 89, + 763, + 482, + 854 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2bf165477525d2a24cc964c43cc7b98f431535217d020b712f2136a2a678a6b8.jpg", + "image_caption": [ + "Figure 1. Existing methods for evaluating MLLMs face various challenges. Our proposed UPME framework addresses these limitations by leveraging a peer review mechanism, reducing annotation costs, and aligning closely with human judgment." + ], + "image_footnote": [], + "bbox": [ + 526, + 335, + 895, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cess known as LLM-as-a-Judge [55]. With the integration of vision encoders [36], Multimodal Large Language Models (MLLMs) [24, 32] can incorporate multiple modalities (e.g., text, image, and audio) [23, 56] and showcase remarkable performance in downstream applications, including visual conversation [7] and embodied intelligence [16]. Furthermore, they can act as agents to interact with GUI [9, 18].", + "bbox": [ + 511, + 628, + 906, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Subsequently, the challenge of conducting objective and comprehensive evaluations of MLLMs while providing users with authoritative guidelines has emerged as a key research focus. A common approach involves constructing benchmarks based on Visual Question Answering (VQA), which assess the performance of MLLMs across various dimensions by setting image-based questions and answers [10, 28, 52, 53]. However, most of these works heavily depend on substantial human involvement in creating Q&A pairs for images [21, 44], leading to limited evaluation volume and making it insufficient for a comprehen", + "bbox": [ + 511, + 734, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution † Corresponding author {maskhui1003, munanning}@gmail.com yuanli-ece@pu.edu.cn", + "bbox": [ + 89, + 862, + 372, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9165", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sive assessment of visual content [6]. Another category of methods is MLLM-as-a-Judge [17, 20, 26, 55], which aims to reduce reliance on human annotations by allowing models to evaluate each other directly. However, these methods often still depend on human-designed questions and cannot entirely eliminate the need for human input. Additionally, a significant challenge is the presence of morbidity and self-preference biases [8, 51], where MLLMs tend to favor longer responses or prefer their own outputs. These biases can lead mutual evaluation mechanisms to deviate from truly understanding visual content.", + "bbox": [ + 89, + 90, + 480, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose UPME, the Unsupervised Peer review MLLM Evaluation framework, which is designed to evaluate MLLMs without requiring human QA annotations, as shown in Figure 1. During each iteration, UPME selects two candidate models and a review model from the MLLM pool. The review model generates questions for a given image and evaluates the responses provided by the candidate models. This evaluation is conducted using a vision-language scoring system that assesses textual response correctness, visual understanding and reasoning, as well as image-text consistency through CLIP scores. Each model is initialized with a confidence weight and alternates roles as candidate and review models. Their estimated scores are calculated based on weighted evaluations provided by the review models. The framework employs dynamic weight optimization, iteratively updating model scores and refining weights through consistency optimization. Ultimately, this process generates a comprehensive and less-biased estimated score list.", + "bbox": [ + 91, + 258, + 482, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We selected the MMStar [10] and ScienceQA [28] datasets as benchmarks because they encompass visual content reasoning and multimodal content across various disciplines. Experimental results show that under unsupervised settings, UPME achieved Pearson correlations of 0.944 and 0.814 with human evaluation results on these two datasets, while Spearman similarities reached 0.972 and 0.886, indicating a high degree of similarity between our approach and human-annotated QA benchmarks. Further experiments demonstrated that UPME achieved higher consistency with human assessments compared to purely peer review-based methods, effectively reducing morbidity and self-preference issues in the MLLM-as-a-Judge framework.", + "bbox": [ + 89, + 547, + 482, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, our contributions can be summarized as follows:", + "bbox": [ + 109, + 744, + 480, + 760 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the first Unsupervised Peer review MLLM Evaluation (UPME) framework, addressing the challenge of reliance on human annotations in MLLM evaluation", + "- Our framework incorporates a Vision-Language Scoring System that targets visual performance and image-text association, providing a visually-focused evaluation.", + "- We conduct extensive experiments demonstrating that UPME achieves high consistency with human QA assessments and strong alignment with human preferences." + ], + "bbox": [ + 89, + 761, + 482, + 897 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 653, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. MLLM Benchmark", + "text_level": 1, + "bbox": [ + 513, + 114, + 702, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "MLLM benchmarks rely on human-annotated QA pairs to measure the accuracy of MLLMs in answering questions. They focus on the different capabilities of MLLMs. ScienceQA [28] evaluate the multi-turn reasoning ability for a broad range of scientific subjects by introducing Chain of Thought (CoT) [47]; MMVet [52] presents an LLM-based open-ended output evaluation method, enabling unified scoring across various question types and answer formats; MMMU [53], inspired by MMLU, primarily assesses model performance in complex reasoning tasks requiring college-level knowledge, covering six major disciplines and 30 fields with multimodal questions; MMStar [10] focuses on creating a benchmark requiring visual content for reasoning, rigorously selecting visual-dependent samples and detecting data leakage to ensure genuine performance gains from multimodal training; RealWorldQA [49] is designed to assess the ability of MLLMs to understand real-world space. Although these benchmarks propose various evaluation methods, their evaluation scope is limited by the reliance on substantial human annotation.", + "bbox": [ + 511, + 137, + 906, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. MLLM-as-a-Judge", + "text_level": 1, + "bbox": [ + 513, + 450, + 696, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To reduce the human annotation workload, LLM-as-a-Judge [55] is proposed to enable models to perform mutual evaluation. This approach primarily involves using LLMs to assess responses in two ways: comparing pairs of answers to determine superiority [20, 25, 55], or directly scoring individual answers based on specific criteria [17, 25]. This strategy is extended to MLLM-as-a-Judge [8] to assess the performance of MLLMs across multiple evaluation tasks. MJ-Bench [11] evaluates the effectiveness of multimodal reward models in text-to-image generation. However, these methods face two main biases [38, 51]: the verbose bias and self-preference bias, which means MLLMs prefer the answers from themselves or with longer text length. Human preferences have also not been effectively incorporated into these automated evaluation systems [31].", + "bbox": [ + 511, + 474, + 906, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Human Preference-based Evaluation", + "text_level": 1, + "bbox": [ + 513, + 710, + 831, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Predefined image and QA-pair benchmarks for MLLMs may encounter issues like data leakage and targeted optimization, which can prevent an accurate reflection of the true capabilities of these models [15, 50]. Therefore, evaluation methods based on human preferences offer a better assessment of whether MLLMs meet human needs [30]. Chatbot Arena [12] provides a public platform specifically for evaluating LLMs based on human preference. This platform uses a pairwise comparison approach, where users interact with two anonymous models and select the one that performs better, generating a dynamic leaderboard via the", + "bbox": [ + 511, + 734, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9166", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c665188dc9e4d133a0633e2a63266255a6c89241ccf47717078e05c42661f46e.jpg", + "image_caption": [ + "Figure 2. The UPME framework consists of three main components: (i) Peer Review Mechanism, where two candidate models and one review model are randomly selected from the MLLM pool. The review model generates questions based on a selected image, and candidate models provide responses. (ii) Vision-Language Judgment Scoring System, which evaluates answers based on textual correctness, visual understanding and reasoning, and image-text correlation. (iii) Dynamic Weight Optimization, ensuring consistency between confidence weights and estimated scores through iterative optimization cycles." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 906, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Elo rating system. Although open-ended user evaluations can more realistically reflect LLM performance in real-world scenarios [46], the high economic cost of collecting large volumes of high-quality user feedback, combined with the long wait time for obtaining such feedback, limits the widespread adoption of this approach.", + "bbox": [ + 88, + 448, + 480, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 550, + 225, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Assigning higher weights to better-performing models leads to a significant improvement in the accuracy of the entire evaluation framework. This conclusion is supported by preliminary experiments detailed in subsection 8.1 and confirmed by recent research [13, 31, 41, 43]. Therefore, we focus on characterizing the relationship between model performance and its judging capability through consistency optimization.", + "bbox": [ + 89, + 575, + 482, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Peer review formulation", + "text_level": 1, + "bbox": [ + 89, + 705, + 312, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem definition. The UPME framework facilitates mutual evaluation among MLLMs via a peer review mechanism, aiming to produce evaluation scores that closely reflect human preferences. Given an unsupervised image dataset $\\mathcal{I} = \\{I_i\\}_{i=1}^n$ of $n$ images (with no human annotated questions and answers) and a model pool $\\mathcal{M} = \\{M_j\\}_{j=1}^m$ consisting of $m$ MLLMs (both open-source and closed-source), we can obtain an UPME-learned score list $\\hat{G}$ :", + "bbox": [ + 89, + 727, + 482, + 849 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {G} := \\left[ \\hat {G} _ {M _ {1}}, \\hat {G} _ {M _ {2}}, \\hat {G} _ {M _ {3}}, \\dots , \\hat {G} _ {M _ {m}} \\right]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 861, + 482, + 880 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "On the other hand, given $G_{M_1}$ represents the score of", + "bbox": [ + 109, + 885, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "model $M_{1}$ on the human-annotated benchmark, we define the human-annotated score list $G$ as:", + "bbox": [ + 511, + 448, + 906, + 477 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG := \\left[ G _ {M _ {1}}, G _ {M _ {2}}, G _ {M _ {3}}, \\dots , G _ {M _ {m}} \\right]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 496, + 906, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our objective is to optimize the UPME-learned scores $\\hat{G}$ by maximizing their similarity to the human-annotated score list $G$ , i.e., $\\mathrm{Sim}(G, \\hat{G}) \\to 1$ , ensuring that UPME-learned scores align closely with human benchmark standards.", + "bbox": [ + 511, + 522, + 905, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview of UPME. Figure 2 shows an overview of our UPME framework. During each review iteration, we randomly select two different candidate models $M_{j}$ and $M_{k}$ from the pool $\\mathcal{M}$ for comparison, along with another model $M_{r}$ to serve as the reviewer. $M_{r}$ generates a question $Q_{i}^{r} = M_{r}(I_{i})$ for each image $I_{i}$ and then prompts the candidate models $M_{j}$ and $M_{k}$ to provide responses $A_{i}^{j,r}$ and $A_{i}^{k,r}$ , respectively. The answers can be written as:", + "bbox": [ + 511, + 622, + 906, + 747 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nA _ {i} ^ {j, r} = M _ {j} \\left(I _ {i}, Q _ {i} ^ {r}\\right), \\quad A _ {i} ^ {k, r} = M _ {k} \\left(I _ {i}, Q _ {i} ^ {r}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 758, + 906, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on the vision-language judgment scoring System $S_{VL}$ (detailed in subsection 3.2), $M_r$ then evaluates these responses, and compares the responses of $M_j$ and $M_k$ to produce a review score $\\mathrm{Review}_i^{j,k,r}$ as:", + "bbox": [ + 511, + 792, + 905, + 856 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {R e v i e w} _ {i} ^ {j, k, r} = S _ {V L} \\left(A _ {i} ^ {j, r}, A _ {i} ^ {k, r}, Q _ {i} ^ {r}, I _ {i} \\mid M _ {r}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 882, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9167", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By assigning each model an initiated confidence weight $w$ , we can calculate the learned score $\\hat{G}_{M_j}$ of each MLLM:", + "bbox": [ + 89, + 90, + 482, + 125 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {G} _ {M _ {j}} = \\sum_ {i} \\sum_ {k \\neq j} \\sum_ {r \\neq k, r \\neq j} \\operatorname {R e v i w} _ {i} ^ {j, k, r} \\times w _ {r}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 132, + 482, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "then we can subsequently obtain the UPME-learned score list $\\hat{G}$ , as illustrated in Equation 1.", + "bbox": [ + 89, + 176, + 482, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Vision-Language Judgment Scoring System", + "text_level": 1, + "bbox": [ + 89, + 215, + 460, + 232 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this subsection, we present our vision-language judgment scoring system, denoted as $S_{\\mathrm{VL}}$ in Equation 9. Our evaluation is structured around three core criteria: Response Correctness, Visual Understanding and Reasoning, and Image-Text Correlation. The first two aspects are evaluated by employing $M_r$ , which reviews responses based on prompts outlined in Section 10. The Image-Text Correlation is measured using CLIP scores.", + "bbox": [ + 89, + 237, + 483, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Response Correctness. The primary criterion is the correctness of the candidate models' responses to questions posed by the review model. Existing studies indicate that paired comparative evaluation yields better accuracy than isolated assessments [26, 27], so we adopt a pairwise scoring approach. Based on the review model $M_r$ 's evaluation $\\mathrm{Judge}_i^r$ for candidates $M_j$ and $M_k$ , we calculate the model's score for Response Accuracy $S_{\\mathrm{Correct}}(j)$ as follows:", + "bbox": [ + 89, + 369, + 483, + 492 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {\\text {C o r r e c t}} (j) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f J u d g e} _ {\\text {C o r r e c t}, j} > \\text {J u d g e} _ {\\text {C o r r e c t}, k} \\\\ 0. 5 & \\text {i f J u d g e} _ {\\text {C o r r e c t}, j} = \\text {J u d g e} _ {\\text {C o r r e c t}, k} \\\\ 0 & \\text {i f J u d g e} _ {\\text {C o r r e c t}, j} < \\text {J u d g e} _ {\\text {C o r r e c t}, k}, \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 501, + 482, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{Judge}_{\\mathrm{Correct}}$ represents the review model's assessment of response correctness. If $\\mathrm{Judge}_{\\mathrm{Correct},j} > \\mathrm{Judge}_{\\mathrm{Correct},k}$ , it indicates that model $M_{j}$ is judged superior in response accuracy to $M_{k}$ , scoring 1 point; a tie yields 0.5 points, and a lower score results in 0 points.", + "bbox": [ + 89, + 566, + 483, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visual Understanding and Reasoning. Inspired by the significant CoT ability demonstrated in the o1 model [34] through adversarial training, we leverage $M_r$ to evaluate the interpretative and analytical capabilities of the model with respect to visual content.", + "bbox": [ + 89, + 654, + 482, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$S_{\\mathrm{Visual}}(j)$ is computed using the function $\\Gamma$ , which evaluates the response across four dimensions: caption, reasoning, grounding, and relationship, as follows:", + "bbox": [ + 89, + 729, + 482, + 776 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {\\text {V i s u a l}} (j) = \\Gamma (\\text {c a p . , r e a . , g r o . , o b j .} \\mid A _ {j}, M _ {r}), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 797, + 482, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where captioning evaluates the ability to generate precise descriptions, reasoning measures logical consistency, grounding assesses accurate object localization, and relationship captures the connections between subjects. Within this structure, the review model $M_r$ provides a combined", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "score via $\\Gamma$ , yielding $S_{\\mathrm{Visual}}(j)$ . $S_{\\mathrm{Visual}}(j)$ reflects a model's capability to go beyond basic VQA, integrating both description and reasoning to demonstrate a deeper level of visual comprehension.", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image-Text Correlation. To capture the alignment between image content and textual responses, we introduce the CLIP [37] score:", + "bbox": [ + 511, + 164, + 905, + 209 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {\\text {C l i p}} (j) = \\operatorname {C L I P} \\left(I _ {i}, A _ {i} ^ {j, r}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 220, + 903, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $S_{\\mathrm{Clip}}(j)$ measures the correlation between image $I_{i}$ and response $A_{i}^{j,r}$ from model $M_{j}$ . The details are described in subsection 9.2.", + "bbox": [ + 511, + 250, + 905, + 295 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Combining the above components, we compute the final vision-language judgment score as a weighted sum of the three criteria: Response Accuracy, Visual Understanding and Reasoning, and Image-Text Correlation:", + "bbox": [ + 511, + 297, + 905, + 357 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {\\mathrm {V L}} = \\gamma_ {1} S _ {\\text {C o r r e c t}} + \\gamma_ {2} S _ {\\text {V i s u a l}} + \\gamma_ {3} S _ {\\text {C l i p}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 369, + 903, + 387 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma_{1},\\gamma_{2}$ , and $\\gamma_{3}$ are weights reflecting the relative importance of each component in the overall evaluation.", + "bbox": [ + 511, + 398, + 905, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Dynamic Weight Optimization", + "text_level": 1, + "bbox": [ + 511, + 436, + 784, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In each iteration, we use Mean Squared Error (MSE) as the loss function to update the weights $w$ , enhancing the consistency between the weights $w$ and the estimated scores $\\hat{G}$ :", + "bbox": [ + 511, + 459, + 905, + 505 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M S E}} (\\hat {G}, w) = \\frac {1}{m} \\sum_ {j = 1} ^ {m} \\left(\\hat {G} _ {M _ {j}} - w _ {M _ {j}}\\right) ^ {2}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 516, + 903, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The entire optimization process is dynamic. After each round of weight updates, the weights and scores continue to change. The loss function computes new scores based on the updated weights, which are then iteratively used to optimize the weights further, forming a continuous iterative optimization loop, as detailed in Algorithm 1.", + "bbox": [ + 511, + 568, + 905, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 672, + 643, + 689 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 696, + 702, + 713 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets. To validate the effectiveness of our framework, we conduct experiments on two distinctly styled datasets: ScienceQA [28] and MMStar [10]. ScienceQA underscores the importance of CoT in scientific reasoning tasks, comprising approximately 21,208 multimodal multiple-choice questions covering various science topics. MMStar focuses on evaluating the multimodal capabilities of MLLMs, and consists of 1,500 meticulously curated samples including 6 core capabilities and 18 detailed axes. Each sample exhibits visual dependency and minimal data leakage.", + "bbox": [ + 511, + 719, + 905, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Candidate MLLMs. Our MLLM pool includes 5 closed-source models and 1 open-source model. The open-source", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9168", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/10a7ce8fa8ec0616e72611fe4ac31b8c446bc99ba8f6c1b1625a96601a76a5b9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsMMstarScienceQA
Pearson (↑)Spearman (↑)Per.Ent. (↓)Pearson (↑)Spearman (↑)Per.Ent. (↓)
LLama-3.2-11b-V0.314±0.07570.550±0.05770.983±0.23100.160±0.04300.225±0.09571.099±0.0000
Claude-3-haiku0.095±0.13010.225±0.15001.099±0.0000-0.145±0.0308-0.525±0.20611.099±0.0000
Claude-3.5-sonnet0.780±0.01460.825±0.05000.752±0.23100.437±0.04520.450±0.17320.752±0.2310
Gemini-1.5-pro0.864±0.00680.850±0.05770.637±0.00000.4147±0.04640.725±0.20620.434±0.5352
GPT-4o-mini0.668±0.00730.725±0.12580.868±0.28860.3354±0.06350.600±0.00001.099±0.0000
GPT-4o0.878±0.00380.875±0.00500.159±0.31830.617±0.00710.625±0.12580.637±0.0000
Methods
Peer Review0.725±0.00440.771±0.16161.040±0.28300.463±0.01930.686±0.17771.040±0.2830
Majority Vote [42]0.757±0.00130.757±0.08571.299±0.17330.509±0.01810.524±0.06601.040±0.0000
Rating Vote [2]0.795±0.00130.743±0.23090.628±0.07550.623±0.00840.629±0.18950.920±0.2387
PRD [22]0.838±0.00270.864±0.03170.427±0.00870.636±0.00420.694±0.07340.746±0.0016
UPME0.944±0.00110.972±0.03300.141±0.28120.814±0.00240.886±0.02860.422±0.2812
", + "bbox": [ + 158, + 88, + 849, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Main results of different models and methods on MMstar and ScienceQA datasets, evaluated by Pearson, Spearman, and Permutation Entropy metrics. Since the model cannot judge itself, scores from the other five models involved in the computations are used for evaluation. Optimal results are highlighted in bold and underlined. Standard deviation (std) is calculated based on four runs.", + "bbox": [ + 89, + 327, + 906, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "model is Llama-3.2-11b-vision-instruct [29], which adapts the Llama-3 architecture to the multimodal domain through visual instruction tuning [24]. The closed-source models consist of GPT-4o [35], GPT-4o-mini [33], Claude-3.5-Sonnet [3], Claude-3-Haiku [4], and Gemini-1.5-Pro [45]. These commercial models employ enhanced human alignment strategies to optimize performance and user experience across various applications.", + "bbox": [ + 91, + 396, + 482, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. For all experiments, we employ three popular metrics to evaluate the setups mentioned above and our UPME method: the Pearson Correlation Coefficient [14] to evaluate the similarity of score lists, Spearman's Rank Correlation Coefficient [54] to measure the correlation of ranking lists, and Permutation Entropy [5] to assess the complexity and unpredictability of the scoring distributions. All experiments are performed for 4 runs with different seeds (seed = 1, 2, 3, 4) and record the average results.", + "bbox": [ + 89, + 517, + 482, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison Methods. Since we are the first to investigate unsupervised peer review in MLLM evaluations, we choose the original peer review mechanism and two popular classical methods, Majority Vote [42] and Rating Vote [2], for our comparative analysis. We also adopt a semi-supervised method, PRD [22], and conduct experiments using individual models as judges to further bolster our comparisons.", + "bbox": [ + 89, + 652, + 482, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Experimental Results", + "text_level": 1, + "bbox": [ + 89, + 768, + 292, + 784 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1. Comparison Results", + "text_level": 1, + "bbox": [ + 89, + 791, + 277, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The experimental results in Table 1 indicate that among the candidate MLLMs, GPT-4o exhibits the highest review capability on the MMstar dataset, achieving Pearson and Spearman correlation coefficients of 0.878 and 0.875, respectively. In contrast, the open-source LLama-3.2 model demonstrates weaker performance. Although Claude-3.5-", + "bbox": [ + 89, + 809, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5eb990f3a52d71235841e3e6a9dae03857815000f623146822506a110345d56a.jpg", + "image_caption": [ + "Figure 3. Convergence experiments." + ], + "image_footnote": [], + "bbox": [ + 524, + 398, + 890, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sonnet and Gemini-1.5-Pro claim to offer performance comparable to GPT-4o, they still show limitations when evaluating other models, falling noticeably short of human-level performance. The original peer review method yields Pearson and Spearman coefficients lower than that of GPT-4o, indicating that the biases of weaker models influence the Peer Review approach. Our UPME method achieves a Pearson correlation of 0.944 and a Spearman correlation of 0.972 on the MMstar dataset, demonstrating a high degree of similarity with human annotations and validating the effectiveness of our optimizations in visual QA comprehension and consistency evaluation.", + "bbox": [ + 511, + 612, + 906, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "On the ScienceQA dataset, even the best-performing GPT-4o achieves Pearson and Spearman correlations of 0.617 and 0.625, reflecting the difficulty of the task. The original peer review method shows more modest results, with Pearson and Spearman of 0.463 and 0.686, respectively. Following a series of optimizations and alignments, UPME achieves Pearson and Spearman correlations", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9169", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c82fec597a9ac77ac103e17938b624e8c8fe06ee19e6546ac79c5a2615255233.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MMstarScienceQA
MethodCor.Vis.Cti.PearsonSpearmanPearsonSpearman
Peer Review0.7270.7140.4570.600
Score Optimization✓0.8540.7710.7130.657
Visual Alignment✓0.8730.8860.7010.771
✓0.7850.8290.5480.771
✓✓0.9030.9430.7750.886
UPME✓✓✓0.9440.9720.8140.886
", + "bbox": [ + 94, + 88, + 482, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Ablation study with Pearson and Spearman similarity metrics for each dataset. Optimal results are highlighted in bold and underlined. 'Cor.' stands for correctness, 'Vis.' stands for visual understanding, and 'Cli.' stands for Clip.", + "bbox": [ + 89, + 241, + 482, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of 0.814 and 0.886, showcasing its robustness and superior performance in complex datasets. While the performance on ScienceQA may not be as high, it is important to note that traditional datasets like ScienceQA might not fully align with the evaluation needs of current MLLMs because they may lack visual dependency or sufficient discrimination [10, 52]. UPME achieves higher Pearson and Spearman correlations, demonstrating its robustness and superior performance.", + "bbox": [ + 88, + 329, + 482, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Moreover, as shown in Figure 3, the loss consistently decreases, and the Pearson similarity metric increases over epochs in 64 different initial settings, which demonstrates the convergence of $w$ and $\\hat{G}$ . Under the current experimental setup (25 images and 1500 judgments), the framework reliably converges within the 30-epoch limit.", + "bbox": [ + 89, + 472, + 483, + 563 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 588, + 243, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the ablation study presented in Table 2, we evaluated the impact of progressively introducing different components on model performance. The Score Optimization phase showed consistent improvements in Pearson and Spearman correlation coefficients as modules were added. For instance, using only the correctness module resulted in Pearson scores of 0.854 for MMstar and 0.713 for ScienceQA. Introducing the visual understanding module alone yielded scores of 0.873 and 0.701, respectively. Adding the Clip component alone gave scores of 0.785 for MMstar and 0.548 for ScienceQA. Combining visual understanding and Clip modules further improved performance to 0.903 and 0.775. Ultimately, integrating all components within the UPME framework resulted in the highest performance, with Pearson and Spearman correlations of 0.944 and 0.972 on MMstar and 0.814 and 0.886 on ScienceQA. These results confirm that optimizing the scoring mechanism and visual alignment significantly boosts evaluation accuracy and robustness across different datasets.", + "bbox": [ + 91, + 613, + 482, + 898 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/73e9364cdc08f25fb677a1f3d656d596947032938587d4d51501487e4168d6d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 90, + 903, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cc6be181d7516aeafea998fd44d2af27f74ac0573fa1603eccdb78fe39fb8c04.jpg", + "image_caption": [ + "Figure 4. The performance of UPME in different sample size." + ], + "image_footnote": [], + "bbox": [ + 517, + 202, + 903, + 313 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.3.Impact of Sample Size", + "text_level": 1, + "bbox": [ + 511, + 359, + 715, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Figure 4, we analyze the impact of sample size on the alignment with human-annotated QA benchmarks, measured through Pearson and Spearman correlation coefficients, as well as Permutation Entropy. As the sample size increases from 5 to 25, there is a clear improvement in both correlation metrics and a decline in Permutation Entropy, indicating that the generated score lists and rank lists show greater alignment with human evaluations. Beyond a sample size of 25, the metrics stabilize, showing minimal changes with further increases from 25 to 100 in sample size. This stabilization suggests that the consistency of the generated outputs relative to human judgments has reached a satisfactory level, making 25 samples a reliable point for evaluation. Therefore, we primarily selected a sample size of 25 for our experiments, as it provides an effective balance between alignment accuracy and data efficiency.", + "bbox": [ + 511, + 378, + 906, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Human Preference Alignment Analysis", + "text_level": 1, + "bbox": [ + 511, + 633, + 864, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The aforementioned experiments primarily reflect the consistency between the UPME framework and the predefined QA-based benchmarks. To further explore the alignment between UPME and human preferences, as well as to examine the two types of biases inherent in MLLM-as-Judge approaches, we selected a subset of responses for human evaluation and analysis of UPME scoring results. The detailed experimental design is provided in subsection 9.4.", + "bbox": [ + 511, + 657, + 906, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Alignment with Human Preference", + "text_level": 1, + "bbox": [ + 511, + 787, + 818, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate whether our proposed UPME method can align with human preferences without manual labeling, we conducted experiments on the MMstar and ScienceQA datasets. The experimental results on Table 3 indicate that the baseline method exhibited relatively low human agreement and model consistency rates, suggesting that the Peer", + "bbox": [ + 511, + 810, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9170", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6b59d2425fda7abd3e5b29e4c063a4c28c08ed5a7168ac3decf25a6e70965bd3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetMethodAgreement (%)Consistency (%)
MMstarPeer Review71.167.5
UPME95.989.8
ScienceQAPeer Review68.261.8
UPME87.482.6
", + "bbox": [ + 91, + 88, + 480, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8b207c8b62b9f05df7ca8f3bb4836617aa410c51583b90498226f094c3398ce0.jpg", + "image_caption": [ + "Figure 5. Model accuracy comparison in peer review framework w/ and w/o UPME, where Peer Review_Cor. represents the correctness of the original peer review, and UPME_Cor. and UPME_Vis. correspond to the two judgment dimensions of response correctness and visual understanding, introduced in subsection 3.2.", + "Figure 6. Heatmap and Table for Peer Review and UPME on Self Preference and Verbosity Bias." + ], + "image_footnote": [], + "bbox": [ + 96, + 227, + 478, + 448 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Review mechanism under an unsupervised setting without weight optimization struggles to align with human preferences. In contrast, UPME demonstrated significant improvements by incorporating Correctness, Visual Understanding, and Clip Correlation. On the MMStar, UPME achieved an agreement rate of $95.9\\%$ and a consistency rate of $89.8\\%$ , showing that the optimized scoring criteria significantly enhance the accuracy of evaluation outputs and alignment with human preferences. By capturing key multimodal understanding metrics without relying on manual labeling, UPME effectively achieved high consistency with human annotations, highlighting its substantial advantages in improving response consistency and accuracy under an unsupervised framework.", + "bbox": [ + 88, + 559, + 482, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Performance w/ and w/o UPME", + "text_level": 1, + "bbox": [ + 89, + 782, + 372, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To assess the improvement in model judgment accuracy within the Peer Review mechanism and the UPME framework, we conducted an experiment, the results are displayed in Figure 5. The calculation of accuracy is based on $ACC = \\frac{\\mathrm{judge}_{\\mathrm{correct}}}{\\mathrm{judge}_{\\mathrm{all}}}$ . Experimental results show that the UPME significantly improves the accuracy of models in", + "bbox": [ + 89, + 806, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cec5f000ca8beaf39c46270dc76f11a7523700688f4ef471fce2898193d490e1.jpg", + "image_caption": [ + "(a) Verbosity Bias", + "(b) Self Preference" + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 906, + 164 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2f9c09aa15170d0f82a7aa5f1fcd3bf3d5542b92b3f702825fd67630781fad62.jpg", + "table_caption": [ + "Table 3. Human Agreement (%) and Human Consistency (%) Rates on Various Datasets." + ], + "table_footnote": [], + "table_body": "
Bias TypeMethodChi-Squarep-valuePhi Coefficient
VerbosityPeer review10.9960.000910.135
BiasUPME0.2800.596960.022
SelfPeer review39.5843.142e-100.257
PreferenceUPME3.4890.06180.076
", + "bbox": [ + 517, + 200, + 905, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "both correctness and visual understanding compared to the original correctness judgment. Specifically, the average accuracy in correctness evaluation increased from $61.56\\%$ in the baseline to $74.48\\%$ with UPME, an improvement of $12.92\\%$ . This indicates a significant advantage of UPME in assessing the correctness of model responses. Furthermore, in the evaluation of visual understanding, UPME achieved an average accuracy of $73.93\\%$ , which is also notably higher than the baseline. This improvement is primarily attributed to UPME's unsupervised evaluation framework, which employs a multi-model peer review mechanism that combines correctness and visual understanding scores to achieve a more efficient and accurate assessment.", + "bbox": [ + 511, + 358, + 906, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Mitigating the Bias of MLLM-as-a-Judge", + "text_level": 1, + "bbox": [ + 511, + 566, + 867, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mitigation of Verbosity bias. As shown in Figure 6, the verbosity-preference bias is significantly alleviated by our method. The original Peer Review approach had a p-value of 0.00091 ( $< 0.05$ ), with a Chi-square value of 10.996 and a Phi coefficient of 0.135, suggesting a significant preference for verbose responses during evaluation. In contrast, the UPME framework yielded a p-value of 0.59696 ( $>0.05$ ), with the Chi-square value reduced to 0.280 and the Phi coefficient decreased to 0.022, indicating that it does not exhibit a significant bias toward selecting verbose responses. Additionally, the number of incorrect selections of verbose responses dropped from 233 to 103 in the UPME framework. Thus, the UPME approach effectively mitigates the bias toward verbose responses in models, enhancing evaluation diversity and alignment with human preferences in an unsupervised setting.", + "bbox": [ + 511, + 589, + 906, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mitigation of Self-preference. Self-preference bias is also significantly alleviated, as shown in Figure 6. In the original Peer Review method, the Chi-square value was 39.584, the Phi coefficient was 0.257, and the p-value was", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9171", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/54aa3c0f1f355d62b9495a9952c98b432f4ff591c547c77f5f6a0d01c519bd03.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 107, + 102, + 290, + 218 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Original Question: Where is the seated person visible in the image?", + "bbox": [ + 295, + 99, + 614, + 111 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Disability of review model", + "bbox": [ + 732, + 99, + 875, + 111 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Options: A: On the fence, B: In the field, C: Outside the field, D: Behind the bench", + "bbox": [ + 297, + 112, + 687, + 121 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Candidate 1 (Gemini-1.5-pro): \"...Therefore, the seated person is visible outside the field. [[C]]\"", + "bbox": [ + 297, + 122, + 754, + 133 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Candidate 2 (Claude-3.5-sonnet): \"...The closest answer would be that the player (though not seated) is in the field. [[B]]\"", + "bbox": [ + 297, + 133, + 870, + 143 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Review model (Claude-3-haiku) [Correctness judgment] \"...However, Assistant B's response correctly concludes that there is no seated person visible in the image. [[2]]\" (Human judgement: [[1]])", + "bbox": [ + 297, + 145, + 893, + 166 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Question generated by review model: What position is the player in the image playing?", + "bbox": [ + 297, + 167, + 707, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Candidate 1 (Gemini-1.5-pro): \"... The player is on a baseball field...\"", + "bbox": [ + 297, + 178, + 633, + 188 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Candidate 2 (Claude-3.5-sonnet) : \"...Overall theme: The image shows a baseball player in action during a game...\"", + "bbox": [ + 297, + 189, + 839, + 199 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Review model (Claude -3-haiku): \"...However, Assistant A provided thorough and insightful analyses of the image is slightly more detailed and precise in its visual assessment. [[1]]\" (Human judgement: [[1]])", + "bbox": [ + 297, + 200, + 893, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/26aa8f5791392d98337dd35d390c6abe17b8a07e501692dce7cfa95724466b3d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 106, + 250, + 290, + 366 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Original Question: Which is right? They are () each other.", + "bbox": [ + 297, + 250, + 571, + 260 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "\\section*{Verbosity bias}", + "bbox": [ + 795, + 250, + 874, + 260 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Options: A: next to, B: far from, C: facing, D: backing\",", + "bbox": [ + 297, + 261, + 560, + 270 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Candidate 1 (Gemini-1.5-pro): \"... Determine the best fit: Two toys are next to each other. [[A]]\" (194 tokens)", + "bbox": [ + 297, + 271, + 816, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Candidate 2 (Claude-3.5-sonnet): \"... Based on this analysis, two toys are next to each other. [[A]]\" (396 tokens)", + "bbox": [ + 297, + 282, + 828, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Review model (GPT-40 mini) [Visual judgment]: \"... Overall, Assistant B demonstrates a deeper level of visual understanding by providing a richer description of the scene [[2]]\" (Human judgement: [[1]])", + "bbox": [ + 297, + 292, + 893, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Question generated by review model: What are the two teddy bears looking at in the image?", + "bbox": [ + 297, + 316, + 728, + 325 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Candidate 1 (Gemini-1.5-pro) : \"...The boats on the water and the mountains in the distance.\" (235 tokens)", + "bbox": [ + 297, + 327, + 795, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Candidate 2 (Claude-3.5-sonnet) : \"...The image shows a baseball player in action during a game...\" (448 tokens)", + "bbox": [ + 297, + 337, + 826, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Review model (GPT-40 mini): \"... I find that Assistant A's response demonstrates better visual understanding and reasoning due to its focus on the emotional and scenic aspects of the view. [[1]]\" (Human judgement: [[1]])", + "bbox": [ + 297, + 349, + 893, + 371 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cb517cf23afb8e7b8289ed69f90f839c5a3f7bcb38858dd1e46910b6cb19dd3b.jpg", + "image_caption": [ + "Figure 7. Case study illustration of UPME. We provide the original human-designed questions and the UPME-generated questions, along with the answer analysis. The upper case presents the Disability of review model, where the review model can not answer the original question itself. The middle case demonstrates cases exhibiting morbidity bias. The bottom case shows self-preference bias." + ], + "image_footnote": [], + "bbox": [ + 107, + 400, + 290, + 515 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Original Question: \" What is the overall mood of this image?", + "bbox": [ + 297, + 398, + 584, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Self-preference", + "bbox": [ + 790, + 397, + 874, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Options: A: Excitement and liveliness, B: Fear and anxiety, C: Calmness and relaxation, D: Confusion and chaos", + "bbox": [ + 297, + 410, + 826, + 420 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Candidate 1 (Claude-3.5-sonnet): \"... Considering these factors, the mood that best describes this image is [[C]].\"", + "bbox": [ + 297, + 420, + 830, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Candidate 2 (GPT-4o): \"... Given the analysis, the most fitting mood is Calmness and relaxation. [[C]]\"", + "bbox": [ + 297, + 431, + 787, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Review model (GPT-40 mini) [Visual judgment]: \"...Assistant B provides a slightly clearer rationale for why the mood aligns with calmness rather than introducing conflicting emotions like melancholy. [[2]]\" (Human judgement: [[1]])", + "bbox": [ + 297, + 441, + 890, + 464 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Question generated by review model: What atmospheric conditions are depicted in the image, and how do they affect visibility?", + "bbox": [ + 297, + 465, + 890, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Candidate 1 (Gemini-1.5-pro): \"... The atmospheric conditions depicted in the image are heavy nighttime rainfall.\"", + "bbox": [ + 297, + 474, + 834, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Candidate 2 (Claude-3.5-sonnet): \"...The most prominent atmospheric condition depicted is falling snow...\"", + "bbox": [ + 297, + 486, + 805, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Review model (GPT-40 mini): \"... I find that Assistant A's response demonstrates better visual understanding and reasoning due to its focus on the emotional and scenic aspects of the view. [[1]]\" (Human judgement: [[1]])", + "bbox": [ + 297, + 497, + 893, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.142e-10, indicating a highly significant bias. In contrast, the UPME framework reduced the Chi-square value to 3.489, the Phi coefficient to 0.076, and the p-value to 0.0618, approaching insignificance. The incorrect selections due to self-preference dropped from 226 to 129, further demonstrating our effectiveness in mitigating this bias.", + "bbox": [ + 89, + 608, + 483, + 699 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Case Study", + "text_level": 1, + "bbox": [ + 89, + 710, + 212, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We selected three cases to demonstrate how the UPME effectively addresses the core issues in traditional MLLM-as-a-judge as shown in Figure 7. In the first case, the review model initially failed to identify the error in Candidate B's response because it was unable to answer the question itself. After applying UPME, the review model was able to generate a question within its capabilities and accurately determine that Candidate A's response showed a deeper understanding of the image, which aligned with human preferences, thus avoiding the misjudgment issue. In the second case, although Candidate B's response was longer, it did not", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "provide additional useful information. Traditional evaluations might wrongly consider the longer answer as more informative, but with UPME, the review model, through precise text evaluation, clearly identified Candidate A's concise and clear answer as more valuable, thus avoiding morbidity bias. Finally, in the third case, UPME effectively avoided self-preference bias, enhancing the fairness and objectivity.", + "bbox": [ + 511, + 608, + 906, + 714 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 739, + 633, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This study proposes an unsupervised peer review-based framework, UPME, effectively reducing the annotation workload in traditional human-designed QA benchmarks. Our vision-language scoring system emphasizes a vision-oriented judgment, addressing issues such as verosity and self-preference biases inherent in MLLM-as-judge evaluation methods. Further experiments validate that our evaluation results achieve higher alignment with human preferences, providing a promising and reliable approach.", + "bbox": [ + 511, + 763, + 906, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9172", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by Alibaba DAMO Academy through the Alibaba Innovative Research Program.", + "bbox": [ + 89, + 114, + 482, + 146 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 157, + 187, + 174 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1", + "[2] Mohammad Allahbakhsh and Aleksandar Ignjatovic. Rating through voting: An iterative method for robust rating. arXiv preprint arXiv:1211.0390, 2012. 5", + "[3] Anthropic. Claude 3.5: A sonnet. https://www.anthropic.com/news/claude-3-5-sonnet, 2024.1,5", + "[4] Anthropic. Claude 3 haiku. https://www.anthropic.com/news/claude-3-haiku, 2024.5, 1", + "[5] Christoph Bandt and Bernd Pompe. Permutation entropy: a natural complexity measure for time series. Physical review letters, 88(17):174102, 2002. 5", + "[6] Han Bao, Yue Huang, Yanbo Wang, Jiayi Ye, Xiangqi Wang, Xiuying Chen, Mohamed Elhoseiny, and Xiangliang Zhang. Autobench-v: Can large vision-language models benchmark themselves? arXiv preprint arXiv:2410.21259, 2024. 2", + "[7] Yuzhe Cai, Shaoguang Mao, Wenshan Wu, Zehua Wang, Yaobo Liang, Tao Ge, Chenfei Wu, Wang You, Ting Song, Yan Xia, et al. Low-code llm: Visual programming over llms. arXiv preprint arXiv:2304.08103, 2, 2023. 1", + "[8] Dongping Chen, Ruoxi Chen, Shilin Zhang, Yinuo Liu, Yaochen Wang, Huichi Zhou, Qihui Zhang, Pan Zhou, Yao Wan, and Lichao Sun. Mllm-as-a-judge: Assessing multimodal llm-as-a-judge with vision-language benchmark. arXiv preprint arXiv:2402.04788, 2024. 2", + "[9] Dongping Chen, Yue Huang, Siyuan Wu, Jingyu Tang, Liuyi Chen, Yilin Bai, Zhigang He, Chenlong Wang, Huichi Zhou, Yiqiang Li, et al. Gui-world: A dataset for gui-oriented multimodal llm-based agents. arXiv preprint arXiv:2406.10819, 2024. 1", + "[10] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 1, 2, 4, 6", + "[11] Zhaorun Chen, Yichao Du, Zichen Wen, Yiyang Zhou, Chenhang Cui, Zhenzhen Weng, Haoqin Tu, Chaoqi Wang, Zhengwei Tong, Leria HUANG, et al. Mj-bench: Is your multimodal reward model really a good judge? In ICML 2024 Workshop on Foundation Models in the Wild, 2024. 2", + "[12] Wei-Lin Chiang, Lianmin Zheng, Ying Sheng, Anastasios Nikolas Angelopoulos, Tianle Li, Dacheng Li, Hao Zhang, Banghua Zhu, Michael Jordan, Joseph E. Gonzalez, and Ion Stoica. Chatbot arena: An open platform for evaluating llms by human preference, 2024. 2", + "[13] Zhumin Chu, Qingyao Ai, Yiteng Tu, Haitao Li, and Yiqun Liu. Pre: A peer review based large language model evaluator. arXiv preprint arXiv:2401.15641, 2024. 3" + ], + "bbox": [ + 93, + 183, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Israel Cohen, Yiteng Huang, Jingdong Chen, Jacob Benesty, Jacob Benesty, Jingdong Chen, Yiteng Huang, and Israel Cohen. Pearson correlation coefficient. Noise reduction in speech processing, pages 1-4, 2009. 5", + "[15] Chunyuan Deng, Yilun Zhao, Xiangru Tang, Mark Gerstein, and Arman Cohan. Benchmark probing: Investigating data leakage in large language models. In NeurIPS 2023 Workshop on Backdoors in Deep Learning-The Good, the Bad, and the Ugly, 2023. 2", + "[16] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 1", + "[17] Chujie Gao, Qihui Zhang, Dongping Chen, Yue Huang, Siyuan Wu, Zhengyan Fu, Yao Wan, Xiangliang Zhang, and Lichao Sun. The best of both worlds: Toward an honest and helpful large language model. arXiv preprint arXiv:2406.00380, 2024. 2", + "[18] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Yuxiao Dong, et al. Co-gagent: A visual language model for gui agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14281-14290, 2024. 1", + "[19] Zdeněk Kasner and Ondřej Dusěk. Beyond traditional benchmarks: Analyzing behaviors of open llms on data-to-text generation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12045–12072, 2024. 1", + "[20] Zdeněk Kasner and Ondřej Dušek. Beyond traditional benchmarks: Analyzing behaviors of open llms on data-to-text generation, 2024. 2", + "[21] Jian Li and Weiheng Lu. A survey on benchmarks of multimodal large language models. arXiv preprint arXiv:2408.08632, 2024. 1", + "[22] Ruosen Li, Teerth Patel, and Xinya Du. Prd: Peer rank and discussion improve large language model based evaluations. arXiv preprint arXiv:2307.02762, 2023. 5, 2", + "[23] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1", + "[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 5", + "[25] Xiao Liu, Xuanyu Lei, Shengyuan Wang, Yue Huang, Zhuoer Feng, Bosi Wen, Jiale Cheng, Pei Ke, et al. Align-bench: Benchmarking chinese alignment of large language models. arXiv preprint arXiv:2311.18743, 2023. 2", + "[26] Yinhong Liu, Han Zhou, Zhijiang Guo, Ehsan Shareghi, Ivan Vulic, Anna Korhonen, and Nigel Collier. Aligning with human judgement: The role of pairwise preference in large language model evaluators. arXiv preprint arXiv:2403.16950, 2024. 2, 4", + "[27] Adian Liusie, Potsawee Manakul, and Mark JF Gales. Zero-shot nlg evaluation through pairware comparisons with llms. arXiv preprint arXiv:2307.07889, 2023. 4" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9173", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 1, 2, 4", + "[29] Meta. Llama 3.2 11b-vision-instruct. https://huggingface.co/meta-llama/Llama-3.2-11B-Vision-Instruct, 2024.5, 1", + "[30] Birger Moell. Evaluating large language models with human feedback: Establishing a swedish benchmark. arXiv preprint arXiv:2405.14006, 2024. 2", + "[31] Kun-Peng Ning, Shuo Yang, Yu-Yang Liu, Jia-Yu Yao, Zhen-Hui Liu, Yu Wang, Ming Pang, and Li Yuan. Peer-review-in-llms: Automatic evaluation method for llms in open-environment. arXiv preprint arXiv:2402.01830, 2024. 2, 3", + "[32] OpenAI. Gpt-4 technical report. https://openai.com/research/gpt-4, 2023.1", + "[33] OpenAI. Gpt-4o mini: Advancing cost-efficient intelligence. https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/, 2024.5,1", + "[34] OpenAI. Introducing openai o1: Advancing model reasoning with visual and logical insights. https://openai.com/index/introducing-openai-o1-preview/, 2024.4", + "[35] OpenAI. Hello gpt-4o, 2024. 5, 1", + "[36] Renjie Pi, Lewei Yao, Jiahui Gao, Jipeng Zhang, and Tong Zhang. Perceptiongpt: Effectively fusing visual perception into lIm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27124-27133, 2024. 1", + "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4, 3", + "[38] Vyas Raina, Adian Liusie, and Mark Gales. Is llm-as-a-judge robust? investigating universal adversarial attacks on zero-shot llm assessment. arXiv preprint arXiv:2402.14016, 2024. 2", + "[39] Refuel.ai. Llm labeling: Technical report, 2025. 2", + "[40] Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaqing Ellen Tan, Yossi Adi, Jingyu Liu, Romain Sauvestre, Tal Remez, et al. Code llama: Open foundation models for code. arXiv preprint arXiv:2308.12950, 2023. 1", + "[41] Andreas Stephan, Dawei Zhu, Matthias Aßenmacher, Xiaoyu Shen, and Benjamin Roth. From calculation to adjudication: Examining llm judges on mathematical reasoning tasks. arXiv preprint arXiv:2409.04168, 2024. 3", + "[42] James Surowiecki. The wisdom of crowds/james surowiecki. NY.: Anchor, 2005. 5", + "[43] Sijun Tan, Siyuan Zhuang, Kyle Montgomery, William Y Tang, Alejandro Cuadron, Chenguang Wang, Raluca Ada" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Popa, and Ion Stoica. Judgebench: A benchmark for evaluating llm-based judges. arXiv preprint arXiv:2410.12784, 2024.3", + "[44] Ruixue Tang, Chao Ma, Wei Emma Zhang, Qi Wu, and Xiaokang Yang. Semantic equivalent adversarial data augmentation for visual question answering. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16, pages 437-453. Springer, 2020. 1", + "[45] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 5, 1", + "[46] Li Wang, Hui Zhang, and Rajesh Patel. A user-centric benchmark for evaluating large language models. arXiv preprint arXiv:2404.13940, 2024. 3", + "[47] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 1, 2", + "[48] Siyuan Wu, Yue Huang, Chujie Gao, Dongping Chen, Qihui Zhang, Yao Wan, Tianyi Zhou, Xiangliang Zhang, Jianfeng Gao, Chaowei Xiao, et al. Unigen: A unified framework for textual dataset generation using large language models. arXiv preprint arXiv:2406.18966, 2024. 1", + "[49] xAI. Realworldqa: A benchmark for evaluating real-world spatial understanding in multimodal ai models, 2024. 2", + "[50] Ruijie Xu, Zengzhi Wang, Run-Ze Fan, and Pengfei Liu. Benchmarking benchmark leakage in large language models. arXiv preprint arXiv:2404.18824, 2024. 2", + "[51] Jiayi Ye, Yanbo Wang, Yue Huang, Dongping Chen, Qihui Zhang, Nuno Moniz, Tian Gao, Werner Geyer, Chao Huang, Pin-Yu Chen, et al. Justice or prejudice? quantifying biases in llm-as-a-judge. arXiv preprint arXiv:2410.02736, 2024. 2", + "[52] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 1, 2, 6", + "[53] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 1, 2", + "[54] Jerrold H Zar. Spearman rank correlation. Encyclopedia of biostatistics, 7, 2005. 5", + "[55] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. 1, 2", + "[56] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, HongFa Wang, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei Li, et al. Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. arXiv preprint arXiv:2310.01852, 2023. 1" + ], + "bbox": [ + 516, + 92, + 903, + 897 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9174", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/UPME_ An Unsupervised Peer Review Framework for Multimodal Large Language Model Evaluation/42fa34eb-b512-4300-b1c0-40e45c00118f_model.json b/2025/UPME_ An Unsupervised Peer Review Framework for Multimodal Large Language Model Evaluation/42fa34eb-b512-4300-b1c0-40e45c00118f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3c4f89b12613f2f672f985f597688dff4928eb1c --- /dev/null +++ b/2025/UPME_ An Unsupervised Peer Review Framework for Multimodal Large Language Model Evaluation/42fa34eb-b512-4300-b1c0-40e45c00118f_model.json @@ -0,0 +1,2541 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.225, + 0.131, + 0.777, + 0.177 + ], + "angle": 0, + "content": "UPME: An Unsupervised Peer Review Framework for Multimodal Large Language Model Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.177, + 0.202, + 0.82, + 0.244 + ], + "angle": 0, + "content": "Qihui Zhang\\(^{1,3*}\\) Munan Ning\\(^{1*}\\) Zheyuan Liu\\(^{1,3}\\) Yue Huang\\(^{4}\\) Shuo Yang\\(^{1}\\) Yanbo Wang\\(^{1}\\) Jiayi Ye\\(^{1}\\) Xiao Chen\\(^{5}\\) Yibing Song\\(^{2,3}\\) Li Yuan\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.254, + 0.846, + 0.294 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Electrical and Computer Engineering, Peking University \\(^{2}\\)Hupan Lab \\(^{3}\\)DAMO Academy, Alibaba Group \\(^{4}\\)University of Notre Dame \\(^{5}\\)Tsinghua University" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.359, + 0.485, + 0.722 + ], + "angle": 0, + "content": "Multimodal Large Language Models (MLLMs) have emerged to tackle the challenges of Visual Question Answering (VQA), sparking a new research focus on conducting objective evaluations of these models. Existing evaluation methods face limitations due to the significant human workload required to design Q&A pairs for visual images, which inherently restricts the scale and scope of evaluations. Although automated MLLM-as-judge approaches attempt to reduce the human workload through automatic evaluations, they often introduce biases. To address these problems, we propose an Unsupervised Peer review MLLM Evaluation framework. It utilizes only image data, allowing models to automatically generate questions and conduct peer review assessments of answers from other models, effectively alleviating the reliance on human workload. Additionally, we introduce the vision-language scoring system to mitigate the bias issues, which focuses on three aspects: (i) response correctness; (ii) visual understanding and reasoning; and (iii) image-text correlation. Experimental results demonstrate that UPME achieves a Pearson correlation of 0.944 with human evaluations on the MMstar dataset and 0.814 on the ScienceQA dataset, indicating that our framework closely aligns with human-designed benchmarks and inherent human preferences." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.739, + 0.222, + 0.755 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.764, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Recently, Large Language Models (LLMs) have made significant strides in reasoning and application capabilities [3, 32, 47], enabling diverse applications such as text and code generation [1, 19, 40, 48]. Notably, due to the exceptional natural language understanding capabilities of LLMs, they are increasingly employed for automated evaluations, a pro" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.336, + 0.896, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.559, + 0.908, + 0.616 + ], + "angle": 0, + "content": "Figure 1. Existing methods for evaluating MLLMs face various challenges. Our proposed UPME framework addresses these limitations by leveraging a peer review mechanism, reducing annotation costs, and aligning closely with human judgment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.735 + ], + "angle": 0, + "content": "cess known as LLM-as-a-Judge [55]. With the integration of vision encoders [36], Multimodal Large Language Models (MLLMs) [24, 32] can incorporate multiple modalities (e.g., text, image, and audio) [23, 56] and showcase remarkable performance in downstream applications, including visual conversation [7] and embodied intelligence [16]. Furthermore, they can act as agents to interact with GUI [9, 18]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Subsequently, the challenge of conducting objective and comprehensive evaluations of MLLMs while providing users with authoritative guidelines has emerged as a key research focus. A common approach involves constructing benchmarks based on Visual Question Answering (VQA), which assess the performance of MLLMs across various dimensions by setting image-based questions and answers [10, 28, 52, 53]. However, most of these works heavily depend on substantial human involvement in creating Q&A pairs for images [21, 44], leading to limited evaluation volume and making it insufficient for a comprehen" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.374, + 0.901 + ], + "angle": 0, + "content": "* Equal contribution † Corresponding author {maskhui1003, munanning}@gmail.com yuanli-ece@pu.edu.cn" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9165" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.257 + ], + "angle": 0, + "content": "sive assessment of visual content [6]. Another category of methods is MLLM-as-a-Judge [17, 20, 26, 55], which aims to reduce reliance on human annotations by allowing models to evaluate each other directly. However, these methods often still depend on human-designed questions and cannot entirely eliminate the need for human input. Additionally, a significant challenge is the presence of morbidity and self-preference biases [8, 51], where MLLMs tend to favor longer responses or prefer their own outputs. These biases can lead mutual evaluation mechanisms to deviate from truly understanding visual content." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.259, + 0.483, + 0.545 + ], + "angle": 0, + "content": "To address these challenges, we propose UPME, the Unsupervised Peer review MLLM Evaluation framework, which is designed to evaluate MLLMs without requiring human QA annotations, as shown in Figure 1. During each iteration, UPME selects two candidate models and a review model from the MLLM pool. The review model generates questions for a given image and evaluates the responses provided by the candidate models. This evaluation is conducted using a vision-language scoring system that assesses textual response correctness, visual understanding and reasoning, as well as image-text consistency through CLIP scores. Each model is initialized with a confidence weight and alternates roles as candidate and review models. Their estimated scores are calculated based on weighted evaluations provided by the review models. The framework employs dynamic weight optimization, iteratively updating model scores and refining weights through consistency optimization. Ultimately, this process generates a comprehensive and less-biased estimated score list." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.548, + 0.483, + 0.744 + ], + "angle": 0, + "content": "We selected the MMStar [10] and ScienceQA [28] datasets as benchmarks because they encompass visual content reasoning and multimodal content across various disciplines. Experimental results show that under unsupervised settings, UPME achieved Pearson correlations of 0.944 and 0.814 with human evaluation results on these two datasets, while Spearman similarities reached 0.972 and 0.886, indicating a high degree of similarity between our approach and human-annotated QA benchmarks. Further experiments demonstrated that UPME achieved higher consistency with human assessments compared to purely peer review-based methods, effectively reducing morbidity and self-preference issues in the MLLM-as-a-Judge framework." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.746, + 0.482, + 0.761 + ], + "angle": 0, + "content": "Overall, our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.762, + 0.483, + 0.806 + ], + "angle": 0, + "content": "- We propose the first Unsupervised Peer review MLLM Evaluation (UPME) framework, addressing the challenge of reliance on human annotations in MLLM evaluation" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.808, + 0.482, + 0.852 + ], + "angle": 0, + "content": "- Our framework incorporates a Vision-Language Scoring System that targets visual performance and image-text association, providing a visually-focused evaluation." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.853, + 0.482, + 0.898 + ], + "angle": 0, + "content": "- We conduct extensive experiments demonstrating that UPME achieves high consistency with human QA assessments and strong alignment with human preferences." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.762, + 0.483, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.655, + 0.105 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.115, + 0.703, + 0.13 + ], + "angle": 0, + "content": "2.1. MLLM Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.907, + 0.44 + ], + "angle": 0, + "content": "MLLM benchmarks rely on human-annotated QA pairs to measure the accuracy of MLLMs in answering questions. They focus on the different capabilities of MLLMs. ScienceQA [28] evaluate the multi-turn reasoning ability for a broad range of scientific subjects by introducing Chain of Thought (CoT) [47]; MMVet [52] presents an LLM-based open-ended output evaluation method, enabling unified scoring across various question types and answer formats; MMMU [53], inspired by MMLU, primarily assesses model performance in complex reasoning tasks requiring college-level knowledge, covering six major disciplines and 30 fields with multimodal questions; MMStar [10] focuses on creating a benchmark requiring visual content for reasoning, rigorously selecting visual-dependent samples and detecting data leakage to ensure genuine performance gains from multimodal training; RealWorldQA [49] is designed to assess the ability of MLLMs to understand real-world space. Although these benchmarks propose various evaluation methods, their evaluation scope is limited by the reliance on substantial human annotation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.452, + 0.697, + 0.468 + ], + "angle": 0, + "content": "2.2. MLLM-as-a-Judge" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.907, + 0.702 + ], + "angle": 0, + "content": "To reduce the human annotation workload, LLM-as-a-Judge [55] is proposed to enable models to perform mutual evaluation. This approach primarily involves using LLMs to assess responses in two ways: comparing pairs of answers to determine superiority [20, 25, 55], or directly scoring individual answers based on specific criteria [17, 25]. This strategy is extended to MLLM-as-a-Judge [8] to assess the performance of MLLMs across multiple evaluation tasks. MJ-Bench [11] evaluates the effectiveness of multimodal reward models in text-to-image generation. However, these methods face two main biases [38, 51]: the verbose bias and self-preference bias, which means MLLMs prefer the answers from themselves or with longer text length. Human preferences have also not been effectively incorporated into these automated evaluation systems [31]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.712, + 0.833, + 0.727 + ], + "angle": 0, + "content": "2.3. Human Preference-based Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Predefined image and QA-pair benchmarks for MLLMs may encounter issues like data leakage and targeted optimization, which can prevent an accurate reflection of the true capabilities of these models [15, 50]. Therefore, evaluation methods based on human preferences offer a better assessment of whether MLLMs meet human needs [30]. Chatbot Arena [12] provides a public platform specifically for evaluating LLMs based on human preference. This platform uses a pairwise comparison approach, where users interact with two anonymous models and select the one that performs better, generating a dynamic leaderboard via the" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9166" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.907, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.351, + 0.907, + 0.424 + ], + "angle": 0, + "content": "Figure 2. The UPME framework consists of three main components: (i) Peer Review Mechanism, where two candidate models and one review model are randomly selected from the MLLM pool. The review model generates questions based on a selected image, and candidate models provide responses. (ii) Vision-Language Judgment Scoring System, which evaluates answers based on textual correctness, visual understanding and reasoning, and image-text correlation. (iii) Dynamic Weight Optimization, ensuring consistency between confidence weights and estimated scores through iterative optimization cycles." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.449, + 0.482, + 0.54 + ], + "angle": 0, + "content": "Elo rating system. Although open-ended user evaluations can more realistically reflect LLM performance in real-world scenarios [46], the high economic cost of collecting large volumes of high-quality user feedback, combined with the long wait time for obtaining such feedback, limits the widespread adoption of this approach." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.551, + 0.226, + 0.569 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.577, + 0.483, + 0.697 + ], + "angle": 0, + "content": "Assigning higher weights to better-performing models leads to a significant improvement in the accuracy of the entire evaluation framework. This conclusion is supported by preliminary experiments detailed in subsection 8.1 and confirmed by recent research [13, 31, 41, 43]. Therefore, we focus on characterizing the relationship between model performance and its judging capability through consistency optimization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.707, + 0.313, + 0.722 + ], + "angle": 0, + "content": "3.1. Peer review formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.728, + 0.483, + 0.85 + ], + "angle": 0, + "content": "Problem definition. The UPME framework facilitates mutual evaluation among MLLMs via a peer review mechanism, aiming to produce evaluation scores that closely reflect human preferences. Given an unsupervised image dataset \\(\\mathcal{I} = \\{I_i\\}_{i=1}^n\\) of \\(n\\) images (with no human annotated questions and answers) and a model pool \\(\\mathcal{M} = \\{M_j\\}_{j=1}^m\\) consisting of \\(m\\) MLLMs (both open-source and closed-source), we can obtain an UPME-learned score list \\(\\hat{G}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.862, + 0.483, + 0.881 + ], + "angle": 0, + "content": "\\[\n\\hat {G} := \\left[ \\hat {G} _ {M _ {1}}, \\hat {G} _ {M _ {2}}, \\hat {G} _ {M _ {3}}, \\dots , \\hat {G} _ {M _ {m}} \\right]. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.484, + 0.902 + ], + "angle": 0, + "content": "On the other hand, given \\( G_{M_1} \\) represents the score of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.449, + 0.907, + 0.478 + ], + "angle": 0, + "content": "model \\(M_{1}\\) on the human-annotated benchmark, we define the human-annotated score list \\(G\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.497, + 0.907, + 0.515 + ], + "angle": 0, + "content": "\\[\nG := \\left[ G _ {M _ {1}}, G _ {M _ {2}}, G _ {M _ {3}}, \\dots , G _ {M _ {m}} \\right]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.906, + 0.598 + ], + "angle": 0, + "content": "Our objective is to optimize the UPME-learned scores \\(\\hat{G}\\) by maximizing their similarity to the human-annotated score list \\(G\\), i.e., \\(\\mathrm{Sim}(G, \\hat{G}) \\to 1\\), ensuring that UPME-learned scores align closely with human benchmark standards." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.623, + 0.907, + 0.748 + ], + "angle": 0, + "content": "Overview of UPME. Figure 2 shows an overview of our UPME framework. During each review iteration, we randomly select two different candidate models \\( M_{j} \\) and \\( M_{k} \\) from the pool \\( \\mathcal{M} \\) for comparison, along with another model \\( M_{r} \\) to serve as the reviewer. \\( M_{r} \\) generates a question \\( Q_{i}^{r} = M_{r}(I_{i}) \\) for each image \\( I_{i} \\) and then prompts the candidate models \\( M_{j} \\) and \\( M_{k} \\) to provide responses \\( A_{i}^{j,r} \\) and \\( A_{i}^{k,r} \\), respectively. The answers can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.76, + 0.907, + 0.78 + ], + "angle": 0, + "content": "\\[\nA _ {i} ^ {j, r} = M _ {j} \\left(I _ {i}, Q _ {i} ^ {r}\\right), \\quad A _ {i} ^ {k, r} = M _ {k} \\left(I _ {i}, Q _ {i} ^ {r}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.793, + 0.906, + 0.857 + ], + "angle": 0, + "content": "Based on the vision-language judgment scoring System \\(S_{VL}\\) (detailed in subsection 3.2), \\(M_r\\) then evaluates these responses, and compares the responses of \\(M_j\\) and \\(M_k\\) to produce a review score \\(\\mathrm{Review}_i^{j,k,r}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.883, + 0.907, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\operatorname {R e v i e w} _ {i} ^ {j, k, r} = S _ {V L} \\left(A _ {i} ^ {j, r}, A _ {i} ^ {k, r}, Q _ {i} ^ {r}, I _ {i} \\mid M _ {r}\\right). \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9167" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.483, + 0.125 + ], + "angle": 0, + "content": "By assigning each model an initiated confidence weight \\( w \\), we can calculate the learned score \\( \\hat{G}_{M_j} \\) of each MLLM:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.133, + 0.483, + 0.168 + ], + "angle": 0, + "content": "\\[\n\\hat {G} _ {M _ {j}} = \\sum_ {i} \\sum_ {k \\neq j} \\sum_ {r \\neq k, r \\neq j} \\operatorname {R e v i w} _ {i} ^ {j, k, r} \\times w _ {r}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.177, + 0.483, + 0.207 + ], + "angle": 0, + "content": "then we can subsequently obtain the UPME-learned score list \\(\\hat{G}\\), as illustrated in Equation 1." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.216, + 0.462, + 0.233 + ], + "angle": 0, + "content": "3.2. Vision-Language Judgment Scoring System" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.238, + 0.484, + 0.359 + ], + "angle": 0, + "content": "In this subsection, we present our vision-language judgment scoring system, denoted as \\( S_{\\mathrm{VL}} \\) in Equation 9. Our evaluation is structured around three core criteria: Response Correctness, Visual Understanding and Reasoning, and Image-Text Correlation. The first two aspects are evaluated by employing \\( M_r \\), which reviews responses based on prompts outlined in Section 10. The Image-Text Correlation is measured using CLIP scores." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.371, + 0.484, + 0.493 + ], + "angle": 0, + "content": "Response Correctness. The primary criterion is the correctness of the candidate models' responses to questions posed by the review model. Existing studies indicate that paired comparative evaluation yields better accuracy than isolated assessments [26, 27], so we adopt a pairwise scoring approach. Based on the review model \\( M_r \\)'s evaluation \\( \\mathrm{Judge}_i^r \\) for candidates \\( M_j \\) and \\( M_k \\), we calculate the model's score for Response Accuracy \\( S_{\\mathrm{Correct}}(j) \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.502, + 0.483, + 0.56 + ], + "angle": 0, + "content": "\\[\nS _ {\\text {C o r r e c t}} (j) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f J u d g e} _ {\\text {C o r r e c t}, j} > \\text {J u d g e} _ {\\text {C o r r e c t}, k} \\\\ 0. 5 & \\text {i f J u d g e} _ {\\text {C o r r e c t}, j} = \\text {J u d g e} _ {\\text {C o r r e c t}, k} \\\\ 0 & \\text {i f J u d g e} _ {\\text {C o r r e c t}, j} < \\text {J u d g e} _ {\\text {C o r r e c t}, k}, \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.484, + 0.645 + ], + "angle": 0, + "content": "where \\(\\mathrm{Judge}_{\\mathrm{Correct}}\\) represents the review model's assessment of response correctness. If \\(\\mathrm{Judge}_{\\mathrm{Correct},j} > \\mathrm{Judge}_{\\mathrm{Correct},k}\\), it indicates that model \\(M_{j}\\) is judged superior in response accuracy to \\(M_{k}\\), scoring 1 point; a tie yields 0.5 points, and a lower score results in 0 points." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.483, + 0.729 + ], + "angle": 0, + "content": "Visual Understanding and Reasoning. Inspired by the significant CoT ability demonstrated in the o1 model [34] through adversarial training, we leverage \\( M_r \\) to evaluate the interpretative and analytical capabilities of the model with respect to visual content." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.731, + 0.483, + 0.777 + ], + "angle": 0, + "content": "\\(S_{\\mathrm{Visual}}(j)\\) is computed using the function \\(\\Gamma\\), which evaluates the response across four dimensions: caption, reasoning, grounding, and relationship, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.799, + 0.483, + 0.816 + ], + "angle": 0, + "content": "\\[\nS _ {\\text {V i s u a l}} (j) = \\Gamma (\\text {c a p . , r e a . , g r o . , o b j .} \\mid A _ {j}, M _ {r}), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where captioning evaluates the ability to generate precise descriptions, reasoning measures logical consistency, grounding assesses accurate object localization, and relationship captures the connections between subjects. Within this structure, the review model \\( M_r \\) provides a combined" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "score via \\(\\Gamma\\), yielding \\(S_{\\mathrm{Visual}}(j)\\). \\(S_{\\mathrm{Visual}}(j)\\) reflects a model's capability to go beyond basic VQA, integrating both description and reasoning to demonstrate a deeper level of visual comprehension." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.165, + 0.906, + 0.21 + ], + "angle": 0, + "content": "Image-Text Correlation. To capture the alignment between image content and textual responses, we introduce the CLIP [37] score:" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.221, + 0.905, + 0.24 + ], + "angle": 0, + "content": "\\[\nS _ {\\text {C l i p}} (j) = \\operatorname {C L I P} \\left(I _ {i}, A _ {i} ^ {j, r}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.25, + 0.906, + 0.296 + ], + "angle": 0, + "content": "where \\( S_{\\mathrm{Clip}}(j) \\) measures the correlation between image \\( I_{i} \\) and response \\( A_{i}^{j,r} \\) from model \\( M_{j} \\). The details are described in subsection 9.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.298, + 0.906, + 0.358 + ], + "angle": 0, + "content": "Combining the above components, we compute the final vision-language judgment score as a weighted sum of the three criteria: Response Accuracy, Visual Understanding and Reasoning, and Image-Text Correlation:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.371, + 0.905, + 0.388 + ], + "angle": 0, + "content": "\\[\nS _ {\\mathrm {V L}} = \\gamma_ {1} S _ {\\text {C o r r e c t}} + \\gamma_ {2} S _ {\\text {V i s u a l}} + \\gamma_ {3} S _ {\\text {C l i p}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.906, + 0.429 + ], + "angle": 0, + "content": "where \\(\\gamma_{1},\\gamma_{2}\\) , and \\(\\gamma_{3}\\) are weights reflecting the relative importance of each component in the overall evaluation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.438, + 0.785, + 0.455 + ], + "angle": 0, + "content": "3.3. Dynamic Weight Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.906, + 0.506 + ], + "angle": 0, + "content": "In each iteration, we use Mean Squared Error (MSE) as the loss function to update the weights \\( w \\), enhancing the consistency between the weights \\( w \\) and the estimated scores \\( \\hat{G} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.517, + 0.905, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M S E}} (\\hat {G}, w) = \\frac {1}{m} \\sum_ {j = 1} ^ {m} \\left(\\hat {G} _ {M _ {j}} - w _ {M _ {j}}\\right) ^ {2}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.906, + 0.66 + ], + "angle": 0, + "content": "The entire optimization process is dynamic. After each round of weight updates, the weights and scores continue to change. The loss function computes new scores based on the updated weights, which are then iteratively used to optimize the weights further, forming a continuous iterative optimization loop, as detailed in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.673, + 0.645, + 0.69 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.698, + 0.704, + 0.714 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.87 + ], + "angle": 0, + "content": "Datasets. To validate the effectiveness of our framework, we conduct experiments on two distinctly styled datasets: ScienceQA [28] and MMStar [10]. ScienceQA underscores the importance of CoT in scientific reasoning tasks, comprising approximately 21,208 multimodal multiple-choice questions covering various science topics. MMStar focuses on evaluating the multimodal capabilities of MLLMs, and consists of 1,500 meticulously curated samples including 6 core capabilities and 18 detailed axes. Each sample exhibits visual dependency and minimal data leakage." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Candidate MLLMs. Our MLLM pool includes 5 closed-source models and 1 open-source model. The open-source" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9168" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.159, + 0.089, + 0.85, + 0.317 + ], + "angle": 0, + "content": "
ModelsMMstarScienceQA
Pearson (↑)Spearman (↑)Per.Ent. (↓)Pearson (↑)Spearman (↑)Per.Ent. (↓)
LLama-3.2-11b-V0.314±0.07570.550±0.05770.983±0.23100.160±0.04300.225±0.09571.099±0.0000
Claude-3-haiku0.095±0.13010.225±0.15001.099±0.0000-0.145±0.0308-0.525±0.20611.099±0.0000
Claude-3.5-sonnet0.780±0.01460.825±0.05000.752±0.23100.437±0.04520.450±0.17320.752±0.2310
Gemini-1.5-pro0.864±0.00680.850±0.05770.637±0.00000.4147±0.04640.725±0.20620.434±0.5352
GPT-4o-mini0.668±0.00730.725±0.12580.868±0.28860.3354±0.06350.600±0.00001.099±0.0000
GPT-4o0.878±0.00380.875±0.00500.159±0.31830.617±0.00710.625±0.12580.637±0.0000
Methods
Peer Review0.725±0.00440.771±0.16161.040±0.28300.463±0.01930.686±0.17771.040±0.2830
Majority Vote [42]0.757±0.00130.757±0.08571.299±0.17330.509±0.01810.524±0.06601.040±0.0000
Rating Vote [2]0.795±0.00130.743±0.23090.628±0.07550.623±0.00840.629±0.18950.920±0.2387
PRD [22]0.838±0.00270.864±0.03170.427±0.00870.636±0.00420.694±0.07340.746±0.0016
UPME0.944±0.00110.972±0.03300.141±0.28120.814±0.00240.886±0.02860.422±0.2812
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.328, + 0.907, + 0.371 + ], + "angle": 0, + "content": "Table 1. Main results of different models and methods on MMstar and ScienceQA datasets, evaluated by Pearson, Spearman, and Permutation Entropy metrics. Since the model cannot judge itself, scores from the other five models involved in the computations are used for evaluation. Optimal results are highlighted in bold and underlined. Standard deviation (std) is calculated based on four runs." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.397, + 0.483, + 0.518 + ], + "angle": 0, + "content": "model is Llama-3.2-11b-vision-instruct [29], which adapts the Llama-3 architecture to the multimodal domain through visual instruction tuning [24]. The closed-source models consist of GPT-4o [35], GPT-4o-mini [33], Claude-3.5-Sonnet [3], Claude-3-Haiku [4], and Gemini-1.5-Pro [45]. These commercial models employ enhanced human alignment strategies to optimize performance and user experience across various applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.483, + 0.653 + ], + "angle": 0, + "content": "Metrics. For all experiments, we employ three popular metrics to evaluate the setups mentioned above and our UPME method: the Pearson Correlation Coefficient [14] to evaluate the similarity of score lists, Spearman's Rank Correlation Coefficient [54] to measure the correlation of ranking lists, and Permutation Entropy [5] to assess the complexity and unpredictability of the scoring distributions. All experiments are performed for 4 runs with different seeds (seed = 1, 2, 3, 4) and record the average results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.483, + 0.76 + ], + "angle": 0, + "content": "Comparison Methods. Since we are the first to investigate unsupervised peer review in MLLM evaluations, we choose the original peer review mechanism and two popular classical methods, Majority Vote [42] and Rating Vote [2], for our comparative analysis. We also adopt a semi-supervised method, PRD [22], and conduct experiments using individual models as judges to further bolster our comparisons." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.769, + 0.293, + 0.785 + ], + "angle": 0, + "content": "4.2. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.792, + 0.279, + 0.806 + ], + "angle": 0, + "content": "4.2.1. Comparison Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The experimental results in Table 1 indicate that among the candidate MLLMs, GPT-4o exhibits the highest review capability on the MMstar dataset, achieving Pearson and Spearman correlation coefficients of 0.878 and 0.875, respectively. In contrast, the open-source LLama-3.2 model demonstrates weaker performance. Although Claude-3.5-" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.399, + 0.891, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.575, + 0.818, + 0.59 + ], + "angle": 0, + "content": "Figure 3. Convergence experiments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.907, + 0.793 + ], + "angle": 0, + "content": "Sonnet and Gemini-1.5-Pro claim to offer performance comparable to GPT-4o, they still show limitations when evaluating other models, falling noticeably short of human-level performance. The original peer review method yields Pearson and Spearman coefficients lower than that of GPT-4o, indicating that the biases of weaker models influence the Peer Review approach. Our UPME method achieves a Pearson correlation of 0.944 and a Spearman correlation of 0.972 on the MMstar dataset, demonstrating a high degree of similarity with human annotations and validating the effectiveness of our optimizations in visual QA comprehension and consistency evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "On the ScienceQA dataset, even the best-performing GPT-4o achieves Pearson and Spearman correlations of 0.617 and 0.625, reflecting the difficulty of the task. The original peer review method shows more modest results, with Pearson and Spearman of 0.463 and 0.686, respectively. Following a series of optimizations and alignments, UPME achieves Pearson and Spearman correlations" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9169" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.483, + 0.23 + ], + "angle": 0, + "content": "
MMstarScienceQA
MethodCor.Vis.Cti.PearsonSpearmanPearsonSpearman
Peer Review0.7270.7140.4570.600
Score Optimization✓0.8540.7710.7130.657
Visual Alignment✓0.8730.8860.7010.771
✓0.7850.8290.5480.771
✓✓0.9030.9430.7750.886
UPME✓✓✓0.9440.9720.8140.886
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.242, + 0.483, + 0.298 + ], + "angle": 0, + "content": "Table 2. Ablation study with Pearson and Spearman similarity metrics for each dataset. Optimal results are highlighted in bold and underlined. 'Cor.' stands for correctness, 'Vis.' stands for visual understanding, and 'Cli.' stands for Clip." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.33, + 0.483, + 0.466 + ], + "angle": 0, + "content": "of 0.814 and 0.886, showcasing its robustness and superior performance in complex datasets. While the performance on ScienceQA may not be as high, it is important to note that traditional datasets like ScienceQA might not fully align with the evaluation needs of current MLLMs because they may lack visual dependency or sufficient discrimination [10, 52]. UPME achieves higher Pearson and Spearman correlations, demonstrating its robustness and superior performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.473, + 0.484, + 0.564 + ], + "angle": 0, + "content": "Moreover, as shown in Figure 3, the loss consistently decreases, and the Pearson similarity metric increases over epochs in 64 different initial settings, which demonstrates the convergence of \\( w \\) and \\( \\hat{G} \\). Under the current experimental setup (25 images and 1500 judgments), the framework reliably converges within the 30-epoch limit." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.589, + 0.245, + 0.605 + ], + "angle": 0, + "content": "4.2.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.614, + 0.483, + 0.9 + ], + "angle": 0, + "content": "In the ablation study presented in Table 2, we evaluated the impact of progressively introducing different components on model performance. The Score Optimization phase showed consistent improvements in Pearson and Spearman correlation coefficients as modules were added. For instance, using only the correctness module resulted in Pearson scores of 0.854 for MMstar and 0.713 for ScienceQA. Introducing the visual understanding module alone yielded scores of 0.873 and 0.701, respectively. Adding the Clip component alone gave scores of 0.785 for MMstar and 0.548 for ScienceQA. Combining visual understanding and Clip modules further improved performance to 0.903 and 0.775. Ultimately, integrating all components within the UPME framework resulted in the highest performance, with Pearson and Spearman correlations of 0.944 and 0.972 on MMstar and 0.814 and 0.886 on ScienceQA. These results confirm that optimizing the scoring mechanism and visual alignment significantly boosts evaluation accuracy and robustness across different datasets." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.905, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.203, + 0.905, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.326, + 0.892, + 0.34 + ], + "angle": 0, + "content": "Figure 4. The performance of UPME in different sample size." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.36, + 0.716, + 0.375 + ], + "angle": 0, + "content": "4.2.3.Impact of Sample Size" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.908, + 0.621 + ], + "angle": 0, + "content": "In Figure 4, we analyze the impact of sample size on the alignment with human-annotated QA benchmarks, measured through Pearson and Spearman correlation coefficients, as well as Permutation Entropy. As the sample size increases from 5 to 25, there is a clear improvement in both correlation metrics and a decline in Permutation Entropy, indicating that the generated score lists and rank lists show greater alignment with human evaluations. Beyond a sample size of 25, the metrics stabilize, showing minimal changes with further increases from 25 to 100 in sample size. This stabilization suggests that the consistency of the generated outputs relative to human judgments has reached a satisfactory level, making 25 samples a reliable point for evaluation. Therefore, we primarily selected a sample size of 25 for our experiments, as it provides an effective balance between alignment accuracy and data efficiency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.634, + 0.866, + 0.651 + ], + "angle": 0, + "content": "5. Human Preference Alignment Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.78 + ], + "angle": 0, + "content": "The aforementioned experiments primarily reflect the consistency between the UPME framework and the predefined QA-based benchmarks. To further explore the alignment between UPME and human preferences, as well as to examine the two types of biases inherent in MLLM-as-Judge approaches, we selected a subset of responses for human evaluation and analysis of UPME scoring results. The detailed experimental design is provided in subsection 9.4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.819, + 0.804 + ], + "angle": 0, + "content": "5.1. Alignment with Human Preference" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To validate whether our proposed UPME method can align with human preferences without manual labeling, we conducted experiments on the MMstar and ScienceQA datasets. The experimental results on Table 3 indicate that the baseline method exhibited relatively low human agreement and model consistency rates, suggesting that the Peer" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9170" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.482, + 0.177 + ], + "angle": 0, + "content": "
DatasetMethodAgreement (%)Consistency (%)
MMstarPeer Review71.167.5
UPME95.989.8
ScienceQAPeer Review68.261.8
UPME87.482.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.187, + 0.483, + 0.214 + ], + "angle": 0, + "content": "Table 3. Human Agreement (%) and Human Consistency (%) Rates on Various Datasets." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.228, + 0.48, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.462, + 0.485, + 0.531 + ], + "angle": 0, + "content": "Figure 5. Model accuracy comparison in peer review framework w/ and w/o UPME, where Peer Review_Cor. represents the correctness of the original peer review, and UPME_Cor. and UPME_Vis. correspond to the two judgment dimensions of response correctness and visual understanding, introduced in subsection 3.2." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.56, + 0.483, + 0.773 + ], + "angle": 0, + "content": "Review mechanism under an unsupervised setting without weight optimization struggles to align with human preferences. In contrast, UPME demonstrated significant improvements by incorporating Correctness, Visual Understanding, and Clip Correlation. On the MMStar, UPME achieved an agreement rate of \\(95.9\\%\\) and a consistency rate of \\(89.8\\%\\), showing that the optimized scoring criteria significantly enhance the accuracy of evaluation outputs and alignment with human preferences. By capturing key multimodal understanding metrics without relying on manual labeling, UPME effectively achieved high consistency with human annotations, highlighting its substantial advantages in improving response consistency and accuracy under an unsupervised framework." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.784, + 0.373, + 0.801 + ], + "angle": 0, + "content": "5.2. Performance w/ and w/o UPME" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.808, + 0.484, + 0.903 + ], + "angle": 0, + "content": "To assess the improvement in model judgment accuracy within the Peer Review mechanism and the UPME framework, we conducted an experiment, the results are displayed in Figure 5. The calculation of accuracy is based on \\(ACC = \\frac{\\mathrm{judge}_{\\mathrm{correct}}}{\\mathrm{judge}_{\\mathrm{all}}}\\). Experimental results show that the UPME significantly improves the accuracy of models in" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.907, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.176, + 0.667, + 0.19 + ], + "angle": 0, + "content": "(a) Verbosity Bias" + }, + { + "type": "image_caption", + "bbox": [ + 0.753, + 0.176, + 0.866, + 0.19 + ], + "angle": 0, + "content": "(b) Self Preference" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.202, + 0.906, + 0.294 + ], + "angle": 0, + "content": "
Bias TypeMethodChi-Squarep-valuePhi Coefficient
VerbosityPeer review10.9960.000910.135
BiasUPME0.2800.596960.022
SelfPeer review39.5843.142e-100.257
PreferenceUPME3.4890.06180.076
" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.304, + 0.907, + 0.333 + ], + "angle": 0, + "content": "Figure 6. Heatmap and Table for Peer Review and UPME on Self Preference and Verbosity Bias." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.359, + 0.907, + 0.556 + ], + "angle": 0, + "content": "both correctness and visual understanding compared to the original correctness judgment. Specifically, the average accuracy in correctness evaluation increased from \\(61.56\\%\\) in the baseline to \\(74.48\\%\\) with UPME, an improvement of \\(12.92\\%\\). This indicates a significant advantage of UPME in assessing the correctness of model responses. Furthermore, in the evaluation of visual understanding, UPME achieved an average accuracy of \\(73.93\\%\\), which is also notably higher than the baseline. This improvement is primarily attributed to UPME's unsupervised evaluation framework, which employs a multi-model peer review mechanism that combines correctness and visual understanding scores to achieve a more efficient and accurate assessment." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.567, + 0.868, + 0.584 + ], + "angle": 0, + "content": "5.3. Mitigating the Bias of MLLM-as-a-Judge" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.907, + 0.832 + ], + "angle": 0, + "content": "Mitigation of Verbosity bias. As shown in Figure 6, the verbosity-preference bias is significantly alleviated by our method. The original Peer Review approach had a p-value of 0.00091 (\\(< 0.05\\)), with a Chi-square value of 10.996 and a Phi coefficient of 0.135, suggesting a significant preference for verbose responses during evaluation. In contrast, the UPME framework yielded a p-value of 0.59696 (\\(>0.05\\)), with the Chi-square value reduced to 0.280 and the Phi coefficient decreased to 0.022, indicating that it does not exhibit a significant bias toward selecting verbose responses. Additionally, the number of incorrect selections of verbose responses dropped from 233 to 103 in the UPME framework. Thus, the UPME approach effectively mitigates the bias toward verbose responses in models, enhancing evaluation diversity and alignment with human preferences in an unsupervised setting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Mitigation of Self-preference. Self-preference bias is also significantly alleviated, as shown in Figure 6. In the original Peer Review method, the Chi-square value was 39.584, the Phi coefficient was 0.257, and the p-value was" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "9171" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.108, + 0.103, + 0.291, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.101, + 0.616, + 0.112 + ], + "angle": 0, + "content": "Original Question: Where is the seated person visible in the image?" + }, + { + "type": "text", + "bbox": [ + 0.733, + 0.1, + 0.876, + 0.112 + ], + "angle": 0, + "content": "Disability of review model" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.113, + 0.688, + 0.122 + ], + "angle": 0, + "content": "Options: A: On the fence, B: In the field, C: Outside the field, D: Behind the bench" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.123, + 0.756, + 0.134 + ], + "angle": 0, + "content": "Candidate 1 (Gemini-1.5-pro): \"...Therefore, the seated person is visible outside the field. [[C]]\"" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.135, + 0.872, + 0.145 + ], + "angle": 0, + "content": "Candidate 2 (Claude-3.5-sonnet): \"...The closest answer would be that the player (though not seated) is in the field. [[B]]\"" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.146, + 0.895, + 0.167 + ], + "angle": 0, + "content": "Review model (Claude-3-haiku) [Correctness judgment] \"...However, Assistant B's response correctly concludes that there is no seated person visible in the image. [[2]]\" (Human judgement: [[1]])" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.168, + 0.708, + 0.178 + ], + "angle": 0, + "content": "Question generated by review model: What position is the player in the image playing?" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.179, + 0.634, + 0.189 + ], + "angle": 0, + "content": "Candidate 1 (Gemini-1.5-pro): \"... The player is on a baseball field...\"" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.19, + 0.84, + 0.2 + ], + "angle": 0, + "content": "Candidate 2 (Claude-3.5-sonnet) : \"...Overall theme: The image shows a baseball player in action during a game...\"" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.201, + 0.895, + 0.223 + ], + "angle": 0, + "content": "Review model (Claude -3-haiku): \"...However, Assistant A provided thorough and insightful analyses of the image is slightly more detailed and precise in its visual assessment. [[1]]\" (Human judgement: [[1]])" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.251, + 0.291, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.25, + 0.573, + 0.261 + ], + "angle": 0, + "content": "Original Question: Which is right? They are () each other." + }, + { + "type": "text", + "bbox": [ + 0.797, + 0.25, + 0.875, + 0.261 + ], + "angle": 0, + "content": "\\section*{Verbosity bias}" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.262, + 0.561, + 0.271 + ], + "angle": 0, + "content": "Options: A: next to, B: far from, C: facing, D: backing\"," + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.272, + 0.818, + 0.283 + ], + "angle": 0, + "content": "Candidate 1 (Gemini-1.5-pro): \"... Determine the best fit: Two toys are next to each other. [[A]]\" (194 tokens)" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.284, + 0.829, + 0.294 + ], + "angle": 0, + "content": "Candidate 2 (Claude-3.5-sonnet): \"... Based on this analysis, two toys are next to each other. [[A]]\" (396 tokens)" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.294, + 0.894, + 0.317 + ], + "angle": 0, + "content": "Review model (GPT-40 mini) [Visual judgment]: \"... Overall, Assistant B demonstrates a deeper level of visual understanding by providing a richer description of the scene [[2]]\" (Human judgement: [[1]])" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.317, + 0.73, + 0.327 + ], + "angle": 0, + "content": "Question generated by review model: What are the two teddy bears looking at in the image?" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.328, + 0.796, + 0.337 + ], + "angle": 0, + "content": "Candidate 1 (Gemini-1.5-pro) : \"...The boats on the water and the mountains in the distance.\" (235 tokens)" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.338, + 0.828, + 0.348 + ], + "angle": 0, + "content": "Candidate 2 (Claude-3.5-sonnet) : \"...The image shows a baseball player in action during a game...\" (448 tokens)" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.35, + 0.895, + 0.372 + ], + "angle": 0, + "content": "Review model (GPT-40 mini): \"... I find that Assistant A's response demonstrates better visual understanding and reasoning due to its focus on the emotional and scenic aspects of the view. [[1]]\" (Human judgement: [[1]])" + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.401, + 0.291, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.399, + 0.585, + 0.41 + ], + "angle": 0, + "content": "Original Question: \" What is the overall mood of this image?" + }, + { + "type": "text", + "bbox": [ + 0.792, + 0.398, + 0.875, + 0.41 + ], + "angle": 0, + "content": "Self-preference" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.411, + 0.828, + 0.421 + ], + "angle": 0, + "content": "Options: A: Excitement and liveliness, B: Fear and anxiety, C: Calmness and relaxation, D: Confusion and chaos" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.421, + 0.831, + 0.431 + ], + "angle": 0, + "content": "Candidate 1 (Claude-3.5-sonnet): \"... Considering these factors, the mood that best describes this image is [[C]].\"" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.432, + 0.788, + 0.442 + ], + "angle": 0, + "content": "Candidate 2 (GPT-4o): \"... Given the analysis, the most fitting mood is Calmness and relaxation. [[C]]\"" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.443, + 0.892, + 0.465 + ], + "angle": 0, + "content": "Review model (GPT-40 mini) [Visual judgment]: \"...Assistant B provides a slightly clearer rationale for why the mood aligns with calmness rather than introducing conflicting emotions like melancholy. [[2]]\" (Human judgement: [[1]])" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.466, + 0.892, + 0.475 + ], + "angle": 0, + "content": "Question generated by review model: What atmospheric conditions are depicted in the image, and how do they affect visibility?" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.476, + 0.836, + 0.486 + ], + "angle": 0, + "content": "Candidate 1 (Gemini-1.5-pro): \"... The atmospheric conditions depicted in the image are heavy nighttime rainfall.\"" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.487, + 0.807, + 0.497 + ], + "angle": 0, + "content": "Candidate 2 (Claude-3.5-sonnet): \"...The most prominent atmospheric condition depicted is falling snow...\"" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.498, + 0.894, + 0.52 + ], + "angle": 0, + "content": "Review model (GPT-40 mini): \"... I find that Assistant A's response demonstrates better visual understanding and reasoning due to its focus on the emotional and scenic aspects of the view. [[1]]\" (Human judgement: [[1]])" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.54, + 0.907, + 0.582 + ], + "angle": 0, + "content": "Figure 7. Case study illustration of UPME. We provide the original human-designed questions and the UPME-generated questions, along with the answer analysis. The upper case presents the Disability of review model, where the review model can not answer the original question itself. The middle case demonstrates cases exhibiting morbidity bias. The bottom case shows self-preference bias." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.609, + 0.484, + 0.7 + ], + "angle": 0, + "content": "3.142e-10, indicating a highly significant bias. In contrast, the UPME framework reduced the Chi-square value to 3.489, the Phi coefficient to 0.076, and the p-value to 0.0618, approaching insignificance. The incorrect selections due to self-preference dropped from 226 to 129, further demonstrating our effectiveness in mitigating this bias." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.712, + 0.214, + 0.728 + ], + "angle": 0, + "content": "5.4. Case Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We selected three cases to demonstrate how the UPME effectively addresses the core issues in traditional MLLM-as-a-judge as shown in Figure 7. In the first case, the review model initially failed to identify the error in Candidate B's response because it was unable to answer the question itself. After applying UPME, the review model was able to generate a question within its capabilities and accurately determine that Candidate A's response showed a deeper understanding of the image, which aligned with human preferences, thus avoiding the misjudgment issue. In the second case, although Candidate B's response was longer, it did not" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.907, + 0.715 + ], + "angle": 0, + "content": "provide additional useful information. Traditional evaluations might wrongly consider the longer answer as more informative, but with UPME, the review model, through precise text evaluation, clearly identified Candidate A's concise and clear answer as more valuable, thus avoiding morbidity bias. Finally, in the third case, UPME effectively avoided self-preference bias, enhancing the fairness and objectivity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.634, + 0.755 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.902 + ], + "angle": 0, + "content": "This study proposes an unsupervised peer review-based framework, UPME, effectively reducing the annotation workload in traditional human-designed QA benchmarks. Our vision-language scoring system emphasizes a vision-oriented judgment, addressing issues such as verosity and self-preference biases inherent in MLLM-as-judge evaluation methods. Further experiments validate that our evaluation results achieve higher alignment with human preferences, providing a promising and reliable approach." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9172" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.483, + 0.147 + ], + "angle": 0, + "content": "This work was supported by Alibaba DAMO Academy through the Alibaba Innovative Research Program." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.158, + 0.188, + 0.175 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.184, + 0.483, + 0.24 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.484, + 0.282 + ], + "angle": 0, + "content": "[2] Mohammad Allahbakhsh and Aleksandar Ignjatovic. Rating through voting: An iterative method for robust rating. arXiv preprint arXiv:1211.0390, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.284, + 0.482, + 0.323 + ], + "angle": 0, + "content": "[3] Anthropic. Claude 3.5: A sonnet. https://www.anthropic.com/news/claude-3-5-sonnet, 2024.1,5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.326, + 0.483, + 0.366 + ], + "angle": 0, + "content": "[4] Anthropic. Claude 3 haiku. https://www.anthropic.com/news/claude-3-haiku, 2024.5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.483, + 0.409 + ], + "angle": 0, + "content": "[5] Christoph Bandt and Bernd Pompe. Permutation entropy: a natural complexity measure for time series. Physical review letters, 88(17):174102, 2002. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.411, + 0.483, + 0.465 + ], + "angle": 0, + "content": "[6] Han Bao, Yue Huang, Yanbo Wang, Jiayi Ye, Xiangqi Wang, Xiuying Chen, Mohamed Elhoseiny, and Xiangliang Zhang. Autobench-v: Can large vision-language models benchmark themselves? arXiv preprint arXiv:2410.21259, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.483, + 0.522 + ], + "angle": 0, + "content": "[7] Yuzhe Cai, Shaoguang Mao, Wenshan Wu, Zehua Wang, Yaobo Liang, Tao Ge, Chenfei Wu, Wang You, Ting Song, Yan Xia, et al. Low-code llm: Visual programming over llms. arXiv preprint arXiv:2304.08103, 2, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.524, + 0.483, + 0.592 + ], + "angle": 0, + "content": "[8] Dongping Chen, Ruoxi Chen, Shilin Zhang, Yinuo Liu, Yaochen Wang, Huichi Zhou, Qihui Zhang, Pan Zhou, Yao Wan, and Lichao Sun. Mllm-as-a-judge: Assessing multimodal llm-as-a-judge with vision-language benchmark. arXiv preprint arXiv:2402.04788, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.594, + 0.483, + 0.661 + ], + "angle": 0, + "content": "[9] Dongping Chen, Yue Huang, Siyuan Wu, Jingyu Tang, Liuyi Chen, Yilin Bai, Zhigang He, Chenlong Wang, Huichi Zhou, Yiqiang Li, et al. Gui-world: A dataset for gui-oriented multimodal llm-based agents. arXiv preprint arXiv:2406.10819, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.664, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[10] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.72, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[11] Zhaorun Chen, Yichao Du, Zichen Wen, Yiyang Zhou, Chenhang Cui, Zhenzhen Weng, Haoqin Tu, Chaoqi Wang, Zhengwei Tong, Leria HUANG, et al. Mj-bench: Is your multimodal reward model really a good judge? In ICML 2024 Workshop on Foundation Models in the Wild, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[12] Wei-Lin Chiang, Lianmin Zheng, Ying Sheng, Anastasios Nikolas Angelopoulos, Tianle Li, Dacheng Li, Hao Zhang, Banghua Zhu, Michael Jordan, Joseph E. Gonzalez, and Ion Stoica. Chatbot arena: An open platform for evaluating llms by human preference, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Zhumin Chu, Qingyao Ai, Yiteng Tu, Haitao Li, and Yiqun Liu. Pre: A peer review based large language model evaluator. arXiv preprint arXiv:2401.15641, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.184, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[14] Israel Cohen, Yiteng Huang, Jingdong Chen, Jacob Benesty, Jacob Benesty, Jingdong Chen, Yiteng Huang, and Israel Cohen. Pearson correlation coefficient. Noise reduction in speech processing, pages 1-4, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[15] Chunyuan Deng, Yilun Zhao, Xiangru Tang, Mark Gerstein, and Arman Cohan. Benchmark probing: Investigating data leakage in large language models. In NeurIPS 2023 Workshop on Backdoors in Deep Learning-The Good, the Bad, and the Ugly, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[16] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.291, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[17] Chujie Gao, Qihui Zhang, Dongping Chen, Yue Huang, Siyuan Wu, Zhengyan Fu, Yao Wan, Xiangliang Zhang, and Lichao Sun. The best of both worlds: Toward an honest and helpful large language model. arXiv preprint arXiv:2406.00380, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[18] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Yuxiao Dong, et al. Co-gagent: A visual language model for gui agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14281-14290, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.433, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[19] Zdeněk Kasner and Ondřej Dusěk. Beyond traditional benchmarks: Analyzing behaviors of open llms on data-to-text generation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12045–12072, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.504, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[20] Zdeněk Kasner and Ondřej Dušek. Beyond traditional benchmarks: Analyzing behaviors of open llms on data-to-text generation, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[21] Jian Li and Weiheng Lu. A survey on benchmarks of multimodal large language models. arXiv preprint arXiv:2408.08632, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.59, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[22] Ruosen Li, Teerth Patel, and Xinya Du. Prd: Peer rank and discussion improve large language model based evaluations. arXiv preprint arXiv:2307.02762, 2023. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.633, + 0.905, + 0.686 + ], + "angle": 0, + "content": "[23] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[25] Xiao Liu, Xuanyu Lei, Shengyuan Wang, Yue Huang, Zhuoer Feng, Bosi Wen, Jiale Cheng, Pei Ke, et al. Align-bench: Benchmarking chinese alignment of large language models. arXiv preprint arXiv:2311.18743, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[26] Yinhong Liu, Han Zhou, Zhijiang Guo, Ehsan Shareghi, Ivan Vulic, Anna Korhonen, and Nigel Collier. Aligning with human judgement: The role of pairwise preference in large language model evaluators. arXiv preprint arXiv:2403.16950, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Adian Liusie, Potsawee Manakul, and Mark JF Gales. Zero-shot nlg evaluation through pairware comparisons with llms. arXiv preprint arXiv:2307.07889, 2023. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9173" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.176 + ], + "angle": 0, + "content": "[28] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.485, + 0.219 + ], + "angle": 0, + "content": "[29] Meta. Llama 3.2 11b-vision-instruct. https://huggingface.co/meta-llama/Llama-3.2-11B-Vision-Instruct, 2024.5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.484, + 0.261 + ], + "angle": 0, + "content": "[30] Birger Moell. Evaluating large language models with human feedback: Establishing a swedish benchmark. arXiv preprint arXiv:2405.14006, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.262, + 0.483, + 0.331 + ], + "angle": 0, + "content": "[31] Kun-Peng Ning, Shuo Yang, Yu-Yang Liu, Jia-Yu Yao, Zhen-Hui Liu, Yu Wang, Ming Pang, and Li Yuan. Peer-review-in-llms: Automatic evaluation method for llms in open-environment. arXiv preprint arXiv:2402.01830, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.48, + 0.361 + ], + "angle": 0, + "content": "[32] OpenAI. Gpt-4 technical report. https://openai.com/research/gpt-4, 2023.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.483, + 0.417 + ], + "angle": 0, + "content": "[33] OpenAI. Gpt-4o mini: Advancing cost-efficient intelligence. https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/, 2024.5,1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.483, + 0.473 + ], + "angle": 0, + "content": "[34] OpenAI. Introducing openai o1: Advancing model reasoning with visual and logical insights. https://openai.com/index/introducing-openai-o1-preview/, 2024.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.475, + 0.327, + 0.489 + ], + "angle": 0, + "content": "[35] OpenAI. Hello gpt-4o, 2024. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.483, + 0.559 + ], + "angle": 0, + "content": "[36] Renjie Pi, Lewei Yao, Jiahui Gao, Jipeng Zhang, and Tong Zhang. Perceptiongpt: Effectively fusing visual perception into lIm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27124-27133, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.483, + 0.643 + ], + "angle": 0, + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.645, + 0.483, + 0.699 + ], + "angle": 0, + "content": "[38] Vyas Raina, Adian Liusie, and Mark Gales. Is llm-as-a-judge robust? investigating universal adversarial attacks on zero-shot llm assessment. arXiv preprint arXiv:2402.14016, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.702, + 0.424, + 0.716 + ], + "angle": 0, + "content": "[39] Refuel.ai. Llm labeling: Technical report, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.717, + 0.483, + 0.785 + ], + "angle": 0, + "content": "[40] Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaqing Ellen Tan, Yossi Adi, Jingyu Liu, Romain Sauvestre, Tal Remez, et al. Code llama: Open foundation models for code. arXiv preprint arXiv:2308.12950, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[41] Andreas Stephan, Dawei Zhu, Matthias Aßenmacher, Xiaoyu Shen, and Benjamin Roth. From calculation to adjudication: Examining llm judges on mathematical reasoning tasks. arXiv preprint arXiv:2409.04168, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[42] James Surowiecki. The wisdom of crowds/james surowiecki. NY.: Anchor, 2005. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[43] Sijun Tan, Siyuan Zhuang, Kyle Montgomery, William Y Tang, Alejandro Cuadron, Chenguang Wang, Raluca Ada" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "Popa, and Ion Stoica. Judgebench: A benchmark for evaluating llm-based judges. arXiv preprint arXiv:2410.12784, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[44] Ruixue Tang, Chao Ma, Wei Emma Zhang, Qi Wu, and Xiaokang Yang. Semantic equivalent adversarial data augmentation for visual question answering. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16, pages 437-453. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[45] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[46] Li Wang, Hui Zhang, and Rajesh Patel. A user-centric benchmark for evaluating large language models. arXiv preprint arXiv:2404.13940, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[47] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[48] Siyuan Wu, Yue Huang, Chujie Gao, Dongping Chen, Qihui Zhang, Yao Wan, Tianyi Zhou, Xiangliang Zhang, Jianfeng Gao, Chaowei Xiao, et al. Unigen: A unified framework for textual dataset generation using large language models. arXiv preprint arXiv:2406.18966, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[49] xAI. Realworldqa: A benchmark for evaluating real-world spatial understanding in multimodal ai models, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[50] Ruijie Xu, Zengzhi Wang, Run-Ze Fan, and Pengfei Liu. Benchmarking benchmark leakage in large language models. arXiv preprint arXiv:2404.18824, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[51] Jiayi Ye, Yanbo Wang, Yue Huang, Dongping Chen, Qihui Zhang, Nuno Moniz, Tian Gao, Werner Geyer, Chao Huang, Pin-Yu Chen, et al. Justice or prejudice? quantifying biases in llm-as-a-judge. arXiv preprint arXiv:2410.02736, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.905, + 0.643 + ], + "angle": 0, + "content": "[52] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.645, + 0.905, + 0.728 + ], + "angle": 0, + "content": "[53] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.729, + 0.905, + 0.756 + ], + "angle": 0, + "content": "[54] Jerrold H Zar. Spearman rank correlation. Encyclopedia of biostatistics, 7, 2005. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.758, + 0.905, + 0.827 + ], + "angle": 0, + "content": "[55] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.829, + 0.905, + 0.898 + ], + "angle": 0, + "content": "[56] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, HongFa Wang, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei Li, et al. Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. arXiv preprint arXiv:2310.01852, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "9174" + } + ] +] \ No newline at end of file diff --git a/2025/UPME_ An Unsupervised Peer Review Framework for Multimodal Large Language Model Evaluation/42fa34eb-b512-4300-b1c0-40e45c00118f_origin.pdf b/2025/UPME_ An Unsupervised Peer Review Framework for Multimodal Large Language Model Evaluation/42fa34eb-b512-4300-b1c0-40e45c00118f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..edb36761574e8dbad0316eed9e123c5660e66645 --- /dev/null +++ b/2025/UPME_ An Unsupervised Peer Review Framework for Multimodal Large Language Model Evaluation/42fa34eb-b512-4300-b1c0-40e45c00118f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:df8cade65a53f4c9e1017bb7434ed4478c730e037992d33e5e3ae51ba529feb2 +size 1057515 diff --git a/2025/UPME_ An Unsupervised Peer Review Framework for Multimodal Large Language Model Evaluation/full.md b/2025/UPME_ An Unsupervised Peer Review Framework for Multimodal Large Language Model Evaluation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1055671caaa6a09abe5912b48b9d83979787341c --- /dev/null +++ b/2025/UPME_ An Unsupervised Peer Review Framework for Multimodal Large Language Model Evaluation/full.md @@ -0,0 +1,377 @@ +# UPME: An Unsupervised Peer Review Framework for Multimodal Large Language Model Evaluation + +Qihui Zhang $^{1,3*}$ Munan Ning $^{1*}$ Zheyuan Liu $^{1,3}$ Yue Huang $^{4}$ Shuo Yang $^{1}$ Yanbo Wang $^{1}$ Jiayi Ye $^{1}$ Xiao Chen $^{5}$ Yibing Song $^{2,3}$ Li Yuan $^{1\dagger}$ + +$^{1}$ School of Electrical and Computer Engineering, Peking University $^{2}$ Hupan Lab $^{3}$ DAMO Academy, Alibaba Group $^{4}$ University of Notre Dame $^{5}$ Tsinghua University + +# Abstract + +Multimodal Large Language Models (MLLMs) have emerged to tackle the challenges of Visual Question Answering (VQA), sparking a new research focus on conducting objective evaluations of these models. Existing evaluation methods face limitations due to the significant human workload required to design Q&A pairs for visual images, which inherently restricts the scale and scope of evaluations. Although automated MLLM-as-judge approaches attempt to reduce the human workload through automatic evaluations, they often introduce biases. To address these problems, we propose an Unsupervised Peer review MLLM Evaluation framework. It utilizes only image data, allowing models to automatically generate questions and conduct peer review assessments of answers from other models, effectively alleviating the reliance on human workload. Additionally, we introduce the vision-language scoring system to mitigate the bias issues, which focuses on three aspects: (i) response correctness; (ii) visual understanding and reasoning; and (iii) image-text correlation. Experimental results demonstrate that UPME achieves a Pearson correlation of 0.944 with human evaluations on the MMstar dataset and 0.814 on the ScienceQA dataset, indicating that our framework closely aligns with human-designed benchmarks and inherent human preferences. + +# 1. Introduction + +Recently, Large Language Models (LLMs) have made significant strides in reasoning and application capabilities [3, 32, 47], enabling diverse applications such as text and code generation [1, 19, 40, 48]. Notably, due to the exceptional natural language understanding capabilities of LLMs, they are increasingly employed for automated evaluations, a pro + +![](images/2bf165477525d2a24cc964c43cc7b98f431535217d020b712f2136a2a678a6b8.jpg) +Figure 1. Existing methods for evaluating MLLMs face various challenges. Our proposed UPME framework addresses these limitations by leveraging a peer review mechanism, reducing annotation costs, and aligning closely with human judgment. + +cess known as LLM-as-a-Judge [55]. With the integration of vision encoders [36], Multimodal Large Language Models (MLLMs) [24, 32] can incorporate multiple modalities (e.g., text, image, and audio) [23, 56] and showcase remarkable performance in downstream applications, including visual conversation [7] and embodied intelligence [16]. Furthermore, they can act as agents to interact with GUI [9, 18]. + +Subsequently, the challenge of conducting objective and comprehensive evaluations of MLLMs while providing users with authoritative guidelines has emerged as a key research focus. A common approach involves constructing benchmarks based on Visual Question Answering (VQA), which assess the performance of MLLMs across various dimensions by setting image-based questions and answers [10, 28, 52, 53]. However, most of these works heavily depend on substantial human involvement in creating Q&A pairs for images [21, 44], leading to limited evaluation volume and making it insufficient for a comprehen + +sive assessment of visual content [6]. Another category of methods is MLLM-as-a-Judge [17, 20, 26, 55], which aims to reduce reliance on human annotations by allowing models to evaluate each other directly. However, these methods often still depend on human-designed questions and cannot entirely eliminate the need for human input. Additionally, a significant challenge is the presence of morbidity and self-preference biases [8, 51], where MLLMs tend to favor longer responses or prefer their own outputs. These biases can lead mutual evaluation mechanisms to deviate from truly understanding visual content. + +To address these challenges, we propose UPME, the Unsupervised Peer review MLLM Evaluation framework, which is designed to evaluate MLLMs without requiring human QA annotations, as shown in Figure 1. During each iteration, UPME selects two candidate models and a review model from the MLLM pool. The review model generates questions for a given image and evaluates the responses provided by the candidate models. This evaluation is conducted using a vision-language scoring system that assesses textual response correctness, visual understanding and reasoning, as well as image-text consistency through CLIP scores. Each model is initialized with a confidence weight and alternates roles as candidate and review models. Their estimated scores are calculated based on weighted evaluations provided by the review models. The framework employs dynamic weight optimization, iteratively updating model scores and refining weights through consistency optimization. Ultimately, this process generates a comprehensive and less-biased estimated score list. + +We selected the MMStar [10] and ScienceQA [28] datasets as benchmarks because they encompass visual content reasoning and multimodal content across various disciplines. Experimental results show that under unsupervised settings, UPME achieved Pearson correlations of 0.944 and 0.814 with human evaluation results on these two datasets, while Spearman similarities reached 0.972 and 0.886, indicating a high degree of similarity between our approach and human-annotated QA benchmarks. Further experiments demonstrated that UPME achieved higher consistency with human assessments compared to purely peer review-based methods, effectively reducing morbidity and self-preference issues in the MLLM-as-a-Judge framework. + +Overall, our contributions can be summarized as follows: + +- We propose the first Unsupervised Peer review MLLM Evaluation (UPME) framework, addressing the challenge of reliance on human annotations in MLLM evaluation +- Our framework incorporates a Vision-Language Scoring System that targets visual performance and image-text association, providing a visually-focused evaluation. +- We conduct extensive experiments demonstrating that UPME achieves high consistency with human QA assessments and strong alignment with human preferences. + +# 2. Related Work + +# 2.1. MLLM Benchmark + +MLLM benchmarks rely on human-annotated QA pairs to measure the accuracy of MLLMs in answering questions. They focus on the different capabilities of MLLMs. ScienceQA [28] evaluate the multi-turn reasoning ability for a broad range of scientific subjects by introducing Chain of Thought (CoT) [47]; MMVet [52] presents an LLM-based open-ended output evaluation method, enabling unified scoring across various question types and answer formats; MMMU [53], inspired by MMLU, primarily assesses model performance in complex reasoning tasks requiring college-level knowledge, covering six major disciplines and 30 fields with multimodal questions; MMStar [10] focuses on creating a benchmark requiring visual content for reasoning, rigorously selecting visual-dependent samples and detecting data leakage to ensure genuine performance gains from multimodal training; RealWorldQA [49] is designed to assess the ability of MLLMs to understand real-world space. Although these benchmarks propose various evaluation methods, their evaluation scope is limited by the reliance on substantial human annotation. + +# 2.2. MLLM-as-a-Judge + +To reduce the human annotation workload, LLM-as-a-Judge [55] is proposed to enable models to perform mutual evaluation. This approach primarily involves using LLMs to assess responses in two ways: comparing pairs of answers to determine superiority [20, 25, 55], or directly scoring individual answers based on specific criteria [17, 25]. This strategy is extended to MLLM-as-a-Judge [8] to assess the performance of MLLMs across multiple evaluation tasks. MJ-Bench [11] evaluates the effectiveness of multimodal reward models in text-to-image generation. However, these methods face two main biases [38, 51]: the verbose bias and self-preference bias, which means MLLMs prefer the answers from themselves or with longer text length. Human preferences have also not been effectively incorporated into these automated evaluation systems [31]. + +# 2.3. Human Preference-based Evaluation + +Predefined image and QA-pair benchmarks for MLLMs may encounter issues like data leakage and targeted optimization, which can prevent an accurate reflection of the true capabilities of these models [15, 50]. Therefore, evaluation methods based on human preferences offer a better assessment of whether MLLMs meet human needs [30]. Chatbot Arena [12] provides a public platform specifically for evaluating LLMs based on human preference. This platform uses a pairwise comparison approach, where users interact with two anonymous models and select the one that performs better, generating a dynamic leaderboard via the + +![](images/c665188dc9e4d133a0633e2a63266255a6c89241ccf47717078e05c42661f46e.jpg) +Figure 2. The UPME framework consists of three main components: (i) Peer Review Mechanism, where two candidate models and one review model are randomly selected from the MLLM pool. The review model generates questions based on a selected image, and candidate models provide responses. (ii) Vision-Language Judgment Scoring System, which evaluates answers based on textual correctness, visual understanding and reasoning, and image-text correlation. (iii) Dynamic Weight Optimization, ensuring consistency between confidence weights and estimated scores through iterative optimization cycles. + +Elo rating system. Although open-ended user evaluations can more realistically reflect LLM performance in real-world scenarios [46], the high economic cost of collecting large volumes of high-quality user feedback, combined with the long wait time for obtaining such feedback, limits the widespread adoption of this approach. + +# 3. Methodology + +Assigning higher weights to better-performing models leads to a significant improvement in the accuracy of the entire evaluation framework. This conclusion is supported by preliminary experiments detailed in subsection 8.1 and confirmed by recent research [13, 31, 41, 43]. Therefore, we focus on characterizing the relationship between model performance and its judging capability through consistency optimization. + +# 3.1. Peer review formulation + +Problem definition. The UPME framework facilitates mutual evaluation among MLLMs via a peer review mechanism, aiming to produce evaluation scores that closely reflect human preferences. Given an unsupervised image dataset $\mathcal{I} = \{I_i\}_{i=1}^n$ of $n$ images (with no human annotated questions and answers) and a model pool $\mathcal{M} = \{M_j\}_{j=1}^m$ consisting of $m$ MLLMs (both open-source and closed-source), we can obtain an UPME-learned score list $\hat{G}$ : + +$$ +\hat {G} := \left[ \hat {G} _ {M _ {1}}, \hat {G} _ {M _ {2}}, \hat {G} _ {M _ {3}}, \dots , \hat {G} _ {M _ {m}} \right]. \tag {1} +$$ + +On the other hand, given $G_{M_1}$ represents the score of + +model $M_{1}$ on the human-annotated benchmark, we define the human-annotated score list $G$ as: + +$$ +G := \left[ G _ {M _ {1}}, G _ {M _ {2}}, G _ {M _ {3}}, \dots , G _ {M _ {m}} \right]. \tag {2} +$$ + +Our objective is to optimize the UPME-learned scores $\hat{G}$ by maximizing their similarity to the human-annotated score list $G$ , i.e., $\mathrm{Sim}(G, \hat{G}) \to 1$ , ensuring that UPME-learned scores align closely with human benchmark standards. + +Overview of UPME. Figure 2 shows an overview of our UPME framework. During each review iteration, we randomly select two different candidate models $M_{j}$ and $M_{k}$ from the pool $\mathcal{M}$ for comparison, along with another model $M_{r}$ to serve as the reviewer. $M_{r}$ generates a question $Q_{i}^{r} = M_{r}(I_{i})$ for each image $I_{i}$ and then prompts the candidate models $M_{j}$ and $M_{k}$ to provide responses $A_{i}^{j,r}$ and $A_{i}^{k,r}$ , respectively. The answers can be written as: + +$$ +A _ {i} ^ {j, r} = M _ {j} \left(I _ {i}, Q _ {i} ^ {r}\right), \quad A _ {i} ^ {k, r} = M _ {k} \left(I _ {i}, Q _ {i} ^ {r}\right). \tag {3} +$$ + +Based on the vision-language judgment scoring System $S_{VL}$ (detailed in subsection 3.2), $M_r$ then evaluates these responses, and compares the responses of $M_j$ and $M_k$ to produce a review score $\mathrm{Review}_i^{j,k,r}$ as: + +$$ +\operatorname {R e v i e w} _ {i} ^ {j, k, r} = S _ {V L} \left(A _ {i} ^ {j, r}, A _ {i} ^ {k, r}, Q _ {i} ^ {r}, I _ {i} \mid M _ {r}\right). \tag {4} +$$ + +By assigning each model an initiated confidence weight $w$ , we can calculate the learned score $\hat{G}_{M_j}$ of each MLLM: + +$$ +\hat {G} _ {M _ {j}} = \sum_ {i} \sum_ {k \neq j} \sum_ {r \neq k, r \neq j} \operatorname {R e v i w} _ {i} ^ {j, k, r} \times w _ {r}, \tag {5} +$$ + +then we can subsequently obtain the UPME-learned score list $\hat{G}$ , as illustrated in Equation 1. + +# 3.2. Vision-Language Judgment Scoring System + +In this subsection, we present our vision-language judgment scoring system, denoted as $S_{\mathrm{VL}}$ in Equation 9. Our evaluation is structured around three core criteria: Response Correctness, Visual Understanding and Reasoning, and Image-Text Correlation. The first two aspects are evaluated by employing $M_r$ , which reviews responses based on prompts outlined in Section 10. The Image-Text Correlation is measured using CLIP scores. + +Response Correctness. The primary criterion is the correctness of the candidate models' responses to questions posed by the review model. Existing studies indicate that paired comparative evaluation yields better accuracy than isolated assessments [26, 27], so we adopt a pairwise scoring approach. Based on the review model $M_r$ 's evaluation $\mathrm{Judge}_i^r$ for candidates $M_j$ and $M_k$ , we calculate the model's score for Response Accuracy $S_{\mathrm{Correct}}(j)$ as follows: + +$$ +S _ {\text {C o r r e c t}} (j) = \left\{ \begin{array}{l l} 1 & \text {i f J u d g e} _ {\text {C o r r e c t}, j} > \text {J u d g e} _ {\text {C o r r e c t}, k} \\ 0. 5 & \text {i f J u d g e} _ {\text {C o r r e c t}, j} = \text {J u d g e} _ {\text {C o r r e c t}, k} \\ 0 & \text {i f J u d g e} _ {\text {C o r r e c t}, j} < \text {J u d g e} _ {\text {C o r r e c t}, k}, \end{array} \right. \tag {6} +$$ + +where $\mathrm{Judge}_{\mathrm{Correct}}$ represents the review model's assessment of response correctness. If $\mathrm{Judge}_{\mathrm{Correct},j} > \mathrm{Judge}_{\mathrm{Correct},k}$ , it indicates that model $M_{j}$ is judged superior in response accuracy to $M_{k}$ , scoring 1 point; a tie yields 0.5 points, and a lower score results in 0 points. + +Visual Understanding and Reasoning. Inspired by the significant CoT ability demonstrated in the o1 model [34] through adversarial training, we leverage $M_r$ to evaluate the interpretative and analytical capabilities of the model with respect to visual content. + +$S_{\mathrm{Visual}}(j)$ is computed using the function $\Gamma$ , which evaluates the response across four dimensions: caption, reasoning, grounding, and relationship, as follows: + +$$ +S _ {\text {V i s u a l}} (j) = \Gamma (\text {c a p . , r e a . , g r o . , o b j .} \mid A _ {j}, M _ {r}), \tag {7} +$$ + +where captioning evaluates the ability to generate precise descriptions, reasoning measures logical consistency, grounding assesses accurate object localization, and relationship captures the connections between subjects. Within this structure, the review model $M_r$ provides a combined + +score via $\Gamma$ , yielding $S_{\mathrm{Visual}}(j)$ . $S_{\mathrm{Visual}}(j)$ reflects a model's capability to go beyond basic VQA, integrating both description and reasoning to demonstrate a deeper level of visual comprehension. + +Image-Text Correlation. To capture the alignment between image content and textual responses, we introduce the CLIP [37] score: + +$$ +S _ {\text {C l i p}} (j) = \operatorname {C L I P} \left(I _ {i}, A _ {i} ^ {j, r}\right), \tag {8} +$$ + +where $S_{\mathrm{Clip}}(j)$ measures the correlation between image $I_{i}$ and response $A_{i}^{j,r}$ from model $M_{j}$ . The details are described in subsection 9.2. + +Combining the above components, we compute the final vision-language judgment score as a weighted sum of the three criteria: Response Accuracy, Visual Understanding and Reasoning, and Image-Text Correlation: + +$$ +S _ {\mathrm {V L}} = \gamma_ {1} S _ {\text {C o r r e c t}} + \gamma_ {2} S _ {\text {V i s u a l}} + \gamma_ {3} S _ {\text {C l i p}}, \tag {9} +$$ + +where $\gamma_{1},\gamma_{2}$ , and $\gamma_{3}$ are weights reflecting the relative importance of each component in the overall evaluation. + +# 3.3. Dynamic Weight Optimization + +In each iteration, we use Mean Squared Error (MSE) as the loss function to update the weights $w$ , enhancing the consistency between the weights $w$ and the estimated scores $\hat{G}$ : + +$$ +\mathcal {L} _ {\mathrm {M S E}} (\hat {G}, w) = \frac {1}{m} \sum_ {j = 1} ^ {m} \left(\hat {G} _ {M _ {j}} - w _ {M _ {j}}\right) ^ {2}. \tag {10} +$$ + +The entire optimization process is dynamic. After each round of weight updates, the weights and scores continue to change. The loss function computes new scores based on the updated weights, which are then iteratively used to optimize the weights further, forming a continuous iterative optimization loop, as detailed in Algorithm 1. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. To validate the effectiveness of our framework, we conduct experiments on two distinctly styled datasets: ScienceQA [28] and MMStar [10]. ScienceQA underscores the importance of CoT in scientific reasoning tasks, comprising approximately 21,208 multimodal multiple-choice questions covering various science topics. MMStar focuses on evaluating the multimodal capabilities of MLLMs, and consists of 1,500 meticulously curated samples including 6 core capabilities and 18 detailed axes. Each sample exhibits visual dependency and minimal data leakage. + +Candidate MLLMs. Our MLLM pool includes 5 closed-source models and 1 open-source model. The open-source + +
ModelsMMstarScienceQA
Pearson (↑)Spearman (↑)Per.Ent. (↓)Pearson (↑)Spearman (↑)Per.Ent. (↓)
LLama-3.2-11b-V0.314±0.07570.550±0.05770.983±0.23100.160±0.04300.225±0.09571.099±0.0000
Claude-3-haiku0.095±0.13010.225±0.15001.099±0.0000-0.145±0.0308-0.525±0.20611.099±0.0000
Claude-3.5-sonnet0.780±0.01460.825±0.05000.752±0.23100.437±0.04520.450±0.17320.752±0.2310
Gemini-1.5-pro0.864±0.00680.850±0.05770.637±0.00000.4147±0.04640.725±0.20620.434±0.5352
GPT-4o-mini0.668±0.00730.725±0.12580.868±0.28860.3354±0.06350.600±0.00001.099±0.0000
GPT-4o0.878±0.00380.875±0.00500.159±0.31830.617±0.00710.625±0.12580.637±0.0000
Methods
Peer Review0.725±0.00440.771±0.16161.040±0.28300.463±0.01930.686±0.17771.040±0.2830
Majority Vote [42]0.757±0.00130.757±0.08571.299±0.17330.509±0.01810.524±0.06601.040±0.0000
Rating Vote [2]0.795±0.00130.743±0.23090.628±0.07550.623±0.00840.629±0.18950.920±0.2387
PRD [22]0.838±0.00270.864±0.03170.427±0.00870.636±0.00420.694±0.07340.746±0.0016
UPME0.944±0.00110.972±0.03300.141±0.28120.814±0.00240.886±0.02860.422±0.2812
+ +Table 1. Main results of different models and methods on MMstar and ScienceQA datasets, evaluated by Pearson, Spearman, and Permutation Entropy metrics. Since the model cannot judge itself, scores from the other five models involved in the computations are used for evaluation. Optimal results are highlighted in bold and underlined. Standard deviation (std) is calculated based on four runs. + +model is Llama-3.2-11b-vision-instruct [29], which adapts the Llama-3 architecture to the multimodal domain through visual instruction tuning [24]. The closed-source models consist of GPT-4o [35], GPT-4o-mini [33], Claude-3.5-Sonnet [3], Claude-3-Haiku [4], and Gemini-1.5-Pro [45]. These commercial models employ enhanced human alignment strategies to optimize performance and user experience across various applications. + +Metrics. For all experiments, we employ three popular metrics to evaluate the setups mentioned above and our UPME method: the Pearson Correlation Coefficient [14] to evaluate the similarity of score lists, Spearman's Rank Correlation Coefficient [54] to measure the correlation of ranking lists, and Permutation Entropy [5] to assess the complexity and unpredictability of the scoring distributions. All experiments are performed for 4 runs with different seeds (seed = 1, 2, 3, 4) and record the average results. + +Comparison Methods. Since we are the first to investigate unsupervised peer review in MLLM evaluations, we choose the original peer review mechanism and two popular classical methods, Majority Vote [42] and Rating Vote [2], for our comparative analysis. We also adopt a semi-supervised method, PRD [22], and conduct experiments using individual models as judges to further bolster our comparisons. + +# 4.2. Experimental Results + +# 4.2.1. Comparison Results + +The experimental results in Table 1 indicate that among the candidate MLLMs, GPT-4o exhibits the highest review capability on the MMstar dataset, achieving Pearson and Spearman correlation coefficients of 0.878 and 0.875, respectively. In contrast, the open-source LLama-3.2 model demonstrates weaker performance. Although Claude-3.5- + +![](images/5eb990f3a52d71235841e3e6a9dae03857815000f623146822506a110345d56a.jpg) +Figure 3. Convergence experiments. + +Sonnet and Gemini-1.5-Pro claim to offer performance comparable to GPT-4o, they still show limitations when evaluating other models, falling noticeably short of human-level performance. The original peer review method yields Pearson and Spearman coefficients lower than that of GPT-4o, indicating that the biases of weaker models influence the Peer Review approach. Our UPME method achieves a Pearson correlation of 0.944 and a Spearman correlation of 0.972 on the MMstar dataset, demonstrating a high degree of similarity with human annotations and validating the effectiveness of our optimizations in visual QA comprehension and consistency evaluation. + +On the ScienceQA dataset, even the best-performing GPT-4o achieves Pearson and Spearman correlations of 0.617 and 0.625, reflecting the difficulty of the task. The original peer review method shows more modest results, with Pearson and Spearman of 0.463 and 0.686, respectively. Following a series of optimizations and alignments, UPME achieves Pearson and Spearman correlations + +
MMstarScienceQA
MethodCor.Vis.Cti.PearsonSpearmanPearsonSpearman
Peer Review0.7270.7140.4570.600
Score Optimization✓0.8540.7710.7130.657
Visual Alignment✓0.8730.8860.7010.771
✓0.7850.8290.5480.771
✓✓0.9030.9430.7750.886
UPME✓✓✓0.9440.9720.8140.886
+ +Table 2. Ablation study with Pearson and Spearman similarity metrics for each dataset. Optimal results are highlighted in bold and underlined. 'Cor.' stands for correctness, 'Vis.' stands for visual understanding, and 'Cli.' stands for Clip. + +of 0.814 and 0.886, showcasing its robustness and superior performance in complex datasets. While the performance on ScienceQA may not be as high, it is important to note that traditional datasets like ScienceQA might not fully align with the evaluation needs of current MLLMs because they may lack visual dependency or sufficient discrimination [10, 52]. UPME achieves higher Pearson and Spearman correlations, demonstrating its robustness and superior performance. + +Moreover, as shown in Figure 3, the loss consistently decreases, and the Pearson similarity metric increases over epochs in 64 different initial settings, which demonstrates the convergence of $w$ and $\hat{G}$ . Under the current experimental setup (25 images and 1500 judgments), the framework reliably converges within the 30-epoch limit. + +# 4.2.2. Ablation Study + +In the ablation study presented in Table 2, we evaluated the impact of progressively introducing different components on model performance. The Score Optimization phase showed consistent improvements in Pearson and Spearman correlation coefficients as modules were added. For instance, using only the correctness module resulted in Pearson scores of 0.854 for MMstar and 0.713 for ScienceQA. Introducing the visual understanding module alone yielded scores of 0.873 and 0.701, respectively. Adding the Clip component alone gave scores of 0.785 for MMstar and 0.548 for ScienceQA. Combining visual understanding and Clip modules further improved performance to 0.903 and 0.775. Ultimately, integrating all components within the UPME framework resulted in the highest performance, with Pearson and Spearman correlations of 0.944 and 0.972 on MMstar and 0.814 and 0.886 on ScienceQA. These results confirm that optimizing the scoring mechanism and visual alignment significantly boosts evaluation accuracy and robustness across different datasets. + +![](images/73e9364cdc08f25fb677a1f3d656d596947032938587d4d51501487e4168d6d4.jpg) + +![](images/cc6be181d7516aeafea998fd44d2af27f74ac0573fa1603eccdb78fe39fb8c04.jpg) +Figure 4. The performance of UPME in different sample size. + +# 4.2.3.Impact of Sample Size + +In Figure 4, we analyze the impact of sample size on the alignment with human-annotated QA benchmarks, measured through Pearson and Spearman correlation coefficients, as well as Permutation Entropy. As the sample size increases from 5 to 25, there is a clear improvement in both correlation metrics and a decline in Permutation Entropy, indicating that the generated score lists and rank lists show greater alignment with human evaluations. Beyond a sample size of 25, the metrics stabilize, showing minimal changes with further increases from 25 to 100 in sample size. This stabilization suggests that the consistency of the generated outputs relative to human judgments has reached a satisfactory level, making 25 samples a reliable point for evaluation. Therefore, we primarily selected a sample size of 25 for our experiments, as it provides an effective balance between alignment accuracy and data efficiency. + +# 5. Human Preference Alignment Analysis + +The aforementioned experiments primarily reflect the consistency between the UPME framework and the predefined QA-based benchmarks. To further explore the alignment between UPME and human preferences, as well as to examine the two types of biases inherent in MLLM-as-Judge approaches, we selected a subset of responses for human evaluation and analysis of UPME scoring results. The detailed experimental design is provided in subsection 9.4. + +# 5.1. Alignment with Human Preference + +To validate whether our proposed UPME method can align with human preferences without manual labeling, we conducted experiments on the MMstar and ScienceQA datasets. The experimental results on Table 3 indicate that the baseline method exhibited relatively low human agreement and model consistency rates, suggesting that the Peer + +
DatasetMethodAgreement (%)Consistency (%)
MMstarPeer Review71.167.5
UPME95.989.8
ScienceQAPeer Review68.261.8
UPME87.482.6
+ +![](images/8b207c8b62b9f05df7ca8f3bb4836617aa410c51583b90498226f094c3398ce0.jpg) +Figure 5. Model accuracy comparison in peer review framework w/ and w/o UPME, where Peer Review_Cor. represents the correctness of the original peer review, and UPME_Cor. and UPME_Vis. correspond to the two judgment dimensions of response correctness and visual understanding, introduced in subsection 3.2. +Figure 6. Heatmap and Table for Peer Review and UPME on Self Preference and Verbosity Bias. + +Review mechanism under an unsupervised setting without weight optimization struggles to align with human preferences. In contrast, UPME demonstrated significant improvements by incorporating Correctness, Visual Understanding, and Clip Correlation. On the MMStar, UPME achieved an agreement rate of $95.9\%$ and a consistency rate of $89.8\%$ , showing that the optimized scoring criteria significantly enhance the accuracy of evaluation outputs and alignment with human preferences. By capturing key multimodal understanding metrics without relying on manual labeling, UPME effectively achieved high consistency with human annotations, highlighting its substantial advantages in improving response consistency and accuracy under an unsupervised framework. + +# 5.2. Performance w/ and w/o UPME + +To assess the improvement in model judgment accuracy within the Peer Review mechanism and the UPME framework, we conducted an experiment, the results are displayed in Figure 5. The calculation of accuracy is based on $ACC = \frac{\mathrm{judge}_{\mathrm{correct}}}{\mathrm{judge}_{\mathrm{all}}}$ . Experimental results show that the UPME significantly improves the accuracy of models in + +![](images/cec5f000ca8beaf39c46270dc76f11a7523700688f4ef471fce2898193d490e1.jpg) +(a) Verbosity Bias +(b) Self Preference + +Table 3. Human Agreement (%) and Human Consistency (%) Rates on Various Datasets. + +
Bias TypeMethodChi-Squarep-valuePhi Coefficient
VerbosityPeer review10.9960.000910.135
BiasUPME0.2800.596960.022
SelfPeer review39.5843.142e-100.257
PreferenceUPME3.4890.06180.076
+ +both correctness and visual understanding compared to the original correctness judgment. Specifically, the average accuracy in correctness evaluation increased from $61.56\%$ in the baseline to $74.48\%$ with UPME, an improvement of $12.92\%$ . This indicates a significant advantage of UPME in assessing the correctness of model responses. Furthermore, in the evaluation of visual understanding, UPME achieved an average accuracy of $73.93\%$ , which is also notably higher than the baseline. This improvement is primarily attributed to UPME's unsupervised evaluation framework, which employs a multi-model peer review mechanism that combines correctness and visual understanding scores to achieve a more efficient and accurate assessment. + +# 5.3. Mitigating the Bias of MLLM-as-a-Judge + +Mitigation of Verbosity bias. As shown in Figure 6, the verbosity-preference bias is significantly alleviated by our method. The original Peer Review approach had a p-value of 0.00091 ( $< 0.05$ ), with a Chi-square value of 10.996 and a Phi coefficient of 0.135, suggesting a significant preference for verbose responses during evaluation. In contrast, the UPME framework yielded a p-value of 0.59696 ( $>0.05$ ), with the Chi-square value reduced to 0.280 and the Phi coefficient decreased to 0.022, indicating that it does not exhibit a significant bias toward selecting verbose responses. Additionally, the number of incorrect selections of verbose responses dropped from 233 to 103 in the UPME framework. Thus, the UPME approach effectively mitigates the bias toward verbose responses in models, enhancing evaluation diversity and alignment with human preferences in an unsupervised setting. + +Mitigation of Self-preference. Self-preference bias is also significantly alleviated, as shown in Figure 6. In the original Peer Review method, the Chi-square value was 39.584, the Phi coefficient was 0.257, and the p-value was + +![](images/54aa3c0f1f355d62b9495a9952c98b432f4ff591c547c77f5f6a0d01c519bd03.jpg) + +Original Question: Where is the seated person visible in the image? + +Disability of review model + +Options: A: On the fence, B: In the field, C: Outside the field, D: Behind the bench + +Candidate 1 (Gemini-1.5-pro): "...Therefore, the seated person is visible outside the field. [[C]]" + +Candidate 2 (Claude-3.5-sonnet): "...The closest answer would be that the player (though not seated) is in the field. [[B]]" + +Review model (Claude-3-haiku) [Correctness judgment] "...However, Assistant B's response correctly concludes that there is no seated person visible in the image. [[2]]" (Human judgement: [[1]]) + +Question generated by review model: What position is the player in the image playing? + +Candidate 1 (Gemini-1.5-pro): "... The player is on a baseball field..." + +Candidate 2 (Claude-3.5-sonnet) : "...Overall theme: The image shows a baseball player in action during a game..." + +Review model (Claude -3-haiku): "...However, Assistant A provided thorough and insightful analyses of the image is slightly more detailed and precise in its visual assessment. [[1]]" (Human judgement: [[1]]) + +![](images/26aa8f5791392d98337dd35d390c6abe17b8a07e501692dce7cfa95724466b3d.jpg) + +Original Question: Which is right? They are () each other. + +\section*{Verbosity bias} + +Options: A: next to, B: far from, C: facing, D: backing", + +Candidate 1 (Gemini-1.5-pro): "... Determine the best fit: Two toys are next to each other. [[A]]" (194 tokens) + +Candidate 2 (Claude-3.5-sonnet): "... Based on this analysis, two toys are next to each other. [[A]]" (396 tokens) + +Review model (GPT-40 mini) [Visual judgment]: "... Overall, Assistant B demonstrates a deeper level of visual understanding by providing a richer description of the scene [[2]]" (Human judgement: [[1]]) + +Question generated by review model: What are the two teddy bears looking at in the image? + +Candidate 1 (Gemini-1.5-pro) : "...The boats on the water and the mountains in the distance." (235 tokens) + +Candidate 2 (Claude-3.5-sonnet) : "...The image shows a baseball player in action during a game..." (448 tokens) + +Review model (GPT-40 mini): "... I find that Assistant A's response demonstrates better visual understanding and reasoning due to its focus on the emotional and scenic aspects of the view. [[1]]" (Human judgement: [[1]]) + +![](images/cb517cf23afb8e7b8289ed69f90f839c5a3f7bcb38858dd1e46910b6cb19dd3b.jpg) +Figure 7. Case study illustration of UPME. We provide the original human-designed questions and the UPME-generated questions, along with the answer analysis. The upper case presents the Disability of review model, where the review model can not answer the original question itself. The middle case demonstrates cases exhibiting morbidity bias. The bottom case shows self-preference bias. + +Original Question: " What is the overall mood of this image? + +Self-preference + +Options: A: Excitement and liveliness, B: Fear and anxiety, C: Calmness and relaxation, D: Confusion and chaos + +Candidate 1 (Claude-3.5-sonnet): "... Considering these factors, the mood that best describes this image is [[C]]." + +Candidate 2 (GPT-4o): "... Given the analysis, the most fitting mood is Calmness and relaxation. [[C]]" + +Review model (GPT-40 mini) [Visual judgment]: "...Assistant B provides a slightly clearer rationale for why the mood aligns with calmness rather than introducing conflicting emotions like melancholy. [[2]]" (Human judgement: [[1]]) + +Question generated by review model: What atmospheric conditions are depicted in the image, and how do they affect visibility? + +Candidate 1 (Gemini-1.5-pro): "... The atmospheric conditions depicted in the image are heavy nighttime rainfall." + +Candidate 2 (Claude-3.5-sonnet): "...The most prominent atmospheric condition depicted is falling snow..." + +Review model (GPT-40 mini): "... I find that Assistant A's response demonstrates better visual understanding and reasoning due to its focus on the emotional and scenic aspects of the view. [[1]]" (Human judgement: [[1]]) + +3.142e-10, indicating a highly significant bias. In contrast, the UPME framework reduced the Chi-square value to 3.489, the Phi coefficient to 0.076, and the p-value to 0.0618, approaching insignificance. The incorrect selections due to self-preference dropped from 226 to 129, further demonstrating our effectiveness in mitigating this bias. + +# 5.4. Case Study + +We selected three cases to demonstrate how the UPME effectively addresses the core issues in traditional MLLM-as-a-judge as shown in Figure 7. In the first case, the review model initially failed to identify the error in Candidate B's response because it was unable to answer the question itself. After applying UPME, the review model was able to generate a question within its capabilities and accurately determine that Candidate A's response showed a deeper understanding of the image, which aligned with human preferences, thus avoiding the misjudgment issue. In the second case, although Candidate B's response was longer, it did not + +provide additional useful information. Traditional evaluations might wrongly consider the longer answer as more informative, but with UPME, the review model, through precise text evaluation, clearly identified Candidate A's concise and clear answer as more valuable, thus avoiding morbidity bias. Finally, in the third case, UPME effectively avoided self-preference bias, enhancing the fairness and objectivity. + +# 6. Conclusion + +This study proposes an unsupervised peer review-based framework, UPME, effectively reducing the annotation workload in traditional human-designed QA benchmarks. Our vision-language scoring system emphasizes a vision-oriented judgment, addressing issues such as verosity and self-preference biases inherent in MLLM-as-judge evaluation methods. Further experiments validate that our evaluation results achieve higher alignment with human preferences, providing a promising and reliable approach. + +# Acknowledgement + +This work was supported by Alibaba DAMO Academy through the Alibaba Innovative Research Program. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1 +[2] Mohammad Allahbakhsh and Aleksandar Ignjatovic. Rating through voting: An iterative method for robust rating. arXiv preprint arXiv:1211.0390, 2012. 5 +[3] Anthropic. Claude 3.5: A sonnet. https://www.anthropic.com/news/claude-3-5-sonnet, 2024.1,5 +[4] Anthropic. Claude 3 haiku. https://www.anthropic.com/news/claude-3-haiku, 2024.5, 1 +[5] Christoph Bandt and Bernd Pompe. Permutation entropy: a natural complexity measure for time series. Physical review letters, 88(17):174102, 2002. 5 +[6] Han Bao, Yue Huang, Yanbo Wang, Jiayi Ye, Xiangqi Wang, Xiuying Chen, Mohamed Elhoseiny, and Xiangliang Zhang. Autobench-v: Can large vision-language models benchmark themselves? arXiv preprint arXiv:2410.21259, 2024. 2 +[7] Yuzhe Cai, Shaoguang Mao, Wenshan Wu, Zehua Wang, Yaobo Liang, Tao Ge, Chenfei Wu, Wang You, Ting Song, Yan Xia, et al. Low-code llm: Visual programming over llms. arXiv preprint arXiv:2304.08103, 2, 2023. 1 +[8] Dongping Chen, Ruoxi Chen, Shilin Zhang, Yinuo Liu, Yaochen Wang, Huichi Zhou, Qihui Zhang, Pan Zhou, Yao Wan, and Lichao Sun. Mllm-as-a-judge: Assessing multimodal llm-as-a-judge with vision-language benchmark. arXiv preprint arXiv:2402.04788, 2024. 2 +[9] Dongping Chen, Yue Huang, Siyuan Wu, Jingyu Tang, Liuyi Chen, Yilin Bai, Zhigang He, Chenlong Wang, Huichi Zhou, Yiqiang Li, et al. Gui-world: A dataset for gui-oriented multimodal llm-based agents. arXiv preprint arXiv:2406.10819, 2024. 1 +[10] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 1, 2, 4, 6 +[11] Zhaorun Chen, Yichao Du, Zichen Wen, Yiyang Zhou, Chenhang Cui, Zhenzhen Weng, Haoqin Tu, Chaoqi Wang, Zhengwei Tong, Leria HUANG, et al. Mj-bench: Is your multimodal reward model really a good judge? In ICML 2024 Workshop on Foundation Models in the Wild, 2024. 2 +[12] Wei-Lin Chiang, Lianmin Zheng, Ying Sheng, Anastasios Nikolas Angelopoulos, Tianle Li, Dacheng Li, Hao Zhang, Banghua Zhu, Michael Jordan, Joseph E. Gonzalez, and Ion Stoica. Chatbot arena: An open platform for evaluating llms by human preference, 2024. 2 +[13] Zhumin Chu, Qingyao Ai, Yiteng Tu, Haitao Li, and Yiqun Liu. Pre: A peer review based large language model evaluator. arXiv preprint arXiv:2401.15641, 2024. 3 + +[14] Israel Cohen, Yiteng Huang, Jingdong Chen, Jacob Benesty, Jacob Benesty, Jingdong Chen, Yiteng Huang, and Israel Cohen. Pearson correlation coefficient. Noise reduction in speech processing, pages 1-4, 2009. 5 +[15] Chunyuan Deng, Yilun Zhao, Xiangru Tang, Mark Gerstein, and Arman Cohan. Benchmark probing: Investigating data leakage in large language models. In NeurIPS 2023 Workshop on Backdoors in Deep Learning-The Good, the Bad, and the Ugly, 2023. 2 +[16] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 1 +[17] Chujie Gao, Qihui Zhang, Dongping Chen, Yue Huang, Siyuan Wu, Zhengyan Fu, Yao Wan, Xiangliang Zhang, and Lichao Sun. The best of both worlds: Toward an honest and helpful large language model. arXiv preprint arXiv:2406.00380, 2024. 2 +[18] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Yuxiao Dong, et al. Co-gagent: A visual language model for gui agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14281-14290, 2024. 1 +[19] Zdeněk Kasner and Ondřej Dusěk. Beyond traditional benchmarks: Analyzing behaviors of open llms on data-to-text generation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12045–12072, 2024. 1 +[20] Zdeněk Kasner and Ondřej Dušek. Beyond traditional benchmarks: Analyzing behaviors of open llms on data-to-text generation, 2024. 2 +[21] Jian Li and Weiheng Lu. A survey on benchmarks of multimodal large language models. arXiv preprint arXiv:2408.08632, 2024. 1 +[22] Ruosen Li, Teerth Patel, and Xinya Du. Prd: Peer rank and discussion improve large language model based evaluations. arXiv preprint arXiv:2307.02762, 2023. 5, 2 +[23] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1 +[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 5 +[25] Xiao Liu, Xuanyu Lei, Shengyuan Wang, Yue Huang, Zhuoer Feng, Bosi Wen, Jiale Cheng, Pei Ke, et al. Align-bench: Benchmarking chinese alignment of large language models. arXiv preprint arXiv:2311.18743, 2023. 2 +[26] Yinhong Liu, Han Zhou, Zhijiang Guo, Ehsan Shareghi, Ivan Vulic, Anna Korhonen, and Nigel Collier. Aligning with human judgement: The role of pairwise preference in large language model evaluators. arXiv preprint arXiv:2403.16950, 2024. 2, 4 +[27] Adian Liusie, Potsawee Manakul, and Mark JF Gales. Zero-shot nlg evaluation through pairware comparisons with llms. arXiv preprint arXiv:2307.07889, 2023. 4 + +[28] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 1, 2, 4 +[29] Meta. Llama 3.2 11b-vision-instruct. https://huggingface.co/meta-llama/Llama-3.2-11B-Vision-Instruct, 2024.5, 1 +[30] Birger Moell. Evaluating large language models with human feedback: Establishing a swedish benchmark. arXiv preprint arXiv:2405.14006, 2024. 2 +[31] Kun-Peng Ning, Shuo Yang, Yu-Yang Liu, Jia-Yu Yao, Zhen-Hui Liu, Yu Wang, Ming Pang, and Li Yuan. Peer-review-in-llms: Automatic evaluation method for llms in open-environment. arXiv preprint arXiv:2402.01830, 2024. 2, 3 +[32] OpenAI. Gpt-4 technical report. https://openai.com/research/gpt-4, 2023.1 +[33] OpenAI. Gpt-4o mini: Advancing cost-efficient intelligence. https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/, 2024.5,1 +[34] OpenAI. Introducing openai o1: Advancing model reasoning with visual and logical insights. https://openai.com/index/introducing-openai-o1-preview/, 2024.4 +[35] OpenAI. Hello gpt-4o, 2024. 5, 1 +[36] Renjie Pi, Lewei Yao, Jiahui Gao, Jipeng Zhang, and Tong Zhang. Perceptiongpt: Effectively fusing visual perception into lIm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27124-27133, 2024. 1 +[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4, 3 +[38] Vyas Raina, Adian Liusie, and Mark Gales. Is llm-as-a-judge robust? investigating universal adversarial attacks on zero-shot llm assessment. arXiv preprint arXiv:2402.14016, 2024. 2 +[39] Refuel.ai. Llm labeling: Technical report, 2025. 2 +[40] Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaqing Ellen Tan, Yossi Adi, Jingyu Liu, Romain Sauvestre, Tal Remez, et al. Code llama: Open foundation models for code. arXiv preprint arXiv:2308.12950, 2023. 1 +[41] Andreas Stephan, Dawei Zhu, Matthias Aßenmacher, Xiaoyu Shen, and Benjamin Roth. From calculation to adjudication: Examining llm judges on mathematical reasoning tasks. arXiv preprint arXiv:2409.04168, 2024. 3 +[42] James Surowiecki. The wisdom of crowds/james surowiecki. NY.: Anchor, 2005. 5 +[43] Sijun Tan, Siyuan Zhuang, Kyle Montgomery, William Y Tang, Alejandro Cuadron, Chenguang Wang, Raluca Ada + +Popa, and Ion Stoica. Judgebench: A benchmark for evaluating llm-based judges. arXiv preprint arXiv:2410.12784, 2024.3 +[44] Ruixue Tang, Chao Ma, Wei Emma Zhang, Qi Wu, and Xiaokang Yang. Semantic equivalent adversarial data augmentation for visual question answering. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16, pages 437-453. Springer, 2020. 1 +[45] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 5, 1 +[46] Li Wang, Hui Zhang, and Rajesh Patel. A user-centric benchmark for evaluating large language models. arXiv preprint arXiv:2404.13940, 2024. 3 +[47] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 1, 2 +[48] Siyuan Wu, Yue Huang, Chujie Gao, Dongping Chen, Qihui Zhang, Yao Wan, Tianyi Zhou, Xiangliang Zhang, Jianfeng Gao, Chaowei Xiao, et al. Unigen: A unified framework for textual dataset generation using large language models. arXiv preprint arXiv:2406.18966, 2024. 1 +[49] xAI. Realworldqa: A benchmark for evaluating real-world spatial understanding in multimodal ai models, 2024. 2 +[50] Ruijie Xu, Zengzhi Wang, Run-Ze Fan, and Pengfei Liu. Benchmarking benchmark leakage in large language models. arXiv preprint arXiv:2404.18824, 2024. 2 +[51] Jiayi Ye, Yanbo Wang, Yue Huang, Dongping Chen, Qihui Zhang, Nuno Moniz, Tian Gao, Werner Geyer, Chao Huang, Pin-Yu Chen, et al. Justice or prejudice? quantifying biases in llm-as-a-judge. arXiv preprint arXiv:2410.02736, 2024. 2 +[52] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 1, 2, 6 +[53] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 1, 2 +[54] Jerrold H Zar. Spearman rank correlation. Encyclopedia of biostatistics, 7, 2005. 5 +[55] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. 1, 2 +[56] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, HongFa Wang, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei Li, et al. Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. arXiv preprint arXiv:2310.01852, 2023. 1 \ No newline at end of file diff --git a/2025/UPME_ An Unsupervised Peer Review Framework for Multimodal Large Language Model Evaluation/images.zip b/2025/UPME_ An Unsupervised Peer Review Framework for Multimodal Large Language Model Evaluation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..739d2ffd22ce51c77f0afaa762f6159a66c4a072 --- /dev/null +++ b/2025/UPME_ An Unsupervised Peer Review Framework for Multimodal Large Language Model Evaluation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d343dc23f0cd86aa5c224ce6d35bd11465634dcbcfbac72a37d23ba2bcf1dc76 +size 563302 diff --git a/2025/UPME_ An Unsupervised Peer Review Framework for Multimodal Large Language Model Evaluation/layout.json b/2025/UPME_ An Unsupervised Peer Review Framework for Multimodal Large Language Model Evaluation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..59b1ef80f6925ecbf6ec6a18644fa7e92ec435c2 --- /dev/null +++ b/2025/UPME_ An Unsupervised Peer Review Framework for Multimodal Large Language Model Evaluation/layout.json @@ -0,0 +1,9390 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 137, + 103, + 475, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 103, + 475, + 140 + ], + "spans": [ + { + "bbox": [ + 137, + 103, + 475, + 140 + ], + "type": "text", + "content": "UPME: An Unsupervised Peer Review Framework for Multimodal Large Language Model Evaluation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 159, + 501, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 159, + 501, + 193 + ], + "spans": [ + { + "bbox": [ + 108, + 159, + 501, + 193 + ], + "type": "text", + "content": "Qihui Zhang" + }, + { + "bbox": [ + 108, + 159, + 501, + 193 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 108, + 159, + 501, + 193 + ], + "type": "text", + "content": " Munan Ning" + }, + { + "bbox": [ + 108, + 159, + 501, + 193 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 108, + 159, + 501, + 193 + ], + "type": "text", + "content": " Zheyuan Liu" + }, + { + "bbox": [ + 108, + 159, + 501, + 193 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 108, + 159, + 501, + 193 + ], + "type": "text", + "content": " Yue Huang" + }, + { + "bbox": [ + 108, + 159, + 501, + 193 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 108, + 159, + 501, + 193 + ], + "type": "text", + "content": " Shuo Yang" + }, + { + "bbox": [ + 108, + 159, + 501, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 159, + 501, + 193 + ], + "type": "text", + "content": " Yanbo Wang" + }, + { + "bbox": [ + 108, + 159, + 501, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 159, + 501, + 193 + ], + "type": "text", + "content": " Jiayi Ye" + }, + { + "bbox": [ + 108, + 159, + 501, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 159, + 501, + 193 + ], + "type": "text", + "content": " Xiao Chen" + }, + { + "bbox": [ + 108, + 159, + 501, + 193 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 108, + 159, + 501, + 193 + ], + "type": "text", + "content": " Yibing Song" + }, + { + "bbox": [ + 108, + 159, + 501, + 193 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 108, + 159, + 501, + 193 + ], + "type": "text", + "content": " Li Yuan" + }, + { + "bbox": [ + 108, + 159, + 501, + 193 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 92, + 201, + 517, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 201, + 517, + 232 + ], + "spans": [ + { + "bbox": [ + 92, + 201, + 517, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 92, + 201, + 517, + 232 + ], + "type": "text", + "content": "School of Electrical and Computer Engineering, Peking University " + }, + { + "bbox": [ + 92, + 201, + 517, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 92, + 201, + 517, + 232 + ], + "type": "text", + "content": "Hupan Lab " + }, + { + "bbox": [ + 92, + 201, + 517, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 92, + 201, + 517, + 232 + ], + "type": "text", + "content": "DAMO Academy, Alibaba Group " + }, + { + "bbox": [ + 92, + 201, + 517, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 92, + 201, + 517, + 232 + ], + "type": "text", + "content": "University of Notre Dame " + }, + { + "bbox": [ + 92, + 201, + 517, + 232 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 92, + 201, + 517, + 232 + ], + "type": "text", + "content": "Tsinghua University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 284, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 296, + 571 + ], + "type": "text", + "content": "Multimodal Large Language Models (MLLMs) have emerged to tackle the challenges of Visual Question Answering (VQA), sparking a new research focus on conducting objective evaluations of these models. Existing evaluation methods face limitations due to the significant human workload required to design Q&A pairs for visual images, which inherently restricts the scale and scope of evaluations. Although automated MLLM-as-judge approaches attempt to reduce the human workload through automatic evaluations, they often introduce biases. To address these problems, we propose an Unsupervised Peer review MLLM Evaluation framework. It utilizes only image data, allowing models to automatically generate questions and conduct peer review assessments of answers from other models, effectively alleviating the reliance on human workload. Additionally, we introduce the vision-language scoring system to mitigate the bias issues, which focuses on three aspects: (i) response correctness; (ii) visual understanding and reasoning; and (iii) image-text correlation. Experimental results demonstrate that UPME achieves a Pearson correlation of 0.944 with human evaluations on the MMstar dataset and 0.814 on the ScienceQA dataset, indicating that our framework closely aligns with human-designed benchmarks and inherent human preferences." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 585, + 135, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 585, + 135, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 585, + 135, + 597 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": "Recently, Large Language Models (LLMs) have made significant strides in reasoning and application capabilities [3, 32, 47], enabling diverse applications such as text and code generation [1, 19, 40, 48]. Notably, due to the exceptional natural language understanding capabilities of LLMs, they are increasingly employed for automated evaluations, a pro" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 322, + 266, + 548, + 430 + ], + "blocks": [ + { + "bbox": [ + 322, + 266, + 548, + 430 + ], + "lines": [ + { + "bbox": [ + 322, + 266, + 548, + 430 + ], + "spans": [ + { + "bbox": [ + 322, + 266, + 548, + 430 + ], + "type": "image", + "image_path": "2bf165477525d2a24cc964c43cc7b98f431535217d020b712f2136a2a678a6b8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 442, + 555, + 487 + ], + "lines": [ + { + "bbox": [ + 313, + 442, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 555, + 487 + ], + "type": "text", + "content": "Figure 1. Existing methods for evaluating MLLMs face various challenges. Our proposed UPME framework addresses these limitations by leveraging a peer review mechanism, reducing annotation costs, and aligning closely with human judgment." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 498, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 582 + ], + "type": "text", + "content": "cess known as LLM-as-a-Judge [55]. With the integration of vision encoders [36], Multimodal Large Language Models (MLLMs) [24, 32] can incorporate multiple modalities (e.g., text, image, and audio) [23, 56] and showcase remarkable performance in downstream applications, including visual conversation [7] and embodied intelligence [16]. Furthermore, they can act as agents to interact with GUI [9, 18]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": "Subsequently, the challenge of conducting objective and comprehensive evaluations of MLLMs while providing users with authoritative guidelines has emerged as a key research focus. A common approach involves constructing benchmarks based on Visual Question Answering (VQA), which assess the performance of MLLMs across various dimensions by setting image-based questions and answers [10, 28, 52, 53]. However, most of these works heavily depend on substantial human involvement in creating Q&A pairs for images [21, 44], leading to limited evaluation volume and making it insufficient for a comprehen" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 683, + 228, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 228, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 228, + 713 + ], + "type": "text", + "content": "* Equal contribution † Corresponding author {maskhui1003, munanning}@gmail.com yuanli-ece@pu.edu.cn" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9165" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "content": "sive assessment of visual content [6]. Another category of methods is MLLM-as-a-Judge [17, 20, 26, 55], which aims to reduce reliance on human annotations by allowing models to evaluate each other directly. However, these methods often still depend on human-designed questions and cannot entirely eliminate the need for human input. Additionally, a significant challenge is the presence of morbidity and self-preference biases [8, 51], where MLLMs tend to favor longer responses or prefer their own outputs. These biases can lead mutual evaluation mechanisms to deviate from truly understanding visual content." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 205, + 295, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 295, + 431 + ], + "type": "text", + "content": "To address these challenges, we propose UPME, the Unsupervised Peer review MLLM Evaluation framework, which is designed to evaluate MLLMs without requiring human QA annotations, as shown in Figure 1. During each iteration, UPME selects two candidate models and a review model from the MLLM pool. The review model generates questions for a given image and evaluates the responses provided by the candidate models. This evaluation is conducted using a vision-language scoring system that assesses textual response correctness, visual understanding and reasoning, as well as image-text consistency through CLIP scores. Each model is initialized with a confidence weight and alternates roles as candidate and review models. Their estimated scores are calculated based on weighted evaluations provided by the review models. The framework employs dynamic weight optimization, iteratively updating model scores and refining weights through consistency optimization. Ultimately, this process generates a comprehensive and less-biased estimated score list." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 434, + 295, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 589 + ], + "type": "text", + "content": "We selected the MMStar [10] and ScienceQA [28] datasets as benchmarks because they encompass visual content reasoning and multimodal content across various disciplines. Experimental results show that under unsupervised settings, UPME achieved Pearson correlations of 0.944 and 0.814 with human evaluation results on these two datasets, while Spearman similarities reached 0.972 and 0.886, indicating a high degree of similarity between our approach and human-annotated QA benchmarks. Further experiments demonstrated that UPME achieved higher consistency with human assessments compared to purely peer review-based methods, effectively reducing morbidity and self-preference issues in the MLLM-as-a-Judge framework." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 590, + 294, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 590, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 67, + 590, + 294, + 602 + ], + "type": "text", + "content": "Overall, our contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 603, + 295, + 711 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 603, + 295, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 295, + 638 + ], + "type": "text", + "content": "- We propose the first Unsupervised Peer review MLLM Evaluation (UPME) framework, addressing the challenge of reliance on human annotations in MLLM evaluation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 639, + 294, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 294, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 294, + 674 + ], + "type": "text", + "content": "- Our framework incorporates a Vision-Language Scoring System that targets visual performance and image-text association, providing a visually-focused evaluation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 675, + 294, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 675, + 294, + 711 + ], + "spans": [ + { + "bbox": [ + 55, + 675, + 294, + 711 + ], + "type": "text", + "content": "- We conduct extensive experiments demonstrating that UPME achieves high consistency with human QA assessments and strong alignment with human preferences." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 91, + 430, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 430, + 102 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 430, + 102 + ], + "type": "text", + "content": "2.1. MLLM Benchmark" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 109, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 555, + 348 + ], + "type": "text", + "content": "MLLM benchmarks rely on human-annotated QA pairs to measure the accuracy of MLLMs in answering questions. They focus on the different capabilities of MLLMs. ScienceQA [28] evaluate the multi-turn reasoning ability for a broad range of scientific subjects by introducing Chain of Thought (CoT) [47]; MMVet [52] presents an LLM-based open-ended output evaluation method, enabling unified scoring across various question types and answer formats; MMMU [53], inspired by MMLU, primarily assesses model performance in complex reasoning tasks requiring college-level knowledge, covering six major disciplines and 30 fields with multimodal questions; MMStar [10] focuses on creating a benchmark requiring visual content for reasoning, rigorously selecting visual-dependent samples and detecting data leakage to ensure genuine performance gains from multimodal training; RealWorldQA [49] is designed to assess the ability of MLLMs to understand real-world space. Although these benchmarks propose various evaluation methods, their evaluation scope is limited by the reliance on substantial human annotation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 357, + 426, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 357, + 426, + 370 + ], + "spans": [ + { + "bbox": [ + 314, + 357, + 426, + 370 + ], + "type": "text", + "content": "2.2. MLLM-as-a-Judge" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 376, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 555 + ], + "type": "text", + "content": "To reduce the human annotation workload, LLM-as-a-Judge [55] is proposed to enable models to perform mutual evaluation. This approach primarily involves using LLMs to assess responses in two ways: comparing pairs of answers to determine superiority [20, 25, 55], or directly scoring individual answers based on specific criteria [17, 25]. This strategy is extended to MLLM-as-a-Judge [8] to assess the performance of MLLMs across multiple evaluation tasks. MJ-Bench [11] evaluates the effectiveness of multimodal reward models in text-to-image generation. However, these methods face two main biases [38, 51]: the verbose bias and self-preference bias, which means MLLMs prefer the answers from themselves or with longer text length. Human preferences have also not been effectively incorporated into these automated evaluation systems [31]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 563, + 509, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 563, + 509, + 575 + ], + "spans": [ + { + "bbox": [ + 314, + 563, + 509, + 575 + ], + "type": "text", + "content": "2.3. Human Preference-based Evaluation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "Predefined image and QA-pair benchmarks for MLLMs may encounter issues like data leakage and targeted optimization, which can prevent an accurate reflection of the true capabilities of these models [15, 50]. Therefore, evaluation methods based on human preferences offer a better assessment of whether MLLMs meet human needs [30]. Chatbot Arena [12] provides a public platform specifically for evaluating LLMs based on human preference. This platform uses a pairwise comparison approach, where users interact with two anonymous models and select the one that performs better, generating a dynamic leaderboard via the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9166" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 555, + 270 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 555, + 270 + ], + "type": "image", + "image_path": "c665188dc9e4d133a0633e2a63266255a6c89241ccf47717078e05c42661f46e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 277, + 555, + 335 + ], + "lines": [ + { + "bbox": [ + 54, + 277, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 54, + 277, + 555, + 335 + ], + "type": "text", + "content": "Figure 2. The UPME framework consists of three main components: (i) Peer Review Mechanism, where two candidate models and one review model are randomly selected from the MLLM pool. The review model generates questions based on a selected image, and candidate models provide responses. (ii) Vision-Language Judgment Scoring System, which evaluates answers based on textual correctness, visual understanding and reasoning, and image-text correlation. (iii) Dynamic Weight Optimization, ensuring consistency between confidence weights and estimated scores through iterative optimization cycles." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 355, + 294, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 355, + 294, + 427 + ], + "spans": [ + { + "bbox": [ + 54, + 355, + 294, + 427 + ], + "type": "text", + "content": "Elo rating system. Although open-ended user evaluations can more realistically reflect LLM performance in real-world scenarios [46], the high economic cost of collecting large volumes of high-quality user feedback, combined with the long wait time for obtaining such feedback, limits the widespread adoption of this approach." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 436, + 138, + 450 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 138, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 138, + 450 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 456, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 552 + ], + "type": "text", + "content": "Assigning higher weights to better-performing models leads to a significant improvement in the accuracy of the entire evaluation framework. This conclusion is supported by preliminary experiments detailed in subsection 8.1 and confirmed by recent research [13, 31, 41, 43]. Therefore, we focus on characterizing the relationship between model performance and its judging capability through consistency optimization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 559, + 191, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 191, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 191, + 571 + ], + "type": "text", + "content": "3.1. Peer review formulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 576, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 295, + 673 + ], + "type": "text", + "content": "Problem definition. The UPME framework facilitates mutual evaluation among MLLMs via a peer review mechanism, aiming to produce evaluation scores that closely reflect human preferences. Given an unsupervised image dataset " + }, + { + "bbox": [ + 55, + 576, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\{I_i\\}_{i=1}^n" + }, + { + "bbox": [ + 55, + 576, + 295, + 673 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 576, + 295, + 673 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 576, + 295, + 673 + ], + "type": "text", + "content": " images (with no human annotated questions and answers) and a model pool " + }, + { + "bbox": [ + 55, + 576, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\{M_j\\}_{j=1}^m" + }, + { + "bbox": [ + 55, + 576, + 295, + 673 + ], + "type": "text", + "content": " consisting of " + }, + { + "bbox": [ + 55, + 576, + 295, + 673 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 576, + 295, + 673 + ], + "type": "text", + "content": " MLLMs (both open-source and closed-source), we can obtain an UPME-learned score list " + }, + { + "bbox": [ + 55, + 576, + 295, + 673 + ], + "type": "inline_equation", + "content": "\\hat{G}" + }, + { + "bbox": [ + 55, + 576, + 295, + 673 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 100, + 682, + 295, + 697 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 682, + 295, + 697 + ], + "spans": [ + { + "bbox": [ + 100, + 682, + 295, + 697 + ], + "type": "interline_equation", + "content": "\\hat {G} := \\left[ \\hat {G} _ {M _ {1}}, \\hat {G} _ {M _ {2}}, \\hat {G} _ {M _ {3}}, \\dots , \\hat {G} _ {M _ {m}} \\right]. \\tag {1}", + "image_path": "1d552bed9a4e4a299e19fc9b530e84ad82704e5a5df952f57b8d91b4e107eddb.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 701, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 296, + 714 + ], + "type": "text", + "content": "On the other hand, given " + }, + { + "bbox": [ + 67, + 701, + 296, + 714 + ], + "type": "inline_equation", + "content": "G_{M_1}" + }, + { + "bbox": [ + 67, + 701, + 296, + 714 + ], + "type": "text", + "content": " represents the score of" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 355, + 555, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 555, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 555, + 378 + ], + "type": "text", + "content": "model " + }, + { + "bbox": [ + 313, + 355, + 555, + 378 + ], + "type": "inline_equation", + "content": "M_{1}" + }, + { + "bbox": [ + 313, + 355, + 555, + 378 + ], + "type": "text", + "content": " on the human-annotated benchmark, we define the human-annotated score list " + }, + { + "bbox": [ + 313, + 355, + 555, + 378 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 355, + 555, + 378 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 359, + 393, + 555, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 393, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 359, + 393, + 555, + 407 + ], + "type": "interline_equation", + "content": "G := \\left[ G _ {M _ {1}}, G _ {M _ {2}}, G _ {M _ {3}}, \\dots , G _ {M _ {m}} \\right]. \\tag {2}", + "image_path": "fbb7775ba11a91229e6eaa831eba2ada5f373f6b9e71eed32bbb4dd7354418e9.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 414, + 554, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 554, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 554, + 473 + ], + "type": "text", + "content": "Our objective is to optimize the UPME-learned scores " + }, + { + "bbox": [ + 313, + 414, + 554, + 473 + ], + "type": "inline_equation", + "content": "\\hat{G}" + }, + { + "bbox": [ + 313, + 414, + 554, + 473 + ], + "type": "text", + "content": " by maximizing their similarity to the human-annotated score list " + }, + { + "bbox": [ + 313, + 414, + 554, + 473 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 414, + 554, + 473 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 414, + 554, + 473 + ], + "type": "inline_equation", + "content": "\\mathrm{Sim}(G, \\hat{G}) \\to 1" + }, + { + "bbox": [ + 313, + 414, + 554, + 473 + ], + "type": "text", + "content": ", ensuring that UPME-learned scores align closely with human benchmark standards." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 493, + 555, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 555, + 592 + ], + "type": "text", + "content": "Overview of UPME. Figure 2 shows an overview of our UPME framework. During each review iteration, we randomly select two different candidate models " + }, + { + "bbox": [ + 313, + 493, + 555, + 592 + ], + "type": "inline_equation", + "content": "M_{j}" + }, + { + "bbox": [ + 313, + 493, + 555, + 592 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 493, + 555, + 592 + ], + "type": "inline_equation", + "content": "M_{k}" + }, + { + "bbox": [ + 313, + 493, + 555, + 592 + ], + "type": "text", + "content": " from the pool " + }, + { + "bbox": [ + 313, + 493, + 555, + 592 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 493, + 555, + 592 + ], + "type": "text", + "content": " for comparison, along with another model " + }, + { + "bbox": [ + 313, + 493, + 555, + 592 + ], + "type": "inline_equation", + "content": "M_{r}" + }, + { + "bbox": [ + 313, + 493, + 555, + 592 + ], + "type": "text", + "content": " to serve as the reviewer. " + }, + { + "bbox": [ + 313, + 493, + 555, + 592 + ], + "type": "inline_equation", + "content": "M_{r}" + }, + { + "bbox": [ + 313, + 493, + 555, + 592 + ], + "type": "text", + "content": " generates a question " + }, + { + "bbox": [ + 313, + 493, + 555, + 592 + ], + "type": "inline_equation", + "content": "Q_{i}^{r} = M_{r}(I_{i})" + }, + { + "bbox": [ + 313, + 493, + 555, + 592 + ], + "type": "text", + "content": " for each image " + }, + { + "bbox": [ + 313, + 493, + 555, + 592 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 313, + 493, + 555, + 592 + ], + "type": "text", + "content": " and then prompts the candidate models " + }, + { + "bbox": [ + 313, + 493, + 555, + 592 + ], + "type": "inline_equation", + "content": "M_{j}" + }, + { + "bbox": [ + 313, + 493, + 555, + 592 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 493, + 555, + 592 + ], + "type": "inline_equation", + "content": "M_{k}" + }, + { + "bbox": [ + 313, + 493, + 555, + 592 + ], + "type": "text", + "content": " to provide responses " + }, + { + "bbox": [ + 313, + 493, + 555, + 592 + ], + "type": "inline_equation", + "content": "A_{i}^{j,r}" + }, + { + "bbox": [ + 313, + 493, + 555, + 592 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 493, + 555, + 592 + ], + "type": "inline_equation", + "content": "A_{i}^{k,r}" + }, + { + "bbox": [ + 313, + 493, + 555, + 592 + ], + "type": "text", + "content": ", respectively. The answers can be written as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 343, + 601, + 555, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 601, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 343, + 601, + 555, + 617 + ], + "type": "interline_equation", + "content": "A _ {i} ^ {j, r} = M _ {j} \\left(I _ {i}, Q _ {i} ^ {r}\\right), \\quad A _ {i} ^ {k, r} = M _ {k} \\left(I _ {i}, Q _ {i} ^ {r}\\right). \\tag {3}", + "image_path": "c03a484bbc5d14be7417563475f9f46b6f9339b6ff431686415452ad3ac9cee0.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 628, + 554, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 554, + 678 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 554, + 678 + ], + "type": "text", + "content": "Based on the vision-language judgment scoring System " + }, + { + "bbox": [ + 313, + 628, + 554, + 678 + ], + "type": "inline_equation", + "content": "S_{VL}" + }, + { + "bbox": [ + 313, + 628, + 554, + 678 + ], + "type": "text", + "content": " (detailed in subsection 3.2), " + }, + { + "bbox": [ + 313, + 628, + 554, + 678 + ], + "type": "inline_equation", + "content": "M_r" + }, + { + "bbox": [ + 313, + 628, + 554, + 678 + ], + "type": "text", + "content": " then evaluates these responses, and compares the responses of " + }, + { + "bbox": [ + 313, + 628, + 554, + 678 + ], + "type": "inline_equation", + "content": "M_j" + }, + { + "bbox": [ + 313, + 628, + 554, + 678 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 628, + 554, + 678 + ], + "type": "inline_equation", + "content": "M_k" + }, + { + "bbox": [ + 313, + 628, + 554, + 678 + ], + "type": "text", + "content": " to produce a review score " + }, + { + "bbox": [ + 313, + 628, + 554, + 678 + ], + "type": "inline_equation", + "content": "\\mathrm{Review}_i^{j,k,r}" + }, + { + "bbox": [ + 313, + 628, + 554, + 678 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 341, + 699, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 699, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 341, + 699, + 555, + 715 + ], + "type": "interline_equation", + "content": "\\operatorname {R e v i e w} _ {i} ^ {j, k, r} = S _ {V L} \\left(A _ {i} ^ {j, r}, A _ {i} ^ {k, r}, Q _ {i} ^ {r}, I _ {i} \\mid M _ {r}\\right). \\tag {4}", + "image_path": "430f0ed6e9c3ac8a55e55c5519992a093c662e1f9e3efe05ac09b676b95cb091.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9167" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 99 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 99 + ], + "type": "text", + "content": "By assigning each model an initiated confidence weight " + }, + { + "bbox": [ + 55, + 72, + 295, + 99 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 72, + 295, + 99 + ], + "type": "text", + "content": ", we can calculate the learned score " + }, + { + "bbox": [ + 55, + 72, + 295, + 99 + ], + "type": "inline_equation", + "content": "\\hat{G}_{M_j}" + }, + { + "bbox": [ + 55, + 72, + 295, + 99 + ], + "type": "text", + "content": " of each MLLM:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 88, + 105, + 295, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 105, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 88, + 105, + 295, + 133 + ], + "type": "interline_equation", + "content": "\\hat {G} _ {M _ {j}} = \\sum_ {i} \\sum_ {k \\neq j} \\sum_ {r \\neq k, r \\neq j} \\operatorname {R e v i w} _ {i} ^ {j, k, r} \\times w _ {r}, \\tag {5}", + "image_path": "7a659873beae4a55104d0441918cd12d6211cb347cbc11b74971c0b7f8189112.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 140, + 295, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 295, + 163 + ], + "type": "text", + "content": "then we can subsequently obtain the UPME-learned score list " + }, + { + "bbox": [ + 55, + 140, + 295, + 163 + ], + "type": "inline_equation", + "content": "\\hat{G}" + }, + { + "bbox": [ + 55, + 140, + 295, + 163 + ], + "type": "text", + "content": ", as illustrated in Equation 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 171, + 282, + 184 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 282, + 184 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 282, + 184 + ], + "type": "text", + "content": "3.2. Vision-Language Judgment Scoring System" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 188, + 296, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 188, + 296, + 284 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 296, + 284 + ], + "type": "text", + "content": "In this subsection, we present our vision-language judgment scoring system, denoted as " + }, + { + "bbox": [ + 55, + 188, + 296, + 284 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{VL}}" + }, + { + "bbox": [ + 55, + 188, + 296, + 284 + ], + "type": "text", + "content": " in Equation 9. Our evaluation is structured around three core criteria: Response Correctness, Visual Understanding and Reasoning, and Image-Text Correlation. The first two aspects are evaluated by employing " + }, + { + "bbox": [ + 55, + 188, + 296, + 284 + ], + "type": "inline_equation", + "content": "M_r" + }, + { + "bbox": [ + 55, + 188, + 296, + 284 + ], + "type": "text", + "content": ", which reviews responses based on prompts outlined in Section 10. The Image-Text Correlation is measured using CLIP scores." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 293, + 296, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 296, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 296, + 390 + ], + "type": "text", + "content": "Response Correctness. The primary criterion is the correctness of the candidate models' responses to questions posed by the review model. Existing studies indicate that paired comparative evaluation yields better accuracy than isolated assessments [26, 27], so we adopt a pairwise scoring approach. Based on the review model " + }, + { + "bbox": [ + 55, + 293, + 296, + 390 + ], + "type": "inline_equation", + "content": "M_r" + }, + { + "bbox": [ + 55, + 293, + 296, + 390 + ], + "type": "text", + "content": "'s evaluation " + }, + { + "bbox": [ + 55, + 293, + 296, + 390 + ], + "type": "inline_equation", + "content": "\\mathrm{Judge}_i^r" + }, + { + "bbox": [ + 55, + 293, + 296, + 390 + ], + "type": "text", + "content": " for candidates " + }, + { + "bbox": [ + 55, + 293, + 296, + 390 + ], + "type": "inline_equation", + "content": "M_j" + }, + { + "bbox": [ + 55, + 293, + 296, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 293, + 296, + 390 + ], + "type": "inline_equation", + "content": "M_k" + }, + { + "bbox": [ + 55, + 293, + 296, + 390 + ], + "type": "text", + "content": ", we calculate the model's score for Response Accuracy " + }, + { + "bbox": [ + 55, + 293, + 296, + 390 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{Correct}}(j)" + }, + { + "bbox": [ + 55, + 293, + 296, + 390 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 397, + 295, + 443 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 397, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 61, + 397, + 295, + 443 + ], + "type": "interline_equation", + "content": "S _ {\\text {C o r r e c t}} (j) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f J u d g e} _ {\\text {C o r r e c t}, j} > \\text {J u d g e} _ {\\text {C o r r e c t}, k} \\\\ 0. 5 & \\text {i f J u d g e} _ {\\text {C o r r e c t}, j} = \\text {J u d g e} _ {\\text {C o r r e c t}, k} \\\\ 0 & \\text {i f J u d g e} _ {\\text {C o r r e c t}, j} < \\text {J u d g e} _ {\\text {C o r r e c t}, k}, \\end{array} \\right. \\tag {6}", + "image_path": "a3855d44f7bb509ff69f2464babb6ebf60d44eace6d2e14a746c4af8f25a9e43.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 449, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 449, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\mathrm{Judge}_{\\mathrm{Correct}}" + }, + { + "bbox": [ + 55, + 449, + 296, + 510 + ], + "type": "text", + "content": " represents the review model's assessment of response correctness. If " + }, + { + "bbox": [ + 55, + 449, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\mathrm{Judge}_{\\mathrm{Correct},j} > \\mathrm{Judge}_{\\mathrm{Correct},k}" + }, + { + "bbox": [ + 55, + 449, + 296, + 510 + ], + "type": "text", + "content": ", it indicates that model " + }, + { + "bbox": [ + 55, + 449, + 296, + 510 + ], + "type": "inline_equation", + "content": "M_{j}" + }, + { + "bbox": [ + 55, + 449, + 296, + 510 + ], + "type": "text", + "content": " is judged superior in response accuracy to " + }, + { + "bbox": [ + 55, + 449, + 296, + 510 + ], + "type": "inline_equation", + "content": "M_{k}" + }, + { + "bbox": [ + 55, + 449, + 296, + 510 + ], + "type": "text", + "content": ", scoring 1 point; a tie yields 0.5 points, and a lower score results in 0 points." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 518, + 295, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 295, + 577 + ], + "type": "text", + "content": "Visual Understanding and Reasoning. Inspired by the significant CoT ability demonstrated in the o1 model [34] through adversarial training, we leverage " + }, + { + "bbox": [ + 55, + 518, + 295, + 577 + ], + "type": "inline_equation", + "content": "M_r" + }, + { + "bbox": [ + 55, + 518, + 295, + 577 + ], + "type": "text", + "content": " to evaluate the interpretative and analytical capabilities of the model with respect to visual content." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 578, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 295, + 615 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{Visual}}(j)" + }, + { + "bbox": [ + 55, + 578, + 295, + 615 + ], + "type": "text", + "content": " is computed using the function " + }, + { + "bbox": [ + 55, + 578, + 295, + 615 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 55, + 578, + 295, + 615 + ], + "type": "text", + "content": ", which evaluates the response across four dimensions: caption, reasoning, grounding, and relationship, as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 83, + 632, + 295, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 632, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 83, + 632, + 295, + 646 + ], + "type": "interline_equation", + "content": "S _ {\\text {V i s u a l}} (j) = \\Gamma (\\text {c a p . , r e a . , g r o . , o b j .} \\mid A _ {j}, M _ {r}), \\tag {7}", + "image_path": "6c652d342eef71d590d10b8c2caab6d2800c62e2bb844fe5741c5aa056c25be9.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "where captioning evaluates the ability to generate precise descriptions, reasoning measures logical consistency, grounding assesses accurate object localization, and relationship captures the connections between subjects. Within this structure, the review model " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "M_r" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " provides a combined" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "score via " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": ", yielding " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{Visual}}(j)" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{Visual}}(j)" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " reflects a model's capability to go beyond basic VQA, integrating both description and reasoning to demonstrate a deeper level of visual comprehension." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 130, + 554, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 130, + 554, + 166 + ], + "spans": [ + { + "bbox": [ + 313, + 130, + 554, + 166 + ], + "type": "text", + "content": "Image-Text Correlation. To capture the alignment between image content and textual responses, we introduce the CLIP [37] score:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 379, + 175, + 553, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 175, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 379, + 175, + 553, + 190 + ], + "type": "interline_equation", + "content": "S _ {\\text {C l i p}} (j) = \\operatorname {C L I P} \\left(I _ {i}, A _ {i} ^ {j, r}\\right), \\tag {8}", + "image_path": "87bd17c806888442e4f48af14597b0cfcefd412c2d66b2ed1939a75be1d86487.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 198, + 554, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 554, + 234 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 554, + 234 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 198, + 554, + 234 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{Clip}}(j)" + }, + { + "bbox": [ + 313, + 198, + 554, + 234 + ], + "type": "text", + "content": " measures the correlation between image " + }, + { + "bbox": [ + 313, + 198, + 554, + 234 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 313, + 198, + 554, + 234 + ], + "type": "text", + "content": " and response " + }, + { + "bbox": [ + 313, + 198, + 554, + 234 + ], + "type": "inline_equation", + "content": "A_{i}^{j,r}" + }, + { + "bbox": [ + 313, + 198, + 554, + 234 + ], + "type": "text", + "content": " from model " + }, + { + "bbox": [ + 313, + 198, + 554, + 234 + ], + "type": "inline_equation", + "content": "M_{j}" + }, + { + "bbox": [ + 313, + 198, + 554, + 234 + ], + "type": "text", + "content": ". The details are described in subsection 9.2." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 236, + 554, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 554, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 554, + 283 + ], + "type": "text", + "content": "Combining the above components, we compute the final vision-language judgment score as a weighted sum of the three criteria: Response Accuracy, Visual Understanding and Reasoning, and Image-Text Correlation:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 354, + 293, + 553, + 307 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 293, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 354, + 293, + 553, + 307 + ], + "type": "interline_equation", + "content": "S _ {\\mathrm {V L}} = \\gamma_ {1} S _ {\\text {C o r r e c t}} + \\gamma_ {2} S _ {\\text {V i s u a l}} + \\gamma_ {3} S _ {\\text {C l i p}}, \\tag {9}", + "image_path": "543c6d51176fdd80700b1914243d27862c115acdcf75e805e776065e48cfb0a7.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 316, + 554, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 554, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 554, + 339 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 316, + 554, + 339 + ], + "type": "inline_equation", + "content": "\\gamma_{1},\\gamma_{2}" + }, + { + "bbox": [ + 313, + 316, + 554, + 339 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 313, + 316, + 554, + 339 + ], + "type": "inline_equation", + "content": "\\gamma_{3}" + }, + { + "bbox": [ + 313, + 316, + 554, + 339 + ], + "type": "text", + "content": " are weights reflecting the relative importance of each component in the overall evaluation." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 346, + 480, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 480, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 480, + 360 + ], + "type": "text", + "content": "3.3. Dynamic Weight Optimization" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 364, + 554, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 554, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 554, + 400 + ], + "type": "text", + "content": "In each iteration, we use Mean Squared Error (MSE) as the loss function to update the weights " + }, + { + "bbox": [ + 313, + 364, + 554, + 400 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 364, + 554, + 400 + ], + "type": "text", + "content": ", enhancing the consistency between the weights " + }, + { + "bbox": [ + 313, + 364, + 554, + 400 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 364, + 554, + 400 + ], + "type": "text", + "content": " and the estimated scores " + }, + { + "bbox": [ + 313, + 364, + 554, + 400 + ], + "type": "inline_equation", + "content": "\\hat{G}" + }, + { + "bbox": [ + 313, + 364, + 554, + 400 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 351, + 409, + 553, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 409, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 351, + 409, + 553, + 441 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M S E}} (\\hat {G}, w) = \\frac {1}{m} \\sum_ {j = 1} ^ {m} \\left(\\hat {G} _ {M _ {j}} - w _ {M _ {j}}\\right) ^ {2}. \\tag {10}", + "image_path": "d1cddedfca58a02ff021715492e21e8db0378dad57e779691f55729df93b1d53.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 450, + 554, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 554, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 554, + 522 + ], + "type": "text", + "content": "The entire optimization process is dynamic. After each round of weight updates, the weights and scores continue to change. The loss function computes new scores based on the updated weights, which are then iteratively used to optimize the weights further, forming a continuous iterative optimization loop, as detailed in Algorithm 1." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 533, + 394, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 394, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 394, + 546 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 552, + 430, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 430, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 430, + 565 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "text", + "content": "Datasets. To validate the effectiveness of our framework, we conduct experiments on two distinctly styled datasets: ScienceQA [28] and MMStar [10]. ScienceQA underscores the importance of CoT in scientific reasoning tasks, comprising approximately 21,208 multimodal multiple-choice questions covering various science topics. MMStar focuses on evaluating the multimodal capabilities of MLLMs, and consists of 1,500 meticulously curated samples including 6 core capabilities and 18 detailed axes. Each sample exhibits visual dependency and minimal data leakage." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Candidate MLLMs. Our MLLM pool includes 5 closed-source models and 1 open-source model. The open-source" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9168" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 97, + 70, + 520, + 251 + ], + "blocks": [ + { + "bbox": [ + 97, + 70, + 520, + 251 + ], + "lines": [ + { + "bbox": [ + 97, + 70, + 520, + 251 + ], + "spans": [ + { + "bbox": [ + 97, + 70, + 520, + 251 + ], + "type": "table", + "html": "
ModelsMMstarScienceQA
Pearson (↑)Spearman (↑)Per.Ent. (↓)Pearson (↑)Spearman (↑)Per.Ent. (↓)
LLama-3.2-11b-V0.314±0.07570.550±0.05770.983±0.23100.160±0.04300.225±0.09571.099±0.0000
Claude-3-haiku0.095±0.13010.225±0.15001.099±0.0000-0.145±0.0308-0.525±0.20611.099±0.0000
Claude-3.5-sonnet0.780±0.01460.825±0.05000.752±0.23100.437±0.04520.450±0.17320.752±0.2310
Gemini-1.5-pro0.864±0.00680.850±0.05770.637±0.00000.4147±0.04640.725±0.20620.434±0.5352
GPT-4o-mini0.668±0.00730.725±0.12580.868±0.28860.3354±0.06350.600±0.00001.099±0.0000
GPT-4o0.878±0.00380.875±0.00500.159±0.31830.617±0.00710.625±0.12580.637±0.0000
Methods
Peer Review0.725±0.00440.771±0.16161.040±0.28300.463±0.01930.686±0.17771.040±0.2830
Majority Vote [42]0.757±0.00130.757±0.08571.299±0.17330.509±0.01810.524±0.06601.040±0.0000
Rating Vote [2]0.795±0.00130.743±0.23090.628±0.07550.623±0.00840.629±0.18950.920±0.2387
PRD [22]0.838±0.00270.864±0.03170.427±0.00870.636±0.00420.694±0.07340.746±0.0016
UPME0.944±0.00110.972±0.03300.141±0.28120.814±0.00240.886±0.02860.422±0.2812
", + "image_path": "10a7ce8fa8ec0616e72611fe4ac31b8c446bc99ba8f6c1b1625a96601a76a5b9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 259, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 259, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 555, + 293 + ], + "type": "text", + "content": "Table 1. Main results of different models and methods on MMstar and ScienceQA datasets, evaluated by Pearson, Spearman, and Permutation Entropy metrics. Since the model cannot judge itself, scores from the other five models involved in the computations are used for evaluation. Optimal results are highlighted in bold and underlined. Standard deviation (std) is calculated based on four runs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 56, + 314, + 295, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 314, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 314, + 295, + 410 + ], + "type": "text", + "content": "model is Llama-3.2-11b-vision-instruct [29], which adapts the Llama-3 architecture to the multimodal domain through visual instruction tuning [24]. The closed-source models consist of GPT-4o [35], GPT-4o-mini [33], Claude-3.5-Sonnet [3], Claude-3-Haiku [4], and Gemini-1.5-Pro [45]. These commercial models employ enhanced human alignment strategies to optimize performance and user experience across various applications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 410, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 295, + 517 + ], + "type": "text", + "content": "Metrics. For all experiments, we employ three popular metrics to evaluate the setups mentioned above and our UPME method: the Pearson Correlation Coefficient [14] to evaluate the similarity of score lists, Spearman's Rank Correlation Coefficient [54] to measure the correlation of ranking lists, and Permutation Entropy [5] to assess the complexity and unpredictability of the scoring distributions. All experiments are performed for 4 runs with different seeds (seed = 1, 2, 3, 4) and record the average results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 601 + ], + "type": "text", + "content": "Comparison Methods. Since we are the first to investigate unsupervised peer review in MLLM evaluations, we choose the original peer review mechanism and two popular classical methods, Majority Vote [42] and Rating Vote [2], for our comparative analysis. We also adopt a semi-supervised method, PRD [22], and conduct experiments using individual models as judges to further bolster our comparisons." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 609, + 179, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 179, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 179, + 621 + ], + "type": "text", + "content": "4.2. Experimental Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 627, + 170, + 638 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 627, + 170, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 627, + 170, + 638 + ], + "type": "text", + "content": "4.2.1. Comparison Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "type": "text", + "content": "The experimental results in Table 1 indicate that among the candidate MLLMs, GPT-4o exhibits the highest review capability on the MMstar dataset, achieving Pearson and Spearman correlation coefficients of 0.878 and 0.875, respectively. In contrast, the open-source LLama-3.2 model demonstrates weaker performance. Although Claude-3.5-" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 321, + 316, + 545, + 449 + ], + "blocks": [ + { + "bbox": [ + 321, + 316, + 545, + 449 + ], + "lines": [ + { + "bbox": [ + 321, + 316, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 321, + 316, + 545, + 449 + ], + "type": "image", + "image_path": "5eb990f3a52d71235841e3e6a9dae03857815000f623146822506a110345d56a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 455, + 500, + 467 + ], + "lines": [ + { + "bbox": [ + 367, + 455, + 500, + 467 + ], + "spans": [ + { + "bbox": [ + 367, + 455, + 500, + 467 + ], + "type": "text", + "content": "Figure 3. Convergence experiments." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 485, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 555, + 628 + ], + "type": "text", + "content": "Sonnet and Gemini-1.5-Pro claim to offer performance comparable to GPT-4o, they still show limitations when evaluating other models, falling noticeably short of human-level performance. The original peer review method yields Pearson and Spearman coefficients lower than that of GPT-4o, indicating that the biases of weaker models influence the Peer Review approach. Our UPME method achieves a Pearson correlation of 0.944 and a Spearman correlation of 0.972 on the MMstar dataset, demonstrating a high degree of similarity with human annotations and validating the effectiveness of our optimizations in visual QA comprehension and consistency evaluation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "On the ScienceQA dataset, even the best-performing GPT-4o achieves Pearson and Spearman correlations of 0.617 and 0.625, reflecting the difficulty of the task. The original peer review method shows more modest results, with Pearson and Spearman of 0.463 and 0.686, respectively. Following a series of optimizations and alignments, UPME achieves Pearson and Spearman correlations" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9169" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 295, + 182 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 295, + 182 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 295, + 182 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 295, + 182 + ], + "type": "table", + "html": "
MMstarScienceQA
MethodCor.Vis.Cti.PearsonSpearmanPearsonSpearman
Peer Review0.7270.7140.4570.600
Score Optimization✓0.8540.7710.7130.657
Visual Alignment✓0.8730.8860.7010.771
✓0.7850.8290.5480.771
✓✓0.9030.9430.7750.886
UPME✓✓✓0.9440.9720.8140.886
", + "image_path": "c82fec597a9ac77ac103e17938b624e8c8fe06ee19e6546ac79c5a2615255233.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 191, + 295, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 191, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 295, + 236 + ], + "type": "text", + "content": "Table 2. Ablation study with Pearson and Spearman similarity metrics for each dataset. Optimal results are highlighted in bold and underlined. 'Cor.' stands for correctness, 'Vis.' stands for visual understanding, and 'Cli.' stands for Clip." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 261, + 295, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 261, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 54, + 261, + 295, + 369 + ], + "type": "text", + "content": "of 0.814 and 0.886, showcasing its robustness and superior performance in complex datasets. While the performance on ScienceQA may not be as high, it is important to note that traditional datasets like ScienceQA might not fully align with the evaluation needs of current MLLMs because they may lack visual dependency or sufficient discrimination [10, 52]. UPME achieves higher Pearson and Spearman correlations, demonstrating its robustness and superior performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 374, + 296, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 296, + 446 + ], + "type": "text", + "content": "Moreover, as shown in Figure 3, the loss consistently decreases, and the Pearson similarity metric increases over epochs in 64 different initial settings, which demonstrates the convergence of " + }, + { + "bbox": [ + 55, + 374, + 296, + 446 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 374, + 296, + 446 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 374, + 296, + 446 + ], + "type": "inline_equation", + "content": "\\hat{G}" + }, + { + "bbox": [ + 55, + 374, + 296, + 446 + ], + "type": "text", + "content": ". Under the current experimental setup (25 images and 1500 judgments), the framework reliably converges within the 30-epoch limit." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 466, + 149, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 149, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 149, + 479 + ], + "type": "text", + "content": "4.2.2. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 486, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 486, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 486, + 295, + 712 + ], + "type": "text", + "content": "In the ablation study presented in Table 2, we evaluated the impact of progressively introducing different components on model performance. The Score Optimization phase showed consistent improvements in Pearson and Spearman correlation coefficients as modules were added. For instance, using only the correctness module resulted in Pearson scores of 0.854 for MMstar and 0.713 for ScienceQA. Introducing the visual understanding module alone yielded scores of 0.873 and 0.701, respectively. Adding the Clip component alone gave scores of 0.785 for MMstar and 0.548 for ScienceQA. Combining visual understanding and Clip modules further improved performance to 0.903 and 0.775. Ultimately, integrating all components within the UPME framework resulted in the highest performance, with Pearson and Spearman correlations of 0.944 and 0.972 on MMstar and 0.814 and 0.886 on ScienceQA. These results confirm that optimizing the scoring mechanism and visual alignment significantly boosts evaluation accuracy and robustness across different datasets." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 553, + 159 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 553, + 159 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 159 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 159 + ], + "type": "image", + "image_path": "73e9364cdc08f25fb677a1f3d656d596947032938587d4d51501487e4168d6d4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 160, + 553, + 248 + ], + "blocks": [ + { + "bbox": [ + 317, + 160, + 553, + 248 + ], + "lines": [ + { + "bbox": [ + 317, + 160, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 317, + 160, + 553, + 248 + ], + "type": "image", + "image_path": "cc6be181d7516aeafea998fd44d2af27f74ac0573fa1603eccdb78fe39fb8c04.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 258, + 545, + 269 + ], + "lines": [ + { + "bbox": [ + 321, + 258, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 321, + 258, + 545, + 269 + ], + "type": "text", + "content": "Figure 4. The performance of UPME in different sample size." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 285, + 438, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 438, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 438, + 297 + ], + "type": "text", + "content": "4.2.3.Impact of Sample Size" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 300, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 491 + ], + "type": "text", + "content": "In Figure 4, we analyze the impact of sample size on the alignment with human-annotated QA benchmarks, measured through Pearson and Spearman correlation coefficients, as well as Permutation Entropy. As the sample size increases from 5 to 25, there is a clear improvement in both correlation metrics and a decline in Permutation Entropy, indicating that the generated score lists and rank lists show greater alignment with human evaluations. Beyond a sample size of 25, the metrics stabilize, showing minimal changes with further increases from 25 to 100 in sample size. This stabilization suggests that the consistency of the generated outputs relative to human judgments has reached a satisfactory level, making 25 samples a reliable point for evaluation. Therefore, we primarily selected a sample size of 25 for our experiments, as it provides an effective balance between alignment accuracy and data efficiency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 502, + 529, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 529, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 529, + 515 + ], + "type": "text", + "content": "5. Human Preference Alignment Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 521, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 617 + ], + "type": "text", + "content": "The aforementioned experiments primarily reflect the consistency between the UPME framework and the predefined QA-based benchmarks. To further explore the alignment between UPME and human preferences, as well as to examine the two types of biases inherent in MLLM-as-Judge approaches, we selected a subset of responses for human evaluation and analysis of UPME scoring results. The detailed experimental design is provided in subsection 9.4." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 624, + 501, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 501, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 501, + 636 + ], + "type": "text", + "content": "5.1. Alignment with Human Preference" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "text", + "content": "To validate whether our proposed UPME method can align with human preferences without manual labeling, we conducted experiments on the MMstar and ScienceQA datasets. The experimental results on Table 3 indicate that the baseline method exhibited relatively low human agreement and model consistency rates, suggesting that the Peer" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9170" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 294, + 140 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 140 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 140 + ], + "type": "table", + "html": "
DatasetMethodAgreement (%)Consistency (%)
MMstarPeer Review71.167.5
UPME95.989.8
ScienceQAPeer Review68.261.8
UPME87.482.6
", + "image_path": "6b59d2425fda7abd3e5b29e4c063a4c28c08ed5a7168ac3decf25a6e70965bd3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 180, + 293, + 355 + ], + "blocks": [ + { + "bbox": [ + 59, + 180, + 293, + 355 + ], + "lines": [ + { + "bbox": [ + 59, + 180, + 293, + 355 + ], + "spans": [ + { + "bbox": [ + 59, + 180, + 293, + 355 + ], + "type": "image", + "image_path": "8b207c8b62b9f05df7ca8f3bb4836617aa410c51583b90498226f094c3398ce0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 365, + 296, + 420 + ], + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 420 + ], + "type": "text", + "content": "Figure 5. Model accuracy comparison in peer review framework w/ and w/o UPME, where Peer Review_Cor. represents the correctness of the original peer review, and UPME_Cor. and UPME_Vis. correspond to the two judgment dimensions of response correctness and visual understanding, introduced in subsection 3.2." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 240, + 555, + 263 + ], + "lines": [ + { + "bbox": [ + 313, + 240, + 555, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 555, + 263 + ], + "type": "text", + "content": "Figure 6. Heatmap and Table for Peer Review and UPME on Self Preference and Verbosity Bias." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 443, + 295, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 443, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 54, + 443, + 295, + 612 + ], + "type": "text", + "content": "Review mechanism under an unsupervised setting without weight optimization struggles to align with human preferences. In contrast, UPME demonstrated significant improvements by incorporating Correctness, Visual Understanding, and Clip Correlation. On the MMStar, UPME achieved an agreement rate of " + }, + { + "bbox": [ + 54, + 443, + 295, + 612 + ], + "type": "inline_equation", + "content": "95.9\\%" + }, + { + "bbox": [ + 54, + 443, + 295, + 612 + ], + "type": "text", + "content": " and a consistency rate of " + }, + { + "bbox": [ + 54, + 443, + 295, + 612 + ], + "type": "inline_equation", + "content": "89.8\\%" + }, + { + "bbox": [ + 54, + 443, + 295, + 612 + ], + "type": "text", + "content": ", showing that the optimized scoring criteria significantly enhance the accuracy of evaluation outputs and alignment with human preferences. By capturing key multimodal understanding metrics without relying on manual labeling, UPME effectively achieved high consistency with human annotations, highlighting its substantial advantages in improving response consistency and accuracy under an unsupervised framework." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 620, + 228, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 228, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 228, + 634 + ], + "type": "text", + "content": "5.2. Performance w/ and w/o UPME" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 639, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 296, + 715 + ], + "type": "text", + "content": "To assess the improvement in model judgment accuracy within the Peer Review mechanism and the UPME framework, we conducted an experiment, the results are displayed in Figure 5. The calculation of accuracy is based on " + }, + { + "bbox": [ + 55, + 639, + 296, + 715 + ], + "type": "inline_equation", + "content": "ACC = \\frac{\\mathrm{judge}_{\\mathrm{correct}}}{\\mathrm{judge}_{\\mathrm{all}}}" + }, + { + "bbox": [ + 55, + 639, + 296, + 715 + ], + "type": "text", + "content": ". Experimental results show that the UPME significantly improves the accuracy of models in" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 555, + 130 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 555, + 130 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 555, + 130 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 555, + 130 + ], + "type": "image", + "image_path": "cec5f000ca8beaf39c46270dc76f11a7523700688f4ef471fce2898193d490e1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 341, + 139, + 408, + 150 + ], + "lines": [ + { + "bbox": [ + 341, + 139, + 408, + 150 + ], + "spans": [ + { + "bbox": [ + 341, + 139, + 408, + 150 + ], + "type": "text", + "content": "(a) Verbosity Bias" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 460, + 139, + 529, + 150 + ], + "lines": [ + { + "bbox": [ + 460, + 139, + 529, + 150 + ], + "spans": [ + { + "bbox": [ + 460, + 139, + 529, + 150 + ], + "type": "text", + "content": "(b) Self Preference" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 159, + 554, + 232 + ], + "blocks": [ + { + "bbox": [ + 55, + 148, + 295, + 169 + ], + "lines": [ + { + "bbox": [ + 55, + 148, + 295, + 169 + ], + "spans": [ + { + "bbox": [ + 55, + 148, + 295, + 169 + ], + "type": "text", + "content": "Table 3. Human Agreement (%) and Human Consistency (%) Rates on Various Datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 159, + 554, + 232 + ], + "lines": [ + { + "bbox": [ + 317, + 159, + 554, + 232 + ], + "spans": [ + { + "bbox": [ + 317, + 159, + 554, + 232 + ], + "type": "table", + "html": "
Bias TypeMethodChi-Squarep-valuePhi Coefficient
VerbosityPeer review10.9960.000910.135
BiasUPME0.2800.596960.022
SelfPeer review39.5843.142e-100.257
PreferenceUPME3.4890.06180.076
", + "image_path": "2f9c09aa15170d0f82a7aa5f1fcd3bf3d5542b92b3f702825fd67630781fad62.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 284, + 555, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 555, + 440 + ], + "type": "text", + "content": "both correctness and visual understanding compared to the original correctness judgment. Specifically, the average accuracy in correctness evaluation increased from " + }, + { + "bbox": [ + 313, + 284, + 555, + 440 + ], + "type": "inline_equation", + "content": "61.56\\%" + }, + { + "bbox": [ + 313, + 284, + 555, + 440 + ], + "type": "text", + "content": " in the baseline to " + }, + { + "bbox": [ + 313, + 284, + 555, + 440 + ], + "type": "inline_equation", + "content": "74.48\\%" + }, + { + "bbox": [ + 313, + 284, + 555, + 440 + ], + "type": "text", + "content": " with UPME, an improvement of " + }, + { + "bbox": [ + 313, + 284, + 555, + 440 + ], + "type": "inline_equation", + "content": "12.92\\%" + }, + { + "bbox": [ + 313, + 284, + 555, + 440 + ], + "type": "text", + "content": ". This indicates a significant advantage of UPME in assessing the correctness of model responses. Furthermore, in the evaluation of visual understanding, UPME achieved an average accuracy of " + }, + { + "bbox": [ + 313, + 284, + 555, + 440 + ], + "type": "inline_equation", + "content": "73.93\\%" + }, + { + "bbox": [ + 313, + 284, + 555, + 440 + ], + "type": "text", + "content": ", which is also notably higher than the baseline. This improvement is primarily attributed to UPME's unsupervised evaluation framework, which employs a multi-model peer review mechanism that combines correctness and visual understanding scores to achieve a more efficient and accurate assessment." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 449, + 531, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 531, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 531, + 462 + ], + "type": "text", + "content": "5.3. Mitigating the Bias of MLLM-as-a-Judge" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 467, + 555, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 555, + 658 + ], + "type": "text", + "content": "Mitigation of Verbosity bias. As shown in Figure 6, the verbosity-preference bias is significantly alleviated by our method. The original Peer Review approach had a p-value of 0.00091 (" + }, + { + "bbox": [ + 313, + 467, + 555, + 658 + ], + "type": "inline_equation", + "content": "< 0.05" + }, + { + "bbox": [ + 313, + 467, + 555, + 658 + ], + "type": "text", + "content": "), with a Chi-square value of 10.996 and a Phi coefficient of 0.135, suggesting a significant preference for verbose responses during evaluation. In contrast, the UPME framework yielded a p-value of 0.59696 (" + }, + { + "bbox": [ + 313, + 467, + 555, + 658 + ], + "type": "inline_equation", + "content": ">0.05" + }, + { + "bbox": [ + 313, + 467, + 555, + 658 + ], + "type": "text", + "content": "), with the Chi-square value reduced to 0.280 and the Phi coefficient decreased to 0.022, indicating that it does not exhibit a significant bias toward selecting verbose responses. Additionally, the number of incorrect selections of verbose responses dropped from 233 to 103 in the UPME framework. Thus, the UPME approach effectively mitigates the bias toward verbose responses in models, enhancing evaluation diversity and alignment with human preferences in an unsupervised setting." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Mitigation of Self-preference. Self-preference bias is also significantly alleviated, as shown in Figure 6. In the original Peer Review method, the Chi-square value was 39.584, the Phi coefficient was 0.257, and the p-value was" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9171" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 81, + 178, + 173 + ], + "blocks": [ + { + "bbox": [ + 66, + 81, + 178, + 173 + ], + "lines": [ + { + "bbox": [ + 66, + 81, + 178, + 173 + ], + "spans": [ + { + "bbox": [ + 66, + 81, + 178, + 173 + ], + "type": "image", + "image_path": "54aa3c0f1f355d62b9495a9952c98b432f4ff591c547c77f5f6a0d01c519bd03.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 181, + 79, + 376, + 88 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 79, + 376, + 88 + ], + "spans": [ + { + "bbox": [ + 181, + 79, + 376, + 88 + ], + "type": "text", + "content": "Original Question: Where is the seated person visible in the image?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 448, + 79, + 536, + 88 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 448, + 79, + 536, + 88 + ], + "spans": [ + { + "bbox": [ + 448, + 79, + 536, + 88 + ], + "type": "text", + "content": "Disability of review model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 182, + 89, + 421, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 89, + 421, + 96 + ], + "spans": [ + { + "bbox": [ + 182, + 89, + 421, + 96 + ], + "type": "text", + "content": "Options: A: On the fence, B: In the field, C: Outside the field, D: Behind the bench" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 182, + 97, + 462, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 97, + 462, + 106 + ], + "spans": [ + { + "bbox": [ + 182, + 97, + 462, + 106 + ], + "type": "text", + "content": "Candidate 1 (Gemini-1.5-pro): \"...Therefore, the seated person is visible outside the field. [[C]]\"" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 182, + 106, + 533, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 106, + 533, + 114 + ], + "spans": [ + { + "bbox": [ + 182, + 106, + 533, + 114 + ], + "type": "text", + "content": "Candidate 2 (Claude-3.5-sonnet): \"...The closest answer would be that the player (though not seated) is in the field. [[B]]\"" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 182, + 115, + 547, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 115, + 547, + 132 + ], + "spans": [ + { + "bbox": [ + 182, + 115, + 547, + 132 + ], + "type": "text", + "content": "Review model (Claude-3-haiku) [Correctness judgment] \"...However, Assistant B's response correctly concludes that there is no seated person visible in the image. [[2]]\" (Human judgement: [[1]])" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 182, + 133, + 433, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 133, + 433, + 140 + ], + "spans": [ + { + "bbox": [ + 182, + 133, + 433, + 140 + ], + "type": "text", + "content": "Question generated by review model: What position is the player in the image playing?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 182, + 141, + 388, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 141, + 388, + 149 + ], + "spans": [ + { + "bbox": [ + 182, + 141, + 388, + 149 + ], + "type": "text", + "content": "Candidate 1 (Gemini-1.5-pro): \"... The player is on a baseball field...\"" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 182, + 150, + 514, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 150, + 514, + 158 + ], + "spans": [ + { + "bbox": [ + 182, + 150, + 514, + 158 + ], + "type": "text", + "content": "Candidate 2 (Claude-3.5-sonnet) : \"...Overall theme: The image shows a baseball player in action during a game...\"" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 182, + 159, + 547, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 159, + 547, + 176 + ], + "spans": [ + { + "bbox": [ + 182, + 159, + 547, + 176 + ], + "type": "text", + "content": "Review model (Claude -3-haiku): \"...However, Assistant A provided thorough and insightful analyses of the image is slightly more detailed and precise in its visual assessment. [[1]]\" (Human judgement: [[1]])" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 65, + 198, + 178, + 290 + ], + "blocks": [ + { + "bbox": [ + 65, + 198, + 178, + 290 + ], + "lines": [ + { + "bbox": [ + 65, + 198, + 178, + 290 + ], + "spans": [ + { + "bbox": [ + 65, + 198, + 178, + 290 + ], + "type": "image", + "image_path": "26aa8f5791392d98337dd35d390c6abe17b8a07e501692dce7cfa95724466b3d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 182, + 198, + 350, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 198, + 350, + 206 + ], + "spans": [ + { + "bbox": [ + 182, + 198, + 350, + 206 + ], + "type": "text", + "content": "Original Question: Which is right? They are () each other." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 487, + 198, + 535, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 487, + 198, + 535, + 206 + ], + "spans": [ + { + "bbox": [ + 487, + 198, + 535, + 206 + ], + "type": "text", + "content": "\\section*{Verbosity bias}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 182, + 207, + 343, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 207, + 343, + 214 + ], + "spans": [ + { + "bbox": [ + 182, + 207, + 343, + 214 + ], + "type": "text", + "content": "Options: A: next to, B: far from, C: facing, D: backing\"," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 182, + 215, + 500, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 215, + 500, + 224 + ], + "spans": [ + { + "bbox": [ + 182, + 215, + 500, + 224 + ], + "type": "text", + "content": "Candidate 1 (Gemini-1.5-pro): \"... Determine the best fit: Two toys are next to each other. [[A]]\" (194 tokens)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 182, + 224, + 507, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 224, + 507, + 232 + ], + "spans": [ + { + "bbox": [ + 182, + 224, + 507, + 232 + ], + "type": "text", + "content": "Candidate 2 (Claude-3.5-sonnet): \"... Based on this analysis, two toys are next to each other. [[A]]\" (396 tokens)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 182, + 232, + 547, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 232, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 182, + 232, + 547, + 251 + ], + "type": "text", + "content": "Review model (GPT-40 mini) [Visual judgment]: \"... Overall, Assistant B demonstrates a deeper level of visual understanding by providing a richer description of the scene [[2]]\" (Human judgement: [[1]])" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 182, + 251, + 446, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 251, + 446, + 258 + ], + "spans": [ + { + "bbox": [ + 182, + 251, + 446, + 258 + ], + "type": "text", + "content": "Question generated by review model: What are the two teddy bears looking at in the image?" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 182, + 259, + 487, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 259, + 487, + 266 + ], + "spans": [ + { + "bbox": [ + 182, + 259, + 487, + 266 + ], + "type": "text", + "content": "Candidate 1 (Gemini-1.5-pro) : \"...The boats on the water and the mountains in the distance.\" (235 tokens)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 182, + 267, + 506, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 267, + 506, + 275 + ], + "spans": [ + { + "bbox": [ + 182, + 267, + 506, + 275 + ], + "type": "text", + "content": "Candidate 2 (Claude-3.5-sonnet) : \"...The image shows a baseball player in action during a game...\" (448 tokens)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 182, + 277, + 547, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 277, + 547, + 294 + ], + "spans": [ + { + "bbox": [ + 182, + 277, + 547, + 294 + ], + "type": "text", + "content": "Review model (GPT-40 mini): \"... I find that Assistant A's response demonstrates better visual understanding and reasoning due to its focus on the emotional and scenic aspects of the view. [[1]]\" (Human judgement: [[1]])" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 66, + 317, + 178, + 408 + ], + "blocks": [ + { + "bbox": [ + 66, + 317, + 178, + 408 + ], + "lines": [ + { + "bbox": [ + 66, + 317, + 178, + 408 + ], + "spans": [ + { + "bbox": [ + 66, + 317, + 178, + 408 + ], + "type": "image", + "image_path": "cb517cf23afb8e7b8289ed69f90f839c5a3f7bcb38858dd1e46910b6cb19dd3b.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 427, + 555, + 460 + ], + "lines": [ + { + "bbox": [ + 55, + 427, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 555, + 460 + ], + "type": "text", + "content": "Figure 7. Case study illustration of UPME. We provide the original human-designed questions and the UPME-generated questions, along with the answer analysis. The upper case presents the Disability of review model, where the review model can not answer the original question itself. The middle case demonstrates cases exhibiting morbidity bias. The bottom case shows self-preference bias." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 182, + 316, + 358, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 316, + 358, + 324 + ], + "spans": [ + { + "bbox": [ + 182, + 316, + 358, + 324 + ], + "type": "text", + "content": "Original Question: \" What is the overall mood of this image?" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 484, + 315, + 535, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 484, + 315, + 535, + 324 + ], + "spans": [ + { + "bbox": [ + 484, + 315, + 535, + 324 + ], + "type": "text", + "content": "Self-preference" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 182, + 325, + 506, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 325, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 182, + 325, + 506, + 333 + ], + "type": "text", + "content": "Options: A: Excitement and liveliness, B: Fear and anxiety, C: Calmness and relaxation, D: Confusion and chaos" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 182, + 333, + 508, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 333, + 508, + 341 + ], + "spans": [ + { + "bbox": [ + 182, + 333, + 508, + 341 + ], + "type": "text", + "content": "Candidate 1 (Claude-3.5-sonnet): \"... Considering these factors, the mood that best describes this image is [[C]].\"" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 182, + 342, + 482, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 342, + 482, + 350 + ], + "spans": [ + { + "bbox": [ + 182, + 342, + 482, + 350 + ], + "type": "text", + "content": "Candidate 2 (GPT-4o): \"... Given the analysis, the most fitting mood is Calmness and relaxation. [[C]]\"" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 182, + 350, + 545, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 350, + 545, + 368 + ], + "spans": [ + { + "bbox": [ + 182, + 350, + 545, + 368 + ], + "type": "text", + "content": "Review model (GPT-40 mini) [Visual judgment]: \"...Assistant B provides a slightly clearer rationale for why the mood aligns with calmness rather than introducing conflicting emotions like melancholy. [[2]]\" (Human judgement: [[1]])" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 182, + 369, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 369, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 182, + 369, + 545, + 376 + ], + "type": "text", + "content": "Question generated by review model: What atmospheric conditions are depicted in the image, and how do they affect visibility?" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 182, + 376, + 511, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 376, + 511, + 384 + ], + "spans": [ + { + "bbox": [ + 182, + 376, + 511, + 384 + ], + "type": "text", + "content": "Candidate 1 (Gemini-1.5-pro): \"... The atmospheric conditions depicted in the image are heavy nighttime rainfall.\"" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 182, + 385, + 493, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 385, + 493, + 393 + ], + "spans": [ + { + "bbox": [ + 182, + 385, + 493, + 393 + ], + "type": "text", + "content": "Candidate 2 (Claude-3.5-sonnet): \"...The most prominent atmospheric condition depicted is falling snow...\"" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 182, + 394, + 547, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 394, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 182, + 394, + 547, + 411 + ], + "type": "text", + "content": "Review model (GPT-40 mini): \"... I find that Assistant A's response demonstrates better visual understanding and reasoning due to its focus on the emotional and scenic aspects of the view. [[1]]\" (Human judgement: [[1]])" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 55, + 482, + 296, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 296, + 554 + ], + "type": "text", + "content": "3.142e-10, indicating a highly significant bias. In contrast, the UPME framework reduced the Chi-square value to 3.489, the Phi coefficient to 0.076, and the p-value to 0.0618, approaching insignificance. The incorrect selections due to self-preference dropped from 226 to 129, further demonstrating our effectiveness in mitigating this bias." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 55, + 563, + 130, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 130, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 130, + 576 + ], + "type": "text", + "content": "5.4. Case Study" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "We selected three cases to demonstrate how the UPME effectively addresses the core issues in traditional MLLM-as-a-judge as shown in Figure 7. In the first case, the review model initially failed to identify the error in Candidate B's response because it was unable to answer the question itself. After applying UPME, the review model was able to generate a question within its capabilities and accurately determine that Candidate A's response showed a deeper understanding of the image, which aligned with human preferences, thus avoiding the misjudgment issue. In the second case, although Candidate B's response was longer, it did not" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 482, + 555, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 555, + 566 + ], + "type": "text", + "content": "provide additional useful information. Traditional evaluations might wrongly consider the longer answer as more informative, but with UPME, the review model, through precise text evaluation, clearly identified Candidate A's concise and clear answer as more valuable, thus avoiding morbidity bias. Finally, in the third case, UPME effectively avoided self-preference bias, enhancing the fairness and objectivity." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 586, + 388, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 388, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 388, + 597 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": "This study proposes an unsupervised peer review-based framework, UPME, effectively reducing the annotation workload in traditional human-designed QA benchmarks. Our vision-language scoring system emphasizes a vision-oriented judgment, addressing issues such as verosity and self-preference biases inherent in MLLM-as-judge evaluation methods. Further experiments validate that our evaluation results achieve higher alignment with human preferences, providing a promising and reliable approach." + } + ] + } + ], + "index": 39 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9172" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 295, + 116 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 295, + 116 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 295, + 116 + ], + "type": "text", + "content": "This work was supported by Alibaba DAMO Academy through the Alibaba Innovative Research Program." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 125, + 115, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 125, + 115, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 125, + 115, + 138 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 145, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 145, + 295, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 145, + 295, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 145, + 295, + 190 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 191, + 296, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 296, + 223 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 296, + 223 + ], + "type": "text", + "content": "[2] Mohammad Allahbakhsh and Aleksandar Ignjatovic. Rating through voting: An iterative method for robust rating. arXiv preprint arXiv:1211.0390, 2012. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 224, + 294, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 224, + 294, + 255 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 294, + 255 + ], + "type": "text", + "content": "[3] Anthropic. Claude 3.5: A sonnet. https://www.anthropic.com/news/claude-3-5-sonnet, 2024.1,5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 258, + 295, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 289 + ], + "type": "text", + "content": "[4] Anthropic. Claude 3 haiku. https://www.anthropic.com/news/claude-3-haiku, 2024.5, 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 291, + 295, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 295, + 323 + ], + "type": "text", + "content": "[5] Christoph Bandt and Bernd Pompe. Permutation entropy: a natural complexity measure for time series. Physical review letters, 88(17):174102, 2002. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 325, + 295, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 325, + 295, + 368 + ], + "spans": [ + { + "bbox": [ + 62, + 325, + 295, + 368 + ], + "type": "text", + "content": "[6] Han Bao, Yue Huang, Yanbo Wang, Jiayi Ye, Xiangqi Wang, Xiuying Chen, Mohamed Elhoseiny, and Xiangliang Zhang. Autobench-v: Can large vision-language models benchmark themselves? arXiv preprint arXiv:2410.21259, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 369, + 295, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 295, + 413 + ], + "type": "text", + "content": "[7] Yuzhe Cai, Shaoguang Mao, Wenshan Wu, Zehua Wang, Yaobo Liang, Tao Ge, Chenfei Wu, Wang You, Ting Song, Yan Xia, et al. Low-code llm: Visual programming over llms. arXiv preprint arXiv:2304.08103, 2, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 415, + 295, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 468 + ], + "type": "text", + "content": "[8] Dongping Chen, Ruoxi Chen, Shilin Zhang, Yinuo Liu, Yaochen Wang, Huichi Zhou, Qihui Zhang, Pan Zhou, Yao Wan, and Lichao Sun. Mllm-as-a-judge: Assessing multimodal llm-as-a-judge with vision-language benchmark. arXiv preprint arXiv:2402.04788, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 470, + 295, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 470, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 470, + 295, + 523 + ], + "type": "text", + "content": "[9] Dongping Chen, Yue Huang, Siyuan Wu, Jingyu Tang, Liuyi Chen, Yilin Bai, Zhigang He, Chenlong Wang, Huichi Zhou, Yiqiang Li, et al. Gui-world: A dataset for gui-oriented multimodal llm-based agents. arXiv preprint arXiv:2406.10819, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 525, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 295, + 568 + ], + "type": "text", + "content": "[10] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 1, 2, 4, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 570, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 570, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 570, + 295, + 624 + ], + "type": "text", + "content": "[11] Zhaorun Chen, Yichao Du, Zichen Wen, Yiyang Zhou, Chenhang Cui, Zhenzhen Weng, Haoqin Tu, Chaoqi Wang, Zhengwei Tong, Leria HUANG, et al. Mj-bench: Is your multimodal reward model really a good judge? In ICML 2024 Workshop on Foundation Models in the Wild, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 625, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 679 + ], + "type": "text", + "content": "[12] Wei-Lin Chiang, Lianmin Zheng, Ying Sheng, Anastasios Nikolas Angelopoulos, Tianle Li, Dacheng Li, Hao Zhang, Banghua Zhu, Michael Jordan, Joseph E. Gonzalez, and Ion Stoica. Chatbot arena: An open platform for evaluating llms by human preference, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[13] Zhumin Chu, Qingyao Ai, Yiteng Tu, Haitao Li, and Yiqun Liu. Pre: A peer review based large language model evaluator. arXiv preprint arXiv:2401.15641, 2024. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[14] Israel Cohen, Yiteng Huang, Jingdong Chen, Jacob Benesty, Jacob Benesty, Jingdong Chen, Yiteng Huang, and Israel Cohen. Pearson correlation coefficient. Noise reduction in speech processing, pages 1-4, 2009. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 119, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 173 + ], + "type": "text", + "content": "[15] Chunyuan Deng, Yilun Zhao, Xiangru Tang, Mark Gerstein, and Arman Cohan. Benchmark probing: Investigating data leakage in large language models. In NeurIPS 2023 Workshop on Backdoors in Deep Learning-The Good, the Bad, and the Ugly, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 228 + ], + "type": "text", + "content": "[16] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 230, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 230, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 553, + 284 + ], + "type": "text", + "content": "[17] Chujie Gao, Qihui Zhang, Dongping Chen, Yue Huang, Siyuan Wu, Zhengyan Fu, Yao Wan, Xiangliang Zhang, and Lichao Sun. The best of both worlds: Toward an honest and helpful large language model. arXiv preprint arXiv:2406.00380, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "type": "text", + "content": "[18] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Yuxiao Dong, et al. Co-gagent: A visual language model for gui agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14281-14290, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 342, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 342, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 317, + 342, + 553, + 396 + ], + "type": "text", + "content": "[19] Zdeněk Kasner and Ondřej Dusěk. Beyond traditional benchmarks: Analyzing behaviors of open llms on data-to-text generation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12045–12072, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 399, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 430 + ], + "type": "text", + "content": "[20] Zdeněk Kasner and Ondřej Dušek. Beyond traditional benchmarks: Analyzing behaviors of open llms on data-to-text generation, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 433, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 464 + ], + "type": "text", + "content": "[21] Jian Li and Weiheng Lu. A survey on benchmarks of multimodal large language models. arXiv preprint arXiv:2408.08632, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 467, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 467, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 467, + 553, + 498 + ], + "type": "text", + "content": "[22] Ruosen Li, Teerth Patel, and Xinya Du. Prd: Peer rank and discussion improve large language model based evaluations. arXiv preprint arXiv:2307.02762, 2023. 5, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 501, + 553, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 501, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 553, + 543 + ], + "type": "text", + "content": "[23] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 545, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 553, + 578 + ], + "type": "text", + "content": "[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 580, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 553, + 624 + ], + "type": "text", + "content": "[25] Xiao Liu, Xuanyu Lei, Shengyuan Wang, Yue Huang, Zhuoer Feng, Bosi Wen, Jiale Cheng, Pei Ke, et al. Align-bench: Benchmarking chinese alignment of large language models. arXiv preprint arXiv:2311.18743, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 625, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 678 + ], + "type": "text", + "content": "[26] Yinhong Liu, Han Zhou, Zhijiang Guo, Ehsan Shareghi, Ivan Vulic, Anna Korhonen, and Nigel Collier. Aligning with human judgement: The role of pairwise preference in large language model evaluators. arXiv preprint arXiv:2403.16950, 2024. 2, 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[27] Adian Liusie, Potsawee Manakul, and Mark JF Gales. Zero-shot nlg evaluation through pairware comparisons with llms. arXiv preprint arXiv:2307.07889, 2023. 4" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9173" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "type": "text", + "content": "[28] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 1, 2, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 296, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 296, + 173 + ], + "type": "text", + "content": "[29] Meta. Llama 3.2 11b-vision-instruct. https://huggingface.co/meta-llama/Llama-3.2-11B-Vision-Instruct, 2024.5, 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 296, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 296, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 296, + 206 + ], + "type": "text", + "content": "[30] Birger Moell. Evaluating large language models with human feedback: Establishing a swedish benchmark. arXiv preprint arXiv:2405.14006, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 207, + 295, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 207, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 295, + 262 + ], + "type": "text", + "content": "[31] Kun-Peng Ning, Shuo Yang, Yu-Yang Liu, Jia-Yu Yao, Zhen-Hui Liu, Yu Wang, Ming Pang, and Li Yuan. Peer-review-in-llms: Automatic evaluation method for llms in open-environment. arXiv preprint arXiv:2402.01830, 2024. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 293, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 293, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 293, + 285 + ], + "type": "text", + "content": "[32] OpenAI. Gpt-4 technical report. https://openai.com/research/gpt-4, 2023.1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 295, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 295, + 330 + ], + "type": "text", + "content": "[33] OpenAI. Gpt-4o mini: Advancing cost-efficient intelligence. https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/, 2024.5,1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 331, + 295, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 295, + 374 + ], + "type": "text", + "content": "[34] OpenAI. Introducing openai o1: Advancing model reasoning with visual and logical insights. https://openai.com/index/introducing-openai-o1-preview/, 2024.4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 376, + 200, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 200, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 200, + 387 + ], + "type": "text", + "content": "[35] OpenAI. Hello gpt-4o, 2024. 5, 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 388, + 295, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 295, + 442 + ], + "type": "text", + "content": "[36] Renjie Pi, Lewei Yao, Jiahui Gao, Jipeng Zhang, and Tong Zhang. Perceptiongpt: Effectively fusing visual perception into lIm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27124-27133, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 445, + 295, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 295, + 509 + ], + "type": "text", + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 510, + 295, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 295, + 553 + ], + "type": "text", + "content": "[38] Vyas Raina, Adian Liusie, and Mark Gales. Is llm-as-a-judge robust? investigating universal adversarial attacks on zero-shot llm assessment. arXiv preprint arXiv:2402.14016, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 555, + 259, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 259, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 259, + 567 + ], + "type": "text", + "content": "[39] Refuel.ai. Llm labeling: Technical report, 2025. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 567, + 295, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 295, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 295, + 621 + ], + "type": "text", + "content": "[40] Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaqing Ellen Tan, Yossi Adi, Jingyu Liu, Romain Sauvestre, Tal Remez, et al. Code llama: Open foundation models for code. arXiv preprint arXiv:2308.12950, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 624, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 667 + ], + "type": "text", + "content": "[41] Andreas Stephan, Dawei Zhu, Matthias Aßenmacher, Xiaoyu Shen, and Benjamin Roth. From calculation to adjudication: Examining llm judges on mathematical reasoning tasks. arXiv preprint arXiv:2409.04168, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 669, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 295, + 689 + ], + "type": "text", + "content": "[42] James Surowiecki. The wisdom of crowds/james surowiecki. NY.: Anchor, 2005. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "text", + "content": "[43] Sijun Tan, Siyuan Zhuang, Kyle Montgomery, William Y Tang, Alejandro Cuadron, Chenguang Wang, Raluca Ada" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 711 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "text", + "content": "Popa, and Ion Stoica. Judgebench: A benchmark for evaluating llm-based judges. arXiv preprint arXiv:2410.12784, 2024.3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 107, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 173 + ], + "type": "text", + "content": "[44] Ruixue Tang, Chao Ma, Wei Emma Zhang, Qi Wu, and Xiaokang Yang. Semantic equivalent adversarial data augmentation for visual question answering. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIX 16, pages 437-453. Springer, 2020. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "type": "text", + "content": "[45] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 5, 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 219, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 251 + ], + "type": "text", + "content": "[46] Li Wang, Hui Zhang, and Rajesh Patel. A user-centric benchmark for evaluating large language models. arXiv preprint arXiv:2404.13940, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 252, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 306 + ], + "type": "text", + "content": "[47] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 308, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 363 + ], + "type": "text", + "content": "[48] Siyuan Wu, Yue Huang, Chujie Gao, Dongping Chen, Qihui Zhang, Yao Wan, Tianyi Zhou, Xiangliang Zhang, Jianfeng Gao, Chaowei Xiao, et al. Unigen: A unified framework for textual dataset generation using large language models. arXiv preprint arXiv:2406.18966, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 365, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 385 + ], + "type": "text", + "content": "[49] xAI. Realworldqa: A benchmark for evaluating real-world spatial understanding in multimodal ai models, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 387, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 553, + 420 + ], + "type": "text", + "content": "[50] Ruijie Xu, Zengzhi Wang, Run-Ze Fan, and Pengfei Liu. Benchmarking benchmark leakage in large language models. arXiv preprint arXiv:2404.18824, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 421, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 464 + ], + "type": "text", + "content": "[51] Jiayi Ye, Yanbo Wang, Yue Huang, Dongping Chen, Qihui Zhang, Nuno Moniz, Tian Gao, Werner Geyer, Chao Huang, Pin-Yu Chen, et al. Justice or prejudice? quantifying biases in llm-as-a-judge. arXiv preprint arXiv:2410.02736, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 466, + 553, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 509 + ], + "type": "text", + "content": "[52] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 1, 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 510, + 553, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 576 + ], + "type": "text", + "content": "[53] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 577, + 553, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 577, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 316, + 577, + 553, + 598 + ], + "type": "text", + "content": "[54] Jerrold H Zar. Spearman rank correlation. Encyclopedia of biostatistics, 7, 2005. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 600, + 553, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 600, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 316, + 600, + 553, + 654 + ], + "type": "text", + "content": "[55] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. 1, 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 656, + 553, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 656, + 553, + 711 + ], + "spans": [ + { + "bbox": [ + 316, + 656, + 553, + 711 + ], + "type": "text", + "content": "[56] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, HongFa Wang, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei Li, et al. Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment. arXiv preprint arXiv:2310.01852, 2023. 1" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "9174" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/URWKV_ Unified RWKV Model with Multi-state Perspective for Low-light Image Restoration/e077664a-8838-4206-9246-0d60f57e032f_content_list.json b/2025/URWKV_ Unified RWKV Model with Multi-state Perspective for Low-light Image Restoration/e077664a-8838-4206-9246-0d60f57e032f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7f4d739dccf453a6c77da44fa74e59314e8a7213 --- /dev/null +++ b/2025/URWKV_ Unified RWKV Model with Multi-state Perspective for Low-light Image Restoration/e077664a-8838-4206-9246-0d60f57e032f_content_list.json @@ -0,0 +1,1664 @@ +[ + { + "type": "text", + "text": "URWKV: Unified RWKV Model with Multi-state Perspective for Low-light Image Restoration", + "text_level": 1, + "bbox": [ + 120, + 128, + 875, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rui Xu $^{1,2}$ , Yuzhen Niu $^{1,2*}$ , Yuezhou Li $^{1,2}$ , Huangbiao Xu $^{1,2}$ , Wenxi Liu $^{1,2}$ , Yuzhong Chen $^{1,2}$ $^{1}$ Fujian Key Laboratory of Network Computing and Intelligent Information Processing, College of Computer and Data Science, Fuzhou University, Fuzhou 350108, China \n $^{2}$ Engineering Research Center of Big Data Intelligence, Ministry of Education, Fuzhou 350108, China", + "bbox": [ + 99, + 202, + 908, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{xurui.ryan.chn, yuzhenniu, liyuezhou.cm, huangbiaoxu.chn}@gmail.com", + "bbox": [ + 202, + 276, + 805, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "wenxi.liu@hotmail.com, yzchen@fzu.edu.cn", + "bbox": [ + 323, + 295, + 681, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 344, + 326, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing low-light image enhancement (LLIE) and joint LLIE and deblurring (LLIE-deblur) models have made strides in addressing predefined degradations, yet they are often constrained by dynamically coupled degradations. To address these challenges, we introduce a Unified Receptance Weighted Key Value (URWKV) model with multi-state perspective, enabling flexible and effective degradation restoration for low-light images. Specifically, we customize the core URWKV block to perceive and analyze complex degradations by leveraging multiple intra- and interstage states. First, inspired by the pupil mechanism in the human visual system, we propose Luminance-adaptive Normalization (LAN) that adjusts normalization parameters based on rich inter-stage states, allowing for adaptive, scene-aware luminance modulation. Second, we aggregate multiple intra-stage states through exponential moving average approach, effectively capturing subtle variations while mitigating information loss inherent in the single-state mechanism. To reduce the degradation effects commonly associated with conventional skip connections, we propose the State-aware Selective Fusion (SSF) module, which dynamically aligns and integrates multi-state features across encoder stages, selectively fusing contextual information. In comparison to state-of-the-art models, our URWKV model achieves superior performance on various benchmarks, while requiring significantly fewer parameters and computational resources. Code is available at: https://github.com/FZU-N/URWKV.", + "bbox": [ + 89, + 376, + 483, + 800 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 816, + 220, + 832 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Low-light environments present significant challenges in computer vision and pattern recognition, impacting appli", + "bbox": [ + 89, + 843, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/53dcb58340066208f8c65566b6cc341cac530221b8f321650fb0e6540e93b547.jpg", + "image_caption": [ + "Figure 1. Compared against existing solutions (including Bi-Former [1] for LLIE, PDHAT [2] for joint LLIE and deblurring, and Restormer [3] as a unified model), our proposed URWKV achieves consistently superior PSNR performance across various degradation scenarios. For better visualization, the maximum PSNR values across all datasets are normalized." + ], + "image_footnote": [], + "bbox": [ + 532, + 344, + 883, + 549 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cations such as nighttime surveillance [4], night photography [5], and remote sensing [6]. In these scenarios, insufficient lighting often leads to severe image degradations, such as increased noise, loss of detail, reduced contrast, and color distortion. Furthermore, another degradation coupled in the low-light images, particularly motion blur induced by prolonged exposure times or camera shake, further compromise the integrity of the images. These challenges not only diminish the visual quality of images but also impede subsequent image analysis and interpretation.", + "bbox": [ + 509, + 654, + 906, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Low-light image enhancement (LLIE) models are foundational and widely researched solutions for addressing image degradations in low-light environments. Recent advancements have integrated sophisticated architectures like CNNs [7, 8], Transformers [9, 10], and Mamba networks [11], enabling more effective handling of complex degrad-", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21267", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion patterns. To address specific challenges such as uneven brightness and spatially varying noise, several specialized methods [1, 12] have been developed, enabling finer adjustments across different image regions. Despite these advancements, another typical coupled degradation, i.e., motion blur, that frequently occurs in low-light conditions, remains a significant challenge for current LLIE models.", + "bbox": [ + 88, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The primary difficulty for addressing coupled degradations arises from the inherent trade-offs among brightness enhancement, noise reduction, and blur restoration. Traditional joint training strategy or the combination strategy of off-the-shelf LLIE and deblurring models often fail to deliver optimal results [2, 13]. To bridge this gap, Zhou et al. [13] introduce a compact encoder-decoder architecture wherein the encoder focuses on low-light enhancement, while the decoder tackles blur restoration, effectively isolating the two tasks. Additionally, Li et al. [2] introduce a perceptual decoupling approach that incorporates auxiliary tasks to achieve task-specific representations. While these methods show efficacy in handling blur under low-light conditions, they are constrained by predefined degradation categories. As a result, their adaptability to the dynamically coupled degradations typically encountered in real-world scenarios is limited, where noise, blur, and other degradations may not occur in fixed, predictable combinations.", + "bbox": [ + 91, + 200, + 482, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To flexibly address various degradations, the unified models [3, 14-16] offer a natural attempt, emerging as a leading approach in image restoration. In particular, since IPT [15] first demonstrates the potential of transformer architectures for low-level vision tasks, the development of transformer-based models [3, 17] has accelerated. However, significant challenges remain, particularly in addressing complex illumination and coupled degradation conditions. Many image restoration models, such as those based on transformers, lack adaptive mechanisms designed for low-light environments, often resulting in amplified degradation and the introduction of new artifacts. Specifically, two key issues persist: 1) the difficulty of exploiting effective adaptive strategies for intricate illumination scenarios and handling coupled degradation factors, and 2) parameter inefficiency, which presents a critical bottleneck for practical deployment in downstream vision tasks.", + "bbox": [ + 91, + 474, + 482, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we extend the existing unified encoder-decoder framework by integrating the recently proposed Receptance Weighted Key Value (RWKV) structure [18, 19], aiming to provide effective solutions for addressing dynamically coupled degradations in low-light conditions from a multi-state perspective. Specifically, we introduce both intra- and inter-stage states to perceive and analyze diverse degradations across multiple stages, enabling a more nuanced restoration throughout the processing pipeline. To achieve this, we first propose luminance-adaptive normalization (LAN), which dynamically adjusts normaliza", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion parameters based on rich inter-stage states, facilitating adaptive, scene-aware luminance modulation. Furthermore, we employ an exponential moving average (EMA) approach to aggregate multiple intra-stage states, effectively capturing subtle variations and mitigating the information loss inherent in single-state mechanisms. Instead of constraining our model to predefined degradation types, we exploit the relationships between multi-state representations to address spatially varying and coupled degradations, thus enabling flexible and dynamic restoration. Additionally, we introduce a state-aware selective fusion (SSF) module, which dynamically aligns and integrates multi-state features across encoder stages, selectively fusing contextual information to enhance restoration quality. As shown in Fig. 1, our model consistently outperforms existing models across various degradation scenarios.", + "bbox": [ + 511, + 90, + 903, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this paper are as follows:", + "bbox": [ + 532, + 334, + 872, + 349 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the unified RWKV model with a novel multi-state perspective, enabling flexible and effective restoration of dynamically coupled degradations in low-light conditions.", + "- We propose LAN to dynamically adjust normalization parameters based on inter-stage states, allowing for adaptive, scene-aware luminance modulation.", + "- We aggregate multiple intra-stage states using an EMA-based approach to capture subtle degradation variations, while mitigating the long-range information loss.", + "- We develop the SSF module, which dynamically aligns and integrates multi-state features across encoder stages, facilitating selective fusion of contextual information.", + "- Our URWKV model outperforms existing models across diverse benchmarks, effectively handling dynamically coupled degradations with fewer parameters and lower computational cost." + ], + "bbox": [ + 513, + 349, + 903, + 606 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 625, + 653, + 640 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Low-light Image Enhancement", + "text_level": 1, + "bbox": [ + 511, + 650, + 785, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared to traditional methods relying on histogram equalization [20-22] or Retinex-based enhancement [23-25], deep learning approaches have gained prominence in low-light image enhancement for their ability to handle complex degradations. Retinex-based models, in particular, have evolved through the integration of advanced architectures such as CNNs [7, 8], Transformers [9], and Mamba networks [11], enabling more precise illumination-reflectance decomposition. Leveraging the long-range dependency modeling capabilities of Transformers, several methods [1, 10] have introduced modifications to overcome the limitations of Transformers in low-light scenarios, achieving improved computational efficiency and enhanced local modeling capabilities. In addition, recognizing that noise characteristics vary spatially, Xu et al. [12] propose to", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21268", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c506fea410fc5e2785ecd9faa72ffbf6bfbb8afc59acb2d23acc1ab36f95ff18.jpg", + "image_caption": [ + "(a) Framework" + ], + "image_footnote": [], + "bbox": [ + 109, + 88, + 509, + 358 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/94c74fa9db9e91bb5f9f7701a8d6a8daa02ba5cc1e7c821903a801e8b923ebe1.jpg", + "image_caption": [ + "Squared ReLU" + ], + "image_footnote": [], + "bbox": [ + 532, + 95, + 883, + 303 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/de7a542251406ef978dad269c459a3d5c43318aab7fdc74d3e8982b6aff6cdc6.jpg", + "image_caption": [ + "Concatenation" + ], + "image_footnote": [], + "bbox": [ + 534, + 311, + 555, + 325 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f97b41b07470492ea5ea202a3fba61893fb2f7f3aaac2d440cf96177c32201bd.jpg", + "image_caption": [ + "Sigmoid" + ], + "image_footnote": [], + "bbox": [ + 666, + 311, + 687, + 325 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3147b06902360a624e91930a280d99747ac4f27eac3d7f2681d03f2059671ec2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 781, + 310, + 800, + 325 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a8c7bab8368351d721f368ba6e4855d2e4b6c0180776e3a78acfaced5d9bbec6.jpg", + "image_caption": [ + "Element-wise Multiplication", + "(b) URWKV Block with Multi-state Spatial and Channel Mixing", + "Figure 2. Overview of the URWKV model. The core URWKV block is integrated into an encoder-decoder framework, featuring a multi-state spatial mixing sub-block and a multi-state channel mixing sub-block. The luminance-adaptive normalization (LAN) incorporates both the current input state $X_{t}$ and multiple inter-stage states $M_{i}$ . Additionally, the multi-state quad-directional token shift (SQ-Shift) aggregates both current input state and multiple intra-stage states $H_{i}$ ( $H_{S_{i}}$ for spatial mixing and $H_{C_{i}}$ for channel mixing, respectively). The state-aware selective fusion (SSF) module further aggregates rich contextual information across encoder stages." + ], + "image_footnote": [], + "bbox": [ + 535, + 330, + 553, + 349 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/297f6290ee9be3c5ad0dad1a19dd303e8b948aea7ababdc1c80cbc1fd042180b.jpg", + "image_caption": [ + "Element-wise Addition" + ], + "image_footnote": [], + "bbox": [ + 666, + 330, + 687, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "separate signal from noise during the enhancement process, enabling more effective brightening while minimizing noise amplification. Furthermore, frequency-domain approaches [26, 27] also contribute to LLIE by decomposing features across distinct frequency bands, proving effective in preserving fine details and reducing noise.", + "bbox": [ + 88, + 474, + 482, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Considering blur as a frequent coupled degradation in low-light scenes, one that conventional LLIE models inadequately address, Zhou et al. [13] and Li et al. [2] propose a specific sequential encoder-decoder architecture and a parallel heterogeneous decoupling framework, respectively. Although these LLIE-deblur models show effectiveness within certain constraints, further exploration is warranted to handle the dynamically coupled degradations typical in real-world scenarios.", + "bbox": [ + 88, + 566, + 482, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Unified Model for Image Restoration", + "text_level": 1, + "bbox": [ + 89, + 710, + 410, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the emergence of deep learning, unified approaches have driven substantial progress, enabling the handling of multiple restoration tasks within a single framework. Compared to CNN-based methods [14, 28], which primarily focus on multi-scale feature extraction and local context enhancement, Transformer-based methods [3, 15, 29, 30] extend relational modeling capabilities by capturing long-range dependencies, leading to improved restoration quality across diverse degradation types. However, the application of Transformers in image restoration often entails high computational costs and memory requirements, which can", + "bbox": [ + 88, + 734, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "hinder their efficiency. In addition, frequency-domain models [31, 32] provide another unified approach by decomposing images into multiple frequency bands, which enables focused noise reduction while preserving fine details across tasks. Recent innovations also utilize Mamba networks [16] to achieve effective image restoration by efficiently balancing local and global feature representations.", + "bbox": [ + 511, + 474, + 905, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While current unified models perform effectively under normal lighting conditions, they encounter significant challenges in low-light environments, specifically involving complex lighting variations and coupled degradations.", + "bbox": [ + 511, + 580, + 905, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 656, + 604, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 681, + 624, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, the URWKV framework follows the typical encoder-decoder architecture, with both the encoder and decoder comprising three stages. In the encoder, each stage is composed of $N_{1}$ URWKV blocks followed by a down-sampling layer, while each stage in the decoder contains $N_{2}$ URWKV blocks and an up-sampling layer. Additionally, each down-sampling or up-sampling layer includes convolutional operations and sampling techniques to adjust either the channel or spatial dimensions.", + "bbox": [ + 511, + 703, + 905, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To effectively handle dynamically coupled degradations, each URWKV block operates as an independent and progressive processing stage, with each stage's output (an implicit restored result, serving as a 'state') encapsulat-", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21269", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/da414007927556da9a4b014cfa97d7c676311ccb7d3d5955bbe129d1df082ada.jpg", + "image_caption": [ + "(a) Luminance-adaptive Normalization" + ], + "image_footnote": [], + "bbox": [ + 107, + 88, + 416, + 236 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b6821a700fc76de291b57e101098d4572a8346af8ad168f077a1cab2c0b94a6e.jpg", + "image_caption": [ + "(b) Multi-kernel aggregation of inter-stage states", + "Figure 3. Illustration of luminance-adaptive normalization (LAN). LAN integrates inter-stage states $M_{i}$ throughout the restoration process, facilitating scene-aware luminance modulation." + ], + "image_footnote": [], + "bbox": [ + 436, + 88, + 890, + 231 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing stage-specific contextual semantics and restoration cues (e.g., luminance or content) for subsequent stages.", + "bbox": [ + 89, + 313, + 480, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, the URWKV block builds on the strengths of original RWKV model [19], while enhancing its capabilities through two tailored sub-blocks: a multi-state spatial mixing sub-block and a multi-state channel mixing sub-block. The former captures spatial dependencies across multiple intra-stage states to enrich spatial information, while the latter enhances feature representations through multi-state channel-wise interactions. Here, the intra-stage states $H_{S_i}$ or $H_{C_i}$ are updated following the SQ-Shift in each sub-block, respectively, where $i \\in \\{1, \\dots, N_1\\}$ (or $i \\in \\{1, \\dots, N_2\\}$ ). To ensure efficient information transfer between encoder and decoder stages without introducing degradation, we incorporate the state-aware selective fusion (SSF) module between these stages. The SSF dynamically aggregates multi-state features across all encoder stages, selectively propagating rich contextual information into the decoder.", + "bbox": [ + 88, + 343, + 482, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. URWKV with Multi-state Perspective", + "text_level": 1, + "bbox": [ + 89, + 606, + 416, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Original RWKV models [18, 19] employ a single-state token shift mechanism, where each token's relationship relies solely on the current and preceding state. This design introduces two primary limitations: First, it struggles to retain long-range dependencies, causing early information to gradually fade during restoration. Second, it fails to effectively capture complex interdependencies among degraded features, as a single state is insufficient to represent intricate coupled degradations. To this end, we propose URWKV devised with the multi-state perspective. Specifically, the URWKV block tailors the luminance-adaptive normalization (LAN) by incorporating multiple inter-stage states to adapt URWKV to low-light scenarios. Additionally, it integrates multiple intra-stage states to facilitate more extensive interactions across long-range dependencies.", + "bbox": [ + 89, + 628, + 482, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Luminance-adaptive Normalization (LAN). Luminance distortion is a major factor leading to other coupled degradations in low-light images. Inspired by the hu", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "man eye's ability to adaptively adjust pupil size in response to varying luminance [33, 34], LAN leverages inter-stage states throughout the restoration process (comprising six stages) to enable scene-aware luminance modulation. This design overcomes limitation of LayerNorm widely used in existing models [3, 17], which relies on fixed parameters $\\gamma$ for scaling the learned representation.", + "bbox": [ + 511, + 313, + 903, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Fig. 3, given the current state input $X_{t}\\in \\mathbb{R}^{C_{t}\\times H_{t}\\times W_{t}}$ at stage $T$ , we first extract 1D luminance vectors from both $X_{t}$ and each historical stage output $M_{i}\\in \\mathbb{R}^{C_{i}\\times H_{i}\\times W_{i}}$ , using the global average pooling (GAP) operation. Here, each $M_{i}$ is obtained from the output of each stage's final URWKV block. $t\\in \\{1,\\dots,N_1\\}$ (or $t\\in \\{1,\\dots,N_2\\}$ ) and $i\\in \\{1,\\dots,T - 1\\}$ . The luminance vectors are then zero-padded to match the dimensionality of the vector with the maximum channel size $C_{max}$ , while preserving the semantic integrity of luminance vector from each individual state. These zero-padded luminance vectors, denoted as $X_{k}^{1D}\\in \\mathbb{R}^{C_{max}}$ , for $k\\in \\{1,\\dots,T\\}$ , are then stacked along the state dimension. Next, a multi-kernel aggregation layer is applied to analyze the stacked 2D luminance map along the channel dimension. Specifically, the aggregation layer uses 1D convolutional kernels of sizes $1\\times T$ , $3\\times T$ , and $5\\times T$ to adaptively capture both local and broader contextual illuminance variations across multiple states. The resulting luminance maps $X_{r}^{2D}\\in \\mathbb{R}^{T\\times C_{max}}$ are expressed as:", + "bbox": [ + 511, + 419, + 905, + 720 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX _ {r} ^ {2 D} = \\operatorname {C o n v} _ {r} ^ {1 D} \\left(\\left[ X _ {1} ^ {1 D},..., X _ {T} ^ {1 D} \\right]\\right), r \\in \\{1, 3, 5 \\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 732, + 903, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{Conv}_r^{1D}$ denotes the set of 1D convolutional kernels, and $[\\cdot ,\\cdot ]$ denotes the stacking operation.", + "bbox": [ + 511, + 763, + 903, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subsequently, the output of each convolutional kernel is concatenated along the channel dimension and projected through a $1\\times 1$ convolution, resulting in the aggregated feature map $X_{\\mathrm{agg}}\\in \\mathbb{R}^{C_t}$ . A multi-layer perceptron (MLP) along with a tanh activation function are then employed to predict the luminance modulator $\\Delta \\gamma_{t}$ . The luminance modulator is computed as:", + "bbox": [ + 511, + 795, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21270", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\gamma_ {t} = \\tanh (\\mathrm {M L P} \\left(X _ {\\text {a g g}}\\right)), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 107, + 482, + 125 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tanh (\\cdot)$ represents the tanh activation function.", + "bbox": [ + 89, + 131, + 444, + 146 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once $\\Delta \\gamma_{t}$ is predicted, the scaling parameter $\\gamma$ can be updated as $\\hat{\\gamma}_{t} = \\gamma +\\Delta \\gamma_{t}$ . Ultimately, the input feature $X_{t}$ at stage $T$ can be adaptively normalized with LAN, expressed as:", + "bbox": [ + 89, + 147, + 483, + 205 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX _ {t} ^ {L A N} = \\frac {X _ {t} - \\mu_ {t}}{\\sigma_ {t}} \\odot \\hat {\\gamma} _ {t} + \\beta , \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 219, + 482, + 251 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mu_t$ and $\\sigma_t$ represent the mean and standard deviation of the input feature $X_t$ , respectively. $\\beta$ is a shift parameter.", + "bbox": [ + 89, + 257, + 482, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-state Quad-directional Token Shift (SQ-Shift). Original quad-directional token shift (Q-shift) [19] effectively captures spatial dependencies among neighboring tokens. However, it neglects the degradation relations between states, making it difficult to maintain consistent degradation restoration across consecutive states. To address this limitation, we aggregate multi-state features to extend Q-Shift, allowing the model to capture long-range dependencies across multiple states while preserving local spatial interactions. Specifically, before applying Q-shift, we employ the exponential moving average (EMA) approach to aggregate the current state with all preceding states within the same stage, enabling richer interactions with historical information. For instance, given the state feature $X_{t}^{LAN} \\in \\mathbb{R}^{C \\times H \\times W}$ normalized by LAN, and multiple historical intra-stage state features $H_{i}, i \\in \\{1, \\dots, t - 1\\}$ , we define the multi-state aggregation MSA as follows:", + "bbox": [ + 89, + 287, + 483, + 559 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {M S A} \\left(X _ {t} ^ {L A N}\\right) = \\alpha \\odot X _ {t} ^ {L A N} + (1 - \\alpha) \\odot \\operatorname {M S A} \\left(H _ {t - 1}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 570, + 482, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ is a decay factor that controls the weight between the current state and the previously aggregated state.", + "bbox": [ + 89, + 601, + 482, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-State Spatial and Channel Mixing. For each input feature in the multi-state spatial mixing sub-block or multi-state channel mixing sub-block, we first apply LAN and SQ-Shift, resulting in features $X_{s}$ for spatial mixing or $X_{c}$ for channel mixing. In the spatial mixing process, $X_{s}$ undergoes three parallel linear projections to generate the matrices for receptance $R_{s}$ , key $K_{s}$ , and value $V_{s}$ . Following the computational approach of [19], we employ the bidirectional attention mechanism Bi-WKV to compute the global attention weights $wkv$ between $K_{s}$ and $V_{s}$ , which are subsequently modulated by the sigmoid-activated receptance $R_{s}$ to control the output. The output feature $O_{s}$ is therefore represented as:", + "bbox": [ + 89, + 632, + 482, + 828 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nO _ {s} = \\left(\\sigma \\left(R _ {s}\\right) \\odot w k v\\right) W _ {O _ {s}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 840, + 482, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $wkv = \\mathrm{Bi - WKV}(K_{\\mathrm{s}},V_{\\mathrm{s}})$ . $\\sigma$ represents the sigmoid function. $W_{O_s}$ is learnable parameter of projection layer.", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/17cc601dad29b54cc16d8c3c43b0df24d4f0f5dbd1f57be0414a1594f9bce023.jpg", + "image_caption": [ + "(a) State-aware Selective Fusion" + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 658, + 282 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fe1e1c9639da0d9e46e7a449a121f9202706f24f6c75dd64998fbb794f8e6c98.jpg", + "image_caption": [ + "(c) State Alignment", + "Figure 4. Illustration of the state-aware fusion (SSF) module." + ], + "image_footnote": [], + "bbox": [ + 683, + 87, + 901, + 284 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For channel mixing, $X_{c}$ undergoes two parallel linear projections to derive the receptance matrix $R_{c}$ and key matrix $K_{c}$ . Then, the output feature $O_{c}$ for channel mixing can be formulated as:", + "bbox": [ + 511, + 332, + 905, + 392 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nO _ {c} = \\left(\\sigma \\left(R _ {c}\\right) \\odot \\operatorname {S q u a r e d R e L U} \\left(K _ {c}\\right) W _ {V _ {c}}\\right) W _ {O _ {c}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 400, + 903, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $W_{V_c}$ and $W_{O_c}$ are learnable parameters in the two projection layers.", + "bbox": [ + 511, + 424, + 905, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. State-aware Selective Fusion (SSF)", + "text_level": 1, + "bbox": [ + 511, + 462, + 815, + 478 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Existing models typically reintroduce early-stage features using navie skip connections (e.g., concatenation or addition). However, they are prone to two major issues, particularly under low-light conditions: (1) the propagation of noise and irrelevant information from the encoder to the decoder [1], and (2) the semantic gap between stages [36], which hinders effective feature fusion.", + "bbox": [ + 511, + 484, + 905, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Rather than adopting the navie or parameter-heavy approaches [3, 17, 36], we propose the lightweight and effective state-aware selective fusion (SSF) module. As illustrated in Fig. 4, the SSF module focuses on capturing the degradation restoration relationships between multistate features to predict the spatial guiding weight, which selectively refines the transfer of encoder features to the decoder. Given three encoder outputs $E_{i} \\in \\mathbb{R}^{C_{i} \\times H_{i} \\times W_{i}}$ obtained from $M_{i}$ , $i \\in \\{1,2,3\\}$ , we first apply a channel-wise mean operation to collapse the channel dimension, resulting in $E_{i}^{M} \\in \\mathbb{R}^{1 \\times H_{i} \\times W_{i}}$ , which reduces inter-channel semantic interference. Next, the adaptive alignment operation is performed on $E_{i}^{M}$ to adjust the spatial dimensions of each feature according to the target dimensions of the decoder at the current fusion stage. The aligned features $E_{i}^{A}$ are then stacked along the channel dimension to produce the concatenated feature map $E^{A} \\in \\mathbb{R}^{3 \\times H_{d} \\times W_{d}}$ , where $H_{d}$ and $W_{d}$ are the height and width of the decoder's feature. This process is formally expressed as:", + "bbox": [ + 511, + 590, + 906, + 878 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE ^ {A} = \\operatorname {S t a c k} (\\operatorname {A d a A l i g n} (\\operatorname {A v g} (E _ {i}))), i \\in \\{1, 2, 3 \\}, \\quad (7)\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 883, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21271", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/58c596f6f5916acdc63cca39ca180aa6a80830005f6745fa5da7db32054a291a.jpg", + "table_caption": [ + "Table 1. Quantitative comparison results on the LOL-v2-real, LOL-v2-syn, SID, SMID, SDSD-indoor, and SDSD-outdoor datasets. Joint LLIE and deblurring models are marked with †, unified models are marked with *, and the rest represent LLIE models." + ], + "table_footnote": [], + "table_body": "
MethodsLOL-v2-realLOL-v2-synSIDSMIDSDSD-indoorSDSD-outdoorComplexity
PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑FLOPs (G)Params (M)
KinD [7]14.740.64113.290.57818.020.58322.180.63421.950.67221.970.65434.998.02
EnlightenGAN [35]18.230.61716.570.73417.230.54322.620.67420.020.60420.100.61661.01114.35
SNR-Net [12]21.480.84924.140.92721.350.55028.520.80726.130.81519.220.65726.354.01
FourLLIE [26]22.350.84724.650.91918.420.51325.640.75024.740.82624.670.787-0.12
UHDFour [27]21.790.85423.600.91322.740.65128.180.80929.130.89523.280.77157.4215.90
LLFormer [10]21.630.80824.130.90822.830.65628.420.79426.820.82529.180.84122.5224.52
Retinexformer [9]22.790.83925.670.92824.440.68029.150.81529.780.89529.830.87815.571.61
BiFormer [1]22.670.86324.870.92122.490.63828.240.81228.730.89926.130.8415.930.83
RetinexMamba [11]22.450.84325.880.93322.450.65628.620.80928.440.89428.520.85934.7524.1
LEDNet†[13]19.020.83524.790.93121.470.63828.420.80727.290.87626.660.85038.577.41
PDHAT†[2]20.160.84124.940.93721.930.65329.190.81726.370.88428.190.862208.197.83
Uformer* [17]18.820.77119.660.87118.540.57727.200.79223.170.85923.850.74812.005.29
MIRNet* [14]22.680.82825.050.92321.360.63226.210.76928.640.88828.990.869785.0031.76
Restormer* [3]18.600.78921.410.83122.010.64528.580.80628.490.89227.990.868140.9926.11
MambaIR* [16]20.450.84525.650.93922.020.65828.410.80525.140.87627.530.85160.664.30
URWKV (Ours)*23.110.87426.360.94423.110.67329.440.82631.240.91129.990.88718.342.25
", + "bbox": [ + 94, + 125, + 903, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where Avg denotes the channel-wise mean operation. AdaAlign combines both upsampling and downsampling to adaptively align multiple states with the target state. Stack is the operation that stacks the aligned features along the channel axis.", + "bbox": [ + 89, + 409, + 480, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To facilitate more nuanced fusion of local details and broader context, we incorporate an inception block [37] with multiple convolutional filters $(1\\times 1,3\\times 3,$ and $5\\times 5)$ to aggregate degradation patterns across various receptive fields. The aggregated feature $E_{agg}$ is then projected via a convolutional layer, followed by a sigmoid function to produce the spatial weight. Therefore, taking the first stage of the decoder as an example, the naive concatenation-based skip connection for the URWKV block in the decoder, following the SSF module, can be extended as follows:", + "bbox": [ + 89, + 484, + 482, + 636 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nD _ {1} ^ {\\prime} = \\left(\\left[ W _ {s} \\odot E _ {3}, D _ {1} \\right]\\right) W _ {p} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 646, + 482, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $W_{s}$ is the spatial weight predicted by the SSF module, $E_{3}$ is the encoder feature, and $D_{1}$ is the input feature to the first stage of the decoder. $\\odot$ denotes element-wise multiplication, and $W_{p}$ represents a convolutional layer applied to refine the features. $[\\cdot ,\\cdot ]$ denote the concatenate operation.", + "bbox": [ + 89, + 674, + 483, + 751 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 763, + 223, + 780 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 787, + 297, + 804 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We implement our model in PyTorch on a workstation equipped with an NVIDIA Tesla A40 GPU. The number of URWKV blocks is set to $N_{1} = 3$ and $N_{2} = 2$ , while the number of input channels is set to $C = 32$ . The decay factor $\\alpha$ is set to 0.5. During training, we use a unified loss function for all datasets, as in [1], which combines L1 Loss,", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SSIM Loss, and Perceptual Loss. The Adam optimizer [38] is employed with $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.99$ , and the learning rate is initialized at $2 \\times 10^{-4}$ and decays gradually to $10^{-6}$ using cosine annealing decay. Following [1, 10], we apply augmentation techniques, such as flipping and rotation, to each image pair. During testing, our URWKV model processes input images of arbitrary shape directly, without the need for cropping or resizing. For evaluation, we employ commonly-used PSNR [39] and SSIM [40] metrics to calculate images in the RGB color space across all datasets.", + "bbox": [ + 511, + 409, + 906, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 511, + 573, + 653, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To assess the performance of various models in scenarios involving dynamically coupled degradations, we compare the LLIE models, LLIE-deblur models, and unified models across 8 typical low-light benchmarks.", + "bbox": [ + 511, + 595, + 905, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LOL. As shown in Table 1, our URWKV demonstrates superior performance compared to LLIE models on LOLv2-real [41], outperforming Retinexformer by 0.32 dB in PSNR and 0.035 in SSIM, respectively. Similarly, on LOLv2-syn [41], our model outperforms RetinexMamba by 0.48 dB in PSNR and 0.011 in SSIM. Notably, LLIE-deblur models [2, 13] and other unified models [3, 14, 16] generally underperform on both datasets, with substantial performance gaps between them and our model.", + "bbox": [ + 511, + 657, + 905, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SID and SMID. As shown in Table 1, our UR-WKV model delivers superior performance, closely trailing Retinexformer on SID [42], while outperforming it by 0.29 dB in PSNR on SMID [42]. In contrast, LLIE-deblur models exhibit inconsistent performance across the datasets. For instance, while PDHAT ranks second on SMID, it significantly underperforms on SID, achieving a PSNR of just", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21272", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/886ba7b69e60c419b10d4989571a9d09570d7b3eb7ad0241a0945d26d5297592.jpg", + "image_caption": [ + "Figure 5. Visual comparison of state-of-the-art models on the SID dataset (top) and the SMID dataset (bottom)." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 209, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a05e1f1bef375b59e77999d908dcfa5d1b7a6043de9e6bfe59d419ec048d0090.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 89, + 325, + 251 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e6c0bfd964b3a757c6ddbb480f90954c1a4e138d640cedfd318170312ecea311.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 325, + 90, + 437, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7bd1cc1bfc5006a005c7d1dac10c4671d890d947b8df48b309f01321d7a15b55.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 90, + 553, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6461dd32c502b1a38d651f6570d1f165a9641a4ec5bf171af68475d661589c66.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 553, + 90, + 666, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8810ec3c47ad147d7173a76a86ed5783e9d4f2aaac910fdad4ca4aeaef928755.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 90, + 779, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2727e65419c06056dd0ae2ef9a8f74e705dedc31b8827668579f898be891ae3e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 781, + 90, + 897, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c5ed6aba216577cf3d04fb777f918cf1d7276c9579a6d87c167ed506dae50421.jpg", + "image_caption": [ + "Figure 6. Visual comparisons on SDSD-indoor dataset." + ], + "image_footnote": [], + "bbox": [ + 94, + 287, + 220, + 396 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/96103177d1aac695356b72edb2d217cba9d1641f4e80c6bb8c68ac315a6e3be2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 287, + 346, + 396 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6f7ca234c2a4cd38834d0efdb1c5e692939069f958f3879088cb64c916b2d88c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 346, + 287, + 475, + 396 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "21.93 dB, which is 1.18 dB lower than our model. Visual comparisons in Fig. 5 also demonstrate the excellent denoising capability of our model when enhancing brightness.", + "bbox": [ + 89, + 431, + 482, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SDSD. Results depicted in Table 1 indicate that even some recent LLIE models [1, 26] still encounter challenges on SDSD [43]. In contrast, our URWKV model delivers state-of-the-art performance across both subsets. Moreover, compared to unified models like Restormer, URWKV outperforms it by $2.75\\mathrm{dB}$ in PSNR on SDSD-indoor and by $2.00\\mathrm{dB}$ in PSNR on SDSD-outdoor. As demonstrated in Fig. 6, our method preserves high-fidelity restoration in the scenario with strong light-dark contrast, whereas other models (e.g. PDHAT) generate unexpected artifacts.", + "bbox": [ + 89, + 477, + 482, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "FiveK. The MIT-Adobe FiveK dataset [44] challenges models with the intricate task of color restoration in lowlight and underexposed images. As reported in Table 2, our model achieves the highest PSNR of $26.08\\mathrm{dB}$ on FiveK dataset, surpassing LLFormer by $0.48\\mathrm{dB}$ and Retinexformer by $1.50\\mathrm{dB}$ . Furthermore, URWKV outperforms the unified model, Restormer, in SSIM by 0.007. These notable improvements underscore URWKV's proficiency in restoring both structural integrity and aesthetic quality.", + "bbox": [ + 89, + 628, + 482, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LOL-blur. The LOL-blur [13] dataset couples varying degrees of low-light and motion blur, which has often been underrepresented in other LLIE datasets. As shown in Table 3, conventional LLIE models typically struggle to address this coupled degradation, while unified models also fail to outperform specialized LLIE-deblur models. In contrast, our URWKV model achieves the highest performance in this scenario, with a PSNR of 27.27 dB and SSIM of 0.890, outperforming the second-best LLIE-Deblur model", + "bbox": [ + 89, + 765, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/40e1b4577bd18624f7195585fc0065c676da7f3f4928669b648b3e0369fbe20a.jpg", + "table_caption": [ + "Table 2. Experimental results on MIT-Adobe FiveK dataset." + ], + "table_footnote": [], + "table_body": "
MethodsEnlightenGAN [35]RUAS [45]RetinexNet [8]SNR-Net [12]LLFormer [10]BiFormer [1]
PSNR ↑13.2620.9116.3224.3925.6025.51
SSIM ↑0.7450.8630.7820.8950.9170.929
MethodsRetinexformer [9]LEDNet [13]PDHAT [2]Restormer [3]MIRNet [14]URWKV (Ours)
PSNR ↑24.5824.3425.0025.4823.7326.08
SSIM ↑0.8860.9150.9290.9290.8950.936
", + "bbox": [ + 517, + 305, + 903, + 415 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/250d159b308e01001fe24d2696326f9e50472d7788dd66c76f4eb1c4f9531188.jpg", + "table_caption": [ + "Table 3. Experimental results on LOL-blur dataset." + ], + "table_footnote": [], + "table_body": "
MethodsFourLLIE [26]UHDFour [27]LLFormer [10]Retinexformer [9]BiFormer [1]GLARE [46]
PSNR ↑19.8125.3424.5525.2524.6723.26
SSIM ↑0.6830.8290.7850.8210.8130.690
MethodsLEDNet [13]PDHAT [2]Restormer [3]MIRNet [14]MambaIR [16]URWKV (Ours)
PSNR ↑26.0626.7126.3823.9926.2827.27
SSIM ↑0.8460.8790.8600.7740.8480.890
", + "bbox": [ + 517, + 444, + 903, + 554 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "PDHAT by 0.56 dB in PSNR and 0.011 in SSIM. Examples in Fig. 7 vividly demonstrate the potential of our model in restoring coupled low-light and blur images.", + "bbox": [ + 511, + 566, + 903, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 621, + 666, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct several ablation studies on the real-world dataset (i.e., LOL-v2-real) to analyze the impact of various components in our model.", + "bbox": [ + 511, + 643, + 903, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of LAN and SSF modules. To validate the individual and combined contributions of the LAN and SSF modules, we evaluate four model variants. As shown in Table 4, models incorporating either LAN or SSF individually show a clear improvement over the baseline (first row in Table 4). Among these, the model with LAN alone achieves a more pronounced gain, with PSNR increasing by $1.38\\mathrm{dB}$ over the baseline, highlighting the critical role of LAN in dynamically adapting to complex luminance variations. Moreover, our final model, which combines both LAN and SSF, incurs only a minor increase in parameters (from 1.64M to 2.25M) and a minimal rise in FLOPs (from 18.25G to 18.34G), yet delivers a significant performance boost compared to baseline. The feature transformation vi", + "bbox": [ + 511, + 689, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21273", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/179d7a682ac682a4cdf73291c73ac81b0c0d28c915755566dc436843f8ca3f4e.jpg", + "image_caption": [ + "Figure 7. Visual comparison of state-of-the-art models on LOL-blur dataset." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 898, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5bdae92335af02b211aed70c9727ef64651c3a7eb7dea982b33b920a03db634d.jpg", + "image_caption": [ + "Figure 8. Top: LAN adaptively enhances the luminance, making details clearer. Bottom: SSF helps mitigate introducing degradation, obtaining fused feature $(D_3^{\\prime})$ based on $E_{1}$ , $E_{2}$ and $E_{3}$ ." + ], + "image_footnote": [], + "bbox": [ + 93, + 294, + 478, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sualizations presented in Fig. 8 also validate the inherent advantages of LAN and SSF.", + "bbox": [ + 89, + 444, + 482, + 473 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-state Quad-directional Token Shift. Our propose multi-state aggregation strategy enhances the original Q-shift, allowing for a more comprehensive understanding of degradation representation. As shown in Table 5, our results indicate that single-state aggregation alone (Table 5a) achieves performance comparable to the model using Q-shift alone (Table 5c). In contrast, multi-state aggregation alone (Table 5b) consistently demonstrates superior performance, attaining a $0.13\\mathrm{dB}$ higher PSNR. When combining single-state aggregation with Q-shift (Table 5d), the model shows an improvement in PSNR. However, it still falls short in SSIM by 0.003 compared to multi-state aggregation alone. Ultimately, our final model (Table 5e), which integrates both Q-shift and multi-state aggregation, demonstrates substantial advantages over all other variants.", + "bbox": [ + 88, + 476, + 482, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SSF vs. Navie Skip Connections. The ablation results in Table 6 highlight the effectiveness of SSF compared to naive skip connections, including simple addition (Add) and concatenation (Cat). When using multi-state encoder features, noise from earlier stages and semantic gaps between states lead to performance degradation in both PSNR and SSIM. In contrast, our SSF module, based on state-aware fusion strategy, avoids directly merging multistate features and predicts a guiding weight to selectively fuse useful patterns, mitigating issues of noise propagation and semantic misalignment. Notably, the \"Multi-state+SSF\" (i.e., our URWKV) yields significant improvements in PSNR and SSIM over \"Single-state+Cat\" without", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/34e1f958d1d0a377a167ecb8c8be554719676c2cabe7a22a9a942ef4fc81a1d2.jpg", + "table_caption": [ + "Table 4. Ablation study on LAN and SSF modules." + ], + "table_footnote": [], + "table_body": "
LANSSFPSNR ↑SSIM ↑ParamsFLOPs
21.330.8561.64M18.25G
✓21.400.8611.65M18.29G
✓22.710.8692.25M18.30G
✓✓23.110.8742.25M18.34G
", + "bbox": [ + 532, + 313, + 879, + 386 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7fc0e2d260f5149d74b89e9f5caf7ac9be2999a66b33551298c31b8a88b4f840.jpg", + "table_caption": [ + "Table 5. Ablation study on multi-state quad-directional token shift." + ], + "table_footnote": [], + "table_body": "
#Single-stateMulti-stateQ-shiftPSNR ↑SSIM ↑
(a)✓22.120.861
(b)✓22.250.870
(c)✓22.190.862
(d)✓✓22.290.867
(e)✓✓23.110.874
", + "bbox": [ + 522, + 416, + 901, + 505 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/354e4fbe8b9aaf4dcfdd7aee3a51abf2ffa5145f38e6bfb31dbdd22190a6a1dd.jpg", + "table_caption": [ + "Table 6. Impacts of SSF and navie skip connections (Skip Conn.) with different state numbers (State No.)." + ], + "table_footnote": [], + "table_body": "
State No.Skip Conn.PSNR ↑SSIM ↑ParamsFLOPs
Single-stateAdd22.660.8721.86M14.68G
Single-stateCat22.710.8692.25M18.30G
Multi-stateCat21.830.8592.50M23.89G
Multi-stateSSF23.110.8742.25M18.34G
", + "bbox": [ + 519, + 549, + 901, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "increasing the model's parameter, demonstrating both enhanced performance and computational efficiency.", + "bbox": [ + 511, + 640, + 903, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 678, + 633, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To address the shortcomings of existing models, we propose URWKV with multi-state perspective, enabling flexible and effective restoration across dynamically coupled degradation scenarios. Specifically, we introduce three key innovations: luminance-adaptive normalization (LAN) for adaptive, scene-aware luminance modulation; an exponential moving average approach to aggregate intra-stage states for spatial-level token-shift; and the state-aware selective fusion (SSF) module for dynamic alignment and integration of multi-state features across encoder stages. In comparison to existing models, our model demonstrates superior performance while requiring significantly fewer parameters and computational resources.", + "bbox": [ + 509, + 704, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21274", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported in part by the National Natural Science Foundation of China under Grant 62471142, in part by the Natural Science Foundation of Fujian Province, China under Grants 2023J01067, in part by Industry-Academy Cooperation Project under Grant 2024H6006, in part by the Collaborative Innovation Platform Project of Fuzhou City under Grant 2023-P-002, in part by the Key Technology Innovation Project for Focused Research and Industrialization in the Software Industry of Fujian Province, and in part by the Key Research and Industrialization Project of Technological Innovation in Fujian Province under Grant 2024XQ002.", + "bbox": [ + 89, + 90, + 485, + 273 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 289, + 187, + 304 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rui Xu, Yuezhou Li, Yuzhen Niu, Huangbiao Xu, Yuzhong Chen, and Tiesong Zhao. Bilateral interaction for local-global collaborative perception in low-light image enhancement. IEEE Transactions on Multimedia, 2024. 1, 2, 5, 6, 7", + "[2] Yuezhou Li, Rui Xu, Yuzhen Niu, Wenzhong Guo, and Tiesong Zhao. Perceptual decoupling with heterogeneous auxiliary tasks for joint low-light image enhancement and deblurring. IEEE Transactions on Multimedia, pages 6663-6675, 2024. 1, 2, 3, 6, 7", + "[3] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. 1, 2, 3, 4, 5, 6, 7", + "[4] Andong Lu, Zhang Zhang, Yan Huang, Yifan Zhang, Chenglong Li, Jin Tang, and Liang Wang. Illumination distillation framework for nighttime person re-identification and a new benchmark. IEEE Transactions on Multimedia, pages 1–14, 2023. 1", + "[5] Xiaopeng Li and Xiaojie Guo. SPN2D-GAN: Semantic prior based night-to-day image-to-image translation. IEEE Transactions on Multimedia, pages 1-15, 2022. 1", + "[6] Xuechen Bai, Xinghua Li, Jianhao Miao, and Huanfeng Shen. Making the rarth clear at night: A high-resolution nighttime light image deblooming network. IEEE Transactions on Geoscience and Remote Sensing, 61:1-13, 2023. 1", + "[7] Yonghua Zhang, Jiawan Zhang, and Xiaojie Guo. Kindling the darkness: A practical low-light image enhancer. In Proceedings of the ACM International Conference on Multimedia, pages 1632-1640, 2019. 1, 2, 6", + "[8] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. In Proceedings of the British Machine Vision Conference, 2018. 1, 2, 7", + "[9] Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang. Retinexformer: One-stage retina-based transformer for low-light image enhancement. In Proceedings of the IEEE International Conference on Computer Vision, pages 12504-12513, 2023. 1, 2, 6, 7" + ], + "bbox": [ + 99, + 314, + 480, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Tao Wang, Kaihao Zhang, Tianrun Shen, Wenhan Luo, Bjorn Stenger, and Tong Lu. Ultra-high-definition low-light image enhancement: A benchmark and transformer-based method. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 2654-2662, 2023. 1, 2, 6, 7", + "[11] Jiesong Bai, Yuhao Yin, and Qiyuan He. Retinexmamba: Retinex-based mamba for low-light image enhancement. arXiv preprint arXiv:2405.03349, 2024. 1, 2, 6", + "[12] Xiaogang Xu, Ruixing Wang, Chi-Wing Fu, and Jiaya Jia. SNR-aware low-light image enhancement. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 17714-17724, 2022. 2, 6, 7", + "[13] Shangchen Zhou, Chongyi Li, and Chen Change Loy. LEDNet: Joint low-light enhancement and deblurring in the dark. In Proceedings of the European Conference on Computer Vision, pages 573-589, 2022. 2, 3, 6, 7", + "[14] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Learning enriched features for real image restoration and enhancement. In Proceedings of the European Conference on Computer Vision, pages 492-511, 2020. 2, 3, 6, 7", + "[15] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12299-12310, 2021. 2, 3", + "[16] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. MambaIR: A simple baseline for image restoration with state-space model. In Proceedings of the European Conference on Computer Vision, pages 222-241, 2025. 2, 3, 6, 7", + "[17] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 17683-17693, 2022. 2, 4, 5, 6", + "[18] Bo Peng, Eric Alcaide, Quentin Anthony, Alon Albalak, Samuel Arcadinho, Stella Biderman, Huanqi Cao, Xin Cheng, Michael Chung, Matteo Grella, et al. RWKV: Reinventing rnns for the transformer era. arXiv preprint arXiv:2305.13048, 2023. 2.4", + "[19] Yuchen Duan, Weiyun Wang, Zhe Chen, Xizhou Zhu, Lewei Lu, Tong Lu, Yu Qiao, Hongsheng Li, Jifeng Dai, and Wenhai Wang. Vision-RWKV: Efficient and scalable visual perception with RWKV-like architectures. arXiv preprint arXiv:2403.02308, 2024. 2, 4, 5", + "[20] J Alex Stark. Adaptive image contrast enhancement using generalizations of histogram equalization. IEEE Transactions on Image Processing, 9:889-896, 2000. 2", + "[21] Mohammad Abdullah-Al-Wadud, Md Hasanul Kabir, M Ali Akber Dewan, and Oksam Chae. A dynamic histogram equalization for image contrast enhancement. IEEE Transactions on Consumer Electronics, 53:593-600, 2007.", + "[22] Turgay Celik and Tardi Tjahjadi. Contextual and variational contrast enhancement. IEEE Transactions on Image Processing, 20:3431-3441, 2011. 2", + "[23] Xueyang Fu, Yinghao Liao, Delu Zeng, Yue Huang, Xiao-Ping Zhang, and Xinghao Ding. A probabilistic method for" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21275", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image enhancement with simultaneous illumination and reflectance estimation. IEEE Transactions on Image Processing, 24:4965-4977, 2015. 2", + "[24] Mading Li, Jiaying Liu, Wenhan Yang, Xiaoyan Sun, and Zongming Guo. Structure-revealing low-light image enhancement via robust retinae model. IEEE Transactions on Image Processing, 27:2828-2841, 2018.", + "[25] Daniel J Jobson, Zia-ur Rahman, and Glenn A Woodell. A multiscale retina for bridging the gap between color images and the human observation of scenes. IEEE Transactions on Image processing, 6:965–976, 1997. 2", + "[26] Chenxi Wang, Hongjun Wu, and Zhi Jin. FourLLIE: Boosting low-light image enhancement by fourier frequency information. In Proceedings of the ACM International Conference on Multimedia, pages 7459-7469, 2023. 3, 6, 7", + "[27] Chongyi Li, Chun-Le Guo, Man Zhou, Zhexin Liang, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Embedding fourier for ultra-high-definition low-light image enhancement. Proceedings of the International Conference on Learning Representations, 2023. 3, 6, 7", + "[28] Yuning Cui, Wenqi Ren, Xiaochun Cao, and Alois Knoll. Revitalizing convolutional network for image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 9423-9438, 2024. 3", + "[29] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image restoration using swin transformer. In Proceedings of the IEEE International Conference on Computer Vision, pages 1833-1844, 2021. 3", + "[30] Akshay Dudhane, Syed Waqas Zamir, Salman Khan, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Burstormer: Burst image restoration and enhancement transformer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5703-5712, 2023. 3", + "[31] Man Zhou, Jie Huang, Chun-Le Guo, and Chongyi Li. Fourmer: An efficient global modeling paradigm for image restoration. In Proceedings of the International Conference on Machine Learning, pages 42589-42601, 2023. 3", + "[32] Yuning Cui, Wenqi Ren, Xiaochun Cao, and Alois Knoll. Image restoration via frequency selection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3", + "[33] Francisco Lara, Paula Bernal-Molina, Vicente Fernandez-Sanchez, and Norberto López-Gil. Changes in the objective amplitude of accommodation with pupil size. Optometry and Vision Science, 91(10):1215-1220, 2014. 4", + "[34] Pablo A Barrionuevo, Luis A Issolio, and Constanza Tripolone. Photoreceptor contributions to the human pupil light reflex. Journal of Photochemistry and Photobiology, 15:100178, 2023. 4", + "[35] Yifan Jiang, Xinyu Gong, Ding Liu, Yu Cheng, Chen Fang, Xiaohui Shen, Jianchao Yang, Pan Zhou, and Zhangyang Wang. EnlightenGAN: Deep light enhancement without paired supervision. IEEE Transactions on Image Processing, 30:2340-2349, 2021. 6, 7", + "[36] Haonan Wang, Peng Cao, Jiaqi Wang, and Osmar R Zaiane. UCTransNet: Rethinking the skip connections in U-Net from a channel-wise perspective with transformer. In Proceedings" + ], + "bbox": [ + 91, + 90, + 483, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of the AAAI conference on artificial intelligence, volume 36, pages 2441-2449, 2022. 5", + "[37] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1-9, 2015. 6", + "[38] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[39] Quan Huynh-Thu and Mohammed Ghanbari. Scope of validity of PSNR in image/video quality assessment. *Electronics Letters*, 44:800–801, 2008. 6", + "[40] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 6", + "[41] Wenhan Yang, Wenjing Wang, Haofeng Huang, Shiqi Wang, and Jiaying Liu. Sparse gradient regularized deep retina network for robust low-light image enhancement. IEEE Transactions on Image Processing, 30:2072-2086, 2021. 6", + "[42] Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3291-3300, 2018. 6", + "[43] Ruixing Wang, Xiaogang Xu, Chi-Wing Fu, Jiangbo Lu, Bei Yu, and Jiaya Jia. Seeing dynamic scene in the dark: A high-quality video dataset with mechatronic alignment. In Proceedings of the IEEE International Conference on Computer Vision, pages 9700-9709, 2021. 7", + "[44] Vladimir Bychkovsky, Sylvain Paris, Eric Chan, and Frédo Durand. Learning photographic global tonal adjustment with a database of input/output image pairs. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 97-104, 2011. 7", + "[45] Risheng Liu, Long Ma, Tengyu Ma, Xin Fan, and Zhongxuan Luo. Learning with nested scene modeling and cooperative architecture search for low-light vision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45:5953-5969, 2022. 7", + "[46] Han Zhou, Wei Dong, Xiaohong Liu, Shuaicheng Liu, Xiongkuo Min, Guangtao Zhai, and Jun Chen. GLARE: Low light image enhancement via generative latent feature based codebook retrieval. In Proceedings of the European Conference on Computer Vision, pages 36-54, 2024. 7" + ], + "bbox": [ + 516, + 92, + 903, + 742 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21276", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/URWKV_ Unified RWKV Model with Multi-state Perspective for Low-light Image Restoration/e077664a-8838-4206-9246-0d60f57e032f_model.json b/2025/URWKV_ Unified RWKV Model with Multi-state Perspective for Low-light Image Restoration/e077664a-8838-4206-9246-0d60f57e032f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..79b3aca8f66e1bf214690a298d72eedee5d19e74 --- /dev/null +++ b/2025/URWKV_ Unified RWKV Model with Multi-state Perspective for Low-light Image Restoration/e077664a-8838-4206-9246-0d60f57e032f_model.json @@ -0,0 +1,2343 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.122, + 0.13, + 0.877, + 0.175 + ], + "angle": 0, + "content": "URWKV: Unified RWKV Model with Multi-state Perspective for Low-light Image Restoration" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.203, + 0.909, + 0.275 + ], + "angle": 0, + "content": "Rui Xu\\(^{1,2}\\), Yuzhen Niu\\(^{1,2*}\\), Yuezhou Li\\(^{1,2}\\), Huangbiao Xu\\(^{1,2}\\), Wenxi Liu\\(^{1,2}\\), Yuzhong Chen\\(^{1,2}\\) \n\\(^{1}\\)Fujian Key Laboratory of Network Computing and Intelligent Information Processing, College of Computer and Data Science, Fuzhou University, Fuzhou 350108, China \n\\(^{2}\\)Engineering Research Center of Big Data Intelligence, Ministry of Education, Fuzhou 350108, China" + }, + { + "type": "text", + "bbox": [ + 0.203, + 0.277, + 0.807, + 0.293 + ], + "angle": 0, + "content": "{xurui.ryan.chn, yuzhenniu, liyuezhou.cm, huangbiaoxu.chn}@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.325, + 0.296, + 0.682, + 0.31 + ], + "angle": 0, + "content": "wenxi.liu@hotmail.com, yzchen@fzu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.345, + 0.327, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.377, + 0.485, + 0.801 + ], + "angle": 0, + "content": "Existing low-light image enhancement (LLIE) and joint LLIE and deblurring (LLIE-deblur) models have made strides in addressing predefined degradations, yet they are often constrained by dynamically coupled degradations. To address these challenges, we introduce a Unified Receptance Weighted Key Value (URWKV) model with multi-state perspective, enabling flexible and effective degradation restoration for low-light images. Specifically, we customize the core URWKV block to perceive and analyze complex degradations by leveraging multiple intra- and interstage states. First, inspired by the pupil mechanism in the human visual system, we propose Luminance-adaptive Normalization (LAN) that adjusts normalization parameters based on rich inter-stage states, allowing for adaptive, scene-aware luminance modulation. Second, we aggregate multiple intra-stage states through exponential moving average approach, effectively capturing subtle variations while mitigating information loss inherent in the single-state mechanism. To reduce the degradation effects commonly associated with conventional skip connections, we propose the State-aware Selective Fusion (SSF) module, which dynamically aligns and integrates multi-state features across encoder stages, selectively fusing contextual information. In comparison to state-of-the-art models, our URWKV model achieves superior performance on various benchmarks, while requiring significantly fewer parameters and computational resources. Code is available at: https://github.com/FZU-N/URWKV." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.222, + 0.833 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.844, + 0.484, + 0.875 + ], + "angle": 0, + "content": "Low-light environments present significant challenges in computer vision and pattern recognition, impacting appli" + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.345, + 0.885, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.559, + 0.908, + 0.643 + ], + "angle": 0, + "content": "Figure 1. Compared against existing solutions (including Bi-Former [1] for LLIE, PDHAT [2] for joint LLIE and deblurring, and Restormer [3] as a unified model), our proposed URWKV achieves consistently superior PSNR performance across various degradation scenarios. For better visualization, the maximum PSNR values across all datasets are normalized." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.655, + 0.907, + 0.807 + ], + "angle": 0, + "content": "cations such as nighttime surveillance [4], night photography [5], and remote sensing [6]. In these scenarios, insufficient lighting often leads to severe image degradations, such as increased noise, loss of detail, reduced contrast, and color distortion. Furthermore, another degradation coupled in the low-light images, particularly motion blur induced by prolonged exposure times or camera shake, further compromise the integrity of the images. These challenges not only diminish the visual quality of images but also impede subsequent image analysis and interpretation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Low-light image enhancement (LLIE) models are foundational and widely researched solutions for addressing image degradations in low-light environments. Recent advancements have integrated sophisticated architectures like CNNs [7, 8], Transformers [9, 10], and Mamba networks [11], enabling more effective handling of complex degrad-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21267" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "tion patterns. To address specific challenges such as uneven brightness and spatially varying noise, several specialized methods [1, 12] have been developed, enabling finer adjustments across different image regions. Despite these advancements, another typical coupled degradation, i.e., motion blur, that frequently occurs in low-light conditions, remains a significant challenge for current LLIE models." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.201, + 0.483, + 0.473 + ], + "angle": 0, + "content": "The primary difficulty for addressing coupled degradations arises from the inherent trade-offs among brightness enhancement, noise reduction, and blur restoration. Traditional joint training strategy or the combination strategy of off-the-shelf LLIE and deblurring models often fail to deliver optimal results [2, 13]. To bridge this gap, Zhou et al. [13] introduce a compact encoder-decoder architecture wherein the encoder focuses on low-light enhancement, while the decoder tackles blur restoration, effectively isolating the two tasks. Additionally, Li et al. [2] introduce a perceptual decoupling approach that incorporates auxiliary tasks to achieve task-specific representations. While these methods show efficacy in handling blur under low-light conditions, they are constrained by predefined degradation categories. As a result, their adaptability to the dynamically coupled degradations typically encountered in real-world scenarios is limited, where noise, blur, and other degradations may not occur in fixed, predictable combinations." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.476, + 0.483, + 0.732 + ], + "angle": 0, + "content": "To flexibly address various degradations, the unified models [3, 14-16] offer a natural attempt, emerging as a leading approach in image restoration. In particular, since IPT [15] first demonstrates the potential of transformer architectures for low-level vision tasks, the development of transformer-based models [3, 17] has accelerated. However, significant challenges remain, particularly in addressing complex illumination and coupled degradation conditions. Many image restoration models, such as those based on transformers, lack adaptive mechanisms designed for low-light environments, often resulting in amplified degradation and the introduction of new artifacts. Specifically, two key issues persist: 1) the difficulty of exploiting effective adaptive strategies for intricate illumination scenarios and handling coupled degradation factors, and 2) parameter inefficiency, which presents a critical bottleneck for practical deployment in downstream vision tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this paper, we extend the existing unified encoder-decoder framework by integrating the recently proposed Receptance Weighted Key Value (RWKV) structure [18, 19], aiming to provide effective solutions for addressing dynamically coupled degradations in low-light conditions from a multi-state perspective. Specifically, we introduce both intra- and inter-stage states to perceive and analyze diverse degradations across multiple stages, enabling a more nuanced restoration throughout the processing pipeline. To achieve this, we first propose luminance-adaptive normalization (LAN), which dynamically adjusts normaliza" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.332 + ], + "angle": 0, + "content": "tion parameters based on rich inter-stage states, facilitating adaptive, scene-aware luminance modulation. Furthermore, we employ an exponential moving average (EMA) approach to aggregate multiple intra-stage states, effectively capturing subtle variations and mitigating the information loss inherent in single-state mechanisms. Instead of constraining our model to predefined degradation types, we exploit the relationships between multi-state representations to address spatially varying and coupled degradations, thus enabling flexible and dynamic restoration. Additionally, we introduce a state-aware selective fusion (SSF) module, which dynamically aligns and integrates multi-state features across encoder stages, selectively fusing contextual information to enhance restoration quality. As shown in Fig. 1, our model consistently outperforms existing models across various degradation scenarios." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.335, + 0.874, + 0.35 + ], + "angle": 0, + "content": "The main contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.351, + 0.905, + 0.41 + ], + "angle": 0, + "content": "- We introduce the unified RWKV model with a novel multi-state perspective, enabling flexible and effective restoration of dynamically coupled degradations in low-light conditions." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.412, + 0.904, + 0.455 + ], + "angle": 0, + "content": "- We propose LAN to dynamically adjust normalization parameters based on inter-stage states, allowing for adaptive, scene-aware luminance modulation." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.457, + 0.905, + 0.501 + ], + "angle": 0, + "content": "- We aggregate multiple intra-stage states using an EMA-based approach to capture subtle degradation variations, while mitigating the long-range information loss." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.502, + 0.904, + 0.546 + ], + "angle": 0, + "content": "- We develop the SSF module, which dynamically aligns and integrates multi-state features across encoder stages, facilitating selective fusion of contextual information." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.548, + 0.905, + 0.607 + ], + "angle": 0, + "content": "- Our URWKV model outperforms existing models across diverse benchmarks, effectively handling dynamically coupled degradations with fewer parameters and lower computational cost." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.351, + 0.905, + 0.607 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.625, + 0.655, + 0.641 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.651, + 0.787, + 0.667 + ], + "angle": 0, + "content": "2.1. Low-light Image Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Compared to traditional methods relying on histogram equalization [20-22] or Retinex-based enhancement [23-25], deep learning approaches have gained prominence in low-light image enhancement for their ability to handle complex degradations. Retinex-based models, in particular, have evolved through the integration of advanced architectures such as CNNs [7, 8], Transformers [9], and Mamba networks [11], enabling more precise illumination-reflectance decomposition. Leveraging the long-range dependency modeling capabilities of Transformers, several methods [1, 10] have introduced modifications to overcome the limitations of Transformers in low-light scenarios, achieving improved computational efficiency and enhanced local modeling capabilities. In addition, recognizing that noise characteristics vary spatially, Xu et al. [12] propose to" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21268" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.089, + 0.511, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.268, + 0.362, + 0.351, + 0.375 + ], + "angle": 0, + "content": "(a) Framework" + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.097, + 0.885, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.312, + 0.556, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.315, + 0.636, + 0.326 + ], + "angle": 0, + "content": "Concatenation" + }, + { + "type": "image", + "bbox": [ + 0.667, + 0.312, + 0.688, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.315, + 0.738, + 0.326 + ], + "angle": 0, + "content": "Sigmoid" + }, + { + "type": "image", + "bbox": [ + 0.782, + 0.311, + 0.802, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.81, + 0.313, + 0.882, + 0.325 + ], + "angle": 0, + "content": "Squared ReLU" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.332, + 0.555, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.334, + 0.634, + 0.353 + ], + "angle": 0, + "content": "Element-wise Multiplication" + }, + { + "type": "image", + "bbox": [ + 0.667, + 0.332, + 0.688, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.333, + 0.764, + 0.353 + ], + "angle": 0, + "content": "Element-wise Addition" + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.362, + 0.89, + 0.375 + ], + "angle": 0, + "content": "(b) URWKV Block with Multi-state Spatial and Channel Mixing" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.386, + 0.907, + 0.457 + ], + "angle": 0, + "content": "Figure 2. Overview of the URWKV model. The core URWKV block is integrated into an encoder-decoder framework, featuring a multi-state spatial mixing sub-block and a multi-state channel mixing sub-block. The luminance-adaptive normalization (LAN) incorporates both the current input state \\( X_{t} \\) and multiple inter-stage states \\( M_{i} \\). Additionally, the multi-state quad-directional token shift (SQ-Shift) aggregates both current input state and multiple intra-stage states \\( H_{i} \\) (\\( H_{S_{i}} \\) for spatial mixing and \\( H_{C_{i}} \\) for channel mixing, respectively). The state-aware selective fusion (SSF) module further aggregates rich contextual information across encoder stages." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.476, + 0.483, + 0.566 + ], + "angle": 0, + "content": "separate signal from noise during the enhancement process, enabling more effective brightening while minimizing noise amplification. Furthermore, frequency-domain approaches [26, 27] also contribute to LLIE by decomposing features across distinct frequency bands, proving effective in preserving fine details and reducing noise." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.567, + 0.483, + 0.703 + ], + "angle": 0, + "content": "Considering blur as a frequent coupled degradation in low-light scenes, one that conventional LLIE models inadequately address, Zhou et al. [13] and Li et al. [2] propose a specific sequential encoder-decoder architecture and a parallel heterogeneous decoupling framework, respectively. Although these LLIE-deblur models show effectiveness within certain constraints, further exploration is warranted to handle the dynamically coupled degradations typical in real-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.712, + 0.411, + 0.729 + ], + "angle": 0, + "content": "2.2. Unified Model for Image Restoration" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "With the emergence of deep learning, unified approaches have driven substantial progress, enabling the handling of multiple restoration tasks within a single framework. Compared to CNN-based methods [14, 28], which primarily focus on multi-scale feature extraction and local context enhancement, Transformer-based methods [3, 15, 29, 30] extend relational modeling capabilities by capturing long-range dependencies, leading to improved restoration quality across diverse degradation types. However, the application of Transformers in image restoration often entails high computational costs and memory requirements, which can" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.906, + 0.581 + ], + "angle": 0, + "content": "hinder their efficiency. In addition, frequency-domain models [31, 32] provide another unified approach by decomposing images into multiple frequency bands, which enables focused noise reduction while preserving fine details across tasks. Recent innovations also utilize Mamba networks [16] to achieve effective image restoration by efficiently balancing local and global feature representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.906, + 0.643 + ], + "angle": 0, + "content": "While current unified models perform effectively under normal lighting conditions, they encounter significant challenges in low-light environments, specifically involving complex lighting variations and coupled degradations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.657, + 0.605, + 0.672 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.625, + 0.697 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.906, + 0.84 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, the URWKV framework follows the typical encoder-decoder architecture, with both the encoder and decoder comprising three stages. In the encoder, each stage is composed of \\( N_{1} \\) URWKV blocks followed by a down-sampling layer, while each stage in the decoder contains \\( N_{2} \\) URWKV blocks and an up-sampling layer. Additionally, each down-sampling or up-sampling layer includes convolutional operations and sampling techniques to adjust either the channel or spatial dimensions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "To effectively handle dynamically coupled degradations, each URWKV block operates as an independent and progressive processing stage, with each stage's output (an implicit restored result, serving as a 'state') encapsulat-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21269" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.089, + 0.418, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.16, + 0.241, + 0.366, + 0.254 + ], + "angle": 0, + "content": "(a) Luminance-adaptive Normalization" + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.089, + 0.891, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.241, + 0.808, + 0.255 + ], + "angle": 0, + "content": "(b) Multi-kernel aggregation of inter-stage states" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.266, + 0.905, + 0.294 + ], + "angle": 0, + "content": "Figure 3. Illustration of luminance-adaptive normalization (LAN). LAN integrates inter-stage states \\( M_{i} \\) throughout the restoration process, facilitating scene-aware luminance modulation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.314, + 0.482, + 0.344 + ], + "angle": 0, + "content": "ing stage-specific contextual semantics and restoration cues (e.g., luminance or content) for subsequent stages." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.344, + 0.483, + 0.6 + ], + "angle": 0, + "content": "Specifically, the URWKV block builds on the strengths of original RWKV model [19], while enhancing its capabilities through two tailored sub-blocks: a multi-state spatial mixing sub-block and a multi-state channel mixing sub-block. The former captures spatial dependencies across multiple intra-stage states to enrich spatial information, while the latter enhances feature representations through multi-state channel-wise interactions. Here, the intra-stage states \\( H_{S_i} \\) or \\( H_{C_i} \\) are updated following the SQ-Shift in each sub-block, respectively, where \\( i \\in \\{1, \\dots, N_1\\} \\) (or \\( i \\in \\{1, \\dots, N_2\\} \\)). To ensure efficient information transfer between encoder and decoder stages without introducing degradation, we incorporate the state-aware selective fusion (SSF) module between these stages. The SSF dynamically aggregates multi-state features across all encoder stages, selectively propagating rich contextual information into the decoder." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.607, + 0.418, + 0.623 + ], + "angle": 0, + "content": "3.2. URWKV with Multi-state Perspective" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Original RWKV models [18, 19] employ a single-state token shift mechanism, where each token's relationship relies solely on the current and preceding state. This design introduces two primary limitations: First, it struggles to retain long-range dependencies, causing early information to gradually fade during restoration. Second, it fails to effectively capture complex interdependencies among degraded features, as a single state is insufficient to represent intricate coupled degradations. To this end, we propose URWKV devised with the multi-state perspective. Specifically, the URWKV block tailors the luminance-adaptive normalization (LAN) by incorporating multiple inter-stage states to adapt URWKV to low-light scenarios. Additionally, it integrates multiple intra-stage states to facilitate more extensive interactions across long-range dependencies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Luminance-adaptive Normalization (LAN). Luminance distortion is a major factor leading to other coupled degradations in low-light images. Inspired by the hu" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.314, + 0.905, + 0.42 + ], + "angle": 0, + "content": "man eye's ability to adaptively adjust pupil size in response to varying luminance [33, 34], LAN leverages inter-stage states throughout the restoration process (comprising six stages) to enable scene-aware luminance modulation. This design overcomes limitation of LayerNorm widely used in existing models [3, 17], which relies on fixed parameters \\(\\gamma\\) for scaling the learned representation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.42, + 0.906, + 0.722 + ], + "angle": 0, + "content": "As illustrated in Fig. 3, given the current state input \\(X_{t}\\in \\mathbb{R}^{C_{t}\\times H_{t}\\times W_{t}}\\) at stage \\(T\\), we first extract 1D luminance vectors from both \\(X_{t}\\) and each historical stage output \\(M_{i}\\in \\mathbb{R}^{C_{i}\\times H_{i}\\times W_{i}}\\), using the global average pooling (GAP) operation. Here, each \\(M_{i}\\) is obtained from the output of each stage's final URWKV block. \\(t\\in \\{1,\\dots,N_1\\}\\) (or \\(t\\in \\{1,\\dots,N_2\\}\\)) and \\(i\\in \\{1,\\dots,T - 1\\}\\). The luminance vectors are then zero-padded to match the dimensionality of the vector with the maximum channel size \\(C_{max}\\), while preserving the semantic integrity of luminance vector from each individual state. These zero-padded luminance vectors, denoted as \\(X_{k}^{1D}\\in \\mathbb{R}^{C_{max}}\\), for \\(k\\in \\{1,\\dots,T\\}\\), are then stacked along the state dimension. Next, a multi-kernel aggregation layer is applied to analyze the stacked 2D luminance map along the channel dimension. Specifically, the aggregation layer uses 1D convolutional kernels of sizes \\(1\\times T\\), \\(3\\times T\\), and \\(5\\times T\\) to adaptively capture both local and broader contextual illuminance variations across multiple states. The resulting luminance maps \\(X_{r}^{2D}\\in \\mathbb{R}^{T\\times C_{max}}\\) are expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.733, + 0.905, + 0.753 + ], + "angle": 0, + "content": "\\[\nX _ {r} ^ {2 D} = \\operatorname {C o n v} _ {r} ^ {1 D} \\left(\\left[ X _ {1} ^ {1 D},..., X _ {T} ^ {1 D} \\right]\\right), r \\in \\{1, 3, 5 \\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.904, + 0.795 + ], + "angle": 0, + "content": "where \\(\\mathrm{Conv}_r^{1D}\\) denotes the set of 1D convolutional kernels, and \\([\\cdot ,\\cdot ]\\) denotes the stacking operation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Subsequently, the output of each convolutional kernel is concatenated along the channel dimension and projected through a \\(1\\times 1\\) convolution, resulting in the aggregated feature map \\(X_{\\mathrm{agg}}\\in \\mathbb{R}^{C_t}\\). A multi-layer perceptron (MLP) along with a tanh activation function are then employed to predict the luminance modulator \\(\\Delta \\gamma_{t}\\). The luminance modulator is computed as:" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21270" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.195, + 0.108, + 0.483, + 0.126 + ], + "angle": 0, + "content": "\\[\n\\Delta \\gamma_ {t} = \\tanh (\\mathrm {M L P} \\left(X _ {\\text {a g g}}\\right)), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.132, + 0.446, + 0.147 + ], + "angle": 0, + "content": "where \\(\\tanh (\\cdot)\\) represents the tanh activation function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.148, + 0.484, + 0.207 + ], + "angle": 0, + "content": "Once \\(\\Delta \\gamma_{t}\\) is predicted, the scaling parameter \\(\\gamma\\) can be updated as \\(\\hat{\\gamma}_{t} = \\gamma +\\Delta \\gamma_{t}\\). Ultimately, the input feature \\(X_{t}\\) at stage \\(T\\) can be adaptively normalized with LAN, expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.22, + 0.483, + 0.252 + ], + "angle": 0, + "content": "\\[\nX _ {t} ^ {L A N} = \\frac {X _ {t} - \\mu_ {t}}{\\sigma_ {t}} \\odot \\hat {\\gamma} _ {t} + \\beta , \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.258, + 0.483, + 0.288 + ], + "angle": 0, + "content": "where \\(\\mu_t\\) and \\(\\sigma_t\\) represent the mean and standard deviation of the input feature \\(X_t\\), respectively. \\(\\beta\\) is a shift parameter." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.288, + 0.484, + 0.56 + ], + "angle": 0, + "content": "Multi-state Quad-directional Token Shift (SQ-Shift). Original quad-directional token shift (Q-shift) [19] effectively captures spatial dependencies among neighboring tokens. However, it neglects the degradation relations between states, making it difficult to maintain consistent degradation restoration across consecutive states. To address this limitation, we aggregate multi-state features to extend Q-Shift, allowing the model to capture long-range dependencies across multiple states while preserving local spatial interactions. Specifically, before applying Q-shift, we employ the exponential moving average (EMA) approach to aggregate the current state with all preceding states within the same stage, enabling richer interactions with historical information. For instance, given the state feature \\( X_{t}^{LAN} \\in \\mathbb{R}^{C \\times H \\times W} \\) normalized by LAN, and multiple historical intra-stage state features \\( H_{i}, i \\in \\{1, \\dots, t - 1\\} \\), we define the multi-state aggregation MSA as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.571, + 0.483, + 0.591 + ], + "angle": 0, + "content": "\\[\n\\operatorname {M S A} \\left(X _ {t} ^ {L A N}\\right) = \\alpha \\odot X _ {t} ^ {L A N} + (1 - \\alpha) \\odot \\operatorname {M S A} \\left(H _ {t - 1}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.602, + 0.483, + 0.632 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a decay factor that controls the weight between the current state and the previously aggregated state." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.633, + 0.483, + 0.829 + ], + "angle": 0, + "content": "Multi-State Spatial and Channel Mixing. For each input feature in the multi-state spatial mixing sub-block or multi-state channel mixing sub-block, we first apply LAN and SQ-Shift, resulting in features \\( X_{s} \\) for spatial mixing or \\( X_{c} \\) for channel mixing. In the spatial mixing process, \\( X_{s} \\) undergoes three parallel linear projections to generate the matrices for receptance \\( R_{s} \\), key \\( K_{s} \\), and value \\( V_{s} \\). Following the computational approach of [19], we employ the bidirectional attention mechanism Bi-WKV to compute the global attention weights \\( wkv \\) between \\( K_{s} \\) and \\( V_{s} \\), which are subsequently modulated by the sigmoid-activated receptance \\( R_{s} \\) to control the output. The output feature \\( O_{s} \\) is therefore represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.842, + 0.483, + 0.86 + ], + "angle": 0, + "content": "\\[\nO _ {s} = \\left(\\sigma \\left(R _ {s}\\right) \\odot w k v\\right) W _ {O _ {s}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where \\( wkv = \\mathrm{Bi - WKV}(K_{\\mathrm{s}},V_{\\mathrm{s}}) \\). \\( \\sigma \\) represents the sigmoid function. \\( W_{O_s} \\) is learnable parameter of projection layer." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.66, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.289, + 0.655, + 0.298 + ], + "angle": 0, + "content": "(a) State-aware Selective Fusion" + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.088, + 0.903, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.749, + 0.289, + 0.831, + 0.299 + ], + "angle": 0, + "content": "(c) State Alignment" + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.306, + 0.892, + 0.321 + ], + "angle": 0, + "content": "Figure 4. Illustration of the state-aware fusion (SSF) module." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.333, + 0.906, + 0.393 + ], + "angle": 0, + "content": "For channel mixing, \\( X_{c} \\) undergoes two parallel linear projections to derive the receptance matrix \\( R_{c} \\) and key matrix \\( K_{c} \\). Then, the output feature \\( O_{c} \\) for channel mixing can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.401, + 0.905, + 0.419 + ], + "angle": 0, + "content": "\\[\nO _ {c} = \\left(\\sigma \\left(R _ {c}\\right) \\odot \\operatorname {S q u a r e d R e L U} \\left(K _ {c}\\right) W _ {V _ {c}}\\right) W _ {O _ {c}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.906, + 0.456 + ], + "angle": 0, + "content": "where \\( W_{V_c} \\) and \\( W_{O_c} \\) are learnable parameters in the two projection layers." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.463, + 0.816, + 0.479 + ], + "angle": 0, + "content": "3.3. State-aware Selective Fusion (SSF)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.906, + 0.59 + ], + "angle": 0, + "content": "Existing models typically reintroduce early-stage features using navie skip connections (e.g., concatenation or addition). However, they are prone to two major issues, particularly under low-light conditions: (1) the propagation of noise and irrelevant information from the encoder to the decoder [1], and (2) the semantic gap between stages [36], which hinders effective feature fusion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.591, + 0.907, + 0.879 + ], + "angle": 0, + "content": "Rather than adopting the navie or parameter-heavy approaches [3, 17, 36], we propose the lightweight and effective state-aware selective fusion (SSF) module. As illustrated in Fig. 4, the SSF module focuses on capturing the degradation restoration relationships between multistate features to predict the spatial guiding weight, which selectively refines the transfer of encoder features to the decoder. Given three encoder outputs \\( E_{i} \\in \\mathbb{R}^{C_{i} \\times H_{i} \\times W_{i}} \\) obtained from \\( M_{i} \\), \\( i \\in \\{1,2,3\\} \\), we first apply a channel-wise mean operation to collapse the channel dimension, resulting in \\( E_{i}^{M} \\in \\mathbb{R}^{1 \\times H_{i} \\times W_{i}} \\), which reduces inter-channel semantic interference. Next, the adaptive alignment operation is performed on \\( E_{i}^{M} \\) to adjust the spatial dimensions of each feature according to the target dimensions of the decoder at the current fusion stage. The aligned features \\( E_{i}^{A} \\) are then stacked along the channel dimension to produce the concatenated feature map \\( E^{A} \\in \\mathbb{R}^{3 \\times H_{d} \\times W_{d}} \\), where \\( H_{d} \\) and \\( W_{d} \\) are the height and width of the decoder's feature. This process is formally expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.884, + 0.907, + 0.903 + ], + "angle": 0, + "content": "\\[\nE ^ {A} = \\operatorname {S t a c k} (\\operatorname {A d a A l i g n} (\\operatorname {A v g} (E _ {i}))), i \\in \\{1, 2, 3 \\}, \\quad (7)\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21271" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison results on the LOL-v2-real, LOL-v2-syn, SID, SMID, SDSD-indoor, and SDSD-outdoor datasets. Joint LLIE and deblurring models are marked with †, unified models are marked with *, and the rest represent LLIE models." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.126, + 0.905, + 0.395 + ], + "angle": 0, + "content": "
MethodsLOL-v2-realLOL-v2-synSIDSMIDSDSD-indoorSDSD-outdoorComplexity
PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑FLOPs (G)Params (M)
KinD [7]14.740.64113.290.57818.020.58322.180.63421.950.67221.970.65434.998.02
EnlightenGAN [35]18.230.61716.570.73417.230.54322.620.67420.020.60420.100.61661.01114.35
SNR-Net [12]21.480.84924.140.92721.350.55028.520.80726.130.81519.220.65726.354.01
FourLLIE [26]22.350.84724.650.91918.420.51325.640.75024.740.82624.670.787-0.12
UHDFour [27]21.790.85423.600.91322.740.65128.180.80929.130.89523.280.77157.4215.90
LLFormer [10]21.630.80824.130.90822.830.65628.420.79426.820.82529.180.84122.5224.52
Retinexformer [9]22.790.83925.670.92824.440.68029.150.81529.780.89529.830.87815.571.61
BiFormer [1]22.670.86324.870.92122.490.63828.240.81228.730.89926.130.8415.930.83
RetinexMamba [11]22.450.84325.880.93322.450.65628.620.80928.440.89428.520.85934.7524.1
LEDNet†[13]19.020.83524.790.93121.470.63828.420.80727.290.87626.660.85038.577.41
PDHAT†[2]20.160.84124.940.93721.930.65329.190.81726.370.88428.190.862208.197.83
Uformer* [17]18.820.77119.660.87118.540.57727.200.79223.170.85923.850.74812.005.29
MIRNet* [14]22.680.82825.050.92321.360.63226.210.76928.640.88828.990.869785.0031.76
Restormer* [3]18.600.78921.410.83122.010.64528.580.80628.490.89227.990.868140.9926.11
MambaIR* [16]20.450.84525.650.93922.020.65828.410.80525.140.87627.530.85160.664.30
URWKV (Ours)*23.110.87426.360.94423.110.67329.440.82631.240.91129.990.88718.342.25
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.482, + 0.485 + ], + "angle": 0, + "content": "where Avg denotes the channel-wise mean operation. AdaAlign combines both upsampling and downsampling to adaptively align multiple states with the target state. Stack is the operation that stacks the aligned features along the channel axis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.483, + 0.637 + ], + "angle": 0, + "content": "To facilitate more nuanced fusion of local details and broader context, we incorporate an inception block [37] with multiple convolutional filters \\((1\\times 1,3\\times 3,\\) and \\(5\\times 5)\\) to aggregate degradation patterns across various receptive fields. The aggregated feature \\(E_{agg}\\) is then projected via a convolutional layer, followed by a sigmoid function to produce the spatial weight. Therefore, taking the first stage of the decoder as an example, the naive concatenation-based skip connection for the URWKV block in the decoder, following the SSF module, can be extended as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.647, + 0.483, + 0.667 + ], + "angle": 0, + "content": "\\[\nD _ {1} ^ {\\prime} = \\left(\\left[ W _ {s} \\odot E _ {3}, D _ {1} \\right]\\right) W _ {p} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.752 + ], + "angle": 0, + "content": "where \\( W_{s} \\) is the spatial weight predicted by the SSF module, \\( E_{3} \\) is the encoder feature, and \\( D_{1} \\) is the input feature to the first stage of the decoder. \\( \\odot \\) denotes element-wise multiplication, and \\( W_{p} \\) represents a convolutional layer applied to refine the features. \\( [\\cdot ,\\cdot ] \\) denote the concatenate operation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.764, + 0.224, + 0.781 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.789, + 0.298, + 0.805 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We implement our model in PyTorch on a workstation equipped with an NVIDIA Tesla A40 GPU. The number of URWKV blocks is set to \\( N_{1} = 3 \\) and \\( N_{2} = 2 \\), while the number of input channels is set to \\( C = 32 \\). The decay factor \\( \\alpha \\) is set to 0.5. During training, we use a unified loss function for all datasets, as in [1], which combines L1 Loss," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.562 + ], + "angle": 0, + "content": "SSIM Loss, and Perceptual Loss. The Adam optimizer [38] is employed with \\(\\beta_{1} = 0.9\\) and \\(\\beta_{2} = 0.99\\), and the learning rate is initialized at \\(2 \\times 10^{-4}\\) and decays gradually to \\(10^{-6}\\) using cosine annealing decay. Following [1, 10], we apply augmentation techniques, such as flipping and rotation, to each image pair. During testing, our URWKV model processes input images of arbitrary shape directly, without the need for cropping or resizing. For evaluation, we employ commonly-used PSNR [39] and SSIM [40] metrics to calculate images in the RGB color space across all datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.574, + 0.654, + 0.589 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.906, + 0.657 + ], + "angle": 0, + "content": "To assess the performance of various models in scenarios involving dynamically coupled degradations, we compare the LLIE models, LLIE-deblur models, and unified models across 8 typical low-light benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.906, + 0.794 + ], + "angle": 0, + "content": "LOL. As shown in Table 1, our URWKV demonstrates superior performance compared to LLIE models on LOLv2-real [41], outperforming Retinexformer by 0.32 dB in PSNR and 0.035 in SSIM, respectively. Similarly, on LOLv2-syn [41], our model outperforms RetinexMamba by 0.48 dB in PSNR and 0.011 in SSIM. Notably, LLIE-deblur models [2, 13] and other unified models [3, 14, 16] generally underperform on both datasets, with substantial performance gaps between them and our model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "SID and SMID. As shown in Table 1, our UR-WKV model delivers superior performance, closely trailing Retinexformer on SID [42], while outperforming it by 0.29 dB in PSNR on SMID [42]. In contrast, LLIE-deblur models exhibit inconsistent performance across the datasets. For instance, while PDHAT ranks second on SMID, it significantly underperforms on SID, achieving a PSNR of just" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21272" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.21, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.09, + 0.326, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.091, + 0.439, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.44, + 0.091, + 0.554, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.091, + 0.668, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.091, + 0.781, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.782, + 0.091, + 0.898, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.26, + 0.826, + 0.274 + ], + "angle": 0, + "content": "Figure 5. Visual comparison of state-of-the-art models on the SID dataset (top) and the SMID dataset (bottom)." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.289, + 0.222, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.289, + 0.348, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.348, + 0.289, + 0.476, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.122, + 0.404, + 0.45, + 0.419 + ], + "angle": 0, + "content": "Figure 6. Visual comparisons on SDSD-indoor dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.483, + 0.478 + ], + "angle": 0, + "content": "21.93 dB, which is 1.18 dB lower than our model. Visual comparisons in Fig. 5 also demonstrate the excellent denoising capability of our model when enhancing brightness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.483, + 0.629 + ], + "angle": 0, + "content": "SDSD. Results depicted in Table 1 indicate that even some recent LLIE models [1, 26] still encounter challenges on SDSD [43]. In contrast, our URWKV model delivers state-of-the-art performance across both subsets. Moreover, compared to unified models like Restormer, URWKV outperforms it by \\(2.75\\mathrm{dB}\\) in PSNR on SDSD-indoor and by \\(2.00\\mathrm{dB}\\) in PSNR on SDSD-outdoor. As demonstrated in Fig. 6, our method preserves high-fidelity restoration in the scenario with strong light-dark contrast, whereas other models (e.g. PDHAT) generate unexpected artifacts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.483, + 0.765 + ], + "angle": 0, + "content": "FiveK. The MIT-Adobe FiveK dataset [44] challenges models with the intricate task of color restoration in lowlight and underexposed images. As reported in Table 2, our model achieves the highest PSNR of \\(26.08\\mathrm{dB}\\) on FiveK dataset, surpassing LLFormer by \\(0.48\\mathrm{dB}\\) and Retinexformer by \\(1.50\\mathrm{dB}\\). Furthermore, URWKV outperforms the unified model, Restormer, in SSIM by 0.007. These notable improvements underscore URWKV's proficiency in restoring both structural integrity and aesthetic quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.766, + 0.483, + 0.901 + ], + "angle": 0, + "content": "LOL-blur. The LOL-blur [13] dataset couples varying degrees of low-light and motion blur, which has often been underrepresented in other LLIE datasets. As shown in Table 3, conventional LLIE models typically struggle to address this coupled degradation, while unified models also fail to outperform specialized LLIE-deblur models. In contrast, our URWKV model achieves the highest performance in this scenario, with a PSNR of 27.27 dB and SSIM of 0.890, outperforming the second-best LLIE-Deblur model" + }, + { + "type": "table_caption", + "bbox": [ + 0.531, + 0.289, + 0.887, + 0.302 + ], + "angle": 0, + "content": "Table 2. Experimental results on MIT-Adobe FiveK dataset." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.306, + 0.905, + 0.416 + ], + "angle": 0, + "content": "
MethodsEnlightenGAN [35]RUAS [45]RetinexNet [8]SNR-Net [12]LLFormer [10]BiFormer [1]
PSNR ↑13.2620.9116.3224.3925.6025.51
SSIM ↑0.7450.8630.7820.8950.9170.929
MethodsRetinexformer [9]LEDNet [13]PDHAT [2]Restormer [3]MIRNet [14]URWKV (Ours)
PSNR ↑24.5824.3425.0025.4823.7326.08
SSIM ↑0.8860.9150.9290.9290.8950.936
" + }, + { + "type": "table_caption", + "bbox": [ + 0.557, + 0.427, + 0.86, + 0.44 + ], + "angle": 0, + "content": "Table 3. Experimental results on LOL-blur dataset." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.445, + 0.905, + 0.555 + ], + "angle": 0, + "content": "
MethodsFourLLIE [26]UHDFour [27]LLFormer [10]Retinexformer [9]BiFormer [1]GLARE [46]
PSNR ↑19.8125.3424.5525.2524.6723.26
SSIM ↑0.6830.8290.7850.8210.8130.690
MethodsLEDNet [13]PDHAT [2]Restormer [3]MIRNet [14]MambaIR [16]URWKV (Ours)
PSNR ↑26.0626.7126.3823.9926.2827.27
SSIM ↑0.8460.8790.8600.7740.8480.890
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.905, + 0.613 + ], + "angle": 0, + "content": "PDHAT by 0.56 dB in PSNR and 0.011 in SSIM. Examples in Fig. 7 vividly demonstrate the potential of our model in restoring coupled low-light and blur images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.622, + 0.667, + 0.639 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.905, + 0.689 + ], + "angle": 0, + "content": "We conduct several ablation studies on the real-world dataset (i.e., LOL-v2-real) to analyze the impact of various components in our model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of LAN and SSF modules. To validate the individual and combined contributions of the LAN and SSF modules, we evaluate four model variants. As shown in Table 4, models incorporating either LAN or SSF individually show a clear improvement over the baseline (first row in Table 4). Among these, the model with LAN alone achieves a more pronounced gain, with PSNR increasing by \\(1.38\\mathrm{dB}\\) over the baseline, highlighting the critical role of LAN in dynamically adapting to complex luminance variations. Moreover, our final model, which combines both LAN and SSF, incurs only a minor increase in parameters (from 1.64M to 2.25M) and a minimal rise in FLOPs (from 18.25G to 18.34G), yet delivers a significant performance boost compared to baseline. The feature transformation vi" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21273" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.089, + 0.9, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.272, + 0.267, + 0.725, + 0.281 + ], + "angle": 0, + "content": "Figure 7. Visual comparison of state-of-the-art models on LOL-blur dataset." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.295, + 0.48, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.39, + 0.483, + 0.434 + ], + "angle": 0, + "content": "Figure 8. Top: LAN adaptively enhances the luminance, making details clearer. Bottom: SSF helps mitigate introducing degradation, obtaining fused feature \\((D_3^{\\prime})\\) based on \\(E_{1}\\), \\(E_{2}\\) and \\(E_{3}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.483, + 0.474 + ], + "angle": 0, + "content": "sualizations presented in Fig. 8 also validate the inherent advantages of LAN and SSF." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.477, + 0.483, + 0.702 + ], + "angle": 0, + "content": "Multi-state Quad-directional Token Shift. Our propose multi-state aggregation strategy enhances the original Q-shift, allowing for a more comprehensive understanding of degradation representation. As shown in Table 5, our results indicate that single-state aggregation alone (Table 5a) achieves performance comparable to the model using Q-shift alone (Table 5c). In contrast, multi-state aggregation alone (Table 5b) consistently demonstrates superior performance, attaining a \\(0.13\\mathrm{dB}\\) higher PSNR. When combining single-state aggregation with Q-shift (Table 5d), the model shows an improvement in PSNR. However, it still falls short in SSIM by 0.003 compared to multi-state aggregation alone. Ultimately, our final model (Table 5e), which integrates both Q-shift and multi-state aggregation, demonstrates substantial advantages over all other variants." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "SSF vs. Navie Skip Connections. The ablation results in Table 6 highlight the effectiveness of SSF compared to naive skip connections, including simple addition (Add) and concatenation (Cat). When using multi-state encoder features, noise from earlier stages and semantic gaps between states lead to performance degradation in both PSNR and SSIM. In contrast, our SSF module, based on state-aware fusion strategy, avoids directly merging multistate features and predicts a guiding weight to selectively fuse useful patterns, mitigating issues of noise propagation and semantic misalignment. Notably, the \"Multi-state+SSF\" (i.e., our URWKV) yields significant improvements in PSNR and SSIM over \"Single-state+Cat\" without" + }, + { + "type": "table_caption", + "bbox": [ + 0.557, + 0.295, + 0.862, + 0.309 + ], + "angle": 0, + "content": "Table 4. Ablation study on LAN and SSF modules." + }, + { + "type": "table", + "bbox": [ + 0.533, + 0.314, + 0.88, + 0.387 + ], + "angle": 0, + "content": "
LANSSFPSNR ↑SSIM ↑ParamsFLOPs
21.330.8561.64M18.25G
✓21.400.8611.65M18.29G
✓22.710.8692.25M18.30G
✓✓23.110.8742.25M18.34G
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.398, + 0.905, + 0.412 + ], + "angle": 0, + "content": "Table 5. Ablation study on multi-state quad-directional token shift." + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.417, + 0.902, + 0.506 + ], + "angle": 0, + "content": "
#Single-stateMulti-stateQ-shiftPSNR ↑SSIM ↑
(a)✓22.120.861
(b)✓22.250.870
(c)✓22.190.862
(d)✓✓22.290.867
(e)✓✓23.110.874
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.517, + 0.905, + 0.545 + ], + "angle": 0, + "content": "Table 6. Impacts of SSF and navie skip connections (Skip Conn.) with different state numbers (State No.)." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.55, + 0.902, + 0.628 + ], + "angle": 0, + "content": "
State No.Skip Conn.PSNR ↑SSIM ↑ParamsFLOPs
Single-stateAdd22.660.8721.86M14.68G
Single-stateCat22.710.8692.25M18.30G
Multi-stateCat21.830.8592.50M23.89G
Multi-stateSSF23.110.8742.25M18.34G
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.641, + 0.905, + 0.672 + ], + "angle": 0, + "content": "increasing the model's parameter, demonstrating both enhanced performance and computational efficiency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.679, + 0.634, + 0.695 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To address the shortcomings of existing models, we propose URWKV with multi-state perspective, enabling flexible and effective restoration across dynamically coupled degradation scenarios. Specifically, we introduce three key innovations: luminance-adaptive normalization (LAN) for adaptive, scene-aware luminance modulation; an exponential moving average approach to aggregate intra-stage states for spatial-level token-shift; and the state-aware selective fusion (SSF) module for dynamic alignment and integration of multi-state features across encoder stages. In comparison to existing models, our model demonstrates superior performance while requiring significantly fewer parameters and computational resources." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21274" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.274 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported in part by the National Natural Science Foundation of China under Grant 62471142, in part by the Natural Science Foundation of Fujian Province, China under Grants 2023J01067, in part by Industry-Academy Cooperation Project under Grant 2024H6006, in part by the Collaborative Innovation Platform Project of Fuzhou City under Grant 2023-P-002, in part by the Key Technology Innovation Project for Focused Research and Industrialization in the Software Industry of Fujian Province, and in part by the Key Research and Industrialization Project of Technological Innovation in Fujian Province under Grant 2024XQ002." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.29, + 0.188, + 0.305 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.315, + 0.482, + 0.381 + ], + "angle": 0, + "content": "[1] Rui Xu, Yuezhou Li, Yuzhen Niu, Huangbiao Xu, Yuzhong Chen, and Tiesong Zhao. Bilateral interaction for local-global collaborative perception in low-light image enhancement. IEEE Transactions on Multimedia, 2024. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.387, + 0.482, + 0.454 + ], + "angle": 0, + "content": "[2] Yuezhou Li, Rui Xu, Yuzhen Niu, Wenzhong Guo, and Tiesong Zhao. Perceptual decoupling with heterogeneous auxiliary tasks for joint low-light image enhancement and deblurring. IEEE Transactions on Multimedia, pages 6663-6675, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.459, + 0.482, + 0.54 + ], + "angle": 0, + "content": "[3] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.544, + 0.482, + 0.611 + ], + "angle": 0, + "content": "[4] Andong Lu, Zhang Zhang, Yan Huang, Yifan Zhang, Chenglong Li, Jin Tang, and Liang Wang. Illumination distillation framework for nighttime person re-identification and a new benchmark. IEEE Transactions on Multimedia, pages 1–14, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.615, + 0.482, + 0.656 + ], + "angle": 0, + "content": "[5] Xiaopeng Li and Xiaojie Guo. SPN2D-GAN: Semantic prior based night-to-day image-to-image translation. IEEE Transactions on Multimedia, pages 1-15, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.66, + 0.482, + 0.713 + ], + "angle": 0, + "content": "[6] Xuechen Bai, Xinghua Li, Jianhao Miao, and Huanfeng Shen. Making the rarth clear at night: A high-resolution nighttime light image deblooming network. IEEE Transactions on Geoscience and Remote Sensing, 61:1-13, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.717, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[7] Yonghua Zhang, Jiawan Zhang, and Xiaojie Guo. Kindling the darkness: A practical low-light image enhancer. In Proceedings of the ACM International Conference on Multimedia, pages 1632-1640, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.775, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[8] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. In Proceedings of the British Machine Vision Conference, 2018. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[9] Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang. Retinexformer: One-stage retina-based transformer for low-light image enhancement. In Proceedings of the IEEE International Conference on Computer Vision, pages 12504-12513, 2023. 1, 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.315, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[10] Tao Wang, Kaihao Zhang, Tianrun Shen, Wenhan Luo, Bjorn Stenger, and Tong Lu. Ultra-high-definition low-light image enhancement: A benchmark and transformer-based method. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 2654-2662, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[11] Jiesong Bai, Yuhao Yin, and Qiyuan He. Retinexmamba: Retinex-based mamba for low-light image enhancement. arXiv preprint arXiv:2405.03349, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.206, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[12] Xiaogang Xu, Ruixing Wang, Chi-Wing Fu, and Jiaya Jia. SNR-aware low-light image enhancement. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 17714-17724, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.261, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[13] Shangchen Zhou, Chongyi Li, and Chen Change Loy. LEDNet: Joint low-light enhancement and deblurring in the dark. In Proceedings of the European Conference on Computer Vision, pages 573-589, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.317, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[14] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Learning enriched features for real image restoration and enhancement. In Proceedings of the European Conference on Computer Vision, pages 492-511, 2020. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.387, + 0.905, + 0.454 + ], + "angle": 0, + "content": "[15] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12299-12310, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.457, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[16] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. MambaIR: A simple baseline for image restoration with state-space model. In Proceedings of the European Conference on Computer Vision, pages 222-241, 2025. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.526, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[17] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 17683-17693, 2022. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.905, + 0.661 + ], + "angle": 0, + "content": "[18] Bo Peng, Eric Alcaide, Quentin Anthony, Alon Albalak, Samuel Arcadinho, Stella Biderman, Huanqi Cao, Xin Cheng, Michael Chung, Matteo Grella, et al. RWKV: Reinventing rnns for the transformer era. arXiv preprint arXiv:2305.13048, 2023. 2.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.732 + ], + "angle": 0, + "content": "[19] Yuchen Duan, Weiyun Wang, Zhe Chen, Xizhou Zhu, Lewei Lu, Tong Lu, Yu Qiao, Hongsheng Li, Jifeng Dai, and Wenhai Wang. Vision-RWKV: Efficient and scalable visual perception with RWKV-like architectures. arXiv preprint arXiv:2403.02308, 2024. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.734, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[20] J Alex Stark. Adaptive image contrast enhancement using generalizations of histogram equalization. IEEE Transactions on Image Processing, 9:889-896, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.777, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[21] Mohammad Abdullah-Al-Wadud, Md Hasanul Kabir, M Ali Akber Dewan, and Oksam Chae. A dynamic histogram equalization for image contrast enhancement. IEEE Transactions on Consumer Electronics, 53:593-600, 2007." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[22] Turgay Celik and Tardi Tjahjadi. Contextual and variational contrast enhancement. IEEE Transactions on Image Processing, 20:3431-3441, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Xueyang Fu, Yinghao Liao, Delu Zeng, Yue Huang, Xiao-Ping Zhang, and Xinghao Ding. A probabilistic method for" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21275" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "image enhancement with simultaneous illumination and reflectance estimation. IEEE Transactions on Image Processing, 24:4965-4977, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.191 + ], + "angle": 0, + "content": "[24] Mading Li, Jiaying Liu, Wenhan Yang, Xiaoyan Sun, and Zongming Guo. Structure-revealing low-light image enhancement via robust retinae model. IEEE Transactions on Image Processing, 27:2828-2841, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.484, + 0.248 + ], + "angle": 0, + "content": "[25] Daniel J Jobson, Zia-ur Rahman, and Glenn A Woodell. A multiscale retina for bridging the gap between color images and the human observation of scenes. IEEE Transactions on Image processing, 6:965–976, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.483, + 0.305 + ], + "angle": 0, + "content": "[26] Chenxi Wang, Hongjun Wu, and Zhi Jin. FourLLIE: Boosting low-light image enhancement by fourier frequency information. In Proceedings of the ACM International Conference on Multimedia, pages 7459-7469, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.483, + 0.375 + ], + "angle": 0, + "content": "[27] Chongyi Li, Chun-Le Guo, Man Zhou, Zhexin Liang, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Embedding fourier for ultra-high-definition low-light image enhancement. Proceedings of the International Conference on Learning Representations, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.483, + 0.433 + ], + "angle": 0, + "content": "[28] Yuning Cui, Wenqi Ren, Xiaochun Cao, and Alois Knoll. Revitalizing convolutional network for image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 9423-9438, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.483, + 0.502 + ], + "angle": 0, + "content": "[29] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image restoration using swin transformer. In Proceedings of the IEEE International Conference on Computer Vision, pages 1833-1844, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.483, + 0.574 + ], + "angle": 0, + "content": "[30] Akshay Dudhane, Syed Waqas Zamir, Salman Khan, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Burstormer: Burst image restoration and enhancement transformer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5703-5712, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.483, + 0.631 + ], + "angle": 0, + "content": "[31] Man Zhou, Jie Huang, Chun-Le Guo, and Chongyi Li. Fourmer: An efficient global modeling paradigm for image restoration. In Proceedings of the International Conference on Machine Learning, pages 42589-42601, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.483, + 0.674 + ], + "angle": 0, + "content": "[32] Yuning Cui, Wenqi Ren, Xiaochun Cao, and Alois Knoll. Image restoration via frequency selection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[33] Francisco Lara, Paula Bernal-Molina, Vicente Fernandez-Sanchez, and Norberto López-Gil. Changes in the objective amplitude of accommodation with pupil size. Optometry and Vision Science, 91(10):1215-1220, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[34] Pablo A Barrionuevo, Luis A Issolio, and Constanza Tripolone. Photoreceptor contributions to the human pupil light reflex. Journal of Photochemistry and Photobiology, 15:100178, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[35] Yifan Jiang, Xinyu Gong, Ding Liu, Yu Cheng, Chen Fang, Xiaohui Shen, Jianchao Yang, Pan Zhou, and Zhangyang Wang. EnlightenGAN: Deep light enhancement without paired supervision. IEEE Transactions on Image Processing, 30:2340-2349, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[36] Haonan Wang, Peng Cao, Jiaqi Wang, and Osmar R Zaiane. UCTransNet: Rethinking the skip connections in U-Net from a channel-wise perspective with transformer. In Proceedings" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "of the AAAI conference on artificial intelligence, volume 36, pages 2441-2449, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[37] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1-9, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[38] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[39] Quan Huynh-Thu and Mohammed Ghanbari. Scope of validity of PSNR in image/video quality assessment. *Electronics Letters*, 44:800–801, 2008. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[40] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[41] Wenhan Yang, Wenjing Wang, Haofeng Huang, Shiqi Wang, and Jiaying Liu. Sparse gradient regularized deep retina network for robust low-light image enhancement. IEEE Transactions on Image Processing, 30:2072-2086, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[42] Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3291-3300, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[43] Ruixing Wang, Xiaogang Xu, Chi-Wing Fu, Jiangbo Lu, Bei Yu, and Jiaya Jia. Seeing dynamic scene in the dark: A high-quality video dataset with mechatronic alignment. In Proceedings of the IEEE International Conference on Computer Vision, pages 9700-9709, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[44] Vladimir Bychkovsky, Sylvain Paris, Eric Chan, and Frédo Durand. Learning photographic global tonal adjustment with a database of input/output image pairs. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 97-104, 2011. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.602, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[45] Risheng Liu, Long Ma, Tengyu Ma, Xin Fan, and Zhongxuan Luo. Learning with nested scene modeling and cooperative architecture search for low-light vision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45:5953-5969, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.673, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[46] Han Zhou, Wei Dong, Xiaohong Liu, Shuaicheng Liu, Xiongkuo Min, Guangtao Zhai, and Jun Chen. GLARE: Low light image enhancement via generative latent feature based codebook retrieval. In Proceedings of the European Conference on Computer Vision, pages 36-54, 2024. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.743 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21276" + } + ] +] \ No newline at end of file diff --git a/2025/URWKV_ Unified RWKV Model with Multi-state Perspective for Low-light Image Restoration/e077664a-8838-4206-9246-0d60f57e032f_origin.pdf b/2025/URWKV_ Unified RWKV Model with Multi-state Perspective for Low-light Image Restoration/e077664a-8838-4206-9246-0d60f57e032f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4b86665a0b25882d9bb91be5db32aac3e2ba42a0 --- /dev/null +++ b/2025/URWKV_ Unified RWKV Model with Multi-state Perspective for Low-light Image Restoration/e077664a-8838-4206-9246-0d60f57e032f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7389e883f0b973fa5ba546e81c51f5ed4a0f354d3f82c2ad0e00245ae00d637b +size 6714605 diff --git a/2025/URWKV_ Unified RWKV Model with Multi-state Perspective for Low-light Image Restoration/full.md b/2025/URWKV_ Unified RWKV Model with Multi-state Perspective for Low-light Image Restoration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..24c87f370c490d8419117693c184b35b34944886 --- /dev/null +++ b/2025/URWKV_ Unified RWKV Model with Multi-state Perspective for Low-light Image Restoration/full.md @@ -0,0 +1,335 @@ +# URWKV: Unified RWKV Model with Multi-state Perspective for Low-light Image Restoration + +Rui Xu $^{1,2}$ , Yuzhen Niu $^{1,2*}$ , Yuezhou Li $^{1,2}$ , Huangbiao Xu $^{1,2}$ , Wenxi Liu $^{1,2}$ , Yuzhong Chen $^{1,2}$ $^{1}$ Fujian Key Laboratory of Network Computing and Intelligent Information Processing, College of Computer and Data Science, Fuzhou University, Fuzhou 350108, China + $^{2}$ Engineering Research Center of Big Data Intelligence, Ministry of Education, Fuzhou 350108, China + +{xurui.ryan.chn, yuzhenniu, liyuezhou.cm, huangbiaoxu.chn}@gmail.com + +wenxi.liu@hotmail.com, yzchen@fzu.edu.cn + +# Abstract + +Existing low-light image enhancement (LLIE) and joint LLIE and deblurring (LLIE-deblur) models have made strides in addressing predefined degradations, yet they are often constrained by dynamically coupled degradations. To address these challenges, we introduce a Unified Receptance Weighted Key Value (URWKV) model with multi-state perspective, enabling flexible and effective degradation restoration for low-light images. Specifically, we customize the core URWKV block to perceive and analyze complex degradations by leveraging multiple intra- and interstage states. First, inspired by the pupil mechanism in the human visual system, we propose Luminance-adaptive Normalization (LAN) that adjusts normalization parameters based on rich inter-stage states, allowing for adaptive, scene-aware luminance modulation. Second, we aggregate multiple intra-stage states through exponential moving average approach, effectively capturing subtle variations while mitigating information loss inherent in the single-state mechanism. To reduce the degradation effects commonly associated with conventional skip connections, we propose the State-aware Selective Fusion (SSF) module, which dynamically aligns and integrates multi-state features across encoder stages, selectively fusing contextual information. In comparison to state-of-the-art models, our URWKV model achieves superior performance on various benchmarks, while requiring significantly fewer parameters and computational resources. Code is available at: https://github.com/FZU-N/URWKV. + +# 1. Introduction + +Low-light environments present significant challenges in computer vision and pattern recognition, impacting appli + +![](images/53dcb58340066208f8c65566b6cc341cac530221b8f321650fb0e6540e93b547.jpg) +Figure 1. Compared against existing solutions (including Bi-Former [1] for LLIE, PDHAT [2] for joint LLIE and deblurring, and Restormer [3] as a unified model), our proposed URWKV achieves consistently superior PSNR performance across various degradation scenarios. For better visualization, the maximum PSNR values across all datasets are normalized. + +cations such as nighttime surveillance [4], night photography [5], and remote sensing [6]. In these scenarios, insufficient lighting often leads to severe image degradations, such as increased noise, loss of detail, reduced contrast, and color distortion. Furthermore, another degradation coupled in the low-light images, particularly motion blur induced by prolonged exposure times or camera shake, further compromise the integrity of the images. These challenges not only diminish the visual quality of images but also impede subsequent image analysis and interpretation. + +Low-light image enhancement (LLIE) models are foundational and widely researched solutions for addressing image degradations in low-light environments. Recent advancements have integrated sophisticated architectures like CNNs [7, 8], Transformers [9, 10], and Mamba networks [11], enabling more effective handling of complex degrad- + +tion patterns. To address specific challenges such as uneven brightness and spatially varying noise, several specialized methods [1, 12] have been developed, enabling finer adjustments across different image regions. Despite these advancements, another typical coupled degradation, i.e., motion blur, that frequently occurs in low-light conditions, remains a significant challenge for current LLIE models. + +The primary difficulty for addressing coupled degradations arises from the inherent trade-offs among brightness enhancement, noise reduction, and blur restoration. Traditional joint training strategy or the combination strategy of off-the-shelf LLIE and deblurring models often fail to deliver optimal results [2, 13]. To bridge this gap, Zhou et al. [13] introduce a compact encoder-decoder architecture wherein the encoder focuses on low-light enhancement, while the decoder tackles blur restoration, effectively isolating the two tasks. Additionally, Li et al. [2] introduce a perceptual decoupling approach that incorporates auxiliary tasks to achieve task-specific representations. While these methods show efficacy in handling blur under low-light conditions, they are constrained by predefined degradation categories. As a result, their adaptability to the dynamically coupled degradations typically encountered in real-world scenarios is limited, where noise, blur, and other degradations may not occur in fixed, predictable combinations. + +To flexibly address various degradations, the unified models [3, 14-16] offer a natural attempt, emerging as a leading approach in image restoration. In particular, since IPT [15] first demonstrates the potential of transformer architectures for low-level vision tasks, the development of transformer-based models [3, 17] has accelerated. However, significant challenges remain, particularly in addressing complex illumination and coupled degradation conditions. Many image restoration models, such as those based on transformers, lack adaptive mechanisms designed for low-light environments, often resulting in amplified degradation and the introduction of new artifacts. Specifically, two key issues persist: 1) the difficulty of exploiting effective adaptive strategies for intricate illumination scenarios and handling coupled degradation factors, and 2) parameter inefficiency, which presents a critical bottleneck for practical deployment in downstream vision tasks. + +In this paper, we extend the existing unified encoder-decoder framework by integrating the recently proposed Receptance Weighted Key Value (RWKV) structure [18, 19], aiming to provide effective solutions for addressing dynamically coupled degradations in low-light conditions from a multi-state perspective. Specifically, we introduce both intra- and inter-stage states to perceive and analyze diverse degradations across multiple stages, enabling a more nuanced restoration throughout the processing pipeline. To achieve this, we first propose luminance-adaptive normalization (LAN), which dynamically adjusts normaliza + +tion parameters based on rich inter-stage states, facilitating adaptive, scene-aware luminance modulation. Furthermore, we employ an exponential moving average (EMA) approach to aggregate multiple intra-stage states, effectively capturing subtle variations and mitigating the information loss inherent in single-state mechanisms. Instead of constraining our model to predefined degradation types, we exploit the relationships between multi-state representations to address spatially varying and coupled degradations, thus enabling flexible and dynamic restoration. Additionally, we introduce a state-aware selective fusion (SSF) module, which dynamically aligns and integrates multi-state features across encoder stages, selectively fusing contextual information to enhance restoration quality. As shown in Fig. 1, our model consistently outperforms existing models across various degradation scenarios. + +The main contributions of this paper are as follows: + +- We introduce the unified RWKV model with a novel multi-state perspective, enabling flexible and effective restoration of dynamically coupled degradations in low-light conditions. +- We propose LAN to dynamically adjust normalization parameters based on inter-stage states, allowing for adaptive, scene-aware luminance modulation. +- We aggregate multiple intra-stage states using an EMA-based approach to capture subtle degradation variations, while mitigating the long-range information loss. +- We develop the SSF module, which dynamically aligns and integrates multi-state features across encoder stages, facilitating selective fusion of contextual information. +- Our URWKV model outperforms existing models across diverse benchmarks, effectively handling dynamically coupled degradations with fewer parameters and lower computational cost. + +# 2. Related Work + +# 2.1. Low-light Image Enhancement + +Compared to traditional methods relying on histogram equalization [20-22] or Retinex-based enhancement [23-25], deep learning approaches have gained prominence in low-light image enhancement for their ability to handle complex degradations. Retinex-based models, in particular, have evolved through the integration of advanced architectures such as CNNs [7, 8], Transformers [9], and Mamba networks [11], enabling more precise illumination-reflectance decomposition. Leveraging the long-range dependency modeling capabilities of Transformers, several methods [1, 10] have introduced modifications to overcome the limitations of Transformers in low-light scenarios, achieving improved computational efficiency and enhanced local modeling capabilities. In addition, recognizing that noise characteristics vary spatially, Xu et al. [12] propose to + +![](images/c506fea410fc5e2785ecd9faa72ffbf6bfbb8afc59acb2d23acc1ab36f95ff18.jpg) +(a) Framework + +![](images/94c74fa9db9e91bb5f9f7701a8d6a8daa02ba5cc1e7c821903a801e8b923ebe1.jpg) +Squared ReLU + +![](images/de7a542251406ef978dad269c459a3d5c43318aab7fdc74d3e8982b6aff6cdc6.jpg) +Concatenation + +![](images/f97b41b07470492ea5ea202a3fba61893fb2f7f3aaac2d440cf96177c32201bd.jpg) +Sigmoid + +![](images/3147b06902360a624e91930a280d99747ac4f27eac3d7f2681d03f2059671ec2.jpg) + +![](images/a8c7bab8368351d721f368ba6e4855d2e4b6c0180776e3a78acfaced5d9bbec6.jpg) +Element-wise Multiplication +(b) URWKV Block with Multi-state Spatial and Channel Mixing +Figure 2. Overview of the URWKV model. The core URWKV block is integrated into an encoder-decoder framework, featuring a multi-state spatial mixing sub-block and a multi-state channel mixing sub-block. The luminance-adaptive normalization (LAN) incorporates both the current input state $X_{t}$ and multiple inter-stage states $M_{i}$ . Additionally, the multi-state quad-directional token shift (SQ-Shift) aggregates both current input state and multiple intra-stage states $H_{i}$ ( $H_{S_{i}}$ for spatial mixing and $H_{C_{i}}$ for channel mixing, respectively). The state-aware selective fusion (SSF) module further aggregates rich contextual information across encoder stages. + +![](images/297f6290ee9be3c5ad0dad1a19dd303e8b948aea7ababdc1c80cbc1fd042180b.jpg) +Element-wise Addition + +separate signal from noise during the enhancement process, enabling more effective brightening while minimizing noise amplification. Furthermore, frequency-domain approaches [26, 27] also contribute to LLIE by decomposing features across distinct frequency bands, proving effective in preserving fine details and reducing noise. + +Considering blur as a frequent coupled degradation in low-light scenes, one that conventional LLIE models inadequately address, Zhou et al. [13] and Li et al. [2] propose a specific sequential encoder-decoder architecture and a parallel heterogeneous decoupling framework, respectively. Although these LLIE-deblur models show effectiveness within certain constraints, further exploration is warranted to handle the dynamically coupled degradations typical in real-world scenarios. + +# 2.2. Unified Model for Image Restoration + +With the emergence of deep learning, unified approaches have driven substantial progress, enabling the handling of multiple restoration tasks within a single framework. Compared to CNN-based methods [14, 28], which primarily focus on multi-scale feature extraction and local context enhancement, Transformer-based methods [3, 15, 29, 30] extend relational modeling capabilities by capturing long-range dependencies, leading to improved restoration quality across diverse degradation types. However, the application of Transformers in image restoration often entails high computational costs and memory requirements, which can + +hinder their efficiency. In addition, frequency-domain models [31, 32] provide another unified approach by decomposing images into multiple frequency bands, which enables focused noise reduction while preserving fine details across tasks. Recent innovations also utilize Mamba networks [16] to achieve effective image restoration by efficiently balancing local and global feature representations. + +While current unified models perform effectively under normal lighting conditions, they encounter significant challenges in low-light environments, specifically involving complex lighting variations and coupled degradations. + +# 3. Method + +# 3.1. Overview + +As illustrated in Fig. 2, the URWKV framework follows the typical encoder-decoder architecture, with both the encoder and decoder comprising three stages. In the encoder, each stage is composed of $N_{1}$ URWKV blocks followed by a down-sampling layer, while each stage in the decoder contains $N_{2}$ URWKV blocks and an up-sampling layer. Additionally, each down-sampling or up-sampling layer includes convolutional operations and sampling techniques to adjust either the channel or spatial dimensions. + +To effectively handle dynamically coupled degradations, each URWKV block operates as an independent and progressive processing stage, with each stage's output (an implicit restored result, serving as a 'state') encapsulat- + +![](images/da414007927556da9a4b014cfa97d7c676311ccb7d3d5955bbe129d1df082ada.jpg) +(a) Luminance-adaptive Normalization + +![](images/b6821a700fc76de291b57e101098d4572a8346af8ad168f077a1cab2c0b94a6e.jpg) +(b) Multi-kernel aggregation of inter-stage states +Figure 3. Illustration of luminance-adaptive normalization (LAN). LAN integrates inter-stage states $M_{i}$ throughout the restoration process, facilitating scene-aware luminance modulation. + +ing stage-specific contextual semantics and restoration cues (e.g., luminance or content) for subsequent stages. + +Specifically, the URWKV block builds on the strengths of original RWKV model [19], while enhancing its capabilities through two tailored sub-blocks: a multi-state spatial mixing sub-block and a multi-state channel mixing sub-block. The former captures spatial dependencies across multiple intra-stage states to enrich spatial information, while the latter enhances feature representations through multi-state channel-wise interactions. Here, the intra-stage states $H_{S_i}$ or $H_{C_i}$ are updated following the SQ-Shift in each sub-block, respectively, where $i \in \{1, \dots, N_1\}$ (or $i \in \{1, \dots, N_2\}$ ). To ensure efficient information transfer between encoder and decoder stages without introducing degradation, we incorporate the state-aware selective fusion (SSF) module between these stages. The SSF dynamically aggregates multi-state features across all encoder stages, selectively propagating rich contextual information into the decoder. + +# 3.2. URWKV with Multi-state Perspective + +Original RWKV models [18, 19] employ a single-state token shift mechanism, where each token's relationship relies solely on the current and preceding state. This design introduces two primary limitations: First, it struggles to retain long-range dependencies, causing early information to gradually fade during restoration. Second, it fails to effectively capture complex interdependencies among degraded features, as a single state is insufficient to represent intricate coupled degradations. To this end, we propose URWKV devised with the multi-state perspective. Specifically, the URWKV block tailors the luminance-adaptive normalization (LAN) by incorporating multiple inter-stage states to adapt URWKV to low-light scenarios. Additionally, it integrates multiple intra-stage states to facilitate more extensive interactions across long-range dependencies. + +Luminance-adaptive Normalization (LAN). Luminance distortion is a major factor leading to other coupled degradations in low-light images. Inspired by the hu + +man eye's ability to adaptively adjust pupil size in response to varying luminance [33, 34], LAN leverages inter-stage states throughout the restoration process (comprising six stages) to enable scene-aware luminance modulation. This design overcomes limitation of LayerNorm widely used in existing models [3, 17], which relies on fixed parameters $\gamma$ for scaling the learned representation. + +As illustrated in Fig. 3, given the current state input $X_{t}\in \mathbb{R}^{C_{t}\times H_{t}\times W_{t}}$ at stage $T$ , we first extract 1D luminance vectors from both $X_{t}$ and each historical stage output $M_{i}\in \mathbb{R}^{C_{i}\times H_{i}\times W_{i}}$ , using the global average pooling (GAP) operation. Here, each $M_{i}$ is obtained from the output of each stage's final URWKV block. $t\in \{1,\dots,N_1\}$ (or $t\in \{1,\dots,N_2\}$ ) and $i\in \{1,\dots,T - 1\}$ . The luminance vectors are then zero-padded to match the dimensionality of the vector with the maximum channel size $C_{max}$ , while preserving the semantic integrity of luminance vector from each individual state. These zero-padded luminance vectors, denoted as $X_{k}^{1D}\in \mathbb{R}^{C_{max}}$ , for $k\in \{1,\dots,T\}$ , are then stacked along the state dimension. Next, a multi-kernel aggregation layer is applied to analyze the stacked 2D luminance map along the channel dimension. Specifically, the aggregation layer uses 1D convolutional kernels of sizes $1\times T$ , $3\times T$ , and $5\times T$ to adaptively capture both local and broader contextual illuminance variations across multiple states. The resulting luminance maps $X_{r}^{2D}\in \mathbb{R}^{T\times C_{max}}$ are expressed as: + +$$ +X _ {r} ^ {2 D} = \operatorname {C o n v} _ {r} ^ {1 D} \left(\left[ X _ {1} ^ {1 D},..., X _ {T} ^ {1 D} \right]\right), r \in \{1, 3, 5 \}, \tag {1} +$$ + +where $\mathrm{Conv}_r^{1D}$ denotes the set of 1D convolutional kernels, and $[\cdot ,\cdot ]$ denotes the stacking operation. + +Subsequently, the output of each convolutional kernel is concatenated along the channel dimension and projected through a $1\times 1$ convolution, resulting in the aggregated feature map $X_{\mathrm{agg}}\in \mathbb{R}^{C_t}$ . A multi-layer perceptron (MLP) along with a tanh activation function are then employed to predict the luminance modulator $\Delta \gamma_{t}$ . The luminance modulator is computed as: + +$$ +\Delta \gamma_ {t} = \tanh (\mathrm {M L P} \left(X _ {\text {a g g}}\right)), \tag {2} +$$ + +where $\tanh (\cdot)$ represents the tanh activation function. + +Once $\Delta \gamma_{t}$ is predicted, the scaling parameter $\gamma$ can be updated as $\hat{\gamma}_{t} = \gamma +\Delta \gamma_{t}$ . Ultimately, the input feature $X_{t}$ at stage $T$ can be adaptively normalized with LAN, expressed as: + +$$ +X _ {t} ^ {L A N} = \frac {X _ {t} - \mu_ {t}}{\sigma_ {t}} \odot \hat {\gamma} _ {t} + \beta , \tag {3} +$$ + +where $\mu_t$ and $\sigma_t$ represent the mean and standard deviation of the input feature $X_t$ , respectively. $\beta$ is a shift parameter. + +Multi-state Quad-directional Token Shift (SQ-Shift). Original quad-directional token shift (Q-shift) [19] effectively captures spatial dependencies among neighboring tokens. However, it neglects the degradation relations between states, making it difficult to maintain consistent degradation restoration across consecutive states. To address this limitation, we aggregate multi-state features to extend Q-Shift, allowing the model to capture long-range dependencies across multiple states while preserving local spatial interactions. Specifically, before applying Q-shift, we employ the exponential moving average (EMA) approach to aggregate the current state with all preceding states within the same stage, enabling richer interactions with historical information. For instance, given the state feature $X_{t}^{LAN} \in \mathbb{R}^{C \times H \times W}$ normalized by LAN, and multiple historical intra-stage state features $H_{i}, i \in \{1, \dots, t - 1\}$ , we define the multi-state aggregation MSA as follows: + +$$ +\operatorname {M S A} \left(X _ {t} ^ {L A N}\right) = \alpha \odot X _ {t} ^ {L A N} + (1 - \alpha) \odot \operatorname {M S A} \left(H _ {t - 1}\right), \tag {4} +$$ + +where $\alpha$ is a decay factor that controls the weight between the current state and the previously aggregated state. + +Multi-State Spatial and Channel Mixing. For each input feature in the multi-state spatial mixing sub-block or multi-state channel mixing sub-block, we first apply LAN and SQ-Shift, resulting in features $X_{s}$ for spatial mixing or $X_{c}$ for channel mixing. In the spatial mixing process, $X_{s}$ undergoes three parallel linear projections to generate the matrices for receptance $R_{s}$ , key $K_{s}$ , and value $V_{s}$ . Following the computational approach of [19], we employ the bidirectional attention mechanism Bi-WKV to compute the global attention weights $wkv$ between $K_{s}$ and $V_{s}$ , which are subsequently modulated by the sigmoid-activated receptance $R_{s}$ to control the output. The output feature $O_{s}$ is therefore represented as: + +$$ +O _ {s} = \left(\sigma \left(R _ {s}\right) \odot w k v\right) W _ {O _ {s}}, \tag {5} +$$ + +where $wkv = \mathrm{Bi - WKV}(K_{\mathrm{s}},V_{\mathrm{s}})$ . $\sigma$ represents the sigmoid function. $W_{O_s}$ is learnable parameter of projection layer. + +![](images/17cc601dad29b54cc16d8c3c43b0df24d4f0f5dbd1f57be0414a1594f9bce023.jpg) +(a) State-aware Selective Fusion + +![](images/fe1e1c9639da0d9e46e7a449a121f9202706f24f6c75dd64998fbb794f8e6c98.jpg) +(c) State Alignment +Figure 4. Illustration of the state-aware fusion (SSF) module. + +For channel mixing, $X_{c}$ undergoes two parallel linear projections to derive the receptance matrix $R_{c}$ and key matrix $K_{c}$ . Then, the output feature $O_{c}$ for channel mixing can be formulated as: + +$$ +O _ {c} = \left(\sigma \left(R _ {c}\right) \odot \operatorname {S q u a r e d R e L U} \left(K _ {c}\right) W _ {V _ {c}}\right) W _ {O _ {c}}, \tag {6} +$$ + +where $W_{V_c}$ and $W_{O_c}$ are learnable parameters in the two projection layers. + +# 3.3. State-aware Selective Fusion (SSF) + +Existing models typically reintroduce early-stage features using navie skip connections (e.g., concatenation or addition). However, they are prone to two major issues, particularly under low-light conditions: (1) the propagation of noise and irrelevant information from the encoder to the decoder [1], and (2) the semantic gap between stages [36], which hinders effective feature fusion. + +Rather than adopting the navie or parameter-heavy approaches [3, 17, 36], we propose the lightweight and effective state-aware selective fusion (SSF) module. As illustrated in Fig. 4, the SSF module focuses on capturing the degradation restoration relationships between multistate features to predict the spatial guiding weight, which selectively refines the transfer of encoder features to the decoder. Given three encoder outputs $E_{i} \in \mathbb{R}^{C_{i} \times H_{i} \times W_{i}}$ obtained from $M_{i}$ , $i \in \{1,2,3\}$ , we first apply a channel-wise mean operation to collapse the channel dimension, resulting in $E_{i}^{M} \in \mathbb{R}^{1 \times H_{i} \times W_{i}}$ , which reduces inter-channel semantic interference. Next, the adaptive alignment operation is performed on $E_{i}^{M}$ to adjust the spatial dimensions of each feature according to the target dimensions of the decoder at the current fusion stage. The aligned features $E_{i}^{A}$ are then stacked along the channel dimension to produce the concatenated feature map $E^{A} \in \mathbb{R}^{3 \times H_{d} \times W_{d}}$ , where $H_{d}$ and $W_{d}$ are the height and width of the decoder's feature. This process is formally expressed as: + +$$ +E ^ {A} = \operatorname {S t a c k} (\operatorname {A d a A l i g n} (\operatorname {A v g} (E _ {i}))), i \in \{1, 2, 3 \}, \quad (7) +$$ + +Table 1. Quantitative comparison results on the LOL-v2-real, LOL-v2-syn, SID, SMID, SDSD-indoor, and SDSD-outdoor datasets. Joint LLIE and deblurring models are marked with †, unified models are marked with *, and the rest represent LLIE models. + +
MethodsLOL-v2-realLOL-v2-synSIDSMIDSDSD-indoorSDSD-outdoorComplexity
PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑FLOPs (G)Params (M)
KinD [7]14.740.64113.290.57818.020.58322.180.63421.950.67221.970.65434.998.02
EnlightenGAN [35]18.230.61716.570.73417.230.54322.620.67420.020.60420.100.61661.01114.35
SNR-Net [12]21.480.84924.140.92721.350.55028.520.80726.130.81519.220.65726.354.01
FourLLIE [26]22.350.84724.650.91918.420.51325.640.75024.740.82624.670.787-0.12
UHDFour [27]21.790.85423.600.91322.740.65128.180.80929.130.89523.280.77157.4215.90
LLFormer [10]21.630.80824.130.90822.830.65628.420.79426.820.82529.180.84122.5224.52
Retinexformer [9]22.790.83925.670.92824.440.68029.150.81529.780.89529.830.87815.571.61
BiFormer [1]22.670.86324.870.92122.490.63828.240.81228.730.89926.130.8415.930.83
RetinexMamba [11]22.450.84325.880.93322.450.65628.620.80928.440.89428.520.85934.7524.1
LEDNet†[13]19.020.83524.790.93121.470.63828.420.80727.290.87626.660.85038.577.41
PDHAT†[2]20.160.84124.940.93721.930.65329.190.81726.370.88428.190.862208.197.83
Uformer* [17]18.820.77119.660.87118.540.57727.200.79223.170.85923.850.74812.005.29
MIRNet* [14]22.680.82825.050.92321.360.63226.210.76928.640.88828.990.869785.0031.76
Restormer* [3]18.600.78921.410.83122.010.64528.580.80628.490.89227.990.868140.9926.11
MambaIR* [16]20.450.84525.650.93922.020.65828.410.80525.140.87627.530.85160.664.30
URWKV (Ours)*23.110.87426.360.94423.110.67329.440.82631.240.91129.990.88718.342.25
+ +where Avg denotes the channel-wise mean operation. AdaAlign combines both upsampling and downsampling to adaptively align multiple states with the target state. Stack is the operation that stacks the aligned features along the channel axis. + +To facilitate more nuanced fusion of local details and broader context, we incorporate an inception block [37] with multiple convolutional filters $(1\times 1,3\times 3,$ and $5\times 5)$ to aggregate degradation patterns across various receptive fields. The aggregated feature $E_{agg}$ is then projected via a convolutional layer, followed by a sigmoid function to produce the spatial weight. Therefore, taking the first stage of the decoder as an example, the naive concatenation-based skip connection for the URWKV block in the decoder, following the SSF module, can be extended as follows: + +$$ +D _ {1} ^ {\prime} = \left(\left[ W _ {s} \odot E _ {3}, D _ {1} \right]\right) W _ {p} \tag {8} +$$ + +where $W_{s}$ is the spatial weight predicted by the SSF module, $E_{3}$ is the encoder feature, and $D_{1}$ is the input feature to the first stage of the decoder. $\odot$ denotes element-wise multiplication, and $W_{p}$ represents a convolutional layer applied to refine the features. $[\cdot ,\cdot ]$ denote the concatenate operation. + +# 4. Experiments + +# 4.1. Experimental Settings + +We implement our model in PyTorch on a workstation equipped with an NVIDIA Tesla A40 GPU. The number of URWKV blocks is set to $N_{1} = 3$ and $N_{2} = 2$ , while the number of input channels is set to $C = 32$ . The decay factor $\alpha$ is set to 0.5. During training, we use a unified loss function for all datasets, as in [1], which combines L1 Loss, + +SSIM Loss, and Perceptual Loss. The Adam optimizer [38] is employed with $\beta_{1} = 0.9$ and $\beta_{2} = 0.99$ , and the learning rate is initialized at $2 \times 10^{-4}$ and decays gradually to $10^{-6}$ using cosine annealing decay. Following [1, 10], we apply augmentation techniques, such as flipping and rotation, to each image pair. During testing, our URWKV model processes input images of arbitrary shape directly, without the need for cropping or resizing. For evaluation, we employ commonly-used PSNR [39] and SSIM [40] metrics to calculate images in the RGB color space across all datasets. + +# 4.2. Main Results + +To assess the performance of various models in scenarios involving dynamically coupled degradations, we compare the LLIE models, LLIE-deblur models, and unified models across 8 typical low-light benchmarks. + +LOL. As shown in Table 1, our URWKV demonstrates superior performance compared to LLIE models on LOLv2-real [41], outperforming Retinexformer by 0.32 dB in PSNR and 0.035 in SSIM, respectively. Similarly, on LOLv2-syn [41], our model outperforms RetinexMamba by 0.48 dB in PSNR and 0.011 in SSIM. Notably, LLIE-deblur models [2, 13] and other unified models [3, 14, 16] generally underperform on both datasets, with substantial performance gaps between them and our model. + +SID and SMID. As shown in Table 1, our UR-WKV model delivers superior performance, closely trailing Retinexformer on SID [42], while outperforming it by 0.29 dB in PSNR on SMID [42]. In contrast, LLIE-deblur models exhibit inconsistent performance across the datasets. For instance, while PDHAT ranks second on SMID, it significantly underperforms on SID, achieving a PSNR of just + +![](images/886ba7b69e60c419b10d4989571a9d09570d7b3eb7ad0241a0945d26d5297592.jpg) +Figure 5. Visual comparison of state-of-the-art models on the SID dataset (top) and the SMID dataset (bottom). + +![](images/a05e1f1bef375b59e77999d908dcfa5d1b7a6043de9e6bfe59d419ec048d0090.jpg) + +![](images/e6c0bfd964b3a757c6ddbb480f90954c1a4e138d640cedfd318170312ecea311.jpg) + +![](images/7bd1cc1bfc5006a005c7d1dac10c4671d890d947b8df48b309f01321d7a15b55.jpg) + +![](images/6461dd32c502b1a38d651f6570d1f165a9641a4ec5bf171af68475d661589c66.jpg) + +![](images/8810ec3c47ad147d7173a76a86ed5783e9d4f2aaac910fdad4ca4aeaef928755.jpg) + +![](images/2727e65419c06056dd0ae2ef9a8f74e705dedc31b8827668579f898be891ae3e.jpg) + +![](images/c5ed6aba216577cf3d04fb777f918cf1d7276c9579a6d87c167ed506dae50421.jpg) +Figure 6. Visual comparisons on SDSD-indoor dataset. + +![](images/96103177d1aac695356b72edb2d217cba9d1641f4e80c6bb8c68ac315a6e3be2.jpg) + +![](images/6f7ca234c2a4cd38834d0efdb1c5e692939069f958f3879088cb64c916b2d88c.jpg) + +21.93 dB, which is 1.18 dB lower than our model. Visual comparisons in Fig. 5 also demonstrate the excellent denoising capability of our model when enhancing brightness. + +SDSD. Results depicted in Table 1 indicate that even some recent LLIE models [1, 26] still encounter challenges on SDSD [43]. In contrast, our URWKV model delivers state-of-the-art performance across both subsets. Moreover, compared to unified models like Restormer, URWKV outperforms it by $2.75\mathrm{dB}$ in PSNR on SDSD-indoor and by $2.00\mathrm{dB}$ in PSNR on SDSD-outdoor. As demonstrated in Fig. 6, our method preserves high-fidelity restoration in the scenario with strong light-dark contrast, whereas other models (e.g. PDHAT) generate unexpected artifacts. + +FiveK. The MIT-Adobe FiveK dataset [44] challenges models with the intricate task of color restoration in lowlight and underexposed images. As reported in Table 2, our model achieves the highest PSNR of $26.08\mathrm{dB}$ on FiveK dataset, surpassing LLFormer by $0.48\mathrm{dB}$ and Retinexformer by $1.50\mathrm{dB}$ . Furthermore, URWKV outperforms the unified model, Restormer, in SSIM by 0.007. These notable improvements underscore URWKV's proficiency in restoring both structural integrity and aesthetic quality. + +LOL-blur. The LOL-blur [13] dataset couples varying degrees of low-light and motion blur, which has often been underrepresented in other LLIE datasets. As shown in Table 3, conventional LLIE models typically struggle to address this coupled degradation, while unified models also fail to outperform specialized LLIE-deblur models. In contrast, our URWKV model achieves the highest performance in this scenario, with a PSNR of 27.27 dB and SSIM of 0.890, outperforming the second-best LLIE-Deblur model + +Table 2. Experimental results on MIT-Adobe FiveK dataset. + +
MethodsEnlightenGAN [35]RUAS [45]RetinexNet [8]SNR-Net [12]LLFormer [10]BiFormer [1]
PSNR ↑13.2620.9116.3224.3925.6025.51
SSIM ↑0.7450.8630.7820.8950.9170.929
MethodsRetinexformer [9]LEDNet [13]PDHAT [2]Restormer [3]MIRNet [14]URWKV (Ours)
PSNR ↑24.5824.3425.0025.4823.7326.08
SSIM ↑0.8860.9150.9290.9290.8950.936
+ +Table 3. Experimental results on LOL-blur dataset. + +
MethodsFourLLIE [26]UHDFour [27]LLFormer [10]Retinexformer [9]BiFormer [1]GLARE [46]
PSNR ↑19.8125.3424.5525.2524.6723.26
SSIM ↑0.6830.8290.7850.8210.8130.690
MethodsLEDNet [13]PDHAT [2]Restormer [3]MIRNet [14]MambaIR [16]URWKV (Ours)
PSNR ↑26.0626.7126.3823.9926.2827.27
SSIM ↑0.8460.8790.8600.7740.8480.890
+ +PDHAT by 0.56 dB in PSNR and 0.011 in SSIM. Examples in Fig. 7 vividly demonstrate the potential of our model in restoring coupled low-light and blur images. + +# 4.3. Ablation Study + +We conduct several ablation studies on the real-world dataset (i.e., LOL-v2-real) to analyze the impact of various components in our model. + +Effectiveness of LAN and SSF modules. To validate the individual and combined contributions of the LAN and SSF modules, we evaluate four model variants. As shown in Table 4, models incorporating either LAN or SSF individually show a clear improvement over the baseline (first row in Table 4). Among these, the model with LAN alone achieves a more pronounced gain, with PSNR increasing by $1.38\mathrm{dB}$ over the baseline, highlighting the critical role of LAN in dynamically adapting to complex luminance variations. Moreover, our final model, which combines both LAN and SSF, incurs only a minor increase in parameters (from 1.64M to 2.25M) and a minimal rise in FLOPs (from 18.25G to 18.34G), yet delivers a significant performance boost compared to baseline. The feature transformation vi + +![](images/179d7a682ac682a4cdf73291c73ac81b0c0d28c915755566dc436843f8ca3f4e.jpg) +Figure 7. Visual comparison of state-of-the-art models on LOL-blur dataset. + +![](images/5bdae92335af02b211aed70c9727ef64651c3a7eb7dea982b33b920a03db634d.jpg) +Figure 8. Top: LAN adaptively enhances the luminance, making details clearer. Bottom: SSF helps mitigate introducing degradation, obtaining fused feature $(D_3^{\prime})$ based on $E_{1}$ , $E_{2}$ and $E_{3}$ . + +sualizations presented in Fig. 8 also validate the inherent advantages of LAN and SSF. + +Multi-state Quad-directional Token Shift. Our propose multi-state aggregation strategy enhances the original Q-shift, allowing for a more comprehensive understanding of degradation representation. As shown in Table 5, our results indicate that single-state aggregation alone (Table 5a) achieves performance comparable to the model using Q-shift alone (Table 5c). In contrast, multi-state aggregation alone (Table 5b) consistently demonstrates superior performance, attaining a $0.13\mathrm{dB}$ higher PSNR. When combining single-state aggregation with Q-shift (Table 5d), the model shows an improvement in PSNR. However, it still falls short in SSIM by 0.003 compared to multi-state aggregation alone. Ultimately, our final model (Table 5e), which integrates both Q-shift and multi-state aggregation, demonstrates substantial advantages over all other variants. + +SSF vs. Navie Skip Connections. The ablation results in Table 6 highlight the effectiveness of SSF compared to naive skip connections, including simple addition (Add) and concatenation (Cat). When using multi-state encoder features, noise from earlier stages and semantic gaps between states lead to performance degradation in both PSNR and SSIM. In contrast, our SSF module, based on state-aware fusion strategy, avoids directly merging multistate features and predicts a guiding weight to selectively fuse useful patterns, mitigating issues of noise propagation and semantic misalignment. Notably, the "Multi-state+SSF" (i.e., our URWKV) yields significant improvements in PSNR and SSIM over "Single-state+Cat" without + +Table 4. Ablation study on LAN and SSF modules. + +
LANSSFPSNR ↑SSIM ↑ParamsFLOPs
21.330.8561.64M18.25G
✓21.400.8611.65M18.29G
✓22.710.8692.25M18.30G
✓✓23.110.8742.25M18.34G
+ +Table 5. Ablation study on multi-state quad-directional token shift. + +
#Single-stateMulti-stateQ-shiftPSNR ↑SSIM ↑
(a)✓22.120.861
(b)✓22.250.870
(c)✓22.190.862
(d)✓✓22.290.867
(e)✓✓23.110.874
+ +Table 6. Impacts of SSF and navie skip connections (Skip Conn.) with different state numbers (State No.). + +
State No.Skip Conn.PSNR ↑SSIM ↑ParamsFLOPs
Single-stateAdd22.660.8721.86M14.68G
Single-stateCat22.710.8692.25M18.30G
Multi-stateCat21.830.8592.50M23.89G
Multi-stateSSF23.110.8742.25M18.34G
+ +increasing the model's parameter, demonstrating both enhanced performance and computational efficiency. + +# 5. Conclusion + +To address the shortcomings of existing models, we propose URWKV with multi-state perspective, enabling flexible and effective restoration across dynamically coupled degradation scenarios. Specifically, we introduce three key innovations: luminance-adaptive normalization (LAN) for adaptive, scene-aware luminance modulation; an exponential moving average approach to aggregate intra-stage states for spatial-level token-shift; and the state-aware selective fusion (SSF) module for dynamic alignment and integration of multi-state features across encoder stages. In comparison to existing models, our model demonstrates superior performance while requiring significantly fewer parameters and computational resources. + +Acknowledgments. This work was supported in part by the National Natural Science Foundation of China under Grant 62471142, in part by the Natural Science Foundation of Fujian Province, China under Grants 2023J01067, in part by Industry-Academy Cooperation Project under Grant 2024H6006, in part by the Collaborative Innovation Platform Project of Fuzhou City under Grant 2023-P-002, in part by the Key Technology Innovation Project for Focused Research and Industrialization in the Software Industry of Fujian Province, and in part by the Key Research and Industrialization Project of Technological Innovation in Fujian Province under Grant 2024XQ002. + +# References + +[1] Rui Xu, Yuezhou Li, Yuzhen Niu, Huangbiao Xu, Yuzhong Chen, and Tiesong Zhao. Bilateral interaction for local-global collaborative perception in low-light image enhancement. IEEE Transactions on Multimedia, 2024. 1, 2, 5, 6, 7 +[2] Yuezhou Li, Rui Xu, Yuzhen Niu, Wenzhong Guo, and Tiesong Zhao. Perceptual decoupling with heterogeneous auxiliary tasks for joint low-light image enhancement and deblurring. IEEE Transactions on Multimedia, pages 6663-6675, 2024. 1, 2, 3, 6, 7 +[3] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. 1, 2, 3, 4, 5, 6, 7 +[4] Andong Lu, Zhang Zhang, Yan Huang, Yifan Zhang, Chenglong Li, Jin Tang, and Liang Wang. Illumination distillation framework for nighttime person re-identification and a new benchmark. IEEE Transactions on Multimedia, pages 1–14, 2023. 1 +[5] Xiaopeng Li and Xiaojie Guo. SPN2D-GAN: Semantic prior based night-to-day image-to-image translation. IEEE Transactions on Multimedia, pages 1-15, 2022. 1 +[6] Xuechen Bai, Xinghua Li, Jianhao Miao, and Huanfeng Shen. Making the rarth clear at night: A high-resolution nighttime light image deblooming network. IEEE Transactions on Geoscience and Remote Sensing, 61:1-13, 2023. 1 +[7] Yonghua Zhang, Jiawan Zhang, and Xiaojie Guo. Kindling the darkness: A practical low-light image enhancer. In Proceedings of the ACM International Conference on Multimedia, pages 1632-1640, 2019. 1, 2, 6 +[8] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. In Proceedings of the British Machine Vision Conference, 2018. 1, 2, 7 +[9] Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang. Retinexformer: One-stage retina-based transformer for low-light image enhancement. In Proceedings of the IEEE International Conference on Computer Vision, pages 12504-12513, 2023. 1, 2, 6, 7 + +[10] Tao Wang, Kaihao Zhang, Tianrun Shen, Wenhan Luo, Bjorn Stenger, and Tong Lu. Ultra-high-definition low-light image enhancement: A benchmark and transformer-based method. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 2654-2662, 2023. 1, 2, 6, 7 +[11] Jiesong Bai, Yuhao Yin, and Qiyuan He. Retinexmamba: Retinex-based mamba for low-light image enhancement. arXiv preprint arXiv:2405.03349, 2024. 1, 2, 6 +[12] Xiaogang Xu, Ruixing Wang, Chi-Wing Fu, and Jiaya Jia. SNR-aware low-light image enhancement. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 17714-17724, 2022. 2, 6, 7 +[13] Shangchen Zhou, Chongyi Li, and Chen Change Loy. LEDNet: Joint low-light enhancement and deblurring in the dark. In Proceedings of the European Conference on Computer Vision, pages 573-589, 2022. 2, 3, 6, 7 +[14] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Learning enriched features for real image restoration and enhancement. In Proceedings of the European Conference on Computer Vision, pages 492-511, 2020. 2, 3, 6, 7 +[15] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12299-12310, 2021. 2, 3 +[16] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. MambaIR: A simple baseline for image restoration with state-space model. In Proceedings of the European Conference on Computer Vision, pages 222-241, 2025. 2, 3, 6, 7 +[17] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 17683-17693, 2022. 2, 4, 5, 6 +[18] Bo Peng, Eric Alcaide, Quentin Anthony, Alon Albalak, Samuel Arcadinho, Stella Biderman, Huanqi Cao, Xin Cheng, Michael Chung, Matteo Grella, et al. RWKV: Reinventing rnns for the transformer era. arXiv preprint arXiv:2305.13048, 2023. 2.4 +[19] Yuchen Duan, Weiyun Wang, Zhe Chen, Xizhou Zhu, Lewei Lu, Tong Lu, Yu Qiao, Hongsheng Li, Jifeng Dai, and Wenhai Wang. Vision-RWKV: Efficient and scalable visual perception with RWKV-like architectures. arXiv preprint arXiv:2403.02308, 2024. 2, 4, 5 +[20] J Alex Stark. Adaptive image contrast enhancement using generalizations of histogram equalization. IEEE Transactions on Image Processing, 9:889-896, 2000. 2 +[21] Mohammad Abdullah-Al-Wadud, Md Hasanul Kabir, M Ali Akber Dewan, and Oksam Chae. A dynamic histogram equalization for image contrast enhancement. IEEE Transactions on Consumer Electronics, 53:593-600, 2007. +[22] Turgay Celik and Tardi Tjahjadi. Contextual and variational contrast enhancement. IEEE Transactions on Image Processing, 20:3431-3441, 2011. 2 +[23] Xueyang Fu, Yinghao Liao, Delu Zeng, Yue Huang, Xiao-Ping Zhang, and Xinghao Ding. A probabilistic method for + +image enhancement with simultaneous illumination and reflectance estimation. IEEE Transactions on Image Processing, 24:4965-4977, 2015. 2 +[24] Mading Li, Jiaying Liu, Wenhan Yang, Xiaoyan Sun, and Zongming Guo. Structure-revealing low-light image enhancement via robust retinae model. IEEE Transactions on Image Processing, 27:2828-2841, 2018. +[25] Daniel J Jobson, Zia-ur Rahman, and Glenn A Woodell. A multiscale retina for bridging the gap between color images and the human observation of scenes. IEEE Transactions on Image processing, 6:965–976, 1997. 2 +[26] Chenxi Wang, Hongjun Wu, and Zhi Jin. FourLLIE: Boosting low-light image enhancement by fourier frequency information. In Proceedings of the ACM International Conference on Multimedia, pages 7459-7469, 2023. 3, 6, 7 +[27] Chongyi Li, Chun-Le Guo, Man Zhou, Zhexin Liang, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Embedding fourier for ultra-high-definition low-light image enhancement. Proceedings of the International Conference on Learning Representations, 2023. 3, 6, 7 +[28] Yuning Cui, Wenqi Ren, Xiaochun Cao, and Alois Knoll. Revitalizing convolutional network for image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 9423-9438, 2024. 3 +[29] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image restoration using swin transformer. In Proceedings of the IEEE International Conference on Computer Vision, pages 1833-1844, 2021. 3 +[30] Akshay Dudhane, Syed Waqas Zamir, Salman Khan, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Burstormer: Burst image restoration and enhancement transformer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5703-5712, 2023. 3 +[31] Man Zhou, Jie Huang, Chun-Le Guo, and Chongyi Li. Fourmer: An efficient global modeling paradigm for image restoration. In Proceedings of the International Conference on Machine Learning, pages 42589-42601, 2023. 3 +[32] Yuning Cui, Wenqi Ren, Xiaochun Cao, and Alois Knoll. Image restoration via frequency selection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3 +[33] Francisco Lara, Paula Bernal-Molina, Vicente Fernandez-Sanchez, and Norberto López-Gil. Changes in the objective amplitude of accommodation with pupil size. Optometry and Vision Science, 91(10):1215-1220, 2014. 4 +[34] Pablo A Barrionuevo, Luis A Issolio, and Constanza Tripolone. Photoreceptor contributions to the human pupil light reflex. Journal of Photochemistry and Photobiology, 15:100178, 2023. 4 +[35] Yifan Jiang, Xinyu Gong, Ding Liu, Yu Cheng, Chen Fang, Xiaohui Shen, Jianchao Yang, Pan Zhou, and Zhangyang Wang. EnlightenGAN: Deep light enhancement without paired supervision. IEEE Transactions on Image Processing, 30:2340-2349, 2021. 6, 7 +[36] Haonan Wang, Peng Cao, Jiaqi Wang, and Osmar R Zaiane. UCTransNet: Rethinking the skip connections in U-Net from a channel-wise perspective with transformer. In Proceedings + +of the AAAI conference on artificial intelligence, volume 36, pages 2441-2449, 2022. 5 +[37] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1-9, 2015. 6 +[38] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[39] Quan Huynh-Thu and Mohammed Ghanbari. Scope of validity of PSNR in image/video quality assessment. *Electronics Letters*, 44:800–801, 2008. 6 +[40] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 6 +[41] Wenhan Yang, Wenjing Wang, Haofeng Huang, Shiqi Wang, and Jiaying Liu. Sparse gradient regularized deep retina network for robust low-light image enhancement. IEEE Transactions on Image Processing, 30:2072-2086, 2021. 6 +[42] Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3291-3300, 2018. 6 +[43] Ruixing Wang, Xiaogang Xu, Chi-Wing Fu, Jiangbo Lu, Bei Yu, and Jiaya Jia. Seeing dynamic scene in the dark: A high-quality video dataset with mechatronic alignment. In Proceedings of the IEEE International Conference on Computer Vision, pages 9700-9709, 2021. 7 +[44] Vladimir Bychkovsky, Sylvain Paris, Eric Chan, and Frédo Durand. Learning photographic global tonal adjustment with a database of input/output image pairs. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 97-104, 2011. 7 +[45] Risheng Liu, Long Ma, Tengyu Ma, Xin Fan, and Zhongxuan Luo. Learning with nested scene modeling and cooperative architecture search for low-light vision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45:5953-5969, 2022. 7 +[46] Han Zhou, Wei Dong, Xiaohong Liu, Shuaicheng Liu, Xiongkuo Min, Guangtao Zhai, and Jun Chen. GLARE: Low light image enhancement via generative latent feature based codebook retrieval. In Proceedings of the European Conference on Computer Vision, pages 36-54, 2024. 7 \ No newline at end of file diff --git a/2025/URWKV_ Unified RWKV Model with Multi-state Perspective for Low-light Image Restoration/images.zip b/2025/URWKV_ Unified RWKV Model with Multi-state Perspective for Low-light Image Restoration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..446793d4cd6341c2da2e4b2108aee3c16f2eb1a2 --- /dev/null +++ b/2025/URWKV_ Unified RWKV Model with Multi-state Perspective for Low-light Image Restoration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1bb7353b3c6e29790d41dc48c479f11a3259e9b03f3c144f8c9691125684e53e +size 778766 diff --git a/2025/URWKV_ Unified RWKV Model with Multi-state Perspective for Low-light Image Restoration/layout.json b/2025/URWKV_ Unified RWKV Model with Multi-state Perspective for Low-light Image Restoration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8c5250578758eeb71f7f5b29eae6f0165671eeae --- /dev/null +++ b/2025/URWKV_ Unified RWKV Model with Multi-state Perspective for Low-light Image Restoration/layout.json @@ -0,0 +1,9568 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 102, + 536, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 102, + 536, + 138 + ], + "spans": [ + { + "bbox": [ + 74, + 102, + 536, + 138 + ], + "type": "text", + "content": "URWKV: Unified RWKV Model with Multi-state Perspective for Low-light Image Restoration" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 160, + 556, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 160, + 556, + 217 + ], + "spans": [ + { + "bbox": [ + 61, + 160, + 556, + 217 + ], + "type": "text", + "content": "Rui Xu" + }, + { + "bbox": [ + 61, + 160, + 556, + 217 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 61, + 160, + 556, + 217 + ], + "type": "text", + "content": ", Yuzhen Niu" + }, + { + "bbox": [ + 61, + 160, + 556, + 217 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 61, + 160, + 556, + 217 + ], + "type": "text", + "content": ", Yuezhou Li" + }, + { + "bbox": [ + 61, + 160, + 556, + 217 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 61, + 160, + 556, + 217 + ], + "type": "text", + "content": ", Huangbiao Xu" + }, + { + "bbox": [ + 61, + 160, + 556, + 217 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 61, + 160, + 556, + 217 + ], + "type": "text", + "content": ", Wenxi Liu" + }, + { + "bbox": [ + 61, + 160, + 556, + 217 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 61, + 160, + 556, + 217 + ], + "type": "text", + "content": ", Yuzhong Chen" + }, + { + "bbox": [ + 61, + 160, + 556, + 217 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 61, + 160, + 556, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 556, + 217 + ], + "type": "text", + "content": "Fujian Key Laboratory of Network Computing and Intelligent Information Processing, College of Computer and Data Science, Fuzhou University, Fuzhou 350108, China \n" + }, + { + "bbox": [ + 61, + 160, + 556, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 160, + 556, + 217 + ], + "type": "text", + "content": "Engineering Research Center of Big Data Intelligence, Ministry of Education, Fuzhou 350108, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 124, + 219, + 493, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 219, + 493, + 232 + ], + "spans": [ + { + "bbox": [ + 124, + 219, + 493, + 232 + ], + "type": "text", + "content": "{xurui.ryan.chn, yuzhenniu, liyuezhou.cm, huangbiaoxu.chn}@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 198, + 234, + 417, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 234, + 417, + 245 + ], + "spans": [ + { + "bbox": [ + 198, + 234, + 417, + 245 + ], + "type": "text", + "content": "wenxi.liu@hotmail.com, yzchen@fzu.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 298, + 296, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 296, + 634 + ], + "type": "text", + "content": "Existing low-light image enhancement (LLIE) and joint LLIE and deblurring (LLIE-deblur) models have made strides in addressing predefined degradations, yet they are often constrained by dynamically coupled degradations. To address these challenges, we introduce a Unified Receptance Weighted Key Value (URWKV) model with multi-state perspective, enabling flexible and effective degradation restoration for low-light images. Specifically, we customize the core URWKV block to perceive and analyze complex degradations by leveraging multiple intra- and interstage states. First, inspired by the pupil mechanism in the human visual system, we propose Luminance-adaptive Normalization (LAN) that adjusts normalization parameters based on rich inter-stage states, allowing for adaptive, scene-aware luminance modulation. Second, we aggregate multiple intra-stage states through exponential moving average approach, effectively capturing subtle variations while mitigating information loss inherent in the single-state mechanism. To reduce the degradation effects commonly associated with conventional skip connections, we propose the State-aware Selective Fusion (SSF) module, which dynamically aligns and integrates multi-state features across encoder stages, selectively fusing contextual information. In comparison to state-of-the-art models, our URWKV model achieves superior performance on various benchmarks, while requiring significantly fewer parameters and computational resources. Code is available at: https://github.com/FZU-N/URWKV." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 647, + 135, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 135, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 135, + 659 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 668, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 668, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 668, + 296, + 693 + ], + "type": "text", + "content": "Low-light environments present significant challenges in computer vision and pattern recognition, impacting appli" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 326, + 273, + 541, + 435 + ], + "blocks": [ + { + "bbox": [ + 326, + 273, + 541, + 435 + ], + "lines": [ + { + "bbox": [ + 326, + 273, + 541, + 435 + ], + "spans": [ + { + "bbox": [ + 326, + 273, + 541, + 435 + ], + "type": "image", + "image_path": "53dcb58340066208f8c65566b6cc341cac530221b8f321650fb0e6540e93b547.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 442, + 555, + 509 + ], + "lines": [ + { + "bbox": [ + 313, + 442, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 555, + 509 + ], + "type": "text", + "content": "Figure 1. Compared against existing solutions (including Bi-Former [1] for LLIE, PDHAT [2] for joint LLIE and deblurring, and Restormer [3] as a unified model), our proposed URWKV achieves consistently superior PSNR performance across various degradation scenarios. For better visualization, the maximum PSNR values across all datasets are normalized." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 518, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 518, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 312, + 518, + 555, + 639 + ], + "type": "text", + "content": "cations such as nighttime surveillance [4], night photography [5], and remote sensing [6]. In these scenarios, insufficient lighting often leads to severe image degradations, such as increased noise, loss of detail, reduced contrast, and color distortion. Furthermore, another degradation coupled in the low-light images, particularly motion blur induced by prolonged exposure times or camera shake, further compromise the integrity of the images. These challenges not only diminish the visual quality of images but also impede subsequent image analysis and interpretation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "Low-light image enhancement (LLIE) models are foundational and widely researched solutions for addressing image degradations in low-light environments. Recent advancements have integrated sophisticated architectures like CNNs [7, 8], Transformers [9, 10], and Mamba networks [11], enabling more effective handling of complex degrad-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21267" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 156 + ], + "type": "text", + "content": "tion patterns. To address specific challenges such as uneven brightness and spatially varying noise, several specialized methods [1, 12] have been developed, enabling finer adjustments across different image regions. Despite these advancements, another typical coupled degradation, i.e., motion blur, that frequently occurs in low-light conditions, remains a significant challenge for current LLIE models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 159, + 295, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 159, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 159, + 295, + 374 + ], + "type": "text", + "content": "The primary difficulty for addressing coupled degradations arises from the inherent trade-offs among brightness enhancement, noise reduction, and blur restoration. Traditional joint training strategy or the combination strategy of off-the-shelf LLIE and deblurring models often fail to deliver optimal results [2, 13]. To bridge this gap, Zhou et al. [13] introduce a compact encoder-decoder architecture wherein the encoder focuses on low-light enhancement, while the decoder tackles blur restoration, effectively isolating the two tasks. Additionally, Li et al. [2] introduce a perceptual decoupling approach that incorporates auxiliary tasks to achieve task-specific representations. While these methods show efficacy in handling blur under low-light conditions, they are constrained by predefined degradation categories. As a result, their adaptability to the dynamically coupled degradations typically encountered in real-world scenarios is limited, where noise, blur, and other degradations may not occur in fixed, predictable combinations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 376, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 295, + 579 + ], + "type": "text", + "content": "To flexibly address various degradations, the unified models [3, 14-16] offer a natural attempt, emerging as a leading approach in image restoration. In particular, since IPT [15] first demonstrates the potential of transformer architectures for low-level vision tasks, the development of transformer-based models [3, 17] has accelerated. However, significant challenges remain, particularly in addressing complex illumination and coupled degradation conditions. Many image restoration models, such as those based on transformers, lack adaptive mechanisms designed for low-light environments, often resulting in amplified degradation and the introduction of new artifacts. Specifically, two key issues persist: 1) the difficulty of exploiting effective adaptive strategies for intricate illumination scenarios and handling coupled degradation factors, and 2) parameter inefficiency, which presents a critical bottleneck for practical deployment in downstream vision tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "In this paper, we extend the existing unified encoder-decoder framework by integrating the recently proposed Receptance Weighted Key Value (RWKV) structure [18, 19], aiming to provide effective solutions for addressing dynamically coupled degradations in low-light conditions from a multi-state perspective. Specifically, we introduce both intra- and inter-stage states to perceive and analyze diverse degradations across multiple stages, enabling a more nuanced restoration throughout the processing pipeline. To achieve this, we first propose luminance-adaptive normalization (LAN), which dynamically adjusts normaliza" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "text", + "content": "tion parameters based on rich inter-stage states, facilitating adaptive, scene-aware luminance modulation. Furthermore, we employ an exponential moving average (EMA) approach to aggregate multiple intra-stage states, effectively capturing subtle variations and mitigating the information loss inherent in single-state mechanisms. Instead of constraining our model to predefined degradation types, we exploit the relationships between multi-state representations to address spatially varying and coupled degradations, thus enabling flexible and dynamic restoration. Additionally, we introduce a state-aware selective fusion (SSF) module, which dynamically aligns and integrates multi-state features across encoder stages, selectively fusing contextual information to enhance restoration quality. As shown in Fig. 1, our model consistently outperforms existing models across various degradation scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 265, + 534, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 265, + 534, + 277 + ], + "spans": [ + { + "bbox": [ + 326, + 265, + 534, + 277 + ], + "type": "text", + "content": "The main contributions of this paper are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 277, + 553, + 480 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 277, + 553, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 277, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 314, + 277, + 553, + 324 + ], + "type": "text", + "content": "- We introduce the unified RWKV model with a novel multi-state perspective, enabling flexible and effective restoration of dynamically coupled degradations in low-light conditions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 326, + 553, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 326, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 314, + 326, + 553, + 360 + ], + "type": "text", + "content": "- We propose LAN to dynamically adjust normalization parameters based on inter-stage states, allowing for adaptive, scene-aware luminance modulation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 361, + 553, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 361, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 314, + 361, + 553, + 396 + ], + "type": "text", + "content": "- We aggregate multiple intra-stage states using an EMA-based approach to capture subtle degradation variations, while mitigating the long-range information loss." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 397, + 553, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 397, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 314, + 397, + 553, + 432 + ], + "type": "text", + "content": "- We develop the SSF module, which dynamically aligns and integrates multi-state features across encoder stages, facilitating selective fusion of contextual information." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 434, + 553, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 434, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 314, + 434, + 553, + 480 + ], + "type": "text", + "content": "- Our URWKV model outperforms existing models across diverse benchmarks, effectively handling dynamically coupled degradations with fewer parameters and lower computational cost." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 495, + 400, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 495, + 400, + 507 + ], + "spans": [ + { + "bbox": [ + 314, + 495, + 400, + 507 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 515, + 481, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 481, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 481, + 528 + ], + "type": "text", + "content": "2.1. Low-light Image Enhancement" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "Compared to traditional methods relying on histogram equalization [20-22] or Retinex-based enhancement [23-25], deep learning approaches have gained prominence in low-light image enhancement for their ability to handle complex degradations. Retinex-based models, in particular, have evolved through the integration of advanced architectures such as CNNs [7, 8], Transformers [9], and Mamba networks [11], enabling more precise illumination-reflectance decomposition. Leveraging the long-range dependency modeling capabilities of Transformers, several methods [1, 10] have introduced modifications to overcome the limitations of Transformers in low-light scenarios, achieving improved computational efficiency and enhanced local modeling capabilities. In addition, recognizing that noise characteristics vary spatially, Xu et al. [12] propose to" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21268" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 70, + 312, + 284 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 312, + 284 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 312, + 284 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 312, + 284 + ], + "type": "image", + "image_path": "c506fea410fc5e2785ecd9faa72ffbf6bfbb8afc59acb2d23acc1ab36f95ff18.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 286, + 214, + 297 + ], + "lines": [ + { + "bbox": [ + 164, + 286, + 214, + 297 + ], + "spans": [ + { + "bbox": [ + 164, + 286, + 214, + 297 + ], + "type": "text", + "content": "(a) Framework" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 326, + 76, + 541, + 240 + ], + "blocks": [ + { + "bbox": [ + 326, + 76, + 541, + 240 + ], + "lines": [ + { + "bbox": [ + 326, + 76, + 541, + 240 + ], + "spans": [ + { + "bbox": [ + 326, + 76, + 541, + 240 + ], + "type": "image", + "image_path": "94c74fa9db9e91bb5f9f7701a8d6a8daa02ba5cc1e7c821903a801e8b923ebe1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 495, + 247, + 539, + 257 + ], + "lines": [ + { + "bbox": [ + 495, + 247, + 539, + 257 + ], + "spans": [ + { + "bbox": [ + 495, + 247, + 539, + 257 + ], + "type": "text", + "content": "Squared ReLU" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 327, + 247, + 340, + 258 + ], + "blocks": [ + { + "bbox": [ + 327, + 247, + 340, + 258 + ], + "lines": [ + { + "bbox": [ + 327, + 247, + 340, + 258 + ], + "spans": [ + { + "bbox": [ + 327, + 247, + 340, + 258 + ], + "type": "image", + "image_path": "de7a542251406ef978dad269c459a3d5c43318aab7fdc74d3e8982b6aff6cdc6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 249, + 389, + 258 + ], + "lines": [ + { + "bbox": [ + 345, + 249, + 389, + 258 + ], + "spans": [ + { + "bbox": [ + 345, + 249, + 389, + 258 + ], + "type": "text", + "content": "Concatenation" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 408, + 247, + 421, + 258 + ], + "blocks": [ + { + "bbox": [ + 408, + 247, + 421, + 258 + ], + "lines": [ + { + "bbox": [ + 408, + 247, + 421, + 258 + ], + "spans": [ + { + "bbox": [ + 408, + 247, + 421, + 258 + ], + "type": "image", + "image_path": "f97b41b07470492ea5ea202a3fba61893fb2f7f3aaac2d440cf96177c32201bd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 426, + 249, + 451, + 258 + ], + "lines": [ + { + "bbox": [ + 426, + 249, + 451, + 258 + ], + "spans": [ + { + "bbox": [ + 426, + 249, + 451, + 258 + ], + "type": "text", + "content": "Sigmoid" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 478, + 246, + 490, + 258 + ], + "blocks": [ + { + "bbox": [ + 478, + 246, + 490, + 258 + ], + "lines": [ + { + "bbox": [ + 478, + 246, + 490, + 258 + ], + "spans": [ + { + "bbox": [ + 478, + 246, + 490, + 258 + ], + "type": "image", + "image_path": "3147b06902360a624e91930a280d99747ac4f27eac3d7f2681d03f2059671ec2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 328, + 262, + 339, + 277 + ], + "blocks": [ + { + "bbox": [ + 328, + 262, + 339, + 277 + ], + "lines": [ + { + "bbox": [ + 328, + 262, + 339, + 277 + ], + "spans": [ + { + "bbox": [ + 328, + 262, + 339, + 277 + ], + "type": "image", + "image_path": "a8c7bab8368351d721f368ba6e4855d2e4b6c0180776e3a78acfaced5d9bbec6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 264, + 388, + 279 + ], + "lines": [ + { + "bbox": [ + 345, + 264, + 388, + 279 + ], + "spans": [ + { + "bbox": [ + 345, + 264, + 388, + 279 + ], + "type": "text", + "content": "Element-wise Multiplication" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 336, + 286, + 544, + 297 + ], + "lines": [ + { + "bbox": [ + 336, + 286, + 544, + 297 + ], + "spans": [ + { + "bbox": [ + 336, + 286, + 544, + 297 + ], + "type": "text", + "content": "(b) URWKV Block with Multi-state Spatial and Channel Mixing" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 305, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 54, + 305, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 54, + 305, + 555, + 361 + ], + "type": "text", + "content": "Figure 2. Overview of the URWKV model. The core URWKV block is integrated into an encoder-decoder framework, featuring a multi-state spatial mixing sub-block and a multi-state channel mixing sub-block. The luminance-adaptive normalization (LAN) incorporates both the current input state " + }, + { + "bbox": [ + 54, + 305, + 555, + 361 + ], + "type": "inline_equation", + "content": "X_{t}" + }, + { + "bbox": [ + 54, + 305, + 555, + 361 + ], + "type": "text", + "content": " and multiple inter-stage states " + }, + { + "bbox": [ + 54, + 305, + 555, + 361 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 54, + 305, + 555, + 361 + ], + "type": "text", + "content": ". Additionally, the multi-state quad-directional token shift (SQ-Shift) aggregates both current input state and multiple intra-stage states " + }, + { + "bbox": [ + 54, + 305, + 555, + 361 + ], + "type": "inline_equation", + "content": "H_{i}" + }, + { + "bbox": [ + 54, + 305, + 555, + 361 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 54, + 305, + 555, + 361 + ], + "type": "inline_equation", + "content": "H_{S_{i}}" + }, + { + "bbox": [ + 54, + 305, + 555, + 361 + ], + "type": "text", + "content": " for spatial mixing and " + }, + { + "bbox": [ + 54, + 305, + 555, + 361 + ], + "type": "inline_equation", + "content": "H_{C_{i}}" + }, + { + "bbox": [ + 54, + 305, + 555, + 361 + ], + "type": "text", + "content": " for channel mixing, respectively). The state-aware selective fusion (SSF) module further aggregates rich contextual information across encoder stages." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 408, + 262, + 421, + 277 + ], + "blocks": [ + { + "bbox": [ + 408, + 262, + 421, + 277 + ], + "lines": [ + { + "bbox": [ + 408, + 262, + 421, + 277 + ], + "spans": [ + { + "bbox": [ + 408, + 262, + 421, + 277 + ], + "type": "image", + "image_path": "297f6290ee9be3c5ad0dad1a19dd303e8b948aea7ababdc1c80cbc1fd042180b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 263, + 467, + 279 + ], + "lines": [ + { + "bbox": [ + 425, + 263, + 467, + 279 + ], + "spans": [ + { + "bbox": [ + 425, + 263, + 467, + 279 + ], + "type": "text", + "content": "Element-wise Addition" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 376, + 295, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 376, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 54, + 376, + 295, + 448 + ], + "type": "text", + "content": "separate signal from noise during the enhancement process, enabling more effective brightening while minimizing noise amplification. Furthermore, frequency-domain approaches [26, 27] also contribute to LLIE by decomposing features across distinct frequency bands, proving effective in preserving fine details and reducing noise." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 54, + 449, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 449, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 54, + 449, + 295, + 556 + ], + "type": "text", + "content": "Considering blur as a frequent coupled degradation in low-light scenes, one that conventional LLIE models inadequately address, Zhou et al. [13] and Li et al. [2] propose a specific sequential encoder-decoder architecture and a parallel heterogeneous decoupling framework, respectively. Although these LLIE-deblur models show effectiveness within certain constraints, further exploration is warranted to handle the dynamically coupled degradations typical in real-world scenarios." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 563, + 251, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 251, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 251, + 577 + ], + "type": "text", + "content": "2.2. Unified Model for Image Restoration" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "content": "With the emergence of deep learning, unified approaches have driven substantial progress, enabling the handling of multiple restoration tasks within a single framework. Compared to CNN-based methods [14, 28], which primarily focus on multi-scale feature extraction and local context enhancement, Transformer-based methods [3, 15, 29, 30] extend relational modeling capabilities by capturing long-range dependencies, leading to improved restoration quality across diverse degradation types. However, the application of Transformers in image restoration often entails high computational costs and memory requirements, which can" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 376, + 554, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 554, + 460 + ], + "type": "text", + "content": "hinder their efficiency. In addition, frequency-domain models [31, 32] provide another unified approach by decomposing images into multiple frequency bands, which enables focused noise reduction while preserving fine details across tasks. Recent innovations also utilize Mamba networks [16] to achieve effective image restoration by efficiently balancing local and global feature representations." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 460, + 554, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 509 + ], + "type": "text", + "content": "While current unified models perform effectively under normal lighting conditions, they encounter significant challenges in low-light environments, specifically involving complex lighting variations and coupled degradations." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 520, + 370, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 370, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 370, + 532 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 540, + 382, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 382, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 382, + 552 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 557, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 554, + 665 + ], + "type": "text", + "content": "As illustrated in Fig. 2, the URWKV framework follows the typical encoder-decoder architecture, with both the encoder and decoder comprising three stages. In the encoder, each stage is composed of " + }, + { + "bbox": [ + 313, + 557, + 554, + 665 + ], + "type": "inline_equation", + "content": "N_{1}" + }, + { + "bbox": [ + 313, + 557, + 554, + 665 + ], + "type": "text", + "content": " URWKV blocks followed by a down-sampling layer, while each stage in the decoder contains " + }, + { + "bbox": [ + 313, + 557, + 554, + 665 + ], + "type": "inline_equation", + "content": "N_{2}" + }, + { + "bbox": [ + 313, + 557, + 554, + 665 + ], + "type": "text", + "content": " URWKV blocks and an up-sampling layer. Additionally, each down-sampling or up-sampling layer includes convolutional operations and sampling techniques to adjust either the channel or spatial dimensions." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "To effectively handle dynamically coupled degradations, each URWKV block operates as an independent and progressive processing stage, with each stage's output (an implicit restored result, serving as a 'state') encapsulat-" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21269" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 70, + 255, + 187 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 255, + 187 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 255, + 187 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 255, + 187 + ], + "type": "image", + "image_path": "da414007927556da9a4b014cfa97d7c676311ccb7d3d5955bbe129d1df082ada.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 190, + 223, + 201 + ], + "lines": [ + { + "bbox": [ + 97, + 190, + 223, + 201 + ], + "spans": [ + { + "bbox": [ + 97, + 190, + 223, + 201 + ], + "type": "text", + "content": "(a) Luminance-adaptive Normalization" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 267, + 70, + 545, + 183 + ], + "blocks": [ + { + "bbox": [ + 267, + 70, + 545, + 183 + ], + "lines": [ + { + "bbox": [ + 267, + 70, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 267, + 70, + 545, + 183 + ], + "type": "image", + "image_path": "b6821a700fc76de291b57e101098d4572a8346af8ad168f077a1cab2c0b94a6e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 190, + 494, + 201 + ], + "lines": [ + { + "bbox": [ + 326, + 190, + 494, + 201 + ], + "spans": [ + { + "bbox": [ + 326, + 190, + 494, + 201 + ], + "type": "text", + "content": "(b) Multi-kernel aggregation of inter-stage states" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 210, + 553, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 210, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 210, + 553, + 232 + ], + "type": "text", + "content": "Figure 3. Illustration of luminance-adaptive normalization (LAN). LAN integrates inter-stage states " + }, + { + "bbox": [ + 55, + 210, + 553, + 232 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 55, + 210, + 553, + 232 + ], + "type": "text", + "content": " throughout the restoration process, facilitating scene-aware luminance modulation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 248, + 294, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 248, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 294, + 272 + ], + "type": "text", + "content": "ing stage-specific contextual semantics and restoration cues (e.g., luminance or content) for subsequent stages." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 272, + 295, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 272, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 54, + 272, + 295, + 475 + ], + "type": "text", + "content": "Specifically, the URWKV block builds on the strengths of original RWKV model [19], while enhancing its capabilities through two tailored sub-blocks: a multi-state spatial mixing sub-block and a multi-state channel mixing sub-block. The former captures spatial dependencies across multiple intra-stage states to enrich spatial information, while the latter enhances feature representations through multi-state channel-wise interactions. Here, the intra-stage states " + }, + { + "bbox": [ + 54, + 272, + 295, + 475 + ], + "type": "inline_equation", + "content": "H_{S_i}" + }, + { + "bbox": [ + 54, + 272, + 295, + 475 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 54, + 272, + 295, + 475 + ], + "type": "inline_equation", + "content": "H_{C_i}" + }, + { + "bbox": [ + 54, + 272, + 295, + 475 + ], + "type": "text", + "content": " are updated following the SQ-Shift in each sub-block, respectively, where " + }, + { + "bbox": [ + 54, + 272, + 295, + 475 + ], + "type": "inline_equation", + "content": "i \\in \\{1, \\dots, N_1\\}" + }, + { + "bbox": [ + 54, + 272, + 295, + 475 + ], + "type": "text", + "content": " (or " + }, + { + "bbox": [ + 54, + 272, + 295, + 475 + ], + "type": "inline_equation", + "content": "i \\in \\{1, \\dots, N_2\\}" + }, + { + "bbox": [ + 54, + 272, + 295, + 475 + ], + "type": "text", + "content": "). To ensure efficient information transfer between encoder and decoder stages without introducing degradation, we incorporate the state-aware selective fusion (SSF) module between these stages. The SSF dynamically aggregates multi-state features across all encoder stages, selectively propagating rich contextual information into the decoder." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 480, + 255, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 255, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 255, + 493 + ], + "type": "text", + "content": "3.2. URWKV with Multi-state Perspective" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "type": "text", + "content": "Original RWKV models [18, 19] employ a single-state token shift mechanism, where each token's relationship relies solely on the current and preceding state. This design introduces two primary limitations: First, it struggles to retain long-range dependencies, causing early information to gradually fade during restoration. Second, it fails to effectively capture complex interdependencies among degraded features, as a single state is insufficient to represent intricate coupled degradations. To this end, we propose URWKV devised with the multi-state perspective. Specifically, the URWKV block tailors the luminance-adaptive normalization (LAN) by incorporating multiple inter-stage states to adapt URWKV to low-light scenarios. Additionally, it integrates multiple intra-stage states to facilitate more extensive interactions across long-range dependencies." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Luminance-adaptive Normalization (LAN). Luminance distortion is a major factor leading to other coupled degradations in low-light images. Inspired by the hu" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 248, + 553, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 553, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 553, + 332 + ], + "type": "text", + "content": "man eye's ability to adaptively adjust pupil size in response to varying luminance [33, 34], LAN leverages inter-stage states throughout the restoration process (comprising six stages) to enable scene-aware luminance modulation. This design overcomes limitation of LayerNorm widely used in existing models [3, 17], which relies on fixed parameters " + }, + { + "bbox": [ + 313, + 248, + 553, + 332 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 248, + 553, + 332 + ], + "type": "text", + "content": " for scaling the learned representation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "text", + "content": "As illustrated in Fig. 3, given the current state input " + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "inline_equation", + "content": "X_{t}\\in \\mathbb{R}^{C_{t}\\times H_{t}\\times W_{t}}" + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "text", + "content": " at stage " + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "text", + "content": ", we first extract 1D luminance vectors from both " + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "inline_equation", + "content": "X_{t}" + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "text", + "content": " and each historical stage output " + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "inline_equation", + "content": "M_{i}\\in \\mathbb{R}^{C_{i}\\times H_{i}\\times W_{i}}" + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "text", + "content": ", using the global average pooling (GAP) operation. Here, each " + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "text", + "content": " is obtained from the output of each stage's final URWKV block. " + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "inline_equation", + "content": "t\\in \\{1,\\dots,N_1\\}" + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "text", + "content": " (or " + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "inline_equation", + "content": "t\\in \\{1,\\dots,N_2\\}" + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "inline_equation", + "content": "i\\in \\{1,\\dots,T - 1\\}" + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "text", + "content": ". The luminance vectors are then zero-padded to match the dimensionality of the vector with the maximum channel size " + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "inline_equation", + "content": "C_{max}" + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "text", + "content": ", while preserving the semantic integrity of luminance vector from each individual state. These zero-padded luminance vectors, denoted as " + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "inline_equation", + "content": "X_{k}^{1D}\\in \\mathbb{R}^{C_{max}}" + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "text", + "content": ", for " + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "inline_equation", + "content": "k\\in \\{1,\\dots,T\\}" + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "text", + "content": ", are then stacked along the state dimension. Next, a multi-kernel aggregation layer is applied to analyze the stacked 2D luminance map along the channel dimension. Specifically, the aggregation layer uses 1D convolutional kernels of sizes " + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "inline_equation", + "content": "1\\times T" + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "inline_equation", + "content": "3\\times T" + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "inline_equation", + "content": "5\\times T" + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "text", + "content": " to adaptively capture both local and broader contextual illuminance variations across multiple states. The resulting luminance maps " + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "inline_equation", + "content": "X_{r}^{2D}\\in \\mathbb{R}^{T\\times C_{max}}" + }, + { + "bbox": [ + 313, + 332, + 554, + 571 + ], + "type": "text", + "content": " are expressed as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 328, + 580, + 553, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 580, + 553, + 596 + ], + "spans": [ + { + "bbox": [ + 328, + 580, + 553, + 596 + ], + "type": "interline_equation", + "content": "X _ {r} ^ {2 D} = \\operatorname {C o n v} _ {r} ^ {1 D} \\left(\\left[ X _ {1} ^ {1 D},..., X _ {T} ^ {1 D} \\right]\\right), r \\in \\{1, 3, 5 \\}, \\tag {1}", + "image_path": "da6b73b5e0849fd9b2b66ddfe120af3b7c69ad44ad2edd5c0771d49bf9088336.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 553, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 605, + 553, + 629 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}_r^{1D}" + }, + { + "bbox": [ + 313, + 605, + 553, + 629 + ], + "type": "text", + "content": " denotes the set of 1D convolutional kernels, and " + }, + { + "bbox": [ + 313, + 605, + 553, + 629 + ], + "type": "inline_equation", + "content": "[\\cdot ,\\cdot ]" + }, + { + "bbox": [ + 313, + 605, + 553, + 629 + ], + "type": "text", + "content": " denotes the stacking operation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": "Subsequently, the output of each convolutional kernel is concatenated along the channel dimension and projected through a " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": " convolution, resulting in the aggregated feature map " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{agg}}\\in \\mathbb{R}^{C_t}" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": ". A multi-layer perceptron (MLP) along with a tanh activation function are then employed to predict the luminance modulator " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\Delta \\gamma_{t}" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": ". The luminance modulator is computed as:" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21270" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 119, + 85, + 295, + 99 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 85, + 295, + 99 + ], + "spans": [ + { + "bbox": [ + 119, + 85, + 295, + 99 + ], + "type": "interline_equation", + "content": "\\Delta \\gamma_ {t} = \\tanh (\\mathrm {M L P} \\left(X _ {\\text {a g g}}\\right)), \\tag {2}", + "image_path": "910840496bcafe6cf3926b086bef77cc405c6e6fa55899409e9cf8383be9dd9a.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 104, + 272, + 116 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 104, + 272, + 116 + ], + "spans": [ + { + "bbox": [ + 55, + 104, + 272, + 116 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 104, + 272, + 116 + ], + "type": "inline_equation", + "content": "\\tanh (\\cdot)" + }, + { + "bbox": [ + 55, + 104, + 272, + 116 + ], + "type": "text", + "content": " represents the tanh activation function." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 117, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 117, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 117, + 296, + 163 + ], + "type": "text", + "content": "Once " + }, + { + "bbox": [ + 55, + 117, + 296, + 163 + ], + "type": "inline_equation", + "content": "\\Delta \\gamma_{t}" + }, + { + "bbox": [ + 55, + 117, + 296, + 163 + ], + "type": "text", + "content": " is predicted, the scaling parameter " + }, + { + "bbox": [ + 55, + 117, + 296, + 163 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 117, + 296, + 163 + ], + "type": "text", + "content": " can be updated as " + }, + { + "bbox": [ + 55, + 117, + 296, + 163 + ], + "type": "inline_equation", + "content": "\\hat{\\gamma}_{t} = \\gamma +\\Delta \\gamma_{t}" + }, + { + "bbox": [ + 55, + 117, + 296, + 163 + ], + "type": "text", + "content": ". Ultimately, the input feature " + }, + { + "bbox": [ + 55, + 117, + 296, + 163 + ], + "type": "inline_equation", + "content": "X_{t}" + }, + { + "bbox": [ + 55, + 117, + 296, + 163 + ], + "type": "text", + "content": " at stage " + }, + { + "bbox": [ + 55, + 117, + 296, + 163 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 117, + 296, + 163 + ], + "type": "text", + "content": " can be adaptively normalized with LAN, expressed as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 114, + 174, + 295, + 199 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 174, + 295, + 199 + ], + "spans": [ + { + "bbox": [ + 114, + 174, + 295, + 199 + ], + "type": "interline_equation", + "content": "X _ {t} ^ {L A N} = \\frac {X _ {t} - \\mu_ {t}}{\\sigma_ {t}} \\odot \\hat {\\gamma} _ {t} + \\beta , \\tag {3}", + "image_path": "a147c88745023b097a6349ba574f6b2ab6836f2ff053f47f514022410ba14c64.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 204, + 295, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 295, + 228 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 204, + 295, + 228 + ], + "type": "inline_equation", + "content": "\\mu_t" + }, + { + "bbox": [ + 55, + 204, + 295, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 204, + 295, + 228 + ], + "type": "inline_equation", + "content": "\\sigma_t" + }, + { + "bbox": [ + 55, + 204, + 295, + 228 + ], + "type": "text", + "content": " represent the mean and standard deviation of the input feature " + }, + { + "bbox": [ + 55, + 204, + 295, + 228 + ], + "type": "inline_equation", + "content": "X_t" + }, + { + "bbox": [ + 55, + 204, + 295, + 228 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 55, + 204, + 295, + 228 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 204, + 295, + 228 + ], + "type": "text", + "content": " is a shift parameter." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 228, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 443 + ], + "type": "text", + "content": "Multi-state Quad-directional Token Shift (SQ-Shift). Original quad-directional token shift (Q-shift) [19] effectively captures spatial dependencies among neighboring tokens. However, it neglects the degradation relations between states, making it difficult to maintain consistent degradation restoration across consecutive states. To address this limitation, we aggregate multi-state features to extend Q-Shift, allowing the model to capture long-range dependencies across multiple states while preserving local spatial interactions. Specifically, before applying Q-shift, we employ the exponential moving average (EMA) approach to aggregate the current state with all preceding states within the same stage, enabling richer interactions with historical information. For instance, given the state feature " + }, + { + "bbox": [ + 55, + 228, + 296, + 443 + ], + "type": "inline_equation", + "content": "X_{t}^{LAN} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 55, + 228, + 296, + 443 + ], + "type": "text", + "content": " normalized by LAN, and multiple historical intra-stage state features " + }, + { + "bbox": [ + 55, + 228, + 296, + 443 + ], + "type": "inline_equation", + "content": "H_{i}, i \\in \\{1, \\dots, t - 1\\}" + }, + { + "bbox": [ + 55, + 228, + 296, + 443 + ], + "type": "text", + "content": ", we define the multi-state aggregation MSA as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 60, + 452, + 295, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 452, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 60, + 452, + 295, + 468 + ], + "type": "interline_equation", + "content": "\\operatorname {M S A} \\left(X _ {t} ^ {L A N}\\right) = \\alpha \\odot X _ {t} ^ {L A N} + (1 - \\alpha) \\odot \\operatorname {M S A} \\left(H _ {t - 1}\\right), \\tag {4}", + "image_path": "0ec58e6215333e8b7faa2fd062207caea343aa18abdb579724be4500439c3721.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 476, + 295, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 476, + 295, + 500 + ], + "spans": [ + { + "bbox": [ + 55, + 476, + 295, + 500 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 476, + 295, + 500 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 476, + 295, + 500 + ], + "type": "text", + "content": " is a decay factor that controls the weight between the current state and the previously aggregated state." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 501, + 295, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 295, + 656 + ], + "type": "text", + "content": "Multi-State Spatial and Channel Mixing. For each input feature in the multi-state spatial mixing sub-block or multi-state channel mixing sub-block, we first apply LAN and SQ-Shift, resulting in features " + }, + { + "bbox": [ + 55, + 501, + 295, + 656 + ], + "type": "inline_equation", + "content": "X_{s}" + }, + { + "bbox": [ + 55, + 501, + 295, + 656 + ], + "type": "text", + "content": " for spatial mixing or " + }, + { + "bbox": [ + 55, + 501, + 295, + 656 + ], + "type": "inline_equation", + "content": "X_{c}" + }, + { + "bbox": [ + 55, + 501, + 295, + 656 + ], + "type": "text", + "content": " for channel mixing. In the spatial mixing process, " + }, + { + "bbox": [ + 55, + 501, + 295, + 656 + ], + "type": "inline_equation", + "content": "X_{s}" + }, + { + "bbox": [ + 55, + 501, + 295, + 656 + ], + "type": "text", + "content": " undergoes three parallel linear projections to generate the matrices for receptance " + }, + { + "bbox": [ + 55, + 501, + 295, + 656 + ], + "type": "inline_equation", + "content": "R_{s}" + }, + { + "bbox": [ + 55, + 501, + 295, + 656 + ], + "type": "text", + "content": ", key " + }, + { + "bbox": [ + 55, + 501, + 295, + 656 + ], + "type": "inline_equation", + "content": "K_{s}" + }, + { + "bbox": [ + 55, + 501, + 295, + 656 + ], + "type": "text", + "content": ", and value " + }, + { + "bbox": [ + 55, + 501, + 295, + 656 + ], + "type": "inline_equation", + "content": "V_{s}" + }, + { + "bbox": [ + 55, + 501, + 295, + 656 + ], + "type": "text", + "content": ". Following the computational approach of [19], we employ the bidirectional attention mechanism Bi-WKV to compute the global attention weights " + }, + { + "bbox": [ + 55, + 501, + 295, + 656 + ], + "type": "inline_equation", + "content": "wkv" + }, + { + "bbox": [ + 55, + 501, + 295, + 656 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 55, + 501, + 295, + 656 + ], + "type": "inline_equation", + "content": "K_{s}" + }, + { + "bbox": [ + 55, + 501, + 295, + 656 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 501, + 295, + 656 + ], + "type": "inline_equation", + "content": "V_{s}" + }, + { + "bbox": [ + 55, + 501, + 295, + 656 + ], + "type": "text", + "content": ", which are subsequently modulated by the sigmoid-activated receptance " + }, + { + "bbox": [ + 55, + 501, + 295, + 656 + ], + "type": "inline_equation", + "content": "R_{s}" + }, + { + "bbox": [ + 55, + 501, + 295, + 656 + ], + "type": "text", + "content": " to control the output. The output feature " + }, + { + "bbox": [ + 55, + 501, + 295, + 656 + ], + "type": "inline_equation", + "content": "O_{s}" + }, + { + "bbox": [ + 55, + 501, + 295, + 656 + ], + "type": "text", + "content": " is therefore represented as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 117, + 666, + 295, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 666, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 117, + 666, + 295, + 681 + ], + "type": "interline_equation", + "content": "O _ {s} = \\left(\\sigma \\left(R _ {s}\\right) \\odot w k v\\right) W _ {O _ {s}}, \\tag {5}", + "image_path": "10dc7d706357d6eca9862d3b96b144b7426154e98d3c5968bb3b643285aa9b7b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "wkv = \\mathrm{Bi - WKV}(K_{\\mathrm{s}},V_{\\mathrm{s}})" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " represents the sigmoid function. " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "W_{O_s}" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " is learnable parameter of projection layer." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 403, + 224 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 403, + 224 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 403, + 224 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 403, + 224 + ], + "type": "image", + "image_path": "17cc601dad29b54cc16d8c3c43b0df24d4f0f5dbd1f57be0414a1594f9bce023.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 228, + 400, + 236 + ], + "lines": [ + { + "bbox": [ + 318, + 228, + 400, + 236 + ], + "spans": [ + { + "bbox": [ + 318, + 228, + 400, + 236 + ], + "type": "text", + "content": "(a) State-aware Selective Fusion" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 418, + 69, + 552, + 225 + ], + "blocks": [ + { + "bbox": [ + 418, + 69, + 552, + 225 + ], + "lines": [ + { + "bbox": [ + 418, + 69, + 552, + 225 + ], + "spans": [ + { + "bbox": [ + 418, + 69, + 552, + 225 + ], + "type": "image", + "image_path": "fe1e1c9639da0d9e46e7a449a121f9202706f24f6c75dd64998fbb794f8e6c98.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 458, + 228, + 508, + 236 + ], + "lines": [ + { + "bbox": [ + 458, + 228, + 508, + 236 + ], + "spans": [ + { + "bbox": [ + 458, + 228, + 508, + 236 + ], + "type": "text", + "content": "(c) State Alignment" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 322, + 242, + 545, + 254 + ], + "lines": [ + { + "bbox": [ + 322, + 242, + 545, + 254 + ], + "spans": [ + { + "bbox": [ + 322, + 242, + 545, + 254 + ], + "type": "text", + "content": "Figure 4. Illustration of the state-aware fusion (SSF) module." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 263, + 554, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 554, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 554, + 311 + ], + "type": "text", + "content": "For channel mixing, " + }, + { + "bbox": [ + 313, + 263, + 554, + 311 + ], + "type": "inline_equation", + "content": "X_{c}" + }, + { + "bbox": [ + 313, + 263, + 554, + 311 + ], + "type": "text", + "content": " undergoes two parallel linear projections to derive the receptance matrix " + }, + { + "bbox": [ + 313, + 263, + 554, + 311 + ], + "type": "inline_equation", + "content": "R_{c}" + }, + { + "bbox": [ + 313, + 263, + 554, + 311 + ], + "type": "text", + "content": " and key matrix " + }, + { + "bbox": [ + 313, + 263, + 554, + 311 + ], + "type": "inline_equation", + "content": "K_{c}" + }, + { + "bbox": [ + 313, + 263, + 554, + 311 + ], + "type": "text", + "content": ". Then, the output feature " + }, + { + "bbox": [ + 313, + 263, + 554, + 311 + ], + "type": "inline_equation", + "content": "O_{c}" + }, + { + "bbox": [ + 313, + 263, + 554, + 311 + ], + "type": "text", + "content": " for channel mixing can be formulated as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 330, + 317, + 553, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 317, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 330, + 317, + 553, + 331 + ], + "type": "interline_equation", + "content": "O _ {c} = \\left(\\sigma \\left(R _ {c}\\right) \\odot \\operatorname {S q u a r e d R e L U} \\left(K _ {c}\\right) W _ {V _ {c}}\\right) W _ {O _ {c}}, \\tag {6}", + "image_path": "c4b37ed0d4691da5e0eebcbb1ca19b51fdde8535f678684a301b99ae24a1c8e5.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 336, + 554, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 554, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 554, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 336, + 554, + 361 + ], + "type": "inline_equation", + "content": "W_{V_c}" + }, + { + "bbox": [ + 313, + 336, + 554, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 336, + 554, + 361 + ], + "type": "inline_equation", + "content": "W_{O_c}" + }, + { + "bbox": [ + 313, + 336, + 554, + 361 + ], + "type": "text", + "content": " are learnable parameters in the two projection layers." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 366, + 499, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 499, + 379 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 499, + 379 + ], + "type": "text", + "content": "3.3. State-aware Selective Fusion (SSF)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 384, + 554, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 554, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 554, + 467 + ], + "type": "text", + "content": "Existing models typically reintroduce early-stage features using navie skip connections (e.g., concatenation or addition). However, they are prone to two major issues, particularly under low-light conditions: (1) the propagation of noise and irrelevant information from the encoder to the decoder [1], and (2) the semantic gap between stages [36], which hinders effective feature fusion." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 468, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 555, + 696 + ], + "type": "text", + "content": "Rather than adopting the navie or parameter-heavy approaches [3, 17, 36], we propose the lightweight and effective state-aware selective fusion (SSF) module. As illustrated in Fig. 4, the SSF module focuses on capturing the degradation restoration relationships between multistate features to predict the spatial guiding weight, which selectively refines the transfer of encoder features to the decoder. Given three encoder outputs " + }, + { + "bbox": [ + 313, + 468, + 555, + 696 + ], + "type": "inline_equation", + "content": "E_{i} \\in \\mathbb{R}^{C_{i} \\times H_{i} \\times W_{i}}" + }, + { + "bbox": [ + 313, + 468, + 555, + 696 + ], + "type": "text", + "content": " obtained from " + }, + { + "bbox": [ + 313, + 468, + 555, + 696 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 313, + 468, + 555, + 696 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 468, + 555, + 696 + ], + "type": "inline_equation", + "content": "i \\in \\{1,2,3\\}" + }, + { + "bbox": [ + 313, + 468, + 555, + 696 + ], + "type": "text", + "content": ", we first apply a channel-wise mean operation to collapse the channel dimension, resulting in " + }, + { + "bbox": [ + 313, + 468, + 555, + 696 + ], + "type": "inline_equation", + "content": "E_{i}^{M} \\in \\mathbb{R}^{1 \\times H_{i} \\times W_{i}}" + }, + { + "bbox": [ + 313, + 468, + 555, + 696 + ], + "type": "text", + "content": ", which reduces inter-channel semantic interference. Next, the adaptive alignment operation is performed on " + }, + { + "bbox": [ + 313, + 468, + 555, + 696 + ], + "type": "inline_equation", + "content": "E_{i}^{M}" + }, + { + "bbox": [ + 313, + 468, + 555, + 696 + ], + "type": "text", + "content": " to adjust the spatial dimensions of each feature according to the target dimensions of the decoder at the current fusion stage. The aligned features " + }, + { + "bbox": [ + 313, + 468, + 555, + 696 + ], + "type": "inline_equation", + "content": "E_{i}^{A}" + }, + { + "bbox": [ + 313, + 468, + 555, + 696 + ], + "type": "text", + "content": " are then stacked along the channel dimension to produce the concatenated feature map " + }, + { + "bbox": [ + 313, + 468, + 555, + 696 + ], + "type": "inline_equation", + "content": "E^{A} \\in \\mathbb{R}^{3 \\times H_{d} \\times W_{d}}" + }, + { + "bbox": [ + 313, + 468, + 555, + 696 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 468, + 555, + 696 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 313, + 468, + 555, + 696 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 468, + 555, + 696 + ], + "type": "inline_equation", + "content": "W_{d}" + }, + { + "bbox": [ + 313, + 468, + 555, + 696 + ], + "type": "text", + "content": " are the height and width of the decoder's feature. This process is formally expressed as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 326, + 700, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 700, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 326, + 700, + 555, + 715 + ], + "type": "interline_equation", + "content": "E ^ {A} = \\operatorname {S t a c k} (\\operatorname {A d a A l i g n} (\\operatorname {A v g} (E _ {i}))), i \\in \\{1, 2, 3 \\}, \\quad (7)", + "image_path": "d67ef1b687ce24b9f1841dc85450b170259c4ba282b97f773df88f0e8a1ced47.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21271" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 99, + 553, + 312 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Quantitative comparison results on the LOL-v2-real, LOL-v2-syn, SID, SMID, SDSD-indoor, and SDSD-outdoor datasets. Joint LLIE and deblurring models are marked with †, unified models are marked with *, and the rest represent LLIE models." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 99, + 553, + 312 + ], + "lines": [ + { + "bbox": [ + 58, + 99, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 58, + 99, + 553, + 312 + ], + "type": "table", + "html": "
MethodsLOL-v2-realLOL-v2-synSIDSMIDSDSD-indoorSDSD-outdoorComplexity
PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑FLOPs (G)Params (M)
KinD [7]14.740.64113.290.57818.020.58322.180.63421.950.67221.970.65434.998.02
EnlightenGAN [35]18.230.61716.570.73417.230.54322.620.67420.020.60420.100.61661.01114.35
SNR-Net [12]21.480.84924.140.92721.350.55028.520.80726.130.81519.220.65726.354.01
FourLLIE [26]22.350.84724.650.91918.420.51325.640.75024.740.82624.670.787-0.12
UHDFour [27]21.790.85423.600.91322.740.65128.180.80929.130.89523.280.77157.4215.90
LLFormer [10]21.630.80824.130.90822.830.65628.420.79426.820.82529.180.84122.5224.52
Retinexformer [9]22.790.83925.670.92824.440.68029.150.81529.780.89529.830.87815.571.61
BiFormer [1]22.670.86324.870.92122.490.63828.240.81228.730.89926.130.8415.930.83
RetinexMamba [11]22.450.84325.880.93322.450.65628.620.80928.440.89428.520.85934.7524.1
LEDNet†[13]19.020.83524.790.93121.470.63828.420.80727.290.87626.660.85038.577.41
PDHAT†[2]20.160.84124.940.93721.930.65329.190.81726.370.88428.190.862208.197.83
Uformer* [17]18.820.77119.660.87118.540.57727.200.79223.170.85923.850.74812.005.29
MIRNet* [14]22.680.82825.050.92321.360.63226.210.76928.640.88828.990.869785.0031.76
Restormer* [3]18.600.78921.410.83122.010.64528.580.80628.490.89227.990.868140.9926.11
MambaIR* [16]20.450.84525.650.93922.020.65828.410.80525.140.87627.530.85160.664.30
URWKV (Ours)*23.110.87426.360.94423.110.67329.440.82631.240.91129.990.88718.342.25
", + "image_path": "58c596f6f5916acdc63cca39ca180aa6a80830005f6745fa5da7db32054a291a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 324, + 294, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 294, + 384 + ], + "type": "text", + "content": "where Avg denotes the channel-wise mean operation. AdaAlign combines both upsampling and downsampling to adaptively align multiple states with the target state. Stack is the operation that stacks the aligned features along the channel axis." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 384, + 295, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 295, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 295, + 504 + ], + "type": "text", + "content": "To facilitate more nuanced fusion of local details and broader context, we incorporate an inception block [37] with multiple convolutional filters " + }, + { + "bbox": [ + 55, + 384, + 295, + 504 + ], + "type": "inline_equation", + "content": "(1\\times 1,3\\times 3," + }, + { + "bbox": [ + 55, + 384, + 295, + 504 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 384, + 295, + 504 + ], + "type": "inline_equation", + "content": "5\\times 5)" + }, + { + "bbox": [ + 55, + 384, + 295, + 504 + ], + "type": "text", + "content": " to aggregate degradation patterns across various receptive fields. The aggregated feature " + }, + { + "bbox": [ + 55, + 384, + 295, + 504 + ], + "type": "inline_equation", + "content": "E_{agg}" + }, + { + "bbox": [ + 55, + 384, + 295, + 504 + ], + "type": "text", + "content": " is then projected via a convolutional layer, followed by a sigmoid function to produce the spatial weight. Therefore, taking the first stage of the decoder as an example, the naive concatenation-based skip connection for the URWKV block in the decoder, following the SSF module, can be extended as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 119, + 512, + 295, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 512, + 295, + 528 + ], + "spans": [ + { + "bbox": [ + 119, + 512, + 295, + 528 + ], + "type": "interline_equation", + "content": "D _ {1} ^ {\\prime} = \\left(\\left[ W _ {s} \\odot E _ {3}, D _ {1} \\right]\\right) W _ {p} \\tag {8}", + "image_path": "ebca6cf8f150d57545ba66790f6149fc8c14467beb4ef01cb4f564a4748a90ed.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 296, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 595 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 595 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 534, + 296, + 595 + ], + "type": "inline_equation", + "content": "W_{s}" + }, + { + "bbox": [ + 55, + 534, + 296, + 595 + ], + "type": "text", + "content": " is the spatial weight predicted by the SSF module, " + }, + { + "bbox": [ + 55, + 534, + 296, + 595 + ], + "type": "inline_equation", + "content": "E_{3}" + }, + { + "bbox": [ + 55, + 534, + 296, + 595 + ], + "type": "text", + "content": " is the encoder feature, and " + }, + { + "bbox": [ + 55, + 534, + 296, + 595 + ], + "type": "inline_equation", + "content": "D_{1}" + }, + { + "bbox": [ + 55, + 534, + 296, + 595 + ], + "type": "text", + "content": " is the input feature to the first stage of the decoder. " + }, + { + "bbox": [ + 55, + 534, + 296, + 595 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 534, + 296, + 595 + ], + "type": "text", + "content": " denotes element-wise multiplication, and " + }, + { + "bbox": [ + 55, + 534, + 296, + 595 + ], + "type": "inline_equation", + "content": "W_{p}" + }, + { + "bbox": [ + 55, + 534, + 296, + 595 + ], + "type": "text", + "content": " represents a convolutional layer applied to refine the features. " + }, + { + "bbox": [ + 55, + 534, + 296, + 595 + ], + "type": "inline_equation", + "content": "[\\cdot ,\\cdot ]" + }, + { + "bbox": [ + 55, + 534, + 296, + 595 + ], + "type": "text", + "content": " denote the concatenate operation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 137, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 137, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 137, + 618 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 624, + 182, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 182, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 182, + 637 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "We implement our model in PyTorch on a workstation equipped with an NVIDIA Tesla A40 GPU. The number of URWKV blocks is set to " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "N_{1} = 3" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "N_{2} = 2" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": ", while the number of input channels is set to " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "C = 32" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": ". The decay factor " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " is set to 0.5. During training, we use a unified loss function for all datasets, as in [1], which combines L1 Loss," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 324, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 445 + ], + "type": "text", + "content": "SSIM Loss, and Perceptual Loss. The Adam optimizer [38] is employed with " + }, + { + "bbox": [ + 313, + 324, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 313, + 324, + 555, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 324, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.99" + }, + { + "bbox": [ + 313, + 324, + 555, + 445 + ], + "type": "text", + "content": ", and the learning rate is initialized at " + }, + { + "bbox": [ + 313, + 324, + 555, + 445 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 324, + 555, + 445 + ], + "type": "text", + "content": " and decays gradually to " + }, + { + "bbox": [ + 313, + 324, + 555, + 445 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 313, + 324, + 555, + 445 + ], + "type": "text", + "content": " using cosine annealing decay. Following [1, 10], we apply augmentation techniques, such as flipping and rotation, to each image pair. During testing, our URWKV model processes input images of arbitrary shape directly, without the need for cropping or resizing. For evaluation, we employ commonly-used PSNR [39] and SSIM [40] metrics to calculate images in the RGB color space across all datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 454, + 400, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 400, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 400, + 466 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 472, + 554, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 554, + 520 + ], + "type": "text", + "content": "To assess the performance of various models in scenarios involving dynamically coupled degradations, we compare the LLIE models, LLIE-deblur models, and unified models across 8 typical low-light benchmarks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 521, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 628 + ], + "type": "text", + "content": "LOL. As shown in Table 1, our URWKV demonstrates superior performance compared to LLIE models on LOLv2-real [41], outperforming Retinexformer by 0.32 dB in PSNR and 0.035 in SSIM, respectively. Similarly, on LOLv2-syn [41], our model outperforms RetinexMamba by 0.48 dB in PSNR and 0.011 in SSIM. Notably, LLIE-deblur models [2, 13] and other unified models [3, 14, 16] generally underperform on both datasets, with substantial performance gaps between them and our model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "SID and SMID. As shown in Table 1, our UR-WKV model delivers superior performance, closely trailing Retinexformer on SID [42], while outperforming it by 0.29 dB in PSNR on SMID [42]. In contrast, LLIE-deblur models exhibit inconsistent performance across the datasets. For instance, while PDHAT ranks second on SMID, it significantly underperforms on SID, achieving a PSNR of just" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21272" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 128, + 200 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 128, + 200 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 128, + 200 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 128, + 200 + ], + "type": "image", + "image_path": "886ba7b69e60c419b10d4989571a9d09570d7b3eb7ad0241a0945d26d5297592.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 205, + 505, + 217 + ], + "lines": [ + { + "bbox": [ + 102, + 205, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 102, + 205, + 505, + 217 + ], + "type": "text", + "content": "Figure 5. Visual comparison of state-of-the-art models on the SID dataset (top) and the SMID dataset (bottom)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 129, + 71, + 199, + 199 + ], + "blocks": [ + { + "bbox": [ + 129, + 71, + 199, + 199 + ], + "lines": [ + { + "bbox": [ + 129, + 71, + 199, + 199 + ], + "spans": [ + { + "bbox": [ + 129, + 71, + 199, + 199 + ], + "type": "image", + "image_path": "a05e1f1bef375b59e77999d908dcfa5d1b7a6043de9e6bfe59d419ec048d0090.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 199, + 72, + 268, + 198 + ], + "blocks": [ + { + "bbox": [ + 199, + 72, + 268, + 198 + ], + "lines": [ + { + "bbox": [ + 199, + 72, + 268, + 198 + ], + "spans": [ + { + "bbox": [ + 199, + 72, + 268, + 198 + ], + "type": "image", + "image_path": "e6c0bfd964b3a757c6ddbb480f90954c1a4e138d640cedfd318170312ecea311.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 269, + 72, + 339, + 198 + ], + "blocks": [ + { + "bbox": [ + 269, + 72, + 339, + 198 + ], + "lines": [ + { + "bbox": [ + 269, + 72, + 339, + 198 + ], + "spans": [ + { + "bbox": [ + 269, + 72, + 339, + 198 + ], + "type": "image", + "image_path": "7bd1cc1bfc5006a005c7d1dac10c4671d890d947b8df48b309f01321d7a15b55.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 339, + 72, + 408, + 198 + ], + "blocks": [ + { + "bbox": [ + 339, + 72, + 408, + 198 + ], + "lines": [ + { + "bbox": [ + 339, + 72, + 408, + 198 + ], + "spans": [ + { + "bbox": [ + 339, + 72, + 408, + 198 + ], + "type": "image", + "image_path": "6461dd32c502b1a38d651f6570d1f165a9641a4ec5bf171af68475d661589c66.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 408, + 72, + 477, + 198 + ], + "blocks": [ + { + "bbox": [ + 408, + 72, + 477, + 198 + ], + "lines": [ + { + "bbox": [ + 408, + 72, + 477, + 198 + ], + "spans": [ + { + "bbox": [ + 408, + 72, + 477, + 198 + ], + "type": "image", + "image_path": "8810ec3c47ad147d7173a76a86ed5783e9d4f2aaac910fdad4ca4aeaef928755.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 478, + 72, + 549, + 198 + ], + "blocks": [ + { + "bbox": [ + 478, + 72, + 549, + 198 + ], + "lines": [ + { + "bbox": [ + 478, + 72, + 549, + 198 + ], + "spans": [ + { + "bbox": [ + 478, + 72, + 549, + 198 + ], + "type": "image", + "image_path": "2727e65419c06056dd0ae2ef9a8f74e705dedc31b8827668579f898be891ae3e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 228, + 135, + 314 + ], + "blocks": [ + { + "bbox": [ + 58, + 228, + 135, + 314 + ], + "lines": [ + { + "bbox": [ + 58, + 228, + 135, + 314 + ], + "spans": [ + { + "bbox": [ + 58, + 228, + 135, + 314 + ], + "type": "image", + "image_path": "c5ed6aba216577cf3d04fb777f918cf1d7276c9579a6d87c167ed506dae50421.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 319, + 275, + 331 + ], + "lines": [ + { + "bbox": [ + 74, + 319, + 275, + 331 + ], + "spans": [ + { + "bbox": [ + 74, + 319, + 275, + 331 + ], + "type": "text", + "content": "Figure 6. Visual comparisons on SDSD-indoor dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 136, + 228, + 212, + 314 + ], + "blocks": [ + { + "bbox": [ + 136, + 228, + 212, + 314 + ], + "lines": [ + { + "bbox": [ + 136, + 228, + 212, + 314 + ], + "spans": [ + { + "bbox": [ + 136, + 228, + 212, + 314 + ], + "type": "image", + "image_path": "96103177d1aac695356b72edb2d217cba9d1641f4e80c6bb8c68ac315a6e3be2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 212, + 228, + 291, + 314 + ], + "blocks": [ + { + "bbox": [ + 212, + 228, + 291, + 314 + ], + "lines": [ + { + "bbox": [ + 212, + 228, + 291, + 314 + ], + "spans": [ + { + "bbox": [ + 212, + 228, + 291, + 314 + ], + "type": "image", + "image_path": "6f7ca234c2a4cd38834d0efdb1c5e692939069f958f3879088cb64c916b2d88c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 342, + 295, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 378 + ], + "type": "text", + "content": "21.93 dB, which is 1.18 dB lower than our model. Visual comparisons in Fig. 5 also demonstrate the excellent denoising capability of our model when enhancing brightness." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 378, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 295, + 498 + ], + "type": "text", + "content": "SDSD. Results depicted in Table 1 indicate that even some recent LLIE models [1, 26] still encounter challenges on SDSD [43]. In contrast, our URWKV model delivers state-of-the-art performance across both subsets. Moreover, compared to unified models like Restormer, URWKV outperforms it by " + }, + { + "bbox": [ + 55, + 378, + 295, + 498 + ], + "type": "inline_equation", + "content": "2.75\\mathrm{dB}" + }, + { + "bbox": [ + 55, + 378, + 295, + 498 + ], + "type": "text", + "content": " in PSNR on SDSD-indoor and by " + }, + { + "bbox": [ + 55, + 378, + 295, + 498 + ], + "type": "inline_equation", + "content": "2.00\\mathrm{dB}" + }, + { + "bbox": [ + 55, + 378, + 295, + 498 + ], + "type": "text", + "content": " in PSNR on SDSD-outdoor. As demonstrated in Fig. 6, our method preserves high-fidelity restoration in the scenario with strong light-dark contrast, whereas other models (e.g. PDHAT) generate unexpected artifacts." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 498, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 605 + ], + "type": "text", + "content": "FiveK. The MIT-Adobe FiveK dataset [44] challenges models with the intricate task of color restoration in lowlight and underexposed images. As reported in Table 2, our model achieves the highest PSNR of " + }, + { + "bbox": [ + 55, + 498, + 295, + 605 + ], + "type": "inline_equation", + "content": "26.08\\mathrm{dB}" + }, + { + "bbox": [ + 55, + 498, + 295, + 605 + ], + "type": "text", + "content": " on FiveK dataset, surpassing LLFormer by " + }, + { + "bbox": [ + 55, + 498, + 295, + 605 + ], + "type": "inline_equation", + "content": "0.48\\mathrm{dB}" + }, + { + "bbox": [ + 55, + 498, + 295, + 605 + ], + "type": "text", + "content": " and Retinexformer by " + }, + { + "bbox": [ + 55, + 498, + 295, + 605 + ], + "type": "inline_equation", + "content": "1.50\\mathrm{dB}" + }, + { + "bbox": [ + 55, + 498, + 295, + 605 + ], + "type": "text", + "content": ". Furthermore, URWKV outperforms the unified model, Restormer, in SSIM by 0.007. These notable improvements underscore URWKV's proficiency in restoring both structural integrity and aesthetic quality." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 606, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 295, + 713 + ], + "type": "text", + "content": "LOL-blur. The LOL-blur [13] dataset couples varying degrees of low-light and motion blur, which has often been underrepresented in other LLIE datasets. As shown in Table 3, conventional LLIE models typically struggle to address this coupled degradation, while unified models also fail to outperform specialized LLIE-deblur models. In contrast, our URWKV model achieves the highest performance in this scenario, with a PSNR of 27.27 dB and SSIM of 0.890, outperforming the second-best LLIE-Deblur model" + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 317, + 242, + 553, + 329 + ], + "blocks": [ + { + "bbox": [ + 324, + 228, + 542, + 239 + ], + "lines": [ + { + "bbox": [ + 324, + 228, + 542, + 239 + ], + "spans": [ + { + "bbox": [ + 324, + 228, + 542, + 239 + ], + "type": "text", + "content": "Table 2. Experimental results on MIT-Adobe FiveK dataset." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 242, + 553, + 329 + ], + "lines": [ + { + "bbox": [ + 317, + 242, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 317, + 242, + 553, + 329 + ], + "type": "table", + "html": "
MethodsEnlightenGAN [35]RUAS [45]RetinexNet [8]SNR-Net [12]LLFormer [10]BiFormer [1]
PSNR ↑13.2620.9116.3224.3925.6025.51
SSIM ↑0.7450.8630.7820.8950.9170.929
MethodsRetinexformer [9]LEDNet [13]PDHAT [2]Restormer [3]MIRNet [14]URWKV (Ours)
PSNR ↑24.5824.3425.0025.4823.7326.08
SSIM ↑0.8860.9150.9290.9290.8950.936
", + "image_path": "40e1b4577bd18624f7195585fc0065c676da7f3f4928669b648b3e0369fbe20a.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 317, + 352, + 553, + 439 + ], + "blocks": [ + { + "bbox": [ + 340, + 338, + 526, + 348 + ], + "lines": [ + { + "bbox": [ + 340, + 338, + 526, + 348 + ], + "spans": [ + { + "bbox": [ + 340, + 338, + 526, + 348 + ], + "type": "text", + "content": "Table 3. Experimental results on LOL-blur dataset." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 352, + 553, + 439 + ], + "lines": [ + { + "bbox": [ + 317, + 352, + 553, + 439 + ], + "spans": [ + { + "bbox": [ + 317, + 352, + 553, + 439 + ], + "type": "table", + "html": "
MethodsFourLLIE [26]UHDFour [27]LLFormer [10]Retinexformer [9]BiFormer [1]GLARE [46]
PSNR ↑19.8125.3424.5525.2524.6723.26
SSIM ↑0.6830.8290.7850.8210.8130.690
MethodsLEDNet [13]PDHAT [2]Restormer [3]MIRNet [14]MambaIR [16]URWKV (Ours)
PSNR ↑26.0626.7126.3823.9926.2827.27
SSIM ↑0.8460.8790.8600.7740.8480.890
", + "image_path": "250d159b308e01001fe24d2696326f9e50472d7788dd66c76f4eb1c4f9531188.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 449, + 553, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 553, + 485 + ], + "type": "text", + "content": "PDHAT by 0.56 dB in PSNR and 0.011 in SSIM. Examples in Fig. 7 vividly demonstrate the potential of our model in restoring coupled low-light and blur images." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 492, + 408, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 408, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 408, + 506 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 510, + 553, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 553, + 545 + ], + "type": "text", + "content": "We conduct several ablation studies on the real-world dataset (i.e., LOL-v2-real) to analyze the impact of various components in our model." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": "Effectiveness of LAN and SSF modules. To validate the individual and combined contributions of the LAN and SSF modules, we evaluate four model variants. As shown in Table 4, models incorporating either LAN or SSF individually show a clear improvement over the baseline (first row in Table 4). Among these, the model with LAN alone achieves a more pronounced gain, with PSNR increasing by " + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "inline_equation", + "content": "1.38\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": " over the baseline, highlighting the critical role of LAN in dynamically adapting to complex luminance variations. Moreover, our final model, which combines both LAN and SSF, incurs only a minor increase in parameters (from 1.64M to 2.25M) and a minimal rise in FLOPs (from 18.25G to 18.34G), yet delivers a significant performance boost compared to baseline. The feature transformation vi" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21273" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 550, + 205 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 550, + 205 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 550, + 205 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 550, + 205 + ], + "type": "image", + "image_path": "179d7a682ac682a4cdf73291c73ac81b0c0d28c915755566dc436843f8ca3f4e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 166, + 211, + 443, + 222 + ], + "lines": [ + { + "bbox": [ + 166, + 211, + 443, + 222 + ], + "spans": [ + { + "bbox": [ + 166, + 211, + 443, + 222 + ], + "type": "text", + "content": "Figure 7. Visual comparison of state-of-the-art models on LOL-blur dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 233, + 293, + 304 + ], + "blocks": [ + { + "bbox": [ + 57, + 233, + 293, + 304 + ], + "lines": [ + { + "bbox": [ + 57, + 233, + 293, + 304 + ], + "spans": [ + { + "bbox": [ + 57, + 233, + 293, + 304 + ], + "type": "image", + "image_path": "5bdae92335af02b211aed70c9727ef64651c3a7eb7dea982b33b920a03db634d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 308, + 295, + 343 + ], + "lines": [ + { + "bbox": [ + 55, + 308, + 295, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 295, + 343 + ], + "type": "text", + "content": "Figure 8. Top: LAN adaptively enhances the luminance, making details clearer. Bottom: SSF helps mitigate introducing degradation, obtaining fused feature " + }, + { + "bbox": [ + 55, + 308, + 295, + 343 + ], + "type": "inline_equation", + "content": "(D_3^{\\prime})" + }, + { + "bbox": [ + 55, + 308, + 295, + 343 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 55, + 308, + 295, + 343 + ], + "type": "inline_equation", + "content": "E_{1}" + }, + { + "bbox": [ + 55, + 308, + 295, + 343 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 308, + 295, + 343 + ], + "type": "inline_equation", + "content": "E_{2}" + }, + { + "bbox": [ + 55, + 308, + 295, + 343 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 308, + 295, + 343 + ], + "type": "inline_equation", + "content": "E_{3}" + }, + { + "bbox": [ + 55, + 308, + 295, + 343 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 352, + 295, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 295, + 375 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 295, + 375 + ], + "type": "text", + "content": "sualizations presented in Fig. 8 also validate the inherent advantages of LAN and SSF." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 377, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 377, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 54, + 377, + 295, + 555 + ], + "type": "text", + "content": "Multi-state Quad-directional Token Shift. Our propose multi-state aggregation strategy enhances the original Q-shift, allowing for a more comprehensive understanding of degradation representation. As shown in Table 5, our results indicate that single-state aggregation alone (Table 5a) achieves performance comparable to the model using Q-shift alone (Table 5c). In contrast, multi-state aggregation alone (Table 5b) consistently demonstrates superior performance, attaining a " + }, + { + "bbox": [ + 54, + 377, + 295, + 555 + ], + "type": "inline_equation", + "content": "0.13\\mathrm{dB}" + }, + { + "bbox": [ + 54, + 377, + 295, + 555 + ], + "type": "text", + "content": " higher PSNR. When combining single-state aggregation with Q-shift (Table 5d), the model shows an improvement in PSNR. However, it still falls short in SSIM by 0.003 compared to multi-state aggregation alone. Ultimately, our final model (Table 5e), which integrates both Q-shift and multi-state aggregation, demonstrates substantial advantages over all other variants." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "SSF vs. Navie Skip Connections. The ablation results in Table 6 highlight the effectiveness of SSF compared to naive skip connections, including simple addition (Add) and concatenation (Cat). When using multi-state encoder features, noise from earlier stages and semantic gaps between states lead to performance degradation in both PSNR and SSIM. In contrast, our SSF module, based on state-aware fusion strategy, avoids directly merging multistate features and predicts a guiding weight to selectively fuse useful patterns, mitigating issues of noise propagation and semantic misalignment. Notably, the \"Multi-state+SSF\" (i.e., our URWKV) yields significant improvements in PSNR and SSIM over \"Single-state+Cat\" without" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 326, + 248, + 538, + 306 + ], + "blocks": [ + { + "bbox": [ + 340, + 233, + 527, + 244 + ], + "lines": [ + { + "bbox": [ + 340, + 233, + 527, + 244 + ], + "spans": [ + { + "bbox": [ + 340, + 233, + 527, + 244 + ], + "type": "text", + "content": "Table 4. Ablation study on LAN and SSF modules." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 248, + 538, + 306 + ], + "lines": [ + { + "bbox": [ + 326, + 248, + 538, + 306 + ], + "spans": [ + { + "bbox": [ + 326, + 248, + 538, + 306 + ], + "type": "table", + "html": "
LANSSFPSNR ↑SSIM ↑ParamsFLOPs
21.330.8561.64M18.25G
✓21.400.8611.65M18.29G
✓22.710.8692.25M18.30G
✓✓23.110.8742.25M18.34G
", + "image_path": "34e1f958d1d0a377a167ecb8c8be554719676c2cabe7a22a9a942ef4fc81a1d2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 320, + 330, + 552, + 400 + ], + "blocks": [ + { + "bbox": [ + 314, + 315, + 553, + 326 + ], + "lines": [ + { + "bbox": [ + 314, + 315, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 314, + 315, + 553, + 326 + ], + "type": "text", + "content": "Table 5. Ablation study on multi-state quad-directional token shift." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 330, + 552, + 400 + ], + "lines": [ + { + "bbox": [ + 320, + 330, + 552, + 400 + ], + "spans": [ + { + "bbox": [ + 320, + 330, + 552, + 400 + ], + "type": "table", + "html": "
#Single-stateMulti-stateQ-shiftPSNR ↑SSIM ↑
(a)✓22.120.861
(b)✓22.250.870
(c)✓22.190.862
(d)✓✓22.290.867
(e)✓✓23.110.874
", + "image_path": "7fc0e2d260f5149d74b89e9f5caf7ac9be2999a66b33551298c31b8a88b4f840.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 318, + 435, + 552, + 497 + ], + "blocks": [ + { + "bbox": [ + 313, + 409, + 553, + 431 + ], + "lines": [ + { + "bbox": [ + 313, + 409, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 553, + 431 + ], + "type": "text", + "content": "Table 6. Impacts of SSF and navie skip connections (Skip Conn.) with different state numbers (State No.)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 435, + 552, + 497 + ], + "lines": [ + { + "bbox": [ + 318, + 435, + 552, + 497 + ], + "spans": [ + { + "bbox": [ + 318, + 435, + 552, + 497 + ], + "type": "table", + "html": "
State No.Skip Conn.PSNR ↑SSIM ↑ParamsFLOPs
Single-stateAdd22.660.8721.86M14.68G
Single-stateCat22.710.8692.25M18.30G
Multi-stateCat21.830.8592.50M23.89G
Multi-stateSSF23.110.8742.25M18.34G
", + "image_path": "354e4fbe8b9aaf4dcfdd7aee3a51abf2ffa5145f38e6bfb31dbdd22190a6a1dd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 507, + 553, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 553, + 532 + ], + "type": "text", + "content": "increasing the model's parameter, demonstrating both enhanced performance and computational efficiency." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 558, + 555, + 713 + ], + "type": "text", + "content": "To address the shortcomings of existing models, we propose URWKV with multi-state perspective, enabling flexible and effective restoration across dynamically coupled degradation scenarios. Specifically, we introduce three key innovations: luminance-adaptive normalization (LAN) for adaptive, scene-aware luminance modulation; an exponential moving average approach to aggregate intra-stage states for spatial-level token-shift; and the state-aware selective fusion (SSF) module for dynamic alignment and integration of multi-state features across encoder stages. In comparison to existing models, our model demonstrates superior performance while requiring significantly fewer parameters and computational resources." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21274" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 217 + ], + "type": "text", + "content": "Acknowledgments. This work was supported in part by the National Natural Science Foundation of China under Grant 62471142, in part by the Natural Science Foundation of Fujian Province, China under Grants 2023J01067, in part by Industry-Academy Cooperation Project under Grant 2024H6006, in part by the Collaborative Innovation Platform Project of Fuzhou City under Grant 2023-P-002, in part by the Key Technology Innovation Project for Focused Research and Industrialization in the Software Industry of Fujian Province, and in part by the Key Research and Industrialization Project of Technological Innovation in Fujian Province under Grant 2024XQ002." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 229, + 115, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 229, + 115, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 115, + 241 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 249, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 61, + 249, + 294, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 249, + 294, + 301 + ], + "spans": [ + { + "bbox": [ + 61, + 249, + 294, + 301 + ], + "type": "text", + "content": "[1] Rui Xu, Yuezhou Li, Yuzhen Niu, Huangbiao Xu, Yuzhong Chen, and Tiesong Zhao. Bilateral interaction for local-global collaborative perception in low-light image enhancement. IEEE Transactions on Multimedia, 2024. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 306, + 294, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 306, + 294, + 359 + ], + "spans": [ + { + "bbox": [ + 61, + 306, + 294, + 359 + ], + "type": "text", + "content": "[2] Yuezhou Li, Rui Xu, Yuzhen Niu, Wenzhong Guo, and Tiesong Zhao. Perceptual decoupling with heterogeneous auxiliary tasks for joint low-light image enhancement and deblurring. IEEE Transactions on Multimedia, pages 6663-6675, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 363, + 294, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 363, + 294, + 427 + ], + "spans": [ + { + "bbox": [ + 62, + 363, + 294, + 427 + ], + "type": "text", + "content": "[3] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 430, + 294, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 430, + 294, + 483 + ], + "spans": [ + { + "bbox": [ + 62, + 430, + 294, + 483 + ], + "type": "text", + "content": "[4] Andong Lu, Zhang Zhang, Yan Huang, Yifan Zhang, Chenglong Li, Jin Tang, and Liang Wang. Illumination distillation framework for nighttime person re-identification and a new benchmark. IEEE Transactions on Multimedia, pages 1–14, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 487, + 294, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 294, + 519 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 294, + 519 + ], + "type": "text", + "content": "[5] Xiaopeng Li and Xiaojie Guo. SPN2D-GAN: Semantic prior based night-to-day image-to-image translation. IEEE Transactions on Multimedia, pages 1-15, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 522, + 294, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 522, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 62, + 522, + 294, + 564 + ], + "type": "text", + "content": "[6] Xuechen Bai, Xinghua Li, Jianhao Miao, and Huanfeng Shen. Making the rarth clear at night: A high-resolution nighttime light image deblooming network. IEEE Transactions on Geoscience and Remote Sensing, 61:1-13, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 567, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 294, + 610 + ], + "type": "text", + "content": "[7] Yonghua Zhang, Jiawan Zhang, and Xiaojie Guo. Kindling the darkness: A practical low-light image enhancer. In Proceedings of the ACM International Conference on Multimedia, pages 1632-1640, 2019. 1, 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 613, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 294, + 655 + ], + "type": "text", + "content": "[8] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. In Proceedings of the British Machine Vision Conference, 2018. 1, 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "type": "text", + "content": "[9] Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang. Retinexformer: One-stage retina-based transformer for low-light image enhancement. In Proceedings of the IEEE International Conference on Computer Vision, pages 12504-12513, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 317, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 317, + 73, + 553, + 127 + ], + "type": "text", + "content": "[10] Tao Wang, Kaihao Zhang, Tianrun Shen, Wenhan Luo, Bjorn Stenger, and Tong Lu. Ultra-high-definition low-light image enhancement: A benchmark and transformer-based method. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 2654-2662, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 129, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 160 + ], + "type": "text", + "content": "[11] Jiesong Bai, Yuhao Yin, and Qiyuan He. Retinexmamba: Retinex-based mamba for low-light image enhancement. arXiv preprint arXiv:2405.03349, 2024. 1, 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 163, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 205 + ], + "type": "text", + "content": "[12] Xiaogang Xu, Ruixing Wang, Chi-Wing Fu, and Jiaya Jia. SNR-aware low-light image enhancement. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 17714-17724, 2022. 2, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 206, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 206, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 317, + 206, + 553, + 249 + ], + "type": "text", + "content": "[13] Shangchen Zhou, Chongyi Li, and Chen Change Loy. LEDNet: Joint low-light enhancement and deblurring in the dark. In Proceedings of the European Conference on Computer Vision, pages 573-589, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 251, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 251, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 553, + 304 + ], + "type": "text", + "content": "[14] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Learning enriched features for real image restoration and enhancement. In Proceedings of the European Conference on Computer Vision, pages 492-511, 2020. 2, 3, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 306, + 553, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 306, + 553, + 359 + ], + "spans": [ + { + "bbox": [ + 317, + 306, + 553, + 359 + ], + "type": "text", + "content": "[15] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12299-12310, 2021. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 361, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 553, + 414 + ], + "type": "text", + "content": "[16] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. MambaIR: A simple baseline for image restoration with state-space model. In Proceedings of the European Conference on Computer Vision, pages 222-241, 2025. 2, 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 416, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 416, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 317, + 416, + 553, + 469 + ], + "type": "text", + "content": "[17] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 17683-17693, 2022. 2, 4, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 472, + 553, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 523 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 523 + ], + "type": "text", + "content": "[18] Bo Peng, Eric Alcaide, Quentin Anthony, Alon Albalak, Samuel Arcadinho, Stella Biderman, Huanqi Cao, Xin Cheng, Michael Chung, Matteo Grella, et al. RWKV: Reinventing rnns for the transformer era. arXiv preprint arXiv:2305.13048, 2023. 2.4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 526, + 553, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 579 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 579 + ], + "type": "text", + "content": "[19] Yuchen Duan, Weiyun Wang, Zhe Chen, Xizhou Zhu, Lewei Lu, Tong Lu, Yu Qiao, Hongsheng Li, Jifeng Dai, and Wenhai Wang. Vision-RWKV: Efficient and scalable visual perception with RWKV-like architectures. arXiv preprint arXiv:2403.02308, 2024. 2, 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 581, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 581, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 581, + 553, + 613 + ], + "type": "text", + "content": "[20] J Alex Stark. Adaptive image contrast enhancement using generalizations of histogram equalization. IEEE Transactions on Image Processing, 9:889-896, 2000. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 615, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 615, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 615, + 553, + 656 + ], + "type": "text", + "content": "[21] Mohammad Abdullah-Al-Wadud, Md Hasanul Kabir, M Ali Akber Dewan, and Oksam Chae. A dynamic histogram equalization for image contrast enhancement. IEEE Transactions on Consumer Electronics, 53:593-600, 2007." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 658, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 689 + ], + "type": "text", + "content": "[22] Turgay Celik and Tardi Tjahjadi. Contextual and variational contrast enhancement. IEEE Transactions on Image Processing, 20:3431-3441, 2011. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[23] Xueyang Fu, Yinghao Liao, Delu Zeng, Yue Huang, Xiao-Ping Zhang, and Xinghao Ding. A probabilistic method for" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21275" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "image enhancement with simultaneous illumination and reflectance estimation. IEEE Transactions on Image Processing, 24:4965-4977, 2015. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 151 + ], + "type": "text", + "content": "[24] Mading Li, Jiaying Liu, Wenhan Yang, Xiaoyan Sun, and Zongming Guo. Structure-revealing low-light image enhancement via robust retinae model. IEEE Transactions on Image Processing, 27:2828-2841, 2018." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 196 + ], + "type": "text", + "content": "[25] Daniel J Jobson, Zia-ur Rahman, and Glenn A Woodell. A multiscale retina for bridging the gap between color images and the human observation of scenes. IEEE Transactions on Image processing, 6:965–976, 1997. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 295, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 241 + ], + "type": "text", + "content": "[26] Chenxi Wang, Hongjun Wu, and Zhi Jin. FourLLIE: Boosting low-light image enhancement by fourier frequency information. In Proceedings of the ACM International Conference on Multimedia, pages 7459-7469, 2023. 3, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 295, + 297 + ], + "type": "text", + "content": "[27] Chongyi Li, Chun-Le Guo, Man Zhou, Zhexin Liang, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Embedding fourier for ultra-high-definition low-light image enhancement. Proceedings of the International Conference on Learning Representations, 2023. 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 298, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 295, + 342 + ], + "type": "text", + "content": "[28] Yuning Cui, Wenqi Ren, Xiaochun Cao, and Alois Knoll. Revitalizing convolutional network for image restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 9423-9438, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 295, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 295, + 397 + ], + "type": "text", + "content": "[29] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image restoration using swin transformer. In Proceedings of the IEEE International Conference on Computer Vision, pages 1833-1844, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 295, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 295, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 295, + 454 + ], + "type": "text", + "content": "[30] Akshay Dudhane, Syed Waqas Zamir, Salman Khan, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Burstormer: Burst image restoration and enhancement transformer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5703-5712, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 295, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 295, + 499 + ], + "type": "text", + "content": "[31] Man Zhou, Jie Huang, Chun-Le Guo, and Chongyi Li. Fourmer: An efficient global modeling paradigm for image restoration. In Proceedings of the International Conference on Machine Learning, pages 42589-42601, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 295, + 533 + ], + "type": "text", + "content": "[32] Yuning Cui, Wenqi Ren, Xiaochun Cao, and Alois Knoll. Image restoration via frequency selection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 295, + 578 + ], + "type": "text", + "content": "[33] Francisco Lara, Paula Bernal-Molina, Vicente Fernandez-Sanchez, and Norberto López-Gil. Changes in the objective amplitude of accommodation with pupil size. Optometry and Vision Science, 91(10):1215-1220, 2014. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 580, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 295, + 623 + ], + "type": "text", + "content": "[34] Pablo A Barrionuevo, Luis A Issolio, and Constanza Tripolone. Photoreceptor contributions to the human pupil light reflex. Journal of Photochemistry and Photobiology, 15:100178, 2023. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 678 + ], + "type": "text", + "content": "[35] Yifan Jiang, Xinyu Gong, Ding Liu, Yu Cheng, Chen Fang, Xiaohui Shen, Jianchao Yang, Pan Zhou, and Zhangyang Wang. EnlightenGAN: Deep light enhancement without paired supervision. IEEE Transactions on Image Processing, 30:2340-2349, 2021. 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 680, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 295, + 714 + ], + "type": "text", + "content": "[36] Haonan Wang, Peng Cao, Jiaqi Wang, and Osmar R Zaiane. UCTransNet: Rethinking the skip connections in U-Net from a channel-wise perspective with transformer. In Proceedings" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 588 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "of the AAAI conference on artificial intelligence, volume 36, pages 2441-2449, 2022. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 97, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 162 + ], + "type": "text", + "content": "[37] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1-9, 2015. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "text", + "content": "[38] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 229 + ], + "type": "text", + "content": "[39] Quan Huynh-Thu and Mohammed Ghanbari. Scope of validity of PSNR in image/video quality assessment. *Electronics Letters*, 44:800–801, 2008. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "text", + "content": "[40] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "text", + "content": "[41] Wenhan Yang, Wenjing Wang, Haofeng Huang, Shiqi Wang, and Jiaying Liu. Sparse gradient regularized deep retina network for robust low-light image enhancement. IEEE Transactions on Image Processing, 30:2072-2086, 2021. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 321, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 363 + ], + "type": "text", + "content": "[42] Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3291-3300, 2018. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 365, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 420 + ], + "type": "text", + "content": "[43] Ruixing Wang, Xiaogang Xu, Chi-Wing Fu, Jiangbo Lu, Bei Yu, and Jiaya Jia. Seeing dynamic scene in the dark: A high-quality video dataset with mechatronic alignment. In Proceedings of the IEEE International Conference on Computer Vision, pages 9700-9709, 2021. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 421, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 475 + ], + "type": "text", + "content": "[44] Vladimir Bychkovsky, Sylvain Paris, Eric Chan, and Frédo Durand. Learning photographic global tonal adjustment with a database of input/output image pairs. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 97-104, 2011. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 476, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 476, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 316, + 476, + 553, + 531 + ], + "type": "text", + "content": "[45] Risheng Liu, Long Ma, Tengyu Ma, Xin Fan, and Zhongxuan Luo. Learning with nested scene modeling and cooperative architecture search for low-light vision. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45:5953-5969, 2022. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 533, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 588 + ], + "type": "text", + "content": "[46] Han Zhou, Wei Dong, Xiaohong Liu, Shuaicheng Liu, Xiongkuo Min, Guangtao Zhai, and Jun Chen. GLARE: Low light image enhancement via generative latent feature based codebook retrieval. In Proceedings of the European Conference on Computer Vision, pages 36-54, 2024. 7" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21276" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/USP-Gaussian_ Unifying Spike-based Image Reconstruction, Pose Correction and Gaussian Splatting/62c2f0f5-8ea0-4959-8e97-f59984bffdfe_content_list.json b/2025/USP-Gaussian_ Unifying Spike-based Image Reconstruction, Pose Correction and Gaussian Splatting/62c2f0f5-8ea0-4959-8e97-f59984bffdfe_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7707e9c18918e43cfd18c3683061ee1ebd3e8faa --- /dev/null +++ b/2025/USP-Gaussian_ Unifying Spike-based Image Reconstruction, Pose Correction and Gaussian Splatting/62c2f0f5-8ea0-4959-8e97-f59984bffdfe_content_list.json @@ -0,0 +1,1758 @@ +[ + { + "type": "text", + "text": "USP-Gaussian: Unifying Spike-based Image Reconstruction, Pose Correction and Gaussian Splatting", + "text_level": 1, + "bbox": [ + 109, + 128, + 887, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kang Chen $^{1,2}$ Jiyuan Zhang $^{1,2}$ Zecheng Hao $^{1,2}$ Yajing Zheng $^{1,2}$ Tiejun Huang $^{1,2,3}$ Zhaofei Yu $^{1,3*}$ $^{1}$ School of Computer Science, Peking University \n $^{2}$ State Key Laboratory for Multimedia Information Processing, Peking University \n $^{3}$ Institute for Artificial Intelligence, Peking University", + "bbox": [ + 107, + 200, + 895, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{mrchenkang, jyzhang}@stu.pku.edu.cn, {haozecheng,yj.zheng,tjhuang,yuzf12}@pku.edu.cn", + "bbox": [ + 125, + 277, + 864, + 292 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/333e4d02bd17b5e8661568d0d4bd749fc6ac67f7fe13fafa5e045a43de43365e.jpg", + "image_caption": [ + "Figure 1. Left. Illustration of our USP-Gaussian framework, where the spike-based image Reconstruction Network (Recon-Net), camera poses, and 3DGS are collaboratively optimized signified by $\\mathbb{O}$ . Mid. Visual ablation showcasing the performance of Recon-Net and 3DGS with and without (w/ & w/o) the joint optimization strategy, with the ablation table depicted in Tab. 4 and the input formulated in Eq. (6). Right. Training curve comparison for Recon-Net and 3DGS with and without joint optimization." + ], + "image_footnote": [], + "bbox": [ + 94, + 303, + 320, + 486 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/77cffedaf199051973ce374e95a5a7d4ebca62ec01f8947bcad9c164c736a129.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 330, + 304, + 640, + 487 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e22ac288bc8fe71cef5efdc783edaa7c29b0c2bce11325bc782d07e9b7a1d359.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 306, + 905, + 486 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 582, + 326, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Spike camera, as an innovative type of neuromorphic camera that captures scenes with 0-1 bit stream at $40\\mathrm{kHz}$ , is increasingly being employed for the novel view synthesis task building on techniques such as Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS). Previous spike-based approaches typically follow a three-stage pipeline: I. Spike-to-image reconstruction based on established algorithms. II. Camera poses estimation. III. Novel view synthesis. However, the cascading framework suffers from substantial cumulative errors, i.e., the quality of the initially reconstructed images will impact pose estimation, ultimately limiting the fidelity of the 3D reconstruction. To address this limitation, we propose a synergistic optimization framework USP-Gaussian, which unifies spike-to-image reconstruction, pose correction, and gaussian splatting into an end-to-end pipeline. Leveraging the multi-view consistency afforded by 3DGS and the motion capture capability of the", + "bbox": [ + 88, + 614, + 485, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "spike camera, our framework enables iterative optimization between the spike-to-image reconstruction network and 3DGS. Experiments on synthetic datasets demonstrate that our method surpasses previous approaches by effectively eliminating cascading errors. Moreover, in real-world scenarios, our method achieves robust 3D reconstruction benefiting from the integration of pose optimization. Our code, data, and trained models are available at https://github.com/chenkang455/USP-Gaussian.", + "bbox": [ + 511, + 569, + 908, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 737, + 645, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Spike cameras [12], inspired by the human retina [49], introduce an innovative paradigm for encoding high-speed dynamic scenes. In contrast to conventional imaging systems [22] that rely on fixed exposure intervals to integrate visual information, spike cameras continuously sense dynamics by recording the spike stream with an impressive firing rate of $40\\mathrm{kHz}$ . Benefiting from the efficient bit representation of spikes, spike cameras can adeptly detect rapid scenes with significantly diminished bandwidth compared", + "bbox": [ + 509, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 114, + 887, + 246, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16609", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to the high-speed cameras such as the Phantom.", + "bbox": [ + 89, + 90, + 406, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Capitalizing on the remarkable capability of the spike camera to capture intricate motion and texture details in high-speed environments, recent research [5, 36, 41, 44, 45, 48, 49] concentrate on establishing the mapping from bit-encoded spike to visually perceivable images. Subsequent investigations delve into the estimation of optical flow [43], facilitation of motion deblur [3, 8, 38] and applications such as high-speed object detection [46], motion magnification [33], depth estimation [34] and occlusion removal [35].", + "bbox": [ + 89, + 108, + 483, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Radiance Fields [21] and 3D Gaussian Splitting [13] have emerged as prominent novel view synthesis technologies benefiting from their designed differential training frameworks. Among them, 3DGS employs Gaussian primitives for explicit scene representation and the splatting algorithm for projection, thereby achieving faster rendering speed compared to the pixel-by-pixel approach implemented in NeRF. Building on NeRF and 3DGS, the high-speed and continuous capturing potential of the spike camera can be fully unleashed in 3D reconstruction tasks [11, 37, 50].", + "bbox": [ + 89, + 244, + 483, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While the idea of applying the spike camera for 3D reconstruction is trivial and straightforward in a sequential manner, i.e., we can reconstruct images from the spike stream, estimate camera poses and subsequently apply NeRF/3DGS to obtain the 3D representation. Nevertheless, the cascading pipeline introduces significant cumulative errors when the quality of initially recovered images is poor, thereby degrading the accuracy of camera pose estimation and limiting the fidelity of the reconstructed scene.", + "bbox": [ + 89, + 412, + 483, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For example, SpikeGS [37] initiates with a self-supervised spike-to-image reconstruction framework built on the Blind Spot Network (BSN) [15], intended to mitigate the performance degradation [6, 7] encountered by supervised learning methods on real-world scenes. With recovered high-quality images, SpikeGS designs a further constraint from the long spike input and integrates it into the vanilla 3DGS training pipeline. While SpikeGS have improved preliminary representation for 3DGS, it remains constrained within the cascading framework thus inheriting the aforementioned cumulative error problem.", + "bbox": [ + 89, + 551, + 482, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose USP-Gaussian, Unifying Spike-based image reconstruction, Pose correction, and Gaussian Splatting within a one-stage optimization framework to mitigate error propagation inherent in the three-stage cascading pipeline. Specifically, USP-Gaussian comprises two synergistic branches as illustrated in Fig. 1, one dedicated to spike-to-image reconstruction via Recon-Net and the other to novel view synthesis implemented with 3DGS. To facilitate the training of Recon-Net, we design $\\mathcal{L}_{\\mathrm{rec}}$ based on the constraint between the input spike and recovered outputs. Besides, we supervise the training of 3DGS based on $\\mathcal{L}_{\\mathrm{gs}}$ , which obviates the require", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ment for pre-recovered high-quality images. While each branch of USP-Gaussian operates independently, we further introduce the joint loss $\\mathcal{L}_{\\mathrm{joint}}$ derived from the outputs of both two branches to facilitate unified optimization.", + "bbox": [ + 511, + 90, + 906, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual ablation comparisons of restored images and the training curve depicted in Fig. 1 demonstrate that the joint optimization of the two branches mutually enhances the recovered textures and the overall training process. We conduct quantitative and qualitative evaluations on the synthetic dataset and the real-world dataset. Experiments showcase that our method achieves superior restoration in both datasets, validating the effectiveness of our joint learning framework. In summary, the contributions of our USP-Gaussian are as follows:", + "bbox": [ + 511, + 152, + 908, + 303 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a unified framework for the joint optimization of Recon-Net, Pose, and 3DGS, mitigating error amplification inherent in cascading frameworks.", + "- We demonstrate that spike-to-image and 3D reconstruction tasks can mutually facilitate and enhance the optimization of each other.", + "- We introduce a novel self-supervised training paradigm for the spike-based image reconstruction network with the multi-view consistency afforded by 3DGS." + ], + "bbox": [ + 511, + 304, + 903, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 455, + 655, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Spike-based Image Reconstruction", + "text_level": 1, + "bbox": [ + 511, + 482, + 815, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While neuromorphic cameras generate high frame rate bitstream information, it is imperative to transform them into multi-bit grayscale images that are perceptible to human observers. To achieve this spike-to-image target, a plethora of methodologies [5-7, 36, 41, 44, 45, 49] have been proposed, which can be broadly classified into model-based, supervised, and self-supervised learning paradigms. Model-based methods, rooted in the sampling principle of the spike camera, aim to infer high-fidelity images drawing from the physical spike-image relationship [48], biological retinal models [49], and short-term synaptic plasticity [45]. Confronted with pervasive dark current noise that is challenging to model, learning-based methods are employed to directly deduce the sharp image from spike input with unknown firing patterns. Specifically, Spk2ImgNet [41] designed a hierarchical architecture engineered to capture the intricate temporal dynamics present within spike streams. SpikeCLIP [5] utilized the textual description of the captured scene as supervision for training the spike-to-image network. Concurrently, self-supervised learning methodologies based on the BSN [6, 7] are developed to counteract the performance deterioration in supervised learning methods owing to the domain gap between synthetic and real-world datasets.", + "bbox": [ + 511, + 507, + 906, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel self-supervised spike-based image reconstruction framework leveraging the multi-view constraints provided by 3D scenes.", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16610", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Novel View Synthesis based on the Neuromorphic Camera", + "text_level": 1, + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent research has extensively employed neuromorphic cameras, including event cameras [10, 16, 18, 27, 47], which encode intensity difference in dynamic scenarios with low-latency and spike cameras [12], for 3D scene reconstruction.", + "bbox": [ + 89, + 128, + 483, + 204 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Event Camera. Klenk et al. [14] proposed the E-NeRF based on the event generation mechanism, which can recover sharp 3D scenes under high-speed camera motion. Rudnev et al. [23] further harnessed a color event camera to learn the color representation of scenes. [1, 20] focused on recovering dynamic scenes with rigid deformations from event streams. In addition to NeRF, EvGGS [29] introduces a generalizable Gaussian splatting framework for event cameras designed to reconstruct 3D scenes without retraining. EF-3DGS [17] develops a free-trajectory framework by aligning the RGB input, event streams, and pose. Event3DGS [31] presents an effective solution for rapid ego-motion robotics scenarios.", + "bbox": [ + 89, + 204, + 483, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spike Camera. Zhu et al. [50] design a differentiable SNN and integrate it into the NeRF training pipeline to simulate the generation of spike streams, where the loss function is formulated based on the rendered spike output and the input spike. Dai et al. [9] further introduce SpikeNVS by extending the spike cameras to RGB-Spike input. Additionally, Guo et al. [11] introduced a color 3D scene reconstruction framework by integrating Bayer-pattern spike streams into the 3DGS training pipeline. SpikeGS [37] initially employs the BSN to recover images from the noisy spike stream and subsequently integrates them into the 3DGS.", + "bbox": [ + 89, + 400, + 483, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While the cascaded framework SpikeGS [37] introduces propagation errors, we propose a joint optimization framework that concurrently handles spike-based image reconstruction, pose estimation, and novel view synthesis.", + "bbox": [ + 89, + 566, + 483, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 638, + 217, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Spike Camera", + "text_level": 1, + "bbox": [ + 89, + 662, + 236, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The spike camera consists of the photodetector, voltage integrator, comparator, and synchronous readout as shown in Fig. 2. Within the designated time interval $\\mathcal{T} = [t_s,t_e]$ , the photodetector continuously captures the light intensity $\\mathbf{I}(t)$ and converts it into the voltage signal $\\mathbf{V}(t) = \\mu \\mathbf{I}(t)$ , where $\\mu$ is the photoelectric conversion coefficient. Concurrently, the comparator assesses the integrator voltage $\\mathbf{A}(t)$ against the predefined threshold $\\Theta$ . When the threshold is met, the spike is emitted and the integrator voltage is reset, formulated as:", + "bbox": [ + 89, + 686, + 483, + 834 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} (t) = \\int_ {t _ {s}} ^ {t} \\mathbf {I} (t) d t \\bmod C, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 834, + 482, + 867 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $C = \\Theta /\\mu$ is a constant proportional to the spike firing threshold $\\Theta$ . While the integrator continuously ag-", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5208623a539375b59ed4bbb1ffea88a52bd7f97c8fed35547893dfd0382f4a5f.jpg", + "image_caption": [ + "Figure 2. (a) The spike camera captures a railway moving 350 km/h. (b) The Recon-Net is designed to establish the mapping between the input spike stream and the sharp image. (c) The working mechanism of the spike camera." + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 906, + 289 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "gregates voltage, the readout circuit synchronously retrieves spike signals from the registers at up to $40\\mathrm{kHz}$ . Spikes fired during the specified interval $\\mathcal{T}$ are represented as a three-dimensional bit stream $S\\in \\{0,1\\}^{K\\times H\\times W}$ , where $H$ and $W$ indicate the image height and width, with $K$ denoting the spike sequence length.", + "bbox": [ + 511, + 382, + 906, + 473 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. 3D Gaussian Splatting", + "text_level": 1, + "bbox": [ + 511, + 483, + 723, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussian Splating leverages a set of three-dimensional Gaussian primitives, symbolized as $\\{\\mathcal{G}\\}_{n=1}^{N_{\\mathrm{gs}}}$ , to effectively represent the spatial characteristics of the captured scene.", + "bbox": [ + 511, + 506, + 905, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Each Gaussian primitive is centered at a point $\\mathbf{p}_n$ , with its spread and orientation influenced by the covariance matrix $\\Sigma_n = RSS^T R^T$ , where $R$ and $T$ denotes the rotation and the scale matrix, respectively. The impact of the Gaussian on any point $\\mathbf{v}$ in the scene is quantified by:", + "bbox": [ + 511, + 555, + 905, + 631 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {n} (\\mathbf {v}) = \\exp \\left(- \\frac {1}{2} (\\mathbf {v} - \\mathbf {p} _ {n}) ^ {T} \\boldsymbol {\\Sigma} _ {n} ^ {- 1} (\\mathbf {v} - \\mathbf {p} _ {n})\\right). (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 641, + 906, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When it comes to rendering for viewing transformation $W$ , these 3D Gaussian primitives are mapped onto 2D primitives $\\mathcal{G}_n^{2D}$ , with the covariance matrix $\\Sigma_n^\\prime$ reformulated as:", + "bbox": [ + 511, + 686, + 905, + 732 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma_ {n} ^ {\\prime} = J W \\Sigma_ {n} W ^ {T} J ^ {T}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 744, + 903, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $J$ represents the Jacobian matrix derived from the affine approximation of the projective transformation.", + "bbox": [ + 511, + 771, + 905, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Further, each pixel on the rendered plane is calculated by rasterizing these 2D Gaussians, sorted by depth, according to the following formulation:", + "bbox": [ + 511, + 801, + 905, + 847 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC (r) = \\sum_ {k = 1} ^ {N _ {\\mathrm {g s}}} T _ {i} \\alpha_ {i} \\mathbf {c} _ {i}, T _ {i} = \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {j}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 859, + 906, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16611", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1f235a959be2d89218507a7cb10908cc685b0584d3ab39210d469737fcc3e0a2.jpg", + "image_caption": [ + "Figure 3. The working pipeline of our USP-Gaussian. For each camera viewpoint, we initially derive a pose sequence at equal time intervals based on the camera model described in Sec. 4.3. Leveraging the 3DGS rendering pipeline detailed in Sec. 4.2, we subsequently generate a corresponding sharp sequence for this viewpoint. Concurrently, Recon-Net is employed to reconstruct the same sequence from the captured spike stream as outlined in Sec. 4.1. While Recon-Net and 3DGS are trained based on the discrepancy between the re-synthesized blurry image with the input long-exposure image, alignment of their outputs for the same scene is further enforced through joint optimization loss as described in Sec. 4.4. With formulated loss functions, our USP-Gaussian enables the collaborative optimization of Gaussian primitives, camera poses, and the parameters of Recon-Net as formulated in Eq. (5)." + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 906, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $D$ represents the indices set reordered by depth, and $D_{k}$ denotes the index of the Gaussian primitive that is the $k$ -th closest to the rendering point.", + "bbox": [ + 89, + 506, + 485, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Methods", + "text_level": 1, + "bbox": [ + 89, + 566, + 191, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Target Clarification. Given the captured spike stream set $\\{S_v\\}_{v=1}^V$ from $V$ different viewpoints with initial inaccurate poses $\\{\\mathcal{P}_v\\}_{v=1}^V$ , the USP-Gaussian aims to optimize the parameters of Recon-Net and Gaussian primitives while outputting refined poses as shown in Fig. 1, formulated as:", + "bbox": [ + 89, + 592, + 483, + 669 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n(\\mathbb {S}, \\mathbb {P}, \\mathbb {G}, \\theta) \\xrightarrow {\\text {U S P - G a u s s i a n}} (\\mathbb {P} ^ {*}, \\mathbb {G} ^ {*}, \\theta^ {*}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 681, + 482, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbb{S}$ and $\\mathbb{P}$ represent the sets of spike streams and poses from $M$ viewpoints, $\\mathbb{G}$ denotes the ensemble of Gaussian primitives, and $\\theta$ encapsulates the Recon-Net parameters. The notation $(^{*})$ indicates the optimized parameters.", + "bbox": [ + 89, + 713, + 483, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Spike-based Image Reconstruction Network", + "text_level": 1, + "bbox": [ + 89, + 786, + 465, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Recon-Net is designed to facilitate the mapping $\\mathcal{F}(\\cdot)$ from the spike input to the sharp image as shown in Fig. 2(b), formulated as $\\mathbf{I}(f) = \\mathcal{F}(S_{\\mathcal{T}};\\theta)$ with $f$ signifying the central moment of the interval $\\mathcal{T}$ . Prior research [6, 36, 41] employs a fixed and brief window (41 frames in their configuration) of the spike stream as the network in", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "put. However, it is insufficient for scenarios with low spike firing rates due to the limited information embedded within.", + "bbox": [ + 511, + 506, + 906, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While extending the window size allows the network to encompass more features, it complicates the feature extraction process and demands large computational resource. To address it, we configure the input of the Recon-Net with a complementary long-short spike stream as shown in Fig. 3.", + "bbox": [ + 511, + 535, + 905, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a long spike stream $S_{\\mathcal{T}}$ , we extract a short spike stream $S_{\\mathcal{T}'}^t \\subset S_{\\mathcal{T}}$ centered around time $t$ , with the network mapping reformulated as $\\mathbf{I}(t) = \\mathcal{F}(S_{\\mathcal{T}}, S_{\\mathcal{T}'}^t, t; \\theta)$ , where time $t$ is incorporated to guide the alignment of the short and long spike. Following the event stream pre-processing in [2, 26, 28], we apply voxelization to the long spike stream $S_{\\mathcal{T}}$ , reducing its dimensions from $K \\times H \\times W$ to $\\frac{K}{4} \\times H \\times W$ accumulating every 4 frames. The network architecture consists of basic convolutional blocks as illustrated in Fig. 3 and detailed in the supplementary materials.", + "bbox": [ + 511, + 611, + 905, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to the spike camera working mechanism described in Eq. (1), we can reconstruct a long-exposure image $\\mathbf{E}(\\mathcal{T})$ from the spike input $S_{\\mathcal{T}}$ with the TFP algorithm [48], mathematically formulated as below:", + "bbox": [ + 511, + 762, + 905, + 823 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E} (\\mathcal {T}) = \\frac {1}{T} \\int_ {\\mathcal {T}} \\mathbf {I} (t) d t = \\frac {C \\cdot N}{T}, \\qquad (6)\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 830, + 906, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $T$ denotes the duration of the interval $\\mathcal{T}$ , and $N$ is the cumulative number of spikes during this period.", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16612", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motivated by the textural similarity between the blurry image and the long-exposure image, we can employ the motion-reblur loss widely employed in self-supervised motion deblurring frameworks [3, 32, 39, 40] to guide the training of the Recon-Net. We initially reconstruct $M$ frames uniformly across the interval to form the image sequence $\\{\\mathbf{I}_{\\mathrm{rec}}(t_m)\\}_{m = 1}^M$ and further average them to synthesize the reblur image. The loss can be formulated by comparing the reblur image to the long-exposure image, expressed as:", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {L} _ {\\mathrm {r e c}} (M, \\mathcal {T}) = \\mathcal {L} \\left(\\frac {1}{M} \\sum_ {m = 1} ^ {M} \\mathbf {I} _ {\\mathrm {r e c}} \\left(t _ {m}\\right), \\mathbf {E} (\\mathcal {T})\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 237, + 483, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the loss $\\mathcal{L}$ [13] combines $L_{1}$ loss with a D-SSIM term, defined as $\\mathcal{L} = (1 - \\lambda)\\mathcal{L}_1 + \\lambda \\mathcal{L}_{\\mathrm{D - SSIM}}$ , effectively integrating both pixel accuracy and structural similarity.", + "bbox": [ + 89, + 286, + 483, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Nevertheless, the input spike stream $S_{\\mathcal{T}}$ sufficiently captures the information to estimate $\\mathbf{E}(\\mathcal{T})$ as detailed in Eq. (6), which will lead the Recon-Net to learn a trivial identity mapping from the spike to the long exposure image. To mitigate it, we propose the multi-reblur loss to regularize the reconstructed sequence. Specifically, we extract $N$ sub-intervals from the long exposure $T_{n} \\subset \\mathcal{T}$ with the same central moment and apply the reblur loss to each sub-interval, formulating the multi-reblur loss as below:", + "bbox": [ + 89, + 333, + 483, + 468 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r e c}} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\mathbb {L} _ {\\mathrm {r e c}} \\left(M _ {n}, \\mathcal {T} _ {n}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 478, + 483, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M_{n}$ represents the number of reconstructed frames within the $n$ -th sub-interval $\\mathcal{T}_n$ .", + "bbox": [ + 89, + 527, + 483, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Spike-based Gaussian Splatting", + "text_level": 1, + "bbox": [ + 89, + 566, + 370, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Suppose that retrieving the pose for each timestamp within the interval is straightforward. Following Sec. 4.1, we uniformly obtain $M$ poses from the interval $\\mathcal{T}$ , yielding a pose sequence $\\{\\mathbf{T}(t_m)\\}_{m=1}^M$ with the corresponding image sequence $\\{\\mathbf{I}_{\\mathrm{gs}}(t_m)\\}_{m=1}^M$ rendered via the 3DGS rasterization pipeline as shown in Fig. 3.", + "bbox": [ + 89, + 590, + 483, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While reconstructing a sequence of images from the spike stream by shortening the exposure interval in Eq. (6) for the training of 3DGS sounds feasible, these short-exposure images are noisy and lack texture details [3]. To this end, we utilize the long-exposure image $\\mathbf{E}(\\mathcal{T})$ as the supervision and design the motion-reblur loss for 3DGS similar to Eq. (7), expressed as:", + "bbox": [ + 89, + 681, + 483, + 787 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {g s}} = \\mathcal {L} \\left(\\frac {1}{M} \\sum_ {m = 1} ^ {M} \\mathbf {I} _ {\\mathrm {g s}} \\left(t _ {m}\\right), \\mathbf {E} (\\mathcal {T})\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 796, + 483, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Spike Camera Trajectory Modeling", + "text_level": 1, + "bbox": [ + 89, + 847, + 401, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Building on the learnable pose representation during camera exposure in previous research [30, 42], we employ a", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "similar approach to optimize the camera pose throughout the interval $\\mathcal{T}$ as shown in Fig. 3. For each spike stream $S_{\\mathcal{T}}$ , we optimize both the start pose $\\mathbf{T}_{\\mathrm{start}}$ and end pose $\\mathbf{T}_{\\mathrm{end}}$ over the interval, while intermediate poses are represented through interpolation between these two.", + "bbox": [ + 511, + 90, + 905, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the pose $\\mathbf{T}(t_m)$ at time $t_m$ , we utilize linear interpolation in the Lie algebra of $\\mathrm{SE}(3)$ to estimate the pose at time $t$ , formulated as:", + "bbox": [ + 511, + 166, + 905, + 210 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} \\left(t _ {m}\\right) = \\mathbf {T} _ {\\text {s t a r t}} \\cdot \\exp \\left(\\frac {t _ {m}}{T} \\cdot \\log \\left(\\mathbf {T} _ {\\text {s t a r t}} ^ {- 1} \\cdot \\mathbf {T} _ {\\text {e n d}}\\right)\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 215, + 905, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the training, the pose set $\\mathbb{P}$ consisting of the start and end poses from $V$ viewpoints is jointly optimized with the Recon-Net and Gaussian primitives as in Eq. (5). For further details on the camera trajectory modelling, please refer to BAD-NeRF [30] and BAD-Gaussian [42].", + "bbox": [ + 511, + 253, + 905, + 329 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Joint Optimization", + "text_level": 1, + "bbox": [ + 511, + 335, + 694, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the spike stream $S_{\\mathcal{T}}$ along with the start pose $\\mathbf{T}_{\\mathrm{start}}$ and the end pose $\\mathbf{T}_{\\mathrm{end}}$ to be optimized, we perform evenly spaced sampling interpolation as in Eq. (10) during the interval, obtaining the pose sequence $\\{\\mathbf{T}(t_m)\\}_{m=1}^M$ .", + "bbox": [ + 511, + 359, + 905, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each pose in the sequence, we select a short spike stream centered around the timestamp $t_m$ , obtaining the image sequence $\\{\\mathbf{I}_{\\mathrm{rec}}(t_m)\\}_{m=1}^M$ based on the Recon-Net in Sec. 4.1 and the image sequence $\\{\\mathbf{I}_{\\mathrm{gs}}(t_m)\\}_{m=1}^M$ rendered from the 3DGS in Sec. 4.2. While both sequences represent the same scene, the joint optimization loss $\\mathcal{L}_{\\mathrm{joint}}$ can be formulated as follows:", + "bbox": [ + 511, + 420, + 905, + 523 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {L} _ {\\text {j o i n t}} = \\frac {1}{M} \\sum_ {m = 1} ^ {M} \\left| \\left| \\mathbf {I} _ {\\mathrm {g s}} \\left(t _ {m}\\right) - \\mathbf {I} _ {\\mathrm {r e c}} \\left(t _ {m}\\right) \\right| \\right| _ {2} ^ {2}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 527, + 905, + 569 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, we found that the joint optimization framework tends to be unstable due to the order difference between the 3DGS and Recon-Net output sequences. Specifically, since the temporal order does not affect the loss in Eq. (9), the optimized pose sequence may reverse relative to the real camera trajectory. To address this, we apply a simple but effective flip-and-minimum operation to align both sequences, with the loss function reformulated as:", + "bbox": [ + 511, + 573, + 905, + 694 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {j o i n t}} = \\min \\left(\\mathbb {L} _ {\\text {j o i n t}}, \\mathbb {L} _ {\\text {j o i n t}} ^ {\\mathrm {r}}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 699, + 903, + 718 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {L} _ {\\text {j o i n t}} ^ {\\mathrm {r}} = \\frac {1}{M} \\sum_ {m = 1} ^ {M} \\left| \\left| \\mathbf {I} _ {\\mathrm {g s}} \\left(t _ {M - m + 1}\\right) - \\mathbf {I} _ {\\mathrm {r e c}} \\left(t _ {m}\\right) \\right| \\right| _ {2} ^ {2}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 720, + 903, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the difference between the $\\mathbb{L}_{\\mathrm{joint}}^{\\mathrm{r}}$ and the $\\mathbb{L}_{\\mathrm{joint}}$ lies in that the rendered sequence via 3DGS is reversed.", + "bbox": [ + 511, + 766, + 905, + 796 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To sum up, the loss function for our USP-Gaussian framework is presented as follows:", + "bbox": [ + 511, + 797, + 905, + 825 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} + \\mathcal {L} _ {\\text {g s}} + \\mathcal {L} _ {\\text {j o i n t}}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 833, + 903, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{rec}}$ optimizes Recon-Net parameters, $\\mathcal{L}_{\\mathrm{gs}}$ refines poses and Gaussian primitives, and $\\mathcal{L}_{\\mathrm{joint}}$ jointly optimizes all components, with gradient flow shown in Fig. 3.", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16613", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/37a8e8279c2392d64e562911ce0358a313ccd3b75da7618652ce045db769c956.jpg", + "image_caption": [ + "Figure 4. 3D reconstruction visual comparison of our USP-Gaussian compared with previous methods on the synthetic dataset, where the input is the long-exposure image defined in Eq. (6)." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 906, + 510 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/415c77045fb304ee71199e9d4d7ba3567e54f3552d7bbac5fd57fb718befc1f9.jpg", + "table_caption": [ + "Table 1. 3D reconstruction quantitative comparison on the synthetic dataset. Color shading indicates the best and second-best result." + ], + "table_footnote": [], + "table_body": "
MethodsWineTanabataFactoryOutdoorpoolAvg
PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
TFP-3DGS [48]24.1380.7960.23924.6260.7780.27427.8450.8520.17928.0240.7860.29426.1580.8030.239
TFI-3DGS [48]23.8460.7330.25121.1680.5500.52925.0190.7340.27824.7550.6370.44623.6970.6630.376
Spk2Img-3DGS [41]21.1500.7890.17820.4510.7580.21425.1150.8060.14323.3620.6470.41722.5200.7500.238
SpikeNeRF [50]25.8880.7980.20025.9650.7250.29527.5310.7800.19125.6420.6570.39626.2570.7400.270
SpikeGS [37]25.6980.8350.24726.1740.8260.24328.6270.8640.21428.2870.7970.27327.1960.8320.244
Ours26.4760.8620.17626.3000.8270.21328.6940.8610.20730.1420.8200.27027.9030.8430.217
", + "bbox": [ + 94, + 589, + 903, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 89, + 715, + 215, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Datasets", + "text_level": 1, + "bbox": [ + 89, + 741, + 192, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Synthetic Dataset. We construct the synthetic dataset based on Deblur-NeRF [19] scenes for quantitative evaluation. We start with rendering 13 sharp images for each viewpoint, with further interpolation algorithm XVFI [25] to insert 7 images between each frame pair. We apply the spike simulator on the interpolated sequence to emulate the real-world spike as in [3], thus generating 97 spike frames for each viewpoint.", + "bbox": [ + 89, + 763, + 482, + 885 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Real-world Dataset. To construct the real-world dataset,", + "bbox": [ + 89, + 885, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we shake the spike camera extremely fast while maintaining a fixed forward-facing orientation and extract 97 spike frames per view, consistent with the configuration in the synthetic dataset. For pose estimation, we convert the raw spike stream in each view to the short-exposure image as defined in Eq. (6) and subsequently employ COLMAP [24] to estimate the pose sequence.", + "bbox": [ + 511, + 718, + 906, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 832, + 718, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our USP-Gaussian with two-stage cascading pipelines and approaches specially designed for the spike camera, i.e., SpikeNeRF [50] and SpikeGS [37]. For the", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16614", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b5401e5ece889c1735c530f2b253351cb8c6da0c9f3640325aaf50f09d67310f.jpg", + "image_caption": [ + "Figure 5. 3D reconstruction visual comparison of our USP-Gaussian compared with previous methods on the real-world dataset." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 906, + 305 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/451a290a7973a7f24a6d3d90f58ab87e6576beb982b2dfa7fa77ce317796057f.jpg", + "table_caption": [ + "Table 2. Comparison of our method with SpikeGS on the wine dataset with inaccurate perturbed poses." + ], + "table_footnote": [], + "table_body": "
MetricsSpikeGSOurs
0%10%20%30%0%10%20%30%
PSNR ↑25.6918.8317.7516.4426.4726.0525.9323.46
SSIM ↑0.8350.6070.5920.5670.8620.8540.8510.772
LPIPS ↓0.2470.3720.4110.4460.1760.1840.1860.245
", + "bbox": [ + 93, + 393, + 483, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "two-stage approach, we initially apply TFP [48], TFI [48], and Spk2ImgNet [41] to reconstruct sharp images from the input spike followed by the common 3DGS processing pipeline, with the code implemented by Spike-Zoo [4]. All Gaussian-based methods are conducted based on the same 3DGS backbone, and identical camera poses input for a fair comparison. While SpikeGS requires several poses per spike stream for the exposure loss, which presents a notable challenge for COLMAP in real-world data, we omit the exposure loss and retain the rest of the framework. All comparative metrics are calculated based on the 3D reconstruction task, with further comparison on the spike-to-image reconstruction task and implementation details available in the supplementary materials.", + "bbox": [ + 88, + 477, + 482, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Synthetic Dataset. Quantitative and qualitative comparisons on the synthetic dataset are shown in Tab. 1 and Fig. 4 respectively. As illustrated in the Tab. 1, our USP-Gaussian achieves the best performance across most metrics in various scenes. Two-stage cascading methods TFP-3DGS and TFI-3DGS yield poor restoration due to the substantial noise embedded in the initially reconstructed sequence. Similarly, although Spk2ImgNet performs well on its original REDS dataset [41], it experiences substantial degradation due to the dataset domain gap. SpikeGS, as the primary comparison algorithm in this paper, heavily relies on the BSN [15] for self-supervised Spike-to-Image reconstruction. Nevertheless, the denoising performance of the BSN deteriorates in the presence of substantial noise,", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/66a58d80780aea1dbd490ce733c76f8de17c71039039fbf4e62018ce76ad5cc6.jpg", + "image_caption": [ + "Pose Trajectory with $10\\%$ Perturbation", + "Figure 6. Visual comparison of initial and optimized poses." + ], + "image_footnote": [], + "bbox": [ + 519, + 368, + 707, + 487 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/36a96113877b65220968cff14aa1c85ae9b68057a14edf353ec6a47b98b11da5.jpg", + "image_caption": [ + "Pose Trajectory with $30\\%$ Perturbation" + ], + "image_footnote": [], + "bbox": [ + 712, + 368, + 900, + 487 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a9c05e603693f1f54d49c5a360d31d420d7e23842a823e19fcd811e9347d3a7d.jpg", + "table_caption": [ + "Table 3. Translation and rotation poses error comparison." + ], + "table_footnote": [], + "table_body": "
StateTranslation ↓Rotation ↓
10%20%30%Avg10%20%30%Avg
Initial1.3022.9674.1932.8201.1322.1503.1592.147
Optimized0.8551.7811.4931.3760.1740.3201.2850.593
Δ (%)34.3%39.9%64.4%51.3%84.6%85.1%59.2%72.4%
", + "bbox": [ + 516, + 547, + 905, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "leading to pronounced grid artifacts and the attenuation of high-frequency details of the SpikeGS. SpikeNeRF faces a comparable challenge, as it requires noisy spike images as input and relies on the SNN for denoising.", + "bbox": [ + 511, + 622, + 906, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To this end, our proposed collaborative learning framework USP-Gaussian, benefiting from the inherent constraints between the spike stream and the multi-view consistency afforded by the 3DGS, effectively addresses the aforementioned problems without relying on the pre-training of the spike-based image reconstruction network.", + "bbox": [ + 511, + 685, + 905, + 775 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the previous part, accurate poses are obtained by applying COLMAP on sharp images rendered by Blender. To demonstrate that the integration of pose correction enhances the robustness of our method, we apply random perturbations of $10\\%$ , $20\\%$ , and $30\\%$ to the accurate poses in the wine scene, generating inaccurate poses for SpikeGS and our method with the comparison shown in Tab. 2.", + "bbox": [ + 511, + 777, + 905, + 882 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, visual trajectories and quantitative errors", + "bbox": [ + 531, + 885, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16615", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fec5786a936e57ca467775dcdddfee9420a0403165e8090c8cfdaa69848bafc8.jpg", + "table_caption": [ + "Table 4. Quantitative ablation on the synthetic dataset." + ], + "table_footnote": [], + "table_body": "
ID\\( \\mathcal{L}_{\\text{gs}} \\)Recon-Net\\( \\mathcal{L}_{\\text{joint}} \\)3DGS OutputsRecon-Net Outputs
\\( \\mathbb{L}_{\\text{rec}} \\)\\( \\mathbb{L}_{\\text{rec}} \\)PSNRSSIMLPIPSPSNRSSIMLPIPS
I✓22.3920.7500.208\\\\\\
II✓\\\\\\24.2550.7080.358
III✓\\\\\\24.5670.7440.332
IV✓✓✓25.6590.8500.18425.5710.8070.244
V✓✓✓26.4760.8620.17627.1380.8550.178
", + "bbox": [ + 94, + 114, + 483, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3b6d9336d9428282244b8c015c97d307596215c9132f3ceeb6839604aa1ca878.jpg", + "image_caption": [ + "Figure 7. Visual ablation on the effectiveness of multi-reblur loss, with the x-axis and y-axis representing step (k) and PSNR." + ], + "image_footnote": [], + "bbox": [ + 94, + 213, + 285, + 337 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c4f9d7ad94ec750ddedc8f23e68394e0223686d5071d530669adfdb4d3601022.jpg", + "image_caption": [ + "Figure 8. Visual ablation on the effectiveness of long spike input." + ], + "image_footnote": [], + "bbox": [ + 285, + 213, + 480, + 337 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of the initial noisy, and optimized poses are depicted in Fig. 6 and Tab. 3 respectively, with MAE measuring translation error and radians measuring rotation error, demonstrating the pose correction capability of our USP-Gaussian.", + "bbox": [ + 89, + 393, + 483, + 453 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Real-world Dataset. We also conduct the visual comparison on the real-world dataset as shown in Fig. 5. Due to the rapid camera ego-motion and the low imaging quality of the spike camera characterized by low spatial resolution, insufficient texture details, and underexposure, the pose sequence estimated by the COLMAP is highly inaccurate, presenting significant challenges for previous methods to recover accurate 3D representation as observed in Fig. 5. Nevertheless, our USP-Gaussian overcomes this limitation benefiting from the unifying optimization of the pose.", + "bbox": [ + 89, + 455, + 483, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 618, + 243, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this subsection, we demonstrate the effectiveness of our proposed joint optimization, complementary spike input, and multi-reblur loss strategies.", + "bbox": [ + 89, + 642, + 482, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "I. Joint Learning. The highlight of USP-Gaussian lies in the design of the joint loss, which aligns outputs from the 3DGS and Recon-Net thus achieving complementary performance improvement. We conduct the ablation experiment on loss functions of $\\mathcal{L}_{\\mathrm{gs}}$ , $\\mathcal{L}_{\\mathrm{rec}}$ and $\\mathcal{L}_{\\mathrm{joint}}$ with quantitative results listed in Tab. 4, which shows that the incorporation of the joint loss (ID-V) yields notable improvements for both 3DGS and Recon-Net compared to independent optimization (ID-I and ID-III).", + "bbox": [ + 89, + 688, + 482, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Further visual ablation comparisons of the restored image and the training curve are presented in Fig. 1. Visual comparison substantiates that the joint optimization of 3DGS and Recon-Net significantly enhances the texture details for both two. Regarding the training curve, the inde", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/880659126e051fac901ed855ca02e65a271eecad1e13d3c0807115b5ae865b80.jpg", + "image_caption": [ + "TFP-Input" + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 645, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/12a12df4201757396d4581728e8baa552e08c98e06ef4579be5bc3b2749bf7fc.jpg", + "image_caption": [ + "w/o Long Spike" + ], + "image_footnote": [], + "bbox": [ + 645, + 90, + 772, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/20556ac5b277f054d730d8f4dff0dcef3ca4c0c9ebffa9ac42441f85d151bda5.jpg", + "image_caption": [ + "w/ Long Spike" + ], + "image_footnote": [], + "bbox": [ + 774, + 90, + 901, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pendent 3DGS training (ID-I) demonstrates slow convergence due to the absence of crucial information in the long-exposure images. Concurrently, the independent training of Recon-Net (ID-III) tends to overfit due to the simplicity of the spike reconstruction task. To this end, our proposed joint optimization framework provides synergistic benefits for both, i.e., Recon-Net effectively compensates for deficient motion features embedded in the long-exposures image while the multi-view consistency afforded by 3DGS culminates in coherent outputs from the Recon-Net.", + "bbox": [ + 511, + 233, + 903, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "II. Multi-reblur Loss. The multi-reblur loss $\\mathcal{L}_{\\mathrm{rec}}$ in Eq. (7) is designed to substitute the single-reblur loss $\\mathbb{L}_{\\mathrm{rec}}$ in Eq. (8). Without further constraints, $\\mathbb{L}_{\\mathrm{rec}}$ will lead ReconNet to learn an identity mapping from the input spike stream to the corresponding long-exposure image as demonstrated in [3]. Comparative analysis in Tab. 4 depicts that $\\mathcal{L}_{\\mathrm{rec}}$ (ID-V) yields beneficial improvement than $\\mathbb{L}_{\\mathrm{rec}}$ (ID-IV). Further comparison on the training curve in Fig. 7 reveals the mechanism of collaborative training, i.e., the multi-reblur loss enhances the performance of Recon-Net, thereby facilitating the training of the 3DGS.", + "bbox": [ + 511, + 385, + 905, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "III. Complementary Input. We utilize complementary long and short spike stream inputs for Recon-Net to address the fixed short window limitation inherent in previous methods [6, 36, 41]. As illustrated in the visual comparison in Fig. 8, the inclusion of the long spike effectively suppresses salt-and-pepper noise in the Recon-Net output, leading to improved SSIM and LPIPS metrics.", + "bbox": [ + 511, + 553, + 905, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 676, + 633, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a unifying framework USP-Gaussian for the collaborative optimization of spike-based image reconstruction, pose correction, and Gaussian Splating. By mitigating error propagation inherent in traditional cascading frameworks, USP-Gaussian attains high reconstruction fidelity across both synthetic and real-world datasets, with further ablation experiments revealing that Spike-to-Image reconstruction and 3D reconstruction tasks furnish mutually complementary information.", + "bbox": [ + 511, + 703, + 905, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation. The limitation of USP-Gaussian is that it requires longer training time and more GPU memory compared to previous methods due to the additional poses and Recon-Net parameters to be optimized.", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16616", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We sincerely appreciate Yuyan Chen (HUST) for her valuable suggestions and for polishing the figures. This work was supported by the National Natural Science Foundation of China (62422601, 62176003, 62088102, 62306015), the China Postdoctoral Science Foundation (2023T160015), the Young Elite Scientists Sponsorship Program by CAST (2023QNRC001), the Beijing Municipal Science and Technology Program (Z241100004224004) and the Beijing Nova Program (20230484362,20240484703).", + "bbox": [ + 89, + 114, + 483, + 251 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 265, + 187, + 281 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anish Bhattacharya, Ratnesh Madaan, Fernando Cladera, Sai Vemprala, Rogerio Bonatti, Kostas Daniilidis, Ashish Kapoor, Vijay Kumar, Nikolai Matni, and Jayesh K Gupta. Evdnerf: Reconstructing event data with dynamic neural radiance fields. In WACV, pages 5846-5855, 2024. 3", + "[2] Kang Chen and Lei Yu. Motion deblur by learning residual from events. IEEE TMM, 2024. 4", + "[3] Kang Chen, Shiyan Chen, Jiyuan Zhang, Baoyue Zhang, Yajing Zheng, Tiejun Huang, and Zhaofei Yu. Spikereveal: Unlocking temporal sequences from real blurry inputs with spike streams. Advances in Neural Information Processing Systems, 37:62673-62696, 2025. 2, 5, 6, 8", + "[4] Kang Chen, Zhiyuan Ye, Tiejun Huang, and Zhaofei Yu. Spike-Zoo: A toolbox for spike-to-image reconstruction. https://github.com/chenkang455/Spike-Zoo, 2025.7", + "[5] Kang Chen, Yajing Zheng, Tiejun Huang, and Zhaofei Yu. Rethinking high-speed image reconstruction framework with spike camera. arXiv preprint arXiv:2501.04477, 2025. 2", + "[6] Shiyan Chen, Chaoteng Duan, Zhaofei Yu, Ruiqin Xiong, and Tiejun Huang. Self-supervised mutual learning for dynamic scene reconstruction of spiking camera. In IJCAI, pages 2859-2866, 2022. 2, 4, 8", + "[7] Shiyan Chen, Zhaofei Yu, and Tiejun Huang. Self-supervised joint dynamic scene reconstruction and optical flow estimation for spiking camera. In AAAI, pages 350-358, 2023. 2", + "[8] Shiyan Chen, Jiyuan Zhang, Yajing Zheng, Tiejun Huang, and Zhaofei Yu. Enhancing motion deblurring in high-speed scenes with spike streams. In NeurIPS, 2023. 2", + "[9] Gaole Dai, Zhenyu Wang, Qinwen Xu, Wen Cheng, Ming Lu, Boxing Shi, Shanghang Zhang, and Tiejun Huang. Spikenvs: Enhancing novel view synthesis from blurry images via spike camera. arXiv preprint arXiv:2404.06710, 2024.3", + "[10] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE TPAMI, 44(1):154-180, 2020. 3", + "[11] Yijia Guo, Liwen Hu, Lei Ma, and Tiejun Huang. Spikegs: Reconstruct 3d scene via fast-moving bio-inspired sensors. arXiv preprint arXiv:2407.03771, 2024. 2, 3" + ], + "bbox": [ + 93, + 290, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Tiejun Huang, Yajing Zheng, Zhaofei Yu, Rui Chen, Yuan Li, Ruiqin Xiong, Lei Ma, Junwei Zhao, Siwei Dong, Lin Zhu, et al. $1000\\times$ faster camera and machine vision with ordinary devices. Engineering, 25:110-119, 2023. 1, 3", + "[13] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 5", + "[14] Simon Klenk, Lukas Koestler, Davide Scaramuzza, and Daniel Cremers. E-nerf: Neural radiance fields from a moving event camera. IEEE Robotics and Automation Letters, 8 (3):1587–1594, 2023. 3", + "[15] Samuli Laine, Tero Karras, Jaakko Lehtinen, and Timo Aila. High-quality self-supervised deep image denoising. NeurIPS, 32, 2019. 2, 7", + "[16] Wenpu Li, Pian Wan, Peng Wang, Jinghang Li, Yi Zhou, and Peidong Liu. Benerf: Neural radiance fields from a single blurry image and event stream. In European Conference on Computer Vision (ECCV), 2024. 3", + "[17] Bohao Liao, Wei Zhai, Zengyu Wan, Tianzhu Zhang, Yang Cao, and Zheng-Jun Zha. Ef-3dgs: Event-aided free-trajectory 3d gaussian splatting. arXiv preprint arXiv:2410.15392, 2024. 3", + "[18] Zibin Liu, Banglei Guan, Yang Shang, Shunkun Liang, Zhenbao Yu, and Qifeng Yu. Optical flow-guided 6dof object pose tracking with an event camera. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6501-6509, 2024. 3", + "[19] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-nerf: Neural radiance fields from blurry images. In CVPR, pages 12861-12870, 2022. 6", + "[20] Qi Ma, Danda Pani Paudel, Ajad Chhatkuli, and Luc Van Gool. Deformable neural radiance fields using rgb and event cameras. In ICCV, pages 3590-3600, 2023. 3", + "[21] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2", + "[22] Bryan Peterson. Understanding exposure: how to shoot great photographs with any camera. AmPhoto books, 2016. 1", + "[23] Viktor Rudnev, Mohamed Elgharib, Christian Theobalt, and Vladislav Golyanik. Eventnerf: Neural radiance fields from a single colour event camera. In CVPR, pages 4992-5002, 2023. 3", + "[24] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, pages 4104-4113, 2016. 6", + "[25] Hyeonjun Sim, Jihyong Oh, and Munchurl Kim. Xvfi: extreme video frame interpolation. In ICCV, pages 14489-14498, 2021. 6", + "[26] Chen Song, Qixing Huang, and Chandrajit Bajaj. E-cir: Event-enhanced continuous intensity recovery. In CVPR, pages 7803-7812, 2022. 4", + "[27] Jing Tao, You Li, Banglei Guan, Yang Shang, and Qifeng Yu. Simultaneous enhancement and noise suppression un" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16617", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "der complex illumination conditions. IEEE Transactions on Instrumentation and Measurement, 2024. 3", + "[28] Bishan Wang, Jingwei He, Lei Yu, Gui-Song Xia, and Wen Yang. Event enhanced high-quality image recovery. In ECCV, pages 155-171. Springer, 2020. 4", + "[29] Jiaxu Wang, Junhao He, Ziyi Zhang, Mingyuan Sun, Jingkai Sun, and Renjing Xu. Evggs: A collaborative learning framework for event-based generalizable gaussian splatting. arXiv preprint arXiv:2405.14959, 2024. 3", + "[30] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. Bad-nerf: Bundle adjusted deblur neural radiance fields. In CVPR, pages 4170–4179, 2023. 5", + "[31] Tianyi Xiong, Jiayi Wu, Botao He, Cornelia Fermuller, Yiannis Aloimonos, Heng Huang, and Christopher Metzler. Event3dgs: Event-based 3d gaussian splatting for high-speed robot egomotion. In 8th Annual Conference on Robot Learning, 2024. 3", + "[32] Fang Xu, Lei Yu, Bishan Wang, Wen Yang, Gui-Song Xia, Xu Jia, Zhendong Qiao, and Jianzhuang Liu. Motion deblurring with real events. In ICCV, pages 2583-2592, 2021. 5", + "[33] Baoyue Zhang, Yajing Zheng, Shiyan Chen, Jiyuan Zhang, Kang Chen, Zhaofei Yu, and Tiejun Huang. Spikemm: Fleximagnification of high-speed micro-motions. arXiv preprint arXiv:2406.00383, 2024. 2", + "[34] Jiyuan Zhang, Lulu Tang, Zhaofei Yu, Jiwen Lu, and Tiejun Huang. Spike transformer: Monocular depth estimation for spiking camera. In ECCV, pages 34-52. Springer, 2022. 2", + "[35] Jiyuan Zhang, Shiyan Chen, Yajing Zheng, Zhaofei Yu, and Tiejun Huang. Unveiling the potential of spike streams for foreground occlusion removal from densely continuous views. arXiv preprint arXiv:2307.00821, 2023. 2", + "[36] Jiyuan Zhang, Shanshan Jia, Zhaofei Yu, and Tiejun Huang. Learning temporal-ordered representation for spike streams based on discrete wavelet transforms. In AAAI, pages 137–147, 2023. 2, 4, 8", + "[37] Jiyuan Zhang, Kang Chen, Shiyan Chen, Yajing Zheng, Tiejun Huang, and Zhaofei Yu. Spikegs: 3d gaussian splatt-ting from spike streams with high-speed camera motion. arXiv preprint arXiv:2407.10062, 2024. 2, 3, 6", + "[38] Jiyuan Zhang, Shiyan Chen, Yajing Zheng, Zhaofei Yu, and Tiejun Huang. Spike-guided motion deblurring with unknown modal spatiotemporal alignment. In CVPR, pages 25047-25057, 2024. 2", + "[39] Xiang Zhang and Lei Yu. Unifying motion deblurring and frame interpolation with events. In CVPR, pages 17765-17774, 2022. 5", + "[40] Xiang Zhang, Lei Yu, Wen Yang, Jianzhuang Liu, and GuiSong Xia. Generalizing event-based motion deblurring in real-world scenarios. In ICCV, pages 10734-10744, 2023. 5", + "[41] Jing Zhao, Ruiqin Xiong, Hangfan Liu, Jian Zhang, and Tiejun Huang. Spk2imgnet: Learning to reconstruct dynamic scene from continuous spike stream. In CVPR, pages 11996-12005, 2021. 2, 4, 6, 7, 8", + "[42] Lingzhe Zhao, Peng Wang, and Peidong Liu. Bad-gaussians: Bundle adjusted deblur gaussian splatting. arXiv preprint arXiv:2403.11831, 2024. 5" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Rui Zhao, Ruiqin Xiong, Jing Zhao, Zhaofei Yu, Xiaopeng Fan, and Tiejun Huang. Learning optical flow from continuous spike streams. NeurIPS, 35:7905-7920, 2022. 2", + "[44] Rui Zhao, Ruiqin Xiong, Jing Zhao, Jian Zhang, Xiaopeng Fan, Zhaofei Yu, and Tiejun Huang. Boosting spike camera image reconstruction from a perspective of dealing with spike fluctuations. In CVPR, pages 24955-24965, 2024. 2", + "[45] Yajing Zheng, Lingxiao Zheng, Zhaofei Yu, Boxin Shi, Yonghong Tian, and Tiejun Huang. High-speed image reconstruction through short-term plasticity for spiking cameras. In CVPR, pages 6358-6367, 2021. 2", + "[46] Yajing Zheng, Zhaofei Yu, Song Wang, and Tiejun Huang. Spike-based motion estimation for object tracking through bio-inspired unsupervised learning. IEEE TIP, 32:335-349, 2022. 2", + "[47] Kangrui Zhou, Taihang Lei, Banglei Guan, and Qifeng Yu. Event-based depth estimation with dense occlusion. Optics Letters, 49(12):3376-3379, 2024. 3", + "[48] Lin Zhu, Siwei Dong, Tiejun Huang, and Yonghong Tian. A retina-inspired sampling method for visual texture reconstruction. In ICME, pages 1432-1437. IEEE, 2019. 2, 4, 6, 7", + "[49] Lin Zhu, Siwei Dong, Jianing Li, Tiejun Huang, and Yonghong Tian. Retina-like visual image reconstruction via spiking neural model. In CVPR, pages 1438-1446, 2020. 1, 2", + "[50] Lin Zhu, Kangmin Jia, Yifan Zhao, Yunshan Qi, Lizhi Wang, and Hua Huang. Spikenerf: Learning neural radiance fields from continuous spike stream. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6285-6295, 2024. 2, 3, 6" + ], + "bbox": [ + 514, + 92, + 906, + 530 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16618", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/USP-Gaussian_ Unifying Spike-based Image Reconstruction, Pose Correction and Gaussian Splatting/62c2f0f5-8ea0-4959-8e97-f59984bffdfe_model.json b/2025/USP-Gaussian_ Unifying Spike-based Image Reconstruction, Pose Correction and Gaussian Splatting/62c2f0f5-8ea0-4959-8e97-f59984bffdfe_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c214599d881b24c91193c2914b494272eedfcdf5 --- /dev/null +++ b/2025/USP-Gaussian_ Unifying Spike-based Image Reconstruction, Pose Correction and Gaussian Splatting/62c2f0f5-8ea0-4959-8e97-f59984bffdfe_model.json @@ -0,0 +1,2387 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.111, + 0.13, + 0.888, + 0.177 + ], + "angle": 0, + "content": "USP-Gaussian: Unifying Spike-based Image Reconstruction, Pose Correction and Gaussian Splatting" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.202, + 0.896, + 0.277 + ], + "angle": 0, + "content": "Kang Chen\\(^{1,2}\\) Jiyuan Zhang\\(^{1,2}\\) Zecheng Hao\\(^{1,2}\\) Yajing Zheng\\(^{1,2}\\) Tiejun Huang\\(^{1,2,3}\\) Zhaofei Yu\\(^{1,3*}\\) \n\\(^{1}\\)School of Computer Science, Peking University \n\\(^{2}\\)State Key Laboratory for Multimedia Information Processing, Peking University \n\\(^{3}\\)Institute for Artificial Intelligence, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.278, + 0.866, + 0.293 + ], + "angle": 0, + "content": "{mrchenkang, jyzhang}@stu.pku.edu.cn, {haozecheng,yj.zheng,tjhuang,yuzf12}@pku.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.304, + 0.321, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.305, + 0.641, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.307, + 0.906, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.499, + 0.907, + 0.556 + ], + "angle": 0, + "content": "Figure 1. Left. Illustration of our USP-Gaussian framework, where the spike-based image Reconstruction Network (Recon-Net), camera poses, and 3DGS are collaboratively optimized signified by \\(\\mathbb{O}\\). Mid. Visual ablation showcasing the performance of Recon-Net and 3DGS with and without (w/ & w/o) the joint optimization strategy, with the ablation table depicted in Tab. 4 and the input formulated in Eq. (6). Right. Training curve comparison for Recon-Net and 3DGS with and without joint optimization." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.583, + 0.327, + 0.599 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.615, + 0.486, + 0.875 + ], + "angle": 0, + "content": "Spike camera, as an innovative type of neuromorphic camera that captures scenes with 0-1 bit stream at \\(40\\mathrm{kHz}\\), is increasingly being employed for the novel view synthesis task building on techniques such as Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS). Previous spike-based approaches typically follow a three-stage pipeline: I. Spike-to-image reconstruction based on established algorithms. II. Camera poses estimation. III. Novel view synthesis. However, the cascading framework suffers from substantial cumulative errors, i.e., the quality of the initially reconstructed images will impact pose estimation, ultimately limiting the fidelity of the 3D reconstruction. To address this limitation, we propose a synergistic optimization framework USP-Gaussian, which unifies spike-to-image reconstruction, pose correction, and gaussian splatting into an end-to-end pipeline. Leveraging the multi-view consistency afforded by 3DGS and the motion capture capability of the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.909, + 0.707 + ], + "angle": 0, + "content": "spike camera, our framework enables iterative optimization between the spike-to-image reconstruction network and 3DGS. Experiments on synthetic datasets demonstrate that our method surpasses previous approaches by effectively eliminating cascading errors. Moreover, in real-world scenarios, our method achieves robust 3D reconstruction benefiting from the integration of pose optimization. Our code, data, and trained models are available at https://github.com/chenkang455/USP-Gaussian." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.738, + 0.646, + 0.753 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Spike cameras [12], inspired by the human retina [49], introduce an innovative paradigm for encoding high-speed dynamic scenes. In contrast to conventional imaging systems [22] that rely on fixed exposure intervals to integrate visual information, spike cameras continuously sense dynamics by recording the spike stream with an impressive firing rate of \\(40\\mathrm{kHz}\\). Benefiting from the efficient bit representation of spikes, spike cameras can adeptly detect rapid scenes with significantly diminished bandwidth compared" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.248, + 0.901 + ], + "angle": 0, + "content": "* Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16609" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.408, + 0.106 + ], + "angle": 0, + "content": "to the high-speed cameras such as the Phantom." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.109, + 0.484, + 0.244 + ], + "angle": 0, + "content": "Capitalizing on the remarkable capability of the spike camera to capture intricate motion and texture details in high-speed environments, recent research [5, 36, 41, 44, 45, 48, 49] concentrate on establishing the mapping from bit-encoded spike to visually perceivable images. Subsequent investigations delve into the estimation of optical flow [43], facilitation of motion deblur [3, 8, 38] and applications such as high-speed object detection [46], motion magnification [33], depth estimation [34] and occlusion removal [35]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.246, + 0.484, + 0.412 + ], + "angle": 0, + "content": "Neural Radiance Fields [21] and 3D Gaussian Splitting [13] have emerged as prominent novel view synthesis technologies benefiting from their designed differential training frameworks. Among them, 3DGS employs Gaussian primitives for explicit scene representation and the splatting algorithm for projection, thereby achieving faster rendering speed compared to the pixel-by-pixel approach implemented in NeRF. Building on NeRF and 3DGS, the high-speed and continuous capturing potential of the spike camera can be fully unleashed in 3D reconstruction tasks [11, 37, 50]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.414, + 0.484, + 0.55 + ], + "angle": 0, + "content": "While the idea of applying the spike camera for 3D reconstruction is trivial and straightforward in a sequential manner, i.e., we can reconstruct images from the spike stream, estimate camera poses and subsequently apply NeRF/3DGS to obtain the 3D representation. Nevertheless, the cascading pipeline introduces significant cumulative errors when the quality of initially recovered images is poor, thereby degrading the accuracy of camera pose estimation and limiting the fidelity of the reconstructed scene." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.483, + 0.717 + ], + "angle": 0, + "content": "For example, SpikeGS [37] initiates with a self-supervised spike-to-image reconstruction framework built on the Blind Spot Network (BSN) [15], intended to mitigate the performance degradation [6, 7] encountered by supervised learning methods on real-world scenes. With recovered high-quality images, SpikeGS designs a further constraint from the long spike input and integrates it into the vanilla 3DGS training pipeline. While SpikeGS have improved preliminary representation for 3DGS, it remains constrained within the cascading framework thus inheriting the aforementioned cumulative error problem." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To address these challenges, we propose USP-Gaussian, Unifying Spike-based image reconstruction, Pose correction, and Gaussian Splatting within a one-stage optimization framework to mitigate error propagation inherent in the three-stage cascading pipeline. Specifically, USP-Gaussian comprises two synergistic branches as illustrated in Fig. 1, one dedicated to spike-to-image reconstruction via Recon-Net and the other to novel view synthesis implemented with 3DGS. To facilitate the training of Recon-Net, we design \\(\\mathcal{L}_{\\mathrm{rec}}\\) based on the constraint between the input spike and recovered outputs. Besides, we supervise the training of 3DGS based on \\(\\mathcal{L}_{\\mathrm{gs}}\\), which obviates the require" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.152 + ], + "angle": 0, + "content": "ment for pre-recovered high-quality images. While each branch of USP-Gaussian operates independently, we further introduce the joint loss \\(\\mathcal{L}_{\\mathrm{joint}}\\) derived from the outputs of both two branches to facilitate unified optimization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.909, + 0.304 + ], + "angle": 0, + "content": "Visual ablation comparisons of restored images and the training curve depicted in Fig. 1 demonstrate that the joint optimization of the two branches mutually enhances the recovered textures and the overall training process. We conduct quantitative and qualitative evaluations on the synthetic dataset and the real-world dataset. Experiments showcase that our method achieves superior restoration in both datasets, validating the effectiveness of our joint learning framework. In summary, the contributions of our USP-Gaussian are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.305, + 0.905, + 0.35 + ], + "angle": 0, + "content": "- We propose a unified framework for the joint optimization of Recon-Net, Pose, and 3DGS, mitigating error amplification inherent in cascading frameworks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.351, + 0.905, + 0.395 + ], + "angle": 0, + "content": "- We demonstrate that spike-to-image and 3D reconstruction tasks can mutually facilitate and enhance the optimization of each other." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.396, + 0.905, + 0.442 + ], + "angle": 0, + "content": "- We introduce a novel self-supervised training paradigm for the spike-based image reconstruction network with the multi-view consistency afforded by 3DGS." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.305, + 0.905, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.457, + 0.656, + 0.473 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.483, + 0.816, + 0.5 + ], + "angle": 0, + "content": "2.1. Spike-based Image Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.854 + ], + "angle": 0, + "content": "While neuromorphic cameras generate high frame rate bitstream information, it is imperative to transform them into multi-bit grayscale images that are perceptible to human observers. To achieve this spike-to-image target, a plethora of methodologies [5-7, 36, 41, 44, 45, 49] have been proposed, which can be broadly classified into model-based, supervised, and self-supervised learning paradigms. Model-based methods, rooted in the sampling principle of the spike camera, aim to infer high-fidelity images drawing from the physical spike-image relationship [48], biological retinal models [49], and short-term synaptic plasticity [45]. Confronted with pervasive dark current noise that is challenging to model, learning-based methods are employed to directly deduce the sharp image from spike input with unknown firing patterns. Specifically, Spk2ImgNet [41] designed a hierarchical architecture engineered to capture the intricate temporal dynamics present within spike streams. SpikeCLIP [5] utilized the textual description of the captured scene as supervision for training the spike-to-image network. Concurrently, self-supervised learning methodologies based on the BSN [6, 7] are developed to counteract the performance deterioration in supervised learning methods owing to the domain gap between synthetic and real-world datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a novel self-supervised spike-based image reconstruction framework leveraging the multi-view constraints provided by 3D scenes." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16610" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.483, + 0.122 + ], + "angle": 0, + "content": "2.2. Novel View Synthesis based on the Neuromorphic Camera" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.129, + 0.484, + 0.205 + ], + "angle": 0, + "content": "Recent research has extensively employed neuromorphic cameras, including event cameras [10, 16, 18, 27, 47], which encode intensity difference in dynamic scenarios with low-latency and spike cameras [12], for 3D scene reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.205, + 0.484, + 0.4 + ], + "angle": 0, + "content": "Event Camera. Klenk et al. [14] proposed the E-NeRF based on the event generation mechanism, which can recover sharp 3D scenes under high-speed camera motion. Rudnev et al. [23] further harnessed a color event camera to learn the color representation of scenes. [1, 20] focused on recovering dynamic scenes with rigid deformations from event streams. In addition to NeRF, EvGGS [29] introduces a generalizable Gaussian splatting framework for event cameras designed to reconstruct 3D scenes without retraining. EF-3DGS [17] develops a free-trajectory framework by aligning the RGB input, event streams, and pose. Event3DGS [31] presents an effective solution for rapid ego-motion robotics scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.401, + 0.484, + 0.567 + ], + "angle": 0, + "content": "Spike Camera. Zhu et al. [50] design a differentiable SNN and integrate it into the NeRF training pipeline to simulate the generation of spike streams, where the loss function is formulated based on the rendered spike output and the input spike. Dai et al. [9] further introduce SpikeNVS by extending the spike cameras to RGB-Spike input. Additionally, Guo et al. [11] introduced a color 3D scene reconstruction framework by integrating Bayer-pattern spike streams into the 3DGS training pipeline. SpikeGS [37] initially employs the BSN to recover images from the noisy spike stream and subsequently integrates them into the 3DGS." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.484, + 0.629 + ], + "angle": 0, + "content": "While the cascaded framework SpikeGS [37] introduces propagation errors, we propose a joint optimization framework that concurrently handles spike-based image reconstruction, pose estimation, and novel view synthesis." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.639, + 0.218, + 0.656 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.663, + 0.238, + 0.68 + ], + "angle": 0, + "content": "3.1. Spike Camera" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.484, + 0.835 + ], + "angle": 0, + "content": "The spike camera consists of the photodetector, voltage integrator, comparator, and synchronous readout as shown in Fig. 2. Within the designated time interval \\(\\mathcal{T} = [t_s,t_e]\\), the photodetector continuously captures the light intensity \\(\\mathbf{I}(t)\\) and converts it into the voltage signal \\(\\mathbf{V}(t) = \\mu \\mathbf{I}(t)\\), where \\(\\mu\\) is the photoelectric conversion coefficient. Concurrently, the comparator assesses the integrator voltage \\(\\mathbf{A}(t)\\) against the predefined threshold \\(\\Theta\\). When the threshold is met, the spike is emitted and the integrator voltage is reset, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.835, + 0.483, + 0.868 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} (t) = \\int_ {t _ {s}} ^ {t} \\mathbf {I} (t) d t \\bmod C, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where \\(C = \\Theta /\\mu\\) is a constant proportional to the spike firing threshold \\(\\Theta\\). While the integrator continuously ag-" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.907, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.3, + 0.907, + 0.356 + ], + "angle": 0, + "content": "Figure 2. (a) The spike camera captures a railway moving 350 km/h. (b) The Recon-Net is designed to establish the mapping between the input spike stream and the sharp image. (c) The working mechanism of the spike camera." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.907, + 0.474 + ], + "angle": 0, + "content": "gregates voltage, the readout circuit synchronously retrieves spike signals from the registers at up to \\(40\\mathrm{kHz}\\). Spikes fired during the specified interval \\(\\mathcal{T}\\) are represented as a three-dimensional bit stream \\(S\\in \\{0,1\\}^{K\\times H\\times W}\\), where \\(H\\) and \\(W\\) indicate the image height and width, with \\(K\\) denoting the spike sequence length." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.484, + 0.724, + 0.501 + ], + "angle": 0, + "content": "3.2. 3D Gaussian Splatting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.906, + 0.555 + ], + "angle": 0, + "content": "3D Gaussian Splating leverages a set of three-dimensional Gaussian primitives, symbolized as \\(\\{\\mathcal{G}\\}_{n=1}^{N_{\\mathrm{gs}}}\\), to effectively represent the spatial characteristics of the captured scene." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.556, + 0.906, + 0.632 + ], + "angle": 0, + "content": "Each Gaussian primitive is centered at a point \\(\\mathbf{p}_n\\), with its spread and orientation influenced by the covariance matrix \\(\\Sigma_n = RSS^T R^T\\), where \\(R\\) and \\(T\\) denotes the rotation and the scale matrix, respectively. The impact of the Gaussian on any point \\(\\mathbf{v}\\) in the scene is quantified by:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.642, + 0.907, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {n} (\\mathbf {v}) = \\exp \\left(- \\frac {1}{2} (\\mathbf {v} - \\mathbf {p} _ {n}) ^ {T} \\boldsymbol {\\Sigma} _ {n} ^ {- 1} (\\mathbf {v} - \\mathbf {p} _ {n})\\right). (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.687, + 0.906, + 0.733 + ], + "angle": 0, + "content": "When it comes to rendering for viewing transformation \\(W\\), these 3D Gaussian primitives are mapped onto 2D primitives \\(\\mathcal{G}_n^{2D}\\), with the covariance matrix \\(\\Sigma_n^\\prime\\) reformulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.745, + 0.905, + 0.765 + ], + "angle": 0, + "content": "\\[\n\\Sigma_ {n} ^ {\\prime} = J W \\Sigma_ {n} W ^ {T} J ^ {T}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.772, + 0.906, + 0.802 + ], + "angle": 0, + "content": "where \\(J\\) represents the Jacobian matrix derived from the affine approximation of the projective transformation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.802, + 0.906, + 0.848 + ], + "angle": 0, + "content": "Further, each pixel on the rendered plane is calculated by rasterizing these 2D Gaussians, sorted by depth, according to the following formulation:" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.86, + 0.907, + 0.905 + ], + "angle": 0, + "content": "\\[\nC (r) = \\sum_ {k = 1} ^ {N _ {\\mathrm {g s}}} T _ {i} \\alpha_ {i} \\mathbf {c} _ {i}, T _ {i} = \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {j}\\right), \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16611" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.907, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.382, + 0.907, + 0.482 + ], + "angle": 0, + "content": "Figure 3. The working pipeline of our USP-Gaussian. For each camera viewpoint, we initially derive a pose sequence at equal time intervals based on the camera model described in Sec. 4.3. Leveraging the 3DGS rendering pipeline detailed in Sec. 4.2, we subsequently generate a corresponding sharp sequence for this viewpoint. Concurrently, Recon-Net is employed to reconstruct the same sequence from the captured spike stream as outlined in Sec. 4.1. While Recon-Net and 3DGS are trained based on the discrepancy between the re-synthesized blurry image with the input long-exposure image, alignment of their outputs for the same scene is further enforced through joint optimization loss as described in Sec. 4.4. With formulated loss functions, our USP-Gaussian enables the collaborative optimization of Gaussian primitives, camera poses, and the parameters of Recon-Net as formulated in Eq. (5)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.486, + 0.553 + ], + "angle": 0, + "content": "where \\( D \\) represents the indices set reordered by depth, and \\( D_{k} \\) denotes the index of the Gaussian primitive that is the \\( k \\)-th closest to the rendering point." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.567, + 0.192, + 0.584 + ], + "angle": 0, + "content": "4. Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.593, + 0.484, + 0.67 + ], + "angle": 0, + "content": "Target Clarification. Given the captured spike stream set \\(\\{S_v\\}_{v=1}^V\\) from \\(V\\) different viewpoints with initial inaccurate poses \\(\\{\\mathcal{P}_v\\}_{v=1}^V\\), the USP-Gaussian aims to optimize the parameters of Recon-Net and Gaussian primitives while outputting refined poses as shown in Fig. 1, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.682, + 0.483, + 0.703 + ], + "angle": 0, + "content": "\\[\n(\\mathbb {S}, \\mathbb {P}, \\mathbb {G}, \\theta) \\xrightarrow {\\text {U S P - G a u s s i a n}} (\\mathbb {P} ^ {*}, \\mathbb {G} ^ {*}, \\theta^ {*}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.714, + 0.484, + 0.777 + ], + "angle": 0, + "content": "where \\(\\mathbb{S}\\) and \\(\\mathbb{P}\\) represent the sets of spike streams and poses from \\(M\\) viewpoints, \\(\\mathbb{G}\\) denotes the ensemble of Gaussian primitives, and \\(\\theta\\) encapsulates the Recon-Net parameters. The notation \\((^{*})\\) indicates the optimized parameters." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.787, + 0.466, + 0.803 + ], + "angle": 0, + "content": "4.1. Spike-based Image Reconstruction Network" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The Recon-Net is designed to facilitate the mapping \\(\\mathcal{F}(\\cdot)\\) from the spike input to the sharp image as shown in Fig. 2(b), formulated as \\(\\mathbf{I}(f) = \\mathcal{F}(S_{\\mathcal{T}};\\theta)\\) with \\(f\\) signifying the central moment of the interval \\(\\mathcal{T}\\). Prior research [6, 36, 41] employs a fixed and brief window (41 frames in their configuration) of the spike stream as the network in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.907, + 0.536 + ], + "angle": 0, + "content": "put. However, it is insufficient for scenarios with low spike firing rates due to the limited information embedded within." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.536, + 0.906, + 0.612 + ], + "angle": 0, + "content": "While extending the window size allows the network to encompass more features, it complicates the feature extraction process and demands large computational resource. To address it, we configure the input of the Recon-Net with a complementary long-short spike stream as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.906, + 0.763 + ], + "angle": 0, + "content": "Given a long spike stream \\( S_{\\mathcal{T}} \\), we extract a short spike stream \\( S_{\\mathcal{T}'}^t \\subset S_{\\mathcal{T}} \\) centered around time \\( t \\), with the network mapping reformulated as \\( \\mathbf{I}(t) = \\mathcal{F}(S_{\\mathcal{T}}, S_{\\mathcal{T}'}^t, t; \\theta) \\), where time \\( t \\) is incorporated to guide the alignment of the short and long spike. Following the event stream pre-processing in [2, 26, 28], we apply voxelization to the long spike stream \\( S_{\\mathcal{T}} \\), reducing its dimensions from \\( K \\times H \\times W \\) to \\( \\frac{K}{4} \\times H \\times W \\) accumulating every 4 frames. The network architecture consists of basic convolutional blocks as illustrated in Fig. 3 and detailed in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.763, + 0.906, + 0.824 + ], + "angle": 0, + "content": "According to the spike camera working mechanism described in Eq. (1), we can reconstruct a long-exposure image \\(\\mathbf{E}(\\mathcal{T})\\) from the spike input \\(S_{\\mathcal{T}}\\) with the TFP algorithm [48], mathematically formulated as below:" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.831, + 0.907, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E} (\\mathcal {T}) = \\frac {1}{T} \\int_ {\\mathcal {T}} \\mathbf {I} (t) d t = \\frac {C \\cdot N}{T}, \\qquad (6)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\(T\\) denotes the duration of the interval \\(\\mathcal{T}\\), and \\(N\\) is the cumulative number of spikes during this period." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16612" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.228 + ], + "angle": 0, + "content": "Motivated by the textural similarity between the blurry image and the long-exposure image, we can employ the motion-reblur loss widely employed in self-supervised motion deblurring frameworks [3, 32, 39, 40] to guide the training of the Recon-Net. We initially reconstruct \\(M\\) frames uniformly across the interval to form the image sequence \\(\\{\\mathbf{I}_{\\mathrm{rec}}(t_m)\\}_{m = 1}^M\\) and further average them to synthesize the reblur image. The loss can be formulated by comparing the reblur image to the long-exposure image, expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.238, + 0.484, + 0.281 + ], + "angle": 0, + "content": "\\[\n\\mathbb {L} _ {\\mathrm {r e c}} (M, \\mathcal {T}) = \\mathcal {L} \\left(\\frac {1}{M} \\sum_ {m = 1} ^ {M} \\mathbf {I} _ {\\mathrm {r e c}} \\left(t _ {m}\\right), \\mathbf {E} (\\mathcal {T})\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.287, + 0.484, + 0.334 + ], + "angle": 0, + "content": "where the loss \\(\\mathcal{L}\\) [13] combines \\(L_{1}\\) loss with a D-SSIM term, defined as \\(\\mathcal{L} = (1 - \\lambda)\\mathcal{L}_1 + \\lambda \\mathcal{L}_{\\mathrm{D - SSIM}}\\), effectively integrating both pixel accuracy and structural similarity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.334, + 0.484, + 0.469 + ], + "angle": 0, + "content": "Nevertheless, the input spike stream \\( S_{\\mathcal{T}} \\) sufficiently captures the information to estimate \\( \\mathbf{E}(\\mathcal{T}) \\) as detailed in Eq. (6), which will lead the Recon-Net to learn a trivial identity mapping from the spike to the long exposure image. To mitigate it, we propose the multi-reblur loss to regularize the reconstructed sequence. Specifically, we extract \\( N \\) sub-intervals from the long exposure \\( T_{n} \\subset \\mathcal{T} \\) with the same central moment and apply the reblur loss to each sub-interval, formulating the multi-reblur loss as below:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.479, + 0.484, + 0.52 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r e c}} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\mathbb {L} _ {\\mathrm {r e c}} \\left(M _ {n}, \\mathcal {T} _ {n}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.529, + 0.484, + 0.56 + ], + "angle": 0, + "content": "where \\(M_{n}\\) represents the number of reconstructed frames within the \\(n\\)-th sub-interval \\(\\mathcal{T}_n\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.568, + 0.371, + 0.584 + ], + "angle": 0, + "content": "4.2. Spike-based Gaussian Splatting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.591, + 0.484, + 0.681 + ], + "angle": 0, + "content": "Suppose that retrieving the pose for each timestamp within the interval is straightforward. Following Sec. 4.1, we uniformly obtain \\(M\\) poses from the interval \\(\\mathcal{T}\\), yielding a pose sequence \\(\\{\\mathbf{T}(t_m)\\}_{m=1}^M\\) with the corresponding image sequence \\(\\{\\mathbf{I}_{\\mathrm{gs}}(t_m)\\}_{m=1}^M\\) rendered via the 3DGS rasterization pipeline as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.682, + 0.484, + 0.788 + ], + "angle": 0, + "content": "While reconstructing a sequence of images from the spike stream by shortening the exposure interval in Eq. (6) for the training of 3DGS sounds feasible, these short-exposure images are noisy and lack texture details [3]. To this end, we utilize the long-exposure image \\(\\mathbf{E}(\\mathcal{T})\\) as the supervision and design the motion-reblur loss for 3DGS similar to Eq. (7), expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.797, + 0.484, + 0.84 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {g s}} = \\mathcal {L} \\left(\\frac {1}{M} \\sum_ {m = 1} ^ {M} \\mathbf {I} _ {\\mathrm {g s}} \\left(t _ {m}\\right), \\mathbf {E} (\\mathcal {T})\\right). \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.402, + 0.865 + ], + "angle": 0, + "content": "4.3. Spike Camera Trajectory Modeling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Building on the learnable pose representation during camera exposure in previous research [30, 42], we employ a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.166 + ], + "angle": 0, + "content": "similar approach to optimize the camera pose throughout the interval \\(\\mathcal{T}\\) as shown in Fig. 3. For each spike stream \\(S_{\\mathcal{T}}\\), we optimize both the start pose \\(\\mathbf{T}_{\\mathrm{start}}\\) and end pose \\(\\mathbf{T}_{\\mathrm{end}}\\) over the interval, while intermediate poses are represented through interpolation between these two." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.167, + 0.906, + 0.212 + ], + "angle": 0, + "content": "For the pose \\(\\mathbf{T}(t_m)\\) at time \\(t_m\\), we utilize linear interpolation in the Lie algebra of \\(\\mathrm{SE}(3)\\) to estimate the pose at time \\(t\\), formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.217, + 0.906, + 0.25 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} \\left(t _ {m}\\right) = \\mathbf {T} _ {\\text {s t a r t}} \\cdot \\exp \\left(\\frac {t _ {m}}{T} \\cdot \\log \\left(\\mathbf {T} _ {\\text {s t a r t}} ^ {- 1} \\cdot \\mathbf {T} _ {\\text {e n d}}\\right)\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.254, + 0.906, + 0.33 + ], + "angle": 0, + "content": "During the training, the pose set \\(\\mathbb{P}\\) consisting of the start and end poses from \\(V\\) viewpoints is jointly optimized with the Recon-Net and Gaussian primitives as in Eq. (5). For further details on the camera trajectory modelling, please refer to BAD-NeRF [30] and BAD-Gaussian [42]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.336, + 0.696, + 0.352 + ], + "angle": 0, + "content": "4.4. Joint Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.36, + 0.906, + 0.42 + ], + "angle": 0, + "content": "Given the spike stream \\( S_{\\mathcal{T}} \\) along with the start pose \\( \\mathbf{T}_{\\mathrm{start}} \\) and the end pose \\( \\mathbf{T}_{\\mathrm{end}} \\) to be optimized, we perform evenly spaced sampling interpolation as in Eq. (10) during the interval, obtaining the pose sequence \\( \\{\\mathbf{T}(t_m)\\}_{m=1}^M \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.421, + 0.906, + 0.525 + ], + "angle": 0, + "content": "For each pose in the sequence, we select a short spike stream centered around the timestamp \\( t_m \\), obtaining the image sequence \\( \\{\\mathbf{I}_{\\mathrm{rec}}(t_m)\\}_{m=1}^M \\) based on the Recon-Net in Sec. 4.1 and the image sequence \\( \\{\\mathbf{I}_{\\mathrm{gs}}(t_m)\\}_{m=1}^M \\) rendered from the 3DGS in Sec. 4.2. While both sequences represent the same scene, the joint optimization loss \\( \\mathcal{L}_{\\mathrm{joint}} \\) can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.529, + 0.906, + 0.57 + ], + "angle": 0, + "content": "\\[\n\\mathbb {L} _ {\\text {j o i n t}} = \\frac {1}{M} \\sum_ {m = 1} ^ {M} \\left| \\left| \\mathbf {I} _ {\\mathrm {g s}} \\left(t _ {m}\\right) - \\mathbf {I} _ {\\mathrm {r e c}} \\left(t _ {m}\\right) \\right| \\right| _ {2} ^ {2}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.574, + 0.906, + 0.695 + ], + "angle": 0, + "content": "However, we found that the joint optimization framework tends to be unstable due to the order difference between the 3DGS and Recon-Net output sequences. Specifically, since the temporal order does not affect the loss in Eq. (9), the optimized pose sequence may reverse relative to the real camera trajectory. To address this, we apply a simple but effective flip-and-minimum operation to align both sequences, with the loss function reformulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.7, + 0.905, + 0.719 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {j o i n t}} = \\min \\left(\\mathbb {L} _ {\\text {j o i n t}}, \\mathbb {L} _ {\\text {j o i n t}} ^ {\\mathrm {r}}\\right), \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.722, + 0.905, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\mathbb {L} _ {\\text {j o i n t}} ^ {\\mathrm {r}} = \\frac {1}{M} \\sum_ {m = 1} ^ {M} \\left| \\left| \\mathbf {I} _ {\\mathrm {g s}} \\left(t _ {M - m + 1}\\right) - \\mathbf {I} _ {\\mathrm {r e c}} \\left(t _ {m}\\right) \\right| \\right| _ {2} ^ {2}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.767, + 0.906, + 0.797 + ], + "angle": 0, + "content": "where the difference between the \\(\\mathbb{L}_{\\mathrm{joint}}^{\\mathrm{r}}\\) and the \\(\\mathbb{L}_{\\mathrm{joint}}\\) lies in that the rendered sequence via 3DGS is reversed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.798, + 0.906, + 0.827 + ], + "angle": 0, + "content": "To sum up, the loss function for our USP-Gaussian framework is presented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.834, + 0.905, + 0.851 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} + \\mathcal {L} _ {\\text {g s}} + \\mathcal {L} _ {\\text {j o i n t}}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{rec}}\\) optimizes Recon-Net parameters, \\(\\mathcal{L}_{\\mathrm{gs}}\\) refines poses and Gaussian primitives, and \\(\\mathcal{L}_{\\mathrm{joint}}\\) jointly optimizes all components, with gradient flow shown in Fig. 3." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16613" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.907, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.52, + 0.907, + 0.551 + ], + "angle": 0, + "content": "Figure 4. 3D reconstruction visual comparison of our USP-Gaussian compared with previous methods on the synthetic dataset, where the input is the long-exposure image defined in Eq. (6)." + }, + { + "type": "table_caption", + "bbox": [ + 0.098, + 0.563, + 0.899, + 0.581 + ], + "angle": 0, + "content": "Table 1. 3D reconstruction quantitative comparison on the synthetic dataset. Color shading indicates the best and second-best result." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.59, + 0.905, + 0.695 + ], + "angle": 0, + "content": "
MethodsWineTanabataFactoryOutdoorpoolAvg
PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
TFP-3DGS [48]24.1380.7960.23924.6260.7780.27427.8450.8520.17928.0240.7860.29426.1580.8030.239
TFI-3DGS [48]23.8460.7330.25121.1680.5500.52925.0190.7340.27824.7550.6370.44623.6970.6630.376
Spk2Img-3DGS [41]21.1500.7890.17820.4510.7580.21425.1150.8060.14323.3620.6470.41722.5200.7500.238
SpikeNeRF [50]25.8880.7980.20025.9650.7250.29527.5310.7800.19125.6420.6570.39626.2570.7400.270
SpikeGS [37]25.6980.8350.24726.1740.8260.24328.6270.8640.21428.2870.7970.27327.1960.8320.244
Ours26.4760.8620.17626.3000.8270.21328.6940.8610.20730.1420.8200.27027.9030.8430.217
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.717, + 0.216, + 0.734 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.742, + 0.194, + 0.756 + ], + "angle": 0, + "content": "5.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.886 + ], + "angle": 0, + "content": "Synthetic Dataset. We construct the synthetic dataset based on Deblur-NeRF [19] scenes for quantitative evaluation. We start with rendering 13 sharp images for each viewpoint, with further interpolation algorithm XVFI [25] to insert 7 images between each frame pair. We apply the spike simulator on the interpolated sequence to emulate the real-world spike as in [3], thus generating 97 spike frames for each viewpoint." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Real-world Dataset. To construct the real-world dataset," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.907, + 0.825 + ], + "angle": 0, + "content": "we shake the spike camera extremely fast while maintaining a fixed forward-facing orientation and extract 97 spike frames per view, consistent with the configuration in the synthetic dataset. For pose estimation, we convert the raw spike stream in each view to the short-exposure image as defined in Eq. (6) and subsequently employ COLMAP [24] to estimate the pose sequence." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.719, + 0.849 + ], + "angle": 0, + "content": "5.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We compare our USP-Gaussian with two-stage cascading pipelines and approaches specially designed for the spike camera, i.e., SpikeNeRF [50] and SpikeGS [37]. For the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16614" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.907, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.316, + 0.877, + 0.331 + ], + "angle": 0, + "content": "Figure 5. 3D reconstruction visual comparison of our USP-Gaussian compared with previous methods on the real-world dataset." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.355, + 0.483, + 0.384 + ], + "angle": 0, + "content": "Table 2. Comparison of our method with SpikeGS on the wine dataset with inaccurate perturbed poses." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.395, + 0.485, + 0.468 + ], + "angle": 0, + "content": "
MetricsSpikeGSOurs
0%10%20%30%0%10%20%30%
PSNR ↑25.6918.8317.7516.4426.4726.0525.9323.46
SSIM ↑0.8350.6070.5920.5670.8620.8540.8510.772
LPIPS ↓0.2470.3720.4110.4460.1760.1840.1860.245
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.478, + 0.483, + 0.688 + ], + "angle": 0, + "content": "two-stage approach, we initially apply TFP [48], TFI [48], and Spk2ImgNet [41] to reconstruct sharp images from the input spike followed by the common 3DGS processing pipeline, with the code implemented by Spike-Zoo [4]. All Gaussian-based methods are conducted based on the same 3DGS backbone, and identical camera poses input for a fair comparison. While SpikeGS requires several poses per spike stream for the exposure loss, which presents a notable challenge for COLMAP in real-world data, we omit the exposure loss and retain the rest of the framework. All comparative metrics are calculated based on the 3D reconstruction task, with further comparison on the spike-to-image reconstruction task and implementation details available in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Synthetic Dataset. Quantitative and qualitative comparisons on the synthetic dataset are shown in Tab. 1 and Fig. 4 respectively. As illustrated in the Tab. 1, our USP-Gaussian achieves the best performance across most metrics in various scenes. Two-stage cascading methods TFP-3DGS and TFI-3DGS yield poor restoration due to the substantial noise embedded in the initially reconstructed sequence. Similarly, although Spk2ImgNet performs well on its original REDS dataset [41], it experiences substantial degradation due to the dataset domain gap. SpikeGS, as the primary comparison algorithm in this paper, heavily relies on the BSN [15] for self-supervised Spike-to-Image reconstruction. Nevertheless, the denoising performance of the BSN deteriorates in the presence of substantial noise," + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.357, + 0.704, + 0.368 + ], + "angle": 0, + "content": "Pose Trajectory with \\(10\\%\\) Perturbation" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.369, + 0.709, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.72, + 0.357, + 0.894, + 0.368 + ], + "angle": 0, + "content": "Pose Trajectory with \\(30\\%\\) Perturbation" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.369, + 0.901, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.502, + 0.885, + 0.516 + ], + "angle": 0, + "content": "Figure 6. Visual comparison of initial and optimized poses." + }, + { + "type": "table_caption", + "bbox": [ + 0.538, + 0.522, + 0.88, + 0.537 + ], + "angle": 0, + "content": "Table 3. Translation and rotation poses error comparison." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.548, + 0.906, + 0.611 + ], + "angle": 0, + "content": "
StateTranslation ↓Rotation ↓
10%20%30%Avg10%20%30%Avg
Initial1.3022.9674.1932.8201.1322.1503.1592.147
Optimized0.8551.7811.4931.3760.1740.3201.2850.593
Δ (%)34.3%39.9%64.4%51.3%84.6%85.1%59.2%72.4%
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.623, + 0.907, + 0.683 + ], + "angle": 0, + "content": "leading to pronounced grid artifacts and the attenuation of high-frequency details of the SpikeGS. SpikeNeRF faces a comparable challenge, as it requires noisy spike images as input and relies on the SNN for denoising." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.906, + 0.776 + ], + "angle": 0, + "content": "To this end, our proposed collaborative learning framework USP-Gaussian, benefiting from the inherent constraints between the spike stream and the multi-view consistency afforded by the 3DGS, effectively addresses the aforementioned problems without relying on the pre-training of the spike-based image reconstruction network." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.779, + 0.906, + 0.883 + ], + "angle": 0, + "content": "In the previous part, accurate poses are obtained by applying COLMAP on sharp images rendered by Blender. To demonstrate that the integration of pose correction enhances the robustness of our method, we apply random perturbations of \\(10\\%\\), \\(20\\%\\), and \\(30\\%\\) to the accurate poses in the wine scene, generating inaccurate poses for SpikeGS and our method with the comparison shown in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Furthermore, visual trajectories and quantitative errors" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16615" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.126, + 0.09, + 0.447, + 0.104 + ], + "angle": 0, + "content": "Table 4. Quantitative ablation on the synthetic dataset." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.115, + 0.484, + 0.2 + ], + "angle": 0, + "content": "
ID\\( \\mathcal{L}_{\\text{gs}} \\)Recon-Net\\( \\mathcal{L}_{\\text{joint}} \\)3DGS OutputsRecon-Net Outputs
\\( \\mathbb{L}_{\\text{rec}} \\)\\( \\mathbb{L}_{\\text{rec}} \\)PSNRSSIMLPIPSPSNRSSIMLPIPS
I✓22.3920.7500.208\\\\\\
II✓\\\\\\24.2550.7080.358
III✓\\\\\\24.5670.7440.332
IV✓✓✓25.6590.8500.18425.5710.8070.244
V✓✓✓26.4760.8620.17627.1380.8550.178
" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.214, + 0.287, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.214, + 0.482, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.35, + 0.483, + 0.379 + ], + "angle": 0, + "content": "Figure 7. Visual ablation on the effectiveness of multi-reblur loss, with the x-axis and y-axis representing step (k) and PSNR." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.394, + 0.484, + 0.454 + ], + "angle": 0, + "content": "of the initial noisy, and optimized poses are depicted in Fig. 6 and Tab. 3 respectively, with MAE measuring translation error and radians measuring rotation error, demonstrating the pose correction capability of our USP-Gaussian." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.456, + 0.484, + 0.609 + ], + "angle": 0, + "content": "Real-world Dataset. We also conduct the visual comparison on the real-world dataset as shown in Fig. 5. Due to the rapid camera ego-motion and the low imaging quality of the spike camera characterized by low spatial resolution, insufficient texture details, and underexposure, the pose sequence estimated by the COLMAP is highly inaccurate, presenting significant challenges for previous methods to recover accurate 3D representation as observed in Fig. 5. Nevertheless, our USP-Gaussian overcomes this limitation benefiting from the unifying optimization of the pose." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.619, + 0.245, + 0.635 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.688 + ], + "angle": 0, + "content": "In this subsection, we demonstrate the effectiveness of our proposed joint optimization, complementary spike input, and multi-reblur loss strategies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.824 + ], + "angle": 0, + "content": "I. Joint Learning. The highlight of USP-Gaussian lies in the design of the joint loss, which aligns outputs from the 3DGS and Recon-Net thus achieving complementary performance improvement. We conduct the ablation experiment on loss functions of \\(\\mathcal{L}_{\\mathrm{gs}}\\), \\(\\mathcal{L}_{\\mathrm{rec}}\\) and \\(\\mathcal{L}_{\\mathrm{joint}}\\) with quantitative results listed in Tab. 4, which shows that the incorporation of the joint loss (ID-V) yields notable improvements for both 3DGS and Recon-Net compared to independent optimization (ID-I and ID-III)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Further visual ablation comparisons of the restored image and the training curve are presented in Fig. 1. Visual comparison substantiates that the joint optimization of 3DGS and Recon-Net significantly enhances the texture details for both two. Regarding the training curve, the inde" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.646, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.56, + 0.179, + 0.606, + 0.189 + ], + "angle": 0, + "content": "TFP-Input" + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.091, + 0.773, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.179, + 0.745, + 0.189 + ], + "angle": 0, + "content": "w/o Long Spike" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.091, + 0.902, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.806, + 0.179, + 0.87, + 0.189 + ], + "angle": 0, + "content": "w/ Long Spike" + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.203, + 0.903, + 0.218 + ], + "angle": 0, + "content": "Figure 8. Visual ablation on the effectiveness of long spike input." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.234, + 0.905, + 0.385 + ], + "angle": 0, + "content": "pendent 3DGS training (ID-I) demonstrates slow convergence due to the absence of crucial information in the long-exposure images. Concurrently, the independent training of Recon-Net (ID-III) tends to overfit due to the simplicity of the spike reconstruction task. To this end, our proposed joint optimization framework provides synergistic benefits for both, i.e., Recon-Net effectively compensates for deficient motion features embedded in the long-exposures image while the multi-view consistency afforded by 3DGS culminates in coherent outputs from the Recon-Net." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.386, + 0.906, + 0.552 + ], + "angle": 0, + "content": "II. Multi-reblur Loss. The multi-reblur loss \\(\\mathcal{L}_{\\mathrm{rec}}\\) in Eq. (7) is designed to substitute the single-reblur loss \\(\\mathbb{L}_{\\mathrm{rec}}\\) in Eq. (8). Without further constraints, \\(\\mathbb{L}_{\\mathrm{rec}}\\) will lead ReconNet to learn an identity mapping from the input spike stream to the corresponding long-exposure image as demonstrated in [3]. Comparative analysis in Tab. 4 depicts that \\(\\mathcal{L}_{\\mathrm{rec}}\\) (ID-V) yields beneficial improvement than \\(\\mathbb{L}_{\\mathrm{rec}}\\) (ID-IV). Further comparison on the training curve in Fig. 7 reveals the mechanism of collaborative training, i.e., the multi-reblur loss enhances the performance of Recon-Net, thereby facilitating the training of the 3DGS." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.906, + 0.659 + ], + "angle": 0, + "content": "III. Complementary Input. We utilize complementary long and short spike stream inputs for Recon-Net to address the fixed short window limitation inherent in previous methods [6, 36, 41]. As illustrated in the visual comparison in Fig. 8, the inclusion of the long spike effectively suppresses salt-and-pepper noise in the Recon-Net output, leading to improved SSIM and LPIPS metrics." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.677, + 0.634, + 0.692 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.906, + 0.839 + ], + "angle": 0, + "content": "In this paper, we propose a unifying framework USP-Gaussian for the collaborative optimization of spike-based image reconstruction, pose correction, and Gaussian Splating. By mitigating error propagation inherent in traditional cascading frameworks, USP-Gaussian attains high reconstruction fidelity across both synthetic and real-world datasets, with further ablation experiments revealing that Spike-to-Image reconstruction and 3D reconstruction tasks furnish mutually complementary information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Limitation. The limitation of USP-Gaussian is that it requires longer training time and more GPU memory compared to previous methods due to the additional poses and Recon-Net parameters to be optimized." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16616" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.252 + ], + "angle": 0, + "content": "We sincerely appreciate Yuyan Chen (HUST) for her valuable suggestions and for polishing the figures. This work was supported by the National Natural Science Foundation of China (62422601, 62176003, 62088102, 62306015), the China Postdoctoral Science Foundation (2023T160015), the Young Elite Scientists Sponsorship Program by CAST (2023QNRC001), the Beijing Municipal Science and Technology Program (Z241100004224004) and the Beijing Nova Program (20230484362,20240484703)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.266, + 0.188, + 0.282 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.291, + 0.484, + 0.36 + ], + "angle": 0, + "content": "[1] Anish Bhattacharya, Ratnesh Madaan, Fernando Cladera, Sai Vemprala, Rogerio Bonatti, Kostas Daniilidis, Ashish Kapoor, Vijay Kumar, Nikolai Matni, and Jayesh K Gupta. Evdnerf: Reconstructing event data with dynamic neural radiance fields. In WACV, pages 5846-5855, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.362, + 0.483, + 0.388 + ], + "angle": 0, + "content": "[2] Kang Chen and Lei Yu. Motion deblur by learning residual from events. IEEE TMM, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.391, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[3] Kang Chen, Shiyan Chen, Jiyuan Zhang, Baoyue Zhang, Yajing Zheng, Tiejun Huang, and Zhaofei Yu. Spikereveal: Unlocking temporal sequences from real blurry inputs with spike streams. Advances in Neural Information Processing Systems, 37:62673-62696, 2025. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.462, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[4] Kang Chen, Zhiyuan Ye, Tiejun Huang, and Zhaofei Yu. Spike-Zoo: A toolbox for spike-to-image reconstruction. https://github.com/chenkang455/Spike-Zoo, 2025.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.519, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[5] Kang Chen, Yajing Zheng, Tiejun Huang, and Zhaofei Yu. Rethinking high-speed image reconstruction framework with spike camera. arXiv preprint arXiv:2501.04477, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.562, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[6] Shiyan Chen, Chaoteng Duan, Zhaofei Yu, Ruiqin Xiong, and Tiejun Huang. Self-supervised mutual learning for dynamic scene reconstruction of spiking camera. In IJCAI, pages 2859-2866, 2022. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.619, + 0.482, + 0.672 + ], + "angle": 0, + "content": "[7] Shiyan Chen, Zhaofei Yu, and Tiejun Huang. Self-supervised joint dynamic scene reconstruction and optical flow estimation for spiking camera. In AAAI, pages 350-358, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.676, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[8] Shiyan Chen, Jiyuan Zhang, Yajing Zheng, Tiejun Huang, and Zhaofei Yu. Enhancing motion deblurring in high-speed scenes with spike streams. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.719, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[9] Gaole Dai, Zhenyu Wang, Qinwen Xu, Wen Cheng, Ming Lu, Boxing Shi, Shanghang Zhang, and Tiejun Huang. Spikenvs: Enhancing novel view synthesis from blurry images via spike camera. arXiv preprint arXiv:2404.06710, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[10] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE TPAMI, 44(1):154-180, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[11] Yijia Guo, Liwen Hu, Lei Ma, and Tiejun Huang. Spikegs: Reconstruct 3d scene via fast-moving bio-inspired sensors. arXiv preprint arXiv:2407.03771, 2024. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.291, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[12] Tiejun Huang, Yajing Zheng, Zhaofei Yu, Rui Chen, Yuan Li, Ruiqin Xiong, Lei Ma, Junwei Zhao, Siwei Dong, Lin Zhu, et al. \\(1000\\times\\) faster camera and machine vision with ordinary devices. Engineering, 25:110-119, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[13] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[14] Simon Klenk, Lukas Koestler, Davide Scaramuzza, and Daniel Cremers. E-nerf: Neural radiance fields from a moving event camera. IEEE Robotics and Automation Letters, 8 (3):1587–1594, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.904, + 0.303 + ], + "angle": 0, + "content": "[15] Samuli Laine, Tero Karras, Jaakko Lehtinen, and Timo Aila. High-quality self-supervised deep image denoising. NeurIPS, 32, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[16] Wenpu Li, Pian Wan, Peng Wang, Jinghang Li, Yi Zhou, and Peidong Liu. Benerf: Neural radiance fields from a single blurry image and event stream. In European Conference on Computer Vision (ECCV), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[17] Bohao Liao, Wei Zhai, Zengyu Wan, Tianzhu Zhang, Yang Cao, and Zheng-Jun Zha. Ef-3dgs: Event-aided free-trajectory 3d gaussian splatting. arXiv preprint arXiv:2410.15392, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[18] Zibin Liu, Banglei Guan, Yang Shang, Shunkun Liang, Zhenbao Yu, and Qifeng Yu. Optical flow-guided 6dof object pose tracking with an event camera. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6501-6509, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.543 + ], + "angle": 0, + "content": "[19] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-nerf: Neural radiance fields from blurry images. In CVPR, pages 12861-12870, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[20] Qi Ma, Danda Pani Paudel, Ajad Chhatkuli, and Luc Van Gool. Deformable neural radiance fields using rgb and event cameras. In ICCV, pages 3590-3600, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.59, + 0.905, + 0.643 + ], + "angle": 0, + "content": "[21] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[22] Bryan Peterson. Understanding exposure: how to shoot great photographs with any camera. AmPhoto books, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.688, + 0.905, + 0.742 + ], + "angle": 0, + "content": "[23] Viktor Rudnev, Mohamed Elgharib, Christian Theobalt, and Vladislav Golyanik. Eventnerf: Neural radiance fields from a single colour event camera. In CVPR, pages 4992-5002, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[24] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, pages 4104-4113, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[25] Hyeonjun Sim, Jihyong Oh, and Munchurl Kim. Xvfi: extreme video frame interpolation. In ICCV, pages 14489-14498, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[26] Chen Song, Qixing Huang, and Chandrajit Bajaj. E-cir: Event-enhanced continuous intensity recovery. In CVPR, pages 7803-7812, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Jing Tao, You Li, Banglei Guan, Yang Shang, and Qifeng Yu. Simultaneous enhancement and noise suppression un" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16617" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "der complex illumination conditions. IEEE Transactions on Instrumentation and Measurement, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.164 + ], + "angle": 0, + "content": "[28] Bishan Wang, Jingwei He, Lei Yu, Gui-Song Xia, and Wen Yang. Event enhanced high-quality image recovery. In ECCV, pages 155-171. Springer, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.167, + 0.482, + 0.222 + ], + "angle": 0, + "content": "[29] Jiaxu Wang, Junhao He, Ziyi Zhang, Mingyuan Sun, Jingkai Sun, and Renjing Xu. Evggs: A collaborative learning framework for event-based generalizable gaussian splatting. arXiv preprint arXiv:2405.14959, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.224, + 0.482, + 0.266 + ], + "angle": 0, + "content": "[30] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. Bad-nerf: Bundle adjusted deblur neural radiance fields. In CVPR, pages 4170–4179, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.268, + 0.482, + 0.336 + ], + "angle": 0, + "content": "[31] Tianyi Xiong, Jiayi Wu, Botao He, Cornelia Fermuller, Yiannis Aloimonos, Heng Huang, and Christopher Metzler. Event3dgs: Event-based 3d gaussian splatting for high-speed robot egomotion. In 8th Annual Conference on Robot Learning, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.339, + 0.482, + 0.381 + ], + "angle": 0, + "content": "[32] Fang Xu, Lei Yu, Bishan Wang, Wen Yang, Gui-Song Xia, Xu Jia, Zhendong Qiao, and Jianzhuang Liu. Motion deblurring with real events. In ICCV, pages 2583-2592, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.383, + 0.482, + 0.437 + ], + "angle": 0, + "content": "[33] Baoyue Zhang, Yajing Zheng, Shiyan Chen, Jiyuan Zhang, Kang Chen, Zhaofei Yu, and Tiejun Huang. Spikemm: Fleximagnification of high-speed micro-motions. arXiv preprint arXiv:2406.00383, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.44, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[34] Jiyuan Zhang, Lulu Tang, Zhaofei Yu, Jiwen Lu, and Tiejun Huang. Spike transformer: Monocular depth estimation for spiking camera. In ECCV, pages 34-52. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.485, + 0.482, + 0.54 + ], + "angle": 0, + "content": "[35] Jiyuan Zhang, Shiyan Chen, Yajing Zheng, Zhaofei Yu, and Tiejun Huang. Unveiling the potential of spike streams for foreground occlusion removal from densely continuous views. arXiv preprint arXiv:2307.00821, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.542, + 0.482, + 0.597 + ], + "angle": 0, + "content": "[36] Jiyuan Zhang, Shanshan Jia, Zhaofei Yu, and Tiejun Huang. Learning temporal-ordered representation for spike streams based on discrete wavelet transforms. In AAAI, pages 137–147, 2023. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.6, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[37] Jiyuan Zhang, Kang Chen, Shiyan Chen, Yajing Zheng, Tiejun Huang, and Zhaofei Yu. Spikegs: 3d gaussian splatt-ting from spike streams with high-speed camera motion. arXiv preprint arXiv:2407.10062, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.482, + 0.711 + ], + "angle": 0, + "content": "[38] Jiyuan Zhang, Shiyan Chen, Yajing Zheng, Zhaofei Yu, and Tiejun Huang. Spike-guided motion deblurring with unknown modal spatiotemporal alignment. In CVPR, pages 25047-25057, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.715, + 0.482, + 0.755 + ], + "angle": 0, + "content": "[39] Xiang Zhang and Lei Yu. Unifying motion deblurring and frame interpolation with events. In CVPR, pages 17765-17774, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.758, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[40] Xiang Zhang, Lei Yu, Wen Yang, Jianzhuang Liu, and GuiSong Xia. Generalizing event-based motion deblurring in real-world scenarios. In ICCV, pages 10734-10744, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[41] Jing Zhao, Ruiqin Xiong, Hangfan Liu, Jian Zhang, and Tiejun Huang. Spk2imgnet: Learning to reconstruct dynamic scene from continuous spike stream. In CVPR, pages 11996-12005, 2021. 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[42] Lingzhe Zhao, Peng Wang, and Peidong Liu. Bad-gaussians: Bundle adjusted deblur gaussian splatting. arXiv preprint arXiv:2403.11831, 2024. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[43] Rui Zhao, Ruiqin Xiong, Jing Zhao, Zhaofei Yu, Xiaopeng Fan, and Tiejun Huang. Learning optical flow from continuous spike streams. NeurIPS, 35:7905-7920, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[44] Rui Zhao, Ruiqin Xiong, Jing Zhao, Jian Zhang, Xiaopeng Fan, Zhaofei Yu, and Tiejun Huang. Boosting spike camera image reconstruction from a perspective of dealing with spike fluctuations. In CVPR, pages 24955-24965, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[45] Yajing Zheng, Lingxiao Zheng, Zhaofei Yu, Boxin Shi, Yonghong Tian, and Tiejun Huang. High-speed image reconstruction through short-term plasticity for spiking cameras. In CVPR, pages 6358-6367, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[46] Yajing Zheng, Zhaofei Yu, Song Wang, and Tiejun Huang. Spike-based motion estimation for object tracking through bio-inspired unsupervised learning. IEEE TIP, 32:335-349, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.306, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[47] Kangrui Zhou, Taihang Lei, Banglei Guan, and Qifeng Yu. Event-based depth estimation with dense occlusion. Optics Letters, 49(12):3376-3379, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[48] Lin Zhu, Siwei Dong, Tiejun Huang, and Yonghong Tian. A retina-inspired sampling method for visual texture reconstruction. In ICME, pages 1432-1437. IEEE, 2019. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[49] Lin Zhu, Siwei Dong, Jianing Li, Tiejun Huang, and Yonghong Tian. Retina-like visual image reconstruction via spiking neural model. In CVPR, pages 1438-1446, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.907, + 0.531 + ], + "angle": 0, + "content": "[50] Lin Zhu, Kangmin Jia, Yifan Zhao, Yunshan Qi, Lizhi Wang, and Hua Huang. Spikenerf: Learning neural radiance fields from continuous spike stream. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6285-6295, 2024. 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.907, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16618" + } + ] +] \ No newline at end of file diff --git a/2025/USP-Gaussian_ Unifying Spike-based Image Reconstruction, Pose Correction and Gaussian Splatting/62c2f0f5-8ea0-4959-8e97-f59984bffdfe_origin.pdf b/2025/USP-Gaussian_ Unifying Spike-based Image Reconstruction, Pose Correction and Gaussian Splatting/62c2f0f5-8ea0-4959-8e97-f59984bffdfe_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..230945649d9112c99e513bae160acca3bece7524 --- /dev/null +++ b/2025/USP-Gaussian_ Unifying Spike-based Image Reconstruction, Pose Correction and Gaussian Splatting/62c2f0f5-8ea0-4959-8e97-f59984bffdfe_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cfe67252ed0be4179025afbf3e79911905b88acfe5d7a6e4b649cab7ce0338d5 +size 2870444 diff --git a/2025/USP-Gaussian_ Unifying Spike-based Image Reconstruction, Pose Correction and Gaussian Splatting/full.md b/2025/USP-Gaussian_ Unifying Spike-based Image Reconstruction, Pose Correction and Gaussian Splatting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..aecece77d39b9751d7ec316725b0cd8f4d9a8573 --- /dev/null +++ b/2025/USP-Gaussian_ Unifying Spike-based Image Reconstruction, Pose Correction and Gaussian Splatting/full.md @@ -0,0 +1,361 @@ +# USP-Gaussian: Unifying Spike-based Image Reconstruction, Pose Correction and Gaussian Splatting + +Kang Chen $^{1,2}$ Jiyuan Zhang $^{1,2}$ Zecheng Hao $^{1,2}$ Yajing Zheng $^{1,2}$ Tiejun Huang $^{1,2,3}$ Zhaofei Yu $^{1,3*}$ $^{1}$ School of Computer Science, Peking University + $^{2}$ State Key Laboratory for Multimedia Information Processing, Peking University + $^{3}$ Institute for Artificial Intelligence, Peking University + +{mrchenkang, jyzhang}@stu.pku.edu.cn, {haozecheng,yj.zheng,tjhuang,yuzf12}@pku.edu.cn + +![](images/333e4d02bd17b5e8661568d0d4bd749fc6ac67f7fe13fafa5e045a43de43365e.jpg) +Figure 1. Left. Illustration of our USP-Gaussian framework, where the spike-based image Reconstruction Network (Recon-Net), camera poses, and 3DGS are collaboratively optimized signified by $\mathbb{O}$ . Mid. Visual ablation showcasing the performance of Recon-Net and 3DGS with and without (w/ & w/o) the joint optimization strategy, with the ablation table depicted in Tab. 4 and the input formulated in Eq. (6). Right. Training curve comparison for Recon-Net and 3DGS with and without joint optimization. + +![](images/77cffedaf199051973ce374e95a5a7d4ebca62ec01f8947bcad9c164c736a129.jpg) + +![](images/e22ac288bc8fe71cef5efdc783edaa7c29b0c2bce11325bc782d07e9b7a1d359.jpg) + +# Abstract + +Spike camera, as an innovative type of neuromorphic camera that captures scenes with 0-1 bit stream at $40\mathrm{kHz}$ , is increasingly being employed for the novel view synthesis task building on techniques such as Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS). Previous spike-based approaches typically follow a three-stage pipeline: I. Spike-to-image reconstruction based on established algorithms. II. Camera poses estimation. III. Novel view synthesis. However, the cascading framework suffers from substantial cumulative errors, i.e., the quality of the initially reconstructed images will impact pose estimation, ultimately limiting the fidelity of the 3D reconstruction. To address this limitation, we propose a synergistic optimization framework USP-Gaussian, which unifies spike-to-image reconstruction, pose correction, and gaussian splatting into an end-to-end pipeline. Leveraging the multi-view consistency afforded by 3DGS and the motion capture capability of the + +spike camera, our framework enables iterative optimization between the spike-to-image reconstruction network and 3DGS. Experiments on synthetic datasets demonstrate that our method surpasses previous approaches by effectively eliminating cascading errors. Moreover, in real-world scenarios, our method achieves robust 3D reconstruction benefiting from the integration of pose optimization. Our code, data, and trained models are available at https://github.com/chenkang455/USP-Gaussian. + +# 1. Introduction + +Spike cameras [12], inspired by the human retina [49], introduce an innovative paradigm for encoding high-speed dynamic scenes. In contrast to conventional imaging systems [22] that rely on fixed exposure intervals to integrate visual information, spike cameras continuously sense dynamics by recording the spike stream with an impressive firing rate of $40\mathrm{kHz}$ . Benefiting from the efficient bit representation of spikes, spike cameras can adeptly detect rapid scenes with significantly diminished bandwidth compared + +to the high-speed cameras such as the Phantom. + +Capitalizing on the remarkable capability of the spike camera to capture intricate motion and texture details in high-speed environments, recent research [5, 36, 41, 44, 45, 48, 49] concentrate on establishing the mapping from bit-encoded spike to visually perceivable images. Subsequent investigations delve into the estimation of optical flow [43], facilitation of motion deblur [3, 8, 38] and applications such as high-speed object detection [46], motion magnification [33], depth estimation [34] and occlusion removal [35]. + +Neural Radiance Fields [21] and 3D Gaussian Splitting [13] have emerged as prominent novel view synthesis technologies benefiting from their designed differential training frameworks. Among them, 3DGS employs Gaussian primitives for explicit scene representation and the splatting algorithm for projection, thereby achieving faster rendering speed compared to the pixel-by-pixel approach implemented in NeRF. Building on NeRF and 3DGS, the high-speed and continuous capturing potential of the spike camera can be fully unleashed in 3D reconstruction tasks [11, 37, 50]. + +While the idea of applying the spike camera for 3D reconstruction is trivial and straightforward in a sequential manner, i.e., we can reconstruct images from the spike stream, estimate camera poses and subsequently apply NeRF/3DGS to obtain the 3D representation. Nevertheless, the cascading pipeline introduces significant cumulative errors when the quality of initially recovered images is poor, thereby degrading the accuracy of camera pose estimation and limiting the fidelity of the reconstructed scene. + +For example, SpikeGS [37] initiates with a self-supervised spike-to-image reconstruction framework built on the Blind Spot Network (BSN) [15], intended to mitigate the performance degradation [6, 7] encountered by supervised learning methods on real-world scenes. With recovered high-quality images, SpikeGS designs a further constraint from the long spike input and integrates it into the vanilla 3DGS training pipeline. While SpikeGS have improved preliminary representation for 3DGS, it remains constrained within the cascading framework thus inheriting the aforementioned cumulative error problem. + +To address these challenges, we propose USP-Gaussian, Unifying Spike-based image reconstruction, Pose correction, and Gaussian Splatting within a one-stage optimization framework to mitigate error propagation inherent in the three-stage cascading pipeline. Specifically, USP-Gaussian comprises two synergistic branches as illustrated in Fig. 1, one dedicated to spike-to-image reconstruction via Recon-Net and the other to novel view synthesis implemented with 3DGS. To facilitate the training of Recon-Net, we design $\mathcal{L}_{\mathrm{rec}}$ based on the constraint between the input spike and recovered outputs. Besides, we supervise the training of 3DGS based on $\mathcal{L}_{\mathrm{gs}}$ , which obviates the require + +ment for pre-recovered high-quality images. While each branch of USP-Gaussian operates independently, we further introduce the joint loss $\mathcal{L}_{\mathrm{joint}}$ derived from the outputs of both two branches to facilitate unified optimization. + +Visual ablation comparisons of restored images and the training curve depicted in Fig. 1 demonstrate that the joint optimization of the two branches mutually enhances the recovered textures and the overall training process. We conduct quantitative and qualitative evaluations on the synthetic dataset and the real-world dataset. Experiments showcase that our method achieves superior restoration in both datasets, validating the effectiveness of our joint learning framework. In summary, the contributions of our USP-Gaussian are as follows: + +- We propose a unified framework for the joint optimization of Recon-Net, Pose, and 3DGS, mitigating error amplification inherent in cascading frameworks. +- We demonstrate that spike-to-image and 3D reconstruction tasks can mutually facilitate and enhance the optimization of each other. +- We introduce a novel self-supervised training paradigm for the spike-based image reconstruction network with the multi-view consistency afforded by 3DGS. + +# 2. Related Work + +# 2.1. Spike-based Image Reconstruction + +While neuromorphic cameras generate high frame rate bitstream information, it is imperative to transform them into multi-bit grayscale images that are perceptible to human observers. To achieve this spike-to-image target, a plethora of methodologies [5-7, 36, 41, 44, 45, 49] have been proposed, which can be broadly classified into model-based, supervised, and self-supervised learning paradigms. Model-based methods, rooted in the sampling principle of the spike camera, aim to infer high-fidelity images drawing from the physical spike-image relationship [48], biological retinal models [49], and short-term synaptic plasticity [45]. Confronted with pervasive dark current noise that is challenging to model, learning-based methods are employed to directly deduce the sharp image from spike input with unknown firing patterns. Specifically, Spk2ImgNet [41] designed a hierarchical architecture engineered to capture the intricate temporal dynamics present within spike streams. SpikeCLIP [5] utilized the textual description of the captured scene as supervision for training the spike-to-image network. Concurrently, self-supervised learning methodologies based on the BSN [6, 7] are developed to counteract the performance deterioration in supervised learning methods owing to the domain gap between synthetic and real-world datasets. + +In this paper, we propose a novel self-supervised spike-based image reconstruction framework leveraging the multi-view constraints provided by 3D scenes. + +# 2.2. Novel View Synthesis based on the Neuromorphic Camera + +Recent research has extensively employed neuromorphic cameras, including event cameras [10, 16, 18, 27, 47], which encode intensity difference in dynamic scenarios with low-latency and spike cameras [12], for 3D scene reconstruction. + +Event Camera. Klenk et al. [14] proposed the E-NeRF based on the event generation mechanism, which can recover sharp 3D scenes under high-speed camera motion. Rudnev et al. [23] further harnessed a color event camera to learn the color representation of scenes. [1, 20] focused on recovering dynamic scenes with rigid deformations from event streams. In addition to NeRF, EvGGS [29] introduces a generalizable Gaussian splatting framework for event cameras designed to reconstruct 3D scenes without retraining. EF-3DGS [17] develops a free-trajectory framework by aligning the RGB input, event streams, and pose. Event3DGS [31] presents an effective solution for rapid ego-motion robotics scenarios. + +Spike Camera. Zhu et al. [50] design a differentiable SNN and integrate it into the NeRF training pipeline to simulate the generation of spike streams, where the loss function is formulated based on the rendered spike output and the input spike. Dai et al. [9] further introduce SpikeNVS by extending the spike cameras to RGB-Spike input. Additionally, Guo et al. [11] introduced a color 3D scene reconstruction framework by integrating Bayer-pattern spike streams into the 3DGS training pipeline. SpikeGS [37] initially employs the BSN to recover images from the noisy spike stream and subsequently integrates them into the 3DGS. + +While the cascaded framework SpikeGS [37] introduces propagation errors, we propose a joint optimization framework that concurrently handles spike-based image reconstruction, pose estimation, and novel view synthesis. + +# 3. Preliminary + +# 3.1. Spike Camera + +The spike camera consists of the photodetector, voltage integrator, comparator, and synchronous readout as shown in Fig. 2. Within the designated time interval $\mathcal{T} = [t_s,t_e]$ , the photodetector continuously captures the light intensity $\mathbf{I}(t)$ and converts it into the voltage signal $\mathbf{V}(t) = \mu \mathbf{I}(t)$ , where $\mu$ is the photoelectric conversion coefficient. Concurrently, the comparator assesses the integrator voltage $\mathbf{A}(t)$ against the predefined threshold $\Theta$ . When the threshold is met, the spike is emitted and the integrator voltage is reset, formulated as: + +$$ +\mathbf {A} (t) = \int_ {t _ {s}} ^ {t} \mathbf {I} (t) d t \bmod C, \tag {1} +$$ + +where $C = \Theta /\mu$ is a constant proportional to the spike firing threshold $\Theta$ . While the integrator continuously ag- + +![](images/5208623a539375b59ed4bbb1ffea88a52bd7f97c8fed35547893dfd0382f4a5f.jpg) +Figure 2. (a) The spike camera captures a railway moving 350 km/h. (b) The Recon-Net is designed to establish the mapping between the input spike stream and the sharp image. (c) The working mechanism of the spike camera. + +gregates voltage, the readout circuit synchronously retrieves spike signals from the registers at up to $40\mathrm{kHz}$ . Spikes fired during the specified interval $\mathcal{T}$ are represented as a three-dimensional bit stream $S\in \{0,1\}^{K\times H\times W}$ , where $H$ and $W$ indicate the image height and width, with $K$ denoting the spike sequence length. + +# 3.2. 3D Gaussian Splatting + +3D Gaussian Splating leverages a set of three-dimensional Gaussian primitives, symbolized as $\{\mathcal{G}\}_{n=1}^{N_{\mathrm{gs}}}$ , to effectively represent the spatial characteristics of the captured scene. + +Each Gaussian primitive is centered at a point $\mathbf{p}_n$ , with its spread and orientation influenced by the covariance matrix $\Sigma_n = RSS^T R^T$ , where $R$ and $T$ denotes the rotation and the scale matrix, respectively. The impact of the Gaussian on any point $\mathbf{v}$ in the scene is quantified by: + +$$ +\mathcal {G} _ {n} (\mathbf {v}) = \exp \left(- \frac {1}{2} (\mathbf {v} - \mathbf {p} _ {n}) ^ {T} \boldsymbol {\Sigma} _ {n} ^ {- 1} (\mathbf {v} - \mathbf {p} _ {n})\right). (2) +$$ + +When it comes to rendering for viewing transformation $W$ , these 3D Gaussian primitives are mapped onto 2D primitives $\mathcal{G}_n^{2D}$ , with the covariance matrix $\Sigma_n^\prime$ reformulated as: + +$$ +\Sigma_ {n} ^ {\prime} = J W \Sigma_ {n} W ^ {T} J ^ {T}, \tag {3} +$$ + +where $J$ represents the Jacobian matrix derived from the affine approximation of the projective transformation. + +Further, each pixel on the rendered plane is calculated by rasterizing these 2D Gaussians, sorted by depth, according to the following formulation: + +$$ +C (r) = \sum_ {k = 1} ^ {N _ {\mathrm {g s}}} T _ {i} \alpha_ {i} \mathbf {c} _ {i}, T _ {i} = \prod_ {j = 1} ^ {i - 1} \left(1 - \alpha_ {j}\right), \tag {4} +$$ + +![](images/1f235a959be2d89218507a7cb10908cc685b0584d3ab39210d469737fcc3e0a2.jpg) +Figure 3. The working pipeline of our USP-Gaussian. For each camera viewpoint, we initially derive a pose sequence at equal time intervals based on the camera model described in Sec. 4.3. Leveraging the 3DGS rendering pipeline detailed in Sec. 4.2, we subsequently generate a corresponding sharp sequence for this viewpoint. Concurrently, Recon-Net is employed to reconstruct the same sequence from the captured spike stream as outlined in Sec. 4.1. While Recon-Net and 3DGS are trained based on the discrepancy between the re-synthesized blurry image with the input long-exposure image, alignment of their outputs for the same scene is further enforced through joint optimization loss as described in Sec. 4.4. With formulated loss functions, our USP-Gaussian enables the collaborative optimization of Gaussian primitives, camera poses, and the parameters of Recon-Net as formulated in Eq. (5). + +where $D$ represents the indices set reordered by depth, and $D_{k}$ denotes the index of the Gaussian primitive that is the $k$ -th closest to the rendering point. + +# 4. Methods + +Target Clarification. Given the captured spike stream set $\{S_v\}_{v=1}^V$ from $V$ different viewpoints with initial inaccurate poses $\{\mathcal{P}_v\}_{v=1}^V$ , the USP-Gaussian aims to optimize the parameters of Recon-Net and Gaussian primitives while outputting refined poses as shown in Fig. 1, formulated as: + +$$ +(\mathbb {S}, \mathbb {P}, \mathbb {G}, \theta) \xrightarrow {\text {U S P - G a u s s i a n}} (\mathbb {P} ^ {*}, \mathbb {G} ^ {*}, \theta^ {*}), \tag {5} +$$ + +where $\mathbb{S}$ and $\mathbb{P}$ represent the sets of spike streams and poses from $M$ viewpoints, $\mathbb{G}$ denotes the ensemble of Gaussian primitives, and $\theta$ encapsulates the Recon-Net parameters. The notation $(^{*})$ indicates the optimized parameters. + +# 4.1. Spike-based Image Reconstruction Network + +The Recon-Net is designed to facilitate the mapping $\mathcal{F}(\cdot)$ from the spike input to the sharp image as shown in Fig. 2(b), formulated as $\mathbf{I}(f) = \mathcal{F}(S_{\mathcal{T}};\theta)$ with $f$ signifying the central moment of the interval $\mathcal{T}$ . Prior research [6, 36, 41] employs a fixed and brief window (41 frames in their configuration) of the spike stream as the network in + +put. However, it is insufficient for scenarios with low spike firing rates due to the limited information embedded within. + +While extending the window size allows the network to encompass more features, it complicates the feature extraction process and demands large computational resource. To address it, we configure the input of the Recon-Net with a complementary long-short spike stream as shown in Fig. 3. + +Given a long spike stream $S_{\mathcal{T}}$ , we extract a short spike stream $S_{\mathcal{T}'}^t \subset S_{\mathcal{T}}$ centered around time $t$ , with the network mapping reformulated as $\mathbf{I}(t) = \mathcal{F}(S_{\mathcal{T}}, S_{\mathcal{T}'}^t, t; \theta)$ , where time $t$ is incorporated to guide the alignment of the short and long spike. Following the event stream pre-processing in [2, 26, 28], we apply voxelization to the long spike stream $S_{\mathcal{T}}$ , reducing its dimensions from $K \times H \times W$ to $\frac{K}{4} \times H \times W$ accumulating every 4 frames. The network architecture consists of basic convolutional blocks as illustrated in Fig. 3 and detailed in the supplementary materials. + +According to the spike camera working mechanism described in Eq. (1), we can reconstruct a long-exposure image $\mathbf{E}(\mathcal{T})$ from the spike input $S_{\mathcal{T}}$ with the TFP algorithm [48], mathematically formulated as below: + +$$ +\mathbf {E} (\mathcal {T}) = \frac {1}{T} \int_ {\mathcal {T}} \mathbf {I} (t) d t = \frac {C \cdot N}{T}, \qquad (6) +$$ + +where $T$ denotes the duration of the interval $\mathcal{T}$ , and $N$ is the cumulative number of spikes during this period. + +Motivated by the textural similarity between the blurry image and the long-exposure image, we can employ the motion-reblur loss widely employed in self-supervised motion deblurring frameworks [3, 32, 39, 40] to guide the training of the Recon-Net. We initially reconstruct $M$ frames uniformly across the interval to form the image sequence $\{\mathbf{I}_{\mathrm{rec}}(t_m)\}_{m = 1}^M$ and further average them to synthesize the reblur image. The loss can be formulated by comparing the reblur image to the long-exposure image, expressed as: + +$$ +\mathbb {L} _ {\mathrm {r e c}} (M, \mathcal {T}) = \mathcal {L} \left(\frac {1}{M} \sum_ {m = 1} ^ {M} \mathbf {I} _ {\mathrm {r e c}} \left(t _ {m}\right), \mathbf {E} (\mathcal {T})\right), \tag {7} +$$ + +where the loss $\mathcal{L}$ [13] combines $L_{1}$ loss with a D-SSIM term, defined as $\mathcal{L} = (1 - \lambda)\mathcal{L}_1 + \lambda \mathcal{L}_{\mathrm{D - SSIM}}$ , effectively integrating both pixel accuracy and structural similarity. + +Nevertheless, the input spike stream $S_{\mathcal{T}}$ sufficiently captures the information to estimate $\mathbf{E}(\mathcal{T})$ as detailed in Eq. (6), which will lead the Recon-Net to learn a trivial identity mapping from the spike to the long exposure image. To mitigate it, we propose the multi-reblur loss to regularize the reconstructed sequence. Specifically, we extract $N$ sub-intervals from the long exposure $T_{n} \subset \mathcal{T}$ with the same central moment and apply the reblur loss to each sub-interval, formulating the multi-reblur loss as below: + +$$ +\mathcal {L} _ {\mathrm {r e c}} = \frac {1}{N} \sum_ {n = 1} ^ {N} \mathbb {L} _ {\mathrm {r e c}} \left(M _ {n}, \mathcal {T} _ {n}\right), \tag {8} +$$ + +where $M_{n}$ represents the number of reconstructed frames within the $n$ -th sub-interval $\mathcal{T}_n$ . + +# 4.2. Spike-based Gaussian Splatting + +Suppose that retrieving the pose for each timestamp within the interval is straightforward. Following Sec. 4.1, we uniformly obtain $M$ poses from the interval $\mathcal{T}$ , yielding a pose sequence $\{\mathbf{T}(t_m)\}_{m=1}^M$ with the corresponding image sequence $\{\mathbf{I}_{\mathrm{gs}}(t_m)\}_{m=1}^M$ rendered via the 3DGS rasterization pipeline as shown in Fig. 3. + +While reconstructing a sequence of images from the spike stream by shortening the exposure interval in Eq. (6) for the training of 3DGS sounds feasible, these short-exposure images are noisy and lack texture details [3]. To this end, we utilize the long-exposure image $\mathbf{E}(\mathcal{T})$ as the supervision and design the motion-reblur loss for 3DGS similar to Eq. (7), expressed as: + +$$ +\mathcal {L} _ {\mathrm {g s}} = \mathcal {L} \left(\frac {1}{M} \sum_ {m = 1} ^ {M} \mathbf {I} _ {\mathrm {g s}} \left(t _ {m}\right), \mathbf {E} (\mathcal {T})\right). \tag {9} +$$ + +# 4.3. Spike Camera Trajectory Modeling + +Building on the learnable pose representation during camera exposure in previous research [30, 42], we employ a + +similar approach to optimize the camera pose throughout the interval $\mathcal{T}$ as shown in Fig. 3. For each spike stream $S_{\mathcal{T}}$ , we optimize both the start pose $\mathbf{T}_{\mathrm{start}}$ and end pose $\mathbf{T}_{\mathrm{end}}$ over the interval, while intermediate poses are represented through interpolation between these two. + +For the pose $\mathbf{T}(t_m)$ at time $t_m$ , we utilize linear interpolation in the Lie algebra of $\mathrm{SE}(3)$ to estimate the pose at time $t$ , formulated as: + +$$ +\mathbf {T} \left(t _ {m}\right) = \mathbf {T} _ {\text {s t a r t}} \cdot \exp \left(\frac {t _ {m}}{T} \cdot \log \left(\mathbf {T} _ {\text {s t a r t}} ^ {- 1} \cdot \mathbf {T} _ {\text {e n d}}\right)\right). \tag {10} +$$ + +During the training, the pose set $\mathbb{P}$ consisting of the start and end poses from $V$ viewpoints is jointly optimized with the Recon-Net and Gaussian primitives as in Eq. (5). For further details on the camera trajectory modelling, please refer to BAD-NeRF [30] and BAD-Gaussian [42]. + +# 4.4. Joint Optimization + +Given the spike stream $S_{\mathcal{T}}$ along with the start pose $\mathbf{T}_{\mathrm{start}}$ and the end pose $\mathbf{T}_{\mathrm{end}}$ to be optimized, we perform evenly spaced sampling interpolation as in Eq. (10) during the interval, obtaining the pose sequence $\{\mathbf{T}(t_m)\}_{m=1}^M$ . + +For each pose in the sequence, we select a short spike stream centered around the timestamp $t_m$ , obtaining the image sequence $\{\mathbf{I}_{\mathrm{rec}}(t_m)\}_{m=1}^M$ based on the Recon-Net in Sec. 4.1 and the image sequence $\{\mathbf{I}_{\mathrm{gs}}(t_m)\}_{m=1}^M$ rendered from the 3DGS in Sec. 4.2. While both sequences represent the same scene, the joint optimization loss $\mathcal{L}_{\mathrm{joint}}$ can be formulated as follows: + +$$ +\mathbb {L} _ {\text {j o i n t}} = \frac {1}{M} \sum_ {m = 1} ^ {M} \left| \left| \mathbf {I} _ {\mathrm {g s}} \left(t _ {m}\right) - \mathbf {I} _ {\mathrm {r e c}} \left(t _ {m}\right) \right| \right| _ {2} ^ {2}. \tag {11} +$$ + +However, we found that the joint optimization framework tends to be unstable due to the order difference between the 3DGS and Recon-Net output sequences. Specifically, since the temporal order does not affect the loss in Eq. (9), the optimized pose sequence may reverse relative to the real camera trajectory. To address this, we apply a simple but effective flip-and-minimum operation to align both sequences, with the loss function reformulated as: + +$$ +\mathcal {L} _ {\text {j o i n t}} = \min \left(\mathbb {L} _ {\text {j o i n t}}, \mathbb {L} _ {\text {j o i n t}} ^ {\mathrm {r}}\right), \tag {12} +$$ + +$$ +\mathbb {L} _ {\text {j o i n t}} ^ {\mathrm {r}} = \frac {1}{M} \sum_ {m = 1} ^ {M} \left| \left| \mathbf {I} _ {\mathrm {g s}} \left(t _ {M - m + 1}\right) - \mathbf {I} _ {\mathrm {r e c}} \left(t _ {m}\right) \right| \right| _ {2} ^ {2}, \tag {13} +$$ + +where the difference between the $\mathbb{L}_{\mathrm{joint}}^{\mathrm{r}}$ and the $\mathbb{L}_{\mathrm{joint}}$ lies in that the rendered sequence via 3DGS is reversed. + +To sum up, the loss function for our USP-Gaussian framework is presented as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {r e c}} + \mathcal {L} _ {\text {g s}} + \mathcal {L} _ {\text {j o i n t}}. \tag {14} +$$ + +where $\mathcal{L}_{\mathrm{rec}}$ optimizes Recon-Net parameters, $\mathcal{L}_{\mathrm{gs}}$ refines poses and Gaussian primitives, and $\mathcal{L}_{\mathrm{joint}}$ jointly optimizes all components, with gradient flow shown in Fig. 3. + +![](images/37a8e8279c2392d64e562911ce0358a313ccd3b75da7618652ce045db769c956.jpg) +Figure 4. 3D reconstruction visual comparison of our USP-Gaussian compared with previous methods on the synthetic dataset, where the input is the long-exposure image defined in Eq. (6). + +Table 1. 3D reconstruction quantitative comparison on the synthetic dataset. Color shading indicates the best and second-best result. + +
MethodsWineTanabataFactoryOutdoorpoolAvg
PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
TFP-3DGS [48]24.1380.7960.23924.6260.7780.27427.8450.8520.17928.0240.7860.29426.1580.8030.239
TFI-3DGS [48]23.8460.7330.25121.1680.5500.52925.0190.7340.27824.7550.6370.44623.6970.6630.376
Spk2Img-3DGS [41]21.1500.7890.17820.4510.7580.21425.1150.8060.14323.3620.6470.41722.5200.7500.238
SpikeNeRF [50]25.8880.7980.20025.9650.7250.29527.5310.7800.19125.6420.6570.39626.2570.7400.270
SpikeGS [37]25.6980.8350.24726.1740.8260.24328.6270.8640.21428.2870.7970.27327.1960.8320.244
Ours26.4760.8620.17626.3000.8270.21328.6940.8610.20730.1420.8200.27027.9030.8430.217
+ +# 5. Experiment + +# 5.1. Datasets + +Synthetic Dataset. We construct the synthetic dataset based on Deblur-NeRF [19] scenes for quantitative evaluation. We start with rendering 13 sharp images for each viewpoint, with further interpolation algorithm XVFI [25] to insert 7 images between each frame pair. We apply the spike simulator on the interpolated sequence to emulate the real-world spike as in [3], thus generating 97 spike frames for each viewpoint. + +Real-world Dataset. To construct the real-world dataset, + +we shake the spike camera extremely fast while maintaining a fixed forward-facing orientation and extract 97 spike frames per view, consistent with the configuration in the synthetic dataset. For pose estimation, we convert the raw spike stream in each view to the short-exposure image as defined in Eq. (6) and subsequently employ COLMAP [24] to estimate the pose sequence. + +# 5.2. Experimental Results + +We compare our USP-Gaussian with two-stage cascading pipelines and approaches specially designed for the spike camera, i.e., SpikeNeRF [50] and SpikeGS [37]. For the + +![](images/b5401e5ece889c1735c530f2b253351cb8c6da0c9f3640325aaf50f09d67310f.jpg) +Figure 5. 3D reconstruction visual comparison of our USP-Gaussian compared with previous methods on the real-world dataset. + +Table 2. Comparison of our method with SpikeGS on the wine dataset with inaccurate perturbed poses. + +
MetricsSpikeGSOurs
0%10%20%30%0%10%20%30%
PSNR ↑25.6918.8317.7516.4426.4726.0525.9323.46
SSIM ↑0.8350.6070.5920.5670.8620.8540.8510.772
LPIPS ↓0.2470.3720.4110.4460.1760.1840.1860.245
+ +two-stage approach, we initially apply TFP [48], TFI [48], and Spk2ImgNet [41] to reconstruct sharp images from the input spike followed by the common 3DGS processing pipeline, with the code implemented by Spike-Zoo [4]. All Gaussian-based methods are conducted based on the same 3DGS backbone, and identical camera poses input for a fair comparison. While SpikeGS requires several poses per spike stream for the exposure loss, which presents a notable challenge for COLMAP in real-world data, we omit the exposure loss and retain the rest of the framework. All comparative metrics are calculated based on the 3D reconstruction task, with further comparison on the spike-to-image reconstruction task and implementation details available in the supplementary materials. + +Synthetic Dataset. Quantitative and qualitative comparisons on the synthetic dataset are shown in Tab. 1 and Fig. 4 respectively. As illustrated in the Tab. 1, our USP-Gaussian achieves the best performance across most metrics in various scenes. Two-stage cascading methods TFP-3DGS and TFI-3DGS yield poor restoration due to the substantial noise embedded in the initially reconstructed sequence. Similarly, although Spk2ImgNet performs well on its original REDS dataset [41], it experiences substantial degradation due to the dataset domain gap. SpikeGS, as the primary comparison algorithm in this paper, heavily relies on the BSN [15] for self-supervised Spike-to-Image reconstruction. Nevertheless, the denoising performance of the BSN deteriorates in the presence of substantial noise, + +![](images/66a58d80780aea1dbd490ce733c76f8de17c71039039fbf4e62018ce76ad5cc6.jpg) +Pose Trajectory with $10\%$ Perturbation +Figure 6. Visual comparison of initial and optimized poses. + +![](images/36a96113877b65220968cff14aa1c85ae9b68057a14edf353ec6a47b98b11da5.jpg) +Pose Trajectory with $30\%$ Perturbation + +Table 3. Translation and rotation poses error comparison. + +
StateTranslation ↓Rotation ↓
10%20%30%Avg10%20%30%Avg
Initial1.3022.9674.1932.8201.1322.1503.1592.147
Optimized0.8551.7811.4931.3760.1740.3201.2850.593
Δ (%)34.3%39.9%64.4%51.3%84.6%85.1%59.2%72.4%
+ +leading to pronounced grid artifacts and the attenuation of high-frequency details of the SpikeGS. SpikeNeRF faces a comparable challenge, as it requires noisy spike images as input and relies on the SNN for denoising. + +To this end, our proposed collaborative learning framework USP-Gaussian, benefiting from the inherent constraints between the spike stream and the multi-view consistency afforded by the 3DGS, effectively addresses the aforementioned problems without relying on the pre-training of the spike-based image reconstruction network. + +In the previous part, accurate poses are obtained by applying COLMAP on sharp images rendered by Blender. To demonstrate that the integration of pose correction enhances the robustness of our method, we apply random perturbations of $10\%$ , $20\%$ , and $30\%$ to the accurate poses in the wine scene, generating inaccurate poses for SpikeGS and our method with the comparison shown in Tab. 2. + +Furthermore, visual trajectories and quantitative errors + +Table 4. Quantitative ablation on the synthetic dataset. + +
ID\( \mathcal{L}_{\text{gs}} \)Recon-Net\( \mathcal{L}_{\text{joint}} \)3DGS OutputsRecon-Net Outputs
\( \mathbb{L}_{\text{rec}} \)\( \mathbb{L}_{\text{rec}} \)PSNRSSIMLPIPSPSNRSSIMLPIPS
I✓22.3920.7500.208\\\
II✓\\\24.2550.7080.358
III✓\\\24.5670.7440.332
IV✓✓✓25.6590.8500.18425.5710.8070.244
V✓✓✓26.4760.8620.17627.1380.8550.178
+ +![](images/3b6d9336d9428282244b8c015c97d307596215c9132f3ceeb6839604aa1ca878.jpg) +Figure 7. Visual ablation on the effectiveness of multi-reblur loss, with the x-axis and y-axis representing step (k) and PSNR. + +![](images/c4f9d7ad94ec750ddedc8f23e68394e0223686d5071d530669adfdb4d3601022.jpg) +Figure 8. Visual ablation on the effectiveness of long spike input. + +of the initial noisy, and optimized poses are depicted in Fig. 6 and Tab. 3 respectively, with MAE measuring translation error and radians measuring rotation error, demonstrating the pose correction capability of our USP-Gaussian. + +Real-world Dataset. We also conduct the visual comparison on the real-world dataset as shown in Fig. 5. Due to the rapid camera ego-motion and the low imaging quality of the spike camera characterized by low spatial resolution, insufficient texture details, and underexposure, the pose sequence estimated by the COLMAP is highly inaccurate, presenting significant challenges for previous methods to recover accurate 3D representation as observed in Fig. 5. Nevertheless, our USP-Gaussian overcomes this limitation benefiting from the unifying optimization of the pose. + +# 5.3. Ablation Study + +In this subsection, we demonstrate the effectiveness of our proposed joint optimization, complementary spike input, and multi-reblur loss strategies. + +I. Joint Learning. The highlight of USP-Gaussian lies in the design of the joint loss, which aligns outputs from the 3DGS and Recon-Net thus achieving complementary performance improvement. We conduct the ablation experiment on loss functions of $\mathcal{L}_{\mathrm{gs}}$ , $\mathcal{L}_{\mathrm{rec}}$ and $\mathcal{L}_{\mathrm{joint}}$ with quantitative results listed in Tab. 4, which shows that the incorporation of the joint loss (ID-V) yields notable improvements for both 3DGS and Recon-Net compared to independent optimization (ID-I and ID-III). + +Further visual ablation comparisons of the restored image and the training curve are presented in Fig. 1. Visual comparison substantiates that the joint optimization of 3DGS and Recon-Net significantly enhances the texture details for both two. Regarding the training curve, the inde + +![](images/880659126e051fac901ed855ca02e65a271eecad1e13d3c0807115b5ae865b80.jpg) +TFP-Input + +![](images/12a12df4201757396d4581728e8baa552e08c98e06ef4579be5bc3b2749bf7fc.jpg) +w/o Long Spike + +![](images/20556ac5b277f054d730d8f4dff0dcef3ca4c0c9ebffa9ac42441f85d151bda5.jpg) +w/ Long Spike + +pendent 3DGS training (ID-I) demonstrates slow convergence due to the absence of crucial information in the long-exposure images. Concurrently, the independent training of Recon-Net (ID-III) tends to overfit due to the simplicity of the spike reconstruction task. To this end, our proposed joint optimization framework provides synergistic benefits for both, i.e., Recon-Net effectively compensates for deficient motion features embedded in the long-exposures image while the multi-view consistency afforded by 3DGS culminates in coherent outputs from the Recon-Net. + +II. Multi-reblur Loss. The multi-reblur loss $\mathcal{L}_{\mathrm{rec}}$ in Eq. (7) is designed to substitute the single-reblur loss $\mathbb{L}_{\mathrm{rec}}$ in Eq. (8). Without further constraints, $\mathbb{L}_{\mathrm{rec}}$ will lead ReconNet to learn an identity mapping from the input spike stream to the corresponding long-exposure image as demonstrated in [3]. Comparative analysis in Tab. 4 depicts that $\mathcal{L}_{\mathrm{rec}}$ (ID-V) yields beneficial improvement than $\mathbb{L}_{\mathrm{rec}}$ (ID-IV). Further comparison on the training curve in Fig. 7 reveals the mechanism of collaborative training, i.e., the multi-reblur loss enhances the performance of Recon-Net, thereby facilitating the training of the 3DGS. + +III. Complementary Input. We utilize complementary long and short spike stream inputs for Recon-Net to address the fixed short window limitation inherent in previous methods [6, 36, 41]. As illustrated in the visual comparison in Fig. 8, the inclusion of the long spike effectively suppresses salt-and-pepper noise in the Recon-Net output, leading to improved SSIM and LPIPS metrics. + +# 6. Conclusion + +In this paper, we propose a unifying framework USP-Gaussian for the collaborative optimization of spike-based image reconstruction, pose correction, and Gaussian Splating. By mitigating error propagation inherent in traditional cascading frameworks, USP-Gaussian attains high reconstruction fidelity across both synthetic and real-world datasets, with further ablation experiments revealing that Spike-to-Image reconstruction and 3D reconstruction tasks furnish mutually complementary information. + +Limitation. The limitation of USP-Gaussian is that it requires longer training time and more GPU memory compared to previous methods due to the additional poses and Recon-Net parameters to be optimized. + +# Acknowledgments + +We sincerely appreciate Yuyan Chen (HUST) for her valuable suggestions and for polishing the figures. This work was supported by the National Natural Science Foundation of China (62422601, 62176003, 62088102, 62306015), the China Postdoctoral Science Foundation (2023T160015), the Young Elite Scientists Sponsorship Program by CAST (2023QNRC001), the Beijing Municipal Science and Technology Program (Z241100004224004) and the Beijing Nova Program (20230484362,20240484703). + +# References + +[1] Anish Bhattacharya, Ratnesh Madaan, Fernando Cladera, Sai Vemprala, Rogerio Bonatti, Kostas Daniilidis, Ashish Kapoor, Vijay Kumar, Nikolai Matni, and Jayesh K Gupta. Evdnerf: Reconstructing event data with dynamic neural radiance fields. In WACV, pages 5846-5855, 2024. 3 +[2] Kang Chen and Lei Yu. Motion deblur by learning residual from events. IEEE TMM, 2024. 4 +[3] Kang Chen, Shiyan Chen, Jiyuan Zhang, Baoyue Zhang, Yajing Zheng, Tiejun Huang, and Zhaofei Yu. Spikereveal: Unlocking temporal sequences from real blurry inputs with spike streams. Advances in Neural Information Processing Systems, 37:62673-62696, 2025. 2, 5, 6, 8 +[4] Kang Chen, Zhiyuan Ye, Tiejun Huang, and Zhaofei Yu. Spike-Zoo: A toolbox for spike-to-image reconstruction. https://github.com/chenkang455/Spike-Zoo, 2025.7 +[5] Kang Chen, Yajing Zheng, Tiejun Huang, and Zhaofei Yu. Rethinking high-speed image reconstruction framework with spike camera. arXiv preprint arXiv:2501.04477, 2025. 2 +[6] Shiyan Chen, Chaoteng Duan, Zhaofei Yu, Ruiqin Xiong, and Tiejun Huang. Self-supervised mutual learning for dynamic scene reconstruction of spiking camera. In IJCAI, pages 2859-2866, 2022. 2, 4, 8 +[7] Shiyan Chen, Zhaofei Yu, and Tiejun Huang. Self-supervised joint dynamic scene reconstruction and optical flow estimation for spiking camera. In AAAI, pages 350-358, 2023. 2 +[8] Shiyan Chen, Jiyuan Zhang, Yajing Zheng, Tiejun Huang, and Zhaofei Yu. Enhancing motion deblurring in high-speed scenes with spike streams. In NeurIPS, 2023. 2 +[9] Gaole Dai, Zhenyu Wang, Qinwen Xu, Wen Cheng, Ming Lu, Boxing Shi, Shanghang Zhang, and Tiejun Huang. Spikenvs: Enhancing novel view synthesis from blurry images via spike camera. arXiv preprint arXiv:2404.06710, 2024.3 +[10] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE TPAMI, 44(1):154-180, 2020. 3 +[11] Yijia Guo, Liwen Hu, Lei Ma, and Tiejun Huang. Spikegs: Reconstruct 3d scene via fast-moving bio-inspired sensors. arXiv preprint arXiv:2407.03771, 2024. 2, 3 + +[12] Tiejun Huang, Yajing Zheng, Zhaofei Yu, Rui Chen, Yuan Li, Ruiqin Xiong, Lei Ma, Junwei Zhao, Siwei Dong, Lin Zhu, et al. $1000\times$ faster camera and machine vision with ordinary devices. Engineering, 25:110-119, 2023. 1, 3 +[13] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 5 +[14] Simon Klenk, Lukas Koestler, Davide Scaramuzza, and Daniel Cremers. E-nerf: Neural radiance fields from a moving event camera. IEEE Robotics and Automation Letters, 8 (3):1587–1594, 2023. 3 +[15] Samuli Laine, Tero Karras, Jaakko Lehtinen, and Timo Aila. High-quality self-supervised deep image denoising. NeurIPS, 32, 2019. 2, 7 +[16] Wenpu Li, Pian Wan, Peng Wang, Jinghang Li, Yi Zhou, and Peidong Liu. Benerf: Neural radiance fields from a single blurry image and event stream. In European Conference on Computer Vision (ECCV), 2024. 3 +[17] Bohao Liao, Wei Zhai, Zengyu Wan, Tianzhu Zhang, Yang Cao, and Zheng-Jun Zha. Ef-3dgs: Event-aided free-trajectory 3d gaussian splatting. arXiv preprint arXiv:2410.15392, 2024. 3 +[18] Zibin Liu, Banglei Guan, Yang Shang, Shunkun Liang, Zhenbao Yu, and Qifeng Yu. Optical flow-guided 6dof object pose tracking with an event camera. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6501-6509, 2024. 3 +[19] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-nerf: Neural radiance fields from blurry images. In CVPR, pages 12861-12870, 2022. 6 +[20] Qi Ma, Danda Pani Paudel, Ajad Chhatkuli, and Luc Van Gool. Deformable neural radiance fields using rgb and event cameras. In ICCV, pages 3590-3600, 2023. 3 +[21] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2 +[22] Bryan Peterson. Understanding exposure: how to shoot great photographs with any camera. AmPhoto books, 2016. 1 +[23] Viktor Rudnev, Mohamed Elgharib, Christian Theobalt, and Vladislav Golyanik. Eventnerf: Neural radiance fields from a single colour event camera. In CVPR, pages 4992-5002, 2023. 3 +[24] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, pages 4104-4113, 2016. 6 +[25] Hyeonjun Sim, Jihyong Oh, and Munchurl Kim. Xvfi: extreme video frame interpolation. In ICCV, pages 14489-14498, 2021. 6 +[26] Chen Song, Qixing Huang, and Chandrajit Bajaj. E-cir: Event-enhanced continuous intensity recovery. In CVPR, pages 7803-7812, 2022. 4 +[27] Jing Tao, You Li, Banglei Guan, Yang Shang, and Qifeng Yu. Simultaneous enhancement and noise suppression un + +der complex illumination conditions. IEEE Transactions on Instrumentation and Measurement, 2024. 3 +[28] Bishan Wang, Jingwei He, Lei Yu, Gui-Song Xia, and Wen Yang. Event enhanced high-quality image recovery. In ECCV, pages 155-171. Springer, 2020. 4 +[29] Jiaxu Wang, Junhao He, Ziyi Zhang, Mingyuan Sun, Jingkai Sun, and Renjing Xu. Evggs: A collaborative learning framework for event-based generalizable gaussian splatting. arXiv preprint arXiv:2405.14959, 2024. 3 +[30] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. Bad-nerf: Bundle adjusted deblur neural radiance fields. In CVPR, pages 4170–4179, 2023. 5 +[31] Tianyi Xiong, Jiayi Wu, Botao He, Cornelia Fermuller, Yiannis Aloimonos, Heng Huang, and Christopher Metzler. Event3dgs: Event-based 3d gaussian splatting for high-speed robot egomotion. In 8th Annual Conference on Robot Learning, 2024. 3 +[32] Fang Xu, Lei Yu, Bishan Wang, Wen Yang, Gui-Song Xia, Xu Jia, Zhendong Qiao, and Jianzhuang Liu. Motion deblurring with real events. In ICCV, pages 2583-2592, 2021. 5 +[33] Baoyue Zhang, Yajing Zheng, Shiyan Chen, Jiyuan Zhang, Kang Chen, Zhaofei Yu, and Tiejun Huang. Spikemm: Fleximagnification of high-speed micro-motions. arXiv preprint arXiv:2406.00383, 2024. 2 +[34] Jiyuan Zhang, Lulu Tang, Zhaofei Yu, Jiwen Lu, and Tiejun Huang. Spike transformer: Monocular depth estimation for spiking camera. In ECCV, pages 34-52. Springer, 2022. 2 +[35] Jiyuan Zhang, Shiyan Chen, Yajing Zheng, Zhaofei Yu, and Tiejun Huang. Unveiling the potential of spike streams for foreground occlusion removal from densely continuous views. arXiv preprint arXiv:2307.00821, 2023. 2 +[36] Jiyuan Zhang, Shanshan Jia, Zhaofei Yu, and Tiejun Huang. Learning temporal-ordered representation for spike streams based on discrete wavelet transforms. In AAAI, pages 137–147, 2023. 2, 4, 8 +[37] Jiyuan Zhang, Kang Chen, Shiyan Chen, Yajing Zheng, Tiejun Huang, and Zhaofei Yu. Spikegs: 3d gaussian splatt-ting from spike streams with high-speed camera motion. arXiv preprint arXiv:2407.10062, 2024. 2, 3, 6 +[38] Jiyuan Zhang, Shiyan Chen, Yajing Zheng, Zhaofei Yu, and Tiejun Huang. Spike-guided motion deblurring with unknown modal spatiotemporal alignment. In CVPR, pages 25047-25057, 2024. 2 +[39] Xiang Zhang and Lei Yu. Unifying motion deblurring and frame interpolation with events. In CVPR, pages 17765-17774, 2022. 5 +[40] Xiang Zhang, Lei Yu, Wen Yang, Jianzhuang Liu, and GuiSong Xia. Generalizing event-based motion deblurring in real-world scenarios. In ICCV, pages 10734-10744, 2023. 5 +[41] Jing Zhao, Ruiqin Xiong, Hangfan Liu, Jian Zhang, and Tiejun Huang. Spk2imgnet: Learning to reconstruct dynamic scene from continuous spike stream. In CVPR, pages 11996-12005, 2021. 2, 4, 6, 7, 8 +[42] Lingzhe Zhao, Peng Wang, and Peidong Liu. Bad-gaussians: Bundle adjusted deblur gaussian splatting. arXiv preprint arXiv:2403.11831, 2024. 5 + +[43] Rui Zhao, Ruiqin Xiong, Jing Zhao, Zhaofei Yu, Xiaopeng Fan, and Tiejun Huang. Learning optical flow from continuous spike streams. NeurIPS, 35:7905-7920, 2022. 2 +[44] Rui Zhao, Ruiqin Xiong, Jing Zhao, Jian Zhang, Xiaopeng Fan, Zhaofei Yu, and Tiejun Huang. Boosting spike camera image reconstruction from a perspective of dealing with spike fluctuations. In CVPR, pages 24955-24965, 2024. 2 +[45] Yajing Zheng, Lingxiao Zheng, Zhaofei Yu, Boxin Shi, Yonghong Tian, and Tiejun Huang. High-speed image reconstruction through short-term plasticity for spiking cameras. In CVPR, pages 6358-6367, 2021. 2 +[46] Yajing Zheng, Zhaofei Yu, Song Wang, and Tiejun Huang. Spike-based motion estimation for object tracking through bio-inspired unsupervised learning. IEEE TIP, 32:335-349, 2022. 2 +[47] Kangrui Zhou, Taihang Lei, Banglei Guan, and Qifeng Yu. Event-based depth estimation with dense occlusion. Optics Letters, 49(12):3376-3379, 2024. 3 +[48] Lin Zhu, Siwei Dong, Tiejun Huang, and Yonghong Tian. A retina-inspired sampling method for visual texture reconstruction. In ICME, pages 1432-1437. IEEE, 2019. 2, 4, 6, 7 +[49] Lin Zhu, Siwei Dong, Jianing Li, Tiejun Huang, and Yonghong Tian. Retina-like visual image reconstruction via spiking neural model. In CVPR, pages 1438-1446, 2020. 1, 2 +[50] Lin Zhu, Kangmin Jia, Yifan Zhao, Yunshan Qi, Lizhi Wang, and Hua Huang. Spikenerf: Learning neural radiance fields from continuous spike stream. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6285-6295, 2024. 2, 3, 6 \ No newline at end of file diff --git a/2025/USP-Gaussian_ Unifying Spike-based Image Reconstruction, Pose Correction and Gaussian Splatting/images.zip b/2025/USP-Gaussian_ Unifying Spike-based Image Reconstruction, Pose Correction and Gaussian Splatting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..91a169342ef5918bb2a46c8fb6ad1095c2d3f11e --- /dev/null +++ b/2025/USP-Gaussian_ Unifying Spike-based Image Reconstruction, Pose Correction and Gaussian Splatting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc574060d985a29998bc60b66f44c0c4f9ebe779201614803d7c8028cdf264b3 +size 909362 diff --git a/2025/USP-Gaussian_ Unifying Spike-based Image Reconstruction, Pose Correction and Gaussian Splatting/layout.json b/2025/USP-Gaussian_ Unifying Spike-based Image Reconstruction, Pose Correction and Gaussian Splatting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1e3935484fd4f9793724fa54adf9deab359944d1 --- /dev/null +++ b/2025/USP-Gaussian_ Unifying Spike-based Image Reconstruction, Pose Correction and Gaussian Splatting/layout.json @@ -0,0 +1,9780 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 102, + 543, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 543, + 140 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 543, + 140 + ], + "type": "text", + "content": "USP-Gaussian: Unifying Spike-based Image Reconstruction, Pose Correction and Gaussian Splatting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 159, + 548, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 159, + 548, + 219 + ], + "spans": [ + { + "bbox": [ + 66, + 159, + 548, + 219 + ], + "type": "text", + "content": "Kang Chen" + }, + { + "bbox": [ + 66, + 159, + 548, + 219 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 66, + 159, + 548, + 219 + ], + "type": "text", + "content": " Jiyuan Zhang" + }, + { + "bbox": [ + 66, + 159, + 548, + 219 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 66, + 159, + 548, + 219 + ], + "type": "text", + "content": " Zecheng Hao" + }, + { + "bbox": [ + 66, + 159, + 548, + 219 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 66, + 159, + 548, + 219 + ], + "type": "text", + "content": " Yajing Zheng" + }, + { + "bbox": [ + 66, + 159, + 548, + 219 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 66, + 159, + 548, + 219 + ], + "type": "text", + "content": " Tiejun Huang" + }, + { + "bbox": [ + 66, + 159, + 548, + 219 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 66, + 159, + 548, + 219 + ], + "type": "text", + "content": " Zhaofei Yu" + }, + { + "bbox": [ + 66, + 159, + 548, + 219 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 66, + 159, + 548, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 159, + 548, + 219 + ], + "type": "text", + "content": "School of Computer Science, Peking University \n" + }, + { + "bbox": [ + 66, + 159, + 548, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 66, + 159, + 548, + 219 + ], + "type": "text", + "content": "State Key Laboratory for Multimedia Information Processing, Peking University \n" + }, + { + "bbox": [ + 66, + 159, + 548, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 66, + 159, + 548, + 219 + ], + "type": "text", + "content": "Institute for Artificial Intelligence, Peking University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 220, + 529, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 220, + 529, + 232 + ], + "spans": [ + { + "bbox": [ + 77, + 220, + 529, + 232 + ], + "type": "text", + "content": "{mrchenkang, jyzhang}@stu.pku.edu.cn, {haozecheng,yj.zheng,tjhuang,yuzf12}@pku.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 240, + 196, + 385 + ], + "blocks": [ + { + "bbox": [ + 58, + 240, + 196, + 385 + ], + "lines": [ + { + "bbox": [ + 58, + 240, + 196, + 385 + ], + "spans": [ + { + "bbox": [ + 58, + 240, + 196, + 385 + ], + "type": "image", + "image_path": "333e4d02bd17b5e8661568d0d4bd749fc6ac67f7fe13fafa5e045a43de43365e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 395, + 555, + 440 + ], + "lines": [ + { + "bbox": [ + 55, + 395, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 555, + 440 + ], + "type": "text", + "content": "Figure 1. Left. Illustration of our USP-Gaussian framework, where the spike-based image Reconstruction Network (Recon-Net), camera poses, and 3DGS are collaboratively optimized signified by " + }, + { + "bbox": [ + 55, + 395, + 555, + 440 + ], + "type": "inline_equation", + "content": "\\mathbb{O}" + }, + { + "bbox": [ + 55, + 395, + 555, + 440 + ], + "type": "text", + "content": ". Mid. Visual ablation showcasing the performance of Recon-Net and 3DGS with and without (w/ & w/o) the joint optimization strategy, with the ablation table depicted in Tab. 4 and the input formulated in Eq. (6). Right. Training curve comparison for Recon-Net and 3DGS with and without joint optimization." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 202, + 241, + 392, + 386 + ], + "blocks": [ + { + "bbox": [ + 202, + 241, + 392, + 386 + ], + "lines": [ + { + "bbox": [ + 202, + 241, + 392, + 386 + ], + "spans": [ + { + "bbox": [ + 202, + 241, + 392, + 386 + ], + "type": "image", + "image_path": "77cffedaf199051973ce374e95a5a7d4ebca62ec01f8947bcad9c164c736a129.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 392, + 243, + 554, + 385 + ], + "blocks": [ + { + "bbox": [ + 392, + 243, + 554, + 385 + ], + "lines": [ + { + "bbox": [ + 392, + 243, + 554, + 385 + ], + "spans": [ + { + "bbox": [ + 392, + 243, + 554, + 385 + ], + "type": "image", + "image_path": "e22ac288bc8fe71cef5efdc783edaa7c29b0c2bce11325bc782d07e9b7a1d359.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 152, + 461, + 200, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 461, + 200, + 474 + ], + "spans": [ + { + "bbox": [ + 152, + 461, + 200, + 474 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 487, + 297, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 487, + 297, + 693 + ], + "spans": [ + { + "bbox": [ + 54, + 487, + 297, + 693 + ], + "type": "text", + "content": "Spike camera, as an innovative type of neuromorphic camera that captures scenes with 0-1 bit stream at " + }, + { + "bbox": [ + 54, + 487, + 297, + 693 + ], + "type": "inline_equation", + "content": "40\\mathrm{kHz}" + }, + { + "bbox": [ + 54, + 487, + 297, + 693 + ], + "type": "text", + "content": ", is increasingly being employed for the novel view synthesis task building on techniques such as Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS). Previous spike-based approaches typically follow a three-stage pipeline: I. Spike-to-image reconstruction based on established algorithms. II. Camera poses estimation. III. Novel view synthesis. However, the cascading framework suffers from substantial cumulative errors, i.e., the quality of the initially reconstructed images will impact pose estimation, ultimately limiting the fidelity of the 3D reconstruction. To address this limitation, we propose a synergistic optimization framework USP-Gaussian, which unifies spike-to-image reconstruction, pose correction, and gaussian splatting into an end-to-end pipeline. Leveraging the multi-view consistency afforded by 3DGS and the motion capture capability of the" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 451, + 556, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 556, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 556, + 559 + ], + "type": "text", + "content": "spike camera, our framework enables iterative optimization between the spike-to-image reconstruction network and 3DGS. Experiments on synthetic datasets demonstrate that our method surpasses previous approaches by effectively eliminating cascading errors. Moreover, in real-world scenarios, our method achieves robust 3D reconstruction benefiting from the integration of pose optimization. Our code, data, and trained models are available at https://github.com/chenkang455/USP-Gaussian." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 584, + 395, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 584, + 395, + 596 + ], + "spans": [ + { + "bbox": [ + 314, + 584, + 395, + 596 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "type": "text", + "content": "Spike cameras [12], inspired by the human retina [49], introduce an innovative paradigm for encoding high-speed dynamic scenes. In contrast to conventional imaging systems [22] that rely on fixed exposure intervals to integrate visual information, spike cameras continuously sense dynamics by recording the spike stream with an impressive firing rate of " + }, + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "40\\mathrm{kHz}" + }, + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "type": "text", + "content": ". Benefiting from the efficient bit representation of spikes, spike cameras can adeptly detect rapid scenes with significantly diminished bandwidth compared" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 151, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 151, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 151, + 713 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16609" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 249, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 249, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 249, + 83 + ], + "type": "text", + "content": "to the high-speed cameras such as the Phantom." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 86, + 296, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 86, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 86, + 296, + 193 + ], + "type": "text", + "content": "Capitalizing on the remarkable capability of the spike camera to capture intricate motion and texture details in high-speed environments, recent research [5, 36, 41, 44, 45, 48, 49] concentrate on establishing the mapping from bit-encoded spike to visually perceivable images. Subsequent investigations delve into the estimation of optical flow [43], facilitation of motion deblur [3, 8, 38] and applications such as high-speed object detection [46], motion magnification [33], depth estimation [34] and occlusion removal [35]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 194, + 296, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 296, + 326 + ], + "type": "text", + "content": "Neural Radiance Fields [21] and 3D Gaussian Splitting [13] have emerged as prominent novel view synthesis technologies benefiting from their designed differential training frameworks. Among them, 3DGS employs Gaussian primitives for explicit scene representation and the splatting algorithm for projection, thereby achieving faster rendering speed compared to the pixel-by-pixel approach implemented in NeRF. Building on NeRF and 3DGS, the high-speed and continuous capturing potential of the spike camera can be fully unleashed in 3D reconstruction tasks [11, 37, 50]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 327, + 296, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 435 + ], + "type": "text", + "content": "While the idea of applying the spike camera for 3D reconstruction is trivial and straightforward in a sequential manner, i.e., we can reconstruct images from the spike stream, estimate camera poses and subsequently apply NeRF/3DGS to obtain the 3D representation. Nevertheless, the cascading pipeline introduces significant cumulative errors when the quality of initially recovered images is poor, thereby degrading the accuracy of camera pose estimation and limiting the fidelity of the reconstructed scene." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 437, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 567 + ], + "type": "text", + "content": "For example, SpikeGS [37] initiates with a self-supervised spike-to-image reconstruction framework built on the Blind Spot Network (BSN) [15], intended to mitigate the performance degradation [6, 7] encountered by supervised learning methods on real-world scenes. With recovered high-quality images, SpikeGS designs a further constraint from the long spike input and integrates it into the vanilla 3DGS training pipeline. While SpikeGS have improved preliminary representation for 3DGS, it remains constrained within the cascading framework thus inheriting the aforementioned cumulative error problem." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "To address these challenges, we propose USP-Gaussian, Unifying Spike-based image reconstruction, Pose correction, and Gaussian Splatting within a one-stage optimization framework to mitigate error propagation inherent in the three-stage cascading pipeline. Specifically, USP-Gaussian comprises two synergistic branches as illustrated in Fig. 1, one dedicated to spike-to-image reconstruction via Recon-Net and the other to novel view synthesis implemented with 3DGS. To facilitate the training of Recon-Net, we design " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rec}}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " based on the constraint between the input spike and recovered outputs. Besides, we supervise the training of 3DGS based on " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{gs}}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": ", which obviates the require" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "type": "text", + "content": "ment for pre-recovered high-quality images. While each branch of USP-Gaussian operates independently, we further introduce the joint loss " + }, + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{joint}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "type": "text", + "content": " derived from the outputs of both two branches to facilitate unified optimization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 121, + 556, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 556, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 556, + 240 + ], + "type": "text", + "content": "Visual ablation comparisons of restored images and the training curve depicted in Fig. 1 demonstrate that the joint optimization of the two branches mutually enhances the recovered textures and the overall training process. We conduct quantitative and qualitative evaluations on the synthetic dataset and the real-world dataset. Experiments showcase that our method achieves superior restoration in both datasets, validating the effectiveness of our joint learning framework. In summary, the contributions of our USP-Gaussian are as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 241, + 553, + 350 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 313, + 241, + 553, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 553, + 277 + ], + "type": "text", + "content": "- We propose a unified framework for the joint optimization of Recon-Net, Pose, and 3DGS, mitigating error amplification inherent in cascading frameworks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 277, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 553, + 312 + ], + "type": "text", + "content": "- We demonstrate that spike-to-image and 3D reconstruction tasks can mutually facilitate and enhance the optimization of each other." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 313, + 553, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 553, + 350 + ], + "type": "text", + "content": "- We introduce a novel self-supervised training paradigm for the spike-based image reconstruction network with the multi-view consistency afforded by 3DGS." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 361, + 401, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 401, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 401, + 374 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 382, + 499, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 499, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 499, + 396 + ], + "type": "text", + "content": "2.1. Spike-based Image Reconstruction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 402, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 676 + ], + "type": "text", + "content": "While neuromorphic cameras generate high frame rate bitstream information, it is imperative to transform them into multi-bit grayscale images that are perceptible to human observers. To achieve this spike-to-image target, a plethora of methodologies [5-7, 36, 41, 44, 45, 49] have been proposed, which can be broadly classified into model-based, supervised, and self-supervised learning paradigms. Model-based methods, rooted in the sampling principle of the spike camera, aim to infer high-fidelity images drawing from the physical spike-image relationship [48], biological retinal models [49], and short-term synaptic plasticity [45]. Confronted with pervasive dark current noise that is challenging to model, learning-based methods are employed to directly deduce the sharp image from spike input with unknown firing patterns. Specifically, Spk2ImgNet [41] designed a hierarchical architecture engineered to capture the intricate temporal dynamics present within spike streams. SpikeCLIP [5] utilized the textual description of the captured scene as supervision for training the spike-to-image network. Concurrently, self-supervised learning methodologies based on the BSN [6, 7] are developed to counteract the performance deterioration in supervised learning methods owing to the domain gap between synthetic and real-world datasets." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "In this paper, we propose a novel self-supervised spike-based image reconstruction framework leveraging the multi-view constraints provided by 3D scenes." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16610" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "2.2. Novel View Synthesis based on the Neuromorphic Camera" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 102, + 296, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 102, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 55, + 102, + 296, + 162 + ], + "type": "text", + "content": "Recent research has extensively employed neuromorphic cameras, including event cameras [10, 16, 18, 27, 47], which encode intensity difference in dynamic scenarios with low-latency and spike cameras [12], for 3D scene reconstruction." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 162, + 296, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 162, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 162, + 296, + 316 + ], + "type": "text", + "content": "Event Camera. Klenk et al. [14] proposed the E-NeRF based on the event generation mechanism, which can recover sharp 3D scenes under high-speed camera motion. Rudnev et al. [23] further harnessed a color event camera to learn the color representation of scenes. [1, 20] focused on recovering dynamic scenes with rigid deformations from event streams. In addition to NeRF, EvGGS [29] introduces a generalizable Gaussian splatting framework for event cameras designed to reconstruct 3D scenes without retraining. EF-3DGS [17] develops a free-trajectory framework by aligning the RGB input, event streams, and pose. Event3DGS [31] presents an effective solution for rapid ego-motion robotics scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 317, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 296, + 449 + ], + "type": "text", + "content": "Spike Camera. Zhu et al. [50] design a differentiable SNN and integrate it into the NeRF training pipeline to simulate the generation of spike streams, where the loss function is formulated based on the rendered spike output and the input spike. Dai et al. [9] further introduce SpikeNVS by extending the spike cameras to RGB-Spike input. Additionally, Guo et al. [11] introduced a color 3D scene reconstruction framework by integrating Bayer-pattern spike streams into the 3DGS training pipeline. SpikeGS [37] initially employs the BSN to recover images from the noisy spike stream and subsequently integrates them into the 3DGS." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 449, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 498 + ], + "type": "text", + "content": "While the cascaded framework SpikeGS [37] introduces propagation errors, we propose a joint optimization framework that concurrently handles spike-based image reconstruction, pose estimation, and novel view synthesis." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 506, + 133, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 133, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 133, + 519 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 525, + 145, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 145, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 145, + 538 + ], + "type": "text", + "content": "3.1. Spike Camera" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 544, + 296, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 296, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 296, + 661 + ], + "type": "text", + "content": "The spike camera consists of the photodetector, voltage integrator, comparator, and synchronous readout as shown in Fig. 2. Within the designated time interval " + }, + { + "bbox": [ + 55, + 544, + 296, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = [t_s,t_e]" + }, + { + "bbox": [ + 55, + 544, + 296, + 661 + ], + "type": "text", + "content": ", the photodetector continuously captures the light intensity " + }, + { + "bbox": [ + 55, + 544, + 296, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{I}(t)" + }, + { + "bbox": [ + 55, + 544, + 296, + 661 + ], + "type": "text", + "content": " and converts it into the voltage signal " + }, + { + "bbox": [ + 55, + 544, + 296, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{V}(t) = \\mu \\mathbf{I}(t)" + }, + { + "bbox": [ + 55, + 544, + 296, + 661 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 544, + 296, + 661 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 544, + 296, + 661 + ], + "type": "text", + "content": " is the photoelectric conversion coefficient. Concurrently, the comparator assesses the integrator voltage " + }, + { + "bbox": [ + 55, + 544, + 296, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{A}(t)" + }, + { + "bbox": [ + 55, + 544, + 296, + 661 + ], + "type": "text", + "content": " against the predefined threshold " + }, + { + "bbox": [ + 55, + 544, + 296, + 661 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 55, + 544, + 296, + 661 + ], + "type": "text", + "content": ". When the threshold is met, the spike is emitted and the integrator voltage is reset, formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 118, + 661, + 295, + 687 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 661, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 118, + 661, + 295, + 687 + ], + "type": "interline_equation", + "content": "\\mathbf {A} (t) = \\int_ {t _ {s}} ^ {t} \\mathbf {I} (t) d t \\bmod C, \\tag {1}", + "image_path": "0e8c5dcca2036295f45ddc038b92cee4850b9be7f2c84e8c293e2529f1fd0109.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "C = \\Theta /\\mu" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " is a constant proportional to the spike firing threshold " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": ". While the integrator continuously ag-" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 555, + 229 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 555, + 229 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 555, + 229 + ], + "type": "image", + "image_path": "5208623a539375b59ed4bbb1ffea88a52bd7f97c8fed35547893dfd0382f4a5f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 237, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 313, + 237, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 555, + 281 + ], + "type": "text", + "content": "Figure 2. (a) The spike camera captures a railway moving 350 km/h. (b) The Recon-Net is designed to establish the mapping between the input spike stream and the sharp image. (c) The working mechanism of the spike camera." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 303, + 555, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 555, + 375 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 555, + 375 + ], + "type": "text", + "content": "gregates voltage, the readout circuit synchronously retrieves spike signals from the registers at up to " + }, + { + "bbox": [ + 313, + 303, + 555, + 375 + ], + "type": "inline_equation", + "content": "40\\mathrm{kHz}" + }, + { + "bbox": [ + 313, + 303, + 555, + 375 + ], + "type": "text", + "content": ". Spikes fired during the specified interval " + }, + { + "bbox": [ + 313, + 303, + 555, + 375 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 303, + 555, + 375 + ], + "type": "text", + "content": " are represented as a three-dimensional bit stream " + }, + { + "bbox": [ + 313, + 303, + 555, + 375 + ], + "type": "inline_equation", + "content": "S\\in \\{0,1\\}^{K\\times H\\times W}" + }, + { + "bbox": [ + 313, + 303, + 555, + 375 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 303, + 555, + 375 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 303, + 555, + 375 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 303, + 555, + 375 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 303, + 555, + 375 + ], + "type": "text", + "content": " indicate the image height and width, with " + }, + { + "bbox": [ + 313, + 303, + 555, + 375 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 303, + 555, + 375 + ], + "type": "text", + "content": " denoting the spike sequence length." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 383, + 443, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 443, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 443, + 396 + ], + "type": "text", + "content": "3.2. 3D Gaussian Splatting" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 401, + 554, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 554, + 439 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 554, + 439 + ], + "type": "text", + "content": "3D Gaussian Splating leverages a set of three-dimensional Gaussian primitives, symbolized as " + }, + { + "bbox": [ + 313, + 401, + 554, + 439 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{G}\\}_{n=1}^{N_{\\mathrm{gs}}}" + }, + { + "bbox": [ + 313, + 401, + 554, + 439 + ], + "type": "text", + "content": ", to effectively represent the spatial characteristics of the captured scene." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 440, + 554, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 440, + 554, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 554, + 500 + ], + "type": "text", + "content": "Each Gaussian primitive is centered at a point " + }, + { + "bbox": [ + 313, + 440, + 554, + 500 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_n" + }, + { + "bbox": [ + 313, + 440, + 554, + 500 + ], + "type": "text", + "content": ", with its spread and orientation influenced by the covariance matrix " + }, + { + "bbox": [ + 313, + 440, + 554, + 500 + ], + "type": "inline_equation", + "content": "\\Sigma_n = RSS^T R^T" + }, + { + "bbox": [ + 313, + 440, + 554, + 500 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 440, + 554, + 500 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 440, + 554, + 500 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 440, + 554, + 500 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 440, + 554, + 500 + ], + "type": "text", + "content": " denotes the rotation and the scale matrix, respectively. The impact of the Gaussian on any point " + }, + { + "bbox": [ + 313, + 440, + 554, + 500 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 313, + 440, + 554, + 500 + ], + "type": "text", + "content": " in the scene is quantified by:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 336, + 508, + 555, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 508, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 336, + 508, + 555, + 536 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {n} (\\mathbf {v}) = \\exp \\left(- \\frac {1}{2} (\\mathbf {v} - \\mathbf {p} _ {n}) ^ {T} \\boldsymbol {\\Sigma} _ {n} ^ {- 1} (\\mathbf {v} - \\mathbf {p} _ {n})\\right). (2)", + "image_path": "769beb40e6f56f1006e3d9be1dad0aa85674753dd3df38bb05df734ba782b2bd.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 544, + 554, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 580 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 580 + ], + "type": "text", + "content": "When it comes to rendering for viewing transformation " + }, + { + "bbox": [ + 313, + 544, + 554, + 580 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 544, + 554, + 580 + ], + "type": "text", + "content": ", these 3D Gaussian primitives are mapped onto 2D primitives " + }, + { + "bbox": [ + 313, + 544, + 554, + 580 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_n^{2D}" + }, + { + "bbox": [ + 313, + 544, + 554, + 580 + ], + "type": "text", + "content": ", with the covariance matrix " + }, + { + "bbox": [ + 313, + 544, + 554, + 580 + ], + "type": "inline_equation", + "content": "\\Sigma_n^\\prime" + }, + { + "bbox": [ + 313, + 544, + 554, + 580 + ], + "type": "text", + "content": " reformulated as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 388, + 590, + 553, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 590, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 388, + 590, + 553, + 605 + ], + "type": "interline_equation", + "content": "\\Sigma_ {n} ^ {\\prime} = J W \\Sigma_ {n} W ^ {T} J ^ {T}, \\tag {3}", + "image_path": "95b48462dd229383071ad25cb426e27c7b5c9fc628a83011c6f62d9553ffaa16.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 611, + 554, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 554, + 635 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 611, + 554, + 635 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 313, + 611, + 554, + 635 + ], + "type": "text", + "content": " represents the Jacobian matrix derived from the affine approximation of the projective transformation." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 635, + 554, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 554, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 554, + 671 + ], + "type": "text", + "content": "Further, each pixel on the rendered plane is calculated by rasterizing these 2D Gaussians, sorted by depth, according to the following formulation:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 354, + 681, + 555, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 681, + 555, + 716 + ], + "spans": [ + { + "bbox": [ + 354, + 681, + 555, + 716 + ], + "type": "interline_equation", + "content": "C (r) = \\sum_ {k = 1} ^ {N _ {\\mathrm {g s}}} T _ {i} \\alpha_ {i} \\mathbf {c} _ {i}, T _ {i} = \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {j}\\right), \\tag {4}", + "image_path": "df1b4609cb4fa30ba4bb1e14a589676438f601247dc0f39d9c1226f00a64b316.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16611" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 555, + 293 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 555, + 293 + ], + "type": "image", + "image_path": "1f235a959be2d89218507a7cb10908cc685b0584d3ab39210d469737fcc3e0a2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 302, + 555, + 381 + ], + "lines": [ + { + "bbox": [ + 54, + 302, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 54, + 302, + 555, + 381 + ], + "type": "text", + "content": "Figure 3. The working pipeline of our USP-Gaussian. For each camera viewpoint, we initially derive a pose sequence at equal time intervals based on the camera model described in Sec. 4.3. Leveraging the 3DGS rendering pipeline detailed in Sec. 4.2, we subsequently generate a corresponding sharp sequence for this viewpoint. Concurrently, Recon-Net is employed to reconstruct the same sequence from the captured spike stream as outlined in Sec. 4.1. While Recon-Net and 3DGS are trained based on the discrepancy between the re-synthesized blurry image with the input long-exposure image, alignment of their outputs for the same scene is further enforced through joint optimization loss as described in Sec. 4.4. With formulated loss functions, our USP-Gaussian enables the collaborative optimization of Gaussian primitives, camera poses, and the parameters of Recon-Net as formulated in Eq. (5)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 401, + 297, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 297, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 297, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 401, + 297, + 437 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 401, + 297, + 437 + ], + "type": "text", + "content": " represents the indices set reordered by depth, and " + }, + { + "bbox": [ + 55, + 401, + 297, + 437 + ], + "type": "inline_equation", + "content": "D_{k}" + }, + { + "bbox": [ + 55, + 401, + 297, + 437 + ], + "type": "text", + "content": " denotes the index of the Gaussian primitive that is the " + }, + { + "bbox": [ + 55, + 401, + 297, + 437 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 401, + 297, + 437 + ], + "type": "text", + "content": "-th closest to the rendering point." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 449, + 117, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 117, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 117, + 462 + ], + "type": "text", + "content": "4. Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 469, + 296, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 296, + 530 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 296, + 530 + ], + "type": "text", + "content": "Target Clarification. Given the captured spike stream set " + }, + { + "bbox": [ + 55, + 469, + 296, + 530 + ], + "type": "inline_equation", + "content": "\\{S_v\\}_{v=1}^V" + }, + { + "bbox": [ + 55, + 469, + 296, + 530 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 469, + 296, + 530 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 469, + 296, + 530 + ], + "type": "text", + "content": " different viewpoints with initial inaccurate poses " + }, + { + "bbox": [ + 55, + 469, + 296, + 530 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{P}_v\\}_{v=1}^V" + }, + { + "bbox": [ + 55, + 469, + 296, + 530 + ], + "type": "text", + "content": ", the USP-Gaussian aims to optimize the parameters of Recon-Net and Gaussian primitives while outputting refined poses as shown in Fig. 1, formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 99, + 540, + 295, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 540, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 99, + 540, + 295, + 556 + ], + "type": "interline_equation", + "content": "(\\mathbb {S}, \\mathbb {P}, \\mathbb {G}, \\theta) \\xrightarrow {\\text {U S P - G a u s s i a n}} (\\mathbb {P} ^ {*}, \\mathbb {G} ^ {*}, \\theta^ {*}), \\tag {5}", + "image_path": "bea6cc1d517df3336ac8475df671c5983a18a780906d3fd59a8091624a4a9977.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 565, + 296, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 296, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 296, + 615 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 565, + 296, + 615 + ], + "type": "inline_equation", + "content": "\\mathbb{S}" + }, + { + "bbox": [ + 55, + 565, + 296, + 615 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 565, + 296, + 615 + ], + "type": "inline_equation", + "content": "\\mathbb{P}" + }, + { + "bbox": [ + 55, + 565, + 296, + 615 + ], + "type": "text", + "content": " represent the sets of spike streams and poses from " + }, + { + "bbox": [ + 55, + 565, + 296, + 615 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 565, + 296, + 615 + ], + "type": "text", + "content": " viewpoints, " + }, + { + "bbox": [ + 55, + 565, + 296, + 615 + ], + "type": "inline_equation", + "content": "\\mathbb{G}" + }, + { + "bbox": [ + 55, + 565, + 296, + 615 + ], + "type": "text", + "content": " denotes the ensemble of Gaussian primitives, and " + }, + { + "bbox": [ + 55, + 565, + 296, + 615 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 565, + 296, + 615 + ], + "type": "text", + "content": " encapsulates the Recon-Net parameters. The notation " + }, + { + "bbox": [ + 55, + 565, + 296, + 615 + ], + "type": "inline_equation", + "content": "(^{*})" + }, + { + "bbox": [ + 55, + 565, + 296, + 615 + ], + "type": "text", + "content": " indicates the optimized parameters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 623, + 285, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 285, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 285, + 635 + ], + "type": "text", + "content": "4.1. Spike-based Image Reconstruction Network" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "The Recon-Net is designed to facilitate the mapping " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot)" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " from the spike input to the sharp image as shown in Fig. 2(b), formulated as " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{I}(f) = \\mathcal{F}(S_{\\mathcal{T}};\\theta)" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " signifying the central moment of the interval " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": ". Prior research [6, 36, 41] employs a fixed and brief window (41 frames in their configuration) of the spike stream as the network in" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 401, + 555, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 555, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 555, + 424 + ], + "type": "text", + "content": "put. However, it is insufficient for scenarios with low spike firing rates due to the limited information embedded within." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 424, + 554, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 554, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 554, + 484 + ], + "type": "text", + "content": "While extending the window size allows the network to encompass more features, it complicates the feature extraction process and demands large computational resource. To address it, we configure the input of the Recon-Net with a complementary long-short spike stream as shown in Fig. 3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 484, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 554, + 604 + ], + "type": "text", + "content": "Given a long spike stream " + }, + { + "bbox": [ + 313, + 484, + 554, + 604 + ], + "type": "inline_equation", + "content": "S_{\\mathcal{T}}" + }, + { + "bbox": [ + 313, + 484, + 554, + 604 + ], + "type": "text", + "content": ", we extract a short spike stream " + }, + { + "bbox": [ + 313, + 484, + 554, + 604 + ], + "type": "inline_equation", + "content": "S_{\\mathcal{T}'}^t \\subset S_{\\mathcal{T}}" + }, + { + "bbox": [ + 313, + 484, + 554, + 604 + ], + "type": "text", + "content": " centered around time " + }, + { + "bbox": [ + 313, + 484, + 554, + 604 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 484, + 554, + 604 + ], + "type": "text", + "content": ", with the network mapping reformulated as " + }, + { + "bbox": [ + 313, + 484, + 554, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{I}(t) = \\mathcal{F}(S_{\\mathcal{T}}, S_{\\mathcal{T}'}^t, t; \\theta)" + }, + { + "bbox": [ + 313, + 484, + 554, + 604 + ], + "type": "text", + "content": ", where time " + }, + { + "bbox": [ + 313, + 484, + 554, + 604 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 484, + 554, + 604 + ], + "type": "text", + "content": " is incorporated to guide the alignment of the short and long spike. Following the event stream pre-processing in [2, 26, 28], we apply voxelization to the long spike stream " + }, + { + "bbox": [ + 313, + 484, + 554, + 604 + ], + "type": "inline_equation", + "content": "S_{\\mathcal{T}}" + }, + { + "bbox": [ + 313, + 484, + 554, + 604 + ], + "type": "text", + "content": ", reducing its dimensions from " + }, + { + "bbox": [ + 313, + 484, + 554, + 604 + ], + "type": "inline_equation", + "content": "K \\times H \\times W" + }, + { + "bbox": [ + 313, + 484, + 554, + 604 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 484, + 554, + 604 + ], + "type": "inline_equation", + "content": "\\frac{K}{4} \\times H \\times W" + }, + { + "bbox": [ + 313, + 484, + 554, + 604 + ], + "type": "text", + "content": " accumulating every 4 frames. The network architecture consists of basic convolutional blocks as illustrated in Fig. 3 and detailed in the supplementary materials." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 604, + 554, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 554, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 554, + 652 + ], + "type": "text", + "content": "According to the spike camera working mechanism described in Eq. (1), we can reconstruct a long-exposure image " + }, + { + "bbox": [ + 313, + 604, + 554, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{E}(\\mathcal{T})" + }, + { + "bbox": [ + 313, + 604, + 554, + 652 + ], + "type": "text", + "content": " from the spike input " + }, + { + "bbox": [ + 313, + 604, + 554, + 652 + ], + "type": "inline_equation", + "content": "S_{\\mathcal{T}}" + }, + { + "bbox": [ + 313, + 604, + 554, + 652 + ], + "type": "text", + "content": " with the TFP algorithm [48], mathematically formulated as below:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 367, + 658, + 555, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 658, + 555, + 685 + ], + "spans": [ + { + "bbox": [ + 367, + 658, + 555, + 685 + ], + "type": "interline_equation", + "content": "\\mathbf {E} (\\mathcal {T}) = \\frac {1}{T} \\int_ {\\mathcal {T}} \\mathbf {I} (t) d t = \\frac {C \\cdot N}{T}, \\qquad (6)", + "image_path": "b74a5a69f67ab3619f16fb0cc6ecb15fd38ec054c526e14e8da6edbc76cdb863.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " denotes the duration of the interval " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " is the cumulative number of spikes during this period." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16612" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "Motivated by the textural similarity between the blurry image and the long-exposure image, we can employ the motion-reblur loss widely employed in self-supervised motion deblurring frameworks [3, 32, 39, 40] to guide the training of the Recon-Net. We initially reconstruct " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " frames uniformly across the interval to form the image sequence " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_{\\mathrm{rec}}(t_m)\\}_{m = 1}^M" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " and further average them to synthesize the reblur image. The loss can be formulated by comparing the reblur image to the long-exposure image, expressed as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 82, + 188, + 296, + 222 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 188, + 296, + 222 + ], + "spans": [ + { + "bbox": [ + 82, + 188, + 296, + 222 + ], + "type": "interline_equation", + "content": "\\mathbb {L} _ {\\mathrm {r e c}} (M, \\mathcal {T}) = \\mathcal {L} \\left(\\frac {1}{M} \\sum_ {m = 1} ^ {M} \\mathbf {I} _ {\\mathrm {r e c}} \\left(t _ {m}\\right), \\mathbf {E} (\\mathcal {T})\\right), \\tag {7}", + "image_path": "2fc83730d782705d62968066b60fef99d024c94793615f43c2a570cb25884962.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 227, + 296, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 227, + 296, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 227, + 296, + 264 + ], + "type": "text", + "content": "where the loss " + }, + { + "bbox": [ + 55, + 227, + 296, + 264 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 227, + 296, + 264 + ], + "type": "text", + "content": " [13] combines " + }, + { + "bbox": [ + 55, + 227, + 296, + 264 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 227, + 296, + 264 + ], + "type": "text", + "content": " loss with a D-SSIM term, defined as " + }, + { + "bbox": [ + 55, + 227, + 296, + 264 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = (1 - \\lambda)\\mathcal{L}_1 + \\lambda \\mathcal{L}_{\\mathrm{D - SSIM}}" + }, + { + "bbox": [ + 55, + 227, + 296, + 264 + ], + "type": "text", + "content": ", effectively integrating both pixel accuracy and structural similarity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 264, + 296, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 296, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 296, + 371 + ], + "type": "text", + "content": "Nevertheless, the input spike stream " + }, + { + "bbox": [ + 55, + 264, + 296, + 371 + ], + "type": "inline_equation", + "content": "S_{\\mathcal{T}}" + }, + { + "bbox": [ + 55, + 264, + 296, + 371 + ], + "type": "text", + "content": " sufficiently captures the information to estimate " + }, + { + "bbox": [ + 55, + 264, + 296, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{E}(\\mathcal{T})" + }, + { + "bbox": [ + 55, + 264, + 296, + 371 + ], + "type": "text", + "content": " as detailed in Eq. (6), which will lead the Recon-Net to learn a trivial identity mapping from the spike to the long exposure image. To mitigate it, we propose the multi-reblur loss to regularize the reconstructed sequence. Specifically, we extract " + }, + { + "bbox": [ + 55, + 264, + 296, + 371 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 264, + 296, + 371 + ], + "type": "text", + "content": " sub-intervals from the long exposure " + }, + { + "bbox": [ + 55, + 264, + 296, + 371 + ], + "type": "inline_equation", + "content": "T_{n} \\subset \\mathcal{T}" + }, + { + "bbox": [ + 55, + 264, + 296, + 371 + ], + "type": "text", + "content": " with the same central moment and apply the reblur loss to each sub-interval, formulating the multi-reblur loss as below:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 379, + 296, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 379, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 116, + 379, + 296, + 411 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r e c}} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\mathbb {L} _ {\\mathrm {r e c}} \\left(M _ {n}, \\mathcal {T} _ {n}\\right), \\tag {8}", + "image_path": "fcaabc5dbcc19d9feb2afe5b145b17a61a731c721815c1253f31eca46c14ca92.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 418, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 443 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 418, + 296, + 443 + ], + "type": "inline_equation", + "content": "M_{n}" + }, + { + "bbox": [ + 55, + 418, + 296, + 443 + ], + "type": "text", + "content": " represents the number of reconstructed frames within the " + }, + { + "bbox": [ + 55, + 418, + 296, + 443 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 418, + 296, + 443 + ], + "type": "text", + "content": "-th sub-interval " + }, + { + "bbox": [ + 55, + 418, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_n" + }, + { + "bbox": [ + 55, + 418, + 296, + 443 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 449, + 227, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 227, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 227, + 462 + ], + "type": "text", + "content": "4.2. Spike-based Gaussian Splatting" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 468, + 296, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 539 + ], + "type": "text", + "content": "Suppose that retrieving the pose for each timestamp within the interval is straightforward. Following Sec. 4.1, we uniformly obtain " + }, + { + "bbox": [ + 55, + 468, + 296, + 539 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 468, + 296, + 539 + ], + "type": "text", + "content": " poses from the interval " + }, + { + "bbox": [ + 55, + 468, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 468, + 296, + 539 + ], + "type": "text", + "content": ", yielding a pose sequence " + }, + { + "bbox": [ + 55, + 468, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{T}(t_m)\\}_{m=1}^M" + }, + { + "bbox": [ + 55, + 468, + 296, + 539 + ], + "type": "text", + "content": " with the corresponding image sequence " + }, + { + "bbox": [ + 55, + 468, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_{\\mathrm{gs}}(t_m)\\}_{m=1}^M" + }, + { + "bbox": [ + 55, + 468, + 296, + 539 + ], + "type": "text", + "content": " rendered via the 3DGS rasterization pipeline as shown in Fig. 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 540, + 296, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 296, + 624 + ], + "type": "text", + "content": "While reconstructing a sequence of images from the spike stream by shortening the exposure interval in Eq. (6) for the training of 3DGS sounds feasible, these short-exposure images are noisy and lack texture details [3]. To this end, we utilize the long-exposure image " + }, + { + "bbox": [ + 55, + 540, + 296, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{E}(\\mathcal{T})" + }, + { + "bbox": [ + 55, + 540, + 296, + 624 + ], + "type": "text", + "content": " as the supervision and design the motion-reblur loss for 3DGS similar to Eq. (7), expressed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 100, + 631, + 296, + 665 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 631, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 100, + 631, + 296, + 665 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {g s}} = \\mathcal {L} \\left(\\frac {1}{M} \\sum_ {m = 1} ^ {M} \\mathbf {I} _ {\\mathrm {g s}} \\left(t _ {m}\\right), \\mathbf {E} (\\mathcal {T})\\right). \\tag {9}", + "image_path": "806ea63f49b407cd2b0cbee36fba06be33724960a3e5b82f77a8642e02949aab.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 671, + 246, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 246, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 246, + 685 + ], + "type": "text", + "content": "4.3. Spike Camera Trajectory Modeling" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Building on the learnable pose representation during camera exposure in previous research [30, 42], we employ a" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "type": "text", + "content": "similar approach to optimize the camera pose throughout the interval " + }, + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "type": "text", + "content": " as shown in Fig. 3. For each spike stream " + }, + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "type": "inline_equation", + "content": "S_{\\mathcal{T}}" + }, + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "type": "text", + "content": ", we optimize both the start pose " + }, + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{start}}" + }, + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "type": "text", + "content": " and end pose " + }, + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{end}}" + }, + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "type": "text", + "content": " over the interval, while intermediate poses are represented through interpolation between these two." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 132, + 554, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 132, + 554, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 132, + 554, + 167 + ], + "type": "text", + "content": "For the pose " + }, + { + "bbox": [ + 313, + 132, + 554, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{T}(t_m)" + }, + { + "bbox": [ + 313, + 132, + 554, + 167 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 313, + 132, + 554, + 167 + ], + "type": "inline_equation", + "content": "t_m" + }, + { + "bbox": [ + 313, + 132, + 554, + 167 + ], + "type": "text", + "content": ", we utilize linear interpolation in the Lie algebra of " + }, + { + "bbox": [ + 313, + 132, + 554, + 167 + ], + "type": "inline_equation", + "content": "\\mathrm{SE}(3)" + }, + { + "bbox": [ + 313, + 132, + 554, + 167 + ], + "type": "text", + "content": " to estimate the pose at time " + }, + { + "bbox": [ + 313, + 132, + 554, + 167 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 132, + 554, + 167 + ], + "type": "text", + "content": ", formulated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 171, + 554, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 171, + 554, + 198 + ], + "spans": [ + { + "bbox": [ + 326, + 171, + 554, + 198 + ], + "type": "interline_equation", + "content": "\\mathbf {T} \\left(t _ {m}\\right) = \\mathbf {T} _ {\\text {s t a r t}} \\cdot \\exp \\left(\\frac {t _ {m}}{T} \\cdot \\log \\left(\\mathbf {T} _ {\\text {s t a r t}} ^ {- 1} \\cdot \\mathbf {T} _ {\\text {e n d}}\\right)\\right). \\tag {10}", + "image_path": "31ae2d1a4dc0c2be9c6e4ee3d0f919edab0732176c29460b2a32fab907a01a03.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 201, + 554, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 554, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 554, + 261 + ], + "type": "text", + "content": "During the training, the pose set " + }, + { + "bbox": [ + 313, + 201, + 554, + 261 + ], + "type": "inline_equation", + "content": "\\mathbb{P}" + }, + { + "bbox": [ + 313, + 201, + 554, + 261 + ], + "type": "text", + "content": " consisting of the start and end poses from " + }, + { + "bbox": [ + 313, + 201, + 554, + 261 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 201, + 554, + 261 + ], + "type": "text", + "content": " viewpoints is jointly optimized with the Recon-Net and Gaussian primitives as in Eq. (5). For further details on the camera trajectory modelling, please refer to BAD-NeRF [30] and BAD-Gaussian [42]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 266, + 425, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 425, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 425, + 278 + ], + "type": "text", + "content": "4.4. Joint Optimization" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 285, + 554, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 554, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 554, + 332 + ], + "type": "text", + "content": "Given the spike stream " + }, + { + "bbox": [ + 313, + 285, + 554, + 332 + ], + "type": "inline_equation", + "content": "S_{\\mathcal{T}}" + }, + { + "bbox": [ + 313, + 285, + 554, + 332 + ], + "type": "text", + "content": " along with the start pose " + }, + { + "bbox": [ + 313, + 285, + 554, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{start}}" + }, + { + "bbox": [ + 313, + 285, + 554, + 332 + ], + "type": "text", + "content": " and the end pose " + }, + { + "bbox": [ + 313, + 285, + 554, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{end}}" + }, + { + "bbox": [ + 313, + 285, + 554, + 332 + ], + "type": "text", + "content": " to be optimized, we perform evenly spaced sampling interpolation as in Eq. (10) during the interval, obtaining the pose sequence " + }, + { + "bbox": [ + 313, + 285, + 554, + 332 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{T}(t_m)\\}_{m=1}^M" + }, + { + "bbox": [ + 313, + 285, + 554, + 332 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 333, + 554, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 554, + 415 + ], + "type": "text", + "content": "For each pose in the sequence, we select a short spike stream centered around the timestamp " + }, + { + "bbox": [ + 313, + 333, + 554, + 415 + ], + "type": "inline_equation", + "content": "t_m" + }, + { + "bbox": [ + 313, + 333, + 554, + 415 + ], + "type": "text", + "content": ", obtaining the image sequence " + }, + { + "bbox": [ + 313, + 333, + 554, + 415 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_{\\mathrm{rec}}(t_m)\\}_{m=1}^M" + }, + { + "bbox": [ + 313, + 333, + 554, + 415 + ], + "type": "text", + "content": " based on the Recon-Net in Sec. 4.1 and the image sequence " + }, + { + "bbox": [ + 313, + 333, + 554, + 415 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_{\\mathrm{gs}}(t_m)\\}_{m=1}^M" + }, + { + "bbox": [ + 313, + 333, + 554, + 415 + ], + "type": "text", + "content": " rendered from the 3DGS in Sec. 4.2. While both sequences represent the same scene, the joint optimization loss " + }, + { + "bbox": [ + 313, + 333, + 554, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{joint}}" + }, + { + "bbox": [ + 313, + 333, + 554, + 415 + ], + "type": "text", + "content": " can be formulated as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 352, + 418, + 554, + 451 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 418, + 554, + 451 + ], + "spans": [ + { + "bbox": [ + 352, + 418, + 554, + 451 + ], + "type": "interline_equation", + "content": "\\mathbb {L} _ {\\text {j o i n t}} = \\frac {1}{M} \\sum_ {m = 1} ^ {M} \\left| \\left| \\mathbf {I} _ {\\mathrm {g s}} \\left(t _ {m}\\right) - \\mathbf {I} _ {\\mathrm {r e c}} \\left(t _ {m}\\right) \\right| \\right| _ {2} ^ {2}. \\tag {11}", + "image_path": "1e12afb4e6c8f435052fd70d38577360ad86811214282cf6af58c167237ce0c9.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 454, + 554, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 554, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 554, + 550 + ], + "type": "text", + "content": "However, we found that the joint optimization framework tends to be unstable due to the order difference between the 3DGS and Recon-Net output sequences. Specifically, since the temporal order does not affect the loss in Eq. (9), the optimized pose sequence may reverse relative to the real camera trajectory. To address this, we apply a simple but effective flip-and-minimum operation to align both sequences, with the loss function reformulated as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 331, + 554, + 553, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 554, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 331, + 554, + 553, + 569 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {j o i n t}} = \\min \\left(\\mathbb {L} _ {\\text {j o i n t}}, \\mathbb {L} _ {\\text {j o i n t}} ^ {\\mathrm {r}}\\right), \\tag {12}", + "image_path": "02ab75f24ea04372039d2cb57e8b9813f2b999379971b77fd1fba0c49253ebe6.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 331, + 571, + 553, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 571, + 553, + 604 + ], + "spans": [ + { + "bbox": [ + 331, + 571, + 553, + 604 + ], + "type": "interline_equation", + "content": "\\mathbb {L} _ {\\text {j o i n t}} ^ {\\mathrm {r}} = \\frac {1}{M} \\sum_ {m = 1} ^ {M} \\left| \\left| \\mathbf {I} _ {\\mathrm {g s}} \\left(t _ {M - m + 1}\\right) - \\mathbf {I} _ {\\mathrm {r e c}} \\left(t _ {m}\\right) \\right| \\right| _ {2} ^ {2}, \\tag {13}", + "image_path": "61197594e724ebabf7a7b0b673fb25668c806e7ed51b48f091d1fd2a5234c4fc.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 607, + 554, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 607, + 554, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 607, + 554, + 631 + ], + "type": "text", + "content": "where the difference between the " + }, + { + "bbox": [ + 313, + 607, + 554, + 631 + ], + "type": "inline_equation", + "content": "\\mathbb{L}_{\\mathrm{joint}}^{\\mathrm{r}}" + }, + { + "bbox": [ + 313, + 607, + 554, + 631 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 313, + 607, + 554, + 631 + ], + "type": "inline_equation", + "content": "\\mathbb{L}_{\\mathrm{joint}}" + }, + { + "bbox": [ + 313, + 607, + 554, + 631 + ], + "type": "text", + "content": " lies in that the rendered sequence via 3DGS is reversed." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 632, + 554, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 632, + 554, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 554, + 654 + ], + "type": "text", + "content": "To sum up, the loss function for our USP-Gaussian framework is presented as follows:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 383, + 660, + 553, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 660, + 553, + 673 + ], + "spans": [ + { + "bbox": [ + 383, + 660, + 553, + 673 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} + \\mathcal {L} _ {\\text {g s}} + \\mathcal {L} _ {\\text {j o i n t}}. \\tag {14}", + "image_path": "742ca334cb28a052bd4d5605e1cd480963371df3001995976e5baa9acefe0f87.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rec}}" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": " optimizes Recon-Net parameters, " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{gs}}" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": " refines poses and Gaussian primitives, and " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{joint}}" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": " jointly optimizes all components, with gradient flow shown in Fig. 3." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16613" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 555, + 404 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 555, + 404 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 555, + 404 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 555, + 404 + ], + "type": "image", + "image_path": "37a8e8279c2392d64e562911ce0358a313ccd3b75da7618652ce045db769c956.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 411, + 555, + 436 + ], + "lines": [ + { + "bbox": [ + 55, + 411, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 555, + 436 + ], + "type": "text", + "content": "Figure 4. 3D reconstruction visual comparison of our USP-Gaussian compared with previous methods on the synthetic dataset, where the input is the long-exposure image defined in Eq. (6)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 467, + 553, + 550 + ], + "blocks": [ + { + "bbox": [ + 59, + 445, + 550, + 460 + ], + "lines": [ + { + "bbox": [ + 59, + 445, + 550, + 460 + ], + "spans": [ + { + "bbox": [ + 59, + 445, + 550, + 460 + ], + "type": "text", + "content": "Table 1. 3D reconstruction quantitative comparison on the synthetic dataset. Color shading indicates the best and second-best result." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 467, + 553, + 550 + ], + "lines": [ + { + "bbox": [ + 58, + 467, + 553, + 550 + ], + "spans": [ + { + "bbox": [ + 58, + 467, + 553, + 550 + ], + "type": "table", + "html": "
MethodsWineTanabataFactoryOutdoorpoolAvg
PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
TFP-3DGS [48]24.1380.7960.23924.6260.7780.27427.8450.8520.17928.0240.7860.29426.1580.8030.239
TFI-3DGS [48]23.8460.7330.25121.1680.5500.52925.0190.7340.27824.7550.6370.44623.6970.6630.376
Spk2Img-3DGS [41]21.1500.7890.17820.4510.7580.21425.1150.8060.14323.3620.6470.41722.5200.7500.238
SpikeNeRF [50]25.8880.7980.20025.9650.7250.29527.5310.7800.19125.6420.6570.39626.2570.7400.270
SpikeGS [37]25.6980.8350.24726.1740.8260.24328.6270.8640.21428.2870.7970.27327.1960.8320.244
Ours26.4760.8620.17626.3000.8270.21328.6940.8610.20730.1420.8200.27027.9030.8430.217
", + "image_path": "415c77045fb304ee71199e9d4d7ba3567e54f3552d7bbac5fd57fb718befc1f9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 567, + 132, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 132, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 132, + 581 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 587, + 118, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 118, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 118, + 598 + ], + "type": "text", + "content": "5.1. Datasets" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 701 + ], + "type": "text", + "content": "Synthetic Dataset. We construct the synthetic dataset based on Deblur-NeRF [19] scenes for quantitative evaluation. We start with rendering 13 sharp images for each viewpoint, with further interpolation algorithm XVFI [25] to insert 7 images between each frame pair. We apply the spike simulator on the interpolated sequence to emulate the real-world spike as in [3], thus generating 97 spike frames for each viewpoint." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "Real-world Dataset. To construct the real-world dataset," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 569, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 555, + 653 + ], + "type": "text", + "content": "we shake the spike camera extremely fast while maintaining a fixed forward-facing orientation and extract 97 spike frames per view, consistent with the configuration in the synthetic dataset. For pose estimation, we convert the raw spike stream in each view to the short-exposure image as defined in Eq. (6) and subsequently employ COLMAP [24] to estimate the pose sequence." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 659, + 440, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 440, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 440, + 672 + ], + "type": "text", + "content": "5.2. Experimental Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "We compare our USP-Gaussian with two-stage cascading pipelines and approaches specially designed for the spike camera, i.e., SpikeNeRF [50] and SpikeGS [37]. For the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16614" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 555, + 242 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 555, + 242 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 555, + 242 + ], + "type": "image", + "image_path": "b5401e5ece889c1735c530f2b253351cb8c6da0c9f3640325aaf50f09d67310f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 250, + 536, + 262 + ], + "lines": [ + { + "bbox": [ + 72, + 250, + 536, + 262 + ], + "spans": [ + { + "bbox": [ + 72, + 250, + 536, + 262 + ], + "type": "text", + "content": "Figure 5. 3D reconstruction visual comparison of our USP-Gaussian compared with previous methods on the real-world dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 312, + 296, + 370 + ], + "blocks": [ + { + "bbox": [ + 55, + 281, + 295, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 281, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 295, + 304 + ], + "type": "text", + "content": "Table 2. Comparison of our method with SpikeGS on the wine dataset with inaccurate perturbed poses." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 312, + 296, + 370 + ], + "lines": [ + { + "bbox": [ + 57, + 312, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 57, + 312, + 296, + 370 + ], + "type": "table", + "html": "
MetricsSpikeGSOurs
0%10%20%30%0%10%20%30%
PSNR ↑25.6918.8317.7516.4426.4726.0525.9323.46
SSIM ↑0.8350.6070.5920.5670.8620.8540.8510.772
LPIPS ↓0.2470.3720.4110.4460.1760.1840.1860.245
", + "image_path": "451a290a7973a7f24a6d3d90f58ab87e6576beb982b2dfa7fa77ce317796057f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 378, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 378, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 54, + 378, + 295, + 544 + ], + "type": "text", + "content": "two-stage approach, we initially apply TFP [48], TFI [48], and Spk2ImgNet [41] to reconstruct sharp images from the input spike followed by the common 3DGS processing pipeline, with the code implemented by Spike-Zoo [4]. All Gaussian-based methods are conducted based on the same 3DGS backbone, and identical camera poses input for a fair comparison. While SpikeGS requires several poses per spike stream for the exposure loss, which presents a notable challenge for COLMAP in real-world data, we omit the exposure loss and retain the rest of the framework. All comparative metrics are calculated based on the 3D reconstruction task, with further comparison on the spike-to-image reconstruction task and implementation details available in the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Synthetic Dataset. Quantitative and qualitative comparisons on the synthetic dataset are shown in Tab. 1 and Fig. 4 respectively. As illustrated in the Tab. 1, our USP-Gaussian achieves the best performance across most metrics in various scenes. Two-stage cascading methods TFP-3DGS and TFI-3DGS yield poor restoration due to the substantial noise embedded in the initially reconstructed sequence. Similarly, although Spk2ImgNet performs well on its original REDS dataset [41], it experiences substantial degradation due to the dataset domain gap. SpikeGS, as the primary comparison algorithm in this paper, heavily relies on the BSN [15] for self-supervised Spike-to-Image reconstruction. Nevertheless, the denoising performance of the BSN deteriorates in the presence of substantial noise," + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 318, + 292, + 433, + 386 + ], + "blocks": [ + { + "bbox": [ + 319, + 282, + 430, + 291 + ], + "lines": [ + { + "bbox": [ + 319, + 282, + 430, + 291 + ], + "spans": [ + { + "bbox": [ + 319, + 282, + 430, + 291 + ], + "type": "text", + "content": "Pose Trajectory with " + }, + { + "bbox": [ + 319, + 282, + 430, + 291 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 319, + 282, + 430, + 291 + ], + "type": "text", + "content": " Perturbation" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 318, + 292, + 433, + 386 + ], + "lines": [ + { + "bbox": [ + 318, + 292, + 433, + 386 + ], + "spans": [ + { + "bbox": [ + 318, + 292, + 433, + 386 + ], + "type": "image", + "image_path": "66a58d80780aea1dbd490ce733c76f8de17c71039039fbf4e62018ce76ad5cc6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 397, + 541, + 408 + ], + "lines": [ + { + "bbox": [ + 326, + 397, + 541, + 408 + ], + "spans": [ + { + "bbox": [ + 326, + 397, + 541, + 408 + ], + "type": "text", + "content": "Figure 6. Visual comparison of initial and optimized poses." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 436, + 292, + 551, + 386 + ], + "blocks": [ + { + "bbox": [ + 440, + 282, + 547, + 291 + ], + "lines": [ + { + "bbox": [ + 440, + 282, + 547, + 291 + ], + "spans": [ + { + "bbox": [ + 440, + 282, + 547, + 291 + ], + "type": "text", + "content": "Pose Trajectory with " + }, + { + "bbox": [ + 440, + 282, + 547, + 291 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 440, + 282, + 547, + 291 + ], + "type": "text", + "content": " Perturbation" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 436, + 292, + 551, + 386 + ], + "lines": [ + { + "bbox": [ + 436, + 292, + 551, + 386 + ], + "spans": [ + { + "bbox": [ + 436, + 292, + 551, + 386 + ], + "type": "image", + "image_path": "36a96113877b65220968cff14aa1c85ae9b68057a14edf353ec6a47b98b11da5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 316, + 434, + 554, + 483 + ], + "blocks": [ + { + "bbox": [ + 329, + 413, + 538, + 425 + ], + "lines": [ + { + "bbox": [ + 329, + 413, + 538, + 425 + ], + "spans": [ + { + "bbox": [ + 329, + 413, + 538, + 425 + ], + "type": "text", + "content": "Table 3. Translation and rotation poses error comparison." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 434, + 554, + 483 + ], + "lines": [ + { + "bbox": [ + 316, + 434, + 554, + 483 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 554, + 483 + ], + "type": "table", + "html": "
StateTranslation ↓Rotation ↓
10%20%30%Avg10%20%30%Avg
Initial1.3022.9674.1932.8201.1322.1503.1592.147
Optimized0.8551.7811.4931.3760.1740.3201.2850.593
Δ (%)34.3%39.9%64.4%51.3%84.6%85.1%59.2%72.4%
", + "image_path": "a9c05e603693f1f54d49c5a360d31d420d7e23842a823e19fcd811e9347d3a7d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 493, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 555, + 540 + ], + "type": "text", + "content": "leading to pronounced grid artifacts and the attenuation of high-frequency details of the SpikeGS. SpikeNeRF faces a comparable challenge, as it requires noisy spike images as input and relies on the SNN for denoising." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 543, + 554, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 554, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 554, + 614 + ], + "type": "text", + "content": "To this end, our proposed collaborative learning framework USP-Gaussian, benefiting from the inherent constraints between the spike stream and the multi-view consistency afforded by the 3DGS, effectively addresses the aforementioned problems without relying on the pre-training of the spike-based image reconstruction network." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 616, + 554, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 554, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 554, + 699 + ], + "type": "text", + "content": "In the previous part, accurate poses are obtained by applying COLMAP on sharp images rendered by Blender. To demonstrate that the integration of pose correction enhances the robustness of our method, we apply random perturbations of " + }, + { + "bbox": [ + 313, + 616, + 554, + 699 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 616, + 554, + 699 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 616, + 554, + 699 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 616, + 554, + 699 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 616, + 554, + 699 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 616, + 554, + 699 + ], + "type": "text", + "content": " to the accurate poses in the wine scene, generating inaccurate poses for SpikeGS and our method with the comparison shown in Tab. 2." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "type": "text", + "content": "Furthermore, visual trajectories and quantitative errors" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16615" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 91, + 296, + 158 + ], + "blocks": [ + { + "bbox": [ + 77, + 71, + 273, + 82 + ], + "lines": [ + { + "bbox": [ + 77, + 71, + 273, + 82 + ], + "spans": [ + { + "bbox": [ + 77, + 71, + 273, + 82 + ], + "type": "text", + "content": "Table 4. Quantitative ablation on the synthetic dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 91, + 296, + 158 + ], + "lines": [ + { + "bbox": [ + 58, + 91, + 296, + 158 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 296, + 158 + ], + "type": "table", + "html": "
ID\\( \\mathcal{L}_{\\text{gs}} \\)Recon-Net\\( \\mathcal{L}_{\\text{joint}} \\)3DGS OutputsRecon-Net Outputs
\\( \\mathbb{L}_{\\text{rec}} \\)\\( \\mathbb{L}_{\\text{rec}} \\)PSNRSSIMLPIPSPSNRSSIMLPIPS
I✓22.3920.7500.208\\\\\\
II✓\\\\\\24.2550.7080.358
III✓\\\\\\24.5670.7440.332
IV✓✓✓25.6590.8500.18425.5710.8070.244
V✓✓✓26.4760.8620.17627.1380.8550.178
", + "image_path": "fec5786a936e57ca467775dcdddfee9420a0403165e8090c8cfdaa69848bafc8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 169, + 175, + 267 + ], + "blocks": [ + { + "bbox": [ + 58, + 169, + 175, + 267 + ], + "lines": [ + { + "bbox": [ + 58, + 169, + 175, + 267 + ], + "spans": [ + { + "bbox": [ + 58, + 169, + 175, + 267 + ], + "type": "image", + "image_path": "3b6d9336d9428282244b8c015c97d307596215c9132f3ceeb6839604aa1ca878.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 277, + 295, + 300 + ], + "lines": [ + { + "bbox": [ + 55, + 277, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 295, + 300 + ], + "type": "text", + "content": "Figure 7. Visual ablation on the effectiveness of multi-reblur loss, with the x-axis and y-axis representing step (k) and PSNR." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 175, + 169, + 294, + 267 + ], + "blocks": [ + { + "bbox": [ + 175, + 169, + 294, + 267 + ], + "lines": [ + { + "bbox": [ + 175, + 169, + 294, + 267 + ], + "spans": [ + { + "bbox": [ + 175, + 169, + 294, + 267 + ], + "type": "image", + "image_path": "c4f9d7ad94ec750ddedc8f23e68394e0223686d5071d530669adfdb4d3601022.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 160, + 552, + 172 + ], + "lines": [ + { + "bbox": [ + 315, + 160, + 552, + 172 + ], + "spans": [ + { + "bbox": [ + 315, + 160, + 552, + 172 + ], + "type": "text", + "content": "Figure 8. Visual ablation on the effectiveness of long spike input." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 312, + 296, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 359 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 359 + ], + "type": "text", + "content": "of the initial noisy, and optimized poses are depicted in Fig. 6 and Tab. 3 respectively, with MAE measuring translation error and radians measuring rotation error, demonstrating the pose correction capability of our USP-Gaussian." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 361, + 296, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 482 + ], + "type": "text", + "content": "Real-world Dataset. We also conduct the visual comparison on the real-world dataset as shown in Fig. 5. Due to the rapid camera ego-motion and the low imaging quality of the spike camera characterized by low spatial resolution, insufficient texture details, and underexposure, the pose sequence estimated by the COLMAP is highly inaccurate, presenting significant challenges for previous methods to recover accurate 3D representation as observed in Fig. 5. Nevertheless, our USP-Gaussian overcomes this limitation benefiting from the unifying optimization of the pose." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 490, + 149, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 149, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 149, + 502 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 509, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 544 + ], + "type": "text", + "content": "In this subsection, we demonstrate the effectiveness of our proposed joint optimization, complementary spike input, and multi-reblur loss strategies." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 545, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 652 + ], + "type": "text", + "content": "I. Joint Learning. The highlight of USP-Gaussian lies in the design of the joint loss, which aligns outputs from the 3DGS and Recon-Net thus achieving complementary performance improvement. We conduct the ablation experiment on loss functions of " + }, + { + "bbox": [ + 55, + 545, + 295, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{gs}}" + }, + { + "bbox": [ + 55, + 545, + 295, + 652 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 545, + 295, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rec}}" + }, + { + "bbox": [ + 55, + 545, + 295, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 545, + 295, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{joint}}" + }, + { + "bbox": [ + 55, + 545, + 295, + 652 + ], + "type": "text", + "content": " with quantitative results listed in Tab. 4, which shows that the incorporation of the joint loss (ID-V) yields notable improvements for both 3DGS and Recon-Net compared to independent optimization (ID-I and ID-III)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Further visual ablation comparisons of the restored image and the training curve are presented in Fig. 1. Visual comparison substantiates that the joint optimization of 3DGS and Recon-Net significantly enhances the texture details for both two. Regarding the training curve, the inde" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 395, + 140 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 395, + 140 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 395, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 395, + 140 + ], + "type": "image", + "image_path": "880659126e051fac901ed855ca02e65a271eecad1e13d3c0807115b5ae865b80.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 141, + 370, + 149 + ], + "lines": [ + { + "bbox": [ + 342, + 141, + 370, + 149 + ], + "spans": [ + { + "bbox": [ + 342, + 141, + 370, + 149 + ], + "type": "text", + "content": "TFP-Input" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 395, + 72, + 473, + 140 + ], + "blocks": [ + { + "bbox": [ + 395, + 72, + 473, + 140 + ], + "lines": [ + { + "bbox": [ + 395, + 72, + 473, + 140 + ], + "spans": [ + { + "bbox": [ + 395, + 72, + 473, + 140 + ], + "type": "image", + "image_path": "12a12df4201757396d4581728e8baa552e08c98e06ef4579be5bc3b2749bf7fc.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 141, + 455, + 149 + ], + "lines": [ + { + "bbox": [ + 413, + 141, + 455, + 149 + ], + "spans": [ + { + "bbox": [ + 413, + 141, + 455, + 149 + ], + "type": "text", + "content": "w/o Long Spike" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 474, + 72, + 552, + 140 + ], + "blocks": [ + { + "bbox": [ + 474, + 72, + 552, + 140 + ], + "lines": [ + { + "bbox": [ + 474, + 72, + 552, + 140 + ], + "spans": [ + { + "bbox": [ + 474, + 72, + 552, + 140 + ], + "type": "image", + "image_path": "20556ac5b277f054d730d8f4dff0dcef3ca4c0c9ebffa9ac42441f85d151bda5.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 141, + 532, + 149 + ], + "lines": [ + { + "bbox": [ + 493, + 141, + 532, + 149 + ], + "spans": [ + { + "bbox": [ + 493, + 141, + 532, + 149 + ], + "type": "text", + "content": "w/ Long Spike" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 185, + 553, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 185, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 185, + 553, + 304 + ], + "type": "text", + "content": "pendent 3DGS training (ID-I) demonstrates slow convergence due to the absence of crucial information in the long-exposure images. Concurrently, the independent training of Recon-Net (ID-III) tends to overfit due to the simplicity of the spike reconstruction task. To this end, our proposed joint optimization framework provides synergistic benefits for both, i.e., Recon-Net effectively compensates for deficient motion features embedded in the long-exposures image while the multi-view consistency afforded by 3DGS culminates in coherent outputs from the Recon-Net." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 305, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 554, + 437 + ], + "type": "text", + "content": "II. Multi-reblur Loss. The multi-reblur loss " + }, + { + "bbox": [ + 313, + 305, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rec}}" + }, + { + "bbox": [ + 313, + 305, + 554, + 437 + ], + "type": "text", + "content": " in Eq. (7) is designed to substitute the single-reblur loss " + }, + { + "bbox": [ + 313, + 305, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\mathbb{L}_{\\mathrm{rec}}" + }, + { + "bbox": [ + 313, + 305, + 554, + 437 + ], + "type": "text", + "content": " in Eq. (8). Without further constraints, " + }, + { + "bbox": [ + 313, + 305, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\mathbb{L}_{\\mathrm{rec}}" + }, + { + "bbox": [ + 313, + 305, + 554, + 437 + ], + "type": "text", + "content": " will lead ReconNet to learn an identity mapping from the input spike stream to the corresponding long-exposure image as demonstrated in [3]. Comparative analysis in Tab. 4 depicts that " + }, + { + "bbox": [ + 313, + 305, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rec}}" + }, + { + "bbox": [ + 313, + 305, + 554, + 437 + ], + "type": "text", + "content": " (ID-V) yields beneficial improvement than " + }, + { + "bbox": [ + 313, + 305, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\mathbb{L}_{\\mathrm{rec}}" + }, + { + "bbox": [ + 313, + 305, + 554, + 437 + ], + "type": "text", + "content": " (ID-IV). Further comparison on the training curve in Fig. 7 reveals the mechanism of collaborative training, i.e., the multi-reblur loss enhances the performance of Recon-Net, thereby facilitating the training of the 3DGS." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 438, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 554, + 521 + ], + "type": "text", + "content": "III. Complementary Input. We utilize complementary long and short spike stream inputs for Recon-Net to address the fixed short window limitation inherent in previous methods [6, 36, 41]. As illustrated in the visual comparison in Fig. 8, the inclusion of the long spike effectively suppresses salt-and-pepper noise in the Recon-Net output, leading to improved SSIM and LPIPS metrics." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 536, + 388, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 536, + 388, + 548 + ], + "spans": [ + { + "bbox": [ + 314, + 536, + 388, + 548 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 557, + 554, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 554, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 554, + 664 + ], + "type": "text", + "content": "In this paper, we propose a unifying framework USP-Gaussian for the collaborative optimization of spike-based image reconstruction, pose correction, and Gaussian Splating. By mitigating error propagation inherent in traditional cascading frameworks, USP-Gaussian attains high reconstruction fidelity across both synthetic and real-world datasets, with further ablation experiments revealing that Spike-to-Image reconstruction and 3D reconstruction tasks furnish mutually complementary information." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Limitation. The limitation of USP-Gaussian is that it requires longer training time and more GPU memory compared to previous methods due to the additional poses and Recon-Net parameters to be optimized." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16616" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 199 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 199 + ], + "type": "text", + "content": "We sincerely appreciate Yuyan Chen (HUST) for her valuable suggestions and for polishing the figures. This work was supported by the National Natural Science Foundation of China (62422601, 62176003, 62088102, 62306015), the China Postdoctoral Science Foundation (2023T160015), the Young Elite Scientists Sponsorship Program by CAST (2023QNRC001), the Beijing Municipal Science and Technology Program (Z241100004224004) and the Beijing Nova Program (20230484362,20240484703)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 210, + 115, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 210, + 115, + 223 + ], + "spans": [ + { + "bbox": [ + 56, + 210, + 115, + 223 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 230, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 230, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 230, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 61, + 230, + 296, + 285 + ], + "type": "text", + "content": "[1] Anish Bhattacharya, Ratnesh Madaan, Fernando Cladera, Sai Vemprala, Rogerio Bonatti, Kostas Daniilidis, Ashish Kapoor, Vijay Kumar, Nikolai Matni, and Jayesh K Gupta. Evdnerf: Reconstructing event data with dynamic neural radiance fields. In WACV, pages 5846-5855, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 286, + 295, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 286, + 295, + 307 + ], + "spans": [ + { + "bbox": [ + 62, + 286, + 295, + 307 + ], + "type": "text", + "content": "[2] Kang Chen and Lei Yu. Motion deblur by learning residual from events. IEEE TMM, 2024. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 309, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 309, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 62, + 309, + 294, + 363 + ], + "type": "text", + "content": "[3] Kang Chen, Shiyan Chen, Jiyuan Zhang, Baoyue Zhang, Yajing Zheng, Tiejun Huang, and Zhaofei Yu. Spikereveal: Unlocking temporal sequences from real blurry inputs with spike streams. Advances in Neural Information Processing Systems, 37:62673-62696, 2025. 2, 5, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 365, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 294, + 408 + ], + "type": "text", + "content": "[4] Kang Chen, Zhiyuan Ye, Tiejun Huang, and Zhaofei Yu. Spike-Zoo: A toolbox for spike-to-image reconstruction. https://github.com/chenkang455/Spike-Zoo, 2025.7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 411, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 411, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 62, + 411, + 294, + 443 + ], + "type": "text", + "content": "[5] Kang Chen, Yajing Zheng, Tiejun Huang, and Zhaofei Yu. Rethinking high-speed image reconstruction framework with spike camera. arXiv preprint arXiv:2501.04477, 2025. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 445, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 294, + 487 + ], + "type": "text", + "content": "[6] Shiyan Chen, Chaoteng Duan, Zhaofei Yu, Ruiqin Xiong, and Tiejun Huang. Self-supervised mutual learning for dynamic scene reconstruction of spiking camera. In IJCAI, pages 2859-2866, 2022. 2, 4, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 490, + 294, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 490, + 294, + 532 + ], + "spans": [ + { + "bbox": [ + 62, + 490, + 294, + 532 + ], + "type": "text", + "content": "[7] Shiyan Chen, Zhaofei Yu, and Tiejun Huang. Self-supervised joint dynamic scene reconstruction and optical flow estimation for spiking camera. In AAAI, pages 350-358, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 535, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 535, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 535, + 294, + 567 + ], + "type": "text", + "content": "[8] Shiyan Chen, Jiyuan Zhang, Yajing Zheng, Tiejun Huang, and Zhaofei Yu. Enhancing motion deblurring in high-speed scenes with spike streams. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 569, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 569, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 62, + 569, + 294, + 621 + ], + "type": "text", + "content": "[9] Gaole Dai, Zhenyu Wang, Qinwen Xu, Wen Cheng, Ming Lu, Boxing Shi, Shanghang Zhang, and Tiejun Huang. Spikenvs: Enhancing novel view synthesis from blurry images via spike camera. arXiv preprint arXiv:2404.06710, 2024.3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 677 + ], + "type": "text", + "content": "[10] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE TPAMI, 44(1):154-180, 2020. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "text", + "content": "[11] Yijia Guo, Liwen Hu, Lei Ma, and Tiejun Huang. Spikegs: Reconstruct 3d scene via fast-moving bio-inspired sensors. arXiv preprint arXiv:2407.03771, 2024. 2, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[12] Tiejun Huang, Yajing Zheng, Zhaofei Yu, Rui Chen, Yuan Li, Ruiqin Xiong, Lei Ma, Junwei Zhao, Siwei Dong, Lin Zhu, et al. " + }, + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "inline_equation", + "content": "1000\\times" + }, + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": " faster camera and machine vision with ordinary devices. Engineering, 25:110-119, 2023. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "text", + "content": "[13] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 163, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 205 + ], + "type": "text", + "content": "[14] Simon Klenk, Lukas Koestler, Davide Scaramuzza, and Daniel Cremers. E-nerf: Neural radiance fields from a moving event camera. IEEE Robotics and Automation Letters, 8 (3):1587–1594, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 208, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 239 + ], + "type": "text", + "content": "[15] Samuli Laine, Tero Karras, Jaakko Lehtinen, and Timo Aila. High-quality self-supervised deep image denoising. NeurIPS, 32, 2019. 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "text", + "content": "[16] Wenpu Li, Pian Wan, Peng Wang, Jinghang Li, Yi Zhou, and Peidong Liu. Benerf: Neural radiance fields from a single blurry image and event stream. In European Conference on Computer Vision (ECCV), 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 287, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 329 + ], + "type": "text", + "content": "[17] Bohao Liao, Wei Zhai, Zengyu Wan, Tianzhu Zhang, Yang Cao, and Zheng-Jun Zha. Ef-3dgs: Event-aided free-trajectory 3d gaussian splatting. arXiv preprint arXiv:2410.15392, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 332, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 385 + ], + "type": "text", + "content": "[18] Zibin Liu, Banglei Guan, Yang Shang, Shunkun Liang, Zhenbao Yu, and Qifeng Yu. Optical flow-guided 6dof object pose tracking with an event camera. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6501-6509, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "text", + "content": "[19] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-nerf: Neural radiance fields from blurry images. In CVPR, pages 12861-12870, 2022. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 433, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 464 + ], + "type": "text", + "content": "[20] Qi Ma, Danda Pani Paudel, Ajad Chhatkuli, and Luc Van Gool. Deformable neural radiance fields using rgb and event cameras. In ICCV, pages 3590-3600, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 467, + 553, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 467, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 317, + 467, + 553, + 509 + ], + "type": "text", + "content": "[21] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 511, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 553, + 542 + ], + "type": "text", + "content": "[22] Bryan Peterson. Understanding exposure: how to shoot great photographs with any camera. AmPhoto books, 2016. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 544, + 553, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 587 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 587 + ], + "type": "text", + "content": "[23] Viktor Rudnev, Mohamed Elgharib, Christian Theobalt, and Vladislav Golyanik. Eventnerf: Neural radiance fields from a single colour event camera. In CVPR, pages 4992-5002, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 590, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 621 + ], + "type": "text", + "content": "[24] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, pages 4104-4113, 2016. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 624, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 655 + ], + "type": "text", + "content": "[25] Hyeonjun Sim, Jihyong Oh, and Munchurl Kim. Xvfi: extreme video frame interpolation. In ICCV, pages 14489-14498, 2021. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "text", + "content": "[26] Chen Song, Qixing Huang, and Chandrajit Bajaj. E-cir: Event-enhanced continuous intensity recovery. In CVPR, pages 7803-7812, 2022. 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[27] Jing Tao, You Li, Banglei Guan, Yang Shang, and Qifeng Yu. Simultaneous enhancement and noise suppression un" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16617" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "der complex illumination conditions. IEEE Transactions on Instrumentation and Measurement, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "text", + "content": "[28] Bishan Wang, Jingwei He, Lei Yu, Gui-Song Xia, and Wen Yang. Event enhanced high-quality image recovery. In ECCV, pages 155-171. Springer, 2020. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 132, + 294, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 132, + 294, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 132, + 294, + 175 + ], + "type": "text", + "content": "[29] Jiaxu Wang, Junhao He, Ziyi Zhang, Mingyuan Sun, Jingkai Sun, and Renjing Xu. Evggs: A collaborative learning framework for event-based generalizable gaussian splatting. arXiv preprint arXiv:2405.14959, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 177, + 294, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 177, + 294, + 210 + ], + "spans": [ + { + "bbox": [ + 56, + 177, + 294, + 210 + ], + "type": "text", + "content": "[30] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. Bad-nerf: Bundle adjusted deblur neural radiance fields. In CVPR, pages 4170–4179, 2023. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 212, + 294, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 212, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 56, + 212, + 294, + 266 + ], + "type": "text", + "content": "[31] Tianyi Xiong, Jiayi Wu, Botao He, Cornelia Fermuller, Yiannis Aloimonos, Heng Huang, and Christopher Metzler. Event3dgs: Event-based 3d gaussian splatting for high-speed robot egomotion. In 8th Annual Conference on Robot Learning, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 268, + 294, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 268, + 294, + 301 + ], + "spans": [ + { + "bbox": [ + 56, + 268, + 294, + 301 + ], + "type": "text", + "content": "[32] Fang Xu, Lei Yu, Bishan Wang, Wen Yang, Gui-Song Xia, Xu Jia, Zhendong Qiao, and Jianzhuang Liu. Motion deblurring with real events. In ICCV, pages 2583-2592, 2021. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 303, + 294, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 303, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 56, + 303, + 294, + 346 + ], + "type": "text", + "content": "[33] Baoyue Zhang, Yajing Zheng, Shiyan Chen, Jiyuan Zhang, Kang Chen, Zhaofei Yu, and Tiejun Huang. Spikemm: Fleximagnification of high-speed micro-motions. arXiv preprint arXiv:2406.00383, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 348, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 348, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 348, + 294, + 381 + ], + "type": "text", + "content": "[34] Jiyuan Zhang, Lulu Tang, Zhaofei Yu, Jiwen Lu, and Tiejun Huang. Spike transformer: Monocular depth estimation for spiking camera. In ECCV, pages 34-52. Springer, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 384, + 294, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 384, + 294, + 427 + ], + "spans": [ + { + "bbox": [ + 56, + 384, + 294, + 427 + ], + "type": "text", + "content": "[35] Jiyuan Zhang, Shiyan Chen, Yajing Zheng, Zhaofei Yu, and Tiejun Huang. Unveiling the potential of spike streams for foreground occlusion removal from densely continuous views. arXiv preprint arXiv:2307.00821, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 429, + 294, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 429, + 294, + 472 + ], + "spans": [ + { + "bbox": [ + 56, + 429, + 294, + 472 + ], + "type": "text", + "content": "[36] Jiyuan Zhang, Shanshan Jia, Zhaofei Yu, and Tiejun Huang. Learning temporal-ordered representation for spike streams based on discrete wavelet transforms. In AAAI, pages 137–147, 2023. 2, 4, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 475, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 294, + 518 + ], + "type": "text", + "content": "[37] Jiyuan Zhang, Kang Chen, Shiyan Chen, Yajing Zheng, Tiejun Huang, and Zhaofei Yu. Spikegs: 3d gaussian splatt-ting from spike streams with high-speed camera motion. arXiv preprint arXiv:2407.10062, 2024. 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "type": "text", + "content": "[38] Jiyuan Zhang, Shiyan Chen, Yajing Zheng, Zhaofei Yu, and Tiejun Huang. Spike-guided motion deblurring with unknown modal spatiotemporal alignment. In CVPR, pages 25047-25057, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 566, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 566, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 566, + 294, + 597 + ], + "type": "text", + "content": "[39] Xiang Zhang and Lei Yu. Unifying motion deblurring and frame interpolation with events. In CVPR, pages 17765-17774, 2022. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 600, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 600, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 600, + 294, + 633 + ], + "type": "text", + "content": "[40] Xiang Zhang, Lei Yu, Wen Yang, Jianzhuang Liu, and GuiSong Xia. Generalizing event-based motion deblurring in real-world scenarios. In ICCV, pages 10734-10744, 2023. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "type": "text", + "content": "[41] Jing Zhao, Ruiqin Xiong, Hangfan Liu, Jian Zhang, and Tiejun Huang. Spk2imgnet: Learning to reconstruct dynamic scene from continuous spike stream. In CVPR, pages 11996-12005, 2021. 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "text", + "content": "[42] Lingzhe Zhao, Peng Wang, and Peidong Liu. Bad-gaussians: Bundle adjusted deblur gaussian splatting. arXiv preprint arXiv:2403.11831, 2024. 5" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 555, + 420 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 106 + ], + "type": "text", + "content": "[43] Rui Zhao, Ruiqin Xiong, Jing Zhao, Zhaofei Yu, Xiaopeng Fan, and Tiejun Huang. Learning optical flow from continuous spike streams. NeurIPS, 35:7905-7920, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 108, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 150 + ], + "type": "text", + "content": "[44] Rui Zhao, Ruiqin Xiong, Jing Zhao, Jian Zhang, Xiaopeng Fan, Zhaofei Yu, and Tiejun Huang. Boosting spike camera image reconstruction from a perspective of dealing with spike fluctuations. In CVPR, pages 24955-24965, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[45] Yajing Zheng, Lingxiao Zheng, Zhaofei Yu, Boxin Shi, Yonghong Tian, and Tiejun Huang. High-speed image reconstruction through short-term plasticity for spiking cameras. In CVPR, pages 6358-6367, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "type": "text", + "content": "[46] Yajing Zheng, Zhaofei Yu, Song Wang, and Tiejun Huang. Spike-based motion estimation for object tracking through bio-inspired unsupervised learning. IEEE TIP, 32:335-349, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 242, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 242, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 242, + 553, + 274 + ], + "type": "text", + "content": "[47] Kangrui Zhou, Taihang Lei, Banglei Guan, and Qifeng Yu. Event-based depth estimation with dense occlusion. Optics Letters, 49(12):3376-3379, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 276, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 317 + ], + "type": "text", + "content": "[48] Lin Zhu, Siwei Dong, Tiejun Huang, and Yonghong Tian. A retina-inspired sampling method for visual texture reconstruction. In ICME, pages 1432-1437. IEEE, 2019. 2, 4, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "type": "text", + "content": "[49] Lin Zhu, Siwei Dong, Jianing Li, Tiejun Huang, and Yonghong Tian. Retina-like visual image reconstruction via spiking neural model. In CVPR, pages 1438-1446, 2020. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 365, + 555, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 555, + 420 + ], + "type": "text", + "content": "[50] Lin Zhu, Kangmin Jia, Yifan Zhao, Yunshan Qi, Lizhi Wang, and Hua Huang. Spikenerf: Learning neural radiance fields from continuous spike stream. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6285-6295, 2024. 2, 3, 6" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16618" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/UVGS_ Reimagining Unstructured 3D Gaussian Splatting using UV Mapping/77087a0b-d1cb-4230-a846-7760f7510d17_content_list.json b/2025/UVGS_ Reimagining Unstructured 3D Gaussian Splatting using UV Mapping/77087a0b-d1cb-4230-a846-7760f7510d17_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..79d4f9183e2c44ac303452bdece3b97e9330c4f8 --- /dev/null +++ b/2025/UVGS_ Reimagining Unstructured 3D Gaussian Splatting using UV Mapping/77087a0b-d1cb-4230-a846-7760f7510d17_content_list.json @@ -0,0 +1,1785 @@ +[ + { + "type": "text", + "text": "UVGS: Reimagining Unstructured 3D Gaussian Splitting using UV Mapping", + "text_level": 1, + "bbox": [ + 109, + 130, + 888, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Aashish Rai $^{1,2}$", + "bbox": [ + 122, + 178, + 241, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dilin Wang $^{2}$ Mihir Jain $^{2}$ Srinath Sridhar $^{1}$", + "bbox": [ + 277, + 179, + 511, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nikolaos Sarafianos2 \nAayush Prakash2", + "bbox": [ + 527, + 180, + 720, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kefan Chen $^{1,2}$", + "bbox": [ + 758, + 180, + 872, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Brown University", + "bbox": [ + 339, + 227, + 491, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meta Reality Labs", + "bbox": [ + 519, + 228, + 676, + 246 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://aashishrai3799.github.io/uvgs", + "bbox": [ + 330, + 257, + 661, + 272 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5f9e3c045a3b9caf6b11549a80329ba7ba55bde2f88692372e9d39e48b98fb56.jpg", + "image_caption": [ + "Figure 1. We propose UVGS - an structured image-like representation for 3DGS obtained by spherical mapping of its primitives. The obtained UVGS maps can be further unified to a 3-channel Super UVGS image to bridging the gap between 3DGS and existing image foundation models. We show Super UVGS can compress the 3DGS assets using pretrained image Autoencoders, and for direct unconditional and conditional 3DGS object generation using diffusion models. We also present one of the first experiments on 3DGS inpainting." + ], + "image_footnote": [], + "bbox": [ + 111, + 287, + 883, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 574, + 326, + 589 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Gaussian Splatting (3DGS) has demonstrated superior quality in modeling 3D objects and scenes. However, generating 3DGS remains challenging due to their discrete, unstructured, and permutation-invariant nature. In this work, we present a simple yet effective method to overcome these challenges. We utilize spherical mapping to transform 3DGS into a structured 2D representation, termed UVGS. UVGS can be viewed as multi-channel images, with feature dimensions as a concatenation of Gaussian attributes such as position, scale, color, opacity, and rotation. We further find that these heterogeneous features can be compressed into a lower-dimensional (e.g., 3-channel) shared feature space using a carefully designed multi-branch network. The compressed UVGS can be treated as typical RGB images. Remarkably, we discover that typical VAEs trained with latent diffusion models can directly generalize to this new representation without additional training. Our novel representation makes it effortless to leverage foundational 2D models, such as diffusion models, to directly model 3DGS. Additionally, one can simply increase the 2D UV resolution", + "bbox": [ + 88, + 598, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to accommodate more Gaussians, making UVGS a scalable solution compared to typical 3D backbones. This approach immediately unlocks various novel generation applications of 3DGS by inherently utilizing the already developed superior 2D generation capabilities. In our experiments, we demonstrate various unconditional, conditional generation, and inpainting applications of 3DGS based on diffusion models, which were previously non-trivial.", + "bbox": [ + 511, + 575, + 906, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 700, + 645, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The creation of high-quality 3D content is essential in applications like virtual reality, game design, robotics, and movie production, where realistic 3D representations play a critical role. Typical 3D representations like Neural Radiance Fields (NeRF) [25] are promising but require substantial computational resources, limiting their scalability for real-time applications. Moreover, NeRF is an implicit representation, which makes editing and manipulation challenging. Recently, 3D Gaussian Splatting (3DGS) [17] emerged as a compelling alternative, enabling efficient and high-fidelity 3D rendering through a large set of Gaussian primitives that model spatial and visual properties. As an explicit represen", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5927", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tation, 3DGS offers several advantages over NeRF. However, while 3DGS offers benefits in terms of speed and visual quality, its unstructured, permutation-invariant nature presents significant challenges for generative tasks. Much like point clouds, it lacks a coherent spatial structure, impeding its integration with conventional image-based generative models. This lack of structure and coherence among primitives hinders the application of image-based generative models [23, 55, 61], which rely on structured data representations.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous methods have tackled these challenges by transforming 3DGS into structured formats, such as voxel grids [8, 14, 55] or image-based representations like Splatter Image [40] or triplanes [64]. Other approaches employ diffusion models to directly predict 3DGS attributes [28]. These methods, while achieving impressive visual results, often require substantial computational resources, memory-intensive multi-view rendering, complex architectures limiting their scalability and flexibility for high-fidelity generation. Generating and processing 3DGS directly by efficiently utilizing modern generative models like Variational Autoencoders (VAEs) and diffusion models is limited as the neural networks are not permutation invariant.", + "bbox": [ + 89, + 244, + 482, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these shortcomings, we introduce UV Gaussian Splatting (UVGS), which provides a structured transformation of 3D Gaussian primitives into a 2D representation while preserving essential 3D information. We use spherical mapping [37] that inscribes Gaussian splats in a spherical surface, and projects attributes like position, rotation, scale, opacity, and color into an organized 14-channel image-like UV map. This mapping introduces spatial structure, resolving issues of permutation invariance by introducing local correspondences between neighboring Gaussians and global coherence across the entire 3D object. The result is a representation that functions as a \"3D representation\" structured in a 2D map format, enabling compatibility with powerful image-based neural network architectures.", + "bbox": [ + 89, + 444, + 482, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While UVGS introduces structure into 3D Gaussian Splitting, its full 14-channel attribute-specific representation presents challenges for direct integration with pretrained 2D generative models, as these models typically expect a simpler, image-compatible data. Each of its heterogeneous attributes—position, color, and transformation—has its own distinct distribution and resides in a separate feature space, making it challenging to represent the 3D object in a unified shared space. To address this, we introduce Super UVGS, a compact 3-channel representation that unifies these diverse attributes into a cohesive format. Using a carefully designed multi-branch mapping network, Super UVGS consolidates the distinct attribute spaces into a shared feature space, enabling a more collective representation of the object. This unified transformation not only facilitates zero-shot compatibility with pretrained 2D models", + "bbox": [ + 89, + 657, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "but also optimizes memory usage and computational efficiency, making Super UVGS highly practical for large-scale 3D tasks. Unlike previous approaches that use Triplanes, voxels, occupancy grid, neural fields etc. and require specialized 3D architectures to train on 3D data, UVGS effortlessly leverages widely available pretrained 2D foundational models. This zero-shot generalization capability allows UVGS to fully benefit from priors learned in 2D domains from large amount of data, improving both flexibility and scalability. To sum up our main contributions are:", + "bbox": [ + 511, + 90, + 903, + 241 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Efficient Structured Representation of 3DGS: We present UVGS, an image-like representation that solves permutation invariance and unstructured nature of discrete 3DGS through spherical mapping, making direct feature extraction possible by organizing unordered points into a coherent 2D representation compatible with 2D models.", + "- Compact and Scalable Super UVGS Representation: To address scalability while dealing with large scale 3DGS points and enabling the direct integration of pre-trained 2D foundation models, we introduce Super UVGS - a low-dimensional version of UVGS maps that retains high fidelity features while reducing memory overhead.", + "- Diverse 3D Applications: Our approach unlocks seamless integration of 3DGS with pre-trained 2D foundation models for various tasks, including unconditional and conditional generation of 3DGS." + ], + "bbox": [ + 513, + 243, + 903, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 497, + 653, + 512 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Generative/Reconstruction Models for Objects: Generation or reconstruction of 3D assets has been a long standing task [3, 21, 24, 26, 27, 31, 34, 35, 39, 41, 43, 50, 54]. Previous reconstruction approaches like NeRF [1, 25] are often slow and do not provide a defining geometry [5, 11, 16, 20, 22, 24, 32, 35, 48, 60, 62]. Advancements in the field led to the emergence of explicit voxel grid based representations that encode colors and opacities directly [8, 29, 45]. These approaches achieve significant speed ups compared to the NeRF based approaches, but they can't produce high fidelity assets due to the low resolution of voxel grids. On the other hand, triplane representation [13, 38, 47, 63] provides a trade-off between the quality and memory utilization. Another line of work [10, 53] splits the input mesh into different patches and simplifies the object generation problem to an image generation problem. However such methods rely on either cutting through the mesh to create a geometry image [10] or rely on an existing UV representation of the geometry and utilize subset of existing UV islands [53] resulting in loss of details. Recently, there has been a notable advancement in 3D Gaussian Splatting (3DGS) for the representation of objects and scenes leading to the emergence of 3DGS showcasing impressive real-time results in reconstruction and generation tasks [14, 17, 19, 30, 42, 49, 51, 52, 54, 56, 63]. Recent", + "bbox": [ + 511, + 523, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5928", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4067950a35b01263b3e660b2b4fc16218bdce22a67eca1d67abe820a53df3b98.jpg", + "image_caption": [ + "Figure 2. The input 3DGS object is first converted to UVGS maps through spherical mapping. We use a multibranch forward mapping network to convert the obtained 14-channel UVGS to a compact 3-channel Super UVGS image. This represents the 3DGS object in a structured manner and can be used with image foundation models for reconstruction or generation. The Super UVGS is mapped back to UVGS through branched inverse mapping, which in turn can be reconstructed back to the 3DGS object through inverse spherical mapping." + ], + "image_footnote": [], + "bbox": [ + 122, + 88, + 872, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "advances in the 3D generative models for asset synthesis using the existing geometries like NeRF, voxel grids, or triplane geometries [2, 12, 42, 52, 56, 63] leverage generative models [15, 36] and the existing 3D datasets [4, 9]. However, most of the works employing 3DGS or other representations use multiview rendering and Score Distillation Sampling (SDS) to achieve convincing generation and reconstruction capabilities [7, 31]. These approaches demand high memory and compute resources and are often quite slow in optimizing due to per scene optimization.", + "bbox": [ + 88, + 397, + 480, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Giving Structures to Discrete Gaussians: Although, 3DGS has led to breakthrough in the reconstruction field by demonstrating superior performance in multiple domains, the generation of 3DGS directly remains challenging due to its discreteness and unstructured nature [55, 61]. These characteristics present substantial challenges when integrating them with conventional computer vision models, like Autoencoders and generative models [61]. The research in direct learning of trained 3DGS primitives is largely unexplored [23]. Some efforts attempt to address this by directly predicting 3DGS attributes using diffusion models [28] while others like Splatter Image [40] project Gaussian objects into image-based representations through direct 3D-unaware projection. These methods struggle with maintaining multiview consistency, as the model only infers seen poses correctly, while hallucinating for unseen poses.", + "bbox": [ + 88, + 550, + 482, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concurrent works [14, 55] follow the voxel-based representations to transport Gaussians into structural voxel grids with volume generation models for generating Gaussians. However, these methods are computationally expensive for high-resolution voxels, face difficulties in preserving high-quality Gaussian reconstructions due to information loss during voxelization. DiffGS [61] tries to solve the above", + "bbox": [ + 88, + 795, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "issues by proposing three continuous functions to represent 3DGS. However, it is limited to only category-level generation and learning generic probability functions for all the categories poses significant compute and design challenges. In contrary, we introduce an efficient way to give structures to discrete Gaussians by taking inspiration from the developments in 3D graphics. Our method does not require any learning to map an unstructured set of Gaussians to this efficient and structured representation (termed UVGS). The proposed representation provides local and global correspondence among different Gaussian points making the widely available existing computer vision frameworks learn and extract underlying features from them.", + "bbox": [ + 511, + 397, + 906, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 601, + 648, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preliminaries: 3DGS represents an object or a scene with a collection of Gaussians primitives to model the geometry and view-dependent appearance. For a 3DGS set, $G = \\{g_i\\}_{i=1}^N$ , representing an object with $N$ individual Gaussians, the geometry of the $i^{th}$ Gaussian is explicitly parameterized via 3D covariance matrix $\\Sigma_i$ and it's center $\\sigma_i \\in \\mathbb{R}^3$ as: $g_i(x) = e^{(-\\frac{1}{2}(x - \\sigma_i)^T\\Sigma^{-1}(x - \\sigma_i))}$ where, the covariance matrix $\\Sigma_i = r_is_is_i^T r_i^T$ is factorized into a rotation matrix $r_i \\in \\mathbb{R}^4$ and a scale matrix $s_i \\in \\mathbb{R}^3$ . The appearance of the $i - th$ Gaussian is represented by a color value $c_i \\in \\mathbb{R}^3$ and an opacity value $o_i \\in R$ . In practice, the color is represented by a series of Spherical Harmonics (SH) coefficients, but for simplicity, we represent the view-independent color by just RGB values. Thus, a single Gaussian can be represented by a set of five attributes as $g_i = \\{\\sigma_i, r_i, s_i, o_i, c_i\\} \\in \\mathbb{R}^{14}$ , and the entire 3DGS can be represented by a set of $N$ such Gaussians as: $G = \\{\\{\\sigma_i, r_i, s_i, o_i, c_i\\}\\}_{i=1}^N$ .", + "bbox": [ + 511, + 627, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5929", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Spherical Mapping", + "text_level": 1, + "bbox": [ + 89, + 90, + 274, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3DGS is represented as a permutation invariant set with no structural correspondence among different Gaussians $g_{i}$ , making it challenging to extract meaningful features from this set containing a few hundred thousands of them using neural networks. To address this, we introduce a novel representation that gives structure to this unstructured set of points and solves the permutation invariance issue for faster and better feature extraction. We propose to accomplish this by employing spherical mapping to map the 3DGS primitives to an image-like representation that is both invariant to random shuffling of 3DGS points and well structured.", + "bbox": [ + 89, + 113, + 483, + 279 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We begin the mapping by inscribing the 3DGS object into a sphere with the same center as the object in the canonical space. Inscribing a 3DGS object into a sphere involves enclosing the object within a sphere. This begins by determining the geometric center of the object. The next step is to calculate the radius of the sphere, which is achieved by measuring the Euclidean distance from the center to the farthest point on the object. The radius of the sphere is defined such that the sphere fully encloses the object. The sphere acts as a bounding volume for the entire object.", + "bbox": [ + 89, + 280, + 483, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We consider each Gaussian $g_{i}$ in 3D to be centered at the mean position represented by $\\sigma_{i}$ with Cartesian coordinates $(x_{i},y_{i},z_{i})$ . The aim is to get the spherical coordinates $(\\rho_{i},\\theta_{i},\\phi_{i})$ for each Gaussian $g_{i}$ . To do so, we calculate the azimuthal $\\theta_{i}$ and polar $\\phi_{i}$ angles for each $g_{i}$ along with the distance from the origin to the point, $\\rho_{i}$ . The spherical radius is defined as $\\rho_{i} = \\sqrt{x_{i}^{2} + y_{i}^{2} + z_{i}^{2}}$ , the azimuthal angle as $\\theta_{i} = \\tan^{-1}(y_{i},x_{i})$ , while the polar angle as $\\phi_{i} = \\cos^{-1}(z_{i},\\rho_{i})$ . The azimuthal and polar angles are then normalized, such that we can map them on a 2D UV map of $M\\times N$ dimensionality with 14-channels. $\\theta_{i}$ and $\\phi_{i}$ are converted to degrees and mapped to UV image coordinates: $\\theta_{i\\mathrm{scaled}} = \\left\\lfloor \\frac{\\pi + \\theta_i}{2\\pi}\\times \\mathrm{M}\\right\\rfloor$ , $\\phi_{i\\mathrm{scaled}} = \\left\\lfloor \\frac{\\phi_i}{\\pi}\\times \\mathrm{N}\\right\\rfloor$ Each channel in the UV map stores 3DGS attributes, including $\\{\\sigma_i,r_i,s_i,o_i,c_i\\} \\in \\mathbb{R}^{14}$ . We refer this 14-channel UV map as UVGS, $U\\in \\mathbb{R}^{M\\times N\\times 14}$ defined as:", + "bbox": [ + 89, + 431, + 483, + 680 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {U} \\left[ \\phi_ {i \\text {s c a l e d}}, \\theta_ {i \\text {s c a l e d}},: \\right] = \\left[ \\sigma_ {i}, r _ {i}, s _ {i}, o _ {i}, c _ {i} \\right]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 684, + 480, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This transformed UVGS representation provides spatial coherence and solves the permutation invariance problem as any random arrangement of points will now map to the same UVGS representation $U$ . It should be noted that this kind of transformation will also preserve the spatial correlation between the Gaussian points in 3D and transform them to 2D UV maps by mapping them to neighboring pixels. This provides both the local level correspondence among the neighboring Gaussians and the overall global correspondence for the object. Thus, solving the unstructured and discreteness problems. This enables standard neural network architectures (e.g. CNNs) to effectively capture correlations among neighboring Gaussians for efficient feature extraction.", + "bbox": [ + 89, + 704, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {g} _ {1}: \\{\\sigma_ {1}, r _ {1}, s _ {1}, o _ {1}, c _ {1} \\} \\\\ \\mathbf {g} _ {2}: \\{\\sigma_ {2}, r _ {2}, s _ {2}, o _ {2}, c _ {2} \\} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 92, + 697, + 135 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5f38a1646a02c3cfbc429eb39ddf0d13ec479cfd6b803ce2154472531c1b93ff.jpg", + "image_caption": [ + "Figure 3. Dynamic Selection. In spherical mapping of 3DGS points to UV maps, multiple points may map to the same pixel, creating a many-to-one issue. Our Dynamic Selection approach addresses this by retaining the attributes of the point with the highest opacity per pixel on the same ray." + ], + "image_footnote": [], + "bbox": [ + 557, + 135, + 870, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further unify the extracted position $(\\sigma)$ , transformation $(r, s)$ , and color $(c, o)$ maps in a same feature space and use the existing image foundation models, we map the obtained UVGS $U \\in \\mathbb{R}^{M \\times N \\times 14}$ further to a 3-channel image $S \\in \\mathbb{R}^{M \\times N \\times 3}$ (termed as Super UVGS), using a Convolutional Neural Network (CNN). A multi-branch forward mapping network is employed to map $U \\in \\mathbb{R}^{M \\times N \\times 14}$ to the 3-channel Super UVGS $S \\in \\mathbb{R}^{M \\times N \\times 3}$ . We provide all technical details in Sec. 3.3 and the supplementary material.", + "bbox": [ + 511, + 303, + 905, + 438 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Super UVGS representation effectively retains all the details of 3DGS attributes and can be directly utilized with existing widely available image-based models. We demonstrate this by showing perfect reconstruction of 3DGS object from Super UVGS image in the experiments section. This semantically structured representation $S$ offers both local and global correspondence in representing Gaussian attributes and needs relatively less storage.", + "bbox": [ + 511, + 439, + 905, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Dynamic GS Selection and Multiple Layers", + "text_level": 1, + "bbox": [ + 511, + 561, + 880, + 577 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When projecting 3DGS points to UV maps using spherical mapping, multiple points may map to the same pixel in UV space as shown in Fig. 3. Two 3DGS points $(g_{1})$ and $(g_{2})$ map to the same pixel on UV map $(P_{a})$ causing many-to-one mapping issue. To address this, we propose a Dynamic Selection approach where each UV pixel retains the 3DGS attribute with the highest opacity intersecting the same ray. Using the same example in Fig. 3, if opacity $o_{1}$ of Gaussian $g_{1}$ is less than opacity $o_{2}$ of $g_{2}$ . Then only $g_{2}$ will be stored in the UV map at pixel $P_{a}$ . We observed that this method helps maintain the geometry and appearance of the 3DGS object while resolving many-to-one mapping issues with minimal quality loss. For more complex objects or real-world scene representation, we stack multiple such layers of UV maps, where each UVGS pixel now holds attributes of the top-K opacity values of 3DGS primitives. This can be accomplished by inscribing the 3DGS object inside multiple spheres where each sphere maps the 3DGS attribute corresponding to the top- $K^{th}$ opacity value along the same ray. More details on this are presented in the supplementary. To show the effectiveness of proposed UVGS", + "bbox": [ + 511, + 583, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5930", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/25f85b471d8849db0732387367ab55d342304451b294b2457187827ee7586f1c.jpg", + "image_caption": [ + "Figure 4. Complex object reconstructions $(\\mathrm{K} = 4)$ using pretrained image-based autoencoder." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 475, + 157 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "maps in capturing the intricacies of a complex objects, we used a pretrained image based autoencoder to reconstruct objects using a 4 layer UVGS as shown in Fig. 4.", + "bbox": [ + 89, + 193, + 483, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Mapping Networks", + "text_level": 1, + "bbox": [ + 89, + 244, + 274, + 260 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our goal is to bring the extracted UVGS maps to a common feature space to better represent the object collectively and to make the 14-channel UVGS representations $U \\in \\mathbb{R}^{M \\times N \\times 14}$ work with the widely available image based foundation models. To accomplish this, we map it to a 3-channel image which can be easily processed by the existing architectures while also maintaining the spatial correspondence. We design a simple yet effective multi-branch CNN to extract features from different UVGS attributes and map them to a 3-channel feature-rich image, termed Super UVGS. The structured UVGS maps provide local and global features that can be learned by a CNN.", + "bbox": [ + 89, + 266, + 483, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Forward Mapping The first layer is a set of three mapping branches for position, transform, and appearance $(\\phi_P^f,\\phi_T^f,\\phi_A^f)$ respectively. We refer to them as position, transformation, and appearance branch. The position branch takes the mean position $(\\sigma)$ as an input and processes it to give a position feature map $M_P$ . Similarly, the transformation branch takes the rotation $(r)$ and scale $(s)$ together to generate another feature map $M_T$ . The last, appearance branch takes the color $(c)$ and opacity $(o)$ together to produce another feature map $M_A$ . All the three features maps from position, transformation, and appearance branch are concatenated to get a final feature map, before passing them to the next module, called the Central Branch. The central branch $(\\phi_C^f)$ is composed of multiple hidden Convolution layers, where each layer is followed by BatchNorm and ReLu activation. The last layer of the central branch is activated using tanh to ensure the Super UVGS does not take any ambiguous value resulting in gradient explosion or undesired artifacts. The obtained Super UVGS $S$ representation squeezes all the 3DGS attributes to a 3 dimensional image while also maintaining local and global structural correspondence among them.", + "bbox": [ + 89, + 448, + 483, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inverse Mapping We design an inverse mapping network that aims to map the obtained 3-channel Super UVGS image $S \\in \\mathbb{R}^{M \\times N \\times 3}$ back to the UVGS maps to obtain each of the five different 3DGS attributes $\\{\\sigma, r, s, o, c\\}$ . The inverse mapping network simply follows the forward mapping network architecture in the reverse order, where at first, we put the Central Branch $(\\phi_{iC})$ followed by attribute specific position, transformation, and appearance branches", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$(\\phi_{iP}, \\phi_{iT}, \\phi_{iA})$ . We provide more details on mapping network in the supplementary.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Branched mapping layers: The rationale behind using branched mapping layers in both forward and reverse mapping networks is to prevent the incompatibility issues arising due the the different value distribution of 3DGS attributes. Note that the disparate distributions of values within each set of attributes in 3D Gaussian Splatting, (i.e., mean position, transformation, and color), pose a challenge to the model when processed collectively. For instance, neighboring Gaussians in UVGS maps show smooth changes in position and color values but typically have large variations in rotation, scale, and opacity values. This results in gradient anomalies and slow convergence. To address this, we propose a multi-branch network architecture, where attribute-specific branches implicitly learn to process these distinct attribute specific properties, focusing on their unique features before passing them to the central branch. The central branch receives a concatenated stack of processed attributes and exploits the correlation between them by extracting local feature correspondences. This information is then mapped to a 3-channel Super UVGS image, effectively capturing the complex relationships between the various attributes. This approach enables our network to manage diverse attribute distributions, resulting in faster convergence, improved accuracy, and specialized processing for each attribute set.", + "bbox": [ + 511, + 123, + 906, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reconstruction Losses: Since the obtained UVGS maps have both local and global features, we opted for image-based losses to train the overall architecture. We use a set of Mean Squared Error (MSE) and Learned Perceptual Image Patch Similarity (LPIPS) [59] between the obtained UVGS from spherical mapping $U$ and the predicted UVGS $\\hat{U}$ from inverse mapping network. We calculate the LPIPS loss over four attributes of 3DGS including mean position $(\\sigma)$ , view independent color $(c)$ , scale $(s)$ , and rotation $(r)$ . The overall LPIPS loss for UV maps can be written as a linear sum of individual attribute loss terms as:", + "bbox": [ + 511, + 500, + 908, + 666 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {U V - l p i p s} = \\mathcal {L} _ {\\sigma} + \\mathcal {L} _ {s} + \\mathcal {L} _ {r} + \\mathcal {L} _ {c} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 670, + 903, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall loss function for the training can be written as: $\\mathcal{L}_{uvgs} = \\mathcal{L}_{mse} + \\lambda \\mathcal{L}_{UV - lpip}$ where $\\lambda$ is a scalar and varied from 0 to 10 during the course of training.", + "bbox": [ + 511, + 689, + 906, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 744, + 645, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3DGS Dataset and UV maps To train the mapping networks and learn a latent space for unconditional and conditional sampling, we need large amount of 3DGS assets. However, there's a lack of such a large-scale dataset for high quality 3DGS assets. To this end, we create a custom large scale dataset by converting the Objaverse [9] meshes into 3DGS representation1. We start by designing a scene of 88", + "bbox": [ + 511, + 767, + 906, + 875 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "Sketchfab data was filtered out of training data due to its license", + "bbox": [ + 529, + 886, + 877, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5931", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1dc5a626ac086a2d02768f0f41c01fbb14266a9fce609b793b8e31f33860ae36.jpg", + "image_caption": [ + "Super UVGS" + ], + "image_footnote": [], + "bbox": [ + 142, + 102, + 218, + 161 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8beda8f23a9bc3c1caaf1bfd49d52e221b90ce4248c166c1d85c074f9b5a33cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 101, + 316, + 160 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b0e1ce15fd53c2e7704c1288b2eef0719f0f1be6a21c0ff1db16d797276155a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 102, + 406, + 159 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6687e906b1b3018f2dfa5af95dd4d2668fa69ec32da4bf312be89ffca22f45ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 428, + 102, + 501, + 159 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/78177fa3c088aad27f8d98524cabc7c526ac1a360d7605e948527a561786e5a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 102, + 599, + 160 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/04d8c3195569f5eaf9a5bf9380fc8b41626cf0a6543a7559151b9ddb142dbc38.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 102, + 691, + 159 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5841673b1bd05609cbc173e2a859e8f1c93a0b5f13b07bf4384fe334df246596.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 102, + 785, + 159 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/65a1810f7586622267b5ec4b77f012e35bd7d97eae625847b5079e7ee338c6b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 807, + 102, + 880, + 160 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ab61707530cf3160e9cd8c366d8970684b15a78a2237bcc6bb9e8c03cf0e7182.jpg", + "image_caption": [ + "Multiview" + ], + "image_footnote": [], + "bbox": [ + 147, + 164, + 215, + 228 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1f4b60a82e998ae541fa551d8a0d44507e52b2229b5ae96487925ef701f3cb02.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 167, + 310, + 232 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/aa3f24f3951e68a4bb79b96328cd391d2f1797d7587555dae6eb6b71d856820b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 169, + 413, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2181461956050c1a50973f4c3c7a8bc54201dc66b3c66e8a442eb0da9ccbbfaa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 169, + 500, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8f4e3156691618f61e31cbb4ef8567fe1a0b27783896476258b0963e08533cc4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 169, + 606, + 233 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/42aacd4f4ded688c31817634696bdc8790fba9fa44403be745688be3057e109d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 178, + 704, + 228 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/09c5b76f22ca499839949af1195aecc4162f0f908164769d3ab94b8b557afabc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 178, + 803, + 244 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bf976d48495428f2ad3081772460cb5594a7770ccaf78138b6de932531834cd9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 803, + 183, + 880, + 247 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eb16a8e08262dd945922a110d24995e7017aaf8a4fa7a0b3c0c8362dd3ca364f.jpg", + "image_caption": [ + "Figure 5. Figure shows a wide variety of high-quality unconditional generation result from our method using Latent Diffusion Model. We train an LDM to randomly sample Super UVGS images from random noise. The Super UVGS can be converted to 3DGS object using inverse mapping network and inverse spherical projection. The unconditional generation model was trained on Objaverse dataset." + ], + "image_footnote": [], + "bbox": [ + 143, + 239, + 218, + 294 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1151107cb1adec15d4e96ec0eb05926ce551844693514ade22960aacac1ea5bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 238, + 320, + 300 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e0578a159631be7f346e7a0101212201fc0a9a8aeb956047446db1297f51178b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 325, + 222, + 413, + 300 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8e08ccbeda13d769a5a7ec266201c75e0025ce62751f7faae2ae650511ce8a8b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 241, + 509, + 300 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a85e5dc621c7b3ae6c8421f1307599da18a1c0ce3ce54339efb51cc5c47db559.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 234, + 607, + 297 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5c8310cfeb5762da162f009b7afca5bb3e703d79674a612be2e209717a2fbcbb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 236, + 699, + 299 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2b0e4f7e92927991c7af1e47af366ca08933e7d807c5c4d18424d48e3cdb5961.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 246, + 784, + 301 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b9ec333bcceb0e99ac1946078b82d4e27cbc23874bb8235162c33d0ea7ab1f7d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 799, + 258, + 880, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cameras in a canonical space and use it to capture Objaverse objects from various angles covering all of the object views. The 88 rendered views from different angles are then used to train a 3DGS for $10\\mathrm{K}$ iterations using [17]. This way, we create a high-quality and large-scale 3DGS dataset of $\\sim 400K$ objects and scenes from Objaverse. We only use static scenes or objects from Objaverse. After fitting all the object to 3DGS representation, we convert the objects to the corresponding UV maps (i.e., UVGS) through Spherical Mapping as illustrated in Fig. 2. For the course of our experiments, we only map the objects to a single layer UV maps as it was sufficient to represent the general purpose Objaverse objects with minimal quality loss. Through mapping, we gathered a UVGS dataset of $\\sim 400K$ maps. We fix the size of the UVGS maps to $512\\times 512$ . Through our experiments, we found that UV maps of size $512\\times 512$ are sufficient to represent objects in our dataset and capable of storing upto $262K$ unique Gaussians. Table 1 compares 1-4 layer UV maps. We also did experiments on ShapeNet [4] cars dataset for evaluation purposes.", + "bbox": [ + 91, + 361, + 483, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines & Metrics: To evaluate the quality of reconstructed 3DGS objects from both Super UVGS image and Autoencoder latent space to 3D, we use PSNR and LPIPS [58]. The aim is to convert the given 3DGS object to UVGS, and then to Super UVGS, and further to Autoencoder's latent space and calculate the metrics from the reconstructions at every step to prove the proposed method doesn't significantly affect the quality of reconstructions, while also providing a structurally meaningful representation that is much compact and easier to use with existing image based models. We compare the generational capabilities of our method against various conditional and unconditional SOTA 3D object generation method including the ones using multiview rendering for optimization DiffTF [2], Get3D [12], methods trying to give structural representa", + "bbox": [ + 89, + 672, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4b7900656acf03689a915809fae75b179603d970107ca3592b479e8d6eb21817.jpg", + "table_caption": [ + "Table 1. PSNR and LPIPS comparison for various reconstruction methods using UVGS and Super UVGS representations on Objverse Cars and Full datasets. AE, VAE, VQVAE are pretrained image based models. $K$ is the number of UVGS layers used. We also report the compression $\\%$ (CP) compared to the fitted 3DGS." + ], + "table_footnote": [], + "table_body": "
MethodPSNR(C/F)LPIPS(C/F)CP(%)
3DGS34.6 / 34.20.02 / 0.020
UVGS (@K=1)31.3 / 31.10.06 / 0.0653.0
UVGS (@K=2)32.8 / 31.90.04 / 0.0545.6
UVGS (@K=4)34.2 / 33.20.02 / 0.0333.3
Super UVGS (@K=1)31.2 / 31.10.07 / 0.0889.7
AE (@K=1)30.9 / 30.80.07 / 0.0999.5
VAE (@K=1)30.6 / 30.90.07 / 0.0999.5
VQVAE (@K=1)30.3 / 30.10.08 / 0.1099.7
", + "bbox": [ + 550, + 440, + 864, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tion to Gaussians, GaussianCube [55], and general purpose SOTA large 3D content generation models like DreamGaussian [42], LGM [44], and EG3D [3]. We also compare the quality of our generation results using FID and KID.", + "bbox": [ + 511, + 590, + 906, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mapping Network Training Details We train the forward and inverse mapping networks to project the obtained UVGS maps $U \\in \\mathbb{R}^{M \\times N \\times 14}$ to Super UVGS image $S \\in \\mathbb{R}^{M \\times N \\times 3}$ , and back to the reconstructed UVGS maps $\\hat{U} \\in \\mathbb{R}^{M \\times N \\times 14}$ . We provide an in-depth discussion of all implementation details in the supplementary material.", + "bbox": [ + 511, + 652, + 906, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. UVGS AutoEncoder and 3DGS Compression", + "text_level": 1, + "bbox": [ + 511, + 755, + 893, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The obtained Super UVGS image is a structurally meaningful representation that can have various applications in the generation and reconstruction of new 3D assets as it contains features that can be learned by the existing image based models. Through our experiments, we show that a 3-channel Super UVGS image can be directly reconstructed using a pretrained image based Autoencoders or VAEs without any fine-tuning. We tested on three different models in", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5932", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/83c9eb6c0339a44fcc598dba59764b39bdde7c43b8a06f7adb4694b8aa648149.jpg", + "image_caption": [ + "DiffTF" + ], + "image_footnote": [], + "bbox": [ + 124, + 126, + 197, + 164 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f8b27dba3749b293c7fcd612abc93e6a5692fdd722615c5523b2220b3661bf2b.jpg", + "image_caption": [ + "Figure 6. Comparison of unconditional 3D asset generation on the cars category with SOTA methods. Figure shows that DiffTF [2] produces low-quality, low-resolution cars lacking detail. While Get3D [12] achieve higher resolution, it suffers from 3D inconsistency, numerous artifacts, and lacks 3D detail. Similar issues are found in GaussianCube [55] along with symmetric inconsistency in the results. In contrast, our method generates high-quality, high-resolution objects that are 3D consistent with sharp and well-defined edges." + ], + "image_footnote": [], + "bbox": [ + 122, + 176, + 197, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5870965b0a011f348bd6406b300281c44a52ef0e9604240a6a8abb734b80b0af.jpg", + "image_caption": [ + "Get3D" + ], + "image_footnote": [], + "bbox": [ + 253, + 114, + 308, + 165 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/10cae7b5547d1e94a4698c66306f6605fb00ce2bdd7d2bf89132787d6e247fce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 250, + 176, + 307, + 213 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eb419d82e0e5c06e732353251b34b4192c4150959452a4d4f001551c659b0e60.jpg", + "image_caption": [ + "GaussianCube" + ], + "image_footnote": [], + "bbox": [ + 313, + 113, + 410, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/98bd3dec6d91c19e4fa3ea1f3ead282225d12a61228adc786a1b1976574945a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 171, + 410, + 219 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8504566ec2d121a5832dae4248bf1c9e5c94d3ef4a5b3bfc24d64b1912390500.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 446, + 128, + 535, + 171 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4924ea712094fd45596345baf08169c651010493e4047c2eb7c7d6a6db3ccd75.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 126, + 596, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/239ff5dad6d501015f121e0ed1457f0d82ddc7d3cbc76c5cf93b672da90b7dce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 185, + 537, + 223 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/62cbbd6a34134c889925976552340213dabdc4fa60cf4a94096a2a5e34ed3a0a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 127, + 651, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8425730ae76e0693e63bf58c18b9bf1988ac9693a8bc038f68565d3062b40028.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 570, + 184, + 651, + 223 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/372dcede4f068eee9c86a879d1329201caa180f78c6b746fa0fc5f7ee45f60f2.jpg", + "image_caption": [ + "UVGS (ours)" + ], + "image_footnote": [], + "bbox": [ + 684, + 125, + 797, + 164 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7088103b939532b32cad1d42762c80488345d78ea510f63ecb785ebfdf45d5ab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 802, + 117, + 875, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/59fdcae91097b03e80b388e4c84b9c9f72c98e6d66420a19d65c733754a7a622.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 683, + 180, + 797, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0a70511b2ed037342fd7e43ca5bd700ca854e7feaa68a04ff3b8e71f47e76d70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 803, + 178, + 875, + 212 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9ccc93f755797d2ed09bbc5b39ddd31e850e21d441d4d40aeaf8236c971807d8.jpg", + "image_caption": [ + "Figure 7. We compare the performance of our model against various SOTA methods for text-conditional object synthesis. Our method not only generates high-quality assets for simpler objects, but also for complicated objects with intricate geometry." + ], + "image_footnote": [], + "bbox": [ + 91, + 295, + 483, + 483 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cluding image AE, KL-VAE [18], VQVAE [46] and each performed quite well without any significant quality loss. The reconstruction PSNR and LPIPS values are presented in Table 1. This means we can now leverage the powerful compression capabilities of image based Autoencoders to compress the storage requirements of 3DGS by more than $99\\%$ . We have shown the storage comparison results in Table 1. It is interesting to note that the Super UVGS representation itself can be used to compress the memory requirement for storing 3DGS object by up to $89.7\\%$ .", + "bbox": [ + 89, + 555, + 482, + 707 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Unconditional & Conditional Generation", + "text_level": 1, + "bbox": [ + 89, + 712, + 441, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We aim to show the effectiveness of Super UVGS representation for directly generating 3DGS objects from a learned latent space. We consider Super UVGS images as a compact and structured proxy for representing 3DGS objects as it maintains the 3D object information while also providing learnable features. The existing methods fail to directly generate a large number of Gaussians (e.g. $100\\mathrm{K}+$ ) to represent objects with sufficient quality either due to the lack of 3D generative model architectures that support such large number of unstructured points, or due to the lack of a large 3DGS dataset [28, 40, 44, 55, 63]. We leverage the", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Super UVGS representation and use the existing 2D image generative models like Diffusion Models [15] for this task. Specifically, to train a generative model capable of randomly sampling new high-quality 3DGS assets for various downstream tasks, we use an unconditional Latent Diffusion Model (LDM) [36] on the obtained Super UVGS images. As illustrated in Section 4.1, we can use a pretrained image VAE to map the Super UVGS image to a latent space and reconstruct back. Hence, we only train a LDM on the latent space. More implementation details are provided in the supplementary.", + "bbox": [ + 511, + 297, + 906, + 463 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Unconditional LDM: To design a generative model capable of randomly sampling new high-quality 3DGS assets for various downstream tasks, we train an unconditional LDM [36] on the learned Super UVGS images. Following [33, 36, 57], we use DDIM [15] for faster and consistent sampling with up to 1000 time steps used in the forward diffusion process, and 20 during denoising. Once trained, the model is used to randomly sample Super UVGS images, resulting in high quality 3DGS assets through inverse mapping. Results are presented in Fig 5. We demonstrate that our method inherently learns to generate multiview consistent images due to the powerful Super UVGS representation unlike most prior works using rendering-based losses.", + "bbox": [ + 511, + 468, + 908, + 664 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Conditional LDM: Similar to unconditional generation, we also trained a text-conditioned LDM following the SD's [33, 36, 57] pipeline and using the predicted text for our dataset. The trained model can be used to generate high-quality text-conditioned 3DGS assets that are multiview consistent. The results are demonstrated in Fig 6.", + "bbox": [ + 511, + 669, + 908, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The above experiments prove the effectiveness of our proposed Super UVGS representation in 3D object synthesis using widely available 2D image models. It also highlights that this compact 3-channel Super UVGS representation stores not just the spatial correspondence among different pixels, but also the rich 3D information of the objects. This way, we can easily convert a 3D asset generation problem into a 2D image generation problem without the use of any complex 3D architecture to handle large amount of un", + "bbox": [ + 511, + 763, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5933", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e5569757356944ee79466c8fb221ce34c48934ad1c6558ee3d05fe8959d08516.jpg", + "image_caption": [ + "Figure 8. 3DGS Inpainting: We present one of the first inpainting results on 3DGS directly leveraging the Super UVGS images and the denoising capabilities of diffusion models." + ], + "image_footnote": [], + "bbox": [ + 99, + 85, + 473, + 286 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "structured and permutation invariant 3DGS primitives, and neither relying upon computationally expensive multiview rendering or SDS loss. Baseline comparison for unconditional and conditional generation is presented in Table 2.", + "bbox": [ + 89, + 340, + 483, + 401 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. 3DGS Inpainting", + "text_level": 1, + "bbox": [ + 89, + 409, + 261, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Leveraging the powerful Super UVGS representation, we present in Fig. 8 one of the first experiments on inpainting 3DGS directly without using any multiview rendering or distilling information from diffusion models. We try to recover the missing Gaussians by leveraging the denoising capabilities of LDM and trying to predict the missing corresponding parts of the Super UVGS image. We believe, this can have potential applications in sparse view reconstruction. More details are given in the supplementary.", + "bbox": [ + 89, + 431, + 483, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Studies & Discussion", + "text_level": 1, + "bbox": [ + 89, + 575, + 359, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct exhaustive ablation studies to justify some of our framework's design choices including the effect of branching in mapping networks, the use of single layer UVGS maps, and the resolution of UVGS maps. We performed our experiments on our custom Objaverse [9] 3DGS dataset and evaluate the performance of our model in terms of PSNR, SSIM, and LPIPS. The results are presented in Table 3. From the table, it can be seen that by using four layers of UVGS maps $(\\mathrm{K} = 4)$ , we can almost match the reconstruction quality of fitted 3DGS results, while we realized that simply with a single layer UVGS, we are able to maintain the overall geometry and appearance of the object for our dataset with a PSNR of more than 30. We also compared the reconstruction performance of our method with and without using branching in the mapping network. It can be clearly seen that using branching network significantly increases the reconstruction quality from Super UVGS space. The main reasoning behind this is the specialization of attributes that the branching provides to individually process each attribute first.", + "bbox": [ + 89, + 598, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/29c2d3ce00e202e64898fc8220ef70edb87833dbc11c6e4a07603ab3ce8fa9d2.jpg", + "table_caption": [ + "Table 2. We compare the FID and KID of unconditional generation using the current SOTA methods on 20K randomly generated samples from each method and ours. We also compare our method against SOTA text-conditioned generation frameworks on CLIP Score for 10K generated objects from each method." + ], + "table_footnote": [], + "table_body": "
Unconditional GenerationText-Conditioned Generation
MethodFID ↓KID ↓MethodCLIP Score ↑
Get3D [12]53.174.19DreamGaussian [42]28.51
DiffTF [2]84.578.73Shap. E [6]30.53
EG3D [3]74.516.62LGM [44]30.74
GaussianCube34.673.72GaussianCube [55]30.34
UVGS (Ours)26.203.24UVGS (Ours)32.62
", + "bbox": [ + 542, + 155, + 874, + 263 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f4d28aafa8c82127eb685605ee87735199243e646870ee52ccb18cb580325036.jpg", + "table_caption": [ + "Table 3. We present quantitative ablation study for number of UVGS layers (K), UVGS map resolution, and the effect of branching in mapping network on the Objaverse 3DGS dataset." + ], + "table_footnote": [], + "table_body": "
MethodPSNRLPIPSUVGS SizePSNRLPIPS
UVGS @K=131.10.06512 × 512 (@K=1)31.10.08
UVGS @K=231.90.05256 × 256 (@K=1)28.20.23
UVGS @K=433.20.03Without Branching27.80.31
", + "bbox": [ + 542, + 316, + 874, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations & Future Work: While single layer UVGS images can recover the geometry of the object, they sometimes suffer in terms of appearance and the generated objects might look washed out. We believe this can be solved by using a multi-layer UVGS maps. Similarly, the single layer UVGS map is limited to representing simpler everyday objects, and may not be sufficient to represent highly-detailed and complex objects or scenes. In the future, we want to extend this framework to learn features for real-world scenes and complex objects like a human head with multi-layer UV mapping. We also want to make this representation more efficient by better utilizing the empty pixels of UVGS maps and Super UVGS images while also maintaining the underlying features and 3D information.", + "bbox": [ + 511, + 388, + 906, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 606, + 633, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced a novel method to solve the underlying issues with 3D Gaussian Splatting (3DGS) that prevent the direct integration of them with the large number of existing image foundational models. We proposed UVGS - a structured representation for 3DGS obtained by spherical mapping of 3DGS primitives to UV maps. We further squeezed the multi-attribute UVGS maps to a 3-channel unified and structured Super UVGS image, which not only maintains the 3D structural information of the object, but also provides a compact feature space for 3DGS attributes. The obtained Super UVGS images are directly integrated with the existing image foundational models for 3DGS compression and unconditional and conditional generation using diffusion models. Leveraging these Super UVGS images, we showed one of the first inpainting experiments on 3DGS.", + "bbox": [ + 511, + 625, + 906, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ACKNOWLEDGEMENTS A part of this work was supported by NSF CAREER grant 2143576 and ONR DURIP grant N00014-23-1-2804.", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5934", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855-5864, 2021. 2", + "[2] Ziang Cao, Fangzhou Hong, Tong Wu, Liang Pan, and Ziwei Liu. Large-vocabulary 3d diffusion model with transformer. arXiv preprint arXiv:2309.07920, 2023. 3, 6, 7, 8", + "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16123-16133, 2022. 2, 6, 8", + "[4] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 3, 6", + "[5] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision, pages 333-350. Springer, 2022. 2", + "[6] Minghao Chen, Junyu Xie, Iro Laina, and Andrea Vedaldi. Shap-editor: Instruction-guided latent 3d editing in seconds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26456-26466, 2024. 8", + "[7] Zilong Chen, Feng Wang, Yikai Wang, and Huaping Liu. Text-to-3d using gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21401-21412, 2024. 3", + "[8] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4456-4465, 2023. 2", + "[9] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli Vanderbilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 3, 5, 8", + "[10] Slava Elizarov, Ciara Rowles, and Simon Donne. Geometry image diffusion: Fast and data-efficient text-to-3d with image-based surface representation. arXiv preprint arXiv:2409.03718, 2024. 2", + "[11] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. 2" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841-31854, 2022. 3, 6, 7, 8", + "[13] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371, 2023. 2", + "[14] Xianglong He, Junyi Chen, Sida Peng, Di Huang, Yangguang Li, Xiaoshui Huang, Chun Yuan, Wanli Ouyang, and Tong He. Gvgen: Text-to-3d generation with volumetric representation. In European Conference on Computer Vision, pages 463–479. Springer, 2025. 2, 3", + "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3, 7", + "[16] Han Huang, Yulun Wu, Junsheng Zhou, Ge Gao, Ming Gu, and Yu-Shen Liu. Neusurf: On-surface priors for neural surface reconstruction from sparse input views. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2312-2320, 2024. 2", + "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2, 6", + "[18] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 7", + "[19] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 2", + "[20] Shujuan Li, Junsheng Zhou, Baorui Ma, Yu-Shen Liu, and Zhizhong Han. Neaf: Learning neural angle fields for point normal estimation. In Proceedings of the AAAI conference on artificial intelligence, pages 1396-1404, 2023. 2", + "[21] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 2", + "[22] Baorui Ma, Haoge Deng, Junsheng Zhou, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Geodream: Disentangling 2d and geometric priors for high-fidelity and consistent 3d generation. arXiv preprint arXiv:2311.17971, 2023. 2", + "[23] Qi Ma, Yue Li, Bin Ren, Nicu Sebe, Ender Konukoglu, Theo Gevers, Luc Van Gool, and Danda Pani Paudel. Shapes-plat: A large-scale dataset of gaussian splats and their self-supervised pretraining. arXiv preprint arXiv:2408.10906, 2024. 2, 3", + "[24] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 2" + ], + "bbox": [ + 93, + 92, + 905, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5935", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2", + "[26] Chaerin Min and Srinath Sridhar. Genheld: Generating and editing handheld objects. arXiv preprint arXiv:2406.05059, 2024. 2", + "[27] Chaerin Min, Sehyun Cha, Changhee Won, and Jongwoo Lim. Tsdf-sampling: Efficient sampling for neural surface field using truncated signed distance field. arXiv preprint arXiv:2311.17878, 2023. 2", + "[28] Yuxuan Mu, Xinxin Zuo, Chuan Guo, Yilin Wang, Juwei Lu, Xiaofeng Wu, Songcen Xu, Peng Dai, Youliang Yan, and Li Cheng. Gsd: View-guided gaussian splatting diffusion for 3d reconstruction. arXiv preprint arXiv:2407.04237, 2024. 2, 3, 7", + "[29] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulo, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4328-4338, 2023. 2", + "[30] Francesco Palandra, Andrea Sanchietti, Daniele Baieri, and Emanuele Rodola. Gsedit: Efficient text-guided editing of 3d objects via gaussian splatting. arXiv preprint arXiv:2403.05154, 2024. 2", + "[31] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 3", + "[32] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 2", + "[33] Aashish Rai and Srinath Sridhar. Egosonics: Generating synchronized audio for silent egocentric videos. arXiv preprint arXiv:2407.20592, 2024. 7", + "[34] Aashish Rai, Hiresh Gupta, Ayush Pandey, Francisco Vicente Carrasco, Shingo Jason Takagi, Amaury Aubel, Daeil Kim, Aayush Prakash, and Fernando De la Torre. Towards realistic generative 3d face models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3738-3748, 2024. 2", + "[35] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject-driven text-to-3d generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2349–2359, 2023. 2", + "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 7", + "[37] Li Shen and Fillia Makedon. Spherical mapping for processing of 3d closed surfaces. Image and vision computing, 24 (7):743-761, 2006. 2" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20875-20886, 2023. 2", + "[39] Jingxiang Sun, Bo Zhang, Ruizhi Shao, Lizhen Wang, Wen Liu, Zhenda Xie, and Yebin Liu. Dreamcraft3d: Hierarchical 3d generation with bootstrapped diffusion prior. arXiv preprint arXiv:2310.16818, 2023. 2", + "[40] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10208-10217, 2024. 2, 3, 7", + "[41] Fariborz Taherkhani, Aashish Rai, Quankai Gao, Shaunak Srivastava, Xuanbai Chen, Fernando De la Torre, Steven Song, Aayush Prakash, and Daeil Kim. Controllable 3d generative adversarial face model via disentangling shape and appearance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 826-836, 2023. 2", + "[42] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 2, 3, 6, 8", + "[43] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22819–22829, 2023. 2", + "[44] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1-18. Springer, 2025. 6, 7, 8", + "[45] Zhicong Tang, Shuyang Gu, Chunyu Wang, Ting Zhang, Jianmin Bao, Dong Chen, and Baining Guo. Volumediffusion: Flexible text-to-3d generation with efficient volumetric encoder. arXiv preprint arXiv:2312.11459, 2023. 2", + "[46] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 7", + "[47] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4563-4573, 2023. 2", + "[48] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3295–3306, 2023. 2", + "[49] Yuxuan Wang, Xuanyu Yi, Zike Wu, Na Zhao, Long Chen, and Hanwang Zhang. View-consistent 3d editing with gaussian splatting. In European Conference on Computer Vision, pages 404-420. Springer, 2025. 2" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5936", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[51] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024. 2", + "[52] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation. arXiv preprint arXiv:2403.14621, 2024. 2, 3", + "[53] Xingguang Yan, Han-Hung Lee, Ziyu Wan, and Angel X Chang. An object is worth 64x64 pixels: Generating 3d object via image diffusion. arXiv preprint arXiv:2408.03178, 2024. 2", + "[54] Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian-dreamer: Fast generation from text to 3d gaussian splattering with point cloud priors. arXiv preprint arXiv:2310.08529, 2023. 2", + "[55] Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: Structuring gaussian splatting using optimal transport for 3d generative modeling. arXiv preprint arXiv:2403.19655, 2024. 2, 3, 6, 7, 8", + "[56] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 2, 3", + "[57] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 7", + "[58] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6", + "[59] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5", + "[60] Junsheng Zhou, Yu-Shen Liu, and Zhizhong Han. Zero-shot scene reconstruction from single images with deep prior assembly. arXiv preprint arXiv:2410.15971, 2024. 2", + "[61] Junsheng Zhou, Weiqi Zhang, and Yu-Shen Liu. Diff-gfs: Functional gaussian splatting diffusion. arXiv preprint arXiv:2410.19657, 2024. 2, 3", + "[62] Junsheng Zhou, Weiqi Zhang, Baorui Ma, Kanle Shi, YuShen Liu, and Zhizhong Han. Udiff: Generating conditional unsigned distance fields with optimal wavelet diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21496-21506, 2024. 2" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10324-10335, 2024. 2, 3, 7", + "[64] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10324-10335, 2024. 2" + ], + "bbox": [ + 516, + 90, + 906, + 260 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5937", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/UVGS_ Reimagining Unstructured 3D Gaussian Splatting using UV Mapping/77087a0b-d1cb-4230-a846-7760f7510d17_model.json b/2025/UVGS_ Reimagining Unstructured 3D Gaussian Splatting using UV Mapping/77087a0b-d1cb-4230-a846-7760f7510d17_model.json new file mode 100644 index 0000000000000000000000000000000000000000..78fad20e17d11dbf351cfa93d59b2ab9b2a221e2 --- /dev/null +++ b/2025/UVGS_ Reimagining Unstructured 3D Gaussian Splatting using UV Mapping/77087a0b-d1cb-4230-a846-7760f7510d17_model.json @@ -0,0 +1,2499 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.131, + 0.89, + 0.155 + ], + "angle": 0, + "content": "UVGS: Reimagining Unstructured 3D Gaussian Splitting using UV Mapping" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.179, + 0.243, + 0.199 + ], + "angle": 0, + "content": "Aashish Rai\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.279, + 0.18, + 0.512, + 0.217 + ], + "angle": 0, + "content": "Dilin Wang\\(^{2}\\) Mihir Jain\\(^{2}\\) Srinath Sridhar\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.181, + 0.721, + 0.218 + ], + "angle": 0, + "content": "Nikolaos Sarafianos2 \nAayush Prakash2" + }, + { + "type": "text", + "bbox": [ + 0.759, + 0.181, + 0.874, + 0.198 + ], + "angle": 0, + "content": "Kefan Chen\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.341, + 0.228, + 0.492, + 0.248 + ], + "angle": 0, + "content": "\\(^{1}\\)Brown University" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.229, + 0.678, + 0.247 + ], + "angle": 0, + "content": "Meta Reality Labs" + }, + { + "type": "text", + "bbox": [ + 0.331, + 0.258, + 0.662, + 0.273 + ], + "angle": 0, + "content": "https://aashishrai3799.github.io/uvgs" + }, + { + "type": "image", + "bbox": [ + 0.112, + 0.289, + 0.885, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.511, + 0.908, + 0.568 + ], + "angle": 0, + "content": "Figure 1. We propose UVGS - an structured image-like representation for 3DGS obtained by spherical mapping of its primitives. The obtained UVGS maps can be further unified to a 3-channel Super UVGS image to bridging the gap between 3DGS and existing image foundation models. We show Super UVGS can compress the 3DGS assets using pretrained image Autoencoders, and for direct unconditional and conditional 3DGS object generation using diffusion models. We also present one of the first experiments on 3DGS inpainting." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.575, + 0.327, + 0.59 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.599, + 0.486, + 0.902 + ], + "angle": 0, + "content": "3D Gaussian Splatting (3DGS) has demonstrated superior quality in modeling 3D objects and scenes. However, generating 3DGS remains challenging due to their discrete, unstructured, and permutation-invariant nature. In this work, we present a simple yet effective method to overcome these challenges. We utilize spherical mapping to transform 3DGS into a structured 2D representation, termed UVGS. UVGS can be viewed as multi-channel images, with feature dimensions as a concatenation of Gaussian attributes such as position, scale, color, opacity, and rotation. We further find that these heterogeneous features can be compressed into a lower-dimensional (e.g., 3-channel) shared feature space using a carefully designed multi-branch network. The compressed UVGS can be treated as typical RGB images. Remarkably, we discover that typical VAEs trained with latent diffusion models can directly generalize to this new representation without additional training. Our novel representation makes it effortless to leverage foundational 2D models, such as diffusion models, to directly model 3DGS. Additionally, one can simply increase the 2D UV resolution" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.576, + 0.907, + 0.697 + ], + "angle": 0, + "content": "to accommodate more Gaussians, making UVGS a scalable solution compared to typical 3D backbones. This approach immediately unlocks various novel generation applications of 3DGS by inherently utilizing the already developed superior 2D generation capabilities. In our experiments, we demonstrate various unconditional, conditional generation, and inpainting applications of 3DGS based on diffusion models, which were previously non-trivial." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.701, + 0.646, + 0.715 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The creation of high-quality 3D content is essential in applications like virtual reality, game design, robotics, and movie production, where realistic 3D representations play a critical role. Typical 3D representations like Neural Radiance Fields (NeRF) [25] are promising but require substantial computational resources, limiting their scalability for real-time applications. Moreover, NeRF is an implicit representation, which makes editing and manipulation challenging. Recently, 3D Gaussian Splatting (3DGS) [17] emerged as a compelling alternative, enabling efficient and high-fidelity 3D rendering through a large set of Gaussian primitives that model spatial and visual properties. As an explicit represen" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5927" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "tation, 3DGS offers several advantages over NeRF. However, while 3DGS offers benefits in terms of speed and visual quality, its unstructured, permutation-invariant nature presents significant challenges for generative tasks. Much like point clouds, it lacks a coherent spatial structure, impeding its integration with conventional image-based generative models. This lack of structure and coherence among primitives hinders the application of image-based generative models [23, 55, 61], which rely on structured data representations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.246, + 0.483, + 0.442 + ], + "angle": 0, + "content": "Previous methods have tackled these challenges by transforming 3DGS into structured formats, such as voxel grids [8, 14, 55] or image-based representations like Splatter Image [40] or triplanes [64]. Other approaches employ diffusion models to directly predict 3DGS attributes [28]. These methods, while achieving impressive visual results, often require substantial computational resources, memory-intensive multi-view rendering, complex architectures limiting their scalability and flexibility for high-fidelity generation. Generating and processing 3DGS directly by efficiently utilizing modern generative models like Variational Autoencoders (VAEs) and diffusion models is limited as the neural networks are not permutation invariant." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.483, + 0.656 + ], + "angle": 0, + "content": "To address these shortcomings, we introduce UV Gaussian Splatting (UVGS), which provides a structured transformation of 3D Gaussian primitives into a 2D representation while preserving essential 3D information. We use spherical mapping [37] that inscribes Gaussian splats in a spherical surface, and projects attributes like position, rotation, scale, opacity, and color into an organized 14-channel image-like UV map. This mapping introduces spatial structure, resolving issues of permutation invariance by introducing local correspondences between neighboring Gaussians and global coherence across the entire 3D object. The result is a representation that functions as a \"3D representation\" structured in a 2D map format, enabling compatibility with powerful image-based neural network architectures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.901 + ], + "angle": 0, + "content": "While UVGS introduces structure into 3D Gaussian Splitting, its full 14-channel attribute-specific representation presents challenges for direct integration with pretrained 2D generative models, as these models typically expect a simpler, image-compatible data. Each of its heterogeneous attributes—position, color, and transformation—has its own distinct distribution and resides in a separate feature space, making it challenging to represent the 3D object in a unified shared space. To address this, we introduce Super UVGS, a compact 3-channel representation that unifies these diverse attributes into a cohesive format. Using a carefully designed multi-branch mapping network, Super UVGS consolidates the distinct attribute spaces into a shared feature space, enabling a more collective representation of the object. This unified transformation not only facilitates zero-shot compatibility with pretrained 2D models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.242 + ], + "angle": 0, + "content": "but also optimizes memory usage and computational efficiency, making Super UVGS highly practical for large-scale 3D tasks. Unlike previous approaches that use Triplanes, voxels, occupancy grid, neural fields etc. and require specialized 3D architectures to train on 3D data, UVGS effortlessly leverages widely available pretrained 2D foundational models. This zero-shot generalization capability allows UVGS to fully benefit from priors learned in 2D domains from large amount of data, improving both flexibility and scalability. To sum up our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.244, + 0.905, + 0.332 + ], + "angle": 0, + "content": "- Efficient Structured Representation of 3DGS: We present UVGS, an image-like representation that solves permutation invariance and unstructured nature of discrete 3DGS through spherical mapping, making direct feature extraction possible by organizing unordered points into a coherent 2D representation compatible with 2D models." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.334, + 0.905, + 0.424 + ], + "angle": 0, + "content": "- Compact and Scalable Super UVGS Representation: To address scalability while dealing with large scale 3DGS points and enabling the direct integration of pre-trained 2D foundation models, we introduce Super UVGS - a low-dimensional version of UVGS maps that retains high fidelity features while reducing memory overhead." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.425, + 0.904, + 0.484 + ], + "angle": 0, + "content": "- Diverse 3D Applications: Our approach unlocks seamless integration of 3DGS with pre-trained 2D foundation models for various tasks, including unconditional and conditional generation of 3DGS." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.244, + 0.905, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.498, + 0.655, + 0.513 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.902 + ], + "angle": 0, + "content": "3D Generative/Reconstruction Models for Objects: Generation or reconstruction of 3D assets has been a long standing task [3, 21, 24, 26, 27, 31, 34, 35, 39, 41, 43, 50, 54]. Previous reconstruction approaches like NeRF [1, 25] are often slow and do not provide a defining geometry [5, 11, 16, 20, 22, 24, 32, 35, 48, 60, 62]. Advancements in the field led to the emergence of explicit voxel grid based representations that encode colors and opacities directly [8, 29, 45]. These approaches achieve significant speed ups compared to the NeRF based approaches, but they can't produce high fidelity assets due to the low resolution of voxel grids. On the other hand, triplane representation [13, 38, 47, 63] provides a trade-off between the quality and memory utilization. Another line of work [10, 53] splits the input mesh into different patches and simplifies the object generation problem to an image generation problem. However such methods rely on either cutting through the mesh to create a geometry image [10] or rely on an existing UV representation of the geometry and utilize subset of existing UV islands [53] resulting in loss of details. Recently, there has been a notable advancement in 3D Gaussian Splatting (3DGS) for the representation of objects and scenes leading to the emergence of 3DGS showcasing impressive real-time results in reconstruction and generation tasks [14, 17, 19, 30, 42, 49, 51, 52, 54, 56, 63]. Recent" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5928" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.089, + 0.873, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.333, + 0.908, + 0.391 + ], + "angle": 0, + "content": "Figure 2. The input 3DGS object is first converted to UVGS maps through spherical mapping. We use a multibranch forward mapping network to convert the obtained 14-channel UVGS to a compact 3-channel Super UVGS image. This represents the 3DGS object in a structured manner and can be used with image foundation models for reconstruction or generation. The Super UVGS is mapped back to UVGS through branched inverse mapping, which in turn can be reconstructed back to the 3DGS object through inverse spherical mapping." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.398, + 0.482, + 0.549 + ], + "angle": 0, + "content": "advances in the 3D generative models for asset synthesis using the existing geometries like NeRF, voxel grids, or triplane geometries [2, 12, 42, 52, 56, 63] leverage generative models [15, 36] and the existing 3D datasets [4, 9]. However, most of the works employing 3DGS or other representations use multiview rendering and Score Distillation Sampling (SDS) to achieve convincing generation and reconstruction capabilities [7, 31]. These approaches demand high memory and compute resources and are often quite slow in optimizing due to per scene optimization." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.551, + 0.483, + 0.793 + ], + "angle": 0, + "content": "Giving Structures to Discrete Gaussians: Although, 3DGS has led to breakthrough in the reconstruction field by demonstrating superior performance in multiple domains, the generation of 3DGS directly remains challenging due to its discreteness and unstructured nature [55, 61]. These characteristics present substantial challenges when integrating them with conventional computer vision models, like Autoencoders and generative models [61]. The research in direct learning of trained 3DGS primitives is largely unexplored [23]. Some efforts attempt to address this by directly predicting 3DGS attributes using diffusion models [28] while others like Splatter Image [40] project Gaussian objects into image-based representations through direct 3D-unaware projection. These methods struggle with maintaining multiview consistency, as the model only infers seen poses correctly, while hallucinating for unseen poses." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Concurrent works [14, 55] follow the voxel-based representations to transport Gaussians into structural voxel grids with volume generation models for generating Gaussians. However, these methods are computationally expensive for high-resolution voxels, face difficulties in preserving high-quality Gaussian reconstructions due to information loss during voxelization. DiffGS [61] tries to solve the above" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.907, + 0.595 + ], + "angle": 0, + "content": "issues by proposing three continuous functions to represent 3DGS. However, it is limited to only category-level generation and learning generic probability functions for all the categories poses significant compute and design challenges. In contrary, we introduce an efficient way to give structures to discrete Gaussians by taking inspiration from the developments in 3D graphics. Our method does not require any learning to map an unstructured set of Gaussians to this efficient and structured representation (termed UVGS). The proposed representation provides local and global correspondence among different Gaussian points making the widely available existing computer vision frameworks learn and extract underlying features from them." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.602, + 0.65, + 0.62 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Preliminaries: 3DGS represents an object or a scene with a collection of Gaussians primitives to model the geometry and view-dependent appearance. For a 3DGS set, \\( G = \\{g_i\\}_{i=1}^N \\), representing an object with \\( N \\) individual Gaussians, the geometry of the \\( i^{th} \\) Gaussian is explicitly parameterized via 3D covariance matrix \\( \\Sigma_i \\) and it's center \\( \\sigma_i \\in \\mathbb{R}^3 \\) as: \\( g_i(x) = e^{(-\\frac{1}{2}(x - \\sigma_i)^T\\Sigma^{-1}(x - \\sigma_i))} \\) where, the covariance matrix \\( \\Sigma_i = r_is_is_i^T r_i^T \\) is factorized into a rotation matrix \\( r_i \\in \\mathbb{R}^4 \\) and a scale matrix \\( s_i \\in \\mathbb{R}^3 \\). The appearance of the \\( i - th \\) Gaussian is represented by a color value \\( c_i \\in \\mathbb{R}^3 \\) and an opacity value \\( o_i \\in R \\). In practice, the color is represented by a series of Spherical Harmonics (SH) coefficients, but for simplicity, we represent the view-independent color by just RGB values. Thus, a single Gaussian can be represented by a set of five attributes as \\( g_i = \\{\\sigma_i, r_i, s_i, o_i, c_i\\} \\in \\mathbb{R}^{14} \\), and the entire 3DGS can be represented by a set of \\( N \\) such Gaussians as: \\( G = \\{\\{\\sigma_i, r_i, s_i, o_i, c_i\\}\\}_{i=1}^N \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5929" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.276, + 0.108 + ], + "angle": 0, + "content": "3.1. Spherical Mapping" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.28 + ], + "angle": 0, + "content": "3DGS is represented as a permutation invariant set with no structural correspondence among different Gaussians \\( g_{i} \\), making it challenging to extract meaningful features from this set containing a few hundred thousands of them using neural networks. To address this, we introduce a novel representation that gives structure to this unstructured set of points and solves the permutation invariance issue for faster and better feature extraction. We propose to accomplish this by employing spherical mapping to map the 3DGS primitives to an image-like representation that is both invariant to random shuffling of 3DGS points and well structured." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.281, + 0.484, + 0.432 + ], + "angle": 0, + "content": "We begin the mapping by inscribing the 3DGS object into a sphere with the same center as the object in the canonical space. Inscribing a 3DGS object into a sphere involves enclosing the object within a sphere. This begins by determining the geometric center of the object. The next step is to calculate the radius of the sphere, which is achieved by measuring the Euclidean distance from the center to the farthest point on the object. The radius of the sphere is defined such that the sphere fully encloses the object. The sphere acts as a bounding volume for the entire object." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.484, + 0.681 + ], + "angle": 0, + "content": "We consider each Gaussian \\(g_{i}\\) in 3D to be centered at the mean position represented by \\(\\sigma_{i}\\) with Cartesian coordinates \\((x_{i},y_{i},z_{i})\\). The aim is to get the spherical coordinates \\((\\rho_{i},\\theta_{i},\\phi_{i})\\) for each Gaussian \\(g_{i}\\). To do so, we calculate the azimuthal \\(\\theta_{i}\\) and polar \\(\\phi_{i}\\) angles for each \\(g_{i}\\) along with the distance from the origin to the point, \\(\\rho_{i}\\). The spherical radius is defined as \\(\\rho_{i} = \\sqrt{x_{i}^{2} + y_{i}^{2} + z_{i}^{2}}\\), the azimuthal angle as \\(\\theta_{i} = \\tan^{-1}(y_{i},x_{i})\\), while the polar angle as \\(\\phi_{i} = \\cos^{-1}(z_{i},\\rho_{i})\\). The azimuthal and polar angles are then normalized, such that we can map them on a 2D UV map of \\(M\\times N\\) dimensionality with 14-channels. \\(\\theta_{i}\\) and \\(\\phi_{i}\\) are converted to degrees and mapped to UV image coordinates: \\(\\theta_{i\\mathrm{scaled}} = \\left\\lfloor \\frac{\\pi + \\theta_i}{2\\pi}\\times \\mathrm{M}\\right\\rfloor\\), \\(\\phi_{i\\mathrm{scaled}} = \\left\\lfloor \\frac{\\phi_i}{\\pi}\\times \\mathrm{N}\\right\\rfloor\\) Each channel in the UV map stores 3DGS attributes, including \\(\\{\\sigma_i,r_i,s_i,o_i,c_i\\} \\in \\mathbb{R}^{14}\\). We refer this 14-channel UV map as UVGS, \\(U\\in \\mathbb{R}^{M\\times N\\times 14}\\) defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.685, + 0.482, + 0.701 + ], + "angle": 0, + "content": "\\[\n\\mathrm {U} \\left[ \\phi_ {i \\text {s c a l e d}}, \\theta_ {i \\text {s c a l e d}},: \\right] = \\left[ \\sigma_ {i}, r _ {i}, s _ {i}, o _ {i}, c _ {i} \\right]. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "This transformed UVGS representation provides spatial coherence and solves the permutation invariance problem as any random arrangement of points will now map to the same UVGS representation \\( U \\). It should be noted that this kind of transformation will also preserve the spatial correlation between the Gaussian points in 3D and transform them to 2D UV maps by mapping them to neighboring pixels. This provides both the local level correspondence among the neighboring Gaussians and the overall global correspondence for the object. Thus, solving the unstructured and discreteness problems. This enables standard neural network architectures (e.g. CNNs) to effectively capture correlations among neighboring Gaussians for efficient feature extraction." + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.093, + 0.698, + 0.136 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {g} _ {1}: \\{\\sigma_ {1}, r _ {1}, s _ {1}, o _ {1}, c _ {1} \\} \\\\ \\mathbf {g} _ {2}: \\{\\sigma_ {2}, r _ {2}, s _ {2}, o _ {2}, c _ {2} \\} \\\\ \\end{array}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.136, + 0.872, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.223, + 0.906, + 0.294 + ], + "angle": 0, + "content": "Figure 3. Dynamic Selection. In spherical mapping of 3DGS points to UV maps, multiple points may map to the same pixel, creating a many-to-one issue. Our Dynamic Selection approach addresses this by retaining the attributes of the point with the highest opacity per pixel on the same ray." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.304, + 0.906, + 0.439 + ], + "angle": 0, + "content": "To further unify the extracted position \\((\\sigma)\\), transformation \\((r, s)\\), and color \\((c, o)\\) maps in a same feature space and use the existing image foundation models, we map the obtained UVGS \\(U \\in \\mathbb{R}^{M \\times N \\times 14}\\) further to a 3-channel image \\(S \\in \\mathbb{R}^{M \\times N \\times 3}\\) (termed as Super UVGS), using a Convolutional Neural Network (CNN). A multi-branch forward mapping network is employed to map \\(U \\in \\mathbb{R}^{M \\times N \\times 14}\\) to the 3-channel Super UVGS \\(S \\in \\mathbb{R}^{M \\times N \\times 3}\\). We provide all technical details in Sec. 3.3 and the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.906, + 0.56 + ], + "angle": 0, + "content": "The Super UVGS representation effectively retains all the details of 3DGS attributes and can be directly utilized with existing widely available image-based models. We demonstrate this by showing perfect reconstruction of 3DGS object from Super UVGS image in the experiments section. This semantically structured representation \\( S \\) offers both local and global correspondence in representing Gaussian attributes and needs relatively less storage." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.562, + 0.882, + 0.578 + ], + "angle": 0, + "content": "3.2. Dynamic GS Selection and Multiple Layers" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.903 + ], + "angle": 0, + "content": "When projecting 3DGS points to UV maps using spherical mapping, multiple points may map to the same pixel in UV space as shown in Fig. 3. Two 3DGS points \\((g_{1})\\) and \\((g_{2})\\) map to the same pixel on UV map \\((P_{a})\\) causing many-to-one mapping issue. To address this, we propose a Dynamic Selection approach where each UV pixel retains the 3DGS attribute with the highest opacity intersecting the same ray. Using the same example in Fig. 3, if opacity \\(o_{1}\\) of Gaussian \\(g_{1}\\) is less than opacity \\(o_{2}\\) of \\(g_{2}\\). Then only \\(g_{2}\\) will be stored in the UV map at pixel \\(P_{a}\\). We observed that this method helps maintain the geometry and appearance of the 3DGS object while resolving many-to-one mapping issues with minimal quality loss. For more complex objects or real-world scene representation, we stack multiple such layers of UV maps, where each UVGS pixel now holds attributes of the top-K opacity values of 3DGS primitives. This can be accomplished by inscribing the 3DGS object inside multiple spheres where each sphere maps the 3DGS attribute corresponding to the top-\\(K^{th}\\) opacity value along the same ray. More details on this are presented in the supplementary. To show the effectiveness of proposed UVGS" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5930" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.477, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.161, + 0.483, + 0.189 + ], + "angle": 0, + "content": "Figure 4. Complex object reconstructions \\((\\mathrm{K} = 4)\\) using pretrained image-based autoencoder." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.194, + 0.484, + 0.24 + ], + "angle": 0, + "content": "maps in capturing the intricacies of a complex objects, we used a pretrained image based autoencoder to reconstruct objects using a 4 layer UVGS as shown in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.245, + 0.276, + 0.261 + ], + "angle": 0, + "content": "3.3. Mapping Networks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.267, + 0.484, + 0.448 + ], + "angle": 0, + "content": "Our goal is to bring the extracted UVGS maps to a common feature space to better represent the object collectively and to make the 14-channel UVGS representations \\( U \\in \\mathbb{R}^{M \\times N \\times 14} \\) work with the widely available image based foundation models. To accomplish this, we map it to a 3-channel image which can be easily processed by the existing architectures while also maintaining the spatial correspondence. We design a simple yet effective multi-branch CNN to extract features from different UVGS attributes and map them to a 3-channel feature-rich image, termed Super UVGS. The structured UVGS maps provide local and global features that can be learned by a CNN." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.449, + 0.484, + 0.78 + ], + "angle": 0, + "content": "Forward Mapping The first layer is a set of three mapping branches for position, transform, and appearance \\((\\phi_P^f,\\phi_T^f,\\phi_A^f)\\) respectively. We refer to them as position, transformation, and appearance branch. The position branch takes the mean position \\((\\sigma)\\) as an input and processes it to give a position feature map \\(M_P\\). Similarly, the transformation branch takes the rotation \\((r)\\) and scale \\((s)\\) together to generate another feature map \\(M_T\\). The last, appearance branch takes the color \\((c)\\) and opacity \\((o)\\) together to produce another feature map \\(M_A\\). All the three features maps from position, transformation, and appearance branch are concatenated to get a final feature map, before passing them to the next module, called the Central Branch. The central branch \\((\\phi_C^f)\\) is composed of multiple hidden Convolution layers, where each layer is followed by BatchNorm and ReLu activation. The last layer of the central branch is activated using tanh to ensure the Super UVGS does not take any ambiguous value resulting in gradient explosion or undesired artifacts. The obtained Super UVGS \\(S\\) representation squeezes all the 3DGS attributes to a 3 dimensional image while also maintaining local and global structural correspondence among them." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Inverse Mapping We design an inverse mapping network that aims to map the obtained 3-channel Super UVGS image \\(S \\in \\mathbb{R}^{M \\times N \\times 3}\\) back to the UVGS maps to obtain each of the five different 3DGS attributes \\(\\{\\sigma, r, s, o, c\\}\\). The inverse mapping network simply follows the forward mapping network architecture in the reverse order, where at first, we put the Central Branch \\((\\phi_{iC})\\) followed by attribute specific position, transformation, and appearance branches" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "\\((\\phi_{iP}, \\phi_{iT}, \\phi_{iA})\\). We provide more details on mapping network in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.124, + 0.907, + 0.501 + ], + "angle": 0, + "content": "Branched mapping layers: The rationale behind using branched mapping layers in both forward and reverse mapping networks is to prevent the incompatibility issues arising due the the different value distribution of 3DGS attributes. Note that the disparate distributions of values within each set of attributes in 3D Gaussian Splatting, (i.e., mean position, transformation, and color), pose a challenge to the model when processed collectively. For instance, neighboring Gaussians in UVGS maps show smooth changes in position and color values but typically have large variations in rotation, scale, and opacity values. This results in gradient anomalies and slow convergence. To address this, we propose a multi-branch network architecture, where attribute-specific branches implicitly learn to process these distinct attribute specific properties, focusing on their unique features before passing them to the central branch. The central branch receives a concatenated stack of processed attributes and exploits the correlation between them by extracting local feature correspondences. This information is then mapped to a 3-channel Super UVGS image, effectively capturing the complex relationships between the various attributes. This approach enables our network to manage diverse attribute distributions, resulting in faster convergence, improved accuracy, and specialized processing for each attribute set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.909, + 0.667 + ], + "angle": 0, + "content": "Reconstruction Losses: Since the obtained UVGS maps have both local and global features, we opted for image-based losses to train the overall architecture. We use a set of Mean Squared Error (MSE) and Learned Perceptual Image Patch Similarity (LPIPS) [59] between the obtained UVGS from spherical mapping \\( U \\) and the predicted UVGS \\( \\hat{U} \\) from inverse mapping network. We calculate the LPIPS loss over four attributes of 3DGS including mean position \\( (\\sigma) \\), view independent color \\( (c) \\), scale \\( (s) \\), and rotation \\( (r) \\). The overall LPIPS loss for UV maps can be written as a linear sum of individual attribute loss terms as:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.671, + 0.905, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {U V - l p i p s} = \\mathcal {L} _ {\\sigma} + \\mathcal {L} _ {s} + \\mathcal {L} _ {r} + \\mathcal {L} _ {c} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.69, + 0.908, + 0.737 + ], + "angle": 0, + "content": "The overall loss function for the training can be written as: \\(\\mathcal{L}_{uvgs} = \\mathcal{L}_{mse} + \\lambda \\mathcal{L}_{UV - lpip}\\) where \\(\\lambda\\) is a scalar and varied from 0 to 10 during the course of training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.746, + 0.646, + 0.763 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.768, + 0.907, + 0.875 + ], + "angle": 0, + "content": "3DGS Dataset and UV maps To train the mapping networks and learn a latent space for unconditional and conditional sampling, we need large amount of 3DGS assets. However, there's a lack of such a large-scale dataset for high quality 3DGS assets. To this end, we create a custom large scale dataset by converting the Objaverse [9] meshes into 3DGS representation1. We start by designing a scene of 88" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.878, + 0.9 + ], + "angle": 0, + "content": "Sketchfab data was filtered out of training data due to its license" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5931" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.09, + 0.127, + 0.171 + ], + "angle": 270, + "content": "Super UVGS" + }, + { + "type": "image", + "bbox": [ + 0.143, + 0.103, + 0.22, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.102, + 0.318, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.103, + 0.408, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.103, + 0.502, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.103, + 0.6, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.103, + 0.692, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.103, + 0.786, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.103, + 0.882, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.192, + 0.126, + 0.257 + ], + "angle": 270, + "content": "Multiview" + }, + { + "type": "image", + "bbox": [ + 0.148, + 0.165, + 0.216, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.168, + 0.311, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.323, + 0.17, + 0.414, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.17, + 0.501, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.17, + 0.607, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.608, + 0.179, + 0.705, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.179, + 0.804, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.805, + 0.184, + 0.882, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.241, + 0.219, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.239, + 0.321, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.223, + 0.415, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.242, + 0.51, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.236, + 0.609, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.237, + 0.7, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.247, + 0.785, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.8, + 0.259, + 0.882, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.307, + 0.907, + 0.351 + ], + "angle": 0, + "content": "Figure 5. Figure shows a wide variety of high-quality unconditional generation result from our method using Latent Diffusion Model. We train an LDM to randomly sample Super UVGS images from random noise. The Super UVGS can be converted to 3DGS object using inverse mapping network and inverse spherical projection. The unconditional generation model was trained on Objaverse dataset." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.362, + 0.485, + 0.665 + ], + "angle": 0, + "content": "cameras in a canonical space and use it to capture Objaverse objects from various angles covering all of the object views. The 88 rendered views from different angles are then used to train a 3DGS for \\(10\\mathrm{K}\\) iterations using [17]. This way, we create a high-quality and large-scale 3DGS dataset of \\(\\sim 400K\\) objects and scenes from Objaverse. We only use static scenes or objects from Objaverse. After fitting all the object to 3DGS representation, we convert the objects to the corresponding UV maps (i.e., UVGS) through Spherical Mapping as illustrated in Fig. 2. For the course of our experiments, we only map the objects to a single layer UV maps as it was sufficient to represent the general purpose Objaverse objects with minimal quality loss. Through mapping, we gathered a UVGS dataset of \\(\\sim 400K\\) maps. We fix the size of the UVGS maps to \\(512\\times 512\\). Through our experiments, we found that UV maps of size \\(512\\times 512\\) are sufficient to represent objects in our dataset and capable of storing upto \\(262K\\) unique Gaussians. Table 1 compares 1-4 layer UV maps. We also did experiments on ShapeNet [4] cars dataset for evaluation purposes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Baselines & Metrics: To evaluate the quality of reconstructed 3DGS objects from both Super UVGS image and Autoencoder latent space to 3D, we use PSNR and LPIPS [58]. The aim is to convert the given 3DGS object to UVGS, and then to Super UVGS, and further to Autoencoder's latent space and calculate the metrics from the reconstructions at every step to prove the proposed method doesn't significantly affect the quality of reconstructions, while also providing a structurally meaningful representation that is much compact and easier to use with existing image based models. We compare the generational capabilities of our method against various conditional and unconditional SOTA 3D object generation method including the ones using multiview rendering for optimization DiffTF [2], Get3D [12], methods trying to give structural representa" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.36, + 0.907, + 0.43 + ], + "angle": 0, + "content": "Table 1. PSNR and LPIPS comparison for various reconstruction methods using UVGS and Super UVGS representations on Objverse Cars and Full datasets. AE, VAE, VQVAE are pretrained image based models. \\(K\\) is the number of UVGS layers used. We also report the compression \\(\\%\\) (CP) compared to the fitted 3DGS." + }, + { + "type": "table", + "bbox": [ + 0.552, + 0.441, + 0.865, + 0.581 + ], + "angle": 0, + "content": "
MethodPSNR(C/F)LPIPS(C/F)CP(%)
3DGS34.6 / 34.20.02 / 0.020
UVGS (@K=1)31.3 / 31.10.06 / 0.0653.0
UVGS (@K=2)32.8 / 31.90.04 / 0.0545.6
UVGS (@K=4)34.2 / 33.20.02 / 0.0333.3
Super UVGS (@K=1)31.2 / 31.10.07 / 0.0889.7
AE (@K=1)30.9 / 30.80.07 / 0.0999.5
VAE (@K=1)30.6 / 30.90.07 / 0.0999.5
VQVAE (@K=1)30.3 / 30.10.08 / 0.1099.7
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.591, + 0.907, + 0.652 + ], + "angle": 0, + "content": "tion to Gaussians, GaussianCube [55], and general purpose SOTA large 3D content generation models like DreamGaussian [42], LGM [44], and EG3D [3]. We also compare the quality of our generation results using FID and KID." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.653, + 0.907, + 0.744 + ], + "angle": 0, + "content": "Mapping Network Training Details We train the forward and inverse mapping networks to project the obtained UVGS maps \\( U \\in \\mathbb{R}^{M \\times N \\times 14} \\) to Super UVGS image \\( S \\in \\mathbb{R}^{M \\times N \\times 3} \\), and back to the reconstructed UVGS maps \\( \\hat{U} \\in \\mathbb{R}^{M \\times N \\times 14} \\). We provide an in-depth discussion of all implementation details in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.756, + 0.895, + 0.773 + ], + "angle": 0, + "content": "4.1. UVGS AutoEncoder and 3DGS Compression" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The obtained Super UVGS image is a structurally meaningful representation that can have various applications in the generation and reconstruction of new 3D assets as it contains features that can be learned by the existing image based models. Through our experiments, we show that a 3-channel Super UVGS image can be directly reconstructed using a pretrained image based Autoencoders or VAEs without any fine-tuning. We tested on three different models in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5932" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.092, + 0.185, + 0.107 + ], + "angle": 0, + "content": "DiffTF" + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.127, + 0.198, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.178, + 0.199, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.314, + 0.094, + 0.36, + 0.106 + ], + "angle": 0, + "content": "Get3D" + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.116, + 0.309, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.177, + 0.308, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.114, + 0.411, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.172, + 0.411, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.489, + 0.094, + 0.591, + 0.106 + ], + "angle": 0, + "content": "GaussianCube" + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.13, + 0.536, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.127, + 0.598, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.186, + 0.539, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.128, + 0.653, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.185, + 0.652, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.094, + 0.824, + 0.108 + ], + "angle": 0, + "content": "UVGS (ours)" + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.126, + 0.799, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.803, + 0.118, + 0.876, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.181, + 0.799, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.805, + 0.179, + 0.877, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.23, + 0.907, + 0.285 + ], + "angle": 0, + "content": "Figure 6. Comparison of unconditional 3D asset generation on the cars category with SOTA methods. Figure shows that DiffTF [2] produces low-quality, low-resolution cars lacking detail. While Get3D [12] achieve higher resolution, it suffers from 3D inconsistency, numerous artifacts, and lacks 3D detail. Similar issues are found in GaussianCube [55] along with symmetric inconsistency in the results. In contrast, our method generates high-quality, high-resolution objects that are 3D consistent with sharp and well-defined edges." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.296, + 0.485, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.495, + 0.483, + 0.551 + ], + "angle": 0, + "content": "Figure 7. We compare the performance of our model against various SOTA methods for text-conditional object synthesis. Our method not only generates high-quality assets for simpler objects, but also for complicated objects with intricate geometry." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.556, + 0.483, + 0.708 + ], + "angle": 0, + "content": "cluding image AE, KL-VAE [18], VQVAE [46] and each performed quite well without any significant quality loss. The reconstruction PSNR and LPIPS values are presented in Table 1. This means we can now leverage the powerful compression capabilities of image based Autoencoders to compress the storage requirements of 3DGS by more than \\(99\\%\\). We have shown the storage comparison results in Table 1. It is interesting to note that the Super UVGS representation itself can be used to compress the memory requirement for storing 3DGS object by up to \\(89.7\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.442, + 0.728 + ], + "angle": 0, + "content": "4.2. Unconditional & Conditional Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We aim to show the effectiveness of Super UVGS representation for directly generating 3DGS objects from a learned latent space. We consider Super UVGS images as a compact and structured proxy for representing 3DGS objects as it maintains the 3D object information while also providing learnable features. The existing methods fail to directly generate a large number of Gaussians (e.g. \\(100\\mathrm{K}+\\)) to represent objects with sufficient quality either due to the lack of 3D generative model architectures that support such large number of unstructured points, or due to the lack of a large 3DGS dataset [28, 40, 44, 55, 63]. We leverage the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.298, + 0.907, + 0.464 + ], + "angle": 0, + "content": "Super UVGS representation and use the existing 2D image generative models like Diffusion Models [15] for this task. Specifically, to train a generative model capable of randomly sampling new high-quality 3DGS assets for various downstream tasks, we use an unconditional Latent Diffusion Model (LDM) [36] on the obtained Super UVGS images. As illustrated in Section 4.1, we can use a pretrained image VAE to map the Super UVGS image to a latent space and reconstruct back. Hence, we only train a LDM on the latent space. More implementation details are provided in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.909, + 0.665 + ], + "angle": 0, + "content": "Unconditional LDM: To design a generative model capable of randomly sampling new high-quality 3DGS assets for various downstream tasks, we train an unconditional LDM [36] on the learned Super UVGS images. Following [33, 36, 57], we use DDIM [15] for faster and consistent sampling with up to 1000 time steps used in the forward diffusion process, and 20 during denoising. Once trained, the model is used to randomly sample Super UVGS images, resulting in high quality 3DGS assets through inverse mapping. Results are presented in Fig 5. We demonstrate that our method inherently learns to generate multiview consistent images due to the powerful Super UVGS representation unlike most prior works using rendering-based losses." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.67, + 0.909, + 0.76 + ], + "angle": 0, + "content": "Conditional LDM: Similar to unconditional generation, we also trained a text-conditioned LDM following the SD's [33, 36, 57] pipeline and using the predicted text for our dataset. The trained model can be used to generate high-quality text-conditioned 3DGS assets that are multiview consistent. The results are demonstrated in Fig 6." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.901 + ], + "angle": 0, + "content": "The above experiments prove the effectiveness of our proposed Super UVGS representation in 3D object synthesis using widely available 2D image models. It also highlights that this compact 3-channel Super UVGS representation stores not just the spatial correspondence among different pixels, but also the rich 3D information of the objects. This way, we can easily convert a 3D asset generation problem into a 2D image generation problem without the use of any complex 3D architecture to handle large amount of un" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5933" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.086, + 0.475, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.288, + 0.483, + 0.331 + ], + "angle": 0, + "content": "Figure 8. 3DGS Inpainting: We present one of the first inpainting results on 3DGS directly leveraging the Super UVGS images and the denoising capabilities of diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.341, + 0.484, + 0.402 + ], + "angle": 0, + "content": "structured and permutation invariant 3DGS primitives, and neither relying upon computationally expensive multiview rendering or SDS loss. Baseline comparison for unconditional and conditional generation is presented in Table 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.41, + 0.262, + 0.427 + ], + "angle": 0, + "content": "4.3. 3DGS Inpainting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.484, + 0.569 + ], + "angle": 0, + "content": "Leveraging the powerful Super UVGS representation, we present in Fig. 8 one of the first experiments on inpainting 3DGS directly without using any multiview rendering or distilling information from diffusion models. We try to recover the missing Gaussians by leveraging the denoising capabilities of LDM and trying to predict the missing corresponding parts of the Super UVGS image. We believe, this can have potential applications in sparse view reconstruction. More details are given in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.577, + 0.36, + 0.592 + ], + "angle": 0, + "content": "4.4. Ablation Studies & Discussion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We conduct exhaustive ablation studies to justify some of our framework's design choices including the effect of branching in mapping networks, the use of single layer UVGS maps, and the resolution of UVGS maps. We performed our experiments on our custom Objaverse [9] 3DGS dataset and evaluate the performance of our model in terms of PSNR, SSIM, and LPIPS. The results are presented in Table 3. From the table, it can be seen that by using four layers of UVGS maps \\((\\mathrm{K} = 4)\\), we can almost match the reconstruction quality of fitted 3DGS results, while we realized that simply with a single layer UVGS, we are able to maintain the overall geometry and appearance of the object for our dataset with a PSNR of more than 30. We also compared the reconstruction performance of our method with and without using branching in the mapping network. It can be clearly seen that using branching network significantly increases the reconstruction quality from Super UVGS space. The main reasoning behind this is the specialization of attributes that the branching provides to individually process each attribute first." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.083, + 0.907, + 0.152 + ], + "angle": 0, + "content": "Table 2. We compare the FID and KID of unconditional generation using the current SOTA methods on 20K randomly generated samples from each method and ours. We also compare our method against SOTA text-conditioned generation frameworks on CLIP Score for 10K generated objects from each method." + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.156, + 0.875, + 0.265 + ], + "angle": 0, + "content": "
Unconditional GenerationText-Conditioned Generation
MethodFID ↓KID ↓MethodCLIP Score ↑
Get3D [12]53.174.19DreamGaussian [42]28.51
DiffTF [2]84.578.73Shap. E [6]30.53
EG3D [3]74.516.62LGM [44]30.74
GaussianCube34.673.72GaussianCube [55]30.34
UVGS (Ours)26.203.24UVGS (Ours)32.62
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.27, + 0.907, + 0.313 + ], + "angle": 0, + "content": "Table 3. We present quantitative ablation study for number of UVGS layers (K), UVGS map resolution, and the effect of branching in mapping network on the Objaverse 3DGS dataset." + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.317, + 0.875, + 0.38 + ], + "angle": 0, + "content": "
MethodPSNRLPIPSUVGS SizePSNRLPIPS
UVGS @K=131.10.06512 × 512 (@K=1)31.10.08
UVGS @K=231.90.05256 × 256 (@K=1)28.20.23
UVGS @K=433.20.03Without Branching27.80.31
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.389, + 0.907, + 0.601 + ], + "angle": 0, + "content": "Limitations & Future Work: While single layer UVGS images can recover the geometry of the object, they sometimes suffer in terms of appearance and the generated objects might look washed out. We believe this can be solved by using a multi-layer UVGS maps. Similarly, the single layer UVGS map is limited to representing simpler everyday objects, and may not be sufficient to represent highly-detailed and complex objects or scenes. In the future, we want to extend this framework to learn features for real-world scenes and complex objects like a human head with multi-layer UV mapping. We also want to make this representation more efficient by better utilizing the empty pixels of UVGS maps and Super UVGS images while also maintaining the underlying features and 3D information." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.634, + 0.622 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.907, + 0.852 + ], + "angle": 0, + "content": "We introduced a novel method to solve the underlying issues with 3D Gaussian Splatting (3DGS) that prevent the direct integration of them with the large number of existing image foundational models. We proposed UVGS - a structured representation for 3DGS obtained by spherical mapping of 3DGS primitives to UV maps. We further squeezed the multi-attribute UVGS maps to a 3-channel unified and structured Super UVGS image, which not only maintains the 3D structural information of the object, but also provides a compact feature space for 3DGS attributes. The obtained Super UVGS images are directly integrated with the existing image foundational models for 3DGS compression and unconditional and conditional generation using diffusion models. Leveraging these Super UVGS images, we showed one of the first inpainting experiments on 3DGS." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "ACKNOWLEDGEMENTS A part of this work was supported by NSF CAREER grant 2143576 and ONR DURIP grant N00014-23-1-2804." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5934" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.198 + ], + "angle": 0, + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855-5864, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.201, + 0.484, + 0.243 + ], + "angle": 0, + "content": "[2] Ziang Cao, Fangzhou Hong, Tong Wu, Liang Pan, and Ziwei Liu. Large-vocabulary 3d diffusion model with transformer. arXiv preprint arXiv:2309.07920, 2023. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.484, + 0.328 + ], + "angle": 0, + "content": "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16123-16133, 2022. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.331, + 0.484, + 0.428 + ], + "angle": 0, + "content": "[4] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.43, + 0.484, + 0.485 + ], + "angle": 0, + "content": "[5] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision, pages 333-350. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.488, + 0.484, + 0.557 + ], + "angle": 0, + "content": "[6] Minghao Chen, Junyu Xie, Iro Laina, and Andrea Vedaldi. Shap-editor: Instruction-guided latent 3d editing in seconds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26456-26466, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.559, + 0.484, + 0.615 + ], + "angle": 0, + "content": "[7] Zilong Chen, Feng Wang, Yikai Wang, and Huaping Liu. Text-to-3d using gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21401-21412, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.484, + 0.686 + ], + "angle": 0, + "content": "[8] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4456-4465, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.688, + 0.484, + 0.771 + ], + "angle": 0, + "content": "[9] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli Vanderbilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[10] Slava Elizarov, Ciara Rowles, and Simon Donne. Geometry image diffusion: Fast and data-efficient text-to-3d with image-based surface representation. arXiv preprint arXiv:2409.03718, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[11] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.175 + ], + "angle": 0, + "content": "[12] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841-31854, 2022. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.177, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[13] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[14] Xianglong He, Junyi Chen, Sida Peng, Di Huang, Yangguang Li, Xiaoshui Huang, Chun Yuan, Wanli Ouyang, and Tong He. Gvgen: Text-to-3d generation with volumetric representation. In European Conference on Computer Vision, pages 463–479. Springer, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.289, + 0.905, + 0.329 + ], + "angle": 0, + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.331, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[16] Han Huang, Yulun Wu, Junsheng Zhou, Ge Gao, Ming Gu, and Yu-Shen Liu. Neusurf: On-surface priors for neural surface reconstruction from sparse input views. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2312-2320, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.4, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.483 + ], + "angle": 0, + "content": "[18] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.905, + 0.565 + ], + "angle": 0, + "content": "[19] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[20] Shujuan Li, Junsheng Zhou, Baorui Ma, Yu-Shen Liu, and Zhizhong Han. Neaf: Learning neural angle fields for point normal estimation. In Proceedings of the AAAI conference on artificial intelligence, pages 1396-1404, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.624, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[21] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[22] Baorui Ma, Haoge Deng, Junsheng Zhou, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Geodream: Disentangling 2d and geometric priors for high-fidelity and consistent 3d generation. arXiv preprint arXiv:2311.17971, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.763, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[23] Qi Ma, Yue Li, Bin Ren, Nicu Sebe, Ender Konukoglu, Theo Gevers, Luc Van Gool, and Danda Pani Paudel. Shapes-plat: A large-scale dataset of gaussian splats and their self-supervised pretraining. arXiv preprint arXiv:2408.10906, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[24] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.906, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5935" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.161, + 0.482, + 0.202 + ], + "angle": 0, + "content": "[26] Chaerin Min and Srinath Sridhar. Genheld: Generating and editing handheld objects. arXiv preprint arXiv:2406.05059, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.204, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[27] Chaerin Min, Sehyun Cha, Changhee Won, and Jongwoo Lim. Tsdf-sampling: Efficient sampling for neural surface field using truncated signed distance field. arXiv preprint arXiv:2311.17878, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.327 + ], + "angle": 0, + "content": "[28] Yuxuan Mu, Xinxin Zuo, Chuan Guo, Yilin Wang, Juwei Lu, Xiaofeng Wu, Songcen Xu, Peng Dai, Youliang Yan, and Li Cheng. Gsd: View-guided gaussian splatting diffusion for 3d reconstruction. arXiv preprint arXiv:2407.04237, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[29] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulo, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4328-4338, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.467 + ], + "angle": 0, + "content": "[30] Francesco Palandra, Andrea Sanchietti, Daniele Baieri, and Emanuele Rodola. Gsedit: Efficient text-guided editing of 3d objects via gaussian splatting. arXiv preprint arXiv:2403.05154, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[31] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[32] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[33] Aashish Rai and Srinath Sridhar. Egosonics: Generating synchronized audio for silent egocentric videos. arXiv preprint arXiv:2407.20592, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[34] Aashish Rai, Hiresh Gupta, Ayush Pandey, Francisco Vicente Carrasco, Shingo Jason Takagi, Amaury Aubel, Daeil Kim, Aayush Prakash, and Fernando De la Torre. Towards realistic generative 3d face models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3738-3748, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[35] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject-driven text-to-3d generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2349–2359, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[37] Li Shen and Fillia Makedon. Spherical mapping for processing of 3d closed surfaces. Image and vision computing, 24 (7):743-761, 2006. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[38] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20875-20886, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[39] Jingxiang Sun, Bo Zhang, Ruizhi Shao, Lizhen Wang, Wen Liu, Zhenda Xie, and Yebin Liu. Dreamcraft3d: Hierarchical 3d generation with bootstrapped diffusion prior. arXiv preprint arXiv:2310.16818, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[40] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10208-10217, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[41] Fariborz Taherkhani, Aashish Rai, Quankai Gao, Shaunak Srivastava, Xuanbai Chen, Fernando De la Torre, Steven Song, Aayush Prakash, and Daeil Kim. Controllable 3d generative adversarial face model via disentangling shape and appearance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 826-836, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[42] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[43] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22819–22829, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.586 + ], + "angle": 0, + "content": "[44] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1-18. Springer, 2025. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[45] Zhicong Tang, Shuyang Gu, Chunyu Wang, Ting Zhang, Jianmin Bao, Dong Chen, and Baining Guo. Volumediffusion: Flexible text-to-3d generation with efficient volumetric encoder. arXiv preprint arXiv:2312.11459, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[46] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[47] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4563-4573, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[48] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3295–3306, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[49] Yuxuan Wang, Xuanyu Yi, Zike Wu, Na Zhao, Long Chen, and Hanwang Zhang. View-consistent 3d editing with gaussian splatting. In European Conference on Computer Vision, pages 404-420. Springer, 2025. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "5936" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[50] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.244 + ], + "angle": 0, + "content": "[51] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.314 + ], + "angle": 0, + "content": "[52] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation. arXiv preprint arXiv:2403.14621, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[53] Xingguang Yan, Han-Hung Lee, Ziyu Wan, and Angel X Chang. An object is worth 64x64 pixels: Generating 3d object via image diffusion. arXiv preprint arXiv:2408.03178, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.439 + ], + "angle": 0, + "content": "[54] Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian-dreamer: Fast generation from text to 3d gaussian splattering with point cloud priors. arXiv preprint arXiv:2310.08529, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.441, + 0.482, + 0.509 + ], + "angle": 0, + "content": "[55] Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: Structuring gaussian splatting using optimal transport for 3d generative modeling. arXiv preprint arXiv:2403.19655, 2024. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[56] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[57] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[58] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.482, + 0.748 + ], + "angle": 0, + "content": "[59] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.749, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[60] Junsheng Zhou, Yu-Shen Liu, and Zhizhong Han. Zero-shot scene reconstruction from single images with deep prior assembly. arXiv preprint arXiv:2410.15971, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[61] Junsheng Zhou, Weiqi Zhang, and Yu-Shen Liu. Diff-gfs: Functional gaussian splatting diffusion. arXiv preprint arXiv:2410.19657, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[62] Junsheng Zhou, Weiqi Zhang, Baorui Ma, Kanle Shi, YuShen Liu, and Zhizhong Han. Udiff: Generating conditional unsigned distance fields with optimal wavelet diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21496-21506, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[63] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10324-10335, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.907, + 0.261 + ], + "angle": 0, + "content": "[64] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10324-10335, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5937" + } + ] +] \ No newline at end of file diff --git a/2025/UVGS_ Reimagining Unstructured 3D Gaussian Splatting using UV Mapping/77087a0b-d1cb-4230-a846-7760f7510d17_origin.pdf b/2025/UVGS_ Reimagining Unstructured 3D Gaussian Splatting using UV Mapping/77087a0b-d1cb-4230-a846-7760f7510d17_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..92769dc66ce0741c51ba30c003d5d6a145f712a8 --- /dev/null +++ b/2025/UVGS_ Reimagining Unstructured 3D Gaussian Splatting using UV Mapping/77087a0b-d1cb-4230-a846-7760f7510d17_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9685d4b7025e72d98b4757baddf280500552f3093b85062a5e360ff035bcf1cd +size 9509467 diff --git a/2025/UVGS_ Reimagining Unstructured 3D Gaussian Splatting using UV Mapping/full.md b/2025/UVGS_ Reimagining Unstructured 3D Gaussian Splatting using UV Mapping/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ce2e993a8fb80b9a62babb15838524ea41c122bd --- /dev/null +++ b/2025/UVGS_ Reimagining Unstructured 3D Gaussian Splatting using UV Mapping/full.md @@ -0,0 +1,340 @@ +# UVGS: Reimagining Unstructured 3D Gaussian Splitting using UV Mapping + +Aashish Rai $^{1,2}$ + +Dilin Wang $^{2}$ Mihir Jain $^{2}$ Srinath Sridhar $^{1}$ + +Nikolaos Sarafianos2 +Aayush Prakash2 + +Kefan Chen $^{1,2}$ + +$^{1}$ Brown University + +Meta Reality Labs + +https://aashishrai3799.github.io/uvgs + +![](images/5f9e3c045a3b9caf6b11549a80329ba7ba55bde2f88692372e9d39e48b98fb56.jpg) +Figure 1. We propose UVGS - an structured image-like representation for 3DGS obtained by spherical mapping of its primitives. The obtained UVGS maps can be further unified to a 3-channel Super UVGS image to bridging the gap between 3DGS and existing image foundation models. We show Super UVGS can compress the 3DGS assets using pretrained image Autoencoders, and for direct unconditional and conditional 3DGS object generation using diffusion models. We also present one of the first experiments on 3DGS inpainting. + +# Abstract + +3D Gaussian Splatting (3DGS) has demonstrated superior quality in modeling 3D objects and scenes. However, generating 3DGS remains challenging due to their discrete, unstructured, and permutation-invariant nature. In this work, we present a simple yet effective method to overcome these challenges. We utilize spherical mapping to transform 3DGS into a structured 2D representation, termed UVGS. UVGS can be viewed as multi-channel images, with feature dimensions as a concatenation of Gaussian attributes such as position, scale, color, opacity, and rotation. We further find that these heterogeneous features can be compressed into a lower-dimensional (e.g., 3-channel) shared feature space using a carefully designed multi-branch network. The compressed UVGS can be treated as typical RGB images. Remarkably, we discover that typical VAEs trained with latent diffusion models can directly generalize to this new representation without additional training. Our novel representation makes it effortless to leverage foundational 2D models, such as diffusion models, to directly model 3DGS. Additionally, one can simply increase the 2D UV resolution + +to accommodate more Gaussians, making UVGS a scalable solution compared to typical 3D backbones. This approach immediately unlocks various novel generation applications of 3DGS by inherently utilizing the already developed superior 2D generation capabilities. In our experiments, we demonstrate various unconditional, conditional generation, and inpainting applications of 3DGS based on diffusion models, which were previously non-trivial. + +# 1. Introduction + +The creation of high-quality 3D content is essential in applications like virtual reality, game design, robotics, and movie production, where realistic 3D representations play a critical role. Typical 3D representations like Neural Radiance Fields (NeRF) [25] are promising but require substantial computational resources, limiting their scalability for real-time applications. Moreover, NeRF is an implicit representation, which makes editing and manipulation challenging. Recently, 3D Gaussian Splatting (3DGS) [17] emerged as a compelling alternative, enabling efficient and high-fidelity 3D rendering through a large set of Gaussian primitives that model spatial and visual properties. As an explicit represen + +tation, 3DGS offers several advantages over NeRF. However, while 3DGS offers benefits in terms of speed and visual quality, its unstructured, permutation-invariant nature presents significant challenges for generative tasks. Much like point clouds, it lacks a coherent spatial structure, impeding its integration with conventional image-based generative models. This lack of structure and coherence among primitives hinders the application of image-based generative models [23, 55, 61], which rely on structured data representations. + +Previous methods have tackled these challenges by transforming 3DGS into structured formats, such as voxel grids [8, 14, 55] or image-based representations like Splatter Image [40] or triplanes [64]. Other approaches employ diffusion models to directly predict 3DGS attributes [28]. These methods, while achieving impressive visual results, often require substantial computational resources, memory-intensive multi-view rendering, complex architectures limiting their scalability and flexibility for high-fidelity generation. Generating and processing 3DGS directly by efficiently utilizing modern generative models like Variational Autoencoders (VAEs) and diffusion models is limited as the neural networks are not permutation invariant. + +To address these shortcomings, we introduce UV Gaussian Splatting (UVGS), which provides a structured transformation of 3D Gaussian primitives into a 2D representation while preserving essential 3D information. We use spherical mapping [37] that inscribes Gaussian splats in a spherical surface, and projects attributes like position, rotation, scale, opacity, and color into an organized 14-channel image-like UV map. This mapping introduces spatial structure, resolving issues of permutation invariance by introducing local correspondences between neighboring Gaussians and global coherence across the entire 3D object. The result is a representation that functions as a "3D representation" structured in a 2D map format, enabling compatibility with powerful image-based neural network architectures. + +While UVGS introduces structure into 3D Gaussian Splitting, its full 14-channel attribute-specific representation presents challenges for direct integration with pretrained 2D generative models, as these models typically expect a simpler, image-compatible data. Each of its heterogeneous attributes—position, color, and transformation—has its own distinct distribution and resides in a separate feature space, making it challenging to represent the 3D object in a unified shared space. To address this, we introduce Super UVGS, a compact 3-channel representation that unifies these diverse attributes into a cohesive format. Using a carefully designed multi-branch mapping network, Super UVGS consolidates the distinct attribute spaces into a shared feature space, enabling a more collective representation of the object. This unified transformation not only facilitates zero-shot compatibility with pretrained 2D models + +but also optimizes memory usage and computational efficiency, making Super UVGS highly practical for large-scale 3D tasks. Unlike previous approaches that use Triplanes, voxels, occupancy grid, neural fields etc. and require specialized 3D architectures to train on 3D data, UVGS effortlessly leverages widely available pretrained 2D foundational models. This zero-shot generalization capability allows UVGS to fully benefit from priors learned in 2D domains from large amount of data, improving both flexibility and scalability. To sum up our main contributions are: + +- Efficient Structured Representation of 3DGS: We present UVGS, an image-like representation that solves permutation invariance and unstructured nature of discrete 3DGS through spherical mapping, making direct feature extraction possible by organizing unordered points into a coherent 2D representation compatible with 2D models. +- Compact and Scalable Super UVGS Representation: To address scalability while dealing with large scale 3DGS points and enabling the direct integration of pre-trained 2D foundation models, we introduce Super UVGS - a low-dimensional version of UVGS maps that retains high fidelity features while reducing memory overhead. +- Diverse 3D Applications: Our approach unlocks seamless integration of 3DGS with pre-trained 2D foundation models for various tasks, including unconditional and conditional generation of 3DGS. + +# 2. Related Work + +3D Generative/Reconstruction Models for Objects: Generation or reconstruction of 3D assets has been a long standing task [3, 21, 24, 26, 27, 31, 34, 35, 39, 41, 43, 50, 54]. Previous reconstruction approaches like NeRF [1, 25] are often slow and do not provide a defining geometry [5, 11, 16, 20, 22, 24, 32, 35, 48, 60, 62]. Advancements in the field led to the emergence of explicit voxel grid based representations that encode colors and opacities directly [8, 29, 45]. These approaches achieve significant speed ups compared to the NeRF based approaches, but they can't produce high fidelity assets due to the low resolution of voxel grids. On the other hand, triplane representation [13, 38, 47, 63] provides a trade-off between the quality and memory utilization. Another line of work [10, 53] splits the input mesh into different patches and simplifies the object generation problem to an image generation problem. However such methods rely on either cutting through the mesh to create a geometry image [10] or rely on an existing UV representation of the geometry and utilize subset of existing UV islands [53] resulting in loss of details. Recently, there has been a notable advancement in 3D Gaussian Splatting (3DGS) for the representation of objects and scenes leading to the emergence of 3DGS showcasing impressive real-time results in reconstruction and generation tasks [14, 17, 19, 30, 42, 49, 51, 52, 54, 56, 63]. Recent + +![](images/4067950a35b01263b3e660b2b4fc16218bdce22a67eca1d67abe820a53df3b98.jpg) +Figure 2. The input 3DGS object is first converted to UVGS maps through spherical mapping. We use a multibranch forward mapping network to convert the obtained 14-channel UVGS to a compact 3-channel Super UVGS image. This represents the 3DGS object in a structured manner and can be used with image foundation models for reconstruction or generation. The Super UVGS is mapped back to UVGS through branched inverse mapping, which in turn can be reconstructed back to the 3DGS object through inverse spherical mapping. + +advances in the 3D generative models for asset synthesis using the existing geometries like NeRF, voxel grids, or triplane geometries [2, 12, 42, 52, 56, 63] leverage generative models [15, 36] and the existing 3D datasets [4, 9]. However, most of the works employing 3DGS or other representations use multiview rendering and Score Distillation Sampling (SDS) to achieve convincing generation and reconstruction capabilities [7, 31]. These approaches demand high memory and compute resources and are often quite slow in optimizing due to per scene optimization. + +Giving Structures to Discrete Gaussians: Although, 3DGS has led to breakthrough in the reconstruction field by demonstrating superior performance in multiple domains, the generation of 3DGS directly remains challenging due to its discreteness and unstructured nature [55, 61]. These characteristics present substantial challenges when integrating them with conventional computer vision models, like Autoencoders and generative models [61]. The research in direct learning of trained 3DGS primitives is largely unexplored [23]. Some efforts attempt to address this by directly predicting 3DGS attributes using diffusion models [28] while others like Splatter Image [40] project Gaussian objects into image-based representations through direct 3D-unaware projection. These methods struggle with maintaining multiview consistency, as the model only infers seen poses correctly, while hallucinating for unseen poses. + +Concurrent works [14, 55] follow the voxel-based representations to transport Gaussians into structural voxel grids with volume generation models for generating Gaussians. However, these methods are computationally expensive for high-resolution voxels, face difficulties in preserving high-quality Gaussian reconstructions due to information loss during voxelization. DiffGS [61] tries to solve the above + +issues by proposing three continuous functions to represent 3DGS. However, it is limited to only category-level generation and learning generic probability functions for all the categories poses significant compute and design challenges. In contrary, we introduce an efficient way to give structures to discrete Gaussians by taking inspiration from the developments in 3D graphics. Our method does not require any learning to map an unstructured set of Gaussians to this efficient and structured representation (termed UVGS). The proposed representation provides local and global correspondence among different Gaussian points making the widely available existing computer vision frameworks learn and extract underlying features from them. + +# 3. Methodology + +Preliminaries: 3DGS represents an object or a scene with a collection of Gaussians primitives to model the geometry and view-dependent appearance. For a 3DGS set, $G = \{g_i\}_{i=1}^N$ , representing an object with $N$ individual Gaussians, the geometry of the $i^{th}$ Gaussian is explicitly parameterized via 3D covariance matrix $\Sigma_i$ and it's center $\sigma_i \in \mathbb{R}^3$ as: $g_i(x) = e^{(-\frac{1}{2}(x - \sigma_i)^T\Sigma^{-1}(x - \sigma_i))}$ where, the covariance matrix $\Sigma_i = r_is_is_i^T r_i^T$ is factorized into a rotation matrix $r_i \in \mathbb{R}^4$ and a scale matrix $s_i \in \mathbb{R}^3$ . The appearance of the $i - th$ Gaussian is represented by a color value $c_i \in \mathbb{R}^3$ and an opacity value $o_i \in R$ . In practice, the color is represented by a series of Spherical Harmonics (SH) coefficients, but for simplicity, we represent the view-independent color by just RGB values. Thus, a single Gaussian can be represented by a set of five attributes as $g_i = \{\sigma_i, r_i, s_i, o_i, c_i\} \in \mathbb{R}^{14}$ , and the entire 3DGS can be represented by a set of $N$ such Gaussians as: $G = \{\{\sigma_i, r_i, s_i, o_i, c_i\}\}_{i=1}^N$ . + +# 3.1. Spherical Mapping + +3DGS is represented as a permutation invariant set with no structural correspondence among different Gaussians $g_{i}$ , making it challenging to extract meaningful features from this set containing a few hundred thousands of them using neural networks. To address this, we introduce a novel representation that gives structure to this unstructured set of points and solves the permutation invariance issue for faster and better feature extraction. We propose to accomplish this by employing spherical mapping to map the 3DGS primitives to an image-like representation that is both invariant to random shuffling of 3DGS points and well structured. + +We begin the mapping by inscribing the 3DGS object into a sphere with the same center as the object in the canonical space. Inscribing a 3DGS object into a sphere involves enclosing the object within a sphere. This begins by determining the geometric center of the object. The next step is to calculate the radius of the sphere, which is achieved by measuring the Euclidean distance from the center to the farthest point on the object. The radius of the sphere is defined such that the sphere fully encloses the object. The sphere acts as a bounding volume for the entire object. + +We consider each Gaussian $g_{i}$ in 3D to be centered at the mean position represented by $\sigma_{i}$ with Cartesian coordinates $(x_{i},y_{i},z_{i})$ . The aim is to get the spherical coordinates $(\rho_{i},\theta_{i},\phi_{i})$ for each Gaussian $g_{i}$ . To do so, we calculate the azimuthal $\theta_{i}$ and polar $\phi_{i}$ angles for each $g_{i}$ along with the distance from the origin to the point, $\rho_{i}$ . The spherical radius is defined as $\rho_{i} = \sqrt{x_{i}^{2} + y_{i}^{2} + z_{i}^{2}}$ , the azimuthal angle as $\theta_{i} = \tan^{-1}(y_{i},x_{i})$ , while the polar angle as $\phi_{i} = \cos^{-1}(z_{i},\rho_{i})$ . The azimuthal and polar angles are then normalized, such that we can map them on a 2D UV map of $M\times N$ dimensionality with 14-channels. $\theta_{i}$ and $\phi_{i}$ are converted to degrees and mapped to UV image coordinates: $\theta_{i\mathrm{scaled}} = \left\lfloor \frac{\pi + \theta_i}{2\pi}\times \mathrm{M}\right\rfloor$ , $\phi_{i\mathrm{scaled}} = \left\lfloor \frac{\phi_i}{\pi}\times \mathrm{N}\right\rfloor$ Each channel in the UV map stores 3DGS attributes, including $\{\sigma_i,r_i,s_i,o_i,c_i\} \in \mathbb{R}^{14}$ . We refer this 14-channel UV map as UVGS, $U\in \mathbb{R}^{M\times N\times 14}$ defined as: + +$$ +\mathrm {U} \left[ \phi_ {i \text {s c a l e d}}, \theta_ {i \text {s c a l e d}},: \right] = \left[ \sigma_ {i}, r _ {i}, s _ {i}, o _ {i}, c _ {i} \right]. \tag {1} +$$ + +This transformed UVGS representation provides spatial coherence and solves the permutation invariance problem as any random arrangement of points will now map to the same UVGS representation $U$ . It should be noted that this kind of transformation will also preserve the spatial correlation between the Gaussian points in 3D and transform them to 2D UV maps by mapping them to neighboring pixels. This provides both the local level correspondence among the neighboring Gaussians and the overall global correspondence for the object. Thus, solving the unstructured and discreteness problems. This enables standard neural network architectures (e.g. CNNs) to effectively capture correlations among neighboring Gaussians for efficient feature extraction. + +$$ +\begin{array}{l} \mathbf {g} _ {1}: \{\sigma_ {1}, r _ {1}, s _ {1}, o _ {1}, c _ {1} \} \\ \mathbf {g} _ {2}: \{\sigma_ {2}, r _ {2}, s _ {2}, o _ {2}, c _ {2} \} \\ \end{array} +$$ + +![](images/5f38a1646a02c3cfbc429eb39ddf0d13ec479cfd6b803ce2154472531c1b93ff.jpg) +Figure 3. Dynamic Selection. In spherical mapping of 3DGS points to UV maps, multiple points may map to the same pixel, creating a many-to-one issue. Our Dynamic Selection approach addresses this by retaining the attributes of the point with the highest opacity per pixel on the same ray. + +To further unify the extracted position $(\sigma)$ , transformation $(r, s)$ , and color $(c, o)$ maps in a same feature space and use the existing image foundation models, we map the obtained UVGS $U \in \mathbb{R}^{M \times N \times 14}$ further to a 3-channel image $S \in \mathbb{R}^{M \times N \times 3}$ (termed as Super UVGS), using a Convolutional Neural Network (CNN). A multi-branch forward mapping network is employed to map $U \in \mathbb{R}^{M \times N \times 14}$ to the 3-channel Super UVGS $S \in \mathbb{R}^{M \times N \times 3}$ . We provide all technical details in Sec. 3.3 and the supplementary material. + +The Super UVGS representation effectively retains all the details of 3DGS attributes and can be directly utilized with existing widely available image-based models. We demonstrate this by showing perfect reconstruction of 3DGS object from Super UVGS image in the experiments section. This semantically structured representation $S$ offers both local and global correspondence in representing Gaussian attributes and needs relatively less storage. + +# 3.2. Dynamic GS Selection and Multiple Layers + +When projecting 3DGS points to UV maps using spherical mapping, multiple points may map to the same pixel in UV space as shown in Fig. 3. Two 3DGS points $(g_{1})$ and $(g_{2})$ map to the same pixel on UV map $(P_{a})$ causing many-to-one mapping issue. To address this, we propose a Dynamic Selection approach where each UV pixel retains the 3DGS attribute with the highest opacity intersecting the same ray. Using the same example in Fig. 3, if opacity $o_{1}$ of Gaussian $g_{1}$ is less than opacity $o_{2}$ of $g_{2}$ . Then only $g_{2}$ will be stored in the UV map at pixel $P_{a}$ . We observed that this method helps maintain the geometry and appearance of the 3DGS object while resolving many-to-one mapping issues with minimal quality loss. For more complex objects or real-world scene representation, we stack multiple such layers of UV maps, where each UVGS pixel now holds attributes of the top-K opacity values of 3DGS primitives. This can be accomplished by inscribing the 3DGS object inside multiple spheres where each sphere maps the 3DGS attribute corresponding to the top- $K^{th}$ opacity value along the same ray. More details on this are presented in the supplementary. To show the effectiveness of proposed UVGS + +![](images/25f85b471d8849db0732387367ab55d342304451b294b2457187827ee7586f1c.jpg) +Figure 4. Complex object reconstructions $(\mathrm{K} = 4)$ using pretrained image-based autoencoder. + +maps in capturing the intricacies of a complex objects, we used a pretrained image based autoencoder to reconstruct objects using a 4 layer UVGS as shown in Fig. 4. + +# 3.3. Mapping Networks + +Our goal is to bring the extracted UVGS maps to a common feature space to better represent the object collectively and to make the 14-channel UVGS representations $U \in \mathbb{R}^{M \times N \times 14}$ work with the widely available image based foundation models. To accomplish this, we map it to a 3-channel image which can be easily processed by the existing architectures while also maintaining the spatial correspondence. We design a simple yet effective multi-branch CNN to extract features from different UVGS attributes and map them to a 3-channel feature-rich image, termed Super UVGS. The structured UVGS maps provide local and global features that can be learned by a CNN. + +Forward Mapping The first layer is a set of three mapping branches for position, transform, and appearance $(\phi_P^f,\phi_T^f,\phi_A^f)$ respectively. We refer to them as position, transformation, and appearance branch. The position branch takes the mean position $(\sigma)$ as an input and processes it to give a position feature map $M_P$ . Similarly, the transformation branch takes the rotation $(r)$ and scale $(s)$ together to generate another feature map $M_T$ . The last, appearance branch takes the color $(c)$ and opacity $(o)$ together to produce another feature map $M_A$ . All the three features maps from position, transformation, and appearance branch are concatenated to get a final feature map, before passing them to the next module, called the Central Branch. The central branch $(\phi_C^f)$ is composed of multiple hidden Convolution layers, where each layer is followed by BatchNorm and ReLu activation. The last layer of the central branch is activated using tanh to ensure the Super UVGS does not take any ambiguous value resulting in gradient explosion or undesired artifacts. The obtained Super UVGS $S$ representation squeezes all the 3DGS attributes to a 3 dimensional image while also maintaining local and global structural correspondence among them. + +Inverse Mapping We design an inverse mapping network that aims to map the obtained 3-channel Super UVGS image $S \in \mathbb{R}^{M \times N \times 3}$ back to the UVGS maps to obtain each of the five different 3DGS attributes $\{\sigma, r, s, o, c\}$ . The inverse mapping network simply follows the forward mapping network architecture in the reverse order, where at first, we put the Central Branch $(\phi_{iC})$ followed by attribute specific position, transformation, and appearance branches + +$(\phi_{iP}, \phi_{iT}, \phi_{iA})$ . We provide more details on mapping network in the supplementary. + +Branched mapping layers: The rationale behind using branched mapping layers in both forward and reverse mapping networks is to prevent the incompatibility issues arising due the the different value distribution of 3DGS attributes. Note that the disparate distributions of values within each set of attributes in 3D Gaussian Splatting, (i.e., mean position, transformation, and color), pose a challenge to the model when processed collectively. For instance, neighboring Gaussians in UVGS maps show smooth changes in position and color values but typically have large variations in rotation, scale, and opacity values. This results in gradient anomalies and slow convergence. To address this, we propose a multi-branch network architecture, where attribute-specific branches implicitly learn to process these distinct attribute specific properties, focusing on their unique features before passing them to the central branch. The central branch receives a concatenated stack of processed attributes and exploits the correlation between them by extracting local feature correspondences. This information is then mapped to a 3-channel Super UVGS image, effectively capturing the complex relationships between the various attributes. This approach enables our network to manage diverse attribute distributions, resulting in faster convergence, improved accuracy, and specialized processing for each attribute set. + +Reconstruction Losses: Since the obtained UVGS maps have both local and global features, we opted for image-based losses to train the overall architecture. We use a set of Mean Squared Error (MSE) and Learned Perceptual Image Patch Similarity (LPIPS) [59] between the obtained UVGS from spherical mapping $U$ and the predicted UVGS $\hat{U}$ from inverse mapping network. We calculate the LPIPS loss over four attributes of 3DGS including mean position $(\sigma)$ , view independent color $(c)$ , scale $(s)$ , and rotation $(r)$ . The overall LPIPS loss for UV maps can be written as a linear sum of individual attribute loss terms as: + +$$ +\mathcal {L} _ {U V - l p i p s} = \mathcal {L} _ {\sigma} + \mathcal {L} _ {s} + \mathcal {L} _ {r} + \mathcal {L} _ {c} \tag {2} +$$ + +The overall loss function for the training can be written as: $\mathcal{L}_{uvgs} = \mathcal{L}_{mse} + \lambda \mathcal{L}_{UV - lpip}$ where $\lambda$ is a scalar and varied from 0 to 10 during the course of training. + +# 4. Experiments + +3DGS Dataset and UV maps To train the mapping networks and learn a latent space for unconditional and conditional sampling, we need large amount of 3DGS assets. However, there's a lack of such a large-scale dataset for high quality 3DGS assets. To this end, we create a custom large scale dataset by converting the Objaverse [9] meshes into 3DGS representation1. We start by designing a scene of 88 + +![](images/1dc5a626ac086a2d02768f0f41c01fbb14266a9fce609b793b8e31f33860ae36.jpg) +Super UVGS + +![](images/8beda8f23a9bc3c1caaf1bfd49d52e221b90ce4248c166c1d85c074f9b5a33cc.jpg) + +![](images/b0e1ce15fd53c2e7704c1288b2eef0719f0f1be6a21c0ff1db16d797276155a7.jpg) + +![](images/6687e906b1b3018f2dfa5af95dd4d2668fa69ec32da4bf312be89ffca22f45ad.jpg) + +![](images/78177fa3c088aad27f8d98524cabc7c526ac1a360d7605e948527a561786e5a8.jpg) + +![](images/04d8c3195569f5eaf9a5bf9380fc8b41626cf0a6543a7559151b9ddb142dbc38.jpg) + +![](images/5841673b1bd05609cbc173e2a859e8f1c93a0b5f13b07bf4384fe334df246596.jpg) + +![](images/65a1810f7586622267b5ec4b77f012e35bd7d97eae625847b5079e7ee338c6b8.jpg) + +![](images/ab61707530cf3160e9cd8c366d8970684b15a78a2237bcc6bb9e8c03cf0e7182.jpg) +Multiview + +![](images/1f4b60a82e998ae541fa551d8a0d44507e52b2229b5ae96487925ef701f3cb02.jpg) + +![](images/aa3f24f3951e68a4bb79b96328cd391d2f1797d7587555dae6eb6b71d856820b.jpg) + +![](images/2181461956050c1a50973f4c3c7a8bc54201dc66b3c66e8a442eb0da9ccbbfaa.jpg) + +![](images/8f4e3156691618f61e31cbb4ef8567fe1a0b27783896476258b0963e08533cc4.jpg) + +![](images/42aacd4f4ded688c31817634696bdc8790fba9fa44403be745688be3057e109d.jpg) + +![](images/09c5b76f22ca499839949af1195aecc4162f0f908164769d3ab94b8b557afabc.jpg) + +![](images/bf976d48495428f2ad3081772460cb5594a7770ccaf78138b6de932531834cd9.jpg) + +![](images/eb16a8e08262dd945922a110d24995e7017aaf8a4fa7a0b3c0c8362dd3ca364f.jpg) +Figure 5. Figure shows a wide variety of high-quality unconditional generation result from our method using Latent Diffusion Model. We train an LDM to randomly sample Super UVGS images from random noise. The Super UVGS can be converted to 3DGS object using inverse mapping network and inverse spherical projection. The unconditional generation model was trained on Objaverse dataset. + +![](images/1151107cb1adec15d4e96ec0eb05926ce551844693514ade22960aacac1ea5bf.jpg) + +![](images/e0578a159631be7f346e7a0101212201fc0a9a8aeb956047446db1297f51178b.jpg) + +![](images/8e08ccbeda13d769a5a7ec266201c75e0025ce62751f7faae2ae650511ce8a8b.jpg) + +![](images/a85e5dc621c7b3ae6c8421f1307599da18a1c0ce3ce54339efb51cc5c47db559.jpg) + +![](images/5c8310cfeb5762da162f009b7afca5bb3e703d79674a612be2e209717a2fbcbb.jpg) + +![](images/2b0e4f7e92927991c7af1e47af366ca08933e7d807c5c4d18424d48e3cdb5961.jpg) + +![](images/b9ec333bcceb0e99ac1946078b82d4e27cbc23874bb8235162c33d0ea7ab1f7d.jpg) + +cameras in a canonical space and use it to capture Objaverse objects from various angles covering all of the object views. The 88 rendered views from different angles are then used to train a 3DGS for $10\mathrm{K}$ iterations using [17]. This way, we create a high-quality and large-scale 3DGS dataset of $\sim 400K$ objects and scenes from Objaverse. We only use static scenes or objects from Objaverse. After fitting all the object to 3DGS representation, we convert the objects to the corresponding UV maps (i.e., UVGS) through Spherical Mapping as illustrated in Fig. 2. For the course of our experiments, we only map the objects to a single layer UV maps as it was sufficient to represent the general purpose Objaverse objects with minimal quality loss. Through mapping, we gathered a UVGS dataset of $\sim 400K$ maps. We fix the size of the UVGS maps to $512\times 512$ . Through our experiments, we found that UV maps of size $512\times 512$ are sufficient to represent objects in our dataset and capable of storing upto $262K$ unique Gaussians. Table 1 compares 1-4 layer UV maps. We also did experiments on ShapeNet [4] cars dataset for evaluation purposes. + +Baselines & Metrics: To evaluate the quality of reconstructed 3DGS objects from both Super UVGS image and Autoencoder latent space to 3D, we use PSNR and LPIPS [58]. The aim is to convert the given 3DGS object to UVGS, and then to Super UVGS, and further to Autoencoder's latent space and calculate the metrics from the reconstructions at every step to prove the proposed method doesn't significantly affect the quality of reconstructions, while also providing a structurally meaningful representation that is much compact and easier to use with existing image based models. We compare the generational capabilities of our method against various conditional and unconditional SOTA 3D object generation method including the ones using multiview rendering for optimization DiffTF [2], Get3D [12], methods trying to give structural representa + +Table 1. PSNR and LPIPS comparison for various reconstruction methods using UVGS and Super UVGS representations on Objverse Cars and Full datasets. AE, VAE, VQVAE are pretrained image based models. $K$ is the number of UVGS layers used. We also report the compression $\%$ (CP) compared to the fitted 3DGS. + +
MethodPSNR(C/F)LPIPS(C/F)CP(%)
3DGS34.6 / 34.20.02 / 0.020
UVGS (@K=1)31.3 / 31.10.06 / 0.0653.0
UVGS (@K=2)32.8 / 31.90.04 / 0.0545.6
UVGS (@K=4)34.2 / 33.20.02 / 0.0333.3
Super UVGS (@K=1)31.2 / 31.10.07 / 0.0889.7
AE (@K=1)30.9 / 30.80.07 / 0.0999.5
VAE (@K=1)30.6 / 30.90.07 / 0.0999.5
VQVAE (@K=1)30.3 / 30.10.08 / 0.1099.7
+ +tion to Gaussians, GaussianCube [55], and general purpose SOTA large 3D content generation models like DreamGaussian [42], LGM [44], and EG3D [3]. We also compare the quality of our generation results using FID and KID. + +Mapping Network Training Details We train the forward and inverse mapping networks to project the obtained UVGS maps $U \in \mathbb{R}^{M \times N \times 14}$ to Super UVGS image $S \in \mathbb{R}^{M \times N \times 3}$ , and back to the reconstructed UVGS maps $\hat{U} \in \mathbb{R}^{M \times N \times 14}$ . We provide an in-depth discussion of all implementation details in the supplementary material. + +# 4.1. UVGS AutoEncoder and 3DGS Compression + +The obtained Super UVGS image is a structurally meaningful representation that can have various applications in the generation and reconstruction of new 3D assets as it contains features that can be learned by the existing image based models. Through our experiments, we show that a 3-channel Super UVGS image can be directly reconstructed using a pretrained image based Autoencoders or VAEs without any fine-tuning. We tested on three different models in + +![](images/83c9eb6c0339a44fcc598dba59764b39bdde7c43b8a06f7adb4694b8aa648149.jpg) +DiffTF + +![](images/f8b27dba3749b293c7fcd612abc93e6a5692fdd722615c5523b2220b3661bf2b.jpg) +Figure 6. Comparison of unconditional 3D asset generation on the cars category with SOTA methods. Figure shows that DiffTF [2] produces low-quality, low-resolution cars lacking detail. While Get3D [12] achieve higher resolution, it suffers from 3D inconsistency, numerous artifacts, and lacks 3D detail. Similar issues are found in GaussianCube [55] along with symmetric inconsistency in the results. In contrast, our method generates high-quality, high-resolution objects that are 3D consistent with sharp and well-defined edges. + +![](images/5870965b0a011f348bd6406b300281c44a52ef0e9604240a6a8abb734b80b0af.jpg) +Get3D + +![](images/10cae7b5547d1e94a4698c66306f6605fb00ce2bdd7d2bf89132787d6e247fce.jpg) + +![](images/eb419d82e0e5c06e732353251b34b4192c4150959452a4d4f001551c659b0e60.jpg) +GaussianCube + +![](images/98bd3dec6d91c19e4fa3ea1f3ead282225d12a61228adc786a1b1976574945a7.jpg) + +![](images/8504566ec2d121a5832dae4248bf1c9e5c94d3ef4a5b3bfc24d64b1912390500.jpg) + +![](images/4924ea712094fd45596345baf08169c651010493e4047c2eb7c7d6a6db3ccd75.jpg) + +![](images/239ff5dad6d501015f121e0ed1457f0d82ddc7d3cbc76c5cf93b672da90b7dce.jpg) + +![](images/62cbbd6a34134c889925976552340213dabdc4fa60cf4a94096a2a5e34ed3a0a.jpg) + +![](images/8425730ae76e0693e63bf58c18b9bf1988ac9693a8bc038f68565d3062b40028.jpg) + +![](images/372dcede4f068eee9c86a879d1329201caa180f78c6b746fa0fc5f7ee45f60f2.jpg) +UVGS (ours) + +![](images/7088103b939532b32cad1d42762c80488345d78ea510f63ecb785ebfdf45d5ab.jpg) + +![](images/59fdcae91097b03e80b388e4c84b9c9f72c98e6d66420a19d65c733754a7a622.jpg) + +![](images/0a70511b2ed037342fd7e43ca5bd700ca854e7feaa68a04ff3b8e71f47e76d70.jpg) + +![](images/9ccc93f755797d2ed09bbc5b39ddd31e850e21d441d4d40aeaf8236c971807d8.jpg) +Figure 7. We compare the performance of our model against various SOTA methods for text-conditional object synthesis. Our method not only generates high-quality assets for simpler objects, but also for complicated objects with intricate geometry. + +cluding image AE, KL-VAE [18], VQVAE [46] and each performed quite well without any significant quality loss. The reconstruction PSNR and LPIPS values are presented in Table 1. This means we can now leverage the powerful compression capabilities of image based Autoencoders to compress the storage requirements of 3DGS by more than $99\%$ . We have shown the storage comparison results in Table 1. It is interesting to note that the Super UVGS representation itself can be used to compress the memory requirement for storing 3DGS object by up to $89.7\%$ . + +# 4.2. Unconditional & Conditional Generation + +We aim to show the effectiveness of Super UVGS representation for directly generating 3DGS objects from a learned latent space. We consider Super UVGS images as a compact and structured proxy for representing 3DGS objects as it maintains the 3D object information while also providing learnable features. The existing methods fail to directly generate a large number of Gaussians (e.g. $100\mathrm{K}+$ ) to represent objects with sufficient quality either due to the lack of 3D generative model architectures that support such large number of unstructured points, or due to the lack of a large 3DGS dataset [28, 40, 44, 55, 63]. We leverage the + +Super UVGS representation and use the existing 2D image generative models like Diffusion Models [15] for this task. Specifically, to train a generative model capable of randomly sampling new high-quality 3DGS assets for various downstream tasks, we use an unconditional Latent Diffusion Model (LDM) [36] on the obtained Super UVGS images. As illustrated in Section 4.1, we can use a pretrained image VAE to map the Super UVGS image to a latent space and reconstruct back. Hence, we only train a LDM on the latent space. More implementation details are provided in the supplementary. + +Unconditional LDM: To design a generative model capable of randomly sampling new high-quality 3DGS assets for various downstream tasks, we train an unconditional LDM [36] on the learned Super UVGS images. Following [33, 36, 57], we use DDIM [15] for faster and consistent sampling with up to 1000 time steps used in the forward diffusion process, and 20 during denoising. Once trained, the model is used to randomly sample Super UVGS images, resulting in high quality 3DGS assets through inverse mapping. Results are presented in Fig 5. We demonstrate that our method inherently learns to generate multiview consistent images due to the powerful Super UVGS representation unlike most prior works using rendering-based losses. + +Conditional LDM: Similar to unconditional generation, we also trained a text-conditioned LDM following the SD's [33, 36, 57] pipeline and using the predicted text for our dataset. The trained model can be used to generate high-quality text-conditioned 3DGS assets that are multiview consistent. The results are demonstrated in Fig 6. + +The above experiments prove the effectiveness of our proposed Super UVGS representation in 3D object synthesis using widely available 2D image models. It also highlights that this compact 3-channel Super UVGS representation stores not just the spatial correspondence among different pixels, but also the rich 3D information of the objects. This way, we can easily convert a 3D asset generation problem into a 2D image generation problem without the use of any complex 3D architecture to handle large amount of un + +![](images/e5569757356944ee79466c8fb221ce34c48934ad1c6558ee3d05fe8959d08516.jpg) +Figure 8. 3DGS Inpainting: We present one of the first inpainting results on 3DGS directly leveraging the Super UVGS images and the denoising capabilities of diffusion models. + +structured and permutation invariant 3DGS primitives, and neither relying upon computationally expensive multiview rendering or SDS loss. Baseline comparison for unconditional and conditional generation is presented in Table 2. + +# 4.3. 3DGS Inpainting + +Leveraging the powerful Super UVGS representation, we present in Fig. 8 one of the first experiments on inpainting 3DGS directly without using any multiview rendering or distilling information from diffusion models. We try to recover the missing Gaussians by leveraging the denoising capabilities of LDM and trying to predict the missing corresponding parts of the Super UVGS image. We believe, this can have potential applications in sparse view reconstruction. More details are given in the supplementary. + +# 4.4. Ablation Studies & Discussion + +We conduct exhaustive ablation studies to justify some of our framework's design choices including the effect of branching in mapping networks, the use of single layer UVGS maps, and the resolution of UVGS maps. We performed our experiments on our custom Objaverse [9] 3DGS dataset and evaluate the performance of our model in terms of PSNR, SSIM, and LPIPS. The results are presented in Table 3. From the table, it can be seen that by using four layers of UVGS maps $(\mathrm{K} = 4)$ , we can almost match the reconstruction quality of fitted 3DGS results, while we realized that simply with a single layer UVGS, we are able to maintain the overall geometry and appearance of the object for our dataset with a PSNR of more than 30. We also compared the reconstruction performance of our method with and without using branching in the mapping network. It can be clearly seen that using branching network significantly increases the reconstruction quality from Super UVGS space. The main reasoning behind this is the specialization of attributes that the branching provides to individually process each attribute first. + +Table 2. We compare the FID and KID of unconditional generation using the current SOTA methods on 20K randomly generated samples from each method and ours. We also compare our method against SOTA text-conditioned generation frameworks on CLIP Score for 10K generated objects from each method. + +
Unconditional GenerationText-Conditioned Generation
MethodFID ↓KID ↓MethodCLIP Score ↑
Get3D [12]53.174.19DreamGaussian [42]28.51
DiffTF [2]84.578.73Shap. E [6]30.53
EG3D [3]74.516.62LGM [44]30.74
GaussianCube34.673.72GaussianCube [55]30.34
UVGS (Ours)26.203.24UVGS (Ours)32.62
+ +Table 3. We present quantitative ablation study for number of UVGS layers (K), UVGS map resolution, and the effect of branching in mapping network on the Objaverse 3DGS dataset. + +
MethodPSNRLPIPSUVGS SizePSNRLPIPS
UVGS @K=131.10.06512 × 512 (@K=1)31.10.08
UVGS @K=231.90.05256 × 256 (@K=1)28.20.23
UVGS @K=433.20.03Without Branching27.80.31
+ +Limitations & Future Work: While single layer UVGS images can recover the geometry of the object, they sometimes suffer in terms of appearance and the generated objects might look washed out. We believe this can be solved by using a multi-layer UVGS maps. Similarly, the single layer UVGS map is limited to representing simpler everyday objects, and may not be sufficient to represent highly-detailed and complex objects or scenes. In the future, we want to extend this framework to learn features for real-world scenes and complex objects like a human head with multi-layer UV mapping. We also want to make this representation more efficient by better utilizing the empty pixels of UVGS maps and Super UVGS images while also maintaining the underlying features and 3D information. + +# 5. Conclusion + +We introduced a novel method to solve the underlying issues with 3D Gaussian Splatting (3DGS) that prevent the direct integration of them with the large number of existing image foundational models. We proposed UVGS - a structured representation for 3DGS obtained by spherical mapping of 3DGS primitives to UV maps. We further squeezed the multi-attribute UVGS maps to a 3-channel unified and structured Super UVGS image, which not only maintains the 3D structural information of the object, but also provides a compact feature space for 3DGS attributes. The obtained Super UVGS images are directly integrated with the existing image foundational models for 3DGS compression and unconditional and conditional generation using diffusion models. Leveraging these Super UVGS images, we showed one of the first inpainting experiments on 3DGS. + +ACKNOWLEDGEMENTS A part of this work was supported by NSF CAREER grant 2143576 and ONR DURIP grant N00014-23-1-2804. + +# References + +[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855-5864, 2021. 2 +[2] Ziang Cao, Fangzhou Hong, Tong Wu, Liang Pan, and Ziwei Liu. Large-vocabulary 3d diffusion model with transformer. arXiv preprint arXiv:2309.07920, 2023. 3, 6, 7, 8 +[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16123-16133, 2022. 2, 6, 8 +[4] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 3, 6 +[5] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision, pages 333-350. Springer, 2022. 2 +[6] Minghao Chen, Junyu Xie, Iro Laina, and Andrea Vedaldi. Shap-editor: Instruction-guided latent 3d editing in seconds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26456-26466, 2024. 8 +[7] Zilong Chen, Feng Wang, Yikai Wang, and Huaping Liu. Text-to-3d using gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21401-21412, 2024. 3 +[8] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4456-4465, 2023. 2 +[9] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli Vanderbilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 3, 5, 8 +[10] Slava Elizarov, Ciara Rowles, and Simon Donne. Geometry image diffusion: Fast and data-efficient text-to-3d with image-based surface representation. arXiv preprint arXiv:2409.03718, 2024. 2 +[11] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. 2 + +[12] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841-31854, 2022. 3, 6, 7, 8 +[13] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371, 2023. 2 +[14] Xianglong He, Junyi Chen, Sida Peng, Di Huang, Yangguang Li, Xiaoshui Huang, Chun Yuan, Wanli Ouyang, and Tong He. Gvgen: Text-to-3d generation with volumetric representation. In European Conference on Computer Vision, pages 463–479. Springer, 2025. 2, 3 +[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3, 7 +[16] Han Huang, Yulun Wu, Junsheng Zhou, Ge Gao, Ming Gu, and Yu-Shen Liu. Neusurf: On-surface priors for neural surface reconstruction from sparse input views. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2312-2320, 2024. 2 +[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2, 6 +[18] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 7 +[19] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 2 +[20] Shujuan Li, Junsheng Zhou, Baorui Ma, Yu-Shen Liu, and Zhizhong Han. Neaf: Learning neural angle fields for point normal estimation. In Proceedings of the AAAI conference on artificial intelligence, pages 1396-1404, 2023. 2 +[21] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 2 +[22] Baorui Ma, Haoge Deng, Junsheng Zhou, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Geodream: Disentangling 2d and geometric priors for high-fidelity and consistent 3d generation. arXiv preprint arXiv:2311.17971, 2023. 2 +[23] Qi Ma, Yue Li, Bin Ren, Nicu Sebe, Ender Konukoglu, Theo Gevers, Luc Van Gool, and Danda Pani Paudel. Shapes-plat: A large-scale dataset of gaussian splats and their self-supervised pretraining. arXiv preprint arXiv:2408.10906, 2024. 2, 3 +[24] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 2 + +[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2 +[26] Chaerin Min and Srinath Sridhar. Genheld: Generating and editing handheld objects. arXiv preprint arXiv:2406.05059, 2024. 2 +[27] Chaerin Min, Sehyun Cha, Changhee Won, and Jongwoo Lim. Tsdf-sampling: Efficient sampling for neural surface field using truncated signed distance field. arXiv preprint arXiv:2311.17878, 2023. 2 +[28] Yuxuan Mu, Xinxin Zuo, Chuan Guo, Yilin Wang, Juwei Lu, Xiaofeng Wu, Songcen Xu, Peng Dai, Youliang Yan, and Li Cheng. Gsd: View-guided gaussian splatting diffusion for 3d reconstruction. arXiv preprint arXiv:2407.04237, 2024. 2, 3, 7 +[29] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulo, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4328-4338, 2023. 2 +[30] Francesco Palandra, Andrea Sanchietti, Daniele Baieri, and Emanuele Rodola. Gsedit: Efficient text-guided editing of 3d objects via gaussian splatting. arXiv preprint arXiv:2403.05154, 2024. 2 +[31] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 3 +[32] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 2 +[33] Aashish Rai and Srinath Sridhar. Egosonics: Generating synchronized audio for silent egocentric videos. arXiv preprint arXiv:2407.20592, 2024. 7 +[34] Aashish Rai, Hiresh Gupta, Ayush Pandey, Francisco Vicente Carrasco, Shingo Jason Takagi, Amaury Aubel, Daeil Kim, Aayush Prakash, and Fernando De la Torre. Towards realistic generative 3d face models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3738-3748, 2024. 2 +[35] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject-driven text-to-3d generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2349–2359, 2023. 2 +[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 7 +[37] Li Shen and Fillia Makedon. Spherical mapping for processing of 3d closed surfaces. Image and vision computing, 24 (7):743-761, 2006. 2 + +[38] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20875-20886, 2023. 2 +[39] Jingxiang Sun, Bo Zhang, Ruizhi Shao, Lizhen Wang, Wen Liu, Zhenda Xie, and Yebin Liu. Dreamcraft3d: Hierarchical 3d generation with bootstrapped diffusion prior. arXiv preprint arXiv:2310.16818, 2023. 2 +[40] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10208-10217, 2024. 2, 3, 7 +[41] Fariborz Taherkhani, Aashish Rai, Quankai Gao, Shaunak Srivastava, Xuanbai Chen, Fernando De la Torre, Steven Song, Aayush Prakash, and Daeil Kim. Controllable 3d generative adversarial face model via disentangling shape and appearance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 826-836, 2023. 2 +[42] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 2, 3, 6, 8 +[43] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22819–22829, 2023. 2 +[44] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1-18. Springer, 2025. 6, 7, 8 +[45] Zhicong Tang, Shuyang Gu, Chunyu Wang, Ting Zhang, Jianmin Bao, Dong Chen, and Baining Guo. Volumediffusion: Flexible text-to-3d generation with efficient volumetric encoder. arXiv preprint arXiv:2312.11459, 2023. 2 +[46] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 7 +[47] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4563-4573, 2023. 2 +[48] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3295–3306, 2023. 2 +[49] Yuxuan Wang, Xuanyu Yi, Zike Wu, Na Zhao, Long Chen, and Hanwang Zhang. View-consistent 3d editing with gaussian splatting. In European Conference on Computer Vision, pages 404-420. Springer, 2025. 2 + +[50] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2 +[51] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024. 2 +[52] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation. arXiv preprint arXiv:2403.14621, 2024. 2, 3 +[53] Xingguang Yan, Han-Hung Lee, Ziyu Wan, and Angel X Chang. An object is worth 64x64 pixels: Generating 3d object via image diffusion. arXiv preprint arXiv:2408.03178, 2024. 2 +[54] Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian-dreamer: Fast generation from text to 3d gaussian splattering with point cloud priors. arXiv preprint arXiv:2310.08529, 2023. 2 +[55] Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: Structuring gaussian splatting using optimal transport for 3d generative modeling. arXiv preprint arXiv:2403.19655, 2024. 2, 3, 6, 7, 8 +[56] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 2, 3 +[57] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 7 +[58] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[59] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5 +[60] Junsheng Zhou, Yu-Shen Liu, and Zhizhong Han. Zero-shot scene reconstruction from single images with deep prior assembly. arXiv preprint arXiv:2410.15971, 2024. 2 +[61] Junsheng Zhou, Weiqi Zhang, and Yu-Shen Liu. Diff-gfs: Functional gaussian splatting diffusion. arXiv preprint arXiv:2410.19657, 2024. 2, 3 +[62] Junsheng Zhou, Weiqi Zhang, Baorui Ma, Kanle Shi, YuShen Liu, and Zhizhong Han. Udiff: Generating conditional unsigned distance fields with optimal wavelet diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21496-21506, 2024. 2 + +[63] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10324-10335, 2024. 2, 3, 7 +[64] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10324-10335, 2024. 2 \ No newline at end of file diff --git a/2025/UVGS_ Reimagining Unstructured 3D Gaussian Splatting using UV Mapping/images.zip b/2025/UVGS_ Reimagining Unstructured 3D Gaussian Splatting using UV Mapping/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e1e70aeeefd31a4ee246ecdf84d9e3d23294ebf9 --- /dev/null +++ b/2025/UVGS_ Reimagining Unstructured 3D Gaussian Splatting using UV Mapping/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a5c425dfdb2fc8789eecabe080628fe00f4a78c0bd9a12f3cbe3f0dff11440d +size 470371 diff --git a/2025/UVGS_ Reimagining Unstructured 3D Gaussian Splatting using UV Mapping/layout.json b/2025/UVGS_ Reimagining Unstructured 3D Gaussian Splatting using UV Mapping/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e337347c9436d16f45ae8bead7fa85b8690d30ba --- /dev/null +++ b/2025/UVGS_ Reimagining Unstructured 3D Gaussian Splatting using UV Mapping/layout.json @@ -0,0 +1,9951 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 103, + 544, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 103, + 544, + 122 + ], + "spans": [ + { + "bbox": [ + 67, + 103, + 544, + 122 + ], + "type": "text", + "content": "UVGS: Reimagining Unstructured 3D Gaussian Splitting using UV Mapping" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 141, + 148, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 141, + 148, + 157 + ], + "spans": [ + { + "bbox": [ + 75, + 141, + 148, + 157 + ], + "type": "text", + "content": "Aashish Rai" + }, + { + "bbox": [ + 75, + 141, + 148, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 170, + 142, + 313, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 142, + 313, + 171 + ], + "spans": [ + { + "bbox": [ + 170, + 142, + 313, + 171 + ], + "type": "text", + "content": "Dilin Wang" + }, + { + "bbox": [ + 170, + 142, + 313, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 170, + 142, + 313, + 171 + ], + "type": "text", + "content": " Mihir Jain" + }, + { + "bbox": [ + 170, + 142, + 313, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 170, + 142, + 313, + 171 + ], + "type": "text", + "content": " Srinath Sridhar" + }, + { + "bbox": [ + 170, + 142, + 313, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 323, + 143, + 441, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 143, + 441, + 172 + ], + "spans": [ + { + "bbox": [ + 323, + 143, + 441, + 172 + ], + "type": "text", + "content": "Nikolaos Sarafianos2 \nAayush Prakash2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 464, + 143, + 534, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 464, + 143, + 534, + 156 + ], + "spans": [ + { + "bbox": [ + 464, + 143, + 534, + 156 + ], + "type": "text", + "content": "Kefan Chen" + }, + { + "bbox": [ + 464, + 143, + 534, + 156 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 208, + 180, + 301, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 180, + 301, + 196 + ], + "spans": [ + { + "bbox": [ + 208, + 180, + 301, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 208, + 180, + 301, + 196 + ], + "type": "text", + "content": "Brown University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 318, + 181, + 414, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 181, + 414, + 195 + ], + "spans": [ + { + "bbox": [ + 318, + 181, + 414, + 195 + ], + "type": "text", + "content": "Meta Reality Labs" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 202, + 204, + 405, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 204, + 405, + 216 + ], + "spans": [ + { + "bbox": [ + 202, + 204, + 405, + 216 + ], + "type": "text", + "content": "https://aashishrai3799.github.io/uvgs" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 68, + 228, + 541, + 395 + ], + "blocks": [ + { + "bbox": [ + 68, + 228, + 541, + 395 + ], + "lines": [ + { + "bbox": [ + 68, + 228, + 541, + 395 + ], + "spans": [ + { + "bbox": [ + 68, + 228, + 541, + 395 + ], + "type": "image", + "image_path": "5f9e3c045a3b9caf6b11549a80329ba7ba55bde2f88692372e9d39e48b98fb56.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 404, + 555, + 449 + ], + "lines": [ + { + "bbox": [ + 54, + 404, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 54, + 404, + 555, + 449 + ], + "type": "text", + "content": "Figure 1. We propose UVGS - an structured image-like representation for 3DGS obtained by spherical mapping of its primitives. The obtained UVGS maps can be further unified to a 3-channel Super UVGS image to bridging the gap between 3DGS and existing image foundation models. We show Super UVGS can compress the 3DGS assets using pretrained image Autoencoders, and for direct unconditional and conditional 3DGS object generation using diffusion models. We also present one of the first experiments on 3DGS inpainting." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 151, + 455, + 200, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 455, + 200, + 467 + ], + "spans": [ + { + "bbox": [ + 151, + 455, + 200, + 467 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 54, + 474, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 474, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 474, + 297, + 714 + ], + "type": "text", + "content": "3D Gaussian Splatting (3DGS) has demonstrated superior quality in modeling 3D objects and scenes. However, generating 3DGS remains challenging due to their discrete, unstructured, and permutation-invariant nature. In this work, we present a simple yet effective method to overcome these challenges. We utilize spherical mapping to transform 3DGS into a structured 2D representation, termed UVGS. UVGS can be viewed as multi-channel images, with feature dimensions as a concatenation of Gaussian attributes such as position, scale, color, opacity, and rotation. We further find that these heterogeneous features can be compressed into a lower-dimensional (e.g., 3-channel) shared feature space using a carefully designed multi-branch network. The compressed UVGS can be treated as typical RGB images. Remarkably, we discover that typical VAEs trained with latent diffusion models can directly generalize to this new representation without additional training. Our novel representation makes it effortless to leverage foundational 2D models, such as diffusion models, to directly model 3DGS. Additionally, one can simply increase the 2D UV resolution" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 456, + 555, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 555, + 552 + ], + "type": "text", + "content": "to accommodate more Gaussians, making UVGS a scalable solution compared to typical 3D backbones. This approach immediately unlocks various novel generation applications of 3DGS by inherently utilizing the already developed superior 2D generation capabilities. In our experiments, we demonstrate various unconditional, conditional generation, and inpainting applications of 3DGS based on diffusion models, which were previously non-trivial." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 555, + 395, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 555, + 395, + 566 + ], + "spans": [ + { + "bbox": [ + 314, + 555, + 395, + 566 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "The creation of high-quality 3D content is essential in applications like virtual reality, game design, robotics, and movie production, where realistic 3D representations play a critical role. Typical 3D representations like Neural Radiance Fields (NeRF) [25] are promising but require substantial computational resources, limiting their scalability for real-time applications. Moreover, NeRF is an implicit representation, which makes editing and manipulation challenging. Recently, 3D Gaussian Splatting (3DGS) [17] emerged as a compelling alternative, enabling efficient and high-fidelity 3D rendering through a large set of Gaussian primitives that model spatial and visual properties. As an explicit represen" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5927" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "tation, 3DGS offers several advantages over NeRF. However, while 3DGS offers benefits in terms of speed and visual quality, its unstructured, permutation-invariant nature presents significant challenges for generative tasks. Much like point clouds, it lacks a coherent spatial structure, impeding its integration with conventional image-based generative models. This lack of structure and coherence among primitives hinders the application of image-based generative models [23, 55, 61], which rely on structured data representations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 194, + 295, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 295, + 350 + ], + "type": "text", + "content": "Previous methods have tackled these challenges by transforming 3DGS into structured formats, such as voxel grids [8, 14, 55] or image-based representations like Splatter Image [40] or triplanes [64]. Other approaches employ diffusion models to directly predict 3DGS attributes [28]. These methods, while achieving impressive visual results, often require substantial computational resources, memory-intensive multi-view rendering, complex architectures limiting their scalability and flexibility for high-fidelity generation. Generating and processing 3DGS directly by efficiently utilizing modern generative models like Variational Autoencoders (VAEs) and diffusion models is limited as the neural networks are not permutation invariant." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 352, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 295, + 519 + ], + "type": "text", + "content": "To address these shortcomings, we introduce UV Gaussian Splatting (UVGS), which provides a structured transformation of 3D Gaussian primitives into a 2D representation while preserving essential 3D information. We use spherical mapping [37] that inscribes Gaussian splats in a spherical surface, and projects attributes like position, rotation, scale, opacity, and color into an organized 14-channel image-like UV map. This mapping introduces spatial structure, resolving issues of permutation invariance by introducing local correspondences between neighboring Gaussians and global coherence across the entire 3D object. The result is a representation that functions as a \"3D representation\" structured in a 2D map format, enabling compatibility with powerful image-based neural network architectures." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": "While UVGS introduces structure into 3D Gaussian Splitting, its full 14-channel attribute-specific representation presents challenges for direct integration with pretrained 2D generative models, as these models typically expect a simpler, image-compatible data. Each of its heterogeneous attributes—position, color, and transformation—has its own distinct distribution and resides in a separate feature space, making it challenging to represent the 3D object in a unified shared space. To address this, we introduce Super UVGS, a compact 3-channel representation that unifies these diverse attributes into a cohesive format. Using a carefully designed multi-branch mapping network, Super UVGS consolidates the distinct attribute spaces into a shared feature space, enabling a more collective representation of the object. This unified transformation not only facilitates zero-shot compatibility with pretrained 2D models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "content": "but also optimizes memory usage and computational efficiency, making Super UVGS highly practical for large-scale 3D tasks. Unlike previous approaches that use Triplanes, voxels, occupancy grid, neural fields etc. and require specialized 3D architectures to train on 3D data, UVGS effortlessly leverages widely available pretrained 2D foundational models. This zero-shot generalization capability allows UVGS to fully benefit from priors learned in 2D domains from large amount of data, improving both flexibility and scalability. To sum up our main contributions are:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 193, + 553, + 383 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 193, + 553, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 193, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 314, + 193, + 553, + 262 + ], + "type": "text", + "content": "- Efficient Structured Representation of 3DGS: We present UVGS, an image-like representation that solves permutation invariance and unstructured nature of discrete 3DGS through spherical mapping, making direct feature extraction possible by organizing unordered points into a coherent 2D representation compatible with 2D models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 264, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 264, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 264, + 553, + 335 + ], + "type": "text", + "content": "- Compact and Scalable Super UVGS Representation: To address scalability while dealing with large scale 3DGS points and enabling the direct integration of pre-trained 2D foundation models, we introduce Super UVGS - a low-dimensional version of UVGS maps that retains high fidelity features while reducing memory overhead." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 336, + 553, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 336, + 553, + 383 + ], + "spans": [ + { + "bbox": [ + 314, + 336, + 553, + 383 + ], + "type": "text", + "content": "- Diverse 3D Applications: Our approach unlocks seamless integration of 3DGS with pre-trained 2D foundation models for various tasks, including unconditional and conditional generation of 3DGS." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 394, + 400, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 394, + 400, + 406 + ], + "spans": [ + { + "bbox": [ + 314, + 394, + 400, + 406 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 415, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 714 + ], + "type": "text", + "content": "3D Generative/Reconstruction Models for Objects: Generation or reconstruction of 3D assets has been a long standing task [3, 21, 24, 26, 27, 31, 34, 35, 39, 41, 43, 50, 54]. Previous reconstruction approaches like NeRF [1, 25] are often slow and do not provide a defining geometry [5, 11, 16, 20, 22, 24, 32, 35, 48, 60, 62]. Advancements in the field led to the emergence of explicit voxel grid based representations that encode colors and opacities directly [8, 29, 45]. These approaches achieve significant speed ups compared to the NeRF based approaches, but they can't produce high fidelity assets due to the low resolution of voxel grids. On the other hand, triplane representation [13, 38, 47, 63] provides a trade-off between the quality and memory utilization. Another line of work [10, 53] splits the input mesh into different patches and simplifies the object generation problem to an image generation problem. However such methods rely on either cutting through the mesh to create a geometry image [10] or rely on an existing UV representation of the geometry and utilize subset of existing UV islands [53] resulting in loss of details. Recently, there has been a notable advancement in 3D Gaussian Splatting (3DGS) for the representation of objects and scenes leading to the emergence of 3DGS showcasing impressive real-time results in reconstruction and generation tasks [14, 17, 19, 30, 42, 49, 51, 52, 54, 56, 63]. Recent" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5928" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 70, + 534, + 259 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 534, + 259 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 534, + 259 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 534, + 259 + ], + "type": "image", + "image_path": "4067950a35b01263b3e660b2b4fc16218bdce22a67eca1d67abe820a53df3b98.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 263, + 555, + 309 + ], + "lines": [ + { + "bbox": [ + 54, + 263, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 54, + 263, + 555, + 309 + ], + "type": "text", + "content": "Figure 2. The input 3DGS object is first converted to UVGS maps through spherical mapping. We use a multibranch forward mapping network to convert the obtained 14-channel UVGS to a compact 3-channel Super UVGS image. This represents the 3DGS object in a structured manner and can be used with image foundation models for reconstruction or generation. The Super UVGS is mapped back to UVGS through branched inverse mapping, which in turn can be reconstructed back to the 3DGS object through inverse spherical mapping." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 315, + 294, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 315, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 54, + 315, + 294, + 434 + ], + "type": "text", + "content": "advances in the 3D generative models for asset synthesis using the existing geometries like NeRF, voxel grids, or triplane geometries [2, 12, 42, 52, 56, 63] leverage generative models [15, 36] and the existing 3D datasets [4, 9]. However, most of the works employing 3DGS or other representations use multiview rendering and Score Distillation Sampling (SDS) to achieve convincing generation and reconstruction capabilities [7, 31]. These approaches demand high memory and compute resources and are often quite slow in optimizing due to per scene optimization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 436, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 436, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 54, + 436, + 295, + 628 + ], + "type": "text", + "content": "Giving Structures to Discrete Gaussians: Although, 3DGS has led to breakthrough in the reconstruction field by demonstrating superior performance in multiple domains, the generation of 3DGS directly remains challenging due to its discreteness and unstructured nature [55, 61]. These characteristics present substantial challenges when integrating them with conventional computer vision models, like Autoencoders and generative models [61]. The research in direct learning of trained 3DGS primitives is largely unexplored [23]. Some efforts attempt to address this by directly predicting 3DGS attributes using diffusion models [28] while others like Splatter Image [40] project Gaussian objects into image-based representations through direct 3D-unaware projection. These methods struggle with maintaining multiview consistency, as the model only infers seen poses correctly, while hallucinating for unseen poses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "text", + "content": "Concurrent works [14, 55] follow the voxel-based representations to transport Gaussians into structural voxel grids with volume generation models for generating Gaussians. However, these methods are computationally expensive for high-resolution voxels, face difficulties in preserving high-quality Gaussian reconstructions due to information loss during voxelization. DiffGS [61] tries to solve the above" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 315, + 555, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 555, + 471 + ], + "type": "text", + "content": "issues by proposing three continuous functions to represent 3DGS. However, it is limited to only category-level generation and learning generic probability functions for all the categories poses significant compute and design challenges. In contrary, we introduce an efficient way to give structures to discrete Gaussians by taking inspiration from the developments in 3D graphics. Our method does not require any learning to map an unstructured set of Gaussians to this efficient and structured representation (termed UVGS). The proposed representation provides local and global correspondence among different Gaussian points making the widely available existing computer vision frameworks learn and extract underlying features from them." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 476, + 397, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 397, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 397, + 491 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "content": "Preliminaries: 3DGS represents an object or a scene with a collection of Gaussians primitives to model the geometry and view-dependent appearance. For a 3DGS set, " + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "inline_equation", + "content": "G = \\{g_i\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "content": ", representing an object with " + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "content": " individual Gaussians, the geometry of the " + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "content": " Gaussian is explicitly parameterized via 3D covariance matrix " + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\Sigma_i" + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "content": " and it's center " + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\sigma_i \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "content": " as: " + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "inline_equation", + "content": "g_i(x) = e^{(-\\frac{1}{2}(x - \\sigma_i)^T\\Sigma^{-1}(x - \\sigma_i))}" + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "content": " where, the covariance matrix " + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\Sigma_i = r_is_is_i^T r_i^T" + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "content": " is factorized into a rotation matrix " + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "inline_equation", + "content": "r_i \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "content": " and a scale matrix " + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "inline_equation", + "content": "s_i \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "content": ". The appearance of the " + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "inline_equation", + "content": "i - th" + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "content": " Gaussian is represented by a color value " + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "inline_equation", + "content": "c_i \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "content": " and an opacity value " + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "inline_equation", + "content": "o_i \\in R" + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "content": ". In practice, the color is represented by a series of Spherical Harmonics (SH) coefficients, but for simplicity, we represent the view-independent color by just RGB values. Thus, a single Gaussian can be represented by a set of five attributes as " + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "inline_equation", + "content": "g_i = \\{\\sigma_i, r_i, s_i, o_i, c_i\\} \\in \\mathbb{R}^{14}" + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "content": ", and the entire 3DGS can be represented by a set of " + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "content": " such Gaussians as: " + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "inline_equation", + "content": "G = \\{\\{\\sigma_i, r_i, s_i, o_i, c_i\\}\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 497, + 556, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5929" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 168, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 168, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 168, + 85 + ], + "type": "text", + "content": "3.1. Spherical Mapping" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 221 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 221 + ], + "type": "text", + "content": "3DGS is represented as a permutation invariant set with no structural correspondence among different Gaussians " + }, + { + "bbox": [ + 55, + 90, + 296, + 221 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 55, + 90, + 296, + 221 + ], + "type": "text", + "content": ", making it challenging to extract meaningful features from this set containing a few hundred thousands of them using neural networks. To address this, we introduce a novel representation that gives structure to this unstructured set of points and solves the permutation invariance issue for faster and better feature extraction. We propose to accomplish this by employing spherical mapping to map the 3DGS primitives to an image-like representation that is both invariant to random shuffling of 3DGS points and well structured." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 222, + 296, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 222, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 296, + 342 + ], + "type": "text", + "content": "We begin the mapping by inscribing the 3DGS object into a sphere with the same center as the object in the canonical space. Inscribing a 3DGS object into a sphere involves enclosing the object within a sphere. This begins by determining the geometric center of the object. The next step is to calculate the radius of the sphere, which is achieved by measuring the Euclidean distance from the center to the farthest point on the object. The radius of the sphere is defined such that the sphere fully encloses the object. The sphere acts as a bounding volume for the entire object." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "text", + "content": "We consider each Gaussian " + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "text", + "content": " in 3D to be centered at the mean position represented by " + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "text", + "content": " with Cartesian coordinates " + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "inline_equation", + "content": "(x_{i},y_{i},z_{i})" + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "text", + "content": ". The aim is to get the spherical coordinates " + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "inline_equation", + "content": "(\\rho_{i},\\theta_{i},\\phi_{i})" + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "text", + "content": " for each Gaussian " + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "text", + "content": ". To do so, we calculate the azimuthal " + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "text", + "content": " and polar " + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "text", + "content": " angles for each " + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "text", + "content": " along with the distance from the origin to the point, " + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\rho_{i}" + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "text", + "content": ". The spherical radius is defined as " + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\rho_{i} = \\sqrt{x_{i}^{2} + y_{i}^{2} + z_{i}^{2}}" + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "text", + "content": ", the azimuthal angle as " + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\theta_{i} = \\tan^{-1}(y_{i},x_{i})" + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "text", + "content": ", while the polar angle as " + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\phi_{i} = \\cos^{-1}(z_{i},\\rho_{i})" + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "text", + "content": ". The azimuthal and polar angles are then normalized, such that we can map them on a 2D UV map of " + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "inline_equation", + "content": "M\\times N" + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "text", + "content": " dimensionality with 14-channels. " + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "text", + "content": " are converted to degrees and mapped to UV image coordinates: " + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\theta_{i\\mathrm{scaled}} = \\left\\lfloor \\frac{\\pi + \\theta_i}{2\\pi}\\times \\mathrm{M}\\right\\rfloor" + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\phi_{i\\mathrm{scaled}} = \\left\\lfloor \\frac{\\phi_i}{\\pi}\\times \\mathrm{N}\\right\\rfloor" + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "text", + "content": " Each channel in the UV map stores 3DGS attributes, including " + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\{\\sigma_i,r_i,s_i,o_i,c_i\\} \\in \\mathbb{R}^{14}" + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "text", + "content": ". We refer this 14-channel UV map as UVGS, " + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "inline_equation", + "content": "U\\in \\mathbb{R}^{M\\times N\\times 14}" + }, + { + "bbox": [ + 55, + 342, + 296, + 539 + ], + "type": "text", + "content": " defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 88, + 542, + 294, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 542, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 88, + 542, + 294, + 555 + ], + "type": "interline_equation", + "content": "\\mathrm {U} \\left[ \\phi_ {i \\text {s c a l e d}}, \\theta_ {i \\text {s c a l e d}},: \\right] = \\left[ \\sigma_ {i}, r _ {i}, s _ {i}, o _ {i}, c _ {i} \\right]. \\tag {1}", + "image_path": "a4f47b1df662f3c68440cc1040d4720d29b2536bce492f2300b9a5e5a4dcfc84.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "This transformed UVGS representation provides spatial coherence and solves the permutation invariance problem as any random arrangement of points will now map to the same UVGS representation " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": ". It should be noted that this kind of transformation will also preserve the spatial correlation between the Gaussian points in 3D and transform them to 2D UV maps by mapping them to neighboring pixels. This provides both the local level correspondence among the neighboring Gaussians and the overall global correspondence for the object. Thus, solving the unstructured and discreteness problems. This enables standard neural network architectures (e.g. CNNs) to effectively capture correlations among neighboring Gaussians for efficient feature extraction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 335, + 73, + 427, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 427, + 107 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 427, + 107 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {g} _ {1}: \\{\\sigma_ {1}, r _ {1}, s _ {1}, o _ {1}, c _ {1} \\} \\\\ \\mathbf {g} _ {2}: \\{\\sigma_ {2}, r _ {2}, s _ {2}, o _ {2}, c _ {2} \\} \\\\ \\end{array}", + "image_path": "012b92faf4a98bf94d076c37efbd2d570b525bf1085502706b029e57cdc4fcf1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 341, + 107, + 533, + 171 + ], + "blocks": [ + { + "bbox": [ + 341, + 107, + 533, + 171 + ], + "lines": [ + { + "bbox": [ + 341, + 107, + 533, + 171 + ], + "spans": [ + { + "bbox": [ + 341, + 107, + 533, + 171 + ], + "type": "image", + "image_path": "5f38a1646a02c3cfbc429eb39ddf0d13ec479cfd6b803ce2154472531c1b93ff.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 176, + 554, + 232 + ], + "lines": [ + { + "bbox": [ + 313, + 176, + 554, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 176, + 554, + 232 + ], + "type": "text", + "content": "Figure 3. Dynamic Selection. In spherical mapping of 3DGS points to UV maps, multiple points may map to the same pixel, creating a many-to-one issue. Our Dynamic Selection approach addresses this by retaining the attributes of the point with the highest opacity per pixel on the same ray." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 240, + 554, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 554, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 554, + 347 + ], + "type": "text", + "content": "To further unify the extracted position " + }, + { + "bbox": [ + 313, + 240, + 554, + 347 + ], + "type": "inline_equation", + "content": "(\\sigma)" + }, + { + "bbox": [ + 313, + 240, + 554, + 347 + ], + "type": "text", + "content": ", transformation " + }, + { + "bbox": [ + 313, + 240, + 554, + 347 + ], + "type": "inline_equation", + "content": "(r, s)" + }, + { + "bbox": [ + 313, + 240, + 554, + 347 + ], + "type": "text", + "content": ", and color " + }, + { + "bbox": [ + 313, + 240, + 554, + 347 + ], + "type": "inline_equation", + "content": "(c, o)" + }, + { + "bbox": [ + 313, + 240, + 554, + 347 + ], + "type": "text", + "content": " maps in a same feature space and use the existing image foundation models, we map the obtained UVGS " + }, + { + "bbox": [ + 313, + 240, + 554, + 347 + ], + "type": "inline_equation", + "content": "U \\in \\mathbb{R}^{M \\times N \\times 14}" + }, + { + "bbox": [ + 313, + 240, + 554, + 347 + ], + "type": "text", + "content": " further to a 3-channel image " + }, + { + "bbox": [ + 313, + 240, + 554, + 347 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{M \\times N \\times 3}" + }, + { + "bbox": [ + 313, + 240, + 554, + 347 + ], + "type": "text", + "content": " (termed as Super UVGS), using a Convolutional Neural Network (CNN). A multi-branch forward mapping network is employed to map " + }, + { + "bbox": [ + 313, + 240, + 554, + 347 + ], + "type": "inline_equation", + "content": "U \\in \\mathbb{R}^{M \\times N \\times 14}" + }, + { + "bbox": [ + 313, + 240, + 554, + 347 + ], + "type": "text", + "content": " to the 3-channel Super UVGS " + }, + { + "bbox": [ + 313, + 240, + 554, + 347 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{M \\times N \\times 3}" + }, + { + "bbox": [ + 313, + 240, + 554, + 347 + ], + "type": "text", + "content": ". We provide all technical details in Sec. 3.3 and the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 348, + 554, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 554, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 554, + 443 + ], + "type": "text", + "content": "The Super UVGS representation effectively retains all the details of 3DGS attributes and can be directly utilized with existing widely available image-based models. We demonstrate this by showing perfect reconstruction of 3DGS object from Super UVGS image in the experiments section. This semantically structured representation " + }, + { + "bbox": [ + 313, + 348, + 554, + 443 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 348, + 554, + 443 + ], + "type": "text", + "content": " offers both local and global correspondence in representing Gaussian attributes and needs relatively less storage." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 445, + 539, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 539, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 539, + 457 + ], + "type": "text", + "content": "3.2. Dynamic GS Selection and Multiple Layers" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": "When projecting 3DGS points to UV maps using spherical mapping, multiple points may map to the same pixel in UV space as shown in Fig. 3. Two 3DGS points " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "(g_{1})" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "(g_{2})" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": " map to the same pixel on UV map " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "(P_{a})" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": " causing many-to-one mapping issue. To address this, we propose a Dynamic Selection approach where each UV pixel retains the 3DGS attribute with the highest opacity intersecting the same ray. Using the same example in Fig. 3, if opacity " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "o_{1}" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": " of Gaussian " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "g_{1}" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": " is less than opacity " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "o_{2}" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "g_{2}" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": ". Then only " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "g_{2}" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": " will be stored in the UV map at pixel " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "P_{a}" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": ". We observed that this method helps maintain the geometry and appearance of the 3DGS object while resolving many-to-one mapping issues with minimal quality loss. For more complex objects or real-world scene representation, we stack multiple such layers of UV maps, where each UVGS pixel now holds attributes of the top-K opacity values of 3DGS primitives. This can be accomplished by inscribing the 3DGS object inside multiple spheres where each sphere maps the 3DGS attribute corresponding to the top-" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "K^{th}" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": " opacity value along the same ray. More details on this are presented in the supplementary. To show the effectiveness of proposed UVGS" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5930" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 291, + 125 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 291, + 125 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 291, + 125 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 291, + 125 + ], + "type": "image", + "image_path": "25f85b471d8849db0732387367ab55d342304451b294b2457187827ee7586f1c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 127, + 295, + 149 + ], + "lines": [ + { + "bbox": [ + 55, + 127, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 55, + 127, + 295, + 149 + ], + "type": "text", + "content": "Figure 4. Complex object reconstructions " + }, + { + "bbox": [ + 55, + 127, + 295, + 149 + ], + "type": "inline_equation", + "content": "(\\mathrm{K} = 4)" + }, + { + "bbox": [ + 55, + 127, + 295, + 149 + ], + "type": "text", + "content": " using pretrained image-based autoencoder." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 153, + 296, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 153, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 153, + 296, + 190 + ], + "type": "text", + "content": "maps in capturing the intricacies of a complex objects, we used a pretrained image based autoencoder to reconstruct objects using a 4 layer UVGS as shown in Fig. 4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 194, + 168, + 206 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 168, + 206 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 168, + 206 + ], + "type": "text", + "content": "3.3. Mapping Networks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 211, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 296, + 354 + ], + "type": "text", + "content": "Our goal is to bring the extracted UVGS maps to a common feature space to better represent the object collectively and to make the 14-channel UVGS representations " + }, + { + "bbox": [ + 55, + 211, + 296, + 354 + ], + "type": "inline_equation", + "content": "U \\in \\mathbb{R}^{M \\times N \\times 14}" + }, + { + "bbox": [ + 55, + 211, + 296, + 354 + ], + "type": "text", + "content": " work with the widely available image based foundation models. To accomplish this, we map it to a 3-channel image which can be easily processed by the existing architectures while also maintaining the spatial correspondence. We design a simple yet effective multi-branch CNN to extract features from different UVGS attributes and map them to a 3-channel feature-rich image, termed Super UVGS. The structured UVGS maps provide local and global features that can be learned by a CNN." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 355, + 296, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 355, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 296, + 617 + ], + "type": "text", + "content": "Forward Mapping The first layer is a set of three mapping branches for position, transform, and appearance " + }, + { + "bbox": [ + 55, + 355, + 296, + 617 + ], + "type": "inline_equation", + "content": "(\\phi_P^f,\\phi_T^f,\\phi_A^f)" + }, + { + "bbox": [ + 55, + 355, + 296, + 617 + ], + "type": "text", + "content": " respectively. We refer to them as position, transformation, and appearance branch. The position branch takes the mean position " + }, + { + "bbox": [ + 55, + 355, + 296, + 617 + ], + "type": "inline_equation", + "content": "(\\sigma)" + }, + { + "bbox": [ + 55, + 355, + 296, + 617 + ], + "type": "text", + "content": " as an input and processes it to give a position feature map " + }, + { + "bbox": [ + 55, + 355, + 296, + 617 + ], + "type": "inline_equation", + "content": "M_P" + }, + { + "bbox": [ + 55, + 355, + 296, + 617 + ], + "type": "text", + "content": ". Similarly, the transformation branch takes the rotation " + }, + { + "bbox": [ + 55, + 355, + 296, + 617 + ], + "type": "inline_equation", + "content": "(r)" + }, + { + "bbox": [ + 55, + 355, + 296, + 617 + ], + "type": "text", + "content": " and scale " + }, + { + "bbox": [ + 55, + 355, + 296, + 617 + ], + "type": "inline_equation", + "content": "(s)" + }, + { + "bbox": [ + 55, + 355, + 296, + 617 + ], + "type": "text", + "content": " together to generate another feature map " + }, + { + "bbox": [ + 55, + 355, + 296, + 617 + ], + "type": "inline_equation", + "content": "M_T" + }, + { + "bbox": [ + 55, + 355, + 296, + 617 + ], + "type": "text", + "content": ". The last, appearance branch takes the color " + }, + { + "bbox": [ + 55, + 355, + 296, + 617 + ], + "type": "inline_equation", + "content": "(c)" + }, + { + "bbox": [ + 55, + 355, + 296, + 617 + ], + "type": "text", + "content": " and opacity " + }, + { + "bbox": [ + 55, + 355, + 296, + 617 + ], + "type": "inline_equation", + "content": "(o)" + }, + { + "bbox": [ + 55, + 355, + 296, + 617 + ], + "type": "text", + "content": " together to produce another feature map " + }, + { + "bbox": [ + 55, + 355, + 296, + 617 + ], + "type": "inline_equation", + "content": "M_A" + }, + { + "bbox": [ + 55, + 355, + 296, + 617 + ], + "type": "text", + "content": ". All the three features maps from position, transformation, and appearance branch are concatenated to get a final feature map, before passing them to the next module, called the Central Branch. The central branch " + }, + { + "bbox": [ + 55, + 355, + 296, + 617 + ], + "type": "inline_equation", + "content": "(\\phi_C^f)" + }, + { + "bbox": [ + 55, + 355, + 296, + 617 + ], + "type": "text", + "content": " is composed of multiple hidden Convolution layers, where each layer is followed by BatchNorm and ReLu activation. The last layer of the central branch is activated using tanh to ensure the Super UVGS does not take any ambiguous value resulting in gradient explosion or undesired artifacts. The obtained Super UVGS " + }, + { + "bbox": [ + 55, + 355, + 296, + 617 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 355, + 296, + 617 + ], + "type": "text", + "content": " representation squeezes all the 3DGS attributes to a 3 dimensional image while also maintaining local and global structural correspondence among them." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "Inverse Mapping We design an inverse mapping network that aims to map the obtained 3-channel Super UVGS image " + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{M \\times N \\times 3}" + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": " back to the UVGS maps to obtain each of the five different 3DGS attributes " + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\{\\sigma, r, s, o, c\\}" + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": ". The inverse mapping network simply follows the forward mapping network architecture in the reverse order, where at first, we put the Central Branch " + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "inline_equation", + "content": "(\\phi_{iC})" + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": " followed by attribute specific position, transformation, and appearance branches" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "(\\phi_{iP}, \\phi_{iT}, \\phi_{iA})" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": ". We provide more details on mapping network in the supplementary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 98, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 98, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 98, + 555, + 396 + ], + "type": "text", + "content": "Branched mapping layers: The rationale behind using branched mapping layers in both forward and reverse mapping networks is to prevent the incompatibility issues arising due the the different value distribution of 3DGS attributes. Note that the disparate distributions of values within each set of attributes in 3D Gaussian Splatting, (i.e., mean position, transformation, and color), pose a challenge to the model when processed collectively. For instance, neighboring Gaussians in UVGS maps show smooth changes in position and color values but typically have large variations in rotation, scale, and opacity values. This results in gradient anomalies and slow convergence. To address this, we propose a multi-branch network architecture, where attribute-specific branches implicitly learn to process these distinct attribute specific properties, focusing on their unique features before passing them to the central branch. The central branch receives a concatenated stack of processed attributes and exploits the correlation between them by extracting local feature correspondences. This information is then mapped to a 3-channel Super UVGS image, effectively capturing the complex relationships between the various attributes. This approach enables our network to manage diverse attribute distributions, resulting in faster convergence, improved accuracy, and specialized processing for each attribute set." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 396, + 556, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 556, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 556, + 528 + ], + "type": "text", + "content": "Reconstruction Losses: Since the obtained UVGS maps have both local and global features, we opted for image-based losses to train the overall architecture. We use a set of Mean Squared Error (MSE) and Learned Perceptual Image Patch Similarity (LPIPS) [59] between the obtained UVGS from spherical mapping " + }, + { + "bbox": [ + 313, + 396, + 556, + 528 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 396, + 556, + 528 + ], + "type": "text", + "content": " and the predicted UVGS " + }, + { + "bbox": [ + 313, + 396, + 556, + 528 + ], + "type": "inline_equation", + "content": "\\hat{U}" + }, + { + "bbox": [ + 313, + 396, + 556, + 528 + ], + "type": "text", + "content": " from inverse mapping network. We calculate the LPIPS loss over four attributes of 3DGS including mean position " + }, + { + "bbox": [ + 313, + 396, + 556, + 528 + ], + "type": "inline_equation", + "content": "(\\sigma)" + }, + { + "bbox": [ + 313, + 396, + 556, + 528 + ], + "type": "text", + "content": ", view independent color " + }, + { + "bbox": [ + 313, + 396, + 556, + 528 + ], + "type": "inline_equation", + "content": "(c)" + }, + { + "bbox": [ + 313, + 396, + 556, + 528 + ], + "type": "text", + "content": ", scale " + }, + { + "bbox": [ + 313, + 396, + 556, + 528 + ], + "type": "inline_equation", + "content": "(s)" + }, + { + "bbox": [ + 313, + 396, + 556, + 528 + ], + "type": "text", + "content": ", and rotation " + }, + { + "bbox": [ + 313, + 396, + 556, + 528 + ], + "type": "inline_equation", + "content": "(r)" + }, + { + "bbox": [ + 313, + 396, + 556, + 528 + ], + "type": "text", + "content": ". The overall LPIPS loss for UV maps can be written as a linear sum of individual attribute loss terms as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 363, + 531, + 553, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 531, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 363, + 531, + 553, + 544 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {U V - l p i p s} = \\mathcal {L} _ {\\sigma} + \\mathcal {L} _ {s} + \\mathcal {L} _ {r} + \\mathcal {L} _ {c} \\tag {2}", + "image_path": "ecdd9d68cbd13b00ed6bd6a544005ac4b5b83f95b53c6f87962c0e28b4c71bf2.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 546, + 555, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 583 + ], + "type": "text", + "content": "The overall loss function for the training can be written as: " + }, + { + "bbox": [ + 313, + 546, + 555, + 583 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{uvgs} = \\mathcal{L}_{mse} + \\lambda \\mathcal{L}_{UV - lpip}" + }, + { + "bbox": [ + 313, + 546, + 555, + 583 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 546, + 555, + 583 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 546, + 555, + 583 + ], + "type": "text", + "content": " is a scalar and varied from 0 to 10 during the course of training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 590, + 395, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 395, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 395, + 604 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 608, + 555, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 608, + 555, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 608, + 555, + 693 + ], + "type": "text", + "content": "3DGS Dataset and UV maps To train the mapping networks and learn a latent space for unconditional and conditional sampling, we need large amount of 3DGS assets. However, there's a lack of such a large-scale dataset for high quality 3DGS assets. To this end, we create a custom large scale dataset by converting the Objaverse [9] meshes into 3DGS representation1. We start by designing a scene of 88" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 702, + 537, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 537, + 712 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 537, + 712 + ], + "type": "text", + "content": "Sketchfab data was filtered out of training data due to its license" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5931" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 87, + 81, + 134, + 128 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 77, + 135 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 77, + 135 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 77, + 135 + ], + "type": "text", + "content": "Super UVGS" + } + ] + } + ], + "index": 0, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 87, + 81, + 134, + 128 + ], + "lines": [ + { + "bbox": [ + 87, + 81, + 134, + 128 + ], + "spans": [ + { + "bbox": [ + 87, + 81, + 134, + 128 + ], + "type": "image", + "image_path": "1dc5a626ac086a2d02768f0f41c01fbb14266a9fce609b793b8e31f33860ae36.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 149, + 80, + 194, + 127 + ], + "blocks": [ + { + "bbox": [ + 149, + 80, + 194, + 127 + ], + "lines": [ + { + "bbox": [ + 149, + 80, + 194, + 127 + ], + "spans": [ + { + "bbox": [ + 149, + 80, + 194, + 127 + ], + "type": "image", + "image_path": "8beda8f23a9bc3c1caaf1bfd49d52e221b90ce4248c166c1d85c074f9b5a33cc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 205, + 81, + 249, + 126 + ], + "blocks": [ + { + "bbox": [ + 205, + 81, + 249, + 126 + ], + "lines": [ + { + "bbox": [ + 205, + 81, + 249, + 126 + ], + "spans": [ + { + "bbox": [ + 205, + 81, + 249, + 126 + ], + "type": "image", + "image_path": "b0e1ce15fd53c2e7704c1288b2eef0719f0f1be6a21c0ff1db16d797276155a7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 262, + 81, + 307, + 126 + ], + "blocks": [ + { + "bbox": [ + 262, + 81, + 307, + 126 + ], + "lines": [ + { + "bbox": [ + 262, + 81, + 307, + 126 + ], + "spans": [ + { + "bbox": [ + 262, + 81, + 307, + 126 + ], + "type": "image", + "image_path": "6687e906b1b3018f2dfa5af95dd4d2668fa69ec32da4bf312be89ffca22f45ad.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 321, + 81, + 367, + 127 + ], + "blocks": [ + { + "bbox": [ + 321, + 81, + 367, + 127 + ], + "lines": [ + { + "bbox": [ + 321, + 81, + 367, + 127 + ], + "spans": [ + { + "bbox": [ + 321, + 81, + 367, + 127 + ], + "type": "image", + "image_path": "78177fa3c088aad27f8d98524cabc7c526ac1a360d7605e948527a561786e5a8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 378, + 81, + 423, + 126 + ], + "blocks": [ + { + "bbox": [ + 378, + 81, + 423, + 126 + ], + "lines": [ + { + "bbox": [ + 378, + 81, + 423, + 126 + ], + "spans": [ + { + "bbox": [ + 378, + 81, + 423, + 126 + ], + "type": "image", + "image_path": "04d8c3195569f5eaf9a5bf9380fc8b41626cf0a6543a7559151b9ddb142dbc38.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 437, + 81, + 481, + 126 + ], + "blocks": [ + { + "bbox": [ + 437, + 81, + 481, + 126 + ], + "lines": [ + { + "bbox": [ + 437, + 81, + 481, + 126 + ], + "spans": [ + { + "bbox": [ + 437, + 81, + 481, + 126 + ], + "type": "image", + "image_path": "5841673b1bd05609cbc173e2a859e8f1c93a0b5f13b07bf4384fe334df246596.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 494, + 81, + 539, + 127 + ], + "blocks": [ + { + "bbox": [ + 494, + 81, + 539, + 127 + ], + "lines": [ + { + "bbox": [ + 494, + 81, + 539, + 127 + ], + "spans": [ + { + "bbox": [ + 494, + 81, + 539, + 127 + ], + "type": "image", + "image_path": "65a1810f7586622267b5ec4b77f012e35bd7d97eae625847b5079e7ee338c6b8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 90, + 130, + 132, + 181 + ], + "blocks": [ + { + "bbox": [ + 62, + 152, + 77, + 203 + ], + "lines": [ + { + "bbox": [ + 62, + 152, + 77, + 203 + ], + "spans": [ + { + "bbox": [ + 62, + 152, + 77, + 203 + ], + "type": "text", + "content": "Multiview" + } + ] + } + ], + "index": 9, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 90, + 130, + 132, + 181 + ], + "lines": [ + { + "bbox": [ + 90, + 130, + 132, + 181 + ], + "spans": [ + { + "bbox": [ + 90, + 130, + 132, + 181 + ], + "type": "image", + "image_path": "ab61707530cf3160e9cd8c366d8970684b15a78a2237bcc6bb9e8c03cf0e7182.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 143, + 133, + 190, + 184 + ], + "blocks": [ + { + "bbox": [ + 143, + 133, + 190, + 184 + ], + "lines": [ + { + "bbox": [ + 143, + 133, + 190, + 184 + ], + "spans": [ + { + "bbox": [ + 143, + 133, + 190, + 184 + ], + "type": "image", + "image_path": "1f4b60a82e998ae541fa551d8a0d44507e52b2229b5ae96487925ef701f3cb02.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 197, + 134, + 253, + 171 + ], + "blocks": [ + { + "bbox": [ + 197, + 134, + 253, + 171 + ], + "lines": [ + { + "bbox": [ + 197, + 134, + 253, + 171 + ], + "spans": [ + { + "bbox": [ + 197, + 134, + 253, + 171 + ], + "type": "image", + "image_path": "aa3f24f3951e68a4bb79b96328cd391d2f1797d7587555dae6eb6b71d856820b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 260, + 134, + 306, + 182 + ], + "blocks": [ + { + "bbox": [ + 260, + 134, + 306, + 182 + ], + "lines": [ + { + "bbox": [ + 260, + 134, + 306, + 182 + ], + "spans": [ + { + "bbox": [ + 260, + 134, + 306, + 182 + ], + "type": "image", + "image_path": "2181461956050c1a50973f4c3c7a8bc54201dc66b3c66e8a442eb0da9ccbbfaa.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 307, + 134, + 371, + 185 + ], + "blocks": [ + { + "bbox": [ + 307, + 134, + 371, + 185 + ], + "lines": [ + { + "bbox": [ + 307, + 134, + 371, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 134, + 371, + 185 + ], + "type": "image", + "image_path": "8f4e3156691618f61e31cbb4ef8567fe1a0b27783896476258b0963e08533cc4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 372, + 141, + 431, + 181 + ], + "blocks": [ + { + "bbox": [ + 372, + 141, + 431, + 181 + ], + "lines": [ + { + "bbox": [ + 372, + 141, + 431, + 181 + ], + "spans": [ + { + "bbox": [ + 372, + 141, + 431, + 181 + ], + "type": "image", + "image_path": "42aacd4f4ded688c31817634696bdc8790fba9fa44403be745688be3057e109d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 432, + 141, + 492, + 194 + ], + "blocks": [ + { + "bbox": [ + 432, + 141, + 492, + 194 + ], + "lines": [ + { + "bbox": [ + 432, + 141, + 492, + 194 + ], + "spans": [ + { + "bbox": [ + 432, + 141, + 492, + 194 + ], + "type": "image", + "image_path": "09c5b76f22ca499839949af1195aecc4162f0f908164769d3ab94b8b557afabc.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 492, + 145, + 539, + 196 + ], + "blocks": [ + { + "bbox": [ + 492, + 145, + 539, + 196 + ], + "lines": [ + { + "bbox": [ + 492, + 145, + 539, + 196 + ], + "spans": [ + { + "bbox": [ + 492, + 145, + 539, + 196 + ], + "type": "image", + "image_path": "bf976d48495428f2ad3081772460cb5594a7770ccaf78138b6de932531834cd9.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 88, + 190, + 134, + 233 + ], + "blocks": [ + { + "bbox": [ + 88, + 190, + 134, + 233 + ], + "lines": [ + { + "bbox": [ + 88, + 190, + 134, + 233 + ], + "spans": [ + { + "bbox": [ + 88, + 190, + 134, + 233 + ], + "type": "image", + "image_path": "eb16a8e08262dd945922a110d24995e7017aaf8a4fa7a0b3c0c8362dd3ca364f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 243, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 55, + 243, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 555, + 277 + ], + "type": "text", + "content": "Figure 5. Figure shows a wide variety of high-quality unconditional generation result from our method using Latent Diffusion Model. We train an LDM to randomly sample Super UVGS images from random noise. The Super UVGS can be converted to 3DGS object using inverse mapping network and inverse spherical projection. The unconditional generation model was trained on Objaverse dataset." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 143, + 189, + 196, + 238 + ], + "blocks": [ + { + "bbox": [ + 143, + 189, + 196, + 238 + ], + "lines": [ + { + "bbox": [ + 143, + 189, + 196, + 238 + ], + "spans": [ + { + "bbox": [ + 143, + 189, + 196, + 238 + ], + "type": "image", + "image_path": "1151107cb1adec15d4e96ec0eb05926ce551844693514ade22960aacac1ea5bf.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 199, + 176, + 253, + 238 + ], + "blocks": [ + { + "bbox": [ + 199, + 176, + 253, + 238 + ], + "lines": [ + { + "bbox": [ + 199, + 176, + 253, + 238 + ], + "spans": [ + { + "bbox": [ + 199, + 176, + 253, + 238 + ], + "type": "image", + "image_path": "e0578a159631be7f346e7a0101212201fc0a9a8aeb956047446db1297f51178b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 257, + 191, + 312, + 238 + ], + "blocks": [ + { + "bbox": [ + 257, + 191, + 312, + 238 + ], + "lines": [ + { + "bbox": [ + 257, + 191, + 312, + 238 + ], + "spans": [ + { + "bbox": [ + 257, + 191, + 312, + 238 + ], + "type": "image", + "image_path": "8e08ccbeda13d769a5a7ec266201c75e0025ce62751f7faae2ae650511ce8a8b.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 315, + 186, + 372, + 236 + ], + "blocks": [ + { + "bbox": [ + 315, + 186, + 372, + 236 + ], + "lines": [ + { + "bbox": [ + 315, + 186, + 372, + 236 + ], + "spans": [ + { + "bbox": [ + 315, + 186, + 372, + 236 + ], + "type": "image", + "image_path": "a85e5dc621c7b3ae6c8421f1307599da18a1c0ce3ce54339efb51cc5c47db559.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 378, + 187, + 428, + 237 + ], + "blocks": [ + { + "bbox": [ + 378, + 187, + 428, + 237 + ], + "lines": [ + { + "bbox": [ + 378, + 187, + 428, + 237 + ], + "spans": [ + { + "bbox": [ + 378, + 187, + 428, + 237 + ], + "type": "image", + "image_path": "5c8310cfeb5762da162f009b7afca5bb3e703d79674a612be2e209717a2fbcbb.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 436, + 195, + 480, + 239 + ], + "blocks": [ + { + "bbox": [ + 436, + 195, + 480, + 239 + ], + "lines": [ + { + "bbox": [ + 436, + 195, + 480, + 239 + ], + "spans": [ + { + "bbox": [ + 436, + 195, + 480, + 239 + ], + "type": "image", + "image_path": "2b0e4f7e92927991c7af1e47af366ca08933e7d807c5c4d18424d48e3cdb5961.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 489, + 205, + 539, + 239 + ], + "blocks": [ + { + "bbox": [ + 489, + 205, + 539, + 239 + ], + "lines": [ + { + "bbox": [ + 489, + 205, + 539, + 239 + ], + "spans": [ + { + "bbox": [ + 489, + 205, + 539, + 239 + ], + "type": "image", + "image_path": "b9ec333bcceb0e99ac1946078b82d4e27cbc23874bb8235162c33d0ea7ab1f7d.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 56, + 286, + 296, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 296, + 526 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 296, + 526 + ], + "type": "text", + "content": "cameras in a canonical space and use it to capture Objaverse objects from various angles covering all of the object views. The 88 rendered views from different angles are then used to train a 3DGS for " + }, + { + "bbox": [ + 56, + 286, + 296, + 526 + ], + "type": "inline_equation", + "content": "10\\mathrm{K}" + }, + { + "bbox": [ + 56, + 286, + 296, + 526 + ], + "type": "text", + "content": " iterations using [17]. This way, we create a high-quality and large-scale 3DGS dataset of " + }, + { + "bbox": [ + 56, + 286, + 296, + 526 + ], + "type": "inline_equation", + "content": "\\sim 400K" + }, + { + "bbox": [ + 56, + 286, + 296, + 526 + ], + "type": "text", + "content": " objects and scenes from Objaverse. We only use static scenes or objects from Objaverse. After fitting all the object to 3DGS representation, we convert the objects to the corresponding UV maps (i.e., UVGS) through Spherical Mapping as illustrated in Fig. 2. For the course of our experiments, we only map the objects to a single layer UV maps as it was sufficient to represent the general purpose Objaverse objects with minimal quality loss. Through mapping, we gathered a UVGS dataset of " + }, + { + "bbox": [ + 56, + 286, + 296, + 526 + ], + "type": "inline_equation", + "content": "\\sim 400K" + }, + { + "bbox": [ + 56, + 286, + 296, + 526 + ], + "type": "text", + "content": " maps. We fix the size of the UVGS maps to " + }, + { + "bbox": [ + 56, + 286, + 296, + 526 + ], + "type": "inline_equation", + "content": "512\\times 512" + }, + { + "bbox": [ + 56, + 286, + 296, + 526 + ], + "type": "text", + "content": ". Through our experiments, we found that UV maps of size " + }, + { + "bbox": [ + 56, + 286, + 296, + 526 + ], + "type": "inline_equation", + "content": "512\\times 512" + }, + { + "bbox": [ + 56, + 286, + 296, + 526 + ], + "type": "text", + "content": " are sufficient to represent objects in our dataset and capable of storing upto " + }, + { + "bbox": [ + 56, + 286, + 296, + 526 + ], + "type": "inline_equation", + "content": "262K" + }, + { + "bbox": [ + 56, + 286, + 296, + 526 + ], + "type": "text", + "content": " unique Gaussians. Table 1 compares 1-4 layer UV maps. We also did experiments on ShapeNet [4] cars dataset for evaluation purposes." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "text", + "content": "Baselines & Metrics: To evaluate the quality of reconstructed 3DGS objects from both Super UVGS image and Autoencoder latent space to 3D, we use PSNR and LPIPS [58]. The aim is to convert the given 3DGS object to UVGS, and then to Super UVGS, and further to Autoencoder's latent space and calculate the metrics from the reconstructions at every step to prove the proposed method doesn't significantly affect the quality of reconstructions, while also providing a structurally meaningful representation that is much compact and easier to use with existing image based models. We compare the generational capabilities of our method against various conditional and unconditional SOTA 3D object generation method including the ones using multiview rendering for optimization DiffTF [2], Get3D [12], methods trying to give structural representa" + } + ] + } + ], + "index": 28 + }, + { + "type": "table", + "bbox": [ + 337, + 349, + 529, + 460 + ], + "blocks": [ + { + "bbox": [ + 313, + 285, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 340 + ], + "type": "text", + "content": "Table 1. PSNR and LPIPS comparison for various reconstruction methods using UVGS and Super UVGS representations on Objverse Cars and Full datasets. AE, VAE, VQVAE are pretrained image based models. " + }, + { + "bbox": [ + 313, + 285, + 555, + 340 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 285, + 555, + 340 + ], + "type": "text", + "content": " is the number of UVGS layers used. We also report the compression " + }, + { + "bbox": [ + 313, + 285, + 555, + 340 + ], + "type": "inline_equation", + "content": "\\%" + }, + { + "bbox": [ + 313, + 285, + 555, + 340 + ], + "type": "text", + "content": " (CP) compared to the fitted 3DGS." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 337, + 349, + 529, + 460 + ], + "lines": [ + { + "bbox": [ + 337, + 349, + 529, + 460 + ], + "spans": [ + { + "bbox": [ + 337, + 349, + 529, + 460 + ], + "type": "table", + "html": "
MethodPSNR(C/F)LPIPS(C/F)CP(%)
3DGS34.6 / 34.20.02 / 0.020
UVGS (@K=1)31.3 / 31.10.06 / 0.0653.0
UVGS (@K=2)32.8 / 31.90.04 / 0.0545.6
UVGS (@K=4)34.2 / 33.20.02 / 0.0333.3
Super UVGS (@K=1)31.2 / 31.10.07 / 0.0889.7
AE (@K=1)30.9 / 30.80.07 / 0.0999.5
VAE (@K=1)30.6 / 30.90.07 / 0.0999.5
VQVAE (@K=1)30.3 / 30.10.08 / 0.1099.7
", + "image_path": "4b7900656acf03689a915809fae75b179603d970107ca3592b479e8d6eb21817.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "table_body" + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 468, + 555, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 555, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 555, + 516 + ], + "type": "text", + "content": "tion to Gaussians, GaussianCube [55], and general purpose SOTA large 3D content generation models like DreamGaussian [42], LGM [44], and EG3D [3]. We also compare the quality of our generation results using FID and KID." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 517, + 555, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 555, + 589 + ], + "type": "text", + "content": "Mapping Network Training Details We train the forward and inverse mapping networks to project the obtained UVGS maps " + }, + { + "bbox": [ + 313, + 517, + 555, + 589 + ], + "type": "inline_equation", + "content": "U \\in \\mathbb{R}^{M \\times N \\times 14}" + }, + { + "bbox": [ + 313, + 517, + 555, + 589 + ], + "type": "text", + "content": " to Super UVGS image " + }, + { + "bbox": [ + 313, + 517, + 555, + 589 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{M \\times N \\times 3}" + }, + { + "bbox": [ + 313, + 517, + 555, + 589 + ], + "type": "text", + "content": ", and back to the reconstructed UVGS maps " + }, + { + "bbox": [ + 313, + 517, + 555, + 589 + ], + "type": "inline_equation", + "content": "\\hat{U} \\in \\mathbb{R}^{M \\times N \\times 14}" + }, + { + "bbox": [ + 313, + 517, + 555, + 589 + ], + "type": "text", + "content": ". We provide an in-depth discussion of all implementation details in the supplementary material." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 598, + 547, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 547, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 547, + 612 + ], + "type": "text", + "content": "4.1. UVGS AutoEncoder and 3DGS Compression" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "The obtained Super UVGS image is a structurally meaningful representation that can have various applications in the generation and reconstruction of new 3D assets as it contains features that can be learned by the existing image based models. Through our experiments, we show that a 3-channel Super UVGS image can be directly reconstructed using a pretrained image based Autoencoders or VAEs without any fine-tuning. We tested on three different models in" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5932" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 76, + 100, + 121, + 130 + ], + "blocks": [ + { + "bbox": [ + 78, + 72, + 113, + 84 + ], + "lines": [ + { + "bbox": [ + 78, + 72, + 113, + 84 + ], + "spans": [ + { + "bbox": [ + 78, + 72, + 113, + 84 + ], + "type": "text", + "content": "DiffTF" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 76, + 100, + 121, + 130 + ], + "lines": [ + { + "bbox": [ + 76, + 100, + 121, + 130 + ], + "spans": [ + { + "bbox": [ + 76, + 100, + 121, + 130 + ], + "type": "image", + "image_path": "83c9eb6c0339a44fcc598dba59764b39bdde7c43b8a06f7adb4694b8aa648149.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 75, + 140, + 121, + 175 + ], + "blocks": [ + { + "bbox": [ + 75, + 140, + 121, + 175 + ], + "lines": [ + { + "bbox": [ + 75, + 140, + 121, + 175 + ], + "spans": [ + { + "bbox": [ + 75, + 140, + 121, + 175 + ], + "type": "image", + "image_path": "f8b27dba3749b293c7fcd612abc93e6a5692fdd722615c5523b2220b3661bf2b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 182, + 555, + 225 + ], + "lines": [ + { + "bbox": [ + 55, + 182, + 555, + 225 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 555, + 225 + ], + "type": "text", + "content": "Figure 6. Comparison of unconditional 3D asset generation on the cars category with SOTA methods. Figure shows that DiffTF [2] produces low-quality, low-resolution cars lacking detail. While Get3D [12] achieve higher resolution, it suffers from 3D inconsistency, numerous artifacts, and lacks 3D detail. Similar issues are found in GaussianCube [55] along with symmetric inconsistency in the results. In contrast, our method generates high-quality, high-resolution objects that are 3D consistent with sharp and well-defined edges." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 155, + 91, + 189, + 131 + ], + "blocks": [ + { + "bbox": [ + 192, + 74, + 220, + 83 + ], + "lines": [ + { + "bbox": [ + 192, + 74, + 220, + 83 + ], + "spans": [ + { + "bbox": [ + 192, + 74, + 220, + 83 + ], + "type": "text", + "content": "Get3D" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 155, + 91, + 189, + 131 + ], + "lines": [ + { + "bbox": [ + 155, + 91, + 189, + 131 + ], + "spans": [ + { + "bbox": [ + 155, + 91, + 189, + 131 + ], + "type": "image", + "image_path": "5870965b0a011f348bd6406b300281c44a52ef0e9604240a6a8abb734b80b0af.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 153, + 140, + 188, + 169 + ], + "blocks": [ + { + "bbox": [ + 153, + 140, + 188, + 169 + ], + "lines": [ + { + "bbox": [ + 153, + 140, + 188, + 169 + ], + "spans": [ + { + "bbox": [ + 153, + 140, + 188, + 169 + ], + "type": "image", + "image_path": "10cae7b5547d1e94a4698c66306f6605fb00ce2bdd7d2bf89132787d6e247fce.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 192, + 90, + 251, + 135 + ], + "blocks": [ + { + "bbox": [ + 192, + 90, + 251, + 135 + ], + "lines": [ + { + "bbox": [ + 192, + 90, + 251, + 135 + ], + "spans": [ + { + "bbox": [ + 192, + 90, + 251, + 135 + ], + "type": "image", + "image_path": "eb419d82e0e5c06e732353251b34b4192c4150959452a4d4f001551c659b0e60.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 74, + 361, + 83 + ], + "lines": [ + { + "bbox": [ + 299, + 74, + 361, + 83 + ], + "spans": [ + { + "bbox": [ + 299, + 74, + 361, + 83 + ], + "type": "text", + "content": "GaussianCube" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 192, + 136, + 251, + 174 + ], + "blocks": [ + { + "bbox": [ + 192, + 136, + 251, + 174 + ], + "lines": [ + { + "bbox": [ + 192, + 136, + 251, + 174 + ], + "spans": [ + { + "bbox": [ + 192, + 136, + 251, + 174 + ], + "type": "image", + "image_path": "98bd3dec6d91c19e4fa3ea1f3ead282225d12a61228adc786a1b1976574945a7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 273, + 102, + 328, + 136 + ], + "blocks": [ + { + "bbox": [ + 273, + 102, + 328, + 136 + ], + "lines": [ + { + "bbox": [ + 273, + 102, + 328, + 136 + ], + "spans": [ + { + "bbox": [ + 273, + 102, + 328, + 136 + ], + "type": "image", + "image_path": "8504566ec2d121a5832dae4248bf1c9e5c94d3ef4a5b3bfc24d64b1912390500.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 331, + 100, + 365, + 137 + ], + "blocks": [ + { + "bbox": [ + 331, + 100, + 365, + 137 + ], + "lines": [ + { + "bbox": [ + 331, + 100, + 365, + 137 + ], + "spans": [ + { + "bbox": [ + 331, + 100, + 365, + 137 + ], + "type": "image", + "image_path": "4924ea712094fd45596345baf08169c651010493e4047c2eb7c7d6a6db3ccd75.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 272, + 147, + 329, + 177 + ], + "blocks": [ + { + "bbox": [ + 272, + 147, + 329, + 177 + ], + "lines": [ + { + "bbox": [ + 272, + 147, + 329, + 177 + ], + "spans": [ + { + "bbox": [ + 272, + 147, + 329, + 177 + ], + "type": "image", + "image_path": "239ff5dad6d501015f121e0ed1457f0d82ddc7d3cbc76c5cf93b672da90b7dce.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 366, + 101, + 399, + 137 + ], + "blocks": [ + { + "bbox": [ + 366, + 101, + 399, + 137 + ], + "lines": [ + { + "bbox": [ + 366, + 101, + 399, + 137 + ], + "spans": [ + { + "bbox": [ + 366, + 101, + 399, + 137 + ], + "type": "image", + "image_path": "62cbbd6a34134c889925976552340213dabdc4fa60cf4a94096a2a5e34ed3a0a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 349, + 146, + 399, + 177 + ], + "blocks": [ + { + "bbox": [ + 349, + 146, + 399, + 177 + ], + "lines": [ + { + "bbox": [ + 349, + 146, + 399, + 177 + ], + "spans": [ + { + "bbox": [ + 349, + 146, + 399, + 177 + ], + "type": "image", + "image_path": "8425730ae76e0693e63bf58c18b9bf1988ac9693a8bc038f68565d3062b40028.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 419, + 99, + 488, + 130 + ], + "blocks": [ + { + "bbox": [ + 447, + 74, + 504, + 85 + ], + "lines": [ + { + "bbox": [ + 447, + 74, + 504, + 85 + ], + "spans": [ + { + "bbox": [ + 447, + 74, + 504, + 85 + ], + "type": "text", + "content": "UVGS (ours)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 419, + 99, + 488, + 130 + ], + "lines": [ + { + "bbox": [ + 419, + 99, + 488, + 130 + ], + "spans": [ + { + "bbox": [ + 419, + 99, + 488, + 130 + ], + "type": "image", + "image_path": "372dcede4f068eee9c86a879d1329201caa180f78c6b746fa0fc5f7ee45f60f2.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 491, + 93, + 536, + 135 + ], + "blocks": [ + { + "bbox": [ + 491, + 93, + 536, + 135 + ], + "lines": [ + { + "bbox": [ + 491, + 93, + 536, + 135 + ], + "spans": [ + { + "bbox": [ + 491, + 93, + 536, + 135 + ], + "type": "image", + "image_path": "7088103b939532b32cad1d42762c80488345d78ea510f63ecb785ebfdf45d5ab.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 418, + 143, + 488, + 171 + ], + "blocks": [ + { + "bbox": [ + 418, + 143, + 488, + 171 + ], + "lines": [ + { + "bbox": [ + 418, + 143, + 488, + 171 + ], + "spans": [ + { + "bbox": [ + 418, + 143, + 488, + 171 + ], + "type": "image", + "image_path": "59fdcae91097b03e80b388e4c84b9c9f72c98e6d66420a19d65c733754a7a622.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 492, + 141, + 536, + 168 + ], + "blocks": [ + { + "bbox": [ + 492, + 141, + 536, + 168 + ], + "lines": [ + { + "bbox": [ + 492, + 141, + 536, + 168 + ], + "spans": [ + { + "bbox": [ + 492, + 141, + 536, + 168 + ], + "type": "image", + "image_path": "0a70511b2ed037342fd7e43ca5bd700ca854e7feaa68a04ff3b8e71f47e76d70.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 56, + 234, + 296, + 383 + ], + "blocks": [ + { + "bbox": [ + 56, + 234, + 296, + 383 + ], + "lines": [ + { + "bbox": [ + 56, + 234, + 296, + 383 + ], + "spans": [ + { + "bbox": [ + 56, + 234, + 296, + 383 + ], + "type": "image", + "image_path": "9ccc93f755797d2ed09bbc5b39ddd31e850e21d441d4d40aeaf8236c971807d8.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 392, + 295, + 436 + ], + "lines": [ + { + "bbox": [ + 55, + 392, + 295, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 295, + 436 + ], + "type": "text", + "content": "Figure 7. We compare the performance of our model against various SOTA methods for text-conditional object synthesis. Our method not only generates high-quality assets for simpler objects, but also for complicated objects with intricate geometry." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 440, + 295, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 440, + 295, + 560 + ], + "spans": [ + { + "bbox": [ + 55, + 440, + 295, + 560 + ], + "type": "text", + "content": "cluding image AE, KL-VAE [18], VQVAE [46] and each performed quite well without any significant quality loss. The reconstruction PSNR and LPIPS values are presented in Table 1. This means we can now leverage the powerful compression capabilities of image based Autoencoders to compress the storage requirements of 3DGS by more than " + }, + { + "bbox": [ + 55, + 440, + 295, + 560 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 55, + 440, + 295, + 560 + ], + "type": "text", + "content": ". We have shown the storage comparison results in Table 1. It is interesting to note that the Super UVGS representation itself can be used to compress the memory requirement for storing 3DGS object by up to " + }, + { + "bbox": [ + 55, + 440, + 295, + 560 + ], + "type": "inline_equation", + "content": "89.7\\%" + }, + { + "bbox": [ + 55, + 440, + 295, + 560 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 564, + 270, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 270, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 270, + 576 + ], + "type": "text", + "content": "4.2. Unconditional & Conditional Generation" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "We aim to show the effectiveness of Super UVGS representation for directly generating 3DGS objects from a learned latent space. We consider Super UVGS images as a compact and structured proxy for representing 3DGS objects as it maintains the 3D object information while also providing learnable features. The existing methods fail to directly generate a large number of Gaussians (e.g. " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "100\\mathrm{K}+" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": ") to represent objects with sufficient quality either due to the lack of 3D generative model architectures that support such large number of unstructured points, or due to the lack of a large 3DGS dataset [28, 40, 44, 55, 63]. We leverage the" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 236, + 555, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 367 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 367 + ], + "type": "text", + "content": "Super UVGS representation and use the existing 2D image generative models like Diffusion Models [15] for this task. Specifically, to train a generative model capable of randomly sampling new high-quality 3DGS assets for various downstream tasks, we use an unconditional Latent Diffusion Model (LDM) [36] on the obtained Super UVGS images. As illustrated in Section 4.1, we can use a pretrained image VAE to map the Super UVGS image to a latent space and reconstruct back. Hence, we only train a LDM on the latent space. More implementation details are provided in the supplementary." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 371, + 556, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 556, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 556, + 526 + ], + "type": "text", + "content": "Unconditional LDM: To design a generative model capable of randomly sampling new high-quality 3DGS assets for various downstream tasks, we train an unconditional LDM [36] on the learned Super UVGS images. Following [33, 36, 57], we use DDIM [15] for faster and consistent sampling with up to 1000 time steps used in the forward diffusion process, and 20 during denoising. Once trained, the model is used to randomly sample Super UVGS images, resulting in high quality 3DGS assets through inverse mapping. Results are presented in Fig 5. We demonstrate that our method inherently learns to generate multiview consistent images due to the powerful Super UVGS representation unlike most prior works using rendering-based losses." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 530, + 556, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 556, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 556, + 601 + ], + "type": "text", + "content": "Conditional LDM: Similar to unconditional generation, we also trained a text-conditioned LDM following the SD's [33, 36, 57] pipeline and using the predicted text for our dataset. The trained model can be used to generate high-quality text-conditioned 3DGS assets that are multiview consistent. The results are demonstrated in Fig 6." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": "The above experiments prove the effectiveness of our proposed Super UVGS representation in 3D object synthesis using widely available 2D image models. It also highlights that this compact 3-channel Super UVGS representation stores not just the spatial correspondence among different pixels, but also the rich 3D information of the objects. This way, we can easily convert a 3D asset generation problem into a 2D image generation problem without the use of any complex 3D architecture to handle large amount of un" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5933" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 68, + 290, + 227 + ], + "blocks": [ + { + "bbox": [ + 61, + 68, + 290, + 227 + ], + "lines": [ + { + "bbox": [ + 61, + 68, + 290, + 227 + ], + "spans": [ + { + "bbox": [ + 61, + 68, + 290, + 227 + ], + "type": "image", + "image_path": "e5569757356944ee79466c8fb221ce34c48934ad1c6558ee3d05fe8959d08516.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 228, + 295, + 262 + ], + "lines": [ + { + "bbox": [ + 55, + 228, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 295, + 262 + ], + "type": "text", + "content": "Figure 8. 3DGS Inpainting: We present one of the first inpainting results on 3DGS directly leveraging the Super UVGS images and the denoising capabilities of diffusion models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 270, + 296, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 318 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 318 + ], + "type": "text", + "content": "structured and permutation invariant 3DGS primitives, and neither relying upon computationally expensive multiview rendering or SDS loss. Baseline comparison for unconditional and conditional generation is presented in Table 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 324, + 160, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 160, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 160, + 338 + ], + "type": "text", + "content": "4.3. 3DGS Inpainting" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "type": "text", + "content": "Leveraging the powerful Super UVGS representation, we present in Fig. 8 one of the first experiments on inpainting 3DGS directly without using any multiview rendering or distilling information from diffusion models. We try to recover the missing Gaussians by leveraging the denoising capabilities of LDM and trying to predict the missing corresponding parts of the Super UVGS image. We believe, this can have potential applications in sparse view reconstruction. More details are given in the supplementary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 456, + 220, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 220, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 220, + 468 + ], + "type": "text", + "content": "4.4. Ablation Studies & Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": "We conduct exhaustive ablation studies to justify some of our framework's design choices including the effect of branching in mapping networks, the use of single layer UVGS maps, and the resolution of UVGS maps. We performed our experiments on our custom Objaverse [9] 3DGS dataset and evaluate the performance of our model in terms of PSNR, SSIM, and LPIPS. The results are presented in Table 3. From the table, it can be seen that by using four layers of UVGS maps " + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "inline_equation", + "content": "(\\mathrm{K} = 4)" + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "text", + "content": ", we can almost match the reconstruction quality of fitted 3DGS results, while we realized that simply with a single layer UVGS, we are able to maintain the overall geometry and appearance of the object for our dataset with a PSNR of more than 30. We also compared the reconstruction performance of our method with and without using branching in the mapping network. It can be clearly seen that using branching network significantly increases the reconstruction quality from Super UVGS space. The main reasoning behind this is the specialization of attributes that the branching provides to individually process each attribute first." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 332, + 123, + 535, + 209 + ], + "blocks": [ + { + "bbox": [ + 313, + 65, + 555, + 120 + ], + "lines": [ + { + "bbox": [ + 313, + 65, + 555, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 65, + 555, + 120 + ], + "type": "text", + "content": "Table 2. We compare the FID and KID of unconditional generation using the current SOTA methods on 20K randomly generated samples from each method and ours. We also compare our method against SOTA text-conditioned generation frameworks on CLIP Score for 10K generated objects from each method." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 123, + 535, + 209 + ], + "lines": [ + { + "bbox": [ + 332, + 123, + 535, + 209 + ], + "spans": [ + { + "bbox": [ + 332, + 123, + 535, + 209 + ], + "type": "table", + "html": "
Unconditional GenerationText-Conditioned Generation
MethodFID ↓KID ↓MethodCLIP Score ↑
Get3D [12]53.174.19DreamGaussian [42]28.51
DiffTF [2]84.578.73Shap. E [6]30.53
EG3D [3]74.516.62LGM [44]30.74
GaussianCube34.673.72GaussianCube [55]30.34
UVGS (Ours)26.203.24UVGS (Ours)32.62
", + "image_path": "29c2d3ce00e202e64898fc8220ef70edb87833dbc11c6e4a07603ab3ce8fa9d2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 332, + 251, + 535, + 300 + ], + "blocks": [ + { + "bbox": [ + 313, + 213, + 555, + 247 + ], + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 247 + ], + "type": "text", + "content": "Table 3. We present quantitative ablation study for number of UVGS layers (K), UVGS map resolution, and the effect of branching in mapping network on the Objaverse 3DGS dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 251, + 535, + 300 + ], + "lines": [ + { + "bbox": [ + 332, + 251, + 535, + 300 + ], + "spans": [ + { + "bbox": [ + 332, + 251, + 535, + 300 + ], + "type": "table", + "html": "
MethodPSNRLPIPSUVGS SizePSNRLPIPS
UVGS @K=131.10.06512 × 512 (@K=1)31.10.08
UVGS @K=231.90.05256 × 256 (@K=1)28.20.23
UVGS @K=433.20.03Without Branching27.80.31
", + "image_path": "f4d28aafa8c82127eb685605ee87735199243e646870ee52ccb18cb580325036.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 308, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 475 + ], + "type": "text", + "content": "Limitations & Future Work: While single layer UVGS images can recover the geometry of the object, they sometimes suffer in terms of appearance and the generated objects might look washed out. We believe this can be solved by using a multi-layer UVGS maps. Similarly, the single layer UVGS map is limited to representing simpler everyday objects, and may not be sufficient to represent highly-detailed and complex objects or scenes. In the future, we want to extend this framework to learn features for real-world scenes and complex objects like a human head with multi-layer UV mapping. We also want to make this representation more efficient by better utilizing the empty pixels of UVGS maps and Super UVGS images while also maintaining the underlying features and 3D information." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 480, + 388, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 388, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 388, + 492 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 495, + 555, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 674 + ], + "type": "text", + "content": "We introduced a novel method to solve the underlying issues with 3D Gaussian Splatting (3DGS) that prevent the direct integration of them with the large number of existing image foundational models. We proposed UVGS - a structured representation for 3DGS obtained by spherical mapping of 3DGS primitives to UV maps. We further squeezed the multi-attribute UVGS maps to a 3-channel unified and structured Super UVGS image, which not only maintains the 3D structural information of the object, but also provides a compact feature space for 3DGS attributes. The obtained Super UVGS images are directly integrated with the existing image foundational models for 3DGS compression and unconditional and conditional generation using diffusion models. Leveraging these Super UVGS images, we showed one of the first inpainting experiments on 3DGS." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "ACKNOWLEDGEMENTS A part of this work was supported by NSF CAREER grant 2143576 and ONR DURIP grant N00014-23-1-2804." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5934" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "text", + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855-5864, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 159, + 296, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 296, + 192 + ], + "type": "text", + "content": "[2] Ziang Cao, Fangzhou Hong, Tong Wu, Liang Pan, and Ziwei Liu. Large-vocabulary 3d diffusion model with transformer. arXiv preprint arXiv:2309.07920, 2023. 3, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 296, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 296, + 259 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 296, + 259 + ], + "type": "text", + "content": "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16123-16133, 2022. 2, 6, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 262, + 296, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 262, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 262, + 296, + 338 + ], + "type": "text", + "content": "[4] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 340, + 296, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 340, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 340, + 296, + 384 + ], + "type": "text", + "content": "[5] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision, pages 333-350. Springer, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 386, + 296, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 386, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 296, + 441 + ], + "type": "text", + "content": "[6] Minghao Chen, Junyu Xie, Iro Laina, and Andrea Vedaldi. Shap-editor: Instruction-guided latent 3d editing in seconds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26456-26466, 2024. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 442, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 442, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 442, + 296, + 487 + ], + "type": "text", + "content": "[7] Zilong Chen, Feng Wang, Yikai Wang, and Huaping Liu. Text-to-3d using gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21401-21412, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 488, + 296, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 296, + 543 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 296, + 543 + ], + "type": "text", + "content": "[8] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4456-4465, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 544, + 296, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 544, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 296, + 610 + ], + "type": "text", + "content": "[9] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli Vanderbilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 3, 5, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 613, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 656 + ], + "type": "text", + "content": "[10] Slava Elizarov, Ciara Rowles, and Simon Donne. Geometry image diffusion: Fast and data-efficient text-to-3d with image-based surface representation. arXiv preprint arXiv:2409.03718, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 658, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 714 + ], + "type": "text", + "content": "[11] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 554, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 138 + ], + "type": "text", + "content": "[12] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841-31854, 2022. 3, 6, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 140, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 172 + ], + "type": "text", + "content": "[13] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 173, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 228 + ], + "type": "text", + "content": "[14] Xianglong He, Junyi Chen, Sida Peng, Di Huang, Yangguang Li, Xiaoshui Huang, Chun Yuan, Wanli Ouyang, and Tong He. Gvgen: Text-to-3d generation with volumetric representation. In European Conference on Computer Vision, pages 463–479. Springer, 2025. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 228, + 553, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 553, + 260 + ], + "type": "text", + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 262, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 262, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 317, + 262, + 553, + 315 + ], + "type": "text", + "content": "[16] Han Huang, Yulun Wu, Junsheng Zhou, Ge Gao, Ming Gu, and Yu-Shen Liu. Neusurf: On-surface priors for neural surface reconstruction from sparse input views. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2312-2320, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 316, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 316, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 317, + 316, + 553, + 360 + ], + "type": "text", + "content": "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 361, + 553, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 382 + ], + "type": "text", + "content": "[18] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 383, + 553, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 553, + 447 + ], + "type": "text", + "content": "[19] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 449, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 553, + 492 + ], + "type": "text", + "content": "[20] Shujuan Li, Junsheng Zhou, Baorui Ma, Yu-Shen Liu, and Zhizhong Han. Neaf: Learning neural angle fields for point normal estimation. In Proceedings of the AAAI conference on artificial intelligence, pages 1396-1404, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 494, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 494, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 317, + 494, + 553, + 558 + ], + "type": "text", + "content": "[21] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 559, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 603 + ], + "type": "text", + "content": "[22] Baorui Ma, Haoge Deng, Junsheng Zhou, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Geodream: Disentangling 2d and geometric priors for high-fidelity and consistent 3d generation. arXiv preprint arXiv:2311.17971, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 604, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 604, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 604, + 553, + 657 + ], + "type": "text", + "content": "[23] Qi Ma, Yue Li, Bin Ren, Nicu Sebe, Ender Konukoglu, Theo Gevers, Luc Van Gool, and Danda Pani Paudel. Shapes-plat: A large-scale dataset of gaussian splats and their self-supervised pretraining. arXiv preprint arXiv:2408.10906, 2024. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 658, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 714 + ], + "type": "text", + "content": "[24] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12663–12673, 2023. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5935" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "text", + "content": "[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 127, + 294, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 127, + 294, + 159 + ], + "spans": [ + { + "bbox": [ + 56, + 127, + 294, + 159 + ], + "type": "text", + "content": "[26] Chaerin Min and Srinath Sridhar. Genheld: Generating and editing handheld objects. arXiv preprint arXiv:2406.05059, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 294, + 205 + ], + "type": "text", + "content": "[27] Chaerin Min, Sehyun Cha, Changhee Won, and Jongwoo Lim. Tsdf-sampling: Efficient sampling for neural surface field using truncated signed distance field. arXiv preprint arXiv:2311.17878, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 258 + ], + "type": "text", + "content": "[28] Yuxuan Mu, Xinxin Zuo, Chuan Guo, Yilin Wang, Juwei Lu, Xiaofeng Wu, Songcen Xu, Peng Dai, Youliang Yan, and Li Cheng. Gsd: View-guided gaussian splatting diffusion for 3d reconstruction. arXiv preprint arXiv:2407.04237, 2024. 2, 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 326 + ], + "type": "text", + "content": "[29] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulo, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4328-4338, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "type": "text", + "content": "[30] Francesco Palandra, Andrea Sanchietti, Daniele Baieri, and Emanuele Rodola. Gsedit: Efficient text-guided editing of 3d objects via gaussian splatting. arXiv preprint arXiv:2403.05154, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 403 + ], + "type": "text", + "content": "[31] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "type": "text", + "content": "[32] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 491 + ], + "type": "text", + "content": "[33] Aashish Rai and Srinath Sridhar. Egosonics: Generating synchronized audio for silent egocentric videos. arXiv preprint arXiv:2407.20592, 2024. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 493, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 558 + ], + "type": "text", + "content": "[34] Aashish Rai, Hiresh Gupta, Ayush Pandey, Francisco Vicente Carrasco, Shingo Jason Takagi, Amaury Aubel, Daeil Kim, Aayush Prakash, and Fernando De la Torre. Towards realistic generative 3d face models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3738-3748, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 559, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 624 + ], + "type": "text", + "content": "[35] Amit Raj, Srinivas Kaza, Ben Poole, Michael Niemeyer, Nataniel Ruiz, Ben Mildenhall, Shiran Zada, Kfir Aberman, Michael Rubinstein, Jonathan Barron, et al. Dreambooth3d: Subject-driven text-to-3d generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2349–2359, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 679 + ], + "type": "text", + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "text", + "content": "[37] Li Shen and Fillia Makedon. Spherical mapping for processing of 3d closed surfaces. Image and vision computing, 24 (7):743-761, 2006. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "text", + "content": "[38] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20875-20886, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "text", + "content": "[39] Jingxiang Sun, Bo Zhang, Ruizhi Shao, Lizhen Wang, Wen Liu, Zhenda Xie, and Yebin Liu. Dreamcraft3d: Hierarchical 3d generation with bootstrapped diffusion prior. arXiv preprint arXiv:2310.16818, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "text", + "content": "[40] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10208-10217, 2024. 2, 3, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 230, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 306 + ], + "type": "text", + "content": "[41] Fariborz Taherkhani, Aashish Rai, Quankai Gao, Shaunak Srivastava, Xuanbai Chen, Fernando De la Torre, Steven Song, Aayush Prakash, and Daeil Kim. Controllable 3d generative adversarial face model via disentangling shape and appearance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 826-836, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "type": "text", + "content": "[42] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 2, 3, 6, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 354, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 409 + ], + "type": "text", + "content": "[43] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22819–22829, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "text", + "content": "[44] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1-18. Springer, 2025. 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 466, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 510 + ], + "type": "text", + "content": "[45] Zhicong Tang, Shuyang Gu, Chunyu Wang, Ting Zhang, Jianmin Bao, Dong Chen, and Baining Guo. Volumediffusion: Flexible text-to-3d generation with efficient volumetric encoder. arXiv preprint arXiv:2312.11459, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 512, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 544 + ], + "type": "text", + "content": "[46] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 545, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 611 + ], + "type": "text", + "content": "[47] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4563-4573, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 613, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 668 + ], + "type": "text", + "content": "[48] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3295–3306, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[49] Yuxuan Wang, Xuanyu Yi, Zike Wu, Na Zhao, Long Chen, and Hanwang Zhang. View-consistent 3d editing with gaussian splatting. In European Conference on Computer Vision, pages 404-420. Springer, 2025. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5936" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[50] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 193 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 193 + ], + "type": "text", + "content": "[51] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 194, + 294, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 248 + ], + "type": "text", + "content": "[52] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation. arXiv preprint arXiv:2403.14621, 2024. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "text", + "content": "[53] Xingguang Yan, Han-Hung Lee, Ziyu Wan, and Angel X Chang. An object is worth 64x64 pixels: Generating 3d object via image diffusion. arXiv preprint arXiv:2408.03178, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 294, + 294, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 347 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 347 + ], + "type": "text", + "content": "[54] Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian-dreamer: Fast generation from text to 3d gaussian splattering with point cloud priors. arXiv preprint arXiv:2310.08529, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 349, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 349, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 349, + 294, + 403 + ], + "type": "text", + "content": "[55] Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: Structuring gaussian splatting using optimal transport for 3d generative modeling. arXiv preprint arXiv:2403.19655, 2024. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "type": "text", + "content": "[56] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "text", + "content": "[57] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "type": "text", + "content": "[58] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 559, + 294, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 592 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 592 + ], + "type": "text", + "content": "[59] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 593, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 593, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 593, + 294, + 624 + ], + "type": "text", + "content": "[60] Junsheng Zhou, Yu-Shen Liu, and Zhizhong Han. Zero-shot scene reconstruction from single images with deep prior assembly. arXiv preprint arXiv:2410.15971, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 657 + ], + "type": "text", + "content": "[61] Junsheng Zhou, Weiqi Zhang, and Yu-Shen Liu. Diff-gfs: Functional gaussian splatting diffusion. arXiv preprint arXiv:2410.19657, 2024. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[62] Junsheng Zhou, Weiqi Zhang, Baorui Ma, Kanle Shi, YuShen Liu, and Zhizhong Han. Udiff: Generating conditional unsigned distance fields with optimal wavelet diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21496-21506, 2024. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 206 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 139 + ], + "type": "text", + "content": "[63] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10324-10335, 2024. 2, 3, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 555, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 555, + 206 + ], + "type": "text", + "content": "[64] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10324-10335, 2024. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5937" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/UWAV_ Uncertainty-weighted Weakly-supervised Audio-Visual Video Parsing/82f7769f-b767-4619-96ea-539823053afb_content_list.json b/2025/UWAV_ Uncertainty-weighted Weakly-supervised Audio-Visual Video Parsing/82f7769f-b767-4619-96ea-539823053afb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2a1247a1d86c2362f0bdafc31c250ec32db03443 --- /dev/null +++ b/2025/UWAV_ Uncertainty-weighted Weakly-supervised Audio-Visual Video Parsing/82f7769f-b767-4619-96ea-539823053afb_content_list.json @@ -0,0 +1,1451 @@ +[ + { + "type": "text", + "text": "UWAV: Uncertainty-weighted Weakly-supervised Audio-Visual Video Parsing", + "text_level": 1, + "bbox": [ + 94, + 130, + 875, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yung-Hsuan Lai $^{1,\\dagger}$ , Janek Ebbers $^{3}$ , Yu-Chiang Frank Wang $^{1,2}$ , François Germain $^{3}$ , Michael Jeffrey Jones $^{3}$ , Moitreya Chatterjee $^{3,\\ddagger}$ , \n $^{1}$ Graduate Institute of Communication Engineering, National Taiwan University $^{2}$ NVIDIA, Taiwan \n $^{3}$ Mitsubishi Electric Research Labs (MERL) \n $\\dagger$ r10942097@ntu.edu.tw $\\ddagger$ metro.smiles@gmail.com", + "bbox": [ + 147, + 179, + 823, + 271 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 313, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Audio-Visual Video Parsing (AVVP) entails the challenging task of localizing both uni-modal events (i.e., those occurring exclusively in either the visual or acoustic modality of a video) and multi-modal events (i.e., those occurring in both modalities concurrently). Moreover, the prohibitive cost of annotating training data with the class labels of all these events, along with their start and end times, imposes constraints on the scalability of AVVP techniques unless they can be trained in a weakly-supervised setting, where only modality-agnostic, video-level labels are available in the training data. To this end, recently proposed approaches seek to generate segment-level pseudolabels to better guide model training. However, the absence of inter-segment dependencies when generating these pseudo-labels and the general bias towards predicting labels that are absent in a segment limit their performance. This work proposes a novel approach towards overcoming these weaknesses called Uncertainty-weighted Weakly-supervised Audio-visual Video Parsing (UWAV). Additionally, our innovative approach factors in the uncertainty associated with these estimated pseudo-labels and incorporates a feature mixup based training regularization for improved training. Empirical results show that UWAV outperforms state-of-the-art methods for the AVVP task on multiple metrics, across two different datasets, attesting to its effectiveness and generalizability.", + "bbox": [ + 75, + 335, + 473, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 746, + 209, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Events that occur in the real world, often leave their imprint on the acoustic and visual modalities. Humans rely heavily on the synergy between their senses of sight and hearing to interpret such events. Audio-visual learning, which seeks to equip machines with a similar synergy, has emerged as one of the most important research areas within the multi", + "bbox": [ + 75, + 772, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/843db44b81fc66a66494d92c620dd3afa3214df81fbb5d2257344d9cd27e586c.jpg", + "image_caption": [ + "Figure 1. A weakly-supervised AVVP task example. Events, considered in this task, might be unimodal or multimodal. Even multimodal events, may not be temporally aligned in the audio and visual modalities, e.g. the cello might only be visible in the first few seconds but might produce music, throughout the video." + ], + "image_footnote": [], + "bbox": [ + 504, + 306, + 890, + 501 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "modal machine learning community. It aims to leverage both these senses (modalities) jointly, to enhance machine perception and understanding of real-world events. Various audio-visual learning tasks have been studied towards this end, including audio-visual segmentation [19, 49], sound source localization [24, 32], audio-visual event localization [29, 34], and audio-visual sound separation [3, 7, 43]. However, many of these tasks assume that audio and visual streams would always be temporally aligned. This assumption often fails in real-world scenarios, where the sonic and visual imprints of events may not perfectly overlap. For instance, one might hear an emergency siren approaching from a distance before it appears in the field of view.", + "bbox": [ + 496, + 595, + 892, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, in order to better understand the events occurring in a video, we explore the task of Audio-Visual Video Parsing (AVVP) [35]. Its goal is to recognize and localize all audio, visual, and audio-visual events occurring in the video. See Figure 1 for an example of this task. The task setup is to perform this prediction for every one-second segment of a video. This task poses two principal challenges,", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Code is available at: https://github.com/Franklin905/UWAV.", + "bbox": [ + 76, + 875, + 467, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13561", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "from a machine learning standpoint: (i) The audio and visual events, that occur, might not be temporally aligned, e.g. if an event becomes audible before its source enters the camera field of view, or the sound source is not visible at all, and (ii) due to the high costs of annotating video segments with per-segment labels, only modality-agnostic, video-level labels are provided during training, i.e., these labels specify which events occur in a video but lack details about the segments or the modality in which they occur.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prior works in the area can be grouped into two orthogonal research directions. The first focuses on enhancing model architectures [23, 45, 51]. Despite advancements in this direction, the absence of fine-grained labels to guide the model during training continues to pose an impediment towards the generalizability of such models. As a result, recent approaches have focused on the second direction of research which aims at generating richer pseudo-labels for improved training, either at the video-level [8, 39] or segment-level [9, 17, 26, 50]. In particular, Rachavarapu et al. [26] propose prototype-based pseudo-labeling (PPL), which seeks to train a pseudo-label generation module in conjunction with a core inference module. However, due to the lack of sufficient training data, this method struggles to generalize. On the other hand, VPLAN [50], VALOR [17], and LSLD [9] leverage large-scale pre-trained foundation models, such as CLIP [28] and CLAP [40], along with ground-truth video-level labels to generate segment-level pseudo-labels for each of the two modalities. Audio/Visual segments (e.g. the audio corresponding to the segment in question and the visual frame at the center of the segment) are fed into CLAP/CLIP, one segment at a time, to generate these pseudo-labels. Despite the significant improvement that these pseudo-label generation methods achieved, the correctness of the generated labels is still limited, constrained primarily by the lack of understanding of intersegment dynamics. For instance, if a crowd is cheering in a segment of the video, it is more likely that the crowd might also be clapping right before or after.", + "bbox": [ + 75, + 231, + 472, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the oversight of inter-segment dependencies and other shortcomings in existing pseudo-label generation methods, we introduce a novel, uncertainty-based, weakly-supervised, video parsing model called UWAV (Uncertainty-weighted Weakly-supervised Audiovisual Video Parsing), capable of generating improved segment-level pseudo-labels for better training of the inference module. We resort to transformer modules [38] to equip UWAV with the ability to capture temporal relationships between segments and pre-train it on a large-scale, supervised audio-visual event localization dataset [12]. Subsequently, this pre-trained model is used to generate segment-level pseudo-labels for each modality on a target, small-scale dataset which only provides weak (i.e., video-level) supervision. Such a design permits a more holis", + "bbox": [ + 75, + 674, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tic understanding of the video, resulting in more accurate pseudo-labels. Additionally, UWAV factors in the uncertainty associated with these estimated pseudo-labels in its optimization. That uncertainty is the result of the shift in the domain of the target dataset, insufficient model capacity, etc. and is computed at training time as the confidence scores associated with these labels. To further enhance the model's ability to learn in this small-scale, weakly-supervised data regime, we also employ a feature mixup strategy. This approach adds more regularization constraints by training on mixed segment features alongside interpolated pseudo-labels, which not only lessens the influence of noise but also enriches the training data, thereby reducing overfitting. Moreover, UWAV addresses a critical class imbalance issue in the pseudo-label enriched training data, viz. most event classes in any given segment of a video are absent/negative (i.e., they do not occur), while only a handful of them do. This creates a natural bias in the training set, making it difficult to learn the positive events. To counter this, we propose a class-frequency aware weighting strategy that lays greater emphasis on the accurate classification and localization of positive events. By incorporating these components into its design, our proposed method (UWAV) outperforms competing state-of-the-art approaches across two different datasets, viz. Look, Listen, and Parse (LLP) [35] and the Audio-Visual Event Localization (AVE) [34], on multiple evaluation metrics. In summary, our contributions are the following:", + "bbox": [ + 496, + 90, + 893, + 512 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel, weakly-supervised method called UWAV, capable of synthesizing temporally coherent pseudo-labels for the AVVP task.", + "- To the best of our knowledge, ours is the first method for the AVVP task, which factors in the uncertainty associated with the estimated pseudo-labels while also regularizing it with a feature mixup strategy.", + "- UWAV outperforms competing state-of-the-art approaches for the AVVP task, across two different datasets on multiple metrics which attest to its generalizability." + ], + "bbox": [ + 500, + 513, + 890, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 678, + 650, + 695 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Audio-Visual Learning: Audio-visual learning has emerged as an area of active research, aiming to develop models that synergistically integrate information from both audio and visual modalities for improved perception and understanding. Towards this end, various audio-visual tasks have been explored by the community so far, such as audio-visual segmentation [19, 22, 42, 49], sound source localization [15, 16, 24, 32], event localization [29, 34, 48], navigation [4-6, 21, 44, 46], generation [25, 30, 41], question answering [11, 18, 33, 47], and sound source separation [2, 7, 36, 43]. In this work, we focus on the task of audio-visual video parsing (AVVP) where the goal is to temporally localize events occurring in a video.", + "bbox": [ + 496, + 704, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13562", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike many other audio-visual learning tasks, AVVP does not assume that events are always aligned across modalities. Some events could be exclusively uni-modal while others may have an audio-visual signature, which requires complex reasoning.", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Audio-Visual Video Parsing (AVVP): To address the challenges of the AVVP task [17, 26, 35], Tian et al. [35] proposed a Hybrid Attention Network (HAN) and a learnable Multi-modal Multiple Instance Learning (MMIL) pooling module. The HAN model facilitates the exchange of information within and across modalities using self-attention and cross-attention layers, while the MMIL pooling module aggregates segment-level event probabilities from both modalities to produce video-level probabilities. Building on this foundation, recent works advanced the field from the following two perspectives. The first group of studies [23, 45, 51] focuses on enhancing model architectures. In particular, Mo and Tian [23] proposed the Multimodal Grouping Network (MGN) to explicitly group semantically similar features within each modality to improve the reasoning process, while Yu et al. [45] proposed the Multi-modal Pyramid Attentional Network (MM-Pyramid) to capture events of varying durations by extracting features at multiple temporal scales. Our proposed method is orthogonal to this line of research and can be integrated with any of these backbones.", + "bbox": [ + 75, + 183, + 472, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The second direction focuses on generating pseudo-labels for improved training, either at the video-level [8, 39] or the segment-level [9, 17, 26, 50]. VPLAN [50], VALOR [17], and LSLD [9] utilize pre-trained CLIP [28] and CLAP [40] along with ground-truth video-level labels to predict pseudo-labels for each modality on a per-segment basis. In contrast, PPL [26] uses the HAN model itself to generate pseudo-labels by constructing prototype features for each class and assigning labels to each segment based on the similarity between its feature and the prototype features. While these pseudo-label generation methods have substantially improved model performance on the AVVP task, they still exhibit some limitations. For instance, to derive accurate pseudo-labels, PPL might require a large enough training set to learn good prototype features, which might pose challenges when applied to smaller datasets. Our proposed method overcomes this problem. On the other hand, methods that leverage CLIP and CLAP to generate pseudo-labels often ignore temporal relationships between segments or the uncertainty associated with these labels. Our work also seeks to plug this void.", + "bbox": [ + 75, + 501, + 472, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 829, + 217, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Formulation: The AVVP task [35] aims to localize all visible and/or audible events in each one-second segment of a video. Specifically, an audible video is split", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "into $T$ one-second segments, denoted as $\\{V_t, A_t\\}_{t=1}^T$ . Each segment is annotated with a pair of ground-truth labels $y_t^v \\in \\{0,1\\}^C, y_t^a \\in \\{0,1\\}^C$ , where $y_t^v$ denotes visual events, $y_t^a$ denotes audio events, and $C$ denotes the total number of events in the pre-defined event set of the data. However, owing to the weakly-supervised nature of the task setup $(y_t^v, y_t^a)$ are unavailable during training. Instead, only the modality-agnostic, video-level labels $y \\in \\{0,1\\}^C$ are available, where 1 indicates the presence of an event at any time (either in the audio or visual stream or both) while 0 indicates an event's absence in the video.", + "bbox": [ + 496, + 90, + 893, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pseudo-Label Based AVVP Framework: The Hybrid Attention Network (HAN) [35] is a commonly used model for the AVVP task. The model works by first utilizing pre-trained visual and audio backbones to extract features from the visual and audio segments respectively, which are then projected to two $d$ -dimensional feature spaces. The resulting visual segment-level features are denoted by $F^v = \\{f_t^v\\}_{t=1}^T \\in \\mathbb{R}^{T \\times d}$ , while the audio segment-level features are denoted by $F^a = \\{f_t^a\\}_{t=1}^T \\in \\mathbb{R}^{T \\times d}$ . These features are provided as input to the HAN model. In the model, information across segments within a modality and across modalities is exchanged through self-attention and cross-attention layers, as shown below:", + "bbox": [ + 496, + 273, + 893, + 472 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {f} _ {t} ^ {v} = f _ {t} ^ {v} + \\underbrace {\\operatorname {A t t n} \\left(f _ {t} ^ {v} , F ^ {v} , F ^ {v}\\right)} _ {\\text {S e l f - A t t e n t i o n}} + \\underbrace {\\operatorname {A t t n} \\left(f _ {t} ^ {v} , F ^ {a} , F ^ {a}\\right)} _ {\\text {C r o s s - A t t e n t i o n}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 479, + 890, + 518 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {f} _ {t} ^ {a} = f _ {t} ^ {a} + \\underbrace {\\operatorname {A t t n} \\left(f _ {t} ^ {a} , F ^ {a} , F ^ {a}\\right)} _ {\\text {S e l f - A t t e n t i o n}} + \\underbrace {\\operatorname {A t t n} \\left(f _ {t} ^ {a} , F ^ {v} , F ^ {v}\\right)} _ {\\text {C r o s s - A t t e n t i o n}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 523, + 890, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{Attn}(Q,K,V)$ denotes the standard multi-head attention mechanism [38]. Finally a classifier, shared across both modalities, transforms the visual segment-level features $\\tilde{F}^v = \\{\\tilde{f}_t^v\\}_{t=1}^T \\in \\mathbb{R}^{T \\times d}$ (resp. audio segment-level features $\\tilde{F}^a = \\{f_t^a\\}_{t=1}^T$ into visual segment-level logits $\\{z_t^v\\}_{t=1}^T \\in \\mathbb{R}^{T \\times C}$ (resp. audio segment-level logits $\\{z_t^a\\}_{t=1}^T$ ). Segment-level probabilities $\\{p_t^v\\}_{t=1}^T, \\{p_t^a\\}_{t=1}^T \\in \\mathbb{R}^{T \\times C}$ are then obtained by applying the sigmoid function on $\\{z_t^v\\}_{t=1}^T$ and $\\{z_t^a\\}_{t=1}^T$ .", + "bbox": [ + 496, + 571, + 890, + 707 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since, only video-level labels $y$ are available during training, Tian et al. [35] introduce an attentive MMIL pooling module to learn to predict video-level probabilities $p \\in \\mathbb{R}^{C}$ :", + "bbox": [ + 496, + 707, + 890, + 765 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nW _ {\\text {m o d a l}} ^ {v, a} = \\operatorname {S o f t m a x} _ {\\text {m o d a l}} \\left(\\operatorname {F C} _ {\\text {m o d a l}} \\left(\\tilde {F} ^ {v, a}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 771, + 890, + 789 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nW _ {\\text {t i m e}} ^ {v, a} = \\operatorname {S o f t m a x} _ {\\text {t i m e}} \\left(\\mathrm {F C} _ {\\text {t i m e}} \\left(\\tilde {F} ^ {v, a}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 801, + 890, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{FC}_{\\mathrm{mod}}$ and $\\mathrm{FC}_{\\mathrm{time}}$ are two learnable fully-connected layers, $\\tilde{F}^{v,a} = \\mathrm{Stack}(\\tilde{F}^v,\\tilde{F}^a)\\in \\mathbb{R}^{2\\times T\\times d}$ denotes the stacked visual and audio features along the first dimension, $\\mathrm{Softmax}_{\\mathrm{mod}}(\\cdot)$ denotes the softmax operation along the modality dimension (i.e., across $v,a$ ), while", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13563", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/029b65d7fd68f8213edf4cde2b94874146ec42b34ecad0188cda9c3d73940c33.jpg", + "image_caption": [ + "Stage 1: Training the pseudo label models on larger dataset" + ], + "image_footnote": [], + "bbox": [ + 78, + 121, + 357, + 273 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a1622defeee030970a2302157682bd239910b8e4aa1869a6824d634183d9653d.jpg", + "image_caption": [ + "Figure 2. UWAV framework: In stage 1, pseudo-label generation modules are equipped with the ability to capture temporal relationships between segments by pre-training on a large-scale, supervised, audio-visual event localization dataset. In stage 2, temporally coherent, uncertainty-weighted pseudo-labels, derived from the pre-trained pseudo-label generation module, are used to guide the learning of the inference model (HAN) aided by a class-balanced loss re-weighting and uncertainty-weighted feature mixup strategy. Note that we use the feature mixup strategy in both modalities while we only show the breakdown of the mixup operation for the audio modality." + ], + "image_footnote": [], + "bbox": [ + 375, + 97, + 727, + 273 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ab9adea9eb46a43f9ff1794ff0eeb3dbf4d2001c9b90e141286f7addd551c025.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 95, + 885, + 271 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\operatorname{Softmax}_{time}(\\cdot)$ denotes the softmax operation along the temporal dimension (i.e., across $1, \\ldots, T$ ). Video-level probabilities $p \\in \\mathbb{R}^C$ are then obtained via:", + "bbox": [ + 75, + 369, + 468, + 415 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np = \\sum_ {m = \\{v, a \\}} \\sum_ {t = 1} ^ {T} W _ {\\text {m o d a l}} ^ {m, t} \\odot W _ {\\text {t i m e}} ^ {m, t} \\odot p _ {t} ^ {m}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 417, + 468, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\odot$ denotes the element-wise product. The HAN model is then optimized with the binary cross-entropy (BCE) loss between the estimated video-level probabilities $p$ and video-level labels $y$ : $\\mathcal{L}_{video} = \\mathrm{BCE}(p, y)$ .", + "bbox": [ + 75, + 468, + 468, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Proposed Approach", + "text_level": 1, + "bbox": [ + 76, + 541, + 267, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we detail our proposed approach (UWAV). At a high level, UWAV works by generating better segment-level pseudo-labels to improve the training of a multi-modal transformer-based inference module, e.g. HAN. Moreover, UWAV factors in the uncertainty associated with these pseudo-labels, addresses the imbalance in the training data, and introduces self-supervised regularization constraints, which all lead to better performance. Figure 2 shows an overview of our proposed framework.", + "bbox": [ + 75, + 566, + 468, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Temporally-Coherent Pseudo-Label Synthesis", + "text_level": 1, + "bbox": [ + 75, + 710, + 465, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One major issue that plagues prior works, based on pseudolabel generation [9, 17, 50], is that they do not capture the temporal dependencies between neighboring segments when generating the pseudo-labels. That is, the generated pseudo-labels are not temporally coherent. To plug this void, we propose to incorporate transformer modules [38] into the pseudo-label generation pipeline, which maps CLIP/CLAP encodings of a segment's visual frame/audio to pseudo-labels. Specifically, two separate transformers are introduced, one each for the visual/audio pseudo-label synthesis modules.", + "bbox": [ + 75, + 734, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pre-Training: Training transformers often requires sufficiently large training data while datasets commonly used for the weakly-supervised AVVP task are relatively small. To mitigate this challenge, we propose to first pre-train the transformer-equipped pseudo-label generation module on a large-scale, supervised, audio-visual event localization dataset - the UnAV [12] dataset. Specifically, given an audible video of duration $T'$ seconds from the pre-training dataset, we split the video into $T'$ one-second segments $\\{V_t', A_t'\\}_{t=1}^{T'}$ , with corresponding audio-visual event labels $y_t^{av'} \\in \\{0,1\\}^{C'}$ , where 1 indicates the presence of an event in both modalities and 0 its absence in at least one modality, while $C'$ denotes the total number of event classes in the pre-training dataset. Next, the video frame at the temporal center of the visual segment is transformed into visual features $G_0^{v'} = \\{g_{0,t}'\\}_{t=1}^{T'} \\in \\mathbb{R}^{T' \\times d_1}$ with CLIP's [28] image encoder. These features are then fed into the corresponding transformer of the visual stream, consisting of $L$ encoder blocks, each block containing a self-attention layer, LayerNorm [1] (LN), and a 2-layer feed-forward network (FFN):", + "bbox": [ + 496, + 369, + 892, + 672 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {G} _ {l} ^ {v ^ {\\prime}} = \\mathrm {L N} \\left(G _ {l} ^ {v ^ {\\prime}} + \\operatorname {A t t n} \\left(G _ {l} ^ {v ^ {\\prime}}, G _ {l} ^ {v ^ {\\prime}}, G _ {l} ^ {v ^ {\\prime}}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 681, + 890, + 700 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nG _ {l + 1} ^ {v ^ {\\prime}} = \\operatorname {L N} \\left(\\tilde {G} _ {l} ^ {v ^ {\\prime}} + \\operatorname {F F N} \\left(\\tilde {G} _ {l} ^ {v ^ {\\prime}}\\right)\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 703, + 890, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concurrently, we convert each event category label in the pre-training dataset into a textual event feature $e_c^{CLIP'} \\in \\mathbb{R}^{d_1}$ by filling in the pre-defined caption template: \"A photo of \" with the corresponding event name and passing it to CLIP's text encoder. Equipped with the visual segment-level features $G_L^{v'} = \\{g_{L,t}'\\}_{t=1}^{T'} \\in \\mathbb{R}^{T' \\times d_1}$ and the textual event features $E^{CLIP'} = \\{e_c^{CLIP'}\\}_{c=1}^{C'} \\in \\mathbb{R}^{C' \\times d_1}$ , we derive visual segment-level logits $\\hat{z}_t^{v'} \\in \\mathbb{R}^{C'}$ and probabilities $\\hat{p}_t^{v'}$ as follows:", + "bbox": [ + 496, + 734, + 892, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {p} _ {t} ^ {v ^ {\\prime}} = \\operatorname {S i g m o i d} \\left(\\hat {z} _ {t} ^ {v ^ {\\prime}}\\right), \\hat {z} _ {t} ^ {v ^ {\\prime}} = E ^ {C L I P ^ {\\prime}} \\cdot g _ {L, t} ^ {v ^ {\\prime} \\top}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 882, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13564", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similar operations are performed in the audio pseudolabel generation pipeline. The raw waveforms corresponding to the 1-second audio segments are transformed into audio features $G_0^{a'} \\in \\mathbb{R}^{T' \\times d_2}$ with CLAP's [12] audio encoder and fed into the corresponding transformer consisting of $L$ encoder blocks. Correspondingly, the textual event features $E^{CLAP'} \\in \\mathbb{R}^{C' \\times d_2}$ are generated with the caption template: \"This is the sound of \" by passing it through CLAP's text encoder. Audio segment-level logits $\\hat{z}_t^{a'} \\in \\mathbb{R}^{C'}$ and probabilities $\\hat{p}_t^{a'}$ can then be derived in the same manner: $\\hat{p}_t^{a'} = \\text{Sigmoid}(\\hat{z}_t^{a'})$ , $\\hat{z}_t^{a'} = E^{CLAP'} \\cdot g_t^{a'}^\\top$ .", + "bbox": [ + 76, + 90, + 468, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the events occurring in the pre-training dataset (UnAV) are audio-visual, we multiply the segment-level visual and audio event probabilities to enforce the predicted labels to be multi-modal in nature: $\\{\\hat{p}_t^{av'}\\}_{t=1}^{T'} \\in \\mathbb{R}^{T' \\times C'}$ . This network is then trained with the binary cross-entropy (BCE) loss:", + "bbox": [ + 76, + 258, + 468, + 348 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t e m p}} = \\mathrm {B C E} \\left(\\hat {p} _ {t} ^ {a v ^ {\\prime}}, y _ {t} ^ {a v ^ {\\prime}}\\right), \\hat {p} _ {t} ^ {a v ^ {\\prime}} = \\tilde {p} _ {t} ^ {v ^ {\\prime}} \\odot \\hat {p} _ {t} ^ {a ^ {\\prime}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 356, + 468, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pseudo-Label Generation on Target Dataset: With the pre-trained pseudo-label generation modules in place, we proceed to employ them for the pseudo-label generation process in the target dataset for the AVVP task. Specifically, the center frame of each of the visual segments $\\{V_t\\}_{t=1}^T$ of the target dataset are passed into CLIP's image encoder, whose output is passed into the pre-trained visual transformer to generate segment features $G_L^v = \\{g_{L,t}^v\\}_{t=1}^T \\in \\mathbb{R}^{T \\times d_1}$ . At the same time, the caption template: \"A photo of \" is used to obtain textual features corresponding to each of the event classes in the target dataset for the AVVP task: $E^{CLIP} \\in \\mathbb{R}^{C \\times d_1}$ . Segment-level visual logits $\\hat{z}_t^v \\in \\mathbb{R}^C$ can be derived by computing their inner product. We also pre-define class-wise visual thresholds $\\theta^v \\in \\mathbb{R}^C$ to transform segment-level visual logits into binary pseudo-labels $\\hat{y}_t^v \\in \\mathbb{R}^C$ :", + "bbox": [ + 76, + 382, + 468, + 626 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} _ {t} ^ {v} = \\mathbb {1} _ {\\left\\{\\hat {z} _ {t} ^ {v} > \\theta^ {v} \\right\\}} \\odot y, \\hat {z} _ {t} ^ {v} = E ^ {C L I P} \\cdot g _ {L, t} ^ {v} ^ {\\top}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 632, + 468, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $y$ denotes the ground-truth video-level labels, $\\mathbb{1}_{\\{\\cdot\\}}$ is the indicator function which returns a value of 1 when the condition is true otherwise 0, and $\\odot$ denotes the elementwise product operation. The $\\odot$ operation zeroes out the predictions of event classes absent in the video-level label.", + "bbox": [ + 76, + 659, + 468, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A similar pseudo-label generation process is employed on the acoustic side. Raw waveforms of audio segments are first fed into CLAP's audio encoder and then into the pretrained audio transformer. The event names of the classes in the target dataset for the AVVP task are filled in the caption template: \"This is the sound of \" to generate textual event features: $E^{CLAP} \\in \\mathbb{R}^{C \\times d_1}$ . Segment-level audio logits $\\hat{z}_t^a \\in \\mathbb{R}^C$ and binary pseudo-labels $\\hat{y}_t^a \\in \\mathbb{R}^C$ are then derived using class-wise thresholds $\\theta^a \\in \\mathbb{R}^C$ .", + "bbox": [ + 76, + 734, + 468, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With binary segment-level pseudo-labels for both modalities $\\hat{y}_t^v, \\hat{y}_t^a$ and the predicted probabilities from the infer", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ence module (HAN) $\\hat{p}_t^v,\\hat{p}_t^a$ in place, the inference module can be trained using the binary cross-entropy loss as shown:", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {h a r d} = \\mathrm {B C E} \\left(p _ {t} ^ {v}, \\hat {y} _ {t} ^ {v}\\right) + \\mathrm {B C E} \\left(p _ {t} ^ {a}, \\hat {y} _ {t} ^ {a}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 132, + 890, + 148 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Training with Pseudo-Label Uncertainty", + "text_level": 1, + "bbox": [ + 498, + 157, + 846, + 175 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While pseudo-labels do provide additional supervision for better training of the inference module, they could potentially be noisy, leading to occasionally incorrect training signals. To ameliorate this problem, we propose an uncertainty-weighted pseudo-label training scheme to improve the robustness of the learning process. Instead of simply training with the binary pseudo-labels $\\hat{y}_t^v,\\hat{y}_t^a$ , we leverage the confidence of the pseudo-label estimation module (associated with the predicted pseudo-label) to weigh the training signal for the inference module. This confidence score serves as a measure of the pseudo-label generation module's uncertainty of its prediction. This may be represented as:", + "bbox": [ + 496, + 181, + 890, + 378 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {p} _ {t} ^ {v} = \\operatorname {S i g m o i d} \\left(\\hat {z} _ {t} ^ {v} - \\theta^ {v}\\right) \\odot y; \\quad \\hat {p} _ {t} ^ {a} = \\operatorname {S i g m o i d} \\left(\\hat {z} _ {t} ^ {a} - \\theta^ {a}\\right) \\odot y. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 386, + 890, + 416 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In other words, considering the visual pseudo-label generation pipeline as an example, the farther the logit $\\hat{z}_t^v$ is from the threshold $\\theta^v$ , whether much lower or higher, the more confident the pseudo-label generation module is about the label it predicts (either approaching 0 or 1). Conversely, the closer the logit is to the threshold, the less the certainty about the correctness of the pseudo-labels (probabilities closer to 0.5). An analogous explanation also holds for the audio pseudo-labels. With the uncertainty-weighted pseudo-labels in place, the inference module (HAN) can be trained with the following uncertainty-weighted pseudo-label loss:", + "bbox": [ + 496, + 417, + 890, + 597 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s o f t}} = \\mathrm {B C E} \\left(p _ {t} ^ {v}, \\hat {p} _ {t} ^ {v}\\right) + \\mathrm {B C E} \\left(p _ {t} ^ {a}, \\hat {p} _ {t} ^ {a}\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 604, + 890, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Uncertainty-weighted Feature Mixup", + "text_level": 1, + "bbox": [ + 498, + 635, + 821, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to the lack of full supervision for the weakly-supervised AVVP task, we explore the efficacy of additional regularization via self-supervision to help the models generalize better. Towards this end, prior pseudo-label generation approaches [26, 39] often employ contrastive learning as a tool to better train the inference module. However, due to the inherent noise in the estimated pseudo-labels, positive samples and negative samples may be mislabeled, decreasing the effectiveness of the self-supervisory training. As an alternative, in this work, we explore the effectiveness of feature mixing, as a self-supervisory training signal for additional regularization. In this setting, we mixup the estimated features of any two segments, additively, and train the model to predict the union of the labels of the two segments. However, since the labels in our setting are noisy, the mixed feature is assigned a label derived from a weighted sum of", + "bbox": [ + 496, + 659, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13565", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the uncertainty-weighted pseudo-labels of each of the two segment features. This is illustrated below:", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {f} _ {t _ {i}, t _ {j}} ^ {v} = \\lambda \\tilde {f} _ {t _ {i}} ^ {v} + (1 - \\lambda) \\tilde {f} _ {t _ {j}} ^ {v}, \\quad \\bar {p} _ {t _ {i}, t _ {j}} ^ {v} = \\lambda \\hat {p} _ {t _ {i}} ^ {v} + (1 - \\lambda) \\hat {p} _ {t _ {j}} ^ {v} \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 128, + 467, + 148 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {f} _ {t _ {i}, t _ {j}} ^ {a} = \\lambda \\tilde {f} _ {t _ {i}} ^ {a} + (1 - \\lambda) \\tilde {f} _ {t _ {j}} ^ {a}, \\quad \\bar {p} _ {t _ {i}, t _ {j}} ^ {a} = \\lambda \\hat {p} _ {t _ {i}} ^ {a} + (1 - \\lambda) \\hat {p} _ {t _ {j}} ^ {a}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 88, + 151, + 467, + 172 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda \\sim \\mathrm{Beta}(\\alpha, \\alpha)$ and $\\alpha$ is a hyper-parameter controlling the Beta distribution, and $t_i$ and $t_j$ indicate two segment indices in a batch of video segments.", + "bbox": [ + 76, + 178, + 468, + 223 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After mixing the uni-modal segment-level features, we pass them through the classifier of the inference module and apply the sigmoid function to the output, obtaining mixed segment-level event probabilities $p_t^{mix - v}$ and $p_t^{mix - a}$ . These are used to train the inference model with the uncertainty-aware mixup loss, as shown below:", + "bbox": [ + 76, + 223, + 468, + 314 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m i x} = \\mathrm {B C E} \\left(p _ {t} ^ {m i x - v}, \\bar {p} _ {t} ^ {v}\\right) + \\mathrm {B C E} \\left(p _ {t} ^ {m i x - a}, \\bar {p} _ {t} ^ {a}\\right). \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 321, + 468, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Class-balanced Loss Re-weighting", + "text_level": 1, + "bbox": [ + 76, + 348, + 375, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Besides the aforementioned challenges of the AVVP task, most of the events in the event set are absent in the pseudolabels of any (segment of a) video (i.e., most event classes are negative events) and only a few events are present (i.e., positive events are much fewer in number). As a result, the model is dominated by the loss from the negative events. When trained without factoring in this bias, the classifier tends to overfit the negative labels and ignore the positive ones. To address this class imbalance issue, we introduce a class-balanced loss re-weighting strategy to re-balance the importance of the losses from the negative and positive events for the uncertainty-weighted pseudo-label loss. Specifically, the loss from the positive events is multiplied by a weight proportional to the frequency of the segments with the negative events in the pseudo-labels, while the loss from the negative events is multiplied by a weight proportional to the frequency of the segments with the positive events in the pseudo-labels, as shown below:", + "bbox": [ + 75, + 371, + 468, + 643 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {w - s o f t} = \\sum_ {m \\in \\{v, a \\}} w _ {p o s} ^ {m} \\cdot y \\cdot \\mathrm {B C E} \\left(p _ {t} ^ {m}, \\hat {p} _ {t} ^ {m}\\right) + \\\\ w _ {n e g} ^ {m} \\cdot (1 - y) \\cdot \\mathrm {B C E} \\left(p _ {t} ^ {m}, \\hat {p} _ {t} ^ {m}\\right), \\tag {17} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 93, + 648, + 468, + 696 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nw _ {p o s} ^ {m} = \\frac {\\sum_ {i = 1} ^ {N} \\sum_ {t = 1} ^ {T} \\sum_ {c = 1} ^ {C} \\left(1 - \\hat {y} _ {i , t , c} ^ {m}\\right)}{N T C} \\times W, \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 705, + 468, + 739 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nw _ {n e g} ^ {m} = \\frac {\\sum_ {i = 1} ^ {N} \\sum_ {t = 1} ^ {T} \\sum_ {c = 1} ^ {C} \\hat {y} _ {i , t , c} ^ {m}}{N T C}, \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 741, + 468, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $N$ denotes the number of videos in the training set, and $W$ is a hyper-parameter.", + "bbox": [ + 76, + 782, + 467, + 813 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In summary, the inference module is trained on the AVVP task with the proposed class-balanced re-weighting, applied to the uncertainty-weighted classification loss, and the uncertainty-weighted feature mixup loss, as shown below:", + "bbox": [ + 76, + 813, + 468, + 886 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {w - s o f t} + \\mathcal {L} _ {\\text {m i x}} + \\mathcal {L} _ {\\text {v i d e o}}. \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 886, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 90, + 632, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We assess the performance of UWAV empirically across two challenging datasets and report its performance, comparing it with existing state-of-the-art approaches both quantitatively and qualitatively. Additionally, through multiple ablation studies, we bring out the effectiveness of the different elements of our proposed approach and the choices of different hyper-parameters. For additional details, ablation studies, and more qualitative results, we refer the reader to our supplementary material.", + "bbox": [ + 496, + 114, + 890, + 251 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 498, + 260, + 689, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets: We evaluate all competing methods on the Look, Listen, and Parse (LLP) dataset [35], which is the principal benchmark dataset for the AVVP task. The dataset consists of 11,849 video clips sourced from YouTube. Each clip is 10 seconds long and represents one or more of 25 diverse event categories, such as human activities, animals, musical instruments, and vehicles. The dataset is split into training, validation, and testing sets, following the official split [35]: 10,000 videos for training, 649 videos for validation, and 1,200 videos for testing. While the training set of this dataset is only associated with video-level labels of the events, the validation and testing split is labeled with segment-level event labels for evaluation purposes. Additionally, to demonstrate the generalizability of our method, we conduct a similar study on the Audio Visual Event (AVE) recognition dataset [34]. The AVE dataset consists of 4,143 video clips crawled from YouTube, each 10 seconds long. It is split into 3,339 videos for training, 402 for validation, and 402 for testing. It includes 29 event categories encompassing human activities, animals, musical instruments, vehicles, and a \"background\" class (i.e., no event occurs). Unlike the LLP dataset, each video in the AVE dataset contains only one audio-visual event. Here too, the training data is only provided with video-level labels while the validation and test splits are annotated with ground-truth event labels for each one-second segment, which either is a specific audio-visual event or \"background\".", + "bbox": [ + 496, + 284, + 890, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics: For the LLP dataset, following the standard evaluation protocol [35], all models are evaluated using macro F1-scores calculated for the following event types: (i) audio-only (A), (ii) visual-only (V), and (iii) audio-visual (AV). Type@AV (Type) and Event@AV (Event) are two additional metrics that evaluate the overall performance of the model, where Type@AV is the mean of the F1-scores for the A, V, and AV events, while Event@AV is the F1-score of all events regardless of the modality in which they occur. Evaluations are conducted at both the segment-level and the event-level. At the segment-level, the model's predictions are compared with the ground truth on a per-segment basis. At the event-level, consecutive positive segments for", + "bbox": [ + 496, + 703, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13566", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a57e368bf5bb9deff42234c3242042e0a6367c55db090aa20d4068a02d8e43f2.jpg", + "table_caption": [ + "Table 1. Comparison with state-of-the-arts methods on the LLP dataset. Best performances are in bold, second-best in underlined." + ], + "table_footnote": [], + "table_body": "
MethodSegment-levelEvent-level
AVAVTypeEventAVAVTypeEvent
HAN [35]60.152.948.954.055.451.348.943.047.748.0
MA [39]60.360.055.158.957.953.656.449.053.050.6
JoMoLD [8]61.363.857.260.859.953.959.949.654.552.5
CMPAE [10]64.266.459.263.362.856.663.751.857.455.7
PoiBin [27]63.163.557.761.460.654.160.351.555.252.3
VPLAN [50]60.564.858.361.259.451.461.551.254.750.8
VALOR [17]61.865.958.462.061.555.462.652.256.754.2
LSLD [9]62.767.159.463.162.255.764.352.657.655.2
PPL [26]65.966.761.964.863.757.364.354.359.957.9
CoLeaf [31]64.264.459.362.662.557.663.254.257.955.6
LEAP [51]62.765.659.362.561.856.463.154.157.855.0
UWAV (Ours)64.270.063.465.963.958.666.757.560.957.4
", + "bbox": [ + 161, + 107, + 803, + 344 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the same event are grouped together as a single event. The F1-score is then computed using a mIoU threshold of 0.5. For the AVE dataset, we follow Tian et al. [34] and use accuracy as the evaluation metric. An event prediction of a segment is considered correct if it matches the ground-truth label for that segment.", + "bbox": [ + 75, + 357, + 468, + 448 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details: In line with prior work [35], each 10-second video in both datasets is split into 10 segments of one second each, where each segment contains 8 frames. For the LLP dataset, pre-trained ResNet-152 [13] and $\\mathrm{R}(2 + 1)\\mathrm{D - }18$ [37] are used to extract 2D and 3D visual features, respectively. The pre-trained VGGish [14] network is used to extract features from the audio, sampled at $16\\mathrm{KHz}$ . For the AVE dataset however, akin to prior work [17], we extract visual features from pre-trained CLIP and $\\mathrm{R}(2 + 1)\\mathrm{D}$ while CLAP is used to embed the audio stream. For both datasets, we set the number of encoder blocks $L$ in the temporal-aware model to 5, $\\alpha$ for the Beta distribution in the feature mixup strategy to 1.7, and $W$ in the class-balanced loss re-weighting step to 0.5. The pseudo-label generation modules and the inference model (HAN) are trained with the AdamW optimizer [20]. For improved performance on the AVE dataset, we replace $\\hat{p}_{t_i}^a,\\hat{p}_{t_j}^a$ in Eq. 15 with $[\\hat{p}_{t_i}^a ],[\\hat{y}_{t_j}^a ]$ and make corresponding modifications in the visual counterparts as well.", + "bbox": [ + 75, + 454, + 468, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines: We demonstrate the effectiveness of UWAV by comparing against an extensive set of baselines. For the LLP dataset, this includes video-level pseudo-label generation methods (MA [39], JoMoLD [8]), segment-level pseudo-label generation methods (VALOR [17], LSLD [9], PPL [26]), and the recently released works (CoLeaf [31], LEAP [51]). On the other hand, for the AVE dataset, baseline approaches with publicly available implementation, which use the state-of-the-art feature backbones (akin to ours), such as HAN [35] and VALOR [17] are used.", + "bbox": [ + 75, + 750, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Results", + "text_level": 1, + "bbox": [ + 500, + 356, + 591, + 371 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Previous Methods on LLP: As shown in Table 1, UWAV surpasses previous methods, across almost all metrics. Notably, we achieve an F-score of 70.0 on the segment-level visual event, 65.9 on the segment-level Type@AV, and 66.7 on the event-level visual event metric. This corresponds to a gain of $1.1\\%$ on segment-level Type@AV F-score and a $1\\%$ improvement on event-level Type@AV F-score, over our closest competitor PPL [26]. Of particular note, is the fact that our segment-level and event-level F-scores improve by more than $3\\%$ , over PPL, for visual events. When compared to other recently published works, such as VALOR [17], CoLeaf [31] and LEAP [51], UWAV outperforms them by up to $3\\%$ on both segment and event-level Type@AV F-score while gains on visual events are up to $5\\%$ on segment-level F-score.", + "bbox": [ + 496, + 380, + 890, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "These observations are consistent with our qualitative comparisons, as well. In the example on the left of Figure 3, our model successfully recognizes and temporally localizes the lawn mower event visually, whereas VALOR [17] (a recent state-of-the-art approach with publicly available implementation) misclassifies it as a chainsaw. Additionally, our model also accurately localize the intermittent sound of the lawn mower. In contrast, VALOR not only misclassifies the sound of the lawn mower as that of a chainsaw but also incorrectly predicts that someone is talking in the video. In the example on the right of Figure 3, our model does not err in recognizing either the visual presence or the audio presence of the telephone, while VALOR fails to accurately predict the events in either modality.", + "bbox": [ + 496, + 625, + 890, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Previous Methods on AVE: To demonstrate the generalizability of our method, we evaluate UWAV on the AVE [34] dataset and compare its per", + "bbox": [ + 496, + 854, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13567", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/884745ea97cc18cb7ad4f14e8126d44f7e5060bd4eeb05af4c5ba4c0e7dfdbad.jpg", + "image_caption": [ + "Figure 3. Comparison between predictions by UWAV and competing AVVP methods on the LLP dataset. \"GT\": ground truth." + ], + "image_footnote": [], + "bbox": [ + 119, + 89, + 504, + 291 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7bb483f21c0388ed8ddd8e93910587c4b0b02a5c664f4c925aacace79b49e052.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 88, + 834, + 290 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0ca1b2f5de95940b4a632767f2a71890cfbc3579ee1bd04dd870f564a2d7b875.jpg", + "table_caption": [ + "Table 2. Model performances on the AVE dataset. CLIP, R(2+1)D-18, and CLAP are used as feature backbones." + ], + "table_footnote": [], + "table_body": "
MethodHAN [35]VALOR [17]UWAV (Ours)
Acc.(%)75.380.480.6
", + "bbox": [ + 96, + 359, + 442, + 397 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/872710bc908d3a91b1bc115e3babd79d766cee77fdde393b417ce71c1ab65eec.jpg", + "table_caption": [ + "Table 3. Accuracy of the generated pseudo-labels on LLP." + ], + "table_footnote": [], + "table_body": "
MethodSegment-level
AVAVTypeEvent
VALOR [17]80.561.755.766.074.6
PPL [26]61.761.857.560.659.4
UWAV (Ours)78.474.565.572.878.4
", + "bbox": [ + 98, + 422, + 444, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "formance with that of previous works. From Table 2, we observe that with the same backbone features, UWAV surpasses VALOR, our closest competitor, even on this small-scale dataset.", + "bbox": [ + 75, + 515, + 468, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Accuracy of the Generated Pseudo-Labels: To evaluate the efficacy of our pseudo-label generation pipeline, we compare the accuracy of our generated pseudo-labels against those obtained from competing methods (with publicly available implementation) [17, 26] on the test set of the LLP dataset. As shown in Table 3, our pre-trained temporally-dependent pseudo-label generation scheme generates more accurate segment-level pseudo-labels than previous methods, by up to $6\\%$ on the segment-level Type@AV F-score, attesting to the advantages of factoring in intersegment temporal dependencies.", + "bbox": [ + 75, + 592, + 468, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 770, + 230, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To demonstrate the potency of the different elements of our proposed method, UWAV, we conduct ablation studies. In particular, the proposed uncertainty-weighted pseudo-label based training, the uncertainty-weighted feature mixup scheme, and the class-balanced loss re-weighting schemes are ablated. As shown in Table 4, incorporating the uncertainty-weighted pseudo-label training step improves", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d5844c5cbc20d84eb9145b330f77ca6e3482b488960220cf53fb5c9c6109838f.jpg", + "table_caption": [ + "Table 4. Ablation study of the proposed components in UWAV. \"Binary\" denotes training with binary pseudo-labels. \"Soft\" denotes training with uncertainty-weighted pseudo-labels." + ], + "table_footnote": [], + "table_body": "
BinarySoftRe-weightMixupSegment-level
AVAVTypeEvent
✓62.767.761.964.262.2
✓63.068.361.864.462.8
✓✓63.669.563.065.463.1
✓✓63.969.062.865.263.4
✓✓✓64.270.063.465.963.9
", + "bbox": [ + 503, + 375, + 888, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the segment-level Type@AV F-score by $2\\%$ , compared to using binary pseudo-labels. This demonstrates the benefit of accounting for the uncertainty in the pseudo-label estimation module. Moreover, sans the class-balanced loss weighting strategy, the model's performance is worse off by $1\\%$ on the Type@AV F-score, revealing the erroneous bias in the model's prediction arising from a skew of the class distribution. On the other hand, introducing the uncertainty-weighted feature mixup results in a gain of $0.8\\%$ on the Type@AV F-score, which underscores the importance of this self-supervised regularization.", + "bbox": [ + 496, + 494, + 892, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 678, + 617, + 693 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we address the challenging task of weakly-supervised audio-visual video parsing (AVVP), which presents a two-fold challenge: (i) potential misalignment between the events of the audio and visual streams, and (ii) the lack of fine-grained labels for each modality. We observe that by considering the temporal relationship between segments, our proposed method (UWAV) is able to provide more reliable pseudo-labels for better training of the inference module. In addition, by factoring in the uncertainty associated with these estimated pseudo-labels, regularizing the training process with a feature mixup strategy, and correcting for class imbalance, UWAV achieves state-of-the-art results on the LLP and AVE datasets.", + "bbox": [ + 496, + 703, + 892, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13568", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4", + "[2] Moitreya Chatterjee, Jonathan Le Roux, Narendra Ahuja, and Anoop Cherian. Visual scene graphs for audio source separation. In ICCV, 2021. 2", + "[3] Moitreya Chatterjee, Narendra Ahuja, and Anoop Cherian. Learning audio-visual dynamics using scene graphs for audio source separation. In NeurIPS, 2022. 1", + "[4] Changan Chen, Unnat Jain, Carl Schissler, Sebastia Vicenc Amengual Gari, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audio-visual navigation in 3d environments. In ECCV, 2020. 2", + "[5] Changan Chen, Ziad Al-Halah, and Kristen Grauman. Semantic audio-visual navigation. In CVPR, 2021.", + "[6] Changan Chen, Sagnik Majumder, Ziad Al-Halah, Ruohan Gao, Santhosh Kumar Ramakrishnan, and Kristen Grauman. Learning to set waypoints for audio-visual navigation. In ICLR, 2021. 2", + "[7] Jiaben Chen, Renrui Zhang, Dongze Lian, Jiaqi Yang, Ziyao Zeng, and Jianbo Shi. iquery: Instruments as queries for audio-visual sound separation. In CVPR, 2023. 1, 2", + "[8] Haoyue Cheng, Zhaoyang Liu, Hang Zhou, Chen Qian, Wayne Wu, and Limin Wang. Joint-modal label denoising for weakly-supervised audio-visual video parsing. In ECCV, 2022. 2, 3, 7", + "[9] Yingying Fan, Yu Wu, Yutian Lin, and Bo Du. Revisit weakly-supervised audio-visual video parsing from the language perspective. In NeurIPS, 2023. 2, 3, 4, 7", + "[10] Junyu Gao, Mengyuan Chen, and Changsheng Xu. Collecting cross-modal presence-absence evidence for weakly-supervised audio-visual event perception. In CVPR, 2023. 7", + "[11] Shijie Geng, Peng Gao, Moitreya Chatterjee, Chiori Hori, Jonathan Le Roux, Yongfeng Zhang, Hongsheng Li, and Anoop Cherian. Dynamic graph representation learning for video dialog via multi-modal shuffled transformers. In AAAI, 2021. 2", + "[12] Tiantian Geng, Teng Wang, Jinming Duan, Runmin Cong, and Feng Zheng. Dense-localizing audio-visual events in untrimmed videos: A large-scale benchmark and baseline. In CVPR, 2023. 2, 4, 5", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 7", + "[14] Shawn Hershey, Sourish Chaudhuri, Daniel PW Ellis, Jort F Gemmeke, Aren Jansen, R Channing Moore, Manoj Plakal, Devin Platt, Rif A Saurous, Bryan Seybold, et al. Cnn architectures for large-scale audio classification. In ICASSP, 2017. 7", + "[15] Xixi Hu, Ziyang Chen, and Andrew Owens. Mix and localize: Localizing sound sources in mixtures. In CVPR, 2022. 2" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Egocentric audio-visual object localization. In CVPR, 2023. 2", + "[17] Yung-Hsuan Lai, Yen-Chun Chen, and Frank Wang. Modality-independent teachers meet weakly-supervised audio-visual event parser. In NeurIPS, 2023. 2, 3, 4, 7, 8", + "[18] Guangyao Li, Yake Wei, Yapeng Tian, Chenliang Xu, JiRong Wen, and Di Hu. Learning to answer questions in dynamic audio-visual scenarios. In CVPR, 2022. 2", + "[19] Jinxiang Liu, Yikun Liu, Fei Zhang, Chen Ju, Ya Zhang, and Yanfeng Wang. Audio-visual segmentation via unlabeled frame exploitation. In CVPR, 2024. 1, 2", + "[20] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 7", + "[21] Sagnik Majumder, Ziad Al-Halah, and Kristen Grauman. Move2hear: Active audio-visual source separation. In ICCV, 2021. 2", + "[22] Yuxin Mao, Jing Zhang, Mochu Xiang, Yiran Zhong, and Yuchao Dai. Multimodal variational auto-encoder based audio-visual segmentation. In ICCV, 2023. 2", + "[23] Shentong Mo and Yapeng Tian. Multi-modal grouping network for weakly-supervised audio-visual video parsing. In NeurIPS, 2022. 2, 3", + "[24] Shentong Mo and Yapeng Tian. Audio-visual grouping network for sound localization from mixtures. In CVPR, 2023. 1, 2", + "[25] Kranti Kumar Parida, Siddharth Srivastava, and Gaurav Sharma. Beyond mono to binaural: Generating binaural audio from mono audio with depth and cross modal attention. In WACV, 2022. 2", + "[26] Kranthi Kumar Rachavarapu, Kalyan Ramakrishnan, et al. Weakly-supervised audio-visual video parsing with prototype-based pseudo-labeling. In CVPR, 2024. 2, 3, 5, 7, 8", + "[27] Kranthi Kumar Rachavarapu et al. Boosting positive segments for weakly-supervised audio-visual video parsing. In ICCV, 2023. 7", + "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 4", + "[29] Varshanth Rao, Md Ibrahim Khalil, Haoda Li, Peng Dai, and Juwei Lu. Dual perspective network for audio-visual event localization. In ECCV, 2022. 1, 2", + "[30] Ludan Ruan, Yiyang Ma, Huan Yang, Huiguo He, Bei Liu, Jianlong Fu, Nicholas Jing Yuan, Qin Jin, and Baining Guo. Mm-diffusion: Learning multi-modal diffusion models for joint audio and video generation. In CVPR, 2023. 2", + "[31] Faegheh Sardari, Armin Mustafa, Philip JB Jackson, and Adrian Hilton. Coleaf: A contrastive-collaborative learning framework for weakly supervised audio-visual video parsing. In ECCV, 2024. 7", + "[32] Arda Senocak, Hyeonggon Ryu, Junsik Kim, Tae-Hyun Oh, Hanspeter Pfister, and Joon Son Chung. Sound source localization is all about cross-modal alignment. In ICCV, 2023. 1, 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13569", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Ankit Shah, Shijie Geng, Peng Gao, Anoop Cherian, Takaaki Hori, Tim K Marks, Jonathan Le Roux, and Chiori Hori. Audio-visual scene-aware dialog and reasoning using audio-visual transformers with joint student-teacher learning. In ICASSP, 2022. 2", + "[34] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-visual event localization in unconstrained videos. In ECCV, 2018. 1, 2, 6, 7", + "[35] Yapeng Tian, Dingzeyu Li, and Chenliang Xu. Unified multisensory perception: Weakly-supervised audio-visual video parsing. In ECCV, 2020. 1, 2, 3, 6, 7, 8", + "[36] Yapeng Tian, Di Hu, and Chenliang Xu. Cyclic co-learning of sounding object visual grounding and sound separation. In CVPR, 2021. 2", + "[37] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In CVPR, 2018. 7", + "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2, 3, 4", + "[39] Yu Wu and Yi Yang. Exploring heterogeneous clues for weakly-supervised audio-visual video parsing. In CVPR, 2021. 2, 3, 5, 7", + "[40] Yusong Wu, Ke Chen, Tianyu Zhang, Yuchen Hui, Taylor Berg-Kirkpatrick, and Shlomo Dubnov. Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation. In ICASSP, 2023. 2, 3", + "[41] Yazhou Xing, Yingqing He, Zeyue Tian, Xintao Wang, and Qifeng Chen. Seeing and hearing: Open-domain visual-audio generation with diffusion latent aligners. In CVPR, 2024. 2", + "[42] Qi Yang, Xing Nie, Tong Li, Pengfei Gao, Ying Guo, Cheng Zhen, Pengfei Yan, and Shiming Xiang. Cooperation does matter: Exploring multi-order bilateral relations for audio-visual segmentation. In CVPR, 2024. 2", + "[43] Yuxin Ye, Wenming Yang, and Yapeng Tian. Lavss: Location-guided audio-visual spatial audio separation. In WACV, 2024. 1, 2", + "[44] Abdelrahman Younes, Daniel Honerkamp, Tim Welschehold, and Abhinav Valada. Catch me if you hear me: Audio-visual navigation in complex unmapped environments with moving sounds. IEEE Robotics and Automation Letters, 2023. 2", + "[45] Jiashuo Yu, Ying Cheng, Rui-Wei Zhao, Rui Feng, and Yuejie Zhang. Mm-pyramid: Multimodal pyramid attentional network for audio-visual event localization and video parsing. In ACM MM, 2022. 2, 3", + "[46] Yinfeng Yu, Wenbing Huang, Fuchun Sun, Changan Chen, Yikai Wang, and Xiaohong Liu. Sound adversarial audio-visual navigation. In ICLR, 2022. 2", + "[47] Heeseung Yun, Youngjae Yu, Wonsuk Yang, Kangil Lee, and Gunhee Kim. Pano-avqa: Grounded audio-visual question answering on 360deg videos. In ICCV, 2021. 2", + "[48] Jinxing Zhou, Liang Zheng, Yiran Zhong, Shijie Hao, and Meng Wang. Positive sample propagation along the audio-visual event line. In CVPR, 2021. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Jinxing Zhou, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audio-visual segmentation. In ECCV, 2022. 1, 2", + "[50] Jinxing Zhou, Dan Guo, Yiran Zhong, and Meng Wang. Improving audio-visual video parsing with pseudo visual labels. arXiv preprint arXiv:2303.02344, 2023. 2, 3, 4, 7", + "[51] Jinxing Zhou, Dan Guo, Yuxin Mao, Yiran Zhong, Xiaojun Chang, and Meng Wang. Label-anticipated event disentanglement for audio-visual video parsing. arXiv preprint arXiv:2407.08126, 2024. 2, 3, 7" + ], + "bbox": [ + 501, + 90, + 892, + 246 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13570", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/UWAV_ Uncertainty-weighted Weakly-supervised Audio-Visual Video Parsing/82f7769f-b767-4619-96ea-539823053afb_model.json b/2025/UWAV_ Uncertainty-weighted Weakly-supervised Audio-Visual Video Parsing/82f7769f-b767-4619-96ea-539823053afb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3c4dd145689e5229ee06b667427baa92a76cd7de --- /dev/null +++ b/2025/UWAV_ Uncertainty-weighted Weakly-supervised Audio-Visual Video Parsing/82f7769f-b767-4619-96ea-539823053afb_model.json @@ -0,0 +1,2013 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.096, + 0.131, + 0.877, + 0.155 + ], + "angle": 0, + "content": "UWAV: Uncertainty-weighted Weakly-supervised Audio-Visual Video Parsing" + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.18, + 0.825, + 0.272 + ], + "angle": 0, + "content": "Yung-Hsuan Lai\\(^{1,\\dagger}\\), Janek Ebbers\\(^{3}\\), Yu-Chiang Frank Wang\\(^{1,2}\\), François Germain\\(^{3}\\), Michael Jeffrey Jones\\(^{3}\\), Moitreya Chatterjee\\(^{3,\\ddagger}\\), \n\\(^{1}\\) Graduate Institute of Communication Engineering, National Taiwan University \\(^{2}\\) NVIDIA, Taiwan \n\\(^{3}\\) Mitsubishi Electric Research Labs (MERL) \n\\(\\dagger\\)r10942097@ntu.edu.tw \\(\\ddagger\\)metro.smiles@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.314, + 0.321 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.474, + 0.732 + ], + "angle": 0, + "content": "Audio-Visual Video Parsing (AVVP) entails the challenging task of localizing both uni-modal events (i.e., those occurring exclusively in either the visual or acoustic modality of a video) and multi-modal events (i.e., those occurring in both modalities concurrently). Moreover, the prohibitive cost of annotating training data with the class labels of all these events, along with their start and end times, imposes constraints on the scalability of AVVP techniques unless they can be trained in a weakly-supervised setting, where only modality-agnostic, video-level labels are available in the training data. To this end, recently proposed approaches seek to generate segment-level pseudolabels to better guide model training. However, the absence of inter-segment dependencies when generating these pseudo-labels and the general bias towards predicting labels that are absent in a segment limit their performance. This work proposes a novel approach towards overcoming these weaknesses called Uncertainty-weighted Weakly-supervised Audio-visual Video Parsing (UWAV). Additionally, our innovative approach factors in the uncertainty associated with these estimated pseudo-labels and incorporates a feature mixup based training regularization for improved training. Empirical results show that UWAV outperforms state-of-the-art methods for the AVVP task on multiple metrics, across two different datasets, attesting to its effectiveness and generalizability." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.747, + 0.21, + 0.763 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.773, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Events that occur in the real world, often leave their imprint on the acoustic and visual modalities. Humans rely heavily on the synergy between their senses of sight and hearing to interpret such events. Audio-visual learning, which seeks to equip machines with a similar synergy, has emerged as one of the most important research areas within the multi" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.307, + 0.891, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.51, + 0.895, + 0.582 + ], + "angle": 0, + "content": "Figure 1. A weakly-supervised AVVP task example. Events, considered in this task, might be unimodal or multimodal. Even multimodal events, may not be temporally aligned in the audio and visual modalities, e.g. the cello might only be visible in the first few seconds but might produce music, throughout the video." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.893, + 0.794 + ], + "angle": 0, + "content": "modal machine learning community. It aims to leverage both these senses (modalities) jointly, to enhance machine perception and understanding of real-world events. Various audio-visual learning tasks have been studied towards this end, including audio-visual segmentation [19, 49], sound source localization [24, 32], audio-visual event localization [29, 34], and audio-visual sound separation [3, 7, 43]. However, many of these tasks assume that audio and visual streams would always be temporally aligned. This assumption often fails in real-world scenarios, where the sonic and visual imprints of events may not perfectly overlap. For instance, one might hear an emergency siren approaching from a distance before it appears in the field of view." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this work, in order to better understand the events occurring in a video, we explore the task of Audio-Visual Video Parsing (AVVP) [35]. Its goal is to recognize and localize all audio, visual, and audio-visual events occurring in the video. See Figure 1 for an example of this task. The task setup is to perform this prediction for every one-second segment of a video. This task poses two principal challenges," + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.875, + 0.468, + 0.9 + ], + "angle": 0, + "content": "\\(^{1}\\)Code is available at: https://github.com/Franklin905/UWAV." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13561" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "from a machine learning standpoint: (i) The audio and visual events, that occur, might not be temporally aligned, e.g. if an event becomes audible before its source enters the camera field of view, or the sound source is not visible at all, and (ii) due to the high costs of annotating video segments with per-segment labels, only modality-agnostic, video-level labels are provided during training, i.e., these labels specify which events occur in a video but lack details about the segments or the modality in which they occur." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.232, + 0.473, + 0.671 + ], + "angle": 0, + "content": "Prior works in the area can be grouped into two orthogonal research directions. The first focuses on enhancing model architectures [23, 45, 51]. Despite advancements in this direction, the absence of fine-grained labels to guide the model during training continues to pose an impediment towards the generalizability of such models. As a result, recent approaches have focused on the second direction of research which aims at generating richer pseudo-labels for improved training, either at the video-level [8, 39] or segment-level [9, 17, 26, 50]. In particular, Rachavarapu et al. [26] propose prototype-based pseudo-labeling (PPL), which seeks to train a pseudo-label generation module in conjunction with a core inference module. However, due to the lack of sufficient training data, this method struggles to generalize. On the other hand, VPLAN [50], VALOR [17], and LSLD [9] leverage large-scale pre-trained foundation models, such as CLIP [28] and CLAP [40], along with ground-truth video-level labels to generate segment-level pseudo-labels for each of the two modalities. Audio/Visual segments (e.g. the audio corresponding to the segment in question and the visual frame at the center of the segment) are fed into CLAP/CLIP, one segment at a time, to generate these pseudo-labels. Despite the significant improvement that these pseudo-label generation methods achieved, the correctness of the generated labels is still limited, constrained primarily by the lack of understanding of intersegment dynamics. For instance, if a crowd is cheering in a segment of the video, it is more likely that the crowd might also be clapping right before or after." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.903 + ], + "angle": 0, + "content": "To address the oversight of inter-segment dependencies and other shortcomings in existing pseudo-label generation methods, we introduce a novel, uncertainty-based, weakly-supervised, video parsing model called UWAV (Uncertainty-weighted Weakly-supervised Audiovisual Video Parsing), capable of generating improved segment-level pseudo-labels for better training of the inference module. We resort to transformer modules [38] to equip UWAV with the ability to capture temporal relationships between segments and pre-train it on a large-scale, supervised audio-visual event localization dataset [12]. Subsequently, this pre-trained model is used to generate segment-level pseudo-labels for each modality on a target, small-scale dataset which only provides weak (i.e., video-level) supervision. Such a design permits a more holis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.513 + ], + "angle": 0, + "content": "tic understanding of the video, resulting in more accurate pseudo-labels. Additionally, UWAV factors in the uncertainty associated with these estimated pseudo-labels in its optimization. That uncertainty is the result of the shift in the domain of the target dataset, insufficient model capacity, etc. and is computed at training time as the confidence scores associated with these labels. To further enhance the model's ability to learn in this small-scale, weakly-supervised data regime, we also employ a feature mixup strategy. This approach adds more regularization constraints by training on mixed segment features alongside interpolated pseudo-labels, which not only lessens the influence of noise but also enriches the training data, thereby reducing overfitting. Moreover, UWAV addresses a critical class imbalance issue in the pseudo-label enriched training data, viz. most event classes in any given segment of a video are absent/negative (i.e., they do not occur), while only a handful of them do. This creates a natural bias in the training set, making it difficult to learn the positive events. To counter this, we propose a class-frequency aware weighting strategy that lays greater emphasis on the accurate classification and localization of positive events. By incorporating these components into its design, our proposed method (UWAV) outperforms competing state-of-the-art approaches across two different datasets, viz. Look, Listen, and Parse (LLP) [35] and the Audio-Visual Event Localization (AVE) [34], on multiple evaluation metrics. In summary, our contributions are the following:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.515, + 0.892, + 0.559 + ], + "angle": 0, + "content": "- We introduce a novel, weakly-supervised method called UWAV, capable of synthesizing temporally coherent pseudo-labels for the AVVP task." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.56, + 0.892, + 0.62 + ], + "angle": 0, + "content": "- To the best of our knowledge, ours is the first method for the AVVP task, which factors in the uncertainty associated with the estimated pseudo-labels while also regularizing it with a feature mixup strategy." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.62, + 0.892, + 0.667 + ], + "angle": 0, + "content": "- UWAV outperforms competing state-of-the-art approaches for the AVVP task, across two different datasets on multiple metrics which attest to its generalizability." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.515, + 0.892, + 0.667 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.679, + 0.651, + 0.696 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Audio-Visual Learning: Audio-visual learning has emerged as an area of active research, aiming to develop models that synergistically integrate information from both audio and visual modalities for improved perception and understanding. Towards this end, various audio-visual tasks have been explored by the community so far, such as audio-visual segmentation [19, 22, 42, 49], sound source localization [15, 16, 24, 32], event localization [29, 34, 48], navigation [4-6, 21, 44, 46], generation [25, 30, 41], question answering [11, 18, 33, 47], and sound source separation [2, 7, 36, 43]. In this work, we focus on the task of audio-visual video parsing (AVVP) where the goal is to temporally localize events occurring in a video." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13562" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.471, + 0.168 + ], + "angle": 0, + "content": "Unlike many other audio-visual learning tasks, AVVP does not assume that events are always aligned across modalities. Some events could be exclusively uni-modal while others may have an audio-visual signature, which requires complex reasoning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.184, + 0.473, + 0.501 + ], + "angle": 0, + "content": "Audio-Visual Video Parsing (AVVP): To address the challenges of the AVVP task [17, 26, 35], Tian et al. [35] proposed a Hybrid Attention Network (HAN) and a learnable Multi-modal Multiple Instance Learning (MMIL) pooling module. The HAN model facilitates the exchange of information within and across modalities using self-attention and cross-attention layers, while the MMIL pooling module aggregates segment-level event probabilities from both modalities to produce video-level probabilities. Building on this foundation, recent works advanced the field from the following two perspectives. The first group of studies [23, 45, 51] focuses on enhancing model architectures. In particular, Mo and Tian [23] proposed the Multimodal Grouping Network (MGN) to explicitly group semantically similar features within each modality to improve the reasoning process, while Yu et al. [45] proposed the Multi-modal Pyramid Attentional Network (MM-Pyramid) to capture events of varying durations by extracting features at multiple temporal scales. Our proposed method is orthogonal to this line of research and can be integrated with any of these backbones." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.502, + 0.473, + 0.82 + ], + "angle": 0, + "content": "The second direction focuses on generating pseudo-labels for improved training, either at the video-level [8, 39] or the segment-level [9, 17, 26, 50]. VPLAN [50], VALOR [17], and LSLD [9] utilize pre-trained CLIP [28] and CLAP [40] along with ground-truth video-level labels to predict pseudo-labels for each modality on a per-segment basis. In contrast, PPL [26] uses the HAN model itself to generate pseudo-labels by constructing prototype features for each class and assigning labels to each segment based on the similarity between its feature and the prototype features. While these pseudo-label generation methods have substantially improved model performance on the AVVP task, they still exhibit some limitations. For instance, to derive accurate pseudo-labels, PPL might require a large enough training set to learn good prototype features, which might pose challenges when applied to smaller datasets. Our proposed method overcomes this problem. On the other hand, methods that leverage CLIP and CLAP to generate pseudo-labels often ignore temporal relationships between segments or the uncertainty associated with these labels. Our work also seeks to plug this void." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.83, + 0.218, + 0.847 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Problem Formulation: The AVVP task [35] aims to localize all visible and/or audible events in each one-second segment of a video. Specifically, an audible video is split" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.895, + 0.258 + ], + "angle": 0, + "content": "into \\(T\\) one-second segments, denoted as \\(\\{V_t, A_t\\}_{t=1}^T\\). Each segment is annotated with a pair of ground-truth labels \\(y_t^v \\in \\{0,1\\}^C, y_t^a \\in \\{0,1\\}^C\\), where \\(y_t^v\\) denotes visual events, \\(y_t^a\\) denotes audio events, and \\(C\\) denotes the total number of events in the pre-defined event set of the data. However, owing to the weakly-supervised nature of the task setup \\((y_t^v, y_t^a)\\) are unavailable during training. Instead, only the modality-agnostic, video-level labels \\(y \\in \\{0,1\\}^C\\) are available, where 1 indicates the presence of an event at any time (either in the audio or visual stream or both) while 0 indicates an event's absence in the video." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.275, + 0.895, + 0.473 + ], + "angle": 0, + "content": "Pseudo-Label Based AVVP Framework: The Hybrid Attention Network (HAN) [35] is a commonly used model for the AVVP task. The model works by first utilizing pre-trained visual and audio backbones to extract features from the visual and audio segments respectively, which are then projected to two \\(d\\)-dimensional feature spaces. The resulting visual segment-level features are denoted by \\(F^v = \\{f_t^v\\}_{t=1}^T \\in \\mathbb{R}^{T \\times d}\\), while the audio segment-level features are denoted by \\(F^a = \\{f_t^a\\}_{t=1}^T \\in \\mathbb{R}^{T \\times d}\\). These features are provided as input to the HAN model. In the model, information across segments within a modality and across modalities is exchanged through self-attention and cross-attention layers, as shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.481, + 0.892, + 0.519 + ], + "angle": 0, + "content": "\\[\n\\tilde {f} _ {t} ^ {v} = f _ {t} ^ {v} + \\underbrace {\\operatorname {A t t n} \\left(f _ {t} ^ {v} , F ^ {v} , F ^ {v}\\right)} _ {\\text {S e l f - A t t e n t i o n}} + \\underbrace {\\operatorname {A t t n} \\left(f _ {t} ^ {v} , F ^ {a} , F ^ {a}\\right)} _ {\\text {C r o s s - A t t e n t i o n}}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.519, + 0.525, + 0.892, + 0.562 + ], + "angle": 0, + "content": "\\[\n\\tilde {f} _ {t} ^ {a} = f _ {t} ^ {a} + \\underbrace {\\operatorname {A t t n} \\left(f _ {t} ^ {a} , F ^ {a} , F ^ {a}\\right)} _ {\\text {S e l f - A t t e n t i o n}} + \\underbrace {\\operatorname {A t t n} \\left(f _ {t} ^ {a} , F ^ {v} , F ^ {v}\\right)} _ {\\text {C r o s s - A t t e n t i o n}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.572, + 0.892, + 0.708 + ], + "angle": 0, + "content": "where \\(\\mathrm{Attn}(Q,K,V)\\) denotes the standard multi-head attention mechanism [38]. Finally a classifier, shared across both modalities, transforms the visual segment-level features \\(\\tilde{F}^v = \\{\\tilde{f}_t^v\\}_{t=1}^T \\in \\mathbb{R}^{T \\times d}\\) (resp. audio segment-level features \\(\\tilde{F}^a = \\{f_t^a\\}_{t=1}^T\\) into visual segment-level logits \\(\\{z_t^v\\}_{t=1}^T \\in \\mathbb{R}^{T \\times C}\\) (resp. audio segment-level logits \\(\\{z_t^a\\}_{t=1}^T\\)). Segment-level probabilities \\(\\{p_t^v\\}_{t=1}^T, \\{p_t^a\\}_{t=1}^T \\in \\mathbb{R}^{T \\times C}\\) are then obtained by applying the sigmoid function on \\(\\{z_t^v\\}_{t=1}^T\\) and \\(\\{z_t^a\\}_{t=1}^T\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.708, + 0.892, + 0.766 + ], + "angle": 0, + "content": "Since, only video-level labels \\(y\\) are available during training, Tian et al. [35] introduce an attentive MMIL pooling module to learn to predict video-level probabilities \\(p \\in \\mathbb{R}^{C}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.772, + 0.891, + 0.79 + ], + "angle": 0, + "content": "\\[\nW _ {\\text {m o d a l}} ^ {v, a} = \\operatorname {S o f t m a x} _ {\\text {m o d a l}} \\left(\\operatorname {F C} _ {\\text {m o d a l}} \\left(\\tilde {F} ^ {v, a}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.802, + 0.891, + 0.82 + ], + "angle": 0, + "content": "\\[\nW _ {\\text {t i m e}} ^ {v, a} = \\operatorname {S o f t m a x} _ {\\text {t i m e}} \\left(\\mathrm {F C} _ {\\text {t i m e}} \\left(\\tilde {F} ^ {v, a}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathrm{FC}_{\\mathrm{mod}}\\) and \\(\\mathrm{FC}_{\\mathrm{time}}\\) are two learnable fully-connected layers, \\(\\tilde{F}^{v,a} = \\mathrm{Stack}(\\tilde{F}^v,\\tilde{F}^a)\\in \\mathbb{R}^{2\\times T\\times d}\\) denotes the stacked visual and audio features along the first dimension, \\(\\mathrm{Softmax}_{\\mathrm{mod}}(\\cdot)\\) denotes the softmax operation along the modality dimension (i.e., across \\(v,a\\)), while" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13563" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.082, + 0.096, + 0.27, + 0.12 + ], + "angle": 0, + "content": "Stage 1: Training the pseudo label models on larger dataset" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.122, + 0.359, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.098, + 0.728, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.734, + 0.097, + 0.887, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.287, + 0.893, + 0.358 + ], + "angle": 0, + "content": "Figure 2. UWAV framework: In stage 1, pseudo-label generation modules are equipped with the ability to capture temporal relationships between segments by pre-training on a large-scale, supervised, audio-visual event localization dataset. In stage 2, temporally coherent, uncertainty-weighted pseudo-labels, derived from the pre-trained pseudo-label generation module, are used to guide the learning of the inference model (HAN) aided by a class-balanced loss re-weighting and uncertainty-weighted feature mixup strategy. Note that we use the feature mixup strategy in both modalities while we only show the breakdown of the mixup operation for the audio modality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.47, + 0.416 + ], + "angle": 0, + "content": "\\(\\operatorname{Softmax}_{time}(\\cdot)\\) denotes the softmax operation along the temporal dimension (i.e., across \\(1, \\ldots, T\\)). Video-level probabilities \\(p \\in \\mathbb{R}^C\\) are then obtained via:" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.419, + 0.469, + 0.462 + ], + "angle": 0, + "content": "\\[\np = \\sum_ {m = \\{v, a \\}} \\sum_ {t = 1} ^ {T} W _ {\\text {m o d a l}} ^ {m, t} \\odot W _ {\\text {t i m e}} ^ {m, t} \\odot p _ {t} ^ {m}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.47, + 0.532 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes the element-wise product. The HAN model is then optimized with the binary cross-entropy (BCE) loss between the estimated video-level probabilities \\(p\\) and video-level labels \\(y\\): \\(\\mathcal{L}_{video} = \\mathrm{BCE}(p, y)\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.542, + 0.269, + 0.56 + ], + "angle": 0, + "content": "4. Proposed Approach" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.47, + 0.704 + ], + "angle": 0, + "content": "In this section, we detail our proposed approach (UWAV). At a high level, UWAV works by generating better segment-level pseudo-labels to improve the training of a multi-modal transformer-based inference module, e.g. HAN. Moreover, UWAV factors in the uncertainty associated with these pseudo-labels, addresses the imbalance in the training data, and introduces self-supervised regularization constraints, which all lead to better performance. Figure 2 shows an overview of our proposed framework." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.711, + 0.466, + 0.728 + ], + "angle": 0, + "content": "4.1. Temporally-Coherent Pseudo-Label Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.901 + ], + "angle": 0, + "content": "One major issue that plagues prior works, based on pseudolabel generation [9, 17, 50], is that they do not capture the temporal dependencies between neighboring segments when generating the pseudo-labels. That is, the generated pseudo-labels are not temporally coherent. To plug this void, we propose to incorporate transformer modules [38] into the pseudo-label generation pipeline, which maps CLIP/CLAP encodings of a segment's visual frame/audio to pseudo-labels. Specifically, two separate transformers are introduced, one each for the visual/audio pseudo-label synthesis modules." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.371, + 0.893, + 0.673 + ], + "angle": 0, + "content": "Pre-Training: Training transformers often requires sufficiently large training data while datasets commonly used for the weakly-supervised AVVP task are relatively small. To mitigate this challenge, we propose to first pre-train the transformer-equipped pseudo-label generation module on a large-scale, supervised, audio-visual event localization dataset - the UnAV [12] dataset. Specifically, given an audible video of duration \\( T' \\) seconds from the pre-training dataset, we split the video into \\( T' \\) one-second segments \\( \\{V_t', A_t'\\}_{t=1}^{T'} \\), with corresponding audio-visual event labels \\( y_t^{av'} \\in \\{0,1\\}^{C'} \\), where 1 indicates the presence of an event in both modalities and 0 its absence in at least one modality, while \\( C' \\) denotes the total number of event classes in the pre-training dataset. Next, the video frame at the temporal center of the visual segment is transformed into visual features \\( G_0^{v'} = \\{g_{0,t}'\\}_{t=1}^{T'} \\in \\mathbb{R}^{T' \\times d_1} \\) with CLIP's [28] image encoder. These features are then fed into the corresponding transformer of the visual stream, consisting of \\( L \\) encoder blocks, each block containing a self-attention layer, LayerNorm [1] (LN), and a 2-layer feed-forward network (FFN):" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.683, + 0.891, + 0.701 + ], + "angle": 0, + "content": "\\[\n\\tilde {G} _ {l} ^ {v ^ {\\prime}} = \\mathrm {L N} \\left(G _ {l} ^ {v ^ {\\prime}} + \\operatorname {A t t n} \\left(G _ {l} ^ {v ^ {\\prime}}, G _ {l} ^ {v ^ {\\prime}}, G _ {l} ^ {v ^ {\\prime}}\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.704, + 0.891, + 0.722 + ], + "angle": 0, + "content": "\\[\nG _ {l + 1} ^ {v ^ {\\prime}} = \\operatorname {L N} \\left(\\tilde {G} _ {l} ^ {v ^ {\\prime}} + \\operatorname {F F N} \\left(\\tilde {G} _ {l} ^ {v ^ {\\prime}}\\right)\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.736, + 0.893, + 0.876 + ], + "angle": 0, + "content": "Concurrently, we convert each event category label in the pre-training dataset into a textual event feature \\(e_c^{CLIP'} \\in \\mathbb{R}^{d_1}\\) by filling in the pre-defined caption template: \"A photo of \" with the corresponding event name and passing it to CLIP's text encoder. Equipped with the visual segment-level features \\(G_L^{v'} = \\{g_{L,t}'\\}_{t=1}^{T'} \\in \\mathbb{R}^{T' \\times d_1}\\) and the textual event features \\(E^{CLIP'} = \\{e_c^{CLIP'}\\}_{c=1}^{C'} \\in \\mathbb{R}^{C' \\times d_1}\\), we derive visual segment-level logits \\(\\hat{z}_t^{v'} \\in \\mathbb{R}^{C'}\\) and probabilities \\(\\hat{p}_t^{v'}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.883, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\hat {p} _ {t} ^ {v ^ {\\prime}} = \\operatorname {S i g m o i d} \\left(\\hat {z} _ {t} ^ {v ^ {\\prime}}\\right), \\hat {z} _ {t} ^ {v ^ {\\prime}} = E ^ {C L I P ^ {\\prime}} \\cdot g _ {L, t} ^ {v ^ {\\prime} \\top}. \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13564" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.26 + ], + "angle": 0, + "content": "Similar operations are performed in the audio pseudolabel generation pipeline. The raw waveforms corresponding to the 1-second audio segments are transformed into audio features \\( G_0^{a'} \\in \\mathbb{R}^{T' \\times d_2} \\) with CLAP's [12] audio encoder and fed into the corresponding transformer consisting of \\( L \\) encoder blocks. Correspondingly, the textual event features \\( E^{CLAP'} \\in \\mathbb{R}^{C' \\times d_2} \\) are generated with the caption template: \"This is the sound of \" by passing it through CLAP's text encoder. Audio segment-level logits \\( \\hat{z}_t^{a'} \\in \\mathbb{R}^{C'} \\) and probabilities \\( \\hat{p}_t^{a'} \\) can then be derived in the same manner: \\( \\hat{p}_t^{a'} = \\text{Sigmoid}(\\hat{z}_t^{a'}) \\), \\( \\hat{z}_t^{a'} = E^{CLAP'} \\cdot g_t^{a'}^\\top \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.26, + 0.47, + 0.349 + ], + "angle": 0, + "content": "Since the events occurring in the pre-training dataset (UnAV) are audio-visual, we multiply the segment-level visual and audio event probabilities to enforce the predicted labels to be multi-modal in nature: \\(\\{\\hat{p}_t^{av'}\\}_{t=1}^{T'} \\in \\mathbb{R}^{T' \\times C'}\\). This network is then trained with the binary cross-entropy (BCE) loss:" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.357, + 0.469, + 0.375 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t e m p}} = \\mathrm {B C E} \\left(\\hat {p} _ {t} ^ {a v ^ {\\prime}}, y _ {t} ^ {a v ^ {\\prime}}\\right), \\hat {p} _ {t} ^ {a v ^ {\\prime}} = \\tilde {p} _ {t} ^ {v ^ {\\prime}} \\odot \\hat {p} _ {t} ^ {a ^ {\\prime}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.383, + 0.47, + 0.627 + ], + "angle": 0, + "content": "Pseudo-Label Generation on Target Dataset: With the pre-trained pseudo-label generation modules in place, we proceed to employ them for the pseudo-label generation process in the target dataset for the AVVP task. Specifically, the center frame of each of the visual segments \\(\\{V_t\\}_{t=1}^T\\) of the target dataset are passed into CLIP's image encoder, whose output is passed into the pre-trained visual transformer to generate segment features \\(G_L^v = \\{g_{L,t}^v\\}_{t=1}^T \\in \\mathbb{R}^{T \\times d_1}\\). At the same time, the caption template: \"A photo of \" is used to obtain textual features corresponding to each of the event classes in the target dataset for the AVVP task: \\(E^{CLIP} \\in \\mathbb{R}^{C \\times d_1}\\). Segment-level visual logits \\(\\hat{z}_t^v \\in \\mathbb{R}^C\\) can be derived by computing their inner product. We also pre-define class-wise visual thresholds \\(\\theta^v \\in \\mathbb{R}^C\\) to transform segment-level visual logits into binary pseudo-labels \\(\\hat{y}_t^v \\in \\mathbb{R}^C\\):" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.633, + 0.469, + 0.653 + ], + "angle": 0, + "content": "\\[\n\\hat {y} _ {t} ^ {v} = \\mathbb {1} _ {\\left\\{\\hat {z} _ {t} ^ {v} > \\theta^ {v} \\right\\}} \\odot y, \\hat {z} _ {t} ^ {v} = E ^ {C L I P} \\cdot g _ {L, t} ^ {v} ^ {\\top}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.66, + 0.469, + 0.735 + ], + "angle": 0, + "content": "where \\(y\\) denotes the ground-truth video-level labels, \\(\\mathbb{1}_{\\{\\cdot\\}}\\) is the indicator function which returns a value of 1 when the condition is true otherwise 0, and \\(\\odot\\) denotes the elementwise product operation. The \\(\\odot\\) operation zeroes out the predictions of event classes absent in the video-level label." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.736, + 0.469, + 0.87 + ], + "angle": 0, + "content": "A similar pseudo-label generation process is employed on the acoustic side. Raw waveforms of audio segments are first fed into CLAP's audio encoder and then into the pretrained audio transformer. The event names of the classes in the target dataset for the AVVP task are filled in the caption template: \"This is the sound of \" to generate textual event features: \\( E^{CLAP} \\in \\mathbb{R}^{C \\times d_1} \\). Segment-level audio logits \\( \\hat{z}_t^a \\in \\mathbb{R}^C \\) and binary pseudo-labels \\( \\hat{y}_t^a \\in \\mathbb{R}^C \\) are then derived using class-wise thresholds \\( \\theta^a \\in \\mathbb{R}^C \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "With binary segment-level pseudo-labels for both modalities \\(\\hat{y}_t^v, \\hat{y}_t^a\\) and the predicted probabilities from the infer" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "ence module (HAN) \\(\\hat{p}_t^v,\\hat{p}_t^a\\) in place, the inference module can be trained using the binary cross-entropy loss as shown:" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.133, + 0.892, + 0.15 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {h a r d} = \\mathrm {B C E} \\left(p _ {t} ^ {v}, \\hat {y} _ {t} ^ {v}\\right) + \\mathrm {B C E} \\left(p _ {t} ^ {a}, \\hat {y} _ {t} ^ {a}\\right). \\tag {11}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.159, + 0.847, + 0.176 + ], + "angle": 0, + "content": "4.2. Training with Pseudo-Label Uncertainty" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.892, + 0.379 + ], + "angle": 0, + "content": "While pseudo-labels do provide additional supervision for better training of the inference module, they could potentially be noisy, leading to occasionally incorrect training signals. To ameliorate this problem, we propose an uncertainty-weighted pseudo-label training scheme to improve the robustness of the learning process. Instead of simply training with the binary pseudo-labels \\(\\hat{y}_t^v,\\hat{y}_t^a\\) , we leverage the confidence of the pseudo-label estimation module (associated with the predicted pseudo-label) to weigh the training signal for the inference module. This confidence score serves as a measure of the pseudo-label generation module's uncertainty of its prediction. This may be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.387, + 0.892, + 0.417 + ], + "angle": 0, + "content": "\\[\n\\hat {p} _ {t} ^ {v} = \\operatorname {S i g m o i d} \\left(\\hat {z} _ {t} ^ {v} - \\theta^ {v}\\right) \\odot y; \\quad \\hat {p} _ {t} ^ {a} = \\operatorname {S i g m o i d} \\left(\\hat {z} _ {t} ^ {a} - \\theta^ {a}\\right) \\odot y. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.892, + 0.598 + ], + "angle": 0, + "content": "In other words, considering the visual pseudo-label generation pipeline as an example, the farther the logit \\(\\hat{z}_t^v\\) is from the threshold \\(\\theta^v\\), whether much lower or higher, the more confident the pseudo-label generation module is about the label it predicts (either approaching 0 or 1). Conversely, the closer the logit is to the threshold, the less the certainty about the correctness of the pseudo-labels (probabilities closer to 0.5). An analogous explanation also holds for the audio pseudo-labels. With the uncertainty-weighted pseudo-labels in place, the inference module (HAN) can be trained with the following uncertainty-weighted pseudo-label loss:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.606, + 0.892, + 0.623 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s o f t}} = \\mathrm {B C E} \\left(p _ {t} ^ {v}, \\hat {p} _ {t} ^ {v}\\right) + \\mathrm {B C E} \\left(p _ {t} ^ {a}, \\hat {p} _ {t} ^ {a}\\right). \\tag {13}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.636, + 0.822, + 0.653 + ], + "angle": 0, + "content": "4.3. Uncertainty-weighted Feature Mixup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Due to the lack of full supervision for the weakly-supervised AVVP task, we explore the efficacy of additional regularization via self-supervision to help the models generalize better. Towards this end, prior pseudo-label generation approaches [26, 39] often employ contrastive learning as a tool to better train the inference module. However, due to the inherent noise in the estimated pseudo-labels, positive samples and negative samples may be mislabeled, decreasing the effectiveness of the self-supervisory training. As an alternative, in this work, we explore the effectiveness of feature mixing, as a self-supervisory training signal for additional regularization. In this setting, we mixup the estimated features of any two segments, additively, and train the model to predict the union of the labels of the two segments. However, since the labels in our setting are noisy, the mixed feature is assigned a label derived from a weighted sum of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13565" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "the uncertainty-weighted pseudo-labels of each of the two segment features. This is illustrated below:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.129, + 0.468, + 0.15 + ], + "angle": 0, + "content": "\\[\n\\bar {f} _ {t _ {i}, t _ {j}} ^ {v} = \\lambda \\tilde {f} _ {t _ {i}} ^ {v} + (1 - \\lambda) \\tilde {f} _ {t _ {j}} ^ {v}, \\quad \\bar {p} _ {t _ {i}, t _ {j}} ^ {v} = \\lambda \\hat {p} _ {t _ {i}} ^ {v} + (1 - \\lambda) \\hat {p} _ {t _ {j}} ^ {v} \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.089, + 0.152, + 0.468, + 0.173 + ], + "angle": 0, + "content": "\\[\n\\bar {f} _ {t _ {i}, t _ {j}} ^ {a} = \\lambda \\tilde {f} _ {t _ {i}} ^ {a} + (1 - \\lambda) \\tilde {f} _ {t _ {j}} ^ {a}, \\quad \\bar {p} _ {t _ {i}, t _ {j}} ^ {a} = \\lambda \\hat {p} _ {t _ {i}} ^ {a} + (1 - \\lambda) \\hat {p} _ {t _ {j}} ^ {a}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.179, + 0.469, + 0.224 + ], + "angle": 0, + "content": "where \\(\\lambda \\sim \\mathrm{Beta}(\\alpha, \\alpha)\\) and \\(\\alpha\\) is a hyper-parameter controlling the Beta distribution, and \\(t_i\\) and \\(t_j\\) indicate two segment indices in a batch of video segments." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.224, + 0.469, + 0.315 + ], + "angle": 0, + "content": "After mixing the uni-modal segment-level features, we pass them through the classifier of the inference module and apply the sigmoid function to the output, obtaining mixed segment-level event probabilities \\( p_t^{mix - v} \\) and \\( p_t^{mix - a} \\). These are used to train the inference model with the uncertainty-aware mixup loss, as shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.323, + 0.469, + 0.341 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m i x} = \\mathrm {B C E} \\left(p _ {t} ^ {m i x - v}, \\bar {p} _ {t} ^ {v}\\right) + \\mathrm {B C E} \\left(p _ {t} ^ {m i x - a}, \\bar {p} _ {t} ^ {a}\\right). \\tag {16}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.349, + 0.376, + 0.365 + ], + "angle": 0, + "content": "4.4. Class-balanced Loss Re-weighting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.47, + 0.644 + ], + "angle": 0, + "content": "Besides the aforementioned challenges of the AVVP task, most of the events in the event set are absent in the pseudolabels of any (segment of a) video (i.e., most event classes are negative events) and only a few events are present (i.e., positive events are much fewer in number). As a result, the model is dominated by the loss from the negative events. When trained without factoring in this bias, the classifier tends to overfit the negative labels and ignore the positive ones. To address this class imbalance issue, we introduce a class-balanced loss re-weighting strategy to re-balance the importance of the losses from the negative and positive events for the uncertainty-weighted pseudo-label loss. Specifically, the loss from the positive events is multiplied by a weight proportional to the frequency of the segments with the negative events in the pseudo-labels, while the loss from the negative events is multiplied by a weight proportional to the frequency of the segments with the positive events in the pseudo-labels, as shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.094, + 0.649, + 0.469, + 0.697 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {w - s o f t} = \\sum_ {m \\in \\{v, a \\}} w _ {p o s} ^ {m} \\cdot y \\cdot \\mathrm {B C E} \\left(p _ {t} ^ {m}, \\hat {p} _ {t} ^ {m}\\right) + \\\\ w _ {n e g} ^ {m} \\cdot (1 - y) \\cdot \\mathrm {B C E} \\left(p _ {t} ^ {m}, \\hat {p} _ {t} ^ {m}\\right), \\tag {17} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.706, + 0.469, + 0.741 + ], + "angle": 0, + "content": "\\[\nw _ {p o s} ^ {m} = \\frac {\\sum_ {i = 1} ^ {N} \\sum_ {t = 1} ^ {T} \\sum_ {c = 1} ^ {C} \\left(1 - \\hat {y} _ {i , t , c} ^ {m}\\right)}{N T C} \\times W, \\tag {18}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.742, + 0.469, + 0.777 + ], + "angle": 0, + "content": "\\[\nw _ {n e g} ^ {m} = \\frac {\\sum_ {i = 1} ^ {N} \\sum_ {t = 1} ^ {T} \\sum_ {c = 1} ^ {C} \\hat {y} _ {i , t , c} ^ {m}}{N T C}, \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.784, + 0.468, + 0.814 + ], + "angle": 0, + "content": "where \\(N\\) denotes the number of videos in the training set, and \\(W\\) is a hyper-parameter." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.814, + 0.469, + 0.887 + ], + "angle": 0, + "content": "In summary, the inference module is trained on the AVVP task with the proposed class-balanced re-weighting, applied to the uncertainty-weighted classification loss, and the uncertainty-weighted feature mixup loss, as shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.887, + 0.469, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {w - s o f t} + \\mathcal {L} _ {\\text {m i x}} + \\mathcal {L} _ {\\text {v i d e o}}. \\tag {20}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.633, + 0.108 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.252 + ], + "angle": 0, + "content": "We assess the performance of UWAV empirically across two challenging datasets and report its performance, comparing it with existing state-of-the-art approaches both quantitatively and qualitatively. Additionally, through multiple ablation studies, we bring out the effectiveness of the different elements of our proposed approach and the choices of different hyper-parameters. For additional details, ablation studies, and more qualitative results, we refer the reader to our supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.261, + 0.691, + 0.278 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.285, + 0.892, + 0.693 + ], + "angle": 0, + "content": "Datasets: We evaluate all competing methods on the Look, Listen, and Parse (LLP) dataset [35], which is the principal benchmark dataset for the AVVP task. The dataset consists of 11,849 video clips sourced from YouTube. Each clip is 10 seconds long and represents one or more of 25 diverse event categories, such as human activities, animals, musical instruments, and vehicles. The dataset is split into training, validation, and testing sets, following the official split [35]: 10,000 videos for training, 649 videos for validation, and 1,200 videos for testing. While the training set of this dataset is only associated with video-level labels of the events, the validation and testing split is labeled with segment-level event labels for evaluation purposes. Additionally, to demonstrate the generalizability of our method, we conduct a similar study on the Audio Visual Event (AVE) recognition dataset [34]. The AVE dataset consists of 4,143 video clips crawled from YouTube, each 10 seconds long. It is split into 3,339 videos for training, 402 for validation, and 402 for testing. It includes 29 event categories encompassing human activities, animals, musical instruments, vehicles, and a \"background\" class (i.e., no event occurs). Unlike the LLP dataset, each video in the AVE dataset contains only one audio-visual event. Here too, the training data is only provided with video-level labels while the validation and test splits are annotated with ground-truth event labels for each one-second segment, which either is a specific audio-visual event or \"background\"." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Metrics: For the LLP dataset, following the standard evaluation protocol [35], all models are evaluated using macro F1-scores calculated for the following event types: (i) audio-only (A), (ii) visual-only (V), and (iii) audio-visual (AV). Type@AV (Type) and Event@AV (Event) are two additional metrics that evaluate the overall performance of the model, where Type@AV is the mean of the F1-scores for the A, V, and AV events, while Event@AV is the F1-score of all events regardless of the modality in which they occur. Evaluations are conducted at both the segment-level and the event-level. At the segment-level, the model's predictions are compared with the ground truth on a per-segment basis. At the event-level, consecutive positive segments for" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13566" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.086, + 0.089, + 0.882, + 0.104 + ], + "angle": 0, + "content": "Table 1. Comparison with state-of-the-arts methods on the LLP dataset. Best performances are in bold, second-best in underlined." + }, + { + "type": "table", + "bbox": [ + 0.162, + 0.108, + 0.805, + 0.345 + ], + "angle": 0, + "content": "
MethodSegment-levelEvent-level
AVAVTypeEventAVAVTypeEvent
HAN [35]60.152.948.954.055.451.348.943.047.748.0
MA [39]60.360.055.158.957.953.656.449.053.050.6
JoMoLD [8]61.363.857.260.859.953.959.949.654.552.5
CMPAE [10]64.266.459.263.362.856.663.751.857.455.7
PoiBin [27]63.163.557.761.460.654.160.351.555.252.3
VPLAN [50]60.564.858.361.259.451.461.551.254.750.8
VALOR [17]61.865.958.462.061.555.462.652.256.754.2
LSLD [9]62.767.159.463.162.255.764.352.657.655.2
PPL [26]65.966.761.964.863.757.364.354.359.957.9
CoLeaf [31]64.264.459.362.662.557.663.254.257.955.6
LEAP [51]62.765.659.362.561.856.463.154.157.855.0
UWAV (Ours)64.270.063.465.963.958.666.757.560.957.4
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.47, + 0.449 + ], + "angle": 0, + "content": "the same event are grouped together as a single event. The F1-score is then computed using a mIoU threshold of 0.5. For the AVE dataset, we follow Tian et al. [34] and use accuracy as the evaluation metric. An event prediction of a segment is considered correct if it matches the ground-truth label for that segment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.455, + 0.47, + 0.744 + ], + "angle": 0, + "content": "Implementation Details: In line with prior work [35], each 10-second video in both datasets is split into 10 segments of one second each, where each segment contains 8 frames. For the LLP dataset, pre-trained ResNet-152 [13] and \\(\\mathrm{R}(2 + 1)\\mathrm{D - }18\\) [37] are used to extract 2D and 3D visual features, respectively. The pre-trained VGGish [14] network is used to extract features from the audio, sampled at \\(16\\mathrm{KHz}\\). For the AVE dataset however, akin to prior work [17], we extract visual features from pre-trained CLIP and \\(\\mathrm{R}(2 + 1)\\mathrm{D}\\) while CLAP is used to embed the audio stream. For both datasets, we set the number of encoder blocks \\(L\\) in the temporal-aware model to 5, \\(\\alpha\\) for the Beta distribution in the feature mixup strategy to 1.7, and \\(W\\) in the class-balanced loss re-weighting step to 0.5. The pseudo-label generation modules and the inference model (HAN) are trained with the AdamW optimizer [20]. For improved performance on the AVE dataset, we replace \\(\\hat{p}_{t_i}^a,\\hat{p}_{t_j}^a\\) in Eq. 15 with \\([\\hat{p}_{t_i}^a ],[\\hat{y}_{t_j}^a ]\\) and make corresponding modifications in the visual counterparts as well." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Baselines: We demonstrate the effectiveness of UWAV by comparing against an extensive set of baselines. For the LLP dataset, this includes video-level pseudo-label generation methods (MA [39], JoMoLD [8]), segment-level pseudo-label generation methods (VALOR [17], LSLD [9], PPL [26]), and the recently released works (CoLeaf [31], LEAP [51]). On the other hand, for the AVE dataset, baseline approaches with publicly available implementation, which use the state-of-the-art feature backbones (akin to ours), such as HAN [35] and VALOR [17] are used." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.357, + 0.593, + 0.372 + ], + "angle": 0, + "content": "5.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.892, + 0.623 + ], + "angle": 0, + "content": "Comparison with Previous Methods on LLP: As shown in Table 1, UWAV surpasses previous methods, across almost all metrics. Notably, we achieve an F-score of 70.0 on the segment-level visual event, 65.9 on the segment-level Type@AV, and 66.7 on the event-level visual event metric. This corresponds to a gain of \\(1.1\\%\\) on segment-level Type@AV F-score and a \\(1\\%\\) improvement on event-level Type@AV F-score, over our closest competitor PPL [26]. Of particular note, is the fact that our segment-level and event-level F-scores improve by more than \\(3\\%\\), over PPL, for visual events. When compared to other recently published works, such as VALOR [17], CoLeaf [31] and LEAP [51], UWAV outperforms them by up to \\(3\\%\\) on both segment and event-level Type@AV F-score while gains on visual events are up to \\(5\\%\\) on segment-level F-score." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.892, + 0.837 + ], + "angle": 0, + "content": "These observations are consistent with our qualitative comparisons, as well. In the example on the left of Figure 3, our model successfully recognizes and temporally localizes the lawn mower event visually, whereas VALOR [17] (a recent state-of-the-art approach with publicly available implementation) misclassifies it as a chainsaw. Additionally, our model also accurately localize the intermittent sound of the lawn mower. In contrast, VALOR not only misclassifies the sound of the lawn mower as that of a chainsaw but also incorrectly predicts that someone is talking in the video. In the example on the right of Figure 3, our model does not err in recognizing either the visual presence or the audio presence of the telephone, while VALOR fails to accurately predict the events in either modality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Comparison with Previous Methods on AVE: To demonstrate the generalizability of our method, we evaluate UWAV on the AVE [34] dataset and compare its per" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13567" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.09, + 0.506, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.089, + 0.835, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.302, + 0.872, + 0.316 + ], + "angle": 0, + "content": "Figure 3. Comparison between predictions by UWAV and competing AVVP methods on the LLP dataset. \"GT\": ground truth." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.328, + 0.47, + 0.356 + ], + "angle": 0, + "content": "Table 2. Model performances on the AVE dataset. CLIP, R(2+1)D-18, and CLAP are used as feature backbones." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.36, + 0.444, + 0.398 + ], + "angle": 0, + "content": "
MethodHAN [35]VALOR [17]UWAV (Ours)
Acc.(%)75.380.480.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.406, + 0.456, + 0.42 + ], + "angle": 0, + "content": "Table 3. Accuracy of the generated pseudo-labels on LLP." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.424, + 0.446, + 0.496 + ], + "angle": 0, + "content": "
MethodSegment-level
AVAVTypeEvent
VALOR [17]80.561.755.766.074.6
PPL [26]61.761.857.560.659.4
UWAV (Ours)78.474.565.572.878.4
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.516, + 0.47, + 0.576 + ], + "angle": 0, + "content": "formance with that of previous works. From Table 2, we observe that with the same backbone features, UWAV surpasses VALOR, our closest competitor, even on this small-scale dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.47, + 0.76 + ], + "angle": 0, + "content": "Accuracy of the Generated Pseudo-Labels: To evaluate the efficacy of our pseudo-label generation pipeline, we compare the accuracy of our generated pseudo-labels against those obtained from competing methods (with publicly available implementation) [17, 26] on the test set of the LLP dataset. As shown in Table 3, our pre-trained temporally-dependent pseudo-label generation scheme generates more accurate segment-level pseudo-labels than previous methods, by up to \\(6\\%\\) on the segment-level Type@AV F-score, attesting to the advantages of factoring in intersegment temporal dependencies." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.231, + 0.787 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To demonstrate the potency of the different elements of our proposed method, UWAV, we conduct ablation studies. In particular, the proposed uncertainty-weighted pseudo-label based training, the uncertainty-weighted feature mixup scheme, and the class-balanced loss re-weighting schemes are ablated. As shown in Table 4, incorporating the uncertainty-weighted pseudo-label training step improves" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.328, + 0.892, + 0.37 + ], + "angle": 0, + "content": "Table 4. Ablation study of the proposed components in UWAV. \"Binary\" denotes training with binary pseudo-labels. \"Soft\" denotes training with uncertainty-weighted pseudo-labels." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.376, + 0.89, + 0.48 + ], + "angle": 0, + "content": "
BinarySoftRe-weightMixupSegment-level
AVAVTypeEvent
✓62.767.761.964.262.2
✓63.068.361.864.462.8
✓✓63.669.563.065.463.1
✓✓63.969.062.865.263.4
✓✓✓64.270.063.465.963.9
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.496, + 0.893, + 0.662 + ], + "angle": 0, + "content": "the segment-level Type@AV F-score by \\(2\\%\\), compared to using binary pseudo-labels. This demonstrates the benefit of accounting for the uncertainty in the pseudo-label estimation module. Moreover, sans the class-balanced loss weighting strategy, the model's performance is worse off by \\(1\\%\\) on the Type@AV F-score, revealing the erroneous bias in the model's prediction arising from a skew of the class distribution. On the other hand, introducing the uncertainty-weighted feature mixup results in a gain of \\(0.8\\%\\) on the Type@AV F-score, which underscores the importance of this self-supervised regularization." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.679, + 0.619, + 0.694 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.893, + 0.9 + ], + "angle": 0, + "content": "In this work, we address the challenging task of weakly-supervised audio-visual video parsing (AVVP), which presents a two-fold challenge: (i) potential misalignment between the events of the audio and visual streams, and (ii) the lack of fine-grained labels for each modality. We observe that by considering the temporal relationship between segments, our proposed method (UWAV) is able to provide more reliable pseudo-labels for better training of the inference module. In addition, by factoring in the uncertainty associated with these estimated pseudo-labels, regularizing the training process with a feature mixup strategy, and correcting for class imbalance, UWAV achieves state-of-the-art results on the LLP and AVE datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13568" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.47, + 0.202 + ], + "angle": 0, + "content": "[2] Moitreya Chatterjee, Jonathan Le Roux, Narendra Ahuja, and Anoop Cherian. Visual scene graphs for audio source separation. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.204, + 0.47, + 0.247 + ], + "angle": 0, + "content": "[3] Moitreya Chatterjee, Narendra Ahuja, and Anoop Cherian. Learning audio-visual dynamics using scene graphs for audio source separation. In NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.249, + 0.47, + 0.316 + ], + "angle": 0, + "content": "[4] Changan Chen, Unnat Jain, Carl Schissler, Sebastia Vicenc Amengual Gari, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audio-visual navigation in 3d environments. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.32, + 0.47, + 0.348 + ], + "angle": 0, + "content": "[5] Changan Chen, Ziad Al-Halah, and Kristen Grauman. Semantic audio-visual navigation. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.351, + 0.47, + 0.405 + ], + "angle": 0, + "content": "[6] Changan Chen, Sagnik Majumder, Ziad Al-Halah, Ruohan Gao, Santhosh Kumar Ramakrishnan, and Kristen Grauman. Learning to set waypoints for audio-visual navigation. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.409, + 0.47, + 0.45 + ], + "angle": 0, + "content": "[7] Jiaben Chen, Renrui Zhang, Dongze Lian, Jiaqi Yang, Ziyao Zeng, and Jianbo Shi. iquery: Instruments as queries for audio-visual sound separation. In CVPR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.47, + 0.508 + ], + "angle": 0, + "content": "[8] Haoyue Cheng, Zhaoyang Liu, Hang Zhou, Chen Qian, Wayne Wu, and Limin Wang. Joint-modal label denoising for weakly-supervised audio-visual video parsing. In ECCV, 2022. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.511, + 0.47, + 0.552 + ], + "angle": 0, + "content": "[9] Yingying Fan, Yu Wu, Yutian Lin, and Bo Du. Revisit weakly-supervised audio-visual video parsing from the language perspective. In NeurIPS, 2023. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.556, + 0.469, + 0.61 + ], + "angle": 0, + "content": "[10] Junyu Gao, Mengyuan Chen, and Changsheng Xu. Collecting cross-modal presence-absence evidence for weakly-supervised audio-visual event perception. In CVPR, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.614, + 0.469, + 0.681 + ], + "angle": 0, + "content": "[11] Shijie Geng, Peng Gao, Moitreya Chatterjee, Chiori Hori, Jonathan Le Roux, Yongfeng Zhang, Hongsheng Li, and Anoop Cherian. Dynamic graph representation learning for video dialog via multi-modal shuffled transformers. In AAAI, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.469, + 0.74 + ], + "angle": 0, + "content": "[12] Tiantian Geng, Teng Wang, Jinming Duan, Runmin Cong, and Feng Zheng. Dense-localizing audio-visual events in untrimmed videos: A large-scale benchmark and baseline. In CVPR, 2023. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.469, + 0.784 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[14] Shawn Hershey, Sourish Chaudhuri, Daniel PW Ellis, Jort F Gemmeke, Aren Jansen, R Channing Moore, Manoj Plakal, Devin Platt, Rif A Saurous, Bryan Seybold, et al. Cnn architectures for large-scale audio classification. In ICASSP, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[15] Xixi Hu, Ziyang Chen, and Andrew Owens. Mix and localize: Localizing sound sources in mixtures. In CVPR, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[16] Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Egocentric audio-visual object localization. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[17] Yung-Hsuan Lai, Yen-Chun Chen, and Frank Wang. Modality-independent teachers meet weakly-supervised audio-visual event parser. In NeurIPS, 2023. 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[18] Guangyao Li, Yake Wei, Yapeng Tian, Chenliang Xu, JiRong Wen, and Di Hu. Learning to answer questions in dynamic audio-visual scenarios. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[19] Jinxiang Liu, Yikun Liu, Fei Zhang, Chen Ju, Ya Zhang, and Yanfeng Wang. Audio-visual segmentation via unlabeled frame exploitation. In CVPR, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[20] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.293, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[21] Sagnik Majumder, Ziad Al-Halah, and Kristen Grauman. Move2hear: Active audio-visual source separation. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.335, + 0.892, + 0.377 + ], + "angle": 0, + "content": "[22] Yuxin Mao, Jing Zhang, Mochu Xiang, Yiran Zhong, and Yuchao Dai. Multimodal variational auto-encoder based audio-visual segmentation. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.379, + 0.892, + 0.419 + ], + "angle": 0, + "content": "[23] Shentong Mo and Yapeng Tian. Multi-modal grouping network for weakly-supervised audio-visual video parsing. In NeurIPS, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.421, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[24] Shentong Mo and Yapeng Tian. Audio-visual grouping network for sound localization from mixtures. In CVPR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.464, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[25] Kranti Kumar Parida, Siddharth Srivastava, and Gaurav Sharma. Beyond mono to binaural: Generating binaural audio from mono audio with depth and cross modal attention. In WACV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.521, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[26] Kranthi Kumar Rachavarapu, Kalyan Ramakrishnan, et al. Weakly-supervised audio-visual video parsing with prototype-based pseudo-labeling. In CVPR, 2024. 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.577, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[27] Kranthi Kumar Rachavarapu et al. Boosting positive segments for weakly-supervised audio-visual video parsing. In ICCV, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.62, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.69, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[29] Varshanth Rao, Md Ibrahim Khalil, Haoda Li, Peng Dai, and Juwei Lu. Dual perspective network for audio-visual event localization. In ECCV, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.733, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[30] Ludan Ruan, Yiyang Ma, Huan Yang, Huiguo He, Bei Liu, Jianlong Fu, Nicholas Jing Yuan, Qin Jin, and Baining Guo. Mm-diffusion: Learning multi-modal diffusion models for joint audio and video generation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[31] Faegheh Sardari, Armin Mustafa, Philip JB Jackson, and Adrian Hilton. Coleaf: A contrastive-collaborative learning framework for weakly supervised audio-visual video parsing. In ECCV, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[32] Arda Senocak, Hyeonggon Ryu, Junsik Kim, Tae-Hyun Oh, Hanspeter Pfister, and Joon Son Chung. Sound source localization is all about cross-modal alignment. In ICCV, 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13569" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[33] Ankit Shah, Shijie Geng, Peng Gao, Anoop Cherian, Takaaki Hori, Tim K Marks, Jonathan Le Roux, and Chiori Hori. Audio-visual scene-aware dialog and reasoning using audio-visual transformers with joint student-teacher learning. In ICASSP, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[34] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-visual event localization in unconstrained videos. In ECCV, 2018. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.206, + 0.471, + 0.248 + ], + "angle": 0, + "content": "[35] Yapeng Tian, Dingzeyu Li, and Chenliang Xu. Unified multisensory perception: Weakly-supervised audio-visual video parsing. In ECCV, 2020. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.47, + 0.29 + ], + "angle": 0, + "content": "[36] Yapeng Tian, Di Hu, and Chenliang Xu. Cyclic co-learning of sounding object visual grounding and sound separation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.47, + 0.333 + ], + "angle": 0, + "content": "[37] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In CVPR, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.47, + 0.389 + ], + "angle": 0, + "content": "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.47, + 0.432 + ], + "angle": 0, + "content": "[39] Yu Wu and Yi Yang. Exploring heterogeneous clues for weakly-supervised audio-visual video parsing. In CVPR, 2021. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.47, + 0.489 + ], + "angle": 0, + "content": "[40] Yusong Wu, Ke Chen, Tianyu Zhang, Yuchen Hui, Taylor Berg-Kirkpatrick, and Shlomo Dubnov. Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation. In ICASSP, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.47, + 0.544 + ], + "angle": 0, + "content": "[41] Yazhou Xing, Yingqing He, Zeyue Tian, Xintao Wang, and Qifeng Chen. Seeing and hearing: Open-domain visual-audio generation with diffusion latent aligners. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.547, + 0.47, + 0.602 + ], + "angle": 0, + "content": "[42] Qi Yang, Xing Nie, Tong Li, Pengfei Gao, Ying Guo, Cheng Zhen, Pengfei Yan, and Shiming Xiang. Cooperation does matter: Exploring multi-order bilateral relations for audio-visual segmentation. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.47, + 0.644 + ], + "angle": 0, + "content": "[43] Yuxin Ye, Wenming Yang, and Yapeng Tian. Lavss: Location-guided audio-visual spatial audio separation. In WACV, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.47, + 0.715 + ], + "angle": 0, + "content": "[44] Abdelrahman Younes, Daniel Honerkamp, Tim Welschehold, and Abhinav Valada. Catch me if you hear me: Audio-visual navigation in complex unmapped environments with moving sounds. IEEE Robotics and Automation Letters, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.47, + 0.772 + ], + "angle": 0, + "content": "[45] Jiashuo Yu, Ying Cheng, Rui-Wei Zhao, Rui Feng, and Yuejie Zhang. Mm-pyramid: Multimodal pyramid attentional network for audio-visual event localization and video parsing. In ACM MM, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[46] Yinfeng Yu, Wenbing Huang, Fuchun Sun, Changan Chen, Yikai Wang, and Xiaohong Liu. Sound adversarial audio-visual navigation. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[47] Heeseung Yun, Youngjae Yu, Wonsuk Yang, Kangil Lee, and Gunhee Kim. Pano-avqa: Grounded audio-visual question answering on 360deg videos. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[48] Jinxing Zhou, Liang Zheng, Yiran Zhong, Shijie Hao, and Meng Wang. Positive sample propagation along the audio-visual event line. In CVPR, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.147 + ], + "angle": 0, + "content": "[49] Jinxing Zhou, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audio-visual segmentation. In ECCV, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.893, + 0.19 + ], + "angle": 0, + "content": "[50] Jinxing Zhou, Dan Guo, Yiran Zhong, and Meng Wang. Improving audio-visual video parsing with pseudo visual labels. arXiv preprint arXiv:2303.02344, 2023. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.893, + 0.247 + ], + "angle": 0, + "content": "[51] Jinxing Zhou, Dan Guo, Yuxin Mao, Yiran Zhong, Xiaojun Chang, and Meng Wang. Label-anticipated event disentanglement for audio-visual video parsing. arXiv preprint arXiv:2407.08126, 2024. 2, 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13570" + } + ] +] \ No newline at end of file diff --git a/2025/UWAV_ Uncertainty-weighted Weakly-supervised Audio-Visual Video Parsing/82f7769f-b767-4619-96ea-539823053afb_origin.pdf b/2025/UWAV_ Uncertainty-weighted Weakly-supervised Audio-Visual Video Parsing/82f7769f-b767-4619-96ea-539823053afb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6c98005ded35efae396c87b202ae8707900aa388 --- /dev/null +++ b/2025/UWAV_ Uncertainty-weighted Weakly-supervised Audio-Visual Video Parsing/82f7769f-b767-4619-96ea-539823053afb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a3c674230beedb945dd2ca83de3194bc8f8c6fefeb56ec4aeca9f33c372ad8c1 +size 1024173 diff --git a/2025/UWAV_ Uncertainty-weighted Weakly-supervised Audio-Visual Video Parsing/full.md b/2025/UWAV_ Uncertainty-weighted Weakly-supervised Audio-Visual Video Parsing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ce4af8cab305fee02106eb76a321f8f3287bd442 --- /dev/null +++ b/2025/UWAV_ Uncertainty-weighted Weakly-supervised Audio-Visual Video Parsing/full.md @@ -0,0 +1,315 @@ +# UWAV: Uncertainty-weighted Weakly-supervised Audio-Visual Video Parsing + +Yung-Hsuan Lai $^{1,\dagger}$ , Janek Ebbers $^{3}$ , Yu-Chiang Frank Wang $^{1,2}$ , François Germain $^{3}$ , Michael Jeffrey Jones $^{3}$ , Moitreya Chatterjee $^{3,\ddagger}$ , + $^{1}$ Graduate Institute of Communication Engineering, National Taiwan University $^{2}$ NVIDIA, Taiwan + $^{3}$ Mitsubishi Electric Research Labs (MERL) + $\dagger$ r10942097@ntu.edu.tw $\ddagger$ metro.smiles@gmail.com + +# Abstract + +Audio-Visual Video Parsing (AVVP) entails the challenging task of localizing both uni-modal events (i.e., those occurring exclusively in either the visual or acoustic modality of a video) and multi-modal events (i.e., those occurring in both modalities concurrently). Moreover, the prohibitive cost of annotating training data with the class labels of all these events, along with their start and end times, imposes constraints on the scalability of AVVP techniques unless they can be trained in a weakly-supervised setting, where only modality-agnostic, video-level labels are available in the training data. To this end, recently proposed approaches seek to generate segment-level pseudolabels to better guide model training. However, the absence of inter-segment dependencies when generating these pseudo-labels and the general bias towards predicting labels that are absent in a segment limit their performance. This work proposes a novel approach towards overcoming these weaknesses called Uncertainty-weighted Weakly-supervised Audio-visual Video Parsing (UWAV). Additionally, our innovative approach factors in the uncertainty associated with these estimated pseudo-labels and incorporates a feature mixup based training regularization for improved training. Empirical results show that UWAV outperforms state-of-the-art methods for the AVVP task on multiple metrics, across two different datasets, attesting to its effectiveness and generalizability. + +# 1. Introduction + +Events that occur in the real world, often leave their imprint on the acoustic and visual modalities. Humans rely heavily on the synergy between their senses of sight and hearing to interpret such events. Audio-visual learning, which seeks to equip machines with a similar synergy, has emerged as one of the most important research areas within the multi + +![](images/843db44b81fc66a66494d92c620dd3afa3214df81fbb5d2257344d9cd27e586c.jpg) +Figure 1. A weakly-supervised AVVP task example. Events, considered in this task, might be unimodal or multimodal. Even multimodal events, may not be temporally aligned in the audio and visual modalities, e.g. the cello might only be visible in the first few seconds but might produce music, throughout the video. + +modal machine learning community. It aims to leverage both these senses (modalities) jointly, to enhance machine perception and understanding of real-world events. Various audio-visual learning tasks have been studied towards this end, including audio-visual segmentation [19, 49], sound source localization [24, 32], audio-visual event localization [29, 34], and audio-visual sound separation [3, 7, 43]. However, many of these tasks assume that audio and visual streams would always be temporally aligned. This assumption often fails in real-world scenarios, where the sonic and visual imprints of events may not perfectly overlap. For instance, one might hear an emergency siren approaching from a distance before it appears in the field of view. + +In this work, in order to better understand the events occurring in a video, we explore the task of Audio-Visual Video Parsing (AVVP) [35]. Its goal is to recognize and localize all audio, visual, and audio-visual events occurring in the video. See Figure 1 for an example of this task. The task setup is to perform this prediction for every one-second segment of a video. This task poses two principal challenges, + +from a machine learning standpoint: (i) The audio and visual events, that occur, might not be temporally aligned, e.g. if an event becomes audible before its source enters the camera field of view, or the sound source is not visible at all, and (ii) due to the high costs of annotating video segments with per-segment labels, only modality-agnostic, video-level labels are provided during training, i.e., these labels specify which events occur in a video but lack details about the segments or the modality in which they occur. + +Prior works in the area can be grouped into two orthogonal research directions. The first focuses on enhancing model architectures [23, 45, 51]. Despite advancements in this direction, the absence of fine-grained labels to guide the model during training continues to pose an impediment towards the generalizability of such models. As a result, recent approaches have focused on the second direction of research which aims at generating richer pseudo-labels for improved training, either at the video-level [8, 39] or segment-level [9, 17, 26, 50]. In particular, Rachavarapu et al. [26] propose prototype-based pseudo-labeling (PPL), which seeks to train a pseudo-label generation module in conjunction with a core inference module. However, due to the lack of sufficient training data, this method struggles to generalize. On the other hand, VPLAN [50], VALOR [17], and LSLD [9] leverage large-scale pre-trained foundation models, such as CLIP [28] and CLAP [40], along with ground-truth video-level labels to generate segment-level pseudo-labels for each of the two modalities. Audio/Visual segments (e.g. the audio corresponding to the segment in question and the visual frame at the center of the segment) are fed into CLAP/CLIP, one segment at a time, to generate these pseudo-labels. Despite the significant improvement that these pseudo-label generation methods achieved, the correctness of the generated labels is still limited, constrained primarily by the lack of understanding of intersegment dynamics. For instance, if a crowd is cheering in a segment of the video, it is more likely that the crowd might also be clapping right before or after. + +To address the oversight of inter-segment dependencies and other shortcomings in existing pseudo-label generation methods, we introduce a novel, uncertainty-based, weakly-supervised, video parsing model called UWAV (Uncertainty-weighted Weakly-supervised Audiovisual Video Parsing), capable of generating improved segment-level pseudo-labels for better training of the inference module. We resort to transformer modules [38] to equip UWAV with the ability to capture temporal relationships between segments and pre-train it on a large-scale, supervised audio-visual event localization dataset [12]. Subsequently, this pre-trained model is used to generate segment-level pseudo-labels for each modality on a target, small-scale dataset which only provides weak (i.e., video-level) supervision. Such a design permits a more holis + +tic understanding of the video, resulting in more accurate pseudo-labels. Additionally, UWAV factors in the uncertainty associated with these estimated pseudo-labels in its optimization. That uncertainty is the result of the shift in the domain of the target dataset, insufficient model capacity, etc. and is computed at training time as the confidence scores associated with these labels. To further enhance the model's ability to learn in this small-scale, weakly-supervised data regime, we also employ a feature mixup strategy. This approach adds more regularization constraints by training on mixed segment features alongside interpolated pseudo-labels, which not only lessens the influence of noise but also enriches the training data, thereby reducing overfitting. Moreover, UWAV addresses a critical class imbalance issue in the pseudo-label enriched training data, viz. most event classes in any given segment of a video are absent/negative (i.e., they do not occur), while only a handful of them do. This creates a natural bias in the training set, making it difficult to learn the positive events. To counter this, we propose a class-frequency aware weighting strategy that lays greater emphasis on the accurate classification and localization of positive events. By incorporating these components into its design, our proposed method (UWAV) outperforms competing state-of-the-art approaches across two different datasets, viz. Look, Listen, and Parse (LLP) [35] and the Audio-Visual Event Localization (AVE) [34], on multiple evaluation metrics. In summary, our contributions are the following: + +- We introduce a novel, weakly-supervised method called UWAV, capable of synthesizing temporally coherent pseudo-labels for the AVVP task. +- To the best of our knowledge, ours is the first method for the AVVP task, which factors in the uncertainty associated with the estimated pseudo-labels while also regularizing it with a feature mixup strategy. +- UWAV outperforms competing state-of-the-art approaches for the AVVP task, across two different datasets on multiple metrics which attest to its generalizability. + +# 2. Related Works + +Audio-Visual Learning: Audio-visual learning has emerged as an area of active research, aiming to develop models that synergistically integrate information from both audio and visual modalities for improved perception and understanding. Towards this end, various audio-visual tasks have been explored by the community so far, such as audio-visual segmentation [19, 22, 42, 49], sound source localization [15, 16, 24, 32], event localization [29, 34, 48], navigation [4-6, 21, 44, 46], generation [25, 30, 41], question answering [11, 18, 33, 47], and sound source separation [2, 7, 36, 43]. In this work, we focus on the task of audio-visual video parsing (AVVP) where the goal is to temporally localize events occurring in a video. + +Unlike many other audio-visual learning tasks, AVVP does not assume that events are always aligned across modalities. Some events could be exclusively uni-modal while others may have an audio-visual signature, which requires complex reasoning. + +Audio-Visual Video Parsing (AVVP): To address the challenges of the AVVP task [17, 26, 35], Tian et al. [35] proposed a Hybrid Attention Network (HAN) and a learnable Multi-modal Multiple Instance Learning (MMIL) pooling module. The HAN model facilitates the exchange of information within and across modalities using self-attention and cross-attention layers, while the MMIL pooling module aggregates segment-level event probabilities from both modalities to produce video-level probabilities. Building on this foundation, recent works advanced the field from the following two perspectives. The first group of studies [23, 45, 51] focuses on enhancing model architectures. In particular, Mo and Tian [23] proposed the Multimodal Grouping Network (MGN) to explicitly group semantically similar features within each modality to improve the reasoning process, while Yu et al. [45] proposed the Multi-modal Pyramid Attentional Network (MM-Pyramid) to capture events of varying durations by extracting features at multiple temporal scales. Our proposed method is orthogonal to this line of research and can be integrated with any of these backbones. + +The second direction focuses on generating pseudo-labels for improved training, either at the video-level [8, 39] or the segment-level [9, 17, 26, 50]. VPLAN [50], VALOR [17], and LSLD [9] utilize pre-trained CLIP [28] and CLAP [40] along with ground-truth video-level labels to predict pseudo-labels for each modality on a per-segment basis. In contrast, PPL [26] uses the HAN model itself to generate pseudo-labels by constructing prototype features for each class and assigning labels to each segment based on the similarity between its feature and the prototype features. While these pseudo-label generation methods have substantially improved model performance on the AVVP task, they still exhibit some limitations. For instance, to derive accurate pseudo-labels, PPL might require a large enough training set to learn good prototype features, which might pose challenges when applied to smaller datasets. Our proposed method overcomes this problem. On the other hand, methods that leverage CLIP and CLAP to generate pseudo-labels often ignore temporal relationships between segments or the uncertainty associated with these labels. Our work also seeks to plug this void. + +# 3. Preliminaries + +Problem Formulation: The AVVP task [35] aims to localize all visible and/or audible events in each one-second segment of a video. Specifically, an audible video is split + +into $T$ one-second segments, denoted as $\{V_t, A_t\}_{t=1}^T$ . Each segment is annotated with a pair of ground-truth labels $y_t^v \in \{0,1\}^C, y_t^a \in \{0,1\}^C$ , where $y_t^v$ denotes visual events, $y_t^a$ denotes audio events, and $C$ denotes the total number of events in the pre-defined event set of the data. However, owing to the weakly-supervised nature of the task setup $(y_t^v, y_t^a)$ are unavailable during training. Instead, only the modality-agnostic, video-level labels $y \in \{0,1\}^C$ are available, where 1 indicates the presence of an event at any time (either in the audio or visual stream or both) while 0 indicates an event's absence in the video. + +Pseudo-Label Based AVVP Framework: The Hybrid Attention Network (HAN) [35] is a commonly used model for the AVVP task. The model works by first utilizing pre-trained visual and audio backbones to extract features from the visual and audio segments respectively, which are then projected to two $d$ -dimensional feature spaces. The resulting visual segment-level features are denoted by $F^v = \{f_t^v\}_{t=1}^T \in \mathbb{R}^{T \times d}$ , while the audio segment-level features are denoted by $F^a = \{f_t^a\}_{t=1}^T \in \mathbb{R}^{T \times d}$ . These features are provided as input to the HAN model. In the model, information across segments within a modality and across modalities is exchanged through self-attention and cross-attention layers, as shown below: + +$$ +\tilde {f} _ {t} ^ {v} = f _ {t} ^ {v} + \underbrace {\operatorname {A t t n} \left(f _ {t} ^ {v} , F ^ {v} , F ^ {v}\right)} _ {\text {S e l f - A t t e n t i o n}} + \underbrace {\operatorname {A t t n} \left(f _ {t} ^ {v} , F ^ {a} , F ^ {a}\right)} _ {\text {C r o s s - A t t e n t i o n}}, \tag {1} +$$ + +$$ +\tilde {f} _ {t} ^ {a} = f _ {t} ^ {a} + \underbrace {\operatorname {A t t n} \left(f _ {t} ^ {a} , F ^ {a} , F ^ {a}\right)} _ {\text {S e l f - A t t e n t i o n}} + \underbrace {\operatorname {A t t n} \left(f _ {t} ^ {a} , F ^ {v} , F ^ {v}\right)} _ {\text {C r o s s - A t t e n t i o n}}, \tag {2} +$$ + +where $\mathrm{Attn}(Q,K,V)$ denotes the standard multi-head attention mechanism [38]. Finally a classifier, shared across both modalities, transforms the visual segment-level features $\tilde{F}^v = \{\tilde{f}_t^v\}_{t=1}^T \in \mathbb{R}^{T \times d}$ (resp. audio segment-level features $\tilde{F}^a = \{f_t^a\}_{t=1}^T$ into visual segment-level logits $\{z_t^v\}_{t=1}^T \in \mathbb{R}^{T \times C}$ (resp. audio segment-level logits $\{z_t^a\}_{t=1}^T$ ). Segment-level probabilities $\{p_t^v\}_{t=1}^T, \{p_t^a\}_{t=1}^T \in \mathbb{R}^{T \times C}$ are then obtained by applying the sigmoid function on $\{z_t^v\}_{t=1}^T$ and $\{z_t^a\}_{t=1}^T$ . + +Since, only video-level labels $y$ are available during training, Tian et al. [35] introduce an attentive MMIL pooling module to learn to predict video-level probabilities $p \in \mathbb{R}^{C}$ : + +$$ +W _ {\text {m o d a l}} ^ {v, a} = \operatorname {S o f t m a x} _ {\text {m o d a l}} \left(\operatorname {F C} _ {\text {m o d a l}} \left(\tilde {F} ^ {v, a}\right)\right), \tag {3} +$$ + +$$ +W _ {\text {t i m e}} ^ {v, a} = \operatorname {S o f t m a x} _ {\text {t i m e}} \left(\mathrm {F C} _ {\text {t i m e}} \left(\tilde {F} ^ {v, a}\right)\right), \tag {4} +$$ + +where $\mathrm{FC}_{\mathrm{mod}}$ and $\mathrm{FC}_{\mathrm{time}}$ are two learnable fully-connected layers, $\tilde{F}^{v,a} = \mathrm{Stack}(\tilde{F}^v,\tilde{F}^a)\in \mathbb{R}^{2\times T\times d}$ denotes the stacked visual and audio features along the first dimension, $\mathrm{Softmax}_{\mathrm{mod}}(\cdot)$ denotes the softmax operation along the modality dimension (i.e., across $v,a$ ), while + +![](images/029b65d7fd68f8213edf4cde2b94874146ec42b34ecad0188cda9c3d73940c33.jpg) +Stage 1: Training the pseudo label models on larger dataset + +![](images/a1622defeee030970a2302157682bd239910b8e4aa1869a6824d634183d9653d.jpg) +Figure 2. UWAV framework: In stage 1, pseudo-label generation modules are equipped with the ability to capture temporal relationships between segments by pre-training on a large-scale, supervised, audio-visual event localization dataset. In stage 2, temporally coherent, uncertainty-weighted pseudo-labels, derived from the pre-trained pseudo-label generation module, are used to guide the learning of the inference model (HAN) aided by a class-balanced loss re-weighting and uncertainty-weighted feature mixup strategy. Note that we use the feature mixup strategy in both modalities while we only show the breakdown of the mixup operation for the audio modality. + +![](images/ab9adea9eb46a43f9ff1794ff0eeb3dbf4d2001c9b90e141286f7addd551c025.jpg) + +$\operatorname{Softmax}_{time}(\cdot)$ denotes the softmax operation along the temporal dimension (i.e., across $1, \ldots, T$ ). Video-level probabilities $p \in \mathbb{R}^C$ are then obtained via: + +$$ +p = \sum_ {m = \{v, a \}} \sum_ {t = 1} ^ {T} W _ {\text {m o d a l}} ^ {m, t} \odot W _ {\text {t i m e}} ^ {m, t} \odot p _ {t} ^ {m}, \tag {5} +$$ + +where $\odot$ denotes the element-wise product. The HAN model is then optimized with the binary cross-entropy (BCE) loss between the estimated video-level probabilities $p$ and video-level labels $y$ : $\mathcal{L}_{video} = \mathrm{BCE}(p, y)$ . + +# 4. Proposed Approach + +In this section, we detail our proposed approach (UWAV). At a high level, UWAV works by generating better segment-level pseudo-labels to improve the training of a multi-modal transformer-based inference module, e.g. HAN. Moreover, UWAV factors in the uncertainty associated with these pseudo-labels, addresses the imbalance in the training data, and introduces self-supervised regularization constraints, which all lead to better performance. Figure 2 shows an overview of our proposed framework. + +# 4.1. Temporally-Coherent Pseudo-Label Synthesis + +One major issue that plagues prior works, based on pseudolabel generation [9, 17, 50], is that they do not capture the temporal dependencies between neighboring segments when generating the pseudo-labels. That is, the generated pseudo-labels are not temporally coherent. To plug this void, we propose to incorporate transformer modules [38] into the pseudo-label generation pipeline, which maps CLIP/CLAP encodings of a segment's visual frame/audio to pseudo-labels. Specifically, two separate transformers are introduced, one each for the visual/audio pseudo-label synthesis modules. + +Pre-Training: Training transformers often requires sufficiently large training data while datasets commonly used for the weakly-supervised AVVP task are relatively small. To mitigate this challenge, we propose to first pre-train the transformer-equipped pseudo-label generation module on a large-scale, supervised, audio-visual event localization dataset - the UnAV [12] dataset. Specifically, given an audible video of duration $T'$ seconds from the pre-training dataset, we split the video into $T'$ one-second segments $\{V_t', A_t'\}_{t=1}^{T'}$ , with corresponding audio-visual event labels $y_t^{av'} \in \{0,1\}^{C'}$ , where 1 indicates the presence of an event in both modalities and 0 its absence in at least one modality, while $C'$ denotes the total number of event classes in the pre-training dataset. Next, the video frame at the temporal center of the visual segment is transformed into visual features $G_0^{v'} = \{g_{0,t}'\}_{t=1}^{T'} \in \mathbb{R}^{T' \times d_1}$ with CLIP's [28] image encoder. These features are then fed into the corresponding transformer of the visual stream, consisting of $L$ encoder blocks, each block containing a self-attention layer, LayerNorm [1] (LN), and a 2-layer feed-forward network (FFN): + +$$ +\tilde {G} _ {l} ^ {v ^ {\prime}} = \mathrm {L N} \left(G _ {l} ^ {v ^ {\prime}} + \operatorname {A t t n} \left(G _ {l} ^ {v ^ {\prime}}, G _ {l} ^ {v ^ {\prime}}, G _ {l} ^ {v ^ {\prime}}\right)\right), \tag {6} +$$ + +$$ +G _ {l + 1} ^ {v ^ {\prime}} = \operatorname {L N} \left(\tilde {G} _ {l} ^ {v ^ {\prime}} + \operatorname {F F N} \left(\tilde {G} _ {l} ^ {v ^ {\prime}}\right)\right). \tag {7} +$$ + +Concurrently, we convert each event category label in the pre-training dataset into a textual event feature $e_c^{CLIP'} \in \mathbb{R}^{d_1}$ by filling in the pre-defined caption template: "A photo of " with the corresponding event name and passing it to CLIP's text encoder. Equipped with the visual segment-level features $G_L^{v'} = \{g_{L,t}'\}_{t=1}^{T'} \in \mathbb{R}^{T' \times d_1}$ and the textual event features $E^{CLIP'} = \{e_c^{CLIP'}\}_{c=1}^{C'} \in \mathbb{R}^{C' \times d_1}$ , we derive visual segment-level logits $\hat{z}_t^{v'} \in \mathbb{R}^{C'}$ and probabilities $\hat{p}_t^{v'}$ as follows: + +$$ +\hat {p} _ {t} ^ {v ^ {\prime}} = \operatorname {S i g m o i d} \left(\hat {z} _ {t} ^ {v ^ {\prime}}\right), \hat {z} _ {t} ^ {v ^ {\prime}} = E ^ {C L I P ^ {\prime}} \cdot g _ {L, t} ^ {v ^ {\prime} \top}. \tag {8} +$$ + +Similar operations are performed in the audio pseudolabel generation pipeline. The raw waveforms corresponding to the 1-second audio segments are transformed into audio features $G_0^{a'} \in \mathbb{R}^{T' \times d_2}$ with CLAP's [12] audio encoder and fed into the corresponding transformer consisting of $L$ encoder blocks. Correspondingly, the textual event features $E^{CLAP'} \in \mathbb{R}^{C' \times d_2}$ are generated with the caption template: "This is the sound of " by passing it through CLAP's text encoder. Audio segment-level logits $\hat{z}_t^{a'} \in \mathbb{R}^{C'}$ and probabilities $\hat{p}_t^{a'}$ can then be derived in the same manner: $\hat{p}_t^{a'} = \text{Sigmoid}(\hat{z}_t^{a'})$ , $\hat{z}_t^{a'} = E^{CLAP'} \cdot g_t^{a'}^\top$ . + +Since the events occurring in the pre-training dataset (UnAV) are audio-visual, we multiply the segment-level visual and audio event probabilities to enforce the predicted labels to be multi-modal in nature: $\{\hat{p}_t^{av'}\}_{t=1}^{T'} \in \mathbb{R}^{T' \times C'}$ . This network is then trained with the binary cross-entropy (BCE) loss: + +$$ +\mathcal {L} _ {\text {t e m p}} = \mathrm {B C E} \left(\hat {p} _ {t} ^ {a v ^ {\prime}}, y _ {t} ^ {a v ^ {\prime}}\right), \hat {p} _ {t} ^ {a v ^ {\prime}} = \tilde {p} _ {t} ^ {v ^ {\prime}} \odot \hat {p} _ {t} ^ {a ^ {\prime}}. \tag {9} +$$ + +Pseudo-Label Generation on Target Dataset: With the pre-trained pseudo-label generation modules in place, we proceed to employ them for the pseudo-label generation process in the target dataset for the AVVP task. Specifically, the center frame of each of the visual segments $\{V_t\}_{t=1}^T$ of the target dataset are passed into CLIP's image encoder, whose output is passed into the pre-trained visual transformer to generate segment features $G_L^v = \{g_{L,t}^v\}_{t=1}^T \in \mathbb{R}^{T \times d_1}$ . At the same time, the caption template: "A photo of " is used to obtain textual features corresponding to each of the event classes in the target dataset for the AVVP task: $E^{CLIP} \in \mathbb{R}^{C \times d_1}$ . Segment-level visual logits $\hat{z}_t^v \in \mathbb{R}^C$ can be derived by computing their inner product. We also pre-define class-wise visual thresholds $\theta^v \in \mathbb{R}^C$ to transform segment-level visual logits into binary pseudo-labels $\hat{y}_t^v \in \mathbb{R}^C$ : + +$$ +\hat {y} _ {t} ^ {v} = \mathbb {1} _ {\left\{\hat {z} _ {t} ^ {v} > \theta^ {v} \right\}} \odot y, \hat {z} _ {t} ^ {v} = E ^ {C L I P} \cdot g _ {L, t} ^ {v} ^ {\top}, \tag {10} +$$ + +where $y$ denotes the ground-truth video-level labels, $\mathbb{1}_{\{\cdot\}}$ is the indicator function which returns a value of 1 when the condition is true otherwise 0, and $\odot$ denotes the elementwise product operation. The $\odot$ operation zeroes out the predictions of event classes absent in the video-level label. + +A similar pseudo-label generation process is employed on the acoustic side. Raw waveforms of audio segments are first fed into CLAP's audio encoder and then into the pretrained audio transformer. The event names of the classes in the target dataset for the AVVP task are filled in the caption template: "This is the sound of " to generate textual event features: $E^{CLAP} \in \mathbb{R}^{C \times d_1}$ . Segment-level audio logits $\hat{z}_t^a \in \mathbb{R}^C$ and binary pseudo-labels $\hat{y}_t^a \in \mathbb{R}^C$ are then derived using class-wise thresholds $\theta^a \in \mathbb{R}^C$ . + +With binary segment-level pseudo-labels for both modalities $\hat{y}_t^v, \hat{y}_t^a$ and the predicted probabilities from the infer + +ence module (HAN) $\hat{p}_t^v,\hat{p}_t^a$ in place, the inference module can be trained using the binary cross-entropy loss as shown: + +$$ +\mathcal {L} _ {h a r d} = \mathrm {B C E} \left(p _ {t} ^ {v}, \hat {y} _ {t} ^ {v}\right) + \mathrm {B C E} \left(p _ {t} ^ {a}, \hat {y} _ {t} ^ {a}\right). \tag {11} +$$ + +# 4.2. Training with Pseudo-Label Uncertainty + +While pseudo-labels do provide additional supervision for better training of the inference module, they could potentially be noisy, leading to occasionally incorrect training signals. To ameliorate this problem, we propose an uncertainty-weighted pseudo-label training scheme to improve the robustness of the learning process. Instead of simply training with the binary pseudo-labels $\hat{y}_t^v,\hat{y}_t^a$ , we leverage the confidence of the pseudo-label estimation module (associated with the predicted pseudo-label) to weigh the training signal for the inference module. This confidence score serves as a measure of the pseudo-label generation module's uncertainty of its prediction. This may be represented as: + +$$ +\hat {p} _ {t} ^ {v} = \operatorname {S i g m o i d} \left(\hat {z} _ {t} ^ {v} - \theta^ {v}\right) \odot y; \quad \hat {p} _ {t} ^ {a} = \operatorname {S i g m o i d} \left(\hat {z} _ {t} ^ {a} - \theta^ {a}\right) \odot y. \tag {12} +$$ + +In other words, considering the visual pseudo-label generation pipeline as an example, the farther the logit $\hat{z}_t^v$ is from the threshold $\theta^v$ , whether much lower or higher, the more confident the pseudo-label generation module is about the label it predicts (either approaching 0 or 1). Conversely, the closer the logit is to the threshold, the less the certainty about the correctness of the pseudo-labels (probabilities closer to 0.5). An analogous explanation also holds for the audio pseudo-labels. With the uncertainty-weighted pseudo-labels in place, the inference module (HAN) can be trained with the following uncertainty-weighted pseudo-label loss: + +$$ +\mathcal {L} _ {\text {s o f t}} = \mathrm {B C E} \left(p _ {t} ^ {v}, \hat {p} _ {t} ^ {v}\right) + \mathrm {B C E} \left(p _ {t} ^ {a}, \hat {p} _ {t} ^ {a}\right). \tag {13} +$$ + +# 4.3. Uncertainty-weighted Feature Mixup + +Due to the lack of full supervision for the weakly-supervised AVVP task, we explore the efficacy of additional regularization via self-supervision to help the models generalize better. Towards this end, prior pseudo-label generation approaches [26, 39] often employ contrastive learning as a tool to better train the inference module. However, due to the inherent noise in the estimated pseudo-labels, positive samples and negative samples may be mislabeled, decreasing the effectiveness of the self-supervisory training. As an alternative, in this work, we explore the effectiveness of feature mixing, as a self-supervisory training signal for additional regularization. In this setting, we mixup the estimated features of any two segments, additively, and train the model to predict the union of the labels of the two segments. However, since the labels in our setting are noisy, the mixed feature is assigned a label derived from a weighted sum of + +the uncertainty-weighted pseudo-labels of each of the two segment features. This is illustrated below: + +$$ +\bar {f} _ {t _ {i}, t _ {j}} ^ {v} = \lambda \tilde {f} _ {t _ {i}} ^ {v} + (1 - \lambda) \tilde {f} _ {t _ {j}} ^ {v}, \quad \bar {p} _ {t _ {i}, t _ {j}} ^ {v} = \lambda \hat {p} _ {t _ {i}} ^ {v} + (1 - \lambda) \hat {p} _ {t _ {j}} ^ {v} \tag {14} +$$ + +$$ +\bar {f} _ {t _ {i}, t _ {j}} ^ {a} = \lambda \tilde {f} _ {t _ {i}} ^ {a} + (1 - \lambda) \tilde {f} _ {t _ {j}} ^ {a}, \quad \bar {p} _ {t _ {i}, t _ {j}} ^ {a} = \lambda \hat {p} _ {t _ {i}} ^ {a} + (1 - \lambda) \hat {p} _ {t _ {j}} ^ {a}, \tag {15} +$$ + +where $\lambda \sim \mathrm{Beta}(\alpha, \alpha)$ and $\alpha$ is a hyper-parameter controlling the Beta distribution, and $t_i$ and $t_j$ indicate two segment indices in a batch of video segments. + +After mixing the uni-modal segment-level features, we pass them through the classifier of the inference module and apply the sigmoid function to the output, obtaining mixed segment-level event probabilities $p_t^{mix - v}$ and $p_t^{mix - a}$ . These are used to train the inference model with the uncertainty-aware mixup loss, as shown below: + +$$ +\mathcal {L} _ {m i x} = \mathrm {B C E} \left(p _ {t} ^ {m i x - v}, \bar {p} _ {t} ^ {v}\right) + \mathrm {B C E} \left(p _ {t} ^ {m i x - a}, \bar {p} _ {t} ^ {a}\right). \tag {16} +$$ + +# 4.4. Class-balanced Loss Re-weighting + +Besides the aforementioned challenges of the AVVP task, most of the events in the event set are absent in the pseudolabels of any (segment of a) video (i.e., most event classes are negative events) and only a few events are present (i.e., positive events are much fewer in number). As a result, the model is dominated by the loss from the negative events. When trained without factoring in this bias, the classifier tends to overfit the negative labels and ignore the positive ones. To address this class imbalance issue, we introduce a class-balanced loss re-weighting strategy to re-balance the importance of the losses from the negative and positive events for the uncertainty-weighted pseudo-label loss. Specifically, the loss from the positive events is multiplied by a weight proportional to the frequency of the segments with the negative events in the pseudo-labels, while the loss from the negative events is multiplied by a weight proportional to the frequency of the segments with the positive events in the pseudo-labels, as shown below: + +$$ +\begin{array}{l} \mathcal {L} _ {w - s o f t} = \sum_ {m \in \{v, a \}} w _ {p o s} ^ {m} \cdot y \cdot \mathrm {B C E} \left(p _ {t} ^ {m}, \hat {p} _ {t} ^ {m}\right) + \\ w _ {n e g} ^ {m} \cdot (1 - y) \cdot \mathrm {B C E} \left(p _ {t} ^ {m}, \hat {p} _ {t} ^ {m}\right), \tag {17} \\ \end{array} +$$ + +$$ +w _ {p o s} ^ {m} = \frac {\sum_ {i = 1} ^ {N} \sum_ {t = 1} ^ {T} \sum_ {c = 1} ^ {C} \left(1 - \hat {y} _ {i , t , c} ^ {m}\right)}{N T C} \times W, \tag {18} +$$ + +$$ +w _ {n e g} ^ {m} = \frac {\sum_ {i = 1} ^ {N} \sum_ {t = 1} ^ {T} \sum_ {c = 1} ^ {C} \hat {y} _ {i , t , c} ^ {m}}{N T C}, \tag {19} +$$ + +where $N$ denotes the number of videos in the training set, and $W$ is a hyper-parameter. + +In summary, the inference module is trained on the AVVP task with the proposed class-balanced re-weighting, applied to the uncertainty-weighted classification loss, and the uncertainty-weighted feature mixup loss, as shown below: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {w - s o f t} + \mathcal {L} _ {\text {m i x}} + \mathcal {L} _ {\text {v i d e o}}. \tag {20} +$$ + +# 5. Experiments + +We assess the performance of UWAV empirically across two challenging datasets and report its performance, comparing it with existing state-of-the-art approaches both quantitatively and qualitatively. Additionally, through multiple ablation studies, we bring out the effectiveness of the different elements of our proposed approach and the choices of different hyper-parameters. For additional details, ablation studies, and more qualitative results, we refer the reader to our supplementary material. + +# 5.1. Experimental Setup + +Datasets: We evaluate all competing methods on the Look, Listen, and Parse (LLP) dataset [35], which is the principal benchmark dataset for the AVVP task. The dataset consists of 11,849 video clips sourced from YouTube. Each clip is 10 seconds long and represents one or more of 25 diverse event categories, such as human activities, animals, musical instruments, and vehicles. The dataset is split into training, validation, and testing sets, following the official split [35]: 10,000 videos for training, 649 videos for validation, and 1,200 videos for testing. While the training set of this dataset is only associated with video-level labels of the events, the validation and testing split is labeled with segment-level event labels for evaluation purposes. Additionally, to demonstrate the generalizability of our method, we conduct a similar study on the Audio Visual Event (AVE) recognition dataset [34]. The AVE dataset consists of 4,143 video clips crawled from YouTube, each 10 seconds long. It is split into 3,339 videos for training, 402 for validation, and 402 for testing. It includes 29 event categories encompassing human activities, animals, musical instruments, vehicles, and a "background" class (i.e., no event occurs). Unlike the LLP dataset, each video in the AVE dataset contains only one audio-visual event. Here too, the training data is only provided with video-level labels while the validation and test splits are annotated with ground-truth event labels for each one-second segment, which either is a specific audio-visual event or "background". + +Metrics: For the LLP dataset, following the standard evaluation protocol [35], all models are evaluated using macro F1-scores calculated for the following event types: (i) audio-only (A), (ii) visual-only (V), and (iii) audio-visual (AV). Type@AV (Type) and Event@AV (Event) are two additional metrics that evaluate the overall performance of the model, where Type@AV is the mean of the F1-scores for the A, V, and AV events, while Event@AV is the F1-score of all events regardless of the modality in which they occur. Evaluations are conducted at both the segment-level and the event-level. At the segment-level, the model's predictions are compared with the ground truth on a per-segment basis. At the event-level, consecutive positive segments for + +Table 1. Comparison with state-of-the-arts methods on the LLP dataset. Best performances are in bold, second-best in underlined. + +
MethodSegment-levelEvent-level
AVAVTypeEventAVAVTypeEvent
HAN [35]60.152.948.954.055.451.348.943.047.748.0
MA [39]60.360.055.158.957.953.656.449.053.050.6
JoMoLD [8]61.363.857.260.859.953.959.949.654.552.5
CMPAE [10]64.266.459.263.362.856.663.751.857.455.7
PoiBin [27]63.163.557.761.460.654.160.351.555.252.3
VPLAN [50]60.564.858.361.259.451.461.551.254.750.8
VALOR [17]61.865.958.462.061.555.462.652.256.754.2
LSLD [9]62.767.159.463.162.255.764.352.657.655.2
PPL [26]65.966.761.964.863.757.364.354.359.957.9
CoLeaf [31]64.264.459.362.662.557.663.254.257.955.6
LEAP [51]62.765.659.362.561.856.463.154.157.855.0
UWAV (Ours)64.270.063.465.963.958.666.757.560.957.4
+ +the same event are grouped together as a single event. The F1-score is then computed using a mIoU threshold of 0.5. For the AVE dataset, we follow Tian et al. [34] and use accuracy as the evaluation metric. An event prediction of a segment is considered correct if it matches the ground-truth label for that segment. + +Implementation Details: In line with prior work [35], each 10-second video in both datasets is split into 10 segments of one second each, where each segment contains 8 frames. For the LLP dataset, pre-trained ResNet-152 [13] and $\mathrm{R}(2 + 1)\mathrm{D - }18$ [37] are used to extract 2D and 3D visual features, respectively. The pre-trained VGGish [14] network is used to extract features from the audio, sampled at $16\mathrm{KHz}$ . For the AVE dataset however, akin to prior work [17], we extract visual features from pre-trained CLIP and $\mathrm{R}(2 + 1)\mathrm{D}$ while CLAP is used to embed the audio stream. For both datasets, we set the number of encoder blocks $L$ in the temporal-aware model to 5, $\alpha$ for the Beta distribution in the feature mixup strategy to 1.7, and $W$ in the class-balanced loss re-weighting step to 0.5. The pseudo-label generation modules and the inference model (HAN) are trained with the AdamW optimizer [20]. For improved performance on the AVE dataset, we replace $\hat{p}_{t_i}^a,\hat{p}_{t_j}^a$ in Eq. 15 with $[\hat{p}_{t_i}^a ],[\hat{y}_{t_j}^a ]$ and make corresponding modifications in the visual counterparts as well. + +Baselines: We demonstrate the effectiveness of UWAV by comparing against an extensive set of baselines. For the LLP dataset, this includes video-level pseudo-label generation methods (MA [39], JoMoLD [8]), segment-level pseudo-label generation methods (VALOR [17], LSLD [9], PPL [26]), and the recently released works (CoLeaf [31], LEAP [51]). On the other hand, for the AVE dataset, baseline approaches with publicly available implementation, which use the state-of-the-art feature backbones (akin to ours), such as HAN [35] and VALOR [17] are used. + +# 5.2. Results + +Comparison with Previous Methods on LLP: As shown in Table 1, UWAV surpasses previous methods, across almost all metrics. Notably, we achieve an F-score of 70.0 on the segment-level visual event, 65.9 on the segment-level Type@AV, and 66.7 on the event-level visual event metric. This corresponds to a gain of $1.1\%$ on segment-level Type@AV F-score and a $1\%$ improvement on event-level Type@AV F-score, over our closest competitor PPL [26]. Of particular note, is the fact that our segment-level and event-level F-scores improve by more than $3\%$ , over PPL, for visual events. When compared to other recently published works, such as VALOR [17], CoLeaf [31] and LEAP [51], UWAV outperforms them by up to $3\%$ on both segment and event-level Type@AV F-score while gains on visual events are up to $5\%$ on segment-level F-score. + +These observations are consistent with our qualitative comparisons, as well. In the example on the left of Figure 3, our model successfully recognizes and temporally localizes the lawn mower event visually, whereas VALOR [17] (a recent state-of-the-art approach with publicly available implementation) misclassifies it as a chainsaw. Additionally, our model also accurately localize the intermittent sound of the lawn mower. In contrast, VALOR not only misclassifies the sound of the lawn mower as that of a chainsaw but also incorrectly predicts that someone is talking in the video. In the example on the right of Figure 3, our model does not err in recognizing either the visual presence or the audio presence of the telephone, while VALOR fails to accurately predict the events in either modality. + +Comparison with Previous Methods on AVE: To demonstrate the generalizability of our method, we evaluate UWAV on the AVE [34] dataset and compare its per + +![](images/884745ea97cc18cb7ad4f14e8126d44f7e5060bd4eeb05af4c5ba4c0e7dfdbad.jpg) +Figure 3. Comparison between predictions by UWAV and competing AVVP methods on the LLP dataset. "GT": ground truth. + +![](images/7bb483f21c0388ed8ddd8e93910587c4b0b02a5c664f4c925aacace79b49e052.jpg) + +Table 2. Model performances on the AVE dataset. CLIP, R(2+1)D-18, and CLAP are used as feature backbones. + +
MethodHAN [35]VALOR [17]UWAV (Ours)
Acc.(%)75.380.480.6
+ +Table 3. Accuracy of the generated pseudo-labels on LLP. + +
MethodSegment-level
AVAVTypeEvent
VALOR [17]80.561.755.766.074.6
PPL [26]61.761.857.560.659.4
UWAV (Ours)78.474.565.572.878.4
+ +formance with that of previous works. From Table 2, we observe that with the same backbone features, UWAV surpasses VALOR, our closest competitor, even on this small-scale dataset. + +Accuracy of the Generated Pseudo-Labels: To evaluate the efficacy of our pseudo-label generation pipeline, we compare the accuracy of our generated pseudo-labels against those obtained from competing methods (with publicly available implementation) [17, 26] on the test set of the LLP dataset. As shown in Table 3, our pre-trained temporally-dependent pseudo-label generation scheme generates more accurate segment-level pseudo-labels than previous methods, by up to $6\%$ on the segment-level Type@AV F-score, attesting to the advantages of factoring in intersegment temporal dependencies. + +# 5.3. Ablation Study + +To demonstrate the potency of the different elements of our proposed method, UWAV, we conduct ablation studies. In particular, the proposed uncertainty-weighted pseudo-label based training, the uncertainty-weighted feature mixup scheme, and the class-balanced loss re-weighting schemes are ablated. As shown in Table 4, incorporating the uncertainty-weighted pseudo-label training step improves + +Table 4. Ablation study of the proposed components in UWAV. "Binary" denotes training with binary pseudo-labels. "Soft" denotes training with uncertainty-weighted pseudo-labels. + +
BinarySoftRe-weightMixupSegment-level
AVAVTypeEvent
✓62.767.761.964.262.2
✓63.068.361.864.462.8
✓✓63.669.563.065.463.1
✓✓63.969.062.865.263.4
✓✓✓64.270.063.465.963.9
+ +the segment-level Type@AV F-score by $2\%$ , compared to using binary pseudo-labels. This demonstrates the benefit of accounting for the uncertainty in the pseudo-label estimation module. Moreover, sans the class-balanced loss weighting strategy, the model's performance is worse off by $1\%$ on the Type@AV F-score, revealing the erroneous bias in the model's prediction arising from a skew of the class distribution. On the other hand, introducing the uncertainty-weighted feature mixup results in a gain of $0.8\%$ on the Type@AV F-score, which underscores the importance of this self-supervised regularization. + +# 6. Conclusion + +In this work, we address the challenging task of weakly-supervised audio-visual video parsing (AVVP), which presents a two-fold challenge: (i) potential misalignment between the events of the audio and visual streams, and (ii) the lack of fine-grained labels for each modality. We observe that by considering the temporal relationship between segments, our proposed method (UWAV) is able to provide more reliable pseudo-labels for better training of the inference module. In addition, by factoring in the uncertainty associated with these estimated pseudo-labels, regularizing the training process with a feature mixup strategy, and correcting for class imbalance, UWAV achieves state-of-the-art results on the LLP and AVE datasets. + +# References + +[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4 +[2] Moitreya Chatterjee, Jonathan Le Roux, Narendra Ahuja, and Anoop Cherian. Visual scene graphs for audio source separation. In ICCV, 2021. 2 +[3] Moitreya Chatterjee, Narendra Ahuja, and Anoop Cherian. Learning audio-visual dynamics using scene graphs for audio source separation. In NeurIPS, 2022. 1 +[4] Changan Chen, Unnat Jain, Carl Schissler, Sebastia Vicenc Amengual Gari, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audio-visual navigation in 3d environments. In ECCV, 2020. 2 +[5] Changan Chen, Ziad Al-Halah, and Kristen Grauman. Semantic audio-visual navigation. In CVPR, 2021. +[6] Changan Chen, Sagnik Majumder, Ziad Al-Halah, Ruohan Gao, Santhosh Kumar Ramakrishnan, and Kristen Grauman. Learning to set waypoints for audio-visual navigation. In ICLR, 2021. 2 +[7] Jiaben Chen, Renrui Zhang, Dongze Lian, Jiaqi Yang, Ziyao Zeng, and Jianbo Shi. iquery: Instruments as queries for audio-visual sound separation. In CVPR, 2023. 1, 2 +[8] Haoyue Cheng, Zhaoyang Liu, Hang Zhou, Chen Qian, Wayne Wu, and Limin Wang. Joint-modal label denoising for weakly-supervised audio-visual video parsing. In ECCV, 2022. 2, 3, 7 +[9] Yingying Fan, Yu Wu, Yutian Lin, and Bo Du. Revisit weakly-supervised audio-visual video parsing from the language perspective. In NeurIPS, 2023. 2, 3, 4, 7 +[10] Junyu Gao, Mengyuan Chen, and Changsheng Xu. Collecting cross-modal presence-absence evidence for weakly-supervised audio-visual event perception. In CVPR, 2023. 7 +[11] Shijie Geng, Peng Gao, Moitreya Chatterjee, Chiori Hori, Jonathan Le Roux, Yongfeng Zhang, Hongsheng Li, and Anoop Cherian. Dynamic graph representation learning for video dialog via multi-modal shuffled transformers. In AAAI, 2021. 2 +[12] Tiantian Geng, Teng Wang, Jinming Duan, Runmin Cong, and Feng Zheng. Dense-localizing audio-visual events in untrimmed videos: A large-scale benchmark and baseline. In CVPR, 2023. 2, 4, 5 +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 7 +[14] Shawn Hershey, Sourish Chaudhuri, Daniel PW Ellis, Jort F Gemmeke, Aren Jansen, R Channing Moore, Manoj Plakal, Devin Platt, Rif A Saurous, Bryan Seybold, et al. Cnn architectures for large-scale audio classification. In ICASSP, 2017. 7 +[15] Xixi Hu, Ziyang Chen, and Andrew Owens. Mix and localize: Localizing sound sources in mixtures. In CVPR, 2022. 2 + +[16] Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Egocentric audio-visual object localization. In CVPR, 2023. 2 +[17] Yung-Hsuan Lai, Yen-Chun Chen, and Frank Wang. Modality-independent teachers meet weakly-supervised audio-visual event parser. In NeurIPS, 2023. 2, 3, 4, 7, 8 +[18] Guangyao Li, Yake Wei, Yapeng Tian, Chenliang Xu, JiRong Wen, and Di Hu. Learning to answer questions in dynamic audio-visual scenarios. In CVPR, 2022. 2 +[19] Jinxiang Liu, Yikun Liu, Fei Zhang, Chen Ju, Ya Zhang, and Yanfeng Wang. Audio-visual segmentation via unlabeled frame exploitation. In CVPR, 2024. 1, 2 +[20] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 7 +[21] Sagnik Majumder, Ziad Al-Halah, and Kristen Grauman. Move2hear: Active audio-visual source separation. In ICCV, 2021. 2 +[22] Yuxin Mao, Jing Zhang, Mochu Xiang, Yiran Zhong, and Yuchao Dai. Multimodal variational auto-encoder based audio-visual segmentation. In ICCV, 2023. 2 +[23] Shentong Mo and Yapeng Tian. Multi-modal grouping network for weakly-supervised audio-visual video parsing. In NeurIPS, 2022. 2, 3 +[24] Shentong Mo and Yapeng Tian. Audio-visual grouping network for sound localization from mixtures. In CVPR, 2023. 1, 2 +[25] Kranti Kumar Parida, Siddharth Srivastava, and Gaurav Sharma. Beyond mono to binaural: Generating binaural audio from mono audio with depth and cross modal attention. In WACV, 2022. 2 +[26] Kranthi Kumar Rachavarapu, Kalyan Ramakrishnan, et al. Weakly-supervised audio-visual video parsing with prototype-based pseudo-labeling. In CVPR, 2024. 2, 3, 5, 7, 8 +[27] Kranthi Kumar Rachavarapu et al. Boosting positive segments for weakly-supervised audio-visual video parsing. In ICCV, 2023. 7 +[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 4 +[29] Varshanth Rao, Md Ibrahim Khalil, Haoda Li, Peng Dai, and Juwei Lu. Dual perspective network for audio-visual event localization. In ECCV, 2022. 1, 2 +[30] Ludan Ruan, Yiyang Ma, Huan Yang, Huiguo He, Bei Liu, Jianlong Fu, Nicholas Jing Yuan, Qin Jin, and Baining Guo. Mm-diffusion: Learning multi-modal diffusion models for joint audio and video generation. In CVPR, 2023. 2 +[31] Faegheh Sardari, Armin Mustafa, Philip JB Jackson, and Adrian Hilton. Coleaf: A contrastive-collaborative learning framework for weakly supervised audio-visual video parsing. In ECCV, 2024. 7 +[32] Arda Senocak, Hyeonggon Ryu, Junsik Kim, Tae-Hyun Oh, Hanspeter Pfister, and Joon Son Chung. Sound source localization is all about cross-modal alignment. In ICCV, 2023. 1, 2 + +[33] Ankit Shah, Shijie Geng, Peng Gao, Anoop Cherian, Takaaki Hori, Tim K Marks, Jonathan Le Roux, and Chiori Hori. Audio-visual scene-aware dialog and reasoning using audio-visual transformers with joint student-teacher learning. In ICASSP, 2022. 2 +[34] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-visual event localization in unconstrained videos. In ECCV, 2018. 1, 2, 6, 7 +[35] Yapeng Tian, Dingzeyu Li, and Chenliang Xu. Unified multisensory perception: Weakly-supervised audio-visual video parsing. In ECCV, 2020. 1, 2, 3, 6, 7, 8 +[36] Yapeng Tian, Di Hu, and Chenliang Xu. Cyclic co-learning of sounding object visual grounding and sound separation. In CVPR, 2021. 2 +[37] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In CVPR, 2018. 7 +[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2, 3, 4 +[39] Yu Wu and Yi Yang. Exploring heterogeneous clues for weakly-supervised audio-visual video parsing. In CVPR, 2021. 2, 3, 5, 7 +[40] Yusong Wu, Ke Chen, Tianyu Zhang, Yuchen Hui, Taylor Berg-Kirkpatrick, and Shlomo Dubnov. Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation. In ICASSP, 2023. 2, 3 +[41] Yazhou Xing, Yingqing He, Zeyue Tian, Xintao Wang, and Qifeng Chen. Seeing and hearing: Open-domain visual-audio generation with diffusion latent aligners. In CVPR, 2024. 2 +[42] Qi Yang, Xing Nie, Tong Li, Pengfei Gao, Ying Guo, Cheng Zhen, Pengfei Yan, and Shiming Xiang. Cooperation does matter: Exploring multi-order bilateral relations for audio-visual segmentation. In CVPR, 2024. 2 +[43] Yuxin Ye, Wenming Yang, and Yapeng Tian. Lavss: Location-guided audio-visual spatial audio separation. In WACV, 2024. 1, 2 +[44] Abdelrahman Younes, Daniel Honerkamp, Tim Welschehold, and Abhinav Valada. Catch me if you hear me: Audio-visual navigation in complex unmapped environments with moving sounds. IEEE Robotics and Automation Letters, 2023. 2 +[45] Jiashuo Yu, Ying Cheng, Rui-Wei Zhao, Rui Feng, and Yuejie Zhang. Mm-pyramid: Multimodal pyramid attentional network for audio-visual event localization and video parsing. In ACM MM, 2022. 2, 3 +[46] Yinfeng Yu, Wenbing Huang, Fuchun Sun, Changan Chen, Yikai Wang, and Xiaohong Liu. Sound adversarial audio-visual navigation. In ICLR, 2022. 2 +[47] Heeseung Yun, Youngjae Yu, Wonsuk Yang, Kangil Lee, and Gunhee Kim. Pano-avqa: Grounded audio-visual question answering on 360deg videos. In ICCV, 2021. 2 +[48] Jinxing Zhou, Liang Zheng, Yiran Zhong, Shijie Hao, and Meng Wang. Positive sample propagation along the audio-visual event line. In CVPR, 2021. 2 + +[49] Jinxing Zhou, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audio-visual segmentation. In ECCV, 2022. 1, 2 +[50] Jinxing Zhou, Dan Guo, Yiran Zhong, and Meng Wang. Improving audio-visual video parsing with pseudo visual labels. arXiv preprint arXiv:2303.02344, 2023. 2, 3, 4, 7 +[51] Jinxing Zhou, Dan Guo, Yuxin Mao, Yiran Zhong, Xiaojun Chang, and Meng Wang. Label-anticipated event disentanglement for audio-visual video parsing. arXiv preprint arXiv:2407.08126, 2024. 2, 3, 7 \ No newline at end of file diff --git a/2025/UWAV_ Uncertainty-weighted Weakly-supervised Audio-Visual Video Parsing/images.zip b/2025/UWAV_ Uncertainty-weighted Weakly-supervised Audio-Visual Video Parsing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6afe9be0d13f7a575d6094ef9326c10c8d012d42 --- /dev/null +++ b/2025/UWAV_ Uncertainty-weighted Weakly-supervised Audio-Visual Video Parsing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:209964a07a23786bdb87cb4600a5ebb5107cb8b17ba126e3a3cfc517c9beba55 +size 474741 diff --git a/2025/UWAV_ Uncertainty-weighted Weakly-supervised Audio-Visual Video Parsing/layout.json b/2025/UWAV_ Uncertainty-weighted Weakly-supervised Audio-Visual Video Parsing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..00f000a2d98a579fa547fcd49ed54eb70c3e3071 --- /dev/null +++ b/2025/UWAV_ Uncertainty-weighted Weakly-supervised Audio-Visual Video Parsing/layout.json @@ -0,0 +1,8298 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 103, + 536, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 103, + 536, + 122 + ], + "spans": [ + { + "bbox": [ + 58, + 103, + 536, + 122 + ], + "type": "text", + "content": "UWAV: Uncertainty-weighted Weakly-supervised Audio-Visual Video Parsing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 90, + 142, + 504, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 142, + 504, + 215 + ], + "spans": [ + { + "bbox": [ + 90, + 142, + 504, + 215 + ], + "type": "text", + "content": "Yung-Hsuan Lai" + }, + { + "bbox": [ + 90, + 142, + 504, + 215 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 90, + 142, + 504, + 215 + ], + "type": "text", + "content": ", Janek Ebbers" + }, + { + "bbox": [ + 90, + 142, + 504, + 215 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 90, + 142, + 504, + 215 + ], + "type": "text", + "content": ", Yu-Chiang Frank Wang" + }, + { + "bbox": [ + 90, + 142, + 504, + 215 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 90, + 142, + 504, + 215 + ], + "type": "text", + "content": ", François Germain" + }, + { + "bbox": [ + 90, + 142, + 504, + 215 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 90, + 142, + 504, + 215 + ], + "type": "text", + "content": ", Michael Jeffrey Jones" + }, + { + "bbox": [ + 90, + 142, + 504, + 215 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 90, + 142, + 504, + 215 + ], + "type": "text", + "content": ", Moitreya Chatterjee" + }, + { + "bbox": [ + 90, + 142, + 504, + 215 + ], + "type": "inline_equation", + "content": "^{3,\\ddagger}" + }, + { + "bbox": [ + 90, + 142, + 504, + 215 + ], + "type": "text", + "content": ", \n" + }, + { + "bbox": [ + 90, + 142, + 504, + 215 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 90, + 142, + 504, + 215 + ], + "type": "text", + "content": " Graduate Institute of Communication Engineering, National Taiwan University " + }, + { + "bbox": [ + 90, + 142, + 504, + 215 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 90, + 142, + 504, + 215 + ], + "type": "text", + "content": " NVIDIA, Taiwan \n" + }, + { + "bbox": [ + 90, + 142, + 504, + 215 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 90, + 142, + 504, + 215 + ], + "type": "text", + "content": " Mitsubishi Electric Research Labs (MERL) \n" + }, + { + "bbox": [ + 90, + 142, + 504, + 215 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 90, + 142, + 504, + 215 + ], + "type": "text", + "content": "r10942097@ntu.edu.tw " + }, + { + "bbox": [ + 90, + 142, + 504, + 215 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 90, + 142, + 504, + 215 + ], + "type": "text", + "content": "metro.smiles@gmail.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 266, + 290, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 290, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 290, + 579 + ], + "type": "text", + "content": "Audio-Visual Video Parsing (AVVP) entails the challenging task of localizing both uni-modal events (i.e., those occurring exclusively in either the visual or acoustic modality of a video) and multi-modal events (i.e., those occurring in both modalities concurrently). Moreover, the prohibitive cost of annotating training data with the class labels of all these events, along with their start and end times, imposes constraints on the scalability of AVVP techniques unless they can be trained in a weakly-supervised setting, where only modality-agnostic, video-level labels are available in the training data. To this end, recently proposed approaches seek to generate segment-level pseudolabels to better guide model training. However, the absence of inter-segment dependencies when generating these pseudo-labels and the general bias towards predicting labels that are absent in a segment limit their performance. This work proposes a novel approach towards overcoming these weaknesses called Uncertainty-weighted Weakly-supervised Audio-visual Video Parsing (UWAV). Additionally, our innovative approach factors in the uncertainty associated with these estimated pseudo-labels and incorporates a feature mixup based training regularization for improved training. Empirical results show that UWAV outperforms state-of-the-art methods for the AVVP task on multiple metrics, across two different datasets, attesting to its effectiveness and generalizability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 591, + 128, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 591, + 128, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 591, + 128, + 604 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 612, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 612, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 612, + 287, + 685 + ], + "type": "text", + "content": "Events that occur in the real world, often leave their imprint on the acoustic and visual modalities. Humans rely heavily on the synergy between their senses of sight and hearing to interpret such events. Audio-visual learning, which seeks to equip machines with a similar synergy, has emerged as one of the most important research areas within the multi" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 243, + 545, + 397 + ], + "blocks": [ + { + "bbox": [ + 309, + 243, + 545, + 397 + ], + "lines": [ + { + "bbox": [ + 309, + 243, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 309, + 243, + 545, + 397 + ], + "type": "image", + "image_path": "843db44b81fc66a66494d92c620dd3afa3214df81fbb5d2257344d9cd27e586c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 403, + 547, + 460 + ], + "lines": [ + { + "bbox": [ + 304, + 403, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 547, + 460 + ], + "type": "text", + "content": "Figure 1. A weakly-supervised AVVP task example. Events, considered in this task, might be unimodal or multimodal. Even multimodal events, may not be temporally aligned in the audio and visual modalities, e.g. the cello might only be visible in the first few seconds but might produce music, throughout the video." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 472, + 546, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 546, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 546, + 628 + ], + "type": "text", + "content": "modal machine learning community. It aims to leverage both these senses (modalities) jointly, to enhance machine perception and understanding of real-world events. Various audio-visual learning tasks have been studied towards this end, including audio-visual segmentation [19, 49], sound source localization [24, 32], audio-visual event localization [29, 34], and audio-visual sound separation [3, 7, 43]. However, many of these tasks assume that audio and visual streams would always be temporally aligned. This assumption often fails in real-world scenarios, where the sonic and visual imprints of events may not perfectly overlap. For instance, one might hear an emergency siren approaching from a distance before it appears in the field of view." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "In this work, in order to better understand the events occurring in a video, we explore the task of Audio-Visual Video Parsing (AVVP) [35]. Its goal is to recognize and localize all audio, visual, and audio-visual events occurring in the video. See Figure 1 for an example of this task. The task setup is to perform this prediction for every one-second segment of a video. This task poses two principal challenges," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "type": "text", + "content": "Code is available at: https://github.com/Franklin905/UWAV." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13561" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "from a machine learning standpoint: (i) The audio and visual events, that occur, might not be temporally aligned, e.g. if an event becomes audible before its source enters the camera field of view, or the sound source is not visible at all, and (ii) due to the high costs of annotating video segments with per-segment labels, only modality-agnostic, video-level labels are provided during training, i.e., these labels specify which events occur in a video but lack details about the segments or the modality in which they occur." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 183, + 289, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 183, + 289, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 289, + 531 + ], + "type": "text", + "content": "Prior works in the area can be grouped into two orthogonal research directions. The first focuses on enhancing model architectures [23, 45, 51]. Despite advancements in this direction, the absence of fine-grained labels to guide the model during training continues to pose an impediment towards the generalizability of such models. As a result, recent approaches have focused on the second direction of research which aims at generating richer pseudo-labels for improved training, either at the video-level [8, 39] or segment-level [9, 17, 26, 50]. In particular, Rachavarapu et al. [26] propose prototype-based pseudo-labeling (PPL), which seeks to train a pseudo-label generation module in conjunction with a core inference module. However, due to the lack of sufficient training data, this method struggles to generalize. On the other hand, VPLAN [50], VALOR [17], and LSLD [9] leverage large-scale pre-trained foundation models, such as CLIP [28] and CLAP [40], along with ground-truth video-level labels to generate segment-level pseudo-labels for each of the two modalities. Audio/Visual segments (e.g. the audio corresponding to the segment in question and the visual frame at the center of the segment) are fed into CLAP/CLIP, one segment at a time, to generate these pseudo-labels. Despite the significant improvement that these pseudo-label generation methods achieved, the correctness of the generated labels is still limited, constrained primarily by the lack of understanding of intersegment dynamics. For instance, if a crowd is cheering in a segment of the video, it is more likely that the crowd might also be clapping right before or after." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": "To address the oversight of inter-segment dependencies and other shortcomings in existing pseudo-label generation methods, we introduce a novel, uncertainty-based, weakly-supervised, video parsing model called UWAV (Uncertainty-weighted Weakly-supervised Audiovisual Video Parsing), capable of generating improved segment-level pseudo-labels for better training of the inference module. We resort to transformer modules [38] to equip UWAV with the ability to capture temporal relationships between segments and pre-train it on a large-scale, supervised audio-visual event localization dataset [12]. Subsequently, this pre-trained model is used to generate segment-level pseudo-labels for each modality on a target, small-scale dataset which only provides weak (i.e., video-level) supervision. Such a design permits a more holis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 406 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 406 + ], + "type": "text", + "content": "tic understanding of the video, resulting in more accurate pseudo-labels. Additionally, UWAV factors in the uncertainty associated with these estimated pseudo-labels in its optimization. That uncertainty is the result of the shift in the domain of the target dataset, insufficient model capacity, etc. and is computed at training time as the confidence scores associated with these labels. To further enhance the model's ability to learn in this small-scale, weakly-supervised data regime, we also employ a feature mixup strategy. This approach adds more regularization constraints by training on mixed segment features alongside interpolated pseudo-labels, which not only lessens the influence of noise but also enriches the training data, thereby reducing overfitting. Moreover, UWAV addresses a critical class imbalance issue in the pseudo-label enriched training data, viz. most event classes in any given segment of a video are absent/negative (i.e., they do not occur), while only a handful of them do. This creates a natural bias in the training set, making it difficult to learn the positive events. To counter this, we propose a class-frequency aware weighting strategy that lays greater emphasis on the accurate classification and localization of positive events. By incorporating these components into its design, our proposed method (UWAV) outperforms competing state-of-the-art approaches across two different datasets, viz. Look, Listen, and Parse (LLP) [35] and the Audio-Visual Event Localization (AVE) [34], on multiple evaluation metrics. In summary, our contributions are the following:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 407, + 545, + 528 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 306, + 407, + 545, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 407, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 545, + 442 + ], + "type": "text", + "content": "- We introduce a novel, weakly-supervised method called UWAV, capable of synthesizing temporally coherent pseudo-labels for the AVVP task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 443, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 545, + 491 + ], + "type": "text", + "content": "- To the best of our knowledge, ours is the first method for the AVVP task, which factors in the uncertainty associated with the estimated pseudo-labels while also regularizing it with a feature mixup strategy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 491, + 545, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 545, + 528 + ], + "type": "text", + "content": "- UWAV outperforms competing state-of-the-art approaches for the AVVP task, across two different datasets on multiple metrics which attest to its generalizability." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 537, + 398, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 537, + 398, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 537, + 398, + 551 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 715 + ], + "type": "text", + "content": "Audio-Visual Learning: Audio-visual learning has emerged as an area of active research, aiming to develop models that synergistically integrate information from both audio and visual modalities for improved perception and understanding. Towards this end, various audio-visual tasks have been explored by the community so far, such as audio-visual segmentation [19, 22, 42, 49], sound source localization [15, 16, 24, 32], event localization [29, 34, 48], navigation [4-6, 21, 44, 46], generation [25, 30, 41], question answering [11, 18, 33, 47], and sound source separation [2, 7, 36, 43]. In this work, we focus on the task of audio-visual video parsing (AVVP) where the goal is to temporally localize events occurring in a video." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13562" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "Unlike many other audio-visual learning tasks, AVVP does not assume that events are always aligned across modalities. Some events could be exclusively uni-modal while others may have an audio-visual signature, which requires complex reasoning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 145, + 289, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 145, + 289, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 145, + 289, + 396 + ], + "type": "text", + "content": "Audio-Visual Video Parsing (AVVP): To address the challenges of the AVVP task [17, 26, 35], Tian et al. [35] proposed a Hybrid Attention Network (HAN) and a learnable Multi-modal Multiple Instance Learning (MMIL) pooling module. The HAN model facilitates the exchange of information within and across modalities using self-attention and cross-attention layers, while the MMIL pooling module aggregates segment-level event probabilities from both modalities to produce video-level probabilities. Building on this foundation, recent works advanced the field from the following two perspectives. The first group of studies [23, 45, 51] focuses on enhancing model architectures. In particular, Mo and Tian [23] proposed the Multimodal Grouping Network (MGN) to explicitly group semantically similar features within each modality to improve the reasoning process, while Yu et al. [45] proposed the Multi-modal Pyramid Attentional Network (MM-Pyramid) to capture events of varying durations by extracting features at multiple temporal scales. Our proposed method is orthogonal to this line of research and can be integrated with any of these backbones." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 397, + 289, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 397, + 289, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 289, + 649 + ], + "type": "text", + "content": "The second direction focuses on generating pseudo-labels for improved training, either at the video-level [8, 39] or the segment-level [9, 17, 26, 50]. VPLAN [50], VALOR [17], and LSLD [9] utilize pre-trained CLIP [28] and CLAP [40] along with ground-truth video-level labels to predict pseudo-labels for each modality on a per-segment basis. In contrast, PPL [26] uses the HAN model itself to generate pseudo-labels by constructing prototype features for each class and assigning labels to each segment based on the similarity between its feature and the prototype features. While these pseudo-label generation methods have substantially improved model performance on the AVVP task, they still exhibit some limitations. For instance, to derive accurate pseudo-labels, PPL might require a large enough training set to learn good prototype features, which might pose challenges when applied to smaller datasets. Our proposed method overcomes this problem. On the other hand, methods that leverage CLIP and CLAP to generate pseudo-labels often ignore temporal relationships between segments or the uncertainty associated with these labels. Our work also seeks to plug this void." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 657, + 133, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 133, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 133, + 670 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Problem Formulation: The AVVP task [35] aims to localize all visible and/or audible events in each one-second segment of a video. Specifically, an audible video is split" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": "into " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": " one-second segments, denoted as " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "\\{V_t, A_t\\}_{t=1}^T" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": ". Each segment is annotated with a pair of ground-truth labels " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "y_t^v \\in \\{0,1\\}^C, y_t^a \\in \\{0,1\\}^C" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "y_t^v" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": " denotes visual events, " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "y_t^a" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": " denotes audio events, and " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": " denotes the total number of events in the pre-defined event set of the data. However, owing to the weakly-supervised nature of the task setup " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "(y_t^v, y_t^a)" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": " are unavailable during training. Instead, only the modality-agnostic, video-level labels " + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "inline_equation", + "content": "y \\in \\{0,1\\}^C" + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": " are available, where 1 indicates the presence of an event at any time (either in the audio or visual stream or both) while 0 indicates an event's absence in the video." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 217, + 547, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 547, + 374 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 547, + 374 + ], + "type": "text", + "content": "Pseudo-Label Based AVVP Framework: The Hybrid Attention Network (HAN) [35] is a commonly used model for the AVVP task. The model works by first utilizing pre-trained visual and audio backbones to extract features from the visual and audio segments respectively, which are then projected to two " + }, + { + "bbox": [ + 304, + 217, + 547, + 374 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 217, + 547, + 374 + ], + "type": "text", + "content": "-dimensional feature spaces. The resulting visual segment-level features are denoted by " + }, + { + "bbox": [ + 304, + 217, + 547, + 374 + ], + "type": "inline_equation", + "content": "F^v = \\{f_t^v\\}_{t=1}^T \\in \\mathbb{R}^{T \\times d}" + }, + { + "bbox": [ + 304, + 217, + 547, + 374 + ], + "type": "text", + "content": ", while the audio segment-level features are denoted by " + }, + { + "bbox": [ + 304, + 217, + 547, + 374 + ], + "type": "inline_equation", + "content": "F^a = \\{f_t^a\\}_{t=1}^T \\in \\mathbb{R}^{T \\times d}" + }, + { + "bbox": [ + 304, + 217, + 547, + 374 + ], + "type": "text", + "content": ". These features are provided as input to the HAN model. In the model, information across segments within a modality and across modalities is exchanged through self-attention and cross-attention layers, as shown below:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 380, + 545, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 380, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 317, + 380, + 545, + 411 + ], + "type": "interline_equation", + "content": "\\tilde {f} _ {t} ^ {v} = f _ {t} ^ {v} + \\underbrace {\\operatorname {A t t n} \\left(f _ {t} ^ {v} , F ^ {v} , F ^ {v}\\right)} _ {\\text {S e l f - A t t e n t i o n}} + \\underbrace {\\operatorname {A t t n} \\left(f _ {t} ^ {v} , F ^ {a} , F ^ {a}\\right)} _ {\\text {C r o s s - A t t e n t i o n}}, \\tag {1}", + "image_path": "b9986c88b2ce65ab4ec2a1bb9bb97ebd16b355b8e34fe4a2549d09df57cfb33c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 415, + 545, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 415, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 317, + 415, + 545, + 445 + ], + "type": "interline_equation", + "content": "\\tilde {f} _ {t} ^ {a} = f _ {t} ^ {a} + \\underbrace {\\operatorname {A t t n} \\left(f _ {t} ^ {a} , F ^ {a} , F ^ {a}\\right)} _ {\\text {S e l f - A t t e n t i o n}} + \\underbrace {\\operatorname {A t t n} \\left(f _ {t} ^ {a} , F ^ {v} , F ^ {v}\\right)} _ {\\text {C r o s s - A t t e n t i o n}}, \\tag {2}", + "image_path": "d3c6325408898faa637dcdc219ce2776daa9a35eefeac7c578357df3fe4a5010.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 453, + 545, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 545, + 560 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 545, + 560 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 453, + 545, + 560 + ], + "type": "inline_equation", + "content": "\\mathrm{Attn}(Q,K,V)" + }, + { + "bbox": [ + 304, + 453, + 545, + 560 + ], + "type": "text", + "content": " denotes the standard multi-head attention mechanism [38]. Finally a classifier, shared across both modalities, transforms the visual segment-level features " + }, + { + "bbox": [ + 304, + 453, + 545, + 560 + ], + "type": "inline_equation", + "content": "\\tilde{F}^v = \\{\\tilde{f}_t^v\\}_{t=1}^T \\in \\mathbb{R}^{T \\times d}" + }, + { + "bbox": [ + 304, + 453, + 545, + 560 + ], + "type": "text", + "content": " (resp. audio segment-level features " + }, + { + "bbox": [ + 304, + 453, + 545, + 560 + ], + "type": "inline_equation", + "content": "\\tilde{F}^a = \\{f_t^a\\}_{t=1}^T" + }, + { + "bbox": [ + 304, + 453, + 545, + 560 + ], + "type": "text", + "content": " into visual segment-level logits " + }, + { + "bbox": [ + 304, + 453, + 545, + 560 + ], + "type": "inline_equation", + "content": "\\{z_t^v\\}_{t=1}^T \\in \\mathbb{R}^{T \\times C}" + }, + { + "bbox": [ + 304, + 453, + 545, + 560 + ], + "type": "text", + "content": " (resp. audio segment-level logits " + }, + { + "bbox": [ + 304, + 453, + 545, + 560 + ], + "type": "inline_equation", + "content": "\\{z_t^a\\}_{t=1}^T" + }, + { + "bbox": [ + 304, + 453, + 545, + 560 + ], + "type": "text", + "content": "). Segment-level probabilities " + }, + { + "bbox": [ + 304, + 453, + 545, + 560 + ], + "type": "inline_equation", + "content": "\\{p_t^v\\}_{t=1}^T, \\{p_t^a\\}_{t=1}^T \\in \\mathbb{R}^{T \\times C}" + }, + { + "bbox": [ + 304, + 453, + 545, + 560 + ], + "type": "text", + "content": " are then obtained by applying the sigmoid function on " + }, + { + "bbox": [ + 304, + 453, + 545, + 560 + ], + "type": "inline_equation", + "content": "\\{z_t^v\\}_{t=1}^T" + }, + { + "bbox": [ + 304, + 453, + 545, + 560 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 453, + 545, + 560 + ], + "type": "inline_equation", + "content": "\\{z_t^a\\}_{t=1}^T" + }, + { + "bbox": [ + 304, + 453, + 545, + 560 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 560, + 545, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 560, + 545, + 606 + ], + "spans": [ + { + "bbox": [ + 304, + 560, + 545, + 606 + ], + "type": "text", + "content": "Since, only video-level labels " + }, + { + "bbox": [ + 304, + 560, + 545, + 606 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 560, + 545, + 606 + ], + "type": "text", + "content": " are available during training, Tian et al. [35] introduce an attentive MMIL pooling module to learn to predict video-level probabilities " + }, + { + "bbox": [ + 304, + 560, + 545, + 606 + ], + "type": "inline_equation", + "content": "p \\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 304, + 560, + 545, + 606 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 337, + 611, + 545, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 611, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 337, + 611, + 545, + 625 + ], + "type": "interline_equation", + "content": "W _ {\\text {m o d a l}} ^ {v, a} = \\operatorname {S o f t m a x} _ {\\text {m o d a l}} \\left(\\operatorname {F C} _ {\\text {m o d a l}} \\left(\\tilde {F} ^ {v, a}\\right)\\right), \\tag {3}", + "image_path": "44c3ab8757263e0aa729ca431be54e682d8c6a4cd378a3b4ad9e951aef73b16a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 345, + 635, + 545, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 635, + 545, + 649 + ], + "spans": [ + { + "bbox": [ + 345, + 635, + 545, + 649 + ], + "type": "interline_equation", + "content": "W _ {\\text {t i m e}} ^ {v, a} = \\operatorname {S o f t m a x} _ {\\text {t i m e}} \\left(\\mathrm {F C} _ {\\text {t i m e}} \\left(\\tilde {F} ^ {v, a}\\right)\\right), \\tag {4}", + "image_path": "67867f03663d72af132ab67220287b00e24c42bcbdbd205f83ee918d383d51b4.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{FC}_{\\mathrm{mod}}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{FC}_{\\mathrm{time}}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " are two learnable fully-connected layers, " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{F}^{v,a} = \\mathrm{Stack}(\\tilde{F}^v,\\tilde{F}^a)\\in \\mathbb{R}^{2\\times T\\times d}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " denotes the stacked visual and audio features along the first dimension, " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{Softmax}_{\\mathrm{mod}}(\\cdot)" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " denotes the softmax operation along the modality dimension (i.e., across " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "v,a" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "), while" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13563" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 96, + 219, + 217 + ], + "blocks": [ + { + "bbox": [ + 50, + 76, + 165, + 95 + ], + "lines": [ + { + "bbox": [ + 50, + 76, + 165, + 95 + ], + "spans": [ + { + "bbox": [ + 50, + 76, + 165, + 95 + ], + "type": "text", + "content": "Stage 1: Training the pseudo label models on larger dataset" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 96, + 219, + 217 + ], + "lines": [ + { + "bbox": [ + 48, + 96, + 219, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 219, + 217 + ], + "type": "image", + "image_path": "029b65d7fd68f8213edf4cde2b94874146ec42b34ecad0188cda9c3d73940c33.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 230, + 77, + 445, + 217 + ], + "blocks": [ + { + "bbox": [ + 230, + 77, + 445, + 217 + ], + "lines": [ + { + "bbox": [ + 230, + 77, + 445, + 217 + ], + "spans": [ + { + "bbox": [ + 230, + 77, + 445, + 217 + ], + "type": "image", + "image_path": "a1622defeee030970a2302157682bd239910b8e4aa1869a6824d634183d9653d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 227, + 546, + 283 + ], + "lines": [ + { + "bbox": [ + 46, + 227, + 546, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 546, + 283 + ], + "type": "text", + "content": "Figure 2. UWAV framework: In stage 1, pseudo-label generation modules are equipped with the ability to capture temporal relationships between segments by pre-training on a large-scale, supervised, audio-visual event localization dataset. In stage 2, temporally coherent, uncertainty-weighted pseudo-labels, derived from the pre-trained pseudo-label generation module, are used to guide the learning of the inference model (HAN) aided by a class-balanced loss re-weighting and uncertainty-weighted feature mixup strategy. Note that we use the feature mixup strategy in both modalities while we only show the breakdown of the mixup operation for the audio modality." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 449, + 76, + 542, + 215 + ], + "blocks": [ + { + "bbox": [ + 449, + 76, + 542, + 215 + ], + "lines": [ + { + "bbox": [ + 449, + 76, + 542, + 215 + ], + "spans": [ + { + "bbox": [ + 449, + 76, + 542, + 215 + ], + "type": "image", + "image_path": "ab9adea9eb46a43f9ff1794ff0eeb3dbf4d2001c9b90e141286f7addd551c025.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 293, + 287, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 329 + ], + "type": "inline_equation", + "content": "\\operatorname{Softmax}_{time}(\\cdot)" + }, + { + "bbox": [ + 46, + 293, + 287, + 329 + ], + "type": "text", + "content": " denotes the softmax operation along the temporal dimension (i.e., across " + }, + { + "bbox": [ + 46, + 293, + 287, + 329 + ], + "type": "inline_equation", + "content": "1, \\ldots, T" + }, + { + "bbox": [ + 46, + 293, + 287, + 329 + ], + "type": "text", + "content": "). Video-level probabilities " + }, + { + "bbox": [ + 46, + 293, + 287, + 329 + ], + "type": "inline_equation", + "content": "p \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 46, + 293, + 287, + 329 + ], + "type": "text", + "content": " are then obtained via:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 81, + 331, + 287, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 331, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 81, + 331, + 287, + 365 + ], + "type": "interline_equation", + "content": "p = \\sum_ {m = \\{v, a \\}} \\sum_ {t = 1} ^ {T} W _ {\\text {m o d a l}} ^ {m, t} \\odot W _ {\\text {t i m e}} ^ {m, t} \\odot p _ {t} ^ {m}, \\tag {5}", + "image_path": "cb34bd9cf7d25a2f91c44e4c8771f95339f077a23f9ee5b3efcff60b40de991a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 371, + 287, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 287, + 421 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 371, + 287, + 421 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 371, + 287, + 421 + ], + "type": "text", + "content": " denotes the element-wise product. The HAN model is then optimized with the binary cross-entropy (BCE) loss between the estimated video-level probabilities " + }, + { + "bbox": [ + 46, + 371, + 287, + 421 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 371, + 287, + 421 + ], + "type": "text", + "content": " and video-level labels " + }, + { + "bbox": [ + 46, + 371, + 287, + 421 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 371, + 287, + 421 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 46, + 371, + 287, + 421 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{video} = \\mathrm{BCE}(p, y)" + }, + { + "bbox": [ + 46, + 371, + 287, + 421 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 429, + 164, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 429, + 164, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 429, + 164, + 443 + ], + "type": "text", + "content": "4. Proposed Approach" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 557 + ], + "type": "text", + "content": "In this section, we detail our proposed approach (UWAV). At a high level, UWAV works by generating better segment-level pseudo-labels to improve the training of a multi-modal transformer-based inference module, e.g. HAN. Moreover, UWAV factors in the uncertainty associated with these pseudo-labels, addresses the imbalance in the training data, and introduces self-supervised regularization constraints, which all lead to better performance. Figure 2 shows an overview of our proposed framework." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 563, + 285, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 285, + 576 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 285, + 576 + ], + "type": "text", + "content": "4.1. Temporally-Coherent Pseudo-Label Synthesis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": "One major issue that plagues prior works, based on pseudolabel generation [9, 17, 50], is that they do not capture the temporal dependencies between neighboring segments when generating the pseudo-labels. That is, the generated pseudo-labels are not temporally coherent. To plug this void, we propose to incorporate transformer modules [38] into the pseudo-label generation pipeline, which maps CLIP/CLAP encodings of a segment's visual frame/audio to pseudo-labels. Specifically, two separate transformers are introduced, one each for the visual/audio pseudo-label synthesis modules." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 293, + 546, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 546, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 546, + 533 + ], + "type": "text", + "content": "Pre-Training: Training transformers often requires sufficiently large training data while datasets commonly used for the weakly-supervised AVVP task are relatively small. To mitigate this challenge, we propose to first pre-train the transformer-equipped pseudo-label generation module on a large-scale, supervised, audio-visual event localization dataset - the UnAV [12] dataset. Specifically, given an audible video of duration " + }, + { + "bbox": [ + 304, + 293, + 546, + 533 + ], + "type": "inline_equation", + "content": "T'" + }, + { + "bbox": [ + 304, + 293, + 546, + 533 + ], + "type": "text", + "content": " seconds from the pre-training dataset, we split the video into " + }, + { + "bbox": [ + 304, + 293, + 546, + 533 + ], + "type": "inline_equation", + "content": "T'" + }, + { + "bbox": [ + 304, + 293, + 546, + 533 + ], + "type": "text", + "content": " one-second segments " + }, + { + "bbox": [ + 304, + 293, + 546, + 533 + ], + "type": "inline_equation", + "content": "\\{V_t', A_t'\\}_{t=1}^{T'}" + }, + { + "bbox": [ + 304, + 293, + 546, + 533 + ], + "type": "text", + "content": ", with corresponding audio-visual event labels " + }, + { + "bbox": [ + 304, + 293, + 546, + 533 + ], + "type": "inline_equation", + "content": "y_t^{av'} \\in \\{0,1\\}^{C'}" + }, + { + "bbox": [ + 304, + 293, + 546, + 533 + ], + "type": "text", + "content": ", where 1 indicates the presence of an event in both modalities and 0 its absence in at least one modality, while " + }, + { + "bbox": [ + 304, + 293, + 546, + 533 + ], + "type": "inline_equation", + "content": "C'" + }, + { + "bbox": [ + 304, + 293, + 546, + 533 + ], + "type": "text", + "content": " denotes the total number of event classes in the pre-training dataset. Next, the video frame at the temporal center of the visual segment is transformed into visual features " + }, + { + "bbox": [ + 304, + 293, + 546, + 533 + ], + "type": "inline_equation", + "content": "G_0^{v'} = \\{g_{0,t}'\\}_{t=1}^{T'} \\in \\mathbb{R}^{T' \\times d_1}" + }, + { + "bbox": [ + 304, + 293, + 546, + 533 + ], + "type": "text", + "content": " with CLIP's [28] image encoder. These features are then fed into the corresponding transformer of the visual stream, consisting of " + }, + { + "bbox": [ + 304, + 293, + 546, + 533 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 293, + 546, + 533 + ], + "type": "text", + "content": " encoder blocks, each block containing a self-attention layer, LayerNorm [1] (LN), and a 2-layer feed-forward network (FFN):" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 347, + 540, + 545, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 540, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 347, + 540, + 545, + 555 + ], + "type": "interline_equation", + "content": "\\tilde {G} _ {l} ^ {v ^ {\\prime}} = \\mathrm {L N} \\left(G _ {l} ^ {v ^ {\\prime}} + \\operatorname {A t t n} \\left(G _ {l} ^ {v ^ {\\prime}}, G _ {l} ^ {v ^ {\\prime}}, G _ {l} ^ {v ^ {\\prime}}\\right)\\right), \\tag {6}", + "image_path": "c05a508bccab26cff3dcfd6fc3701a61a71b220336dd8fabc8d668c8069e4060.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 340, + 557, + 545, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 557, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 340, + 557, + 545, + 571 + ], + "type": "interline_equation", + "content": "G _ {l + 1} ^ {v ^ {\\prime}} = \\operatorname {L N} \\left(\\tilde {G} _ {l} ^ {v ^ {\\prime}} + \\operatorname {F F N} \\left(\\tilde {G} _ {l} ^ {v ^ {\\prime}}\\right)\\right). \\tag {7}", + "image_path": "8ac943e2b0caac588f13cf48851ae13558ee66850afbb13085d2e8d2c6262183.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 546, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 546, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 546, + 693 + ], + "type": "text", + "content": "Concurrently, we convert each event category label in the pre-training dataset into a textual event feature " + }, + { + "bbox": [ + 304, + 582, + 546, + 693 + ], + "type": "inline_equation", + "content": "e_c^{CLIP'} \\in \\mathbb{R}^{d_1}" + }, + { + "bbox": [ + 304, + 582, + 546, + 693 + ], + "type": "text", + "content": " by filling in the pre-defined caption template: \"A photo of \" with the corresponding event name and passing it to CLIP's text encoder. Equipped with the visual segment-level features " + }, + { + "bbox": [ + 304, + 582, + 546, + 693 + ], + "type": "inline_equation", + "content": "G_L^{v'} = \\{g_{L,t}'\\}_{t=1}^{T'} \\in \\mathbb{R}^{T' \\times d_1}" + }, + { + "bbox": [ + 304, + 582, + 546, + 693 + ], + "type": "text", + "content": " and the textual event features " + }, + { + "bbox": [ + 304, + 582, + 546, + 693 + ], + "type": "inline_equation", + "content": "E^{CLIP'} = \\{e_c^{CLIP'}\\}_{c=1}^{C'} \\in \\mathbb{R}^{C' \\times d_1}" + }, + { + "bbox": [ + 304, + 582, + 546, + 693 + ], + "type": "text", + "content": ", we derive visual segment-level logits " + }, + { + "bbox": [ + 304, + 582, + 546, + 693 + ], + "type": "inline_equation", + "content": "\\hat{z}_t^{v'} \\in \\mathbb{R}^{C'}" + }, + { + "bbox": [ + 304, + 582, + 546, + 693 + ], + "type": "text", + "content": " and probabilities " + }, + { + "bbox": [ + 304, + 582, + 546, + 693 + ], + "type": "inline_equation", + "content": "\\hat{p}_t^{v'}" + }, + { + "bbox": [ + 304, + 582, + 546, + 693 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 336, + 699, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 699, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 336, + 699, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\hat {p} _ {t} ^ {v ^ {\\prime}} = \\operatorname {S i g m o i d} \\left(\\hat {z} _ {t} ^ {v ^ {\\prime}}\\right), \\hat {z} _ {t} ^ {v ^ {\\prime}} = E ^ {C L I P ^ {\\prime}} \\cdot g _ {L, t} ^ {v ^ {\\prime} \\top}. \\tag {8}", + "image_path": "ce612c17c5efd20258941d53514edbeda557766bce3313d331e5991e94deb0e0.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13564" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 205 + ], + "type": "text", + "content": "Similar operations are performed in the audio pseudolabel generation pipeline. The raw waveforms corresponding to the 1-second audio segments are transformed into audio features " + }, + { + "bbox": [ + 47, + 72, + 287, + 205 + ], + "type": "inline_equation", + "content": "G_0^{a'} \\in \\mathbb{R}^{T' \\times d_2}" + }, + { + "bbox": [ + 47, + 72, + 287, + 205 + ], + "type": "text", + "content": " with CLAP's [12] audio encoder and fed into the corresponding transformer consisting of " + }, + { + "bbox": [ + 47, + 72, + 287, + 205 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 72, + 287, + 205 + ], + "type": "text", + "content": " encoder blocks. Correspondingly, the textual event features " + }, + { + "bbox": [ + 47, + 72, + 287, + 205 + ], + "type": "inline_equation", + "content": "E^{CLAP'} \\in \\mathbb{R}^{C' \\times d_2}" + }, + { + "bbox": [ + 47, + 72, + 287, + 205 + ], + "type": "text", + "content": " are generated with the caption template: \"This is the sound of \" by passing it through CLAP's text encoder. Audio segment-level logits " + }, + { + "bbox": [ + 47, + 72, + 287, + 205 + ], + "type": "inline_equation", + "content": "\\hat{z}_t^{a'} \\in \\mathbb{R}^{C'}" + }, + { + "bbox": [ + 47, + 72, + 287, + 205 + ], + "type": "text", + "content": " and probabilities " + }, + { + "bbox": [ + 47, + 72, + 287, + 205 + ], + "type": "inline_equation", + "content": "\\hat{p}_t^{a'}" + }, + { + "bbox": [ + 47, + 72, + 287, + 205 + ], + "type": "text", + "content": " can then be derived in the same manner: " + }, + { + "bbox": [ + 47, + 72, + 287, + 205 + ], + "type": "inline_equation", + "content": "\\hat{p}_t^{a'} = \\text{Sigmoid}(\\hat{z}_t^{a'})" + }, + { + "bbox": [ + 47, + 72, + 287, + 205 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 72, + 287, + 205 + ], + "type": "inline_equation", + "content": "\\hat{z}_t^{a'} = E^{CLAP'} \\cdot g_t^{a'}^\\top" + }, + { + "bbox": [ + 47, + 72, + 287, + 205 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 205, + 287, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 205, + 287, + 276 + ], + "spans": [ + { + "bbox": [ + 47, + 205, + 287, + 276 + ], + "type": "text", + "content": "Since the events occurring in the pre-training dataset (UnAV) are audio-visual, we multiply the segment-level visual and audio event probabilities to enforce the predicted labels to be multi-modal in nature: " + }, + { + "bbox": [ + 47, + 205, + 287, + 276 + ], + "type": "inline_equation", + "content": "\\{\\hat{p}_t^{av'}\\}_{t=1}^{T'} \\in \\mathbb{R}^{T' \\times C'}" + }, + { + "bbox": [ + 47, + 205, + 287, + 276 + ], + "type": "text", + "content": ". This network is then trained with the binary cross-entropy (BCE) loss:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 74, + 282, + 287, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 282, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 74, + 282, + 287, + 297 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t e m p}} = \\mathrm {B C E} \\left(\\hat {p} _ {t} ^ {a v ^ {\\prime}}, y _ {t} ^ {a v ^ {\\prime}}\\right), \\hat {p} _ {t} ^ {a v ^ {\\prime}} = \\tilde {p} _ {t} ^ {v ^ {\\prime}} \\odot \\hat {p} _ {t} ^ {a ^ {\\prime}}. \\tag {9}", + "image_path": "38279722e3697f1dd5368333d60e3abb25ec3dbd890afd78e1bc45e029db5272.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 303, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 303, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 47, + 303, + 287, + 496 + ], + "type": "text", + "content": "Pseudo-Label Generation on Target Dataset: With the pre-trained pseudo-label generation modules in place, we proceed to employ them for the pseudo-label generation process in the target dataset for the AVVP task. Specifically, the center frame of each of the visual segments " + }, + { + "bbox": [ + 47, + 303, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\{V_t\\}_{t=1}^T" + }, + { + "bbox": [ + 47, + 303, + 287, + 496 + ], + "type": "text", + "content": " of the target dataset are passed into CLIP's image encoder, whose output is passed into the pre-trained visual transformer to generate segment features " + }, + { + "bbox": [ + 47, + 303, + 287, + 496 + ], + "type": "inline_equation", + "content": "G_L^v = \\{g_{L,t}^v\\}_{t=1}^T \\in \\mathbb{R}^{T \\times d_1}" + }, + { + "bbox": [ + 47, + 303, + 287, + 496 + ], + "type": "text", + "content": ". At the same time, the caption template: \"A photo of \" is used to obtain textual features corresponding to each of the event classes in the target dataset for the AVVP task: " + }, + { + "bbox": [ + 47, + 303, + 287, + 496 + ], + "type": "inline_equation", + "content": "E^{CLIP} \\in \\mathbb{R}^{C \\times d_1}" + }, + { + "bbox": [ + 47, + 303, + 287, + 496 + ], + "type": "text", + "content": ". Segment-level visual logits " + }, + { + "bbox": [ + 47, + 303, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\hat{z}_t^v \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 47, + 303, + 287, + 496 + ], + "type": "text", + "content": " can be derived by computing their inner product. We also pre-define class-wise visual thresholds " + }, + { + "bbox": [ + 47, + 303, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\theta^v \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 47, + 303, + 287, + 496 + ], + "type": "text", + "content": " to transform segment-level visual logits into binary pseudo-labels " + }, + { + "bbox": [ + 47, + 303, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\hat{y}_t^v \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 47, + 303, + 287, + 496 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 501, + 287, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 501, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 82, + 501, + 287, + 517 + ], + "type": "interline_equation", + "content": "\\hat {y} _ {t} ^ {v} = \\mathbb {1} _ {\\left\\{\\hat {z} _ {t} ^ {v} > \\theta^ {v} \\right\\}} \\odot y, \\hat {z} _ {t} ^ {v} = E ^ {C L I P} \\cdot g _ {L, t} ^ {v} ^ {\\top}, \\tag {10}", + "image_path": "57fe2782ed516bb2c7c7379415708bbbfae0767f0c208ed609bb6d3615c9a527.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 522, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 522, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 522, + 287, + 582 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 522, + 287, + 582 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 47, + 522, + 287, + 582 + ], + "type": "text", + "content": " denotes the ground-truth video-level labels, " + }, + { + "bbox": [ + 47, + 522, + 287, + 582 + ], + "type": "inline_equation", + "content": "\\mathbb{1}_{\\{\\cdot\\}}" + }, + { + "bbox": [ + 47, + 522, + 287, + 582 + ], + "type": "text", + "content": " is the indicator function which returns a value of 1 when the condition is true otherwise 0, and " + }, + { + "bbox": [ + 47, + 522, + 287, + 582 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 47, + 522, + 287, + 582 + ], + "type": "text", + "content": " denotes the elementwise product operation. The " + }, + { + "bbox": [ + 47, + 522, + 287, + 582 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 47, + 522, + 287, + 582 + ], + "type": "text", + "content": " operation zeroes out the predictions of event classes absent in the video-level label." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 582, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 287, + 689 + ], + "type": "text", + "content": "A similar pseudo-label generation process is employed on the acoustic side. Raw waveforms of audio segments are first fed into CLAP's audio encoder and then into the pretrained audio transformer. The event names of the classes in the target dataset for the AVVP task are filled in the caption template: \"This is the sound of \" to generate textual event features: " + }, + { + "bbox": [ + 47, + 582, + 287, + 689 + ], + "type": "inline_equation", + "content": "E^{CLAP} \\in \\mathbb{R}^{C \\times d_1}" + }, + { + "bbox": [ + 47, + 582, + 287, + 689 + ], + "type": "text", + "content": ". Segment-level audio logits " + }, + { + "bbox": [ + 47, + 582, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\hat{z}_t^a \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 47, + 582, + 287, + 689 + ], + "type": "text", + "content": " and binary pseudo-labels " + }, + { + "bbox": [ + 47, + 582, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\hat{y}_t^a \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 47, + 582, + 287, + 689 + ], + "type": "text", + "content": " are then derived using class-wise thresholds " + }, + { + "bbox": [ + 47, + 582, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\theta^a \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 47, + 582, + 287, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "With binary segment-level pseudo-labels for both modalities " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\hat{y}_t^v, \\hat{y}_t^a" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " and the predicted probabilities from the infer" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "ence module (HAN) " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "\\hat{p}_t^v,\\hat{p}_t^a" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": " in place, the inference module can be trained using the binary cross-entropy loss as shown:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 346, + 105, + 545, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 105, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 346, + 105, + 545, + 118 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {h a r d} = \\mathrm {B C E} \\left(p _ {t} ^ {v}, \\hat {y} _ {t} ^ {v}\\right) + \\mathrm {B C E} \\left(p _ {t} ^ {a}, \\hat {y} _ {t} ^ {a}\\right). \\tag {11}", + "image_path": "72ea79e0635416e4843f5784df54cee0e1eafeba73a1301046ca272186dcf2e9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 125, + 518, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 125, + 518, + 139 + ], + "spans": [ + { + "bbox": [ + 305, + 125, + 518, + 139 + ], + "type": "text", + "content": "4.2. Training with Pseudo-Label Uncertainty" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 144, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 545, + 300 + ], + "type": "text", + "content": "While pseudo-labels do provide additional supervision for better training of the inference module, they could potentially be noisy, leading to occasionally incorrect training signals. To ameliorate this problem, we propose an uncertainty-weighted pseudo-label training scheme to improve the robustness of the learning process. Instead of simply training with the binary pseudo-labels " + }, + { + "bbox": [ + 304, + 144, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\hat{y}_t^v,\\hat{y}_t^a" + }, + { + "bbox": [ + 304, + 144, + 545, + 300 + ], + "type": "text", + "content": " , we leverage the confidence of the pseudo-label estimation module (associated with the predicted pseudo-label) to weigh the training signal for the inference module. This confidence score serves as a measure of the pseudo-label generation module's uncertainty of its prediction. This may be represented as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 306, + 545, + 330 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 306, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 311, + 306, + 545, + 330 + ], + "type": "interline_equation", + "content": "\\hat {p} _ {t} ^ {v} = \\operatorname {S i g m o i d} \\left(\\hat {z} _ {t} ^ {v} - \\theta^ {v}\\right) \\odot y; \\quad \\hat {p} _ {t} ^ {a} = \\operatorname {S i g m o i d} \\left(\\hat {z} _ {t} ^ {a} - \\theta^ {a}\\right) \\odot y. \\tag {12}", + "image_path": "ff5d05d75929f4666b3549d586d16b752fc88d71460829b37f5301888a0883e7.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 331, + 545, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 473 + ], + "type": "text", + "content": "In other words, considering the visual pseudo-label generation pipeline as an example, the farther the logit " + }, + { + "bbox": [ + 304, + 331, + 545, + 473 + ], + "type": "inline_equation", + "content": "\\hat{z}_t^v" + }, + { + "bbox": [ + 304, + 331, + 545, + 473 + ], + "type": "text", + "content": " is from the threshold " + }, + { + "bbox": [ + 304, + 331, + 545, + 473 + ], + "type": "inline_equation", + "content": "\\theta^v" + }, + { + "bbox": [ + 304, + 331, + 545, + 473 + ], + "type": "text", + "content": ", whether much lower or higher, the more confident the pseudo-label generation module is about the label it predicts (either approaching 0 or 1). Conversely, the closer the logit is to the threshold, the less the certainty about the correctness of the pseudo-labels (probabilities closer to 0.5). An analogous explanation also holds for the audio pseudo-labels. With the uncertainty-weighted pseudo-labels in place, the inference module (HAN) can be trained with the following uncertainty-weighted pseudo-label loss:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 347, + 479, + 545, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 479, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 347, + 479, + 545, + 493 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s o f t}} = \\mathrm {B C E} \\left(p _ {t} ^ {v}, \\hat {p} _ {t} ^ {v}\\right) + \\mathrm {B C E} \\left(p _ {t} ^ {a}, \\hat {p} _ {t} ^ {a}\\right). \\tag {13}", + "image_path": "2b8fddda03dd8d447ac6166830878c20aae834c3a1242241662c0512368a149c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 503, + 503, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 503, + 503, + 517 + ], + "spans": [ + { + "bbox": [ + 305, + 503, + 503, + 517 + ], + "type": "text", + "content": "4.3. Uncertainty-weighted Feature Mixup" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": "Due to the lack of full supervision for the weakly-supervised AVVP task, we explore the efficacy of additional regularization via self-supervision to help the models generalize better. Towards this end, prior pseudo-label generation approaches [26, 39] often employ contrastive learning as a tool to better train the inference module. However, due to the inherent noise in the estimated pseudo-labels, positive samples and negative samples may be mislabeled, decreasing the effectiveness of the self-supervisory training. As an alternative, in this work, we explore the effectiveness of feature mixing, as a self-supervisory training signal for additional regularization. In this setting, we mixup the estimated features of any two segments, additively, and train the model to predict the union of the labels of the two segments. However, since the labels in our setting are noisy, the mixed feature is assigned a label derived from a weighted sum of" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13565" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "the uncertainty-weighted pseudo-labels of each of the two segment features. This is illustrated below:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 102, + 286, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 102, + 286, + 118 + ], + "spans": [ + { + "bbox": [ + 53, + 102, + 286, + 118 + ], + "type": "interline_equation", + "content": "\\bar {f} _ {t _ {i}, t _ {j}} ^ {v} = \\lambda \\tilde {f} _ {t _ {i}} ^ {v} + (1 - \\lambda) \\tilde {f} _ {t _ {j}} ^ {v}, \\quad \\bar {p} _ {t _ {i}, t _ {j}} ^ {v} = \\lambda \\hat {p} _ {t _ {i}} ^ {v} + (1 - \\lambda) \\hat {p} _ {t _ {j}} ^ {v} \\tag {14}", + "image_path": "cf330da62b3fc6c9094e7ab89bfb1759a0f277e62177da877d09318dd7132185.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 120, + 286, + 137 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 120, + 286, + 137 + ], + "spans": [ + { + "bbox": [ + 54, + 120, + 286, + 137 + ], + "type": "interline_equation", + "content": "\\bar {f} _ {t _ {i}, t _ {j}} ^ {a} = \\lambda \\tilde {f} _ {t _ {i}} ^ {a} + (1 - \\lambda) \\tilde {f} _ {t _ {j}} ^ {a}, \\quad \\bar {p} _ {t _ {i}, t _ {j}} ^ {a} = \\lambda \\hat {p} _ {t _ {i}} ^ {a} + (1 - \\lambda) \\hat {p} _ {t _ {j}} ^ {a}, \\tag {15}", + "image_path": "b4ed7e8ffea7b31c7514effe5fe2f081ce0996b93830cb82b0d308fc7e525bb9.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 141, + 287, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 141, + 287, + 177 + ], + "spans": [ + { + "bbox": [ + 47, + 141, + 287, + 177 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 141, + 287, + 177 + ], + "type": "inline_equation", + "content": "\\lambda \\sim \\mathrm{Beta}(\\alpha, \\alpha)" + }, + { + "bbox": [ + 47, + 141, + 287, + 177 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 141, + 287, + 177 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 141, + 287, + 177 + ], + "type": "text", + "content": " is a hyper-parameter controlling the Beta distribution, and " + }, + { + "bbox": [ + 47, + 141, + 287, + 177 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 47, + 141, + 287, + 177 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 141, + 287, + 177 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 47, + 141, + 287, + 177 + ], + "type": "text", + "content": " indicate two segment indices in a batch of video segments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 177, + 287, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 177, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 47, + 177, + 287, + 249 + ], + "type": "text", + "content": "After mixing the uni-modal segment-level features, we pass them through the classifier of the inference module and apply the sigmoid function to the output, obtaining mixed segment-level event probabilities " + }, + { + "bbox": [ + 47, + 177, + 287, + 249 + ], + "type": "inline_equation", + "content": "p_t^{mix - v}" + }, + { + "bbox": [ + 47, + 177, + 287, + 249 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 177, + 287, + 249 + ], + "type": "inline_equation", + "content": "p_t^{mix - a}" + }, + { + "bbox": [ + 47, + 177, + 287, + 249 + ], + "type": "text", + "content": ". These are used to train the inference model with the uncertainty-aware mixup loss, as shown below:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 60, + 255, + 287, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 255, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 60, + 255, + 287, + 270 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m i x} = \\mathrm {B C E} \\left(p _ {t} ^ {m i x - v}, \\bar {p} _ {t} ^ {v}\\right) + \\mathrm {B C E} \\left(p _ {t} ^ {m i x - a}, \\bar {p} _ {t} ^ {a}\\right). \\tag {16}", + "image_path": "689cf393d4aec8e5397e34ef2d7a03c69c12267313afd11ad6bfad44a68fa828.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 276, + 230, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 276, + 230, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 276, + 230, + 289 + ], + "type": "text", + "content": "4.4. Class-balanced Loss Re-weighting" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 294, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 287, + 510 + ], + "type": "text", + "content": "Besides the aforementioned challenges of the AVVP task, most of the events in the event set are absent in the pseudolabels of any (segment of a) video (i.e., most event classes are negative events) and only a few events are present (i.e., positive events are much fewer in number). As a result, the model is dominated by the loss from the negative events. When trained without factoring in this bias, the classifier tends to overfit the negative labels and ignore the positive ones. To address this class imbalance issue, we introduce a class-balanced loss re-weighting strategy to re-balance the importance of the losses from the negative and positive events for the uncertainty-weighted pseudo-label loss. Specifically, the loss from the positive events is multiplied by a weight proportional to the frequency of the segments with the negative events in the pseudo-labels, while the loss from the negative events is multiplied by a weight proportional to the frequency of the segments with the positive events in the pseudo-labels, as shown below:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 514, + 287, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 287, + 552 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {w - s o f t} = \\sum_ {m \\in \\{v, a \\}} w _ {p o s} ^ {m} \\cdot y \\cdot \\mathrm {B C E} \\left(p _ {t} ^ {m}, \\hat {p} _ {t} ^ {m}\\right) + \\\\ w _ {n e g} ^ {m} \\cdot (1 - y) \\cdot \\mathrm {B C E} \\left(p _ {t} ^ {m}, \\hat {p} _ {t} ^ {m}\\right), \\tag {17} \\\\ \\end{array}", + "image_path": "9906c1cc49795834e52220b11df5a30efc6d5e6ea9655dba8e31a6b6b65cecdd.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 559, + 287, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 559, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 67, + 559, + 287, + 586 + ], + "type": "interline_equation", + "content": "w _ {p o s} ^ {m} = \\frac {\\sum_ {i = 1} ^ {N} \\sum_ {t = 1} ^ {T} \\sum_ {c = 1} ^ {C} \\left(1 - \\hat {y} _ {i , t , c} ^ {m}\\right)}{N T C} \\times W, \\tag {18}", + "image_path": "45606bb7f12d939fa3cf74f460d6a95c398a418580eb84329f08acb4815cdf44.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 587, + 287, + 615 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 587, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 67, + 587, + 287, + 615 + ], + "type": "interline_equation", + "content": "w _ {n e g} ^ {m} = \\frac {\\sum_ {i = 1} ^ {N} \\sum_ {t = 1} ^ {T} \\sum_ {c = 1} ^ {C} \\hat {y} _ {i , t , c} ^ {m}}{N T C}, \\tag {19}", + "image_path": "ed0f58379a67af86c02cea19295374af13bc4dff9807022b4ccbd4cbaa88539f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 620, + 286, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 286, + 644 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 286, + 644 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 620, + 286, + 644 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 620, + 286, + 644 + ], + "type": "text", + "content": " denotes the number of videos in the training set, and " + }, + { + "bbox": [ + 47, + 620, + 286, + 644 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 47, + 620, + 286, + 644 + ], + "type": "text", + "content": " is a hyper-parameter." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 644, + 287, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 644, + 287, + 702 + ], + "spans": [ + { + "bbox": [ + 47, + 644, + 287, + 702 + ], + "type": "text", + "content": "In summary, the inference module is trained on the AVVP task with the proposed class-balanced re-weighting, applied to the uncertainty-weighted classification loss, and the uncertainty-weighted feature mixup loss, as shown below:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 91, + 702, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 702, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 91, + 702, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {w - s o f t} + \\mathcal {L} _ {\\text {m i x}} + \\mathcal {L} _ {\\text {v i d e o}}. \\tag {20}", + "image_path": "a7e5178eaceee1721ab7cffa44a95761e9fa9663cce309771b37fc0015cda54f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 72, + 387, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 387, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 387, + 85 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 91, + 545, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 199 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 199 + ], + "type": "text", + "content": "We assess the performance of UWAV empirically across two challenging datasets and report its performance, comparing it with existing state-of-the-art approaches both quantitatively and qualitatively. Additionally, through multiple ablation studies, we bring out the effectiveness of the different elements of our proposed approach and the choices of different hyper-parameters. For additional details, ablation studies, and more qualitative results, we refer the reader to our supplementary material." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 206, + 422, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 206, + 422, + 220 + ], + "spans": [ + { + "bbox": [ + 305, + 206, + 422, + 220 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 225, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 225, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 225, + 545, + 548 + ], + "type": "text", + "content": "Datasets: We evaluate all competing methods on the Look, Listen, and Parse (LLP) dataset [35], which is the principal benchmark dataset for the AVVP task. The dataset consists of 11,849 video clips sourced from YouTube. Each clip is 10 seconds long and represents one or more of 25 diverse event categories, such as human activities, animals, musical instruments, and vehicles. The dataset is split into training, validation, and testing sets, following the official split [35]: 10,000 videos for training, 649 videos for validation, and 1,200 videos for testing. While the training set of this dataset is only associated with video-level labels of the events, the validation and testing split is labeled with segment-level event labels for evaluation purposes. Additionally, to demonstrate the generalizability of our method, we conduct a similar study on the Audio Visual Event (AVE) recognition dataset [34]. The AVE dataset consists of 4,143 video clips crawled from YouTube, each 10 seconds long. It is split into 3,339 videos for training, 402 for validation, and 402 for testing. It includes 29 event categories encompassing human activities, animals, musical instruments, vehicles, and a \"background\" class (i.e., no event occurs). Unlike the LLP dataset, each video in the AVE dataset contains only one audio-visual event. Here too, the training data is only provided with video-level labels while the validation and test splits are annotated with ground-truth event labels for each one-second segment, which either is a specific audio-visual event or \"background\"." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "content": "Metrics: For the LLP dataset, following the standard evaluation protocol [35], all models are evaluated using macro F1-scores calculated for the following event types: (i) audio-only (A), (ii) visual-only (V), and (iii) audio-visual (AV). Type@AV (Type) and Event@AV (Event) are two additional metrics that evaluate the overall performance of the model, where Type@AV is the mean of the F1-scores for the A, V, and AV events, while Event@AV is the F1-score of all events regardless of the modality in which they occur. Evaluations are conducted at both the segment-level and the event-level. At the segment-level, the model's predictions are compared with the ground truth on a per-segment basis. At the event-level, consecutive positive segments for" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13566" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 99, + 85, + 492, + 273 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 539, + 82 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 539, + 82 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 539, + 82 + ], + "type": "text", + "content": "Table 1. Comparison with state-of-the-arts methods on the LLP dataset. Best performances are in bold, second-best in underlined." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 99, + 85, + 492, + 273 + ], + "lines": [ + { + "bbox": [ + 99, + 85, + 492, + 273 + ], + "spans": [ + { + "bbox": [ + 99, + 85, + 492, + 273 + ], + "type": "table", + "html": "
MethodSegment-levelEvent-level
AVAVTypeEventAVAVTypeEvent
HAN [35]60.152.948.954.055.451.348.943.047.748.0
MA [39]60.360.055.158.957.953.656.449.053.050.6
JoMoLD [8]61.363.857.260.859.953.959.949.654.552.5
CMPAE [10]64.266.459.263.362.856.663.751.857.455.7
PoiBin [27]63.163.557.761.460.654.160.351.555.252.3
VPLAN [50]60.564.858.361.259.451.461.551.254.750.8
VALOR [17]61.865.958.462.061.555.462.652.256.754.2
LSLD [9]62.767.159.463.162.255.764.352.657.655.2
PPL [26]65.966.761.964.863.757.364.354.359.957.9
CoLeaf [31]64.264.459.362.662.557.663.254.257.955.6
LEAP [51]62.765.659.362.561.856.463.154.157.855.0
UWAV (Ours)64.270.063.465.963.958.666.757.560.957.4
", + "image_path": "a57e368bf5bb9deff42234c3242042e0a6367c55db090aa20d4068a02d8e43f2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 283, + 287, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 287, + 355 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 287, + 355 + ], + "type": "text", + "content": "the same event are grouped together as a single event. The F1-score is then computed using a mIoU threshold of 0.5. For the AVE dataset, we follow Tian et al. [34] and use accuracy as the evaluation metric. An event prediction of a segment is considered correct if it matches the ground-truth label for that segment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 360, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 360, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 287, + 589 + ], + "type": "text", + "content": "Implementation Details: In line with prior work [35], each 10-second video in both datasets is split into 10 segments of one second each, where each segment contains 8 frames. For the LLP dataset, pre-trained ResNet-152 [13] and " + }, + { + "bbox": [ + 46, + 360, + 287, + 589 + ], + "type": "inline_equation", + "content": "\\mathrm{R}(2 + 1)\\mathrm{D - }18" + }, + { + "bbox": [ + 46, + 360, + 287, + 589 + ], + "type": "text", + "content": " [37] are used to extract 2D and 3D visual features, respectively. The pre-trained VGGish [14] network is used to extract features from the audio, sampled at " + }, + { + "bbox": [ + 46, + 360, + 287, + 589 + ], + "type": "inline_equation", + "content": "16\\mathrm{KHz}" + }, + { + "bbox": [ + 46, + 360, + 287, + 589 + ], + "type": "text", + "content": ". For the AVE dataset however, akin to prior work [17], we extract visual features from pre-trained CLIP and " + }, + { + "bbox": [ + 46, + 360, + 287, + 589 + ], + "type": "inline_equation", + "content": "\\mathrm{R}(2 + 1)\\mathrm{D}" + }, + { + "bbox": [ + 46, + 360, + 287, + 589 + ], + "type": "text", + "content": " while CLAP is used to embed the audio stream. For both datasets, we set the number of encoder blocks " + }, + { + "bbox": [ + 46, + 360, + 287, + 589 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 360, + 287, + 589 + ], + "type": "text", + "content": " in the temporal-aware model to 5, " + }, + { + "bbox": [ + 46, + 360, + 287, + 589 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 360, + 287, + 589 + ], + "type": "text", + "content": " for the Beta distribution in the feature mixup strategy to 1.7, and " + }, + { + "bbox": [ + 46, + 360, + 287, + 589 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 360, + 287, + 589 + ], + "type": "text", + "content": " in the class-balanced loss re-weighting step to 0.5. The pseudo-label generation modules and the inference model (HAN) are trained with the AdamW optimizer [20]. For improved performance on the AVE dataset, we replace " + }, + { + "bbox": [ + 46, + 360, + 287, + 589 + ], + "type": "inline_equation", + "content": "\\hat{p}_{t_i}^a,\\hat{p}_{t_j}^a" + }, + { + "bbox": [ + 46, + 360, + 287, + 589 + ], + "type": "text", + "content": " in Eq. 15 with " + }, + { + "bbox": [ + 46, + 360, + 287, + 589 + ], + "type": "inline_equation", + "content": "[\\hat{p}_{t_i}^a ],[\\hat{y}_{t_j}^a ]" + }, + { + "bbox": [ + 46, + 360, + 287, + 589 + ], + "type": "text", + "content": " and make corresponding modifications in the visual counterparts as well." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "text", + "content": "Baselines: We demonstrate the effectiveness of UWAV by comparing against an extensive set of baselines. For the LLP dataset, this includes video-level pseudo-label generation methods (MA [39], JoMoLD [8]), segment-level pseudo-label generation methods (VALOR [17], LSLD [9], PPL [26]), and the recently released works (CoLeaf [31], LEAP [51]). On the other hand, for the AVE dataset, baseline approaches with publicly available implementation, which use the state-of-the-art feature backbones (akin to ours), such as HAN [35] and VALOR [17] are used." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 282, + 362, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 282, + 362, + 294 + ], + "spans": [ + { + "bbox": [ + 306, + 282, + 362, + 294 + ], + "type": "text", + "content": "5.2. Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 301, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 545, + 493 + ], + "type": "text", + "content": "Comparison with Previous Methods on LLP: As shown in Table 1, UWAV surpasses previous methods, across almost all metrics. Notably, we achieve an F-score of 70.0 on the segment-level visual event, 65.9 on the segment-level Type@AV, and 66.7 on the event-level visual event metric. This corresponds to a gain of " + }, + { + "bbox": [ + 304, + 301, + 545, + 493 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 304, + 301, + 545, + 493 + ], + "type": "text", + "content": " on segment-level Type@AV F-score and a " + }, + { + "bbox": [ + 304, + 301, + 545, + 493 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 301, + 545, + 493 + ], + "type": "text", + "content": " improvement on event-level Type@AV F-score, over our closest competitor PPL [26]. Of particular note, is the fact that our segment-level and event-level F-scores improve by more than " + }, + { + "bbox": [ + 304, + 301, + 545, + 493 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 304, + 301, + 545, + 493 + ], + "type": "text", + "content": ", over PPL, for visual events. When compared to other recently published works, such as VALOR [17], CoLeaf [31] and LEAP [51], UWAV outperforms them by up to " + }, + { + "bbox": [ + 304, + 301, + 545, + 493 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 304, + 301, + 545, + 493 + ], + "type": "text", + "content": " on both segment and event-level Type@AV F-score while gains on visual events are up to " + }, + { + "bbox": [ + 304, + 301, + 545, + 493 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 301, + 545, + 493 + ], + "type": "text", + "content": " on segment-level F-score." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 495, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 662 + ], + "type": "text", + "content": "These observations are consistent with our qualitative comparisons, as well. In the example on the left of Figure 3, our model successfully recognizes and temporally localizes the lawn mower event visually, whereas VALOR [17] (a recent state-of-the-art approach with publicly available implementation) misclassifies it as a chainsaw. Additionally, our model also accurately localize the intermittent sound of the lawn mower. In contrast, VALOR not only misclassifies the sound of the lawn mower as that of a chainsaw but also incorrectly predicts that someone is talking in the video. In the example on the right of Figure 3, our model does not err in recognizing either the visual presence or the audio presence of the telephone, while VALOR fails to accurately predict the events in either modality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 677, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 715 + ], + "type": "text", + "content": "Comparison with Previous Methods on AVE: To demonstrate the generalizability of our method, we evaluate UWAV on the AVE [34] dataset and compare its per" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13567" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 71, + 309, + 231 + ], + "blocks": [ + { + "bbox": [ + 73, + 71, + 309, + 231 + ], + "lines": [ + { + "bbox": [ + 73, + 71, + 309, + 231 + ], + "spans": [ + { + "bbox": [ + 73, + 71, + 309, + 231 + ], + "type": "image", + "image_path": "884745ea97cc18cb7ad4f14e8126d44f7e5060bd4eeb05af4c5ba4c0e7dfdbad.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 239, + 533, + 250 + ], + "lines": [ + { + "bbox": [ + 58, + 239, + 533, + 250 + ], + "spans": [ + { + "bbox": [ + 58, + 239, + 533, + 250 + ], + "type": "text", + "content": "Figure 3. Comparison between predictions by UWAV and competing AVVP methods on the LLP dataset. \"GT\": ground truth." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 511, + 230 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 511, + 230 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 511, + 230 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 511, + 230 + ], + "type": "image", + "image_path": "7bb483f21c0388ed8ddd8e93910587c4b0b02a5c664f4c925aacace79b49e052.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 59, + 285, + 271, + 315 + ], + "blocks": [ + { + "bbox": [ + 47, + 259, + 287, + 281 + ], + "lines": [ + { + "bbox": [ + 47, + 259, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 259, + 287, + 281 + ], + "type": "text", + "content": "Table 2. Model performances on the AVE dataset. CLIP, R(2+1)D-18, and CLAP are used as feature backbones." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 285, + 271, + 315 + ], + "lines": [ + { + "bbox": [ + 59, + 285, + 271, + 315 + ], + "spans": [ + { + "bbox": [ + 59, + 285, + 271, + 315 + ], + "type": "table", + "html": "
MethodHAN [35]VALOR [17]UWAV (Ours)
Acc.(%)75.380.480.6
", + "image_path": "0ca1b2f5de95940b4a632767f2a71890cfbc3579ee1bd04dd870f564a2d7b875.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 60, + 335, + 272, + 392 + ], + "blocks": [ + { + "bbox": [ + 55, + 321, + 279, + 332 + ], + "lines": [ + { + "bbox": [ + 55, + 321, + 279, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 321, + 279, + 332 + ], + "type": "text", + "content": "Table 3. Accuracy of the generated pseudo-labels on LLP." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 335, + 272, + 392 + ], + "lines": [ + { + "bbox": [ + 60, + 335, + 272, + 392 + ], + "spans": [ + { + "bbox": [ + 60, + 335, + 272, + 392 + ], + "type": "table", + "html": "
MethodSegment-level
AVAVTypeEvent
VALOR [17]80.561.755.766.074.6
PPL [26]61.761.857.560.659.4
UWAV (Ours)78.474.565.572.878.4
", + "image_path": "872710bc908d3a91b1bc115e3babd79d766cee77fdde393b417ce71c1ab65eec.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 408, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 408, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 287, + 456 + ], + "type": "text", + "content": "formance with that of previous works. From Table 2, we observe that with the same backbone features, UWAV surpasses VALOR, our closest competitor, even on this small-scale dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 469, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 287, + 601 + ], + "type": "text", + "content": "Accuracy of the Generated Pseudo-Labels: To evaluate the efficacy of our pseudo-label generation pipeline, we compare the accuracy of our generated pseudo-labels against those obtained from competing methods (with publicly available implementation) [17, 26] on the test set of the LLP dataset. As shown in Table 3, our pre-trained temporally-dependent pseudo-label generation scheme generates more accurate segment-level pseudo-labels than previous methods, by up to " + }, + { + "bbox": [ + 46, + 469, + 287, + 601 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 46, + 469, + 287, + 601 + ], + "type": "text", + "content": " on the segment-level Type@AV F-score, attesting to the advantages of factoring in intersegment temporal dependencies." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 610, + 141, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 141, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 141, + 623 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "To demonstrate the potency of the different elements of our proposed method, UWAV, we conduct ablation studies. In particular, the proposed uncertainty-weighted pseudo-label based training, the uncertainty-weighted feature mixup scheme, and the class-balanced loss re-weighting schemes are ablated. As shown in Table 4, incorporating the uncertainty-weighted pseudo-label training step improves" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 308, + 297, + 544, + 380 + ], + "blocks": [ + { + "bbox": [ + 305, + 259, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 305, + 259, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 305, + 259, + 545, + 293 + ], + "type": "text", + "content": "Table 4. Ablation study of the proposed components in UWAV. \"Binary\" denotes training with binary pseudo-labels. \"Soft\" denotes training with uncertainty-weighted pseudo-labels." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 297, + 544, + 380 + ], + "lines": [ + { + "bbox": [ + 308, + 297, + 544, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 544, + 380 + ], + "type": "table", + "html": "
BinarySoftRe-weightMixupSegment-level
AVAVTypeEvent
✓62.767.761.964.262.2
✓63.068.361.864.462.8
✓✓63.669.563.065.463.1
✓✓63.969.062.865.263.4
✓✓✓64.270.063.465.963.9
", + "image_path": "d5844c5cbc20d84eb9145b330f77ca6e3482b488960220cf53fb5c9c6109838f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 392, + 546, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 546, + 524 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 546, + 524 + ], + "type": "text", + "content": "the segment-level Type@AV F-score by " + }, + { + "bbox": [ + 304, + 392, + 546, + 524 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 304, + 392, + 546, + 524 + ], + "type": "text", + "content": ", compared to using binary pseudo-labels. This demonstrates the benefit of accounting for the uncertainty in the pseudo-label estimation module. Moreover, sans the class-balanced loss weighting strategy, the model's performance is worse off by " + }, + { + "bbox": [ + 304, + 392, + 546, + 524 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 392, + 546, + 524 + ], + "type": "text", + "content": " on the Type@AV F-score, revealing the erroneous bias in the model's prediction arising from a skew of the class distribution. On the other hand, introducing the uncertainty-weighted feature mixup results in a gain of " + }, + { + "bbox": [ + 304, + 392, + 546, + 524 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 304, + 392, + 546, + 524 + ], + "type": "text", + "content": " on the Type@AV F-score, which underscores the importance of this self-supervised regularization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 537, + 378, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 537, + 378, + 549 + ], + "spans": [ + { + "bbox": [ + 306, + 537, + 378, + 549 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 557, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 546, + 712 + ], + "type": "text", + "content": "In this work, we address the challenging task of weakly-supervised audio-visual video parsing (AVVP), which presents a two-fold challenge: (i) potential misalignment between the events of the audio and visual streams, and (ii) the lack of fine-grained labels for each modality. We observe that by considering the temporal relationship between segments, our proposed method (UWAV) is able to provide more reliable pseudo-labels for better training of the inference module. In addition, by factoring in the uncertainty associated with these estimated pseudo-labels, regularizing the training process with a feature mixup strategy, and correcting for class imbalance, UWAV achieves state-of-the-art results on the LLP and AVE datasets." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13568" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 126, + 287, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 287, + 159 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 287, + 159 + ], + "type": "text", + "content": "[2] Moitreya Chatterjee, Jonathan Le Roux, Narendra Ahuja, and Anoop Cherian. Visual scene graphs for audio source separation. In ICCV, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 161, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 161, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 53, + 161, + 287, + 195 + ], + "type": "text", + "content": "[3] Moitreya Chatterjee, Narendra Ahuja, and Anoop Cherian. Learning audio-visual dynamics using scene graphs for audio source separation. In NeurIPS, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 197, + 287, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 197, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 53, + 197, + 287, + 250 + ], + "type": "text", + "content": "[4] Changan Chen, Unnat Jain, Carl Schissler, Sebastia Vicenc Amengual Gari, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audio-visual navigation in 3d environments. In ECCV, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 253, + 287, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 253, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 53, + 253, + 287, + 275 + ], + "type": "text", + "content": "[5] Changan Chen, Ziad Al-Halah, and Kristen Grauman. Semantic audio-visual navigation. In CVPR, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 277, + 287, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 277, + 287, + 320 + ], + "spans": [ + { + "bbox": [ + 53, + 277, + 287, + 320 + ], + "type": "text", + "content": "[6] Changan Chen, Sagnik Majumder, Ziad Al-Halah, Ruohan Gao, Santhosh Kumar Ramakrishnan, and Kristen Grauman. Learning to set waypoints for audio-visual navigation. In ICLR, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 323, + 287, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 323, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 323, + 287, + 356 + ], + "type": "text", + "content": "[7] Jiaben Chen, Renrui Zhang, Dongze Lian, Jiaqi Yang, Ziyao Zeng, and Jianbo Shi. iquery: Instruments as queries for audio-visual sound separation. In CVPR, 2023. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 358, + 287, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 287, + 402 + ], + "type": "text", + "content": "[8] Haoyue Cheng, Zhaoyang Liu, Hang Zhou, Chen Qian, Wayne Wu, and Limin Wang. Joint-modal label denoising for weakly-supervised audio-visual video parsing. In ECCV, 2022. 2, 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 404, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 404, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 53, + 404, + 287, + 437 + ], + "type": "text", + "content": "[9] Yingying Fan, Yu Wu, Yutian Lin, and Bo Du. Revisit weakly-supervised audio-visual video parsing from the language perspective. In NeurIPS, 2023. 2, 3, 4, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 440, + 287, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 440, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 440, + 287, + 483 + ], + "type": "text", + "content": "[10] Junyu Gao, Mengyuan Chen, and Changsheng Xu. Collecting cross-modal presence-absence evidence for weakly-supervised audio-visual event perception. In CVPR, 2023. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 486, + 287, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 486, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 48, + 486, + 287, + 539 + ], + "type": "text", + "content": "[11] Shijie Geng, Peng Gao, Moitreya Chatterjee, Chiori Hori, Jonathan Le Roux, Yongfeng Zhang, Hongsheng Li, and Anoop Cherian. Dynamic graph representation learning for video dialog via multi-modal shuffled transformers. In AAAI, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 287, + 586 + ], + "type": "text", + "content": "[12] Tiantian Geng, Teng Wang, Jinming Duan, Runmin Cong, and Feng Zheng. Dense-localizing audio-visual events in untrimmed videos: A large-scale benchmark and baseline. In CVPR, 2023. 2, 4, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 589, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 287, + 620 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 677 + ], + "type": "text", + "content": "[14] Shawn Hershey, Sourish Chaudhuri, Daniel PW Ellis, Jort F Gemmeke, Aren Jansen, R Channing Moore, Manoj Plakal, Devin Platt, Rif A Saurous, Bryan Seybold, et al. Cnn architectures for large-scale audio classification. In ICASSP, 2017. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[15] Xixi Hu, Ziyang Chen, and Andrew Owens. Mix and localize: Localizing sound sources in mixtures. In CVPR, 2022. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[16] Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Egocentric audio-visual object localization. In CVPR, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "text", + "content": "[17] Yung-Hsuan Lai, Yen-Chun Chen, and Frank Wang. Modality-independent teachers meet weakly-supervised audio-visual event parser. In NeurIPS, 2023. 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "text", + "content": "[18] Guangyao Li, Yake Wei, Yapeng Tian, Chenliang Xu, JiRong Wen, and Di Hu. Learning to answer questions in dynamic audio-visual scenarios. In CVPR, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "text", + "content": "[19] Jinxiang Liu, Yikun Liu, Fei Zhang, Chen Ju, Ya Zhang, and Yanfeng Wang. Audio-visual segmentation via unlabeled frame exploitation. In CVPR, 2024. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 209, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 230 + ], + "type": "text", + "content": "[20] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 232, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 232, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 545, + 262 + ], + "type": "text", + "content": "[21] Sagnik Majumder, Ziad Al-Halah, and Kristen Grauman. Move2hear: Active audio-visual source separation. In ICCV, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 265, + 545, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 265, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 307, + 265, + 545, + 298 + ], + "type": "text", + "content": "[22] Yuxin Mao, Jing Zhang, Mochu Xiang, Yiran Zhong, and Yuchao Dai. Multimodal variational auto-encoder based audio-visual segmentation. In ICCV, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 300, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 300, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 300, + 545, + 331 + ], + "type": "text", + "content": "[23] Shentong Mo and Yapeng Tian. Multi-modal grouping network for weakly-supervised audio-visual video parsing. In NeurIPS, 2022. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 333, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 333, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 333, + 545, + 365 + ], + "type": "text", + "content": "[24] Shentong Mo and Yapeng Tian. Audio-visual grouping network for sound localization from mixtures. In CVPR, 2023. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 367, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 367, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 367, + 545, + 409 + ], + "type": "text", + "content": "[25] Kranti Kumar Parida, Siddharth Srivastava, and Gaurav Sharma. Beyond mono to binaural: Generating binaural audio from mono audio with depth and cross modal attention. In WACV, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 412, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 412, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 412, + 545, + 453 + ], + "type": "text", + "content": "[26] Kranthi Kumar Rachavarapu, Kalyan Ramakrishnan, et al. Weakly-supervised audio-visual video parsing with prototype-based pseudo-labeling. In CVPR, 2024. 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "type": "text", + "content": "[27] Kranthi Kumar Rachavarapu et al. Boosting positive segments for weakly-supervised audio-visual video parsing. In ICCV, 2023. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 491, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 491, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 491, + 545, + 544 + ], + "type": "text", + "content": "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 546, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 546, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 546, + 545, + 578 + ], + "type": "text", + "content": "[29] Varshanth Rao, Md Ibrahim Khalil, Haoda Li, Peng Dai, and Juwei Lu. Dual perspective network for audio-visual event localization. In ECCV, 2022. 1, 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 580, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 580, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 580, + 545, + 624 + ], + "type": "text", + "content": "[30] Ludan Ruan, Yiyang Ma, Huan Yang, Huiguo He, Bei Liu, Jianlong Fu, Nicholas Jing Yuan, Qin Jin, and Baining Guo. Mm-diffusion: Learning multi-modal diffusion models for joint audio and video generation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 625, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 668 + ], + "type": "text", + "content": "[31] Faegheh Sardari, Armin Mustafa, Philip JB Jackson, and Adrian Hilton. Coleaf: A contrastive-collaborative learning framework for weakly supervised audio-visual video parsing. In ECCV, 2024. 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[32] Arda Senocak, Hyeonggon Ryu, Junsik Kim, Tae-Hyun Oh, Hanspeter Pfister, and Joon Son Chung. Sound source localization is all about cross-modal alignment. In ICCV, 2023. 1, 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13569" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[33] Ankit Shah, Shijie Geng, Peng Gao, Anoop Cherian, Takaaki Hori, Tim K Marks, Jonathan Le Roux, and Chiori Hori. Audio-visual scene-aware dialog and reasoning using audio-visual transformers with joint student-teacher learning. In ICASSP, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 161 + ], + "type": "text", + "content": "[34] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-visual event localization in unconstrained videos. In ECCV, 2018. 1, 2, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 196 + ], + "type": "text", + "content": "[35] Yapeng Tian, Dingzeyu Li, and Chenliang Xu. Unified multisensory perception: Weakly-supervised audio-visual video parsing. In ECCV, 2020. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 229 + ], + "type": "text", + "content": "[36] Yapeng Tian, Di Hu, and Chenliang Xu. Cyclic co-learning of sounding object visual grounding and sound separation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 230, + 287, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 263 + ], + "type": "text", + "content": "[37] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In CVPR, 2018. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "text", + "content": "[38] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2, 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "type": "text", + "content": "[39] Yu Wu and Yi Yang. Exploring heterogeneous clues for weakly-supervised audio-visual video parsing. In CVPR, 2021. 2, 3, 5, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 343, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 387 + ], + "type": "text", + "content": "[40] Yusong Wu, Ke Chen, Tianyu Zhang, Yuchen Hui, Taylor Berg-Kirkpatrick, and Shlomo Dubnov. Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation. In ICASSP, 2023. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 388, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 430 + ], + "type": "text", + "content": "[41] Yazhou Xing, Yingqing He, Zeyue Tian, Xintao Wang, and Qifeng Chen. Seeing and hearing: Open-domain visual-audio generation with diffusion latent aligners. In CVPR, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 433, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 433, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 287, + 476 + ], + "type": "text", + "content": "[42] Qi Yang, Xing Nie, Tong Li, Pengfei Gao, Ying Guo, Cheng Zhen, Pengfei Yan, and Shiming Xiang. Cooperation does matter: Exploring multi-order bilateral relations for audio-visual segmentation. In CVPR, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 478, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 287, + 510 + ], + "type": "text", + "content": "[43] Yuxin Ye, Wenming Yang, and Yapeng Tian. Lavss: Location-guided audio-visual spatial audio separation. In WACV, 2024. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 511, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 566 + ], + "type": "text", + "content": "[44] Abdelrahman Younes, Daniel Honerkamp, Tim Welschehold, and Abhinav Valada. Catch me if you hear me: Audio-visual navigation in complex unmapped environments with moving sounds. IEEE Robotics and Automation Letters, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 567, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 611 + ], + "type": "text", + "content": "[45] Jiashuo Yu, Ying Cheng, Rui-Wei Zhao, Rui Feng, and Yuejie Zhang. Mm-pyramid: Multimodal pyramid attentional network for audio-visual event localization and video parsing. In ACM MM, 2022. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "type": "text", + "content": "[46] Yinfeng Yu, Wenbing Huang, Fuchun Sun, Changan Chen, Yikai Wang, and Xiaohong Liu. Sound adversarial audio-visual navigation. In ICLR, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "text", + "content": "[47] Heeseung Yun, Youngjae Yu, Wonsuk Yang, Kangil Lee, and Gunhee Kim. Pano-avqa: Grounded audio-visual question answering on 360deg videos. In ICCV, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "text", + "content": "[48] Jinxing Zhou, Liang Zheng, Yiran Zhong, Shijie Hao, and Meng Wang. Positive sample propagation along the audio-visual event line. In CVPR, 2021. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 195 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "type": "text", + "content": "[49] Jinxing Zhou, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audio-visual segmentation. In ECCV, 2022. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 546, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 546, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 546, + 150 + ], + "type": "text", + "content": "[50] Jinxing Zhou, Dan Guo, Yiran Zhong, and Meng Wang. Improving audio-visual video parsing with pseudo visual labels. arXiv preprint arXiv:2303.02344, 2023. 2, 3, 4, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 152, + 546, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 546, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 546, + 195 + ], + "type": "text", + "content": "[51] Jinxing Zhou, Dan Guo, Yuxin Mao, Yiran Zhong, Xiaojun Chang, and Meng Wang. Label-anticipated event disentanglement for audio-visual video parsing. arXiv preprint arXiv:2407.08126, 2024. 2, 3, 7" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13570" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/UltraFusion_ Ultra High Dynamic Imaging using Exposure Fusion/c2c26ec1-c8af-49f8-87ba-b3c749531067_content_list.json b/2025/UltraFusion_ Ultra High Dynamic Imaging using Exposure Fusion/c2c26ec1-c8af-49f8-87ba-b3c749531067_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3a9daa02c7146e832b594cdb57bae4f414cc10ff --- /dev/null +++ b/2025/UltraFusion_ Ultra High Dynamic Imaging using Exposure Fusion/c2c26ec1-c8af-49f8-87ba-b3c749531067_content_list.json @@ -0,0 +1,2058 @@ +[ + { + "type": "text", + "text": "UltraFusion: Ultra High Dynamic Imaging using Exposure Fusion", + "text_level": 1, + "bbox": [ + 163, + 130, + 833, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zixuan Chen $^{1,3*}$ Yujin Wang $^{1*}$ Xin Cai $^{2}$ Zhiyuan You $^{2}$ Zheming Lu $^{3}$ Fan Zhang $^{1}$ Shi Guo $^{1}$ Tianfan Xue $^{2,1}$ $^{1}$ Shanghai AI Laboratory $^{2}$ The Chinese University of Hong Kong $^{3}$ Zhejiang University $\\{\\mathrm{zxchen}, \\mathrm{zheminglu}\\} @ \\mathrm{zju.edu.cn}, \\{\\mathrm{wangyujin}, \\mathrm{zhangfan}, \\mathrm{guoshi}\\} @ \\mathrm{pjlab.org.cn}$ caixin@link.cuhk.edu.hk, zhiyuanyou@foxmail.com, tfxue@ie.cuhk.edu.hk", + "bbox": [ + 135, + 178, + 856, + 268 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8c0cd337a93d395280ba4e556888a77ce0c9e50727e6122374703d8a3e498ce7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 282, + 225, + 349 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/36d3e4ee10e59d8b48b0e69ed2af2771da8cffbef7d1ef45698dcfcc478c5487.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 353, + 225, + 420 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/63d68dd668dfee57726428ac7abc43378a731407eaa586b283c9c68858466568.jpg", + "image_caption": [ + "Under Exposure" + ], + "image_footnote": [], + "bbox": [ + 96, + 422, + 169, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f63f14dc9e6b83e9e95f9a46eb4855ca19e8a03694217921e2bf0cb8e2520d9c.jpg", + "image_caption": [ + "Over Exposure" + ], + "image_footnote": [], + "bbox": [ + 169, + 422, + 245, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/86651fcc803b59ecfa03880f24637a9bb3261e0666a6e9ac4f759db0266b63af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 282, + 493, + 419 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fcb181836efee2444eee333e4a85eea9eb756749bd591ab59362ec48cb6d60e9.jpg", + "image_caption": [ + "Previous HDR Reconstruction" + ], + "image_footnote": [], + "bbox": [ + 267, + 422, + 339, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7e496caaf6c6680156378f487b44035a95744e62423e91d5cf000f060bf49374.jpg", + "image_caption": [ + "Exposure Multi-", + "Figure 1. Comparing our guided inpainting HDR imaging method with a state-of-the-art HDR reconstruction [25] and multi-exposure fusion [48] methods. Both scenes are selected from our captured new real-world benchmark. Left: night cityscape with large exposure difference. Right: afternoon street with motion-induced occlusion. Previous methods struggle to handle these scenes. By modeling HDR as an inpainting problem, our method can produce visually appealing results without ghosting artifacts in these challenging scenes." + ], + "image_footnote": [], + "bbox": [ + 343, + 422, + 416, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b835c1fbb72c1f1486a1043c9d220255d1fe832bfb91412c607a4b18e587846a.jpg", + "image_caption": [ + "Our Guided Inpainting" + ], + "image_footnote": [], + "bbox": [ + 418, + 422, + 493, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c831a0057bd40a4adbbca202cdafe7fb9330090748ae062db1032b3a29990e3a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 282, + 630, + 349 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3c455bfe6e36334b0a72ff3d20b997495663f57fc6e97ca232bb72dffa9b5076.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 353, + 632, + 419 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3ff29fa01c3875d6d0c40d2def81b6b154728a7cce43e84c94b453d690d37b9e.jpg", + "image_caption": [ + "Under Exposure" + ], + "image_footnote": [], + "bbox": [ + 503, + 422, + 576, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2f02524b899cfa3a741c8f5d984376a57a8bb172d9d4fdde124d26e59fc916ec.jpg", + "image_caption": [ + "Over Exposure" + ], + "image_footnote": [], + "bbox": [ + 580, + 422, + 651, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d0c3555145f968fd5a365347564bd2000c2ac22711b6e14cd5290fdf41cf82ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 282, + 901, + 419 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a501619ead788a1050d16faa5643ac3b2bcd7522b4c0abecb555b05be4ad8f1f.jpg", + "image_caption": [ + "Previous HDR Reconstruction" + ], + "image_footnote": [], + "bbox": [ + 679, + 422, + 751, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6d8560561528f07f0acafd374af04b22d9d42f3f60c94fed3d091372e6291349.jpg", + "image_caption": [ + "Previous MultiExposure Fusion" + ], + "image_footnote": [], + "bbox": [ + 754, + 422, + 826, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2e320d4ccb5ca97842a1e91ba555728ec8993b80e833b63cbedd42c8f1e7927c.jpg", + "image_caption": [ + "Our Guided Inpainting" + ], + "image_footnote": [], + "bbox": [ + 828, + 422, + 901, + 479 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 580, + 326, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Capturing high dynamic range (HDR) scenes is one of the most important issues in camera design. Majority of cameras use exposure fusion, which fuses images captured by different exposure levels, to increase dynamic range. However, this approach can only handle images with limited exposure difference, normally 3-4 stops. When applying to very high dynamic range scenes where a large exposure difference is required, this approach often fails due to incorrect alignment or inconsistent lighting between inputs, or tone mapping artifacts. In this work, we propose UltraFusion, the first exposure fusion technique that can merge inputs with 9 stops differences. The key idea is that we model exposure fusion as a guided inpainting problem, where the under-exposed image is used as a guidance to fill the missing information of over-exposed highlights in the overexposed region. Using an under-exposed image as a soft guidance, instead of a hard constraint, our model is robust", + "bbox": [ + 88, + 607, + 485, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to potential alignment issue or lighting variations. Moreover, by utilizing the image prior of the generative model, our model also generates natural tone mapping, even for very high-dynamic range scenes. Our approach outperforms HDR-Transformer on latest HDR benchmarks. Moreover, to test its performance in ultra high dynamic range scenes, we capture a new real-world exposure fusion benchmark, UltraFusion dataset, with exposure differences up to 9 stops, and experiments show that UltraFusion can generate beautiful and high-quality fusion results under various scenarios. Code and data will be available at https://openimaginglab.github.io/UltraFusion.", + "bbox": [ + 511, + 582, + 906, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 782, + 645, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "High dynamic range (HDR) imaging is one of the fundamental problems in the modern camera design. Due to hardware limitation, camera sensors have a much smaller dynamic range compared with the real-world. To increase it, majority of HDR solutions merge multiple images, either with the same [9] or different exposure lev", + "bbox": [ + 511, + 809, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. This work was done during Zixuan Chen's internship at Shanghai Artificial Intelligence Laboratory.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16111", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3a4b4e41840596b085cca383f3ac0e28058089416fec9396c9695a8e73b7af40.jpg", + "image_caption": [ + "Figure 2. Visual comparison on directly utilizing ControlNet [65] and our UltraFusion. Without pre-aligned data, ControlNet struggles to fix a frame for reference. However, our method fixes the over-exposed image as the reference and the under-exposed image as the guidance for inpainting, thereby avoiding artifacts." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 480, + 222 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "els [3, 15, 17, 24, 34, 40, 43, 49, 56, 57]. Despite all recent advances, majority HDR algorithms can only bring limited increase in dynamic range. For example, $\\mathrm{HDR + }$ [9], the first HDR algorithm used by commercial cameras, can only robustly increase the dynamic range by 8 times (3 stops). Therefore, in this work, we study the following question: can we drastically increase the dynamic range of a camera by fusing two images with very large exposure difference, like 9 stops in Fig. 1?", + "bbox": [ + 88, + 309, + 482, + 445 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This is a fundamental challenging problem, due to the following three issues. First, to handle dynamic scenes, most of HDR fusion algorithms will first align input frames, which is very challenging when input has large brightness difference. As result, a ghosting issue happens when alignment fails, indicated by zoom-in patches in the right scene of Fig. 1. Second, most of HDR algorithms assume that under-exposed image is simply a darker version of the normal image. However, the appearance of an object may change when exposure levels change, like ship in the left scene of Fig. 1, resulting unnatural fusion result. Third, sometimes the result of fusion is an HDR image, which cannot be directly shown on normal low-dynamic-range display. Therefore, these HDR images will be further compressed through a tone-mapping process. When dynamic range is high, tone-mapping may introduce additional issue. Maintaining a natural contrast and rich details in the final output is challenging, as shown in the zoom-in patches of previous HDR reconstruction method in Fig. 1.", + "bbox": [ + 91, + 446, + 482, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a completely different fusion method, UltraFusion, which models it as a guided inpainting problem. In this setup, the user captures two images, one normal exposed image where brighter objects are overexposed, and another under exposed image, which only captures the very brighter parts of the scene. We use the normal exposed image as a reference, and inpainting the missing information in the highlight. Unlike the traditional inpainting, we use the information from under-exposed frame as guidance, so inpainted highlight is not completely generated, but stays consistent with under-exposed frame.", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There are three advantages of this approach when handling large exposure difference. First, it follows the exposure fusion [30] setup. Unlike HDR fusion techniques, which first generate HDR image and then compress to a low-dynamic-range (LDR) image through tone mapping, exposure fusion directly generates LDR output, avoiding cascading errors. That being said, our method outputs a tone-mapped LDR image instead of a linear HDR. Second, the under-exposed image is used as a soft guidance, instead of a hard constraint. Therefore, UltraFusion is robust to alignment error (see the right scene of Fig. 1) and lighting variation (see the left scene of Fig. 1). Third, image prior of the generative model ensures the natural look of the output image, reducing the potential artifacts.", + "bbox": [ + 511, + 90, + 903, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To train a guided inpainting, a simple solution is to train a ControlNet [65] using two input images. However, such paradigm cannot handle dynamic HDR scenes, as ControlNet may not know which frame to choose as the reference, increasing the difficulty to fuse the information from over and short-exposed images. For instance, as shown in Fig. 2, ControlNet selects the over-exposed image as the reference frame in red boxed region, but chooses the under-exposed image as the reference frame in green boxed region, leading to substantial artifacts in the result. Additionally, as a generative model, it may inevitably generate fake image content, as indicated by green box in Fig. 2.", + "bbox": [ + 511, + 305, + 903, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we design UltraFusion as follows. First, we warp the short-exposed image to the long-exposed one, and mask out the occluded regions. Then, we utilize the diffusion prior to inpaint the long-exposed image guided by partial short-exposed information. To make the network retain more details to generate guidance information, we propose a new decompose-and-fuse control branch, which eliminates the luminance component of the short-exposed image, extracting structure and color information instead, and employs a multi-scale cross-attention to improve the feature fusion with long-exposed image. Second, as there is no existing large-scale training data for exposure fusion of dynamic scene, we also propose a novel training data synthesis pipeline, utilizing existing high-quality multi-exposure datasets (pre-aligned) of static scene and video datasets. At last, to ensure the generated output maintains the fidelity to the reality, we also train an additional fidelity control branch, using the same decompose-and-fuse strategy and multi-scale cross-attention.", + "bbox": [ + 511, + 489, + 903, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "At last, to evaluate the effectiveness of our framework, we capture 100 under/over-exposed image pairs, covering daytime, nighttime, indoor, outdoor scenes with local and global motion patterns. Experiments on both latest HDR imaging datasets and our captured benchmark demonstrate that, comparing to existing methods, our UltraFusion is more robust to scenes with large exposure differences and large motion, as shown in Fig. 1.", + "bbox": [ + 511, + 779, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16112", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 91, + 89, + 227, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. HDR imaging", + "text_level": 1, + "bbox": [ + 89, + 116, + 235, + 132 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "HDR Imaging can be deided into HDR reconstruction and Multi-Exposure Fusion (MEF) typically depends on the domain where fusion occurs [19]. HDR reconstruction methods [3, 15, 17, 24, 34, 40, 43, 49, 51, 56, 57] inverts the camera response function (CRF) to merge exposure brackets in the linear HDR domain [28]. In most cases, tone mapping is necessary to display the reconstructed HDR image properly on standard LDR monitors. As a cost-effective alternative [13], MEF methods [13, 19, 22, 28, 30, 33, 35, 48, 52, 68, 69] directly fuse images in the LDR domain, sidestepping CRF calibration and sophisticated tone mapping process [30]. Regardless of the type of HDR Imaging methods, they contend with ghosting artifacts caused by camera shake and object movement [39]. Previous methods have attempted explicit or implicit alignment using optical flow [15, 33, 34, 49] or attention mechanism [3, 17, 24, 40, 43, 56, 57]. However, most HDR Imaging methods suffer from unpleasing artifacts when large motion causes occlusion in the complementary region.", + "bbox": [ + 89, + 140, + 482, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models. Recently, it has been witnessed with the rapid rise of diffusion models [10, 37] and their successful application in various tasks, including controllable image generation [59, 65], image restoration [21, 23, 31, 47, 63], image editing [4, 14, 29, 38] and image inpainting [5, 26, 50, 64]. In the field of HDR imaging, the application of diffusion models has primarily focused on HDR deghosting [8, 12, 58]. However, since these methods do not leverage the diffusion priors learned from large-scale datasets, their generalization ability is limited by the scale of the HDR dataset. While some recent works [6, 18] have employed diffusion priors, they tend to focus on single image HDR. Without another differently-exposed image as reference, the results generated by these methods lack sufficient reliability. Unlike previous methods, we utilize diffusion priors and use the short-exposed image as a reference to perform reliable inpainting in the highlight regions of the overexposed image, thereby achieving natural and reliable HDR scene reconstruction. Compared to diffusion-based inpainting methods [5, 26, 50, 64] that perform inpainting solely from scratch, we leverage information from short-exposure images to guide a more accurate inpainting process.", + "bbox": [ + 89, + 430, + 482, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Tone mapping methods. The goal of tone mapping is to convert HDR images to LDR for display on standard screens while enhancing visual detail. Due to the challenge of obtaining ground-truth tone mapping results, unsupervised deep learning approaches have been developed using adversarial [45] and contrastive [2] learning. To address data limitations, Cai et al. [1] manually curated training data by selecting the best results from 13 tone mapping methods for network learning [1, 11]. However, previous tone", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mapping algorithms, lacking robust image priors and facing data constraints, struggle with visually pleasing results and generalization in extreme high dynamic range scenes. By incorporating diffusion-based image priors, our method achieves aesthetic results even in challenging high dynamic range scenarios (see Fig. 1).", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 194, + 648, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an over-exposed image $I_{oe}$ and an under-exposed image $I_{ue}$ , traditional exposure fusion algorithms directly aggregate different frequency band of both images, which are sensitive to misalignment error or lighting variation. Instead, we treat this as an inpainting problem. Specifically, we use the over-exposed image $I_{oe}$ as the base image and inpainting missing information in the highlight region. To ensure inpainted highlights are real, we also use highlights from under-exposed image as guidance.", + "bbox": [ + 511, + 218, + 903, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on this idea, we design a 2-stage newtwork shown in Fig. 3, which consists of the pre-alignment stage and the guided inpainting stage. The pre-alignment stage outputs a coarse-aligned version of $I_{ue}$ , which is used as the soft guidance in the following guided inpainting stage. Details of each stage is described below.", + "bbox": [ + 511, + 354, + 905, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Pre-alignment stage", + "text_level": 1, + "bbox": [ + 511, + 453, + 702, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Most of optical flow alignment assumes input have the similar brightness. Therefore, we first adjust the brightness $I_{ue}$ to match the distribution of $I_{oe}$ through intensity mapping function [7]. Then, we adopt RAFT [42], a pre-trained optical flow network, to estimate the bidirectional flow $f_{oe \\rightarrow ue}$ and $f_{ue \\rightarrow oe}$ and align $I_{ue}$ to $I_{oe}$ using backward warping. However, backward warping will result in ghosting at the occlusion boundary [67], leading to artifacts in the next guided inpainting stage. To solve that, we utilize the forward-and-backward consistency check [54] to estimate occluded regions $\\mathcal{M}$ and mask them out in the warped output. Finally, we can obtain a pre-aligned output $I_{ue \\rightarrow oe}$ of the first stage:", + "bbox": [ + 511, + 474, + 905, + 672 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {u e \\rightarrow o e} = (1 - \\mathcal {M}) \\cdot \\mathcal {W} \\left(I _ {u e}, f _ {o e \\rightarrow u e}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 680, + 903, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{W}$ denotes the backward warping. Fig. 3 (a) shows the output is a masked and aligned under-exposed image.", + "bbox": [ + 511, + 704, + 903, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Guided inpainting stage", + "text_level": 1, + "bbox": [ + 511, + 742, + 733, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We build our guided inpainting model based on the Stable Diffusion Model [36] because its powerful generative prior can help to resolve ambiguity during inpainting. Similar to other diffusion-based image enhancement techniques [21], we also inject the following information through an additional control branch, as shown in Fig. 3(b): 1) the image to be inpainted, which is the over-exposed image $I_{oe}$ , 2) the additional guidance of highlight, which is the underexposed image $I_{ue}$ , and 3) the diffusion latent at the current", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16113", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8fadd714b2953a61ee44d5c446f878dc6e4cc7cac1082b3d009fbf9faa54d7de.jpg", + "image_caption": [ + "(a) Pre-Alignment Stage", + "Figure 3. The whole backbone of UltraFusion. Our method is a 2-stage framework, consisting of (a) pre-alignment stage and (b) guided inpainting stage. The first stage pre-aligns the under-exposed image $I_{ue}$ to the over-exposed image $I_{oe}$ and masks the occluded regions. In the subsequent guided inpainting stage, we propose a new decompose-and-fuse control branch to utilize the diffusion priors, together with a fidelity control branch for fidelity keeping." + ], + "image_footnote": [], + "bbox": [ + 98, + 116, + 491, + 280 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/11a54cfe6cf82e099f81cf5ed02503946f06f32d6bb8edb3f9ca63e8b19d3f0e.jpg", + "image_caption": [ + "(b) Guided Inpainting Stage" + ], + "image_footnote": [], + "bbox": [ + 504, + 104, + 903, + 280 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ac6aa4d034a334b9edfb1f03808169d4ce337e20f0e5bbde26014db38a946ca2.jpg", + "image_caption": [ + "Decompose-and-Fuse Control Branch", + "Figure 4. The detailed architecture of our proposed decompose-and-fuse control branch." + ], + "image_footnote": [], + "bbox": [ + 94, + 375, + 478, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "diffusion step $z_{t}$ , as previous work [21] shows that including diffusion latent as an additional condition can improve image quality. The main diffusion denoising network is a pretrained U-Net. The over-exposed image is first encoded using a pretrained VAE before entering the diffusion module, and the outputs are converted back to the image space using the pretrained decoder.", + "bbox": [ + 89, + 595, + 483, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The main differences between our solution and general diffusion-based image enhancement are two-fold. First, we propose a novel decompose-and-fuse control branch to inject two input images and diffusion latent as a control signal, as Fig. 11 (b) shows that naively injecting this information may not be able to guide diffusion to faithfully inpaint missing highlights obtained from the under-exposed image. Second, we train an additional fidelity control branch to provide faithful structure and color information for the decoding process via shortcuts. Details are described below.", + "bbox": [ + 89, + 703, + 482, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Decompose-and-fuse control branch. Fig. 4 shows our control branch. We use the over-exposed image $I_{oe}$ as the main control signal and the under-exposed image $I_{ue}$ as the", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "soft guidance. Following ControlNet [65], we copy the encoder and middle blocks of the denoising U-Net as the main extractor, but update their weights during training. A simple soft guidance is to use the encoded under-exposed image latent $y_{ue}$ from the VAE encoder, combined with the overexposed latent $y_{oe}$ . However, the underexposed images are often too dark to be directly used as soft guidance, since the model may entirely ignore this guidance.", + "bbox": [ + 511, + 351, + 906, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, we decompose the under-exposed image into the color and structure information, both of which are robust to brightness changes. Specifically, we use the normalized image as the structure component, similar to SSIM [46], as:", + "bbox": [ + 511, + 473, + 906, + 532 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {u e} = \\left(Y _ {u e} - \\mu \\left(Y _ {u e}\\right)\\right) / \\sigma \\left(Y _ {u e}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 547, + 903, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Y_{ue}$ represents the luminance channel of $I_{ue}$ under YUV space, $\\mu(\\cdot)$ and $\\sigma(\\cdot)$ denote the mean intensity and standard deviation, respectively. The chroma channels (UV) are used as color information. The extracted structure and color information are further encoded using trained color and structure extractors (gray GE block in Fig. 4). Following [65], we implement GE using several simple convolution layers to extract multi-scale features.", + "bbox": [ + 511, + 568, + 905, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At last, the extracted features are injected into the main extractor with a multi-scale cross attention, as shown in the bottom part of Fig. 4. The output of each cross attention module is fed into both the next level of the main extractor and the corresponding U-Net block, using zero convolution. Fidelity control branch. Even with the control block, we sometimes observe undesirable modifications of texture introduced by VAE. An example is shown in Fig. 11 (c). Therefore, to further improve fidelity, we design a Fidelity Control Branch (FCB), inspired by [47]. FCB mitigates texture distortions by injecting features into the VAE decoder. It has a similar architecture as the decompose-and-fuse control branch, with two main differences: 1) the main extractor of FCB adopts the same structure as the VAE Encoder,", + "bbox": [ + 511, + 689, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16114", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e7c2a68532deed2df235cb26f482d510940848446e1c592acbe6f8aa5ccf6695.jpg", + "image_caption": [ + "Figure 5. The illustration of our training data synthesis pipeline." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 486, + 217 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "rather than the denoising U-Net, to provide corresponding shortcuts to the VAE Decoder, with adjustments to the soft guidance extractor as well, and 2) the main extractor of the FCB directly takes the over-exposed image as input. To train the fidelity control branch, we freeze the pre-trained VAE Encoder and Decoder and encode the ground truth $I_{gt}$ to latent space, simulating the denoised latent $z_0$ during the inference. Then we input the corresponding overexposed image and under-exposed image to the fidelity control branch to extract faithful features. At last, the VAE Decoder decodes the compressed latent to a reconstructed image $\\hat{I}_{gt}$ . We adopt the reconstruction loss $||\\hat{I}_{gt} - I_{gt}||_1$ as an additional loss term.", + "bbox": [ + 89, + 237, + 483, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training data synthesis", + "text_level": 1, + "bbox": [ + 89, + 441, + 305, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Preparing the training data for the proposed guided inpainting network is challenging. To train the model, ideally we need a large-scale HDR dataset that 1) covers different dynamic scenes, 2) has large exposure variance up to 9 stops, and 3) contains ground truth fusion result. However, no existing dataset satisfy all requirements.", + "bbox": [ + 89, + 464, + 483, + 554 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To solve this issue, we propose a novel training data synthesis pipeline. Specifically, as shown in Fig. 5, we first randomly sample an image sequence with $N$ frames from a video dataset [55]. To model large motion, we select the first and the last frames. Then, similar to the pre-alignment stage, we estimate the bidirectional optical flow between two selected frames using [42] and compute a pseudo occlusion mask via forward-backward consistency check. Subsequently, we randomly sample an under-exposed image patch from the high-quality static dataset [1] (ground truths are pre-aligned), resize the pseudo occlusion mask to match the patch size, and mask out the pseudo occluded region to synthesize a pseudo pre-aligned output. With our synthesized training data, our UltraFusion learns to handle dynamic scenes with only static multi-exposure image pairs.", + "bbox": [ + 89, + 555, + 483, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 792, + 215, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental setting", + "text_level": 1, + "bbox": [ + 89, + 818, + 287, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We utilize SICE [1] dataset and Video-90K [55] dataset to synthesize our training data. Following previous works [1, 22], we select images with the highest and lowest brightness in each exposure bracket of the SICE dataset", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/68b45325848c2b6028dde807d2de1ff012aa2f924f636e9d09b76224060d7c5c.jpg", + "image_caption": [ + "Figure 6. (a) The data distribution of our benchmark. Coordinate value of exposure time represents the upper boundary. (b) The user study result on our benchmark." + ], + "image_footnote": [], + "bbox": [ + 514, + 88, + 903, + 172 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "as inputs. We evaluate our method on both static datasets and dynamic datasets. For evaluation, we use the MEFB dataset [66] with 100 static under/over-exposed image pairs and RealHDRV [39], a dynamic HDR deghosting test set containing 50 scenes with varying motion patterns.", + "bbox": [ + 511, + 229, + 903, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "UltraFusion benchmark. Existing exposure fusion benchmark cannot fully evaluate fusing in real-world challenging conditions, as they either lack realistic motion [66] or have limited dynamic range [15, 39, 44]. Therefore, we collect a new real-world UltraFusion benchmark, which contains 100 real-captured under/over-exposed image pairs. Compared to previous datasets, our benchmark is more challenging for three reasons: 1) Our benchmark features larger exposure differences between the two input images (up to 9 stops). 2) It includes more realistic motion, with many scenes containing extensive and unintentional foreground movement. 3) Our benchmark is highly diverse, encompassing daytime, nighttime, indoor, and outdoor scenes captured by DSLR Camera (Canon R8) and mobile phones (iPhone12, iPhone13, Redmi K50 Pro and OPPO Reno8 Pro). We summarize the exposure difference distribution and exposure time distribution of our benchmark in Fig. 6 (a). It can be observed that our benchmark covers a wide range of exposure differences and diverse exposure times, which can be used to effectively test the robustness of the HDR methods.", + "bbox": [ + 511, + 306, + 906, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. We leverage the generative prior encapsulated in Stable Diffusion V2.1 [37]. The decompose-and-fuse control branch (DFCB) is trained for 140k iterations with batch size $bs = 32$ on 8 NVIDIA RTX 4090 GPUs. We also train the fidelity control branch for 1000K iteration with batch size $bs = 1$ on a single NVIDIA RTX 4090 GPU. Adam is adopted as the optimizer and the learning rate is fixed to 0.0001. To adapt HDR reconstruction methods to 2 differently exposed inputs, we re-implement them by following their default settings.", + "bbox": [ + 511, + 611, + 908, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. We utilize four widely used non-reference metric MUSIQ [16], DeQA-Score [62], PAQ2PIQ [61] and HyperIQA [41] for quantitative comparison. Moreover, for static dataset [66], as no ground truths are provided, we select the task-specific MEF-SSIM [27] for structure retention evaluation. For dynamic dataset [39] with HDR ground truths, TMQI [60] is used to evaluate the performance of fidelity and naturalness. In addition, we conduct a user study on our Ultrafusion benchmark to", + "bbox": [ + 511, + 763, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16115", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/153282c10c34b54527e5a1ccf53178e9554ede352ff25c6775ea5ccd054a9eb8.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons on the static MEFB dataset [66]." + ], + "table_footnote": [], + "table_body": "
TypeMethodMEFB [66]
MUSIQ↑DeQA-Score↑PAQ2PIQ↑HyperIQA↑MEF-SSIM↑
HDR Rec.HDR-Transformer [25]63.102.98371.360.59960.8626
SCTNet [44]63.133.02171.480.60680.8777
SAFNet [17]61.702.51872.670.56460.7711
MEFDeepfuse [35]52.582.94567.960.43990.8968
MEF-GAN [53]50.592.81869.990.37450.7722
U2Fusion [52]63.393.21972.230.51590.9304
Defusion [20]62.703.11872.820.54550.9062
MEFLUT [13]65.713.27771.210.52670.8608
HSDS-MEF [48]66.763.54472.600.60260.9520
TC-MoA [69]64.603.35571.850.53940.9636
OursUltraFusion68.823.88173.800.64820.9385
", + "bbox": [ + 98, + 114, + 658, + 291 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a060de34805c5ccb86e319fe2ec960ca25d98d5de577432f89cfa9d0c546e0cf.jpg", + "image_caption": [ + "Figure 7. Trade-off curve between MEF-SSIM and MUSIQ on MEFB dataset [66]. Our UltraFusion achieves the best trade-off between image quality and information preservation." + ], + "image_footnote": [], + "bbox": [ + 679, + 90, + 890, + 210 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/16f1be99088e9f719c2de384b169b8a827fb51a78931608bb59b421ac0c46522.jpg", + "table_caption": [ + "Table 2. Quantitative comparisons on dynamic RealHDRV dataset [39] and our challenging UltraFusion benchmark." + ], + "table_footnote": [], + "table_body": "
TypeMethodRealHDRV [39]UltraFusion Benchmark
TMQI↑MUSIQ↑DeQA-Score↑PAQ2PIQ↑HyperIQA↑MUSIQ↑DeQA-Score↑PAQ2PIQ↑HyperIQA↑
HDR Rec.HDR-Transformer [25]0.868062.243.49670.330.522563.662.90972.830.5619
SCTNet [44]0.871562.693.53270.740.527261.843.10272.940.5888
SAFNet [17]0.872662.073.50670.480.515661.502.17973.150.5487
MEFDefusion [20]0.818756.603.30268.380.485660.313.35271.870.5463
MEFLUT [13]0.829762.423.31570.040.502063.623.34371.730.5074
HSDS-MEF [48]0.832361.763.36071.110.505464.543.62773.420.5923
OursUltraFusion0.892567.513.83073.400.583368.413.95775.180.6214
", + "bbox": [ + 94, + 316, + 903, + 444 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/93fae921e5e476884387d948973f1976642988310f9507a1a08887028e930a3e.jpg", + "image_caption": [ + "Inputs" + ], + "image_footnote": [], + "bbox": [ + 93, + 452, + 227, + 609 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f63c111f10620fd0b58b32182118a0267571a37ccdfb6e10254bf33f7d5b2f90.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 227, + 452, + 362, + 530 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4b2c30a02693391b912120a96138c5f9c5a574fafb51e686b2611241e193ea96.jpg", + "image_caption": [ + "HSDS-MEF" + ], + "image_footnote": [], + "bbox": [ + 228, + 531, + 362, + 609 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4a76d138e85d212c3ca80e0004d77fd085d52211103a356828fcfbbfb4b74b71.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 452, + 498, + 530 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/50091b5090b783499c421556760e57a9ce03aadb7eb1f33d0b04aa83ed869fc8.jpg", + "image_caption": [ + "TC-MoA" + ], + "image_footnote": [], + "bbox": [ + 364, + 531, + 498, + 609 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9b6897253144233b47e28c2683745d54bdc929f8f82eb245058545b65d401c7d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 498, + 452, + 633, + 530 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b93b39a851c7892229290213dfe78c449b4845a73dbd35ddf953521c833b168b.jpg", + "image_caption": [ + "HDR-Transformer", + "Figure 8. Visual comparisons of different exposure fusion methods on static MEFB dataset [66]." + ], + "image_footnote": [], + "bbox": [ + 498, + 531, + 633, + 609 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1973d59ef8ebbd782431021e85f5f6b24221bc452bbc94b5a823cd3b3e31eebb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 452, + 767, + 530 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/57d20f7dd140b1ddc5d5f1e2ca505fabc93d0bbb08787715cf6646b1bfb52a41.jpg", + "image_caption": [ + "SCTNet" + ], + "image_footnote": [], + "bbox": [ + 633, + 531, + 767, + 609 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f0f1c525374f4e41b8fe2a761b284a5c327d3dc005aa45f4debf56473acb5de5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 452, + 903, + 530 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1da5b84a86915fb13d404279e971376aa047b4317358a70ecb05753c10a2259e.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 769, + 531, + 903, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "perform subjective evaluation.", + "bbox": [ + 89, + 650, + 292, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons with HDR imaging methods", + "text_level": 1, + "bbox": [ + 89, + 679, + 447, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method with state-of-the-art HDR Imaging methods, including HDR reconstruction methods HDRTransformer [25], SCTNet [44], SAFNet [17], and multi-exposure fusion methods Deepfuse [35], MEF-GAN [53], U2Fusion [52], Defusion [20], MEFLUT [13], HSDS-MEF [48], TC-MoA [69]. As HDR reconstruction methods cannot output an LDR image directly, we use professional software Photomatix [32] to perform tone mapping.", + "bbox": [ + 89, + 702, + 482, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on static dataset. We evaluate the fusion performance on the MEFB dataset [66], focusing on large exposure differences. As shown in Tab. 1, our method outperforms other methods on all four non-reference metrics (MUSIQ, DeQA-Score, PAQ2PIQ, and HyperIQA). Specif", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ically, our proposed UltraFusion achieves 2.06 gain in terms of MUSIQ compared to HSDS-MEF. In terms of MEFSSIM, as shown in Fig. 7, our baseline model (ControlNet [65]) achieves better image quality but lacks fidelity, while HSDS-MEF [48] retains more information from inputs at the cost of quality. Our UltraFusion outperforms most algorithms, and achieves similar fidelity scores compared to HSDS-MEF and TC-MoA but with much higher image quality (non-reference metrics), indicating the best trade-off between visual quality and information preservation. The qualitative comparison in Fig. 8 further validates this. In contrast, HDR reconstruction methods (HDRTransformer and SCTNet) often miss some detail in highlights and MEF methods (HSDS-MEF and TC-MoA) introduce unnatural transition from brighter to dark regions.", + "bbox": [ + 511, + 650, + 906, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on dynamic dataset. To further illustrate the", + "bbox": [ + 513, + 885, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16116", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bc95d2be8f4d67ad38d53524cfc904f00289bfc22f5c7218c3eeb62ee2fcf447.jpg", + "image_caption": [ + "Figure 9. Visual comparisons on our captured UltraFusion benchmark." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 496, + 212 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0c5de0821e51669d6cfebf1021f7caf4ebfee2582ec8eb6a78669ed62e7d1548.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 89, + 903, + 212 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fc03860d1f45d750b32b003f2f060649ac00a6b85631b5fc9e857954e105b739.jpg", + "image_caption": [ + "Figure 10. Visual results on dynamic RealHDRV dataset [39]." + ], + "image_footnote": [], + "bbox": [ + 94, + 237, + 480, + 356 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "robustness of UltraFusion to global and local motion, we use the RealHDRV dataset [39]. We extract the corresponding over-exposed image from the HDR ground truth as input. Tab. 2 demonstrates that our UltraFusion achieve state-of-the-art performance in terms of all metrics. For dynamic scenes, TMQI metric is a particularly important metric, as it is specially designed for HDR evaluation by assessing structural similarity between fusion output and ground truth. Since MEF methods are mainly designed for static scenes, they lack the capability to handle motion, resulting in low TMQI scores and overlay artifacts, as shown in Fig. 10. HDR reconstruction methods trained on dynamic datasets achieve better performance, but still produce noticeable artifact. On the other side, due to our soft inpainting guidance, UltraFusion is much more robust to misalignment and occlusion and the fusion output contains almost no artifacts. It achieves the highest TMQI in Tab. 2 and the best visual result in Fig. 10.", + "bbox": [ + 88, + 383, + 482, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation on our UltraFusion benchmark. At last, the evaluation on our benchmark validates the robustness of UltraFusion in the most challenging scenes. On all four non-reference metrics, our method outperforms competitors by a large margin, as shown in Tab. 2. The qualitative comparison in Fig. 9 is also consistent with the quantitative metrics. For example, in the red zoom-in patch, integrating the highly bright sun from the under-exposed image into the over-exposed image is extremely challenging. Other methods fail to maintain the shape of the sun or preserve high contrast in the fused region, while our method naturally reconstructs the sun, preserving its appearance and visual pleasing tone of the whole image. Moreover, we conduct a user study on our proposed benchmark. Specifically, we randomly select 20 scenes from our benchmark and invite 136 different users to participate. For each scene, each user", + "bbox": [ + 89, + 659, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fb5aa52949fbbc1df2d002e6a492dc3a4a686669c3ec05ab5caa99f8bcc23f90.jpg", + "table_caption": [ + "Table 3. Ablation studies of three key components of our proposed UltraFusion on RealHDRV dataset [39]." + ], + "table_footnote": [], + "table_body": "
ModelTMQI↑MUSIQ↑
w/o Alignment Strategy0.742763.67
w/o Decompose-and-Fuse Control Branch0.887266.94
w/o Fidelity Control Branch0.876367.36
Ours UltraFusion0.892567.51
", + "bbox": [ + 517, + 268, + 901, + 348 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "is asked to compare our method with a randomly chosen baseline. The user study in Fig. 6 (b) indicates our method is more favored by users than competitors. This outcome aligns with the non-reference metric evaluation, showing that our method produces more natural images. More visual comparisons are available in the supplementary.", + "bbox": [ + 511, + 356, + 906, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation studies", + "text_level": 1, + "bbox": [ + 511, + 455, + 673, + 469 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To validate the effectiveness of our UltraFusion, we conduct ablation studies on the three key components, followed by an in-depth exploration of the designs of them.", + "bbox": [ + 511, + 477, + 903, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Key components. First, we perform the ablation study of the proposed three key components, including the alignment strategy, decompose-and-fuse control branch, and fidelity control branch, on the RealHDRV [39] dataset, as shown in Tab. 3. We first remove our alignment strategy, opting instead to train the model directly on the original SICE [1] dataset and input differently exposed image pairs without coarse alignment. Without the pre-alignment stage and the data synthesis pipeline designed for it, performance drops significantly in terms of TMQI. Fig. 11 (a) also shows that the model fails to implicitly align large motions. Then, we replace the decompose-and-fuse control branch (DFCB) with the vanilla ControlNet [65], and it fails to fuse features with large exposure differences, which leads to the loss of details (see Fig. 11 (b)) and a decrease in TMQI. Finally, when we exclude the fidelity control branch (FCB), Ultra-Fusion loses the capability to maintain the detailed structure and vivid color, as indicated by Fig. 11 (c).", + "bbox": [ + 511, + 523, + 905, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Alignment strategy. Second, we perform ablation of the proposed alignment strategy, which is the most critical design for our framework. Our alignment strategy mainly consists of a data synthesis pipeline at training time and a pre-alignment module at testing time. Without both of them, the whole model degrades to a multi-exposure fusion method as shown in Fig. 12 (c). Performing pre-alignment can reduce", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16117", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5abb710dbb0e448dd45cb1770ae17b51a45e59abde7b099cebaa43fd2be006ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 90, + 220, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e787c4b7688cc487b7374b6bb6f2522faf7659013dde49b96d56eef566d19ac8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 159, + 220, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6e0f7a164f0a723e511fa6f2a2dde3cab4370632af847f02973953b46220f0ce.jpg", + "image_caption": [ + "Under-Exposure Over-Exposure" + ], + "image_footnote": [], + "bbox": [ + 94, + 223, + 220, + 270 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8b152c2bbf15ea46b11b42ef9e184c4d92cce9513e098abbfab65abf19dbe7f9.jpg", + "image_caption": [ + "Align Strategy" + ], + "image_footnote": [], + "bbox": [ + 94, + 279, + 220, + 327 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/acc38ca9ba787f3e692385b170efda8b705adbd9f32bc6db7d501a053f18f9b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 90, + 480, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/024f5917bf46a81601c7326afcd47098d80a7b0b6624c032a70d8d8dd8f4d399.jpg", + "image_caption": [ + "Under-Exposure Over-Exposure" + ], + "image_footnote": [], + "bbox": [ + 223, + 223, + 349, + 270 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a9a3606b36d94eb2244055ecd14827cb0787d0afc33163268628c466fc9a0669.jpg", + "image_caption": [ + "w/o DFCB", + "w/o DFCB Ours", + "w/o FCB" + ], + "image_footnote": [], + "bbox": [ + 223, + 279, + 349, + 325 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0ed35ecb9248ccc15261403a4788d48b614a161c45496a7dedd22d482d0741ba.jpg", + "image_caption": [ + "Under-Exposure" + ], + "image_footnote": [], + "bbox": [ + 357, + 223, + 418, + 270 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a0584271bf223fb9120b7cbe2268dcfea0b8b1e6238c885c3098bb2a6243505b.jpg", + "image_caption": [ + "Over-Exposure" + ], + "image_footnote": [], + "bbox": [ + 419, + 223, + 478, + 270 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c10b1b7f5e11a52a91f74342a4176e312873317f2edf197febdeb6239222bf0b.jpg", + "image_caption": [ + "Figure 11. Visual results of ablating each key component of our method. Each key component contributes to the final results.", + "(a) Under Exposure" + ], + "image_footnote": [], + "bbox": [ + 94, + 378, + 222, + 446 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3a56d6a8e5d4d08ace351fa76c5620bab4e4ec3e9a22c4c92f328bfaa425edf9.jpg", + "image_caption": [ + "(b) Over Exposure" + ], + "image_footnote": [], + "bbox": [ + 222, + 378, + 349, + 446 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cc8d9ed2386622e856315b3863ae14254d27e91cbe1e66801442da63f7f09fd2.jpg", + "image_caption": [ + "(c) $w / o$ Data Syn. & Pre-Alignment" + ], + "image_footnote": [], + "bbox": [ + 351, + 378, + 480, + 446 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6e8091871fb133a04c7174c7f44afe54d8861bc0a0db0eefd4837526ec2c7eb7.jpg", + "image_caption": [ + "(d) $w / o$ Data Syn.", + "Figure 12. Effectiveness of our alignment strategy." + ], + "image_footnote": [], + "bbox": [ + 94, + 455, + 222, + 522 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8672368948b28c961db0260e2c6c7affda4d6a2b9bd28200b0f3f6477080f800.jpg", + "image_caption": [ + "(e) $w / o$ Pre-Alignment" + ], + "image_footnote": [], + "bbox": [ + 222, + 455, + 349, + 522 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a9a6fcbf963dd66ee7dd71cd7c787b29ff98798b7e186771ac1656ebf13bec01.jpg", + "image_caption": [ + "(f) Ours Alignment Strategy" + ], + "image_footnote": [], + "bbox": [ + 351, + 455, + 480, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "artifacts to some extent (see Fig. 12 (d)), but may still be prone to alignment errors. Our data synthesis pipeline can improves the robustness of our algorithm to unaligned conditions, but motion in the occluded regions still cannot be solved, as shown in Fig. 12. After combining data synthesis pipeline and pre-alignment stage, our algorithm demonstrates strong capability towards dynamic scenes.", + "bbox": [ + 89, + 551, + 482, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Decompose-and-fuse control branch. At last, we evaluate how the form of soft guidance provided by the underexposed image impacts the recovery of the highlight regions. When using the RGB under-exposed image $I_{ue}$ as guidance, the model will ignore some details in the output (Fig. 13 (c)). Replacing $I_{ue}$ with its structure information $S_{ue}$ retains more details (Fig. 13 (d)), but without generating vivid color. By incorporating both under-exposed color information $C_{ue}$ and structure information $S_{ue}$ , the reconstructed results can maintain more details and color consistency (see Fig. 13 (e)). At last, multi-scale cross-attention further improves the fusion result (see Fig. 13 (f)).", + "bbox": [ + 89, + 657, + 482, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Application on general image fusion", + "text_level": 1, + "bbox": [ + 89, + 848, + 401, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "One advantage of the UltraFusion is that it can be extended to general image fusion, thanks to the flexibility of our", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d2e126ab4c4120a57e9b0c1262da9e33a36254f8a2d6d8a19706c9077e7584e7.jpg", + "image_caption": [ + "(a) Under Exposure" + ], + "image_footnote": [], + "bbox": [ + 516, + 90, + 645, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1234e3eac74e46d5b3d1ddd6537297a776e49e8a65e0af8cf33e6d103351186f.jpg", + "image_caption": [ + "(b) Over Exposure" + ], + "image_footnote": [], + "bbox": [ + 645, + 90, + 774, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e70afd959c2dfca63614c9e5376b855b8494bbfea371877023ac947966c1f2fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 90, + 903, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/418f6375d1bb490a9b7f192e3db8751a8d4df423921c802fa2d980cb5bad9348.jpg", + "image_caption": [ + "(d) $S_{ue}$" + ], + "image_footnote": [], + "bbox": [ + 516, + 166, + 645, + 232 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0e7e24886d1a941348040818f60e36884a30d1ac14e1cc1eb74b10cd7e609740.jpg", + "image_caption": [ + "(e) $S_{ue} + C_{ue}$" + ], + "image_footnote": [], + "bbox": [ + 645, + 166, + 774, + 232 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/52273cdf2c8c480e0843f8f4c08895e69350fea189c8c30ad70cff61494c1731.jpg", + "image_caption": [ + "(c) $I_{ue}$", + "(f) DFCB" + ], + "image_footnote": [], + "bbox": [ + 776, + 166, + 903, + 232 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/93a6f208f33de5c9abd764e12cb155a7145b6c512100e50a6a7eb6496a34a3bb.jpg", + "image_caption": [ + "Figure 13. Detailed ablation study on the design choices of decompose-and-fuse control branch.", + "(a) Input Over Exposure", + "Figure 14. Extension to general fusion. Given totally different underexposed images as guidance (upper right corner), our method can also generate different fusion results." + ], + "image_footnote": [], + "bbox": [ + 514, + 282, + 640, + 349 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/34b8effd521b6019697b2527d039a37a14246e5cc88117b100cd7084c79e6305.jpg", + "image_caption": [ + "(b) Our Fusion Result", + "(c) Another Fusion w/ Diff. Highlight" + ], + "image_footnote": [], + "bbox": [ + 640, + 284, + 766, + 349 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/766deb6635d7215f24f16182405356b0287fcd7dccdefca626f397f52bb120e1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 284, + 893, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "guided inpainting. To illustrate this potential, we explore one additional interesting demo to fuse two irrelevant images captured by different cameras at different locations. As shown in Fig. 14, UltraFusion successfully copies the moon (b) or blue sky (c) to the over-exposed image, unlocking many interesting potential applications (e.g., image harmonization).", + "bbox": [ + 511, + 415, + 905, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 531, + 633, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce a novel approach to HDR imaging, tackling challenges presented by significant exposure differences and large motion. By modeling the fusion process as a guided inpainting problem and utilizing the under-exposed image for soft guidance, our method is robust with alignment errors and circumvents tone mapping, resulting in natural, artifact-free outputs. We also propose decompose-and-fuse control branch and fidelity control branch to improve feature modulation and fidelity preservation of ControlNet. Extensive experiments on existing datasets and our captured benchmark demonstrate the robustness and effectiveness of our method compared to previous HDR methods.", + "bbox": [ + 511, + 551, + 905, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With extreme exposure differences and challenging non-rigid motion, occlusion mask estimation may introduce errors, causing the restoration of certain highlight regions degrades to single image HDR. While our method is able to use diffusion priors to restore these highlights, restoration without under-exposed information remains unreliable. Moreover, it takes almost 3.3s to fuse $512 \\times 512$ inputs on an NVIDIA RTX 4090 GPU. A more exposure-robust optical flow algorithm and a faster implementation are highly desirable, and we leave them for future work.", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16118", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Key R&D Program of China No.2022ZD0160201, Shanghai Artificial Intelligence Laboratory and RGC Early Career Scheme (ECS) No. 24209224.", + "bbox": [ + 89, + 114, + 482, + 175 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 191, + 186, + 207 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jianrui Cai, Shuhang Gu, and Lei Zhang. Learning a deep single image contrast enhancer from multi-exposure images. IEEE Transactions on Image Processing, 27(4):2049-2062, 2018. 3, 5, 7", + "[2] Cong Cao, Huanjing Yue, Xin Liu, and Jingyu Yang. Unsupervised HDR image and video tone mapping via contrastive learning. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 3", + "[3] Jie Chen, Zaifeng Yang, Tsz Nam Chan, Hui Li, Junhui Hou, and Lap-Pui Chau. Attention-guided progressive neural texture fusion for high dynamic range image restoration. IEEE Transactions on Image Processing, 31:2661-2672, 2022. 2, 3", + "[4] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024. 3", + "[5] Ciprian Corneanu, Raghudeep Gadde, and Aleix M Martinez. Latentpaint: Image inpainting in latent space with diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4334-4343, 2024. 3", + "[6] Abhishek Goswami, Aru Ranjan Singh, Francesco Banterle, Kurt Debattista, and Thomas Bashford-Rogers. Semantic aware diffusion inverse tone mapping. arXiv preprint arXiv:2405.15468, 2024.3", + "[7] Michael D Grossberg and Shree K Nayar. Determining the camera response from images: What is knowable? IEEE Transactions on pattern analysis and machine intelligence, 25(11):1455-1467, 2003. 3", + "[8] Yuanshen Guan, Ruikang Xu, Mingde Yao, Ruisheng Gao, Lizhi Wang, and Zhiwei Xiong. Diffusion-promoted HDR video reconstruction. arXiv preprint arXiv:2406.08204, 2024.3", + "[9] Samuel W Hasinoff, Dillon Sharlet, Ryan Geiss, Andrew Adams, Jonathan T Barron, Florian Kainz, Jiawen Chen, and Marc Levoy. Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM Transactions on Graphics (ToG), 35(6):1-12, 2016. 1, 2", + "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3", + "[11] Litao Hu, Huaijin Chen, and Jan P Allebach. Joint multiscale tone mapping and denoising for HDR image enhancement. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 729-738, 2022. 3" + ], + "bbox": [ + 93, + 215, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Tao Hu, Qingsen Yan, Yuankai Qi, and Yanning Zhang. Generating content for HDR deghosting from frequency view. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25732-25741, 2024. 3", + "[13] Ting Jiang, Chuan Wang, Xinpeng Li, Ru Li, Haoqiang Fan, and Shuaicheng Liu. Meflut: Unsupervised 1d lookup tables for multi-exposure image fusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10542-10551, 2023. 3, 6", + "[14] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. arXiv preprint arXiv:2403.06976, 2024. 3", + "[15] Nima Khademi Kalantari, Ravi Ramamoorthi, et al. Deep high dynamic range imaging of dynamic scenes. ACM Trans. Graph., 36(4):144-1, 2017. 2, 3, 5", + "[16] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 5", + "[17] Lingtong Kong, Bo Li, Yike Xiong, Hao Zhang, Hong Gu, and Jinwei Chen. Safnet: Selective alignment fusion network for efficient HDR imaging. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2, 3, 6", + "[18] Baiang Li, Sizhuo Ma, Yanhong Zeng, Xiaogang Xu, Youqing Fang, Zhao Zhang, Jian Wang, and Kai Chen. Sagiri: Low dynamic range image enhancement with generative diffusion prior. arXiv preprint arXiv:2406.09389, 2024. 3", + "[19] Hui Li, Kede Ma, Hongwei Yong, and Lei Zhang. Fast multiscale structural patch decomposition for multi-exposure image fusion. IEEE Transactions on Image Processing, 29: 5805-5816, 2020. 3", + "[20] Pengwei Liang, Junjun Jiang, Xianming Liu, and Jiayi Ma. Fusion from decomposition: A self-supervised decomposition approach for image fusion. In European Conference on Computer Vision, pages 719-735. Springer, 2022. 6", + "[21] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Wanli Ouyang, Yu Qiao, and Chao Dong. Diff-bir: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070, 2023. 3, 4", + "[22] Renshuai Liu, Chengyang Li, Haitao Cao, Yinglin Zheng, Ming Zeng, and Xuan Cheng. Emef: ensemble multi-exposure image fusion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1710-1718, 2023. 3, 5", + "[23] Yuhao Liu, Zhanghan Ke, Fang Liu, Nanxuan Zhao, and Rynson WH Lau. Diff-plugin: Revitalizing details for diffusion-based low-level tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4197-4208, 2024. 3", + "[24] Zhen Liu, Yinglong Wang, Bing Zeng, and Shuaicheng Liu. Ghost-free high dynamic range imaging with context-aware transformer. In European Conference on computer vision, pages 344-360. Springer, 2022. 2, 3", + "[25] Zhen Liu, Yinglong Wang, Bing Zeng, and Shuaicheng Liu. Ghost-free high dynamic range imaging with context-aware" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16119", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "transformer. In European Conference on computer vision, pages 344-360. Springer, 2022. 1, 6", + "[26] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 3", + "[27] Kede Ma, Kai Zeng, and Zhou Wang. Perceptual quality assessment for multi-exposure image fusion. IEEE Transactions on Image Processing, 24(11):3345-3356, 2015. 5", + "[28] Kede Ma, Hui Li, Hongwei Yong, Zhou Wang, Deyu Meng, and Lei Zhang. Robust multi-exposure image fusion: a structural patch decomposition approach. IEEE Transactions on Image Processing, 26(5):2519-2532, 2017. 3", + "[29] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 3", + "[30] Tom Mertens, Jan Kautz, and Frank Van Reeth. Exposure fusion. In 15th Pacific Conference on Computer Graphics and Applications (PG'07), pages 382-390. IEEE, 2007. 2, 3", + "[31] Ozan Özdenizci and Robert Legenstein. Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10346-10357, 2023. 3", + "[32] Photomatix. Commercially-available hdr processing software. https://www.hdrsoft.com/.6", + "[33] K Ram Prabhakar, Rajat Arora, Adhitya Swaminathan, Kunal Pratap Singh, and R Venkatesh Babu. A fast, scalable, and reliable deghosting method for extreme exposure fusion. In 2019 IEEE International Conference on Computational Photography (ICCP), pages 1-8. IEEE, 2019. 3", + "[34] K Ram Prabhakar, Susmit Agrawal, Durgesh Kumar Singh, Balraj Ashwath, and R Venkatesh Babu. Towards practical and efficient high-resolution HDR deghosting with cnn. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16, pages 497-513. Springer, 2020. 2, 3", + "[35] K Ram Prabhakar, V Sai Srikar, and R Venkatesh Babu. Deepfuse: A deep unsupervised approach for exposure fusion with extreme exposure image pairs. In Proceedings of the IEEE international conference on computer vision, pages 4714-4722, 2017. 3, 6", + "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 3", + "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 5", + "[38] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition, pages 8839-8849, 2024. 3", + "[39] Yong Shu, Liquan Shen, Xiangyu Hu, Mengyao Li, and Zihao Zhou. Towards real-world HDR video reconstruction: A large-scale benchmark dataset and a two-stage alignment network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2879-2888, 2024. 3, 5, 6, 7", + "[40] Jou Won Song, Ye-In Park, Kyeongbo Kong, Jaeho Kwak, and Suk-Ju Kang. Selective transhdr: Transformer-based selective HDR imaging using ghost region mask. In European Conference on Computer Vision, pages 288-304. Springer, 2022. 2, 3", + "[41] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3667-3676, 2020. 5", + "[42] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 3, 5", + "[43] Steven Tel, Zongwei Wu, Yulun Zhang, Barthélémy Heyrman, Cédric Demonceaux, Radu Timofte, and Dominique Gingac. Alignment-free HDR deghosting with semantics consistent transformer. arXiv preprint arXiv:2305.18135, 2023. 2, 3", + "[44] Steven Tel, Zongwei Wu, Yulun Zhang, Barthélémy Heyrman, Cédric Demonceaux, Radu Timofte, and Dominique Gingac. Alignment-free HDR deghosting with semantics consistent transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12836-12845, 2023. 5, 6", + "[45] Yael Vinker, Inbar Huberman-Spiegelglas, and Raanan Fattal. Unpaired learning for high dynamic range image tone mapping. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14657-14666, 2021. 3", + "[46] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 4", + "[47] Shuchen Weng, Peixuan Zhang, Yu Li, Si Li, Boxin Shi, et al. L-cad: Language-based colorization with any-level descriptions using diffusion priors. Advances in Neural Information Processing Systems, 36, 2024. 3, 4", + "[48] Guanyao Wu, Hongming Fu, Jinyuan Liu, Long Ma, Xin Fan, and Risheng Liu. Hybrid-supervised dual-search: Leveraging automatic learning for loss-free multi-exposure image fusion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5985-5993, 2024. 1, 3, 6", + "[49] Shangzhe Wu, Jiarui Xu, Yu-Wing Tai, and Chi-Keung Tang. Deep high dynamic range imaging with large foreground motions. In Proceedings of the European Conference on Computer Vision (ECCV), pages 117–132, 2018. 2, 3", + "[50] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16120", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22428-22437, 2023. 3", + "[51] Gangwei Xu, Yujin Wang, Jinwei Gu, Tianfan Xue, and Xin Yang. Hdrflow: Real-time HDR video reconstruction with large motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24851-24860, 2024. 3", + "[52] Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502-518, 2020. 3, 6", + "[53] Han Xu, Jiayi Ma, and Xiao-Ping Zhang. Mef-gan: Multiexposure image fusion via generative adversarial networks. IEEE Transactions on Image Processing, 29:7203-7216, 2020. 6", + "[54] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8121-8130, 2022. 3", + "[55] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. International Journal of Computer Vision, 127:1106-1125, 2019. 5", + "[56] Qingsen Yan, Dong Gong, Qinfeng Shi, Anton van den Hengel, Chunhua Shen, Ian Reid, and Yanning Zhang. Attention-guided network for ghost-free high dynamic range imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1751-1760, 2019. 2, 3", + "[57] Qingsen Yan, Lei Zhang, Yu Liu, Yu Zhu, Jinqiu Sun, Qinfeng Shi, and Yanning Zhang. Deep HDR imaging via a non-local network. IEEE Transactions on Image Processing, 29: 4308-4322, 2020. 2, 3", + "[58] Qingsen Yan, Tao Hu, Yuan Sun, Hao Tang, Yu Zhu, Wei Dong, Luc Van Gool, and Yanning Zhang. Towards high-quality HDR dehosting with conditional diffusion models. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 3", + "[59] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3", + "[60] Hojatollah Yeganeh and Zhou Wang. Objective quality assessment of tone-mapped images. IEEE Transactions on Image processing, 22(2):657-667, 2012. 5", + "[61] Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3575–3585, 2020. 5", + "[62] Zhiyuan You, Xin Cai, Jinjin Gu, Tianfan Xue, and Chao Dong. Teaching large language models to regress accurate image quality scores using score distribution. arXiv preprint arXiv:2501.11561, 2025. 5" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photorealistic image restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25669-25680, 2024. 3", + "[64] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 3", + "[65] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3, 4, 6, 7", + "[66] Xingchen Zhang. Benchmarking and comparing multi-exposure image fusion algorithms. Information Fusion, 74: 111-131, 2021. 5, 6", + "[67] Shengyu Zhao, Yilun Sheng, Yue Dong, Eric I Chang, Yan Xu, et al. Maskflownet: Asymmetric feature matching with learnable occlusion mask. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6278-6287, 2020. 3", + "[68] Zixiang Zhao, Lilun Deng, Haowen Bai, Yukun Cui, Zhipeng Zhang, Yulun Zhang, Haotong Qin, Dongdong Chen, Jiangshe Zhang, Peng Wang, et al. Image fusion via vision-language model. arXiv preprint arXiv:2402.02235, 2024.3", + "[69] Pengfei Zhu, Yang Sun, Bing Cao, and Qinghua Hu. Task-customized mixture of adapters for general image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7099-7108, 2024. 3, 6" + ], + "bbox": [ + 516, + 90, + 906, + 542 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16121", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/UltraFusion_ Ultra High Dynamic Imaging using Exposure Fusion/c2c26ec1-c8af-49f8-87ba-b3c749531067_model.json b/2025/UltraFusion_ Ultra High Dynamic Imaging using Exposure Fusion/c2c26ec1-c8af-49f8-87ba-b3c749531067_model.json new file mode 100644 index 0000000000000000000000000000000000000000..95787df593d121bb72a35dbb678d5a20530449a9 --- /dev/null +++ b/2025/UltraFusion_ Ultra High Dynamic Imaging using Exposure Fusion/c2c26ec1-c8af-49f8-87ba-b3c749531067_model.json @@ -0,0 +1,3181 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.165, + 0.131, + 0.834, + 0.154 + ], + "angle": 0, + "content": "UltraFusion: Ultra High Dynamic Imaging using Exposure Fusion" + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.179, + 0.857, + 0.27 + ], + "angle": 0, + "content": "Zixuan Chen\\(^{1,3*}\\) Yujin Wang\\(^{1*}\\) Xin Cai\\(^{2}\\) Zhiyuan You\\(^{2}\\) Zheming Lu\\(^{3}\\) Fan Zhang\\(^{1}\\) Shi Guo\\(^{1}\\) Tianfan Xue\\(^{2,1}\\) \n\\(^{1}\\)Shanghai AI Laboratory \\(^{2}\\)The Chinese University of Hong Kong \\(^{3}\\)Zhejiang University \\(\\{\\mathrm{zxchen}, \\mathrm{zheminglu}\\} @ \\mathrm{zju.edu.cn}, \\{\\mathrm{wangyujin}, \\mathrm{zhangfan}, \\mathrm{guoshi}\\} @ \\mathrm{pjlab.org.cn}\\) caixin@link.cuhk.edu.hk, zhiyuanyou@foxmail.com, tfxue@ie.cuhk.edu.hk" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.283, + 0.226, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.354, + 0.226, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.423, + 0.17, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.482, + 0.168, + 0.491 + ], + "angle": 0, + "content": "Under Exposure" + }, + { + "type": "image", + "bbox": [ + 0.171, + 0.423, + 0.246, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.177, + 0.482, + 0.24, + 0.491 + ], + "angle": 0, + "content": "Over Exposure" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.283, + 0.495, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.423, + 0.341, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.274, + 0.482, + 0.339, + 0.499 + ], + "angle": 0, + "content": "Previous HDR Reconstruction" + }, + { + "type": "image", + "bbox": [ + 0.344, + 0.423, + 0.417, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.482, + 0.415, + 0.5 + ], + "angle": 0, + "content": "Exposure Multi-" + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.423, + 0.495, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.434, + 0.482, + 0.483, + 0.5 + ], + "angle": 0, + "content": "Our Guided Inpainting" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.283, + 0.632, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.354, + 0.633, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.423, + 0.577, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.482, + 0.575, + 0.49 + ], + "angle": 0, + "content": "Under Exposure" + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.423, + 0.653, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.584, + 0.482, + 0.647, + 0.49 + ], + "angle": 0, + "content": "Over Exposure" + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.283, + 0.903, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.423, + 0.753, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.482, + 0.749, + 0.499 + ], + "angle": 0, + "content": "Previous HDR Reconstruction" + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.423, + 0.828, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.482, + 0.827, + 0.499 + ], + "angle": 0, + "content": "Previous MultiExposure Fusion" + }, + { + "type": "image", + "bbox": [ + 0.83, + 0.423, + 0.903, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.842, + 0.482, + 0.89, + 0.5 + ], + "angle": 0, + "content": "Our Guided Inpainting" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.513, + 0.907, + 0.569 + ], + "angle": 0, + "content": "Figure 1. Comparing our guided inpainting HDR imaging method with a state-of-the-art HDR reconstruction [25] and multi-exposure fusion [48] methods. Both scenes are selected from our captured new real-world benchmark. Left: night cityscape with large exposure difference. Right: afternoon street with motion-induced occlusion. Previous methods struggle to handle these scenes. By modeling HDR as an inpainting problem, our method can produce visually appealing results without ghosting artifacts in these challenging scenes." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.581, + 0.327, + 0.597 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.608, + 0.486, + 0.864 + ], + "angle": 0, + "content": "Capturing high dynamic range (HDR) scenes is one of the most important issues in camera design. Majority of cameras use exposure fusion, which fuses images captured by different exposure levels, to increase dynamic range. However, this approach can only handle images with limited exposure difference, normally 3-4 stops. When applying to very high dynamic range scenes where a large exposure difference is required, this approach often fails due to incorrect alignment or inconsistent lighting between inputs, or tone mapping artifacts. In this work, we propose UltraFusion, the first exposure fusion technique that can merge inputs with 9 stops differences. The key idea is that we model exposure fusion as a guided inpainting problem, where the under-exposed image is used as a guidance to fill the missing information of over-exposed highlights in the overexposed region. Using an under-exposed image as a soft guidance, instead of a hard constraint, our model is robust" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.907, + 0.765 + ], + "angle": 0, + "content": "to potential alignment issue or lighting variations. Moreover, by utilizing the image prior of the generative model, our model also generates natural tone mapping, even for very high-dynamic range scenes. Our approach outperforms HDR-Transformer on latest HDR benchmarks. Moreover, to test its performance in ultra high dynamic range scenes, we capture a new real-world exposure fusion benchmark, UltraFusion dataset, with exposure differences up to 9 stops, and experiments show that UltraFusion can generate beautiful and high-quality fusion results under various scenarios. Code and data will be available at https://openimaginglab.github.io/UltraFusion." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.784, + 0.646, + 0.799 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.902 + ], + "angle": 0, + "content": "High dynamic range (HDR) imaging is one of the fundamental problems in the modern camera design. Due to hardware limitation, camera sensors have a much smaller dynamic range compared with the real-world. To increase it, majority of HDR solutions merge multiple images, either with the same [9] or different exposure lev" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution. This work was done during Zixuan Chen's internship at Shanghai Artificial Intelligence Laboratory." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16111" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.23, + 0.483, + 0.3 + ], + "angle": 0, + "content": "Figure 2. Visual comparison on directly utilizing ControlNet [65] and our UltraFusion. Without pre-aligned data, ControlNet struggles to fix a frame for reference. However, our method fixes the over-exposed image as the reference and the under-exposed image as the guidance for inpainting, thereby avoiding artifacts." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.31, + 0.483, + 0.446 + ], + "angle": 0, + "content": "els [3, 15, 17, 24, 34, 40, 43, 49, 56, 57]. Despite all recent advances, majority HDR algorithms can only bring limited increase in dynamic range. For example, \\(\\mathrm{HDR + }\\) [9], the first HDR algorithm used by commercial cameras, can only robustly increase the dynamic range by 8 times (3 stops). Therefore, in this work, we study the following question: can we drastically increase the dynamic range of a camera by fusing two images with very large exposure difference, like 9 stops in Fig. 1?" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.447, + 0.483, + 0.734 + ], + "angle": 0, + "content": "This is a fundamental challenging problem, due to the following three issues. First, to handle dynamic scenes, most of HDR fusion algorithms will first align input frames, which is very challenging when input has large brightness difference. As result, a ghosting issue happens when alignment fails, indicated by zoom-in patches in the right scene of Fig. 1. Second, most of HDR algorithms assume that under-exposed image is simply a darker version of the normal image. However, the appearance of an object may change when exposure levels change, like ship in the left scene of Fig. 1, resulting unnatural fusion result. Third, sometimes the result of fusion is an HDR image, which cannot be directly shown on normal low-dynamic-range display. Therefore, these HDR images will be further compressed through a tone-mapping process. When dynamic range is high, tone-mapping may introduce additional issue. Maintaining a natural contrast and rich details in the final output is challenging, as shown in the zoom-in patches of previous HDR reconstruction method in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose a completely different fusion method, UltraFusion, which models it as a guided inpainting problem. In this setup, the user captures two images, one normal exposed image where brighter objects are overexposed, and another under exposed image, which only captures the very brighter parts of the scene. We use the normal exposed image as a reference, and inpainting the missing information in the highlight. Unlike the traditional inpainting, we use the information from under-exposed frame as guidance, so inpainted highlight is not completely generated, but stays consistent with under-exposed frame." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.304 + ], + "angle": 0, + "content": "There are three advantages of this approach when handling large exposure difference. First, it follows the exposure fusion [30] setup. Unlike HDR fusion techniques, which first generate HDR image and then compress to a low-dynamic-range (LDR) image through tone mapping, exposure fusion directly generates LDR output, avoiding cascading errors. That being said, our method outputs a tone-mapped LDR image instead of a linear HDR. Second, the under-exposed image is used as a soft guidance, instead of a hard constraint. Therefore, UltraFusion is robust to alignment error (see the right scene of Fig. 1) and lighting variation (see the left scene of Fig. 1). Third, image prior of the generative model ensures the natural look of the output image, reducing the potential artifacts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.306, + 0.905, + 0.488 + ], + "angle": 0, + "content": "To train a guided inpainting, a simple solution is to train a ControlNet [65] using two input images. However, such paradigm cannot handle dynamic HDR scenes, as ControlNet may not know which frame to choose as the reference, increasing the difficulty to fuse the information from over and short-exposed images. For instance, as shown in Fig. 2, ControlNet selects the over-exposed image as the reference frame in red boxed region, but chooses the under-exposed image as the reference frame in green boxed region, leading to substantial artifacts in the result. Additionally, as a generative model, it may inevitably generate fake image content, as indicated by green box in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.905, + 0.777 + ], + "angle": 0, + "content": "To address these challenges, we design UltraFusion as follows. First, we warp the short-exposed image to the long-exposed one, and mask out the occluded regions. Then, we utilize the diffusion prior to inpaint the long-exposed image guided by partial short-exposed information. To make the network retain more details to generate guidance information, we propose a new decompose-and-fuse control branch, which eliminates the luminance component of the short-exposed image, extracting structure and color information instead, and employs a multi-scale cross-attention to improve the feature fusion with long-exposed image. Second, as there is no existing large-scale training data for exposure fusion of dynamic scene, we also propose a novel training data synthesis pipeline, utilizing existing high-quality multi-exposure datasets (pre-aligned) of static scene and video datasets. At last, to ensure the generated output maintains the fidelity to the reality, we also train an additional fidelity control branch, using the same decompose-and-fuse strategy and multi-scale cross-attention." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.902 + ], + "angle": 0, + "content": "At last, to evaluate the effectiveness of our framework, we capture 100 under/over-exposed image pairs, covering daytime, nighttime, indoor, outdoor scenes with local and global motion patterns. Experiments on both latest HDR imaging datasets and our captured benchmark demonstrate that, comparing to existing methods, our UltraFusion is more robust to scenes with large exposure differences and large motion, as shown in Fig. 1." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16112" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.228, + 0.106 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.117, + 0.236, + 0.133 + ], + "angle": 0, + "content": "2.1. HDR imaging" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.141, + 0.483, + 0.429 + ], + "angle": 0, + "content": "HDR Imaging can be deided into HDR reconstruction and Multi-Exposure Fusion (MEF) typically depends on the domain where fusion occurs [19]. HDR reconstruction methods [3, 15, 17, 24, 34, 40, 43, 49, 51, 56, 57] inverts the camera response function (CRF) to merge exposure brackets in the linear HDR domain [28]. In most cases, tone mapping is necessary to display the reconstructed HDR image properly on standard LDR monitors. As a cost-effective alternative [13], MEF methods [13, 19, 22, 28, 30, 33, 35, 48, 52, 68, 69] directly fuse images in the LDR domain, sidestepping CRF calibration and sophisticated tone mapping process [30]. Regardless of the type of HDR Imaging methods, they contend with ghosting artifacts caused by camera shake and object movement [39]. Previous methods have attempted explicit or implicit alignment using optical flow [15, 33, 34, 49] or attention mechanism [3, 17, 24, 40, 43, 56, 57]. However, most HDR Imaging methods suffer from unpleasing artifacts when large motion causes occlusion in the complementary region." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.431, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Diffusion models. Recently, it has been witnessed with the rapid rise of diffusion models [10, 37] and their successful application in various tasks, including controllable image generation [59, 65], image restoration [21, 23, 31, 47, 63], image editing [4, 14, 29, 38] and image inpainting [5, 26, 50, 64]. In the field of HDR imaging, the application of diffusion models has primarily focused on HDR deghosting [8, 12, 58]. However, since these methods do not leverage the diffusion priors learned from large-scale datasets, their generalization ability is limited by the scale of the HDR dataset. While some recent works [6, 18] have employed diffusion priors, they tend to focus on single image HDR. Without another differently-exposed image as reference, the results generated by these methods lack sufficient reliability. Unlike previous methods, we utilize diffusion priors and use the short-exposed image as a reference to perform reliable inpainting in the highlight regions of the overexposed image, thereby achieving natural and reliable HDR scene reconstruction. Compared to diffusion-based inpainting methods [5, 26, 50, 64] that perform inpainting solely from scratch, we leverage information from short-exposure images to guide a more accurate inpainting process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Tone mapping methods. The goal of tone mapping is to convert HDR images to LDR for display on standard screens while enhancing visual detail. Due to the challenge of obtaining ground-truth tone mapping results, unsupervised deep learning approaches have been developed using adversarial [45] and contrastive [2] learning. To address data limitations, Cai et al. [1] manually curated training data by selecting the best results from 13 tone mapping methods for network learning [1, 11]. However, previous tone" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.183 + ], + "angle": 0, + "content": "mapping algorithms, lacking robust image priors and facing data constraints, struggle with visually pleasing results and generalization in extreme high dynamic range scenes. By incorporating diffusion-based image priors, our method achieves aesthetic results even in challenging high dynamic range scenarios (see Fig. 1)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.195, + 0.649, + 0.212 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.219, + 0.905, + 0.355 + ], + "angle": 0, + "content": "Given an over-exposed image \\( I_{oe} \\) and an under-exposed image \\( I_{ue} \\), traditional exposure fusion algorithms directly aggregate different frequency band of both images, which are sensitive to misalignment error or lighting variation. Instead, we treat this as an inpainting problem. Specifically, we use the over-exposed image \\( I_{oe} \\) as the base image and inpainting missing information in the highlight region. To ensure inpainted highlights are real, we also use highlights from under-exposed image as guidance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.356, + 0.906, + 0.446 + ], + "angle": 0, + "content": "Based on this idea, we design a 2-stage newtwork shown in Fig. 3, which consists of the pre-alignment stage and the guided inpainting stage. The pre-alignment stage outputs a coarse-aligned version of \\( I_{ue} \\), which is used as the soft guidance in the following guided inpainting stage. Details of each stage is described below." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.454, + 0.704, + 0.47 + ], + "angle": 0, + "content": "3.1. Pre-alignment stage" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.906, + 0.673 + ], + "angle": 0, + "content": "Most of optical flow alignment assumes input have the similar brightness. Therefore, we first adjust the brightness \\( I_{ue} \\) to match the distribution of \\( I_{oe} \\) through intensity mapping function [7]. Then, we adopt RAFT [42], a pre-trained optical flow network, to estimate the bidirectional flow \\( f_{oe \\rightarrow ue} \\) and \\( f_{ue \\rightarrow oe} \\) and align \\( I_{ue} \\) to \\( I_{oe} \\) using backward warping. However, backward warping will result in ghosting at the occlusion boundary [67], leading to artifacts in the next guided inpainting stage. To solve that, we utilize the forward-and-backward consistency check [54] to estimate occluded regions \\( \\mathcal{M} \\) and mask them out in the warped output. Finally, we can obtain a pre-aligned output \\( I_{ue \\rightarrow oe} \\) of the first stage:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.681, + 0.905, + 0.699 + ], + "angle": 0, + "content": "\\[\nI _ {u e \\rightarrow o e} = (1 - \\mathcal {M}) \\cdot \\mathcal {W} \\left(I _ {u e}, f _ {o e \\rightarrow u e}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.905, + 0.736 + ], + "angle": 0, + "content": "where \\(\\mathcal{W}\\) denotes the backward warping. Fig. 3 (a) shows the output is a masked and aligned under-exposed image." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.734, + 0.759 + ], + "angle": 0, + "content": "3.2. Guided inpainting stage" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We build our guided inpainting model based on the Stable Diffusion Model [36] because its powerful generative prior can help to resolve ambiguity during inpainting. Similar to other diffusion-based image enhancement techniques [21], we also inject the following information through an additional control branch, as shown in Fig. 3(b): 1) the image to be inpainted, which is the over-exposed image \\( I_{oe} \\), 2) the additional guidance of highlight, which is the underexposed image \\( I_{ue} \\), and 3) the diffusion latent at the current" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16113" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.093, + 0.384, + 0.108 + ], + "angle": 0, + "content": "(a) Pre-Alignment Stage" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.117, + 0.493, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.598, + 0.093, + 0.789, + 0.108 + ], + "angle": 0, + "content": "(b) Guided Inpainting Stage" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.106, + 0.904, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.289, + 0.908, + 0.345 + ], + "angle": 0, + "content": "Figure 3. The whole backbone of UltraFusion. Our method is a 2-stage framework, consisting of (a) pre-alignment stage and (b) guided inpainting stage. The first stage pre-aligns the under-exposed image \\( I_{ue} \\) to the over-exposed image \\( I_{oe} \\) and masks the occluded regions. In the subsequent guided inpainting stage, we propose a new decompose-and-fuse control branch to utilize the diffusion priors, together with a fidelity control branch for fidelity keeping." + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.357, + 0.43, + 0.373 + ], + "angle": 0, + "content": "Decompose-and-Fuse Control Branch" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.376, + 0.48, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.565, + 0.483, + 0.592 + ], + "angle": 0, + "content": "Figure 4. The detailed architecture of our proposed decompose-and-fuse control branch." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.484, + 0.702 + ], + "angle": 0, + "content": "diffusion step \\( z_{t} \\), as previous work [21] shows that including diffusion latent as an additional condition can improve image quality. The main diffusion denoising network is a pretrained U-Net. The over-exposed image is first encoded using a pretrained VAE before entering the diffusion module, and the outputs are converted back to the image space using the pretrained decoder." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.854 + ], + "angle": 0, + "content": "The main differences between our solution and general diffusion-based image enhancement are two-fold. First, we propose a novel decompose-and-fuse control branch to inject two input images and diffusion latent as a control signal, as Fig. 11 (b) shows that naively injecting this information may not be able to guide diffusion to faithfully inpaint missing highlights obtained from the under-exposed image. Second, we train an additional fidelity control branch to provide faithful structure and color information for the decoding process via shortcuts. Details are described below." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Decompose-and-fuse control branch. Fig. 4 shows our control branch. We use the over-exposed image \\( I_{oe} \\) as the main control signal and the under-exposed image \\( I_{ue} \\) as the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.352, + 0.907, + 0.473 + ], + "angle": 0, + "content": "soft guidance. Following ControlNet [65], we copy the encoder and middle blocks of the denoising U-Net as the main extractor, but update their weights during training. A simple soft guidance is to use the encoded under-exposed image latent \\( y_{ue} \\) from the VAE encoder, combined with the overexposed latent \\( y_{oe} \\). However, the underexposed images are often too dark to be directly used as soft guidance, since the model may entirely ignore this guidance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.907, + 0.534 + ], + "angle": 0, + "content": "Therefore, we decompose the under-exposed image into the color and structure information, both of which are robust to brightness changes. Specifically, we use the normalized image as the structure component, similar to SSIM [46], as:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.548, + 0.905, + 0.565 + ], + "angle": 0, + "content": "\\[\nS _ {u e} = \\left(Y _ {u e} - \\mu \\left(Y _ {u e}\\right)\\right) / \\sigma \\left(Y _ {u e}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.906, + 0.689 + ], + "angle": 0, + "content": "where \\( Y_{ue} \\) represents the luminance channel of \\( I_{ue} \\) under YUV space, \\( \\mu(\\cdot) \\) and \\( \\sigma(\\cdot) \\) denote the mean intensity and standard deviation, respectively. The chroma channels (UV) are used as color information. The extracted structure and color information are further encoded using trained color and structure extractors (gray GE block in Fig. 4). Following [65], we implement GE using several simple convolution layers to extract multi-scale features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.901 + ], + "angle": 0, + "content": "At last, the extracted features are injected into the main extractor with a multi-scale cross attention, as shown in the bottom part of Fig. 4. The output of each cross attention module is fed into both the next level of the main extractor and the corresponding U-Net block, using zero convolution. Fidelity control branch. Even with the control block, we sometimes observe undesirable modifications of texture introduced by VAE. An example is shown in Fig. 11 (c). Therefore, to further improve fidelity, we design a Fidelity Control Branch (FCB), inspired by [47]. FCB mitigates texture distortions by injecting features into the VAE decoder. It has a similar architecture as the decompose-and-fuse control branch, with two main differences: 1) the main extractor of FCB adopts the same structure as the VAE Encoder," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16114" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.488, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.219, + 0.478, + 0.233 + ], + "angle": 0, + "content": "Figure 5. The illustration of our training data synthesis pipeline." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.238, + 0.485, + 0.435 + ], + "angle": 0, + "content": "rather than the denoising U-Net, to provide corresponding shortcuts to the VAE Decoder, with adjustments to the soft guidance extractor as well, and 2) the main extractor of the FCB directly takes the over-exposed image as input. To train the fidelity control branch, we freeze the pre-trained VAE Encoder and Decoder and encode the ground truth \\( I_{gt} \\) to latent space, simulating the denoised latent \\( z_0 \\) during the inference. Then we input the corresponding overexposed image and under-exposed image to the fidelity control branch to extract faithful features. At last, the VAE Decoder decodes the compressed latent to a reconstructed image \\( \\hat{I}_{gt} \\). We adopt the reconstruction loss \\( ||\\hat{I}_{gt} - I_{gt}||_1 \\) as an additional loss term." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.443, + 0.307, + 0.459 + ], + "angle": 0, + "content": "3.3. Training data synthesis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.465, + 0.484, + 0.555 + ], + "angle": 0, + "content": "Preparing the training data for the proposed guided inpainting network is challenging. To train the model, ideally we need a large-scale HDR dataset that 1) covers different dynamic scenes, 2) has large exposure variance up to 9 stops, and 3) contains ground truth fusion result. However, no existing dataset satisfy all requirements." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.556, + 0.484, + 0.783 + ], + "angle": 0, + "content": "To solve this issue, we propose a novel training data synthesis pipeline. Specifically, as shown in Fig. 5, we first randomly sample an image sequence with \\(N\\) frames from a video dataset [55]. To model large motion, we select the first and the last frames. Then, similar to the pre-alignment stage, we estimate the bidirectional optical flow between two selected frames using [42] and compute a pseudo occlusion mask via forward-backward consistency check. Subsequently, we randomly sample an under-exposed image patch from the high-quality static dataset [1] (ground truths are pre-aligned), resize the pseudo occlusion mask to match the patch size, and mask out the pseudo occluded region to synthesize a pseudo pre-aligned output. With our synthesized training data, our UltraFusion learns to handle dynamic scenes with only static multi-exposure image pairs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.794, + 0.217, + 0.811 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.819, + 0.289, + 0.835 + ], + "angle": 0, + "content": "4.1. Experimental setting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Datasets. We utilize SICE [1] dataset and Video-90K [55] dataset to synthesize our training data. Following previous works [1, 22], we select images with the highest and lowest brightness in each exposure bracket of the SICE dataset" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.089, + 0.905, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.178, + 0.907, + 0.22 + ], + "angle": 0, + "content": "Figure 6. (a) The data distribution of our benchmark. Coordinate value of exposure time represents the upper boundary. (b) The user study result on our benchmark." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.23, + 0.905, + 0.306 + ], + "angle": 0, + "content": "as inputs. We evaluate our method on both static datasets and dynamic datasets. For evaluation, we use the MEFB dataset [66] with 100 static under/over-exposed image pairs and RealHDRV [39], a dynamic HDR deghosting test set containing 50 scenes with varying motion patterns." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.308, + 0.907, + 0.61 + ], + "angle": 0, + "content": "UltraFusion benchmark. Existing exposure fusion benchmark cannot fully evaluate fusing in real-world challenging conditions, as they either lack realistic motion [66] or have limited dynamic range [15, 39, 44]. Therefore, we collect a new real-world UltraFusion benchmark, which contains 100 real-captured under/over-exposed image pairs. Compared to previous datasets, our benchmark is more challenging for three reasons: 1) Our benchmark features larger exposure differences between the two input images (up to 9 stops). 2) It includes more realistic motion, with many scenes containing extensive and unintentional foreground movement. 3) Our benchmark is highly diverse, encompassing daytime, nighttime, indoor, and outdoor scenes captured by DSLR Camera (Canon R8) and mobile phones (iPhone12, iPhone13, Redmi K50 Pro and OPPO Reno8 Pro). We summarize the exposure difference distribution and exposure time distribution of our benchmark in Fig. 6 (a). It can be observed that our benchmark covers a wide range of exposure differences and diverse exposure times, which can be used to effectively test the robustness of the HDR methods." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.612, + 0.909, + 0.764 + ], + "angle": 0, + "content": "Implementation details. We leverage the generative prior encapsulated in Stable Diffusion V2.1 [37]. The decompose-and-fuse control branch (DFCB) is trained for 140k iterations with batch size \\(bs = 32\\) on 8 NVIDIA RTX 4090 GPUs. We also train the fidelity control branch for 1000K iteration with batch size \\(bs = 1\\) on a single NVIDIA RTX 4090 GPU. Adam is adopted as the optimizer and the learning rate is fixed to 0.0001. To adapt HDR reconstruction methods to 2 differently exposed inputs, we re-implement them by following their default settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Evaluation metrics. We utilize four widely used non-reference metric MUSIQ [16], DeQA-Score [62], PAQ2PIQ [61] and HyperIQA [41] for quantitative comparison. Moreover, for static dataset [66], as no ground truths are provided, we select the task-specific MEF-SSIM [27] for structure retention evaluation. For dynamic dataset [39] with HDR ground truths, TMQI [60] is used to evaluate the performance of fidelity and naturalness. In addition, we conduct a user study on our Ultrafusion benchmark to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16115" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.178, + 0.09, + 0.578, + 0.105 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons on the static MEFB dataset [66]." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.115, + 0.66, + 0.292 + ], + "angle": 0, + "content": "
TypeMethodMEFB [66]
MUSIQ↑DeQA-Score↑PAQ2PIQ↑HyperIQA↑MEF-SSIM↑
HDR Rec.HDR-Transformer [25]63.102.98371.360.59960.8626
SCTNet [44]63.133.02171.480.60680.8777
SAFNet [17]61.702.51872.670.56460.7711
MEFDeepfuse [35]52.582.94567.960.43990.8968
MEF-GAN [53]50.592.81869.990.37450.7722
U2Fusion [52]63.393.21972.230.51590.9304
Defusion [20]62.703.11872.820.54550.9062
MEFLUT [13]65.713.27771.210.52670.8608
HSDS-MEF [48]66.763.54472.600.60260.9520
TC-MoA [69]64.603.35571.850.53940.9636
OursUltraFusion68.823.88173.800.64820.9385
" + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.091, + 0.891, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.221, + 0.906, + 0.292 + ], + "angle": 0, + "content": "Figure 7. Trade-off curve between MEF-SSIM and MUSIQ on MEFB dataset [66]. Our UltraFusion achieves the best trade-off between image quality and information preservation." + }, + { + "type": "table_caption", + "bbox": [ + 0.154, + 0.3, + 0.842, + 0.314 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons on dynamic RealHDRV dataset [39] and our challenging UltraFusion benchmark." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.318, + 0.905, + 0.445 + ], + "angle": 0, + "content": "
TypeMethodRealHDRV [39]UltraFusion Benchmark
TMQI↑MUSIQ↑DeQA-Score↑PAQ2PIQ↑HyperIQA↑MUSIQ↑DeQA-Score↑PAQ2PIQ↑HyperIQA↑
HDR Rec.HDR-Transformer [25]0.868062.243.49670.330.522563.662.90972.830.5619
SCTNet [44]0.871562.693.53270.740.527261.843.10272.940.5888
SAFNet [17]0.872662.073.50670.480.515661.502.17973.150.5487
MEFDefusion [20]0.818756.603.30268.380.485660.313.35271.870.5463
MEFLUT [13]0.829762.423.31570.040.502063.623.34371.730.5074
HSDS-MEF [48]0.832361.763.36071.110.505464.543.62773.420.5923
OursUltraFusion0.892567.513.83073.400.583368.413.95775.180.6214
" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.453, + 0.228, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.614, + 0.18, + 0.626 + ], + "angle": 0, + "content": "Inputs" + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.453, + 0.364, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.532, + 0.364, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.613, + 0.332, + 0.625 + ], + "angle": 0, + "content": "HSDS-MEF" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.453, + 0.499, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.532, + 0.499, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.613, + 0.459, + 0.625 + ], + "angle": 0, + "content": "TC-MoA" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.453, + 0.635, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.532, + 0.634, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.613, + 0.62, + 0.625 + ], + "angle": 0, + "content": "HDR-Transformer" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.453, + 0.769, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.532, + 0.769, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.613, + 0.726, + 0.625 + ], + "angle": 0, + "content": "SCTNet" + }, + { + "type": "image", + "bbox": [ + 0.77, + 0.453, + 0.905, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.77, + 0.532, + 0.905, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.822, + 0.614, + 0.853, + 0.625 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.212, + 0.629, + 0.783, + 0.643 + ], + "angle": 0, + "content": "Figure 8. Visual comparisons of different exposure fusion methods on static MEFB dataset [66]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.651, + 0.294, + 0.667 + ], + "angle": 0, + "content": "perform subjective evaluation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.68, + 0.449, + 0.696 + ], + "angle": 0, + "content": "4.2. Comparisons with HDR imaging methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.483, + 0.824 + ], + "angle": 0, + "content": "We compare our method with state-of-the-art HDR Imaging methods, including HDR reconstruction methods HDRTransformer [25], SCTNet [44], SAFNet [17], and multi-exposure fusion methods Deepfuse [35], MEF-GAN [53], U2Fusion [52], Defusion [20], MEFLUT [13], HSDS-MEF [48], TC-MoA [69]. As HDR reconstruction methods cannot output an LDR image directly, we use professional software Photomatix [32] to perform tone mapping." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Evaluation on static dataset. We evaluate the fusion performance on the MEFB dataset [66], focusing on large exposure differences. As shown in Tab. 1, our method outperforms other methods on all four non-reference metrics (MUSIQ, DeQA-Score, PAQ2PIQ, and HyperIQA). Specif" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.651, + 0.907, + 0.879 + ], + "angle": 0, + "content": "ically, our proposed UltraFusion achieves 2.06 gain in terms of MUSIQ compared to HSDS-MEF. In terms of MEFSSIM, as shown in Fig. 7, our baseline model (ControlNet [65]) achieves better image quality but lacks fidelity, while HSDS-MEF [48] retains more information from inputs at the cost of quality. Our UltraFusion outperforms most algorithms, and achieves similar fidelity scores compared to HSDS-MEF and TC-MoA but with much higher image quality (non-reference metrics), indicating the best trade-off between visual quality and information preservation. The qualitative comparison in Fig. 8 further validates this. In contrast, HDR reconstruction methods (HDRTransformer and SCTNet) often miss some detail in highlights and MEF methods (HSDS-MEF and TC-MoA) introduce unnatural transition from brighter to dark regions." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Evaluation on dynamic dataset. To further illustrate the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16116" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.497, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.09, + 0.905, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.218, + 0.707, + 0.232 + ], + "angle": 0, + "content": "Figure 9. Visual comparisons on our captured UltraFusion benchmark." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.238, + 0.481, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.361, + 0.471, + 0.375 + ], + "angle": 0, + "content": "Figure 10. Visual results on dynamic RealHDRV dataset [39]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.384, + 0.483, + 0.656 + ], + "angle": 0, + "content": "robustness of UltraFusion to global and local motion, we use the RealHDRV dataset [39]. We extract the corresponding over-exposed image from the HDR ground truth as input. Tab. 2 demonstrates that our UltraFusion achieve state-of-the-art performance in terms of all metrics. For dynamic scenes, TMQI metric is a particularly important metric, as it is specially designed for HDR evaluation by assessing structural similarity between fusion output and ground truth. Since MEF methods are mainly designed for static scenes, they lack the capability to handle motion, resulting in low TMQI scores and overlay artifacts, as shown in Fig. 10. HDR reconstruction methods trained on dynamic datasets achieve better performance, but still produce noticeable artifact. On the other side, due to our soft inpainting guidance, UltraFusion is much more robust to misalignment and occlusion and the fusion output contains almost no artifacts. It achieves the highest TMQI in Tab. 2 and the best visual result in Fig. 10." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Evaluation on our UltraFusion benchmark. At last, the evaluation on our benchmark validates the robustness of UltraFusion in the most challenging scenes. On all four non-reference metrics, our method outperforms competitors by a large margin, as shown in Tab. 2. The qualitative comparison in Fig. 9 is also consistent with the quantitative metrics. For example, in the red zoom-in patch, integrating the highly bright sun from the under-exposed image into the over-exposed image is extremely challenging. Other methods fail to maintain the shape of the sun or preserve high contrast in the fused region, while our method naturally reconstructs the sun, preserving its appearance and visual pleasing tone of the whole image. Moreover, we conduct a user study on our proposed benchmark. Specifically, we randomly select 20 scenes from our benchmark and invite 136 different users to participate. For each scene, each user" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.238, + 0.905, + 0.266 + ], + "angle": 0, + "content": "Table 3. Ablation studies of three key components of our proposed UltraFusion on RealHDRV dataset [39]." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.269, + 0.903, + 0.349 + ], + "angle": 0, + "content": "
ModelTMQI↑MUSIQ↑
w/o Alignment Strategy0.742763.67
w/o Decompose-and-Fuse Control Branch0.887266.94
w/o Fidelity Control Branch0.876367.36
Ours UltraFusion0.892567.51
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.907, + 0.448 + ], + "angle": 0, + "content": "is asked to compare our method with a randomly chosen baseline. The user study in Fig. 6 (b) indicates our method is more favored by users than competitors. This outcome aligns with the non-reference metric evaluation, showing that our method produces more natural images. More visual comparisons are available in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.456, + 0.674, + 0.47 + ], + "angle": 0, + "content": "4.3. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.905, + 0.524 + ], + "angle": 0, + "content": "To validate the effectiveness of our UltraFusion, we conduct ablation studies on the three key components, followed by an in-depth exploration of the designs of them." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.906, + 0.795 + ], + "angle": 0, + "content": "Key components. First, we perform the ablation study of the proposed three key components, including the alignment strategy, decompose-and-fuse control branch, and fidelity control branch, on the RealHDRV [39] dataset, as shown in Tab. 3. We first remove our alignment strategy, opting instead to train the model directly on the original SICE [1] dataset and input differently exposed image pairs without coarse alignment. Without the pre-alignment stage and the data synthesis pipeline designed for it, performance drops significantly in terms of TMQI. Fig. 11 (a) also shows that the model fails to implicitly align large motions. Then, we replace the decompose-and-fuse control branch (DFCB) with the vanilla ControlNet [65], and it fails to fuse features with large exposure differences, which leads to the loss of details (see Fig. 11 (b)) and a decrease in TMQI. Finally, when we exclude the fidelity control branch (FCB), Ultra-Fusion loses the capability to maintain the detailed structure and vivid color, as indicated by Fig. 11 (c)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Alignment strategy. Second, we perform ablation of the proposed alignment strategy, which is the most critical design for our framework. Our alignment strategy mainly consists of a data synthesis pipeline at training time and a pre-alignment module at testing time. Without both of them, the whole model degrades to a multi-exposure fusion method as shown in Fig. 12 (c). Performing pre-alignment can reduce" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16117" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.091, + 0.222, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.16, + 0.222, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.224, + 0.221, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.272, + 0.218, + 0.279 + ], + "angle": 0, + "content": "Under-Exposure Over-Exposure" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.28, + 0.221, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.329, + 0.153, + 0.341 + ], + "angle": 0, + "content": "Align Strategy" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.091, + 0.481, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.224, + 0.351, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.272, + 0.349, + 0.279 + ], + "angle": 0, + "content": "Under-Exposure Over-Exposure" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.28, + 0.351, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.237, + 0.328, + 0.278, + 0.336 + ], + "angle": 0, + "content": "w/o DFCB" + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.224, + 0.419, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.358, + 0.272, + 0.419, + 0.279 + ], + "angle": 0, + "content": "Under-Exposure" + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.224, + 0.48, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.42, + 0.272, + 0.48, + 0.279 + ], + "angle": 0, + "content": "Over-Exposure" + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.329, + 0.351, + 0.336 + ], + "angle": 0, + "content": "w/o DFCB Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.358, + 0.329, + 0.406, + 0.336 + ], + "angle": 0, + "content": "w/o FCB" + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.347, + 0.483, + 0.374 + ], + "angle": 0, + "content": "Figure 11. Visual results of ablating each key component of our method. Each key component contributes to the final results." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.38, + 0.223, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.448, + 0.195, + 0.456 + ], + "angle": 0, + "content": "(a) Under Exposure" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.38, + 0.351, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.448, + 0.319, + 0.456 + ], + "angle": 0, + "content": "(b) Over Exposure" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.38, + 0.481, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.448, + 0.481, + 0.456 + ], + "angle": 0, + "content": "(c) \\( w / o \\) Data Syn. & Pre-Alignment" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.456, + 0.223, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.127, + 0.524, + 0.191, + 0.532 + ], + "angle": 0, + "content": "(d) \\( w / o \\) Data Syn." + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.456, + 0.351, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.246, + 0.524, + 0.33, + 0.532 + ], + "angle": 0, + "content": "(e) \\( w / o \\) Pre-Alignment" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.456, + 0.481, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.524, + 0.468, + 0.533 + ], + "angle": 0, + "content": "(f) Ours Alignment Strategy" + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.537, + 0.437, + 0.55 + ], + "angle": 0, + "content": "Figure 12. Effectiveness of our alignment strategy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.658 + ], + "angle": 0, + "content": "artifacts to some extent (see Fig. 12 (d)), but may still be prone to alignment errors. Our data synthesis pipeline can improves the robustness of our algorithm to unaligned conditions, but motion in the occluded regions still cannot be solved, as shown in Fig. 12. After combining data synthesis pipeline and pre-alignment stage, our algorithm demonstrates strong capability towards dynamic scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Decompose-and-fuse control branch. At last, we evaluate how the form of soft guidance provided by the underexposed image impacts the recovery of the highlight regions. When using the RGB under-exposed image \\( I_{ue} \\) as guidance, the model will ignore some details in the output (Fig. 13 (c)). Replacing \\( I_{ue} \\) with its structure information \\( S_{ue} \\) retains more details (Fig. 13 (d)), but without generating vivid color. By incorporating both under-exposed color information \\( C_{ue} \\) and structure information \\( S_{ue} \\), the reconstructed results can maintain more details and color consistency (see Fig. 13 (e)). At last, multi-scale cross-attention further improves the fusion result (see Fig. 13 (f))." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.402, + 0.866 + ], + "angle": 0, + "content": "4.4. Application on general image fusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "One advantage of the UltraFusion is that it can be extended to general image fusion, thanks to the flexibility of our" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.091, + 0.647, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.157, + 0.625, + 0.166 + ], + "angle": 0, + "content": "(a) Under Exposure" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.091, + 0.776, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.667, + 0.157, + 0.749, + 0.166 + ], + "angle": 0, + "content": "(b) Over Exposure" + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.091, + 0.904, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.825, + 0.157, + 0.855, + 0.167 + ], + "angle": 0, + "content": "(c) \\(I_{ue}\\)" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.167, + 0.647, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.566, + 0.235, + 0.597, + 0.245 + ], + "angle": 0, + "content": "(d) \\(S_{ue}\\)" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.167, + 0.776, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.68, + 0.235, + 0.741, + 0.245 + ], + "angle": 0, + "content": "(e) \\(S_{ue} + C_{ue}\\)" + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.167, + 0.904, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.819, + 0.235, + 0.862, + 0.244 + ], + "angle": 0, + "content": "(f) DFCB" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.25, + 0.907, + 0.277 + ], + "angle": 0, + "content": "Figure 13. Detailed ablation study on the design choices of decompose-and-fuse control branch." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.284, + 0.641, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.352, + 0.627, + 0.361 + ], + "angle": 0, + "content": "(a) Input Over Exposure" + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.285, + 0.767, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.663, + 0.352, + 0.747, + 0.36 + ], + "angle": 0, + "content": "(b) Our Fusion Result" + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.285, + 0.895, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.758, + 0.352, + 0.903, + 0.361 + ], + "angle": 0, + "content": "(c) Another Fusion w/ Diff. Highlight" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.365, + 0.906, + 0.405 + ], + "angle": 0, + "content": "Figure 14. Extension to general fusion. Given totally different underexposed images as guidance (upper right corner), our method can also generate different fusion results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.906, + 0.52 + ], + "angle": 0, + "content": "guided inpainting. To illustrate this potential, we explore one additional interesting demo to fuse two irrelevant images captured by different cameras at different locations. As shown in Fig. 14, UltraFusion successfully copies the moon (b) or blue sky (c) to the over-exposed image, unlocking many interesting potential applications (e.g., image harmonization)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.532, + 0.634, + 0.547 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.748 + ], + "angle": 0, + "content": "In this work, we introduce a novel approach to HDR imaging, tackling challenges presented by significant exposure differences and large motion. By modeling the fusion process as a guided inpainting problem and utilizing the under-exposed image for soft guidance, our method is robust with alignment errors and circumvents tone mapping, resulting in natural, artifact-free outputs. We also propose decompose-and-fuse control branch and fidelity control branch to improve feature modulation and fidelity preservation of ControlNet. Extensive experiments on existing datasets and our captured benchmark demonstrate the robustness and effectiveness of our method compared to previous HDR methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "With extreme exposure differences and challenging non-rigid motion, occlusion mask estimation may introduce errors, causing the restoration of certain highlight regions degrades to single image HDR. While our method is able to use diffusion priors to restore these highlights, restoration without under-exposed information remains unreliable. Moreover, it takes almost 3.3s to fuse \\(512 \\times 512\\) inputs on an NVIDIA RTX 4090 GPU. A more exposure-robust optical flow algorithm and a faster implementation are highly desirable, and we leave them for future work." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16118" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.483, + 0.176 + ], + "angle": 0, + "content": "This work was supported by the National Key R&D Program of China No.2022ZD0160201, Shanghai Artificial Intelligence Laboratory and RGC Early Career Scheme (ECS) No. 24209224." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.192, + 0.187, + 0.208 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.217, + 0.484, + 0.272 + ], + "angle": 0, + "content": "[1] Jianrui Cai, Shuhang Gu, and Lei Zhang. Learning a deep single image contrast enhancer from multi-exposure images. IEEE Transactions on Image Processing, 27(4):2049-2062, 2018. 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.274, + 0.484, + 0.33 + ], + "angle": 0, + "content": "[2] Cong Cao, Huanjing Yue, Xin Liu, and Jingyu Yang. Unsupervised HDR image and video tone mapping via contrastive learning. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.332, + 0.483, + 0.399 + ], + "angle": 0, + "content": "[3] Jie Chen, Zaifeng Yang, Tsz Nam Chan, Hui Li, Junhui Hou, and Lap-Pui Chau. Attention-guided progressive neural texture fusion for high dynamic range image restoration. IEEE Transactions on Image Processing, 31:2661-2672, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.403, + 0.483, + 0.471 + ], + "angle": 0, + "content": "[4] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.474, + 0.484, + 0.543 + ], + "angle": 0, + "content": "[5] Ciprian Corneanu, Raghudeep Gadde, and Aleix M Martinez. Latentpaint: Image inpainting in latent space with diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4334-4343, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.546, + 0.484, + 0.6 + ], + "angle": 0, + "content": "[6] Abhishek Goswami, Aru Ranjan Singh, Francesco Banterle, Kurt Debattista, and Thomas Bashford-Rogers. Semantic aware diffusion inverse tone mapping. arXiv preprint arXiv:2405.15468, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.603, + 0.484, + 0.657 + ], + "angle": 0, + "content": "[7] Michael D Grossberg and Shree K Nayar. Determining the camera response from images: What is knowable? IEEE Transactions on pattern analysis and machine intelligence, 25(11):1455-1467, 2003. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.66, + 0.484, + 0.714 + ], + "angle": 0, + "content": "[8] Yuanshen Guan, Ruikang Xu, Mingde Yao, Ruisheng Gao, Lizhi Wang, and Zhiwei Xiong. Diffusion-promoted HDR video reconstruction. arXiv preprint arXiv:2406.08204, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.717, + 0.484, + 0.786 + ], + "angle": 0, + "content": "[9] Samuel W Hasinoff, Dillon Sharlet, Ryan Geiss, Andrew Adams, Jonathan T Barron, Florian Kainz, Jiawen Chen, and Marc Levoy. Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM Transactions on Graphics (ToG), 35(6):1-12, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.899 + ], + "angle": 0, + "content": "[11] Litao Hu, Huaijin Chen, and Jan P Allebach. Joint multiscale tone mapping and denoising for HDR image enhancement. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 729-738, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.217, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[12] Tao Hu, Qingsen Yan, Yuankai Qi, and Yanning Zhang. Generating content for HDR deghosting from frequency view. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25732-25741, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[13] Ting Jiang, Chuan Wang, Xinpeng Li, Ru Li, Haoqiang Fan, and Shuaicheng Liu. Meflut: Unsupervised 1d lookup tables for multi-exposure image fusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10542-10551, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[14] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. arXiv preprint arXiv:2403.06976, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.904, + 0.318 + ], + "angle": 0, + "content": "[15] Nima Khademi Kalantari, Ravi Ramamoorthi, et al. Deep high dynamic range imaging of dynamic scenes. ACM Trans. Graph., 36(4):144-1, 2017. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[16] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.378, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[17] Lingtong Kong, Bo Li, Yike Xiong, Hao Zhang, Hong Gu, and Jinwei Chen. Safnet: Selective alignment fusion network for efficient HDR imaging. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.434, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[18] Baiang Li, Sizhuo Ma, Yanhong Zeng, Xiaogang Xu, Youqing Fang, Zhao Zhang, Jian Wang, and Kai Chen. Sagiri: Low dynamic range image enhancement with generative diffusion prior. arXiv preprint arXiv:2406.09389, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.504, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[19] Hui Li, Kede Ma, Hongwei Yong, and Lei Zhang. Fast multiscale structural patch decomposition for multi-exposure image fusion. IEEE Transactions on Image Processing, 29: 5805-5816, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[20] Pengwei Liang, Junjun Jiang, Xianming Liu, and Jiayi Ma. Fusion from decomposition: A self-supervised decomposition approach for image fusion. In European Conference on Computer Vision, pages 719-735. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.619, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[21] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Wanli Ouyang, Yu Qiao, and Chao Dong. Diff-bir: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[22] Renshuai Liu, Chengyang Li, Haitao Cao, Yinglin Zheng, Ming Zeng, and Xuan Cheng. Emef: ensemble multi-exposure image fusion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1710-1718, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[23] Yuhao Liu, Zhanghan Ke, Fang Liu, Nanxuan Zhao, and Rynson WH Lau. Diff-plugin: Revitalizing details for diffusion-based low-level tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4197-4208, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.817, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[24] Zhen Liu, Yinglong Wang, Bing Zeng, and Shuaicheng Liu. Ghost-free high dynamic range imaging with context-aware transformer. In European Conference on computer vision, pages 344-360. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Zhen Liu, Yinglong Wang, Bing Zeng, and Shuaicheng Liu. Ghost-free high dynamic range imaging with context-aware" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16119" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "transformer. In European Conference on computer vision, pages 344-360. Springer, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[26] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[27] Kede Ma, Kai Zeng, and Zhou Wang. Perceptual quality assessment for multi-exposure image fusion. IEEE Transactions on Image Processing, 24(11):3345-3356, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.483, + 0.291 + ], + "angle": 0, + "content": "[28] Kede Ma, Hui Li, Hongwei Yong, Zhou Wang, Deyu Meng, and Lei Zhang. Robust multi-exposure image fusion: a structural patch decomposition approach. IEEE Transactions on Image Processing, 26(5):2519-2532, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[29] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.483, + 0.391 + ], + "angle": 0, + "content": "[30] Tom Mertens, Jan Kautz, and Frank Van Reeth. Exposure fusion. In 15th Pacific Conference on Computer Graphics and Applications (PG'07), pages 382-390. IEEE, 2007. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.393, + 0.483, + 0.447 + ], + "angle": 0, + "content": "[31] Ozan Özdenizci and Robert Legenstein. Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10346-10357, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.483, + 0.476 + ], + "angle": 0, + "content": "[32] Photomatix. Commercially-available hdr processing software. https://www.hdrsoft.com/.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.478, + 0.483, + 0.548 + ], + "angle": 0, + "content": "[33] K Ram Prabhakar, Rajat Arora, Adhitya Swaminathan, Kunal Pratap Singh, and R Venkatesh Babu. A fast, scalable, and reliable deghosting method for extreme exposure fusion. In 2019 IEEE International Conference on Computational Photography (ICCP), pages 1-8. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.549, + 0.483, + 0.632 + ], + "angle": 0, + "content": "[34] K Ram Prabhakar, Susmit Agrawal, Durgesh Kumar Singh, Balraj Ashwath, and R Venkatesh Babu. Towards practical and efficient high-resolution HDR deghosting with cnn. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16, pages 497-513. Springer, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.634, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[35] K Ram Prabhakar, V Sai Srikar, and R Venkatesh Babu. Deepfuse: A deep unsupervised approach for exposure fusion with extreme exposure image pairs. In Proceedings of the IEEE international conference on computer vision, pages 4714-4722, 2017. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[38] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition, pages 8839-8849, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[39] Yong Shu, Liquan Shen, Xiangyu Hu, Mengyao Li, and Zihao Zhou. Towards real-world HDR video reconstruction: A large-scale benchmark dataset and a two-stage alignment network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2879-2888, 2024. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[40] Jou Won Song, Ye-In Park, Kyeongbo Kong, Jaeho Kwak, and Suk-Ju Kang. Selective transhdr: Transformer-based selective HDR imaging using ghost region mask. In European Conference on Computer Vision, pages 288-304. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[41] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3667-3676, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.411 + ], + "angle": 0, + "content": "[42] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.905, + 0.481 + ], + "angle": 0, + "content": "[43] Steven Tel, Zongwei Wu, Yulun Zhang, Barthélémy Heyrman, Cédric Demonceaux, Radu Timofte, and Dominique Gingac. Alignment-free HDR deghosting with semantics consistent transformer. arXiv preprint arXiv:2305.18135, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.483, + 0.905, + 0.565 + ], + "angle": 0, + "content": "[44] Steven Tel, Zongwei Wu, Yulun Zhang, Barthélémy Heyrman, Cédric Demonceaux, Radu Timofte, and Dominique Gingac. Alignment-free HDR deghosting with semantics consistent transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12836-12845, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.633 + ], + "angle": 0, + "content": "[45] Yael Vinker, Inbar Huberman-Spiegelglas, and Raanan Fattal. Unpaired learning for high dynamic range image tone mapping. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14657-14666, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.636, + 0.905, + 0.69 + ], + "angle": 0, + "content": "[46] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[47] Shuchen Weng, Peixuan Zhang, Yu Li, Si Li, Boxin Shi, et al. L-cad: Language-based colorization with any-level descriptions using diffusion priors. Advances in Neural Information Processing Systems, 36, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[48] Guanyao Wu, Hongming Fu, Jinyuan Liu, Long Ma, Xin Fan, and Risheng Liu. Hybrid-supervised dual-search: Leveraging automatic learning for loss-free multi-exposure image fusion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5985-5993, 2024. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[49] Shangzhe Wu, Jiarui Xu, Yu-Wing Tai, and Chi-Keung Tang. Deep high dynamic range imaging with large foreground motions. In Proceedings of the European Conference on Computer Vision (ECCV), pages 117–132, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[50] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16120" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22428-22437, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[51] Gangwei Xu, Yujin Wang, Jinwei Gu, Tianfan Xue, and Xin Yang. Hdrflow: Real-time HDR video reconstruction with large motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24851-24860, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.263 + ], + "angle": 0, + "content": "[52] Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502-518, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.484, + 0.319 + ], + "angle": 0, + "content": "[53] Han Xu, Jiayi Ma, and Xiao-Ping Zhang. Mef-gan: Multiexposure image fusion via generative adversarial networks. IEEE Transactions on Image Processing, 29:7203-7216, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.484, + 0.389 + ], + "angle": 0, + "content": "[54] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8121-8130, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.484, + 0.447 + ], + "angle": 0, + "content": "[55] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. International Journal of Computer Vision, 127:1106-1125, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.484, + 0.531 + ], + "angle": 0, + "content": "[56] Qingsen Yan, Dong Gong, Qinfeng Shi, Anton van den Hengel, Chunhua Shen, Ian Reid, and Yanning Zhang. Attention-guided network for ghost-free high dynamic range imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1751-1760, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.484, + 0.588 + ], + "angle": 0, + "content": "[57] Qingsen Yan, Lei Zhang, Yu Liu, Yu Zhu, Jinqiu Sun, Qinfeng Shi, and Yanning Zhang. Deep HDR imaging via a non-local network. IEEE Transactions on Image Processing, 29: 4308-4322, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.484, + 0.659 + ], + "angle": 0, + "content": "[58] Qingsen Yan, Tao Hu, Yuan Sun, Hao Tang, Yu Zhu, Wei Dong, Luc Van Gool, and Yanning Zhang. Towards high-quality HDR dehosting with conditional diffusion models. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.484, + 0.716 + ], + "angle": 0, + "content": "[59] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.484, + 0.761 + ], + "angle": 0, + "content": "[60] Hojatollah Yeganeh and Zhou Wang. Objective quality assessment of tone-mapped images. IEEE Transactions on Image processing, 22(2):657-667, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[61] Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3575–3585, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[62] Zhiyuan You, Xin Cai, Jinjin Gu, Tianfan Xue, and Chao Dong. Teaching large language models to regress accurate image quality scores using score distribution. arXiv preprint arXiv:2501.11561, 2025. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.176 + ], + "angle": 0, + "content": "[63] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photorealistic image restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25669-25680, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.908, + 0.232 + ], + "angle": 0, + "content": "[64] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.908, + 0.29 + ], + "angle": 0, + "content": "[65] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.908, + 0.331 + ], + "angle": 0, + "content": "[66] Xingchen Zhang. Benchmarking and comparing multi-exposure image fusion algorithms. Information Fusion, 74: 111-131, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.908, + 0.403 + ], + "angle": 0, + "content": "[67] Shengyu Zhao, Yilun Sheng, Yue Dong, Eric I Chang, Yan Xu, et al. Maskflownet: Asymmetric feature matching with learnable occlusion mask. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6278-6287, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.908, + 0.471 + ], + "angle": 0, + "content": "[68] Zixiang Zhao, Lilun Deng, Haowen Bai, Yukun Cui, Zhipeng Zhang, Yulun Zhang, Haotong Qin, Dongdong Chen, Jiangshe Zhang, Peng Wang, et al. Image fusion via vision-language model. arXiv preprint arXiv:2402.02235, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.473, + 0.908, + 0.543 + ], + "angle": 0, + "content": "[69] Pengfei Zhu, Yang Sun, Bing Cao, and Qinghua Hu. Task-customized mixture of adapters for general image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7099-7108, 2024. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16121" + } + ] +] \ No newline at end of file diff --git a/2025/UltraFusion_ Ultra High Dynamic Imaging using Exposure Fusion/c2c26ec1-c8af-49f8-87ba-b3c749531067_origin.pdf b/2025/UltraFusion_ Ultra High Dynamic Imaging using Exposure Fusion/c2c26ec1-c8af-49f8-87ba-b3c749531067_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8b74e80bbfb645293e4aabc46424ae2c0943e266 --- /dev/null +++ b/2025/UltraFusion_ Ultra High Dynamic Imaging using Exposure Fusion/c2c26ec1-c8af-49f8-87ba-b3c749531067_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f4612f4f3e6812f540bdf5f2a902ca1f6b0b1e2345f2386c14d3e0d19defd40f +size 9950133 diff --git a/2025/UltraFusion_ Ultra High Dynamic Imaging using Exposure Fusion/full.md b/2025/UltraFusion_ Ultra High Dynamic Imaging using Exposure Fusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8a1a00a82892d8be6b3762ab53566029450d839f --- /dev/null +++ b/2025/UltraFusion_ Ultra High Dynamic Imaging using Exposure Fusion/full.md @@ -0,0 +1,415 @@ +# UltraFusion: Ultra High Dynamic Imaging using Exposure Fusion + +Zixuan Chen $^{1,3*}$ Yujin Wang $^{1*}$ Xin Cai $^{2}$ Zhiyuan You $^{2}$ Zheming Lu $^{3}$ Fan Zhang $^{1}$ Shi Guo $^{1}$ Tianfan Xue $^{2,1}$ $^{1}$ Shanghai AI Laboratory $^{2}$ The Chinese University of Hong Kong $^{3}$ Zhejiang University $\{\mathrm{zxchen}, \mathrm{zheminglu}\} @ \mathrm{zju.edu.cn}, \{\mathrm{wangyujin}, \mathrm{zhangfan}, \mathrm{guoshi}\} @ \mathrm{pjlab.org.cn}$ caixin@link.cuhk.edu.hk, zhiyuanyou@foxmail.com, tfxue@ie.cuhk.edu.hk + +![](images/8c0cd337a93d395280ba4e556888a77ce0c9e50727e6122374703d8a3e498ce7.jpg) + +![](images/36d3e4ee10e59d8b48b0e69ed2af2771da8cffbef7d1ef45698dcfcc478c5487.jpg) + +![](images/63d68dd668dfee57726428ac7abc43378a731407eaa586b283c9c68858466568.jpg) +Under Exposure + +![](images/f63f14dc9e6b83e9e95f9a46eb4855ca19e8a03694217921e2bf0cb8e2520d9c.jpg) +Over Exposure + +![](images/86651fcc803b59ecfa03880f24637a9bb3261e0666a6e9ac4f759db0266b63af.jpg) + +![](images/fcb181836efee2444eee333e4a85eea9eb756749bd591ab59362ec48cb6d60e9.jpg) +Previous HDR Reconstruction + +![](images/7e496caaf6c6680156378f487b44035a95744e62423e91d5cf000f060bf49374.jpg) +Exposure Multi- +Figure 1. Comparing our guided inpainting HDR imaging method with a state-of-the-art HDR reconstruction [25] and multi-exposure fusion [48] methods. Both scenes are selected from our captured new real-world benchmark. Left: night cityscape with large exposure difference. Right: afternoon street with motion-induced occlusion. Previous methods struggle to handle these scenes. By modeling HDR as an inpainting problem, our method can produce visually appealing results without ghosting artifacts in these challenging scenes. + +![](images/b835c1fbb72c1f1486a1043c9d220255d1fe832bfb91412c607a4b18e587846a.jpg) +Our Guided Inpainting + +![](images/c831a0057bd40a4adbbca202cdafe7fb9330090748ae062db1032b3a29990e3a.jpg) + +![](images/3c455bfe6e36334b0a72ff3d20b997495663f57fc6e97ca232bb72dffa9b5076.jpg) + +![](images/3ff29fa01c3875d6d0c40d2def81b6b154728a7cce43e84c94b453d690d37b9e.jpg) +Under Exposure + +![](images/2f02524b899cfa3a741c8f5d984376a57a8bb172d9d4fdde124d26e59fc916ec.jpg) +Over Exposure + +![](images/d0c3555145f968fd5a365347564bd2000c2ac22711b6e14cd5290fdf41cf82ca.jpg) + +![](images/a501619ead788a1050d16faa5643ac3b2bcd7522b4c0abecb555b05be4ad8f1f.jpg) +Previous HDR Reconstruction + +![](images/6d8560561528f07f0acafd374af04b22d9d42f3f60c94fed3d091372e6291349.jpg) +Previous MultiExposure Fusion + +![](images/2e320d4ccb5ca97842a1e91ba555728ec8993b80e833b63cbedd42c8f1e7927c.jpg) +Our Guided Inpainting + +# Abstract + +Capturing high dynamic range (HDR) scenes is one of the most important issues in camera design. Majority of cameras use exposure fusion, which fuses images captured by different exposure levels, to increase dynamic range. However, this approach can only handle images with limited exposure difference, normally 3-4 stops. When applying to very high dynamic range scenes where a large exposure difference is required, this approach often fails due to incorrect alignment or inconsistent lighting between inputs, or tone mapping artifacts. In this work, we propose UltraFusion, the first exposure fusion technique that can merge inputs with 9 stops differences. The key idea is that we model exposure fusion as a guided inpainting problem, where the under-exposed image is used as a guidance to fill the missing information of over-exposed highlights in the overexposed region. Using an under-exposed image as a soft guidance, instead of a hard constraint, our model is robust + +to potential alignment issue or lighting variations. Moreover, by utilizing the image prior of the generative model, our model also generates natural tone mapping, even for very high-dynamic range scenes. Our approach outperforms HDR-Transformer on latest HDR benchmarks. Moreover, to test its performance in ultra high dynamic range scenes, we capture a new real-world exposure fusion benchmark, UltraFusion dataset, with exposure differences up to 9 stops, and experiments show that UltraFusion can generate beautiful and high-quality fusion results under various scenarios. Code and data will be available at https://openimaginglab.github.io/UltraFusion. + +# 1. Introduction + +High dynamic range (HDR) imaging is one of the fundamental problems in the modern camera design. Due to hardware limitation, camera sensors have a much smaller dynamic range compared with the real-world. To increase it, majority of HDR solutions merge multiple images, either with the same [9] or different exposure lev + +![](images/3a4b4e41840596b085cca383f3ac0e28058089416fec9396c9695a8e73b7af40.jpg) +Figure 2. Visual comparison on directly utilizing ControlNet [65] and our UltraFusion. Without pre-aligned data, ControlNet struggles to fix a frame for reference. However, our method fixes the over-exposed image as the reference and the under-exposed image as the guidance for inpainting, thereby avoiding artifacts. + +els [3, 15, 17, 24, 34, 40, 43, 49, 56, 57]. Despite all recent advances, majority HDR algorithms can only bring limited increase in dynamic range. For example, $\mathrm{HDR + }$ [9], the first HDR algorithm used by commercial cameras, can only robustly increase the dynamic range by 8 times (3 stops). Therefore, in this work, we study the following question: can we drastically increase the dynamic range of a camera by fusing two images with very large exposure difference, like 9 stops in Fig. 1? + +This is a fundamental challenging problem, due to the following three issues. First, to handle dynamic scenes, most of HDR fusion algorithms will first align input frames, which is very challenging when input has large brightness difference. As result, a ghosting issue happens when alignment fails, indicated by zoom-in patches in the right scene of Fig. 1. Second, most of HDR algorithms assume that under-exposed image is simply a darker version of the normal image. However, the appearance of an object may change when exposure levels change, like ship in the left scene of Fig. 1, resulting unnatural fusion result. Third, sometimes the result of fusion is an HDR image, which cannot be directly shown on normal low-dynamic-range display. Therefore, these HDR images will be further compressed through a tone-mapping process. When dynamic range is high, tone-mapping may introduce additional issue. Maintaining a natural contrast and rich details in the final output is challenging, as shown in the zoom-in patches of previous HDR reconstruction method in Fig. 1. + +In this work, we propose a completely different fusion method, UltraFusion, which models it as a guided inpainting problem. In this setup, the user captures two images, one normal exposed image where brighter objects are overexposed, and another under exposed image, which only captures the very brighter parts of the scene. We use the normal exposed image as a reference, and inpainting the missing information in the highlight. Unlike the traditional inpainting, we use the information from under-exposed frame as guidance, so inpainted highlight is not completely generated, but stays consistent with under-exposed frame. + +There are three advantages of this approach when handling large exposure difference. First, it follows the exposure fusion [30] setup. Unlike HDR fusion techniques, which first generate HDR image and then compress to a low-dynamic-range (LDR) image through tone mapping, exposure fusion directly generates LDR output, avoiding cascading errors. That being said, our method outputs a tone-mapped LDR image instead of a linear HDR. Second, the under-exposed image is used as a soft guidance, instead of a hard constraint. Therefore, UltraFusion is robust to alignment error (see the right scene of Fig. 1) and lighting variation (see the left scene of Fig. 1). Third, image prior of the generative model ensures the natural look of the output image, reducing the potential artifacts. + +To train a guided inpainting, a simple solution is to train a ControlNet [65] using two input images. However, such paradigm cannot handle dynamic HDR scenes, as ControlNet may not know which frame to choose as the reference, increasing the difficulty to fuse the information from over and short-exposed images. For instance, as shown in Fig. 2, ControlNet selects the over-exposed image as the reference frame in red boxed region, but chooses the under-exposed image as the reference frame in green boxed region, leading to substantial artifacts in the result. Additionally, as a generative model, it may inevitably generate fake image content, as indicated by green box in Fig. 2. + +To address these challenges, we design UltraFusion as follows. First, we warp the short-exposed image to the long-exposed one, and mask out the occluded regions. Then, we utilize the diffusion prior to inpaint the long-exposed image guided by partial short-exposed information. To make the network retain more details to generate guidance information, we propose a new decompose-and-fuse control branch, which eliminates the luminance component of the short-exposed image, extracting structure and color information instead, and employs a multi-scale cross-attention to improve the feature fusion with long-exposed image. Second, as there is no existing large-scale training data for exposure fusion of dynamic scene, we also propose a novel training data synthesis pipeline, utilizing existing high-quality multi-exposure datasets (pre-aligned) of static scene and video datasets. At last, to ensure the generated output maintains the fidelity to the reality, we also train an additional fidelity control branch, using the same decompose-and-fuse strategy and multi-scale cross-attention. + +At last, to evaluate the effectiveness of our framework, we capture 100 under/over-exposed image pairs, covering daytime, nighttime, indoor, outdoor scenes with local and global motion patterns. Experiments on both latest HDR imaging datasets and our captured benchmark demonstrate that, comparing to existing methods, our UltraFusion is more robust to scenes with large exposure differences and large motion, as shown in Fig. 1. + +# 2. Related work + +# 2.1. HDR imaging + +HDR Imaging can be deided into HDR reconstruction and Multi-Exposure Fusion (MEF) typically depends on the domain where fusion occurs [19]. HDR reconstruction methods [3, 15, 17, 24, 34, 40, 43, 49, 51, 56, 57] inverts the camera response function (CRF) to merge exposure brackets in the linear HDR domain [28]. In most cases, tone mapping is necessary to display the reconstructed HDR image properly on standard LDR monitors. As a cost-effective alternative [13], MEF methods [13, 19, 22, 28, 30, 33, 35, 48, 52, 68, 69] directly fuse images in the LDR domain, sidestepping CRF calibration and sophisticated tone mapping process [30]. Regardless of the type of HDR Imaging methods, they contend with ghosting artifacts caused by camera shake and object movement [39]. Previous methods have attempted explicit or implicit alignment using optical flow [15, 33, 34, 49] or attention mechanism [3, 17, 24, 40, 43, 56, 57]. However, most HDR Imaging methods suffer from unpleasing artifacts when large motion causes occlusion in the complementary region. + +Diffusion models. Recently, it has been witnessed with the rapid rise of diffusion models [10, 37] and their successful application in various tasks, including controllable image generation [59, 65], image restoration [21, 23, 31, 47, 63], image editing [4, 14, 29, 38] and image inpainting [5, 26, 50, 64]. In the field of HDR imaging, the application of diffusion models has primarily focused on HDR deghosting [8, 12, 58]. However, since these methods do not leverage the diffusion priors learned from large-scale datasets, their generalization ability is limited by the scale of the HDR dataset. While some recent works [6, 18] have employed diffusion priors, they tend to focus on single image HDR. Without another differently-exposed image as reference, the results generated by these methods lack sufficient reliability. Unlike previous methods, we utilize diffusion priors and use the short-exposed image as a reference to perform reliable inpainting in the highlight regions of the overexposed image, thereby achieving natural and reliable HDR scene reconstruction. Compared to diffusion-based inpainting methods [5, 26, 50, 64] that perform inpainting solely from scratch, we leverage information from short-exposure images to guide a more accurate inpainting process. + +Tone mapping methods. The goal of tone mapping is to convert HDR images to LDR for display on standard screens while enhancing visual detail. Due to the challenge of obtaining ground-truth tone mapping results, unsupervised deep learning approaches have been developed using adversarial [45] and contrastive [2] learning. To address data limitations, Cai et al. [1] manually curated training data by selecting the best results from 13 tone mapping methods for network learning [1, 11]. However, previous tone + +mapping algorithms, lacking robust image priors and facing data constraints, struggle with visually pleasing results and generalization in extreme high dynamic range scenes. By incorporating diffusion-based image priors, our method achieves aesthetic results even in challenging high dynamic range scenarios (see Fig. 1). + +# 3. Methodology + +Given an over-exposed image $I_{oe}$ and an under-exposed image $I_{ue}$ , traditional exposure fusion algorithms directly aggregate different frequency band of both images, which are sensitive to misalignment error or lighting variation. Instead, we treat this as an inpainting problem. Specifically, we use the over-exposed image $I_{oe}$ as the base image and inpainting missing information in the highlight region. To ensure inpainted highlights are real, we also use highlights from under-exposed image as guidance. + +Based on this idea, we design a 2-stage newtwork shown in Fig. 3, which consists of the pre-alignment stage and the guided inpainting stage. The pre-alignment stage outputs a coarse-aligned version of $I_{ue}$ , which is used as the soft guidance in the following guided inpainting stage. Details of each stage is described below. + +# 3.1. Pre-alignment stage + +Most of optical flow alignment assumes input have the similar brightness. Therefore, we first adjust the brightness $I_{ue}$ to match the distribution of $I_{oe}$ through intensity mapping function [7]. Then, we adopt RAFT [42], a pre-trained optical flow network, to estimate the bidirectional flow $f_{oe \rightarrow ue}$ and $f_{ue \rightarrow oe}$ and align $I_{ue}$ to $I_{oe}$ using backward warping. However, backward warping will result in ghosting at the occlusion boundary [67], leading to artifacts in the next guided inpainting stage. To solve that, we utilize the forward-and-backward consistency check [54] to estimate occluded regions $\mathcal{M}$ and mask them out in the warped output. Finally, we can obtain a pre-aligned output $I_{ue \rightarrow oe}$ of the first stage: + +$$ +I _ {u e \rightarrow o e} = (1 - \mathcal {M}) \cdot \mathcal {W} \left(I _ {u e}, f _ {o e \rightarrow u e}\right), \tag {1} +$$ + +where $\mathcal{W}$ denotes the backward warping. Fig. 3 (a) shows the output is a masked and aligned under-exposed image. + +# 3.2. Guided inpainting stage + +We build our guided inpainting model based on the Stable Diffusion Model [36] because its powerful generative prior can help to resolve ambiguity during inpainting. Similar to other diffusion-based image enhancement techniques [21], we also inject the following information through an additional control branch, as shown in Fig. 3(b): 1) the image to be inpainted, which is the over-exposed image $I_{oe}$ , 2) the additional guidance of highlight, which is the underexposed image $I_{ue}$ , and 3) the diffusion latent at the current + +![](images/8fadd714b2953a61ee44d5c446f878dc6e4cc7cac1082b3d009fbf9faa54d7de.jpg) +(a) Pre-Alignment Stage +Figure 3. The whole backbone of UltraFusion. Our method is a 2-stage framework, consisting of (a) pre-alignment stage and (b) guided inpainting stage. The first stage pre-aligns the under-exposed image $I_{ue}$ to the over-exposed image $I_{oe}$ and masks the occluded regions. In the subsequent guided inpainting stage, we propose a new decompose-and-fuse control branch to utilize the diffusion priors, together with a fidelity control branch for fidelity keeping. + +![](images/11a54cfe6cf82e099f81cf5ed02503946f06f32d6bb8edb3f9ca63e8b19d3f0e.jpg) +(b) Guided Inpainting Stage + +![](images/ac6aa4d034a334b9edfb1f03808169d4ce337e20f0e5bbde26014db38a946ca2.jpg) +Decompose-and-Fuse Control Branch +Figure 4. The detailed architecture of our proposed decompose-and-fuse control branch. + +diffusion step $z_{t}$ , as previous work [21] shows that including diffusion latent as an additional condition can improve image quality. The main diffusion denoising network is a pretrained U-Net. The over-exposed image is first encoded using a pretrained VAE before entering the diffusion module, and the outputs are converted back to the image space using the pretrained decoder. + +The main differences between our solution and general diffusion-based image enhancement are two-fold. First, we propose a novel decompose-and-fuse control branch to inject two input images and diffusion latent as a control signal, as Fig. 11 (b) shows that naively injecting this information may not be able to guide diffusion to faithfully inpaint missing highlights obtained from the under-exposed image. Second, we train an additional fidelity control branch to provide faithful structure and color information for the decoding process via shortcuts. Details are described below. + +Decompose-and-fuse control branch. Fig. 4 shows our control branch. We use the over-exposed image $I_{oe}$ as the main control signal and the under-exposed image $I_{ue}$ as the + +soft guidance. Following ControlNet [65], we copy the encoder and middle blocks of the denoising U-Net as the main extractor, but update their weights during training. A simple soft guidance is to use the encoded under-exposed image latent $y_{ue}$ from the VAE encoder, combined with the overexposed latent $y_{oe}$ . However, the underexposed images are often too dark to be directly used as soft guidance, since the model may entirely ignore this guidance. + +Therefore, we decompose the under-exposed image into the color and structure information, both of which are robust to brightness changes. Specifically, we use the normalized image as the structure component, similar to SSIM [46], as: + +$$ +S _ {u e} = \left(Y _ {u e} - \mu \left(Y _ {u e}\right)\right) / \sigma \left(Y _ {u e}\right), \tag {2} +$$ + +where $Y_{ue}$ represents the luminance channel of $I_{ue}$ under YUV space, $\mu(\cdot)$ and $\sigma(\cdot)$ denote the mean intensity and standard deviation, respectively. The chroma channels (UV) are used as color information. The extracted structure and color information are further encoded using trained color and structure extractors (gray GE block in Fig. 4). Following [65], we implement GE using several simple convolution layers to extract multi-scale features. + +At last, the extracted features are injected into the main extractor with a multi-scale cross attention, as shown in the bottom part of Fig. 4. The output of each cross attention module is fed into both the next level of the main extractor and the corresponding U-Net block, using zero convolution. Fidelity control branch. Even with the control block, we sometimes observe undesirable modifications of texture introduced by VAE. An example is shown in Fig. 11 (c). Therefore, to further improve fidelity, we design a Fidelity Control Branch (FCB), inspired by [47]. FCB mitigates texture distortions by injecting features into the VAE decoder. It has a similar architecture as the decompose-and-fuse control branch, with two main differences: 1) the main extractor of FCB adopts the same structure as the VAE Encoder, + +![](images/e7c2a68532deed2df235cb26f482d510940848446e1c592acbe6f8aa5ccf6695.jpg) +Figure 5. The illustration of our training data synthesis pipeline. + +rather than the denoising U-Net, to provide corresponding shortcuts to the VAE Decoder, with adjustments to the soft guidance extractor as well, and 2) the main extractor of the FCB directly takes the over-exposed image as input. To train the fidelity control branch, we freeze the pre-trained VAE Encoder and Decoder and encode the ground truth $I_{gt}$ to latent space, simulating the denoised latent $z_0$ during the inference. Then we input the corresponding overexposed image and under-exposed image to the fidelity control branch to extract faithful features. At last, the VAE Decoder decodes the compressed latent to a reconstructed image $\hat{I}_{gt}$ . We adopt the reconstruction loss $||\hat{I}_{gt} - I_{gt}||_1$ as an additional loss term. + +# 3.3. Training data synthesis + +Preparing the training data for the proposed guided inpainting network is challenging. To train the model, ideally we need a large-scale HDR dataset that 1) covers different dynamic scenes, 2) has large exposure variance up to 9 stops, and 3) contains ground truth fusion result. However, no existing dataset satisfy all requirements. + +To solve this issue, we propose a novel training data synthesis pipeline. Specifically, as shown in Fig. 5, we first randomly sample an image sequence with $N$ frames from a video dataset [55]. To model large motion, we select the first and the last frames. Then, similar to the pre-alignment stage, we estimate the bidirectional optical flow between two selected frames using [42] and compute a pseudo occlusion mask via forward-backward consistency check. Subsequently, we randomly sample an under-exposed image patch from the high-quality static dataset [1] (ground truths are pre-aligned), resize the pseudo occlusion mask to match the patch size, and mask out the pseudo occluded region to synthesize a pseudo pre-aligned output. With our synthesized training data, our UltraFusion learns to handle dynamic scenes with only static multi-exposure image pairs. + +# 4. Experiment + +# 4.1. Experimental setting + +Datasets. We utilize SICE [1] dataset and Video-90K [55] dataset to synthesize our training data. Following previous works [1, 22], we select images with the highest and lowest brightness in each exposure bracket of the SICE dataset + +![](images/68b45325848c2b6028dde807d2de1ff012aa2f924f636e9d09b76224060d7c5c.jpg) +Figure 6. (a) The data distribution of our benchmark. Coordinate value of exposure time represents the upper boundary. (b) The user study result on our benchmark. + +as inputs. We evaluate our method on both static datasets and dynamic datasets. For evaluation, we use the MEFB dataset [66] with 100 static under/over-exposed image pairs and RealHDRV [39], a dynamic HDR deghosting test set containing 50 scenes with varying motion patterns. + +UltraFusion benchmark. Existing exposure fusion benchmark cannot fully evaluate fusing in real-world challenging conditions, as they either lack realistic motion [66] or have limited dynamic range [15, 39, 44]. Therefore, we collect a new real-world UltraFusion benchmark, which contains 100 real-captured under/over-exposed image pairs. Compared to previous datasets, our benchmark is more challenging for three reasons: 1) Our benchmark features larger exposure differences between the two input images (up to 9 stops). 2) It includes more realistic motion, with many scenes containing extensive and unintentional foreground movement. 3) Our benchmark is highly diverse, encompassing daytime, nighttime, indoor, and outdoor scenes captured by DSLR Camera (Canon R8) and mobile phones (iPhone12, iPhone13, Redmi K50 Pro and OPPO Reno8 Pro). We summarize the exposure difference distribution and exposure time distribution of our benchmark in Fig. 6 (a). It can be observed that our benchmark covers a wide range of exposure differences and diverse exposure times, which can be used to effectively test the robustness of the HDR methods. + +Implementation details. We leverage the generative prior encapsulated in Stable Diffusion V2.1 [37]. The decompose-and-fuse control branch (DFCB) is trained for 140k iterations with batch size $bs = 32$ on 8 NVIDIA RTX 4090 GPUs. We also train the fidelity control branch for 1000K iteration with batch size $bs = 1$ on a single NVIDIA RTX 4090 GPU. Adam is adopted as the optimizer and the learning rate is fixed to 0.0001. To adapt HDR reconstruction methods to 2 differently exposed inputs, we re-implement them by following their default settings. + +Evaluation metrics. We utilize four widely used non-reference metric MUSIQ [16], DeQA-Score [62], PAQ2PIQ [61] and HyperIQA [41] for quantitative comparison. Moreover, for static dataset [66], as no ground truths are provided, we select the task-specific MEF-SSIM [27] for structure retention evaluation. For dynamic dataset [39] with HDR ground truths, TMQI [60] is used to evaluate the performance of fidelity and naturalness. In addition, we conduct a user study on our Ultrafusion benchmark to + +Table 1. Quantitative comparisons on the static MEFB dataset [66]. + +
TypeMethodMEFB [66]
MUSIQ↑DeQA-Score↑PAQ2PIQ↑HyperIQA↑MEF-SSIM↑
HDR Rec.HDR-Transformer [25]63.102.98371.360.59960.8626
SCTNet [44]63.133.02171.480.60680.8777
SAFNet [17]61.702.51872.670.56460.7711
MEFDeepfuse [35]52.582.94567.960.43990.8968
MEF-GAN [53]50.592.81869.990.37450.7722
U2Fusion [52]63.393.21972.230.51590.9304
Defusion [20]62.703.11872.820.54550.9062
MEFLUT [13]65.713.27771.210.52670.8608
HSDS-MEF [48]66.763.54472.600.60260.9520
TC-MoA [69]64.603.35571.850.53940.9636
OursUltraFusion68.823.88173.800.64820.9385
+ +![](images/a060de34805c5ccb86e319fe2ec960ca25d98d5de577432f89cfa9d0c546e0cf.jpg) +Figure 7. Trade-off curve between MEF-SSIM and MUSIQ on MEFB dataset [66]. Our UltraFusion achieves the best trade-off between image quality and information preservation. + +Table 2. Quantitative comparisons on dynamic RealHDRV dataset [39] and our challenging UltraFusion benchmark. + +
TypeMethodRealHDRV [39]UltraFusion Benchmark
TMQI↑MUSIQ↑DeQA-Score↑PAQ2PIQ↑HyperIQA↑MUSIQ↑DeQA-Score↑PAQ2PIQ↑HyperIQA↑
HDR Rec.HDR-Transformer [25]0.868062.243.49670.330.522563.662.90972.830.5619
SCTNet [44]0.871562.693.53270.740.527261.843.10272.940.5888
SAFNet [17]0.872662.073.50670.480.515661.502.17973.150.5487
MEFDefusion [20]0.818756.603.30268.380.485660.313.35271.870.5463
MEFLUT [13]0.829762.423.31570.040.502063.623.34371.730.5074
HSDS-MEF [48]0.832361.763.36071.110.505464.543.62773.420.5923
OursUltraFusion0.892567.513.83073.400.583368.413.95775.180.6214
+ +![](images/93fae921e5e476884387d948973f1976642988310f9507a1a08887028e930a3e.jpg) +Inputs + +![](images/f63c111f10620fd0b58b32182118a0267571a37ccdfb6e10254bf33f7d5b2f90.jpg) + +![](images/4b2c30a02693391b912120a96138c5f9c5a574fafb51e686b2611241e193ea96.jpg) +HSDS-MEF + +![](images/4a76d138e85d212c3ca80e0004d77fd085d52211103a356828fcfbbfb4b74b71.jpg) + +![](images/50091b5090b783499c421556760e57a9ce03aadb7eb1f33d0b04aa83ed869fc8.jpg) +TC-MoA + +![](images/9b6897253144233b47e28c2683745d54bdc929f8f82eb245058545b65d401c7d.jpg) + +![](images/b93b39a851c7892229290213dfe78c449b4845a73dbd35ddf953521c833b168b.jpg) +HDR-Transformer +Figure 8. Visual comparisons of different exposure fusion methods on static MEFB dataset [66]. + +![](images/1973d59ef8ebbd782431021e85f5f6b24221bc452bbc94b5a823cd3b3e31eebb.jpg) + +![](images/57d20f7dd140b1ddc5d5f1e2ca505fabc93d0bbb08787715cf6646b1bfb52a41.jpg) +SCTNet + +![](images/f0f1c525374f4e41b8fe2a761b284a5c327d3dc005aa45f4debf56473acb5de5.jpg) + +![](images/1da5b84a86915fb13d404279e971376aa047b4317358a70ecb05753c10a2259e.jpg) +Ours + +perform subjective evaluation. + +# 4.2. Comparisons with HDR imaging methods + +We compare our method with state-of-the-art HDR Imaging methods, including HDR reconstruction methods HDRTransformer [25], SCTNet [44], SAFNet [17], and multi-exposure fusion methods Deepfuse [35], MEF-GAN [53], U2Fusion [52], Defusion [20], MEFLUT [13], HSDS-MEF [48], TC-MoA [69]. As HDR reconstruction methods cannot output an LDR image directly, we use professional software Photomatix [32] to perform tone mapping. + +Evaluation on static dataset. We evaluate the fusion performance on the MEFB dataset [66], focusing on large exposure differences. As shown in Tab. 1, our method outperforms other methods on all four non-reference metrics (MUSIQ, DeQA-Score, PAQ2PIQ, and HyperIQA). Specif + +ically, our proposed UltraFusion achieves 2.06 gain in terms of MUSIQ compared to HSDS-MEF. In terms of MEFSSIM, as shown in Fig. 7, our baseline model (ControlNet [65]) achieves better image quality but lacks fidelity, while HSDS-MEF [48] retains more information from inputs at the cost of quality. Our UltraFusion outperforms most algorithms, and achieves similar fidelity scores compared to HSDS-MEF and TC-MoA but with much higher image quality (non-reference metrics), indicating the best trade-off between visual quality and information preservation. The qualitative comparison in Fig. 8 further validates this. In contrast, HDR reconstruction methods (HDRTransformer and SCTNet) often miss some detail in highlights and MEF methods (HSDS-MEF and TC-MoA) introduce unnatural transition from brighter to dark regions. + +Evaluation on dynamic dataset. To further illustrate the + +![](images/bc95d2be8f4d67ad38d53524cfc904f00289bfc22f5c7218c3eeb62ee2fcf447.jpg) +Figure 9. Visual comparisons on our captured UltraFusion benchmark. + +![](images/0c5de0821e51669d6cfebf1021f7caf4ebfee2582ec8eb6a78669ed62e7d1548.jpg) + +![](images/fc03860d1f45d750b32b003f2f060649ac00a6b85631b5fc9e857954e105b739.jpg) +Figure 10. Visual results on dynamic RealHDRV dataset [39]. + +robustness of UltraFusion to global and local motion, we use the RealHDRV dataset [39]. We extract the corresponding over-exposed image from the HDR ground truth as input. Tab. 2 demonstrates that our UltraFusion achieve state-of-the-art performance in terms of all metrics. For dynamic scenes, TMQI metric is a particularly important metric, as it is specially designed for HDR evaluation by assessing structural similarity between fusion output and ground truth. Since MEF methods are mainly designed for static scenes, they lack the capability to handle motion, resulting in low TMQI scores and overlay artifacts, as shown in Fig. 10. HDR reconstruction methods trained on dynamic datasets achieve better performance, but still produce noticeable artifact. On the other side, due to our soft inpainting guidance, UltraFusion is much more robust to misalignment and occlusion and the fusion output contains almost no artifacts. It achieves the highest TMQI in Tab. 2 and the best visual result in Fig. 10. + +Evaluation on our UltraFusion benchmark. At last, the evaluation on our benchmark validates the robustness of UltraFusion in the most challenging scenes. On all four non-reference metrics, our method outperforms competitors by a large margin, as shown in Tab. 2. The qualitative comparison in Fig. 9 is also consistent with the quantitative metrics. For example, in the red zoom-in patch, integrating the highly bright sun from the under-exposed image into the over-exposed image is extremely challenging. Other methods fail to maintain the shape of the sun or preserve high contrast in the fused region, while our method naturally reconstructs the sun, preserving its appearance and visual pleasing tone of the whole image. Moreover, we conduct a user study on our proposed benchmark. Specifically, we randomly select 20 scenes from our benchmark and invite 136 different users to participate. For each scene, each user + +Table 3. Ablation studies of three key components of our proposed UltraFusion on RealHDRV dataset [39]. + +
ModelTMQI↑MUSIQ↑
w/o Alignment Strategy0.742763.67
w/o Decompose-and-Fuse Control Branch0.887266.94
w/o Fidelity Control Branch0.876367.36
Ours UltraFusion0.892567.51
+ +is asked to compare our method with a randomly chosen baseline. The user study in Fig. 6 (b) indicates our method is more favored by users than competitors. This outcome aligns with the non-reference metric evaluation, showing that our method produces more natural images. More visual comparisons are available in the supplementary. + +# 4.3. Ablation studies + +To validate the effectiveness of our UltraFusion, we conduct ablation studies on the three key components, followed by an in-depth exploration of the designs of them. + +Key components. First, we perform the ablation study of the proposed three key components, including the alignment strategy, decompose-and-fuse control branch, and fidelity control branch, on the RealHDRV [39] dataset, as shown in Tab. 3. We first remove our alignment strategy, opting instead to train the model directly on the original SICE [1] dataset and input differently exposed image pairs without coarse alignment. Without the pre-alignment stage and the data synthesis pipeline designed for it, performance drops significantly in terms of TMQI. Fig. 11 (a) also shows that the model fails to implicitly align large motions. Then, we replace the decompose-and-fuse control branch (DFCB) with the vanilla ControlNet [65], and it fails to fuse features with large exposure differences, which leads to the loss of details (see Fig. 11 (b)) and a decrease in TMQI. Finally, when we exclude the fidelity control branch (FCB), Ultra-Fusion loses the capability to maintain the detailed structure and vivid color, as indicated by Fig. 11 (c). + +Alignment strategy. Second, we perform ablation of the proposed alignment strategy, which is the most critical design for our framework. Our alignment strategy mainly consists of a data synthesis pipeline at training time and a pre-alignment module at testing time. Without both of them, the whole model degrades to a multi-exposure fusion method as shown in Fig. 12 (c). Performing pre-alignment can reduce + +![](images/5abb710dbb0e448dd45cb1770ae17b51a45e59abde7b099cebaa43fd2be006ba.jpg) + +![](images/e787c4b7688cc487b7374b6bb6f2522faf7659013dde49b96d56eef566d19ac8.jpg) + +![](images/6e0f7a164f0a723e511fa6f2a2dde3cab4370632af847f02973953b46220f0ce.jpg) +Under-Exposure Over-Exposure + +![](images/8b152c2bbf15ea46b11b42ef9e184c4d92cce9513e098abbfab65abf19dbe7f9.jpg) +Align Strategy + +![](images/acc38ca9ba787f3e692385b170efda8b705adbd9f32bc6db7d501a053f18f9b2.jpg) + +![](images/024f5917bf46a81601c7326afcd47098d80a7b0b6624c032a70d8d8dd8f4d399.jpg) +Under-Exposure Over-Exposure + +![](images/a9a3606b36d94eb2244055ecd14827cb0787d0afc33163268628c466fc9a0669.jpg) +w/o DFCB +w/o DFCB Ours +w/o FCB + +![](images/0ed35ecb9248ccc15261403a4788d48b614a161c45496a7dedd22d482d0741ba.jpg) +Under-Exposure + +![](images/a0584271bf223fb9120b7cbe2268dcfea0b8b1e6238c885c3098bb2a6243505b.jpg) +Over-Exposure + +![](images/c10b1b7f5e11a52a91f74342a4176e312873317f2edf197febdeb6239222bf0b.jpg) +Figure 11. Visual results of ablating each key component of our method. Each key component contributes to the final results. +(a) Under Exposure + +![](images/3a56d6a8e5d4d08ace351fa76c5620bab4e4ec3e9a22c4c92f328bfaa425edf9.jpg) +(b) Over Exposure + +![](images/cc8d9ed2386622e856315b3863ae14254d27e91cbe1e66801442da63f7f09fd2.jpg) +(c) $w / o$ Data Syn. & Pre-Alignment + +![](images/6e8091871fb133a04c7174c7f44afe54d8861bc0a0db0eefd4837526ec2c7eb7.jpg) +(d) $w / o$ Data Syn. +Figure 12. Effectiveness of our alignment strategy. + +![](images/8672368948b28c961db0260e2c6c7affda4d6a2b9bd28200b0f3f6477080f800.jpg) +(e) $w / o$ Pre-Alignment + +![](images/a9a6fcbf963dd66ee7dd71cd7c787b29ff98798b7e186771ac1656ebf13bec01.jpg) +(f) Ours Alignment Strategy + +artifacts to some extent (see Fig. 12 (d)), but may still be prone to alignment errors. Our data synthesis pipeline can improves the robustness of our algorithm to unaligned conditions, but motion in the occluded regions still cannot be solved, as shown in Fig. 12. After combining data synthesis pipeline and pre-alignment stage, our algorithm demonstrates strong capability towards dynamic scenes. + +Decompose-and-fuse control branch. At last, we evaluate how the form of soft guidance provided by the underexposed image impacts the recovery of the highlight regions. When using the RGB under-exposed image $I_{ue}$ as guidance, the model will ignore some details in the output (Fig. 13 (c)). Replacing $I_{ue}$ with its structure information $S_{ue}$ retains more details (Fig. 13 (d)), but without generating vivid color. By incorporating both under-exposed color information $C_{ue}$ and structure information $S_{ue}$ , the reconstructed results can maintain more details and color consistency (see Fig. 13 (e)). At last, multi-scale cross-attention further improves the fusion result (see Fig. 13 (f)). + +# 4.4. Application on general image fusion + +One advantage of the UltraFusion is that it can be extended to general image fusion, thanks to the flexibility of our + +![](images/d2e126ab4c4120a57e9b0c1262da9e33a36254f8a2d6d8a19706c9077e7584e7.jpg) +(a) Under Exposure + +![](images/1234e3eac74e46d5b3d1ddd6537297a776e49e8a65e0af8cf33e6d103351186f.jpg) +(b) Over Exposure + +![](images/e70afd959c2dfca63614c9e5376b855b8494bbfea371877023ac947966c1f2fe.jpg) + +![](images/418f6375d1bb490a9b7f192e3db8751a8d4df423921c802fa2d980cb5bad9348.jpg) +(d) $S_{ue}$ + +![](images/0e7e24886d1a941348040818f60e36884a30d1ac14e1cc1eb74b10cd7e609740.jpg) +(e) $S_{ue} + C_{ue}$ + +![](images/52273cdf2c8c480e0843f8f4c08895e69350fea189c8c30ad70cff61494c1731.jpg) +(c) $I_{ue}$ +(f) DFCB + +![](images/93a6f208f33de5c9abd764e12cb155a7145b6c512100e50a6a7eb6496a34a3bb.jpg) +Figure 13. Detailed ablation study on the design choices of decompose-and-fuse control branch. +(a) Input Over Exposure +Figure 14. Extension to general fusion. Given totally different underexposed images as guidance (upper right corner), our method can also generate different fusion results. + +![](images/34b8effd521b6019697b2527d039a37a14246e5cc88117b100cd7084c79e6305.jpg) +(b) Our Fusion Result +(c) Another Fusion w/ Diff. Highlight + +![](images/766deb6635d7215f24f16182405356b0287fcd7dccdefca626f397f52bb120e1.jpg) + +guided inpainting. To illustrate this potential, we explore one additional interesting demo to fuse two irrelevant images captured by different cameras at different locations. As shown in Fig. 14, UltraFusion successfully copies the moon (b) or blue sky (c) to the over-exposed image, unlocking many interesting potential applications (e.g., image harmonization). + +# 5. Conclusion + +In this work, we introduce a novel approach to HDR imaging, tackling challenges presented by significant exposure differences and large motion. By modeling the fusion process as a guided inpainting problem and utilizing the under-exposed image for soft guidance, our method is robust with alignment errors and circumvents tone mapping, resulting in natural, artifact-free outputs. We also propose decompose-and-fuse control branch and fidelity control branch to improve feature modulation and fidelity preservation of ControlNet. Extensive experiments on existing datasets and our captured benchmark demonstrate the robustness and effectiveness of our method compared to previous HDR methods. + +With extreme exposure differences and challenging non-rigid motion, occlusion mask estimation may introduce errors, causing the restoration of certain highlight regions degrades to single image HDR. While our method is able to use diffusion priors to restore these highlights, restoration without under-exposed information remains unreliable. Moreover, it takes almost 3.3s to fuse $512 \times 512$ inputs on an NVIDIA RTX 4090 GPU. A more exposure-robust optical flow algorithm and a faster implementation are highly desirable, and we leave them for future work. + +# Acknowledgment + +This work was supported by the National Key R&D Program of China No.2022ZD0160201, Shanghai Artificial Intelligence Laboratory and RGC Early Career Scheme (ECS) No. 24209224. + +# References + +[1] Jianrui Cai, Shuhang Gu, and Lei Zhang. Learning a deep single image contrast enhancer from multi-exposure images. IEEE Transactions on Image Processing, 27(4):2049-2062, 2018. 3, 5, 7 +[2] Cong Cao, Huanjing Yue, Xin Liu, and Jingyu Yang. Unsupervised HDR image and video tone mapping via contrastive learning. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 3 +[3] Jie Chen, Zaifeng Yang, Tsz Nam Chan, Hui Li, Junhui Hou, and Lap-Pui Chau. Attention-guided progressive neural texture fusion for high dynamic range image restoration. IEEE Transactions on Image Processing, 31:2661-2672, 2022. 2, 3 +[4] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024. 3 +[5] Ciprian Corneanu, Raghudeep Gadde, and Aleix M Martinez. Latentpaint: Image inpainting in latent space with diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4334-4343, 2024. 3 +[6] Abhishek Goswami, Aru Ranjan Singh, Francesco Banterle, Kurt Debattista, and Thomas Bashford-Rogers. Semantic aware diffusion inverse tone mapping. arXiv preprint arXiv:2405.15468, 2024.3 +[7] Michael D Grossberg and Shree K Nayar. Determining the camera response from images: What is knowable? IEEE Transactions on pattern analysis and machine intelligence, 25(11):1455-1467, 2003. 3 +[8] Yuanshen Guan, Ruikang Xu, Mingde Yao, Ruisheng Gao, Lizhi Wang, and Zhiwei Xiong. Diffusion-promoted HDR video reconstruction. arXiv preprint arXiv:2406.08204, 2024.3 +[9] Samuel W Hasinoff, Dillon Sharlet, Ryan Geiss, Andrew Adams, Jonathan T Barron, Florian Kainz, Jiawen Chen, and Marc Levoy. Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM Transactions on Graphics (ToG), 35(6):1-12, 2016. 1, 2 +[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3 +[11] Litao Hu, Huaijin Chen, and Jan P Allebach. Joint multiscale tone mapping and denoising for HDR image enhancement. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 729-738, 2022. 3 + +[12] Tao Hu, Qingsen Yan, Yuankai Qi, and Yanning Zhang. Generating content for HDR deghosting from frequency view. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25732-25741, 2024. 3 +[13] Ting Jiang, Chuan Wang, Xinpeng Li, Ru Li, Haoqiang Fan, and Shuaicheng Liu. Meflut: Unsupervised 1d lookup tables for multi-exposure image fusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10542-10551, 2023. 3, 6 +[14] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. arXiv preprint arXiv:2403.06976, 2024. 3 +[15] Nima Khademi Kalantari, Ravi Ramamoorthi, et al. Deep high dynamic range imaging of dynamic scenes. ACM Trans. Graph., 36(4):144-1, 2017. 2, 3, 5 +[16] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 5 +[17] Lingtong Kong, Bo Li, Yike Xiong, Hao Zhang, Hong Gu, and Jinwei Chen. Safnet: Selective alignment fusion network for efficient HDR imaging. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2, 3, 6 +[18] Baiang Li, Sizhuo Ma, Yanhong Zeng, Xiaogang Xu, Youqing Fang, Zhao Zhang, Jian Wang, and Kai Chen. Sagiri: Low dynamic range image enhancement with generative diffusion prior. arXiv preprint arXiv:2406.09389, 2024. 3 +[19] Hui Li, Kede Ma, Hongwei Yong, and Lei Zhang. Fast multiscale structural patch decomposition for multi-exposure image fusion. IEEE Transactions on Image Processing, 29: 5805-5816, 2020. 3 +[20] Pengwei Liang, Junjun Jiang, Xianming Liu, and Jiayi Ma. Fusion from decomposition: A self-supervised decomposition approach for image fusion. In European Conference on Computer Vision, pages 719-735. Springer, 2022. 6 +[21] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Wanli Ouyang, Yu Qiao, and Chao Dong. Diff-bir: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070, 2023. 3, 4 +[22] Renshuai Liu, Chengyang Li, Haitao Cao, Yinglin Zheng, Ming Zeng, and Xuan Cheng. Emef: ensemble multi-exposure image fusion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1710-1718, 2023. 3, 5 +[23] Yuhao Liu, Zhanghan Ke, Fang Liu, Nanxuan Zhao, and Rynson WH Lau. Diff-plugin: Revitalizing details for diffusion-based low-level tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4197-4208, 2024. 3 +[24] Zhen Liu, Yinglong Wang, Bing Zeng, and Shuaicheng Liu. Ghost-free high dynamic range imaging with context-aware transformer. In European Conference on computer vision, pages 344-360. Springer, 2022. 2, 3 +[25] Zhen Liu, Yinglong Wang, Bing Zeng, and Shuaicheng Liu. Ghost-free high dynamic range imaging with context-aware + +transformer. In European Conference on computer vision, pages 344-360. Springer, 2022. 1, 6 +[26] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 3 +[27] Kede Ma, Kai Zeng, and Zhou Wang. Perceptual quality assessment for multi-exposure image fusion. IEEE Transactions on Image Processing, 24(11):3345-3356, 2015. 5 +[28] Kede Ma, Hui Li, Hongwei Yong, Zhou Wang, Deyu Meng, and Lei Zhang. Robust multi-exposure image fusion: a structural patch decomposition approach. IEEE Transactions on Image Processing, 26(5):2519-2532, 2017. 3 +[29] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 3 +[30] Tom Mertens, Jan Kautz, and Frank Van Reeth. Exposure fusion. In 15th Pacific Conference on Computer Graphics and Applications (PG'07), pages 382-390. IEEE, 2007. 2, 3 +[31] Ozan Özdenizci and Robert Legenstein. Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10346-10357, 2023. 3 +[32] Photomatix. Commercially-available hdr processing software. https://www.hdrsoft.com/.6 +[33] K Ram Prabhakar, Rajat Arora, Adhitya Swaminathan, Kunal Pratap Singh, and R Venkatesh Babu. A fast, scalable, and reliable deghosting method for extreme exposure fusion. In 2019 IEEE International Conference on Computational Photography (ICCP), pages 1-8. IEEE, 2019. 3 +[34] K Ram Prabhakar, Susmit Agrawal, Durgesh Kumar Singh, Balraj Ashwath, and R Venkatesh Babu. Towards practical and efficient high-resolution HDR deghosting with cnn. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16, pages 497-513. Springer, 2020. 2, 3 +[35] K Ram Prabhakar, V Sai Srikar, and R Venkatesh Babu. Deepfuse: A deep unsupervised approach for exposure fusion with extreme exposure image pairs. In Proceedings of the IEEE international conference on computer vision, pages 4714-4722, 2017. 3, 6 +[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 3 +[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 5 +[38] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In Proceedings of the IEEE/CVF + +Conference on Computer Vision and Pattern Recognition, pages 8839-8849, 2024. 3 +[39] Yong Shu, Liquan Shen, Xiangyu Hu, Mengyao Li, and Zihao Zhou. Towards real-world HDR video reconstruction: A large-scale benchmark dataset and a two-stage alignment network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2879-2888, 2024. 3, 5, 6, 7 +[40] Jou Won Song, Ye-In Park, Kyeongbo Kong, Jaeho Kwak, and Suk-Ju Kang. Selective transhdr: Transformer-based selective HDR imaging using ghost region mask. In European Conference on Computer Vision, pages 288-304. Springer, 2022. 2, 3 +[41] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3667-3676, 2020. 5 +[42] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 3, 5 +[43] Steven Tel, Zongwei Wu, Yulun Zhang, Barthélémy Heyrman, Cédric Demonceaux, Radu Timofte, and Dominique Gingac. Alignment-free HDR deghosting with semantics consistent transformer. arXiv preprint arXiv:2305.18135, 2023. 2, 3 +[44] Steven Tel, Zongwei Wu, Yulun Zhang, Barthélémy Heyrman, Cédric Demonceaux, Radu Timofte, and Dominique Gingac. Alignment-free HDR deghosting with semantics consistent transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12836-12845, 2023. 5, 6 +[45] Yael Vinker, Inbar Huberman-Spiegelglas, and Raanan Fattal. Unpaired learning for high dynamic range image tone mapping. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14657-14666, 2021. 3 +[46] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 4 +[47] Shuchen Weng, Peixuan Zhang, Yu Li, Si Li, Boxin Shi, et al. L-cad: Language-based colorization with any-level descriptions using diffusion priors. Advances in Neural Information Processing Systems, 36, 2024. 3, 4 +[48] Guanyao Wu, Hongming Fu, Jinyuan Liu, Long Ma, Xin Fan, and Risheng Liu. Hybrid-supervised dual-search: Leveraging automatic learning for loss-free multi-exposure image fusion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5985-5993, 2024. 1, 3, 6 +[49] Shangzhe Wu, Jiarui Xu, Yu-Wing Tai, and Chi-Keung Tang. Deep high dynamic range imaging with large foreground motions. In Proceedings of the European Conference on Computer Vision (ECCV), pages 117–132, 2018. 2, 3 +[50] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting + +with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22428-22437, 2023. 3 +[51] Gangwei Xu, Yujin Wang, Jinwei Gu, Tianfan Xue, and Xin Yang. Hdrflow: Real-time HDR video reconstruction with large motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24851-24860, 2024. 3 +[52] Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502-518, 2020. 3, 6 +[53] Han Xu, Jiayi Ma, and Xiao-Ping Zhang. Mef-gan: Multiexposure image fusion via generative adversarial networks. IEEE Transactions on Image Processing, 29:7203-7216, 2020. 6 +[54] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8121-8130, 2022. 3 +[55] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. International Journal of Computer Vision, 127:1106-1125, 2019. 5 +[56] Qingsen Yan, Dong Gong, Qinfeng Shi, Anton van den Hengel, Chunhua Shen, Ian Reid, and Yanning Zhang. Attention-guided network for ghost-free high dynamic range imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1751-1760, 2019. 2, 3 +[57] Qingsen Yan, Lei Zhang, Yu Liu, Yu Zhu, Jinqiu Sun, Qinfeng Shi, and Yanning Zhang. Deep HDR imaging via a non-local network. IEEE Transactions on Image Processing, 29: 4308-4322, 2020. 2, 3 +[58] Qingsen Yan, Tao Hu, Yuan Sun, Hao Tang, Yu Zhu, Wei Dong, Luc Van Gool, and Yanning Zhang. Towards high-quality HDR dehosting with conditional diffusion models. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 3 +[59] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3 +[60] Hojatollah Yeganeh and Zhou Wang. Objective quality assessment of tone-mapped images. IEEE Transactions on Image processing, 22(2):657-667, 2012. 5 +[61] Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3575–3585, 2020. 5 +[62] Zhiyuan You, Xin Cai, Jinjin Gu, Tianfan Xue, and Chao Dong. Teaching large language models to regress accurate image quality scores using score distribution. arXiv preprint arXiv:2501.11561, 2025. 5 + +[63] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photorealistic image restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25669-25680, 2024. 3 +[64] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 3 +[65] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3, 4, 6, 7 +[66] Xingchen Zhang. Benchmarking and comparing multi-exposure image fusion algorithms. Information Fusion, 74: 111-131, 2021. 5, 6 +[67] Shengyu Zhao, Yilun Sheng, Yue Dong, Eric I Chang, Yan Xu, et al. Maskflownet: Asymmetric feature matching with learnable occlusion mask. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6278-6287, 2020. 3 +[68] Zixiang Zhao, Lilun Deng, Haowen Bai, Yukun Cui, Zhipeng Zhang, Yulun Zhang, Haotong Qin, Dongdong Chen, Jiangshe Zhang, Peng Wang, et al. Image fusion via vision-language model. arXiv preprint arXiv:2402.02235, 2024.3 +[69] Pengfei Zhu, Yang Sun, Bing Cao, and Qinghua Hu. Task-customized mixture of adapters for general image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7099-7108, 2024. 3, 6 \ No newline at end of file diff --git a/2025/UltraFusion_ Ultra High Dynamic Imaging using Exposure Fusion/images.zip b/2025/UltraFusion_ Ultra High Dynamic Imaging using Exposure Fusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d1e248cabe65256020d03eefe233de95465feb64 --- /dev/null +++ b/2025/UltraFusion_ Ultra High Dynamic Imaging using Exposure Fusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05cea2ec58f9a6a0a5ae14627edb4902e3c7d72b1ff0a00ed2e3c5794c012654 +size 908639 diff --git a/2025/UltraFusion_ Ultra High Dynamic Imaging using Exposure Fusion/layout.json b/2025/UltraFusion_ Ultra High Dynamic Imaging using Exposure Fusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cad0b348dc0336df3b95b50ca18dda88ab99440b --- /dev/null +++ b/2025/UltraFusion_ Ultra High Dynamic Imaging using Exposure Fusion/layout.json @@ -0,0 +1,11377 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 100, + 103, + 510, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 103, + 510, + 121 + ], + "spans": [ + { + "bbox": [ + 100, + 103, + 510, + 121 + ], + "type": "text", + "content": "UltraFusion: Ultra High Dynamic Imaging using Exposure Fusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 141, + 524, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 141, + 524, + 213 + ], + "spans": [ + { + "bbox": [ + 83, + 141, + 524, + 213 + ], + "type": "text", + "content": "Zixuan Chen" + }, + { + "bbox": [ + 83, + 141, + 524, + 213 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 83, + 141, + 524, + 213 + ], + "type": "text", + "content": " Yujin Wang" + }, + { + "bbox": [ + 83, + 141, + 524, + 213 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 83, + 141, + 524, + 213 + ], + "type": "text", + "content": " Xin Cai" + }, + { + "bbox": [ + 83, + 141, + 524, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 83, + 141, + 524, + 213 + ], + "type": "text", + "content": " Zhiyuan You" + }, + { + "bbox": [ + 83, + 141, + 524, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 83, + 141, + 524, + 213 + ], + "type": "text", + "content": " Zheming Lu" + }, + { + "bbox": [ + 83, + 141, + 524, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 83, + 141, + 524, + 213 + ], + "type": "text", + "content": " Fan Zhang" + }, + { + "bbox": [ + 83, + 141, + 524, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 141, + 524, + 213 + ], + "type": "text", + "content": " Shi Guo" + }, + { + "bbox": [ + 83, + 141, + 524, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 141, + 524, + 213 + ], + "type": "text", + "content": " Tianfan Xue" + }, + { + "bbox": [ + 83, + 141, + 524, + 213 + ], + "type": "inline_equation", + "content": "^{2,1}" + }, + { + "bbox": [ + 83, + 141, + 524, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 141, + 524, + 213 + ], + "type": "text", + "content": "Shanghai AI Laboratory " + }, + { + "bbox": [ + 83, + 141, + 524, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 83, + 141, + 524, + 213 + ], + "type": "text", + "content": "The Chinese University of Hong Kong " + }, + { + "bbox": [ + 83, + 141, + 524, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 83, + 141, + 524, + 213 + ], + "type": "text", + "content": "Zhejiang University " + }, + { + "bbox": [ + 83, + 141, + 524, + 213 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{zxchen}, \\mathrm{zheminglu}\\} @ \\mathrm{zju.edu.cn}, \\{\\mathrm{wangyujin}, \\mathrm{zhangfan}, \\mathrm{guoshi}\\} @ \\mathrm{pjlab.org.cn}" + }, + { + "bbox": [ + 83, + 141, + 524, + 213 + ], + "type": "text", + "content": " caixin@link.cuhk.edu.hk, zhiyuanyou@foxmail.com, tfxue@ie.cuhk.edu.hk" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 59, + 224, + 138, + 277 + ], + "blocks": [ + { + "bbox": [ + 59, + 224, + 138, + 277 + ], + "lines": [ + { + "bbox": [ + 59, + 224, + 138, + 277 + ], + "spans": [ + { + "bbox": [ + 59, + 224, + 138, + 277 + ], + "type": "image", + "image_path": "8c0cd337a93d395280ba4e556888a77ce0c9e50727e6122374703d8a3e498ce7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 59, + 280, + 138, + 333 + ], + "blocks": [ + { + "bbox": [ + 59, + 280, + 138, + 333 + ], + "lines": [ + { + "bbox": [ + 59, + 280, + 138, + 333 + ], + "spans": [ + { + "bbox": [ + 59, + 280, + 138, + 333 + ], + "type": "image", + "image_path": "36d3e4ee10e59d8b48b0e69ed2af2771da8cffbef7d1ef45698dcfcc478c5487.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 59, + 335, + 104, + 380 + ], + "blocks": [ + { + "bbox": [ + 59, + 335, + 104, + 380 + ], + "lines": [ + { + "bbox": [ + 59, + 335, + 104, + 380 + ], + "spans": [ + { + "bbox": [ + 59, + 335, + 104, + 380 + ], + "type": "image", + "image_path": "63d68dd668dfee57726428ac7abc43378a731407eaa586b283c9c68858466568.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 381, + 102, + 388 + ], + "lines": [ + { + "bbox": [ + 59, + 381, + 102, + 388 + ], + "spans": [ + { + "bbox": [ + 59, + 381, + 102, + 388 + ], + "type": "text", + "content": "Under Exposure" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 104, + 335, + 150, + 380 + ], + "blocks": [ + { + "bbox": [ + 104, + 335, + 150, + 380 + ], + "lines": [ + { + "bbox": [ + 104, + 335, + 150, + 380 + ], + "spans": [ + { + "bbox": [ + 104, + 335, + 150, + 380 + ], + "type": "image", + "image_path": "f63f14dc9e6b83e9e95f9a46eb4855ca19e8a03694217921e2bf0cb8e2520d9c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 381, + 146, + 388 + ], + "lines": [ + { + "bbox": [ + 108, + 381, + 146, + 388 + ], + "spans": [ + { + "bbox": [ + 108, + 381, + 146, + 388 + ], + "type": "text", + "content": "Over Exposure" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 141, + 224, + 302, + 332 + ], + "blocks": [ + { + "bbox": [ + 141, + 224, + 302, + 332 + ], + "lines": [ + { + "bbox": [ + 141, + 224, + 302, + 332 + ], + "spans": [ + { + "bbox": [ + 141, + 224, + 302, + 332 + ], + "type": "image", + "image_path": "86651fcc803b59ecfa03880f24637a9bb3261e0666a6e9ac4f759db0266b63af.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 164, + 335, + 208, + 380 + ], + "blocks": [ + { + "bbox": [ + 164, + 335, + 208, + 380 + ], + "lines": [ + { + "bbox": [ + 164, + 335, + 208, + 380 + ], + "spans": [ + { + "bbox": [ + 164, + 335, + 208, + 380 + ], + "type": "image", + "image_path": "fcb181836efee2444eee333e4a85eea9eb756749bd591ab59362ec48cb6d60e9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 381, + 207, + 395 + ], + "lines": [ + { + "bbox": [ + 167, + 381, + 207, + 395 + ], + "spans": [ + { + "bbox": [ + 167, + 381, + 207, + 395 + ], + "type": "text", + "content": "Previous HDR Reconstruction" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 210, + 335, + 255, + 380 + ], + "blocks": [ + { + "bbox": [ + 210, + 335, + 255, + 380 + ], + "lines": [ + { + "bbox": [ + 210, + 335, + 255, + 380 + ], + "spans": [ + { + "bbox": [ + 210, + 335, + 255, + 380 + ], + "type": "image", + "image_path": "7e496caaf6c6680156378f487b44035a95744e62423e91d5cf000f060bf49374.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 381, + 253, + 396 + ], + "lines": [ + { + "bbox": [ + 211, + 381, + 253, + 396 + ], + "spans": [ + { + "bbox": [ + 211, + 381, + 253, + 396 + ], + "type": "text", + "content": "Exposure Multi-" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 406, + 555, + 450 + ], + "lines": [ + { + "bbox": [ + 54, + 406, + 555, + 450 + ], + "spans": [ + { + "bbox": [ + 54, + 406, + 555, + 450 + ], + "type": "text", + "content": "Figure 1. Comparing our guided inpainting HDR imaging method with a state-of-the-art HDR reconstruction [25] and multi-exposure fusion [48] methods. Both scenes are selected from our captured new real-world benchmark. Left: night cityscape with large exposure difference. Right: afternoon street with motion-induced occlusion. Previous methods struggle to handle these scenes. By modeling HDR as an inpainting problem, our method can produce visually appealing results without ghosting artifacts in these challenging scenes." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 256, + 335, + 302, + 380 + ], + "blocks": [ + { + "bbox": [ + 256, + 335, + 302, + 380 + ], + "lines": [ + { + "bbox": [ + 256, + 335, + 302, + 380 + ], + "spans": [ + { + "bbox": [ + 256, + 335, + 302, + 380 + ], + "type": "image", + "image_path": "b835c1fbb72c1f1486a1043c9d220255d1fe832bfb91412c607a4b18e587846a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 265, + 381, + 295, + 396 + ], + "lines": [ + { + "bbox": [ + 265, + 381, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 265, + 381, + 295, + 396 + ], + "type": "text", + "content": "Our Guided Inpainting" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 308, + 224, + 386, + 277 + ], + "blocks": [ + { + "bbox": [ + 308, + 224, + 386, + 277 + ], + "lines": [ + { + "bbox": [ + 308, + 224, + 386, + 277 + ], + "spans": [ + { + "bbox": [ + 308, + 224, + 386, + 277 + ], + "type": "image", + "image_path": "c831a0057bd40a4adbbca202cdafe7fb9330090748ae062db1032b3a29990e3a.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 309, + 280, + 387, + 332 + ], + "blocks": [ + { + "bbox": [ + 309, + 280, + 387, + 332 + ], + "lines": [ + { + "bbox": [ + 309, + 280, + 387, + 332 + ], + "spans": [ + { + "bbox": [ + 309, + 280, + 387, + 332 + ], + "type": "image", + "image_path": "3c455bfe6e36334b0a72ff3d20b997495663f57fc6e97ca232bb72dffa9b5076.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 308, + 335, + 353, + 380 + ], + "blocks": [ + { + "bbox": [ + 308, + 335, + 353, + 380 + ], + "lines": [ + { + "bbox": [ + 308, + 335, + 353, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 335, + 353, + 380 + ], + "type": "image", + "image_path": "3ff29fa01c3875d6d0c40d2def81b6b154728a7cce43e84c94b453d690d37b9e.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 381, + 351, + 388 + ], + "lines": [ + { + "bbox": [ + 309, + 381, + 351, + 388 + ], + "spans": [ + { + "bbox": [ + 309, + 381, + 351, + 388 + ], + "type": "text", + "content": "Under Exposure" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 355, + 335, + 399, + 380 + ], + "blocks": [ + { + "bbox": [ + 355, + 335, + 399, + 380 + ], + "lines": [ + { + "bbox": [ + 355, + 335, + 399, + 380 + ], + "spans": [ + { + "bbox": [ + 355, + 335, + 399, + 380 + ], + "type": "image", + "image_path": "2f02524b899cfa3a741c8f5d984376a57a8bb172d9d4fdde124d26e59fc916ec.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 357, + 381, + 395, + 388 + ], + "lines": [ + { + "bbox": [ + 357, + 381, + 395, + 388 + ], + "spans": [ + { + "bbox": [ + 357, + 381, + 395, + 388 + ], + "type": "text", + "content": "Over Exposure" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 391, + 224, + 552, + 332 + ], + "blocks": [ + { + "bbox": [ + 391, + 224, + 552, + 332 + ], + "lines": [ + { + "bbox": [ + 391, + 224, + 552, + 332 + ], + "spans": [ + { + "bbox": [ + 391, + 224, + 552, + 332 + ], + "type": "image", + "image_path": "d0c3555145f968fd5a365347564bd2000c2ac22711b6e14cd5290fdf41cf82ca.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 416, + 335, + 460, + 380 + ], + "blocks": [ + { + "bbox": [ + 416, + 335, + 460, + 380 + ], + "lines": [ + { + "bbox": [ + 416, + 335, + 460, + 380 + ], + "spans": [ + { + "bbox": [ + 416, + 335, + 460, + 380 + ], + "type": "image", + "image_path": "a501619ead788a1050d16faa5643ac3b2bcd7522b4c0abecb555b05be4ad8f1f.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 381, + 458, + 395 + ], + "lines": [ + { + "bbox": [ + 417, + 381, + 458, + 395 + ], + "spans": [ + { + "bbox": [ + 417, + 381, + 458, + 395 + ], + "type": "text", + "content": "Previous HDR Reconstruction" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 462, + 335, + 506, + 380 + ], + "blocks": [ + { + "bbox": [ + 462, + 335, + 506, + 380 + ], + "lines": [ + { + "bbox": [ + 462, + 335, + 506, + 380 + ], + "spans": [ + { + "bbox": [ + 462, + 335, + 506, + 380 + ], + "type": "image", + "image_path": "6d8560561528f07f0acafd374af04b22d9d42f3f60c94fed3d091372e6291349.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 462, + 381, + 506, + 395 + ], + "lines": [ + { + "bbox": [ + 462, + 381, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 462, + 381, + 506, + 395 + ], + "type": "text", + "content": "Previous MultiExposure Fusion" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 507, + 335, + 552, + 380 + ], + "blocks": [ + { + "bbox": [ + 507, + 335, + 552, + 380 + ], + "lines": [ + { + "bbox": [ + 507, + 335, + 552, + 380 + ], + "spans": [ + { + "bbox": [ + 507, + 335, + 552, + 380 + ], + "type": "image", + "image_path": "2e320d4ccb5ca97842a1e91ba555728ec8993b80e833b63cbedd42c8f1e7927c.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 515, + 381, + 544, + 396 + ], + "lines": [ + { + "bbox": [ + 515, + 381, + 544, + 396 + ], + "spans": [ + { + "bbox": [ + 515, + 381, + 544, + 396 + ], + "type": "text", + "content": "Our Guided Inpainting" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "bbox": [ + 151, + 460, + 200, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 460, + 200, + 472 + ], + "spans": [ + { + "bbox": [ + 151, + 460, + 200, + 472 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 54, + 481, + 297, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 481, + 297, + 684 + ], + "spans": [ + { + "bbox": [ + 54, + 481, + 297, + 684 + ], + "type": "text", + "content": "Capturing high dynamic range (HDR) scenes is one of the most important issues in camera design. Majority of cameras use exposure fusion, which fuses images captured by different exposure levels, to increase dynamic range. However, this approach can only handle images with limited exposure difference, normally 3-4 stops. When applying to very high dynamic range scenes where a large exposure difference is required, this approach often fails due to incorrect alignment or inconsistent lighting between inputs, or tone mapping artifacts. In this work, we propose UltraFusion, the first exposure fusion technique that can merge inputs with 9 stops differences. The key idea is that we model exposure fusion as a guided inpainting problem, where the under-exposed image is used as a guidance to fill the missing information of over-exposed highlights in the overexposed region. Using an under-exposed image as a soft guidance, instead of a hard constraint, our model is robust" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 461, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 555, + 605 + ], + "type": "text", + "content": "to potential alignment issue or lighting variations. Moreover, by utilizing the image prior of the generative model, our model also generates natural tone mapping, even for very high-dynamic range scenes. Our approach outperforms HDR-Transformer on latest HDR benchmarks. Moreover, to test its performance in ultra high dynamic range scenes, we capture a new real-world exposure fusion benchmark, UltraFusion dataset, with exposure differences up to 9 stops, and experiments show that UltraFusion can generate beautiful and high-quality fusion results under various scenarios. Code and data will be available at https://openimaginglab.github.io/UltraFusion." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 314, + 620, + 395, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 620, + 395, + 632 + ], + "spans": [ + { + "bbox": [ + 314, + 620, + 395, + 632 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "content": "High dynamic range (HDR) imaging is one of the fundamental problems in the modern camera design. Due to hardware limitation, camera sensors have a much smaller dynamic range compared with the real-world. To increase it, majority of HDR solutions merge multiple images, either with the same [9] or different exposure lev" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "*Equal contribution. This work was done during Zixuan Chen's internship at Shanghai Artificial Intelligence Laboratory." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16111" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 294, + 176 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 176 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 176 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 176 + ], + "type": "image", + "image_path": "3a4b4e41840596b085cca383f3ac0e28058089416fec9396c9695a8e73b7af40.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 182, + 295, + 237 + ], + "lines": [ + { + "bbox": [ + 55, + 182, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 295, + 237 + ], + "type": "text", + "content": "Figure 2. Visual comparison on directly utilizing ControlNet [65] and our UltraFusion. Without pre-aligned data, ControlNet struggles to fix a frame for reference. However, our method fixes the over-exposed image as the reference and the under-exposed image as the guidance for inpainting, thereby avoiding artifacts." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 245, + 295, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 245, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 54, + 245, + 295, + 353 + ], + "type": "text", + "content": "els [3, 15, 17, 24, 34, 40, 43, 49, 56, 57]. Despite all recent advances, majority HDR algorithms can only bring limited increase in dynamic range. For example, " + }, + { + "bbox": [ + 54, + 245, + 295, + 353 + ], + "type": "inline_equation", + "content": "\\mathrm{HDR + }" + }, + { + "bbox": [ + 54, + 245, + 295, + 353 + ], + "type": "text", + "content": " [9], the first HDR algorithm used by commercial cameras, can only robustly increase the dynamic range by 8 times (3 stops). Therefore, in this work, we study the following question: can we drastically increase the dynamic range of a camera by fusing two images with very large exposure difference, like 9 stops in Fig. 1?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 354, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 295, + 581 + ], + "type": "text", + "content": "This is a fundamental challenging problem, due to the following three issues. First, to handle dynamic scenes, most of HDR fusion algorithms will first align input frames, which is very challenging when input has large brightness difference. As result, a ghosting issue happens when alignment fails, indicated by zoom-in patches in the right scene of Fig. 1. Second, most of HDR algorithms assume that under-exposed image is simply a darker version of the normal image. However, the appearance of an object may change when exposure levels change, like ship in the left scene of Fig. 1, resulting unnatural fusion result. Third, sometimes the result of fusion is an HDR image, which cannot be directly shown on normal low-dynamic-range display. Therefore, these HDR images will be further compressed through a tone-mapping process. When dynamic range is high, tone-mapping may introduce additional issue. Maintaining a natural contrast and rich details in the final output is challenging, as shown in the zoom-in patches of previous HDR reconstruction method in Fig. 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "In this work, we propose a completely different fusion method, UltraFusion, which models it as a guided inpainting problem. In this setup, the user captures two images, one normal exposed image where brighter objects are overexposed, and another under exposed image, which only captures the very brighter parts of the scene. We use the normal exposed image as a reference, and inpainting the missing information in the highlight. Unlike the traditional inpainting, we use the information from under-exposed frame as guidance, so inpainted highlight is not completely generated, but stays consistent with under-exposed frame." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "content": "There are three advantages of this approach when handling large exposure difference. First, it follows the exposure fusion [30] setup. Unlike HDR fusion techniques, which first generate HDR image and then compress to a low-dynamic-range (LDR) image through tone mapping, exposure fusion directly generates LDR output, avoiding cascading errors. That being said, our method outputs a tone-mapped LDR image instead of a linear HDR. Second, the under-exposed image is used as a soft guidance, instead of a hard constraint. Therefore, UltraFusion is robust to alignment error (see the right scene of Fig. 1) and lighting variation (see the left scene of Fig. 1). Third, image prior of the generative model ensures the natural look of the output image, reducing the potential artifacts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 242, + 553, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 553, + 386 + ], + "type": "text", + "content": "To train a guided inpainting, a simple solution is to train a ControlNet [65] using two input images. However, such paradigm cannot handle dynamic HDR scenes, as ControlNet may not know which frame to choose as the reference, increasing the difficulty to fuse the information from over and short-exposed images. For instance, as shown in Fig. 2, ControlNet selects the over-exposed image as the reference frame in red boxed region, but chooses the under-exposed image as the reference frame in green boxed region, leading to substantial artifacts in the result. Additionally, as a generative model, it may inevitably generate fake image content, as indicated by green box in Fig. 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 388, + 553, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 553, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 553, + 615 + ], + "type": "text", + "content": "To address these challenges, we design UltraFusion as follows. First, we warp the short-exposed image to the long-exposed one, and mask out the occluded regions. Then, we utilize the diffusion prior to inpaint the long-exposed image guided by partial short-exposed information. To make the network retain more details to generate guidance information, we propose a new decompose-and-fuse control branch, which eliminates the luminance component of the short-exposed image, extracting structure and color information instead, and employs a multi-scale cross-attention to improve the feature fusion with long-exposed image. Second, as there is no existing large-scale training data for exposure fusion of dynamic scene, we also propose a novel training data synthesis pipeline, utilizing existing high-quality multi-exposure datasets (pre-aligned) of static scene and video datasets. At last, to ensure the generated output maintains the fidelity to the reality, we also train an additional fidelity control branch, using the same decompose-and-fuse strategy and multi-scale cross-attention." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": "At last, to evaluate the effectiveness of our framework, we capture 100 under/over-exposed image pairs, covering daytime, nighttime, indoor, outdoor scenes with local and global motion patterns. Experiments on both latest HDR imaging datasets and our captured benchmark demonstrate that, comparing to existing methods, our UltraFusion is more robust to scenes with large exposure differences and large motion, as shown in Fig. 1." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16112" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 139, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 139, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 139, + 83 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 144, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 144, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 144, + 105 + ], + "type": "text", + "content": "2.1. HDR imaging" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 111, + 295, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 111, + 295, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 111, + 295, + 339 + ], + "type": "text", + "content": "HDR Imaging can be deided into HDR reconstruction and Multi-Exposure Fusion (MEF) typically depends on the domain where fusion occurs [19]. HDR reconstruction methods [3, 15, 17, 24, 34, 40, 43, 49, 51, 56, 57] inverts the camera response function (CRF) to merge exposure brackets in the linear HDR domain [28]. In most cases, tone mapping is necessary to display the reconstructed HDR image properly on standard LDR monitors. As a cost-effective alternative [13], MEF methods [13, 19, 22, 28, 30, 33, 35, 48, 52, 68, 69] directly fuse images in the LDR domain, sidestepping CRF calibration and sophisticated tone mapping process [30]. Regardless of the type of HDR Imaging methods, they contend with ghosting artifacts caused by camera shake and object movement [39]. Previous methods have attempted explicit or implicit alignment using optical flow [15, 33, 34, 49] or attention mechanism [3, 17, 24, 40, 43, 56, 57]. However, most HDR Imaging methods suffer from unpleasing artifacts when large motion causes occlusion in the complementary region." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 341, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 295, + 604 + ], + "type": "text", + "content": "Diffusion models. Recently, it has been witnessed with the rapid rise of diffusion models [10, 37] and their successful application in various tasks, including controllable image generation [59, 65], image restoration [21, 23, 31, 47, 63], image editing [4, 14, 29, 38] and image inpainting [5, 26, 50, 64]. In the field of HDR imaging, the application of diffusion models has primarily focused on HDR deghosting [8, 12, 58]. However, since these methods do not leverage the diffusion priors learned from large-scale datasets, their generalization ability is limited by the scale of the HDR dataset. While some recent works [6, 18] have employed diffusion priors, they tend to focus on single image HDR. Without another differently-exposed image as reference, the results generated by these methods lack sufficient reliability. Unlike previous methods, we utilize diffusion priors and use the short-exposed image as a reference to perform reliable inpainting in the highlight regions of the overexposed image, thereby achieving natural and reliable HDR scene reconstruction. Compared to diffusion-based inpainting methods [5, 26, 50, 64] that perform inpainting solely from scratch, we leverage information from short-exposure images to guide a more accurate inpainting process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "Tone mapping methods. The goal of tone mapping is to convert HDR images to LDR for display on standard screens while enhancing visual detail. Due to the challenge of obtaining ground-truth tone mapping results, unsupervised deep learning approaches have been developed using adversarial [45] and contrastive [2] learning. To address data limitations, Cai et al. [1] manually curated training data by selecting the best results from 13 tone mapping methods for network learning [1, 11]. However, previous tone" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "mapping algorithms, lacking robust image priors and facing data constraints, struggle with visually pleasing results and generalization in extreme high dynamic range scenes. By incorporating diffusion-based image priors, our method achieves aesthetic results even in challenging high dynamic range scenarios (see Fig. 1)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 154, + 397, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 154, + 397, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 397, + 167 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 173, + 553, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 173, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 173, + 553, + 281 + ], + "type": "text", + "content": "Given an over-exposed image " + }, + { + "bbox": [ + 313, + 173, + 553, + 281 + ], + "type": "inline_equation", + "content": "I_{oe}" + }, + { + "bbox": [ + 313, + 173, + 553, + 281 + ], + "type": "text", + "content": " and an under-exposed image " + }, + { + "bbox": [ + 313, + 173, + 553, + 281 + ], + "type": "inline_equation", + "content": "I_{ue}" + }, + { + "bbox": [ + 313, + 173, + 553, + 281 + ], + "type": "text", + "content": ", traditional exposure fusion algorithms directly aggregate different frequency band of both images, which are sensitive to misalignment error or lighting variation. Instead, we treat this as an inpainting problem. Specifically, we use the over-exposed image " + }, + { + "bbox": [ + 313, + 173, + 553, + 281 + ], + "type": "inline_equation", + "content": "I_{oe}" + }, + { + "bbox": [ + 313, + 173, + 553, + 281 + ], + "type": "text", + "content": " as the base image and inpainting missing information in the highlight region. To ensure inpainted highlights are real, we also use highlights from under-exposed image as guidance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 281, + 554, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 554, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 554, + 353 + ], + "type": "text", + "content": "Based on this idea, we design a 2-stage newtwork shown in Fig. 3, which consists of the pre-alignment stage and the guided inpainting stage. The pre-alignment stage outputs a coarse-aligned version of " + }, + { + "bbox": [ + 313, + 281, + 554, + 353 + ], + "type": "inline_equation", + "content": "I_{ue}" + }, + { + "bbox": [ + 313, + 281, + 554, + 353 + ], + "type": "text", + "content": ", which is used as the soft guidance in the following guided inpainting stage. Details of each stage is described below." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 359, + 430, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 430, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 430, + 372 + ], + "type": "text", + "content": "3.1. Pre-alignment stage" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 376, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 554, + 533 + ], + "type": "text", + "content": "Most of optical flow alignment assumes input have the similar brightness. Therefore, we first adjust the brightness " + }, + { + "bbox": [ + 313, + 376, + 554, + 533 + ], + "type": "inline_equation", + "content": "I_{ue}" + }, + { + "bbox": [ + 313, + 376, + 554, + 533 + ], + "type": "text", + "content": " to match the distribution of " + }, + { + "bbox": [ + 313, + 376, + 554, + 533 + ], + "type": "inline_equation", + "content": "I_{oe}" + }, + { + "bbox": [ + 313, + 376, + 554, + 533 + ], + "type": "text", + "content": " through intensity mapping function [7]. Then, we adopt RAFT [42], a pre-trained optical flow network, to estimate the bidirectional flow " + }, + { + "bbox": [ + 313, + 376, + 554, + 533 + ], + "type": "inline_equation", + "content": "f_{oe \\rightarrow ue}" + }, + { + "bbox": [ + 313, + 376, + 554, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 376, + 554, + 533 + ], + "type": "inline_equation", + "content": "f_{ue \\rightarrow oe}" + }, + { + "bbox": [ + 313, + 376, + 554, + 533 + ], + "type": "text", + "content": " and align " + }, + { + "bbox": [ + 313, + 376, + 554, + 533 + ], + "type": "inline_equation", + "content": "I_{ue}" + }, + { + "bbox": [ + 313, + 376, + 554, + 533 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 376, + 554, + 533 + ], + "type": "inline_equation", + "content": "I_{oe}" + }, + { + "bbox": [ + 313, + 376, + 554, + 533 + ], + "type": "text", + "content": " using backward warping. However, backward warping will result in ghosting at the occlusion boundary [67], leading to artifacts in the next guided inpainting stage. To solve that, we utilize the forward-and-backward consistency check [54] to estimate occluded regions " + }, + { + "bbox": [ + 313, + 376, + 554, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 376, + 554, + 533 + ], + "type": "text", + "content": " and mask them out in the warped output. Finally, we can obtain a pre-aligned output " + }, + { + "bbox": [ + 313, + 376, + 554, + 533 + ], + "type": "inline_equation", + "content": "I_{ue \\rightarrow oe}" + }, + { + "bbox": [ + 313, + 376, + 554, + 533 + ], + "type": "text", + "content": " of the first stage:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 354, + 539, + 553, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 539, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 354, + 539, + 553, + 553 + ], + "type": "interline_equation", + "content": "I _ {u e \\rightarrow o e} = (1 - \\mathcal {M}) \\cdot \\mathcal {W} \\left(I _ {u e}, f _ {o e \\rightarrow u e}\\right), \\tag {1}", + "image_path": "fd651dca1ac660c9d727e9630c57fccaf6dd6afba3b0c260f482508c6363a49a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 558, + 553, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 582 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 558, + 553, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 313, + 558, + 553, + 582 + ], + "type": "text", + "content": " denotes the backward warping. Fig. 3 (a) shows the output is a masked and aligned under-exposed image." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 588, + 449, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 449, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 449, + 601 + ], + "type": "text", + "content": "3.2. Guided inpainting stage" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "We build our guided inpainting model based on the Stable Diffusion Model [36] because its powerful generative prior can help to resolve ambiguity during inpainting. Similar to other diffusion-based image enhancement techniques [21], we also inject the following information through an additional control branch, as shown in Fig. 3(b): 1) the image to be inpainted, which is the over-exposed image " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "I_{oe}" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": ", 2) the additional guidance of highlight, which is the underexposed image " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "I_{ue}" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": ", and 3) the diffusion latent at the current" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16113" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 92, + 301, + 222 + ], + "blocks": [ + { + "bbox": [ + 133, + 73, + 235, + 85 + ], + "lines": [ + { + "bbox": [ + 133, + 73, + 235, + 85 + ], + "spans": [ + { + "bbox": [ + 133, + 73, + 235, + 85 + ], + "type": "text", + "content": "(a) Pre-Alignment Stage" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 92, + 301, + 222 + ], + "lines": [ + { + "bbox": [ + 60, + 92, + 301, + 222 + ], + "spans": [ + { + "bbox": [ + 60, + 92, + 301, + 222 + ], + "type": "image", + "image_path": "8fadd714b2953a61ee44d5c446f878dc6e4cc7cac1082b3d009fbf9faa54d7de.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 228, + 555, + 273 + ], + "lines": [ + { + "bbox": [ + 55, + 228, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 555, + 273 + ], + "type": "text", + "content": "Figure 3. The whole backbone of UltraFusion. Our method is a 2-stage framework, consisting of (a) pre-alignment stage and (b) guided inpainting stage. The first stage pre-aligns the under-exposed image " + }, + { + "bbox": [ + 55, + 228, + 555, + 273 + ], + "type": "inline_equation", + "content": "I_{ue}" + }, + { + "bbox": [ + 55, + 228, + 555, + 273 + ], + "type": "text", + "content": " to the over-exposed image " + }, + { + "bbox": [ + 55, + 228, + 555, + 273 + ], + "type": "inline_equation", + "content": "I_{oe}" + }, + { + "bbox": [ + 55, + 228, + 555, + 273 + ], + "type": "text", + "content": " and masks the occluded regions. In the subsequent guided inpainting stage, we propose a new decompose-and-fuse control branch to utilize the diffusion priors, together with a fidelity control branch for fidelity keeping." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 309, + 83, + 553, + 222 + ], + "blocks": [ + { + "bbox": [ + 365, + 73, + 482, + 85 + ], + "lines": [ + { + "bbox": [ + 365, + 73, + 482, + 85 + ], + "spans": [ + { + "bbox": [ + 365, + 73, + 482, + 85 + ], + "type": "text", + "content": "(b) Guided Inpainting Stage" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 83, + 553, + 222 + ], + "lines": [ + { + "bbox": [ + 309, + 83, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 309, + 83, + 553, + 222 + ], + "type": "image", + "image_path": "11a54cfe6cf82e099f81cf5ed02503946f06f32d6bb8edb3f9ca63e8b19d3f0e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 297, + 293, + 437 + ], + "blocks": [ + { + "bbox": [ + 89, + 282, + 263, + 295 + ], + "lines": [ + { + "bbox": [ + 89, + 282, + 263, + 295 + ], + "spans": [ + { + "bbox": [ + 89, + 282, + 263, + 295 + ], + "type": "text", + "content": "Decompose-and-Fuse Control Branch" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 297, + 293, + 437 + ], + "lines": [ + { + "bbox": [ + 58, + 297, + 293, + 437 + ], + "spans": [ + { + "bbox": [ + 58, + 297, + 293, + 437 + ], + "type": "image", + "image_path": "ac6aa4d034a334b9edfb1f03808169d4ce337e20f0e5bbde26014db38a946ca2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 447, + 295, + 468 + ], + "lines": [ + { + "bbox": [ + 55, + 447, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 295, + 468 + ], + "type": "text", + "content": "Figure 4. The detailed architecture of our proposed decompose-and-fuse control branch." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 472, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 555 + ], + "type": "text", + "content": "diffusion step " + }, + { + "bbox": [ + 55, + 472, + 296, + 555 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 55, + 472, + 296, + 555 + ], + "type": "text", + "content": ", as previous work [21] shows that including diffusion latent as an additional condition can improve image quality. The main diffusion denoising network is a pretrained U-Net. The over-exposed image is first encoded using a pretrained VAE before entering the diffusion module, and the outputs are converted back to the image space using the pretrained decoder." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 557, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 676 + ], + "type": "text", + "content": "The main differences between our solution and general diffusion-based image enhancement are two-fold. First, we propose a novel decompose-and-fuse control branch to inject two input images and diffusion latent as a control signal, as Fig. 11 (b) shows that naively injecting this information may not be able to guide diffusion to faithfully inpaint missing highlights obtained from the under-exposed image. Second, we train an additional fidelity control branch to provide faithful structure and color information for the decoding process via shortcuts. Details are described below." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Decompose-and-fuse control branch. Fig. 4 shows our control branch. We use the over-exposed image " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "I_{oe}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " as the main control signal and the under-exposed image " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "I_{ue}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " as the" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 278, + 555, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 555, + 374 + ], + "type": "text", + "content": "soft guidance. Following ControlNet [65], we copy the encoder and middle blocks of the denoising U-Net as the main extractor, but update their weights during training. A simple soft guidance is to use the encoded under-exposed image latent " + }, + { + "bbox": [ + 313, + 278, + 555, + 374 + ], + "type": "inline_equation", + "content": "y_{ue}" + }, + { + "bbox": [ + 313, + 278, + 555, + 374 + ], + "type": "text", + "content": " from the VAE encoder, combined with the overexposed latent " + }, + { + "bbox": [ + 313, + 278, + 555, + 374 + ], + "type": "inline_equation", + "content": "y_{oe}" + }, + { + "bbox": [ + 313, + 278, + 555, + 374 + ], + "type": "text", + "content": ". However, the underexposed images are often too dark to be directly used as soft guidance, since the model may entirely ignore this guidance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 375, + 555, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 555, + 422 + ], + "type": "text", + "content": "Therefore, we decompose the under-exposed image into the color and structure information, both of which are robust to brightness changes. Specifically, we use the normalized image as the structure component, similar to SSIM [46], as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 368, + 434, + 553, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 434, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 368, + 434, + 553, + 447 + ], + "type": "interline_equation", + "content": "S _ {u e} = \\left(Y _ {u e} - \\mu \\left(Y _ {u e}\\right)\\right) / \\sigma \\left(Y _ {u e}\\right), \\tag {2}", + "image_path": "4f2a552a997751ca3b74fd42fc976a5d63c9b0bfd8e3672ecab78af88891b541.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 450, + 554, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 554, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 450, + 554, + 545 + ], + "type": "inline_equation", + "content": "Y_{ue}" + }, + { + "bbox": [ + 313, + 450, + 554, + 545 + ], + "type": "text", + "content": " represents the luminance channel of " + }, + { + "bbox": [ + 313, + 450, + 554, + 545 + ], + "type": "inline_equation", + "content": "I_{ue}" + }, + { + "bbox": [ + 313, + 450, + 554, + 545 + ], + "type": "text", + "content": " under YUV space, " + }, + { + "bbox": [ + 313, + 450, + 554, + 545 + ], + "type": "inline_equation", + "content": "\\mu(\\cdot)" + }, + { + "bbox": [ + 313, + 450, + 554, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 450, + 554, + 545 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 313, + 450, + 554, + 545 + ], + "type": "text", + "content": " denote the mean intensity and standard deviation, respectively. The chroma channels (UV) are used as color information. The extracted structure and color information are further encoded using trained color and structure extractors (gray GE block in Fig. 4). Following [65], we implement GE using several simple convolution layers to extract multi-scale features." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": "At last, the extracted features are injected into the main extractor with a multi-scale cross attention, as shown in the bottom part of Fig. 4. The output of each cross attention module is fed into both the next level of the main extractor and the corresponding U-Net block, using zero convolution. Fidelity control branch. Even with the control block, we sometimes observe undesirable modifications of texture introduced by VAE. An example is shown in Fig. 11 (c). Therefore, to further improve fidelity, we design a Fidelity Control Branch (FCB), inspired by [47]. FCB mitigates texture distortions by injecting features into the VAE decoder. It has a similar architecture as the decompose-and-fuse control branch, with two main differences: 1) the main extractor of FCB adopts the same structure as the VAE Encoder," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16114" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 298, + 172 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 298, + 172 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 298, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 298, + 172 + ], + "type": "image", + "image_path": "e7c2a68532deed2df235cb26f482d510940848446e1c592acbe6f8aa5ccf6695.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 173, + 292, + 184 + ], + "lines": [ + { + "bbox": [ + 58, + 173, + 292, + 184 + ], + "spans": [ + { + "bbox": [ + 58, + 173, + 292, + 184 + ], + "type": "text", + "content": "Figure 5. The illustration of our training data synthesis pipeline." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 188, + 296, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 188, + 296, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 296, + 344 + ], + "type": "text", + "content": "rather than the denoising U-Net, to provide corresponding shortcuts to the VAE Decoder, with adjustments to the soft guidance extractor as well, and 2) the main extractor of the FCB directly takes the over-exposed image as input. To train the fidelity control branch, we freeze the pre-trained VAE Encoder and Decoder and encode the ground truth " + }, + { + "bbox": [ + 55, + 188, + 296, + 344 + ], + "type": "inline_equation", + "content": "I_{gt}" + }, + { + "bbox": [ + 55, + 188, + 296, + 344 + ], + "type": "text", + "content": " to latent space, simulating the denoised latent " + }, + { + "bbox": [ + 55, + 188, + 296, + 344 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 55, + 188, + 296, + 344 + ], + "type": "text", + "content": " during the inference. Then we input the corresponding overexposed image and under-exposed image to the fidelity control branch to extract faithful features. At last, the VAE Decoder decodes the compressed latent to a reconstructed image " + }, + { + "bbox": [ + 55, + 188, + 296, + 344 + ], + "type": "inline_equation", + "content": "\\hat{I}_{gt}" + }, + { + "bbox": [ + 55, + 188, + 296, + 344 + ], + "type": "text", + "content": ". We adopt the reconstruction loss " + }, + { + "bbox": [ + 55, + 188, + 296, + 344 + ], + "type": "inline_equation", + "content": "||\\hat{I}_{gt} - I_{gt}||_1" + }, + { + "bbox": [ + 55, + 188, + 296, + 344 + ], + "type": "text", + "content": " as an additional loss term." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 350, + 187, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 187, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 187, + 363 + ], + "type": "text", + "content": "3.3. Training data synthesis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 368, + 296, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 368, + 296, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 368, + 296, + 439 + ], + "type": "text", + "content": "Preparing the training data for the proposed guided inpainting network is challenging. To train the model, ideally we need a large-scale HDR dataset that 1) covers different dynamic scenes, 2) has large exposure variance up to 9 stops, and 3) contains ground truth fusion result. However, no existing dataset satisfy all requirements." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 440, + 296, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 440, + 296, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 440, + 296, + 620 + ], + "type": "text", + "content": "To solve this issue, we propose a novel training data synthesis pipeline. Specifically, as shown in Fig. 5, we first randomly sample an image sequence with " + }, + { + "bbox": [ + 55, + 440, + 296, + 620 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 440, + 296, + 620 + ], + "type": "text", + "content": " frames from a video dataset [55]. To model large motion, we select the first and the last frames. Then, similar to the pre-alignment stage, we estimate the bidirectional optical flow between two selected frames using [42] and compute a pseudo occlusion mask via forward-backward consistency check. Subsequently, we randomly sample an under-exposed image patch from the high-quality static dataset [1] (ground truths are pre-aligned), resize the pseudo occlusion mask to match the patch size, and mask out the pseudo occluded region to synthesize a pseudo pre-aligned output. With our synthesized training data, our UltraFusion learns to handle dynamic scenes with only static multi-exposure image pairs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 628, + 132, + 642 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 132, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 132, + 642 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 648, + 176, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 176, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 176, + 661 + ], + "type": "text", + "content": "4.1. Experimental setting" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Datasets. We utilize SICE [1] dataset and Video-90K [55] dataset to synthesize our training data. Following previous works [1, 22], we select images with the highest and lowest brightness in each exposure bracket of the SICE dataset" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 70, + 553, + 137 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 553, + 137 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 553, + 137 + ], + "type": "image", + "image_path": "68b45325848c2b6028dde807d2de1ff012aa2f924f636e9d09b76224060d7c5c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 140, + 555, + 174 + ], + "lines": [ + { + "bbox": [ + 313, + 140, + 555, + 174 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 555, + 174 + ], + "type": "text", + "content": "Figure 6. (a) The data distribution of our benchmark. Coordinate value of exposure time represents the upper boundary. (b) The user study result on our benchmark." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 182, + 553, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 553, + 242 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 553, + 242 + ], + "type": "text", + "content": "as inputs. We evaluate our method on both static datasets and dynamic datasets. For evaluation, we use the MEFB dataset [66] with 100 static under/over-exposed image pairs and RealHDRV [39], a dynamic HDR deghosting test set containing 50 scenes with varying motion patterns." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 243, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 483 + ], + "type": "text", + "content": "UltraFusion benchmark. Existing exposure fusion benchmark cannot fully evaluate fusing in real-world challenging conditions, as they either lack realistic motion [66] or have limited dynamic range [15, 39, 44]. Therefore, we collect a new real-world UltraFusion benchmark, which contains 100 real-captured under/over-exposed image pairs. Compared to previous datasets, our benchmark is more challenging for three reasons: 1) Our benchmark features larger exposure differences between the two input images (up to 9 stops). 2) It includes more realistic motion, with many scenes containing extensive and unintentional foreground movement. 3) Our benchmark is highly diverse, encompassing daytime, nighttime, indoor, and outdoor scenes captured by DSLR Camera (Canon R8) and mobile phones (iPhone12, iPhone13, Redmi K50 Pro and OPPO Reno8 Pro). We summarize the exposure difference distribution and exposure time distribution of our benchmark in Fig. 6 (a). It can be observed that our benchmark covers a wide range of exposure differences and diverse exposure times, which can be used to effectively test the robustness of the HDR methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 484, + 556, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 556, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 556, + 605 + ], + "type": "text", + "content": "Implementation details. We leverage the generative prior encapsulated in Stable Diffusion V2.1 [37]. The decompose-and-fuse control branch (DFCB) is trained for 140k iterations with batch size " + }, + { + "bbox": [ + 313, + 484, + 556, + 605 + ], + "type": "inline_equation", + "content": "bs = 32" + }, + { + "bbox": [ + 313, + 484, + 556, + 605 + ], + "type": "text", + "content": " on 8 NVIDIA RTX 4090 GPUs. We also train the fidelity control branch for 1000K iteration with batch size " + }, + { + "bbox": [ + 313, + 484, + 556, + 605 + ], + "type": "inline_equation", + "content": "bs = 1" + }, + { + "bbox": [ + 313, + 484, + 556, + 605 + ], + "type": "text", + "content": " on a single NVIDIA RTX 4090 GPU. Adam is adopted as the optimizer and the learning rate is fixed to 0.0001. To adapt HDR reconstruction methods to 2 differently exposed inputs, we re-implement them by following their default settings." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": "Evaluation metrics. We utilize four widely used non-reference metric MUSIQ [16], DeQA-Score [62], PAQ2PIQ [61] and HyperIQA [41] for quantitative comparison. Moreover, for static dataset [66], as no ground truths are provided, we select the task-specific MEF-SSIM [27] for structure retention evaluation. For dynamic dataset [39] with HDR ground truths, TMQI [60] is used to evaluate the performance of fidelity and naturalness. In addition, we conduct a user study on our Ultrafusion benchmark to" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16115" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 91, + 403, + 231 + ], + "blocks": [ + { + "bbox": [ + 108, + 71, + 353, + 83 + ], + "lines": [ + { + "bbox": [ + 108, + 71, + 353, + 83 + ], + "spans": [ + { + "bbox": [ + 108, + 71, + 353, + 83 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons on the static MEFB dataset [66]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 91, + 403, + 231 + ], + "lines": [ + { + "bbox": [ + 60, + 91, + 403, + 231 + ], + "spans": [ + { + "bbox": [ + 60, + 91, + 403, + 231 + ], + "type": "table", + "html": "
TypeMethodMEFB [66]
MUSIQ↑DeQA-Score↑PAQ2PIQ↑HyperIQA↑MEF-SSIM↑
HDR Rec.HDR-Transformer [25]63.102.98371.360.59960.8626
SCTNet [44]63.133.02171.480.60680.8777
SAFNet [17]61.702.51872.670.56460.7711
MEFDeepfuse [35]52.582.94567.960.43990.8968
MEF-GAN [53]50.592.81869.990.37450.7722
U2Fusion [52]63.393.21972.230.51590.9304
Defusion [20]62.703.11872.820.54550.9062
MEFLUT [13]65.713.27771.210.52670.8608
HSDS-MEF [48]66.763.54472.600.60260.9520
TC-MoA [69]64.603.35571.850.53940.9636
OursUltraFusion68.823.88173.800.64820.9385
", + "image_path": "153282c10c34b54527e5a1ccf53178e9554ede352ff25c6775ea5ccd054a9eb8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 416, + 72, + 545, + 167 + ], + "blocks": [ + { + "bbox": [ + 416, + 72, + 545, + 167 + ], + "lines": [ + { + "bbox": [ + 416, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 416, + 72, + 545, + 167 + ], + "type": "image", + "image_path": "a060de34805c5ccb86e319fe2ec960ca25d98d5de577432f89cfa9d0c546e0cf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 175, + 554, + 231 + ], + "lines": [ + { + "bbox": [ + 406, + 175, + 554, + 231 + ], + "spans": [ + { + "bbox": [ + 406, + 175, + 554, + 231 + ], + "type": "text", + "content": "Figure 7. Trade-off curve between MEF-SSIM and MUSIQ on MEFB dataset [66]. Our UltraFusion achieves the best trade-off between image quality and information preservation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 251, + 553, + 352 + ], + "blocks": [ + { + "bbox": [ + 94, + 237, + 515, + 248 + ], + "lines": [ + { + "bbox": [ + 94, + 237, + 515, + 248 + ], + "spans": [ + { + "bbox": [ + 94, + 237, + 515, + 248 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons on dynamic RealHDRV dataset [39] and our challenging UltraFusion benchmark." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 251, + 553, + 352 + ], + "lines": [ + { + "bbox": [ + 58, + 251, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 58, + 251, + 553, + 352 + ], + "type": "table", + "html": "
TypeMethodRealHDRV [39]UltraFusion Benchmark
TMQI↑MUSIQ↑DeQA-Score↑PAQ2PIQ↑HyperIQA↑MUSIQ↑DeQA-Score↑PAQ2PIQ↑HyperIQA↑
HDR Rec.HDR-Transformer [25]0.868062.243.49670.330.522563.662.90972.830.5619
SCTNet [44]0.871562.693.53270.740.527261.843.10272.940.5888
SAFNet [17]0.872662.073.50670.480.515661.502.17973.150.5487
MEFDefusion [20]0.818756.603.30268.380.485660.313.35271.870.5463
MEFLUT [13]0.829762.423.31570.040.502063.623.34371.730.5074
HSDS-MEF [48]0.832361.763.36071.110.505464.543.62773.420.5923
OursUltraFusion0.892567.513.83073.400.583368.413.95775.180.6214
", + "image_path": "16f1be99088e9f719c2de384b169b8a827fb51a78931608bb59b421ac0c46522.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 57, + 358, + 139, + 483 + ], + "blocks": [ + { + "bbox": [ + 57, + 358, + 139, + 483 + ], + "lines": [ + { + "bbox": [ + 57, + 358, + 139, + 483 + ], + "spans": [ + { + "bbox": [ + 57, + 358, + 139, + 483 + ], + "type": "image", + "image_path": "93fae921e5e476884387d948973f1976642988310f9507a1a08887028e930a3e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 486, + 110, + 495 + ], + "lines": [ + { + "bbox": [ + 86, + 486, + 110, + 495 + ], + "spans": [ + { + "bbox": [ + 86, + 486, + 110, + 495 + ], + "type": "text", + "content": "Inputs" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 139, + 358, + 222, + 420 + ], + "blocks": [ + { + "bbox": [ + 139, + 358, + 222, + 420 + ], + "lines": [ + { + "bbox": [ + 139, + 358, + 222, + 420 + ], + "spans": [ + { + "bbox": [ + 139, + 358, + 222, + 420 + ], + "type": "image", + "image_path": "f63c111f10620fd0b58b32182118a0267571a37ccdfb6e10254bf33f7d5b2f90.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 140, + 421, + 222, + 483 + ], + "blocks": [ + { + "bbox": [ + 140, + 421, + 222, + 483 + ], + "lines": [ + { + "bbox": [ + 140, + 421, + 222, + 483 + ], + "spans": [ + { + "bbox": [ + 140, + 421, + 222, + 483 + ], + "type": "image", + "image_path": "4b2c30a02693391b912120a96138c5f9c5a574fafb51e686b2611241e193ea96.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 485, + 203, + 495 + ], + "lines": [ + { + "bbox": [ + 159, + 485, + 203, + 495 + ], + "spans": [ + { + "bbox": [ + 159, + 485, + 203, + 495 + ], + "type": "text", + "content": "HSDS-MEF" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 223, + 358, + 305, + 420 + ], + "blocks": [ + { + "bbox": [ + 223, + 358, + 305, + 420 + ], + "lines": [ + { + "bbox": [ + 223, + 358, + 305, + 420 + ], + "spans": [ + { + "bbox": [ + 223, + 358, + 305, + 420 + ], + "type": "image", + "image_path": "4a76d138e85d212c3ca80e0004d77fd085d52211103a356828fcfbbfb4b74b71.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 223, + 421, + 305, + 483 + ], + "blocks": [ + { + "bbox": [ + 223, + 421, + 305, + 483 + ], + "lines": [ + { + "bbox": [ + 223, + 421, + 305, + 483 + ], + "spans": [ + { + "bbox": [ + 223, + 421, + 305, + 483 + ], + "type": "image", + "image_path": "50091b5090b783499c421556760e57a9ce03aadb7eb1f33d0b04aa83ed869fc8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 485, + 280, + 495 + ], + "lines": [ + { + "bbox": [ + 247, + 485, + 280, + 495 + ], + "spans": [ + { + "bbox": [ + 247, + 485, + 280, + 495 + ], + "type": "text", + "content": "TC-MoA" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 305, + 358, + 388, + 420 + ], + "blocks": [ + { + "bbox": [ + 305, + 358, + 388, + 420 + ], + "lines": [ + { + "bbox": [ + 305, + 358, + 388, + 420 + ], + "spans": [ + { + "bbox": [ + 305, + 358, + 388, + 420 + ], + "type": "image", + "image_path": "9b6897253144233b47e28c2683745d54bdc929f8f82eb245058545b65d401c7d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 305, + 421, + 388, + 483 + ], + "blocks": [ + { + "bbox": [ + 305, + 421, + 388, + 483 + ], + "lines": [ + { + "bbox": [ + 305, + 421, + 388, + 483 + ], + "spans": [ + { + "bbox": [ + 305, + 421, + 388, + 483 + ], + "type": "image", + "image_path": "b93b39a851c7892229290213dfe78c449b4845a73dbd35ddf953521c833b168b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 485, + 379, + 495 + ], + "lines": [ + { + "bbox": [ + 314, + 485, + 379, + 495 + ], + "spans": [ + { + "bbox": [ + 314, + 485, + 379, + 495 + ], + "type": "text", + "content": "HDR-Transformer" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 129, + 498, + 479, + 509 + ], + "lines": [ + { + "bbox": [ + 129, + 498, + 479, + 509 + ], + "spans": [ + { + "bbox": [ + 129, + 498, + 479, + 509 + ], + "type": "text", + "content": "Figure 8. Visual comparisons of different exposure fusion methods on static MEFB dataset [66]." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 388, + 358, + 470, + 420 + ], + "blocks": [ + { + "bbox": [ + 388, + 358, + 470, + 420 + ], + "lines": [ + { + "bbox": [ + 388, + 358, + 470, + 420 + ], + "spans": [ + { + "bbox": [ + 388, + 358, + 470, + 420 + ], + "type": "image", + "image_path": "1973d59ef8ebbd782431021e85f5f6b24221bc452bbc94b5a823cd3b3e31eebb.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 388, + 421, + 470, + 483 + ], + "blocks": [ + { + "bbox": [ + 388, + 421, + 470, + 483 + ], + "lines": [ + { + "bbox": [ + 388, + 421, + 470, + 483 + ], + "spans": [ + { + "bbox": [ + 388, + 421, + 470, + 483 + ], + "type": "image", + "image_path": "57d20f7dd140b1ddc5d5f1e2ca505fabc93d0bbb08787715cf6646b1bfb52a41.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 485, + 444, + 495 + ], + "lines": [ + { + "bbox": [ + 414, + 485, + 444, + 495 + ], + "spans": [ + { + "bbox": [ + 414, + 485, + 444, + 495 + ], + "type": "text", + "content": "SCTNet" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 471, + 358, + 553, + 420 + ], + "blocks": [ + { + "bbox": [ + 471, + 358, + 553, + 420 + ], + "lines": [ + { + "bbox": [ + 471, + 358, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 471, + 358, + 553, + 420 + ], + "type": "image", + "image_path": "f0f1c525374f4e41b8fe2a761b284a5c327d3dc005aa45f4debf56473acb5de5.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 471, + 421, + 553, + 483 + ], + "blocks": [ + { + "bbox": [ + 471, + 421, + 553, + 483 + ], + "lines": [ + { + "bbox": [ + 471, + 421, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 471, + 421, + 553, + 483 + ], + "type": "image", + "image_path": "1da5b84a86915fb13d404279e971376aa047b4317358a70ecb05753c10a2259e.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 503, + 486, + 522, + 495 + ], + "lines": [ + { + "bbox": [ + 503, + 486, + 522, + 495 + ], + "spans": [ + { + "bbox": [ + 503, + 486, + 522, + 495 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 515, + 179, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 179, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 179, + 528 + ], + "type": "text", + "content": "perform subjective evaluation." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 538, + 274, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 274, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 274, + 551 + ], + "type": "text", + "content": "4.2. Comparisons with HDR imaging methods" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 556, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 295, + 652 + ], + "type": "text", + "content": "We compare our method with state-of-the-art HDR Imaging methods, including HDR reconstruction methods HDRTransformer [25], SCTNet [44], SAFNet [17], and multi-exposure fusion methods Deepfuse [35], MEF-GAN [53], U2Fusion [52], Defusion [20], MEFLUT [13], HSDS-MEF [48], TC-MoA [69]. As HDR reconstruction methods cannot output an LDR image directly, we use professional software Photomatix [32] to perform tone mapping." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Evaluation on static dataset. We evaluate the fusion performance on the MEFB dataset [66], focusing on large exposure differences. As shown in Tab. 1, our method outperforms other methods on all four non-reference metrics (MUSIQ, DeQA-Score, PAQ2PIQ, and HyperIQA). Specif" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 515, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 555, + 696 + ], + "type": "text", + "content": "ically, our proposed UltraFusion achieves 2.06 gain in terms of MUSIQ compared to HSDS-MEF. In terms of MEFSSIM, as shown in Fig. 7, our baseline model (ControlNet [65]) achieves better image quality but lacks fidelity, while HSDS-MEF [48] retains more information from inputs at the cost of quality. Our UltraFusion outperforms most algorithms, and achieves similar fidelity scores compared to HSDS-MEF and TC-MoA but with much higher image quality (non-reference metrics), indicating the best trade-off between visual quality and information preservation. The qualitative comparison in Fig. 8 further validates this. In contrast, HDR reconstruction methods (HDRTransformer and SCTNet) often miss some detail in highlights and MEF methods (HSDS-MEF and TC-MoA) introduce unnatural transition from brighter to dark regions." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 314, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 314, + 701, + 555, + 714 + ], + "type": "text", + "content": "Evaluation on dynamic dataset. To further illustrate the" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16116" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 304, + 168 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 304, + 168 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 304, + 168 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 304, + 168 + ], + "type": "image", + "image_path": "bc95d2be8f4d67ad38d53524cfc904f00289bfc22f5c7218c3eeb62ee2fcf447.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 172, + 432, + 183 + ], + "lines": [ + { + "bbox": [ + 176, + 172, + 432, + 183 + ], + "spans": [ + { + "bbox": [ + 176, + 172, + 432, + 183 + ], + "type": "text", + "content": "Figure 9. Visual comparisons on our captured UltraFusion benchmark." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 553, + 168 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 553, + 168 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 553, + 168 + ], + "type": "image", + "image_path": "0c5de0821e51669d6cfebf1021f7caf4ebfee2582ec8eb6a78669ed62e7d1548.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 188, + 294, + 282 + ], + "blocks": [ + { + "bbox": [ + 58, + 188, + 294, + 282 + ], + "lines": [ + { + "bbox": [ + 58, + 188, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 58, + 188, + 294, + 282 + ], + "type": "image", + "image_path": "fc03860d1f45d750b32b003f2f060649ac00a6b85631b5fc9e857954e105b739.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 285, + 288, + 297 + ], + "lines": [ + { + "bbox": [ + 62, + 285, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 288, + 297 + ], + "type": "text", + "content": "Figure 10. Visual results on dynamic RealHDRV dataset [39]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 304, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 304, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 54, + 304, + 295, + 519 + ], + "type": "text", + "content": "robustness of UltraFusion to global and local motion, we use the RealHDRV dataset [39]. We extract the corresponding over-exposed image from the HDR ground truth as input. Tab. 2 demonstrates that our UltraFusion achieve state-of-the-art performance in terms of all metrics. For dynamic scenes, TMQI metric is a particularly important metric, as it is specially designed for HDR evaluation by assessing structural similarity between fusion output and ground truth. Since MEF methods are mainly designed for static scenes, they lack the capability to handle motion, resulting in low TMQI scores and overlay artifacts, as shown in Fig. 10. HDR reconstruction methods trained on dynamic datasets achieve better performance, but still produce noticeable artifact. On the other side, due to our soft inpainting guidance, UltraFusion is much more robust to misalignment and occlusion and the fusion output contains almost no artifacts. It achieves the highest TMQI in Tab. 2 and the best visual result in Fig. 10." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": "Evaluation on our UltraFusion benchmark. At last, the evaluation on our benchmark validates the robustness of UltraFusion in the most challenging scenes. On all four non-reference metrics, our method outperforms competitors by a large margin, as shown in Tab. 2. The qualitative comparison in Fig. 9 is also consistent with the quantitative metrics. For example, in the red zoom-in patch, integrating the highly bright sun from the under-exposed image into the over-exposed image is extremely challenging. Other methods fail to maintain the shape of the sun or preserve high contrast in the fused region, while our method naturally reconstructs the sun, preserving its appearance and visual pleasing tone of the whole image. Moreover, we conduct a user study on our proposed benchmark. Specifically, we randomly select 20 scenes from our benchmark and invite 136 different users to participate. For each scene, each user" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 213, + 552, + 276 + ], + "blocks": [ + { + "bbox": [ + 313, + 188, + 553, + 210 + ], + "lines": [ + { + "bbox": [ + 313, + 188, + 553, + 210 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 553, + 210 + ], + "type": "text", + "content": "Table 3. Ablation studies of three key components of our proposed UltraFusion on RealHDRV dataset [39]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 213, + 552, + 276 + ], + "lines": [ + { + "bbox": [ + 317, + 213, + 552, + 276 + ], + "spans": [ + { + "bbox": [ + 317, + 213, + 552, + 276 + ], + "type": "table", + "html": "
ModelTMQI↑MUSIQ↑
w/o Alignment Strategy0.742763.67
w/o Decompose-and-Fuse Control Branch0.887266.94
w/o Fidelity Control Branch0.876367.36
Ours UltraFusion0.892567.51
", + "image_path": "fb5aa52949fbbc1df2d002e6a492dc3a4a686669c3ec05ab5caa99f8bcc23f90.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 282, + 555, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 555, + 354 + ], + "type": "text", + "content": "is asked to compare our method with a randomly chosen baseline. The user study in Fig. 6 (b) indicates our method is more favored by users than competitors. This outcome aligns with the non-reference metric evaluation, showing that our method produces more natural images. More visual comparisons are available in the supplementary." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 361, + 412, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 412, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 412, + 372 + ], + "type": "text", + "content": "4.3. Ablation studies" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 378, + 553, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 553, + 415 + ], + "type": "text", + "content": "To validate the effectiveness of our UltraFusion, we conduct ablation studies on the three key components, followed by an in-depth exploration of the designs of them." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 415, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 629 + ], + "type": "text", + "content": "Key components. First, we perform the ablation study of the proposed three key components, including the alignment strategy, decompose-and-fuse control branch, and fidelity control branch, on the RealHDRV [39] dataset, as shown in Tab. 3. We first remove our alignment strategy, opting instead to train the model directly on the original SICE [1] dataset and input differently exposed image pairs without coarse alignment. Without the pre-alignment stage and the data synthesis pipeline designed for it, performance drops significantly in terms of TMQI. Fig. 11 (a) also shows that the model fails to implicitly align large motions. Then, we replace the decompose-and-fuse control branch (DFCB) with the vanilla ControlNet [65], and it fails to fuse features with large exposure differences, which leads to the loss of details (see Fig. 11 (b)) and a decrease in TMQI. Finally, when we exclude the fidelity control branch (FCB), Ultra-Fusion loses the capability to maintain the detailed structure and vivid color, as indicated by Fig. 11 (c)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "Alignment strategy. Second, we perform ablation of the proposed alignment strategy, which is the most critical design for our framework. Our alignment strategy mainly consists of a data synthesis pipeline at training time and a pre-alignment module at testing time. Without both of them, the whole model degrades to a multi-exposure fusion method as shown in Fig. 12 (c). Performing pre-alignment can reduce" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16117" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 135, + 125 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 135, + 125 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 135, + 125 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 135, + 125 + ], + "type": "image", + "image_path": "5abb710dbb0e448dd45cb1770ae17b51a45e59abde7b099cebaa43fd2be006ba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 126, + 135, + 176 + ], + "blocks": [ + { + "bbox": [ + 58, + 126, + 135, + 176 + ], + "lines": [ + { + "bbox": [ + 58, + 126, + 135, + 176 + ], + "spans": [ + { + "bbox": [ + 58, + 126, + 135, + 176 + ], + "type": "image", + "image_path": "e787c4b7688cc487b7374b6bb6f2522faf7659013dde49b96d56eef566d19ac8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 177, + 135, + 214 + ], + "blocks": [ + { + "bbox": [ + 58, + 177, + 135, + 214 + ], + "lines": [ + { + "bbox": [ + 58, + 177, + 135, + 214 + ], + "spans": [ + { + "bbox": [ + 58, + 177, + 135, + 214 + ], + "type": "image", + "image_path": "6e0f7a164f0a723e511fa6f2a2dde3cab4370632af847f02973953b46220f0ce.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 215, + 133, + 220 + ], + "lines": [ + { + "bbox": [ + 58, + 215, + 133, + 220 + ], + "spans": [ + { + "bbox": [ + 58, + 215, + 133, + 220 + ], + "type": "text", + "content": "Under-Exposure Over-Exposure" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 221, + 135, + 259 + ], + "blocks": [ + { + "bbox": [ + 58, + 221, + 135, + 259 + ], + "lines": [ + { + "bbox": [ + 58, + 221, + 135, + 259 + ], + "spans": [ + { + "bbox": [ + 58, + 221, + 135, + 259 + ], + "type": "image", + "image_path": "8b152c2bbf15ea46b11b42ef9e184c4d92cce9513e098abbfab65abf19dbe7f9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 260, + 93, + 270 + ], + "lines": [ + { + "bbox": [ + 59, + 260, + 93, + 270 + ], + "spans": [ + { + "bbox": [ + 59, + 260, + 93, + 270 + ], + "type": "text", + "content": "Align Strategy" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 136, + 72, + 294, + 176 + ], + "blocks": [ + { + "bbox": [ + 136, + 72, + 294, + 176 + ], + "lines": [ + { + "bbox": [ + 136, + 72, + 294, + 176 + ], + "spans": [ + { + "bbox": [ + 136, + 72, + 294, + 176 + ], + "type": "image", + "image_path": "acc38ca9ba787f3e692385b170efda8b705adbd9f32bc6db7d501a053f18f9b2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 137, + 177, + 214, + 214 + ], + "blocks": [ + { + "bbox": [ + 137, + 177, + 214, + 214 + ], + "lines": [ + { + "bbox": [ + 137, + 177, + 214, + 214 + ], + "spans": [ + { + "bbox": [ + 137, + 177, + 214, + 214 + ], + "type": "image", + "image_path": "024f5917bf46a81601c7326afcd47098d80a7b0b6624c032a70d8d8dd8f4d399.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 215, + 213, + 220 + ], + "lines": [ + { + "bbox": [ + 138, + 215, + 213, + 220 + ], + "spans": [ + { + "bbox": [ + 138, + 215, + 213, + 220 + ], + "type": "text", + "content": "Under-Exposure Over-Exposure" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 137, + 221, + 214, + 258 + ], + "blocks": [ + { + "bbox": [ + 137, + 221, + 214, + 258 + ], + "lines": [ + { + "bbox": [ + 137, + 221, + 214, + 258 + ], + "spans": [ + { + "bbox": [ + 137, + 221, + 214, + 258 + ], + "type": "image", + "image_path": "a9a3606b36d94eb2244055ecd14827cb0787d0afc33163268628c466fc9a0669.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 259, + 170, + 266 + ], + "lines": [ + { + "bbox": [ + 145, + 259, + 170, + 266 + ], + "spans": [ + { + "bbox": [ + 145, + 259, + 170, + 266 + ], + "type": "text", + "content": "w/o DFCB" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 260, + 214, + 266 + ], + "lines": [ + { + "bbox": [ + 137, + 260, + 214, + 266 + ], + "spans": [ + { + "bbox": [ + 137, + 260, + 214, + 266 + ], + "type": "text", + "content": "w/o DFCB Ours" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 219, + 260, + 248, + 266 + ], + "lines": [ + { + "bbox": [ + 219, + 260, + 248, + 266 + ], + "spans": [ + { + "bbox": [ + 219, + 260, + 248, + 266 + ], + "type": "text", + "content": "w/o FCB" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 219, + 177, + 256, + 214 + ], + "blocks": [ + { + "bbox": [ + 219, + 177, + 256, + 214 + ], + "lines": [ + { + "bbox": [ + 219, + 177, + 256, + 214 + ], + "spans": [ + { + "bbox": [ + 219, + 177, + 256, + 214 + ], + "type": "image", + "image_path": "0ed35ecb9248ccc15261403a4788d48b614a161c45496a7dedd22d482d0741ba.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 219, + 215, + 256, + 220 + ], + "lines": [ + { + "bbox": [ + 219, + 215, + 256, + 220 + ], + "spans": [ + { + "bbox": [ + 219, + 215, + 256, + 220 + ], + "type": "text", + "content": "Under-Exposure" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 257, + 177, + 293, + 214 + ], + "blocks": [ + { + "bbox": [ + 257, + 177, + 293, + 214 + ], + "lines": [ + { + "bbox": [ + 257, + 177, + 293, + 214 + ], + "spans": [ + { + "bbox": [ + 257, + 177, + 293, + 214 + ], + "type": "image", + "image_path": "a0584271bf223fb9120b7cbe2268dcfea0b8b1e6238c885c3098bb2a6243505b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 215, + 293, + 220 + ], + "lines": [ + { + "bbox": [ + 257, + 215, + 293, + 220 + ], + "spans": [ + { + "bbox": [ + 257, + 215, + 293, + 220 + ], + "type": "text", + "content": "Over-Exposure" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 58, + 300, + 136, + 354 + ], + "blocks": [ + { + "bbox": [ + 55, + 274, + 295, + 296 + ], + "lines": [ + { + "bbox": [ + 55, + 274, + 295, + 296 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 295, + 296 + ], + "type": "text", + "content": "Figure 11. Visual results of ablating each key component of our method. Each key component contributes to the final results." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 300, + 136, + 354 + ], + "lines": [ + { + "bbox": [ + 58, + 300, + 136, + 354 + ], + "spans": [ + { + "bbox": [ + 58, + 300, + 136, + 354 + ], + "type": "image", + "image_path": "c10b1b7f5e11a52a91f74342a4176e312873317f2edf197febdeb6239222bf0b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 354, + 119, + 361 + ], + "lines": [ + { + "bbox": [ + 75, + 354, + 119, + 361 + ], + "spans": [ + { + "bbox": [ + 75, + 354, + 119, + 361 + ], + "type": "text", + "content": "(a) Under Exposure" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 136, + 300, + 214, + 354 + ], + "blocks": [ + { + "bbox": [ + 136, + 300, + 214, + 354 + ], + "lines": [ + { + "bbox": [ + 136, + 300, + 214, + 354 + ], + "spans": [ + { + "bbox": [ + 136, + 300, + 214, + 354 + ], + "type": "image", + "image_path": "3a56d6a8e5d4d08ace351fa76c5620bab4e4ec3e9a22c4c92f328bfaa425edf9.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 354, + 195, + 361 + ], + "lines": [ + { + "bbox": [ + 153, + 354, + 195, + 361 + ], + "spans": [ + { + "bbox": [ + 153, + 354, + 195, + 361 + ], + "type": "text", + "content": "(b) Over Exposure" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 215, + 300, + 294, + 354 + ], + "blocks": [ + { + "bbox": [ + 215, + 300, + 294, + 354 + ], + "lines": [ + { + "bbox": [ + 215, + 300, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 215, + 300, + 294, + 354 + ], + "type": "image", + "image_path": "cc8d9ed2386622e856315b3863ae14254d27e91cbe1e66801442da63f7f09fd2.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 354, + 294, + 361 + ], + "lines": [ + { + "bbox": [ + 215, + 354, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 215, + 354, + 294, + 361 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 215, + 354, + 294, + 361 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 215, + 354, + 294, + 361 + ], + "type": "text", + "content": " Data Syn. & Pre-Alignment" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 58, + 361, + 136, + 414 + ], + "blocks": [ + { + "bbox": [ + 58, + 361, + 136, + 414 + ], + "lines": [ + { + "bbox": [ + 58, + 361, + 136, + 414 + ], + "spans": [ + { + "bbox": [ + 58, + 361, + 136, + 414 + ], + "type": "image", + "image_path": "6e8091871fb133a04c7174c7f44afe54d8861bc0a0db0eefd4837526ec2c7eb7.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 415, + 116, + 421 + ], + "lines": [ + { + "bbox": [ + 77, + 415, + 116, + 421 + ], + "spans": [ + { + "bbox": [ + 77, + 415, + 116, + 421 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 77, + 415, + 116, + 421 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 77, + 415, + 116, + 421 + ], + "type": "text", + "content": " Data Syn." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 83, + 425, + 267, + 435 + ], + "lines": [ + { + "bbox": [ + 83, + 425, + 267, + 435 + ], + "spans": [ + { + "bbox": [ + 83, + 425, + 267, + 435 + ], + "type": "text", + "content": "Figure 12. Effectiveness of our alignment strategy." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 136, + 361, + 214, + 414 + ], + "blocks": [ + { + "bbox": [ + 136, + 361, + 214, + 414 + ], + "lines": [ + { + "bbox": [ + 136, + 361, + 214, + 414 + ], + "spans": [ + { + "bbox": [ + 136, + 361, + 214, + 414 + ], + "type": "image", + "image_path": "8672368948b28c961db0260e2c6c7affda4d6a2b9bd28200b0f3f6477080f800.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 150, + 415, + 201, + 421 + ], + "lines": [ + { + "bbox": [ + 150, + 415, + 201, + 421 + ], + "spans": [ + { + "bbox": [ + 150, + 415, + 201, + 421 + ], + "type": "text", + "content": "(e) " + }, + { + "bbox": [ + 150, + 415, + 201, + 421 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 150, + 415, + 201, + 421 + ], + "type": "text", + "content": " Pre-Alignment" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 215, + 361, + 294, + 414 + ], + "blocks": [ + { + "bbox": [ + 215, + 361, + 294, + 414 + ], + "lines": [ + { + "bbox": [ + 215, + 361, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 215, + 361, + 294, + 414 + ], + "type": "image", + "image_path": "a9a6fcbf963dd66ee7dd71cd7c787b29ff98798b7e186771ac1656ebf13bec01.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 415, + 286, + 422 + ], + "lines": [ + { + "bbox": [ + 223, + 415, + 286, + 422 + ], + "spans": [ + { + "bbox": [ + 223, + 415, + 286, + 422 + ], + "type": "text", + "content": "(f) Ours Alignment Strategy" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 437, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 521 + ], + "type": "text", + "content": "artifacts to some extent (see Fig. 12 (d)), but may still be prone to alignment errors. Our data synthesis pipeline can improves the robustness of our algorithm to unaligned conditions, but motion in the occluded regions still cannot be solved, as shown in Fig. 12. After combining data synthesis pipeline and pre-alignment stage, our algorithm demonstrates strong capability towards dynamic scenes." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 55, + 521, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 665 + ], + "type": "text", + "content": "Decompose-and-fuse control branch. At last, we evaluate how the form of soft guidance provided by the underexposed image impacts the recovery of the highlight regions. When using the RGB under-exposed image " + }, + { + "bbox": [ + 55, + 521, + 295, + 665 + ], + "type": "inline_equation", + "content": "I_{ue}" + }, + { + "bbox": [ + 55, + 521, + 295, + 665 + ], + "type": "text", + "content": " as guidance, the model will ignore some details in the output (Fig. 13 (c)). Replacing " + }, + { + "bbox": [ + 55, + 521, + 295, + 665 + ], + "type": "inline_equation", + "content": "I_{ue}" + }, + { + "bbox": [ + 55, + 521, + 295, + 665 + ], + "type": "text", + "content": " with its structure information " + }, + { + "bbox": [ + 55, + 521, + 295, + 665 + ], + "type": "inline_equation", + "content": "S_{ue}" + }, + { + "bbox": [ + 55, + 521, + 295, + 665 + ], + "type": "text", + "content": " retains more details (Fig. 13 (d)), but without generating vivid color. By incorporating both under-exposed color information " + }, + { + "bbox": [ + 55, + 521, + 295, + 665 + ], + "type": "inline_equation", + "content": "C_{ue}" + }, + { + "bbox": [ + 55, + 521, + 295, + 665 + ], + "type": "text", + "content": " and structure information " + }, + { + "bbox": [ + 55, + 521, + 295, + 665 + ], + "type": "inline_equation", + "content": "S_{ue}" + }, + { + "bbox": [ + 55, + 521, + 295, + 665 + ], + "type": "text", + "content": ", the reconstructed results can maintain more details and color consistency (see Fig. 13 (e)). At last, multi-scale cross-attention further improves the fusion result (see Fig. 13 (f))." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 55, + 672, + 246, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 246, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 246, + 685 + ], + "type": "text", + "content": "4.4. Application on general image fusion" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "One advantage of the UltraFusion is that it can be extended to general image fusion, thanks to the flexibility of our" + } + ] + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 395, + 123 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 395, + 123 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 395, + 123 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 395, + 123 + ], + "type": "image", + "image_path": "d2e126ab4c4120a57e9b0c1262da9e33a36254f8a2d6d8a19706c9077e7584e7.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 124, + 382, + 131 + ], + "lines": [ + { + "bbox": [ + 329, + 124, + 382, + 131 + ], + "spans": [ + { + "bbox": [ + 329, + 124, + 382, + 131 + ], + "type": "text", + "content": "(a) Under Exposure" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 395, + 72, + 474, + 123 + ], + "blocks": [ + { + "bbox": [ + 395, + 72, + 474, + 123 + ], + "lines": [ + { + "bbox": [ + 395, + 72, + 474, + 123 + ], + "spans": [ + { + "bbox": [ + 395, + 72, + 474, + 123 + ], + "type": "image", + "image_path": "1234e3eac74e46d5b3d1ddd6537297a776e49e8a65e0af8cf33e6d103351186f.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 124, + 458, + 131 + ], + "lines": [ + { + "bbox": [ + 408, + 124, + 458, + 131 + ], + "spans": [ + { + "bbox": [ + 408, + 124, + 458, + 131 + ], + "type": "text", + "content": "(b) Over Exposure" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 474, + 72, + 553, + 123 + ], + "blocks": [ + { + "bbox": [ + 474, + 72, + 553, + 123 + ], + "lines": [ + { + "bbox": [ + 474, + 72, + 553, + 123 + ], + "spans": [ + { + "bbox": [ + 474, + 72, + 553, + 123 + ], + "type": "image", + "image_path": "e70afd959c2dfca63614c9e5376b855b8494bbfea371877023ac947966c1f2fe.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 316, + 132, + 395, + 184 + ], + "blocks": [ + { + "bbox": [ + 316, + 132, + 395, + 184 + ], + "lines": [ + { + "bbox": [ + 316, + 132, + 395, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 132, + 395, + 184 + ], + "type": "image", + "image_path": "418f6375d1bb490a9b7f192e3db8751a8d4df423921c802fa2d980cb5bad9348.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 346, + 186, + 365, + 194 + ], + "lines": [ + { + "bbox": [ + 346, + 186, + 365, + 194 + ], + "spans": [ + { + "bbox": [ + 346, + 186, + 365, + 194 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 346, + 186, + 365, + 194 + ], + "type": "inline_equation", + "content": "S_{ue}" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 395, + 132, + 474, + 184 + ], + "blocks": [ + { + "bbox": [ + 395, + 132, + 474, + 184 + ], + "lines": [ + { + "bbox": [ + 395, + 132, + 474, + 184 + ], + "spans": [ + { + "bbox": [ + 395, + 132, + 474, + 184 + ], + "type": "image", + "image_path": "0e7e24886d1a941348040818f60e36884a30d1ac14e1cc1eb74b10cd7e609740.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 186, + 453, + 194 + ], + "lines": [ + { + "bbox": [ + 416, + 186, + 453, + 194 + ], + "spans": [ + { + "bbox": [ + 416, + 186, + 453, + 194 + ], + "type": "text", + "content": "(e) " + }, + { + "bbox": [ + 416, + 186, + 453, + 194 + ], + "type": "inline_equation", + "content": "S_{ue} + C_{ue}" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 475, + 132, + 553, + 184 + ], + "blocks": [ + { + "bbox": [ + 504, + 124, + 523, + 132 + ], + "lines": [ + { + "bbox": [ + 504, + 124, + 523, + 132 + ], + "spans": [ + { + "bbox": [ + 504, + 124, + 523, + 132 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 504, + 124, + 523, + 132 + ], + "type": "inline_equation", + "content": "I_{ue}" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 475, + 132, + 553, + 184 + ], + "lines": [ + { + "bbox": [ + 475, + 132, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 475, + 132, + 553, + 184 + ], + "type": "image", + "image_path": "52273cdf2c8c480e0843f8f4c08895e69350fea189c8c30ad70cff61494c1731.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 501, + 186, + 527, + 193 + ], + "lines": [ + { + "bbox": [ + 501, + 186, + 527, + 193 + ], + "spans": [ + { + "bbox": [ + 501, + 186, + 527, + 193 + ], + "type": "text", + "content": "(f) DFCB" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 315, + 224, + 392, + 277 + ], + "blocks": [ + { + "bbox": [ + 313, + 198, + 555, + 219 + ], + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 219 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 219 + ], + "type": "text", + "content": "Figure 13. Detailed ablation study on the design choices of decompose-and-fuse control branch." + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 224, + 392, + 277 + ], + "lines": [ + { + "bbox": [ + 315, + 224, + 392, + 277 + ], + "spans": [ + { + "bbox": [ + 315, + 224, + 392, + 277 + ], + "type": "image", + "image_path": "93a6f208f33de5c9abd764e12cb155a7145b6c512100e50a6a7eb6496a34a3bb.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 278, + 383, + 285 + ], + "lines": [ + { + "bbox": [ + 325, + 278, + 383, + 285 + ], + "spans": [ + { + "bbox": [ + 325, + 278, + 383, + 285 + ], + "type": "text", + "content": "(a) Input Over Exposure" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 289, + 554, + 320 + ], + "lines": [ + { + "bbox": [ + 313, + 289, + 554, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 554, + 320 + ], + "type": "text", + "content": "Figure 14. Extension to general fusion. Given totally different underexposed images as guidance (upper right corner), our method can also generate different fusion results." + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 392, + 225, + 469, + 277 + ], + "blocks": [ + { + "bbox": [ + 392, + 225, + 469, + 277 + ], + "lines": [ + { + "bbox": [ + 392, + 225, + 469, + 277 + ], + "spans": [ + { + "bbox": [ + 392, + 225, + 469, + 277 + ], + "type": "image", + "image_path": "34b8effd521b6019697b2527d039a37a14246e5cc88117b100cd7084c79e6305.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 278, + 457, + 285 + ], + "lines": [ + { + "bbox": [ + 405, + 278, + 457, + 285 + ], + "spans": [ + { + "bbox": [ + 405, + 278, + 457, + 285 + ], + "type": "text", + "content": "(b) Our Fusion Result" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 463, + 278, + 552, + 285 + ], + "lines": [ + { + "bbox": [ + 463, + 278, + 552, + 285 + ], + "spans": [ + { + "bbox": [ + 463, + 278, + 552, + 285 + ], + "type": "text", + "content": "(c) Another Fusion w/ Diff. Highlight" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 470, + 225, + 547, + 277 + ], + "blocks": [ + { + "bbox": [ + 470, + 225, + 547, + 277 + ], + "lines": [ + { + "bbox": [ + 470, + 225, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 470, + 225, + 547, + 277 + ], + "type": "image", + "image_path": "766deb6635d7215f24f16182405356b0287fcd7dccdefca626f397f52bb120e1.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + } + ], + "index": 52 + }, + { + "bbox": [ + 313, + 329, + 554, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 554, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 554, + 411 + ], + "type": "text", + "content": "guided inpainting. To illustrate this potential, we explore one additional interesting demo to fuse two irrelevant images captured by different cameras at different locations. As shown in Fig. 14, UltraFusion successfully copies the moon (b) or blue sky (c) to the over-exposed image, unlocking many interesting potential applications (e.g., image harmonization)." + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 314, + 421, + 388, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 421, + 388, + 433 + ], + "spans": [ + { + "bbox": [ + 314, + 421, + 388, + 433 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 56 + }, + { + "bbox": [ + 313, + 437, + 554, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 592 + ], + "type": "text", + "content": "In this work, we introduce a novel approach to HDR imaging, tackling challenges presented by significant exposure differences and large motion. By modeling the fusion process as a guided inpainting problem and utilizing the under-exposed image for soft guidance, our method is robust with alignment errors and circumvents tone mapping, resulting in natural, artifact-free outputs. We also propose decompose-and-fuse control branch and fidelity control branch to improve feature modulation and fidelity preservation of ControlNet. Extensive experiments on existing datasets and our captured benchmark demonstrate the robustness and effectiveness of our method compared to previous HDR methods." + } + ] + } + ], + "index": 57 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "With extreme exposure differences and challenging non-rigid motion, occlusion mask estimation may introduce errors, causing the restoration of certain highlight regions degrades to single image HDR. While our method is able to use diffusion priors to restore these highlights, restoration without under-exposed information remains unreliable. Moreover, it takes almost 3.3s to fuse " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " inputs on an NVIDIA RTX 4090 GPU. A more exposure-robust optical flow algorithm and a faster implementation are highly desirable, and we leave them for future work." + } + ] + } + ], + "index": 58 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16118" + } + ] + } + ], + "index": 59 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 295, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 295, + 139 + ], + "type": "text", + "content": "This work was supported by the National Key R&D Program of China No.2022ZD0160201, Shanghai Artificial Intelligence Laboratory and RGC Early Career Scheme (ECS) No. 24209224." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 114, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 114, + 164 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 114, + 164 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 171, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 171, + 296, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 171, + 296, + 215 + ], + "spans": [ + { + "bbox": [ + 61, + 171, + 296, + 215 + ], + "type": "text", + "content": "[1] Jianrui Cai, Shuhang Gu, and Lei Zhang. Learning a deep single image contrast enhancer from multi-exposure images. IEEE Transactions on Image Processing, 27(4):2049-2062, 2018. 3, 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 217, + 296, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 217, + 296, + 261 + ], + "spans": [ + { + "bbox": [ + 61, + 217, + 296, + 261 + ], + "type": "text", + "content": "[2] Cong Cao, Huanjing Yue, Xin Liu, and Jingyu Yang. Unsupervised HDR image and video tone mapping via contrastive learning. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 262, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 262, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 61, + 262, + 295, + 316 + ], + "type": "text", + "content": "[3] Jie Chen, Zaifeng Yang, Tsz Nam Chan, Hui Li, Junhui Hou, and Lap-Pui Chau. Attention-guided progressive neural texture fusion for high dynamic range image restoration. IEEE Transactions on Image Processing, 31:2661-2672, 2022. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 319, + 295, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 295, + 373 + ], + "type": "text", + "content": "[4] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 375, + 296, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 375, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 375, + 296, + 430 + ], + "type": "text", + "content": "[5] Ciprian Corneanu, Raghudeep Gadde, and Aleix M Martinez. Latentpaint: Image inpainting in latent space with diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4334-4343, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 432, + 296, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 296, + 475 + ], + "type": "text", + "content": "[6] Abhishek Goswami, Aru Ranjan Singh, Francesco Banterle, Kurt Debattista, and Thomas Bashford-Rogers. Semantic aware diffusion inverse tone mapping. arXiv preprint arXiv:2405.15468, 2024.3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 477, + 296, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 477, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 477, + 296, + 520 + ], + "type": "text", + "content": "[7] Michael D Grossberg and Shree K Nayar. Determining the camera response from images: What is knowable? IEEE Transactions on pattern analysis and machine intelligence, 25(11):1455-1467, 2003. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 522, + 296, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 522, + 296, + 565 + ], + "spans": [ + { + "bbox": [ + 62, + 522, + 296, + 565 + ], + "type": "text", + "content": "[8] Yuanshen Guan, Ruikang Xu, Mingde Yao, Ruisheng Gao, Lizhi Wang, and Zhiwei Xiong. Diffusion-promoted HDR video reconstruction. arXiv preprint arXiv:2406.08204, 2024.3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 567, + 296, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 296, + 622 + ], + "type": "text", + "content": "[9] Samuel W Hasinoff, Dillon Sharlet, Ryan Geiss, Andrew Adams, Jonathan T Barron, Florian Kainz, Jiawen Chen, and Marc Levoy. Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM Transactions on Graphics (ToG), 35(6):1-12, 2016. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 657 + ], + "type": "text", + "content": "[10] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "text", + "content": "[11] Litao Hu, Huaijin Chen, and Jan P Allebach. Joint multiscale tone mapping and denoising for HDR image enhancement. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 729-738, 2022. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[12] Tao Hu, Qingsen Yan, Yuankai Qi, and Yanning Zhang. Generating content for HDR deghosting from frequency view. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25732-25741, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 119, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 172 + ], + "type": "text", + "content": "[13] Ting Jiang, Chuan Wang, Xinpeng Li, Ru Li, Haoqiang Fan, and Shuaicheng Liu. Meflut: Unsupervised 1d lookup tables for multi-exposure image fusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10542-10551, 2023. 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "type": "text", + "content": "[14] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. arXiv preprint arXiv:2403.06976, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 251 + ], + "type": "text", + "content": "[15] Nima Khademi Kalantari, Ravi Ramamoorthi, et al. Deep high dynamic range imaging of dynamic scenes. ACM Trans. Graph., 36(4):144-1, 2017. 2, 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "type": "text", + "content": "[16] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 299, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 299, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 317, + 299, + 553, + 342 + ], + "type": "text", + "content": "[17] Lingtong Kong, Bo Li, Yike Xiong, Hao Zhang, Hong Gu, and Jinwei Chen. Safnet: Selective alignment fusion network for efficient HDR imaging. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2, 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 343, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 343, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 317, + 343, + 553, + 396 + ], + "type": "text", + "content": "[18] Baiang Li, Sizhuo Ma, Yanhong Zeng, Xiaogang Xu, Youqing Fang, Zhao Zhang, Jian Wang, and Kai Chen. Sagiri: Low dynamic range image enhancement with generative diffusion prior. arXiv preprint arXiv:2406.09389, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 399, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 442 + ], + "type": "text", + "content": "[19] Hui Li, Kede Ma, Hongwei Yong, and Lei Zhang. Fast multiscale structural patch decomposition for multi-exposure image fusion. IEEE Transactions on Image Processing, 29: 5805-5816, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 445, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 488 + ], + "type": "text", + "content": "[20] Pengwei Liang, Junjun Jiang, Xianming Liu, and Jiayi Ma. Fusion from decomposition: A self-supervised decomposition approach for image fusion. In European Conference on Computer Vision, pages 719-735. Springer, 2022. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 490, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 490, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 490, + 553, + 533 + ], + "type": "text", + "content": "[21] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Wanli Ouyang, Yu Qiao, and Chao Dong. Diff-bir: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070, 2023. 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "type": "text", + "content": "[22] Renshuai Liu, Chengyang Li, Haitao Cao, Yinglin Zheng, Ming Zeng, and Xuan Cheng. Emef: ensemble multi-exposure image fusion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1710-1718, 2023. 3, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 590, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 645 + ], + "type": "text", + "content": "[23] Yuhao Liu, Zhanghan Ke, Fang Liu, Nanxuan Zhao, and Rynson WH Lau. Diff-plugin: Revitalizing details for diffusion-based low-level tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4197-4208, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 690 + ], + "type": "text", + "content": "[24] Zhen Liu, Yinglong Wang, Bing Zeng, and Shuaicheng Liu. Ghost-free high dynamic range imaging with context-aware transformer. In European Conference on computer vision, pages 344-360. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[25] Zhen Liu, Yinglong Wang, Bing Zeng, and Shuaicheng Liu. Ghost-free high dynamic range imaging with context-aware" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16119" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "transformer. In European Conference on computer vision, pages 344-360. Springer, 2022. 1, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "text", + "content": "[26] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "text", + "content": "[27] Kede Ma, Kai Zeng, and Zhou Wang. Perceptual quality assessment for multi-exposure image fusion. IEEE Transactions on Image Processing, 24(11):3345-3356, 2015. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 230 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 230 + ], + "type": "text", + "content": "[28] Kede Ma, Hui Li, Hongwei Yong, Zhou Wang, Deyu Meng, and Lei Zhang. Robust multi-exposure image fusion: a structural patch decomposition approach. IEEE Transactions on Image Processing, 26(5):2519-2532, 2017. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 294, + 274 + ], + "type": "text", + "content": "[29] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 295, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 295, + 309 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 295, + 309 + ], + "type": "text", + "content": "[30] Tom Mertens, Jan Kautz, and Frank Van Reeth. Exposure fusion. In 15th Pacific Conference on Computer Graphics and Applications (PG'07), pages 382-390. IEEE, 2007. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 311, + 295, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 311, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 311, + 295, + 354 + ], + "type": "text", + "content": "[31] Ozan Özdenizci and Robert Legenstein. Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10346-10357, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 355, + 295, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 295, + 376 + ], + "type": "text", + "content": "[32] Photomatix. Commercially-available hdr processing software. https://www.hdrsoft.com/.6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 378, + 295, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 378, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 56, + 378, + 295, + 434 + ], + "type": "text", + "content": "[33] K Ram Prabhakar, Rajat Arora, Adhitya Swaminathan, Kunal Pratap Singh, and R Venkatesh Babu. A fast, scalable, and reliable deghosting method for extreme exposure fusion. In 2019 IEEE International Conference on Computational Photography (ICCP), pages 1-8. IEEE, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 434, + 295, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 295, + 500 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 295, + 500 + ], + "type": "text", + "content": "[34] K Ram Prabhakar, Susmit Agrawal, Durgesh Kumar Singh, Balraj Ashwath, and R Venkatesh Babu. Towards practical and efficient high-resolution HDR deghosting with cnn. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16, pages 497-513. Springer, 2020. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 502, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 502, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 502, + 295, + 555 + ], + "type": "text", + "content": "[35] K Ram Prabhakar, V Sai Srikar, and R Venkatesh Babu. Deepfuse: A deep unsupervised approach for exposure fusion with extreme exposure image pairs. In Proceedings of the IEEE international conference on computer vision, pages 4714-4722, 2017. 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 557, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 295, + 612 + ], + "type": "text", + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 613, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 295, + 668 + ], + "type": "text", + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "text", + "content": "[38] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition, pages 8839-8849, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 160 + ], + "type": "text", + "content": "[39] Yong Shu, Liquan Shen, Xiangyu Hu, Mengyao Li, and Zihao Zhou. Towards real-world HDR video reconstruction: A large-scale benchmark dataset and a two-stage alignment network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2879-2888, 2024. 3, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 162, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 215 + ], + "type": "text", + "content": "[40] Jou Won Song, Ye-In Park, Kyeongbo Kong, Jaeho Kwak, and Suk-Ju Kang. Selective transhdr: Transformer-based selective HDR imaging using ghost region mask. In European Conference on Computer Vision, pages 288-304. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "type": "text", + "content": "[41] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3667-3676, 2020. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 272, + 553, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 325 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 325 + ], + "type": "text", + "content": "[42] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 327, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 380 + ], + "type": "text", + "content": "[43] Steven Tel, Zongwei Wu, Yulun Zhang, Barthélémy Heyrman, Cédric Demonceaux, Radu Timofte, and Dominique Gingac. Alignment-free HDR deghosting with semantics consistent transformer. arXiv preprint arXiv:2305.18135, 2023. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 382, + 553, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 382, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 316, + 382, + 553, + 447 + ], + "type": "text", + "content": "[44] Steven Tel, Zongwei Wu, Yulun Zhang, Barthélémy Heyrman, Cédric Demonceaux, Radu Timofte, and Dominique Gingac. Alignment-free HDR deghosting with semantics consistent transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12836-12845, 2023. 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 449, + 553, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 501 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 501 + ], + "type": "text", + "content": "[45] Yael Vinker, Inbar Huberman-Spiegelglas, and Raanan Fattal. Unpaired learning for high dynamic range image tone mapping. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14657-14666, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 503, + 553, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 503, + 553, + 546 + ], + "spans": [ + { + "bbox": [ + 316, + 503, + 553, + 546 + ], + "type": "text", + "content": "[46] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 548, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 591 + ], + "type": "text", + "content": "[47] Shuchen Weng, Peixuan Zhang, Yu Li, Si Li, Boxin Shi, et al. L-cad: Language-based colorization with any-level descriptions using diffusion priors. Advances in Neural Information Processing Systems, 36, 2024. 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 593, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 646 + ], + "type": "text", + "content": "[48] Guanyao Wu, Hongming Fu, Jinyuan Liu, Long Ma, Xin Fan, and Risheng Liu. Hybrid-supervised dual-search: Leveraging automatic learning for loss-free multi-exposure image fusion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5985-5993, 2024. 1, 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "text", + "content": "[49] Shangzhe Wu, Jiarui Xu, Yu-Wing Tai, and Chi-Keung Tang. Deep high dynamic range imaging with large foreground motions. In Proceedings of the European Conference on Computer Vision (ECCV), pages 117–132, 2018. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[50] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16120" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "type": "text", + "content": "with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22428-22437, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "text", + "content": "[51] Gangwei Xu, Yujin Wang, Jinwei Gu, Tianfan Xue, and Xin Yang. Hdrflow: Real-time HDR video reconstruction with large motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24851-24860, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 208 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 208 + ], + "type": "text", + "content": "[52] Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502-518, 2020. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 296, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 296, + 252 + ], + "type": "text", + "content": "[53] Han Xu, Jiayi Ma, and Xiao-Ping Zhang. Mef-gan: Multiexposure image fusion via generative adversarial networks. IEEE Transactions on Image Processing, 29:7203-7216, 2020. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 296, + 308 + ], + "type": "text", + "content": "[54] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8121-8130, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 296, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 296, + 354 + ], + "type": "text", + "content": "[55] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. International Journal of Computer Vision, 127:1106-1125, 2019. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 355, + 296, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 296, + 420 + ], + "type": "text", + "content": "[56] Qingsen Yan, Dong Gong, Qinfeng Shi, Anton van den Hengel, Chunhua Shen, Ian Reid, and Yanning Zhang. Attention-guided network for ghost-free high dynamic range imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1751-1760, 2019. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 296, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 296, + 465 + ], + "type": "text", + "content": "[57] Qingsen Yan, Lei Zhang, Yu Liu, Yu Zhu, Jinqiu Sun, Qinfeng Shi, and Yanning Zhang. Deep HDR imaging via a non-local network. IEEE Transactions on Image Processing, 29: 4308-4322, 2020. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 296, + 521 + ], + "type": "text", + "content": "[58] Qingsen Yan, Tao Hu, Yuan Sun, Hao Tang, Yu Zhu, Wei Dong, Luc Van Gool, and Yanning Zhang. Towards high-quality HDR dehosting with conditional diffusion models. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 296, + 567 + ], + "type": "text", + "content": "[59] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 296, + 602 + ], + "type": "text", + "content": "[60] Hojatollah Yeganeh and Zhou Wang. Objective quality assessment of tone-mapped images. IEEE Transactions on Image processing, 22(2):657-667, 2012. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 602, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 296, + 667 + ], + "type": "text", + "content": "[61] Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3575–3585, 2020. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "type": "text", + "content": "[62] Zhiyuan You, Xin Cai, Jinjin Gu, Tianfan Xue, and Chao Dong. Teaching large language models to regress accurate image quality scores using score distribution. arXiv preprint arXiv:2501.11561, 2025. 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 430 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 139 + ], + "type": "text", + "content": "[63] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photorealistic image restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25669-25680, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 555, + 183 + ], + "type": "text", + "content": "[64] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 185, + 555, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 555, + 229 + ], + "type": "text", + "content": "[65] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 230, + 555, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 555, + 262 + ], + "type": "text", + "content": "[66] Xingchen Zhang. Benchmarking and comparing multi-exposure image fusion algorithms. Information Fusion, 74: 111-131, 2021. 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 263, + 555, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 555, + 319 + ], + "type": "text", + "content": "[67] Shengyu Zhao, Yilun Sheng, Yue Dong, Eric I Chang, Yan Xu, et al. Maskflownet: Asymmetric feature matching with learnable occlusion mask. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6278-6287, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 319, + 555, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 555, + 373 + ], + "type": "text", + "content": "[68] Zixiang Zhao, Lilun Deng, Haowen Bai, Yukun Cui, Zhipeng Zhang, Yulun Zhang, Haotong Qin, Dongdong Chen, Jiangshe Zhang, Peng Wang, et al. Image fusion via vision-language model. arXiv preprint arXiv:2402.02235, 2024.3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 374, + 555, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 374, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 374, + 555, + 430 + ], + "type": "text", + "content": "[69] Pengfei Zhu, Yang Sun, Bing Cao, and Qinghua Hu. Task-customized mixture of adapters for general image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7099-7108, 2024. 3, 6" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16121" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/UnCommon Objects in 3D/4d2bd976-41fd-45e5-bfc6-8a4578f73725_content_list.json b/2025/UnCommon Objects in 3D/4d2bd976-41fd-45e5-bfc6-8a4578f73725_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..14285d66879edb5915b5da366c49fabf8266e4a8 --- /dev/null +++ b/2025/UnCommon Objects in 3D/4d2bd976-41fd-45e5-bfc6-8a4578f73725_content_list.json @@ -0,0 +1,1347 @@ +[ + { + "type": "text", + "text": "UnCommon Objects in 3D", + "text_level": 1, + "bbox": [ + 362, + 130, + 633, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xingchen Liu $^{\\infty}$ Piyush Tayal $^{\\infty}$ Jianyuan Wang $^{\\infty}$ Jesus Zarzar $^{+\\infty}$ Tom Monnier $^{\\infty}$ Konstantinos Tertikas $^{\\infty*}$ Jiali Duan $^{\\infty}$ Antoine Toisoul $^{\\infty}$ Jason Y. Zhang $^{\\dagger}$ Natalia Neverova $^{\\infty}$ Andrea Vedaldi $^{\\infty}$ Roman Shapovalov $^{\\infty}$ David Novotny $^{\\infty}$", + "bbox": [ + 150, + 179, + 849, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "*NKUA, Greece †KAUST †Carnegie Mellon University †Meta AI", + "bbox": [ + 214, + 263, + 777, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "uCO3D https://uco3d.github.io https://github.com/facebookresearch/uco3d", + "bbox": [ + 109, + 303, + 885, + 323 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/948844b57f01375a98ddd77397dac02a46e50d0415ca6e1316512cc7c65dbf5b.jpg", + "image_caption": [ + "Figure 1. We introduce UnCommon Objects in 3D (uCO3D), a large and diverse dataset of high-quality $360^{\\circ}$ videos covering over 1,000 object categories. Each video frame is 3D-annotated with accurate SfM cameras, point cloud, and a 3D Gaussian Splitting reconstruction." + ], + "image_footnote": [], + "bbox": [ + 93, + 348, + 906, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 753, + 328, + 771 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce Uncommon Objects in 3D (uCO3D), a new object-centric dataset for 3D deep learning and 3D generative AI. uCO3D is the largest publicly-available collection of high-resolution videos of objects with 3D annotations that ensures full- $360^{\\circ}$ coverage. uCO3D is significantly more diverse than MVImgNet and CO3Dv2, covering more than 1,000 object categories. It is also of higher quality, due", + "bbox": [ + 89, + 794, + 485, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to extensive quality checks of both the collected videos and the 3D annotations. Similar to analogous datasets, uCO3D contains annotations for 3D camera poses, depth maps and sparse point clouds. In addition, each object is equipped with a caption and a 3D Gaussian Splat reconstruction. We train several large 3D models on MVImgNet, CO3Dv2, and uCO3D and obtain superior results using the latter, showing that uCO3D is better for learning applications.", + "bbox": [ + 511, + 755, + 908, + 878 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14102", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 89, + 222, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The primacy of data has been the defining characteristic of the last decade of machine learning, alongside deep learning. The most powerful models in language, speech, and computer vision are large deep networks trained on massive amounts of data, and then further fine-tuned on its high-quality subset. This paradigm is expected to extend to all machine learning applications, including 3D vision.", + "bbox": [ + 89, + 114, + 482, + 220 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, 3D training data is much harder to come by than data for text, audio, and image processing.", + "bbox": [ + 89, + 220, + 482, + 251 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Seeking training data for large 3D neural networks [27], many have turned to synthetic datasets like Objaverse [13]. However, synthetic data is a poor substitute for real data in applications like digital twinning, which creates 3D models of real-life objects. This is why many photorealistic reconstruction networks [6, 23, 57, 58, 63, 67] are trained using real object-centric datasets like CO3D [47], MVImgNet [76], GSO [15], and OmniObject3D [68]. Real data is also crucial for generalization, as demonstrated by DUSt3r [64] for point map prediction and DepthAnything [74] for depth prediction, both of which are trained on numerous real datasets. Even non-curated image datasets like the billion-scale LAION [52] are applicable to 3D vision. For instance, text-to-3D generators [35, 45, 53, 54] build on LAION-trained large text-to-image models [5, 11].", + "bbox": [ + 89, + 251, + 482, + 477 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given the importance of 3D datasets, but also their relative scarcity, in this paper we ask what is the next step for real data in 3D vision. To answer this question, we note that, while the size of a dataset is crucial, in most cases its quality is just as important. For example, text-to-3D models [35, 55] are notoriously sensitive to fine-tuning data quality, training only with the best-looking models (e.g., Instant3D [35] uses only about $1\\%$ of Objaverse). We conclude that high-quality datasets balancing scale and fidelity are more valuable than simply amassing low-quality data.", + "bbox": [ + 89, + 478, + 482, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on this observation, we argue that there is a gap in the real object-centric 3D datasets that are currently available, as none strikes the optimal balance between quality and scale. For example, the 3D object scans in OmniObject3D [68] and GSO [15] provide very accurate geometry and textures, but only count a few thousand objects. Conversely, datasets like CO3D [47] and MvImgNet [76] contain orders-of-magnitude more objects, but lack reliable 3D scans. Instead, they provide many views of the objects together with lower-quality 3D cameras and point clouds reconstructed with structure-from-motion (SfM).", + "bbox": [ + 89, + 628, + 482, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we address this gap with a new dataset, Uncommon Objects in 3D (uCO3D), which better balances data quality and size (Tab. 1). Similar to CO3D, it comprises full- $360^{\\circ}$ crowd-sourced videos capturing objects from all sides, annotated with cameras and point clouds using SfM. Furthermore, uCO3D has much greater data diversity (Fig. 2) than prior alternatives as it contains objects", + "bbox": [ + 89, + 795, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/f5a214261e7fb05d8096d82a21d1fd86b13c1c40199307bf4a217eb12dcf6ed4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RealCount# ClassesData typeAnnotations
ShapeNet [8]X51k553D meshesmesh
Objaverse [13]X800k21k3D meshesmesh
Objaverse-XL [12]X10M2M3D meshesmesh
ABO [10]X8k633D meshesmesh
OmniObject3D [68]✓6k190Videos w/ meshescameras, mesh
GSO [14]✓1k17Views w/ meshescameras, mesh
Objectron [2]✓15k9Limited vp. videoscameras, 3D box
MVImgNet [76]✓220k238Limited vp. videoscameras, pcl
CO3D [47]✓19k50360° videoscameras, pcl
CO3Dv2 [47]✓40k50360° videoscameras, pcl
uCO3D (ours)✓170k1k360° videoscameras, 3DGS, caption
", + "bbox": [ + 514, + 89, + 903, + 219 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Overview of 3D object datasets. We compare the number of objects / classes, the type of data and associated annotations.", + "bbox": [ + 511, + 233, + 903, + 262 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "from the 1,070 visual object categories of the LVIS [22] taxonomy, which has long tails. For reference, MVImgNet and CO3Dv2 contain only 238 and 50 categories, respectively. These fine-grained categories are organized in supercategories, also shown in Fig. 2. Furthermore, uCO3D contains 170k scenes, which is more than four times larger than CO3Dv2's 40k. While this is less than MVImgNet's 220k, uCO3D's videos cover each object from all sides, as opposed to MVImgNet's partial object captures.", + "bbox": [ + 511, + 299, + 906, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Besides improving size and diversity, uCO3D also raises the quality bar. This was achieved by checking extensively both the collected videos and their 3D annotations. Differently from datasets like CO3Dv2 that still contain a certain portion of low-quality videos, in uCO3D we manually verified that each video provides full $360^{\\circ}$ turn-table covering all sides of the object. Additionally, $60\\%+$ of the videos have $1080p+$ resolution, higher than CO3Dv2. To ensure 3D-annotation quality, we improved both the reconstruction algorithm and the reconstruction validation. For camera reconstruction, we used VGGSfM [62], which is currently the best SfM system available, and is more robust and accurate than COLMAP [50], used in CO3Dv2 and MvImgNet. We also improve on CO3Dv2's active-learning camera quality evaluation by combining it with novel-view synthesis accuracy after reconstructing each scene using 3D Gaussian Splatting (3DGS) [29]. The latter also guarantees that scenes are reconstructible to a high quality, which is important for training of 3D models.", + "bbox": [ + 511, + 441, + 908, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We validate uCO3D's benefits in applications. We train two popular 3D models, LRM [27] and CAT3D [18], using uCO3D and demonstrate improved results compared to training on MVImgNet and CO3Dv2, which makes uCO3D the better data source for real object-centric 3D learning. We also use uCO3D to train a text-to-3D model following Instant3D's [35] two-stage design. The latter requires objects to be rendered from canonical viewpoints, and thus, so far, was limited to synthetic data. By using our 3DGS reconstructions, we 're-shoot' uCO3D's from these viewpoints, which allows to train a more realistic generator.", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14103", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1c9a9d2aec5e61eea77375b106e8b94f4d471dc011a18f91314b3318ac994d63.jpg", + "image_caption": [ + "Figure 2. Statistics of uCO3D. (Left) We plot the number of objects per super-category. In total, the dataset contains 50 super-categories, each gathering around 20 sub-categories. (Right) We show a word cloud of all 1,070 visual categories represented in the dataset." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 903, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 287, + 232, + 304 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Datasets of synthetic 3D objects. Historically, object-centric 3D datasets have predominantly been synthetic, composed of artist-generated 3D models. A prominent example is ShapeNet [8], with 51,000 meshes across 55 object categories. The meshes have detailed geometries, but relatively simplistic textures. Datasets such as 3DFUTURE [17], IKEA [36], Pix3D [56], and ABO [10] are less diverse, concentrating on furniture and other consumer goods. In contrast, ModelNet [69], DeepCAD [66], and ABC [33] provide CAD models with clean geometry but no texture. Objaverse [13] is perhaps the most impactful dataset following ShapeNet. It is significantly larger, comprising 800,000 3D meshes. Objaverse-XL [12] expands this collection to 10M objects. These datasets have been pivotal in enabling the first 3D deep generative models, including text-to-3D [35, 53, 54] and image-to-3D [6, 27, 73] models.", + "bbox": [ + 91, + 316, + 483, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Real 3D object datasets. Acquiring real-world 3D data presents many challenges, resulting in only a few real 3D-object datasets. Early datasets like Pascal3D [70] include multiple categories but provide only isolated object views and approximate 3D annotations. Conversely, DTU [28], BlendedMVS [51], GSO [15], OmniObject3D [68], Aria Digital Twin [44], and Digital Twin Catalog [1] provide 3D scans of objects, featuring high-quality 3D geometry and textures, but have only a few thousand objects.", + "bbox": [ + 89, + 580, + 482, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The use of 3D scanners significantly restricts the scale of data acquisition; consequently, other datasets capture multiview turntable-like videos of objects using consumer cameras. CO3D and CO3Dv2 [47] crowd-sourced $40,000360^{\\circ}$ object videos, providing 3D annotations by reconstructing point clouds and cameras using COLMAP SfM [50]. MvImgNet [76] collected even more videos (220,000) across more object categories (238), but their videos capture objects only partially, preventing full reconstruction. Objectron [2] is similar to MvImgNet, but with fewer videos (10,000). A common challenge is that large-scale datasets often rely on SfM for video reconstruction, which can", + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "lead to imprecise 3D annotations. uCO3D also employs SfM, but using VGGSfM, which has greater accuracy than COLMAP, and with a more reliable data validation setup. Furthermore, uCO3D is five times larger and significantly more diverse than CO3Dv2, encompassing 20 times more visual categories, and provides caption and 3D Gaussian Splat reconstructions of each object.", + "bbox": [ + 511, + 289, + 906, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Applications. In order to assess the quality of uCO3D, we measure how it benefits a number of popular downstream applications. First, we consider feedforward few-view 3D reconstruction models. Among those, LRM [27] is a transformer that maps an input image to a neural radiance field supported by a triplane [7]. LightplaneLRM [6] adds splatting layers and a memory-efficient renderer. Further extensions use different representations like 3D Gaussian Splats [59, 73, 77, 80] and meshes [65, 71].", + "bbox": [ + 511, + 402, + 905, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We also consider text-to-3D generators, which create 3D assets from text, and focus on the two-stages approach of Instant3D [35]. This is based on training a text-to-multi-view diffusion model [11, 48] which generates several 2D views of the object, followed by a 3D reconstruction network that outputs the 3D asset, all in a matter of seconds. The multi-view diffusion is improved in ViewDiff [26], MVDiffusion [60], IM-3D [42], CAT3D [18] and many others. AssetGen [55] further extends Instant3D by modelling material properties instead of baking in the radiance function and adds a texture refiner that outputs relightable PBR textures. As an illustration, we use uCO3D to train a model like CAT3D, which results in better new-view synthesis than the one trained on alternative datasets. We also show that the Gaussian Splat reconstructions provided with uCO3D can supervise, for the first time, an Instant3D-like pipeline using solely real-life data.", + "bbox": [ + 511, + 539, + 906, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Uncommon Objects in 3D", + "text_level": 1, + "bbox": [ + 511, + 813, + 756, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce uCO3D, our new dataset of real-life 3D objects. uCO3D comprises $360^{\\circ}$ turn-table-like videos of objects, crowdsourced and annotated with 3D cameras, point clouds, 3D Gaussians, and textual captions.", + "bbox": [ + 511, + 839, + 905, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14104", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1fe67a25a5d1ab72ff15d61286823e3b7b97aa7df867541a99fc374591138628.jpg", + "image_caption": [ + "Video" + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 145, + 375 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0190d353e987df0730825c5db96919ce00b367b182a5e27137b0c4d88a877fe5.jpg", + "image_caption": [ + "a) Sparse SfM reconstruction" + ], + "image_footnote": [], + "bbox": [ + 148, + 89, + 411, + 373 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0b7647e4b46ffe6a23cc4a62672c65288cebac29a5fd0de5ff4f01fd177cfb0a.jpg", + "image_caption": [ + "b) Densified SfM point cloud" + ], + "image_footnote": [], + "bbox": [ + 413, + 88, + 661, + 375 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/50e0c9f3d8d42436ee19e288476b00f646d7b940fa2509d7b181deced7d14e20.jpg", + "image_caption": [ + "c) 3D Gaussian Splats", + "Figure 3. Data annotation overview. Each scene in uCO3D is reconstructed in three different ways: a) per-frame cameras with sparse point cloud calculated by VGGSfM [62], b) semi-dense point cloud comprising triangulations of additional denser tracks from VGGSfM's tracker, c) 3D Gaussian Splat [29] reconstruction optimized separately for each scene." + ], + "image_footnote": [], + "bbox": [ + 666, + 97, + 903, + 372 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/335b00b451c85871155c25bb326249a4455b9d40dd1c807515e4153ae4468f97.jpg", + "image_caption": [ + "a) Input video" + ], + "image_footnote": [], + "bbox": [ + 112, + 457, + 460, + 489 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7b94ce4c01501416c8417140602cdde0b2a17cc8c0bd545c34466c70c2960190.jpg", + "image_caption": [ + "b) Sine-wave camera trajectory", + "Figure 4. Data collection example. For each video, the cameras follow a sine-wave trajectory to ensure good viewpoint coverage." + ], + "image_footnote": [], + "bbox": [ + 94, + 505, + 480, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared to older datasets like CO3Dv2 [47], uCO3D comes with many improvements. First, uCO3D is much larger and more diverse than CO3Dv2: it contains more than 1k different categories and more than 170k objects, compared to the 50 and 38k of CO3Dv2. While CO3Dv2's categories are taken from MS COCO [38], the categories in uCO3D are taken from the LVIS [21] taxonomy. Hence, we inherit the LVIS focus on covering the long-tail of the visual-category distribution. To simplify data analysis, we grouped the $1\\mathrm{k}+$ LVIS categories to 50 super-categories, each containing approximately 20 subcategories. Figure 2 shows the number of videos collected per super-category, and the LVIS category distribution.", + "bbox": [ + 89, + 669, + 482, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Second, uCO3D improves quality significantly compared to CO3Dv2, ensuring that videos are of high reso", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "lution, cover each object well, and that the 3D annotations are accurate. uCO3D also contains rich textual descriptions of each object, missing in other datasets, and useful to train large generative models. It also comes with additional 3D Gaussian Splat reconstructions of all objects, each rigidly aligned to a canonical object-centric reference, which make it possible to re-render the dataset from a fixed, canonical set of cameras, simulating synthetic data acquisition, which is very useful for training generative models [35, 55].", + "bbox": [ + 511, + 460, + 906, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dataset collection. Videos of objects were captured by workers on Amazon Mechanical Turk. To ensure high video quality, workers were required to submit videos of a sufficient resolution. As a result, more than $60\\%$ of videos in uCO3D are of 1080p resolution or higher, compared to $33\\%$ in CO3Dv2. Furthermore, to aid the 3D reconstruction, workers followed a sine-wave capture trajectory instead of the plain circular trajectory of CO3Dv2, ensuring varying camera elevations (cf. Fig. 4). Finally, each video was individually manually assessed to make sure that it adheres to these requirements, a process more rigorous than the rough eyeballing used in CO3Dv2 [47].", + "bbox": [ + 511, + 603, + 908, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video object segmentation. We used text-conditioned Segment-Anything (langSAM) [20, 32] to segment the object of interest in each video frame given text-conditioning in form of the object-category name, which had been provided by Turkers at collection time.", + "bbox": [ + 511, + 792, + 906, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To improve frame-to-frame consistency, CO3Dv2 used a simple Viterbi algorithm, which often led to segmentation", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14105", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2506e58f0b6465728af23e0f29d70188eb225e0b69f86010fd0a44e4e085c24f.jpg", + "image_caption": [ + "Figure 5. 3D reconstruction comparison. We show results of LightplaneLRM [6] models trained on MVIImgNet, CO3Dv2 and uCO3D." + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 890, + 486 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ff988240d2de12af08ea30d5bf6e666e3b9268cf1a2e1afe13bcc52d04e969e3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Train datasetOmniObject3DStanfordORB
LPIPS↓PSNR↑IoU↑LPIPS↓PSNR↑IoU↑
MVIImgNet [76]0.10923.390.9280.07024.4510.939
CO3Dv2 [47]0.09523.620.9260.05625.6170.956
uCO3D (ours)0.09324.610.9460.05725.7150.957
", + "bbox": [ + 91, + 532, + 480, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. 3D reconstruction benchmark. We compare LightplaneLRM [6] models trained on CO3Dv2, MVImgNet, and uCO3D. We report novel-view synthesis performances on OmniObject3D [68] and StanfordORB [34].", + "bbox": [ + 89, + 625, + 482, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "flickering, impairing the final 3D reconstruction quality. Instead, in uCO3D, we refine the SAM segmentations with state-of-the-art deep video-segmenter based on XMem [9], leading to more temporally-stable object segmentations.", + "bbox": [ + 89, + 710, + 482, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D annotation with VGGSfM. For each video, we use the state-of-the-art VGGSfM [62] Structure from Motion (SfM) system to estimate the parameters of the cameras (intrinsic and extrinsic) for 200 frames sampled uniformly. VGGSfM also outputs a sparse 3D point cloud, and its denser version obtained by triangulating additional 3D points from VGGSfM's tracker. Examples of sparse and densified SfM point clouds are shown in Fig. 3.", + "bbox": [ + 89, + 779, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Scene alignment. While the coordinate system of VG-GSfM cameras is defined only up to a rigid transformation, it is crucial for applications like generation and reconstruction to train on a dataset of rigidly aligned objects. We thus align all objects so they have a horizontal ground plane, similar scale, centring, and orientation. Details of the scene alignment procedure are in the supplementary material.", + "bbox": [ + 511, + 535, + 906, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gaussian Splat reconstruction. Sparse and even dense SfM point clouds provide an accurate but still quite sparse 3D reconstruction of the scene's surface. To further densify it, uCO3D provides a 3D Gaussian Splat reconstruction [29] for each scene, fitted using gsplat [75].", + "bbox": [ + 511, + 648, + 906, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Scene captioning. uCO3D also provides textual captions for all scenes, useful for generative modelling. Motivated by Cap3D [41], we first caption each view separately using a vision-language model, and then summarise these into a single scene caption using LLAMA3 [16].", + "bbox": [ + 511, + 732, + 908, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Applications", + "text_level": 1, + "bbox": [ + 511, + 828, + 645, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we demonstrate uCO3D's merit on three popular 3D learning tasks: feedforward sparse-view 3D reconstruction (Sec. 4.1), new-view synthesis using diffusion", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14106", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7fbbd9a4c29a229d341999b5d317e801372b7d676572ace261639cfc958e115f.jpg", + "image_caption": [ + "Figure 6. Novel-view synthesis comparison. We compare results of CAT3D-like [19] models trained on different datasets (MVIImgNet, CO3D, uCO3D) and evaluated on standard NVs datasets (top-to-bottom: RealEstate10K [79], LLFF [43], DTU [28], Mip-NeRF 360 [4])." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 898, + 467 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9f4c70c27f9387c7a3ca326ee1e7b84b4f9891d0e857220e6eec4963868d70dc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Train datasetEasierDataset DifficultyHarder
←Re10K [79]LLFF [43]DTU [28]Mip-NeRF [4]
LPIPS↓PSNR↑LPIPS↓PSNR↑LPIPS↓PSNR↑LPIPS↓PSNR↑
MVImgNet [76]0.31018.770.42614.380.37712.790.60512.39
CO3Dv2 [47]0.28120.020.41814.950.32916.420.53214.19
uCO3D (ours)0.27819.770.41815.160.31516.970.52814.37
", + "bbox": [ + 91, + 529, + 480, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Novel-view synthesis benchmark. We evaluate CAT3D-like [19] models trained on MVImgNet, CO3Dv2 or uCO3D. We report NVS performances on RealEstate10K [79], LLFF [43], DTU [28] and Mip-NeRF 360 [4].", + "bbox": [ + 89, + 623, + 482, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(Sec. 4.2), and text-to-3D (Sec. 4.3).", + "bbox": [ + 89, + 709, + 333, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Few-view 3D Object Reconstruction", + "text_level": 1, + "bbox": [ + 89, + 739, + 403, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Traditionally, multi-view 3D-annotated datasets such as CO3D or MVImgNet have been used to supervise few-view 3D reconstructors. In this section, we train LightplaneLRM [6], an evolution of the seminal LRM [27], and show that doing so on uCO3D leads to better performance than training on alternative datasets.", + "bbox": [ + 89, + 762, + 482, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LRM is a large transformer [61] that accepts few input images of an object and predicts a 3D representation of the latter. The transformer, conditioned on the tokens of the ob", + "bbox": [ + 89, + 854, + 483, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "served images via cross attention, converts a set of learnable input tokens to a 3D representation. The 3D representation is a triplane [7] supporting an opacity/radiance implicit shape. LightplaneLRM improves LRM by adding so called \"splattering layers\" and a memory-efficient renderer.", + "bbox": [ + 511, + 532, + 906, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During training, LightplaneLRM receives four random source frames from a training uCO3D video sequence, and renders the predicted triplane into held-out target views. Learning minimizes the photometric loss between the renders and the corresponding ground-truth targets. Both source and target views are masked using the extracted segmentation masks to make sure that LightplaneLRM only reconstructs the foreground object. Training uses the Adam optimizer and is warm-started following the original LRM training protocol by pre-training the model on a large dataset of synthetic objects similar to Objaverse [13].", + "bbox": [ + 511, + 609, + 908, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. Our main goal is to demonstrate that uCO3D contains higher quality data than existing object-centric datasets. As such, starting from the model pre-trained on the synthetic data, we finetune either on uCO3D, or on two other baseline datasets, namely MVImgNet and CO3Dv2.", + "bbox": [ + 511, + 785, + 908, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation protocol. We evaluate each trained model in a novel-view synthesis setting on two small-scale high-", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14107", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "quality object-centric datasets: OmniObject3D [68] and Stanford-ORB [34]. Given four views of a held-out test scene, the model reconstructs the scene which is then rendered to unseen target views. We report the average LPIPS [78] loss and Peak-signal-to-noise ratio (PSNR) between each render and the corresponding ground-truth image. We also report the intersection-over-union (IoU) between the rendered object alpha mask and the target view segmentation mask.", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. Table 2 and Fig. 5 report the quantitative and qualitative results, respectively. The LightplaneLRM trained on uCO3D is better than the other baselines in most metrics on both datasets. The latter confirms that uCO3D is currently the most reliable source of real data for training feedforward few-view 3D reconstructors.", + "bbox": [ + 89, + 238, + 483, + 328 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Novel-view synthesis using diffusion", + "text_level": 1, + "bbox": [ + 89, + 353, + 403, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We now consider application of uCO3D to training new-view image diffusion generators. These generators can, given one or a few views of an object and a target camera pose, output new arbitrary views as observed from the target camera, hallucinating missing details based on a statistical prior they learn. They can thus complement and integrate the feed-forward reconstruction models of the previous section, which are deterministic and thus unable to deal with ambiguity well. To this end, we train a diffusion model similar to the recent CAT3D [19], but reimplement it from scratch given lack of source code (see details in the supplementary). We call this model CAT3D-like.", + "bbox": [ + 89, + 380, + 483, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation protocol. As in Sec. 4.1, we compare versions of CAT3D-like trained using uCO3D, MVImgNet, and CO3Dv2 and test them on held-out datasets. A feature of CAT3D is the ability to reconstruct both the principal object in the images as well as the background. We thus benchmark the method using new-view synthesis datasets that do contain background, namely DTU [28], containing structured light scans of various objects, LLFF [43], containing scenes captured from fronto-parallel camera trajectories, RealEstate10k [79], containing real-estate walk-throughs, and Mip-NeRF 360 [3] with complex indoor and outdoor scenes. For evaluation, we take three known views as input and use the model to predict a new view. We report LPIPS and PSNR but not the IoU since CAT3D only generates new RGB views without reconstructing the 3D shape.", + "bbox": [ + 89, + 571, + 483, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. Table 3 and Fig. 6 contain the results: training CAT3D-like on uCO3D leads to the best performance across all four datasets. Even when compared to MVImgNet, which is slightly larger than uCO3D, the latter improves PSNR by 3-4 points, and reduces the LPIPS error by $5\\%$ to $20\\%$ .", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0288ee55c550563816bdacb45ca723a0fb48b07c37bc1972bd1fc319bb942999.jpg", + "image_caption": [ + "Figure 7. Supervising Instant3D with 3DGS. For each training scene, its 3DGS is rendered from 4 canonical views yielding a captioned image dataset for finetuning an image diffuser. Samples from the latter are then reconstructed with LRM." + ], + "image_footnote": [], + "bbox": [ + 514, + 89, + 906, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Photorealistic Text-to-3D", + "text_level": 1, + "bbox": [ + 513, + 455, + 741, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we show that uCO3D enables training photorealistic text-to-3D generators. Methods like CAT3D and others [39, 42, 53] generate several views of the object first, and then fit a 3D model, such as a NeRF or 3DGS, via optimization. This can work well, but it is not particularly robust or fast. An alternative, popularized by Instant3D [35] and follow-ups [54, 72], is to use a feedforward reconstructor in the second step, similar to LightplaneLRM from Sec. 4.1, which is faster and more robust. However, these models require canonical views of the objects — for example, Instant3D considers 4 orthogonal viewpoints, covering all 'sides'. The requirement of such training canonical views complicates training on real data, where viewpoints are arbitrary, and explains why such models are usually trained on synthetic data, limiting realism.", + "bbox": [ + 511, + 481, + 906, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Imaging 3DGS from canonical views. Our new idea is to 're-shoot' the 3DGS reconstructions provided with uCO3D from canonical viewpoints, making our data compatible with any method requiring canonical views for training (Fig. 7). To do so, we render the normalized reference frames (Sec. 3) into four views for each object, and arrange them in a grid as a target for the text-to-4-view generator. We double check the quality of the renders by calculating their CLIP similarity [24] to the object caption, and discard a sample if this is below 0.3.", + "bbox": [ + 511, + 715, + 906, + 867 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use this data to fine-tune a text-to-4-view image diffusion model using these 4-view image grids and the corre", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14108", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e2b0fa420a90fe166e3f20c3ef071f3b28905f69dddf06737320c964708c87d4.jpg", + "image_caption": [ + "Figure 8. Qualitative results for text-to-3D generation displaying the 4-view grids generated by our Instant3D-like model given the input caption, and the 3D asset obtained by reconstructing the latter. The 4-view grid generator was trained using the canonical 4-view renders of uCO3D's 3DGS scene reconstructions." + ], + "image_footnote": [], + "bbox": [ + 94, + 85, + 906, + 409 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5e45e4f957637f0c5999755a3825c467ca6ccaaefae7eb53ea4bb5c09c1ee729.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Train datasetReal - FID↓Surreal - FID↓
Synthetic82.842.3
uCO3D (ours)63.968.9
", + "bbox": [ + 155, + 483, + 419, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Text-to-3D evaluation. We compare Instant3D-like models trained on uCO3D or a dataset of synthetic renders from artist-created meshes. We report FID on two sets of data corresponding to real and surreal objects, see text for details.", + "bbox": [ + 89, + 547, + 482, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sponding scene captions. At inference time, given a caption describing the desired object, we use the model to sample a new 4-view grid and feed the latter, together with the corresponding cameras, to the LightplaneLRM model (Sec. 4.1) for 3D reconstruction.", + "bbox": [ + 89, + 630, + 482, + 704 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Baselines. We train another 4-view generator on a dataset of synthetic assets similar to Objaverse [13] and use it with the original LightplaneLRM model [6] trained on the same data and thus optimally matched to it.", + "bbox": [ + 89, + 712, + 482, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation protocol. We report metrics evaluating the alignment between the distributions of the generations and the ground-truth objects. Specifically, we report the Frechet Inception Distance (FID) [25] between the renders of the generated 3D shapes and images of ground-truth objects.", + "bbox": [ + 89, + 779, + 482, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The main purpose of this experiment is to show that, by training on the uCO3D dataset, the 3D generations are more realistic. We assess this using two sets of prompts: Sur-", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "real, containing 100 captions of objects from the synthetic dataset, and Real, containing 100 random captions form the held-out evaluation sequences of uCO3D. We report FID between the generated 3D shapes and the images/renders corresponding to the objects of each prompt-set.", + "bbox": [ + 511, + 487, + 906, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. Table 4 and Fig. 8 contain the quantitative and qualitative results respectively. The table reveals that the uCO3D-trained generator outperforms the synthetic generator when evaluated on real prompts. The latter verifies our hypothesis that a generator trained on uCO3D yields more realistic samples than a model trained on synthetic data.", + "bbox": [ + 511, + 569, + 906, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 679, + 638, + 695 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have introduced uCO3D, a new object-centric 3D dataset of real-life objects. uCO3D strikes a balance between size and quality, ensuring the quality of the collected turntable-like videos and of the 3D annotations, while at the same time significantly expanding the scale of the data compared to CO3Dv2 and the diversity compared to CO3Dv2 and MVImgNet. We have shown the benefits of using this dataset compared to alternatives when training models for feedforward few-view 3D reconstruction, multi-view generation, and text-to-3D generation. Equipped with 3D cameras, point clouds, masks, textual captions, and 3DGS reconstructions of objects, uCO3D is a ready-to-use resource for training large generative models and for exploring 3D deep learning.", + "bbox": [ + 511, + 705, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14109", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 118, + 187, + 133 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Aria digital twin catalog. https://www.projectaria.com/datasets/dtc/. Accessed: 2024-12-20.3", + "[2] Adel Ahmadyan, Liangkai Zhang, Artsiom Ablavatski, Jianing Wei, and Matthias Grundmann. Objectron: A large scale dataset of object-centric videos in the wild with pose annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7822-7831, 2021. 3", + "[3] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proc. ICCV, 2021. 7", + "[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In Proc. CVPR, 2022. 6", + "[5] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv.cs, abs/2311.15127, 2023. 2", + "[6] Ang Cao, Justin Johnson, Andrea Vedaldi, and David Novotny. Lightplane: Highly-scalable components for neural 3d fields. arXiv, 2024. 2, 3, 5, 6, 8, 1", + "[7] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In Proc. CVPR, 2022. 3, 6", + "[8] Angel X. Chang, Thomas A. Funkhouser, Leonidas J. Guibas, Pat Hanrahan, Qi-Xing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet an information-rich 3d model repository. arXiv.cs, abs/1512.03012, 2015. 3", + "[9] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In European Conference on Computer Vision, pages 640-658. Springer, 2022. 5", + "[10] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21126-21136, 2022. 3", + "[11] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam S. Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, Matthew Yu, Abhishek Kadian, Filip Radenovic, Dhruv Mahajan, Kunpeng Li, Yue Zhao, Vladan Petrovic, Mitesh Kumar Singh, Simran Motwani, Yi Wen, Yiwen Song, Roshan Sumbaly, Vignesh Ramanathan, Zijian He, Peter Vajda, and Devi Parikh. Emu:" + ], + "bbox": [ + 93, + 143, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Enhancing image generation models using photogenic needles in a haystack. CoRR, abs/2309.15807, 2023. 2, 3, 1", + "[12] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-XL: A universe of $10\\mathrm{M} + 3\\mathrm{D}$ objects. CoRR, abs/2307.05663, 2023. 3", + "[13] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In Proc. CVPR, 2023. 2, 3, 6, 8, 1", + "[14] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022.", + "[15] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B. McHugh, and Vincent Vanhoucke. Google Scanned Objects: A high-quality dataset of 3D scanned household items. In Proc. ICRA, 2022. 2, 3", + "[16] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aurélien Rodriguez, Austen Gregerson, Ava Spataru, Baptiste Rozière, Bethany Biron, Binh Tang, Bobbie Chern, Charlotte Caucheteux, Chaya Nayak, Chloe Bi, Chris Marra, Chris McConnell, Christian Keller, Christophe Touret, Chunyang Wu, Corinne Wong, Cristian Canton Ferrer, Cyrus Nikolaidis, Damien Allonsius, Daniel Song, Danielle Pintz, Danny Livshits, David Esiobu, Dhruv Choudhary, Dhruv Mahajan, Diego Garcia-Olano, Diego Perino, Dieuwke Hupkes, Egor Lakomkin, Ehab AlBadawy, Elina Lobanova, Emily Dinan, Eric Michael Smith, Filip Radenovic, Frank Zhang, Gabriel Synnaeve, Gabrielle Lee, Georgia Lewis Anderson, Graeme Nail, Grégoire Mialon, Guan Pang, Guillem Cucurell, Hailey Nguyen, Hannah Ko-revaar, Hu Xu, Hugo Touvron, Iliyan Zarov, Imanol Arrieta Ibarra, Isabel M. Kloumann, Ishan Misra, Ivan Evtimov, Jade Copet, Jaewon Lee, Jan Geffert, Jana Vranes, Jason Park, Jay Mahadeokar, Jeet Shah, Jelmer van der Linde, Jennifer Billock, Jenny Hong, Jenya Lee, Jeremy Fu, Jianfeng Chi, Jianyu Huang, Jiawen Liu, Jie Wang, Jiecao Yu, Joanna Bitton, Joe Spisak, Jongsoo Park, Joseph Rocca, Joshua Johnstun, Joshua Saxe, Junteng Jia, Kalyan Vasuden Alwala Kartikeya Upasani Kate Plawiak Ke Li Kenneth Heafield and Kevin Stone. The Llama 3 herd of models. arXiv 2407.21783, 2024. 5", + "[17] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d furniture shape with texture. International Journal of Computer Vision, 129:3313-3337, 2021. 3" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14110", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T. Barron, and Ben Poole. CAT3D: Create Anything in 3D with Multi-View Diffusion Models. arXiv.cs, 2024. 2, 3", + "[19] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T. Barron, and Ben Poole. CAT3D: create anything in 3d with multi-view diffusion models. arXiv, 2405.10314, 2024. 6, 7, 1", + "[20] Paul Guerrero. Lang-sam. https://github.com/paulguerrero/lang-sam, 2024.4", + "[21] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019. 4", + "[22] Agrim Gupta, Piotr Dólar, and Ross B. Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Proc. CVPR, 2019. 2", + "[23] Philipp Henzler, Jeremy Reizenstein, Patrick Labatut, Roman Shapovalov, Tobias Ritschel, Andrea Vedaldi, and David Novotny. Unsupervised learning of 3d object categories from videos in the wild. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[24] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 7", + "[25] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In Proc. NeurIPS, 2017. 8, 1", + "[26] Lukas Hollein, Aljaž Božić, Norman Müller, David Novotny, Hung-Yu Tseng, Christian Richardt, Michael Zollhöfer, and Matthias Nießner. ViewDiff: 3D-Consistent Image Generation with Text-to-Image Models. arXiv preprint, 2024. 3", + "[27] Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. In Proc. ICLR, 2024. 2, 3, 6", + "[28] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, pages 406-413, 2014. 3, 6, 7", + "[29] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian Splating for real-time radiance field rendering. Proc. SIGGRAPH, 42(4), 2023. 2, 4, 5", + "[30] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 1", + "[31] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. Proc. ICLR, 2015. 1", + "[32] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proc. CVPR, 2023. 4" + ], + "bbox": [ + 91, + 90, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. ABC: A big CAD model dataset for geometric deep learning. In Proc. CVPR, 2019. 3", + "[34] Zhengfei Kuang, Yunzhi Zhang, Hong-Xing Yu, Samir Agarwala, Elliott Wu, Jiajun Wu, et al. Stanford-orb: a real-world 3d object inverse rendering benchmark. In NeurIPS, 2023. 5, 7", + "[35] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3D: Fast text-to-3D with sparse-view generation and large reconstruction model. Proc. ICLR, 2024. 2, 3, 4, 7", + "[36] Joseph J Lim, Hamed Pirsiavash, and Antonio Torralba. Parsing intela objects: Fine pose estimation. In Proceedings of the IEEE international conference on computer vision, pages 2992-2999, 2013. 3", + "[37] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. arXiv.cs, abs/2305.08891, 2023. 1", + "[38] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In Proc. ECCV, 2014. 4", + "[39] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3D object. In Proc. ICCV, 2023. 7", + "[40] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-Solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In Proc. NeurIPS, 2022. 1", + "[41] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. arXiv preprint, 2023. 5", + "[42] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, Natalia Neverova, Andrea Vedaldi, Oran Gafni, and Filippos Kokkinos. IM-3D: Iterative multiview diffusion and reconstruction for high-quality 3D generation. In Proceedings of the International Conference on Machine Learning (ICML), 2024. 3, 7", + "[43] Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 2019. 6, 7", + "[44] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Carl Yuheng Ren. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception, 2023. 3", + "[45] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. In Proc. ICLR, 2023. 2", + "[46] Christian Reiser, Richard Szeliski, Dor Verbin, Pratul P. Srinivasan, Ben Mildenhall, Andreas Geiger, Jonathan T. Barron, and Peter Hedman. MERF: memory-efficient radiance fields for real-time view synthesis in unbounded scenes. abs/2302.12249, 2023. 1" + ], + "bbox": [ + 516, + 90, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14111", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common Objects in 3D: Large-scale learning and evaluation of real-life 3D category reconstruction. In Proc. ICCV, 2021. 2, 3, 4, 5, 6", + "[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proc. CVPR, 2022. 3, 1", + "[49] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In Proc. ICLR, 2022. 1", + "[50] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proc. CVPR, 2016. 2, 3", + "[51] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In Proc. ECCV, 2016. 3", + "[52] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2", + "[53] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. MVDream: Multi-view diffusion for 3D generation. In Proc. ICLR, 2024. 2, 3, 7, 1", + "[54] Yawar Siddiqui, Filippos Kokkinos, Tom Monnier, Mahendra Kariya, Yanir Kleiman, Emilien Garreau, Oran Gafni, Natalia Neverova, Andrea Vedaldi, Roman Shapovalov, and David Novotny. Meta 3D Asset Gen: Text-to-mesh generation with high-quality geometry, texture, and PBR materials. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2024. 2, 3, 7", + "[55] Yawar Siddiqui, Tom Monnier, Filippos Kokkinos, Mahendra Kariya, Yanir Kleiman, Emilien Garreau, Oran Gafni, Natalia Neverova, Andrea Vedaldi, Roman Shapovalov, et al. Meta 3d assetgen: Text-to-mesh generation with high-quality geometry, texture, and pbr materials. NeurIPS, 2024. 2, 3, 4", + "[56] Xingyuan Sun, Jiajun Wu, Xiuming Zhang, Zhoutong Zhang, Chengkai Zhang, Tianfan Xue, Joshua B Tenenbaum, and William T Freeman. Pix3d: Dataset and methods for single-image 3d shape modeling. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3", + "[57] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Viewset diffusion: (0-)image-conditioned 3D generative models from 2D data. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 2, 1", + "[58] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Splatter Image: Ultra-fast single-view 3D reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[59] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. LGM: Large multi-view Gaussian model for high-resolution 3D content creation. arXiv, 2402.05054, 2024. 3" + ], + "bbox": [ + 91, + 90, + 485, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[60] Shitao Tang, Jiacheng Chen, Dilin Wang, Chengzhou Tang, Fuyang Zhang, Yuchen Fan, Vikas Chandra, Yasutaka Furukawa, and Rakesh Ranjan. MVDiffusion++: A dense high-resolution multi-view diffusion model for single or sparse-view 3d object reconstruction. arXiv, 2402.12712, 2024. 3", + "[61] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. NeurIPS, 2017. 6", + "[62] Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. Vggsfm: Visual geometry grounded deep structure from motion. In CVPR, pages 21686-21697, 2024. 2, 4, 5", + "[63] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P. Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas A. Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proc. CVPR, 2021. 2", + "[64] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. DUSt3R: Geometric 3D vision made easy. In Proc. CVPR, 2024. 2", + "[65] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. MeshLRM: large reconstruction model for high-quality mesh. arXiv, 2404.12385, 2024. 3", + "[66] Rundi Wu, Chang Xiao, and Changxi Zheng. Deepcad: A deep generative network for computer-aided design models. in 2021 IEEE. In CVF International Conference on Computer Vision (ICCV), pages 6772-6782, 2021. 3", + "[67] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P. Srinivasan, Dor Verbin, Jonathan T. Barron, Ben Poole, and Aleksander Holynski. ReconFusion: 3D Reconstruction with Diffusion Priors. arXiv preprint, 2023. 2, 1", + "[68] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, et al. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 803-814, 2023. 2, 3, 5, 7", + "[69] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguuang Zhang, Xiaou Tang, and Jianxiong Xiao. 3D ShapeNets: A deep representation for volumetric shapes. In Proc. CVPR, 2015. 3", + "[70] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond PASCAL: A benchmark for 3D object detection in the wild. In Proc. WACV, 2014. 3", + "[71] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. InstantMesh: efficient 3D mesh generation from a single image with sparse-view large reconstruction models. arXiv, 2404.07191, 2024. 3", + "[72] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation. arXiv preprint arXiv:2403.14621, 2024. 7" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14112", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[73] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. GRM: Large gaussian reconstruction model for efficient 3D reconstruction and generation. arXiv, 2403.14621, 2024. 3", + "[74] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proc. CVPR, 2024. 2", + "[75] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, and Angjoo Kanazawa. gsplat: An open-source library for Gaussian splatting. arXiv preprint arXiv:2409.06765, 2024. 5", + "[76] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, et al. Mvmgnet: A large-scale dataset of multi-view images. In CVPR, pages 9150-9161, 2023. 2, 3, 5, 6", + "[77] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. GS-LRM: large reconstruction model for 3D Gaussian splatting. arXiv, 2404.19702, 2024. 3", + "[78] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proc. CVPR, pages 586-595, 2018. 7", + "[79] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snively. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 6, 7", + "[80] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets Gaussian splatting: Fast and generalizable single-view 3D reconstruction with transformers. arXiv.cs, abs/2312.09147, 2023. 3" + ], + "bbox": [ + 91, + 92, + 480, + 598 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "14113", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/UnCommon Objects in 3D/4d2bd976-41fd-45e5-bfc6-8a4578f73725_model.json b/2025/UnCommon Objects in 3D/4d2bd976-41fd-45e5-bfc6-8a4578f73725_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cbcb6bb27bdfda38ade682411d2616d8042a6391 --- /dev/null +++ b/2025/UnCommon Objects in 3D/4d2bd976-41fd-45e5-bfc6-8a4578f73725_model.json @@ -0,0 +1,2248 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.364, + 0.131, + 0.635, + 0.154 + ], + "angle": 0, + "content": "UnCommon Objects in 3D" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.18, + 0.851, + 0.245 + ], + "angle": 0, + "content": "Xingchen Liu\\(^{\\infty}\\) Piyush Tayal\\(^{\\infty}\\) Jianyuan Wang\\(^{\\infty}\\) Jesus Zarzar\\(^{+\\infty}\\) Tom Monnier\\(^{\\infty}\\) Konstantinos Tertikas\\(^{\\infty*}\\) Jiali Duan\\(^{\\infty}\\) Antoine Toisoul\\(^{\\infty}\\) Jason Y. Zhang\\(^{\\dagger}\\) Natalia Neverova\\(^{\\infty}\\) Andrea Vedaldi\\(^{\\infty}\\) Roman Shapovalov\\(^{\\infty}\\) David Novotny\\(^{\\infty}\\)" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.264, + 0.778, + 0.283 + ], + "angle": 0, + "content": "*NKUA, Greece †KAUST †Carnegie Mellon University †Meta AI" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.304, + 0.887, + 0.324 + ], + "angle": 0, + "content": "uCO3D https://uco3d.github.io https://github.com/facebookresearch/uco3d" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.349, + 0.907, + 0.701 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.706, + 0.908, + 0.737 + ], + "angle": 0, + "content": "Figure 1. We introduce UnCommon Objects in 3D (uCO3D), a large and diverse dataset of high-quality \\(360^{\\circ}\\) videos covering over 1,000 object categories. Each video frame is 3D-annotated with accurate SfM cameras, point cloud, and a 3D Gaussian Splitting reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.755, + 0.329, + 0.772 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.486, + 0.903 + ], + "angle": 0, + "content": "We introduce Uncommon Objects in 3D (uCO3D), a new object-centric dataset for 3D deep learning and 3D generative AI. uCO3D is the largest publicly-available collection of high-resolution videos of objects with 3D annotations that ensures full- \\(360^{\\circ}\\) coverage. uCO3D is significantly more diverse than MVImgNet and CO3Dv2, covering more than 1,000 object categories. It is also of higher quality, due" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.756, + 0.909, + 0.879 + ], + "angle": 0, + "content": "to extensive quality checks of both the collected videos and the 3D annotations. Similar to analogous datasets, uCO3D contains annotations for 3D camera poses, depth maps and sparse point clouds. In addition, each object is equipped with a caption and a 3D Gaussian Splat reconstruction. We train several large 3D models on MVImgNet, CO3Dv2, and uCO3D and obtain superior results using the latter, showing that uCO3D is better for learning applications." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "14102" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.223, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.483, + 0.221 + ], + "angle": 0, + "content": "The primacy of data has been the defining characteristic of the last decade of machine learning, alongside deep learning. The most powerful models in language, speech, and computer vision are large deep networks trained on massive amounts of data, and then further fine-tuned on its high-quality subset. This paradigm is expected to extend to all machine learning applications, including 3D vision." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.222, + 0.483, + 0.252 + ], + "angle": 0, + "content": "However, 3D training data is much harder to come by than data for text, audio, and image processing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.252, + 0.483, + 0.478 + ], + "angle": 0, + "content": "Seeking training data for large 3D neural networks [27], many have turned to synthetic datasets like Objaverse [13]. However, synthetic data is a poor substitute for real data in applications like digital twinning, which creates 3D models of real-life objects. This is why many photorealistic reconstruction networks [6, 23, 57, 58, 63, 67] are trained using real object-centric datasets like CO3D [47], MVImgNet [76], GSO [15], and OmniObject3D [68]. Real data is also crucial for generalization, as demonstrated by DUSt3r [64] for point map prediction and DepthAnything [74] for depth prediction, both of which are trained on numerous real datasets. Even non-curated image datasets like the billion-scale LAION [52] are applicable to 3D vision. For instance, text-to-3D generators [35, 45, 53, 54] build on LAION-trained large text-to-image models [5, 11]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.483, + 0.63 + ], + "angle": 0, + "content": "Given the importance of 3D datasets, but also their relative scarcity, in this paper we ask what is the next step for real data in 3D vision. To answer this question, we note that, while the size of a dataset is crucial, in most cases its quality is just as important. For example, text-to-3D models [35, 55] are notoriously sensitive to fine-tuning data quality, training only with the best-looking models (e.g., Instant3D [35] uses only about \\(1\\%\\) of Objaverse). We conclude that high-quality datasets balancing scale and fidelity are more valuable than simply amassing low-quality data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.483, + 0.795 + ], + "angle": 0, + "content": "Based on this observation, we argue that there is a gap in the real object-centric 3D datasets that are currently available, as none strikes the optimal balance between quality and scale. For example, the 3D object scans in OmniObject3D [68] and GSO [15] provide very accurate geometry and textures, but only count a few thousand objects. Conversely, datasets like CO3D [47] and MvImgNet [76] contain orders-of-magnitude more objects, but lack reliable 3D scans. Instead, they provide many views of the objects together with lower-quality 3D cameras and point clouds reconstructed with structure-from-motion (SfM)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this paper, we address this gap with a new dataset, Uncommon Objects in 3D (uCO3D), which better balances data quality and size (Tab. 1). Similar to CO3D, it comprises full- \\(360^{\\circ}\\) crowd-sourced videos capturing objects from all sides, annotated with cameras and point clouds using SfM. Furthermore, uCO3D has much greater data diversity (Fig. 2) than prior alternatives as it contains objects" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.09, + 0.905, + 0.22 + ], + "angle": 0, + "content": "
RealCount# ClassesData typeAnnotations
ShapeNet [8]X51k553D meshesmesh
Objaverse [13]X800k21k3D meshesmesh
Objaverse-XL [12]X10M2M3D meshesmesh
ABO [10]X8k633D meshesmesh
OmniObject3D [68]✓6k190Videos w/ meshescameras, mesh
GSO [14]✓1k17Views w/ meshescameras, mesh
Objectron [2]✓15k9Limited vp. videoscameras, 3D box
MVImgNet [76]✓220k238Limited vp. videoscameras, pcl
CO3D [47]✓19k50360° videoscameras, pcl
CO3Dv2 [47]✓40k50360° videoscameras, pcl
uCO3D (ours)✓170k1k360° videoscameras, 3DGS, caption
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.234, + 0.905, + 0.263 + ], + "angle": 0, + "content": "Table 1. Overview of 3D object datasets. We compare the number of objects / classes, the type of data and associated annotations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.3, + 0.907, + 0.437 + ], + "angle": 0, + "content": "from the 1,070 visual object categories of the LVIS [22] taxonomy, which has long tails. For reference, MVImgNet and CO3Dv2 contain only 238 and 50 categories, respectively. These fine-grained categories are organized in supercategories, also shown in Fig. 2. Furthermore, uCO3D contains 170k scenes, which is more than four times larger than CO3Dv2's 40k. While this is less than MVImgNet's 220k, uCO3D's videos cover each object from all sides, as opposed to MVImgNet's partial object captures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.442, + 0.909, + 0.73 + ], + "angle": 0, + "content": "Besides improving size and diversity, uCO3D also raises the quality bar. This was achieved by checking extensively both the collected videos and their 3D annotations. Differently from datasets like CO3Dv2 that still contain a certain portion of low-quality videos, in uCO3D we manually verified that each video provides full \\(360^{\\circ}\\) turn-table covering all sides of the object. Additionally, \\(60\\%+\\) of the videos have \\(1080p+\\) resolution, higher than CO3Dv2. To ensure 3D-annotation quality, we improved both the reconstruction algorithm and the reconstruction validation. For camera reconstruction, we used VGGSfM [62], which is currently the best SfM system available, and is more robust and accurate than COLMAP [50], used in CO3Dv2 and MvImgNet. We also improve on CO3Dv2's active-learning camera quality evaluation by combining it with novel-view synthesis accuracy after reconstructing each scene using 3D Gaussian Splatting (3DGS) [29]. The latter also guarantees that scenes are reconstructible to a high quality, which is important for training of 3D models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We validate uCO3D's benefits in applications. We train two popular 3D models, LRM [27] and CAT3D [18], using uCO3D and demonstrate improved results compared to training on MVImgNet and CO3Dv2, which makes uCO3D the better data source for real object-centric 3D learning. We also use uCO3D to train a text-to-3D model following Instant3D's [35] two-stage design. The latter requires objects to be rendered from canonical viewpoints, and thus, so far, was limited to synthetic data. By using our 3DGS reconstructions, we 're-shoot' uCO3D's from these viewpoints, which allows to train a more realistic generator." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14103" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.905, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.235, + 0.905, + 0.266 + ], + "angle": 0, + "content": "Figure 2. Statistics of uCO3D. (Left) We plot the number of objects per super-category. In total, the dataset contains 50 super-categories, each gathering around 20 sub-categories. (Right) We show a word cloud of all 1,070 visual categories represented in the dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.289, + 0.233, + 0.305 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.317, + 0.485, + 0.573 + ], + "angle": 0, + "content": "Datasets of synthetic 3D objects. Historically, object-centric 3D datasets have predominantly been synthetic, composed of artist-generated 3D models. A prominent example is ShapeNet [8], with 51,000 meshes across 55 object categories. The meshes have detailed geometries, but relatively simplistic textures. Datasets such as 3DFUTURE [17], IKEA [36], Pix3D [56], and ABO [10] are less diverse, concentrating on furniture and other consumer goods. In contrast, ModelNet [69], DeepCAD [66], and ABC [33] provide CAD models with clean geometry but no texture. Objaverse [13] is perhaps the most impactful dataset following ShapeNet. It is significantly larger, comprising 800,000 3D meshes. Objaverse-XL [12] expands this collection to 10M objects. These datasets have been pivotal in enabling the first 3D deep generative models, including text-to-3D [35, 53, 54] and image-to-3D [6, 27, 73] models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.718 + ], + "angle": 0, + "content": "Real 3D object datasets. Acquiring real-world 3D data presents many challenges, resulting in only a few real 3D-object datasets. Early datasets like Pascal3D [70] include multiple categories but provide only isolated object views and approximate 3D annotations. Conversely, DTU [28], BlendedMVS [51], GSO [15], OmniObject3D [68], Aria Digital Twin [44], and Digital Twin Catalog [1] provide 3D scans of objects, featuring high-quality 3D geometry and textures, but have only a few thousand objects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The use of 3D scanners significantly restricts the scale of data acquisition; consequently, other datasets capture multiview turntable-like videos of objects using consumer cameras. CO3D and CO3Dv2 [47] crowd-sourced \\(40,000360^{\\circ}\\) object videos, providing 3D annotations by reconstructing point clouds and cameras using COLMAP SfM [50]. MvImgNet [76] collected even more videos (220,000) across more object categories (238), but their videos capture objects only partially, preventing full reconstruction. Objectron [2] is similar to MvImgNet, but with fewer videos (10,000). A common challenge is that large-scale datasets often rely on SfM for video reconstruction, which can" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.29, + 0.907, + 0.397 + ], + "angle": 0, + "content": "lead to imprecise 3D annotations. uCO3D also employs SfM, but using VGGSfM, which has greater accuracy than COLMAP, and with a more reliable data validation setup. Furthermore, uCO3D is five times larger and significantly more diverse than CO3Dv2, encompassing 20 times more visual categories, and provides caption and 3D Gaussian Splat reconstructions of each object." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.906, + 0.54 + ], + "angle": 0, + "content": "Applications. In order to assess the quality of uCO3D, we measure how it benefits a number of popular downstream applications. First, we consider feedforward few-view 3D reconstruction models. Among those, LRM [27] is a transformer that maps an input image to a neural radiance field supported by a triplane [7]. LightplaneLRM [6] adds splatting layers and a memory-efficient renderer. Further extensions use different representations like 3D Gaussian Splats [59, 73, 77, 80] and meshes [65, 71]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.54, + 0.907, + 0.798 + ], + "angle": 0, + "content": "We also consider text-to-3D generators, which create 3D assets from text, and focus on the two-stages approach of Instant3D [35]. This is based on training a text-to-multi-view diffusion model [11, 48] which generates several 2D views of the object, followed by a 3D reconstruction network that outputs the 3D asset, all in a matter of seconds. The multi-view diffusion is improved in ViewDiff [26], MVDiffusion [60], IM-3D [42], CAT3D [18] and many others. AssetGen [55] further extends Instant3D by modelling material properties instead of baking in the radiance function and adds a texture refiner that outputs relightable PBR textures. As an illustration, we use uCO3D to train a model like CAT3D, which results in better new-view synthesis than the one trained on alternative datasets. We also show that the Gaussian Splat reconstructions provided with uCO3D can supervise, for the first time, an Instant3D-like pipeline using solely real-life data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.814, + 0.757, + 0.831 + ], + "angle": 0, + "content": "3. Uncommon Objects in 3D" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.903 + ], + "angle": 0, + "content": "In this section, we introduce uCO3D, our new dataset of real-life 3D objects. uCO3D comprises \\(360^{\\circ}\\) turn-table-like videos of objects, crowdsourced and annotated with 3D cameras, point clouds, 3D Gaussians, and textual captions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14104" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.147, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.377, + 0.136, + 0.387 + ], + "angle": 0, + "content": "Video" + }, + { + "type": "image", + "bbox": [ + 0.149, + 0.09, + 0.413, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.209, + 0.376, + 0.353, + 0.387 + ], + "angle": 0, + "content": "a) Sparse SfM reconstruction" + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.089, + 0.663, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.466, + 0.376, + 0.611, + 0.387 + ], + "angle": 0, + "content": "b) Densified SfM point cloud" + }, + { + "type": "image", + "bbox": [ + 0.667, + 0.098, + 0.904, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.728, + 0.375, + 0.837, + 0.387 + ], + "angle": 0, + "content": "c) 3D Gaussian Splats" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.392, + 0.908, + 0.435 + ], + "angle": 0, + "content": "Figure 3. Data annotation overview. Each scene in uCO3D is reconstructed in three different ways: a) per-frame cameras with sparse point cloud calculated by VGGSfM [62], b) semi-dense point cloud comprising triangulations of additional denser tracks from VGGSfM's tracker, c) 3D Gaussian Splat [29] reconstruction optimized separately for each scene." + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.458, + 0.462, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.491, + 0.331, + 0.503 + ], + "angle": 0, + "content": "a) Input video" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.506, + 0.482, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.195, + 0.591, + 0.381, + 0.604 + ], + "angle": 0, + "content": "b) Sine-wave camera trajectory" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.607, + 0.483, + 0.636 + ], + "angle": 0, + "content": "Figure 4. Data collection example. For each video, the cameras follow a sine-wave trajectory to ensure good viewpoint coverage." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.67, + 0.483, + 0.867 + ], + "angle": 0, + "content": "Compared to older datasets like CO3Dv2 [47], uCO3D comes with many improvements. First, uCO3D is much larger and more diverse than CO3Dv2: it contains more than 1k different categories and more than 170k objects, compared to the 50 and 38k of CO3Dv2. While CO3Dv2's categories are taken from MS COCO [38], the categories in uCO3D are taken from the LVIS [21] taxonomy. Hence, we inherit the LVIS focus on covering the long-tail of the visual-category distribution. To simplify data analysis, we grouped the \\(1\\mathrm{k}+\\) LVIS categories to 50 super-categories, each containing approximately 20 subcategories. Figure 2 shows the number of videos collected per super-category, and the LVIS category distribution." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Second, uCO3D improves quality significantly compared to CO3Dv2, ensuring that videos are of high reso" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.907, + 0.598 + ], + "angle": 0, + "content": "lution, cover each object well, and that the 3D annotations are accurate. uCO3D also contains rich textual descriptions of each object, missing in other datasets, and useful to train large generative models. It also comes with additional 3D Gaussian Splat reconstructions of all objects, each rigidly aligned to a canonical object-centric reference, which make it possible to re-render the dataset from a fixed, canonical set of cameras, simulating synthetic data acquisition, which is very useful for training generative models [35, 55]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.604, + 0.909, + 0.788 + ], + "angle": 0, + "content": "Dataset collection. Videos of objects were captured by workers on Amazon Mechanical Turk. To ensure high video quality, workers were required to submit videos of a sufficient resolution. As a result, more than \\(60\\%\\) of videos in uCO3D are of 1080p resolution or higher, compared to \\(33\\%\\) in CO3Dv2. Furthermore, to aid the 3D reconstruction, workers followed a sine-wave capture trajectory instead of the plain circular trajectory of CO3Dv2, ensuring varying camera elevations (cf. Fig. 4). Finally, each video was individually manually assessed to make sure that it adheres to these requirements, a process more rigorous than the rough eyeballing used in CO3Dv2 [47]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.793, + 0.908, + 0.869 + ], + "angle": 0, + "content": "Video object segmentation. We used text-conditioned Segment-Anything (langSAM) [20, 32] to segment the object of interest in each video frame given text-conditioning in form of the object-category name, which had been provided by Turkers at collection time." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "To improve frame-to-frame consistency, CO3Dv2 used a simple Viterbi algorithm, which often led to segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14105" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.089, + 0.891, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.494, + 0.904, + 0.51 + ], + "angle": 0, + "content": "Figure 5. 3D reconstruction comparison. We show results of LightplaneLRM [6] models trained on MVIImgNet, CO3Dv2 and uCO3D." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.533, + 0.481, + 0.616 + ], + "angle": 0, + "content": "
Train datasetOmniObject3DStanfordORB
LPIPS↓PSNR↑IoU↑LPIPS↓PSNR↑IoU↑
MVIImgNet [76]0.10923.390.9280.07024.4510.939
CO3Dv2 [47]0.09523.620.9260.05625.6170.956
uCO3D (ours)0.09324.610.9460.05725.7150.957
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.626, + 0.483, + 0.682 + ], + "angle": 0, + "content": "Table 2. 3D reconstruction benchmark. We compare LightplaneLRM [6] models trained on CO3Dv2, MVImgNet, and uCO3D. We report novel-view synthesis performances on OmniObject3D [68] and StanfordORB [34]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.712, + 0.483, + 0.773 + ], + "angle": 0, + "content": "flickering, impairing the final 3D reconstruction quality. Instead, in uCO3D, we refine the SAM segmentations with state-of-the-art deep video-segmenter based on XMem [9], leading to more temporally-stable object segmentations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.903 + ], + "angle": 0, + "content": "3D annotation with VGGSfM. For each video, we use the state-of-the-art VGGSfM [62] Structure from Motion (SfM) system to estimate the parameters of the cameras (intrinsic and extrinsic) for 200 frames sampled uniformly. VGGSfM also outputs a sparse 3D point cloud, and its denser version obtained by triangulating additional 3D points from VGGSfM's tracker. Examples of sparse and densified SfM point clouds are shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.536, + 0.907, + 0.644 + ], + "angle": 0, + "content": "Scene alignment. While the coordinate system of VG-GSfM cameras is defined only up to a rigid transformation, it is crucial for applications like generation and reconstruction to train on a dataset of rigidly aligned objects. We thus align all objects so they have a horizontal ground plane, similar scale, centring, and orientation. Details of the scene alignment procedure are in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.65, + 0.907, + 0.727 + ], + "angle": 0, + "content": "Gaussian Splat reconstruction. Sparse and even dense SfM point clouds provide an accurate but still quite sparse 3D reconstruction of the scene's surface. To further densify it, uCO3D provides a 3D Gaussian Splat reconstruction [29] for each scene, fitted using gsplat [75]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.909, + 0.81 + ], + "angle": 0, + "content": "Scene captioning. uCO3D also provides textual captions for all scenes, useful for generative modelling. Motivated by Cap3D [41], we first caption each view separately using a vision-language model, and then summarise these into a single scene caption using LLAMA3 [16]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.829, + 0.646, + 0.846 + ], + "angle": 0, + "content": "4. Applications" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "In this section, we demonstrate uCO3D's merit on three popular 3D learning tasks: feedforward sparse-view 3D reconstruction (Sec. 4.1), new-view synthesis using diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14106" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.089, + 0.899, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.477, + 0.908, + 0.507 + ], + "angle": 0, + "content": "Figure 6. Novel-view synthesis comparison. We compare results of CAT3D-like [19] models trained on different datasets (MVIImgNet, CO3D, uCO3D) and evaluated on standard NVs datasets (top-to-bottom: RealEstate10K [79], LLFF [43], DTU [28], Mip-NeRF 360 [4])." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.53, + 0.481, + 0.616 + ], + "angle": 0, + "content": "
Train datasetEasierDataset DifficultyHarder
←Re10K [79]LLFF [43]DTU [28]Mip-NeRF [4]
LPIPS↓PSNR↑LPIPS↓PSNR↑LPIPS↓PSNR↑LPIPS↓PSNR↑
MVImgNet [76]0.31018.770.42614.380.37712.790.60512.39
CO3Dv2 [47]0.28120.020.41814.950.32916.420.53214.19
uCO3D (ours)0.27819.770.41815.160.31516.970.52814.37
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.624, + 0.483, + 0.682 + ], + "angle": 0, + "content": "Table 3. Novel-view synthesis benchmark. We evaluate CAT3D-like [19] models trained on MVImgNet, CO3Dv2 or uCO3D. We report NVS performances on RealEstate10K [79], LLFF [43], DTU [28] and Mip-NeRF 360 [4]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.71, + 0.334, + 0.726 + ], + "angle": 0, + "content": "(Sec. 4.2), and text-to-3D (Sec. 4.3)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.74, + 0.405, + 0.756 + ], + "angle": 0, + "content": "4.1. Few-view 3D Object Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.763, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Traditionally, multi-view 3D-annotated datasets such as CO3D or MVImgNet have been used to supervise few-view 3D reconstructors. In this section, we train LightplaneLRM [6], an evolution of the seminal LRM [27], and show that doing so on uCO3D leads to better performance than training on alternative datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.9 + ], + "angle": 0, + "content": "LRM is a large transformer [61] that accepts few input images of an object and predicts a 3D representation of the latter. The transformer, conditioned on the tokens of the ob" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.907, + 0.608 + ], + "angle": 0, + "content": "served images via cross attention, converts a set of learnable input tokens to a 3D representation. The 3D representation is a triplane [7] supporting an opacity/radiance implicit shape. LightplaneLRM improves LRM by adding so called \"splattering layers\" and a memory-efficient renderer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.909, + 0.779 + ], + "angle": 0, + "content": "During training, LightplaneLRM receives four random source frames from a training uCO3D video sequence, and renders the predicted triplane into held-out target views. Learning minimizes the photometric loss between the renders and the corresponding ground-truth targets. Both source and target views are masked using the extracted segmentation masks to make sure that LightplaneLRM only reconstructs the foreground object. Training uses the Adam optimizer and is warm-started following the original LRM training protocol by pre-training the model on a large dataset of synthetic objects similar to Objaverse [13]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.786, + 0.909, + 0.862 + ], + "angle": 0, + "content": "Baselines. Our main goal is to demonstrate that uCO3D contains higher quality data than existing object-centric datasets. As such, starting from the model pre-trained on the synthetic data, we finetune either on uCO3D, or on two other baseline datasets, namely MVImgNet and CO3Dv2." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Evaluation protocol. We evaluate each trained model in a novel-view synthesis setting on two small-scale high-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14107" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.228 + ], + "angle": 0, + "content": "quality object-centric datasets: OmniObject3D [68] and Stanford-ORB [34]. Given four views of a held-out test scene, the model reconstructs the scene which is then rendered to unseen target views. We report the average LPIPS [78] loss and Peak-signal-to-noise ratio (PSNR) between each render and the corresponding ground-truth image. We also report the intersection-over-union (IoU) between the rendered object alpha mask and the target view segmentation mask." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.239, + 0.484, + 0.329 + ], + "angle": 0, + "content": "Results. Table 2 and Fig. 5 report the quantitative and qualitative results, respectively. The LightplaneLRM trained on uCO3D is better than the other baselines in most metrics on both datasets. The latter confirms that uCO3D is currently the most reliable source of real data for training feedforward few-view 3D reconstructors." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.354, + 0.405, + 0.37 + ], + "angle": 0, + "content": "4.2. Novel-view synthesis using diffusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.381, + 0.484, + 0.562 + ], + "angle": 0, + "content": "We now consider application of uCO3D to training new-view image diffusion generators. These generators can, given one or a few views of an object and a target camera pose, output new arbitrary views as observed from the target camera, hallucinating missing details based on a statistical prior they learn. They can thus complement and integrate the feed-forward reconstruction models of the previous section, which are deterministic and thus unable to deal with ambiguity well. To this end, we train a diffusion model similar to the recent CAT3D [19], but reimplement it from scratch given lack of source code (see details in the supplementary). We call this model CAT3D-like." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.573, + 0.484, + 0.801 + ], + "angle": 0, + "content": "Evaluation protocol. As in Sec. 4.1, we compare versions of CAT3D-like trained using uCO3D, MVImgNet, and CO3Dv2 and test them on held-out datasets. A feature of CAT3D is the ability to reconstruct both the principal object in the images as well as the background. We thus benchmark the method using new-view synthesis datasets that do contain background, namely DTU [28], containing structured light scans of various objects, LLFF [43], containing scenes captured from fronto-parallel camera trajectories, RealEstate10k [79], containing real-estate walk-throughs, and Mip-NeRF 360 [3] with complex indoor and outdoor scenes. For evaluation, we take three known views as input and use the model to predict a new view. We report LPIPS and PSNR but not the IoU since CAT3D only generates new RGB views without reconstructing the 3D shape." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Results. Table 3 and Fig. 6 contain the results: training CAT3D-like on uCO3D leads to the best performance across all four datasets. Even when compared to MVImgNet, which is slightly larger than uCO3D, the latter improves PSNR by 3-4 points, and reduces the LPIPS error by \\(5\\%\\) to \\(20\\%\\)." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.09, + 0.907, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.371, + 0.907, + 0.426 + ], + "angle": 0, + "content": "Figure 7. Supervising Instant3D with 3DGS. For each training scene, its 3DGS is rendered from 4 canonical views yielding a captioned image dataset for finetuning an image diffuser. Samples from the latter are then reconstructed with LRM." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.457, + 0.743, + 0.472 + ], + "angle": 0, + "content": "4.3. Photorealistic Text-to-3D" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.482, + 0.907, + 0.709 + ], + "angle": 0, + "content": "Next, we show that uCO3D enables training photorealistic text-to-3D generators. Methods like CAT3D and others [39, 42, 53] generate several views of the object first, and then fit a 3D model, such as a NeRF or 3DGS, via optimization. This can work well, but it is not particularly robust or fast. An alternative, popularized by Instant3D [35] and follow-ups [54, 72], is to use a feedforward reconstructor in the second step, similar to LightplaneLRM from Sec. 4.1, which is faster and more robust. However, these models require canonical views of the objects — for example, Instant3D considers 4 orthogonal viewpoints, covering all 'sides'. The requirement of such training canonical views complicates training on real data, where viewpoints are arbitrary, and explains why such models are usually trained on synthetic data, limiting realism." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.717, + 0.907, + 0.868 + ], + "angle": 0, + "content": "Imaging 3DGS from canonical views. Our new idea is to 're-shoot' the 3DGS reconstructions provided with uCO3D from canonical viewpoints, making our data compatible with any method requiring canonical views for training (Fig. 7). To do so, we render the normalized reference frames (Sec. 3) into four views for each object, and arrange them in a grid as a target for the text-to-4-view generator. We double check the quality of the renders by calculating their CLIP similarity [24] to the object caption, and discard a sample if this is below 0.3." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "We use this data to fine-tune a text-to-4-view image diffusion model using these 4-view image grids and the corre" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14108" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.087, + 0.907, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.419, + 0.908, + 0.461 + ], + "angle": 0, + "content": "Figure 8. Qualitative results for text-to-3D generation displaying the 4-view grids generated by our Instant3D-like model given the input caption, and the 3D asset obtained by reconstructing the latter. The 4-view grid generator was trained using the canonical 4-view renders of uCO3D's 3DGS scene reconstructions." + }, + { + "type": "table", + "bbox": [ + 0.156, + 0.484, + 0.42, + 0.538 + ], + "angle": 0, + "content": "
Train datasetReal - FID↓Surreal - FID↓
Synthetic82.842.3
uCO3D (ours)63.968.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.548, + 0.483, + 0.604 + ], + "angle": 0, + "content": "Table 4. Text-to-3D evaluation. We compare Instant3D-like models trained on uCO3D or a dataset of synthetic renders from artist-created meshes. We report FID on two sets of data corresponding to real and surreal objects, see text for details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.631, + 0.483, + 0.705 + ], + "angle": 0, + "content": "sponding scene captions. At inference time, given a caption describing the desired object, we use the model to sample a new 4-view grid and feed the latter, together with the corresponding cameras, to the LightplaneLRM model (Sec. 4.1) for 3D reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.713, + 0.483, + 0.773 + ], + "angle": 0, + "content": "Baselines. We train another 4-view generator on a dataset of synthetic assets similar to Objaverse [13] and use it with the original LightplaneLRM model [6] trained on the same data and thus optimally matched to it." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Evaluation protocol. We report metrics evaluating the alignment between the distributions of the generations and the ground-truth objects. Specifically, we report the Frechet Inception Distance (FID) [25] between the renders of the generated 3D shapes and images of ground-truth objects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The main purpose of this experiment is to show that, by training on the uCO3D dataset, the 3D generations are more realistic. We assess this using two sets of prompts: Sur-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.907, + 0.563 + ], + "angle": 0, + "content": "real, containing 100 captions of objects from the synthetic dataset, and Real, containing 100 random captions form the held-out evaluation sequences of uCO3D. We report FID between the generated 3D shapes and the images/renders corresponding to the objects of each prompt-set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.907, + 0.661 + ], + "angle": 0, + "content": "Results. Table 4 and Fig. 8 contain the quantitative and qualitative results respectively. The table reveals that the uCO3D-trained generator outperforms the synthetic generator when evaluated on real prompts. The latter verifies our hypothesis that a generator trained on uCO3D yields more realistic samples than a model trained on synthetic data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.64, + 0.696 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.706, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We have introduced uCO3D, a new object-centric 3D dataset of real-life objects. uCO3D strikes a balance between size and quality, ensuring the quality of the collected turntable-like videos and of the 3D annotations, while at the same time significantly expanding the scale of the data compared to CO3Dv2 and the diversity compared to CO3Dv2 and MVImgNet. We have shown the benefits of using this dataset compared to alternatives when training models for feedforward few-view 3D reconstruction, multi-view generation, and text-to-3D generation. Equipped with 3D cameras, point clouds, masks, textual captions, and 3DGS reconstructions of objects, uCO3D is a ready-to-use resource for training large generative models and for exploring 3D deep learning." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14109" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.119, + 0.188, + 0.135 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.145, + 0.482, + 0.185 + ], + "angle": 0, + "content": "[1] Aria digital twin catalog. https://www.projectaria.com/datasets/dtc/. Accessed: 2024-12-20.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.482, + 0.269 + ], + "angle": 0, + "content": "[2] Adel Ahmadyan, Liangkai Zhang, Artsiom Ablavatski, Jianing Wei, and Matthias Grundmann. Objectron: A large scale dataset of object-centric videos in the wild with pose annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7822-7831, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.483, + 0.326 + ], + "angle": 0, + "content": "[3] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proc. ICCV, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.328, + 0.483, + 0.369 + ], + "angle": 0, + "content": "[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In Proc. CVPR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.37, + 0.482, + 0.451 + ], + "angle": 0, + "content": "[5] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv.cs, abs/2311.15127, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.453, + 0.482, + 0.495 + ], + "angle": 0, + "content": "[6] Ang Cao, Justin Johnson, Andrea Vedaldi, and David Novotny. Lightplane: Highly-scalable components for neural 3d fields. arXiv, 2024. 2, 3, 5, 6, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.496, + 0.482, + 0.578 + ], + "angle": 0, + "content": "[7] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In Proc. CVPR, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[8] Angel X. Chang, Thomas A. Funkhouser, Leonidas J. Guibas, Pat Hanrahan, Qi-Xing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet an information-rich 3d model repository. arXiv.cs, abs/1512.03012, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.65, + 0.482, + 0.706 + ], + "angle": 0, + "content": "[9] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In European Conference on Computer Vision, pages 640-658. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.707, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[10] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21126-21136, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[11] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam S. Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, Matthew Yu, Abhishek Kadian, Filip Radenovic, Dhruv Mahajan, Kunpeng Li, Yue Zhao, Vladan Petrovic, Mitesh Kumar Singh, Simran Motwani, Yi Wen, Yiwen Song, Roshan Sumbaly, Vignesh Ramanathan, Zijian He, Peter Vajda, and Devi Parikh. Emu:" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.145, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Enhancing image generation models using photogenic needles in a haystack. CoRR, abs/2309.15807, 2023. 2, 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[12] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-XL: A universe of \\(10\\mathrm{M} + 3\\mathrm{D}\\) objects. CoRR, abs/2307.05663, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[13] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In Proc. CVPR, 2023. 2, 3, 6, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.29, + 0.905, + 0.372 + ], + "angle": 0, + "content": "[14] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.373, + 0.904, + 0.441 + ], + "angle": 0, + "content": "[15] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B. McHugh, and Vincent Vanhoucke. Google Scanned Objects: A high-quality dataset of 3D scanned household items. In Proc. ICRA, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.444, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[16] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aurélien Rodriguez, Austen Gregerson, Ava Spataru, Baptiste Rozière, Bethany Biron, Binh Tang, Bobbie Chern, Charlotte Caucheteux, Chaya Nayak, Chloe Bi, Chris Marra, Chris McConnell, Christian Keller, Christophe Touret, Chunyang Wu, Corinne Wong, Cristian Canton Ferrer, Cyrus Nikolaidis, Damien Allonsius, Daniel Song, Danielle Pintz, Danny Livshits, David Esiobu, Dhruv Choudhary, Dhruv Mahajan, Diego Garcia-Olano, Diego Perino, Dieuwke Hupkes, Egor Lakomkin, Ehab AlBadawy, Elina Lobanova, Emily Dinan, Eric Michael Smith, Filip Radenovic, Frank Zhang, Gabriel Synnaeve, Gabrielle Lee, Georgia Lewis Anderson, Graeme Nail, Grégoire Mialon, Guan Pang, Guillem Cucurell, Hailey Nguyen, Hannah Ko-revaar, Hu Xu, Hugo Touvron, Iliyan Zarov, Imanol Arrieta Ibarra, Isabel M. Kloumann, Ishan Misra, Ivan Evtimov, Jade Copet, Jaewon Lee, Jan Geffert, Jana Vranes, Jason Park, Jay Mahadeokar, Jeet Shah, Jelmer van der Linde, Jennifer Billock, Jenny Hong, Jenya Lee, Jeremy Fu, Jianfeng Chi, Jianyu Huang, Jiawen Liu, Jie Wang, Jiecao Yu, Joanna Bitton, Joe Spisak, Jongsoo Park, Joseph Rocca, Joshua Johnstun, Joshua Saxe, Junteng Jia, Kalyan Vasuden Alwala Kartikeya Upasani Kate Plawiak Ke Li Kenneth Heafield and Kevin Stone. The Llama 3 herd of models. arXiv 2407.21783, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[17] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d furniture shape with texture. International Journal of Computer Vision, 129:3313-3337, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14110" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[18] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T. Barron, and Ben Poole. CAT3D: Create Anything in 3D with Multi-View Diffusion Models. arXiv.cs, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.485, + 0.232 + ], + "angle": 0, + "content": "[19] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T. Barron, and Ben Poole. CAT3D: create anything in 3d with multi-view diffusion models. arXiv, 2405.10314, 2024. 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.484, + 0.26 + ], + "angle": 0, + "content": "[20] Paul Guerrero. Lang-sam. https://github.com/paulguerrero/lang-sam, 2024.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.484, + 0.315 + ], + "angle": 0, + "content": "[21] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[22] Agrim Gupta, Piotr Dólar, and Ross B. Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Proc. CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.484, + 0.44 + ], + "angle": 0, + "content": "[23] Philipp Henzler, Jeremy Reizenstein, Patrick Labatut, Roman Shapovalov, Tobias Ritschel, Andrea Vedaldi, and David Novotny. Unsupervised learning of 3d object categories from videos in the wild. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.441, + 0.482, + 0.495 + ], + "angle": 0, + "content": "[24] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[25] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In Proc. NeurIPS, 2017. 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.609 + ], + "angle": 0, + "content": "[26] Lukas Hollein, Aljaž Božić, Norman Müller, David Novotny, Hung-Yu Tseng, Christian Richardt, Michael Zollhöfer, and Matthias Nießner. ViewDiff: 3D-Consistent Image Generation with Text-to-Image Models. arXiv preprint, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.61, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[27] Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. In Proc. ICLR, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.482, + 0.706 + ], + "angle": 0, + "content": "[28] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, pages 406-413, 2014. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[29] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian Splating for real-time radiance field rendering. Proc. SIGGRAPH, 42(4), 2023. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[30] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[31] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. Proc. ICLR, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[32] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proc. CVPR, 2023. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[33] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. ABC: A big CAD model dataset for geometric deep learning. In Proc. CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.906, + 0.203 + ], + "angle": 0, + "content": "[34] Zhengfei Kuang, Yunzhi Zhang, Hong-Xing Yu, Samir Agarwala, Elliott Wu, Jiajun Wu, et al. Stanford-orb: a real-world 3d object inverse rendering benchmark. In NeurIPS, 2023. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.204, + 0.906, + 0.273 + ], + "angle": 0, + "content": "[35] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3D: Fast text-to-3D with sparse-view generation and large reconstruction model. Proc. ICLR, 2024. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.908, + 0.328 + ], + "angle": 0, + "content": "[36] Joseph J Lim, Hamed Pirsiavash, and Antonio Torralba. Parsing intela objects: Fine pose estimation. In Proceedings of the IEEE international conference on computer vision, pages 2992-2999, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.906, + 0.371 + ], + "angle": 0, + "content": "[37] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. arXiv.cs, abs/2305.08891, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.906, + 0.426 + ], + "angle": 0, + "content": "[38] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In Proc. ECCV, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.906, + 0.469 + ], + "angle": 0, + "content": "[39] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3D object. In Proc. ICCV, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.906, + 0.524 + ], + "angle": 0, + "content": "[40] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-Solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In Proc. NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.906, + 0.567 + ], + "angle": 0, + "content": "[41] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. arXiv preprint, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.906, + 0.648 + ], + "angle": 0, + "content": "[42] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, Natalia Neverova, Andrea Vedaldi, Oran Gafni, and Filippos Kokkinos. IM-3D: Iterative multiview diffusion and reconstruction for high-quality 3D generation. In Proceedings of the International Conference on Machine Learning (ICML), 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.65, + 0.906, + 0.72 + ], + "angle": 0, + "content": "[43] Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.906, + 0.788 + ], + "angle": 0, + "content": "[44] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Carl Yuheng Ren. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.906, + 0.83 + ], + "angle": 0, + "content": "[45] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. In Proc. ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[46] Christian Reiser, Richard Szeliski, Dor Verbin, Pratul P. Srinivasan, Ben Mildenhall, Andreas Geiger, Jonathan T. Barron, and Peter Hedman. MERF: memory-efficient radiance fields for real-time view synthesis in unbounded scenes. abs/2302.12249, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14111" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[47] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common Objects in 3D: Large-scale learning and evaluation of real-life 3D category reconstruction. In Proc. ICCV, 2021. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proc. CVPR, 2022. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.249 + ], + "angle": 0, + "content": "[49] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In Proc. ICLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[50] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proc. CVPR, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.483, + 0.334 + ], + "angle": 0, + "content": "[51] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In Proc. ECCV, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.336, + 0.483, + 0.419 + ], + "angle": 0, + "content": "[52] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.483, + 0.463 + ], + "angle": 0, + "content": "[53] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. MVDream: Multi-view diffusion for 3D generation. In Proc. ICLR, 2024. 2, 3, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.464, + 0.483, + 0.561 + ], + "angle": 0, + "content": "[54] Yawar Siddiqui, Filippos Kokkinos, Tom Monnier, Mahendra Kariya, Yanir Kleiman, Emilien Garreau, Oran Gafni, Natalia Neverova, Andrea Vedaldi, Roman Shapovalov, and David Novotny. Meta 3D Asset Gen: Text-to-mesh generation with high-quality geometry, texture, and PBR materials. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.483, + 0.645 + ], + "angle": 0, + "content": "[55] Yawar Siddiqui, Tom Monnier, Filippos Kokkinos, Mahendra Kariya, Yanir Kleiman, Emilien Garreau, Oran Gafni, Natalia Neverova, Andrea Vedaldi, Roman Shapovalov, et al. Meta 3d assetgen: Text-to-mesh generation with high-quality geometry, texture, and pbr materials. NeurIPS, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.483, + 0.716 + ], + "angle": 0, + "content": "[56] Xingyuan Sun, Jiajun Wu, Xiuming Zhang, Zhoutong Zhang, Chengkai Zhang, Tianfan Xue, Joshua B Tenenbaum, and William T Freeman. Pix3d: Dataset and methods for single-image 3d shape modeling. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[57] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Viewset diffusion: (0-)image-conditioned 3D generative models from 2D data. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[58] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Splatter Image: Ultra-fast single-view 3D reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[59] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. LGM: Large multi-view Gaussian model for high-resolution 3D content creation. arXiv, 2402.05054, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[60] Shitao Tang, Jiacheng Chen, Dilin Wang, Chengzhou Tang, Fuyang Zhang, Yuchen Fan, Vikas Chandra, Yasutaka Furukawa, and Rakesh Ranjan. MVDiffusion++: A dense high-resolution multi-view diffusion model for single or sparse-view 3d object reconstruction. arXiv, 2402.12712, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[61] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. NeurIPS, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[62] Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. Vggsfm: Visual geometry grounded deep structure from motion. In CVPR, pages 21686-21697, 2024. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.281, + 0.905, + 0.349 + ], + "angle": 0, + "content": "[63] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P. Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas A. Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proc. CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.353, + 0.905, + 0.395 + ], + "angle": 0, + "content": "[64] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. DUSt3R: Geometric 3D vision made easy. In Proc. CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.397, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[65] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. MeshLRM: large reconstruction model for high-quality mesh. arXiv, 2404.12385, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[66] Rundi Wu, Chang Xiao, and Changxi Zheng. Deepcad: A deep generative network for computer-aided design models. in 2021 IEEE. In CVF International Conference on Computer Vision (ICCV), pages 6772-6782, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.513, + 0.905, + 0.583 + ], + "angle": 0, + "content": "[67] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P. Srinivasan, Dor Verbin, Jonathan T. Barron, Ben Poole, and Aleksander Holynski. ReconFusion: 3D Reconstruction with Diffusion Priors. arXiv preprint, 2023. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.585, + 0.905, + 0.669 + ], + "angle": 0, + "content": "[68] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, et al. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 803-814, 2023. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.671, + 0.905, + 0.726 + ], + "angle": 0, + "content": "[69] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguuang Zhang, Xiaou Tang, and Jianxiong Xiao. 3D ShapeNets: A deep representation for volumetric shapes. In Proc. CVPR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.729, + 0.905, + 0.77 + ], + "angle": 0, + "content": "[70] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond PASCAL: A benchmark for 3D object detection in the wild. In Proc. WACV, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.773, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[71] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. InstantMesh: efficient 3D mesh generation from a single image with sparse-view large reconstruction models. arXiv, 2404.07191, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[72] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation. arXiv preprint arXiv:2403.14621, 2024. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14112" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[73] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. GRM: Large gaussian reconstruction model for efficient 3D reconstruction and generation. arXiv, 2403.14621, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.217 + ], + "angle": 0, + "content": "[74] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proc. CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[75] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, and Angjoo Kanazawa. gsplat: An open-source library for Gaussian splatting. arXiv preprint arXiv:2409.06765, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.358 + ], + "angle": 0, + "content": "[76] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, et al. Mvmgnet: A large-scale dataset of multi-view images. In CVPR, pages 9150-9161, 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.482, + 0.415 + ], + "angle": 0, + "content": "[77] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. GS-LRM: large reconstruction model for 3D Gaussian splatting. arXiv, 2404.19702, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.471 + ], + "angle": 0, + "content": "[78] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proc. CVPR, pages 586-595, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.474, + 0.482, + 0.528 + ], + "angle": 0, + "content": "[79] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snively. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.531, + 0.482, + 0.599 + ], + "angle": 0, + "content": "[80] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets Gaussian splatting: Fast and generalizable single-view 3D reconstruction with transformers. arXiv.cs, abs/2312.09147, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "14113" + } + ] +] \ No newline at end of file diff --git a/2025/UnCommon Objects in 3D/4d2bd976-41fd-45e5-bfc6-8a4578f73725_origin.pdf b/2025/UnCommon Objects in 3D/4d2bd976-41fd-45e5-bfc6-8a4578f73725_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b091861434b76dc5cfc632c72e178c9290c10c84 --- /dev/null +++ b/2025/UnCommon Objects in 3D/4d2bd976-41fd-45e5-bfc6-8a4578f73725_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:232444dc97af808821cc1d906831f85ae5bd8987bdca0fdf9bb7784f55aebf2d +size 4732705 diff --git a/2025/UnCommon Objects in 3D/full.md b/2025/UnCommon Objects in 3D/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d3757c65e93f62a728a68dbbaf382f5d88a60831 --- /dev/null +++ b/2025/UnCommon Objects in 3D/full.md @@ -0,0 +1,273 @@ +# UnCommon Objects in 3D + +Xingchen Liu $^{\infty}$ Piyush Tayal $^{\infty}$ Jianyuan Wang $^{\infty}$ Jesus Zarzar $^{+\infty}$ Tom Monnier $^{\infty}$ Konstantinos Tertikas $^{\infty*}$ Jiali Duan $^{\infty}$ Antoine Toisoul $^{\infty}$ Jason Y. Zhang $^{\dagger}$ Natalia Neverova $^{\infty}$ Andrea Vedaldi $^{\infty}$ Roman Shapovalov $^{\infty}$ David Novotny $^{\infty}$ + +*NKUA, Greece †KAUST †Carnegie Mellon University †Meta AI + +uCO3D https://uco3d.github.io https://github.com/facebookresearch/uco3d + +![](images/948844b57f01375a98ddd77397dac02a46e50d0415ca6e1316512cc7c65dbf5b.jpg) +Figure 1. We introduce UnCommon Objects in 3D (uCO3D), a large and diverse dataset of high-quality $360^{\circ}$ videos covering over 1,000 object categories. Each video frame is 3D-annotated with accurate SfM cameras, point cloud, and a 3D Gaussian Splitting reconstruction. + +# Abstract + +We introduce Uncommon Objects in 3D (uCO3D), a new object-centric dataset for 3D deep learning and 3D generative AI. uCO3D is the largest publicly-available collection of high-resolution videos of objects with 3D annotations that ensures full- $360^{\circ}$ coverage. uCO3D is significantly more diverse than MVImgNet and CO3Dv2, covering more than 1,000 object categories. It is also of higher quality, due + +to extensive quality checks of both the collected videos and the 3D annotations. Similar to analogous datasets, uCO3D contains annotations for 3D camera poses, depth maps and sparse point clouds. In addition, each object is equipped with a caption and a 3D Gaussian Splat reconstruction. We train several large 3D models on MVImgNet, CO3Dv2, and uCO3D and obtain superior results using the latter, showing that uCO3D is better for learning applications. + +# 1. Introduction + +The primacy of data has been the defining characteristic of the last decade of machine learning, alongside deep learning. The most powerful models in language, speech, and computer vision are large deep networks trained on massive amounts of data, and then further fine-tuned on its high-quality subset. This paradigm is expected to extend to all machine learning applications, including 3D vision. + +However, 3D training data is much harder to come by than data for text, audio, and image processing. + +Seeking training data for large 3D neural networks [27], many have turned to synthetic datasets like Objaverse [13]. However, synthetic data is a poor substitute for real data in applications like digital twinning, which creates 3D models of real-life objects. This is why many photorealistic reconstruction networks [6, 23, 57, 58, 63, 67] are trained using real object-centric datasets like CO3D [47], MVImgNet [76], GSO [15], and OmniObject3D [68]. Real data is also crucial for generalization, as demonstrated by DUSt3r [64] for point map prediction and DepthAnything [74] for depth prediction, both of which are trained on numerous real datasets. Even non-curated image datasets like the billion-scale LAION [52] are applicable to 3D vision. For instance, text-to-3D generators [35, 45, 53, 54] build on LAION-trained large text-to-image models [5, 11]. + +Given the importance of 3D datasets, but also their relative scarcity, in this paper we ask what is the next step for real data in 3D vision. To answer this question, we note that, while the size of a dataset is crucial, in most cases its quality is just as important. For example, text-to-3D models [35, 55] are notoriously sensitive to fine-tuning data quality, training only with the best-looking models (e.g., Instant3D [35] uses only about $1\%$ of Objaverse). We conclude that high-quality datasets balancing scale and fidelity are more valuable than simply amassing low-quality data. + +Based on this observation, we argue that there is a gap in the real object-centric 3D datasets that are currently available, as none strikes the optimal balance between quality and scale. For example, the 3D object scans in OmniObject3D [68] and GSO [15] provide very accurate geometry and textures, but only count a few thousand objects. Conversely, datasets like CO3D [47] and MvImgNet [76] contain orders-of-magnitude more objects, but lack reliable 3D scans. Instead, they provide many views of the objects together with lower-quality 3D cameras and point clouds reconstructed with structure-from-motion (SfM). + +In this paper, we address this gap with a new dataset, Uncommon Objects in 3D (uCO3D), which better balances data quality and size (Tab. 1). Similar to CO3D, it comprises full- $360^{\circ}$ crowd-sourced videos capturing objects from all sides, annotated with cameras and point clouds using SfM. Furthermore, uCO3D has much greater data diversity (Fig. 2) than prior alternatives as it contains objects + +
RealCount# ClassesData typeAnnotations
ShapeNet [8]X51k553D meshesmesh
Objaverse [13]X800k21k3D meshesmesh
Objaverse-XL [12]X10M2M3D meshesmesh
ABO [10]X8k633D meshesmesh
OmniObject3D [68]✓6k190Videos w/ meshescameras, mesh
GSO [14]✓1k17Views w/ meshescameras, mesh
Objectron [2]✓15k9Limited vp. videoscameras, 3D box
MVImgNet [76]✓220k238Limited vp. videoscameras, pcl
CO3D [47]✓19k50360° videoscameras, pcl
CO3Dv2 [47]✓40k50360° videoscameras, pcl
uCO3D (ours)✓170k1k360° videoscameras, 3DGS, caption
+ +Table 1. Overview of 3D object datasets. We compare the number of objects / classes, the type of data and associated annotations. + +from the 1,070 visual object categories of the LVIS [22] taxonomy, which has long tails. For reference, MVImgNet and CO3Dv2 contain only 238 and 50 categories, respectively. These fine-grained categories are organized in supercategories, also shown in Fig. 2. Furthermore, uCO3D contains 170k scenes, which is more than four times larger than CO3Dv2's 40k. While this is less than MVImgNet's 220k, uCO3D's videos cover each object from all sides, as opposed to MVImgNet's partial object captures. + +Besides improving size and diversity, uCO3D also raises the quality bar. This was achieved by checking extensively both the collected videos and their 3D annotations. Differently from datasets like CO3Dv2 that still contain a certain portion of low-quality videos, in uCO3D we manually verified that each video provides full $360^{\circ}$ turn-table covering all sides of the object. Additionally, $60\%+$ of the videos have $1080p+$ resolution, higher than CO3Dv2. To ensure 3D-annotation quality, we improved both the reconstruction algorithm and the reconstruction validation. For camera reconstruction, we used VGGSfM [62], which is currently the best SfM system available, and is more robust and accurate than COLMAP [50], used in CO3Dv2 and MvImgNet. We also improve on CO3Dv2's active-learning camera quality evaluation by combining it with novel-view synthesis accuracy after reconstructing each scene using 3D Gaussian Splatting (3DGS) [29]. The latter also guarantees that scenes are reconstructible to a high quality, which is important for training of 3D models. + +We validate uCO3D's benefits in applications. We train two popular 3D models, LRM [27] and CAT3D [18], using uCO3D and demonstrate improved results compared to training on MVImgNet and CO3Dv2, which makes uCO3D the better data source for real object-centric 3D learning. We also use uCO3D to train a text-to-3D model following Instant3D's [35] two-stage design. The latter requires objects to be rendered from canonical viewpoints, and thus, so far, was limited to synthetic data. By using our 3DGS reconstructions, we 're-shoot' uCO3D's from these viewpoints, which allows to train a more realistic generator. + +![](images/1c9a9d2aec5e61eea77375b106e8b94f4d471dc011a18f91314b3318ac994d63.jpg) +Figure 2. Statistics of uCO3D. (Left) We plot the number of objects per super-category. In total, the dataset contains 50 super-categories, each gathering around 20 sub-categories. (Right) We show a word cloud of all 1,070 visual categories represented in the dataset. + +# 2. Related Work + +Datasets of synthetic 3D objects. Historically, object-centric 3D datasets have predominantly been synthetic, composed of artist-generated 3D models. A prominent example is ShapeNet [8], with 51,000 meshes across 55 object categories. The meshes have detailed geometries, but relatively simplistic textures. Datasets such as 3DFUTURE [17], IKEA [36], Pix3D [56], and ABO [10] are less diverse, concentrating on furniture and other consumer goods. In contrast, ModelNet [69], DeepCAD [66], and ABC [33] provide CAD models with clean geometry but no texture. Objaverse [13] is perhaps the most impactful dataset following ShapeNet. It is significantly larger, comprising 800,000 3D meshes. Objaverse-XL [12] expands this collection to 10M objects. These datasets have been pivotal in enabling the first 3D deep generative models, including text-to-3D [35, 53, 54] and image-to-3D [6, 27, 73] models. + +Real 3D object datasets. Acquiring real-world 3D data presents many challenges, resulting in only a few real 3D-object datasets. Early datasets like Pascal3D [70] include multiple categories but provide only isolated object views and approximate 3D annotations. Conversely, DTU [28], BlendedMVS [51], GSO [15], OmniObject3D [68], Aria Digital Twin [44], and Digital Twin Catalog [1] provide 3D scans of objects, featuring high-quality 3D geometry and textures, but have only a few thousand objects. + +The use of 3D scanners significantly restricts the scale of data acquisition; consequently, other datasets capture multiview turntable-like videos of objects using consumer cameras. CO3D and CO3Dv2 [47] crowd-sourced $40,000360^{\circ}$ object videos, providing 3D annotations by reconstructing point clouds and cameras using COLMAP SfM [50]. MvImgNet [76] collected even more videos (220,000) across more object categories (238), but their videos capture objects only partially, preventing full reconstruction. Objectron [2] is similar to MvImgNet, but with fewer videos (10,000). A common challenge is that large-scale datasets often rely on SfM for video reconstruction, which can + +lead to imprecise 3D annotations. uCO3D also employs SfM, but using VGGSfM, which has greater accuracy than COLMAP, and with a more reliable data validation setup. Furthermore, uCO3D is five times larger and significantly more diverse than CO3Dv2, encompassing 20 times more visual categories, and provides caption and 3D Gaussian Splat reconstructions of each object. + +Applications. In order to assess the quality of uCO3D, we measure how it benefits a number of popular downstream applications. First, we consider feedforward few-view 3D reconstruction models. Among those, LRM [27] is a transformer that maps an input image to a neural radiance field supported by a triplane [7]. LightplaneLRM [6] adds splatting layers and a memory-efficient renderer. Further extensions use different representations like 3D Gaussian Splats [59, 73, 77, 80] and meshes [65, 71]. + +We also consider text-to-3D generators, which create 3D assets from text, and focus on the two-stages approach of Instant3D [35]. This is based on training a text-to-multi-view diffusion model [11, 48] which generates several 2D views of the object, followed by a 3D reconstruction network that outputs the 3D asset, all in a matter of seconds. The multi-view diffusion is improved in ViewDiff [26], MVDiffusion [60], IM-3D [42], CAT3D [18] and many others. AssetGen [55] further extends Instant3D by modelling material properties instead of baking in the radiance function and adds a texture refiner that outputs relightable PBR textures. As an illustration, we use uCO3D to train a model like CAT3D, which results in better new-view synthesis than the one trained on alternative datasets. We also show that the Gaussian Splat reconstructions provided with uCO3D can supervise, for the first time, an Instant3D-like pipeline using solely real-life data. + +# 3. Uncommon Objects in 3D + +In this section, we introduce uCO3D, our new dataset of real-life 3D objects. uCO3D comprises $360^{\circ}$ turn-table-like videos of objects, crowdsourced and annotated with 3D cameras, point clouds, 3D Gaussians, and textual captions. + +![](images/1fe67a25a5d1ab72ff15d61286823e3b7b97aa7df867541a99fc374591138628.jpg) +Video + +![](images/0190d353e987df0730825c5db96919ce00b367b182a5e27137b0c4d88a877fe5.jpg) +a) Sparse SfM reconstruction + +![](images/0b7647e4b46ffe6a23cc4a62672c65288cebac29a5fd0de5ff4f01fd177cfb0a.jpg) +b) Densified SfM point cloud + +![](images/50e0c9f3d8d42436ee19e288476b00f646d7b940fa2509d7b181deced7d14e20.jpg) +c) 3D Gaussian Splats +Figure 3. Data annotation overview. Each scene in uCO3D is reconstructed in three different ways: a) per-frame cameras with sparse point cloud calculated by VGGSfM [62], b) semi-dense point cloud comprising triangulations of additional denser tracks from VGGSfM's tracker, c) 3D Gaussian Splat [29] reconstruction optimized separately for each scene. + +![](images/335b00b451c85871155c25bb326249a4455b9d40dd1c807515e4153ae4468f97.jpg) +a) Input video + +![](images/7b94ce4c01501416c8417140602cdde0b2a17cc8c0bd545c34466c70c2960190.jpg) +b) Sine-wave camera trajectory +Figure 4. Data collection example. For each video, the cameras follow a sine-wave trajectory to ensure good viewpoint coverage. + +Compared to older datasets like CO3Dv2 [47], uCO3D comes with many improvements. First, uCO3D is much larger and more diverse than CO3Dv2: it contains more than 1k different categories and more than 170k objects, compared to the 50 and 38k of CO3Dv2. While CO3Dv2's categories are taken from MS COCO [38], the categories in uCO3D are taken from the LVIS [21] taxonomy. Hence, we inherit the LVIS focus on covering the long-tail of the visual-category distribution. To simplify data analysis, we grouped the $1\mathrm{k}+$ LVIS categories to 50 super-categories, each containing approximately 20 subcategories. Figure 2 shows the number of videos collected per super-category, and the LVIS category distribution. + +Second, uCO3D improves quality significantly compared to CO3Dv2, ensuring that videos are of high reso + +lution, cover each object well, and that the 3D annotations are accurate. uCO3D also contains rich textual descriptions of each object, missing in other datasets, and useful to train large generative models. It also comes with additional 3D Gaussian Splat reconstructions of all objects, each rigidly aligned to a canonical object-centric reference, which make it possible to re-render the dataset from a fixed, canonical set of cameras, simulating synthetic data acquisition, which is very useful for training generative models [35, 55]. + +Dataset collection. Videos of objects were captured by workers on Amazon Mechanical Turk. To ensure high video quality, workers were required to submit videos of a sufficient resolution. As a result, more than $60\%$ of videos in uCO3D are of 1080p resolution or higher, compared to $33\%$ in CO3Dv2. Furthermore, to aid the 3D reconstruction, workers followed a sine-wave capture trajectory instead of the plain circular trajectory of CO3Dv2, ensuring varying camera elevations (cf. Fig. 4). Finally, each video was individually manually assessed to make sure that it adheres to these requirements, a process more rigorous than the rough eyeballing used in CO3Dv2 [47]. + +Video object segmentation. We used text-conditioned Segment-Anything (langSAM) [20, 32] to segment the object of interest in each video frame given text-conditioning in form of the object-category name, which had been provided by Turkers at collection time. + +To improve frame-to-frame consistency, CO3Dv2 used a simple Viterbi algorithm, which often led to segmentation + +![](images/2506e58f0b6465728af23e0f29d70188eb225e0b69f86010fd0a44e4e085c24f.jpg) +Figure 5. 3D reconstruction comparison. We show results of LightplaneLRM [6] models trained on MVIImgNet, CO3Dv2 and uCO3D. + +
Train datasetOmniObject3DStanfordORB
LPIPS↓PSNR↑IoU↑LPIPS↓PSNR↑IoU↑
MVIImgNet [76]0.10923.390.9280.07024.4510.939
CO3Dv2 [47]0.09523.620.9260.05625.6170.956
uCO3D (ours)0.09324.610.9460.05725.7150.957
+ +Table 2. 3D reconstruction benchmark. We compare LightplaneLRM [6] models trained on CO3Dv2, MVImgNet, and uCO3D. We report novel-view synthesis performances on OmniObject3D [68] and StanfordORB [34]. + +flickering, impairing the final 3D reconstruction quality. Instead, in uCO3D, we refine the SAM segmentations with state-of-the-art deep video-segmenter based on XMem [9], leading to more temporally-stable object segmentations. + +3D annotation with VGGSfM. For each video, we use the state-of-the-art VGGSfM [62] Structure from Motion (SfM) system to estimate the parameters of the cameras (intrinsic and extrinsic) for 200 frames sampled uniformly. VGGSfM also outputs a sparse 3D point cloud, and its denser version obtained by triangulating additional 3D points from VGGSfM's tracker. Examples of sparse and densified SfM point clouds are shown in Fig. 3. + +Scene alignment. While the coordinate system of VG-GSfM cameras is defined only up to a rigid transformation, it is crucial for applications like generation and reconstruction to train on a dataset of rigidly aligned objects. We thus align all objects so they have a horizontal ground plane, similar scale, centring, and orientation. Details of the scene alignment procedure are in the supplementary material. + +Gaussian Splat reconstruction. Sparse and even dense SfM point clouds provide an accurate but still quite sparse 3D reconstruction of the scene's surface. To further densify it, uCO3D provides a 3D Gaussian Splat reconstruction [29] for each scene, fitted using gsplat [75]. + +Scene captioning. uCO3D also provides textual captions for all scenes, useful for generative modelling. Motivated by Cap3D [41], we first caption each view separately using a vision-language model, and then summarise these into a single scene caption using LLAMA3 [16]. + +# 4. Applications + +In this section, we demonstrate uCO3D's merit on three popular 3D learning tasks: feedforward sparse-view 3D reconstruction (Sec. 4.1), new-view synthesis using diffusion + +![](images/7fbbd9a4c29a229d341999b5d317e801372b7d676572ace261639cfc958e115f.jpg) +Figure 6. Novel-view synthesis comparison. We compare results of CAT3D-like [19] models trained on different datasets (MVIImgNet, CO3D, uCO3D) and evaluated on standard NVs datasets (top-to-bottom: RealEstate10K [79], LLFF [43], DTU [28], Mip-NeRF 360 [4]). + +
Train datasetEasierDataset DifficultyHarder
←Re10K [79]LLFF [43]DTU [28]Mip-NeRF [4]
LPIPS↓PSNR↑LPIPS↓PSNR↑LPIPS↓PSNR↑LPIPS↓PSNR↑
MVImgNet [76]0.31018.770.42614.380.37712.790.60512.39
CO3Dv2 [47]0.28120.020.41814.950.32916.420.53214.19
uCO3D (ours)0.27819.770.41815.160.31516.970.52814.37
+ +Table 3. Novel-view synthesis benchmark. We evaluate CAT3D-like [19] models trained on MVImgNet, CO3Dv2 or uCO3D. We report NVS performances on RealEstate10K [79], LLFF [43], DTU [28] and Mip-NeRF 360 [4]. + +(Sec. 4.2), and text-to-3D (Sec. 4.3). + +# 4.1. Few-view 3D Object Reconstruction + +Traditionally, multi-view 3D-annotated datasets such as CO3D or MVImgNet have been used to supervise few-view 3D reconstructors. In this section, we train LightplaneLRM [6], an evolution of the seminal LRM [27], and show that doing so on uCO3D leads to better performance than training on alternative datasets. + +LRM is a large transformer [61] that accepts few input images of an object and predicts a 3D representation of the latter. The transformer, conditioned on the tokens of the ob + +served images via cross attention, converts a set of learnable input tokens to a 3D representation. The 3D representation is a triplane [7] supporting an opacity/radiance implicit shape. LightplaneLRM improves LRM by adding so called "splattering layers" and a memory-efficient renderer. + +During training, LightplaneLRM receives four random source frames from a training uCO3D video sequence, and renders the predicted triplane into held-out target views. Learning minimizes the photometric loss between the renders and the corresponding ground-truth targets. Both source and target views are masked using the extracted segmentation masks to make sure that LightplaneLRM only reconstructs the foreground object. Training uses the Adam optimizer and is warm-started following the original LRM training protocol by pre-training the model on a large dataset of synthetic objects similar to Objaverse [13]. + +Baselines. Our main goal is to demonstrate that uCO3D contains higher quality data than existing object-centric datasets. As such, starting from the model pre-trained on the synthetic data, we finetune either on uCO3D, or on two other baseline datasets, namely MVImgNet and CO3Dv2. + +Evaluation protocol. We evaluate each trained model in a novel-view synthesis setting on two small-scale high- + +quality object-centric datasets: OmniObject3D [68] and Stanford-ORB [34]. Given four views of a held-out test scene, the model reconstructs the scene which is then rendered to unseen target views. We report the average LPIPS [78] loss and Peak-signal-to-noise ratio (PSNR) between each render and the corresponding ground-truth image. We also report the intersection-over-union (IoU) between the rendered object alpha mask and the target view segmentation mask. + +Results. Table 2 and Fig. 5 report the quantitative and qualitative results, respectively. The LightplaneLRM trained on uCO3D is better than the other baselines in most metrics on both datasets. The latter confirms that uCO3D is currently the most reliable source of real data for training feedforward few-view 3D reconstructors. + +# 4.2. Novel-view synthesis using diffusion + +We now consider application of uCO3D to training new-view image diffusion generators. These generators can, given one or a few views of an object and a target camera pose, output new arbitrary views as observed from the target camera, hallucinating missing details based on a statistical prior they learn. They can thus complement and integrate the feed-forward reconstruction models of the previous section, which are deterministic and thus unable to deal with ambiguity well. To this end, we train a diffusion model similar to the recent CAT3D [19], but reimplement it from scratch given lack of source code (see details in the supplementary). We call this model CAT3D-like. + +Evaluation protocol. As in Sec. 4.1, we compare versions of CAT3D-like trained using uCO3D, MVImgNet, and CO3Dv2 and test them on held-out datasets. A feature of CAT3D is the ability to reconstruct both the principal object in the images as well as the background. We thus benchmark the method using new-view synthesis datasets that do contain background, namely DTU [28], containing structured light scans of various objects, LLFF [43], containing scenes captured from fronto-parallel camera trajectories, RealEstate10k [79], containing real-estate walk-throughs, and Mip-NeRF 360 [3] with complex indoor and outdoor scenes. For evaluation, we take three known views as input and use the model to predict a new view. We report LPIPS and PSNR but not the IoU since CAT3D only generates new RGB views without reconstructing the 3D shape. + +Results. Table 3 and Fig. 6 contain the results: training CAT3D-like on uCO3D leads to the best performance across all four datasets. Even when compared to MVImgNet, which is slightly larger than uCO3D, the latter improves PSNR by 3-4 points, and reduces the LPIPS error by $5\%$ to $20\%$ . + +![](images/0288ee55c550563816bdacb45ca723a0fb48b07c37bc1972bd1fc319bb942999.jpg) +Figure 7. Supervising Instant3D with 3DGS. For each training scene, its 3DGS is rendered from 4 canonical views yielding a captioned image dataset for finetuning an image diffuser. Samples from the latter are then reconstructed with LRM. + +# 4.3. Photorealistic Text-to-3D + +Next, we show that uCO3D enables training photorealistic text-to-3D generators. Methods like CAT3D and others [39, 42, 53] generate several views of the object first, and then fit a 3D model, such as a NeRF or 3DGS, via optimization. This can work well, but it is not particularly robust or fast. An alternative, popularized by Instant3D [35] and follow-ups [54, 72], is to use a feedforward reconstructor in the second step, similar to LightplaneLRM from Sec. 4.1, which is faster and more robust. However, these models require canonical views of the objects — for example, Instant3D considers 4 orthogonal viewpoints, covering all 'sides'. The requirement of such training canonical views complicates training on real data, where viewpoints are arbitrary, and explains why such models are usually trained on synthetic data, limiting realism. + +Imaging 3DGS from canonical views. Our new idea is to 're-shoot' the 3DGS reconstructions provided with uCO3D from canonical viewpoints, making our data compatible with any method requiring canonical views for training (Fig. 7). To do so, we render the normalized reference frames (Sec. 3) into four views for each object, and arrange them in a grid as a target for the text-to-4-view generator. We double check the quality of the renders by calculating their CLIP similarity [24] to the object caption, and discard a sample if this is below 0.3. + +We use this data to fine-tune a text-to-4-view image diffusion model using these 4-view image grids and the corre + +![](images/e2b0fa420a90fe166e3f20c3ef071f3b28905f69dddf06737320c964708c87d4.jpg) +Figure 8. Qualitative results for text-to-3D generation displaying the 4-view grids generated by our Instant3D-like model given the input caption, and the 3D asset obtained by reconstructing the latter. The 4-view grid generator was trained using the canonical 4-view renders of uCO3D's 3DGS scene reconstructions. + +
Train datasetReal - FID↓Surreal - FID↓
Synthetic82.842.3
uCO3D (ours)63.968.9
+ +Table 4. Text-to-3D evaluation. We compare Instant3D-like models trained on uCO3D or a dataset of synthetic renders from artist-created meshes. We report FID on two sets of data corresponding to real and surreal objects, see text for details. + +sponding scene captions. At inference time, given a caption describing the desired object, we use the model to sample a new 4-view grid and feed the latter, together with the corresponding cameras, to the LightplaneLRM model (Sec. 4.1) for 3D reconstruction. + +Baselines. We train another 4-view generator on a dataset of synthetic assets similar to Objaverse [13] and use it with the original LightplaneLRM model [6] trained on the same data and thus optimally matched to it. + +Evaluation protocol. We report metrics evaluating the alignment between the distributions of the generations and the ground-truth objects. Specifically, we report the Frechet Inception Distance (FID) [25] between the renders of the generated 3D shapes and images of ground-truth objects. + +The main purpose of this experiment is to show that, by training on the uCO3D dataset, the 3D generations are more realistic. We assess this using two sets of prompts: Sur- + +real, containing 100 captions of objects from the synthetic dataset, and Real, containing 100 random captions form the held-out evaluation sequences of uCO3D. We report FID between the generated 3D shapes and the images/renders corresponding to the objects of each prompt-set. + +Results. Table 4 and Fig. 8 contain the quantitative and qualitative results respectively. The table reveals that the uCO3D-trained generator outperforms the synthetic generator when evaluated on real prompts. The latter verifies our hypothesis that a generator trained on uCO3D yields more realistic samples than a model trained on synthetic data. + +# 5. Conclusions + +We have introduced uCO3D, a new object-centric 3D dataset of real-life objects. uCO3D strikes a balance between size and quality, ensuring the quality of the collected turntable-like videos and of the 3D annotations, while at the same time significantly expanding the scale of the data compared to CO3Dv2 and the diversity compared to CO3Dv2 and MVImgNet. We have shown the benefits of using this dataset compared to alternatives when training models for feedforward few-view 3D reconstruction, multi-view generation, and text-to-3D generation. Equipped with 3D cameras, point clouds, masks, textual captions, and 3DGS reconstructions of objects, uCO3D is a ready-to-use resource for training large generative models and for exploring 3D deep learning. + +# References + +[1] Aria digital twin catalog. https://www.projectaria.com/datasets/dtc/. Accessed: 2024-12-20.3 +[2] Adel Ahmadyan, Liangkai Zhang, Artsiom Ablavatski, Jianing Wei, and Matthias Grundmann. Objectron: A large scale dataset of object-centric videos in the wild with pose annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7822-7831, 2021. 3 +[3] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proc. ICCV, 2021. 7 +[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In Proc. CVPR, 2022. 6 +[5] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv.cs, abs/2311.15127, 2023. 2 +[6] Ang Cao, Justin Johnson, Andrea Vedaldi, and David Novotny. Lightplane: Highly-scalable components for neural 3d fields. arXiv, 2024. 2, 3, 5, 6, 8, 1 +[7] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In Proc. CVPR, 2022. 3, 6 +[8] Angel X. Chang, Thomas A. Funkhouser, Leonidas J. Guibas, Pat Hanrahan, Qi-Xing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet an information-rich 3d model repository. arXiv.cs, abs/1512.03012, 2015. 3 +[9] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In European Conference on Computer Vision, pages 640-658. Springer, 2022. 5 +[10] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21126-21136, 2022. 3 +[11] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam S. Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, Matthew Yu, Abhishek Kadian, Filip Radenovic, Dhruv Mahajan, Kunpeng Li, Yue Zhao, Vladan Petrovic, Mitesh Kumar Singh, Simran Motwani, Yi Wen, Yiwen Song, Roshan Sumbaly, Vignesh Ramanathan, Zijian He, Peter Vajda, and Devi Parikh. Emu: + +Enhancing image generation models using photogenic needles in a haystack. CoRR, abs/2309.15807, 2023. 2, 3, 1 +[12] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-XL: A universe of $10\mathrm{M} + 3\mathrm{D}$ objects. CoRR, abs/2307.05663, 2023. 3 +[13] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In Proc. CVPR, 2023. 2, 3, 6, 8, 1 +[14] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022. +[15] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B. McHugh, and Vincent Vanhoucke. Google Scanned Objects: A high-quality dataset of 3D scanned household items. In Proc. ICRA, 2022. 2, 3 +[16] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aurélien Rodriguez, Austen Gregerson, Ava Spataru, Baptiste Rozière, Bethany Biron, Binh Tang, Bobbie Chern, Charlotte Caucheteux, Chaya Nayak, Chloe Bi, Chris Marra, Chris McConnell, Christian Keller, Christophe Touret, Chunyang Wu, Corinne Wong, Cristian Canton Ferrer, Cyrus Nikolaidis, Damien Allonsius, Daniel Song, Danielle Pintz, Danny Livshits, David Esiobu, Dhruv Choudhary, Dhruv Mahajan, Diego Garcia-Olano, Diego Perino, Dieuwke Hupkes, Egor Lakomkin, Ehab AlBadawy, Elina Lobanova, Emily Dinan, Eric Michael Smith, Filip Radenovic, Frank Zhang, Gabriel Synnaeve, Gabrielle Lee, Georgia Lewis Anderson, Graeme Nail, Grégoire Mialon, Guan Pang, Guillem Cucurell, Hailey Nguyen, Hannah Ko-revaar, Hu Xu, Hugo Touvron, Iliyan Zarov, Imanol Arrieta Ibarra, Isabel M. Kloumann, Ishan Misra, Ivan Evtimov, Jade Copet, Jaewon Lee, Jan Geffert, Jana Vranes, Jason Park, Jay Mahadeokar, Jeet Shah, Jelmer van der Linde, Jennifer Billock, Jenny Hong, Jenya Lee, Jeremy Fu, Jianfeng Chi, Jianyu Huang, Jiawen Liu, Jie Wang, Jiecao Yu, Joanna Bitton, Joe Spisak, Jongsoo Park, Joseph Rocca, Joshua Johnstun, Joshua Saxe, Junteng Jia, Kalyan Vasuden Alwala Kartikeya Upasani Kate Plawiak Ke Li Kenneth Heafield and Kevin Stone. The Llama 3 herd of models. arXiv 2407.21783, 2024. 5 +[17] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d furniture shape with texture. International Journal of Computer Vision, 129:3313-3337, 2021. 3 + +[18] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T. Barron, and Ben Poole. CAT3D: Create Anything in 3D with Multi-View Diffusion Models. arXiv.cs, 2024. 2, 3 +[19] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T. Barron, and Ben Poole. CAT3D: create anything in 3d with multi-view diffusion models. arXiv, 2405.10314, 2024. 6, 7, 1 +[20] Paul Guerrero. Lang-sam. https://github.com/paulguerrero/lang-sam, 2024.4 +[21] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019. 4 +[22] Agrim Gupta, Piotr Dólar, and Ross B. Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Proc. CVPR, 2019. 2 +[23] Philipp Henzler, Jeremy Reizenstein, Patrick Labatut, Roman Shapovalov, Tobias Ritschel, Andrea Vedaldi, and David Novotny. Unsupervised learning of 3d object categories from videos in the wild. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[24] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 7 +[25] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In Proc. NeurIPS, 2017. 8, 1 +[26] Lukas Hollein, Aljaž Božić, Norman Müller, David Novotny, Hung-Yu Tseng, Christian Richardt, Michael Zollhöfer, and Matthias Nießner. ViewDiff: 3D-Consistent Image Generation with Text-to-Image Models. arXiv preprint, 2024. 3 +[27] Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. In Proc. ICLR, 2024. 2, 3, 6 +[28] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, pages 406-413, 2014. 3, 6, 7 +[29] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian Splating for real-time radiance field rendering. Proc. SIGGRAPH, 42(4), 2023. 2, 4, 5 +[30] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 1 +[31] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. Proc. ICLR, 2015. 1 +[32] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proc. CVPR, 2023. 4 + +[33] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. ABC: A big CAD model dataset for geometric deep learning. In Proc. CVPR, 2019. 3 +[34] Zhengfei Kuang, Yunzhi Zhang, Hong-Xing Yu, Samir Agarwala, Elliott Wu, Jiajun Wu, et al. Stanford-orb: a real-world 3d object inverse rendering benchmark. In NeurIPS, 2023. 5, 7 +[35] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3D: Fast text-to-3D with sparse-view generation and large reconstruction model. Proc. ICLR, 2024. 2, 3, 4, 7 +[36] Joseph J Lim, Hamed Pirsiavash, and Antonio Torralba. Parsing intela objects: Fine pose estimation. In Proceedings of the IEEE international conference on computer vision, pages 2992-2999, 2013. 3 +[37] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. arXiv.cs, abs/2305.08891, 2023. 1 +[38] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In Proc. ECCV, 2014. 4 +[39] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3D object. In Proc. ICCV, 2023. 7 +[40] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-Solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In Proc. NeurIPS, 2022. 1 +[41] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. arXiv preprint, 2023. 5 +[42] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, Natalia Neverova, Andrea Vedaldi, Oran Gafni, and Filippos Kokkinos. IM-3D: Iterative multiview diffusion and reconstruction for high-quality 3D generation. In Proceedings of the International Conference on Machine Learning (ICML), 2024. 3, 7 +[43] Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 2019. 6, 7 +[44] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Carl Yuheng Ren. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception, 2023. 3 +[45] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. In Proc. ICLR, 2023. 2 +[46] Christian Reiser, Richard Szeliski, Dor Verbin, Pratul P. Srinivasan, Ben Mildenhall, Andreas Geiger, Jonathan T. Barron, and Peter Hedman. MERF: memory-efficient radiance fields for real-time view synthesis in unbounded scenes. abs/2302.12249, 2023. 1 + +[47] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common Objects in 3D: Large-scale learning and evaluation of real-life 3D category reconstruction. In Proc. ICCV, 2021. 2, 3, 4, 5, 6 +[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proc. CVPR, 2022. 3, 1 +[49] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In Proc. ICLR, 2022. 1 +[50] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proc. CVPR, 2016. 2, 3 +[51] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In Proc. ECCV, 2016. 3 +[52] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2 +[53] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. MVDream: Multi-view diffusion for 3D generation. In Proc. ICLR, 2024. 2, 3, 7, 1 +[54] Yawar Siddiqui, Filippos Kokkinos, Tom Monnier, Mahendra Kariya, Yanir Kleiman, Emilien Garreau, Oran Gafni, Natalia Neverova, Andrea Vedaldi, Roman Shapovalov, and David Novotny. Meta 3D Asset Gen: Text-to-mesh generation with high-quality geometry, texture, and PBR materials. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2024. 2, 3, 7 +[55] Yawar Siddiqui, Tom Monnier, Filippos Kokkinos, Mahendra Kariya, Yanir Kleiman, Emilien Garreau, Oran Gafni, Natalia Neverova, Andrea Vedaldi, Roman Shapovalov, et al. Meta 3d assetgen: Text-to-mesh generation with high-quality geometry, texture, and pbr materials. NeurIPS, 2024. 2, 3, 4 +[56] Xingyuan Sun, Jiajun Wu, Xiuming Zhang, Zhoutong Zhang, Chengkai Zhang, Tianfan Xue, Joshua B Tenenbaum, and William T Freeman. Pix3d: Dataset and methods for single-image 3d shape modeling. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3 +[57] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Viewset diffusion: (0-)image-conditioned 3D generative models from 2D data. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 2, 1 +[58] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Splatter Image: Ultra-fast single-view 3D reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[59] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. LGM: Large multi-view Gaussian model for high-resolution 3D content creation. arXiv, 2402.05054, 2024. 3 + +[60] Shitao Tang, Jiacheng Chen, Dilin Wang, Chengzhou Tang, Fuyang Zhang, Yuchen Fan, Vikas Chandra, Yasutaka Furukawa, and Rakesh Ranjan. MVDiffusion++: A dense high-resolution multi-view diffusion model for single or sparse-view 3d object reconstruction. arXiv, 2402.12712, 2024. 3 +[61] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. NeurIPS, 2017. 6 +[62] Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. Vggsfm: Visual geometry grounded deep structure from motion. In CVPR, pages 21686-21697, 2024. 2, 4, 5 +[63] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P. Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas A. Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proc. CVPR, 2021. 2 +[64] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. DUSt3R: Geometric 3D vision made easy. In Proc. CVPR, 2024. 2 +[65] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. MeshLRM: large reconstruction model for high-quality mesh. arXiv, 2404.12385, 2024. 3 +[66] Rundi Wu, Chang Xiao, and Changxi Zheng. Deepcad: A deep generative network for computer-aided design models. in 2021 IEEE. In CVF International Conference on Computer Vision (ICCV), pages 6772-6782, 2021. 3 +[67] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P. Srinivasan, Dor Verbin, Jonathan T. Barron, Ben Poole, and Aleksander Holynski. ReconFusion: 3D Reconstruction with Diffusion Priors. arXiv preprint, 2023. 2, 1 +[68] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, et al. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 803-814, 2023. 2, 3, 5, 7 +[69] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguuang Zhang, Xiaou Tang, and Jianxiong Xiao. 3D ShapeNets: A deep representation for volumetric shapes. In Proc. CVPR, 2015. 3 +[70] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond PASCAL: A benchmark for 3D object detection in the wild. In Proc. WACV, 2014. 3 +[71] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. InstantMesh: efficient 3D mesh generation from a single image with sparse-view large reconstruction models. arXiv, 2404.07191, 2024. 3 +[72] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation. arXiv preprint arXiv:2403.14621, 2024. 7 + +[73] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. GRM: Large gaussian reconstruction model for efficient 3D reconstruction and generation. arXiv, 2403.14621, 2024. 3 +[74] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proc. CVPR, 2024. 2 +[75] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, and Angjoo Kanazawa. gsplat: An open-source library for Gaussian splatting. arXiv preprint arXiv:2409.06765, 2024. 5 +[76] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, et al. Mvmgnet: A large-scale dataset of multi-view images. In CVPR, pages 9150-9161, 2023. 2, 3, 5, 6 +[77] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. GS-LRM: large reconstruction model for 3D Gaussian splatting. arXiv, 2404.19702, 2024. 3 +[78] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proc. CVPR, pages 586-595, 2018. 7 +[79] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snively. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 6, 7 +[80] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets Gaussian splatting: Fast and generalizable single-view 3D reconstruction with transformers. arXiv.cs, abs/2312.09147, 2023. 3 \ No newline at end of file diff --git a/2025/UnCommon Objects in 3D/images.zip b/2025/UnCommon Objects in 3D/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c7a8662d94b398853ab38626be95fdc3b79bfb8a --- /dev/null +++ b/2025/UnCommon Objects in 3D/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b3353a92e611b06425ba7719ee1b10b4fcbb9c540e6f99311e9138180cf39fb1 +size 1047458 diff --git a/2025/UnCommon Objects in 3D/layout.json b/2025/UnCommon Objects in 3D/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d6ccfcc4d5bf3b23e0e1b6c0f8239e11b41ce524 --- /dev/null +++ b/2025/UnCommon Objects in 3D/layout.json @@ -0,0 +1,7417 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 222, + 103, + 388, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 103, + 388, + 121 + ], + "spans": [ + { + "bbox": [ + 222, + 103, + 388, + 121 + ], + "type": "text", + "content": "UnCommon Objects in 3D" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 92, + 142, + 520, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 142, + 520, + 194 + ], + "spans": [ + { + "bbox": [ + 92, + 142, + 520, + 194 + ], + "type": "text", + "content": "Xingchen Liu" + }, + { + "bbox": [ + 92, + 142, + 520, + 194 + ], + "type": "inline_equation", + "content": "^{\\infty}" + }, + { + "bbox": [ + 92, + 142, + 520, + 194 + ], + "type": "text", + "content": " Piyush Tayal" + }, + { + "bbox": [ + 92, + 142, + 520, + 194 + ], + "type": "inline_equation", + "content": "^{\\infty}" + }, + { + "bbox": [ + 92, + 142, + 520, + 194 + ], + "type": "text", + "content": " Jianyuan Wang" + }, + { + "bbox": [ + 92, + 142, + 520, + 194 + ], + "type": "inline_equation", + "content": "^{\\infty}" + }, + { + "bbox": [ + 92, + 142, + 520, + 194 + ], + "type": "text", + "content": " Jesus Zarzar" + }, + { + "bbox": [ + 92, + 142, + 520, + 194 + ], + "type": "inline_equation", + "content": "^{+\\infty}" + }, + { + "bbox": [ + 92, + 142, + 520, + 194 + ], + "type": "text", + "content": " Tom Monnier" + }, + { + "bbox": [ + 92, + 142, + 520, + 194 + ], + "type": "inline_equation", + "content": "^{\\infty}" + }, + { + "bbox": [ + 92, + 142, + 520, + 194 + ], + "type": "text", + "content": " Konstantinos Tertikas" + }, + { + "bbox": [ + 92, + 142, + 520, + 194 + ], + "type": "inline_equation", + "content": "^{\\infty*}" + }, + { + "bbox": [ + 92, + 142, + 520, + 194 + ], + "type": "text", + "content": " Jiali Duan" + }, + { + "bbox": [ + 92, + 142, + 520, + 194 + ], + "type": "inline_equation", + "content": "^{\\infty}" + }, + { + "bbox": [ + 92, + 142, + 520, + 194 + ], + "type": "text", + "content": " Antoine Toisoul" + }, + { + "bbox": [ + 92, + 142, + 520, + 194 + ], + "type": "inline_equation", + "content": "^{\\infty}" + }, + { + "bbox": [ + 92, + 142, + 520, + 194 + ], + "type": "text", + "content": " Jason Y. Zhang" + }, + { + "bbox": [ + 92, + 142, + 520, + 194 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 92, + 142, + 520, + 194 + ], + "type": "text", + "content": " Natalia Neverova" + }, + { + "bbox": [ + 92, + 142, + 520, + 194 + ], + "type": "inline_equation", + "content": "^{\\infty}" + }, + { + "bbox": [ + 92, + 142, + 520, + 194 + ], + "type": "text", + "content": " Andrea Vedaldi" + }, + { + "bbox": [ + 92, + 142, + 520, + 194 + ], + "type": "inline_equation", + "content": "^{\\infty}" + }, + { + "bbox": [ + 92, + 142, + 520, + 194 + ], + "type": "text", + "content": " Roman Shapovalov" + }, + { + "bbox": [ + 92, + 142, + 520, + 194 + ], + "type": "inline_equation", + "content": "^{\\infty}" + }, + { + "bbox": [ + 92, + 142, + 520, + 194 + ], + "type": "text", + "content": " David Novotny" + }, + { + "bbox": [ + 92, + 142, + 520, + 194 + ], + "type": "inline_equation", + "content": "^{\\infty}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 209, + 476, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 209, + 476, + 224 + ], + "spans": [ + { + "bbox": [ + 131, + 209, + 476, + 224 + ], + "type": "text", + "content": "*NKUA, Greece †KAUST †Carnegie Mellon University †Meta AI" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 240, + 542, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 240, + 542, + 256 + ], + "spans": [ + { + "bbox": [ + 67, + 240, + 542, + 256 + ], + "type": "text", + "content": "uCO3D https://uco3d.github.io https://github.com/facebookresearch/uco3d" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 57, + 276, + 555, + 555 + ], + "blocks": [ + { + "bbox": [ + 57, + 276, + 555, + 555 + ], + "lines": [ + { + "bbox": [ + 57, + 276, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 57, + 276, + 555, + 555 + ], + "type": "image", + "image_path": "948844b57f01375a98ddd77397dac02a46e50d0415ca6e1316512cc7c65dbf5b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 559, + 555, + 583 + ], + "lines": [ + { + "bbox": [ + 55, + 559, + 555, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 555, + 583 + ], + "type": "text", + "content": "Figure 1. We introduce UnCommon Objects in 3D (uCO3D), a large and diverse dataset of high-quality " + }, + { + "bbox": [ + 55, + 559, + 555, + 583 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 55, + 559, + 555, + 583 + ], + "type": "text", + "content": " videos covering over 1,000 object categories. Each video frame is 3D-annotated with accurate SfM cameras, point cloud, and a 3D Gaussian Splitting reconstruction." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 597, + 201, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 597, + 201, + 611 + ], + "spans": [ + { + "bbox": [ + 151, + 597, + 201, + 611 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "type": "text", + "content": "We introduce Uncommon Objects in 3D (uCO3D), a new object-centric dataset for 3D deep learning and 3D generative AI. uCO3D is the largest publicly-available collection of high-resolution videos of objects with 3D annotations that ensures full- " + }, + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "type": "text", + "content": " coverage. uCO3D is significantly more diverse than MVImgNet and CO3Dv2, covering more than 1,000 object categories. It is also of higher quality, due" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 598, + 556, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 556, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 556, + 696 + ], + "type": "text", + "content": "to extensive quality checks of both the collected videos and the 3D annotations. Similar to analogous datasets, uCO3D contains annotations for 3D camera poses, depth maps and sparse point clouds. In addition, each object is equipped with a caption and a 3D Gaussian Splat reconstruction. We train several large 3D models on MVImgNet, CO3Dv2, and uCO3D and obtain superior results using the latter, showing that uCO3D is better for learning applications." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "14102" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 295, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 295, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 295, + 175 + ], + "type": "text", + "content": "The primacy of data has been the defining characteristic of the last decade of machine learning, alongside deep learning. The most powerful models in language, speech, and computer vision are large deep networks trained on massive amounts of data, and then further fine-tuned on its high-quality subset. This paradigm is expected to extend to all machine learning applications, including 3D vision." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 175, + 295, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 295, + 199 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 295, + 199 + ], + "type": "text", + "content": "However, 3D training data is much harder to come by than data for text, audio, and image processing." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 199, + 295, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 199, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 199, + 295, + 378 + ], + "type": "text", + "content": "Seeking training data for large 3D neural networks [27], many have turned to synthetic datasets like Objaverse [13]. However, synthetic data is a poor substitute for real data in applications like digital twinning, which creates 3D models of real-life objects. This is why many photorealistic reconstruction networks [6, 23, 57, 58, 63, 67] are trained using real object-centric datasets like CO3D [47], MVImgNet [76], GSO [15], and OmniObject3D [68]. Real data is also crucial for generalization, as demonstrated by DUSt3r [64] for point map prediction and DepthAnything [74] for depth prediction, both of which are trained on numerous real datasets. Even non-curated image datasets like the billion-scale LAION [52] are applicable to 3D vision. For instance, text-to-3D generators [35, 45, 53, 54] build on LAION-trained large text-to-image models [5, 11]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 379, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 295, + 498 + ], + "type": "text", + "content": "Given the importance of 3D datasets, but also their relative scarcity, in this paper we ask what is the next step for real data in 3D vision. To answer this question, we note that, while the size of a dataset is crucial, in most cases its quality is just as important. For example, text-to-3D models [35, 55] are notoriously sensitive to fine-tuning data quality, training only with the best-looking models (e.g., Instant3D [35] uses only about " + }, + { + "bbox": [ + 55, + 379, + 295, + 498 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 379, + 295, + 498 + ], + "type": "text", + "content": " of Objaverse). We conclude that high-quality datasets balancing scale and fidelity are more valuable than simply amassing low-quality data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 498, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 629 + ], + "type": "text", + "content": "Based on this observation, we argue that there is a gap in the real object-centric 3D datasets that are currently available, as none strikes the optimal balance between quality and scale. For example, the 3D object scans in OmniObject3D [68] and GSO [15] provide very accurate geometry and textures, but only count a few thousand objects. Conversely, datasets like CO3D [47] and MvImgNet [76] contain orders-of-magnitude more objects, but lack reliable 3D scans. Instead, they provide many views of the objects together with lower-quality 3D cameras and point clouds reconstructed with structure-from-motion (SfM)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "In this paper, we address this gap with a new dataset, Uncommon Objects in 3D (uCO3D), which better balances data quality and size (Tab. 1). Similar to CO3D, it comprises full- " + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": " crowd-sourced videos capturing objects from all sides, annotated with cameras and point clouds using SfM. Furthermore, uCO3D has much greater data diversity (Fig. 2) than prior alternatives as it contains objects" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 71, + 553, + 174 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 553, + 174 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 553, + 174 + ], + "type": "table", + "html": "
RealCount# ClassesData typeAnnotations
ShapeNet [8]X51k553D meshesmesh
Objaverse [13]X800k21k3D meshesmesh
Objaverse-XL [12]X10M2M3D meshesmesh
ABO [10]X8k633D meshesmesh
OmniObject3D [68]✓6k190Videos w/ meshescameras, mesh
GSO [14]✓1k17Views w/ meshescameras, mesh
Objectron [2]✓15k9Limited vp. videoscameras, 3D box
MVImgNet [76]✓220k238Limited vp. videoscameras, pcl
CO3D [47]✓19k50360° videoscameras, pcl
CO3Dv2 [47]✓40k50360° videoscameras, pcl
uCO3D (ours)✓170k1k360° videoscameras, 3DGS, caption
", + "image_path": "f5a214261e7fb05d8096d82a21d1fd86b13c1c40199307bf4a217eb12dcf6ed4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 185, + 553, + 208 + ], + "lines": [ + { + "bbox": [ + 313, + 185, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 313, + 185, + 553, + 208 + ], + "type": "text", + "content": "Table 1. Overview of 3D object datasets. We compare the number of objects / classes, the type of data and associated annotations." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 237, + 555, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 237, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 555, + 346 + ], + "type": "text", + "content": "from the 1,070 visual object categories of the LVIS [22] taxonomy, which has long tails. For reference, MVImgNet and CO3Dv2 contain only 238 and 50 categories, respectively. These fine-grained categories are organized in supercategories, also shown in Fig. 2. Furthermore, uCO3D contains 170k scenes, which is more than four times larger than CO3Dv2's 40k. While this is less than MVImgNet's 220k, uCO3D's videos cover each object from all sides, as opposed to MVImgNet's partial object captures." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "text", + "content": "Besides improving size and diversity, uCO3D also raises the quality bar. This was achieved by checking extensively both the collected videos and their 3D annotations. Differently from datasets like CO3Dv2 that still contain a certain portion of low-quality videos, in uCO3D we manually verified that each video provides full " + }, + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "text", + "content": " turn-table covering all sides of the object. Additionally, " + }, + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "inline_equation", + "content": "60\\%+" + }, + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "text", + "content": " of the videos have " + }, + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "inline_equation", + "content": "1080p+" + }, + { + "bbox": [ + 313, + 350, + 556, + 578 + ], + "type": "text", + "content": " resolution, higher than CO3Dv2. To ensure 3D-annotation quality, we improved both the reconstruction algorithm and the reconstruction validation. For camera reconstruction, we used VGGSfM [62], which is currently the best SfM system available, and is more robust and accurate than COLMAP [50], used in CO3Dv2 and MvImgNet. We also improve on CO3Dv2's active-learning camera quality evaluation by combining it with novel-view synthesis accuracy after reconstructing each scene using 3D Gaussian Splatting (3DGS) [29]. The latter also guarantees that scenes are reconstructible to a high quality, which is important for training of 3D models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "We validate uCO3D's benefits in applications. We train two popular 3D models, LRM [27] and CAT3D [18], using uCO3D and demonstrate improved results compared to training on MVImgNet and CO3Dv2, which makes uCO3D the better data source for real object-centric 3D learning. We also use uCO3D to train a text-to-3D model following Instant3D's [35] two-stage design. The latter requires objects to be rendered from canonical viewpoints, and thus, so far, was limited to synthetic data. By using our 3DGS reconstructions, we 're-shoot' uCO3D's from these viewpoints, which allows to train a more realistic generator." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14103" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 553, + 178 + ], + "type": "image", + "image_path": "1c9a9d2aec5e61eea77375b106e8b94f4d471dc011a18f91314b3318ac994d63.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 186, + 553, + 210 + ], + "lines": [ + { + "bbox": [ + 54, + 186, + 553, + 210 + ], + "spans": [ + { + "bbox": [ + 54, + 186, + 553, + 210 + ], + "type": "text", + "content": "Figure 2. Statistics of uCO3D. (Left) We plot the number of objects per super-category. In total, the dataset contains 50 super-categories, each gathering around 20 sub-categories. (Right) We show a word cloud of all 1,070 visual categories represented in the dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 228, + 142, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 142, + 241 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 142, + 241 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 251, + 296, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 296, + 453 + ], + "type": "text", + "content": "Datasets of synthetic 3D objects. Historically, object-centric 3D datasets have predominantly been synthetic, composed of artist-generated 3D models. A prominent example is ShapeNet [8], with 51,000 meshes across 55 object categories. The meshes have detailed geometries, but relatively simplistic textures. Datasets such as 3DFUTURE [17], IKEA [36], Pix3D [56], and ABO [10] are less diverse, concentrating on furniture and other consumer goods. In contrast, ModelNet [69], DeepCAD [66], and ABC [33] provide CAD models with clean geometry but no texture. Objaverse [13] is perhaps the most impactful dataset following ShapeNet. It is significantly larger, comprising 800,000 3D meshes. Objaverse-XL [12] expands this collection to 10M objects. These datasets have been pivotal in enabling the first 3D deep generative models, including text-to-3D [35, 53, 54] and image-to-3D [6, 27, 73] models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 460, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 568 + ], + "type": "text", + "content": "Real 3D object datasets. Acquiring real-world 3D data presents many challenges, resulting in only a few real 3D-object datasets. Early datasets like Pascal3D [70] include multiple categories but provide only isolated object views and approximate 3D annotations. Conversely, DTU [28], BlendedMVS [51], GSO [15], OmniObject3D [68], Aria Digital Twin [44], and Digital Twin Catalog [1] provide 3D scans of objects, featuring high-quality 3D geometry and textures, but have only a few thousand objects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": "The use of 3D scanners significantly restricts the scale of data acquisition; consequently, other datasets capture multiview turntable-like videos of objects using consumer cameras. CO3D and CO3Dv2 [47] crowd-sourced " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "40,000360^{\\circ}" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " object videos, providing 3D annotations by reconstructing point clouds and cameras using COLMAP SfM [50]. MvImgNet [76] collected even more videos (220,000) across more object categories (238), but their videos capture objects only partially, preventing full reconstruction. Objectron [2] is similar to MvImgNet, but with fewer videos (10,000). A common challenge is that large-scale datasets often rely on SfM for video reconstruction, which can" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 229, + 555, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 229, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 555, + 314 + ], + "type": "text", + "content": "lead to imprecise 3D annotations. uCO3D also employs SfM, but using VGGSfM, which has greater accuracy than COLMAP, and with a more reliable data validation setup. Furthermore, uCO3D is five times larger and significantly more diverse than CO3Dv2, encompassing 20 times more visual categories, and provides caption and 3D Gaussian Splat reconstructions of each object." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 319, + 554, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 554, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 554, + 427 + ], + "type": "text", + "content": "Applications. In order to assess the quality of uCO3D, we measure how it benefits a number of popular downstream applications. First, we consider feedforward few-view 3D reconstruction models. Among those, LRM [27] is a transformer that maps an input image to a neural radiance field supported by a triplane [7]. LightplaneLRM [6] adds splatting layers and a memory-efficient renderer. Further extensions use different representations like 3D Gaussian Splats [59, 73, 77, 80] and meshes [65, 71]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 427, + 555, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 555, + 632 + ], + "type": "text", + "content": "We also consider text-to-3D generators, which create 3D assets from text, and focus on the two-stages approach of Instant3D [35]. This is based on training a text-to-multi-view diffusion model [11, 48] which generates several 2D views of the object, followed by a 3D reconstruction network that outputs the 3D asset, all in a matter of seconds. The multi-view diffusion is improved in ViewDiff [26], MVDiffusion [60], IM-3D [42], CAT3D [18] and many others. AssetGen [55] further extends Instant3D by modelling material properties instead of baking in the radiance function and adds a texture refiner that outputs relightable PBR textures. As an illustration, we use uCO3D to train a model like CAT3D, which results in better new-view synthesis than the one trained on alternative datasets. We also show that the Gaussian Splat reconstructions provided with uCO3D can supervise, for the first time, an Instant3D-like pipeline using solely real-life data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 644, + 463, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 644, + 463, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 644, + 463, + 658 + ], + "type": "text", + "content": "3. Uncommon Objects in 3D" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "content": "In this section, we introduce uCO3D, our new dataset of real-life 3D objects. uCO3D comprises " + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "content": " turn-table-like videos of objects, crowdsourced and annotated with 3D cameras, point clouds, 3D Gaussians, and textual captions." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14104" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 89, + 297 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 89, + 297 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 89, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 89, + 297 + ], + "type": "image", + "image_path": "1fe67a25a5d1ab72ff15d61286823e3b7b97aa7df867541a99fc374591138628.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 298, + 83, + 306 + ], + "lines": [ + { + "bbox": [ + 64, + 298, + 83, + 306 + ], + "spans": [ + { + "bbox": [ + 64, + 298, + 83, + 306 + ], + "type": "text", + "content": "Video" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 91, + 71, + 252, + 296 + ], + "blocks": [ + { + "bbox": [ + 91, + 71, + 252, + 296 + ], + "lines": [ + { + "bbox": [ + 91, + 71, + 252, + 296 + ], + "spans": [ + { + "bbox": [ + 91, + 71, + 252, + 296 + ], + "type": "image", + "image_path": "0190d353e987df0730825c5db96919ce00b367b182a5e27137b0c4d88a877fe5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 297, + 216, + 306 + ], + "lines": [ + { + "bbox": [ + 127, + 297, + 216, + 306 + ], + "spans": [ + { + "bbox": [ + 127, + 297, + 216, + 306 + ], + "type": "text", + "content": "a) Sparse SfM reconstruction" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 253, + 70, + 405, + 297 + ], + "blocks": [ + { + "bbox": [ + 253, + 70, + 405, + 297 + ], + "lines": [ + { + "bbox": [ + 253, + 70, + 405, + 297 + ], + "spans": [ + { + "bbox": [ + 253, + 70, + 405, + 297 + ], + "type": "image", + "image_path": "0b7647e4b46ffe6a23cc4a62672c65288cebac29a5fd0de5ff4f01fd177cfb0a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 285, + 297, + 373, + 306 + ], + "lines": [ + { + "bbox": [ + 285, + 297, + 373, + 306 + ], + "spans": [ + { + "bbox": [ + 285, + 297, + 373, + 306 + ], + "type": "text", + "content": "b) Densified SfM point cloud" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 408, + 77, + 553, + 295 + ], + "blocks": [ + { + "bbox": [ + 408, + 77, + 553, + 295 + ], + "lines": [ + { + "bbox": [ + 408, + 77, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 408, + 77, + 553, + 295 + ], + "type": "image", + "image_path": "50e0c9f3d8d42436ee19e288476b00f646d7b940fa2509d7b181deced7d14e20.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 445, + 297, + 512, + 306 + ], + "lines": [ + { + "bbox": [ + 445, + 297, + 512, + 306 + ], + "spans": [ + { + "bbox": [ + 445, + 297, + 512, + 306 + ], + "type": "text", + "content": "c) 3D Gaussian Splats" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 310, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 55, + 310, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 555, + 344 + ], + "type": "text", + "content": "Figure 3. Data annotation overview. Each scene in uCO3D is reconstructed in three different ways: a) per-frame cameras with sparse point cloud calculated by VGGSfM [62], b) semi-dense point cloud comprising triangulations of additional denser tracks from VGGSfM's tracker, c) 3D Gaussian Splat [29] reconstruction optimized separately for each scene." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 69, + 362, + 282, + 388 + ], + "blocks": [ + { + "bbox": [ + 69, + 362, + 282, + 388 + ], + "lines": [ + { + "bbox": [ + 69, + 362, + 282, + 388 + ], + "spans": [ + { + "bbox": [ + 69, + 362, + 282, + 388 + ], + "type": "image", + "image_path": "335b00b451c85871155c25bb326249a4455b9d40dd1c807515e4153ae4468f97.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 388, + 202, + 398 + ], + "lines": [ + { + "bbox": [ + 149, + 388, + 202, + 398 + ], + "spans": [ + { + "bbox": [ + 149, + 388, + 202, + 398 + ], + "type": "text", + "content": "a) Input video" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 58, + 400, + 294, + 467 + ], + "blocks": [ + { + "bbox": [ + 58, + 400, + 294, + 467 + ], + "lines": [ + { + "bbox": [ + 58, + 400, + 294, + 467 + ], + "spans": [ + { + "bbox": [ + 58, + 400, + 294, + 467 + ], + "type": "image", + "image_path": "7b94ce4c01501416c8417140602cdde0b2a17cc8c0bd545c34466c70c2960190.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 119, + 468, + 233, + 478 + ], + "lines": [ + { + "bbox": [ + 119, + 468, + 233, + 478 + ], + "spans": [ + { + "bbox": [ + 119, + 468, + 233, + 478 + ], + "type": "text", + "content": "b) Sine-wave camera trajectory" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 480, + 295, + 503 + ], + "lines": [ + { + "bbox": [ + 55, + 480, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 295, + 503 + ], + "type": "text", + "content": "Figure 4. Data collection example. For each video, the cameras follow a sine-wave trajectory to ensure good viewpoint coverage." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 530, + 295, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 295, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 295, + 686 + ], + "type": "text", + "content": "Compared to older datasets like CO3Dv2 [47], uCO3D comes with many improvements. First, uCO3D is much larger and more diverse than CO3Dv2: it contains more than 1k different categories and more than 170k objects, compared to the 50 and 38k of CO3Dv2. While CO3Dv2's categories are taken from MS COCO [38], the categories in uCO3D are taken from the LVIS [21] taxonomy. Hence, we inherit the LVIS focus on covering the long-tail of the visual-category distribution. To simplify data analysis, we grouped the " + }, + { + "bbox": [ + 55, + 530, + 295, + 686 + ], + "type": "inline_equation", + "content": "1\\mathrm{k}+" + }, + { + "bbox": [ + 55, + 530, + 295, + 686 + ], + "type": "text", + "content": " LVIS categories to 50 super-categories, each containing approximately 20 subcategories. Figure 2 shows the number of videos collected per super-category, and the LVIS category distribution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Second, uCO3D improves quality significantly compared to CO3Dv2, ensuring that videos are of high reso" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 365, + 555, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 473 + ], + "type": "text", + "content": "lution, cover each object well, and that the 3D annotations are accurate. uCO3D also contains rich textual descriptions of each object, missing in other datasets, and useful to train large generative models. It also comes with additional 3D Gaussian Splat reconstructions of all objects, each rigidly aligned to a canonical object-centric reference, which make it possible to re-render the dataset from a fixed, canonical set of cameras, simulating synthetic data acquisition, which is very useful for training generative models [35, 55]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 478, + 556, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 556, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 556, + 624 + ], + "type": "text", + "content": "Dataset collection. Videos of objects were captured by workers on Amazon Mechanical Turk. To ensure high video quality, workers were required to submit videos of a sufficient resolution. As a result, more than " + }, + { + "bbox": [ + 313, + 478, + 556, + 624 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 313, + 478, + 556, + 624 + ], + "type": "text", + "content": " of videos in uCO3D are of 1080p resolution or higher, compared to " + }, + { + "bbox": [ + 313, + 478, + 556, + 624 + ], + "type": "inline_equation", + "content": "33\\%" + }, + { + "bbox": [ + 313, + 478, + 556, + 624 + ], + "type": "text", + "content": " in CO3Dv2. Furthermore, to aid the 3D reconstruction, workers followed a sine-wave capture trajectory instead of the plain circular trajectory of CO3Dv2, ensuring varying camera elevations (cf. Fig. 4). Finally, each video was individually manually assessed to make sure that it adheres to these requirements, a process more rigorous than the rough eyeballing used in CO3Dv2 [47]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 628, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 555, + 688 + ], + "type": "text", + "content": "Video object segmentation. We used text-conditioned Segment-Anything (langSAM) [20, 32] to segment the object of interest in each video frame given text-conditioning in form of the object-category name, which had been provided by Turkers at collection time." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "To improve frame-to-frame consistency, CO3Dv2 used a simple Viterbi algorithm, which often led to segmentation" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14105" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 545, + 385 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 545, + 385 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 545, + 385 + ], + "type": "image", + "image_path": "2506e58f0b6465728af23e0f29d70188eb225e0b69f86010fd0a44e4e085c24f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 391, + 553, + 403 + ], + "lines": [ + { + "bbox": [ + 56, + 391, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 391, + 553, + 403 + ], + "type": "text", + "content": "Figure 5. 3D reconstruction comparison. We show results of LightplaneLRM [6] models trained on MVIImgNet, CO3Dv2 and uCO3D." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 422, + 294, + 487 + ], + "blocks": [ + { + "bbox": [ + 56, + 422, + 294, + 487 + ], + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 487 + ], + "type": "table", + "html": "
Train datasetOmniObject3DStanfordORB
LPIPS↓PSNR↑IoU↑LPIPS↓PSNR↑IoU↑
MVIImgNet [76]0.10923.390.9280.07024.4510.939
CO3Dv2 [47]0.09523.620.9260.05625.6170.956
uCO3D (ours)0.09324.610.9460.05725.7150.957
", + "image_path": "ff988240d2de12af08ea30d5bf6e666e3b9268cf1a2e1afe13bcc52d04e969e3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 495, + 295, + 540 + ], + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 540 + ], + "type": "text", + "content": "Table 2. 3D reconstruction benchmark. We compare LightplaneLRM [6] models trained on CO3Dv2, MVImgNet, and uCO3D. We report novel-view synthesis performances on OmniObject3D [68] and StanfordORB [34]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 563, + 295, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 295, + 612 + ], + "type": "text", + "content": "flickering, impairing the final 3D reconstruction quality. Instead, in uCO3D, we refine the SAM segmentations with state-of-the-art deep video-segmenter based on XMem [9], leading to more temporally-stable object segmentations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "content": "3D annotation with VGGSfM. For each video, we use the state-of-the-art VGGSfM [62] Structure from Motion (SfM) system to estimate the parameters of the cameras (intrinsic and extrinsic) for 200 frames sampled uniformly. VGGSfM also outputs a sparse 3D point cloud, and its denser version obtained by triangulating additional 3D points from VGGSfM's tracker. Examples of sparse and densified SfM point clouds are shown in Fig. 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 424, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 555, + 510 + ], + "type": "text", + "content": "Scene alignment. While the coordinate system of VG-GSfM cameras is defined only up to a rigid transformation, it is crucial for applications like generation and reconstruction to train on a dataset of rigidly aligned objects. We thus align all objects so they have a horizontal ground plane, similar scale, centring, and orientation. Details of the scene alignment procedure are in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 514, + 555, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 555, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 555, + 575 + ], + "type": "text", + "content": "Gaussian Splat reconstruction. Sparse and even dense SfM point clouds provide an accurate but still quite sparse 3D reconstruction of the scene's surface. To further densify it, uCO3D provides a 3D Gaussian Splat reconstruction [29] for each scene, fitted using gsplat [75]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 580, + 556, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 556, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 556, + 641 + ], + "type": "text", + "content": "Scene captioning. uCO3D also provides textual captions for all scenes, useful for generative modelling. Motivated by Cap3D [41], we first caption each view separately using a vision-language model, and then summarise these into a single scene caption using LLAMA3 [16]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 656, + 395, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 656, + 395, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 656, + 395, + 670 + ], + "type": "text", + "content": "4. Applications" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "In this section, we demonstrate uCO3D's merit on three popular 3D learning tasks: feedforward sparse-view 3D reconstruction (Sec. 4.1), new-view synthesis using diffusion" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14106" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 550, + 370 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 550, + 370 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 550, + 370 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 550, + 370 + ], + "type": "image", + "image_path": "7fbbd9a4c29a229d341999b5d317e801372b7d676572ace261639cfc958e115f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 377, + 555, + 401 + ], + "lines": [ + { + "bbox": [ + 55, + 377, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 555, + 401 + ], + "type": "text", + "content": "Figure 6. Novel-view synthesis comparison. We compare results of CAT3D-like [19] models trained on different datasets (MVIImgNet, CO3D, uCO3D) and evaluated on standard NVs datasets (top-to-bottom: RealEstate10K [79], LLFF [43], DTU [28], Mip-NeRF 360 [4])." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 419, + 294, + 487 + ], + "blocks": [ + { + "bbox": [ + 56, + 419, + 294, + 487 + ], + "lines": [ + { + "bbox": [ + 56, + 419, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 419, + 294, + 487 + ], + "type": "table", + "html": "
Train datasetEasierDataset DifficultyHarder
←Re10K [79]LLFF [43]DTU [28]Mip-NeRF [4]
LPIPS↓PSNR↑LPIPS↓PSNR↑LPIPS↓PSNR↑LPIPS↓PSNR↑
MVImgNet [76]0.31018.770.42614.380.37712.790.60512.39
CO3Dv2 [47]0.28120.020.41814.950.32916.420.53214.19
uCO3D (ours)0.27819.770.41815.160.31516.970.52814.37
", + "image_path": "9f4c70c27f9387c7a3ca326ee1e7b84b4f9891d0e857220e6eec4963868d70dc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 494, + 295, + 540 + ], + "lines": [ + { + "bbox": [ + 55, + 494, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 295, + 540 + ], + "type": "text", + "content": "Table 3. Novel-view synthesis benchmark. We evaluate CAT3D-like [19] models trained on MVImgNet, CO3Dv2 or uCO3D. We report NVS performances on RealEstate10K [79], LLFF [43], DTU [28] and Mip-NeRF 360 [4]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 562, + 204, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 204, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 204, + 574 + ], + "type": "text", + "content": "(Sec. 4.2), and text-to-3D (Sec. 4.3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 586, + 247, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 247, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 247, + 598 + ], + "type": "text", + "content": "4.1. Few-view 3D Object Reconstruction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 604, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 295, + 676 + ], + "type": "text", + "content": "Traditionally, multi-view 3D-annotated datasets such as CO3D or MVImgNet have been used to supervise few-view 3D reconstructors. In this section, we train LightplaneLRM [6], an evolution of the seminal LRM [27], and show that doing so on uCO3D leads to better performance than training on alternative datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 712 + ], + "type": "text", + "content": "LRM is a large transformer [61] that accepts few input images of an object and predicts a 3D representation of the latter. The transformer, conditioned on the tokens of the ob" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 422, + 555, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 555, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 555, + 481 + ], + "type": "text", + "content": "served images via cross attention, converts a set of learnable input tokens to a 3D representation. The 3D representation is a triplane [7] supporting an opacity/radiance implicit shape. LightplaneLRM improves LRM by adding so called \"splattering layers\" and a memory-efficient renderer." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 483, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 556, + 616 + ], + "type": "text", + "content": "During training, LightplaneLRM receives four random source frames from a training uCO3D video sequence, and renders the predicted triplane into held-out target views. Learning minimizes the photometric loss between the renders and the corresponding ground-truth targets. Both source and target views are masked using the extracted segmentation masks to make sure that LightplaneLRM only reconstructs the foreground object. Training uses the Adam optimizer and is warm-started following the original LRM training protocol by pre-training the model on a large dataset of synthetic objects similar to Objaverse [13]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 622, + 556, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 622, + 556, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 622, + 556, + 682 + ], + "type": "text", + "content": "Baselines. Our main goal is to demonstrate that uCO3D contains higher quality data than existing object-centric datasets. As such, starting from the model pre-trained on the synthetic data, we finetune either on uCO3D, or on two other baseline datasets, namely MVImgNet and CO3Dv2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Evaluation protocol. We evaluate each trained model in a novel-view synthesis setting on two small-scale high-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14107" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "quality object-centric datasets: OmniObject3D [68] and Stanford-ORB [34]. Given four views of a held-out test scene, the model reconstructs the scene which is then rendered to unseen target views. We report the average LPIPS [78] loss and Peak-signal-to-noise ratio (PSNR) between each render and the corresponding ground-truth image. We also report the intersection-over-union (IoU) between the rendered object alpha mask and the target view segmentation mask." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 189, + 296, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 189, + 296, + 260 + ], + "spans": [ + { + "bbox": [ + 55, + 189, + 296, + 260 + ], + "type": "text", + "content": "Results. Table 2 and Fig. 5 report the quantitative and qualitative results, respectively. The LightplaneLRM trained on uCO3D is better than the other baselines in most metrics on both datasets. The latter confirms that uCO3D is currently the most reliable source of real data for training feedforward few-view 3D reconstructors." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 280, + 247, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 280, + 247, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 247, + 293 + ], + "type": "text", + "content": "4.2. Novel-view synthesis using diffusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 301, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 296, + 445 + ], + "type": "text", + "content": "We now consider application of uCO3D to training new-view image diffusion generators. These generators can, given one or a few views of an object and a target camera pose, output new arbitrary views as observed from the target camera, hallucinating missing details based on a statistical prior they learn. They can thus complement and integrate the feed-forward reconstruction models of the previous section, which are deterministic and thus unable to deal with ambiguity well. To this end, we train a diffusion model similar to the recent CAT3D [19], but reimplement it from scratch given lack of source code (see details in the supplementary). We call this model CAT3D-like." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 453, + 296, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 296, + 634 + ], + "type": "text", + "content": "Evaluation protocol. As in Sec. 4.1, we compare versions of CAT3D-like trained using uCO3D, MVImgNet, and CO3Dv2 and test them on held-out datasets. A feature of CAT3D is the ability to reconstruct both the principal object in the images as well as the background. We thus benchmark the method using new-view synthesis datasets that do contain background, namely DTU [28], containing structured light scans of various objects, LLFF [43], containing scenes captured from fronto-parallel camera trajectories, RealEstate10k [79], containing real-estate walk-throughs, and Mip-NeRF 360 [3] with complex indoor and outdoor scenes. For evaluation, we take three known views as input and use the model to predict a new view. We report LPIPS and PSNR but not the IoU since CAT3D only generates new RGB views without reconstructing the 3D shape." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Results. Table 3 and Fig. 6 contain the results: training CAT3D-like on uCO3D leads to the best performance across all four datasets. Even when compared to MVImgNet, which is slightly larger than uCO3D, the latter improves PSNR by 3-4 points, and reduces the LPIPS error by " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 315, + 71, + 555, + 285 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 555, + 285 + ], + "type": "image", + "image_path": "0288ee55c550563816bdacb45ca723a0fb48b07c37bc1972bd1fc319bb942999.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 293, + 555, + 337 + ], + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 337 + ], + "type": "text", + "content": "Figure 7. Supervising Instant3D with 3DGS. For each training scene, its 3DGS is rendered from 4 canonical views yielding a captioned image dataset for finetuning an image diffuser. Samples from the latter are then reconstructed with LRM." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 361, + 454, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 361, + 454, + 373 + ], + "spans": [ + { + "bbox": [ + 314, + 361, + 454, + 373 + ], + "type": "text", + "content": "4.3. Photorealistic Text-to-3D" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 381, + 555, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 555, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 555, + 561 + ], + "type": "text", + "content": "Next, we show that uCO3D enables training photorealistic text-to-3D generators. Methods like CAT3D and others [39, 42, 53] generate several views of the object first, and then fit a 3D model, such as a NeRF or 3DGS, via optimization. This can work well, but it is not particularly robust or fast. An alternative, popularized by Instant3D [35] and follow-ups [54, 72], is to use a feedforward reconstructor in the second step, similar to LightplaneLRM from Sec. 4.1, which is faster and more robust. However, these models require canonical views of the objects — for example, Instant3D considers 4 orthogonal viewpoints, covering all 'sides'. The requirement of such training canonical views complicates training on real data, where viewpoints are arbitrary, and explains why such models are usually trained on synthetic data, limiting realism." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 567, + 555, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 555, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 555, + 687 + ], + "type": "text", + "content": "Imaging 3DGS from canonical views. Our new idea is to 're-shoot' the 3DGS reconstructions provided with uCO3D from canonical viewpoints, making our data compatible with any method requiring canonical views for training (Fig. 7). To do so, we render the normalized reference frames (Sec. 3) into four views for each object, and arrange them in a grid as a target for the text-to-4-view generator. We double check the quality of the renders by calculating their CLIP similarity [24] to the object caption, and discard a sample if this is below 0.3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "We use this data to fine-tune a text-to-4-view image diffusion model using these 4-view image grids and the corre" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14108" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 68, + 555, + 324 + ], + "blocks": [ + { + "bbox": [ + 58, + 68, + 555, + 324 + ], + "lines": [ + { + "bbox": [ + 58, + 68, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 58, + 68, + 555, + 324 + ], + "type": "image", + "image_path": "e2b0fa420a90fe166e3f20c3ef071f3b28905f69dddf06737320c964708c87d4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 331, + 555, + 365 + ], + "lines": [ + { + "bbox": [ + 55, + 331, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 555, + 365 + ], + "type": "text", + "content": "Figure 8. Qualitative results for text-to-3D generation displaying the 4-view grids generated by our Instant3D-like model given the input caption, and the 3D asset obtained by reconstructing the latter. The 4-view grid generator was trained using the canonical 4-view renders of uCO3D's 3DGS scene reconstructions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 95, + 383, + 257, + 426 + ], + "blocks": [ + { + "bbox": [ + 95, + 383, + 257, + 426 + ], + "lines": [ + { + "bbox": [ + 95, + 383, + 257, + 426 + ], + "spans": [ + { + "bbox": [ + 95, + 383, + 257, + 426 + ], + "type": "table", + "html": "
Train datasetReal - FID↓Surreal - FID↓
Synthetic82.842.3
uCO3D (ours)63.968.9
", + "image_path": "5e45e4f957637f0c5999755a3825c467ca6ccaaefae7eb53ea4bb5c09c1ee729.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 434, + 295, + 478 + ], + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 478 + ], + "type": "text", + "content": "Table 4. Text-to-3D evaluation. We compare Instant3D-like models trained on uCO3D or a dataset of synthetic renders from artist-created meshes. We report FID on two sets of data corresponding to real and surreal objects, see text for details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 499, + 295, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 499, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 295, + 558 + ], + "type": "text", + "content": "sponding scene captions. At inference time, given a caption describing the desired object, we use the model to sample a new 4-view grid and feed the latter, together with the corresponding cameras, to the LightplaneLRM model (Sec. 4.1) for 3D reconstruction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 564, + 295, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 295, + 612 + ], + "type": "text", + "content": "Baselines. We train another 4-view generator on a dataset of synthetic assets similar to Objaverse [13] and use it with the original LightplaneLRM model [6] trained on the same data and thus optimally matched to it." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "content": "Evaluation protocol. We report metrics evaluating the alignment between the distributions of the generations and the ground-truth objects. Specifically, we report the Frechet Inception Distance (FID) [25] between the renders of the generated 3D shapes and images of ground-truth objects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "The main purpose of this experiment is to show that, by training on the uCO3D dataset, the 3D generations are more realistic. We assess this using two sets of prompts: Sur-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 386, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 555, + 445 + ], + "type": "text", + "content": "real, containing 100 captions of objects from the synthetic dataset, and Real, containing 100 random captions form the held-out evaluation sequences of uCO3D. We report FID between the generated 3D shapes and the images/renders corresponding to the objects of each prompt-set." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 451, + 555, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 555, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 555, + 523 + ], + "type": "text", + "content": "Results. Table 4 and Fig. 8 contain the quantitative and qualitative results respectively. The table reveals that the uCO3D-trained generator outperforms the synthetic generator when evaluated on real prompts. The latter verifies our hypothesis that a generator trained on uCO3D yields more realistic samples than a model trained on synthetic data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 538, + 391, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 391, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 391, + 551 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 559, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 555, + 713 + ], + "type": "text", + "content": "We have introduced uCO3D, a new object-centric 3D dataset of real-life objects. uCO3D strikes a balance between size and quality, ensuring the quality of the collected turntable-like videos and of the 3D annotations, while at the same time significantly expanding the scale of the data compared to CO3Dv2 and the diversity compared to CO3Dv2 and MVImgNet. We have shown the benefits of using this dataset compared to alternatives when training models for feedforward few-view 3D reconstruction, multi-view generation, and text-to-3D generation. Equipped with 3D cameras, point clouds, masks, textual captions, and 3DGS reconstructions of objects, uCO3D is a ready-to-use resource for training large generative models and for exploring 3D deep learning." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14109" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 114, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 114, + 294, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 114, + 294, + 146 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 294, + 146 + ], + "type": "text", + "content": "[1] Aria digital twin catalog. https://www.projectaria.com/datasets/dtc/. Accessed: 2024-12-20.3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 294, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 294, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 294, + 213 + ], + "type": "text", + "content": "[2] Adel Ahmadyan, Liangkai Zhang, Artsiom Ablavatski, Jianing Wei, and Matthias Grundmann. Objectron: A large scale dataset of object-centric videos in the wild with pose annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7822-7831, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 214, + 295, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 295, + 258 + ], + "type": "text", + "content": "[3] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proc. ICCV, 2021. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 259, + 295, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 259, + 295, + 292 + ], + "spans": [ + { + "bbox": [ + 62, + 259, + 295, + 292 + ], + "type": "text", + "content": "[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In Proc. CVPR, 2022. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 293, + 294, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 293, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 293, + 294, + 357 + ], + "type": "text", + "content": "[5] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv.cs, abs/2311.15127, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 358, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 358, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 358, + 294, + 392 + ], + "type": "text", + "content": "[6] Ang Cao, Justin Johnson, Andrea Vedaldi, and David Novotny. Lightplane: Highly-scalable components for neural 3d fields. arXiv, 2024. 2, 3, 5, 6, 8, 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 392, + 294, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 294, + 457 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 294, + 457 + ], + "type": "text", + "content": "[7] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In Proc. CVPR, 2022. 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 459, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 294, + 514 + ], + "type": "text", + "content": "[8] Angel X. Chang, Thomas A. Funkhouser, Leonidas J. Guibas, Pat Hanrahan, Qi-Xing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet an information-rich 3d model repository. arXiv.cs, abs/1512.03012, 2015. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 514, + 294, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 294, + 559 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 294, + 559 + ], + "type": "text", + "content": "[9] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In European Conference on Computer Vision, pages 640-658. Springer, 2022. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 559, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 559, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 559, + 294, + 635 + ], + "type": "text", + "content": "[10] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21126-21136, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 636, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 294, + 713 + ], + "type": "text", + "content": "[11] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam S. Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, Matthew Yu, Abhishek Kadian, Filip Radenovic, Dhruv Mahajan, Kunpeng Li, Yue Zhao, Vladan Petrovic, Mitesh Kumar Singh, Simran Motwani, Yi Wen, Yiwen Song, Roshan Sumbaly, Vignesh Ramanathan, Zijian He, Peter Vajda, and Devi Parikh. Emu:" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Enhancing image generation models using photogenic needles in a haystack. CoRR, abs/2309.15807, 2023. 2, 3, 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 96, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 171 + ], + "type": "text", + "content": "[12] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, Eli VanderBilt, Aniruddha Kembhavi, Carl Vondrick, Georgia Gkioxari, Kiana Ehsani, Ludwig Schmidt, and Ali Farhadi. Objverse-XL: A universe of " + }, + { + "bbox": [ + 317, + 96, + 553, + 171 + ], + "type": "inline_equation", + "content": "10\\mathrm{M} + 3\\mathrm{D}" + }, + { + "bbox": [ + 317, + 96, + 553, + 171 + ], + "type": "text", + "content": " objects. CoRR, abs/2307.05663, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 173, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 228 + ], + "type": "text", + "content": "[13] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In Proc. CVPR, 2023. 2, 3, 6, 8, 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 229, + 553, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 229, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 316, + 229, + 553, + 294 + ], + "type": "text", + "content": "[14] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 295, + 553, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 295, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 316, + 295, + 553, + 349 + ], + "type": "text", + "content": "[15] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B. McHugh, and Vincent Vanhoucke. Google Scanned Objects: A high-quality dataset of 3D scanned household items. In Proc. ICRA, 2022. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 351, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 351, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 351, + 553, + 668 + ], + "type": "text", + "content": "[16] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aurélien Rodriguez, Austen Gregerson, Ava Spataru, Baptiste Rozière, Bethany Biron, Binh Tang, Bobbie Chern, Charlotte Caucheteux, Chaya Nayak, Chloe Bi, Chris Marra, Chris McConnell, Christian Keller, Christophe Touret, Chunyang Wu, Corinne Wong, Cristian Canton Ferrer, Cyrus Nikolaidis, Damien Allonsius, Daniel Song, Danielle Pintz, Danny Livshits, David Esiobu, Dhruv Choudhary, Dhruv Mahajan, Diego Garcia-Olano, Diego Perino, Dieuwke Hupkes, Egor Lakomkin, Ehab AlBadawy, Elina Lobanova, Emily Dinan, Eric Michael Smith, Filip Radenovic, Frank Zhang, Gabriel Synnaeve, Gabrielle Lee, Georgia Lewis Anderson, Graeme Nail, Grégoire Mialon, Guan Pang, Guillem Cucurell, Hailey Nguyen, Hannah Ko-revaar, Hu Xu, Hugo Touvron, Iliyan Zarov, Imanol Arrieta Ibarra, Isabel M. Kloumann, Ishan Misra, Ivan Evtimov, Jade Copet, Jaewon Lee, Jan Geffert, Jana Vranes, Jason Park, Jay Mahadeokar, Jeet Shah, Jelmer van der Linde, Jennifer Billock, Jenny Hong, Jenya Lee, Jeremy Fu, Jianfeng Chi, Jianyu Huang, Jiawen Liu, Jie Wang, Jiecao Yu, Joanna Bitton, Joe Spisak, Jongsoo Park, Joseph Rocca, Joshua Johnstun, Joshua Saxe, Junteng Jia, Kalyan Vasuden Alwala Kartikeya Upasani Kate Plawiak Ke Li Kenneth Heafield and Kevin Stone. The Llama 3 herd of models. arXiv 2407.21783, 2024. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[17] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d furniture shape with texture. International Journal of Computer Vision, 129:3313-3337, 2021. 3" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14110" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[18] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T. Barron, and Ben Poole. CAT3D: Create Anything in 3D with Multi-View Diffusion Models. arXiv.cs, 2024. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 296, + 183 + ], + "type": "text", + "content": "[19] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T. Barron, and Ben Poole. CAT3D: create anything in 3d with multi-view diffusion models. arXiv, 2405.10314, 2024. 6, 7, 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 296, + 205 + ], + "type": "text", + "content": "[20] Paul Guerrero. Lang-sam. https://github.com/paulguerrero/lang-sam, 2024.4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 296, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 296, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 296, + 249 + ], + "type": "text", + "content": "[21] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 282 + ], + "type": "text", + "content": "[22] Agrim Gupta, Piotr Dólar, and Ross B. Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Proc. CVPR, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 283, + 296, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 296, + 348 + ], + "type": "text", + "content": "[23] Philipp Henzler, Jeremy Reizenstein, Patrick Labatut, Roman Shapovalov, Tobias Ritschel, Andrea Vedaldi, and David Novotny. Unsupervised learning of 3d object categories from videos in the wild. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 349, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 349, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 349, + 294, + 392 + ], + "type": "text", + "content": "[24] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "type": "text", + "content": "[25] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In Proc. NeurIPS, 2017. 8, 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 294, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 482 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 482 + ], + "type": "text", + "content": "[26] Lukas Hollein, Aljaž Božić, Norman Müller, David Novotny, Hung-Yu Tseng, Christian Richardt, Michael Zollhöfer, and Matthias Nießner. ViewDiff: 3D-Consistent Image Generation with Text-to-Image Models. arXiv preprint, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 483, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 483, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 483, + 294, + 525 + ], + "type": "text", + "content": "[27] Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: Large reconstruction model for single image to 3D. In Proc. ICLR, 2024. 2, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 525, + 294, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 294, + 559 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 294, + 559 + ], + "type": "text", + "content": "[28] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In CVPR, pages 406-413, 2014. 3, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "type": "text", + "content": "[29] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian Splating for real-time radiance field rendering. Proc. SIGGRAPH, 42(4), 2023. 2, 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 603, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 635 + ], + "type": "text", + "content": "[30] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 658 + ], + "type": "text", + "content": "[31] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. Proc. ICLR, 2015. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[32] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proc. CVPR, 2023. 4" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 117 + ], + "type": "text", + "content": "[33] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. ABC: A big CAD model dataset for geometric deep learning. In Proc. CVPR, 2019. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 118, + 554, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 160 + ], + "type": "text", + "content": "[34] Zhengfei Kuang, Yunzhi Zhang, Hong-Xing Yu, Samir Agarwala, Elliott Wu, Jiajun Wu, et al. Stanford-orb: a real-world 3d object inverse rendering benchmark. In NeurIPS, 2023. 5, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 161, + 554, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 161, + 554, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 161, + 554, + 216 + ], + "type": "text", + "content": "[35] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3D: Fast text-to-3D with sparse-view generation and large reconstruction model. Proc. ICLR, 2024. 2, 3, 4, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 217, + 555, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 555, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 555, + 259 + ], + "type": "text", + "content": "[36] Joseph J Lim, Hamed Pirsiavash, and Antonio Torralba. Parsing intela objects: Fine pose estimation. In Proceedings of the IEEE international conference on computer vision, pages 2992-2999, 2013. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 261, + 554, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 554, + 293 + ], + "type": "text", + "content": "[37] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. arXiv.cs, abs/2305.08891, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 294, + 554, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 554, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 554, + 337 + ], + "type": "text", + "content": "[38] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In Proc. ECCV, 2014. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 338, + 554, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 554, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 554, + 371 + ], + "type": "text", + "content": "[39] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3D object. In Proc. ICCV, 2023. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 372, + 554, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 554, + 415 + ], + "type": "text", + "content": "[40] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-Solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In Proc. NeurIPS, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 415, + 554, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 554, + 449 + ], + "type": "text", + "content": "[41] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. arXiv preprint, 2023. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 449, + 554, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 554, + 513 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 554, + 513 + ], + "type": "text", + "content": "[42] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, Natalia Neverova, Andrea Vedaldi, Oran Gafni, and Filippos Kokkinos. IM-3D: Iterative multiview diffusion and reconstruction for high-quality 3D generation. In Proceedings of the International Conference on Machine Learning (ICML), 2024. 3, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 514, + 554, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 554, + 570 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 554, + 570 + ], + "type": "text", + "content": "[43] Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 2019. 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 571, + 554, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 554, + 624 + ], + "type": "text", + "content": "[44] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Carl Yuheng Ren. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 625, + 554, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 554, + 657 + ], + "type": "text", + "content": "[45] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. In Proc. ICLR, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "type": "text", + "content": "[46] Christian Reiser, Richard Szeliski, Dor Verbin, Pratul P. Srinivasan, Ben Mildenhall, Andreas Geiger, Jonathan T. Barron, and Peter Hedman. MERF: memory-efficient radiance fields for real-time view synthesis in unbounded scenes. abs/2302.12249, 2023. 1" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14111" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[47] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common Objects in 3D: Large-scale learning and evaluation of real-life 3D category reconstruction. In Proc. ICCV, 2021. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 173 + ], + "type": "text", + "content": "[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proc. CVPR, 2022. 3, 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 295, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 197 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 197 + ], + "type": "text", + "content": "[49] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In Proc. ICLR, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 229 + ], + "type": "text", + "content": "[50] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proc. CVPR, 2016. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 295, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 295, + 264 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 295, + 264 + ], + "type": "text", + "content": "[51] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In Proc. ECCV, 2016. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 266, + 295, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 266, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 266, + 295, + 331 + ], + "type": "text", + "content": "[52] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 333, + 295, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 295, + 366 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 295, + 366 + ], + "type": "text", + "content": "[53] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. MVDream: Multi-view diffusion for 3D generation. In Proc. ICLR, 2024. 2, 3, 7, 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 367, + 295, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 367, + 295, + 444 + ], + "spans": [ + { + "bbox": [ + 56, + 367, + 295, + 444 + ], + "type": "text", + "content": "[54] Yawar Siddiqui, Filippos Kokkinos, Tom Monnier, Mahendra Kariya, Yanir Kleiman, Emilien Garreau, Oran Gafni, Natalia Neverova, Andrea Vedaldi, Roman Shapovalov, and David Novotny. Meta 3D Asset Gen: Text-to-mesh generation with high-quality geometry, texture, and PBR materials. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 2024. 2, 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 295, + 510 + ], + "type": "text", + "content": "[55] Yawar Siddiqui, Tom Monnier, Filippos Kokkinos, Mahendra Kariya, Yanir Kleiman, Emilien Garreau, Oran Gafni, Natalia Neverova, Andrea Vedaldi, Roman Shapovalov, et al. Meta 3d assetgen: Text-to-mesh generation with high-quality geometry, texture, and pbr materials. NeurIPS, 2024. 2, 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 295, + 567 + ], + "type": "text", + "content": "[56] Xingyuan Sun, Jiajun Wu, Xiuming Zhang, Zhoutong Zhang, Chengkai Zhang, Tianfan Xue, Joshua B Tenenbaum, and William T Freeman. Pix3d: Dataset and methods for single-image 3d shape modeling. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 295, + 622 + ], + "type": "text", + "content": "[57] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Viewset diffusion: (0-)image-conditioned 3D generative models from 2D data. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 2, 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 668 + ], + "type": "text", + "content": "[58] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Splatter Image: Ultra-fast single-view 3D reconstruction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "text", + "content": "[59] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. LGM: Large multi-view Gaussian model for high-resolution 3D content creation. arXiv, 2402.05054, 2024. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "text", + "content": "[60] Shitao Tang, Jiacheng Chen, Dilin Wang, Chengzhou Tang, Fuyang Zhang, Yuchen Fan, Vikas Chandra, Yasutaka Furukawa, and Rakesh Ranjan. MVDiffusion++: A dense high-resolution multi-view diffusion model for single or sparse-view 3d object reconstruction. arXiv, 2402.12712, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "text", + "content": "[61] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. NeurIPS, 2017. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "type": "text", + "content": "[62] Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. Vggsfm: Visual geometry grounded deep structure from motion. In CVPR, pages 21686-21697, 2024. 2, 4, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 222, + 553, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 222, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 316, + 222, + 553, + 276 + ], + "type": "text", + "content": "[63] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P. Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas A. Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proc. CVPR, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 279, + 553, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 279, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 316, + 279, + 553, + 312 + ], + "type": "text", + "content": "[64] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. DUSt3R: Geometric 3D vision made easy. In Proc. CVPR, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 314, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 314, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 314, + 553, + 358 + ], + "type": "text", + "content": "[65] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. MeshLRM: large reconstruction model for high-quality mesh. arXiv, 2404.12385, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 360, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 553, + 404 + ], + "type": "text", + "content": "[66] Rundi Wu, Chang Xiao, and Changxi Zheng. Deepcad: A deep generative network for computer-aided design models. in 2021 IEEE. In CVF International Conference on Computer Vision (ICCV), pages 6772-6782, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 406, + 553, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 406, + 553, + 461 + ], + "spans": [ + { + "bbox": [ + 316, + 406, + 553, + 461 + ], + "type": "text", + "content": "[67] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P. Srinivasan, Dor Verbin, Jonathan T. Barron, Ben Poole, and Aleksander Holynski. ReconFusion: 3D Reconstruction with Diffusion Priors. arXiv preprint, 2023. 2, 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 463, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 463, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 463, + 553, + 529 + ], + "type": "text", + "content": "[68] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, et al. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 803-814, 2023. 2, 3, 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 531, + 553, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 531, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 316, + 531, + 553, + 574 + ], + "type": "text", + "content": "[69] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguuang Zhang, Xiaou Tang, and Jianxiong Xiao. 3D ShapeNets: A deep representation for volumetric shapes. In Proc. CVPR, 2015. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 577, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 577, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 577, + 553, + 609 + ], + "type": "text", + "content": "[70] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond PASCAL: A benchmark for 3D object detection in the wild. In Proc. WACV, 2014. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 612, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 612, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 612, + 553, + 656 + ], + "type": "text", + "content": "[71] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. InstantMesh: efficient 3D mesh generation from a single image with sparse-view large reconstruction models. arXiv, 2404.07191, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[72] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation. arXiv preprint arXiv:2403.14621, 2024. 7" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14112" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 474 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "text", + "content": "[73] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. GRM: Large gaussian reconstruction model for efficient 3D reconstruction and generation. arXiv, 2403.14621, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 171 + ], + "type": "text", + "content": "[74] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proc. CVPR, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "text", + "content": "[75] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, and Angjoo Kanazawa. gsplat: An open-source library for Gaussian splatting. arXiv preprint arXiv:2409.06765, 2024. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 283 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 283 + ], + "type": "text", + "content": "[76] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, et al. Mvmgnet: A large-scale dataset of multi-view images. In CVPR, pages 9150-9161, 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 294, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 294, + 328 + ], + "type": "text", + "content": "[77] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. GS-LRM: large reconstruction model for 3D Gaussian splatting. arXiv, 2404.19702, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "type": "text", + "content": "[78] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proc. CVPR, pages 586-595, 2018. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 375, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 375, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 375, + 294, + 418 + ], + "type": "text", + "content": "[79] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snively. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 420, + 294, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 420, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 420, + 294, + 474 + ], + "type": "text", + "content": "[80] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets Gaussian splatting: Fast and generalizable single-view 3D reconstruction with transformers. arXiv.cs, abs/2312.09147, 2023. 3" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14113" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Unbiased Video Scene Graph Generation via Visual and Semantic Dual Debiasing/168f707b-a7fa-4e7c-93c1-aac7eef770eb_content_list.json b/2025/Unbiased Video Scene Graph Generation via Visual and Semantic Dual Debiasing/168f707b-a7fa-4e7c-93c1-aac7eef770eb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2f24c458526948f15aa67fbfb6029047d970a39d --- /dev/null +++ b/2025/Unbiased Video Scene Graph Generation via Visual and Semantic Dual Debiasing/168f707b-a7fa-4e7c-93c1-aac7eef770eb_content_list.json @@ -0,0 +1,1837 @@ +[ + { + "type": "text", + "text": "Unbiased Video Scene Graph Generation via Visual and Semantic Dual Debiasing", + "text_level": 1, + "bbox": [ + 140, + 128, + 856, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yanjun $\\mathrm{Li}^{1\\dagger}$ , Zhaoyang $\\mathrm{Li}^1$ , Honghui Chen $^2$ , Lizhi $\\mathrm{Xu}^1$", + "bbox": [ + 277, + 202, + 712, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Science and Technology of China", + "bbox": [ + 305, + 222, + 691, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ College of Physics and Information Engineering, Fuzhou University", + "bbox": [ + 225, + 239, + 769, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ballinli@mail.ustc.edu.cn, lizhaoyang@mail.ustc.edu.cn, chh5840996@gmail.com", + "bbox": [ + 158, + 260, + 834, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "xulizhi@mail.ustc.edu.cn † Corresponding Author", + "bbox": [ + 305, + 277, + 689, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 327, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video Scene Graph Generation (VidSGG) aims to capture dynamic relationships among entities by sequentially analyzing video frames and integrating visual and semantic information. However, VidSGG is challenged by significant biases that skew predictions. To mitigate these biases, we propose a Visual and Semantic Awareness (VISA) framework for unbiased VidSGG. VISA addresses visual bias through memory-enhanced temporal integration that enhances object representations and concurrently reduces semantic bias by iteratively integrating object features with comprehensive semantic information derived from triplet relationships. This visual-semantics dual debiasing approach results in more unbiased representations of complex scene dynamics. Extensive experiments demonstrate the effectiveness of our method, where VISA outperforms existing unbiased VidSGG approaches by a substantial margin (e.g., +13.1% improvement in mR@20 and mR@50 for the SGCLS task under Semi Constraint).", + "bbox": [ + 86, + 359, + 483, + 632 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 646, + 220, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scene Graph Generation (SGG) [22] constructs structured scene representations by identifying objects and their relationships as triplets (). This structure enables various downstream tasks [7, 10, 14, 27], highlighting SGG as a pivotal area in computer vision. Initially applied to static images (ImgSGG) [36], SGG has been extended to video data (VidSGG) [3] to model dynamic interactions over time. However, VidSGG faces challenges due to biases favoring overrepresented classes, leading to the underrepresentation of rare classes.", + "bbox": [ + 89, + 672, + 482, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent methods have adopted attention-based strategies to address class imbalance and mitigate classification biases [23, 25]. However, these approaches, primarily relying on learnable attention layers, often amplify biases towards", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7a95470c773dfec5b061732ac1e65d7855e271b9276a9c57f5294d0d65768004.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 325, + 905, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3b0f00d18e9a3436f02d76a5e55c3f4c5ba3995031bcd04ea410b8416891640f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 425, + 906, + 481 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/200af16ffb35782c7aea1ac3b49c1b8f67073dc9584fe99db26228d743c6ca52.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 486, + 906, + 563 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/937af58526537528cb6ef03ffefb600550d29734e4175c11ff69e9f661e835d1.jpg", + "image_caption": [ + "Figure 1. Framework Comparison. We identify visual and semantic biases in VidSGG. Figures (a) and (c) show that prior methods rely on weak visual features and limited context, leading to inaccurate attention maps and skewed predicate distributions. In (b) and (d), our memory-enhanced visual debiasing and iterative semantic fusion mitigate both biases, resulting in stronger feature representations and more balanced predicate outcomes." + ], + "image_footnote": [], + "bbox": [ + 517, + 563, + 905, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "high-frequency classes [12]. To overcome this limitation, alternative methods have been proposed [12, 18] aiming to reduce spurious correlations and better account for label correlations. Although these methods alleviate bias to some extent, they often fall short of achieving the desired performance. A key oversight of these approaches is their heavy reliance on debiasing techniques, which neglect the", + "bbox": [ + 511, + 794, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19047", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "intrinsic visual-semantic nature of scene graphs.", + "bbox": [ + 89, + 90, + 410, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Scene graphs inherently integrate both visual and semantic information, allowing biases to be categorized into visual biases and semantic biases. Visual biases arise from low visual quality in videos, such as blurring and occlusion [23], which degrade object feature representations. For instance, as shown in Fig. 1(a), objects may be partially visible or obscured over time, hindering models from capturing robust visual features for classification. Additionally, as illustrated in Fig. 1(c), the VidSGG task faces semantic biases. Previous methods [5, 23] rely solely on visual features to guide the relation generator for predicate prediction. This dependence on single-modal information leads to biased assumptions, similar to humans inferring relationships without sufficient context, thereby favoring more frequent predicate classes [21].", + "bbox": [ + 88, + 107, + 480, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address scene graph biases, we categorize them into visual and semantic types, guiding our proposed framework, VISA (VIual and Semantic Awareness), for unbiased VidSGG, as illustrated in Fig. 1 (b) and (d). VISA mitigates visual bias through a Memory-Enhanced Temporal Integrator that enriches the representations of entities in the current frame with historical representations from previous frames. It employs a relation generator to construct initial scene graphs and a hierarchical semantic extractor to iteratively refine semantic information. Our hierarchical fusion strategy integrates multimodal visual and semantic contexts to address semantic biases. Specifically, the visual debiasing method reduces visual feature variance, while the semantic debiasing method increases the Kullback-Leibler (KL) divergence. This dual-debiasing approach effectively mitigates visual biases caused by poor visual feature quality and semantic biases arising from insufficient contextual information, while also addressing model bias due to data skew. Extensive experiments demonstrate the effectiveness of VISA, achieving superior performance in unbiased VidSGG compared to previous methods. Our main contributions are summarized as follows:", + "bbox": [ + 91, + 333, + 483, + 662 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We identify and explicitly define the visual and semantic biases affecting VidSGG, highlighting their impact on model performance.", + "- We propose VISA, a debiasing framework for VidSGG that simultaneously addresses visual and semantic biases through memory-enhanced temporal integration and hierarchical semantic extraction.", + "- We validate the effectiveness of our approach through extensive experiments and theoretical analyses, demonstrating state-of-the-art performance in unbiased VidSGG." + ], + "bbox": [ + 89, + 666, + 482, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 823, + 232, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Scene Graph Generation", + "text_level": 1, + "bbox": [ + 89, + 848, + 316, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Scene Graph Generation (SGG) approaches primarily focus on static image-based SGG (ImgSGG) and video-based", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SGG (VidSGG). For ImgSGG, sequential encoders like LSTM [39] and attention mechanisms [32] capture global context, with attention methods [6, 16, 30] advancing standalone entity and predicate proposals (e.g., SGTR [16]).", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "VidSGG extends SGG to dynamic contexts, requiring intra-frame relationships [11]. Recent methods improve temporal coherence [20] and efficiency by leveraging adaptive structures [31] and social context modeling [2].", + "bbox": [ + 511, + 152, + 903, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite their success, existing methods ignore unbiasedness. This results in biased SGG with inaccurate and misleading relationships in scenes. Consequently, such biases make models less reliable and less generalizable to real-world applications.", + "bbox": [ + 511, + 214, + 903, + 289 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Unbiased Scene Graph Generation", + "text_level": 1, + "bbox": [ + 511, + 301, + 815, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike conventional SGG, unbiased SGG aims to reduce biases toward frequent objects and relationships. Key approaches include knowledge distillation [15] (e.g., LS-KD, which addresses multi-predicate challenges) and dual-branch architectures [41] (e.g., DHL, which balances attention between head and tail classes).", + "bbox": [ + 511, + 324, + 903, + 414 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For unbiased VidSGG, Transformer-based models [5] and Gaussian Mixture Models [23] have been explored to capture video dynamics better and reduce bias in visual relations. Iterative approaches [28] use conditional variables to improve video relation detection but lack a hierarchical strategy that integrates both visual and semantic representations, as in our method.", + "bbox": [ + 511, + 415, + 903, + 520 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike existing approaches that merely apply traditional debiasing strategies to VidSGG, our framework leverages the inherent visual-semantic nature of scene graphs. This enables a solution that is both experimentally validated and theoretically robust, effectively addressing the fundamental challenges in achieving unbiased VidSGG.", + "bbox": [ + 511, + 522, + 903, + 612 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 627, + 602, + 642 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 651, + 622, + 665 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The VISA framework, depicted in Fig. 2, comprises Memory Guided Sequence Modeling (MGSM), and the Iterative Relation Generator (IRG). It begins with an off-the-shelf object detector to identify objects in video frames, $\\mathbf{b}_i^t$ represents the object anchors of the $i$ -th entity in the $t$ -th frame. In bounding-box approaches (e.g., VidSGG), $\\mathbf{b}_i^t \\in \\mathbb{R}^4$ denotes a box, while in panoptic segmentation, $\\mathbf{b}_i^t \\in \\mathbb{R}^n$ represents the $n$ boundary points. RoI-Align operations [9] are then applied to obtain feature representations, denoted as $\\mathbf{v}_i^t$ . The MGSM refines these features to produce subject and object representations, yielding enhanced features $\\hat{\\mathbf{v}}_i^t$ , while two separate FFNs generate distinct subject and object embeddings. The IRG leverages these visual representations, using the Hierarchical Semantic Extractor (HSE) to fuse visual and semantic information. Through iterative", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19048", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fde13c76350966c85ca971b89014b75f7f449b601256093d62f6106143b8073f.jpg", + "image_caption": [ + "Figure 2. Framework of VISA. Given an input video, VISA extracts object features using an off-the-shelf detector. Subsequently, the Memory Guided Sequence Modeling leverages these features to develop robust object representations. Following this, the Iterative Relation Generator employs the object features to predict relationships between subject-object pairs. It is important to note that the Hierarchical Semantic Extractor is not engaged during the initial phase of scene graph generation." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 911, + 344 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "processing, it deduces unbiased semantic relations and generates relation predictions.", + "bbox": [ + 89, + 424, + 482, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Memory Guided Sequence Modeling", + "text_level": 1, + "bbox": [ + 89, + 462, + 408, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We observe that as video duration increases, the Transformer tends to generate insufficient visual representations, particularly for small objects prone to blurring and occlusion (see Fig. 1 (a)). This insufficiency leads to high variance in feature representations due to the instability of visual information, causing the model to favor frequent entities and overlook less common ones.", + "bbox": [ + 89, + 484, + 483, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our theoretical analysis reveals that this instability introduces noise into the object representations, modeled as:", + "bbox": [ + 89, + 590, + 483, + 621 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {i} ^ {t} = \\mathbf {v} _ {i} + \\boldsymbol {\\epsilon} _ {i} ^ {t}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 628, + 480, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon_i^t$ denotes zero-mean Gaussian noise with covariance $\\pmb{\\Sigma}$ , i.e., $\\mathbb{E}[\\epsilon_i^t] = 0$ and $\\mathrm{Cov}[\\epsilon_i^t] = \\pmb{\\Sigma}$ . We assume these noise terms are independent across different time steps and objects. The observed features from the $t$ -th frame are used directly as the object's representation:", + "bbox": [ + 89, + 655, + 483, + 731 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {v}} _ {i} ^ {t} = \\mathbf {v} _ {i} ^ {t}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 254, + 739, + 480, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consequently, the expectation and variance of the visual representation are computed as follows:", + "bbox": [ + 89, + 766, + 482, + 796 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} \\left[ \\hat {\\mathbf {v}} _ {i} ^ {t} \\right] = \\mathbf {v} _ {i}, \\quad \\operatorname {V a r} \\left[ \\hat {\\mathbf {v}} _ {i} ^ {t} \\right] = \\boldsymbol {\\Sigma}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 805, + 480, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While the single-frame estimation is unbiased, the noise introduced due to visual instability leads to high variance in the feature representations. This high variance destabilizes the feature representations and impairs subsequent relational predictions.", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/28d8dd22ed2cbb3e84b95763a3621ddf683af7821cde69e4cd992ccf8c461a6c.jpg", + "image_caption": [ + "Figure 3. Structure of Memory Guided Sequence Modeling." + ], + "image_footnote": [], + "bbox": [ + 526, + 421, + 893, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Understanding that the high variance stems from the instability of visual information motivated us to design the Memory Guided Sequence Modeling (MGSM) module (illustrated in Fig. 3). The MGSM leverages temporal visual information from preceding frames to enhance individual feature representations of frames, effectively reducing the variance caused by insufficient visual representations.", + "bbox": [ + 511, + 592, + 903, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, we compute adaptive weights $\\mathbf{W}_i^t$ based on the current frame's features:", + "bbox": [ + 511, + 696, + 903, + 727 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {W} _ {i} ^ {t} = \\sigma (\\mathrm {M L P} (\\mathbf {v} _ {i} ^ {t})), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 739, + 903, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\sigma$ is the sigmoid function, and MLP denotes a multi-layer perceptron. Using these weights, we obtain the weighted feature representation:", + "bbox": [ + 511, + 768, + 906, + 814 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {i} ^ {t *} = \\mathbf {W} _ {i} ^ {t} \\odot \\mathbf {v} _ {i} ^ {t} \\| \\mathbf {v} _ {i} ^ {t - 1}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 825, + 903, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\odot$ denotes element-wise multiplication, and $\\parallel$ indicates feature concatenation.", + "bbox": [ + 511, + 854, + 903, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To accumulate visual information and refine the current", + "bbox": [ + 532, + 885, + 906, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19049", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "representations, we introduce an update mechanism:", + "bbox": [ + 89, + 90, + 437, + 106 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {i} ^ {t + 1} = (1 - \\lambda) \\mathbf {M} _ {i} ^ {t} + \\lambda \\mathbf {v} _ {i} ^ {t}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 113, + 482, + 132 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As a result, the expectation and variance of the memory representation $\\mathbf{M}_i^t$ stabilize over time:", + "bbox": [ + 89, + 141, + 482, + 170 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} \\left[ \\mathbf {M} _ {i} ^ {t} \\right] = \\mathbf {v} _ {i}, \\operatorname {V a r} \\left[ \\mathbf {M} _ {i} ^ {t} \\right] = \\frac {\\lambda \\boldsymbol {\\Sigma}}{2 - \\lambda} \\approx \\frac {\\lambda \\boldsymbol {\\Sigma}}{2}, \\text {f o r s m a l l} \\lambda . \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 178, + 482, + 209 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This analysis shows that a smaller $\\lambda$ reduces variance, enhancing the robustness of feature estimation. However, $\\lambda$ cannot be too small, as excessively small values slow the adaptation to new information and may introduce bias. We provide detailed proofs of the variance calculation and the associated trade-off in the supplementary materials.", + "bbox": [ + 89, + 215, + 483, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The final enhanced feature representation is obtained through a dual attention layer:", + "bbox": [ + 89, + 306, + 483, + 338 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {v}} _ {i} ^ {t} = \\operatorname {A t t e n t i o n} \\left(\\mathbf {Q} = \\mathbf {v} _ {i} ^ {t *}, \\mathbf {K} = \\mathbf {M} _ {i} ^ {t}, \\mathbf {V} = \\mathbf {v} _ {i} ^ {t}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 345, + 482, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, the final enhanced feature representation is processed by the Transformer Encoder, as in [23], to generate the subject and object representations. By adaptively enhancing features and leveraging the reduced variance from the memory representation, the MGSM addresses the high variance caused by insufficient visual representations, thereby mitigating visual bias.", + "bbox": [ + 89, + 372, + 483, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our experimental results (Fig. 6) confirm our theoretical insights by demonstrating that MGSM enhances visual representations. This enhancement reduces the high variance caused by insufficient visual information, which is crucial for addressing visual bias in VidSGG.", + "bbox": [ + 89, + 478, + 483, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Iterative Relation Generator", + "text_level": 1, + "bbox": [ + 89, + 561, + 346, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing methods for unbiased SGG predominantly predict predicates directly from visual representations [5, 23]. However, this approach is akin to humans attempting to deduce relationships solely based on visual cues without sufficient contextual information, which easily introduces semantic bias [21]. Specifically, when lacking context, models tend to over-rely on the imbalanced and biased prior distributions present in the training data, leading to biased relationship predictions.", + "bbox": [ + 89, + 583, + 482, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To alleviate semantic bias, we start by analyzing the problem from an information-theoretic perspective. Semantic bias occurs when a model lacks sufficient contextual information, leading to high uncertainty in predicting relationships. This uncertainty causes the model to over-rely on biased prior distributions from the training data. We use entropy $H$ to quantify this uncertainty [29]. We consider the conditional entropy $H(r_{ij} \\mid v_i, v_j, S)$ , where $v_i$ and $v_j$ are the feature representations of the subject and object, respectively. Based on the properties of conditional entropy [29], the additional context provided by $S$ decreases the uncertainty in predicting relationships:", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nH \\left(r _ {i j} \\mid v _ {i}, v _ {j}, S\\right) \\leq H \\left(r _ {i j} \\mid v _ {i}, v _ {j}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 90, + 906, + 108 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This reduction in conditional entropy means the model is less uncertain and relies less on biased priors. By incorporating context $S$ , the likelihood term $P(v_{i}, v_{j}, S \\mid r_{ij})$ in Bayes' theorem becomes more informative, effectively reducing the relative influence of the prior $P(r_{ij})$ . According to Bayes' theorem:", + "bbox": [ + 511, + 116, + 906, + 207 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP \\left(r _ {i j} \\mid v _ {i}, v _ {j}, S\\right) = \\frac {P \\left(v _ {i} , v _ {j} , S \\mid r _ {i j}\\right) \\cdot P \\left(r _ {i j}\\right)}{P \\left(v _ {i} , v _ {j} , S\\right)}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 214, + 906, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To quantify the change in reliance on the prior, we consider the Kullback-Leibler (KL) divergence between the posterior $P(r_{ij} \\mid v_i, v_j, S)$ and the prior $P(r_{ij})$ :", + "bbox": [ + 511, + 256, + 905, + 303 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} D _ {\\mathrm {K L}} \\left(P \\left(r _ {i j} \\mid v _ {i}, v _ {j}, S\\right) \\| P \\left(r _ {i j}\\right)\\right) \\\\ = \\sum_ {r _ {i j}} P (r _ {i j} \\mid v _ {i}, v _ {j}, S) \\log \\frac {P (r _ {i j} \\mid v _ {i} , v _ {j} , S)}{P (r _ {i j})} \\\\ = H \\left(r _ {i j} \\mid v _ {i}, v _ {j}, P \\left(r _ {i j}\\right)\\right) - H \\left(r _ {i j} \\mid v _ {i}, v _ {j}, S\\right). \\tag {11} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 306, + 903, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As the conditional entropy $H(r_{ij} \\mid v_i, v_j, S)$ decreases due to the additional context, the KL divergence increases. This indicates that the posterior distribution diverges from the biased prior, reflecting a reduced reliance on the prior and effectively mitigating semantic bias.", + "bbox": [ + 511, + 402, + 905, + 478 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This theoretical insight motivates our approach to incorporate contextual information into the model. Based on this, we propose the Iterative Relation Generator (IRG), an iterative semantic debiasing component within our framework. The IRG iteratively introduces contextual information to refine relationship predictions, thereby reducing uncertainty and dependence on biased priors.", + "bbox": [ + 511, + 479, + 905, + 584 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the first iteration, the IRG infers semantic information from each subject-object pair to generate semantic embeddings. For each subject-object pair $(j,i)$ at frame $t$ , we construct the composite objects features $\\mathbf{p}_{j,i}^{t}$ :", + "bbox": [ + 511, + 584, + 905, + 646 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {j, i} ^ {t} = f _ {v} \\left(\\mathbf {v} _ {j} ^ {t}\\right) \\| f _ {v} \\left(\\mathbf {v} _ {i} ^ {t}\\right) \\| \\left(f _ {u} \\left(\\mathbf {u} _ {j i} ^ {t}\\right) + f _ {\\text {b o x}} \\left(b _ {j} ^ {t}, b _ {i} ^ {t}\\right)\\right) \\| \\mathbf {s} _ {j} ^ {t} \\| \\mathbf {s} _ {i} ^ {t}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 648, + 903, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{v}_j^t$ and $\\mathbf{v}_i^t$ are the visual feature representations of the subject and object. The tensor $\\mathbf{u}_{ji}^{t}$ is the union box feature map extracted via RoIAlign [9], and $\\mathbf{s}_j^t$ and $\\mathbf{s}_i^t$ are the semantic embeddings obtained from GloVe [24]. The functions $f_{v}, f_{u}$ , and $f_{\\mathrm{box}}$ are feature transformation functions.", + "bbox": [ + 511, + 686, + 905, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then perform initial relationship prediction, obtaining predicate categories such as attention, spatial, and contacting. We select the semantic embeddings with the highest probabilities from these distributions to form integrated triplet embeddings $\\mathbf{C}_{pre,(j,i)}^t$ :", + "bbox": [ + 511, + 763, + 905, + 843 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} _ {p r e, (j, i)} ^ {t} = \\left[ \\begin{array}{l} f _ {v} \\left(\\mathbf {v} _ {j} ^ {t}\\right) \\| f _ {a, (j, i)} ^ {t} \\| f _ {v} \\left(\\mathbf {v} _ {i} ^ {t}\\right) \\\\ f _ {v} \\left(\\mathbf {v} _ {j} ^ {t}\\right) \\| f _ {s, (j, i)} ^ {t} \\| f _ {v} \\left(\\mathbf {v} _ {i} ^ {t}\\right) \\\\ f _ {v} \\left(\\mathbf {v} _ {j} ^ {t}\\right) \\| f _ {c, (j, i)} ^ {t} \\| f _ {v} \\left(\\mathbf {v} _ {i} ^ {t}\\right) \\end{array} \\right], \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 848, + 903, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19050", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/66fe30543f6d0373c4cce5d852ce2fc5fbcb7cd07591580a913fa30b880826d3.jpg", + "image_caption": [ + "Figure 4. Structure of Hierarchical Semantics Extractor." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 486, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f_{a,(j,i)}^t, f_{s,(j,i)}^t$ , and $f_{c,(j,i)}^t$ are the semantic embeddings corresponding to the respective predicates.", + "bbox": [ + 89, + 260, + 482, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further exploit hierarchical contextual information, we design the Hierarchical Semantics Extractor (HSE) (as shown in Fig. 4). Inspired by the multi-hierarchical feature fusion method [35], the HSE combines composite object features with integrated triplets to obtain fine-grained visual and semantic representations. We decompose the composite objects features $\\mathbf{p}_{j,i}^{t}$ into fine-grained subject and object representations as follows:", + "bbox": [ + 89, + 290, + 483, + 411 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {S} _ {\\mathbf {p} _ {j, i} ^ {t}} ^ {t} = f _ {v} (\\mathbf {v} _ {j} ^ {t}) \\| \\left(f _ {u} (\\mathbf {u} _ {j} ^ {t}) + f _ {\\text {b o x}} \\left(b _ {j} ^ {t}, b _ {i} ^ {t}\\right)\\right) \\| \\mathbf {s} _ {j} ^ {t}, \\\\ \\mathbf {G} _ {i} ^ {t} = f _ {v} (t ^ {t}) \\| \\left(f _ {v} (t ^ {t}) + f _ {\\text {b o x}} (t ^ {t}, t ^ {t})\\right) \\| \\end{array} \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 419, + 483, + 449 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {O} _ {\\mathbf {p} _ {j, i} ^ {t}} ^ {t} = f _ {v} \\left(\\mathbf {v} _ {i} ^ {t}\\right) \\| \\left(f _ {u} \\left(\\mathbf {u} _ {j i} ^ {t}\\right) + f _ {\\text {b o x}} \\left(b _ {j} ^ {t}, b _ {i} ^ {t}\\right)\\right) \\| \\mathbf {s} _ {i} ^ {t}.\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 441, + 424, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly, we decompose the integrated triplet embeddings $\\mathbf{C}_{pre,(j,i)}^t$ :", + "bbox": [ + 89, + 474, + 482, + 508 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {S} _ {\\mathbf {C} _ {(j, i)} ^ {t}} ^ {t} = \\left\\{f _ {v} \\left(\\mathbf {v} _ {j} ^ {t}\\right) \\| f _ {k} ^ {t} \\mid k \\in \\{a, s, c \\} \\right\\}, \\\\ \\text {C} _ {i} ^ {t} = \\left\\{f _ {v} ^ {t + 1} \\mid f _ {k} ^ {t + 1} \\mid k \\in \\{a, c \\} \\right\\}. \\end{array} \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 515, + 483, + 545 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {O} _ {\\mathbf {C} _ {(j, i)} ^ {t}} ^ {t} = \\left\\{f _ {k} ^ {t} \\| f _ {v} \\left(\\mathbf {v} _ {i} ^ {t}\\right) \\mid k \\in \\{a, s, c \\} \\right\\}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 532, + 480, + 563 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "These fine-grained representations are concatenated with downsampled embeddings from the previous cross-attention layer via a stride-two convolutional layer, serving as inputs to the next cross-attention layer. This hierarchical structure enables the model to capture multi-level contextual information, effectively reducing uncertainty and reliance on biased priors. The HSE is active for iteration counts $N \\geq 1$ .", + "bbox": [ + 89, + 566, + 483, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After $N$ iterations of the IRG, the final inter-object predicates are generated by the Spatial Encoder and Temporal Decoder modules adopted from [5]. By iteratively leveraging contextual information, our method effectively mitigates semantic bias. Experimental results (as shown in Fig. 7) validate that incorporating contextual information reduces the model's reliance on biased priors and improves relationship predictions.", + "bbox": [ + 89, + 688, + 483, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training and Testing", + "text_level": 1, + "bbox": [ + 89, + 818, + 287, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. In the training phase, VISA synthesizes object features from MGSM for preliminary scene graph construction. Upon reaching iteration $N \\geq 1$ , it activates the HSE to generate a hierarchical visual-semantics fusion embedding.", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Section 3.2, the memory representation $\\mathbf{M}_i^t$ is initialized as a zero tensor, while the embeddings for the three relationships are randomly initialized as described in Section 3.3. We optimize the comprehensive loss function:", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {t o t a l}} = L _ {p} + L _ {e} + L _ {\\text {c o n t r a}}. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 162, + 906, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $L_{p}$ and $L_{e}$ denote the predicate and entity classification losses, both computed via cross-entropy, and $L_{\\mathrm{contra}}$ is the contrastive loss, all defined as in [23].", + "bbox": [ + 511, + 191, + 905, + 236 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Testing. After training, the proposed MGSM produces object predictions through the visual debiasing component, while the IRG performs semantic debiasing.", + "bbox": [ + 511, + 237, + 905, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 297, + 645, + 314 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset and Metrics", + "text_level": 1, + "bbox": [ + 511, + 323, + 705, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We conducted experiments on the Action Genome (AG) dataset [11], the largest benchmark for VidSGG. AG consists of 234, 253 annotated frames containing 476, 229 bounding boxes across 35 object categories. It provides 1, 715, 568 predicate instances spanning 26 distinct relationship classes.", + "bbox": [ + 511, + 345, + 905, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics and Evaluation Setup. We employed the standard metrics for unbiased VidSGG evaluation, mean Recall@K (mR@K), considering $K \\in \\{10, 20, 50\\}$ . Following established protocols [5, 11, 23], we engaged in three SGG tasks, PREDCLS, SGCLS and SGDET. Evaluation was conducted across three settings, With Constraint, Semi Constraint, and No Constraint. Due to page constraints, further details are provided in the supplementary materials.", + "bbox": [ + 511, + 439, + 906, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation details", + "text_level": 1, + "bbox": [ + 511, + 571, + 728, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following prior work [5, 19, 23], we adopted Faster R-CNN [26] with ResNet-101 [8] as the object detector, initially trained on the AG dataset. To ensure a fair comparison, we utilized the official implementations of these methods. For our MGSM module, we set the $\\lambda$ parameter to 0.04 for the SGCLS task and 0.06 for the SGDET task. In the IRG module, we implemented a dual-procedure setup, enabling iterative relational inference with the number of iterations $N$ set to 1. Due to page constraints, further details are provided in the supplementary materials.", + "bbox": [ + 511, + 594, + 905, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Comparison to existing methods", + "text_level": 1, + "bbox": [ + 511, + 757, + 797, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We began with a qualitative analysis, shown in Fig. 5, illustrating results under the Semi Constraint condition, which closely resembles real-world scenarios [5] where multiple predicates typically describe scenes. False positive relationships and object classifications were highlighted in purple blocks. The attention intensity demonstrated that our framework effectively enhanced object representations (e.g., pilow) through visual debiasing across frames from $t$ to $t + 1$ ,", + "bbox": [ + 511, + 779, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19051", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/842803436adecc37ed630b478be44fea1d1fd42eb435440b42fca2436dce67d5.jpg", + "image_caption": [ + "Figure 5. Qualitative results of our proposed method. Comparison with the open-sourced state-of-the-art model TEMPURA [23] under Semi Constraint, closely matching ground truth [5]. Visualizations show attention intensity and video scene graphs from Frame $t$ to $t + 1$ . False positive relationships and object classifications are highlighted in purple blocks." + ], + "image_footnote": [], + "bbox": [ + 130, + 73, + 867, + 354 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/15881c717a6f834ae30982eed0951ca86ade7efa683fa2d0fe687f0e1b9924c8.jpg", + "table_caption": [ + "Table 1. Quantitative results comparisons with unbiased SGG methods under With, Semi, and No Constraints. The experiments are conducted on the dataset AG [11]. The best results are highlighted." + ], + "table_footnote": [], + "table_body": "
ConstraintMethodPREDCLSSGCLSSGDET
mR@10mR@20mR@50mR@10mR@20mR@50mR@10mR@20mR@50
WithRelDN [40]6.26.26.23.43.43.43.33.33.3
HCRD supervised [1]-------8.39.1
TRACE [31]15.215.215.28.98.98.98.28.28.2
ISGG [13]------19.722.9-
STTran [5]37.840.140.227.228.028.020.820.822.2
STTran-TPI [33]37.340.640.628.329.329.320.220.221.8
TEMPURA [23]42.946.346.334.035.235.222.622.623.7
FloCoDe [12]44.849.249.337.439.239.424.224.227.9
VISA (Ours)46.952.052.040.842.542.627.327.330.7
SemiTEMPURA [23]40.744.544.636.939.539.521.821.822.5
VISA (Ours)51.356.356.447.852.652.631.731.733.2
NoRelDN [40]31.263.175.518.636.942.67.518.833.7
TRACE [31]50.973.682.731.942.746.322.831.341.8
STTran [5]51.467.782.740.750.158.820.929.739.2
TEMPURA [23]61.585.198.048.361.166.424.733.943.7
FloCoDe [12]63.286.998.649.763.869.228.635.447.2
VISA (Ours)65.889.199.852.066.371.430.736.750.4
", + "bbox": [ + 194, + 453, + 802, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "compared with the TEMPURA [23]. Additionally, our approach inferred more accurate semantic predicates (e.g., sit on) within a single frame via contextual information. These visualizations confirm the correctness of partitioning scene graph biases into visual and semantic biases and demonstrate that our framework, VISA, significantly mitigates both in VidSGG.", + "bbox": [ + 88, + 705, + 480, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 6 and Fig. 7 further validated our theoretical analysis. Fig. 6 illustrated the variance of feature representations across training epochs, with shaded regions indicating the standard deviation. TEMPURA exhibited higher variance, suggesting less stable feature representations,", + "bbox": [ + 88, + 824, + 485, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "whereas VISA demonstrated significantly reduced variance and tighter standard deviation regions. This reduction confirmed the effectiveness of the MGSM module in stabilizing feature representations. Additionally, Fig. 7 presented the confusion matrix comparing relationship predictions between TEMPURA and VISA. The reduced influence from prior dataset biases in VISA validated our approach to mitigate the semantics bias in relationship predictions using KL divergence principles.", + "bbox": [ + 511, + 705, + 906, + 843 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our approach achieved state-of-the-art performance across all mR@K metrics, as shown in Table 1. VISA demonstrated improvements ranging from a minimum of", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19052", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/17ce4d345dd13c711a4a394d8f3e7c823aece04b3a829a6448c5b9182e238132.jpg", + "image_caption": [ + "Figure 6. Comparison of TEMPURA and VISA visual feature representation variance. Visualization of visual feature representation variance across training epochs. The shaded regions represent the standard deviation." + ], + "image_footnote": [], + "bbox": [ + 140, + 87, + 433, + 309 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4d0dcbc7f963dffd0ec61a351219a352fe413a4adfc858209a5c6be36fe451d9.jpg", + "image_caption": [ + "Figure 8. Quantitative results of our proposed method. Compare our proposed method with open-sourced methods, TEMPURA [23] and STTran [5] on the mR@10 metric under With Constraint and No Constraint." + ], + "image_footnote": [], + "bbox": [ + 91, + 377, + 488, + 518 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/89946733d5ef7ce47a859b7d5bff943903792db96ea2913dc34efb6222fc1384.jpg", + "image_caption": [ + "Figure 7. Qualitative results for relationship confusion matrix. Results for the subject-object pairs, person-sofa, and person-floor. $+1.2\\%$ in mR@50 for the PREDCLS task under the No Constraint scenario to a maximum of $+13.1\\%$ in both mR@20 and mR@50 for the SGCLS task under the Semi Constraint scenario. These results validate the correctness of partitioning scene graph biases into visual and semantic biases and demonstrate the effectiveness of VISA in mitigating them, thereby advancing unbiased VidSGG." + ], + "image_footnote": [], + "bbox": [ + 96, + 518, + 480, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, as shown in Fig. 8, our VISA model surpassed open-source unbiased VidSGG methods in mR@10 across the three predicate class distributions (HEAD, BODY,TAIL) on AG under both With Constraint and No Constraint settings. Here, HEAD represented classes", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ca2a2efcf5ab3e2815e7f5f5ee63fe739b1cc5172774ebb0b12232956884ac5.jpg", + "image_caption": [ + "(a) PredCLS" + ], + "image_footnote": [], + "bbox": [ + 514, + 87, + 658, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5e95f0a704aa7b90ea1e7d8219afd17bf8b515ddcae46e1766c55dd35f44e119.jpg", + "image_caption": [ + "(b) SGCLS" + ], + "image_footnote": [], + "bbox": [ + 668, + 88, + 785, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d41735c28576e1cd660f80b9badcba4ee86e8673b4416f03a5f17aae3bacfd92.jpg", + "image_caption": [ + "(c) SGDET" + ], + "image_footnote": [], + "bbox": [ + 792, + 88, + 906, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dd42a5d5698a3fd623c939ce78765f3ea37db7f6b92d9134a7f08fd4bad917f7.jpg", + "image_caption": [ + "(a) PredCLS" + ], + "image_footnote": [], + "bbox": [ + 513, + 199, + 658, + 294 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cf95ea171e1f498e27044bdf69fa41a7a0713fcbe8b8f3b1c3ba902a24c2f5d6.jpg", + "image_caption": [ + "(b) SGCLS" + ], + "image_footnote": [], + "bbox": [ + 668, + 199, + 784, + 294 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2b662f01b3733e90d24769529b3161ab65569a8249c2478f10fe75cf1cda3136.jpg", + "image_caption": [ + "(c) SGDET" + ], + "image_footnote": [], + "bbox": [ + 792, + 199, + 906, + 294 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with $\\geq 100,000$ training samples, $BODY$ with fewer than 100,000, and $TAIL$ with fewer than 8,000. VISA consistently outperformed all competitors across mR@10 metrics, achieving a notable $+11.0\\%$ improvement for the tail class in the SGDET task under No Constraint. This strong performance on tail classes highlighted the effectiveness of VISA's dual debiasing strategy.", + "bbox": [ + 511, + 383, + 906, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 500, + 679, + 515 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation study on VISA modules. Table 2 evaluates the effectiveness of VISA modules. Removing MGSM resulted in at least a $-2.2\\%$ drop in mR@10 (SGCLS, Semi Constraint), confirming its capability to alleviate visual bias. Similarly, removing IRG caused at least a $-0.7\\%$ decline in mR@10 (SGCLS, No Constraint), validating its role in reducing semantic bias. These results justify the division of biases into visual and semantic components and demonstrate VISA's effectiveness in mitigating both. Note that MGSM was not applied to the PREDCLS task, as it utilizes ground-truth visual inputs, rendering MGSM unnecessary.", + "bbox": [ + 511, + 522, + 906, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Influence of increasing iteration count $N$ . We analyzed the effect of iteration number $N$ on VISA's performance. Due to computational constraints, we conducted experiments with $N = 2$ and $N = 3$ , denoted as $\\mathrm{VISA}_{N=2}$ and $\\mathrm{VISA}_{N=3}$ , respectively. Results in Table 3 show that increasing $N$ consistently enhances the mR@K performance for unbiased VidSGG, though the gain from $N = 2$ to $N = 3$ is smaller compared to $N = 1$ to $N = 2$ . Additionally, training time approximately doubled when increasing from $N = 1$ to $N = 2$ . Further details are provided in the supplementary materials.", + "bbox": [ + 511, + 688, + 908, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The value of the $\\lambda$ . Table 4 showed results for different values of $\\lambda$ . VISA achieved optimal performance for SGCLS at $\\lambda = 0.04$ , while the best results for SGDET occur", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19053", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6ccfe805fa8f2658a024a9ca87aeffe2fbaac7097d6771cda944af795555f5aa.jpg", + "table_caption": [ + "Table 2. Ablation study of MGSM and IRG under With Constraint, Semi Constraint, No Constraint." + ], + "table_footnote": [], + "table_body": "
With ConstraintSemi ConstraintNo Constraint
PREDCLSSGCLSSGDETPREDCLSSGCLSSGDETPREDCLSSGCLSSGDETPREDCLSSGCLSSGDETPREDCLSSGCLSSGDET
mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20
VISA46.952.040.842.527.327.351.356.347.852.631.731.765.889.152.066.330.736.7
w/o MGSM46.952.035.738.218.824.651.356.345.649.324.127.865.889.149.062.027.934.7
w/o MGSM & IRG42.946.334.035.218.522.640.744.536.939.518.521.861.585.148.361.124.733.9
", + "bbox": [ + 94, + 108, + 911, + 196 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fdda9d2d94d159ee2f138eaedc57f0d0e0fb407cb99dbf07888fca7c4470fb2b.jpg", + "table_caption": [ + "Table 3. Influence of increasing iteration count $N$" + ], + "table_footnote": [], + "table_body": "
With ConstraintSemi ConstraintNo Constraint
PREDCLSSGCLSSGDETPREDCLSSGCLSSGDETPREDCLSSGCLSSGDET
mR@10mR@20mR@10mR@20mR@10mR@20mR@10mR@20mR@10mR@20mR@10mR@20mR@10mR@20mR@10mR@20mR@10mR@20
VISA46.952.040.842.527.327.351.356.347.852.631.731.765.889.152.066.330.736.7
VISA_N=247.953.041.743.425.528.752.556.948.253.627.531.966.090.853.167.031.837.2
VISA_N=348.153.342.244.025.828.952.857.148.553.927.832.166.291.253.567.232.037.5
", + "bbox": [ + 94, + 223, + 903, + 313 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a641d9423aef549e5095ef3f2855227dc0afd207a808719ed0d162449daf1170.jpg", + "table_caption": [ + "Table 4. Different values of $\\lambda$ on SGCLS and SGDET,we summarize the results in mean-Recall@K,with best in bold." + ], + "table_footnote": [], + "table_body": "
λWith ConstraintSemi ConstraintNo Constraint
SGCLSSGDETSGCLSSGDETSGCLSSGDET
mR@10mR@20mR@50mR@10mR@20mR@50mR@10mR@20mR@50mR@10mR@20mR@50mR@10mR@20mR@50mR@10mR@20mR@50
0.00534.236.236.217.921.522.540.142.642.721.526.628.146.457.957.924.832.542.5
0.0136.037.437.418.622.223.240.744.644.723.728.630.047.259.865.225.434.043.0
0.0239.241.241.818.422.123.043.147.047.122.527.128.148.362.467.524.532.841.7
0.0440.842.542.619.823.324.347.852.652.624.529.531.152.066.371.426.733.641.7
0.0635.136.436.427.327.330.741.845.845.931.731.733.246.959.364.730.736.750.4
0.0834.335.635.618.321.622.540.243.743.723.327.828.947.159.665.724.732.740.6
0.133.234.434.417.520.322.139.142.542.522.425.427.246.259.664.322.131.439.2
", + "bbox": [ + 94, + 340, + 903, + 463 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/18ffb617f798f52f61eb2a32229e1967d041913a1aaa05611142b7eb2a81c1e8.jpg", + "table_caption": [ + "Table 5. Performance evaluation on the PVSG [38] Dataset" + ], + "table_footnote": [], + "table_body": "
MethodPVSG Metrics
Stage-1Stage-2mR@20mR@50mR@100
VPS [4, 17]Transformer0.360.390.40
IPS+T [4, 34]Transformer1.751.862.03
MGSMIRG8.828.9410.51
", + "bbox": [ + 94, + 493, + 480, + 560 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8b33206568b2804b6ba9cf38fc0412af82815f1c8f617fc912fbea8f7854057c.jpg", + "table_caption": [ + "Table 6. Performance evaluation on the PSG4D-GTA and PSG4D-HOI [37] 4DPVSG Datasets." + ], + "table_footnote": [], + "table_body": "
MethodPSG4D-GTAPSG4D-HOI
mR@20mR@50mR@100mR@20mR@50mR@100
PSG4DFounder (P) [37]2.102.933.133.103.954.17
PSG4DFounder (R) [37]3.313.854.023.654.164.97
MGSM + IRG10.8311.5111.9810.9512.3212.87
", + "bbox": [ + 94, + 606, + 483, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "at $\\lambda = 0.06$ . The differences in optimal $\\lambda$ values reflect the distinct characteristics of these tasks. Specifically, SGCLS benefits from ground-truth data, enabling effective updates to the memory representations with a smaller $\\lambda$ . In contrast, SGDET, which builds scene graphs from scratch, performs better with a slightly higher $\\lambda$ . Notably, since the MSGM module was not applied to the PREDCLS task, the parameter $\\lambda$ is also not involved in PREDCLS.", + "bbox": [ + 88, + 672, + 482, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Exploration on PVSG and 4DPVSG Datasets. We extended our framework to Panoptic Video Scene Graph Generation (PVSG) and 4D Panoptic Scene Graph Generation (4DPVSG) tasks [37, 38]. Unlike VidSGG, PVSG utilizes pixel-level masks for grounding nodes. Using the IPS+T+Transformer codebase, we integrated MGSM with a Memory-Enhanced Temporal Integrator and employed", + "bbox": [ + 88, + 794, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "UniTrack [34] for temporal linking. For relation classification, IRG was adapted for iterative semantic debiasing. The 4DPVSG task further generalized PVSG to dynamic 4D contexts, employing PSG4DFormer variants with point cloud (P) and RGB-D (R) inputs. As shown in Table 5, our method significantly outperformed baselines, achieving minimum improvements of $+7.52\\%$ mR@20 (PVSG) and $+7.30\\%$ mR@20 (4DPVSG, PSG4D-HOI). These results confirm the broad applicability of our dual debiasing approach across VidSGG and indicate potential benefits for other generative and embodied intelligence applications.", + "bbox": [ + 511, + 477, + 906, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 654, + 633, + 669 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we present VISA, an approach for unbiased video scene graph generation that addresses both visual and semantic biases inherent in scene graphs. VISA reduces visual bias by enriching current frame representations with historical context to decrease visual feature variance, and mitigates semantic bias by increasing the KL divergence between model predictions and biased priors. Our framework integrates hierarchical semantic contexts with visual features, enabling more unbiased scene graph generation. Extensive experimental and theoretical analyses demonstrate the effectiveness of VISA. Future work will explore extending our visual-semantic debiasing approach to other multimodal domains such as visual question answering and video captioning, as well as applications in embodied intelligence, including robotics perception and interactive agents.", + "bbox": [ + 511, + 674, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19054", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We gratefully acknowledge the three anonymous reviewers for their constructive feedback on our CVPR 2025 submission. This work was supported by personal funding from the corresponding author.", + "bbox": [ + 89, + 113, + 485, + 172 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 196, + 187, + 213 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] João Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 6", + "[2] Shizhe Chen, Zhiyuan Shi, Pascal Mettes, and Cees G. M. Snoek. Social fabric: Tubelet compositions for video relation detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR), 2021. 2", + "[3] Shizhe Chen, Jing Xiao, and Lin Chen. Video scene graph generation from single-frame weak supervision. In Proceedings of the Eleventh International Conference on Learning Representations (ICLR), 2022. 1", + "[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 8", + "[5] Yunhan Cong, Wentao Liao, Heiko Ackermann, Bodo Rosenhahn, and Michael Ying Yang. Spatial-temporal transformer for dynamic scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 4, 5, 6, 7", + "[6] Yunhan Cong, Michael Ying Yang, and Bodo Rosenhahn. Reltr: Relation transformer for scene graph generation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023. 2", + "[7] Linfeng Gao, Bin Wang, and Wenmin Wang. Image captioning with scene-graph based semantic concepts. In Proceedings of the International Conference on Machine Learning and Computing, 2018. 1", + "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5", + "[9] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 2, 4", + "[10] Marcel Hildebrandt, Haoyang Li, Raphael Koner, Volker Tresp, and Stephan Gunnemann. Scene graph reasoning for visual question answering. arXiv preprint arXiv:2007.01072, 2020. 1", + "[11] Jiading Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as compositions of spatiotemporal scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 5, 6", + "[12] Anant Khandelwal. Flocode: Unbiased dynamic scene graph generation with temporal consistency and correlation debias" + ], + "bbox": [ + 93, + 222, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2024. 1, 6", + "[13] Shreyank Khandelwal and Leonid Sigal. Iterative scene graph generation. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 6", + "[14] Hengshuang Li, Xin Li, Behnam Karimi, Jie Chen, and Ming Sun. Joint learning of object graph and relation graph for visual question answering. In Proceedings of the IEEE International Conference on Multimedia and Expo (ICME), 2022. 1", + "[15] Lifeng Li, Jing Xiao, Huan Shi, Wei Wang, Jing Shao, Alan Liu, Yueting Yang, and Lin Chen. Label semantic knowledge distillation for unbiased scene graph generation. IEEE Transactions on Circuits and Systems for Video Technology (TCSVT), 2023. 2", + "[16] Ruofei Li, Shuhui Zhang, and Xiaochun He. Sgtr: End-to-end scene graph generation with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[17] Xiangtai Li, Wenwei Zhang, Jiangmiao Pang, Kai Chen, Guangliang Cheng, Yunhai Tong, and Chen Change Loy. Video k-net: A simple, strong, and unified baseline for video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 8", + "[18] Yicong Li, Xun Yang, Xindi Shang, and Tat-Seng Chua. Interventional video relation detection. In Proceedings of the 29th ACM International Conference on Multimedia (ACMMM), 2021. 1", + "[19] Yifei Li, Xinyi Yang, and Chenglin Xu. Dynamic scene graph generation via anticipatory pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 5", + "[20] Chang Liu, Yining Jin, Kaixuan Xu, Guohao Gong, and Yuesheng Mu. Beyond short-term snippet: Video relation detection with spatio-temporal global context. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[21] Yang Liu, Guanbin Li, and Liang Lin. Cross-modal causal relational reasoning for event-level visual question answering. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023. 2, 4", + "[22] Cewu Lu, Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Visual relationship detection with language priors. In Proceedings of the European Conference on Computer Vision (ECCV), 2016. 1", + "[23] Sourya Nag, Kyusong Min, Swami Tripathi, and Amit K. Roy-Chowdhury. Unbiased scene graph generation in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 4, 5, 6, 7", + "[24] Jeffrey Pennington, Richard Socher, and Christopher D. Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2014. 4", + "[25] Tao Pu, Tianshui Chen, Hefeng Wu, Yongyi Lu, and Liang Lin. Spatial-temporal knowledge-embedded transformer for" + ], + "bbox": [ + 516, + 92, + 905, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19055", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "video scene graph generation. IEEE Transactions on Image Processing (TIP), 2023. 1", + "[26] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2017. 5", + "[27] Scott Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D. Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the Fourth Workshop on Vision and Language, 2015. 1", + "[28] Xun Shang, Yicheng Li, Jing Xiao, Weiqing Ji, and Tat-Seng Chua. Video visual relation detection via iterative inference. In Proceedings of the 29th ACM International Conference on Multimedia (ACMMM), 2021. 2", + "[29] Claude E Shannon. A mathematical theory of communication. Bell System Technical Journal, 1948. 4", + "[30] Shubham Shit, Raphael Koner, Benedikt Wittmann, Johannes Paetzold, Ivan Ezhov, Haoyang Li, Jindong Pan, Sina Sharifzadeh, Georgios Kaissis, and Volker Tresp. Relationformer: A unified framework for image-to-graph generation. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2", + "[31] Yudong Teng, Liang Wang, Zhenyu Li, and Gang Wu. Target adaptive context aggregation for video scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR), 2021. 2, 6", + "[32] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 2", + "[33] Shizhe Wang, Linfeng Gao, Xinyi Lyu, Yulan Guo, Ping Zeng, and Jie Song. Dynamic scene graph generation via temporal prior inference. In Proceedings of the 30th ACM International Conference on Multimedia (ACMMM), 2022. 6", + "[34] Zhongdao Wang, Hengshuang Zhao, Ya-Li Li, Shengjin Wang, Philip Torr, and Luca Bertinetto. Do different tracking tasks require different appearance models? Advances in Neural Information Processing Systems (NeurIPS), 2021. 8", + "[35] Ziyang Wang, Yi-Lin Sung, Feng Cheng, Gedas Bertasius, and Mohit Bansal. Unified coarse-to-fine alignment for video-text retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 5", + "[36] Danfei Xu, Yuke Zhu, Christopher B. Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1", + "[37] Jingkang Yang, Jun Cen, Wenxuan Peng, Fangzhou Liu, Shuai amd Hong, Xiangtai Li, Kaiyang Zhou, Qifeng Chen, and Ziwei Liu. 4d panoptic scene graph generation. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 8", + "[38] Jingkang Yang, Wenxuan Peng, Xiangtai Li, Zujin Guo, Liangyu Chen, Bo Li, Zheng Ma, Kaiyang Zhou, Wayne Zhang, Chen Change Loy, et al. Panoptic video scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 8" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2", + "[40] Jiaxian Zhang, Kuan-Jui Shih, Ahmed Elgammal, Alexander Tao, and Bryan Catanzaro. Graphical contrastive losses for scene graph parsing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6", + "[41] Cheng Zheng, Linfeng Gao, Xinyi Lyu, Ping Zeng, Abdulmotaleb El Saddik, and Heng Tao Shen. Dual-branch hybrid learning network for unbiased scene graph generation. IEEE Transactions on Circuits and Systems for Video Technology (TCSVT), 2023. 2" + ], + "bbox": [ + 516, + 92, + 905, + 287 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19056", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Unbiased Video Scene Graph Generation via Visual and Semantic Dual Debiasing/168f707b-a7fa-4e7c-93c1-aac7eef770eb_model.json b/2025/Unbiased Video Scene Graph Generation via Visual and Semantic Dual Debiasing/168f707b-a7fa-4e7c-93c1-aac7eef770eb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8a03fc8b95b7418ab43476300ffb9571d94d899c --- /dev/null +++ b/2025/Unbiased Video Scene Graph Generation via Visual and Semantic Dual Debiasing/168f707b-a7fa-4e7c-93c1-aac7eef770eb_model.json @@ -0,0 +1,2398 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.141, + 0.13, + 0.857, + 0.176 + ], + "angle": 0, + "content": "Unbiased Video Scene Graph Generation via Visual and Semantic Dual Debiasing" + }, + { + "type": "text", + "bbox": [ + 0.278, + 0.203, + 0.714, + 0.222 + ], + "angle": 0, + "content": "Yanjun \\(\\mathrm{Li}^{1\\dagger}\\), Zhaoyang \\(\\mathrm{Li}^1\\), Honghui Chen\\(^2\\), Lizhi \\(\\mathrm{Xu}^1\\)" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.223, + 0.692, + 0.241 + ], + "angle": 0, + "content": "1University of Science and Technology of China" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.241, + 0.771, + 0.259 + ], + "angle": 0, + "content": "\\(^{2}\\)College of Physics and Information Engineering, Fuzhou University" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.261, + 0.835, + 0.276 + ], + "angle": 0, + "content": "ballinli@mail.ustc.edu.cn, lizhaoyang@mail.ustc.edu.cn, chh5840996@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.278, + 0.691, + 0.293 + ], + "angle": 0, + "content": "xulizhi@mail.ustc.edu.cn † Corresponding Author" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.328, + 0.327, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.36, + 0.485, + 0.633 + ], + "angle": 0, + "content": "Video Scene Graph Generation (VidSGG) aims to capture dynamic relationships among entities by sequentially analyzing video frames and integrating visual and semantic information. However, VidSGG is challenged by significant biases that skew predictions. To mitigate these biases, we propose a Visual and Semantic Awareness (VISA) framework for unbiased VidSGG. VISA addresses visual bias through memory-enhanced temporal integration that enhances object representations and concurrently reduces semantic bias by iteratively integrating object features with comprehensive semantic information derived from triplet relationships. This visual-semantics dual debiasing approach results in more unbiased representations of complex scene dynamics. Extensive experiments demonstrate the effectiveness of our method, where VISA outperforms existing unbiased VidSGG approaches by a substantial margin (e.g., +13.1% improvement in mR@20 and mR@50 for the SGCLS task under Semi Constraint)." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.647, + 0.222, + 0.663 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.838 + ], + "angle": 0, + "content": "Scene Graph Generation (SGG) [22] constructs structured scene representations by identifying objects and their relationships as triplets (). This structure enables various downstream tasks [7, 10, 14, 27], highlighting SGG as a pivotal area in computer vision. Initially applied to static images (ImgSGG) [36], SGG has been extended to video data (VidSGG) [3] to model dynamic interactions over time. However, VidSGG faces challenges due to biases favoring overrepresented classes, leading to the underrepresentation of rare classes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Recent methods have adopted attention-based strategies to address class imbalance and mitigate classification biases [23, 25]. However, these approaches, primarily relying on learnable attention layers, often amplify biases towards" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.327, + 0.906, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.426, + 0.907, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.487, + 0.907, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.564, + 0.906, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.677, + 0.907, + 0.776 + ], + "angle": 0, + "content": "Figure 1. Framework Comparison. We identify visual and semantic biases in VidSGG. Figures (a) and (c) show that prior methods rely on weak visual features and limited context, leading to inaccurate attention maps and skewed predicate distributions. In (b) and (d), our memory-enhanced visual debiasing and iterative semantic fusion mitigate both biases, resulting in stronger feature representations and more balanced predicate outcomes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.903 + ], + "angle": 0, + "content": "high-frequency classes [12]. To overcome this limitation, alternative methods have been proposed [12, 18] aiming to reduce spurious correlations and better account for label correlations. Although these methods alleviate bias to some extent, they often fall short of achieving the desired performance. A key oversight of these approaches is their heavy reliance on debiasing techniques, which neglect the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19047" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.411, + 0.107 + ], + "angle": 0, + "content": "intrinsic visual-semantic nature of scene graphs." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.108, + 0.482, + 0.333 + ], + "angle": 0, + "content": "Scene graphs inherently integrate both visual and semantic information, allowing biases to be categorized into visual biases and semantic biases. Visual biases arise from low visual quality in videos, such as blurring and occlusion [23], which degrade object feature representations. For instance, as shown in Fig. 1(a), objects may be partially visible or obscured over time, hindering models from capturing robust visual features for classification. Additionally, as illustrated in Fig. 1(c), the VidSGG task faces semantic biases. Previous methods [5, 23] rely solely on visual features to guide the relation generator for predicate prediction. This dependence on single-modal information leads to biased assumptions, similar to humans inferring relationships without sufficient context, thereby favoring more frequent predicate classes [21]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.334, + 0.485, + 0.664 + ], + "angle": 0, + "content": "To address scene graph biases, we categorize them into visual and semantic types, guiding our proposed framework, VISA (VIual and Semantic Awareness), for unbiased VidSGG, as illustrated in Fig. 1 (b) and (d). VISA mitigates visual bias through a Memory-Enhanced Temporal Integrator that enriches the representations of entities in the current frame with historical representations from previous frames. It employs a relation generator to construct initial scene graphs and a hierarchical semantic extractor to iteratively refine semantic information. Our hierarchical fusion strategy integrates multimodal visual and semantic contexts to address semantic biases. Specifically, the visual debiasing method reduces visual feature variance, while the semantic debiasing method increases the Kullback-Leibler (KL) divergence. This dual-debiasing approach effectively mitigates visual biases caused by poor visual feature quality and semantic biases arising from insufficient contextual information, while also addressing model bias due to data skew. Extensive experiments demonstrate the effectiveness of VISA, achieving superior performance in unbiased VidSGG compared to previous methods. Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.667, + 0.483, + 0.711 + ], + "angle": 0, + "content": "- We identify and explicitly define the visual and semantic biases affecting VidSGG, highlighting their impact on model performance." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.712, + 0.483, + 0.771 + ], + "angle": 0, + "content": "- We propose VISA, a debiasing framework for VidSGG that simultaneously addresses visual and semantic biases through memory-enhanced temporal integration and hierarchical semantic extraction." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.772, + 0.483, + 0.817 + ], + "angle": 0, + "content": "- We validate the effectiveness of our approach through extensive experiments and theoretical analyses, demonstrating state-of-the-art performance in unbiased VidSGG." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.667, + 0.483, + 0.817 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.824, + 0.233, + 0.839 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.318, + 0.864 + ], + "angle": 0, + "content": "2.1. Scene Graph Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Scene Graph Generation (SGG) approaches primarily focus on static image-based SGG (ImgSGG) and video-based" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "SGG (VidSGG). For ImgSGG, sequential encoders like LSTM [39] and attention mechanisms [32] capture global context, with attention methods [6, 16, 30] advancing standalone entity and predicate proposals (e.g., SGTR [16])." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.905, + 0.213 + ], + "angle": 0, + "content": "VidSGG extends SGG to dynamic contexts, requiring intra-frame relationships [11]. Recent methods improve temporal coherence [20] and efficiency by leveraging adaptive structures [31] and social context modeling [2]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.215, + 0.905, + 0.29 + ], + "angle": 0, + "content": "Despite their success, existing methods ignore unbiasedness. This results in biased SGG with inaccurate and misleading relationships in scenes. Consequently, such biases make models less reliable and less generalizable to real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.302, + 0.816, + 0.318 + ], + "angle": 0, + "content": "2.2. Unbiased Scene Graph Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.325, + 0.905, + 0.415 + ], + "angle": 0, + "content": "Unlike conventional SGG, unbiased SGG aims to reduce biases toward frequent objects and relationships. Key approaches include knowledge distillation [15] (e.g., LS-KD, which addresses multi-predicate challenges) and dual-branch architectures [41] (e.g., DHL, which balances attention between head and tail classes)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.905, + 0.521 + ], + "angle": 0, + "content": "For unbiased VidSGG, Transformer-based models [5] and Gaussian Mixture Models [23] have been explored to capture video dynamics better and reduce bias in visual relations. Iterative approaches [28] use conditional variables to improve video relation detection but lack a hierarchical strategy that integrates both visual and semantic representations, as in our method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.905, + 0.613 + ], + "angle": 0, + "content": "Unlike existing approaches that merely apply traditional debiasing strategies to VidSGG, our framework leverages the inherent visual-semantic nature of scene graphs. This enables a solution that is both experimentally validated and theoretically robust, effectively addressing the fundamental challenges in achieving unbiased VidSGG." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.628, + 0.604, + 0.643 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.652, + 0.624, + 0.666 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The VISA framework, depicted in Fig. 2, comprises Memory Guided Sequence Modeling (MGSM), and the Iterative Relation Generator (IRG). It begins with an off-the-shelf object detector to identify objects in video frames, \\(\\mathbf{b}_i^t\\) represents the object anchors of the \\(i\\)-th entity in the \\(t\\)-th frame. In bounding-box approaches (e.g., VidSGG), \\(\\mathbf{b}_i^t \\in \\mathbb{R}^4\\) denotes a box, while in panoptic segmentation, \\(\\mathbf{b}_i^t \\in \\mathbb{R}^n\\) represents the \\(n\\) boundary points. RoI-Align operations [9] are then applied to obtain feature representations, denoted as \\(\\mathbf{v}_i^t\\). The MGSM refines these features to produce subject and object representations, yielding enhanced features \\(\\hat{\\mathbf{v}}_i^t\\), while two separate FFNs generate distinct subject and object embeddings. The IRG leverages these visual representations, using the Hierarchical Semantic Extractor (HSE) to fuse visual and semantic information. Through iterative" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19048" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.086, + 0.912, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.354, + 0.908, + 0.412 + ], + "angle": 0, + "content": "Figure 2. Framework of VISA. Given an input video, VISA extracts object features using an off-the-shelf detector. Subsequently, the Memory Guided Sequence Modeling leverages these features to develop robust object representations. Following this, the Iterative Relation Generator employs the object features to predict relationships between subject-object pairs. It is important to note that the Hierarchical Semantic Extractor is not engaged during the initial phase of scene graph generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.425, + 0.483, + 0.455 + ], + "angle": 0, + "content": "processing, it deduces unbiased semantic relations and generates relation predictions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.463, + 0.409, + 0.48 + ], + "angle": 0, + "content": "3.2. Memory Guided Sequence Modeling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.484, + 0.59 + ], + "angle": 0, + "content": "We observe that as video duration increases, the Transformer tends to generate insufficient visual representations, particularly for small objects prone to blurring and occlusion (see Fig. 1 (a)). This insufficiency leads to high variance in feature representations due to the instability of visual information, causing the model to favor frequent entities and overlook less common ones." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.591, + 0.484, + 0.622 + ], + "angle": 0, + "content": "Our theoretical analysis reveals that this instability introduces noise into the object representations, modeled as:" + }, + { + "type": "equation", + "bbox": [ + 0.238, + 0.63, + 0.482, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {i} ^ {t} = \\mathbf {v} _ {i} + \\boldsymbol {\\epsilon} _ {i} ^ {t}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.484, + 0.732 + ], + "angle": 0, + "content": "where \\(\\epsilon_i^t\\) denotes zero-mean Gaussian noise with covariance \\(\\pmb{\\Sigma}\\), i.e., \\(\\mathbb{E}[\\epsilon_i^t] = 0\\) and \\(\\mathrm{Cov}[\\epsilon_i^t] = \\pmb{\\Sigma}\\). We assume these noise terms are independent across different time steps and objects. The observed features from the \\(t\\)-th frame are used directly as the object's representation:" + }, + { + "type": "equation", + "bbox": [ + 0.255, + 0.74, + 0.482, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {v}} _ {i} ^ {t} = \\mathbf {v} _ {i} ^ {t}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.767, + 0.483, + 0.797 + ], + "angle": 0, + "content": "Consequently, the expectation and variance of the visual representation are computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.806, + 0.482, + 0.823 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} \\left[ \\hat {\\mathbf {v}} _ {i} ^ {t} \\right] = \\mathbf {v} _ {i}, \\quad \\operatorname {V a r} \\left[ \\hat {\\mathbf {v}} _ {i} ^ {t} \\right] = \\boldsymbol {\\Sigma}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "While the single-frame estimation is unbiased, the noise introduced due to visual instability leads to high variance in the feature representations. This high variance destabilizes the feature representations and impairs subsequent relational predictions." + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.422, + 0.895, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.564, + 0.897, + 0.58 + ], + "angle": 0, + "content": "Figure 3. Structure of Memory Guided Sequence Modeling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.593, + 0.905, + 0.698 + ], + "angle": 0, + "content": "Understanding that the high variance stems from the instability of visual information motivated us to design the Memory Guided Sequence Modeling (MGSM) module (illustrated in Fig. 3). The MGSM leverages temporal visual information from preceding frames to enhance individual feature representations of frames, effectively reducing the variance caused by insufficient visual representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.698, + 0.905, + 0.728 + ], + "angle": 0, + "content": "Specifically, we compute adaptive weights \\(\\mathbf{W}_i^t\\) based on the current frame's features:" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.74, + 0.905, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\mathbf {W} _ {i} ^ {t} = \\sigma (\\mathrm {M L P} (\\mathbf {v} _ {i} ^ {t})), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.77, + 0.907, + 0.815 + ], + "angle": 0, + "content": "where \\(\\sigma\\) is the sigmoid function, and MLP denotes a multi-layer perceptron. Using these weights, we obtain the weighted feature representation:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.826, + 0.905, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {i} ^ {t *} = \\mathbf {W} _ {i} ^ {t} \\odot \\mathbf {v} _ {i} ^ {t} \\| \\mathbf {v} _ {i} ^ {t - 1}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.855, + 0.905, + 0.884 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes element-wise multiplication, and \\(\\parallel\\) indicates feature concatenation." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.907, + 0.9 + ], + "angle": 0, + "content": "To accumulate visual information and refine the current" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19049" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.438, + 0.107 + ], + "angle": 0, + "content": "representations, we introduce an update mechanism:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.114, + 0.483, + 0.133 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {i} ^ {t + 1} = (1 - \\lambda) \\mathbf {M} _ {i} ^ {t} + \\lambda \\mathbf {v} _ {i} ^ {t}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.142, + 0.483, + 0.171 + ], + "angle": 0, + "content": "As a result, the expectation and variance of the memory representation \\(\\mathbf{M}_i^t\\) stabilize over time:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.179, + 0.483, + 0.21 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} \\left[ \\mathbf {M} _ {i} ^ {t} \\right] = \\mathbf {v} _ {i}, \\operatorname {V a r} \\left[ \\mathbf {M} _ {i} ^ {t} \\right] = \\frac {\\lambda \\boldsymbol {\\Sigma}}{2 - \\lambda} \\approx \\frac {\\lambda \\boldsymbol {\\Sigma}}{2}, \\text {f o r s m a l l} \\lambda . \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.217, + 0.484, + 0.307 + ], + "angle": 0, + "content": "This analysis shows that a smaller \\(\\lambda\\) reduces variance, enhancing the robustness of feature estimation. However, \\(\\lambda\\) cannot be too small, as excessively small values slow the adaptation to new information and may introduce bias. We provide detailed proofs of the variance calculation and the associated trade-off in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.308, + 0.484, + 0.339 + ], + "angle": 0, + "content": "The final enhanced feature representation is obtained through a dual attention layer:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.347, + 0.483, + 0.365 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {v}} _ {i} ^ {t} = \\operatorname {A t t e n t i o n} \\left(\\mathbf {Q} = \\mathbf {v} _ {i} ^ {t *}, \\mathbf {K} = \\mathbf {M} _ {i} ^ {t}, \\mathbf {V} = \\mathbf {v} _ {i} ^ {t}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.373, + 0.484, + 0.478 + ], + "angle": 0, + "content": "Finally, the final enhanced feature representation is processed by the Transformer Encoder, as in [23], to generate the subject and object representations. By adaptively enhancing features and leveraging the reduced variance from the memory representation, the MGSM addresses the high variance caused by insufficient visual representations, thereby mitigating visual bias." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.484, + 0.554 + ], + "angle": 0, + "content": "Our experimental results (Fig. 6) confirm our theoretical insights by demonstrating that MGSM enhances visual representations. This enhancement reduces the high variance caused by insufficient visual information, which is crucial for addressing visual bias in VidSGG." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.562, + 0.347, + 0.577 + ], + "angle": 0, + "content": "3.3. Iterative Relation Generator" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.719 + ], + "angle": 0, + "content": "Existing methods for unbiased SGG predominantly predict predicates directly from visual representations [5, 23]. However, this approach is akin to humans attempting to deduce relationships solely based on visual cues without sufficient contextual information, which easily introduces semantic bias [21]. Specifically, when lacking context, models tend to over-rely on the imbalanced and biased prior distributions present in the training data, leading to biased relationship predictions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To alleviate semantic bias, we start by analyzing the problem from an information-theoretic perspective. Semantic bias occurs when a model lacks sufficient contextual information, leading to high uncertainty in predicting relationships. This uncertainty causes the model to over-rely on biased prior distributions from the training data. We use entropy \\( H \\) to quantify this uncertainty [29]. We consider the conditional entropy \\( H(r_{ij} \\mid v_i, v_j, S) \\), where \\( v_i \\) and \\( v_j \\) are the feature representations of the subject and object, respectively. Based on the properties of conditional entropy [29], the additional context provided by \\( S \\) decreases the uncertainty in predicting relationships:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.092, + 0.907, + 0.109 + ], + "angle": 0, + "content": "\\[\nH \\left(r _ {i j} \\mid v _ {i}, v _ {j}, S\\right) \\leq H \\left(r _ {i j} \\mid v _ {i}, v _ {j}\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.117, + 0.907, + 0.208 + ], + "angle": 0, + "content": "This reduction in conditional entropy means the model is less uncertain and relies less on biased priors. By incorporating context \\( S \\), the likelihood term \\( P(v_{i}, v_{j}, S \\mid r_{ij}) \\) in Bayes' theorem becomes more informative, effectively reducing the relative influence of the prior \\( P(r_{ij}) \\). According to Bayes' theorem:" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.215, + 0.907, + 0.25 + ], + "angle": 0, + "content": "\\[\nP \\left(r _ {i j} \\mid v _ {i}, v _ {j}, S\\right) = \\frac {P \\left(v _ {i} , v _ {j} , S \\mid r _ {i j}\\right) \\cdot P \\left(r _ {i j}\\right)}{P \\left(v _ {i} , v _ {j} , S\\right)}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.257, + 0.906, + 0.304 + ], + "angle": 0, + "content": "To quantify the change in reliance on the prior, we consider the Kullback-Leibler (KL) divergence between the posterior \\( P(r_{ij} \\mid v_i, v_j, S) \\) and the prior \\( P(r_{ij}) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.308, + 0.905, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} D _ {\\mathrm {K L}} \\left(P \\left(r _ {i j} \\mid v _ {i}, v _ {j}, S\\right) \\| P \\left(r _ {i j}\\right)\\right) \\\\ = \\sum_ {r _ {i j}} P (r _ {i j} \\mid v _ {i}, v _ {j}, S) \\log \\frac {P (r _ {i j} \\mid v _ {i} , v _ {j} , S)}{P (r _ {i j})} \\\\ = H \\left(r _ {i j} \\mid v _ {i}, v _ {j}, P \\left(r _ {i j}\\right)\\right) - H \\left(r _ {i j} \\mid v _ {i}, v _ {j}, S\\right). \\tag {11} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.404, + 0.906, + 0.479 + ], + "angle": 0, + "content": "As the conditional entropy \\( H(r_{ij} \\mid v_i, v_j, S) \\) decreases due to the additional context, the KL divergence increases. This indicates that the posterior distribution diverges from the biased prior, reflecting a reduced reliance on the prior and effectively mitigating semantic bias." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.48, + 0.906, + 0.585 + ], + "angle": 0, + "content": "This theoretical insight motivates our approach to incorporate contextual information into the model. Based on this, we propose the Iterative Relation Generator (IRG), an iterative semantic debiasing component within our framework. The IRG iteratively introduces contextual information to refine relationship predictions, thereby reducing uncertainty and dependence on biased priors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.906, + 0.647 + ], + "angle": 0, + "content": "In the first iteration, the IRG infers semantic information from each subject-object pair to generate semantic embeddings. For each subject-object pair \\((j,i)\\) at frame \\(t\\), we construct the composite objects features \\(\\mathbf{p}_{j,i}^{t}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.65, + 0.905, + 0.682 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {j, i} ^ {t} = f _ {v} \\left(\\mathbf {v} _ {j} ^ {t}\\right) \\| f _ {v} \\left(\\mathbf {v} _ {i} ^ {t}\\right) \\| \\left(f _ {u} \\left(\\mathbf {u} _ {j i} ^ {t}\\right) + f _ {\\text {b o x}} \\left(b _ {j} ^ {t}, b _ {i} ^ {t}\\right)\\right) \\| \\mathbf {s} _ {j} ^ {t} \\| \\mathbf {s} _ {i} ^ {t}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.906, + 0.764 + ], + "angle": 0, + "content": "where \\(\\mathbf{v}_j^t\\) and \\(\\mathbf{v}_i^t\\) are the visual feature representations of the subject and object. The tensor \\(\\mathbf{u}_{ji}^{t}\\) is the union box feature map extracted via RoIAlign [9], and \\(\\mathbf{s}_j^t\\) and \\(\\mathbf{s}_i^t\\) are the semantic embeddings obtained from GloVe [24]. The functions \\(f_{v}, f_{u}\\), and \\(f_{\\mathrm{box}}\\) are feature transformation functions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.906, + 0.844 + ], + "angle": 0, + "content": "We then perform initial relationship prediction, obtaining predicate categories such as attention, spatial, and contacting. We select the semantic embeddings with the highest probabilities from these distributions to form integrated triplet embeddings \\(\\mathbf{C}_{pre,(j,i)}^t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.849, + 0.905, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} _ {p r e, (j, i)} ^ {t} = \\left[ \\begin{array}{l} f _ {v} \\left(\\mathbf {v} _ {j} ^ {t}\\right) \\| f _ {a, (j, i)} ^ {t} \\| f _ {v} \\left(\\mathbf {v} _ {i} ^ {t}\\right) \\\\ f _ {v} \\left(\\mathbf {v} _ {j} ^ {t}\\right) \\| f _ {s, (j, i)} ^ {t} \\| f _ {v} \\left(\\mathbf {v} _ {i} ^ {t}\\right) \\\\ f _ {v} \\left(\\mathbf {v} _ {j} ^ {t}\\right) \\| f _ {c, (j, i)} ^ {t} \\| f _ {v} \\left(\\mathbf {v} _ {i} ^ {t}\\right) \\end{array} \\right], \\tag {13}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19050" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.487, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.233, + 0.466, + 0.248 + ], + "angle": 0, + "content": "Figure 4. Structure of Hierarchical Semantics Extractor." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.261, + 0.483, + 0.291 + ], + "angle": 0, + "content": "where \\( f_{a,(j,i)}^t, f_{s,(j,i)}^t \\), and \\( f_{c,(j,i)}^t \\) are the semantic embeddings corresponding to the respective predicates." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.291, + 0.484, + 0.412 + ], + "angle": 0, + "content": "To further exploit hierarchical contextual information, we design the Hierarchical Semantics Extractor (HSE) (as shown in Fig. 4). Inspired by the multi-hierarchical feature fusion method [35], the HSE combines composite object features with integrated triplets to obtain fine-grained visual and semantic representations. We decompose the composite objects features \\(\\mathbf{p}_{j,i}^{t}\\) into fine-grained subject and object representations as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.42, + 0.484, + 0.45 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {S} _ {\\mathbf {p} _ {j, i} ^ {t}} ^ {t} = f _ {v} (\\mathbf {v} _ {j} ^ {t}) \\| \\left(f _ {u} (\\mathbf {u} _ {j} ^ {t}) + f _ {\\text {b o x}} \\left(b _ {j} ^ {t}, b _ {i} ^ {t}\\right)\\right) \\| \\mathbf {s} _ {j} ^ {t}, \\\\ \\mathbf {G} _ {i} ^ {t} = f _ {v} (t ^ {t}) \\| \\left(f _ {v} (t ^ {t}) + f _ {\\text {b o x}} (t ^ {t}, t ^ {t})\\right) \\| \\end{array} \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.442, + 0.425, + 0.466 + ], + "angle": 0, + "content": "\\[\n\\mathbf {O} _ {\\mathbf {p} _ {j, i} ^ {t}} ^ {t} = f _ {v} \\left(\\mathbf {v} _ {i} ^ {t}\\right) \\| \\left(f _ {u} \\left(\\mathbf {u} _ {j i} ^ {t}\\right) + f _ {\\text {b o x}} \\left(b _ {j} ^ {t}, b _ {i} ^ {t}\\right)\\right) \\| \\mathbf {s} _ {i} ^ {t}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.483, + 0.51 + ], + "angle": 0, + "content": "Similarly, we decompose the integrated triplet embeddings \\(\\mathbf{C}_{pre,(j,i)}^t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.516, + 0.484, + 0.546 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {S} _ {\\mathbf {C} _ {(j, i)} ^ {t}} ^ {t} = \\left\\{f _ {v} \\left(\\mathbf {v} _ {j} ^ {t}\\right) \\| f _ {k} ^ {t} \\mid k \\in \\{a, s, c \\} \\right\\}, \\\\ \\text {C} _ {i} ^ {t} = \\left\\{f _ {v} ^ {t + 1} \\mid f _ {k} ^ {t + 1} \\mid k \\in \\{a, c \\} \\right\\}. \\end{array} \\tag {15}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.533, + 0.482, + 0.564 + ], + "angle": 0, + "content": "\\[\n\\mathbf {O} _ {\\mathbf {C} _ {(j, i)} ^ {t}} ^ {t} = \\left\\{f _ {k} ^ {t} \\| f _ {v} \\left(\\mathbf {v} _ {i} ^ {t}\\right) \\mid k \\in \\{a, s, c \\} \\right\\}. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.484, + 0.688 + ], + "angle": 0, + "content": "These fine-grained representations are concatenated with downsampled embeddings from the previous cross-attention layer via a stride-two convolutional layer, serving as inputs to the next cross-attention layer. This hierarchical structure enables the model to capture multi-level contextual information, effectively reducing uncertainty and reliance on biased priors. The HSE is active for iteration counts \\( N \\geq 1 \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.81 + ], + "angle": 0, + "content": "After \\(N\\) iterations of the IRG, the final inter-object predicates are generated by the Spatial Encoder and Temporal Decoder modules adopted from [5]. By iteratively leveraging contextual information, our method effectively mitigates semantic bias. Experimental results (as shown in Fig. 7) validate that incorporating contextual information reduces the model's reliance on biased priors and improves relationship predictions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.819, + 0.288, + 0.835 + ], + "angle": 0, + "content": "3.4. Training and Testing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Training. In the training phase, VISA synthesizes object features from MGSM for preliminary scene graph construction. Upon reaching iteration \\( N \\geq 1 \\), it activates the HSE to generate a hierarchical visual-semantics fusion embedding." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "In Section 3.2, the memory representation \\(\\mathbf{M}_i^t\\) is initialized as a zero tensor, while the embeddings for the three relationships are randomly initialized as described in Section 3.3. We optimize the comprehensive loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.164, + 0.907, + 0.181 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {t o t a l}} = L _ {p} + L _ {e} + L _ {\\text {c o n t r a}}. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.192, + 0.906, + 0.237 + ], + "angle": 0, + "content": "Here, \\( L_{p} \\) and \\( L_{e} \\) denote the predicate and entity classification losses, both computed via cross-entropy, and \\( L_{\\mathrm{contra}} \\) is the contrastive loss, all defined as in [23]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.238, + 0.906, + 0.283 + ], + "angle": 0, + "content": "Testing. After training, the proposed MGSM produces object predictions through the visual debiasing component, while the IRG performs semantic debiasing." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.299, + 0.646, + 0.315 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.324, + 0.706, + 0.339 + ], + "angle": 0, + "content": "4.1. Dataset and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.347, + 0.906, + 0.437 + ], + "angle": 0, + "content": "Dataset. We conducted experiments on the Action Genome (AG) dataset [11], the largest benchmark for VidSGG. AG consists of 234, 253 annotated frames containing 476, 229 bounding boxes across 35 object categories. It provides 1, 715, 568 predicate instances spanning 26 distinct relationship classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.907, + 0.562 + ], + "angle": 0, + "content": "Metrics and Evaluation Setup. We employed the standard metrics for unbiased VidSGG evaluation, mean Recall@K (mR@K), considering \\( K \\in \\{10, 20, 50\\} \\). Following established protocols [5, 11, 23], we engaged in three SGG tasks, PREDCLS, SGCLS and SGDET. Evaluation was conducted across three settings, With Constraint, Semi Constraint, and No Constraint. Due to page constraints, further details are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.573, + 0.729, + 0.589 + ], + "angle": 0, + "content": "4.2. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.906, + 0.747 + ], + "angle": 0, + "content": "Following prior work [5, 19, 23], we adopted Faster R-CNN [26] with ResNet-101 [8] as the object detector, initially trained on the AG dataset. To ensure a fair comparison, we utilized the official implementations of these methods. For our MGSM module, we set the \\(\\lambda\\) parameter to 0.04 for the SGCLS task and 0.06 for the SGDET task. In the IRG module, we implemented a dual-procedure setup, enabling iterative relational inference with the number of iterations \\(N\\) set to 1. Due to page constraints, further details are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.798, + 0.774 + ], + "angle": 0, + "content": "4.3. Comparison to existing methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We began with a qualitative analysis, shown in Fig. 5, illustrating results under the Semi Constraint condition, which closely resembles real-world scenarios [5] where multiple predicates typically describe scenes. False positive relationships and object classifications were highlighted in purple blocks. The attention intensity demonstrated that our framework effectively enhanced object representations (e.g., pilow) through visual debiasing across frames from \\( t \\) to \\( t + 1 \\)," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19051" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.131, + 0.074, + 0.868, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.365, + 0.908, + 0.409 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of our proposed method. Comparison with the open-sourced state-of-the-art model TEMPURA [23] under Semi Constraint, closely matching ground truth [5]. Visualizations show attention intensity and video scene graphs from Frame \\(t\\) to \\(t + 1\\). False positive relationships and object classifications are highlighted in purple blocks." + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.415, + 0.908, + 0.445 + ], + "angle": 0, + "content": "Table 1. Quantitative results comparisons with unbiased SGG methods under With, Semi, and No Constraints. The experiments are conducted on the dataset AG [11]. The best results are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.195, + 0.454, + 0.803, + 0.684 + ], + "angle": 0, + "content": "
ConstraintMethodPREDCLSSGCLSSGDET
mR@10mR@20mR@50mR@10mR@20mR@50mR@10mR@20mR@50
WithRelDN [40]6.26.26.23.43.43.43.33.33.3
HCRD supervised [1]-------8.39.1
TRACE [31]15.215.215.28.98.98.98.28.28.2
ISGG [13]------19.722.9-
STTran [5]37.840.140.227.228.028.020.820.822.2
STTran-TPI [33]37.340.640.628.329.329.320.220.221.8
TEMPURA [23]42.946.346.334.035.235.222.622.623.7
FloCoDe [12]44.849.249.337.439.239.424.224.227.9
VISA (Ours)46.952.052.040.842.542.627.327.330.7
SemiTEMPURA [23]40.744.544.636.939.539.521.821.822.5
VISA (Ours)51.356.356.447.852.652.631.731.733.2
NoRelDN [40]31.263.175.518.636.942.67.518.833.7
TRACE [31]50.973.682.731.942.746.322.831.341.8
STTran [5]51.467.782.740.750.158.820.929.739.2
TEMPURA [23]61.585.198.048.361.166.424.733.943.7
FloCoDe [12]63.286.998.649.763.869.228.635.447.2
VISA (Ours)65.889.199.852.066.371.430.736.750.4
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.706, + 0.482, + 0.812 + ], + "angle": 0, + "content": "compared with the TEMPURA [23]. Additionally, our approach inferred more accurate semantic predicates (e.g., sit on) within a single frame via contextual information. These visualizations confirm the correctness of partitioning scene graph biases into visual and semantic biases and demonstrate that our framework, VISA, significantly mitigates both in VidSGG." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.825, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Fig. 6 and Fig. 7 further validated our theoretical analysis. Fig. 6 illustrated the variance of feature representations across training epochs, with shaded regions indicating the standard deviation. TEMPURA exhibited higher variance, suggesting less stable feature representations," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.706, + 0.907, + 0.844 + ], + "angle": 0, + "content": "whereas VISA demonstrated significantly reduced variance and tighter standard deviation regions. This reduction confirmed the effectiveness of the MGSM module in stabilizing feature representations. Additionally, Fig. 7 presented the confusion matrix comparing relationship predictions between TEMPURA and VISA. The reduced influence from prior dataset biases in VISA validated our approach to mitigate the semantics bias in relationship predictions using KL divergence principles." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Our approach achieved state-of-the-art performance across all mR@K metrics, as shown in Table 1. VISA demonstrated improvements ranging from a minimum of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19052" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.088, + 0.434, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.314, + 0.487, + 0.371 + ], + "angle": 0, + "content": "Figure 6. Comparison of TEMPURA and VISA visual feature representation variance. Visualization of visual feature representation variance across training epochs. The shaded regions represent the standard deviation." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.378, + 0.489, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.519, + 0.482, + 0.677 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Figure 7. Qualitative results for relationship confusion matrix. Results for the subject-object pairs, person-sofa, and person-floor. \\(+1.2\\%\\) in mR@50 for the PREDCLS task under the No Constraint scenario to a maximum of \\(+13.1\\%\\) in both mR@20 and mR@50 for the SGCLS task under the Semi Constraint scenario. These results validate the correctness of partitioning scene graph biases into visual and semantic biases and demonstrate the effectiveness of VISA in mitigating them, thereby advancing unbiased VidSGG." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Furthermore, as shown in Fig. 8, our VISA model surpassed open-source unbiased VidSGG methods in mR@10 across the three predicate class distributions (HEAD, BODY,TAIL) on AG under both With Constraint and No Constraint settings. Here, HEAD represented classes" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.088, + 0.66, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.186, + 0.627, + 0.196 + ], + "angle": 0, + "content": "(a) PredCLS" + }, + { + "type": "image", + "bbox": [ + 0.669, + 0.089, + 0.787, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.711, + 0.185, + 0.753, + 0.196 + ], + "angle": 0, + "content": "(b) SGCLS" + }, + { + "type": "image", + "bbox": [ + 0.794, + 0.089, + 0.907, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.836, + 0.186, + 0.877, + 0.196 + ], + "angle": 0, + "content": "(c) SGDET" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.2, + 0.659, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.295, + 0.626, + 0.304 + ], + "angle": 0, + "content": "(a) PredCLS" + }, + { + "type": "image", + "bbox": [ + 0.669, + 0.2, + 0.785, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.712, + 0.295, + 0.753, + 0.304 + ], + "angle": 0, + "content": "(b) SGCLS" + }, + { + "type": "image", + "bbox": [ + 0.794, + 0.2, + 0.907, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.836, + 0.295, + 0.877, + 0.304 + ], + "angle": 0, + "content": "(c) SGDET" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.315, + 0.907, + 0.37 + ], + "angle": 0, + "content": "Figure 8. Quantitative results of our proposed method. Compare our proposed method with open-sourced methods, TEMPURA [23] and STTran [5] on the mR@10 metric under With Constraint and No Constraint." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.385, + 0.907, + 0.492 + ], + "angle": 0, + "content": "with \\(\\geq 100,000\\) training samples, \\(BODY\\) with fewer than 100,000, and \\(TAIL\\) with fewer than 8,000. VISA consistently outperformed all competitors across mR@10 metrics, achieving a notable \\(+11.0\\%\\) improvement for the tail class in the SGDET task under No Constraint. This strong performance on tail classes highlighted the effectiveness of VISA's dual debiasing strategy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.501, + 0.68, + 0.516 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.907, + 0.689 + ], + "angle": 0, + "content": "Ablation study on VISA modules. Table 2 evaluates the effectiveness of VISA modules. Removing MGSM resulted in at least a \\(-2.2\\%\\) drop in mR@10 (SGCLS, Semi Constraint), confirming its capability to alleviate visual bias. Similarly, removing IRG caused at least a \\(-0.7\\%\\) decline in mR@10 (SGCLS, No Constraint), validating its role in reducing semantic bias. These results justify the division of biases into visual and semantic components and demonstrate VISA's effectiveness in mitigating both. Note that MGSM was not applied to the PREDCLS task, as it utilizes ground-truth visual inputs, rendering MGSM unnecessary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.856 + ], + "angle": 0, + "content": "Influence of increasing iteration count \\(N\\). We analyzed the effect of iteration number \\(N\\) on VISA's performance. Due to computational constraints, we conducted experiments with \\(N = 2\\) and \\(N = 3\\), denoted as \\(\\mathrm{VISA}_{N=2}\\) and \\(\\mathrm{VISA}_{N=3}\\), respectively. Results in Table 3 show that increasing \\(N\\) consistently enhances the mR@K performance for unbiased VidSGG, though the gain from \\(N = 2\\) to \\(N = 3\\) is smaller compared to \\(N = 1\\) to \\(N = 2\\). Additionally, training time approximately doubled when increasing from \\(N = 1\\) to \\(N = 2\\). Further details are provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "The value of the \\(\\lambda\\). Table 4 showed results for different values of \\(\\lambda\\). VISA achieved optimal performance for SGCLS at \\(\\lambda = 0.04\\), while the best results for SGDET occur" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19053" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.202, + 0.09, + 0.794, + 0.104 + ], + "angle": 0, + "content": "Table 2. Ablation study of MGSM and IRG under With Constraint, Semi Constraint, No Constraint." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.109, + 0.912, + 0.198 + ], + "angle": 0, + "content": "
With ConstraintSemi ConstraintNo Constraint
PREDCLSSGCLSSGDETPREDCLSSGCLSSGDETPREDCLSSGCLSSGDETPREDCLSSGCLSSGDETPREDCLSSGCLSSGDET
mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20
VISA46.952.040.842.527.327.351.356.347.852.631.731.765.889.152.066.330.736.7
w/o MGSM46.952.035.738.218.824.651.356.345.649.324.127.865.889.149.062.027.934.7
w/o MGSM & IRG42.946.334.035.218.522.640.744.536.939.518.521.861.585.148.361.124.733.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.35, + 0.206, + 0.645, + 0.219 + ], + "angle": 0, + "content": "Table 3. Influence of increasing iteration count \\(N\\)" + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.224, + 0.905, + 0.314 + ], + "angle": 0, + "content": "
With ConstraintSemi ConstraintNo Constraint
PREDCLSSGCLSSGDETPREDCLSSGCLSSGDETPREDCLSSGCLSSGDET
mR@10mR@20mR@10mR@20mR@10mR@20mR@10mR@20mR@10mR@20mR@10mR@20mR@10mR@20mR@10mR@20mR@10mR@20
VISA46.952.040.842.527.327.351.356.347.852.631.731.765.889.152.066.330.736.7
VISA_N=247.953.041.743.425.528.752.556.948.253.627.531.966.090.853.167.031.837.2
VISA_N=348.153.342.244.025.828.952.857.148.553.927.832.166.291.253.567.232.037.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.145, + 0.323, + 0.851, + 0.337 + ], + "angle": 0, + "content": "Table 4. Different values of \\( \\lambda \\) on SGCLS and SGDET,we summarize the results in mean-Recall@K,with best in bold." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.342, + 0.905, + 0.464 + ], + "angle": 0, + "content": "
λWith ConstraintSemi ConstraintNo Constraint
SGCLSSGDETSGCLSSGDETSGCLSSGDET
mR@10mR@20mR@50mR@10mR@20mR@50mR@10mR@20mR@50mR@10mR@20mR@50mR@10mR@20mR@50mR@10mR@20mR@50
0.00534.236.236.217.921.522.540.142.642.721.526.628.146.457.957.924.832.542.5
0.0136.037.437.418.622.223.240.744.644.723.728.630.047.259.865.225.434.043.0
0.0239.241.241.818.422.123.043.147.047.122.527.128.148.362.467.524.532.841.7
0.0440.842.542.619.823.324.347.852.652.624.529.531.152.066.371.426.733.641.7
0.0635.136.436.427.327.330.741.845.845.931.731.733.246.959.364.730.736.750.4
0.0834.335.635.618.321.622.540.243.743.723.327.828.947.159.665.724.732.740.6
0.133.234.434.417.520.322.139.142.542.522.425.427.246.259.664.322.131.439.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.111, + 0.476, + 0.465, + 0.49 + ], + "angle": 0, + "content": "Table 5. Performance evaluation on the PVSG [38] Dataset" + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.494, + 0.482, + 0.561 + ], + "angle": 0, + "content": "
MethodPVSG Metrics
Stage-1Stage-2mR@20mR@50mR@100
VPS [4, 17]Transformer0.360.390.40
IPS+T [4, 34]Transformer1.751.862.03
MGSMIRG8.828.9410.51
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.574, + 0.483, + 0.601 + ], + "angle": 0, + "content": "Table 6. Performance evaluation on the PSG4D-GTA and PSG4D-HOI [37] 4DPVSG Datasets." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.607, + 0.485, + 0.658 + ], + "angle": 0, + "content": "
MethodPSG4D-GTAPSG4D-HOI
mR@20mR@50mR@100mR@20mR@50mR@100
PSG4DFounder (P) [37]2.102.933.133.103.954.17
PSG4DFounder (R) [37]3.313.854.023.654.164.97
MGSM + IRG10.8311.5111.9810.9512.3212.87
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.673, + 0.483, + 0.793 + ], + "angle": 0, + "content": "at \\(\\lambda = 0.06\\). The differences in optimal \\(\\lambda\\) values reflect the distinct characteristics of these tasks. Specifically, SGCLS benefits from ground-truth data, enabling effective updates to the memory representations with a smaller \\(\\lambda\\). In contrast, SGDET, which builds scene graphs from scratch, performs better with a slightly higher \\(\\lambda\\). Notably, since the MSGM module was not applied to the PREDCLS task, the parameter \\(\\lambda\\) is also not involved in PREDCLS." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.795, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Exploration on PVSG and 4DPVSG Datasets. We extended our framework to Panoptic Video Scene Graph Generation (PVSG) and 4D Panoptic Scene Graph Generation (4DPVSG) tasks [37, 38]. Unlike VidSGG, PVSG utilizes pixel-level masks for grounding nodes. Using the IPS+T+Transformer codebase, we integrated MGSM with a Memory-Enhanced Temporal Integrator and employed" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.907, + 0.645 + ], + "angle": 0, + "content": "UniTrack [34] for temporal linking. For relation classification, IRG was adapted for iterative semantic debiasing. The 4DPVSG task further generalized PVSG to dynamic 4D contexts, employing PSG4DFormer variants with point cloud (P) and RGB-D (R) inputs. As shown in Table 5, our method significantly outperformed baselines, achieving minimum improvements of \\(+7.52\\%\\) mR@20 (PVSG) and \\(+7.30\\%\\) mR@20 (4DPVSG, PSG4D-HOI). These results confirm the broad applicability of our dual debiasing approach across VidSGG and indicate potential benefits for other generative and embodied intelligence applications." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.655, + 0.634, + 0.67 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this work, we present VISA, an approach for unbiased video scene graph generation that addresses both visual and semantic biases inherent in scene graphs. VISA reduces visual bias by enriching current frame representations with historical context to decrease visual feature variance, and mitigates semantic bias by increasing the KL divergence between model predictions and biased priors. Our framework integrates hierarchical semantic contexts with visual features, enabling more unbiased scene graph generation. Extensive experimental and theoretical analyses demonstrate the effectiveness of VISA. Future work will explore extending our visual-semantic debiasing approach to other multimodal domains such as visual question answering and video captioning, as well as applications in embodied intelligence, including robotics perception and interactive agents." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19054" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.173 + ], + "angle": 0, + "content": "We gratefully acknowledge the three anonymous reviewers for their constructive feedback on our CVPR 2025 submission. This work was supported by personal funding from the corresponding author." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.198, + 0.188, + 0.214 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.223, + 0.484, + 0.279 + ], + "angle": 0, + "content": "[1] João Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.281, + 0.484, + 0.336 + ], + "angle": 0, + "content": "[2] Shizhe Chen, Zhiyuan Shi, Pascal Mettes, and Cees G. M. Snoek. Social fabric: Tubelet compositions for video relation detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.337, + 0.483, + 0.392 + ], + "angle": 0, + "content": "[3] Shizhe Chen, Jing Xiao, and Lin Chen. Video scene graph generation from single-frame weak supervision. In Proceedings of the Eleventh International Conference on Learning Representations (ICLR), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.393, + 0.484, + 0.463 + ], + "angle": 0, + "content": "[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.464, + 0.484, + 0.533 + ], + "angle": 0, + "content": "[5] Yunhan Cong, Wentao Liao, Heiko Ackermann, Bodo Rosenhahn, and Michael Ying Yang. Spatial-temporal transformer for dynamic scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.534, + 0.484, + 0.588 + ], + "angle": 0, + "content": "[6] Yunhan Cong, Michael Ying Yang, and Bodo Rosenhahn. Reltr: Relation transformer for scene graph generation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.591, + 0.484, + 0.646 + ], + "angle": 0, + "content": "[7] Linfeng Gao, Bin Wang, and Wenmin Wang. Image captioning with scene-graph based semantic concepts. In Proceedings of the International Conference on Machine Learning and Computing, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.484, + 0.745 + ], + "angle": 0, + "content": "[9] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.484, + 0.801 + ], + "angle": 0, + "content": "[10] Marcel Hildebrandt, Haoyang Li, Raphael Koner, Volker Tresp, and Stephan Gunnemann. Scene graph reasoning for visual question answering. arXiv preprint arXiv:2007.01072, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[11] Jiading Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as compositions of spatiotemporal scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[12] Anant Khandelwal. Flocode: Unbiased dynamic scene graph generation with temporal consistency and correlation debias" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.223, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.134 + ], + "angle": 0, + "content": "ing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2024. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.906, + 0.176 + ], + "angle": 0, + "content": "[13] Shreyank Khandelwal and Leonid Sigal. Iterative scene graph generation. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.906, + 0.245 + ], + "angle": 0, + "content": "[14] Hengshuang Li, Xin Li, Behnam Karimi, Jie Chen, and Ming Sun. Joint learning of object graph and relation graph for visual question answering. In Proceedings of the IEEE International Conference on Multimedia and Expo (ICME), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.906, + 0.315 + ], + "angle": 0, + "content": "[15] Lifeng Li, Jing Xiao, Huan Shi, Wei Wang, Jing Shao, Alan Liu, Yueting Yang, and Lin Chen. Label semantic knowledge distillation for unbiased scene graph generation. IEEE Transactions on Circuits and Systems for Video Technology (TCSVT), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.906, + 0.372 + ], + "angle": 0, + "content": "[16] Ruofei Li, Shuhui Zhang, and Xiaochun He. Sgtr: End-to-end scene graph generation with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.372, + 0.906, + 0.453 + ], + "angle": 0, + "content": "[17] Xiangtai Li, Wenwei Zhang, Jiangmiao Pang, Kai Chen, Guangliang Cheng, Yunhai Tong, and Chen Change Loy. Video k-net: A simple, strong, and unified baseline for video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.906, + 0.511 + ], + "angle": 0, + "content": "[18] Yicong Li, Xun Yang, Xindi Shang, and Tat-Seng Chua. Interventional video relation detection. In Proceedings of the 29th ACM International Conference on Multimedia (ACMMM), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.906, + 0.567 + ], + "angle": 0, + "content": "[19] Yifei Li, Xinyi Yang, and Chenglin Xu. Dynamic scene graph generation via anticipatory pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.906, + 0.636 + ], + "angle": 0, + "content": "[20] Chang Liu, Yining Jin, Kaixuan Xu, Guohao Gong, and Yuesheng Mu. Beyond short-term snippet: Video relation detection with spatio-temporal global context. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.636, + 0.906, + 0.692 + ], + "angle": 0, + "content": "[21] Yang Liu, Guanbin Li, and Liang Lin. Cross-modal causal relational reasoning for event-level visual question answering. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.906, + 0.748 + ], + "angle": 0, + "content": "[22] Cewu Lu, Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Visual relationship detection with language priors. In Proceedings of the European Conference on Computer Vision (ECCV), 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.906, + 0.816 + ], + "angle": 0, + "content": "[23] Sourya Nag, Kyusong Min, Swami Tripathi, and Amit K. Roy-Chowdhury. Unbiased scene graph generation in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.906, + 0.873 + ], + "angle": 0, + "content": "[24] Jeffrey Pennington, Richard Socher, and Christopher D. Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[25] Tao Pu, Tianshui Chen, Hefeng Wu, Yongyi Lu, and Liang Lin. Spatial-temporal knowledge-embedded transformer for" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19055" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "video scene graph generation. IEEE Transactions on Image Processing (TIP), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[26] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[27] Scott Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D. Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the Fourth Workshop on Vision and Language, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[28] Xun Shang, Yicheng Li, Jing Xiao, Weiqing Ji, and Tat-Seng Chua. Video visual relation detection via iterative inference. In Proceedings of the 29th ACM International Conference on Multimedia (ACMMM), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.303, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[29] Claude E Shannon. A mathematical theory of communication. Bell System Technical Journal, 1948. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[30] Shubham Shit, Raphael Koner, Benedikt Wittmann, Johannes Paetzold, Ivan Ezhov, Haoyang Li, Jindong Pan, Sina Sharifzadeh, Georgios Kaissis, and Volker Tresp. Relationformer: A unified framework for image-to-graph generation. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.469 + ], + "angle": 0, + "content": "[31] Yudong Teng, Liang Wang, Zhenyu Li, and Gang Wu. Target adaptive context aggregation for video scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR), 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.525 + ], + "angle": 0, + "content": "[32] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.482, + 0.593 + ], + "angle": 0, + "content": "[33] Shizhe Wang, Linfeng Gao, Xinyi Lyu, Yulan Guo, Ping Zeng, and Jie Song. Dynamic scene graph generation via temporal prior inference. In Proceedings of the 30th ACM International Conference on Multimedia (ACMMM), 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[34] Zhongdao Wang, Hengshuang Zhao, Ya-Li Li, Shengjin Wang, Philip Torr, and Luca Bertinetto. Do different tracking tasks require different appearance models? Advances in Neural Information Processing Systems (NeurIPS), 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.706 + ], + "angle": 0, + "content": "[35] Ziyang Wang, Yi-Lin Sung, Feng Cheng, Gedas Bertasius, and Mohit Bansal. Unified coarse-to-fine alignment for video-text retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.762 + ], + "angle": 0, + "content": "[36] Danfei Xu, Yuke Zhu, Christopher B. Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.763, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[37] Jingkang Yang, Jun Cen, Wenxuan Peng, Fangzhou Liu, Shuai amd Hong, Xiangtai Li, Kaiyang Zhou, Qifeng Chen, and Ziwei Liu. 4d panoptic scene graph generation. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Jingkang Yang, Wenxuan Peng, Xiangtai Li, Zujin Guo, Liangyu Chen, Bo Li, Zheng Ma, Kaiyang Zhou, Wayne Zhang, Chen Change Loy, et al. Panoptic video scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[39] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.906, + 0.218 + ], + "angle": 0, + "content": "[40] Jiaxian Zhang, Kuan-Jui Shih, Ahmed Elgammal, Alexander Tao, and Bryan Catanzaro. Graphical contrastive losses for scene graph parsing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.906, + 0.289 + ], + "angle": 0, + "content": "[41] Cheng Zheng, Linfeng Gao, Xinyi Lyu, Ping Zeng, Abdulmotaleb El Saddik, and Heng Tao Shen. Dual-branch hybrid learning network for unbiased scene graph generation. IEEE Transactions on Circuits and Systems for Video Technology (TCSVT), 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19056" + } + ] +] \ No newline at end of file diff --git a/2025/Unbiased Video Scene Graph Generation via Visual and Semantic Dual Debiasing/168f707b-a7fa-4e7c-93c1-aac7eef770eb_origin.pdf b/2025/Unbiased Video Scene Graph Generation via Visual and Semantic Dual Debiasing/168f707b-a7fa-4e7c-93c1-aac7eef770eb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..38ac2e146c6e5e7ec5c7fd5807b81c11753f446e --- /dev/null +++ b/2025/Unbiased Video Scene Graph Generation via Visual and Semantic Dual Debiasing/168f707b-a7fa-4e7c-93c1-aac7eef770eb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:560af0ecc9eaec9f944396507e9522e11406e1b7d4c81da60debb56675e31454 +size 4298535 diff --git a/2025/Unbiased Video Scene Graph Generation via Visual and Semantic Dual Debiasing/full.md b/2025/Unbiased Video Scene Graph Generation via Visual and Semantic Dual Debiasing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b1f2e469d24841727fba738359faaadd13dc8d01 --- /dev/null +++ b/2025/Unbiased Video Scene Graph Generation via Visual and Semantic Dual Debiasing/full.md @@ -0,0 +1,378 @@ +# Unbiased Video Scene Graph Generation via Visual and Semantic Dual Debiasing + +Yanjun $\mathrm{Li}^{1\dagger}$ , Zhaoyang $\mathrm{Li}^1$ , Honghui Chen $^2$ , Lizhi $\mathrm{Xu}^1$ + +1University of Science and Technology of China + +$^{2}$ College of Physics and Information Engineering, Fuzhou University + +ballinli@mail.ustc.edu.cn, lizhaoyang@mail.ustc.edu.cn, chh5840996@gmail.com + +xulizhi@mail.ustc.edu.cn † Corresponding Author + +# Abstract + +Video Scene Graph Generation (VidSGG) aims to capture dynamic relationships among entities by sequentially analyzing video frames and integrating visual and semantic information. However, VidSGG is challenged by significant biases that skew predictions. To mitigate these biases, we propose a Visual and Semantic Awareness (VISA) framework for unbiased VidSGG. VISA addresses visual bias through memory-enhanced temporal integration that enhances object representations and concurrently reduces semantic bias by iteratively integrating object features with comprehensive semantic information derived from triplet relationships. This visual-semantics dual debiasing approach results in more unbiased representations of complex scene dynamics. Extensive experiments demonstrate the effectiveness of our method, where VISA outperforms existing unbiased VidSGG approaches by a substantial margin (e.g., +13.1% improvement in mR@20 and mR@50 for the SGCLS task under Semi Constraint). + +# 1. Introduction + +Scene Graph Generation (SGG) [22] constructs structured scene representations by identifying objects and their relationships as triplets (). This structure enables various downstream tasks [7, 10, 14, 27], highlighting SGG as a pivotal area in computer vision. Initially applied to static images (ImgSGG) [36], SGG has been extended to video data (VidSGG) [3] to model dynamic interactions over time. However, VidSGG faces challenges due to biases favoring overrepresented classes, leading to the underrepresentation of rare classes. + +Recent methods have adopted attention-based strategies to address class imbalance and mitigate classification biases [23, 25]. However, these approaches, primarily relying on learnable attention layers, often amplify biases towards + +![](images/7a95470c773dfec5b061732ac1e65d7855e271b9276a9c57f5294d0d65768004.jpg) + +![](images/3b0f00d18e9a3436f02d76a5e55c3f4c5ba3995031bcd04ea410b8416891640f.jpg) + +![](images/200af16ffb35782c7aea1ac3b49c1b8f67073dc9584fe99db26228d743c6ca52.jpg) + +![](images/937af58526537528cb6ef03ffefb600550d29734e4175c11ff69e9f661e835d1.jpg) +Figure 1. Framework Comparison. We identify visual and semantic biases in VidSGG. Figures (a) and (c) show that prior methods rely on weak visual features and limited context, leading to inaccurate attention maps and skewed predicate distributions. In (b) and (d), our memory-enhanced visual debiasing and iterative semantic fusion mitigate both biases, resulting in stronger feature representations and more balanced predicate outcomes. + +high-frequency classes [12]. To overcome this limitation, alternative methods have been proposed [12, 18] aiming to reduce spurious correlations and better account for label correlations. Although these methods alleviate bias to some extent, they often fall short of achieving the desired performance. A key oversight of these approaches is their heavy reliance on debiasing techniques, which neglect the + +intrinsic visual-semantic nature of scene graphs. + +Scene graphs inherently integrate both visual and semantic information, allowing biases to be categorized into visual biases and semantic biases. Visual biases arise from low visual quality in videos, such as blurring and occlusion [23], which degrade object feature representations. For instance, as shown in Fig. 1(a), objects may be partially visible or obscured over time, hindering models from capturing robust visual features for classification. Additionally, as illustrated in Fig. 1(c), the VidSGG task faces semantic biases. Previous methods [5, 23] rely solely on visual features to guide the relation generator for predicate prediction. This dependence on single-modal information leads to biased assumptions, similar to humans inferring relationships without sufficient context, thereby favoring more frequent predicate classes [21]. + +To address scene graph biases, we categorize them into visual and semantic types, guiding our proposed framework, VISA (VIual and Semantic Awareness), for unbiased VidSGG, as illustrated in Fig. 1 (b) and (d). VISA mitigates visual bias through a Memory-Enhanced Temporal Integrator that enriches the representations of entities in the current frame with historical representations from previous frames. It employs a relation generator to construct initial scene graphs and a hierarchical semantic extractor to iteratively refine semantic information. Our hierarchical fusion strategy integrates multimodal visual and semantic contexts to address semantic biases. Specifically, the visual debiasing method reduces visual feature variance, while the semantic debiasing method increases the Kullback-Leibler (KL) divergence. This dual-debiasing approach effectively mitigates visual biases caused by poor visual feature quality and semantic biases arising from insufficient contextual information, while also addressing model bias due to data skew. Extensive experiments demonstrate the effectiveness of VISA, achieving superior performance in unbiased VidSGG compared to previous methods. Our main contributions are summarized as follows: + +- We identify and explicitly define the visual and semantic biases affecting VidSGG, highlighting their impact on model performance. +- We propose VISA, a debiasing framework for VidSGG that simultaneously addresses visual and semantic biases through memory-enhanced temporal integration and hierarchical semantic extraction. +- We validate the effectiveness of our approach through extensive experiments and theoretical analyses, demonstrating state-of-the-art performance in unbiased VidSGG. + +# 2. Related Work + +# 2.1. Scene Graph Generation + +Scene Graph Generation (SGG) approaches primarily focus on static image-based SGG (ImgSGG) and video-based + +SGG (VidSGG). For ImgSGG, sequential encoders like LSTM [39] and attention mechanisms [32] capture global context, with attention methods [6, 16, 30] advancing standalone entity and predicate proposals (e.g., SGTR [16]). + +VidSGG extends SGG to dynamic contexts, requiring intra-frame relationships [11]. Recent methods improve temporal coherence [20] and efficiency by leveraging adaptive structures [31] and social context modeling [2]. + +Despite their success, existing methods ignore unbiasedness. This results in biased SGG with inaccurate and misleading relationships in scenes. Consequently, such biases make models less reliable and less generalizable to real-world applications. + +# 2.2. Unbiased Scene Graph Generation + +Unlike conventional SGG, unbiased SGG aims to reduce biases toward frequent objects and relationships. Key approaches include knowledge distillation [15] (e.g., LS-KD, which addresses multi-predicate challenges) and dual-branch architectures [41] (e.g., DHL, which balances attention between head and tail classes). + +For unbiased VidSGG, Transformer-based models [5] and Gaussian Mixture Models [23] have been explored to capture video dynamics better and reduce bias in visual relations. Iterative approaches [28] use conditional variables to improve video relation detection but lack a hierarchical strategy that integrates both visual and semantic representations, as in our method. + +Unlike existing approaches that merely apply traditional debiasing strategies to VidSGG, our framework leverages the inherent visual-semantic nature of scene graphs. This enables a solution that is both experimentally validated and theoretically robust, effectively addressing the fundamental challenges in achieving unbiased VidSGG. + +# 3. Method + +# 3.1. Overview + +The VISA framework, depicted in Fig. 2, comprises Memory Guided Sequence Modeling (MGSM), and the Iterative Relation Generator (IRG). It begins with an off-the-shelf object detector to identify objects in video frames, $\mathbf{b}_i^t$ represents the object anchors of the $i$ -th entity in the $t$ -th frame. In bounding-box approaches (e.g., VidSGG), $\mathbf{b}_i^t \in \mathbb{R}^4$ denotes a box, while in panoptic segmentation, $\mathbf{b}_i^t \in \mathbb{R}^n$ represents the $n$ boundary points. RoI-Align operations [9] are then applied to obtain feature representations, denoted as $\mathbf{v}_i^t$ . The MGSM refines these features to produce subject and object representations, yielding enhanced features $\hat{\mathbf{v}}_i^t$ , while two separate FFNs generate distinct subject and object embeddings. The IRG leverages these visual representations, using the Hierarchical Semantic Extractor (HSE) to fuse visual and semantic information. Through iterative + +![](images/fde13c76350966c85ca971b89014b75f7f449b601256093d62f6106143b8073f.jpg) +Figure 2. Framework of VISA. Given an input video, VISA extracts object features using an off-the-shelf detector. Subsequently, the Memory Guided Sequence Modeling leverages these features to develop robust object representations. Following this, the Iterative Relation Generator employs the object features to predict relationships between subject-object pairs. It is important to note that the Hierarchical Semantic Extractor is not engaged during the initial phase of scene graph generation. + +processing, it deduces unbiased semantic relations and generates relation predictions. + +# 3.2. Memory Guided Sequence Modeling + +We observe that as video duration increases, the Transformer tends to generate insufficient visual representations, particularly for small objects prone to blurring and occlusion (see Fig. 1 (a)). This insufficiency leads to high variance in feature representations due to the instability of visual information, causing the model to favor frequent entities and overlook less common ones. + +Our theoretical analysis reveals that this instability introduces noise into the object representations, modeled as: + +$$ +\mathbf {v} _ {i} ^ {t} = \mathbf {v} _ {i} + \boldsymbol {\epsilon} _ {i} ^ {t}, \tag {1} +$$ + +where $\epsilon_i^t$ denotes zero-mean Gaussian noise with covariance $\pmb{\Sigma}$ , i.e., $\mathbb{E}[\epsilon_i^t] = 0$ and $\mathrm{Cov}[\epsilon_i^t] = \pmb{\Sigma}$ . We assume these noise terms are independent across different time steps and objects. The observed features from the $t$ -th frame are used directly as the object's representation: + +$$ +\hat {\mathbf {v}} _ {i} ^ {t} = \mathbf {v} _ {i} ^ {t}. \tag {2} +$$ + +Consequently, the expectation and variance of the visual representation are computed as follows: + +$$ +\mathbb {E} \left[ \hat {\mathbf {v}} _ {i} ^ {t} \right] = \mathbf {v} _ {i}, \quad \operatorname {V a r} \left[ \hat {\mathbf {v}} _ {i} ^ {t} \right] = \boldsymbol {\Sigma}. \tag {3} +$$ + +While the single-frame estimation is unbiased, the noise introduced due to visual instability leads to high variance in the feature representations. This high variance destabilizes the feature representations and impairs subsequent relational predictions. + +![](images/28d8dd22ed2cbb3e84b95763a3621ddf683af7821cde69e4cd992ccf8c461a6c.jpg) +Figure 3. Structure of Memory Guided Sequence Modeling. + +Understanding that the high variance stems from the instability of visual information motivated us to design the Memory Guided Sequence Modeling (MGSM) module (illustrated in Fig. 3). The MGSM leverages temporal visual information from preceding frames to enhance individual feature representations of frames, effectively reducing the variance caused by insufficient visual representations. + +Specifically, we compute adaptive weights $\mathbf{W}_i^t$ based on the current frame's features: + +$$ +\mathbf {W} _ {i} ^ {t} = \sigma (\mathrm {M L P} (\mathbf {v} _ {i} ^ {t})), \tag {4} +$$ + +where $\sigma$ is the sigmoid function, and MLP denotes a multi-layer perceptron. Using these weights, we obtain the weighted feature representation: + +$$ +\mathbf {v} _ {i} ^ {t *} = \mathbf {W} _ {i} ^ {t} \odot \mathbf {v} _ {i} ^ {t} \| \mathbf {v} _ {i} ^ {t - 1}, \tag {5} +$$ + +where $\odot$ denotes element-wise multiplication, and $\parallel$ indicates feature concatenation. + +To accumulate visual information and refine the current + +representations, we introduce an update mechanism: + +$$ +\mathbf {M} _ {i} ^ {t + 1} = (1 - \lambda) \mathbf {M} _ {i} ^ {t} + \lambda \mathbf {v} _ {i} ^ {t}. \tag {6} +$$ + +As a result, the expectation and variance of the memory representation $\mathbf{M}_i^t$ stabilize over time: + +$$ +\mathbb {E} \left[ \mathbf {M} _ {i} ^ {t} \right] = \mathbf {v} _ {i}, \operatorname {V a r} \left[ \mathbf {M} _ {i} ^ {t} \right] = \frac {\lambda \boldsymbol {\Sigma}}{2 - \lambda} \approx \frac {\lambda \boldsymbol {\Sigma}}{2}, \text {f o r s m a l l} \lambda . \tag {7} +$$ + +This analysis shows that a smaller $\lambda$ reduces variance, enhancing the robustness of feature estimation. However, $\lambda$ cannot be too small, as excessively small values slow the adaptation to new information and may introduce bias. We provide detailed proofs of the variance calculation and the associated trade-off in the supplementary materials. + +The final enhanced feature representation is obtained through a dual attention layer: + +$$ +\hat {\mathbf {v}} _ {i} ^ {t} = \operatorname {A t t e n t i o n} \left(\mathbf {Q} = \mathbf {v} _ {i} ^ {t *}, \mathbf {K} = \mathbf {M} _ {i} ^ {t}, \mathbf {V} = \mathbf {v} _ {i} ^ {t}\right). \tag {8} +$$ + +Finally, the final enhanced feature representation is processed by the Transformer Encoder, as in [23], to generate the subject and object representations. By adaptively enhancing features and leveraging the reduced variance from the memory representation, the MGSM addresses the high variance caused by insufficient visual representations, thereby mitigating visual bias. + +Our experimental results (Fig. 6) confirm our theoretical insights by demonstrating that MGSM enhances visual representations. This enhancement reduces the high variance caused by insufficient visual information, which is crucial for addressing visual bias in VidSGG. + +# 3.3. Iterative Relation Generator + +Existing methods for unbiased SGG predominantly predict predicates directly from visual representations [5, 23]. However, this approach is akin to humans attempting to deduce relationships solely based on visual cues without sufficient contextual information, which easily introduces semantic bias [21]. Specifically, when lacking context, models tend to over-rely on the imbalanced and biased prior distributions present in the training data, leading to biased relationship predictions. + +To alleviate semantic bias, we start by analyzing the problem from an information-theoretic perspective. Semantic bias occurs when a model lacks sufficient contextual information, leading to high uncertainty in predicting relationships. This uncertainty causes the model to over-rely on biased prior distributions from the training data. We use entropy $H$ to quantify this uncertainty [29]. We consider the conditional entropy $H(r_{ij} \mid v_i, v_j, S)$ , where $v_i$ and $v_j$ are the feature representations of the subject and object, respectively. Based on the properties of conditional entropy [29], the additional context provided by $S$ decreases the uncertainty in predicting relationships: + +$$ +H \left(r _ {i j} \mid v _ {i}, v _ {j}, S\right) \leq H \left(r _ {i j} \mid v _ {i}, v _ {j}\right). \tag {9} +$$ + +This reduction in conditional entropy means the model is less uncertain and relies less on biased priors. By incorporating context $S$ , the likelihood term $P(v_{i}, v_{j}, S \mid r_{ij})$ in Bayes' theorem becomes more informative, effectively reducing the relative influence of the prior $P(r_{ij})$ . According to Bayes' theorem: + +$$ +P \left(r _ {i j} \mid v _ {i}, v _ {j}, S\right) = \frac {P \left(v _ {i} , v _ {j} , S \mid r _ {i j}\right) \cdot P \left(r _ {i j}\right)}{P \left(v _ {i} , v _ {j} , S\right)}. \tag {10} +$$ + +To quantify the change in reliance on the prior, we consider the Kullback-Leibler (KL) divergence between the posterior $P(r_{ij} \mid v_i, v_j, S)$ and the prior $P(r_{ij})$ : + +$$ +\begin{array}{l} D _ {\mathrm {K L}} \left(P \left(r _ {i j} \mid v _ {i}, v _ {j}, S\right) \| P \left(r _ {i j}\right)\right) \\ = \sum_ {r _ {i j}} P (r _ {i j} \mid v _ {i}, v _ {j}, S) \log \frac {P (r _ {i j} \mid v _ {i} , v _ {j} , S)}{P (r _ {i j})} \\ = H \left(r _ {i j} \mid v _ {i}, v _ {j}, P \left(r _ {i j}\right)\right) - H \left(r _ {i j} \mid v _ {i}, v _ {j}, S\right). \tag {11} \\ \end{array} +$$ + +As the conditional entropy $H(r_{ij} \mid v_i, v_j, S)$ decreases due to the additional context, the KL divergence increases. This indicates that the posterior distribution diverges from the biased prior, reflecting a reduced reliance on the prior and effectively mitigating semantic bias. + +This theoretical insight motivates our approach to incorporate contextual information into the model. Based on this, we propose the Iterative Relation Generator (IRG), an iterative semantic debiasing component within our framework. The IRG iteratively introduces contextual information to refine relationship predictions, thereby reducing uncertainty and dependence on biased priors. + +In the first iteration, the IRG infers semantic information from each subject-object pair to generate semantic embeddings. For each subject-object pair $(j,i)$ at frame $t$ , we construct the composite objects features $\mathbf{p}_{j,i}^{t}$ : + +$$ +\mathbf {p} _ {j, i} ^ {t} = f _ {v} \left(\mathbf {v} _ {j} ^ {t}\right) \| f _ {v} \left(\mathbf {v} _ {i} ^ {t}\right) \| \left(f _ {u} \left(\mathbf {u} _ {j i} ^ {t}\right) + f _ {\text {b o x}} \left(b _ {j} ^ {t}, b _ {i} ^ {t}\right)\right) \| \mathbf {s} _ {j} ^ {t} \| \mathbf {s} _ {i} ^ {t}, \tag {12} +$$ + +where $\mathbf{v}_j^t$ and $\mathbf{v}_i^t$ are the visual feature representations of the subject and object. The tensor $\mathbf{u}_{ji}^{t}$ is the union box feature map extracted via RoIAlign [9], and $\mathbf{s}_j^t$ and $\mathbf{s}_i^t$ are the semantic embeddings obtained from GloVe [24]. The functions $f_{v}, f_{u}$ , and $f_{\mathrm{box}}$ are feature transformation functions. + +We then perform initial relationship prediction, obtaining predicate categories such as attention, spatial, and contacting. We select the semantic embeddings with the highest probabilities from these distributions to form integrated triplet embeddings $\mathbf{C}_{pre,(j,i)}^t$ : + +$$ +\mathbf {C} _ {p r e, (j, i)} ^ {t} = \left[ \begin{array}{l} f _ {v} \left(\mathbf {v} _ {j} ^ {t}\right) \| f _ {a, (j, i)} ^ {t} \| f _ {v} \left(\mathbf {v} _ {i} ^ {t}\right) \\ f _ {v} \left(\mathbf {v} _ {j} ^ {t}\right) \| f _ {s, (j, i)} ^ {t} \| f _ {v} \left(\mathbf {v} _ {i} ^ {t}\right) \\ f _ {v} \left(\mathbf {v} _ {j} ^ {t}\right) \| f _ {c, (j, i)} ^ {t} \| f _ {v} \left(\mathbf {v} _ {i} ^ {t}\right) \end{array} \right], \tag {13} +$$ + +![](images/66fe30543f6d0373c4cce5d852ce2fc5fbcb7cd07591580a913fa30b880826d3.jpg) +Figure 4. Structure of Hierarchical Semantics Extractor. + +where $f_{a,(j,i)}^t, f_{s,(j,i)}^t$ , and $f_{c,(j,i)}^t$ are the semantic embeddings corresponding to the respective predicates. + +To further exploit hierarchical contextual information, we design the Hierarchical Semantics Extractor (HSE) (as shown in Fig. 4). Inspired by the multi-hierarchical feature fusion method [35], the HSE combines composite object features with integrated triplets to obtain fine-grained visual and semantic representations. We decompose the composite objects features $\mathbf{p}_{j,i}^{t}$ into fine-grained subject and object representations as follows: + +$$ +\begin{array}{l} \mathbf {S} _ {\mathbf {p} _ {j, i} ^ {t}} ^ {t} = f _ {v} (\mathbf {v} _ {j} ^ {t}) \| \left(f _ {u} (\mathbf {u} _ {j} ^ {t}) + f _ {\text {b o x}} \left(b _ {j} ^ {t}, b _ {i} ^ {t}\right)\right) \| \mathbf {s} _ {j} ^ {t}, \\ \mathbf {G} _ {i} ^ {t} = f _ {v} (t ^ {t}) \| \left(f _ {v} (t ^ {t}) + f _ {\text {b o x}} (t ^ {t}, t ^ {t})\right) \| \end{array} \tag {14} +$$ + +$$ +\mathbf {O} _ {\mathbf {p} _ {j, i} ^ {t}} ^ {t} = f _ {v} \left(\mathbf {v} _ {i} ^ {t}\right) \| \left(f _ {u} \left(\mathbf {u} _ {j i} ^ {t}\right) + f _ {\text {b o x}} \left(b _ {j} ^ {t}, b _ {i} ^ {t}\right)\right) \| \mathbf {s} _ {i} ^ {t}. +$$ + +Similarly, we decompose the integrated triplet embeddings $\mathbf{C}_{pre,(j,i)}^t$ : + +$$ +\begin{array}{l} \mathbf {S} _ {\mathbf {C} _ {(j, i)} ^ {t}} ^ {t} = \left\{f _ {v} \left(\mathbf {v} _ {j} ^ {t}\right) \| f _ {k} ^ {t} \mid k \in \{a, s, c \} \right\}, \\ \text {C} _ {i} ^ {t} = \left\{f _ {v} ^ {t + 1} \mid f _ {k} ^ {t + 1} \mid k \in \{a, c \} \right\}. \end{array} \tag {15} +$$ + +$$ +\mathbf {O} _ {\mathbf {C} _ {(j, i)} ^ {t}} ^ {t} = \left\{f _ {k} ^ {t} \| f _ {v} \left(\mathbf {v} _ {i} ^ {t}\right) \mid k \in \{a, s, c \} \right\}. \tag {15} +$$ + +These fine-grained representations are concatenated with downsampled embeddings from the previous cross-attention layer via a stride-two convolutional layer, serving as inputs to the next cross-attention layer. This hierarchical structure enables the model to capture multi-level contextual information, effectively reducing uncertainty and reliance on biased priors. The HSE is active for iteration counts $N \geq 1$ . + +After $N$ iterations of the IRG, the final inter-object predicates are generated by the Spatial Encoder and Temporal Decoder modules adopted from [5]. By iteratively leveraging contextual information, our method effectively mitigates semantic bias. Experimental results (as shown in Fig. 7) validate that incorporating contextual information reduces the model's reliance on biased priors and improves relationship predictions. + +# 3.4. Training and Testing + +Training. In the training phase, VISA synthesizes object features from MGSM for preliminary scene graph construction. Upon reaching iteration $N \geq 1$ , it activates the HSE to generate a hierarchical visual-semantics fusion embedding. + +In Section 3.2, the memory representation $\mathbf{M}_i^t$ is initialized as a zero tensor, while the embeddings for the three relationships are randomly initialized as described in Section 3.3. We optimize the comprehensive loss function: + +$$ +L _ {\text {t o t a l}} = L _ {p} + L _ {e} + L _ {\text {c o n t r a}}. \tag {16} +$$ + +Here, $L_{p}$ and $L_{e}$ denote the predicate and entity classification losses, both computed via cross-entropy, and $L_{\mathrm{contra}}$ is the contrastive loss, all defined as in [23]. + +Testing. After training, the proposed MGSM produces object predictions through the visual debiasing component, while the IRG performs semantic debiasing. + +# 4. Experiments + +# 4.1. Dataset and Metrics + +Dataset. We conducted experiments on the Action Genome (AG) dataset [11], the largest benchmark for VidSGG. AG consists of 234, 253 annotated frames containing 476, 229 bounding boxes across 35 object categories. It provides 1, 715, 568 predicate instances spanning 26 distinct relationship classes. + +Metrics and Evaluation Setup. We employed the standard metrics for unbiased VidSGG evaluation, mean Recall@K (mR@K), considering $K \in \{10, 20, 50\}$ . Following established protocols [5, 11, 23], we engaged in three SGG tasks, PREDCLS, SGCLS and SGDET. Evaluation was conducted across three settings, With Constraint, Semi Constraint, and No Constraint. Due to page constraints, further details are provided in the supplementary materials. + +# 4.2. Implementation details + +Following prior work [5, 19, 23], we adopted Faster R-CNN [26] with ResNet-101 [8] as the object detector, initially trained on the AG dataset. To ensure a fair comparison, we utilized the official implementations of these methods. For our MGSM module, we set the $\lambda$ parameter to 0.04 for the SGCLS task and 0.06 for the SGDET task. In the IRG module, we implemented a dual-procedure setup, enabling iterative relational inference with the number of iterations $N$ set to 1. Due to page constraints, further details are provided in the supplementary materials. + +# 4.3. Comparison to existing methods + +We began with a qualitative analysis, shown in Fig. 5, illustrating results under the Semi Constraint condition, which closely resembles real-world scenarios [5] where multiple predicates typically describe scenes. False positive relationships and object classifications were highlighted in purple blocks. The attention intensity demonstrated that our framework effectively enhanced object representations (e.g., pilow) through visual debiasing across frames from $t$ to $t + 1$ , + +![](images/842803436adecc37ed630b478be44fea1d1fd42eb435440b42fca2436dce67d5.jpg) +Figure 5. Qualitative results of our proposed method. Comparison with the open-sourced state-of-the-art model TEMPURA [23] under Semi Constraint, closely matching ground truth [5]. Visualizations show attention intensity and video scene graphs from Frame $t$ to $t + 1$ . False positive relationships and object classifications are highlighted in purple blocks. + +Table 1. Quantitative results comparisons with unbiased SGG methods under With, Semi, and No Constraints. The experiments are conducted on the dataset AG [11]. The best results are highlighted. + +
ConstraintMethodPREDCLSSGCLSSGDET
mR@10mR@20mR@50mR@10mR@20mR@50mR@10mR@20mR@50
WithRelDN [40]6.26.26.23.43.43.43.33.33.3
HCRD supervised [1]-------8.39.1
TRACE [31]15.215.215.28.98.98.98.28.28.2
ISGG [13]------19.722.9-
STTran [5]37.840.140.227.228.028.020.820.822.2
STTran-TPI [33]37.340.640.628.329.329.320.220.221.8
TEMPURA [23]42.946.346.334.035.235.222.622.623.7
FloCoDe [12]44.849.249.337.439.239.424.224.227.9
VISA (Ours)46.952.052.040.842.542.627.327.330.7
SemiTEMPURA [23]40.744.544.636.939.539.521.821.822.5
VISA (Ours)51.356.356.447.852.652.631.731.733.2
NoRelDN [40]31.263.175.518.636.942.67.518.833.7
TRACE [31]50.973.682.731.942.746.322.831.341.8
STTran [5]51.467.782.740.750.158.820.929.739.2
TEMPURA [23]61.585.198.048.361.166.424.733.943.7
FloCoDe [12]63.286.998.649.763.869.228.635.447.2
VISA (Ours)65.889.199.852.066.371.430.736.750.4
+ +compared with the TEMPURA [23]. Additionally, our approach inferred more accurate semantic predicates (e.g., sit on) within a single frame via contextual information. These visualizations confirm the correctness of partitioning scene graph biases into visual and semantic biases and demonstrate that our framework, VISA, significantly mitigates both in VidSGG. + +Fig. 6 and Fig. 7 further validated our theoretical analysis. Fig. 6 illustrated the variance of feature representations across training epochs, with shaded regions indicating the standard deviation. TEMPURA exhibited higher variance, suggesting less stable feature representations, + +whereas VISA demonstrated significantly reduced variance and tighter standard deviation regions. This reduction confirmed the effectiveness of the MGSM module in stabilizing feature representations. Additionally, Fig. 7 presented the confusion matrix comparing relationship predictions between TEMPURA and VISA. The reduced influence from prior dataset biases in VISA validated our approach to mitigate the semantics bias in relationship predictions using KL divergence principles. + +Our approach achieved state-of-the-art performance across all mR@K metrics, as shown in Table 1. VISA demonstrated improvements ranging from a minimum of + +![](images/17ce4d345dd13c711a4a394d8f3e7c823aece04b3a829a6448c5b9182e238132.jpg) +Figure 6. Comparison of TEMPURA and VISA visual feature representation variance. Visualization of visual feature representation variance across training epochs. The shaded regions represent the standard deviation. + +![](images/4d0dcbc7f963dffd0ec61a351219a352fe413a4adfc858209a5c6be36fe451d9.jpg) +Figure 8. Quantitative results of our proposed method. Compare our proposed method with open-sourced methods, TEMPURA [23] and STTran [5] on the mR@10 metric under With Constraint and No Constraint. + +![](images/89946733d5ef7ce47a859b7d5bff943903792db96ea2913dc34efb6222fc1384.jpg) +Figure 7. Qualitative results for relationship confusion matrix. Results for the subject-object pairs, person-sofa, and person-floor. $+1.2\%$ in mR@50 for the PREDCLS task under the No Constraint scenario to a maximum of $+13.1\%$ in both mR@20 and mR@50 for the SGCLS task under the Semi Constraint scenario. These results validate the correctness of partitioning scene graph biases into visual and semantic biases and demonstrate the effectiveness of VISA in mitigating them, thereby advancing unbiased VidSGG. + +Furthermore, as shown in Fig. 8, our VISA model surpassed open-source unbiased VidSGG methods in mR@10 across the three predicate class distributions (HEAD, BODY,TAIL) on AG under both With Constraint and No Constraint settings. Here, HEAD represented classes + +![](images/2ca2a2efcf5ab3e2815e7f5f5ee63fe739b1cc5172774ebb0b12232956884ac5.jpg) +(a) PredCLS + +![](images/5e95f0a704aa7b90ea1e7d8219afd17bf8b515ddcae46e1766c55dd35f44e119.jpg) +(b) SGCLS + +![](images/d41735c28576e1cd660f80b9badcba4ee86e8673b4416f03a5f17aae3bacfd92.jpg) +(c) SGDET + +![](images/dd42a5d5698a3fd623c939ce78765f3ea37db7f6b92d9134a7f08fd4bad917f7.jpg) +(a) PredCLS + +![](images/cf95ea171e1f498e27044bdf69fa41a7a0713fcbe8b8f3b1c3ba902a24c2f5d6.jpg) +(b) SGCLS + +![](images/2b662f01b3733e90d24769529b3161ab65569a8249c2478f10fe75cf1cda3136.jpg) +(c) SGDET + +with $\geq 100,000$ training samples, $BODY$ with fewer than 100,000, and $TAIL$ with fewer than 8,000. VISA consistently outperformed all competitors across mR@10 metrics, achieving a notable $+11.0\%$ improvement for the tail class in the SGDET task under No Constraint. This strong performance on tail classes highlighted the effectiveness of VISA's dual debiasing strategy. + +# 4.4. Ablation Studies + +Ablation study on VISA modules. Table 2 evaluates the effectiveness of VISA modules. Removing MGSM resulted in at least a $-2.2\%$ drop in mR@10 (SGCLS, Semi Constraint), confirming its capability to alleviate visual bias. Similarly, removing IRG caused at least a $-0.7\%$ decline in mR@10 (SGCLS, No Constraint), validating its role in reducing semantic bias. These results justify the division of biases into visual and semantic components and demonstrate VISA's effectiveness in mitigating both. Note that MGSM was not applied to the PREDCLS task, as it utilizes ground-truth visual inputs, rendering MGSM unnecessary. + +Influence of increasing iteration count $N$ . We analyzed the effect of iteration number $N$ on VISA's performance. Due to computational constraints, we conducted experiments with $N = 2$ and $N = 3$ , denoted as $\mathrm{VISA}_{N=2}$ and $\mathrm{VISA}_{N=3}$ , respectively. Results in Table 3 show that increasing $N$ consistently enhances the mR@K performance for unbiased VidSGG, though the gain from $N = 2$ to $N = 3$ is smaller compared to $N = 1$ to $N = 2$ . Additionally, training time approximately doubled when increasing from $N = 1$ to $N = 2$ . Further details are provided in the supplementary materials. + +The value of the $\lambda$ . Table 4 showed results for different values of $\lambda$ . VISA achieved optimal performance for SGCLS at $\lambda = 0.04$ , while the best results for SGDET occur + +Table 2. Ablation study of MGSM and IRG under With Constraint, Semi Constraint, No Constraint. + +
With ConstraintSemi ConstraintNo Constraint
PREDCLSSGCLSSGDETPREDCLSSGCLSSGDETPREDCLSSGCLSSGDETPREDCLSSGCLSSGDETPREDCLSSGCLSSGDET
mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20
VISA46.952.040.842.527.327.351.356.347.852.631.731.765.889.152.066.330.736.7
w/o MGSM46.952.035.738.218.824.651.356.345.649.324.127.865.889.149.062.027.934.7
w/o MGSM & IRG42.946.334.035.218.522.640.744.536.939.518.521.861.585.148.361.124.733.9
+ +Table 3. Influence of increasing iteration count $N$ + +
With ConstraintSemi ConstraintNo Constraint
PREDCLSSGCLSSGDETPREDCLSSGCLSSGDETPREDCLSSGCLSSGDET
mR@10mR@20mR@10mR@20mR@10mR@20mR@10mR@20mR@10mR@20mR@10mR@20mR@10mR@20mR@10mR@20mR@10mR@20
VISA46.952.040.842.527.327.351.356.347.852.631.731.765.889.152.066.330.736.7
VISA_N=247.953.041.743.425.528.752.556.948.253.627.531.966.090.853.167.031.837.2
VISA_N=348.153.342.244.025.828.952.857.148.553.927.832.166.291.253.567.232.037.5
+ +Table 4. Different values of $\lambda$ on SGCLS and SGDET,we summarize the results in mean-Recall@K,with best in bold. + +
λWith ConstraintSemi ConstraintNo Constraint
SGCLSSGDETSGCLSSGDETSGCLSSGDET
mR@10mR@20mR@50mR@10mR@20mR@50mR@10mR@20mR@50mR@10mR@20mR@50mR@10mR@20mR@50mR@10mR@20mR@50
0.00534.236.236.217.921.522.540.142.642.721.526.628.146.457.957.924.832.542.5
0.0136.037.437.418.622.223.240.744.644.723.728.630.047.259.865.225.434.043.0
0.0239.241.241.818.422.123.043.147.047.122.527.128.148.362.467.524.532.841.7
0.0440.842.542.619.823.324.347.852.652.624.529.531.152.066.371.426.733.641.7
0.0635.136.436.427.327.330.741.845.845.931.731.733.246.959.364.730.736.750.4
0.0834.335.635.618.321.622.540.243.743.723.327.828.947.159.665.724.732.740.6
0.133.234.434.417.520.322.139.142.542.522.425.427.246.259.664.322.131.439.2
+ +Table 5. Performance evaluation on the PVSG [38] Dataset + +
MethodPVSG Metrics
Stage-1Stage-2mR@20mR@50mR@100
VPS [4, 17]Transformer0.360.390.40
IPS+T [4, 34]Transformer1.751.862.03
MGSMIRG8.828.9410.51
+ +Table 6. Performance evaluation on the PSG4D-GTA and PSG4D-HOI [37] 4DPVSG Datasets. + +
MethodPSG4D-GTAPSG4D-HOI
mR@20mR@50mR@100mR@20mR@50mR@100
PSG4DFounder (P) [37]2.102.933.133.103.954.17
PSG4DFounder (R) [37]3.313.854.023.654.164.97
MGSM + IRG10.8311.5111.9810.9512.3212.87
+ +at $\lambda = 0.06$ . The differences in optimal $\lambda$ values reflect the distinct characteristics of these tasks. Specifically, SGCLS benefits from ground-truth data, enabling effective updates to the memory representations with a smaller $\lambda$ . In contrast, SGDET, which builds scene graphs from scratch, performs better with a slightly higher $\lambda$ . Notably, since the MSGM module was not applied to the PREDCLS task, the parameter $\lambda$ is also not involved in PREDCLS. + +Exploration on PVSG and 4DPVSG Datasets. We extended our framework to Panoptic Video Scene Graph Generation (PVSG) and 4D Panoptic Scene Graph Generation (4DPVSG) tasks [37, 38]. Unlike VidSGG, PVSG utilizes pixel-level masks for grounding nodes. Using the IPS+T+Transformer codebase, we integrated MGSM with a Memory-Enhanced Temporal Integrator and employed + +UniTrack [34] for temporal linking. For relation classification, IRG was adapted for iterative semantic debiasing. The 4DPVSG task further generalized PVSG to dynamic 4D contexts, employing PSG4DFormer variants with point cloud (P) and RGB-D (R) inputs. As shown in Table 5, our method significantly outperformed baselines, achieving minimum improvements of $+7.52\%$ mR@20 (PVSG) and $+7.30\%$ mR@20 (4DPVSG, PSG4D-HOI). These results confirm the broad applicability of our dual debiasing approach across VidSGG and indicate potential benefits for other generative and embodied intelligence applications. + +# 5. Conclusion + +In this work, we present VISA, an approach for unbiased video scene graph generation that addresses both visual and semantic biases inherent in scene graphs. VISA reduces visual bias by enriching current frame representations with historical context to decrease visual feature variance, and mitigates semantic bias by increasing the KL divergence between model predictions and biased priors. Our framework integrates hierarchical semantic contexts with visual features, enabling more unbiased scene graph generation. Extensive experimental and theoretical analyses demonstrate the effectiveness of VISA. Future work will explore extending our visual-semantic debiasing approach to other multimodal domains such as visual question answering and video captioning, as well as applications in embodied intelligence, including robotics perception and interactive agents. + +# Acknowledgments + +We gratefully acknowledge the three anonymous reviewers for their constructive feedback on our CVPR 2025 submission. This work was supported by personal funding from the corresponding author. + +# References + +[1] João Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 6 +[2] Shizhe Chen, Zhiyuan Shi, Pascal Mettes, and Cees G. M. Snoek. Social fabric: Tubelet compositions for video relation detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR), 2021. 2 +[3] Shizhe Chen, Jing Xiao, and Lin Chen. Video scene graph generation from single-frame weak supervision. In Proceedings of the Eleventh International Conference on Learning Representations (ICLR), 2022. 1 +[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 8 +[5] Yunhan Cong, Wentao Liao, Heiko Ackermann, Bodo Rosenhahn, and Michael Ying Yang. Spatial-temporal transformer for dynamic scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 4, 5, 6, 7 +[6] Yunhan Cong, Michael Ying Yang, and Bodo Rosenhahn. Reltr: Relation transformer for scene graph generation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023. 2 +[7] Linfeng Gao, Bin Wang, and Wenmin Wang. Image captioning with scene-graph based semantic concepts. In Proceedings of the International Conference on Machine Learning and Computing, 2018. 1 +[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5 +[9] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 2, 4 +[10] Marcel Hildebrandt, Haoyang Li, Raphael Koner, Volker Tresp, and Stephan Gunnemann. Scene graph reasoning for visual question answering. arXiv preprint arXiv:2007.01072, 2020. 1 +[11] Jiading Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as compositions of spatiotemporal scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 5, 6 +[12] Anant Khandelwal. Flocode: Unbiased dynamic scene graph generation with temporal consistency and correlation debias + +ing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2024. 1, 6 +[13] Shreyank Khandelwal and Leonid Sigal. Iterative scene graph generation. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 6 +[14] Hengshuang Li, Xin Li, Behnam Karimi, Jie Chen, and Ming Sun. Joint learning of object graph and relation graph for visual question answering. In Proceedings of the IEEE International Conference on Multimedia and Expo (ICME), 2022. 1 +[15] Lifeng Li, Jing Xiao, Huan Shi, Wei Wang, Jing Shao, Alan Liu, Yueting Yang, and Lin Chen. Label semantic knowledge distillation for unbiased scene graph generation. IEEE Transactions on Circuits and Systems for Video Technology (TCSVT), 2023. 2 +[16] Ruofei Li, Shuhui Zhang, and Xiaochun He. Sgtr: End-to-end scene graph generation with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[17] Xiangtai Li, Wenwei Zhang, Jiangmiao Pang, Kai Chen, Guangliang Cheng, Yunhai Tong, and Chen Change Loy. Video k-net: A simple, strong, and unified baseline for video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 8 +[18] Yicong Li, Xun Yang, Xindi Shang, and Tat-Seng Chua. Interventional video relation detection. In Proceedings of the 29th ACM International Conference on Multimedia (ACMMM), 2021. 1 +[19] Yifei Li, Xinyi Yang, and Chenglin Xu. Dynamic scene graph generation via anticipatory pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 5 +[20] Chang Liu, Yining Jin, Kaixuan Xu, Guohao Gong, and Yuesheng Mu. Beyond short-term snippet: Video relation detection with spatio-temporal global context. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[21] Yang Liu, Guanbin Li, and Liang Lin. Cross-modal causal relational reasoning for event-level visual question answering. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023. 2, 4 +[22] Cewu Lu, Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Visual relationship detection with language priors. In Proceedings of the European Conference on Computer Vision (ECCV), 2016. 1 +[23] Sourya Nag, Kyusong Min, Swami Tripathi, and Amit K. Roy-Chowdhury. Unbiased scene graph generation in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 4, 5, 6, 7 +[24] Jeffrey Pennington, Richard Socher, and Christopher D. Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2014. 4 +[25] Tao Pu, Tianshui Chen, Hefeng Wu, Yongyi Lu, and Liang Lin. Spatial-temporal knowledge-embedded transformer for + +video scene graph generation. IEEE Transactions on Image Processing (TIP), 2023. 1 +[26] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2017. 5 +[27] Scott Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D. Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the Fourth Workshop on Vision and Language, 2015. 1 +[28] Xun Shang, Yicheng Li, Jing Xiao, Weiqing Ji, and Tat-Seng Chua. Video visual relation detection via iterative inference. In Proceedings of the 29th ACM International Conference on Multimedia (ACMMM), 2021. 2 +[29] Claude E Shannon. A mathematical theory of communication. Bell System Technical Journal, 1948. 4 +[30] Shubham Shit, Raphael Koner, Benedikt Wittmann, Johannes Paetzold, Ivan Ezhov, Haoyang Li, Jindong Pan, Sina Sharifzadeh, Georgios Kaissis, and Volker Tresp. Relationformer: A unified framework for image-to-graph generation. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2 +[31] Yudong Teng, Liang Wang, Zhenyu Li, and Gang Wu. Target adaptive context aggregation for video scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR), 2021. 2, 6 +[32] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 2 +[33] Shizhe Wang, Linfeng Gao, Xinyi Lyu, Yulan Guo, Ping Zeng, and Jie Song. Dynamic scene graph generation via temporal prior inference. In Proceedings of the 30th ACM International Conference on Multimedia (ACMMM), 2022. 6 +[34] Zhongdao Wang, Hengshuang Zhao, Ya-Li Li, Shengjin Wang, Philip Torr, and Luca Bertinetto. Do different tracking tasks require different appearance models? Advances in Neural Information Processing Systems (NeurIPS), 2021. 8 +[35] Ziyang Wang, Yi-Lin Sung, Feng Cheng, Gedas Bertasius, and Mohit Bansal. Unified coarse-to-fine alignment for video-text retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 5 +[36] Danfei Xu, Yuke Zhu, Christopher B. Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1 +[37] Jingkang Yang, Jun Cen, Wenxuan Peng, Fangzhou Liu, Shuai amd Hong, Xiangtai Li, Kaiyang Zhou, Qifeng Chen, and Ziwei Liu. 4d panoptic scene graph generation. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 8 +[38] Jingkang Yang, Wenxuan Peng, Xiangtai Li, Zujin Guo, Liangyu Chen, Bo Li, Zheng Ma, Kaiyang Zhou, Wayne Zhang, Chen Change Loy, et al. Panoptic video scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 8 + +[39] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2 +[40] Jiaxian Zhang, Kuan-Jui Shih, Ahmed Elgammal, Alexander Tao, and Bryan Catanzaro. Graphical contrastive losses for scene graph parsing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6 +[41] Cheng Zheng, Linfeng Gao, Xinyi Lyu, Ping Zeng, Abdulmotaleb El Saddik, and Heng Tao Shen. Dual-branch hybrid learning network for unbiased scene graph generation. IEEE Transactions on Circuits and Systems for Video Technology (TCSVT), 2023. 2 \ No newline at end of file diff --git a/2025/Unbiased Video Scene Graph Generation via Visual and Semantic Dual Debiasing/images.zip b/2025/Unbiased Video Scene Graph Generation via Visual and Semantic Dual Debiasing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c9ab2e8348d9e43df40f97477a35e3d2bef853a3 --- /dev/null +++ b/2025/Unbiased Video Scene Graph Generation via Visual and Semantic Dual Debiasing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:18c192c425c9dad82f2183976637108cdbfc9ab743e6a6041c22ec4eba9f365b +size 880660 diff --git a/2025/Unbiased Video Scene Graph Generation via Visual and Semantic Dual Debiasing/layout.json b/2025/Unbiased Video Scene Graph Generation via Visual and Semantic Dual Debiasing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d6d6a0da8a07cd25b38836678ae41078c0028bf6 --- /dev/null +++ b/2025/Unbiased Video Scene Graph Generation via Visual and Semantic Dual Debiasing/layout.json @@ -0,0 +1,9379 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 86, + 102, + 524, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 102, + 524, + 139 + ], + "spans": [ + { + "bbox": [ + 86, + 102, + 524, + 139 + ], + "type": "text", + "content": "Unbiased Video Scene Graph Generation via Visual and Semantic Dual Debiasing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 170, + 160, + 436, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 160, + 436, + 175 + ], + "spans": [ + { + "bbox": [ + 170, + 160, + 436, + 175 + ], + "type": "text", + "content": "Yanjun " + }, + { + "bbox": [ + 170, + 160, + 436, + 175 + ], + "type": "inline_equation", + "content": "\\mathrm{Li}^{1\\dagger}" + }, + { + "bbox": [ + 170, + 160, + 436, + 175 + ], + "type": "text", + "content": ", Zhaoyang " + }, + { + "bbox": [ + 170, + 160, + 436, + 175 + ], + "type": "inline_equation", + "content": "\\mathrm{Li}^1" + }, + { + "bbox": [ + 170, + 160, + 436, + 175 + ], + "type": "text", + "content": ", Honghui Chen" + }, + { + "bbox": [ + 170, + 160, + 436, + 175 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 170, + 160, + 436, + 175 + ], + "type": "text", + "content": ", Lizhi " + }, + { + "bbox": [ + 170, + 160, + 436, + 175 + ], + "type": "inline_equation", + "content": "\\mathrm{Xu}^1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 187, + 176, + 423, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 176, + 423, + 190 + ], + "spans": [ + { + "bbox": [ + 187, + 176, + 423, + 190 + ], + "type": "text", + "content": "1University of Science and Technology of China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 190, + 471, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 190, + 471, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 190, + 471, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 190, + 471, + 205 + ], + "type": "text", + "content": "College of Physics and Information Engineering, Fuzhou University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 97, + 206, + 511, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 206, + 511, + 218 + ], + "spans": [ + { + "bbox": [ + 97, + 206, + 511, + 218 + ], + "type": "text", + "content": "ballinli@mail.ustc.edu.cn, lizhaoyang@mail.ustc.edu.cn, chh5840996@gmail.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 187, + 220, + 422, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 220, + 422, + 232 + ], + "spans": [ + { + "bbox": [ + 187, + 220, + 422, + 232 + ], + "type": "text", + "content": "xulizhi@mail.ustc.edu.cn † Corresponding Author" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 259, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 259, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 259, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 285, + 296, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 296, + 501 + ], + "type": "text", + "content": "Video Scene Graph Generation (VidSGG) aims to capture dynamic relationships among entities by sequentially analyzing video frames and integrating visual and semantic information. However, VidSGG is challenged by significant biases that skew predictions. To mitigate these biases, we propose a Visual and Semantic Awareness (VISA) framework for unbiased VidSGG. VISA addresses visual bias through memory-enhanced temporal integration that enhances object representations and concurrently reduces semantic bias by iteratively integrating object features with comprehensive semantic information derived from triplet relationships. This visual-semantics dual debiasing approach results in more unbiased representations of complex scene dynamics. Extensive experiments demonstrate the effectiveness of our method, where VISA outperforms existing unbiased VidSGG approaches by a substantial margin (e.g., +13.1% improvement in mR@20 and mR@50 for the SGCLS task under Semi Constraint)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 512, + 135, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 135, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 135, + 525 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 533, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 663 + ], + "type": "text", + "content": "Scene Graph Generation (SGG) [22] constructs structured scene representations by identifying objects and their relationships as triplets (). This structure enables various downstream tasks [7, 10, 14, 27], highlighting SGG as a pivotal area in computer vision. Initially applied to static images (ImgSGG) [36], SGG has been extended to video data (VidSGG) [3] to model dynamic interactions over time. However, VidSGG faces challenges due to biases favoring overrepresented classes, leading to the underrepresentation of rare classes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Recent methods have adopted attention-based strategies to address class imbalance and mitigate classification biases [23, 25]. However, these approaches, primarily relying on learnable attention layers, often amplify biases towards" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 258, + 554, + 335 + ], + "blocks": [ + { + "bbox": [ + 317, + 258, + 554, + 335 + ], + "lines": [ + { + "bbox": [ + 317, + 258, + 554, + 335 + ], + "spans": [ + { + "bbox": [ + 317, + 258, + 554, + 335 + ], + "type": "image", + "image_path": "7a95470c773dfec5b061732ac1e65d7855e271b9276a9c57f5294d0d65768004.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 316, + 337, + 555, + 381 + ], + "blocks": [ + { + "bbox": [ + 316, + 337, + 555, + 381 + ], + "lines": [ + { + "bbox": [ + 316, + 337, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 337, + 555, + 381 + ], + "type": "image", + "image_path": "3b0f00d18e9a3436f02d76a5e55c3f4c5ba3995031bcd04ea410b8416891640f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 317, + 385, + 555, + 446 + ], + "blocks": [ + { + "bbox": [ + 317, + 385, + 555, + 446 + ], + "lines": [ + { + "bbox": [ + 317, + 385, + 555, + 446 + ], + "spans": [ + { + "bbox": [ + 317, + 385, + 555, + 446 + ], + "type": "image", + "image_path": "200af16ffb35782c7aea1ac3b49c1b8f67073dc9584fe99db26228d743c6ca52.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 317, + 446, + 554, + 529 + ], + "blocks": [ + { + "bbox": [ + 317, + 446, + 554, + 529 + ], + "lines": [ + { + "bbox": [ + 317, + 446, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 317, + 446, + 554, + 529 + ], + "type": "image", + "image_path": "937af58526537528cb6ef03ffefb600550d29734e4175c11ff69e9f661e835d1.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 536, + 555, + 614 + ], + "lines": [ + { + "bbox": [ + 313, + 536, + 555, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 555, + 614 + ], + "type": "text", + "content": "Figure 1. Framework Comparison. We identify visual and semantic biases in VidSGG. Figures (a) and (c) show that prior methods rely on weak visual features and limited context, leading to inaccurate attention maps and skewed predicate distributions. In (b) and (d), our memory-enhanced visual debiasing and iterative semantic fusion mitigate both biases, resulting in stronger feature representations and more balanced predicate outcomes." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "high-frequency classes [12]. To overcome this limitation, alternative methods have been proposed [12, 18] aiming to reduce spurious correlations and better account for label correlations. Although these methods alleviate bias to some extent, they often fall short of achieving the desired performance. A key oversight of these approaches is their heavy reliance on debiasing techniques, which neglect the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19047" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 251, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 251, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 251, + 84 + ], + "type": "text", + "content": "intrinsic visual-semantic nature of scene graphs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 85, + 294, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 85, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 54, + 85, + 294, + 263 + ], + "type": "text", + "content": "Scene graphs inherently integrate both visual and semantic information, allowing biases to be categorized into visual biases and semantic biases. Visual biases arise from low visual quality in videos, such as blurring and occlusion [23], which degrade object feature representations. For instance, as shown in Fig. 1(a), objects may be partially visible or obscured over time, hindering models from capturing robust visual features for classification. Additionally, as illustrated in Fig. 1(c), the VidSGG task faces semantic biases. Previous methods [5, 23] rely solely on visual features to guide the relation generator for predicate prediction. This dependence on single-modal information leads to biased assumptions, similar to humans inferring relationships without sufficient context, thereby favoring more frequent predicate classes [21]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 264, + 296, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 525 + ], + "type": "text", + "content": "To address scene graph biases, we categorize them into visual and semantic types, guiding our proposed framework, VISA (VIual and Semantic Awareness), for unbiased VidSGG, as illustrated in Fig. 1 (b) and (d). VISA mitigates visual bias through a Memory-Enhanced Temporal Integrator that enriches the representations of entities in the current frame with historical representations from previous frames. It employs a relation generator to construct initial scene graphs and a hierarchical semantic extractor to iteratively refine semantic information. Our hierarchical fusion strategy integrates multimodal visual and semantic contexts to address semantic biases. Specifically, the visual debiasing method reduces visual feature variance, while the semantic debiasing method increases the Kullback-Leibler (KL) divergence. This dual-debiasing approach effectively mitigates visual biases caused by poor visual feature quality and semantic biases arising from insufficient contextual information, while also addressing model bias due to data skew. Extensive experiments demonstrate the effectiveness of VISA, achieving superior performance in unbiased VidSGG compared to previous methods. Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 528, + 295, + 647 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 528, + 295, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 295, + 563 + ], + "type": "text", + "content": "- We identify and explicitly define the visual and semantic biases affecting VidSGG, highlighting their impact on model performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 563, + 295, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 295, + 610 + ], + "type": "text", + "content": "- We propose VISA, a debiasing framework for VidSGG that simultaneously addresses visual and semantic biases through memory-enhanced temporal integration and hierarchical semantic extraction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 611, + 295, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 295, + 647 + ], + "type": "text", + "content": "- We validate the effectiveness of our approach through extensive experiments and theoretical analyses, demonstrating state-of-the-art performance in unbiased VidSGG." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 652, + 142, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 142, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 142, + 664 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 672, + 194, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 194, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 194, + 684 + ], + "type": "text", + "content": "2.1. Scene Graph Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Scene Graph Generation (SGG) approaches primarily focus on static image-based SGG (ImgSGG) and video-based" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "SGG (VidSGG). For ImgSGG, sequential encoders like LSTM [39] and attention mechanisms [32] capture global context, with attention methods [6, 16, 30] advancing standalone entity and predicate proposals (e.g., SGTR [16])." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 121, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 168 + ], + "type": "text", + "content": "VidSGG extends SGG to dynamic contexts, requiring intra-frame relationships [11]. Recent methods improve temporal coherence [20] and efficiency by leveraging adaptive structures [31] and social context modeling [2]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 170, + 553, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 170, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 553, + 229 + ], + "type": "text", + "content": "Despite their success, existing methods ignore unbiasedness. This results in biased SGG with inaccurate and misleading relationships in scenes. Consequently, such biases make models less reliable and less generalizable to real-world applications." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 239, + 499, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 499, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 499, + 251 + ], + "type": "text", + "content": "2.2. Unbiased Scene Graph Generation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 257, + 553, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 553, + 328 + ], + "type": "text", + "content": "Unlike conventional SGG, unbiased SGG aims to reduce biases toward frequent objects and relationships. Key approaches include knowledge distillation [15] (e.g., LS-KD, which addresses multi-predicate challenges) and dual-branch architectures [41] (e.g., DHL, which balances attention between head and tail classes)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 329, + 553, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 553, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 553, + 412 + ], + "type": "text", + "content": "For unbiased VidSGG, Transformer-based models [5] and Gaussian Mixture Models [23] have been explored to capture video dynamics better and reduce bias in visual relations. Iterative approaches [28] use conditional variables to improve video relation detection but lack a hierarchical strategy that integrates both visual and semantic representations, as in our method." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 414, + 553, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 553, + 485 + ], + "type": "text", + "content": "Unlike existing approaches that merely apply traditional debiasing strategies to VidSGG, our framework leverages the inherent visual-semantic nature of scene graphs. This enables a solution that is both experimentally validated and theoretically robust, effectively addressing the fundamental challenges in achieving unbiased VidSGG." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 497, + 369, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 369, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 369, + 509 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 516, + 381, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 381, + 527 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 381, + 527 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "The VISA framework, depicted in Fig. 2, comprises Memory Guided Sequence Modeling (MGSM), and the Iterative Relation Generator (IRG). It begins with an off-the-shelf object detector to identify objects in video frames, " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_i^t" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " represents the object anchors of the " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "-th entity in the " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "-th frame. In bounding-box approaches (e.g., VidSGG), " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_i^t \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " denotes a box, while in panoptic segmentation, " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_i^t \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " boundary points. RoI-Align operations [9] are then applied to obtain feature representations, denoted as " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i^t" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": ". The MGSM refines these features to produce subject and object representations, yielding enhanced features " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{v}}_i^t" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": ", while two separate FFNs generate distinct subject and object embeddings. The IRG leverages these visual representations, using the Hierarchical Semantic Extractor (HSE) to fuse visual and semantic information. Through iterative" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19048" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 558, + 273 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 558, + 273 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 558, + 273 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 558, + 273 + ], + "type": "image", + "image_path": "fde13c76350966c85ca971b89014b75f7f449b601256093d62f6106143b8073f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 280, + 555, + 326 + ], + "lines": [ + { + "bbox": [ + 54, + 280, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 54, + 280, + 555, + 326 + ], + "type": "text", + "content": "Figure 2. Framework of VISA. Given an input video, VISA extracts object features using an off-the-shelf detector. Subsequently, the Memory Guided Sequence Modeling leverages these features to develop robust object representations. Following this, the Iterative Relation Generator employs the object features to predict relationships between subject-object pairs. It is important to note that the Hierarchical Semantic Extractor is not engaged during the initial phase of scene graph generation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 336, + 295, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 295, + 360 + ], + "type": "text", + "content": "processing, it deduces unbiased semantic relations and generates relation predictions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 366, + 250, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 250, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 250, + 380 + ], + "type": "text", + "content": "3.2. Memory Guided Sequence Modeling" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 384, + 296, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 467 + ], + "type": "text", + "content": "We observe that as video duration increases, the Transformer tends to generate insufficient visual representations, particularly for small objects prone to blurring and occlusion (see Fig. 1 (a)). This insufficiency leads to high variance in feature representations due to the instability of visual information, causing the model to favor frequent entities and overlook less common ones." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 468, + 296, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 492 + ], + "type": "text", + "content": "Our theoretical analysis reveals that this instability introduces noise into the object representations, modeled as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 145, + 498, + 294, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 498, + 294, + 513 + ], + "spans": [ + { + "bbox": [ + 145, + 498, + 294, + 513 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {i} ^ {t} = \\mathbf {v} _ {i} + \\boldsymbol {\\epsilon} _ {i} ^ {t}, \\tag {1}", + "image_path": "abb9eb4153175e376c61c9b46ed29bbd25aebdb5a054f9604d346acc5ee16d15.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 519, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 296, + 579 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 519, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\epsilon_i^t" + }, + { + "bbox": [ + 55, + 519, + 296, + 579 + ], + "type": "text", + "content": " denotes zero-mean Gaussian noise with covariance " + }, + { + "bbox": [ + 55, + 519, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}" + }, + { + "bbox": [ + 55, + 519, + 296, + 579 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 519, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[\\epsilon_i^t] = 0" + }, + { + "bbox": [ + 55, + 519, + 296, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 519, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\mathrm{Cov}[\\epsilon_i^t] = \\pmb{\\Sigma}" + }, + { + "bbox": [ + 55, + 519, + 296, + 579 + ], + "type": "text", + "content": ". We assume these noise terms are independent across different time steps and objects. The observed features from the " + }, + { + "bbox": [ + 55, + 519, + 296, + 579 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 519, + 296, + 579 + ], + "type": "text", + "content": "-th frame are used directly as the object's representation:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 156, + 586, + 294, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 586, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 156, + 586, + 294, + 601 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {v}} _ {i} ^ {t} = \\mathbf {v} _ {i} ^ {t}. \\tag {2}", + "image_path": "cf0e37807b00ca0068cd9c380e67bd947e0d8d52878efd4e14c46cdee7ab47f3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 607, + 295, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 607, + 295, + 631 + ], + "spans": [ + { + "bbox": [ + 55, + 607, + 295, + 631 + ], + "type": "text", + "content": "Consequently, the expectation and variance of the visual representation are computed as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 111, + 638, + 294, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 638, + 294, + 651 + ], + "spans": [ + { + "bbox": [ + 111, + 638, + 294, + 651 + ], + "type": "interline_equation", + "content": "\\mathbb {E} \\left[ \\hat {\\mathbf {v}} _ {i} ^ {t} \\right] = \\mathbf {v} _ {i}, \\quad \\operatorname {V a r} \\left[ \\hat {\\mathbf {v}} _ {i} ^ {t} \\right] = \\boldsymbol {\\Sigma}. \\tag {3}", + "image_path": "d70d13ebefd24ced591c5d4c4f15a4e07a85df4b712008b186a1139a29c3ecb1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "While the single-frame estimation is unbiased, the noise introduced due to visual instability leads to high variance in the feature representations. This high variance destabilizes the feature representations and impairs subsequent relational predictions." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 322, + 334, + 547, + 439 + ], + "blocks": [ + { + "bbox": [ + 322, + 334, + 547, + 439 + ], + "lines": [ + { + "bbox": [ + 322, + 334, + 547, + 439 + ], + "spans": [ + { + "bbox": [ + 322, + 334, + 547, + 439 + ], + "type": "image", + "image_path": "28d8dd22ed2cbb3e84b95763a3621ddf683af7821cde69e4cd992ccf8c461a6c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 319, + 446, + 548, + 459 + ], + "lines": [ + { + "bbox": [ + 319, + 446, + 548, + 459 + ], + "spans": [ + { + "bbox": [ + 319, + 446, + 548, + 459 + ], + "type": "text", + "content": "Figure 3. Structure of Memory Guided Sequence Modeling." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 469, + 553, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 553, + 552 + ], + "type": "text", + "content": "Understanding that the high variance stems from the instability of visual information motivated us to design the Memory Guided Sequence Modeling (MGSM) module (illustrated in Fig. 3). The MGSM leverages temporal visual information from preceding frames to enhance individual feature representations of frames, effectively reducing the variance caused by insufficient visual representations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 552, + 553, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 553, + 576 + ], + "type": "text", + "content": "Specifically, we compute adaptive weights " + }, + { + "bbox": [ + 313, + 552, + 553, + 576 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_i^t" + }, + { + "bbox": [ + 313, + 552, + 553, + 576 + ], + "type": "text", + "content": " based on the current frame's features:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 390, + 586, + 553, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 586, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 390, + 586, + 553, + 601 + ], + "type": "interline_equation", + "content": "\\mathbf {W} _ {i} ^ {t} = \\sigma (\\mathrm {M L P} (\\mathbf {v} _ {i} ^ {t})), \\tag {4}", + "image_path": "b936bc24ed2a83e65b51498f0d01963625499d9a6202d5199ae5683674f35350.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 609, + 555, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 555, + 645 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 609, + 555, + 645 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 609, + 555, + 645 + ], + "type": "text", + "content": " is the sigmoid function, and MLP denotes a multi-layer perceptron. Using these weights, we obtain the weighted feature representation:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 386, + 654, + 553, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 654, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 386, + 654, + 553, + 669 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {i} ^ {t *} = \\mathbf {W} _ {i} ^ {t} \\odot \\mathbf {v} _ {i} ^ {t} \\| \\mathbf {v} _ {i} ^ {t - 1}, \\tag {5}", + "image_path": "9322981865af19b428febb6f5677fc5a99ed4a638c5379f960e4f035827257b6.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 677, + 553, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 700 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 553, + 700 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 677, + 553, + 700 + ], + "type": "text", + "content": " denotes element-wise multiplication, and " + }, + { + "bbox": [ + 313, + 677, + 553, + 700 + ], + "type": "inline_equation", + "content": "\\parallel" + }, + { + "bbox": [ + 313, + 677, + 553, + 700 + ], + "type": "text", + "content": " indicates feature concatenation." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 326, + 701, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 712 + ], + "type": "text", + "content": "To accumulate visual information and refine the current" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19049" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 268, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 268, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 268, + 84 + ], + "type": "text", + "content": "representations, we introduce an update mechanism:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 90, + 295, + 105 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 90, + 295, + 105 + ], + "spans": [ + { + "bbox": [ + 116, + 90, + 295, + 105 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {i} ^ {t + 1} = (1 - \\lambda) \\mathbf {M} _ {i} ^ {t} + \\lambda \\mathbf {v} _ {i} ^ {t}. \\tag {6}", + "image_path": "4ceefb470a1209a15363ec7d368a507298d3f26778c7a3138f5fcf5813e69315.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 112, + 295, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 112, + 295, + 135 + ], + "spans": [ + { + "bbox": [ + 55, + 112, + 295, + 135 + ], + "type": "text", + "content": "As a result, the expectation and variance of the memory representation " + }, + { + "bbox": [ + 55, + 112, + 295, + 135 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i^t" + }, + { + "bbox": [ + 55, + 112, + 295, + 135 + ], + "type": "text", + "content": " stabilize over time:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 141, + 295, + 166 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 141, + 295, + 166 + ], + "spans": [ + { + "bbox": [ + 63, + 141, + 295, + 166 + ], + "type": "interline_equation", + "content": "\\mathbb {E} \\left[ \\mathbf {M} _ {i} ^ {t} \\right] = \\mathbf {v} _ {i}, \\operatorname {V a r} \\left[ \\mathbf {M} _ {i} ^ {t} \\right] = \\frac {\\lambda \\boldsymbol {\\Sigma}}{2 - \\lambda} \\approx \\frac {\\lambda \\boldsymbol {\\Sigma}}{2}, \\text {f o r s m a l l} \\lambda . \\tag {7}", + "image_path": "856ba6cf5100baa79b993787c66dfd2d612701b9e50aa65600b5c18e09eaf225.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 171, + 296, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 296, + 243 + ], + "type": "text", + "content": "This analysis shows that a smaller " + }, + { + "bbox": [ + 55, + 171, + 296, + 243 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 171, + 296, + 243 + ], + "type": "text", + "content": " reduces variance, enhancing the robustness of feature estimation. However, " + }, + { + "bbox": [ + 55, + 171, + 296, + 243 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 171, + 296, + 243 + ], + "type": "text", + "content": " cannot be too small, as excessively small values slow the adaptation to new information and may introduce bias. We provide detailed proofs of the variance calculation and the associated trade-off in the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 243, + 296, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 296, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 296, + 268 + ], + "type": "text", + "content": "The final enhanced feature representation is obtained through a dual attention layer:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 79, + 274, + 295, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 274, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 79, + 274, + 295, + 289 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {v}} _ {i} ^ {t} = \\operatorname {A t t e n t i o n} \\left(\\mathbf {Q} = \\mathbf {v} _ {i} ^ {t *}, \\mathbf {K} = \\mathbf {M} _ {i} ^ {t}, \\mathbf {V} = \\mathbf {v} _ {i} ^ {t}\\right). \\tag {8}", + "image_path": "9c456793db9b40f011d2a78e6d8128f2e391844d971b0f39e8c90c1e2e4cb648.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 295, + 296, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 296, + 378 + ], + "type": "text", + "content": "Finally, the final enhanced feature representation is processed by the Transformer Encoder, as in [23], to generate the subject and object representations. By adaptively enhancing features and leveraging the reduced variance from the memory representation, the MGSM addresses the high variance caused by insufficient visual representations, thereby mitigating visual bias." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 379, + 296, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 296, + 438 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 296, + 438 + ], + "type": "text", + "content": "Our experimental results (Fig. 6) confirm our theoretical insights by demonstrating that MGSM enhances visual representations. This enhancement reduces the high variance caused by insufficient visual information, which is crucial for addressing visual bias in VidSGG." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 445, + 212, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 212, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 212, + 456 + ], + "type": "text", + "content": "3.3. Iterative Relation Generator" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 462, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 569 + ], + "type": "text", + "content": "Existing methods for unbiased SGG predominantly predict predicates directly from visual representations [5, 23]. However, this approach is akin to humans attempting to deduce relationships solely based on visual cues without sufficient contextual information, which easily introduces semantic bias [21]. Specifically, when lacking context, models tend to over-rely on the imbalanced and biased prior distributions present in the training data, leading to biased relationship predictions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "To alleviate semantic bias, we start by analyzing the problem from an information-theoretic perspective. Semantic bias occurs when a model lacks sufficient contextual information, leading to high uncertainty in predicting relationships. This uncertainty causes the model to over-rely on biased prior distributions from the training data. We use entropy " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " to quantify this uncertainty [29]. We consider the conditional entropy " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "H(r_{ij} \\mid v_i, v_j, S)" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "v_j" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " are the feature representations of the subject and object, respectively. Based on the properties of conditional entropy [29], the additional context provided by " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " decreases the uncertainty in predicting relationships:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 361, + 72, + 555, + 86 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 72, + 555, + 86 + ], + "spans": [ + { + "bbox": [ + 361, + 72, + 555, + 86 + ], + "type": "interline_equation", + "content": "H \\left(r _ {i j} \\mid v _ {i}, v _ {j}, S\\right) \\leq H \\left(r _ {i j} \\mid v _ {i}, v _ {j}\\right). \\tag {9}", + "image_path": "8a8e4f545ea2bda0bf55da8f563bbb4c9acc32e172b342f816c145c10983b4a1.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 92, + 555, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 92, + 555, + 164 + ], + "spans": [ + { + "bbox": [ + 313, + 92, + 555, + 164 + ], + "type": "text", + "content": "This reduction in conditional entropy means the model is less uncertain and relies less on biased priors. By incorporating context " + }, + { + "bbox": [ + 313, + 92, + 555, + 164 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 92, + 555, + 164 + ], + "type": "text", + "content": ", the likelihood term " + }, + { + "bbox": [ + 313, + 92, + 555, + 164 + ], + "type": "inline_equation", + "content": "P(v_{i}, v_{j}, S \\mid r_{ij})" + }, + { + "bbox": [ + 313, + 92, + 555, + 164 + ], + "type": "text", + "content": " in Bayes' theorem becomes more informative, effectively reducing the relative influence of the prior " + }, + { + "bbox": [ + 313, + 92, + 555, + 164 + ], + "type": "inline_equation", + "content": "P(r_{ij})" + }, + { + "bbox": [ + 313, + 92, + 555, + 164 + ], + "type": "text", + "content": ". According to Bayes' theorem:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 329, + 170, + 555, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 170, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 329, + 170, + 555, + 198 + ], + "type": "interline_equation", + "content": "P \\left(r _ {i j} \\mid v _ {i}, v _ {j}, S\\right) = \\frac {P \\left(v _ {i} , v _ {j} , S \\mid r _ {i j}\\right) \\cdot P \\left(r _ {i j}\\right)}{P \\left(v _ {i} , v _ {j} , S\\right)}. \\tag {10}", + "image_path": "dab7c3f7de073a236c9a40e0090e3523eba58d58f66240a8a5cd23331d4bbb2b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 203, + 554, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 203, + 554, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 554, + 240 + ], + "type": "text", + "content": "To quantify the change in reliance on the prior, we consider the Kullback-Leibler (KL) divergence between the posterior " + }, + { + "bbox": [ + 313, + 203, + 554, + 240 + ], + "type": "inline_equation", + "content": "P(r_{ij} \\mid v_i, v_j, S)" + }, + { + "bbox": [ + 313, + 203, + 554, + 240 + ], + "type": "text", + "content": " and the prior " + }, + { + "bbox": [ + 313, + 203, + 554, + 240 + ], + "type": "inline_equation", + "content": "P(r_{ij})" + }, + { + "bbox": [ + 313, + 203, + 554, + 240 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 326, + 243, + 553, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 243, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 326, + 243, + 553, + 316 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} D _ {\\mathrm {K L}} \\left(P \\left(r _ {i j} \\mid v _ {i}, v _ {j}, S\\right) \\| P \\left(r _ {i j}\\right)\\right) \\\\ = \\sum_ {r _ {i j}} P (r _ {i j} \\mid v _ {i}, v _ {j}, S) \\log \\frac {P (r _ {i j} \\mid v _ {i} , v _ {j} , S)}{P (r _ {i j})} \\\\ = H \\left(r _ {i j} \\mid v _ {i}, v _ {j}, P \\left(r _ {i j}\\right)\\right) - H \\left(r _ {i j} \\mid v _ {i}, v _ {j}, S\\right). \\tag {11} \\\\ \\end{array}", + "image_path": "11000368f7a5be4ac50516850ff1d5091b0fd98b440b3ef217f949958ad6b9f0.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 319, + 554, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 554, + 379 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 554, + 379 + ], + "type": "text", + "content": "As the conditional entropy " + }, + { + "bbox": [ + 313, + 319, + 554, + 379 + ], + "type": "inline_equation", + "content": "H(r_{ij} \\mid v_i, v_j, S)" + }, + { + "bbox": [ + 313, + 319, + 554, + 379 + ], + "type": "text", + "content": " decreases due to the additional context, the KL divergence increases. This indicates that the posterior distribution diverges from the biased prior, reflecting a reduced reliance on the prior and effectively mitigating semantic bias." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 380, + 554, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 554, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 554, + 463 + ], + "type": "text", + "content": "This theoretical insight motivates our approach to incorporate contextual information into the model. Based on this, we propose the Iterative Relation Generator (IRG), an iterative semantic debiasing component within our framework. The IRG iteratively introduces contextual information to refine relationship predictions, thereby reducing uncertainty and dependence on biased priors." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 463, + 554, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 554, + 512 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 554, + 512 + ], + "type": "text", + "content": "In the first iteration, the IRG infers semantic information from each subject-object pair to generate semantic embeddings. For each subject-object pair " + }, + { + "bbox": [ + 313, + 463, + 554, + 512 + ], + "type": "inline_equation", + "content": "(j,i)" + }, + { + "bbox": [ + 313, + 463, + 554, + 512 + ], + "type": "text", + "content": " at frame " + }, + { + "bbox": [ + 313, + 463, + 554, + 512 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 463, + 554, + 512 + ], + "type": "text", + "content": ", we construct the composite objects features " + }, + { + "bbox": [ + 313, + 463, + 554, + 512 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{j,i}^{t}" + }, + { + "bbox": [ + 313, + 463, + 554, + 512 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 514, + 553, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 514, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 320, + 514, + 553, + 540 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {j, i} ^ {t} = f _ {v} \\left(\\mathbf {v} _ {j} ^ {t}\\right) \\| f _ {v} \\left(\\mathbf {v} _ {i} ^ {t}\\right) \\| \\left(f _ {u} \\left(\\mathbf {u} _ {j i} ^ {t}\\right) + f _ {\\text {b o x}} \\left(b _ {j} ^ {t}, b _ {i} ^ {t}\\right)\\right) \\| \\mathbf {s} _ {j} ^ {t} \\| \\mathbf {s} _ {i} ^ {t}, \\tag {12}", + "image_path": "0ee069705d80bded0d20493160c852dbaf0d51575af1373161e5c1e4cd373107.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_j^t" + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i^t" + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "text", + "content": " are the visual feature representations of the subject and object. The tensor " + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_{ji}^{t}" + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "text", + "content": " is the union box feature map extracted via RoIAlign [9], and " + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_j^t" + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i^t" + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "text", + "content": " are the semantic embeddings obtained from GloVe [24]. The functions " + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "inline_equation", + "content": "f_{v}, f_{u}" + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{box}}" + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "text", + "content": " are feature transformation functions." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 605, + 554, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 668 + ], + "type": "text", + "content": "We then perform initial relationship prediction, obtaining predicate categories such as attention, spatial, and contacting. We select the semantic embeddings with the highest probabilities from these distributions to form integrated triplet embeddings " + }, + { + "bbox": [ + 313, + 605, + 554, + 668 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{pre,(j,i)}^t" + }, + { + "bbox": [ + 313, + 605, + 554, + 668 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 339, + 672, + 553, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 672, + 553, + 717 + ], + "spans": [ + { + "bbox": [ + 339, + 672, + 553, + 717 + ], + "type": "interline_equation", + "content": "\\mathbf {C} _ {p r e, (j, i)} ^ {t} = \\left[ \\begin{array}{l} f _ {v} \\left(\\mathbf {v} _ {j} ^ {t}\\right) \\| f _ {a, (j, i)} ^ {t} \\| f _ {v} \\left(\\mathbf {v} _ {i} ^ {t}\\right) \\\\ f _ {v} \\left(\\mathbf {v} _ {j} ^ {t}\\right) \\| f _ {s, (j, i)} ^ {t} \\| f _ {v} \\left(\\mathbf {v} _ {i} ^ {t}\\right) \\\\ f _ {v} \\left(\\mathbf {v} _ {j} ^ {t}\\right) \\| f _ {c, (j, i)} ^ {t} \\| f _ {v} \\left(\\mathbf {v} _ {i} ^ {t}\\right) \\end{array} \\right], \\tag {13}", + "image_path": "3d73b3bf2ee4419ec8bb35fcd3abf3b7ffcebafaac32167ada10118f34303b5f.jpg" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19050" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 298, + 178 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 298, + 178 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 298, + 178 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 298, + 178 + ], + "type": "image", + "image_path": "66fe30543f6d0373c4cce5d852ce2fc5fbcb7cd07591580a913fa30b880826d3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 184, + 285, + 196 + ], + "lines": [ + { + "bbox": [ + 66, + 184, + 285, + 196 + ], + "spans": [ + { + "bbox": [ + 66, + 184, + 285, + 196 + ], + "type": "text", + "content": "Figure 4. Structure of Hierarchical Semantics Extractor." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 206, + 295, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 206, + 295, + 230 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 295, + 230 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 206, + 295, + 230 + ], + "type": "inline_equation", + "content": "f_{a,(j,i)}^t, f_{s,(j,i)}^t" + }, + { + "bbox": [ + 55, + 206, + 295, + 230 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 206, + 295, + 230 + ], + "type": "inline_equation", + "content": "f_{c,(j,i)}^t" + }, + { + "bbox": [ + 55, + 206, + 295, + 230 + ], + "type": "text", + "content": " are the semantic embeddings corresponding to the respective predicates." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 230, + 296, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 230, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 296, + 326 + ], + "type": "text", + "content": "To further exploit hierarchical contextual information, we design the Hierarchical Semantics Extractor (HSE) (as shown in Fig. 4). Inspired by the multi-hierarchical feature fusion method [35], the HSE combines composite object features with integrated triplets to obtain fine-grained visual and semantic representations. We decompose the composite objects features " + }, + { + "bbox": [ + 55, + 230, + 296, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{j,i}^{t}" + }, + { + "bbox": [ + 55, + 230, + 296, + 326 + ], + "type": "text", + "content": " into fine-grained subject and object representations as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 73, + 332, + 296, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 332, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 73, + 332, + 296, + 356 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {S} _ {\\mathbf {p} _ {j, i} ^ {t}} ^ {t} = f _ {v} (\\mathbf {v} _ {j} ^ {t}) \\| \\left(f _ {u} (\\mathbf {u} _ {j} ^ {t}) + f _ {\\text {b o x}} \\left(b _ {j} ^ {t}, b _ {i} ^ {t}\\right)\\right) \\| \\mathbf {s} _ {j} ^ {t}, \\\\ \\mathbf {G} _ {i} ^ {t} = f _ {v} (t ^ {t}) \\| \\left(f _ {v} (t ^ {t}) + f _ {\\text {b o x}} (t ^ {t}, t ^ {t})\\right) \\| \\end{array} \\tag {14}", + "image_path": "949dfe5a9958a6474670a00d22e3750f009b33fe96e54d57b76b1d39473948dc.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 74, + 350, + 260, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 350, + 260, + 369 + ], + "spans": [ + { + "bbox": [ + 74, + 350, + 260, + 369 + ], + "type": "interline_equation", + "content": "\\mathbf {O} _ {\\mathbf {p} _ {j, i} ^ {t}} ^ {t} = f _ {v} \\left(\\mathbf {v} _ {i} ^ {t}\\right) \\| \\left(f _ {u} \\left(\\mathbf {u} _ {j i} ^ {t}\\right) + f _ {\\text {b o x}} \\left(b _ {j} ^ {t}, b _ {i} ^ {t}\\right)\\right) \\| \\mathbf {s} _ {i} ^ {t}.", + "image_path": "e63f42a975e5677961b675fd594a7f83981542f267083b01c443f8b5d188cb8e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 376, + 295, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 403 + ], + "type": "text", + "content": "Similarly, we decompose the integrated triplet embeddings " + }, + { + "bbox": [ + 55, + 376, + 295, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{pre,(j,i)}^t" + }, + { + "bbox": [ + 55, + 376, + 295, + 403 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 95, + 408, + 296, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 408, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 95, + 408, + 296, + 432 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {S} _ {\\mathbf {C} _ {(j, i)} ^ {t}} ^ {t} = \\left\\{f _ {v} \\left(\\mathbf {v} _ {j} ^ {t}\\right) \\| f _ {k} ^ {t} \\mid k \\in \\{a, s, c \\} \\right\\}, \\\\ \\text {C} _ {i} ^ {t} = \\left\\{f _ {v} ^ {t + 1} \\mid f _ {k} ^ {t + 1} \\mid k \\in \\{a, c \\} \\right\\}. \\end{array} \\tag {15}", + "image_path": "e02c07625799fe0ff49d39e8a8c8a509f18f07880d6871abb7eaa592af0d752c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 94, + 422, + 294, + 446 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 422, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 94, + 422, + 294, + 446 + ], + "type": "interline_equation", + "content": "\\mathbf {O} _ {\\mathbf {C} _ {(j, i)} ^ {t}} ^ {t} = \\left\\{f _ {k} ^ {t} \\| f _ {v} \\left(\\mathbf {v} _ {i} ^ {t}\\right) \\mid k \\in \\{a, s, c \\} \\right\\}. \\tag {15}", + "image_path": "a43019a5611b1ed7f28b36e3dcf2d3cb0484a917a85ceb8c5b78c719bf81d006.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 449, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 544 + ], + "type": "text", + "content": "These fine-grained representations are concatenated with downsampled embeddings from the previous cross-attention layer via a stride-two convolutional layer, serving as inputs to the next cross-attention layer. This hierarchical structure enables the model to capture multi-level contextual information, effectively reducing uncertainty and reliance on biased priors. The HSE is active for iteration counts " + }, + { + "bbox": [ + 55, + 449, + 296, + 544 + ], + "type": "inline_equation", + "content": "N \\geq 1" + }, + { + "bbox": [ + 55, + 449, + 296, + 544 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 545, + 296, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 641 + ], + "type": "text", + "content": "After " + }, + { + "bbox": [ + 55, + 545, + 296, + 641 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 545, + 296, + 641 + ], + "type": "text", + "content": " iterations of the IRG, the final inter-object predicates are generated by the Spatial Encoder and Temporal Decoder modules adopted from [5]. By iteratively leveraging contextual information, our method effectively mitigates semantic bias. Experimental results (as shown in Fig. 7) validate that incorporating contextual information reduces the model's reliance on biased priors and improves relationship predictions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 648, + 176, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 176, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 176, + 661 + ], + "type": "text", + "content": "3.4. Training and Testing" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Training. In the training phase, VISA synthesizes object features from MGSM for preliminary scene graph construction. Upon reaching iteration " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "N \\geq 1" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": ", it activates the HSE to generate a hierarchical visual-semantics fusion embedding." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "In Section 3.2, the memory representation " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_i^t" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " is initialized as a zero tensor, while the embeddings for the three relationships are randomly initialized as described in Section 3.3. We optimize the comprehensive loss function:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 377, + 129, + 555, + 143 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 129, + 555, + 143 + ], + "spans": [ + { + "bbox": [ + 377, + 129, + 555, + 143 + ], + "type": "interline_equation", + "content": "L _ {\\text {t o t a l}} = L _ {p} + L _ {e} + L _ {\\text {c o n t r a}}. \\tag {16}", + "image_path": "bddf4d7a6a5638dc602caac11927d43e9bb02cd883d96ff5e2d614bb53dd3348.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 152, + 554, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 554, + 187 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 554, + 187 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 152, + 554, + 187 + ], + "type": "inline_equation", + "content": "L_{p}" + }, + { + "bbox": [ + 313, + 152, + 554, + 187 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 152, + 554, + 187 + ], + "type": "inline_equation", + "content": "L_{e}" + }, + { + "bbox": [ + 313, + 152, + 554, + 187 + ], + "type": "text", + "content": " denote the predicate and entity classification losses, both computed via cross-entropy, and " + }, + { + "bbox": [ + 313, + 152, + 554, + 187 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{contra}}" + }, + { + "bbox": [ + 313, + 152, + 554, + 187 + ], + "type": "text", + "content": " is the contrastive loss, all defined as in [23]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 188, + 554, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 188, + 554, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 554, + 224 + ], + "type": "text", + "content": "Testing. After training, the proposed MGSM produces object predictions through the visual debiasing component, while the IRG performs semantic debiasing." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 236, + 395, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 395, + 249 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 395, + 249 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 256, + 432, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 432, + 268 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 432, + 268 + ], + "type": "text", + "content": "4.1. Dataset and Metrics" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 274, + 554, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 554, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 554, + 346 + ], + "type": "text", + "content": "Dataset. We conducted experiments on the Action Genome (AG) dataset [11], the largest benchmark for VidSGG. AG consists of 234, 253 annotated frames containing 476, 229 bounding boxes across 35 object categories. It provides 1, 715, 568 predicate instances spanning 26 distinct relationship classes." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 348, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 555, + 445 + ], + "type": "text", + "content": "Metrics and Evaluation Setup. We employed the standard metrics for unbiased VidSGG evaluation, mean Recall@K (mR@K), considering " + }, + { + "bbox": [ + 313, + 348, + 555, + 445 + ], + "type": "inline_equation", + "content": "K \\in \\{10, 20, 50\\}" + }, + { + "bbox": [ + 313, + 348, + 555, + 445 + ], + "type": "text", + "content": ". Following established protocols [5, 11, 23], we engaged in three SGG tasks, PREDCLS, SGCLS and SGDET. Evaluation was conducted across three settings, With Constraint, Semi Constraint, and No Constraint. Due to page constraints, further details are provided in the supplementary materials." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 453, + 446, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 446, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 446, + 466 + ], + "type": "text", + "content": "4.2. Implementation details" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 471, + 554, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 554, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 554, + 591 + ], + "type": "text", + "content": "Following prior work [5, 19, 23], we adopted Faster R-CNN [26] with ResNet-101 [8] as the object detector, initially trained on the AG dataset. To ensure a fair comparison, we utilized the official implementations of these methods. For our MGSM module, we set the " + }, + { + "bbox": [ + 313, + 471, + 554, + 591 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 471, + 554, + 591 + ], + "type": "text", + "content": " parameter to 0.04 for the SGCLS task and 0.06 for the SGDET task. In the IRG module, we implemented a dual-procedure setup, enabling iterative relational inference with the number of iterations " + }, + { + "bbox": [ + 313, + 471, + 554, + 591 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 471, + 554, + 591 + ], + "type": "text", + "content": " set to 1. Due to page constraints, further details are provided in the supplementary materials." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 600, + 488, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 488, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 488, + 613 + ], + "type": "text", + "content": "4.3. Comparison to existing methods" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": "We began with a qualitative analysis, shown in Fig. 5, illustrating results under the Semi Constraint condition, which closely resembles real-world scenarios [5] where multiple predicates typically describe scenes. False positive relationships and object classifications were highlighted in purple blocks. The attention intensity demonstrated that our framework effectively enhanced object representations (e.g., pilow) through visual debiasing across frames from " + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19051" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 58, + 531, + 281 + ], + "blocks": [ + { + "bbox": [ + 80, + 58, + 531, + 281 + ], + "lines": [ + { + "bbox": [ + 80, + 58, + 531, + 281 + ], + "spans": [ + { + "bbox": [ + 80, + 58, + 531, + 281 + ], + "type": "image", + "image_path": "842803436adecc37ed630b478be44fea1d1fd42eb435440b42fca2436dce67d5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 289, + 555, + 323 + ], + "lines": [ + { + "bbox": [ + 54, + 289, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 54, + 289, + 555, + 323 + ], + "type": "text", + "content": "Figure 5. Qualitative results of our proposed method. Comparison with the open-sourced state-of-the-art model TEMPURA [23] under Semi Constraint, closely matching ground truth [5]. Visualizations show attention intensity and video scene graphs from Frame " + }, + { + "bbox": [ + 54, + 289, + 555, + 323 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 289, + 555, + 323 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 289, + 555, + 323 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 54, + 289, + 555, + 323 + ], + "type": "text", + "content": ". False positive relationships and object classifications are highlighted in purple blocks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 119, + 359, + 491, + 541 + ], + "blocks": [ + { + "bbox": [ + 54, + 328, + 555, + 352 + ], + "lines": [ + { + "bbox": [ + 54, + 328, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 54, + 328, + 555, + 352 + ], + "type": "text", + "content": "Table 1. Quantitative results comparisons with unbiased SGG methods under With, Semi, and No Constraints. The experiments are conducted on the dataset AG [11]. The best results are highlighted." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 119, + 359, + 491, + 541 + ], + "lines": [ + { + "bbox": [ + 119, + 359, + 491, + 541 + ], + "spans": [ + { + "bbox": [ + 119, + 359, + 491, + 541 + ], + "type": "table", + "html": "
ConstraintMethodPREDCLSSGCLSSGDET
mR@10mR@20mR@50mR@10mR@20mR@50mR@10mR@20mR@50
WithRelDN [40]6.26.26.23.43.43.43.33.33.3
HCRD supervised [1]-------8.39.1
TRACE [31]15.215.215.28.98.98.98.28.28.2
ISGG [13]------19.722.9-
STTran [5]37.840.140.227.228.028.020.820.822.2
STTran-TPI [33]37.340.640.628.329.329.320.220.221.8
TEMPURA [23]42.946.346.334.035.235.222.622.623.7
FloCoDe [12]44.849.249.337.439.239.424.224.227.9
VISA (Ours)46.952.052.040.842.542.627.327.330.7
SemiTEMPURA [23]40.744.544.636.939.539.521.821.822.5
VISA (Ours)51.356.356.447.852.652.631.731.733.2
NoRelDN [40]31.263.175.518.636.942.67.518.833.7
TRACE [31]50.973.682.731.942.746.322.831.341.8
STTran [5]51.467.782.740.750.158.820.929.739.2
TEMPURA [23]61.585.198.048.361.166.424.733.943.7
FloCoDe [12]63.286.998.649.763.869.228.635.447.2
VISA (Ours)65.889.199.852.066.371.430.736.750.4
", + "image_path": "15881c717a6f834ae30982eed0951ca86ade7efa683fa2d0fe687f0e1b9924c8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 559, + 294, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 559, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 54, + 559, + 294, + 643 + ], + "type": "text", + "content": "compared with the TEMPURA [23]. Additionally, our approach inferred more accurate semantic predicates (e.g., sit on) within a single frame via contextual information. These visualizations confirm the correctness of partitioning scene graph biases into visual and semantic biases and demonstrate that our framework, VISA, significantly mitigates both in VidSGG." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 653, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 653, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 653, + 297, + 715 + ], + "type": "text", + "content": "Fig. 6 and Fig. 7 further validated our theoretical analysis. Fig. 6 illustrated the variance of feature representations across training epochs, with shaded regions indicating the standard deviation. TEMPURA exhibited higher variance, suggesting less stable feature representations," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 559, + 555, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 555, + 668 + ], + "type": "text", + "content": "whereas VISA demonstrated significantly reduced variance and tighter standard deviation regions. This reduction confirmed the effectiveness of the MGSM module in stabilizing feature representations. Additionally, Fig. 7 presented the confusion matrix comparing relationship predictions between TEMPURA and VISA. The reduced influence from prior dataset biases in VISA validated our approach to mitigate the semantics bias in relationship predictions using KL divergence principles." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Our approach achieved state-of-the-art performance across all mR@K metrics, as shown in Table 1. VISA demonstrated improvements ranging from a minimum of" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19052" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 69, + 265, + 245 + ], + "blocks": [ + { + "bbox": [ + 86, + 69, + 265, + 245 + ], + "lines": [ + { + "bbox": [ + 86, + 69, + 265, + 245 + ], + "spans": [ + { + "bbox": [ + 86, + 69, + 265, + 245 + ], + "type": "image", + "image_path": "17ce4d345dd13c711a4a394d8f3e7c823aece04b3a829a6448c5b9182e238132.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 248, + 298, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 248, + 298, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 298, + 293 + ], + "type": "text", + "content": "Figure 6. Comparison of TEMPURA and VISA visual feature representation variance. Visualization of visual feature representation variance across training epochs. The shaded regions represent the standard deviation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 299, + 299, + 411 + ], + "blocks": [ + { + "bbox": [ + 56, + 299, + 299, + 411 + ], + "lines": [ + { + "bbox": [ + 56, + 299, + 299, + 411 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 299, + 411 + ], + "type": "image", + "image_path": "4d0dcbc7f963dffd0ec61a351219a352fe413a4adfc858209a5c6be36fe451d9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 249, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 313, + 249, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 555, + 293 + ], + "type": "text", + "content": "Figure 8. Quantitative results of our proposed method. Compare our proposed method with open-sourced methods, TEMPURA [23] and STTran [5] on the mR@10 metric under With Constraint and No Constraint." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 411, + 294, + 536 + ], + "blocks": [ + { + "bbox": [ + 59, + 411, + 294, + 536 + ], + "lines": [ + { + "bbox": [ + 59, + 411, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 59, + 411, + 294, + 536 + ], + "type": "image", + "image_path": "89946733d5ef7ce47a859b7d5bff943903792db96ea2913dc34efb6222fc1384.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "text", + "content": "Figure 7. Qualitative results for relationship confusion matrix. Results for the subject-object pairs, person-sofa, and person-floor. " + }, + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "inline_equation", + "content": "+1.2\\%" + }, + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "text", + "content": " in mR@50 for the PREDCLS task under the No Constraint scenario to a maximum of " + }, + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "inline_equation", + "content": "+13.1\\%" + }, + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "text", + "content": " in both mR@20 and mR@50 for the SGCLS task under the Semi Constraint scenario. These results validate the correctness of partitioning scene graph biases into visual and semantic biases and demonstrate the effectiveness of VISA in mitigating them, thereby advancing unbiased VidSGG." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Furthermore, as shown in Fig. 8, our VISA model surpassed open-source unbiased VidSGG methods in mR@10 across the three predicate class distributions (HEAD, BODY,TAIL) on AG under both With Constraint and No Constraint settings. Here, HEAD represented classes" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 315, + 69, + 403, + 145 + ], + "blocks": [ + { + "bbox": [ + 315, + 69, + 403, + 145 + ], + "lines": [ + { + "bbox": [ + 315, + 69, + 403, + 145 + ], + "spans": [ + { + "bbox": [ + 315, + 69, + 403, + 145 + ], + "type": "image", + "image_path": "2ca2a2efcf5ab3e2815e7f5f5ee63fe739b1cc5172774ebb0b12232956884ac5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 147, + 383, + 155 + ], + "lines": [ + { + "bbox": [ + 356, + 147, + 383, + 155 + ], + "spans": [ + { + "bbox": [ + 356, + 147, + 383, + 155 + ], + "type": "text", + "content": "(a) PredCLS" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 409, + 70, + 481, + 145 + ], + "blocks": [ + { + "bbox": [ + 409, + 70, + 481, + 145 + ], + "lines": [ + { + "bbox": [ + 409, + 70, + 481, + 145 + ], + "spans": [ + { + "bbox": [ + 409, + 70, + 481, + 145 + ], + "type": "image", + "image_path": "5e95f0a704aa7b90ea1e7d8219afd17bf8b515ddcae46e1766c55dd35f44e119.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 146, + 460, + 155 + ], + "lines": [ + { + "bbox": [ + 435, + 146, + 460, + 155 + ], + "spans": [ + { + "bbox": [ + 435, + 146, + 460, + 155 + ], + "type": "text", + "content": "(b) SGCLS" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 485, + 70, + 555, + 145 + ], + "blocks": [ + { + "bbox": [ + 485, + 70, + 555, + 145 + ], + "lines": [ + { + "bbox": [ + 485, + 70, + 555, + 145 + ], + "spans": [ + { + "bbox": [ + 485, + 70, + 555, + 145 + ], + "type": "image", + "image_path": "d41735c28576e1cd660f80b9badcba4ee86e8673b4416f03a5f17aae3bacfd92.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 511, + 147, + 536, + 155 + ], + "lines": [ + { + "bbox": [ + 511, + 147, + 536, + 155 + ], + "spans": [ + { + "bbox": [ + 511, + 147, + 536, + 155 + ], + "type": "text", + "content": "(c) SGDET" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 314, + 158, + 403, + 233 + ], + "blocks": [ + { + "bbox": [ + 314, + 158, + 403, + 233 + ], + "lines": [ + { + "bbox": [ + 314, + 158, + 403, + 233 + ], + "spans": [ + { + "bbox": [ + 314, + 158, + 403, + 233 + ], + "type": "image", + "image_path": "dd42a5d5698a3fd623c939ce78765f3ea37db7f6b92d9134a7f08fd4bad917f7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 233, + 383, + 240 + ], + "lines": [ + { + "bbox": [ + 356, + 233, + 383, + 240 + ], + "spans": [ + { + "bbox": [ + 356, + 233, + 383, + 240 + ], + "type": "text", + "content": "(a) PredCLS" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 409, + 158, + 480, + 233 + ], + "blocks": [ + { + "bbox": [ + 409, + 158, + 480, + 233 + ], + "lines": [ + { + "bbox": [ + 409, + 158, + 480, + 233 + ], + "spans": [ + { + "bbox": [ + 409, + 158, + 480, + 233 + ], + "type": "image", + "image_path": "cf95ea171e1f498e27044bdf69fa41a7a0713fcbe8b8f3b1c3ba902a24c2f5d6.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 233, + 460, + 240 + ], + "lines": [ + { + "bbox": [ + 435, + 233, + 460, + 240 + ], + "spans": [ + { + "bbox": [ + 435, + 233, + 460, + 240 + ], + "type": "text", + "content": "(b) SGCLS" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 485, + 158, + 555, + 233 + ], + "blocks": [ + { + "bbox": [ + 485, + 158, + 555, + 233 + ], + "lines": [ + { + "bbox": [ + 485, + 158, + 555, + 233 + ], + "spans": [ + { + "bbox": [ + 485, + 158, + 555, + 233 + ], + "type": "image", + "image_path": "2b662f01b3733e90d24769529b3161ab65569a8249c2478f10fe75cf1cda3136.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 511, + 233, + 536, + 240 + ], + "lines": [ + { + "bbox": [ + 511, + 233, + 536, + 240 + ], + "spans": [ + { + "bbox": [ + 511, + 233, + 536, + 240 + ], + "type": "text", + "content": "(c) SGDET" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 304, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 389 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 313, + 304, + 555, + 389 + ], + "type": "inline_equation", + "content": "\\geq 100,000" + }, + { + "bbox": [ + 313, + 304, + 555, + 389 + ], + "type": "text", + "content": " training samples, " + }, + { + "bbox": [ + 313, + 304, + 555, + 389 + ], + "type": "inline_equation", + "content": "BODY" + }, + { + "bbox": [ + 313, + 304, + 555, + 389 + ], + "type": "text", + "content": " with fewer than 100,000, and " + }, + { + "bbox": [ + 313, + 304, + 555, + 389 + ], + "type": "inline_equation", + "content": "TAIL" + }, + { + "bbox": [ + 313, + 304, + 555, + 389 + ], + "type": "text", + "content": " with fewer than 8,000. VISA consistently outperformed all competitors across mR@10 metrics, achieving a notable " + }, + { + "bbox": [ + 313, + 304, + 555, + 389 + ], + "type": "inline_equation", + "content": "+11.0\\%" + }, + { + "bbox": [ + 313, + 304, + 555, + 389 + ], + "type": "text", + "content": " improvement for the tail class in the SGDET task under No Constraint. This strong performance on tail classes highlighted the effectiveness of VISA's dual debiasing strategy." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 396, + 416, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 416, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 416, + 408 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 414, + 555, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 555, + 545 + ], + "type": "text", + "content": "Ablation study on VISA modules. Table 2 evaluates the effectiveness of VISA modules. Removing MGSM resulted in at least a " + }, + { + "bbox": [ + 313, + 414, + 555, + 545 + ], + "type": "inline_equation", + "content": "-2.2\\%" + }, + { + "bbox": [ + 313, + 414, + 555, + 545 + ], + "type": "text", + "content": " drop in mR@10 (SGCLS, Semi Constraint), confirming its capability to alleviate visual bias. Similarly, removing IRG caused at least a " + }, + { + "bbox": [ + 313, + 414, + 555, + 545 + ], + "type": "inline_equation", + "content": "-0.7\\%" + }, + { + "bbox": [ + 313, + 414, + 555, + 545 + ], + "type": "text", + "content": " decline in mR@10 (SGCLS, No Constraint), validating its role in reducing semantic bias. These results justify the division of biases into visual and semantic components and demonstrate VISA's effectiveness in mitigating both. Note that MGSM was not applied to the PREDCLS task, as it utilizes ground-truth visual inputs, rendering MGSM unnecessary." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 545, + 556, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 677 + ], + "type": "text", + "content": "Influence of increasing iteration count " + }, + { + "bbox": [ + 313, + 545, + 556, + 677 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 545, + 556, + 677 + ], + "type": "text", + "content": ". We analyzed the effect of iteration number " + }, + { + "bbox": [ + 313, + 545, + 556, + 677 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 545, + 556, + 677 + ], + "type": "text", + "content": " on VISA's performance. Due to computational constraints, we conducted experiments with " + }, + { + "bbox": [ + 313, + 545, + 556, + 677 + ], + "type": "inline_equation", + "content": "N = 2" + }, + { + "bbox": [ + 313, + 545, + 556, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 545, + 556, + 677 + ], + "type": "inline_equation", + "content": "N = 3" + }, + { + "bbox": [ + 313, + 545, + 556, + 677 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 313, + 545, + 556, + 677 + ], + "type": "inline_equation", + "content": "\\mathrm{VISA}_{N=2}" + }, + { + "bbox": [ + 313, + 545, + 556, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 545, + 556, + 677 + ], + "type": "inline_equation", + "content": "\\mathrm{VISA}_{N=3}" + }, + { + "bbox": [ + 313, + 545, + 556, + 677 + ], + "type": "text", + "content": ", respectively. Results in Table 3 show that increasing " + }, + { + "bbox": [ + 313, + 545, + 556, + 677 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 545, + 556, + 677 + ], + "type": "text", + "content": " consistently enhances the mR@K performance for unbiased VidSGG, though the gain from " + }, + { + "bbox": [ + 313, + 545, + 556, + 677 + ], + "type": "inline_equation", + "content": "N = 2" + }, + { + "bbox": [ + 313, + 545, + 556, + 677 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 545, + 556, + 677 + ], + "type": "inline_equation", + "content": "N = 3" + }, + { + "bbox": [ + 313, + 545, + 556, + 677 + ], + "type": "text", + "content": " is smaller compared to " + }, + { + "bbox": [ + 313, + 545, + 556, + 677 + ], + "type": "inline_equation", + "content": "N = 1" + }, + { + "bbox": [ + 313, + 545, + 556, + 677 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 545, + 556, + 677 + ], + "type": "inline_equation", + "content": "N = 2" + }, + { + "bbox": [ + 313, + 545, + 556, + 677 + ], + "type": "text", + "content": ". Additionally, training time approximately doubled when increasing from " + }, + { + "bbox": [ + 313, + 545, + 556, + 677 + ], + "type": "inline_equation", + "content": "N = 1" + }, + { + "bbox": [ + 313, + 545, + 556, + 677 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 545, + 556, + 677 + ], + "type": "inline_equation", + "content": "N = 2" + }, + { + "bbox": [ + 313, + 545, + 556, + 677 + ], + "type": "text", + "content": ". Further details are provided in the supplementary materials." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "The value of the " + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": ". Table 4 showed results for different values of " + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": ". VISA achieved optimal performance for SGCLS at " + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\lambda = 0.04" + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": ", while the best results for SGDET occur" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19053" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 86, + 558, + 156 + ], + "blocks": [ + { + "bbox": [ + 123, + 71, + 485, + 82 + ], + "lines": [ + { + "bbox": [ + 123, + 71, + 485, + 82 + ], + "spans": [ + { + "bbox": [ + 123, + 71, + 485, + 82 + ], + "type": "text", + "content": "Table 2. Ablation study of MGSM and IRG under With Constraint, Semi Constraint, No Constraint." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 86, + 558, + 156 + ], + "lines": [ + { + "bbox": [ + 58, + 86, + 558, + 156 + ], + "spans": [ + { + "bbox": [ + 58, + 86, + 558, + 156 + ], + "type": "table", + "html": "
With ConstraintSemi ConstraintNo Constraint
PREDCLSSGCLSSGDETPREDCLSSGCLSSGDETPREDCLSSGCLSSGDETPREDCLSSGCLSSGDETPREDCLSSGCLSSGDET
mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20 mR@10 mR@20
VISA46.952.040.842.527.327.351.356.347.852.631.731.765.889.152.066.330.736.7
w/o MGSM46.952.035.738.218.824.651.356.345.649.324.127.865.889.149.062.027.934.7
w/o MGSM & IRG42.946.334.035.218.522.640.744.536.939.518.521.861.585.148.361.124.733.9
", + "image_path": "6ccfe805fa8f2658a024a9ca87aeffe2fbaac7097d6771cda944af795555f5aa.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 177, + 553, + 248 + ], + "blocks": [ + { + "bbox": [ + 214, + 163, + 394, + 173 + ], + "lines": [ + { + "bbox": [ + 214, + 163, + 394, + 173 + ], + "spans": [ + { + "bbox": [ + 214, + 163, + 394, + 173 + ], + "type": "text", + "content": "Table 3. Influence of increasing iteration count " + }, + { + "bbox": [ + 214, + 163, + 394, + 173 + ], + "type": "inline_equation", + "content": "N" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 177, + 553, + 248 + ], + "lines": [ + { + "bbox": [ + 58, + 177, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 58, + 177, + 553, + 248 + ], + "type": "table", + "html": "
With ConstraintSemi ConstraintNo Constraint
PREDCLSSGCLSSGDETPREDCLSSGCLSSGDETPREDCLSSGCLSSGDET
mR@10mR@20mR@10mR@20mR@10mR@20mR@10mR@20mR@10mR@20mR@10mR@20mR@10mR@20mR@10mR@20mR@10mR@20
VISA46.952.040.842.527.327.351.356.347.852.631.731.765.889.152.066.330.736.7
VISA_N=247.953.041.743.425.528.752.556.948.253.627.531.966.090.853.167.031.837.2
VISA_N=348.153.342.244.025.828.952.857.148.553.927.832.166.291.253.567.232.037.5
", + "image_path": "fdda9d2d94d159ee2f138eaedc57f0d0e0fb407cb99dbf07888fca7c4470fb2b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 58, + 270, + 553, + 367 + ], + "blocks": [ + { + "bbox": [ + 88, + 255, + 520, + 266 + ], + "lines": [ + { + "bbox": [ + 88, + 255, + 520, + 266 + ], + "spans": [ + { + "bbox": [ + 88, + 255, + 520, + 266 + ], + "type": "text", + "content": "Table 4. Different values of " + }, + { + "bbox": [ + 88, + 255, + 520, + 266 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 88, + 255, + 520, + 266 + ], + "type": "text", + "content": " on SGCLS and SGDET,we summarize the results in mean-Recall@K,with best in bold." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 270, + 553, + 367 + ], + "lines": [ + { + "bbox": [ + 58, + 270, + 553, + 367 + ], + "spans": [ + { + "bbox": [ + 58, + 270, + 553, + 367 + ], + "type": "table", + "html": "
λWith ConstraintSemi ConstraintNo Constraint
SGCLSSGDETSGCLSSGDETSGCLSSGDET
mR@10mR@20mR@50mR@10mR@20mR@50mR@10mR@20mR@50mR@10mR@20mR@50mR@10mR@20mR@50mR@10mR@20mR@50
0.00534.236.236.217.921.522.540.142.642.721.526.628.146.457.957.924.832.542.5
0.0136.037.437.418.622.223.240.744.644.723.728.630.047.259.865.225.434.043.0
0.0239.241.241.818.422.123.043.147.047.122.527.128.148.362.467.524.532.841.7
0.0440.842.542.619.823.324.347.852.652.624.529.531.152.066.371.426.733.641.7
0.0635.136.436.427.327.330.741.845.845.931.731.733.246.959.364.730.736.750.4
0.0834.335.635.618.321.622.540.243.743.723.327.828.947.159.665.724.732.740.6
0.133.234.434.417.520.322.139.142.542.522.425.427.246.259.664.322.131.439.2
", + "image_path": "a641d9423aef549e5095ef3f2855227dc0afd207a808719ed0d162449daf1170.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 58, + 391, + 294, + 444 + ], + "blocks": [ + { + "bbox": [ + 67, + 376, + 284, + 388 + ], + "lines": [ + { + "bbox": [ + 67, + 376, + 284, + 388 + ], + "spans": [ + { + "bbox": [ + 67, + 376, + 284, + 388 + ], + "type": "text", + "content": "Table 5. Performance evaluation on the PVSG [38] Dataset" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 391, + 294, + 444 + ], + "lines": [ + { + "bbox": [ + 58, + 391, + 294, + 444 + ], + "spans": [ + { + "bbox": [ + 58, + 391, + 294, + 444 + ], + "type": "table", + "html": "
MethodPVSG Metrics
Stage-1Stage-2mR@20mR@50mR@100
VPS [4, 17]Transformer0.360.390.40
IPS+T [4, 34]Transformer1.751.862.03
MGSMIRG8.828.9410.51
", + "image_path": "18ffb617f798f52f61eb2a32229e1967d041913a1aaa05611142b7eb2a81c1e8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 58, + 480, + 296, + 521 + ], + "blocks": [ + { + "bbox": [ + 55, + 454, + 295, + 475 + ], + "lines": [ + { + "bbox": [ + 55, + 454, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 295, + 475 + ], + "type": "text", + "content": "Table 6. Performance evaluation on the PSG4D-GTA and PSG4D-HOI [37] 4DPVSG Datasets." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 480, + 296, + 521 + ], + "lines": [ + { + "bbox": [ + 58, + 480, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 58, + 480, + 296, + 521 + ], + "type": "table", + "html": "
MethodPSG4D-GTAPSG4D-HOI
mR@20mR@50mR@100mR@20mR@50mR@100
PSG4DFounder (P) [37]2.102.933.133.103.954.17
PSG4DFounder (R) [37]3.313.854.023.654.164.97
MGSM + IRG10.8311.5111.9810.9512.3212.87
", + "image_path": "8b33206568b2804b6ba9cf38fc0412af82815f1c8f617fc912fbea8f7854057c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 533, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 533, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 54, + 533, + 295, + 628 + ], + "type": "text", + "content": "at " + }, + { + "bbox": [ + 54, + 533, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\lambda = 0.06" + }, + { + "bbox": [ + 54, + 533, + 295, + 628 + ], + "type": "text", + "content": ". The differences in optimal " + }, + { + "bbox": [ + 54, + 533, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 54, + 533, + 295, + 628 + ], + "type": "text", + "content": " values reflect the distinct characteristics of these tasks. Specifically, SGCLS benefits from ground-truth data, enabling effective updates to the memory representations with a smaller " + }, + { + "bbox": [ + 54, + 533, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 54, + 533, + 295, + 628 + ], + "type": "text", + "content": ". In contrast, SGDET, which builds scene graphs from scratch, performs better with a slightly higher " + }, + { + "bbox": [ + 54, + 533, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 54, + 533, + 295, + 628 + ], + "type": "text", + "content": ". Notably, since the MSGM module was not applied to the PREDCLS task, the parameter " + }, + { + "bbox": [ + 54, + 533, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 54, + 533, + 295, + 628 + ], + "type": "text", + "content": " is also not involved in PREDCLS." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 629, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 629, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 629, + 295, + 715 + ], + "type": "text", + "content": "Exploration on PVSG and 4DPVSG Datasets. We extended our framework to Panoptic Video Scene Graph Generation (PVSG) and 4D Panoptic Scene Graph Generation (4DPVSG) tasks [37, 38]. Unlike VidSGG, PVSG utilizes pixel-level masks for grounding nodes. Using the IPS+T+Transformer codebase, we integrated MGSM with a Memory-Enhanced Temporal Integrator and employed" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 378, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 555, + 510 + ], + "type": "text", + "content": "UniTrack [34] for temporal linking. For relation classification, IRG was adapted for iterative semantic debiasing. The 4DPVSG task further generalized PVSG to dynamic 4D contexts, employing PSG4DFormer variants with point cloud (P) and RGB-D (R) inputs. As shown in Table 5, our method significantly outperformed baselines, achieving minimum improvements of " + }, + { + "bbox": [ + 313, + 378, + 555, + 510 + ], + "type": "inline_equation", + "content": "+7.52\\%" + }, + { + "bbox": [ + 313, + 378, + 555, + 510 + ], + "type": "text", + "content": " mR@20 (PVSG) and " + }, + { + "bbox": [ + 313, + 378, + 555, + 510 + ], + "type": "inline_equation", + "content": "+7.30\\%" + }, + { + "bbox": [ + 313, + 378, + 555, + 510 + ], + "type": "text", + "content": " mR@20 (4DPVSG, PSG4D-HOI). These results confirm the broad applicability of our dual debiasing approach across VidSGG and indicate potential benefits for other generative and embodied intelligence applications." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 518, + 388, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 388, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 388, + 530 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 534, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 715 + ], + "type": "text", + "content": "In this work, we present VISA, an approach for unbiased video scene graph generation that addresses both visual and semantic biases inherent in scene graphs. VISA reduces visual bias by enriching current frame representations with historical context to decrease visual feature variance, and mitigates semantic bias by increasing the KL divergence between model predictions and biased priors. Our framework integrates hierarchical semantic contexts with visual features, enabling more unbiased scene graph generation. Extensive experimental and theoretical analyses demonstrate the effectiveness of VISA. Future work will explore extending our visual-semantic debiasing approach to other multimodal domains such as visual question answering and video captioning, as well as applications in embodied intelligence, including robotics perception and interactive agents." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19054" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 137 + ], + "type": "text", + "content": "We gratefully acknowledge the three anonymous reviewers for their constructive feedback on our CVPR 2025 submission. This work was supported by personal funding from the corresponding author." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "spans": [ + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 176, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 176, + 296, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 176, + 296, + 220 + ], + "spans": [ + { + "bbox": [ + 61, + 176, + 296, + 220 + ], + "type": "text", + "content": "[1] João Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 222, + 296, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 222, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 61, + 222, + 296, + 266 + ], + "type": "text", + "content": "[2] Shizhe Chen, Zhiyuan Shi, Pascal Mettes, and Cees G. M. Snoek. Social fabric: Tubelet compositions for video relation detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR), 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 266, + 295, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 266, + 295, + 310 + ], + "spans": [ + { + "bbox": [ + 61, + 266, + 295, + 310 + ], + "type": "text", + "content": "[3] Shizhe Chen, Jing Xiao, and Lin Chen. Video scene graph generation from single-frame weak supervision. In Proceedings of the Eleventh International Conference on Learning Representations (ICLR), 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 311, + 296, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 311, + 296, + 366 + ], + "spans": [ + { + "bbox": [ + 61, + 311, + 296, + 366 + ], + "type": "text", + "content": "[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 367, + 296, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 367, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 61, + 367, + 296, + 422 + ], + "type": "text", + "content": "[5] Yunhan Cong, Wentao Liao, Heiko Ackermann, Bodo Rosenhahn, and Michael Ying Yang. Spatial-temporal transformer for dynamic scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 422, + 296, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 296, + 465 + ], + "type": "text", + "content": "[6] Yunhan Cong, Michael Ying Yang, and Bodo Rosenhahn. Reltr: Relation transformer for scene graph generation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 468, + 296, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 296, + 511 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 296, + 511 + ], + "type": "text", + "content": "[7] Linfeng Gao, Bin Wang, and Wenmin Wang. Image captioning with scene-graph based semantic concepts. In Proceedings of the International Conference on Machine Learning and Computing, 2018. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 512, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 296, + 555 + ], + "type": "text", + "content": "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 557, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 296, + 590 + ], + "type": "text", + "content": "[9] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017. 2, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 590, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 296, + 634 + ], + "type": "text", + "content": "[10] Marcel Hildebrandt, Haoyang Li, Raphael Koner, Volker Tresp, and Stephan Gunnemann. Scene graph reasoning for visual question answering. arXiv preprint arXiv:2007.01072, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 635, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 296, + 689 + ], + "type": "text", + "content": "[11] Jiading Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as compositions of spatiotemporal scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "text", + "content": "[12] Anant Khandelwal. Flocode: Unbiased dynamic scene graph generation with temporal consistency and correlation debias" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 106 + ], + "type": "text", + "content": "ing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2024. 1, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 106, + 554, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 554, + 139 + ], + "type": "text", + "content": "[13] Shreyank Khandelwal and Leonid Sigal. Iterative scene graph generation. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 140, + 554, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 554, + 194 + ], + "type": "text", + "content": "[14] Hengshuang Li, Xin Li, Behnam Karimi, Jie Chen, and Ming Sun. Joint learning of object graph and relation graph for visual question answering. In Proceedings of the IEEE International Conference on Multimedia and Expo (ICME), 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 195, + 554, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 554, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 554, + 249 + ], + "type": "text", + "content": "[15] Lifeng Li, Jing Xiao, Huan Shi, Wei Wang, Jing Shao, Alan Liu, Yueting Yang, and Lin Chen. Label semantic knowledge distillation for unbiased scene graph generation. IEEE Transactions on Circuits and Systems for Video Technology (TCSVT), 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 250, + 554, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 554, + 294 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 554, + 294 + ], + "type": "text", + "content": "[16] Ruofei Li, Shuhui Zhang, and Xiaochun He. Sgtr: End-to-end scene graph generation with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 294, + 554, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 554, + 358 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 554, + 358 + ], + "type": "text", + "content": "[17] Xiangtai Li, Wenwei Zhang, Jiangmiao Pang, Kai Chen, Guangliang Cheng, Yunhai Tong, and Chen Change Loy. Video k-net: A simple, strong, and unified baseline for video segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 360, + 554, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 554, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 554, + 404 + ], + "type": "text", + "content": "[18] Yicong Li, Xun Yang, Xindi Shang, and Tat-Seng Chua. Interventional video relation detection. In Proceedings of the 29th ACM International Conference on Multimedia (ACMMM), 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 404, + 554, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 554, + 449 + ], + "type": "text", + "content": "[19] Yifei Li, Xinyi Yang, and Chenglin Xu. Dynamic scene graph generation via anticipatory pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 449, + 554, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 554, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 554, + 503 + ], + "type": "text", + "content": "[20] Chang Liu, Yining Jin, Kaixuan Xu, Guohao Gong, and Yuesheng Mu. Beyond short-term snippet: Video relation detection with spatio-temporal global context. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 503, + 554, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 503, + 554, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 503, + 554, + 548 + ], + "type": "text", + "content": "[21] Yang Liu, Guanbin Li, and Liang Lin. Cross-modal causal relational reasoning for event-level visual question answering. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023. 2, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 548, + 554, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 554, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 554, + 592 + ], + "type": "text", + "content": "[22] Cewu Lu, Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Visual relationship detection with language priors. In Proceedings of the European Conference on Computer Vision (ECCV), 2016. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 592, + 554, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 554, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 554, + 646 + ], + "type": "text", + "content": "[23] Sourya Nag, Kyusong Min, Swami Tripathi, and Amit K. Roy-Chowdhury. Unbiased scene graph generation in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 554, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 691 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 691 + ], + "type": "text", + "content": "[24] Jeffrey Pennington, Richard Socher, and Christopher D. Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2014. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "type": "text", + "content": "[25] Tao Pu, Tianshui Chen, Hefeng Wu, Yongyi Lu, and Liang Lin. Spatial-temporal knowledge-embedded transformer for" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19055" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "video scene graph generation. IEEE Transactions on Image Processing (TIP), 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "type": "text", + "content": "[26] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2017. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "type": "text", + "content": "[27] Scott Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D. Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the Fourth Workshop on Vision and Language, 2015. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 239 + ], + "type": "text", + "content": "[28] Xun Shang, Yicheng Li, Jing Xiao, Weiqing Ji, and Tat-Seng Chua. Video visual relation detection via iterative inference. In Proceedings of the 29th ACM International Conference on Multimedia (ACMMM), 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 259 + ], + "type": "text", + "content": "[29] Claude E Shannon. A mathematical theory of communication. Bell System Technical Journal, 1948. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 261, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 327 + ], + "type": "text", + "content": "[30] Shubham Shit, Raphael Koner, Benedikt Wittmann, Johannes Paetzold, Ivan Ezhov, Haoyang Li, Jindong Pan, Sina Sharifzadeh, Georgios Kaissis, and Volker Tresp. Relationformer: A unified framework for image-to-graph generation. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 294, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 371 + ], + "type": "text", + "content": "[31] Yudong Teng, Liang Wang, Zhenyu Li, and Gang Wu. Target adaptive context aggregation for video scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR), 2021. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 372, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 415 + ], + "type": "text", + "content": "[32] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 416, + 294, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 294, + 469 + ], + "type": "text", + "content": "[33] Shizhe Wang, Linfeng Gao, Xinyi Lyu, Yulan Guo, Ping Zeng, and Jie Song. Dynamic scene graph generation via temporal prior inference. In Proceedings of the 30th ACM International Conference on Multimedia (ACMMM), 2022. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "type": "text", + "content": "[34] Zhongdao Wang, Hengshuang Zhao, Ya-Li Li, Shengjin Wang, Philip Torr, and Luca Bertinetto. Do different tracking tasks require different appearance models? Advances in Neural Information Processing Systems (NeurIPS), 2021. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 515, + 294, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 559 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 559 + ], + "type": "text", + "content": "[35] Ziyang Wang, Yi-Lin Sung, Feng Cheng, Gedas Bertasius, and Mohit Bansal. Unified coarse-to-fine alignment for video-text retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 559, + 294, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 603 + ], + "type": "text", + "content": "[36] Danfei Xu, Yuke Zhu, Christopher B. Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 604, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 294, + 657 + ], + "type": "text", + "content": "[37] Jingkang Yang, Jun Cen, Wenxuan Peng, Fangzhou Liu, Shuai amd Hong, Xiangtai Li, Kaiyang Zhou, Qifeng Chen, and Ziwei Liu. 4d panoptic scene graph generation. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[38] Jingkang Yang, Wenxuan Peng, Xiangtai Li, Zujin Guo, Liangyu Chen, Bo Li, Zheng Ma, Kaiyang Zhou, Wayne Zhang, Chen Change Loy, et al. Panoptic video scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 228 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "text", + "content": "[39] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 554, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 172 + ], + "type": "text", + "content": "[40] Jiaxian Zhang, Kuan-Jui Shih, Ahmed Elgammal, Alexander Tao, and Bryan Catanzaro. Graphical contrastive losses for scene graph parsing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 174, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 554, + 228 + ], + "type": "text", + "content": "[41] Cheng Zheng, Linfeng Gao, Xinyi Lyu, Ping Zeng, Abdulmotaleb El Saddik, and Heng Tao Shen. Dual-branch hybrid learning network for unbiased scene graph generation. IEEE Transactions on Circuits and Systems for Video Technology (TCSVT), 2023. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "19056" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Unbiasing through Textual Descriptions_ Mitigating Representation Bias in Video Benchmarks/5edfcfce-ad23-41c6-9a43-d3cb9f8bcb90_content_list.json b/2025/Unbiasing through Textual Descriptions_ Mitigating Representation Bias in Video Benchmarks/5edfcfce-ad23-41c6-9a43-d3cb9f8bcb90_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ab8ba56528cbc443b9301f81e1e2ab4af2cf47e6 --- /dev/null +++ b/2025/Unbiasing through Textual Descriptions_ Mitigating Representation Bias in Video Benchmarks/5edfcfce-ad23-41c6-9a43-d3cb9f8bcb90_content_list.json @@ -0,0 +1,1601 @@ +[ + { + "type": "text", + "text": "Unbiasing through Textual Descriptions: Mitigating Representation Bias in Video Benchmarks", + "text_level": 1, + "bbox": [ + 228, + 130, + 769, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nina Shvetsova $^{1,2,3*}$ , Arsha Nagrani $^{4}$ , Bernt Schiele $^{3}$ , Hilde Kuehne $^{1,2,5}$ , Christian Rupprecht $^{4}$ $^{1}$ Goethe University Frankfurt, $^{2}$ Tuebingen AI Center/University of Tuebingen, $^{3}$ MPI for Informatics, Saarland Informatics Campus, $^{4}$ University of Oxford, $^{5}$ MIT-IBM Watson AI Lab shvetsov@uni-frankfurt.de", + "bbox": [ + 102, + 202, + 893, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 309, + 326, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a new \"Unbiased through Textual Description (UTD)\" video benchmark based on unbiased subsets of existing video classification and retrieval datasets to enable a more robust assessment of video understanding capabilities. Namely, we tackle the problem that current video benchmarks may suffer from different representation biases, e.g., object bias or single-frame bias, where mere recognition of objects or utilization of only a single frame is sufficient for correct prediction. We leverage VLMs and LLMs to analyze and debias benchmarks from such representation biases. Specifically, we generate frame-wise textual descriptions of videos, filter them for specific information (e.g. only objects) and leverage them to examine representation biases across three dimensions: 1) concept bias—determining if a specific concept (e.g., objects) alone suffice for prediction; 2) temporal bias—assessing if temporal information contributes to prediction; and 3) common sense vs. dataset bias—evaluating whether zero-shot reasoning or dataset correlations contribute to prediction. We conduct a systematic analysis of 12 popular video classification and retrieval datasets and create new object-debiased test splits for these datasets. Moreover, we benchmark 30 state-of-the-art video models on original and debiased splits and analyze biases in the models. To facilitate the future development of more robust video understanding benchmarks and models, we release: \"UTD-descriptions\", a dataset with our rich structured descriptions for each dataset, and \"UTD-splits\", a dataset of object-debiased test splits.1", + "bbox": [ + 86, + 340, + 483, + 765 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 795, + 220, + 810 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ever since action recognition moved from simple, staged clip classification to real-life video understanding [18, 29,", + "bbox": [ + 89, + 821, + 482, + 852 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guess the activity from the following objects:", + "text_level": 1, + "bbox": [ + 562, + 310, + 854, + 323 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video 1:", + "bbox": [ + 588, + 328, + 635, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Person, makeup brush, makeup sponge, makeup palette", + "bbox": [ + 519, + 342, + 699, + 364 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Answer:", + "bbox": [ + 542, + 368, + 589, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video 2:", + "bbox": [ + 781, + 328, + 830, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Person, stage, wall, curtain", + "bbox": [ + 730, + 345, + 883, + 358 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Answer:", + "bbox": [ + 736, + 368, + 785, + 377 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guess the activity from a single frame:", + "text_level": 1, + "bbox": [ + 584, + 392, + 833, + 405 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video 3:", + "bbox": [ + 588, + 406, + 635, + 416 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7cca5a13202945120aa3cf187b0849c1491a2df52b3caf83675a1e8cba9d6439.jpg", + "image_caption": [ + "Answer:" + ], + "image_footnote": [], + "bbox": [ + 542, + 417, + 683, + 488 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video 4:", + "bbox": [ + 785, + 407, + 833, + 416 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1f2f04ebfc6cb72bbb6f1fef87983da65e430507e89d9a0b00e3a589d3877ca5.jpg", + "image_caption": [ + "Answer:", + "Figure 1. Can you guess activity on a video based solely on the objects or a single frame? (The answer is on the next page in Fig. 2.)" + ], + "image_footnote": [], + "bbox": [ + 738, + 417, + 877, + 488 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "36], there has been an ongoing discussion about how much of a system's performance can be attributed to the recognition of the action itself compared to related or unrelated visual cues such as objects or scene information in the data [11, 15, 47], namely object/scene representation bias (see Fig. 1). Different strategies have been proposed to separate action recognition from scene and object influences such as segmenting the human figure in a video and placing it against a gray or black background [11, 15] as well as mimicking actions in unrelated contexts [47]. Some works focused on collecting datasets designed to minimize scene or object representation bias [3, 14, 23]. However, debiased solutions have so far rarely been adopted for benchmarking. Reasons for that may be that modified videos in case of [11, 15] often introduce artifacts and benchmarks with explicitly low scene bias [14, 23], might be considered out-of-domain relative to standard training setups. Another line of research questions whether current video benchmarks truly require temporal understanding, as single-frame models have shown comparable performance on several benchmarks [5, 19]. While new benchmarks are emerging to assess temporal understanding [2, 7, 28], temporal representen", + "bbox": [ + 511, + 568, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*The work was done during PhD visit to the University of Oxford within the ELLIS PhD program. \n1Project page: https://utd-project.github.io/", + "bbox": [ + 89, + 862, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29050", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tation bias in standard benchmarks remains underexplored.", + "bbox": [ + 89, + 90, + 480, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a new scalable method, Unbiasing through Textual Descriptions (UTD), to systematically analyze and mitigate different types of representation bias, such as object bias, in existing video benchmarks through automatic textual descriptions. Since most datasets lack textual narrations of videos, we leverage Vision-Language Models (VLMs) to generate detailed descriptions of video frames. We then use Large Language Models (LLMs) to extract different concepts from those captions, namely objects, activities, and verbs (an overview of this process is shown in Fig. 3). Based on generated descriptions, we examine representation biases across three independent dimensions. First, concept representation bias — assessing if a specific concept (e.g., objects) alone is sufficient for accurate prediction of the video. Second, temporal representation bias — evaluating if temporal information is needed for correct prediction or if single-frame information is sufficient. Lastly, we distinguish between common sense and dataset bias. Common sense bias refers to the implicit association of certain concepts with specific activities, such as associating an object “a piano” with the label “playing piano.” This differs from dataset bias, where spurious correlations arise within the dataset, e.g., “apply makeup” label frequently occurs with objects like “mirror” and “flowers” in the background. We propose to leverage the zero-shot reasoning capabilities of LLMs to assess common sense bias and train a linear classifier for the dataset bias.", + "bbox": [ + 89, + 109, + 483, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In our study, we found that many samples of popular benchmarks have a significant object representation bias, meaning that recognizing only objects on the video is already enough to predict the correct label. Therefore, using such data for benchmarking video backbones may give models with strong object recognition ability an advantage (i.e. a shortcut), unintentionally hindering video understanding evaluation with respect to temporal reasoning or action recognition. Therefore, by excluding test samples with the highest object bias, we introduce UTD-splits, debiased evaluation splits for considered datasets, which we believe provide a more robust assessment of video understanding capabilities. Importantly, compared to other approaches that rely on modifying videos [11, 15], this approach neither requires any additional processing of evaluation data, as no new videos are generated nor does it require any fine-tuning to adjust for any domain bias that could be introduced by debiasing video frames at a pixel level.", + "bbox": [ + 89, + 520, + 483, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We apply the proposed UTD method to analyze and debias six action classification datasets, namely UCF101 [36], SomethingSomethingv2 [14], Kinetics-400 [17], -600 [8], and -700 [9], and Moments In Time [31], and six text-to-video retrieval datasets, namely MSRVTT [48], DiDeMo [1], ActivityNet [6], LSMDC [33], YouCook2 [12], and Spoken Moments In Time [32]. Fur", + "bbox": [ + 89, + 795, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/834fcd624e91a485f644e93d98bb548065be8c122eceebe071789c400e19f283.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 90, + 705, + 188 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f2f50bbbf6cf8ed897ca6a7a880cef0a794ca46d3adf699c7747b77996b8529a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 89, + 895, + 186 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dadd4b2751de16f84e3c1f01fd49e0f814e5c466ca9a25f98aa424b095ee8702.jpg", + "image_caption": [ + "Figure 2. The answer to Fig. 1. Some videos may exhibit an object representation bias (allowing predictions based solely on objects) or single-frame representation bias (solely on a single frame), while others require more information for prediction." + ], + "image_footnote": [], + "bbox": [ + 516, + 189, + 705, + 294 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/31d354e5d76d81af4e06ea7fe752cb877a5b166085ca9c5da74f63ac7725b900.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 189, + 901, + 294 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ther, we benchmark 30 state-of-the-art video models of six different model designs, namely VideoMAE [38], VideoMAE2 [42], All-in-one [40], Unmasked Teacher [21], VideoMamba [22], and InternVid [44], on the original and UTD-debiased splits. We show that debiasing reduces performance saturation on current benchmarks. Considering model performance drops between original and debiased splits also allows us to analyze models on object bias, e.g., in action classification, larger backbones or those trained on more data appear more robust to object bias.", + "bbox": [ + 511, + 364, + 906, + 515 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize the contributions of this work as follows: (1) We introduce Unbiasing through Textual Descriptions (UTD), a scalable, automated method to analyze and mitigate representation biases in video benchmarks without the need for human annotators in the loop. (2) Using UTD, we analyze representation biases across 12 popular classification and retrieval video benchmarks and release UTD-debiased test splits for each, UTD-splits, promoting a more robust evaluation of video understanding. (3) We benchmark 30 state-of-the-art video models on both the original and UTD-debiased splits to assess their robustness against object bias under varying conditions. (4) Finally, we release UTD-descriptions, a dataset of detailed structured descriptions (including objects, activities, and verb categories) for all 12 considered datasets, to support further dataset analysis and the development of less biased benchmarks.", + "bbox": [ + 511, + 515, + 908, + 757 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 768, + 656, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Debiasing Video Datasets. While most visual datasets are prone to various biases [13], one unique effect in video understanding is the impact of object and scene information on the overall performance. Different strategies have so far been explored to mitigate this bias. He et al. [15] and HAT [11] proposed to separate the performance of scene appearance from human motion by segmenting human fig-", + "bbox": [ + 511, + 794, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29051", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7e412135a17f8098fd8ef3f6745550ec7642f0351cc698c3e24509721f75ac04.jpg", + "image_caption": [ + "Figure 3. The proposed UTD method involves generating textual descriptions of different concepts in video frames using VLMs and LLMs, combining them in various temporal configurations, and evaluating the performance of these concept-temporal representations with strong text embedding models. For each representation, we distinguish between common sense bias, which relies on zero-shot reasoning by text embedding models, and dataset bias, assessed using a linear model trained on the dataset's training set." + ], + "image_footnote": [], + "bbox": [ + 89, + 89, + 903, + 358 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ures and placing them into black/gray background, or alternatively by cropping or inpainting human figures and classifying videos based on background only. Weinzaepfel et al. [47] proposed the Mimetics dataset by extending certain classes from the Kinetics dataset [17] with videos of activities in unrelated background. Zhai et al. [51] proposed an approach for scene-debiased action recognition via adversarial scene reconstruction and classification. Another line of work focused on capturing datasets with explicitly low object and scene bias. SomethingSomethingv1 and later v2 dataset [14] were proposed as an almost scene-free dataset, focusing on simple manipulative actions only. Li et al. [23] proposed a procedure for designing representation unbiased datasets and introduced Diving48 as an example. However, this specialization usually also implies that those cases are rather suitable for specialized approaches but not for, e.g., zero-shot evaluation of generalist video models. Other studies [5, 19] showed that for several benchmarks, using single frames during training or inference can achieve performance comparable to models that incorporate temporal information. However, recent efforts have mainly focused on developing new benchmarks to assess temporal understanding in videos [2], e.g., visual question-answering benchmarks [7, 28], while temporal representation bias in standard benchmarks has not been thoroughly investigated. In contrast, our UTD method allows for a systematic analysis of representation biases across various dimensions, including different concept and temporal configurations. Additionally, we are the first to propose mitigating object representation bias in popular benchmarks by discarding object-biased samples, which helps us to avoid adding artifacts to", + "bbox": [ + 91, + 426, + 483, + 893 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "test samples (as with cropping or inpainting in [11, 15]) and keeps the dataset domain unconstrained, unlike [14, 23].", + "bbox": [ + 511, + 426, + 906, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LLMs for Video Understanding. The progress of LLMs prompted a wide range of applications in various fields, including video understanding. Here, one line of work utilizes pre-trained LLMs in video-language tasks, such as video captioning [37, 46, 53], video question answering [21, 27, 46, 53], text-video retrieval [35, 53], or dialog-based user interaction [46, 52]. Some works [10, 34, 44] create new video-language datasets using image/audio captioning models or subtitles and using LLM summarization and reasoning ability to post-process these captions. Other works [20, 39, 45, 50] use expert perception models to extract information from the video that is later processed by the LLM for tasks such as video captioning. In our work, we leverage VLM- and LLM-generated frame descriptions to evaluate different representation biases in video datasets.", + "bbox": [ + 511, + 455, + 908, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Discovering Representation Biases", + "text_level": 1, + "bbox": [ + 511, + 695, + 826, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we define representation bias in Sec. 3.1, outline the considered representation biases in Sec. 3.2, and detail our UTD method in Sec. 3.3.", + "bbox": [ + 511, + 719, + 906, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Definition: Representation Bias", + "text_level": 1, + "bbox": [ + 511, + 773, + 792, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To formalize the definition of representation bias in datasets, we adopt the notation from RESOUND [23].", + "bbox": [ + 511, + 795, + 906, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $D = \\{v_{n}\\}_{n\\leq N}$ be a video dataset $D$ where each video $v_{n}$ consists of a sequence of frames $v_{n} = \\{f_{n,i}\\}_{i\\leq |v_n|}$ . A representation $\\phi$ is a function $\\phi (v)$ of the video input $v$ . Let $R = \\{\\phi \\}$ be a family of representations that share specific characteristics. For example, a fam", + "bbox": [ + 511, + 825, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "29052", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ily of single-frame representations includes representations that encode information from only a single frame of a video $\\phi (v_{n}) = \\phi (f_{n,i})$ where $i\\leq |v_{n}|$ or a family of object-based representations encodes only the objects in the video without considering motion or activity information.", + "bbox": [ + 89, + 90, + 480, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The representation $\\phi$ can be used to train different models $\\gamma_{\\phi}$ for desired tasks, e.g. video classification. Let $M(v_{n},\\gamma_{\\phi})$ be the performance (e.g., accuracy) of a model $\\gamma_{\\phi}$ on a video $v_{n}$ , then the performance of $\\gamma_{\\phi}$ on $D$ is", + "bbox": [ + 89, + 167, + 482, + 228 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM (D, \\gamma_ {\\phi}) = \\frac {1}{N} \\sum_ {n = 1} ^ {N} M \\left(v _ {n}, \\gamma_ {\\phi}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 233, + 482, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We define the performance of a representation $\\phi$ on $D$ as", + "bbox": [ + 89, + 277, + 467, + 292 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM (D, \\phi) = \\max _ {\\gamma_ {\\phi}} M (D, \\gamma_ {\\phi}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 300, + 482, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the maximum is taken over all models $\\gamma_{\\phi}$ that are based on the representation $\\phi$", + "bbox": [ + 89, + 328, + 482, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When the performance $M(D, \\phi)$ of a representation $\\phi$ is high, the dataset exhibits a preference for this representation, indicating that the dataset is more biased toward this representation. In this work, we use $M(D, \\phi)$ as a measure of bias towards representation $\\phi$ . However, a high bias toward a specific representation is not necessarily a negative thing. Ideally, a dataset should exhibit the highest bias toward the \"ground truth representation\": the representation truly required to solve the task. For instance, in video understanding, one could argue that the optimal representation should capture both static (e.g., objects) and dynamic (e.g., actions) properties, as well as encode temporal aspects. However, if a video dataset used to evaluate video understanding exhibits a strong object or single-frame representation bias, it may unintentionally favor models with strong object recognition or image understanding capabilities, thus hindering the evaluation of true video understanding. Therefore, in this work, we aim to estimate representation biases in popular video benchmarks.", + "bbox": [ + 89, + 358, + 482, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Different Representation Biases", + "text_level": 1, + "bbox": [ + 89, + 652, + 369, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our work, we consider different types of representation biases with respect to three independent aspects:", + "bbox": [ + 89, + 674, + 482, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concept Representations: First, we consider representations that encode four different concept categories: objects $\\phi_{o}(v_{n})$ , activities $\\phi_{a}(v_{n})$ (actions along with tied objects), verbs $\\phi_{\\nu}(v_{n})$ (actions alone, without objects), and objects+composition+activities $\\phi_{d}(v_{n})$ (objects, their spatial compositions in the video, and activities), see an example in Fig. 3. We aim to assess biases in classification and retrieval benchmarks toward particular concepts. For example, in retrieval, we ideally would like to see that benchmarks are not biased solely toward objects or activities.", + "bbox": [ + 88, + 704, + 482, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temporal Representations: Second, we consider the bias of single-frame representations compared to representations that consider multiple frames. Let $\\phi_x$ be", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "one of the objects $\\phi_o$ , activities $\\phi_a$ , verbs $\\phi_\\nu$ , or objects+composition+activities $\\phi_d$ representations. Similarly to [41], we consider the following options to encode information from multiple frames: middle-frame refers to the case that a video is represented only by a single middle frame of this video, e.g., $\\phi_{x,\\mathrm{mid}}(v_n) = \\phi_x(f_{n,|v_n| / 2})$ , max-score frame refers to the case that a video is represented using a single frame $\\phi_{x,\\max}(v_n) = \\phi_x(f_{n,i})$ where the frame $f_{n,i}$ has the highest performance score $M(f_{n,i},\\phi_x)$ over all frames, average-overframes refers to the case that the video is represented as an average representation of frames ignoring the order of frames $\\phi_{x,\\mathrm{avg}}(v_n) = \\frac{1}{|v_n|}\\sum_{i\\leq |v_n|}\\phi_x(f_{n,i})$ , and, finally, sequence-of-frames refers to the case that the video representation encodes concepts and their order in the frames $\\phi_{x,\\mathrm{seq}}(v_n) = \\phi_x(f_{n,1},f_{n,2},..,f_{n,|v_n|})$ .", + "bbox": [ + 511, + 90, + 903, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Common Sense Bias vs. Dataset Bias: Finally, for any concept-temporal representation $\\phi$ (e.g., object middle-frame representation $\\phi_{o,mid}$ ), we can consider two types of representation biases: common sense bias or dataset bias. Common sense representation bias refers to bias based on \"common sense reasoning\", e.g., a video featuring an object \"piano\" would most likely be classified under the \"playing piano\" activity label. Formally, in Eq. (2), this bias $M(D,\\phi)$ is measured by taking the maximum performance only over zero-shot models $\\gamma_{\\phi}$ that have not seen the correlations in the training set of $D$ . We distinguish this from dataset bias, which refers to biases caused by spurious correlations within the dataset, e.g., where certain objects are unintentionally associated with specific labels, such as \"mirror\" or \"flower bouquet\" frequently appear with the \"apply makeup\" label. In this case, in Eq. (2), we measure dataset bias by taking the maximum performance over models $\\gamma_{\\phi}$ trained or fine-tuned on the training set of $D$ .", + "bbox": [ + 511, + 335, + 903, + 607 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We note that the concept, temporal, and common sense vs. dataset bias dimensions are independent, allowing us to combine them in any configuration. For example, we can evaluate object middle-frame common sense bias or activity sequence-of-frames dataset bias.", + "bbox": [ + 511, + 608, + 903, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Unbiasing through Textual Descriptions (UTD)", + "text_level": 1, + "bbox": [ + 511, + 695, + 903, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since measuring bias for every representation within a family (e.g., a family of object representations) is infeasible, we approximate it by considering a single representative from the family [23] and propose a way to precisely control what each representation encodes by using textual descriptions.", + "bbox": [ + 511, + 718, + 903, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The key idea of our method is to use textual descriptions as an intermediate video representation. This approach allows us to easily control which concepts in the video are included or excluded in the representation and to combine information into different temporal representations. Namely, to obtain representation $\\phi$ that encodes only a specific concept category, we propose to encode visible concepts of this", + "bbox": [ + 511, + 795, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29053", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "category into textual descriptions, e.g., encode objects with a list of visible objects, e.g., \"table, lamp, ..\". This approach, compared to cropping or inpainting [11, 15], allows us to isolate specific concepts more precisely, avoiding information leakage from cropped areas or adding artifacts caused by imperfect inpainting. Additionally, textual descriptions allow us to use off-the-shelf LLMs with strong common sense reasoning abilities to estimate common sense bias. Since metadata with visible objects, activities, etc., is usually not available, we propose using recent powerful VLMs and LLMs to extract this information.", + "bbox": [ + 89, + 90, + 480, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the next sections, we first describe our method of obtaining textual description with respect to different concept and temporal categories (Sec. 3.3.1), then we describe our method of discovering common sense and dataset biases (Sec. 3.3.2), and lastly, we describe our method of unbiaseding benchmarks (Sec. 3.3.3).", + "bbox": [ + 89, + 257, + 480, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.1. Obtaining Textual Descriptions", + "text_level": 1, + "bbox": [ + 89, + 353, + 354, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While current VLMs [24-26], trained to perform complex image reasoning, are able to generate detailed descriptions of video frames, they still face challenges when prompted to directly extract specific concepts from a frame, e.g., only visible objects. Therefore, we decomposed the task of concept extraction into two parts: extraction of detailed frame description using VLMs and then extraction of different concepts from frame description using LLMs (see Fig. 3).", + "bbox": [ + 89, + 372, + 482, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Obtaining Textual Descriptions Using VLMs. Given train/test videos $v_{n} = \\{f_{n,i}\\}_{i\\leq |v_{n}|}, n < N$ of the dataset, we feed each frame $f_{n,i}$ of each video $v_{n}$ into the VLM model $d$ to generate a detailed textual description $d_{n,i} = d(f_{n_i})$ (Fig. 3). We found that without complex prompt engineering or few-shot in-context learning strategies [4] VLMs produce long, detailed descriptions. While our method is not restricted to a specific VLM, for reproducibility and feasibility, we select an open-sourced strong middle-scale VLM model, LLaVA-1.6-Mistral-7B [26] with the prompt: \"Describe the objects' relationships in the photo.\" Since the resulting textual descriptions include objects, their composition, relationships, and visible activities, we use them as objects+composition+activities descriptions.", + "bbox": [ + 89, + 493, + 482, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Extracting Concepts Categories Using LLMs. Each of the obtained detailed descriptions $d_{n,i}$ is further fed to the LLM $\\zeta$ to extract only objects $o_{n,i} = \\zeta(d_{n,i}, p_o)$ , only activities $a_{n,i} = \\zeta(d_{n,i}, p_a)$ information, with $p_o, p_a$ as respective prompts. Then, to obtain information about verbs, we task the LLM with a prompt $p_\\nu$ to delete all object information from activities $\\nu_{n,i} = \\zeta(a_{n,i}, p_\\nu)$ . Since such text processing requires understanding what objects and activities are, our prompt includes a few examples following a few-shot in-context learning strategy [4]. We provide all prompts in the supplementary material.", + "bbox": [ + 89, + 704, + 482, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Aggregating Temporal Information. We use the obtained textual descriptions $x_{n,i}$ of video $v_{n}$ (where $x_{n,i}$ is one", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of $o_{n,i}$ $a_{n,i}$ , $\\nu_{n,i}$ , or $d_{n,i}$ ) in different temporal setups: first, we keep them as single frame descriptions that might be further aggregated when transformed into vector representations (e.g. average-over-frames), as we describe in Sec. 3.3.2, second, we combine them into sequence-over-frame $x_{n}$ setup with the template: \"Frame 1: $x_{n,1}$ Frame 2: $x_{n,2} \\ldots$ Frame $|v_{n}| \\colon x_{n,|v_{n}|}$ \" (Fig. 3). Since objects+composition+activities descriptions $d_{n,i}$ when combined with sequence-over-frame templates become too long, we summarize them to 15-word descriptions $d_{n,i}'$ via an LLM before inputting them into the sequence prompt.", + "bbox": [ + 511, + 90, + 903, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2. Calculating Representation Biases", + "text_level": 1, + "bbox": [ + 511, + 265, + 797, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After obtaining a specific concept-temporal representation $\\phi$ , we assess common sense and dataset bias for dataset $D$ toward $\\phi$ by measuring the performance $M(D, \\phi)$ , as stated Eq. (2), where maximum is taken over zero-shot models $\\gamma_{\\phi}$ for common sense bias and over fine-tuned models $\\gamma_{\\phi}$ for dataset bias (see Sec. 3.2). Since maximizing over all models is infeasible in practice, we design strong models for common sense and dataset-based reasoning by utilizing state-of-the-art LLMs. Our ablation study verifies our design choice (see supplementary material).", + "bbox": [ + 511, + 284, + 903, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Common Sense Bias. We leverage recent strong LLM-based text embedding models [30, 43]. Such models are trained on large datasets to encode text into vector representations that can be used for various tasks, such as information retrieval. We encode our textual descriptions and all class labels/captions using the pre-trained text embedding model $\\chi$ . Specifically, we prompt the model with the relevant concept and task. For example, we use prompts like \"Given a list of objects visible on the video frame, retrieve the activity depicted in this video\" for object description and \"Given an activity, retrieve a video frame description that may depict this activity\" for class labels in action classification. We use cosine similarity between obtained embeddings for zero-shot classification/retrieval. Prompts are detailed in the supplement.", + "bbox": [ + 511, + 435, + 905, + 661 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset Bias. For action classification datasets, we evaluate dataset bias, as action classification models are often tested in fine-tuned setups and might exploit dataset representation biases learned from training data. For this we a linear logistic regression model trained on our strong text embeddings used in the common sense bias analysis and train model on the official train splits of the datasets.", + "bbox": [ + 511, + 662, + 903, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. For the performance measure $M(D, \\phi)$ , we use accuracy for action classification and recall@1 and text-to-video retrieval as the most popular metrics for these tasks.", + "bbox": [ + 511, + 768, + 903, + 813 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.3. Unbiasing Datasets", + "text_level": 1, + "bbox": [ + 511, + 821, + 692, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Sec. 4, we find that many video datasets exhibit a significant object representation bias. To support robust video model evaluation, we construct UTD-splits, object-debiased test splits, by excluding object-biased samples from the", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29054", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/02ddb50ab1245ecda2edb9771f14246906ee39074b27cefe88a91f7e3b8047b5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
UCFSSv2K400K600K700MiTMSRDDMActNLSMDCYC2S-MiT
obj+comp+act66.36.448.044.139.022.636.627.226.617.08.445.9
objects63.3-3.05.3-1.145.9-2.141.8-2.337.0-2.021.0-1.632.1-4.527.0-0.224.8-1.813.6-3.47.9-0.529.8-16.1
activities67.4+1.16.445.2-2.841.4-2.736.7-2.321.0-1.625.1-11.521.1-6.121.4-5.214.7-2.38.1-0.341.1-4.8
verbs50.8-15.55.8-0.624.8-23.221.4-22.717.6-21.416.2-6.410.5-26.17.0-20.27.4-19.25.5-11.51.2-7.213.1-32.8
", + "bbox": [ + 93, + 88, + 911, + 157 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Evaluation of concept representation bias. We report the performance of our model in sequence-of-frames temporal setup and common sense bias setup and color code with respect to the difference to objects+composition+activities (obj+comp+act) concepts. It shows that the overall classification performance drops only slightly when classifying actions based on objects, compared to classification and retrieval on verbs only, with SSv2 as the only dataset with less drop when classifying on verbs only.", + "bbox": [ + 89, + 162, + 906, + 219 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "original test sets. Since our method allows us to measure the performance of the representation $M(v_{n},\\phi)$ on individual samples $v_{n}$ , we exclude samples that are classified or retrieved correctly ( $M(v_{n},\\phi) = 1$ ) based on object sequence-of-frame representation. For text-to-video retrieval, we debias with respect to common-sense bias, but for action classification, we debias with respect to dataset bias since video models are mostly tested in fine-tuned setup. To minimize the impact of random fluctuations in the process, we use three different prompts for labels/captions in the text embedding models and train three linear models for each classification dataset using bootstrapped training sets (we provide details and analysis in the supplement) and exclude only samples when all nine models agreed on correct classification/retrieval. Note that the percentage of removed samples is determined automatically based on the extent of object representation bias.", + "bbox": [ + 88, + 237, + 480, + 493 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, since debiasing may disproportionately remove certain label classes in classification datasets, we additionally construct balanced UTD splits. While maintaining the total number of removed samples as in the original debiasing method, we adjust the number of samples removed from each class based on their average confidence (across nine models) to preserve the original class proportions.", + "bbox": [ + 89, + 494, + 482, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Benchmarking Datasets", + "text_level": 1, + "bbox": [ + 89, + 614, + 316, + 632 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we analyze action classification and text-to-video retrieval datasets on different representation biases using our UTD method.", + "bbox": [ + 89, + 641, + 482, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 89, + 696, + 171, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We consider six common action classification datasets: UCF101 (UCF) [36], SomethingSomethingv2 (SSv2) [14], Kinetics400 (K400) [17], -600 (K600) [8], -700 (K700) [9], Moments In Time (MiT) [31] and six text-to-video retrieval datasets: MSRVTT (MSR) [48], YouCook2 (YC2) [12], DiDeMo (DDM) [1], LSMDC [33], ActivityNet (ANet) [6], and Spoken Moments in Time (S-MiT) [32].", + "bbox": [ + 89, + 719, + 482, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. For all train/test videos, we use $|v_{n}| = 8$ uniformly sampled frames. We use the LLaVA-1.6-7B-mistral model [26] as VLM, Mistral-7B-Instructv0.2 [16] as LLM, and SFR-Employing-Mistral [30] as text", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "embedding model. The temperature is set to 0.", + "bbox": [ + 511, + 237, + 821, + 252 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. UTD-descriptions", + "text_level": 1, + "bbox": [ + 511, + 266, + 687, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following our UTD pipeline, first, we generate textual descriptions for 12 considered datasets, resulting in UTD-descriptions dataset (see qualitative examples in the supplement). We conducted a user study to validate the quality of object, activity, and verb predictions. Participants were asked to verify which objects from a provided list were visible in a given image. The study used 100 random frames from 100 videos of the UCF dataset (which has noisy, low-resolution frames that might be an additional challenge to the VLM). Seven participants validated 761 annotated objects in total, where $87.6\\%$ (667 out of 761) were identified as visible, with only 94 not selected as visible. To better understand the VLM's errors, we manually classified these 94 objects into five categories: 1) attribute error (e.g., the object is \"right hand\" instead of \"left hand\"), 2) misclassification (the object is present but incorrectly identified as a different object, e.g., a \"snowboard\" instead of \"snow slide\"), 3) hallucination, 4) human annotation mistake – the object is visible, and 5) other. We found that only 45 objects were classified as hallucinations, 13 as misclassifications, 20 as \"object visible\" 8 as attribute errors, and 8 as other (see examples in the supplement). This results in a hallucination rate of only $6\\%$ , which we find acceptable for our analysis.", + "bbox": [ + 511, + 289, + 906, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Concept Representation Bias", + "text_level": 1, + "bbox": [ + 511, + 651, + 772, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Further, based on generated descriptions, we evaluate representation bias with respect to four concept categories, namely objects, activities, verbs, and objects+composition+activities, as described in Sec. 3.2. For this, we consider common sense bias in sequence-of-frames temporal setup in Tab. 1 (all concept-temporal combination results can be found in the supplement). It shows that using only object information results in only a slight drop in performance compared to objects+composition+activities, indicating that most datasets can, to a certain extent, be addressed with object information, thus showing an object representation bias. Interestingly, classification based on only objects results in 63.3pp accuracy for UCF101 and 45.9pp for Kinetics400. This may be due to the fact that some action classes are tied to the objects (e.g. \"playing", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29055", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/66fdd7ef658d5fa65d841de94e5476bd8ed6fbb7ed4ab1b4b9a79018d4924662.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
UCFSSv2K400K600K700MiT
obj.act.verbsobj.act.verbsobj.act.verbsobj.act.verbsobj.act.verbsobj.act.verbs
common sense b.63.367.450.85.36.45.845.945.224.841.841.421.437.036.717.621.021.016.2
dataset bias80.3+17.083.4+16.073.5+22.713.6+8.316.0+9.612.2+6.460.7+14.860.7+15.542.2+17.460.4+18.660.3+18.942.5+21.152.3+15.349.7+13.034.9+17.327.8+6.827.3+6.322.9+6.7
", + "bbox": [ + 93, + 89, + 903, + 143 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ce33ea7a5e0c211572ceb598acb70148d07e359227cf812b979a188491cd3411.jpg", + "table_caption": [ + "Table 2. Comparison of common sense bias vs dataset bias. We report the performance of our model in sequence of frame temporal setup and color code with respect to difference to common sense bias. It shows that the overall performance growth is significant for all concepts, including object concepts, indicating that datasets might contain a significant correlation between labels and visible objects." + ], + "table_footnote": [], + "table_body": "
UCF SSv2 K400 K600 K700 MiTMSR DDM ActN LSMDC YC2 S-MiT
middle-frame61.36.039.735.631.120.123.419.813.512.76.135.5
max-score frame66.57.448.043.338.721.433.429.121.518.38.946.1
avg.-over-frames66.76.846.642.137.123.431.926.526.616.48.643.8
seq.-of-frames66.36.448.044.139.022.636.627.226.617.08.445.9
", + "bbox": [ + 91, + 199, + 480, + 272 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f6ef2d9fc0ca893e6ac72ad2b20b8722a8a92129793c7a83f45aa1fe829695a0.jpg", + "table_caption": [ + "Table 3. Evaluation of temporal representation bias. We report the performance of our model in objects+composition+activities concepts setup and common sense bias setup." + ], + "table_footnote": [], + "table_body": "
UCF SSv2 K400 K600 K700 MiTMSR DDM ActN LSMDC YC2 S-MiT
Fleiss' kappa86.177.891.290.792.591.292.091.790.992.591.393.2
% object-biased samples7375555482428232011636
% of full test split in UTD-split279345455276727780899464
", + "bbox": [ + 91, + 327, + 480, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. UTD-splits statistics. Fleiss' kappa coefficient measures the agreement among 9 models used to create UTD-splits.", + "bbox": [ + 89, + 402, + 482, + 433 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "piano\"), making visible objects equally or more reliable indicators of a class than activities. This becomes clear when we consider the verbs-based representation. Here, the performance drops significantly for most action classification datasets, except for SSv2 and MiT. Also, in the case of text-to-video retrieval, it shows that objects are better indicators than activities for some datasets (MSRVTT, DiDeMo, ActivityNet), and activities are better indicators for others (YouCook2, LSMDC, S-MiT).", + "bbox": [ + 88, + 441, + 482, + 578 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Common Sense vs. Dataset Bias", + "text_level": 1, + "bbox": [ + 89, + 588, + 372, + 603 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further consider common sense bias compared to dataset bias for the action classification datasets in Tab. 2. Overall, across all benchmarks and concepts, dataset bias is significantly higher than common sense bias, and the results indicate that considered benchmarks also contain a dataset-specific object bias in the training set. The performance on UCF using only objects increases from 63.3pp to 80.3pp, and even on SSv2, which was developed to minimize object representation bias, using only correlation of seen objects in the train set to the classes, we were able to predict 13.6pp of videos correctly. This significant increase based on available object information only shows that some actions might be connected to unrelated objects that hint at a certain class without being directly associated with it.", + "bbox": [ + 88, + 609, + 482, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Temporal Representation Bias", + "text_level": 1, + "bbox": [ + 89, + 832, + 362, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, we consider representation bias with respect to the temporal aspect in Tab. 3. Namely, we consider which temporal representations, single frame vs. multiple frames,", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "best-score frame vs. middle frame, as well as average-over-frames vs. sequence-of-frames embeddings, perform best. For all datasets except ActivityNet, max-score-frame performance is comparable (< 3.5% difference) to multiple frames performance (average of frames/sequence over frames). Moreover, for SSv2, Kinetics400, DiDeMo, LSMDC, YouCook2, and S-MiT, we found that max-score-frame is a better indicator of class/captions than all frames.", + "bbox": [ + 511, + 200, + 906, + 323 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6.UTD-splits", + "text_level": 1, + "bbox": [ + 511, + 333, + 635, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since our analysis reveals significant object bias in benchmarks, we generate object-debiased UTD-splits for considered benchmarks, as described in Sec. 3.3.3. Namely, we remove object-biased samples where nine models that use different prompts in the text embedding model or trained on the bootstrapped train set uniformly agree that the sample can be predicted correctly based solely on objects. Tab. 4 summarizes the percentage of object-biased samples identified in the datasets, the level of model agreement (measured by Fleiss' kappa coefficient, which is notably high), and the number of samples retained from the original datasets.", + "bbox": [ + 511, + 356, + 906, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Benchmarking Video Models", + "text_level": 1, + "bbox": [ + 511, + 537, + 782, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we benchmark state-of-the-art video encoders on considered datasets and their UTD-splits.", + "bbox": [ + 511, + 564, + 905, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Setup", + "text_level": 1, + "bbox": [ + 511, + 604, + 596, + 621 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Video Models. We consider state-of-the-art video encoders with varying model sizes and pre-training configurations with open-sourced weights. For action classification, we benchmark 24 different variants of VideoMAE [38], VideoMAE2 [42], All-in-one [40], Unmasked Teacher (UMT) [21], VideoMamba [22], and InternVid [44] families. For text-to-video retrieval, we consider 17 videolanguage models of UMT [21], VideoMamba [22], and InternVid [44] families. In total, we analyze 30 models and list all considered variants in the supplement.", + "bbox": [ + 511, + 628, + 906, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. To evaluate action classification performance in a standard aligned setup, we follow the VideoGLUE [49] benchmark of using a frozen video backbone and lightweight pooling head as a test of the strength of video model representation. To adapt each video backbone for a specific dataset, we create a model that 1) uses frozen video backbone to extract spatial-temporal features, and then 2) uses a single attention pooling layer [49] to ag-", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29056", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ba7e118cb7a89b496d812f4989f36b4fb8354d7b4658f29cc2a4b71fd6c46cff.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
UCFUCF-UTD-splitUCF-UTD-s. balancedSSv2SSv2-UTD-splitSSv2-UTD-s. balancedK400K400-UTD-splitK400-UTD-s. balanced
videomae-B-K40089.281.3–7.982.8–6.454.252.1–2.153.0–1.264.345.7–18.649.5–14.8
videomae-B-UH88.479.2–9.280.4–8.056.254.0–2.254.9–1.364.345.4–18.949.6–14.7
videomae-L-UH95.790.9–4.891.0–4.767.365.4–1.966.2–1.175.859.3–16.563.3–12.5
videomae-H-UH95.290.9–4.391.0–4.267.465.5–1.966.3–1.176.559.5–17.063.7–12.8
videomaev2-B-K710-fnK71099.098.1–0.997.6–1.457.154.7–2.455.9–1.285.071.9–13.175.5–9.5
allinone-B-WV2M+CC84.563.1–21.472.9–11.626.222.5–3.724.7–1.566.942.6–24.350.8–16.1
allinone-B-WV2M+HT81.457.7–23.768.3–13.122.719.3–3.421.3–1.461.436.9–24.545.2–16.2
allinone-B-WV2M+HT+CC+YTT+81.358.5–22.866.2–15.122.018.8–3.220.7–1.361.236.4–24.844.6–16.6
umt-B-K71091.376.8–14.582.4–8.950.647.9–2.749.1–1.577.758.3–19.463.7–14.0
umt-B-K710-fnK71099.096.8–2.297.1–1.949.446.6–2.848.1–1.385.672.4–13.276.2–9.4
umt-L-K71095.787.2–8.589.2–6.559.457.0–2.458.1–1.384.068.4–15.673.0–11.0
umt-L-K710-fnK71098.996.7–2.297.1–1.857.955.5–2.456.7–1.289.078.7–10.381.6–7.4
videomamba-vm-K40090.974.6–16.381.0–9.935.932.6–3.334.4–1.575.354.4–20.960.5–14.8
videomamba-vm-5M93.782.0–11.786.7–7.048.545.6–2.946.9–1.677.558.1–19.464.0–13.5
videomamba-vm-17M93.280.5–12.784.3–8.947.444.7–2.745.9–1.577.758.1–19.663.6–14.1
videomamba-vm-25M94.383.0–11.386.5–7.848.745.9–2.847.1–1.678.459.1–19.364.6–13.8
internid-B-10M-FLT94.081.1–12.984.7–9.348.145.2–2.946.7–1.478.659.3–19.364.9–13.7
internid-B-200M94.582.9–11.685.2–9.354.451.8–2.652.8–1.680.061.7–18.366.8–13.2
internid-L-10M93.280.2–13.085.3–7.952.650.0–2.651.1–1.577.457.4–20.063.8–13.6
internid-L-WebVid10M94.182.5–11.686.3–7.854.952.4–2.553.4–1.578.959.7–19.265.7–13.2
internid-L-10M-DIV94.883.2–11.687.6–7.251.748.8–2.950.1–1.680.362.1–18.267.4–12.9
internid-L-10M-FLT95.586.3–9.288.4–7.153.650.9–2.752.1–1.581.964.7–17.269.6–12.3
internid-L-50M95.285.5–9.788.4–6.860.658.2–2.459.3–1.381.363.7–17.669.1–12.2
internid-L-200M96.890.5–6.391.9–4.963.161.0–2.161.9–1.284.669.4–15.273.9–10.7
", + "bbox": [ + 91, + 99, + 580, + 368 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2cd727989c63efb7cea568496a59c55df589e5a9e1feb9762638400b947c9264.jpg", + "table_caption": [ + "Table 5. Benchmarking video models in action classification. We report accuracy on full test/val sets and our debiased UTD-splits. The accuracy differences with respect to the full test/val sets are color-coded." + ], + "table_footnote": [], + "table_body": "
MSRMSR-UTD-splitDDMDDM-UTD-splitANetANet-UTD-split
umt-b-5M30.020.5–9.530.220.6–9.628.622.3–6.3
umt-b-17M35.626.3–9.337.727.6–10.134.227.3–6.9
umt-b-25M35.324.8–10.534.224.6–9.625.119.7–5.4
umt-l-5M34.824.9–9.933.521.8–11.734.828.7–6.1
umt-l-17M43.631.1–12.546.335.6–10.745.938.7–7.2
umt-l-25M42.330.6–11.743.633.5–10.136.730.4–6.3
videomamba-vm-5M33.323.0–10.337.127.1–10.037.130.1–7.0
videomamba-vm-17M34.925.5–9.440.628.9–11.740.433.0–7.4
videomamba-vm-25M34.925.5–9.441.430.5–10.941.133.8–7.3
internid-B-10M-FLT37.925.4–12.528.617.2–11.424.418.8–5.6
internid-B-200M38.124.7–13.430.219.4–10.826.220.1–6.1
internid-L-10M26.718.7–8.022.615.6–7.021.516.3–5.2
internid-L-WV10M26.517.6–8.922.213.8–8.423.117.5–5.6
internid-L-10M-DIV37.324.2–13.126.915.9–11.023.217.7–5.5
internid-L-10M-FLT38.726.3–12.429.218.8–10.424.519.0–5.5
internid-L-50M32.422.0–10.426.518.3–8.224.418.0–6.4
internid-L-200M38.224.8–13.430.320.2–10.128.722.1–6.6
", + "bbox": [ + 586, + 99, + 911, + 315 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Benchmarking video-language models in text-to-video retrieval. We report accuracy on full test/val sets and our debiased UTD-splits. The accuracy differences with respect to the full test/val sets are color-coded.", + "bbox": [ + 581, + 319, + 911, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ggregate spatial-temporal features into a single vector representation and 3) a single linear projection layer to classify on the required number of classes. We train the model on a full train split of the corresponding dataset. For video model evaluation in text-to-video retrieval, we reported their zero-shot performance on our debiased splits. We follow the corresponding model recipes in using re-ranking and how many examples are reranked. We report the accuracy for classification and the recall@1 for text-to-video retrieval. Additional details are provided in the supplement.", + "bbox": [ + 89, + 426, + 483, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Results", + "text_level": 1, + "bbox": [ + 89, + 589, + 184, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate the performance of video models for action classification on the UCF, SSv2, and K400 datasets and their UTD-splits in Tab. 5 and for text-to-video retrieval on the MSRVTT, DiDeMo, and ActivityNet datasets and their UTD-splits in Tab. 6. Performance for other datasets is reported in the supplement. Across all models, performance drops notably on debiased splits (except for SSv2), suggesting that current models strongly rely on object representation bias. Evaluation on UTD debiased splits shows a higher range of scores of models, revealing that some models (e.g. all-in-one models) were more heavily relying on object bias rather than utilizing action/motion cues for video understanding. Further, we can observe that the best performance flips for some models for the debiased case compared to the full test set, e.g., for videomae-B-K400 compared to videomamba-vm-K400 on the UCF dataset.", + "bbox": [ + 89, + 612, + 483, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additionally, by comparing the performance decline between the original datasets and the UTD-splits, we can assess models' sensitivity to object bias. Overall, we see in the", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "case of action classification, Tab. 5, that VideoMAE models are most robust to object bias, whereas Allinone models are most impacted by object bias. We also note that performance drops correlate between the UTD-split and UTD-split-balanced versions. We can further observe that larger backbones/architectures, and to a certain extent, more training data, lead to a reduced bias compared to smaller models, e.g. InternVid and VideoMAE. This pattern does not hold for text-to-video retrieval, shown in Tab. 6. Here, larger backbones and backbones trained with more data turn out to have higher object bias than smaller models or models trained with less data (see InternVid). We attribute this behavior, to the fact that retrieval models leverage information from both the vision as well as the text backbone. Thus, if larger text backbones also learn more object properties or a larger vocabulary of objects, this might further encourage object-centering learning in the visual stream as well.", + "bbox": [ + 511, + 426, + 906, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 694, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we presented a new way to examine representation biases in video classification and retrieval datasets by leveraging recent, powerful VLMs and LLMs. Our investigation confirms that most video datasets are heavily focused on object recognition. Based on our bias analysis, we propose a method to reduce object bias in these benchmarks, enabling more robust video understanding evaluation. With that, we are able to benchmark state-of-the-art video models and show an analysis of the object bias in these models. We hope that our findings, together with the presented benchmark, offer valuable insights for the future development of more robust video understanding benchmarks and models.", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29057", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Nina Shvetsova is supported in part by German Federal Ministry of Education and Research (BMBF) project STCL - 01IS22067. Nina Shvetsova also acknowledges travel support from ELISE (GA no 951847). Prof. Hilde Kuehne is supported in part by the ERC Starting Grant GraViLa 101117556.", + "bbox": [ + 89, + 114, + 483, + 204 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 217, + 187, + 233 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, 2017. 2, 6", + "[2] Piyush Bagad, Makarand Tapaswi, and Cees GM Snoek. Test of time: Instilling video-language models with a sense of time. In CVPR, pages 2503-2516, 2023. 1, 3", + "[3] Yizhak Ben-Shabat, Xin Yu, Fatemeh Saleh, Dylan Campbell, Cristian Rodriguez-Opazo, Hongdong Li, and Stephen Gould. The ika asm dataset: Understanding people assembling furniture through actions, objects and pose. 2021. 1", + "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 33:1877-1901, 2020. 5", + "[5] Shyamal Buch, Cristóbal Eyzaguirre, Adrien Gaidon, Jiajun Wu, Li Fei-Fei, and Juan Carlos Niebles. Revisiting the\" video\" in video-language understanding. In CVPR, pages 2917-2927, 2022. 1, 3", + "[6] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 2, 6", + "[7] Mu Cai, Reuben Tan, Jianrui Zhang, Bocheng Zou, Kai Zhang, Feng Yao, Fangrui Zhu, Jing Gu, Yiwu Zhong, Yuzhang Shang, et al. Temporalbench: Benchmarking fine-grained temporal understanding for multimodal video models. arXiv preprint arXiv:2410.10818, 2024. 1, 3", + "[8] João Carreira, Eric Noland, Andras Banki-Horvath, Chloe Hillier, and Andrew Zisserman. A short note about kinetics-600. arxiv, 1808.01340, 2018. 2, 6", + "[9] João Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. arxiv, 1907.06987, 2019. 2, 6", + "[10] Sihan Chen, Handong Li, Qunbo Wang, Zijia Zhao, Mingzhen Sun, Xinxin Zhu, and Jing Liu. Vast: A vision-audio-subtitle-text omni-modality foundation model and dataset. NeurIPS, 36, 2024. 3", + "[11] Jihoon Chung, Yu Wu, and Olga Russakovsky. Enabling detailed action recognition evaluation through video dataset augmentation. In NeurIPS, pages 39020-39033. Curran Associates, Inc., 2022. 1, 2, 3, 5", + "[12] Pradipto Das, Chenliang Xu, Richard F. Doell, and Jason J. Corso. A thousand frames in just a few words: Linguual description of videos through latent topics and sparse object stitching. In CVPR, 2013. 2, 6" + ], + "bbox": [ + 93, + 243, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Simone Fabbrizzi, Symeon Papadopoulos, Eirini Ntoutsi, and Ioannis Kompatsiaris. A survey on bias in visual datasets. Computer Vision and Image Understanding, 223, 2022. 2", + "[14] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, Florian Hoppe, Christian Thurau, Ingo Bax, and Roland Memisevic. The \"something something\" video database for learning and evaluating visual common sense. In ICCV, 2017. 1, 2, 3, 6", + "[15] Yun He, Soma Shirakabe, Yutaka Satoh, and Hirokatsu Kataoka. Human action recognition without human. 2016. 1, 2, 3, 5", + "[16] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 6", + "[17] Will Kay, João Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arxiv, 1705.06950, 2017. 2, 3, 6", + "[18] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In ICCV, 2011. 1", + "[19] Jie Lei, Tamara L Berg, and Mohit Bansal. Revealing single frame bias for video-and-language learning. In ACL, 2023. 1, 3", + "[20] Kunchang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 3", + "[21] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. In ICCV, pages 19948-19960, 2023. 2, 3, 7", + "[22] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. arXiv preprint arXiv:2403.06977, 2024. 2, 7", + "[23] Yingwei Li, Yi Li, and Nuno Vasconcelos. Resound: Towards action recognition without representation bias. In ECCV, 2018. 1, 3, 4", + "[24] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 5", + "[25] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023.", + "[26] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024.5,6", + "[27] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In ACL, 2024. 3" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29058", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. NeurIPS, 36:46212-46244, 2023. 1, 3", + "[29] Marcin Marszalek, Ivan Laptev, and Cordelia Schmid. Actions in context. In CVPR, 2009. 1", + "[30] Rui Meng, Ye Liu, Shafiq Rayhan Joty, Caiming Xiong, Yingbo Zhou, and Semih Yavuz. Sfrembedding-mistral: enhance text retrieval with transfer learning. *Salesforce AI Research Blog*, 3, 2024. 5, 6", + "[31] Mathew Monfort, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, Dan Gutfreund, Carl Vondrick, et al. Moments in time dataset: one million videos for event understanding. TPAMI, 42(2):502-508, 2019. 2, 6", + "[32] Mathew Monfort, SouYoung Jin, Alexander Liu, David Harwath, Rogerio Feris, James Glass, and Aude Oliva. Spoken moments: Learning joint audio-visual representations from video descriptions. In CVPR, 2021. 2, 6", + "[33] Anna Rohrbach, Marcus Rohrbach, Niket Tandon, and Bernt Schiele. A dataset for movie description. In CVPR, 2015. 2, 6", + "[34] Nina Shvetsova, Anna Kukleva, Xudong Hong, Christian Rupprecht, Bernt Schiele, and Hilde Kuehne. Howtocap-tion: Prompting llms to transform video annotations at scale. arXiv preprint arXiv:2310.04900, 2023. 3", + "[35] Nina Shvetsova, Anna Kukleva, Bernt Schiele, and Hilde Kuehne. In-style: Bridging text and uncurated videos with style transfer for text-video retrieval. In ICCV, 2023. 3", + "[36] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arxiv, 1705.06950, 2017. 1, 2, 6", + "[37] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In ICCV, pages 7464-7473, 2019. 3", + "[38] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. NeurIPS, 35:10078-10093, 2022. 2, 7", + "[39] Junke Wang, Dongdong Chen, Chong Luo, Xiyang Dai, Lu Yuan, Zuxuan Wu, and Yu-Gang Jiang. Chatvideo: A tracklet-centric multimodal and versatile video understanding system. arXiv preprint arXiv:2304.14407, 2023. 3", + "[40] Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Kevin Qinghong Lin, Satoshi Tsutsui, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, et al. All in one: Exploring unified video-language pre-training. In CVPR, pages 6598-6608, 2023. 2, 7", + "[41] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In ECCV, 2016. 4", + "[42] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In CVPR, pages 14549-14560, 2023. 2, 7" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. Improving text embeddings with large language models. arXiv preprint arXiv:2401.00368, 2023. 5", + "[44] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Intervid: A large-scale video-text dataset for multimodal understanding and generation. In ICLR, 2023. 2, 3, 7", + "[45] Zhenhailong Wang, Manling Li, Ruochen Xu, Luowei Zhou, Jie Lei, Xudong Lin, Shuohang Wang, Ziyi Yang, Chenguang Zhu, Derek Hoiem, et al. Language models with image descriptors are strong few-shot video-language learners. NeurIPS, 35:8483-8497, 2022. 3", + "[46] Zhanyu Wang, Longyue Wang, Zhen Zhao, Minghao Wu, Chenyang Lyu, Huayang Li, Deng Cai, Luping Zhou, Shuming Shi, and Zhaopeng Tu. Gpt4video: A unified multimodal large language model for Instruction-followed understanding and safety-aware generation. arXiv preprint arXiv:2311.16511, 2023. 3", + "[47] Philippe Weinzaepfel and Grégory Rogez. Mimetics: Towards understanding human actions out of context. *IJCV*, 129, 2021. 1, 3", + "[48] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. 2, 6", + "[49] Liangzhe Yuan, Nitesh Bharadwaj Gundavarapu, Long Zhao, Hao Zhou, Yin Cui, Lu Jiang, Xuan Yang, Menglin Jia, Tobias Weyand, Luke Friedman, et al. Videoglue: Video general understanding evaluation of foundation models. arXiv preprint arXiv:2307.03166, 2023. 7", + "[50] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, et al. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 3", + "[51] Yuanhao Zhai, Ziyi Liu, Zhenyu Wu, Yi Wu, Chunluan Zhou, David Doermann, Junsong Yuan, and Gang Hua. Soar: Scene-debiasing open-set action recognition. In ICCV, pages 10244–10254, 2023. 3", + "[52] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 3", + "[53] Long Zhao, Nitesh B Gundavarapu, Liangzhe Yuan, Hao Zhou, Shen Yan, Jennifer J Sun, Luke Friedman, Rui Qian, Tobias Weyand, Yue Zhao, et al. Videoprism: A foundational visual encoder for video understanding. arXiv preprint arXiv:2402.13217, 2024. 3" + ], + "bbox": [ + 516, + 92, + 903, + 767 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29059", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Unbiasing through Textual Descriptions_ Mitigating Representation Bias in Video Benchmarks/5edfcfce-ad23-41c6-9a43-d3cb9f8bcb90_model.json b/2025/Unbiasing through Textual Descriptions_ Mitigating Representation Bias in Video Benchmarks/5edfcfce-ad23-41c6-9a43-d3cb9f8bcb90_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d4533c8bad83db62f0c3cd8730e16e68b9715abd --- /dev/null +++ b/2025/Unbiasing through Textual Descriptions_ Mitigating Representation Bias in Video Benchmarks/5edfcfce-ad23-41c6-9a43-d3cb9f8bcb90_model.json @@ -0,0 +1,2156 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.23, + 0.131, + 0.77, + 0.177 + ], + "angle": 0, + "content": "Unbiasing through Textual Descriptions: Mitigating Representation Bias in Video Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.203, + 0.894, + 0.275 + ], + "angle": 0, + "content": "Nina Shvetsova\\(^{1,2,3*}\\), Arsha Nagrani\\(^{4}\\), Bernt Schiele\\(^{3}\\), Hilde Kuehne\\(^{1,2,5}\\), Christian Rupprecht\\(^{4}\\) \n\\(^{1}\\) Goethe University Frankfurt, \\(^{2}\\) Tuebingen AI Center/University of Tuebingen, \\(^{3}\\) MPI for Informatics, Saarland Informatics Campus, \\(^{4}\\) University of Oxford, \\(^{5}\\) MIT-IBM Watson AI Lab shvetsov@uni-frankfurt.de" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.31, + 0.327, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.342, + 0.485, + 0.766 + ], + "angle": 0, + "content": "We propose a new \"Unbiased through Textual Description (UTD)\" video benchmark based on unbiased subsets of existing video classification and retrieval datasets to enable a more robust assessment of video understanding capabilities. Namely, we tackle the problem that current video benchmarks may suffer from different representation biases, e.g., object bias or single-frame bias, where mere recognition of objects or utilization of only a single frame is sufficient for correct prediction. We leverage VLMs and LLMs to analyze and debias benchmarks from such representation biases. Specifically, we generate frame-wise textual descriptions of videos, filter them for specific information (e.g. only objects) and leverage them to examine representation biases across three dimensions: 1) concept bias—determining if a specific concept (e.g., objects) alone suffice for prediction; 2) temporal bias—assessing if temporal information contributes to prediction; and 3) common sense vs. dataset bias—evaluating whether zero-shot reasoning or dataset correlations contribute to prediction. We conduct a systematic analysis of 12 popular video classification and retrieval datasets and create new object-debiased test splits for these datasets. Moreover, we benchmark 30 state-of-the-art video models on original and debiased splits and analyze biases in the models. To facilitate the future development of more robust video understanding benchmarks and models, we release: \"UTD-descriptions\", a dataset with our rich structured descriptions for each dataset, and \"UTD-splits\", a dataset of object-debiased test splits.1" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.796, + 0.222, + 0.811 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.822, + 0.483, + 0.853 + ], + "angle": 0, + "content": "Ever since action recognition moved from simple, staged clip classification to real-life video understanding [18, 29," + }, + { + "type": "title", + "bbox": [ + 0.563, + 0.311, + 0.856, + 0.324 + ], + "angle": 0, + "content": "Guess the activity from the following objects:" + }, + { + "type": "text", + "bbox": [ + 0.589, + 0.329, + 0.636, + 0.339 + ], + "angle": 0, + "content": "Video 1:" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.343, + 0.7, + 0.366 + ], + "angle": 0, + "content": "Person, makeup brush, makeup sponge, makeup palette" + }, + { + "type": "text", + "bbox": [ + 0.544, + 0.369, + 0.591, + 0.379 + ], + "angle": 0, + "content": "Answer:" + }, + { + "type": "text", + "bbox": [ + 0.782, + 0.329, + 0.831, + 0.34 + ], + "angle": 0, + "content": "Video 2:" + }, + { + "type": "text", + "bbox": [ + 0.732, + 0.347, + 0.885, + 0.359 + ], + "angle": 0, + "content": "Person, stage, wall, curtain" + }, + { + "type": "text", + "bbox": [ + 0.737, + 0.369, + 0.786, + 0.378 + ], + "angle": 0, + "content": "Answer:" + }, + { + "type": "title", + "bbox": [ + 0.585, + 0.393, + 0.834, + 0.406 + ], + "angle": 0, + "content": "Guess the activity from a single frame:" + }, + { + "type": "text", + "bbox": [ + 0.589, + 0.407, + 0.636, + 0.417 + ], + "angle": 0, + "content": "Video 3:" + }, + { + "type": "image", + "bbox": [ + 0.544, + 0.419, + 0.684, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.493, + 0.593, + 0.502 + ], + "angle": 0, + "content": "Answer:" + }, + { + "type": "text", + "bbox": [ + 0.786, + 0.408, + 0.834, + 0.417 + ], + "angle": 0, + "content": "Video 4:" + }, + { + "type": "image", + "bbox": [ + 0.739, + 0.419, + 0.878, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.493, + 0.788, + 0.502 + ], + "angle": 0, + "content": "Answer:" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.514, + 0.907, + 0.542 + ], + "angle": 0, + "content": "Figure 1. Can you guess activity on a video based solely on the objects or a single frame? (The answer is on the next page in Fig. 2.)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.903 + ], + "angle": 0, + "content": "36], there has been an ongoing discussion about how much of a system's performance can be attributed to the recognition of the action itself compared to related or unrelated visual cues such as objects or scene information in the data [11, 15, 47], namely object/scene representation bias (see Fig. 1). Different strategies have been proposed to separate action recognition from scene and object influences such as segmenting the human figure in a video and placing it against a gray or black background [11, 15] as well as mimicking actions in unrelated contexts [47]. Some works focused on collecting datasets designed to minimize scene or object representation bias [3, 14, 23]. However, debiased solutions have so far rarely been adopted for benchmarking. Reasons for that may be that modified videos in case of [11, 15] often introduce artifacts and benchmarks with explicitly low scene bias [14, 23], might be considered out-of-domain relative to standard training setups. Another line of research questions whether current video benchmarks truly require temporal understanding, as single-frame models have shown comparable performance on several benchmarks [5, 19]. While new benchmarks are emerging to assess temporal understanding [2, 7, 28], temporal representen" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.863, + 0.483, + 0.901 + ], + "angle": 0, + "content": "*The work was done during PhD visit to the University of Oxford within the ELLIS PhD program. \n1Project page: https://utd-project.github.io/" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29050" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.482, + 0.106 + ], + "angle": 0, + "content": "tation bias in standard benchmarks remains underexplored." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.111, + 0.484, + 0.517 + ], + "angle": 0, + "content": "In this work, we propose a new scalable method, Unbiasing through Textual Descriptions (UTD), to systematically analyze and mitigate different types of representation bias, such as object bias, in existing video benchmarks through automatic textual descriptions. Since most datasets lack textual narrations of videos, we leverage Vision-Language Models (VLMs) to generate detailed descriptions of video frames. We then use Large Language Models (LLMs) to extract different concepts from those captions, namely objects, activities, and verbs (an overview of this process is shown in Fig. 3). Based on generated descriptions, we examine representation biases across three independent dimensions. First, concept representation bias — assessing if a specific concept (e.g., objects) alone is sufficient for accurate prediction of the video. Second, temporal representation bias — evaluating if temporal information is needed for correct prediction or if single-frame information is sufficient. Lastly, we distinguish between common sense and dataset bias. Common sense bias refers to the implicit association of certain concepts with specific activities, such as associating an object “a piano” with the label “playing piano.” This differs from dataset bias, where spurious correlations arise within the dataset, e.g., “apply makeup” label frequently occurs with objects like “mirror” and “flowers” in the background. We propose to leverage the zero-shot reasoning capabilities of LLMs to assess common sense bias and train a linear classifier for the dataset bias." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.484, + 0.793 + ], + "angle": 0, + "content": "In our study, we found that many samples of popular benchmarks have a significant object representation bias, meaning that recognizing only objects on the video is already enough to predict the correct label. Therefore, using such data for benchmarking video backbones may give models with strong object recognition ability an advantage (i.e. a shortcut), unintentionally hindering video understanding evaluation with respect to temporal reasoning or action recognition. Therefore, by excluding test samples with the highest object bias, we introduce UTD-splits, debiased evaluation splits for considered datasets, which we believe provide a more robust assessment of video understanding capabilities. Importantly, compared to other approaches that rely on modifying videos [11, 15], this approach neither requires any additional processing of evaluation data, as no new videos are generated nor does it require any fine-tuning to adjust for any domain bias that could be introduced by debiasing video frames at a pixel level." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We apply the proposed UTD method to analyze and debias six action classification datasets, namely UCF101 [36], SomethingSomethingv2 [14], Kinetics-400 [17], -600 [8], and -700 [9], and Moments In Time [31], and six text-to-video retrieval datasets, namely MSRVTT [48], DiDeMo [1], ActivityNet [6], LSMDC [33], YouCook2 [12], and Spoken Moments In Time [32]. Fur" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.706, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.09, + 0.897, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.19, + 0.706, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.19, + 0.902, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.301, + 0.907, + 0.357 + ], + "angle": 0, + "content": "Figure 2. The answer to Fig. 1. Some videos may exhibit an object representation bias (allowing predictions based solely on objects) or single-frame representation bias (solely on a single frame), while others require more information for prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.365, + 0.907, + 0.516 + ], + "angle": 0, + "content": "ther, we benchmark 30 state-of-the-art video models of six different model designs, namely VideoMAE [38], VideoMAE2 [42], All-in-one [40], Unmasked Teacher [21], VideoMamba [22], and InternVid [44], on the original and UTD-debiased splits. We show that debiasing reduces performance saturation on current benchmarks. Considering model performance drops between original and debiased splits also allows us to analyze models on object bias, e.g., in action classification, larger backbones or those trained on more data appear more robust to object bias." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.909, + 0.758 + ], + "angle": 0, + "content": "We summarize the contributions of this work as follows: (1) We introduce Unbiasing through Textual Descriptions (UTD), a scalable, automated method to analyze and mitigate representation biases in video benchmarks without the need for human annotators in the loop. (2) Using UTD, we analyze representation biases across 12 popular classification and retrieval video benchmarks and release UTD-debiased test splits for each, UTD-splits, promoting a more robust evaluation of video understanding. (3) We benchmark 30 state-of-the-art video models on both the original and UTD-debiased splits to assess their robustness against object bias under varying conditions. (4) Finally, we release UTD-descriptions, a dataset of detailed structured descriptions (including objects, activities, and verb categories) for all 12 considered datasets, to support further dataset analysis and the development of less biased benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.77, + 0.658, + 0.785 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Debiasing Video Datasets. While most visual datasets are prone to various biases [13], one unique effect in video understanding is the impact of object and scene information on the overall performance. Different strategies have so far been explored to mitigate this bias. He et al. [15] and HAT [11] proposed to separate the performance of scene appearance from human motion by segmenting human fig-" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29051" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.09, + 0.905, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.362, + 0.908, + 0.419 + ], + "angle": 0, + "content": "Figure 3. The proposed UTD method involves generating textual descriptions of different concepts in video frames using VLMs and LLMs, combining them in various temporal configurations, and evaluating the performance of these concept-temporal representations with strong text embedding models. For each representation, we distinguish between common sense bias, which relies on zero-shot reasoning by text embedding models, and dataset bias, assessed using a linear model trained on the dataset's training set." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.427, + 0.485, + 0.895 + ], + "angle": 0, + "content": "ures and placing them into black/gray background, or alternatively by cropping or inpainting human figures and classifying videos based on background only. Weinzaepfel et al. [47] proposed the Mimetics dataset by extending certain classes from the Kinetics dataset [17] with videos of activities in unrelated background. Zhai et al. [51] proposed an approach for scene-debiased action recognition via adversarial scene reconstruction and classification. Another line of work focused on capturing datasets with explicitly low object and scene bias. SomethingSomethingv1 and later v2 dataset [14] were proposed as an almost scene-free dataset, focusing on simple manipulative actions only. Li et al. [23] proposed a procedure for designing representation unbiased datasets and introduced Diving48 as an example. However, this specialization usually also implies that those cases are rather suitable for specialized approaches but not for, e.g., zero-shot evaluation of generalist video models. Other studies [5, 19] showed that for several benchmarks, using single frames during training or inference can achieve performance comparable to models that incorporate temporal information. However, recent efforts have mainly focused on developing new benchmarks to assess temporal understanding in videos [2], e.g., visual question-answering benchmarks [7, 28], while temporal representation bias in standard benchmarks has not been thoroughly investigated. In contrast, our UTD method allows for a systematic analysis of representation biases across various dimensions, including different concept and temporal configurations. Additionally, we are the first to propose mitigating object representation bias in popular benchmarks by discarding object-biased samples, which helps us to avoid adding artifacts to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.907, + 0.457 + ], + "angle": 0, + "content": "test samples (as with cropping or inpainting in [11, 15]) and keeps the dataset domain unconstrained, unlike [14, 23]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.909, + 0.683 + ], + "angle": 0, + "content": "LLMs for Video Understanding. The progress of LLMs prompted a wide range of applications in various fields, including video understanding. Here, one line of work utilizes pre-trained LLMs in video-language tasks, such as video captioning [37, 46, 53], video question answering [21, 27, 46, 53], text-video retrieval [35, 53], or dialog-based user interaction [46, 52]. Some works [10, 34, 44] create new video-language datasets using image/audio captioning models or subtitles and using LLM summarization and reasoning ability to post-process these captions. Other works [20, 39, 45, 50] use expert perception models to extract information from the video that is later processed by the LLM for tasks such as video captioning. In our work, we leverage VLM- and LLM-generated frame descriptions to evaluate different representation biases in video datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.696, + 0.828, + 0.713 + ], + "angle": 0, + "content": "3. Discovering Representation Biases" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.908, + 0.765 + ], + "angle": 0, + "content": "In this section, we define representation bias in Sec. 3.1, outline the considered representation biases in Sec. 3.2, and detail our UTD method in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.774, + 0.794, + 0.79 + ], + "angle": 0, + "content": "3.1. Definition: Representation Bias" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.825 + ], + "angle": 0, + "content": "To formalize the definition of representation bias in datasets, we adopt the notation from RESOUND [23]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Let \\(D = \\{v_{n}\\}_{n\\leq N}\\) be a video dataset \\(D\\) where each video \\(v_{n}\\) consists of a sequence of frames \\(v_{n} = \\{f_{n,i}\\}_{i\\leq |v_n|}\\). A representation \\(\\phi\\) is a function \\(\\phi (v)\\) of the video input \\(v\\). Let \\(R = \\{\\phi \\}\\) be a family of representations that share specific characteristics. For example, a fam" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29052" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.166 + ], + "angle": 0, + "content": "ily of single-frame representations includes representations that encode information from only a single frame of a video \\(\\phi (v_{n}) = \\phi (f_{n,i})\\) where \\(i\\leq |v_{n}|\\) or a family of object-based representations encodes only the objects in the video without considering motion or activity information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.483, + 0.229 + ], + "angle": 0, + "content": "The representation \\(\\phi\\) can be used to train different models \\(\\gamma_{\\phi}\\) for desired tasks, e.g. video classification. Let \\(M(v_{n},\\gamma_{\\phi})\\) be the performance (e.g., accuracy) of a model \\(\\gamma_{\\phi}\\) on a video \\(v_{n}\\), then the performance of \\(\\gamma_{\\phi}\\) on \\(D\\) is" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.234, + 0.483, + 0.274 + ], + "angle": 0, + "content": "\\[\nM (D, \\gamma_ {\\phi}) = \\frac {1}{N} \\sum_ {n = 1} ^ {N} M \\left(v _ {n}, \\gamma_ {\\phi}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.279, + 0.468, + 0.294 + ], + "angle": 0, + "content": "We define the performance of a representation \\(\\phi\\) on \\(D\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.301, + 0.483, + 0.323 + ], + "angle": 0, + "content": "\\[\nM (D, \\phi) = \\max _ {\\gamma_ {\\phi}} M (D, \\gamma_ {\\phi}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.329, + 0.483, + 0.358 + ], + "angle": 0, + "content": "where the maximum is taken over all models \\(\\gamma_{\\phi}\\) that are based on the representation \\(\\phi\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.359, + 0.483, + 0.646 + ], + "angle": 0, + "content": "When the performance \\( M(D, \\phi) \\) of a representation \\( \\phi \\) is high, the dataset exhibits a preference for this representation, indicating that the dataset is more biased toward this representation. In this work, we use \\( M(D, \\phi) \\) as a measure of bias towards representation \\( \\phi \\). However, a high bias toward a specific representation is not necessarily a negative thing. Ideally, a dataset should exhibit the highest bias toward the \"ground truth representation\": the representation truly required to solve the task. For instance, in video understanding, one could argue that the optimal representation should capture both static (e.g., objects) and dynamic (e.g., actions) properties, as well as encode temporal aspects. However, if a video dataset used to evaluate video understanding exhibits a strong object or single-frame representation bias, it may unintentionally favor models with strong object recognition or image understanding capabilities, thus hindering the evaluation of true video understanding. Therefore, in this work, we aim to estimate representation biases in popular video benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.653, + 0.37, + 0.669 + ], + "angle": 0, + "content": "3.2. Different Representation Biases" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.704 + ], + "angle": 0, + "content": "In our work, we consider different types of representation biases with respect to three independent aspects:" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.705, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Concept Representations: First, we consider representations that encode four different concept categories: objects \\(\\phi_{o}(v_{n})\\), activities \\(\\phi_{a}(v_{n})\\) (actions along with tied objects), verbs \\(\\phi_{\\nu}(v_{n})\\) (actions alone, without objects), and objects+composition+activities \\(\\phi_{d}(v_{n})\\) (objects, their spatial compositions in the video, and activities), see an example in Fig. 3. We aim to assess biases in classification and retrieval benchmarks toward particular concepts. For example, in retrieval, we ideally would like to see that benchmarks are not biased solely toward objects or activities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Temporal Representations: Second, we consider the bias of single-frame representations compared to representations that consider multiple frames. Let \\(\\phi_x\\) be" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.336 + ], + "angle": 0, + "content": "one of the objects \\(\\phi_o\\), activities \\(\\phi_a\\), verbs \\(\\phi_\\nu\\), or objects+composition+activities \\(\\phi_d\\) representations. Similarly to [41], we consider the following options to encode information from multiple frames: middle-frame refers to the case that a video is represented only by a single middle frame of this video, e.g., \\(\\phi_{x,\\mathrm{mid}}(v_n) = \\phi_x(f_{n,|v_n| / 2})\\), max-score frame refers to the case that a video is represented using a single frame \\(\\phi_{x,\\max}(v_n) = \\phi_x(f_{n,i})\\) where the frame \\(f_{n,i}\\) has the highest performance score \\(M(f_{n,i},\\phi_x)\\) over all frames, average-overframes refers to the case that the video is represented as an average representation of frames ignoring the order of frames \\(\\phi_{x,\\mathrm{avg}}(v_n) = \\frac{1}{|v_n|}\\sum_{i\\leq |v_n|}\\phi_x(f_{n,i})\\), and, finally, sequence-of-frames refers to the case that the video representation encodes concepts and their order in the frames \\(\\phi_{x,\\mathrm{seq}}(v_n) = \\phi_x(f_{n,1},f_{n,2},..,f_{n,|v_n|})\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.336, + 0.905, + 0.608 + ], + "angle": 0, + "content": "Common Sense Bias vs. Dataset Bias: Finally, for any concept-temporal representation \\(\\phi\\) (e.g., object middle-frame representation \\(\\phi_{o,mid}\\)), we can consider two types of representation biases: common sense bias or dataset bias. Common sense representation bias refers to bias based on \"common sense reasoning\", e.g., a video featuring an object \"piano\" would most likely be classified under the \"playing piano\" activity label. Formally, in Eq. (2), this bias \\(M(D,\\phi)\\) is measured by taking the maximum performance only over zero-shot models \\(\\gamma_{\\phi}\\) that have not seen the correlations in the training set of \\(D\\). We distinguish this from dataset bias, which refers to biases caused by spurious correlations within the dataset, e.g., where certain objects are unintentionally associated with specific labels, such as \"mirror\" or \"flower bouquet\" frequently appear with the \"apply makeup\" label. In this case, in Eq. (2), we measure dataset bias by taking the maximum performance over models \\(\\gamma_{\\phi}\\) trained or fine-tuned on the training set of \\(D\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.905, + 0.683 + ], + "angle": 0, + "content": "We note that the concept, temporal, and common sense vs. dataset bias dimensions are independent, allowing us to combine them in any configuration. For example, we can evaluate object middle-frame common sense bias or activity sequence-of-frames dataset bias." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.696, + 0.905, + 0.712 + ], + "angle": 0, + "content": "3.3. Unbiasing through Textual Descriptions (UTD)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.905, + 0.794 + ], + "angle": 0, + "content": "Since measuring bias for every representation within a family (e.g., a family of object representations) is infeasible, we approximate it by considering a single representative from the family [23] and propose a way to precisely control what each representation encodes by using textual descriptions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.901 + ], + "angle": 0, + "content": "The key idea of our method is to use textual descriptions as an intermediate video representation. This approach allows us to easily control which concepts in the video are included or excluded in the representation and to combine information into different temporal representations. Namely, to obtain representation \\(\\phi\\) that encodes only a specific concept category, we propose to encode visible concepts of this" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29053" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.257 + ], + "angle": 0, + "content": "category into textual descriptions, e.g., encode objects with a list of visible objects, e.g., \"table, lamp, ..\". This approach, compared to cropping or inpainting [11, 15], allows us to isolate specific concepts more precisely, avoiding information leakage from cropped areas or adding artifacts caused by imperfect inpainting. Additionally, textual descriptions allow us to use off-the-shelf LLMs with strong common sense reasoning abilities to estimate common sense bias. Since metadata with visible objects, activities, etc., is usually not available, we propose using recent powerful VLMs and LLMs to extract this information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.258, + 0.482, + 0.349 + ], + "angle": 0, + "content": "In the next sections, we first describe our method of obtaining textual description with respect to different concept and temporal categories (Sec. 3.3.1), then we describe our method of discovering common sense and dataset biases (Sec. 3.3.2), and lastly, we describe our method of unbiaseding benchmarks (Sec. 3.3.3)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.354, + 0.356, + 0.369 + ], + "angle": 0, + "content": "3.3.1. Obtaining Textual Descriptions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.373, + 0.483, + 0.493 + ], + "angle": 0, + "content": "While current VLMs [24-26], trained to perform complex image reasoning, are able to generate detailed descriptions of video frames, they still face challenges when prompted to directly extract specific concepts from a frame, e.g., only visible objects. Therefore, we decomposed the task of concept extraction into two parts: extraction of detailed frame description using VLMs and then extraction of different concepts from frame description using LLMs (see Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.494, + 0.483, + 0.705 + ], + "angle": 0, + "content": "Obtaining Textual Descriptions Using VLMs. Given train/test videos \\( v_{n} = \\{f_{n,i}\\}_{i\\leq |v_{n}|}, n < N \\) of the dataset, we feed each frame \\( f_{n,i} \\) of each video \\( v_{n} \\) into the VLM model \\( d \\) to generate a detailed textual description \\( d_{n,i} = d(f_{n_i}) \\) (Fig. 3). We found that without complex prompt engineering or few-shot in-context learning strategies [4] VLMs produce long, detailed descriptions. While our method is not restricted to a specific VLM, for reproducibility and feasibility, we select an open-sourced strong middle-scale VLM model, LLaVA-1.6-Mistral-7B [26] with the prompt: \"Describe the objects' relationships in the photo.\" Since the resulting textual descriptions include objects, their composition, relationships, and visible activities, we use them as objects+composition+activities descriptions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Extracting Concepts Categories Using LLMs. Each of the obtained detailed descriptions \\( d_{n,i} \\) is further fed to the LLM \\( \\zeta \\) to extract only objects \\( o_{n,i} = \\zeta(d_{n,i}, p_o) \\), only activities \\( a_{n,i} = \\zeta(d_{n,i}, p_a) \\) information, with \\( p_o, p_a \\) as respective prompts. Then, to obtain information about verbs, we task the LLM with a prompt \\( p_\\nu \\) to delete all object information from activities \\( \\nu_{n,i} = \\zeta(a_{n,i}, p_\\nu) \\). Since such text processing requires understanding what objects and activities are, our prompt includes a few examples following a few-shot in-context learning strategy [4]. We provide all prompts in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Aggregating Temporal Information. We use the obtained textual descriptions \\( x_{n,i} \\) of video \\( v_{n} \\) (where \\( x_{n,i} \\) is one" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.259 + ], + "angle": 0, + "content": "of \\( o_{n,i} \\) \\( a_{n,i} \\), \\( \\nu_{n,i} \\), or \\( d_{n,i} \\)) in different temporal setups: first, we keep them as single frame descriptions that might be further aggregated when transformed into vector representations (e.g. average-over-frames), as we describe in Sec. 3.3.2, second, we combine them into sequence-over-frame \\( x_{n} \\) setup with the template: \"Frame 1: \\( x_{n,1} \\) Frame 2: \\( x_{n,2} \\ldots \\) Frame \\( |v_{n}| \\colon x_{n,|v_{n}|} \\)\" (Fig. 3). Since objects+composition+activities descriptions \\( d_{n,i} \\) when combined with sequence-over-frame templates become too long, we summarize them to 15-word descriptions \\( d_{n,i}' \\) via an LLM before inputting them into the sequence prompt." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.266, + 0.798, + 0.281 + ], + "angle": 0, + "content": "3.3.2. Calculating Representation Biases" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.285, + 0.905, + 0.435 + ], + "angle": 0, + "content": "After obtaining a specific concept-temporal representation \\(\\phi\\), we assess common sense and dataset bias for dataset \\(D\\) toward \\(\\phi\\) by measuring the performance \\(M(D, \\phi)\\), as stated Eq. (2), where maximum is taken over zero-shot models \\(\\gamma_{\\phi}\\) for common sense bias and over fine-tuned models \\(\\gamma_{\\phi}\\) for dataset bias (see Sec. 3.2). Since maximizing over all models is infeasible in practice, we design strong models for common sense and dataset-based reasoning by utilizing state-of-the-art LLMs. Our ablation study verifies our design choice (see supplementary material)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.436, + 0.906, + 0.662 + ], + "angle": 0, + "content": "Common Sense Bias. We leverage recent strong LLM-based text embedding models [30, 43]. Such models are trained on large datasets to encode text into vector representations that can be used for various tasks, such as information retrieval. We encode our textual descriptions and all class labels/captions using the pre-trained text embedding model \\(\\chi\\). Specifically, we prompt the model with the relevant concept and task. For example, we use prompts like \"Given a list of objects visible on the video frame, retrieve the activity depicted in this video\" for object description and \"Given an activity, retrieve a video frame description that may depict this activity\" for class labels in action classification. We use cosine similarity between obtained embeddings for zero-shot classification/retrieval. Prompts are detailed in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.663, + 0.905, + 0.768 + ], + "angle": 0, + "content": "Dataset Bias. For action classification datasets, we evaluate dataset bias, as action classification models are often tested in fine-tuned setups and might exploit dataset representation biases learned from training data. For this we a linear logistic regression model trained on our strong text embeddings used in the common sense bias analysis and train model on the official train splits of the datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.769, + 0.905, + 0.814 + ], + "angle": 0, + "content": "Metrics. For the performance measure \\( M(D, \\phi) \\), we use accuracy for action classification and recall@1 and text-to-video retrieval as the most popular metrics for these tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.822, + 0.694, + 0.837 + ], + "angle": 0, + "content": "3.3.3. Unbiasing Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In Sec. 4, we find that many video datasets exhibit a significant object representation bias. To support robust video model evaluation, we construct UTD-splits, object-debiased test splits, by excluding object-biased samples from the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29054" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.912, + 0.158 + ], + "angle": 0, + "content": "
UCFSSv2K400K600K700MiTMSRDDMActNLSMDCYC2S-MiT
obj+comp+act66.36.448.044.139.022.636.627.226.617.08.445.9
objects63.3-3.05.3-1.145.9-2.141.8-2.337.0-2.021.0-1.632.1-4.527.0-0.224.8-1.813.6-3.47.9-0.529.8-16.1
activities67.4+1.16.445.2-2.841.4-2.736.7-2.321.0-1.625.1-11.521.1-6.121.4-5.214.7-2.38.1-0.341.1-4.8
verbs50.8-15.55.8-0.624.8-23.221.4-22.717.6-21.416.2-6.410.5-26.17.0-20.27.4-19.25.5-11.51.2-7.213.1-32.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.163, + 0.908, + 0.22 + ], + "angle": 0, + "content": "Table 1. Evaluation of concept representation bias. We report the performance of our model in sequence-of-frames temporal setup and common sense bias setup and color code with respect to the difference to objects+composition+activities (obj+comp+act) concepts. It shows that the overall classification performance drops only slightly when classifying actions based on objects, compared to classification and retrieval on verbs only, with SSv2 as the only dataset with less drop when classifying on verbs only." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.238, + 0.482, + 0.494 + ], + "angle": 0, + "content": "original test sets. Since our method allows us to measure the performance of the representation \\( M(v_{n},\\phi) \\) on individual samples \\( v_{n} \\), we exclude samples that are classified or retrieved correctly (\\( M(v_{n},\\phi) = 1 \\)) based on object sequence-of-frame representation. For text-to-video retrieval, we debias with respect to common-sense bias, but for action classification, we debias with respect to dataset bias since video models are mostly tested in fine-tuned setup. To minimize the impact of random fluctuations in the process, we use three different prompts for labels/captions in the text embedding models and train three linear models for each classification dataset using bootstrapped training sets (we provide details and analysis in the supplement) and exclude only samples when all nine models agreed on correct classification/retrieval. Note that the percentage of removed samples is determined automatically based on the extent of object representation bias." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.496, + 0.483, + 0.603 + ], + "angle": 0, + "content": "Finally, since debiasing may disproportionately remove certain label classes in classification datasets, we additionally construct balanced UTD splits. While maintaining the total number of removed samples as in the original debiasing method, we adjust the number of samples removed from each class based on their average confidence (across nine models) to preserve the original class proportions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.616, + 0.317, + 0.633 + ], + "angle": 0, + "content": "4. Benchmarking Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.483, + 0.687 + ], + "angle": 0, + "content": "In this section, we analyze action classification and text-to-video retrieval datasets on different representation biases using our UTD method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.697, + 0.173, + 0.714 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Datasets. We consider six common action classification datasets: UCF101 (UCF) [36], SomethingSomethingv2 (SSv2) [14], Kinetics400 (K400) [17], -600 (K600) [8], -700 (K700) [9], Moments In Time (MiT) [31] and six text-to-video retrieval datasets: MSRVTT (MSR) [48], YouCook2 (YC2) [12], DiDeMo (DDM) [1], LSMDC [33], ActivityNet (ANet) [6], and Spoken Moments in Time (S-MiT) [32]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. For all train/test videos, we use \\( |v_{n}| = 8 \\) uniformly sampled frames. We use the LLaVA-1.6-7B-mistral model [26] as VLM, Mistral-7B-Instructv0.2 [16] as LLM, and SFR-Employing-Mistral [30] as text" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.238, + 0.822, + 0.253 + ], + "angle": 0, + "content": "embedding model. The temperature is set to 0." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.267, + 0.688, + 0.284 + ], + "angle": 0, + "content": "4.2. UTD-descriptions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.29, + 0.907, + 0.637 + ], + "angle": 0, + "content": "Following our UTD pipeline, first, we generate textual descriptions for 12 considered datasets, resulting in UTD-descriptions dataset (see qualitative examples in the supplement). We conducted a user study to validate the quality of object, activity, and verb predictions. Participants were asked to verify which objects from a provided list were visible in a given image. The study used 100 random frames from 100 videos of the UCF dataset (which has noisy, low-resolution frames that might be an additional challenge to the VLM). Seven participants validated 761 annotated objects in total, where \\(87.6\\%\\) (667 out of 761) were identified as visible, with only 94 not selected as visible. To better understand the VLM's errors, we manually classified these 94 objects into five categories: 1) attribute error (e.g., the object is \"right hand\" instead of \"left hand\"), 2) misclassification (the object is present but incorrectly identified as a different object, e.g., a \"snowboard\" instead of \"snow slide\"), 3) hallucination, 4) human annotation mistake – the object is visible, and 5) other. We found that only 45 objects were classified as hallucinations, 13 as misclassifications, 20 as \"object visible\" 8 as attribute errors, and 8 as other (see examples in the supplement). This results in a hallucination rate of only \\(6\\%\\), which we find acceptable for our analysis." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.652, + 0.773, + 0.667 + ], + "angle": 0, + "content": "4.3. Concept Representation Bias" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Further, based on generated descriptions, we evaluate representation bias with respect to four concept categories, namely objects, activities, verbs, and objects+composition+activities, as described in Sec. 3.2. For this, we consider common sense bias in sequence-of-frames temporal setup in Tab. 1 (all concept-temporal combination results can be found in the supplement). It shows that using only object information results in only a slight drop in performance compared to objects+composition+activities, indicating that most datasets can, to a certain extent, be addressed with object information, thus showing an object representation bias. Interestingly, classification based on only objects results in 63.3pp accuracy for UCF101 and 45.9pp for Kinetics400. This may be due to the fact that some action classes are tied to the objects (e.g. \"playing" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29055" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.09, + 0.905, + 0.145 + ], + "angle": 0, + "content": "
UCFSSv2K400K600K700MiT
obj.act.verbsobj.act.verbsobj.act.verbsobj.act.verbsobj.act.verbsobj.act.verbs
common sense b.63.367.450.85.36.45.845.945.224.841.841.421.437.036.717.621.021.016.2
dataset bias80.3+17.083.4+16.073.5+22.713.6+8.316.0+9.612.2+6.460.7+14.860.7+15.542.2+17.460.4+18.660.3+18.942.5+21.152.3+15.349.7+13.034.9+17.327.8+6.827.3+6.322.9+6.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.151, + 0.907, + 0.194 + ], + "angle": 0, + "content": "Table 2. Comparison of common sense bias vs dataset bias. We report the performance of our model in sequence of frame temporal setup and color code with respect to difference to common sense bias. It shows that the overall performance growth is significant for all concepts, including object concepts, indicating that datasets might contain a significant correlation between labels and visible objects." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.2, + 0.482, + 0.273 + ], + "angle": 0, + "content": "
UCF SSv2 K400 K600 K700 MiTMSR DDM ActN LSMDC YC2 S-MiT
middle-frame61.36.039.735.631.120.123.419.813.512.76.135.5
max-score frame66.57.448.043.338.721.433.429.121.518.38.946.1
avg.-over-frames66.76.846.642.137.123.431.926.526.616.48.643.8
seq.-of-frames66.36.448.044.139.022.636.627.226.617.08.445.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.278, + 0.483, + 0.32 + ], + "angle": 0, + "content": "Table 3. Evaluation of temporal representation bias. We report the performance of our model in objects+composition+activities concepts setup and common sense bias setup." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.328, + 0.482, + 0.399 + ], + "angle": 0, + "content": "
UCF SSv2 K400 K600 K700 MiTMSR DDM ActN LSMDC YC2 S-MiT
Fleiss' kappa86.177.891.290.792.591.292.091.790.992.591.393.2
% object-biased samples7375555482428232011636
% of full test split in UTD-split279345455276727780899464
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.404, + 0.483, + 0.434 + ], + "angle": 0, + "content": "Table 4. UTD-splits statistics. Fleiss' kappa coefficient measures the agreement among 9 models used to create UTD-splits." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.442, + 0.483, + 0.579 + ], + "angle": 0, + "content": "piano\"), making visible objects equally or more reliable indicators of a class than activities. This becomes clear when we consider the verbs-based representation. Here, the performance drops significantly for most action classification datasets, except for SSv2 and MiT. Also, in the case of text-to-video retrieval, it shows that objects are better indicators than activities for some datasets (MSRVTT, DiDeMo, ActivityNet), and activities are better indicators for others (YouCook2, LSMDC, S-MiT)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.589, + 0.373, + 0.604 + ], + "angle": 0, + "content": "4.4. Common Sense vs. Dataset Bias" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.611, + 0.483, + 0.822 + ], + "angle": 0, + "content": "We further consider common sense bias compared to dataset bias for the action classification datasets in Tab. 2. Overall, across all benchmarks and concepts, dataset bias is significantly higher than common sense bias, and the results indicate that considered benchmarks also contain a dataset-specific object bias in the training set. The performance on UCF using only objects increases from 63.3pp to 80.3pp, and even on SSv2, which was developed to minimize object representation bias, using only correlation of seen objects in the train set to the classes, we were able to predict 13.6pp of videos correctly. This significant increase based on available object information only shows that some actions might be connected to unrelated objects that hint at a certain class without being directly associated with it." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.833, + 0.363, + 0.851 + ], + "angle": 0, + "content": "4.5. Temporal Representation Bias" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Finally, we consider representation bias with respect to the temporal aspect in Tab. 3. Namely, we consider which temporal representations, single frame vs. multiple frames," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.202, + 0.907, + 0.324 + ], + "angle": 0, + "content": "best-score frame vs. middle frame, as well as average-over-frames vs. sequence-of-frames embeddings, perform best. For all datasets except ActivityNet, max-score-frame performance is comparable (< 3.5% difference) to multiple frames performance (average of frames/sequence over frames). Moreover, for SSv2, Kinetics400, DiDeMo, LSMDC, YouCook2, and S-MiT, we found that max-score-frame is a better indicator of class/captions than all frames." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.334, + 0.637, + 0.35 + ], + "angle": 0, + "content": "4.6.UTD-splits" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.907, + 0.525 + ], + "angle": 0, + "content": "Since our analysis reveals significant object bias in benchmarks, we generate object-debiased UTD-splits for considered benchmarks, as described in Sec. 3.3.3. Namely, we remove object-biased samples where nine models that use different prompts in the text embedding model or trained on the bootstrapped train set uniformly agree that the sample can be predicted correctly based solely on objects. Tab. 4 summarizes the percentage of object-biased samples identified in the datasets, the level of model agreement (measured by Fleiss' kappa coefficient, which is notably high), and the number of samples retained from the original datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.538, + 0.783, + 0.556 + ], + "angle": 0, + "content": "5. Benchmarking Video Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.906, + 0.595 + ], + "angle": 0, + "content": "In this section, we benchmark state-of-the-art video encoders on considered datasets and their UTD-splits." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.606, + 0.597, + 0.622 + ], + "angle": 0, + "content": "5.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.78 + ], + "angle": 0, + "content": "Video Models. We consider state-of-the-art video encoders with varying model sizes and pre-training configurations with open-sourced weights. For action classification, we benchmark 24 different variants of VideoMAE [38], VideoMAE2 [42], All-in-one [40], Unmasked Teacher (UMT) [21], VideoMamba [22], and InternVid [44] families. For text-to-video retrieval, we consider 17 videolanguage models of UMT [21], VideoMamba [22], and InternVid [44] families. In total, we analyze 30 models and list all considered variants in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Implementation Details. To evaluate action classification performance in a standard aligned setup, we follow the VideoGLUE [49] benchmark of using a frozen video backbone and lightweight pooling head as a test of the strength of video model representation. To adapt each video backbone for a specific dataset, we create a model that 1) uses frozen video backbone to extract spatial-temporal features, and then 2) uses a single attention pooling layer [49] to ag-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29056" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.1, + 0.581, + 0.369 + ], + "angle": 0, + "content": "
UCFUCF-UTD-splitUCF-UTD-s. balancedSSv2SSv2-UTD-splitSSv2-UTD-s. balancedK400K400-UTD-splitK400-UTD-s. balanced
videomae-B-K40089.281.3–7.982.8–6.454.252.1–2.153.0–1.264.345.7–18.649.5–14.8
videomae-B-UH88.479.2–9.280.4–8.056.254.0–2.254.9–1.364.345.4–18.949.6–14.7
videomae-L-UH95.790.9–4.891.0–4.767.365.4–1.966.2–1.175.859.3–16.563.3–12.5
videomae-H-UH95.290.9–4.391.0–4.267.465.5–1.966.3–1.176.559.5–17.063.7–12.8
videomaev2-B-K710-fnK71099.098.1–0.997.6–1.457.154.7–2.455.9–1.285.071.9–13.175.5–9.5
allinone-B-WV2M+CC84.563.1–21.472.9–11.626.222.5–3.724.7–1.566.942.6–24.350.8–16.1
allinone-B-WV2M+HT81.457.7–23.768.3–13.122.719.3–3.421.3–1.461.436.9–24.545.2–16.2
allinone-B-WV2M+HT+CC+YTT+81.358.5–22.866.2–15.122.018.8–3.220.7–1.361.236.4–24.844.6–16.6
umt-B-K71091.376.8–14.582.4–8.950.647.9–2.749.1–1.577.758.3–19.463.7–14.0
umt-B-K710-fnK71099.096.8–2.297.1–1.949.446.6–2.848.1–1.385.672.4–13.276.2–9.4
umt-L-K71095.787.2–8.589.2–6.559.457.0–2.458.1–1.384.068.4–15.673.0–11.0
umt-L-K710-fnK71098.996.7–2.297.1–1.857.955.5–2.456.7–1.289.078.7–10.381.6–7.4
videomamba-vm-K40090.974.6–16.381.0–9.935.932.6–3.334.4–1.575.354.4–20.960.5–14.8
videomamba-vm-5M93.782.0–11.786.7–7.048.545.6–2.946.9–1.677.558.1–19.464.0–13.5
videomamba-vm-17M93.280.5–12.784.3–8.947.444.7–2.745.9–1.577.758.1–19.663.6–14.1
videomamba-vm-25M94.383.0–11.386.5–7.848.745.9–2.847.1–1.678.459.1–19.364.6–13.8
internid-B-10M-FLT94.081.1–12.984.7–9.348.145.2–2.946.7–1.478.659.3–19.364.9–13.7
internid-B-200M94.582.9–11.685.2–9.354.451.8–2.652.8–1.680.061.7–18.366.8–13.2
internid-L-10M93.280.2–13.085.3–7.952.650.0–2.651.1–1.577.457.4–20.063.8–13.6
internid-L-WebVid10M94.182.5–11.686.3–7.854.952.4–2.553.4–1.578.959.7–19.265.7–13.2
internid-L-10M-DIV94.883.2–11.687.6–7.251.748.8–2.950.1–1.680.362.1–18.267.4–12.9
internid-L-10M-FLT95.586.3–9.288.4–7.153.650.9–2.752.1–1.581.964.7–17.269.6–12.3
internid-L-50M95.285.5–9.788.4–6.860.658.2–2.459.3–1.381.363.7–17.669.1–12.2
internid-L-200M96.890.5–6.391.9–4.963.161.0–2.161.9–1.284.669.4–15.273.9–10.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.374, + 0.584, + 0.417 + ], + "angle": 0, + "content": "Table 5. Benchmarking video models in action classification. We report accuracy on full test/val sets and our debiased UTD-splits. The accuracy differences with respect to the full test/val sets are color-coded." + }, + { + "type": "table", + "bbox": [ + 0.587, + 0.1, + 0.912, + 0.316 + ], + "angle": 0, + "content": "
MSRMSR-UTD-splitDDMDDM-UTD-splitANetANet-UTD-split
umt-b-5M30.020.5–9.530.220.6–9.628.622.3–6.3
umt-b-17M35.626.3–9.337.727.6–10.134.227.3–6.9
umt-b-25M35.324.8–10.534.224.6–9.625.119.7–5.4
umt-l-5M34.824.9–9.933.521.8–11.734.828.7–6.1
umt-l-17M43.631.1–12.546.335.6–10.745.938.7–7.2
umt-l-25M42.330.6–11.743.633.5–10.136.730.4–6.3
videomamba-vm-5M33.323.0–10.337.127.1–10.037.130.1–7.0
videomamba-vm-17M34.925.5–9.440.628.9–11.740.433.0–7.4
videomamba-vm-25M34.925.5–9.441.430.5–10.941.133.8–7.3
internid-B-10M-FLT37.925.4–12.528.617.2–11.424.418.8–5.6
internid-B-200M38.124.7–13.430.219.4–10.826.220.1–6.1
internid-L-10M26.718.7–8.022.615.6–7.021.516.3–5.2
internid-L-WV10M26.517.6–8.922.213.8–8.423.117.5–5.6
internid-L-10M-DIV37.324.2–13.126.915.9–11.023.217.7–5.5
internid-L-10M-FLT38.726.3–12.429.218.8–10.424.519.0–5.5
internid-L-50M32.422.0–10.426.518.3–8.224.418.0–6.4
internid-L-200M38.224.8–13.430.320.2–10.128.722.1–6.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.582, + 0.32, + 0.913, + 0.389 + ], + "angle": 0, + "content": "Table 6. Benchmarking video-language models in text-to-video retrieval. We report accuracy on full test/val sets and our debiased UTD-splits. The accuracy differences with respect to the full test/val sets are color-coded." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.484, + 0.579 + ], + "angle": 0, + "content": "ggregate spatial-temporal features into a single vector representation and 3) a single linear projection layer to classify on the required number of classes. We train the model on a full train split of the corresponding dataset. For video model evaluation in text-to-video retrieval, we reported their zero-shot performance on our debiased splits. We follow the corresponding model recipes in using re-ranking and how many examples are reranked. We report the accuracy for classification and the recall@1 for text-to-video retrieval. Additional details are provided in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.59, + 0.185, + 0.605 + ], + "angle": 0, + "content": "5.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.484, + 0.854 + ], + "angle": 0, + "content": "We evaluate the performance of video models for action classification on the UCF, SSv2, and K400 datasets and their UTD-splits in Tab. 5 and for text-to-video retrieval on the MSRVTT, DiDeMo, and ActivityNet datasets and their UTD-splits in Tab. 6. Performance for other datasets is reported in the supplement. Across all models, performance drops notably on debiased splits (except for SSv2), suggesting that current models strongly rely on object representation bias. Evaluation on UTD debiased splits shows a higher range of scores of models, revealing that some models (e.g. all-in-one models) were more heavily relying on object bias rather than utilizing action/motion cues for video understanding. Further, we can observe that the best performance flips for some models for the debiased case compared to the full test set, e.g., for videomae-B-K400 compared to videomamba-vm-K400 on the UCF dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Additionally, by comparing the performance decline between the original datasets and the UTD-splits, we can assess models' sensitivity to object bias. Overall, we see in the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.908, + 0.683 + ], + "angle": 0, + "content": "case of action classification, Tab. 5, that VideoMAE models are most robust to object bias, whereas Allinone models are most impacted by object bias. We also note that performance drops correlate between the UTD-split and UTD-split-balanced versions. We can further observe that larger backbones/architectures, and to a certain extent, more training data, lead to a reduced bias compared to smaller models, e.g. InternVid and VideoMAE. This pattern does not hold for text-to-video retrieval, shown in Tab. 6. Here, larger backbones and backbones trained with more data turn out to have higher object bias than smaller models or models trained with less data (see InternVid). We attribute this behavior, to the fact that retrieval models leverage information from both the vision as well as the text backbone. Thus, if larger text backbones also learn more object properties or a larger vocabulary of objects, this might further encourage object-centering learning in the visual stream as well." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.634, + 0.71 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this paper, we presented a new way to examine representation biases in video classification and retrieval datasets by leveraging recent, powerful VLMs and LLMs. Our investigation confirms that most video datasets are heavily focused on object recognition. Based on our bias analysis, we propose a method to reduce object bias in these benchmarks, enabling more robust video understanding evaluation. With that, we are able to benchmark state-of-the-art video models and show an analysis of the object bias in these models. We hope that our findings, together with the presented benchmark, offer valuable insights for the future development of more robust video understanding benchmarks and models." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29057" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.205 + ], + "angle": 0, + "content": "Nina Shvetsova is supported in part by German Federal Ministry of Education and Research (BMBF) project STCL - 01IS22067. Nina Shvetsova also acknowledges travel support from ELISE (GA no 951847). Prof. Hilde Kuehne is supported in part by the ERC Starting Grant GraViLa 101117556." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.218, + 0.188, + 0.234 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.244, + 0.484, + 0.298 + ], + "angle": 0, + "content": "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.3, + 0.484, + 0.341 + ], + "angle": 0, + "content": "[2] Piyush Bagad, Makarand Tapaswi, and Cees GM Snoek. Test of time: Instilling video-language models with a sense of time. In CVPR, pages 2503-2516, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.342, + 0.483, + 0.397 + ], + "angle": 0, + "content": "[3] Yizhak Ben-Shabat, Xin Yu, Fatemeh Saleh, Dylan Campbell, Cristian Rodriguez-Opazo, Hongdong Li, and Stephen Gould. The ika asm dataset: Understanding people assembling furniture through actions, objects and pose. 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.398, + 0.483, + 0.466 + ], + "angle": 0, + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 33:1877-1901, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.468, + 0.483, + 0.523 + ], + "angle": 0, + "content": "[5] Shyamal Buch, Cristóbal Eyzaguirre, Adrien Gaidon, Jiajun Wu, Li Fei-Fei, and Juan Carlos Niebles. Revisiting the\" video\" in video-language understanding. In CVPR, pages 2917-2927, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.524, + 0.483, + 0.579 + ], + "angle": 0, + "content": "[6] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.483, + 0.648 + ], + "angle": 0, + "content": "[7] Mu Cai, Reuben Tan, Jianrui Zhang, Bocheng Zou, Kai Zhang, Feng Yao, Fangrui Zhu, Jing Gu, Yiwu Zhong, Yuzhang Shang, et al. Temporalbench: Benchmarking fine-grained temporal understanding for multimodal video models. arXiv preprint arXiv:2410.10818, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.483, + 0.69 + ], + "angle": 0, + "content": "[8] João Carreira, Eric Noland, Andras Banki-Horvath, Chloe Hillier, and Andrew Zisserman. A short note about kinetics-600. arxiv, 1808.01340, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.692, + 0.483, + 0.734 + ], + "angle": 0, + "content": "[9] João Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. arxiv, 1907.06987, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.735, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[10] Sihan Chen, Handong Li, Qunbo Wang, Zijia Zhao, Mingzhen Sun, Xinxin Zhu, and Jing Liu. Vast: A vision-audio-subtitle-text omni-modality foundation model and dataset. NeurIPS, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[11] Jihoon Chung, Yu Wu, and Olga Russakovsky. Enabling detailed action recognition evaluation through video dataset augmentation. In NeurIPS, pages 39020-39033. Curran Associates, Inc., 2022. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Pradipto Das, Chenliang Xu, Richard F. Doell, and Jason J. Corso. A thousand frames in just a few words: Linguual description of videos through latent topics and sparse object stitching. In CVPR, 2013. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.244, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[13] Simone Fabbrizzi, Symeon Papadopoulos, Eirini Ntoutsi, and Ioannis Kompatsiaris. A survey on bias in visual datasets. Computer Vision and Image Understanding, 223, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.246 + ], + "angle": 0, + "content": "[14] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, Florian Hoppe, Christian Thurau, Ingo Bax, and Roland Memisevic. The \"something something\" video database for learning and evaluating visual common sense. In ICCV, 2017. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[15] Yun He, Soma Shirakabe, Yutaka Satoh, and Hirokatsu Kataoka. Human action recognition without human. 2016. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.294, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[16] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.366, + 0.905, + 0.435 + ], + "angle": 0, + "content": "[17] Will Kay, João Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arxiv, 1705.06950, 2017. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.437, + 0.905, + 0.478 + ], + "angle": 0, + "content": "[18] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In ICCV, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.481, + 0.905, + 0.521 + ], + "angle": 0, + "content": "[19] Jie Lei, Tamara L Berg, and Mohit Bansal. Revealing single frame bias for video-and-language learning. In ACL, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[20] Kunchang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.582, + 0.905, + 0.637 + ], + "angle": 0, + "content": "[21] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. In ICCV, pages 19948-19960, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.64, + 0.905, + 0.694 + ], + "angle": 0, + "content": "[22] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. arXiv preprint arXiv:2403.06977, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.698, + 0.905, + 0.739 + ], + "angle": 0, + "content": "[23] Yingwei Li, Yi Li, and Nuno Vasconcelos. Resound: Towards action recognition without representation bias. In ECCV, 2018. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.742, + 0.904, + 0.769 + ], + "angle": 0, + "content": "[24] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.772, + 0.904, + 0.799 + ], + "angle": 0, + "content": "[25] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[26] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024.5,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[27] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In ACL, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29058" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[28] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. NeurIPS, 36:46212-46244, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.175 + ], + "angle": 0, + "content": "[29] Marcin Marszalek, Ivan Laptev, and Cordelia Schmid. Actions in context. In CVPR, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[30] Rui Meng, Ye Liu, Shafiq Rayhan Joty, Caiming Xiong, Yingbo Zhou, and Semih Yavuz. Sfrembedding-mistral: enhance text retrieval with transfer learning. *Salesforce AI Research Blog*, 3, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[31] Mathew Monfort, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, Dan Gutfreund, Carl Vondrick, et al. Moments in time dataset: one million videos for event understanding. TPAMI, 42(2):502-508, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[32] Mathew Monfort, SouYoung Jin, Alexander Liu, David Harwath, Rogerio Feris, James Glass, and Aude Oliva. Spoken moments: Learning joint audio-visual representations from video descriptions. In CVPR, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[33] Anna Rohrbach, Marcus Rohrbach, Niket Tandon, and Bernt Schiele. A dataset for movie description. In CVPR, 2015. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[34] Nina Shvetsova, Anna Kukleva, Xudong Hong, Christian Rupprecht, Bernt Schiele, and Hilde Kuehne. Howtocap-tion: Prompting llms to transform video annotations at scale. arXiv preprint arXiv:2310.04900, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[35] Nina Shvetsova, Anna Kukleva, Bernt Schiele, and Hilde Kuehne. In-style: Bridging text and uncurated videos with style transfer for text-video retrieval. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.506, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[36] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arxiv, 1705.06950, 2017. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.549, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[37] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In ICCV, pages 7464-7473, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.606, + 0.482, + 0.658 + ], + "angle": 0, + "content": "[38] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. NeurIPS, 35:10078-10093, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.662, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[39] Junke Wang, Dongdong Chen, Chong Luo, Xiyang Dai, Lu Yuan, Zuxuan Wu, and Yu-Gang Jiang. Chatvideo: A tracklet-centric multimodal and versatile video understanding system. arXiv preprint arXiv:2304.14407, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.719, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[40] Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Kevin Qinghong Lin, Satoshi Tsutsui, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, et al. All in one: Exploring unified video-language pre-training. In CVPR, pages 6598-6608, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[41] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In ECCV, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[42] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In CVPR, pages 14549-14560, 2023. 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[43] Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. Improving text embeddings with large language models. arXiv preprint arXiv:2401.00368, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[44] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Intervid: A large-scale video-text dataset for multimodal understanding and generation. In ICLR, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[45] Zhenhailong Wang, Manling Li, Ruochen Xu, Luowei Zhou, Jie Lei, Xudong Lin, Shuohang Wang, Ziyi Yang, Chenguang Zhu, Derek Hoiem, et al. Language models with image descriptors are strong few-shot video-language learners. NeurIPS, 35:8483-8497, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[46] Zhanyu Wang, Longyue Wang, Zhen Zhao, Minghao Wu, Chenyang Lyu, Huayang Li, Deng Cai, Luping Zhou, Shuming Shi, and Zhaopeng Tu. Gpt4video: A unified multimodal large language model for Instruction-followed understanding and safety-aware generation. arXiv preprint arXiv:2311.16511, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.905, + 0.415 + ], + "angle": 0, + "content": "[47] Philippe Weinzaepfel and Grégory Rogez. Mimetics: Towards understanding human actions out of context. *IJCV*, 129, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.458 + ], + "angle": 0, + "content": "[48] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.529 + ], + "angle": 0, + "content": "[49] Liangzhe Yuan, Nitesh Bharadwaj Gundavarapu, Long Zhao, Hao Zhou, Yin Cui, Lu Jiang, Xuan Yang, Menglin Jia, Tobias Weyand, Luke Friedman, et al. Videoglue: Video general understanding evaluation of foundation models. arXiv preprint arXiv:2307.03166, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.531, + 0.905, + 0.599 + ], + "angle": 0, + "content": "[50] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, et al. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.602, + 0.905, + 0.655 + ], + "angle": 0, + "content": "[51] Yuanhao Zhai, Ziyi Liu, Zhenyu Wu, Yi Wu, Chunluan Zhou, David Doermann, Junsong Yuan, and Gang Hua. Soar: Scene-debiasing open-set action recognition. In ICCV, pages 10244–10254, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.658, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[52] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.701, + 0.905, + 0.768 + ], + "angle": 0, + "content": "[53] Long Zhao, Nitesh B Gundavarapu, Liangzhe Yuan, Hao Zhou, Shen Yan, Jennifer J Sun, Luke Friedman, Rui Qian, Tobias Weyand, Yue Zhao, et al. Videoprism: A foundational visual encoder for video understanding. arXiv preprint arXiv:2402.13217, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.768 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "29059" + } + ] +] \ No newline at end of file diff --git a/2025/Unbiasing through Textual Descriptions_ Mitigating Representation Bias in Video Benchmarks/5edfcfce-ad23-41c6-9a43-d3cb9f8bcb90_origin.pdf b/2025/Unbiasing through Textual Descriptions_ Mitigating Representation Bias in Video Benchmarks/5edfcfce-ad23-41c6-9a43-d3cb9f8bcb90_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5ef22534382fa6727914d5fd50d65c3077472a72 --- /dev/null +++ b/2025/Unbiasing through Textual Descriptions_ Mitigating Representation Bias in Video Benchmarks/5edfcfce-ad23-41c6-9a43-d3cb9f8bcb90_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:602e57c51981bc3e3eea55b329392d1a0e44ae8c978fe1d349509fb3cf9e44df +size 6605857 diff --git a/2025/Unbiasing through Textual Descriptions_ Mitigating Representation Bias in Video Benchmarks/full.md b/2025/Unbiasing through Textual Descriptions_ Mitigating Representation Bias in Video Benchmarks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4a34323d91990eb34f5f8a021f273b2802c567ab --- /dev/null +++ b/2025/Unbiasing through Textual Descriptions_ Mitigating Representation Bias in Video Benchmarks/full.md @@ -0,0 +1,303 @@ +# Unbiasing through Textual Descriptions: Mitigating Representation Bias in Video Benchmarks + +Nina Shvetsova $^{1,2,3*}$ , Arsha Nagrani $^{4}$ , Bernt Schiele $^{3}$ , Hilde Kuehne $^{1,2,5}$ , Christian Rupprecht $^{4}$ $^{1}$ Goethe University Frankfurt, $^{2}$ Tuebingen AI Center/University of Tuebingen, $^{3}$ MPI for Informatics, Saarland Informatics Campus, $^{4}$ University of Oxford, $^{5}$ MIT-IBM Watson AI Lab shvetsov@uni-frankfurt.de + +# Abstract + +We propose a new "Unbiased through Textual Description (UTD)" video benchmark based on unbiased subsets of existing video classification and retrieval datasets to enable a more robust assessment of video understanding capabilities. Namely, we tackle the problem that current video benchmarks may suffer from different representation biases, e.g., object bias or single-frame bias, where mere recognition of objects or utilization of only a single frame is sufficient for correct prediction. We leverage VLMs and LLMs to analyze and debias benchmarks from such representation biases. Specifically, we generate frame-wise textual descriptions of videos, filter them for specific information (e.g. only objects) and leverage them to examine representation biases across three dimensions: 1) concept bias—determining if a specific concept (e.g., objects) alone suffice for prediction; 2) temporal bias—assessing if temporal information contributes to prediction; and 3) common sense vs. dataset bias—evaluating whether zero-shot reasoning or dataset correlations contribute to prediction. We conduct a systematic analysis of 12 popular video classification and retrieval datasets and create new object-debiased test splits for these datasets. Moreover, we benchmark 30 state-of-the-art video models on original and debiased splits and analyze biases in the models. To facilitate the future development of more robust video understanding benchmarks and models, we release: "UTD-descriptions", a dataset with our rich structured descriptions for each dataset, and "UTD-splits", a dataset of object-debiased test splits.1 + +# 1. Introduction + +Ever since action recognition moved from simple, staged clip classification to real-life video understanding [18, 29, + +# Guess the activity from the following objects: + +Video 1: + +Person, makeup brush, makeup sponge, makeup palette + +Answer: + +Video 2: + +Person, stage, wall, curtain + +Answer: + +# Guess the activity from a single frame: + +Video 3: + +![](images/7cca5a13202945120aa3cf187b0849c1491a2df52b3caf83675a1e8cba9d6439.jpg) +Answer: + +Video 4: + +![](images/1f2f04ebfc6cb72bbb6f1fef87983da65e430507e89d9a0b00e3a589d3877ca5.jpg) +Answer: +Figure 1. Can you guess activity on a video based solely on the objects or a single frame? (The answer is on the next page in Fig. 2.) + +36], there has been an ongoing discussion about how much of a system's performance can be attributed to the recognition of the action itself compared to related or unrelated visual cues such as objects or scene information in the data [11, 15, 47], namely object/scene representation bias (see Fig. 1). Different strategies have been proposed to separate action recognition from scene and object influences such as segmenting the human figure in a video and placing it against a gray or black background [11, 15] as well as mimicking actions in unrelated contexts [47]. Some works focused on collecting datasets designed to minimize scene or object representation bias [3, 14, 23]. However, debiased solutions have so far rarely been adopted for benchmarking. Reasons for that may be that modified videos in case of [11, 15] often introduce artifacts and benchmarks with explicitly low scene bias [14, 23], might be considered out-of-domain relative to standard training setups. Another line of research questions whether current video benchmarks truly require temporal understanding, as single-frame models have shown comparable performance on several benchmarks [5, 19]. While new benchmarks are emerging to assess temporal understanding [2, 7, 28], temporal representen + +tation bias in standard benchmarks remains underexplored. + +In this work, we propose a new scalable method, Unbiasing through Textual Descriptions (UTD), to systematically analyze and mitigate different types of representation bias, such as object bias, in existing video benchmarks through automatic textual descriptions. Since most datasets lack textual narrations of videos, we leverage Vision-Language Models (VLMs) to generate detailed descriptions of video frames. We then use Large Language Models (LLMs) to extract different concepts from those captions, namely objects, activities, and verbs (an overview of this process is shown in Fig. 3). Based on generated descriptions, we examine representation biases across three independent dimensions. First, concept representation bias — assessing if a specific concept (e.g., objects) alone is sufficient for accurate prediction of the video. Second, temporal representation bias — evaluating if temporal information is needed for correct prediction or if single-frame information is sufficient. Lastly, we distinguish between common sense and dataset bias. Common sense bias refers to the implicit association of certain concepts with specific activities, such as associating an object “a piano” with the label “playing piano.” This differs from dataset bias, where spurious correlations arise within the dataset, e.g., “apply makeup” label frequently occurs with objects like “mirror” and “flowers” in the background. We propose to leverage the zero-shot reasoning capabilities of LLMs to assess common sense bias and train a linear classifier for the dataset bias. + +In our study, we found that many samples of popular benchmarks have a significant object representation bias, meaning that recognizing only objects on the video is already enough to predict the correct label. Therefore, using such data for benchmarking video backbones may give models with strong object recognition ability an advantage (i.e. a shortcut), unintentionally hindering video understanding evaluation with respect to temporal reasoning or action recognition. Therefore, by excluding test samples with the highest object bias, we introduce UTD-splits, debiased evaluation splits for considered datasets, which we believe provide a more robust assessment of video understanding capabilities. Importantly, compared to other approaches that rely on modifying videos [11, 15], this approach neither requires any additional processing of evaluation data, as no new videos are generated nor does it require any fine-tuning to adjust for any domain bias that could be introduced by debiasing video frames at a pixel level. + +We apply the proposed UTD method to analyze and debias six action classification datasets, namely UCF101 [36], SomethingSomethingv2 [14], Kinetics-400 [17], -600 [8], and -700 [9], and Moments In Time [31], and six text-to-video retrieval datasets, namely MSRVTT [48], DiDeMo [1], ActivityNet [6], LSMDC [33], YouCook2 [12], and Spoken Moments In Time [32]. Fur + +![](images/834fcd624e91a485f644e93d98bb548065be8c122eceebe071789c400e19f283.jpg) + +![](images/f2f50bbbf6cf8ed897ca6a7a880cef0a794ca46d3adf699c7747b77996b8529a.jpg) + +![](images/dadd4b2751de16f84e3c1f01fd49e0f814e5c466ca9a25f98aa424b095ee8702.jpg) +Figure 2. The answer to Fig. 1. Some videos may exhibit an object representation bias (allowing predictions based solely on objects) or single-frame representation bias (solely on a single frame), while others require more information for prediction. + +![](images/31d354e5d76d81af4e06ea7fe752cb877a5b166085ca9c5da74f63ac7725b900.jpg) + +ther, we benchmark 30 state-of-the-art video models of six different model designs, namely VideoMAE [38], VideoMAE2 [42], All-in-one [40], Unmasked Teacher [21], VideoMamba [22], and InternVid [44], on the original and UTD-debiased splits. We show that debiasing reduces performance saturation on current benchmarks. Considering model performance drops between original and debiased splits also allows us to analyze models on object bias, e.g., in action classification, larger backbones or those trained on more data appear more robust to object bias. + +We summarize the contributions of this work as follows: (1) We introduce Unbiasing through Textual Descriptions (UTD), a scalable, automated method to analyze and mitigate representation biases in video benchmarks without the need for human annotators in the loop. (2) Using UTD, we analyze representation biases across 12 popular classification and retrieval video benchmarks and release UTD-debiased test splits for each, UTD-splits, promoting a more robust evaluation of video understanding. (3) We benchmark 30 state-of-the-art video models on both the original and UTD-debiased splits to assess their robustness against object bias under varying conditions. (4) Finally, we release UTD-descriptions, a dataset of detailed structured descriptions (including objects, activities, and verb categories) for all 12 considered datasets, to support further dataset analysis and the development of less biased benchmarks. + +# 2. Related Work + +Debiasing Video Datasets. While most visual datasets are prone to various biases [13], one unique effect in video understanding is the impact of object and scene information on the overall performance. Different strategies have so far been explored to mitigate this bias. He et al. [15] and HAT [11] proposed to separate the performance of scene appearance from human motion by segmenting human fig- + +![](images/7e412135a17f8098fd8ef3f6745550ec7642f0351cc698c3e24509721f75ac04.jpg) +Figure 3. The proposed UTD method involves generating textual descriptions of different concepts in video frames using VLMs and LLMs, combining them in various temporal configurations, and evaluating the performance of these concept-temporal representations with strong text embedding models. For each representation, we distinguish between common sense bias, which relies on zero-shot reasoning by text embedding models, and dataset bias, assessed using a linear model trained on the dataset's training set. + +ures and placing them into black/gray background, or alternatively by cropping or inpainting human figures and classifying videos based on background only. Weinzaepfel et al. [47] proposed the Mimetics dataset by extending certain classes from the Kinetics dataset [17] with videos of activities in unrelated background. Zhai et al. [51] proposed an approach for scene-debiased action recognition via adversarial scene reconstruction and classification. Another line of work focused on capturing datasets with explicitly low object and scene bias. SomethingSomethingv1 and later v2 dataset [14] were proposed as an almost scene-free dataset, focusing on simple manipulative actions only. Li et al. [23] proposed a procedure for designing representation unbiased datasets and introduced Diving48 as an example. However, this specialization usually also implies that those cases are rather suitable for specialized approaches but not for, e.g., zero-shot evaluation of generalist video models. Other studies [5, 19] showed that for several benchmarks, using single frames during training or inference can achieve performance comparable to models that incorporate temporal information. However, recent efforts have mainly focused on developing new benchmarks to assess temporal understanding in videos [2], e.g., visual question-answering benchmarks [7, 28], while temporal representation bias in standard benchmarks has not been thoroughly investigated. In contrast, our UTD method allows for a systematic analysis of representation biases across various dimensions, including different concept and temporal configurations. Additionally, we are the first to propose mitigating object representation bias in popular benchmarks by discarding object-biased samples, which helps us to avoid adding artifacts to + +test samples (as with cropping or inpainting in [11, 15]) and keeps the dataset domain unconstrained, unlike [14, 23]. + +LLMs for Video Understanding. The progress of LLMs prompted a wide range of applications in various fields, including video understanding. Here, one line of work utilizes pre-trained LLMs in video-language tasks, such as video captioning [37, 46, 53], video question answering [21, 27, 46, 53], text-video retrieval [35, 53], or dialog-based user interaction [46, 52]. Some works [10, 34, 44] create new video-language datasets using image/audio captioning models or subtitles and using LLM summarization and reasoning ability to post-process these captions. Other works [20, 39, 45, 50] use expert perception models to extract information from the video that is later processed by the LLM for tasks such as video captioning. In our work, we leverage VLM- and LLM-generated frame descriptions to evaluate different representation biases in video datasets. + +# 3. Discovering Representation Biases + +In this section, we define representation bias in Sec. 3.1, outline the considered representation biases in Sec. 3.2, and detail our UTD method in Sec. 3.3. + +# 3.1. Definition: Representation Bias + +To formalize the definition of representation bias in datasets, we adopt the notation from RESOUND [23]. + +Let $D = \{v_{n}\}_{n\leq N}$ be a video dataset $D$ where each video $v_{n}$ consists of a sequence of frames $v_{n} = \{f_{n,i}\}_{i\leq |v_n|}$ . A representation $\phi$ is a function $\phi (v)$ of the video input $v$ . Let $R = \{\phi \}$ be a family of representations that share specific characteristics. For example, a fam + +ily of single-frame representations includes representations that encode information from only a single frame of a video $\phi (v_{n}) = \phi (f_{n,i})$ where $i\leq |v_{n}|$ or a family of object-based representations encodes only the objects in the video without considering motion or activity information. + +The representation $\phi$ can be used to train different models $\gamma_{\phi}$ for desired tasks, e.g. video classification. Let $M(v_{n},\gamma_{\phi})$ be the performance (e.g., accuracy) of a model $\gamma_{\phi}$ on a video $v_{n}$ , then the performance of $\gamma_{\phi}$ on $D$ is + +$$ +M (D, \gamma_ {\phi}) = \frac {1}{N} \sum_ {n = 1} ^ {N} M \left(v _ {n}, \gamma_ {\phi}\right). \tag {1} +$$ + +We define the performance of a representation $\phi$ on $D$ as + +$$ +M (D, \phi) = \max _ {\gamma_ {\phi}} M (D, \gamma_ {\phi}), \tag {2} +$$ + +where the maximum is taken over all models $\gamma_{\phi}$ that are based on the representation $\phi$ + +When the performance $M(D, \phi)$ of a representation $\phi$ is high, the dataset exhibits a preference for this representation, indicating that the dataset is more biased toward this representation. In this work, we use $M(D, \phi)$ as a measure of bias towards representation $\phi$ . However, a high bias toward a specific representation is not necessarily a negative thing. Ideally, a dataset should exhibit the highest bias toward the "ground truth representation": the representation truly required to solve the task. For instance, in video understanding, one could argue that the optimal representation should capture both static (e.g., objects) and dynamic (e.g., actions) properties, as well as encode temporal aspects. However, if a video dataset used to evaluate video understanding exhibits a strong object or single-frame representation bias, it may unintentionally favor models with strong object recognition or image understanding capabilities, thus hindering the evaluation of true video understanding. Therefore, in this work, we aim to estimate representation biases in popular video benchmarks. + +# 3.2. Different Representation Biases + +In our work, we consider different types of representation biases with respect to three independent aspects: + +Concept Representations: First, we consider representations that encode four different concept categories: objects $\phi_{o}(v_{n})$ , activities $\phi_{a}(v_{n})$ (actions along with tied objects), verbs $\phi_{\nu}(v_{n})$ (actions alone, without objects), and objects+composition+activities $\phi_{d}(v_{n})$ (objects, their spatial compositions in the video, and activities), see an example in Fig. 3. We aim to assess biases in classification and retrieval benchmarks toward particular concepts. For example, in retrieval, we ideally would like to see that benchmarks are not biased solely toward objects or activities. + +Temporal Representations: Second, we consider the bias of single-frame representations compared to representations that consider multiple frames. Let $\phi_x$ be + +one of the objects $\phi_o$ , activities $\phi_a$ , verbs $\phi_\nu$ , or objects+composition+activities $\phi_d$ representations. Similarly to [41], we consider the following options to encode information from multiple frames: middle-frame refers to the case that a video is represented only by a single middle frame of this video, e.g., $\phi_{x,\mathrm{mid}}(v_n) = \phi_x(f_{n,|v_n| / 2})$ , max-score frame refers to the case that a video is represented using a single frame $\phi_{x,\max}(v_n) = \phi_x(f_{n,i})$ where the frame $f_{n,i}$ has the highest performance score $M(f_{n,i},\phi_x)$ over all frames, average-overframes refers to the case that the video is represented as an average representation of frames ignoring the order of frames $\phi_{x,\mathrm{avg}}(v_n) = \frac{1}{|v_n|}\sum_{i\leq |v_n|}\phi_x(f_{n,i})$ , and, finally, sequence-of-frames refers to the case that the video representation encodes concepts and their order in the frames $\phi_{x,\mathrm{seq}}(v_n) = \phi_x(f_{n,1},f_{n,2},..,f_{n,|v_n|})$ . + +Common Sense Bias vs. Dataset Bias: Finally, for any concept-temporal representation $\phi$ (e.g., object middle-frame representation $\phi_{o,mid}$ ), we can consider two types of representation biases: common sense bias or dataset bias. Common sense representation bias refers to bias based on "common sense reasoning", e.g., a video featuring an object "piano" would most likely be classified under the "playing piano" activity label. Formally, in Eq. (2), this bias $M(D,\phi)$ is measured by taking the maximum performance only over zero-shot models $\gamma_{\phi}$ that have not seen the correlations in the training set of $D$ . We distinguish this from dataset bias, which refers to biases caused by spurious correlations within the dataset, e.g., where certain objects are unintentionally associated with specific labels, such as "mirror" or "flower bouquet" frequently appear with the "apply makeup" label. In this case, in Eq. (2), we measure dataset bias by taking the maximum performance over models $\gamma_{\phi}$ trained or fine-tuned on the training set of $D$ . + +We note that the concept, temporal, and common sense vs. dataset bias dimensions are independent, allowing us to combine them in any configuration. For example, we can evaluate object middle-frame common sense bias or activity sequence-of-frames dataset bias. + +# 3.3. Unbiasing through Textual Descriptions (UTD) + +Since measuring bias for every representation within a family (e.g., a family of object representations) is infeasible, we approximate it by considering a single representative from the family [23] and propose a way to precisely control what each representation encodes by using textual descriptions. + +The key idea of our method is to use textual descriptions as an intermediate video representation. This approach allows us to easily control which concepts in the video are included or excluded in the representation and to combine information into different temporal representations. Namely, to obtain representation $\phi$ that encodes only a specific concept category, we propose to encode visible concepts of this + +category into textual descriptions, e.g., encode objects with a list of visible objects, e.g., "table, lamp, ..". This approach, compared to cropping or inpainting [11, 15], allows us to isolate specific concepts more precisely, avoiding information leakage from cropped areas or adding artifacts caused by imperfect inpainting. Additionally, textual descriptions allow us to use off-the-shelf LLMs with strong common sense reasoning abilities to estimate common sense bias. Since metadata with visible objects, activities, etc., is usually not available, we propose using recent powerful VLMs and LLMs to extract this information. + +In the next sections, we first describe our method of obtaining textual description with respect to different concept and temporal categories (Sec. 3.3.1), then we describe our method of discovering common sense and dataset biases (Sec. 3.3.2), and lastly, we describe our method of unbiaseding benchmarks (Sec. 3.3.3). + +# 3.3.1. Obtaining Textual Descriptions + +While current VLMs [24-26], trained to perform complex image reasoning, are able to generate detailed descriptions of video frames, they still face challenges when prompted to directly extract specific concepts from a frame, e.g., only visible objects. Therefore, we decomposed the task of concept extraction into two parts: extraction of detailed frame description using VLMs and then extraction of different concepts from frame description using LLMs (see Fig. 3). + +Obtaining Textual Descriptions Using VLMs. Given train/test videos $v_{n} = \{f_{n,i}\}_{i\leq |v_{n}|}, n < N$ of the dataset, we feed each frame $f_{n,i}$ of each video $v_{n}$ into the VLM model $d$ to generate a detailed textual description $d_{n,i} = d(f_{n_i})$ (Fig. 3). We found that without complex prompt engineering or few-shot in-context learning strategies [4] VLMs produce long, detailed descriptions. While our method is not restricted to a specific VLM, for reproducibility and feasibility, we select an open-sourced strong middle-scale VLM model, LLaVA-1.6-Mistral-7B [26] with the prompt: "Describe the objects' relationships in the photo." Since the resulting textual descriptions include objects, their composition, relationships, and visible activities, we use them as objects+composition+activities descriptions. + +Extracting Concepts Categories Using LLMs. Each of the obtained detailed descriptions $d_{n,i}$ is further fed to the LLM $\zeta$ to extract only objects $o_{n,i} = \zeta(d_{n,i}, p_o)$ , only activities $a_{n,i} = \zeta(d_{n,i}, p_a)$ information, with $p_o, p_a$ as respective prompts. Then, to obtain information about verbs, we task the LLM with a prompt $p_\nu$ to delete all object information from activities $\nu_{n,i} = \zeta(a_{n,i}, p_\nu)$ . Since such text processing requires understanding what objects and activities are, our prompt includes a few examples following a few-shot in-context learning strategy [4]. We provide all prompts in the supplementary material. + +Aggregating Temporal Information. We use the obtained textual descriptions $x_{n,i}$ of video $v_{n}$ (where $x_{n,i}$ is one + +of $o_{n,i}$ $a_{n,i}$ , $\nu_{n,i}$ , or $d_{n,i}$ ) in different temporal setups: first, we keep them as single frame descriptions that might be further aggregated when transformed into vector representations (e.g. average-over-frames), as we describe in Sec. 3.3.2, second, we combine them into sequence-over-frame $x_{n}$ setup with the template: "Frame 1: $x_{n,1}$ Frame 2: $x_{n,2} \ldots$ Frame $|v_{n}| \colon x_{n,|v_{n}|}$ " (Fig. 3). Since objects+composition+activities descriptions $d_{n,i}$ when combined with sequence-over-frame templates become too long, we summarize them to 15-word descriptions $d_{n,i}'$ via an LLM before inputting them into the sequence prompt. + +# 3.3.2. Calculating Representation Biases + +After obtaining a specific concept-temporal representation $\phi$ , we assess common sense and dataset bias for dataset $D$ toward $\phi$ by measuring the performance $M(D, \phi)$ , as stated Eq. (2), where maximum is taken over zero-shot models $\gamma_{\phi}$ for common sense bias and over fine-tuned models $\gamma_{\phi}$ for dataset bias (see Sec. 3.2). Since maximizing over all models is infeasible in practice, we design strong models for common sense and dataset-based reasoning by utilizing state-of-the-art LLMs. Our ablation study verifies our design choice (see supplementary material). + +Common Sense Bias. We leverage recent strong LLM-based text embedding models [30, 43]. Such models are trained on large datasets to encode text into vector representations that can be used for various tasks, such as information retrieval. We encode our textual descriptions and all class labels/captions using the pre-trained text embedding model $\chi$ . Specifically, we prompt the model with the relevant concept and task. For example, we use prompts like "Given a list of objects visible on the video frame, retrieve the activity depicted in this video" for object description and "Given an activity, retrieve a video frame description that may depict this activity" for class labels in action classification. We use cosine similarity between obtained embeddings for zero-shot classification/retrieval. Prompts are detailed in the supplement. + +Dataset Bias. For action classification datasets, we evaluate dataset bias, as action classification models are often tested in fine-tuned setups and might exploit dataset representation biases learned from training data. For this we a linear logistic regression model trained on our strong text embeddings used in the common sense bias analysis and train model on the official train splits of the datasets. + +Metrics. For the performance measure $M(D, \phi)$ , we use accuracy for action classification and recall@1 and text-to-video retrieval as the most popular metrics for these tasks. + +# 3.3.3. Unbiasing Datasets + +In Sec. 4, we find that many video datasets exhibit a significant object representation bias. To support robust video model evaluation, we construct UTD-splits, object-debiased test splits, by excluding object-biased samples from the + +
UCFSSv2K400K600K700MiTMSRDDMActNLSMDCYC2S-MiT
obj+comp+act66.36.448.044.139.022.636.627.226.617.08.445.9
objects63.3-3.05.3-1.145.9-2.141.8-2.337.0-2.021.0-1.632.1-4.527.0-0.224.8-1.813.6-3.47.9-0.529.8-16.1
activities67.4+1.16.445.2-2.841.4-2.736.7-2.321.0-1.625.1-11.521.1-6.121.4-5.214.7-2.38.1-0.341.1-4.8
verbs50.8-15.55.8-0.624.8-23.221.4-22.717.6-21.416.2-6.410.5-26.17.0-20.27.4-19.25.5-11.51.2-7.213.1-32.8
+ +Table 1. Evaluation of concept representation bias. We report the performance of our model in sequence-of-frames temporal setup and common sense bias setup and color code with respect to the difference to objects+composition+activities (obj+comp+act) concepts. It shows that the overall classification performance drops only slightly when classifying actions based on objects, compared to classification and retrieval on verbs only, with SSv2 as the only dataset with less drop when classifying on verbs only. + +original test sets. Since our method allows us to measure the performance of the representation $M(v_{n},\phi)$ on individual samples $v_{n}$ , we exclude samples that are classified or retrieved correctly ( $M(v_{n},\phi) = 1$ ) based on object sequence-of-frame representation. For text-to-video retrieval, we debias with respect to common-sense bias, but for action classification, we debias with respect to dataset bias since video models are mostly tested in fine-tuned setup. To minimize the impact of random fluctuations in the process, we use three different prompts for labels/captions in the text embedding models and train three linear models for each classification dataset using bootstrapped training sets (we provide details and analysis in the supplement) and exclude only samples when all nine models agreed on correct classification/retrieval. Note that the percentage of removed samples is determined automatically based on the extent of object representation bias. + +Finally, since debiasing may disproportionately remove certain label classes in classification datasets, we additionally construct balanced UTD splits. While maintaining the total number of removed samples as in the original debiasing method, we adjust the number of samples removed from each class based on their average confidence (across nine models) to preserve the original class proportions. + +# 4. Benchmarking Datasets + +In this section, we analyze action classification and text-to-video retrieval datasets on different representation biases using our UTD method. + +# 4.1. Setup + +Datasets. We consider six common action classification datasets: UCF101 (UCF) [36], SomethingSomethingv2 (SSv2) [14], Kinetics400 (K400) [17], -600 (K600) [8], -700 (K700) [9], Moments In Time (MiT) [31] and six text-to-video retrieval datasets: MSRVTT (MSR) [48], YouCook2 (YC2) [12], DiDeMo (DDM) [1], LSMDC [33], ActivityNet (ANet) [6], and Spoken Moments in Time (S-MiT) [32]. + +Implementation Details. For all train/test videos, we use $|v_{n}| = 8$ uniformly sampled frames. We use the LLaVA-1.6-7B-mistral model [26] as VLM, Mistral-7B-Instructv0.2 [16] as LLM, and SFR-Employing-Mistral [30] as text + +embedding model. The temperature is set to 0. + +# 4.2. UTD-descriptions + +Following our UTD pipeline, first, we generate textual descriptions for 12 considered datasets, resulting in UTD-descriptions dataset (see qualitative examples in the supplement). We conducted a user study to validate the quality of object, activity, and verb predictions. Participants were asked to verify which objects from a provided list were visible in a given image. The study used 100 random frames from 100 videos of the UCF dataset (which has noisy, low-resolution frames that might be an additional challenge to the VLM). Seven participants validated 761 annotated objects in total, where $87.6\%$ (667 out of 761) were identified as visible, with only 94 not selected as visible. To better understand the VLM's errors, we manually classified these 94 objects into five categories: 1) attribute error (e.g., the object is "right hand" instead of "left hand"), 2) misclassification (the object is present but incorrectly identified as a different object, e.g., a "snowboard" instead of "snow slide"), 3) hallucination, 4) human annotation mistake – the object is visible, and 5) other. We found that only 45 objects were classified as hallucinations, 13 as misclassifications, 20 as "object visible" 8 as attribute errors, and 8 as other (see examples in the supplement). This results in a hallucination rate of only $6\%$ , which we find acceptable for our analysis. + +# 4.3. Concept Representation Bias + +Further, based on generated descriptions, we evaluate representation bias with respect to four concept categories, namely objects, activities, verbs, and objects+composition+activities, as described in Sec. 3.2. For this, we consider common sense bias in sequence-of-frames temporal setup in Tab. 1 (all concept-temporal combination results can be found in the supplement). It shows that using only object information results in only a slight drop in performance compared to objects+composition+activities, indicating that most datasets can, to a certain extent, be addressed with object information, thus showing an object representation bias. Interestingly, classification based on only objects results in 63.3pp accuracy for UCF101 and 45.9pp for Kinetics400. This may be due to the fact that some action classes are tied to the objects (e.g. "playing + +
UCFSSv2K400K600K700MiT
obj.act.verbsobj.act.verbsobj.act.verbsobj.act.verbsobj.act.verbsobj.act.verbs
common sense b.63.367.450.85.36.45.845.945.224.841.841.421.437.036.717.621.021.016.2
dataset bias80.3+17.083.4+16.073.5+22.713.6+8.316.0+9.612.2+6.460.7+14.860.7+15.542.2+17.460.4+18.660.3+18.942.5+21.152.3+15.349.7+13.034.9+17.327.8+6.827.3+6.322.9+6.7
+ +Table 2. Comparison of common sense bias vs dataset bias. We report the performance of our model in sequence of frame temporal setup and color code with respect to difference to common sense bias. It shows that the overall performance growth is significant for all concepts, including object concepts, indicating that datasets might contain a significant correlation between labels and visible objects. + +
UCF SSv2 K400 K600 K700 MiTMSR DDM ActN LSMDC YC2 S-MiT
middle-frame61.36.039.735.631.120.123.419.813.512.76.135.5
max-score frame66.57.448.043.338.721.433.429.121.518.38.946.1
avg.-over-frames66.76.846.642.137.123.431.926.526.616.48.643.8
seq.-of-frames66.36.448.044.139.022.636.627.226.617.08.445.9
+ +Table 3. Evaluation of temporal representation bias. We report the performance of our model in objects+composition+activities concepts setup and common sense bias setup. + +
UCF SSv2 K400 K600 K700 MiTMSR DDM ActN LSMDC YC2 S-MiT
Fleiss' kappa86.177.891.290.792.591.292.091.790.992.591.393.2
% object-biased samples7375555482428232011636
% of full test split in UTD-split279345455276727780899464
+ +Table 4. UTD-splits statistics. Fleiss' kappa coefficient measures the agreement among 9 models used to create UTD-splits. + +piano"), making visible objects equally or more reliable indicators of a class than activities. This becomes clear when we consider the verbs-based representation. Here, the performance drops significantly for most action classification datasets, except for SSv2 and MiT. Also, in the case of text-to-video retrieval, it shows that objects are better indicators than activities for some datasets (MSRVTT, DiDeMo, ActivityNet), and activities are better indicators for others (YouCook2, LSMDC, S-MiT). + +# 4.4. Common Sense vs. Dataset Bias + +We further consider common sense bias compared to dataset bias for the action classification datasets in Tab. 2. Overall, across all benchmarks and concepts, dataset bias is significantly higher than common sense bias, and the results indicate that considered benchmarks also contain a dataset-specific object bias in the training set. The performance on UCF using only objects increases from 63.3pp to 80.3pp, and even on SSv2, which was developed to minimize object representation bias, using only correlation of seen objects in the train set to the classes, we were able to predict 13.6pp of videos correctly. This significant increase based on available object information only shows that some actions might be connected to unrelated objects that hint at a certain class without being directly associated with it. + +# 4.5. Temporal Representation Bias + +Finally, we consider representation bias with respect to the temporal aspect in Tab. 3. Namely, we consider which temporal representations, single frame vs. multiple frames, + +best-score frame vs. middle frame, as well as average-over-frames vs. sequence-of-frames embeddings, perform best. For all datasets except ActivityNet, max-score-frame performance is comparable (< 3.5% difference) to multiple frames performance (average of frames/sequence over frames). Moreover, for SSv2, Kinetics400, DiDeMo, LSMDC, YouCook2, and S-MiT, we found that max-score-frame is a better indicator of class/captions than all frames. + +# 4.6.UTD-splits + +Since our analysis reveals significant object bias in benchmarks, we generate object-debiased UTD-splits for considered benchmarks, as described in Sec. 3.3.3. Namely, we remove object-biased samples where nine models that use different prompts in the text embedding model or trained on the bootstrapped train set uniformly agree that the sample can be predicted correctly based solely on objects. Tab. 4 summarizes the percentage of object-biased samples identified in the datasets, the level of model agreement (measured by Fleiss' kappa coefficient, which is notably high), and the number of samples retained from the original datasets. + +# 5. Benchmarking Video Models + +In this section, we benchmark state-of-the-art video encoders on considered datasets and their UTD-splits. + +# 5.1. Setup + +Video Models. We consider state-of-the-art video encoders with varying model sizes and pre-training configurations with open-sourced weights. For action classification, we benchmark 24 different variants of VideoMAE [38], VideoMAE2 [42], All-in-one [40], Unmasked Teacher (UMT) [21], VideoMamba [22], and InternVid [44] families. For text-to-video retrieval, we consider 17 videolanguage models of UMT [21], VideoMamba [22], and InternVid [44] families. In total, we analyze 30 models and list all considered variants in the supplement. + +Implementation Details. To evaluate action classification performance in a standard aligned setup, we follow the VideoGLUE [49] benchmark of using a frozen video backbone and lightweight pooling head as a test of the strength of video model representation. To adapt each video backbone for a specific dataset, we create a model that 1) uses frozen video backbone to extract spatial-temporal features, and then 2) uses a single attention pooling layer [49] to ag- + +
UCFUCF-UTD-splitUCF-UTD-s. balancedSSv2SSv2-UTD-splitSSv2-UTD-s. balancedK400K400-UTD-splitK400-UTD-s. balanced
videomae-B-K40089.281.3–7.982.8–6.454.252.1–2.153.0–1.264.345.7–18.649.5–14.8
videomae-B-UH88.479.2–9.280.4–8.056.254.0–2.254.9–1.364.345.4–18.949.6–14.7
videomae-L-UH95.790.9–4.891.0–4.767.365.4–1.966.2–1.175.859.3–16.563.3–12.5
videomae-H-UH95.290.9–4.391.0–4.267.465.5–1.966.3–1.176.559.5–17.063.7–12.8
videomaev2-B-K710-fnK71099.098.1–0.997.6–1.457.154.7–2.455.9–1.285.071.9–13.175.5–9.5
allinone-B-WV2M+CC84.563.1–21.472.9–11.626.222.5–3.724.7–1.566.942.6–24.350.8–16.1
allinone-B-WV2M+HT81.457.7–23.768.3–13.122.719.3–3.421.3–1.461.436.9–24.545.2–16.2
allinone-B-WV2M+HT+CC+YTT+81.358.5–22.866.2–15.122.018.8–3.220.7–1.361.236.4–24.844.6–16.6
umt-B-K71091.376.8–14.582.4–8.950.647.9–2.749.1–1.577.758.3–19.463.7–14.0
umt-B-K710-fnK71099.096.8–2.297.1–1.949.446.6–2.848.1–1.385.672.4–13.276.2–9.4
umt-L-K71095.787.2–8.589.2–6.559.457.0–2.458.1–1.384.068.4–15.673.0–11.0
umt-L-K710-fnK71098.996.7–2.297.1–1.857.955.5–2.456.7–1.289.078.7–10.381.6–7.4
videomamba-vm-K40090.974.6–16.381.0–9.935.932.6–3.334.4–1.575.354.4–20.960.5–14.8
videomamba-vm-5M93.782.0–11.786.7–7.048.545.6–2.946.9–1.677.558.1–19.464.0–13.5
videomamba-vm-17M93.280.5–12.784.3–8.947.444.7–2.745.9–1.577.758.1–19.663.6–14.1
videomamba-vm-25M94.383.0–11.386.5–7.848.745.9–2.847.1–1.678.459.1–19.364.6–13.8
internid-B-10M-FLT94.081.1–12.984.7–9.348.145.2–2.946.7–1.478.659.3–19.364.9–13.7
internid-B-200M94.582.9–11.685.2–9.354.451.8–2.652.8–1.680.061.7–18.366.8–13.2
internid-L-10M93.280.2–13.085.3–7.952.650.0–2.651.1–1.577.457.4–20.063.8–13.6
internid-L-WebVid10M94.182.5–11.686.3–7.854.952.4–2.553.4–1.578.959.7–19.265.7–13.2
internid-L-10M-DIV94.883.2–11.687.6–7.251.748.8–2.950.1–1.680.362.1–18.267.4–12.9
internid-L-10M-FLT95.586.3–9.288.4–7.153.650.9–2.752.1–1.581.964.7–17.269.6–12.3
internid-L-50M95.285.5–9.788.4–6.860.658.2–2.459.3–1.381.363.7–17.669.1–12.2
internid-L-200M96.890.5–6.391.9–4.963.161.0–2.161.9–1.284.669.4–15.273.9–10.7
+ +Table 5. Benchmarking video models in action classification. We report accuracy on full test/val sets and our debiased UTD-splits. The accuracy differences with respect to the full test/val sets are color-coded. + +
MSRMSR-UTD-splitDDMDDM-UTD-splitANetANet-UTD-split
umt-b-5M30.020.5–9.530.220.6–9.628.622.3–6.3
umt-b-17M35.626.3–9.337.727.6–10.134.227.3–6.9
umt-b-25M35.324.8–10.534.224.6–9.625.119.7–5.4
umt-l-5M34.824.9–9.933.521.8–11.734.828.7–6.1
umt-l-17M43.631.1–12.546.335.6–10.745.938.7–7.2
umt-l-25M42.330.6–11.743.633.5–10.136.730.4–6.3
videomamba-vm-5M33.323.0–10.337.127.1–10.037.130.1–7.0
videomamba-vm-17M34.925.5–9.440.628.9–11.740.433.0–7.4
videomamba-vm-25M34.925.5–9.441.430.5–10.941.133.8–7.3
internid-B-10M-FLT37.925.4–12.528.617.2–11.424.418.8–5.6
internid-B-200M38.124.7–13.430.219.4–10.826.220.1–6.1
internid-L-10M26.718.7–8.022.615.6–7.021.516.3–5.2
internid-L-WV10M26.517.6–8.922.213.8–8.423.117.5–5.6
internid-L-10M-DIV37.324.2–13.126.915.9–11.023.217.7–5.5
internid-L-10M-FLT38.726.3–12.429.218.8–10.424.519.0–5.5
internid-L-50M32.422.0–10.426.518.3–8.224.418.0–6.4
internid-L-200M38.224.8–13.430.320.2–10.128.722.1–6.6
+ +Table 6. Benchmarking video-language models in text-to-video retrieval. We report accuracy on full test/val sets and our debiased UTD-splits. The accuracy differences with respect to the full test/val sets are color-coded. + +ggregate spatial-temporal features into a single vector representation and 3) a single linear projection layer to classify on the required number of classes. We train the model on a full train split of the corresponding dataset. For video model evaluation in text-to-video retrieval, we reported their zero-shot performance on our debiased splits. We follow the corresponding model recipes in using re-ranking and how many examples are reranked. We report the accuracy for classification and the recall@1 for text-to-video retrieval. Additional details are provided in the supplement. + +# 5.2. Results + +We evaluate the performance of video models for action classification on the UCF, SSv2, and K400 datasets and their UTD-splits in Tab. 5 and for text-to-video retrieval on the MSRVTT, DiDeMo, and ActivityNet datasets and their UTD-splits in Tab. 6. Performance for other datasets is reported in the supplement. Across all models, performance drops notably on debiased splits (except for SSv2), suggesting that current models strongly rely on object representation bias. Evaluation on UTD debiased splits shows a higher range of scores of models, revealing that some models (e.g. all-in-one models) were more heavily relying on object bias rather than utilizing action/motion cues for video understanding. Further, we can observe that the best performance flips for some models for the debiased case compared to the full test set, e.g., for videomae-B-K400 compared to videomamba-vm-K400 on the UCF dataset. + +Additionally, by comparing the performance decline between the original datasets and the UTD-splits, we can assess models' sensitivity to object bias. Overall, we see in the + +case of action classification, Tab. 5, that VideoMAE models are most robust to object bias, whereas Allinone models are most impacted by object bias. We also note that performance drops correlate between the UTD-split and UTD-split-balanced versions. We can further observe that larger backbones/architectures, and to a certain extent, more training data, lead to a reduced bias compared to smaller models, e.g. InternVid and VideoMAE. This pattern does not hold for text-to-video retrieval, shown in Tab. 6. Here, larger backbones and backbones trained with more data turn out to have higher object bias than smaller models or models trained with less data (see InternVid). We attribute this behavior, to the fact that retrieval models leverage information from both the vision as well as the text backbone. Thus, if larger text backbones also learn more object properties or a larger vocabulary of objects, this might further encourage object-centering learning in the visual stream as well. + +# 6. Conclusion + +In this paper, we presented a new way to examine representation biases in video classification and retrieval datasets by leveraging recent, powerful VLMs and LLMs. Our investigation confirms that most video datasets are heavily focused on object recognition. Based on our bias analysis, we propose a method to reduce object bias in these benchmarks, enabling more robust video understanding evaluation. With that, we are able to benchmark state-of-the-art video models and show an analysis of the object bias in these models. We hope that our findings, together with the presented benchmark, offer valuable insights for the future development of more robust video understanding benchmarks and models. + +# 7. Acknowledgments + +Nina Shvetsova is supported in part by German Federal Ministry of Education and Research (BMBF) project STCL - 01IS22067. Nina Shvetsova also acknowledges travel support from ELISE (GA no 951847). Prof. Hilde Kuehne is supported in part by the ERC Starting Grant GraViLa 101117556. + +# References + +[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, 2017. 2, 6 +[2] Piyush Bagad, Makarand Tapaswi, and Cees GM Snoek. Test of time: Instilling video-language models with a sense of time. In CVPR, pages 2503-2516, 2023. 1, 3 +[3] Yizhak Ben-Shabat, Xin Yu, Fatemeh Saleh, Dylan Campbell, Cristian Rodriguez-Opazo, Hongdong Li, and Stephen Gould. The ika asm dataset: Understanding people assembling furniture through actions, objects and pose. 2021. 1 +[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 33:1877-1901, 2020. 5 +[5] Shyamal Buch, Cristóbal Eyzaguirre, Adrien Gaidon, Jiajun Wu, Li Fei-Fei, and Juan Carlos Niebles. Revisiting the" video" in video-language understanding. In CVPR, pages 2917-2927, 2022. 1, 3 +[6] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 2, 6 +[7] Mu Cai, Reuben Tan, Jianrui Zhang, Bocheng Zou, Kai Zhang, Feng Yao, Fangrui Zhu, Jing Gu, Yiwu Zhong, Yuzhang Shang, et al. Temporalbench: Benchmarking fine-grained temporal understanding for multimodal video models. arXiv preprint arXiv:2410.10818, 2024. 1, 3 +[8] João Carreira, Eric Noland, Andras Banki-Horvath, Chloe Hillier, and Andrew Zisserman. A short note about kinetics-600. arxiv, 1808.01340, 2018. 2, 6 +[9] João Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. arxiv, 1907.06987, 2019. 2, 6 +[10] Sihan Chen, Handong Li, Qunbo Wang, Zijia Zhao, Mingzhen Sun, Xinxin Zhu, and Jing Liu. Vast: A vision-audio-subtitle-text omni-modality foundation model and dataset. NeurIPS, 36, 2024. 3 +[11] Jihoon Chung, Yu Wu, and Olga Russakovsky. Enabling detailed action recognition evaluation through video dataset augmentation. In NeurIPS, pages 39020-39033. Curran Associates, Inc., 2022. 1, 2, 3, 5 +[12] Pradipto Das, Chenliang Xu, Richard F. Doell, and Jason J. Corso. A thousand frames in just a few words: Linguual description of videos through latent topics and sparse object stitching. In CVPR, 2013. 2, 6 + +[13] Simone Fabbrizzi, Symeon Papadopoulos, Eirini Ntoutsi, and Ioannis Kompatsiaris. A survey on bias in visual datasets. Computer Vision and Image Understanding, 223, 2022. 2 +[14] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, Florian Hoppe, Christian Thurau, Ingo Bax, and Roland Memisevic. The "something something" video database for learning and evaluating visual common sense. In ICCV, 2017. 1, 2, 3, 6 +[15] Yun He, Soma Shirakabe, Yutaka Satoh, and Hirokatsu Kataoka. Human action recognition without human. 2016. 1, 2, 3, 5 +[16] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 6 +[17] Will Kay, João Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arxiv, 1705.06950, 2017. 2, 3, 6 +[18] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In ICCV, 2011. 1 +[19] Jie Lei, Tamara L Berg, and Mohit Bansal. Revealing single frame bias for video-and-language learning. In ACL, 2023. 1, 3 +[20] Kunchang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 3 +[21] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. In ICCV, pages 19948-19960, 2023. 2, 3, 7 +[22] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. arXiv preprint arXiv:2403.06977, 2024. 2, 7 +[23] Yingwei Li, Yi Li, and Nuno Vasconcelos. Resound: Towards action recognition without representation bias. In ECCV, 2018. 1, 3, 4 +[24] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 5 +[25] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. +[26] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024.5,6 +[27] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In ACL, 2024. 3 + +[28] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. NeurIPS, 36:46212-46244, 2023. 1, 3 +[29] Marcin Marszalek, Ivan Laptev, and Cordelia Schmid. Actions in context. In CVPR, 2009. 1 +[30] Rui Meng, Ye Liu, Shafiq Rayhan Joty, Caiming Xiong, Yingbo Zhou, and Semih Yavuz. Sfrembedding-mistral: enhance text retrieval with transfer learning. *Salesforce AI Research Blog*, 3, 2024. 5, 6 +[31] Mathew Monfort, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, Dan Gutfreund, Carl Vondrick, et al. Moments in time dataset: one million videos for event understanding. TPAMI, 42(2):502-508, 2019. 2, 6 +[32] Mathew Monfort, SouYoung Jin, Alexander Liu, David Harwath, Rogerio Feris, James Glass, and Aude Oliva. Spoken moments: Learning joint audio-visual representations from video descriptions. In CVPR, 2021. 2, 6 +[33] Anna Rohrbach, Marcus Rohrbach, Niket Tandon, and Bernt Schiele. A dataset for movie description. In CVPR, 2015. 2, 6 +[34] Nina Shvetsova, Anna Kukleva, Xudong Hong, Christian Rupprecht, Bernt Schiele, and Hilde Kuehne. Howtocap-tion: Prompting llms to transform video annotations at scale. arXiv preprint arXiv:2310.04900, 2023. 3 +[35] Nina Shvetsova, Anna Kukleva, Bernt Schiele, and Hilde Kuehne. In-style: Bridging text and uncurated videos with style transfer for text-video retrieval. In ICCV, 2023. 3 +[36] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arxiv, 1705.06950, 2017. 1, 2, 6 +[37] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In ICCV, pages 7464-7473, 2019. 3 +[38] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. NeurIPS, 35:10078-10093, 2022. 2, 7 +[39] Junke Wang, Dongdong Chen, Chong Luo, Xiyang Dai, Lu Yuan, Zuxuan Wu, and Yu-Gang Jiang. Chatvideo: A tracklet-centric multimodal and versatile video understanding system. arXiv preprint arXiv:2304.14407, 2023. 3 +[40] Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Kevin Qinghong Lin, Satoshi Tsutsui, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, et al. All in one: Exploring unified video-language pre-training. In CVPR, pages 6598-6608, 2023. 2, 7 +[41] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In ECCV, 2016. 4 +[42] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In CVPR, pages 14549-14560, 2023. 2, 7 + +[43] Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. Improving text embeddings with large language models. arXiv preprint arXiv:2401.00368, 2023. 5 +[44] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Intervid: A large-scale video-text dataset for multimodal understanding and generation. In ICLR, 2023. 2, 3, 7 +[45] Zhenhailong Wang, Manling Li, Ruochen Xu, Luowei Zhou, Jie Lei, Xudong Lin, Shuohang Wang, Ziyi Yang, Chenguang Zhu, Derek Hoiem, et al. Language models with image descriptors are strong few-shot video-language learners. NeurIPS, 35:8483-8497, 2022. 3 +[46] Zhanyu Wang, Longyue Wang, Zhen Zhao, Minghao Wu, Chenyang Lyu, Huayang Li, Deng Cai, Luping Zhou, Shuming Shi, and Zhaopeng Tu. Gpt4video: A unified multimodal large language model for Instruction-followed understanding and safety-aware generation. arXiv preprint arXiv:2311.16511, 2023. 3 +[47] Philippe Weinzaepfel and Grégory Rogez. Mimetics: Towards understanding human actions out of context. *IJCV*, 129, 2021. 1, 3 +[48] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. 2, 6 +[49] Liangzhe Yuan, Nitesh Bharadwaj Gundavarapu, Long Zhao, Hao Zhou, Yin Cui, Lu Jiang, Xuan Yang, Menglin Jia, Tobias Weyand, Luke Friedman, et al. Videoglue: Video general understanding evaluation of foundation models. arXiv preprint arXiv:2307.03166, 2023. 7 +[50] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, et al. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 3 +[51] Yuanhao Zhai, Ziyi Liu, Zhenyu Wu, Yi Wu, Chunluan Zhou, David Doermann, Junsong Yuan, and Gang Hua. Soar: Scene-debiasing open-set action recognition. In ICCV, pages 10244–10254, 2023. 3 +[52] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 3 +[53] Long Zhao, Nitesh B Gundavarapu, Liangzhe Yuan, Hao Zhou, Shen Yan, Jennifer J Sun, Luke Friedman, Rui Qian, Tobias Weyand, Yue Zhao, et al. Videoprism: A foundational visual encoder for video understanding. arXiv preprint arXiv:2402.13217, 2024. 3 \ No newline at end of file diff --git a/2025/Unbiasing through Textual Descriptions_ Mitigating Representation Bias in Video Benchmarks/images.zip b/2025/Unbiasing through Textual Descriptions_ Mitigating Representation Bias in Video Benchmarks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1ca5b0c1d63fe21238f61a968e8ba92243d01c94 --- /dev/null +++ b/2025/Unbiasing through Textual Descriptions_ Mitigating Representation Bias in Video Benchmarks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1514da24080e639eb4823720ad0b48f33de0971758f1a6adaee983f5afda72eb +size 523552 diff --git a/2025/Unbiasing through Textual Descriptions_ Mitigating Representation Bias in Video Benchmarks/layout.json b/2025/Unbiasing through Textual Descriptions_ Mitigating Representation Bias in Video Benchmarks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5abd1ad89991e53c45343f5a74964ee97a7be55c --- /dev/null +++ b/2025/Unbiasing through Textual Descriptions_ Mitigating Representation Bias in Video Benchmarks/layout.json @@ -0,0 +1,8516 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 103, + 471, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 103, + 471, + 140 + ], + "spans": [ + { + "bbox": [ + 140, + 103, + 471, + 140 + ], + "type": "text", + "content": "Unbiasing through Textual Descriptions: Mitigating Representation Bias in Video Benchmarks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 160, + 547, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 160, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 63, + 160, + 547, + 217 + ], + "type": "text", + "content": "Nina Shvetsova" + }, + { + "bbox": [ + 63, + 160, + 547, + 217 + ], + "type": "inline_equation", + "content": "^{1,2,3*}" + }, + { + "bbox": [ + 63, + 160, + 547, + 217 + ], + "type": "text", + "content": ", Arsha Nagrani" + }, + { + "bbox": [ + 63, + 160, + 547, + 217 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 63, + 160, + 547, + 217 + ], + "type": "text", + "content": ", Bernt Schiele" + }, + { + "bbox": [ + 63, + 160, + 547, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 63, + 160, + 547, + 217 + ], + "type": "text", + "content": ", Hilde Kuehne" + }, + { + "bbox": [ + 63, + 160, + 547, + 217 + ], + "type": "inline_equation", + "content": "^{1,2,5}" + }, + { + "bbox": [ + 63, + 160, + 547, + 217 + ], + "type": "text", + "content": ", Christian Rupprecht" + }, + { + "bbox": [ + 63, + 160, + 547, + 217 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 63, + 160, + 547, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 160, + 547, + 217 + ], + "type": "text", + "content": " Goethe University Frankfurt, " + }, + { + "bbox": [ + 63, + 160, + 547, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 63, + 160, + 547, + 217 + ], + "type": "text", + "content": " Tuebingen AI Center/University of Tuebingen, " + }, + { + "bbox": [ + 63, + 160, + 547, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 63, + 160, + 547, + 217 + ], + "type": "text", + "content": " MPI for Informatics, Saarland Informatics Campus, " + }, + { + "bbox": [ + 63, + 160, + 547, + 217 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 63, + 160, + 547, + 217 + ], + "type": "text", + "content": " University of Oxford, " + }, + { + "bbox": [ + 63, + 160, + 547, + 217 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 63, + 160, + 547, + 217 + ], + "type": "text", + "content": " MIT-IBM Watson AI Lab shvetsov@uni-frankfurt.de" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 245, + 200, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 245, + 200, + 257 + ], + "spans": [ + { + "bbox": [ + 152, + 245, + 200, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 270, + 296, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 296, + 606 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 296, + 606 + ], + "type": "text", + "content": "We propose a new \"Unbiased through Textual Description (UTD)\" video benchmark based on unbiased subsets of existing video classification and retrieval datasets to enable a more robust assessment of video understanding capabilities. Namely, we tackle the problem that current video benchmarks may suffer from different representation biases, e.g., object bias or single-frame bias, where mere recognition of objects or utilization of only a single frame is sufficient for correct prediction. We leverage VLMs and LLMs to analyze and debias benchmarks from such representation biases. Specifically, we generate frame-wise textual descriptions of videos, filter them for specific information (e.g. only objects) and leverage them to examine representation biases across three dimensions: 1) concept bias—determining if a specific concept (e.g., objects) alone suffice for prediction; 2) temporal bias—assessing if temporal information contributes to prediction; and 3) common sense vs. dataset bias—evaluating whether zero-shot reasoning or dataset correlations contribute to prediction. We conduct a systematic analysis of 12 popular video classification and retrieval datasets and create new object-debiased test splits for these datasets. Moreover, we benchmark 30 state-of-the-art video models on original and debiased splits and analyze biases in the models. To facilitate the future development of more robust video understanding benchmarks and models, we release: \"UTD-descriptions\", a dataset with our rich structured descriptions for each dataset, and \"UTD-splits\", a dataset of object-debiased test splits.1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 630, + 135, + 642 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 630, + 135, + 642 + ], + "spans": [ + { + "bbox": [ + 56, + 630, + 135, + 642 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 651, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 295, + 675 + ], + "type": "text", + "content": "Ever since action recognition moved from simple, staged clip classification to real-life video understanding [18, 29," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 344, + 246, + 523, + 256 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 246, + 523, + 256 + ], + "spans": [ + { + "bbox": [ + 344, + 246, + 523, + 256 + ], + "type": "text", + "content": "Guess the activity from the following objects:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 360, + 260, + 389, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 260, + 389, + 268 + ], + "spans": [ + { + "bbox": [ + 360, + 260, + 389, + 268 + ], + "type": "text", + "content": "Video 1:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 318, + 271, + 428, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 271, + 428, + 289 + ], + "spans": [ + { + "bbox": [ + 318, + 271, + 428, + 289 + ], + "type": "text", + "content": "Person, makeup brush, makeup sponge, makeup palette" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 332, + 292, + 361, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 292, + 361, + 300 + ], + "spans": [ + { + "bbox": [ + 332, + 292, + 361, + 300 + ], + "type": "text", + "content": "Answer:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 478, + 260, + 508, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 478, + 260, + 508, + 269 + ], + "spans": [ + { + "bbox": [ + 478, + 260, + 508, + 269 + ], + "type": "text", + "content": "Video 2:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 447, + 274, + 541, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 447, + 274, + 541, + 284 + ], + "spans": [ + { + "bbox": [ + 447, + 274, + 541, + 284 + ], + "type": "text", + "content": "Person, stage, wall, curtain" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 451, + 292, + 481, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 451, + 292, + 481, + 299 + ], + "spans": [ + { + "bbox": [ + 451, + 292, + 481, + 299 + ], + "type": "text", + "content": "Answer:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 358, + 311, + 510, + 321 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 311, + 510, + 321 + ], + "spans": [ + { + "bbox": [ + 358, + 311, + 510, + 321 + ], + "type": "text", + "content": "Guess the activity from a single frame:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 360, + 322, + 389, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 322, + 389, + 330 + ], + "spans": [ + { + "bbox": [ + 360, + 322, + 389, + 330 + ], + "type": "text", + "content": "Video 3:" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 332, + 331, + 418, + 387 + ], + "blocks": [ + { + "bbox": [ + 332, + 331, + 418, + 387 + ], + "lines": [ + { + "bbox": [ + 332, + 331, + 418, + 387 + ], + "spans": [ + { + "bbox": [ + 332, + 331, + 418, + 387 + ], + "type": "image", + "image_path": "7cca5a13202945120aa3cf187b0849c1491a2df52b3caf83675a1e8cba9d6439.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 390, + 362, + 397 + ], + "lines": [ + { + "bbox": [ + 333, + 390, + 362, + 397 + ], + "spans": [ + { + "bbox": [ + 333, + 390, + 362, + 397 + ], + "type": "text", + "content": "Answer:" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 481, + 323, + 510, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 481, + 323, + 510, + 330 + ], + "spans": [ + { + "bbox": [ + 481, + 323, + 510, + 330 + ], + "type": "text", + "content": "Video 4:" + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 452, + 331, + 537, + 387 + ], + "blocks": [ + { + "bbox": [ + 452, + 331, + 537, + 387 + ], + "lines": [ + { + "bbox": [ + 452, + 331, + 537, + 387 + ], + "spans": [ + { + "bbox": [ + 452, + 331, + 537, + 387 + ], + "type": "image", + "image_path": "1f2f04ebfc6cb72bbb6f1fef87983da65e430507e89d9a0b00e3a589d3877ca5.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 390, + 482, + 397 + ], + "lines": [ + { + "bbox": [ + 453, + 390, + 482, + 397 + ], + "spans": [ + { + "bbox": [ + 453, + 390, + 482, + 397 + ], + "type": "text", + "content": "Answer:" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 407, + 555, + 429 + ], + "lines": [ + { + "bbox": [ + 313, + 407, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 555, + 429 + ], + "type": "text", + "content": "Figure 1. Can you guess activity on a video based solely on the objects or a single frame? (The answer is on the next page in Fig. 2.)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": "36], there has been an ongoing discussion about how much of a system's performance can be attributed to the recognition of the action itself compared to related or unrelated visual cues such as objects or scene information in the data [11, 15, 47], namely object/scene representation bias (see Fig. 1). Different strategies have been proposed to separate action recognition from scene and object influences such as segmenting the human figure in a video and placing it against a gray or black background [11, 15] as well as mimicking actions in unrelated contexts [47]. Some works focused on collecting datasets designed to minimize scene or object representation bias [3, 14, 23]. However, debiased solutions have so far rarely been adopted for benchmarking. Reasons for that may be that modified videos in case of [11, 15] often introduce artifacts and benchmarks with explicitly low scene bias [14, 23], might be considered out-of-domain relative to standard training setups. Another line of research questions whether current video benchmarks truly require temporal understanding, as single-frame models have shown comparable performance on several benchmarks [5, 19]. While new benchmarks are emerging to assess temporal understanding [2, 7, 28], temporal representen" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": "*The work was done during PhD visit to the University of Oxford within the ELLIS PhD program. \n1Project page: https://utd-project.github.io/" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29050" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 83 + ], + "type": "text", + "content": "tation bias in standard benchmarks remains underexplored." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 296, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 296, + 409 + ], + "type": "text", + "content": "In this work, we propose a new scalable method, Unbiasing through Textual Descriptions (UTD), to systematically analyze and mitigate different types of representation bias, such as object bias, in existing video benchmarks through automatic textual descriptions. Since most datasets lack textual narrations of videos, we leverage Vision-Language Models (VLMs) to generate detailed descriptions of video frames. We then use Large Language Models (LLMs) to extract different concepts from those captions, namely objects, activities, and verbs (an overview of this process is shown in Fig. 3). Based on generated descriptions, we examine representation biases across three independent dimensions. First, concept representation bias — assessing if a specific concept (e.g., objects) alone is sufficient for accurate prediction of the video. Second, temporal representation bias — evaluating if temporal information is needed for correct prediction or if single-frame information is sufficient. Lastly, we distinguish between common sense and dataset bias. Common sense bias refers to the implicit association of certain concepts with specific activities, such as associating an object “a piano” with the label “playing piano.” This differs from dataset bias, where spurious correlations arise within the dataset, e.g., “apply makeup” label frequently occurs with objects like “mirror” and “flowers” in the background. We propose to leverage the zero-shot reasoning capabilities of LLMs to assess common sense bias and train a linear classifier for the dataset bias." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 412, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 296, + 628 + ], + "type": "text", + "content": "In our study, we found that many samples of popular benchmarks have a significant object representation bias, meaning that recognizing only objects on the video is already enough to predict the correct label. Therefore, using such data for benchmarking video backbones may give models with strong object recognition ability an advantage (i.e. a shortcut), unintentionally hindering video understanding evaluation with respect to temporal reasoning or action recognition. Therefore, by excluding test samples with the highest object bias, we introduce UTD-splits, debiased evaluation splits for considered datasets, which we believe provide a more robust assessment of video understanding capabilities. Importantly, compared to other approaches that rely on modifying videos [11, 15], this approach neither requires any additional processing of evaluation data, as no new videos are generated nor does it require any fine-tuning to adjust for any domain bias that could be introduced by debiasing video frames at a pixel level." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": "We apply the proposed UTD method to analyze and debias six action classification datasets, namely UCF101 [36], SomethingSomethingv2 [14], Kinetics-400 [17], -600 [8], and -700 [9], and Moments In Time [31], and six text-to-video retrieval datasets, namely MSRVTT [48], DiDeMo [1], ActivityNet [6], LSMDC [33], YouCook2 [12], and Spoken Moments In Time [32]. Fur" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 432, + 149 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 432, + 149 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 432, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 432, + 149 + ], + "type": "image", + "image_path": "834fcd624e91a485f644e93d98bb548065be8c122eceebe071789c400e19f283.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 433, + 71, + 548, + 148 + ], + "blocks": [ + { + "bbox": [ + 433, + 71, + 548, + 148 + ], + "lines": [ + { + "bbox": [ + 433, + 71, + 548, + 148 + ], + "spans": [ + { + "bbox": [ + 433, + 71, + 548, + 148 + ], + "type": "image", + "image_path": "f2f50bbbf6cf8ed897ca6a7a880cef0a794ca46d3adf699c7747b77996b8529a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 150, + 432, + 233 + ], + "blocks": [ + { + "bbox": [ + 316, + 150, + 432, + 233 + ], + "lines": [ + { + "bbox": [ + 316, + 150, + 432, + 233 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 432, + 233 + ], + "type": "image", + "image_path": "dadd4b2751de16f84e3c1f01fd49e0f814e5c466ca9a25f98aa424b095ee8702.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 238, + 555, + 282 + ], + "lines": [ + { + "bbox": [ + 313, + 238, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 555, + 282 + ], + "type": "text", + "content": "Figure 2. The answer to Fig. 1. Some videos may exhibit an object representation bias (allowing predictions based solely on objects) or single-frame representation bias (solely on a single frame), while others require more information for prediction." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 433, + 150, + 552, + 233 + ], + "blocks": [ + { + "bbox": [ + 433, + 150, + 552, + 233 + ], + "lines": [ + { + "bbox": [ + 433, + 150, + 552, + 233 + ], + "spans": [ + { + "bbox": [ + 433, + 150, + 552, + 233 + ], + "type": "image", + "image_path": "31d354e5d76d81af4e06ea7fe752cb877a5b166085ca9c5da74f63ac7725b900.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 289, + 555, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 408 + ], + "type": "text", + "content": "ther, we benchmark 30 state-of-the-art video models of six different model designs, namely VideoMAE [38], VideoMAE2 [42], All-in-one [40], Unmasked Teacher [21], VideoMamba [22], and InternVid [44], on the original and UTD-debiased splits. We show that debiasing reduces performance saturation on current benchmarks. Considering model performance drops between original and debiased splits also allows us to analyze models on object bias, e.g., in action classification, larger backbones or those trained on more data appear more robust to object bias." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 408, + 556, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 556, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 556, + 600 + ], + "type": "text", + "content": "We summarize the contributions of this work as follows: (1) We introduce Unbiasing through Textual Descriptions (UTD), a scalable, automated method to analyze and mitigate representation biases in video benchmarks without the need for human annotators in the loop. (2) Using UTD, we analyze representation biases across 12 popular classification and retrieval video benchmarks and release UTD-debiased test splits for each, UTD-splits, promoting a more robust evaluation of video understanding. (3) We benchmark 30 state-of-the-art video models on both the original and UTD-debiased splits to assess their robustness against object bias under varying conditions. (4) Finally, we release UTD-descriptions, a dataset of detailed structured descriptions (including objects, activities, and verb categories) for all 12 considered datasets, to support further dataset analysis and the development of less biased benchmarks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 609, + 402, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 609, + 402, + 621 + ], + "spans": [ + { + "bbox": [ + 314, + 609, + 402, + 621 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "Debiasing Video Datasets. While most visual datasets are prone to various biases [13], one unique effect in video understanding is the impact of object and scene information on the overall performance. Different strategies have so far been explored to mitigate this bias. He et al. [15] and HAT [11] proposed to separate the performance of scene appearance from human motion by segmenting human fig-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29051" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 71, + 553, + 284 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 553, + 284 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 553, + 284 + ], + "type": "image", + "image_path": "7e412135a17f8098fd8ef3f6745550ec7642f0351cc698c3e24509721f75ac04.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 286, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 54, + 286, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 54, + 286, + 555, + 331 + ], + "type": "text", + "content": "Figure 3. The proposed UTD method involves generating textual descriptions of different concepts in video frames using VLMs and LLMs, combining them in various temporal configurations, and evaluating the performance of these concept-temporal representations with strong text embedding models. For each representation, we distinguish between common sense bias, which relies on zero-shot reasoning by text embedding models, and dataset bias, assessed using a linear model trained on the dataset's training set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 338, + 296, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 296, + 708 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 296, + 708 + ], + "type": "text", + "content": "ures and placing them into black/gray background, or alternatively by cropping or inpainting human figures and classifying videos based on background only. Weinzaepfel et al. [47] proposed the Mimetics dataset by extending certain classes from the Kinetics dataset [17] with videos of activities in unrelated background. Zhai et al. [51] proposed an approach for scene-debiased action recognition via adversarial scene reconstruction and classification. Another line of work focused on capturing datasets with explicitly low object and scene bias. SomethingSomethingv1 and later v2 dataset [14] were proposed as an almost scene-free dataset, focusing on simple manipulative actions only. Li et al. [23] proposed a procedure for designing representation unbiased datasets and introduced Diving48 as an example. However, this specialization usually also implies that those cases are rather suitable for specialized approaches but not for, e.g., zero-shot evaluation of generalist video models. Other studies [5, 19] showed that for several benchmarks, using single frames during training or inference can achieve performance comparable to models that incorporate temporal information. However, recent efforts have mainly focused on developing new benchmarks to assess temporal understanding in videos [2], e.g., visual question-answering benchmarks [7, 28], while temporal representation bias in standard benchmarks has not been thoroughly investigated. In contrast, our UTD method allows for a systematic analysis of representation biases across various dimensions, including different concept and temporal configurations. Additionally, we are the first to propose mitigating object representation bias in popular benchmarks by discarding object-biased samples, which helps us to avoid adding artifacts to" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 338, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 361 + ], + "type": "text", + "content": "test samples (as with cropping or inpainting in [11, 15]) and keeps the dataset domain unconstrained, unlike [14, 23]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 361, + 556, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 556, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 556, + 540 + ], + "type": "text", + "content": "LLMs for Video Understanding. The progress of LLMs prompted a wide range of applications in various fields, including video understanding. Here, one line of work utilizes pre-trained LLMs in video-language tasks, such as video captioning [37, 46, 53], video question answering [21, 27, 46, 53], text-video retrieval [35, 53], or dialog-based user interaction [46, 52]. Some works [10, 34, 44] create new video-language datasets using image/audio captioning models or subtitles and using LLM summarization and reasoning ability to post-process these captions. Other works [20, 39, 45, 50] use expert perception models to extract information from the video that is later processed by the LLM for tasks such as video captioning. In our work, we leverage VLM- and LLM-generated frame descriptions to evaluate different representation biases in video datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 551, + 506, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 506, + 564 + ], + "type": "text", + "content": "3. Discovering Representation Biases" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 570, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 605 + ], + "type": "text", + "content": "In this section, we define representation bias in Sec. 3.1, outline the considered representation biases in Sec. 3.2, and detail our UTD method in Sec. 3.3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 613, + 485, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 485, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 485, + 625 + ], + "type": "text", + "content": "3.1. Definition: Representation Bias" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 630, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 653 + ], + "type": "text", + "content": "To formalize the definition of representation bias in datasets, we adopt the notation from RESOUND [23]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "D = \\{v_{n}\\}_{n\\leq N}" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": " be a video dataset " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": " where each video " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "v_{n}" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": " consists of a sequence of frames " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "v_{n} = \\{f_{n,i}\\}_{i\\leq |v_n|}" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": ". A representation " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": " is a function " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\phi (v)" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": " of the video input " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "R = \\{\\phi \\}" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": " be a family of representations that share specific characteristics. For example, a fam" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29052" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "content": "ily of single-frame representations includes representations that encode information from only a single frame of a video " + }, + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "inline_equation", + "content": "\\phi (v_{n}) = \\phi (f_{n,i})" + }, + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "inline_equation", + "content": "i\\leq |v_{n}|" + }, + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "content": " or a family of object-based representations encodes only the objects in the video without considering motion or activity information." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 295, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 181 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 181 + ], + "type": "text", + "content": "The representation " + }, + { + "bbox": [ + 55, + 133, + 295, + 181 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 133, + 295, + 181 + ], + "type": "text", + "content": " can be used to train different models " + }, + { + "bbox": [ + 55, + 133, + 295, + 181 + ], + "type": "inline_equation", + "content": "\\gamma_{\\phi}" + }, + { + "bbox": [ + 55, + 133, + 295, + 181 + ], + "type": "text", + "content": " for desired tasks, e.g. video classification. Let " + }, + { + "bbox": [ + 55, + 133, + 295, + 181 + ], + "type": "inline_equation", + "content": "M(v_{n},\\gamma_{\\phi})" + }, + { + "bbox": [ + 55, + 133, + 295, + 181 + ], + "type": "text", + "content": " be the performance (e.g., accuracy) of a model " + }, + { + "bbox": [ + 55, + 133, + 295, + 181 + ], + "type": "inline_equation", + "content": "\\gamma_{\\phi}" + }, + { + "bbox": [ + 55, + 133, + 295, + 181 + ], + "type": "text", + "content": " on a video " + }, + { + "bbox": [ + 55, + 133, + 295, + 181 + ], + "type": "inline_equation", + "content": "v_{n}" + }, + { + "bbox": [ + 55, + 133, + 295, + 181 + ], + "type": "text", + "content": ", then the performance of " + }, + { + "bbox": [ + 55, + 133, + 295, + 181 + ], + "type": "inline_equation", + "content": "\\gamma_{\\phi}" + }, + { + "bbox": [ + 55, + 133, + 295, + 181 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 55, + 133, + 295, + 181 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 133, + 295, + 181 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 108, + 185, + 295, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 185, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 108, + 185, + 295, + 217 + ], + "type": "interline_equation", + "content": "M (D, \\gamma_ {\\phi}) = \\frac {1}{N} \\sum_ {n = 1} ^ {N} M \\left(v _ {n}, \\gamma_ {\\phi}\\right). \\tag {1}", + "image_path": "4a79f03bec4e11454d4158e9e1508c1993e263a6249efeb2e56c39dfbafc25d3.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 220, + 286, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 220, + 286, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 286, + 232 + ], + "type": "text", + "content": "We define the performance of a representation " + }, + { + "bbox": [ + 55, + 220, + 286, + 232 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 220, + 286, + 232 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 55, + 220, + 286, + 232 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 220, + 286, + 232 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 238, + 295, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 238, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 116, + 238, + 295, + 255 + ], + "type": "interline_equation", + "content": "M (D, \\phi) = \\max _ {\\gamma_ {\\phi}} M (D, \\gamma_ {\\phi}), \\tag {2}", + "image_path": "716d59ca687adc24c14d572a5da6751876969a9d955ecf39d0dd6489b5cdee2a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 260, + 295, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 260, + 295, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 260, + 295, + 283 + ], + "type": "text", + "content": "where the maximum is taken over all models " + }, + { + "bbox": [ + 55, + 260, + 295, + 283 + ], + "type": "inline_equation", + "content": "\\gamma_{\\phi}" + }, + { + "bbox": [ + 55, + 260, + 295, + 283 + ], + "type": "text", + "content": " that are based on the representation " + }, + { + "bbox": [ + 55, + 260, + 295, + 283 + ], + "type": "inline_equation", + "content": "\\phi" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 284, + 295, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 295, + 511 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 295, + 511 + ], + "type": "text", + "content": "When the performance " + }, + { + "bbox": [ + 55, + 284, + 295, + 511 + ], + "type": "inline_equation", + "content": "M(D, \\phi)" + }, + { + "bbox": [ + 55, + 284, + 295, + 511 + ], + "type": "text", + "content": " of a representation " + }, + { + "bbox": [ + 55, + 284, + 295, + 511 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 284, + 295, + 511 + ], + "type": "text", + "content": " is high, the dataset exhibits a preference for this representation, indicating that the dataset is more biased toward this representation. In this work, we use " + }, + { + "bbox": [ + 55, + 284, + 295, + 511 + ], + "type": "inline_equation", + "content": "M(D, \\phi)" + }, + { + "bbox": [ + 55, + 284, + 295, + 511 + ], + "type": "text", + "content": " as a measure of bias towards representation " + }, + { + "bbox": [ + 55, + 284, + 295, + 511 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 284, + 295, + 511 + ], + "type": "text", + "content": ". However, a high bias toward a specific representation is not necessarily a negative thing. Ideally, a dataset should exhibit the highest bias toward the \"ground truth representation\": the representation truly required to solve the task. For instance, in video understanding, one could argue that the optimal representation should capture both static (e.g., objects) and dynamic (e.g., actions) properties, as well as encode temporal aspects. However, if a video dataset used to evaluate video understanding exhibits a strong object or single-frame representation bias, it may unintentionally favor models with strong object recognition or image understanding capabilities, thus hindering the evaluation of true video understanding. Therefore, in this work, we aim to estimate representation biases in popular video benchmarks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 517, + 226, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 226, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 226, + 529 + ], + "type": "text", + "content": "3.2. Different Representation Biases" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 534, + 295, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 557 + ], + "type": "text", + "content": "In our work, we consider different types of representation biases with respect to three independent aspects:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 558, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 558, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 54, + 558, + 295, + 677 + ], + "type": "text", + "content": "Concept Representations: First, we consider representations that encode four different concept categories: objects " + }, + { + "bbox": [ + 54, + 558, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\phi_{o}(v_{n})" + }, + { + "bbox": [ + 54, + 558, + 295, + 677 + ], + "type": "text", + "content": ", activities " + }, + { + "bbox": [ + 54, + 558, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\phi_{a}(v_{n})" + }, + { + "bbox": [ + 54, + 558, + 295, + 677 + ], + "type": "text", + "content": " (actions along with tied objects), verbs " + }, + { + "bbox": [ + 54, + 558, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\phi_{\\nu}(v_{n})" + }, + { + "bbox": [ + 54, + 558, + 295, + 677 + ], + "type": "text", + "content": " (actions alone, without objects), and objects+composition+activities " + }, + { + "bbox": [ + 54, + 558, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\phi_{d}(v_{n})" + }, + { + "bbox": [ + 54, + 558, + 295, + 677 + ], + "type": "text", + "content": " (objects, their spatial compositions in the video, and activities), see an example in Fig. 3. We aim to assess biases in classification and retrieval benchmarks toward particular concepts. For example, in retrieval, we ideally would like to see that benchmarks are not biased solely toward objects or activities." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Temporal Representations: Second, we consider the bias of single-frame representations compared to representations that consider multiple frames. Let " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\phi_x" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " be" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 266 + ], + "type": "text", + "content": "one of the objects " + }, + { + "bbox": [ + 313, + 72, + 553, + 266 + ], + "type": "inline_equation", + "content": "\\phi_o" + }, + { + "bbox": [ + 313, + 72, + 553, + 266 + ], + "type": "text", + "content": ", activities " + }, + { + "bbox": [ + 313, + 72, + 553, + 266 + ], + "type": "inline_equation", + "content": "\\phi_a" + }, + { + "bbox": [ + 313, + 72, + 553, + 266 + ], + "type": "text", + "content": ", verbs " + }, + { + "bbox": [ + 313, + 72, + 553, + 266 + ], + "type": "inline_equation", + "content": "\\phi_\\nu" + }, + { + "bbox": [ + 313, + 72, + 553, + 266 + ], + "type": "text", + "content": ", or objects+composition+activities " + }, + { + "bbox": [ + 313, + 72, + 553, + 266 + ], + "type": "inline_equation", + "content": "\\phi_d" + }, + { + "bbox": [ + 313, + 72, + 553, + 266 + ], + "type": "text", + "content": " representations. Similarly to [41], we consider the following options to encode information from multiple frames: middle-frame refers to the case that a video is represented only by a single middle frame of this video, e.g., " + }, + { + "bbox": [ + 313, + 72, + 553, + 266 + ], + "type": "inline_equation", + "content": "\\phi_{x,\\mathrm{mid}}(v_n) = \\phi_x(f_{n,|v_n| / 2})" + }, + { + "bbox": [ + 313, + 72, + 553, + 266 + ], + "type": "text", + "content": ", max-score frame refers to the case that a video is represented using a single frame " + }, + { + "bbox": [ + 313, + 72, + 553, + 266 + ], + "type": "inline_equation", + "content": "\\phi_{x,\\max}(v_n) = \\phi_x(f_{n,i})" + }, + { + "bbox": [ + 313, + 72, + 553, + 266 + ], + "type": "text", + "content": " where the frame " + }, + { + "bbox": [ + 313, + 72, + 553, + 266 + ], + "type": "inline_equation", + "content": "f_{n,i}" + }, + { + "bbox": [ + 313, + 72, + 553, + 266 + ], + "type": "text", + "content": " has the highest performance score " + }, + { + "bbox": [ + 313, + 72, + 553, + 266 + ], + "type": "inline_equation", + "content": "M(f_{n,i},\\phi_x)" + }, + { + "bbox": [ + 313, + 72, + 553, + 266 + ], + "type": "text", + "content": " over all frames, average-overframes refers to the case that the video is represented as an average representation of frames ignoring the order of frames " + }, + { + "bbox": [ + 313, + 72, + 553, + 266 + ], + "type": "inline_equation", + "content": "\\phi_{x,\\mathrm{avg}}(v_n) = \\frac{1}{|v_n|}\\sum_{i\\leq |v_n|}\\phi_x(f_{n,i})" + }, + { + "bbox": [ + 313, + 72, + 553, + 266 + ], + "type": "text", + "content": ", and, finally, sequence-of-frames refers to the case that the video representation encodes concepts and their order in the frames " + }, + { + "bbox": [ + 313, + 72, + 553, + 266 + ], + "type": "inline_equation", + "content": "\\phi_{x,\\mathrm{seq}}(v_n) = \\phi_x(f_{n,1},f_{n,2},..,f_{n,|v_n|})" + }, + { + "bbox": [ + 313, + 72, + 553, + 266 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 266, + 553, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 553, + 481 + ], + "type": "text", + "content": "Common Sense Bias vs. Dataset Bias: Finally, for any concept-temporal representation " + }, + { + "bbox": [ + 313, + 266, + 553, + 481 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 266, + 553, + 481 + ], + "type": "text", + "content": " (e.g., object middle-frame representation " + }, + { + "bbox": [ + 313, + 266, + 553, + 481 + ], + "type": "inline_equation", + "content": "\\phi_{o,mid}" + }, + { + "bbox": [ + 313, + 266, + 553, + 481 + ], + "type": "text", + "content": "), we can consider two types of representation biases: common sense bias or dataset bias. Common sense representation bias refers to bias based on \"common sense reasoning\", e.g., a video featuring an object \"piano\" would most likely be classified under the \"playing piano\" activity label. Formally, in Eq. (2), this bias " + }, + { + "bbox": [ + 313, + 266, + 553, + 481 + ], + "type": "inline_equation", + "content": "M(D,\\phi)" + }, + { + "bbox": [ + 313, + 266, + 553, + 481 + ], + "type": "text", + "content": " is measured by taking the maximum performance only over zero-shot models " + }, + { + "bbox": [ + 313, + 266, + 553, + 481 + ], + "type": "inline_equation", + "content": "\\gamma_{\\phi}" + }, + { + "bbox": [ + 313, + 266, + 553, + 481 + ], + "type": "text", + "content": " that have not seen the correlations in the training set of " + }, + { + "bbox": [ + 313, + 266, + 553, + 481 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 266, + 553, + 481 + ], + "type": "text", + "content": ". We distinguish this from dataset bias, which refers to biases caused by spurious correlations within the dataset, e.g., where certain objects are unintentionally associated with specific labels, such as \"mirror\" or \"flower bouquet\" frequently appear with the \"apply makeup\" label. In this case, in Eq. (2), we measure dataset bias by taking the maximum performance over models " + }, + { + "bbox": [ + 313, + 266, + 553, + 481 + ], + "type": "inline_equation", + "content": "\\gamma_{\\phi}" + }, + { + "bbox": [ + 313, + 266, + 553, + 481 + ], + "type": "text", + "content": " trained or fine-tuned on the training set of " + }, + { + "bbox": [ + 313, + 266, + 553, + 481 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 266, + 553, + 481 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 482, + 553, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 553, + 540 + ], + "type": "text", + "content": "We note that the concept, temporal, and common sense vs. dataset bias dimensions are independent, allowing us to combine them in any configuration. For example, we can evaluate object middle-frame common sense bias or activity sequence-of-frames dataset bias." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 551, + 553, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 553, + 563 + ], + "type": "text", + "content": "3.3. Unbiasing through Textual Descriptions (UTD)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 569, + 553, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 553, + 628 + ], + "type": "text", + "content": "Since measuring bias for every representation within a family (e.g., a family of object representations) is infeasible, we approximate it by considering a single representative from the family [23] and propose a way to precisely control what each representation encodes by using textual descriptions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": "The key idea of our method is to use textual descriptions as an intermediate video representation. This approach allows us to easily control which concepts in the video are included or excluded in the representation and to combine information into different temporal representations. Namely, to obtain representation " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": " that encodes only a specific concept category, we propose to encode visible concepts of this" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29053" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "content": "category into textual descriptions, e.g., encode objects with a list of visible objects, e.g., \"table, lamp, ..\". This approach, compared to cropping or inpainting [11, 15], allows us to isolate specific concepts more precisely, avoiding information leakage from cropped areas or adding artifacts caused by imperfect inpainting. Additionally, textual descriptions allow us to use off-the-shelf LLMs with strong common sense reasoning abilities to estimate common sense bias. Since metadata with visible objects, activities, etc., is usually not available, we propose using recent powerful VLMs and LLMs to extract this information." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 204, + 294, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 294, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 294, + 276 + ], + "type": "text", + "content": "In the next sections, we first describe our method of obtaining textual description with respect to different concept and temporal categories (Sec. 3.3.1), then we describe our method of discovering common sense and dataset biases (Sec. 3.3.2), and lastly, we describe our method of unbiaseding benchmarks (Sec. 3.3.3)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 280, + 217, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 280, + 217, + 292 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 217, + 292 + ], + "type": "text", + "content": "3.3.1. Obtaining Textual Descriptions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 295, + 295, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 295, + 390 + ], + "type": "text", + "content": "While current VLMs [24-26], trained to perform complex image reasoning, are able to generate detailed descriptions of video frames, they still face challenges when prompted to directly extract specific concepts from a frame, e.g., only visible objects. Therefore, we decomposed the task of concept extraction into two parts: extraction of detailed frame description using VLMs and then extraction of different concepts from frame description using LLMs (see Fig. 3)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 391, + 295, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 295, + 558 + ], + "type": "text", + "content": "Obtaining Textual Descriptions Using VLMs. Given train/test videos " + }, + { + "bbox": [ + 55, + 391, + 295, + 558 + ], + "type": "inline_equation", + "content": "v_{n} = \\{f_{n,i}\\}_{i\\leq |v_{n}|}, n < N" + }, + { + "bbox": [ + 55, + 391, + 295, + 558 + ], + "type": "text", + "content": " of the dataset, we feed each frame " + }, + { + "bbox": [ + 55, + 391, + 295, + 558 + ], + "type": "inline_equation", + "content": "f_{n,i}" + }, + { + "bbox": [ + 55, + 391, + 295, + 558 + ], + "type": "text", + "content": " of each video " + }, + { + "bbox": [ + 55, + 391, + 295, + 558 + ], + "type": "inline_equation", + "content": "v_{n}" + }, + { + "bbox": [ + 55, + 391, + 295, + 558 + ], + "type": "text", + "content": " into the VLM model " + }, + { + "bbox": [ + 55, + 391, + 295, + 558 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 391, + 295, + 558 + ], + "type": "text", + "content": " to generate a detailed textual description " + }, + { + "bbox": [ + 55, + 391, + 295, + 558 + ], + "type": "inline_equation", + "content": "d_{n,i} = d(f_{n_i})" + }, + { + "bbox": [ + 55, + 391, + 295, + 558 + ], + "type": "text", + "content": " (Fig. 3). We found that without complex prompt engineering or few-shot in-context learning strategies [4] VLMs produce long, detailed descriptions. While our method is not restricted to a specific VLM, for reproducibility and feasibility, we select an open-sourced strong middle-scale VLM model, LLaVA-1.6-Mistral-7B [26] with the prompt: \"Describe the objects' relationships in the photo.\" Since the resulting textual descriptions include objects, their composition, relationships, and visible activities, we use them as objects+composition+activities descriptions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "text", + "content": "Extracting Concepts Categories Using LLMs. Each of the obtained detailed descriptions " + }, + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "inline_equation", + "content": "d_{n,i}" + }, + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "text", + "content": " is further fed to the LLM " + }, + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "text", + "content": " to extract only objects " + }, + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "inline_equation", + "content": "o_{n,i} = \\zeta(d_{n,i}, p_o)" + }, + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "text", + "content": ", only activities " + }, + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "inline_equation", + "content": "a_{n,i} = \\zeta(d_{n,i}, p_a)" + }, + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "text", + "content": " information, with " + }, + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "inline_equation", + "content": "p_o, p_a" + }, + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "text", + "content": " as respective prompts. Then, to obtain information about verbs, we task the LLM with a prompt " + }, + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "inline_equation", + "content": "p_\\nu" + }, + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "text", + "content": " to delete all object information from activities " + }, + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\nu_{n,i} = \\zeta(a_{n,i}, p_\\nu)" + }, + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "text", + "content": ". Since such text processing requires understanding what objects and activities are, our prompt includes a few examples following a few-shot in-context learning strategy [4]. We provide all prompts in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Aggregating Temporal Information. We use the obtained textual descriptions " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "x_{n,i}" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " of video " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "v_{n}" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "x_{n,i}" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " is one" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "text", + "content": "of " + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "inline_equation", + "content": "o_{n,i}" + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "inline_equation", + "content": "a_{n,i}" + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "inline_equation", + "content": "\\nu_{n,i}" + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "text", + "content": ", or " + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "inline_equation", + "content": "d_{n,i}" + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "text", + "content": ") in different temporal setups: first, we keep them as single frame descriptions that might be further aggregated when transformed into vector representations (e.g. average-over-frames), as we describe in Sec. 3.3.2, second, we combine them into sequence-over-frame " + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "inline_equation", + "content": "x_{n}" + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "text", + "content": " setup with the template: \"Frame 1: " + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "inline_equation", + "content": "x_{n,1}" + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "text", + "content": " Frame 2: " + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "inline_equation", + "content": "x_{n,2} \\ldots" + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "text", + "content": " Frame " + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "inline_equation", + "content": "|v_{n}| \\colon x_{n,|v_{n}|}" + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "text", + "content": "\" (Fig. 3). Since objects+composition+activities descriptions " + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "inline_equation", + "content": "d_{n,i}" + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "text", + "content": " when combined with sequence-over-frame templates become too long, we summarize them to 15-word descriptions " + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "inline_equation", + "content": "d_{n,i}'" + }, + { + "bbox": [ + 313, + 72, + 553, + 205 + ], + "type": "text", + "content": " via an LLM before inputting them into the sequence prompt." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 210, + 488, + 222 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 210, + 488, + 222 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 488, + 222 + ], + "type": "text", + "content": "3.3.2. Calculating Representation Biases" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 225, + 553, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 225, + 553, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 553, + 344 + ], + "type": "text", + "content": "After obtaining a specific concept-temporal representation " + }, + { + "bbox": [ + 313, + 225, + 553, + 344 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 225, + 553, + 344 + ], + "type": "text", + "content": ", we assess common sense and dataset bias for dataset " + }, + { + "bbox": [ + 313, + 225, + 553, + 344 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 225, + 553, + 344 + ], + "type": "text", + "content": " toward " + }, + { + "bbox": [ + 313, + 225, + 553, + 344 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 225, + 553, + 344 + ], + "type": "text", + "content": " by measuring the performance " + }, + { + "bbox": [ + 313, + 225, + 553, + 344 + ], + "type": "inline_equation", + "content": "M(D, \\phi)" + }, + { + "bbox": [ + 313, + 225, + 553, + 344 + ], + "type": "text", + "content": ", as stated Eq. (2), where maximum is taken over zero-shot models " + }, + { + "bbox": [ + 313, + 225, + 553, + 344 + ], + "type": "inline_equation", + "content": "\\gamma_{\\phi}" + }, + { + "bbox": [ + 313, + 225, + 553, + 344 + ], + "type": "text", + "content": " for common sense bias and over fine-tuned models " + }, + { + "bbox": [ + 313, + 225, + 553, + 344 + ], + "type": "inline_equation", + "content": "\\gamma_{\\phi}" + }, + { + "bbox": [ + 313, + 225, + 553, + 344 + ], + "type": "text", + "content": " for dataset bias (see Sec. 3.2). Since maximizing over all models is infeasible in practice, we design strong models for common sense and dataset-based reasoning by utilizing state-of-the-art LLMs. Our ablation study verifies our design choice (see supplementary material)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 345, + 554, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 554, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 554, + 524 + ], + "type": "text", + "content": "Common Sense Bias. We leverage recent strong LLM-based text embedding models [30, 43]. Such models are trained on large datasets to encode text into vector representations that can be used for various tasks, such as information retrieval. We encode our textual descriptions and all class labels/captions using the pre-trained text embedding model " + }, + { + "bbox": [ + 313, + 345, + 554, + 524 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 313, + 345, + 554, + 524 + ], + "type": "text", + "content": ". Specifically, we prompt the model with the relevant concept and task. For example, we use prompts like \"Given a list of objects visible on the video frame, retrieve the activity depicted in this video\" for object description and \"Given an activity, retrieve a video frame description that may depict this activity\" for class labels in action classification. We use cosine similarity between obtained embeddings for zero-shot classification/retrieval. Prompts are detailed in the supplement." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 525, + 553, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 553, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 553, + 608 + ], + "type": "text", + "content": "Dataset Bias. For action classification datasets, we evaluate dataset bias, as action classification models are often tested in fine-tuned setups and might exploit dataset representation biases learned from training data. For this we a linear logistic regression model trained on our strong text embeddings used in the common sense bias analysis and train model on the official train splits of the datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 609, + 553, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 553, + 644 + ], + "type": "text", + "content": "Metrics. For the performance measure " + }, + { + "bbox": [ + 313, + 609, + 553, + 644 + ], + "type": "inline_equation", + "content": "M(D, \\phi)" + }, + { + "bbox": [ + 313, + 609, + 553, + 644 + ], + "type": "text", + "content": ", we use accuracy for action classification and recall@1 and text-to-video retrieval as the most popular metrics for these tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 651, + 424, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 651, + 424, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 651, + 424, + 662 + ], + "type": "text", + "content": "3.3.3. Unbiasing Datasets" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "In Sec. 4, we find that many video datasets exhibit a significant object representation bias. To support robust video model evaluation, we construct UTD-splits, object-debiased test splits, by excluding object-biased samples from the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29054" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 558, + 125 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 558, + 125 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 558, + 125 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 558, + 125 + ], + "type": "table", + "html": "
UCFSSv2K400K600K700MiTMSRDDMActNLSMDCYC2S-MiT
obj+comp+act66.36.448.044.139.022.636.627.226.617.08.445.9
objects63.3-3.05.3-1.145.9-2.141.8-2.337.0-2.021.0-1.632.1-4.527.0-0.224.8-1.813.6-3.47.9-0.529.8-16.1
activities67.4+1.16.445.2-2.841.4-2.736.7-2.321.0-1.625.1-11.521.1-6.121.4-5.214.7-2.38.1-0.341.1-4.8
verbs50.8-15.55.8-0.624.8-23.221.4-22.717.6-21.416.2-6.410.5-26.17.0-20.27.4-19.25.5-11.51.2-7.213.1-32.8
", + "image_path": "02ddb50ab1245ecda2edb9771f14246906ee39074b27cefe88a91f7e3b8047b5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 129, + 555, + 174 + ], + "lines": [ + { + "bbox": [ + 55, + 129, + 555, + 174 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 555, + 174 + ], + "type": "text", + "content": "Table 1. Evaluation of concept representation bias. We report the performance of our model in sequence-of-frames temporal setup and common sense bias setup and color code with respect to the difference to objects+composition+activities (obj+comp+act) concepts. It shows that the overall classification performance drops only slightly when classifying actions based on objects, compared to classification and retrieval on verbs only, with SSv2 as the only dataset with less drop when classifying on verbs only." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 188, + 294, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 188, + 294, + 391 + ], + "spans": [ + { + "bbox": [ + 54, + 188, + 294, + 391 + ], + "type": "text", + "content": "original test sets. Since our method allows us to measure the performance of the representation " + }, + { + "bbox": [ + 54, + 188, + 294, + 391 + ], + "type": "inline_equation", + "content": "M(v_{n},\\phi)" + }, + { + "bbox": [ + 54, + 188, + 294, + 391 + ], + "type": "text", + "content": " on individual samples " + }, + { + "bbox": [ + 54, + 188, + 294, + 391 + ], + "type": "inline_equation", + "content": "v_{n}" + }, + { + "bbox": [ + 54, + 188, + 294, + 391 + ], + "type": "text", + "content": ", we exclude samples that are classified or retrieved correctly (" + }, + { + "bbox": [ + 54, + 188, + 294, + 391 + ], + "type": "inline_equation", + "content": "M(v_{n},\\phi) = 1" + }, + { + "bbox": [ + 54, + 188, + 294, + 391 + ], + "type": "text", + "content": ") based on object sequence-of-frame representation. For text-to-video retrieval, we debias with respect to common-sense bias, but for action classification, we debias with respect to dataset bias since video models are mostly tested in fine-tuned setup. To minimize the impact of random fluctuations in the process, we use three different prompts for labels/captions in the text embedding models and train three linear models for each classification dataset using bootstrapped training sets (we provide details and analysis in the supplement) and exclude only samples when all nine models agreed on correct classification/retrieval. Note that the percentage of removed samples is determined automatically based on the extent of object representation bias." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 392, + 295, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 295, + 477 + ], + "type": "text", + "content": "Finally, since debiasing may disproportionately remove certain label classes in classification datasets, we additionally construct balanced UTD splits. While maintaining the total number of removed samples as in the original debiasing method, we adjust the number of samples removed from each class based on their average confidence (across nine models) to preserve the original class proportions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 487, + 194, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 194, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 194, + 501 + ], + "type": "text", + "content": "4. Benchmarking Datasets" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 508, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 295, + 544 + ], + "type": "text", + "content": "In this section, we analyze action classification and text-to-video retrieval datasets on different representation biases using our UTD method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 552, + 105, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 105, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 105, + 565 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "text", + "content": "Datasets. We consider six common action classification datasets: UCF101 (UCF) [36], SomethingSomethingv2 (SSv2) [14], Kinetics400 (K400) [17], -600 (K600) [8], -700 (K700) [9], Moments In Time (MiT) [31] and six text-to-video retrieval datasets: MSRVTT (MSR) [48], YouCook2 (YC2) [12], DiDeMo (DDM) [1], LSMDC [33], ActivityNet (ANet) [6], and Spoken Moments in Time (S-MiT) [32]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Implementation Details. For all train/test videos, we use " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "|v_{n}| = 8" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " uniformly sampled frames. We use the LLaVA-1.6-7B-mistral model [26] as VLM, Mistral-7B-Instructv0.2 [16] as LLM, and SFR-Employing-Mistral [30] as text" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 188, + 503, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 188, + 503, + 200 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 503, + 200 + ], + "type": "text", + "content": "embedding model. The temperature is set to 0." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 211, + 421, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 421, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 421, + 224 + ], + "type": "text", + "content": "4.2. UTD-descriptions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 229, + 555, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 229, + 555, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 555, + 504 + ], + "type": "text", + "content": "Following our UTD pipeline, first, we generate textual descriptions for 12 considered datasets, resulting in UTD-descriptions dataset (see qualitative examples in the supplement). We conducted a user study to validate the quality of object, activity, and verb predictions. Participants were asked to verify which objects from a provided list were visible in a given image. The study used 100 random frames from 100 videos of the UCF dataset (which has noisy, low-resolution frames that might be an additional challenge to the VLM). Seven participants validated 761 annotated objects in total, where " + }, + { + "bbox": [ + 313, + 229, + 555, + 504 + ], + "type": "inline_equation", + "content": "87.6\\%" + }, + { + "bbox": [ + 313, + 229, + 555, + 504 + ], + "type": "text", + "content": " (667 out of 761) were identified as visible, with only 94 not selected as visible. To better understand the VLM's errors, we manually classified these 94 objects into five categories: 1) attribute error (e.g., the object is \"right hand\" instead of \"left hand\"), 2) misclassification (the object is present but incorrectly identified as a different object, e.g., a \"snowboard\" instead of \"snow slide\"), 3) hallucination, 4) human annotation mistake – the object is visible, and 5) other. We found that only 45 objects were classified as hallucinations, 13 as misclassifications, 20 as \"object visible\" 8 as attribute errors, and 8 as other (see examples in the supplement). This results in a hallucination rate of only " + }, + { + "bbox": [ + 313, + 229, + 555, + 504 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 313, + 229, + 555, + 504 + ], + "type": "text", + "content": ", which we find acceptable for our analysis." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 516, + 473, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 473, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 473, + 528 + ], + "type": "text", + "content": "4.3. Concept Representation Bias" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "Further, based on generated descriptions, we evaluate representation bias with respect to four concept categories, namely objects, activities, verbs, and objects+composition+activities, as described in Sec. 3.2. For this, we consider common sense bias in sequence-of-frames temporal setup in Tab. 1 (all concept-temporal combination results can be found in the supplement). It shows that using only object information results in only a slight drop in performance compared to objects+composition+activities, indicating that most datasets can, to a certain extent, be addressed with object information, thus showing an object representation bias. Interestingly, classification based on only objects results in 63.3pp accuracy for UCF101 and 45.9pp for Kinetics400. This may be due to the fact that some action classes are tied to the objects (e.g. \"playing" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29055" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 71, + 553, + 114 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 553, + 114 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 553, + 114 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 553, + 114 + ], + "type": "table", + "html": "
UCFSSv2K400K600K700MiT
obj.act.verbsobj.act.verbsobj.act.verbsobj.act.verbsobj.act.verbsobj.act.verbs
common sense b.63.367.450.85.36.45.845.945.224.841.841.421.437.036.717.621.021.016.2
dataset bias80.3+17.083.4+16.073.5+22.713.6+8.316.0+9.612.2+6.460.7+14.860.7+15.542.2+17.460.4+18.660.3+18.942.5+21.152.3+15.349.7+13.034.9+17.327.8+6.827.3+6.322.9+6.7
", + "image_path": "66fdd7ef658d5fa65d841de94e5476bd8ed6fbb7ed4ab1b4b9a79018d4924662.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 158, + 294, + 216 + ], + "blocks": [ + { + "bbox": [ + 55, + 119, + 555, + 153 + ], + "lines": [ + { + "bbox": [ + 55, + 119, + 555, + 153 + ], + "spans": [ + { + "bbox": [ + 55, + 119, + 555, + 153 + ], + "type": "text", + "content": "Table 2. Comparison of common sense bias vs dataset bias. We report the performance of our model in sequence of frame temporal setup and color code with respect to difference to common sense bias. It shows that the overall performance growth is significant for all concepts, including object concepts, indicating that datasets might contain a significant correlation between labels and visible objects." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 158, + 294, + 216 + ], + "lines": [ + { + "bbox": [ + 56, + 158, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 158, + 294, + 216 + ], + "type": "table", + "html": "
UCF SSv2 K400 K600 K700 MiTMSR DDM ActN LSMDC YC2 S-MiT
middle-frame61.36.039.735.631.120.123.419.813.512.76.135.5
max-score frame66.57.448.043.338.721.433.429.121.518.38.946.1
avg.-over-frames66.76.846.642.137.123.431.926.526.616.48.643.8
seq.-of-frames66.36.448.044.139.022.636.627.226.617.08.445.9
", + "image_path": "ce33ea7a5e0c211572ceb598acb70148d07e359227cf812b979a188491cd3411.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 259, + 294, + 316 + ], + "blocks": [ + { + "bbox": [ + 55, + 220, + 295, + 253 + ], + "lines": [ + { + "bbox": [ + 55, + 220, + 295, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 295, + 253 + ], + "type": "text", + "content": "Table 3. Evaluation of temporal representation bias. We report the performance of our model in objects+composition+activities concepts setup and common sense bias setup." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 259, + 294, + 316 + ], + "lines": [ + { + "bbox": [ + 56, + 259, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 259, + 294, + 316 + ], + "type": "table", + "html": "
UCF SSv2 K400 K600 K700 MiTMSR DDM ActN LSMDC YC2 S-MiT
Fleiss' kappa86.177.891.290.792.591.292.091.790.992.591.393.2
% object-biased samples7375555482428232011636
% of full test split in UTD-split279345455276727780899464
", + "image_path": "f6ef2d9fc0ca893e6ac72ad2b20b8722a8a92129793c7a83f45aa1fe829695a0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 319, + 295, + 343 + ], + "lines": [ + { + "bbox": [ + 55, + 319, + 295, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 295, + 343 + ], + "type": "text", + "content": "Table 4. UTD-splits statistics. Fleiss' kappa coefficient measures the agreement among 9 models used to create UTD-splits." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 350, + 295, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 350, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 54, + 350, + 295, + 458 + ], + "type": "text", + "content": "piano\"), making visible objects equally or more reliable indicators of a class than activities. This becomes clear when we consider the verbs-based representation. Here, the performance drops significantly for most action classification datasets, except for SSv2 and MiT. Also, in the case of text-to-video retrieval, it shows that objects are better indicators than activities for some datasets (MSRVTT, DiDeMo, ActivityNet), and activities are better indicators for others (YouCook2, LSMDC, S-MiT)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 466, + 228, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 228, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 228, + 478 + ], + "type": "text", + "content": "4.4. Common Sense vs. Dataset Bias" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 483, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 483, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 54, + 483, + 295, + 651 + ], + "type": "text", + "content": "We further consider common sense bias compared to dataset bias for the action classification datasets in Tab. 2. Overall, across all benchmarks and concepts, dataset bias is significantly higher than common sense bias, and the results indicate that considered benchmarks also contain a dataset-specific object bias in the training set. The performance on UCF using only objects increases from 63.3pp to 80.3pp, and even on SSv2, which was developed to minimize object representation bias, using only correlation of seen objects in the train set to the classes, we were able to predict 13.6pp of videos correctly. This significant increase based on available object information only shows that some actions might be connected to unrelated objects that hint at a certain class without being directly associated with it." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 659, + 222, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 222, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 222, + 673 + ], + "type": "text", + "content": "4.5. Temporal Representation Bias" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Finally, we consider representation bias with respect to the temporal aspect in Tab. 3. Namely, we consider which temporal representations, single frame vs. multiple frames," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 159, + 555, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 159, + 555, + 256 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 555, + 256 + ], + "type": "text", + "content": "best-score frame vs. middle frame, as well as average-over-frames vs. sequence-of-frames embeddings, perform best. For all datasets except ActivityNet, max-score-frame performance is comparable (< 3.5% difference) to multiple frames performance (average of frames/sequence over frames). Moreover, for SSv2, Kinetics400, DiDeMo, LSMDC, YouCook2, and S-MiT, we found that max-score-frame is a better indicator of class/captions than all frames." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 264, + 389, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 389, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 389, + 277 + ], + "type": "text", + "content": "4.6.UTD-splits" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 282, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 555, + 415 + ], + "type": "text", + "content": "Since our analysis reveals significant object bias in benchmarks, we generate object-debiased UTD-splits for considered benchmarks, as described in Sec. 3.3.3. Namely, we remove object-biased samples where nine models that use different prompts in the text embedding model or trained on the bootstrapped train set uniformly agree that the sample can be predicted correctly based solely on objects. Tab. 4 summarizes the percentage of object-biased samples identified in the datasets, the level of model agreement (measured by Fleiss' kappa coefficient, which is notably high), and the number of samples retained from the original datasets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 426, + 479, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 479, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 479, + 440 + ], + "type": "text", + "content": "5. Benchmarking Video Models" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 447, + 554, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 554, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 554, + 471 + ], + "type": "text", + "content": "In this section, we benchmark state-of-the-art video encoders on considered datasets and their UTD-splits." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 479, + 365, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 365, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 365, + 492 + ], + "type": "text", + "content": "5.1. Setup" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 498, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 617 + ], + "type": "text", + "content": "Video Models. We consider state-of-the-art video encoders with varying model sizes and pre-training configurations with open-sourced weights. For action classification, we benchmark 24 different variants of VideoMAE [38], VideoMAE2 [42], All-in-one [40], Unmasked Teacher (UMT) [21], VideoMamba [22], and InternVid [44] families. For text-to-video retrieval, we consider 17 videolanguage models of UMT [21], VideoMamba [22], and InternVid [44] families. In total, we analyze 30 models and list all considered variants in the supplement." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Implementation Details. To evaluate action classification performance in a standard aligned setup, we follow the VideoGLUE [49] benchmark of using a frozen video backbone and lightweight pooling head as a test of the strength of video model representation. To adapt each video backbone for a specific dataset, we create a model that 1) uses frozen video backbone to extract spatial-temporal features, and then 2) uses a single attention pooling layer [49] to ag-" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29056" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 79, + 355, + 292 + ], + "blocks": [ + { + "bbox": [ + 56, + 79, + 355, + 292 + ], + "lines": [ + { + "bbox": [ + 56, + 79, + 355, + 292 + ], + "spans": [ + { + "bbox": [ + 56, + 79, + 355, + 292 + ], + "type": "table", + "html": "
UCFUCF-UTD-splitUCF-UTD-s. balancedSSv2SSv2-UTD-splitSSv2-UTD-s. balancedK400K400-UTD-splitK400-UTD-s. balanced
videomae-B-K40089.281.3–7.982.8–6.454.252.1–2.153.0–1.264.345.7–18.649.5–14.8
videomae-B-UH88.479.2–9.280.4–8.056.254.0–2.254.9–1.364.345.4–18.949.6–14.7
videomae-L-UH95.790.9–4.891.0–4.767.365.4–1.966.2–1.175.859.3–16.563.3–12.5
videomae-H-UH95.290.9–4.391.0–4.267.465.5–1.966.3–1.176.559.5–17.063.7–12.8
videomaev2-B-K710-fnK71099.098.1–0.997.6–1.457.154.7–2.455.9–1.285.071.9–13.175.5–9.5
allinone-B-WV2M+CC84.563.1–21.472.9–11.626.222.5–3.724.7–1.566.942.6–24.350.8–16.1
allinone-B-WV2M+HT81.457.7–23.768.3–13.122.719.3–3.421.3–1.461.436.9–24.545.2–16.2
allinone-B-WV2M+HT+CC+YTT+81.358.5–22.866.2–15.122.018.8–3.220.7–1.361.236.4–24.844.6–16.6
umt-B-K71091.376.8–14.582.4–8.950.647.9–2.749.1–1.577.758.3–19.463.7–14.0
umt-B-K710-fnK71099.096.8–2.297.1–1.949.446.6–2.848.1–1.385.672.4–13.276.2–9.4
umt-L-K71095.787.2–8.589.2–6.559.457.0–2.458.1–1.384.068.4–15.673.0–11.0
umt-L-K710-fnK71098.996.7–2.297.1–1.857.955.5–2.456.7–1.289.078.7–10.381.6–7.4
videomamba-vm-K40090.974.6–16.381.0–9.935.932.6–3.334.4–1.575.354.4–20.960.5–14.8
videomamba-vm-5M93.782.0–11.786.7–7.048.545.6–2.946.9–1.677.558.1–19.464.0–13.5
videomamba-vm-17M93.280.5–12.784.3–8.947.444.7–2.745.9–1.577.758.1–19.663.6–14.1
videomamba-vm-25M94.383.0–11.386.5–7.848.745.9–2.847.1–1.678.459.1–19.364.6–13.8
internid-B-10M-FLT94.081.1–12.984.7–9.348.145.2–2.946.7–1.478.659.3–19.364.9–13.7
internid-B-200M94.582.9–11.685.2–9.354.451.8–2.652.8–1.680.061.7–18.366.8–13.2
internid-L-10M93.280.2–13.085.3–7.952.650.0–2.651.1–1.577.457.4–20.063.8–13.6
internid-L-WebVid10M94.182.5–11.686.3–7.854.952.4–2.553.4–1.578.959.7–19.265.7–13.2
internid-L-10M-DIV94.883.2–11.687.6–7.251.748.8–2.950.1–1.680.362.1–18.267.4–12.9
internid-L-10M-FLT95.586.3–9.288.4–7.153.650.9–2.752.1–1.581.964.7–17.269.6–12.3
internid-L-50M95.285.5–9.788.4–6.860.658.2–2.459.3–1.381.363.7–17.669.1–12.2
internid-L-200M96.890.5–6.391.9–4.963.161.0–2.161.9–1.284.669.4–15.273.9–10.7
", + "image_path": "ba7e118cb7a89b496d812f4989f36b4fb8354d7b4658f29cc2a4b71fd6c46cff.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 359, + 79, + 558, + 250 + ], + "blocks": [ + { + "bbox": [ + 55, + 296, + 357, + 330 + ], + "lines": [ + { + "bbox": [ + 55, + 296, + 357, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 357, + 330 + ], + "type": "text", + "content": "Table 5. Benchmarking video models in action classification. We report accuracy on full test/val sets and our debiased UTD-splits. The accuracy differences with respect to the full test/val sets are color-coded." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 359, + 79, + 558, + 250 + ], + "lines": [ + { + "bbox": [ + 359, + 79, + 558, + 250 + ], + "spans": [ + { + "bbox": [ + 359, + 79, + 558, + 250 + ], + "type": "table", + "html": "
MSRMSR-UTD-splitDDMDDM-UTD-splitANetANet-UTD-split
umt-b-5M30.020.5–9.530.220.6–9.628.622.3–6.3
umt-b-17M35.626.3–9.337.727.6–10.134.227.3–6.9
umt-b-25M35.324.8–10.534.224.6–9.625.119.7–5.4
umt-l-5M34.824.9–9.933.521.8–11.734.828.7–6.1
umt-l-17M43.631.1–12.546.335.6–10.745.938.7–7.2
umt-l-25M42.330.6–11.743.633.5–10.136.730.4–6.3
videomamba-vm-5M33.323.0–10.337.127.1–10.037.130.1–7.0
videomamba-vm-17M34.925.5–9.440.628.9–11.740.433.0–7.4
videomamba-vm-25M34.925.5–9.441.430.5–10.941.133.8–7.3
internid-B-10M-FLT37.925.4–12.528.617.2–11.424.418.8–5.6
internid-B-200M38.124.7–13.430.219.4–10.826.220.1–6.1
internid-L-10M26.718.7–8.022.615.6–7.021.516.3–5.2
internid-L-WV10M26.517.6–8.922.213.8–8.423.117.5–5.6
internid-L-10M-DIV37.324.2–13.126.915.9–11.023.217.7–5.5
internid-L-10M-FLT38.726.3–12.429.218.8–10.424.519.0–5.5
internid-L-50M32.422.0–10.426.518.3–8.224.418.0–6.4
internid-L-200M38.224.8–13.430.320.2–10.128.722.1–6.6
", + "image_path": "2cd727989c63efb7cea568496a59c55df589e5a9e1feb9762638400b947c9264.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 356, + 253, + 558, + 308 + ], + "lines": [ + { + "bbox": [ + 356, + 253, + 558, + 308 + ], + "spans": [ + { + "bbox": [ + 356, + 253, + 558, + 308 + ], + "type": "text", + "content": "Table 6. Benchmarking video-language models in text-to-video retrieval. We report accuracy on full test/val sets and our debiased UTD-splits. The accuracy differences with respect to the full test/val sets are color-coded." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 338, + 296, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 458 + ], + "type": "text", + "content": "ggregate spatial-temporal features into a single vector representation and 3) a single linear projection layer to classify on the required number of classes. We train the model on a full train split of the corresponding dataset. For video model evaluation in text-to-video retrieval, we reported their zero-shot performance on our debiased splits. We follow the corresponding model recipes in using re-ranking and how many examples are reranked. We report the accuracy for classification and the recall@1 for text-to-video retrieval. Additional details are provided in the supplement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 467, + 113, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 113, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 113, + 479 + ], + "type": "text", + "content": "5.2. Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 485, + 296, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 296, + 676 + ], + "type": "text", + "content": "We evaluate the performance of video models for action classification on the UCF, SSv2, and K400 datasets and their UTD-splits in Tab. 5 and for text-to-video retrieval on the MSRVTT, DiDeMo, and ActivityNet datasets and their UTD-splits in Tab. 6. Performance for other datasets is reported in the supplement. Across all models, performance drops notably on debiased splits (except for SSv2), suggesting that current models strongly rely on object representation bias. Evaluation on UTD debiased splits shows a higher range of scores of models, revealing that some models (e.g. all-in-one models) were more heavily relying on object bias rather than utilizing action/motion cues for video understanding. Further, we can observe that the best performance flips for some models for the debiased case compared to the full test set, e.g., for videomae-B-K400 compared to videomamba-vm-K400 on the UCF dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Additionally, by comparing the performance decline between the original datasets and the UTD-splits, we can assess models' sensitivity to object bias. Overall, we see in the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 338, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 540 + ], + "type": "text", + "content": "case of action classification, Tab. 5, that VideoMAE models are most robust to object bias, whereas Allinone models are most impacted by object bias. We also note that performance drops correlate between the UTD-split and UTD-split-balanced versions. We can further observe that larger backbones/architectures, and to a certain extent, more training data, lead to a reduced bias compared to smaller models, e.g. InternVid and VideoMAE. This pattern does not hold for text-to-video retrieval, shown in Tab. 6. Here, larger backbones and backbones trained with more data turn out to have higher object bias than smaller models or models trained with less data (see InternVid). We attribute this behavior, to the fact that retrieval models leverage information from both the vision as well as the text backbone. Thus, if larger text backbones also learn more object properties or a larger vocabulary of objects, this might further encourage object-centering learning in the visual stream as well." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "In this paper, we presented a new way to examine representation biases in video classification and retrieval datasets by leveraging recent, powerful VLMs and LLMs. Our investigation confirms that most video datasets are heavily focused on object recognition. Based on our bias analysis, we propose a method to reduce object bias in these benchmarks, enabling more robust video understanding evaluation. With that, we are able to benchmark state-of-the-art video models and show an analysis of the object bias in these models. We hope that our findings, together with the presented benchmark, offer valuable insights for the future development of more robust video understanding benchmarks and models." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29057" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "7. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 162 + ], + "type": "text", + "content": "Nina Shvetsova is supported in part by German Federal Ministry of Education and Research (BMBF) project STCL - 01IS22067. Nina Shvetsova also acknowledges travel support from ELISE (GA no 951847). Prof. Hilde Kuehne is supported in part by the ERC Starting Grant GraViLa 101117556." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 193, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 193, + 296, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 193, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 61, + 193, + 296, + 236 + ], + "type": "text", + "content": "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, 2017. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 237, + 296, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 237, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 61, + 237, + 296, + 270 + ], + "type": "text", + "content": "[2] Piyush Bagad, Makarand Tapaswi, and Cees GM Snoek. Test of time: Instilling video-language models with a sense of time. In CVPR, pages 2503-2516, 2023. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 270, + 295, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 270, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 61, + 270, + 295, + 314 + ], + "type": "text", + "content": "[3] Yizhak Ben-Shabat, Xin Yu, Fatemeh Saleh, Dylan Campbell, Cristian Rodriguez-Opazo, Hongdong Li, and Stephen Gould. The ika asm dataset: Understanding people assembling furniture through actions, objects and pose. 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 315, + 295, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 315, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 315, + 295, + 369 + ], + "type": "text", + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 33:1877-1901, 2020. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 370, + 295, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 370, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 62, + 370, + 295, + 414 + ], + "type": "text", + "content": "[5] Shyamal Buch, Cristóbal Eyzaguirre, Adrien Gaidon, Jiajun Wu, Li Fei-Fei, and Juan Carlos Niebles. Revisiting the\" video\" in video-language understanding. In CVPR, pages 2917-2927, 2022. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 415, + 295, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 458 + ], + "type": "text", + "content": "[6] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 459, + 295, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 295, + 513 + ], + "type": "text", + "content": "[7] Mu Cai, Reuben Tan, Jianrui Zhang, Bocheng Zou, Kai Zhang, Feng Yao, Fangrui Zhu, Jing Gu, Yiwu Zhong, Yuzhang Shang, et al. Temporalbench: Benchmarking fine-grained temporal understanding for multimodal video models. arXiv preprint arXiv:2410.10818, 2024. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 514, + 295, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 295, + 546 + ], + "type": "text", + "content": "[8] João Carreira, Eric Noland, Andras Banki-Horvath, Chloe Hillier, and Andrew Zisserman. A short note about kinetics-600. arxiv, 1808.01340, 2018. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 548, + 295, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 548, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 62, + 548, + 295, + 581 + ], + "type": "text", + "content": "[9] João Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. arxiv, 1907.06987, 2019. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 582, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 582, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 582, + 295, + 624 + ], + "type": "text", + "content": "[10] Sihan Chen, Handong Li, Qunbo Wang, Zijia Zhao, Mingzhen Sun, Xinxin Zhu, and Jing Liu. Vast: A vision-audio-subtitle-text omni-modality foundation model and dataset. NeurIPS, 36, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 625, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 668 + ], + "type": "text", + "content": "[11] Jihoon Chung, Yu Wu, and Olga Russakovsky. Enabling detailed action recognition evaluation through video dataset augmentation. In NeurIPS, pages 39020-39033. Curran Associates, Inc., 2022. 1, 2, 3, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[12] Pradipto Das, Chenliang Xu, Richard F. Doell, and Jason J. Corso. A thousand frames in just a few words: Linguual description of videos through latent topics and sparse object stitching. In CVPR, 2013. 2, 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[13] Simone Fabbrizzi, Symeon Papadopoulos, Eirini Ntoutsi, and Ioannis Kompatsiaris. A survey on bias in visual datasets. Computer Vision and Image Understanding, 223, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 119, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 194 + ], + "type": "text", + "content": "[14] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, Florian Hoppe, Christian Thurau, Ingo Bax, and Roland Memisevic. The \"something something\" video database for learning and evaluating visual common sense. In ICCV, 2017. 1, 2, 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 229 + ], + "type": "text", + "content": "[15] Yun He, Soma Shirakabe, Yutaka Satoh, and Hirokatsu Kataoka. Human action recognition without human. 2016. 1, 2, 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 232, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 285 + ], + "type": "text", + "content": "[16] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 289, + 553, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 289, + 553, + 344 + ], + "spans": [ + { + "bbox": [ + 316, + 289, + 553, + 344 + ], + "type": "text", + "content": "[17] Will Kay, João Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arxiv, 1705.06950, 2017. 2, 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 346, + 553, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 346, + 553, + 378 + ], + "spans": [ + { + "bbox": [ + 316, + 346, + 553, + 378 + ], + "type": "text", + "content": "[18] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In ICCV, 2011. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 380, + 553, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 380, + 553, + 412 + ], + "spans": [ + { + "bbox": [ + 316, + 380, + 553, + 412 + ], + "type": "text", + "content": "[19] Jie Lei, Tamara L Berg, and Mohit Bansal. Revealing single frame bias for video-and-language learning. In ACL, 2023. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 415, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 458 + ], + "type": "text", + "content": "[20] Kunchang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 460, + 553, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 504 + ], + "type": "text", + "content": "[21] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. In ICCV, pages 19948-19960, 2023. 2, 3, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 506, + 553, + 549 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 506, + 553, + 549 + ], + "spans": [ + { + "bbox": [ + 317, + 506, + 553, + 549 + ], + "type": "text", + "content": "[22] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. arXiv preprint arXiv:2403.06977, 2024. 2, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 552, + 553, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 552, + 553, + 585 + ], + "spans": [ + { + "bbox": [ + 317, + 552, + 553, + 585 + ], + "type": "text", + "content": "[23] Yingwei Li, Yi Li, and Nuno Vasconcelos. Resound: Towards action recognition without representation bias. In ECCV, 2018. 1, 3, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 587, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 587, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 587, + 553, + 609 + ], + "type": "text", + "content": "[24] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 611, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 611, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 611, + 553, + 632 + ], + "type": "text", + "content": "[25] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 635, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 667 + ], + "type": "text", + "content": "[26] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024.5,6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[27] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In ACL, 2024. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29058" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[28] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. NeurIPS, 36:46212-46244, 2023. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 138 + ], + "type": "text", + "content": "[29] Marcin Marszalek, Ivan Laptev, and Cordelia Schmid. Actions in context. In CVPR, 2009. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "text", + "content": "[30] Rui Meng, Ye Liu, Shafiq Rayhan Joty, Caiming Xiong, Yingbo Zhou, and Semih Yavuz. Sfrembedding-mistral: enhance text retrieval with transfer learning. *Salesforce AI Research Blog*, 3, 2024. 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 239 + ], + "type": "text", + "content": "[31] Mathew Monfort, Alex Andonian, Bolei Zhou, Kandan Ramakrishnan, Sarah Adel Bargal, Tom Yan, Lisa Brown, Quanfu Fan, Dan Gutfreund, Carl Vondrick, et al. Moments in time dataset: one million videos for event understanding. TPAMI, 42(2):502-508, 2019. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 285 + ], + "type": "text", + "content": "[32] Mathew Monfort, SouYoung Jin, Alexander Liu, David Harwath, Rogerio Feris, James Glass, and Aude Oliva. Spoken moments: Learning joint audio-visual representations from video descriptions. In CVPR, 2021. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 287, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 294, + 319 + ], + "type": "text", + "content": "[33] Anna Rohrbach, Marcus Rohrbach, Niket Tandon, and Bernt Schiele. A dataset for movie description. In CVPR, 2015. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 321, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 364 + ], + "type": "text", + "content": "[34] Nina Shvetsova, Anna Kukleva, Xudong Hong, Christian Rupprecht, Bernt Schiele, and Hilde Kuehne. Howtocap-tion: Prompting llms to transform video annotations at scale. arXiv preprint arXiv:2310.04900, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 366, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 294, + 398 + ], + "type": "text", + "content": "[35] Nina Shvetsova, Anna Kukleva, Bernt Schiele, and Hilde Kuehne. In-style: Bridging text and uncurated videos with style transfer for text-video retrieval. In ICCV, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 400, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 400, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 400, + 294, + 431 + ], + "type": "text", + "content": "[36] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arxiv, 1705.06950, 2017. 1, 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "type": "text", + "content": "[37] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In ICCV, pages 7464-7473, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 479, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 521 + ], + "type": "text", + "content": "[38] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. NeurIPS, 35:10078-10093, 2022. 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 524, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 524, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 524, + 294, + 567 + ], + "type": "text", + "content": "[39] Junke Wang, Dongdong Chen, Chong Luo, Xiyang Dai, Lu Yuan, Zuxuan Wu, and Yu-Gang Jiang. Chatvideo: A tracklet-centric multimodal and versatile video understanding system. arXiv preprint arXiv:2304.14407, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 569, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 569, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 569, + 294, + 623 + ], + "type": "text", + "content": "[40] Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Kevin Qinghong Lin, Satoshi Tsutsui, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, et al. All in one: Exploring unified video-language pre-training. In CVPR, pages 6598-6608, 2023. 2, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 625, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 667 + ], + "type": "text", + "content": "[41] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In ECCV, 2016. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[42] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In CVPR, pages 14549-14560, 2023. 2, 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 608 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[43] Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. Improving text embeddings with large language models. arXiv preprint arXiv:2401.00368, 2023. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "text", + "content": "[44] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Intervid: A large-scale video-text dataset for multimodal understanding and generation. In ICLR, 2023. 2, 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "text", + "content": "[45] Zhenhailong Wang, Manling Li, Ruochen Xu, Luowei Zhou, Jie Lei, Xudong Lin, Shuohang Wang, Ziyi Yang, Chenguang Zhu, Derek Hoiem, et al. Language models with image descriptors are strong few-shot video-language learners. NeurIPS, 35:8483-8497, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 230, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 293 + ], + "type": "text", + "content": "[46] Zhanyu Wang, Longyue Wang, Zhen Zhao, Minghao Wu, Chenyang Lyu, Huayang Li, Deng Cai, Luping Zhou, Shuming Shi, and Zhaopeng Tu. Gpt4video: A unified multimodal large language model for Instruction-followed understanding and safety-aware generation. arXiv preprint arXiv:2311.16511, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 297, + 553, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 328 + ], + "type": "text", + "content": "[47] Philippe Weinzaepfel and Grégory Rogez. Mimetics: Towards understanding human actions out of context. *IJCV*, 129, 2021. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 331, + 553, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 362 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 362 + ], + "type": "text", + "content": "[48] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In CVPR, 2016. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 365, + 553, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 418 + ], + "type": "text", + "content": "[49] Liangzhe Yuan, Nitesh Bharadwaj Gundavarapu, Long Zhao, Hao Zhou, Yin Cui, Lu Jiang, Xuan Yang, Menglin Jia, Tobias Weyand, Luke Friedman, et al. Videoglue: Video general understanding evaluation of foundation models. arXiv preprint arXiv:2307.03166, 2023. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 420, + 553, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 420, + 553, + 474 + ], + "spans": [ + { + "bbox": [ + 316, + 420, + 553, + 474 + ], + "type": "text", + "content": "[50] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, et al. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 476, + 553, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 476, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 316, + 476, + 553, + 518 + ], + "type": "text", + "content": "[51] Yuanhao Zhai, Ziyi Liu, Zhenyu Wu, Yi Wu, Chunluan Zhou, David Doermann, Junsong Yuan, and Gang Hua. Soar: Scene-debiasing open-set action recognition. In ICCV, pages 10244–10254, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 521, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 521, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 521, + 553, + 553 + ], + "type": "text", + "content": "[52] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 555, + 553, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 608 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 608 + ], + "type": "text", + "content": "[53] Long Zhao, Nitesh B Gundavarapu, Liangzhe Yuan, Hao Zhou, Shen Yan, Jennifer J Sun, Luke Friedman, Rui Qian, Tobias Weyand, Yue Zhao, et al. Videoprism: A foundational visual encoder for video understanding. arXiv preprint arXiv:2402.13217, 2024. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29059" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Unboxed_ Geometrically and Temporally Consistent Video Outpainting/3a973275-71b1-47b0-9c32-b162722059fe_content_list.json b/2025/Unboxed_ Geometrically and Temporally Consistent Video Outpainting/3a973275-71b1-47b0-9c32-b162722059fe_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3c40d90442c3f750bd08083ac4ce713474bd5c10 --- /dev/null +++ b/2025/Unboxed_ Geometrically and Temporally Consistent Video Outpainting/3a973275-71b1-47b0-9c32-b162722059fe_content_list.json @@ -0,0 +1,1591 @@ +[ + { + "type": "text", + "text": "Unboxed: Geometrically and Temporally Consistent Video Outpainting", + "text_level": 1, + "bbox": [ + 125, + 130, + 844, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhongrui $\\mathrm{Yu}^1$", + "bbox": [ + 135, + 179, + 246, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Martina Megaro-Boldini²", + "bbox": [ + 263, + 180, + 467, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ ETH Zürich", + "bbox": [ + 281, + 204, + 388, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhonyu@ethz.ch", + "bbox": [ + 272, + 229, + 398, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Robert W. Sumner1,2", + "bbox": [ + 483, + 180, + 648, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abdelaziz Djelouah2", + "bbox": [ + 665, + 180, + 831, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2DisneyResearch|Studios", + "bbox": [ + 485, + 204, + 687, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "adelaziz.djelouah@disney.com", + "bbox": [ + 439, + 229, + 697, + 243 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e95dfc66ca20e059d3146175c4d41d9edd181e90cfed1507dd72e46babf01c15.jpg", + "image_caption": [ + "Figure 1. Video outpainting results on two different resolutions. In both case, the area outside the red box is outpainted. Our method achieves results of better quality and more temporally stable, compared to recent works such as Dehan [13], M3DDM [14] and MOTIA [48]." + ], + "image_footnote": [], + "bbox": [ + 50, + 260, + 379, + 476 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/294ffb16ba03e9e73345713154b66df8e1b93d6c74c910b8f745bae7733d51e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 393, + 260, + 924, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 529, + 313, + 545 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Extending the field of view of video content beyond its original version has many applications: immersive viewing experience with VR devices, reformatting 4:3 legacy content to today's viewing conditions with wide screens, or simply extending vertically captured phone videos. Many existing works focus on synthesizing the video using generative models only. Despite promising results, this strategy seems at the moment limited in terms of quality. In this work, we address this problem using two key ideas: 3D supported outpainting for the static regions of the images, and leveraging pre-trained video diffusion model to ensure realistic and temporally coherent results, particularly for the dynamic parts. In the first stage, we iterate between image outpainting and updating the 3D scene representation - we use 3D Gaussian Splatting. Then we consider dynamic objects independently per frame and inpaint missing pixels. Finally, we propose a denoising scheme that allows to maintain known reliable regions and update the dynamic parts to obtain temporally realistic results. We achieve state-of-the-art video outpainting. This is validated quantitatively and through a user study. We are also able to extend the field of view largely beyond the limits reached by existing methods.", + "bbox": [ + 75, + 569, + 472, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 555, + 630, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Expanding video frames beyond their original field of view (FoV), known as Video Outpainting, is a technique with broad applications in modern media. One example is 4:3 legacy content re-targeting where video outpainting can be used to fill modern widescreen devices seamlessly while avoiding any unnatural stretching. A more forward-looking application is movie viewing with virtual reality (VR) headsets. Video outpainting can enhance the immersion by embedding the viewer in the scene, eliminating distractions from unrelated backgrounds. However, these applications require extremely high-quality results, which is particularly challenging since the synthesized images must integrate harmoniously with the original content and maintain strong temporal coherence across frames.", + "bbox": [ + 496, + 582, + 890, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although a wide variety of methods have been proposed, using a patch-based approach [1, 2, 12, 46], optical flow [13] or Generative Adversarial Networks (GAN) [24, 53], most recent methods have clearly identified Diffusion Models [14, 48] as the most promising direction. M3DDM [14] trains a 3D diffusion model on large-scale data for video outpainting with a mask modeling approach.", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7309", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In contrast, MOTIA [48] proposes to fine-tune a pre-trained video diffusion model [19] on each test video to capture data-specific motion and content patterns. However, both of these methods have limitations. First, the temporal consistency of the outpainted regions is still unsatisfactory. We can observe duplications, distortions or objects appearing and disappearing. Second, relying entirely on video diffusion models has a strong impact in terms of memory and limits the methods to low resolution and small field of view extension. Some of these issues can be observed in Figure 1. Both M3DDM [14] and MOTIA [48] have strong visual artifacts and only the latter could be used to outpaint the video to a larger field of view.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these issues we propose two core ideas: 3D-supported outpainting for static regions and conditional video synthesis with a pre-trained video diffusion model to ensure realistic, temporally coherent results for dynamic areas. Following these key ideas, we decompose the video into static and dynamic regions and break down the video outpainting task into three stages. First, for static regions, we rely on 3D Gaussians [23] as a supporting 3D representation, and we alternate between outpainting key frames and updating the 3D representation. Second, for dynamic objects, we implement an object-wise inpainting. Here, among other things, we leverage object-tracking masks in each frame to define inpainting regions and use a fine-tuned diffusion model to inpaint. Lastly, using the resulting frames as a starting point, we synthesize temporally consistent results for all frames using a pre-trained video diffusion model [7].", + "bbox": [ + 75, + 297, + 472, + 555 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Figure 1 demonstrates the effectiveness of our method. We achieve better results than prior methods both in terms of video quality and temporal consistency. This is confirmed by both quantitative and qualitative experiments. Furthermore, by using GPU-efficient Gaussian Splatting for static region outpainting and avoiding fine-tuning of the video diffusion model, we can go from an input $480 \\times 480$ video ( $\\mathrm{FoV} = 31^{\\circ}$ ) to an output resolution of $2560 \\times 720$ ( $\\mathrm{FoV} = 120^{\\circ}$ ) with a GPU memory usage of less than 16GB for the entire pipeline.", + "bbox": [ + 75, + 566, + 472, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 96, + 729, + 398, + 743 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- a novel video outpainting method that combines a 3D representation with a pre-trained video diffusion model;", + "- an optimization strategy with an additional loss to enable seamless integration of the outpainting with the original static content;", + "- demonstrating that a pre-trained video diffusion model can function as a temporal filter, making discontinuous frames temporally coherent and simplifying the inpainting task for the dynamic parts.", + "- State-of-the-art results for video outpainting" + ], + "bbox": [ + 76, + 756, + 467, + 906 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Immersive Scene Generation. To create immersive experiences for users in VR applications, several methods are proposed to generate realistic 3D or 4D scenes that allow for arbitrary viewpoint exploration. LucidDreamer [10] and the following works [16, 35, 43, 54, 55] lift a single image to 3D space, project it from different viewpoints, and inpaint them using diffusion models to create a 3D scene. However, directly applying their method to consecutive video frames would cause severe geometric distortion between frames due to the simple depth alignment and optimization techniques. Dreamscene360 [63] and 4D4KGen [30] create panoramic 3D and 4D scenes separately from a text-generated panoramic image. VividDream [27] generates multi-view videos from a single image and reconstructs a 4D scene with 4D Gaussians. These methods are primarily designed for static or simply animated images and cannot handle videos with complex camera and object motion. Methods such as [9, 28, 29, 49, 50] reconstruct 4D scenes from casual videos, but they are unable to complete regions beyond what is captured in the video.", + "bbox": [ + 496, + 114, + 890, + 416 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image completion. Image completion [6, 39], including image inpainting and outpainting, is a task that restores missing part of an image, and it serves as a basis for video outpainting. Traditional methods [6, 18, 21] use low-level features from the incomplete image to propagate information in known regions to fill missing areas. Early deep learning based methods [32, 42, 44] introduced end-to-end frameworks that learned mappings from corrupted to complete images. Recent methods have increasingly leveraged generative models, including GAN [17, 56, 57, 62], and Diffusion Models [4, 11, 22, 34, 40, 41]. Reference-based image completion methods utilize reference images as generation conditions [52] or fine-tune models on a set of reference images [45] to produce completions that are more faithful to the references. However, even fine-tuned on all input video frames, frame-by-frame completion introduces substantial temporal inconsistency in video outpainting.", + "bbox": [ + 496, + 417, + 890, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Outpainting. In the field of video completion, video outpainting is challenging but less explored. Early patch-based methods [1, 2] select the nearest patch from the spatio-temporal domain of the video to expand each frame. Lee et al. [26] employ 3D information to warp neighboring frames to the current frame. Dehan et al. [13] utilize estimated optical flow to warp the background from adjacent frames and use an image completion network to fill any remaining uncompleted areas. While Zhang et al. [61] target video in-painting motion module and structure guidance. M3DDM [14] extends the image diffusion model framework to 3D and trains it on extensive video data for the outpainting task. A global frame is introduced as a conditioning input to improve global information perception. MOTIA [48] proposes a method that fine-tunes a pre-trained", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7310", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b5c993a86ea967cea348375fe72fd26011ea7a2a7d2c501568417d548a47784f.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 91, + 99, + 236, + 148 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/09a54f17834f1d8eec1f0d39bfb2bcacdbc37c3dfe82d38740419aec20612394.jpg", + "image_caption": [ + "Pre-process" + ], + "image_footnote": [], + "bbox": [ + 91, + 190, + 142, + 229 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3405ae143188f7bed6ed6f3a352b9f5ef03dc44aad903a2064c2aa01b39507c0.jpg", + "image_caption": [ + "Dynamic Object Masks" + ], + "image_footnote": [], + "bbox": [ + 143, + 190, + 194, + 229 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/41fbf127238cb8c67fb260b2b1d9232e72dcb00a37224087e6f655a114ffdacd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 194, + 190, + 241, + 229 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/72cf0259f7d1001528b546c82cbcfb4b45dd3d8c71acc92cd9114d67e825d918.jpg", + "image_caption": [ + "Static Scene Reconstruction" + ], + "image_footnote": [], + "bbox": [ + 86, + 253, + 236, + 335 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ab7e21691daaa99ca318ffecf57409926e41322488202fc41bf425c28a2ed81a.jpg", + "image_caption": [ + "(1) Static Region" + ], + "image_footnote": [], + "bbox": [ + 256, + 108, + 457, + 335 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6a1f9e6b97856706050a05db1c0930f537c9e6ec14c573710aed566bb8e78b87.jpg", + "image_caption": [ + "(2) Dynamic Objects", + "Figure 2. Overview of the proposed method. Starting from the input video sequence, dynamic objects are separated from the static parts of the image, and a static scene reconstruction is made using Gaussian Splatting (GS). In the first stage of our method, the static part is outpainted by iterating between image outpainting, depth prediction and updating the GS model. After this stage, we identify parts of the dynamic object that may require inpainting. Here, an image diffusion model is finetuned on the input frames and used for inpainting each frame independently, which may result in temporal inconsistencies as seen of the car windows. In the last step, a pre-trained video diffusion model is used to fix temporal inconsistencies. Details about each step are provided in the text." + ], + "image_footnote": [], + "bbox": [ + 467, + 108, + 691, + 344 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6e319d107356edd0ba8fc84e351a4f791deec53c22212860a34e9453f404ba15.jpg", + "image_caption": [ + "(3) Final Video Synthesis", + "Temporally Consistent Results" + ], + "image_footnote": [], + "bbox": [ + 705, + 116, + 888, + 344 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "video diffusion model on test samples to learn the intrinsic patterns of a video. In general, diffusion-based approaches rely on priors learned by temporal modules to maintain continuity in the outpainted regions, which is not sufficient. We demonstrate that a pipeline integrating constraints from a 3D representation achieves results of much higher quality.", + "bbox": [ + 75, + 448, + 468, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 555, + 166, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a sequence of video frames $\\{I_i\\in \\mathbb{R}^{h\\times w\\times c}\\}_{i = 1}^N$ , our goal is to expand each frame beyond its original field of view to $\\{\\tilde{I}_i\\in \\mathbb{R}^{H\\times W\\times c}\\}_{i = 1}^N$ , while maintaining both intra-frame and cross-frame consistency. Although the problem is more challenging than image outpainting, due to the temporal consistency requirements, the presence of multiple frames offers additional observations that can be leveraged. More specifically, it is possible to build a 3D representation of the scene using 3D Gaussians.", + "bbox": [ + 75, + 580, + 468, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Figure 2, our proposed method can be divided into 3 distinct stages. In the first stage, we focus on the static parts of the scene and alternate between image outpainting and updating the 3D scene representation. In the second stage, we focus on the dynamic objects in the scene. Here, we fine-tune an image diffusion model on the available images and use it to inpaint the missing parts of the object in each frame. At this point, all the frames are outpainted, but temporal inconsistencies remain. These frames serve as a starting point for a video synthesis scheme leveraging pre-trained stable video diffusion to synthesize a temporally consistent video sequence. Next, we detail each", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of these main steps.", + "bbox": [ + 500, + 448, + 632, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Pre-processing and Static Scene Outpainting", + "text_level": 1, + "bbox": [ + 500, + 474, + 877, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From the static parts of the video, it is possible to estimate a 3D reconstruction of the scene and use the different views across time to provide information about the regions to outpaint. In the first step, we identify the dynamic parts of the scene. A wide variety of methods can be employed here, but we simply rely on object segmentation and, in particular, object tracking models [33]. Figure 2 shows an example of the binary segmentation masks. These masks are expanded to avoid including incorrectly labeled pixels into the background. We also compute a 2D bounding box around the object to help keep track of the object size in case it partially or completely leaves the field of view. The static parts are all the pixels that are not part of any dynamic object.", + "bbox": [ + 496, + 500, + 890, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Static Scene Reconstruction. We propose to use Gaussian Splatting (GS) [23] to reconstruct the static part of the video. This approach allows regions visible in other frames to be rendered into the current frame. For the input video frames $\\{I_i\\}_{i=1}^N$ , we first estimate the corresponding camera poses $\\{\\mathbf{p}_i\\}_{i=1}^N$ and intrinsic parameters $\\{\\mathbf{K}_i\\}_{i=1}^N$ using an off-the-shelf camera pose estimator [60]. We compute a depth map for each frame using parallax and project the background pixels into 3D space, yielding an initial pointcloud for GS training. The training process is supervised by the image reconstruction loss and the depth loss. The image reconstruction loss is the combination of the L1 loss and", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7311", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/89a119bdb865156ce67a37a4ea8c2d23e43ad9c77436f0065b7782e27badc70e.jpg", + "image_caption": [ + "First key-frame $(k = 1)$", + "$I_{k = 1}^{\\mathrm{static}}$" + ], + "image_footnote": [], + "bbox": [ + 80, + 104, + 199, + 170 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/72dc690b512c7129f381d8742b4619f55f25810f9ead2305a88e0b6b336ede05.jpg", + "image_caption": [ + "$I_{k = 1}^{\\mathrm{static}}$" + ], + "image_footnote": [], + "bbox": [ + 204, + 104, + 325, + 170 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/af585d567554915f32c7b929fd1f75dfa311d581fa21d8d13643f4b6bd019e48.jpg", + "image_caption": [ + "$L_{\\mathrm{recon}}$" + ], + "image_footnote": [], + "bbox": [ + 331, + 104, + 450, + 170 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/37bc47ebb604cf9485529c50ea63149efe391394487854107dcb2f2bbb709de5.jpg", + "image_caption": [ + "Depth after GS finetuning", + "$L_{\\mathrm{recon}} + L_{\\mathrm{depth}}$" + ], + "image_footnote": [], + "bbox": [ + 454, + 104, + 576, + 170 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a6a9c7eb4e65893f26ad32169c9f36d5d7327a6bd1ba64c71bc5834ed31c2756.jpg", + "image_caption": [ + "$L_{\\mathrm{recon}} + L_{\\mathrm{depth}} + L_{\\mathrm{depth}}^{\\mathrm{Sp}}$" + ], + "image_footnote": [], + "bbox": [ + 583, + 104, + 702, + 170 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2d843ea0a1c70843f385b3d8bba6b6d0eb60188295e4810ac20f9ff8c35f72b8.jpg", + "image_caption": [ + "Next key-frame $(k = 2)$", + "$I_{k = 2}^{\\mathrm{static}}$", + "Figure 3. Effect of the loss terms on the GS fine-tuning. Starting from the rasterization of the 3D Gaussians for the first key-frame $k = 1$ , we use image outpainting to obtain $\\hat{I}_k^{\\mathrm{static}}$ . This image is subdivided into super-pixels as illustrated here. During the optimization to update the GS model, we use a combination of losses. We can see that all are needed to avoid discontinuities and obtain a realistic smooth depth map. After this update, the process is repeated with the next key-frame using the updated 3D Gaussians." + ], + "image_footnote": [], + "bbox": [ + 763, + 104, + 883, + 171 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the structural similarity loss (SSIM). The depth loss acts as a regularization and uses Pearson Correlation [64] between the rendered depth and the depth estimated from a monocular depth estimation model [37]. Apart from the dynamic object masking, this is similar to the optimization proposed by Zhu et al. [64].", + "bbox": [ + 75, + 263, + 468, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Single Image Outpainting. Starting from the GS model $GS_{\\mathrm{static}}$ representing the static parts of the scene, we uniformly sample $k \\in \\{1, \\dots, K\\}$ key frames. Our objective is to outpaint them one by one, while we progressively update the GS model. Here we first detail outpainting for a single image. Considering frame $I_k^{\\mathrm{static}}$ , we adjust its field of view to obtain new intrinsic parameters $\\hat{\\mathbf{K}}_k$ , and render $GS_{\\mathrm{static}}$ based on the new parameters to get the image $\\hat{I}_k^{\\mathrm{static}}$ . The mask $\\hat{\\mathbf{m}}_k^{\\mathrm{static}}$ for outpainting is defined by checking whether each pixel is covered by the projection of Gaussian points. Any image outpainting method can be used at this level, but we find that the Stable Diffusion XL (SDXL) image outpainting model [38] performs well for our task. This can be summarized as:", + "bbox": [ + 75, + 371, + 470, + 583 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {k} ^ {\\text {s t a t i c}} = \\phi_ {G S _ {\\text {s t a t i c}}} \\left(\\hat {\\mathbf {K}} _ {k}, \\mathbf {p} _ {k}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 594, + 468, + 612 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {I} _ {k} ^ {\\text {s t a t i c}} = \\phi_ {\\mathrm {S D X L}} \\left(I _ {k} ^ {\\text {s t a t i c}}, \\hat {\\mathbf {m}} _ {k} ^ {\\text {s t a t i c}}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 614, + 468, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Updating the GS model. To ensure the consistency of the outpainting across frames, we update the GS model each time we outpaint a key-frame. Denote by $K'$ the number of key-frames that have been outpainted so far. First, we estimate the depth for the newly outpainted image using a single image depth model [37]. Then we initialize new Gaussian points for the outpainted pixel and fine-tune the GS model with this additional data. We use an image reconstruction loss", + "bbox": [ + 75, + 650, + 468, + 785 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {\\text {r e c o n}} = \\sum_ {k = 1} ^ {K ^ {\\prime}} \\lambda_ {1} L _ {1} \\left(\\hat {I} _ {k} ^ {\\text {s t a t i c}}, \\phi_ {\\mathrm {G S}} \\left(\\hat {K} _ {k}, p _ {k}\\right)\\right) \\tag {3} \\\\ + \\lambda_ {2} L _ {\\mathrm {S S I M}} (\\hat {I} _ {k} ^ {\\text {s t a t i c}}, \\phi_ {\\mathrm {G S}} (\\hat {K} _ {k}, p _ {k})) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 794, + 468, + 858 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "between the rasterization from the model $\\phi_{\\mathrm{GS}}(\\hat{K}_k,p_k)$ and the outpainting result. We also use 2 depth based losses, the", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "first one on the full depth map:", + "bbox": [ + 500, + 263, + 705, + 279 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {d e p t h}} = \\sum_ {k = 1} ^ {K ^ {\\prime}} \\frac {\\operatorname {C o v} \\left(D \\left(\\hat {I} _ {k} ^ {\\text {s t a t i c}}\\right) , D _ {\\mathrm {G S}} \\left(\\hat {K} _ {k} , p _ {k}\\right)\\right)}{\\sqrt {\\operatorname {V a r} \\left(D \\left(\\hat {I} _ {k} ^ {\\text {s t a t i c}}\\right)\\right) \\operatorname {V a r} \\left(D _ {\\mathrm {G S}} \\left(\\hat {K} _ {k} , p _ {k}\\right)\\right)}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 291, + 890, + 339 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $D(\\hat{I}_k^{\\mathrm{static}})$ is the depth predicted by the single image depth model [37], and $D_{\\mathrm{GS}}(\\hat{K}_k,p_k)$ is the depth from GS model that we optimize. For every image, we extract $S$ super-pixels, and use the second depth loss at this level:", + "bbox": [ + 498, + 351, + 892, + 415 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {d e p t h}} ^ {\\mathrm {S p}} = \\sum_ {k = 1} ^ {K ^ {\\prime}} \\sum_ {s = 1} ^ {S} \\frac {\\operatorname {C o v} _ {s} \\left(D \\left(\\hat {I} _ {k} ^ {\\text {s t a t i c}}\\right) , D _ {\\mathrm {G S}} \\left(\\hat {K} _ {k} , p _ {k}\\right)\\right)}{\\sqrt {\\operatorname {V a r} _ {s} \\left(D \\left(\\hat {I} _ {k} ^ {\\text {s t a t i c}}\\right)\\right) \\operatorname {V a r} _ {s} \\left(D _ {\\mathrm {G S}} \\left(\\hat {K} _ {k} , p _ {k}\\right)\\right)}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 428, + 890, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 3 demonstrate this process and shows the importance of the different loss terms for the GS update. Starting from the rasterization of the initial GS model, we can see that the 3D model does not cover the new field of view. The frame is outpainted and superpixels are illustrated in yellow. Using only the reconstruction loss $L_{\\mathrm{recon}}$ leads to strong discontinuities between the original GS model and the new outpainted areas. Adding the depth loss $L_{\\mathrm{depth}}$ helps, but the bottom part of the image still has important discontinuities that would be revealed when projecting to other frames (i.e. other views). Using all the proposed loss terms provides the best results, and the process can now be iterated by adding new views.", + "bbox": [ + 496, + 489, + 892, + 684 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Initial Inpainting for the Dynamic Object(s)", + "text_level": 1, + "bbox": [ + 500, + 695, + 872, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After the static scene outpainting, we can render any frame with $\\phi_{\\mathrm{GS}}(\\hat{K}_i,p_i)$ . As illustrated in Figure 2, we miss part of the object when it is partially outside the filed of view. We need to inpaint this missing part. We use the 2D tracking bounding box to identify the areas that require inpainting for each frame of the video.", + "bbox": [ + 496, + 719, + 890, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For this inpainting problem we use a similar strategy as ReaFill [45]. An image diffusion model is fine-tuned for the inpainting task on the input frames (in our case SDXL). It is then applied independently on each frame. Thanks to the fine-tuning, the resulting inpainting is very similar across frames. However temporal inconsistencies remain as it is", + "bbox": [ + 496, + 810, + 892, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7312", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cd18af78ee96da840a66cb6044ef3e42c6573947b61da9d570a129dd0e21c650.jpg", + "image_caption": [ + "Figure 4. Details of the video latent denoising strategy. Starting from the independently outpainted frames $\\tilde{I}_{i\\in 1,\\dots,N}$ , we obtained the corresponding latents in the forward process $(\\tilde{x}_{i\\in 1,\\dots,N})$ . In the denoising process we blend known latent with denoising latents according to the mask $\\mathbf{m}$ . The mask and the time step $t_{\\mathrm{end}}$ respectively control which region of the frames get updated by the video diffusion model and how large the changes are." + ], + "image_footnote": [], + "bbox": [ + 68, + 88, + 477, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "clearly visible on the inpainting results in Figure 2. Next we describe how a Video Diffusion Model [7, 8] is used to fix all these temporal artifacts.", + "bbox": [ + 75, + 382, + 468, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Guided Video Synthesis", + "text_level": 1, + "bbox": [ + 76, + 435, + 295, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "If we consider the original input sequences of frames $I_{i\\in 1,\\dots,N}$ , after the previous stages we obtain fully out-painted frames $\\tilde{I}_{i\\in 1,\\dots,N}$ . Static regions are consistent thanks to the supporting GS model, while the dynamic objects are independently inpainted for each frame. Starting from this data, our objective is to synthesize a realistic sequence of frames $I_{i\\in 1,\\dots,N}^{*}$ . Not only the temporal inconsistencies on the dynamic objects need to be fixed, but the entire outpainted regions need to appear realistic. Indeed, real-world videos often contain subtle background movements, such as leaves swaying in the wind or water rippling.", + "bbox": [ + 75, + 459, + 468, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We leverage pre-trained video diffusion models for this video synthesis. The idea is that video diffusion models are trained for generating real video sequences, with consistent content and realistic motion. We propose using the previously generate frames in a denoising scheme that modifies the different frame areas based on the expected inconsistency level. On the dynamic objects, we expect more changes (as the inpainting is independent), while less change should be allowed on the rest of the image as it is supported by the 3D consistent GS model.", + "bbox": [ + 75, + 627, + 468, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "More specifically, if we consider a latent video diffusion model, a sequence of input frames $I_{i}$ is encoded into the corresponding sequence of latent values $x_{i}$ . The forward process can be applied to the sequence of frames $\\tilde{I}_{i\\in 1,\\dots,N}$ to obtain known latents", + "bbox": [ + 75, + 777, + 468, + 851 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {x} _ {t - 1} \\sim \\mathcal {N} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\tilde {x} _ {0}, \\sqrt {1 - \\bar {\\alpha} _ {t}} \\mathbf {1}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 859, + 468, + 877 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For simplicity, we dropped the frame index $i$ .", + "bbox": [ + 76, + 885, + 377, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the reverse process to synthesize the final video sequence. At each time step, we denoise previous latents using the latent video diffusion model with parameters $\\theta$ , to obtain temporary latents", + "bbox": [ + 498, + 90, + 890, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1} ^ {\\operatorname {t m p}} \\sim \\mathcal {N} \\left(\\mu_ {\\theta} \\left(x _ {t} ^ {*}, t\\right), \\Sigma_ {\\theta} \\left(x _ {t} ^ {*}, t\\right)\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 162, + 890, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To obtain the final latents for this time step, we blend in the known latents according to", + "bbox": [ + 498, + 191, + 890, + 222 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1} ^ {*} = (1 - \\mathbf {m}) \\odot \\tilde {x} _ {t - 1} + \\mathbf {m} \\odot x _ {t - 1} ^ {\\mathrm {t m p}} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 234, + 890, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the binary mask $\\mathbf{m}$ is defined as", + "bbox": [ + 500, + 263, + 754, + 277 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {m} = \\left\\{ \\begin{array}{l l} \\mathbf {m} ^ {\\text {d y n a m i c}} & \\text {i f} t \\geq t _ {\\text {e n d}} \\\\ \\mathbf {m} ^ {\\text {d y n a m i c}} + \\mathbf {m} ^ {\\text {s t a t i c}} & \\text {i f} t < t _ {\\text {e n d}} \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 286, + 890, + 324 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The mask, $\\mathbf{m}^{\\mathrm{dynamic}}$ , corresponds to the dynamic objects in the scene, and the mask $\\mathbf{m}^{\\mathrm{static}}$ corresponds to the static parts in the outpainted regions. This scheme is similar to ideas already explored in the context of image diffusion models [3, 34], but applied to video diffusion models. If we observe the changes in Figure 2, we can see that the pre-trained video diffusion model can function as a temporal filter, making discontinuous frames temporally coherent. This simplifies the inpainting task since we do not rely on expensive training of the video model and simply use it at inference time.", + "bbox": [ + 496, + 335, + 892, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 515, + 632, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation details", + "text_level": 1, + "bbox": [ + 500, + 540, + 714, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Static Scene. In pre-processing, we estimate the camera poses of the source video [31, 60]. Our GS training process consists of $1 + K + 1$ rounds in total, where $K$ is the number of key frames we sample from the video. In the first training round, we train the GS model only on the static part of the video for 3000 iterations. Then we iterate between outpainting the $K$ key frames and $K$ rounds of GS training to update the outpainted regions into the GS model. To avoid overfitting, we employ a sliding window training strategy: following outpainting of the $i$ -th frame, we train only on frames within the window $[i - 5, i + 5]$ for 500 iterations. This window shifts progressively as outpainting and training continues. After outpainting all $K$ key frames, we conduct a final round of training for refinement on all out-painted and original background frames for 1000 iterations. All training and outpainting can be completed on a single 4090 GPU.", + "bbox": [ + 496, + 564, + 890, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Initial Object Inpainting. Before initial inpainting on the missing regions of dynamic objects, we fine-tune the SDXL model on the masked source video. The approach is similar to the one proposed by Tang et al. [45]. Masks are randomly", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7313", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/08f698742f583338b3b845b1861c522ccbc43d7e250713da970ba9a84f8fdb37.jpg", + "image_caption": [ + "Figure 5. Qualitative Comparison with State-of-the-art methods on the DAVIS dataset [36]. The area outside of the red lines (only shown for the first frame) is outpainted. We show the results for 3 frames from the each video sequence. Compared to MOTIA [48] and M3DDM[14], our results are of better quality and temporally stable. Please pay attention to the changes of the outpainted area between the frames of the same sequence." + ], + "image_footnote": [], + "bbox": [ + 68, + 89, + 877, + 570 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "generated for each input frame, with random distributed positions and mask ratios varying from 0.2 to 0.8. For each object, a prompt \"a photo of [v]\" is applied. During fine-tuning, a Low-Rank Adaption (LoRA) [20] with rank 8 is trained for both the text-encoder and UNet with a batch size of 4 and gradient accumulation steps of 4 for 2000 steps.", + "bbox": [ + 75, + 654, + 472, + 761 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Video Diffusion Model. We use a pretrained image-to-video model [7] to complete object motion. The number of denoising steps is set to 25, with the level of noise added to the reconstructed frames generally set to 0.8 in our experiments. We apply a resampling strategy similar to Repaint [34] five times after each denoising step to improve the quality and stability of the objects. For videos where the object fully appears in the middle or last frame, we em", + "bbox": [ + 75, + 779, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ploy the Time-Reversal strategy from [15], conditioning the generation on additional frames to better retain the object identity.", + "bbox": [ + 498, + 654, + 893, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experiment Setup", + "text_level": 1, + "bbox": [ + 500, + 710, + 676, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Competitors. We compare our method with the following methods. (1) Dehan et al. [13] separate foreground and background components. They use estimated optical flow to warp the background, with missing regions filled in using an image completion network. (2) M3DDM [14] proposes a masked 3D Diffusion Model architecture, which is trained on WebVid [5] and a self-collected 5M e-commerce dataset for video outpaint tasks. (3) MOTIA [48] is the current state-of-the-art method for video outpainting. It is based on the ControlNet [58] inpainting model and the temporal module fromAnimateDiff [19]. It trains a LoRA [20] for", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7314", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "each test sample to learn the intrinsic patterns.", + "bbox": [ + 76, + 90, + 383, + 104 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets. Following M3DDM [14] and MOTIA [48], we evaluate our method on the DAVIS [36] and YouTube-VOS [51] datasets. For the YouTube-VOS [51] dataset, due to its large number of videos and the lack of clarity in prior works on which videos were tested, we selected a high-quality subset and evaluated all methods on it. Consistent with prior works, we compare the horizontal outpainting results for each test video at mask ratios of 0.25 and 0.66.", + "bbox": [ + 75, + 106, + 468, + 226 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metrics. The quality of outpainting is primarily evaluated in terms of visual quality, realism, and temporal consistency. For this purpose, we employ two well-established metrics: Fréchet Video Distance (FVD) [47] and flow warping error $(E_{\\mathrm{warp}})$ [25]. FVD assesses the similarity between the distribution of the outpainted video and the original video, while flow warping error evaluates temporal consistency by measuring the error when outpainted frames are warped to adjacent frames using estimated optical flow. In addition, to align with previous works, we also report Peak Signal to Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) [59] between the outpainted and ground-truth videos, although we consider these metrics limited in capturing the quality of outpainting. We further conducted a user study to gather subjective evaluations of outpainting quality, which we detail in Sec. 4.3.", + "bbox": [ + 75, + 244, + 472, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Evaluation & Results", + "text_level": 1, + "bbox": [ + 76, + 511, + 277, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. Fig. 5 and Fig. 6 illustrate the qualitative comparison of our method against other approaches on the DAVIS and YouTube-VOS datasets. M3DDM [14] frequently suffers from generation failures and produces blurred content, as shown in the first samples of Fig. 5 and Fig. 6. MOTIA [48], while fine-tuning on each test sample to capture input-specific patterns, often generates repetitive objects and background patterns. More importantly, both methods struggle to accurately capture contextual information across frames, resulting in inconsistency when outpainting elements seen in adjacent frames. In contrast, our method overcomes these issues, achieving better cross-frame consistency. Furthermore, our approach's object-specific completion approach prevents objects from duplicating or disappearing beyond the frame.", + "bbox": [ + 75, + 534, + 472, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generalization and multiple objects. Our strategy, differentiating between static and dynamic elements, can naturally extend to multiple dynamic objects by decomposing the frame into object layers based on their depth. Each object is processed separately using our pipeline and then seamlessly blended to produce the final result. Figure 7 shows that our method still performs best on such complex scenarios.", + "bbox": [ + 75, + 779, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4f7813febd35ba3dc97b6dd27df7e655f65b2258e25554056c6402fa69fe10d3.jpg", + "image_caption": [ + "Figure 6. Qualitative Comparison with State-of-the-art methods on Youtube-VOS dataset [51]. The area outside of the red lines (only shown for the first frame) is outpainted. We show the results for 3 frames from the each video sequence. Compared to Dehan [13], MOTIA [48] and M3DDM[14], our results are of better quality and temporally stable. Please pay attention to the changes of the outpainted area between the frames of the same sequence." + ], + "image_footnote": [], + "bbox": [ + 488, + 90, + 888, + 328 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2e853b0c067821d1f93277b2e0bcd86eebf05e0d9e85f313ae569af3879a75eb.jpg", + "image_caption": [ + "Figure 7. Qualitative Comparison with State-of-the-art methods on challenging scenarios (multiple dynamic objects, complex static regions, etc.)." + ], + "image_footnote": [], + "bbox": [ + 486, + 455, + 890, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Results. Tab. 1 presents a quantitative comparison of our method with other approaches. Our method outperforms the current state-of-the-art in both FVD and $E_{\\text{warp}}$ , achieving improvements of $18.8\\%$ and $11.7\\%$ on the DAVIS dataset and $26.9\\%$ and $33.3\\%$ on the YouTube-VOS dataset, respectively. We also achieve the highest PSNR and SSIM scores on both datasets. For LPIPS, while our method is the second behind the number reported in MOTIA [48] on the DAVIS dataset, it surpasses the score obtained from using their public implementation and provided model.", + "bbox": [ + 496, + 698, + 890, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study We conducted ablation studies for the different components of our method. Quantitative evalua", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7315", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/68141a09f0fd4cf9b1baa109b167e8ab7b0fca9f3bce894513b55ef60abf1188.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DAVISYouTube-VOS
PSNR↑SSIM↑LPIPS↓FVD↓E_warp↓PSNR↑SSIM↑LPIPS↓FVD↓E_warp↓
Dehan [13]17.960.6270.233363.10.03018.340.7610.287353.60.043
M3DDM [14]20.260.7080.204300.00.01718.330.7250.424283.10.024
MOTIA (model) [48]18.250.6990.282294.80.02118.540.7900.250156.40.027
MOTIA (paper) [48]20.360.7580.160286.3------
Ours22.620.8060.200232.50.01519.880.8320.236114.20.016
", + "bbox": [ + 83, + 88, + 885, + 218 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e2f132b6f53fd63eadb810ca1adfe2159bf50c7f32fababcedbe92bec2c4ec2f.jpg", + "table_caption": [ + "Table 1. Quantitative Comparison with State-of-the-art methods on DAVIS dataset [36] and YouTube-VOS [51]. Our method achieves the best performance overall. In the case of MOTIA [48], we have 2 rows as we report the evaluation using the model and code available online(model) and the numbers from the original paper (paper). See the text for details." + ], + "table_footnote": [], + "table_body": "
PSNR↑SSIM↑LPIPS↓FVD↓E_warp↓
(a)18.150.6940.248448.80.032
(b)18.050.7160.274323.20.022
(c)20.680.7720.218293.30.020
(d)20.050.7760.223319.80.017
(e)22.620.8060.200232.50.015
", + "bbox": [ + 112, + 292, + 433, + 401 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Quantitative ablation results. (a). Outpainting with fine-tuned Image Model; (b). (a) + Guided Video Synthesis; (c). Our model without guided video synthesis; (d). Our model without the initialization from object inpainting; (e). Our complete model.", + "bbox": [ + 75, + 411, + 470, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tion is shown in Table 2, while the visual results are provided in supplementary material. When only using single image outpainting (SDXL fine-tuned on content [45]), the images achieve good realism but lack temporal continuity. When applying video diffusion guided synthesis to these outpainted frames directly, the temporal consistency within the outpainted areas improves individually. However, coherence between the outpainted regions and the original video content remains weak. This issue is difficult to resolve solely with Video Diffusion Model priors. When using the static region outpainting (assisted with GS) and either the image inpainting or video diffusion alone are applied for object completion, the objects suffer from significant temporal variations in appearance. Using all the components achieves the best results.", + "bbox": [ + 75, + 486, + 468, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User Study. Given the subjective nature of evaluating outpainting quality, we conducted a user study to compare our method with competing approaches. Users were asked to evaluate each method across several dimensions, including realism (such as whether the outpainted results look natural and harmonious), temporal consistency (including background consistency, object motion consistency, etc.), and overall visual quality (including color fidelity, smoothness of boundaries, blurriness etc.). The results, shown in Fig. 8, indicate that our method was clearly preferred by users compared to others, achieving more than $80\\%$ of the votes in all dimensions. Further details of the user study can", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/41c904528a4bc15d586ad920aa1283aa241aca5003efce01f59358301a495c4a.jpg", + "image_caption": [ + "Figure 8. The result of the user study includes 619 votes from 37 participants. Our method is preferred by more than $80\\%$ of the votes in all dimensions." + ], + "image_footnote": [], + "bbox": [ + 506, + 284, + 885, + 388 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6b4dda42f6c045166c29ae37338ef75bf1fbcc304b49f64a015ea09bfa96a043.jpg", + "image_caption": [ + "Figure 9. Challenges with dynamic objects moving in and out of the original frame: Our current solution of tracking and extrapolating the trajectory of the bounding box can be improved." + ], + "image_footnote": [], + "bbox": [ + 500, + 445, + 890, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "be found in the supplementary materials.", + "bbox": [ + 500, + 619, + 769, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussions", + "text_level": 1, + "bbox": [ + 500, + 648, + 620, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Building on Gaussian Splatting and a Video Diffusion Model, we propose a new state-of-the-art method for video outpainting. Despite our strong results, some issues remain, especially around objects moving in and out of the original frame. An example is shown in Fig. 9. Currently tracking and extrapolating the trajectory of the bounding box offers a solution, but limitations in the single image inpainting negatively impact the quality of the final result. Another area of improvement is the outpainting involving multiple occluded dynamic objects. Our current solution of inpainting each object separately and blending them based on the estimated depth is limited. Exploring the use of more advanced 4D dynamic Gaussian representations jointly with video models offers a promising future direction towards higher and higher quality video outpainting.", + "bbox": [ + 496, + 672, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7316", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Amit Aides, Tamar Avraham, and Yoav Y Schechner. Multiscale ultrawide foveated video extrapolation. In 2011 IEEE International Conference on Computational Photography (ICCP), pages 1-8. IEEE, 2011. 1, 2", + "[2] Tamar Avraham and Yoav Y Schechner. Ultrawide foveated video extrapolation. IEEE Journal of Selected Topics in Signal Processing, 5(2):321-334, 2010. 1, 2", + "[3] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18208-18218, 2022. 5", + "[4] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18208-18218, 2022. 2", + "[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 6", + "[6] M Bertalmio. Image inpainting, 2000. 2", + "[7] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 5, 6", + "[8] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 5", + "[9] Wen-Hsuan Chu, Lei Ke, and Katerina Fragkiadaki. Dreamscene4d: Dynamic multi-object scene generation from monocular videos. arXiv preprint arXiv:2405.02280, 2024. 2", + "[10] Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384, 2023. 2", + "[11] Ciprian Corneanu, Raghudeep Gadde, and Aleix M Martínez. Latentpaint: Image inpainting in latent space with diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4334-4343, 2024. 2", + "[12] Antonio Criminisi, Patrick Pérez, and Kentaro Toyama. Region filling and object removal by exemplar-based image inpainting. IEEE Transactions on image processing, 13(9): 1200-1212, 2004. 1", + "[13] Loic Dehan, Wiebe Van Ranst, Patrick Vandewalle, and Toon Goedemé. Complete and temporally consistent video outpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 687-695, 2022. 1, 2, 6, 7, 8", + "[14] Fanda Fan, Chaoxu Guo, Litong Gong, Biao Wang, Tiezheng Ge, Yuning Jiang, Chunjie Luo, and Jianfeng Zhan. Hierar-" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "chical masked 3d diffusion model for video outpainting. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7890-7900, 2023. 1, 2, 6, 7, 8", + "[15] Haiwen Feng, Zheng Ding, Zhihao Xia, Simon Niklaus, Victoria Abrevaya, Michael J Black, and Xuaner Zhang. Explorative inbetweening of time and space. In European Conference on Computer Vision, pages 378-395. Springer, 2025. 6", + "[16] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2", + "[18] Qiang Guo, Shanshan Gao, Xiaofeng Zhang, Yilong Yin, and Caiming Zhang. Patch-based image inpainting via two-stage low rank approximation. IEEE transactions on visualization and computer graphics, 24(6):2023-2036, 2017. 2", + "[19] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2, 6", + "[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6", + "[21] Jia-Bin Huang, Sing Bing Kang, Narendra Ahuja, and Johannes Kopf. Image completion using planar structure guidance. ACM Transactions on graphics (TOG), 33(4):1-10, 2014. 2", + "[22] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. arXiv preprint arXiv:2403.06976, 2024. 2", + "[23] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3", + "[24] Naoki Kimura and Jun Rekimoto. Extvision: augmentation of visual experiences with generation of context images for a peripheral vision using deep neural network. In Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems, pages 1-10, 2018. 1", + "[25] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In European Conference on Computer Vision, 2018. 7", + "[26] Minhyeok Lee, Suhwan Cho, Chajin Shin, Jungho Lee, Sunghun Yang, and Sangyoun Lee. Video diffusion models are strong video inpainter. arXiv preprint arXiv:2408.11402, 2024. 2", + "[27] Yao-Chih Lee, Yi-Ting Chen, Andrew Wang, Ting-Hsuan Liao, Brandon Y Feng, and Jia-Bin Huang. Vividdream: Generating 3d scene with ambient dynamics. arXiv preprint arXiv:2405.20334, 2024. 2" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7317", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Yao-Chih Lee, Zhoutong Zhang, Kevin Blackburn-Matzen, Simon Niklaus, Jianming Zhang, Jia-Bin Huang, and Feng Liu. Fast view synthesis of casual videos with soup-of-planes. In European Conference on Computer Vision, pages 278–296. Springer, 2025. 2", + "[29] Jiahui Lei, Yijia Weng, Adam Harley, Leonidas Guibas, and Kostas Daniilidis. Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds. arXiv preprint arXiv:2405.17421, 2024. 2", + "[30] Renjie Li, Panwang Pan, Bangbang Yang, Dejia Xu, Shijie Zhou, Xuanyang Zhang, Zeming Li, Achuta Kadambi, Zhangyang Wang, and Zhiwen Fan. 4k4dgen: Panoramic 4d generation at 4k resolution. arXiv preprint arXiv:2406.13527, 2024. 2", + "[31] Zhengqi Li, Richard Tucker, Forrester Cole, Qianqian Wang, Linyi Jin, Vickie Ye, Angjoo Kanazawa, Aleksander Holynski, and Noah Snively. Megasam: Accurate, fast, and robust structure and motion from casual dynamic videos, 2024. 5", + "[32] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In Proceedings of the European conference on computer vision (ECCV), pages 85-100, 2018. 2", + "[33] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 3", + "[34] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 2, 5, 6", + "[35] Hao Ouyang, Kathryn Heal, Stephen Lombardi, and Tiancheng Sun. Text2immersion: Generative immersive scene with 3d gaussians. arXiv preprint arXiv:2312.09242, 2023. 2", + "[36] F. Perazzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Computer Vision and Pattern Recognition, 2016. 6, 7, 8", + "[37] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. UniDepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 4", + "[38] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 4", + "[39] Weize Quan, Jiaxi Chen, Yanli Liu, Dong-Ming Yan, and Peter Wonka. Deep learning-based image and video inpainting: A survey. International Journal of Computer Vision, 132(7): 2367-2400, 2024. 2", + "[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gener" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2", + "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2", + "[42] Iizuka Satoshi, Simo-Serra Edgar, and Ishikawa Hiroshi. Globally and locally consistent image completion. ACM Trans. Graph, 36(4):3073659, 2017. 2", + "[43] Jaidev Shriram, Alex Trevithick, Lingjie Liu, and Ravi Ramamoorthi. Realmdreamer: Text-driven 3d scene generation with inpainting and depth diffusion. arXiv preprint arXiv:2404.07199, 2024. 2", + "[44] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Armenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2149–2159, 2022. 2", + "[45] Luming Tang, Nataniel Ruiz, Qinghao Chu, Yuanzhen Li, Aleksander Holynski, David E Jacobs, Bharath Hariharan, Yael Pritch, Neal Wadhwa, Kfir Aberman, et al. Realfill: Reference-driven generation for authentic image completion. ACM Transactions on Graphics (TOG), 43(4):1-12, 2024. 2, 4, 5, 8", + "[46] Laura Turban, Fabrice Urban, and Philippe Guillotel. Extrafoveal video extension for an immersive viewing experience. IEEE transactions on visualization and computer graphics, 23(5):1520-1533, 2016. 1", + "[47] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 7", + "[48] Fu-Yun Wang, Xiaoshi Wu, Zhaoyang Huang, Xiaoyu Shi, Dazhong Shen, Guanglu Song, Yu Liu, and Hongsheng Li. Be-your-outpainter: Mastering video outpainting through input-specific adaptation. arXiv preprint arXiv:2403.13745, 2024. 1, 2, 6, 7, 8", + "[49] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 2", + "[50] Shizun Wang, Xingyi Yang, Qiuhong Shen, Zhenxiang Jiang, and Xinchao Wang. Gflow: Recovering 4d world from monocular video. arXiv preprint arXiv:2405.18426, 2024. 2", + "[51] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas S. Huang. Youtube-vos: A large-scale video object segmentation benchmark. CoRR, abs/1809.03327, 2018. 7, 8", + "[52] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18381-18391, 2023. 2" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7318", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Zhenqiang Ying and Alan Bovik. 180-degree outpainting from a single image. arXiv preprint arXiv:2001.04568, 2020. 1", + "[54] Hong-Xing Yu, Haoyi Duan, Charles Herrmann, William T Freeman, and Jiajun Wu. Wonderworld: Interactive 3d scene generation from a single image. arXiv preprint arXiv:2406.09394, 2024. 2", + "[55] Hong-Xing Yu, Haoyi Duan, Junhwa Hur, Kyle Sargent, Michael Rubinstein, William T Freeman, Forrester Cole, Deqing Sun, Noah Snavely, Jiajun Wu, et al. Wonderjourney: Going from anywhere to everywhere. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6658-6667, 2024. 2", + "[56] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contextual attention. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5505-5514, 2018. 2", + "[57] Yanhong Zeng, Jianlong Fu, Hongyang Chao, and Baining Guo. Aggregated contextual transformations for high-resolution image inpainting. IEEE Transactions on Visualization and Computer Graphics, 29(7):3266-3280, 2022. 2", + "[58] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 6", + "[59] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7", + "[60] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T Freeman. Structure and motion from casual videos. In European Conference on Computer Vision, pages 20-37. Springer, 2022. 3, 5", + "[61] Zhixing Zhang, Bichen Wu, Xiaoyan Wang, Yaqiao Luo, Luxin Zhang, Yinan Zhao, Peter Vajda, Dimitris Metaxas, and Licheng Yu. Avid: Any-length video inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7162-7172, 2024. 2", + "[62] Haitian Zheng, Zhe Lin, Jingwan Lu, Scott Cohen, Eli Shechtman, Connelly Barnes, Jianming Zhang, Ning Xu, Sohrab Amirghodsi, and Jiebo Luo. Image inpainting with cascaded modulation gan and object-aware training. In European Conference on Computer Vision, pages 277-296. Springer, 2022. 2", + "[63] Shijie Zhou, Zhiwen Fan, Dejia Xu, Haoran Chang, Pradyumna Chari, Suya Bharadwaj, Tejas You, Zhangyang Wang, and Achuta Kadambi. Dreamscene360: Unconstrained text-to-3d scene generation with panoramic gaussian splatting. arXiv preprint arXiv:2404.06903, 2024. 2", + "[64] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. arXiv preprint arXiv:2312.00451, 2023. 4" + ], + "bbox": [ + 78, + 90, + 470, + 867 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7319", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Unboxed_ Geometrically and Temporally Consistent Video Outpainting/3a973275-71b1-47b0-9c32-b162722059fe_model.json b/2025/Unboxed_ Geometrically and Temporally Consistent Video Outpainting/3a973275-71b1-47b0-9c32-b162722059fe_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e9cc04d2e6d72c6b2f745f7968d36732f2a01d80 --- /dev/null +++ b/2025/Unboxed_ Geometrically and Temporally Consistent Video Outpainting/3a973275-71b1-47b0-9c32-b162722059fe_model.json @@ -0,0 +1,2477 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.126, + 0.131, + 0.846, + 0.155 + ], + "angle": 0, + "content": "Unboxed: Geometrically and Temporally Consistent Video Outpainting" + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.18, + 0.248, + 0.2 + ], + "angle": 0, + "content": "Zhongrui \\(\\mathrm{Yu}^1\\)" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.181, + 0.468, + 0.2 + ], + "angle": 0, + "content": "Martina Megaro-Boldini²" + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.205, + 0.389, + 0.222 + ], + "angle": 0, + "content": "\\(^{1}\\)ETH Zürich" + }, + { + "type": "text", + "bbox": [ + 0.273, + 0.23, + 0.399, + 0.244 + ], + "angle": 0, + "content": "zhonyu@ethz.ch" + }, + { + "type": "text", + "bbox": [ + 0.484, + 0.181, + 0.65, + 0.199 + ], + "angle": 0, + "content": "Robert W. Sumner1,2" + }, + { + "type": "text", + "bbox": [ + 0.666, + 0.181, + 0.832, + 0.2 + ], + "angle": 0, + "content": "Abdelaziz Djelouah2" + }, + { + "type": "text", + "bbox": [ + 0.486, + 0.205, + 0.689, + 0.224 + ], + "angle": 0, + "content": "2DisneyResearch|Studios" + }, + { + "type": "text", + "bbox": [ + 0.44, + 0.23, + 0.699, + 0.244 + ], + "angle": 0, + "content": "adelaziz.djelouah@disney.com" + }, + { + "type": "image", + "bbox": [ + 0.052, + 0.261, + 0.38, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.394, + 0.261, + 0.925, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.488, + 0.893, + 0.517 + ], + "angle": 0, + "content": "Figure 1. Video outpainting results on two different resolutions. In both case, the area outside the red box is outpainted. Our method achieves results of better quality and more temporally stable, compared to recent works such as Dehan [13], M3DDM [14] and MOTIA [48]." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.53, + 0.314, + 0.546 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Extending the field of view of video content beyond its original version has many applications: immersive viewing experience with VR devices, reformatting 4:3 legacy content to today's viewing conditions with wide screens, or simply extending vertically captured phone videos. Many existing works focus on synthesizing the video using generative models only. Despite promising results, this strategy seems at the moment limited in terms of quality. In this work, we address this problem using two key ideas: 3D supported outpainting for the static regions of the images, and leveraging pre-trained video diffusion model to ensure realistic and temporally coherent results, particularly for the dynamic parts. In the first stage, we iterate between image outpainting and updating the 3D scene representation - we use 3D Gaussian Splatting. Then we consider dynamic objects independently per frame and inpaint missing pixels. Finally, we propose a denoising scheme that allows to maintain known reliable regions and update the dynamic parts to obtain temporally realistic results. We achieve state-of-the-art video outpainting. This is validated quantitatively and through a user study. We are also able to extend the field of view largely beyond the limits reached by existing methods." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.556, + 0.631, + 0.571 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.793 + ], + "angle": 0, + "content": "Expanding video frames beyond their original field of view (FoV), known as Video Outpainting, is a technique with broad applications in modern media. One example is 4:3 legacy content re-targeting where video outpainting can be used to fill modern widescreen devices seamlessly while avoiding any unnatural stretching. A more forward-looking application is movie viewing with virtual reality (VR) headsets. Video outpainting can enhance the immersion by embedding the viewer in the scene, eliminating distractions from unrelated backgrounds. However, these applications require extremely high-quality results, which is particularly challenging since the synthesized images must integrate harmoniously with the original content and maintain strong temporal coherence across frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Although a wide variety of methods have been proposed, using a patch-based approach [1, 2, 12, 46], optical flow [13] or Generative Adversarial Networks (GAN) [24, 53], most recent methods have clearly identified Diffusion Models [14, 48] as the most promising direction. M3DDM [14] trains a 3D diffusion model on large-scale data for video outpainting with a mask modeling approach." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7309" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "In contrast, MOTIA [48] proposes to fine-tune a pre-trained video diffusion model [19] on each test video to capture data-specific motion and content patterns. However, both of these methods have limitations. First, the temporal consistency of the outpainted regions is still unsatisfactory. We can observe duplications, distortions or objects appearing and disappearing. Second, relying entirely on video diffusion models has a strong impact in terms of memory and limits the methods to low resolution and small field of view extension. Some of these issues can be observed in Figure 1. Both M3DDM [14] and MOTIA [48] have strong visual artifacts and only the latter could be used to outpaint the video to a larger field of view." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.299, + 0.473, + 0.556 + ], + "angle": 0, + "content": "To address these issues we propose two core ideas: 3D-supported outpainting for static regions and conditional video synthesis with a pre-trained video diffusion model to ensure realistic, temporally coherent results for dynamic areas. Following these key ideas, we decompose the video into static and dynamic regions and break down the video outpainting task into three stages. First, for static regions, we rely on 3D Gaussians [23] as a supporting 3D representation, and we alternate between outpainting key frames and updating the 3D representation. Second, for dynamic objects, we implement an object-wise inpainting. Here, among other things, we leverage object-tracking masks in each frame to define inpainting regions and use a fine-tuned diffusion model to inpaint. Lastly, using the resulting frames as a starting point, we synthesize temporally consistent results for all frames using a pre-trained video diffusion model [7]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.473, + 0.719 + ], + "angle": 0, + "content": "Figure 1 demonstrates the effectiveness of our method. We achieve better results than prior methods both in terms of video quality and temporal consistency. This is confirmed by both quantitative and qualitative experiments. Furthermore, by using GPU-efficient Gaussian Splatting for static region outpainting and avoiding fine-tuning of the video diffusion model, we can go from an input \\(480 \\times 480\\) video (\\(\\mathrm{FoV} = 31^{\\circ}\\)) to an output resolution of \\(2560 \\times 720\\) (\\(\\mathrm{FoV} = 120^{\\circ}\\)) with a GPU memory usage of less than 16GB for the entire pipeline." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.73, + 0.399, + 0.744 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.757, + 0.468, + 0.785 + ], + "angle": 0, + "content": "- a novel video outpainting method that combines a 3D representation with a pre-trained video diffusion model;" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.787, + 0.468, + 0.83 + ], + "angle": 0, + "content": "- an optimization strategy with an additional loss to enable seamless integration of the outpainting with the original static content;" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.832, + 0.468, + 0.891 + ], + "angle": 0, + "content": "- demonstrating that a pre-trained video diffusion model can function as a temporal filter, making discontinuous frames temporally coherent and simplifying the inpainting task for the dynamic parts." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.893, + 0.384, + 0.907 + ], + "angle": 0, + "content": "- State-of-the-art results for video outpainting" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.757, + 0.468, + 0.907 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.105 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.417 + ], + "angle": 0, + "content": "Immersive Scene Generation. To create immersive experiences for users in VR applications, several methods are proposed to generate realistic 3D or 4D scenes that allow for arbitrary viewpoint exploration. LucidDreamer [10] and the following works [16, 35, 43, 54, 55] lift a single image to 3D space, project it from different viewpoints, and inpaint them using diffusion models to create a 3D scene. However, directly applying their method to consecutive video frames would cause severe geometric distortion between frames due to the simple depth alignment and optimization techniques. Dreamscene360 [63] and 4D4KGen [30] create panoramic 3D and 4D scenes separately from a text-generated panoramic image. VividDream [27] generates multi-view videos from a single image and reconstructs a 4D scene with 4D Gaussians. These methods are primarily designed for static or simply animated images and cannot handle videos with complex camera and object motion. Methods such as [9, 28, 29, 49, 50] reconstruct 4D scenes from casual videos, but they are unable to complete regions beyond what is captured in the video." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.892, + 0.674 + ], + "angle": 0, + "content": "Image completion. Image completion [6, 39], including image inpainting and outpainting, is a task that restores missing part of an image, and it serves as a basis for video outpainting. Traditional methods [6, 18, 21] use low-level features from the incomplete image to propagate information in known regions to fill missing areas. Early deep learning based methods [32, 42, 44] introduced end-to-end frameworks that learned mappings from corrupted to complete images. Recent methods have increasingly leveraged generative models, including GAN [17, 56, 57, 62], and Diffusion Models [4, 11, 22, 34, 40, 41]. Reference-based image completion methods utilize reference images as generation conditions [52] or fine-tune models on a set of reference images [45] to produce completions that are more faithful to the references. However, even fine-tuned on all input video frames, frame-by-frame completion introduces substantial temporal inconsistency in video outpainting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Video Outpainting. In the field of video completion, video outpainting is challenging but less explored. Early patch-based methods [1, 2] select the nearest patch from the spatio-temporal domain of the video to expand each frame. Lee et al. [26] employ 3D information to warp neighboring frames to the current frame. Dehan et al. [13] utilize estimated optical flow to warp the background from adjacent frames and use an image completion network to fill any remaining uncompleted areas. While Zhang et al. [61] target video in-painting motion module and structure guidance. M3DDM [14] extends the image diffusion model framework to 3D and trains it on extensive video data for the outpainting task. A global frame is introduced as a conditioning input to improve global information perception. MOTIA [48] proposes a method that fine-tunes a pre-trained" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7310" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.09, + 0.121, + 0.099 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.101, + 0.237, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.165, + 0.165, + 0.177 + ], + "angle": 0, + "content": "Pre-process" + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.181, + 0.182, + 0.19 + ], + "angle": 0, + "content": "Dynamic Object Masks" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.191, + 0.143, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.145, + 0.191, + 0.195, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.196, + 0.191, + 0.242, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.242, + 0.2, + 0.25 + ], + "angle": 0, + "content": "Static Scene Reconstruction" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.255, + 0.237, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.09, + 0.352, + 0.104 + ], + "angle": 0, + "content": "(1) Static Region" + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.109, + 0.458, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.465, + 0.091, + 0.587, + 0.104 + ], + "angle": 0, + "content": "(2) Dynamic Objects" + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.109, + 0.692, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.708, + 0.091, + 0.858, + 0.104 + ], + "angle": 0, + "content": "(3) Final Video Synthesis" + }, + { + "type": "image_caption", + "bbox": [ + 0.73, + 0.108, + 0.84, + 0.116 + ], + "angle": 0, + "content": "Temporally Consistent Results" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.117, + 0.889, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.356, + 0.894, + 0.44 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed method. Starting from the input video sequence, dynamic objects are separated from the static parts of the image, and a static scene reconstruction is made using Gaussian Splatting (GS). In the first stage of our method, the static part is outpainted by iterating between image outpainting, depth prediction and updating the GS model. After this stage, we identify parts of the dynamic object that may require inpainting. Here, an image diffusion model is finetuned on the input frames and used for inpainting each frame independently, which may result in temporal inconsistencies as seen of the car windows. In the last step, a pre-trained video diffusion model is used to fix temporal inconsistencies. Details about each step are provided in the text." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.449, + 0.47, + 0.54 + ], + "angle": 0, + "content": "video diffusion model on test samples to learn the intrinsic patterns of a video. In general, diffusion-based approaches rely on priors learned by temporal modules to maintain continuity in the outpainted regions, which is not sufficient. We demonstrate that a pipeline integrating constraints from a 3D representation achieves results of much higher quality." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.556, + 0.168, + 0.571 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.47, + 0.718 + ], + "angle": 0, + "content": "Given a sequence of video frames \\(\\{I_i\\in \\mathbb{R}^{h\\times w\\times c}\\}_{i = 1}^N\\), our goal is to expand each frame beyond its original field of view to \\(\\{\\tilde{I}_i\\in \\mathbb{R}^{H\\times W\\times c}\\}_{i = 1}^N\\), while maintaining both intra-frame and cross-frame consistency. Although the problem is more challenging than image outpainting, due to the temporal consistency requirements, the presence of multiple frames offers additional observations that can be leveraged. More specifically, it is possible to build a 3D representation of the scene using 3D Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "As illustrated in Figure 2, our proposed method can be divided into 3 distinct stages. In the first stage, we focus on the static parts of the scene and alternate between image outpainting and updating the 3D scene representation. In the second stage, we focus on the dynamic objects in the scene. Here, we fine-tune an image diffusion model on the available images and use it to inpaint the missing parts of the object in each frame. At this point, all the frames are outpainted, but temporal inconsistencies remain. These frames serve as a starting point for a video synthesis scheme leveraging pre-trained stable video diffusion to synthesize a temporally consistent video sequence. Next, we detail each" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.449, + 0.633, + 0.464 + ], + "angle": 0, + "content": "of these main steps." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.475, + 0.879, + 0.493 + ], + "angle": 0, + "content": "3.1. Pre-processing and Static Scene Outpainting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.892, + 0.697 + ], + "angle": 0, + "content": "From the static parts of the video, it is possible to estimate a 3D reconstruction of the scene and use the different views across time to provide information about the regions to outpaint. In the first step, we identify the dynamic parts of the scene. A wide variety of methods can be employed here, but we simply rely on object segmentation and, in particular, object tracking models [33]. Figure 2 shows an example of the binary segmentation masks. These masks are expanded to avoid including incorrectly labeled pixels into the background. We also compute a 2D bounding box around the object to help keep track of the object size in case it partially or completely leaves the field of view. The static parts are all the pixels that are not part of any dynamic object." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Static Scene Reconstruction. We propose to use Gaussian Splatting (GS) [23] to reconstruct the static part of the video. This approach allows regions visible in other frames to be rendered into the current frame. For the input video frames \\(\\{I_i\\}_{i=1}^N\\), we first estimate the corresponding camera poses \\(\\{\\mathbf{p}_i\\}_{i=1}^N\\) and intrinsic parameters \\(\\{\\mathbf{K}_i\\}_{i=1}^N\\) using an off-the-shelf camera pose estimator [60]. We compute a depth map for each frame using parallax and project the background pixels into 3D space, yielding an initial pointcloud for GS training. The training process is supervised by the image reconstruction loss and the depth loss. The image reconstruction loss is the combination of the L1 loss and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7311" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.15, + 0.092, + 0.26, + 0.104 + ], + "angle": 0, + "content": "First key-frame \\((k = 1)\\)" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.106, + 0.2, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.173, + 0.155, + 0.186 + ], + "angle": 0, + "content": "\\(I_{k = 1}^{\\mathrm{static}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.205, + 0.106, + 0.326, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.173, + 0.28, + 0.185 + ], + "angle": 0, + "content": "\\(I_{k = 1}^{\\mathrm{static}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.106, + 0.452, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.172, + 0.409, + 0.183 + ], + "angle": 0, + "content": "\\(L_{\\mathrm{recon}}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.092, + 0.584, + 0.104 + ], + "angle": 0, + "content": "Depth after GS finetuning" + }, + { + "type": "image", + "bbox": [ + 0.455, + 0.106, + 0.578, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.475, + 0.172, + 0.559, + 0.184 + ], + "angle": 0, + "content": "\\(L_{\\mathrm{recon}} + L_{\\mathrm{depth}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.106, + 0.704, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.585, + 0.172, + 0.702, + 0.185 + ], + "angle": 0, + "content": "\\(L_{\\mathrm{recon}} + L_{\\mathrm{depth}} + L_{\\mathrm{depth}}^{\\mathrm{Sp}}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.093, + 0.878, + 0.104 + ], + "angle": 0, + "content": "Next key-frame \\((k = 2)\\)" + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.106, + 0.885, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.174, + 0.842, + 0.19 + ], + "angle": 0, + "content": "\\(I_{k = 2}^{\\mathrm{static}}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.192, + 0.894, + 0.249 + ], + "angle": 0, + "content": "Figure 3. Effect of the loss terms on the GS fine-tuning. Starting from the rasterization of the 3D Gaussians for the first key-frame \\( k = 1 \\), we use image outpainting to obtain \\( \\hat{I}_k^{\\mathrm{static}} \\). This image is subdivided into super-pixels as illustrated here. During the optimization to update the GS model, we use a combination of losses. We can see that all are needed to avoid discontinuities and obtain a realistic smooth depth map. After this update, the process is repeated with the next key-frame using the updated 3D Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.264, + 0.47, + 0.355 + ], + "angle": 0, + "content": "the structural similarity loss (SSIM). The depth loss acts as a regularization and uses Pearson Correlation [64] between the rendered depth and the depth estimated from a monocular depth estimation model [37]. Apart from the dynamic object masking, this is similar to the optimization proposed by Zhu et al. [64]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.471, + 0.584 + ], + "angle": 0, + "content": "Single Image Outpainting. Starting from the GS model \\(GS_{\\mathrm{static}}\\) representing the static parts of the scene, we uniformly sample \\(k \\in \\{1, \\dots, K\\}\\) key frames. Our objective is to outpaint them one by one, while we progressively update the GS model. Here we first detail outpainting for a single image. Considering frame \\(I_k^{\\mathrm{static}}\\), we adjust its field of view to obtain new intrinsic parameters \\(\\hat{\\mathbf{K}}_k\\), and render \\(GS_{\\mathrm{static}}\\) based on the new parameters to get the image \\(\\hat{I}_k^{\\mathrm{static}}\\). The mask \\(\\hat{\\mathbf{m}}_k^{\\mathrm{static}}\\) for outpainting is defined by checking whether each pixel is covered by the projection of Gaussian points. Any image outpainting method can be used at this level, but we find that the Stable Diffusion XL (SDXL) image outpainting model [38] performs well for our task. This can be summarized as:" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.595, + 0.469, + 0.613 + ], + "angle": 0, + "content": "\\[\nI _ {k} ^ {\\text {s t a t i c}} = \\phi_ {G S _ {\\text {s t a t i c}}} \\left(\\hat {\\mathbf {K}} _ {k}, \\mathbf {p} _ {k}\\right) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.616, + 0.469, + 0.635 + ], + "angle": 0, + "content": "\\[\n\\hat {I} _ {k} ^ {\\text {s t a t i c}} = \\phi_ {\\mathrm {S D X L}} \\left(I _ {k} ^ {\\text {s t a t i c}}, \\hat {\\mathbf {m}} _ {k} ^ {\\text {s t a t i c}}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.651, + 0.47, + 0.786 + ], + "angle": 0, + "content": "Updating the GS model. To ensure the consistency of the outpainting across frames, we update the GS model each time we outpaint a key-frame. Denote by \\( K' \\) the number of key-frames that have been outpainted so far. First, we estimate the depth for the newly outpainted image using a single image depth model [37]. Then we initialize new Gaussian points for the outpainted pixel and fine-tune the GS model with this additional data. We use an image reconstruction loss" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.795, + 0.469, + 0.859 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {\\text {r e c o n}} = \\sum_ {k = 1} ^ {K ^ {\\prime}} \\lambda_ {1} L _ {1} \\left(\\hat {I} _ {k} ^ {\\text {s t a t i c}}, \\phi_ {\\mathrm {G S}} \\left(\\hat {K} _ {k}, p _ {k}\\right)\\right) \\tag {3} \\\\ + \\lambda_ {2} L _ {\\mathrm {S S I M}} (\\hat {I} _ {k} ^ {\\text {s t a t i c}}, \\phi_ {\\mathrm {G S}} (\\hat {K} _ {k}, p _ {k})) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.87, + 0.47, + 0.901 + ], + "angle": 0, + "content": "between the rasterization from the model \\(\\phi_{\\mathrm{GS}}(\\hat{K}_k,p_k)\\) and the outpainting result. We also use 2 depth based losses, the" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.264, + 0.706, + 0.28 + ], + "angle": 0, + "content": "first one on the full depth map:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.292, + 0.892, + 0.34 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {d e p t h}} = \\sum_ {k = 1} ^ {K ^ {\\prime}} \\frac {\\operatorname {C o v} \\left(D \\left(\\hat {I} _ {k} ^ {\\text {s t a t i c}}\\right) , D _ {\\mathrm {G S}} \\left(\\hat {K} _ {k} , p _ {k}\\right)\\right)}{\\sqrt {\\operatorname {V a r} \\left(D \\left(\\hat {I} _ {k} ^ {\\text {s t a t i c}}\\right)\\right) \\operatorname {V a r} \\left(D _ {\\mathrm {G S}} \\left(\\hat {K} _ {k} , p _ {k}\\right)\\right)}} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.352, + 0.893, + 0.416 + ], + "angle": 0, + "content": "where \\(D(\\hat{I}_k^{\\mathrm{static}})\\) is the depth predicted by the single image depth model [37], and \\(D_{\\mathrm{GS}}(\\hat{K}_k,p_k)\\) is the depth from GS model that we optimize. For every image, we extract \\(S\\) super-pixels, and use the second depth loss at this level:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.429, + 0.892, + 0.489 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {d e p t h}} ^ {\\mathrm {S p}} = \\sum_ {k = 1} ^ {K ^ {\\prime}} \\sum_ {s = 1} ^ {S} \\frac {\\operatorname {C o v} _ {s} \\left(D \\left(\\hat {I} _ {k} ^ {\\text {s t a t i c}}\\right) , D _ {\\mathrm {G S}} \\left(\\hat {K} _ {k} , p _ {k}\\right)\\right)}{\\sqrt {\\operatorname {V a r} _ {s} \\left(D \\left(\\hat {I} _ {k} ^ {\\text {s t a t i c}}\\right)\\right) \\operatorname {V a r} _ {s} \\left(D _ {\\mathrm {G S}} \\left(\\hat {K} _ {k} , p _ {k}\\right)\\right)}} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.893, + 0.685 + ], + "angle": 0, + "content": "Figure 3 demonstrate this process and shows the importance of the different loss terms for the GS update. Starting from the rasterization of the initial GS model, we can see that the 3D model does not cover the new field of view. The frame is outpainted and superpixels are illustrated in yellow. Using only the reconstruction loss \\( L_{\\mathrm{recon}} \\) leads to strong discontinuities between the original GS model and the new outpainted areas. Adding the depth loss \\( L_{\\mathrm{depth}} \\) helps, but the bottom part of the image still has important discontinuities that would be revealed when projecting to other frames (i.e. other views). Using all the proposed loss terms provides the best results, and the process can now be iterated by adding new views." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.696, + 0.874, + 0.713 + ], + "angle": 0, + "content": "3.2. Initial Inpainting for the Dynamic Object(s)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.809 + ], + "angle": 0, + "content": "After the static scene outpainting, we can render any frame with \\(\\phi_{\\mathrm{GS}}(\\hat{K}_i,p_i)\\). As illustrated in Figure 2, we miss part of the object when it is partially outside the filed of view. We need to inpaint this missing part. We use the 2D tracking bounding box to identify the areas that require inpainting for each frame of the video." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.893, + 0.9 + ], + "angle": 0, + "content": "For this inpainting problem we use a similar strategy as ReaFill [45]. An image diffusion model is fine-tuned for the inpainting task on the input frames (in our case SDXL). It is then applied independently on each frame. Thanks to the fine-tuning, the resulting inpainting is very similar across frames. However temporal inconsistencies remain as it is" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7312" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.07, + 0.089, + 0.478, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.273, + 0.47, + 0.37 + ], + "angle": 0, + "content": "Figure 4. Details of the video latent denoising strategy. Starting from the independently outpainted frames \\(\\tilde{I}_{i\\in 1,\\dots,N}\\), we obtained the corresponding latents in the forward process \\((\\tilde{x}_{i\\in 1,\\dots,N})\\). In the denoising process we blend known latent with denoising latents according to the mask \\(\\mathbf{m}\\). The mask and the time step \\(t_{\\mathrm{end}}\\) respectively control which region of the frames get updated by the video diffusion model and how large the changes are." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.47, + 0.428 + ], + "angle": 0, + "content": "clearly visible on the inpainting results in Figure 2. Next we describe how a Video Diffusion Model [7, 8] is used to fix all these temporal artifacts." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.436, + 0.296, + 0.451 + ], + "angle": 0, + "content": "3.3. Guided Video Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.46, + 0.469, + 0.627 + ], + "angle": 0, + "content": "If we consider the original input sequences of frames \\( I_{i\\in 1,\\dots,N} \\), after the previous stages we obtain fully out-painted frames \\( \\tilde{I}_{i\\in 1,\\dots,N} \\). Static regions are consistent thanks to the supporting GS model, while the dynamic objects are independently inpainted for each frame. Starting from this data, our objective is to synthesize a realistic sequence of frames \\( I_{i\\in 1,\\dots,N}^{*} \\). Not only the temporal inconsistencies on the dynamic objects need to be fixed, but the entire outpainted regions need to appear realistic. Indeed, real-world videos often contain subtle background movements, such as leaves swaying in the wind or water rippling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.469, + 0.777 + ], + "angle": 0, + "content": "We leverage pre-trained video diffusion models for this video synthesis. The idea is that video diffusion models are trained for generating real video sequences, with consistent content and realistic motion. We propose using the previously generate frames in a denoising scheme that modifies the different frame areas based on the expected inconsistency level. On the dynamic objects, we expect more changes (as the inpainting is independent), while less change should be allowed on the rest of the image as it is supported by the 3D consistent GS model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.778, + 0.469, + 0.852 + ], + "angle": 0, + "content": "More specifically, if we consider a latent video diffusion model, a sequence of input frames \\( I_{i} \\) is encoded into the corresponding sequence of latent values \\( x_{i} \\). The forward process can be applied to the sequence of frames \\( \\tilde{I}_{i\\in 1,\\dots,N} \\) to obtain known latents" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.861, + 0.469, + 0.878 + ], + "angle": 0, + "content": "\\[\n\\tilde {x} _ {t - 1} \\sim \\mathcal {N} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\tilde {x} _ {0}, \\sqrt {1 - \\bar {\\alpha} _ {t}} \\mathbf {1}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.378, + 0.901 + ], + "angle": 0, + "content": "For simplicity, we dropped the frame index \\(i\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "We use the reverse process to synthesize the final video sequence. At each time step, we denoise previous latents using the latent video diffusion model with parameters \\(\\theta\\), to obtain temporary latents" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.163, + 0.892, + 0.181 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1} ^ {\\operatorname {t m p}} \\sim \\mathcal {N} \\left(\\mu_ {\\theta} \\left(x _ {t} ^ {*}, t\\right), \\Sigma_ {\\theta} \\left(x _ {t} ^ {*}, t\\right)\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.193, + 0.891, + 0.223 + ], + "angle": 0, + "content": "To obtain the final latents for this time step, we blend in the known latents according to" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.235, + 0.892, + 0.253 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1} ^ {*} = (1 - \\mathbf {m}) \\odot \\tilde {x} _ {t - 1} + \\mathbf {m} \\odot x _ {t - 1} ^ {\\mathrm {t m p}} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.264, + 0.756, + 0.278 + ], + "angle": 0, + "content": "where the binary mask \\(\\mathbf{m}\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.287, + 0.892, + 0.325 + ], + "angle": 0, + "content": "\\[\n\\mathbf {m} = \\left\\{ \\begin{array}{l l} \\mathbf {m} ^ {\\text {d y n a m i c}} & \\text {i f} t \\geq t _ {\\text {e n d}} \\\\ \\mathbf {m} ^ {\\text {d y n a m i c}} + \\mathbf {m} ^ {\\text {s t a t i c}} & \\text {i f} t < t _ {\\text {e n d}} \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.336, + 0.893, + 0.501 + ], + "angle": 0, + "content": "The mask, \\(\\mathbf{m}^{\\mathrm{dynamic}}\\), corresponds to the dynamic objects in the scene, and the mask \\(\\mathbf{m}^{\\mathrm{static}}\\) corresponds to the static parts in the outpainted regions. This scheme is similar to ideas already explored in the context of image diffusion models [3, 34], but applied to video diffusion models. If we observe the changes in Figure 2, we can see that the pre-trained video diffusion model can function as a temporal filter, making discontinuous frames temporally coherent. This simplifies the inpainting task since we do not rely on expensive training of the video model and simply use it at inference time." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.516, + 0.633, + 0.533 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.541, + 0.715, + 0.557 + ], + "angle": 0, + "content": "4.1. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.892, + 0.82 + ], + "angle": 0, + "content": "Static Scene. In pre-processing, we estimate the camera poses of the source video [31, 60]. Our GS training process consists of \\(1 + K + 1\\) rounds in total, where \\(K\\) is the number of key frames we sample from the video. In the first training round, we train the GS model only on the static part of the video for 3000 iterations. Then we iterate between outpainting the \\(K\\) key frames and \\(K\\) rounds of GS training to update the outpainted regions into the GS model. To avoid overfitting, we employ a sliding window training strategy: following outpainting of the \\(i\\)-th frame, we train only on frames within the window \\([i - 5, i + 5]\\) for 500 iterations. This window shifts progressively as outpainting and training continues. After outpainting all \\(K\\) key frames, we conduct a final round of training for refinement on all out-painted and original background frames for 1000 iterations. All training and outpainting can be completed on a single 4090 GPU." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Initial Object Inpainting. Before initial inpainting on the missing regions of dynamic objects, we fine-tune the SDXL model on the masked source video. The approach is similar to the one proposed by Tang et al. [45]. Masks are randomly" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7313" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.07, + 0.09, + 0.878, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.582, + 0.893, + 0.64 + ], + "angle": 0, + "content": "Figure 5. Qualitative Comparison with State-of-the-art methods on the DAVIS dataset [36]. The area outside of the red lines (only shown for the first frame) is outpainted. We show the results for 3 frames from the each video sequence. Compared to MOTIA [48] and M3DDM[14], our results are of better quality and temporally stable. Please pay attention to the changes of the outpainted area between the frames of the same sequence." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.473, + 0.762 + ], + "angle": 0, + "content": "generated for each input frame, with random distributed positions and mask ratios varying from 0.2 to 0.8. For each object, a prompt \"a photo of [v]\" is applied. During fine-tuning, a Low-Rank Adaption (LoRA) [20] with rank 8 is trained for both the text-encoder and UNet with a batch size of 4 and gradient accumulation steps of 4 for 2000 steps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Video Diffusion Model. We use a pretrained image-to-video model [7] to complete object motion. The number of denoising steps is set to 25, with the level of noise added to the reconstructed frames generally set to 0.8 in our experiments. We apply a resampling strategy similar to Repaint [34] five times after each denoising step to improve the quality and stability of the objects. For videos where the object fully appears in the middle or last frame, we em" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.655, + 0.895, + 0.701 + ], + "angle": 0, + "content": "ploy the Time-Reversal strategy from [15], conditioning the generation on additional frames to better retain the object identity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.711, + 0.678, + 0.728 + ], + "angle": 0, + "content": "4.2. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Competitors. We compare our method with the following methods. (1) Dehan et al. [13] separate foreground and background components. They use estimated optical flow to warp the background, with missing regions filled in using an image completion network. (2) M3DDM [14] proposes a masked 3D Diffusion Model architecture, which is trained on WebVid [5] and a self-collected 5M e-commerce dataset for video outpaint tasks. (3) MOTIA [48] is the current state-of-the-art method for video outpainting. It is based on the ControlNet [58] inpainting model and the temporal module fromAnimateDiff [19]. It trains a LoRA [20] for" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7314" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.384, + 0.106 + ], + "angle": 0, + "content": "each test sample to learn the intrinsic patterns." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.47, + 0.227 + ], + "angle": 0, + "content": "Datasets. Following M3DDM [14] and MOTIA [48], we evaluate our method on the DAVIS [36] and YouTube-VOS [51] datasets. For the YouTube-VOS [51] dataset, due to its large number of videos and the lack of clarity in prior works on which videos were tested, we selected a high-quality subset and evaluated all methods on it. Consistent with prior works, we compare the horizontal outpainting results for each test video at mask ratios of 0.25 and 0.66." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.246, + 0.473, + 0.504 + ], + "angle": 0, + "content": "Evaluation Metrics. The quality of outpainting is primarily evaluated in terms of visual quality, realism, and temporal consistency. For this purpose, we employ two well-established metrics: Fréchet Video Distance (FVD) [47] and flow warping error \\((E_{\\mathrm{warp}})\\) [25]. FVD assesses the similarity between the distribution of the outpainted video and the original video, while flow warping error evaluates temporal consistency by measuring the error when outpainted frames are warped to adjacent frames using estimated optical flow. In addition, to align with previous works, we also report Peak Signal to Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) [59] between the outpainted and ground-truth videos, although we consider these metrics limited in capturing the quality of outpainting. We further conducted a user study to gather subjective evaluations of outpainting quality, which we detail in Sec. 4.3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.512, + 0.279, + 0.527 + ], + "angle": 0, + "content": "4.3. Evaluation & Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.473, + 0.763 + ], + "angle": 0, + "content": "Qualitative Results. Fig. 5 and Fig. 6 illustrate the qualitative comparison of our method against other approaches on the DAVIS and YouTube-VOS datasets. M3DDM [14] frequently suffers from generation failures and produces blurred content, as shown in the first samples of Fig. 5 and Fig. 6. MOTIA [48], while fine-tuning on each test sample to capture input-specific patterns, often generates repetitive objects and background patterns. More importantly, both methods struggle to accurately capture contextual information across frames, resulting in inconsistency when outpainting elements seen in adjacent frames. In contrast, our method overcomes these issues, achieving better cross-frame consistency. Furthermore, our approach's object-specific completion approach prevents objects from duplicating or disappearing beyond the frame." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Generalization and multiple objects. Our strategy, differentiating between static and dynamic elements, can naturally extend to multiple dynamic objects by decomposing the frame into object layers based on their depth. Each object is processed separately using our pipeline and then seamlessly blended to produce the final result. Figure 7 shows that our method still performs best on such complex scenarios." + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.091, + 0.89, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.341, + 0.893, + 0.44 + ], + "angle": 0, + "content": "Figure 6. Qualitative Comparison with State-of-the-art methods on Youtube-VOS dataset [51]. The area outside of the red lines (only shown for the first frame) is outpainted. We show the results for 3 frames from the each video sequence. Compared to Dehan [13], MOTIA [48] and M3DDM[14], our results are of better quality and temporally stable. Please pay attention to the changes of the outpainted area between the frames of the same sequence." + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.456, + 0.891, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.639, + 0.892, + 0.682 + ], + "angle": 0, + "content": "Figure 7. Qualitative Comparison with State-of-the-art methods on challenging scenarios (multiple dynamic objects, complex static regions, etc.)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.699, + 0.892, + 0.851 + ], + "angle": 0, + "content": "Quantitative Results. Tab. 1 presents a quantitative comparison of our method with other approaches. Our method outperforms the current state-of-the-art in both FVD and \\( E_{\\text{warp}} \\), achieving improvements of \\( 18.8\\% \\) and \\( 11.7\\% \\) on the DAVIS dataset and \\( 26.9\\% \\) and \\( 33.3\\% \\) on the YouTube-VOS dataset, respectively. We also achieve the highest PSNR and SSIM scores on both datasets. For LPIPS, while our method is the second behind the number reported in MOTIA [48] on the DAVIS dataset, it surpasses the score obtained from using their public implementation and provided model." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Ablation Study We conducted ablation studies for the different components of our method. Quantitative evalua" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7315" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.089, + 0.887, + 0.219 + ], + "angle": 0, + "content": "
DAVISYouTube-VOS
PSNR↑SSIM↑LPIPS↓FVD↓E_warp↓PSNR↑SSIM↑LPIPS↓FVD↓E_warp↓
Dehan [13]17.960.6270.233363.10.03018.340.7610.287353.60.043
M3DDM [14]20.260.7080.204300.00.01718.330.7250.424283.10.024
MOTIA (model) [48]18.250.6990.282294.80.02118.540.7900.250156.40.027
MOTIA (paper) [48]20.360.7580.160286.3------
Ours22.620.8060.200232.50.01519.880.8320.236114.20.016
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.228, + 0.895, + 0.272 + ], + "angle": 0, + "content": "Table 1. Quantitative Comparison with State-of-the-art methods on DAVIS dataset [36] and YouTube-VOS [51]. Our method achieves the best performance overall. In the case of MOTIA [48], we have 2 rows as we report the evaluation using the model and code available online(model) and the numbers from the original paper (paper). See the text for details." + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.293, + 0.434, + 0.402 + ], + "angle": 0, + "content": "
PSNR↑SSIM↑LPIPS↓FVD↓E_warp↓
(a)18.150.6940.248448.80.032
(b)18.050.7160.274323.20.022
(c)20.680.7720.218293.30.020
(d)20.050.7760.223319.80.017
(e)22.620.8060.200232.50.015
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.412, + 0.472, + 0.468 + ], + "angle": 0, + "content": "Table 2. Quantitative ablation results. (a). Outpainting with fine-tuned Image Model; (b). (a) + Guided Video Synthesis; (c). Our model without guided video synthesis; (d). Our model without the initialization from object inpainting; (e). Our complete model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.47, + 0.713 + ], + "angle": 0, + "content": "tion is shown in Table 2, while the visual results are provided in supplementary material. When only using single image outpainting (SDXL fine-tuned on content [45]), the images achieve good realism but lack temporal continuity. When applying video diffusion guided synthesis to these outpainted frames directly, the temporal consistency within the outpainted areas improves individually. However, coherence between the outpainted regions and the original video content remains weak. This issue is difficult to resolve solely with Video Diffusion Model priors. When using the static region outpainting (assisted with GS) and either the image inpainting or video diffusion alone are applied for object completion, the objects suffer from significant temporal variations in appearance. Using all the components achieves the best results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "User Study. Given the subjective nature of evaluating outpainting quality, we conducted a user study to compare our method with competing approaches. Users were asked to evaluate each method across several dimensions, including realism (such as whether the outpainted results look natural and harmonious), temporal consistency (including background consistency, object motion consistency, etc.), and overall visual quality (including color fidelity, smoothness of boundaries, blurriness etc.). The results, shown in Fig. 8, indicate that our method was clearly preferred by users compared to others, achieving more than \\(80\\%\\) of the votes in all dimensions. Further details of the user study can" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.285, + 0.887, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.391, + 0.893, + 0.432 + ], + "angle": 0, + "content": "Figure 8. The result of the user study includes 619 votes from 37 participants. Our method is preferred by more than \\(80\\%\\) of the votes in all dimensions." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.446, + 0.892, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.571, + 0.894, + 0.613 + ], + "angle": 0, + "content": "Figure 9. Challenges with dynamic objects moving in and out of the original frame: Our current solution of tracking and extrapolating the trajectory of the bounding box can be improved." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.621, + 0.771, + 0.636 + ], + "angle": 0, + "content": "be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.649, + 0.621, + 0.664 + ], + "angle": 0, + "content": "5. Discussions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Building on Gaussian Splatting and a Video Diffusion Model, we propose a new state-of-the-art method for video outpainting. Despite our strong results, some issues remain, especially around objects moving in and out of the original frame. An example is shown in Fig. 9. Currently tracking and extrapolating the trajectory of the bounding box offers a solution, but limitations in the single image inpainting negatively impact the quality of the final result. Another area of improvement is the outpainting involving multiple occluded dynamic objects. Our current solution of inpainting each object separately and blending them based on the estimated depth is limited. Exploring the use of more advanced 4D dynamic Gaussian representations jointly with video models offers a promising future direction towards higher and higher quality video outpainting." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7316" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Amit Aides, Tamar Avraham, and Yoav Y Schechner. Multiscale ultrawide foveated video extrapolation. In 2011 IEEE International Conference on Computational Photography (ICCP), pages 1-8. IEEE, 2011. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.47, + 0.213 + ], + "angle": 0, + "content": "[2] Tamar Avraham and Yoav Y Schechner. Ultrawide foveated video extrapolation. IEEE Journal of Selected Topics in Signal Processing, 5(2):321-334, 2010. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.471, + 0.282 + ], + "angle": 0, + "content": "[3] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18208-18218, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.283, + 0.471, + 0.34 + ], + "angle": 0, + "content": "[4] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18208-18218, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.471, + 0.395 + ], + "angle": 0, + "content": "[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.353, + 0.41 + ], + "angle": 0, + "content": "[6] M Bertalmio. Image inpainting, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.411, + 0.471, + 0.48 + ], + "angle": 0, + "content": "[7] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.471, + 0.563 + ], + "angle": 0, + "content": "[8] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.471, + 0.618 + ], + "angle": 0, + "content": "[9] Wen-Hsuan Chu, Lei Ke, and Katerina Fragkiadaki. Dreamscene4d: Dynamic multi-object scene generation from monocular videos. arXiv preprint arXiv:2405.02280, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.621, + 0.471, + 0.675 + ], + "angle": 0, + "content": "[10] Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.677, + 0.471, + 0.745 + ], + "angle": 0, + "content": "[11] Ciprian Corneanu, Raghudeep Gadde, and Aleix M Martínez. Latentpaint: Image inpainting in latent space with diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4334-4343, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.747, + 0.471, + 0.802 + ], + "angle": 0, + "content": "[12] Antonio Criminisi, Patrick Pérez, and Kentaro Toyama. Region filling and object removal by exemplar-based image inpainting. IEEE Transactions on image processing, 13(9): 1200-1212, 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.803, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[13] Loic Dehan, Wiebe Van Ranst, Patrick Vandewalle, and Toon Goedemé. Complete and temporally consistent video outpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 687-695, 2022. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.872, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[14] Fanda Fan, Chaoxu Guo, Litong Gong, Biao Wang, Tiezheng Ge, Yuning Jiang, Chunjie Luo, and Jianfeng Zhan. Hierar-" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "chical masked 3d diffusion model for video outpainting. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7890-7900, 2023. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.135, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[15] Haiwen Feng, Zheng Ding, Zhihao Xia, Simon Niklaus, Victoria Abrevaya, Michael J Black, and Xuaner Zhang. Explorative inbetweening of time and space. In European Conference on Computer Vision, pages 378-395. Springer, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.204, + 0.892, + 0.258 + ], + "angle": 0, + "content": "[16] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.26, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.372 + ], + "angle": 0, + "content": "[18] Qiang Guo, Shanshan Gao, Xiaofeng Zhang, Yilong Yin, and Caiming Zhang. Patch-based image inpainting via two-stage low rank approximation. IEEE transactions on visualization and computer graphics, 24(6):2023-2036, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[19] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.441, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.497, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[21] Jia-Bin Huang, Sing Bing Kang, Narendra Ahuja, and Johannes Kopf. Image completion using planar structure guidance. ACM Transactions on graphics (TOG), 33(4):1-10, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.608 + ], + "angle": 0, + "content": "[22] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. arXiv preprint arXiv:2403.06976, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[23] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.664, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[24] Naoki Kimura and Jun Rekimoto. Extvision: augmentation of visual experiences with generation of context images for a peripheral vision using deep neural network. In Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems, pages 1-10, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[25] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In European Conference on Computer Vision, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[26] Minhyeok Lee, Suhwan Cho, Chajin Shin, Jungho Lee, Sunghun Yang, and Sangyoun Lee. Video diffusion models are strong video inpainter. arXiv preprint arXiv:2408.11402, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Yao-Chih Lee, Yi-Ting Chen, Andrew Wang, Ting-Hsuan Liao, Brandon Y Feng, and Jia-Bin Huang. Vividdream: Generating 3d scene with ambient dynamics. arXiv preprint arXiv:2405.20334, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7317" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[28] Yao-Chih Lee, Zhoutong Zhang, Kevin Blackburn-Matzen, Simon Niklaus, Jianming Zhang, Jia-Bin Huang, and Feng Liu. Fast view synthesis of casual videos with soup-of-planes. In European Conference on Computer Vision, pages 278–296. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.472, + 0.219 + ], + "angle": 0, + "content": "[29] Jiahui Lei, Yijia Weng, Adam Harley, Leonidas Guibas, and Kostas Daniilidis. Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds. arXiv preprint arXiv:2405.17421, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[30] Renjie Li, Panwang Pan, Bangbang Yang, Dejia Xu, Shijie Zhou, Xuanyang Zhang, Zeming Li, Achuta Kadambi, Zhangyang Wang, and Zhiwen Fan. 4k4dgen: Panoramic 4d generation at 4k resolution. arXiv preprint arXiv:2406.13527, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.471, + 0.347 + ], + "angle": 0, + "content": "[31] Zhengqi Li, Richard Tucker, Forrester Cole, Qianqian Wang, Linyi Jin, Vickie Ye, Angjoo Kanazawa, Aleksander Holynski, and Noah Snively. Megasam: Accurate, fast, and robust structure and motion from casual dynamic videos, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.472, + 0.417 + ], + "angle": 0, + "content": "[32] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In Proceedings of the European conference on computer vision (ECCV), pages 85-100, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.471, + 0.488 + ], + "angle": 0, + "content": "[33] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.471, + 0.559 + ], + "angle": 0, + "content": "[34] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.471, + 0.615 + ], + "angle": 0, + "content": "[35] Hao Ouyang, Kathryn Heal, Stephen Lombardi, and Tiancheng Sun. Text2immersion: Generative immersive scene with 3d gaussians. arXiv preprint arXiv:2312.09242, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.617, + 0.471, + 0.673 + ], + "angle": 0, + "content": "[36] F. Perazzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Computer Vision and Pattern Recognition, 2016. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.471, + 0.743 + ], + "angle": 0, + "content": "[37] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. UniDepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.471, + 0.814 + ], + "angle": 0, + "content": "[38] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[39] Weize Quan, Jiaxi Chen, Yanli Liu, Dong-Ming Yan, and Peter Wonka. Deep learning-based image and video inpainting: A survey. International Journal of Computer Vision, 132(7): 2367-2400, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gener" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.192 + ], + "angle": 0, + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.894, + 0.235 + ], + "angle": 0, + "content": "[42] Iizuka Satoshi, Simo-Serra Edgar, and Ishikawa Hiroshi. Globally and locally consistent image completion. ACM Trans. Graph, 36(4):3073659, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.894, + 0.291 + ], + "angle": 0, + "content": "[43] Jaidev Shriram, Alex Trevithick, Lingjie Liu, and Ravi Ramamoorthi. Realmdreamer: Text-driven 3d scene generation with inpainting and depth diffusion. arXiv preprint arXiv:2404.07199, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.293, + 0.894, + 0.389 + ], + "angle": 0, + "content": "[44] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Armenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2149–2159, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[45] Luming Tang, Nataniel Ruiz, Qinghao Chu, Yuanzhen Li, Aleksander Holynski, David E Jacobs, Bharath Hariharan, Yael Pritch, Neal Wadhwa, Kfir Aberman, et al. Realfill: Reference-driven generation for authentic image completion. ACM Transactions on Graphics (TOG), 43(4):1-12, 2024. 2, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[46] Laura Turban, Fabrice Urban, and Philippe Guillotel. Extrafoveal video extension for an immersive viewing experience. IEEE transactions on visualization and computer graphics, 23(5):1520-1533, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.589 + ], + "angle": 0, + "content": "[47] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.59, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[48] Fu-Yun Wang, Xiaoshi Wu, Zhaoyang Huang, Xiaoyu Shi, Dazhong Shen, Guanglu Song, Yu Liu, and Hongsheng Li. Be-your-outpainter: Mastering video outpainting through input-specific adaptation. arXiv preprint arXiv:2403.13745, 2024. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[49] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[50] Shizun Wang, Xingyi Yang, Qiuhong Shen, Zhenxiang Jiang, and Xinchao Wang. Gflow: Recovering 4d world from monocular video. arXiv preprint arXiv:2405.18426, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[51] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas S. Huang. Youtube-vos: A large-scale video object segmentation benchmark. CoRR, abs/1809.03327, 2018. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[52] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18381-18391, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7318" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[53] Zhenqiang Ying and Alan Bovik. 180-degree outpainting from a single image. arXiv preprint arXiv:2001.04568, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[54] Hong-Xing Yu, Haoyi Duan, Charles Herrmann, William T Freeman, and Jiajun Wu. Wonderworld: Interactive 3d scene generation from a single image. arXiv preprint arXiv:2406.09394, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.471, + 0.275 + ], + "angle": 0, + "content": "[55] Hong-Xing Yu, Haoyi Duan, Junhwa Hur, Kyle Sargent, Michael Rubinstein, William T Freeman, Forrester Cole, Deqing Sun, Noah Snavely, Jiajun Wu, et al. Wonderjourney: Going from anywhere to everywhere. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6658-6667, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.47, + 0.344 + ], + "angle": 0, + "content": "[56] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contextual attention. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5505-5514, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.347, + 0.469, + 0.403 + ], + "angle": 0, + "content": "[57] Yanhong Zeng, Jianlong Fu, Hongyang Chao, and Baining Guo. Aggregated contextual transformations for high-resolution image inpainting. IEEE Transactions on Visualization and Computer Graphics, 29(7):3266-3280, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.404, + 0.469, + 0.459 + ], + "angle": 0, + "content": "[58] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.46, + 0.47, + 0.529 + ], + "angle": 0, + "content": "[59] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.531, + 0.469, + 0.586 + ], + "angle": 0, + "content": "[60] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T Freeman. Structure and motion from casual videos. In European Conference on Computer Vision, pages 20-37. Springer, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.588, + 0.469, + 0.669 + ], + "angle": 0, + "content": "[61] Zhixing Zhang, Bichen Wu, Xiaoyan Wang, Yaqiao Luo, Luxin Zhang, Yinan Zhao, Peter Vajda, Dimitris Metaxas, and Licheng Yu. Avid: Any-length video inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7162-7172, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.672, + 0.469, + 0.755 + ], + "angle": 0, + "content": "[62] Haitian Zheng, Zhe Lin, Jingwan Lu, Scott Cohen, Eli Shechtman, Connelly Barnes, Jianming Zhang, Ning Xu, Sohrab Amirghodsi, and Jiebo Luo. Image inpainting with cascaded modulation gan and object-aware training. In European Conference on Computer Vision, pages 277-296. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.756, + 0.469, + 0.825 + ], + "angle": 0, + "content": "[63] Shijie Zhou, Zhiwen Fan, Dejia Xu, Haoran Chang, Pradyumna Chari, Suya Bharadwaj, Tejas You, Zhangyang Wang, and Achuta Kadambi. Dreamscene360: Unconstrained text-to-3d scene generation with panoramic gaussian splatting. arXiv preprint arXiv:2404.06903, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.827, + 0.469, + 0.868 + ], + "angle": 0, + "content": "[64] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. arXiv preprint arXiv:2312.00451, 2023. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.868 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7319" + } + ] +] \ No newline at end of file diff --git a/2025/Unboxed_ Geometrically and Temporally Consistent Video Outpainting/3a973275-71b1-47b0-9c32-b162722059fe_origin.pdf b/2025/Unboxed_ Geometrically and Temporally Consistent Video Outpainting/3a973275-71b1-47b0-9c32-b162722059fe_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1a02f681a7b89cf721290e0fdd4ca53aa13b3442 --- /dev/null +++ b/2025/Unboxed_ Geometrically and Temporally Consistent Video Outpainting/3a973275-71b1-47b0-9c32-b162722059fe_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f1fa76f83017daf50561949a0cb22d8a3ee720e5f57f56ffa6fbb3d25d3d9b9e +size 3255405 diff --git a/2025/Unboxed_ Geometrically and Temporally Consistent Video Outpainting/full.md b/2025/Unboxed_ Geometrically and Temporally Consistent Video Outpainting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3ab520ec2890c9ac00fe4cd2bc513fca2feb9d51 --- /dev/null +++ b/2025/Unboxed_ Geometrically and Temporally Consistent Video Outpainting/full.md @@ -0,0 +1,335 @@ +# Unboxed: Geometrically and Temporally Consistent Video Outpainting + +Zhongrui $\mathrm{Yu}^1$ + +Martina Megaro-Boldini² + +$^{1}$ ETH Zürich + +zhonyu@ethz.ch + +Robert W. Sumner1,2 + +Abdelaziz Djelouah2 + +2DisneyResearch|Studios + +adelaziz.djelouah@disney.com + +![](images/e95dfc66ca20e059d3146175c4d41d9edd181e90cfed1507dd72e46babf01c15.jpg) +Figure 1. Video outpainting results on two different resolutions. In both case, the area outside the red box is outpainted. Our method achieves results of better quality and more temporally stable, compared to recent works such as Dehan [13], M3DDM [14] and MOTIA [48]. + +![](images/294ffb16ba03e9e73345713154b66df8e1b93d6c74c910b8f745bae7733d51e0.jpg) + +# Abstract + +Extending the field of view of video content beyond its original version has many applications: immersive viewing experience with VR devices, reformatting 4:3 legacy content to today's viewing conditions with wide screens, or simply extending vertically captured phone videos. Many existing works focus on synthesizing the video using generative models only. Despite promising results, this strategy seems at the moment limited in terms of quality. In this work, we address this problem using two key ideas: 3D supported outpainting for the static regions of the images, and leveraging pre-trained video diffusion model to ensure realistic and temporally coherent results, particularly for the dynamic parts. In the first stage, we iterate between image outpainting and updating the 3D scene representation - we use 3D Gaussian Splatting. Then we consider dynamic objects independently per frame and inpaint missing pixels. Finally, we propose a denoising scheme that allows to maintain known reliable regions and update the dynamic parts to obtain temporally realistic results. We achieve state-of-the-art video outpainting. This is validated quantitatively and through a user study. We are also able to extend the field of view largely beyond the limits reached by existing methods. + +# 1. Introduction + +Expanding video frames beyond their original field of view (FoV), known as Video Outpainting, is a technique with broad applications in modern media. One example is 4:3 legacy content re-targeting where video outpainting can be used to fill modern widescreen devices seamlessly while avoiding any unnatural stretching. A more forward-looking application is movie viewing with virtual reality (VR) headsets. Video outpainting can enhance the immersion by embedding the viewer in the scene, eliminating distractions from unrelated backgrounds. However, these applications require extremely high-quality results, which is particularly challenging since the synthesized images must integrate harmoniously with the original content and maintain strong temporal coherence across frames. + +Although a wide variety of methods have been proposed, using a patch-based approach [1, 2, 12, 46], optical flow [13] or Generative Adversarial Networks (GAN) [24, 53], most recent methods have clearly identified Diffusion Models [14, 48] as the most promising direction. M3DDM [14] trains a 3D diffusion model on large-scale data for video outpainting with a mask modeling approach. + +In contrast, MOTIA [48] proposes to fine-tune a pre-trained video diffusion model [19] on each test video to capture data-specific motion and content patterns. However, both of these methods have limitations. First, the temporal consistency of the outpainted regions is still unsatisfactory. We can observe duplications, distortions or objects appearing and disappearing. Second, relying entirely on video diffusion models has a strong impact in terms of memory and limits the methods to low resolution and small field of view extension. Some of these issues can be observed in Figure 1. Both M3DDM [14] and MOTIA [48] have strong visual artifacts and only the latter could be used to outpaint the video to a larger field of view. + +To address these issues we propose two core ideas: 3D-supported outpainting for static regions and conditional video synthesis with a pre-trained video diffusion model to ensure realistic, temporally coherent results for dynamic areas. Following these key ideas, we decompose the video into static and dynamic regions and break down the video outpainting task into three stages. First, for static regions, we rely on 3D Gaussians [23] as a supporting 3D representation, and we alternate between outpainting key frames and updating the 3D representation. Second, for dynamic objects, we implement an object-wise inpainting. Here, among other things, we leverage object-tracking masks in each frame to define inpainting regions and use a fine-tuned diffusion model to inpaint. Lastly, using the resulting frames as a starting point, we synthesize temporally consistent results for all frames using a pre-trained video diffusion model [7]. + +Figure 1 demonstrates the effectiveness of our method. We achieve better results than prior methods both in terms of video quality and temporal consistency. This is confirmed by both quantitative and qualitative experiments. Furthermore, by using GPU-efficient Gaussian Splatting for static region outpainting and avoiding fine-tuning of the video diffusion model, we can go from an input $480 \times 480$ video ( $\mathrm{FoV} = 31^{\circ}$ ) to an output resolution of $2560 \times 720$ ( $\mathrm{FoV} = 120^{\circ}$ ) with a GPU memory usage of less than 16GB for the entire pipeline. + +In summary, our contributions are as follows: + +- a novel video outpainting method that combines a 3D representation with a pre-trained video diffusion model; +- an optimization strategy with an additional loss to enable seamless integration of the outpainting with the original static content; +- demonstrating that a pre-trained video diffusion model can function as a temporal filter, making discontinuous frames temporally coherent and simplifying the inpainting task for the dynamic parts. +- State-of-the-art results for video outpainting + +# 2. Related Work + +Immersive Scene Generation. To create immersive experiences for users in VR applications, several methods are proposed to generate realistic 3D or 4D scenes that allow for arbitrary viewpoint exploration. LucidDreamer [10] and the following works [16, 35, 43, 54, 55] lift a single image to 3D space, project it from different viewpoints, and inpaint them using diffusion models to create a 3D scene. However, directly applying their method to consecutive video frames would cause severe geometric distortion between frames due to the simple depth alignment and optimization techniques. Dreamscene360 [63] and 4D4KGen [30] create panoramic 3D and 4D scenes separately from a text-generated panoramic image. VividDream [27] generates multi-view videos from a single image and reconstructs a 4D scene with 4D Gaussians. These methods are primarily designed for static or simply animated images and cannot handle videos with complex camera and object motion. Methods such as [9, 28, 29, 49, 50] reconstruct 4D scenes from casual videos, but they are unable to complete regions beyond what is captured in the video. + +Image completion. Image completion [6, 39], including image inpainting and outpainting, is a task that restores missing part of an image, and it serves as a basis for video outpainting. Traditional methods [6, 18, 21] use low-level features from the incomplete image to propagate information in known regions to fill missing areas. Early deep learning based methods [32, 42, 44] introduced end-to-end frameworks that learned mappings from corrupted to complete images. Recent methods have increasingly leveraged generative models, including GAN [17, 56, 57, 62], and Diffusion Models [4, 11, 22, 34, 40, 41]. Reference-based image completion methods utilize reference images as generation conditions [52] or fine-tune models on a set of reference images [45] to produce completions that are more faithful to the references. However, even fine-tuned on all input video frames, frame-by-frame completion introduces substantial temporal inconsistency in video outpainting. + +Video Outpainting. In the field of video completion, video outpainting is challenging but less explored. Early patch-based methods [1, 2] select the nearest patch from the spatio-temporal domain of the video to expand each frame. Lee et al. [26] employ 3D information to warp neighboring frames to the current frame. Dehan et al. [13] utilize estimated optical flow to warp the background from adjacent frames and use an image completion network to fill any remaining uncompleted areas. While Zhang et al. [61] target video in-painting motion module and structure guidance. M3DDM [14] extends the image diffusion model framework to 3D and trains it on extensive video data for the outpainting task. A global frame is introduced as a conditioning input to improve global information perception. MOTIA [48] proposes a method that fine-tunes a pre-trained + +![](images/b5c993a86ea967cea348375fe72fd26011ea7a2a7d2c501568417d548a47784f.jpg) +Input + +![](images/09a54f17834f1d8eec1f0d39bfb2bcacdbc37c3dfe82d38740419aec20612394.jpg) +Pre-process + +![](images/3405ae143188f7bed6ed6f3a352b9f5ef03dc44aad903a2064c2aa01b39507c0.jpg) +Dynamic Object Masks + +![](images/41fbf127238cb8c67fb260b2b1d9232e72dcb00a37224087e6f655a114ffdacd.jpg) + +![](images/72cf0259f7d1001528b546c82cbcfb4b45dd3d8c71acc92cd9114d67e825d918.jpg) +Static Scene Reconstruction + +![](images/ab7e21691daaa99ca318ffecf57409926e41322488202fc41bf425c28a2ed81a.jpg) +(1) Static Region + +![](images/6a1f9e6b97856706050a05db1c0930f537c9e6ec14c573710aed566bb8e78b87.jpg) +(2) Dynamic Objects +Figure 2. Overview of the proposed method. Starting from the input video sequence, dynamic objects are separated from the static parts of the image, and a static scene reconstruction is made using Gaussian Splatting (GS). In the first stage of our method, the static part is outpainted by iterating between image outpainting, depth prediction and updating the GS model. After this stage, we identify parts of the dynamic object that may require inpainting. Here, an image diffusion model is finetuned on the input frames and used for inpainting each frame independently, which may result in temporal inconsistencies as seen of the car windows. In the last step, a pre-trained video diffusion model is used to fix temporal inconsistencies. Details about each step are provided in the text. + +![](images/6e319d107356edd0ba8fc84e351a4f791deec53c22212860a34e9453f404ba15.jpg) +(3) Final Video Synthesis +Temporally Consistent Results + +video diffusion model on test samples to learn the intrinsic patterns of a video. In general, diffusion-based approaches rely on priors learned by temporal modules to maintain continuity in the outpainted regions, which is not sufficient. We demonstrate that a pipeline integrating constraints from a 3D representation achieves results of much higher quality. + +# 3. Method + +Given a sequence of video frames $\{I_i\in \mathbb{R}^{h\times w\times c}\}_{i = 1}^N$ , our goal is to expand each frame beyond its original field of view to $\{\tilde{I}_i\in \mathbb{R}^{H\times W\times c}\}_{i = 1}^N$ , while maintaining both intra-frame and cross-frame consistency. Although the problem is more challenging than image outpainting, due to the temporal consistency requirements, the presence of multiple frames offers additional observations that can be leveraged. More specifically, it is possible to build a 3D representation of the scene using 3D Gaussians. + +As illustrated in Figure 2, our proposed method can be divided into 3 distinct stages. In the first stage, we focus on the static parts of the scene and alternate between image outpainting and updating the 3D scene representation. In the second stage, we focus on the dynamic objects in the scene. Here, we fine-tune an image diffusion model on the available images and use it to inpaint the missing parts of the object in each frame. At this point, all the frames are outpainted, but temporal inconsistencies remain. These frames serve as a starting point for a video synthesis scheme leveraging pre-trained stable video diffusion to synthesize a temporally consistent video sequence. Next, we detail each + +of these main steps. + +# 3.1. Pre-processing and Static Scene Outpainting + +From the static parts of the video, it is possible to estimate a 3D reconstruction of the scene and use the different views across time to provide information about the regions to outpaint. In the first step, we identify the dynamic parts of the scene. A wide variety of methods can be employed here, but we simply rely on object segmentation and, in particular, object tracking models [33]. Figure 2 shows an example of the binary segmentation masks. These masks are expanded to avoid including incorrectly labeled pixels into the background. We also compute a 2D bounding box around the object to help keep track of the object size in case it partially or completely leaves the field of view. The static parts are all the pixels that are not part of any dynamic object. + +Static Scene Reconstruction. We propose to use Gaussian Splatting (GS) [23] to reconstruct the static part of the video. This approach allows regions visible in other frames to be rendered into the current frame. For the input video frames $\{I_i\}_{i=1}^N$ , we first estimate the corresponding camera poses $\{\mathbf{p}_i\}_{i=1}^N$ and intrinsic parameters $\{\mathbf{K}_i\}_{i=1}^N$ using an off-the-shelf camera pose estimator [60]. We compute a depth map for each frame using parallax and project the background pixels into 3D space, yielding an initial pointcloud for GS training. The training process is supervised by the image reconstruction loss and the depth loss. The image reconstruction loss is the combination of the L1 loss and + +![](images/89a119bdb865156ce67a37a4ea8c2d23e43ad9c77436f0065b7782e27badc70e.jpg) +First key-frame $(k = 1)$ +$I_{k = 1}^{\mathrm{static}}$ + +![](images/72dc690b512c7129f381d8742b4619f55f25810f9ead2305a88e0b6b336ede05.jpg) +$I_{k = 1}^{\mathrm{static}}$ + +![](images/af585d567554915f32c7b929fd1f75dfa311d581fa21d8d13643f4b6bd019e48.jpg) +$L_{\mathrm{recon}}$ + +![](images/37bc47ebb604cf9485529c50ea63149efe391394487854107dcb2f2bbb709de5.jpg) +Depth after GS finetuning +$L_{\mathrm{recon}} + L_{\mathrm{depth}}$ + +![](images/a6a9c7eb4e65893f26ad32169c9f36d5d7327a6bd1ba64c71bc5834ed31c2756.jpg) +$L_{\mathrm{recon}} + L_{\mathrm{depth}} + L_{\mathrm{depth}}^{\mathrm{Sp}}$ + +![](images/2d843ea0a1c70843f385b3d8bba6b6d0eb60188295e4810ac20f9ff8c35f72b8.jpg) +Next key-frame $(k = 2)$ +$I_{k = 2}^{\mathrm{static}}$ +Figure 3. Effect of the loss terms on the GS fine-tuning. Starting from the rasterization of the 3D Gaussians for the first key-frame $k = 1$ , we use image outpainting to obtain $\hat{I}_k^{\mathrm{static}}$ . This image is subdivided into super-pixels as illustrated here. During the optimization to update the GS model, we use a combination of losses. We can see that all are needed to avoid discontinuities and obtain a realistic smooth depth map. After this update, the process is repeated with the next key-frame using the updated 3D Gaussians. + +the structural similarity loss (SSIM). The depth loss acts as a regularization and uses Pearson Correlation [64] between the rendered depth and the depth estimated from a monocular depth estimation model [37]. Apart from the dynamic object masking, this is similar to the optimization proposed by Zhu et al. [64]. + +Single Image Outpainting. Starting from the GS model $GS_{\mathrm{static}}$ representing the static parts of the scene, we uniformly sample $k \in \{1, \dots, K\}$ key frames. Our objective is to outpaint them one by one, while we progressively update the GS model. Here we first detail outpainting for a single image. Considering frame $I_k^{\mathrm{static}}$ , we adjust its field of view to obtain new intrinsic parameters $\hat{\mathbf{K}}_k$ , and render $GS_{\mathrm{static}}$ based on the new parameters to get the image $\hat{I}_k^{\mathrm{static}}$ . The mask $\hat{\mathbf{m}}_k^{\mathrm{static}}$ for outpainting is defined by checking whether each pixel is covered by the projection of Gaussian points. Any image outpainting method can be used at this level, but we find that the Stable Diffusion XL (SDXL) image outpainting model [38] performs well for our task. This can be summarized as: + +$$ +I _ {k} ^ {\text {s t a t i c}} = \phi_ {G S _ {\text {s t a t i c}}} \left(\hat {\mathbf {K}} _ {k}, \mathbf {p} _ {k}\right) \tag {1} +$$ + +$$ +\hat {I} _ {k} ^ {\text {s t a t i c}} = \phi_ {\mathrm {S D X L}} \left(I _ {k} ^ {\text {s t a t i c}}, \hat {\mathbf {m}} _ {k} ^ {\text {s t a t i c}}\right) \tag {2} +$$ + +Updating the GS model. To ensure the consistency of the outpainting across frames, we update the GS model each time we outpaint a key-frame. Denote by $K'$ the number of key-frames that have been outpainted so far. First, we estimate the depth for the newly outpainted image using a single image depth model [37]. Then we initialize new Gaussian points for the outpainted pixel and fine-tune the GS model with this additional data. We use an image reconstruction loss + +$$ +\begin{array}{l} L _ {\text {r e c o n}} = \sum_ {k = 1} ^ {K ^ {\prime}} \lambda_ {1} L _ {1} \left(\hat {I} _ {k} ^ {\text {s t a t i c}}, \phi_ {\mathrm {G S}} \left(\hat {K} _ {k}, p _ {k}\right)\right) \tag {3} \\ + \lambda_ {2} L _ {\mathrm {S S I M}} (\hat {I} _ {k} ^ {\text {s t a t i c}}, \phi_ {\mathrm {G S}} (\hat {K} _ {k}, p _ {k})) \\ \end{array} +$$ + +between the rasterization from the model $\phi_{\mathrm{GS}}(\hat{K}_k,p_k)$ and the outpainting result. We also use 2 depth based losses, the + +first one on the full depth map: + +$$ +L _ {\text {d e p t h}} = \sum_ {k = 1} ^ {K ^ {\prime}} \frac {\operatorname {C o v} \left(D \left(\hat {I} _ {k} ^ {\text {s t a t i c}}\right) , D _ {\mathrm {G S}} \left(\hat {K} _ {k} , p _ {k}\right)\right)}{\sqrt {\operatorname {V a r} \left(D \left(\hat {I} _ {k} ^ {\text {s t a t i c}}\right)\right) \operatorname {V a r} \left(D _ {\mathrm {G S}} \left(\hat {K} _ {k} , p _ {k}\right)\right)}} \tag {4} +$$ + +where $D(\hat{I}_k^{\mathrm{static}})$ is the depth predicted by the single image depth model [37], and $D_{\mathrm{GS}}(\hat{K}_k,p_k)$ is the depth from GS model that we optimize. For every image, we extract $S$ super-pixels, and use the second depth loss at this level: + +$$ +L _ {\text {d e p t h}} ^ {\mathrm {S p}} = \sum_ {k = 1} ^ {K ^ {\prime}} \sum_ {s = 1} ^ {S} \frac {\operatorname {C o v} _ {s} \left(D \left(\hat {I} _ {k} ^ {\text {s t a t i c}}\right) , D _ {\mathrm {G S}} \left(\hat {K} _ {k} , p _ {k}\right)\right)}{\sqrt {\operatorname {V a r} _ {s} \left(D \left(\hat {I} _ {k} ^ {\text {s t a t i c}}\right)\right) \operatorname {V a r} _ {s} \left(D _ {\mathrm {G S}} \left(\hat {K} _ {k} , p _ {k}\right)\right)}} \tag {5} +$$ + +Figure 3 demonstrate this process and shows the importance of the different loss terms for the GS update. Starting from the rasterization of the initial GS model, we can see that the 3D model does not cover the new field of view. The frame is outpainted and superpixels are illustrated in yellow. Using only the reconstruction loss $L_{\mathrm{recon}}$ leads to strong discontinuities between the original GS model and the new outpainted areas. Adding the depth loss $L_{\mathrm{depth}}$ helps, but the bottom part of the image still has important discontinuities that would be revealed when projecting to other frames (i.e. other views). Using all the proposed loss terms provides the best results, and the process can now be iterated by adding new views. + +# 3.2. Initial Inpainting for the Dynamic Object(s) + +After the static scene outpainting, we can render any frame with $\phi_{\mathrm{GS}}(\hat{K}_i,p_i)$ . As illustrated in Figure 2, we miss part of the object when it is partially outside the filed of view. We need to inpaint this missing part. We use the 2D tracking bounding box to identify the areas that require inpainting for each frame of the video. + +For this inpainting problem we use a similar strategy as ReaFill [45]. An image diffusion model is fine-tuned for the inpainting task on the input frames (in our case SDXL). It is then applied independently on each frame. Thanks to the fine-tuning, the resulting inpainting is very similar across frames. However temporal inconsistencies remain as it is + +![](images/cd18af78ee96da840a66cb6044ef3e42c6573947b61da9d570a129dd0e21c650.jpg) +Figure 4. Details of the video latent denoising strategy. Starting from the independently outpainted frames $\tilde{I}_{i\in 1,\dots,N}$ , we obtained the corresponding latents in the forward process $(\tilde{x}_{i\in 1,\dots,N})$ . In the denoising process we blend known latent with denoising latents according to the mask $\mathbf{m}$ . The mask and the time step $t_{\mathrm{end}}$ respectively control which region of the frames get updated by the video diffusion model and how large the changes are. + +clearly visible on the inpainting results in Figure 2. Next we describe how a Video Diffusion Model [7, 8] is used to fix all these temporal artifacts. + +# 3.3. Guided Video Synthesis + +If we consider the original input sequences of frames $I_{i\in 1,\dots,N}$ , after the previous stages we obtain fully out-painted frames $\tilde{I}_{i\in 1,\dots,N}$ . Static regions are consistent thanks to the supporting GS model, while the dynamic objects are independently inpainted for each frame. Starting from this data, our objective is to synthesize a realistic sequence of frames $I_{i\in 1,\dots,N}^{*}$ . Not only the temporal inconsistencies on the dynamic objects need to be fixed, but the entire outpainted regions need to appear realistic. Indeed, real-world videos often contain subtle background movements, such as leaves swaying in the wind or water rippling. + +We leverage pre-trained video diffusion models for this video synthesis. The idea is that video diffusion models are trained for generating real video sequences, with consistent content and realistic motion. We propose using the previously generate frames in a denoising scheme that modifies the different frame areas based on the expected inconsistency level. On the dynamic objects, we expect more changes (as the inpainting is independent), while less change should be allowed on the rest of the image as it is supported by the 3D consistent GS model. + +More specifically, if we consider a latent video diffusion model, a sequence of input frames $I_{i}$ is encoded into the corresponding sequence of latent values $x_{i}$ . The forward process can be applied to the sequence of frames $\tilde{I}_{i\in 1,\dots,N}$ to obtain known latents + +$$ +\tilde {x} _ {t - 1} \sim \mathcal {N} \left(\sqrt {\bar {\alpha} _ {t}} \tilde {x} _ {0}, \sqrt {1 - \bar {\alpha} _ {t}} \mathbf {1}\right). \tag {6} +$$ + +For simplicity, we dropped the frame index $i$ . + +We use the reverse process to synthesize the final video sequence. At each time step, we denoise previous latents using the latent video diffusion model with parameters $\theta$ , to obtain temporary latents + +$$ +x _ {t - 1} ^ {\operatorname {t m p}} \sim \mathcal {N} \left(\mu_ {\theta} \left(x _ {t} ^ {*}, t\right), \Sigma_ {\theta} \left(x _ {t} ^ {*}, t\right)\right) \tag {7} +$$ + +To obtain the final latents for this time step, we blend in the known latents according to + +$$ +x _ {t - 1} ^ {*} = (1 - \mathbf {m}) \odot \tilde {x} _ {t - 1} + \mathbf {m} \odot x _ {t - 1} ^ {\mathrm {t m p}} \tag {8} +$$ + +where the binary mask $\mathbf{m}$ is defined as + +$$ +\mathbf {m} = \left\{ \begin{array}{l l} \mathbf {m} ^ {\text {d y n a m i c}} & \text {i f} t \geq t _ {\text {e n d}} \\ \mathbf {m} ^ {\text {d y n a m i c}} + \mathbf {m} ^ {\text {s t a t i c}} & \text {i f} t < t _ {\text {e n d}} \end{array} \right. \tag {9} +$$ + +The mask, $\mathbf{m}^{\mathrm{dynamic}}$ , corresponds to the dynamic objects in the scene, and the mask $\mathbf{m}^{\mathrm{static}}$ corresponds to the static parts in the outpainted regions. This scheme is similar to ideas already explored in the context of image diffusion models [3, 34], but applied to video diffusion models. If we observe the changes in Figure 2, we can see that the pre-trained video diffusion model can function as a temporal filter, making discontinuous frames temporally coherent. This simplifies the inpainting task since we do not rely on expensive training of the video model and simply use it at inference time. + +# 4. Experiments + +# 4.1. Implementation details + +Static Scene. In pre-processing, we estimate the camera poses of the source video [31, 60]. Our GS training process consists of $1 + K + 1$ rounds in total, where $K$ is the number of key frames we sample from the video. In the first training round, we train the GS model only on the static part of the video for 3000 iterations. Then we iterate between outpainting the $K$ key frames and $K$ rounds of GS training to update the outpainted regions into the GS model. To avoid overfitting, we employ a sliding window training strategy: following outpainting of the $i$ -th frame, we train only on frames within the window $[i - 5, i + 5]$ for 500 iterations. This window shifts progressively as outpainting and training continues. After outpainting all $K$ key frames, we conduct a final round of training for refinement on all out-painted and original background frames for 1000 iterations. All training and outpainting can be completed on a single 4090 GPU. + +Initial Object Inpainting. Before initial inpainting on the missing regions of dynamic objects, we fine-tune the SDXL model on the masked source video. The approach is similar to the one proposed by Tang et al. [45]. Masks are randomly + +![](images/08f698742f583338b3b845b1861c522ccbc43d7e250713da970ba9a84f8fdb37.jpg) +Figure 5. Qualitative Comparison with State-of-the-art methods on the DAVIS dataset [36]. The area outside of the red lines (only shown for the first frame) is outpainted. We show the results for 3 frames from the each video sequence. Compared to MOTIA [48] and M3DDM[14], our results are of better quality and temporally stable. Please pay attention to the changes of the outpainted area between the frames of the same sequence. + +generated for each input frame, with random distributed positions and mask ratios varying from 0.2 to 0.8. For each object, a prompt "a photo of [v]" is applied. During fine-tuning, a Low-Rank Adaption (LoRA) [20] with rank 8 is trained for both the text-encoder and UNet with a batch size of 4 and gradient accumulation steps of 4 for 2000 steps. + +Video Diffusion Model. We use a pretrained image-to-video model [7] to complete object motion. The number of denoising steps is set to 25, with the level of noise added to the reconstructed frames generally set to 0.8 in our experiments. We apply a resampling strategy similar to Repaint [34] five times after each denoising step to improve the quality and stability of the objects. For videos where the object fully appears in the middle or last frame, we em + +ploy the Time-Reversal strategy from [15], conditioning the generation on additional frames to better retain the object identity. + +# 4.2. Experiment Setup + +Competitors. We compare our method with the following methods. (1) Dehan et al. [13] separate foreground and background components. They use estimated optical flow to warp the background, with missing regions filled in using an image completion network. (2) M3DDM [14] proposes a masked 3D Diffusion Model architecture, which is trained on WebVid [5] and a self-collected 5M e-commerce dataset for video outpaint tasks. (3) MOTIA [48] is the current state-of-the-art method for video outpainting. It is based on the ControlNet [58] inpainting model and the temporal module fromAnimateDiff [19]. It trains a LoRA [20] for + +each test sample to learn the intrinsic patterns. + +Datasets. Following M3DDM [14] and MOTIA [48], we evaluate our method on the DAVIS [36] and YouTube-VOS [51] datasets. For the YouTube-VOS [51] dataset, due to its large number of videos and the lack of clarity in prior works on which videos were tested, we selected a high-quality subset and evaluated all methods on it. Consistent with prior works, we compare the horizontal outpainting results for each test video at mask ratios of 0.25 and 0.66. + +Evaluation Metrics. The quality of outpainting is primarily evaluated in terms of visual quality, realism, and temporal consistency. For this purpose, we employ two well-established metrics: Fréchet Video Distance (FVD) [47] and flow warping error $(E_{\mathrm{warp}})$ [25]. FVD assesses the similarity between the distribution of the outpainted video and the original video, while flow warping error evaluates temporal consistency by measuring the error when outpainted frames are warped to adjacent frames using estimated optical flow. In addition, to align with previous works, we also report Peak Signal to Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) [59] between the outpainted and ground-truth videos, although we consider these metrics limited in capturing the quality of outpainting. We further conducted a user study to gather subjective evaluations of outpainting quality, which we detail in Sec. 4.3. + +# 4.3. Evaluation & Results + +Qualitative Results. Fig. 5 and Fig. 6 illustrate the qualitative comparison of our method against other approaches on the DAVIS and YouTube-VOS datasets. M3DDM [14] frequently suffers from generation failures and produces blurred content, as shown in the first samples of Fig. 5 and Fig. 6. MOTIA [48], while fine-tuning on each test sample to capture input-specific patterns, often generates repetitive objects and background patterns. More importantly, both methods struggle to accurately capture contextual information across frames, resulting in inconsistency when outpainting elements seen in adjacent frames. In contrast, our method overcomes these issues, achieving better cross-frame consistency. Furthermore, our approach's object-specific completion approach prevents objects from duplicating or disappearing beyond the frame. + +Generalization and multiple objects. Our strategy, differentiating between static and dynamic elements, can naturally extend to multiple dynamic objects by decomposing the frame into object layers based on their depth. Each object is processed separately using our pipeline and then seamlessly blended to produce the final result. Figure 7 shows that our method still performs best on such complex scenarios. + +![](images/4f7813febd35ba3dc97b6dd27df7e655f65b2258e25554056c6402fa69fe10d3.jpg) +Figure 6. Qualitative Comparison with State-of-the-art methods on Youtube-VOS dataset [51]. The area outside of the red lines (only shown for the first frame) is outpainted. We show the results for 3 frames from the each video sequence. Compared to Dehan [13], MOTIA [48] and M3DDM[14], our results are of better quality and temporally stable. Please pay attention to the changes of the outpainted area between the frames of the same sequence. + +![](images/2e853b0c067821d1f93277b2e0bcd86eebf05e0d9e85f313ae569af3879a75eb.jpg) +Figure 7. Qualitative Comparison with State-of-the-art methods on challenging scenarios (multiple dynamic objects, complex static regions, etc.). + +Quantitative Results. Tab. 1 presents a quantitative comparison of our method with other approaches. Our method outperforms the current state-of-the-art in both FVD and $E_{\text{warp}}$ , achieving improvements of $18.8\%$ and $11.7\%$ on the DAVIS dataset and $26.9\%$ and $33.3\%$ on the YouTube-VOS dataset, respectively. We also achieve the highest PSNR and SSIM scores on both datasets. For LPIPS, while our method is the second behind the number reported in MOTIA [48] on the DAVIS dataset, it surpasses the score obtained from using their public implementation and provided model. + +Ablation Study We conducted ablation studies for the different components of our method. Quantitative evalua + +
DAVISYouTube-VOS
PSNR↑SSIM↑LPIPS↓FVD↓E_warp↓PSNR↑SSIM↑LPIPS↓FVD↓E_warp↓
Dehan [13]17.960.6270.233363.10.03018.340.7610.287353.60.043
M3DDM [14]20.260.7080.204300.00.01718.330.7250.424283.10.024
MOTIA (model) [48]18.250.6990.282294.80.02118.540.7900.250156.40.027
MOTIA (paper) [48]20.360.7580.160286.3------
Ours22.620.8060.200232.50.01519.880.8320.236114.20.016
+ +Table 1. Quantitative Comparison with State-of-the-art methods on DAVIS dataset [36] and YouTube-VOS [51]. Our method achieves the best performance overall. In the case of MOTIA [48], we have 2 rows as we report the evaluation using the model and code available online(model) and the numbers from the original paper (paper). See the text for details. + +
PSNR↑SSIM↑LPIPS↓FVD↓E_warp↓
(a)18.150.6940.248448.80.032
(b)18.050.7160.274323.20.022
(c)20.680.7720.218293.30.020
(d)20.050.7760.223319.80.017
(e)22.620.8060.200232.50.015
+ +Table 2. Quantitative ablation results. (a). Outpainting with fine-tuned Image Model; (b). (a) + Guided Video Synthesis; (c). Our model without guided video synthesis; (d). Our model without the initialization from object inpainting; (e). Our complete model. + +tion is shown in Table 2, while the visual results are provided in supplementary material. When only using single image outpainting (SDXL fine-tuned on content [45]), the images achieve good realism but lack temporal continuity. When applying video diffusion guided synthesis to these outpainted frames directly, the temporal consistency within the outpainted areas improves individually. However, coherence between the outpainted regions and the original video content remains weak. This issue is difficult to resolve solely with Video Diffusion Model priors. When using the static region outpainting (assisted with GS) and either the image inpainting or video diffusion alone are applied for object completion, the objects suffer from significant temporal variations in appearance. Using all the components achieves the best results. + +User Study. Given the subjective nature of evaluating outpainting quality, we conducted a user study to compare our method with competing approaches. Users were asked to evaluate each method across several dimensions, including realism (such as whether the outpainted results look natural and harmonious), temporal consistency (including background consistency, object motion consistency, etc.), and overall visual quality (including color fidelity, smoothness of boundaries, blurriness etc.). The results, shown in Fig. 8, indicate that our method was clearly preferred by users compared to others, achieving more than $80\%$ of the votes in all dimensions. Further details of the user study can + +![](images/41c904528a4bc15d586ad920aa1283aa241aca5003efce01f59358301a495c4a.jpg) +Figure 8. The result of the user study includes 619 votes from 37 participants. Our method is preferred by more than $80\%$ of the votes in all dimensions. + +![](images/6b4dda42f6c045166c29ae37338ef75bf1fbcc304b49f64a015ea09bfa96a043.jpg) +Figure 9. Challenges with dynamic objects moving in and out of the original frame: Our current solution of tracking and extrapolating the trajectory of the bounding box can be improved. + +be found in the supplementary materials. + +# 5. Discussions + +Building on Gaussian Splatting and a Video Diffusion Model, we propose a new state-of-the-art method for video outpainting. Despite our strong results, some issues remain, especially around objects moving in and out of the original frame. An example is shown in Fig. 9. Currently tracking and extrapolating the trajectory of the bounding box offers a solution, but limitations in the single image inpainting negatively impact the quality of the final result. Another area of improvement is the outpainting involving multiple occluded dynamic objects. Our current solution of inpainting each object separately and blending them based on the estimated depth is limited. Exploring the use of more advanced 4D dynamic Gaussian representations jointly with video models offers a promising future direction towards higher and higher quality video outpainting. + +# References + +[1] Amit Aides, Tamar Avraham, and Yoav Y Schechner. Multiscale ultrawide foveated video extrapolation. In 2011 IEEE International Conference on Computational Photography (ICCP), pages 1-8. IEEE, 2011. 1, 2 +[2] Tamar Avraham and Yoav Y Schechner. Ultrawide foveated video extrapolation. IEEE Journal of Selected Topics in Signal Processing, 5(2):321-334, 2010. 1, 2 +[3] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18208-18218, 2022. 5 +[4] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18208-18218, 2022. 2 +[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 6 +[6] M Bertalmio. Image inpainting, 2000. 2 +[7] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 5, 6 +[8] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 5 +[9] Wen-Hsuan Chu, Lei Ke, and Katerina Fragkiadaki. Dreamscene4d: Dynamic multi-object scene generation from monocular videos. arXiv preprint arXiv:2405.02280, 2024. 2 +[10] Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384, 2023. 2 +[11] Ciprian Corneanu, Raghudeep Gadde, and Aleix M Martínez. Latentpaint: Image inpainting in latent space with diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4334-4343, 2024. 2 +[12] Antonio Criminisi, Patrick Pérez, and Kentaro Toyama. Region filling and object removal by exemplar-based image inpainting. IEEE Transactions on image processing, 13(9): 1200-1212, 2004. 1 +[13] Loic Dehan, Wiebe Van Ranst, Patrick Vandewalle, and Toon Goedemé. Complete and temporally consistent video outpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 687-695, 2022. 1, 2, 6, 7, 8 +[14] Fanda Fan, Chaoxu Guo, Litong Gong, Biao Wang, Tiezheng Ge, Yuning Jiang, Chunjie Luo, and Jianfeng Zhan. Hierar- + +chical masked 3d diffusion model for video outpainting. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7890-7900, 2023. 1, 2, 6, 7, 8 +[15] Haiwen Feng, Zheng Ding, Zhihao Xia, Simon Niklaus, Victoria Abrevaya, Michael J Black, and Xuaner Zhang. Explorative inbetweening of time and space. In European Conference on Computer Vision, pages 378-395. Springer, 2025. 6 +[16] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. Advances in Neural Information Processing Systems, 36, 2024. 2 +[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2 +[18] Qiang Guo, Shanshan Gao, Xiaofeng Zhang, Yilong Yin, and Caiming Zhang. Patch-based image inpainting via two-stage low rank approximation. IEEE transactions on visualization and computer graphics, 24(6):2023-2036, 2017. 2 +[19] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2, 6 +[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6 +[21] Jia-Bin Huang, Sing Bing Kang, Narendra Ahuja, and Johannes Kopf. Image completion using planar structure guidance. ACM Transactions on graphics (TOG), 33(4):1-10, 2014. 2 +[22] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. arXiv preprint arXiv:2403.06976, 2024. 2 +[23] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3 +[24] Naoki Kimura and Jun Rekimoto. Extvision: augmentation of visual experiences with generation of context images for a peripheral vision using deep neural network. In Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems, pages 1-10, 2018. 1 +[25] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In European Conference on Computer Vision, 2018. 7 +[26] Minhyeok Lee, Suhwan Cho, Chajin Shin, Jungho Lee, Sunghun Yang, and Sangyoun Lee. Video diffusion models are strong video inpainter. arXiv preprint arXiv:2408.11402, 2024. 2 +[27] Yao-Chih Lee, Yi-Ting Chen, Andrew Wang, Ting-Hsuan Liao, Brandon Y Feng, and Jia-Bin Huang. Vividdream: Generating 3d scene with ambient dynamics. arXiv preprint arXiv:2405.20334, 2024. 2 + +[28] Yao-Chih Lee, Zhoutong Zhang, Kevin Blackburn-Matzen, Simon Niklaus, Jianming Zhang, Jia-Bin Huang, and Feng Liu. Fast view synthesis of casual videos with soup-of-planes. In European Conference on Computer Vision, pages 278–296. Springer, 2025. 2 +[29] Jiahui Lei, Yijia Weng, Adam Harley, Leonidas Guibas, and Kostas Daniilidis. Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds. arXiv preprint arXiv:2405.17421, 2024. 2 +[30] Renjie Li, Panwang Pan, Bangbang Yang, Dejia Xu, Shijie Zhou, Xuanyang Zhang, Zeming Li, Achuta Kadambi, Zhangyang Wang, and Zhiwen Fan. 4k4dgen: Panoramic 4d generation at 4k resolution. arXiv preprint arXiv:2406.13527, 2024. 2 +[31] Zhengqi Li, Richard Tucker, Forrester Cole, Qianqian Wang, Linyi Jin, Vickie Ye, Angjoo Kanazawa, Aleksander Holynski, and Noah Snively. Megasam: Accurate, fast, and robust structure and motion from casual dynamic videos, 2024. 5 +[32] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In Proceedings of the European conference on computer vision (ECCV), pages 85-100, 2018. 2 +[33] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 3 +[34] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 2, 5, 6 +[35] Hao Ouyang, Kathryn Heal, Stephen Lombardi, and Tiancheng Sun. Text2immersion: Generative immersive scene with 3d gaussians. arXiv preprint arXiv:2312.09242, 2023. 2 +[36] F. Perazzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Computer Vision and Pattern Recognition, 2016. 6, 7, 8 +[37] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. UniDepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 4 +[38] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 4 +[39] Weize Quan, Jiaxi Chen, Yanli Liu, Dong-Ming Yan, and Peter Wonka. Deep learning-based image and video inpainting: A survey. International Journal of Computer Vision, 132(7): 2367-2400, 2024. 2 +[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gener + +ation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2 +[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2 +[42] Iizuka Satoshi, Simo-Serra Edgar, and Ishikawa Hiroshi. Globally and locally consistent image completion. ACM Trans. Graph, 36(4):3073659, 2017. 2 +[43] Jaidev Shriram, Alex Trevithick, Lingjie Liu, and Ravi Ramamoorthi. Realmdreamer: Text-driven 3d scene generation with inpainting and depth diffusion. arXiv preprint arXiv:2404.07199, 2024. 2 +[44] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Armenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2149–2159, 2022. 2 +[45] Luming Tang, Nataniel Ruiz, Qinghao Chu, Yuanzhen Li, Aleksander Holynski, David E Jacobs, Bharath Hariharan, Yael Pritch, Neal Wadhwa, Kfir Aberman, et al. Realfill: Reference-driven generation for authentic image completion. ACM Transactions on Graphics (TOG), 43(4):1-12, 2024. 2, 4, 5, 8 +[46] Laura Turban, Fabrice Urban, and Philippe Guillotel. Extrafoveal video extension for an immersive viewing experience. IEEE transactions on visualization and computer graphics, 23(5):1520-1533, 2016. 1 +[47] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 7 +[48] Fu-Yun Wang, Xiaoshi Wu, Zhaoyang Huang, Xiaoyu Shi, Dazhong Shen, Guanglu Song, Yu Liu, and Hongsheng Li. Be-your-outpainter: Mastering video outpainting through input-specific adaptation. arXiv preprint arXiv:2403.13745, 2024. 1, 2, 6, 7, 8 +[49] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 2 +[50] Shizun Wang, Xingyi Yang, Qiuhong Shen, Zhenxiang Jiang, and Xinchao Wang. Gflow: Recovering 4d world from monocular video. arXiv preprint arXiv:2405.18426, 2024. 2 +[51] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas S. Huang. Youtube-vos: A large-scale video object segmentation benchmark. CoRR, abs/1809.03327, 2018. 7, 8 +[52] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18381-18391, 2023. 2 + +[53] Zhenqiang Ying and Alan Bovik. 180-degree outpainting from a single image. arXiv preprint arXiv:2001.04568, 2020. 1 +[54] Hong-Xing Yu, Haoyi Duan, Charles Herrmann, William T Freeman, and Jiajun Wu. Wonderworld: Interactive 3d scene generation from a single image. arXiv preprint arXiv:2406.09394, 2024. 2 +[55] Hong-Xing Yu, Haoyi Duan, Junhwa Hur, Kyle Sargent, Michael Rubinstein, William T Freeman, Forrester Cole, Deqing Sun, Noah Snavely, Jiajun Wu, et al. Wonderjourney: Going from anywhere to everywhere. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6658-6667, 2024. 2 +[56] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contextual attention. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5505-5514, 2018. 2 +[57] Yanhong Zeng, Jianlong Fu, Hongyang Chao, and Baining Guo. Aggregated contextual transformations for high-resolution image inpainting. IEEE Transactions on Visualization and Computer Graphics, 29(7):3266-3280, 2022. 2 +[58] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 6 +[59] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7 +[60] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T Freeman. Structure and motion from casual videos. In European Conference on Computer Vision, pages 20-37. Springer, 2022. 3, 5 +[61] Zhixing Zhang, Bichen Wu, Xiaoyan Wang, Yaqiao Luo, Luxin Zhang, Yinan Zhao, Peter Vajda, Dimitris Metaxas, and Licheng Yu. Avid: Any-length video inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7162-7172, 2024. 2 +[62] Haitian Zheng, Zhe Lin, Jingwan Lu, Scott Cohen, Eli Shechtman, Connelly Barnes, Jianming Zhang, Ning Xu, Sohrab Amirghodsi, and Jiebo Luo. Image inpainting with cascaded modulation gan and object-aware training. In European Conference on Computer Vision, pages 277-296. Springer, 2022. 2 +[63] Shijie Zhou, Zhiwen Fan, Dejia Xu, Haoran Chang, Pradyumna Chari, Suya Bharadwaj, Tejas You, Zhangyang Wang, and Achuta Kadambi. Dreamscene360: Unconstrained text-to-3d scene generation with panoramic gaussian splatting. arXiv preprint arXiv:2404.06903, 2024. 2 +[64] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. arXiv preprint arXiv:2312.00451, 2023. 4 \ No newline at end of file diff --git a/2025/Unboxed_ Geometrically and Temporally Consistent Video Outpainting/images.zip b/2025/Unboxed_ Geometrically and Temporally Consistent Video Outpainting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d8ca1e85e40ccc71f0b5fd0a92ad8c5dd5d1fb23 --- /dev/null +++ b/2025/Unboxed_ Geometrically and Temporally Consistent Video Outpainting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73696966813d0b81a9ecdb5032c4810122dd28034d8fba684fdbfb0d82b79d73 +size 869568 diff --git a/2025/Unboxed_ Geometrically and Temporally Consistent Video Outpainting/layout.json b/2025/Unboxed_ Geometrically and Temporally Consistent Video Outpainting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cbbedfd308bd2f329cb2fb20c25b636de57717de --- /dev/null +++ b/2025/Unboxed_ Geometrically and Temporally Consistent Video Outpainting/layout.json @@ -0,0 +1,8831 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 103, + 517, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 103, + 517, + 122 + ], + "spans": [ + { + "bbox": [ + 77, + 103, + 517, + 122 + ], + "type": "text", + "content": "Unboxed: Geometrically and Temporally Consistent Video Outpainting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 142, + 151, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 142, + 151, + 158 + ], + "spans": [ + { + "bbox": [ + 83, + 142, + 151, + 158 + ], + "type": "text", + "content": "Zhongrui " + }, + { + "bbox": [ + 83, + 142, + 151, + 158 + ], + "type": "inline_equation", + "content": "\\mathrm{Yu}^1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 161, + 143, + 286, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 143, + 286, + 158 + ], + "spans": [ + { + "bbox": [ + 161, + 143, + 286, + 158 + ], + "type": "text", + "content": "Martina Megaro-Boldini²" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 172, + 162, + 238, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 162, + 238, + 175 + ], + "spans": [ + { + "bbox": [ + 172, + 162, + 238, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 172, + 162, + 238, + 175 + ], + "type": "text", + "content": "ETH Zürich" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 167, + 182, + 244, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 182, + 244, + 193 + ], + "spans": [ + { + "bbox": [ + 167, + 182, + 244, + 193 + ], + "type": "text", + "content": "zhonyu@ethz.ch" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 143, + 397, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 143, + 397, + 157 + ], + "spans": [ + { + "bbox": [ + 296, + 143, + 397, + 157 + ], + "type": "text", + "content": "Robert W. Sumner1,2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 407, + 143, + 509, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 407, + 143, + 509, + 158 + ], + "spans": [ + { + "bbox": [ + 407, + 143, + 509, + 158 + ], + "type": "text", + "content": "Abdelaziz Djelouah2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 297, + 162, + 421, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 162, + 421, + 177 + ], + "spans": [ + { + "bbox": [ + 297, + 162, + 421, + 177 + ], + "type": "text", + "content": "2DisneyResearch|Studios" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 269, + 182, + 427, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 182, + 427, + 193 + ], + "spans": [ + { + "bbox": [ + 269, + 182, + 427, + 193 + ], + "type": "text", + "content": "adelaziz.djelouah@disney.com" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 31, + 206, + 232, + 377 + ], + "blocks": [ + { + "bbox": [ + 31, + 206, + 232, + 377 + ], + "lines": [ + { + "bbox": [ + 31, + 206, + 232, + 377 + ], + "spans": [ + { + "bbox": [ + 31, + 206, + 232, + 377 + ], + "type": "image", + "image_path": "e95dfc66ca20e059d3146175c4d41d9edd181e90cfed1507dd72e46babf01c15.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 386, + 546, + 409 + ], + "lines": [ + { + "bbox": [ + 46, + 386, + 546, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 386, + 546, + 409 + ], + "type": "text", + "content": "Figure 1. Video outpainting results on two different resolutions. In both case, the area outside the red box is outpainted. Our method achieves results of better quality and more temporally stable, compared to recent works such as Dehan [13], M3DDM [14] and MOTIA [48]." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 241, + 206, + 566, + 373 + ], + "blocks": [ + { + "bbox": [ + 241, + 206, + 566, + 373 + ], + "lines": [ + { + "bbox": [ + 241, + 206, + 566, + 373 + ], + "spans": [ + { + "bbox": [ + 241, + 206, + 566, + 373 + ], + "type": "image", + "image_path": "294ffb16ba03e9e73345713154b66df8e1b93d6c74c910b8f745bae7733d51e0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 143, + 419, + 192, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 419, + 192, + 432 + ], + "spans": [ + { + "bbox": [ + 143, + 419, + 192, + 432 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 451, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 289, + 713 + ], + "type": "text", + "content": "Extending the field of view of video content beyond its original version has many applications: immersive viewing experience with VR devices, reformatting 4:3 legacy content to today's viewing conditions with wide screens, or simply extending vertically captured phone videos. Many existing works focus on synthesizing the video using generative models only. Despite promising results, this strategy seems at the moment limited in terms of quality. In this work, we address this problem using two key ideas: 3D supported outpainting for the static regions of the images, and leveraging pre-trained video diffusion model to ensure realistic and temporally coherent results, particularly for the dynamic parts. In the first stage, we iterate between image outpainting and updating the 3D scene representation - we use 3D Gaussian Splatting. Then we consider dynamic objects independently per frame and inpaint missing pixels. Finally, we propose a denoising scheme that allows to maintain known reliable regions and update the dynamic parts to obtain temporally realistic results. We achieve state-of-the-art video outpainting. This is validated quantitatively and through a user study. We are also able to extend the field of view largely beyond the limits reached by existing methods." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 440, + 386, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 440, + 386, + 452 + ], + "spans": [ + { + "bbox": [ + 306, + 440, + 386, + 452 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 461, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 628 + ], + "type": "text", + "content": "Expanding video frames beyond their original field of view (FoV), known as Video Outpainting, is a technique with broad applications in modern media. One example is 4:3 legacy content re-targeting where video outpainting can be used to fill modern widescreen devices seamlessly while avoiding any unnatural stretching. A more forward-looking application is movie viewing with virtual reality (VR) headsets. Video outpainting can enhance the immersion by embedding the viewer in the scene, eliminating distractions from unrelated backgrounds. However, these applications require extremely high-quality results, which is particularly challenging since the synthesized images must integrate harmoniously with the original content and maintain strong temporal coherence across frames." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Although a wide variety of methods have been proposed, using a patch-based approach [1, 2, 12, 46], optical flow [13] or Generative Adversarial Networks (GAN) [24, 53], most recent methods have clearly identified Diffusion Models [14, 48] as the most promising direction. M3DDM [14] trains a 3D diffusion model on large-scale data for video outpainting with a mask modeling approach." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7309" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "In contrast, MOTIA [48] proposes to fine-tune a pre-trained video diffusion model [19] on each test video to capture data-specific motion and content patterns. However, both of these methods have limitations. First, the temporal consistency of the outpainted regions is still unsatisfactory. We can observe duplications, distortions or objects appearing and disappearing. Second, relying entirely on video diffusion models has a strong impact in terms of memory and limits the methods to low resolution and small field of view extension. Some of these issues can be observed in Figure 1. Both M3DDM [14] and MOTIA [48] have strong visual artifacts and only the latter could be used to outpaint the video to a larger field of view." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 236, + 289, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 236, + 289, + 440 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 289, + 440 + ], + "type": "text", + "content": "To address these issues we propose two core ideas: 3D-supported outpainting for static regions and conditional video synthesis with a pre-trained video diffusion model to ensure realistic, temporally coherent results for dynamic areas. Following these key ideas, we decompose the video into static and dynamic regions and break down the video outpainting task into three stages. First, for static regions, we rely on 3D Gaussians [23] as a supporting 3D representation, and we alternate between outpainting key frames and updating the 3D representation. Second, for dynamic objects, we implement an object-wise inpainting. Here, among other things, we leverage object-tracking masks in each frame to define inpainting regions and use a fine-tuned diffusion model to inpaint. Lastly, using the resulting frames as a starting point, we synthesize temporally consistent results for all frames using a pre-trained video diffusion model [7]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 449, + 289, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 289, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 289, + 569 + ], + "type": "text", + "content": "Figure 1 demonstrates the effectiveness of our method. We achieve better results than prior methods both in terms of video quality and temporal consistency. This is confirmed by both quantitative and qualitative experiments. Furthermore, by using GPU-efficient Gaussian Splatting for static region outpainting and avoiding fine-tuning of the video diffusion model, we can go from an input " + }, + { + "bbox": [ + 46, + 449, + 289, + 569 + ], + "type": "inline_equation", + "content": "480 \\times 480" + }, + { + "bbox": [ + 46, + 449, + 289, + 569 + ], + "type": "text", + "content": " video (" + }, + { + "bbox": [ + 46, + 449, + 289, + 569 + ], + "type": "inline_equation", + "content": "\\mathrm{FoV} = 31^{\\circ}" + }, + { + "bbox": [ + 46, + 449, + 289, + 569 + ], + "type": "text", + "content": ") to an output resolution of " + }, + { + "bbox": [ + 46, + 449, + 289, + 569 + ], + "type": "inline_equation", + "content": "2560 \\times 720" + }, + { + "bbox": [ + 46, + 449, + 289, + 569 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 449, + 289, + 569 + ], + "type": "inline_equation", + "content": "\\mathrm{FoV} = 120^{\\circ}" + }, + { + "bbox": [ + 46, + 449, + 289, + 569 + ], + "type": "text", + "content": ") with a GPU memory usage of less than 16GB for the entire pipeline." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 578, + 244, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 578, + 244, + 589 + ], + "spans": [ + { + "bbox": [ + 59, + 578, + 244, + 589 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 599, + 286, + 718 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 47, + 599, + 286, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 286, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 286, + 621 + ], + "type": "text", + "content": "- a novel video outpainting method that combines a 3D representation with a pre-trained video diffusion model;" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 623, + 286, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 286, + 657 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 286, + 657 + ], + "type": "text", + "content": "- an optimization strategy with an additional loss to enable seamless integration of the outpainting with the original static content;" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 658, + 286, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 286, + 705 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 286, + 705 + ], + "type": "text", + "content": "- demonstrating that a pre-trained video diffusion model can function as a temporal filter, making discontinuous frames temporally coherent and simplifying the inpainting task for the dynamic parts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 707, + 235, + 718 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 707, + 235, + 718 + ], + "spans": [ + { + "bbox": [ + 47, + 707, + 235, + 718 + ], + "type": "text", + "content": "- State-of-the-art results for video outpainting" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 91, + 545, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 330 + ], + "type": "text", + "content": "Immersive Scene Generation. To create immersive experiences for users in VR applications, several methods are proposed to generate realistic 3D or 4D scenes that allow for arbitrary viewpoint exploration. LucidDreamer [10] and the following works [16, 35, 43, 54, 55] lift a single image to 3D space, project it from different viewpoints, and inpaint them using diffusion models to create a 3D scene. However, directly applying their method to consecutive video frames would cause severe geometric distortion between frames due to the simple depth alignment and optimization techniques. Dreamscene360 [63] and 4D4KGen [30] create panoramic 3D and 4D scenes separately from a text-generated panoramic image. VividDream [27] generates multi-view videos from a single image and reconstructs a 4D scene with 4D Gaussians. These methods are primarily designed for static or simply animated images and cannot handle videos with complex camera and object motion. Methods such as [9, 28, 29, 49, 50] reconstruct 4D scenes from casual videos, but they are unable to complete regions beyond what is captured in the video." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 331, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 533 + ], + "type": "text", + "content": "Image completion. Image completion [6, 39], including image inpainting and outpainting, is a task that restores missing part of an image, and it serves as a basis for video outpainting. Traditional methods [6, 18, 21] use low-level features from the incomplete image to propagate information in known regions to fill missing areas. Early deep learning based methods [32, 42, 44] introduced end-to-end frameworks that learned mappings from corrupted to complete images. Recent methods have increasingly leveraged generative models, including GAN [17, 56, 57, 62], and Diffusion Models [4, 11, 22, 34, 40, 41]. Reference-based image completion methods utilize reference images as generation conditions [52] or fine-tune models on a set of reference images [45] to produce completions that are more faithful to the references. However, even fine-tuned on all input video frames, frame-by-frame completion introduces substantial temporal inconsistency in video outpainting." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "Video Outpainting. In the field of video completion, video outpainting is challenging but less explored. Early patch-based methods [1, 2] select the nearest patch from the spatio-temporal domain of the video to expand each frame. Lee et al. [26] employ 3D information to warp neighboring frames to the current frame. Dehan et al. [13] utilize estimated optical flow to warp the background from adjacent frames and use an image completion network to fill any remaining uncompleted areas. While Zhang et al. [61] target video in-painting motion module and structure guidance. M3DDM [14] extends the image diffusion model framework to 3D and trains it on extensive video data for the outpainting task. A global frame is introduced as a conditioning input to improve global information perception. MOTIA [48] proposes a method that fine-tunes a pre-trained" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7310" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 79, + 145, + 118 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 74, + 78 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 74, + 78 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 74, + 78 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 79, + 145, + 118 + ], + "lines": [ + { + "bbox": [ + 56, + 79, + 145, + 118 + ], + "spans": [ + { + "bbox": [ + 56, + 79, + 145, + 118 + ], + "type": "image", + "image_path": "b5c993a86ea967cea348375fe72fd26011ea7a2a7d2c501568417d548a47784f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 151, + 87, + 182 + ], + "blocks": [ + { + "bbox": [ + 56, + 130, + 100, + 140 + ], + "lines": [ + { + "bbox": [ + 56, + 130, + 100, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 100, + 140 + ], + "type": "text", + "content": "Pre-process" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 151, + 87, + 182 + ], + "lines": [ + { + "bbox": [ + 56, + 151, + 87, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 87, + 182 + ], + "type": "image", + "image_path": "09a54f17834f1d8eec1f0d39bfb2bcacdbc37c3dfe82d38740419aec20612394.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 88, + 151, + 119, + 182 + ], + "blocks": [ + { + "bbox": [ + 56, + 143, + 111, + 150 + ], + "lines": [ + { + "bbox": [ + 56, + 143, + 111, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 143, + 111, + 150 + ], + "type": "text", + "content": "Dynamic Object Masks" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 88, + 151, + 119, + 182 + ], + "lines": [ + { + "bbox": [ + 88, + 151, + 119, + 182 + ], + "spans": [ + { + "bbox": [ + 88, + 151, + 119, + 182 + ], + "type": "image", + "image_path": "3405ae143188f7bed6ed6f3a352b9f5ef03dc44aad903a2064c2aa01b39507c0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 119, + 151, + 148, + 182 + ], + "blocks": [ + { + "bbox": [ + 119, + 151, + 148, + 182 + ], + "lines": [ + { + "bbox": [ + 119, + 151, + 148, + 182 + ], + "spans": [ + { + "bbox": [ + 119, + 151, + 148, + 182 + ], + "type": "image", + "image_path": "41fbf127238cb8c67fb260b2b1d9232e72dcb00a37224087e6f655a114ffdacd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 53, + 201, + 145, + 266 + ], + "blocks": [ + { + "bbox": [ + 59, + 191, + 122, + 198 + ], + "lines": [ + { + "bbox": [ + 59, + 191, + 122, + 198 + ], + "spans": [ + { + "bbox": [ + 59, + 191, + 122, + 198 + ], + "type": "text", + "content": "Static Scene Reconstruction" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 201, + 145, + 266 + ], + "lines": [ + { + "bbox": [ + 53, + 201, + 145, + 266 + ], + "spans": [ + { + "bbox": [ + 53, + 201, + 145, + 266 + ], + "type": "image", + "image_path": "72cf0259f7d1001528b546c82cbcfb4b45dd3d8c71acc92cd9114d67e825d918.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 157, + 86, + 280, + 266 + ], + "blocks": [ + { + "bbox": [ + 153, + 71, + 215, + 82 + ], + "lines": [ + { + "bbox": [ + 153, + 71, + 215, + 82 + ], + "spans": [ + { + "bbox": [ + 153, + 71, + 215, + 82 + ], + "type": "text", + "content": "(1) Static Region" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 157, + 86, + 280, + 266 + ], + "lines": [ + { + "bbox": [ + 157, + 86, + 280, + 266 + ], + "spans": [ + { + "bbox": [ + 157, + 86, + 280, + 266 + ], + "type": "image", + "image_path": "ab7e21691daaa99ca318ffecf57409926e41322488202fc41bf425c28a2ed81a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 286, + 86, + 423, + 273 + ], + "blocks": [ + { + "bbox": [ + 284, + 72, + 359, + 82 + ], + "lines": [ + { + "bbox": [ + 284, + 72, + 359, + 82 + ], + "spans": [ + { + "bbox": [ + 284, + 72, + 359, + 82 + ], + "type": "text", + "content": "(2) Dynamic Objects" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 286, + 86, + 423, + 273 + ], + "lines": [ + { + "bbox": [ + 286, + 86, + 423, + 273 + ], + "spans": [ + { + "bbox": [ + 286, + 86, + 423, + 273 + ], + "type": "image", + "image_path": "6a1f9e6b97856706050a05db1c0930f537c9e6ec14c573710aed566bb8e78b87.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 281, + 547, + 348 + ], + "lines": [ + { + "bbox": [ + 46, + 281, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 547, + 348 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed method. Starting from the input video sequence, dynamic objects are separated from the static parts of the image, and a static scene reconstruction is made using Gaussian Splatting (GS). In the first stage of our method, the static part is outpainted by iterating between image outpainting, depth prediction and updating the GS model. After this stage, we identify parts of the dynamic object that may require inpainting. Here, an image diffusion model is finetuned on the input frames and used for inpainting each frame independently, which may result in temporal inconsistencies as seen of the car windows. In the last step, a pre-trained video diffusion model is used to fix temporal inconsistencies. Details about each step are provided in the text." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 432, + 92, + 544, + 273 + ], + "blocks": [ + { + "bbox": [ + 433, + 72, + 525, + 82 + ], + "lines": [ + { + "bbox": [ + 433, + 72, + 525, + 82 + ], + "spans": [ + { + "bbox": [ + 433, + 72, + 525, + 82 + ], + "type": "text", + "content": "(3) Final Video Synthesis" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 446, + 85, + 514, + 91 + ], + "lines": [ + { + "bbox": [ + 446, + 85, + 514, + 91 + ], + "spans": [ + { + "bbox": [ + 446, + 85, + 514, + 91 + ], + "type": "text", + "content": "Temporally Consistent Results" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 432, + 92, + 544, + 273 + ], + "lines": [ + { + "bbox": [ + 432, + 92, + 544, + 273 + ], + "spans": [ + { + "bbox": [ + 432, + 92, + 544, + 273 + ], + "type": "image", + "image_path": "6e319d107356edd0ba8fc84e351a4f791deec53c22212860a34e9453f404ba15.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 355, + 287, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 355, + 287, + 427 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 287, + 427 + ], + "type": "text", + "content": "video diffusion model on test samples to learn the intrinsic patterns of a video. In general, diffusion-based approaches rely on priors learned by temporal modules to maintain continuity in the outpainted regions, which is not sufficient. We demonstrate that a pipeline integrating constraints from a 3D representation achieves results of much higher quality." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 440, + 102, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 440, + 102, + 452 + ], + "spans": [ + { + "bbox": [ + 47, + 440, + 102, + 452 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 460, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 568 + ], + "type": "text", + "content": "Given a sequence of video frames " + }, + { + "bbox": [ + 46, + 460, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\{I_i\\in \\mathbb{R}^{h\\times w\\times c}\\}_{i = 1}^N" + }, + { + "bbox": [ + 46, + 460, + 287, + 568 + ], + "type": "text", + "content": ", our goal is to expand each frame beyond its original field of view to " + }, + { + "bbox": [ + 46, + 460, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\{\\tilde{I}_i\\in \\mathbb{R}^{H\\times W\\times c}\\}_{i = 1}^N" + }, + { + "bbox": [ + 46, + 460, + 287, + 568 + ], + "type": "text", + "content": ", while maintaining both intra-frame and cross-frame consistency. Although the problem is more challenging than image outpainting, due to the temporal consistency requirements, the presence of multiple frames offers additional observations that can be leveraged. More specifically, it is possible to build a 3D representation of the scene using 3D Gaussians." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "As illustrated in Figure 2, our proposed method can be divided into 3 distinct stages. In the first stage, we focus on the static parts of the scene and alternate between image outpainting and updating the 3D scene representation. In the second stage, we focus on the dynamic objects in the scene. Here, we fine-tune an image diffusion model on the available images and use it to inpaint the missing parts of the object in each frame. At this point, all the frames are outpainted, but temporal inconsistencies remain. These frames serve as a starting point for a video synthesis scheme leveraging pre-trained stable video diffusion to synthesize a temporally consistent video sequence. Next, we detail each" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 355, + 387, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 355, + 387, + 367 + ], + "spans": [ + { + "bbox": [ + 306, + 355, + 387, + 367 + ], + "type": "text", + "content": "of these main steps." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 376, + 537, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 376, + 537, + 390 + ], + "spans": [ + { + "bbox": [ + 306, + 376, + 537, + 390 + ], + "type": "text", + "content": "3.1. Pre-processing and Static Scene Outpainting" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 396, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 552 + ], + "type": "text", + "content": "From the static parts of the video, it is possible to estimate a 3D reconstruction of the scene and use the different views across time to provide information about the regions to outpaint. In the first step, we identify the dynamic parts of the scene. A wide variety of methods can be employed here, but we simply rely on object segmentation and, in particular, object tracking models [33]. Figure 2 shows an example of the binary segmentation masks. These masks are expanded to avoid including incorrectly labeled pixels into the background. We also compute a 2D bounding box around the object to help keep track of the object size in case it partially or completely leaves the field of view. The static parts are all the pixels that are not part of any dynamic object." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "Static Scene Reconstruction. We propose to use Gaussian Splatting (GS) [23] to reconstruct the static part of the video. This approach allows regions visible in other frames to be rendered into the current frame. For the input video frames " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\{I_i\\}_{i=1}^N" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": ", we first estimate the corresponding camera poses " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{p}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " and intrinsic parameters " + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{K}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": " using an off-the-shelf camera pose estimator [60]. We compute a depth map for each frame using parallax and project the background pixels into 3D space, yielding an initial pointcloud for GS training. The training process is supervised by the image reconstruction loss and the depth loss. The image reconstruction loss is the combination of the L1 loss and" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7311" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 83, + 122, + 135 + ], + "blocks": [ + { + "bbox": [ + 91, + 72, + 159, + 82 + ], + "lines": [ + { + "bbox": [ + 91, + 72, + 159, + 82 + ], + "spans": [ + { + "bbox": [ + 91, + 72, + 159, + 82 + ], + "type": "text", + "content": "First key-frame " + }, + { + "bbox": [ + 91, + 72, + 159, + 82 + ], + "type": "inline_equation", + "content": "(k = 1)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 83, + 122, + 135 + ], + "lines": [ + { + "bbox": [ + 49, + 83, + 122, + 135 + ], + "spans": [ + { + "bbox": [ + 49, + 83, + 122, + 135 + ], + "type": "image", + "image_path": "89a119bdb865156ce67a37a4ea8c2d23e43ad9c77436f0065b7782e27badc70e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 137, + 94, + 147 + ], + "lines": [ + { + "bbox": [ + 77, + 137, + 94, + 147 + ], + "spans": [ + { + "bbox": [ + 77, + 137, + 94, + 147 + ], + "type": "inline_equation", + "content": "I_{k = 1}^{\\mathrm{static}}" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 125, + 83, + 199, + 135 + ], + "blocks": [ + { + "bbox": [ + 125, + 83, + 199, + 135 + ], + "lines": [ + { + "bbox": [ + 125, + 83, + 199, + 135 + ], + "spans": [ + { + "bbox": [ + 125, + 83, + 199, + 135 + ], + "type": "image", + "image_path": "72dc690b512c7129f381d8742b4619f55f25810f9ead2305a88e0b6b336ede05.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 137, + 171, + 146 + ], + "lines": [ + { + "bbox": [ + 154, + 137, + 171, + 146 + ], + "spans": [ + { + "bbox": [ + 154, + 137, + 171, + 146 + ], + "type": "inline_equation", + "content": "I_{k = 1}^{\\mathrm{static}}" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 203, + 83, + 276, + 135 + ], + "blocks": [ + { + "bbox": [ + 203, + 83, + 276, + 135 + ], + "lines": [ + { + "bbox": [ + 203, + 83, + 276, + 135 + ], + "spans": [ + { + "bbox": [ + 203, + 83, + 276, + 135 + ], + "type": "image", + "image_path": "af585d567554915f32c7b929fd1f75dfa311d581fa21d8d13643f4b6bd019e48.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 136, + 250, + 144 + ], + "lines": [ + { + "bbox": [ + 229, + 136, + 250, + 144 + ], + "spans": [ + { + "bbox": [ + 229, + 136, + 250, + 144 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{recon}}" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 278, + 83, + 353, + 135 + ], + "blocks": [ + { + "bbox": [ + 277, + 72, + 357, + 82 + ], + "lines": [ + { + "bbox": [ + 277, + 72, + 357, + 82 + ], + "spans": [ + { + "bbox": [ + 277, + 72, + 357, + 82 + ], + "type": "text", + "content": "Depth after GS finetuning" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 278, + 83, + 353, + 135 + ], + "lines": [ + { + "bbox": [ + 278, + 83, + 353, + 135 + ], + "spans": [ + { + "bbox": [ + 278, + 83, + 353, + 135 + ], + "type": "image", + "image_path": "37bc47ebb604cf9485529c50ea63149efe391394487854107dcb2f2bbb709de5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 136, + 342, + 145 + ], + "lines": [ + { + "bbox": [ + 290, + 136, + 342, + 145 + ], + "spans": [ + { + "bbox": [ + 290, + 136, + 342, + 145 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{recon}} + L_{\\mathrm{depth}}" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 357, + 83, + 430, + 135 + ], + "blocks": [ + { + "bbox": [ + 357, + 83, + 430, + 135 + ], + "lines": [ + { + "bbox": [ + 357, + 83, + 430, + 135 + ], + "spans": [ + { + "bbox": [ + 357, + 83, + 430, + 135 + ], + "type": "image", + "image_path": "a6a9c7eb4e65893f26ad32169c9f36d5d7327a6bd1ba64c71bc5834ed31c2756.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 136, + 429, + 146 + ], + "lines": [ + { + "bbox": [ + 358, + 136, + 429, + 146 + ], + "spans": [ + { + "bbox": [ + 358, + 136, + 429, + 146 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{recon}} + L_{\\mathrm{depth}} + L_{\\mathrm{depth}}^{\\mathrm{Sp}}" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 467, + 83, + 541, + 136 + ], + "blocks": [ + { + "bbox": [ + 469, + 73, + 537, + 82 + ], + "lines": [ + { + "bbox": [ + 469, + 73, + 537, + 82 + ], + "spans": [ + { + "bbox": [ + 469, + 73, + 537, + 82 + ], + "type": "text", + "content": "Next key-frame " + }, + { + "bbox": [ + 469, + 73, + 537, + 82 + ], + "type": "inline_equation", + "content": "(k = 2)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 467, + 83, + 541, + 136 + ], + "lines": [ + { + "bbox": [ + 467, + 83, + 541, + 136 + ], + "spans": [ + { + "bbox": [ + 467, + 83, + 541, + 136 + ], + "type": "image", + "image_path": "2d843ea0a1c70843f385b3d8bba6b6d0eb60188295e4810ac20f9ff8c35f72b8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 137, + 515, + 150 + ], + "lines": [ + { + "bbox": [ + 493, + 137, + 515, + 150 + ], + "spans": [ + { + "bbox": [ + 493, + 137, + 515, + 150 + ], + "type": "inline_equation", + "content": "I_{k = 2}^{\\mathrm{static}}" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 152, + 547, + 197 + ], + "lines": [ + { + "bbox": [ + 46, + 152, + 547, + 197 + ], + "spans": [ + { + "bbox": [ + 46, + 152, + 547, + 197 + ], + "type": "text", + "content": "Figure 3. Effect of the loss terms on the GS fine-tuning. Starting from the rasterization of the 3D Gaussians for the first key-frame " + }, + { + "bbox": [ + 46, + 152, + 547, + 197 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 46, + 152, + 547, + 197 + ], + "type": "text", + "content": ", we use image outpainting to obtain " + }, + { + "bbox": [ + 46, + 152, + 547, + 197 + ], + "type": "inline_equation", + "content": "\\hat{I}_k^{\\mathrm{static}}" + }, + { + "bbox": [ + 46, + 152, + 547, + 197 + ], + "type": "text", + "content": ". This image is subdivided into super-pixels as illustrated here. During the optimization to update the GS model, we use a combination of losses. We can see that all are needed to avoid discontinuities and obtain a realistic smooth depth map. After this update, the process is repeated with the next key-frame using the updated 3D Gaussians." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 209, + 287, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 209, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 287, + 281 + ], + "type": "text", + "content": "the structural similarity loss (SSIM). The depth loss acts as a regularization and uses Pearson Correlation [64] between the rendered depth and the depth estimated from a monocular depth estimation model [37]. Apart from the dynamic object masking, this is similar to the optimization proposed by Zhu et al. [64]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 294, + 288, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 288, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 288, + 462 + ], + "type": "text", + "content": "Single Image Outpainting. Starting from the GS model " + }, + { + "bbox": [ + 46, + 294, + 288, + 462 + ], + "type": "inline_equation", + "content": "GS_{\\mathrm{static}}" + }, + { + "bbox": [ + 46, + 294, + 288, + 462 + ], + "type": "text", + "content": " representing the static parts of the scene, we uniformly sample " + }, + { + "bbox": [ + 46, + 294, + 288, + 462 + ], + "type": "inline_equation", + "content": "k \\in \\{1, \\dots, K\\}" + }, + { + "bbox": [ + 46, + 294, + 288, + 462 + ], + "type": "text", + "content": " key frames. Our objective is to outpaint them one by one, while we progressively update the GS model. Here we first detail outpainting for a single image. Considering frame " + }, + { + "bbox": [ + 46, + 294, + 288, + 462 + ], + "type": "inline_equation", + "content": "I_k^{\\mathrm{static}}" + }, + { + "bbox": [ + 46, + 294, + 288, + 462 + ], + "type": "text", + "content": ", we adjust its field of view to obtain new intrinsic parameters " + }, + { + "bbox": [ + 46, + 294, + 288, + 462 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{K}}_k" + }, + { + "bbox": [ + 46, + 294, + 288, + 462 + ], + "type": "text", + "content": ", and render " + }, + { + "bbox": [ + 46, + 294, + 288, + 462 + ], + "type": "inline_equation", + "content": "GS_{\\mathrm{static}}" + }, + { + "bbox": [ + 46, + 294, + 288, + 462 + ], + "type": "text", + "content": " based on the new parameters to get the image " + }, + { + "bbox": [ + 46, + 294, + 288, + 462 + ], + "type": "inline_equation", + "content": "\\hat{I}_k^{\\mathrm{static}}" + }, + { + "bbox": [ + 46, + 294, + 288, + 462 + ], + "type": "text", + "content": ". The mask " + }, + { + "bbox": [ + 46, + 294, + 288, + 462 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{m}}_k^{\\mathrm{static}}" + }, + { + "bbox": [ + 46, + 294, + 288, + 462 + ], + "type": "text", + "content": " for outpainting is defined by checking whether each pixel is covered by the projection of Gaussian points. Any image outpainting method can be used at this level, but we find that the Stable Diffusion XL (SDXL) image outpainting model [38] performs well for our task. This can be summarized as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 471, + 287, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 471, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 107, + 471, + 287, + 485 + ], + "type": "interline_equation", + "content": "I _ {k} ^ {\\text {s t a t i c}} = \\phi_ {G S _ {\\text {s t a t i c}}} \\left(\\hat {\\mathbf {K}} _ {k}, \\mathbf {p} _ {k}\\right) \\tag {1}", + "image_path": "b541a4e641f3b3edbca11b93df4ae6786b169dc61f930b62cc19c74b79a1dff8.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 108, + 487, + 287, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 487, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 108, + 487, + 287, + 502 + ], + "type": "interline_equation", + "content": "\\hat {I} _ {k} ^ {\\text {s t a t i c}} = \\phi_ {\\mathrm {S D X L}} \\left(I _ {k} ^ {\\text {s t a t i c}}, \\hat {\\mathbf {m}} _ {k} ^ {\\text {s t a t i c}}\\right) \\tag {2}", + "image_path": "f4f360be32f52e04621fee42f51b81b0612f234d81f4912074be3c780b15ba47.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 515, + 287, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 515, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 287, + 622 + ], + "type": "text", + "content": "Updating the GS model. To ensure the consistency of the outpainting across frames, we update the GS model each time we outpaint a key-frame. Denote by " + }, + { + "bbox": [ + 46, + 515, + 287, + 622 + ], + "type": "inline_equation", + "content": "K'" + }, + { + "bbox": [ + 46, + 515, + 287, + 622 + ], + "type": "text", + "content": " the number of key-frames that have been outpainted so far. First, we estimate the depth for the newly outpainted image using a single image depth model [37]. Then we initialize new Gaussian points for the outpainted pixel and fine-tune the GS model with this additional data. We use an image reconstruction loss" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 84, + 629, + 287, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 629, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 84, + 629, + 287, + 680 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {\\text {r e c o n}} = \\sum_ {k = 1} ^ {K ^ {\\prime}} \\lambda_ {1} L _ {1} \\left(\\hat {I} _ {k} ^ {\\text {s t a t i c}}, \\phi_ {\\mathrm {G S}} \\left(\\hat {K} _ {k}, p _ {k}\\right)\\right) \\tag {3} \\\\ + \\lambda_ {2} L _ {\\mathrm {S S I M}} (\\hat {I} _ {k} ^ {\\text {s t a t i c}}, \\phi_ {\\mathrm {G S}} (\\hat {K} _ {k}, p _ {k})) \\\\ \\end{array}", + "image_path": "6a8f38f320aaa303377afe5933081f986991503f74f83bace024885d7ac5589f.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "between the rasterization from the model " + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{GS}}(\\hat{K}_k,p_k)" + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": " and the outpainting result. We also use 2 depth based losses, the" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 209, + 432, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 209, + 432, + 221 + ], + "spans": [ + { + "bbox": [ + 306, + 209, + 432, + 221 + ], + "type": "text", + "content": "first one on the full depth map:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 319, + 231, + 545, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 231, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 319, + 231, + 545, + 269 + ], + "type": "interline_equation", + "content": "L _ {\\text {d e p t h}} = \\sum_ {k = 1} ^ {K ^ {\\prime}} \\frac {\\operatorname {C o v} \\left(D \\left(\\hat {I} _ {k} ^ {\\text {s t a t i c}}\\right) , D _ {\\mathrm {G S}} \\left(\\hat {K} _ {k} , p _ {k}\\right)\\right)}{\\sqrt {\\operatorname {V a r} \\left(D \\left(\\hat {I} _ {k} ^ {\\text {s t a t i c}}\\right)\\right) \\operatorname {V a r} \\left(D _ {\\mathrm {G S}} \\left(\\hat {K} _ {k} , p _ {k}\\right)\\right)}} \\tag {4}", + "image_path": "51316925900b9feb616a30a8aead11e0e886233e04b45bc5a55981e752a728f9.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 278, + 546, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 278, + 546, + 329 + ], + "spans": [ + { + "bbox": [ + 305, + 278, + 546, + 329 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 278, + 546, + 329 + ], + "type": "inline_equation", + "content": "D(\\hat{I}_k^{\\mathrm{static}})" + }, + { + "bbox": [ + 305, + 278, + 546, + 329 + ], + "type": "text", + "content": " is the depth predicted by the single image depth model [37], and " + }, + { + "bbox": [ + 305, + 278, + 546, + 329 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{GS}}(\\hat{K}_k,p_k)" + }, + { + "bbox": [ + 305, + 278, + 546, + 329 + ], + "type": "text", + "content": " is the depth from GS model that we optimize. For every image, we extract " + }, + { + "bbox": [ + 305, + 278, + 546, + 329 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 278, + 546, + 329 + ], + "type": "text", + "content": " super-pixels, and use the second depth loss at this level:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 339, + 545, + 387 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 545, + 387 + ], + "type": "interline_equation", + "content": "L _ {\\text {d e p t h}} ^ {\\mathrm {S p}} = \\sum_ {k = 1} ^ {K ^ {\\prime}} \\sum_ {s = 1} ^ {S} \\frac {\\operatorname {C o v} _ {s} \\left(D \\left(\\hat {I} _ {k} ^ {\\text {s t a t i c}}\\right) , D _ {\\mathrm {G S}} \\left(\\hat {K} _ {k} , p _ {k}\\right)\\right)}{\\sqrt {\\operatorname {V a r} _ {s} \\left(D \\left(\\hat {I} _ {k} ^ {\\text {s t a t i c}}\\right)\\right) \\operatorname {V a r} _ {s} \\left(D _ {\\mathrm {G S}} \\left(\\hat {K} _ {k} , p _ {k}\\right)\\right)}} \\tag {5}", + "image_path": "870afaab44898bfbc30c935388666a72d7972de4971424dd15b9ca04f611a577.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 388, + 546, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 546, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 546, + 542 + ], + "type": "text", + "content": "Figure 3 demonstrate this process and shows the importance of the different loss terms for the GS update. Starting from the rasterization of the initial GS model, we can see that the 3D model does not cover the new field of view. The frame is outpainted and superpixels are illustrated in yellow. Using only the reconstruction loss " + }, + { + "bbox": [ + 304, + 388, + 546, + 542 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{recon}}" + }, + { + "bbox": [ + 304, + 388, + 546, + 542 + ], + "type": "text", + "content": " leads to strong discontinuities between the original GS model and the new outpainted areas. Adding the depth loss " + }, + { + "bbox": [ + 304, + 388, + 546, + 542 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{depth}}" + }, + { + "bbox": [ + 304, + 388, + 546, + 542 + ], + "type": "text", + "content": " helps, but the bottom part of the image still has important discontinuities that would be revealed when projecting to other frames (i.e. other views). Using all the proposed loss terms provides the best results, and the process can now be iterated by adding new views." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 551, + 534, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 534, + 564 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 534, + 564 + ], + "type": "text", + "content": "3.2. Initial Inpainting for the Dynamic Object(s)" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 570, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 640 + ], + "type": "text", + "content": "After the static scene outpainting, we can render any frame with " + }, + { + "bbox": [ + 304, + 570, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{GS}}(\\hat{K}_i,p_i)" + }, + { + "bbox": [ + 304, + 570, + 545, + 640 + ], + "type": "text", + "content": ". As illustrated in Figure 2, we miss part of the object when it is partially outside the filed of view. We need to inpaint this missing part. We use the 2D tracking bounding box to identify the areas that require inpainting for each frame of the video." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 642, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 546, + 712 + ], + "type": "text", + "content": "For this inpainting problem we use a similar strategy as ReaFill [45]. An image diffusion model is fine-tuned for the inpainting task on the input frames (in our case SDXL). It is then applied independently on each frame. Thanks to the fine-tuning, the resulting inpainting is very similar across frames. However temporal inconsistencies remain as it is" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7312" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 42, + 70, + 292, + 208 + ], + "blocks": [ + { + "bbox": [ + 42, + 70, + 292, + 208 + ], + "lines": [ + { + "bbox": [ + 42, + 70, + 292, + 208 + ], + "spans": [ + { + "bbox": [ + 42, + 70, + 292, + 208 + ], + "type": "image", + "image_path": "cd18af78ee96da840a66cb6044ef3e42c6573947b61da9d570a129dd0e21c650.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 216, + 287, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 216, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 287, + 293 + ], + "type": "text", + "content": "Figure 4. Details of the video latent denoising strategy. Starting from the independently outpainted frames " + }, + { + "bbox": [ + 46, + 216, + 287, + 293 + ], + "type": "inline_equation", + "content": "\\tilde{I}_{i\\in 1,\\dots,N}" + }, + { + "bbox": [ + 46, + 216, + 287, + 293 + ], + "type": "text", + "content": ", we obtained the corresponding latents in the forward process " + }, + { + "bbox": [ + 46, + 216, + 287, + 293 + ], + "type": "inline_equation", + "content": "(\\tilde{x}_{i\\in 1,\\dots,N})" + }, + { + "bbox": [ + 46, + 216, + 287, + 293 + ], + "type": "text", + "content": ". In the denoising process we blend known latent with denoising latents according to the mask " + }, + { + "bbox": [ + 46, + 216, + 287, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 46, + 216, + 287, + 293 + ], + "type": "text", + "content": ". The mask and the time step " + }, + { + "bbox": [ + 46, + 216, + 287, + 293 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{end}}" + }, + { + "bbox": [ + 46, + 216, + 287, + 293 + ], + "type": "text", + "content": " respectively control which region of the frames get updated by the video diffusion model and how large the changes are." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 303, + 287, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 287, + 338 + ], + "type": "text", + "content": "clearly visible on the inpainting results in Figure 2. Next we describe how a Video Diffusion Model [7, 8] is used to fix all these temporal artifacts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 345, + 181, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 345, + 181, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 345, + 181, + 357 + ], + "type": "text", + "content": "3.3. Guided Video Synthesis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 364, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 364, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 287, + 496 + ], + "type": "text", + "content": "If we consider the original input sequences of frames " + }, + { + "bbox": [ + 46, + 364, + 287, + 496 + ], + "type": "inline_equation", + "content": "I_{i\\in 1,\\dots,N}" + }, + { + "bbox": [ + 46, + 364, + 287, + 496 + ], + "type": "text", + "content": ", after the previous stages we obtain fully out-painted frames " + }, + { + "bbox": [ + 46, + 364, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\tilde{I}_{i\\in 1,\\dots,N}" + }, + { + "bbox": [ + 46, + 364, + 287, + 496 + ], + "type": "text", + "content": ". Static regions are consistent thanks to the supporting GS model, while the dynamic objects are independently inpainted for each frame. Starting from this data, our objective is to synthesize a realistic sequence of frames " + }, + { + "bbox": [ + 46, + 364, + 287, + 496 + ], + "type": "inline_equation", + "content": "I_{i\\in 1,\\dots,N}^{*}" + }, + { + "bbox": [ + 46, + 364, + 287, + 496 + ], + "type": "text", + "content": ". Not only the temporal inconsistencies on the dynamic objects need to be fixed, but the entire outpainted regions need to appear realistic. Indeed, real-world videos often contain subtle background movements, such as leaves swaying in the wind or water rippling." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 497, + 287, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 615 + ], + "type": "text", + "content": "We leverage pre-trained video diffusion models for this video synthesis. The idea is that video diffusion models are trained for generating real video sequences, with consistent content and realistic motion. We propose using the previously generate frames in a denoising scheme that modifies the different frame areas based on the expected inconsistency level. On the dynamic objects, we expect more changes (as the inpainting is independent), while less change should be allowed on the rest of the image as it is supported by the 3D consistent GS model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 616, + 287, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 287, + 674 + ], + "type": "text", + "content": "More specifically, if we consider a latent video diffusion model, a sequence of input frames " + }, + { + "bbox": [ + 46, + 616, + 287, + 674 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 46, + 616, + 287, + 674 + ], + "type": "text", + "content": " is encoded into the corresponding sequence of latent values " + }, + { + "bbox": [ + 46, + 616, + 287, + 674 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 616, + 287, + 674 + ], + "type": "text", + "content": ". The forward process can be applied to the sequence of frames " + }, + { + "bbox": [ + 46, + 616, + 287, + 674 + ], + "type": "inline_equation", + "content": "\\tilde{I}_{i\\in 1,\\dots,N}" + }, + { + "bbox": [ + 46, + 616, + 287, + 674 + ], + "type": "text", + "content": " to obtain known latents" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 681, + 287, + 695 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 681, + 287, + 695 + ], + "spans": [ + { + "bbox": [ + 102, + 681, + 287, + 695 + ], + "type": "interline_equation", + "content": "\\tilde {x} _ {t - 1} \\sim \\mathcal {N} \\left(\\sqrt {\\bar {\\alpha} _ {t}} \\tilde {x} _ {0}, \\sqrt {1 - \\bar {\\alpha} _ {t}} \\mathbf {1}\\right). \\tag {6}", + "image_path": "27c1ea871641fdf05d49fc57dc953c45668898c33aa8bcd5b68a4e63d2edf262.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 701, + 231, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 231, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 231, + 713 + ], + "type": "text", + "content": "For simplicity, we dropped the frame index " + }, + { + "bbox": [ + 47, + 701, + 231, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 701, + 231, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "We use the reverse process to synthesize the final video sequence. At each time step, we denoise previous latents using the latent video diffusion model with parameters " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": ", to obtain temporary latents" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 359, + 129, + 545, + 143 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 129, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 359, + 129, + 545, + 143 + ], + "type": "interline_equation", + "content": "x _ {t - 1} ^ {\\operatorname {t m p}} \\sim \\mathcal {N} \\left(\\mu_ {\\theta} \\left(x _ {t} ^ {*}, t\\right), \\Sigma_ {\\theta} \\left(x _ {t} ^ {*}, t\\right)\\right) \\tag {7}", + "image_path": "f7027aa9965d8685a44552703ec749f41b3037323d62e6434446918f30548e1e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 152, + 545, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 152, + 545, + 176 + ], + "spans": [ + { + "bbox": [ + 305, + 152, + 545, + 176 + ], + "type": "text", + "content": "To obtain the final latents for this time step, we blend in the known latents according to" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 349, + 186, + 545, + 200 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 186, + 545, + 200 + ], + "spans": [ + { + "bbox": [ + 349, + 186, + 545, + 200 + ], + "type": "interline_equation", + "content": "x _ {t - 1} ^ {*} = (1 - \\mathbf {m}) \\odot \\tilde {x} _ {t - 1} + \\mathbf {m} \\odot x _ {t - 1} ^ {\\mathrm {t m p}} \\tag {8}", + "image_path": "9f151105efb29803f14c238ae55492724fb893079cf35a0b73bc1197244f7b40.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 209, + 462, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 209, + 462, + 220 + ], + "spans": [ + { + "bbox": [ + 306, + 209, + 462, + 220 + ], + "type": "text", + "content": "where the binary mask " + }, + { + "bbox": [ + 306, + 209, + 462, + 220 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 306, + 209, + 462, + 220 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 354, + 227, + 545, + 257 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 227, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 354, + 227, + 545, + 257 + ], + "type": "interline_equation", + "content": "\\mathbf {m} = \\left\\{ \\begin{array}{l l} \\mathbf {m} ^ {\\text {d y n a m i c}} & \\text {i f} t \\geq t _ {\\text {e n d}} \\\\ \\mathbf {m} ^ {\\text {d y n a m i c}} + \\mathbf {m} ^ {\\text {s t a t i c}} & \\text {i f} t < t _ {\\text {e n d}} \\end{array} \\right. \\tag {9}", + "image_path": "1ca1e55346ee05cd3bf414a8e9bba02e04ca68703c0850cb696060f72b5cd32a.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 266, + 546, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 546, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 546, + 396 + ], + "type": "text", + "content": "The mask, " + }, + { + "bbox": [ + 304, + 266, + 546, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^{\\mathrm{dynamic}}" + }, + { + "bbox": [ + 304, + 266, + 546, + 396 + ], + "type": "text", + "content": ", corresponds to the dynamic objects in the scene, and the mask " + }, + { + "bbox": [ + 304, + 266, + 546, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^{\\mathrm{static}}" + }, + { + "bbox": [ + 304, + 266, + 546, + 396 + ], + "type": "text", + "content": " corresponds to the static parts in the outpainted regions. This scheme is similar to ideas already explored in the context of image diffusion models [3, 34], but applied to video diffusion models. If we observe the changes in Figure 2, we can see that the pre-trained video diffusion model can function as a temporal filter, making discontinuous frames temporally coherent. This simplifies the inpainting task since we do not rely on expensive training of the video model and simply use it at inference time." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 408, + 387, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 408, + 387, + 422 + ], + "spans": [ + { + "bbox": [ + 306, + 408, + 387, + 422 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 428, + 437, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 428, + 437, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 428, + 437, + 441 + ], + "type": "text", + "content": "4.1. Implementation details" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 447, + 545, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 545, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 545, + 649 + ], + "type": "text", + "content": "Static Scene. In pre-processing, we estimate the camera poses of the source video [31, 60]. Our GS training process consists of " + }, + { + "bbox": [ + 304, + 447, + 545, + 649 + ], + "type": "inline_equation", + "content": "1 + K + 1" + }, + { + "bbox": [ + 304, + 447, + 545, + 649 + ], + "type": "text", + "content": " rounds in total, where " + }, + { + "bbox": [ + 304, + 447, + 545, + 649 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 447, + 545, + 649 + ], + "type": "text", + "content": " is the number of key frames we sample from the video. In the first training round, we train the GS model only on the static part of the video for 3000 iterations. Then we iterate between outpainting the " + }, + { + "bbox": [ + 304, + 447, + 545, + 649 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 447, + 545, + 649 + ], + "type": "text", + "content": " key frames and " + }, + { + "bbox": [ + 304, + 447, + 545, + 649 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 447, + 545, + 649 + ], + "type": "text", + "content": " rounds of GS training to update the outpainted regions into the GS model. To avoid overfitting, we employ a sliding window training strategy: following outpainting of the " + }, + { + "bbox": [ + 304, + 447, + 545, + 649 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 447, + 545, + 649 + ], + "type": "text", + "content": "-th frame, we train only on frames within the window " + }, + { + "bbox": [ + 304, + 447, + 545, + 649 + ], + "type": "inline_equation", + "content": "[i - 5, i + 5]" + }, + { + "bbox": [ + 304, + 447, + 545, + 649 + ], + "type": "text", + "content": " for 500 iterations. This window shifts progressively as outpainting and training continues. After outpainting all " + }, + { + "bbox": [ + 304, + 447, + 545, + 649 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 447, + 545, + 649 + ], + "type": "text", + "content": " key frames, we conduct a final round of training for refinement on all out-painted and original background frames for 1000 iterations. All training and outpainting can be completed on a single 4090 GPU." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Initial Object Inpainting. Before initial inpainting on the missing regions of dynamic objects, we fine-tune the SDXL model on the masked source video. The approach is similar to the one proposed by Tang et al. [45]. Masks are randomly" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7313" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 42, + 71, + 537, + 452 + ], + "blocks": [ + { + "bbox": [ + 42, + 71, + 537, + 452 + ], + "lines": [ + { + "bbox": [ + 42, + 71, + 537, + 452 + ], + "spans": [ + { + "bbox": [ + 42, + 71, + 537, + 452 + ], + "type": "image", + "image_path": "08f698742f583338b3b845b1861c522ccbc43d7e250713da970ba9a84f8fdb37.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 460, + 546, + 506 + ], + "lines": [ + { + "bbox": [ + 46, + 460, + 546, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 546, + 506 + ], + "type": "text", + "content": "Figure 5. Qualitative Comparison with State-of-the-art methods on the DAVIS dataset [36]. The area outside of the red lines (only shown for the first frame) is outpainted. We show the results for 3 frames from the each video sequence. Compared to MOTIA [48] and M3DDM[14], our results are of better quality and temporally stable. Please pay attention to the changes of the outpainted area between the frames of the same sequence." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 518, + 289, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 289, + 603 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 289, + 603 + ], + "type": "text", + "content": "generated for each input frame, with random distributed positions and mask ratios varying from 0.2 to 0.8. For each object, a prompt \"a photo of [v]\" is applied. During fine-tuning, a Low-Rank Adaption (LoRA) [20] with rank 8 is trained for both the text-encoder and UNet with a batch size of 4 and gradient accumulation steps of 4 for 2000 steps." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": "Video Diffusion Model. We use a pretrained image-to-video model [7] to complete object motion. The number of denoising steps is set to 25, with the level of noise added to the reconstructed frames generally set to 0.8 in our experiments. We apply a resampling strategy similar to Repaint [34] five times after each denoising step to improve the quality and stability of the objects. For videos where the object fully appears in the middle or last frame, we em" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 518, + 547, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 518, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 305, + 518, + 547, + 555 + ], + "type": "text", + "content": "ploy the Time-Reversal strategy from [15], conditioning the generation on additional frames to better retain the object identity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 563, + 414, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 414, + 576 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 414, + 576 + ], + "type": "text", + "content": "4.2. Experiment Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "Competitors. We compare our method with the following methods. (1) Dehan et al. [13] separate foreground and background components. They use estimated optical flow to warp the background, with missing regions filled in using an image completion network. (2) M3DDM [14] proposes a masked 3D Diffusion Model architecture, which is trained on WebVid [5] and a self-collected 5M e-commerce dataset for video outpaint tasks. (3) MOTIA [48] is the current state-of-the-art method for video outpainting. It is based on the ControlNet [58] inpainting model and the temporal module fromAnimateDiff [19]. It trains a LoRA [20] for" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7314" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 235, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 235, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 235, + 83 + ], + "type": "text", + "content": "each test sample to learn the intrinsic patterns." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 287, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 287, + 179 + ], + "type": "text", + "content": "Datasets. Following M3DDM [14] and MOTIA [48], we evaluate our method on the DAVIS [36] and YouTube-VOS [51] datasets. For the YouTube-VOS [51] dataset, due to its large number of videos and the lack of clarity in prior works on which videos were tested, we selected a high-quality subset and evaluated all methods on it. Consistent with prior works, we compare the horizontal outpainting results for each test video at mask ratios of 0.25 and 0.66." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 194, + 289, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 289, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 289, + 399 + ], + "type": "text", + "content": "Evaluation Metrics. The quality of outpainting is primarily evaluated in terms of visual quality, realism, and temporal consistency. For this purpose, we employ two well-established metrics: Fréchet Video Distance (FVD) [47] and flow warping error " + }, + { + "bbox": [ + 46, + 194, + 289, + 399 + ], + "type": "inline_equation", + "content": "(E_{\\mathrm{warp}})" + }, + { + "bbox": [ + 46, + 194, + 289, + 399 + ], + "type": "text", + "content": " [25]. FVD assesses the similarity between the distribution of the outpainted video and the original video, while flow warping error evaluates temporal consistency by measuring the error when outpainted frames are warped to adjacent frames using estimated optical flow. In addition, to align with previous works, we also report Peak Signal to Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) [59] between the outpainted and ground-truth videos, although we consider these metrics limited in capturing the quality of outpainting. We further conducted a user study to gather subjective evaluations of outpainting quality, which we detail in Sec. 4.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 405, + 170, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 405, + 170, + 417 + ], + "spans": [ + { + "bbox": [ + 47, + 405, + 170, + 417 + ], + "type": "text", + "content": "4.3. Evaluation & Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 423, + 289, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 289, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 289, + 604 + ], + "type": "text", + "content": "Qualitative Results. Fig. 5 and Fig. 6 illustrate the qualitative comparison of our method against other approaches on the DAVIS and YouTube-VOS datasets. M3DDM [14] frequently suffers from generation failures and produces blurred content, as shown in the first samples of Fig. 5 and Fig. 6. MOTIA [48], while fine-tuning on each test sample to capture input-specific patterns, often generates repetitive objects and background patterns. More importantly, both methods struggle to accurately capture contextual information across frames, resulting in inconsistency when outpainting elements seen in adjacent frames. In contrast, our method overcomes these issues, achieving better cross-frame consistency. Furthermore, our approach's object-specific completion approach prevents objects from duplicating or disappearing beyond the frame." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": "Generalization and multiple objects. Our strategy, differentiating between static and dynamic elements, can naturally extend to multiple dynamic objects by decomposing the frame into object layers based on their depth. Each object is processed separately using our pipeline and then seamlessly blended to produce the final result. Figure 7 shows that our method still performs best on such complex scenarios." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 299, + 72, + 544, + 260 + ], + "blocks": [ + { + "bbox": [ + 299, + 72, + 544, + 260 + ], + "lines": [ + { + "bbox": [ + 299, + 72, + 544, + 260 + ], + "spans": [ + { + "bbox": [ + 299, + 72, + 544, + 260 + ], + "type": "image", + "image_path": "4f7813febd35ba3dc97b6dd27df7e655f65b2258e25554056c6402fa69fe10d3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 270, + 546, + 348 + ], + "lines": [ + { + "bbox": [ + 304, + 270, + 546, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 546, + 348 + ], + "type": "text", + "content": "Figure 6. Qualitative Comparison with State-of-the-art methods on Youtube-VOS dataset [51]. The area outside of the red lines (only shown for the first frame) is outpainted. We show the results for 3 frames from the each video sequence. Compared to Dehan [13], MOTIA [48] and M3DDM[14], our results are of better quality and temporally stable. Please pay attention to the changes of the outpainted area between the frames of the same sequence." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 298, + 361, + 545, + 495 + ], + "blocks": [ + { + "bbox": [ + 298, + 361, + 545, + 495 + ], + "lines": [ + { + "bbox": [ + 298, + 361, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 298, + 361, + 545, + 495 + ], + "type": "image", + "image_path": "2e853b0c067821d1f93277b2e0bcd86eebf05e0d9e85f313ae569af3879a75eb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 506, + 545, + 540 + ], + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 540 + ], + "type": "text", + "content": "Figure 7. Qualitative Comparison with State-of-the-art methods on challenging scenarios (multiple dynamic objects, complex static regions, etc.)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 553, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 553, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 553, + 545, + 673 + ], + "type": "text", + "content": "Quantitative Results. Tab. 1 presents a quantitative comparison of our method with other approaches. Our method outperforms the current state-of-the-art in both FVD and " + }, + { + "bbox": [ + 304, + 553, + 545, + 673 + ], + "type": "inline_equation", + "content": "E_{\\text{warp}}" + }, + { + "bbox": [ + 304, + 553, + 545, + 673 + ], + "type": "text", + "content": ", achieving improvements of " + }, + { + "bbox": [ + 304, + 553, + 545, + 673 + ], + "type": "inline_equation", + "content": "18.8\\%" + }, + { + "bbox": [ + 304, + 553, + 545, + 673 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 553, + 545, + 673 + ], + "type": "inline_equation", + "content": "11.7\\%" + }, + { + "bbox": [ + 304, + 553, + 545, + 673 + ], + "type": "text", + "content": " on the DAVIS dataset and " + }, + { + "bbox": [ + 304, + 553, + 545, + 673 + ], + "type": "inline_equation", + "content": "26.9\\%" + }, + { + "bbox": [ + 304, + 553, + 545, + 673 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 553, + 545, + 673 + ], + "type": "inline_equation", + "content": "33.3\\%" + }, + { + "bbox": [ + 304, + 553, + 545, + 673 + ], + "type": "text", + "content": " on the YouTube-VOS dataset, respectively. We also achieve the highest PSNR and SSIM scores on both datasets. For LPIPS, while our method is the second behind the number reported in MOTIA [48] on the DAVIS dataset, it surpasses the score obtained from using their public implementation and provided model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Ablation Study We conducted ablation studies for the different components of our method. Quantitative evalua" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7315" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 70, + 542, + 173 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 542, + 173 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 542, + 173 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 542, + 173 + ], + "type": "table", + "html": "
DAVISYouTube-VOS
PSNR↑SSIM↑LPIPS↓FVD↓E_warp↓PSNR↑SSIM↑LPIPS↓FVD↓E_warp↓
Dehan [13]17.960.6270.233363.10.03018.340.7610.287353.60.043
M3DDM [14]20.260.7080.204300.00.01718.330.7250.424283.10.024
MOTIA (model) [48]18.250.6990.282294.80.02118.540.7900.250156.40.027
MOTIA (paper) [48]20.360.7580.160286.3------
Ours22.620.8060.200232.50.01519.880.8320.236114.20.016
", + "image_path": "68141a09f0fd4cf9b1baa109b167e8ab7b0fca9f3bce894513b55ef60abf1188.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 232, + 265, + 318 + ], + "blocks": [ + { + "bbox": [ + 46, + 180, + 547, + 215 + ], + "lines": [ + { + "bbox": [ + 46, + 180, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 547, + 215 + ], + "type": "text", + "content": "Table 1. Quantitative Comparison with State-of-the-art methods on DAVIS dataset [36] and YouTube-VOS [51]. Our method achieves the best performance overall. In the case of MOTIA [48], we have 2 rows as we report the evaluation using the model and code available online(model) and the numbers from the original paper (paper). See the text for details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 232, + 265, + 318 + ], + "lines": [ + { + "bbox": [ + 69, + 232, + 265, + 318 + ], + "spans": [ + { + "bbox": [ + 69, + 232, + 265, + 318 + ], + "type": "table", + "html": "
PSNR↑SSIM↑LPIPS↓FVD↓E_warp↓
(a)18.150.6940.248448.80.032
(b)18.050.7160.274323.20.022
(c)20.680.7720.218293.30.020
(d)20.050.7760.223319.80.017
(e)22.620.8060.200232.50.015
", + "image_path": "e2f132b6f53fd63eadb810ca1adfe2159bf50c7f32fababcedbe92bec2c4ec2f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 326, + 288, + 370 + ], + "lines": [ + { + "bbox": [ + 46, + 326, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 288, + 370 + ], + "type": "text", + "content": "Table 2. Quantitative ablation results. (a). Outpainting with fine-tuned Image Model; (b). (a) + Guided Video Synthesis; (c). Our model without guided video synthesis; (d). Our model without the initialization from object inpainting; (e). Our complete model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 385, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 287, + 564 + ], + "type": "text", + "content": "tion is shown in Table 2, while the visual results are provided in supplementary material. When only using single image outpainting (SDXL fine-tuned on content [45]), the images achieve good realism but lack temporal continuity. When applying video diffusion guided synthesis to these outpainted frames directly, the temporal consistency within the outpainted areas improves individually. However, coherence between the outpainted regions and the original video content remains weak. This issue is difficult to resolve solely with Video Diffusion Model priors. When using the static region outpainting (assisted with GS) and either the image inpainting or video diffusion alone are applied for object completion, the objects suffer from significant temporal variations in appearance. Using all the components achieves the best results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "User Study. Given the subjective nature of evaluating outpainting quality, we conducted a user study to compare our method with competing approaches. Users were asked to evaluate each method across several dimensions, including realism (such as whether the outpainted results look natural and harmonious), temporal consistency (including background consistency, object motion consistency, etc.), and overall visual quality (including color fidelity, smoothness of boundaries, blurriness etc.). The results, shown in Fig. 8, indicate that our method was clearly preferred by users compared to others, achieving more than " + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": " of the votes in all dimensions. Further details of the user study can" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 225, + 542, + 308 + ], + "blocks": [ + { + "bbox": [ + 310, + 225, + 542, + 308 + ], + "lines": [ + { + "bbox": [ + 310, + 225, + 542, + 308 + ], + "spans": [ + { + "bbox": [ + 310, + 225, + 542, + 308 + ], + "type": "image", + "image_path": "41c904528a4bc15d586ad920aa1283aa241aca5003efce01f59358301a495c4a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 309, + 546, + 342 + ], + "lines": [ + { + "bbox": [ + 305, + 309, + 546, + 342 + ], + "spans": [ + { + "bbox": [ + 305, + 309, + 546, + 342 + ], + "type": "text", + "content": "Figure 8. The result of the user study includes 619 votes from 37 participants. Our method is preferred by more than " + }, + { + "bbox": [ + 305, + 309, + 546, + 342 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 305, + 309, + 546, + 342 + ], + "type": "text", + "content": " of the votes in all dimensions." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 353, + 545, + 444 + ], + "blocks": [ + { + "bbox": [ + 306, + 353, + 545, + 444 + ], + "lines": [ + { + "bbox": [ + 306, + 353, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 306, + 353, + 545, + 444 + ], + "type": "image", + "image_path": "6b4dda42f6c045166c29ae37338ef75bf1fbcc304b49f64a015ea09bfa96a043.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 452, + 547, + 485 + ], + "lines": [ + { + "bbox": [ + 305, + 452, + 547, + 485 + ], + "spans": [ + { + "bbox": [ + 305, + 452, + 547, + 485 + ], + "type": "text", + "content": "Figure 9. Challenges with dynamic objects moving in and out of the original frame: Our current solution of tracking and extrapolating the trajectory of the bounding box can be improved." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 491, + 471, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 471, + 503 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 471, + 503 + ], + "type": "text", + "content": "be found in the supplementary materials." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 514, + 380, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 380, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 380, + 525 + ], + "type": "text", + "content": "5. Discussions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 533, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 715 + ], + "type": "text", + "content": "Building on Gaussian Splatting and a Video Diffusion Model, we propose a new state-of-the-art method for video outpainting. Despite our strong results, some issues remain, especially around objects moving in and out of the original frame. An example is shown in Fig. 9. Currently tracking and extrapolating the trajectory of the bounding box offers a solution, but limitations in the single image inpainting negatively impact the quality of the final result. Another area of improvement is the outpainting involving multiple occluded dynamic objects. Our current solution of inpainting each object separately and blending them based on the estimated depth is limited. Exploring the use of more advanced 4D dynamic Gaussian representations jointly with video models offers a promising future direction towards higher and higher quality video outpainting." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7316" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Amit Aides, Tamar Avraham, and Yoav Y Schechner. Multiscale ultrawide foveated video extrapolation. In 2011 IEEE International Conference on Computational Photography (ICCP), pages 1-8. IEEE, 2011. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 287, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 287, + 168 + ], + "type": "text", + "content": "[2] Tamar Avraham and Yoav Y Schechner. Ultrawide foveated video extrapolation. IEEE Journal of Selected Topics in Signal Processing, 5(2):321-334, 2010. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 288, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 288, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 288, + 223 + ], + "type": "text", + "content": "[3] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18208-18218, 2022. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 288, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 288, + 269 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 288, + 269 + ], + "type": "text", + "content": "[4] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18208-18218, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 269, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 288, + 312 + ], + "type": "text", + "content": "[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 313, + 216, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 216, + 324 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 216, + 324 + ], + "type": "text", + "content": "[6] M Bertalmio. Image inpainting, 2000. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 325, + 288, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 325, + 288, + 380 + ], + "spans": [ + { + "bbox": [ + 53, + 325, + 288, + 380 + ], + "type": "text", + "content": "[7] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 380, + 288, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 445 + ], + "type": "text", + "content": "[8] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 447, + 288, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 489 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 489 + ], + "type": "text", + "content": "[9] Wen-Hsuan Chu, Lei Ke, and Katerina Fragkiadaki. Dreamscene4d: Dynamic multi-object scene generation from monocular videos. arXiv preprint arXiv:2405.02280, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 491, + 288, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 288, + 534 + ], + "type": "text", + "content": "[10] Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 536, + 288, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 288, + 590 + ], + "type": "text", + "content": "[11] Ciprian Corneanu, Raghudeep Gadde, and Aleix M Martínez. Latentpaint: Image inpainting in latent space with diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4334-4343, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 591, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 288, + 635 + ], + "type": "text", + "content": "[12] Antonio Criminisi, Patrick Pérez, and Kentaro Toyama. Region filling and object removal by exemplar-based image inpainting. IEEE Transactions on image processing, 13(9): 1200-1212, 2004. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "text", + "content": "[13] Loic Dehan, Wiebe Van Ranst, Patrick Vandewalle, and Toon Goedemé. Complete and temporally consistent video outpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 687-695, 2022. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "type": "text", + "content": "[14] Fanda Fan, Chaoxu Guo, Litong Gong, Biao Wang, Tiezheng Ge, Yuning Jiang, Chunjie Luo, and Jianfeng Zhan. Hierar-" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "chical masked 3d diffusion model for video outpainting. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7890-7900, 2023. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 106, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 545, + 160 + ], + "type": "text", + "content": "[15] Haiwen Feng, Zheng Ding, Zhihao Xia, Simon Niklaus, Victoria Abrevaya, Michael J Black, and Xuaner Zhang. Explorative inbetweening of time and space. In European Conference on Computer Vision, pages 378-395. Springer, 2025. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 161, + 545, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 161, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 308, + 161, + 545, + 204 + ], + "type": "text", + "content": "[16] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 205, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 205, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 205, + 545, + 249 + ], + "type": "text", + "content": "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 250, + 545, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 294 + ], + "type": "text", + "content": "[18] Qiang Guo, Shanshan Gao, Xiaofeng Zhang, Yilong Yin, and Caiming Zhang. Patch-based image inpainting via two-stage low rank approximation. IEEE transactions on visualization and computer graphics, 24(6):2023-2036, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 294, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 348 + ], + "type": "text", + "content": "[19] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 349, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 349, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 349, + 545, + 392 + ], + "type": "text", + "content": "[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 393, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 393, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 393, + 545, + 436 + ], + "type": "text", + "content": "[21] Jia-Bin Huang, Sing Bing Kang, Narendra Ahuja, and Johannes Kopf. Image completion using planar structure guidance. ACM Transactions on graphics (TOG), 33(4):1-10, 2014. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 437, + 545, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 481 + ], + "type": "text", + "content": "[22] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. arXiv preprint arXiv:2403.06976, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "type": "text", + "content": "[23] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 525, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 545, + 581 + ], + "type": "text", + "content": "[24] Naoki Kimura and Jun Rekimoto. Extvision: augmentation of visual experiences with generation of context images for a peripheral vision using deep neural network. In Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems, pages 1-10, 2018. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "text", + "content": "[25] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In European Conference on Computer Vision, 2018. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 625, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 667 + ], + "type": "text", + "content": "[26] Minhyeok Lee, Suhwan Cho, Chajin Shin, Jungho Lee, Sunghun Yang, and Sangyoun Lee. Video diffusion models are strong video inpainter. arXiv preprint arXiv:2408.11402, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "text", + "content": "[27] Yao-Chih Lee, Yi-Ting Chen, Andrew Wang, Ting-Hsuan Liao, Brandon Y Feng, and Jia-Bin Huang. Vividdream: Generating 3d scene with ambient dynamics. arXiv preprint arXiv:2405.20334, 2024. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7317" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[28] Yao-Chih Lee, Zhoutong Zhang, Kevin Blackburn-Matzen, Simon Niklaus, Jianming Zhang, Jia-Bin Huang, and Feng Liu. Fast view synthesis of casual videos with soup-of-planes. In European Conference on Computer Vision, pages 278–296. Springer, 2025. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "type": "text", + "content": "[29] Jiahui Lei, Yijia Weng, Adam Harley, Leonidas Guibas, and Kostas Daniilidis. Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds. arXiv preprint arXiv:2405.17421, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 288, + 229 + ], + "type": "text", + "content": "[30] Renjie Li, Panwang Pan, Bangbang Yang, Dejia Xu, Shijie Zhou, Xuanyang Zhang, Zeming Li, Achuta Kadambi, Zhangyang Wang, and Zhiwen Fan. 4k4dgen: Panoramic 4d generation at 4k resolution. arXiv preprint arXiv:2406.13527, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 288, + 274 + ], + "type": "text", + "content": "[31] Zhengqi Li, Richard Tucker, Forrester Cole, Qianqian Wang, Linyi Jin, Vickie Ye, Angjoo Kanazawa, Aleksander Holynski, and Noah Snively. Megasam: Accurate, fast, and robust structure and motion from casual dynamic videos, 2024. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 276, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 288, + 330 + ], + "type": "text", + "content": "[32] Guilin Liu, Fitsum A Reda, Kevin J Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In Proceedings of the European conference on computer vision (ECCV), pages 85-100, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 331, + 288, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 288, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 288, + 386 + ], + "type": "text", + "content": "[33] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 388, + 288, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 288, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 288, + 442 + ], + "type": "text", + "content": "[34] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 2, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 444, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 288, + 487 + ], + "type": "text", + "content": "[35] Hao Ouyang, Kathryn Heal, Stephen Lombardi, and Tiancheng Sun. Text2immersion: Generative immersive scene with 3d gaussians. arXiv preprint arXiv:2312.09242, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 488, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 488, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 488, + 288, + 533 + ], + "type": "text", + "content": "[36] F. Perazzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Computer Vision and Pattern Recognition, 2016. 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 534, + 288, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 288, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 288, + 588 + ], + "type": "text", + "content": "[37] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. UniDepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 590, + 288, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 288, + 644 + ], + "type": "text", + "content": "[38] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "text", + "content": "[39] Weize Quan, Jiaxi Chen, Yanli Liu, Dong-Ming Yan, and Peter Wonka. Deep learning-based image and video inpainting: A survey. International Journal of Computer Vision, 132(7): 2367-2400, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "text", + "content": "[40] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gener" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "ation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 547, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 152 + ], + "type": "text", + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 547, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 547, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 547, + 186 + ], + "type": "text", + "content": "[42] Iizuka Satoshi, Simo-Serra Edgar, and Ishikawa Hiroshi. Globally and locally consistent image completion. ACM Trans. Graph, 36(4):3073659, 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 186, + 547, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 547, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 547, + 230 + ], + "type": "text", + "content": "[43] Jaidev Shriram, Alex Trevithick, Lingjie Liu, and Ravi Ramamoorthi. Realmdreamer: Text-driven 3d scene generation with inpainting and depth diffusion. arXiv preprint arXiv:2404.07199, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 232, + 547, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 232, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 547, + 308 + ], + "type": "text", + "content": "[44] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Armenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2149–2159, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 309, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 374 + ], + "type": "text", + "content": "[45] Luming Tang, Nataniel Ruiz, Qinghao Chu, Yuanzhen Li, Aleksander Holynski, David E Jacobs, Bharath Hariharan, Yael Pritch, Neal Wadhwa, Kfir Aberman, et al. Realfill: Reference-driven generation for authentic image completion. ACM Transactions on Graphics (TOG), 43(4):1-12, 2024. 2, 4, 5, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 376, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 421 + ], + "type": "text", + "content": "[46] Laura Turban, Fabrice Urban, and Philippe Guillotel. Extrafoveal video extension for an immersive viewing experience. IEEE transactions on visualization and computer graphics, 23(5):1520-1533, 2016. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 422, + 545, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 466 + ], + "type": "text", + "content": "[47] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 467, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 545, + 521 + ], + "type": "text", + "content": "[48] Fu-Yun Wang, Xiaoshi Wu, Zhaoyang Huang, Xiaoyu Shi, Dazhong Shen, Guanglu Song, Yu Liu, and Hongsheng Li. Be-your-outpainter: Mastering video outpainting through input-specific adaptation. arXiv preprint arXiv:2403.13745, 2024. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 522, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 567 + ], + "type": "text", + "content": "[49] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 568, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 601 + ], + "type": "text", + "content": "[50] Shizun Wang, Xingyi Yang, Qiuhong Shen, Zhenxiang Jiang, and Xinchao Wang. Gflow: Recovering 4d world from monocular video. arXiv preprint arXiv:2405.18426, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "type": "text", + "content": "[51] Ning Xu, Linjie Yang, Yuchen Fan, Dingcheng Yue, Yuchen Liang, Jianchao Yang, and Thomas S. Huang. Youtube-vos: A large-scale video object segmentation benchmark. CoRR, abs/1809.03327, 2018. 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "text", + "content": "[52] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18381-18391, 2023. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7318" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 687 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[53] Zhenqiang Ying and Alan Bovik. 180-degree outpainting from a single image. arXiv preprint arXiv:2001.04568, 2020. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "type": "text", + "content": "[54] Hong-Xing Yu, Haoyi Duan, Charles Herrmann, William T Freeman, and Jiajun Wu. Wonderworld: Interactive 3d scene generation from a single image. arXiv preprint arXiv:2406.09394, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 217 + ], + "type": "text", + "content": "[55] Hong-Xing Yu, Haoyi Duan, Junhwa Hur, Kyle Sargent, Michael Rubinstein, William T Freeman, Forrester Cole, Deqing Sun, Noah Snavely, Jiajun Wu, et al. Wonderjourney: Going from anywhere to everywhere. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6658-6667, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 272 + ], + "type": "text", + "content": "[56] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contextual attention. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5505-5514, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 274, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 274, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 274, + 287, + 319 + ], + "type": "text", + "content": "[57] Yanhong Zeng, Jianlong Fu, Hongyang Chao, and Baining Guo. Aggregated contextual transformations for high-resolution image inpainting. IEEE Transactions on Visualization and Computer Graphics, 29(7):3266-3280, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 319, + 287, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 287, + 363 + ], + "type": "text", + "content": "[58] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 364, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 364, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 48, + 364, + 287, + 418 + ], + "type": "text", + "content": "[59] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 420, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 420, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 420, + 287, + 464 + ], + "type": "text", + "content": "[60] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T Freeman. Structure and motion from casual videos. In European Conference on Computer Vision, pages 20-37. Springer, 2022. 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 465, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 465, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 465, + 287, + 529 + ], + "type": "text", + "content": "[61] Zhixing Zhang, Bichen Wu, Xiaoyan Wang, Yaqiao Luo, Luxin Zhang, Yinan Zhao, Peter Vajda, Dimitris Metaxas, and Licheng Yu. Avid: Any-length video inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7162-7172, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 532, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 532, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 532, + 287, + 597 + ], + "type": "text", + "content": "[62] Haitian Zheng, Zhe Lin, Jingwan Lu, Scott Cohen, Eli Shechtman, Connelly Barnes, Jianming Zhang, Ning Xu, Sohrab Amirghodsi, and Jiebo Luo. Image inpainting with cascaded modulation gan and object-aware training. In European Conference on Computer Vision, pages 277-296. Springer, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 598, + 287, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 598, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 48, + 598, + 287, + 653 + ], + "type": "text", + "content": "[63] Shijie Zhou, Zhiwen Fan, Dejia Xu, Haoran Chang, Pradyumna Chari, Suya Bharadwaj, Tejas You, Zhangyang Wang, and Achuta Kadambi. Dreamscene360: Unconstrained text-to-3d scene generation with panoramic gaussian splatting. arXiv preprint arXiv:2404.06903, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 654, + 287, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 654, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 48, + 654, + 287, + 687 + ], + "type": "text", + "content": "[64] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. arXiv preprint arXiv:2312.00451, 2023. 4" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7319" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Uncertain Multimodal Intention and Emotion Understanding in the Wild/50195041-da23-4b5c-8721-3858d2b20dc9_content_list.json b/2025/Uncertain Multimodal Intention and Emotion Understanding in the Wild/50195041-da23-4b5c-8721-3858d2b20dc9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a52401e9997816ff6e34477a3f8b0b72814648c6 --- /dev/null +++ b/2025/Uncertain Multimodal Intention and Emotion Understanding in the Wild/50195041-da23-4b5c-8721-3858d2b20dc9_content_list.json @@ -0,0 +1,1476 @@ +[ + { + "type": "text", + "text": "Uncertain Multimodal Intention and Emotion Understanding in the Wild", + "text_level": 1, + "bbox": [ + 129, + 130, + 867, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qu Yang $^{1}$ , Qinghongya Shi $^{1}$ , Tongxin Wang $^{1}$ and Mang Ye $^{1,2}$", + "bbox": [ + 241, + 178, + 751, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science, Wuhan University, Wuhan, China", + "bbox": [ + 241, + 199, + 754, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Taikang Center for Life and Medical Sciences, Wuhan University, Wuhan, China", + "bbox": [ + 173, + 215, + 823, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yangqu, shiqinghongya, wangtx, yemang}@whu.edu.cn", + "bbox": [ + 277, + 236, + 720, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/yan9qu/CVPR25-MINE", + "bbox": [ + 328, + 253, + 660, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 303, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding intention and emotion from social media poses unique challenges due to the inherent uncertainty in multimodal data, where posts often contain incomplete or missing modalities. While this uncertainty reflects real-world scenarios, it remains underexplored within the computer vision community, particularly in conjunction with the intrinsic relationship between emotion and intention. To address these challenges, we introduce the Multimodal IntentioN and Emotion Understanding in the Wild (MINE) dataset, comprising over 20,000 topic-specific social media posts with natural modality variations across text, image, video, and audio. MINE is distinctively constructed to capture both the uncertain nature of multimodal data and the implicit correlations between intentions and emotions, providing extensive annotations for both aspects. To tackle these scenarios, we propose the Bridging Emotion-Intention via Implicit Label Reasoning (BEAR) framework. BEAR consists of two key components: a BEIFormer that leverages emotion-intention correlations, and a Modality Asynchronous Prompt that handles modality uncertainty. Experiments show that BEAR outperforms existing methods in processing uncertain multimodal data while effectively mining emotion-intention relationships for social media content understanding.", + "bbox": [ + 89, + 335, + 483, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 729, + 220, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Intention and emotion are fundamental in shaping human social behaviors and decision-making processes [25, 32, 38]. Understanding their interplay is crucial for the advancement of artificial general intelligence. While prior studies [6, 42, 43] in the vision community have independently explored intention or emotion recognition in multimodal contexts (as illustrated in Fig. 1), employing datasets like MIntRec [47] and IEMOCAP [2], these datasets fail", + "bbox": [ + 89, + 755, + 483, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/749a53af190299b99909b56b3b41296da0364779ff4ab30992111945795123fa.jpg", + "image_caption": [ + "Figure 1. Overview of multi-modal emotion and intention understanding from social media platforms." + ], + "image_footnote": [], + "bbox": [ + 519, + 304, + 901, + 486 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to capture the inherent link between intention and emotion and overlook the challenges of uncertain modalities and thematic diversity in real-world scenarios. Addressing these gaps is vital for a comprehensive understanding of intentions and emotions in natural environments. In this paper, we introduce the Multimodal Intention and EmotioN Understanding in the Wild (MINE) dataset. MINE provides detailed annotations of both intention and emotion, establishing a platform to explore their correlation. It presents the challenges of uncertain modalities, thus mirroring the intricacies of real-life social interactions, as shown in Fig. 2.", + "bbox": [ + 511, + 534, + 906, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The design of MINE is guided by several key principles: 1) MINE addresses a critical gap in existing datasets by enabling comprehensive insights into the interactions between intention and emotion within social contexts. 2) In contrast to data typically curated under controlled conditions (such as scripted dialogues or TV series clips [2, 30, 35, 44, 46, 47]), which exhibit significant disparities with real-world situations, MINE leverages authentic user-generated content from social platforms, thus mirroring real-world dynamics of intention and emotion expression. 3) Unlike related datasets that typically assume fixed modalities [2, 47], MINE incorporates the challenge of uncertain modalities in social media, enhancing the realism and applicability of", + "bbox": [ + 511, + 703, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding author.", + "bbox": [ + 112, + 886, + 251, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24700", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e537ae4431384b95c75fc6e28318bcc76935fd041d1a6f5107a7e4369c3ffea3.jpg", + "image_caption": [ + "Figure 2. Overview of the MINE dataset, illustrating the multimodal data collection from social media, categorized by topics and including text, image, video, and audio content, each reflecting various intentions and emotions. The modality combination is uncertain." + ], + "image_footnote": [], + "bbox": [ + 181, + 93, + 823, + 250 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "intention and emotion understanding tasks. At the algorithmic level, this manifests as the missing modality problem, where modality heterogeneity is primarily reflected in the absence of certain modal inputs. Previous research on missing modalities [3, 50] often relies on artificial sampling methods, creating biases and noise due to incomplete information (e.g., an image showing “news about something” with the text “That’s so ironic!”, where removing text leads to misinterpretation). MINE, however, preserves native modal information, eliminating potential errors and offering a reliable benchmark for uncertain modality research. 4) By utilizing topical tags (such as #CAMPUS) from social media, MINE provides objective context for analyzing intentions and emotions, enabling a more nuanced understanding of these elements within specific thematic frames.", + "bbox": [ + 88, + 289, + 485, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To harness the intricate relationship between emotions and intentions within the MINE dataset, we propose Bridging Emotion-Intention via Implicit Label Reasoning (BEAR) framework, comprising two key components: BEIFormer and Modality Asynchronous Prompt (MAP). The BEIFormer introduces an innovative contrastive task, utilizing [MASK] tokens to conceal intention and emotion labels within training prompts. This design enables the model to reconstruct textual data that is in alignment with ground truth labels, leveraging multimodal inputs and the process of unmasking text. This implicit reconstruction and alignment process in BEAR enhances the understanding by delving into the relationship between emotion and intention. Furthermore, the MAP is specifically engineered to tackle the challenge of uncertain modalities. It provides customized cues tailored for diverse uncertain scenarios, thereby substantially improving understanding capacity in contexts where modality data may be incomplete.", + "bbox": [ + 89, + 518, + 483, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contribution is threefold:", + "bbox": [ + 109, + 792, + 390, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(1) We introduce MINE, the first large-scale dataset that captures the inherent relationship between intention and emotion. Comprising over 20,000 topic-specific posts with natural modality distribution, MINE uniquely provides accurate annotations for intention and emotion, enabling systematic study of their correlations in realistic scenarios.", + "bbox": [ + 89, + 809, + 485, + 901 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(2) We propose Bridging Emotion-Intention via Implicit Label Reasoning, a novel framework that bridges emotion and intention understanding. It incorporates two key components: (i) BEIFormer that implicitly mines emotion-intention relationships through contrastive learning, and (ii) Modality Asynchronous Prompt that handles uncertain modalities without compromising existing features.", + "(3) We establish comprehensive benchmarks on MINE by evaluating multiple baselines. Through extensive experiments, we demonstrate that BEAR consistently outperforms existing methods by significant margins (e.g., $+2.65\\%$ Macro F1 for intention recognition and $+3.54\\%$ accuracy for emotion recognition), validating the importance of joint emotion-intention modeling." + ], + "bbox": [ + 511, + 289, + 908, + 506 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 535, + 655, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Intention and Emotion Understanding", + "text_level": 1, + "bbox": [ + 511, + 565, + 844, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Intention recognition is to understand the goals behind human actions. While early datasets like ATIS [9, 37] were limited to specific domains, recent datasets like CLINC150[15], Intentonomy [11], and MIntRec [47, 48] have expanded to diverse sources including social media and TV series.", + "bbox": [ + 511, + 593, + 906, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For emotion recognition, datasets have evolved from controlled environments (IEMOCAP [2]) to more naturalistic settings with CMU-MOSEI [46], PanoSent [21], and MELD [26], encompassing multimodal expressions across text, speech, and visual cues. However, as shown in Tab. 1, existing datasets typically focus on either intention or emotion in isolation. While recent works such as EMO-TyDA [31], MC-EIU [17], and EmoInt-MD [35] attempt to combine both aspects, they are constrained by controlled settings and fixed modalities, unlike real-world social media posts which often contain varying combinations of modalities. Our MINE dataset addresses these limitations by capturing authentic emotion-intention relationships while preserving the natural complexity of missing modalities.", + "bbox": [ + 511, + 689, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24701", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/9b031e4cb4f32ab34cbd179a2ce223f7ed830e914ee2bb939280ec93be1133f3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypeDatasetModalityMissingClass NumsData NumsSource
####Emo.Int.
IntentionCLINC150 [15]✓✗✗✗✗-15023,700Acted
HWU64 [18]✓✗✗✗✗-6425,716Acted
Intentonomy [11]✓✓✗✗✗-2814,455In the wild
MDID [14]✓✓✗✗✗-81,299In the wild
MSED [10]✓✓✗✗✗-69,190In the wild
MIntRec [47]✓✗✓✓✗-202,224TV
MIntRec2.0 [48]✓✗✓✓✗-3015,040TV
EmotionIEMOCAP [2]✓✗✓✓✗6-7,433Acted
CMU-MOSEI [46]✓✗✓✓✗6-23,453In the wild
GoEmotions [4]✓✗✗✗✗28-58,000In the wild
EmotionRoI [24]✓✗✗✗✗6-1,980In the wild
MELD [26]✓✗✓✓✗8-13,708TV
AllEMOTyDA [31]✓✗✓✓✗101219,365Acted
MINE✓✓✓✓✓112020,168In the wild
", + "bbox": [ + 178, + 88, + 821, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Comparison of existing datasets for intention (Int), emotion (Emo), and combined (All) understanding. MINE uniquely addresses uncertain modalities while combining both emotion and intention analysis in real-world scenarios.", + "bbox": [ + 89, + 349, + 906, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Uncertain Modality in the Wild", + "text_level": 1, + "bbox": [ + 89, + 402, + 369, + 419 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Uncertain modality (or missing modality) presents a fundamental challenge in multimodal understanding. Contemporary studies [3, 16, 22, 23, 50] have investigated this challenge using datasets like IEMOCAP [2] and CMUMOSEI [46]. Researchers have proposed various approaches, ranging from automated fusion strategy identification [22] and meta-sampling integration [3] to unified frameworks for modality inference [29, 50] and Bayesian meta-learning applications [23].", + "bbox": [ + 88, + 425, + 482, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, these methods typically rely on random sampling to simulate missing modalities, which can introduce bias and alter the original content meaning (refer to Appendix G). In contrast, MINE captures naturally occurring uncertain modalities in the wild, providing a more authentic foundation for multimodal research.", + "bbox": [ + 89, + 561, + 482, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. MINE Dataset", + "text_level": 1, + "bbox": [ + 89, + 664, + 236, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we outline the creation of the MINE dataset, including steps such as data preparation, establishing categories for intentions and emotions, annotating multimodal content with these categories, and analyzing the data. Detailed metrics of the MINE dataset are provided in Tab. 2.", + "bbox": [ + 89, + 689, + 482, + 765 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Topic Definition and Data Preparation", + "text_level": 1, + "bbox": [ + 89, + 772, + 421, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To explore how topics influence the understanding of intentions and emotions in real-world scenarios, we obtained permission from the consent of data subjects and collected high-quality, multimodal data from social media platforms—venues teeming with genuine user expressions. This approach significantly diverges from the artificial confines of laboratory settings or the limited diversity found", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in sources like movie clips, frequently utilized in prior research. Our strategy involved organizing a wide range of multimodal data via clusters based on topic-specific hashtags, ensuring a rich and varied data source.", + "bbox": [ + 511, + 404, + 906, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The MINE dataset comprehensively covers ten topics that are integral to daily life. Following the methodology of Kapoor et al. [12], we selected these topics from the top 100 most frequently discussed subjects on social media platforms. MINE explores these topics in depth, paying particular attention to intersections identified through related keywords. For instance, the keyword 'Education' encompasses associated themes such as 'Social Media Impact on Teachers and Students' and 'Use of Social Media in Academic Learning'. To ensure thematic diversity, we selected ten distinct topics for our study and developed a comprehensive keyword framework for each one, with detailed documentation available in Appendix A. For data collection, we employed a hashtag-based method (e.g., #Education) rather than general topic searches to capture a broader variety of content within each topic.", + "bbox": [ + 511, + 470, + 908, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The data collection spans Twitter posts from 2019 to 2022, with all content obtained with explicit user consent. To ensure data quality, we implemented multiple preprocessing steps, including the removal of nonsensical texts (those consisting solely of links or random characters) and the filtering of harmful content (related to pornography and drugs). While visual content plays a crucial role in conveying emotional and intentional nuances, we maintained a deliberate balance with textual content to ensure comprehensive multimodal analysis. The final dataset composition reflects this approach, with visual content constituting approximately $70\\%$ of the data, creating a diverse yet imbal", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24702", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/210fa498550e476eb090992ad03b9dd3a6dca2bd9912c30460c64d409de358f8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ItemNumber
Number of intentions20
Number of emotions11
Number of text20044
Number of images12428
Number of videos3895
Number of audios1615
Average length of text utterances (words)20.0
Total number of words489069
Average duration of videos (seconds)24.4
Total duration of videos (hours)21.9
", + "bbox": [ + 130, + 88, + 467, + 256 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/30be5adc7b0e4a197dc632727ffb8a7789550089f8ee9fa612b3f8bfce119a1f.jpg", + "table_caption": [ + "(a) Data Statistics" + ], + "table_footnote": [], + "table_body": "
TextImageVideoAudioNumber
✓3913
✓99
✓4
✓✓12261
✓✓2248
✓✓3
✓✓18
✓✓✓26
✓✓✓1557
✓✓✓✓39
", + "bbox": [ + 560, + 88, + 846, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(b) Distribution of Modality Combinations", + "bbox": [ + 589, + 258, + 816, + 271 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2. Data and Modality Statistics of the MINE dataset.", + "bbox": [ + 321, + 273, + 671, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "anced multimodal distribution, as detailed in Tab. 2.", + "bbox": [ + 89, + 297, + 433, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Taxonomy Definition and Data Annotation", + "text_level": 1, + "bbox": [ + 89, + 319, + 454, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Defining labels for intentions and emotions presents a unique challenge due to their subjective nature. While previous work [47] treated intentions as mutually exclusive categories ('Expressing emotions' or 'Achieving goals'), we argue that social interactions naturally blend both elements, aligning with philosophical theories [32].", + "bbox": [ + 89, + 340, + 483, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MINE incorporates 20 intention and 11 emotion labels per data point, drawing from literature on both dialogue and visual intention analysis [11, 47] and emotion research [2, 34]. We frame intention recognition as a multi-label task to capture the nuanced nature of social media interactions, while structuring emotion recognition as a single-label task to identify dominant emotions, consistent with established practices [2]. To address inherent subjectivity, our annotation employs a crowd-sourcing platform with dual annotations. Detailed category definitions and annotation guidelines are provided in Appendices A and H.", + "bbox": [ + 89, + 431, + 483, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The annotation workflow involved two independent groups of annotators (15 and 13 members) under domain expert supervision. Through iterative training and validation, we reduced the initial annotation discrepancy rate from over $25\\%$ to below $3\\%$ . The final dataset comprises 16,000 training samples, 2,000 validation samples, and 2,000 test samples, with $20\\%$ undergoing additional expert review. Privacy protection measures and detailed annotation protocols are documented in Appendices C and H.", + "bbox": [ + 89, + 598, + 483, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Analysis of the annotated data reveals strong correlations between intentions and emotions. For instance, aggressive intentions (\"Criticize\", \"Disagree\") frequently co-occur with negative emotions. We also observe that certain intentions can complement emotional states - for example, \"Comfort\" intentions often appear in posts with \"Sad\" emotions, suggesting an attempt to provide emotional support. Additionally, while topic-intention correlations exist (e.g., \"Teach\" with education-related content), they are more one-to-one and less pronounced than emotion-intention relationships. Detailed analyses are provided in Appendix B.", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Feature Pre-extraction", + "text_level": 1, + "bbox": [ + 511, + 296, + 723, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We pre-extract features from each modality (text, image, video, and audio) in the dataset to facilitate multimodal fusion. This approach offers three key advantages: it reduces signal complexity and computational costs for researchers, enables consistent and fair comparisons across different multimodal fusion techniques, and minimizes privacy risks in our data. For feature extraction, we employ state-of-the-art pre-trained models: BERT [13] for text, ViT-S [5] for images, Video Swin Transformer [20] for video, and wav2vec 2.0 [1] for audio. All extracted features are aligned to a dimension of 768 to ensure consistent multimodal fusion. Detailed feature extraction procedures are documented in Appendix C.", + "bbox": [ + 511, + 318, + 906, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Proposed Method", + "text_level": 1, + "bbox": [ + 511, + 527, + 687, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Bridging the Relation of Emotion and Intention", + "text_level": 1, + "bbox": [ + 511, + 551, + 905, + 569 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As previously highlighted, the interconnection of intentions and emotions is important in both real-life situations and neural network analyses. Existing methodologies that focus solely on either intention or emotion often overlook the valuable insights from their interaction, limiting comprehensive understanding in both domains. To bridge this gap, we introduce the Bridging Emotion-Intention Former (BEIFormer), as depicted in Fig. 3. This model unifies these two critical domains. Specifically, we first establish the baseline for multimodal emotion and intention understanding by pre-extracting features from multimodal data to produce text, visual, and audio modality features $X_{L}$ , $X_{V}$ , and $X_{A}$ , respectively (as detailed in Sec. 3.3). We then employ Multi-modal Transformer [36] to aggregate features from any two modalities, extracting joint information, demonstrated here using visual and audio modalities:", + "bbox": [ + 511, + 574, + 906, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} Y _ {V} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {V} K _ {A} ^ {\\top}}{\\sqrt {d _ {k}}}\\right) V _ {A} (1) \\\\ = \\operatorname {s o f t m a x} \\left(\\frac {X _ {V} W _ {Q _ {V}} W _ {K _ {A}} ^ {\\top} X _ {A} ^ {\\top}}{\\sqrt {d _ {k}}}\\right) X _ {A} W _ {V _ {A}}, (1) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 828, + 903, + 904 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24703", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a6cca9e7cb725d8a11ccf96c4182bb6d5d2cf1e23375f69dece666d6d3f6c951.jpg", + "image_caption": [ + "Figure 3. Architecture of the Bridging Emotion-Intention via Implicit Label Reasoning (BEAR) framework. The framework consists of two main components: (1) a MulT Former that aggregates multimodal features into a unified representation, and (2) a Bridging Emotion Intention Former that fuses aggregated and prompt features. During training, ground truth emotion and intention labels are encoded to match the answer feature, enabling effective integration of emotion-intention signals and enhanced reasoning capabilities." + ], + "image_footnote": [], + "bbox": [ + 91, + 90, + 898, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Q_{V} = X_{V}W_{Q_{V}}$ , $K_{A} = X_{A}W_{K_{A}}$ , and $V_{A} = X_{A}W_{V_{A}}$ represent Query, Key, and Value component from the Audio modality, respectively. The term $W$ denotes the weights. Here, $Y_{V}$ maintains the same length as $Q_{V}$ (i.e., $T_{V}$ ), but exists in the feature space of $V_{A}$ . The softmax function in Eq. (1), scaled by $\\sqrt{d_k}$ , computes a score matrix $\\mathrm{softmax}(\\cdot) \\in \\mathbb{R}^{T_V\\times T_A}$ , with its $(i,j)$ -th entry quantifying the attention from the $i$ -th time step in modality $V$ to the $j$ -th time step in modality $A$ . Consequently, each $i$ -th time step of $Y_{V}$ represents a weighted summary of $V_{A}$ , weighted by the $i$ -th row in $\\mathrm{softmax}(\\cdot)$ .", + "bbox": [ + 89, + 335, + 483, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After applying two-pair cross-modal adaptation to three modalities, we obtain the modality aggregated representation $f_{ma}$ . However, this approach primarily reveals intention or emotion-related features within the multimodal data, driven by a singular training objective. It overlooks the intrinsic correlations between intentions and emotions that could serve as additional supervisory information. To leverage this relational information, we devise an implicit BEIFormer that integrates all-level intermediate tokens from both the Multi-modal Former, establishing implicit Emotion-Intention associations to enhance the understanding of these subjective dimensions. Specifically, we define the BEI Prompt $T_{prom}$ as a textual modality prior (as shown in Fig. 3), using [MASK] tokens to erase Emotion-Intention related priors. This process generates the hint feature $f_{prom}$ via the BEI Former (a transformer-based subnetwork), serving as the Query input for the BEIFormer. The BEIFormer then utilizes the modality aggregated feature $f_{ma}$ as Key and Value inputs to generate the final interaction output $f_{bef}$ .", + "bbox": [ + 89, + 505, + 483, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subsequently, we transform the emotion and intention labels into textual information $T_{gt}$ and use text encoder to align these labels within the same representation space as the hint feature, producing the ground truth feature $f_{gt}$ . The ground truth feature and the BEIFormer interaction output $f_{bef}$ are then used for contrastive learning during training:", + "bbox": [ + 89, + 810, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {N C E} \\left(x _ {i}, y _ {i}\\right) = - \\log \\frac {\\exp \\left(x _ {i} ^ {\\top} y _ {i} / \\tau\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(x _ {i} ^ {\\top} y _ {j} / \\tau\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 343, + 905, + 382 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x$ and $y$ are elements from $f_{bef}$ and $f_{gt}$ , respectively, $B$ denotes the batch size, and the temperature hyperparameter $\\tau$ is empirically set to 0.05. The full training loss is defined as:", + "bbox": [ + 511, + 383, + 906, + 441 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\left\\{ \\begin{array}{l l} \\mathcal {L} _ {\\mathrm {e m o}} + \\mathcal {L} _ {\\mathrm {N C E}}, & C _ {t} = \\mathrm {e m o}, \\\\ \\mathcal {L} _ {\\text {i n t}} + \\mathcal {L} _ {\\mathrm {N C E}}, & C _ {t} = \\text {i n t}, \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 446, + 905, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $C_t$ determines the current task. In our model, $\\mathcal{L}_{\\mathrm{emo/int}}$ represents the conventional loss for emotion or intention classification. For emotion recognition, we employ the standard multi-class cross-entropy loss:", + "bbox": [ + 511, + 492, + 905, + 553 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {e m o}} = \\frac {1}{N} \\sum_ {i} \\mathcal {L} _ {i} = - \\frac {1}{N} \\sum_ {i} \\sum_ {c = 1} ^ {M} y _ {i c} \\log \\left(p _ {i c}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 561, + 905, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N$ is the number of samples, $M$ is the number of classes, $y_{ic}$ indicates the ground truth label (1 if the $i$ -th sample belongs to class $c$ , and 0 otherwise), and $p_{ic}$ is the predicted probability of the $i$ -th sample belonging to class $c$ . This loss function also applies to multi-label intention understanding, ensuring a consistent approach across both emotion and intention classification tasks. Through implicit engagement with semantic correlations between intention and emotion during training, the BEIForner enhances the network's capacity to discern these subjective dimensions, enabling more nuanced understanding.", + "bbox": [ + 511, + 612, + 906, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Modality Asynchronous Prompt", + "text_level": 1, + "bbox": [ + 511, + 787, + 797, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to the interplay between Emotion-Intention, the challenge of uncertain modalities presents a significant hurdle. Existing methods fail to provide a reasonable solution. For instance, MMIN [50] focuses on learning inter-modal correlations during multimodal data training to infer missing modal information at deployment. However, this strat", + "bbox": [ + 511, + 810, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24704", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/158a6da33e75921e3bf0cf2f10fd9b7c8aff1fb86cff621d29903bba3bb66d81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 89, + 480, + 193 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4d25cea6030d62ff691a0b4c3f8adcd4e1ff284182c2298379bee9616889104e.jpg", + "image_caption": [ + "Figure 4. Comparison of Modal Asynchronous Prompt strategy with MPMM [16]. MAP preserves original representations in existing modalities while only prompting missing modalities, whereas the full modality prompt can corrupt existing features." + ], + "image_footnote": [], + "bbox": [ + 102, + 199, + 480, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "egy heavily relies on extensive multimodal priors, a requirement rarely met in social media contexts. While methods like MPMM [3, 16] take an alternative approach by using prompts to integrate missing information into neural networks, their strategy of indiscriminately concatenating all modalities, as shown in Fig. 4, risks compromising feature integrity. This approach can lead to data distortion, reducing model effectiveness in handling missing modalities.", + "bbox": [ + 89, + 381, + 482, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In response, we introduce the Modality Asynchronous Prompt (MAP), which dynamically provides asynchronous cues tailored to various types of missing modality. For each missing type, we develop a unique set of prompts that supply contextually relevant cues, enriching the missing modalities while preserving features of existing modalities. For example, when text and visual inputs are missing, our approach focuses exclusively on enhancing the missing modalities, ensuring that audio features remain undisturbed.", + "bbox": [ + 89, + 503, + 482, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, we integrate these prompts with existing features through element-wise addition (sums prompt features with missing modality features) rather than concatenation, minimizing parameter count and computational overhead by avoiding the additional Multi Layer Perceptron required in concatenation approaches.", + "bbox": [ + 89, + 640, + 483, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Benchmark and Experiments", + "text_level": 1, + "bbox": [ + 89, + 744, + 362, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Benchmark Methods", + "text_level": 1, + "bbox": [ + 89, + 770, + 290, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For our benchmark, we select established baseline methods in multimodal intention recognition [8, 36, 41] and missing modality challenges [16, 50], chosen for their demonstrated efficacy and influence. They are all trained on MINE.", + "bbox": [ + 89, + 792, + 482, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MulT. The Multimodal Transformer (MulT) [36] employs the Transformer [39] for learning multimodal language representations through directional cross-modal attention. For", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "consistency, we combine image and video features for visual representation across all multimodal approaches.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MISA. Modality-Invariant and -Specific Representations (MISA) [8] differentiates modality features into invariant and specific subspaces for enhanced fusion and prediction.", + "bbox": [ + 511, + 123, + 903, + 169 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MFSA. This approach [41] leverages a transformer-based model with adversarial learning to create modality-specific and -agnostic representations for sentiment recognition.", + "bbox": [ + 511, + 171, + 903, + 217 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MMIN. The Missing Modality Imagination Network (MMIN) [50] predicts absent modality representations through joint multimodal representation learning.", + "bbox": [ + 511, + 219, + 903, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MPMM [16] introduces a prompt-based learning framework, enhancing robustness in visual recognition tasks under various missing-modality conditions.", + "bbox": [ + 511, + 267, + 903, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All missing modality experiments use MulT [36] as the baseline framework for consistency, with detailed configurations that can be found in Appendix F.", + "bbox": [ + 511, + 314, + 903, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Experimental Setup", + "text_level": 1, + "bbox": [ + 513, + 376, + 702, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following [11, 33], we utilize three multi-label metrics for intention understanding: Macro F1, Micro F1, and Samples F1, chosen for comprehensive performance assessment. Detailed metric definitions appear in Appendix E. For emotion recognition, following [47], we use accuracy, F1 score, precision, and recall to evaluate model performance comprehensively.", + "bbox": [ + 511, + 401, + 905, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Main Experiments", + "text_level": 1, + "bbox": [ + 513, + 523, + 692, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3 presents the performance comparison of various methods on intention and emotion understanding tasks. Our approach consistently outperforms existing techniques across both validation and test splits, demonstrating significant advances in handling missing modalities and mining the Emotion-Intention relationship in social media analysis.", + "bbox": [ + 511, + 547, + 905, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On the validation set, our method achieves a Macro F1 score of $36.02\\%$ for intention understanding, surpassing the MulT baseline by $2.65\\%$ . For emotion recognition, it reaches $64.82\\%$ accuracy, exceeding MulT by $3.54\\%$ . These improvements persist in the test split, with our method achieving $35.79\\%$ Macro F1 and $64.53\\%$ accuracy, outperforming the baseline by $2.68\\%$ and $1.81\\%$ .", + "bbox": [ + 511, + 641, + 905, + 746 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "These results highlight the effectiveness of our strategy, particularly our Bridging Emotion-Intention via Implicit Label Reasoning (BEAR) approach. BEAR significantly enhances the accuracy of intention and emotion recognition, achieving substantial gains over the metrics compared to conventional training methods that treat these tasks separately. This underscores our method's capacity to offer a deeper and more accurate analysis of intention and emotion, validating its superiority in multimodal data analysis within the context of social media.", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24705", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4689a79a0876153d7c0cf8e61b6291040569d85c07c3fcc77ebc326b61c26e52.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SplitMethodIntentionEmotion
Macro F1Micro F1Samples F1Acc.F1 ScorePrecisionRecall
ValidationRandom15.1218.0317.694.911.854.822.25
MISA [8]23.8953.3253.6363.9318.6020.2217.55
MFSA [41]31.5553.8853.5264.1519.5620.0717.91
MulT [36]34.3752.4753.1861.2819.2920.8818.79
TCL(Map [51])34.2553.1453.7163.1519.1420.0118.79
MMIN [50]27.9148.8549.5059.3017.4520.6716.97
MPMM [16]32.8852.0753.0063.1517.6521.1017.58
Ours36.0254.0653.8864.8219.8921.3119.85
TestingRandom15.1817.9617.616.252.065.172.00
MISA [8]23.7452.2752.5463.2815.5523.9816.52
MFSA [41]32.4053.7553.0164.0517.6926.5817.91
MulT [36]34.1152.0552.7762.7220.3823.8319.58
TCL(Map [51])34.7351.6452.4163.3120.1425.2719.32
MMIN [50]27.1548.1449.0257.7115.4717.6015.17
MPMM [16]32.8152.1352.7462.8217.7825.2717.94
Ours35.2552.9053.3564.6720.5529.6419.95
", + "bbox": [ + 148, + 88, + 846, + 361 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 401, + 243, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Component Analysis. Our study delves into the advantages of combining the Modality Asynchronous Prompt (MAP) with the Bridging Emotion-Intention Former (BEIF), offering a detailed examination in Tab. 4. Our results reveal that while each component individually contributes to performance improvement—MAP elevating the Macro F1 score to $33.36\\%$ and BEIF to $35.25\\%$ for intention understanding—it's their combined application that significantly amplifies effectiveness, as evidenced by the compelling outcomes achieved (35.79% Macro F1, 53.92% Micro F1, and 54.88% Samples F1). The ablation study on MAP particularly highlights the effectiveness of our method, especially in addressing the challenge of missing modalities. The asynchronous prompting of missing information, crucially, does not compromise the feature space of the existing modalities. This approach, combined with our preference for element-wise addition over concatenation, ensures the dimensional consistency and integrity of the data representation remain intact. These findings indicate that our holistic approach is adept at offering a nuanced and thorough understanding of complex multimodal data.", + "bbox": [ + 88, + 430, + 485, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Emotion-Intention Correlation Mining. To validate the effectiveness of various strategies for mining intention and emotion correlations, we conduct ablation studies, as shown in Tab. 5. The experiments explore various methods, including a generic multi-task learning paradigm that directly combines emotion and intention losses. This paradigm, while straightforward, shows limited improvement (Macro F1: +1.21%) due to the absence of a specialized Emotion-Intention mechanism during training. It can even harm the ability to understand intention in certain scenarios, with decreases in Micro F1 (-1.06%) and Samples F1 (-0.30%) compared to the baseline. While the Complete Information method (Comp. Info.), which employs one aspect as a prior during both training and inference, exhibits notable gains (+9.06% Macro F1, +9.80% Micro F1), its practical application is limited by available prior knowledge in real-world scenarios. The limitations of these alternative approaches further emphasize the importance of BEAR, which achieves consistent improvements (+2.44% Macro F1) without requiring additional inference-time information.", + "bbox": [ + 511, + 401, + 906, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Feature Extractor Impact. Beyond model architecture, we investigate how different feature extractors influence in", + "bbox": [ + 511, + 717, + 908, + 750 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/104f5c2140859823d175394e43f91c195552b321dea3e646b40e16e15b3b850c.jpg", + "table_caption": [ + "Table 3. Performance comparison of multimodal fusion and missing modality methods on intention and emotion understanding tasks. Results show consistent improvements of our approach across all metrics on both validation and test sets." + ], + "table_footnote": [], + "table_body": "
ComponentIntentionEmotion
MAPEABEIFMacro F1Micro F1Samples F1Acc.F1 ScorePrecisionRecall
32.8152.1352.7462.8217.7825.2717.94
✓33.3652.7453.4362.7218.9528.4319.30
✓✓34.3752.4753.1861.2819.2925.8818.79
✓35.2552.9053.3564.6720.0229.1919.35
✓✓✓35.7953.9254.8864.5320.5529.6420.35
", + "bbox": [ + 174, + 767, + 820, + 881 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation study showing the contribution of each component. Integration of both Modality Asynchronous Prompt (MAP) and Bridging Emo-Int Former (BEIF) yields optimal performance across all metrics. EA means element-wise addition.", + "bbox": [ + 89, + 886, + 906, + 916 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24706", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f21729a5b1e2bf20557f546e842d3eeeff6e1b3c4b50957bf0757b7e38ffab1e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMacro F1Intention
Micro F1Samples F1
Baseline32.8152.1352.74
w/ Multi-task34.0251.0752.44
(+1.21)(-1.06)(-0.30)
w/ EmoInt-Trans [35]31.5952.0651.32
(-1.22)(-0.06)(-1.42)
w/ BEIFormer35.2552.9053.35
(+2.44)(+0.77)(+0.61)
w/ Comp. Info.†41.8761.9362.09
(+9.06)(+9.80)(+8.54)
", + "bbox": [ + 91, + 88, + 485, + 257 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tention and emotion understanding performance. We evaluate various state-of-the-art extractors for visual [20, 27], audio [7], and textual [49] modalities. As shown in Tab. 6, more advanced feature extractors consistently improve performance. CLIP-ViT-S [27] provides the most substantial gains for visual features (+1.81% Macro F1), while SemBERT [49] significantly enhances text representation (+1.42% Micro F1). Notably, while AST shows modest improvements in the metrics (+0.76% Macro F1), its impact is more significant than these numbers suggest, given that audio data comprises only a small portion of the test set.", + "bbox": [ + 88, + 315, + 480, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generalization Analysis on CMU-MOSEI. To evaluate the generalizability of our MAP approach, we conduct experiments on the CMU-MOSEI dataset across various baseline architectures, as detailed in Tab. 7. MAP demonstrates consistent improvements across all tested configurations. These consistent gains across different architectures and modality combinations validate MAP's robustness in handling missing modalities.", + "bbox": [ + 88, + 482, + 480, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Data Imbalance. The modality imbalance in MINE is intentional, reflecting the natural, in-the-wild modality distribution of social media content. To evaluate its impact, we conduct experiments with modality resampling. As shown in Tab. 8, resampling (the rare modalities by $2\\mathrm{x}$ and $3\\mathrm{x}$ ), especially feature-level contribution resampling [40], can improve the understanding performance.", + "bbox": [ + 88, + 603, + 480, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 89, + 724, + 339, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce MINE, a novel dataset capturing the inherent relationship between intention and emotion in natural social media settings. Unlike existing datasets that treat these aspects in isolation or rely on controlled environments, MINE preserves the authentic challenges of real-world social media data—particularly uncertain modalities and diverse thematic contexts. By providing detailed annotations for both intention and emotion across multiple modalities, MINE establishes a comprehensive benchmark for understanding human social behavior in naturalistic settings.", + "bbox": [ + 88, + 750, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1a49365f489a3abba56f533a0e8f5e63bf2343d71d59eaf890be7cba47d0d12b.jpg", + "table_caption": [ + "Table 5. Advanced relation mining strategies are important in enhancing emotion-intention analysis. †represents the utilization of supplementary information during the inference process." + ], + "table_footnote": [], + "table_body": "
MethodMacro F1Intention
Micro F1Samples F1
Baseline32.8152.1352.74
w/ AST [7]33.5752.9553.34
(+0.76)(+0.82)(+0.60)
w/ CLIP-ViT-S [27]34.6253.7054.02
(+1.81)(+1.57)(+1.28)
w/ SwinT-S [20]33.1452.6752.98
(+0.33)(+0.54)(+0.24)
w/ SemBERT [49]33.8953.5554.02
(+1.08)(+1.42)(+1.28)
", + "bbox": [ + 517, + 88, + 898, + 257 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e44363322b3adb1a24bb89bd16fae5b6755b1aaf8140f9f03c02888bd89da760.jpg", + "table_caption": [ + "Table 6. Impact of various feature extractors." + ], + "table_footnote": [], + "table_body": "
ModelUtilized modality
T, VT, AT
MulT [36]37.1538.2841.91
w/ MAP38.93+1.7839.14+0.8643.35+1.44
MFN [45]58.4860.5459.74
w/ MAP61.33+2.8561.95+1.4163.72+3.98
BERT_MAG [28]62.8361.3962.83
w/ MAP64.51+1.6862.55+1.1664.42+1.59
LMF [19]42.3857.1555.94
w/ MAP44.25+1.8758.67+1.5257.12+1.18
", + "bbox": [ + 517, + 301, + 898, + 474 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6c7a7042726c39e02b7316659c93b26b14882592a8b7246eaaa408a3ab58b4cf.jpg", + "table_caption": [ + "Table 7. Evaluation MAP across various architectures and modality combinations on CMU-MOSEI. T: Text, V: Visual, A: Audio." + ], + "table_footnote": [], + "table_body": "
Methodsours2*rare3*rareWei et al [40]
Average51.4551.7951.0853.20
", + "bbox": [ + 535, + 527, + 880, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Results of various methods for resisting modal imbalance.", + "bbox": [ + 513, + 575, + 903, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To effectively process MINE's complex multimodal data, we propose BEAR (Bridging Emotion-Intention via Implicit Label Reasoning). It addresses two critical challenges: the intrinsic relationship between emotion and intention, and the uncertainty of modalities. The BEAR employs an alignment task with masked emotion and intention categories, enabling implicit mining of their relationships. Meanwhile, the Modality Asynchronous Prompt handles missing modalities through targeted prompting that preserves the integrity of available features. Experimental results demonstrate BEAR's effectiveness, outperforming existing methods across various metrics. In summary, MINE opens new avenues for understanding human social behavior through multimodal content.", + "bbox": [ + 509, + 609, + 903, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work is partially supported by the National Key Research and Development Program of China (2024YFC3308400), National Natural Science Foundation of China under Grant (62176188, 62361166629), Major Project of Science and Technology Innovation of Hubei Province (2024BCA003) and Key Project of the Health Commission of Hubei Province (WJ2023Z003).", + "bbox": [ + 511, + 825, + 903, + 924 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24707", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. In NIPS, 2020. 4", + "[2] Carlos Busso, Murtaza Bulut, Chi-Chun Lee, Abe Kazemzadeh, Emily Mower, Samuel Kim, Jeannette N Chang, Sungbok Lee, and Shrikanth S Narayanan. Iemocap: Interactive emotional dyadic motion capture database. Language resources and evaluation, 2008. 1, 2, 3, 4", + "[3] Haozhe Chi, Minghua Yang, Junhao Zhu, Guanhong Wang, and Gaoang Wang. Missing modality meets meta sampling (m3s): An efficient universal approach for multimodal sentiment analysis with missing modality. arXiv, 2022. 2, 3, 6", + "[4] Dorottya Demszky, Dana Movshovitz-Attias, Jeongwoo Ko, Alan Cowen, Gaurav Nemade, and Sujith Ravi. Goemotions: A dataset of fine-grained emotions. arXiv, 2020. 3", + "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv, 2020. 4", + "[6] Yiyang Fang, Wenke Huang, Guancheng Wan, Kehua Su, and Mang Ye. Emoe: Modality-specific enhanced dynamic emotion experts. In CVPR, 2025. 1", + "[7] Yuan Gong, Yu-An Chung, and James Glass. Ast: Audio spectrogram transformer. arXiv, 2021. 8", + "[8] Devamanyu Hazarika, Roger Zimmermann, and Soujanya Poria. Misa: Modality-invariant and-specific representations for multimodal sentiment analysis. In ACM MM, 2020. 6, 7", + "[9] Charles T Hemphill, John J Godfrey, and George R Doddington. The atis spoken language systems pilot corpus. In Speech and Natural Language: Proceedings of a Workshop, 1990. 2", + "[10] Ao Jia, Yu He, Yazhou Zhang, Sagar Uprety, Dawei Song, and Christina Lioma. Beyond emotion: A multi-modal dataset for human desire understanding. In NACCL, 2022. 3", + "[11] Menglin Jia, Zuxuan Wu, Austin Reiter, Claire Cardie, Serge Belongie, and Ser-Nam Lim. Intentonomy: a dataset and study towards human intent understanding. In CVPR, 2021. 2, 3, 4, 6", + "[12] Kawaljeet Kaur Kapoor, Kuttimani Tamilmani, Nripendra P Rana, Pushp Patil, Yogesh K Dwivedi, and Sridhar Nerur. Advances in social media research: Past, present and future. Information Systems Frontiers, 2018. 3", + "[13] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NACCL, 2019. 4", + "[14] Julia Kruk, Jonah Lubin, Karan Sikka, Xiao Lin, Dan Jurafsky, and Ajay Divakaran. Integrating text and image: Determining multimodal document intent in Instagram posts. In EMNLP-IJCNLP, 2019. 3", + "[15] Stefan Larson, Anish Mahendran, Joseph J Peper, Christopher Clarke, Andrew Lee, Parker Hill, Jonathan K Kummerfeld, Kevin Leach, Michael A Laurenzano, Lingjia Tang," + ], + "bbox": [ + 93, + 114, + 482, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "et al. An evaluation dataset for intent classification and out-of-scope prediction. In EMNLP-IJCNLP, 2019. 2, 3", + "[16] Yi-Lun Lee, Yi-Hsuan Tsai, Wei-Chen Chiu, and Chen-Yu Lee. Multimodal prompting with missing modalities for visual recognition. In CVPR, 2023. 3, 6, 7", + "[17] Rui Liu, Haolin Zuo, Zheng Lian, Xiaofen Xing, Bjorn W Schuller, and Haizhou Li. Emotion and intent joint understanding in multimodal conversation: A benchmarking dataset. arXiv, 2024. 2", + "[18] Xingkun Liu, Arash Eshghi, Pawel Swietojanski, and Verena Rieser. Benchmarking natural language understanding services for building conversational agents. arXiv, 2019. 3", + "[19] Zhun Liu, Ying Shen, Varun Bharadhwaj Lakshminarasimhan, Paul Pu Liang, Amir Zadeh, and LouisPhilippe Morency. Efficient low-rank multimodal fusion with modality-specific factors. arXiv, 2018. 8", + "[20] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In CVPR, 2022. 4, 8", + "[21] Meng Luo, Hao Fei, Bobo Li, Shengqiong Wu, Qian Liu, Soujanya Poria, Erik Cambria, Mong-Li Lee, and Wynne Hsu. Panosent: A panoptic sextuple extraction benchmark for multimodal conversational aspect-based sentiment analysis. In ACM MM, 2024. 2", + "[22] Mengmeng Ma, Jian Ren, Long Zhao, Sergey Tulyakov, Cathy Wu, and Xi Peng. Smil: Multimodal learning with severely missing modality. In AAAI, 2021. 3", + "[23] Mengmeng Ma, Jian Ren, Long Zhao, Davide Testuggine, and Xi Peng. Are multimodal transformers robust to missing modality? In CVPR, 2022. 3", + "[24] Kuan-Chuan Peng, Amir Sadovnik, Andrew Gallagher, and Tsuhan Chen. Where do emotions come from? predicting the emotion stimuli map. In IEEE ICIP, 2016. 3", + "[25] Wei Peng, Yue Hu, Yuqiang Xie, Luxi Xing, and Yajing Sun. Cogintac: Modeling the relationships between intention, emotion and action in interactive process from cognitive perspective. In 2022 IEEE Congress on Evolutionary Computation, 2022. 1", + "[26] Soujanya Poria, Devamanyu Hazarika, Navonil Majumder, Gautam Naik, Erik Cambria, and Rada Mihalcea. Meld: A multimodal multi-party dataset for emotion recognition in conversations. In ACL, 2019. 2, 3", + "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 8", + "[28] Wasifur Rahman, Md Kamrul Hasan, Sangwu Lee, AmirAli Bagher Zadeh, Chengfeng Mao, Louis-Philippe Morency, and Ehsan Hoque. Integrating multimodal information in large pretrained transformers. In ACL, 2020. 8", + "[29] Md Kaykobad Reza, Ashley Prater-Bennette, and M Salman Asif. Robust multimodal learning with missing modalities via parameter-efficient adaptation. IEEE TPAMI, 2024. 3", + "[30] Tulika Saha, Aditya Patra, Sriparna Saha, and Pushpak Bhattacharyya. Towards emotion-aided multi-modal dialogue act classification. In ACL, 2020. 1" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24708", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Tulika Saha, Aditya Patra, Sriparna Saha, and Pushpak Bhattacharyya. Towards emotion-aided multi-modal dialogue act classification. In ACL, 2020. 2, 3", + "[32] Tobias Schröder, Terrence C Stewart, and Paul Thagard. Intention, emotion, and action: A neural theory based on semantic pointers. Cognitive science, 2014. 1, 4", + "[33] QingHongYa Shi, Mang Ye, Ziyi Zhang, and Bo Du. Learnable hierarchical label embedding and grouping for visual intention understanding. IEEE Transactions on Affective Computing, 2023. 6", + "[34] Mohammad Faridul Haque Siddiqui, Parashar Dhakal, Xiaoli Yang, and Ahmad Y Javaid. A survey on databases for multimodal emotion recognition and an introduction to the viri (visible and infrared image) database. Multimodal Technologies and Interaction, 2022. 4", + "[35] Gopendra Vikram Singh, Mauajama Firdaus, Asif Ekbal, and Pushpak Bhattacharyya. Emoin-trans: A multimodal transformer for identifying emotions and intents in social conversations. IEEE/ACM TASLP, 2022. 1, 2, 8", + "[36] Yao-Hung Hubert Tsai, Shaojie Bai, Paul Pu Liang, J. Zico Kolter, Louis-Philippe Morency, and Ruslan Salakhutdinov. Multimodal transformer for unaligned multimodal language sequences. In ACL, 2019. 4, 6, 7, 8", + "[37] Gokhan Tur, Dilek Hakkani-Tur, and Larry Heck. What is left to be understood in atis? In IEEE Spoken Language Technology Workshop, 2010. 2", + "[38] Chai M Tyng, Hafeez U Amin, Mohammad NM Saad, and Aamir S Malik. The influences of emotion on learning and memory. Frontiers in psychology, 2017. 1", + "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, 2017. 6", + "[40] Yake Wei, Ruoxuan Feng, Zihe Wang, and Di Hu. Enhancing multimodal cooperation via sample-level modality valuation. In CVPR, 2024. 8", + "[41] Dingkang Yang, Haopeng Kuang, Shuai Huang, and Lihua Zhang. Learning modality-specific and-agnostic representations for asynchronous multimodal language sequences. In ACM MM, 2022. 6, 7", + "[42] Qu Yang, Mang Ye, and Bo Du. Emollm: Multimodal emotional understanding meets large language models. arXiv, 2024. 1", + "[43] Qu Yang, Mang Ye, and Dacheng Tao. Synergy of sight and semantics: visual intention understanding with clip. In ECCV, 2024. 1", + "[44] Amir Zadeh, Rowan Zellers, Eli Pincus, and Louis-Philippe Morency. Mosi: multimodal corpus of sentiment intensity and subjectivity analysis in online opinion videos. arXiv, 2016. 1", + "[45] Amir Zadeh, Paul Pu Liang, Navonil Mazumder, Soujanya Poria, Erik Cambria, and Louis-Philippe Morency. Memory fusion network for multi-view sequential learning. In AAAI, 2018. 8", + "[46] AmirAli Bagher Zadeh, Paul Pu Liang, Soujanya Poria, Erik Cambria, and Louis-Philippe Morency. Multimodal language analysis in the wild: Cmu-mosei dataset and interpretable dynamic fusion graph. In ACL, 2018. 1, 2, 3" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Hanlei Zhang, Hua Xu, Xin Wang, Qianrui Zhou, Shaojie Zhao, and Jiayan Teng. Mintrec: A new dataset for multimodal intent recognition. In ACM MM, 2022. 1, 2, 3, 4, 6", + "[48] Hanlei Zhang, Xin Wang, Hua Xu, Qianrui Zhou, Kai Gao, Jianhua Su, Wenrui Li, Yanting Chen, et al. Mintrec2.0: A large-scale benchmark dataset for multimodal intent recognition and out-of-scope detection in conversations. In ICLR, 2024. 2, 3", + "[49] Zhuosheng Zhang, Yuwei Wu, Hai Zhao, Zuchao Li, Shuai-liang Zhang, Xi Zhou, and Xiang Zhou. Semantics-aware bert for language understanding. In AAAI, 2020. 8", + "[50] Jinming Zhao, Ruichen Li, and Qin Jin. Missing modality imagination network for emotion recognition with uncertain missing modalities. In ACL, 2021. 2, 3, 5, 6, 7", + "[51] Qianrui Zhou, Hua Xu, Hao Li, Hanlei Zhang, Xiaohan Zhang, Yifan Wang, and Kai Gao. Token-level contrastive learning with modality-aware prompting for multimodal intent recognition. In AAAI, 2024. 7" + ], + "bbox": [ + 514, + 92, + 903, + 359 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24709", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Uncertain Multimodal Intention and Emotion Understanding in the Wild/50195041-da23-4b5c-8721-3858d2b20dc9_model.json b/2025/Uncertain Multimodal Intention and Emotion Understanding in the Wild/50195041-da23-4b5c-8721-3858d2b20dc9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..32731241484e19db79ca81ab2cd9f4b5e908c658 --- /dev/null +++ b/2025/Uncertain Multimodal Intention and Emotion Understanding in the Wild/50195041-da23-4b5c-8721-3858d2b20dc9_model.json @@ -0,0 +1,2046 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.13, + 0.131, + 0.868, + 0.154 + ], + "angle": 0, + "content": "Uncertain Multimodal Intention and Emotion Understanding in the Wild" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.179, + 0.753, + 0.2 + ], + "angle": 0, + "content": "Qu Yang\\(^{1}\\), Qinghongya Shi\\(^{1}\\), Tongxin Wang\\(^{1}\\) and Mang Ye\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.2, + 0.756, + 0.217 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Computer Science, Wuhan University, Wuhan, China" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.217, + 0.824, + 0.235 + ], + "angle": 0, + "content": "\\(^{2}\\) Taikang Center for Life and Medical Sciences, Wuhan University, Wuhan, China" + }, + { + "type": "text", + "bbox": [ + 0.278, + 0.237, + 0.721, + 0.253 + ], + "angle": 0, + "content": "{yangqu, shiqinghongya, wangtx, yemang}@whu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.329, + 0.255, + 0.661, + 0.269 + ], + "angle": 0, + "content": "https://github.com/yan9qu/CVPR25-MINE" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.304, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.485, + 0.701 + ], + "angle": 0, + "content": "Understanding intention and emotion from social media poses unique challenges due to the inherent uncertainty in multimodal data, where posts often contain incomplete or missing modalities. While this uncertainty reflects real-world scenarios, it remains underexplored within the computer vision community, particularly in conjunction with the intrinsic relationship between emotion and intention. To address these challenges, we introduce the Multimodal IntentioN and Emotion Understanding in the Wild (MINE) dataset, comprising over 20,000 topic-specific social media posts with natural modality variations across text, image, video, and audio. MINE is distinctively constructed to capture both the uncertain nature of multimodal data and the implicit correlations between intentions and emotions, providing extensive annotations for both aspects. To tackle these scenarios, we propose the Bridging Emotion-Intention via Implicit Label Reasoning (BEAR) framework. BEAR consists of two key components: a BEIFormer that leverages emotion-intention correlations, and a Modality Asynchronous Prompt that handles modality uncertainty. Experiments show that BEAR outperforms existing methods in processing uncertain multimodal data while effectively mining emotion-intention relationships for social media content understanding." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.73, + 0.222, + 0.745 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.756, + 0.484, + 0.877 + ], + "angle": 0, + "content": "Intention and emotion are fundamental in shaping human social behaviors and decision-making processes [25, 32, 38]. Understanding their interplay is crucial for the advancement of artificial general intelligence. While prior studies [6, 42, 43] in the vision community have independently explored intention or emotion recognition in multimodal contexts (as illustrated in Fig. 1), employing datasets like MIntRec [47] and IEMOCAP [2], these datasets fail" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.305, + 0.903, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.491, + 0.907, + 0.52 + ], + "angle": 0, + "content": "Figure 1. Overview of multi-modal emotion and intention understanding from social media platforms." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.701 + ], + "angle": 0, + "content": "to capture the inherent link between intention and emotion and overlook the challenges of uncertain modalities and thematic diversity in real-world scenarios. Addressing these gaps is vital for a comprehensive understanding of intentions and emotions in natural environments. In this paper, we introduce the Multimodal Intention and EmotioN Understanding in the Wild (MINE) dataset. MINE provides detailed annotations of both intention and emotion, establishing a platform to explore their correlation. It presents the challenges of uncertain modalities, thus mirroring the intricacies of real-life social interactions, as shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The design of MINE is guided by several key principles: 1) MINE addresses a critical gap in existing datasets by enabling comprehensive insights into the interactions between intention and emotion within social contexts. 2) In contrast to data typically curated under controlled conditions (such as scripted dialogues or TV series clips [2, 30, 35, 44, 46, 47]), which exhibit significant disparities with real-world situations, MINE leverages authentic user-generated content from social platforms, thus mirroring real-world dynamics of intention and emotion expression. 3) Unlike related datasets that typically assume fixed modalities [2, 47], MINE incorporates the challenge of uncertain modalities in social media, enhancing the realism and applicability of" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.887, + 0.253, + 0.901 + ], + "angle": 0, + "content": "Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24700" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.182, + 0.094, + 0.825, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.253, + 0.908, + 0.283 + ], + "angle": 0, + "content": "Figure 2. Overview of the MINE dataset, illustrating the multimodal data collection from social media, categorized by topics and including text, image, video, and audio content, each reflecting various intentions and emotions. The modality combination is uncertain." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.29, + 0.486, + 0.518 + ], + "angle": 0, + "content": "intention and emotion understanding tasks. At the algorithmic level, this manifests as the missing modality problem, where modality heterogeneity is primarily reflected in the absence of certain modal inputs. Previous research on missing modalities [3, 50] often relies on artificial sampling methods, creating biases and noise due to incomplete information (e.g., an image showing “news about something” with the text “That’s so ironic!”, where removing text leads to misinterpretation). MINE, however, preserves native modal information, eliminating potential errors and offering a reliable benchmark for uncertain modality research. 4) By utilizing topical tags (such as #CAMPUS) from social media, MINE provides objective context for analyzing intentions and emotions, enabling a more nuanced understanding of these elements within specific thematic frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.519, + 0.485, + 0.791 + ], + "angle": 0, + "content": "To harness the intricate relationship between emotions and intentions within the MINE dataset, we propose Bridging Emotion-Intention via Implicit Label Reasoning (BEAR) framework, comprising two key components: BEIFormer and Modality Asynchronous Prompt (MAP). The BEIFormer introduces an innovative contrastive task, utilizing [MASK] tokens to conceal intention and emotion labels within training prompts. This design enables the model to reconstruct textual data that is in alignment with ground truth labels, leveraging multimodal inputs and the process of unmasking text. This implicit reconstruction and alignment process in BEAR enhances the understanding by delving into the relationship between emotion and intention. Furthermore, the MAP is specifically engineered to tackle the challenge of uncertain modalities. It provides customized cues tailored for diverse uncertain scenarios, thereby substantially improving understanding capacity in contexts where modality data may be incomplete." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.793, + 0.391, + 0.808 + ], + "angle": 0, + "content": "In summary, our contribution is threefold:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.486, + 0.902 + ], + "angle": 0, + "content": "(1) We introduce MINE, the first large-scale dataset that captures the inherent relationship between intention and emotion. Comprising over 20,000 topic-specific posts with natural modality distribution, MINE uniquely provides accurate annotations for intention and emotion, enabling systematic study of their correlations in realistic scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.29, + 0.907, + 0.397 + ], + "angle": 0, + "content": "(2) We propose Bridging Emotion-Intention via Implicit Label Reasoning, a novel framework that bridges emotion and intention understanding. It incorporates two key components: (i) BEIFormer that implicitly mines emotion-intention relationships through contrastive learning, and (ii) Modality Asynchronous Prompt that handles uncertain modalities without compromising existing features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.401, + 0.909, + 0.507 + ], + "angle": 0, + "content": "(3) We establish comprehensive benchmarks on MINE by evaluating multiple baselines. Through extensive experiments, we demonstrate that BEAR consistently outperforms existing methods by significant margins (e.g., \\(+2.65\\%\\) Macro F1 for intention recognition and \\(+3.54\\%\\) accuracy for emotion recognition), validating the importance of joint emotion-intention modeling." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.29, + 0.909, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.536, + 0.656, + 0.552 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.566, + 0.845, + 0.583 + ], + "angle": 0, + "content": "2.1. Intention and Emotion Understanding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.907, + 0.683 + ], + "angle": 0, + "content": "Intention recognition is to understand the goals behind human actions. While early datasets like ATIS [9, 37] were limited to specific domains, recent datasets like CLINC150[15], Intentonomy [11], and MIntRec [47, 48] have expanded to diverse sources including social media and TV series." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.902 + ], + "angle": 0, + "content": "For emotion recognition, datasets have evolved from controlled environments (IEMOCAP [2]) to more naturalistic settings with CMU-MOSEI [46], PanoSent [21], and MELD [26], encompassing multimodal expressions across text, speech, and visual cues. However, as shown in Tab. 1, existing datasets typically focus on either intention or emotion in isolation. While recent works such as EMO-TyDA [31], MC-EIU [17], and EmoInt-MD [35] attempt to combine both aspects, they are constrained by controlled settings and fixed modalities, unlike real-world social media posts which often contain varying combinations of modalities. Our MINE dataset addresses these limitations by capturing authentic emotion-intention relationships while preserving the natural complexity of missing modalities." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "24701" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.179, + 0.089, + 0.823, + 0.34 + ], + "angle": 0, + "content": "
TypeDatasetModalityMissingClass NumsData NumsSource
####Emo.Int.
IntentionCLINC150 [15]✓✗✗✗✗-15023,700Acted
HWU64 [18]✓✗✗✗✗-6425,716Acted
Intentonomy [11]✓✓✗✗✗-2814,455In the wild
MDID [14]✓✓✗✗✗-81,299In the wild
MSED [10]✓✓✗✗✗-69,190In the wild
MIntRec [47]✓✗✓✓✗-202,224TV
MIntRec2.0 [48]✓✗✓✓✗-3015,040TV
EmotionIEMOCAP [2]✓✗✓✓✗6-7,433Acted
CMU-MOSEI [46]✓✗✓✓✗6-23,453In the wild
GoEmotions [4]✓✗✗✗✗28-58,000In the wild
EmotionRoI [24]✓✗✗✗✗6-1,980In the wild
MELD [26]✓✗✓✓✗8-13,708TV
AllEMOTyDA [31]✓✗✓✓✗101219,365Acted
MINE✓✓✓✓✓112020,168In the wild
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.35, + 0.907, + 0.379 + ], + "angle": 0, + "content": "Table 1. Comparison of existing datasets for intention (Int), emotion (Emo), and combined (All) understanding. MINE uniquely addresses uncertain modalities while combining both emotion and intention analysis in real-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.404, + 0.37, + 0.42 + ], + "angle": 0, + "content": "2.2. Uncertain Modality in the Wild" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.426, + 0.483, + 0.563 + ], + "angle": 0, + "content": "Uncertain modality (or missing modality) presents a fundamental challenge in multimodal understanding. Contemporary studies [3, 16, 22, 23, 50] have investigated this challenge using datasets like IEMOCAP [2] and CMUMOSEI [46]. Researchers have proposed various approaches, ranging from automated fusion strategy identification [22] and meta-sampling integration [3] to unified frameworks for modality inference [29, 50] and Bayesian meta-learning applications [23]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.563, + 0.483, + 0.651 + ], + "angle": 0, + "content": "However, these methods typically rely on random sampling to simulate missing modalities, which can introduce bias and alter the original content meaning (refer to Appendix G). In contrast, MINE captures naturally occurring uncertain modalities in the wild, providing a more authentic foundation for multimodal research." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.665, + 0.238, + 0.68 + ], + "angle": 0, + "content": "3. MINE Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.766 + ], + "angle": 0, + "content": "In this section, we outline the creation of the MINE dataset, including steps such as data preparation, establishing categories for intentions and emotions, annotating multimodal content with these categories, and analyzing the data. Detailed metrics of the MINE dataset are provided in Tab. 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.422, + 0.79 + ], + "angle": 0, + "content": "3.1. Topic Definition and Data Preparation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To explore how topics influence the understanding of intentions and emotions in real-world scenarios, we obtained permission from the consent of data subjects and collected high-quality, multimodal data from social media platforms—venues teeming with genuine user expressions. This approach significantly diverges from the artificial confines of laboratory settings or the limited diversity found" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.405, + 0.907, + 0.466 + ], + "angle": 0, + "content": "in sources like movie clips, frequently utilized in prior research. Our strategy involved organizing a wide range of multimodal data via clusters based on topic-specific hashtags, ensuring a rich and varied data source." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.909, + 0.715 + ], + "angle": 0, + "content": "The MINE dataset comprehensively covers ten topics that are integral to daily life. Following the methodology of Kapoor et al. [12], we selected these topics from the top 100 most frequently discussed subjects on social media platforms. MINE explores these topics in depth, paying particular attention to intersections identified through related keywords. For instance, the keyword 'Education' encompasses associated themes such as 'Social Media Impact on Teachers and Students' and 'Use of Social Media in Academic Learning'. To ensure thematic diversity, we selected ten distinct topics for our study and developed a comprehensive keyword framework for each one, with detailed documentation available in Appendix A. For data collection, we employed a hashtag-based method (e.g., #Education) rather than general topic searches to capture a broader variety of content within each topic." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The data collection spans Twitter posts from 2019 to 2022, with all content obtained with explicit user consent. To ensure data quality, we implemented multiple preprocessing steps, including the removal of nonsensical texts (those consisting solely of links or random characters) and the filtering of harmful content (related to pornography and drugs). While visual content plays a crucial role in conveying emotional and intentional nuances, we maintained a deliberate balance with textual content to ensure comprehensive multimodal analysis. The final dataset composition reflects this approach, with visual content constituting approximately \\(70\\%\\) of the data, creating a diverse yet imbal" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24702" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.131, + 0.089, + 0.468, + 0.257 + ], + "angle": 0, + "content": "
ItemNumber
Number of intentions20
Number of emotions11
Number of text20044
Number of images12428
Number of videos3895
Number of audios1615
Average length of text utterances (words)20.0
Total number of words489069
Average duration of videos (seconds)24.4
Total duration of videos (hours)21.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.248, + 0.26, + 0.345, + 0.271 + ], + "angle": 0, + "content": "(a) Data Statistics" + }, + { + "type": "table", + "bbox": [ + 0.562, + 0.089, + 0.848, + 0.258 + ], + "angle": 0, + "content": "
TextImageVideoAudioNumber
✓3913
✓99
✓4
✓✓12261
✓✓2248
✓✓3
✓✓18
✓✓✓26
✓✓✓1557
✓✓✓✓39
" + }, + { + "type": "table_caption", + "bbox": [ + 0.591, + 0.26, + 0.817, + 0.272 + ], + "angle": 0, + "content": "(b) Distribution of Modality Combinations" + }, + { + "type": "table_caption", + "bbox": [ + 0.323, + 0.275, + 0.673, + 0.288 + ], + "angle": 0, + "content": "Table 2. Data and Modality Statistics of the MINE dataset." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.298, + 0.434, + 0.312 + ], + "angle": 0, + "content": "anced multimodal distribution, as detailed in Tab. 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.32, + 0.455, + 0.336 + ], + "angle": 0, + "content": "3.2. Taxonomy Definition and Data Annotation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.484, + 0.433 + ], + "angle": 0, + "content": "Defining labels for intentions and emotions presents a unique challenge due to their subjective nature. While previous work [47] treated intentions as mutually exclusive categories ('Expressing emotions' or 'Achieving goals'), we argue that social interactions naturally blend both elements, aligning with philosophical theories [32]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.484, + 0.599 + ], + "angle": 0, + "content": "MINE incorporates 20 intention and 11 emotion labels per data point, drawing from literature on both dialogue and visual intention analysis [11, 47] and emotion research [2, 34]. We frame intention recognition as a multi-label task to capture the nuanced nature of social media interactions, while structuring emotion recognition as a single-label task to identify dominant emotions, consistent with established practices [2]. To address inherent subjectivity, our annotation employs a crowd-sourcing platform with dual annotations. Detailed category definitions and annotation guidelines are provided in Appendices A and H." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.484, + 0.734 + ], + "angle": 0, + "content": "The annotation workflow involved two independent groups of annotators (15 and 13 members) under domain expert supervision. Through iterative training and validation, we reduced the initial annotation discrepancy rate from over \\(25\\%\\) to below \\(3\\%\\). The final dataset comprises 16,000 training samples, 2,000 validation samples, and 2,000 test samples, with \\(20\\%\\) undergoing additional expert review. Privacy protection measures and detailed annotation protocols are documented in Appendices C and H." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Analysis of the annotated data reveals strong correlations between intentions and emotions. For instance, aggressive intentions (\"Criticize\", \"Disagree\") frequently co-occur with negative emotions. We also observe that certain intentions can complement emotional states - for example, \"Comfort\" intentions often appear in posts with \"Sad\" emotions, suggesting an attempt to provide emotional support. Additionally, while topic-intention correlations exist (e.g., \"Teach\" with education-related content), they are more one-to-one and less pronounced than emotion-intention relationships. Detailed analyses are provided in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.297, + 0.725, + 0.312 + ], + "angle": 0, + "content": "3.3. Feature Pre-extraction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.319, + 0.907, + 0.516 + ], + "angle": 0, + "content": "We pre-extract features from each modality (text, image, video, and audio) in the dataset to facilitate multimodal fusion. This approach offers three key advantages: it reduces signal complexity and computational costs for researchers, enables consistent and fair comparisons across different multimodal fusion techniques, and minimizes privacy risks in our data. For feature extraction, we employ state-of-the-art pre-trained models: BERT [13] for text, ViT-S [5] for images, Video Swin Transformer [20] for video, and wav2vec 2.0 [1] for audio. All extracted features are aligned to a dimension of 768 to ensure consistent multimodal fusion. Detailed feature extraction procedures are documented in Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.529, + 0.688, + 0.546 + ], + "angle": 0, + "content": "4. Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.553, + 0.906, + 0.57 + ], + "angle": 0, + "content": "4.1. Bridging the Relation of Emotion and Intention" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.575, + 0.907, + 0.818 + ], + "angle": 0, + "content": "As previously highlighted, the interconnection of intentions and emotions is important in both real-life situations and neural network analyses. Existing methodologies that focus solely on either intention or emotion often overlook the valuable insights from their interaction, limiting comprehensive understanding in both domains. To bridge this gap, we introduce the Bridging Emotion-Intention Former (BEIFormer), as depicted in Fig. 3. This model unifies these two critical domains. Specifically, we first establish the baseline for multimodal emotion and intention understanding by pre-extracting features from multimodal data to produce text, visual, and audio modality features \\( X_{L} \\), \\( X_{V} \\), and \\( X_{A} \\), respectively (as detailed in Sec. 3.3). We then employ Multi-modal Transformer [36] to aggregate features from any two modalities, extracting joint information, demonstrated here using visual and audio modalities:" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.829, + 0.905, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} Y _ {V} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {V} K _ {A} ^ {\\top}}{\\sqrt {d _ {k}}}\\right) V _ {A} (1) \\\\ = \\operatorname {s o f t m a x} \\left(\\frac {X _ {V} W _ {Q _ {V}} W _ {K _ {A}} ^ {\\top} X _ {A} ^ {\\top}}{\\sqrt {d _ {k}}}\\right) X _ {A} W _ {V _ {A}}, (1) \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24703" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.092, + 0.9, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.267, + 0.908, + 0.325 + ], + "angle": 0, + "content": "Figure 3. Architecture of the Bridging Emotion-Intention via Implicit Label Reasoning (BEAR) framework. The framework consists of two main components: (1) a MulT Former that aggregates multimodal features into a unified representation, and (2) a Bridging Emotion Intention Former that fuses aggregated and prompt features. During training, ground truth emotion and intention labels are encoded to match the answer feature, enabling effective integration of emotion-intention signals and enhanced reasoning capabilities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.484, + 0.504 + ], + "angle": 0, + "content": "where \\( Q_{V} = X_{V}W_{Q_{V}} \\), \\( K_{A} = X_{A}W_{K_{A}} \\), and \\( V_{A} = X_{A}W_{V_{A}} \\) represent Query, Key, and Value component from the Audio modality, respectively. The term \\( W \\) denotes the weights. Here, \\( Y_{V} \\) maintains the same length as \\( Q_{V} \\) (i.e., \\( T_{V} \\)), but exists in the feature space of \\( V_{A} \\). The softmax function in Eq. (1), scaled by \\( \\sqrt{d_k} \\), computes a score matrix \\( \\mathrm{softmax}(\\cdot) \\in \\mathbb{R}^{T_V\\times T_A} \\), with its \\( (i,j) \\)-th entry quantifying the attention from the \\( i \\)-th time step in modality \\( V \\) to the \\( j \\)-th time step in modality \\( A \\). Consequently, each \\( i \\)-th time step of \\( Y_{V} \\) represents a weighted summary of \\( V_{A} \\), weighted by the \\( i \\)-th row in \\( \\mathrm{softmax}(\\cdot) \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.484, + 0.809 + ], + "angle": 0, + "content": "After applying two-pair cross-modal adaptation to three modalities, we obtain the modality aggregated representation \\( f_{ma} \\). However, this approach primarily reveals intention or emotion-related features within the multimodal data, driven by a singular training objective. It overlooks the intrinsic correlations between intentions and emotions that could serve as additional supervisory information. To leverage this relational information, we devise an implicit BEIFormer that integrates all-level intermediate tokens from both the Multi-modal Former, establishing implicit Emotion-Intention associations to enhance the understanding of these subjective dimensions. Specifically, we define the BEI Prompt \\( T_{prom} \\) as a textual modality prior (as shown in Fig. 3), using [MASK] tokens to erase Emotion-Intention related priors. This process generates the hint feature \\( f_{prom} \\) via the BEI Former (a transformer-based subnetwork), serving as the Query input for the BEIFormer. The BEIFormer then utilizes the modality aggregated feature \\( f_{ma} \\) as Key and Value inputs to generate the final interaction output \\( f_{bef} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.904 + ], + "angle": 0, + "content": "Subsequently, we transform the emotion and intention labels into textual information \\( T_{gt} \\) and use text encoder to align these labels within the same representation space as the hint feature, producing the ground truth feature \\( f_{gt} \\). The ground truth feature and the BEIFormer interaction output \\( f_{bef} \\) are then used for contrastive learning during training:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.344, + 0.906, + 0.383 + ], + "angle": 0, + "content": "\\[\n\\mathbf {N C E} \\left(x _ {i}, y _ {i}\\right) = - \\log \\frac {\\exp \\left(x _ {i} ^ {\\top} y _ {i} / \\tau\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(x _ {i} ^ {\\top} y _ {j} / \\tau\\right)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.384, + 0.907, + 0.442 + ], + "angle": 0, + "content": "where \\( x \\) and \\( y \\) are elements from \\( f_{bef} \\) and \\( f_{gt} \\), respectively, \\( B \\) denotes the batch size, and the temperature hyperparameter \\( \\tau \\) is empirically set to 0.05. The full training loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.447, + 0.906, + 0.488 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\left\\{ \\begin{array}{l l} \\mathcal {L} _ {\\mathrm {e m o}} + \\mathcal {L} _ {\\mathrm {N C E}}, & C _ {t} = \\mathrm {e m o}, \\\\ \\mathcal {L} _ {\\text {i n t}} + \\mathcal {L} _ {\\mathrm {N C E}}, & C _ {t} = \\text {i n t}, \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.906, + 0.554 + ], + "angle": 0, + "content": "where \\(C_t\\) determines the current task. In our model, \\(\\mathcal{L}_{\\mathrm{emo/int}}\\) represents the conventional loss for emotion or intention classification. For emotion recognition, we employ the standard multi-class cross-entropy loss:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.563, + 0.906, + 0.605 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {e m o}} = \\frac {1}{N} \\sum_ {i} \\mathcal {L} _ {i} = - \\frac {1}{N} \\sum_ {i} \\sum_ {c = 1} ^ {M} y _ {i c} \\log \\left(p _ {i c}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.907, + 0.78 + ], + "angle": 0, + "content": "where \\(N\\) is the number of samples, \\(M\\) is the number of classes, \\(y_{ic}\\) indicates the ground truth label (1 if the \\(i\\)-th sample belongs to class \\(c\\), and 0 otherwise), and \\(p_{ic}\\) is the predicted probability of the \\(i\\)-th sample belonging to class \\(c\\). This loss function also applies to multi-label intention understanding, ensuring a consistent approach across both emotion and intention classification tasks. Through implicit engagement with semantic correlations between intention and emotion during training, the BEIForner enhances the network's capacity to discern these subjective dimensions, enabling more nuanced understanding." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.798, + 0.805 + ], + "angle": 0, + "content": "4.2. Modality Asynchronous Prompt" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.902 + ], + "angle": 0, + "content": "In addition to the interplay between Emotion-Intention, the challenge of uncertain modalities presents a significant hurdle. Existing methods fail to provide a reasonable solution. For instance, MMIN [50] focuses on learning inter-modal correlations during multimodal data training to infer missing modal information at deployment. However, this strat" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24704" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.09, + 0.482, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.2, + 0.482, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.316, + 0.485, + 0.372 + ], + "angle": 0, + "content": "Figure 4. Comparison of Modal Asynchronous Prompt strategy with MPMM [16]. MAP preserves original representations in existing modalities while only prompting missing modalities, whereas the full modality prompt can corrupt existing features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.382, + 0.483, + 0.502 + ], + "angle": 0, + "content": "egy heavily relies on extensive multimodal priors, a requirement rarely met in social media contexts. While methods like MPMM [3, 16] take an alternative approach by using prompts to integrate missing information into neural networks, their strategy of indiscriminately concatenating all modalities, as shown in Fig. 4, risks compromising feature integrity. This approach can lead to data distortion, reducing model effectiveness in handling missing modalities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.504, + 0.483, + 0.639 + ], + "angle": 0, + "content": "In response, we introduce the Modality Asynchronous Prompt (MAP), which dynamically provides asynchronous cues tailored to various types of missing modality. For each missing type, we develop a unique set of prompts that supply contextually relevant cues, enriching the missing modalities while preserving features of existing modalities. For example, when text and visual inputs are missing, our approach focuses exclusively on enhancing the missing modalities, ensuring that audio features remain undisturbed." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.641, + 0.484, + 0.732 + ], + "angle": 0, + "content": "Furthermore, we integrate these prompts with existing features through element-wise addition (sums prompt features with missing modality features) rather than concatenation, minimizing parameter count and computational overhead by avoiding the additional Multi Layer Perceptron required in concatenation approaches." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.746, + 0.363, + 0.763 + ], + "angle": 0, + "content": "5. Benchmark and Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.771, + 0.291, + 0.787 + ], + "angle": 0, + "content": "5.1. Benchmark Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.794, + 0.483, + 0.854 + ], + "angle": 0, + "content": "For our benchmark, we select established baseline methods in multimodal intention recognition [8, 36, 41] and missing modality challenges [16, 50], chosen for their demonstrated efficacy and influence. They are all trained on MINE." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "MulT. The Multimodal Transformer (MulT) [36] employs the Transformer [39] for learning multimodal language representations through directional cross-modal attention. For" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "consistency, we combine image and video features for visual representation across all multimodal approaches." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.124, + 0.905, + 0.17 + ], + "angle": 0, + "content": "MISA. Modality-Invariant and -Specific Representations (MISA) [8] differentiates modality features into invariant and specific subspaces for enhanced fusion and prediction." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.172, + 0.905, + 0.218 + ], + "angle": 0, + "content": "MFSA. This approach [41] leverages a transformer-based model with adversarial learning to create modality-specific and -agnostic representations for sentiment recognition." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.22, + 0.905, + 0.265 + ], + "angle": 0, + "content": "MMIN. The Missing Modality Imagination Network (MMIN) [50] predicts absent modality representations through joint multimodal representation learning." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.268, + 0.905, + 0.313 + ], + "angle": 0, + "content": "MPMM [16] introduces a prompt-based learning framework, enhancing robustness in visual recognition tasks under various missing-modality conditions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.315, + 0.905, + 0.361 + ], + "angle": 0, + "content": "All missing modality experiments use MulT [36] as the baseline framework for consistency, with detailed configurations that can be found in Appendix F." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.377, + 0.704, + 0.394 + ], + "angle": 0, + "content": "5.2. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.906, + 0.508 + ], + "angle": 0, + "content": "Evaluation Metrics. Following [11, 33], we utilize three multi-label metrics for intention understanding: Macro F1, Micro F1, and Samples F1, chosen for comprehensive performance assessment. Detailed metric definitions appear in Appendix E. For emotion recognition, following [47], we use accuracy, F1 score, precision, and recall to evaluate model performance comprehensively." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.524, + 0.694, + 0.54 + ], + "angle": 0, + "content": "5.3. Main Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.906, + 0.64 + ], + "angle": 0, + "content": "Table 3 presents the performance comparison of various methods on intention and emotion understanding tasks. Our approach consistently outperforms existing techniques across both validation and test splits, demonstrating significant advances in handling missing modalities and mining the Emotion-Intention relationship in social media analysis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.906, + 0.747 + ], + "angle": 0, + "content": "On the validation set, our method achieves a Macro F1 score of \\(36.02\\%\\) for intention understanding, surpassing the MulT baseline by \\(2.65\\%\\). For emotion recognition, it reaches \\(64.82\\%\\) accuracy, exceeding MulT by \\(3.54\\%\\). These improvements persist in the test split, with our method achieving \\(35.79\\%\\) Macro F1 and \\(64.53\\%\\) accuracy, outperforming the baseline by \\(2.68\\%\\) and \\(1.81\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "These results highlight the effectiveness of our strategy, particularly our Bridging Emotion-Intention via Implicit Label Reasoning (BEAR) approach. BEAR significantly enhances the accuracy of intention and emotion recognition, achieving substantial gains over the metrics compared to conventional training methods that treat these tasks separately. This underscores our method's capacity to offer a deeper and more accurate analysis of intention and emotion, validating its superiority in multimodal data analysis within the context of social media." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24705" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.15, + 0.089, + 0.847, + 0.362 + ], + "angle": 0, + "content": "
SplitMethodIntentionEmotion
Macro F1Micro F1Samples F1Acc.F1 ScorePrecisionRecall
ValidationRandom15.1218.0317.694.911.854.822.25
MISA [8]23.8953.3253.6363.9318.6020.2217.55
MFSA [41]31.5553.8853.5264.1519.5620.0717.91
MulT [36]34.3752.4753.1861.2819.2920.8818.79
TCL(Map [51])34.2553.1453.7163.1519.1420.0118.79
MMIN [50]27.9148.8549.5059.3017.4520.6716.97
MPMM [16]32.8852.0753.0063.1517.6521.1017.58
Ours36.0254.0653.8864.8219.8921.3119.85
TestingRandom15.1817.9617.616.252.065.172.00
MISA [8]23.7452.2752.5463.2815.5523.9816.52
MFSA [41]32.4053.7553.0164.0517.6926.5817.91
MulT [36]34.1152.0552.7762.7220.3823.8319.58
TCL(Map [51])34.7351.6452.4163.3120.1425.2719.32
MMIN [50]27.1548.1449.0257.7115.4717.6015.17
MPMM [16]32.8152.1352.7462.8217.7825.2717.94
Ours35.2552.9053.3564.6720.5529.6419.95
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.366, + 0.905, + 0.396 + ], + "angle": 0, + "content": "Table 3. Performance comparison of multimodal fusion and missing modality methods on intention and emotion understanding tasks. Results show consistent improvements of our approach across all metrics on both validation and test sets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.402, + 0.245, + 0.419 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.431, + 0.486, + 0.751 + ], + "angle": 0, + "content": "Component Analysis. Our study delves into the advantages of combining the Modality Asynchronous Prompt (MAP) with the Bridging Emotion-Intention Former (BEIF), offering a detailed examination in Tab. 4. Our results reveal that while each component individually contributes to performance improvement—MAP elevating the Macro F1 score to \\(33.36\\%\\) and BEIF to \\(35.25\\%\\) for intention understanding—it's their combined application that significantly amplifies effectiveness, as evidenced by the compelling outcomes achieved (35.79% Macro F1, 53.92% Micro F1, and 54.88% Samples F1). The ablation study on MAP particularly highlights the effectiveness of our method, especially in addressing the challenge of missing modalities. The asynchronous prompting of missing information, crucially, does not compromise the feature space of the existing modalities. This approach, combined with our preference for element-wise addition over concatenation, ensures the dimensional consistency and integrity of the data representation remain intact. These findings indicate that our holistic approach is adept at offering a nuanced and thorough understanding of complex multimodal data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.908, + 0.706 + ], + "angle": 0, + "content": "Emotion-Intention Correlation Mining. To validate the effectiveness of various strategies for mining intention and emotion correlations, we conduct ablation studies, as shown in Tab. 5. The experiments explore various methods, including a generic multi-task learning paradigm that directly combines emotion and intention losses. This paradigm, while straightforward, shows limited improvement (Macro F1: +1.21%) due to the absence of a specialized Emotion-Intention mechanism during training. It can even harm the ability to understand intention in certain scenarios, with decreases in Micro F1 (-1.06%) and Samples F1 (-0.30%) compared to the baseline. While the Complete Information method (Comp. Info.), which employs one aspect as a prior during both training and inference, exhibits notable gains (+9.06% Macro F1, +9.80% Micro F1), its practical application is limited by available prior knowledge in real-world scenarios. The limitations of these alternative approaches further emphasize the importance of BEAR, which achieves consistent improvements (+2.44% Macro F1) without requiring additional inference-time information." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.718, + 0.909, + 0.75 + ], + "angle": 0, + "content": "Feature Extractor Impact. Beyond model architecture, we investigate how different feature extractors influence in" + }, + { + "type": "table", + "bbox": [ + 0.176, + 0.768, + 0.821, + 0.882 + ], + "angle": 0, + "content": "
ComponentIntentionEmotion
MAPEABEIFMacro F1Micro F1Samples F1Acc.F1 ScorePrecisionRecall
32.8152.1352.7462.8217.7825.2717.94
✓33.3652.7453.4362.7218.9528.4319.30
✓✓34.3752.4753.1861.2819.2925.8818.79
✓35.2552.9053.3564.6720.0229.1919.35
✓✓✓35.7953.9254.8864.5320.5529.6420.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.887, + 0.908, + 0.917 + ], + "angle": 0, + "content": "Table 4. Ablation study showing the contribution of each component. Integration of both Modality Asynchronous Prompt (MAP) and Bridging Emo-Int Former (BEIF) yields optimal performance across all metrics. EA means element-wise addition." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24706" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.486, + 0.258 + ], + "angle": 0, + "content": "
MethodMacro F1Intention
Micro F1Samples F1
Baseline32.8152.1352.74
w/ Multi-task34.0251.0752.44
(+1.21)(-1.06)(-0.30)
w/ EmoInt-Trans [35]31.5952.0651.32
(-1.22)(-0.06)(-1.42)
w/ BEIFormer35.2552.9053.35
(+2.44)(+0.77)(+0.61)
w/ Comp. Info.†41.8761.9362.09
(+9.06)(+9.80)(+8.54)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.264, + 0.485, + 0.307 + ], + "angle": 0, + "content": "Table 5. Advanced relation mining strategies are important in enhancing emotion-intention analysis. †represents the utilization of supplementary information during the inference process." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.316, + 0.482, + 0.481 + ], + "angle": 0, + "content": "tention and emotion understanding performance. We evaluate various state-of-the-art extractors for visual [20, 27], audio [7], and textual [49] modalities. As shown in Tab. 6, more advanced feature extractors consistently improve performance. CLIP-ViT-S [27] provides the most substantial gains for visual features (+1.81% Macro F1), while SemBERT [49] significantly enhances text representation (+1.42% Micro F1). Notably, while AST shows modest improvements in the metrics (+0.76% Macro F1), its impact is more significant than these numbers suggest, given that audio data comprises only a small portion of the test set." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.483, + 0.482, + 0.603 + ], + "angle": 0, + "content": "Generalization Analysis on CMU-MOSEI. To evaluate the generalizability of our MAP approach, we conduct experiments on the CMU-MOSEI dataset across various baseline architectures, as detailed in Tab. 7. MAP demonstrates consistent improvements across all tested configurations. These consistent gains across different architectures and modality combinations validate MAP's robustness in handling missing modalities." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.604, + 0.482, + 0.71 + ], + "angle": 0, + "content": "Data Imbalance. The modality imbalance in MINE is intentional, reflecting the natural, in-the-wild modality distribution of social media content. To evaluate its impact, we conduct experiments with modality resampling. As shown in Tab. 8, resampling (the rare modalities by \\(2\\mathrm{x}\\) and \\(3\\mathrm{x}\\)), especially feature-level contribution resampling [40], can improve the understanding performance." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.725, + 0.34, + 0.74 + ], + "angle": 0, + "content": "6. Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We introduce MINE, a novel dataset capturing the inherent relationship between intention and emotion in natural social media settings. Unlike existing datasets that treat these aspects in isolation or rely on controlled environments, MINE preserves the authentic challenges of real-world social media data—particularly uncertain modalities and diverse thematic contexts. By providing detailed annotations for both intention and emotion across multiple modalities, MINE establishes a comprehensive benchmark for understanding human social behavior in naturalistic settings." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.9, + 0.258 + ], + "angle": 0, + "content": "
MethodMacro F1Intention
Micro F1Samples F1
Baseline32.8152.1352.74
w/ AST [7]33.5752.9553.34
(+0.76)(+0.82)(+0.60)
w/ CLIP-ViT-S [27]34.6253.7054.02
(+1.81)(+1.57)(+1.28)
w/ SwinT-S [20]33.1452.6752.98
(+0.33)(+0.54)(+0.24)
w/ SemBERT [49]33.8953.5554.02
(+1.08)(+1.42)(+1.28)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.575, + 0.263, + 0.844, + 0.277 + ], + "angle": 0, + "content": "Table 6. Impact of various feature extractors." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.302, + 0.9, + 0.476 + ], + "angle": 0, + "content": "
ModelUtilized modality
T, VT, AT
MulT [36]37.1538.2841.91
w/ MAP38.93+1.7839.14+0.8643.35+1.44
MFN [45]58.4860.5459.74
w/ MAP61.33+2.8561.95+1.4163.72+3.98
BERT_MAG [28]62.8361.3962.83
w/ MAP64.51+1.6862.55+1.1664.42+1.59
LMF [19]42.3857.1555.94
w/ MAP44.25+1.8758.67+1.5257.12+1.18
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.48, + 0.905, + 0.509 + ], + "angle": 0, + "content": "Table 7. Evaluation MAP across various architectures and modality combinations on CMU-MOSEI. T: Text, V: Visual, A: Audio." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.528, + 0.882, + 0.572 + ], + "angle": 0, + "content": "
Methodsours2*rare3*rareWei et al [40]
Average51.4551.7951.0853.20
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.577, + 0.904, + 0.591 + ], + "angle": 0, + "content": "Table 8. Results of various methods for resisting modal imbalance." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.611, + 0.905, + 0.822 + ], + "angle": 0, + "content": "To effectively process MINE's complex multimodal data, we propose BEAR (Bridging Emotion-Intention via Implicit Label Reasoning). It addresses two critical challenges: the intrinsic relationship between emotion and intention, and the uncertainty of modalities. The BEAR employs an alignment task with masked emotion and intention categories, enabling implicit mining of their relationships. Meanwhile, the Modality Asynchronous Prompt handles missing modalities through targeted prompting that preserves the integrity of available features. Experimental results demonstrate BEAR's effectiveness, outperforming existing methods across various metrics. In summary, MINE opens new avenues for understanding human social behavior through multimodal content." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.827, + 0.905, + 0.925 + ], + "angle": 0, + "content": "Acknowledgements. This work is partially supported by the National Key Research and Development Program of China (2024YFC3308400), National Natural Science Foundation of China under Grant (62176188, 62361166629), Major Project of Science and Technology Innovation of Hubei Province (2024BCA003) and Key Project of the Health Commission of Hubei Province (WJ2023Z003)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24707" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.483, + 0.157 + ], + "angle": 0, + "content": "[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. In NIPS, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.483, + 0.227 + ], + "angle": 0, + "content": "[2] Carlos Busso, Murtaza Bulut, Chi-Chun Lee, Abe Kazemzadeh, Emily Mower, Samuel Kim, Jeannette N Chang, Sungbok Lee, and Shrikanth S Narayanan. Iemocap: Interactive emotional dyadic motion capture database. Language resources and evaluation, 2008. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.23, + 0.483, + 0.298 + ], + "angle": 0, + "content": "[3] Haozhe Chi, Minghua Yang, Junhao Zhu, Guanhong Wang, and Gaoang Wang. Missing modality meets meta sampling (m3s): An efficient universal approach for multimodal sentiment analysis with missing modality. arXiv, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.301, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[4] Dorottya Demszky, Dana Movshovitz-Attias, Jeongwoo Ko, Alan Cowen, Gaurav Nemade, and Sujith Ravi. Goemotions: A dataset of fine-grained emotions. arXiv, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.344, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.415, + 0.482, + 0.456 + ], + "angle": 0, + "content": "[6] Yiyang Fang, Wenke Huang, Guancheng Wan, Kehua Su, and Mang Ye. Emoe: Modality-specific enhanced dynamic emotion experts. In CVPR, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.459, + 0.482, + 0.486 + ], + "angle": 0, + "content": "[7] Yuan Gong, Yu-An Chung, and James Glass. Ast: Audio spectrogram transformer. arXiv, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.488, + 0.482, + 0.529 + ], + "angle": 0, + "content": "[8] Devamanyu Hazarika, Roger Zimmermann, and Soujanya Poria. Misa: Modality-invariant and-specific representations for multimodal sentiment analysis. In ACM MM, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.482, + 0.584 + ], + "angle": 0, + "content": "[9] Charles T Hemphill, John J Godfrey, and George R Doddington. The atis spoken language systems pilot corpus. In Speech and Natural Language: Proceedings of a Workshop, 1990. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.588, + 0.482, + 0.641 + ], + "angle": 0, + "content": "[10] Ao Jia, Yu He, Yazhou Zhang, Sagar Uprety, Dawei Song, and Christina Lioma. Beyond emotion: A multi-modal dataset for human desire understanding. In NACCL, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.645, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[11] Menglin Jia, Zuxuan Wu, Austin Reiter, Claire Cardie, Serge Belongie, and Ser-Nam Lim. Intentonomy: a dataset and study towards human intent understanding. In CVPR, 2021. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.702, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[12] Kawaljeet Kaur Kapoor, Kuttimani Tamilmani, Nripendra P Rana, Pushp Patil, Yogesh K Dwivedi, and Sridhar Nerur. Advances in social media research: Past, present and future. Information Systems Frontiers, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.759, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[13] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NACCL, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[14] Julia Kruk, Jonah Lubin, Karan Sikka, Xiao Lin, Dan Jurafsky, and Ajay Divakaran. Integrating text and image: Determining multimodal document intent in Instagram posts. In EMNLP-IJCNLP, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[15] Stefan Larson, Anish Mahendran, Joseph J Peper, Christopher Clarke, Andrew Lee, Parker Hill, Jonathan K Kummerfeld, Kevin Leach, Michael A Laurenzano, Lingjia Tang," + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "et al. An evaluation dataset for intent classification and out-of-scope prediction. In EMNLP-IJCNLP, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[16] Yi-Lun Lee, Yi-Hsuan Tsai, Wei-Chen Chiu, and Chen-Yu Lee. Multimodal prompting with missing modalities for visual recognition. In CVPR, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[17] Rui Liu, Haolin Zuo, Zheng Lian, Xiaofen Xing, Bjorn W Schuller, and Haizhou Li. Emotion and intent joint understanding in multimodal conversation: A benchmarking dataset. arXiv, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[18] Xingkun Liu, Arash Eshghi, Pawel Swietojanski, and Verena Rieser. Benchmarking natural language understanding services for building conversational agents. arXiv, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[19] Zhun Liu, Ying Shen, Varun Bharadhwaj Lakshminarasimhan, Paul Pu Liang, Amir Zadeh, and LouisPhilippe Morency. Efficient low-rank multimodal fusion with modality-specific factors. arXiv, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[20] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In CVPR, 2022. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[21] Meng Luo, Hao Fei, Bobo Li, Shengqiong Wu, Qian Liu, Soujanya Poria, Erik Cambria, Mong-Li Lee, and Wynne Hsu. Panosent: A panoptic sextuple extraction benchmark for multimodal conversational aspect-based sentiment analysis. In ACM MM, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[22] Mengmeng Ma, Jian Ren, Long Zhao, Sergey Tulyakov, Cathy Wu, and Xi Peng. Smil: Multimodal learning with severely missing modality. In AAAI, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[23] Mengmeng Ma, Jian Ren, Long Zhao, Davide Testuggine, and Xi Peng. Are multimodal transformers robust to missing modality? In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[24] Kuan-Chuan Peng, Amir Sadovnik, Andrew Gallagher, and Tsuhan Chen. Where do emotions come from? predicting the emotion stimuli map. In IEEE ICIP, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.563, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[25] Wei Peng, Yue Hu, Yuqiang Xie, Luxi Xing, and Yajing Sun. Cogintac: Modeling the relationships between intention, emotion and action in interactive process from cognitive perspective. In 2022 IEEE Congress on Evolutionary Computation, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[26] Soujanya Poria, Devamanyu Hazarika, Navonil Majumder, Gautam Naik, Erik Cambria, and Rada Mihalcea. Meld: A multimodal multi-party dataset for emotion recognition in conversations. In ACL, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[28] Wasifur Rahman, Md Kamrul Hasan, Sangwu Lee, AmirAli Bagher Zadeh, Chengfeng Mao, Louis-Philippe Morency, and Ehsan Hoque. Integrating multimodal information in large pretrained transformers. In ACL, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[29] Md Kaykobad Reza, Ashley Prater-Bennette, and M Salman Asif. Robust multimodal learning with missing modalities via parameter-efficient adaptation. IEEE TPAMI, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[30] Tulika Saha, Aditya Patra, Sriparna Saha, and Pushpak Bhattacharyya. Towards emotion-aided multi-modal dialogue act classification. In ACL, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24708" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[31] Tulika Saha, Aditya Patra, Sriparna Saha, and Pushpak Bhattacharyya. Towards emotion-aided multi-modal dialogue act classification. In ACL, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[32] Tobias Schröder, Terrence C Stewart, and Paul Thagard. Intention, emotion, and action: A neural theory based on semantic pointers. Cognitive science, 2014. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[33] QingHongYa Shi, Mang Ye, Ziyi Zhang, and Bo Du. Learnable hierarchical label embedding and grouping for visual intention understanding. IEEE Transactions on Affective Computing, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.483, + 0.303 + ], + "angle": 0, + "content": "[34] Mohammad Faridul Haque Siddiqui, Parashar Dhakal, Xiaoli Yang, and Ahmad Y Javaid. A survey on databases for multimodal emotion recognition and an introduction to the viri (visible and infrared image) database. Multimodal Technologies and Interaction, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.483, + 0.36 + ], + "angle": 0, + "content": "[35] Gopendra Vikram Singh, Mauajama Firdaus, Asif Ekbal, and Pushpak Bhattacharyya. Emoin-trans: A multimodal transformer for identifying emotions and intents in social conversations. IEEE/ACM TASLP, 2022. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[36] Yao-Hung Hubert Tsai, Shaojie Bai, Paul Pu Liang, J. Zico Kolter, Louis-Philippe Morency, and Ruslan Salakhutdinov. Multimodal transformer for unaligned multimodal language sequences. In ACL, 2019. 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[37] Gokhan Tur, Dilek Hakkani-Tur, and Larry Heck. What is left to be understood in atis? In IEEE Spoken Language Technology Workshop, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[38] Chai M Tyng, Hafeez U Amin, Mohammad NM Saad, and Aamir S Malik. The influences of emotion on learning and memory. Frontiers in psychology, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[40] Yake Wei, Ruoxuan Feng, Zihe Wang, and Di Hu. Enhancing multimodal cooperation via sample-level modality valuation. In CVPR, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[41] Dingkang Yang, Haopeng Kuang, Shuai Huang, and Lihua Zhang. Learning modality-specific and-agnostic representations for asynchronous multimodal language sequences. In ACM MM, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.686 + ], + "angle": 0, + "content": "[42] Qu Yang, Mang Ye, and Bo Du. Emollm: Multimodal emotional understanding meets large language models. arXiv, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[43] Qu Yang, Mang Ye, and Dacheng Tao. Synergy of sight and semantics: visual intention understanding with clip. In ECCV, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[44] Amir Zadeh, Rowan Zellers, Eli Pincus, and Louis-Philippe Morency. Mosi: multimodal corpus of sentiment intensity and subjectivity analysis in online opinion videos. arXiv, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[45] Amir Zadeh, Paul Pu Liang, Navonil Mazumder, Soujanya Poria, Erik Cambria, and Louis-Philippe Morency. Memory fusion network for multi-view sequential learning. In AAAI, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[46] AmirAli Bagher Zadeh, Paul Pu Liang, Soujanya Poria, Erik Cambria, and Louis-Philippe Morency. Multimodal language analysis in the wild: Cmu-mosei dataset and interpretable dynamic fusion graph. In ACL, 2018. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.146 + ], + "angle": 0, + "content": "[47] Hanlei Zhang, Hua Xu, Xin Wang, Qianrui Zhou, Shaojie Zhao, and Jiayan Teng. Mintrec: A new dataset for multimodal intent recognition. In ACM MM, 2022. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[48] Hanlei Zhang, Xin Wang, Hua Xu, Qianrui Zhou, Kai Gao, Jianhua Su, Wenrui Li, Yanting Chen, et al. Mintrec2.0: A large-scale benchmark dataset for multimodal intent recognition and out-of-scope detection in conversations. In ICLR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[49] Zhuosheng Zhang, Yuwei Wu, Hai Zhao, Zuchao Li, Shuai-liang Zhang, Xi Zhou, and Xiang Zhou. Semantics-aware bert for language understanding. In AAAI, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[50] Jinming Zhao, Ruichen Li, and Qin Jin. Missing modality imagination network for emotion recognition with uncertain missing modalities. In ACL, 2021. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[51] Qianrui Zhou, Hua Xu, Hao Li, Hanlei Zhang, Xiaohan Zhang, Yifan Wang, and Kai Gao. Token-level contrastive learning with modality-aware prompting for multimodal intent recognition. In AAAI, 2024. 7" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24709" + } + ] +] \ No newline at end of file diff --git a/2025/Uncertain Multimodal Intention and Emotion Understanding in the Wild/50195041-da23-4b5c-8721-3858d2b20dc9_origin.pdf b/2025/Uncertain Multimodal Intention and Emotion Understanding in the Wild/50195041-da23-4b5c-8721-3858d2b20dc9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f55cb584b35916797957b5a21e6c523788fa7850 --- /dev/null +++ b/2025/Uncertain Multimodal Intention and Emotion Understanding in the Wild/50195041-da23-4b5c-8721-3858d2b20dc9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:79779771c225c1913682034388f40ec4430d1e9a3527b4f333c01de7122bc291 +size 978211 diff --git a/2025/Uncertain Multimodal Intention and Emotion Understanding in the Wild/full.md b/2025/Uncertain Multimodal Intention and Emotion Understanding in the Wild/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8ff368aa9779af9f45e3671ee1944159f776a5d9 --- /dev/null +++ b/2025/Uncertain Multimodal Intention and Emotion Understanding in the Wild/full.md @@ -0,0 +1,286 @@ +# Uncertain Multimodal Intention and Emotion Understanding in the Wild + +Qu Yang $^{1}$ , Qinghongya Shi $^{1}$ , Tongxin Wang $^{1}$ and Mang Ye $^{1,2}$ + +$^{1}$ School of Computer Science, Wuhan University, Wuhan, China + +$^{2}$ Taikang Center for Life and Medical Sciences, Wuhan University, Wuhan, China + +{yangqu, shiqinghongya, wangtx, yemang}@whu.edu.cn + +https://github.com/yan9qu/CVPR25-MINE + +# Abstract + +Understanding intention and emotion from social media poses unique challenges due to the inherent uncertainty in multimodal data, where posts often contain incomplete or missing modalities. While this uncertainty reflects real-world scenarios, it remains underexplored within the computer vision community, particularly in conjunction with the intrinsic relationship between emotion and intention. To address these challenges, we introduce the Multimodal IntentioN and Emotion Understanding in the Wild (MINE) dataset, comprising over 20,000 topic-specific social media posts with natural modality variations across text, image, video, and audio. MINE is distinctively constructed to capture both the uncertain nature of multimodal data and the implicit correlations between intentions and emotions, providing extensive annotations for both aspects. To tackle these scenarios, we propose the Bridging Emotion-Intention via Implicit Label Reasoning (BEAR) framework. BEAR consists of two key components: a BEIFormer that leverages emotion-intention correlations, and a Modality Asynchronous Prompt that handles modality uncertainty. Experiments show that BEAR outperforms existing methods in processing uncertain multimodal data while effectively mining emotion-intention relationships for social media content understanding. + +# 1. Introduction + +Intention and emotion are fundamental in shaping human social behaviors and decision-making processes [25, 32, 38]. Understanding their interplay is crucial for the advancement of artificial general intelligence. While prior studies [6, 42, 43] in the vision community have independently explored intention or emotion recognition in multimodal contexts (as illustrated in Fig. 1), employing datasets like MIntRec [47] and IEMOCAP [2], these datasets fail + +![](images/749a53af190299b99909b56b3b41296da0364779ff4ab30992111945795123fa.jpg) +Figure 1. Overview of multi-modal emotion and intention understanding from social media platforms. + +to capture the inherent link between intention and emotion and overlook the challenges of uncertain modalities and thematic diversity in real-world scenarios. Addressing these gaps is vital for a comprehensive understanding of intentions and emotions in natural environments. In this paper, we introduce the Multimodal Intention and EmotioN Understanding in the Wild (MINE) dataset. MINE provides detailed annotations of both intention and emotion, establishing a platform to explore their correlation. It presents the challenges of uncertain modalities, thus mirroring the intricacies of real-life social interactions, as shown in Fig. 2. + +The design of MINE is guided by several key principles: 1) MINE addresses a critical gap in existing datasets by enabling comprehensive insights into the interactions between intention and emotion within social contexts. 2) In contrast to data typically curated under controlled conditions (such as scripted dialogues or TV series clips [2, 30, 35, 44, 46, 47]), which exhibit significant disparities with real-world situations, MINE leverages authentic user-generated content from social platforms, thus mirroring real-world dynamics of intention and emotion expression. 3) Unlike related datasets that typically assume fixed modalities [2, 47], MINE incorporates the challenge of uncertain modalities in social media, enhancing the realism and applicability of + +![](images/e537ae4431384b95c75fc6e28318bcc76935fd041d1a6f5107a7e4369c3ffea3.jpg) +Figure 2. Overview of the MINE dataset, illustrating the multimodal data collection from social media, categorized by topics and including text, image, video, and audio content, each reflecting various intentions and emotions. The modality combination is uncertain. + +intention and emotion understanding tasks. At the algorithmic level, this manifests as the missing modality problem, where modality heterogeneity is primarily reflected in the absence of certain modal inputs. Previous research on missing modalities [3, 50] often relies on artificial sampling methods, creating biases and noise due to incomplete information (e.g., an image showing “news about something” with the text “That’s so ironic!”, where removing text leads to misinterpretation). MINE, however, preserves native modal information, eliminating potential errors and offering a reliable benchmark for uncertain modality research. 4) By utilizing topical tags (such as #CAMPUS) from social media, MINE provides objective context for analyzing intentions and emotions, enabling a more nuanced understanding of these elements within specific thematic frames. + +To harness the intricate relationship between emotions and intentions within the MINE dataset, we propose Bridging Emotion-Intention via Implicit Label Reasoning (BEAR) framework, comprising two key components: BEIFormer and Modality Asynchronous Prompt (MAP). The BEIFormer introduces an innovative contrastive task, utilizing [MASK] tokens to conceal intention and emotion labels within training prompts. This design enables the model to reconstruct textual data that is in alignment with ground truth labels, leveraging multimodal inputs and the process of unmasking text. This implicit reconstruction and alignment process in BEAR enhances the understanding by delving into the relationship between emotion and intention. Furthermore, the MAP is specifically engineered to tackle the challenge of uncertain modalities. It provides customized cues tailored for diverse uncertain scenarios, thereby substantially improving understanding capacity in contexts where modality data may be incomplete. + +In summary, our contribution is threefold: + +(1) We introduce MINE, the first large-scale dataset that captures the inherent relationship between intention and emotion. Comprising over 20,000 topic-specific posts with natural modality distribution, MINE uniquely provides accurate annotations for intention and emotion, enabling systematic study of their correlations in realistic scenarios. + +(2) We propose Bridging Emotion-Intention via Implicit Label Reasoning, a novel framework that bridges emotion and intention understanding. It incorporates two key components: (i) BEIFormer that implicitly mines emotion-intention relationships through contrastive learning, and (ii) Modality Asynchronous Prompt that handles uncertain modalities without compromising existing features. +(3) We establish comprehensive benchmarks on MINE by evaluating multiple baselines. Through extensive experiments, we demonstrate that BEAR consistently outperforms existing methods by significant margins (e.g., $+2.65\%$ Macro F1 for intention recognition and $+3.54\%$ accuracy for emotion recognition), validating the importance of joint emotion-intention modeling. + +# 2. Related Work + +# 2.1. Intention and Emotion Understanding + +Intention recognition is to understand the goals behind human actions. While early datasets like ATIS [9, 37] were limited to specific domains, recent datasets like CLINC150[15], Intentonomy [11], and MIntRec [47, 48] have expanded to diverse sources including social media and TV series. + +For emotion recognition, datasets have evolved from controlled environments (IEMOCAP [2]) to more naturalistic settings with CMU-MOSEI [46], PanoSent [21], and MELD [26], encompassing multimodal expressions across text, speech, and visual cues. However, as shown in Tab. 1, existing datasets typically focus on either intention or emotion in isolation. While recent works such as EMO-TyDA [31], MC-EIU [17], and EmoInt-MD [35] attempt to combine both aspects, they are constrained by controlled settings and fixed modalities, unlike real-world social media posts which often contain varying combinations of modalities. Our MINE dataset addresses these limitations by capturing authentic emotion-intention relationships while preserving the natural complexity of missing modalities. + +
TypeDatasetModalityMissingClass NumsData NumsSource
####Emo.Int.
IntentionCLINC150 [15]✓✗✗✗✗-15023,700Acted
HWU64 [18]✓✗✗✗✗-6425,716Acted
Intentonomy [11]✓✓✗✗✗-2814,455In the wild
MDID [14]✓✓✗✗✗-81,299In the wild
MSED [10]✓✓✗✗✗-69,190In the wild
MIntRec [47]✓✗✓✓✗-202,224TV
MIntRec2.0 [48]✓✗✓✓✗-3015,040TV
EmotionIEMOCAP [2]✓✗✓✓✗6-7,433Acted
CMU-MOSEI [46]✓✗✓✓✗6-23,453In the wild
GoEmotions [4]✓✗✗✗✗28-58,000In the wild
EmotionRoI [24]✓✗✗✗✗6-1,980In the wild
MELD [26]✓✗✓✓✗8-13,708TV
AllEMOTyDA [31]✓✗✓✓✗101219,365Acted
MINE✓✓✓✓✓112020,168In the wild
+ +Table 1. Comparison of existing datasets for intention (Int), emotion (Emo), and combined (All) understanding. MINE uniquely addresses uncertain modalities while combining both emotion and intention analysis in real-world scenarios. + +# 2.2. Uncertain Modality in the Wild + +Uncertain modality (or missing modality) presents a fundamental challenge in multimodal understanding. Contemporary studies [3, 16, 22, 23, 50] have investigated this challenge using datasets like IEMOCAP [2] and CMUMOSEI [46]. Researchers have proposed various approaches, ranging from automated fusion strategy identification [22] and meta-sampling integration [3] to unified frameworks for modality inference [29, 50] and Bayesian meta-learning applications [23]. + +However, these methods typically rely on random sampling to simulate missing modalities, which can introduce bias and alter the original content meaning (refer to Appendix G). In contrast, MINE captures naturally occurring uncertain modalities in the wild, providing a more authentic foundation for multimodal research. + +# 3. MINE Dataset + +In this section, we outline the creation of the MINE dataset, including steps such as data preparation, establishing categories for intentions and emotions, annotating multimodal content with these categories, and analyzing the data. Detailed metrics of the MINE dataset are provided in Tab. 2. + +# 3.1. Topic Definition and Data Preparation + +To explore how topics influence the understanding of intentions and emotions in real-world scenarios, we obtained permission from the consent of data subjects and collected high-quality, multimodal data from social media platforms—venues teeming with genuine user expressions. This approach significantly diverges from the artificial confines of laboratory settings or the limited diversity found + +in sources like movie clips, frequently utilized in prior research. Our strategy involved organizing a wide range of multimodal data via clusters based on topic-specific hashtags, ensuring a rich and varied data source. + +The MINE dataset comprehensively covers ten topics that are integral to daily life. Following the methodology of Kapoor et al. [12], we selected these topics from the top 100 most frequently discussed subjects on social media platforms. MINE explores these topics in depth, paying particular attention to intersections identified through related keywords. For instance, the keyword 'Education' encompasses associated themes such as 'Social Media Impact on Teachers and Students' and 'Use of Social Media in Academic Learning'. To ensure thematic diversity, we selected ten distinct topics for our study and developed a comprehensive keyword framework for each one, with detailed documentation available in Appendix A. For data collection, we employed a hashtag-based method (e.g., #Education) rather than general topic searches to capture a broader variety of content within each topic. + +The data collection spans Twitter posts from 2019 to 2022, with all content obtained with explicit user consent. To ensure data quality, we implemented multiple preprocessing steps, including the removal of nonsensical texts (those consisting solely of links or random characters) and the filtering of harmful content (related to pornography and drugs). While visual content plays a crucial role in conveying emotional and intentional nuances, we maintained a deliberate balance with textual content to ensure comprehensive multimodal analysis. The final dataset composition reflects this approach, with visual content constituting approximately $70\%$ of the data, creating a diverse yet imbal + +
ItemNumber
Number of intentions20
Number of emotions11
Number of text20044
Number of images12428
Number of videos3895
Number of audios1615
Average length of text utterances (words)20.0
Total number of words489069
Average duration of videos (seconds)24.4
Total duration of videos (hours)21.9
+ +(a) Data Statistics + +
TextImageVideoAudioNumber
✓3913
✓99
✓4
✓✓12261
✓✓2248
✓✓3
✓✓18
✓✓✓26
✓✓✓1557
✓✓✓✓39
+ +(b) Distribution of Modality Combinations + +Table 2. Data and Modality Statistics of the MINE dataset. + +anced multimodal distribution, as detailed in Tab. 2. + +# 3.2. Taxonomy Definition and Data Annotation + +Defining labels for intentions and emotions presents a unique challenge due to their subjective nature. While previous work [47] treated intentions as mutually exclusive categories ('Expressing emotions' or 'Achieving goals'), we argue that social interactions naturally blend both elements, aligning with philosophical theories [32]. + +MINE incorporates 20 intention and 11 emotion labels per data point, drawing from literature on both dialogue and visual intention analysis [11, 47] and emotion research [2, 34]. We frame intention recognition as a multi-label task to capture the nuanced nature of social media interactions, while structuring emotion recognition as a single-label task to identify dominant emotions, consistent with established practices [2]. To address inherent subjectivity, our annotation employs a crowd-sourcing platform with dual annotations. Detailed category definitions and annotation guidelines are provided in Appendices A and H. + +The annotation workflow involved two independent groups of annotators (15 and 13 members) under domain expert supervision. Through iterative training and validation, we reduced the initial annotation discrepancy rate from over $25\%$ to below $3\%$ . The final dataset comprises 16,000 training samples, 2,000 validation samples, and 2,000 test samples, with $20\%$ undergoing additional expert review. Privacy protection measures and detailed annotation protocols are documented in Appendices C and H. + +Analysis of the annotated data reveals strong correlations between intentions and emotions. For instance, aggressive intentions ("Criticize", "Disagree") frequently co-occur with negative emotions. We also observe that certain intentions can complement emotional states - for example, "Comfort" intentions often appear in posts with "Sad" emotions, suggesting an attempt to provide emotional support. Additionally, while topic-intention correlations exist (e.g., "Teach" with education-related content), they are more one-to-one and less pronounced than emotion-intention relationships. Detailed analyses are provided in Appendix B. + +# 3.3. Feature Pre-extraction + +We pre-extract features from each modality (text, image, video, and audio) in the dataset to facilitate multimodal fusion. This approach offers three key advantages: it reduces signal complexity and computational costs for researchers, enables consistent and fair comparisons across different multimodal fusion techniques, and minimizes privacy risks in our data. For feature extraction, we employ state-of-the-art pre-trained models: BERT [13] for text, ViT-S [5] for images, Video Swin Transformer [20] for video, and wav2vec 2.0 [1] for audio. All extracted features are aligned to a dimension of 768 to ensure consistent multimodal fusion. Detailed feature extraction procedures are documented in Appendix C. + +# 4. Proposed Method + +# 4.1. Bridging the Relation of Emotion and Intention + +As previously highlighted, the interconnection of intentions and emotions is important in both real-life situations and neural network analyses. Existing methodologies that focus solely on either intention or emotion often overlook the valuable insights from their interaction, limiting comprehensive understanding in both domains. To bridge this gap, we introduce the Bridging Emotion-Intention Former (BEIFormer), as depicted in Fig. 3. This model unifies these two critical domains. Specifically, we first establish the baseline for multimodal emotion and intention understanding by pre-extracting features from multimodal data to produce text, visual, and audio modality features $X_{L}$ , $X_{V}$ , and $X_{A}$ , respectively (as detailed in Sec. 3.3). We then employ Multi-modal Transformer [36] to aggregate features from any two modalities, extracting joint information, demonstrated here using visual and audio modalities: + +$$ +\begin{array}{l} Y _ {V} = \operatorname {s o f t m a x} \left(\frac {Q _ {V} K _ {A} ^ {\top}}{\sqrt {d _ {k}}}\right) V _ {A} (1) \\ = \operatorname {s o f t m a x} \left(\frac {X _ {V} W _ {Q _ {V}} W _ {K _ {A}} ^ {\top} X _ {A} ^ {\top}}{\sqrt {d _ {k}}}\right) X _ {A} W _ {V _ {A}}, (1) \\ \end{array} +$$ + +![](images/a6cca9e7cb725d8a11ccf96c4182bb6d5d2cf1e23375f69dece666d6d3f6c951.jpg) +Figure 3. Architecture of the Bridging Emotion-Intention via Implicit Label Reasoning (BEAR) framework. The framework consists of two main components: (1) a MulT Former that aggregates multimodal features into a unified representation, and (2) a Bridging Emotion Intention Former that fuses aggregated and prompt features. During training, ground truth emotion and intention labels are encoded to match the answer feature, enabling effective integration of emotion-intention signals and enhanced reasoning capabilities. + +where $Q_{V} = X_{V}W_{Q_{V}}$ , $K_{A} = X_{A}W_{K_{A}}$ , and $V_{A} = X_{A}W_{V_{A}}$ represent Query, Key, and Value component from the Audio modality, respectively. The term $W$ denotes the weights. Here, $Y_{V}$ maintains the same length as $Q_{V}$ (i.e., $T_{V}$ ), but exists in the feature space of $V_{A}$ . The softmax function in Eq. (1), scaled by $\sqrt{d_k}$ , computes a score matrix $\mathrm{softmax}(\cdot) \in \mathbb{R}^{T_V\times T_A}$ , with its $(i,j)$ -th entry quantifying the attention from the $i$ -th time step in modality $V$ to the $j$ -th time step in modality $A$ . Consequently, each $i$ -th time step of $Y_{V}$ represents a weighted summary of $V_{A}$ , weighted by the $i$ -th row in $\mathrm{softmax}(\cdot)$ . + +After applying two-pair cross-modal adaptation to three modalities, we obtain the modality aggregated representation $f_{ma}$ . However, this approach primarily reveals intention or emotion-related features within the multimodal data, driven by a singular training objective. It overlooks the intrinsic correlations between intentions and emotions that could serve as additional supervisory information. To leverage this relational information, we devise an implicit BEIFormer that integrates all-level intermediate tokens from both the Multi-modal Former, establishing implicit Emotion-Intention associations to enhance the understanding of these subjective dimensions. Specifically, we define the BEI Prompt $T_{prom}$ as a textual modality prior (as shown in Fig. 3), using [MASK] tokens to erase Emotion-Intention related priors. This process generates the hint feature $f_{prom}$ via the BEI Former (a transformer-based subnetwork), serving as the Query input for the BEIFormer. The BEIFormer then utilizes the modality aggregated feature $f_{ma}$ as Key and Value inputs to generate the final interaction output $f_{bef}$ . + +Subsequently, we transform the emotion and intention labels into textual information $T_{gt}$ and use text encoder to align these labels within the same representation space as the hint feature, producing the ground truth feature $f_{gt}$ . The ground truth feature and the BEIFormer interaction output $f_{bef}$ are then used for contrastive learning during training: + +$$ +\mathbf {N C E} \left(x _ {i}, y _ {i}\right) = - \log \frac {\exp \left(x _ {i} ^ {\top} y _ {i} / \tau\right)}{\sum_ {j = 1} ^ {B} \exp \left(x _ {i} ^ {\top} y _ {j} / \tau\right)}, \tag {2} +$$ + +where $x$ and $y$ are elements from $f_{bef}$ and $f_{gt}$ , respectively, $B$ denotes the batch size, and the temperature hyperparameter $\tau$ is empirically set to 0.05. The full training loss is defined as: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \left\{ \begin{array}{l l} \mathcal {L} _ {\mathrm {e m o}} + \mathcal {L} _ {\mathrm {N C E}}, & C _ {t} = \mathrm {e m o}, \\ \mathcal {L} _ {\text {i n t}} + \mathcal {L} _ {\mathrm {N C E}}, & C _ {t} = \text {i n t}, \end{array} \right. \tag {3} +$$ + +where $C_t$ determines the current task. In our model, $\mathcal{L}_{\mathrm{emo/int}}$ represents the conventional loss for emotion or intention classification. For emotion recognition, we employ the standard multi-class cross-entropy loss: + +$$ +\mathcal {L} _ {\mathrm {e m o}} = \frac {1}{N} \sum_ {i} \mathcal {L} _ {i} = - \frac {1}{N} \sum_ {i} \sum_ {c = 1} ^ {M} y _ {i c} \log \left(p _ {i c}\right), \tag {4} +$$ + +where $N$ is the number of samples, $M$ is the number of classes, $y_{ic}$ indicates the ground truth label (1 if the $i$ -th sample belongs to class $c$ , and 0 otherwise), and $p_{ic}$ is the predicted probability of the $i$ -th sample belonging to class $c$ . This loss function also applies to multi-label intention understanding, ensuring a consistent approach across both emotion and intention classification tasks. Through implicit engagement with semantic correlations between intention and emotion during training, the BEIForner enhances the network's capacity to discern these subjective dimensions, enabling more nuanced understanding. + +# 4.2. Modality Asynchronous Prompt + +In addition to the interplay between Emotion-Intention, the challenge of uncertain modalities presents a significant hurdle. Existing methods fail to provide a reasonable solution. For instance, MMIN [50] focuses on learning inter-modal correlations during multimodal data training to infer missing modal information at deployment. However, this strat + +![](images/158a6da33e75921e3bf0cf2f10fd9b7c8aff1fb86cff621d29903bba3bb66d81.jpg) + +![](images/4d25cea6030d62ff691a0b4c3f8adcd4e1ff284182c2298379bee9616889104e.jpg) +Figure 4. Comparison of Modal Asynchronous Prompt strategy with MPMM [16]. MAP preserves original representations in existing modalities while only prompting missing modalities, whereas the full modality prompt can corrupt existing features. + +egy heavily relies on extensive multimodal priors, a requirement rarely met in social media contexts. While methods like MPMM [3, 16] take an alternative approach by using prompts to integrate missing information into neural networks, their strategy of indiscriminately concatenating all modalities, as shown in Fig. 4, risks compromising feature integrity. This approach can lead to data distortion, reducing model effectiveness in handling missing modalities. + +In response, we introduce the Modality Asynchronous Prompt (MAP), which dynamically provides asynchronous cues tailored to various types of missing modality. For each missing type, we develop a unique set of prompts that supply contextually relevant cues, enriching the missing modalities while preserving features of existing modalities. For example, when text and visual inputs are missing, our approach focuses exclusively on enhancing the missing modalities, ensuring that audio features remain undisturbed. + +Furthermore, we integrate these prompts with existing features through element-wise addition (sums prompt features with missing modality features) rather than concatenation, minimizing parameter count and computational overhead by avoiding the additional Multi Layer Perceptron required in concatenation approaches. + +# 5. Benchmark and Experiments + +# 5.1. Benchmark Methods + +For our benchmark, we select established baseline methods in multimodal intention recognition [8, 36, 41] and missing modality challenges [16, 50], chosen for their demonstrated efficacy and influence. They are all trained on MINE. + +MulT. The Multimodal Transformer (MulT) [36] employs the Transformer [39] for learning multimodal language representations through directional cross-modal attention. For + +consistency, we combine image and video features for visual representation across all multimodal approaches. + +MISA. Modality-Invariant and -Specific Representations (MISA) [8] differentiates modality features into invariant and specific subspaces for enhanced fusion and prediction. + +MFSA. This approach [41] leverages a transformer-based model with adversarial learning to create modality-specific and -agnostic representations for sentiment recognition. + +MMIN. The Missing Modality Imagination Network (MMIN) [50] predicts absent modality representations through joint multimodal representation learning. + +MPMM [16] introduces a prompt-based learning framework, enhancing robustness in visual recognition tasks under various missing-modality conditions. + +All missing modality experiments use MulT [36] as the baseline framework for consistency, with detailed configurations that can be found in Appendix F. + +# 5.2. Experimental Setup + +Evaluation Metrics. Following [11, 33], we utilize three multi-label metrics for intention understanding: Macro F1, Micro F1, and Samples F1, chosen for comprehensive performance assessment. Detailed metric definitions appear in Appendix E. For emotion recognition, following [47], we use accuracy, F1 score, precision, and recall to evaluate model performance comprehensively. + +# 5.3. Main Experiments + +Table 3 presents the performance comparison of various methods on intention and emotion understanding tasks. Our approach consistently outperforms existing techniques across both validation and test splits, demonstrating significant advances in handling missing modalities and mining the Emotion-Intention relationship in social media analysis. + +On the validation set, our method achieves a Macro F1 score of $36.02\%$ for intention understanding, surpassing the MulT baseline by $2.65\%$ . For emotion recognition, it reaches $64.82\%$ accuracy, exceeding MulT by $3.54\%$ . These improvements persist in the test split, with our method achieving $35.79\%$ Macro F1 and $64.53\%$ accuracy, outperforming the baseline by $2.68\%$ and $1.81\%$ . + +These results highlight the effectiveness of our strategy, particularly our Bridging Emotion-Intention via Implicit Label Reasoning (BEAR) approach. BEAR significantly enhances the accuracy of intention and emotion recognition, achieving substantial gains over the metrics compared to conventional training methods that treat these tasks separately. This underscores our method's capacity to offer a deeper and more accurate analysis of intention and emotion, validating its superiority in multimodal data analysis within the context of social media. + +
SplitMethodIntentionEmotion
Macro F1Micro F1Samples F1Acc.F1 ScorePrecisionRecall
ValidationRandom15.1218.0317.694.911.854.822.25
MISA [8]23.8953.3253.6363.9318.6020.2217.55
MFSA [41]31.5553.8853.5264.1519.5620.0717.91
MulT [36]34.3752.4753.1861.2819.2920.8818.79
TCL(Map [51])34.2553.1453.7163.1519.1420.0118.79
MMIN [50]27.9148.8549.5059.3017.4520.6716.97
MPMM [16]32.8852.0753.0063.1517.6521.1017.58
Ours36.0254.0653.8864.8219.8921.3119.85
TestingRandom15.1817.9617.616.252.065.172.00
MISA [8]23.7452.2752.5463.2815.5523.9816.52
MFSA [41]32.4053.7553.0164.0517.6926.5817.91
MulT [36]34.1152.0552.7762.7220.3823.8319.58
TCL(Map [51])34.7351.6452.4163.3120.1425.2719.32
MMIN [50]27.1548.1449.0257.7115.4717.6015.17
MPMM [16]32.8152.1352.7462.8217.7825.2717.94
Ours35.2552.9053.3564.6720.5529.6419.95
+ +# 5.4. Ablation Study + +Component Analysis. Our study delves into the advantages of combining the Modality Asynchronous Prompt (MAP) with the Bridging Emotion-Intention Former (BEIF), offering a detailed examination in Tab. 4. Our results reveal that while each component individually contributes to performance improvement—MAP elevating the Macro F1 score to $33.36\%$ and BEIF to $35.25\%$ for intention understanding—it's their combined application that significantly amplifies effectiveness, as evidenced by the compelling outcomes achieved (35.79% Macro F1, 53.92% Micro F1, and 54.88% Samples F1). The ablation study on MAP particularly highlights the effectiveness of our method, especially in addressing the challenge of missing modalities. The asynchronous prompting of missing information, crucially, does not compromise the feature space of the existing modalities. This approach, combined with our preference for element-wise addition over concatenation, ensures the dimensional consistency and integrity of the data representation remain intact. These findings indicate that our holistic approach is adept at offering a nuanced and thorough understanding of complex multimodal data. + +Emotion-Intention Correlation Mining. To validate the effectiveness of various strategies for mining intention and emotion correlations, we conduct ablation studies, as shown in Tab. 5. The experiments explore various methods, including a generic multi-task learning paradigm that directly combines emotion and intention losses. This paradigm, while straightforward, shows limited improvement (Macro F1: +1.21%) due to the absence of a specialized Emotion-Intention mechanism during training. It can even harm the ability to understand intention in certain scenarios, with decreases in Micro F1 (-1.06%) and Samples F1 (-0.30%) compared to the baseline. While the Complete Information method (Comp. Info.), which employs one aspect as a prior during both training and inference, exhibits notable gains (+9.06% Macro F1, +9.80% Micro F1), its practical application is limited by available prior knowledge in real-world scenarios. The limitations of these alternative approaches further emphasize the importance of BEAR, which achieves consistent improvements (+2.44% Macro F1) without requiring additional inference-time information. + +Feature Extractor Impact. Beyond model architecture, we investigate how different feature extractors influence in + +Table 3. Performance comparison of multimodal fusion and missing modality methods on intention and emotion understanding tasks. Results show consistent improvements of our approach across all metrics on both validation and test sets. + +
ComponentIntentionEmotion
MAPEABEIFMacro F1Micro F1Samples F1Acc.F1 ScorePrecisionRecall
32.8152.1352.7462.8217.7825.2717.94
✓33.3652.7453.4362.7218.9528.4319.30
✓✓34.3752.4753.1861.2819.2925.8818.79
✓35.2552.9053.3564.6720.0229.1919.35
✓✓✓35.7953.9254.8864.5320.5529.6420.35
+ +Table 4. Ablation study showing the contribution of each component. Integration of both Modality Asynchronous Prompt (MAP) and Bridging Emo-Int Former (BEIF) yields optimal performance across all metrics. EA means element-wise addition. + +
MethodMacro F1Intention
Micro F1Samples F1
Baseline32.8152.1352.74
w/ Multi-task34.0251.0752.44
(+1.21)(-1.06)(-0.30)
w/ EmoInt-Trans [35]31.5952.0651.32
(-1.22)(-0.06)(-1.42)
w/ BEIFormer35.2552.9053.35
(+2.44)(+0.77)(+0.61)
w/ Comp. Info.†41.8761.9362.09
(+9.06)(+9.80)(+8.54)
+ +tention and emotion understanding performance. We evaluate various state-of-the-art extractors for visual [20, 27], audio [7], and textual [49] modalities. As shown in Tab. 6, more advanced feature extractors consistently improve performance. CLIP-ViT-S [27] provides the most substantial gains for visual features (+1.81% Macro F1), while SemBERT [49] significantly enhances text representation (+1.42% Micro F1). Notably, while AST shows modest improvements in the metrics (+0.76% Macro F1), its impact is more significant than these numbers suggest, given that audio data comprises only a small portion of the test set. + +Generalization Analysis on CMU-MOSEI. To evaluate the generalizability of our MAP approach, we conduct experiments on the CMU-MOSEI dataset across various baseline architectures, as detailed in Tab. 7. MAP demonstrates consistent improvements across all tested configurations. These consistent gains across different architectures and modality combinations validate MAP's robustness in handling missing modalities. + +Data Imbalance. The modality imbalance in MINE is intentional, reflecting the natural, in-the-wild modality distribution of social media content. To evaluate its impact, we conduct experiments with modality resampling. As shown in Tab. 8, resampling (the rare modalities by $2\mathrm{x}$ and $3\mathrm{x}$ ), especially feature-level contribution resampling [40], can improve the understanding performance. + +# 6. Conclusion and Discussion + +We introduce MINE, a novel dataset capturing the inherent relationship between intention and emotion in natural social media settings. Unlike existing datasets that treat these aspects in isolation or rely on controlled environments, MINE preserves the authentic challenges of real-world social media data—particularly uncertain modalities and diverse thematic contexts. By providing detailed annotations for both intention and emotion across multiple modalities, MINE establishes a comprehensive benchmark for understanding human social behavior in naturalistic settings. + +Table 5. Advanced relation mining strategies are important in enhancing emotion-intention analysis. †represents the utilization of supplementary information during the inference process. + +
MethodMacro F1Intention
Micro F1Samples F1
Baseline32.8152.1352.74
w/ AST [7]33.5752.9553.34
(+0.76)(+0.82)(+0.60)
w/ CLIP-ViT-S [27]34.6253.7054.02
(+1.81)(+1.57)(+1.28)
w/ SwinT-S [20]33.1452.6752.98
(+0.33)(+0.54)(+0.24)
w/ SemBERT [49]33.8953.5554.02
(+1.08)(+1.42)(+1.28)
+ +Table 6. Impact of various feature extractors. + +
ModelUtilized modality
T, VT, AT
MulT [36]37.1538.2841.91
w/ MAP38.93+1.7839.14+0.8643.35+1.44
MFN [45]58.4860.5459.74
w/ MAP61.33+2.8561.95+1.4163.72+3.98
BERT_MAG [28]62.8361.3962.83
w/ MAP64.51+1.6862.55+1.1664.42+1.59
LMF [19]42.3857.1555.94
w/ MAP44.25+1.8758.67+1.5257.12+1.18
+ +Table 7. Evaluation MAP across various architectures and modality combinations on CMU-MOSEI. T: Text, V: Visual, A: Audio. + +
Methodsours2*rare3*rareWei et al [40]
Average51.4551.7951.0853.20
+ +Table 8. Results of various methods for resisting modal imbalance. + +To effectively process MINE's complex multimodal data, we propose BEAR (Bridging Emotion-Intention via Implicit Label Reasoning). It addresses two critical challenges: the intrinsic relationship between emotion and intention, and the uncertainty of modalities. The BEAR employs an alignment task with masked emotion and intention categories, enabling implicit mining of their relationships. Meanwhile, the Modality Asynchronous Prompt handles missing modalities through targeted prompting that preserves the integrity of available features. Experimental results demonstrate BEAR's effectiveness, outperforming existing methods across various metrics. In summary, MINE opens new avenues for understanding human social behavior through multimodal content. + +Acknowledgements. This work is partially supported by the National Key Research and Development Program of China (2024YFC3308400), National Natural Science Foundation of China under Grant (62176188, 62361166629), Major Project of Science and Technology Innovation of Hubei Province (2024BCA003) and Key Project of the Health Commission of Hubei Province (WJ2023Z003). + +# References + +[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. In NIPS, 2020. 4 +[2] Carlos Busso, Murtaza Bulut, Chi-Chun Lee, Abe Kazemzadeh, Emily Mower, Samuel Kim, Jeannette N Chang, Sungbok Lee, and Shrikanth S Narayanan. Iemocap: Interactive emotional dyadic motion capture database. Language resources and evaluation, 2008. 1, 2, 3, 4 +[3] Haozhe Chi, Minghua Yang, Junhao Zhu, Guanhong Wang, and Gaoang Wang. Missing modality meets meta sampling (m3s): An efficient universal approach for multimodal sentiment analysis with missing modality. arXiv, 2022. 2, 3, 6 +[4] Dorottya Demszky, Dana Movshovitz-Attias, Jeongwoo Ko, Alan Cowen, Gaurav Nemade, and Sujith Ravi. Goemotions: A dataset of fine-grained emotions. arXiv, 2020. 3 +[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv, 2020. 4 +[6] Yiyang Fang, Wenke Huang, Guancheng Wan, Kehua Su, and Mang Ye. Emoe: Modality-specific enhanced dynamic emotion experts. In CVPR, 2025. 1 +[7] Yuan Gong, Yu-An Chung, and James Glass. Ast: Audio spectrogram transformer. arXiv, 2021. 8 +[8] Devamanyu Hazarika, Roger Zimmermann, and Soujanya Poria. Misa: Modality-invariant and-specific representations for multimodal sentiment analysis. In ACM MM, 2020. 6, 7 +[9] Charles T Hemphill, John J Godfrey, and George R Doddington. The atis spoken language systems pilot corpus. In Speech and Natural Language: Proceedings of a Workshop, 1990. 2 +[10] Ao Jia, Yu He, Yazhou Zhang, Sagar Uprety, Dawei Song, and Christina Lioma. Beyond emotion: A multi-modal dataset for human desire understanding. In NACCL, 2022. 3 +[11] Menglin Jia, Zuxuan Wu, Austin Reiter, Claire Cardie, Serge Belongie, and Ser-Nam Lim. Intentonomy: a dataset and study towards human intent understanding. In CVPR, 2021. 2, 3, 4, 6 +[12] Kawaljeet Kaur Kapoor, Kuttimani Tamilmani, Nripendra P Rana, Pushp Patil, Yogesh K Dwivedi, and Sridhar Nerur. Advances in social media research: Past, present and future. Information Systems Frontiers, 2018. 3 +[13] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NACCL, 2019. 4 +[14] Julia Kruk, Jonah Lubin, Karan Sikka, Xiao Lin, Dan Jurafsky, and Ajay Divakaran. Integrating text and image: Determining multimodal document intent in Instagram posts. In EMNLP-IJCNLP, 2019. 3 +[15] Stefan Larson, Anish Mahendran, Joseph J Peper, Christopher Clarke, Andrew Lee, Parker Hill, Jonathan K Kummerfeld, Kevin Leach, Michael A Laurenzano, Lingjia Tang, + +et al. An evaluation dataset for intent classification and out-of-scope prediction. In EMNLP-IJCNLP, 2019. 2, 3 +[16] Yi-Lun Lee, Yi-Hsuan Tsai, Wei-Chen Chiu, and Chen-Yu Lee. Multimodal prompting with missing modalities for visual recognition. In CVPR, 2023. 3, 6, 7 +[17] Rui Liu, Haolin Zuo, Zheng Lian, Xiaofen Xing, Bjorn W Schuller, and Haizhou Li. Emotion and intent joint understanding in multimodal conversation: A benchmarking dataset. arXiv, 2024. 2 +[18] Xingkun Liu, Arash Eshghi, Pawel Swietojanski, and Verena Rieser. Benchmarking natural language understanding services for building conversational agents. arXiv, 2019. 3 +[19] Zhun Liu, Ying Shen, Varun Bharadhwaj Lakshminarasimhan, Paul Pu Liang, Amir Zadeh, and LouisPhilippe Morency. Efficient low-rank multimodal fusion with modality-specific factors. arXiv, 2018. 8 +[20] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In CVPR, 2022. 4, 8 +[21] Meng Luo, Hao Fei, Bobo Li, Shengqiong Wu, Qian Liu, Soujanya Poria, Erik Cambria, Mong-Li Lee, and Wynne Hsu. Panosent: A panoptic sextuple extraction benchmark for multimodal conversational aspect-based sentiment analysis. In ACM MM, 2024. 2 +[22] Mengmeng Ma, Jian Ren, Long Zhao, Sergey Tulyakov, Cathy Wu, and Xi Peng. Smil: Multimodal learning with severely missing modality. In AAAI, 2021. 3 +[23] Mengmeng Ma, Jian Ren, Long Zhao, Davide Testuggine, and Xi Peng. Are multimodal transformers robust to missing modality? In CVPR, 2022. 3 +[24] Kuan-Chuan Peng, Amir Sadovnik, Andrew Gallagher, and Tsuhan Chen. Where do emotions come from? predicting the emotion stimuli map. In IEEE ICIP, 2016. 3 +[25] Wei Peng, Yue Hu, Yuqiang Xie, Luxi Xing, and Yajing Sun. Cogintac: Modeling the relationships between intention, emotion and action in interactive process from cognitive perspective. In 2022 IEEE Congress on Evolutionary Computation, 2022. 1 +[26] Soujanya Poria, Devamanyu Hazarika, Navonil Majumder, Gautam Naik, Erik Cambria, and Rada Mihalcea. Meld: A multimodal multi-party dataset for emotion recognition in conversations. In ACL, 2019. 2, 3 +[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 8 +[28] Wasifur Rahman, Md Kamrul Hasan, Sangwu Lee, AmirAli Bagher Zadeh, Chengfeng Mao, Louis-Philippe Morency, and Ehsan Hoque. Integrating multimodal information in large pretrained transformers. In ACL, 2020. 8 +[29] Md Kaykobad Reza, Ashley Prater-Bennette, and M Salman Asif. Robust multimodal learning with missing modalities via parameter-efficient adaptation. IEEE TPAMI, 2024. 3 +[30] Tulika Saha, Aditya Patra, Sriparna Saha, and Pushpak Bhattacharyya. Towards emotion-aided multi-modal dialogue act classification. In ACL, 2020. 1 + +[31] Tulika Saha, Aditya Patra, Sriparna Saha, and Pushpak Bhattacharyya. Towards emotion-aided multi-modal dialogue act classification. In ACL, 2020. 2, 3 +[32] Tobias Schröder, Terrence C Stewart, and Paul Thagard. Intention, emotion, and action: A neural theory based on semantic pointers. Cognitive science, 2014. 1, 4 +[33] QingHongYa Shi, Mang Ye, Ziyi Zhang, and Bo Du. Learnable hierarchical label embedding and grouping for visual intention understanding. IEEE Transactions on Affective Computing, 2023. 6 +[34] Mohammad Faridul Haque Siddiqui, Parashar Dhakal, Xiaoli Yang, and Ahmad Y Javaid. A survey on databases for multimodal emotion recognition and an introduction to the viri (visible and infrared image) database. Multimodal Technologies and Interaction, 2022. 4 +[35] Gopendra Vikram Singh, Mauajama Firdaus, Asif Ekbal, and Pushpak Bhattacharyya. Emoin-trans: A multimodal transformer for identifying emotions and intents in social conversations. IEEE/ACM TASLP, 2022. 1, 2, 8 +[36] Yao-Hung Hubert Tsai, Shaojie Bai, Paul Pu Liang, J. Zico Kolter, Louis-Philippe Morency, and Ruslan Salakhutdinov. Multimodal transformer for unaligned multimodal language sequences. In ACL, 2019. 4, 6, 7, 8 +[37] Gokhan Tur, Dilek Hakkani-Tur, and Larry Heck. What is left to be understood in atis? In IEEE Spoken Language Technology Workshop, 2010. 2 +[38] Chai M Tyng, Hafeez U Amin, Mohammad NM Saad, and Aamir S Malik. The influences of emotion on learning and memory. Frontiers in psychology, 2017. 1 +[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, 2017. 6 +[40] Yake Wei, Ruoxuan Feng, Zihe Wang, and Di Hu. Enhancing multimodal cooperation via sample-level modality valuation. In CVPR, 2024. 8 +[41] Dingkang Yang, Haopeng Kuang, Shuai Huang, and Lihua Zhang. Learning modality-specific and-agnostic representations for asynchronous multimodal language sequences. In ACM MM, 2022. 6, 7 +[42] Qu Yang, Mang Ye, and Bo Du. Emollm: Multimodal emotional understanding meets large language models. arXiv, 2024. 1 +[43] Qu Yang, Mang Ye, and Dacheng Tao. Synergy of sight and semantics: visual intention understanding with clip. In ECCV, 2024. 1 +[44] Amir Zadeh, Rowan Zellers, Eli Pincus, and Louis-Philippe Morency. Mosi: multimodal corpus of sentiment intensity and subjectivity analysis in online opinion videos. arXiv, 2016. 1 +[45] Amir Zadeh, Paul Pu Liang, Navonil Mazumder, Soujanya Poria, Erik Cambria, and Louis-Philippe Morency. Memory fusion network for multi-view sequential learning. In AAAI, 2018. 8 +[46] AmirAli Bagher Zadeh, Paul Pu Liang, Soujanya Poria, Erik Cambria, and Louis-Philippe Morency. Multimodal language analysis in the wild: Cmu-mosei dataset and interpretable dynamic fusion graph. In ACL, 2018. 1, 2, 3 + +[47] Hanlei Zhang, Hua Xu, Xin Wang, Qianrui Zhou, Shaojie Zhao, and Jiayan Teng. Mintrec: A new dataset for multimodal intent recognition. In ACM MM, 2022. 1, 2, 3, 4, 6 +[48] Hanlei Zhang, Xin Wang, Hua Xu, Qianrui Zhou, Kai Gao, Jianhua Su, Wenrui Li, Yanting Chen, et al. Mintrec2.0: A large-scale benchmark dataset for multimodal intent recognition and out-of-scope detection in conversations. In ICLR, 2024. 2, 3 +[49] Zhuosheng Zhang, Yuwei Wu, Hai Zhao, Zuchao Li, Shuai-liang Zhang, Xi Zhou, and Xiang Zhou. Semantics-aware bert for language understanding. In AAAI, 2020. 8 +[50] Jinming Zhao, Ruichen Li, and Qin Jin. Missing modality imagination network for emotion recognition with uncertain missing modalities. In ACL, 2021. 2, 3, 5, 6, 7 +[51] Qianrui Zhou, Hua Xu, Hao Li, Hanlei Zhang, Xiaohan Zhang, Yifan Wang, and Kai Gao. Token-level contrastive learning with modality-aware prompting for multimodal intent recognition. In AAAI, 2024. 7 \ No newline at end of file diff --git a/2025/Uncertain Multimodal Intention and Emotion Understanding in the Wild/images.zip b/2025/Uncertain Multimodal Intention and Emotion Understanding in the Wild/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c40e18a3ec462ab5872c7214ac2fd017a6095cf8 --- /dev/null +++ b/2025/Uncertain Multimodal Intention and Emotion Understanding in the Wild/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8211fbd9c58c57459f66396f6fc23013b5f8ffd9e07a18d69a954ce42dd3dfc9 +size 570287 diff --git a/2025/Uncertain Multimodal Intention and Emotion Understanding in the Wild/layout.json b/2025/Uncertain Multimodal Intention and Emotion Understanding in the Wild/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..21c57bda081f68fa5c0dc8a4a9bfc822c2fd9614 --- /dev/null +++ b/2025/Uncertain Multimodal Intention and Emotion Understanding in the Wild/layout.json @@ -0,0 +1,7664 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 79, + 103, + 531, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 103, + 531, + 121 + ], + "spans": [ + { + "bbox": [ + 79, + 103, + 531, + 121 + ], + "type": "text", + "content": "Uncertain Multimodal Intention and Emotion Understanding in the Wild" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 148, + 141, + 460, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 141, + 460, + 158 + ], + "spans": [ + { + "bbox": [ + 148, + 141, + 460, + 158 + ], + "type": "text", + "content": "Qu Yang" + }, + { + "bbox": [ + 148, + 141, + 460, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 141, + 460, + 158 + ], + "type": "text", + "content": ", Qinghongya Shi" + }, + { + "bbox": [ + 148, + 141, + 460, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 141, + 460, + 158 + ], + "type": "text", + "content": ", Tongxin Wang" + }, + { + "bbox": [ + 148, + 141, + 460, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 141, + 460, + 158 + ], + "type": "text", + "content": " and Mang Ye" + }, + { + "bbox": [ + 148, + 141, + 460, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 148, + 158, + 462, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 158, + 462, + 171 + ], + "spans": [ + { + "bbox": [ + 148, + 158, + 462, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 148, + 158, + 462, + 171 + ], + "type": "text", + "content": " School of Computer Science, Wuhan University, Wuhan, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 171, + 504, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 171, + 504, + 186 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 504, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 106, + 171, + 504, + 186 + ], + "type": "text", + "content": " Taikang Center for Life and Medical Sciences, Wuhan University, Wuhan, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 170, + 187, + 441, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 187, + 441, + 200 + ], + "spans": [ + { + "bbox": [ + 170, + 187, + 441, + 200 + ], + "type": "text", + "content": "{yangqu, shiqinghongya, wangtx, yemang}@whu.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 201, + 201, + 404, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 201, + 404, + 213 + ], + "spans": [ + { + "bbox": [ + 201, + 201, + 404, + 213 + ], + "type": "text", + "content": "https://github.com/yan9qu/CVPR25-MINE" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 240, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 240, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 240, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 266, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 555 + ], + "type": "text", + "content": "Understanding intention and emotion from social media poses unique challenges due to the inherent uncertainty in multimodal data, where posts often contain incomplete or missing modalities. While this uncertainty reflects real-world scenarios, it remains underexplored within the computer vision community, particularly in conjunction with the intrinsic relationship between emotion and intention. To address these challenges, we introduce the Multimodal IntentioN and Emotion Understanding in the Wild (MINE) dataset, comprising over 20,000 topic-specific social media posts with natural modality variations across text, image, video, and audio. MINE is distinctively constructed to capture both the uncertain nature of multimodal data and the implicit correlations between intentions and emotions, providing extensive annotations for both aspects. To tackle these scenarios, we propose the Bridging Emotion-Intention via Implicit Label Reasoning (BEAR) framework. BEAR consists of two key components: a BEIFormer that leverages emotion-intention correlations, and a Modality Asynchronous Prompt that handles modality uncertainty. Experiments show that BEAR outperforms existing methods in processing uncertain multimodal data while effectively mining emotion-intention relationships for social media content understanding." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 578, + 135, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 135, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 135, + 590 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 598, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 296, + 694 + ], + "type": "text", + "content": "Intention and emotion are fundamental in shaping human social behaviors and decision-making processes [25, 32, 38]. Understanding their interplay is crucial for the advancement of artificial general intelligence. While prior studies [6, 42, 43] in the vision community have independently explored intention or emotion recognition in multimodal contexts (as illustrated in Fig. 1), employing datasets like MIntRec [47] and IEMOCAP [2], these datasets fail" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 318, + 241, + 552, + 385 + ], + "blocks": [ + { + "bbox": [ + 318, + 241, + 552, + 385 + ], + "lines": [ + { + "bbox": [ + 318, + 241, + 552, + 385 + ], + "spans": [ + { + "bbox": [ + 318, + 241, + 552, + 385 + ], + "type": "image", + "image_path": "749a53af190299b99909b56b3b41296da0364779ff4ab30992111945795123fa.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 388, + 555, + 411 + ], + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 411 + ], + "type": "text", + "content": "Figure 1. Overview of multi-modal emotion and intention understanding from social media platforms." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 423, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 555 + ], + "type": "text", + "content": "to capture the inherent link between intention and emotion and overlook the challenges of uncertain modalities and thematic diversity in real-world scenarios. Addressing these gaps is vital for a comprehensive understanding of intentions and emotions in natural environments. In this paper, we introduce the Multimodal Intention and EmotioN Understanding in the Wild (MINE) dataset. MINE provides detailed annotations of both intention and emotion, establishing a platform to explore their correlation. It presents the challenges of uncertain modalities, thus mirroring the intricacies of real-life social interactions, as shown in Fig. 2." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "content": "The design of MINE is guided by several key principles: 1) MINE addresses a critical gap in existing datasets by enabling comprehensive insights into the interactions between intention and emotion within social contexts. 2) In contrast to data typically curated under controlled conditions (such as scripted dialogues or TV series clips [2, 30, 35, 44, 46, 47]), which exhibit significant disparities with real-world situations, MINE leverages authentic user-generated content from social platforms, thus mirroring real-world dynamics of intention and emotion expression. 3) Unlike related datasets that typically assume fixed modalities [2, 47], MINE incorporates the challenge of uncertain modalities in social media, enhancing the realism and applicability of" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 702, + 154, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 702, + 154, + 713 + ], + "spans": [ + { + "bbox": [ + 69, + 702, + 154, + 713 + ], + "type": "text", + "content": "Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24700" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 74, + 504, + 198 + ], + "blocks": [ + { + "bbox": [ + 111, + 74, + 504, + 198 + ], + "lines": [ + { + "bbox": [ + 111, + 74, + 504, + 198 + ], + "spans": [ + { + "bbox": [ + 111, + 74, + 504, + 198 + ], + "type": "image", + "image_path": "e537ae4431384b95c75fc6e28318bcc76935fd041d1a6f5107a7e4369c3ffea3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 200, + 555, + 224 + ], + "lines": [ + { + "bbox": [ + 54, + 200, + 555, + 224 + ], + "spans": [ + { + "bbox": [ + 54, + 200, + 555, + 224 + ], + "type": "text", + "content": "Figure 2. Overview of the MINE dataset, illustrating the multimodal data collection from social media, categorized by topics and including text, image, video, and audio content, each reflecting various intentions and emotions. The modality combination is uncertain." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 229, + 297, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 229, + 297, + 410 + ], + "spans": [ + { + "bbox": [ + 54, + 229, + 297, + 410 + ], + "type": "text", + "content": "intention and emotion understanding tasks. At the algorithmic level, this manifests as the missing modality problem, where modality heterogeneity is primarily reflected in the absence of certain modal inputs. Previous research on missing modalities [3, 50] often relies on artificial sampling methods, creating biases and noise due to incomplete information (e.g., an image showing “news about something” with the text “That’s so ironic!”, where removing text leads to misinterpretation). MINE, however, preserves native modal information, eliminating potential errors and offering a reliable benchmark for uncertain modality research. 4) By utilizing topical tags (such as #CAMPUS) from social media, MINE provides objective context for analyzing intentions and emotions, enabling a more nuanced understanding of these elements within specific thematic frames." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 411, + 296, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 296, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 296, + 626 + ], + "type": "text", + "content": "To harness the intricate relationship between emotions and intentions within the MINE dataset, we propose Bridging Emotion-Intention via Implicit Label Reasoning (BEAR) framework, comprising two key components: BEIFormer and Modality Asynchronous Prompt (MAP). The BEIFormer introduces an innovative contrastive task, utilizing [MASK] tokens to conceal intention and emotion labels within training prompts. This design enables the model to reconstruct textual data that is in alignment with ground truth labels, leveraging multimodal inputs and the process of unmasking text. This implicit reconstruction and alignment process in BEAR enhances the understanding by delving into the relationship between emotion and intention. Furthermore, the MAP is specifically engineered to tackle the challenge of uncertain modalities. It provides customized cues tailored for diverse uncertain scenarios, thereby substantially improving understanding capacity in contexts where modality data may be incomplete." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 628, + 239, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 628, + 239, + 639 + ], + "spans": [ + { + "bbox": [ + 67, + 628, + 239, + 639 + ], + "type": "text", + "content": "In summary, our contribution is threefold:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "type": "text", + "content": "(1) We introduce MINE, the first large-scale dataset that captures the inherent relationship between intention and emotion. Comprising over 20,000 topic-specific posts with natural modality distribution, MINE uniquely provides accurate annotations for intention and emotion, enabling systematic study of their correlations in realistic scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 229, + 556, + 401 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 313, + 229, + 555, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 229, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 555, + 314 + ], + "type": "text", + "content": "(2) We propose Bridging Emotion-Intention via Implicit Label Reasoning, a novel framework that bridges emotion and intention understanding. It incorporates two key components: (i) BEIFormer that implicitly mines emotion-intention relationships through contrastive learning, and (ii) Modality Asynchronous Prompt that handles uncertain modalities without compromising existing features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 317, + 556, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 556, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 556, + 401 + ], + "type": "text", + "content": "(3) We establish comprehensive benchmarks on MINE by evaluating multiple baselines. Through extensive experiments, we demonstrate that BEAR consistently outperforms existing methods by significant margins (e.g., " + }, + { + "bbox": [ + 313, + 317, + 556, + 401 + ], + "type": "inline_equation", + "content": "+2.65\\%" + }, + { + "bbox": [ + 313, + 317, + 556, + 401 + ], + "type": "text", + "content": " Macro F1 for intention recognition and " + }, + { + "bbox": [ + 313, + 317, + 556, + 401 + ], + "type": "inline_equation", + "content": "+3.54\\%" + }, + { + "bbox": [ + 313, + 317, + 556, + 401 + ], + "type": "text", + "content": " accuracy for emotion recognition), validating the importance of joint emotion-intention modeling." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 424, + 401, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 424, + 401, + 437 + ], + "spans": [ + { + "bbox": [ + 314, + 424, + 401, + 437 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 448, + 517, + 461 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 517, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 517, + 461 + ], + "type": "text", + "content": "2.1. Intention and Emotion Understanding" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 470, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 555, + 540 + ], + "type": "text", + "content": "Intention recognition is to understand the goals behind human actions. While early datasets like ATIS [9, 37] were limited to specific domains, recent datasets like CLINC150[15], Intentonomy [11], and MIntRec [47, 48] have expanded to diverse sources including social media and TV series." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "text", + "content": "For emotion recognition, datasets have evolved from controlled environments (IEMOCAP [2]) to more naturalistic settings with CMU-MOSEI [46], PanoSent [21], and MELD [26], encompassing multimodal expressions across text, speech, and visual cues. However, as shown in Tab. 1, existing datasets typically focus on either intention or emotion in isolation. While recent works such as EMO-TyDA [31], MC-EIU [17], and EmoInt-MD [35] attempt to combine both aspects, they are constrained by controlled settings and fixed modalities, unlike real-world social media posts which often contain varying combinations of modalities. Our MINE dataset addresses these limitations by capturing authentic emotion-intention relationships while preserving the natural complexity of missing modalities." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "24701" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 109, + 70, + 503, + 269 + ], + "blocks": [ + { + "bbox": [ + 109, + 70, + 503, + 269 + ], + "lines": [ + { + "bbox": [ + 109, + 70, + 503, + 269 + ], + "spans": [ + { + "bbox": [ + 109, + 70, + 503, + 269 + ], + "type": "table", + "html": "
TypeDatasetModalityMissingClass NumsData NumsSource
####Emo.Int.
IntentionCLINC150 [15]✓✗✗✗✗-15023,700Acted
HWU64 [18]✓✗✗✗✗-6425,716Acted
Intentonomy [11]✓✓✗✗✗-2814,455In the wild
MDID [14]✓✓✗✗✗-81,299In the wild
MSED [10]✓✓✗✗✗-69,190In the wild
MIntRec [47]✓✗✓✓✗-202,224TV
MIntRec2.0 [48]✓✗✓✓✗-3015,040TV
EmotionIEMOCAP [2]✓✗✓✓✗6-7,433Acted
CMU-MOSEI [46]✓✗✓✓✗6-23,453In the wild
GoEmotions [4]✓✗✗✗✗28-58,000In the wild
EmotionRoI [24]✓✗✗✗✗6-1,980In the wild
MELD [26]✓✗✓✓✗8-13,708TV
AllEMOTyDA [31]✓✗✓✓✗101219,365Acted
MINE✓✓✓✓✓112020,168In the wild
", + "image_path": "9b031e4cb4f32ab34cbd179a2ce223f7ed830e914ee2bb939280ec93be1133f3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 277, + 555, + 300 + ], + "lines": [ + { + "bbox": [ + 55, + 277, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 555, + 300 + ], + "type": "text", + "content": "Table 1. Comparison of existing datasets for intention (Int), emotion (Emo), and combined (All) understanding. MINE uniquely addresses uncertain modalities while combining both emotion and intention analysis in real-world scenarios." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 319, + 226, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 226, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 226, + 332 + ], + "type": "text", + "content": "2.2. Uncertain Modality in the Wild" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 337, + 295, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 337, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 54, + 337, + 295, + 445 + ], + "type": "text", + "content": "Uncertain modality (or missing modality) presents a fundamental challenge in multimodal understanding. Contemporary studies [3, 16, 22, 23, 50] have investigated this challenge using datasets like IEMOCAP [2] and CMUMOSEI [46]. Researchers have proposed various approaches, ranging from automated fusion strategy identification [22] and meta-sampling integration [3] to unified frameworks for modality inference [29, 50] and Bayesian meta-learning applications [23]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 445, + 295, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 295, + 515 + ], + "type": "text", + "content": "However, these methods typically rely on random sampling to simulate missing modalities, which can introduce bias and alter the original content meaning (refer to Appendix G). In contrast, MINE captures naturally occurring uncertain modalities in the wild, providing a more authentic foundation for multimodal research." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 526, + 145, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 145, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 145, + 538 + ], + "type": "text", + "content": "3. MINE Dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 546, + 295, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 606 + ], + "type": "text", + "content": "In this section, we outline the creation of the MINE dataset, including steps such as data preparation, establishing categories for intentions and emotions, annotating multimodal content with these categories, and analyzing the data. Detailed metrics of the MINE dataset are provided in Tab. 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 612, + 258, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 258, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 258, + 625 + ], + "type": "text", + "content": "3.1. Topic Definition and Data Preparation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "To explore how topics influence the understanding of intentions and emotions in real-world scenarios, we obtained permission from the consent of data subjects and collected high-quality, multimodal data from social media platforms—venues teeming with genuine user expressions. This approach significantly diverges from the artificial confines of laboratory settings or the limited diversity found" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 320, + 555, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 555, + 369 + ], + "type": "text", + "content": "in sources like movie clips, frequently utilized in prior research. Our strategy involved organizing a wide range of multimodal data via clusters based on topic-specific hashtags, ensuring a rich and varied data source." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 373, + 556, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 556, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 556, + 566 + ], + "type": "text", + "content": "The MINE dataset comprehensively covers ten topics that are integral to daily life. Following the methodology of Kapoor et al. [12], we selected these topics from the top 100 most frequently discussed subjects on social media platforms. MINE explores these topics in depth, paying particular attention to intersections identified through related keywords. For instance, the keyword 'Education' encompasses associated themes such as 'Social Media Impact on Teachers and Students' and 'Use of Social Media in Academic Learning'. To ensure thematic diversity, we selected ten distinct topics for our study and developed a comprehensive keyword framework for each one, with detailed documentation available in Appendix A. For data collection, we employed a hashtag-based method (e.g., #Education) rather than general topic searches to capture a broader variety of content within each topic." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "The data collection spans Twitter posts from 2019 to 2022, with all content obtained with explicit user consent. To ensure data quality, we implemented multiple preprocessing steps, including the removal of nonsensical texts (those consisting solely of links or random characters) and the filtering of harmful content (related to pornography and drugs). While visual content plays a crucial role in conveying emotional and intentional nuances, we maintained a deliberate balance with textual content to ensure comprehensive multimodal analysis. The final dataset composition reflects this approach, with visual content constituting approximately " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " of the data, creating a diverse yet imbal" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24702" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 70, + 286, + 203 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 286, + 203 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 286, + 203 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 286, + 203 + ], + "type": "table", + "html": "
ItemNumber
Number of intentions20
Number of emotions11
Number of text20044
Number of images12428
Number of videos3895
Number of audios1615
Average length of text utterances (words)20.0
Total number of words489069
Average duration of videos (seconds)24.4
Total duration of videos (hours)21.9
", + "image_path": "210fa498550e476eb090992ad03b9dd3a6dca2bd9912c30460c64d409de358f8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 343, + 70, + 518, + 204 + ], + "blocks": [ + { + "bbox": [ + 151, + 205, + 211, + 214 + ], + "lines": [ + { + "bbox": [ + 151, + 205, + 211, + 214 + ], + "spans": [ + { + "bbox": [ + 151, + 205, + 211, + 214 + ], + "type": "text", + "content": "(a) Data Statistics" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 343, + 70, + 518, + 204 + ], + "lines": [ + { + "bbox": [ + 343, + 70, + 518, + 204 + ], + "spans": [ + { + "bbox": [ + 343, + 70, + 518, + 204 + ], + "type": "table", + "html": "
TextImageVideoAudioNumber
✓3913
✓99
✓4
✓✓12261
✓✓2248
✓✓3
✓✓18
✓✓✓26
✓✓✓1557
✓✓✓✓39
", + "image_path": "30be5adc7b0e4a197dc632727ffb8a7789550089f8ee9fa612b3f8bfce119a1f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 361, + 205, + 500, + 215 + ], + "lines": [ + { + "bbox": [ + 361, + 205, + 500, + 215 + ], + "spans": [ + { + "bbox": [ + 361, + 205, + 500, + 215 + ], + "type": "text", + "content": "(b) Distribution of Modality Combinations" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 197, + 217, + 411, + 228 + ], + "lines": [ + { + "bbox": [ + 197, + 217, + 411, + 228 + ], + "spans": [ + { + "bbox": [ + 197, + 217, + 411, + 228 + ], + "type": "text", + "content": "Table 2. Data and Modality Statistics of the MINE dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 236, + 265, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 265, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 265, + 247 + ], + "type": "text", + "content": "anced multimodal distribution, as detailed in Tab. 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 253, + 278, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 278, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 278, + 266 + ], + "type": "text", + "content": "3.2. Taxonomy Definition and Data Annotation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 270, + 296, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 342 + ], + "type": "text", + "content": "Defining labels for intentions and emotions presents a unique challenge due to their subjective nature. While previous work [47] treated intentions as mutually exclusive categories ('Expressing emotions' or 'Achieving goals'), we argue that social interactions naturally blend both elements, aligning with philosophical theories [32]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 342, + 296, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 474 + ], + "type": "text", + "content": "MINE incorporates 20 intention and 11 emotion labels per data point, drawing from literature on both dialogue and visual intention analysis [11, 47] and emotion research [2, 34]. We frame intention recognition as a multi-label task to capture the nuanced nature of social media interactions, while structuring emotion recognition as a single-label task to identify dominant emotions, consistent with established practices [2]. To address inherent subjectivity, our annotation employs a crowd-sourcing platform with dual annotations. Detailed category definitions and annotation guidelines are provided in Appendices A and H." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 474, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 296, + 581 + ], + "type": "text", + "content": "The annotation workflow involved two independent groups of annotators (15 and 13 members) under domain expert supervision. Through iterative training and validation, we reduced the initial annotation discrepancy rate from over " + }, + { + "bbox": [ + 55, + 474, + 296, + 581 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 55, + 474, + 296, + 581 + ], + "type": "text", + "content": " to below " + }, + { + "bbox": [ + 55, + 474, + 296, + 581 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 55, + 474, + 296, + 581 + ], + "type": "text", + "content": ". The final dataset comprises 16,000 training samples, 2,000 validation samples, and 2,000 test samples, with " + }, + { + "bbox": [ + 55, + 474, + 296, + 581 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 474, + 296, + 581 + ], + "type": "text", + "content": " undergoing additional expert review. Privacy protection measures and detailed annotation protocols are documented in Appendices C and H." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "Analysis of the annotated data reveals strong correlations between intentions and emotions. For instance, aggressive intentions (\"Criticize\", \"Disagree\") frequently co-occur with negative emotions. We also observe that certain intentions can complement emotional states - for example, \"Comfort\" intentions often appear in posts with \"Sad\" emotions, suggesting an attempt to provide emotional support. Additionally, while topic-intention correlations exist (e.g., \"Teach\" with education-related content), they are more one-to-one and less pronounced than emotion-intention relationships. Detailed analyses are provided in Appendix B." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 235, + 443, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 443, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 443, + 247 + ], + "type": "text", + "content": "3.3. Feature Pre-extraction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 252, + 555, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 252, + 555, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 252, + 555, + 408 + ], + "type": "text", + "content": "We pre-extract features from each modality (text, image, video, and audio) in the dataset to facilitate multimodal fusion. This approach offers three key advantages: it reduces signal complexity and computational costs for researchers, enables consistent and fair comparisons across different multimodal fusion techniques, and minimizes privacy risks in our data. For feature extraction, we employ state-of-the-art pre-trained models: BERT [13] for text, ViT-S [5] for images, Video Swin Transformer [20] for video, and wav2vec 2.0 [1] for audio. All extracted features are aligned to a dimension of 768 to ensure consistent multimodal fusion. Detailed feature extraction procedures are documented in Appendix C." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 418, + 421, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 421, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 421, + 432 + ], + "type": "text", + "content": "4. Proposed Method" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 437, + 554, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 451 + ], + "type": "text", + "content": "4.1. Bridging the Relation of Emotion and Intention" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 455, + 555, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 555, + 647 + ], + "type": "text", + "content": "As previously highlighted, the interconnection of intentions and emotions is important in both real-life situations and neural network analyses. Existing methodologies that focus solely on either intention or emotion often overlook the valuable insights from their interaction, limiting comprehensive understanding in both domains. To bridge this gap, we introduce the Bridging Emotion-Intention Former (BEIFormer), as depicted in Fig. 3. This model unifies these two critical domains. Specifically, we first establish the baseline for multimodal emotion and intention understanding by pre-extracting features from multimodal data to produce text, visual, and audio modality features " + }, + { + "bbox": [ + 313, + 455, + 555, + 647 + ], + "type": "inline_equation", + "content": "X_{L}" + }, + { + "bbox": [ + 313, + 455, + 555, + 647 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 455, + 555, + 647 + ], + "type": "inline_equation", + "content": "X_{V}" + }, + { + "bbox": [ + 313, + 455, + 555, + 647 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 455, + 555, + 647 + ], + "type": "inline_equation", + "content": "X_{A}" + }, + { + "bbox": [ + 313, + 455, + 555, + 647 + ], + "type": "text", + "content": ", respectively (as detailed in Sec. 3.3). We then employ Multi-modal Transformer [36] to aggregate features from any two modalities, extracting joint information, demonstrated here using visual and audio modalities:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 332, + 656, + 553, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 656, + 553, + 716 + ], + "spans": [ + { + "bbox": [ + 332, + 656, + 553, + 716 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} Y _ {V} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {V} K _ {A} ^ {\\top}}{\\sqrt {d _ {k}}}\\right) V _ {A} (1) \\\\ = \\operatorname {s o f t m a x} \\left(\\frac {X _ {V} W _ {Q _ {V}} W _ {K _ {A}} ^ {\\top} X _ {A} ^ {\\top}}{\\sqrt {d _ {k}}}\\right) X _ {A} W _ {V _ {A}}, (1) \\\\ \\end{array}", + "image_path": "a013a980e9a2fa387b2f5b594ab828fb9a74bfdf314724e257789f853171df32.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24703" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 72, + 550, + 208 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 550, + 208 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 550, + 208 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 550, + 208 + ], + "type": "image", + "image_path": "a6cca9e7cb725d8a11ccf96c4182bb6d5d2cf1e23375f69dece666d6d3f6c951.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 211, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 211, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 555, + 257 + ], + "type": "text", + "content": "Figure 3. Architecture of the Bridging Emotion-Intention via Implicit Label Reasoning (BEAR) framework. The framework consists of two main components: (1) a MulT Former that aggregates multimodal features into a unified representation, and (2) a Bridging Emotion Intention Former that fuses aggregated and prompt features. During training, ground truth emotion and intention labels are encoded to match the answer feature, enabling effective integration of emotion-intention signals and enhanced reasoning capabilities." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "inline_equation", + "content": "Q_{V} = X_{V}W_{Q_{V}}" + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "inline_equation", + "content": "K_{A} = X_{A}W_{K_{A}}" + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "inline_equation", + "content": "V_{A} = X_{A}W_{V_{A}}" + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "text", + "content": " represent Query, Key, and Value component from the Audio modality, respectively. The term " + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "text", + "content": " denotes the weights. Here, " + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "inline_equation", + "content": "Y_{V}" + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "text", + "content": " maintains the same length as " + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "inline_equation", + "content": "Q_{V}" + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "inline_equation", + "content": "T_{V}" + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "text", + "content": "), but exists in the feature space of " + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "inline_equation", + "content": "V_{A}" + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "text", + "content": ". The softmax function in Eq. (1), scaled by " + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "inline_equation", + "content": "\\sqrt{d_k}" + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "text", + "content": ", computes a score matrix " + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "inline_equation", + "content": "\\mathrm{softmax}(\\cdot) \\in \\mathbb{R}^{T_V\\times T_A}" + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "text", + "content": ", with its " + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "text", + "content": "-th entry quantifying the attention from the " + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "text", + "content": "-th time step in modality " + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "text", + "content": " to the " + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "text", + "content": "-th time step in modality " + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "text", + "content": ". Consequently, each " + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "text", + "content": "-th time step of " + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "inline_equation", + "content": "Y_{V}" + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "text", + "content": " represents a weighted summary of " + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "inline_equation", + "content": "V_{A}" + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "text", + "content": ", weighted by the " + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "text", + "content": "-th row in " + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "inline_equation", + "content": "\\mathrm{softmax}(\\cdot)" + }, + { + "bbox": [ + 55, + 266, + 296, + 399 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 400, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 296, + 640 + ], + "type": "text", + "content": "After applying two-pair cross-modal adaptation to three modalities, we obtain the modality aggregated representation " + }, + { + "bbox": [ + 55, + 400, + 296, + 640 + ], + "type": "inline_equation", + "content": "f_{ma}" + }, + { + "bbox": [ + 55, + 400, + 296, + 640 + ], + "type": "text", + "content": ". However, this approach primarily reveals intention or emotion-related features within the multimodal data, driven by a singular training objective. It overlooks the intrinsic correlations between intentions and emotions that could serve as additional supervisory information. To leverage this relational information, we devise an implicit BEIFormer that integrates all-level intermediate tokens from both the Multi-modal Former, establishing implicit Emotion-Intention associations to enhance the understanding of these subjective dimensions. Specifically, we define the BEI Prompt " + }, + { + "bbox": [ + 55, + 400, + 296, + 640 + ], + "type": "inline_equation", + "content": "T_{prom}" + }, + { + "bbox": [ + 55, + 400, + 296, + 640 + ], + "type": "text", + "content": " as a textual modality prior (as shown in Fig. 3), using [MASK] tokens to erase Emotion-Intention related priors. This process generates the hint feature " + }, + { + "bbox": [ + 55, + 400, + 296, + 640 + ], + "type": "inline_equation", + "content": "f_{prom}" + }, + { + "bbox": [ + 55, + 400, + 296, + 640 + ], + "type": "text", + "content": " via the BEI Former (a transformer-based subnetwork), serving as the Query input for the BEIFormer. The BEIFormer then utilizes the modality aggregated feature " + }, + { + "bbox": [ + 55, + 400, + 296, + 640 + ], + "type": "inline_equation", + "content": "f_{ma}" + }, + { + "bbox": [ + 55, + 400, + 296, + 640 + ], + "type": "text", + "content": " as Key and Value inputs to generate the final interaction output " + }, + { + "bbox": [ + 55, + 400, + 296, + 640 + ], + "type": "inline_equation", + "content": "f_{bef}" + }, + { + "bbox": [ + 55, + 400, + 296, + 640 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": "Subsequently, we transform the emotion and intention labels into textual information " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "T_{gt}" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": " and use text encoder to align these labels within the same representation space as the hint feature, producing the ground truth feature " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "f_{gt}" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": ". The ground truth feature and the BEIFormer interaction output " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "f_{bef}" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": " are then used for contrastive learning during training:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 345, + 272, + 554, + 303 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 272, + 554, + 303 + ], + "spans": [ + { + "bbox": [ + 345, + 272, + 554, + 303 + ], + "type": "interline_equation", + "content": "\\mathbf {N C E} \\left(x _ {i}, y _ {i}\\right) = - \\log \\frac {\\exp \\left(x _ {i} ^ {\\top} y _ {i} / \\tau\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(x _ {i} ^ {\\top} y _ {j} / \\tau\\right)}, \\tag {2}", + "image_path": "09a690b6ef5d1c46741b71ba8e07be6b8268de650044de818c46bc9855637f4a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 304, + 555, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 350 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 304, + 555, + 350 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 304, + 555, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 304, + 555, + 350 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 304, + 555, + 350 + ], + "type": "text", + "content": " are elements from " + }, + { + "bbox": [ + 313, + 304, + 555, + 350 + ], + "type": "inline_equation", + "content": "f_{bef}" + }, + { + "bbox": [ + 313, + 304, + 555, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 304, + 555, + 350 + ], + "type": "inline_equation", + "content": "f_{gt}" + }, + { + "bbox": [ + 313, + 304, + 555, + 350 + ], + "type": "text", + "content": ", respectively, " + }, + { + "bbox": [ + 313, + 304, + 555, + 350 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 304, + 555, + 350 + ], + "type": "text", + "content": " denotes the batch size, and the temperature hyperparameter " + }, + { + "bbox": [ + 313, + 304, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 304, + 555, + 350 + ], + "type": "text", + "content": " is empirically set to 0.05. The full training loss is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 356, + 354, + 554, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 354, + 554, + 386 + ], + "spans": [ + { + "bbox": [ + 356, + 354, + 554, + 386 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\left\\{ \\begin{array}{l l} \\mathcal {L} _ {\\mathrm {e m o}} + \\mathcal {L} _ {\\mathrm {N C E}}, & C _ {t} = \\mathrm {e m o}, \\\\ \\mathcal {L} _ {\\text {i n t}} + \\mathcal {L} _ {\\mathrm {N C E}}, & C _ {t} = \\text {i n t}, \\end{array} \\right. \\tag {3}", + "image_path": "6d8f6949bc8e8e717675fac751381b2474dc57262f1057bc6a9c70a576c8e39d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 390, + 554, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 554, + 438 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 554, + 438 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 390, + 554, + 438 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 313, + 390, + 554, + 438 + ], + "type": "text", + "content": " determines the current task. In our model, " + }, + { + "bbox": [ + 313, + 390, + 554, + 438 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{emo/int}}" + }, + { + "bbox": [ + 313, + 390, + 554, + 438 + ], + "type": "text", + "content": " represents the conventional loss for emotion or intention classification. For emotion recognition, we employ the standard multi-class cross-entropy loss:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 339, + 445, + 554, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 445, + 554, + 479 + ], + "spans": [ + { + "bbox": [ + 339, + 445, + 554, + 479 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {e m o}} = \\frac {1}{N} \\sum_ {i} \\mathcal {L} _ {i} = - \\frac {1}{N} \\sum_ {i} \\sum_ {c = 1} ^ {M} y _ {i c} \\log \\left(p _ {i c}\\right), \\tag {4}", + "image_path": "d49f63415473b16e231f0f8c316a3ae9b43331e2ccda7b5f786b82328c729083.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 485, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 555, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 485, + 555, + 617 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 485, + 555, + 617 + ], + "type": "text", + "content": " is the number of samples, " + }, + { + "bbox": [ + 313, + 485, + 555, + 617 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 485, + 555, + 617 + ], + "type": "text", + "content": " is the number of classes, " + }, + { + "bbox": [ + 313, + 485, + 555, + 617 + ], + "type": "inline_equation", + "content": "y_{ic}" + }, + { + "bbox": [ + 313, + 485, + 555, + 617 + ], + "type": "text", + "content": " indicates the ground truth label (1 if the " + }, + { + "bbox": [ + 313, + 485, + 555, + 617 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 485, + 555, + 617 + ], + "type": "text", + "content": "-th sample belongs to class " + }, + { + "bbox": [ + 313, + 485, + 555, + 617 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 485, + 555, + 617 + ], + "type": "text", + "content": ", and 0 otherwise), and " + }, + { + "bbox": [ + 313, + 485, + 555, + 617 + ], + "type": "inline_equation", + "content": "p_{ic}" + }, + { + "bbox": [ + 313, + 485, + 555, + 617 + ], + "type": "text", + "content": " is the predicted probability of the " + }, + { + "bbox": [ + 313, + 485, + 555, + 617 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 485, + 555, + 617 + ], + "type": "text", + "content": "-th sample belonging to class " + }, + { + "bbox": [ + 313, + 485, + 555, + 617 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 485, + 555, + 617 + ], + "type": "text", + "content": ". This loss function also applies to multi-label intention understanding, ensuring a consistent approach across both emotion and intention classification tasks. Through implicit engagement with semantic correlations between intention and emotion during training, the BEIForner enhances the network's capacity to discern these subjective dimensions, enabling more nuanced understanding." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 624, + 488, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 488, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 488, + 637 + ], + "type": "text", + "content": "4.2. Modality Asynchronous Prompt" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": "In addition to the interplay between Emotion-Intention, the challenge of uncertain modalities presents a significant hurdle. Existing methods fail to provide a reasonable solution. For instance, MMIN [50] focuses on learning inter-modal correlations during multimodal data training to infer missing modal information at deployment. However, this strat" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24704" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 71, + 294, + 153 + ], + "blocks": [ + { + "bbox": [ + 63, + 71, + 294, + 153 + ], + "lines": [ + { + "bbox": [ + 63, + 71, + 294, + 153 + ], + "spans": [ + { + "bbox": [ + 63, + 71, + 294, + 153 + ], + "type": "image", + "image_path": "158a6da33e75921e3bf0cf2f10fd9b7c8aff1fb86cff621d29903bba3bb66d81.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 63, + 158, + 294, + 240 + ], + "blocks": [ + { + "bbox": [ + 63, + 158, + 294, + 240 + ], + "lines": [ + { + "bbox": [ + 63, + 158, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 63, + 158, + 294, + 240 + ], + "type": "image", + "image_path": "4d25cea6030d62ff691a0b4c3f8adcd4e1ff284182c2298379bee9616889104e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 250, + 296, + 294 + ], + "lines": [ + { + "bbox": [ + 55, + 250, + 296, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 296, + 294 + ], + "type": "text", + "content": "Figure 4. Comparison of Modal Asynchronous Prompt strategy with MPMM [16]. MAP preserves original representations in existing modalities while only prompting missing modalities, whereas the full modality prompt can corrupt existing features." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 302, + 295, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 302, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 295, + 397 + ], + "type": "text", + "content": "egy heavily relies on extensive multimodal priors, a requirement rarely met in social media contexts. While methods like MPMM [3, 16] take an alternative approach by using prompts to integrate missing information into neural networks, their strategy of indiscriminately concatenating all modalities, as shown in Fig. 4, risks compromising feature integrity. This approach can lead to data distortion, reducing model effectiveness in handling missing modalities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 399, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 295, + 506 + ], + "type": "text", + "content": "In response, we introduce the Modality Asynchronous Prompt (MAP), which dynamically provides asynchronous cues tailored to various types of missing modality. For each missing type, we develop a unique set of prompts that supply contextually relevant cues, enriching the missing modalities while preserving features of existing modalities. For example, when text and visual inputs are missing, our approach focuses exclusively on enhancing the missing modalities, ensuring that audio features remain undisturbed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 507, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 296, + 579 + ], + "type": "text", + "content": "Furthermore, we integrate these prompts with existing features through element-wise addition (sums prompt features with missing modality features) rather than concatenation, minimizing parameter count and computational overhead by avoiding the additional Multi Layer Perceptron required in concatenation approaches." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 590, + 222, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 222, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 222, + 604 + ], + "type": "text", + "content": "5. Benchmark and Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 610, + 178, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 178, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 178, + 623 + ], + "type": "text", + "content": "5.1. Benchmark Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 628, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 295, + 676 + ], + "type": "text", + "content": "For our benchmark, we select established baseline methods in multimodal intention recognition [8, 36, 41] and missing modality challenges [16, 50], chosen for their demonstrated efficacy and influence. They are all trained on MINE." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "MulT. The Multimodal Transformer (MulT) [36] employs the Transformer [39] for learning multimodal language representations through directional cross-modal attention. For" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "consistency, we combine image and video features for visual representation across all multimodal approaches." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 98, + 553, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 98, + 553, + 134 + ], + "spans": [ + { + "bbox": [ + 313, + 98, + 553, + 134 + ], + "type": "text", + "content": "MISA. Modality-Invariant and -Specific Representations (MISA) [8] differentiates modality features into invariant and specific subspaces for enhanced fusion and prediction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 136, + 553, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 136, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 313, + 136, + 553, + 172 + ], + "type": "text", + "content": "MFSA. This approach [41] leverages a transformer-based model with adversarial learning to create modality-specific and -agnostic representations for sentiment recognition." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 174, + 553, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 174, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 553, + 209 + ], + "type": "text", + "content": "MMIN. The Missing Modality Imagination Network (MMIN) [50] predicts absent modality representations through joint multimodal representation learning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 212, + 553, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 553, + 247 + ], + "type": "text", + "content": "MPMM [16] introduces a prompt-based learning framework, enhancing robustness in visual recognition tasks under various missing-modality conditions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 249, + 553, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 553, + 285 + ], + "type": "text", + "content": "All missing modality experiments use MulT [36] as the baseline framework for consistency, with detailed configurations that can be found in Appendix F." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 298, + 430, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 298, + 430, + 312 + ], + "spans": [ + { + "bbox": [ + 314, + 298, + 430, + 312 + ], + "type": "text", + "content": "5.2. Experimental Setup" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 318, + 554, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 554, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 554, + 402 + ], + "type": "text", + "content": "Evaluation Metrics. Following [11, 33], we utilize three multi-label metrics for intention understanding: Macro F1, Micro F1, and Samples F1, chosen for comprehensive performance assessment. Detailed metric definitions appear in Appendix E. For emotion recognition, following [47], we use accuracy, F1 score, precision, and recall to evaluate model performance comprehensively." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 415, + 424, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 415, + 424, + 427 + ], + "spans": [ + { + "bbox": [ + 314, + 415, + 424, + 427 + ], + "type": "text", + "content": "5.3. Main Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 434, + 554, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 554, + 506 + ], + "type": "text", + "content": "Table 3 presents the performance comparison of various methods on intention and emotion understanding tasks. Our approach consistently outperforms existing techniques across both validation and test splits, demonstrating significant advances in handling missing modalities and mining the Emotion-Intention relationship in social media analysis." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "type": "text", + "content": "On the validation set, our method achieves a Macro F1 score of " + }, + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "type": "inline_equation", + "content": "36.02\\%" + }, + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "type": "text", + "content": " for intention understanding, surpassing the MulT baseline by " + }, + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "type": "inline_equation", + "content": "2.65\\%" + }, + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "type": "text", + "content": ". For emotion recognition, it reaches " + }, + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "type": "inline_equation", + "content": "64.82\\%" + }, + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "type": "text", + "content": " accuracy, exceeding MulT by " + }, + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "type": "inline_equation", + "content": "3.54\\%" + }, + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "type": "text", + "content": ". These improvements persist in the test split, with our method achieving " + }, + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "type": "inline_equation", + "content": "35.79\\%" + }, + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "type": "text", + "content": " Macro F1 and " + }, + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "type": "inline_equation", + "content": "64.53\\%" + }, + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "type": "text", + "content": " accuracy, outperforming the baseline by " + }, + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "type": "inline_equation", + "content": "2.68\\%" + }, + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "type": "inline_equation", + "content": "1.81\\%" + }, + { + "bbox": [ + 313, + 508, + 554, + 591 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "These results highlight the effectiveness of our strategy, particularly our Bridging Emotion-Intention via Implicit Label Reasoning (BEAR) approach. BEAR significantly enhances the accuracy of intention and emotion recognition, achieving substantial gains over the metrics compared to conventional training methods that treat these tasks separately. This underscores our method's capacity to offer a deeper and more accurate analysis of intention and emotion, validating its superiority in multimodal data analysis within the context of social media." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24705" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 91, + 70, + 518, + 286 + ], + "blocks": [ + { + "bbox": [ + 91, + 70, + 518, + 286 + ], + "lines": [ + { + "bbox": [ + 91, + 70, + 518, + 286 + ], + "spans": [ + { + "bbox": [ + 91, + 70, + 518, + 286 + ], + "type": "table", + "html": "
SplitMethodIntentionEmotion
Macro F1Micro F1Samples F1Acc.F1 ScorePrecisionRecall
ValidationRandom15.1218.0317.694.911.854.822.25
MISA [8]23.8953.3253.6363.9318.6020.2217.55
MFSA [41]31.5553.8853.5264.1519.5620.0717.91
MulT [36]34.3752.4753.1861.2819.2920.8818.79
TCL(Map [51])34.2553.1453.7163.1519.1420.0118.79
MMIN [50]27.9148.8549.5059.3017.4520.6716.97
MPMM [16]32.8852.0753.0063.1517.6521.1017.58
Ours36.0254.0653.8864.8219.8921.3119.85
TestingRandom15.1817.9617.616.252.065.172.00
MISA [8]23.7452.2752.5463.2815.5523.9816.52
MFSA [41]32.4053.7553.0164.0517.6926.5817.91
MulT [36]34.1152.0552.7762.7220.3823.8319.58
TCL(Map [51])34.7351.6452.4163.3120.1425.2719.32
MMIN [50]27.1548.1449.0257.7115.4717.6015.17
MPMM [16]32.8152.1352.7462.8217.7825.2717.94
Ours35.2552.9053.3564.6720.5529.6419.95
", + "image_path": "4689a79a0876153d7c0cf8e61b6291040569d85c07c3fcc77ebc326b61c26e52.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 318, + 149, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 149, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 149, + 331 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 341, + 297, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 341, + 297, + 594 + ], + "spans": [ + { + "bbox": [ + 54, + 341, + 297, + 594 + ], + "type": "text", + "content": "Component Analysis. Our study delves into the advantages of combining the Modality Asynchronous Prompt (MAP) with the Bridging Emotion-Intention Former (BEIF), offering a detailed examination in Tab. 4. Our results reveal that while each component individually contributes to performance improvement—MAP elevating the Macro F1 score to " + }, + { + "bbox": [ + 54, + 341, + 297, + 594 + ], + "type": "inline_equation", + "content": "33.36\\%" + }, + { + "bbox": [ + 54, + 341, + 297, + 594 + ], + "type": "text", + "content": " and BEIF to " + }, + { + "bbox": [ + 54, + 341, + 297, + 594 + ], + "type": "inline_equation", + "content": "35.25\\%" + }, + { + "bbox": [ + 54, + 341, + 297, + 594 + ], + "type": "text", + "content": " for intention understanding—it's their combined application that significantly amplifies effectiveness, as evidenced by the compelling outcomes achieved (35.79% Macro F1, 53.92% Micro F1, and 54.88% Samples F1). The ablation study on MAP particularly highlights the effectiveness of our method, especially in addressing the challenge of missing modalities. The asynchronous prompting of missing information, crucially, does not compromise the feature space of the existing modalities. This approach, combined with our preference for element-wise addition over concatenation, ensures the dimensional consistency and integrity of the data representation remain intact. These findings indicate that our holistic approach is adept at offering a nuanced and thorough understanding of complex multimodal data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 318, + 555, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 555, + 559 + ], + "type": "text", + "content": "Emotion-Intention Correlation Mining. To validate the effectiveness of various strategies for mining intention and emotion correlations, we conduct ablation studies, as shown in Tab. 5. The experiments explore various methods, including a generic multi-task learning paradigm that directly combines emotion and intention losses. This paradigm, while straightforward, shows limited improvement (Macro F1: +1.21%) due to the absence of a specialized Emotion-Intention mechanism during training. It can even harm the ability to understand intention in certain scenarios, with decreases in Micro F1 (-1.06%) and Samples F1 (-0.30%) compared to the baseline. While the Complete Information method (Comp. Info.), which employs one aspect as a prior during both training and inference, exhibits notable gains (+9.06% Macro F1, +9.80% Micro F1), its practical application is limited by available prior knowledge in real-world scenarios. The limitations of these alternative approaches further emphasize the importance of BEAR, which achieves consistent improvements (+2.44% Macro F1) without requiring additional inference-time information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 568, + 556, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 556, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 556, + 594 + ], + "type": "text", + "content": "Feature Extractor Impact. Beyond model architecture, we investigate how different feature extractors influence in" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 107, + 608, + 502, + 698 + ], + "blocks": [ + { + "bbox": [ + 55, + 289, + 553, + 313 + ], + "lines": [ + { + "bbox": [ + 55, + 289, + 553, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 553, + 313 + ], + "type": "text", + "content": "Table 3. Performance comparison of multimodal fusion and missing modality methods on intention and emotion understanding tasks. Results show consistent improvements of our approach across all metrics on both validation and test sets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 107, + 608, + 502, + 698 + ], + "lines": [ + { + "bbox": [ + 107, + 608, + 502, + 698 + ], + "spans": [ + { + "bbox": [ + 107, + 608, + 502, + 698 + ], + "type": "table", + "html": "
ComponentIntentionEmotion
MAPEABEIFMacro F1Micro F1Samples F1Acc.F1 ScorePrecisionRecall
32.8152.1352.7462.8217.7825.2717.94
✓33.3652.7453.4362.7218.9528.4319.30
✓✓34.3752.4753.1861.2819.2925.8818.79
✓35.2552.9053.3564.6720.0229.1919.35
✓✓✓35.7953.9254.8864.5320.5529.6420.35
", + "image_path": "104f5c2140859823d175394e43f91c195552b321dea3e646b40e16e15b3b850c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 702, + 555, + 726 + ], + "lines": [ + { + "bbox": [ + 55, + 702, + 555, + 726 + ], + "spans": [ + { + "bbox": [ + 55, + 702, + 555, + 726 + ], + "type": "text", + "content": "Table 4. Ablation study showing the contribution of each component. Integration of both Modality Asynchronous Prompt (MAP) and Bridging Emo-Int Former (BEIF) yields optimal performance across all metrics. EA means element-wise addition." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24706" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 297, + 204 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 297, + 204 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 297, + 204 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 297, + 204 + ], + "type": "table", + "html": "
MethodMacro F1Intention
Micro F1Samples F1
Baseline32.8152.1352.74
w/ Multi-task34.0251.0752.44
(+1.21)(-1.06)(-0.30)
w/ EmoInt-Trans [35]31.5952.0651.32
(-1.22)(-0.06)(-1.42)
w/ BEIFormer35.2552.9053.35
(+2.44)(+0.77)(+0.61)
w/ Comp. Info.†41.8761.9362.09
(+9.06)(+9.80)(+8.54)
", + "image_path": "f21729a5b1e2bf20557f546e842d3eeeff6e1b3c4b50957bf0757b7e38ffab1e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 250, + 294, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 250, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 54, + 250, + 294, + 380 + ], + "type": "text", + "content": "tention and emotion understanding performance. We evaluate various state-of-the-art extractors for visual [20, 27], audio [7], and textual [49] modalities. As shown in Tab. 6, more advanced feature extractors consistently improve performance. CLIP-ViT-S [27] provides the most substantial gains for visual features (+1.81% Macro F1), while SemBERT [49] significantly enhances text representation (+1.42% Micro F1). Notably, while AST shows modest improvements in the metrics (+0.76% Macro F1), its impact is more significant than these numbers suggest, given that audio data comprises only a small portion of the test set." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 382, + 294, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 382, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 54, + 382, + 294, + 477 + ], + "type": "text", + "content": "Generalization Analysis on CMU-MOSEI. To evaluate the generalizability of our MAP approach, we conduct experiments on the CMU-MOSEI dataset across various baseline architectures, as detailed in Tab. 7. MAP demonstrates consistent improvements across all tested configurations. These consistent gains across different architectures and modality combinations validate MAP's robustness in handling missing modalities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 478, + 294, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 478, + 294, + 562 + ], + "spans": [ + { + "bbox": [ + 54, + 478, + 294, + 562 + ], + "type": "text", + "content": "Data Imbalance. The modality imbalance in MINE is intentional, reflecting the natural, in-the-wild modality distribution of social media content. To evaluate its impact, we conduct experiments with modality resampling. As shown in Tab. 8, resampling (the rare modalities by " + }, + { + "bbox": [ + 54, + 478, + 294, + 562 + ], + "type": "inline_equation", + "content": "2\\mathrm{x}" + }, + { + "bbox": [ + 54, + 478, + 294, + 562 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 478, + 294, + 562 + ], + "type": "inline_equation", + "content": "3\\mathrm{x}" + }, + { + "bbox": [ + 54, + 478, + 294, + 562 + ], + "type": "text", + "content": "), especially feature-level contribution resampling [40], can improve the understanding performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 574, + 208, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 208, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 208, + 586 + ], + "type": "text", + "content": "6. Conclusion and Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "content": "We introduce MINE, a novel dataset capturing the inherent relationship between intention and emotion in natural social media settings. Unlike existing datasets that treat these aspects in isolation or rely on controlled environments, MINE preserves the authentic challenges of real-world social media data—particularly uncertain modalities and diverse thematic contexts. By providing detailed annotations for both intention and emotion across multiple modalities, MINE establishes a comprehensive benchmark for understanding human social behavior in naturalistic settings." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 550, + 204 + ], + "blocks": [ + { + "bbox": [ + 55, + 209, + 296, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 296, + 243 + ], + "type": "text", + "content": "Table 5. Advanced relation mining strategies are important in enhancing emotion-intention analysis. †represents the utilization of supplementary information during the inference process." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 70, + 550, + 204 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 550, + 204 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 550, + 204 + ], + "type": "table", + "html": "
MethodMacro F1Intention
Micro F1Samples F1
Baseline32.8152.1352.74
w/ AST [7]33.5752.9553.34
(+0.76)(+0.82)(+0.60)
w/ CLIP-ViT-S [27]34.6253.7054.02
(+1.81)(+1.57)(+1.28)
w/ SwinT-S [20]33.1452.6752.98
(+0.33)(+0.54)(+0.24)
w/ SemBERT [49]33.8953.5554.02
(+1.08)(+1.42)(+1.28)
", + "image_path": "1a49365f489a3abba56f533a0e8f5e63bf2343d71d59eaf890be7cba47d0d12b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 239, + 550, + 376 + ], + "blocks": [ + { + "bbox": [ + 351, + 208, + 516, + 219 + ], + "lines": [ + { + "bbox": [ + 351, + 208, + 516, + 219 + ], + "spans": [ + { + "bbox": [ + 351, + 208, + 516, + 219 + ], + "type": "text", + "content": "Table 6. Impact of various feature extractors." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 239, + 550, + 376 + ], + "lines": [ + { + "bbox": [ + 317, + 239, + 550, + 376 + ], + "spans": [ + { + "bbox": [ + 317, + 239, + 550, + 376 + ], + "type": "table", + "html": "
ModelUtilized modality
T, VT, AT
MulT [36]37.1538.2841.91
w/ MAP38.93+1.7839.14+0.8643.35+1.44
MFN [45]58.4860.5459.74
w/ MAP61.33+2.8561.95+1.4163.72+3.98
BERT_MAG [28]62.8361.3962.83
w/ MAP64.51+1.6862.55+1.1664.42+1.59
LMF [19]42.3857.1555.94
w/ MAP44.25+1.8758.67+1.5257.12+1.18
", + "image_path": "e44363322b3adb1a24bb89bd16fae5b6755b1aaf8140f9f03c02888bd89da760.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 328, + 418, + 539, + 453 + ], + "blocks": [ + { + "bbox": [ + 313, + 380, + 553, + 403 + ], + "lines": [ + { + "bbox": [ + 313, + 380, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 553, + 403 + ], + "type": "text", + "content": "Table 7. Evaluation MAP across various architectures and modality combinations on CMU-MOSEI. T: Text, V: Visual, A: Audio." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 418, + 539, + 453 + ], + "lines": [ + { + "bbox": [ + 328, + 418, + 539, + 453 + ], + "spans": [ + { + "bbox": [ + 328, + 418, + 539, + 453 + ], + "type": "table", + "html": "
Methodsours2*rare3*rareWei et al [40]
Average51.4551.7951.0853.20
", + "image_path": "6c7a7042726c39e02b7316659c93b26b14882592a8b7246eaaa408a3ab58b4cf.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 456, + 553, + 468 + ], + "lines": [ + { + "bbox": [ + 314, + 456, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 314, + 456, + 553, + 468 + ], + "type": "text", + "content": "Table 8. Results of various methods for resisting modal imbalance." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 312, + 483, + 553, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 483, + 553, + 651 + ], + "spans": [ + { + "bbox": [ + 312, + 483, + 553, + 651 + ], + "type": "text", + "content": "To effectively process MINE's complex multimodal data, we propose BEAR (Bridging Emotion-Intention via Implicit Label Reasoning). It addresses two critical challenges: the intrinsic relationship between emotion and intention, and the uncertainty of modalities. The BEAR employs an alignment task with masked emotion and intention categories, enabling implicit mining of their relationships. Meanwhile, the Modality Asynchronous Prompt handles missing modalities through targeted prompting that preserves the integrity of available features. Experimental results demonstrate BEAR's effectiveness, outperforming existing methods across various metrics. In summary, MINE opens new avenues for understanding human social behavior through multimodal content." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 553, + 732 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 732 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 732 + ], + "type": "text", + "content": "Acknowledgements. This work is partially supported by the National Key Research and Development Program of China (2024YFC3308400), National Natural Science Foundation of China under Grant (62176188, 62361166629), Major Project of Science and Technology Innovation of Hubei Province (2024BCA003) and Key Project of the Health Commission of Hubei Province (WJ2023Z003)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24707" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 295, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 295, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 295, + 124 + ], + "type": "text", + "content": "[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. In NIPS, 2020. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 179 + ], + "type": "text", + "content": "[2] Carlos Busso, Murtaza Bulut, Chi-Chun Lee, Abe Kazemzadeh, Emily Mower, Samuel Kim, Jeannette N Chang, Sungbok Lee, and Shrikanth S Narayanan. Iemocap: Interactive emotional dyadic motion capture database. Language resources and evaluation, 2008. 1, 2, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 295, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 295, + 236 + ], + "type": "text", + "content": "[3] Haozhe Chi, Minghua Yang, Junhao Zhu, Guanhong Wang, and Gaoang Wang. Missing modality meets meta sampling (m3s): An efficient universal approach for multimodal sentiment analysis with missing modality. arXiv, 2022. 2, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 238, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 238, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 62, + 238, + 294, + 270 + ], + "type": "text", + "content": "[4] Dorottya Demszky, Dana Movshovitz-Attias, Jeongwoo Ko, Alan Cowen, Gaurav Nemade, and Sujith Ravi. Goemotions: A dataset of fine-grained emotions. arXiv, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 272, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 272, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 272, + 295, + 327 + ], + "type": "text", + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv, 2020. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 328, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 328, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 328, + 294, + 361 + ], + "type": "text", + "content": "[6] Yiyang Fang, Wenke Huang, Guancheng Wan, Kehua Su, and Mang Ye. Emoe: Modality-specific enhanced dynamic emotion experts. In CVPR, 2025. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 363, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 363, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 363, + 294, + 384 + ], + "type": "text", + "content": "[7] Yuan Gong, Yu-An Chung, and James Glass. Ast: Audio spectrogram transformer. arXiv, 2021. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 386, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 386, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 294, + 418 + ], + "type": "text", + "content": "[8] Devamanyu Hazarika, Roger Zimmermann, and Soujanya Poria. Misa: Modality-invariant and-specific representations for multimodal sentiment analysis. In ACM MM, 2020. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 420, + 294, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 294, + 462 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 294, + 462 + ], + "type": "text", + "content": "[9] Charles T Hemphill, John J Godfrey, and George R Doddington. The atis spoken language systems pilot corpus. In Speech and Natural Language: Proceedings of a Workshop, 1990. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 465, + 294, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 465, + 294, + 507 + ], + "spans": [ + { + "bbox": [ + 57, + 465, + 294, + 507 + ], + "type": "text", + "content": "[10] Ao Jia, Yu He, Yazhou Zhang, Sagar Uprety, Dawei Song, and Christina Lioma. Beyond emotion: A multi-modal dataset for human desire understanding. In NACCL, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 510, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 510, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 57, + 510, + 294, + 553 + ], + "type": "text", + "content": "[11] Menglin Jia, Zuxuan Wu, Austin Reiter, Claire Cardie, Serge Belongie, and Ser-Nam Lim. Intentonomy: a dataset and study towards human intent understanding. In CVPR, 2021. 2, 3, 4, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 555, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 555, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 555, + 294, + 599 + ], + "type": "text", + "content": "[12] Kawaljeet Kaur Kapoor, Kuttimani Tamilmani, Nripendra P Rana, Pushp Patil, Yogesh K Dwivedi, and Sridhar Nerur. Advances in social media research: Past, present and future. Information Systems Frontiers, 2018. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 601, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 294, + 634 + ], + "type": "text", + "content": "[13] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NACCL, 2019. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 635, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 294, + 678 + ], + "type": "text", + "content": "[14] Julia Kruk, Jonah Lubin, Karan Sikka, Xiao Lin, Dan Jurafsky, and Ajay Divakaran. Integrating text and image: Determining multimodal document intent in Instagram posts. In EMNLP-IJCNLP, 2019. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 680, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 294, + 714 + ], + "type": "text", + "content": "[15] Stefan Larson, Anish Mahendran, Joseph J Peper, Christopher Clarke, Andrew Lee, Parker Hill, Jonathan K Kummerfeld, Kevin Leach, Michael A Laurenzano, Lingjia Tang," + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "et al. An evaluation dataset for intent classification and out-of-scope prediction. In EMNLP-IJCNLP, 2019. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "text", + "content": "[16] Yi-Lun Lee, Yi-Hsuan Tsai, Wei-Chen Chiu, and Chen-Yu Lee. Multimodal prompting with missing modalities for visual recognition. In CVPR, 2023. 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 130, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 173 + ], + "type": "text", + "content": "[17] Rui Liu, Haolin Zuo, Zheng Lian, Xiaofen Xing, Bjorn W Schuller, and Haizhou Li. Emotion and intent joint understanding in multimodal conversation: A benchmarking dataset. arXiv, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 207 + ], + "type": "text", + "content": "[18] Xingkun Liu, Arash Eshghi, Pawel Swietojanski, and Verena Rieser. Benchmarking natural language understanding services for building conversational agents. arXiv, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "text", + "content": "[19] Zhun Liu, Ying Shen, Varun Bharadhwaj Lakshminarasimhan, Paul Pu Liang, Amir Zadeh, and LouisPhilippe Morency. Efficient low-rank multimodal fusion with modality-specific factors. arXiv, 2018. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "type": "text", + "content": "[20] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In CVPR, 2022. 4, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 287, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 342 + ], + "type": "text", + "content": "[21] Meng Luo, Hao Fei, Bobo Li, Shengqiong Wu, Qian Liu, Soujanya Poria, Erik Cambria, Mong-Li Lee, and Wynne Hsu. Panosent: A panoptic sextuple extraction benchmark for multimodal conversational aspect-based sentiment analysis. In ACM MM, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 343, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 376 + ], + "type": "text", + "content": "[22] Mengmeng Ma, Jian Ren, Long Zhao, Sergey Tulyakov, Cathy Wu, and Xi Peng. Smil: Multimodal learning with severely missing modality. In AAAI, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 377, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 409 + ], + "type": "text", + "content": "[23] Mengmeng Ma, Jian Ren, Long Zhao, Davide Testuggine, and Xi Peng. Are multimodal transformers robust to missing modality? In CVPR, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 443 + ], + "type": "text", + "content": "[24] Kuan-Chuan Peng, Amir Sadovnik, Andrew Gallagher, and Tsuhan Chen. Where do emotions come from? predicting the emotion stimuli map. In IEEE ICIP, 2016. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 445, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 499 + ], + "type": "text", + "content": "[25] Wei Peng, Yue Hu, Yuqiang Xie, Luxi Xing, and Yajing Sun. Cogintac: Modeling the relationships between intention, emotion and action in interactive process from cognitive perspective. In 2022 IEEE Congress on Evolutionary Computation, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "type": "text", + "content": "[26] Soujanya Poria, Devamanyu Hazarika, Navonil Majumder, Gautam Naik, Erik Cambria, and Rada Mihalcea. Meld: A multimodal multi-party dataset for emotion recognition in conversations. In ACL, 2019. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 545, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 599 + ], + "type": "text", + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "text", + "content": "[28] Wasifur Rahman, Md Kamrul Hasan, Sangwu Lee, AmirAli Bagher Zadeh, Chengfeng Mao, Louis-Philippe Morency, and Ehsan Hoque. Integrating multimodal information in large pretrained transformers. In ACL, 2020. 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "type": "text", + "content": "[29] Md Kaykobad Reza, Ashley Prater-Bennette, and M Salman Asif. Robust multimodal learning with missing modalities via parameter-efficient adaptation. IEEE TPAMI, 2024. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "text", + "content": "[30] Tulika Saha, Aditya Patra, Sriparna Saha, and Pushpak Bhattacharyya. Towards emotion-aided multi-modal dialogue act classification. In ACL, 2020. 1" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24708" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[31] Tulika Saha, Aditya Patra, Sriparna Saha, and Pushpak Bhattacharyya. Towards emotion-aided multi-modal dialogue act classification. In ACL, 2020. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "text", + "content": "[32] Tobias Schröder, Terrence C Stewart, and Paul Thagard. Intention, emotion, and action: A neural theory based on semantic pointers. Cognitive science, 2014. 1, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "text", + "content": "[33] QingHongYa Shi, Mang Ye, Ziyi Zhang, and Bo Du. Learnable hierarchical label embedding and grouping for visual intention understanding. IEEE Transactions on Affective Computing, 2023. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 239 + ], + "type": "text", + "content": "[34] Mohammad Faridul Haque Siddiqui, Parashar Dhakal, Xiaoli Yang, and Ahmad Y Javaid. A survey on databases for multimodal emotion recognition and an introduction to the viri (visible and infrared image) database. Multimodal Technologies and Interaction, 2022. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 295, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 295, + 285 + ], + "type": "text", + "content": "[35] Gopendra Vikram Singh, Mauajama Firdaus, Asif Ekbal, and Pushpak Bhattacharyya. Emoin-trans: A multimodal transformer for identifying emotions and intents in social conversations. IEEE/ACM TASLP, 2022. 1, 2, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "text", + "content": "[36] Yao-Hung Hubert Tsai, Shaojie Bai, Paul Pu Liang, J. Zico Kolter, Louis-Philippe Morency, and Ruslan Salakhutdinov. Multimodal transformer for unaligned multimodal language sequences. In ACL, 2019. 4, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 331, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 364 + ], + "type": "text", + "content": "[37] Gokhan Tur, Dilek Hakkani-Tur, and Larry Heck. What is left to be understood in atis? In IEEE Spoken Language Technology Workshop, 2010. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 398 + ], + "type": "text", + "content": "[38] Chai M Tyng, Hafeez U Amin, Mohammad NM Saad, and Aamir S Malik. The influences of emotion on learning and memory. Frontiers in psychology, 2017. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 399, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 431 + ], + "type": "text", + "content": "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, 2017. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 433, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 294, + 464 + ], + "type": "text", + "content": "[40] Yake Wei, Ruoxuan Feng, Zihe Wang, and Di Hu. Enhancing multimodal cooperation via sample-level modality valuation. In CVPR, 2024. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "text", + "content": "[41] Dingkang Yang, Haopeng Kuang, Shuai Huang, and Lihua Zhang. Learning modality-specific and-agnostic representations for asynchronous multimodal language sequences. In ACM MM, 2022. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 512, + 294, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 543 + ], + "type": "text", + "content": "[42] Qu Yang, Mang Ye, and Bo Du. Emollm: Multimodal emotional understanding meets large language models. arXiv, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "type": "text", + "content": "[43] Qu Yang, Mang Ye, and Dacheng Tao. Synergy of sight and semantics: visual intention understanding with clip. In ECCV, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 580, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 294, + 622 + ], + "type": "text", + "content": "[44] Amir Zadeh, Rowan Zellers, Eli Pincus, and Louis-Philippe Morency. Mosi: multimodal corpus of sentiment intensity and subjectivity analysis in online opinion videos. arXiv, 2016. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 624, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 666 + ], + "type": "text", + "content": "[45] Amir Zadeh, Paul Pu Liang, Navonil Mazumder, Soujanya Poria, Erik Cambria, and Louis-Philippe Morency. Memory fusion network for multi-view sequential learning. In AAAI, 2018. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[46] AmirAli Bagher Zadeh, Paul Pu Liang, Soujanya Poria, Erik Cambria, and Louis-Philippe Morency. Multimodal language analysis in the wild: Cmu-mosei dataset and interpretable dynamic fusion graph. In ACL, 2018. 1, 2, 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 285 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 115 + ], + "type": "text", + "content": "[47] Hanlei Zhang, Hua Xu, Xin Wang, Qianrui Zhou, Shaojie Zhao, and Jiayan Teng. Mintrec: A new dataset for multimodal intent recognition. In ACM MM, 2022. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "text", + "content": "[48] Hanlei Zhang, Xin Wang, Hua Xu, Qianrui Zhou, Kai Gao, Jianhua Su, Wenrui Li, Yanting Chen, et al. Mintrec2.0: A large-scale benchmark dataset for multimodal intent recognition and out-of-scope detection in conversations. In ICLR, 2024. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 174, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 206 + ], + "type": "text", + "content": "[49] Zhuosheng Zhang, Yuwei Wu, Hai Zhao, Zuchao Li, Shuai-liang Zhang, Xi Zhou, and Xiang Zhou. Semantics-aware bert for language understanding. In AAAI, 2020. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 208, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 239 + ], + "type": "text", + "content": "[50] Jinming Zhao, Ruichen Li, and Qin Jin. Missing modality imagination network for emotion recognition with uncertain missing modalities. In ACL, 2021. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "text", + "content": "[51] Qianrui Zhou, Hua Xu, Hao Li, Hanlei Zhang, Xiaohan Zhang, Yifan Wang, and Kai Gao. Token-level contrastive learning with modality-aware prompting for multimodal intent recognition. In AAAI, 2024. 7" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24709" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Uncertainty Meets Diversity_ A Comprehensive Active Learning Framework for Indoor 3D Object Detection/d2e21437-1e7e-4dc1-839a-f40a2bf9489e_content_list.json b/2025/Uncertainty Meets Diversity_ A Comprehensive Active Learning Framework for Indoor 3D Object Detection/d2e21437-1e7e-4dc1-839a-f40a2bf9489e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3f0a4ce3286b29ebd44b47098a99917a06641530 --- /dev/null +++ b/2025/Uncertainty Meets Diversity_ A Comprehensive Active Learning Framework for Indoor 3D Object Detection/d2e21437-1e7e-4dc1-839a-f40a2bf9489e_content_list.json @@ -0,0 +1,1419 @@ +[ + { + "type": "text", + "text": "Uncertainty Meets Diversity: A Comprehensive Active Learning Framework for Indoor 3D Object Detection", + "text_level": 1, + "bbox": [ + 96, + 128, + 903, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiangyi Wang, Na Zhao* \nSingapore University of Technology and Design (SUTD) \nwangjiangyi0519@gmail.com, na.zhao@sutd.edu.sg", + "bbox": [ + 272, + 203, + 725, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Active learning has emerged as a promising approach to reduce the substantial annotation burden in 3D object detection tasks, spurring several initiatives in outdoor environments. However, its application in indoor environments remains unexplored. Compared to outdoor 3D datasets, indoor datasets face significant challenges, including fewer training samples per class, a greater number of classes, more severe class imbalance, and more diverse scene types and intra-class variances. This paper presents the first study on active learning for indoor 3D object detection, where we propose a novel framework tailored for this task. Our method incorporates two key criteria - uncertainty and diversity - to actively select the most ambiguous and informative unlabeled samples for annotation. The uncertainty criterion accounts for both inaccurate detections and undetected objects, ensuring that the most ambiguous samples are prioritized. Meanwhile, the diversity criterion is formulated as a joint optimization problem that maximizes the diversity of both object class distributions and scene types, using a new Class-aware Adaptive Prototype (CAP) bank. The CAP bank dynamically allocates representative prototypes to each class, helping to capture varying intra-class diversity across different categories. We evaluate our method on SUN RGB-D and ScanNetV2, where it outperforms baselines by a significant margin, achieving over $85\\%$ of fully-supervised performance with just $10\\%$ of the annotation budget. Our code is available at: https://github.com/JoeWang-0519/CVPR25_UMD.", + "bbox": [ + 89, + 323, + 483, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 762, + 220, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D object detection, which aims to localize and identify objects within a 3D scene, is crucial for various applications, including robotics [10], autonomous driving [20, 58], and augmented/virtual reality [51]. While deep learning techniques have achieved significant success in 3D object detection [31, 41, 48], most existing approaches are data-", + "bbox": [ + 89, + 787, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/218ac544ffc061776612f813e9b811f331f308366de023c2783b1287a760b51b.jpg", + "image_caption": [ + "Figure 1. Challenges in active learning for indoor 3D object detection, including both uncertainty and diversity aspects. We show two indoor scenes on ScanNetV2, with red boxes for ground truths and cyan for predictions (10% data using CAGroup3D detector), displaying only 'chair' and 'door' categories for clarity. For uncertainty, the presence of undetected objects and inaccurate detections undermines the quality of uncertainty estimation. For diversity, high scene-type diversity and varying intra-class variances in indoor environments complicate diverse sample selection. Our work presents a hybrid approach to address these challenges." + ], + "image_footnote": [], + "bbox": [ + 522, + 291, + 897, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hungry [41], requiring large amounts of annotated data for training. However, annotating objects in a 3D scene is a labor-intensive process. For example, annotating a single scene in the SUN RGB-D dataset takes over 1.9 minutes per object, with an average of approximately 14 objects per scene [46]. To mitigate the high costs of 3D annotation, active learning (AL) [30, 34, 38] offers a promising solution. AL seeks to maximize the utility of limited labeling resources by actively selecting the most informative samples for human annotation. Through this process, AL can achieve performance comparable to fully-supervised methods, but with significantly fewer labeled samples.", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author: na_zhao@sutd.edu.sg", + "bbox": [ + 109, + 887, + 357, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20329", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although active learning has proven effective in image-based tasks such as image classification and object detection, using uncertainty [5, 16, 37] or/and diversity [34, 38, 60] as selection criteria, its application to 3D object detection remains largely underexplored. To date, only a few studies [12, 24, 25] have applied AL techniques to outdoor 3D object detection. For instance, CRB [25] uses label conciseness to capture uncertainty, along with feature representativeness and geometric balance to enhance diversity for sample selection. KECOR [24] favors samples whose latent features have the highest coding length, thereby ensuring diversity. While these methods demonstrate effectiveness in outdoor scenarios, they are suboptimal for indoor environments, as illustrated in Tab. 2, due to the unique challenges posed by 3D object detection in indoor settings.", + "bbox": [ + 89, + 90, + 480, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the one hand, a 3D indoor dataset [6] could be $6.2\\mathbf{x}$ smaller in scale but contain $3.3\\mathbf{x}$ more number of categories, with more severe class imbalance, compared to an outdoor dataset [11], as shown in [14]. This scarcity of training samples, coupled with pronounced class imbalance, impairs 3D detection accuracy in indoor scenarios, leading to more frequent occurrences of inaccurate detection and undetected objects, as illustrated in Fig. 1. In this figure, three detections are inaccurate, incorrectly localizing background objects, while three other objects remain undetected. These challenges underscore the need to account for both inaccurately detected proposals and undetected objects when selecting informative samples from unlabeled data.", + "bbox": [ + 89, + 320, + 482, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, unlike outdoor scenes, which primarily feature road environments with a limited set of object classes, indoor environments exhibit greater scene-type diversity and intra-class variances, caused by diverse layouts and variations in object shapes within each class. For example, as shown in Fig. 1, five 'chair' instances can form three clusters in the embedding space due to the variations in shape and occlusion. This phenomenon in indoor environments complicate the selection of diverse samples.", + "bbox": [ + 89, + 520, + 482, + 656 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we present the first solution to apply active learning in indoor 3D object detection, utilizing both uncertainty and diversity as selection criteria. For uncertainty, we design a two-pronged epistemic uncertainty estimation strategy to address inaccurate detections and undetected objects, which first assesses uncertainty separately for each case, then integrates them into a unified criterion. Specifically, we introduce a localization-aware uncertainty score that combines the IoU score – measured between the original detected object and its perturbed version – with the Shannon entropy. This helps mitigate the negative effects of invalid proposals, such as background or partial detections, when estimating the classification uncertainty of detected proposals. To quantify the epistemic uncertainty arising from undetectability, we develop a lightweight network that estimates the number of undetected objects. For di", + "bbox": [ + 89, + 659, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "versity, we frame the diverse sample selection problem as an optimization task that jointly maximizes scene-type diversity and intra-class variances, leveraging a novel Class-aware Adaptive Prototype (CAP) bank. This CAP bank is designed to capture varying intra-class variances by generating an adaptive number of representative prototypes per class. Observing that different scene types exhibit distinct histograms of object categories as shown in Fig. 1, each composed of various prototypes, we use the CAP bank to model both class-level and scene-level diversity.", + "bbox": [ + 511, + 90, + 903, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate the proposed method on SUN RGB-D [46] and ScanNet [6] datasets. Using CAGroup3D [48] as base detection model, our approach achieves $56.11\\%$ , $61.77\\%$ $\\mathrm{mAP}@0.25$ on SUN RGB-D and ScanNet, respectively, reaching $87.24\\%$ and $85.02\\%$ of the fully-supervised performance with only $10\\%$ of the annotation budget.", + "bbox": [ + 511, + 242, + 903, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 345, + 653, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generic Active Learning. The majority of existing works have explored active learning (AL) strategies on image classification task, which can be categorized into uncertainty-based [5, 9, 16, 37, 50] and diversity-based [1, 7, 34, 38, 54, 60] methods based on the selection criteria used. Uncertainty-based approaches select ambiguous samples for annotation by maximizing uncertainty scores, which can be computed in various ways, such as entropy [9, 17, 39], minimum posterior probability [50], or multi-class margin [5, 16, 37]. Shannon entropy [39] is a commonly used metric in classification that quantifies uncertainty in the probability distribution, while multi-class margin [37] measures the uncertainty through the difference between the top two predicted probabilities. MC Dropout [9] employs Monte Carlo Dropout to approximate the posterior predicted probability and use Shannon entropy to assess the underlying uncertainty. Diversity-based methods [1, 38, 60] select a compact subset of samples that best represent the global data distribution. For example, Core-set [38] formulates active learning as a core-set selection problem and solves it using the $k$ -centroid algorithm. Recent works [4, 22, 45] combine both uncertainty and diversity through hybrid approaches. A representative approach, BADGE [4], balances uncertainty and diversity by applying the $k$ -means++ [52] algorithm in the gradient space, which considers both magnitude and direction of gradient vectors.", + "bbox": [ + 511, + 369, + 906, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Object Detection. We focus on indoor 3D object detection [15, 63, 64, 66] and exclude outdoor methods [3, 42-44], which differ significantly. Prior indoor 3D object detection methods can be categorized into two lines in terms of underlying representation, i.e., voxel-based methods [13, 28, 40], and point-based methods [36, 49, 55, 56, 67]. Voxel-based methods convert point cloud to voxel grids and use 3D convolution networks to address the irregularity of point cloud. In contrast, point-based methods extract point", + "bbox": [ + 511, + 763, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20330", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/73f15c81cc263313548d559a4e6947e1c8a867cadcf8d24854e64352e73c1cac.jpg", + "image_caption": [ + "Figure 2. Overview of our proposed AL framework for indoor 3D object detection, exploiting both uncertainty and diversity. In the $r$ -th round, we first construct a candidate pool of size $\\delta \\cdot B_r$ using the Top- $K$ unified uncertainty score, which accounts for both inaccurate detections and undetections. Then, we jointly optimize the intra-class and scene-type diversity to select the $r$ -th selected dataset $\\mathcal{D}_r$ . Lastly, along with the previous labeled point clouds $\\mathcal{D}_L^{r-1}$ , we retrain the model until the total labeled dataset reaches the annotation budget $B$ ." + ], + "image_footnote": [], + "bbox": [ + 130, + 88, + 867, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "level features and use a voting mechanism [36, 56] to encourage points to vote towards their corresponding object centers, followed by grouping the vote points into clusters and then generating object proposals. In this work, we adopt the state-of-the-art method in indoor 3D object detection, CAGroup3D [48], which takes advantage of the two lines of work through a two-stage fully sparse convolutional 3D detection framework. It extracts voxel-wise features from raw point clouds and leverages 'voting and grouping' mechanism to generate object proposals.", + "bbox": [ + 88, + 395, + 483, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Active Learning for Object Detection. Despite extensive studies on AL in image classification, applying it to image-based object detection is challenging due to the task's complexity, which involves both localization and classification of multiple objects. Several works [1, 18, 30, 53, 59, 61] have investigated active object detection by adapting generic AL techniques to this task. For example, CDAL [1] designs contextual diversity to capture spatial and semantic context diversity within a dataset, while DivProto [53] and PPAL [59] incorporate both diversity and uncertainty. While these methods yield promising results in 2D object detection, they cannot be directly applied to 3D object detection due to the significant modality gap. Recently, driven by the need to reduce annotation burden in 3D object detection for autonomous driving, a few studies [8, 12, 21, 24, 25, 27, 29, 32] have explored active learning for 3D object detection in outdoor environments. Notably, CRB [25] selects active samples with concise labels, representative features, and balanced density progressively. KECOR [24] leverages kernel coding rate [26] to measure the richness of latent features. However, these outdoor methods are suboptimal for indoor environments, where 3D object detection faces unique challenges. To ad", + "bbox": [ + 91, + 553, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "dress these, this paper presents the first solution for applying active learning to indoor 3D object detection, overcoming these challenges by incorporating designs to estimate uncertainty in both inaccurate detections and undetections, while simultaneously optimizing diversity with respect to scene-type and intra-class variances.", + "bbox": [ + 511, + 395, + 906, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 506, + 648, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a point cloud $\\mathcal{P}_i = \\{(x,y,z)\\}$ , a 3D object detection model aims to predict the locations and class labels for a set of objects $\\mathfrak{B}_i = \\{(\\mathbf{b}_k,y_k)\\}$ , where $\\mathbf{b}_k\\in \\mathbb{R}^7$ represents the parameter of the $k$ -th bounding box, including the center position, box size and yaw angle, and $y_{k}$ indicates its predicted class label. In the active learning setting, our objective is to train the 3D detection model by constructing a small labeled dataset $\\mathcal{D}_L$ , which starts with a small portion of random samples $\\mathcal{D}_L^0$ and is progressively expanded over $R$ rounds from a large pool $\\mathcal{D}$ . In each active learning round, $r\\in \\{1,\\dots,R\\}$ , $B_{r}$ samples are selected from the unlabeled dataset $\\mathcal{D}_U = \\mathcal{D} / \\mathcal{D}_L^{r - 1}$ by an active learning policy, and labeled by an oracle, which constitute the $r$ -th selected dataset $\\mathcal{D}_r$ . Then, $\\mathcal{D}_L$ will be updated as $\\mathcal{D}_L^r\\gets \\mathcal{D}_L^{r - 1}\\cup \\mathcal{D}_r$ until the size of $\\mathcal{D}_L$ reaches the final budget $B$ , i.e., $|\\mathcal{D}_L^0 | + \\sum_{r = 1}^{R}B_r = B$ where $B\\ll |\\mathcal{D}|$ .", + "bbox": [ + 511, + 532, + 906, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To tackle active 3D object detection in indoor scenarios, we propose a novel framework, illustrated in Fig. 2, that leverages both uncertainty and diversity as selection criteria. Following previous work [21, 53, 59] that considers both uncertainty and diversity, we adopt a budget expansion ratio $\\delta \\in \\mathbb{Z}^{+}$ to first select a candidate pool $\\mathcal{D}_r^{\\mathrm{cand}}$ of size $K = \\delta \\cdot B_r$ based on the uncertainty criterion for $r$ -th active learning round. Then, from this pool, the $\\mathcal{D}_r$ of size", + "bbox": [ + 511, + 780, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20331", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/959d9e7a7148fab5539adbc4f1411f8e4bba96da6d6c854a854023714cc89ba1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetmAP@.25mAP@.50mAR@.25mAR@.50
SUN RGB-D [46]28.9112.4867.1226.09
ScanNetV2 [6]35.7520.4076.6542.48
", + "bbox": [ + 94, + 89, + 482, + 133 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Motivation for designing epistemic uncertainty estimation. We report the mean average precision (mAP) and mean average recall (mAR) at two IoU thresholds (0.25 and 0.50) for the initial detection model, which is trained using $2\\%$ labeled data.", + "bbox": [ + 89, + 140, + 482, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$B_{r}$ is chosen based on the diversity criterion. For the uncertainty criterion, we propose a two-pronged epistemic uncertainty estimation strategy that generates a unified score, accounting for uncertainty in both detected and undetected objects. This is achieved through a localization-aware uncertainty score and an estimate of the undetected object count (Sec. 3.1). For the diversity criterion, we frame the diverse sample selection as an optimization problem and propose an efficient solution that jointly optimizes intra-class variances and scene-type diversity (Sec. 3.2).", + "bbox": [ + 88, + 204, + 483, + 356 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Epistemic Uncertainty Estimation", + "text_level": 1, + "bbox": [ + 89, + 364, + 387, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Tab. 1, we evaluate the detection performance of a 3D object detection model trained on the initial random samples $D_{L}^{0}$ (2% labeled data). As shown, the model achieves 20.40% mAP@.50 and 42.48% mAR@.50 on ScanNet [6], indicating that approximately 80% of the predictions are inaccurate and 60% of ground-truth objects remain undetected. The results are even worse on the more challenging SUN RGB-D [46] dataset. These findings highlight the need to address the two types of epistemic uncertainty in 3D object detection: inaccurate detections and undetected objects. Given that the detection task involves both localization and classification, we estimate the epistemic uncertainty of inaccurate detections by considering both aspects. Specifically, we introduce a localization-aware uncertainty score that exploits an IoU score for localization reliability to combine with Shannon entropy. To handle undetections, we design a lightweight network to estimate the number of undetected objects, which serves as the uncertainty measure for undetections. These two uncertainty scores are then integrated into a unified uncertainty score, which is used for Top-K selection to generate the candidate pool $\\mathcal{D}_r^{\\mathrm{cand}}$ .", + "bbox": [ + 91, + 386, + 483, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Localization-aware Uncertainty Score. To estimate the uncertainty of detected objects from a point cloud $\\mathcal{P}_i$ , we first compute the classification uncertainty using the Shannon entropy [39] of the predicted class probabilities for all detected object proposals. For the $j$ -th proposal, with predicted class probabilities over $C$ classes denoted as $\\mathbf{p}_j = \\{p_{j,c}\\}_{c=1}^C \\in \\mathbb{R}^C$ , the Shannon entropy is calculated as:", + "bbox": [ + 89, + 704, + 482, + 811 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE (\\mathbf {p} _ {j}) = - \\sum_ {c = 1} ^ {C} p _ {j, c} \\log (p _ {j, c}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 820, + 482, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, not all detected proposals are valid, as some may correspond to background or partial objects. To mitigate", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the impact of these invalid proposals, we incorporate an IoU score that measures the similarity between the original detection and its perturbed version, obtained by applying weak augmentations (e.g., random rotation and scaling) to the input point cloud. The intuition is that if the model is confident about an object, small perturbations should not significantly affect its predicted position, while background or partial proposals will likely produce inconsistent predictions. Consequently, the localization-aware uncertainty score for detected objects in $\\mathcal{P}_i$ is defined as:", + "bbox": [ + 511, + 90, + 906, + 243 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {U} _ {\\det } \\left(\\mathcal {P} _ {i}\\right) := \\frac {\\sum_ {j = 1} ^ {\\left| \\mathfrak {B} _ {i} \\right|} \\operatorname {I o U} _ {j} \\cdot E \\left(\\mathbf {p} _ {j}\\right)}{\\sum_ {j = 1} ^ {\\left| \\mathfrak {B} _ {i} \\right|} \\operatorname {I o U} _ {j}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 250, + 906, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our IoU scores serve as soft weights, adaptively reducing the contribution of invalid objects in the epistemic uncertainty estimation.", + "bbox": [ + 511, + 297, + 906, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Undetected Number Prediction Network. When estimating uncertainty for active object detection, previous works [25, 53, 59] typically focus only on the uncertainty within the detected objects. However, the uncertainty of undetected objects is equally important, especially in the context of active learning for indoor 3D object detection, where model performance is often hindered by the scarcity of training samples and severe class imbalance. As shown by the mean average recall (mAR) scores in Tab. 1, more than half of the ground truth objects remain undetected in both 3D indoor datasets. This observation highlights the need to account for the uncertainty of undetected objects. To address this, we design a lightweight network to estimate the number of undetected objects.", + "bbox": [ + 511, + 344, + 906, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As depicted in Fig. 2, we first extract the feature map $\\mathbf{f}_i$ from the point cloud $\\mathcal{P}_i$ using the feature extractor. We then apply global average pooling (GAP) on $\\mathbf{f}_i$ to obtain the global feature $\\mathbf{g}_i$ . Simultaneously, we create a masked feature map by masking out the regions corresponding to the detected bounding boxes produced by RoI refinement. This masked feature map is passed through GAP to yield the undetected feature $\\mathbf{u}_i$ . We concatenate these two features, $\\mathbf{g}_i$ and $\\mathbf{u}_i$ , and use the concatenation as input to our undetected object number prediction network $\\Theta_{\\mathrm{undet}}$ , a lightweight network consisting of three-layer MLPs. The output of the network, i.e. $\\Theta_{\\mathrm{undet}}([\\mathbf{g}_i; \\mathbf{u}_i])$ represents the uncertainty score for undetections, denoted as $\\mathbf{U}_{\\mathrm{undet}}$ . This output is compared with the true number of undetected objects, computed based on the detected objects $\\mathfrak{B}_i$ and the ground-truth objects $\\mathfrak{B}_i^*$ , to compute the loss for training $\\Theta_{\\mathrm{undet}}$ :", + "bbox": [ + 511, + 556, + 908, + 797 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {u n d e t}} = \\sum_ {i = 1} ^ {| \\mathcal {D} _ {L} |} \\left(\\Theta_ {\\text {u n d e t}} \\left(\\left[ \\mathbf {g} _ {i}; \\mathbf {u} _ {i} \\right]\\right) - \\mathcal {N} \\left(\\mathfrak {B} _ {i}, \\mathfrak {B} _ {i} ^ {*}\\right)\\right) ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 805, + 906, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{N}(\\cdot, \\cdot)$ is a function that computes the number of undetected objects, with details provided in the supplementary material. Additionally, to enhance the generalizability", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20332", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of the network $\\Theta_{\\mathrm{undet}}$ , we employ random masking of the predicted boxes during the training phase, encouraging the network to better capture the undetectability in point clouds. Unified Uncertainty Score. After obtaining the two uncertainty scores, $\\mathbf{U}_{\\mathrm{det}}$ and $\\mathbf{U}_{\\mathrm{undet}}$ , we aim to combine them into a unified uncertainty score. However, these two scores are on different scales, making their direct product infeasible. To address this issue, we introduce a normalization step, inspired by [33], using a hyperparameter $k$ . Specifically, we normalize each score as follows:", + "bbox": [ + 89, + 90, + 480, + 243 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {U}} _ {\\star} (\\mathcal {P} _ {i}) = \\max (0, \\frac {\\mathbf {U} _ {\\star} (\\mathcal {P} _ {i}) - \\mu (\\mathbf {U} _ {\\star})}{k \\cdot \\sigma (\\mathbf {U} _ {\\star})} + 0. 5), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 251, + 482, + 284 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\star \\in \\{\\mathrm{det},\\mathrm{undet}\\}$ , and $\\mu (\\cdot)$ and $\\sigma (\\cdot)$ denote the mean and standard deviation, respectively. Once the scores are normalized, we compute their product to yield the unified uncertainty score.", + "bbox": [ + 89, + 295, + 483, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Prototype-based Diversity Optimization", + "text_level": 1, + "bbox": [ + 89, + 364, + 431, + 382 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Fig. 1, indoor environments exhibit significant intra-class variances and scene-type diversity, both of which motivate our approach to the diversity criterion. In 3D indoor scenes, the intra-class variances can vary greatly across object categories due to the differences in shape and occlusion. For example, the variance in 'chairs' is typically high due to the wide variety of shapes and designs, whereas 'doors' tend to have lower variance due to their more standardized geometry. This variability makes methods that model each class with a fixed number of prototypes [57, 65] insufficient to capture the intra-class distributions. To address this, we propose a novel Class-aware Adaptive Prototype (CAP) bank, which dynamically allocates a varying number of prototypes for each class, inferred from the predicted object proposals. Additionally, we observe that indoor scenes naturally form distinct clusters based on the distribution of object categories. For instance, bathrooms typically contain toilets and bathtubs, while bedrooms feature beds and dressers. Through this observation, we represent each scene as a histogram of the prototypes corresponding to the predicted object categories within the scene. As both intra-class variances and scene-type diversity can be modeled using the CAP bank, we frame diverse sample selection as an optimization problem that jointly maximizes both aspects, and propose an efficient solution.", + "bbox": [ + 91, + 387, + 483, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Class-aware Adaptive Prototype Bank. The design of our Class-aware Adaptive Prototype (CAP) bank is inspired by infinite mixture prototypes (IMP) [2], which was originally proposed for few-shot learning. IMP allows for inferring the number of clusters from the data, rather than relying on a pre-defined number of clusters. To build the CAP bank, we first iterate over the initial labeled set $D_{L}^{r - 1}$ to initialize one prototype per class. Specifically, we leverage the ground-truth annotations to extract features for all objects belong-", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ing to the $c$ -th class in $D_L^{r-1}$ , and compute their average as to form the initial prototype $\\mu_{1,c}$ for class $c$ . Subsequently, we iterate over the unlabeled dataset $\\mathcal{D}_U$ and update the prototypes for each class, resulting in our CAP bank, denoted as $\\mathbf{M} = \\{\\{\\pmb{\\mu}_{k,c}\\}_{k=1}^{M_c}\\}_{c=1}^C$ , where $\\pmb{\\mu}_{k,c}$ represents the $k$ -th prototype for class $c$ , $M_c$ is the number of prototypes for class $c$ , which may vary between classes, and $C$ is the total number of classes. In each iteration (one batch), we use the detection model from the previous round to generate object proposals and extract their features and predicted class labels. For each predicted object, we compute the cosine similarity between its feature embedding $\\mathbf{o}_j$ and each of the existing prototypes in its predicted class. This similarity score is formulated as:", + "bbox": [ + 511, + 90, + 906, + 303 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S i m} \\left(\\mathbf {o} _ {j}, \\boldsymbol {\\mu} _ {k, c}\\right) = \\frac {\\mathbf {o} _ {j} ^ {T} \\cdot \\boldsymbol {\\mu} _ {k , c}}{\\left| \\left| \\mathbf {o} _ {j} \\right| \\right| _ {2} \\left| \\left| \\boldsymbol {\\mu} _ {k , c} \\right| \\right| _ {2}} \\in [ - 1, 1 ]. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 309, + 906, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After calculating the similarity scores, each object is either assigned to an existing prototype or used to form a new prototype, according to the following assignment rule:", + "bbox": [ + 511, + 356, + 906, + 402 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} \\left(\\mathbf {o} _ {j}\\right) = \\left\\{ \\begin{array}{l} M _ {c} + 1, \\text {i f} \\max _ {k} \\operatorname {S i m} \\left(\\mathbf {o} _ {j}, \\boldsymbol {\\mu} _ {k, c}\\right) < \\tau_ {\\text {s i m}}, \\\\ \\arg \\max _ {k} \\operatorname {S i m} \\left(\\mathbf {o} _ {j}, \\boldsymbol {\\mu} _ {k, c}\\right), \\quad \\text {o t h e r w i s e}. \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 412, + 905, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\tau_{\\mathrm{sim}}$ is similarity threshold, which is set to 0.3 by default. Finally, the prototypes - whether assigned or newly formed - are updated using the new object proposals in each batch. However, since many predicted bounding boxes might be invalid, especially in the early stages of training (e.g., as shown in Tab. 1), we seek to mitigate the influence of these invalid proposals when updating the prototypes. To this end, we incorporate the IoU score from Eq. 2 into the prototype update process, effectively weighing each object's contribution to the prototype updates as follows:", + "bbox": [ + 511, + 474, + 908, + 626 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} _ {k, c} = \\frac {\\sum_ {j = 1} ^ {| \\mathfrak {B} _ {i} |} \\mathbb {1} \\left\\{\\mathcal {A} \\left(\\mathbf {o} _ {j}\\right) , k \\right\\} \\cdot \\operatorname {I o U} _ {j} \\cdot \\mathbf {o} _ {j}}{\\sum_ {j = 1} ^ {| \\mathfrak {B} _ {i} |} \\mathbb {1} \\left\\{\\mathcal {A} \\left(\\mathbf {o} _ {j}\\right) , k \\right\\} \\cdot \\operatorname {I o U} _ {j}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 635, + 906, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the indicator function $\\mathbb{1}(m,n)$ equals to 1 if $m = n$ , and 0 otherwise. The complete CAP bank creation procedure is outlined in Algo. 1 of the supplementary material.", + "bbox": [ + 511, + 689, + 905, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prototype-based Diversity Optimization. With the generated CAP bank, we can characterize each class through a set of prototypes, and further represent objects as histograms of these prototypes. These histograms allow us to capture the intra-class variance for each class, as objects within the same class are reflected in the variation of their corresponding histograms. Furthermore, based on the observation that each scene can be characterized by the distribution of object categories, which distinguishes one scene type from another, we also model scenes as histograms of prototypes. Therefore, this histogram-based representation", + "bbox": [ + 511, + 734, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20333", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "implicitly captures the scene type. Since both intra-class variance and scene-type diversity are encoded in the prototype histograms, we design a novel and efficient optimization method to quantify the diversity criterion by considering both these two aspects. Specifically, we formulate the optimization for diverse sample selection as follows:", + "bbox": [ + 89, + 90, + 483, + 183 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\max _ {\\mathcal {D} _ {r} \\subset \\mathcal {D} _ {r} ^ {\\text {c a n d}}} E (\\mathbf {h} (\\mathcal {D} _ {r})) + \\beta | | \\mathbf {h} (\\mathcal {D} _ {r}) | | _ {1}, \\\\ \\begin{array}{l} \\text {s . t .} \\quad \\mathbf {h} (\\mathcal {D} _ {r}) = \\sum_ {\\mathcal {P} _ {i} \\in \\mathcal {D} _ {r}} \\mathbf {h} (\\mathcal {P} _ {i}) = \\sum_ {\\mathcal {P} _ {i} \\in \\mathcal {D} _ {r}} \\sum_ {\\mathbf {o} _ {j} \\in \\mathcal {P} _ {i}} \\mathbf {h} (\\mathbf {o} _ {j}), \\end{array} \\\\ \\mathbf {h} \\left(\\mathbf {o} _ {j}\\right) = \\sum_ {c = 1} ^ {C} \\sum_ {k = 1} ^ {M _ {c}} [ \\text {o n e} \\_ \\text {h o t} (\\mathcal {A} \\left(\\mathbf {o} _ {j}\\right)) ] ^ {c, k}, \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 191, + 480, + 310 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{h}(\\mathbf{o}_j)$ represents the assignment histogram of $j$ -th object w.r.t. the CAP bank $\\mathbf{M}$ , while $\\mathbf{h}(\\mathcal{P}_i)$ and $\\mathbf{h}(\\mathcal{D}_r)$ represent the histograms of all objects in the individual point cloud $\\mathcal{P}_i$ and the selected labeled samples $\\mathcal{D}_r$ at the $r$ -th round, respectively. The two terms in this objective function are designed to select samples that are both diverse and holistic across all prototypes in the CAP bank. The first term captures the intra-class variance for all classes, ensuring that the selected samples represent the diversity within each class as well as across different classes. The second term promotes scene-level diversity by encouraging coverage across different object categories, which implicitly increases the diversity in scene types. Given that the first term (entropy) and the second term ( $l_1$ -norm) have different magnitudes, we introduce a hyperparameter $\\beta$ to rescale the second term. In our experiments, we set $\\beta = 0.01$ .", + "bbox": [ + 89, + 311, + 483, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, directly solving the optimization in Eq. 8 is computationally expensive due to the combinatorial search over $|\\mathcal{D}_r^{\\mathrm{cand}}|$ samples. To efficiently narrow the search space while maintaining solution quality, we leverage one key property of the objective function over $S$ disjoint partitions $\\{\\mathcal{D}_{r,s}\\}_{s = 1}^{S}$ of $\\mathcal{D}_r$ , i.e., $\\cup_{s}\\mathcal{D}_{r,s} = \\mathcal{D}_r$ :", + "bbox": [ + 89, + 553, + 483, + 643 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} E \\left(\\mathbf {h} \\left(\\mathcal {D} _ {r}\\right)\\right) + \\beta | | \\mathbf {h} \\left(\\mathcal {D} _ {r}\\right) | | _ {1} \\\\ \\stackrel {c} {=} \\sum_ {s = 1} ^ {S} \\left\\{\\frac {N _ {s}}{N _ {\\text {t o t a l}}} E \\left(\\mathbf {h} \\left(\\mathcal {D} _ {r, s}\\right)\\right) + \\beta \\left\\| \\mathbf {h} \\left(\\mathcal {D} _ {r, s}\\right) \\right\\| _ {1} \\right\\} + \\mathcal {O} \\left(\\frac {N _ {\\text {i n t e r}}}{N _ {\\text {t o t a l}}}\\right), \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 646, + 480, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{h}(\\mathcal{D}_{r,s})$ represents the histograms of all objects in $s$ -th partition $\\mathcal{D}_{r,s}$ , $N_{s}$ and $N_{\\mathrm{total}}$ are the number of objects in the $s$ -th partition $\\mathcal{D}_{r,s}$ and selected labeled samples $\\mathcal{D}_r$ , $N_{\\mathrm{inter}}$ is the number of objects that share the same prototypes but belong to different partitions, and $\\mathcal{O}(\\cdot)$ is the Big O notation from mathematics. To align with the $S$ disjoint partitions of $\\mathcal{D}_r$ , we use k-means++ [52] to group the candidate pool $\\mathcal{D}_r^{\\mathrm{cand}}$ into $S$ clusters, denoted as $\\{\\mathcal{D}_{r,s}^{\\mathrm{cand}}\\}_{s = 1}^{S}$ , each of which is approximately disjoint with each other w.r.t. prototype assignments. This guarantees that $\\frac{N_{\\mathrm{inter}}}{N_{\\mathrm{total}}} \\to 0$ , and then allows us to decompose the original optimization problem into $S$ sub-problems within each cluster $\\mathcal{D}_{r,s}^{\\mathrm{cand}}$ :", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\max _ {\\mathcal {D} _ {r, s} \\subset \\mathcal {D} _ {r, s} ^ {\\text {c a n d}}} E (\\mathbf {h} (\\mathcal {D} _ {r, s})) + \\beta \\frac {N _ {\\text {t o t a l}}}{N _ {s}} | | \\mathbf {h} (\\mathcal {D} _ {r, s}) | | _ {1}, \\\\ \\begin{array}{l} \\text {s . t .} \\quad \\mathbf {h} (\\mathcal {D} _ {r, s}) = \\sum_ {\\mathcal {P} _ {i} \\in \\mathcal {D} _ {r, s}} \\mathbf {h} (\\mathcal {P} _ {i}) = \\sum_ {\\mathcal {P} _ {i} \\in \\mathcal {D} _ {r, s}} \\sum_ {\\mathbf {o} _ {j} \\in \\mathcal {P} _ {i}} \\mathbf {h} (\\mathbf {o} _ {j}), \\end{array} \\\\ \\mathbf {h} \\left(\\mathbf {o} _ {j}\\right) = \\sum_ {c = 1} ^ {C} \\sum_ {k = 1} ^ {M _ {c}} [ \\text {o n e} \\text {h o t} (\\mathcal {A} \\left(\\mathbf {o} _ {j}\\right)) ] ^ {c, k}. \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 87, + 908, + 214 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use greedy search to solve the $S$ sub-problems to obtain the diverse samples $\\mathcal{D}_{r,s}$ for each cluster, and combine them together as the selected dataset $\\mathcal{D}_r$ in the $r$ -th active learning round. We provide the detailed derivation of Eq. 9 in the supplementary material.", + "bbox": [ + 511, + 215, + 906, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 306, + 645, + 324 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setups", + "text_level": 1, + "bbox": [ + 511, + 333, + 710, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D Detection Datasets. We evaluate our proposed method on two 3D indoor object detection benchmarks: SUN RGB-D [46] and ScanNet [6]. SUN RGB-D is an indoor dataset for 3D object detection, which contains 10,335 RGB-D images with 3D room layouts from four different sensors. Officially, they are split into 5,285 training samples and 5,050 testing samples. Following the standard evaluation protocol [19, 35, 36], we evaluate the performance on 10 most common categories. ScanNetV2 is a large 3D dataset with 1,513 RGB-D scans of 707 indoor scenes. These scans are split into 1,201 samples for training and 312 samples for testing, each of which is annotated with dense semantic segmentation masks. We follow [36] to derive the axis-aligned bounding boxes without considering the orientation, and adopt 18 object categories for evaluation.", + "bbox": [ + 511, + 354, + 906, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We use the state-of-the-art CA-Group3D [48] as the detection model, and implement our code base using OpenPCDet [47] toolkit. Following the training strategies in [48], we set the voxel size as $0.02\\mathrm{m}$ . For both datasets, the network is trained by AdamW optimizer [23] with an initial learning rate and weight decay rate of 0.001 and 0.0001 respectively. The gradnorm clip [62] is applied to stabilize the training process. For each active learning round, SUN RGB-D takes 48 epochs to train on the updated labeled dataset, and the learning rate decays on [32, 44] epochs with decay rate of 0.1. ScanNet requires 120 epochs for training and decays the learning rate on [80, 110] epochs by the same decay rate. All experiments are conducted using one 3090Ti NVIDIA GPU. We set the similarity threshold $\\tau_{\\mathrm{sim}}$ to 0.3, balance ratio $\\beta$ to 0.01, and budget expanding ratio $\\delta$ to 6 empirically.", + "bbox": [ + 511, + 582, + 905, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation. Following the evaluation protocol in [31, 36, 48], we report mean average precision (mAP) with different IoU thresholds, i.e., 0.25 and 0.50. Additionally, to investigate the effectiveness of different active learning strategies, we record the mAP with $2\\%$ , $4\\%$ , $6\\%$ , $8\\%$ , $10\\%$ of the train", + "bbox": [ + 511, + 825, + 905, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20334", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e0077951725b1c4d58b78bb74c2ef871237634e438de2ca776082a3d9e9d46db.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsPresened atSUN RGB-DScanNetV2
mAP@0.25 (%)mAP@0.50 (%)mAP@0.25 (%)mAP@0.50 (%)
2D-GenerieRAND-42.7125.1455.0939.74
MC-DROPOUT [9]ICML'1747.4828.3645.7128.72
CORESET [38]ICLR'1846.9928.1255.3440.08
BADGE [4]ICLR'2049.1230.5457.1041.97
2D-ODCDAL [1]ECCV'2050.8630.7557.9942.33
DIVPROTO [53]CVPR'2252.3132.5459.0143.10
PPAL [59]CVPR'2452.5533.5159.4443.15
3D-ODCRB [25]ICLR'2350.8931.3058.7342.71
KECOR [24]ICCV'2351.0731.4158.5442.56
OURS-56.1135.4261.7745.93
ORACLE†-64.3147.4372.6558.81
", + "bbox": [ + 130, + 82, + 859, + 280 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. mAP@0.25 (%) and mAP@0.50 (%) scores of the proposed method and AL baselines on the SUN RGB-D and ScanNetV2 benchmarks with $10\\%$ queried point clouds. CAGroup3D is used as the detection model for all approaches. $\\dagger$ : indicates the detection results using fully-annotated data, and serves as the upper bound for all active learning strategies.", + "bbox": [ + 89, + 281, + 906, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4b3db0e785a91ae609319f4207bc12e3e9a37f45af9668c402f6502417cddcbf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 337, + 480, + 441 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/893b7c2e70ae723b239d06c00a749ac8fd40db81019eeefdaa34a6be217db612.jpg", + "image_caption": [ + "Figure 3. mAP@0.25 (\\%) score of the proposed method and AL baselines on SUN RGB-D and ScanNetV2 benchmarks." + ], + "image_footnote": [], + "bbox": [ + 93, + 441, + 480, + 545 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing set for both datasets, where the first $2\\%$ samples are selected randomly. In each active learning round, $2\\%$ of the training point clouds are acquired from the rest unlabeled dataset via active learning strategies for annotation.", + "bbox": [ + 89, + 606, + 482, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We evaluate nine AL baselines: random sampling (RAND), three 2D generic AL strategies (McDROPOUT, CORESET, BADGE) representing uncertainty-based, diversity-based, and hybrid methods, three 2D AL detection strategies (CDAL, DIVPROTO, PPAL) adapted to 3D detection tasks by us, and two state-of-the-art 3D outdoor AL detection strategies (CRB and KECOR).", + "bbox": [ + 89, + 667, + 482, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Experimental Results", + "text_level": 1, + "bbox": [ + 89, + 786, + 292, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SUN RGB-D. In Fig. 3, we compare our proposed method with different baselines through an increasing percentage of selected point clouds [25, 59]. As demonstrated by the results, our proposed strategy consistently outperforms all competing baselines by significant margins for different percentages of selected point clouds. Notably, with only", + "bbox": [ + 89, + 809, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$10\\%$ annotated data, our approach achieves over $85\\%$ of the fully-annotated performance (red dashed line in Fig. 3). Specifically, as detailed in Tab. 2, it surpasses the state-of-the-art 2D active learning strategy, PPAI, by $3.56\\%$ in $\\mathrm{mAP}@0.25$ and $1.91\\%$ in $\\mathrm{mAP}@0.50$ , with a $10\\%$ annotation budget. To elaborate further, it is observed that AL methods for 2D object detection generally outperform those for 3D outdoor counterpart. We hypothesize that current 3D outdoor approaches overlook the intra-class variances across different categories, leading to redundancy in selecting active samples. This redundancy may hinder the capability of the learned detection model.", + "bbox": [ + 511, + 339, + 906, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ScanNetV2. To study the effectiveness of our proposed approach on point clouds with complicated room layouts, we conduct experiments using the ScanNetV2 dataset, a more comprehensive benchmark for indoor environments. The results, illustrated in Fig. 3 and Tab. 2, demonstrate the performance comparison across varying annotation budgets. Our method surpasses all existing AL approaches by a huge margin, achieving approximately $85\\%$ of the fully-annotated performance with $10\\%$ annotation budget. In Fig. 4, we present a qualitative comparison with random sampling. Random sampling captures only partial object characteristics (e.g., 'bookshelf') due to limited training data. In contrast, our approach successfully detects most of the target objects in challenging scenarios by effectively selecting informative unlabeled samples based on uncertainty and diversity criteria. Additional qualitative results can be found in the supplementary materials.", + "bbox": [ + 511, + 521, + 908, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 787, + 666, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We ablate some critical designs for an in-depth exploration. Design Choices. We start by exploring the effectiveness of different designs in our proposed method, as shown in Tab. 3. When equipped with the IoU score in Eq. 2 and the uncertainty score for undetections $\\mathbf{U}_{\\mathrm{undet}}$ , the mAP@0.25 increases by $1.1\\%$ and $0.7\\%$ on SUN RGB-D respectively,", + "bbox": [ + 511, + 809, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20335", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/51eaf9d3af05c0cd3219f4eeeca08f10a5ff3a2503f573af0c3e0ca42712e519.jpg", + "image_caption": [ + "(a) Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 101, + 108, + 359, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1d46073ab58a5c741e33fc66e70304697c46b4beed11564d1f6bf39200fca216.jpg", + "image_caption": [ + "(b) Random (10%)", + "Figure 4. Qualitative comparison between random sampling and our proposed active learning method on ScanNetV2 val set." + ], + "image_footnote": [], + "bbox": [ + 370, + 108, + 625, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/65b23165b211db0a2779d1c4e0c12e1d0ac6ad785f86eebff6930c77efe9d9d8.jpg", + "image_caption": [ + "(c) Ours (10%)" + ], + "image_footnote": [], + "bbox": [ + 638, + 108, + 893, + 200 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6fac66c582b7b7fc4c7aa165570a4b93e4b74b9de1c47c524d556ae2c180424a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
UncertaintyDiversitySUN RGB-DScanNetV2
mAP@.25mAP@.50mAP@.25mAP@.50
XX42.725.155.139.7
X✓54.533.958.742.9
w/ IoU score✓55.634.760.843.8
w/ Uundet✓55.234.360.143.3
✓X54.333.558.442.8
✓w/ K-means55.334.160.443.5
✓w/ Greedy-O54.533.358.743.0
✓✓56.135.461.845.9
", + "bbox": [ + 94, + 232, + 482, + 348 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ad46ab98f854f555d1ea5412ba19e53d8a1cbfcf337b53e7a7baae95227a814d.jpg", + "image_caption": [ + "Figure 5. Effects of hyper-parameters in our proposed method on SUN RGB-D dataset." + ], + "image_footnote": [], + "bbox": [ + 93, + 380, + 285, + 500 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6f98d1a12c2a1327561d2808e3b1504c7d2c3cf47c160036b20fa5d1fe186b8b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 380, + 478, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "validating its effectiveness as an uncertainty-based strategy. Additionally, we ablate the diversity branch through two alternatives: (1) Use K-means [52] to select a fixed number of prototypes to replace CAP Bank, denoted as 'K-means'. (2) Apply greedy search to the original optimization problem as defined in Eq. 8, denoted as 'Greedy-O'. It demonstrates that the mAP@0.50 of 'K-means' decreases by $2.4\\%$ for ScanNetV2 dataset, which suggests that a fixed prototype number leads to a sub-optimal outcome. Furthermore, 'Greedy-O' solution only achieves minor performance improvements. A possible reason is that we restrict the searching space of the original optimization problem (Eq. 8) to several regularized sub-spaces (Eq. 10), eliminating a vast number of sub-optimal solutions.", + "bbox": [ + 88, + 537, + 482, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hyper-parameters. In Fig. 5, we report the ablation study of two hyper-parameters in our proposed method, which are the scale $k$ in Eq. 4 and the budget expanding ratio $\\delta$ for determining the candidate pool [59]. Overall, the proposed method is relatively insensitive to the choices of hyperparameters. As depicted in Fig. 5 (a), the best performance is reached with a moderate scale value of $k = 3$ , which effectively suppresses extreme values while preserving essential information from different sources of uncertainty. Additionally, we observe that excessively large values of $\\delta$ will", + "bbox": [ + 88, + 750, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/bb1eb788b13952bbacdf23678562238dff1479ba646220402bf1a6ccb3b5aca8.jpg", + "table_caption": [ + "Table 3. Ablation studies of different designs in our proposed method on SUN RGB-D and ScanNetV2 datasets." + ], + "table_footnote": [], + "table_body": "
MeasuresSUN RGB-DScanNetV2
mAP@.25mAP@.50mAP@.25mAP@.50
Prob. Margin [37]55.835.061.545.5
MC-Dropout [9]56.035.261.645.6
Entropy [39]56.135.461.845.9
", + "bbox": [ + 514, + 232, + 905, + 296 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation studies of different uncertainty measures in our proposed method on SUN RGB-D and ScanNetV2 datasets.", + "bbox": [ + 511, + 297, + 905, + 325 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "significantly degrade model performance in the early stages, which is shown in Fig. 5 (b). We hypothesize that for big $\\delta$ , the candidate pool contains many samples the model is already confident about, thereby harming the quality of the final selected samples. Conversely, when $\\delta$ is very small, the selection process is dominated by uncertainty, which restricts the impact of diversity branch. We find that setting $\\delta = 6$ achieves a better tradeoff between uncertainty- and diversity-based selection, optimizing overall performance.", + "bbox": [ + 511, + 335, + 906, + 472 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Uncertainty Measures. We replace the Shannon entropy in Eq. 2 with two alternative uncertainty measures: MC-Dropout [9] and probability margin [37], and compare their performance in Tab. 4. The results show that all three measures yield similar performance, with entropy performing slightly better. Therefore, we select entropy as our uncertainty measure, as it is both more efficient and effective.", + "bbox": [ + 511, + 472, + 905, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 592, + 633, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents a novel active learning approach for 3D indoor object detection, improving annotation efficiency by combining uncertainty estimation with diversity-driven sampling. Specifically, our approach estimates epistemic uncertainty, accounting for both inaccurate detections and undetected objects, and selects diverse samples through an optimization method based on a Class-aware Adaptive Prototype bank. Our method significantly outperforms state-of-the-art baselines, achieving over $85\\%$ of fully-supervised performance with just $10\\%$ of the annotation budget.", + "bbox": [ + 511, + 618, + 906, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgment", + "text_level": 1, + "bbox": [ + 511, + 784, + 684, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research is supported by the Ministry of Education, Singapore, under its MOE Academic Research Fund Tier 2 (MOE-T2EP20124-0013), it is also supported by the Agency for Science, Technology and Research (A*STAR) under its MTC Programmatic Funds (Grant No. M23L7b0021).", + "bbox": [ + 511, + 809, + 905, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20336", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sharat Agarwal, Himanshu Arora, Saket Anand, and Chetan Arora. Contextual diversity for active learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16, pages 137-153. Springer, 2020. 2, 3, 7", + "[2] Kelsey Allen, Evan Shelhamer, Hanul Shin, and Joshua Tenenbaum. Infinite mixture prototypes for few-shot learning. In International conference on machine learning, pages 232-241. PMLR, 2019. 5", + "[3] Eduardo Arnold, Omar Y Al-Jarrah, Mehrdad Dianati, Saber Fallah, David Oxtoby, and Alex Mouzakitis. A survey on 3d object detection methods for autonomous driving applications. IEEE Transactions on Intelligent Transportation Systems, 20(10):3782-3795, 2019. 2", + "[4] Jordan T Ash, Chicheng Zhang, Akshay Krishnamurthy, John Langford, and Alekh Agarwal. Deep batch active learning by diverse, uncertain gradient lower bounds. arXiv preprint arXiv:1906.03671, 2019. 2, 7", + "[5] Gui Citovsky, Giulia DeSalvo, Claudio Gentile, Lazaros Karydas, Anand Rajagopalan, Afshin Rostamizadeh, and Sanjiv Kumar. Batch active learning at scale. Advances in Neural Information Processing Systems, 34:11933-11944, 2021. 2", + "[6] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 2, 4, 6", + "[7] Ehsan Elhamifar, Guillermo Sapiro, Allen Yang, and S Shankar Sasrty. A convex optimization framework for active learning. In Proceedings of the IEEE International Conference on Computer Vision, pages 209-216, 2013. 2", + "[8] Di Feng, Xiao Wei, Lars Rosenbaum, Atsuto Maki, and Klaus Dietmayer. Deep active learning for efficient training of a lidar 3d object detector. In 2019 IEEE Intelligent Vehicles Symposium (IV), pages 667-674. IEEE, 2019. 3", + "[9] Yarin Gal, Riashat Islam, and Zoubin Ghahramani. Deep Bayesian active learning with image data. In Proceedings of the 34th International Conference on Machine Learning, pages 1183-1192. PMLR, 2017. 2, 7, 8", + "[10] Mike Gashler and Tony Martinez. Temporal nonlinear dimensionality reduction. In The 2011 International Joint Conference on Neural Networks, pages 1959-1966. IEEE, 2011. 1", + "[11] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 2", + "[12] Ahmed Ghita, Björk Antoniussen, Walter Zimmer, Ross Greer, Christian Creß, Andreas Møgelmose, Mohan M Trivedi, and Alois C Knoll. Activeanno3d—an active learning framework for multi-modal 3d object detection. arXiv preprint arXiv:2402.03235, 2024. 2, 3", + "[13] JunYoung Gwak, Christopher Choy, and Silvio Savarese. Generative sparse detection networks for 3d single-shot object detection. In Computer Vision-ECCV 2020: 16th Euro-" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "pean Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 297-313. Springer, 2020. 2", + "[14] Yucheng Han, Na Zhao, Weiling Chen, Keng Teck Ma, and Hanwang Zhang. Dual-perspective knowledge enrichment for semi-supervised 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2049-2057, 2024. 2", + "[15] Pengkun Jiao, Na Zhao, Jingjing Chen, and Yu-Gang Jiang. Unlocking textual and visual wisdom: Open-vocabulary 3d object detection enhanced by comprehensive guidance from text and image. In European Conference on Computer Vision, pages 376-392. Springer, 2024. 2", + "[16] Ajay J Joshi, Fatih Porikli, and Nikolaos Papanikolopoulos. Multi-class active learning for image classification. In 2009 IEEE conference on computer vision and pattern recognition, pages 2372-2379. IEEE, 2009. 2", + "[17] Yoon-Yeong Kim, Kyungwoo Song, JoonHo Jang, and Il-Chul Moon. Lada: Look-ahead data acquisition via augmentation for deep active learning. Advances in Neural Information Processing Systems, 34:22919-22930, 2021. 2", + "[18] Seong Min Kye, Kwanghee Choi, Hyeongmin Byun, and Buru Chang. Tidal: Learning training dynamics for active learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22335-22345, 2023. 3", + "[19] Jean Lahoud and Bernard Ghanem. 2d-driven 3d object detection in rgb-d images. In Proceedings of the IEEE international conference on computer vision, pages 4622-4630, 2017. 6", + "[20] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1", + "[21] Jinpeng Lin, Zhihao Liang, Shengheng Deng, Lile Cai, Tao Jiang, Tianrui Li, Kui Jia, and Xun Xu. Exploring diversity-based active learning for 3d object detection in autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 2024. 3", + "[22] Zhuoming Liu, Hao Ding, Huaping Zhong, Weijia Li, Jifeng Dai, and Conghui He. Influence selection for active learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9274-9283, 2021. 2", + "[23] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[24] Yadan Luo, Zhuoxiao Chen, Zhen Fang, Zheng Zhang, Mahsa Baktashmotlagh, and Zi Huang. Kecor: Kernel coding rate maximization for active 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18279-18290, 2023. 2, 3, 7", + "[25] Yadan Luo, Zhuoxiao Chen, Zijian Wang, Xin Yu, Zi Huang, and Mahsa Baktashmotlagh. Exploring active 3d object detection from a generalization perspective. arXiv preprint arXiv:2301.09249, 2023. 2, 3, 4, 7", + "[26] Yi Ma, Harm Derksen, Wei Hong, and John Wright. Segmentation of multivariate mixed data via lossy data coding and compression. IEEE transactions on pattern analysis and machine intelligence, 29(9):1546-1562, 2007. 3" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20337", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Roberta Maisano, Valeria Tomaselli, Alessandro Capra, Francesco Longo, and Antonio Puliafito. Reducing complexity of 3d indoor object detection. In 2018 IEEE 4th International Forum on Research and Technology for Society and Industry (RTSI), pages 1-6. IEEE, 2018. 3", + "[28] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In 2015 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 922-928. IEEE, 2015. 2", + "[29] Michael Meyer and Georg Kuschk. Automotive radar dataset for deep learning based 3d object detection. In 2019 16th European radar conference (EuRAD), pages 129-132. IEEE, 2019. 3", + "[30] Peng Mi, Jianghang Lin, Yiyi Zhou, Yunhang Shen, Gen Luo, Xiaoshuai Sun, Liujuan Cao, Rongrong Fu, Qiang Xu, and Rongrong Ji. Active teacher for semi-supervised object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14482-14491, 2022. 1, 3", + "[31] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2906-2917, 2021. 1, 6", + "[32] Aimee Moses, Srikanth Jakkampudi, Cheryl Danner, and Derek Biega. Localization-based active learning (local) for object detection in 3d point clouds. In Geospatial Informatics XII, pages 44-58. SPIE, 2022. 3", + "[33] Yuzuru Nakamura, Yasunori Ishii, and Takayoshi Yamashita. Active domain adaptation with false negative prediction for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28782-28792, 2024. 5", + "[34] Hieu T Nguyen and Arnold Smeulders. Active learning using pre-clustering. In Proceedings of the twenty-first international conference on Machine learning, page 79, 2004. 1, 2", + "[35] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb-d data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 918-927, 2018. 6", + "[36] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 2, 3, 6", + "[37] Dan Roth and Kevin Small. Margin-based active learning for structured output spaces. In Machine Learning: ECML 2006: 17th European Conference on Machine Learning Berlin, Germany, September 18-22, 2006 Proceedings 17, pages 413-424. Springer, 2006. 2, 8", + "[38] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. In International Conference on Learning Representations, 2018. 1, 2, 7", + "[39] Claude Elwood Shannon. A mathematical theory of communication. The Bell system technical journal, 27(3):379-423, 1948. 2, 4, 8" + ], + "bbox": [ + 91, + 90, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Xiaoke Shen and Ioannis Stamos. Frustum voxelnet for 3d object detection from rgb-d or depth images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1698-1706, 2020. 2", + "[41] Yichao Shen, Zigang Geng, Yuhui Yuan, Yutong Lin, Ze Liu, Chunyu Wang, Han Hu, Nanning Zheng, and Baining Guo. V-detr: Detr with vertex relative position encoding for 3d object detection. arXiv preprint arXiv:2308.04409, 2023. 1", + "[42] Hualian Sheng, Sijia Cai, Na Zhao, Bing Deng, Jianqiang Huang, Xian-Sheng Hua, Min-Jian Zhao, and Gim Hee Lee. Rethinking iou-based optimization for single-stage 3d object detection. In European Conference on Computer Vision, pages 544-561. Springer, 2022. 2", + "[43] Hualian Sheng, Sijia Cai, Na Zhao, Bing Deng, Min-Jian Zhao, and Gim Hee Lee. Pdr: Progressive depth regularization for monocular 3d object detection. IEEE Transactions on Circuits and Systems for Video Technology, 33(12):7591-7603, 2023.", + "[44] Hualian Sheng, Sijia Cai, Na Zhao, Bing Deng, Qiao Liang, Min-Jian Zhao, and Jieping Ye. Ct3d++: Improving 3d object detection with keypoint-induced channel-wise transformer. International Journal on Computer Vision, 2025. 2", + "[45] Samarth Sinha, Sayna Ebrahimi, and Trevor Darrell. Variational adversarial active learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5972-5981, 2019. 2", + "[46] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 567-576, 2015. 1, 2, 4, 6", + "[47] OpenPCDet Development Team. Openpcdet: An opensource toolbox for 3d object detection from point clouds. https://github.com/open-mmlab/OpenPCDet, 2020.6", + "[48] Haiyang Wang, Lihe Ding, Shaocong Dong, Shaoshuai Shi, Aoxue Li, Jianan Li, Zhenguo Li, and Liwei Wang. Cagroup3d: Class-aware grouping for 3d object detection on point clouds. Advances in Neural Information Processing Systems, 35:29975-29988, 2022. 1, 2, 3, 6", + "[49] Jiangyi Wang, Zhongyao Cheng, Na Zhao, Jun Cheng, and Xulei Yang. On-the-fly point feature representation for point clouds analysis. arXiv preprint arXiv:2407.21335, 2024. 2", + "[50] Keze Wang, Dongyu Zhang, Ya Li, Ruimao Zhang, and Liang Lin. Cost-effective active learning for deep image classification. IEEE Transactions on Circuits and Systems for Video Technology, 27(12):2591-2600, 2016. 2", + "[51] Jamie Watson, Mohamed Sayed, Zawar Qureshi, Gabriel J. Brostow, Sara Vicente, Oisin Mac Aodha, and Michael Firman. Virtual occlusions through implicit depth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9053-9064, 2023. 1", + "[52] Junjie Wu. Advances in $K$ -means clustering: a data mining thinking. Springer Science & Business Media, 2012. 2, 6, 8", + "[53] Jiaxi Wu, Jiaxin Chen, and Di Huang. Entropy-based active learning for object detection with progressive diversity constraint. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9397-9406, 2022. 3, 4, 7" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20338", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Tsung-Han Wu, Yueh-Cheng Liu, Yu-Kai Huang, Hsin-Ying Lee, Hung-Ting Su, Ping-Chia Huang, and Winston H Hsu. Redal: Region-based and diversity-aware active learning for point cloud semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15510-15519, 2021. 2", + "[55] Qian Xie, Yu-Kun Lai, Jing Wu, Zhoutao Wang, Yiming Zhang, Kai Xu, and Jun Wang. Mlcvnet: Multi-level context votenet for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10447-10456, 2020. 2", + "[56] Qian Xie, Yu-Kun Lai, Jing Wu, Zhoutao Wang, Dening Lu, Mingqiang Wei, and Jun Wang. Venet: Voting enhancement network for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3712-3721, 2021. 2, 3", + "[57] Yating Xu, Conghui Hu, Na Zhao, and Gim Hee Lee. Generalized few-shot point cloud segmentation via geometric words. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21506-21515, 2023. 5", + "[58] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17830-17839, 2023. 1", + "[59] Chenhongyi Yang, Lichao Huang, and Elliot J Crowley. Plug and play active learning for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17784-17793, 2024. 3, 4, 7, 8", + "[60] Yi Yang, Zhigang Ma, Feiping Nie, Xiaojun Chang, and Alexander G Hauptmann. Multi-class active learning by uncertainty sampling with diversity maximization. International Journal of Computer Vision, 113:113-127, 2015. 2", + "[61] Tianning Yuan, Fang Wan, Mengying Fu, Jianzhuang Liu, Songcen Xu, Xiangyang Ji, and Qixiang Ye. Multiple instance active learning for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5330-5339, 2021. 3", + "[62] Bohang Zhang, Jikai Jin, Cong Fang, and Liwei Wang. Improved analysis of clipping algorithms for non-convex optimization. Advances in Neural Information Processing Systems, 33:15511-15521, 2020. 6", + "[63] Na Zhao and Gim Hee Lee. Static-dynamic co-teaching for class-incremental 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3436-3445, 2022. 2", + "[64] Na Zhao, Tat-Seng Chua, and Gim Hee Lee. Sess: Self-ensembling semi-supervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11079–11087, 2020. 2", + "[65] Na Zhao, Tat-Seng Chua, and Gim Hee Lee. Few-shot 3d point cloud semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8873-8882, 2021. 5", + "[66] Na Zhao, Peisheng Qian, Fang Wu, Xun Xu, Xulei Yang, and Gim Hee Lee. Sdcot++: Improved static-dynamic co" + ], + "bbox": [ + 91, + 92, + 485, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "teaching for class-incremental 3d object detection. IEEE Transactions on Image Processing, 2024. 2", + "[67] Qiang Zhou and Chaohui Yu. Point rcnn: An angle-free framework for rotated object detection. Remote Sensing, 14 (11):2605, 2022. 2" + ], + "bbox": [ + 516, + 92, + 903, + 161 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20339", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Uncertainty Meets Diversity_ A Comprehensive Active Learning Framework for Indoor 3D Object Detection/d2e21437-1e7e-4dc1-839a-f40a2bf9489e_model.json b/2025/Uncertainty Meets Diversity_ A Comprehensive Active Learning Framework for Indoor 3D Object Detection/d2e21437-1e7e-4dc1-839a-f40a2bf9489e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2cd88e12a75a7618ebc215d8366824b85fbc4884 --- /dev/null +++ b/2025/Uncertainty Meets Diversity_ A Comprehensive Active Learning Framework for Indoor 3D Object Detection/d2e21437-1e7e-4dc1-839a-f40a2bf9489e_model.json @@ -0,0 +1,2147 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.097, + 0.13, + 0.904, + 0.177 + ], + "angle": 0, + "content": "Uncertainty Meets Diversity: A Comprehensive Active Learning Framework for Indoor 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.273, + 0.204, + 0.726, + 0.258 + ], + "angle": 0, + "content": "Jiangyi Wang, Na Zhao* \nSingapore University of Technology and Design (SUTD) \nwangjiangyi0519@gmail.com, na.zhao@sutd.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.485, + 0.746 + ], + "angle": 0, + "content": "Active learning has emerged as a promising approach to reduce the substantial annotation burden in 3D object detection tasks, spurring several initiatives in outdoor environments. However, its application in indoor environments remains unexplored. Compared to outdoor 3D datasets, indoor datasets face significant challenges, including fewer training samples per class, a greater number of classes, more severe class imbalance, and more diverse scene types and intra-class variances. This paper presents the first study on active learning for indoor 3D object detection, where we propose a novel framework tailored for this task. Our method incorporates two key criteria - uncertainty and diversity - to actively select the most ambiguous and informative unlabeled samples for annotation. The uncertainty criterion accounts for both inaccurate detections and undetected objects, ensuring that the most ambiguous samples are prioritized. Meanwhile, the diversity criterion is formulated as a joint optimization problem that maximizes the diversity of both object class distributions and scene types, using a new Class-aware Adaptive Prototype (CAP) bank. The CAP bank dynamically allocates representative prototypes to each class, helping to capture varying intra-class diversity across different categories. We evaluate our method on SUN RGB-D and ScanNetV2, where it outperforms baselines by a significant margin, achieving over \\(85\\%\\) of fully-supervised performance with just \\(10\\%\\) of the annotation budget. Our code is available at: https://github.com/JoeWang-0519/CVPR25_UMD." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.763, + 0.222, + 0.779 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.788, + 0.484, + 0.88 + ], + "angle": 0, + "content": "3D object detection, which aims to localize and identify objects within a 3D scene, is crucial for various applications, including robotics [10], autonomous driving [20, 58], and augmented/virtual reality [51]. While deep learning techniques have achieved significant success in 3D object detection [31, 41, 48], most existing approaches are data-" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.292, + 0.898, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.566, + 0.907, + 0.707 + ], + "angle": 0, + "content": "Figure 1. Challenges in active learning for indoor 3D object detection, including both uncertainty and diversity aspects. We show two indoor scenes on ScanNetV2, with red boxes for ground truths and cyan for predictions (10% data using CAGroup3D detector), displaying only 'chair' and 'door' categories for clarity. For uncertainty, the presence of undetected objects and inaccurate detections undermines the quality of uncertainty estimation. For diversity, high scene-type diversity and varying intra-class variances in indoor environments complicate diverse sample selection. Our work presents a hybrid approach to address these challenges." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "hungry [41], requiring large amounts of annotated data for training. However, annotating objects in a 3D scene is a labor-intensive process. For example, annotating a single scene in the SUN RGB-D dataset takes over 1.9 minutes per object, with an average of approximately 14 objects per scene [46]. To mitigate the high costs of 3D annotation, active learning (AL) [30, 34, 38] offers a promising solution. AL seeks to maximize the utility of limited labeling resources by actively selecting the most informative samples for human annotation. Through this process, AL can achieve performance comparable to fully-supervised methods, but with significantly fewer labeled samples." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.359, + 0.901 + ], + "angle": 0, + "content": "*Corresponding Author: na_zhao@sutd.edu.sg" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20329" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.318 + ], + "angle": 0, + "content": "Although active learning has proven effective in image-based tasks such as image classification and object detection, using uncertainty [5, 16, 37] or/and diversity [34, 38, 60] as selection criteria, its application to 3D object detection remains largely underexplored. To date, only a few studies [12, 24, 25] have applied AL techniques to outdoor 3D object detection. For instance, CRB [25] uses label conciseness to capture uncertainty, along with feature representativeness and geometric balance to enhance diversity for sample selection. KECOR [24] favors samples whose latent features have the highest coding length, thereby ensuring diversity. While these methods demonstrate effectiveness in outdoor scenarios, they are suboptimal for indoor environments, as illustrated in Tab. 2, due to the unique challenges posed by 3D object detection in indoor settings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.321, + 0.483, + 0.518 + ], + "angle": 0, + "content": "On the one hand, a 3D indoor dataset [6] could be \\(6.2\\mathbf{x}\\) smaller in scale but contain \\(3.3\\mathbf{x}\\) more number of categories, with more severe class imbalance, compared to an outdoor dataset [11], as shown in [14]. This scarcity of training samples, coupled with pronounced class imbalance, impairs 3D detection accuracy in indoor scenarios, leading to more frequent occurrences of inaccurate detection and undetected objects, as illustrated in Fig. 1. In this figure, three detections are inaccurate, incorrectly localizing background objects, while three other objects remain undetected. These challenges underscore the need to account for both inaccurately detected proposals and undetected objects when selecting informative samples from unlabeled data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.483, + 0.657 + ], + "angle": 0, + "content": "On the other hand, unlike outdoor scenes, which primarily feature road environments with a limited set of object classes, indoor environments exhibit greater scene-type diversity and intra-class variances, caused by diverse layouts and variations in object shapes within each class. For example, as shown in Fig. 1, five 'chair' instances can form three clusters in the embedding space due to the variations in shape and occlusion. This phenomenon in indoor environments complicate the selection of diverse samples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this work, we present the first solution to apply active learning in indoor 3D object detection, utilizing both uncertainty and diversity as selection criteria. For uncertainty, we design a two-pronged epistemic uncertainty estimation strategy to address inaccurate detections and undetected objects, which first assesses uncertainty separately for each case, then integrates them into a unified criterion. Specifically, we introduce a localization-aware uncertainty score that combines the IoU score – measured between the original detected object and its perturbed version – with the Shannon entropy. This helps mitigate the negative effects of invalid proposals, such as background or partial detections, when estimating the classification uncertainty of detected proposals. To quantify the epistemic uncertainty arising from undetectability, we develop a lightweight network that estimates the number of undetected objects. For di" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.242 + ], + "angle": 0, + "content": "versity, we frame the diverse sample selection problem as an optimization task that jointly maximizes scene-type diversity and intra-class variances, leveraging a novel Class-aware Adaptive Prototype (CAP) bank. This CAP bank is designed to capture varying intra-class variances by generating an adaptive number of representative prototypes per class. Observing that different scene types exhibit distinct histograms of object categories as shown in Fig. 1, each composed of various prototypes, we use the CAP bank to model both class-level and scene-level diversity." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.243, + 0.905, + 0.334 + ], + "angle": 0, + "content": "We evaluate the proposed method on SUN RGB-D [46] and ScanNet [6] datasets. Using CAGroup3D [48] as base detection model, our approach achieves \\(56.11\\%\\), \\(61.77\\%\\) \\(\\mathrm{mAP}@0.25\\) on SUN RGB-D and ScanNet, respectively, reaching \\(87.24\\%\\) and \\(85.02\\%\\) of the fully-supervised performance with only \\(10\\%\\) of the annotation budget." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.346, + 0.655, + 0.362 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.371, + 0.907, + 0.763 + ], + "angle": 0, + "content": "Generic Active Learning. The majority of existing works have explored active learning (AL) strategies on image classification task, which can be categorized into uncertainty-based [5, 9, 16, 37, 50] and diversity-based [1, 7, 34, 38, 54, 60] methods based on the selection criteria used. Uncertainty-based approaches select ambiguous samples for annotation by maximizing uncertainty scores, which can be computed in various ways, such as entropy [9, 17, 39], minimum posterior probability [50], or multi-class margin [5, 16, 37]. Shannon entropy [39] is a commonly used metric in classification that quantifies uncertainty in the probability distribution, while multi-class margin [37] measures the uncertainty through the difference between the top two predicted probabilities. MC Dropout [9] employs Monte Carlo Dropout to approximate the posterior predicted probability and use Shannon entropy to assess the underlying uncertainty. Diversity-based methods [1, 38, 60] select a compact subset of samples that best represent the global data distribution. For example, Core-set [38] formulates active learning as a core-set selection problem and solves it using the \\(k\\)-centroid algorithm. Recent works [4, 22, 45] combine both uncertainty and diversity through hybrid approaches. A representative approach, BADGE [4], balances uncertainty and diversity by applying the \\(k\\)-means++ [52] algorithm in the gradient space, which considers both magnitude and direction of gradient vectors." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.765, + 0.909, + 0.902 + ], + "angle": 0, + "content": "3D Object Detection. We focus on indoor 3D object detection [15, 63, 64, 66] and exclude outdoor methods [3, 42-44], which differ significantly. Prior indoor 3D object detection methods can be categorized into two lines in terms of underlying representation, i.e., voxel-based methods [13, 28, 40], and point-based methods [36, 49, 55, 56, 67]. Voxel-based methods convert point cloud to voxel grids and use 3D convolution networks to address the irregularity of point cloud. In contrast, point-based methods extract point" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20330" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.089, + 0.868, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.322, + 0.908, + 0.381 + ], + "angle": 0, + "content": "Figure 2. Overview of our proposed AL framework for indoor 3D object detection, exploiting both uncertainty and diversity. In the \\(r\\)-th round, we first construct a candidate pool of size \\(\\delta \\cdot B_r\\) using the Top-\\(K\\) unified uncertainty score, which accounts for both inaccurate detections and undetections. Then, we jointly optimize the intra-class and scene-type diversity to select the \\(r\\)-th selected dataset \\(\\mathcal{D}_r\\). Lastly, along with the previous labeled point clouds \\(\\mathcal{D}_L^{r-1}\\), we retrain the model until the total labeled dataset reaches the annotation budget \\(B\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.396, + 0.484, + 0.547 + ], + "angle": 0, + "content": "level features and use a voting mechanism [36, 56] to encourage points to vote towards their corresponding object centers, followed by grouping the vote points into clusters and then generating object proposals. In this work, we adopt the state-of-the-art method in indoor 3D object detection, CAGroup3D [48], which takes advantage of the two lines of work through a two-stage fully sparse convolutional 3D detection framework. It extracts voxel-wise features from raw point clouds and leverages 'voting and grouping' mechanism to generate object proposals." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.554, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Active Learning for Object Detection. Despite extensive studies on AL in image classification, applying it to image-based object detection is challenging due to the task's complexity, which involves both localization and classification of multiple objects. Several works [1, 18, 30, 53, 59, 61] have investigated active object detection by adapting generic AL techniques to this task. For example, CDAL [1] designs contextual diversity to capture spatial and semantic context diversity within a dataset, while DivProto [53] and PPAL [59] incorporate both diversity and uncertainty. While these methods yield promising results in 2D object detection, they cannot be directly applied to 3D object detection due to the significant modality gap. Recently, driven by the need to reduce annotation burden in 3D object detection for autonomous driving, a few studies [8, 12, 21, 24, 25, 27, 29, 32] have explored active learning for 3D object detection in outdoor environments. Notably, CRB [25] selects active samples with concise labels, representative features, and balanced density progressively. KECOR [24] leverages kernel coding rate [26] to measure the richness of latent features. However, these outdoor methods are suboptimal for indoor environments, where 3D object detection faces unique challenges. To ad" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.396, + 0.907, + 0.487 + ], + "angle": 0, + "content": "dress these, this paper presents the first solution for applying active learning to indoor 3D object detection, overcoming these challenges by incorporating designs to estimate uncertainty in both inaccurate detections and undetections, while simultaneously optimizing diversity with respect to scene-type and intra-class variances." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.507, + 0.649, + 0.525 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.534, + 0.907, + 0.78 + ], + "angle": 0, + "content": "Given a point cloud \\(\\mathcal{P}_i = \\{(x,y,z)\\}\\), a 3D object detection model aims to predict the locations and class labels for a set of objects \\(\\mathfrak{B}_i = \\{(\\mathbf{b}_k,y_k)\\}\\), where \\(\\mathbf{b}_k\\in \\mathbb{R}^7\\) represents the parameter of the \\(k\\)-th bounding box, including the center position, box size and yaw angle, and \\(y_{k}\\) indicates its predicted class label. In the active learning setting, our objective is to train the 3D detection model by constructing a small labeled dataset \\(\\mathcal{D}_L\\), which starts with a small portion of random samples \\(\\mathcal{D}_L^0\\) and is progressively expanded over \\(R\\) rounds from a large pool \\(\\mathcal{D}\\). In each active learning round, \\(r\\in \\{1,\\dots,R\\}\\), \\(B_{r}\\) samples are selected from the unlabeled dataset \\(\\mathcal{D}_U = \\mathcal{D} / \\mathcal{D}_L^{r - 1}\\) by an active learning policy, and labeled by an oracle, which constitute the \\(r\\)-th selected dataset \\(\\mathcal{D}_r\\). Then, \\(\\mathcal{D}_L\\) will be updated as \\(\\mathcal{D}_L^r\\gets \\mathcal{D}_L^{r - 1}\\cup \\mathcal{D}_r\\) until the size of \\(\\mathcal{D}_L\\) reaches the final budget \\(B\\), i.e., \\(|\\mathcal{D}_L^0 | + \\sum_{r = 1}^{R}B_r = B\\) where \\(B\\ll |\\mathcal{D}|\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.902 + ], + "angle": 0, + "content": "To tackle active 3D object detection in indoor scenarios, we propose a novel framework, illustrated in Fig. 2, that leverages both uncertainty and diversity as selection criteria. Following previous work [21, 53, 59] that considers both uncertainty and diversity, we adopt a budget expansion ratio \\(\\delta \\in \\mathbb{Z}^{+}\\) to first select a candidate pool \\(\\mathcal{D}_r^{\\mathrm{cand}}\\) of size \\(K = \\delta \\cdot B_r\\) based on the uncertainty criterion for \\(r\\)-th active learning round. Then, from this pool, the \\(\\mathcal{D}_r\\) of size" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20331" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.09, + 0.483, + 0.134 + ], + "angle": 0, + "content": "
DatasetmAP@.25mAP@.50mAR@.25mAR@.50
SUN RGB-D [46]28.9112.4867.1226.09
ScanNetV2 [6]35.7520.4076.6542.48
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.141, + 0.483, + 0.198 + ], + "angle": 0, + "content": "Table 1. Motivation for designing epistemic uncertainty estimation. We report the mean average precision (mAP) and mean average recall (mAR) at two IoU thresholds (0.25 and 0.50) for the initial detection model, which is trained using \\(2\\%\\) labeled data." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.205, + 0.484, + 0.357 + ], + "angle": 0, + "content": "\\(B_{r}\\) is chosen based on the diversity criterion. For the uncertainty criterion, we propose a two-pronged epistemic uncertainty estimation strategy that generates a unified score, accounting for uncertainty in both detected and undetected objects. This is achieved through a localization-aware uncertainty score and an estimate of the undetected object count (Sec. 3.1). For the diversity criterion, we frame the diverse sample selection as an optimization problem and propose an efficient solution that jointly optimizes intra-class variances and scene-type diversity (Sec. 3.2)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.365, + 0.388, + 0.382 + ], + "angle": 0, + "content": "3.1. Epistemic Uncertainty Estimation" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.387, + 0.484, + 0.705 + ], + "angle": 0, + "content": "In Tab. 1, we evaluate the detection performance of a 3D object detection model trained on the initial random samples \\( D_{L}^{0} \\) (2% labeled data). As shown, the model achieves 20.40% mAP@.50 and 42.48% mAR@.50 on ScanNet [6], indicating that approximately 80% of the predictions are inaccurate and 60% of ground-truth objects remain undetected. The results are even worse on the more challenging SUN RGB-D [46] dataset. These findings highlight the need to address the two types of epistemic uncertainty in 3D object detection: inaccurate detections and undetected objects. Given that the detection task involves both localization and classification, we estimate the epistemic uncertainty of inaccurate detections by considering both aspects. Specifically, we introduce a localization-aware uncertainty score that exploits an IoU score for localization reliability to combine with Shannon entropy. To handle undetections, we design a lightweight network to estimate the number of undetected objects, which serves as the uncertainty measure for undetections. These two uncertainty scores are then integrated into a unified uncertainty score, which is used for Top-K selection to generate the candidate pool \\( \\mathcal{D}_r^{\\mathrm{cand}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.812 + ], + "angle": 0, + "content": "Localization-aware Uncertainty Score. To estimate the uncertainty of detected objects from a point cloud \\(\\mathcal{P}_i\\), we first compute the classification uncertainty using the Shannon entropy [39] of the predicted class probabilities for all detected object proposals. For the \\(j\\)-th proposal, with predicted class probabilities over \\(C\\) classes denoted as \\(\\mathbf{p}_j = \\{p_{j,c}\\}_{c=1}^C \\in \\mathbb{R}^C\\), the Shannon entropy is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.821, + 0.483, + 0.862 + ], + "angle": 0, + "content": "\\[\nE (\\mathbf {p} _ {j}) = - \\sum_ {c = 1} ^ {C} p _ {j, c} \\log (p _ {j, c}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "However, not all detected proposals are valid, as some may correspond to background or partial objects. To mitigate" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.244 + ], + "angle": 0, + "content": "the impact of these invalid proposals, we incorporate an IoU score that measures the similarity between the original detection and its perturbed version, obtained by applying weak augmentations (e.g., random rotation and scaling) to the input point cloud. The intuition is that if the model is confident about an object, small perturbations should not significantly affect its predicted position, while background or partial proposals will likely produce inconsistent predictions. Consequently, the localization-aware uncertainty score for detected objects in \\(\\mathcal{P}_i\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.25, + 0.907, + 0.294 + ], + "angle": 0, + "content": "\\[\n\\mathbf {U} _ {\\det } \\left(\\mathcal {P} _ {i}\\right) := \\frac {\\sum_ {j = 1} ^ {\\left| \\mathfrak {B} _ {i} \\right|} \\operatorname {I o U} _ {j} \\cdot E \\left(\\mathbf {p} _ {j}\\right)}{\\sum_ {j = 1} ^ {\\left| \\mathfrak {B} _ {i} \\right|} \\operatorname {I o U} _ {j}}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.299, + 0.907, + 0.344 + ], + "angle": 0, + "content": "Our IoU scores serve as soft weights, adaptively reducing the contribution of invalid objects in the epistemic uncertainty estimation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.345, + 0.908, + 0.556 + ], + "angle": 0, + "content": "Undetected Number Prediction Network. When estimating uncertainty for active object detection, previous works [25, 53, 59] typically focus only on the uncertainty within the detected objects. However, the uncertainty of undetected objects is equally important, especially in the context of active learning for indoor 3D object detection, where model performance is often hindered by the scarcity of training samples and severe class imbalance. As shown by the mean average recall (mAR) scores in Tab. 1, more than half of the ground truth objects remain undetected in both 3D indoor datasets. This observation highlights the need to account for the uncertainty of undetected objects. To address this, we design a lightweight network to estimate the number of undetected objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.557, + 0.909, + 0.799 + ], + "angle": 0, + "content": "As depicted in Fig. 2, we first extract the feature map \\(\\mathbf{f}_i\\) from the point cloud \\(\\mathcal{P}_i\\) using the feature extractor. We then apply global average pooling (GAP) on \\(\\mathbf{f}_i\\) to obtain the global feature \\(\\mathbf{g}_i\\). Simultaneously, we create a masked feature map by masking out the regions corresponding to the detected bounding boxes produced by RoI refinement. This masked feature map is passed through GAP to yield the undetected feature \\(\\mathbf{u}_i\\). We concatenate these two features, \\(\\mathbf{g}_i\\) and \\(\\mathbf{u}_i\\), and use the concatenation as input to our undetected object number prediction network \\(\\Theta_{\\mathrm{undet}}\\), a lightweight network consisting of three-layer MLPs. The output of the network, i.e. \\(\\Theta_{\\mathrm{undet}}([\\mathbf{g}_i; \\mathbf{u}_i])\\) represents the uncertainty score for undetections, denoted as \\(\\mathbf{U}_{\\mathrm{undet}}\\). This output is compared with the true number of undetected objects, computed based on the detected objects \\(\\mathfrak{B}_i\\) and the ground-truth objects \\(\\mathfrak{B}_i^*\\), to compute the loss for training \\(\\Theta_{\\mathrm{undet}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.806, + 0.907, + 0.849 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {u n d e t}} = \\sum_ {i = 1} ^ {| \\mathcal {D} _ {L} |} \\left(\\Theta_ {\\text {u n d e t}} \\left(\\left[ \\mathbf {g} _ {i}; \\mathbf {u} _ {i} \\right]\\right) - \\mathcal {N} \\left(\\mathfrak {B} _ {i}, \\mathfrak {B} _ {i} ^ {*}\\right)\\right) ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}(\\cdot, \\cdot)\\) is a function that computes the number of undetected objects, with details provided in the supplementary material. Additionally, to enhance the generalizability" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20332" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.244 + ], + "angle": 0, + "content": "of the network \\(\\Theta_{\\mathrm{undet}}\\), we employ random masking of the predicted boxes during the training phase, encouraging the network to better capture the undetectability in point clouds. Unified Uncertainty Score. After obtaining the two uncertainty scores, \\(\\mathbf{U}_{\\mathrm{det}}\\) and \\(\\mathbf{U}_{\\mathrm{undet}}\\), we aim to combine them into a unified uncertainty score. However, these two scores are on different scales, making their direct product infeasible. To address this issue, we introduce a normalization step, inspired by [33], using a hyperparameter \\(k\\). Specifically, we normalize each score as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.252, + 0.483, + 0.285 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {U}} _ {\\star} (\\mathcal {P} _ {i}) = \\max (0, \\frac {\\mathbf {U} _ {\\star} (\\mathcal {P} _ {i}) - \\mu (\\mathbf {U} _ {\\star})}{k \\cdot \\sigma (\\mathbf {U} _ {\\star})} + 0. 5), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.296, + 0.484, + 0.357 + ], + "angle": 0, + "content": "where \\(\\star \\in \\{\\mathrm{det},\\mathrm{undet}\\}\\), and \\(\\mu (\\cdot)\\) and \\(\\sigma (\\cdot)\\) denote the mean and standard deviation, respectively. Once the scores are normalized, we compute their product to yield the unified uncertainty score." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.366, + 0.433, + 0.383 + ], + "angle": 0, + "content": "3.2. Prototype-based Diversity Optimization" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.388, + 0.484, + 0.764 + ], + "angle": 0, + "content": "As illustrated in Fig. 1, indoor environments exhibit significant intra-class variances and scene-type diversity, both of which motivate our approach to the diversity criterion. In 3D indoor scenes, the intra-class variances can vary greatly across object categories due to the differences in shape and occlusion. For example, the variance in 'chairs' is typically high due to the wide variety of shapes and designs, whereas 'doors' tend to have lower variance due to their more standardized geometry. This variability makes methods that model each class with a fixed number of prototypes [57, 65] insufficient to capture the intra-class distributions. To address this, we propose a novel Class-aware Adaptive Prototype (CAP) bank, which dynamically allocates a varying number of prototypes for each class, inferred from the predicted object proposals. Additionally, we observe that indoor scenes naturally form distinct clusters based on the distribution of object categories. For instance, bathrooms typically contain toilets and bathtubs, while bedrooms feature beds and dressers. Through this observation, we represent each scene as a histogram of the prototypes corresponding to the predicted object categories within the scene. As both intra-class variances and scene-type diversity can be modeled using the CAP bank, we frame diverse sample selection as an optimization problem that jointly maximizes both aspects, and propose an efficient solution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Class-aware Adaptive Prototype Bank. The design of our Class-aware Adaptive Prototype (CAP) bank is inspired by infinite mixture prototypes (IMP) [2], which was originally proposed for few-shot learning. IMP allows for inferring the number of clusters from the data, rather than relying on a pre-defined number of clusters. To build the CAP bank, we first iterate over the initial labeled set \\( D_{L}^{r - 1} \\) to initialize one prototype per class. Specifically, we leverage the ground-truth annotations to extract features for all objects belong-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.907, + 0.304 + ], + "angle": 0, + "content": "ing to the \\(c\\)-th class in \\(D_L^{r-1}\\), and compute their average as to form the initial prototype \\(\\mu_{1,c}\\) for class \\(c\\). Subsequently, we iterate over the unlabeled dataset \\(\\mathcal{D}_U\\) and update the prototypes for each class, resulting in our CAP bank, denoted as \\(\\mathbf{M} = \\{\\{\\pmb{\\mu}_{k,c}\\}_{k=1}^{M_c}\\}_{c=1}^C\\), where \\(\\pmb{\\mu}_{k,c}\\) represents the \\(k\\)-th prototype for class \\(c\\), \\(M_c\\) is the number of prototypes for class \\(c\\), which may vary between classes, and \\(C\\) is the total number of classes. In each iteration (one batch), we use the detection model from the previous round to generate object proposals and extract their features and predicted class labels. For each predicted object, we compute the cosine similarity between its feature embedding \\(\\mathbf{o}_j\\) and each of the existing prototypes in its predicted class. This similarity score is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.31, + 0.907, + 0.348 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S i m} \\left(\\mathbf {o} _ {j}, \\boldsymbol {\\mu} _ {k, c}\\right) = \\frac {\\mathbf {o} _ {j} ^ {T} \\cdot \\boldsymbol {\\mu} _ {k , c}}{\\left| \\left| \\mathbf {o} _ {j} \\right| \\right| _ {2} \\left| \\left| \\boldsymbol {\\mu} _ {k , c} \\right| \\right| _ {2}} \\in [ - 1, 1 ]. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.357, + 0.907, + 0.403 + ], + "angle": 0, + "content": "After calculating the similarity scores, each object is either assigned to an existing prototype or used to form a new prototype, according to the following assignment rule:" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.413, + 0.906, + 0.475 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} \\left(\\mathbf {o} _ {j}\\right) = \\left\\{ \\begin{array}{l} M _ {c} + 1, \\text {i f} \\max _ {k} \\operatorname {S i m} \\left(\\mathbf {o} _ {j}, \\boldsymbol {\\mu} _ {k, c}\\right) < \\tau_ {\\text {s i m}}, \\\\ \\arg \\max _ {k} \\operatorname {S i m} \\left(\\mathbf {o} _ {j}, \\boldsymbol {\\mu} _ {k, c}\\right), \\quad \\text {o t h e r w i s e}. \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.909, + 0.627 + ], + "angle": 0, + "content": "Here, \\(\\tau_{\\mathrm{sim}}\\) is similarity threshold, which is set to 0.3 by default. Finally, the prototypes - whether assigned or newly formed - are updated using the new object proposals in each batch. However, since many predicted bounding boxes might be invalid, especially in the early stages of training (e.g., as shown in Tab. 1), we seek to mitigate the influence of these invalid proposals when updating the prototypes. To this end, we incorporate the IoU score from Eq. 2 into the prototype update process, effectively weighing each object's contribution to the prototype updates as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.636, + 0.907, + 0.68 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} _ {k, c} = \\frac {\\sum_ {j = 1} ^ {| \\mathfrak {B} _ {i} |} \\mathbb {1} \\left\\{\\mathcal {A} \\left(\\mathbf {o} _ {j}\\right) , k \\right\\} \\cdot \\operatorname {I o U} _ {j} \\cdot \\mathbf {o} _ {j}}{\\sum_ {j = 1} ^ {| \\mathfrak {B} _ {i} |} \\mathbb {1} \\left\\{\\mathcal {A} \\left(\\mathbf {o} _ {j}\\right) , k \\right\\} \\cdot \\operatorname {I o U} _ {j}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.906, + 0.735 + ], + "angle": 0, + "content": "where the indicator function \\(\\mathbb{1}(m,n)\\) equals to 1 if \\(m = n\\), and 0 otherwise. The complete CAP bank creation procedure is outlined in Algo. 1 of the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Prototype-based Diversity Optimization. With the generated CAP bank, we can characterize each class through a set of prototypes, and further represent objects as histograms of these prototypes. These histograms allow us to capture the intra-class variance for each class, as objects within the same class are reflected in the variation of their corresponding histograms. Furthermore, based on the observation that each scene can be characterized by the distribution of object categories, which distinguishes one scene type from another, we also model scenes as histograms of prototypes. Therefore, this histogram-based representation" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20333" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.184 + ], + "angle": 0, + "content": "implicitly captures the scene type. Since both intra-class variance and scene-type diversity are encoded in the prototype histograms, we design a novel and efficient optimization method to quantify the diversity criterion by considering both these two aspects. Specifically, we formulate the optimization for diverse sample selection as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.192, + 0.482, + 0.311 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\max _ {\\mathcal {D} _ {r} \\subset \\mathcal {D} _ {r} ^ {\\text {c a n d}}} E (\\mathbf {h} (\\mathcal {D} _ {r})) + \\beta | | \\mathbf {h} (\\mathcal {D} _ {r}) | | _ {1}, \\\\ \\begin{array}{l} \\text {s . t .} \\quad \\mathbf {h} (\\mathcal {D} _ {r}) = \\sum_ {\\mathcal {P} _ {i} \\in \\mathcal {D} _ {r}} \\mathbf {h} (\\mathcal {P} _ {i}) = \\sum_ {\\mathcal {P} _ {i} \\in \\mathcal {D} _ {r}} \\sum_ {\\mathbf {o} _ {j} \\in \\mathcal {P} _ {i}} \\mathbf {h} (\\mathbf {o} _ {j}), \\end{array} \\\\ \\mathbf {h} \\left(\\mathbf {o} _ {j}\\right) = \\sum_ {c = 1} ^ {C} \\sum_ {k = 1} ^ {M _ {c}} [ \\text {o n e} \\_ \\text {h o t} (\\mathcal {A} \\left(\\mathbf {o} _ {j}\\right)) ] ^ {c, k}, \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.312, + 0.484, + 0.552 + ], + "angle": 0, + "content": "where \\(\\mathbf{h}(\\mathbf{o}_j)\\) represents the assignment histogram of \\(j\\)-th object w.r.t. the CAP bank \\(\\mathbf{M}\\), while \\(\\mathbf{h}(\\mathcal{P}_i)\\) and \\(\\mathbf{h}(\\mathcal{D}_r)\\) represent the histograms of all objects in the individual point cloud \\(\\mathcal{P}_i\\) and the selected labeled samples \\(\\mathcal{D}_r\\) at the \\(r\\)-th round, respectively. The two terms in this objective function are designed to select samples that are both diverse and holistic across all prototypes in the CAP bank. The first term captures the intra-class variance for all classes, ensuring that the selected samples represent the diversity within each class as well as across different classes. The second term promotes scene-level diversity by encouraging coverage across different object categories, which implicitly increases the diversity in scene types. Given that the first term (entropy) and the second term (\\(l_1\\)-norm) have different magnitudes, we introduce a hyperparameter \\(\\beta\\) to rescale the second term. In our experiments, we set \\(\\beta = 0.01\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.645 + ], + "angle": 0, + "content": "However, directly solving the optimization in Eq. 8 is computationally expensive due to the combinatorial search over \\( |\\mathcal{D}_r^{\\mathrm{cand}}| \\) samples. To efficiently narrow the search space while maintaining solution quality, we leverage one key property of the objective function over \\( S \\) disjoint partitions \\( \\{\\mathcal{D}_{r,s}\\}_{s = 1}^{S} \\) of \\( \\mathcal{D}_r \\), i.e., \\( \\cup_{s}\\mathcal{D}_{r,s} = \\mathcal{D}_r \\):" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.647, + 0.482, + 0.719 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} E \\left(\\mathbf {h} \\left(\\mathcal {D} _ {r}\\right)\\right) + \\beta | | \\mathbf {h} \\left(\\mathcal {D} _ {r}\\right) | | _ {1} \\\\ \\stackrel {c} {=} \\sum_ {s = 1} ^ {S} \\left\\{\\frac {N _ {s}}{N _ {\\text {t o t a l}}} E \\left(\\mathbf {h} \\left(\\mathcal {D} _ {r, s}\\right)\\right) + \\beta \\left\\| \\mathbf {h} \\left(\\mathcal {D} _ {r, s}\\right) \\right\\| _ {1} \\right\\} + \\mathcal {O} \\left(\\frac {N _ {\\text {i n t e r}}}{N _ {\\text {t o t a l}}}\\right), \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{h}(\\mathcal{D}_{r,s})\\) represents the histograms of all objects in \\(s\\)-th partition \\(\\mathcal{D}_{r,s}\\), \\(N_{s}\\) and \\(N_{\\mathrm{total}}\\) are the number of objects in the \\(s\\)-th partition \\(\\mathcal{D}_{r,s}\\) and selected labeled samples \\(\\mathcal{D}_r\\), \\(N_{\\mathrm{inter}}\\) is the number of objects that share the same prototypes but belong to different partitions, and \\(\\mathcal{O}(\\cdot)\\) is the Big O notation from mathematics. To align with the \\(S\\) disjoint partitions of \\(\\mathcal{D}_r\\), we use k-means++ [52] to group the candidate pool \\(\\mathcal{D}_r^{\\mathrm{cand}}\\) into \\(S\\) clusters, denoted as \\(\\{\\mathcal{D}_{r,s}^{\\mathrm{cand}}\\}_{s = 1}^{S}\\), each of which is approximately disjoint with each other w.r.t. prototype assignments. This guarantees that \\(\\frac{N_{\\mathrm{inter}}}{N_{\\mathrm{total}}} \\to 0\\), and then allows us to decompose the original optimization problem into \\(S\\) sub-problems within each cluster \\(\\mathcal{D}_{r,s}^{\\mathrm{cand}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.088, + 0.91, + 0.215 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\max _ {\\mathcal {D} _ {r, s} \\subset \\mathcal {D} _ {r, s} ^ {\\text {c a n d}}} E (\\mathbf {h} (\\mathcal {D} _ {r, s})) + \\beta \\frac {N _ {\\text {t o t a l}}}{N _ {s}} | | \\mathbf {h} (\\mathcal {D} _ {r, s}) | | _ {1}, \\\\ \\begin{array}{l} \\text {s . t .} \\quad \\mathbf {h} (\\mathcal {D} _ {r, s}) = \\sum_ {\\mathcal {P} _ {i} \\in \\mathcal {D} _ {r, s}} \\mathbf {h} (\\mathcal {P} _ {i}) = \\sum_ {\\mathcal {P} _ {i} \\in \\mathcal {D} _ {r, s}} \\sum_ {\\mathbf {o} _ {j} \\in \\mathcal {P} _ {i}} \\mathbf {h} (\\mathbf {o} _ {j}), \\end{array} \\\\ \\mathbf {h} \\left(\\mathbf {o} _ {j}\\right) = \\sum_ {c = 1} ^ {C} \\sum_ {k = 1} ^ {M _ {c}} [ \\text {o n e} \\text {h o t} (\\mathcal {A} \\left(\\mathbf {o} _ {j}\\right)) ] ^ {c, k}. \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.217, + 0.907, + 0.293 + ], + "angle": 0, + "content": "We use greedy search to solve the \\(S\\) sub-problems to obtain the diverse samples \\(\\mathcal{D}_{r,s}\\) for each cluster, and combine them together as the selected dataset \\(\\mathcal{D}_r\\) in the \\(r\\)-th active learning round. We provide the detailed derivation of Eq. 9 in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.308, + 0.646, + 0.325 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.334, + 0.712, + 0.35 + ], + "angle": 0, + "content": "4.1. Experimental Setups" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.356, + 0.907, + 0.582 + ], + "angle": 0, + "content": "3D Detection Datasets. We evaluate our proposed method on two 3D indoor object detection benchmarks: SUN RGB-D [46] and ScanNet [6]. SUN RGB-D is an indoor dataset for 3D object detection, which contains 10,335 RGB-D images with 3D room layouts from four different sensors. Officially, they are split into 5,285 training samples and 5,050 testing samples. Following the standard evaluation protocol [19, 35, 36], we evaluate the performance on 10 most common categories. ScanNetV2 is a large 3D dataset with 1,513 RGB-D scans of 707 indoor scenes. These scans are split into 1,201 samples for training and 312 samples for testing, each of which is annotated with dense semantic segmentation masks. We follow [36] to derive the axis-aligned bounding boxes without considering the orientation, and adopt 18 object categories for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.906, + 0.825 + ], + "angle": 0, + "content": "Implementation Details. We use the state-of-the-art CA-Group3D [48] as the detection model, and implement our code base using OpenPCDet [47] toolkit. Following the training strategies in [48], we set the voxel size as \\(0.02\\mathrm{m}\\). For both datasets, the network is trained by AdamW optimizer [23] with an initial learning rate and weight decay rate of 0.001 and 0.0001 respectively. The gradnorm clip [62] is applied to stabilize the training process. For each active learning round, SUN RGB-D takes 48 epochs to train on the updated labeled dataset, and the learning rate decays on [32, 44] epochs with decay rate of 0.1. ScanNet requires 120 epochs for training and decays the learning rate on [80, 110] epochs by the same decay rate. All experiments are conducted using one 3090Ti NVIDIA GPU. We set the similarity threshold \\(\\tau_{\\mathrm{sim}}\\) to 0.3, balance ratio \\(\\beta\\) to 0.01, and budget expanding ratio \\(\\delta\\) to 6 empirically." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.9 + ], + "angle": 0, + "content": "Evaluation. Following the evaluation protocol in [31, 36, 48], we report mean average precision (mAP) with different IoU thresholds, i.e., 0.25 and 0.50. Additionally, to investigate the effectiveness of different active learning strategies, we record the mAP with \\(2\\%\\), \\(4\\%\\), \\(6\\%\\), \\(8\\%\\), \\(10\\%\\) of the train" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20334" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.132, + 0.083, + 0.861, + 0.281 + ], + "angle": 0, + "content": "
MethodsPresened atSUN RGB-DScanNetV2
mAP@0.25 (%)mAP@0.50 (%)mAP@0.25 (%)mAP@0.50 (%)
2D-GenerieRAND-42.7125.1455.0939.74
MC-DROPOUT [9]ICML'1747.4828.3645.7128.72
CORESET [38]ICLR'1846.9928.1255.3440.08
BADGE [4]ICLR'2049.1230.5457.1041.97
2D-ODCDAL [1]ECCV'2050.8630.7557.9942.33
DIVPROTO [53]CVPR'2252.3132.5459.0143.10
PPAL [59]CVPR'2452.5533.5159.4443.15
3D-ODCRB [25]ICLR'2350.8931.3058.7342.71
KECOR [24]ICCV'2351.0731.4158.5442.56
OURS-56.1135.4261.7745.93
ORACLE†-64.3147.4372.6558.81
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.282, + 0.907, + 0.327 + ], + "angle": 0, + "content": "Table 2. mAP@0.25 (%) and mAP@0.50 (%) scores of the proposed method and AL baselines on the SUN RGB-D and ScanNetV2 benchmarks with \\(10\\%\\) queried point clouds. CAGroup3D is used as the detection model for all approaches. \\(\\dagger\\): indicates the detection results using fully-annotated data, and serves as the upper bound for all active learning strategies." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.338, + 0.482, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.442, + 0.482, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.563, + 0.483, + 0.593 + ], + "angle": 0, + "content": "Figure 3. mAP@0.25 (\\%) score of the proposed method and AL baselines on SUN RGB-D and ScanNetV2 benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.483, + 0.667 + ], + "angle": 0, + "content": "ing set for both datasets, where the first \\(2\\%\\) samples are selected randomly. In each active learning round, \\(2\\%\\) of the training point clouds are acquired from the rest unlabeled dataset via active learning strategies for annotation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.669, + 0.483, + 0.775 + ], + "angle": 0, + "content": "Baselines. We evaluate nine AL baselines: random sampling (RAND), three 2D generic AL strategies (McDROPOUT, CORESET, BADGE) representing uncertainty-based, diversity-based, and hybrid methods, three 2D AL detection strategies (CDAL, DIVPROTO, PPAL) adapted to 3D detection tasks by us, and two state-of-the-art 3D outdoor AL detection strategies (CRB and KECOR)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.787, + 0.294, + 0.803 + ], + "angle": 0, + "content": "4.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.903 + ], + "angle": 0, + "content": "SUN RGB-D. In Fig. 3, we compare our proposed method with different baselines through an increasing percentage of selected point clouds [25, 59]. As demonstrated by the results, our proposed strategy consistently outperforms all competing baselines by significant margins for different percentages of selected point clouds. Notably, with only" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.907, + 0.521 + ], + "angle": 0, + "content": "\\(10\\%\\) annotated data, our approach achieves over \\(85\\%\\) of the fully-annotated performance (red dashed line in Fig. 3). Specifically, as detailed in Tab. 2, it surpasses the state-of-the-art 2D active learning strategy, PPAI, by \\(3.56\\%\\) in \\(\\mathrm{mAP}@0.25\\) and \\(1.91\\%\\) in \\(\\mathrm{mAP}@0.50\\), with a \\(10\\%\\) annotation budget. To elaborate further, it is observed that AL methods for 2D object detection generally outperform those for 3D outdoor counterpart. We hypothesize that current 3D outdoor approaches overlook the intra-class variances across different categories, leading to redundancy in selecting active samples. This redundancy may hinder the capability of the learned detection model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.909, + 0.78 + ], + "angle": 0, + "content": "ScanNetV2. To study the effectiveness of our proposed approach on point clouds with complicated room layouts, we conduct experiments using the ScanNetV2 dataset, a more comprehensive benchmark for indoor environments. The results, illustrated in Fig. 3 and Tab. 2, demonstrate the performance comparison across varying annotation budgets. Our method surpasses all existing AL approaches by a huge margin, achieving approximately \\(85\\%\\) of the fully-annotated performance with \\(10\\%\\) annotation budget. In Fig. 4, we present a qualitative comparison with random sampling. Random sampling captures only partial object characteristics (e.g., 'bookshelf') due to limited training data. In contrast, our approach successfully detects most of the target objects in challenging scenarios by effectively selecting informative unlabeled samples based on uncertainty and diversity criteria. Additional qualitative results can be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.667, + 0.804 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We ablate some critical designs for an in-depth exploration. Design Choices. We start by exploring the effectiveness of different designs in our proposed method, as shown in Tab. 3. When equipped with the IoU score in Eq. 2 and the uncertainty score for undetections \\(\\mathbf{U}_{\\mathrm{undet}}\\), the mAP@0.25 increases by \\(1.1\\%\\) and \\(0.7\\%\\) on SUN RGB-D respectively," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20335" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.167, + 0.093, + 0.281, + 0.107 + ], + "angle": 0, + "content": "(a) Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.109, + 0.36, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.431, + 0.093, + 0.558, + 0.107 + ], + "angle": 0, + "content": "(b) Random (10%)" + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.109, + 0.627, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.093, + 0.817, + 0.107 + ], + "angle": 0, + "content": "(c) Ours (10%)" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.109, + 0.894, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.207, + 0.865, + 0.222 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison between random sampling and our proposed active learning method on ScanNetV2 val set." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.233, + 0.483, + 0.349 + ], + "angle": 0, + "content": "
UncertaintyDiversitySUN RGB-DScanNetV2
mAP@.25mAP@.50mAP@.25mAP@.50
XX42.725.155.139.7
X✓54.533.958.742.9
w/ IoU score✓55.634.760.843.8
w/ Uundet✓55.234.360.143.3
✓X54.333.558.442.8
✓w/ K-means55.334.160.443.5
✓w/ Greedy-O54.533.358.743.0
✓✓56.135.461.845.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.351, + 0.483, + 0.378 + ], + "angle": 0, + "content": "Table 3. Ablation studies of different designs in our proposed method on SUN RGB-D and ScanNetV2 datasets." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.381, + 0.287, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.381, + 0.48, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.502, + 0.483, + 0.529 + ], + "angle": 0, + "content": "Figure 5. Effects of hyper-parameters in our proposed method on SUN RGB-D dataset." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.539, + 0.483, + 0.749 + ], + "angle": 0, + "content": "validating its effectiveness as an uncertainty-based strategy. Additionally, we ablate the diversity branch through two alternatives: (1) Use K-means [52] to select a fixed number of prototypes to replace CAP Bank, denoted as 'K-means'. (2) Apply greedy search to the original optimization problem as defined in Eq. 8, denoted as 'Greedy-O'. It demonstrates that the mAP@0.50 of 'K-means' decreases by \\(2.4\\%\\) for ScanNetV2 dataset, which suggests that a fixed prototype number leads to a sub-optimal outcome. Furthermore, 'Greedy-O' solution only achieves minor performance improvements. A possible reason is that we restrict the searching space of the original optimization problem (Eq. 8) to several regularized sub-spaces (Eq. 10), eliminating a vast number of sub-optimal solutions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Hyper-parameters. In Fig. 5, we report the ablation study of two hyper-parameters in our proposed method, which are the scale \\( k \\) in Eq. 4 and the budget expanding ratio \\( \\delta \\) for determining the candidate pool [59]. Overall, the proposed method is relatively insensitive to the choices of hyperparameters. As depicted in Fig. 5 (a), the best performance is reached with a moderate scale value of \\( k = 3 \\), which effectively suppresses extreme values while preserving essential information from different sources of uncertainty. Additionally, we observe that excessively large values of \\( \\delta \\) will" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.233, + 0.906, + 0.297 + ], + "angle": 0, + "content": "
MeasuresSUN RGB-DScanNetV2
mAP@.25mAP@.50mAP@.25mAP@.50
Prob. Margin [37]55.835.061.545.5
MC-Dropout [9]56.035.261.645.6
Entropy [39]56.135.461.845.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.299, + 0.906, + 0.327 + ], + "angle": 0, + "content": "Table 4. Ablation studies of different uncertainty measures in our proposed method on SUN RGB-D and ScanNetV2 datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.336, + 0.907, + 0.473 + ], + "angle": 0, + "content": "significantly degrade model performance in the early stages, which is shown in Fig. 5 (b). We hypothesize that for big \\(\\delta\\), the candidate pool contains many samples the model is already confident about, thereby harming the quality of the final selected samples. Conversely, when \\(\\delta\\) is very small, the selection process is dominated by uncertainty, which restricts the impact of diversity branch. We find that setting \\(\\delta = 6\\) achieves a better tradeoff between uncertainty- and diversity-based selection, optimizing overall performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.906, + 0.579 + ], + "angle": 0, + "content": "Uncertainty Measures. We replace the Shannon entropy in Eq. 2 with two alternative uncertainty measures: MC-Dropout [9] and probability margin [37], and compare their performance in Tab. 4. The results show that all three measures yield similar performance, with entropy performing slightly better. Therefore, we select entropy as our uncertainty measure, as it is both more efficient and effective." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.593, + 0.634, + 0.608 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.619, + 0.907, + 0.772 + ], + "angle": 0, + "content": "This paper presents a novel active learning approach for 3D indoor object detection, improving annotation efficiency by combining uncertainty estimation with diversity-driven sampling. Specifically, our approach estimates epistemic uncertainty, accounting for both inaccurate detections and undetected objects, and selects diverse samples through an optimization method based on a Class-aware Adaptive Prototype bank. Our method significantly outperforms state-of-the-art baselines, achieving over \\(85\\%\\) of fully-supervised performance with just \\(10\\%\\) of the annotation budget." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.785, + 0.686, + 0.802 + ], + "angle": 0, + "content": "6. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.9 + ], + "angle": 0, + "content": "This research is supported by the Ministry of Education, Singapore, under its MOE Academic Research Fund Tier 2 (MOE-T2EP20124-0013), it is also supported by the Agency for Science, Technology and Research (A*STAR) under its MTC Programmatic Funds (Grant No. M23L7b0021)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20336" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Sharat Agarwal, Himanshu Arora, Saket Anand, and Chetan Arora. Contextual diversity for active learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16, pages 137-153. Springer, 2020. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.24 + ], + "angle": 0, + "content": "[2] Kelsey Allen, Evan Shelhamer, Hanul Shin, and Joshua Tenenbaum. Infinite mixture prototypes for few-shot learning. In International conference on machine learning, pages 232-241. PMLR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.484, + 0.31 + ], + "angle": 0, + "content": "[3] Eduardo Arnold, Omar Y Al-Jarrah, Mehrdad Dianati, Saber Fallah, David Oxtoby, and Alex Mouzakitis. A survey on 3d object detection methods for autonomous driving applications. IEEE Transactions on Intelligent Transportation Systems, 20(10):3782-3795, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.311, + 0.483, + 0.367 + ], + "angle": 0, + "content": "[4] Jordan T Ash, Chicheng Zhang, Akshay Krishnamurthy, John Langford, and Alekh Agarwal. Deep batch active learning by diverse, uncertain gradient lower bounds. arXiv preprint arXiv:1906.03671, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.483, + 0.436 + ], + "angle": 0, + "content": "[5] Gui Citovsky, Giulia DeSalvo, Claudio Gentile, Lazaros Karydas, Anand Rajagopalan, Afshin Rostamizadeh, and Sanjiv Kumar. Batch active learning at scale. Advances in Neural Information Processing Systems, 34:11933-11944, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.483, + 0.507 + ], + "angle": 0, + "content": "[6] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.508, + 0.483, + 0.564 + ], + "angle": 0, + "content": "[7] Ehsan Elhamifar, Guillermo Sapiro, Allen Yang, and S Shankar Sasrty. A convex optimization framework for active learning. In Proceedings of the IEEE International Conference on Computer Vision, pages 209-216, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.483, + 0.62 + ], + "angle": 0, + "content": "[8] Di Feng, Xiao Wei, Lars Rosenbaum, Atsuto Maki, and Klaus Dietmayer. Deep active learning for efficient training of a lidar 3d object detector. In 2019 IEEE Intelligent Vehicles Symposium (IV), pages 667-674. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.483, + 0.676 + ], + "angle": 0, + "content": "[9] Yarin Gal, Riashat Islam, and Zoubin Ghahramani. Deep Bayesian active learning with image data. In Proceedings of the 34th International Conference on Machine Learning, pages 1183-1192. PMLR, 2017. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.677, + 0.483, + 0.732 + ], + "angle": 0, + "content": "[10] Mike Gashler and Tony Martinez. Temporal nonlinear dimensionality reduction. In The 2011 International Joint Conference on Neural Networks, pages 1959-1966. IEEE, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[11] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[12] Ahmed Ghita, Björk Antoniussen, Walter Zimmer, Ross Greer, Christian Creß, Andreas Møgelmose, Mohan M Trivedi, and Alois C Knoll. Activeanno3d—an active learning framework for multi-modal 3d object detection. arXiv preprint arXiv:2402.03235, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] JunYoung Gwak, Christopher Choy, and Silvio Savarese. Generative sparse detection networks for 3d single-shot object detection. In Computer Vision-ECCV 2020: 16th Euro-" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "pean Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 297-313. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.189 + ], + "angle": 0, + "content": "[14] Yucheng Han, Na Zhao, Weiling Chen, Keng Teck Ma, and Hanwang Zhang. Dual-perspective knowledge enrichment for semi-supervised 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2049-2057, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.906, + 0.26 + ], + "angle": 0, + "content": "[15] Pengkun Jiao, Na Zhao, Jingjing Chen, and Yu-Gang Jiang. Unlocking textual and visual wisdom: Open-vocabulary 3d object detection enhanced by comprehensive guidance from text and image. In European Conference on Computer Vision, pages 376-392. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.906, + 0.315 + ], + "angle": 0, + "content": "[16] Ajay J Joshi, Fatih Porikli, and Nikolaos Papanikolopoulos. Multi-class active learning for image classification. In 2009 IEEE conference on computer vision and pattern recognition, pages 2372-2379. IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.906, + 0.372 + ], + "angle": 0, + "content": "[17] Yoon-Yeong Kim, Kyungwoo Song, JoonHo Jang, and Il-Chul Moon. Lada: Look-ahead data acquisition via augmentation for deep active learning. Advances in Neural Information Processing Systems, 34:22919-22930, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.906, + 0.439 + ], + "angle": 0, + "content": "[18] Seong Min Kye, Kwanghee Choi, Hyeongmin Byun, and Buru Chang. Tidal: Learning training dynamics for active learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22335-22345, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.441, + 0.906, + 0.495 + ], + "angle": 0, + "content": "[19] Jean Lahoud and Bernard Ghanem. 2d-driven 3d object detection in rgb-d images. In Proceedings of the IEEE international conference on computer vision, pages 4622-4630, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.906, + 0.567 + ], + "angle": 0, + "content": "[20] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.906, + 0.636 + ], + "angle": 0, + "content": "[21] Jinpeng Lin, Zhihao Liang, Shengheng Deng, Lile Cai, Tao Jiang, Tianrui Li, Kui Jia, and Xun Xu. Exploring diversity-based active learning for 3d object detection in autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.906, + 0.692 + ], + "angle": 0, + "content": "[22] Zhuoming Liu, Hao Ding, Huaping Zhong, Weijia Li, Jifeng Dai, and Conghui He. Influence selection for active learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9274-9283, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.906, + 0.72 + ], + "angle": 0, + "content": "[23] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.906, + 0.789 + ], + "angle": 0, + "content": "[24] Yadan Luo, Zhuoxiao Chen, Zhen Fang, Zheng Zhang, Mahsa Baktashmotlagh, and Zi Huang. Kecor: Kernel coding rate maximization for active 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18279-18290, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.906, + 0.845 + ], + "angle": 0, + "content": "[25] Yadan Luo, Zhuoxiao Chen, Zijian Wang, Xin Yu, Zi Huang, and Mahsa Baktashmotlagh. Exploring active 3d object detection from a generalization perspective. arXiv preprint arXiv:2301.09249, 2023. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[26] Yi Ma, Harm Derksen, Wei Hong, and John Wright. Segmentation of multivariate mixed data via lossy data coding and compression. IEEE transactions on pattern analysis and machine intelligence, 29(9):1546-1562, 2007. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20337" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[27] Roberta Maisano, Valeria Tomaselli, Alessandro Capra, Francesco Longo, and Antonio Puliafito. Reducing complexity of 3d indoor object detection. In 2018 IEEE 4th International Forum on Research and Technology for Society and Industry (RTSI), pages 1-6. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.483, + 0.22 + ], + "angle": 0, + "content": "[28] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In 2015 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 922-928. IEEE, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.275 + ], + "angle": 0, + "content": "[29] Michael Meyer and Georg Kuschk. Automotive radar dataset for deep learning based 3d object detection. In 2019 16th European radar conference (EuRAD), pages 129-132. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.483, + 0.359 + ], + "angle": 0, + "content": "[30] Peng Mi, Jianghang Lin, Yiyi Zhou, Yunhang Shen, Gen Luo, Xiaoshuai Sun, Liujuan Cao, Rongrong Fu, Qiang Xu, and Rongrong Ji. Active teacher for semi-supervised object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14482-14491, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.483, + 0.417 + ], + "angle": 0, + "content": "[31] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2906-2917, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.483, + 0.473 + ], + "angle": 0, + "content": "[32] Aimee Moses, Srikanth Jakkampudi, Cheryl Danner, and Derek Biega. Localization-based active learning (local) for object detection in 3d point clouds. In Geospatial Informatics XII, pages 44-58. SPIE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.483, + 0.545 + ], + "angle": 0, + "content": "[33] Yuzuru Nakamura, Yasunori Ishii, and Takayoshi Yamashita. Active domain adaptation with false negative prediction for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28782-28792, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.483, + 0.6 + ], + "angle": 0, + "content": "[34] Hieu T Nguyen and Arnold Smeulders. Active learning using pre-clustering. In Proceedings of the twenty-first international conference on Machine learning, page 79, 2004. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.603, + 0.483, + 0.659 + ], + "angle": 0, + "content": "[35] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb-d data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 918-927, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.483, + 0.729 + ], + "angle": 0, + "content": "[36] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[37] Dan Roth and Kevin Small. Margin-based active learning for structured output spaces. In Machine Learning: ECML 2006: 17th European Conference on Machine Learning Berlin, Germany, September 18-22, 2006 Proceedings 17, pages 413-424. Springer, 2006. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.483, + 0.855 + ], + "angle": 0, + "content": "[38] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. In International Conference on Learning Representations, 2018. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[39] Claude Elwood Shannon. A mathematical theory of communication. The Bell system technical journal, 27(3):379-423, 1948. 2, 4, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[40] Xiaoke Shen and Ioannis Stamos. Frustum voxelnet for 3d object detection from rgb-d or depth images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1698-1706, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[41] Yichao Shen, Zigang Geng, Yuhui Yuan, Yutong Lin, Ze Liu, Chunyu Wang, Han Hu, Nanning Zheng, and Baining Guo. V-detr: Detr with vertex relative position encoding for 3d object detection. arXiv preprint arXiv:2308.04409, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[42] Hualian Sheng, Sijia Cai, Na Zhao, Bing Deng, Jianqiang Huang, Xian-Sheng Hua, Min-Jian Zhao, and Gim Hee Lee. Rethinking iou-based optimization for single-stage 3d object detection. In European Conference on Computer Vision, pages 544-561. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[43] Hualian Sheng, Sijia Cai, Na Zhao, Bing Deng, Min-Jian Zhao, and Gim Hee Lee. Pdr: Progressive depth regularization for monocular 3d object detection. IEEE Transactions on Circuits and Systems for Video Technology, 33(12):7591-7603, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.343, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[44] Hualian Sheng, Sijia Cai, Na Zhao, Bing Deng, Qiao Liang, Min-Jian Zhao, and Jieping Ye. Ct3d++: Improving 3d object detection with keypoint-induced channel-wise transformer. International Journal on Computer Vision, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[45] Samarth Sinha, Sayna Ebrahimi, and Trevor Darrell. Variational adversarial active learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5972-5981, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[46] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 567-576, 2015. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.565 + ], + "angle": 0, + "content": "[47] OpenPCDet Development Team. Openpcdet: An opensource toolbox for 3d object detection from point clouds. https://github.com/open-mmlab/OpenPCDet, 2020.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[48] Haiyang Wang, Lihe Ding, Shaocong Dong, Shaoshuai Shi, Aoxue Li, Jianan Li, Zhenguo Li, and Liwei Wang. Cagroup3d: Class-aware grouping for 3d object detection on point clouds. Advances in Neural Information Processing Systems, 35:29975-29988, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.636, + 0.905, + 0.678 + ], + "angle": 0, + "content": "[49] Jiangyi Wang, Zhongyao Cheng, Na Zhao, Jun Cheng, and Xulei Yang. On-the-fly point feature representation for point clouds analysis. arXiv preprint arXiv:2407.21335, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[50] Keze Wang, Dongyu Zhang, Ya Li, Ruimao Zhang, and Liang Lin. Cost-effective active learning for deep image classification. IEEE Transactions on Circuits and Systems for Video Technology, 27(12):2591-2600, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[51] Jamie Watson, Mohamed Sayed, Zawar Qureshi, Gabriel J. Brostow, Sara Vicente, Oisin Mac Aodha, and Michael Firman. Virtual occlusions through implicit depth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9053-9064, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[52] Junjie Wu. Advances in \\( K \\)-means clustering: a data mining thinking. Springer Science & Business Media, 2012. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[53] Jiaxi Wu, Jiaxin Chen, and Di Huang. Entropy-based active learning for object detection with progressive diversity constraint. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9397-9406, 2022. 3, 4, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20338" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.176 + ], + "angle": 0, + "content": "[54] Tsung-Han Wu, Yueh-Cheng Liu, Yu-Kai Huang, Hsin-Ying Lee, Hung-Ting Su, Ping-Chia Huang, and Winston H Hsu. Redal: Region-based and diversity-aware active learning for point cloud semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15510-15519, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.484, + 0.245 + ], + "angle": 0, + "content": "[55] Qian Xie, Yu-Kun Lai, Jing Wu, Zhoutao Wang, Yiming Zhang, Kai Xu, and Jun Wang. Mlcvnet: Multi-level context votenet for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10447-10456, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.484, + 0.314 + ], + "angle": 0, + "content": "[56] Qian Xie, Yu-Kun Lai, Jing Wu, Zhoutao Wang, Dening Lu, Mingqiang Wei, and Jun Wang. Venet: Voting enhancement network for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3712-3721, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.372 + ], + "angle": 0, + "content": "[57] Yating Xu, Conghui Hu, Na Zhao, and Gim Hee Lee. Generalized few-shot point cloud segmentation via geometric words. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21506-21515, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.467 + ], + "angle": 0, + "content": "[58] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17830-17839, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[59] Chenhongyi Yang, Lichao Huang, and Elliot J Crowley. Plug and play active learning for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17784-17793, 2024. 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[60] Yi Yang, Zhigang Ma, Feiping Nie, Xiaojun Chang, and Alexander G Hauptmann. Multi-class active learning by uncertainty sampling with diversity maximization. International Journal of Computer Vision, 113:113-127, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.58, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[61] Tianning Yuan, Fang Wan, Mengying Fu, Jianzhuang Liu, Songcen Xu, Xiangyang Ji, and Qixiang Ye. Multiple instance active learning for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5330-5339, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[62] Bohang Zhang, Jikai Jin, Cong Fang, and Liwei Wang. Improved analysis of clipping algorithms for non-convex optimization. Advances in Neural Information Processing Systems, 33:15511-15521, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[63] Na Zhao and Gim Hee Lee. Static-dynamic co-teaching for class-incremental 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3436-3445, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[64] Na Zhao, Tat-Seng Chua, and Gim Hee Lee. Sess: Self-ensembling semi-supervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11079–11087, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.873 + ], + "angle": 0, + "content": "[65] Na Zhao, Tat-Seng Chua, and Gim Hee Lee. Few-shot 3d point cloud semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8873-8882, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[66] Na Zhao, Peisheng Qian, Fang Wu, Xun Xu, Xulei Yang, and Gim Hee Lee. Sdcot++: Improved static-dynamic co" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "teaching for class-incremental 3d object detection. IEEE Transactions on Image Processing, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[67] Qiang Zhou and Chaohui Yu. Point rcnn: An angle-free framework for rotated object detection. Remote Sensing, 14 (11):2605, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20339" + } + ] +] \ No newline at end of file diff --git a/2025/Uncertainty Meets Diversity_ A Comprehensive Active Learning Framework for Indoor 3D Object Detection/d2e21437-1e7e-4dc1-839a-f40a2bf9489e_origin.pdf b/2025/Uncertainty Meets Diversity_ A Comprehensive Active Learning Framework for Indoor 3D Object Detection/d2e21437-1e7e-4dc1-839a-f40a2bf9489e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e4c251177426a10324a3bf075ff96b2a2b394784 --- /dev/null +++ b/2025/Uncertainty Meets Diversity_ A Comprehensive Active Learning Framework for Indoor 3D Object Detection/d2e21437-1e7e-4dc1-839a-f40a2bf9489e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2216bac3ce3c636505165af10ec1c48bab45c8e6897c87dfd12fc85fb1cffa8a +size 2054224 diff --git a/2025/Uncertainty Meets Diversity_ A Comprehensive Active Learning Framework for Indoor 3D Object Detection/full.md b/2025/Uncertainty Meets Diversity_ A Comprehensive Active Learning Framework for Indoor 3D Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..63bce1fd8d07d0ee1fb37a4d963b27dff817d687 --- /dev/null +++ b/2025/Uncertainty Meets Diversity_ A Comprehensive Active Learning Framework for Indoor 3D Object Detection/full.md @@ -0,0 +1,296 @@ +# Uncertainty Meets Diversity: A Comprehensive Active Learning Framework for Indoor 3D Object Detection + +Jiangyi Wang, Na Zhao* +Singapore University of Technology and Design (SUTD) +wangjiangyi0519@gmail.com, na.zhao@sutd.edu.sg + +# Abstract + +Active learning has emerged as a promising approach to reduce the substantial annotation burden in 3D object detection tasks, spurring several initiatives in outdoor environments. However, its application in indoor environments remains unexplored. Compared to outdoor 3D datasets, indoor datasets face significant challenges, including fewer training samples per class, a greater number of classes, more severe class imbalance, and more diverse scene types and intra-class variances. This paper presents the first study on active learning for indoor 3D object detection, where we propose a novel framework tailored for this task. Our method incorporates two key criteria - uncertainty and diversity - to actively select the most ambiguous and informative unlabeled samples for annotation. The uncertainty criterion accounts for both inaccurate detections and undetected objects, ensuring that the most ambiguous samples are prioritized. Meanwhile, the diversity criterion is formulated as a joint optimization problem that maximizes the diversity of both object class distributions and scene types, using a new Class-aware Adaptive Prototype (CAP) bank. The CAP bank dynamically allocates representative prototypes to each class, helping to capture varying intra-class diversity across different categories. We evaluate our method on SUN RGB-D and ScanNetV2, where it outperforms baselines by a significant margin, achieving over $85\%$ of fully-supervised performance with just $10\%$ of the annotation budget. Our code is available at: https://github.com/JoeWang-0519/CVPR25_UMD. + +# 1. Introduction + +3D object detection, which aims to localize and identify objects within a 3D scene, is crucial for various applications, including robotics [10], autonomous driving [20, 58], and augmented/virtual reality [51]. While deep learning techniques have achieved significant success in 3D object detection [31, 41, 48], most existing approaches are data- + +![](images/218ac544ffc061776612f813e9b811f331f308366de023c2783b1287a760b51b.jpg) +Figure 1. Challenges in active learning for indoor 3D object detection, including both uncertainty and diversity aspects. We show two indoor scenes on ScanNetV2, with red boxes for ground truths and cyan for predictions (10% data using CAGroup3D detector), displaying only 'chair' and 'door' categories for clarity. For uncertainty, the presence of undetected objects and inaccurate detections undermines the quality of uncertainty estimation. For diversity, high scene-type diversity and varying intra-class variances in indoor environments complicate diverse sample selection. Our work presents a hybrid approach to address these challenges. + +hungry [41], requiring large amounts of annotated data for training. However, annotating objects in a 3D scene is a labor-intensive process. For example, annotating a single scene in the SUN RGB-D dataset takes over 1.9 minutes per object, with an average of approximately 14 objects per scene [46]. To mitigate the high costs of 3D annotation, active learning (AL) [30, 34, 38] offers a promising solution. AL seeks to maximize the utility of limited labeling resources by actively selecting the most informative samples for human annotation. Through this process, AL can achieve performance comparable to fully-supervised methods, but with significantly fewer labeled samples. + +Although active learning has proven effective in image-based tasks such as image classification and object detection, using uncertainty [5, 16, 37] or/and diversity [34, 38, 60] as selection criteria, its application to 3D object detection remains largely underexplored. To date, only a few studies [12, 24, 25] have applied AL techniques to outdoor 3D object detection. For instance, CRB [25] uses label conciseness to capture uncertainty, along with feature representativeness and geometric balance to enhance diversity for sample selection. KECOR [24] favors samples whose latent features have the highest coding length, thereby ensuring diversity. While these methods demonstrate effectiveness in outdoor scenarios, they are suboptimal for indoor environments, as illustrated in Tab. 2, due to the unique challenges posed by 3D object detection in indoor settings. + +On the one hand, a 3D indoor dataset [6] could be $6.2\mathbf{x}$ smaller in scale but contain $3.3\mathbf{x}$ more number of categories, with more severe class imbalance, compared to an outdoor dataset [11], as shown in [14]. This scarcity of training samples, coupled with pronounced class imbalance, impairs 3D detection accuracy in indoor scenarios, leading to more frequent occurrences of inaccurate detection and undetected objects, as illustrated in Fig. 1. In this figure, three detections are inaccurate, incorrectly localizing background objects, while three other objects remain undetected. These challenges underscore the need to account for both inaccurately detected proposals and undetected objects when selecting informative samples from unlabeled data. + +On the other hand, unlike outdoor scenes, which primarily feature road environments with a limited set of object classes, indoor environments exhibit greater scene-type diversity and intra-class variances, caused by diverse layouts and variations in object shapes within each class. For example, as shown in Fig. 1, five 'chair' instances can form three clusters in the embedding space due to the variations in shape and occlusion. This phenomenon in indoor environments complicate the selection of diverse samples. + +In this work, we present the first solution to apply active learning in indoor 3D object detection, utilizing both uncertainty and diversity as selection criteria. For uncertainty, we design a two-pronged epistemic uncertainty estimation strategy to address inaccurate detections and undetected objects, which first assesses uncertainty separately for each case, then integrates them into a unified criterion. Specifically, we introduce a localization-aware uncertainty score that combines the IoU score – measured between the original detected object and its perturbed version – with the Shannon entropy. This helps mitigate the negative effects of invalid proposals, such as background or partial detections, when estimating the classification uncertainty of detected proposals. To quantify the epistemic uncertainty arising from undetectability, we develop a lightweight network that estimates the number of undetected objects. For di + +versity, we frame the diverse sample selection problem as an optimization task that jointly maximizes scene-type diversity and intra-class variances, leveraging a novel Class-aware Adaptive Prototype (CAP) bank. This CAP bank is designed to capture varying intra-class variances by generating an adaptive number of representative prototypes per class. Observing that different scene types exhibit distinct histograms of object categories as shown in Fig. 1, each composed of various prototypes, we use the CAP bank to model both class-level and scene-level diversity. + +We evaluate the proposed method on SUN RGB-D [46] and ScanNet [6] datasets. Using CAGroup3D [48] as base detection model, our approach achieves $56.11\%$ , $61.77\%$ $\mathrm{mAP}@0.25$ on SUN RGB-D and ScanNet, respectively, reaching $87.24\%$ and $85.02\%$ of the fully-supervised performance with only $10\%$ of the annotation budget. + +# 2. Related Work + +Generic Active Learning. The majority of existing works have explored active learning (AL) strategies on image classification task, which can be categorized into uncertainty-based [5, 9, 16, 37, 50] and diversity-based [1, 7, 34, 38, 54, 60] methods based on the selection criteria used. Uncertainty-based approaches select ambiguous samples for annotation by maximizing uncertainty scores, which can be computed in various ways, such as entropy [9, 17, 39], minimum posterior probability [50], or multi-class margin [5, 16, 37]. Shannon entropy [39] is a commonly used metric in classification that quantifies uncertainty in the probability distribution, while multi-class margin [37] measures the uncertainty through the difference between the top two predicted probabilities. MC Dropout [9] employs Monte Carlo Dropout to approximate the posterior predicted probability and use Shannon entropy to assess the underlying uncertainty. Diversity-based methods [1, 38, 60] select a compact subset of samples that best represent the global data distribution. For example, Core-set [38] formulates active learning as a core-set selection problem and solves it using the $k$ -centroid algorithm. Recent works [4, 22, 45] combine both uncertainty and diversity through hybrid approaches. A representative approach, BADGE [4], balances uncertainty and diversity by applying the $k$ -means++ [52] algorithm in the gradient space, which considers both magnitude and direction of gradient vectors. + +3D Object Detection. We focus on indoor 3D object detection [15, 63, 64, 66] and exclude outdoor methods [3, 42-44], which differ significantly. Prior indoor 3D object detection methods can be categorized into two lines in terms of underlying representation, i.e., voxel-based methods [13, 28, 40], and point-based methods [36, 49, 55, 56, 67]. Voxel-based methods convert point cloud to voxel grids and use 3D convolution networks to address the irregularity of point cloud. In contrast, point-based methods extract point + +![](images/73f15c81cc263313548d559a4e6947e1c8a867cadcf8d24854e64352e73c1cac.jpg) +Figure 2. Overview of our proposed AL framework for indoor 3D object detection, exploiting both uncertainty and diversity. In the $r$ -th round, we first construct a candidate pool of size $\delta \cdot B_r$ using the Top- $K$ unified uncertainty score, which accounts for both inaccurate detections and undetections. Then, we jointly optimize the intra-class and scene-type diversity to select the $r$ -th selected dataset $\mathcal{D}_r$ . Lastly, along with the previous labeled point clouds $\mathcal{D}_L^{r-1}$ , we retrain the model until the total labeled dataset reaches the annotation budget $B$ . + +level features and use a voting mechanism [36, 56] to encourage points to vote towards their corresponding object centers, followed by grouping the vote points into clusters and then generating object proposals. In this work, we adopt the state-of-the-art method in indoor 3D object detection, CAGroup3D [48], which takes advantage of the two lines of work through a two-stage fully sparse convolutional 3D detection framework. It extracts voxel-wise features from raw point clouds and leverages 'voting and grouping' mechanism to generate object proposals. + +Active Learning for Object Detection. Despite extensive studies on AL in image classification, applying it to image-based object detection is challenging due to the task's complexity, which involves both localization and classification of multiple objects. Several works [1, 18, 30, 53, 59, 61] have investigated active object detection by adapting generic AL techniques to this task. For example, CDAL [1] designs contextual diversity to capture spatial and semantic context diversity within a dataset, while DivProto [53] and PPAL [59] incorporate both diversity and uncertainty. While these methods yield promising results in 2D object detection, they cannot be directly applied to 3D object detection due to the significant modality gap. Recently, driven by the need to reduce annotation burden in 3D object detection for autonomous driving, a few studies [8, 12, 21, 24, 25, 27, 29, 32] have explored active learning for 3D object detection in outdoor environments. Notably, CRB [25] selects active samples with concise labels, representative features, and balanced density progressively. KECOR [24] leverages kernel coding rate [26] to measure the richness of latent features. However, these outdoor methods are suboptimal for indoor environments, where 3D object detection faces unique challenges. To ad + +dress these, this paper presents the first solution for applying active learning to indoor 3D object detection, overcoming these challenges by incorporating designs to estimate uncertainty in both inaccurate detections and undetections, while simultaneously optimizing diversity with respect to scene-type and intra-class variances. + +# 3. Methodology + +Given a point cloud $\mathcal{P}_i = \{(x,y,z)\}$ , a 3D object detection model aims to predict the locations and class labels for a set of objects $\mathfrak{B}_i = \{(\mathbf{b}_k,y_k)\}$ , where $\mathbf{b}_k\in \mathbb{R}^7$ represents the parameter of the $k$ -th bounding box, including the center position, box size and yaw angle, and $y_{k}$ indicates its predicted class label. In the active learning setting, our objective is to train the 3D detection model by constructing a small labeled dataset $\mathcal{D}_L$ , which starts with a small portion of random samples $\mathcal{D}_L^0$ and is progressively expanded over $R$ rounds from a large pool $\mathcal{D}$ . In each active learning round, $r\in \{1,\dots,R\}$ , $B_{r}$ samples are selected from the unlabeled dataset $\mathcal{D}_U = \mathcal{D} / \mathcal{D}_L^{r - 1}$ by an active learning policy, and labeled by an oracle, which constitute the $r$ -th selected dataset $\mathcal{D}_r$ . Then, $\mathcal{D}_L$ will be updated as $\mathcal{D}_L^r\gets \mathcal{D}_L^{r - 1}\cup \mathcal{D}_r$ until the size of $\mathcal{D}_L$ reaches the final budget $B$ , i.e., $|\mathcal{D}_L^0 | + \sum_{r = 1}^{R}B_r = B$ where $B\ll |\mathcal{D}|$ . + +To tackle active 3D object detection in indoor scenarios, we propose a novel framework, illustrated in Fig. 2, that leverages both uncertainty and diversity as selection criteria. Following previous work [21, 53, 59] that considers both uncertainty and diversity, we adopt a budget expansion ratio $\delta \in \mathbb{Z}^{+}$ to first select a candidate pool $\mathcal{D}_r^{\mathrm{cand}}$ of size $K = \delta \cdot B_r$ based on the uncertainty criterion for $r$ -th active learning round. Then, from this pool, the $\mathcal{D}_r$ of size + +
DatasetmAP@.25mAP@.50mAR@.25mAR@.50
SUN RGB-D [46]28.9112.4867.1226.09
ScanNetV2 [6]35.7520.4076.6542.48
+ +Table 1. Motivation for designing epistemic uncertainty estimation. We report the mean average precision (mAP) and mean average recall (mAR) at two IoU thresholds (0.25 and 0.50) for the initial detection model, which is trained using $2\%$ labeled data. + +$B_{r}$ is chosen based on the diversity criterion. For the uncertainty criterion, we propose a two-pronged epistemic uncertainty estimation strategy that generates a unified score, accounting for uncertainty in both detected and undetected objects. This is achieved through a localization-aware uncertainty score and an estimate of the undetected object count (Sec. 3.1). For the diversity criterion, we frame the diverse sample selection as an optimization problem and propose an efficient solution that jointly optimizes intra-class variances and scene-type diversity (Sec. 3.2). + +# 3.1. Epistemic Uncertainty Estimation + +In Tab. 1, we evaluate the detection performance of a 3D object detection model trained on the initial random samples $D_{L}^{0}$ (2% labeled data). As shown, the model achieves 20.40% mAP@.50 and 42.48% mAR@.50 on ScanNet [6], indicating that approximately 80% of the predictions are inaccurate and 60% of ground-truth objects remain undetected. The results are even worse on the more challenging SUN RGB-D [46] dataset. These findings highlight the need to address the two types of epistemic uncertainty in 3D object detection: inaccurate detections and undetected objects. Given that the detection task involves both localization and classification, we estimate the epistemic uncertainty of inaccurate detections by considering both aspects. Specifically, we introduce a localization-aware uncertainty score that exploits an IoU score for localization reliability to combine with Shannon entropy. To handle undetections, we design a lightweight network to estimate the number of undetected objects, which serves as the uncertainty measure for undetections. These two uncertainty scores are then integrated into a unified uncertainty score, which is used for Top-K selection to generate the candidate pool $\mathcal{D}_r^{\mathrm{cand}}$ . + +Localization-aware Uncertainty Score. To estimate the uncertainty of detected objects from a point cloud $\mathcal{P}_i$ , we first compute the classification uncertainty using the Shannon entropy [39] of the predicted class probabilities for all detected object proposals. For the $j$ -th proposal, with predicted class probabilities over $C$ classes denoted as $\mathbf{p}_j = \{p_{j,c}\}_{c=1}^C \in \mathbb{R}^C$ , the Shannon entropy is calculated as: + +$$ +E (\mathbf {p} _ {j}) = - \sum_ {c = 1} ^ {C} p _ {j, c} \log (p _ {j, c}). \tag {1} +$$ + +However, not all detected proposals are valid, as some may correspond to background or partial objects. To mitigate + +the impact of these invalid proposals, we incorporate an IoU score that measures the similarity between the original detection and its perturbed version, obtained by applying weak augmentations (e.g., random rotation and scaling) to the input point cloud. The intuition is that if the model is confident about an object, small perturbations should not significantly affect its predicted position, while background or partial proposals will likely produce inconsistent predictions. Consequently, the localization-aware uncertainty score for detected objects in $\mathcal{P}_i$ is defined as: + +$$ +\mathbf {U} _ {\det } \left(\mathcal {P} _ {i}\right) := \frac {\sum_ {j = 1} ^ {\left| \mathfrak {B} _ {i} \right|} \operatorname {I o U} _ {j} \cdot E \left(\mathbf {p} _ {j}\right)}{\sum_ {j = 1} ^ {\left| \mathfrak {B} _ {i} \right|} \operatorname {I o U} _ {j}}. \tag {2} +$$ + +Our IoU scores serve as soft weights, adaptively reducing the contribution of invalid objects in the epistemic uncertainty estimation. + +Undetected Number Prediction Network. When estimating uncertainty for active object detection, previous works [25, 53, 59] typically focus only on the uncertainty within the detected objects. However, the uncertainty of undetected objects is equally important, especially in the context of active learning for indoor 3D object detection, where model performance is often hindered by the scarcity of training samples and severe class imbalance. As shown by the mean average recall (mAR) scores in Tab. 1, more than half of the ground truth objects remain undetected in both 3D indoor datasets. This observation highlights the need to account for the uncertainty of undetected objects. To address this, we design a lightweight network to estimate the number of undetected objects. + +As depicted in Fig. 2, we first extract the feature map $\mathbf{f}_i$ from the point cloud $\mathcal{P}_i$ using the feature extractor. We then apply global average pooling (GAP) on $\mathbf{f}_i$ to obtain the global feature $\mathbf{g}_i$ . Simultaneously, we create a masked feature map by masking out the regions corresponding to the detected bounding boxes produced by RoI refinement. This masked feature map is passed through GAP to yield the undetected feature $\mathbf{u}_i$ . We concatenate these two features, $\mathbf{g}_i$ and $\mathbf{u}_i$ , and use the concatenation as input to our undetected object number prediction network $\Theta_{\mathrm{undet}}$ , a lightweight network consisting of three-layer MLPs. The output of the network, i.e. $\Theta_{\mathrm{undet}}([\mathbf{g}_i; \mathbf{u}_i])$ represents the uncertainty score for undetections, denoted as $\mathbf{U}_{\mathrm{undet}}$ . This output is compared with the true number of undetected objects, computed based on the detected objects $\mathfrak{B}_i$ and the ground-truth objects $\mathfrak{B}_i^*$ , to compute the loss for training $\Theta_{\mathrm{undet}}$ : + +$$ +\mathcal {L} _ {\text {u n d e t}} = \sum_ {i = 1} ^ {| \mathcal {D} _ {L} |} \left(\Theta_ {\text {u n d e t}} \left(\left[ \mathbf {g} _ {i}; \mathbf {u} _ {i} \right]\right) - \mathcal {N} \left(\mathfrak {B} _ {i}, \mathfrak {B} _ {i} ^ {*}\right)\right) ^ {2}, \tag {3} +$$ + +where $\mathcal{N}(\cdot, \cdot)$ is a function that computes the number of undetected objects, with details provided in the supplementary material. Additionally, to enhance the generalizability + +of the network $\Theta_{\mathrm{undet}}$ , we employ random masking of the predicted boxes during the training phase, encouraging the network to better capture the undetectability in point clouds. Unified Uncertainty Score. After obtaining the two uncertainty scores, $\mathbf{U}_{\mathrm{det}}$ and $\mathbf{U}_{\mathrm{undet}}$ , we aim to combine them into a unified uncertainty score. However, these two scores are on different scales, making their direct product infeasible. To address this issue, we introduce a normalization step, inspired by [33], using a hyperparameter $k$ . Specifically, we normalize each score as follows: + +$$ +\tilde {\mathbf {U}} _ {\star} (\mathcal {P} _ {i}) = \max (0, \frac {\mathbf {U} _ {\star} (\mathcal {P} _ {i}) - \mu (\mathbf {U} _ {\star})}{k \cdot \sigma (\mathbf {U} _ {\star})} + 0. 5), \tag {4} +$$ + +where $\star \in \{\mathrm{det},\mathrm{undet}\}$ , and $\mu (\cdot)$ and $\sigma (\cdot)$ denote the mean and standard deviation, respectively. Once the scores are normalized, we compute their product to yield the unified uncertainty score. + +# 3.2. Prototype-based Diversity Optimization + +As illustrated in Fig. 1, indoor environments exhibit significant intra-class variances and scene-type diversity, both of which motivate our approach to the diversity criterion. In 3D indoor scenes, the intra-class variances can vary greatly across object categories due to the differences in shape and occlusion. For example, the variance in 'chairs' is typically high due to the wide variety of shapes and designs, whereas 'doors' tend to have lower variance due to their more standardized geometry. This variability makes methods that model each class with a fixed number of prototypes [57, 65] insufficient to capture the intra-class distributions. To address this, we propose a novel Class-aware Adaptive Prototype (CAP) bank, which dynamically allocates a varying number of prototypes for each class, inferred from the predicted object proposals. Additionally, we observe that indoor scenes naturally form distinct clusters based on the distribution of object categories. For instance, bathrooms typically contain toilets and bathtubs, while bedrooms feature beds and dressers. Through this observation, we represent each scene as a histogram of the prototypes corresponding to the predicted object categories within the scene. As both intra-class variances and scene-type diversity can be modeled using the CAP bank, we frame diverse sample selection as an optimization problem that jointly maximizes both aspects, and propose an efficient solution. + +Class-aware Adaptive Prototype Bank. The design of our Class-aware Adaptive Prototype (CAP) bank is inspired by infinite mixture prototypes (IMP) [2], which was originally proposed for few-shot learning. IMP allows for inferring the number of clusters from the data, rather than relying on a pre-defined number of clusters. To build the CAP bank, we first iterate over the initial labeled set $D_{L}^{r - 1}$ to initialize one prototype per class. Specifically, we leverage the ground-truth annotations to extract features for all objects belong- + +ing to the $c$ -th class in $D_L^{r-1}$ , and compute their average as to form the initial prototype $\mu_{1,c}$ for class $c$ . Subsequently, we iterate over the unlabeled dataset $\mathcal{D}_U$ and update the prototypes for each class, resulting in our CAP bank, denoted as $\mathbf{M} = \{\{\pmb{\mu}_{k,c}\}_{k=1}^{M_c}\}_{c=1}^C$ , where $\pmb{\mu}_{k,c}$ represents the $k$ -th prototype for class $c$ , $M_c$ is the number of prototypes for class $c$ , which may vary between classes, and $C$ is the total number of classes. In each iteration (one batch), we use the detection model from the previous round to generate object proposals and extract their features and predicted class labels. For each predicted object, we compute the cosine similarity between its feature embedding $\mathbf{o}_j$ and each of the existing prototypes in its predicted class. This similarity score is formulated as: + +$$ +\operatorname {S i m} \left(\mathbf {o} _ {j}, \boldsymbol {\mu} _ {k, c}\right) = \frac {\mathbf {o} _ {j} ^ {T} \cdot \boldsymbol {\mu} _ {k , c}}{\left| \left| \mathbf {o} _ {j} \right| \right| _ {2} \left| \left| \boldsymbol {\mu} _ {k , c} \right| \right| _ {2}} \in [ - 1, 1 ]. \tag {5} +$$ + +After calculating the similarity scores, each object is either assigned to an existing prototype or used to form a new prototype, according to the following assignment rule: + +$$ +\mathcal {A} \left(\mathbf {o} _ {j}\right) = \left\{ \begin{array}{l} M _ {c} + 1, \text {i f} \max _ {k} \operatorname {S i m} \left(\mathbf {o} _ {j}, \boldsymbol {\mu} _ {k, c}\right) < \tau_ {\text {s i m}}, \\ \arg \max _ {k} \operatorname {S i m} \left(\mathbf {o} _ {j}, \boldsymbol {\mu} _ {k, c}\right), \quad \text {o t h e r w i s e}. \end{array} \right. \tag {6} +$$ + +Here, $\tau_{\mathrm{sim}}$ is similarity threshold, which is set to 0.3 by default. Finally, the prototypes - whether assigned or newly formed - are updated using the new object proposals in each batch. However, since many predicted bounding boxes might be invalid, especially in the early stages of training (e.g., as shown in Tab. 1), we seek to mitigate the influence of these invalid proposals when updating the prototypes. To this end, we incorporate the IoU score from Eq. 2 into the prototype update process, effectively weighing each object's contribution to the prototype updates as follows: + +$$ +\boldsymbol {\mu} _ {k, c} = \frac {\sum_ {j = 1} ^ {| \mathfrak {B} _ {i} |} \mathbb {1} \left\{\mathcal {A} \left(\mathbf {o} _ {j}\right) , k \right\} \cdot \operatorname {I o U} _ {j} \cdot \mathbf {o} _ {j}}{\sum_ {j = 1} ^ {| \mathfrak {B} _ {i} |} \mathbb {1} \left\{\mathcal {A} \left(\mathbf {o} _ {j}\right) , k \right\} \cdot \operatorname {I o U} _ {j}}, \tag {7} +$$ + +where the indicator function $\mathbb{1}(m,n)$ equals to 1 if $m = n$ , and 0 otherwise. The complete CAP bank creation procedure is outlined in Algo. 1 of the supplementary material. + +Prototype-based Diversity Optimization. With the generated CAP bank, we can characterize each class through a set of prototypes, and further represent objects as histograms of these prototypes. These histograms allow us to capture the intra-class variance for each class, as objects within the same class are reflected in the variation of their corresponding histograms. Furthermore, based on the observation that each scene can be characterized by the distribution of object categories, which distinguishes one scene type from another, we also model scenes as histograms of prototypes. Therefore, this histogram-based representation + +implicitly captures the scene type. Since both intra-class variance and scene-type diversity are encoded in the prototype histograms, we design a novel and efficient optimization method to quantify the diversity criterion by considering both these two aspects. Specifically, we formulate the optimization for diverse sample selection as follows: + +$$ +\begin{array}{l} \max _ {\mathcal {D} _ {r} \subset \mathcal {D} _ {r} ^ {\text {c a n d}}} E (\mathbf {h} (\mathcal {D} _ {r})) + \beta | | \mathbf {h} (\mathcal {D} _ {r}) | | _ {1}, \\ \begin{array}{l} \text {s . t .} \quad \mathbf {h} (\mathcal {D} _ {r}) = \sum_ {\mathcal {P} _ {i} \in \mathcal {D} _ {r}} \mathbf {h} (\mathcal {P} _ {i}) = \sum_ {\mathcal {P} _ {i} \in \mathcal {D} _ {r}} \sum_ {\mathbf {o} _ {j} \in \mathcal {P} _ {i}} \mathbf {h} (\mathbf {o} _ {j}), \end{array} \\ \mathbf {h} \left(\mathbf {o} _ {j}\right) = \sum_ {c = 1} ^ {C} \sum_ {k = 1} ^ {M _ {c}} [ \text {o n e} \_ \text {h o t} (\mathcal {A} \left(\mathbf {o} _ {j}\right)) ] ^ {c, k}, \tag {8} \\ \end{array} +$$ + +where $\mathbf{h}(\mathbf{o}_j)$ represents the assignment histogram of $j$ -th object w.r.t. the CAP bank $\mathbf{M}$ , while $\mathbf{h}(\mathcal{P}_i)$ and $\mathbf{h}(\mathcal{D}_r)$ represent the histograms of all objects in the individual point cloud $\mathcal{P}_i$ and the selected labeled samples $\mathcal{D}_r$ at the $r$ -th round, respectively. The two terms in this objective function are designed to select samples that are both diverse and holistic across all prototypes in the CAP bank. The first term captures the intra-class variance for all classes, ensuring that the selected samples represent the diversity within each class as well as across different classes. The second term promotes scene-level diversity by encouraging coverage across different object categories, which implicitly increases the diversity in scene types. Given that the first term (entropy) and the second term ( $l_1$ -norm) have different magnitudes, we introduce a hyperparameter $\beta$ to rescale the second term. In our experiments, we set $\beta = 0.01$ . + +However, directly solving the optimization in Eq. 8 is computationally expensive due to the combinatorial search over $|\mathcal{D}_r^{\mathrm{cand}}|$ samples. To efficiently narrow the search space while maintaining solution quality, we leverage one key property of the objective function over $S$ disjoint partitions $\{\mathcal{D}_{r,s}\}_{s = 1}^{S}$ of $\mathcal{D}_r$ , i.e., $\cup_{s}\mathcal{D}_{r,s} = \mathcal{D}_r$ : + +$$ +\begin{array}{l} E \left(\mathbf {h} \left(\mathcal {D} _ {r}\right)\right) + \beta | | \mathbf {h} \left(\mathcal {D} _ {r}\right) | | _ {1} \\ \stackrel {c} {=} \sum_ {s = 1} ^ {S} \left\{\frac {N _ {s}}{N _ {\text {t o t a l}}} E \left(\mathbf {h} \left(\mathcal {D} _ {r, s}\right)\right) + \beta \left\| \mathbf {h} \left(\mathcal {D} _ {r, s}\right) \right\| _ {1} \right\} + \mathcal {O} \left(\frac {N _ {\text {i n t e r}}}{N _ {\text {t o t a l}}}\right), \tag {9} \\ \end{array} +$$ + +where $\mathbf{h}(\mathcal{D}_{r,s})$ represents the histograms of all objects in $s$ -th partition $\mathcal{D}_{r,s}$ , $N_{s}$ and $N_{\mathrm{total}}$ are the number of objects in the $s$ -th partition $\mathcal{D}_{r,s}$ and selected labeled samples $\mathcal{D}_r$ , $N_{\mathrm{inter}}$ is the number of objects that share the same prototypes but belong to different partitions, and $\mathcal{O}(\cdot)$ is the Big O notation from mathematics. To align with the $S$ disjoint partitions of $\mathcal{D}_r$ , we use k-means++ [52] to group the candidate pool $\mathcal{D}_r^{\mathrm{cand}}$ into $S$ clusters, denoted as $\{\mathcal{D}_{r,s}^{\mathrm{cand}}\}_{s = 1}^{S}$ , each of which is approximately disjoint with each other w.r.t. prototype assignments. This guarantees that $\frac{N_{\mathrm{inter}}}{N_{\mathrm{total}}} \to 0$ , and then allows us to decompose the original optimization problem into $S$ sub-problems within each cluster $\mathcal{D}_{r,s}^{\mathrm{cand}}$ : + +$$ +\begin{array}{l} \max _ {\mathcal {D} _ {r, s} \subset \mathcal {D} _ {r, s} ^ {\text {c a n d}}} E (\mathbf {h} (\mathcal {D} _ {r, s})) + \beta \frac {N _ {\text {t o t a l}}}{N _ {s}} | | \mathbf {h} (\mathcal {D} _ {r, s}) | | _ {1}, \\ \begin{array}{l} \text {s . t .} \quad \mathbf {h} (\mathcal {D} _ {r, s}) = \sum_ {\mathcal {P} _ {i} \in \mathcal {D} _ {r, s}} \mathbf {h} (\mathcal {P} _ {i}) = \sum_ {\mathcal {P} _ {i} \in \mathcal {D} _ {r, s}} \sum_ {\mathbf {o} _ {j} \in \mathcal {P} _ {i}} \mathbf {h} (\mathbf {o} _ {j}), \end{array} \\ \mathbf {h} \left(\mathbf {o} _ {j}\right) = \sum_ {c = 1} ^ {C} \sum_ {k = 1} ^ {M _ {c}} [ \text {o n e} \text {h o t} (\mathcal {A} \left(\mathbf {o} _ {j}\right)) ] ^ {c, k}. \tag {10} \\ \end{array} +$$ + +We use greedy search to solve the $S$ sub-problems to obtain the diverse samples $\mathcal{D}_{r,s}$ for each cluster, and combine them together as the selected dataset $\mathcal{D}_r$ in the $r$ -th active learning round. We provide the detailed derivation of Eq. 9 in the supplementary material. + +# 4. Experiments + +# 4.1. Experimental Setups + +3D Detection Datasets. We evaluate our proposed method on two 3D indoor object detection benchmarks: SUN RGB-D [46] and ScanNet [6]. SUN RGB-D is an indoor dataset for 3D object detection, which contains 10,335 RGB-D images with 3D room layouts from four different sensors. Officially, they are split into 5,285 training samples and 5,050 testing samples. Following the standard evaluation protocol [19, 35, 36], we evaluate the performance on 10 most common categories. ScanNetV2 is a large 3D dataset with 1,513 RGB-D scans of 707 indoor scenes. These scans are split into 1,201 samples for training and 312 samples for testing, each of which is annotated with dense semantic segmentation masks. We follow [36] to derive the axis-aligned bounding boxes without considering the orientation, and adopt 18 object categories for evaluation. + +Implementation Details. We use the state-of-the-art CA-Group3D [48] as the detection model, and implement our code base using OpenPCDet [47] toolkit. Following the training strategies in [48], we set the voxel size as $0.02\mathrm{m}$ . For both datasets, the network is trained by AdamW optimizer [23] with an initial learning rate and weight decay rate of 0.001 and 0.0001 respectively. The gradnorm clip [62] is applied to stabilize the training process. For each active learning round, SUN RGB-D takes 48 epochs to train on the updated labeled dataset, and the learning rate decays on [32, 44] epochs with decay rate of 0.1. ScanNet requires 120 epochs for training and decays the learning rate on [80, 110] epochs by the same decay rate. All experiments are conducted using one 3090Ti NVIDIA GPU. We set the similarity threshold $\tau_{\mathrm{sim}}$ to 0.3, balance ratio $\beta$ to 0.01, and budget expanding ratio $\delta$ to 6 empirically. + +Evaluation. Following the evaluation protocol in [31, 36, 48], we report mean average precision (mAP) with different IoU thresholds, i.e., 0.25 and 0.50. Additionally, to investigate the effectiveness of different active learning strategies, we record the mAP with $2\%$ , $4\%$ , $6\%$ , $8\%$ , $10\%$ of the train + +
MethodsPresened atSUN RGB-DScanNetV2
mAP@0.25 (%)mAP@0.50 (%)mAP@0.25 (%)mAP@0.50 (%)
2D-GenerieRAND-42.7125.1455.0939.74
MC-DROPOUT [9]ICML'1747.4828.3645.7128.72
CORESET [38]ICLR'1846.9928.1255.3440.08
BADGE [4]ICLR'2049.1230.5457.1041.97
2D-ODCDAL [1]ECCV'2050.8630.7557.9942.33
DIVPROTO [53]CVPR'2252.3132.5459.0143.10
PPAL [59]CVPR'2452.5533.5159.4443.15
3D-ODCRB [25]ICLR'2350.8931.3058.7342.71
KECOR [24]ICCV'2351.0731.4158.5442.56
OURS-56.1135.4261.7745.93
ORACLE†-64.3147.4372.6558.81
+ +Table 2. mAP@0.25 (%) and mAP@0.50 (%) scores of the proposed method and AL baselines on the SUN RGB-D and ScanNetV2 benchmarks with $10\%$ queried point clouds. CAGroup3D is used as the detection model for all approaches. $\dagger$ : indicates the detection results using fully-annotated data, and serves as the upper bound for all active learning strategies. + +![](images/4b3db0e785a91ae609319f4207bc12e3e9a37f45af9668c402f6502417cddcbf.jpg) + +![](images/893b7c2e70ae723b239d06c00a749ac8fd40db81019eeefdaa34a6be217db612.jpg) +Figure 3. mAP@0.25 (\%) score of the proposed method and AL baselines on SUN RGB-D and ScanNetV2 benchmarks. + +ing set for both datasets, where the first $2\%$ samples are selected randomly. In each active learning round, $2\%$ of the training point clouds are acquired from the rest unlabeled dataset via active learning strategies for annotation. + +Baselines. We evaluate nine AL baselines: random sampling (RAND), three 2D generic AL strategies (McDROPOUT, CORESET, BADGE) representing uncertainty-based, diversity-based, and hybrid methods, three 2D AL detection strategies (CDAL, DIVPROTO, PPAL) adapted to 3D detection tasks by us, and two state-of-the-art 3D outdoor AL detection strategies (CRB and KECOR). + +# 4.2. Experimental Results + +SUN RGB-D. In Fig. 3, we compare our proposed method with different baselines through an increasing percentage of selected point clouds [25, 59]. As demonstrated by the results, our proposed strategy consistently outperforms all competing baselines by significant margins for different percentages of selected point clouds. Notably, with only + +$10\%$ annotated data, our approach achieves over $85\%$ of the fully-annotated performance (red dashed line in Fig. 3). Specifically, as detailed in Tab. 2, it surpasses the state-of-the-art 2D active learning strategy, PPAI, by $3.56\%$ in $\mathrm{mAP}@0.25$ and $1.91\%$ in $\mathrm{mAP}@0.50$ , with a $10\%$ annotation budget. To elaborate further, it is observed that AL methods for 2D object detection generally outperform those for 3D outdoor counterpart. We hypothesize that current 3D outdoor approaches overlook the intra-class variances across different categories, leading to redundancy in selecting active samples. This redundancy may hinder the capability of the learned detection model. + +ScanNetV2. To study the effectiveness of our proposed approach on point clouds with complicated room layouts, we conduct experiments using the ScanNetV2 dataset, a more comprehensive benchmark for indoor environments. The results, illustrated in Fig. 3 and Tab. 2, demonstrate the performance comparison across varying annotation budgets. Our method surpasses all existing AL approaches by a huge margin, achieving approximately $85\%$ of the fully-annotated performance with $10\%$ annotation budget. In Fig. 4, we present a qualitative comparison with random sampling. Random sampling captures only partial object characteristics (e.g., 'bookshelf') due to limited training data. In contrast, our approach successfully detects most of the target objects in challenging scenarios by effectively selecting informative unlabeled samples based on uncertainty and diversity criteria. Additional qualitative results can be found in the supplementary materials. + +# 4.3. Ablation Study + +We ablate some critical designs for an in-depth exploration. Design Choices. We start by exploring the effectiveness of different designs in our proposed method, as shown in Tab. 3. When equipped with the IoU score in Eq. 2 and the uncertainty score for undetections $\mathbf{U}_{\mathrm{undet}}$ , the mAP@0.25 increases by $1.1\%$ and $0.7\%$ on SUN RGB-D respectively, + +![](images/51eaf9d3af05c0cd3219f4eeeca08f10a5ff3a2503f573af0c3e0ca42712e519.jpg) +(a) Ground Truth + +![](images/1d46073ab58a5c741e33fc66e70304697c46b4beed11564d1f6bf39200fca216.jpg) +(b) Random (10%) +Figure 4. Qualitative comparison between random sampling and our proposed active learning method on ScanNetV2 val set. + +![](images/65b23165b211db0a2779d1c4e0c12e1d0ac6ad785f86eebff6930c77efe9d9d8.jpg) +(c) Ours (10%) + +
UncertaintyDiversitySUN RGB-DScanNetV2
mAP@.25mAP@.50mAP@.25mAP@.50
XX42.725.155.139.7
X✓54.533.958.742.9
w/ IoU score✓55.634.760.843.8
w/ Uundet✓55.234.360.143.3
✓X54.333.558.442.8
✓w/ K-means55.334.160.443.5
✓w/ Greedy-O54.533.358.743.0
✓✓56.135.461.845.9
+ +![](images/ad46ab98f854f555d1ea5412ba19e53d8a1cbfcf337b53e7a7baae95227a814d.jpg) +Figure 5. Effects of hyper-parameters in our proposed method on SUN RGB-D dataset. + +![](images/6f98d1a12c2a1327561d2808e3b1504c7d2c3cf47c160036b20fa5d1fe186b8b.jpg) + +validating its effectiveness as an uncertainty-based strategy. Additionally, we ablate the diversity branch through two alternatives: (1) Use K-means [52] to select a fixed number of prototypes to replace CAP Bank, denoted as 'K-means'. (2) Apply greedy search to the original optimization problem as defined in Eq. 8, denoted as 'Greedy-O'. It demonstrates that the mAP@0.50 of 'K-means' decreases by $2.4\%$ for ScanNetV2 dataset, which suggests that a fixed prototype number leads to a sub-optimal outcome. Furthermore, 'Greedy-O' solution only achieves minor performance improvements. A possible reason is that we restrict the searching space of the original optimization problem (Eq. 8) to several regularized sub-spaces (Eq. 10), eliminating a vast number of sub-optimal solutions. + +Hyper-parameters. In Fig. 5, we report the ablation study of two hyper-parameters in our proposed method, which are the scale $k$ in Eq. 4 and the budget expanding ratio $\delta$ for determining the candidate pool [59]. Overall, the proposed method is relatively insensitive to the choices of hyperparameters. As depicted in Fig. 5 (a), the best performance is reached with a moderate scale value of $k = 3$ , which effectively suppresses extreme values while preserving essential information from different sources of uncertainty. Additionally, we observe that excessively large values of $\delta$ will + +Table 3. Ablation studies of different designs in our proposed method on SUN RGB-D and ScanNetV2 datasets. + +
MeasuresSUN RGB-DScanNetV2
mAP@.25mAP@.50mAP@.25mAP@.50
Prob. Margin [37]55.835.061.545.5
MC-Dropout [9]56.035.261.645.6
Entropy [39]56.135.461.845.9
+ +Table 4. Ablation studies of different uncertainty measures in our proposed method on SUN RGB-D and ScanNetV2 datasets. + +significantly degrade model performance in the early stages, which is shown in Fig. 5 (b). We hypothesize that for big $\delta$ , the candidate pool contains many samples the model is already confident about, thereby harming the quality of the final selected samples. Conversely, when $\delta$ is very small, the selection process is dominated by uncertainty, which restricts the impact of diversity branch. We find that setting $\delta = 6$ achieves a better tradeoff between uncertainty- and diversity-based selection, optimizing overall performance. + +Uncertainty Measures. We replace the Shannon entropy in Eq. 2 with two alternative uncertainty measures: MC-Dropout [9] and probability margin [37], and compare their performance in Tab. 4. The results show that all three measures yield similar performance, with entropy performing slightly better. Therefore, we select entropy as our uncertainty measure, as it is both more efficient and effective. + +# 5. Conclusion + +This paper presents a novel active learning approach for 3D indoor object detection, improving annotation efficiency by combining uncertainty estimation with diversity-driven sampling. Specifically, our approach estimates epistemic uncertainty, accounting for both inaccurate detections and undetected objects, and selects diverse samples through an optimization method based on a Class-aware Adaptive Prototype bank. Our method significantly outperforms state-of-the-art baselines, achieving over $85\%$ of fully-supervised performance with just $10\%$ of the annotation budget. + +# 6. Acknowledgment + +This research is supported by the Ministry of Education, Singapore, under its MOE Academic Research Fund Tier 2 (MOE-T2EP20124-0013), it is also supported by the Agency for Science, Technology and Research (A*STAR) under its MTC Programmatic Funds (Grant No. M23L7b0021). + +# References + +[1] Sharat Agarwal, Himanshu Arora, Saket Anand, and Chetan Arora. Contextual diversity for active learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16, pages 137-153. Springer, 2020. 2, 3, 7 +[2] Kelsey Allen, Evan Shelhamer, Hanul Shin, and Joshua Tenenbaum. Infinite mixture prototypes for few-shot learning. In International conference on machine learning, pages 232-241. PMLR, 2019. 5 +[3] Eduardo Arnold, Omar Y Al-Jarrah, Mehrdad Dianati, Saber Fallah, David Oxtoby, and Alex Mouzakitis. A survey on 3d object detection methods for autonomous driving applications. IEEE Transactions on Intelligent Transportation Systems, 20(10):3782-3795, 2019. 2 +[4] Jordan T Ash, Chicheng Zhang, Akshay Krishnamurthy, John Langford, and Alekh Agarwal. Deep batch active learning by diverse, uncertain gradient lower bounds. arXiv preprint arXiv:1906.03671, 2019. 2, 7 +[5] Gui Citovsky, Giulia DeSalvo, Claudio Gentile, Lazaros Karydas, Anand Rajagopalan, Afshin Rostamizadeh, and Sanjiv Kumar. Batch active learning at scale. Advances in Neural Information Processing Systems, 34:11933-11944, 2021. 2 +[6] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 2, 4, 6 +[7] Ehsan Elhamifar, Guillermo Sapiro, Allen Yang, and S Shankar Sasrty. A convex optimization framework for active learning. In Proceedings of the IEEE International Conference on Computer Vision, pages 209-216, 2013. 2 +[8] Di Feng, Xiao Wei, Lars Rosenbaum, Atsuto Maki, and Klaus Dietmayer. Deep active learning for efficient training of a lidar 3d object detector. In 2019 IEEE Intelligent Vehicles Symposium (IV), pages 667-674. IEEE, 2019. 3 +[9] Yarin Gal, Riashat Islam, and Zoubin Ghahramani. Deep Bayesian active learning with image data. In Proceedings of the 34th International Conference on Machine Learning, pages 1183-1192. PMLR, 2017. 2, 7, 8 +[10] Mike Gashler and Tony Martinez. Temporal nonlinear dimensionality reduction. In The 2011 International Joint Conference on Neural Networks, pages 1959-1966. IEEE, 2011. 1 +[11] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 2 +[12] Ahmed Ghita, Björk Antoniussen, Walter Zimmer, Ross Greer, Christian Creß, Andreas Møgelmose, Mohan M Trivedi, and Alois C Knoll. Activeanno3d—an active learning framework for multi-modal 3d object detection. arXiv preprint arXiv:2402.03235, 2024. 2, 3 +[13] JunYoung Gwak, Christopher Choy, and Silvio Savarese. Generative sparse detection networks for 3d single-shot object detection. In Computer Vision-ECCV 2020: 16th Euro- + +pean Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 297-313. Springer, 2020. 2 +[14] Yucheng Han, Na Zhao, Weiling Chen, Keng Teck Ma, and Hanwang Zhang. Dual-perspective knowledge enrichment for semi-supervised 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2049-2057, 2024. 2 +[15] Pengkun Jiao, Na Zhao, Jingjing Chen, and Yu-Gang Jiang. Unlocking textual and visual wisdom: Open-vocabulary 3d object detection enhanced by comprehensive guidance from text and image. In European Conference on Computer Vision, pages 376-392. Springer, 2024. 2 +[16] Ajay J Joshi, Fatih Porikli, and Nikolaos Papanikolopoulos. Multi-class active learning for image classification. In 2009 IEEE conference on computer vision and pattern recognition, pages 2372-2379. IEEE, 2009. 2 +[17] Yoon-Yeong Kim, Kyungwoo Song, JoonHo Jang, and Il-Chul Moon. Lada: Look-ahead data acquisition via augmentation for deep active learning. Advances in Neural Information Processing Systems, 34:22919-22930, 2021. 2 +[18] Seong Min Kye, Kwanghee Choi, Hyeongmin Byun, and Buru Chang. Tidal: Learning training dynamics for active learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22335-22345, 2023. 3 +[19] Jean Lahoud and Bernard Ghanem. 2d-driven 3d object detection in rgb-d images. In Proceedings of the IEEE international conference on computer vision, pages 4622-4630, 2017. 6 +[20] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1 +[21] Jinpeng Lin, Zhihao Liang, Shengheng Deng, Lile Cai, Tao Jiang, Tianrui Li, Kui Jia, and Xun Xu. Exploring diversity-based active learning for 3d object detection in autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 2024. 3 +[22] Zhuoming Liu, Hao Ding, Huaping Zhong, Weijia Li, Jifeng Dai, and Conghui He. Influence selection for active learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9274-9283, 2021. 2 +[23] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[24] Yadan Luo, Zhuoxiao Chen, Zhen Fang, Zheng Zhang, Mahsa Baktashmotlagh, and Zi Huang. Kecor: Kernel coding rate maximization for active 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18279-18290, 2023. 2, 3, 7 +[25] Yadan Luo, Zhuoxiao Chen, Zijian Wang, Xin Yu, Zi Huang, and Mahsa Baktashmotlagh. Exploring active 3d object detection from a generalization perspective. arXiv preprint arXiv:2301.09249, 2023. 2, 3, 4, 7 +[26] Yi Ma, Harm Derksen, Wei Hong, and John Wright. Segmentation of multivariate mixed data via lossy data coding and compression. IEEE transactions on pattern analysis and machine intelligence, 29(9):1546-1562, 2007. 3 + +[27] Roberta Maisano, Valeria Tomaselli, Alessandro Capra, Francesco Longo, and Antonio Puliafito. Reducing complexity of 3d indoor object detection. In 2018 IEEE 4th International Forum on Research and Technology for Society and Industry (RTSI), pages 1-6. IEEE, 2018. 3 +[28] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In 2015 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 922-928. IEEE, 2015. 2 +[29] Michael Meyer and Georg Kuschk. Automotive radar dataset for deep learning based 3d object detection. In 2019 16th European radar conference (EuRAD), pages 129-132. IEEE, 2019. 3 +[30] Peng Mi, Jianghang Lin, Yiyi Zhou, Yunhang Shen, Gen Luo, Xiaoshuai Sun, Liujuan Cao, Rongrong Fu, Qiang Xu, and Rongrong Ji. Active teacher for semi-supervised object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14482-14491, 2022. 1, 3 +[31] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2906-2917, 2021. 1, 6 +[32] Aimee Moses, Srikanth Jakkampudi, Cheryl Danner, and Derek Biega. Localization-based active learning (local) for object detection in 3d point clouds. In Geospatial Informatics XII, pages 44-58. SPIE, 2022. 3 +[33] Yuzuru Nakamura, Yasunori Ishii, and Takayoshi Yamashita. Active domain adaptation with false negative prediction for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28782-28792, 2024. 5 +[34] Hieu T Nguyen and Arnold Smeulders. Active learning using pre-clustering. In Proceedings of the twenty-first international conference on Machine learning, page 79, 2004. 1, 2 +[35] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb-d data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 918-927, 2018. 6 +[36] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 2, 3, 6 +[37] Dan Roth and Kevin Small. Margin-based active learning for structured output spaces. In Machine Learning: ECML 2006: 17th European Conference on Machine Learning Berlin, Germany, September 18-22, 2006 Proceedings 17, pages 413-424. Springer, 2006. 2, 8 +[38] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. In International Conference on Learning Representations, 2018. 1, 2, 7 +[39] Claude Elwood Shannon. A mathematical theory of communication. The Bell system technical journal, 27(3):379-423, 1948. 2, 4, 8 + +[40] Xiaoke Shen and Ioannis Stamos. Frustum voxelnet for 3d object detection from rgb-d or depth images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1698-1706, 2020. 2 +[41] Yichao Shen, Zigang Geng, Yuhui Yuan, Yutong Lin, Ze Liu, Chunyu Wang, Han Hu, Nanning Zheng, and Baining Guo. V-detr: Detr with vertex relative position encoding for 3d object detection. arXiv preprint arXiv:2308.04409, 2023. 1 +[42] Hualian Sheng, Sijia Cai, Na Zhao, Bing Deng, Jianqiang Huang, Xian-Sheng Hua, Min-Jian Zhao, and Gim Hee Lee. Rethinking iou-based optimization for single-stage 3d object detection. In European Conference on Computer Vision, pages 544-561. Springer, 2022. 2 +[43] Hualian Sheng, Sijia Cai, Na Zhao, Bing Deng, Min-Jian Zhao, and Gim Hee Lee. Pdr: Progressive depth regularization for monocular 3d object detection. IEEE Transactions on Circuits and Systems for Video Technology, 33(12):7591-7603, 2023. +[44] Hualian Sheng, Sijia Cai, Na Zhao, Bing Deng, Qiao Liang, Min-Jian Zhao, and Jieping Ye. Ct3d++: Improving 3d object detection with keypoint-induced channel-wise transformer. International Journal on Computer Vision, 2025. 2 +[45] Samarth Sinha, Sayna Ebrahimi, and Trevor Darrell. Variational adversarial active learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5972-5981, 2019. 2 +[46] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 567-576, 2015. 1, 2, 4, 6 +[47] OpenPCDet Development Team. Openpcdet: An opensource toolbox for 3d object detection from point clouds. https://github.com/open-mmlab/OpenPCDet, 2020.6 +[48] Haiyang Wang, Lihe Ding, Shaocong Dong, Shaoshuai Shi, Aoxue Li, Jianan Li, Zhenguo Li, and Liwei Wang. Cagroup3d: Class-aware grouping for 3d object detection on point clouds. Advances in Neural Information Processing Systems, 35:29975-29988, 2022. 1, 2, 3, 6 +[49] Jiangyi Wang, Zhongyao Cheng, Na Zhao, Jun Cheng, and Xulei Yang. On-the-fly point feature representation for point clouds analysis. arXiv preprint arXiv:2407.21335, 2024. 2 +[50] Keze Wang, Dongyu Zhang, Ya Li, Ruimao Zhang, and Liang Lin. Cost-effective active learning for deep image classification. IEEE Transactions on Circuits and Systems for Video Technology, 27(12):2591-2600, 2016. 2 +[51] Jamie Watson, Mohamed Sayed, Zawar Qureshi, Gabriel J. Brostow, Sara Vicente, Oisin Mac Aodha, and Michael Firman. Virtual occlusions through implicit depth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9053-9064, 2023. 1 +[52] Junjie Wu. Advances in $K$ -means clustering: a data mining thinking. Springer Science & Business Media, 2012. 2, 6, 8 +[53] Jiaxi Wu, Jiaxin Chen, and Di Huang. Entropy-based active learning for object detection with progressive diversity constraint. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9397-9406, 2022. 3, 4, 7 + +[54] Tsung-Han Wu, Yueh-Cheng Liu, Yu-Kai Huang, Hsin-Ying Lee, Hung-Ting Su, Ping-Chia Huang, and Winston H Hsu. Redal: Region-based and diversity-aware active learning for point cloud semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15510-15519, 2021. 2 +[55] Qian Xie, Yu-Kun Lai, Jing Wu, Zhoutao Wang, Yiming Zhang, Kai Xu, and Jun Wang. Mlcvnet: Multi-level context votenet for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10447-10456, 2020. 2 +[56] Qian Xie, Yu-Kun Lai, Jing Wu, Zhoutao Wang, Dening Lu, Mingqiang Wei, and Jun Wang. Venet: Voting enhancement network for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3712-3721, 2021. 2, 3 +[57] Yating Xu, Conghui Hu, Na Zhao, and Gim Hee Lee. Generalized few-shot point cloud segmentation via geometric words. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21506-21515, 2023. 5 +[58] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17830-17839, 2023. 1 +[59] Chenhongyi Yang, Lichao Huang, and Elliot J Crowley. Plug and play active learning for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17784-17793, 2024. 3, 4, 7, 8 +[60] Yi Yang, Zhigang Ma, Feiping Nie, Xiaojun Chang, and Alexander G Hauptmann. Multi-class active learning by uncertainty sampling with diversity maximization. International Journal of Computer Vision, 113:113-127, 2015. 2 +[61] Tianning Yuan, Fang Wan, Mengying Fu, Jianzhuang Liu, Songcen Xu, Xiangyang Ji, and Qixiang Ye. Multiple instance active learning for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5330-5339, 2021. 3 +[62] Bohang Zhang, Jikai Jin, Cong Fang, and Liwei Wang. Improved analysis of clipping algorithms for non-convex optimization. Advances in Neural Information Processing Systems, 33:15511-15521, 2020. 6 +[63] Na Zhao and Gim Hee Lee. Static-dynamic co-teaching for class-incremental 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3436-3445, 2022. 2 +[64] Na Zhao, Tat-Seng Chua, and Gim Hee Lee. Sess: Self-ensembling semi-supervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11079–11087, 2020. 2 +[65] Na Zhao, Tat-Seng Chua, and Gim Hee Lee. Few-shot 3d point cloud semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8873-8882, 2021. 5 +[66] Na Zhao, Peisheng Qian, Fang Wu, Xun Xu, Xulei Yang, and Gim Hee Lee. Sdcot++: Improved static-dynamic co + +teaching for class-incremental 3d object detection. IEEE Transactions on Image Processing, 2024. 2 +[67] Qiang Zhou and Chaohui Yu. Point rcnn: An angle-free framework for rotated object detection. Remote Sensing, 14 (11):2605, 2022. 2 \ No newline at end of file diff --git a/2025/Uncertainty Meets Diversity_ A Comprehensive Active Learning Framework for Indoor 3D Object Detection/images.zip b/2025/Uncertainty Meets Diversity_ A Comprehensive Active Learning Framework for Indoor 3D Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7b43ca0ca7c64bb1458399031335c4c01928b246 --- /dev/null +++ b/2025/Uncertainty Meets Diversity_ A Comprehensive Active Learning Framework for Indoor 3D Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:18db5a7ce24d13611fc0f8bd234694079d38960250dd9e62442dad3890060eba +size 550795 diff --git a/2025/Uncertainty Meets Diversity_ A Comprehensive Active Learning Framework for Indoor 3D Object Detection/layout.json b/2025/Uncertainty Meets Diversity_ A Comprehensive Active Learning Framework for Indoor 3D Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3ec640c36e13b5e0d9bbe5b9fb8b29dab35518a9 --- /dev/null +++ b/2025/Uncertainty Meets Diversity_ A Comprehensive Active Learning Framework for Indoor 3D Object Detection/layout.json @@ -0,0 +1,9822 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 102, + 553, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 102, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 59, + 102, + 553, + 140 + ], + "type": "text", + "content": "Uncertainty Meets Diversity: A Comprehensive Active Learning Framework for Indoor 3D Object Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 167, + 161, + 444, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 161, + 444, + 204 + ], + "spans": [ + { + "bbox": [ + 167, + 161, + 444, + 204 + ], + "type": "text", + "content": "Jiangyi Wang, Na Zhao* \nSingapore University of Technology and Design (SUTD) \nwangjiangyi0519@gmail.com, na.zhao@sutd.edu.sg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 256, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 296, + 590 + ], + "type": "text", + "content": "Active learning has emerged as a promising approach to reduce the substantial annotation burden in 3D object detection tasks, spurring several initiatives in outdoor environments. However, its application in indoor environments remains unexplored. Compared to outdoor 3D datasets, indoor datasets face significant challenges, including fewer training samples per class, a greater number of classes, more severe class imbalance, and more diverse scene types and intra-class variances. This paper presents the first study on active learning for indoor 3D object detection, where we propose a novel framework tailored for this task. Our method incorporates two key criteria - uncertainty and diversity - to actively select the most ambiguous and informative unlabeled samples for annotation. The uncertainty criterion accounts for both inaccurate detections and undetected objects, ensuring that the most ambiguous samples are prioritized. Meanwhile, the diversity criterion is formulated as a joint optimization problem that maximizes the diversity of both object class distributions and scene types, using a new Class-aware Adaptive Prototype (CAP) bank. The CAP bank dynamically allocates representative prototypes to each class, helping to capture varying intra-class diversity across different categories. We evaluate our method on SUN RGB-D and ScanNetV2, where it outperforms baselines by a significant margin, achieving over " + }, + { + "bbox": [ + 55, + 256, + 296, + 590 + ], + "type": "inline_equation", + "content": "85\\%" + }, + { + "bbox": [ + 55, + 256, + 296, + 590 + ], + "type": "text", + "content": " of fully-supervised performance with just " + }, + { + "bbox": [ + 55, + 256, + 296, + 590 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 256, + 296, + 590 + ], + "type": "text", + "content": " of the annotation budget. Our code is available at: https://github.com/JoeWang-0519/CVPR25_UMD." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 604, + 135, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 135, + 616 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 135, + 616 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 624, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 296, + 696 + ], + "type": "text", + "content": "3D object detection, which aims to localize and identify objects within a 3D scene, is crucial for various applications, including robotics [10], autonomous driving [20, 58], and augmented/virtual reality [51]. While deep learning techniques have achieved significant success in 3D object detection [31, 41, 48], most existing approaches are data-" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 320, + 231, + 549, + 439 + ], + "blocks": [ + { + "bbox": [ + 320, + 231, + 549, + 439 + ], + "lines": [ + { + "bbox": [ + 320, + 231, + 549, + 439 + ], + "spans": [ + { + "bbox": [ + 320, + 231, + 549, + 439 + ], + "type": "image", + "image_path": "218ac544ffc061776612f813e9b811f331f308366de023c2783b1287a760b51b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 448, + 555, + 559 + ], + "lines": [ + { + "bbox": [ + 313, + 448, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 555, + 559 + ], + "type": "text", + "content": "Figure 1. Challenges in active learning for indoor 3D object detection, including both uncertainty and diversity aspects. We show two indoor scenes on ScanNetV2, with red boxes for ground truths and cyan for predictions (10% data using CAGroup3D detector), displaying only 'chair' and 'door' categories for clarity. For uncertainty, the presence of undetected objects and inaccurate detections undermines the quality of uncertainty estimation. For diversity, high scene-type diversity and varying intra-class variances in indoor environments complicate diverse sample selection. Our work presents a hybrid approach to address these challenges." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "hungry [41], requiring large amounts of annotated data for training. However, annotating objects in a 3D scene is a labor-intensive process. For example, annotating a single scene in the SUN RGB-D dataset takes over 1.9 minutes per object, with an average of approximately 14 objects per scene [46]. To mitigate the high costs of 3D annotation, active learning (AL) [30, 34, 38] offers a promising solution. AL seeks to maximize the utility of limited labeling resources by actively selecting the most informative samples for human annotation. Through this process, AL can achieve performance comparable to fully-supervised methods, but with significantly fewer labeled samples." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 219, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 219, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 219, + 713 + ], + "type": "text", + "content": "*Corresponding Author: na_zhao@sutd.edu.sg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20329" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": "Although active learning has proven effective in image-based tasks such as image classification and object detection, using uncertainty [5, 16, 37] or/and diversity [34, 38, 60] as selection criteria, its application to 3D object detection remains largely underexplored. To date, only a few studies [12, 24, 25] have applied AL techniques to outdoor 3D object detection. For instance, CRB [25] uses label conciseness to capture uncertainty, along with feature representativeness and geometric balance to enhance diversity for sample selection. KECOR [24] favors samples whose latent features have the highest coding length, thereby ensuring diversity. While these methods demonstrate effectiveness in outdoor scenarios, they are suboptimal for indoor environments, as illustrated in Tab. 2, due to the unique challenges posed by 3D object detection in indoor settings." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 254, + 295, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 295, + 410 + ], + "type": "text", + "content": "On the one hand, a 3D indoor dataset [6] could be " + }, + { + "bbox": [ + 55, + 254, + 295, + 410 + ], + "type": "inline_equation", + "content": "6.2\\mathbf{x}" + }, + { + "bbox": [ + 55, + 254, + 295, + 410 + ], + "type": "text", + "content": " smaller in scale but contain " + }, + { + "bbox": [ + 55, + 254, + 295, + 410 + ], + "type": "inline_equation", + "content": "3.3\\mathbf{x}" + }, + { + "bbox": [ + 55, + 254, + 295, + 410 + ], + "type": "text", + "content": " more number of categories, with more severe class imbalance, compared to an outdoor dataset [11], as shown in [14]. This scarcity of training samples, coupled with pronounced class imbalance, impairs 3D detection accuracy in indoor scenarios, leading to more frequent occurrences of inaccurate detection and undetected objects, as illustrated in Fig. 1. In this figure, three detections are inaccurate, incorrectly localizing background objects, while three other objects remain undetected. These challenges underscore the need to account for both inaccurately detected proposals and undetected objects when selecting informative samples from unlabeled data." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 412, + 295, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 295, + 520 + ], + "type": "text", + "content": "On the other hand, unlike outdoor scenes, which primarily feature road environments with a limited set of object classes, indoor environments exhibit greater scene-type diversity and intra-class variances, caused by diverse layouts and variations in object shapes within each class. For example, as shown in Fig. 1, five 'chair' instances can form three clusters in the embedding space due to the variations in shape and occlusion. This phenomenon in indoor environments complicate the selection of diverse samples." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": "In this work, we present the first solution to apply active learning in indoor 3D object detection, utilizing both uncertainty and diversity as selection criteria. For uncertainty, we design a two-pronged epistemic uncertainty estimation strategy to address inaccurate detections and undetected objects, which first assesses uncertainty separately for each case, then integrates them into a unified criterion. Specifically, we introduce a localization-aware uncertainty score that combines the IoU score – measured between the original detected object and its perturbed version – with the Shannon entropy. This helps mitigate the negative effects of invalid proposals, such as background or partial detections, when estimating the classification uncertainty of detected proposals. To quantify the epistemic uncertainty arising from undetectability, we develop a lightweight network that estimates the number of undetected objects. For di" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "content": "versity, we frame the diverse sample selection problem as an optimization task that jointly maximizes scene-type diversity and intra-class variances, leveraging a novel Class-aware Adaptive Prototype (CAP) bank. This CAP bank is designed to capture varying intra-class variances by generating an adaptive number of representative prototypes per class. Observing that different scene types exhibit distinct histograms of object categories as shown in Fig. 1, each composed of various prototypes, we use the CAP bank to model both class-level and scene-level diversity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 192, + 553, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 553, + 264 + ], + "type": "text", + "content": "We evaluate the proposed method on SUN RGB-D [46] and ScanNet [6] datasets. Using CAGroup3D [48] as base detection model, our approach achieves " + }, + { + "bbox": [ + 313, + 192, + 553, + 264 + ], + "type": "inline_equation", + "content": "56.11\\%" + }, + { + "bbox": [ + 313, + 192, + 553, + 264 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 192, + 553, + 264 + ], + "type": "inline_equation", + "content": "61.77\\%" + }, + { + "bbox": [ + 313, + 192, + 553, + 264 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}@0.25" + }, + { + "bbox": [ + 313, + 192, + 553, + 264 + ], + "type": "text", + "content": " on SUN RGB-D and ScanNet, respectively, reaching " + }, + { + "bbox": [ + 313, + 192, + 553, + 264 + ], + "type": "inline_equation", + "content": "87.24\\%" + }, + { + "bbox": [ + 313, + 192, + 553, + 264 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 192, + 553, + 264 + ], + "type": "inline_equation", + "content": "85.02\\%" + }, + { + "bbox": [ + 313, + 192, + 553, + 264 + ], + "type": "text", + "content": " of the fully-supervised performance with only " + }, + { + "bbox": [ + 313, + 192, + 553, + 264 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 192, + 553, + 264 + ], + "type": "text", + "content": " of the annotation budget." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 274, + 400, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 274, + 400, + 286 + ], + "spans": [ + { + "bbox": [ + 314, + 274, + 400, + 286 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 293, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 604 + ], + "type": "text", + "content": "Generic Active Learning. The majority of existing works have explored active learning (AL) strategies on image classification task, which can be categorized into uncertainty-based [5, 9, 16, 37, 50] and diversity-based [1, 7, 34, 38, 54, 60] methods based on the selection criteria used. Uncertainty-based approaches select ambiguous samples for annotation by maximizing uncertainty scores, which can be computed in various ways, such as entropy [9, 17, 39], minimum posterior probability [50], or multi-class margin [5, 16, 37]. Shannon entropy [39] is a commonly used metric in classification that quantifies uncertainty in the probability distribution, while multi-class margin [37] measures the uncertainty through the difference between the top two predicted probabilities. MC Dropout [9] employs Monte Carlo Dropout to approximate the posterior predicted probability and use Shannon entropy to assess the underlying uncertainty. Diversity-based methods [1, 38, 60] select a compact subset of samples that best represent the global data distribution. For example, Core-set [38] formulates active learning as a core-set selection problem and solves it using the " + }, + { + "bbox": [ + 313, + 293, + 555, + 604 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 293, + 555, + 604 + ], + "type": "text", + "content": "-centroid algorithm. Recent works [4, 22, 45] combine both uncertainty and diversity through hybrid approaches. A representative approach, BADGE [4], balances uncertainty and diversity by applying the " + }, + { + "bbox": [ + 313, + 293, + 555, + 604 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 293, + 555, + 604 + ], + "type": "text", + "content": "-means++ [52] algorithm in the gradient space, which considers both magnitude and direction of gradient vectors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": "3D Object Detection. We focus on indoor 3D object detection [15, 63, 64, 66] and exclude outdoor methods [3, 42-44], which differ significantly. Prior indoor 3D object detection methods can be categorized into two lines in terms of underlying representation, i.e., voxel-based methods [13, 28, 40], and point-based methods [36, 49, 55, 56, 67]. Voxel-based methods convert point cloud to voxel grids and use 3D convolution networks to address the irregularity of point cloud. In contrast, point-based methods extract point" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20330" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 70, + 531, + 243 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 531, + 243 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 531, + 243 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 531, + 243 + ], + "type": "image", + "image_path": "73f15c81cc263313548d559a4e6947e1c8a867cadcf8d24854e64352e73c1cac.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 255, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 54, + 255, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 555, + 301 + ], + "type": "text", + "content": "Figure 2. Overview of our proposed AL framework for indoor 3D object detection, exploiting both uncertainty and diversity. In the " + }, + { + "bbox": [ + 54, + 255, + 555, + 301 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 54, + 255, + 555, + 301 + ], + "type": "text", + "content": "-th round, we first construct a candidate pool of size " + }, + { + "bbox": [ + 54, + 255, + 555, + 301 + ], + "type": "inline_equation", + "content": "\\delta \\cdot B_r" + }, + { + "bbox": [ + 54, + 255, + 555, + 301 + ], + "type": "text", + "content": " using the Top-" + }, + { + "bbox": [ + 54, + 255, + 555, + 301 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 255, + 555, + 301 + ], + "type": "text", + "content": " unified uncertainty score, which accounts for both inaccurate detections and undetections. Then, we jointly optimize the intra-class and scene-type diversity to select the " + }, + { + "bbox": [ + 54, + 255, + 555, + 301 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 54, + 255, + 555, + 301 + ], + "type": "text", + "content": "-th selected dataset " + }, + { + "bbox": [ + 54, + 255, + 555, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 54, + 255, + 555, + 301 + ], + "type": "text", + "content": ". Lastly, along with the previous labeled point clouds " + }, + { + "bbox": [ + 54, + 255, + 555, + 301 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_L^{r-1}" + }, + { + "bbox": [ + 54, + 255, + 555, + 301 + ], + "type": "text", + "content": ", we retrain the model until the total labeled dataset reaches the annotation budget " + }, + { + "bbox": [ + 54, + 255, + 555, + 301 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 54, + 255, + 555, + 301 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 313, + 296, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 313, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 54, + 313, + 296, + 433 + ], + "type": "text", + "content": "level features and use a voting mechanism [36, 56] to encourage points to vote towards their corresponding object centers, followed by grouping the vote points into clusters and then generating object proposals. In this work, we adopt the state-of-the-art method in indoor 3D object detection, CAGroup3D [48], which takes advantage of the two lines of work through a two-stage fully sparse convolutional 3D detection framework. It extracts voxel-wise features from raw point clouds and leverages 'voting and grouping' mechanism to generate object proposals." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 438, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 438, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 438, + 296, + 714 + ], + "type": "text", + "content": "Active Learning for Object Detection. Despite extensive studies on AL in image classification, applying it to image-based object detection is challenging due to the task's complexity, which involves both localization and classification of multiple objects. Several works [1, 18, 30, 53, 59, 61] have investigated active object detection by adapting generic AL techniques to this task. For example, CDAL [1] designs contextual diversity to capture spatial and semantic context diversity within a dataset, while DivProto [53] and PPAL [59] incorporate both diversity and uncertainty. While these methods yield promising results in 2D object detection, they cannot be directly applied to 3D object detection due to the significant modality gap. Recently, driven by the need to reduce annotation burden in 3D object detection for autonomous driving, a few studies [8, 12, 21, 24, 25, 27, 29, 32] have explored active learning for 3D object detection in outdoor environments. Notably, CRB [25] selects active samples with concise labels, representative features, and balanced density progressively. KECOR [24] leverages kernel coding rate [26] to measure the richness of latent features. However, these outdoor methods are suboptimal for indoor environments, where 3D object detection faces unique challenges. To ad" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 313, + 555, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 555, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 555, + 385 + ], + "type": "text", + "content": "dress these, this paper presents the first solution for applying active learning to indoor 3D object detection, overcoming these challenges by incorporating designs to estimate uncertainty in both inaccurate detections and undetections, while simultaneously optimizing diversity with respect to scene-type and intra-class variances." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 401, + 397, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 397, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 397, + 415 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "text", + "content": "Given a point cloud " + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i = \\{(x,y,z)\\}" + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "text", + "content": ", a 3D object detection model aims to predict the locations and class labels for a set of objects " + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\mathfrak{B}_i = \\{(\\mathbf{b}_k,y_k)\\}" + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_k\\in \\mathbb{R}^7" + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "text", + "content": " represents the parameter of the " + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "text", + "content": "-th bounding box, including the center position, box size and yaw angle, and " + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "inline_equation", + "content": "y_{k}" + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "text", + "content": " indicates its predicted class label. In the active learning setting, our objective is to train the 3D detection model by constructing a small labeled dataset " + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_L" + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "text", + "content": ", which starts with a small portion of random samples " + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_L^0" + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "text", + "content": " and is progressively expanded over " + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "text", + "content": " rounds from a large pool " + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "text", + "content": ". In each active learning round, " + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "inline_equation", + "content": "r\\in \\{1,\\dots,R\\}" + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "inline_equation", + "content": "B_{r}" + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "text", + "content": " samples are selected from the unlabeled dataset " + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_U = \\mathcal{D} / \\mathcal{D}_L^{r - 1}" + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "text", + "content": " by an active learning policy, and labeled by an oracle, which constitute the " + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "text", + "content": "-th selected dataset " + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "text", + "content": ". Then, " + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_L" + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "text", + "content": " will be updated as " + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_L^r\\gets \\mathcal{D}_L^{r - 1}\\cup \\mathcal{D}_r" + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "text", + "content": " until the size of " + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_L" + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "text", + "content": " reaches the final budget " + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "inline_equation", + "content": "|\\mathcal{D}_L^0 | + \\sum_{r = 1}^{R}B_r = B" + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "inline_equation", + "content": "B\\ll |\\mathcal{D}|" + }, + { + "bbox": [ + 313, + 422, + 555, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": "To tackle active 3D object detection in indoor scenarios, we propose a novel framework, illustrated in Fig. 2, that leverages both uncertainty and diversity as selection criteria. Following previous work [21, 53, 59] that considers both uncertainty and diversity, we adopt a budget expansion ratio " + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\delta \\in \\mathbb{Z}^{+}" + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": " to first select a candidate pool " + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r^{\\mathrm{cand}}" + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "inline_equation", + "content": "K = \\delta \\cdot B_r" + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": " based on the uncertainty criterion for " + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": "-th active learning round. Then, from this pool, the " + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": " of size" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20331" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 71, + 295, + 106 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 295, + 106 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 295, + 106 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 295, + 106 + ], + "type": "table", + "html": "
DatasetmAP@.25mAP@.50mAR@.25mAR@.50
SUN RGB-D [46]28.9112.4867.1226.09
ScanNetV2 [6]35.7520.4076.6542.48
", + "image_path": "959d9e7a7148fab5539adbc4f1411f8e4bba96da6d6c854a854023714cc89ba1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 111, + 295, + 156 + ], + "lines": [ + { + "bbox": [ + 55, + 111, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 111, + 295, + 156 + ], + "type": "text", + "content": "Table 1. Motivation for designing epistemic uncertainty estimation. We report the mean average precision (mAP) and mean average recall (mAR) at two IoU thresholds (0.25 and 0.50) for the initial detection model, which is trained using " + }, + { + "bbox": [ + 55, + 111, + 295, + 156 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 55, + 111, + 295, + 156 + ], + "type": "text", + "content": " labeled data." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 162, + 296, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 162, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 54, + 162, + 296, + 282 + ], + "type": "inline_equation", + "content": "B_{r}" + }, + { + "bbox": [ + 54, + 162, + 296, + 282 + ], + "type": "text", + "content": " is chosen based on the diversity criterion. For the uncertainty criterion, we propose a two-pronged epistemic uncertainty estimation strategy that generates a unified score, accounting for uncertainty in both detected and undetected objects. This is achieved through a localization-aware uncertainty score and an estimate of the undetected object count (Sec. 3.1). For the diversity criterion, we frame the diverse sample selection as an optimization problem and propose an efficient solution that jointly optimizes intra-class variances and scene-type diversity (Sec. 3.2)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 289, + 237, + 302 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 237, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 237, + 302 + ], + "type": "text", + "content": "3.1. Epistemic Uncertainty Estimation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 306, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 306, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 306, + 296, + 558 + ], + "type": "text", + "content": "In Tab. 1, we evaluate the detection performance of a 3D object detection model trained on the initial random samples " + }, + { + "bbox": [ + 56, + 306, + 296, + 558 + ], + "type": "inline_equation", + "content": "D_{L}^{0}" + }, + { + "bbox": [ + 56, + 306, + 296, + 558 + ], + "type": "text", + "content": " (2% labeled data). As shown, the model achieves 20.40% mAP@.50 and 42.48% mAR@.50 on ScanNet [6], indicating that approximately 80% of the predictions are inaccurate and 60% of ground-truth objects remain undetected. The results are even worse on the more challenging SUN RGB-D [46] dataset. These findings highlight the need to address the two types of epistemic uncertainty in 3D object detection: inaccurate detections and undetected objects. Given that the detection task involves both localization and classification, we estimate the epistemic uncertainty of inaccurate detections by considering both aspects. Specifically, we introduce a localization-aware uncertainty score that exploits an IoU score for localization reliability to combine with Shannon entropy. To handle undetections, we design a lightweight network to estimate the number of undetected objects, which serves as the uncertainty measure for undetections. These two uncertainty scores are then integrated into a unified uncertainty score, which is used for Top-K selection to generate the candidate pool " + }, + { + "bbox": [ + 56, + 306, + 296, + 558 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r^{\\mathrm{cand}}" + }, + { + "bbox": [ + 56, + 306, + 296, + 558 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 295, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 643 + ], + "type": "text", + "content": "Localization-aware Uncertainty Score. To estimate the uncertainty of detected objects from a point cloud " + }, + { + "bbox": [ + 55, + 558, + 295, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 55, + 558, + 295, + 643 + ], + "type": "text", + "content": ", we first compute the classification uncertainty using the Shannon entropy [39] of the predicted class probabilities for all detected object proposals. For the " + }, + { + "bbox": [ + 55, + 558, + 295, + 643 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 558, + 295, + 643 + ], + "type": "text", + "content": "-th proposal, with predicted class probabilities over " + }, + { + "bbox": [ + 55, + 558, + 295, + 643 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 558, + 295, + 643 + ], + "type": "text", + "content": " classes denoted as " + }, + { + "bbox": [ + 55, + 558, + 295, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_j = \\{p_{j,c}\\}_{c=1}^C \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 55, + 558, + 295, + 643 + ], + "type": "text", + "content": ", the Shannon entropy is calculated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 114, + 650, + 295, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 650, + 295, + 682 + ], + "spans": [ + { + "bbox": [ + 114, + 650, + 295, + 682 + ], + "type": "interline_equation", + "content": "E (\\mathbf {p} _ {j}) = - \\sum_ {c = 1} ^ {C} p _ {j, c} \\log (p _ {j, c}). \\tag {1}", + "image_path": "5596b006e7cd8f670c0962738a84160efcc6b1818acc8fcde7ccd942bf5da9d5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "However, not all detected proposals are valid, as some may correspond to background or partial objects. To mitigate" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": "the impact of these invalid proposals, we incorporate an IoU score that measures the similarity between the original detection and its perturbed version, obtained by applying weak augmentations (e.g., random rotation and scaling) to the input point cloud. The intuition is that if the model is confident about an object, small perturbations should not significantly affect its predicted position, while background or partial proposals will likely produce inconsistent predictions. Consequently, the localization-aware uncertainty score for detected objects in " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 363, + 198, + 555, + 232 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 198, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 363, + 198, + 555, + 232 + ], + "type": "interline_equation", + "content": "\\mathbf {U} _ {\\det } \\left(\\mathcal {P} _ {i}\\right) := \\frac {\\sum_ {j = 1} ^ {\\left| \\mathfrak {B} _ {i} \\right|} \\operatorname {I o U} _ {j} \\cdot E \\left(\\mathbf {p} _ {j}\\right)}{\\sum_ {j = 1} ^ {\\left| \\mathfrak {B} _ {i} \\right|} \\operatorname {I o U} _ {j}}. \\tag {2}", + "image_path": "d457c74f5ee1a290b0866906dba90737bd2f64009dd9736b35a3c2c6c5cd66f7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 236, + 555, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 272 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 272 + ], + "type": "text", + "content": "Our IoU scores serve as soft weights, adaptively reducing the contribution of invalid objects in the epistemic uncertainty estimation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 273, + 555, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 555, + 440 + ], + "type": "text", + "content": "Undetected Number Prediction Network. When estimating uncertainty for active object detection, previous works [25, 53, 59] typically focus only on the uncertainty within the detected objects. However, the uncertainty of undetected objects is equally important, especially in the context of active learning for indoor 3D object detection, where model performance is often hindered by the scarcity of training samples and severe class imbalance. As shown by the mean average recall (mAR) scores in Tab. 1, more than half of the ground truth objects remain undetected in both 3D indoor datasets. This observation highlights the need to account for the uncertainty of undetected objects. To address this, we design a lightweight network to estimate the number of undetected objects." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 441, + 556, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 556, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 556, + 632 + ], + "type": "text", + "content": "As depicted in Fig. 2, we first extract the feature map " + }, + { + "bbox": [ + 313, + 441, + 556, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 313, + 441, + 556, + 632 + ], + "type": "text", + "content": " from the point cloud " + }, + { + "bbox": [ + 313, + 441, + 556, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 313, + 441, + 556, + 632 + ], + "type": "text", + "content": " using the feature extractor. We then apply global average pooling (GAP) on " + }, + { + "bbox": [ + 313, + 441, + 556, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 313, + 441, + 556, + 632 + ], + "type": "text", + "content": " to obtain the global feature " + }, + { + "bbox": [ + 313, + 441, + 556, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_i" + }, + { + "bbox": [ + 313, + 441, + 556, + 632 + ], + "type": "text", + "content": ". Simultaneously, we create a masked feature map by masking out the regions corresponding to the detected bounding boxes produced by RoI refinement. This masked feature map is passed through GAP to yield the undetected feature " + }, + { + "bbox": [ + 313, + 441, + 556, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_i" + }, + { + "bbox": [ + 313, + 441, + 556, + 632 + ], + "type": "text", + "content": ". We concatenate these two features, " + }, + { + "bbox": [ + 313, + 441, + 556, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_i" + }, + { + "bbox": [ + 313, + 441, + 556, + 632 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 441, + 556, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_i" + }, + { + "bbox": [ + 313, + 441, + 556, + 632 + ], + "type": "text", + "content": ", and use the concatenation as input to our undetected object number prediction network " + }, + { + "bbox": [ + 313, + 441, + 556, + 632 + ], + "type": "inline_equation", + "content": "\\Theta_{\\mathrm{undet}}" + }, + { + "bbox": [ + 313, + 441, + 556, + 632 + ], + "type": "text", + "content": ", a lightweight network consisting of three-layer MLPs. The output of the network, i.e. " + }, + { + "bbox": [ + 313, + 441, + 556, + 632 + ], + "type": "inline_equation", + "content": "\\Theta_{\\mathrm{undet}}([\\mathbf{g}_i; \\mathbf{u}_i])" + }, + { + "bbox": [ + 313, + 441, + 556, + 632 + ], + "type": "text", + "content": " represents the uncertainty score for undetections, denoted as " + }, + { + "bbox": [ + 313, + 441, + 556, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{\\mathrm{undet}}" + }, + { + "bbox": [ + 313, + 441, + 556, + 632 + ], + "type": "text", + "content": ". This output is compared with the true number of undetected objects, computed based on the detected objects " + }, + { + "bbox": [ + 313, + 441, + 556, + 632 + ], + "type": "inline_equation", + "content": "\\mathfrak{B}_i" + }, + { + "bbox": [ + 313, + 441, + 556, + 632 + ], + "type": "text", + "content": " and the ground-truth objects " + }, + { + "bbox": [ + 313, + 441, + 556, + 632 + ], + "type": "inline_equation", + "content": "\\mathfrak{B}_i^*" + }, + { + "bbox": [ + 313, + 441, + 556, + 632 + ], + "type": "text", + "content": ", to compute the loss for training " + }, + { + "bbox": [ + 313, + 441, + 556, + 632 + ], + "type": "inline_equation", + "content": "\\Theta_{\\mathrm{undet}}" + }, + { + "bbox": [ + 313, + 441, + 556, + 632 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 339, + 638, + 555, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 638, + 555, + 672 + ], + "spans": [ + { + "bbox": [ + 339, + 638, + 555, + 672 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {u n d e t}} = \\sum_ {i = 1} ^ {| \\mathcal {D} _ {L} |} \\left(\\Theta_ {\\text {u n d e t}} \\left(\\left[ \\mathbf {g} _ {i}; \\mathbf {u} _ {i} \\right]\\right) - \\mathcal {N} \\left(\\mathfrak {B} _ {i}, \\mathfrak {B} _ {i} ^ {*}\\right)\\right) ^ {2}, \\tag {3}", + "image_path": "3b4fafde96043956164806cfbe779ab63c6a581e5cf75dd4a931bd0286ffcda7.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": " is a function that computes the number of undetected objects, with details provided in the supplementary material. Additionally, to enhance the generalizability" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20332" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "text", + "content": "of the network " + }, + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "inline_equation", + "content": "\\Theta_{\\mathrm{undet}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "text", + "content": ", we employ random masking of the predicted boxes during the training phase, encouraging the network to better capture the undetectability in point clouds. Unified Uncertainty Score. After obtaining the two uncertainty scores, " + }, + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{\\mathrm{det}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{\\mathrm{undet}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "text", + "content": ", we aim to combine them into a unified uncertainty score. However, these two scores are on different scales, making their direct product infeasible. To address this issue, we introduce a normalization step, inspired by [33], using a hyperparameter " + }, + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "text", + "content": ". Specifically, we normalize each score as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 83, + 199, + 295, + 225 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 199, + 295, + 225 + ], + "spans": [ + { + "bbox": [ + 83, + 199, + 295, + 225 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {U}} _ {\\star} (\\mathcal {P} _ {i}) = \\max (0, \\frac {\\mathbf {U} _ {\\star} (\\mathcal {P} _ {i}) - \\mu (\\mathbf {U} _ {\\star})}{k \\cdot \\sigma (\\mathbf {U} _ {\\star})} + 0. 5), \\tag {4}", + "image_path": "b6e3f3574c101072f467e1de20c846925798aae51ede8d3ddd731e2d68a4cd92.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 234, + 296, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 234, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 296, + 282 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 234, + 296, + 282 + ], + "type": "inline_equation", + "content": "\\star \\in \\{\\mathrm{det},\\mathrm{undet}\\}" + }, + { + "bbox": [ + 55, + 234, + 296, + 282 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 234, + 296, + 282 + ], + "type": "inline_equation", + "content": "\\mu (\\cdot)" + }, + { + "bbox": [ + 55, + 234, + 296, + 282 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 234, + 296, + 282 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 55, + 234, + 296, + 282 + ], + "type": "text", + "content": " denote the mean and standard deviation, respectively. Once the scores are normalized, we compute their product to yield the unified uncertainty score." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 289, + 264, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 264, + 303 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 264, + 303 + ], + "type": "text", + "content": "3.2. Prototype-based Diversity Optimization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 307, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 307, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 56, + 307, + 296, + 605 + ], + "type": "text", + "content": "As illustrated in Fig. 1, indoor environments exhibit significant intra-class variances and scene-type diversity, both of which motivate our approach to the diversity criterion. In 3D indoor scenes, the intra-class variances can vary greatly across object categories due to the differences in shape and occlusion. For example, the variance in 'chairs' is typically high due to the wide variety of shapes and designs, whereas 'doors' tend to have lower variance due to their more standardized geometry. This variability makes methods that model each class with a fixed number of prototypes [57, 65] insufficient to capture the intra-class distributions. To address this, we propose a novel Class-aware Adaptive Prototype (CAP) bank, which dynamically allocates a varying number of prototypes for each class, inferred from the predicted object proposals. Additionally, we observe that indoor scenes naturally form distinct clusters based on the distribution of object categories. For instance, bathrooms typically contain toilets and bathtubs, while bedrooms feature beds and dressers. Through this observation, we represent each scene as a histogram of the prototypes corresponding to the predicted object categories within the scene. As both intra-class variances and scene-type diversity can be modeled using the CAP bank, we frame diverse sample selection as an optimization problem that jointly maximizes both aspects, and propose an efficient solution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "Class-aware Adaptive Prototype Bank. The design of our Class-aware Adaptive Prototype (CAP) bank is inspired by infinite mixture prototypes (IMP) [2], which was originally proposed for few-shot learning. IMP allows for inferring the number of clusters from the data, rather than relying on a pre-defined number of clusters. To build the CAP bank, we first iterate over the initial labeled set " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "D_{L}^{r - 1}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " to initialize one prototype per class. Specifically, we leverage the ground-truth annotations to extract features for all objects belong-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": "ing to the " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": "-th class in " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "D_L^{r-1}" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": ", and compute their average as to form the initial prototype " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "\\mu_{1,c}" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " for class " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": ". Subsequently, we iterate over the unlabeled dataset " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_U" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " and update the prototypes for each class, resulting in our CAP bank, denoted as " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{M} = \\{\\{\\pmb{\\mu}_{k,c}\\}_{k=1}^{M_c}\\}_{c=1}^C" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{k,c}" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": "-th prototype for class " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "M_c" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " is the number of prototypes for class " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": ", which may vary between classes, and " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " is the total number of classes. In each iteration (one batch), we use the detection model from the previous round to generate object proposals and extract their features and predicted class labels. For each predicted object, we compute the cosine similarity between its feature embedding " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_j" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " and each of the existing prototypes in its predicted class. This similarity score is formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 346, + 245, + 555, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 245, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 346, + 245, + 555, + 275 + ], + "type": "interline_equation", + "content": "\\operatorname {S i m} \\left(\\mathbf {o} _ {j}, \\boldsymbol {\\mu} _ {k, c}\\right) = \\frac {\\mathbf {o} _ {j} ^ {T} \\cdot \\boldsymbol {\\mu} _ {k , c}}{\\left| \\left| \\mathbf {o} _ {j} \\right| \\right| _ {2} \\left| \\left| \\boldsymbol {\\mu} _ {k , c} \\right| \\right| _ {2}} \\in [ - 1, 1 ]. \\tag {5}", + "image_path": "a5d3447449e3e33553d06914ce08ba3b8279d65c18643490a4013b7b85d2244c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 282, + 555, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 555, + 319 + ], + "type": "text", + "content": "After calculating the similarity scores, each object is either assigned to an existing prototype or used to form a new prototype, according to the following assignment rule:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 324, + 327, + 554, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 327, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 324, + 327, + 554, + 376 + ], + "type": "interline_equation", + "content": "\\mathcal {A} \\left(\\mathbf {o} _ {j}\\right) = \\left\\{ \\begin{array}{l} M _ {c} + 1, \\text {i f} \\max _ {k} \\operatorname {S i m} \\left(\\mathbf {o} _ {j}, \\boldsymbol {\\mu} _ {k, c}\\right) < \\tau_ {\\text {s i m}}, \\\\ \\arg \\max _ {k} \\operatorname {S i m} \\left(\\mathbf {o} _ {j}, \\boldsymbol {\\mu} _ {k, c}\\right), \\quad \\text {o t h e r w i s e}. \\end{array} \\right. \\tag {6}", + "image_path": "b94465bf7931e1e89884dabdef6be17286254582ba77fbecd5decd67299fe712.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 376, + 556, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 556, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 556, + 496 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 376, + 556, + 496 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{sim}}" + }, + { + "bbox": [ + 313, + 376, + 556, + 496 + ], + "type": "text", + "content": " is similarity threshold, which is set to 0.3 by default. Finally, the prototypes - whether assigned or newly formed - are updated using the new object proposals in each batch. However, since many predicted bounding boxes might be invalid, especially in the early stages of training (e.g., as shown in Tab. 1), we seek to mitigate the influence of these invalid proposals when updating the prototypes. To this end, we incorporate the IoU score from Eq. 2 into the prototype update process, effectively weighing each object's contribution to the prototype updates as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 353, + 503, + 555, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 503, + 555, + 538 + ], + "spans": [ + { + "bbox": [ + 353, + 503, + 555, + 538 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} _ {k, c} = \\frac {\\sum_ {j = 1} ^ {| \\mathfrak {B} _ {i} |} \\mathbb {1} \\left\\{\\mathcal {A} \\left(\\mathbf {o} _ {j}\\right) , k \\right\\} \\cdot \\operatorname {I o U} _ {j} \\cdot \\mathbf {o} _ {j}}{\\sum_ {j = 1} ^ {| \\mathfrak {B} _ {i} |} \\mathbb {1} \\left\\{\\mathcal {A} \\left(\\mathbf {o} _ {j}\\right) , k \\right\\} \\cdot \\operatorname {I o U} _ {j}}, \\tag {7}", + "image_path": "51b0939a944d4b5e5ae5abd4912c34969e766493e118f08e9d69d69b7a90ef3a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 546, + 554, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 554, + 582 + ], + "type": "text", + "content": "where the indicator function " + }, + { + "bbox": [ + 313, + 546, + 554, + 582 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(m,n)" + }, + { + "bbox": [ + 313, + 546, + 554, + 582 + ], + "type": "text", + "content": " equals to 1 if " + }, + { + "bbox": [ + 313, + 546, + 554, + 582 + ], + "type": "inline_equation", + "content": "m = n" + }, + { + "bbox": [ + 313, + 546, + 554, + 582 + ], + "type": "text", + "content": ", and 0 otherwise. The complete CAP bank creation procedure is outlined in Algo. 1 of the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": "Prototype-based Diversity Optimization. With the generated CAP bank, we can characterize each class through a set of prototypes, and further represent objects as histograms of these prototypes. These histograms allow us to capture the intra-class variance for each class, as objects within the same class are reflected in the variation of their corresponding histograms. Furthermore, based on the observation that each scene can be characterized by the distribution of object categories, which distinguishes one scene type from another, we also model scenes as histograms of prototypes. Therefore, this histogram-based representation" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20333" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": "implicitly captures the scene type. Since both intra-class variance and scene-type diversity are encoded in the prototype histograms, we design a novel and efficient optimization method to quantify the diversity criterion by considering both these two aspects. Specifically, we formulate the optimization for diverse sample selection as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 152, + 294, + 246 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 152, + 294, + 246 + ], + "spans": [ + { + "bbox": [ + 65, + 152, + 294, + 246 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\max _ {\\mathcal {D} _ {r} \\subset \\mathcal {D} _ {r} ^ {\\text {c a n d}}} E (\\mathbf {h} (\\mathcal {D} _ {r})) + \\beta | | \\mathbf {h} (\\mathcal {D} _ {r}) | | _ {1}, \\\\ \\begin{array}{l} \\text {s . t .} \\quad \\mathbf {h} (\\mathcal {D} _ {r}) = \\sum_ {\\mathcal {P} _ {i} \\in \\mathcal {D} _ {r}} \\mathbf {h} (\\mathcal {P} _ {i}) = \\sum_ {\\mathcal {P} _ {i} \\in \\mathcal {D} _ {r}} \\sum_ {\\mathbf {o} _ {j} \\in \\mathcal {P} _ {i}} \\mathbf {h} (\\mathbf {o} _ {j}), \\end{array} \\\\ \\mathbf {h} \\left(\\mathbf {o} _ {j}\\right) = \\sum_ {c = 1} ^ {C} \\sum_ {k = 1} ^ {M _ {c}} [ \\text {o n e} \\_ \\text {h o t} (\\mathcal {A} \\left(\\mathbf {o} _ {j}\\right)) ] ^ {c, k}, \\tag {8} \\\\ \\end{array}", + "image_path": "878a05cc3b17205a65375403d11f4c4feaa98e07d008205929f1c8d6bc920e61.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 247, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 247, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{h}(\\mathbf{o}_j)" + }, + { + "bbox": [ + 55, + 247, + 296, + 437 + ], + "type": "text", + "content": " represents the assignment histogram of " + }, + { + "bbox": [ + 55, + 247, + 296, + 437 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 247, + 296, + 437 + ], + "type": "text", + "content": "-th object w.r.t. the CAP bank " + }, + { + "bbox": [ + 55, + 247, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 55, + 247, + 296, + 437 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 55, + 247, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{h}(\\mathcal{P}_i)" + }, + { + "bbox": [ + 55, + 247, + 296, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 247, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{h}(\\mathcal{D}_r)" + }, + { + "bbox": [ + 55, + 247, + 296, + 437 + ], + "type": "text", + "content": " represent the histograms of all objects in the individual point cloud " + }, + { + "bbox": [ + 55, + 247, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 55, + 247, + 296, + 437 + ], + "type": "text", + "content": " and the selected labeled samples " + }, + { + "bbox": [ + 55, + 247, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 55, + 247, + 296, + 437 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 55, + 247, + 296, + 437 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 247, + 296, + 437 + ], + "type": "text", + "content": "-th round, respectively. The two terms in this objective function are designed to select samples that are both diverse and holistic across all prototypes in the CAP bank. The first term captures the intra-class variance for all classes, ensuring that the selected samples represent the diversity within each class as well as across different classes. The second term promotes scene-level diversity by encouraging coverage across different object categories, which implicitly increases the diversity in scene types. Given that the first term (entropy) and the second term (" + }, + { + "bbox": [ + 55, + 247, + 296, + 437 + ], + "type": "inline_equation", + "content": "l_1" + }, + { + "bbox": [ + 55, + 247, + 296, + 437 + ], + "type": "text", + "content": "-norm) have different magnitudes, we introduce a hyperparameter " + }, + { + "bbox": [ + 55, + 247, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 247, + 296, + 437 + ], + "type": "text", + "content": " to rescale the second term. In our experiments, we set " + }, + { + "bbox": [ + 55, + 247, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\beta = 0.01" + }, + { + "bbox": [ + 55, + 247, + 296, + 437 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 438, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 510 + ], + "type": "text", + "content": "However, directly solving the optimization in Eq. 8 is computationally expensive due to the combinatorial search over " + }, + { + "bbox": [ + 55, + 438, + 296, + 510 + ], + "type": "inline_equation", + "content": "|\\mathcal{D}_r^{\\mathrm{cand}}|" + }, + { + "bbox": [ + 55, + 438, + 296, + 510 + ], + "type": "text", + "content": " samples. To efficiently narrow the search space while maintaining solution quality, we leverage one key property of the objective function over " + }, + { + "bbox": [ + 55, + 438, + 296, + 510 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 438, + 296, + 510 + ], + "type": "text", + "content": " disjoint partitions " + }, + { + "bbox": [ + 55, + 438, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{D}_{r,s}\\}_{s = 1}^{S}" + }, + { + "bbox": [ + 55, + 438, + 296, + 510 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 438, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 55, + 438, + 296, + 510 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 438, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\cup_{s}\\mathcal{D}_{r,s} = \\mathcal{D}_r" + }, + { + "bbox": [ + 55, + 438, + 296, + 510 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 512, + 294, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 294, + 569 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} E \\left(\\mathbf {h} \\left(\\mathcal {D} _ {r}\\right)\\right) + \\beta | | \\mathbf {h} \\left(\\mathcal {D} _ {r}\\right) | | _ {1} \\\\ \\stackrel {c} {=} \\sum_ {s = 1} ^ {S} \\left\\{\\frac {N _ {s}}{N _ {\\text {t o t a l}}} E \\left(\\mathbf {h} \\left(\\mathcal {D} _ {r, s}\\right)\\right) + \\beta \\left\\| \\mathbf {h} \\left(\\mathcal {D} _ {r, s}\\right) \\right\\| _ {1} \\right\\} + \\mathcal {O} \\left(\\frac {N _ {\\text {i n t e r}}}{N _ {\\text {t o t a l}}}\\right), \\tag {9} \\\\ \\end{array}", + "image_path": "4c5f928f163fc33b56abcd39489038583fd873da8c2c49f033d185aaa789be73.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{h}(\\mathcal{D}_{r,s})" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " represents the histograms of all objects in " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "-th partition " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{r,s}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{total}}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " are the number of objects in the " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "-th partition " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{r,s}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " and selected labeled samples " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{inter}}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " is the number of objects that share the same prototypes but belong to different partitions, and " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\cdot)" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " is the Big O notation from mathematics. To align with the " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " disjoint partitions of " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": ", we use k-means++ [52] to group the candidate pool " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r^{\\mathrm{cand}}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " clusters, denoted as " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{D}_{r,s}^{\\mathrm{cand}}\\}_{s = 1}^{S}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": ", each of which is approximately disjoint with each other w.r.t. prototype assignments. This guarantees that " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\frac{N_{\\mathrm{inter}}}{N_{\\mathrm{total}}} \\to 0" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": ", and then allows us to decompose the original optimization problem into " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " sub-problems within each cluster " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{r,s}^{\\mathrm{cand}}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 315, + 69, + 556, + 170 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 69, + 556, + 170 + ], + "spans": [ + { + "bbox": [ + 315, + 69, + 556, + 170 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\max _ {\\mathcal {D} _ {r, s} \\subset \\mathcal {D} _ {r, s} ^ {\\text {c a n d}}} E (\\mathbf {h} (\\mathcal {D} _ {r, s})) + \\beta \\frac {N _ {\\text {t o t a l}}}{N _ {s}} | | \\mathbf {h} (\\mathcal {D} _ {r, s}) | | _ {1}, \\\\ \\begin{array}{l} \\text {s . t .} \\quad \\mathbf {h} (\\mathcal {D} _ {r, s}) = \\sum_ {\\mathcal {P} _ {i} \\in \\mathcal {D} _ {r, s}} \\mathbf {h} (\\mathcal {P} _ {i}) = \\sum_ {\\mathcal {P} _ {i} \\in \\mathcal {D} _ {r, s}} \\sum_ {\\mathbf {o} _ {j} \\in \\mathcal {P} _ {i}} \\mathbf {h} (\\mathbf {o} _ {j}), \\end{array} \\\\ \\mathbf {h} \\left(\\mathbf {o} _ {j}\\right) = \\sum_ {c = 1} ^ {C} \\sum_ {k = 1} ^ {M _ {c}} [ \\text {o n e} \\text {h o t} (\\mathcal {A} \\left(\\mathbf {o} _ {j}\\right)) ] ^ {c, k}. \\tag {10} \\\\ \\end{array}", + "image_path": "f16a5b4301b20c52b0e297b1721c5e8e5260e6f001798b97847e17536e872eed.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 171, + 555, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 555, + 232 + ], + "type": "text", + "content": "We use greedy search to solve the " + }, + { + "bbox": [ + 313, + 171, + 555, + 232 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 171, + 555, + 232 + ], + "type": "text", + "content": " sub-problems to obtain the diverse samples " + }, + { + "bbox": [ + 313, + 171, + 555, + 232 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{r,s}" + }, + { + "bbox": [ + 313, + 171, + 555, + 232 + ], + "type": "text", + "content": " for each cluster, and combine them together as the selected dataset " + }, + { + "bbox": [ + 313, + 171, + 555, + 232 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 313, + 171, + 555, + 232 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 313, + 171, + 555, + 232 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 171, + 555, + 232 + ], + "type": "text", + "content": "-th active learning round. We provide the detailed derivation of Eq. 9 in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 243, + 395, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 395, + 257 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 395, + 257 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 264, + 435, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 435, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 435, + 277 + ], + "type": "text", + "content": "4.1. Experimental Setups" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 281, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 460 + ], + "type": "text", + "content": "3D Detection Datasets. We evaluate our proposed method on two 3D indoor object detection benchmarks: SUN RGB-D [46] and ScanNet [6]. SUN RGB-D is an indoor dataset for 3D object detection, which contains 10,335 RGB-D images with 3D room layouts from four different sensors. Officially, they are split into 5,285 training samples and 5,050 testing samples. Following the standard evaluation protocol [19, 35, 36], we evaluate the performance on 10 most common categories. ScanNetV2 is a large 3D dataset with 1,513 RGB-D scans of 707 indoor scenes. These scans are split into 1,201 samples for training and 312 samples for testing, each of which is annotated with dense semantic segmentation masks. We follow [36] to derive the axis-aligned bounding boxes without considering the orientation, and adopt 18 object categories for evaluation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 461, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 554, + 653 + ], + "type": "text", + "content": "Implementation Details. We use the state-of-the-art CA-Group3D [48] as the detection model, and implement our code base using OpenPCDet [47] toolkit. Following the training strategies in [48], we set the voxel size as " + }, + { + "bbox": [ + 313, + 461, + 554, + 653 + ], + "type": "inline_equation", + "content": "0.02\\mathrm{m}" + }, + { + "bbox": [ + 313, + 461, + 554, + 653 + ], + "type": "text", + "content": ". For both datasets, the network is trained by AdamW optimizer [23] with an initial learning rate and weight decay rate of 0.001 and 0.0001 respectively. The gradnorm clip [62] is applied to stabilize the training process. For each active learning round, SUN RGB-D takes 48 epochs to train on the updated labeled dataset, and the learning rate decays on [32, 44] epochs with decay rate of 0.1. ScanNet requires 120 epochs for training and decays the learning rate on [80, 110] epochs by the same decay rate. All experiments are conducted using one 3090Ti NVIDIA GPU. We set the similarity threshold " + }, + { + "bbox": [ + 313, + 461, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{sim}}" + }, + { + "bbox": [ + 313, + 461, + 554, + 653 + ], + "type": "text", + "content": " to 0.3, balance ratio " + }, + { + "bbox": [ + 313, + 461, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 461, + 554, + 653 + ], + "type": "text", + "content": " to 0.01, and budget expanding ratio " + }, + { + "bbox": [ + 313, + 461, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 461, + 554, + 653 + ], + "type": "text", + "content": " to 6 empirically." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "text", + "content": "Evaluation. Following the evaluation protocol in [31, 36, 48], we report mean average precision (mAP) with different IoU thresholds, i.e., 0.25 and 0.50. Additionally, to investigate the effectiveness of different active learning strategies, we record the mAP with " + }, + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "text", + "content": " of the train" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20334" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 65, + 526, + 222 + ], + "blocks": [ + { + "bbox": [ + 80, + 65, + 526, + 222 + ], + "lines": [ + { + "bbox": [ + 80, + 65, + 526, + 222 + ], + "spans": [ + { + "bbox": [ + 80, + 65, + 526, + 222 + ], + "type": "table", + "html": "
MethodsPresened atSUN RGB-DScanNetV2
mAP@0.25 (%)mAP@0.50 (%)mAP@0.25 (%)mAP@0.50 (%)
2D-GenerieRAND-42.7125.1455.0939.74
MC-DROPOUT [9]ICML'1747.4828.3645.7128.72
CORESET [38]ICLR'1846.9928.1255.3440.08
BADGE [4]ICLR'2049.1230.5457.1041.97
2D-ODCDAL [1]ECCV'2050.8630.7557.9942.33
DIVPROTO [53]CVPR'2252.3132.5459.0143.10
PPAL [59]CVPR'2452.5533.5159.4443.15
3D-ODCRB [25]ICLR'2350.8931.3058.7342.71
KECOR [24]ICCV'2351.0731.4158.5442.56
OURS-56.1135.4261.7745.93
ORACLE†-64.3147.4372.6558.81
", + "image_path": "e0077951725b1c4d58b78bb74c2ef871237634e438de2ca776082a3d9e9d46db.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 223, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 223, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 555, + 258 + ], + "type": "text", + "content": "Table 2. mAP@0.25 (%) and mAP@0.50 (%) scores of the proposed method and AL baselines on the SUN RGB-D and ScanNetV2 benchmarks with " + }, + { + "bbox": [ + 55, + 223, + 555, + 258 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 223, + 555, + 258 + ], + "type": "text", + "content": " queried point clouds. CAGroup3D is used as the detection model for all approaches. " + }, + { + "bbox": [ + 55, + 223, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 223, + 555, + 258 + ], + "type": "text", + "content": ": indicates the detection results using fully-annotated data, and serves as the upper bound for all active learning strategies." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 57, + 267, + 294, + 350 + ], + "blocks": [ + { + "bbox": [ + 57, + 267, + 294, + 350 + ], + "lines": [ + { + "bbox": [ + 57, + 267, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 57, + 267, + 294, + 350 + ], + "type": "image", + "image_path": "4b3db0e785a91ae609319f4207bc12e3e9a37f45af9668c402f6502417cddcbf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 57, + 350, + 294, + 432 + ], + "blocks": [ + { + "bbox": [ + 57, + 350, + 294, + 432 + ], + "lines": [ + { + "bbox": [ + 57, + 350, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 57, + 350, + 294, + 432 + ], + "type": "image", + "image_path": "893b7c2e70ae723b239d06c00a749ac8fd40db81019eeefdaa34a6be217db612.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 445, + 295, + 469 + ], + "lines": [ + { + "bbox": [ + 55, + 445, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 295, + 469 + ], + "type": "text", + "content": "Figure 3. mAP@0.25 (\\%) score of the proposed method and AL baselines on SUN RGB-D and ScanNetV2 benchmarks." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 480, + 295, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 295, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 295, + 528 + ], + "type": "text", + "content": "ing set for both datasets, where the first " + }, + { + "bbox": [ + 55, + 480, + 295, + 528 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 55, + 480, + 295, + 528 + ], + "type": "text", + "content": " samples are selected randomly. In each active learning round, " + }, + { + "bbox": [ + 55, + 480, + 295, + 528 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 55, + 480, + 295, + 528 + ], + "type": "text", + "content": " of the training point clouds are acquired from the rest unlabeled dataset via active learning strategies for annotation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 529, + 295, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 295, + 613 + ], + "type": "text", + "content": "Baselines. We evaluate nine AL baselines: random sampling (RAND), three 2D generic AL strategies (McDROPOUT, CORESET, BADGE) representing uncertainty-based, diversity-based, and hybrid methods, three 2D AL detection strategies (CDAL, DIVPROTO, PPAL) adapted to 3D detection tasks by us, and two state-of-the-art 3D outdoor AL detection strategies (CRB and KECOR)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 623, + 179, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 179, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 179, + 635 + ], + "type": "text", + "content": "4.2. Experimental Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 715 + ], + "type": "text", + "content": "SUN RGB-D. In Fig. 3, we compare our proposed method with different baselines through an increasing percentage of selected point clouds [25, 59]. As demonstrated by the results, our proposed strategy consistently outperforms all competing baselines by significant margins for different percentages of selected point clouds. Notably, with only" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 269, + 555, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 555, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 555, + 412 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 269, + 555, + 412 + ], + "type": "text", + "content": " annotated data, our approach achieves over " + }, + { + "bbox": [ + 313, + 269, + 555, + 412 + ], + "type": "inline_equation", + "content": "85\\%" + }, + { + "bbox": [ + 313, + 269, + 555, + 412 + ], + "type": "text", + "content": " of the fully-annotated performance (red dashed line in Fig. 3). Specifically, as detailed in Tab. 2, it surpasses the state-of-the-art 2D active learning strategy, PPAI, by " + }, + { + "bbox": [ + 313, + 269, + 555, + 412 + ], + "type": "inline_equation", + "content": "3.56\\%" + }, + { + "bbox": [ + 313, + 269, + 555, + 412 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 269, + 555, + 412 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}@0.25" + }, + { + "bbox": [ + 313, + 269, + 555, + 412 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 269, + 555, + 412 + ], + "type": "inline_equation", + "content": "1.91\\%" + }, + { + "bbox": [ + 313, + 269, + 555, + 412 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 269, + 555, + 412 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}@0.50" + }, + { + "bbox": [ + 313, + 269, + 555, + 412 + ], + "type": "text", + "content": ", with a " + }, + { + "bbox": [ + 313, + 269, + 555, + 412 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 269, + 555, + 412 + ], + "type": "text", + "content": " annotation budget. To elaborate further, it is observed that AL methods for 2D object detection generally outperform those for 3D outdoor counterpart. We hypothesize that current 3D outdoor approaches overlook the intra-class variances across different categories, leading to redundancy in selecting active samples. This redundancy may hinder the capability of the learned detection model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 413, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 556, + 617 + ], + "type": "text", + "content": "ScanNetV2. To study the effectiveness of our proposed approach on point clouds with complicated room layouts, we conduct experiments using the ScanNetV2 dataset, a more comprehensive benchmark for indoor environments. The results, illustrated in Fig. 3 and Tab. 2, demonstrate the performance comparison across varying annotation budgets. Our method surpasses all existing AL approaches by a huge margin, achieving approximately " + }, + { + "bbox": [ + 313, + 413, + 556, + 617 + ], + "type": "inline_equation", + "content": "85\\%" + }, + { + "bbox": [ + 313, + 413, + 556, + 617 + ], + "type": "text", + "content": " of the fully-annotated performance with " + }, + { + "bbox": [ + 313, + 413, + 556, + 617 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 413, + 556, + 617 + ], + "type": "text", + "content": " annotation budget. In Fig. 4, we present a qualitative comparison with random sampling. Random sampling captures only partial object characteristics (e.g., 'bookshelf') due to limited training data. In contrast, our approach successfully detects most of the target objects in challenging scenarios by effectively selecting informative unlabeled samples based on uncertainty and diversity criteria. Additional qualitative results can be found in the supplementary materials." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 624, + 408, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 408, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 408, + 636 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": "We ablate some critical designs for an in-depth exploration. Design Choices. We start by exploring the effectiveness of different designs in our proposed method, as shown in Tab. 3. When equipped with the IoU score in Eq. 2 and the uncertainty score for undetections " + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{\\mathrm{undet}}" + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": ", the mAP@0.25 increases by " + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": " on SUN RGB-D respectively," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20335" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 86, + 220, + 159 + ], + "blocks": [ + { + "bbox": [ + 102, + 73, + 171, + 84 + ], + "lines": [ + { + "bbox": [ + 102, + 73, + 171, + 84 + ], + "spans": [ + { + "bbox": [ + 102, + 73, + 171, + 84 + ], + "type": "text", + "content": "(a) Ground Truth" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 86, + 220, + 159 + ], + "lines": [ + { + "bbox": [ + 62, + 86, + 220, + 159 + ], + "spans": [ + { + "bbox": [ + 62, + 86, + 220, + 159 + ], + "type": "image", + "image_path": "51eaf9d3af05c0cd3219f4eeeca08f10a5ff3a2503f573af0c3e0ca42712e519.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 227, + 86, + 383, + 159 + ], + "blocks": [ + { + "bbox": [ + 263, + 73, + 341, + 84 + ], + "lines": [ + { + "bbox": [ + 263, + 73, + 341, + 84 + ], + "spans": [ + { + "bbox": [ + 263, + 73, + 341, + 84 + ], + "type": "text", + "content": "(b) Random (10%)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 227, + 86, + 383, + 159 + ], + "lines": [ + { + "bbox": [ + 227, + 86, + 383, + 159 + ], + "spans": [ + { + "bbox": [ + 227, + 86, + 383, + 159 + ], + "type": "image", + "image_path": "1d46073ab58a5c741e33fc66e70304697c46b4beed11564d1f6bf39200fca216.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 79, + 163, + 529, + 175 + ], + "lines": [ + { + "bbox": [ + 79, + 163, + 529, + 175 + ], + "spans": [ + { + "bbox": [ + 79, + 163, + 529, + 175 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison between random sampling and our proposed active learning method on ScanNetV2 val set." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 391, + 86, + 547, + 159 + ], + "blocks": [ + { + "bbox": [ + 436, + 73, + 500, + 84 + ], + "lines": [ + { + "bbox": [ + 436, + 73, + 500, + 84 + ], + "spans": [ + { + "bbox": [ + 436, + 73, + 500, + 84 + ], + "type": "text", + "content": "(c) Ours (10%)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 86, + 547, + 159 + ], + "lines": [ + { + "bbox": [ + 391, + 86, + 547, + 159 + ], + "spans": [ + { + "bbox": [ + 391, + 86, + 547, + 159 + ], + "type": "image", + "image_path": "65b23165b211db0a2779d1c4e0c12e1d0ac6ad785f86eebff6930c77efe9d9d8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 58, + 184, + 295, + 276 + ], + "blocks": [ + { + "bbox": [ + 58, + 184, + 295, + 276 + ], + "lines": [ + { + "bbox": [ + 58, + 184, + 295, + 276 + ], + "spans": [ + { + "bbox": [ + 58, + 184, + 295, + 276 + ], + "type": "table", + "html": "
UncertaintyDiversitySUN RGB-DScanNetV2
mAP@.25mAP@.50mAP@.25mAP@.50
XX42.725.155.139.7
X✓54.533.958.742.9
w/ IoU score✓55.634.760.843.8
w/ Uundet✓55.234.360.143.3
✓X54.333.558.442.8
✓w/ K-means55.334.160.443.5
✓w/ Greedy-O54.533.358.743.0
✓✓56.135.461.845.9
", + "image_path": "6fac66c582b7b7fc4c7aa165570a4b93e4b74b9de1c47c524d556ae2c180424a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 57, + 301, + 175, + 396 + ], + "blocks": [ + { + "bbox": [ + 57, + 301, + 175, + 396 + ], + "lines": [ + { + "bbox": [ + 57, + 301, + 175, + 396 + ], + "spans": [ + { + "bbox": [ + 57, + 301, + 175, + 396 + ], + "type": "image", + "image_path": "ad46ab98f854f555d1ea5412ba19e53d8a1cbfcf337b53e7a7baae95227a814d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 397, + 295, + 418 + ], + "lines": [ + { + "bbox": [ + 55, + 397, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 295, + 418 + ], + "type": "text", + "content": "Figure 5. Effects of hyper-parameters in our proposed method on SUN RGB-D dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 175, + 301, + 293, + 396 + ], + "blocks": [ + { + "bbox": [ + 175, + 301, + 293, + 396 + ], + "lines": [ + { + "bbox": [ + 175, + 301, + 293, + 396 + ], + "spans": [ + { + "bbox": [ + 175, + 301, + 293, + 396 + ], + "type": "image", + "image_path": "6f98d1a12c2a1327561d2808e3b1504c7d2c3cf47c160036b20fa5d1fe186b8b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 426, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 426, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 54, + 426, + 295, + 593 + ], + "type": "text", + "content": "validating its effectiveness as an uncertainty-based strategy. Additionally, we ablate the diversity branch through two alternatives: (1) Use K-means [52] to select a fixed number of prototypes to replace CAP Bank, denoted as 'K-means'. (2) Apply greedy search to the original optimization problem as defined in Eq. 8, denoted as 'Greedy-O'. It demonstrates that the mAP@0.50 of 'K-means' decreases by " + }, + { + "bbox": [ + 54, + 426, + 295, + 593 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 54, + 426, + 295, + 593 + ], + "type": "text", + "content": " for ScanNetV2 dataset, which suggests that a fixed prototype number leads to a sub-optimal outcome. Furthermore, 'Greedy-O' solution only achieves minor performance improvements. A possible reason is that we restrict the searching space of the original optimization problem (Eq. 8) to several regularized sub-spaces (Eq. 10), eliminating a vast number of sub-optimal solutions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": "Hyper-parameters. In Fig. 5, we report the ablation study of two hyper-parameters in our proposed method, which are the scale " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": " in Eq. 4 and the budget expanding ratio " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": " for determining the candidate pool [59]. Overall, the proposed method is relatively insensitive to the choices of hyperparameters. As depicted in Fig. 5 (a), the best performance is reached with a moderate scale value of " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "k = 3" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": ", which effectively suppresses extreme values while preserving essential information from different sources of uncertainty. Additionally, we observe that excessively large values of " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": " will" + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 315, + 184, + 554, + 235 + ], + "blocks": [ + { + "bbox": [ + 55, + 277, + 295, + 299 + ], + "lines": [ + { + "bbox": [ + 55, + 277, + 295, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 295, + 299 + ], + "type": "text", + "content": "Table 3. Ablation studies of different designs in our proposed method on SUN RGB-D and ScanNetV2 datasets." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 184, + 554, + 235 + ], + "lines": [ + { + "bbox": [ + 315, + 184, + 554, + 235 + ], + "spans": [ + { + "bbox": [ + 315, + 184, + 554, + 235 + ], + "type": "table", + "html": "
MeasuresSUN RGB-DScanNetV2
mAP@.25mAP@.50mAP@.25mAP@.50
Prob. Margin [37]55.835.061.545.5
MC-Dropout [9]56.035.261.645.6
Entropy [39]56.135.461.845.9
", + "image_path": "bb1eb788b13952bbacdf23678562238dff1479ba646220402bf1a6ccb3b5aca8.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 236, + 554, + 258 + ], + "lines": [ + { + "bbox": [ + 313, + 236, + 554, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 554, + 258 + ], + "type": "text", + "content": "Table 4. Ablation studies of different uncertainty measures in our proposed method on SUN RGB-D and ScanNetV2 datasets." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 266, + 555, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 555, + 374 + ], + "type": "text", + "content": "significantly degrade model performance in the early stages, which is shown in Fig. 5 (b). We hypothesize that for big " + }, + { + "bbox": [ + 313, + 266, + 555, + 374 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 266, + 555, + 374 + ], + "type": "text", + "content": ", the candidate pool contains many samples the model is already confident about, thereby harming the quality of the final selected samples. Conversely, when " + }, + { + "bbox": [ + 313, + 266, + 555, + 374 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 266, + 555, + 374 + ], + "type": "text", + "content": " is very small, the selection process is dominated by uncertainty, which restricts the impact of diversity branch. We find that setting " + }, + { + "bbox": [ + 313, + 266, + 555, + 374 + ], + "type": "inline_equation", + "content": "\\delta = 6" + }, + { + "bbox": [ + 313, + 266, + 555, + 374 + ], + "type": "text", + "content": " achieves a better tradeoff between uncertainty- and diversity-based selection, optimizing overall performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 374, + 554, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 554, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 554, + 458 + ], + "type": "text", + "content": "Uncertainty Measures. We replace the Shannon entropy in Eq. 2 with two alternative uncertainty measures: MC-Dropout [9] and probability margin [37], and compare their performance in Tab. 4. The results show that all three measures yield similar performance, with entropy performing slightly better. Therefore, we select entropy as our uncertainty measure, as it is both more efficient and effective." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 469, + 388, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 388, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 388, + 481 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 490, + 555, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 490, + 555, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 555, + 611 + ], + "type": "text", + "content": "This paper presents a novel active learning approach for 3D indoor object detection, improving annotation efficiency by combining uncertainty estimation with diversity-driven sampling. Specifically, our approach estimates epistemic uncertainty, accounting for both inaccurate detections and undetected objects, and selects diverse samples through an optimization method based on a Class-aware Adaptive Prototype bank. Our method significantly outperforms state-of-the-art baselines, achieving over " + }, + { + "bbox": [ + 313, + 490, + 555, + 611 + ], + "type": "inline_equation", + "content": "85\\%" + }, + { + "bbox": [ + 313, + 490, + 555, + 611 + ], + "type": "text", + "content": " of fully-supervised performance with just " + }, + { + "bbox": [ + 313, + 490, + 555, + 611 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 490, + 555, + 611 + ], + "type": "text", + "content": " of the annotation budget." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 621, + 419, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 419, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 419, + 635 + ], + "type": "text", + "content": "6. Acknowledgment" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 641, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 712 + ], + "type": "text", + "content": "This research is supported by the Ministry of Education, Singapore, under its MOE Academic Research Fund Tier 2 (MOE-T2EP20124-0013), it is also supported by the Agency for Science, Technology and Research (A*STAR) under its MTC Programmatic Funds (Grant No. M23L7b0021)." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20336" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Sharat Agarwal, Himanshu Arora, Saket Anand, and Chetan Arora. Contextual diversity for active learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16, pages 137-153. Springer, 2020. 2, 3, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 190 + ], + "type": "text", + "content": "[2] Kelsey Allen, Evan Shelhamer, Hanul Shin, and Joshua Tenenbaum. Infinite mixture prototypes for few-shot learning. In International conference on machine learning, pages 232-241. PMLR, 2019. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 296, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 296, + 245 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 296, + 245 + ], + "type": "text", + "content": "[3] Eduardo Arnold, Omar Y Al-Jarrah, Mehrdad Dianati, Saber Fallah, David Oxtoby, and Alex Mouzakitis. A survey on 3d object detection methods for autonomous driving applications. IEEE Transactions on Intelligent Transportation Systems, 20(10):3782-3795, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 246, + 295, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 246, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 246, + 295, + 290 + ], + "type": "text", + "content": "[4] Jordan T Ash, Chicheng Zhang, Akshay Krishnamurthy, John Langford, and Alekh Agarwal. Deep batch active learning by diverse, uncertain gradient lower bounds. arXiv preprint arXiv:1906.03671, 2019. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 291, + 295, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 295, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 295, + 345 + ], + "type": "text", + "content": "[5] Gui Citovsky, Giulia DeSalvo, Claudio Gentile, Lazaros Karydas, Anand Rajagopalan, Afshin Rostamizadeh, and Sanjiv Kumar. Batch active learning at scale. Advances in Neural Information Processing Systems, 34:11933-11944, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 346, + 295, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 295, + 401 + ], + "type": "text", + "content": "[6] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 2, 4, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 402, + 295, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 402, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 402, + 295, + 446 + ], + "type": "text", + "content": "[7] Ehsan Elhamifar, Guillermo Sapiro, Allen Yang, and S Shankar Sasrty. A convex optimization framework for active learning. In Proceedings of the IEEE International Conference on Computer Vision, pages 209-216, 2013. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 447, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 295, + 491 + ], + "type": "text", + "content": "[8] Di Feng, Xiao Wei, Lars Rosenbaum, Atsuto Maki, and Klaus Dietmayer. Deep active learning for efficient training of a lidar 3d object detector. In 2019 IEEE Intelligent Vehicles Symposium (IV), pages 667-674. IEEE, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 491, + 295, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 295, + 535 + ], + "type": "text", + "content": "[9] Yarin Gal, Riashat Islam, and Zoubin Ghahramani. Deep Bayesian active learning with image data. In Proceedings of the 34th International Conference on Machine Learning, pages 1183-1192. PMLR, 2017. 2, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 536, + 295, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 295, + 579 + ], + "type": "text", + "content": "[10] Mike Gashler and Tony Martinez. Temporal nonlinear dimensionality reduction. In The 2011 International Joint Conference on Neural Networks, pages 1959-1966. IEEE, 2011. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 580, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 295, + 624 + ], + "type": "text", + "content": "[11] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 680 + ], + "type": "text", + "content": "[12] Ahmed Ghita, Björk Antoniussen, Walter Zimmer, Ross Greer, Christian Creß, Andreas Møgelmose, Mohan M Trivedi, and Alois C Knoll. Activeanno3d—an active learning framework for multi-modal 3d object detection. arXiv preprint arXiv:2402.03235, 2024. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[13] JunYoung Gwak, Christopher Choy, and Silvio Savarese. Generative sparse detection networks for 3d single-shot object detection. In Computer Vision-ECCV 2020: 16th Euro-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "pean Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 297-313. Springer, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 555, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 149 + ], + "type": "text", + "content": "[14] Yucheng Han, Na Zhao, Weiling Chen, Keng Teck Ma, and Hanwang Zhang. Dual-perspective knowledge enrichment for semi-supervised 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2049-2057, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 150, + 554, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 554, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 554, + 205 + ], + "type": "text", + "content": "[15] Pengkun Jiao, Na Zhao, Jingjing Chen, and Yu-Gang Jiang. Unlocking textual and visual wisdom: Open-vocabulary 3d object detection enhanced by comprehensive guidance from text and image. In European Conference on Computer Vision, pages 376-392. Springer, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 206, + 554, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 554, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 554, + 249 + ], + "type": "text", + "content": "[16] Ajay J Joshi, Fatih Porikli, and Nikolaos Papanikolopoulos. Multi-class active learning for image classification. In 2009 IEEE conference on computer vision and pattern recognition, pages 2372-2379. IEEE, 2009. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 250, + 554, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 554, + 294 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 554, + 294 + ], + "type": "text", + "content": "[17] Yoon-Yeong Kim, Kyungwoo Song, JoonHo Jang, and Il-Chul Moon. Lada: Look-ahead data acquisition via augmentation for deep active learning. Advances in Neural Information Processing Systems, 34:22919-22930, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 294, + 554, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 554, + 347 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 554, + 347 + ], + "type": "text", + "content": "[18] Seong Min Kye, Kwanghee Choi, Hyeongmin Byun, and Buru Chang. Tidal: Learning training dynamics for active learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22335-22345, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 349, + 554, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 349, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 349, + 554, + 392 + ], + "type": "text", + "content": "[19] Jean Lahoud and Bernard Ghanem. 2d-driven 3d object detection in rgb-d images. In Proceedings of the IEEE international conference on computer vision, pages 4622-4630, 2017. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 393, + 554, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 554, + 449 + ], + "type": "text", + "content": "[20] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 449, + 554, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 554, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 554, + 503 + ], + "type": "text", + "content": "[21] Jinpeng Lin, Zhihao Liang, Shengheng Deng, Lile Cai, Tao Jiang, Tianrui Li, Kui Jia, and Xun Xu. Exploring diversity-based active learning for 3d object detection in autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 504, + 554, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 554, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 554, + 548 + ], + "type": "text", + "content": "[22] Zhuoming Liu, Hao Ding, Huaping Zhong, Weijia Li, Jifeng Dai, and Conghui He. Influence selection for active learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9274-9283, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 548, + 554, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 554, + 570 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 554, + 570 + ], + "type": "text", + "content": "[23] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 571, + 554, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 554, + 624 + ], + "type": "text", + "content": "[24] Yadan Luo, Zhuoxiao Chen, Zhen Fang, Zheng Zhang, Mahsa Baktashmotlagh, and Zi Huang. Kecor: Kernel coding rate maximization for active 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18279-18290, 2023. 2, 3, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 625, + 554, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 554, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 554, + 669 + ], + "type": "text", + "content": "[25] Yadan Luo, Zhuoxiao Chen, Zijian Wang, Xin Yu, Zi Huang, and Mahsa Baktashmotlagh. Exploring active 3d object detection from a generalization perspective. arXiv preprint arXiv:2301.09249, 2023. 2, 3, 4, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "text", + "content": "[26] Yi Ma, Harm Derksen, Wei Hong, and John Wright. Segmentation of multivariate mixed data via lossy data coding and compression. IEEE transactions on pattern analysis and machine intelligence, 29(9):1546-1562, 2007. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20337" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[27] Roberta Maisano, Valeria Tomaselli, Alessandro Capra, Francesco Longo, and Antonio Puliafito. Reducing complexity of 3d indoor object detection. In 2018 IEEE 4th International Forum on Research and Technology for Society and Industry (RTSI), pages 1-6. IEEE, 2018. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 295, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 174 + ], + "type": "text", + "content": "[28] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In 2015 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 922-928. IEEE, 2015. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 217 + ], + "type": "text", + "content": "[29] Michael Meyer and Georg Kuschk. Automotive radar dataset for deep learning based 3d object detection. In 2019 16th European radar conference (EuRAD), pages 129-132. IEEE, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 295, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 295, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 295, + 284 + ], + "type": "text", + "content": "[30] Peng Mi, Jianghang Lin, Yiyi Zhou, Yunhang Shen, Gen Luo, Xiaoshuai Sun, Liujuan Cao, Rongrong Fu, Qiang Xu, and Rongrong Ji. Active teacher for semi-supervised object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14482-14491, 2022. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 295, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 295, + 330 + ], + "type": "text", + "content": "[31] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2906-2917, 2021. 1, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 295, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 295, + 374 + ], + "type": "text", + "content": "[32] Aimee Moses, Srikanth Jakkampudi, Cheryl Danner, and Derek Biega. Localization-based active learning (local) for object detection in 3d point clouds. In Geospatial Informatics XII, pages 44-58. SPIE, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 295, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 295, + 431 + ], + "type": "text", + "content": "[33] Yuzuru Nakamura, Yasunori Ishii, and Takayoshi Yamashita. Active domain adaptation with false negative prediction for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28782-28792, 2024. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 433, + 295, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 295, + 475 + ], + "type": "text", + "content": "[34] Hieu T Nguyen and Arnold Smeulders. Active learning using pre-clustering. In Proceedings of the twenty-first international conference on Machine learning, page 79, 2004. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 477, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 477, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 477, + 295, + 521 + ], + "type": "text", + "content": "[35] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb-d data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 918-927, 2018. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 295, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 295, + 577 + ], + "type": "text", + "content": "[36] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 2, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 295, + 634 + ], + "type": "text", + "content": "[37] Dan Roth and Kevin Small. Margin-based active learning for structured output spaces. In Machine Learning: ECML 2006: 17th European Conference on Machine Learning Berlin, Germany, September 18-22, 2006 Proceedings 17, pages 413-424. Springer, 2006. 2, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 295, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 295, + 677 + ], + "type": "text", + "content": "[38] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. In International Conference on Learning Representations, 2018. 1, 2, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 295, + 712 + ], + "type": "text", + "content": "[39] Claude Elwood Shannon. A mathematical theory of communication. The Bell system technical journal, 27(3):379-423, 1948. 2, 4, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[40] Xiaoke Shen and Ioannis Stamos. Frustum voxelnet for 3d object detection from rgb-d or depth images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1698-1706, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[41] Yichao Shen, Zigang Geng, Yuhui Yuan, Yutong Lin, Ze Liu, Chunyu Wang, Han Hu, Nanning Zheng, and Baining Guo. V-detr: Detr with vertex relative position encoding for 3d object detection. arXiv preprint arXiv:2308.04409, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 162, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 216 + ], + "type": "text", + "content": "[42] Hualian Sheng, Sijia Cai, Na Zhao, Bing Deng, Jianqiang Huang, Xian-Sheng Hua, Min-Jian Zhao, and Gim Hee Lee. Rethinking iou-based optimization for single-stage 3d object detection. In European Conference on Computer Vision, pages 544-561. Springer, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 217, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 270 + ], + "type": "text", + "content": "[43] Hualian Sheng, Sijia Cai, Na Zhao, Bing Deng, Min-Jian Zhao, and Gim Hee Lee. Pdr: Progressive depth regularization for monocular 3d object detection. IEEE Transactions on Circuits and Systems for Video Technology, 33(12):7591-7603, 2023." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 271, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 271, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 271, + 553, + 316 + ], + "type": "text", + "content": "[44] Hualian Sheng, Sijia Cai, Na Zhao, Bing Deng, Qiao Liang, Min-Jian Zhao, and Jieping Ye. Ct3d++: Improving 3d object detection with keypoint-induced channel-wise transformer. International Journal on Computer Vision, 2025. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "type": "text", + "content": "[45] Samarth Sinha, Sayna Ebrahimi, and Trevor Darrell. Variational adversarial active learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5972-5981, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 360, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 553, + 404 + ], + "type": "text", + "content": "[46] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 567-576, 2015. 1, 2, 4, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 404, + 553, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 447 + ], + "type": "text", + "content": "[47] OpenPCDet Development Team. Openpcdet: An opensource toolbox for 3d object detection from point clouds. https://github.com/open-mmlab/OpenPCDet, 2020.6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "type": "text", + "content": "[48] Haiyang Wang, Lihe Ding, Shaocong Dong, Shaoshuai Shi, Aoxue Li, Jianan Li, Zhenguo Li, and Liwei Wang. Cagroup3d: Class-aware grouping for 3d object detection on point clouds. Advances in Neural Information Processing Systems, 35:29975-29988, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 503, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 503, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 503, + 553, + 536 + ], + "type": "text", + "content": "[49] Jiangyi Wang, Zhongyao Cheng, Na Zhao, Jun Cheng, and Xulei Yang. On-the-fly point feature representation for point clouds analysis. arXiv preprint arXiv:2407.21335, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "type": "text", + "content": "[50] Keze Wang, Dongyu Zhang, Ya Li, Ruimao Zhang, and Liang Lin. Cost-effective active learning for deep image classification. IEEE Transactions on Circuits and Systems for Video Technology, 27(12):2591-2600, 2016. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "text", + "content": "[51] Jamie Watson, Mohamed Sayed, Zawar Qureshi, Gabriel J. Brostow, Sara Vicente, Oisin Mac Aodha, and Michael Firman. Virtual occlusions through implicit depth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9053-9064, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 636, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 658 + ], + "type": "text", + "content": "[52] Junjie Wu. Advances in " + }, + { + "bbox": [ + 316, + 636, + 553, + 658 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 316, + 636, + 553, + 658 + ], + "type": "text", + "content": "-means clustering: a data mining thinking. Springer Science & Business Media, 2012. 2, 6, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[53] Jiaxi Wu, Jiaxin Chen, and Di Huang. Entropy-based active learning for object detection with progressive diversity constraint. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9397-9406, 2022. 3, 4, 7" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20338" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 139 + ], + "type": "text", + "content": "[54] Tsung-Han Wu, Yueh-Cheng Liu, Yu-Kai Huang, Hsin-Ying Lee, Hung-Ting Su, Ping-Chia Huang, and Winston H Hsu. Redal: Region-based and diversity-aware active learning for point cloud semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15510-15519, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 296, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 296, + 194 + ], + "type": "text", + "content": "[55] Qian Xie, Yu-Kun Lai, Jing Wu, Zhoutao Wang, Yiming Zhang, Kai Xu, and Jun Wang. Mlcvnet: Multi-level context votenet for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10447-10456, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 195, + 296, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 296, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 296, + 248 + ], + "type": "text", + "content": "[56] Qian Xie, Yu-Kun Lai, Jing Wu, Zhoutao Wang, Dening Lu, Mingqiang Wei, and Jun Wang. Venet: Voting enhancement network for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3712-3721, 2021. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 250, + 294, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 294 + ], + "type": "text", + "content": "[57] Yating Xu, Conghui Hu, Na Zhao, and Gim Hee Lee. Generalized few-shot point cloud segmentation via geometric words. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21506-21515, 2023. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 294, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 369 + ], + "type": "text", + "content": "[58] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17830-17839, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "type": "text", + "content": "[59] Chenhongyi Yang, Lichao Huang, and Elliot J Crowley. Plug and play active learning for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17784-17793, 2024. 3, 4, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 415, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 459 + ], + "type": "text", + "content": "[60] Yi Yang, Zhigang Ma, Feiping Nie, Xiaojun Chang, and Alexander G Hauptmann. Multi-class active learning by uncertainty sampling with diversity maximization. International Journal of Computer Vision, 113:113-127, 2015. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 459, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 459, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 459, + 294, + 514 + ], + "type": "text", + "content": "[61] Tianning Yuan, Fang Wan, Mengying Fu, Jianzhuang Liu, Songcen Xu, Xiangyang Ji, and Qixiang Ye. Multiple instance active learning for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5330-5339, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 514, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 557 + ], + "type": "text", + "content": "[62] Bohang Zhang, Jikai Jin, Cong Fang, and Liwei Wang. Improved analysis of clipping algorithms for non-convex optimization. Advances in Neural Information Processing Systems, 33:15511-15521, 2020. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "type": "text", + "content": "[63] Na Zhao and Gim Hee Lee. Static-dynamic co-teaching for class-incremental 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3436-3445, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 602, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 647 + ], + "type": "text", + "content": "[64] Na Zhao, Tat-Seng Chua, and Gim Hee Lee. Sess: Self-ensembling semi-supervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11079–11087, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 691 + ], + "type": "text", + "content": "[65] Na Zhao, Tat-Seng Chua, and Gim Hee Lee. Few-shot 3d point cloud semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8873-8882, 2021. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[66] Na Zhao, Peisheng Qian, Fang Wu, Xun Xu, Xulei Yang, and Gim Hee Lee. Sdcot++: Improved static-dynamic co" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "teaching for class-incremental 3d object detection. IEEE Transactions on Image Processing, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "text", + "content": "[67] Qiang Zhou and Chaohui Yu. Point rcnn: An angle-free framework for rotated object detection. Remote Sensing, 14 (11):2605, 2022. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20339" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Uncertainty Weighted Gradients for Model Calibration/792a362b-85b5-4092-b750-5040ac391b64_content_list.json b/2025/Uncertainty Weighted Gradients for Model Calibration/792a362b-85b5-4092-b750-5040ac391b64_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d1b662d8a9af46d1af28f15eece8d4b0d315863d --- /dev/null +++ b/2025/Uncertainty Weighted Gradients for Model Calibration/792a362b-85b5-4092-b750-5040ac391b64_content_list.json @@ -0,0 +1,1724 @@ +[ + { + "type": "text", + "text": "Uncertainty Weighted Gradients for Model Calibration", + "text_level": 1, + "bbox": [ + 218, + 130, + 779, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinxu Lin $^{1}$ Linwei Tao $^{1}$ Minjing Dong $^{2}$ Chang Xu $^{1}$ $^{1}$ The University of Sydney, $^{2}$ City University of Hong Kong \n{jinxu.lin, linwei.tao, c.xu}@sydney.edu.au, minjdong@cityu.edu.hk", + "bbox": [ + 205, + 179, + 787, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 328, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Model calibration is essential for ensuring that the predictions of deep neural networks accurately reflect true probabilities in real-world classification tasks. However, deep networks often produce over-confident or under-confident predictions, leading to miscalibration. Various methods have been proposed to address this issue by designing effective loss functions for calibration, such as focal loss. In this paper, we analyze its effectiveness and provide a unified loss framework of focal loss and its variants, where we mainly attribute their superiority in model calibration to the loss weighting factor that estimates sample-wise uncertainty. Based on our analysis, existing loss functions fail to achieve optimal calibration performance due to two main issues: including misalignment during optimization and insufficient precision in uncertainty estimation. Specifically, focal loss cannot align sample uncertainty with gradient scaling and the single logit cannot indicate the uncertainty. To address these issues, we reformulate the optimization from the perspective of gradients, which focuses on uncertain samples. Meanwhile, we propose using the Brier Score as the loss weight factor, which provides a more accurate uncertainty estimation via all the logits. Extensive experiments on various models and datasets demonstrate that our method achieves state-of-the-art (SOTA) performance. $^{1}$", + "bbox": [ + 86, + 301, + 485, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 695, + 223, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep Neural Networks (DNNs) have achieved remarkable success in various domains, including image classification. However, recent studies [5, 30] reveal that DNNs often suffer from mis-calibration in classification task, exhibiting over-confidence or under-confidence in their predictions. For example, a model might output a confidence score of 0.8 for a particular prediction, which does not necessarily correspond to an $80\\%$ probability of correctness.", + "bbox": [ + 89, + 720, + 482, + 842 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Calibrating DNNs, which aligns the predicted confidence with the true probabilities, is therefore crucial to en", + "bbox": [ + 89, + 843, + 482, + 873 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hancing their reliability in practical applications, such as autonomous driving, medical imaging, and weather forecasting. To address mis-calibration issues, several methods [26, 28] have been proposed, many of which focus on modifying the loss function during training. A common approach involves adding regularization terms with the Cross-Entropy (CE) loss to improve calibration, as seen in methods like Maximum Mean Calibration Error (MMCE) [15] and Meta Calibration [1]. Additionally, Focal Loss (FL) [17], which adjusts per-sample loss weights based on prediction difficulty, has been shown to improve calibration performance. Mukhoti et al. [18] attribute the effectiveness of FL to its implicit regularization on the entropy of predicted probabilities, which mitigates overconfidence. Similarly, Dual Focal Loss (DFL) [27] incorporates the second-most-probable class in the weighting mechanism to address model under-confidence. Another research interest regarding calibration is the evaluation of calibration performance. Metrics commonly employed include Expected Calibration Error (ECE)[5] and Brier Score[2].", + "bbox": [ + 511, + 268, + 906, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The focal-loss-based methods can be unified under a general loss framework expressed as $u \\cdot \\mathrm{CE}$ , where $u$ is a loss weighting factor. In FL, the factor $u_{\\mathrm{FL}}$ can be treated as an estimation for sample-wise difficulty. Interestingly, we observe that $u_{\\mathrm{FL}}$ , originally designed for binary classification, is mathematically equivalent to the Brier Score for binary cases, which can measure the uncertainty. Furthermore, the weighting term $u_{\\mathrm{DFL}}$ in DFL aligns with the Brier Score in three-class classification scenarios. This observation motivates the hypothesis that weighting the loss with sample calibration metrics, such as the Brier Score, could better identify uncertain samples and facilitate targeted training.", + "bbox": [ + 511, + 577, + 908, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Furthermore, our investigation reveals a limitation of directly applying the weighting factor $u$ to the loss term. Specifically, the vanilla optimization of focal loss would achieve a misalignment with the objective of scaling gradients for harder samples with larger magnitudes. Based on our analysis, we mainly attribute this issue to a differentiable loss weighting factor $u$ , which could disrupt the positive correlation between the CE and its gradient magnitude during backpropagation, impeding the model's ability", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Code is available at https://github.com/Jinxu-Lin/BSCE-GRA.", + "bbox": [ + 107, + 886, + 444, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15497", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to prioritize higher-uncertain samples effectively.", + "bbox": [ + 89, + 90, + 415, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these issues, we first propose applying the weighting factor to scaling the gradient rather than the loss itself. This allows us directly aligns the gradient optimization with sample uncertainty, which ensures that harder samples receive appropriately scaled updates without disrupting the optimization process. Within this framework, we then introduce a generalized form of the Brier Score as a gradient weighting factor, which provides a more accurate estimation of sample uncertainty that considers all categories. This leads to the development of the BSCE-GRA loss function, which adjusts gradient to directly scale optimization based on sample uncertainty by Brier Score.", + "bbox": [ + 89, + 106, + 480, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct extensive experiments to validate the effectiveness of the proposed method, achieving state-of-the-art results across various datasets and model architectures, which demonstrate the effectiveness of our methods. The contributions of this paper are summarized as:", + "bbox": [ + 89, + 287, + 480, + 362 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We provide a new perspective by unifying some existing loss modification techniques for model calibration under a sample-weighting framework. With this framework, we provide extensive analysis of their limitations.", + "2. We propose a simple yet effective optimization framework for model calibration with a gradient weighting factor, where we scale the gradients to encourage the model to focus on uncertain samples effectively.", + "3. We analyze the use of different uncertainty metrics within this framework and introduce BSCE-GRA, a new loss function based on the Brier Score that provides an accurate sample-wise uncertainty estimation.", + "4. We conduct extensive experiments under different settings to validate our proposed methods. Our uncertainty-weighted framework shows consistent effectiveness across various uncertainty metrics, and the proposed BSCE-GRA loss achieves state-of-the-art results." + ], + "bbox": [ + 89, + 364, + 482, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 633, + 240, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, numerous techniques have been proposed to address the problem of network miscalibration, which can generally be classified into three categories.", + "bbox": [ + 89, + 659, + 482, + 704 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The first category is post-hoc calibration techniques, which adjust model predictions after training by optimizing additional parameters on a held-out validation set. These methods include Platt Scaling [23], which performs a linear transformation on the original prediction logits; Isotonic Regression [34], which uses piecewise functions to transform logits; Bayesian Binning into Quantiles (BBQ)[20], which extends histogram binning with Bayesian model averaging; and Beta Calibration[13], initially designed for binary classification and later generalized to multi-class settings with Dirichlet distributions by Kull et al. [14]. Temperature Scaling [5], one of the most widely used post-hoc calibration methods, tunes the temperature parameter in the", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SoftMax function to minimize negative log-likelihood on a held-out validation set. In this work, we report calibration performance with post-temperature scaling results.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The second category includes regularization techniques, which are known to effectively calibrate DNNs. Data augmentation methods, such as Mixup [31] and AugMix [7], train DNNs on mixed samples to mitigate overconfident predictions. Model ensemble techniques, which involve independently training multiple DNNs and averaging their predictions, have been shown to enhance both accuracy and predictive uncertainty by aggregating outputs from multiple models [16, 24, 36]. Label smoothing [19], which replaces one-hot labels with soft labels, encourages the model to make less confident predictions, thereby reducing overconfidence. Additionally, weight decay has also been demonstrated to improve confidence calibration [5].", + "bbox": [ + 511, + 137, + 903, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The third category focuses on modifying the training loss to improve calibration. These methods include adding a differentiable auxiliary surrogate loss for Expected Calibration Error (ECE)[1, 10, 11] or replacing the training loss with other loss functions, such as Mean Squared Error (MSE)[9], Focal Loss [18], Inverse Focal Loss [32], and Dual Focal Loss [27]. Among these, Focal Loss [18], which adds a modulation term to the Cross-Entropy loss to focus on hard-to-classify samples, provides a simple and effective way to train calibrated models. Focal Loss and Dual Focal Loss can be categorized as margin-based losses, similar to Hinge Loss [3], Triplet Loss [25], and Margin Ranking Loss [33].", + "bbox": [ + 511, + 335, + 903, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 511, + 532, + 638, + 550 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 511, + 559, + 717, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given a $K$ -class dataset $\\mathbb{S} = \\{\\pmb{z}^{(1)},\\dots,\\pmb{z}^{(N)}\\}$ , where each training sample $\\pmb{z}^{(n)}\\coloneqq (x^{(n)},y^{(n)})$ is an input-label pair. Let $\\mathcal{X}$ and $\\mathcal{V}$ represent the input space and the label space, respectively. The ground truth label $y\\in \\mathcal{V}$ is encoded in a one-hot vector format, where $y_{i} = 1$ if $i\\in K$ represents the actual class. We define a classifier $f_{\\theta}$ trained on $\\mathbb{S}$ that maps an input $\\pmb{x}\\in \\mathcal{X}$ to a probability distribution $\\hat{p} (\\pmb {x})$ . The classifier then provides a prediction $k = \\arg \\max_{i\\in K}\\hat{p}_i$ , which indicates the index of the predicted label. The predicted label $\\hat{y}\\in \\mathcal{Y}$ is similarly represented as $y$ in the one-hot encoding of prediction $k$ . The confidence $\\hat{p}_c$ is defined as the predicted probability associated with the prediction $k$ .", + "bbox": [ + 511, + 580, + 903, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A well-calibrated model ensures that the provided confidence $\\hat{p}_c$ accurately reflects the true probability of correct classification. We define the true class-posterior probability vector as $\\boldsymbol{\\eta}(\\boldsymbol{x}) = [\\eta_1(\\boldsymbol{x}),\\dots,\\eta_K(\\boldsymbol{x})]$ , where $\\eta_k(\\boldsymbol{x}) = \\mathbb{P}(y = k|\\boldsymbol{x})$ is the true probability of class $k$ given input $\\boldsymbol{x}$ . Formally, a network is considered perfectly calibrated if $\\mathbb{P}(\\hat{y} = y|\\hat{p}_c = p) = p$ for all $p\\in [0,1]$ [5], which can be also written as $\\hat{p} (\\pmb {x}) = \\pmb {\\eta}(\\pmb {x})$ . The uncertainty of a model for a sample $\\pmb{x}$ , also referred to the calibrated error $c(x)$", + "bbox": [ + 511, + 763, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15498", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "can be computed as the difference between $\\pmb{\\eta}(\\pmb{x})$ and $\\hat{p}(\\pmb{x})$ :", + "bbox": [ + 89, + 90, + 482, + 107 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nc (\\boldsymbol {x}) = \\| \\eta (\\boldsymbol {x}) - \\hat {p} _ {c} (\\boldsymbol {x}) \\|. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 119, + 482, + 138 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Metric for Calibrated Error", + "text_level": 1, + "bbox": [ + 89, + 145, + 344, + 160 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In practice, it is hard to access the ground truth probability $\\pmb{\\eta}(\\pmb{x})$ on real-world samples and thus the calibration error cannot be directly computed by Eq. 1, some alternative methods have been proposed to evaluate the uncertainty.", + "bbox": [ + 89, + 167, + 482, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ECE. The Expected Calibration Error (ECE) is defined as $\\mathbb{E}_{\\hat{p}_c}[\\| \\mathbb{P}(\\hat{y} = y|\\hat{p}_c) - \\hat{p}_c]$ . Guo et al. [5] propose an approximation of ECE. Specifically, all samples are divided into $M$ bins $\\{B_m\\}_{m=1}^M$ of equal width based on their confidence, where each bin $B_m$ contains all samples whose confidences fall within the range $\\hat{p}_c \\in \\left[\\frac{m}{M}, \\frac{m+1}{M}\\right)$ . For each bin $B_m$ , the average confidence is computed as $C_m = \\frac{1}{|B_m|} \\sum_{i \\in B_m} \\hat{p}_c^{(i)}$ and the bin accuracy is computed as $A_m = \\frac{1}{|B_m|} \\sum_{i \\in B_m} \\mathbb{1}(\\hat{y}_k^{(i)} = y_k^{(i)})$ , where $\\mathbb{1}$ is the indicator function. The ECE can then be computed as the average $L_1$ difference between bin accuracy and confidence:", + "bbox": [ + 89, + 228, + 483, + 402 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {E C E} = \\sum_ {m = 1} ^ {M} \\frac {\\left| B _ {m} \\right|}{N} \\left| A _ {m} - C _ {m} \\right|, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 414, + 482, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $N$ denotes the number of samples in each bin. In addition to the ECE in Eq.2, there are several variants used to measure calibration error. For example, AdaECE [21] groups samples into bins $B_{m}$ with an equal number of smaples such that $|B_m| = |B_n|$ for all bins. Another variant, ClasswiesECE [14] measure the calibration error by considering each of the $K$ classes separately.", + "bbox": [ + 89, + 460, + 483, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Brier Score. Besides ECE computing the expectation of calibrated error among whole datasets, some other techniques have also been used to evaluate the sample-wise uncertainty. Accuracy and calibration are distinct concepts, and one cannot be inferred from the other unequivocally. The Brier Score (BS) [2] unify these two concepts, which is commonly used in calibration literature. It has been shown that this family of metrics can be decomposed into a calibration term and a refinement term. Achieving an optimal score requires both accurate predictions and appropriate confidence levels. For a given sample, the Brier Score is mathematically defined as the Mean Squared Error between the predicted probability distribution $\\hat{p}$ and one-hot encoded ground truth label $y$ . We introduce a generalized Brier Score (gBS) form as follows:", + "bbox": [ + 89, + 566, + 483, + 792 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {g B S} = \\sum_ {i = 1} ^ {K} \\| \\hat {p} _ {i} (\\boldsymbol {x}) - y _ {i} \\| _ {\\beta} ^ {\\gamma} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 803, + 482, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\gamma$ and $\\beta$ are the hyperparameters for the exponent and norm order, respectively. When $\\gamma = 2$ and $\\beta = 2$ , this formulation reduces to the original Brier Score.", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Calibrating Method", + "text_level": 1, + "bbox": [ + 513, + 90, + 705, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Temperature Scaling A widely used post-hoc technique for improving classification calibration is temperature scaling. It adjusts the sharpness of the output probability distribution by scaling the logits in the SoftMax function using a temperature parameter, defined as $\\hat{p}_i = \\frac{\\exp(\\hat{g}_i / T)}{\\sum_{k=1}^{K} \\exp(\\hat{g}_k / T)}$ , where $\\hat{g}$ represents the logits before applying the SoftMax function, and $T$ is the temperature that controls the scaling. Calibration performance can be enhanced by tuning $T$ on a held-out validation set.", + "bbox": [ + 511, + 112, + 906, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Focal Loss and Dual Focal Loss Some other previous works have attempted to improve model calibration by modifying the loss function. Focal Loss [17] was initially introduced to address the foreground-background imbalance problem in object detection. It addresses the issue by incorporating a loss weighting factor based on sample complexity, reducing the weight of easy samples and allowing the model to focus on harder-to-classify instances.", + "bbox": [ + 511, + 253, + 905, + 373 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, given the predicted probability $\\hat{p} (\\pmb {x})$ on sample $\\pmb{x}$ , the focal loss is defined as:", + "bbox": [ + 511, + 375, + 905, + 402 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {F L}} (\\boldsymbol {x}, y) = - \\sum_ {i = 1} ^ {K} y _ {i} (1 - \\hat {p} _ {i}) ^ {\\gamma} \\log \\hat {p} _ {i} (\\boldsymbol {x}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 412, + 905, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\gamma$ is a pre-defined hyperparameter. Previous studies [18] have demonstrated that optimizing models using Focal Loss results in better calibration compared to Cross-Entropy. This improvement is partly due to the entropy-based regularization effect introduced by Focal Loss, while complexity-based weighting also likely plays a role. However, we noticed that Focal Loss was initially applied to binary classification, where the complexity term $(1 - \\hat{p}_i)$ is related to the Brier Score, which can also be used for evaluating uncertainty. In multi-class classification, although complexity and uncertainty are not strictly equivalent, the complexity term still captures some class-level uncertainty.", + "bbox": [ + 511, + 463, + 905, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Several variants of Focal Loss exist, including Dual Focal Loss [27], which incorporates the probability from the second most probable class into the scaling factor to address the under-confidence issue caused by Focal Loss:", + "bbox": [ + 511, + 643, + 905, + 704 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D F L}} (\\boldsymbol {x}, y) = - \\sum_ {i = 1} ^ {K} y _ {i} (1 - \\hat {p} _ {i} (\\boldsymbol {x}) + \\hat {p} _ {j} (\\boldsymbol {x})) ^ {\\gamma} \\log \\hat {p} _ {i} (\\boldsymbol {x}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 714, + 906, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{p}_j(x) = \\max_i\\{\\hat{p}_i(x)|\\hat{p}_i(x) < \\hat{p}_{gt}(x)\\}$", + "bbox": [ + 511, + 763, + 816, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 513, + 792, + 604, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Weighting with Sample-wise Uncertainty", + "text_level": 1, + "bbox": [ + 511, + 816, + 862, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Focal Loss and Dual Focal Loss both support the concept that uncertainty-based weighting in conjunction with Cross Entropy can enhance model calibration. We propose a generalized loss function framework, termed Uncertainty CE", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15499", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Loss, which incorporates a sample-wise uncertainty metric into Cross Entropy as a weighting factor:", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {U n c e r t a i n t y}} (\\boldsymbol {x}, y) = - \\sum_ {i = 1} ^ {K} u \\left(\\hat {p} _ {i} (\\boldsymbol {x})\\right) \\cdot y _ {i} \\log \\hat {p} _ {i} (\\boldsymbol {x}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 132, + 482, + 172 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $u(\\hat{p}_i)$ is the adaptive term that evaluates the uncertainty of sample $x$ . The motivation behind this design is to use the predicted calibration error for each sample to scale the loss function, thereby directing the model's optimization more effectively towards samples with higher uncertainty.", + "bbox": [ + 89, + 184, + 483, + 260 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, the weights used in Focal Loss is given by:", + "bbox": [ + 109, + 260, + 467, + 275 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nu _ {\\mathrm {F L}} (\\hat {p} (\\boldsymbol {x})) = (1 - \\hat {p} _ {c}) ^ {\\gamma}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 286, + 482, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "And the scaling term in Dual Focal Loss is defined as:", + "bbox": [ + 89, + 314, + 449, + 329 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nu _ {\\mathrm {D F L}} (\\hat {p} (\\boldsymbol {x})) = (1 - \\hat {p} _ {c} (x) + \\hat {p} _ {j} (x)) ^ {\\gamma}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 340, + 482, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{p}_j(x)$ is the second maximum confidence in prediction. The weight used in Focal Loss can be seen as derived from the Brier Score, focusing on the error of the ground true class. In contrast, the scaling term in Dual Focal Loss incorporates uncertainty information from the second most probable class. However, directly weighting loss function with certain uncertainty metrics, as those used in Focal Loss and Dual Focal Loss, presents several challenges.", + "bbox": [ + 89, + 369, + 483, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reviewing our objectives, we aim to weight the loss function value by each sample's uncertainty, resulting in greater optimization steps for those samples with higher uncertainty. Our focus actually is not on the loss value itself, but on the optimization. The purpose of uncertainty weighting is to focus the model's attention on samples with higher uncertainty. Thus, the key factor is the gradient: higher weights should lead to larger gradients, resulting in more substantial model updates for those samples. For Cross Entropy, its value is positively correlated with the gradient: a higher CE value yields a larger gradient, leading to more significant updates for optimization. When applying a scalar weight to CE, this relationship remains intact, as it does not disrupt the positive correlation.", + "bbox": [ + 89, + 489, + 483, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, for differentiable weights, directly applying them to the loss can disrupt the positive correlation between the loss and the gradient, ultimately impairing effective model optimization. To further analyze this, let us consider Focal Loss. For Focal Loss $\\mathcal{L}_{\\mathrm{FL}}$ and Cross Entropy $\\mathcal{L}_{\\mathrm{CE}}$ , their gradients are given by: $\\frac{\\partial}{\\partial w}\\mathcal{L}_{\\mathrm{FL}} = g(\\hat{p}_i,\\gamma)\\frac{\\partial}{\\partial w}\\mathcal{L}_{\\mathrm{CE}}$ , where $g(p,\\gamma) = (1 - p)^{\\gamma} - \\gamma p(1 - p)^{\\gamma -1}\\log (p)$ , $\\gamma$ is a predefined hyperparameter, and $\\boldsymbol{w}$ represents the parameters of the final linear layer. From a gradient perspective, $g(p,\\gamma)$ acts as a weight on the CE gradient, which is illustrated in Figure 1. There exists a point $p_0\\in [0,1]$ such that within the range $[0,p_0],\\frac{\\partial}{\\partial p} g(p,\\gamma) > 0$ and within the range $[p_0,1],\\frac{\\partial}{\\partial p} g(p,\\gamma) < 0$ . In binary classification, $1 - \\hat{p}_i$ is related to", + "bbox": [ + 89, + 700, + 483, + 904 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3b60aa11e502d750f2ad60558338e62d216fc72b2bec4d7340aa3b64027b7fd1.jpg", + "image_caption": [ + "Plot of $g(p)$ for Different $\\gamma$ Values in Focal Loss", + "Figure 1. $g(p, \\gamma)$ of Focal Loss vs predicted confidence $\\hat{p}_c$ ." + ], + "image_footnote": [], + "bbox": [ + 517, + 102, + 903, + 266 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the Brier Score that reveals calibrated error, indicating that the predicted calibrated error is not always positively correlated with the weight on gradient. As uncertainty decreases, the weight on gradient initially increases, before eventually aligning with the level of uncertainty. This causes the model to focus more on samples with moderate uncertainty rather than those with the highest uncertainty, which contradicts our original goal of emphasizing the most uncertain samples. Further, changes in uncertainty are not immediately reflected in the sample weights; instead, multiple training iterations are required for these changes to align properly. Thus, directly applying some kind of uncertainty metrics into Eq. 6 would pose issues to the uncertainty weighting.", + "bbox": [ + 511, + 316, + 906, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Sample-wise Uncertainty on Gradients", + "text_level": 1, + "bbox": [ + 511, + 522, + 846, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The discussion in Sec. 4.1 motivates us to ensure the weight on gradients for a sample to align with its uncertainty. This could be intuitively solved by directly applying the model uncertainty on samples as a weight of the gradient. We define the modified gradient as:", + "bbox": [ + 511, + 544, + 905, + 619 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial}{\\partial \\theta} \\mathcal {L} _ {\\text {U n c e r t a i n t y - G R A}} (\\boldsymbol {x}, y) = u (\\hat {p} (\\boldsymbol {x})) \\frac {\\partial}{\\partial \\theta} \\mathcal {L} _ {\\mathrm {C E}} (\\boldsymbol {x}, y). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 627, + 903, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Taking a SGD optimizer as an example, the model update can be expressed as: $\\theta_{t + 1} = \\theta +\\alpha \\cdot \\frac{\\partial}{\\partial\\theta}\\mathcal{L}_{\\mathrm{Uncertainty - GRA}}(\\boldsymbol {x},y)$ where $\\theta$ is the model parameters and $\\alpha$ is the learning rate. When the predicted uncertainty on sample $\\pmb{x}$ is higher, the optimization step results in a larger update, encouraging the model to focus more on uncertain samples. A general form of the loss function for our uncertainty-weighted gradient framework, termed Uncertainty-GRA CE Loss, is defined as:", + "bbox": [ + 511, + 664, + 906, + 797 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {U n c e r t a i n t y - G R A}} (\\boldsymbol {x}, y) = - \\int \\sum_ {i = 1} ^ {K} u (\\hat {p} (\\boldsymbol {x})) \\cdot \\frac {y _ {i}}{\\hat {p} _ {i} (\\boldsymbol {x})} \\mathrm {d} \\hat {p} (\\boldsymbol {x}), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 806, + 903, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practice, to compute the $\\mathcal{L}_{\\mathrm{Uncertainty - GRA}}$ , we can detach the gradient of $u(\\hat{p} (\\pmb {x}))$ and multiply it with the Cross Entropy, instead of calculating the gradient integration.", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15500", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6924a014be14d1c385e5f18451224290bdbdf369c7df357b0a3ff91bfd7eb54d.jpg", + "image_caption": [ + "(a) An illustration of the $u_{\\mathrm{FL}}$" + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 321, + 244 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/452fd0f06819ecff3714e8a32ff5da4ade46079b2ab28b66986f13ffe461ca0b.jpg", + "image_caption": [ + "(b) An illustration of the $u_{\\mathrm{DFL}}$ .", + "Figure 2. An illustration of value of gradient weight function on a 4 class classification. It is obvious that $u_{\\mathrm{FL}}$ varies only along the $p_i$ axis and $u_{\\mathrm{DFL}}$ changes along the $p_i$ and $p_j$ axes. $u_{\\mathrm{BS}}$ responds to changes across all axes, providing a more complete uncertainty evaluation." + ], + "image_footnote": [], + "bbox": [ + 370, + 85, + 604, + 244 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/50d805ec096e62c5b10963250182e4b37bf2a9c9de00e3a8313ea604220013c0.jpg", + "image_caption": [ + "(c) An illustration of the $u_{\\mathrm{BS}}$" + ], + "image_footnote": [], + "bbox": [ + 656, + 85, + 885, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Sample-wise Uncertainty Metrics", + "text_level": 1, + "bbox": [ + 89, + 327, + 382, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With Eq. 10, calibration can be achieved by employing the sample's ground-truth uncertainty as the scaling term. However, accessing the ground-truth uncertainty for real-world image samples is impractical. Therefore, alternative metrics are required to estimate the sample-wise uncertainty $u(x)$ .", + "bbox": [ + 89, + 349, + 483, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "One approach to evaluate sample-wise uncertainty is based on computing the generalized Brier Score (gBS):", + "bbox": [ + 89, + 426, + 483, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nu _ {\\mathrm {g B S}} (\\hat {p} (\\boldsymbol {x})) = \\sum_ {i = 1} ^ {K} \\| \\hat {p} _ {i} (\\boldsymbol {x}) - y _ {i} \\| _ {\\beta} ^ {\\gamma}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 465, + 482, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We provide a brief discussion on the effectiveness of the generalized Brier Score (gBS) as a measure of calibration error. When using the Brier Score (BS) to evaluate calibration error, with $\\beta = 2$ and $\\gamma = 2$ , the difference between the expected and predicted calibration error is given by:", + "bbox": [ + 89, + 513, + 483, + 590 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} c (\\boldsymbol {x}) - u _ {\\mathrm {B S}} (\\hat {p} (\\boldsymbol {x})) = \\mathbb {E} _ {y \\sim \\eta (x)} [ \\| \\hat {p} (\\boldsymbol {x}) - \\boldsymbol {\\eta} (\\boldsymbol {x}) \\| _ {2} ^ {2} - \\| \\hat {p} (\\boldsymbol {x}) - y \\| _ {2} ^ {2} ] \\\\ = \\sum_ {k = 1} ^ {K} \\boldsymbol {\\eta} _ {k} (\\boldsymbol {x}) \\left(\\boldsymbol {\\eta} _ {k} (\\boldsymbol {x}) - 1\\right). \\tag {12} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 599, + 506, + 661 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thus, the term $\\sum_{k=1}^{K} \\eta_k(x) (\\eta_k(x) - 1)$ depends solely on $\\eta(x)$ which is fixed for a given sample $x$ . Consequently, the error remains constant for a specific sample $x$ . We provide a detailed computation of Eq. 12 in Appendix.", + "bbox": [ + 89, + 672, + 482, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Several variants of $u_{\\mathrm{gBS}}(\\hat{p})$ have demonstrated their effectiveness in previous studies. When $\\beta = 1$ and only the actual class is considered, the generalized Brier Score can be interpreted as the scaling term in Focal Loss in Eq. 7. When $\\beta = 1$ and both the actual class and the maximum predicted class (excluding the actual class) are considered, the generalized Brier Score can be interpreted as the scaling term in Dual Focal Loss in Eq. 8.", + "bbox": [ + 89, + 734, + 482, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, these two metrics only consider several classes in the predicted probability. In the case of Focal Loss, which was originally designed for binary classification,", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "considering one-dimensional uncertainty is sufficient due to the constraint that the sum of probabilities must equal 1. When extended to multi-class classification, these metrics are not accurate enough to evaluate uncertainty.", + "bbox": [ + 511, + 328, + 903, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We visualize the gradient weights $u_{\\mathrm{FL}}$ , $u_{\\mathrm{DFL}}$ and $u_{\\mathrm{BS}}$ in Figure 2 for a 4-class scenario. The three axes represent three dimensions of the predicted probability, while the fourth is implied by the probability sum constraint. From the figure, it is evident that $u_{\\mathrm{FL}}$ and $u_{\\mathrm{DFL}}$ are sensitive only to changes in one or two dimensions. However, different points in the coordinate system have different uncertainties. They fail to adequately capture uncertainty change across all dimensions. But the $u_{\\mathrm{BS}}$ can accurately respond to changes along any coordinate axis.", + "bbox": [ + 511, + 388, + 905, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further evaluate whether these metrics accurately measure the ground truth uncertainty of samples, we conducted experiments on a toy dataset. The toy dataset consists of 5 two-dimensional Gaussian distributions, representing 5 groups of data: $\\mathcal{N}(\\mu_i,\\Sigma),i\\in \\{0,\\dots ,5\\}$ . The mean vectors $\\mu_{i}$ were randomly sampled from the range $[-10,10]$ and $\\Sigma = I$ was used as the shared covariance matrix for all groups. We generated 10,000 data points from each group to form the training dataset, which was used to train a two layer CNN model for 5 epochs. An additional 1,000 samples from each distribution were used to create the test dataset. Therefore, the ground truth probability $\\eta (\\pmb {x})$ can be computed using the probability density function (PDF) of each distribution:", + "bbox": [ + 511, + 539, + 906, + 750 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\eta (\\boldsymbol {x}) = \\frac {p ^ {n} (\\boldsymbol {x})}{\\sum_ {i} ^ {N} p ^ {i} (\\boldsymbol {x})}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 762, + 903, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N = 5$ in this case, $p^n (\\pmb {x})$ and $p^i (\\pmb {x})$ are the PDF of corresponds class for $\\pmb{x}$ and remaining classes, respectively. The ground truth uncertainty of each sample is calculated by Eq.1, representing the model's ground truth uncertainty for each sample, along with the sample-wise uncertainty metrics $u_{\\mathrm{FL}},u_{\\mathrm{DFL}}$ and $u_{\\mathrm{gBS}}$ . To validate the accuracy", + "bbox": [ + 511, + 809, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15501", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/aca6666bed3413b3be9605a798fff6ec0073552017e27adcab00ea1518a3294a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetModelCEBLMMCEFLSDDFLBSCEBSCE-GRA
Pre TPost TPre TPost TPre TPost TPre TPost TPre TPost TPre TPost TPre TPost T
CIFAR10ResNet504.361.324.291.504.481.411.261.151.001.000.880.880.740.74
ResNet1104.71.564.481.664.801.291.811.171.011.010.990.990.870.87
WideResNet3.350.942.861.103.651.281.841.043.321.161.70.951.461.12
DenseNet4.641.463.961.464.811.671.371.170.870.771.011.010.870.87
CIFAR100ResNet5018.053.057.874.2715.873.325.532.572.542.561.901.901.591.59
ResNet11018.844.6316.774.3018.653.936.883.713.473.472.752.752.532.53
WideResNet14.813.277.744.4614.582.992.702.715.452.522.632.422.462.46
DenseNet19.13.438.132.9917.562.873.381.304.681.831.631.621.621.61
TinyImageNetResNet5014.945.167.811.4714.582.992.182.186.712.284.01.764.571.47
", + "bbox": [ + 91, + 88, + 906, + 241 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison of Calibration Methods Using ECE Across Various Datasets and Models. ECE values are reported using 15 bins, with the best-performing method for each dataset-model combination highlighted in bold. Results are averaged over three runs with different random seeds.", + "bbox": [ + 89, + 252, + 906, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/aa2b34465d503e26385b986d94702aec25884608307ea87fb7ffbe260d14c3fc.jpg", + "image_caption": [ + "Figure 3. Comparison of different ECE metrics. The first three plots show the uncertainty for CIFAR-10 using ResNet-50, while the remaining plots represent ResNet-110 on CIFAR-10." + ], + "image_footnote": [], + "bbox": [ + 99, + 308, + 218, + 382 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b1a0cafc6b558ccd1f52e7538a2fce08d5c2ce37372c99c37642608df313b262.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 308, + 356, + 382 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e8ea14360f6549b8709c684c222ab15b20a8e06334bb4750e505779ca5499dbf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 309, + 491, + 382 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f7871714057ae1a14355edb664b5c02e74c3151e359785b44af2a093a17fd064.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 309, + 627, + 382 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/87d1358feb434fb5f3fffaae88fbbb5e79a53bfb5edcf9f25f820823a2c0073a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 309, + 764, + 382 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/017e996c9608e71808b491d3dbfad8f3c6d79b6a46c14f907f0bb51928b41168.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 777, + 310, + 898, + 381 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of these metrics, we used Pearson correlations between the calibrated error and the uncertainty metrics to verify their positive correlation among 5 runs. We use grid-search to find the optimal hyper-parameter for each methods. Among the methods, gBS achieves the highest Pearson correlation coefficient of 0.664, indicating the strongest linear relationship between the predicted values and the true targets. DFL follows with a correlation of 0.638, while FL has the lowest correlation at 0.550. This suggests that gBS provides the most accurate predicted uncertainty alignment with ground-truth uncertainty, followed by DFL and FL.", + "bbox": [ + 88, + 450, + 482, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on the results, we conclude that although ground-truth uncertainty is inaccessible for real-world datasets, these alternative metrics provide a reliable means of estimating uncertainty, making them suitable for use during training. Besides, the metric $u_{\\mathrm{gBS}}$ has the best performance compared to the uncertainty metric used in Focal Loss and Dual Focal Loss. We incorporate $u_{\\mathrm{gBS}}$ into the gradient-weighted framework discussed in Section 4.2 and introduce a new loss function called BSCE-GRA. This loss function uses the generalized Brier Score as an adaptive uncertainty metric to weight the Cross Entropy gradients, defined as:", + "bbox": [ + 89, + 618, + 482, + 784 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {B S C E - G R A}} (\\boldsymbol {x}, y) = - \\int \\sum_ {i = 1} ^ {K} u _ {\\mathrm {g B S}} (\\hat {p} (\\boldsymbol {x})) \\cdot \\frac {y _ {i}}{\\hat {p} _ {i} (\\boldsymbol {x})} \\mathrm {d} \\hat {p} (\\boldsymbol {x}). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 792, + 482, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further validate the effectiveness of proposed method, we provide a comprehensive theoretical evidence that optimizing with BSCE-GRA, the K-class predicted probability $q$ would equal the actual class-posterior probability", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\eta$ , thereby preventing over/under-confidence when convergence. Due to page limitation, the proof is provided in Appendix.", + "bbox": [ + 511, + 450, + 905, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Besides, the Uncertainty-GRA Loss makes it possible to use non-differentiable uncertainty metrics for model calibration. However, it requires the weights to capture sample-wise uncertainty, whereas some existing measurements like ECE can only compute the uncertainty for a group of samples, making them unsuitable for the proposed framework.", + "bbox": [ + 511, + 496, + 905, + 587 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 511, + 617, + 637, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our methods on multiple deep neural networks (DNNs), including ResNet50, ResNet110 [6], WideResNet [35], and DenseNet [8]. Our experiments are conducted on CIFAR-10, CIFAR-100 [12], and Tiny-ImageNet [4] to assess calibration performance. Further details about the datasets can be found in the appendix.", + "bbox": [ + 511, + 643, + 905, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our methods, BSCE, BSCE-GRA, and ECE-CE, with multiple existing approaches, including training with Cross Entropy (CE), Brier Loss (BL)[2], MMCE Loss[15], Focal Loss with Adaptive Exponent (FLSD)[18], and Dual Focal Loss[27]. For Focal Loss, we employ the FLSD-53 strategy [18] to adaptively adjust the gamma value sample-wise, setting $\\gamma_{\\mathrm{FL}} = 5$ for $\\hat{p}_c \\in [0,0.2)$ and $\\gamma_{\\mathrm{FL}} = 3$ for $\\hat{p}_c \\in [0.2,1)$ . For Dual Focal Loss, the gamma value is set to 5, as reported in the original work.", + "bbox": [ + 511, + 734, + 905, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Setup. Our training setup follows prior works [18, 27]. We train CIFAR-10 and CIFAR-100 for 350", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15502", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e353d8870b219815d2cee7f87df3da1e70f68397c8e672e44560104b5602fb24.jpg", + "image_caption": [ + "(a) Gradient norm distribution at epoch 50" + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 282, + 175 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/348cade48d66fc01a8b12e3d67982302da7cf23b4169562c1524d92d12f3cb78.jpg", + "image_caption": [ + "(b) Gradient norm distribution at epoch 150", + "Figure 4. Evolution of gradient norm distributions across different training epochs for various loss functions. The scatter plots show the relationship between gradient norm and Brier Score for different loss functions (Focal Loss, Dual Focal Loss, BSCE-GRA)." + ], + "image_footnote": [], + "bbox": [ + 300, + 88, + 488, + 175 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a07661b82435e63d1529d9abf5fcea3f2d8cc682cc060d3bdfa3624e1207687b.jpg", + "image_caption": [ + "(c) Gradient norm distribution at epoch 250" + ], + "image_footnote": [], + "bbox": [ + 506, + 88, + 696, + 175 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/22beb2d17ba9c0f3ac1c97e94a7f6ca3bb131c95d596ca070edfc7b23ee2476b.jpg", + "image_caption": [ + "(d) Gradient norm distribution at epoch 350" + ], + "image_footnote": [], + "bbox": [ + 715, + 88, + 901, + 175 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d22b5198ef5dbd0bc0976c7940e6b40b01d78dca536cc5d41dccc371aae4a76a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MetricsFLSDFLSD-GRADFLDFL-GRABSCEBSCE-GRA
pre Tpost Tpre Tpost Tpre Tpost Tpre Tpost Tpre Tpost Tpre Tpost T
Acc95.04%95.04%94.72%94.72%94.63%94.63%94.76%94.76%95.03%95.03%94.69%94.69%
ECE1.261.150.880.881.001.000.930.840.880.880.740.74
Ada ECE1.561.451.191.191.221.220.770.810.960.960.710.71
", + "bbox": [ + 91, + 255, + 906, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison of weighting different uncertainty metrics on gradient or loss function, including $u_{\\mathrm{FL}}$ and $u_{\\mathrm{FL}}$ . The results validate the effectiveness of the gradient-weighting strategy among different uncertainty metrics.", + "bbox": [ + 88, + 348, + 906, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "epochs, using 5,000 images from the training set for validation. The learning rate is initially set to 0.1 for the first 150 epochs, then reduced to 0.01 for the next 100 epochs, and further reduced to 0.001 for the remaining epochs. For Tiny-ImageNet, we train for 100 epochs, with the learning rate set to 0.1 for the first 40 epochs, 0.01 for the next 20 epochs, and 0.001 for the remaining epochs. All experiments are conducted using SGD with a weight decay of $5 \\times 10^{-4}$ and a momentum of 0.9. The training and testing batch sizes for all datasets are set to 128. We re-run all baseline methods using three different random seeds (1, 42, and 71), and report the average results. All experiments are performed on a single Nvidia 4090 GPU. For temperature scaling, the temperature parameter $T$ is optimized through a grid search with $T \\in [0, 0.1, 0.2, \\dots, 10]$ on the validation set, selecting the value that yields the best post-temperature-scaling Expected Calibration Error (ECE). The same optimized temperature parameter is applied to other metrics, such as AdaECE. Further details on the datasets and additional experimental setup are provided in the appendix.", + "bbox": [ + 88, + 402, + 485, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Calibration Performance", + "text_level": 1, + "bbox": [ + 89, + 724, + 318, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We report the average ECE before and after temperature scaling among three random seeds, along with the corresponding optimal temperatures, in Table 1. BSCE-GRA achieves state-of-the-art ECE performance in most cases, particularly in the pre-temperature-scaling results. Notably, the fact that most of the optimal temperatures for BSCE-GRA are found to be 1 indicates that BSCE-GRA trains an inherently calibrated model, capable of achieving strong calibration performance without the need for additional temperature scaling. This is a crucial advantage", + "bbox": [ + 88, + 750, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "for developing accurate and reliable models that are efficient and require minimal post-processing. The results on CIFAR-10 generally show better calibration performance compared to datasets with more labels (e.g., CIFAR-100 and Tiny-ImageNet) across multiple models. Regarding network architecture, ResNet-50 demonstrates the best calibration performance among the four DNNs tested (ResNet-50, ResNet-110, Wide-ResNet-26-10, and DenseNet-121) on both CIFAR-10 and CIFAR-100 datasets.", + "bbox": [ + 511, + 402, + 906, + 539 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Different Metrics. The methods are further evaluated using several widely-accepted metrics to assess calibration performance across models, including Adaptive ECE and Classwise-ECE. Adaptive ECE measures the expected calibration error while accounting for the distribution of the data, whereas Classwise-ECE is a variant that evaluates calibration error for each class individually. Figure 3 presents the results of multiple methods using ResNet-50 and ResNet-110 on the CIFAR-10 dataset. The figure demonstrates that BSCE-GRA is the only method that achieves both inherently calibrated models and state-of-the-art performance across various metrics. Additional results are presented in the appendix, providing further evidence of the effectiveness of our method in model calibration.", + "bbox": [ + 511, + 546, + 908, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Calibration over Training. Figure 5a presents the ECE on the test set for models trained with Focal Loss and Cross-Entropy loss over the entire training period on CIFAR-10 using ResNet-50. To improve visualization, the ECE values are smoothed using an exponential moving average. The figure suggests that, after the initial warm-up epochs, where predicted probabilities are unstable, the ECE of models trained with BSCE-GRA and Focal Loss consistently remains lower compared to models trained with", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15503", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f226010bb8659982c60fc861a275f563ccbe17bdaea90cba375cd84ba8778ebf.jpg", + "image_caption": [ + "(a) ECE over epochs using CE, Focal Loss and BSCE-GRA." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 339, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b782b0e452c63074844c72f91585b72b3f4f6b60c49927e187f68034089a55ea.jpg", + "image_caption": [ + "(b) Gradient Density at epoch 150 using CE, Focal Loss and BSCE-GRA.", + "Figure 5. Figure 5a presents the evolution of ECE throughout the training process, demonstrating that our method rapidly converges to the best result by epoch 250. The subsequent figures depict the gradient magnitudes of various methods between epochs 150 and 250." + ], + "image_footnote": [], + "bbox": [ + 375, + 90, + 620, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1f347dbf8a6d7a902fc393558799e170acc7e9d3bf84dc8076516982c7d6fc3f.jpg", + "image_caption": [ + "(c) Gradient Density at epoch 250 using CE, Focal Loss and BSCE-GRA." + ], + "image_footnote": [], + "bbox": [ + 658, + 90, + 903, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cross-Entropy. It also indicates that, during training with a moderate learning rate from epochs 150 to 200, Focal Loss tends to produce better-calibrated models than BSCE-GRA. This may be due to the fact that, during these epochs, the model makes more predictions with mid-range confidence levels, and as shown in Figure 1, Focal Loss directs the model's attention towards these moderately uncertain samples. In contrast, BSCE-GRA imposes stronger regularization on the gradient, resulting in smaller optimization steps compared to Focal Loss. To further validate this hypothesis, we present the gradient density of the last linear layer across the entire training set at epochs 150 and 250 in Figure 5b and Fig. 5c, which shows that BSCE-GRA results in smaller gradient magnitudes during these epochs. After 250 epochs, when the learning rate undergoes its second reduction, the model trained with BSCE-GRA soon achieves better calibration performance compared to FL and CE.", + "bbox": [ + 88, + 295, + 483, + 553 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Gradient Value among Uncertainty", + "text_level": 1, + "bbox": [ + 89, + 561, + 398, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We extend our analysis to examine the relationship between gradient magnitude and uncertainty. Specifically, we compute the gradient norms of the last linear layer for all samples in the training set at epochs 50, 150, 250, and 350. To quantify uncertainty, we calculate the Brier Score for each training sample. We present the results for Focal Loss (FL), Dual Focal Loss (DFL), and BSCE-GRA. Figure 4 illustrates the relationship between gradient magnitude and Brier Score. The figure clearly shows that the gradients produced by BSCE-GRA are the most sensitive to changes in the Brier Score, as indicated by the narrow distribution compared to other loss functions. This aligns with our goal: we aim for BSCE-GRA to exhibit sufficient sensitivity to uncertainty, allowing the model to adjust its gradient values based on changes in uncertainty, thereby focusing more on highly uncertain samples. Notably, the gradient distribution for Dual Focal Loss is relatively more dispersed concerning the Brier Score, with a wider range of possible values compared to other loss functions. This could be attributed to the partial derivatives involving other classes, as Dual Focal Loss takes the second most probable class into account", + "bbox": [ + 89, + 583, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in its loss calculation. In contrast, Focal Loss and BSCE-GRA only involve gradients along a single dimension.", + "bbox": [ + 511, + 295, + 906, + 325 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Weighting FL and DFL on Gradients", + "text_level": 1, + "bbox": [ + 511, + 352, + 834, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate the performance of directly applying the uncertainty terms $u_{\\mathrm{FL}}$ and $u_{\\mathrm{DFL}}$ as weights on the gradients. Experiments are conducted using the default settings discussed above, on CIFAR-10 with ResNet-50. The results are presented in Table 2. It is evident that applying weights on gradients results in performance improvement for both $u_{\\mathrm{FL}}$ and $u_{\\mathrm{DFL}}$ , further validating its effectiveness.", + "bbox": [ + 511, + 378, + 906, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct additional experiments to comprehensively validate our proposed method, BSCE-GRA, under various settings. Due to page limitations, the extended experiments are provided in the Appendix.", + "bbox": [ + 511, + 491, + 906, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 582, + 633, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed a novel approach to model calibration by directly weighting gradients based on uncertainty. We analyzed the strengths and limitations of Focal Loss and Dual Focal Loss from the perspective of sample weighting and introduced a framework that scales gradient magnitudes based on model uncertainty to focus more on uncertain samples. Additionally, we introduced BSCEGRA, a loss function incorporating uncertainty metrics to enhance model calibration. Extensive experiments on various datasets and network architectures demonstrated significant improvements in calibration performance, achieving state-of-the-art results. Our findings emphasize the value of integrating uncertainty-aware mechanisms directly into the optimization process, providing a reliable framework for training calibrated deep neural networks suitable for real-world applications requiring trustworthy predictions. Future work may explore further optimization strategies for uncertainty-weighted approaches and their impact on tasks like active learning and robustness to adversarial attacks.", + "bbox": [ + 509, + 613, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15504", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgment", + "text_level": 1, + "bbox": [ + 94, + 89, + 259, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the Start-up Grant (No. 9610680) of the City University of Hong Kong, Young Scientist Fund (No. 62406265) of NSFC, and the Australian Research Council under Projects DP240101848 and FT230100549.", + "bbox": [ + 93, + 114, + 480, + 188 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15505", + "bbox": [ + 480, + 946, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ondrej Bohdal, Yongxin Yang, and Timothy Hospedales. Meta-calibration: Learning of model calibration using differentiable expected calibration error. arXiv preprint arXiv:2106.09613, 2021. 1, 2", + "[2] Glenn W Brier. Verification of forecasts expressed in terms of probability. Monthly weather review, 78(1):1-3, 1950. 1, 3, 6", + "[3] Corinna Cortes. Support-vector networks. Machine Learning, 1995. 2", + "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6, 1", + "[5] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In International conference on machine learning, pages 1321-1330. PMLR, 2017. 1, 2, 3", + "[6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[7] Dan Hendrycks*, Norman Mu*, Ekin Dogus Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. Aug-mix: A simple method to improve robustness and uncertainty under data shift. In International Conference on Learning Representations, 2020. 2", + "[8] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700-4708, 2017. 6", + "[9] Like Hui and Mikhail Belkin. {EVALUATION} {of} {neural} {architectures} {trained} {with} {square} {loss} {vs} {cross}-{entropy} {in} {classification} {tasks}. In International Conference on Learning Representations, 2021. 2", + "[10] Archit Karandikar, Nicholas Cain, Dustin Tran, Balaji Lakshminarayanan, Jonathon Shlens, Michael C Mozer, and Becca Roelofs. Soft calibration objectives for neural networks. Advances in Neural Information Processing Systems, 34:29768-29779, 2021. 2", + "[11] Ranganath Krishnan and Omesh Tickoo. Improving model calibration with accuracy versus uncertainty optimization. Advances in Neural Information Processing Systems, 33: 18237-18248, 2020. 2", + "[12] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6, 1", + "[13] Meelis Kull, Telmo Silva Filho, and Peter Flach. Beta calibration: a well-founded and easily implemented improvement on logistic calibration for binary classifiers. In Artificial intelligence and statistics, pages 623-631. PMLR, 2017. 2", + "[14] Meelis Kull, Miquel Perello Nieto, Markus Kängsepp, Telmo Silva Filho, Hao Song, and Peter Flach. Beyond temperature scaling: Obtaining well-calibrated multi-class probabilities with dirichlet calibration. Advances in neural information processing systems, 32, 2019. 2, 3" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Aviral Kumar, Sunita Sarawagi, and Ujjwal Jain. Trainable calibration measures for neural networks from kernel mean embeddings. In International Conference on Machine Learning, pages 2805-2814. PMLR, 2018. 1, 6", + "[16] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems, 30, 2017. 2", + "[17] T Lin. Focal loss for dense object detection. arXiv preprint arXiv:1708.02002, 2017. 1, 3", + "[18] Jishnu Mukhoti, Viveka Kulharia, Amartya Sanyal, Stuart Golodetz, Philip Torr, and Puneet Dokania. Calibrating deep neural networks using focal loss. Advances in Neural Information Processing Systems, 33:15288-15299, 2020. 1, 2, 3, 6", + "[19] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? Advances in neural information processing systems, 32, 2019. 2", + "[20] Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. In Proceedings of the AAAI conference on artificial intelligence, 2015. 2", + "[21] Khanh Nguyen and Brendan O'Connor. Posterior calibration and exploratory analysis for natural language processing models. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1587-1598, 2015. 3", + "[22] Joshua C Peterson, Ruairidh M Battleday, Thomas L Grifths, and Olga Russakovsky. Human uncertainty makes classification more robust. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9617-9626, 2019. 6", + "[23] John Platt et al. Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Advances in large margin classifiers, 10(3):61-74, 1999. 2", + "[24] Rahul Rahaman et al. Uncertainty quantification and deep ensembles. Advances in neural information processing systems, 34:20063-20075, 2021. 2", + "[25] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 2", + "[26] Linwei Tao, Minjing Dong, Daochang Liu, Changming Sun, and Chang Xu. Calibrating a deep neural network with its predecessors. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 4271–4279, 2023. 1", + "[27] Linwei Tao, Minjing Dong, and Chang Xu. Dual focal loss for calibration. In International Conference on Machine Learning, pages 33833-33849. PMLR, 2023. 1, 2, 3, 6", + "[28] Linwei Tao, Minjing Dong, and Chang Xu. Feature clipping for uncertainty calibration. arXiv preprint arXiv:2410.19796, 2024. 1", + "[29] Linwei Tao, Haolan Guo, Minjing Dong, and Chang Xu. Consistency calibration: Improving uncertainty calibration via consistency among perturbed neighbors. arXiv preprint arXiv:2410.12295, 2024. 6" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15506", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Linwei Tao, Younan Zhu, Haolan Guo, Minjing Dong, and Chang Xu. A benchmark study on calibration. In The Twelfth International Conference on Learning Representations, 2024. 1", + "[31] Sunil Thulasidasan, Gopinath Chennupati, Jeff A Bilmes, Tanmoy Bhattacharya, and Sarah Michalak. On mixup training: Improved calibration and predictive uncertainty for deep neural networks. Advances in neural information processing systems, 32, 2019. 2", + "[32] Deng-Bao Wang, Lei Feng, and Min-Ling Zhang. Rethinking calibration of deep neural networks: Do not be afraid of overconfidence. Advances in Neural Information Processing Systems, 34:11809-11820, 2021. 2", + "[33] Jason Weston, Samy Bengio, and Nicolas Usunier. Wsabie: Scaling up to large vocabulary image annotation. In Twenty-Second International Joint Conference on Artificial Intelligence. CiteSeer, 2011. 2", + "[34] Bianca Zadrozny and Charles Elkan. Transforming classifier scores into accurate multiclass probability estimates. In Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining, pages 694-699, 2002. 2", + "[35] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In *British Machine Vision Conference* 2016. British Machine Vision Association, 2016. 6", + "[36] Jize Zhang, Bhavya Kailkhura, and T Yong-Jin Han. Mix-n-match: Ensemble and compositional methods for uncertainty calibration in deep learning. In International conference on machine learning, pages 11117-11128. PMLR, 2020. 2" + ], + "bbox": [ + 91, + 92, + 480, + 500 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15507", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Uncertainty Weighted Gradients for Model Calibration/792a362b-85b5-4092-b750-5040ac391b64_model.json b/2025/Uncertainty Weighted Gradients for Model Calibration/792a362b-85b5-4092-b750-5040ac391b64_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6f044f0a6a1dd8bcff753893576f24c073214945 --- /dev/null +++ b/2025/Uncertainty Weighted Gradients for Model Calibration/792a362b-85b5-4092-b750-5040ac391b64_model.json @@ -0,0 +1,2213 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.22, + 0.131, + 0.78, + 0.154 + ], + "angle": 0, + "content": "Uncertainty Weighted Gradients for Model Calibration" + }, + { + "type": "text", + "bbox": [ + 0.206, + 0.18, + 0.788, + 0.234 + ], + "angle": 0, + "content": "Jinxu Lin\\(^{1}\\) Linwei Tao\\(^{1}\\) Minjing Dong\\(^{2}\\) Chang Xu\\(^{1}\\) \n\\(^{1}\\)The University of Sydney, \\(^{2}\\)City University of Hong Kong \n{jinxu.lin, linwei.tao, c.xu}@sydney.edu.au, minjdong@cityu.edu.hk" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.329, + 0.287 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.302, + 0.486, + 0.665 + ], + "angle": 0, + "content": "Model calibration is essential for ensuring that the predictions of deep neural networks accurately reflect true probabilities in real-world classification tasks. However, deep networks often produce over-confident or under-confident predictions, leading to miscalibration. Various methods have been proposed to address this issue by designing effective loss functions for calibration, such as focal loss. In this paper, we analyze its effectiveness and provide a unified loss framework of focal loss and its variants, where we mainly attribute their superiority in model calibration to the loss weighting factor that estimates sample-wise uncertainty. Based on our analysis, existing loss functions fail to achieve optimal calibration performance due to two main issues: including misalignment during optimization and insufficient precision in uncertainty estimation. Specifically, focal loss cannot align sample uncertainty with gradient scaling and the single logit cannot indicate the uncertainty. To address these issues, we reformulate the optimization from the perspective of gradients, which focuses on uncertain samples. Meanwhile, we propose using the Brier Score as the loss weight factor, which provides a more accurate uncertainty estimation via all the logits. Extensive experiments on various models and datasets demonstrate that our method achieves state-of-the-art (SOTA) performance.\\(^{1}\\)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.696, + 0.224, + 0.712 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.722, + 0.483, + 0.843 + ], + "angle": 0, + "content": "Deep Neural Networks (DNNs) have achieved remarkable success in various domains, including image classification. However, recent studies [5, 30] reveal that DNNs often suffer from mis-calibration in classification task, exhibiting over-confidence or under-confidence in their predictions. For example, a model might output a confidence score of 0.8 for a particular prediction, which does not necessarily correspond to an \\(80\\%\\) probability of correctness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.844, + 0.483, + 0.874 + ], + "angle": 0, + "content": "Calibrating DNNs, which aligns the predicted confidence with the true probabilities, is therefore crucial to en" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.27, + 0.908, + 0.573 + ], + "angle": 0, + "content": "hancing their reliability in practical applications, such as autonomous driving, medical imaging, and weather forecasting. To address mis-calibration issues, several methods [26, 28] have been proposed, many of which focus on modifying the loss function during training. A common approach involves adding regularization terms with the Cross-Entropy (CE) loss to improve calibration, as seen in methods like Maximum Mean Calibration Error (MMCE) [15] and Meta Calibration [1]. Additionally, Focal Loss (FL) [17], which adjusts per-sample loss weights based on prediction difficulty, has been shown to improve calibration performance. Mukhoti et al. [18] attribute the effectiveness of FL to its implicit regularization on the entropy of predicted probabilities, which mitigates overconfidence. Similarly, Dual Focal Loss (DFL) [27] incorporates the second-most-probable class in the weighting mechanism to address model under-confidence. Another research interest regarding calibration is the evaluation of calibration performance. Metrics commonly employed include Expected Calibration Error (ECE)[5] and Brier Score[2]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.578, + 0.909, + 0.761 + ], + "angle": 0, + "content": "The focal-loss-based methods can be unified under a general loss framework expressed as \\( u \\cdot \\mathrm{CE} \\), where \\( u \\) is a loss weighting factor. In FL, the factor \\( u_{\\mathrm{FL}} \\) can be treated as an estimation for sample-wise difficulty. Interestingly, we observe that \\( u_{\\mathrm{FL}} \\), originally designed for binary classification, is mathematically equivalent to the Brier Score for binary cases, which can measure the uncertainty. Furthermore, the weighting term \\( u_{\\mathrm{DFL}} \\) in DFL aligns with the Brier Score in three-class classification scenarios. This observation motivates the hypothesis that weighting the loss with sample calibration metrics, such as the Brier Score, could better identify uncertain samples and facilitate targeted training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Furthermore, our investigation reveals a limitation of directly applying the weighting factor \\( u \\) to the loss term. Specifically, the vanilla optimization of focal loss would achieve a misalignment with the objective of scaling gradients for harder samples with larger magnitudes. Based on our analysis, we mainly attribute this issue to a differentiable loss weighting factor \\( u \\), which could disrupt the positive correlation between the CE and its gradient magnitude during backpropagation, impeding the model's ability" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.446, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)Code is available at https://github.com/Jinxu-Lin/BSCE-GRA." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15497" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.416, + 0.106 + ], + "angle": 0, + "content": "to prioritize higher-uncertain samples effectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.107, + 0.482, + 0.287 + ], + "angle": 0, + "content": "To address these issues, we first propose applying the weighting factor to scaling the gradient rather than the loss itself. This allows us directly aligns the gradient optimization with sample uncertainty, which ensures that harder samples receive appropriately scaled updates without disrupting the optimization process. Within this framework, we then introduce a generalized form of the Brier Score as a gradient weighting factor, which provides a more accurate estimation of sample uncertainty that considers all categories. This leads to the development of the BSCE-GRA loss function, which adjusts gradient to directly scale optimization based on sample uncertainty by Brier Score." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.288, + 0.482, + 0.363 + ], + "angle": 0, + "content": "We conduct extensive experiments to validate the effectiveness of the proposed method, achieving state-of-the-art results across various datasets and model architectures, which demonstrate the effectiveness of our methods. The contributions of this paper are summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.365, + 0.483, + 0.425 + ], + "angle": 0, + "content": "1. We provide a new perspective by unifying some existing loss modification techniques for model calibration under a sample-weighting framework. With this framework, we provide extensive analysis of their limitations." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.426, + 0.482, + 0.486 + ], + "angle": 0, + "content": "2. We propose a simple yet effective optimization framework for model calibration with a gradient weighting factor, where we scale the gradients to encourage the model to focus on uncertain samples effectively." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.486, + 0.482, + 0.546 + ], + "angle": 0, + "content": "3. We analyze the use of different uncertainty metrics within this framework and introduce BSCE-GRA, a new loss function based on the Brier Score that provides an accurate sample-wise uncertainty estimation." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.547, + 0.482, + 0.621 + ], + "angle": 0, + "content": "4. We conduct extensive experiments under different settings to validate our proposed methods. Our uncertainty-weighted framework shows consistent effectiveness across various uncertainty metrics, and the proposed BSCE-GRA loss achieves state-of-the-art results." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.365, + 0.483, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.634, + 0.241, + 0.649 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.705 + ], + "angle": 0, + "content": "In recent years, numerous techniques have been proposed to address the problem of network miscalibration, which can generally be classified into three categories." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The first category is post-hoc calibration techniques, which adjust model predictions after training by optimizing additional parameters on a held-out validation set. These methods include Platt Scaling [23], which performs a linear transformation on the original prediction logits; Isotonic Regression [34], which uses piecewise functions to transform logits; Bayesian Binning into Quantiles (BBQ)[20], which extends histogram binning with Bayesian model averaging; and Beta Calibration[13], initially designed for binary classification and later generalized to multi-class settings with Dirichlet distributions by Kull et al. [14]. Temperature Scaling [5], one of the most widely used post-hoc calibration methods, tunes the temperature parameter in the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "SoftMax function to minimize negative log-likelihood on a held-out validation set. In this work, we report calibration performance with post-temperature scaling results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.905, + 0.334 + ], + "angle": 0, + "content": "The second category includes regularization techniques, which are known to effectively calibrate DNNs. Data augmentation methods, such as Mixup [31] and AugMix [7], train DNNs on mixed samples to mitigate overconfident predictions. Model ensemble techniques, which involve independently training multiple DNNs and averaging their predictions, have been shown to enhance both accuracy and predictive uncertainty by aggregating outputs from multiple models [16, 24, 36]. Label smoothing [19], which replaces one-hot labels with soft labels, encourages the model to make less confident predictions, thereby reducing overconfidence. Additionally, weight decay has also been demonstrated to improve confidence calibration [5]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.336, + 0.905, + 0.518 + ], + "angle": 0, + "content": "The third category focuses on modifying the training loss to improve calibration. These methods include adding a differentiable auxiliary surrogate loss for Expected Calibration Error (ECE)[1, 10, 11] or replacing the training loss with other loss functions, such as Mean Squared Error (MSE)[9], Focal Loss [18], Inverse Focal Loss [32], and Dual Focal Loss [27]. Among these, Focal Loss [18], which adds a modulation term to the Cross-Entropy loss to focus on hard-to-classify samples, provides a simple and effective way to train calibrated models. Focal Loss and Dual Focal Loss can be categorized as margin-based losses, similar to Hinge Loss [3], Triplet Loss [25], and Margin Ranking Loss [33]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.534, + 0.64, + 0.551 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.56, + 0.718, + 0.575 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.905, + 0.764 + ], + "angle": 0, + "content": "Given a \\(K\\)-class dataset \\(\\mathbb{S} = \\{\\pmb{z}^{(1)},\\dots,\\pmb{z}^{(N)}\\}\\), where each training sample \\(\\pmb{z}^{(n)}\\coloneqq (x^{(n)},y^{(n)})\\) is an input-label pair. Let \\(\\mathcal{X}\\) and \\(\\mathcal{V}\\) represent the input space and the label space, respectively. The ground truth label \\(y\\in \\mathcal{V}\\) is encoded in a one-hot vector format, where \\(y_{i} = 1\\) if \\(i\\in K\\) represents the actual class. We define a classifier \\(f_{\\theta}\\) trained on \\(\\mathbb{S}\\) that maps an input \\(\\pmb{x}\\in \\mathcal{X}\\) to a probability distribution \\(\\hat{p} (\\pmb {x})\\). The classifier then provides a prediction \\(k = \\arg \\max_{i\\in K}\\hat{p}_i\\), which indicates the index of the predicted label. The predicted label \\(\\hat{y}\\in \\mathcal{Y}\\) is similarly represented as \\(y\\) in the one-hot encoding of prediction \\(k\\). The confidence \\(\\hat{p}_c\\) is defined as the predicted probability associated with the prediction \\(k\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.905, + 0.901 + ], + "angle": 0, + "content": "A well-calibrated model ensures that the provided confidence \\(\\hat{p}_c\\) accurately reflects the true probability of correct classification. We define the true class-posterior probability vector as \\(\\boldsymbol{\\eta}(\\boldsymbol{x}) = [\\eta_1(\\boldsymbol{x}),\\dots,\\eta_K(\\boldsymbol{x})]\\), where \\(\\eta_k(\\boldsymbol{x}) = \\mathbb{P}(y = k|\\boldsymbol{x})\\) is the true probability of class \\(k\\) given input \\(\\boldsymbol{x}\\). Formally, a network is considered perfectly calibrated if \\(\\mathbb{P}(\\hat{y} = y|\\hat{p}_c = p) = p\\) for all \\(p\\in [0,1]\\) [5], which can be also written as \\(\\hat{p} (\\pmb {x}) = \\pmb {\\eta}(\\pmb {x})\\). The uncertainty of a model for a sample \\(\\pmb{x}\\), also referred to the calibrated error \\(c(x)\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15498" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.483, + 0.108 + ], + "angle": 0, + "content": "can be computed as the difference between \\(\\pmb{\\eta}(\\pmb{x})\\) and \\(\\hat{p}(\\pmb{x})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.121, + 0.483, + 0.139 + ], + "angle": 0, + "content": "\\[\nc (\\boldsymbol {x}) = \\| \\eta (\\boldsymbol {x}) - \\hat {p} _ {c} (\\boldsymbol {x}) \\|. \\tag {1}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.146, + 0.345, + 0.161 + ], + "angle": 0, + "content": "3.2. Metric for Calibrated Error" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.483, + 0.228 + ], + "angle": 0, + "content": "In practice, it is hard to access the ground truth probability \\(\\pmb{\\eta}(\\pmb{x})\\) on real-world samples and thus the calibration error cannot be directly computed by Eq. 1, some alternative methods have been proposed to evaluate the uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.229, + 0.484, + 0.404 + ], + "angle": 0, + "content": "ECE. The Expected Calibration Error (ECE) is defined as \\(\\mathbb{E}_{\\hat{p}_c}[\\| \\mathbb{P}(\\hat{y} = y|\\hat{p}_c) - \\hat{p}_c]\\). Guo et al. [5] propose an approximation of ECE. Specifically, all samples are divided into \\(M\\) bins \\(\\{B_m\\}_{m=1}^M\\) of equal width based on their confidence, where each bin \\(B_m\\) contains all samples whose confidences fall within the range \\(\\hat{p}_c \\in \\left[\\frac{m}{M}, \\frac{m+1}{M}\\right)\\). For each bin \\(B_m\\), the average confidence is computed as \\(C_m = \\frac{1}{|B_m|} \\sum_{i \\in B_m} \\hat{p}_c^{(i)}\\) and the bin accuracy is computed as \\(A_m = \\frac{1}{|B_m|} \\sum_{i \\in B_m} \\mathbb{1}(\\hat{y}_k^{(i)} = y_k^{(i)})\\), where \\(\\mathbb{1}\\) is the indicator function. The ECE can then be computed as the average \\(L_1\\) difference between bin accuracy and confidence:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.415, + 0.483, + 0.456 + ], + "angle": 0, + "content": "\\[\n\\mathrm {E C E} = \\sum_ {m = 1} ^ {M} \\frac {\\left| B _ {m} \\right|}{N} \\left| A _ {m} - C _ {m} \\right|, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.461, + 0.484, + 0.567 + ], + "angle": 0, + "content": "where \\(N\\) denotes the number of samples in each bin. In addition to the ECE in Eq.2, there are several variants used to measure calibration error. For example, AdaECE [21] groups samples into bins \\(B_{m}\\) with an equal number of smaples such that \\(|B_m| = |B_n|\\) for all bins. Another variant, ClasswiesECE [14] measure the calibration error by considering each of the \\(K\\) classes separately." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.484, + 0.793 + ], + "angle": 0, + "content": "Brier Score. Besides ECE computing the expectation of calibrated error among whole datasets, some other techniques have also been used to evaluate the sample-wise uncertainty. Accuracy and calibration are distinct concepts, and one cannot be inferred from the other unequivocally. The Brier Score (BS) [2] unify these two concepts, which is commonly used in calibration literature. It has been shown that this family of metrics can be decomposed into a calibration term and a refinement term. Achieving an optimal score requires both accurate predictions and appropriate confidence levels. For a given sample, the Brier Score is mathematically defined as the Mean Squared Error between the predicted probability distribution \\(\\hat{p}\\) and one-hot encoded ground truth label \\(y\\). We introduce a generalized Brier Score (gBS) form as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.804, + 0.483, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathrm {g B S} = \\sum_ {i = 1} ^ {K} \\| \\hat {p} _ {i} (\\boldsymbol {x}) - y _ {i} \\| _ {\\beta} ^ {\\gamma} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\gamma\\) and \\(\\beta\\) are the hyperparameters for the exponent and norm order, respectively. When \\(\\gamma = 2\\) and \\(\\beta = 2\\), this formulation reduces to the original Brier Score." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.706, + 0.107 + ], + "angle": 0, + "content": "3.3. Calibrating Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.907, + 0.253 + ], + "angle": 0, + "content": "Temperature Scaling A widely used post-hoc technique for improving classification calibration is temperature scaling. It adjusts the sharpness of the output probability distribution by scaling the logits in the SoftMax function using a temperature parameter, defined as \\(\\hat{p}_i = \\frac{\\exp(\\hat{g}_i / T)}{\\sum_{k=1}^{K} \\exp(\\hat{g}_k / T)}\\), where \\(\\hat{g}\\) represents the logits before applying the SoftMax function, and \\(T\\) is the temperature that controls the scaling. Calibration performance can be enhanced by tuning \\(T\\) on a held-out validation set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.254, + 0.906, + 0.374 + ], + "angle": 0, + "content": "Focal Loss and Dual Focal Loss Some other previous works have attempted to improve model calibration by modifying the loss function. Focal Loss [17] was initially introduced to address the foreground-background imbalance problem in object detection. It addresses the issue by incorporating a loss weighting factor based on sample complexity, reducing the weight of easy samples and allowing the model to focus on harder-to-classify instances." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.375, + 0.906, + 0.404 + ], + "angle": 0, + "content": "Formally, given the predicted probability \\(\\hat{p} (\\pmb {x})\\) on sample \\(\\pmb{x}\\), the focal loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.414, + 0.906, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {F L}} (\\boldsymbol {x}, y) = - \\sum_ {i = 1} ^ {K} y _ {i} (1 - \\hat {p} _ {i}) ^ {\\gamma} \\log \\hat {p} _ {i} (\\boldsymbol {x}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.906, + 0.645 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is a pre-defined hyperparameter. Previous studies [18] have demonstrated that optimizing models using Focal Loss results in better calibration compared to Cross-Entropy. This improvement is partly due to the entropy-based regularization effect introduced by Focal Loss, while complexity-based weighting also likely plays a role. However, we noticed that Focal Loss was initially applied to binary classification, where the complexity term \\((1 - \\hat{p}_i)\\) is related to the Brier Score, which can also be used for evaluating uncertainty. In multi-class classification, although complexity and uncertainty are not strictly equivalent, the complexity term still captures some class-level uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.906, + 0.705 + ], + "angle": 0, + "content": "Several variants of Focal Loss exist, including Dual Focal Loss [27], which incorporates the probability from the second most probable class into the scaling factor to address the under-confidence issue caused by Focal Loss:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.715, + 0.907, + 0.756 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D F L}} (\\boldsymbol {x}, y) = - \\sum_ {i = 1} ^ {K} y _ {i} (1 - \\hat {p} _ {i} (\\boldsymbol {x}) + \\hat {p} _ {j} (\\boldsymbol {x})) ^ {\\gamma} \\log \\hat {p} _ {i} (\\boldsymbol {x}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.765, + 0.818, + 0.783 + ], + "angle": 0, + "content": "where \\(\\hat{p}_j(x) = \\max_i\\{\\hat{p}_i(x)|\\hat{p}_i(x) < \\hat{p}_{gt}(x)\\}\\)" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.794, + 0.605, + 0.809 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.864, + 0.835 + ], + "angle": 0, + "content": "4.1. Weighting with Sample-wise Uncertainty" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Focal Loss and Dual Focal Loss both support the concept that uncertainty-based weighting in conjunction with Cross Entropy can enhance model calibration. We propose a generalized loss function framework, termed Uncertainty CE" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15499" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "Loss, which incorporates a sample-wise uncertainty metric into Cross Entropy as a weighting factor:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.133, + 0.483, + 0.174 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {U n c e r t a i n t y}} (\\boldsymbol {x}, y) = - \\sum_ {i = 1} ^ {K} u \\left(\\hat {p} _ {i} (\\boldsymbol {x})\\right) \\cdot y _ {i} \\log \\hat {p} _ {i} (\\boldsymbol {x}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.185, + 0.484, + 0.261 + ], + "angle": 0, + "content": "where \\( u(\\hat{p}_i) \\) is the adaptive term that evaluates the uncertainty of sample \\( x \\). The motivation behind this design is to use the predicted calibration error for each sample to scale the loss function, thereby directing the model's optimization more effectively towards samples with higher uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.261, + 0.468, + 0.276 + ], + "angle": 0, + "content": "Therefore, the weights used in Focal Loss is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.287, + 0.483, + 0.304 + ], + "angle": 0, + "content": "\\[\nu _ {\\mathrm {F L}} (\\hat {p} (\\boldsymbol {x})) = (1 - \\hat {p} _ {c}) ^ {\\gamma}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.315, + 0.45, + 0.33 + ], + "angle": 0, + "content": "And the scaling term in Dual Focal Loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.342, + 0.483, + 0.359 + ], + "angle": 0, + "content": "\\[\nu _ {\\mathrm {D F L}} (\\hat {p} (\\boldsymbol {x})) = (1 - \\hat {p} _ {c} (x) + \\hat {p} _ {j} (x)) ^ {\\gamma}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.37, + 0.484, + 0.49 + ], + "angle": 0, + "content": "where \\(\\hat{p}_j(x)\\) is the second maximum confidence in prediction. The weight used in Focal Loss can be seen as derived from the Brier Score, focusing on the error of the ground true class. In contrast, the scaling term in Dual Focal Loss incorporates uncertainty information from the second most probable class. However, directly weighting loss function with certain uncertainty metrics, as those used in Focal Loss and Dual Focal Loss, presents several challenges." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.484, + 0.701 + ], + "angle": 0, + "content": "Reviewing our objectives, we aim to weight the loss function value by each sample's uncertainty, resulting in greater optimization steps for those samples with higher uncertainty. Our focus actually is not on the loss value itself, but on the optimization. The purpose of uncertainty weighting is to focus the model's attention on samples with higher uncertainty. Thus, the key factor is the gradient: higher weights should lead to larger gradients, resulting in more substantial model updates for those samples. For Cross Entropy, its value is positively correlated with the gradient: a higher CE value yields a larger gradient, leading to more significant updates for optimization. When applying a scalar weight to CE, this relationship remains intact, as it does not disrupt the positive correlation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.702, + 0.484, + 0.905 + ], + "angle": 0, + "content": "However, for differentiable weights, directly applying them to the loss can disrupt the positive correlation between the loss and the gradient, ultimately impairing effective model optimization. To further analyze this, let us consider Focal Loss. For Focal Loss \\(\\mathcal{L}_{\\mathrm{FL}}\\) and Cross Entropy \\(\\mathcal{L}_{\\mathrm{CE}}\\), their gradients are given by: \\(\\frac{\\partial}{\\partial w}\\mathcal{L}_{\\mathrm{FL}} = g(\\hat{p}_i,\\gamma)\\frac{\\partial}{\\partial w}\\mathcal{L}_{\\mathrm{CE}}\\), where \\(g(p,\\gamma) = (1 - p)^{\\gamma} - \\gamma p(1 - p)^{\\gamma -1}\\log (p)\\), \\(\\gamma\\) is a predefined hyperparameter, and \\(\\boldsymbol{w}\\) represents the parameters of the final linear layer. From a gradient perspective, \\(g(p,\\gamma)\\) acts as a weight on the CE gradient, which is illustrated in Figure 1. There exists a point \\(p_0\\in [0,1]\\) such that within the range \\([0,p_0],\\frac{\\partial}{\\partial p} g(p,\\gamma) > 0\\) and within the range \\([p_0,1],\\frac{\\partial}{\\partial p} g(p,\\gamma) < 0\\). In binary classification, \\(1 - \\hat{p}_i\\) is related to" + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.091, + 0.882, + 0.104 + ], + "angle": 0, + "content": "Plot of \\( g(p) \\) for Different \\( \\gamma \\) Values in Focal Loss" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.103, + 0.905, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.28, + 0.885, + 0.295 + ], + "angle": 0, + "content": "Figure 1. \\( g(p, \\gamma) \\) of Focal Loss vs predicted confidence \\( \\hat{p}_c \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.907, + 0.516 + ], + "angle": 0, + "content": "the Brier Score that reveals calibrated error, indicating that the predicted calibrated error is not always positively correlated with the weight on gradient. As uncertainty decreases, the weight on gradient initially increases, before eventually aligning with the level of uncertainty. This causes the model to focus more on samples with moderate uncertainty rather than those with the highest uncertainty, which contradicts our original goal of emphasizing the most uncertain samples. Further, changes in uncertainty are not immediately reflected in the sample weights; instead, multiple training iterations are required for these changes to align properly. Thus, directly applying some kind of uncertainty metrics into Eq. 6 would pose issues to the uncertainty weighting." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.523, + 0.848, + 0.539 + ], + "angle": 0, + "content": "4.2. Sample-wise Uncertainty on Gradients" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.906, + 0.62 + ], + "angle": 0, + "content": "The discussion in Sec. 4.1 motivates us to ensure the weight on gradients for a sample to align with its uncertainty. This could be intuitively solved by directly applying the model uncertainty on samples as a weight of the gradient. We define the modified gradient as:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.628, + 0.905, + 0.658 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial}{\\partial \\theta} \\mathcal {L} _ {\\text {U n c e r t a i n t y - G R A}} (\\boldsymbol {x}, y) = u (\\hat {p} (\\boldsymbol {x})) \\frac {\\partial}{\\partial \\theta} \\mathcal {L} _ {\\mathrm {C E}} (\\boldsymbol {x}, y). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.665, + 0.907, + 0.799 + ], + "angle": 0, + "content": "Taking a SGD optimizer as an example, the model update can be expressed as: \\(\\theta_{t + 1} = \\theta +\\alpha \\cdot \\frac{\\partial}{\\partial\\theta}\\mathcal{L}_{\\mathrm{Uncertainty - GRA}}(\\boldsymbol {x},y)\\) where \\(\\theta\\) is the model parameters and \\(\\alpha\\) is the learning rate. When the predicted uncertainty on sample \\(\\pmb{x}\\) is higher, the optimization step results in a larger update, encouraging the model to focus more on uncertain samples. A general form of the loss function for our uncertainty-weighted gradient framework, termed Uncertainty-GRA CE Loss, is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.807, + 0.905, + 0.847 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {U n c e r t a i n t y - G R A}} (\\boldsymbol {x}, y) = - \\int \\sum_ {i = 1} ^ {K} u (\\hat {p} (\\boldsymbol {x})) \\cdot \\frac {y _ {i}}{\\hat {p} _ {i} (\\boldsymbol {x})} \\mathrm {d} \\hat {p} (\\boldsymbol {x}), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In practice, to compute the \\(\\mathcal{L}_{\\mathrm{Uncertainty - GRA}}\\), we can detach the gradient of \\(u(\\hat{p} (\\pmb {x}))\\) and multiply it with the Cross Entropy, instead of calculating the gradient integration." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15500" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.086, + 0.322, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.25, + 0.293, + 0.262 + ], + "angle": 0, + "content": "(a) An illustration of the \\(u_{\\mathrm{FL}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.087, + 0.605, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.25, + 0.578, + 0.262 + ], + "angle": 0, + "content": "(b) An illustration of the \\( u_{\\mathrm{DFL}} \\)." + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.087, + 0.887, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.25, + 0.859, + 0.262 + ], + "angle": 0, + "content": "(c) An illustration of the \\(u_{\\mathrm{BS}}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.274, + 0.908, + 0.304 + ], + "angle": 0, + "content": "Figure 2. An illustration of value of gradient weight function on a 4 class classification. It is obvious that \\( u_{\\mathrm{FL}} \\) varies only along the \\( p_i \\) axis and \\( u_{\\mathrm{DFL}} \\) changes along the \\( p_i \\) and \\( p_j \\) axes. \\( u_{\\mathrm{BS}} \\) responds to changes across all axes, providing a more complete uncertainty evaluation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.328, + 0.383, + 0.344 + ], + "angle": 0, + "content": "4.3. Sample-wise Uncertainty Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.35, + 0.484, + 0.426 + ], + "angle": 0, + "content": "With Eq. 10, calibration can be achieved by employing the sample's ground-truth uncertainty as the scaling term. However, accessing the ground-truth uncertainty for real-world image samples is impractical. Therefore, alternative metrics are required to estimate the sample-wise uncertainty \\( u(x) \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.484, + 0.456 + ], + "angle": 0, + "content": "One approach to evaluate sample-wise uncertainty is based on computing the generalized Brier Score (gBS):" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.466, + 0.483, + 0.507 + ], + "angle": 0, + "content": "\\[\nu _ {\\mathrm {g B S}} (\\hat {p} (\\boldsymbol {x})) = \\sum_ {i = 1} ^ {K} \\| \\hat {p} _ {i} (\\boldsymbol {x}) - y _ {i} \\| _ {\\beta} ^ {\\gamma}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.484, + 0.591 + ], + "angle": 0, + "content": "We provide a brief discussion on the effectiveness of the generalized Brier Score (gBS) as a measure of calibration error. When using the Brier Score (BS) to evaluate calibration error, with \\(\\beta = 2\\) and \\(\\gamma = 2\\), the difference between the expected and predicted calibration error is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.6, + 0.508, + 0.662 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} c (\\boldsymbol {x}) - u _ {\\mathrm {B S}} (\\hat {p} (\\boldsymbol {x})) = \\mathbb {E} _ {y \\sim \\eta (x)} [ \\| \\hat {p} (\\boldsymbol {x}) - \\boldsymbol {\\eta} (\\boldsymbol {x}) \\| _ {2} ^ {2} - \\| \\hat {p} (\\boldsymbol {x}) - y \\| _ {2} ^ {2} ] \\\\ = \\sum_ {k = 1} ^ {K} \\boldsymbol {\\eta} _ {k} (\\boldsymbol {x}) \\left(\\boldsymbol {\\eta} _ {k} (\\boldsymbol {x}) - 1\\right). \\tag {12} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.735 + ], + "angle": 0, + "content": "Thus, the term \\(\\sum_{k=1}^{K} \\eta_k(x) (\\eta_k(x) - 1)\\) depends solely on \\(\\eta(x)\\) which is fixed for a given sample \\(x\\). Consequently, the error remains constant for a specific sample \\(x\\). We provide a detailed computation of Eq. 12 in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Several variants of \\( u_{\\mathrm{gBS}}(\\hat{p}) \\) have demonstrated their effectiveness in previous studies. When \\( \\beta = 1 \\) and only the actual class is considered, the generalized Brier Score can be interpreted as the scaling term in Focal Loss in Eq. 7. When \\( \\beta = 1 \\) and both the actual class and the maximum predicted class (excluding the actual class) are considered, the generalized Brier Score can be interpreted as the scaling term in Dual Focal Loss in Eq. 8." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "However, these two metrics only consider several classes in the predicted probability. In the case of Focal Loss, which was originally designed for binary classification," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.329, + 0.905, + 0.389 + ], + "angle": 0, + "content": "considering one-dimensional uncertainty is sufficient due to the constraint that the sum of probabilities must equal 1. When extended to multi-class classification, these metrics are not accurate enough to evaluate uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.39, + 0.906, + 0.54 + ], + "angle": 0, + "content": "We visualize the gradient weights \\( u_{\\mathrm{FL}} \\), \\( u_{\\mathrm{DFL}} \\) and \\( u_{\\mathrm{BS}} \\) in Figure 2 for a 4-class scenario. The three axes represent three dimensions of the predicted probability, while the fourth is implied by the probability sum constraint. From the figure, it is evident that \\( u_{\\mathrm{FL}} \\) and \\( u_{\\mathrm{DFL}} \\) are sensitive only to changes in one or two dimensions. However, different points in the coordinate system have different uncertainties. They fail to adequately capture uncertainty change across all dimensions. But the \\( u_{\\mathrm{BS}} \\) can accurately respond to changes along any coordinate axis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.54, + 0.907, + 0.751 + ], + "angle": 0, + "content": "To further evaluate whether these metrics accurately measure the ground truth uncertainty of samples, we conducted experiments on a toy dataset. The toy dataset consists of 5 two-dimensional Gaussian distributions, representing 5 groups of data: \\(\\mathcal{N}(\\mu_i,\\Sigma),i\\in \\{0,\\dots ,5\\}\\). The mean vectors \\(\\mu_{i}\\) were randomly sampled from the range \\([-10,10]\\) and \\(\\Sigma = I\\) was used as the shared covariance matrix for all groups. We generated 10,000 data points from each group to form the training dataset, which was used to train a two layer CNN model for 5 epochs. An additional 1,000 samples from each distribution were used to create the test dataset. Therefore, the ground truth probability \\(\\eta (\\pmb {x})\\) can be computed using the probability density function (PDF) of each distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.763, + 0.905, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\eta (\\boldsymbol {x}) = \\frac {p ^ {n} (\\boldsymbol {x})}{\\sum_ {i} ^ {N} p ^ {i} (\\boldsymbol {x})}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\(N = 5\\) in this case, \\(p^n (\\pmb {x})\\) and \\(p^i (\\pmb {x})\\) are the PDF of corresponds class for \\(\\pmb{x}\\) and remaining classes, respectively. The ground truth uncertainty of each sample is calculated by Eq.1, representing the model's ground truth uncertainty for each sample, along with the sample-wise uncertainty metrics \\(u_{\\mathrm{FL}},u_{\\mathrm{DFL}}\\) and \\(u_{\\mathrm{gBS}}\\). To validate the accuracy" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15501" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.242 + ], + "angle": 0, + "content": "
DatasetModelCEBLMMCEFLSDDFLBSCEBSCE-GRA
Pre TPost TPre TPost TPre TPost TPre TPost TPre TPost TPre TPost TPre TPost T
CIFAR10ResNet504.361.324.291.504.481.411.261.151.001.000.880.880.740.74
ResNet1104.71.564.481.664.801.291.811.171.011.010.990.990.870.87
WideResNet3.350.942.861.103.651.281.841.043.321.161.70.951.461.12
DenseNet4.641.463.961.464.811.671.371.170.870.771.011.010.870.87
CIFAR100ResNet5018.053.057.874.2715.873.325.532.572.542.561.901.901.591.59
ResNet11018.844.6316.774.3018.653.936.883.713.473.472.752.752.532.53
WideResNet14.813.277.744.4614.582.992.702.715.452.522.632.422.462.46
DenseNet19.13.438.132.9917.562.873.381.304.681.831.631.621.621.61
TinyImageNetResNet5014.945.167.811.4714.582.992.182.186.712.284.01.764.571.47
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.253, + 0.907, + 0.294 + ], + "angle": 0, + "content": "Table 1. Comparison of Calibration Methods Using ECE Across Various Datasets and Models. ECE values are reported using 15 bins, with the best-performing method for each dataset-model combination highlighted in bold. Results are averaged over three runs with different random seeds." + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.309, + 0.22, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.309, + 0.357, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.31, + 0.492, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.31, + 0.629, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.31, + 0.766, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.311, + 0.899, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.397, + 0.907, + 0.425 + ], + "angle": 0, + "content": "Figure 3. Comparison of different ECE metrics. The first three plots show the uncertainty for CIFAR-10 using ResNet-50, while the remaining plots represent ResNet-110 on CIFAR-10." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.452, + 0.483, + 0.618 + ], + "angle": 0, + "content": "of these metrics, we used Pearson correlations between the calibrated error and the uncertainty metrics to verify their positive correlation among 5 runs. We use grid-search to find the optimal hyper-parameter for each methods. Among the methods, gBS achieves the highest Pearson correlation coefficient of 0.664, indicating the strongest linear relationship between the predicted values and the true targets. DFL follows with a correlation of 0.638, while FL has the lowest correlation at 0.550. This suggests that gBS provides the most accurate predicted uncertainty alignment with ground-truth uncertainty, followed by DFL and FL." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.619, + 0.483, + 0.785 + ], + "angle": 0, + "content": "Based on the results, we conclude that although ground-truth uncertainty is inaccessible for real-world datasets, these alternative metrics provide a reliable means of estimating uncertainty, making them suitable for use during training. Besides, the metric \\( u_{\\mathrm{gBS}} \\) has the best performance compared to the uncertainty metric used in Focal Loss and Dual Focal Loss. We incorporate \\( u_{\\mathrm{gBS}} \\) into the gradient-weighted framework discussed in Section 4.2 and introduce a new loss function called BSCE-GRA. This loss function uses the generalized Brier Score as an adaptive uncertainty metric to weight the Cross Entropy gradients, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.793, + 0.483, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {B S C E - G R A}} (\\boldsymbol {x}, y) = - \\int \\sum_ {i = 1} ^ {K} u _ {\\mathrm {g B S}} (\\hat {p} (\\boldsymbol {x})) \\cdot \\frac {y _ {i}}{\\hat {p} _ {i} (\\boldsymbol {x})} \\mathrm {d} \\hat {p} (\\boldsymbol {x}). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To further validate the effectiveness of proposed method, we provide a comprehensive theoretical evidence that optimizing with BSCE-GRA, the K-class predicted probability \\(q\\) would equal the actual class-posterior probability" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.906, + 0.496 + ], + "angle": 0, + "content": "\\(\\eta\\), thereby preventing over/under-confidence when convergence. Due to page limitation, the proof is provided in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.906, + 0.588 + ], + "angle": 0, + "content": "Besides, the Uncertainty-GRA Loss makes it possible to use non-differentiable uncertainty metrics for model calibration. However, it requires the weights to capture sample-wise uncertainty, whereas some existing measurements like ECE can only compute the uncertainty for a group of samples, making them unsuitable for the proposed framework." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.618, + 0.638, + 0.635 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.906, + 0.734 + ], + "angle": 0, + "content": "We evaluate our methods on multiple deep neural networks (DNNs), including ResNet50, ResNet110 [6], WideResNet [35], and DenseNet [8]. Our experiments are conducted on CIFAR-10, CIFAR-100 [12], and Tiny-ImageNet [4] to assess calibration performance. Further details about the datasets can be found in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.871 + ], + "angle": 0, + "content": "Baselines. We compare our methods, BSCE, BSCE-GRA, and ECE-CE, with multiple existing approaches, including training with Cross Entropy (CE), Brier Loss (BL)[2], MMCE Loss[15], Focal Loss with Adaptive Exponent (FLSD)[18], and Dual Focal Loss[27]. For Focal Loss, we employ the FLSD-53 strategy [18] to adaptively adjust the gamma value sample-wise, setting \\(\\gamma_{\\mathrm{FL}} = 5\\) for \\(\\hat{p}_c \\in [0,0.2)\\) and \\(\\gamma_{\\mathrm{FL}} = 3\\) for \\(\\hat{p}_c \\in [0.2,1)\\). For Dual Focal Loss, the gamma value is set to 5, as reported in the original work." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Training Setup. Our training setup follows prior works [18, 27]. We train CIFAR-10 and CIFAR-100 for 350" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15502" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.283, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.179, + 0.283, + 0.203 + ], + "angle": 0, + "content": "(a) Gradient norm distribution at epoch 50" + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.089, + 0.49, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.179, + 0.49, + 0.203 + ], + "angle": 0, + "content": "(b) Gradient norm distribution at epoch 150" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.089, + 0.697, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.179, + 0.698, + 0.203 + ], + "angle": 0, + "content": "(c) Gradient norm distribution at epoch 250" + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.089, + 0.903, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.179, + 0.905, + 0.203 + ], + "angle": 0, + "content": "(d) Gradient norm distribution at epoch 350" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.215, + 0.907, + 0.245 + ], + "angle": 0, + "content": "Figure 4. Evolution of gradient norm distributions across different training epochs for various loss functions. The scatter plots show the relationship between gradient norm and Brier Score for different loss functions (Focal Loss, Dual Focal Loss, BSCE-GRA)." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.256, + 0.907, + 0.34 + ], + "angle": 0, + "content": "
MetricsFLSDFLSD-GRADFLDFL-GRABSCEBSCE-GRA
pre Tpost Tpre Tpost Tpre Tpost Tpre Tpost Tpre Tpost Tpre Tpost T
Acc95.04%95.04%94.72%94.72%94.63%94.63%94.76%94.76%95.03%95.03%94.69%94.69%
ECE1.261.150.880.881.001.000.930.840.880.880.740.74
Ada ECE1.561.451.191.191.221.220.770.810.960.960.710.71
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.349, + 0.907, + 0.379 + ], + "angle": 0, + "content": "Table 2. Comparison of weighting different uncertainty metrics on gradient or loss function, including \\( u_{\\mathrm{FL}} \\) and \\( u_{\\mathrm{FL}} \\). The results validate the effectiveness of the gradient-weighting strategy among different uncertainty metrics." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.404, + 0.486, + 0.707 + ], + "angle": 0, + "content": "epochs, using 5,000 images from the training set for validation. The learning rate is initially set to 0.1 for the first 150 epochs, then reduced to 0.01 for the next 100 epochs, and further reduced to 0.001 for the remaining epochs. For Tiny-ImageNet, we train for 100 epochs, with the learning rate set to 0.1 for the first 40 epochs, 0.01 for the next 20 epochs, and 0.001 for the remaining epochs. All experiments are conducted using SGD with a weight decay of \\(5 \\times 10^{-4}\\) and a momentum of 0.9. The training and testing batch sizes for all datasets are set to 128. We re-run all baseline methods using three different random seeds (1, 42, and 71), and report the average results. All experiments are performed on a single Nvidia 4090 GPU. For temperature scaling, the temperature parameter \\(T\\) is optimized through a grid search with \\(T \\in [0, 0.1, 0.2, \\dots, 10]\\) on the validation set, selecting the value that yields the best post-temperature-scaling Expected Calibration Error (ECE). The same optimized temperature parameter is applied to other metrics, such as AdaECE. Further details on the datasets and additional experimental setup are provided in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.725, + 0.32, + 0.74 + ], + "angle": 0, + "content": "5.1. Calibration Performance" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.485, + 0.902 + ], + "angle": 0, + "content": "We report the average ECE before and after temperature scaling among three random seeds, along with the corresponding optimal temperatures, in Table 1. BSCE-GRA achieves state-of-the-art ECE performance in most cases, particularly in the pre-temperature-scaling results. Notably, the fact that most of the optimal temperatures for BSCE-GRA are found to be 1 indicates that BSCE-GRA trains an inherently calibrated model, capable of achieving strong calibration performance without the need for additional temperature scaling. This is a crucial advantage" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.404, + 0.907, + 0.54 + ], + "angle": 0, + "content": "for developing accurate and reliable models that are efficient and require minimal post-processing. The results on CIFAR-10 generally show better calibration performance compared to datasets with more labels (e.g., CIFAR-100 and Tiny-ImageNet) across multiple models. Regarding network architecture, ResNet-50 demonstrates the best calibration performance among the four DNNs tested (ResNet-50, ResNet-110, Wide-ResNet-26-10, and DenseNet-121) on both CIFAR-10 and CIFAR-100 datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.909, + 0.759 + ], + "angle": 0, + "content": "Different Metrics. The methods are further evaluated using several widely-accepted metrics to assess calibration performance across models, including Adaptive ECE and Classwise-ECE. Adaptive ECE measures the expected calibration error while accounting for the distribution of the data, whereas Classwise-ECE is a variant that evaluates calibration error for each class individually. Figure 3 presents the results of multiple methods using ResNet-50 and ResNet-110 on the CIFAR-10 dataset. The figure demonstrates that BSCE-GRA is the only method that achieves both inherently calibrated models and state-of-the-art performance across various metrics. Additional results are presented in the appendix, providing further evidence of the effectiveness of our method in model calibration." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Calibration over Training. Figure 5a presents the ECE on the test set for models trained with Focal Loss and Cross-Entropy loss over the entire training period on CIFAR-10 using ResNet-50. To improve visualization, the ECE values are smoothed using an exponential moving average. The figure suggests that, after the initial warm-up epochs, where predicted probabilities are unstable, the ECE of models trained with BSCE-GRA and Focal Loss consistently remains lower compared to models trained with" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15503" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.341, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.205, + 0.34, + 0.228 + ], + "angle": 0, + "content": "(a) ECE over epochs using CE, Focal Loss and BSCE-GRA." + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.092, + 0.621, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.376, + 0.205, + 0.622, + 0.228 + ], + "angle": 0, + "content": "(b) Gradient Density at epoch 150 using CE, Focal Loss and BSCE-GRA." + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.092, + 0.905, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.205, + 0.905, + 0.228 + ], + "angle": 0, + "content": "(c) Gradient Density at epoch 250 using CE, Focal Loss and BSCE-GRA." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.241, + 0.908, + 0.27 + ], + "angle": 0, + "content": "Figure 5. Figure 5a presents the evolution of ECE throughout the training process, demonstrating that our method rapidly converges to the best result by epoch 250. The subsequent figures depict the gradient magnitudes of various methods between epochs 150 and 250." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.296, + 0.485, + 0.554 + ], + "angle": 0, + "content": "Cross-Entropy. It also indicates that, during training with a moderate learning rate from epochs 150 to 200, Focal Loss tends to produce better-calibrated models than BSCE-GRA. This may be due to the fact that, during these epochs, the model makes more predictions with mid-range confidence levels, and as shown in Figure 1, Focal Loss directs the model's attention towards these moderately uncertain samples. In contrast, BSCE-GRA imposes stronger regularization on the gradient, resulting in smaller optimization steps compared to Focal Loss. To further validate this hypothesis, we present the gradient density of the last linear layer across the entire training set at epochs 150 and 250 in Figure 5b and Fig. 5c, which shows that BSCE-GRA results in smaller gradient magnitudes during these epochs. After 250 epochs, when the learning rate undergoes its second reduction, the model trained with BSCE-GRA soon achieves better calibration performance compared to FL and CE." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.562, + 0.399, + 0.579 + ], + "angle": 0, + "content": "5.2. Gradient Value among Uncertainty" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We extend our analysis to examine the relationship between gradient magnitude and uncertainty. Specifically, we compute the gradient norms of the last linear layer for all samples in the training set at epochs 50, 150, 250, and 350. To quantify uncertainty, we calculate the Brier Score for each training sample. We present the results for Focal Loss (FL), Dual Focal Loss (DFL), and BSCE-GRA. Figure 4 illustrates the relationship between gradient magnitude and Brier Score. The figure clearly shows that the gradients produced by BSCE-GRA are the most sensitive to changes in the Brier Score, as indicated by the narrow distribution compared to other loss functions. This aligns with our goal: we aim for BSCE-GRA to exhibit sufficient sensitivity to uncertainty, allowing the model to adjust its gradient values based on changes in uncertainty, thereby focusing more on highly uncertain samples. Notably, the gradient distribution for Dual Focal Loss is relatively more dispersed concerning the Brier Score, with a wider range of possible values compared to other loss functions. This could be attributed to the partial derivatives involving other classes, as Dual Focal Loss takes the second most probable class into account" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.296, + 0.907, + 0.327 + ], + "angle": 0, + "content": "in its loss calculation. In contrast, Focal Loss and BSCE-GRA only involve gradients along a single dimension." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.353, + 0.836, + 0.37 + ], + "angle": 0, + "content": "5.3. Weighting FL and DFL on Gradients" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.907, + 0.487 + ], + "angle": 0, + "content": "We evaluate the performance of directly applying the uncertainty terms \\( u_{\\mathrm{FL}} \\) and \\( u_{\\mathrm{DFL}} \\) as weights on the gradients. Experiments are conducted using the default settings discussed above, on CIFAR-10 with ResNet-50. The results are presented in Table 2. It is evident that applying weights on gradients results in performance improvement for both \\( u_{\\mathrm{FL}} \\) and \\( u_{\\mathrm{DFL}} \\), further validating its effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.907, + 0.553 + ], + "angle": 0, + "content": "We conduct additional experiments to comprehensively validate our proposed method, BSCE-GRA, under various settings. Due to page limitations, the extended experiments are provided in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.583, + 0.634, + 0.599 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.614, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we proposed a novel approach to model calibration by directly weighting gradients based on uncertainty. We analyzed the strengths and limitations of Focal Loss and Dual Focal Loss from the perspective of sample weighting and introduced a framework that scales gradient magnitudes based on model uncertainty to focus more on uncertain samples. Additionally, we introduced BSCEGRA, a loss function incorporating uncertainty metrics to enhance model calibration. Extensive experiments on various datasets and network architectures demonstrated significant improvements in calibration performance, achieving state-of-the-art results. Our findings emphasize the value of integrating uncertainty-aware mechanisms directly into the optimization process, providing a reliable framework for training calibrated deep neural networks suitable for real-world applications requiring trustworthy predictions. Future work may explore further optimization strategies for uncertainty-weighted approaches and their impact on tasks like active learning and robustness to adversarial attacks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15504" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.095, + 0.09, + 0.261, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.116, + 0.482, + 0.189 + ], + "angle": 0, + "content": "This work was supported in part by the Start-up Grant (No. 9610680) of the City University of Hong Kong, Young Scientist Fund (No. 62406265) of NSFC, and the Australian Research Council under Projects DP240101848 and FT230100549." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.947, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15505" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Ondrej Bohdal, Yongxin Yang, and Timothy Hospedales. Meta-calibration: Learning of model calibration using differentiable expected calibration error. arXiv preprint arXiv:2106.09613, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.171, + 0.484, + 0.212 + ], + "angle": 0, + "content": "[2] Glenn W Brier. Verification of forecasts expressed in terms of probability. Monthly weather review, 78(1):1-3, 1950. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.213, + 0.483, + 0.242 + ], + "angle": 0, + "content": "[3] Corinna Cortes. Support-vector networks. Machine Learning, 1995. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.243, + 0.483, + 0.298 + ], + "angle": 0, + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.299, + 0.483, + 0.354 + ], + "angle": 0, + "content": "[5] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In International conference on machine learning, pages 1321-1330. PMLR, 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.355, + 0.483, + 0.411 + ], + "angle": 0, + "content": "[6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.411, + 0.483, + 0.48 + ], + "angle": 0, + "content": "[7] Dan Hendrycks*, Norman Mu*, Ekin Dogus Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. Aug-mix: A simple method to improve robustness and uncertainty under data shift. In International Conference on Learning Representations, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.483, + 0.537 + ], + "angle": 0, + "content": "[8] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700-4708, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.538, + 0.483, + 0.605 + ], + "angle": 0, + "content": "[9] Like Hui and Mikhail Belkin. {EVALUATION} {of} {neural} {architectures} {trained} {with} {square} {loss} {vs} {cross}-{entropy} {in} {classification} {tasks}. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.607, + 0.483, + 0.675 + ], + "angle": 0, + "content": "[10] Archit Karandikar, Nicholas Cain, Dustin Tran, Balaji Lakshminarayanan, Jonathon Shlens, Michael C Mozer, and Becca Roelofs. Soft calibration objectives for neural networks. Advances in Neural Information Processing Systems, 34:29768-29779, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.677, + 0.483, + 0.732 + ], + "angle": 0, + "content": "[11] Ranganath Krishnan and Omesh Tickoo. Improving model calibration with accuracy versus uncertainty optimization. Advances in Neural Information Processing Systems, 33: 18237-18248, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[12] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[13] Meelis Kull, Telmo Silva Filho, and Peter Flach. Beta calibration: a well-founded and easily implemented improvement on logistic calibration for binary classifiers. In Artificial intelligence and statistics, pages 623-631. PMLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Meelis Kull, Miquel Perello Nieto, Markus Kängsepp, Telmo Silva Filho, Hao Song, and Peter Flach. Beyond temperature scaling: Obtaining well-calibrated multi-class probabilities with dirichlet calibration. Advances in neural information processing systems, 32, 2019. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[15] Aviral Kumar, Sunita Sarawagi, and Ujjwal Jain. Trainable calibration measures for neural networks from kernel mean embeddings. In International Conference on Machine Learning, pages 2805-2814. PMLR, 2018. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.206 + ], + "angle": 0, + "content": "[16] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[17] T Lin. Focal loss for dense object detection. arXiv preprint arXiv:1708.02002, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[18] Jishnu Mukhoti, Viveka Kulharia, Amartya Sanyal, Stuart Golodetz, Philip Torr, and Puneet Dokania. Calibrating deep neural networks using focal loss. Advances in Neural Information Processing Systems, 33:15288-15299, 2020. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.904, + 0.348 + ], + "angle": 0, + "content": "[19] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[20] Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. In Proceedings of the AAAI conference on artificial intelligence, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[21] Khanh Nguyen and Brendan O'Connor. Posterior calibration and exploratory analysis for natural language processing models. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1587-1598, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[22] Joshua C Peterson, Ruairidh M Battleday, Thomas L Grifths, and Olga Russakovsky. Human uncertainty makes classification more robust. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9617-9626, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.905, + 0.589 + ], + "angle": 0, + "content": "[23] John Platt et al. Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Advances in large margin classifiers, 10(3):61-74, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.59, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[24] Rahul Rahaman et al. Uncertainty quantification and deep ensembles. Advances in neural information processing systems, 34:20063-20075, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[25] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[26] Linwei Tao, Minjing Dong, Daochang Liu, Changming Sun, and Chang Xu. Calibrating a deep neural network with its predecessors. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 4271–4279, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[27] Linwei Tao, Minjing Dong, and Chang Xu. Dual focal loss for calibration. In International Conference on Machine Learning, pages 33833-33849. PMLR, 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[28] Linwei Tao, Minjing Dong, and Chang Xu. Feature clipping for uncertainty calibration. arXiv preprint arXiv:2410.19796, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[29] Linwei Tao, Haolan Guo, Minjing Dong, and Chang Xu. Consistency calibration: Improving uncertainty calibration via consistency among perturbed neighbors. arXiv preprint arXiv:2410.12295, 2024. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15506" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[30] Linwei Tao, Younan Zhu, Haolan Guo, Minjing Dong, and Chang Xu. A benchmark study on calibration. In The Twelfth International Conference on Learning Representations, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[31] Sunil Thulasidasan, Gopinath Chennupati, Jeff A Bilmes, Tanmoy Bhattacharya, and Sarah Michalak. On mixup training: Improved calibration and predictive uncertainty for deep neural networks. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[32] Deng-Bao Wang, Lei Feng, and Min-Ling Zhang. Rethinking calibration of deep neural networks: Do not be afraid of overconfidence. Advances in Neural Information Processing Systems, 34:11809-11820, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[33] Jason Weston, Samy Bengio, and Nicolas Usunier. Wsabie: Scaling up to large vocabulary image annotation. In Twenty-Second International Joint Conference on Artificial Intelligence. CiteSeer, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[34] Bianca Zadrozny and Charles Elkan. Transforming classifier scores into accurate multiclass probability estimates. In Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining, pages 694-699, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.403, + 0.482, + 0.444 + ], + "angle": 0, + "content": "[35] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In *British Machine Vision Conference* 2016. British Machine Vision Association, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.482, + 0.501 + ], + "angle": 0, + "content": "[36] Jize Zhang, Bhavya Kailkhura, and T Yong-Jin Han. Mix-n-match: Ensemble and compositional methods for uncertainty calibration in deep learning. In International conference on machine learning, pages 11117-11128. PMLR, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "15507" + } + ] +] \ No newline at end of file diff --git a/2025/Uncertainty Weighted Gradients for Model Calibration/792a362b-85b5-4092-b750-5040ac391b64_origin.pdf b/2025/Uncertainty Weighted Gradients for Model Calibration/792a362b-85b5-4092-b750-5040ac391b64_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..71f9c30a48640acf25a92373b80ccd2406789170 --- /dev/null +++ b/2025/Uncertainty Weighted Gradients for Model Calibration/792a362b-85b5-4092-b750-5040ac391b64_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9976d97d1d25e8ac305b0dadf58b819817903e14a96b6f377c16443ed20c0ec9 +size 2256902 diff --git a/2025/Uncertainty Weighted Gradients for Model Calibration/full.md b/2025/Uncertainty Weighted Gradients for Model Calibration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fe63addb286101431a6604268e5d9a92aa087e99 --- /dev/null +++ b/2025/Uncertainty Weighted Gradients for Model Calibration/full.md @@ -0,0 +1,338 @@ +# Uncertainty Weighted Gradients for Model Calibration + +Jinxu Lin $^{1}$ Linwei Tao $^{1}$ Minjing Dong $^{2}$ Chang Xu $^{1}$ $^{1}$ The University of Sydney, $^{2}$ City University of Hong Kong +{jinxu.lin, linwei.tao, c.xu}@sydney.edu.au, minjdong@cityu.edu.hk + +# Abstract + +Model calibration is essential for ensuring that the predictions of deep neural networks accurately reflect true probabilities in real-world classification tasks. However, deep networks often produce over-confident or under-confident predictions, leading to miscalibration. Various methods have been proposed to address this issue by designing effective loss functions for calibration, such as focal loss. In this paper, we analyze its effectiveness and provide a unified loss framework of focal loss and its variants, where we mainly attribute their superiority in model calibration to the loss weighting factor that estimates sample-wise uncertainty. Based on our analysis, existing loss functions fail to achieve optimal calibration performance due to two main issues: including misalignment during optimization and insufficient precision in uncertainty estimation. Specifically, focal loss cannot align sample uncertainty with gradient scaling and the single logit cannot indicate the uncertainty. To address these issues, we reformulate the optimization from the perspective of gradients, which focuses on uncertain samples. Meanwhile, we propose using the Brier Score as the loss weight factor, which provides a more accurate uncertainty estimation via all the logits. Extensive experiments on various models and datasets demonstrate that our method achieves state-of-the-art (SOTA) performance. $^{1}$ + +# 1. Introduction + +Deep Neural Networks (DNNs) have achieved remarkable success in various domains, including image classification. However, recent studies [5, 30] reveal that DNNs often suffer from mis-calibration in classification task, exhibiting over-confidence or under-confidence in their predictions. For example, a model might output a confidence score of 0.8 for a particular prediction, which does not necessarily correspond to an $80\%$ probability of correctness. + +Calibrating DNNs, which aligns the predicted confidence with the true probabilities, is therefore crucial to en + +hancing their reliability in practical applications, such as autonomous driving, medical imaging, and weather forecasting. To address mis-calibration issues, several methods [26, 28] have been proposed, many of which focus on modifying the loss function during training. A common approach involves adding regularization terms with the Cross-Entropy (CE) loss to improve calibration, as seen in methods like Maximum Mean Calibration Error (MMCE) [15] and Meta Calibration [1]. Additionally, Focal Loss (FL) [17], which adjusts per-sample loss weights based on prediction difficulty, has been shown to improve calibration performance. Mukhoti et al. [18] attribute the effectiveness of FL to its implicit regularization on the entropy of predicted probabilities, which mitigates overconfidence. Similarly, Dual Focal Loss (DFL) [27] incorporates the second-most-probable class in the weighting mechanism to address model under-confidence. Another research interest regarding calibration is the evaluation of calibration performance. Metrics commonly employed include Expected Calibration Error (ECE)[5] and Brier Score[2]. + +The focal-loss-based methods can be unified under a general loss framework expressed as $u \cdot \mathrm{CE}$ , where $u$ is a loss weighting factor. In FL, the factor $u_{\mathrm{FL}}$ can be treated as an estimation for sample-wise difficulty. Interestingly, we observe that $u_{\mathrm{FL}}$ , originally designed for binary classification, is mathematically equivalent to the Brier Score for binary cases, which can measure the uncertainty. Furthermore, the weighting term $u_{\mathrm{DFL}}$ in DFL aligns with the Brier Score in three-class classification scenarios. This observation motivates the hypothesis that weighting the loss with sample calibration metrics, such as the Brier Score, could better identify uncertain samples and facilitate targeted training. + +Furthermore, our investigation reveals a limitation of directly applying the weighting factor $u$ to the loss term. Specifically, the vanilla optimization of focal loss would achieve a misalignment with the objective of scaling gradients for harder samples with larger magnitudes. Based on our analysis, we mainly attribute this issue to a differentiable loss weighting factor $u$ , which could disrupt the positive correlation between the CE and its gradient magnitude during backpropagation, impeding the model's ability + +to prioritize higher-uncertain samples effectively. + +To address these issues, we first propose applying the weighting factor to scaling the gradient rather than the loss itself. This allows us directly aligns the gradient optimization with sample uncertainty, which ensures that harder samples receive appropriately scaled updates without disrupting the optimization process. Within this framework, we then introduce a generalized form of the Brier Score as a gradient weighting factor, which provides a more accurate estimation of sample uncertainty that considers all categories. This leads to the development of the BSCE-GRA loss function, which adjusts gradient to directly scale optimization based on sample uncertainty by Brier Score. + +We conduct extensive experiments to validate the effectiveness of the proposed method, achieving state-of-the-art results across various datasets and model architectures, which demonstrate the effectiveness of our methods. The contributions of this paper are summarized as: + +1. We provide a new perspective by unifying some existing loss modification techniques for model calibration under a sample-weighting framework. With this framework, we provide extensive analysis of their limitations. +2. We propose a simple yet effective optimization framework for model calibration with a gradient weighting factor, where we scale the gradients to encourage the model to focus on uncertain samples effectively. +3. We analyze the use of different uncertainty metrics within this framework and introduce BSCE-GRA, a new loss function based on the Brier Score that provides an accurate sample-wise uncertainty estimation. +4. We conduct extensive experiments under different settings to validate our proposed methods. Our uncertainty-weighted framework shows consistent effectiveness across various uncertainty metrics, and the proposed BSCE-GRA loss achieves state-of-the-art results. + +# 2. Related Works + +In recent years, numerous techniques have been proposed to address the problem of network miscalibration, which can generally be classified into three categories. + +The first category is post-hoc calibration techniques, which adjust model predictions after training by optimizing additional parameters on a held-out validation set. These methods include Platt Scaling [23], which performs a linear transformation on the original prediction logits; Isotonic Regression [34], which uses piecewise functions to transform logits; Bayesian Binning into Quantiles (BBQ)[20], which extends histogram binning with Bayesian model averaging; and Beta Calibration[13], initially designed for binary classification and later generalized to multi-class settings with Dirichlet distributions by Kull et al. [14]. Temperature Scaling [5], one of the most widely used post-hoc calibration methods, tunes the temperature parameter in the + +SoftMax function to minimize negative log-likelihood on a held-out validation set. In this work, we report calibration performance with post-temperature scaling results. + +The second category includes regularization techniques, which are known to effectively calibrate DNNs. Data augmentation methods, such as Mixup [31] and AugMix [7], train DNNs on mixed samples to mitigate overconfident predictions. Model ensemble techniques, which involve independently training multiple DNNs and averaging their predictions, have been shown to enhance both accuracy and predictive uncertainty by aggregating outputs from multiple models [16, 24, 36]. Label smoothing [19], which replaces one-hot labels with soft labels, encourages the model to make less confident predictions, thereby reducing overconfidence. Additionally, weight decay has also been demonstrated to improve confidence calibration [5]. + +The third category focuses on modifying the training loss to improve calibration. These methods include adding a differentiable auxiliary surrogate loss for Expected Calibration Error (ECE)[1, 10, 11] or replacing the training loss with other loss functions, such as Mean Squared Error (MSE)[9], Focal Loss [18], Inverse Focal Loss [32], and Dual Focal Loss [27]. Among these, Focal Loss [18], which adds a modulation term to the Cross-Entropy loss to focus on hard-to-classify samples, provides a simple and effective way to train calibrated models. Focal Loss and Dual Focal Loss can be categorized as margin-based losses, similar to Hinge Loss [3], Triplet Loss [25], and Margin Ranking Loss [33]. + +# 3. Preliminary + +# 3.1. Problem Formulation + +Given a $K$ -class dataset $\mathbb{S} = \{\pmb{z}^{(1)},\dots,\pmb{z}^{(N)}\}$ , where each training sample $\pmb{z}^{(n)}\coloneqq (x^{(n)},y^{(n)})$ is an input-label pair. Let $\mathcal{X}$ and $\mathcal{V}$ represent the input space and the label space, respectively. The ground truth label $y\in \mathcal{V}$ is encoded in a one-hot vector format, where $y_{i} = 1$ if $i\in K$ represents the actual class. We define a classifier $f_{\theta}$ trained on $\mathbb{S}$ that maps an input $\pmb{x}\in \mathcal{X}$ to a probability distribution $\hat{p} (\pmb {x})$ . The classifier then provides a prediction $k = \arg \max_{i\in K}\hat{p}_i$ , which indicates the index of the predicted label. The predicted label $\hat{y}\in \mathcal{Y}$ is similarly represented as $y$ in the one-hot encoding of prediction $k$ . The confidence $\hat{p}_c$ is defined as the predicted probability associated with the prediction $k$ . + +A well-calibrated model ensures that the provided confidence $\hat{p}_c$ accurately reflects the true probability of correct classification. We define the true class-posterior probability vector as $\boldsymbol{\eta}(\boldsymbol{x}) = [\eta_1(\boldsymbol{x}),\dots,\eta_K(\boldsymbol{x})]$ , where $\eta_k(\boldsymbol{x}) = \mathbb{P}(y = k|\boldsymbol{x})$ is the true probability of class $k$ given input $\boldsymbol{x}$ . Formally, a network is considered perfectly calibrated if $\mathbb{P}(\hat{y} = y|\hat{p}_c = p) = p$ for all $p\in [0,1]$ [5], which can be also written as $\hat{p} (\pmb {x}) = \pmb {\eta}(\pmb {x})$ . The uncertainty of a model for a sample $\pmb{x}$ , also referred to the calibrated error $c(x)$ + +can be computed as the difference between $\pmb{\eta}(\pmb{x})$ and $\hat{p}(\pmb{x})$ : + +$$ +c (\boldsymbol {x}) = \| \eta (\boldsymbol {x}) - \hat {p} _ {c} (\boldsymbol {x}) \|. \tag {1} +$$ + +# 3.2. Metric for Calibrated Error + +In practice, it is hard to access the ground truth probability $\pmb{\eta}(\pmb{x})$ on real-world samples and thus the calibration error cannot be directly computed by Eq. 1, some alternative methods have been proposed to evaluate the uncertainty. + +ECE. The Expected Calibration Error (ECE) is defined as $\mathbb{E}_{\hat{p}_c}[\| \mathbb{P}(\hat{y} = y|\hat{p}_c) - \hat{p}_c]$ . Guo et al. [5] propose an approximation of ECE. Specifically, all samples are divided into $M$ bins $\{B_m\}_{m=1}^M$ of equal width based on their confidence, where each bin $B_m$ contains all samples whose confidences fall within the range $\hat{p}_c \in \left[\frac{m}{M}, \frac{m+1}{M}\right)$ . For each bin $B_m$ , the average confidence is computed as $C_m = \frac{1}{|B_m|} \sum_{i \in B_m} \hat{p}_c^{(i)}$ and the bin accuracy is computed as $A_m = \frac{1}{|B_m|} \sum_{i \in B_m} \mathbb{1}(\hat{y}_k^{(i)} = y_k^{(i)})$ , where $\mathbb{1}$ is the indicator function. The ECE can then be computed as the average $L_1$ difference between bin accuracy and confidence: + +$$ +\mathrm {E C E} = \sum_ {m = 1} ^ {M} \frac {\left| B _ {m} \right|}{N} \left| A _ {m} - C _ {m} \right|, \tag {2} +$$ + +where $N$ denotes the number of samples in each bin. In addition to the ECE in Eq.2, there are several variants used to measure calibration error. For example, AdaECE [21] groups samples into bins $B_{m}$ with an equal number of smaples such that $|B_m| = |B_n|$ for all bins. Another variant, ClasswiesECE [14] measure the calibration error by considering each of the $K$ classes separately. + +Brier Score. Besides ECE computing the expectation of calibrated error among whole datasets, some other techniques have also been used to evaluate the sample-wise uncertainty. Accuracy and calibration are distinct concepts, and one cannot be inferred from the other unequivocally. The Brier Score (BS) [2] unify these two concepts, which is commonly used in calibration literature. It has been shown that this family of metrics can be decomposed into a calibration term and a refinement term. Achieving an optimal score requires both accurate predictions and appropriate confidence levels. For a given sample, the Brier Score is mathematically defined as the Mean Squared Error between the predicted probability distribution $\hat{p}$ and one-hot encoded ground truth label $y$ . We introduce a generalized Brier Score (gBS) form as follows: + +$$ +\mathrm {g B S} = \sum_ {i = 1} ^ {K} \| \hat {p} _ {i} (\boldsymbol {x}) - y _ {i} \| _ {\beta} ^ {\gamma} \tag {3} +$$ + +where $\gamma$ and $\beta$ are the hyperparameters for the exponent and norm order, respectively. When $\gamma = 2$ and $\beta = 2$ , this formulation reduces to the original Brier Score. + +# 3.3. Calibrating Method + +Temperature Scaling A widely used post-hoc technique for improving classification calibration is temperature scaling. It adjusts the sharpness of the output probability distribution by scaling the logits in the SoftMax function using a temperature parameter, defined as $\hat{p}_i = \frac{\exp(\hat{g}_i / T)}{\sum_{k=1}^{K} \exp(\hat{g}_k / T)}$ , where $\hat{g}$ represents the logits before applying the SoftMax function, and $T$ is the temperature that controls the scaling. Calibration performance can be enhanced by tuning $T$ on a held-out validation set. + +Focal Loss and Dual Focal Loss Some other previous works have attempted to improve model calibration by modifying the loss function. Focal Loss [17] was initially introduced to address the foreground-background imbalance problem in object detection. It addresses the issue by incorporating a loss weighting factor based on sample complexity, reducing the weight of easy samples and allowing the model to focus on harder-to-classify instances. + +Formally, given the predicted probability $\hat{p} (\pmb {x})$ on sample $\pmb{x}$ , the focal loss is defined as: + +$$ +\mathcal {L} _ {\mathrm {F L}} (\boldsymbol {x}, y) = - \sum_ {i = 1} ^ {K} y _ {i} (1 - \hat {p} _ {i}) ^ {\gamma} \log \hat {p} _ {i} (\boldsymbol {x}), \tag {4} +$$ + +where $\gamma$ is a pre-defined hyperparameter. Previous studies [18] have demonstrated that optimizing models using Focal Loss results in better calibration compared to Cross-Entropy. This improvement is partly due to the entropy-based regularization effect introduced by Focal Loss, while complexity-based weighting also likely plays a role. However, we noticed that Focal Loss was initially applied to binary classification, where the complexity term $(1 - \hat{p}_i)$ is related to the Brier Score, which can also be used for evaluating uncertainty. In multi-class classification, although complexity and uncertainty are not strictly equivalent, the complexity term still captures some class-level uncertainty. + +Several variants of Focal Loss exist, including Dual Focal Loss [27], which incorporates the probability from the second most probable class into the scaling factor to address the under-confidence issue caused by Focal Loss: + +$$ +\mathcal {L} _ {\mathrm {D F L}} (\boldsymbol {x}, y) = - \sum_ {i = 1} ^ {K} y _ {i} (1 - \hat {p} _ {i} (\boldsymbol {x}) + \hat {p} _ {j} (\boldsymbol {x})) ^ {\gamma} \log \hat {p} _ {i} (\boldsymbol {x}), \tag {5} +$$ + +where $\hat{p}_j(x) = \max_i\{\hat{p}_i(x)|\hat{p}_i(x) < \hat{p}_{gt}(x)\}$ + +# 4. Method + +# 4.1. Weighting with Sample-wise Uncertainty + +Focal Loss and Dual Focal Loss both support the concept that uncertainty-based weighting in conjunction with Cross Entropy can enhance model calibration. We propose a generalized loss function framework, termed Uncertainty CE + +Loss, which incorporates a sample-wise uncertainty metric into Cross Entropy as a weighting factor: + +$$ +\mathcal {L} _ {\text {U n c e r t a i n t y}} (\boldsymbol {x}, y) = - \sum_ {i = 1} ^ {K} u \left(\hat {p} _ {i} (\boldsymbol {x})\right) \cdot y _ {i} \log \hat {p} _ {i} (\boldsymbol {x}), \tag {6} +$$ + +where $u(\hat{p}_i)$ is the adaptive term that evaluates the uncertainty of sample $x$ . The motivation behind this design is to use the predicted calibration error for each sample to scale the loss function, thereby directing the model's optimization more effectively towards samples with higher uncertainty. + +Therefore, the weights used in Focal Loss is given by: + +$$ +u _ {\mathrm {F L}} (\hat {p} (\boldsymbol {x})) = (1 - \hat {p} _ {c}) ^ {\gamma}. \tag {7} +$$ + +And the scaling term in Dual Focal Loss is defined as: + +$$ +u _ {\mathrm {D F L}} (\hat {p} (\boldsymbol {x})) = (1 - \hat {p} _ {c} (x) + \hat {p} _ {j} (x)) ^ {\gamma}, \tag {8} +$$ + +where $\hat{p}_j(x)$ is the second maximum confidence in prediction. The weight used in Focal Loss can be seen as derived from the Brier Score, focusing on the error of the ground true class. In contrast, the scaling term in Dual Focal Loss incorporates uncertainty information from the second most probable class. However, directly weighting loss function with certain uncertainty metrics, as those used in Focal Loss and Dual Focal Loss, presents several challenges. + +Reviewing our objectives, we aim to weight the loss function value by each sample's uncertainty, resulting in greater optimization steps for those samples with higher uncertainty. Our focus actually is not on the loss value itself, but on the optimization. The purpose of uncertainty weighting is to focus the model's attention on samples with higher uncertainty. Thus, the key factor is the gradient: higher weights should lead to larger gradients, resulting in more substantial model updates for those samples. For Cross Entropy, its value is positively correlated with the gradient: a higher CE value yields a larger gradient, leading to more significant updates for optimization. When applying a scalar weight to CE, this relationship remains intact, as it does not disrupt the positive correlation. + +However, for differentiable weights, directly applying them to the loss can disrupt the positive correlation between the loss and the gradient, ultimately impairing effective model optimization. To further analyze this, let us consider Focal Loss. For Focal Loss $\mathcal{L}_{\mathrm{FL}}$ and Cross Entropy $\mathcal{L}_{\mathrm{CE}}$ , their gradients are given by: $\frac{\partial}{\partial w}\mathcal{L}_{\mathrm{FL}} = g(\hat{p}_i,\gamma)\frac{\partial}{\partial w}\mathcal{L}_{\mathrm{CE}}$ , where $g(p,\gamma) = (1 - p)^{\gamma} - \gamma p(1 - p)^{\gamma -1}\log (p)$ , $\gamma$ is a predefined hyperparameter, and $\boldsymbol{w}$ represents the parameters of the final linear layer. From a gradient perspective, $g(p,\gamma)$ acts as a weight on the CE gradient, which is illustrated in Figure 1. There exists a point $p_0\in [0,1]$ such that within the range $[0,p_0],\frac{\partial}{\partial p} g(p,\gamma) > 0$ and within the range $[p_0,1],\frac{\partial}{\partial p} g(p,\gamma) < 0$ . In binary classification, $1 - \hat{p}_i$ is related to + +![](images/3b60aa11e502d750f2ad60558338e62d216fc72b2bec4d7340aa3b64027b7fd1.jpg) +Plot of $g(p)$ for Different $\gamma$ Values in Focal Loss +Figure 1. $g(p, \gamma)$ of Focal Loss vs predicted confidence $\hat{p}_c$ . + +the Brier Score that reveals calibrated error, indicating that the predicted calibrated error is not always positively correlated with the weight on gradient. As uncertainty decreases, the weight on gradient initially increases, before eventually aligning with the level of uncertainty. This causes the model to focus more on samples with moderate uncertainty rather than those with the highest uncertainty, which contradicts our original goal of emphasizing the most uncertain samples. Further, changes in uncertainty are not immediately reflected in the sample weights; instead, multiple training iterations are required for these changes to align properly. Thus, directly applying some kind of uncertainty metrics into Eq. 6 would pose issues to the uncertainty weighting. + +# 4.2. Sample-wise Uncertainty on Gradients + +The discussion in Sec. 4.1 motivates us to ensure the weight on gradients for a sample to align with its uncertainty. This could be intuitively solved by directly applying the model uncertainty on samples as a weight of the gradient. We define the modified gradient as: + +$$ +\frac {\partial}{\partial \theta} \mathcal {L} _ {\text {U n c e r t a i n t y - G R A}} (\boldsymbol {x}, y) = u (\hat {p} (\boldsymbol {x})) \frac {\partial}{\partial \theta} \mathcal {L} _ {\mathrm {C E}} (\boldsymbol {x}, y). \tag {9} +$$ + +Taking a SGD optimizer as an example, the model update can be expressed as: $\theta_{t + 1} = \theta +\alpha \cdot \frac{\partial}{\partial\theta}\mathcal{L}_{\mathrm{Uncertainty - GRA}}(\boldsymbol {x},y)$ where $\theta$ is the model parameters and $\alpha$ is the learning rate. When the predicted uncertainty on sample $\pmb{x}$ is higher, the optimization step results in a larger update, encouraging the model to focus more on uncertain samples. A general form of the loss function for our uncertainty-weighted gradient framework, termed Uncertainty-GRA CE Loss, is defined as: + +$$ +\mathcal {L} _ {\text {U n c e r t a i n t y - G R A}} (\boldsymbol {x}, y) = - \int \sum_ {i = 1} ^ {K} u (\hat {p} (\boldsymbol {x})) \cdot \frac {y _ {i}}{\hat {p} _ {i} (\boldsymbol {x})} \mathrm {d} \hat {p} (\boldsymbol {x}), \tag {10} +$$ + +In practice, to compute the $\mathcal{L}_{\mathrm{Uncertainty - GRA}}$ , we can detach the gradient of $u(\hat{p} (\pmb {x}))$ and multiply it with the Cross Entropy, instead of calculating the gradient integration. + +![](images/6924a014be14d1c385e5f18451224290bdbdf369c7df357b0a3ff91bfd7eb54d.jpg) +(a) An illustration of the $u_{\mathrm{FL}}$ + +![](images/452fd0f06819ecff3714e8a32ff5da4ade46079b2ab28b66986f13ffe461ca0b.jpg) +(b) An illustration of the $u_{\mathrm{DFL}}$ . +Figure 2. An illustration of value of gradient weight function on a 4 class classification. It is obvious that $u_{\mathrm{FL}}$ varies only along the $p_i$ axis and $u_{\mathrm{DFL}}$ changes along the $p_i$ and $p_j$ axes. $u_{\mathrm{BS}}$ responds to changes across all axes, providing a more complete uncertainty evaluation. + +![](images/50d805ec096e62c5b10963250182e4b37bf2a9c9de00e3a8313ea604220013c0.jpg) +(c) An illustration of the $u_{\mathrm{BS}}$ + +# 4.3. Sample-wise Uncertainty Metrics + +With Eq. 10, calibration can be achieved by employing the sample's ground-truth uncertainty as the scaling term. However, accessing the ground-truth uncertainty for real-world image samples is impractical. Therefore, alternative metrics are required to estimate the sample-wise uncertainty $u(x)$ . + +One approach to evaluate sample-wise uncertainty is based on computing the generalized Brier Score (gBS): + +$$ +u _ {\mathrm {g B S}} (\hat {p} (\boldsymbol {x})) = \sum_ {i = 1} ^ {K} \| \hat {p} _ {i} (\boldsymbol {x}) - y _ {i} \| _ {\beta} ^ {\gamma}. \tag {11} +$$ + +We provide a brief discussion on the effectiveness of the generalized Brier Score (gBS) as a measure of calibration error. When using the Brier Score (BS) to evaluate calibration error, with $\beta = 2$ and $\gamma = 2$ , the difference between the expected and predicted calibration error is given by: + +$$ +\begin{array}{l} c (\boldsymbol {x}) - u _ {\mathrm {B S}} (\hat {p} (\boldsymbol {x})) = \mathbb {E} _ {y \sim \eta (x)} [ \| \hat {p} (\boldsymbol {x}) - \boldsymbol {\eta} (\boldsymbol {x}) \| _ {2} ^ {2} - \| \hat {p} (\boldsymbol {x}) - y \| _ {2} ^ {2} ] \\ = \sum_ {k = 1} ^ {K} \boldsymbol {\eta} _ {k} (\boldsymbol {x}) \left(\boldsymbol {\eta} _ {k} (\boldsymbol {x}) - 1\right). \tag {12} \\ \end{array} +$$ + +Thus, the term $\sum_{k=1}^{K} \eta_k(x) (\eta_k(x) - 1)$ depends solely on $\eta(x)$ which is fixed for a given sample $x$ . Consequently, the error remains constant for a specific sample $x$ . We provide a detailed computation of Eq. 12 in Appendix. + +Several variants of $u_{\mathrm{gBS}}(\hat{p})$ have demonstrated their effectiveness in previous studies. When $\beta = 1$ and only the actual class is considered, the generalized Brier Score can be interpreted as the scaling term in Focal Loss in Eq. 7. When $\beta = 1$ and both the actual class and the maximum predicted class (excluding the actual class) are considered, the generalized Brier Score can be interpreted as the scaling term in Dual Focal Loss in Eq. 8. + +However, these two metrics only consider several classes in the predicted probability. In the case of Focal Loss, which was originally designed for binary classification, + +considering one-dimensional uncertainty is sufficient due to the constraint that the sum of probabilities must equal 1. When extended to multi-class classification, these metrics are not accurate enough to evaluate uncertainty. + +We visualize the gradient weights $u_{\mathrm{FL}}$ , $u_{\mathrm{DFL}}$ and $u_{\mathrm{BS}}$ in Figure 2 for a 4-class scenario. The three axes represent three dimensions of the predicted probability, while the fourth is implied by the probability sum constraint. From the figure, it is evident that $u_{\mathrm{FL}}$ and $u_{\mathrm{DFL}}$ are sensitive only to changes in one or two dimensions. However, different points in the coordinate system have different uncertainties. They fail to adequately capture uncertainty change across all dimensions. But the $u_{\mathrm{BS}}$ can accurately respond to changes along any coordinate axis. + +To further evaluate whether these metrics accurately measure the ground truth uncertainty of samples, we conducted experiments on a toy dataset. The toy dataset consists of 5 two-dimensional Gaussian distributions, representing 5 groups of data: $\mathcal{N}(\mu_i,\Sigma),i\in \{0,\dots ,5\}$ . The mean vectors $\mu_{i}$ were randomly sampled from the range $[-10,10]$ and $\Sigma = I$ was used as the shared covariance matrix for all groups. We generated 10,000 data points from each group to form the training dataset, which was used to train a two layer CNN model for 5 epochs. An additional 1,000 samples from each distribution were used to create the test dataset. Therefore, the ground truth probability $\eta (\pmb {x})$ can be computed using the probability density function (PDF) of each distribution: + +$$ +\eta (\boldsymbol {x}) = \frac {p ^ {n} (\boldsymbol {x})}{\sum_ {i} ^ {N} p ^ {i} (\boldsymbol {x})}, \tag {13} +$$ + +where $N = 5$ in this case, $p^n (\pmb {x})$ and $p^i (\pmb {x})$ are the PDF of corresponds class for $\pmb{x}$ and remaining classes, respectively. The ground truth uncertainty of each sample is calculated by Eq.1, representing the model's ground truth uncertainty for each sample, along with the sample-wise uncertainty metrics $u_{\mathrm{FL}},u_{\mathrm{DFL}}$ and $u_{\mathrm{gBS}}$ . To validate the accuracy + +
DatasetModelCEBLMMCEFLSDDFLBSCEBSCE-GRA
Pre TPost TPre TPost TPre TPost TPre TPost TPre TPost TPre TPost TPre TPost T
CIFAR10ResNet504.361.324.291.504.481.411.261.151.001.000.880.880.740.74
ResNet1104.71.564.481.664.801.291.811.171.011.010.990.990.870.87
WideResNet3.350.942.861.103.651.281.841.043.321.161.70.951.461.12
DenseNet4.641.463.961.464.811.671.371.170.870.771.011.010.870.87
CIFAR100ResNet5018.053.057.874.2715.873.325.532.572.542.561.901.901.591.59
ResNet11018.844.6316.774.3018.653.936.883.713.473.472.752.752.532.53
WideResNet14.813.277.744.4614.582.992.702.715.452.522.632.422.462.46
DenseNet19.13.438.132.9917.562.873.381.304.681.831.631.621.621.61
TinyImageNetResNet5014.945.167.811.4714.582.992.182.186.712.284.01.764.571.47
+ +Table 1. Comparison of Calibration Methods Using ECE Across Various Datasets and Models. ECE values are reported using 15 bins, with the best-performing method for each dataset-model combination highlighted in bold. Results are averaged over three runs with different random seeds. + +![](images/aa2b34465d503e26385b986d94702aec25884608307ea87fb7ffbe260d14c3fc.jpg) +Figure 3. Comparison of different ECE metrics. The first three plots show the uncertainty for CIFAR-10 using ResNet-50, while the remaining plots represent ResNet-110 on CIFAR-10. + +![](images/b1a0cafc6b558ccd1f52e7538a2fce08d5c2ce37372c99c37642608df313b262.jpg) + +![](images/e8ea14360f6549b8709c684c222ab15b20a8e06334bb4750e505779ca5499dbf.jpg) + +![](images/f7871714057ae1a14355edb664b5c02e74c3151e359785b44af2a093a17fd064.jpg) + +![](images/87d1358feb434fb5f3fffaae88fbbb5e79a53bfb5edcf9f25f820823a2c0073a.jpg) + +![](images/017e996c9608e71808b491d3dbfad8f3c6d79b6a46c14f907f0bb51928b41168.jpg) + +of these metrics, we used Pearson correlations between the calibrated error and the uncertainty metrics to verify their positive correlation among 5 runs. We use grid-search to find the optimal hyper-parameter for each methods. Among the methods, gBS achieves the highest Pearson correlation coefficient of 0.664, indicating the strongest linear relationship between the predicted values and the true targets. DFL follows with a correlation of 0.638, while FL has the lowest correlation at 0.550. This suggests that gBS provides the most accurate predicted uncertainty alignment with ground-truth uncertainty, followed by DFL and FL. + +Based on the results, we conclude that although ground-truth uncertainty is inaccessible for real-world datasets, these alternative metrics provide a reliable means of estimating uncertainty, making them suitable for use during training. Besides, the metric $u_{\mathrm{gBS}}$ has the best performance compared to the uncertainty metric used in Focal Loss and Dual Focal Loss. We incorporate $u_{\mathrm{gBS}}$ into the gradient-weighted framework discussed in Section 4.2 and introduce a new loss function called BSCE-GRA. This loss function uses the generalized Brier Score as an adaptive uncertainty metric to weight the Cross Entropy gradients, defined as: + +$$ +\mathcal {L} _ {\mathrm {B S C E - G R A}} (\boldsymbol {x}, y) = - \int \sum_ {i = 1} ^ {K} u _ {\mathrm {g B S}} (\hat {p} (\boldsymbol {x})) \cdot \frac {y _ {i}}{\hat {p} _ {i} (\boldsymbol {x})} \mathrm {d} \hat {p} (\boldsymbol {x}). \tag {14} +$$ + +To further validate the effectiveness of proposed method, we provide a comprehensive theoretical evidence that optimizing with BSCE-GRA, the K-class predicted probability $q$ would equal the actual class-posterior probability + +$\eta$ , thereby preventing over/under-confidence when convergence. Due to page limitation, the proof is provided in Appendix. + +Besides, the Uncertainty-GRA Loss makes it possible to use non-differentiable uncertainty metrics for model calibration. However, it requires the weights to capture sample-wise uncertainty, whereas some existing measurements like ECE can only compute the uncertainty for a group of samples, making them unsuitable for the proposed framework. + +# 5. Experiment + +We evaluate our methods on multiple deep neural networks (DNNs), including ResNet50, ResNet110 [6], WideResNet [35], and DenseNet [8]. Our experiments are conducted on CIFAR-10, CIFAR-100 [12], and Tiny-ImageNet [4] to assess calibration performance. Further details about the datasets can be found in the appendix. + +Baselines. We compare our methods, BSCE, BSCE-GRA, and ECE-CE, with multiple existing approaches, including training with Cross Entropy (CE), Brier Loss (BL)[2], MMCE Loss[15], Focal Loss with Adaptive Exponent (FLSD)[18], and Dual Focal Loss[27]. For Focal Loss, we employ the FLSD-53 strategy [18] to adaptively adjust the gamma value sample-wise, setting $\gamma_{\mathrm{FL}} = 5$ for $\hat{p}_c \in [0,0.2)$ and $\gamma_{\mathrm{FL}} = 3$ for $\hat{p}_c \in [0.2,1)$ . For Dual Focal Loss, the gamma value is set to 5, as reported in the original work. + +Training Setup. Our training setup follows prior works [18, 27]. We train CIFAR-10 and CIFAR-100 for 350 + +![](images/e353d8870b219815d2cee7f87df3da1e70f68397c8e672e44560104b5602fb24.jpg) +(a) Gradient norm distribution at epoch 50 + +![](images/348cade48d66fc01a8b12e3d67982302da7cf23b4169562c1524d92d12f3cb78.jpg) +(b) Gradient norm distribution at epoch 150 +Figure 4. Evolution of gradient norm distributions across different training epochs for various loss functions. The scatter plots show the relationship between gradient norm and Brier Score for different loss functions (Focal Loss, Dual Focal Loss, BSCE-GRA). + +![](images/a07661b82435e63d1529d9abf5fcea3f2d8cc682cc060d3bdfa3624e1207687b.jpg) +(c) Gradient norm distribution at epoch 250 + +![](images/22beb2d17ba9c0f3ac1c97e94a7f6ca3bb131c95d596ca070edfc7b23ee2476b.jpg) +(d) Gradient norm distribution at epoch 350 + +
MetricsFLSDFLSD-GRADFLDFL-GRABSCEBSCE-GRA
pre Tpost Tpre Tpost Tpre Tpost Tpre Tpost Tpre Tpost Tpre Tpost T
Acc95.04%95.04%94.72%94.72%94.63%94.63%94.76%94.76%95.03%95.03%94.69%94.69%
ECE1.261.150.880.881.001.000.930.840.880.880.740.74
Ada ECE1.561.451.191.191.221.220.770.810.960.960.710.71
+ +Table 2. Comparison of weighting different uncertainty metrics on gradient or loss function, including $u_{\mathrm{FL}}$ and $u_{\mathrm{FL}}$ . The results validate the effectiveness of the gradient-weighting strategy among different uncertainty metrics. + +epochs, using 5,000 images from the training set for validation. The learning rate is initially set to 0.1 for the first 150 epochs, then reduced to 0.01 for the next 100 epochs, and further reduced to 0.001 for the remaining epochs. For Tiny-ImageNet, we train for 100 epochs, with the learning rate set to 0.1 for the first 40 epochs, 0.01 for the next 20 epochs, and 0.001 for the remaining epochs. All experiments are conducted using SGD with a weight decay of $5 \times 10^{-4}$ and a momentum of 0.9. The training and testing batch sizes for all datasets are set to 128. We re-run all baseline methods using three different random seeds (1, 42, and 71), and report the average results. All experiments are performed on a single Nvidia 4090 GPU. For temperature scaling, the temperature parameter $T$ is optimized through a grid search with $T \in [0, 0.1, 0.2, \dots, 10]$ on the validation set, selecting the value that yields the best post-temperature-scaling Expected Calibration Error (ECE). The same optimized temperature parameter is applied to other metrics, such as AdaECE. Further details on the datasets and additional experimental setup are provided in the appendix. + +# 5.1. Calibration Performance + +We report the average ECE before and after temperature scaling among three random seeds, along with the corresponding optimal temperatures, in Table 1. BSCE-GRA achieves state-of-the-art ECE performance in most cases, particularly in the pre-temperature-scaling results. Notably, the fact that most of the optimal temperatures for BSCE-GRA are found to be 1 indicates that BSCE-GRA trains an inherently calibrated model, capable of achieving strong calibration performance without the need for additional temperature scaling. This is a crucial advantage + +for developing accurate and reliable models that are efficient and require minimal post-processing. The results on CIFAR-10 generally show better calibration performance compared to datasets with more labels (e.g., CIFAR-100 and Tiny-ImageNet) across multiple models. Regarding network architecture, ResNet-50 demonstrates the best calibration performance among the four DNNs tested (ResNet-50, ResNet-110, Wide-ResNet-26-10, and DenseNet-121) on both CIFAR-10 and CIFAR-100 datasets. + +Different Metrics. The methods are further evaluated using several widely-accepted metrics to assess calibration performance across models, including Adaptive ECE and Classwise-ECE. Adaptive ECE measures the expected calibration error while accounting for the distribution of the data, whereas Classwise-ECE is a variant that evaluates calibration error for each class individually. Figure 3 presents the results of multiple methods using ResNet-50 and ResNet-110 on the CIFAR-10 dataset. The figure demonstrates that BSCE-GRA is the only method that achieves both inherently calibrated models and state-of-the-art performance across various metrics. Additional results are presented in the appendix, providing further evidence of the effectiveness of our method in model calibration. + +Calibration over Training. Figure 5a presents the ECE on the test set for models trained with Focal Loss and Cross-Entropy loss over the entire training period on CIFAR-10 using ResNet-50. To improve visualization, the ECE values are smoothed using an exponential moving average. The figure suggests that, after the initial warm-up epochs, where predicted probabilities are unstable, the ECE of models trained with BSCE-GRA and Focal Loss consistently remains lower compared to models trained with + +![](images/f226010bb8659982c60fc861a275f563ccbe17bdaea90cba375cd84ba8778ebf.jpg) +(a) ECE over epochs using CE, Focal Loss and BSCE-GRA. + +![](images/b782b0e452c63074844c72f91585b72b3f4f6b60c49927e187f68034089a55ea.jpg) +(b) Gradient Density at epoch 150 using CE, Focal Loss and BSCE-GRA. +Figure 5. Figure 5a presents the evolution of ECE throughout the training process, demonstrating that our method rapidly converges to the best result by epoch 250. The subsequent figures depict the gradient magnitudes of various methods between epochs 150 and 250. + +![](images/1f347dbf8a6d7a902fc393558799e170acc7e9d3bf84dc8076516982c7d6fc3f.jpg) +(c) Gradient Density at epoch 250 using CE, Focal Loss and BSCE-GRA. + +Cross-Entropy. It also indicates that, during training with a moderate learning rate from epochs 150 to 200, Focal Loss tends to produce better-calibrated models than BSCE-GRA. This may be due to the fact that, during these epochs, the model makes more predictions with mid-range confidence levels, and as shown in Figure 1, Focal Loss directs the model's attention towards these moderately uncertain samples. In contrast, BSCE-GRA imposes stronger regularization on the gradient, resulting in smaller optimization steps compared to Focal Loss. To further validate this hypothesis, we present the gradient density of the last linear layer across the entire training set at epochs 150 and 250 in Figure 5b and Fig. 5c, which shows that BSCE-GRA results in smaller gradient magnitudes during these epochs. After 250 epochs, when the learning rate undergoes its second reduction, the model trained with BSCE-GRA soon achieves better calibration performance compared to FL and CE. + +# 5.2. Gradient Value among Uncertainty + +We extend our analysis to examine the relationship between gradient magnitude and uncertainty. Specifically, we compute the gradient norms of the last linear layer for all samples in the training set at epochs 50, 150, 250, and 350. To quantify uncertainty, we calculate the Brier Score for each training sample. We present the results for Focal Loss (FL), Dual Focal Loss (DFL), and BSCE-GRA. Figure 4 illustrates the relationship between gradient magnitude and Brier Score. The figure clearly shows that the gradients produced by BSCE-GRA are the most sensitive to changes in the Brier Score, as indicated by the narrow distribution compared to other loss functions. This aligns with our goal: we aim for BSCE-GRA to exhibit sufficient sensitivity to uncertainty, allowing the model to adjust its gradient values based on changes in uncertainty, thereby focusing more on highly uncertain samples. Notably, the gradient distribution for Dual Focal Loss is relatively more dispersed concerning the Brier Score, with a wider range of possible values compared to other loss functions. This could be attributed to the partial derivatives involving other classes, as Dual Focal Loss takes the second most probable class into account + +in its loss calculation. In contrast, Focal Loss and BSCE-GRA only involve gradients along a single dimension. + +# 5.3. Weighting FL and DFL on Gradients + +We evaluate the performance of directly applying the uncertainty terms $u_{\mathrm{FL}}$ and $u_{\mathrm{DFL}}$ as weights on the gradients. Experiments are conducted using the default settings discussed above, on CIFAR-10 with ResNet-50. The results are presented in Table 2. It is evident that applying weights on gradients results in performance improvement for both $u_{\mathrm{FL}}$ and $u_{\mathrm{DFL}}$ , further validating its effectiveness. + +We conduct additional experiments to comprehensively validate our proposed method, BSCE-GRA, under various settings. Due to page limitations, the extended experiments are provided in the Appendix. + +# 6. Conclusion + +In this paper, we proposed a novel approach to model calibration by directly weighting gradients based on uncertainty. We analyzed the strengths and limitations of Focal Loss and Dual Focal Loss from the perspective of sample weighting and introduced a framework that scales gradient magnitudes based on model uncertainty to focus more on uncertain samples. Additionally, we introduced BSCEGRA, a loss function incorporating uncertainty metrics to enhance model calibration. Extensive experiments on various datasets and network architectures demonstrated significant improvements in calibration performance, achieving state-of-the-art results. Our findings emphasize the value of integrating uncertainty-aware mechanisms directly into the optimization process, providing a reliable framework for training calibrated deep neural networks suitable for real-world applications requiring trustworthy predictions. Future work may explore further optimization strategies for uncertainty-weighted approaches and their impact on tasks like active learning and robustness to adversarial attacks. + +# 7. Acknowledgment + +This work was supported in part by the Start-up Grant (No. 9610680) of the City University of Hong Kong, Young Scientist Fund (No. 62406265) of NSFC, and the Australian Research Council under Projects DP240101848 and FT230100549. + +# References + +[1] Ondrej Bohdal, Yongxin Yang, and Timothy Hospedales. Meta-calibration: Learning of model calibration using differentiable expected calibration error. arXiv preprint arXiv:2106.09613, 2021. 1, 2 +[2] Glenn W Brier. Verification of forecasts expressed in terms of probability. Monthly weather review, 78(1):1-3, 1950. 1, 3, 6 +[3] Corinna Cortes. Support-vector networks. Machine Learning, 1995. 2 +[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6, 1 +[5] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In International conference on machine learning, pages 1321-1330. PMLR, 2017. 1, 2, 3 +[6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[7] Dan Hendrycks*, Norman Mu*, Ekin Dogus Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. Aug-mix: A simple method to improve robustness and uncertainty under data shift. In International Conference on Learning Representations, 2020. 2 +[8] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700-4708, 2017. 6 +[9] Like Hui and Mikhail Belkin. {EVALUATION} {of} {neural} {architectures} {trained} {with} {square} {loss} {vs} {cross}-{entropy} {in} {classification} {tasks}. In International Conference on Learning Representations, 2021. 2 +[10] Archit Karandikar, Nicholas Cain, Dustin Tran, Balaji Lakshminarayanan, Jonathon Shlens, Michael C Mozer, and Becca Roelofs. Soft calibration objectives for neural networks. Advances in Neural Information Processing Systems, 34:29768-29779, 2021. 2 +[11] Ranganath Krishnan and Omesh Tickoo. Improving model calibration with accuracy versus uncertainty optimization. Advances in Neural Information Processing Systems, 33: 18237-18248, 2020. 2 +[12] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6, 1 +[13] Meelis Kull, Telmo Silva Filho, and Peter Flach. Beta calibration: a well-founded and easily implemented improvement on logistic calibration for binary classifiers. In Artificial intelligence and statistics, pages 623-631. PMLR, 2017. 2 +[14] Meelis Kull, Miquel Perello Nieto, Markus Kängsepp, Telmo Silva Filho, Hao Song, and Peter Flach. Beyond temperature scaling: Obtaining well-calibrated multi-class probabilities with dirichlet calibration. Advances in neural information processing systems, 32, 2019. 2, 3 + +[15] Aviral Kumar, Sunita Sarawagi, and Ujjwal Jain. Trainable calibration measures for neural networks from kernel mean embeddings. In International Conference on Machine Learning, pages 2805-2814. PMLR, 2018. 1, 6 +[16] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems, 30, 2017. 2 +[17] T Lin. Focal loss for dense object detection. arXiv preprint arXiv:1708.02002, 2017. 1, 3 +[18] Jishnu Mukhoti, Viveka Kulharia, Amartya Sanyal, Stuart Golodetz, Philip Torr, and Puneet Dokania. Calibrating deep neural networks using focal loss. Advances in Neural Information Processing Systems, 33:15288-15299, 2020. 1, 2, 3, 6 +[19] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? Advances in neural information processing systems, 32, 2019. 2 +[20] Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. In Proceedings of the AAAI conference on artificial intelligence, 2015. 2 +[21] Khanh Nguyen and Brendan O'Connor. Posterior calibration and exploratory analysis for natural language processing models. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1587-1598, 2015. 3 +[22] Joshua C Peterson, Ruairidh M Battleday, Thomas L Grifths, and Olga Russakovsky. Human uncertainty makes classification more robust. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9617-9626, 2019. 6 +[23] John Platt et al. Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Advances in large margin classifiers, 10(3):61-74, 1999. 2 +[24] Rahul Rahaman et al. Uncertainty quantification and deep ensembles. Advances in neural information processing systems, 34:20063-20075, 2021. 2 +[25] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 2 +[26] Linwei Tao, Minjing Dong, Daochang Liu, Changming Sun, and Chang Xu. Calibrating a deep neural network with its predecessors. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 4271–4279, 2023. 1 +[27] Linwei Tao, Minjing Dong, and Chang Xu. Dual focal loss for calibration. In International Conference on Machine Learning, pages 33833-33849. PMLR, 2023. 1, 2, 3, 6 +[28] Linwei Tao, Minjing Dong, and Chang Xu. Feature clipping for uncertainty calibration. arXiv preprint arXiv:2410.19796, 2024. 1 +[29] Linwei Tao, Haolan Guo, Minjing Dong, and Chang Xu. Consistency calibration: Improving uncertainty calibration via consistency among perturbed neighbors. arXiv preprint arXiv:2410.12295, 2024. 6 + +[30] Linwei Tao, Younan Zhu, Haolan Guo, Minjing Dong, and Chang Xu. A benchmark study on calibration. In The Twelfth International Conference on Learning Representations, 2024. 1 +[31] Sunil Thulasidasan, Gopinath Chennupati, Jeff A Bilmes, Tanmoy Bhattacharya, and Sarah Michalak. On mixup training: Improved calibration and predictive uncertainty for deep neural networks. Advances in neural information processing systems, 32, 2019. 2 +[32] Deng-Bao Wang, Lei Feng, and Min-Ling Zhang. Rethinking calibration of deep neural networks: Do not be afraid of overconfidence. Advances in Neural Information Processing Systems, 34:11809-11820, 2021. 2 +[33] Jason Weston, Samy Bengio, and Nicolas Usunier. Wsabie: Scaling up to large vocabulary image annotation. In Twenty-Second International Joint Conference on Artificial Intelligence. CiteSeer, 2011. 2 +[34] Bianca Zadrozny and Charles Elkan. Transforming classifier scores into accurate multiclass probability estimates. In Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining, pages 694-699, 2002. 2 +[35] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In *British Machine Vision Conference* 2016. British Machine Vision Association, 2016. 6 +[36] Jize Zhang, Bhavya Kailkhura, and T Yong-Jin Han. Mix-n-match: Ensemble and compositional methods for uncertainty calibration in deep learning. In International conference on machine learning, pages 11117-11128. PMLR, 2020. 2 \ No newline at end of file diff --git a/2025/Uncertainty Weighted Gradients for Model Calibration/images.zip b/2025/Uncertainty Weighted Gradients for Model Calibration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..20de56a9b1ff11ccf0515d5b62fef12cbc0fc286 --- /dev/null +++ b/2025/Uncertainty Weighted Gradients for Model Calibration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9060075b2987ebf21f188e9bf60574f08ea3e0d773aee7ab9a5803526cb06120 +size 393255 diff --git a/2025/Uncertainty Weighted Gradients for Model Calibration/layout.json b/2025/Uncertainty Weighted Gradients for Model Calibration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3129ed0dff8dc51c45c9567c249784fddfdc6fb2 --- /dev/null +++ b/2025/Uncertainty Weighted Gradients for Model Calibration/layout.json @@ -0,0 +1,9842 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 103, + 477, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 103, + 477, + 121 + ], + "spans": [ + { + "bbox": [ + 134, + 103, + 477, + 121 + ], + "type": "text", + "content": "Uncertainty Weighted Gradients for Model Calibration" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 142, + 482, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 142, + 482, + 185 + ], + "spans": [ + { + "bbox": [ + 126, + 142, + 482, + 185 + ], + "type": "text", + "content": "Jinxu Lin" + }, + { + "bbox": [ + 126, + 142, + 482, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 126, + 142, + 482, + 185 + ], + "type": "text", + "content": " Linwei Tao" + }, + { + "bbox": [ + 126, + 142, + 482, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 126, + 142, + 482, + 185 + ], + "type": "text", + "content": " Minjing Dong" + }, + { + "bbox": [ + 126, + 142, + 482, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 126, + 142, + 482, + 185 + ], + "type": "text", + "content": " Chang Xu" + }, + { + "bbox": [ + 126, + 142, + 482, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 126, + 142, + 482, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 126, + 142, + 482, + 185 + ], + "type": "text", + "content": "The University of Sydney, " + }, + { + "bbox": [ + 126, + 142, + 482, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 126, + 142, + 482, + 185 + ], + "type": "text", + "content": "City University of Hong Kong \n{jinxu.lin, linwei.tao, c.xu}@sydney.edu.au, minjdong@cityu.edu.hk" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 213, + 201, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 201, + 227 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 201, + 227 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 239, + 297, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 297, + 526 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 297, + 526 + ], + "type": "text", + "content": "Model calibration is essential for ensuring that the predictions of deep neural networks accurately reflect true probabilities in real-world classification tasks. However, deep networks often produce over-confident or under-confident predictions, leading to miscalibration. Various methods have been proposed to address this issue by designing effective loss functions for calibration, such as focal loss. In this paper, we analyze its effectiveness and provide a unified loss framework of focal loss and its variants, where we mainly attribute their superiority in model calibration to the loss weighting factor that estimates sample-wise uncertainty. Based on our analysis, existing loss functions fail to achieve optimal calibration performance due to two main issues: including misalignment during optimization and insufficient precision in uncertainty estimation. Specifically, focal loss cannot align sample uncertainty with gradient scaling and the single logit cannot indicate the uncertainty. To address these issues, we reformulate the optimization from the perspective of gradients, which focuses on uncertain samples. Meanwhile, we propose using the Brier Score as the loss weight factor, which provides a more accurate uncertainty estimation via all the logits. Extensive experiments on various models and datasets demonstrate that our method achieves state-of-the-art (SOTA) performance." + }, + { + "bbox": [ + 53, + 239, + 297, + 526 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 551, + 137, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 137, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 137, + 563 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 571, + 295, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 295, + 667 + ], + "type": "text", + "content": "Deep Neural Networks (DNNs) have achieved remarkable success in various domains, including image classification. However, recent studies [5, 30] reveal that DNNs often suffer from mis-calibration in classification task, exhibiting over-confidence or under-confidence in their predictions. For example, a model might output a confidence score of 0.8 for a particular prediction, which does not necessarily correspond to an " + }, + { + "bbox": [ + 55, + 571, + 295, + 667 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 55, + 571, + 295, + 667 + ], + "type": "text", + "content": " probability of correctness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 668, + 295, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 668, + 295, + 692 + ], + "spans": [ + { + "bbox": [ + 55, + 668, + 295, + 692 + ], + "type": "text", + "content": "Calibrating DNNs, which aligns the predicted confidence with the true probabilities, is therefore crucial to en" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 213, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 453 + ], + "type": "text", + "content": "hancing their reliability in practical applications, such as autonomous driving, medical imaging, and weather forecasting. To address mis-calibration issues, several methods [26, 28] have been proposed, many of which focus on modifying the loss function during training. A common approach involves adding regularization terms with the Cross-Entropy (CE) loss to improve calibration, as seen in methods like Maximum Mean Calibration Error (MMCE) [15] and Meta Calibration [1]. Additionally, Focal Loss (FL) [17], which adjusts per-sample loss weights based on prediction difficulty, has been shown to improve calibration performance. Mukhoti et al. [18] attribute the effectiveness of FL to its implicit regularization on the entropy of predicted probabilities, which mitigates overconfidence. Similarly, Dual Focal Loss (DFL) [27] incorporates the second-most-probable class in the weighting mechanism to address model under-confidence. Another research interest regarding calibration is the evaluation of calibration performance. Metrics commonly employed include Expected Calibration Error (ECE)[5] and Brier Score[2]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 457, + 556, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 556, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 556, + 602 + ], + "type": "text", + "content": "The focal-loss-based methods can be unified under a general loss framework expressed as " + }, + { + "bbox": [ + 313, + 457, + 556, + 602 + ], + "type": "inline_equation", + "content": "u \\cdot \\mathrm{CE}" + }, + { + "bbox": [ + 313, + 457, + 556, + 602 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 457, + 556, + 602 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 313, + 457, + 556, + 602 + ], + "type": "text", + "content": " is a loss weighting factor. In FL, the factor " + }, + { + "bbox": [ + 313, + 457, + 556, + 602 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{FL}}" + }, + { + "bbox": [ + 313, + 457, + 556, + 602 + ], + "type": "text", + "content": " can be treated as an estimation for sample-wise difficulty. Interestingly, we observe that " + }, + { + "bbox": [ + 313, + 457, + 556, + 602 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{FL}}" + }, + { + "bbox": [ + 313, + 457, + 556, + 602 + ], + "type": "text", + "content": ", originally designed for binary classification, is mathematically equivalent to the Brier Score for binary cases, which can measure the uncertainty. Furthermore, the weighting term " + }, + { + "bbox": [ + 313, + 457, + 556, + 602 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{DFL}}" + }, + { + "bbox": [ + 313, + 457, + 556, + 602 + ], + "type": "text", + "content": " in DFL aligns with the Brier Score in three-class classification scenarios. This observation motivates the hypothesis that weighting the loss with sample calibration metrics, such as the Brier Score, could better identify uncertain samples and facilitate targeted training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Furthermore, our investigation reveals a limitation of directly applying the weighting factor " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": " to the loss term. Specifically, the vanilla optimization of focal loss would achieve a misalignment with the objective of scaling gradients for harder samples with larger magnitudes. Based on our analysis, we mainly attribute this issue to a differentiable loss weighting factor " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": ", which could disrupt the positive correlation between the CE and its gradient magnitude during backpropagation, impeding the model's ability" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 272, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 272, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 272, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 702, + 272, + 713 + ], + "type": "text", + "content": "Code is available at https://github.com/Jinxu-Lin/BSCE-GRA." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15497" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 254, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 254, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 254, + 83 + ], + "type": "text", + "content": "to prioritize higher-uncertain samples effectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 84, + 294, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 84, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 84, + 294, + 227 + ], + "type": "text", + "content": "To address these issues, we first propose applying the weighting factor to scaling the gradient rather than the loss itself. This allows us directly aligns the gradient optimization with sample uncertainty, which ensures that harder samples receive appropriately scaled updates without disrupting the optimization process. Within this framework, we then introduce a generalized form of the Brier Score as a gradient weighting factor, which provides a more accurate estimation of sample uncertainty that considers all categories. This leads to the development of the BSCE-GRA loss function, which adjusts gradient to directly scale optimization based on sample uncertainty by Brier Score." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 228, + 294, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 294, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 294, + 287 + ], + "type": "text", + "content": "We conduct extensive experiments to validate the effectiveness of the proposed method, achieving state-of-the-art results across various datasets and model architectures, which demonstrate the effectiveness of our methods. The contributions of this paper are summarized as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 289, + 295, + 491 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 289, + 295, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 289, + 295, + 336 + ], + "spans": [ + { + "bbox": [ + 56, + 289, + 295, + 336 + ], + "type": "text", + "content": "1. We provide a new perspective by unifying some existing loss modification techniques for model calibration under a sample-weighting framework. With this framework, we provide extensive analysis of their limitations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 337, + 294, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 294, + 384 + ], + "type": "text", + "content": "2. We propose a simple yet effective optimization framework for model calibration with a gradient weighting factor, where we scale the gradients to encourage the model to focus on uncertain samples effectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 384, + 294, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 294, + 432 + ], + "type": "text", + "content": "3. We analyze the use of different uncertainty metrics within this framework and introduce BSCE-GRA, a new loss function based on the Brier Score that provides an accurate sample-wise uncertainty estimation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 433, + 294, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 294, + 491 + ], + "type": "text", + "content": "4. We conduct extensive experiments under different settings to validate our proposed methods. Our uncertainty-weighted framework shows consistent effectiveness across various uncertainty metrics, and the proposed BSCE-GRA loss achieves state-of-the-art results." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 502, + 147, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 147, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 147, + 514 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 522, + 295, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 558 + ], + "type": "text", + "content": "In recent years, numerous techniques have been proposed to address the problem of network miscalibration, which can generally be classified into three categories." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "The first category is post-hoc calibration techniques, which adjust model predictions after training by optimizing additional parameters on a held-out validation set. These methods include Platt Scaling [23], which performs a linear transformation on the original prediction logits; Isotonic Regression [34], which uses piecewise functions to transform logits; Bayesian Binning into Quantiles (BBQ)[20], which extends histogram binning with Bayesian model averaging; and Beta Calibration[13], initially designed for binary classification and later generalized to multi-class settings with Dirichlet distributions by Kull et al. [14]. Temperature Scaling [5], one of the most widely used post-hoc calibration methods, tunes the temperature parameter in the" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "SoftMax function to minimize negative log-likelihood on a held-out validation set. In this work, we report calibration performance with post-temperature scaling results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 109, + 553, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 553, + 264 + ], + "type": "text", + "content": "The second category includes regularization techniques, which are known to effectively calibrate DNNs. Data augmentation methods, such as Mixup [31] and AugMix [7], train DNNs on mixed samples to mitigate overconfident predictions. Model ensemble techniques, which involve independently training multiple DNNs and averaging their predictions, have been shown to enhance both accuracy and predictive uncertainty by aggregating outputs from multiple models [16, 24, 36]. Label smoothing [19], which replaces one-hot labels with soft labels, encourages the model to make less confident predictions, thereby reducing overconfidence. Additionally, weight decay has also been demonstrated to improve confidence calibration [5]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 266, + 553, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 553, + 410 + ], + "type": "text", + "content": "The third category focuses on modifying the training loss to improve calibration. These methods include adding a differentiable auxiliary surrogate loss for Expected Calibration Error (ECE)[1, 10, 11] or replacing the training loss with other loss functions, such as Mean Squared Error (MSE)[9], Focal Loss [18], Inverse Focal Loss [32], and Dual Focal Loss [27]. Among these, Focal Loss [18], which adds a modulation term to the Cross-Entropy loss to focus on hard-to-classify samples, provides a simple and effective way to train calibrated models. Focal Loss and Dual Focal Loss can be categorized as margin-based losses, similar to Hinge Loss [3], Triplet Loss [25], and Margin Ranking Loss [33]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 422, + 391, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 391, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 391, + 436 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 443, + 439, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 439, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 439, + 455 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "text", + "content": "Given a " + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "text", + "content": "-class dataset " + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\mathbb{S} = \\{\\pmb{z}^{(1)},\\dots,\\pmb{z}^{(N)}\\}" + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "text", + "content": ", where each training sample " + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{z}^{(n)}\\coloneqq (x^{(n)},y^{(n)})" + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "text", + "content": " is an input-label pair. Let " + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "text", + "content": " represent the input space and the label space, respectively. The ground truth label " + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "inline_equation", + "content": "y\\in \\mathcal{V}" + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "text", + "content": " is encoded in a one-hot vector format, where " + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "inline_equation", + "content": "y_{i} = 1" + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "inline_equation", + "content": "i\\in K" + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "text", + "content": " represents the actual class. We define a classifier " + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "text", + "content": " trained on " + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\mathbb{S}" + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "text", + "content": " that maps an input " + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{x}\\in \\mathcal{X}" + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "text", + "content": " to a probability distribution " + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\hat{p} (\\pmb {x})" + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "text", + "content": ". The classifier then provides a prediction " + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "inline_equation", + "content": "k = \\arg \\max_{i\\in K}\\hat{p}_i" + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "text", + "content": ", which indicates the index of the predicted label. The predicted label " + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\hat{y}\\in \\mathcal{Y}" + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "text", + "content": " is similarly represented as " + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "text", + "content": " in the one-hot encoding of prediction " + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "text", + "content": ". The confidence " + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\hat{p}_c" + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "text", + "content": " is defined as the predicted probability associated with the prediction " + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 460, + 553, + 605 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": "A well-calibrated model ensures that the provided confidence " + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\hat{p}_c" + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": " accurately reflects the true probability of correct classification. We define the true class-posterior probability vector as " + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\eta}(\\boldsymbol{x}) = [\\eta_1(\\boldsymbol{x}),\\dots,\\eta_K(\\boldsymbol{x})]" + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\eta_k(\\boldsymbol{x}) = \\mathbb{P}(y = k|\\boldsymbol{x})" + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": " is the true probability of class " + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": " given input " + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}" + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": ". Formally, a network is considered perfectly calibrated if " + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{P}(\\hat{y} = y|\\hat{p}_c = p) = p" + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "inline_equation", + "content": "p\\in [0,1]" + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": " [5], which can be also written as " + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\hat{p} (\\pmb {x}) = \\pmb {\\eta}(\\pmb {x})" + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": ". The uncertainty of a model for a sample " + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": ", also referred to the calibrated error " + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "inline_equation", + "content": "c(x)" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15498" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 85 + ], + "type": "text", + "content": "can be computed as the difference between " + }, + { + "bbox": [ + 55, + 72, + 295, + 85 + ], + "type": "inline_equation", + "content": "\\pmb{\\eta}(\\pmb{x})" + }, + { + "bbox": [ + 55, + 72, + 295, + 85 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 295, + 85 + ], + "type": "inline_equation", + "content": "\\hat{p}(\\pmb{x})" + }, + { + "bbox": [ + 55, + 72, + 295, + 85 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 123, + 95, + 295, + 110 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 95, + 295, + 110 + ], + "spans": [ + { + "bbox": [ + 123, + 95, + 295, + 110 + ], + "type": "interline_equation", + "content": "c (\\boldsymbol {x}) = \\| \\eta (\\boldsymbol {x}) - \\hat {p} _ {c} (\\boldsymbol {x}) \\|. \\tag {1}", + "image_path": "c7d2c8881e7ef5808dc16db46e958966d22dd161e14dbb6c3856725d089e22f2.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 115, + 211, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 115, + 211, + 127 + ], + "spans": [ + { + "bbox": [ + 55, + 115, + 211, + 127 + ], + "type": "text", + "content": "3.2. Metric for Calibrated Error" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 133, + 295, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 180 + ], + "type": "text", + "content": "In practice, it is hard to access the ground truth probability " + }, + { + "bbox": [ + 55, + 133, + 295, + 180 + ], + "type": "inline_equation", + "content": "\\pmb{\\eta}(\\pmb{x})" + }, + { + "bbox": [ + 55, + 133, + 295, + 180 + ], + "type": "text", + "content": " on real-world samples and thus the calibration error cannot be directly computed by Eq. 1, some alternative methods have been proposed to evaluate the uncertainty." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 181, + 296, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 181, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 181, + 296, + 319 + ], + "type": "text", + "content": "ECE. The Expected Calibration Error (ECE) is defined as " + }, + { + "bbox": [ + 55, + 181, + 296, + 319 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\hat{p}_c}[\\| \\mathbb{P}(\\hat{y} = y|\\hat{p}_c) - \\hat{p}_c]" + }, + { + "bbox": [ + 55, + 181, + 296, + 319 + ], + "type": "text", + "content": ". Guo et al. [5] propose an approximation of ECE. Specifically, all samples are divided into " + }, + { + "bbox": [ + 55, + 181, + 296, + 319 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 181, + 296, + 319 + ], + "type": "text", + "content": " bins " + }, + { + "bbox": [ + 55, + 181, + 296, + 319 + ], + "type": "inline_equation", + "content": "\\{B_m\\}_{m=1}^M" + }, + { + "bbox": [ + 55, + 181, + 296, + 319 + ], + "type": "text", + "content": " of equal width based on their confidence, where each bin " + }, + { + "bbox": [ + 55, + 181, + 296, + 319 + ], + "type": "inline_equation", + "content": "B_m" + }, + { + "bbox": [ + 55, + 181, + 296, + 319 + ], + "type": "text", + "content": " contains all samples whose confidences fall within the range " + }, + { + "bbox": [ + 55, + 181, + 296, + 319 + ], + "type": "inline_equation", + "content": "\\hat{p}_c \\in \\left[\\frac{m}{M}, \\frac{m+1}{M}\\right)" + }, + { + "bbox": [ + 55, + 181, + 296, + 319 + ], + "type": "text", + "content": ". For each bin " + }, + { + "bbox": [ + 55, + 181, + 296, + 319 + ], + "type": "inline_equation", + "content": "B_m" + }, + { + "bbox": [ + 55, + 181, + 296, + 319 + ], + "type": "text", + "content": ", the average confidence is computed as " + }, + { + "bbox": [ + 55, + 181, + 296, + 319 + ], + "type": "inline_equation", + "content": "C_m = \\frac{1}{|B_m|} \\sum_{i \\in B_m} \\hat{p}_c^{(i)}" + }, + { + "bbox": [ + 55, + 181, + 296, + 319 + ], + "type": "text", + "content": " and the bin accuracy is computed as " + }, + { + "bbox": [ + 55, + 181, + 296, + 319 + ], + "type": "inline_equation", + "content": "A_m = \\frac{1}{|B_m|} \\sum_{i \\in B_m} \\mathbb{1}(\\hat{y}_k^{(i)} = y_k^{(i)})" + }, + { + "bbox": [ + 55, + 181, + 296, + 319 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 181, + 296, + 319 + ], + "type": "inline_equation", + "content": "\\mathbb{1}" + }, + { + "bbox": [ + 55, + 181, + 296, + 319 + ], + "type": "text", + "content": " is the indicator function. The ECE can then be computed as the average " + }, + { + "bbox": [ + 55, + 181, + 296, + 319 + ], + "type": "inline_equation", + "content": "L_1" + }, + { + "bbox": [ + 55, + 181, + 296, + 319 + ], + "type": "text", + "content": " difference between bin accuracy and confidence:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 111, + 328, + 295, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 328, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 111, + 328, + 295, + 361 + ], + "type": "interline_equation", + "content": "\\mathrm {E C E} = \\sum_ {m = 1} ^ {M} \\frac {\\left| B _ {m} \\right|}{N} \\left| A _ {m} - C _ {m} \\right|, \\tag {2}", + "image_path": "d310c1f6ab7f7624472b337790c79bd916041c0ef7b947d43f7b7f42245a8e75.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 365, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 449 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 365, + 296, + 449 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 365, + 296, + 449 + ], + "type": "text", + "content": " denotes the number of samples in each bin. In addition to the ECE in Eq.2, there are several variants used to measure calibration error. For example, AdaECE [21] groups samples into bins " + }, + { + "bbox": [ + 55, + 365, + 296, + 449 + ], + "type": "inline_equation", + "content": "B_{m}" + }, + { + "bbox": [ + 55, + 365, + 296, + 449 + ], + "type": "text", + "content": " with an equal number of smaples such that " + }, + { + "bbox": [ + 55, + 365, + 296, + 449 + ], + "type": "inline_equation", + "content": "|B_m| = |B_n|" + }, + { + "bbox": [ + 55, + 365, + 296, + 449 + ], + "type": "text", + "content": " for all bins. Another variant, ClasswiesECE [14] measure the calibration error by considering each of the " + }, + { + "bbox": [ + 55, + 365, + 296, + 449 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 365, + 296, + 449 + ], + "type": "text", + "content": " classes separately." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 449, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 628 + ], + "type": "text", + "content": "Brier Score. Besides ECE computing the expectation of calibrated error among whole datasets, some other techniques have also been used to evaluate the sample-wise uncertainty. Accuracy and calibration are distinct concepts, and one cannot be inferred from the other unequivocally. The Brier Score (BS) [2] unify these two concepts, which is commonly used in calibration literature. It has been shown that this family of metrics can be decomposed into a calibration term and a refinement term. Achieving an optimal score requires both accurate predictions and appropriate confidence levels. For a given sample, the Brier Score is mathematically defined as the Mean Squared Error between the predicted probability distribution " + }, + { + "bbox": [ + 55, + 449, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\hat{p}" + }, + { + "bbox": [ + 55, + 449, + 296, + 628 + ], + "type": "text", + "content": " and one-hot encoded ground truth label " + }, + { + "bbox": [ + 55, + 449, + 296, + 628 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 449, + 296, + 628 + ], + "type": "text", + "content": ". We introduce a generalized Brier Score (gBS) form as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 121, + 636, + 295, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 636, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 121, + 636, + 295, + 669 + ], + "type": "interline_equation", + "content": "\\mathrm {g B S} = \\sum_ {i = 1} ^ {K} \\| \\hat {p} _ {i} (\\boldsymbol {x}) - y _ {i} \\| _ {\\beta} ^ {\\gamma} \\tag {3}", + "image_path": "c02f25647617d53de6531267ca968103bfd7d1f871db1e87c08037278b0bdd65.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " are the hyperparameters for the exponent and norm order, respectively. When " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma = 2" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\beta = 2" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ", this formulation reduces to the original Brier Score." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 72, + 432, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 432, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 432, + 84 + ], + "type": "text", + "content": "3.3. Calibrating Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 89, + 555, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 200 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 200 + ], + "type": "text", + "content": "Temperature Scaling A widely used post-hoc technique for improving classification calibration is temperature scaling. It adjusts the sharpness of the output probability distribution by scaling the logits in the SoftMax function using a temperature parameter, defined as " + }, + { + "bbox": [ + 313, + 89, + 555, + 200 + ], + "type": "inline_equation", + "content": "\\hat{p}_i = \\frac{\\exp(\\hat{g}_i / T)}{\\sum_{k=1}^{K} \\exp(\\hat{g}_k / T)}" + }, + { + "bbox": [ + 313, + 89, + 555, + 200 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 89, + 555, + 200 + ], + "type": "inline_equation", + "content": "\\hat{g}" + }, + { + "bbox": [ + 313, + 89, + 555, + 200 + ], + "type": "text", + "content": " represents the logits before applying the SoftMax function, and " + }, + { + "bbox": [ + 313, + 89, + 555, + 200 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 89, + 555, + 200 + ], + "type": "text", + "content": " is the temperature that controls the scaling. Calibration performance can be enhanced by tuning " + }, + { + "bbox": [ + 313, + 89, + 555, + 200 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 89, + 555, + 200 + ], + "type": "text", + "content": " on a held-out validation set." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 201, + 554, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 554, + 296 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 554, + 296 + ], + "type": "text", + "content": "Focal Loss and Dual Focal Loss Some other previous works have attempted to improve model calibration by modifying the loss function. Focal Loss [17] was initially introduced to address the foreground-background imbalance problem in object detection. It addresses the issue by incorporating a loss weighting factor based on sample complexity, reducing the weight of easy samples and allowing the model to focus on harder-to-classify instances." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 297, + 554, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 554, + 319 + ], + "type": "text", + "content": "Formally, given the predicted probability " + }, + { + "bbox": [ + 313, + 297, + 554, + 319 + ], + "type": "inline_equation", + "content": "\\hat{p} (\\pmb {x})" + }, + { + "bbox": [ + 313, + 297, + 554, + 319 + ], + "type": "text", + "content": " on sample " + }, + { + "bbox": [ + 313, + 297, + 554, + 319 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 297, + 554, + 319 + ], + "type": "text", + "content": ", the focal loss is defined as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 349, + 327, + 554, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 327, + 554, + 360 + ], + "spans": [ + { + "bbox": [ + 349, + 327, + 554, + 360 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {F L}} (\\boldsymbol {x}, y) = - \\sum_ {i = 1} ^ {K} y _ {i} (1 - \\hat {p} _ {i}) ^ {\\gamma} \\log \\hat {p} _ {i} (\\boldsymbol {x}), \\tag {4}", + "image_path": "4c699487197955f79c3d06c5c384c21b487f7395379b0f74ba8d41e0bf1fc81a.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 367, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 554, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 367, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 367, + 554, + 510 + ], + "type": "text", + "content": " is a pre-defined hyperparameter. Previous studies [18] have demonstrated that optimizing models using Focal Loss results in better calibration compared to Cross-Entropy. This improvement is partly due to the entropy-based regularization effect introduced by Focal Loss, while complexity-based weighting also likely plays a role. However, we noticed that Focal Loss was initially applied to binary classification, where the complexity term " + }, + { + "bbox": [ + 313, + 367, + 554, + 510 + ], + "type": "inline_equation", + "content": "(1 - \\hat{p}_i)" + }, + { + "bbox": [ + 313, + 367, + 554, + 510 + ], + "type": "text", + "content": " is related to the Brier Score, which can also be used for evaluating uncertainty. In multi-class classification, although complexity and uncertainty are not strictly equivalent, the complexity term still captures some class-level uncertainty." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 510, + 554, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 558 + ], + "type": "text", + "content": "Several variants of Focal Loss exist, including Dual Focal Loss [27], which incorporates the probability from the second most probable class into the scaling factor to address the under-confidence issue caused by Focal Loss:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 319, + 566, + 555, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 566, + 555, + 598 + ], + "spans": [ + { + "bbox": [ + 319, + 566, + 555, + 598 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D F L}} (\\boldsymbol {x}, y) = - \\sum_ {i = 1} ^ {K} y _ {i} (1 - \\hat {p} _ {i} (\\boldsymbol {x}) + \\hat {p} _ {j} (\\boldsymbol {x})) ^ {\\gamma} \\log \\hat {p} _ {i} (\\boldsymbol {x}), \\tag {5}", + "image_path": "2e4a18b2ad6604cf0e564649fc272adb1a3c6ec9fc11bfe7301d3566f63f2756.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 605, + 500, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 500, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 500, + 620 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 605, + 500, + 620 + ], + "type": "inline_equation", + "content": "\\hat{p}_j(x) = \\max_i\\{\\hat{p}_i(x)|\\hat{p}_i(x) < \\hat{p}_{gt}(x)\\}" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 628, + 370, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 628, + 370, + 640 + ], + "spans": [ + { + "bbox": [ + 314, + 628, + 370, + 640 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 647, + 528, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 528, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 528, + 661 + ], + "type": "text", + "content": "4.1. Weighting with Sample-wise Uncertainty" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "Focal Loss and Dual Focal Loss both support the concept that uncertainty-based weighting in conjunction with Cross Entropy can enhance model calibration. We propose a generalized loss function framework, termed Uncertainty CE" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15499" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "Loss, which incorporates a sample-wise uncertainty metric into Cross Entropy as a weighting factor:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 80, + 105, + 295, + 137 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 105, + 295, + 137 + ], + "spans": [ + { + "bbox": [ + 80, + 105, + 295, + 137 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {U n c e r t a i n t y}} (\\boldsymbol {x}, y) = - \\sum_ {i = 1} ^ {K} u \\left(\\hat {p} _ {i} (\\boldsymbol {x})\\right) \\cdot y _ {i} \\log \\hat {p} _ {i} (\\boldsymbol {x}), \\tag {6}", + "image_path": "71dec31bb5e8caa452b884c57d1ab3182811795c8be34fd92ea52cf3b707bd6a.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 146, + 296, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 146, + 296, + 206 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 296, + 206 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 146, + 296, + 206 + ], + "type": "inline_equation", + "content": "u(\\hat{p}_i)" + }, + { + "bbox": [ + 55, + 146, + 296, + 206 + ], + "type": "text", + "content": " is the adaptive term that evaluates the uncertainty of sample " + }, + { + "bbox": [ + 55, + 146, + 296, + 206 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 146, + 296, + 206 + ], + "type": "text", + "content": ". The motivation behind this design is to use the predicted calibration error for each sample to scale the loss function, thereby directing the model's optimization more effectively towards samples with higher uncertainty." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 206, + 286, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 206, + 286, + 218 + ], + "spans": [ + { + "bbox": [ + 67, + 206, + 286, + 218 + ], + "type": "text", + "content": "Therefore, the weights used in Focal Loss is given by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 227, + 295, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 227, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 127, + 227, + 295, + 240 + ], + "type": "interline_equation", + "content": "u _ {\\mathrm {F L}} (\\hat {p} (\\boldsymbol {x})) = (1 - \\hat {p} _ {c}) ^ {\\gamma}. \\tag {7}", + "image_path": "566e9cbae7ffdbcd5b035a497a8bda37d28021aea22b49a26738d4a080478db6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 249, + 275, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 249, + 275, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 275, + 261 + ], + "type": "text", + "content": "And the scaling term in Dual Focal Loss is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 100, + 270, + 295, + 284 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 270, + 295, + 284 + ], + "spans": [ + { + "bbox": [ + 100, + 270, + 295, + 284 + ], + "type": "interline_equation", + "content": "u _ {\\mathrm {D F L}} (\\hat {p} (\\boldsymbol {x})) = (1 - \\hat {p} _ {c} (x) + \\hat {p} _ {j} (x)) ^ {\\gamma}, \\tag {8}", + "image_path": "45ed220636e0fa6d43b8d9baf067a9ccbd453ad19065cd646e39bc4e077ddd7a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 293, + 296, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 296, + 388 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 293, + 296, + 388 + ], + "type": "inline_equation", + "content": "\\hat{p}_j(x)" + }, + { + "bbox": [ + 55, + 293, + 296, + 388 + ], + "type": "text", + "content": " is the second maximum confidence in prediction. The weight used in Focal Loss can be seen as derived from the Brier Score, focusing on the error of the ground true class. In contrast, the scaling term in Dual Focal Loss incorporates uncertainty information from the second most probable class. However, directly weighting loss function with certain uncertainty metrics, as those used in Focal Loss and Dual Focal Loss, presents several challenges." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 388, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 555 + ], + "type": "text", + "content": "Reviewing our objectives, we aim to weight the loss function value by each sample's uncertainty, resulting in greater optimization steps for those samples with higher uncertainty. Our focus actually is not on the loss value itself, but on the optimization. The purpose of uncertainty weighting is to focus the model's attention on samples with higher uncertainty. Thus, the key factor is the gradient: higher weights should lead to larger gradients, resulting in more substantial model updates for those samples. For Cross Entropy, its value is positively correlated with the gradient: a higher CE value yields a larger gradient, leading to more significant updates for optimization. When applying a scalar weight to CE, this relationship remains intact, as it does not disrupt the positive correlation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 555, + 296, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 296, + 716 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 296, + 716 + ], + "type": "text", + "content": "However, for differentiable weights, directly applying them to the loss can disrupt the positive correlation between the loss and the gradient, ultimately impairing effective model optimization. To further analyze this, let us consider Focal Loss. For Focal Loss " + }, + { + "bbox": [ + 55, + 555, + 296, + 716 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{FL}}" + }, + { + "bbox": [ + 55, + 555, + 296, + 716 + ], + "type": "text", + "content": " and Cross Entropy " + }, + { + "bbox": [ + 55, + 555, + 296, + 716 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 55, + 555, + 296, + 716 + ], + "type": "text", + "content": ", their gradients are given by: " + }, + { + "bbox": [ + 55, + 555, + 296, + 716 + ], + "type": "inline_equation", + "content": "\\frac{\\partial}{\\partial w}\\mathcal{L}_{\\mathrm{FL}} = g(\\hat{p}_i,\\gamma)\\frac{\\partial}{\\partial w}\\mathcal{L}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 55, + 555, + 296, + 716 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 555, + 296, + 716 + ], + "type": "inline_equation", + "content": "g(p,\\gamma) = (1 - p)^{\\gamma} - \\gamma p(1 - p)^{\\gamma -1}\\log (p)" + }, + { + "bbox": [ + 55, + 555, + 296, + 716 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 555, + 296, + 716 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 555, + 296, + 716 + ], + "type": "text", + "content": " is a predefined hyperparameter, and " + }, + { + "bbox": [ + 55, + 555, + 296, + 716 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}" + }, + { + "bbox": [ + 55, + 555, + 296, + 716 + ], + "type": "text", + "content": " represents the parameters of the final linear layer. From a gradient perspective, " + }, + { + "bbox": [ + 55, + 555, + 296, + 716 + ], + "type": "inline_equation", + "content": "g(p,\\gamma)" + }, + { + "bbox": [ + 55, + 555, + 296, + 716 + ], + "type": "text", + "content": " acts as a weight on the CE gradient, which is illustrated in Figure 1. There exists a point " + }, + { + "bbox": [ + 55, + 555, + 296, + 716 + ], + "type": "inline_equation", + "content": "p_0\\in [0,1]" + }, + { + "bbox": [ + 55, + 555, + 296, + 716 + ], + "type": "text", + "content": " such that within the range " + }, + { + "bbox": [ + 55, + 555, + 296, + 716 + ], + "type": "inline_equation", + "content": "[0,p_0],\\frac{\\partial}{\\partial p} g(p,\\gamma) > 0" + }, + { + "bbox": [ + 55, + 555, + 296, + 716 + ], + "type": "text", + "content": " and within the range " + }, + { + "bbox": [ + 55, + 555, + 296, + 716 + ], + "type": "inline_equation", + "content": "[p_0,1],\\frac{\\partial}{\\partial p} g(p,\\gamma) < 0" + }, + { + "bbox": [ + 55, + 555, + 296, + 716 + ], + "type": "text", + "content": ". In binary classification, " + }, + { + "bbox": [ + 55, + 555, + 296, + 716 + ], + "type": "inline_equation", + "content": "1 - \\hat{p}_i" + }, + { + "bbox": [ + 55, + 555, + 296, + 716 + ], + "type": "text", + "content": " is related to" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 81, + 553, + 211 + ], + "blocks": [ + { + "bbox": [ + 345, + 72, + 539, + 82 + ], + "lines": [ + { + "bbox": [ + 345, + 72, + 539, + 82 + ], + "spans": [ + { + "bbox": [ + 345, + 72, + 539, + 82 + ], + "type": "text", + "content": "Plot of " + }, + { + "bbox": [ + 345, + 72, + 539, + 82 + ], + "type": "inline_equation", + "content": "g(p)" + }, + { + "bbox": [ + 345, + 72, + 539, + 82 + ], + "type": "text", + "content": " for Different " + }, + { + "bbox": [ + 345, + 72, + 539, + 82 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 345, + 72, + 539, + 82 + ], + "type": "text", + "content": " Values in Focal Loss" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 81, + 553, + 211 + ], + "lines": [ + { + "bbox": [ + 317, + 81, + 553, + 211 + ], + "spans": [ + { + "bbox": [ + 317, + 81, + 553, + 211 + ], + "type": "image", + "image_path": "3b60aa11e502d750f2ad60558338e62d216fc72b2bec4d7340aa3b64027b7fd1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 221, + 541, + 233 + ], + "lines": [ + { + "bbox": [ + 326, + 221, + 541, + 233 + ], + "spans": [ + { + "bbox": [ + 326, + 221, + 541, + 233 + ], + "type": "text", + "content": "Figure 1. " + }, + { + "bbox": [ + 326, + 221, + 541, + 233 + ], + "type": "inline_equation", + "content": "g(p, \\gamma)" + }, + { + "bbox": [ + 326, + 221, + 541, + 233 + ], + "type": "text", + "content": " of Focal Loss vs predicted confidence " + }, + { + "bbox": [ + 326, + 221, + 541, + 233 + ], + "type": "inline_equation", + "content": "\\hat{p}_c" + }, + { + "bbox": [ + 326, + 221, + 541, + 233 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 251, + 555, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 408 + ], + "type": "text", + "content": "the Brier Score that reveals calibrated error, indicating that the predicted calibrated error is not always positively correlated with the weight on gradient. As uncertainty decreases, the weight on gradient initially increases, before eventually aligning with the level of uncertainty. This causes the model to focus more on samples with moderate uncertainty rather than those with the highest uncertainty, which contradicts our original goal of emphasizing the most uncertain samples. Further, changes in uncertainty are not immediately reflected in the sample weights; instead, multiple training iterations are required for these changes to align properly. Thus, directly applying some kind of uncertainty metrics into Eq. 6 would pose issues to the uncertainty weighting." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 414, + 518, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 518, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 518, + 426 + ], + "type": "text", + "content": "4.2. Sample-wise Uncertainty on Gradients" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 431, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 554, + 491 + ], + "type": "text", + "content": "The discussion in Sec. 4.1 motivates us to ensure the weight on gradients for a sample to align with its uncertainty. This could be intuitively solved by directly applying the model uncertainty on samples as a weight of the gradient. We define the modified gradient as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 343, + 497, + 553, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 497, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 343, + 497, + 553, + 521 + ], + "type": "interline_equation", + "content": "\\frac {\\partial}{\\partial \\theta} \\mathcal {L} _ {\\text {U n c e r t a i n t y - G R A}} (\\boldsymbol {x}, y) = u (\\hat {p} (\\boldsymbol {x})) \\frac {\\partial}{\\partial \\theta} \\mathcal {L} _ {\\mathrm {C E}} (\\boldsymbol {x}, y). \\tag {9}", + "image_path": "e858f85ac2779cb39464abf997e9b2d7b75d6c8fe98980d655b75061be18dcf3.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 526, + 555, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 555, + 632 + ], + "type": "text", + "content": "Taking a SGD optimizer as an example, the model update can be expressed as: " + }, + { + "bbox": [ + 313, + 526, + 555, + 632 + ], + "type": "inline_equation", + "content": "\\theta_{t + 1} = \\theta +\\alpha \\cdot \\frac{\\partial}{\\partial\\theta}\\mathcal{L}_{\\mathrm{Uncertainty - GRA}}(\\boldsymbol {x},y)" + }, + { + "bbox": [ + 313, + 526, + 555, + 632 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 526, + 555, + 632 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 526, + 555, + 632 + ], + "type": "text", + "content": " is the model parameters and " + }, + { + "bbox": [ + 313, + 526, + 555, + 632 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 526, + 555, + 632 + ], + "type": "text", + "content": " is the learning rate. When the predicted uncertainty on sample " + }, + { + "bbox": [ + 313, + 526, + 555, + 632 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 526, + 555, + 632 + ], + "type": "text", + "content": " is higher, the optimization step results in a larger update, encouraging the model to focus more on uncertain samples. A general form of the loss function for our uncertainty-weighted gradient framework, termed Uncertainty-GRA CE Loss, is defined as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 319, + 639, + 553, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 639, + 553, + 670 + ], + "spans": [ + { + "bbox": [ + 319, + 639, + 553, + 670 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {U n c e r t a i n t y - G R A}} (\\boldsymbol {x}, y) = - \\int \\sum_ {i = 1} ^ {K} u (\\hat {p} (\\boldsymbol {x})) \\cdot \\frac {y _ {i}}{\\hat {p} _ {i} (\\boldsymbol {x})} \\mathrm {d} \\hat {p} (\\boldsymbol {x}), \\tag {10}", + "image_path": "04c8439a377174b64f3e3d6cef7b16b41e472d61e70400c4394f26b636f21582.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "In practice, to compute the " + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Uncertainty - GRA}}" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": ", we can detach the gradient of " + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "u(\\hat{p} (\\pmb {x}))" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": " and multiply it with the Cross Entropy, instead of calculating the gradient integration." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15500" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 197, + 194 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 197, + 194 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 197, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 197, + 194 + ], + "type": "image", + "image_path": "6924a014be14d1c385e5f18451224290bdbdf369c7df357b0a3ff91bfd7eb54d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 198, + 179, + 207 + ], + "lines": [ + { + "bbox": [ + 84, + 198, + 179, + 207 + ], + "spans": [ + { + "bbox": [ + 84, + 198, + 179, + 207 + ], + "type": "text", + "content": "(a) An illustration of the " + }, + { + "bbox": [ + 84, + 198, + 179, + 207 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{FL}}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 227, + 68, + 370, + 194 + ], + "blocks": [ + { + "bbox": [ + 227, + 68, + 370, + 194 + ], + "lines": [ + { + "bbox": [ + 227, + 68, + 370, + 194 + ], + "spans": [ + { + "bbox": [ + 227, + 68, + 370, + 194 + ], + "type": "image", + "image_path": "452fd0f06819ecff3714e8a32ff5da4ade46079b2ab28b66986f13ffe461ca0b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 198, + 353, + 207 + ], + "lines": [ + { + "bbox": [ + 255, + 198, + 353, + 207 + ], + "spans": [ + { + "bbox": [ + 255, + 198, + 353, + 207 + ], + "type": "text", + "content": "(b) An illustration of the " + }, + { + "bbox": [ + 255, + 198, + 353, + 207 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{DFL}}" + }, + { + "bbox": [ + 255, + 198, + 353, + 207 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 217, + 555, + 240 + ], + "lines": [ + { + "bbox": [ + 55, + 217, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 555, + 240 + ], + "type": "text", + "content": "Figure 2. An illustration of value of gradient weight function on a 4 class classification. It is obvious that " + }, + { + "bbox": [ + 55, + 217, + 555, + 240 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{FL}}" + }, + { + "bbox": [ + 55, + 217, + 555, + 240 + ], + "type": "text", + "content": " varies only along the " + }, + { + "bbox": [ + 55, + 217, + 555, + 240 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 217, + 555, + 240 + ], + "type": "text", + "content": " axis and " + }, + { + "bbox": [ + 55, + 217, + 555, + 240 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{DFL}}" + }, + { + "bbox": [ + 55, + 217, + 555, + 240 + ], + "type": "text", + "content": " changes along the " + }, + { + "bbox": [ + 55, + 217, + 555, + 240 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 217, + 555, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 217, + 555, + 240 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 55, + 217, + 555, + 240 + ], + "type": "text", + "content": " axes. " + }, + { + "bbox": [ + 55, + 217, + 555, + 240 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{BS}}" + }, + { + "bbox": [ + 55, + 217, + 555, + 240 + ], + "type": "text", + "content": " responds to changes across all axes, providing a more complete uncertainty evaluation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 402, + 68, + 542, + 194 + ], + "blocks": [ + { + "bbox": [ + 402, + 68, + 542, + 194 + ], + "lines": [ + { + "bbox": [ + 402, + 68, + 542, + 194 + ], + "spans": [ + { + "bbox": [ + 402, + 68, + 542, + 194 + ], + "type": "image", + "image_path": "50d805ec096e62c5b10963250182e4b37bf2a9c9de00e3a8313ea604220013c0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 198, + 525, + 207 + ], + "lines": [ + { + "bbox": [ + 430, + 198, + 525, + 207 + ], + "spans": [ + { + "bbox": [ + 430, + 198, + 525, + 207 + ], + "type": "text", + "content": "(c) An illustration of the " + }, + { + "bbox": [ + 430, + 198, + 525, + 207 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{BS}}" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 259, + 234, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 259, + 234, + 272 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 234, + 272 + ], + "type": "text", + "content": "4.3. Sample-wise Uncertainty Metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 277, + 296, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 296, + 337 + ], + "type": "text", + "content": "With Eq. 10, calibration can be achieved by employing the sample's ground-truth uncertainty as the scaling term. However, accessing the ground-truth uncertainty for real-world image samples is impractical. Therefore, alternative metrics are required to estimate the sample-wise uncertainty " + }, + { + "bbox": [ + 55, + 277, + 296, + 337 + ], + "type": "inline_equation", + "content": "u(x)" + }, + { + "bbox": [ + 55, + 277, + 296, + 337 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 338, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 361 + ], + "type": "text", + "content": "One approach to evaluate sample-wise uncertainty is based on computing the generalized Brier Score (gBS):" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 107, + 369, + 295, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 369, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 107, + 369, + 295, + 401 + ], + "type": "interline_equation", + "content": "u _ {\\mathrm {g B S}} (\\hat {p} (\\boldsymbol {x})) = \\sum_ {i = 1} ^ {K} \\| \\hat {p} _ {i} (\\boldsymbol {x}) - y _ {i} \\| _ {\\beta} ^ {\\gamma}. \\tag {11}", + "image_path": "cf40f2f980a622412c37e5b3983e0b99dfc9ab6de7d24277c9bfcac601a9364d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 407, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 468 + ], + "type": "text", + "content": "We provide a brief discussion on the effectiveness of the generalized Brier Score (gBS) as a measure of calibration error. When using the Brier Score (BS) to evaluate calibration error, with " + }, + { + "bbox": [ + 55, + 407, + 296, + 468 + ], + "type": "inline_equation", + "content": "\\beta = 2" + }, + { + "bbox": [ + 55, + 407, + 296, + 468 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 407, + 296, + 468 + ], + "type": "inline_equation", + "content": "\\gamma = 2" + }, + { + "bbox": [ + 55, + 407, + 296, + 468 + ], + "type": "text", + "content": ", the difference between the expected and predicted calibration error is given by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 475, + 310, + 524 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 310, + 524 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 310, + 524 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} c (\\boldsymbol {x}) - u _ {\\mathrm {B S}} (\\hat {p} (\\boldsymbol {x})) = \\mathbb {E} _ {y \\sim \\eta (x)} [ \\| \\hat {p} (\\boldsymbol {x}) - \\boldsymbol {\\eta} (\\boldsymbol {x}) \\| _ {2} ^ {2} - \\| \\hat {p} (\\boldsymbol {x}) - y \\| _ {2} ^ {2} ] \\\\ = \\sum_ {k = 1} ^ {K} \\boldsymbol {\\eta} _ {k} (\\boldsymbol {x}) \\left(\\boldsymbol {\\eta} _ {k} (\\boldsymbol {x}) - 1\\right). \\tag {12} \\\\ \\end{array}", + "image_path": "1d6b1e081612f4e229ee43c99c0379fe6ae22873255ec4b8788c20c175469dde.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "text", + "content": "Thus, the term " + }, + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\sum_{k=1}^{K} \\eta_k(x) (\\eta_k(x) - 1)" + }, + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "text", + "content": " depends solely on " + }, + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\eta(x)" + }, + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "text", + "content": " which is fixed for a given sample " + }, + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "text", + "content": ". Consequently, the error remains constant for a specific sample " + }, + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "text", + "content": ". We provide a detailed computation of Eq. 12 in Appendix." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 582, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 677 + ], + "type": "text", + "content": "Several variants of " + }, + { + "bbox": [ + 55, + 582, + 295, + 677 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{gBS}}(\\hat{p})" + }, + { + "bbox": [ + 55, + 582, + 295, + 677 + ], + "type": "text", + "content": " have demonstrated their effectiveness in previous studies. When " + }, + { + "bbox": [ + 55, + 582, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\beta = 1" + }, + { + "bbox": [ + 55, + 582, + 295, + 677 + ], + "type": "text", + "content": " and only the actual class is considered, the generalized Brier Score can be interpreted as the scaling term in Focal Loss in Eq. 7. When " + }, + { + "bbox": [ + 55, + 582, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\beta = 1" + }, + { + "bbox": [ + 55, + 582, + 295, + 677 + ], + "type": "text", + "content": " and both the actual class and the maximum predicted class (excluding the actual class) are considered, the generalized Brier Score can be interpreted as the scaling term in Dual Focal Loss in Eq. 8." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "However, these two metrics only consider several classes in the predicted probability. In the case of Focal Loss, which was originally designed for binary classification," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 260, + 553, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 553, + 308 + ], + "type": "text", + "content": "considering one-dimensional uncertainty is sufficient due to the constraint that the sum of probabilities must equal 1. When extended to multi-class classification, these metrics are not accurate enough to evaluate uncertainty." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 308, + 554, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 554, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 554, + 427 + ], + "type": "text", + "content": "We visualize the gradient weights " + }, + { + "bbox": [ + 313, + 308, + 554, + 427 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{FL}}" + }, + { + "bbox": [ + 313, + 308, + 554, + 427 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 308, + 554, + 427 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{DFL}}" + }, + { + "bbox": [ + 313, + 308, + 554, + 427 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 308, + 554, + 427 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{BS}}" + }, + { + "bbox": [ + 313, + 308, + 554, + 427 + ], + "type": "text", + "content": " in Figure 2 for a 4-class scenario. The three axes represent three dimensions of the predicted probability, while the fourth is implied by the probability sum constraint. From the figure, it is evident that " + }, + { + "bbox": [ + 313, + 308, + 554, + 427 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{FL}}" + }, + { + "bbox": [ + 313, + 308, + 554, + 427 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 308, + 554, + 427 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{DFL}}" + }, + { + "bbox": [ + 313, + 308, + 554, + 427 + ], + "type": "text", + "content": " are sensitive only to changes in one or two dimensions. However, different points in the coordinate system have different uncertainties. They fail to adequately capture uncertainty change across all dimensions. But the " + }, + { + "bbox": [ + 313, + 308, + 554, + 427 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{BS}}" + }, + { + "bbox": [ + 313, + 308, + 554, + 427 + ], + "type": "text", + "content": " can accurately respond to changes along any coordinate axis." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 427, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 555, + 594 + ], + "type": "text", + "content": "To further evaluate whether these metrics accurately measure the ground truth uncertainty of samples, we conducted experiments on a toy dataset. The toy dataset consists of 5 two-dimensional Gaussian distributions, representing 5 groups of data: " + }, + { + "bbox": [ + 313, + 427, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mu_i,\\Sigma),i\\in \\{0,\\dots ,5\\}" + }, + { + "bbox": [ + 313, + 427, + 555, + 594 + ], + "type": "text", + "content": ". The mean vectors " + }, + { + "bbox": [ + 313, + 427, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\mu_{i}" + }, + { + "bbox": [ + 313, + 427, + 555, + 594 + ], + "type": "text", + "content": " were randomly sampled from the range " + }, + { + "bbox": [ + 313, + 427, + 555, + 594 + ], + "type": "inline_equation", + "content": "[-10,10]" + }, + { + "bbox": [ + 313, + 427, + 555, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 427, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\Sigma = I" + }, + { + "bbox": [ + 313, + 427, + 555, + 594 + ], + "type": "text", + "content": " was used as the shared covariance matrix for all groups. We generated 10,000 data points from each group to form the training dataset, which was used to train a two layer CNN model for 5 epochs. An additional 1,000 samples from each distribution were used to create the test dataset. Therefore, the ground truth probability " + }, + { + "bbox": [ + 313, + 427, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\eta (\\pmb {x})" + }, + { + "bbox": [ + 313, + 427, + 555, + 594 + ], + "type": "text", + "content": " can be computed using the probability density function (PDF) of each distribution:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 392, + 604, + 553, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 604, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 392, + 604, + 553, + 632 + ], + "type": "interline_equation", + "content": "\\eta (\\boldsymbol {x}) = \\frac {p ^ {n} (\\boldsymbol {x})}{\\sum_ {i} ^ {N} p ^ {i} (\\boldsymbol {x})}, \\tag {13}", + "image_path": "8f9696cd1a9163a87a83528019ae346b09f5f0e7e5e2120a9af8c4151bf2a946.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "inline_equation", + "content": "N = 5" + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": " in this case, " + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "inline_equation", + "content": "p^n (\\pmb {x})" + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "inline_equation", + "content": "p^i (\\pmb {x})" + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": " are the PDF of corresponds class for " + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": " and remaining classes, respectively. The ground truth uncertainty of each sample is calculated by Eq.1, representing the model's ground truth uncertainty for each sample, along with the sample-wise uncertainty metrics " + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{FL}},u_{\\mathrm{DFL}}" + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{gBS}}" + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": ". To validate the accuracy" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15501" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 191 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 191 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 191 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 191 + ], + "type": "table", + "html": "
DatasetModelCEBLMMCEFLSDDFLBSCEBSCE-GRA
Pre TPost TPre TPost TPre TPost TPre TPost TPre TPost TPre TPost TPre TPost T
CIFAR10ResNet504.361.324.291.504.481.411.261.151.001.000.880.880.740.74
ResNet1104.71.564.481.664.801.291.811.171.011.010.990.990.870.87
WideResNet3.350.942.861.103.651.281.841.043.321.161.70.951.461.12
DenseNet4.641.463.961.464.811.671.371.170.870.771.011.010.870.87
CIFAR100ResNet5018.053.057.874.2715.873.325.532.572.542.561.901.901.591.59
ResNet11018.844.6316.774.3018.653.936.883.713.473.472.752.752.532.53
WideResNet14.813.277.744.4614.582.992.702.715.452.522.632.422.462.46
DenseNet19.13.438.132.9917.562.873.381.304.681.831.631.621.621.61
TinyImageNetResNet5014.945.167.811.4714.582.992.182.186.712.284.01.764.571.47
", + "image_path": "aca6666bed3413b3be9605a798fff6ec0073552017e27adcab00ea1518a3294a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 200, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 200, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 200, + 555, + 232 + ], + "type": "text", + "content": "Table 1. Comparison of Calibration Methods Using ECE Across Various Datasets and Models. ECE values are reported using 15 bins, with the best-performing method for each dataset-model combination highlighted in bold. Results are averaged over three runs with different random seeds." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 61, + 244, + 134, + 303 + ], + "blocks": [ + { + "bbox": [ + 61, + 244, + 134, + 303 + ], + "lines": [ + { + "bbox": [ + 61, + 244, + 134, + 303 + ], + "spans": [ + { + "bbox": [ + 61, + 244, + 134, + 303 + ], + "type": "image", + "image_path": "aa2b34465d503e26385b986d94702aec25884608307ea87fb7ffbe260d14c3fc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 314, + 555, + 336 + ], + "lines": [ + { + "bbox": [ + 55, + 314, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 555, + 336 + ], + "type": "text", + "content": "Figure 3. Comparison of different ECE metrics. The first three plots show the uncertainty for CIFAR-10 using ResNet-50, while the remaining plots represent ResNet-110 on CIFAR-10." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 143, + 244, + 218, + 303 + ], + "blocks": [ + { + "bbox": [ + 143, + 244, + 218, + 303 + ], + "lines": [ + { + "bbox": [ + 143, + 244, + 218, + 303 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 218, + 303 + ], + "type": "image", + "image_path": "b1a0cafc6b558ccd1f52e7538a2fce08d5c2ce37372c99c37642608df313b262.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 225, + 245, + 301, + 303 + ], + "blocks": [ + { + "bbox": [ + 225, + 245, + 301, + 303 + ], + "lines": [ + { + "bbox": [ + 225, + 245, + 301, + 303 + ], + "spans": [ + { + "bbox": [ + 225, + 245, + 301, + 303 + ], + "type": "image", + "image_path": "e8ea14360f6549b8709c684c222ab15b20a8e06334bb4750e505779ca5499dbf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 245, + 384, + 303 + ], + "blocks": [ + { + "bbox": [ + 309, + 245, + 384, + 303 + ], + "lines": [ + { + "bbox": [ + 309, + 245, + 384, + 303 + ], + "spans": [ + { + "bbox": [ + 309, + 245, + 384, + 303 + ], + "type": "image", + "image_path": "f7871714057ae1a14355edb664b5c02e74c3151e359785b44af2a093a17fd064.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 394, + 245, + 468, + 303 + ], + "blocks": [ + { + "bbox": [ + 394, + 245, + 468, + 303 + ], + "lines": [ + { + "bbox": [ + 394, + 245, + 468, + 303 + ], + "spans": [ + { + "bbox": [ + 394, + 245, + 468, + 303 + ], + "type": "image", + "image_path": "87d1358feb434fb5f3fffaae88fbbb5e79a53bfb5edcf9f25f820823a2c0073a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 476, + 246, + 550, + 302 + ], + "blocks": [ + { + "bbox": [ + 476, + 246, + 550, + 302 + ], + "lines": [ + { + "bbox": [ + 476, + 246, + 550, + 302 + ], + "spans": [ + { + "bbox": [ + 476, + 246, + 550, + 302 + ], + "type": "image", + "image_path": "017e996c9608e71808b491d3dbfad8f3c6d79b6a46c14f907f0bb51928b41168.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 357, + 295, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 357, + 295, + 489 + ], + "spans": [ + { + "bbox": [ + 54, + 357, + 295, + 489 + ], + "type": "text", + "content": "of these metrics, we used Pearson correlations between the calibrated error and the uncertainty metrics to verify their positive correlation among 5 runs. We use grid-search to find the optimal hyper-parameter for each methods. Among the methods, gBS achieves the highest Pearson correlation coefficient of 0.664, indicating the strongest linear relationship between the predicted values and the true targets. DFL follows with a correlation of 0.638, while FL has the lowest correlation at 0.550. This suggests that gBS provides the most accurate predicted uncertainty alignment with ground-truth uncertainty, followed by DFL and FL." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 490, + 295, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 295, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 295, + 621 + ], + "type": "text", + "content": "Based on the results, we conclude that although ground-truth uncertainty is inaccessible for real-world datasets, these alternative metrics provide a reliable means of estimating uncertainty, making them suitable for use during training. Besides, the metric " + }, + { + "bbox": [ + 55, + 490, + 295, + 621 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{gBS}}" + }, + { + "bbox": [ + 55, + 490, + 295, + 621 + ], + "type": "text", + "content": " has the best performance compared to the uncertainty metric used in Focal Loss and Dual Focal Loss. We incorporate " + }, + { + "bbox": [ + 55, + 490, + 295, + 621 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{gBS}}" + }, + { + "bbox": [ + 55, + 490, + 295, + 621 + ], + "type": "text", + "content": " into the gradient-weighted framework discussed in Section 4.2 and introduce a new loss function called BSCE-GRA. This loss function uses the generalized Brier Score as an adaptive uncertainty metric to weight the Cross Entropy gradients, defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 628, + 295, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 628, + 295, + 659 + ], + "spans": [ + { + "bbox": [ + 61, + 628, + 295, + 659 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {B S C E - G R A}} (\\boldsymbol {x}, y) = - \\int \\sum_ {i = 1} ^ {K} u _ {\\mathrm {g B S}} (\\hat {p} (\\boldsymbol {x})) \\cdot \\frac {y _ {i}}{\\hat {p} _ {i} (\\boldsymbol {x})} \\mathrm {d} \\hat {p} (\\boldsymbol {x}). \\tag {14}", + "image_path": "e8fc2aecb8f7b90ff688ba38fa7b377515a872cb0d67c5e43646c268cc552085.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "To further validate the effectiveness of proposed method, we provide a comprehensive theoretical evidence that optimizing with BSCE-GRA, the K-class predicted probability " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " would equal the actual class-posterior probability" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 357, + 554, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 554, + 392 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 357, + 554, + 392 + ], + "type": "text", + "content": ", thereby preventing over/under-confidence when convergence. Due to page limitation, the proof is provided in Appendix." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 393, + 554, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 554, + 465 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 554, + 465 + ], + "type": "text", + "content": "Besides, the Uncertainty-GRA Loss makes it possible to use non-differentiable uncertainty metrics for model calibration. However, it requires the weights to capture sample-wise uncertainty, whereas some existing measurements like ECE can only compute the uncertainty for a group of samples, making them unsuitable for the proposed framework." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 489, + 390, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 390, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 390, + 502 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 510, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 581 + ], + "type": "text", + "content": "We evaluate our methods on multiple deep neural networks (DNNs), including ResNet50, ResNet110 [6], WideResNet [35], and DenseNet [8]. Our experiments are conducted on CIFAR-10, CIFAR-100 [12], and Tiny-ImageNet [4] to assess calibration performance. Further details about the datasets can be found in the appendix." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "text", + "content": "Baselines. We compare our methods, BSCE, BSCE-GRA, and ECE-CE, with multiple existing approaches, including training with Cross Entropy (CE), Brier Loss (BL)[2], MMCE Loss[15], Focal Loss with Adaptive Exponent (FLSD)[18], and Dual Focal Loss[27]. For Focal Loss, we employ the FLSD-53 strategy [18] to adaptively adjust the gamma value sample-wise, setting " + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\gamma_{\\mathrm{FL}} = 5" + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\hat{p}_c \\in [0,0.2)" + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\gamma_{\\mathrm{FL}} = 3" + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\hat{p}_c \\in [0.2,1)" + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "text", + "content": ". For Dual Focal Loss, the gamma value is set to 5, as reported in the original work." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Training Setup. Our training setup follows prior works [18, 27]. We train CIFAR-10 and CIFAR-100 for 350" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15502" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 173, + 139 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 173, + 139 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 173, + 139 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 173, + 139 + ], + "type": "image", + "image_path": "e353d8870b219815d2cee7f87df3da1e70f68397c8e672e44560104b5602fb24.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 141, + 173, + 160 + ], + "lines": [ + { + "bbox": [ + 56, + 141, + 173, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 173, + 160 + ], + "type": "text", + "content": "(a) Gradient norm distribution at epoch 50" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 184, + 70, + 299, + 139 + ], + "blocks": [ + { + "bbox": [ + 184, + 70, + 299, + 139 + ], + "lines": [ + { + "bbox": [ + 184, + 70, + 299, + 139 + ], + "spans": [ + { + "bbox": [ + 184, + 70, + 299, + 139 + ], + "type": "image", + "image_path": "348cade48d66fc01a8b12e3d67982302da7cf23b4169562c1524d92d12f3cb78.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 141, + 299, + 160 + ], + "lines": [ + { + "bbox": [ + 184, + 141, + 299, + 160 + ], + "spans": [ + { + "bbox": [ + 184, + 141, + 299, + 160 + ], + "type": "text", + "content": "(b) Gradient norm distribution at epoch 150" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 170, + 555, + 194 + ], + "lines": [ + { + "bbox": [ + 55, + 170, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 170, + 555, + 194 + ], + "type": "text", + "content": "Figure 4. Evolution of gradient norm distributions across different training epochs for various loss functions. The scatter plots show the relationship between gradient norm and Brier Score for different loss functions (Focal Loss, Dual Focal Loss, BSCE-GRA)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 310, + 70, + 426, + 139 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 426, + 139 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 426, + 139 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 426, + 139 + ], + "type": "image", + "image_path": "a07661b82435e63d1529d9abf5fcea3f2d8cc682cc060d3bdfa3624e1207687b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 141, + 427, + 160 + ], + "lines": [ + { + "bbox": [ + 310, + 141, + 427, + 160 + ], + "spans": [ + { + "bbox": [ + 310, + 141, + 427, + 160 + ], + "type": "text", + "content": "(c) Gradient norm distribution at epoch 250" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 438, + 70, + 552, + 139 + ], + "blocks": [ + { + "bbox": [ + 438, + 70, + 552, + 139 + ], + "lines": [ + { + "bbox": [ + 438, + 70, + 552, + 139 + ], + "spans": [ + { + "bbox": [ + 438, + 70, + 552, + 139 + ], + "type": "image", + "image_path": "22beb2d17ba9c0f3ac1c97e94a7f6ca3bb131c95d596ca070edfc7b23ee2476b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 437, + 141, + 553, + 160 + ], + "lines": [ + { + "bbox": [ + 437, + 141, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 437, + 141, + 553, + 160 + ], + "type": "text", + "content": "(d) Gradient norm distribution at epoch 350" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 56, + 202, + 555, + 269 + ], + "blocks": [ + { + "bbox": [ + 56, + 202, + 555, + 269 + ], + "lines": [ + { + "bbox": [ + 56, + 202, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 56, + 202, + 555, + 269 + ], + "type": "table", + "html": "
MetricsFLSDFLSD-GRADFLDFL-GRABSCEBSCE-GRA
pre Tpost Tpre Tpost Tpre Tpost Tpre Tpost Tpre Tpost Tpre Tpost T
Acc95.04%95.04%94.72%94.72%94.63%94.63%94.76%94.76%95.03%95.03%94.69%94.69%
ECE1.261.150.880.881.001.000.930.840.880.880.740.74
Ada ECE1.561.451.191.191.221.220.770.810.960.960.710.71
", + "image_path": "d22b5198ef5dbd0bc0976c7940e6b40b01d78dca536cc5d41dccc371aae4a76a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 276, + 555, + 300 + ], + "lines": [ + { + "bbox": [ + 54, + 276, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 54, + 276, + 555, + 300 + ], + "type": "text", + "content": "Table 2. Comparison of weighting different uncertainty metrics on gradient or loss function, including " + }, + { + "bbox": [ + 54, + 276, + 555, + 300 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{FL}}" + }, + { + "bbox": [ + 54, + 276, + 555, + 300 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 276, + 555, + 300 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{FL}}" + }, + { + "bbox": [ + 54, + 276, + 555, + 300 + ], + "type": "text", + "content": ". The results validate the effectiveness of the gradient-weighting strategy among different uncertainty metrics." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 319, + 297, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 319, + 297, + 559 + ], + "spans": [ + { + "bbox": [ + 54, + 319, + 297, + 559 + ], + "type": "text", + "content": "epochs, using 5,000 images from the training set for validation. The learning rate is initially set to 0.1 for the first 150 epochs, then reduced to 0.01 for the next 100 epochs, and further reduced to 0.001 for the remaining epochs. For Tiny-ImageNet, we train for 100 epochs, with the learning rate set to 0.1 for the first 40 epochs, 0.01 for the next 20 epochs, and 0.001 for the remaining epochs. All experiments are conducted using SGD with a weight decay of " + }, + { + "bbox": [ + 54, + 319, + 297, + 559 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 54, + 319, + 297, + 559 + ], + "type": "text", + "content": " and a momentum of 0.9. The training and testing batch sizes for all datasets are set to 128. We re-run all baseline methods using three different random seeds (1, 42, and 71), and report the average results. All experiments are performed on a single Nvidia 4090 GPU. For temperature scaling, the temperature parameter " + }, + { + "bbox": [ + 54, + 319, + 297, + 559 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 319, + 297, + 559 + ], + "type": "text", + "content": " is optimized through a grid search with " + }, + { + "bbox": [ + 54, + 319, + 297, + 559 + ], + "type": "inline_equation", + "content": "T \\in [0, 0.1, 0.2, \\dots, 10]" + }, + { + "bbox": [ + 54, + 319, + 297, + 559 + ], + "type": "text", + "content": " on the validation set, selecting the value that yields the best post-temperature-scaling Expected Calibration Error (ECE). The same optimized temperature parameter is applied to other metrics, such as AdaECE. Further details on the datasets and additional experimental setup are provided in the appendix." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 574, + 195, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 195, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 195, + 586 + ], + "type": "text", + "content": "5.1. Calibration Performance" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 54, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 296, + 714 + ], + "type": "text", + "content": "We report the average ECE before and after temperature scaling among three random seeds, along with the corresponding optimal temperatures, in Table 1. BSCE-GRA achieves state-of-the-art ECE performance in most cases, particularly in the pre-temperature-scaling results. Notably, the fact that most of the optimal temperatures for BSCE-GRA are found to be 1 indicates that BSCE-GRA trains an inherently calibrated model, capable of achieving strong calibration performance without the need for additional temperature scaling. This is a crucial advantage" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 319, + 555, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 427 + ], + "type": "text", + "content": "for developing accurate and reliable models that are efficient and require minimal post-processing. The results on CIFAR-10 generally show better calibration performance compared to datasets with more labels (e.g., CIFAR-100 and Tiny-ImageNet) across multiple models. Regarding network architecture, ResNet-50 demonstrates the best calibration performance among the four DNNs tested (ResNet-50, ResNet-110, Wide-ResNet-26-10, and DenseNet-121) on both CIFAR-10 and CIFAR-100 datasets." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 433, + 556, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 556, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 556, + 601 + ], + "type": "text", + "content": "Different Metrics. The methods are further evaluated using several widely-accepted metrics to assess calibration performance across models, including Adaptive ECE and Classwise-ECE. Adaptive ECE measures the expected calibration error while accounting for the distribution of the data, whereas Classwise-ECE is a variant that evaluates calibration error for each class individually. Figure 3 presents the results of multiple methods using ResNet-50 and ResNet-110 on the CIFAR-10 dataset. The figure demonstrates that BSCE-GRA is the only method that achieves both inherently calibrated models and state-of-the-art performance across various metrics. Additional results are presented in the appendix, providing further evidence of the effectiveness of our method in model calibration." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Calibration over Training. Figure 5a presents the ECE on the test set for models trained with Focal Loss and Cross-Entropy loss over the entire training period on CIFAR-10 using ResNet-50. To improve visualization, the ECE values are smoothed using an exponential moving average. The figure suggests that, after the initial warm-up epochs, where predicted probabilities are unstable, the ECE of models trained with BSCE-GRA and Focal Loss consistently remains lower compared to models trained with" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15503" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 208, + 159 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 208, + 159 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 208, + 159 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 208, + 159 + ], + "type": "image", + "image_path": "f226010bb8659982c60fc861a275f563ccbe17bdaea90cba375cd84ba8778ebf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 162, + 208, + 180 + ], + "lines": [ + { + "bbox": [ + 56, + 162, + 208, + 180 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 208, + 180 + ], + "type": "text", + "content": "(a) ECE over epochs using CE, Focal Loss and BSCE-GRA." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 230, + 72, + 380, + 159 + ], + "blocks": [ + { + "bbox": [ + 230, + 72, + 380, + 159 + ], + "lines": [ + { + "bbox": [ + 230, + 72, + 380, + 159 + ], + "spans": [ + { + "bbox": [ + 230, + 72, + 380, + 159 + ], + "type": "image", + "image_path": "b782b0e452c63074844c72f91585b72b3f4f6b60c49927e187f68034089a55ea.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 230, + 162, + 380, + 180 + ], + "lines": [ + { + "bbox": [ + 230, + 162, + 380, + 180 + ], + "spans": [ + { + "bbox": [ + 230, + 162, + 380, + 180 + ], + "type": "text", + "content": "(b) Gradient Density at epoch 150 using CE, Focal Loss and BSCE-GRA." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 190, + 555, + 213 + ], + "lines": [ + { + "bbox": [ + 55, + 190, + 555, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 555, + 213 + ], + "type": "text", + "content": "Figure 5. Figure 5a presents the evolution of ECE throughout the training process, demonstrating that our method rapidly converges to the best result by epoch 250. The subsequent figures depict the gradient magnitudes of various methods between epochs 150 and 250." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 403, + 72, + 553, + 159 + ], + "blocks": [ + { + "bbox": [ + 403, + 72, + 553, + 159 + ], + "lines": [ + { + "bbox": [ + 403, + 72, + 553, + 159 + ], + "spans": [ + { + "bbox": [ + 403, + 72, + 553, + 159 + ], + "type": "image", + "image_path": "1f347dbf8a6d7a902fc393558799e170acc7e9d3bf84dc8076516982c7d6fc3f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 162, + 553, + 180 + ], + "lines": [ + { + "bbox": [ + 403, + 162, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 403, + 162, + 553, + 180 + ], + "type": "text", + "content": "(c) Gradient Density at epoch 250 using CE, Focal Loss and BSCE-GRA." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 234, + 296, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 234, + 296, + 438 + ], + "spans": [ + { + "bbox": [ + 54, + 234, + 296, + 438 + ], + "type": "text", + "content": "Cross-Entropy. It also indicates that, during training with a moderate learning rate from epochs 150 to 200, Focal Loss tends to produce better-calibrated models than BSCE-GRA. This may be due to the fact that, during these epochs, the model makes more predictions with mid-range confidence levels, and as shown in Figure 1, Focal Loss directs the model's attention towards these moderately uncertain samples. In contrast, BSCE-GRA imposes stronger regularization on the gradient, resulting in smaller optimization steps compared to Focal Loss. To further validate this hypothesis, we present the gradient density of the last linear layer across the entire training set at epochs 150 and 250 in Figure 5b and Fig. 5c, which shows that BSCE-GRA results in smaller gradient magnitudes during these epochs. After 250 epochs, when the learning rate undergoes its second reduction, the model trained with BSCE-GRA soon achieves better calibration performance compared to FL and CE." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 445, + 244, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 244, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 244, + 458 + ], + "type": "text", + "content": "5.2. Gradient Value among Uncertainty" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 462, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 296, + 714 + ], + "type": "text", + "content": "We extend our analysis to examine the relationship between gradient magnitude and uncertainty. Specifically, we compute the gradient norms of the last linear layer for all samples in the training set at epochs 50, 150, 250, and 350. To quantify uncertainty, we calculate the Brier Score for each training sample. We present the results for Focal Loss (FL), Dual Focal Loss (DFL), and BSCE-GRA. Figure 4 illustrates the relationship between gradient magnitude and Brier Score. The figure clearly shows that the gradients produced by BSCE-GRA are the most sensitive to changes in the Brier Score, as indicated by the narrow distribution compared to other loss functions. This aligns with our goal: we aim for BSCE-GRA to exhibit sufficient sensitivity to uncertainty, allowing the model to adjust its gradient values based on changes in uncertainty, thereby focusing more on highly uncertain samples. Notably, the gradient distribution for Dual Focal Loss is relatively more dispersed concerning the Brier Score, with a wider range of possible values compared to other loss functions. This could be attributed to the partial derivatives involving other classes, as Dual Focal Loss takes the second most probable class into account" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 234, + 555, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 555, + 258 + ], + "type": "text", + "content": "in its loss calculation. In contrast, Focal Loss and BSCE-GRA only involve gradients along a single dimension." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 279, + 511, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 511, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 511, + 293 + ], + "type": "text", + "content": "5.3. Weighting FL and DFL on Gradients" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 300, + 555, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 385 + ], + "type": "text", + "content": "We evaluate the performance of directly applying the uncertainty terms " + }, + { + "bbox": [ + 313, + 300, + 555, + 385 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{FL}}" + }, + { + "bbox": [ + 313, + 300, + 555, + 385 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 300, + 555, + 385 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{DFL}}" + }, + { + "bbox": [ + 313, + 300, + 555, + 385 + ], + "type": "text", + "content": " as weights on the gradients. Experiments are conducted using the default settings discussed above, on CIFAR-10 with ResNet-50. The results are presented in Table 2. It is evident that applying weights on gradients results in performance improvement for both " + }, + { + "bbox": [ + 313, + 300, + 555, + 385 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{FL}}" + }, + { + "bbox": [ + 313, + 300, + 555, + 385 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 300, + 555, + 385 + ], + "type": "inline_equation", + "content": "u_{\\mathrm{DFL}}" + }, + { + "bbox": [ + 313, + 300, + 555, + 385 + ], + "type": "text", + "content": ", further validating its effectiveness." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 389, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 437 + ], + "type": "text", + "content": "We conduct additional experiments to comprehensively validate our proposed method, BSCE-GRA, under various settings. Due to page limitations, the extended experiments are provided in the Appendix." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 461, + 388, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 388, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 388, + 474 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 486, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 486, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 486, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we proposed a novel approach to model calibration by directly weighting gradients based on uncertainty. We analyzed the strengths and limitations of Focal Loss and Dual Focal Loss from the perspective of sample weighting and introduced a framework that scales gradient magnitudes based on model uncertainty to focus more on uncertain samples. Additionally, we introduced BSCEGRA, a loss function incorporating uncertainty metrics to enhance model calibration. Extensive experiments on various datasets and network architectures demonstrated significant improvements in calibration performance, achieving state-of-the-art results. Our findings emphasize the value of integrating uncertainty-aware mechanisms directly into the optimization process, providing a reliable framework for training calibrated deep neural networks suitable for real-world applications requiring trustworthy predictions. Future work may explore further optimization strategies for uncertainty-weighted approaches and their impact on tasks like active learning and robustness to adversarial attacks." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15504" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 58, + 71, + 159, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 71, + 159, + 85 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 159, + 85 + ], + "type": "text", + "content": "7. Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 294, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 91, + 294, + 149 + ], + "spans": [ + { + "bbox": [ + 57, + 91, + 294, + 149 + ], + "type": "text", + "content": "This work was supported in part by the Start-up Grant (No. 9610680) of the City University of Hong Kong, Young Scientist Fund (No. 62406265) of NSFC, and the Australian Research Council under Projects DP240101848 and FT230100549." + } + ] + } + ], + "index": 1 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "text", + "content": "15505" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Ondrej Bohdal, Yongxin Yang, and Timothy Hospedales. Meta-calibration: Learning of model calibration using differentiable expected calibration error. arXiv preprint arXiv:2106.09613, 2021. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 135, + 296, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 135, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 135, + 296, + 167 + ], + "type": "text", + "content": "[2] Glenn W Brier. Verification of forecasts expressed in terms of probability. Monthly weather review, 78(1):1-3, 1950. 1, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 168, + 295, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 168, + 295, + 191 + ], + "spans": [ + { + "bbox": [ + 62, + 168, + 295, + 191 + ], + "type": "text", + "content": "[3] Corinna Cortes. Support-vector networks. Machine Learning, 1995. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 192, + 295, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 192, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 192, + 295, + 236 + ], + "type": "text", + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 236, + 295, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 295, + 280 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 295, + 280 + ], + "type": "text", + "content": "[5] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In International conference on machine learning, pages 1321-1330. PMLR, 2017. 1, 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 281, + 295, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 281, + 295, + 325 + ], + "spans": [ + { + "bbox": [ + 62, + 281, + 295, + 325 + ], + "type": "text", + "content": "[6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 325, + 295, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 325, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 62, + 325, + 295, + 380 + ], + "type": "text", + "content": "[7] Dan Hendrycks*, Norman Mu*, Ekin Dogus Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. Aug-mix: A simple method to improve robustness and uncertainty under data shift. In International Conference on Learning Representations, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 380, + 295, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 425 + ], + "type": "text", + "content": "[8] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700-4708, 2017. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 426, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 426, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 426, + 295, + 479 + ], + "type": "text", + "content": "[9] Like Hui and Mikhail Belkin. {EVALUATION} {of} {neural} {architectures} {trained} {with} {square} {loss} {vs} {cross}-{entropy} {in} {classification} {tasks}. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 480, + 295, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 480, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 57, + 480, + 295, + 534 + ], + "type": "text", + "content": "[10] Archit Karandikar, Nicholas Cain, Dustin Tran, Balaji Lakshminarayanan, Jonathon Shlens, Michael C Mozer, and Becca Roelofs. Soft calibration objectives for neural networks. Advances in Neural Information Processing Systems, 34:29768-29779, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 536, + 295, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 295, + 579 + ], + "type": "text", + "content": "[11] Ranganath Krishnan and Omesh Tickoo. Improving model calibration with accuracy versus uncertainty optimization. Advances in Neural Information Processing Systems, 33: 18237-18248, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 581, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 295, + 602 + ], + "type": "text", + "content": "[12] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6, 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 603, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 295, + 656 + ], + "type": "text", + "content": "[13] Meelis Kull, Telmo Silva Filho, and Peter Flach. Beta calibration: a well-founded and easily implemented improvement on logistic calibration for binary classifiers. In Artificial intelligence and statistics, pages 623-631. PMLR, 2017. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "text", + "content": "[14] Meelis Kull, Miquel Perello Nieto, Markus Kängsepp, Telmo Silva Filho, Hao Song, and Peter Flach. Beyond temperature scaling: Obtaining well-calibrated multi-class probabilities with dirichlet calibration. Advances in neural information processing systems, 32, 2019. 2, 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[15] Aviral Kumar, Sunita Sarawagi, and Ujjwal Jain. Trainable calibration measures for neural networks from kernel mean embeddings. In International Conference on Machine Learning, pages 2805-2814. PMLR, 2018. 1, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 163 + ], + "type": "text", + "content": "[16] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 163, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 185 + ], + "type": "text", + "content": "[17] T Lin. Focal loss for dense object detection. arXiv preprint arXiv:1708.02002, 2017. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "type": "text", + "content": "[18] Jishnu Mukhoti, Viveka Kulharia, Amartya Sanyal, Stuart Golodetz, Philip Torr, and Puneet Dokania. Calibrating deep neural networks using focal loss. Advances in Neural Information Processing Systems, 33:15288-15299, 2020. 1, 2, 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 241, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 275 + ], + "type": "text", + "content": "[19] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "text", + "content": "[20] Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. In Proceedings of the AAAI conference on artificial intelligence, 2015. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 321, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 374 + ], + "type": "text", + "content": "[21] Khanh Nguyen and Brendan O'Connor. Posterior calibration and exploratory analysis for natural language processing models. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1587-1598, 2015. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 376, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 430 + ], + "type": "text", + "content": "[22] Joshua C Peterson, Ruairidh M Battleday, Thomas L Grifths, and Olga Russakovsky. Human uncertainty makes classification more robust. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9617-9626, 2019. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 432, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 553, + 466 + ], + "type": "text", + "content": "[23] John Platt et al. Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Advances in large margin classifiers, 10(3):61-74, 1999. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 467, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 467, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 467, + 553, + 499 + ], + "type": "text", + "content": "[24] Rahul Rahaman et al. Uncertainty quantification and deep ensembles. Advances in neural information processing systems, 34:20063-20075, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "type": "text", + "content": "[25] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 545, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 553, + 599 + ], + "type": "text", + "content": "[26] Linwei Tao, Minjing Dong, Daochang Liu, Changming Sun, and Chang Xu. Calibrating a deep neural network with its predecessors. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 4271–4279, 2023. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "type": "text", + "content": "[27] Linwei Tao, Minjing Dong, and Chang Xu. Dual focal loss for calibration. In International Conference on Machine Learning, pages 33833-33849. PMLR, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "text", + "content": "[28] Linwei Tao, Minjing Dong, and Chang Xu. Feature clipping for uncertainty calibration. arXiv preprint arXiv:2410.19796, 2024. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "text", + "content": "[29] Linwei Tao, Haolan Guo, Minjing Dong, and Chang Xu. Consistency calibration: Improving uncertainty calibration via consistency among perturbed neighbors. arXiv preprint arXiv:2410.12295, 2024. 6" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15506" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 396 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[30] Linwei Tao, Younan Zhu, Haolan Guo, Minjing Dong, and Chang Xu. A benchmark study on calibration. In The Twelfth International Conference on Learning Representations, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "text", + "content": "[31] Sunil Thulasidasan, Gopinath Chennupati, Jeff A Bilmes, Tanmoy Bhattacharya, and Sarah Michalak. On mixup training: Improved calibration and predictive uncertainty for deep neural networks. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "type": "text", + "content": "[32] Deng-Bao Wang, Lei Feng, and Min-Ling Zhang. Rethinking calibration of deep neural networks: Do not be afraid of overconfidence. Advances in Neural Information Processing Systems, 34:11809-11820, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 261 + ], + "type": "text", + "content": "[33] Jason Weston, Samy Bengio, and Nicolas Usunier. Wsabie: Scaling up to large vocabulary image annotation. In Twenty-Second International Joint Conference on Artificial Intelligence. CiteSeer, 2011. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 317 + ], + "type": "text", + "content": "[34] Bianca Zadrozny and Charles Elkan. Transforming classifier scores into accurate multiclass probability estimates. In Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining, pages 694-699, 2002. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 294, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 351 + ], + "type": "text", + "content": "[35] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In *British Machine Vision Conference* 2016. British Machine Vision Association, 2016. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 353, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 294, + 396 + ], + "type": "text", + "content": "[36] Jize Zhang, Bhavya Kailkhura, and T Yong-Jin Han. Mix-n-match: Ensemble and compositional methods for uncertainty calibration in deep learning. In International conference on machine learning, pages 11117-11128. PMLR, 2020. 2" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15507" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction/308073a9-0b23-46de-a5fd-816ea2a49dc7_content_list.json b/2025/Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction/308073a9-0b23-46de-a5fd-816ea2a49dc7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d81027080aef84651bf7e073741d698287814c4a --- /dev/null +++ b/2025/Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction/308073a9-0b23-46de-a5fd-816ea2a49dc7_content_list.json @@ -0,0 +1,1674 @@ +[ + { + "type": "text", + "text": "Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction", + "text_level": 1, + "bbox": [ + 192, + 130, + 805, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaolu Liu $^{1*}$ , Ruizi Yang $^{1*}$ , Song Wang $^{1}$ , Wentong Li $^{2}$ , Junbo Chen $^{3\\dagger}$ , Jianke Zhu $^{1\\dagger}$ $^{1}$ Zhejiang University $^{2}$ NUAA $^{3}$ Udeer.ai", + "bbox": [ + 168, + 202, + 833, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{xiaoluliu, ruiziyang, songw, jkzhu}@zju.edu.cn, wentong.li@nuaa.edu.cn, junbo@udeer.ai", + "bbox": [ + 112, + 241, + 877, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reliable high-definition (HD) map construction is crucial for the driving safety of autonomous vehicles. Although recent studies demonstrate improved performance, their generalization capability across unfamiliar driving scenes remains unexplored. To tackle this issue, we propose UIGenMap, an uncertainty-instructed structure injection approach for generalizable HD map vectorization, which concerns the uncertainty resampling in statistical distribution and employs explicit instance features to reduce excessive reliance on training data. Specifically, we introduce the perspective-view (PV) detection branch to obtain explicit structural features, in which the uncertainty-aware decoder is designed to dynamically sample probability distributions considering the difference in scenes. With probabilistic embedding and selection, UI2DPrompt is proposed to construct PV-learnable prompts. These PV prompts are integrated into the map decoder by designed hybrid injection to compensate for neglected instance structures. To ensure real-time inference, a lightweight Mimic Query Distillation is designed to learn from PV prompts, which can serve as an efficient alternative to the flow of PV branches. Extensive experiments on challenging geographically disjoint (geo-based) data splits demonstrate that our UIGenMap achieves superior performance, with $+5.7$ mAP improvement on the nuScenes dataset. Source code is available at https://github.com/xiaolul2/UIGenMap.", + "bbox": [ + 88, + 323, + 485, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 746, + 220, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reliable HD map construction is crucial for driving safety of autonomous vehicles, which provides essential geometric information for ego-localization, perceptions, path planning, and predictions [10, 13, 23, 45]. Recently, online high-definition (HD) map vectorization has attracted considerable attention for its cost-effectiveness and ability to", + "bbox": [ + 89, + 772, + 483, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5656bbe841dca3b885a7c186417c1316f1e3d9ed1f752b34ffcabae4674ef3af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 290, + 694, + 439 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9156bb1d437084df267619f6a5d8a538c3c80a72c9065125b6e56e780b50e116.jpg", + "image_caption": [ + "Figure 1. (a) Comparisons on original and different geo-based data partitions among previous methods and ours. (b) Visual comparisons with circles on BEV map to represent the learned uncertainties, in which PV instances serve as compensation and provide a detailed understanding of map perception with generalization." + ], + "image_footnote": [], + "bbox": [ + 696, + 290, + 905, + 441 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "provide real-time updates [25, 31, 56]. Given images captured by onboard cameras, current map construction relies on transformer scheme [57] to convert perspective-view (PV) image features into bird's-eye-view (BEV) space. Then, a decoder follows to predict the map elements.", + "bbox": [ + 511, + 531, + 906, + 607 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the achievements on widely adopted benchmarks [4, 26, 28], little attention has been paid to investigating the generalization capabilities, which refer to the model's performance on unfamiliar driving scenarios that are different from the training sets. As discussed in [27, 50], in the widely used public datasets [1, 46], there are a large number of geographical overlaps between the training and validation scenes. The models are then more likely to memorize similar driving scenes rather than fully capture their underlying road structures. Variations in driving scenes, such as the contrast between bustling city centers and suburbs with simple road structures, highlight significant shifts in feature distributions and reveal the model's limitations in robustness. Besides, learning-based BEV transformation will inevitably lead to geometric errors and loss of texture details. As shown in Figure 1(a), previous methods [25, 50] exhibit notable performance degradation on region-based and city-based data splits under changing driving scenes.", + "bbox": [ + 511, + 609, + 908, + 882 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To tackle the above issue, we propose the innovative em", + "bbox": [ + 532, + 885, + 903, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution", + "bbox": [ + 107, + 875, + 243, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors", + "bbox": [ + 109, + 887, + 238, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22359", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "employment of model uncertainty targeted for more generalizable HD map constructions. On the one hand, the uncertainty estimation process involves learning both the statistical mean and variance value, enabling dynamic resampling based on probabilistic distributions. Given the discrepancy in feature distribution across different driving scenes, such a statistical resampling strategy enhances the model's dynamic adaptability in unfamiliar environments. However, the estimated model uncertainty quantifies the confidence in its predictions, which can serve as a reliable indicator for further feature selection and enhancement. As illustrated in Figure 1(b), direct 2D perception can capture more intuitive semantic and angular structural information, which might be missed or incorrectly learned during the implicit PV-to-BEV conversions. Thus, combined with uncertainty-based feature selection, more reliable explicit information can be emphasized as compensation to boost the robustness of BEV-level map construction.", + "bbox": [ + 89, + 90, + 480, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the above discussion, in this work, we propose UIGenMap, an uncertainty-instructed PV structure injection strategy for more generalizable HD map vectorization. Specifically, a PV detection branch is introduced to provide explicit structural instances. In both PV and BEV branches, we move beyond the vanilla transformer structure by designing the statistically sensitive uncertainty-aware decoder (UA-Decoder), which constructs the probabilistic outputs by uncertainty-aware attention and finer point-level uncertainty head. Leveraging the PV uncertainty output, we introduce the UI2DPrompt to construct learnable PV prompts. These PV prompts are integrated into the BEV decoder by the designed hybrid injection mechanism, which acts as structural compensation for the BEV map prediction. Consequently, reliable structural representation from PV space can be injected to enhance the model's capability, thereby improving its perceptual ability in complex environments. Considering the demands for real-time inference, we also design a lightweight mimic query distillation module to learn from constructed PV prompts. Then during inference, the mimicked query can replace the PV detection flow to reduce extra computational burdens.", + "bbox": [ + 91, + 366, + 482, + 698 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To prioritize the evaluation of generalization capability, we assess our method using geographically disjoint (geo-based) data partitions on the nuScenes [1] and Argoverse2 [46] datasets. Extensive experiments demonstrate that our UIGenMap achieves state-of-the-art performance with higher adaptive ability, especially with $+5.7$ mAP improvement on the region-based partition and $+4.3$ mAP on the more challenging city-based split on nuScenes. The main contributions can be summarized as follows:", + "bbox": [ + 89, + 700, + 480, + 835 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose UIGenMap, an uncertainty-instructed approach to more generalizable HD map vectorization with explicit structure injection.", + "- The UA-Decoder and UI2DPrompt are proposed to con" + ], + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "struct structural PV prompts, which are dynamically integrated with BEV perception by hybrid injection.", + "bbox": [ + 526, + 90, + 903, + 119 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Lightweight MQ-Distillation design supports real-time inference without the additional computational overhead of the PV detection branches.", + "- Promising results on more challenging geo-based data partitions show that UIGenMap outperforms previous approaches and achieves powerful generalization capability." + ], + "bbox": [ + 511, + 121, + 903, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 224, + 653, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vectorized HD Map Construction. Conventional HD map construction approaches [38, 39, 58] rely on SLAM-based annotations, which require tedious offline manual labeling at high costs. Recent studies [6, 11, 19, 20, 29, 37, 49] have explored online approaches with the input of onboard sensor data. In [18, 22, 24, 35, 42], HDMap construction is treated as a segmentation task, relying on post-processing to achieve map vectorization. Later, Liao et al. [25, 26] represent map elements as ordered points and employ transformers [41, 57] after the BEV encoder for auto-regressive detection. Other modeling methods, such as Bezier curves and pivot points, are explored in BeMapNet [37] and PivotNet [6] for geometric representation. Several works [4, 19, 40, 43, 50] are proposed for different temporal fusion approaches. GeMap [54] and MapQR [31] improve mapping performance from the perspectives of geometric relationships and decoder design, respectively. In [5, 28, 51, 52], rasterized maps are utilized for structure compensation. Despite improved performance, these approaches seldom consider the generalization capability of models in unfamiliar driving scenarios. In this work, we incorporate uncertainty-instructed PV detection retracing, aiming for better performance across diverse scenarios.", + "bbox": [ + 511, + 251, + 906, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "PV Detection for 3D Perception. Instances from PV detection can be effective compensations for 3D perceptions [12, 14, 32, 47, 48, 53]. For 3D object detection, BEVFormerv2 [48] utilizes a two-stage BEV detector that integrates the query from the PV head into the BEV space. In SimMoD [53], 2D detection is utilized to construct proposals for 3D fine-grained perception. Far3D [14] combines 2D detection boxes with estimated depth information to improve long-range perception performance. For model generalization, Lu et al. [32] propose a generalizable framework based on perspective debiasing, in which PV consistency learning is employed among different domains. Applying for 3D lanes, Li et al. [17] use 2D priors for lane detection and topology reasoning. Inspired by the above methods, we employ PV structural features for more generalizable HD map vectorization, in which distillations from 2D instances are utilized to ensure inference efficiency.", + "bbox": [ + 511, + 599, + 905, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Uncertainty in Model. Generally, uncertainty refers to the model's confidence level for its predictions, which can be a statistical representation to indicate the model's under", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22360", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b6bbcd069bbfbe9d38546f9957fe490aa974ee4bdbc101c565d45ed913b89ba3.jpg", + "image_caption": [ + "(a) Overview", + "Figure 2. (a) Overview of Our UIGenMap. For training, the PV branch is introduced with the uncertainty-instructed structural injection, in which the MQ-Distillation is designed to mimic PV structural features. (b) Uncertainty-Aware Decoder Architecture. For each layer, UA-Decoder comprises probabilistic UA-Attention and UA-Head for reliable output. (c) UI2DPrompt Design. We construct PV prompts from PV-detected elements and their corresponding uncertainties, which are integrated with the main branch by hybrid injection." + ], + "image_footnote": [], + "bbox": [ + 99, + 107, + 390, + 383 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b3d856e728e09cd8b9f9e9daa043ffb7d42a9757d215f705d8449fdb9a96b0b6.jpg", + "image_caption": [ + "(b) UA-Decoder" + ], + "image_footnote": [], + "bbox": [ + 401, + 102, + 893, + 222 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ad0eba1419bc33d3c8b2b80bd59af6aa08f0016479f848884af8fecaee102b92.jpg", + "image_caption": [ + "(c) UI2DPrompt" + ], + "image_footnote": [], + "bbox": [ + 403, + 237, + 887, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "standing of scenes [3, 15, 44]. Uncertainty has gained intensive attention due to the potential for reliability evaluation and performance enhancement [2, 16]. Specifically, the generalization problem with uncertainty modeling is investigated in [21, 36], which expects uncertainty estimation for out-of-distribution generalization. For the transformer structure [8, 34], uncertainty has been explored to guide the training and inference processes for more robust model performance. In the field of autonomous driving, significant uncertainty has been introduced for 3D occupancy [2, 44], optical flow [33], and object tracking [55]. Recently, Gu et al. [7] present uncertainty estimation for the vectorized map, which is targeted for trajectory prediction. In our design, we utilize uncertainty for more generalized HD map construction, which considers the variability in feature distributions and incorporates explicit instance features.", + "bbox": [ + 88, + 460, + 485, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. UIGenMap", + "text_level": 1, + "bbox": [ + 89, + 720, + 210, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given surrounding-view images captured from onboard cameras, we aim to construct more robust vectorized map elements in the BEV space. Each map element comprises a class label $c_{\\mathrm{map}}$ and an ordered sequence of points $\\mathbf{P}_{\\mathrm{map}} = \\{(p_x^i,p_y^i)\\}_{i = 1}^N$ that represents the structure of the instance. $N$ denotes the total number of points for each map element.", + "bbox": [ + 89, + 747, + 482, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The architecture of UIGenMap is presented in Figure 2. Based on the main BEV detection, we obtain PV instance from an introduced pre-trained PV branch, which is injected as structural priors for BEV perception. Firstly, we extract", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "image features for the PV and BEV branches separately. As in [22], BEV features $\\mathrm{F_{bev}}$ are constructed from the interaction between PV features and learnable BEV queries. Then, we design the uncertainty-aware (UA-) decoder from inner-layer UA-Attention and point-level heads to achieve reliable probabilistic outputs. Based on the PV coordinates and uncertainty output, we propose the UI2DPrompt module to construct reliable PV prompts. These prompts are incorporated into BEV features and map instance queries via hybrid injection, where uncertainties are utilized as the basis for feature selection and enhancement. Finally, the UA-Decoder employs enhanced BEV features and instance queries for reliable map predictions. To ensure real-time inference, the lightweight MQ-Distillation module performs query distillation from learned PV prompts to imitate explicit structural features, aiming for efficient computation.", + "bbox": [ + 511, + 460, + 906, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Uncertainty-Aware Decoder", + "text_level": 1, + "bbox": [ + 511, + 712, + 764, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We design UA-Decoder for both PV and BEV space. Systematic uncertainty-aware design is employed to improve the statistical understanding of features and outputs to achieve more dynamic self-adjustment for different driving scenarios. To ensure comprehensive perception, each decoder layer comprises UA-Attention at the instance feature level and UA-Head for point-level uncertainty output along with positions, respectively.", + "bbox": [ + 511, + 734, + 906, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "UA-Attention. In conventional deformable attention [57], the weight $\\alpha_{i}$ for each sampled feature values from BEV feature $F_{\\mathrm{BEV}}$ is determined by linear layer (Linear(·)) on", + "bbox": [ + 511, + 854, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22361", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e4226113f26f405ad6d4e9364b9a37a32d401984fdacb846dc0b5bcc270199f6.jpg", + "image_caption": [ + "Figure 3. UA-Attention Design. The learned probabilistic weights $\\alpha_{i}$ are dynamically resampled from Gaussian distribution with stochastic disturbance, which is multiplied with sampled feature values to update map queries." + ], + "image_footnote": [], + "bbox": [ + 94, + 90, + 482, + 261 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$Q_{\\mathrm{map}}^i$ , which can be expressed as $\\alpha_{i} = \\mathrm{Linear}(Q_{\\mathrm{map}}^{i})$ where $Q_{\\mathrm{map}}^i$ is the $i$ -th instance query of totally $M$ map elements. Then, the updated part of query $\\Delta Q_{\\mathrm{map}}^i$ is aggregated by the dot-product between the learned weight $\\alpha_{i}$ and sampled feature values from $\\mathbf{F}_{\\mathrm{bev}}$ . Considering complex driving scenarios, such deterministic weights are inadaptable for challenging situations. More dynamic weights are required to avoid potential ambiguities. Inspired by [8, 34], we explicitly model the attention scores under a Gaussian distribution $\\alpha_{i}\\sim \\mathcal{N}(\\mu_{i},\\sigma_{i}^{2})$ , as shown in Figure 3. We adopt the re-parameterization strategy to perform the probabilistic attention, which is constructed by the corresponding mean $\\mu_{i}$ and variance $\\sigma_{i}$ . Such probabilistic parameters are predicted through multilayer perception (MLP(·)). Thus, the construction of dynamic resampled attention weights can be formulated as below:", + "bbox": [ + 89, + 329, + 483, + 573 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\mu_ {i}, \\sigma_ {i} ^ {2} \\right\\} = \\mathrm {M L P} \\left(Q _ {\\text {m a p}} ^ {i}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 587, + 480, + 604 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i} = \\mu_ {i} + \\sigma_ {i} \\epsilon , \\epsilon \\sim \\mathcal {N} (0, 1), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 609, + 480, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon$ is the Gaussian noise conforming to normal distribution. During inference, attention weight $\\alpha_{i}$ can be dynamically resampled from learned statistical distribution to suit for driving scenarios with large variations. Then query features can be obtained by dot production with the sampled feature values in order to enable reliable query updates.", + "bbox": [ + 89, + 636, + 482, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "UA-Head with Uncertainty Loss. Based on the robust design of UA-Attention, quantized uncertainty for each lane point output is also essential, which is point-specific at a fine-grained level to explicitly reflect the reliability and uncertainty information within the internal transformer layers. Inspired by [7], we replace the vanilla deterministic output with the UA-Head that predicts coordinates $\\mathbf{P}_{\\mathrm{map}} = \\{(\\hat{p}_x^i,\\hat{p}_y^i)\\}_{i = 1}^N$ along with their corresponding uncertainties $\\sigma_{\\mathrm{map}} = \\{(\\sigma_x^i,\\sigma_y^i)\\}_{i = 1}^N$ for each map element, as shown in Figure 2 (b). These parameters represent the location and scale of the joint Laplace distribution $Laplace(\\hat{p}_j^i,\\frac{\\sigma_j^i}{\\sqrt{2}})$ at", + "bbox": [ + 89, + 728, + 483, + 904 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the $j$ -th dimension $(x$ or $y)$ of the $i$ -th point, which refers to a probabilistic representation for stable output.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For uncertainty-aware training, we combine probabilistic Negative Log-Likelihood (NLL) loss $\\mathcal{L}_{\\mathrm{nll}}$ with point regression loss $\\mathcal{L}_{\\mathrm{pts}}$ in [50] to balance uncertainties and coordinate outputs, in which $\\mathcal{L}_{\\mathrm{nll}}$ can be formulated as follows:", + "bbox": [ + 511, + 121, + 903, + 181 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {n l l}} = \\lambda_ {\\mathrm {n l l}} \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {2} \\left(\\log \\left(2 \\sigma_ {j} ^ {i}\\right) + \\frac {\\left| p _ {j} ^ {i} - \\hat {p} _ {j} ^ {i} \\right|}{\\sigma_ {j} ^ {i}}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 190, + 903, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $\\lambda_{\\mathrm{nl}}$ as the loss weight setting.", + "bbox": [ + 511, + 242, + 745, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Figure 2 (a), UA-Decoder is used for the construction of the PV branch and the BEV map. To provide more robust structural features in PV branch, PV coordinates $\\mathrm{P_{pv}} = \\{x_{pv}^i,y_{pv}^i\\}_{i = 1}^N$ and the uncertainty parameters $\\sigma_{\\mathrm{pv}} = \\{\\sigma_{\\mathrm{pv,x}}^{i},\\sigma_{\\mathrm{pv,y}}^{i}\\}_{i = 1}^{N}$ are constructed by the UA-Decoder structure, which is later utilized in our UI2DPrompt design. Thus, the systematic combination of uncertainty-aware design provides dynamic sampling during the learning process, along with established uncertainty, to construct a reliable basis for confidence scoring.", + "bbox": [ + 511, + 257, + 906, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. UI2DPrompt", + "text_level": 1, + "bbox": [ + 511, + 417, + 651, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "PV branches contain abundant structural and geometric details. To mitigate information loss and misinterpretation during query-based PV-to-BEV transformation, reliable structural PV instances can be injected as compensation for BEV prediction. This section focuses on constructing PV prompts based on uncertainty-aware PV predictions, which are dynamically integrated with the map decoder for BEV map element detection.", + "bbox": [ + 511, + 440, + 905, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prompt Construction. Given the learned PV instance coordinates and the corresponding uncertainty parameters $\\{\\mathrm{P_{pv}},\\sigma_{\\mathrm{pv}}\\}$ , PV prompts are constructed to provide dynamic 2D priors. Firstly, we employ the estimated classification score $c_{\\mathrm{map}}$ for PV candidates selection. When $c_{\\mathrm{map}} > c_{\\mathrm{thr}}$ , PV instances are chosen as candidates for reliable localized information. Considering the sparse depth information and the characteristics of static map elements on the ground, we utilize Inverse Perspective Mapping (IPM) to transform $\\mathrm{P_{pv}}$ from the PV image coordinate to the BEV coordinate system $\\mathrm{P}_{\\mathrm{trans}} = \\{p_x^{\\prime},p_y^{\\prime}\\}$ , which makes a flat grounding assumption and utilizes the affine transformation between the PV and BEV coordinates. Given the corresponding intrinsic parameters $\\mathbf{K}_{\\mathrm{cam}}$ and extrinsic transformation matrix $\\mathrm{T}_{\\mathrm{ego2cam}}$ , the projection $\\mathrm{P_{pv}}$ can be defined as:", + "bbox": [ + 511, + 561, + 906, + 789 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {\\text {t r a n s}} = \\operatorname {P r o j} _ {2 D} \\left(\\mathrm {T} _ {\\text {e g o 2 c a m}} ^ {- 1} \\cdot \\left[ \\begin{array}{c} \\mathrm {K} _ {\\text {c a m}} ^ {- 1} \\cdot \\mathrm {P} _ {\\mathrm {p v}} \\\\ 1 \\end{array} \\right]\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 797, + 903, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\operatorname{Proj}_{2\\mathrm{D}}(\\cdot)$ denotes the conversion from homogeneous coordinates into 2D plane coordinates.", + "bbox": [ + 511, + 839, + 903, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To ensure reliable enhanced features, the corresponding uncertainty parameters $\\sigma_{\\mathrm{pv}}$ cooperate as weights and feature", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22362", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "embeddings. We encode the transformed coordinates $\\mathbf{P}_{\\mathrm{trans}}$ and uncertainty parameters $\\sigma_{\\mathrm{pv}}$ by $\\phi_p(\\cdot)$ and $\\phi_{\\sigma}(\\cdot)$ , respectively. Point-specific embeddings $\\{e_{pv}^i\\}_{i = 1}^N$ are constructed by concatenation:", + "bbox": [ + 89, + 90, + 482, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {e} _ {\\mathrm {p v}} ^ {i} = \\operatorname {C o n c a t} \\left(\\left[ \\phi_ {p} \\left(\\mathrm {P} _ {\\text {t r a n s}}\\right), \\phi_ {\\sigma} \\left(\\sigma_ {\\mathrm {p v}}\\right) \\right]\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 157, + 482, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thus, the structural PV output coordinates and uncertainty information are dynamically encoded.", + "bbox": [ + 89, + 181, + 482, + 213 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the above, PV uncertainties can serve as the quantized weights for dynamic feature strengthening and adjustment. Reliable confidence indicator is able to be formulated as the weight $\\{\\omega_{\\mathrm{pv}}^i\\}_{i = 1}^N$ concerning each point-specific prompt. Then, the enhanced PV prompt $\\tilde{\\pmb{e}}_{\\mathrm{pv}}^{i}$ can be constructed by:", + "bbox": [ + 89, + 213, + 483, + 305 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\omega_ {\\mathrm {p v}} ^ {i} = \\exp \\left(\\frac {\\left(\\| \\sigma_ {\\mathrm {p v}} ^ {i} \\| _ {2}\\right) ^ {- 1}}{\\sum_ {i = 1} ^ {N} \\left(\\| \\sigma_ {\\mathrm {p v}} ^ {i} \\| _ {2}\\right) ^ {- 1}}\\right), \\quad \\tilde {\\boldsymbol {e}} _ {\\mathrm {p v}} ^ {i} = \\omega_ {\\mathrm {p v}} ^ {i} \\boldsymbol {e} _ {\\mathrm {p v}} ^ {i} + \\boldsymbol {e} _ {\\mathrm {m}} ^ {i}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 311, + 482, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\exp (\\cdot)$ represents the exponential amplification of the normalized inverse uncertainty value. $\\{e_{\\mathrm{m}}^{i}\\}_{i = 1}^{N}$ is the predefined mimic query, which is utilized for MQ-Distillation. This approach reduces the introduction of PV errors, ensuring that injecting PV instances amplifies reliable PV information while minimizing uncertainty-induced errors.", + "bbox": [ + 89, + 364, + 482, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hybrid Injection. Constructed from PV instances with uncertainty output, PV prompts $\\{\\tilde{e}_{\\mathrm{pv}}^i\\}_{i = 1}^N$ are incorporated as compensation for the construction of the BEV map. These prompts are injected into the BEV feature $\\mathbf{F}_{\\mathrm{bev}}$ and the map queries $Q_{\\mathrm{map}}^i$ for the map decoder by the designed hybrid injection mechanism, which includes both prompt-to-BEV (P2BEV) and prompt-to-query (P2Q) attention.", + "bbox": [ + 89, + 455, + 483, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Figure 4, point-level PV prompts $\\{\\tilde{e}_{\\mathrm{pv}}^i\\}_{i = 1}^N$ are integrated with $\\mathbf{F}_{\\mathrm{bev}}$ through cross-attention of P2BEV. This step enriches BEV features by integrating detailed, point-specific information from PV prompts, enhancing instance-specific detail in the BEV representation:", + "bbox": [ + 89, + 561, + 483, + 637 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {F} _ {\\mathrm {b e v}} ^ {\\prime} = \\operatorname {A t t n} \\left(W _ {q} \\mathrm {F} _ {\\mathrm {b e v}}, W _ {k} \\tilde {\\boldsymbol {e}} _ {\\mathrm {p v}} ^ {i}, W _ {v} \\tilde {\\boldsymbol {e}} _ {\\mathrm {p v}} ^ {i}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 643, + 482, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $W_{\\{q,k,v\\}}$ are the corresponding projection matrix. Then, we encode point-level prompts into instance-level representations through a linear block $\\phi_e(\\cdot)$ , which aggregates point-specific information to create a higher-level representation for subsequent processing. These instance-level prompts $\\tilde{e}_{\\mathrm{inst}}$ are used to integrate with $Q_{\\mathrm{map}}$ , which inject reliable priors for the initialized queries and enhance the overall map construction:", + "bbox": [ + 89, + 667, + 483, + 787 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {e}} _ {\\text {i n s t}} = \\phi_ {e} \\left(\\left\\{\\tilde {\\boldsymbol {e}} _ {\\mathrm {p v}} ^ {i} \\right\\} _ {i = 1} ^ {N}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 794, + 482, + 814 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {\\text {m a p}} ^ {\\prime} = \\operatorname {A t t n} \\left(W _ {q} Q _ {\\text {m a p}}, W _ {k} \\tilde {\\mathbf {e}} _ {\\text {i n s t}}, W _ {v} \\tilde {\\mathbf {e}} _ {\\text {i n s t}}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 834, + 482, + 852 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By injecting PV priors into both BEV and decoder queries, this hybrid mechanism improves the reliability of BEV perception and reduces transition errors. Explicit structural", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/12ff6682f09d2c79c4a074e8d7670d60298783929293de060d1f9cd29885c73e.jpg", + "image_caption": [ + "Figure 4. Design of Hybrid Injection. P2BEV-Attention and P2Q-Attention are designed for PV prompts to integrate with BEV features and map instance queries as structural compensation." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 903, + 266 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "features enhance the generalization capability of the learned model and perform robustly in various driving environments, ensuring better accuracy and consistency.", + "bbox": [ + 511, + 324, + 906, + 371 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Mimic Query Distillation", + "text_level": 1, + "bbox": [ + 511, + 380, + 741, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address the extra computation introduced by the PV branch, we propose the lightweight MQ-Distillation to distill from PV prompt construction, where the mimic query can serve as an alternative to achieve real-time inference.", + "bbox": [ + 511, + 404, + 905, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In MQ-Distillation, we define a set of mimic queries $\\{e_m^i\\}_{i = 1}^N$ , along with an MLP-based query learner $h(\\cdot)$ which is designed to capture and distill the structural features of the PV prompt construction process $\\tilde{e}_{\\mathrm{pv}}^i$ . As shown in Equation 6, during training, the predefined mimic queries $e_{m}^{i}$ and the constructed PV prompts $\\tilde{e}_{\\mathrm{pv}}^i$ are jointly combined by addition, which are sent to the hybrid injection for further operations.", + "bbox": [ + 511, + 465, + 906, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To obtain sufficient structural features and achieve progressive alignment with PV prompts, we design the loss function that guides mimic queries for distillation. We use Mean Squared Error (MSE) as the distillation loss to measure the difference between the constructed PV prompts and the mimicked queries, which can be formulated as follows:", + "bbox": [ + 511, + 585, + 905, + 676 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i s t i l l}} = \\lambda_ {\\text {d i s t i l l}} \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\tilde {\\boldsymbol {e}} _ {\\mathrm {p v}} ^ {i} - h \\left(\\boldsymbol {e} _ {m} ^ {i}\\right)\\right) ^ {2}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 685, + 903, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{distill}}$ is a weight coefficient that balances the impact of the distillation loss on the overall training process. Such loss function guides the mimic queries to accurately replicate the PV prompts generated by the PV branch, making them reliable substitutes for inference.", + "bbox": [ + 511, + 734, + 905, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To this end, only the mimic queries can be utilized during inference, which eliminates the requirement for PV prompt computation. This approach maintains real-time inference capability by efficiently mimicking PV structural features with minimal computational cost, enabling streamlined yet robust map construction.", + "bbox": [ + 511, + 810, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22363", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/87d1c68ff54e10d012ce3dba3df427d174d4e973f536a7c94475c245044bc48b.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Quantitative comparisons on nuScenes validation dataset, utilizing region-based and city-based dataset partitions. Methods with “*” denote that data results are obtained from corresponding papers [27, 50], and others are reproduced by official public codes. “R50” and “SwinT” denotes the image backbone of ResNet50 [9] and Swin-Tiny [30]. “UIGenMap-d” represents the utilization of mimic query from MQ-Distillation. FPS is measured on a single RTX A40 GPU." + ], + "table_body": "
MethodBackboneEpochsRegion-BasedCity-BasedFPS
APpedAPdivAPboumAPAPpedAPdivAPboumAP
VectorMapNet* [ICML'23]R5012015.817.021.218.0-----
MapTR* [ICLR'23]R50246.420.735.520.911.112.921.115.015.8
MGMap [CVPR'24]R50249.427.038.124.810.917.824.717.811.5
MapQR [ECCV'24]R50248.425.739.924.710.215.825.517.212.1
GeMap [ECCV'24]R502411.826.643.627.312.916.326.718.611.6
MapTRv2* [IJCV'24]R502414.728.743.328.918.918.627.921.812.9
StreamMapNet* [WACV'24]R502432.229.340.834.118.716.422.719.313.3
UIGenMap-d (Ours)R502440.330.846.839.3 (+5.2)23.119.625.422.7 (+3.4)12.2
UIGenMap (Ours)R502440.331.847.239.8 (+5.7)24.120.726.123.6 (+4.3)8.2
StreamMapNet [WACV'24]SwinT2432.527.441.733.921.819.725.822.49.3
UIGenMap-d (Ours)SwinT2442.631.947.240.6 (+6.7)24.721.228.824.9 (+2.5)8.5
", + "bbox": [ + 99, + 88, + 903, + 280 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Training and Inference", + "text_level": 1, + "bbox": [ + 89, + 348, + 305, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During the training process, combined with $\\mathcal{L}_{\\mathrm{nl}}$ for uncertainty output and $\\mathcal{L}_{\\mathrm{distill}}$ for mimic query distillation, the overall loss setting $\\mathcal{L}_{\\mathrm{map}}$ can be represented as:", + "bbox": [ + 89, + 369, + 482, + 417 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a p}} = \\lambda_ {1} \\mathcal {L} _ {\\text {p t s}} + \\lambda_ {2} \\mathcal {L} _ {\\text {c l s}} + \\mathcal {L} _ {\\text {n l l}} + \\mathcal {L} _ {\\text {d i s t i l l}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 425, + 482, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{pts}}$ denotes the point-to-point Manhattan distance between the regressed points and the ground-truth. $L_{\\mathrm{cls}}$ represents the focal loss for map classification. $\\lambda_{1}$ and $\\lambda_{2}$ are the corresponding loss weights for point regression and map classification, respectively.", + "bbox": [ + 89, + 449, + 482, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For inference, the well-learned mimic query can only be employed as a substitute for PV prompts, avoiding the computational burden for the PV detection branch. Thus, our approach ensures real-time inference without introducing unnecessary computational overhead, while preserving effective performance. Additionally, the model's uncertainty learning enables dynamic sampling during inference, enhancing the adaptability to diverse driving environments.", + "bbox": [ + 89, + 523, + 482, + 647 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 657, + 223, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 89, + 681, + 289, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We conduct extensive experiments on widely used autonomous driving datasets, including nuScenes [1] and Argoverse2 [46]. nuScenes dataset comprises 1,000 driving scenes from Boston and Singapore. Each scene lasts approximately 20 seconds with a $2\\mathrm{Hz}$ sampling rate. For each sample, there are six surrounding images and 32-beam LiDAR point clouds. Argoverse2 contains 1,000 scenes from 6 cities, with each scene capturing 15 seconds of data at $20\\mathrm{Hz}$ from 7 cameras and $10\\mathrm{Hz}$ LiDAR sweeps. In our experiments, the sampling frequency for Argoverse2 is unified into $2\\mathrm{Hz}$ as in StreamMapNet [50].", + "bbox": [ + 89, + 703, + 482, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Geo-based Data Partitions. Considering the overlap of widely used training and validation partitions, we adopt", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "geographically disjoint (geo-based) dataset partitions as benchmarks to evaluate the generalization capability of the presented model. Experiments are conducted on both region-based and city-based data partitions. The region-based partition is adopted from [50], where the driving locations geographically disjoin the training and validation sets. In the city-based data partition, we apply city-wise splits in [27], where the training and validation data are from different cities with a larger distribution change. On nuScenes, Boston and Onenorth are used for training, while Queenstown and Holland Village are used for validation. For Ar-goverse2, Miami and Pittsburgh are used in training, and the remains are employed for validation. For different dataset partitions, models are separately trained on that split's training set and evaluated on their corresponding validation set.", + "bbox": [ + 511, + 349, + 906, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. In line with the existing works [26, 50], we conduct experiments on three static map elements, namely lane dividers, pedestrian crossings, and road boundaries. Perception range is set to $[-15.0m, 15.0m]$ for X-axis and $[-30.0m, 30.0m]$ for Y-axis. We use average precision as the evaluation metric, which calculates the Chamfer distance between predictions and ground truths. The mean Average Precision (mAP) is obtained by averaging AP across the distinct thresholds: $\\{0.5m, 1m, 1.5m\\}$ .", + "bbox": [ + 511, + 577, + 908, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 726, + 732, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For fair comparisons, we employ ResNet50 [9] as the backbone. As in [50], the size of BEV features is set to $50 \\times 100$ . The number of map queries and query dimensions are set to 100 and 512. We use AdamW optimizer with a learning rate of $4e^{-4}$ . Training epochs are limited to prevent potential overfitting. All models are trained on 4 NVIDIA Tesla V100 GPUs with a batch size of 4. The weighted factors $\\lambda_{1}$ and $\\lambda_{2}$ are set to 50.0 and 5.0, separately. $\\lambda_{\\mathrm{nil}}$ and $\\lambda_{\\mathrm{distill}}$ are set to 0.05 and 10.0. We adopt a pre-trained PV detection branch for our UIGenMap training process. More details", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22364", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f0a0b16b1b793824fb076adc41a1021d42430872f8c4563aecf4930ed6e1f3cf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneEpochsRegion-BasedCity-BasedFPS
\\(AP_{ped}\\)\\(AP_{div}\\)\\(AP_{bou}\\)mAP\\(AP_{ped}\\)\\(AP_{div}\\)\\(AP_{bou}\\)mAP
\\(VectorMapNet^{*}\\)[ICML'23]R5012035.634.937.836.1-----
\\(MapTR^{*}\\)[ICLR'23]R503048.150.455.051.134.741.734.737.915.8
\\(GeMap\\)[ECCV'24]R503052.559.258.256.634.541.140.438.711.6
\\(MapTRv2^{*}\\)[JCV'24]R503053.358.758.856.937.442.241.940.512.9
\\(MapQR\\)[ECCV'24]R503054.258.060.457.533.841.541.638.912.1
\\(StreamMapNet^{*}\\)[WACV'24]R503057.955.761.358.340.243.443.342.313.3
\\(UIGenMap-d\\)(Ours)R503058.358.462.459.7 (+1.4)42.944.444.744.0 (+1.7)12.2
\\(UIGenMap\\)(Ours)R503059.858.762.760.4 (+2.1)43.745.745.144.8 (+2.5)8.2
", + "bbox": [ + 99, + 88, + 903, + 234 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/72381be7d513c296c77390d45546205b8e92454ee30e29826efe313a14baa39d.jpg", + "table_caption": [ + "Table 2. Quantitative comparisons on Argoverse2 validation datasets, utilizing region-based and city-based dataset partitions. Methods with “*” denote that data results are obtained from corresponding papers [27, 50], and others are reproduced by official public codes. Models are trained for 30 epochs and follow the same setting. FPS is measured on a single RTX A40 GPU." + ], + "table_footnote": [], + "table_body": "
DatasetMethodEpochs\\(AP_{ped}\\)\\(AP_{div}\\)\\(AP_{bou}\\)mAP
nuScenesMapTRV2 [26]2459.862.462.461.5
StreamMapNet [50]2460.461.958.960.4
SQD-MapNet [43]2463.065.563.363.9
UIGenMap(Ours)2464.265.464.964.8
Argoverse2StreamMapNet [50]3062.959.563.061.5
SQD-MapNet [43]3064.960.264.963.3
UIGenMap(Ours)3064.560.866.163.8
", + "bbox": [ + 93, + 287, + 485, + 404 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/85edb8fc76eefbefc43b11e48a4c76780b057157945aa7b952438728cad5edc4.jpg", + "table_caption": [ + "Table 3. Performance on nuScenes and Argoverse2 datasets with the original dataset partition at ${30}\\mathrm{\\;m}$ range." + ], + "table_footnote": [], + "table_body": "
bkbUI2DPromptUA-DecoderAPpedAPdivAPboumAP
32.229.340.834.1
✓39.928.643.837.4
✓36.131.344.037.1
✓✓41.629.744.838.7
✓✓41.429.847.439.5
✓✓39.230.146.238.5
✓✓✓40.331.847.239.8
", + "bbox": [ + 93, + 441, + 478, + 553 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablations on the impact of component designs, 'bkb' means the shared image backbone from PV branches.", + "bbox": [ + 89, + 556, + 482, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "are provided in the supplementary materials.", + "bbox": [ + 89, + 595, + 387, + 611 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Main Results", + "text_level": 1, + "bbox": [ + 89, + 619, + 230, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on nuScenes. We compare the generalization capability with previous methods [25, 26, 28, 31, 50, 54] on geospatial data partitions, in which the image size is set to $480 \\times 800$ . As shown in Table 1, our UIGenMap outperforms previous methods at a large margin. For the ResNet50 backbone, UIGenMap achieves $5.7\\mathrm{mAP}$ improvement on the region-based partition, and $4.3\\mathrm{mAP}$ gain is obtained on the city-based splits. For more efficient design, UIGenMapd also archives $39.3\\mathrm{mAP}$ with a faster inference speed. Experiments on the Swin-T [30] backbone show significant performance with $40.6\\mathrm{mAP}$ achievement. Furthermore, we show some visual examples in Figure 5, indicating more intuitive achievements for UIGenMap.", + "bbox": [ + 88, + 643, + 482, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on Argoverse2. More experiments are conducted on Argoverse2 dataset. The image size is $608 \\times 608$ , and the sampling rate is $2\\mathrm{Hz}$ . For a fair comparison, the models are trained for 30 epochs. As shown in Table 2, UIGen", + "bbox": [ + 88, + 839, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Map achieves the best results of $60.4\\mathrm{mAP}$ on region-based data partitions and $43.4\\mathrm{mAP}$ on city-based splits, which performs better than previous approaches. Upon further investigation, calibration errors and dataset distribution also affect model performance. More visualization can be found in our supplementary materials.", + "bbox": [ + 511, + 291, + 906, + 381 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on Original Data Partitions. For thorough evaluation, we conducted experiments on the original data partitions with similar driving scenarios, which are widely used in previous studies. As shown in Table 3, we compare our methods with others utilizing the same baseline. UIGenMap still outperforms the existing approaches, which outperforms StreamMapNet with $+4.4$ mAP and $+2.3$ mAP separately, validating the effectiveness of our approach.", + "bbox": [ + 511, + 382, + 908, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 513, + 679, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, ablations are conducted to verify the effectiveness of our designs. We perform experiments on nuScenes with region-based partition using ResNet50 backbone. All models are trained for 24 epochs. More ablations are provided in the supplementary materials.", + "bbox": [ + 511, + 537, + 906, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Contributions of Main Components. Table 4 demonstrates the impact of each component in our design. Based on the pre-trained PV backbone, we design the UA-Decoder for more reliable uncertainty output and UI2DPrompt to dynamically inject PV structural information, which achieves $+2.7$ mAP and $+2.4$ mAP improvement, respectively. The combination of these designs gains $+4.0$ mAP total. The ablations are conducted on UA-Decoder at BEV space. Besides, the pre-trained PV backbone further emphasizes the effectiveness to $39.8$ mAP for better scene understanding.", + "bbox": [ + 511, + 612, + 905, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Design of UA-Decoder. We design the UA-Attention for probabilistic attention and the UA-Head to construct the uncertainty output from the instance and finer point levels. Table 5 shows the ablation without other modules. Our uncertainty-aware design supports dynamic adaptation to achieve more robust performance, which achieves $+3.0$ mAP improvements based on systematic combination.", + "bbox": [ + 511, + 763, + 905, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Design for UI2DPrompt. UI2DPrompt consists of prompt construction and a hybrid injection module. Table 6 shows", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22365", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/55491a0f775a128dc58c15508ee62993ff71f113d1dc12b8ded035fe71d74f4b.jpg", + "image_caption": [ + "Surrounding Views with PV Detection" + ], + "image_footnote": [], + "bbox": [ + 174, + 89, + 566, + 339 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1353fbd0974062a11dcd003a74e21b249732bba0d5f9e1fe3839ca7360f4af46.jpg", + "image_caption": [ + "Baseline", + "GT", + "Figure 5. Qualitative results on the nuScenes dataset with the region-based data partition. Uncertainty outputs are represented as circles." + ], + "image_footnote": [], + "bbox": [ + 566, + 90, + 818, + 339 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/35bba4c96e67f47fb67ecfa2bc4d8ed42d54b67aea630cfa5b9e6a0492cfc17f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
UA-Att.UA-HeadAPpedAPdivAPboumAP
32.229.340.834.1
✓31.730.442.935.0
✓34.830.543.436.3
✓✓36.131.344.037.1
", + "bbox": [ + 135, + 375, + 439, + 450 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/bb1e8c656fed9e0e7bdbc4fe46d5379f5ba84c597befe8606519906236d2a9a1.jpg", + "table_caption": [ + "Table 5. Ablations on UA-Decoder without other modules. UAAtt and UA-Head represent the utilization of uncertainty-aware attention and output head respectively." + ], + "table_footnote": [], + "table_body": "
Coords. Trans.Unc. UtilizationmAP
//37.6
IPMembedding only39.3
IPMweight & embedding39.8
3D Trans.weight & embedding38.7
", + "bbox": [ + 138, + 503, + 437, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the strategy selection for constructing PV prompts. Compared to a depth-based 3D transition, IPM is more suitable for coordinating the transformation based on the assumption of a flattened road surface. When uncertainty is applied both as dynamic weighting and as learning-based embeddings, performance is further enhanced by reliable PV prompts, achieving an improvement of $+2.2\\mathrm{mAP}$ .", + "bbox": [ + 89, + 640, + 482, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7 demonstrates the impact of the hybrid injection design, with a total performance of $39.8\\mathrm{mAP}$ for combining prompt-to-BEV and prompt-to-query attention. This result confirms the optimal utilization of reliable structural information, effectively compensating for information loss and errors introduced in the basic learning process.", + "bbox": [ + 89, + 747, + 482, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablations on MQ-Distillation. To verify the effectiveness of the distillation module, we conduct ablations on the hyperparameter of distillation loss weight $\\lambda_{\\mathrm{distill}}$ , as this affects the imitation learning of PV structural features. As", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1371390fca46ceb8263cb3c1796ecbd8703ef199dec52ee3fe4ee5a78ccf21f5.jpg", + "table_caption": [ + "Table 6. Ablations on the construction of PV prompts. \"3D Trans.\" means the depth-based transformation. The first row with $\\frac{1}{r}$ indicates direct PV coordinate embedding." + ], + "table_footnote": [], + "table_body": "
P2BEVP2QAPpedAPdivAPboumAP
39.330.445.538.4
✓39.931.146.039.2
✓39.630.746.138.8
✓✓40.331.847.239.8
", + "bbox": [ + 553, + 375, + 867, + 450 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/816f98bf545daf2f9b2a456fabc8f0d59ee24b843fded633758a719c6942bf8c.jpg", + "table_caption": [ + "Table 7. Ablations on hybrid injection. P2BEV and P2Q mean the cross-attention for $\\mathrm{F}_{\\mathrm{bev}}$ and $Q_{\\mathrm{map}}$ separately." + ], + "table_footnote": [], + "table_body": "
λdistill571015/
mAP37.538.239.338.839.8
", + "bbox": [ + 563, + 488, + 856, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Ablations on $\\lambda_{\\mathrm{distill}}$ for MQ-Distillation Design. The last column refers to the use of the best teacher of PV prompts.", + "bbox": [ + 511, + 532, + 905, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "illustrated in Table 8, the best performance is achieved by setting the weight to 10.0, which refers to the best learning process of the mimic query for explicit structural features.", + "bbox": [ + 511, + 570, + 905, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 511, + 632, + 782, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose UIGenMap, an uncertainty-instructed structure injection approach to achieve a more generalized HD map construction. Specifically, we introduce the PV detection branch to obtain explicit structural features, in which the UA-Decoder produces adaptive probabilistic outputs. The UI2DPrompt module is proposed to construct reliable PV prompts and achieve dynamic integration through the hybrid injection designed. For real-time inference, MQ-Distillation is designed to mimic the structural prompts of PV. Comprehensive experiments on the challenging geo-based partitions of nuScenes and Argoverse2 datasets demonstrate the generalization of our approach.", + "bbox": [ + 511, + 657, + 906, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For limitations, current methods are designed and evaluated primarily on specific datasets. In the future, developing larger and more diverse datasets and benchmarks across various locations remains a crucial area for exploration.", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22366", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by National Natural Science Foundation of China under Grant No.62376244. It is also supported by Information Technology Center and State Key Lab of CAD&CG, Zhejiang University.", + "bbox": [ + 89, + 114, + 483, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 186, + 187, + 203 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11621-11631, 2020. 1, 2, 6", + "[2] Anh-Quan Cao, Angela Dai, and Raoul de Charette. Pasco: Urban 3d panoptic scene completion with uncertainty awareness. In CVPR, pages 14554–14564, 2024. 3", + "[3] Jie Chang, Zhonghao Lan, Changmao Cheng, and Yichen Wei. Data uncertainty learning in face recognition. In CVPR, pages 5710-5719, 2020. 3", + "[4] Jiacheng Chen, Yuefan Wu, Jiaqi Tan, Hang Ma, and Yasutaka Furukawa. Maptracker: Tracking with strided memory fusion for consistent vector hd mapping. In ECCV, pages 90-107. Springer, 2025. 1, 2", + "[5] Sehwan Choi, Jungho Kim, Hongjae Shin, and Jun Won Choi. Mask2map: Vectorized hd map construction using bird's eye view segmentation masks. In ECCV, 2024. 2", + "[6] Wenjie Ding, Limeng Qiao, Xi Qiu, and Chi Zhang. Pivotnet: Vectorized pivot learning for end-to-end hd map construction. In ICCV, pages 3672-3682, 2023. 2", + "[7] Xunjiang Gu, Guanyu Song, Igor Gilitschenski, Marco Pavone, and Boris Ivanovic. Producing and leveraging online map uncertainty in trajectory prediction. In CVPR, pages 14521-14530, 2024. 3, 4", + "[8] Hongji Guo, Hanjing Wang, and Qiang Ji. Uncertainty-guided probabilistic transformer for complex action recognition. In CVPR, pages 20052-20061, 2022. 3, 4", + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6", + "[10] Yuzhe He, Shuang Liang, Xiaofei Rui, Chengying Cai, and Guowei Wan. Egovm: Achieving precise eg-localization using lightweight vectorized maps. arXiv preprint arXiv:2307.08991, 2023. 1", + "[11] Haotian Hu, Fanyi Wang, Yaonong Wang, Laifeng Hu, Jingwei Xu, and Zhiwang Zhang. Admap: Anti-disturbance framework for reconstructing online vectorized hd map. arXiv preprint arXiv:2401.13172, 2024. 2", + "[12] Haoxuanye Ji, Pengpeng Liang, and Erkang Cheng. Enhancing 3d object detection with 2d detection-guided query anchors. In CVPR, pages 21178-21187, 2024. 2", + "[13] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, pages 8340-8350, 2023. 1", + "[14] Xiaohui Jiang, Shuai Lin Li, Yingfei Liu, Shihao Wang, Fan Jia, Tiancai Wang, Lijin Han, and Xiangyu Zhang. Far3d:" + ], + "bbox": [ + 93, + 212, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Expanding the horizon for surround-view 3d object detection. In AAAI, pages 2561-2569, 2024. 2", + "[15] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? In NeuIPS, 2017. 3", + "[16] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. In NeuIPS, 2017. 3", + "[17] Han Li, Zehao Huang, Zitian Wang, Wenge Rong, Naiyan Wang, and Si Liu. Enhancing 3d lane detection and topology reasoning with 2d lane priors. arXiv preprint arXiv:2406.03105, 2024. 2", + "[18] Qi Li, Yue Wang, Yilun Wang, and Hang Zhao. Hdmapnet: An online hd map construction and evaluation framework. In ICRA, pages 4628-4634. IEEE, 2022. 2", + "[19] Siyu Li, Jiacheng Lin, Hao Shi, Jiaming Zhang, Song Wang, You Yao, Zhiyong Li, and Kailun Yang. Dtlmapper: Dual temporal consistent learning for vectorized hd map construction. arXiv preprint arXiv:2405.05518, 2024. 2", + "[20] Siyu Li, Kailun Yang, Hao Shi, Song Wang, You Yao, and Zhiyong Li. Genmapping: Unleashing the potential of inverse perspective mapping for robust online hd map construction. arXiv preprint arXiv:2409.08688, 2024. 2", + "[21] Xiaotong Li, Yongxing Dai, Yixiao Ge, Jun Liu, Ying Shan, and Lingyu Duan. Uncertainty modeling for out-of-distribution generalization. In ICLR, 2022. 3", + "[22] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In ECCV, pages 1-18. Springer, 2022. 2, 3", + "[23] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In ICCV, pages 14864-14873, 2024. 1", + "[24] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. In NeuIPS, pages 10421-10434, 2022. 2", + "[25] Benchcheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. Maptr: Structured modeling and learning for online vectorized hd map construction. In ICLR, 2022. 1, 2, 7", + "[26] Benchcheng Liao, Shaoyu Chen, Yunchi Zhang, Bo Jiang, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Maptrv2: An end-to-end framework for online vectorized hd map construction. IJCV, pages 1-23, 2024. 1, 2, 6, 7", + "[27] Adam Lilja, Junsheng Fu, Erik Stenberg, and Lars Hammarstrand. Localization is all you evaluate: Data leakage in online mapping datasets and how to fix it. In CVPR, pages 22150-22159, 2024. 1, 6, 7", + "[28] Xiaolu Liu, Song Wang, Wentong Li, Ruizi Yang, Junbo Chen, and Jianke Zhu. Mgmt: Mask-guided learning for online vectorized hd map construction. In CVPR, pages 14812-14821, 2024. 1, 2, 7", + "[29] Yicheng Liu, Tianyuan Yuan, Yue Wang, Yilun Wang, and Hang Zhao. Vectormapnet: End-to-end vectorized hd map learning. In ICML, pages 22352-22369. PMLR, 2023. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22367", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 6, 7", + "[31] Zihao Liu, Xiaoyu Zhang, Guangwei Liu, Ji Zhao, and Ningyi Xu. Leveraging enhanced queries of point sets for vectorized map construction. arXiv preprint arXiv:2402.17430, 2024. 1, 2, 7", + "[32] Hao Lu, Yunpeng Zhang, Qing Lian, Dalong Du, and Yingcong Chen. Towards generalizable multi-camera 3d object detection via perspective debiasing. arXiv preprint arXiv:2310.11346, 2023. 2", + "[33] Ao Luo, Fan Yang, Xin Li, Lang Nie, Chunyu Lin, Haoqiang Fan, and Shuaicheng Liu. Gaflow: Incorporating gaussian attention into optical flow. In ICCV, pages 9642–9651, 2023. 3", + "[34] Jiahuan Pei, Cheng Wang, and György Szarvas. Transformer uncertainty estimation with hierarchical stochastic attention. In AAAI, pages 11147-11155, 2022. 3, 4", + "[35] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, pages 194-210. Springer, 2020. 2", + "[36] Fengchun Qiao and Xi Peng. Uncertainty-guided model generalization to unseen domains. In CVPR, pages 6790-6800, 2021. 3", + "[37] Limeng Qiao, Wenjie Ding, Xi Qiu, and Chi Zhang. End-to-end vectorized hd-map construction with piecewiseBezier curve. In CVPR, pages 13218-13228, 2023. 2", + "[38] Tixiao Shan and Brendan Englot. Lego-loam: Lightweight and ground-optimized lidar odometry and mapping on variable terrain. In IROS, pages 4758–4765. IEEE, 2018. 2", + "[39] Tixiao Shan, Brendan Englot, Drew Meyers, Wei Wang, Carlo Ratti, and Daniela Rus. Lio-sam: Tightly-coupled lidar inertial odometry via smoothing and mapping. In IROS, pages 5135-5142. IEEE, 2020. 2", + "[40] Jingyu Song, Xudong Chen, Liupei Lu, Jie Li, and Katherine A Skinner. Memfusionmap: Working memory fusion for online vectorized hd map construction. arXiv preprint arXiv:2409.18737, 2024. 2", + "[41] A Vaswani. Attention is all you need. In NeuIPS, 2017. 2", + "[42] Song Wang, Wentong Li, Wenyu Liu, Xiaolu Liu, and Jianke Zhu. Lidar2map: In defense of lidar-based semantic map construction using online camera distillation. In CVPR, pages 5186-5195, 2023. 2", + "[43] Shuo Wang, Fan Jia, Yingfei Liu, Yucheng Zhao, Zehui Chen, Tiancai Wang, Chi Zhang, Xiangyu Zhang, and Feng Zhao. Stream query denoising for vectorized hd map construction. In ECCV, 2024. 2, 7", + "[44] Song Wang, Zhongdao Wang, Jiawei Yu, Wentong Li, Bailan Feng, Junbo Chen, and Jianke Zhu. Relioco: Towards reliable semantic occupancy prediction via uncertainty learning. arXiv preprint arXiv:2409.18026, 2024. 3", + "[45] Song Wang, Jiawei Yu, Wentong Li, Wenyu Liu, Xiaolu Liu, Junbo Chen, and Jianke Zhu. Not all voxels are equal: Hardness-aware semantic scene completion with self-distillation. In CVPR, pages 14792-14801, 2024. 1" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In NeurIPS, 2021. 1, 2, 6", + "[47] Qiangqiang Wu, Yan Xia, Jia Wan, and Antoni B Chan. Boosting 3d single object tracking with 2d matching distillation and 3d pre-training. In ECCV, 2024. 2", + "[48] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In CVPR, pages 17830-17839, 2023. 2", + "[49] Jing Yang, Minyue Jiang, Sen Yang, Xiao Tan, Yingying Li, Errui Ding, Hanli Wang, and Jingdong Wang. Mgmtapnet: Multi-granularity representation learning for end-to-end vectorized hd map construction. arXiv preprint arXiv:2410.07733, 2024. 2", + "[50] Tianyuan Yuan, Yicheng Liu, Yue Wang, Yilun Wang, and Hang Zhao. Streammapnet: Streaming mapping network for vectorized online hd map construction. In WACV, pages 7356-7365, 2024. 1, 2, 4, 6, 7", + "[51] Gongjie Zhang, Jiahao Lin, Shuang Wu, Zhipeng Luo, Yang Xue, Shijian Lu, Zuoguan Wang, et al. Online map vectorization for autonomous driving: A rasterization perspective. In NeuIPS, 2024. 2", + "[52] Xiaoyu Zhang, Guangwei Liu, Zihao Liu, Ningyi Xu, Yunhui Liu, and Ji Zhao. Enhancing vectorized map perception with historical rasterized maps. In ECCV, 2024. 2", + "[53] Yunpeng Zhang, Wenzhao Zheng, Zheng Zhu, Guan Huang, Jiwen Lu, and Jie Zhou. A simple baseline for multi-camera 3d object detection. In AAAI, pages 3507-3515, 2023. 2", + "[54] Zhixin Zhang, Yiyuan Zhang, Xiaohan Ding, Fusheng Jin, and Xiangyu Yue. Online vectorized hd map construction using geometry. arXiv preprint arXiv:2312.03341, 2023. 2, 7", + "[55] Lijun Zhou, Tao Tang, Pengkun Hao, Zihang He, Kalok Ho, Shuo Gu, Wenbo Hou, Zhihui Hao, Haiyang Sun, Kun Zhan, et al. Ua-track: Uncertainty-aware end-to-end 3d multi-object tracking. arXiv preprint arXiv:2406.02147, 2024. 3", + "[56] Yi Zhou, Hui Zhang, Jiaqian Yu, Yifan Yang, Sangil Jung, Seung-In Park, and ByungIn Yoo. Himap: Hybrid representation learning for end-to-end vectorized hd map construction. In CVPR, pages 15396-15406, 2024. 1", + "[57] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 1, 2, 3", + "[58] Yanjin Zhu, Xin Zheng, and Jianke Zhu. Mesh-loam: Real-time mesh-based lidar odometry and mapping. TIV, 2024. 2" + ], + "bbox": [ + 516, + 90, + 903, + 825 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22368", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction/308073a9-0b23-46de-a5fd-816ea2a49dc7_model.json b/2025/Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction/308073a9-0b23-46de-a5fd-816ea2a49dc7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7f2bf4959909cf283bcdf7cfb496cec284de858f --- /dev/null +++ b/2025/Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction/308073a9-0b23-46de-a5fd-816ea2a49dc7_model.json @@ -0,0 +1,2398 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.194, + 0.131, + 0.806, + 0.176 + ], + "angle": 0, + "content": "Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction" + }, + { + "type": "text", + "bbox": [ + 0.169, + 0.203, + 0.834, + 0.24 + ], + "angle": 0, + "content": "Xiaolu Liu\\(^{1*}\\), Ruizi Yang\\(^{1*}\\), Song Wang\\(^{1}\\), Wentong Li\\(^{2}\\), Junbo Chen\\(^{3\\dagger}\\), Jianke Zhu\\(^{1\\dagger}\\) \n\\(^{1}\\)Zhejiang University \\(^{2}\\)NUAA \\(^{3}\\)Udeer.ai" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.242, + 0.878, + 0.258 + ], + "angle": 0, + "content": "{xiaoluliu, ruiziyang, songw, jkzhu}@zju.edu.cn, wentong.li@nuaa.edu.cn, junbo@udeer.ai" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.324, + 0.486, + 0.718 + ], + "angle": 0, + "content": "Reliable high-definition (HD) map construction is crucial for the driving safety of autonomous vehicles. Although recent studies demonstrate improved performance, their generalization capability across unfamiliar driving scenes remains unexplored. To tackle this issue, we propose UIGenMap, an uncertainty-instructed structure injection approach for generalizable HD map vectorization, which concerns the uncertainty resampling in statistical distribution and employs explicit instance features to reduce excessive reliance on training data. Specifically, we introduce the perspective-view (PV) detection branch to obtain explicit structural features, in which the uncertainty-aware decoder is designed to dynamically sample probability distributions considering the difference in scenes. With probabilistic embedding and selection, UI2DPrompt is proposed to construct PV-learnable prompts. These PV prompts are integrated into the map decoder by designed hybrid injection to compensate for neglected instance structures. To ensure real-time inference, a lightweight Mimic Query Distillation is designed to learn from PV prompts, which can serve as an efficient alternative to the flow of PV branches. Extensive experiments on challenging geographically disjoint (geo-based) data splits demonstrate that our UIGenMap achieves superior performance, with \\(+5.7\\) mAP improvement on the nuScenes dataset. Source code is available at https://github.com/xiaolul2/UIGenMap." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.747, + 0.222, + 0.763 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.773, + 0.484, + 0.864 + ], + "angle": 0, + "content": "Reliable HD map construction is crucial for driving safety of autonomous vehicles, which provides essential geometric information for ego-localization, perceptions, path planning, and predictions [10, 13, 23, 45]. Recently, online high-definition (HD) map vectorization has attracted considerable attention for its cost-effectiveness and ability to" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.291, + 0.696, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.291, + 0.906, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.444, + 0.907, + 0.515 + ], + "angle": 0, + "content": "Figure 1. (a) Comparisons on original and different geo-based data partitions among previous methods and ours. (b) Visual comparisons with circles on BEV map to represent the learned uncertainties, in which PV instances serve as compensation and provide a detailed understanding of map perception with generalization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.907, + 0.608 + ], + "angle": 0, + "content": "provide real-time updates [25, 31, 56]. Given images captured by onboard cameras, current map construction relies on transformer scheme [57] to convert perspective-view (PV) image features into bird's-eye-view (BEV) space. Then, a decoder follows to predict the map elements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.909, + 0.883 + ], + "angle": 0, + "content": "Despite the achievements on widely adopted benchmarks [4, 26, 28], little attention has been paid to investigating the generalization capabilities, which refer to the model's performance on unfamiliar driving scenarios that are different from the training sets. As discussed in [27, 50], in the widely used public datasets [1, 46], there are a large number of geographical overlaps between the training and validation scenes. The models are then more likely to memorize similar driving scenes rather than fully capture their underlying road structures. Variations in driving scenes, such as the contrast between bustling city centers and suburbs with simple road structures, highlight significant shifts in feature distributions and reveal the model's limitations in robustness. Besides, learning-based BEV transformation will inevitably lead to geometric errors and loss of texture details. As shown in Figure 1(a), previous methods [25, 50] exhibit notable performance degradation on region-based and city-based data splits under changing driving scenes." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "To tackle the above issue, we propose the innovative em" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.245, + 0.888 + ], + "angle": 0, + "content": "*Equal Contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.24, + 0.9 + ], + "angle": 0, + "content": "† Corresponding authors" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "22359" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.363 + ], + "angle": 0, + "content": "employment of model uncertainty targeted for more generalizable HD map constructions. On the one hand, the uncertainty estimation process involves learning both the statistical mean and variance value, enabling dynamic resampling based on probabilistic distributions. Given the discrepancy in feature distribution across different driving scenes, such a statistical resampling strategy enhances the model's dynamic adaptability in unfamiliar environments. However, the estimated model uncertainty quantifies the confidence in its predictions, which can serve as a reliable indicator for further feature selection and enhancement. As illustrated in Figure 1(b), direct 2D perception can capture more intuitive semantic and angular structural information, which might be missed or incorrectly learned during the implicit PV-to-BEV conversions. Thus, combined with uncertainty-based feature selection, more reliable explicit information can be emphasized as compensation to boost the robustness of BEV-level map construction." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.367, + 0.483, + 0.699 + ], + "angle": 0, + "content": "Based on the above discussion, in this work, we propose UIGenMap, an uncertainty-instructed PV structure injection strategy for more generalizable HD map vectorization. Specifically, a PV detection branch is introduced to provide explicit structural instances. In both PV and BEV branches, we move beyond the vanilla transformer structure by designing the statistically sensitive uncertainty-aware decoder (UA-Decoder), which constructs the probabilistic outputs by uncertainty-aware attention and finer point-level uncertainty head. Leveraging the PV uncertainty output, we introduce the UI2DPrompt to construct learnable PV prompts. These PV prompts are integrated into the BEV decoder by the designed hybrid injection mechanism, which acts as structural compensation for the BEV map prediction. Consequently, reliable structural representation from PV space can be injected to enhance the model's capability, thereby improving its perceptual ability in complex environments. Considering the demands for real-time inference, we also design a lightweight mimic query distillation module to learn from constructed PV prompts. Then during inference, the mimicked query can replace the PV detection flow to reduce extra computational burdens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.702, + 0.482, + 0.837 + ], + "angle": 0, + "content": "To prioritize the evaluation of generalization capability, we assess our method using geographically disjoint (geo-based) data partitions on the nuScenes [1] and Argoverse2 [46] datasets. Extensive experiments demonstrate that our UIGenMap achieves state-of-the-art performance with higher adaptive ability, especially with \\(+5.7\\) mAP improvement on the region-based partition and \\(+4.3\\) mAP on the more challenging city-based split on nuScenes. The main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.84, + 0.483, + 0.884 + ], + "angle": 0, + "content": "- We propose UIGenMap, an uncertainty-instructed approach to more generalizable HD map vectorization with explicit structure injection." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- The UA-Decoder and UI2DPrompt are proposed to con" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "struct structural PV prompts, which are dynamically integrated with BEV perception by hybrid injection." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.122, + 0.905, + 0.166 + ], + "angle": 0, + "content": "- Lightweight MQ-Distillation design supports real-time inference without the additional computational overhead of the PV detection branches." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.168, + 0.905, + 0.213 + ], + "angle": 0, + "content": "- Promising results on more challenging geo-based data partitions show that UIGenMap outperforms previous approaches and achieves powerful generalization capability." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.122, + 0.905, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.226, + 0.655, + 0.241 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.252, + 0.907, + 0.599 + ], + "angle": 0, + "content": "Vectorized HD Map Construction. Conventional HD map construction approaches [38, 39, 58] rely on SLAM-based annotations, which require tedious offline manual labeling at high costs. Recent studies [6, 11, 19, 20, 29, 37, 49] have explored online approaches with the input of onboard sensor data. In [18, 22, 24, 35, 42], HDMap construction is treated as a segmentation task, relying on post-processing to achieve map vectorization. Later, Liao et al. [25, 26] represent map elements as ordered points and employ transformers [41, 57] after the BEV encoder for auto-regressive detection. Other modeling methods, such as Bezier curves and pivot points, are explored in BeMapNet [37] and PivotNet [6] for geometric representation. Several works [4, 19, 40, 43, 50] are proposed for different temporal fusion approaches. GeMap [54] and MapQR [31] improve mapping performance from the perspectives of geometric relationships and decoder design, respectively. In [5, 28, 51, 52], rasterized maps are utilized for structure compensation. Despite improved performance, these approaches seldom consider the generalization capability of models in unfamiliar driving scenarios. In this work, we incorporate uncertainty-instructed PV detection retracing, aiming for better performance across diverse scenarios." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.6, + 0.906, + 0.855 + ], + "angle": 0, + "content": "PV Detection for 3D Perception. Instances from PV detection can be effective compensations for 3D perceptions [12, 14, 32, 47, 48, 53]. For 3D object detection, BEVFormerv2 [48] utilizes a two-stage BEV detector that integrates the query from the PV head into the BEV space. In SimMoD [53], 2D detection is utilized to construct proposals for 3D fine-grained perception. Far3D [14] combines 2D detection boxes with estimated depth information to improve long-range perception performance. For model generalization, Lu et al. [32] propose a generalizable framework based on perspective debiasing, in which PV consistency learning is employed among different domains. Applying for 3D lanes, Li et al. [17] use 2D priors for lane detection and topology reasoning. Inspired by the above methods, we employ PV structural features for more generalizable HD map vectorization, in which distillations from 2D instances are utilized to ensure inference efficiency." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Uncertainty in Model. Generally, uncertainty refers to the model's confidence level for its predictions, which can be a statistical representation to indicate the model's under" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22360" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.101, + 0.186, + 0.114 + ], + "angle": 0, + "content": "(a) Overview" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.108, + 0.391, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.403, + 0.096, + 0.49, + 0.108 + ], + "angle": 0, + "content": "(b) UA-Decoder" + }, + { + "type": "image", + "bbox": [ + 0.402, + 0.103, + 0.894, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.407, + 0.23, + 0.491, + 0.242 + ], + "angle": 0, + "content": "(c) UI2DPrompt" + }, + { + "type": "image", + "bbox": [ + 0.405, + 0.238, + 0.888, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.388, + 0.907, + 0.445 + ], + "angle": 0, + "content": "Figure 2. (a) Overview of Our UIGenMap. For training, the PV branch is introduced with the uncertainty-instructed structural injection, in which the MQ-Distillation is designed to mimic PV structural features. (b) Uncertainty-Aware Decoder Architecture. For each layer, UA-Decoder comprises probabilistic UA-Attention and UA-Head for reliable output. (c) UI2DPrompt Design. We construct PV prompts from PV-detected elements and their corresponding uncertainties, which are integrated with the main branch by hybrid injection." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.462, + 0.486, + 0.705 + ], + "angle": 0, + "content": "standing of scenes [3, 15, 44]. Uncertainty has gained intensive attention due to the potential for reliability evaluation and performance enhancement [2, 16]. Specifically, the generalization problem with uncertainty modeling is investigated in [21, 36], which expects uncertainty estimation for out-of-distribution generalization. For the transformer structure [8, 34], uncertainty has been explored to guide the training and inference processes for more robust model performance. In the field of autonomous driving, significant uncertainty has been introduced for 3D occupancy [2, 44], optical flow [33], and object tracking [55]. Recently, Gu et al. [7] present uncertainty estimation for the vectorized map, which is targeted for trajectory prediction. In our design, we utilize uncertainty for more generalized HD map construction, which considers the variability in feature distributions and incorporates explicit instance features." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.721, + 0.212, + 0.74 + ], + "angle": 0, + "content": "3. UIGenMap" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.748, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Given surrounding-view images captured from onboard cameras, we aim to construct more robust vectorized map elements in the BEV space. Each map element comprises a class label \\( c_{\\mathrm{map}} \\) and an ordered sequence of points \\( \\mathbf{P}_{\\mathrm{map}} = \\{(p_x^i,p_y^i)\\}_{i = 1}^N \\) that represents the structure of the instance. \\( N \\) denotes the total number of points for each map element." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The architecture of UIGenMap is presented in Figure 2. Based on the main BEV detection, we obtain PV instance from an introduced pre-trained PV branch, which is injected as structural priors for BEV perception. Firstly, we extract" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.907, + 0.706 + ], + "angle": 0, + "content": "image features for the PV and BEV branches separately. As in [22], BEV features \\(\\mathrm{F_{bev}}\\) are constructed from the interaction between PV features and learnable BEV queries. Then, we design the uncertainty-aware (UA-) decoder from inner-layer UA-Attention and point-level heads to achieve reliable probabilistic outputs. Based on the PV coordinates and uncertainty output, we propose the UI2DPrompt module to construct reliable PV prompts. These prompts are incorporated into BEV features and map instance queries via hybrid injection, where uncertainties are utilized as the basis for feature selection and enhancement. Finally, the UA-Decoder employs enhanced BEV features and instance queries for reliable map predictions. To ensure real-time inference, the lightweight MQ-Distillation module performs query distillation from learned PV prompts to imitate explicit structural features, aiming for efficient computation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.713, + 0.766, + 0.729 + ], + "angle": 0, + "content": "3.1. Uncertainty-Aware Decoder" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.856 + ], + "angle": 0, + "content": "We design UA-Decoder for both PV and BEV space. Systematic uncertainty-aware design is employed to improve the statistical understanding of features and outputs to achieve more dynamic self-adjustment for different driving scenarios. To ensure comprehensive perception, each decoder layer comprises UA-Attention at the instance feature level and UA-Head for point-level uncertainty output along with positions, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.902 + ], + "angle": 0, + "content": "UA-Attention. In conventional deformable attention [57], the weight \\(\\alpha_{i}\\) for each sampled feature values from BEV feature \\(F_{\\mathrm{BEV}}\\) is determined by linear layer (Linear(·)) on" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22361" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.092, + 0.483, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.264, + 0.484, + 0.322 + ], + "angle": 0, + "content": "Figure 3. UA-Attention Design. The learned probabilistic weights \\(\\alpha_{i}\\) are dynamically resampled from Gaussian distribution with stochastic disturbance, which is multiplied with sampled feature values to update map queries." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.33, + 0.484, + 0.574 + ], + "angle": 0, + "content": "\\(Q_{\\mathrm{map}}^i\\) , which can be expressed as \\(\\alpha_{i} = \\mathrm{Linear}(Q_{\\mathrm{map}}^{i})\\) where \\(Q_{\\mathrm{map}}^i\\) is the \\(i\\) -th instance query of totally \\(M\\) map elements. Then, the updated part of query \\(\\Delta Q_{\\mathrm{map}}^i\\) is aggregated by the dot-product between the learned weight \\(\\alpha_{i}\\) and sampled feature values from \\(\\mathbf{F}_{\\mathrm{bev}}\\) . Considering complex driving scenarios, such deterministic weights are inadaptable for challenging situations. More dynamic weights are required to avoid potential ambiguities. Inspired by [8, 34], we explicitly model the attention scores under a Gaussian distribution \\(\\alpha_{i}\\sim \\mathcal{N}(\\mu_{i},\\sigma_{i}^{2})\\) , as shown in Figure 3. We adopt the re-parameterization strategy to perform the probabilistic attention, which is constructed by the corresponding mean \\(\\mu_{i}\\) and variance \\(\\sigma_{i}\\) . Such probabilistic parameters are predicted through multilayer perception (MLP(·)). Thus, the construction of dynamic resampled attention weights can be formulated as below:" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.588, + 0.482, + 0.605 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\mu_ {i}, \\sigma_ {i} ^ {2} \\right\\} = \\mathrm {M L P} \\left(Q _ {\\text {m a p}} ^ {i}\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.61, + 0.482, + 0.625 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i} = \\mu_ {i} + \\sigma_ {i} \\epsilon , \\epsilon \\sim \\mathcal {N} (0, 1), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.637, + 0.483, + 0.728 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) is the Gaussian noise conforming to normal distribution. During inference, attention weight \\(\\alpha_{i}\\) can be dynamically resampled from learned statistical distribution to suit for driving scenarios with large variations. Then query features can be obtained by dot production with the sampled feature values in order to enable reliable query updates." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.729, + 0.484, + 0.905 + ], + "angle": 0, + "content": "UA-Head with Uncertainty Loss. Based on the robust design of UA-Attention, quantized uncertainty for each lane point output is also essential, which is point-specific at a fine-grained level to explicitly reflect the reliability and uncertainty information within the internal transformer layers. Inspired by [7], we replace the vanilla deterministic output with the UA-Head that predicts coordinates \\(\\mathbf{P}_{\\mathrm{map}} = \\{(\\hat{p}_x^i,\\hat{p}_y^i)\\}_{i = 1}^N\\) along with their corresponding uncertainties \\(\\sigma_{\\mathrm{map}} = \\{(\\sigma_x^i,\\sigma_y^i)\\}_{i = 1}^N\\) for each map element, as shown in Figure 2 (b). These parameters represent the location and scale of the joint Laplace distribution \\(Laplace(\\hat{p}_j^i,\\frac{\\sigma_j^i}{\\sqrt{2}})\\) at" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "the \\(j\\)-th dimension \\((x\\) or \\(y)\\) of the \\(i\\)-th point, which refers to a probabilistic representation for stable output." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.905, + 0.183 + ], + "angle": 0, + "content": "For uncertainty-aware training, we combine probabilistic Negative Log-Likelihood (NLL) loss \\(\\mathcal{L}_{\\mathrm{nll}}\\) with point regression loss \\(\\mathcal{L}_{\\mathrm{pts}}\\) in [50] to balance uncertainties and coordinate outputs, in which \\(\\mathcal{L}_{\\mathrm{nll}}\\) can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.191, + 0.905, + 0.235 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {n l l}} = \\lambda_ {\\mathrm {n l l}} \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {2} \\left(\\log \\left(2 \\sigma_ {j} ^ {i}\\right) + \\frac {\\left| p _ {j} ^ {i} - \\hat {p} _ {j} ^ {i} \\right|}{\\sigma_ {j} ^ {i}}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.243, + 0.746, + 0.258 + ], + "angle": 0, + "content": "with \\(\\lambda_{\\mathrm{nl}}\\) as the loss weight setting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.907, + 0.412 + ], + "angle": 0, + "content": "As illustrated in Figure 2 (a), UA-Decoder is used for the construction of the PV branch and the BEV map. To provide more robust structural features in PV branch, PV coordinates \\(\\mathrm{P_{pv}} = \\{x_{pv}^i,y_{pv}^i\\}_{i = 1}^N\\) and the uncertainty parameters \\(\\sigma_{\\mathrm{pv}} = \\{\\sigma_{\\mathrm{pv,x}}^{i},\\sigma_{\\mathrm{pv,y}}^{i}\\}_{i = 1}^{N}\\) are constructed by the UA-Decoder structure, which is later utilized in our UI2DPrompt design. Thus, the systematic combination of uncertainty-aware design provides dynamic sampling during the learning process, along with established uncertainty, to construct a reliable basis for confidence scoring." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.419, + 0.652, + 0.436 + ], + "angle": 0, + "content": "3.2. UI2DPrompt" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.906, + 0.561 + ], + "angle": 0, + "content": "PV branches contain abundant structural and geometric details. To mitigate information loss and misinterpretation during query-based PV-to-BEV transformation, reliable structural PV instances can be injected as compensation for BEV prediction. This section focuses on constructing PV prompts based on uncertainty-aware PV predictions, which are dynamically integrated with the map decoder for BEV map element detection." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.562, + 0.907, + 0.79 + ], + "angle": 0, + "content": "Prompt Construction. Given the learned PV instance coordinates and the corresponding uncertainty parameters \\(\\{\\mathrm{P_{pv}},\\sigma_{\\mathrm{pv}}\\}\\), PV prompts are constructed to provide dynamic 2D priors. Firstly, we employ the estimated classification score \\(c_{\\mathrm{map}}\\) for PV candidates selection. When \\(c_{\\mathrm{map}} > c_{\\mathrm{thr}}\\), PV instances are chosen as candidates for reliable localized information. Considering the sparse depth information and the characteristics of static map elements on the ground, we utilize Inverse Perspective Mapping (IPM) to transform \\(\\mathrm{P_{pv}}\\) from the PV image coordinate to the BEV coordinate system \\(\\mathrm{P}_{\\mathrm{trans}} = \\{p_x^{\\prime},p_y^{\\prime}\\}\\), which makes a flat grounding assumption and utilizes the affine transformation between the PV and BEV coordinates. Given the corresponding intrinsic parameters \\(\\mathbf{K}_{\\mathrm{cam}}\\) and extrinsic transformation matrix \\(\\mathrm{T}_{\\mathrm{ego2cam}}\\), the projection \\(\\mathrm{P_{pv}}\\) can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.798, + 0.905, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {\\text {t r a n s}} = \\operatorname {P r o j} _ {2 D} \\left(\\mathrm {T} _ {\\text {e g o 2 c a m}} ^ {- 1} \\cdot \\left[ \\begin{array}{c} \\mathrm {K} _ {\\text {c a m}} ^ {- 1} \\cdot \\mathrm {P} _ {\\mathrm {p v}} \\\\ 1 \\end{array} \\right]\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.871 + ], + "angle": 0, + "content": "where \\(\\operatorname{Proj}_{2\\mathrm{D}}(\\cdot)\\) denotes the conversion from homogeneous coordinates into 2D plane coordinates." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "To ensure reliable enhanced features, the corresponding uncertainty parameters \\(\\sigma_{\\mathrm{pv}}\\) cooperate as weights and feature" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22362" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.152 + ], + "angle": 0, + "content": "embeddings. We encode the transformed coordinates \\(\\mathbf{P}_{\\mathrm{trans}}\\) and uncertainty parameters \\(\\sigma_{\\mathrm{pv}}\\) by \\(\\phi_p(\\cdot)\\) and \\(\\phi_{\\sigma}(\\cdot)\\), respectively. Point-specific embeddings \\(\\{e_{pv}^i\\}_{i = 1}^N\\) are constructed by concatenation:" + }, + { + "type": "equation", + "bbox": [ + 0.159, + 0.159, + 0.483, + 0.178 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {e} _ {\\mathrm {p v}} ^ {i} = \\operatorname {C o n c a t} \\left(\\left[ \\phi_ {p} \\left(\\mathrm {P} _ {\\text {t r a n s}}\\right), \\phi_ {\\sigma} \\left(\\sigma_ {\\mathrm {p v}}\\right) \\right]\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.483, + 0.214 + ], + "angle": 0, + "content": "Thus, the structural PV output coordinates and uncertainty information are dynamically encoded." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.214, + 0.484, + 0.306 + ], + "angle": 0, + "content": "Based on the above, PV uncertainties can serve as the quantized weights for dynamic feature strengthening and adjustment. Reliable confidence indicator is able to be formulated as the weight \\(\\{\\omega_{\\mathrm{pv}}^i\\}_{i = 1}^N\\) concerning each point-specific prompt. Then, the enhanced PV prompt \\(\\tilde{\\pmb{e}}_{\\mathrm{pv}}^{i}\\) can be constructed by:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.312, + 0.483, + 0.365 + ], + "angle": 0, + "content": "\\[\n\\omega_ {\\mathrm {p v}} ^ {i} = \\exp \\left(\\frac {\\left(\\| \\sigma_ {\\mathrm {p v}} ^ {i} \\| _ {2}\\right) ^ {- 1}}{\\sum_ {i = 1} ^ {N} \\left(\\| \\sigma_ {\\mathrm {p v}} ^ {i} \\| _ {2}\\right) ^ {- 1}}\\right), \\quad \\tilde {\\boldsymbol {e}} _ {\\mathrm {p v}} ^ {i} = \\omega_ {\\mathrm {p v}} ^ {i} \\boldsymbol {e} _ {\\mathrm {p v}} ^ {i} + \\boldsymbol {e} _ {\\mathrm {m}} ^ {i}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.366, + 0.483, + 0.456 + ], + "angle": 0, + "content": "where \\(\\exp (\\cdot)\\) represents the exponential amplification of the normalized inverse uncertainty value. \\(\\{e_{\\mathrm{m}}^{i}\\}_{i = 1}^{N}\\) is the predefined mimic query, which is utilized for MQ-Distillation. This approach reduces the introduction of PV errors, ensuring that injecting PV instances amplifies reliable PV information while minimizing uncertainty-induced errors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.484, + 0.562 + ], + "angle": 0, + "content": "Hybrid Injection. Constructed from PV instances with uncertainty output, PV prompts \\(\\{\\tilde{e}_{\\mathrm{pv}}^i\\}_{i = 1}^N\\) are incorporated as compensation for the construction of the BEV map. These prompts are injected into the BEV feature \\(\\mathbf{F}_{\\mathrm{bev}}\\) and the map queries \\(Q_{\\mathrm{map}}^i\\) for the map decoder by the designed hybrid injection mechanism, which includes both prompt-to-BEV (P2BEV) and prompt-to-query (P2Q) attention." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.562, + 0.484, + 0.638 + ], + "angle": 0, + "content": "As shown in Figure 4, point-level PV prompts \\(\\{\\tilde{e}_{\\mathrm{pv}}^i\\}_{i = 1}^N\\) are integrated with \\(\\mathbf{F}_{\\mathrm{bev}}\\) through cross-attention of P2BEV. This step enriches BEV features by integrating detailed, point-specific information from PV prompts, enhancing instance-specific detail in the BEV representation:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.644, + 0.483, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\mathrm {F} _ {\\mathrm {b e v}} ^ {\\prime} = \\operatorname {A t t n} \\left(W _ {q} \\mathrm {F} _ {\\mathrm {b e v}}, W _ {k} \\tilde {\\boldsymbol {e}} _ {\\mathrm {p v}} ^ {i}, W _ {v} \\tilde {\\boldsymbol {e}} _ {\\mathrm {p v}} ^ {i}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.669, + 0.484, + 0.789 + ], + "angle": 0, + "content": "where \\( W_{\\{q,k,v\\}} \\) are the corresponding projection matrix. Then, we encode point-level prompts into instance-level representations through a linear block \\( \\phi_e(\\cdot) \\), which aggregates point-specific information to create a higher-level representation for subsequent processing. These instance-level prompts \\( \\tilde{e}_{\\mathrm{inst}} \\) are used to integrate with \\( Q_{\\mathrm{map}} \\), which inject reliable priors for the initialized queries and enhance the overall map construction:" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.795, + 0.483, + 0.815 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {e}} _ {\\text {i n s t}} = \\phi_ {e} \\left(\\left\\{\\tilde {\\boldsymbol {e}} _ {\\mathrm {p v}} ^ {i} \\right\\} _ {i = 1} ^ {N}\\right), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.835, + 0.483, + 0.853 + ], + "angle": 0, + "content": "\\[\nQ _ {\\text {m a p}} ^ {\\prime} = \\operatorname {A t t n} \\left(W _ {q} Q _ {\\text {m a p}}, W _ {k} \\tilde {\\mathbf {e}} _ {\\text {i n s t}}, W _ {v} \\tilde {\\mathbf {e}} _ {\\text {i n s t}}\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "By injecting PV priors into both BEV and decoder queries, this hybrid mechanism improves the reliability of BEV perception and reduces transition errors. Explicit structural" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.905, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.27, + 0.907, + 0.313 + ], + "angle": 0, + "content": "Figure 4. Design of Hybrid Injection. P2BEV-Attention and P2Q-Attention are designed for PV prompts to integrate with BEV features and map instance queries as structural compensation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.325, + 0.907, + 0.372 + ], + "angle": 0, + "content": "features enhance the generalization capability of the learned model and perform robustly in various driving environments, ensuring better accuracy and consistency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.381, + 0.743, + 0.398 + ], + "angle": 0, + "content": "3.3. Mimic Query Distillation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.405, + 0.906, + 0.464 + ], + "angle": 0, + "content": "To address the extra computation introduced by the PV branch, we propose the lightweight MQ-Distillation to distill from PV prompt construction, where the mimic query can serve as an alternative to achieve real-time inference." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.466, + 0.907, + 0.586 + ], + "angle": 0, + "content": "In MQ-Distillation, we define a set of mimic queries \\(\\{e_m^i\\}_{i = 1}^N\\), along with an MLP-based query learner \\(h(\\cdot)\\) which is designed to capture and distill the structural features of the PV prompt construction process \\(\\tilde{e}_{\\mathrm{pv}}^i\\). As shown in Equation 6, during training, the predefined mimic queries \\(e_{m}^{i}\\) and the constructed PV prompts \\(\\tilde{e}_{\\mathrm{pv}}^i\\) are jointly combined by addition, which are sent to the hybrid injection for further operations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.587, + 0.906, + 0.677 + ], + "angle": 0, + "content": "To obtain sufficient structural features and achieve progressive alignment with PV prompts, we design the loss function that guides mimic queries for distillation. We use Mean Squared Error (MSE) as the distillation loss to measure the difference between the constructed PV prompts and the mimicked queries, which can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.686, + 0.905, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i s t i l l}} = \\lambda_ {\\text {d i s t i l l}} \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\tilde {\\boldsymbol {e}} _ {\\mathrm {p v}} ^ {i} - h \\left(\\boldsymbol {e} _ {m} ^ {i}\\right)\\right) ^ {2}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.809 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{distill}}\\) is a weight coefficient that balances the impact of the distillation loss on the overall training process. Such loss function guides the mimic queries to accurately replicate the PV prompts generated by the PV branch, making them reliable substitutes for inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To this end, only the mimic queries can be utilized during inference, which eliminates the requirement for PV prompt computation. This approach maintains real-time inference capability by efficiently mimicking PV structural features with minimal computational cost, enabling streamlined yet robust map construction." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22363" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.905, + 0.281 + ], + "angle": 0, + "content": "
MethodBackboneEpochsRegion-BasedCity-BasedFPS
APpedAPdivAPboumAPAPpedAPdivAPboumAP
VectorMapNet* [ICML'23]R5012015.817.021.218.0-----
MapTR* [ICLR'23]R50246.420.735.520.911.112.921.115.015.8
MGMap [CVPR'24]R50249.427.038.124.810.917.824.717.811.5
MapQR [ECCV'24]R50248.425.739.924.710.215.825.517.212.1
GeMap [ECCV'24]R502411.826.643.627.312.916.326.718.611.6
MapTRv2* [IJCV'24]R502414.728.743.328.918.918.627.921.812.9
StreamMapNet* [WACV'24]R502432.229.340.834.118.716.422.719.313.3
UIGenMap-d (Ours)R502440.330.846.839.3 (+5.2)23.119.625.422.7 (+3.4)12.2
UIGenMap (Ours)R502440.331.847.239.8 (+5.7)24.120.726.123.6 (+4.3)8.2
StreamMapNet [WACV'24]SwinT2432.527.441.733.921.819.725.822.49.3
UIGenMap-d (Ours)SwinT2442.631.947.240.6 (+6.7)24.721.228.824.9 (+2.5)8.5
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.282, + 0.907, + 0.339 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons on nuScenes validation dataset, utilizing region-based and city-based dataset partitions. Methods with “*” denote that data results are obtained from corresponding papers [27, 50], and others are reproduced by official public codes. “R50” and “SwinT” denotes the image backbone of ResNet50 [9] and Swin-Tiny [30]. “UIGenMap-d” represents the utilization of mimic query from MQ-Distillation. FPS is measured on a single RTX A40 GPU." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.349, + 0.307, + 0.366 + ], + "angle": 0, + "content": "3.4. Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.371, + 0.483, + 0.418 + ], + "angle": 0, + "content": "During the training process, combined with \\(\\mathcal{L}_{\\mathrm{nl}}\\) for uncertainty output and \\(\\mathcal{L}_{\\mathrm{distill}}\\) for mimic query distillation, the overall loss setting \\(\\mathcal{L}_{\\mathrm{map}}\\) can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.426, + 0.483, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a p}} = \\lambda_ {1} \\mathcal {L} _ {\\text {p t s}} + \\lambda_ {2} \\mathcal {L} _ {\\text {c l s}} + \\mathcal {L} _ {\\text {n l l}} + \\mathcal {L} _ {\\text {d i s t i l l}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.45, + 0.483, + 0.525 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{pts}}\\) denotes the point-to-point Manhattan distance between the regressed points and the ground-truth. \\(L_{\\mathrm{cls}}\\) represents the focal loss for map classification. \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are the corresponding loss weights for point regression and map classification, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.483, + 0.648 + ], + "angle": 0, + "content": "For inference, the well-learned mimic query can only be employed as a substitute for PV prompts, avoiding the computational burden for the PV detection branch. Thus, our approach ensures real-time inference without introducing unnecessary computational overhead, while preserving effective performance. Additionally, the model's uncertainty learning enables dynamic sampling during inference, enhancing the adaptability to diverse driving environments." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.658, + 0.224, + 0.675 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.29, + 0.697 + ], + "angle": 0, + "content": "4.1. Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Datasets. We conduct extensive experiments on widely used autonomous driving datasets, including nuScenes [1] and Argoverse2 [46]. nuScenes dataset comprises 1,000 driving scenes from Boston and Singapore. Each scene lasts approximately 20 seconds with a \\(2\\mathrm{Hz}\\) sampling rate. For each sample, there are six surrounding images and 32-beam LiDAR point clouds. Argoverse2 contains 1,000 scenes from 6 cities, with each scene capturing 15 seconds of data at \\(20\\mathrm{Hz}\\) from 7 cameras and \\(10\\mathrm{Hz}\\) LiDAR sweeps. In our experiments, the sampling frequency for Argoverse2 is unified into \\(2\\mathrm{Hz}\\) as in StreamMapNet [50]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Geo-based Data Partitions. Considering the overlap of widely used training and validation partitions, we adopt" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.35, + 0.907, + 0.577 + ], + "angle": 0, + "content": "geographically disjoint (geo-based) dataset partitions as benchmarks to evaluate the generalization capability of the presented model. Experiments are conducted on both region-based and city-based data partitions. The region-based partition is adopted from [50], where the driving locations geographically disjoin the training and validation sets. In the city-based data partition, we apply city-wise splits in [27], where the training and validation data are from different cities with a larger distribution change. On nuScenes, Boston and Onenorth are used for training, while Queenstown and Holland Village are used for validation. For Ar-goverse2, Miami and Pittsburgh are used in training, and the remains are employed for validation. For different dataset partitions, models are separately trained on that split's training set and evaluated on their corresponding validation set." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.578, + 0.909, + 0.716 + ], + "angle": 0, + "content": "Evaluation Metrics. In line with the existing works [26, 50], we conduct experiments on three static map elements, namely lane dividers, pedestrian crossings, and road boundaries. Perception range is set to \\([-15.0m, 15.0m]\\) for X-axis and \\([-30.0m, 30.0m]\\) for Y-axis. We use average precision as the evaluation metric, which calculates the Chamfer distance between predictions and ground truths. The mean Average Precision (mAP) is obtained by averaging AP across the distinct thresholds: \\(\\{0.5m, 1m, 1.5m\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.727, + 0.733, + 0.743 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "For fair comparisons, we employ ResNet50 [9] as the backbone. As in [50], the size of BEV features is set to \\( 50 \\times 100 \\). The number of map queries and query dimensions are set to 100 and 512. We use AdamW optimizer with a learning rate of \\( 4e^{-4} \\). Training epochs are limited to prevent potential overfitting. All models are trained on 4 NVIDIA Tesla V100 GPUs with a batch size of 4. The weighted factors \\( \\lambda_{1} \\) and \\( \\lambda_{2} \\) are set to 50.0 and 5.0, separately. \\( \\lambda_{\\mathrm{nil}} \\) and \\( \\lambda_{\\mathrm{distill}} \\) are set to 0.05 and 10.0. We adopt a pre-trained PV detection branch for our UIGenMap training process. More details" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22364" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.905, + 0.235 + ], + "angle": 0, + "content": "
MethodBackboneEpochsRegion-BasedCity-BasedFPS
\\(AP_{ped}\\)\\(AP_{div}\\)\\(AP_{bou}\\)mAP\\(AP_{ped}\\)\\(AP_{div}\\)\\(AP_{bou}\\)mAP
\\(VectorMapNet^{*}\\)[ICML'23]R5012035.634.937.836.1-----
\\(MapTR^{*}\\)[ICLR'23]R503048.150.455.051.134.741.734.737.915.8
\\(GeMap\\)[ECCV'24]R503052.559.258.256.634.541.140.438.711.6
\\(MapTRv2^{*}\\)[JCV'24]R503053.358.758.856.937.442.241.940.512.9
\\(MapQR\\)[ECCV'24]R503054.258.060.457.533.841.541.638.912.1
\\(StreamMapNet^{*}\\)[WACV'24]R503057.955.761.358.340.243.443.342.313.3
\\(UIGenMap-d\\)(Ours)R503058.358.462.459.7 (+1.4)42.944.444.744.0 (+1.7)12.2
\\(UIGenMap\\)(Ours)R503059.858.762.760.4 (+2.1)43.745.745.144.8 (+2.5)8.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.237, + 0.905, + 0.281 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons on Argoverse2 validation datasets, utilizing region-based and city-based dataset partitions. Methods with “*” denote that data results are obtained from corresponding papers [27, 50], and others are reproduced by official public codes. Models are trained for 30 epochs and follow the same setting. FPS is measured on a single RTX A40 GPU." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.289, + 0.486, + 0.405 + ], + "angle": 0, + "content": "
DatasetMethodEpochs\\(AP_{ped}\\)\\(AP_{div}\\)\\(AP_{bou}\\)mAP
nuScenesMapTRV2 [26]2459.862.462.461.5
StreamMapNet [50]2460.461.958.960.4
SQD-MapNet [43]2463.065.563.363.9
UIGenMap(Ours)2464.265.464.964.8
Argoverse2StreamMapNet [50]3062.959.563.061.5
SQD-MapNet [43]3064.960.264.963.3
UIGenMap(Ours)3064.560.866.163.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.41, + 0.483, + 0.439 + ], + "angle": 0, + "content": "Table 3. Performance on nuScenes and Argoverse2 datasets with the original dataset partition at \\( {30}\\mathrm{\\;m} \\) range." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.442, + 0.48, + 0.554 + ], + "angle": 0, + "content": "
bkbUI2DPromptUA-DecoderAPpedAPdivAPboumAP
32.229.340.834.1
✓39.928.643.837.4
✓36.131.344.037.1
✓✓41.629.744.838.7
✓✓41.429.847.439.5
✓✓39.230.146.238.5
✓✓✓40.331.847.239.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.558, + 0.483, + 0.587 + ], + "angle": 0, + "content": "Table 4. Ablations on the impact of component designs, 'bkb' means the shared image backbone from PV branches." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.596, + 0.388, + 0.612 + ], + "angle": 0, + "content": "are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.621, + 0.231, + 0.637 + ], + "angle": 0, + "content": "4.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.644, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Results on nuScenes. We compare the generalization capability with previous methods [25, 26, 28, 31, 50, 54] on geospatial data partitions, in which the image size is set to \\(480 \\times 800\\). As shown in Table 1, our UIGenMap outperforms previous methods at a large margin. For the ResNet50 backbone, UIGenMap achieves \\(5.7\\mathrm{mAP}\\) improvement on the region-based partition, and \\(4.3\\mathrm{mAP}\\) gain is obtained on the city-based splits. For more efficient design, UIGenMapd also archives \\(39.3\\mathrm{mAP}\\) with a faster inference speed. Experiments on the Swin-T [30] backbone show significant performance with \\(40.6\\mathrm{mAP}\\) achievement. Furthermore, we show some visual examples in Figure 5, indicating more intuitive achievements for UIGenMap." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Results on Argoverse2. More experiments are conducted on Argoverse2 dataset. The image size is \\(608 \\times 608\\), and the sampling rate is \\(2\\mathrm{Hz}\\). For a fair comparison, the models are trained for 30 epochs. As shown in Table 2, UIGen" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.292, + 0.907, + 0.382 + ], + "angle": 0, + "content": "Map achieves the best results of \\(60.4\\mathrm{mAP}\\) on region-based data partitions and \\(43.4\\mathrm{mAP}\\) on city-based splits, which performs better than previous approaches. Upon further investigation, calibration errors and dataset distribution also affect model performance. More visualization can be found in our supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.909, + 0.504 + ], + "angle": 0, + "content": "Results on Original Data Partitions. For thorough evaluation, we conducted experiments on the original data partitions with similar driving scenarios, which are widely used in previous studies. As shown in Table 3, we compare our methods with others utilizing the same baseline. UIGenMap still outperforms the existing approaches, which outperforms StreamMapNet with \\(+4.4\\) mAP and \\(+2.3\\) mAP separately, validating the effectiveness of our approach." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.514, + 0.68, + 0.53 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.613 + ], + "angle": 0, + "content": "In this section, ablations are conducted to verify the effectiveness of our designs. We perform experiments on nuScenes with region-based partition using ResNet50 backbone. All models are trained for 24 epochs. More ablations are provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.906, + 0.764 + ], + "angle": 0, + "content": "Contributions of Main Components. Table 4 demonstrates the impact of each component in our design. Based on the pre-trained PV backbone, we design the UA-Decoder for more reliable uncertainty output and UI2DPrompt to dynamically inject PV structural information, which achieves \\(+2.7\\) mAP and \\(+2.4\\) mAP improvement, respectively. The combination of these designs gains \\(+4.0\\) mAP total. The ablations are conducted on UA-Decoder at BEV space. Besides, the pre-trained PV backbone further emphasizes the effectiveness to \\(39.8\\) mAP for better scene understanding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.906, + 0.87 + ], + "angle": 0, + "content": "Design of UA-Decoder. We design the UA-Attention for probabilistic attention and the UA-Head to construct the uncertainty output from the instance and finer point levels. Table 5 shows the ablation without other modules. Our uncertainty-aware design supports dynamic adaptation to achieve more robust performance, which achieves \\(+3.0\\) mAP improvements based on systematic combination." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Design for UI2DPrompt. UI2DPrompt consists of prompt construction and a hybrid injection module. Table 6 shows" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22365" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.175, + 0.09, + 0.568, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.339, + 0.453, + 0.351 + ], + "angle": 0, + "content": "Surrounding Views with PV Detection" + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.091, + 0.819, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.339, + 0.635, + 0.349 + ], + "angle": 0, + "content": "Baseline" + }, + { + "type": "image_caption", + "bbox": [ + 0.761, + 0.339, + 0.78, + 0.348 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.353, + 0.9, + 0.368 + ], + "angle": 0, + "content": "Figure 5. Qualitative results on the nuScenes dataset with the region-based data partition. Uncertainty outputs are represented as circles." + }, + { + "type": "table", + "bbox": [ + 0.136, + 0.375, + 0.44, + 0.452 + ], + "angle": 0, + "content": "
UA-Att.UA-HeadAPpedAPdivAPboumAP
32.229.340.834.1
✓31.730.442.935.0
✓34.830.543.436.3
✓✓36.131.344.037.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.456, + 0.483, + 0.498 + ], + "angle": 0, + "content": "Table 5. Ablations on UA-Decoder without other modules. UAAtt and UA-Head represent the utilization of uncertainty-aware attention and output head respectively." + }, + { + "type": "table", + "bbox": [ + 0.139, + 0.505, + 0.438, + 0.582 + ], + "angle": 0, + "content": "
Coords. Trans.Unc. UtilizationmAP
//37.6
IPMembedding only39.3
IPMweight & embedding39.8
3D Trans.weight & embedding38.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.585, + 0.483, + 0.628 + ], + "angle": 0, + "content": "Table 6. Ablations on the construction of PV prompts. \"3D Trans.\" means the depth-based transformation. The first row with \\( \\frac{1}{r} \\) indicates direct PV coordinate embedding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.641, + 0.483, + 0.746 + ], + "angle": 0, + "content": "the strategy selection for constructing PV prompts. Compared to a depth-based 3D transition, IPM is more suitable for coordinating the transformation based on the assumption of a flattened road surface. When uncertainty is applied both as dynamic weighting and as learning-based embeddings, performance is further enhanced by reliable PV prompts, achieving an improvement of \\(+2.2\\mathrm{mAP}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.748, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Table 7 demonstrates the impact of the hybrid injection design, with a total performance of \\(39.8\\mathrm{mAP}\\) for combining prompt-to-BEV and prompt-to-query attention. This result confirms the optimal utilization of reliable structural information, effectively compensating for information loss and errors introduced in the basic learning process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Ablations on MQ-Distillation. To verify the effectiveness of the distillation module, we conduct ablations on the hyperparameter of distillation loss weight \\(\\lambda_{\\mathrm{distill}}\\), as this affects the imitation learning of PV structural features. As" + }, + { + "type": "table", + "bbox": [ + 0.554, + 0.375, + 0.868, + 0.452 + ], + "angle": 0, + "content": "
P2BEVP2QAPpedAPdivAPboumAP
39.330.445.538.4
✓39.931.146.039.2
✓39.630.746.138.8
✓✓40.331.847.239.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.456, + 0.906, + 0.484 + ], + "angle": 0, + "content": "Table 7. Ablations on hybrid injection. P2BEV and P2Q mean the cross-attention for \\( \\mathrm{F}_{\\mathrm{bev}} \\) and \\( Q_{\\mathrm{map}} \\) separately." + }, + { + "type": "table", + "bbox": [ + 0.565, + 0.489, + 0.857, + 0.53 + ], + "angle": 0, + "content": "
λdistill571015/
mAP37.538.239.338.839.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.533, + 0.906, + 0.562 + ], + "angle": 0, + "content": "Table 8. Ablations on \\(\\lambda_{\\mathrm{distill}}\\) for MQ-Distillation Design. The last column refers to the use of the best teacher of PV prompts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.571, + 0.906, + 0.618 + ], + "angle": 0, + "content": "illustrated in Table 8, the best performance is achieved by setting the weight to 10.0, which refers to the best learning process of the mimic query for explicit structural features." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.633, + 0.783, + 0.648 + ], + "angle": 0, + "content": "5. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.84 + ], + "angle": 0, + "content": "In this paper, we propose UIGenMap, an uncertainty-instructed structure injection approach to achieve a more generalized HD map construction. Specifically, we introduce the PV detection branch to obtain explicit structural features, in which the UA-Decoder produces adaptive probabilistic outputs. The UI2DPrompt module is proposed to construct reliable PV prompts and achieve dynamic integration through the hybrid injection designed. For real-time inference, MQ-Distillation is designed to mimic the structural prompts of PV. Comprehensive experiments on the challenging geo-based partitions of nuScenes and Argoverse2 datasets demonstrate the generalization of our approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "For limitations, current methods are designed and evaluated primarily on specific datasets. In the future, developing larger and more diverse datasets and benchmarks across various locations remains a crucial area for exploration." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22366" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.177 + ], + "angle": 0, + "content": "This work is supported by National Natural Science Foundation of China under Grant No.62376244. It is also supported by Information Technology Center and State Key Lab of CAD&CG, Zhejiang University." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.188, + 0.188, + 0.204 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.213, + 0.484, + 0.282 + ], + "angle": 0, + "content": "[1] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11621-11631, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.284, + 0.484, + 0.325 + ], + "angle": 0, + "content": "[2] Anh-Quan Cao, Angela Dai, and Raoul de Charette. Pasco: Urban 3d panoptic scene completion with uncertainty awareness. In CVPR, pages 14554–14564, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.327, + 0.483, + 0.367 + ], + "angle": 0, + "content": "[3] Jie Chang, Zhonghao Lan, Changmao Cheng, and Yichen Wei. Data uncertainty learning in face recognition. In CVPR, pages 5710-5719, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.369, + 0.482, + 0.422 + ], + "angle": 0, + "content": "[4] Jiacheng Chen, Yuefan Wu, Jiaqi Tan, Hang Ma, and Yasutaka Furukawa. Maptracker: Tracking with strided memory fusion for consistent vector hd mapping. In ECCV, pages 90-107. Springer, 2025. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.483, + 0.465 + ], + "angle": 0, + "content": "[5] Sehwan Choi, Jungho Kim, Hongjae Shin, and Jun Won Choi. Mask2map: Vectorized hd map construction using bird's eye view segmentation masks. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.482, + 0.508 + ], + "angle": 0, + "content": "[6] Wenjie Ding, Limeng Qiao, Xi Qiu, and Chi Zhang. Pivotnet: Vectorized pivot learning for end-to-end hd map construction. In ICCV, pages 3672-3682, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.51, + 0.482, + 0.563 + ], + "angle": 0, + "content": "[7] Xunjiang Gu, Guanyu Song, Igor Gilitschenski, Marco Pavone, and Boris Ivanovic. Producing and leveraging online map uncertainty in trajectory prediction. In CVPR, pages 14521-14530, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.482, + 0.606 + ], + "angle": 0, + "content": "[8] Hongji Guo, Hanjing Wang, and Qiang Ji. Uncertainty-guided probabilistic transformer for complex action recognition. In CVPR, pages 20052-20061, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.482, + 0.648 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.65, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[10] Yuzhe He, Shuang Liang, Xiaofei Rui, Chengying Cai, and Guowei Wan. Egovm: Achieving precise eg-localization using lightweight vectorized maps. arXiv preprint arXiv:2307.08991, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.706, + 0.482, + 0.76 + ], + "angle": 0, + "content": "[11] Haotian Hu, Fanyi Wang, Yaonong Wang, Laifeng Hu, Jingwei Xu, and Zhiwang Zhang. Admap: Anti-disturbance framework for reconstructing online vectorized hd map. arXiv preprint arXiv:2401.13172, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[12] Haoxuanye Ji, Pengpeng Liang, and Erkang Cheng. Enhancing 3d object detection with 2d detection-guided query anchors. In CVPR, pages 21178-21187, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.805, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[13] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, pages 8340-8350, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[14] Xiaohui Jiang, Shuai Lin Li, Yingfei Liu, Shihao Wang, Fan Jia, Tiancai Wang, Lijin Han, and Xiangyu Zhang. Far3d:" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.213, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Expanding the horizon for surround-view 3d object detection. In AAAI, pages 2561-2569, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.16 + ], + "angle": 0, + "content": "[15] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? In NeuIPS, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[16] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. In NeuIPS, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.206, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[17] Han Li, Zehao Huang, Zitian Wang, Wenge Rong, Naiyan Wang, and Si Liu. Enhancing 3d lane detection and topology reasoning with 2d lane priors. arXiv preprint arXiv:2406.03105, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.904, + 0.302 + ], + "angle": 0, + "content": "[18] Qi Li, Yue Wang, Yilun Wang, and Hang Zhao. Hdmapnet: An online hd map construction and evaluation framework. In ICRA, pages 4628-4634. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.904, + 0.357 + ], + "angle": 0, + "content": "[19] Siyu Li, Jiacheng Lin, Hao Shi, Jiaming Zhang, Song Wang, You Yao, Zhiyong Li, and Kailun Yang. Dtlmapper: Dual temporal consistent learning for vectorized hd map construction. arXiv preprint arXiv:2405.05518, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.904, + 0.413 + ], + "angle": 0, + "content": "[20] Siyu Li, Kailun Yang, Hao Shi, Song Wang, You Yao, and Zhiyong Li. Genmapping: Unleashing the potential of inverse perspective mapping for robust online hd map construction. arXiv preprint arXiv:2409.08688, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.415, + 0.904, + 0.455 + ], + "angle": 0, + "content": "[21] Xiaotong Li, Yongxing Dai, Yixiao Ge, Jun Liu, Ying Shan, and Lingyu Duan. Uncertainty modeling for out-of-distribution generalization. In ICLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.457, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[22] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In ECCV, pages 1-18. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[23] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In ICCV, pages 14864-14873, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.582, + 0.904, + 0.636 + ], + "angle": 0, + "content": "[24] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. In NeuIPS, pages 10421-10434, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.638, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[25] Benchcheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. Maptr: Structured modeling and learning for online vectorized hd map construction. In ICLR, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.693, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[26] Benchcheng Liao, Shaoyu Chen, Yunchi Zhang, Bo Jiang, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Maptrv2: An end-to-end framework for online vectorized hd map construction. IJCV, pages 1-23, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[27] Adam Lilja, Junsheng Fu, Erik Stenberg, and Lars Hammarstrand. Localization is all you evaluate: Data leakage in online mapping datasets and how to fix it. In CVPR, pages 22150-22159, 2024. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[28] Xiaolu Liu, Song Wang, Wentong Li, Ruizi Yang, Junbo Chen, and Jianke Zhu. Mgmt: Mask-guided learning for online vectorized hd map construction. In CVPR, pages 14812-14821, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[29] Yicheng Liu, Tianyuan Yuan, Yue Wang, Yilun Wang, and Hang Zhao. Vectormapnet: End-to-end vectorized hd map learning. In ICML, pages 22352-22369. PMLR, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22367" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[30] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[31] Zihao Liu, Xiaoyu Zhang, Guangwei Liu, Ji Zhao, and Ningyi Xu. Leveraging enhanced queries of point sets for vectorized map construction. arXiv preprint arXiv:2402.17430, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[32] Hao Lu, Yunpeng Zhang, Qing Lian, Dalong Du, and Yingcong Chen. Towards generalizable multi-camera 3d object detection via perspective debiasing. arXiv preprint arXiv:2310.11346, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.265, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[33] Ao Luo, Fan Yang, Xin Li, Lang Nie, Chunyu Lin, Haoqiang Fan, and Shuaicheng Liu. Gaflow: Incorporating gaussian attention into optical flow. In ICCV, pages 9642–9651, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.323, + 0.482, + 0.365 + ], + "angle": 0, + "content": "[34] Jiahuan Pei, Cheng Wang, and György Szarvas. Transformer uncertainty estimation with hierarchical stochastic attention. In AAAI, pages 11147-11155, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.367, + 0.482, + 0.409 + ], + "angle": 0, + "content": "[35] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, pages 194-210. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.411, + 0.482, + 0.45 + ], + "angle": 0, + "content": "[36] Fengchun Qiao and Xi Peng. Uncertainty-guided model generalization to unseen domains. In CVPR, pages 6790-6800, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.454, + 0.482, + 0.495 + ], + "angle": 0, + "content": "[37] Limeng Qiao, Wenjie Ding, Xi Qiu, and Chi Zhang. End-to-end vectorized hd-map construction with piecewiseBezier curve. In CVPR, pages 13218-13228, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.482, + 0.539 + ], + "angle": 0, + "content": "[38] Tixiao Shan and Brendan Englot. Lego-loam: Lightweight and ground-optimized lidar odometry and mapping on variable terrain. In IROS, pages 4758–4765. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.541, + 0.482, + 0.597 + ], + "angle": 0, + "content": "[39] Tixiao Shan, Brendan Englot, Drew Meyers, Wei Wang, Carlo Ratti, and Daniela Rus. Lio-sam: Tightly-coupled lidar inertial odometry via smoothing and mapping. In IROS, pages 5135-5142. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.6, + 0.482, + 0.654 + ], + "angle": 0, + "content": "[40] Jingyu Song, Xudong Chen, Liupei Lu, Jie Li, and Katherine A Skinner. Memfusionmap: Working memory fusion for online vectorized hd map construction. arXiv preprint arXiv:2409.18737, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.468, + 0.671 + ], + "angle": 0, + "content": "[41] A Vaswani. Attention is all you need. In NeuIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.673, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[42] Song Wang, Wentong Li, Wenyu Liu, Xiaolu Liu, and Jianke Zhu. Lidar2map: In defense of lidar-based semantic map construction using online camera distillation. In CVPR, pages 5186-5195, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.731, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[43] Shuo Wang, Fan Jia, Yingfei Liu, Yucheng Zhao, Zehui Chen, Tiancai Wang, Chi Zhang, Xiangyu Zhang, and Feng Zhao. Stream query denoising for vectorized hd map construction. In ECCV, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[44] Song Wang, Zhongdao Wang, Jiawei Yu, Wentong Li, Bailan Feng, Junbo Chen, and Jianke Zhu. Relioco: Towards reliable semantic occupancy prediction via uncertainty learning. arXiv preprint arXiv:2409.18026, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[45] Song Wang, Jiawei Yu, Wentong Li, Wenyu Liu, Xiaolu Liu, Junbo Chen, and Jianke Zhu. Not all voxels are equal: Hardness-aware semantic scene completion with self-distillation. In CVPR, pages 14792-14801, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[46] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In NeurIPS, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[47] Qiangqiang Wu, Yan Xia, Jia Wan, and Antoni B Chan. Boosting 3d single object tracking with 2d matching distillation and 3d pre-training. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[48] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In CVPR, pages 17830-17839, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.358 + ], + "angle": 0, + "content": "[49] Jing Yang, Minyue Jiang, Sen Yang, Xiao Tan, Yingying Li, Errui Ding, Hanli Wang, and Jingdong Wang. Mgmtapnet: Multi-granularity representation learning for end-to-end vectorized hd map construction. arXiv preprint arXiv:2410.07733, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.905, + 0.415 + ], + "angle": 0, + "content": "[50] Tianyuan Yuan, Yicheng Liu, Yue Wang, Yilun Wang, and Hang Zhao. Streammapnet: Streaming mapping network for vectorized online hd map construction. In WACV, pages 7356-7365, 2024. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.471 + ], + "angle": 0, + "content": "[51] Gongjie Zhang, Jiahao Lin, Shuang Wu, Zhipeng Luo, Yang Xue, Shijian Lu, Zuoguan Wang, et al. Online map vectorization for autonomous driving: A rasterization perspective. In NeuIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.474, + 0.905, + 0.515 + ], + "angle": 0, + "content": "[52] Xiaoyu Zhang, Guangwei Liu, Zihao Liu, Ningyi Xu, Yunhui Liu, and Ji Zhao. Enhancing vectorized map perception with historical rasterized maps. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.517, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[53] Yunpeng Zhang, Wenzhao Zheng, Zheng Zhu, Guan Huang, Jiwen Lu, and Jie Zhou. A simple baseline for multi-camera 3d object detection. In AAAI, pages 3507-3515, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.56, + 0.905, + 0.613 + ], + "angle": 0, + "content": "[54] Zhixin Zhang, Yiyuan Zhang, Xiaohan Ding, Fusheng Jin, and Xiangyu Yue. Online vectorized hd map construction using geometry. arXiv preprint arXiv:2312.03341, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.616, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[55] Lijun Zhou, Tao Tang, Pengkun Hao, Zihang He, Kalok Ho, Shuo Gu, Wenbo Hou, Zhihui Hao, Haiyang Sun, Kun Zhan, et al. Ua-track: Uncertainty-aware end-to-end 3d multi-object tracking. arXiv preprint arXiv:2406.02147, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.673, + 0.905, + 0.728 + ], + "angle": 0, + "content": "[56] Yi Zhou, Hui Zhang, Jiaqian Yu, Yifan Yang, Sangil Jung, Seung-In Park, and ByungIn Yoo. Himap: Hybrid representation learning for end-to-end vectorized hd map construction. In CVPR, pages 15396-15406, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.729, + 0.905, + 0.784 + ], + "angle": 0, + "content": "[57] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.786, + 0.905, + 0.826 + ], + "angle": 0, + "content": "[58] Yanjin Zhu, Xin Zheng, and Jianke Zhu. Mesh-loam: Real-time mesh-based lidar odometry and mapping. TIV, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.826 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22368" + } + ] +] \ No newline at end of file diff --git a/2025/Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction/308073a9-0b23-46de-a5fd-816ea2a49dc7_origin.pdf b/2025/Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction/308073a9-0b23-46de-a5fd-816ea2a49dc7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..37be1978f30fe26a982b529d651934d8c1851081 --- /dev/null +++ b/2025/Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction/308073a9-0b23-46de-a5fd-816ea2a49dc7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:19cfef8c11fd73545cbeea33bd50945e6feae419adfbc222f67b6a3ffdee31f4 +size 1873880 diff --git a/2025/Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction/full.md b/2025/Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e0ffda78c54d2b888bd587aadc77ae34a76667ff --- /dev/null +++ b/2025/Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction/full.md @@ -0,0 +1,345 @@ +# Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction + +Xiaolu Liu $^{1*}$ , Ruizi Yang $^{1*}$ , Song Wang $^{1}$ , Wentong Li $^{2}$ , Junbo Chen $^{3\dagger}$ , Jianke Zhu $^{1\dagger}$ $^{1}$ Zhejiang University $^{2}$ NUAA $^{3}$ Udeer.ai + +{xiaoluliu, ruiziyang, songw, jkzhu}@zju.edu.cn, wentong.li@nuaa.edu.cn, junbo@udeer.ai + +# Abstract + +Reliable high-definition (HD) map construction is crucial for the driving safety of autonomous vehicles. Although recent studies demonstrate improved performance, their generalization capability across unfamiliar driving scenes remains unexplored. To tackle this issue, we propose UIGenMap, an uncertainty-instructed structure injection approach for generalizable HD map vectorization, which concerns the uncertainty resampling in statistical distribution and employs explicit instance features to reduce excessive reliance on training data. Specifically, we introduce the perspective-view (PV) detection branch to obtain explicit structural features, in which the uncertainty-aware decoder is designed to dynamically sample probability distributions considering the difference in scenes. With probabilistic embedding and selection, UI2DPrompt is proposed to construct PV-learnable prompts. These PV prompts are integrated into the map decoder by designed hybrid injection to compensate for neglected instance structures. To ensure real-time inference, a lightweight Mimic Query Distillation is designed to learn from PV prompts, which can serve as an efficient alternative to the flow of PV branches. Extensive experiments on challenging geographically disjoint (geo-based) data splits demonstrate that our UIGenMap achieves superior performance, with $+5.7$ mAP improvement on the nuScenes dataset. Source code is available at https://github.com/xiaolul2/UIGenMap. + +# 1. Introduction + +Reliable HD map construction is crucial for driving safety of autonomous vehicles, which provides essential geometric information for ego-localization, perceptions, path planning, and predictions [10, 13, 23, 45]. Recently, online high-definition (HD) map vectorization has attracted considerable attention for its cost-effectiveness and ability to + +![](images/5656bbe841dca3b885a7c186417c1316f1e3d9ed1f752b34ffcabae4674ef3af.jpg) + +![](images/9156bb1d437084df267619f6a5d8a538c3c80a72c9065125b6e56e780b50e116.jpg) +Figure 1. (a) Comparisons on original and different geo-based data partitions among previous methods and ours. (b) Visual comparisons with circles on BEV map to represent the learned uncertainties, in which PV instances serve as compensation and provide a detailed understanding of map perception with generalization. + +provide real-time updates [25, 31, 56]. Given images captured by onboard cameras, current map construction relies on transformer scheme [57] to convert perspective-view (PV) image features into bird's-eye-view (BEV) space. Then, a decoder follows to predict the map elements. + +Despite the achievements on widely adopted benchmarks [4, 26, 28], little attention has been paid to investigating the generalization capabilities, which refer to the model's performance on unfamiliar driving scenarios that are different from the training sets. As discussed in [27, 50], in the widely used public datasets [1, 46], there are a large number of geographical overlaps between the training and validation scenes. The models are then more likely to memorize similar driving scenes rather than fully capture their underlying road structures. Variations in driving scenes, such as the contrast between bustling city centers and suburbs with simple road structures, highlight significant shifts in feature distributions and reveal the model's limitations in robustness. Besides, learning-based BEV transformation will inevitably lead to geometric errors and loss of texture details. As shown in Figure 1(a), previous methods [25, 50] exhibit notable performance degradation on region-based and city-based data splits under changing driving scenes. + +To tackle the above issue, we propose the innovative em + +employment of model uncertainty targeted for more generalizable HD map constructions. On the one hand, the uncertainty estimation process involves learning both the statistical mean and variance value, enabling dynamic resampling based on probabilistic distributions. Given the discrepancy in feature distribution across different driving scenes, such a statistical resampling strategy enhances the model's dynamic adaptability in unfamiliar environments. However, the estimated model uncertainty quantifies the confidence in its predictions, which can serve as a reliable indicator for further feature selection and enhancement. As illustrated in Figure 1(b), direct 2D perception can capture more intuitive semantic and angular structural information, which might be missed or incorrectly learned during the implicit PV-to-BEV conversions. Thus, combined with uncertainty-based feature selection, more reliable explicit information can be emphasized as compensation to boost the robustness of BEV-level map construction. + +Based on the above discussion, in this work, we propose UIGenMap, an uncertainty-instructed PV structure injection strategy for more generalizable HD map vectorization. Specifically, a PV detection branch is introduced to provide explicit structural instances. In both PV and BEV branches, we move beyond the vanilla transformer structure by designing the statistically sensitive uncertainty-aware decoder (UA-Decoder), which constructs the probabilistic outputs by uncertainty-aware attention and finer point-level uncertainty head. Leveraging the PV uncertainty output, we introduce the UI2DPrompt to construct learnable PV prompts. These PV prompts are integrated into the BEV decoder by the designed hybrid injection mechanism, which acts as structural compensation for the BEV map prediction. Consequently, reliable structural representation from PV space can be injected to enhance the model's capability, thereby improving its perceptual ability in complex environments. Considering the demands for real-time inference, we also design a lightweight mimic query distillation module to learn from constructed PV prompts. Then during inference, the mimicked query can replace the PV detection flow to reduce extra computational burdens. + +To prioritize the evaluation of generalization capability, we assess our method using geographically disjoint (geo-based) data partitions on the nuScenes [1] and Argoverse2 [46] datasets. Extensive experiments demonstrate that our UIGenMap achieves state-of-the-art performance with higher adaptive ability, especially with $+5.7$ mAP improvement on the region-based partition and $+4.3$ mAP on the more challenging city-based split on nuScenes. The main contributions can be summarized as follows: + +- We propose UIGenMap, an uncertainty-instructed approach to more generalizable HD map vectorization with explicit structure injection. +- The UA-Decoder and UI2DPrompt are proposed to con + +struct structural PV prompts, which are dynamically integrated with BEV perception by hybrid injection. + +- Lightweight MQ-Distillation design supports real-time inference without the additional computational overhead of the PV detection branches. +- Promising results on more challenging geo-based data partitions show that UIGenMap outperforms previous approaches and achieves powerful generalization capability. + +# 2. Related Work + +Vectorized HD Map Construction. Conventional HD map construction approaches [38, 39, 58] rely on SLAM-based annotations, which require tedious offline manual labeling at high costs. Recent studies [6, 11, 19, 20, 29, 37, 49] have explored online approaches with the input of onboard sensor data. In [18, 22, 24, 35, 42], HDMap construction is treated as a segmentation task, relying on post-processing to achieve map vectorization. Later, Liao et al. [25, 26] represent map elements as ordered points and employ transformers [41, 57] after the BEV encoder for auto-regressive detection. Other modeling methods, such as Bezier curves and pivot points, are explored in BeMapNet [37] and PivotNet [6] for geometric representation. Several works [4, 19, 40, 43, 50] are proposed for different temporal fusion approaches. GeMap [54] and MapQR [31] improve mapping performance from the perspectives of geometric relationships and decoder design, respectively. In [5, 28, 51, 52], rasterized maps are utilized for structure compensation. Despite improved performance, these approaches seldom consider the generalization capability of models in unfamiliar driving scenarios. In this work, we incorporate uncertainty-instructed PV detection retracing, aiming for better performance across diverse scenarios. + +PV Detection for 3D Perception. Instances from PV detection can be effective compensations for 3D perceptions [12, 14, 32, 47, 48, 53]. For 3D object detection, BEVFormerv2 [48] utilizes a two-stage BEV detector that integrates the query from the PV head into the BEV space. In SimMoD [53], 2D detection is utilized to construct proposals for 3D fine-grained perception. Far3D [14] combines 2D detection boxes with estimated depth information to improve long-range perception performance. For model generalization, Lu et al. [32] propose a generalizable framework based on perspective debiasing, in which PV consistency learning is employed among different domains. Applying for 3D lanes, Li et al. [17] use 2D priors for lane detection and topology reasoning. Inspired by the above methods, we employ PV structural features for more generalizable HD map vectorization, in which distillations from 2D instances are utilized to ensure inference efficiency. + +Uncertainty in Model. Generally, uncertainty refers to the model's confidence level for its predictions, which can be a statistical representation to indicate the model's under + +![](images/b6bbcd069bbfbe9d38546f9957fe490aa974ee4bdbc101c565d45ed913b89ba3.jpg) +(a) Overview +Figure 2. (a) Overview of Our UIGenMap. For training, the PV branch is introduced with the uncertainty-instructed structural injection, in which the MQ-Distillation is designed to mimic PV structural features. (b) Uncertainty-Aware Decoder Architecture. For each layer, UA-Decoder comprises probabilistic UA-Attention and UA-Head for reliable output. (c) UI2DPrompt Design. We construct PV prompts from PV-detected elements and their corresponding uncertainties, which are integrated with the main branch by hybrid injection. + +![](images/b3d856e728e09cd8b9f9e9daa043ffb7d42a9757d215f705d8449fdb9a96b0b6.jpg) +(b) UA-Decoder + +![](images/ad0eba1419bc33d3c8b2b80bd59af6aa08f0016479f848884af8fecaee102b92.jpg) +(c) UI2DPrompt + +standing of scenes [3, 15, 44]. Uncertainty has gained intensive attention due to the potential for reliability evaluation and performance enhancement [2, 16]. Specifically, the generalization problem with uncertainty modeling is investigated in [21, 36], which expects uncertainty estimation for out-of-distribution generalization. For the transformer structure [8, 34], uncertainty has been explored to guide the training and inference processes for more robust model performance. In the field of autonomous driving, significant uncertainty has been introduced for 3D occupancy [2, 44], optical flow [33], and object tracking [55]. Recently, Gu et al. [7] present uncertainty estimation for the vectorized map, which is targeted for trajectory prediction. In our design, we utilize uncertainty for more generalized HD map construction, which considers the variability in feature distributions and incorporates explicit instance features. + +# 3. UIGenMap + +Given surrounding-view images captured from onboard cameras, we aim to construct more robust vectorized map elements in the BEV space. Each map element comprises a class label $c_{\mathrm{map}}$ and an ordered sequence of points $\mathbf{P}_{\mathrm{map}} = \{(p_x^i,p_y^i)\}_{i = 1}^N$ that represents the structure of the instance. $N$ denotes the total number of points for each map element. + +The architecture of UIGenMap is presented in Figure 2. Based on the main BEV detection, we obtain PV instance from an introduced pre-trained PV branch, which is injected as structural priors for BEV perception. Firstly, we extract + +image features for the PV and BEV branches separately. As in [22], BEV features $\mathrm{F_{bev}}$ are constructed from the interaction between PV features and learnable BEV queries. Then, we design the uncertainty-aware (UA-) decoder from inner-layer UA-Attention and point-level heads to achieve reliable probabilistic outputs. Based on the PV coordinates and uncertainty output, we propose the UI2DPrompt module to construct reliable PV prompts. These prompts are incorporated into BEV features and map instance queries via hybrid injection, where uncertainties are utilized as the basis for feature selection and enhancement. Finally, the UA-Decoder employs enhanced BEV features and instance queries for reliable map predictions. To ensure real-time inference, the lightweight MQ-Distillation module performs query distillation from learned PV prompts to imitate explicit structural features, aiming for efficient computation. + +# 3.1. Uncertainty-Aware Decoder + +We design UA-Decoder for both PV and BEV space. Systematic uncertainty-aware design is employed to improve the statistical understanding of features and outputs to achieve more dynamic self-adjustment for different driving scenarios. To ensure comprehensive perception, each decoder layer comprises UA-Attention at the instance feature level and UA-Head for point-level uncertainty output along with positions, respectively. + +UA-Attention. In conventional deformable attention [57], the weight $\alpha_{i}$ for each sampled feature values from BEV feature $F_{\mathrm{BEV}}$ is determined by linear layer (Linear(·)) on + +![](images/e4226113f26f405ad6d4e9364b9a37a32d401984fdacb846dc0b5bcc270199f6.jpg) +Figure 3. UA-Attention Design. The learned probabilistic weights $\alpha_{i}$ are dynamically resampled from Gaussian distribution with stochastic disturbance, which is multiplied with sampled feature values to update map queries. + +$Q_{\mathrm{map}}^i$ , which can be expressed as $\alpha_{i} = \mathrm{Linear}(Q_{\mathrm{map}}^{i})$ where $Q_{\mathrm{map}}^i$ is the $i$ -th instance query of totally $M$ map elements. Then, the updated part of query $\Delta Q_{\mathrm{map}}^i$ is aggregated by the dot-product between the learned weight $\alpha_{i}$ and sampled feature values from $\mathbf{F}_{\mathrm{bev}}$ . Considering complex driving scenarios, such deterministic weights are inadaptable for challenging situations. More dynamic weights are required to avoid potential ambiguities. Inspired by [8, 34], we explicitly model the attention scores under a Gaussian distribution $\alpha_{i}\sim \mathcal{N}(\mu_{i},\sigma_{i}^{2})$ , as shown in Figure 3. We adopt the re-parameterization strategy to perform the probabilistic attention, which is constructed by the corresponding mean $\mu_{i}$ and variance $\sigma_{i}$ . Such probabilistic parameters are predicted through multilayer perception (MLP(·)). Thus, the construction of dynamic resampled attention weights can be formulated as below: + +$$ +\left\{\mu_ {i}, \sigma_ {i} ^ {2} \right\} = \mathrm {M L P} \left(Q _ {\text {m a p}} ^ {i}\right), \tag {1} +$$ + +$$ +\alpha_ {i} = \mu_ {i} + \sigma_ {i} \epsilon , \epsilon \sim \mathcal {N} (0, 1), \tag {2} +$$ + +where $\epsilon$ is the Gaussian noise conforming to normal distribution. During inference, attention weight $\alpha_{i}$ can be dynamically resampled from learned statistical distribution to suit for driving scenarios with large variations. Then query features can be obtained by dot production with the sampled feature values in order to enable reliable query updates. + +UA-Head with Uncertainty Loss. Based on the robust design of UA-Attention, quantized uncertainty for each lane point output is also essential, which is point-specific at a fine-grained level to explicitly reflect the reliability and uncertainty information within the internal transformer layers. Inspired by [7], we replace the vanilla deterministic output with the UA-Head that predicts coordinates $\mathbf{P}_{\mathrm{map}} = \{(\hat{p}_x^i,\hat{p}_y^i)\}_{i = 1}^N$ along with their corresponding uncertainties $\sigma_{\mathrm{map}} = \{(\sigma_x^i,\sigma_y^i)\}_{i = 1}^N$ for each map element, as shown in Figure 2 (b). These parameters represent the location and scale of the joint Laplace distribution $Laplace(\hat{p}_j^i,\frac{\sigma_j^i}{\sqrt{2}})$ at + +the $j$ -th dimension $(x$ or $y)$ of the $i$ -th point, which refers to a probabilistic representation for stable output. + +For uncertainty-aware training, we combine probabilistic Negative Log-Likelihood (NLL) loss $\mathcal{L}_{\mathrm{nll}}$ with point regression loss $\mathcal{L}_{\mathrm{pts}}$ in [50] to balance uncertainties and coordinate outputs, in which $\mathcal{L}_{\mathrm{nll}}$ can be formulated as follows: + +$$ +\mathcal {L} _ {\mathrm {n l l}} = \lambda_ {\mathrm {n l l}} \frac {1}{N} \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {2} \left(\log \left(2 \sigma_ {j} ^ {i}\right) + \frac {\left| p _ {j} ^ {i} - \hat {p} _ {j} ^ {i} \right|}{\sigma_ {j} ^ {i}}\right), \tag {3} +$$ + +with $\lambda_{\mathrm{nl}}$ as the loss weight setting. + +As illustrated in Figure 2 (a), UA-Decoder is used for the construction of the PV branch and the BEV map. To provide more robust structural features in PV branch, PV coordinates $\mathrm{P_{pv}} = \{x_{pv}^i,y_{pv}^i\}_{i = 1}^N$ and the uncertainty parameters $\sigma_{\mathrm{pv}} = \{\sigma_{\mathrm{pv,x}}^{i},\sigma_{\mathrm{pv,y}}^{i}\}_{i = 1}^{N}$ are constructed by the UA-Decoder structure, which is later utilized in our UI2DPrompt design. Thus, the systematic combination of uncertainty-aware design provides dynamic sampling during the learning process, along with established uncertainty, to construct a reliable basis for confidence scoring. + +# 3.2. UI2DPrompt + +PV branches contain abundant structural and geometric details. To mitigate information loss and misinterpretation during query-based PV-to-BEV transformation, reliable structural PV instances can be injected as compensation for BEV prediction. This section focuses on constructing PV prompts based on uncertainty-aware PV predictions, which are dynamically integrated with the map decoder for BEV map element detection. + +Prompt Construction. Given the learned PV instance coordinates and the corresponding uncertainty parameters $\{\mathrm{P_{pv}},\sigma_{\mathrm{pv}}\}$ , PV prompts are constructed to provide dynamic 2D priors. Firstly, we employ the estimated classification score $c_{\mathrm{map}}$ for PV candidates selection. When $c_{\mathrm{map}} > c_{\mathrm{thr}}$ , PV instances are chosen as candidates for reliable localized information. Considering the sparse depth information and the characteristics of static map elements on the ground, we utilize Inverse Perspective Mapping (IPM) to transform $\mathrm{P_{pv}}$ from the PV image coordinate to the BEV coordinate system $\mathrm{P}_{\mathrm{trans}} = \{p_x^{\prime},p_y^{\prime}\}$ , which makes a flat grounding assumption and utilizes the affine transformation between the PV and BEV coordinates. Given the corresponding intrinsic parameters $\mathbf{K}_{\mathrm{cam}}$ and extrinsic transformation matrix $\mathrm{T}_{\mathrm{ego2cam}}$ , the projection $\mathrm{P_{pv}}$ can be defined as: + +$$ +\mathbf {P} _ {\text {t r a n s}} = \operatorname {P r o j} _ {2 D} \left(\mathrm {T} _ {\text {e g o 2 c a m}} ^ {- 1} \cdot \left[ \begin{array}{c} \mathrm {K} _ {\text {c a m}} ^ {- 1} \cdot \mathrm {P} _ {\mathrm {p v}} \\ 1 \end{array} \right]\right), \tag {4} +$$ + +where $\operatorname{Proj}_{2\mathrm{D}}(\cdot)$ denotes the conversion from homogeneous coordinates into 2D plane coordinates. + +To ensure reliable enhanced features, the corresponding uncertainty parameters $\sigma_{\mathrm{pv}}$ cooperate as weights and feature + +embeddings. We encode the transformed coordinates $\mathbf{P}_{\mathrm{trans}}$ and uncertainty parameters $\sigma_{\mathrm{pv}}$ by $\phi_p(\cdot)$ and $\phi_{\sigma}(\cdot)$ , respectively. Point-specific embeddings $\{e_{pv}^i\}_{i = 1}^N$ are constructed by concatenation: + +$$ +\boldsymbol {e} _ {\mathrm {p v}} ^ {i} = \operatorname {C o n c a t} \left(\left[ \phi_ {p} \left(\mathrm {P} _ {\text {t r a n s}}\right), \phi_ {\sigma} \left(\sigma_ {\mathrm {p v}}\right) \right]\right). \tag {5} +$$ + +Thus, the structural PV output coordinates and uncertainty information are dynamically encoded. + +Based on the above, PV uncertainties can serve as the quantized weights for dynamic feature strengthening and adjustment. Reliable confidence indicator is able to be formulated as the weight $\{\omega_{\mathrm{pv}}^i\}_{i = 1}^N$ concerning each point-specific prompt. Then, the enhanced PV prompt $\tilde{\pmb{e}}_{\mathrm{pv}}^{i}$ can be constructed by: + +$$ +\omega_ {\mathrm {p v}} ^ {i} = \exp \left(\frac {\left(\| \sigma_ {\mathrm {p v}} ^ {i} \| _ {2}\right) ^ {- 1}}{\sum_ {i = 1} ^ {N} \left(\| \sigma_ {\mathrm {p v}} ^ {i} \| _ {2}\right) ^ {- 1}}\right), \quad \tilde {\boldsymbol {e}} _ {\mathrm {p v}} ^ {i} = \omega_ {\mathrm {p v}} ^ {i} \boldsymbol {e} _ {\mathrm {p v}} ^ {i} + \boldsymbol {e} _ {\mathrm {m}} ^ {i}, \tag {6} +$$ + +where $\exp (\cdot)$ represents the exponential amplification of the normalized inverse uncertainty value. $\{e_{\mathrm{m}}^{i}\}_{i = 1}^{N}$ is the predefined mimic query, which is utilized for MQ-Distillation. This approach reduces the introduction of PV errors, ensuring that injecting PV instances amplifies reliable PV information while minimizing uncertainty-induced errors. + +Hybrid Injection. Constructed from PV instances with uncertainty output, PV prompts $\{\tilde{e}_{\mathrm{pv}}^i\}_{i = 1}^N$ are incorporated as compensation for the construction of the BEV map. These prompts are injected into the BEV feature $\mathbf{F}_{\mathrm{bev}}$ and the map queries $Q_{\mathrm{map}}^i$ for the map decoder by the designed hybrid injection mechanism, which includes both prompt-to-BEV (P2BEV) and prompt-to-query (P2Q) attention. + +As shown in Figure 4, point-level PV prompts $\{\tilde{e}_{\mathrm{pv}}^i\}_{i = 1}^N$ are integrated with $\mathbf{F}_{\mathrm{bev}}$ through cross-attention of P2BEV. This step enriches BEV features by integrating detailed, point-specific information from PV prompts, enhancing instance-specific detail in the BEV representation: + +$$ +\mathrm {F} _ {\mathrm {b e v}} ^ {\prime} = \operatorname {A t t n} \left(W _ {q} \mathrm {F} _ {\mathrm {b e v}}, W _ {k} \tilde {\boldsymbol {e}} _ {\mathrm {p v}} ^ {i}, W _ {v} \tilde {\boldsymbol {e}} _ {\mathrm {p v}} ^ {i}\right), \tag {7} +$$ + +where $W_{\{q,k,v\}}$ are the corresponding projection matrix. Then, we encode point-level prompts into instance-level representations through a linear block $\phi_e(\cdot)$ , which aggregates point-specific information to create a higher-level representation for subsequent processing. These instance-level prompts $\tilde{e}_{\mathrm{inst}}$ are used to integrate with $Q_{\mathrm{map}}$ , which inject reliable priors for the initialized queries and enhance the overall map construction: + +$$ +\tilde {\boldsymbol {e}} _ {\text {i n s t}} = \phi_ {e} \left(\left\{\tilde {\boldsymbol {e}} _ {\mathrm {p v}} ^ {i} \right\} _ {i = 1} ^ {N}\right), \tag {8} +$$ + +$$ +Q _ {\text {m a p}} ^ {\prime} = \operatorname {A t t n} \left(W _ {q} Q _ {\text {m a p}}, W _ {k} \tilde {\mathbf {e}} _ {\text {i n s t}}, W _ {v} \tilde {\mathbf {e}} _ {\text {i n s t}}\right). \tag {9} +$$ + +By injecting PV priors into both BEV and decoder queries, this hybrid mechanism improves the reliability of BEV perception and reduces transition errors. Explicit structural + +![](images/12ff6682f09d2c79c4a074e8d7670d60298783929293de060d1f9cd29885c73e.jpg) +Figure 4. Design of Hybrid Injection. P2BEV-Attention and P2Q-Attention are designed for PV prompts to integrate with BEV features and map instance queries as structural compensation. + +features enhance the generalization capability of the learned model and perform robustly in various driving environments, ensuring better accuracy and consistency. + +# 3.3. Mimic Query Distillation + +To address the extra computation introduced by the PV branch, we propose the lightweight MQ-Distillation to distill from PV prompt construction, where the mimic query can serve as an alternative to achieve real-time inference. + +In MQ-Distillation, we define a set of mimic queries $\{e_m^i\}_{i = 1}^N$ , along with an MLP-based query learner $h(\cdot)$ which is designed to capture and distill the structural features of the PV prompt construction process $\tilde{e}_{\mathrm{pv}}^i$ . As shown in Equation 6, during training, the predefined mimic queries $e_{m}^{i}$ and the constructed PV prompts $\tilde{e}_{\mathrm{pv}}^i$ are jointly combined by addition, which are sent to the hybrid injection for further operations. + +To obtain sufficient structural features and achieve progressive alignment with PV prompts, we design the loss function that guides mimic queries for distillation. We use Mean Squared Error (MSE) as the distillation loss to measure the difference between the constructed PV prompts and the mimicked queries, which can be formulated as follows: + +$$ +\mathcal {L} _ {\text {d i s t i l l}} = \lambda_ {\text {d i s t i l l}} \frac {1}{N} \sum_ {i = 1} ^ {N} \left(\tilde {\boldsymbol {e}} _ {\mathrm {p v}} ^ {i} - h \left(\boldsymbol {e} _ {m} ^ {i}\right)\right) ^ {2}, \tag {10} +$$ + +where $\lambda_{\mathrm{distill}}$ is a weight coefficient that balances the impact of the distillation loss on the overall training process. Such loss function guides the mimic queries to accurately replicate the PV prompts generated by the PV branch, making them reliable substitutes for inference. + +To this end, only the mimic queries can be utilized during inference, which eliminates the requirement for PV prompt computation. This approach maintains real-time inference capability by efficiently mimicking PV structural features with minimal computational cost, enabling streamlined yet robust map construction. + +
MethodBackboneEpochsRegion-BasedCity-BasedFPS
APpedAPdivAPboumAPAPpedAPdivAPboumAP
VectorMapNet* [ICML'23]R5012015.817.021.218.0-----
MapTR* [ICLR'23]R50246.420.735.520.911.112.921.115.015.8
MGMap [CVPR'24]R50249.427.038.124.810.917.824.717.811.5
MapQR [ECCV'24]R50248.425.739.924.710.215.825.517.212.1
GeMap [ECCV'24]R502411.826.643.627.312.916.326.718.611.6
MapTRv2* [IJCV'24]R502414.728.743.328.918.918.627.921.812.9
StreamMapNet* [WACV'24]R502432.229.340.834.118.716.422.719.313.3
UIGenMap-d (Ours)R502440.330.846.839.3 (+5.2)23.119.625.422.7 (+3.4)12.2
UIGenMap (Ours)R502440.331.847.239.8 (+5.7)24.120.726.123.6 (+4.3)8.2
StreamMapNet [WACV'24]SwinT2432.527.441.733.921.819.725.822.49.3
UIGenMap-d (Ours)SwinT2442.631.947.240.6 (+6.7)24.721.228.824.9 (+2.5)8.5
+ +Table 1. Quantitative comparisons on nuScenes validation dataset, utilizing region-based and city-based dataset partitions. Methods with “*” denote that data results are obtained from corresponding papers [27, 50], and others are reproduced by official public codes. “R50” and “SwinT” denotes the image backbone of ResNet50 [9] and Swin-Tiny [30]. “UIGenMap-d” represents the utilization of mimic query from MQ-Distillation. FPS is measured on a single RTX A40 GPU. + +# 3.4. Training and Inference + +During the training process, combined with $\mathcal{L}_{\mathrm{nl}}$ for uncertainty output and $\mathcal{L}_{\mathrm{distill}}$ for mimic query distillation, the overall loss setting $\mathcal{L}_{\mathrm{map}}$ can be represented as: + +$$ +\mathcal {L} _ {\text {m a p}} = \lambda_ {1} \mathcal {L} _ {\text {p t s}} + \lambda_ {2} \mathcal {L} _ {\text {c l s}} + \mathcal {L} _ {\text {n l l}} + \mathcal {L} _ {\text {d i s t i l l}}, \tag {11} +$$ + +where $\mathcal{L}_{\mathrm{pts}}$ denotes the point-to-point Manhattan distance between the regressed points and the ground-truth. $L_{\mathrm{cls}}$ represents the focal loss for map classification. $\lambda_{1}$ and $\lambda_{2}$ are the corresponding loss weights for point regression and map classification, respectively. + +For inference, the well-learned mimic query can only be employed as a substitute for PV prompts, avoiding the computational burden for the PV detection branch. Thus, our approach ensures real-time inference without introducing unnecessary computational overhead, while preserving effective performance. Additionally, the model's uncertainty learning enables dynamic sampling during inference, enhancing the adaptability to diverse driving environments. + +# 4. Experiments + +# 4.1. Datasets and Metrics + +Datasets. We conduct extensive experiments on widely used autonomous driving datasets, including nuScenes [1] and Argoverse2 [46]. nuScenes dataset comprises 1,000 driving scenes from Boston and Singapore. Each scene lasts approximately 20 seconds with a $2\mathrm{Hz}$ sampling rate. For each sample, there are six surrounding images and 32-beam LiDAR point clouds. Argoverse2 contains 1,000 scenes from 6 cities, with each scene capturing 15 seconds of data at $20\mathrm{Hz}$ from 7 cameras and $10\mathrm{Hz}$ LiDAR sweeps. In our experiments, the sampling frequency for Argoverse2 is unified into $2\mathrm{Hz}$ as in StreamMapNet [50]. + +Geo-based Data Partitions. Considering the overlap of widely used training and validation partitions, we adopt + +geographically disjoint (geo-based) dataset partitions as benchmarks to evaluate the generalization capability of the presented model. Experiments are conducted on both region-based and city-based data partitions. The region-based partition is adopted from [50], where the driving locations geographically disjoin the training and validation sets. In the city-based data partition, we apply city-wise splits in [27], where the training and validation data are from different cities with a larger distribution change. On nuScenes, Boston and Onenorth are used for training, while Queenstown and Holland Village are used for validation. For Ar-goverse2, Miami and Pittsburgh are used in training, and the remains are employed for validation. For different dataset partitions, models are separately trained on that split's training set and evaluated on their corresponding validation set. + +Evaluation Metrics. In line with the existing works [26, 50], we conduct experiments on three static map elements, namely lane dividers, pedestrian crossings, and road boundaries. Perception range is set to $[-15.0m, 15.0m]$ for X-axis and $[-30.0m, 30.0m]$ for Y-axis. We use average precision as the evaluation metric, which calculates the Chamfer distance between predictions and ground truths. The mean Average Precision (mAP) is obtained by averaging AP across the distinct thresholds: $\{0.5m, 1m, 1.5m\}$ . + +# 4.2. Implementation Details + +For fair comparisons, we employ ResNet50 [9] as the backbone. As in [50], the size of BEV features is set to $50 \times 100$ . The number of map queries and query dimensions are set to 100 and 512. We use AdamW optimizer with a learning rate of $4e^{-4}$ . Training epochs are limited to prevent potential overfitting. All models are trained on 4 NVIDIA Tesla V100 GPUs with a batch size of 4. The weighted factors $\lambda_{1}$ and $\lambda_{2}$ are set to 50.0 and 5.0, separately. $\lambda_{\mathrm{nil}}$ and $\lambda_{\mathrm{distill}}$ are set to 0.05 and 10.0. We adopt a pre-trained PV detection branch for our UIGenMap training process. More details + +
MethodBackboneEpochsRegion-BasedCity-BasedFPS
\(AP_{ped}\)\(AP_{div}\)\(AP_{bou}\)mAP\(AP_{ped}\)\(AP_{div}\)\(AP_{bou}\)mAP
\(VectorMapNet^{*}\)[ICML'23]R5012035.634.937.836.1-----
\(MapTR^{*}\)[ICLR'23]R503048.150.455.051.134.741.734.737.915.8
\(GeMap\)[ECCV'24]R503052.559.258.256.634.541.140.438.711.6
\(MapTRv2^{*}\)[JCV'24]R503053.358.758.856.937.442.241.940.512.9
\(MapQR\)[ECCV'24]R503054.258.060.457.533.841.541.638.912.1
\(StreamMapNet^{*}\)[WACV'24]R503057.955.761.358.340.243.443.342.313.3
\(UIGenMap-d\)(Ours)R503058.358.462.459.7 (+1.4)42.944.444.744.0 (+1.7)12.2
\(UIGenMap\)(Ours)R503059.858.762.760.4 (+2.1)43.745.745.144.8 (+2.5)8.2
+ +Table 2. Quantitative comparisons on Argoverse2 validation datasets, utilizing region-based and city-based dataset partitions. Methods with “*” denote that data results are obtained from corresponding papers [27, 50], and others are reproduced by official public codes. Models are trained for 30 epochs and follow the same setting. FPS is measured on a single RTX A40 GPU. + +
DatasetMethodEpochs\(AP_{ped}\)\(AP_{div}\)\(AP_{bou}\)mAP
nuScenesMapTRV2 [26]2459.862.462.461.5
StreamMapNet [50]2460.461.958.960.4
SQD-MapNet [43]2463.065.563.363.9
UIGenMap(Ours)2464.265.464.964.8
Argoverse2StreamMapNet [50]3062.959.563.061.5
SQD-MapNet [43]3064.960.264.963.3
UIGenMap(Ours)3064.560.866.163.8
+ +Table 3. Performance on nuScenes and Argoverse2 datasets with the original dataset partition at ${30}\mathrm{\;m}$ range. + +
bkbUI2DPromptUA-DecoderAPpedAPdivAPboumAP
32.229.340.834.1
✓39.928.643.837.4
✓36.131.344.037.1
✓✓41.629.744.838.7
✓✓41.429.847.439.5
✓✓39.230.146.238.5
✓✓✓40.331.847.239.8
+ +Table 4. Ablations on the impact of component designs, 'bkb' means the shared image backbone from PV branches. + +are provided in the supplementary materials. + +# 4.3. Main Results + +Results on nuScenes. We compare the generalization capability with previous methods [25, 26, 28, 31, 50, 54] on geospatial data partitions, in which the image size is set to $480 \times 800$ . As shown in Table 1, our UIGenMap outperforms previous methods at a large margin. For the ResNet50 backbone, UIGenMap achieves $5.7\mathrm{mAP}$ improvement on the region-based partition, and $4.3\mathrm{mAP}$ gain is obtained on the city-based splits. For more efficient design, UIGenMapd also archives $39.3\mathrm{mAP}$ with a faster inference speed. Experiments on the Swin-T [30] backbone show significant performance with $40.6\mathrm{mAP}$ achievement. Furthermore, we show some visual examples in Figure 5, indicating more intuitive achievements for UIGenMap. + +Results on Argoverse2. More experiments are conducted on Argoverse2 dataset. The image size is $608 \times 608$ , and the sampling rate is $2\mathrm{Hz}$ . For a fair comparison, the models are trained for 30 epochs. As shown in Table 2, UIGen + +Map achieves the best results of $60.4\mathrm{mAP}$ on region-based data partitions and $43.4\mathrm{mAP}$ on city-based splits, which performs better than previous approaches. Upon further investigation, calibration errors and dataset distribution also affect model performance. More visualization can be found in our supplementary materials. + +Results on Original Data Partitions. For thorough evaluation, we conducted experiments on the original data partitions with similar driving scenarios, which are widely used in previous studies. As shown in Table 3, we compare our methods with others utilizing the same baseline. UIGenMap still outperforms the existing approaches, which outperforms StreamMapNet with $+4.4$ mAP and $+2.3$ mAP separately, validating the effectiveness of our approach. + +# 4.4. Ablation Studies + +In this section, ablations are conducted to verify the effectiveness of our designs. We perform experiments on nuScenes with region-based partition using ResNet50 backbone. All models are trained for 24 epochs. More ablations are provided in the supplementary materials. + +Contributions of Main Components. Table 4 demonstrates the impact of each component in our design. Based on the pre-trained PV backbone, we design the UA-Decoder for more reliable uncertainty output and UI2DPrompt to dynamically inject PV structural information, which achieves $+2.7$ mAP and $+2.4$ mAP improvement, respectively. The combination of these designs gains $+4.0$ mAP total. The ablations are conducted on UA-Decoder at BEV space. Besides, the pre-trained PV backbone further emphasizes the effectiveness to $39.8$ mAP for better scene understanding. + +Design of UA-Decoder. We design the UA-Attention for probabilistic attention and the UA-Head to construct the uncertainty output from the instance and finer point levels. Table 5 shows the ablation without other modules. Our uncertainty-aware design supports dynamic adaptation to achieve more robust performance, which achieves $+3.0$ mAP improvements based on systematic combination. + +Design for UI2DPrompt. UI2DPrompt consists of prompt construction and a hybrid injection module. Table 6 shows + +![](images/55491a0f775a128dc58c15508ee62993ff71f113d1dc12b8ded035fe71d74f4b.jpg) +Surrounding Views with PV Detection + +![](images/1353fbd0974062a11dcd003a74e21b249732bba0d5f9e1fe3839ca7360f4af46.jpg) +Baseline +GT +Figure 5. Qualitative results on the nuScenes dataset with the region-based data partition. Uncertainty outputs are represented as circles. + +
UA-Att.UA-HeadAPpedAPdivAPboumAP
32.229.340.834.1
✓31.730.442.935.0
✓34.830.543.436.3
✓✓36.131.344.037.1
+ +Table 5. Ablations on UA-Decoder without other modules. UAAtt and UA-Head represent the utilization of uncertainty-aware attention and output head respectively. + +
Coords. Trans.Unc. UtilizationmAP
//37.6
IPMembedding only39.3
IPMweight & embedding39.8
3D Trans.weight & embedding38.7
+ +the strategy selection for constructing PV prompts. Compared to a depth-based 3D transition, IPM is more suitable for coordinating the transformation based on the assumption of a flattened road surface. When uncertainty is applied both as dynamic weighting and as learning-based embeddings, performance is further enhanced by reliable PV prompts, achieving an improvement of $+2.2\mathrm{mAP}$ . + +Table 7 demonstrates the impact of the hybrid injection design, with a total performance of $39.8\mathrm{mAP}$ for combining prompt-to-BEV and prompt-to-query attention. This result confirms the optimal utilization of reliable structural information, effectively compensating for information loss and errors introduced in the basic learning process. + +Ablations on MQ-Distillation. To verify the effectiveness of the distillation module, we conduct ablations on the hyperparameter of distillation loss weight $\lambda_{\mathrm{distill}}$ , as this affects the imitation learning of PV structural features. As + +Table 6. Ablations on the construction of PV prompts. "3D Trans." means the depth-based transformation. The first row with $\frac{1}{r}$ indicates direct PV coordinate embedding. + +
P2BEVP2QAPpedAPdivAPboumAP
39.330.445.538.4
✓39.931.146.039.2
✓39.630.746.138.8
✓✓40.331.847.239.8
+ +Table 7. Ablations on hybrid injection. P2BEV and P2Q mean the cross-attention for $\mathrm{F}_{\mathrm{bev}}$ and $Q_{\mathrm{map}}$ separately. + +
λdistill571015/
mAP37.538.239.338.839.8
+ +Table 8. Ablations on $\lambda_{\mathrm{distill}}$ for MQ-Distillation Design. The last column refers to the use of the best teacher of PV prompts. + +illustrated in Table 8, the best performance is achieved by setting the weight to 10.0, which refers to the best learning process of the mimic query for explicit structural features. + +# 5. Conclusion and Future Work + +In this paper, we propose UIGenMap, an uncertainty-instructed structure injection approach to achieve a more generalized HD map construction. Specifically, we introduce the PV detection branch to obtain explicit structural features, in which the UA-Decoder produces adaptive probabilistic outputs. The UI2DPrompt module is proposed to construct reliable PV prompts and achieve dynamic integration through the hybrid injection designed. For real-time inference, MQ-Distillation is designed to mimic the structural prompts of PV. Comprehensive experiments on the challenging geo-based partitions of nuScenes and Argoverse2 datasets demonstrate the generalization of our approach. + +For limitations, current methods are designed and evaluated primarily on specific datasets. In the future, developing larger and more diverse datasets and benchmarks across various locations remains a crucial area for exploration. + +# Acknowledgments + +This work is supported by National Natural Science Foundation of China under Grant No.62376244. It is also supported by Information Technology Center and State Key Lab of CAD&CG, Zhejiang University. + +# References + +[1] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11621-11631, 2020. 1, 2, 6 +[2] Anh-Quan Cao, Angela Dai, and Raoul de Charette. Pasco: Urban 3d panoptic scene completion with uncertainty awareness. In CVPR, pages 14554–14564, 2024. 3 +[3] Jie Chang, Zhonghao Lan, Changmao Cheng, and Yichen Wei. Data uncertainty learning in face recognition. In CVPR, pages 5710-5719, 2020. 3 +[4] Jiacheng Chen, Yuefan Wu, Jiaqi Tan, Hang Ma, and Yasutaka Furukawa. Maptracker: Tracking with strided memory fusion for consistent vector hd mapping. In ECCV, pages 90-107. Springer, 2025. 1, 2 +[5] Sehwan Choi, Jungho Kim, Hongjae Shin, and Jun Won Choi. Mask2map: Vectorized hd map construction using bird's eye view segmentation masks. In ECCV, 2024. 2 +[6] Wenjie Ding, Limeng Qiao, Xi Qiu, and Chi Zhang. Pivotnet: Vectorized pivot learning for end-to-end hd map construction. In ICCV, pages 3672-3682, 2023. 2 +[7] Xunjiang Gu, Guanyu Song, Igor Gilitschenski, Marco Pavone, and Boris Ivanovic. Producing and leveraging online map uncertainty in trajectory prediction. In CVPR, pages 14521-14530, 2024. 3, 4 +[8] Hongji Guo, Hanjing Wang, and Qiang Ji. Uncertainty-guided probabilistic transformer for complex action recognition. In CVPR, pages 20052-20061, 2022. 3, 4 +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6 +[10] Yuzhe He, Shuang Liang, Xiaofei Rui, Chengying Cai, and Guowei Wan. Egovm: Achieving precise eg-localization using lightweight vectorized maps. arXiv preprint arXiv:2307.08991, 2023. 1 +[11] Haotian Hu, Fanyi Wang, Yaonong Wang, Laifeng Hu, Jingwei Xu, and Zhiwang Zhang. Admap: Anti-disturbance framework for reconstructing online vectorized hd map. arXiv preprint arXiv:2401.13172, 2024. 2 +[12] Haoxuanye Ji, Pengpeng Liang, and Erkang Cheng. Enhancing 3d object detection with 2d detection-guided query anchors. In CVPR, pages 21178-21187, 2024. 2 +[13] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, pages 8340-8350, 2023. 1 +[14] Xiaohui Jiang, Shuai Lin Li, Yingfei Liu, Shihao Wang, Fan Jia, Tiancai Wang, Lijin Han, and Xiangyu Zhang. Far3d: + +Expanding the horizon for surround-view 3d object detection. In AAAI, pages 2561-2569, 2024. 2 +[15] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? In NeuIPS, 2017. 3 +[16] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. In NeuIPS, 2017. 3 +[17] Han Li, Zehao Huang, Zitian Wang, Wenge Rong, Naiyan Wang, and Si Liu. Enhancing 3d lane detection and topology reasoning with 2d lane priors. arXiv preprint arXiv:2406.03105, 2024. 2 +[18] Qi Li, Yue Wang, Yilun Wang, and Hang Zhao. Hdmapnet: An online hd map construction and evaluation framework. In ICRA, pages 4628-4634. IEEE, 2022. 2 +[19] Siyu Li, Jiacheng Lin, Hao Shi, Jiaming Zhang, Song Wang, You Yao, Zhiyong Li, and Kailun Yang. Dtlmapper: Dual temporal consistent learning for vectorized hd map construction. arXiv preprint arXiv:2405.05518, 2024. 2 +[20] Siyu Li, Kailun Yang, Hao Shi, Song Wang, You Yao, and Zhiyong Li. Genmapping: Unleashing the potential of inverse perspective mapping for robust online hd map construction. arXiv preprint arXiv:2409.08688, 2024. 2 +[21] Xiaotong Li, Yongxing Dai, Yixiao Ge, Jun Liu, Ying Shan, and Lingyu Duan. Uncertainty modeling for out-of-distribution generalization. In ICLR, 2022. 3 +[22] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In ECCV, pages 1-18. Springer, 2022. 2, 3 +[23] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In ICCV, pages 14864-14873, 2024. 1 +[24] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. In NeuIPS, pages 10421-10434, 2022. 2 +[25] Benchcheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. Maptr: Structured modeling and learning for online vectorized hd map construction. In ICLR, 2022. 1, 2, 7 +[26] Benchcheng Liao, Shaoyu Chen, Yunchi Zhang, Bo Jiang, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Maptrv2: An end-to-end framework for online vectorized hd map construction. IJCV, pages 1-23, 2024. 1, 2, 6, 7 +[27] Adam Lilja, Junsheng Fu, Erik Stenberg, and Lars Hammarstrand. Localization is all you evaluate: Data leakage in online mapping datasets and how to fix it. In CVPR, pages 22150-22159, 2024. 1, 6, 7 +[28] Xiaolu Liu, Song Wang, Wentong Li, Ruizi Yang, Junbo Chen, and Jianke Zhu. Mgmt: Mask-guided learning for online vectorized hd map construction. In CVPR, pages 14812-14821, 2024. 1, 2, 7 +[29] Yicheng Liu, Tianyuan Yuan, Yue Wang, Yilun Wang, and Hang Zhao. Vectormapnet: End-to-end vectorized hd map learning. In ICML, pages 22352-22369. PMLR, 2023. 2 + +[30] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 6, 7 +[31] Zihao Liu, Xiaoyu Zhang, Guangwei Liu, Ji Zhao, and Ningyi Xu. Leveraging enhanced queries of point sets for vectorized map construction. arXiv preprint arXiv:2402.17430, 2024. 1, 2, 7 +[32] Hao Lu, Yunpeng Zhang, Qing Lian, Dalong Du, and Yingcong Chen. Towards generalizable multi-camera 3d object detection via perspective debiasing. arXiv preprint arXiv:2310.11346, 2023. 2 +[33] Ao Luo, Fan Yang, Xin Li, Lang Nie, Chunyu Lin, Haoqiang Fan, and Shuaicheng Liu. Gaflow: Incorporating gaussian attention into optical flow. In ICCV, pages 9642–9651, 2023. 3 +[34] Jiahuan Pei, Cheng Wang, and György Szarvas. Transformer uncertainty estimation with hierarchical stochastic attention. In AAAI, pages 11147-11155, 2022. 3, 4 +[35] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, pages 194-210. Springer, 2020. 2 +[36] Fengchun Qiao and Xi Peng. Uncertainty-guided model generalization to unseen domains. In CVPR, pages 6790-6800, 2021. 3 +[37] Limeng Qiao, Wenjie Ding, Xi Qiu, and Chi Zhang. End-to-end vectorized hd-map construction with piecewiseBezier curve. In CVPR, pages 13218-13228, 2023. 2 +[38] Tixiao Shan and Brendan Englot. Lego-loam: Lightweight and ground-optimized lidar odometry and mapping on variable terrain. In IROS, pages 4758–4765. IEEE, 2018. 2 +[39] Tixiao Shan, Brendan Englot, Drew Meyers, Wei Wang, Carlo Ratti, and Daniela Rus. Lio-sam: Tightly-coupled lidar inertial odometry via smoothing and mapping. In IROS, pages 5135-5142. IEEE, 2020. 2 +[40] Jingyu Song, Xudong Chen, Liupei Lu, Jie Li, and Katherine A Skinner. Memfusionmap: Working memory fusion for online vectorized hd map construction. arXiv preprint arXiv:2409.18737, 2024. 2 +[41] A Vaswani. Attention is all you need. In NeuIPS, 2017. 2 +[42] Song Wang, Wentong Li, Wenyu Liu, Xiaolu Liu, and Jianke Zhu. Lidar2map: In defense of lidar-based semantic map construction using online camera distillation. In CVPR, pages 5186-5195, 2023. 2 +[43] Shuo Wang, Fan Jia, Yingfei Liu, Yucheng Zhao, Zehui Chen, Tiancai Wang, Chi Zhang, Xiangyu Zhang, and Feng Zhao. Stream query denoising for vectorized hd map construction. In ECCV, 2024. 2, 7 +[44] Song Wang, Zhongdao Wang, Jiawei Yu, Wentong Li, Bailan Feng, Junbo Chen, and Jianke Zhu. Relioco: Towards reliable semantic occupancy prediction via uncertainty learning. arXiv preprint arXiv:2409.18026, 2024. 3 +[45] Song Wang, Jiawei Yu, Wentong Li, Wenyu Liu, Xiaolu Liu, Junbo Chen, and Jianke Zhu. Not all voxels are equal: Hardness-aware semantic scene completion with self-distillation. In CVPR, pages 14792-14801, 2024. 1 + +[46] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In NeurIPS, 2021. 1, 2, 6 +[47] Qiangqiang Wu, Yan Xia, Jia Wan, and Antoni B Chan. Boosting 3d single object tracking with 2d matching distillation and 3d pre-training. In ECCV, 2024. 2 +[48] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In CVPR, pages 17830-17839, 2023. 2 +[49] Jing Yang, Minyue Jiang, Sen Yang, Xiao Tan, Yingying Li, Errui Ding, Hanli Wang, and Jingdong Wang. Mgmtapnet: Multi-granularity representation learning for end-to-end vectorized hd map construction. arXiv preprint arXiv:2410.07733, 2024. 2 +[50] Tianyuan Yuan, Yicheng Liu, Yue Wang, Yilun Wang, and Hang Zhao. Streammapnet: Streaming mapping network for vectorized online hd map construction. In WACV, pages 7356-7365, 2024. 1, 2, 4, 6, 7 +[51] Gongjie Zhang, Jiahao Lin, Shuang Wu, Zhipeng Luo, Yang Xue, Shijian Lu, Zuoguan Wang, et al. Online map vectorization for autonomous driving: A rasterization perspective. In NeuIPS, 2024. 2 +[52] Xiaoyu Zhang, Guangwei Liu, Zihao Liu, Ningyi Xu, Yunhui Liu, and Ji Zhao. Enhancing vectorized map perception with historical rasterized maps. In ECCV, 2024. 2 +[53] Yunpeng Zhang, Wenzhao Zheng, Zheng Zhu, Guan Huang, Jiwen Lu, and Jie Zhou. A simple baseline for multi-camera 3d object detection. In AAAI, pages 3507-3515, 2023. 2 +[54] Zhixin Zhang, Yiyuan Zhang, Xiaohan Ding, Fusheng Jin, and Xiangyu Yue. Online vectorized hd map construction using geometry. arXiv preprint arXiv:2312.03341, 2023. 2, 7 +[55] Lijun Zhou, Tao Tang, Pengkun Hao, Zihang He, Kalok Ho, Shuo Gu, Wenbo Hou, Zhihui Hao, Haiyang Sun, Kun Zhan, et al. Ua-track: Uncertainty-aware end-to-end 3d multi-object tracking. arXiv preprint arXiv:2406.02147, 2024. 3 +[56] Yi Zhou, Hui Zhang, Jiaqian Yu, Yifan Yang, Sangil Jung, Seung-In Park, and ByungIn Yoo. Himap: Hybrid representation learning for end-to-end vectorized hd map construction. In CVPR, pages 15396-15406, 2024. 1 +[57] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 1, 2, 3 +[58] Yanjin Zhu, Xin Zheng, and Jianke Zhu. Mesh-loam: Real-time mesh-based lidar odometry and mapping. TIV, 2024. 2 \ No newline at end of file diff --git a/2025/Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction/images.zip b/2025/Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..eb8d84f14cdcd9a89444400897caa0e0b2e10762 --- /dev/null +++ b/2025/Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5acc9c4fe32ca72cbecde80711c52233f9bd806134c6b29f6c40de9ec1ca6691 +size 631835 diff --git a/2025/Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction/layout.json b/2025/Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..82554952088d673f9113071caf7a14a3ffe9f3b7 --- /dev/null +++ b/2025/Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction/layout.json @@ -0,0 +1,9793 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 118, + 103, + 493, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 103, + 493, + 139 + ], + "spans": [ + { + "bbox": [ + 118, + 103, + 493, + 139 + ], + "type": "text", + "content": "Uncertainty-Instructed Structure Injection for Generalizable HD Map Construction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 103, + 160, + 510, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 160, + 510, + 190 + ], + "spans": [ + { + "bbox": [ + 103, + 160, + 510, + 190 + ], + "type": "text", + "content": "Xiaolu Liu" + }, + { + "bbox": [ + 103, + 160, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 103, + 160, + 510, + 190 + ], + "type": "text", + "content": ", Ruizi Yang" + }, + { + "bbox": [ + 103, + 160, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 103, + 160, + 510, + 190 + ], + "type": "text", + "content": ", Song Wang" + }, + { + "bbox": [ + 103, + 160, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 103, + 160, + 510, + 190 + ], + "type": "text", + "content": ", Wentong Li" + }, + { + "bbox": [ + 103, + 160, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 103, + 160, + 510, + 190 + ], + "type": "text", + "content": ", Junbo Chen" + }, + { + "bbox": [ + 103, + 160, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + }, + { + "bbox": [ + 103, + 160, + 510, + 190 + ], + "type": "text", + "content": ", Jianke Zhu" + }, + { + "bbox": [ + 103, + 160, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 103, + 160, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 103, + 160, + 510, + 190 + ], + "type": "text", + "content": "Zhejiang University " + }, + { + "bbox": [ + 103, + 160, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 103, + 160, + 510, + 190 + ], + "type": "text", + "content": "NUAA " + }, + { + "bbox": [ + 103, + 160, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 103, + 160, + 510, + 190 + ], + "type": "text", + "content": "Udeer.ai" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 191, + 537, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 191, + 537, + 204 + ], + "spans": [ + { + "bbox": [ + 69, + 191, + 537, + 204 + ], + "type": "text", + "content": "{xiaoluliu, ruiziyang, songw, jkzhu}@zju.edu.cn, wentong.li@nuaa.edu.cn, junbo@udeer.ai" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 256, + 297, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 256, + 297, + 568 + ], + "spans": [ + { + "bbox": [ + 54, + 256, + 297, + 568 + ], + "type": "text", + "content": "Reliable high-definition (HD) map construction is crucial for the driving safety of autonomous vehicles. Although recent studies demonstrate improved performance, their generalization capability across unfamiliar driving scenes remains unexplored. To tackle this issue, we propose UIGenMap, an uncertainty-instructed structure injection approach for generalizable HD map vectorization, which concerns the uncertainty resampling in statistical distribution and employs explicit instance features to reduce excessive reliance on training data. Specifically, we introduce the perspective-view (PV) detection branch to obtain explicit structural features, in which the uncertainty-aware decoder is designed to dynamically sample probability distributions considering the difference in scenes. With probabilistic embedding and selection, UI2DPrompt is proposed to construct PV-learnable prompts. These PV prompts are integrated into the map decoder by designed hybrid injection to compensate for neglected instance structures. To ensure real-time inference, a lightweight Mimic Query Distillation is designed to learn from PV prompts, which can serve as an efficient alternative to the flow of PV branches. Extensive experiments on challenging geographically disjoint (geo-based) data splits demonstrate that our UIGenMap achieves superior performance, with " + }, + { + "bbox": [ + 54, + 256, + 297, + 568 + ], + "type": "inline_equation", + "content": "+5.7" + }, + { + "bbox": [ + 54, + 256, + 297, + 568 + ], + "type": "text", + "content": " mAP improvement on the nuScenes dataset. Source code is available at https://github.com/xiaolul2/UIGenMap." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 591, + 135, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 135, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 135, + 604 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 612, + 296, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 296, + 684 + ], + "type": "text", + "content": "Reliable HD map construction is crucial for driving safety of autonomous vehicles, which provides essential geometric information for ego-localization, perceptions, path planning, and predictions [10, 13, 23, 45]. Recently, online high-definition (HD) map vectorization has attracted considerable attention for its cost-effectiveness and ability to" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 319, + 230, + 425, + 348 + ], + "blocks": [ + { + "bbox": [ + 319, + 230, + 425, + 348 + ], + "lines": [ + { + "bbox": [ + 319, + 230, + 425, + 348 + ], + "spans": [ + { + "bbox": [ + 319, + 230, + 425, + 348 + ], + "type": "image", + "image_path": "5656bbe841dca3b885a7c186417c1316f1e3d9ed1f752b34ffcabae4674ef3af.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 426, + 230, + 554, + 350 + ], + "blocks": [ + { + "bbox": [ + 426, + 230, + 554, + 350 + ], + "lines": [ + { + "bbox": [ + 426, + 230, + 554, + 350 + ], + "spans": [ + { + "bbox": [ + 426, + 230, + 554, + 350 + ], + "type": "image", + "image_path": "9156bb1d437084df267619f6a5d8a538c3c80a72c9065125b6e56e780b50e116.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 351, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 313, + 351, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 555, + 407 + ], + "type": "text", + "content": "Figure 1. (a) Comparisons on original and different geo-based data partitions among previous methods and ours. (b) Visual comparisons with circles on BEV map to represent the learned uncertainties, in which PV instances serve as compensation and provide a detailed understanding of map perception with generalization." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 421, + 555, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 555, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 555, + 481 + ], + "type": "text", + "content": "provide real-time updates [25, 31, 56]. Given images captured by onboard cameras, current map construction relies on transformer scheme [57] to convert perspective-view (PV) image features into bird's-eye-view (BEV) space. Then, a decoder follows to predict the map elements." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 483, + 556, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 556, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 556, + 699 + ], + "type": "text", + "content": "Despite the achievements on widely adopted benchmarks [4, 26, 28], little attention has been paid to investigating the generalization capabilities, which refer to the model's performance on unfamiliar driving scenarios that are different from the training sets. As discussed in [27, 50], in the widely used public datasets [1, 46], there are a large number of geographical overlaps between the training and validation scenes. The models are then more likely to memorize similar driving scenes rather than fully capture their underlying road structures. Variations in driving scenes, such as the contrast between bustling city centers and suburbs with simple road structures, highlight significant shifts in feature distributions and reveal the model's limitations in robustness. Besides, learning-based BEV transformation will inevitably lead to geometric errors and loss of texture details. As shown in Figure 1(a), previous methods [25, 50] exhibit notable performance degradation on region-based and city-based data splits under changing driving scenes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "content": "To tackle the above issue, we propose the innovative em" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 149, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 149, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 149, + 703 + ], + "type": "text", + "content": "*Equal Contribution" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "text", + "content": "† Corresponding authors" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22359" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 287 + ], + "type": "text", + "content": "employment of model uncertainty targeted for more generalizable HD map constructions. On the one hand, the uncertainty estimation process involves learning both the statistical mean and variance value, enabling dynamic resampling based on probabilistic distributions. Given the discrepancy in feature distribution across different driving scenes, such a statistical resampling strategy enhances the model's dynamic adaptability in unfamiliar environments. However, the estimated model uncertainty quantifies the confidence in its predictions, which can serve as a reliable indicator for further feature selection and enhancement. As illustrated in Figure 1(b), direct 2D perception can capture more intuitive semantic and angular structural information, which might be missed or incorrectly learned during the implicit PV-to-BEV conversions. Thus, combined with uncertainty-based feature selection, more reliable explicit information can be emphasized as compensation to boost the robustness of BEV-level map construction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 290, + 295, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 290, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 290, + 295, + 553 + ], + "type": "text", + "content": "Based on the above discussion, in this work, we propose UIGenMap, an uncertainty-instructed PV structure injection strategy for more generalizable HD map vectorization. Specifically, a PV detection branch is introduced to provide explicit structural instances. In both PV and BEV branches, we move beyond the vanilla transformer structure by designing the statistically sensitive uncertainty-aware decoder (UA-Decoder), which constructs the probabilistic outputs by uncertainty-aware attention and finer point-level uncertainty head. Leveraging the PV uncertainty output, we introduce the UI2DPrompt to construct learnable PV prompts. These PV prompts are integrated into the BEV decoder by the designed hybrid injection mechanism, which acts as structural compensation for the BEV map prediction. Consequently, reliable structural representation from PV space can be injected to enhance the model's capability, thereby improving its perceptual ability in complex environments. Considering the demands for real-time inference, we also design a lightweight mimic query distillation module to learn from constructed PV prompts. Then during inference, the mimicked query can replace the PV detection flow to reduce extra computational burdens." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 555, + 294, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 294, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 294, + 662 + ], + "type": "text", + "content": "To prioritize the evaluation of generalization capability, we assess our method using geographically disjoint (geo-based) data partitions on the nuScenes [1] and Argoverse2 [46] datasets. Extensive experiments demonstrate that our UIGenMap achieves state-of-the-art performance with higher adaptive ability, especially with " + }, + { + "bbox": [ + 55, + 555, + 294, + 662 + ], + "type": "inline_equation", + "content": "+5.7" + }, + { + "bbox": [ + 55, + 555, + 294, + 662 + ], + "type": "text", + "content": " mAP improvement on the region-based partition and " + }, + { + "bbox": [ + 55, + 555, + 294, + 662 + ], + "type": "inline_equation", + "content": "+4.3" + }, + { + "bbox": [ + 55, + 555, + 294, + 662 + ], + "type": "text", + "content": " mAP on the more challenging city-based split on nuScenes. The main contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 665, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 700 + ], + "type": "text", + "content": "- We propose UIGenMap, an uncertainty-instructed approach to more generalizable HD map vectorization with explicit structure injection." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "- The UA-Decoder and UI2DPrompt are proposed to con" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 95 + ], + "type": "text", + "content": "struct structural PV prompts, which are dynamically integrated with BEV perception by hybrid injection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 96, + 553, + 168 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 96, + 553, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 96, + 553, + 131 + ], + "spans": [ + { + "bbox": [ + 314, + 96, + 553, + 131 + ], + "type": "text", + "content": "- Lightweight MQ-Distillation design supports real-time inference without the additional computational overhead of the PV detection branches." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 133, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 553, + 168 + ], + "type": "text", + "content": "- Promising results on more challenging geo-based data partitions show that UIGenMap outperforms previous approaches and achieves powerful generalization capability." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 178, + 400, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 178, + 400, + 190 + ], + "spans": [ + { + "bbox": [ + 314, + 178, + 400, + 190 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 199, + 555, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 555, + 474 + ], + "type": "text", + "content": "Vectorized HD Map Construction. Conventional HD map construction approaches [38, 39, 58] rely on SLAM-based annotations, which require tedious offline manual labeling at high costs. Recent studies [6, 11, 19, 20, 29, 37, 49] have explored online approaches with the input of onboard sensor data. In [18, 22, 24, 35, 42], HDMap construction is treated as a segmentation task, relying on post-processing to achieve map vectorization. Later, Liao et al. [25, 26] represent map elements as ordered points and employ transformers [41, 57] after the BEV encoder for auto-regressive detection. Other modeling methods, such as Bezier curves and pivot points, are explored in BeMapNet [37] and PivotNet [6] for geometric representation. Several works [4, 19, 40, 43, 50] are proposed for different temporal fusion approaches. GeMap [54] and MapQR [31] improve mapping performance from the perspectives of geometric relationships and decoder design, respectively. In [5, 28, 51, 52], rasterized maps are utilized for structure compensation. Despite improved performance, these approaches seldom consider the generalization capability of models in unfamiliar driving scenarios. In this work, we incorporate uncertainty-instructed PV detection retracing, aiming for better performance across diverse scenarios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 475, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 554, + 677 + ], + "type": "text", + "content": "PV Detection for 3D Perception. Instances from PV detection can be effective compensations for 3D perceptions [12, 14, 32, 47, 48, 53]. For 3D object detection, BEVFormerv2 [48] utilizes a two-stage BEV detector that integrates the query from the PV head into the BEV space. In SimMoD [53], 2D detection is utilized to construct proposals for 3D fine-grained perception. Far3D [14] combines 2D detection boxes with estimated depth information to improve long-range perception performance. For model generalization, Lu et al. [32] propose a generalizable framework based on perspective debiasing, in which PV consistency learning is employed among different domains. Applying for 3D lanes, Li et al. [17] use 2D priors for lane detection and topology reasoning. Inspired by the above methods, we employ PV structural features for more generalizable HD map vectorization, in which distillations from 2D instances are utilized to ensure inference efficiency." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Uncertainty in Model. Generally, uncertainty refers to the model's confidence level for its predictions, which can be a statistical representation to indicate the model's under" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22360" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 85, + 239, + 304 + ], + "blocks": [ + { + "bbox": [ + 63, + 79, + 113, + 90 + ], + "lines": [ + { + "bbox": [ + 63, + 79, + 113, + 90 + ], + "spans": [ + { + "bbox": [ + 63, + 79, + 113, + 90 + ], + "type": "text", + "content": "(a) Overview" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 85, + 239, + 304 + ], + "lines": [ + { + "bbox": [ + 61, + 85, + 239, + 304 + ], + "spans": [ + { + "bbox": [ + 61, + 85, + 239, + 304 + ], + "type": "image", + "image_path": "b6bbcd069bbfbe9d38546f9957fe490aa974ee4bdbc101c565d45ed913b89ba3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 307, + 555, + 352 + ], + "lines": [ + { + "bbox": [ + 55, + 307, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 555, + 352 + ], + "type": "text", + "content": "Figure 2. (a) Overview of Our UIGenMap. For training, the PV branch is introduced with the uncertainty-instructed structural injection, in which the MQ-Distillation is designed to mimic PV structural features. (b) Uncertainty-Aware Decoder Architecture. For each layer, UA-Decoder comprises probabilistic UA-Attention and UA-Head for reliable output. (c) UI2DPrompt Design. We construct PV prompts from PV-detected elements and their corresponding uncertainties, which are integrated with the main branch by hybrid injection." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 246, + 81, + 547, + 176 + ], + "blocks": [ + { + "bbox": [ + 246, + 76, + 299, + 85 + ], + "lines": [ + { + "bbox": [ + 246, + 76, + 299, + 85 + ], + "spans": [ + { + "bbox": [ + 246, + 76, + 299, + 85 + ], + "type": "text", + "content": "(b) UA-Decoder" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 246, + 81, + 547, + 176 + ], + "lines": [ + { + "bbox": [ + 246, + 81, + 547, + 176 + ], + "spans": [ + { + "bbox": [ + 246, + 81, + 547, + 176 + ], + "type": "image", + "image_path": "b3d856e728e09cd8b9f9e9daa043ffb7d42a9757d215f705d8449fdb9a96b0b6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 247, + 188, + 543, + 304 + ], + "blocks": [ + { + "bbox": [ + 249, + 182, + 300, + 191 + ], + "lines": [ + { + "bbox": [ + 249, + 182, + 300, + 191 + ], + "spans": [ + { + "bbox": [ + 249, + 182, + 300, + 191 + ], + "type": "text", + "content": "(c) UI2DPrompt" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 247, + 188, + 543, + 304 + ], + "lines": [ + { + "bbox": [ + 247, + 188, + 543, + 304 + ], + "spans": [ + { + "bbox": [ + 247, + 188, + 543, + 304 + ], + "type": "image", + "image_path": "ad0eba1419bc33d3c8b2b80bd59af6aa08f0016479f848884af8fecaee102b92.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 365, + 297, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 297, + 558 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 297, + 558 + ], + "type": "text", + "content": "standing of scenes [3, 15, 44]. Uncertainty has gained intensive attention due to the potential for reliability evaluation and performance enhancement [2, 16]. Specifically, the generalization problem with uncertainty modeling is investigated in [21, 36], which expects uncertainty estimation for out-of-distribution generalization. For the transformer structure [8, 34], uncertainty has been explored to guide the training and inference processes for more robust model performance. In the field of autonomous driving, significant uncertainty has been introduced for 3D occupancy [2, 44], optical flow [33], and object tracking [55]. Recently, Gu et al. [7] present uncertainty estimation for the vectorized map, which is targeted for trajectory prediction. In our design, we utilize uncertainty for more generalized HD map construction, which considers the variability in feature distributions and incorporates explicit instance features." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 571, + 129, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 129, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 129, + 586 + ], + "type": "text", + "content": "3. UIGenMap" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 592, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 295, + 664 + ], + "type": "text", + "content": "Given surrounding-view images captured from onboard cameras, we aim to construct more robust vectorized map elements in the BEV space. Each map element comprises a class label " + }, + { + "bbox": [ + 55, + 592, + 295, + 664 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{map}}" + }, + { + "bbox": [ + 55, + 592, + 295, + 664 + ], + "type": "text", + "content": " and an ordered sequence of points " + }, + { + "bbox": [ + 55, + 592, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{map}} = \\{(p_x^i,p_y^i)\\}_{i = 1}^N" + }, + { + "bbox": [ + 55, + 592, + 295, + 664 + ], + "type": "text", + "content": " that represents the structure of the instance. " + }, + { + "bbox": [ + 55, + 592, + 295, + 664 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 592, + 295, + 664 + ], + "type": "text", + "content": " denotes the total number of points for each map element." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "The architecture of UIGenMap is presented in Figure 2. Based on the main BEV detection, we obtain PV instance from an introduced pre-trained PV branch, which is injected as structural priors for BEV perception. Firstly, we extract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 365, + 555, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 559 + ], + "type": "text", + "content": "image features for the PV and BEV branches separately. As in [22], BEV features " + }, + { + "bbox": [ + 313, + 365, + 555, + 559 + ], + "type": "inline_equation", + "content": "\\mathrm{F_{bev}}" + }, + { + "bbox": [ + 313, + 365, + 555, + 559 + ], + "type": "text", + "content": " are constructed from the interaction between PV features and learnable BEV queries. Then, we design the uncertainty-aware (UA-) decoder from inner-layer UA-Attention and point-level heads to achieve reliable probabilistic outputs. Based on the PV coordinates and uncertainty output, we propose the UI2DPrompt module to construct reliable PV prompts. These prompts are incorporated into BEV features and map instance queries via hybrid injection, where uncertainties are utilized as the basis for feature selection and enhancement. Finally, the UA-Decoder employs enhanced BEV features and instance queries for reliable map predictions. To ensure real-time inference, the lightweight MQ-Distillation module performs query distillation from learned PV prompts to imitate explicit structural features, aiming for efficient computation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 564, + 468, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 468, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 468, + 577 + ], + "type": "text", + "content": "3.1. Uncertainty-Aware Decoder" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 677 + ], + "type": "text", + "content": "We design UA-Decoder for both PV and BEV space. Systematic uncertainty-aware design is employed to improve the statistical understanding of features and outputs to achieve more dynamic self-adjustment for different driving scenarios. To ensure comprehensive perception, each decoder layer comprises UA-Attention at the instance feature level and UA-Head for point-level uncertainty output along with positions, respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": "UA-Attention. In conventional deformable attention [57], the weight " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " for each sampled feature values from BEV feature " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{BEV}}" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " is determined by linear layer (Linear(·)) on" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22361" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 295, + 207 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 295, + 207 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 295, + 207 + ], + "type": "image", + "image_path": "e4226113f26f405ad6d4e9364b9a37a32d401984fdacb846dc0b5bcc270199f6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 209, + 296, + 255 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 296, + 255 + ], + "type": "text", + "content": "Figure 3. UA-Attention Design. The learned probabilistic weights " + }, + { + "bbox": [ + 55, + 209, + 296, + 255 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 55, + 209, + 296, + 255 + ], + "type": "text", + "content": " are dynamically resampled from Gaussian distribution with stochastic disturbance, which is multiplied with sampled feature values to update map queries." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 261, + 296, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 261, + 296, + 454 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 296, + 454 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{map}}^i" + }, + { + "bbox": [ + 55, + 261, + 296, + 454 + ], + "type": "text", + "content": " , which can be expressed as " + }, + { + "bbox": [ + 55, + 261, + 296, + 454 + ], + "type": "inline_equation", + "content": "\\alpha_{i} = \\mathrm{Linear}(Q_{\\mathrm{map}}^{i})" + }, + { + "bbox": [ + 55, + 261, + 296, + 454 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 261, + 296, + 454 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{map}}^i" + }, + { + "bbox": [ + 55, + 261, + 296, + 454 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 55, + 261, + 296, + 454 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 261, + 296, + 454 + ], + "type": "text", + "content": " -th instance query of totally " + }, + { + "bbox": [ + 55, + 261, + 296, + 454 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 261, + 296, + 454 + ], + "type": "text", + "content": " map elements. Then, the updated part of query " + }, + { + "bbox": [ + 55, + 261, + 296, + 454 + ], + "type": "inline_equation", + "content": "\\Delta Q_{\\mathrm{map}}^i" + }, + { + "bbox": [ + 55, + 261, + 296, + 454 + ], + "type": "text", + "content": " is aggregated by the dot-product between the learned weight " + }, + { + "bbox": [ + 55, + 261, + 296, + 454 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 55, + 261, + 296, + 454 + ], + "type": "text", + "content": " and sampled feature values from " + }, + { + "bbox": [ + 55, + 261, + 296, + 454 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{bev}}" + }, + { + "bbox": [ + 55, + 261, + 296, + 454 + ], + "type": "text", + "content": " . Considering complex driving scenarios, such deterministic weights are inadaptable for challenging situations. More dynamic weights are required to avoid potential ambiguities. Inspired by [8, 34], we explicitly model the attention scores under a Gaussian distribution " + }, + { + "bbox": [ + 55, + 261, + 296, + 454 + ], + "type": "inline_equation", + "content": "\\alpha_{i}\\sim \\mathcal{N}(\\mu_{i},\\sigma_{i}^{2})" + }, + { + "bbox": [ + 55, + 261, + 296, + 454 + ], + "type": "text", + "content": " , as shown in Figure 3. We adopt the re-parameterization strategy to perform the probabilistic attention, which is constructed by the corresponding mean " + }, + { + "bbox": [ + 55, + 261, + 296, + 454 + ], + "type": "inline_equation", + "content": "\\mu_{i}" + }, + { + "bbox": [ + 55, + 261, + 296, + 454 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 55, + 261, + 296, + 454 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 55, + 261, + 296, + 454 + ], + "type": "text", + "content": " . Such probabilistic parameters are predicted through multilayer perception (MLP(·)). Thus, the construction of dynamic resampled attention weights can be formulated as below:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 465, + 294, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 465, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 136, + 465, + 294, + 479 + ], + "type": "interline_equation", + "content": "\\left\\{\\mu_ {i}, \\sigma_ {i} ^ {2} \\right\\} = \\mathrm {M L P} \\left(Q _ {\\text {m a p}} ^ {i}\\right), \\tag {1}", + "image_path": "3595f05cc7df1b9738166b78482d22b73d6fb10629966bae256f587ae3b80da0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 118, + 483, + 294, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 483, + 294, + 495 + ], + "spans": [ + { + "bbox": [ + 118, + 483, + 294, + 495 + ], + "type": "interline_equation", + "content": "\\alpha_ {i} = \\mu_ {i} + \\sigma_ {i} \\epsilon , \\epsilon \\sim \\mathcal {N} (0, 1), \\tag {2}", + "image_path": "5e427de776c3385e9bd192c40492461216f78792a4b2a0f9d2e5633ad4469032.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 504, + 295, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 295, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 295, + 576 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 504, + 295, + 576 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 504, + 295, + 576 + ], + "type": "text", + "content": " is the Gaussian noise conforming to normal distribution. During inference, attention weight " + }, + { + "bbox": [ + 55, + 504, + 295, + 576 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 55, + 504, + 295, + 576 + ], + "type": "text", + "content": " can be dynamically resampled from learned statistical distribution to suit for driving scenarios with large variations. Then query features can be obtained by dot production with the sampled feature values in order to enable reliable query updates." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 577, + 296, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 577, + 296, + 716 + ], + "spans": [ + { + "bbox": [ + 55, + 577, + 296, + 716 + ], + "type": "text", + "content": "UA-Head with Uncertainty Loss. Based on the robust design of UA-Attention, quantized uncertainty for each lane point output is also essential, which is point-specific at a fine-grained level to explicitly reflect the reliability and uncertainty information within the internal transformer layers. Inspired by [7], we replace the vanilla deterministic output with the UA-Head that predicts coordinates " + }, + { + "bbox": [ + 55, + 577, + 296, + 716 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{map}} = \\{(\\hat{p}_x^i,\\hat{p}_y^i)\\}_{i = 1}^N" + }, + { + "bbox": [ + 55, + 577, + 296, + 716 + ], + "type": "text", + "content": " along with their corresponding uncertainties " + }, + { + "bbox": [ + 55, + 577, + 296, + 716 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{map}} = \\{(\\sigma_x^i,\\sigma_y^i)\\}_{i = 1}^N" + }, + { + "bbox": [ + 55, + 577, + 296, + 716 + ], + "type": "text", + "content": " for each map element, as shown in Figure 2 (b). These parameters represent the location and scale of the joint Laplace distribution " + }, + { + "bbox": [ + 55, + 577, + 296, + 716 + ], + "type": "inline_equation", + "content": "Laplace(\\hat{p}_j^i,\\frac{\\sigma_j^i}{\\sqrt{2}})" + }, + { + "bbox": [ + 55, + 577, + 296, + 716 + ], + "type": "text", + "content": " at" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "the " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "-th dimension " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "(x" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "y)" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "-th point, which refers to a probabilistic representation for stable output." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 96, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 553, + 144 + ], + "type": "text", + "content": "For uncertainty-aware training, we combine probabilistic Negative Log-Likelihood (NLL) loss " + }, + { + "bbox": [ + 313, + 96, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{nll}}" + }, + { + "bbox": [ + 313, + 96, + 553, + 144 + ], + "type": "text", + "content": " with point regression loss " + }, + { + "bbox": [ + 313, + 96, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pts}}" + }, + { + "bbox": [ + 313, + 96, + 553, + 144 + ], + "type": "text", + "content": " in [50] to balance uncertainties and coordinate outputs, in which " + }, + { + "bbox": [ + 313, + 96, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{nll}}" + }, + { + "bbox": [ + 313, + 96, + 553, + 144 + ], + "type": "text", + "content": " can be formulated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 330, + 151, + 553, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 151, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 330, + 151, + 553, + 186 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {n l l}} = \\lambda_ {\\mathrm {n l l}} \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {2} \\left(\\log \\left(2 \\sigma_ {j} ^ {i}\\right) + \\frac {\\left| p _ {j} ^ {i} - \\hat {p} _ {j} ^ {i} \\right|}{\\sigma_ {j} ^ {i}}\\right), \\tag {3}", + "image_path": "6c64c4706cf34e5f03c834fcdee0459b04a7279e7897ca4973ee30d4e4eec9ed.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 192, + 456, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 456, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 456, + 204 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 313, + 192, + 456, + 204 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{nl}}" + }, + { + "bbox": [ + 313, + 192, + 456, + 204 + ], + "type": "text", + "content": " as the loss weight setting." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 204, + 555, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 555, + 326 + ], + "type": "text", + "content": "As illustrated in Figure 2 (a), UA-Decoder is used for the construction of the PV branch and the BEV map. To provide more robust structural features in PV branch, PV coordinates " + }, + { + "bbox": [ + 313, + 204, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\mathrm{P_{pv}} = \\{x_{pv}^i,y_{pv}^i\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 204, + 555, + 326 + ], + "type": "text", + "content": " and the uncertainty parameters " + }, + { + "bbox": [ + 313, + 204, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{pv}} = \\{\\sigma_{\\mathrm{pv,x}}^{i},\\sigma_{\\mathrm{pv,y}}^{i}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 313, + 204, + 555, + 326 + ], + "type": "text", + "content": " are constructed by the UA-Decoder structure, which is later utilized in our UI2DPrompt design. Thus, the systematic combination of uncertainty-aware design provides dynamic sampling during the learning process, along with established uncertainty, to construct a reliable basis for confidence scoring." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 331, + 399, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 399, + 345 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 399, + 345 + ], + "type": "text", + "content": "3.2. UI2DPrompt" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 349, + 554, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 554, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 554, + 444 + ], + "type": "text", + "content": "PV branches contain abundant structural and geometric details. To mitigate information loss and misinterpretation during query-based PV-to-BEV transformation, reliable structural PV instances can be injected as compensation for BEV prediction. This section focuses on constructing PV prompts based on uncertainty-aware PV predictions, which are dynamically integrated with the map decoder for BEV map element detection." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 445, + 555, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 625 + ], + "type": "text", + "content": "Prompt Construction. Given the learned PV instance coordinates and the corresponding uncertainty parameters " + }, + { + "bbox": [ + 313, + 445, + 555, + 625 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{P_{pv}},\\sigma_{\\mathrm{pv}}\\}" + }, + { + "bbox": [ + 313, + 445, + 555, + 625 + ], + "type": "text", + "content": ", PV prompts are constructed to provide dynamic 2D priors. Firstly, we employ the estimated classification score " + }, + { + "bbox": [ + 313, + 445, + 555, + 625 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{map}}" + }, + { + "bbox": [ + 313, + 445, + 555, + 625 + ], + "type": "text", + "content": " for PV candidates selection. When " + }, + { + "bbox": [ + 313, + 445, + 555, + 625 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{map}} > c_{\\mathrm{thr}}" + }, + { + "bbox": [ + 313, + 445, + 555, + 625 + ], + "type": "text", + "content": ", PV instances are chosen as candidates for reliable localized information. Considering the sparse depth information and the characteristics of static map elements on the ground, we utilize Inverse Perspective Mapping (IPM) to transform " + }, + { + "bbox": [ + 313, + 445, + 555, + 625 + ], + "type": "inline_equation", + "content": "\\mathrm{P_{pv}}" + }, + { + "bbox": [ + 313, + 445, + 555, + 625 + ], + "type": "text", + "content": " from the PV image coordinate to the BEV coordinate system " + }, + { + "bbox": [ + 313, + 445, + 555, + 625 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_{\\mathrm{trans}} = \\{p_x^{\\prime},p_y^{\\prime}\\}" + }, + { + "bbox": [ + 313, + 445, + 555, + 625 + ], + "type": "text", + "content": ", which makes a flat grounding assumption and utilizes the affine transformation between the PV and BEV coordinates. Given the corresponding intrinsic parameters " + }, + { + "bbox": [ + 313, + 445, + 555, + 625 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{\\mathrm{cam}}" + }, + { + "bbox": [ + 313, + 445, + 555, + 625 + ], + "type": "text", + "content": " and extrinsic transformation matrix " + }, + { + "bbox": [ + 313, + 445, + 555, + 625 + ], + "type": "inline_equation", + "content": "\\mathrm{T}_{\\mathrm{ego2cam}}" + }, + { + "bbox": [ + 313, + 445, + 555, + 625 + ], + "type": "text", + "content": ", the projection " + }, + { + "bbox": [ + 313, + 445, + 555, + 625 + ], + "type": "inline_equation", + "content": "\\mathrm{P_{pv}}" + }, + { + "bbox": [ + 313, + 445, + 555, + 625 + ], + "type": "text", + "content": " can be defined as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 344, + 632, + 553, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 632, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 344, + 632, + 553, + 658 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {\\text {t r a n s}} = \\operatorname {P r o j} _ {2 D} \\left(\\mathrm {T} _ {\\text {e g o 2 c a m}} ^ {- 1} \\cdot \\left[ \\begin{array}{c} \\mathrm {K} _ {\\text {c a m}} ^ {- 1} \\cdot \\mathrm {P} _ {\\mathrm {p v}} \\\\ 1 \\end{array} \\right]\\right), \\tag {4}", + "image_path": "53fdb31e63299f12eeb65cab50929110258666d5e14a3e5c3a2bda90b4067028.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\operatorname{Proj}_{2\\mathrm{D}}(\\cdot)" + }, + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "type": "text", + "content": " denotes the conversion from homogeneous coordinates into 2D plane coordinates." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "To ensure reliable enhanced features, the corresponding uncertainty parameters " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{pv}}" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " cooperate as weights and feature" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22362" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": "embeddings. We encode the transformed coordinates " + }, + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{trans}}" + }, + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": " and uncertainty parameters " + }, + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{pv}}" + }, + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "inline_equation", + "content": "\\phi_p(\\cdot)" + }, + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "inline_equation", + "content": "\\phi_{\\sigma}(\\cdot)" + }, + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": ", respectively. Point-specific embeddings " + }, + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "inline_equation", + "content": "\\{e_{pv}^i\\}_{i = 1}^N" + }, + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": " are constructed by concatenation:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 97, + 125, + 295, + 140 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 125, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 97, + 125, + 295, + 140 + ], + "type": "interline_equation", + "content": "\\boldsymbol {e} _ {\\mathrm {p v}} ^ {i} = \\operatorname {C o n c a t} \\left(\\left[ \\phi_ {p} \\left(\\mathrm {P} _ {\\text {t r a n s}}\\right), \\phi_ {\\sigma} \\left(\\sigma_ {\\mathrm {p v}}\\right) \\right]\\right). \\tag {5}", + "image_path": "6e374426785f0e6854c3114ec740d6d59bd42b876038fd24bc7397fb34c19c2b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 144, + 295, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 169 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 169 + ], + "type": "text", + "content": "Thus, the structural PV output coordinates and uncertainty information are dynamically encoded." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 169, + 296, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 169, + 296, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 296, + 242 + ], + "type": "text", + "content": "Based on the above, PV uncertainties can serve as the quantized weights for dynamic feature strengthening and adjustment. Reliable confidence indicator is able to be formulated as the weight " + }, + { + "bbox": [ + 55, + 169, + 296, + 242 + ], + "type": "inline_equation", + "content": "\\{\\omega_{\\mathrm{pv}}^i\\}_{i = 1}^N" + }, + { + "bbox": [ + 55, + 169, + 296, + 242 + ], + "type": "text", + "content": " concerning each point-specific prompt. Then, the enhanced PV prompt " + }, + { + "bbox": [ + 55, + 169, + 296, + 242 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{e}}_{\\mathrm{pv}}^{i}" + }, + { + "bbox": [ + 55, + 169, + 296, + 242 + ], + "type": "text", + "content": " can be constructed by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 247, + 295, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 247, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 61, + 247, + 295, + 289 + ], + "type": "interline_equation", + "content": "\\omega_ {\\mathrm {p v}} ^ {i} = \\exp \\left(\\frac {\\left(\\| \\sigma_ {\\mathrm {p v}} ^ {i} \\| _ {2}\\right) ^ {- 1}}{\\sum_ {i = 1} ^ {N} \\left(\\| \\sigma_ {\\mathrm {p v}} ^ {i} \\| _ {2}\\right) ^ {- 1}}\\right), \\quad \\tilde {\\boldsymbol {e}} _ {\\mathrm {p v}} ^ {i} = \\omega_ {\\mathrm {p v}} ^ {i} \\boldsymbol {e} _ {\\mathrm {p v}} ^ {i} + \\boldsymbol {e} _ {\\mathrm {m}} ^ {i}, \\tag {6}", + "image_path": "4500a4cc3fe09dda75c7d4a8fb1ad89fec5c344750b5ffa10d84d99bc0d0c46c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 289, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 295, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 289, + 295, + 361 + ], + "type": "inline_equation", + "content": "\\exp (\\cdot)" + }, + { + "bbox": [ + 55, + 289, + 295, + 361 + ], + "type": "text", + "content": " represents the exponential amplification of the normalized inverse uncertainty value. " + }, + { + "bbox": [ + 55, + 289, + 295, + 361 + ], + "type": "inline_equation", + "content": "\\{e_{\\mathrm{m}}^{i}\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 55, + 289, + 295, + 361 + ], + "type": "text", + "content": " is the predefined mimic query, which is utilized for MQ-Distillation. This approach reduces the introduction of PV errors, ensuring that injecting PV instances amplifies reliable PV information while minimizing uncertainty-induced errors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 361, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 445 + ], + "type": "text", + "content": "Hybrid Injection. Constructed from PV instances with uncertainty output, PV prompts " + }, + { + "bbox": [ + 55, + 361, + 296, + 445 + ], + "type": "inline_equation", + "content": "\\{\\tilde{e}_{\\mathrm{pv}}^i\\}_{i = 1}^N" + }, + { + "bbox": [ + 55, + 361, + 296, + 445 + ], + "type": "text", + "content": " are incorporated as compensation for the construction of the BEV map. These prompts are injected into the BEV feature " + }, + { + "bbox": [ + 55, + 361, + 296, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{bev}}" + }, + { + "bbox": [ + 55, + 361, + 296, + 445 + ], + "type": "text", + "content": " and the map queries " + }, + { + "bbox": [ + 55, + 361, + 296, + 445 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{map}}^i" + }, + { + "bbox": [ + 55, + 361, + 296, + 445 + ], + "type": "text", + "content": " for the map decoder by the designed hybrid injection mechanism, which includes both prompt-to-BEV (P2BEV) and prompt-to-query (P2Q) attention." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 445, + 296, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 296, + 505 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 296, + 505 + ], + "type": "text", + "content": "As shown in Figure 4, point-level PV prompts " + }, + { + "bbox": [ + 55, + 445, + 296, + 505 + ], + "type": "inline_equation", + "content": "\\{\\tilde{e}_{\\mathrm{pv}}^i\\}_{i = 1}^N" + }, + { + "bbox": [ + 55, + 445, + 296, + 505 + ], + "type": "text", + "content": " are integrated with " + }, + { + "bbox": [ + 55, + 445, + 296, + 505 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{bev}}" + }, + { + "bbox": [ + 55, + 445, + 296, + 505 + ], + "type": "text", + "content": " through cross-attention of P2BEV. This step enriches BEV features by integrating detailed, point-specific information from PV prompts, enhancing instance-specific detail in the BEV representation:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 92, + 510, + 295, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 510, + 295, + 526 + ], + "spans": [ + { + "bbox": [ + 92, + 510, + 295, + 526 + ], + "type": "interline_equation", + "content": "\\mathrm {F} _ {\\mathrm {b e v}} ^ {\\prime} = \\operatorname {A t t n} \\left(W _ {q} \\mathrm {F} _ {\\mathrm {b e v}}, W _ {k} \\tilde {\\boldsymbol {e}} _ {\\mathrm {p v}} ^ {i}, W _ {v} \\tilde {\\boldsymbol {e}} _ {\\mathrm {p v}} ^ {i}\\right), \\tag {7}", + "image_path": "f6d0f8bb6e92f257bf61b6e5f6269c62de788f215b91351d1dcde065b68b4bec.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 529, + 296, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 296, + 624 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 529, + 296, + 624 + ], + "type": "inline_equation", + "content": "W_{\\{q,k,v\\}}" + }, + { + "bbox": [ + 55, + 529, + 296, + 624 + ], + "type": "text", + "content": " are the corresponding projection matrix. Then, we encode point-level prompts into instance-level representations through a linear block " + }, + { + "bbox": [ + 55, + 529, + 296, + 624 + ], + "type": "inline_equation", + "content": "\\phi_e(\\cdot)" + }, + { + "bbox": [ + 55, + 529, + 296, + 624 + ], + "type": "text", + "content": ", which aggregates point-specific information to create a higher-level representation for subsequent processing. These instance-level prompts " + }, + { + "bbox": [ + 55, + 529, + 296, + 624 + ], + "type": "inline_equation", + "content": "\\tilde{e}_{\\mathrm{inst}}" + }, + { + "bbox": [ + 55, + 529, + 296, + 624 + ], + "type": "text", + "content": " are used to integrate with " + }, + { + "bbox": [ + 55, + 529, + 296, + 624 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{map}}" + }, + { + "bbox": [ + 55, + 529, + 296, + 624 + ], + "type": "text", + "content": ", which inject reliable priors for the initialized queries and enhance the overall map construction:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 127, + 629, + 295, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 629, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 127, + 629, + 295, + 645 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {e}} _ {\\text {i n s t}} = \\phi_ {e} \\left(\\left\\{\\tilde {\\boldsymbol {e}} _ {\\mathrm {p v}} ^ {i} \\right\\} _ {i = 1} ^ {N}\\right), \\tag {8}", + "image_path": "0bcae6764553051bc735654aa96d853b60354eccd70eb3ddaeae8edc71165afb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 83, + 661, + 295, + 675 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 661, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 83, + 661, + 295, + 675 + ], + "type": "interline_equation", + "content": "Q _ {\\text {m a p}} ^ {\\prime} = \\operatorname {A t t n} \\left(W _ {q} Q _ {\\text {m a p}}, W _ {k} \\tilde {\\mathbf {e}} _ {\\text {i n s t}}, W _ {v} \\tilde {\\mathbf {e}} _ {\\text {i n s t}}\\right). \\tag {9}", + "image_path": "a61122792e1259c4b19e214f543d69367f0e8580f4f43e57a06ae7d1fed44679.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "By injecting PV priors into both BEV and decoder queries, this hybrid mechanism improves the reliability of BEV perception and reduces transition errors. Explicit structural" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 553, + 211 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 553, + 211 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 211 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 211 + ], + "type": "image", + "image_path": "12ff6682f09d2c79c4a074e8d7670d60298783929293de060d1f9cd29885c73e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 213, + 555, + 247 + ], + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 247 + ], + "type": "text", + "content": "Figure 4. Design of Hybrid Injection. P2BEV-Attention and P2Q-Attention are designed for PV prompts to integrate with BEV features and map instance queries as structural compensation." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 257, + 555, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 555, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 555, + 294 + ], + "type": "text", + "content": "features enhance the generalization capability of the learned model and perform robustly in various driving environments, ensuring better accuracy and consistency." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 301, + 454, + 315 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 454, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 454, + 315 + ], + "type": "text", + "content": "3.3. Mimic Query Distillation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 320, + 554, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 554, + 367 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 554, + 367 + ], + "type": "text", + "content": "To address the extra computation introduced by the PV branch, we propose the lightweight MQ-Distillation to distill from PV prompt construction, where the mimic query can serve as an alternative to achieve real-time inference." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 369, + 555, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 464 + ], + "type": "text", + "content": "In MQ-Distillation, we define a set of mimic queries " + }, + { + "bbox": [ + 313, + 369, + 555, + 464 + ], + "type": "inline_equation", + "content": "\\{e_m^i\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 369, + 555, + 464 + ], + "type": "text", + "content": ", along with an MLP-based query learner " + }, + { + "bbox": [ + 313, + 369, + 555, + 464 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 313, + 369, + 555, + 464 + ], + "type": "text", + "content": " which is designed to capture and distill the structural features of the PV prompt construction process " + }, + { + "bbox": [ + 313, + 369, + 555, + 464 + ], + "type": "inline_equation", + "content": "\\tilde{e}_{\\mathrm{pv}}^i" + }, + { + "bbox": [ + 313, + 369, + 555, + 464 + ], + "type": "text", + "content": ". As shown in Equation 6, during training, the predefined mimic queries " + }, + { + "bbox": [ + 313, + 369, + 555, + 464 + ], + "type": "inline_equation", + "content": "e_{m}^{i}" + }, + { + "bbox": [ + 313, + 369, + 555, + 464 + ], + "type": "text", + "content": " and the constructed PV prompts " + }, + { + "bbox": [ + 313, + 369, + 555, + 464 + ], + "type": "inline_equation", + "content": "\\tilde{e}_{\\mathrm{pv}}^i" + }, + { + "bbox": [ + 313, + 369, + 555, + 464 + ], + "type": "text", + "content": " are jointly combined by addition, which are sent to the hybrid injection for further operations." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 464, + 554, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 554, + 536 + ], + "type": "text", + "content": "To obtain sufficient structural features and achieve progressive alignment with PV prompts, we design the loss function that guides mimic queries for distillation. We use Mean Squared Error (MSE) as the distillation loss to measure the difference between the constructed PV prompts and the mimicked queries, which can be formulated as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 353, + 543, + 553, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 543, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 353, + 543, + 553, + 576 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i s t i l l}} = \\lambda_ {\\text {d i s t i l l}} \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\tilde {\\boldsymbol {e}} _ {\\mathrm {p v}} ^ {i} - h \\left(\\boldsymbol {e} _ {m} ^ {i}\\right)\\right) ^ {2}, \\tag {10}", + "image_path": "c90c8c6d26fd811a452ef8c7d90106ff8691d6149b0c6ecf55cafe429c386a74.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 582, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 582, + 554, + 640 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{distill}}" + }, + { + "bbox": [ + 313, + 582, + 554, + 640 + ], + "type": "text", + "content": " is a weight coefficient that balances the impact of the distillation loss on the overall training process. Such loss function guides the mimic queries to accurately replicate the PV prompts generated by the PV branch, making them reliable substitutes for inference." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": "To this end, only the mimic queries can be utilized during inference, which eliminates the requirement for PV prompt computation. This approach maintains real-time inference capability by efficiently mimicking PV structural features with minimal computational cost, enabling streamlined yet robust map construction." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22363" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 553, + 222 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 553, + 222 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 553, + 222 + ], + "type": "table", + "html": "
MethodBackboneEpochsRegion-BasedCity-BasedFPS
APpedAPdivAPboumAPAPpedAPdivAPboumAP
VectorMapNet* [ICML'23]R5012015.817.021.218.0-----
MapTR* [ICLR'23]R50246.420.735.520.911.112.921.115.015.8
MGMap [CVPR'24]R50249.427.038.124.810.917.824.717.811.5
MapQR [ECCV'24]R50248.425.739.924.710.215.825.517.212.1
GeMap [ECCV'24]R502411.826.643.627.312.916.326.718.611.6
MapTRv2* [IJCV'24]R502414.728.743.328.918.918.627.921.812.9
StreamMapNet* [WACV'24]R502432.229.340.834.118.716.422.719.313.3
UIGenMap-d (Ours)R502440.330.846.839.3 (+5.2)23.119.625.422.7 (+3.4)12.2
UIGenMap (Ours)R502440.331.847.239.8 (+5.7)24.120.726.123.6 (+4.3)8.2
StreamMapNet [WACV'24]SwinT2432.527.441.733.921.819.725.822.49.3
UIGenMap-d (Ours)SwinT2442.631.947.240.6 (+6.7)24.721.228.824.9 (+2.5)8.5
", + "image_path": "87d1c68ff54e10d012ce3dba3df427d174d4e973f536a7c94475c245044bc48b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 223, + 555, + 268 + ], + "lines": [ + { + "bbox": [ + 55, + 223, + 555, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 555, + 268 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons on nuScenes validation dataset, utilizing region-based and city-based dataset partitions. Methods with “*” denote that data results are obtained from corresponding papers [27, 50], and others are reproduced by official public codes. “R50” and “SwinT” denotes the image backbone of ResNet50 [9] and Swin-Tiny [30]. “UIGenMap-d” represents the utilization of mimic query from MQ-Distillation. FPS is measured on a single RTX A40 GPU." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 276, + 187, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 187, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 187, + 289 + ], + "type": "text", + "content": "3.4. Training and Inference" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 293, + 295, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 295, + 331 + ], + "type": "text", + "content": "During the training process, combined with " + }, + { + "bbox": [ + 55, + 293, + 295, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{nl}}" + }, + { + "bbox": [ + 55, + 293, + 295, + 331 + ], + "type": "text", + "content": " for uncertainty output and " + }, + { + "bbox": [ + 55, + 293, + 295, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{distill}}" + }, + { + "bbox": [ + 55, + 293, + 295, + 331 + ], + "type": "text", + "content": " for mimic query distillation, the overall loss setting " + }, + { + "bbox": [ + 55, + 293, + 295, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{map}}" + }, + { + "bbox": [ + 55, + 293, + 295, + 331 + ], + "type": "text", + "content": " can be represented as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 337, + 295, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 337, + 295, + 352 + ], + "spans": [ + { + "bbox": [ + 93, + 337, + 295, + 352 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a p}} = \\lambda_ {1} \\mathcal {L} _ {\\text {p t s}} + \\lambda_ {2} \\mathcal {L} _ {\\text {c l s}} + \\mathcal {L} _ {\\text {n l l}} + \\mathcal {L} _ {\\text {d i s t i l l}}, \\tag {11}", + "image_path": "8785aa7f55c9d4b1fd4b9dca1829767e57839a07d381c0b2694d7ed0815f3fb8.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 356, + 295, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 356, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 295, + 415 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 356, + 295, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pts}}" + }, + { + "bbox": [ + 55, + 356, + 295, + 415 + ], + "type": "text", + "content": " denotes the point-to-point Manhattan distance between the regressed points and the ground-truth. " + }, + { + "bbox": [ + 55, + 356, + 295, + 415 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{cls}}" + }, + { + "bbox": [ + 55, + 356, + 295, + 415 + ], + "type": "text", + "content": " represents the focal loss for map classification. " + }, + { + "bbox": [ + 55, + 356, + 295, + 415 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 55, + 356, + 295, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 356, + 295, + 415 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 55, + 356, + 295, + 415 + ], + "type": "text", + "content": " are the corresponding loss weights for point regression and map classification, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 415, + 295, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 513 + ], + "type": "text", + "content": "For inference, the well-learned mimic query can only be employed as a substitute for PV prompts, avoiding the computational burden for the PV detection branch. Thus, our approach ensures real-time inference without introducing unnecessary computational overhead, while preserving effective performance. Additionally, the model's uncertainty learning enables dynamic sampling during inference, enhancing the adaptability to diverse driving environments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 521, + 137, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 137, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 137, + 534 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 540, + 177, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 177, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 177, + 552 + ], + "type": "text", + "content": "4.1. Datasets and Metrics" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 557, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 689 + ], + "type": "text", + "content": "Datasets. We conduct extensive experiments on widely used autonomous driving datasets, including nuScenes [1] and Argoverse2 [46]. nuScenes dataset comprises 1,000 driving scenes from Boston and Singapore. Each scene lasts approximately 20 seconds with a " + }, + { + "bbox": [ + 55, + 557, + 295, + 689 + ], + "type": "inline_equation", + "content": "2\\mathrm{Hz}" + }, + { + "bbox": [ + 55, + 557, + 295, + 689 + ], + "type": "text", + "content": " sampling rate. For each sample, there are six surrounding images and 32-beam LiDAR point clouds. Argoverse2 contains 1,000 scenes from 6 cities, with each scene capturing 15 seconds of data at " + }, + { + "bbox": [ + 55, + 557, + 295, + 689 + ], + "type": "inline_equation", + "content": "20\\mathrm{Hz}" + }, + { + "bbox": [ + 55, + 557, + 295, + 689 + ], + "type": "text", + "content": " from 7 cameras and " + }, + { + "bbox": [ + 55, + 557, + 295, + 689 + ], + "type": "inline_equation", + "content": "10\\mathrm{Hz}" + }, + { + "bbox": [ + 55, + 557, + 295, + 689 + ], + "type": "text", + "content": " LiDAR sweeps. In our experiments, the sampling frequency for Argoverse2 is unified into " + }, + { + "bbox": [ + 55, + 557, + 295, + 689 + ], + "type": "inline_equation", + "content": "2\\mathrm{Hz}" + }, + { + "bbox": [ + 55, + 557, + 295, + 689 + ], + "type": "text", + "content": " as in StreamMapNet [50]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Geo-based Data Partitions. Considering the overlap of widely used training and validation partitions, we adopt" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 277, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 456 + ], + "type": "text", + "content": "geographically disjoint (geo-based) dataset partitions as benchmarks to evaluate the generalization capability of the presented model. Experiments are conducted on both region-based and city-based data partitions. The region-based partition is adopted from [50], where the driving locations geographically disjoin the training and validation sets. In the city-based data partition, we apply city-wise splits in [27], where the training and validation data are from different cities with a larger distribution change. On nuScenes, Boston and Onenorth are used for training, while Queenstown and Holland Village are used for validation. For Ar-goverse2, Miami and Pittsburgh are used in training, and the remains are employed for validation. For different dataset partitions, models are separately trained on that split's training set and evaluated on their corresponding validation set." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 457, + 556, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 556, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 556, + 567 + ], + "type": "text", + "content": "Evaluation Metrics. In line with the existing works [26, 50], we conduct experiments on three static map elements, namely lane dividers, pedestrian crossings, and road boundaries. Perception range is set to " + }, + { + "bbox": [ + 313, + 457, + 556, + 567 + ], + "type": "inline_equation", + "content": "[-15.0m, 15.0m]" + }, + { + "bbox": [ + 313, + 457, + 556, + 567 + ], + "type": "text", + "content": " for X-axis and " + }, + { + "bbox": [ + 313, + 457, + 556, + 567 + ], + "type": "inline_equation", + "content": "[-30.0m, 30.0m]" + }, + { + "bbox": [ + 313, + 457, + 556, + 567 + ], + "type": "text", + "content": " for Y-axis. We use average precision as the evaluation metric, which calculates the Chamfer distance between predictions and ground truths. The mean Average Precision (mAP) is obtained by averaging AP across the distinct thresholds: " + }, + { + "bbox": [ + 313, + 457, + 556, + 567 + ], + "type": "inline_equation", + "content": "\\{0.5m, 1m, 1.5m\\}" + }, + { + "bbox": [ + 313, + 457, + 556, + 567 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 575, + 448, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 448, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 448, + 588 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "For fair comparisons, we employ ResNet50 [9] as the backbone. As in [50], the size of BEV features is set to " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "50 \\times 100" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": ". The number of map queries and query dimensions are set to 100 and 512. We use AdamW optimizer with a learning rate of " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "4e^{-4}" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": ". Training epochs are limited to prevent potential overfitting. All models are trained on 4 NVIDIA Tesla V100 GPUs with a batch size of 4. The weighted factors " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " are set to 50.0 and 5.0, separately. " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{nil}}" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{distill}}" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " are set to 0.05 and 10.0. We adopt a pre-trained PV detection branch for our UIGenMap training process. More details" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22364" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 553, + 186 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 553, + 186 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 553, + 186 + ], + "type": "table", + "html": "
MethodBackboneEpochsRegion-BasedCity-BasedFPS
\\(AP_{ped}\\)\\(AP_{div}\\)\\(AP_{bou}\\)mAP\\(AP_{ped}\\)\\(AP_{div}\\)\\(AP_{bou}\\)mAP
\\(VectorMapNet^{*}\\)[ICML'23]R5012035.634.937.836.1-----
\\(MapTR^{*}\\)[ICLR'23]R503048.150.455.051.134.741.734.737.915.8
\\(GeMap\\)[ECCV'24]R503052.559.258.256.634.541.140.438.711.6
\\(MapTRv2^{*}\\)[JCV'24]R503053.358.758.856.937.442.241.940.512.9
\\(MapQR\\)[ECCV'24]R503054.258.060.457.533.841.541.638.912.1
\\(StreamMapNet^{*}\\)[WACV'24]R503057.955.761.358.340.243.443.342.313.3
\\(UIGenMap-d\\)(Ours)R503058.358.462.459.7 (+1.4)42.944.444.744.0 (+1.7)12.2
\\(UIGenMap\\)(Ours)R503059.858.762.760.4 (+2.1)43.745.745.144.8 (+2.5)8.2
", + "image_path": "f0a0b16b1b793824fb076adc41a1021d42430872f8c4563aecf4930ed6e1f3cf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 228, + 297, + 320 + ], + "blocks": [ + { + "bbox": [ + 55, + 187, + 553, + 222 + ], + "lines": [ + { + "bbox": [ + 55, + 187, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 187, + 553, + 222 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons on Argoverse2 validation datasets, utilizing region-based and city-based dataset partitions. Methods with “*” denote that data results are obtained from corresponding papers [27, 50], and others are reproduced by official public codes. Models are trained for 30 epochs and follow the same setting. FPS is measured on a single RTX A40 GPU." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 228, + 297, + 320 + ], + "lines": [ + { + "bbox": [ + 57, + 228, + 297, + 320 + ], + "spans": [ + { + "bbox": [ + 57, + 228, + 297, + 320 + ], + "type": "table", + "html": "
DatasetMethodEpochs\\(AP_{ped}\\)\\(AP_{div}\\)\\(AP_{bou}\\)mAP
nuScenesMapTRV2 [26]2459.862.462.461.5
StreamMapNet [50]2460.461.958.960.4
SQD-MapNet [43]2463.065.563.363.9
UIGenMap(Ours)2464.265.464.964.8
Argoverse2StreamMapNet [50]3062.959.563.061.5
SQD-MapNet [43]3064.960.264.963.3
UIGenMap(Ours)3064.560.866.163.8
", + "image_path": "72381be7d513c296c77390d45546205b8e92454ee30e29826efe313a14baa39d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 57, + 350, + 293, + 438 + ], + "blocks": [ + { + "bbox": [ + 55, + 324, + 295, + 347 + ], + "lines": [ + { + "bbox": [ + 55, + 324, + 295, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 295, + 347 + ], + "type": "text", + "content": "Table 3. Performance on nuScenes and Argoverse2 datasets with the original dataset partition at " + }, + { + "bbox": [ + 55, + 324, + 295, + 347 + ], + "type": "inline_equation", + "content": "{30}\\mathrm{\\;m}" + }, + { + "bbox": [ + 55, + 324, + 295, + 347 + ], + "type": "text", + "content": " range." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 350, + 293, + 438 + ], + "lines": [ + { + "bbox": [ + 57, + 350, + 293, + 438 + ], + "spans": [ + { + "bbox": [ + 57, + 350, + 293, + 438 + ], + "type": "table", + "html": "
bkbUI2DPromptUA-DecoderAPpedAPdivAPboumAP
32.229.340.834.1
✓39.928.643.837.4
✓36.131.344.037.1
✓✓41.629.744.838.7
✓✓41.429.847.439.5
✓✓39.230.146.238.5
✓✓✓40.331.847.239.8
", + "image_path": "85edb8fc76eefbefc43b11e48a4c76780b057157945aa7b952438728cad5edc4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 441, + 295, + 464 + ], + "lines": [ + { + "bbox": [ + 55, + 441, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 295, + 464 + ], + "type": "text", + "content": "Table 4. Ablations on the impact of component designs, 'bkb' means the shared image backbone from PV branches." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 472, + 237, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 237, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 237, + 484 + ], + "type": "text", + "content": "are provided in the supplementary materials." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 491, + 141, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 141, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 141, + 504 + ], + "type": "text", + "content": "4.3. Main Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 510, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 510, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 54, + 510, + 295, + 664 + ], + "type": "text", + "content": "Results on nuScenes. We compare the generalization capability with previous methods [25, 26, 28, 31, 50, 54] on geospatial data partitions, in which the image size is set to " + }, + { + "bbox": [ + 54, + 510, + 295, + 664 + ], + "type": "inline_equation", + "content": "480 \\times 800" + }, + { + "bbox": [ + 54, + 510, + 295, + 664 + ], + "type": "text", + "content": ". As shown in Table 1, our UIGenMap outperforms previous methods at a large margin. For the ResNet50 backbone, UIGenMap achieves " + }, + { + "bbox": [ + 54, + 510, + 295, + 664 + ], + "type": "inline_equation", + "content": "5.7\\mathrm{mAP}" + }, + { + "bbox": [ + 54, + 510, + 295, + 664 + ], + "type": "text", + "content": " improvement on the region-based partition, and " + }, + { + "bbox": [ + 54, + 510, + 295, + 664 + ], + "type": "inline_equation", + "content": "4.3\\mathrm{mAP}" + }, + { + "bbox": [ + 54, + 510, + 295, + 664 + ], + "type": "text", + "content": " gain is obtained on the city-based splits. For more efficient design, UIGenMapd also archives " + }, + { + "bbox": [ + 54, + 510, + 295, + 664 + ], + "type": "inline_equation", + "content": "39.3\\mathrm{mAP}" + }, + { + "bbox": [ + 54, + 510, + 295, + 664 + ], + "type": "text", + "content": " with a faster inference speed. Experiments on the Swin-T [30] backbone show significant performance with " + }, + { + "bbox": [ + 54, + 510, + 295, + 664 + ], + "type": "inline_equation", + "content": "40.6\\mathrm{mAP}" + }, + { + "bbox": [ + 54, + 510, + 295, + 664 + ], + "type": "text", + "content": " achievement. Furthermore, we show some visual examples in Figure 5, indicating more intuitive achievements for UIGenMap." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 665, + 295, + 713 + ], + "type": "text", + "content": "Results on Argoverse2. More experiments are conducted on Argoverse2 dataset. The image size is " + }, + { + "bbox": [ + 54, + 665, + 295, + 713 + ], + "type": "inline_equation", + "content": "608 \\times 608" + }, + { + "bbox": [ + 54, + 665, + 295, + 713 + ], + "type": "text", + "content": ", and the sampling rate is " + }, + { + "bbox": [ + 54, + 665, + 295, + 713 + ], + "type": "inline_equation", + "content": "2\\mathrm{Hz}" + }, + { + "bbox": [ + 54, + 665, + 295, + 713 + ], + "type": "text", + "content": ". For a fair comparison, the models are trained for 30 epochs. As shown in Table 2, UIGen" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 231, + 555, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 555, + 302 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 555, + 302 + ], + "type": "text", + "content": "Map achieves the best results of " + }, + { + "bbox": [ + 313, + 231, + 555, + 302 + ], + "type": "inline_equation", + "content": "60.4\\mathrm{mAP}" + }, + { + "bbox": [ + 313, + 231, + 555, + 302 + ], + "type": "text", + "content": " on region-based data partitions and " + }, + { + "bbox": [ + 313, + 231, + 555, + 302 + ], + "type": "inline_equation", + "content": "43.4\\mathrm{mAP}" + }, + { + "bbox": [ + 313, + 231, + 555, + 302 + ], + "type": "text", + "content": " on city-based splits, which performs better than previous approaches. Upon further investigation, calibration errors and dataset distribution also affect model performance. More visualization can be found in our supplementary materials." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 303, + 556, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 556, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 556, + 399 + ], + "type": "text", + "content": "Results on Original Data Partitions. For thorough evaluation, we conducted experiments on the original data partitions with similar driving scenarios, which are widely used in previous studies. As shown in Table 3, we compare our methods with others utilizing the same baseline. UIGenMap still outperforms the existing approaches, which outperforms StreamMapNet with " + }, + { + "bbox": [ + 313, + 303, + 556, + 399 + ], + "type": "inline_equation", + "content": "+4.4" + }, + { + "bbox": [ + 313, + 303, + 556, + 399 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 313, + 303, + 556, + 399 + ], + "type": "inline_equation", + "content": "+2.3" + }, + { + "bbox": [ + 313, + 303, + 556, + 399 + ], + "type": "text", + "content": " mAP separately, validating the effectiveness of our approach." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 407, + 416, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 416, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 416, + 419 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 426, + 555, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 485 + ], + "type": "text", + "content": "In this section, ablations are conducted to verify the effectiveness of our designs. We perform experiments on nuScenes with region-based partition using ResNet50 backbone. All models are trained for 24 epochs. More ablations are provided in the supplementary materials." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 485, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 554, + 605 + ], + "type": "text", + "content": "Contributions of Main Components. Table 4 demonstrates the impact of each component in our design. Based on the pre-trained PV backbone, we design the UA-Decoder for more reliable uncertainty output and UI2DPrompt to dynamically inject PV structural information, which achieves " + }, + { + "bbox": [ + 313, + 485, + 554, + 605 + ], + "type": "inline_equation", + "content": "+2.7" + }, + { + "bbox": [ + 313, + 485, + 554, + 605 + ], + "type": "text", + "content": " mAP and " + }, + { + "bbox": [ + 313, + 485, + 554, + 605 + ], + "type": "inline_equation", + "content": "+2.4" + }, + { + "bbox": [ + 313, + 485, + 554, + 605 + ], + "type": "text", + "content": " mAP improvement, respectively. The combination of these designs gains " + }, + { + "bbox": [ + 313, + 485, + 554, + 605 + ], + "type": "inline_equation", + "content": "+4.0" + }, + { + "bbox": [ + 313, + 485, + 554, + 605 + ], + "type": "text", + "content": " mAP total. The ablations are conducted on UA-Decoder at BEV space. Besides, the pre-trained PV backbone further emphasizes the effectiveness to " + }, + { + "bbox": [ + 313, + 485, + 554, + 605 + ], + "type": "inline_equation", + "content": "39.8" + }, + { + "bbox": [ + 313, + 485, + 554, + 605 + ], + "type": "text", + "content": " mAP for better scene understanding." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "type": "text", + "content": "Design of UA-Decoder. We design the UA-Attention for probabilistic attention and the UA-Head to construct the uncertainty output from the instance and finer point levels. Table 5 shows the ablation without other modules. Our uncertainty-aware design supports dynamic adaptation to achieve more robust performance, which achieves " + }, + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "type": "inline_equation", + "content": "+3.0" + }, + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "type": "text", + "content": " mAP improvements based on systematic combination." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Design for UI2DPrompt. UI2DPrompt consists of prompt construction and a hybrid injection module. Table 6 shows" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22365" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 71, + 347, + 269 + ], + "blocks": [ + { + "bbox": [ + 107, + 71, + 347, + 269 + ], + "lines": [ + { + "bbox": [ + 107, + 71, + 347, + 269 + ], + "spans": [ + { + "bbox": [ + 107, + 71, + 347, + 269 + ], + "type": "image", + "image_path": "55491a0f775a128dc58c15508ee62993ff71f113d1dc12b8ded035fe71d74f4b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 268, + 277, + 277 + ], + "lines": [ + { + "bbox": [ + 159, + 268, + 277, + 277 + ], + "spans": [ + { + "bbox": [ + 159, + 268, + 277, + 277 + ], + "type": "text", + "content": "Surrounding Views with PV Detection" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 347, + 72, + 501, + 269 + ], + "blocks": [ + { + "bbox": [ + 347, + 72, + 501, + 269 + ], + "lines": [ + { + "bbox": [ + 347, + 72, + 501, + 269 + ], + "spans": [ + { + "bbox": [ + 347, + 72, + 501, + 269 + ], + "type": "image", + "image_path": "1353fbd0974062a11dcd003a74e21b249732bba0d5f9e1fe3839ca7360f4af46.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 268, + 388, + 276 + ], + "lines": [ + { + "bbox": [ + 360, + 268, + 388, + 276 + ], + "spans": [ + { + "bbox": [ + 360, + 268, + 388, + 276 + ], + "type": "text", + "content": "Baseline" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 465, + 268, + 477, + 275 + ], + "lines": [ + { + "bbox": [ + 465, + 268, + 477, + 275 + ], + "spans": [ + { + "bbox": [ + 465, + 268, + 477, + 275 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 279, + 550, + 291 + ], + "lines": [ + { + "bbox": [ + 58, + 279, + 550, + 291 + ], + "spans": [ + { + "bbox": [ + 58, + 279, + 550, + 291 + ], + "type": "text", + "content": "Figure 5. Qualitative results on the nuScenes dataset with the region-based data partition. Uncertainty outputs are represented as circles." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 83, + 297, + 269, + 357 + ], + "blocks": [ + { + "bbox": [ + 83, + 297, + 269, + 357 + ], + "lines": [ + { + "bbox": [ + 83, + 297, + 269, + 357 + ], + "spans": [ + { + "bbox": [ + 83, + 297, + 269, + 357 + ], + "type": "table", + "html": "
UA-Att.UA-HeadAPpedAPdivAPboumAP
32.229.340.834.1
✓31.730.442.935.0
✓34.830.543.436.3
✓✓36.131.344.037.1
", + "image_path": "35bba4c96e67f47fb67ecfa2bc4d8ed42d54b67aea630cfa5b9e6a0492cfc17f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 85, + 399, + 268, + 460 + ], + "blocks": [ + { + "bbox": [ + 55, + 361, + 295, + 394 + ], + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 394 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 394 + ], + "type": "text", + "content": "Table 5. Ablations on UA-Decoder without other modules. UAAtt and UA-Head represent the utilization of uncertainty-aware attention and output head respectively." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 399, + 268, + 460 + ], + "lines": [ + { + "bbox": [ + 85, + 399, + 268, + 460 + ], + "spans": [ + { + "bbox": [ + 85, + 399, + 268, + 460 + ], + "type": "table", + "html": "
Coords. Trans.Unc. UtilizationmAP
//37.6
IPMembedding only39.3
IPMweight & embedding39.8
3D Trans.weight & embedding38.7
", + "image_path": "bb1e8c656fed9e0e7bdbc4fe46d5379f5ba84c597befe8606519906236d2a9a1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 507, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 295, + 590 + ], + "type": "text", + "content": "the strategy selection for constructing PV prompts. Compared to a depth-based 3D transition, IPM is more suitable for coordinating the transformation based on the assumption of a flattened road surface. When uncertainty is applied both as dynamic weighting and as learning-based embeddings, performance is further enhanced by reliable PV prompts, achieving an improvement of " + }, + { + "bbox": [ + 55, + 507, + 295, + 590 + ], + "type": "inline_equation", + "content": "+2.2\\mathrm{mAP}" + }, + { + "bbox": [ + 55, + 507, + 295, + 590 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 592, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 295, + 664 + ], + "type": "text", + "content": "Table 7 demonstrates the impact of the hybrid injection design, with a total performance of " + }, + { + "bbox": [ + 55, + 592, + 295, + 664 + ], + "type": "inline_equation", + "content": "39.8\\mathrm{mAP}" + }, + { + "bbox": [ + 55, + 592, + 295, + 664 + ], + "type": "text", + "content": " for combining prompt-to-BEV and prompt-to-query attention. This result confirms the optimal utilization of reliable structural information, effectively compensating for information loss and errors introduced in the basic learning process." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Ablations on MQ-Distillation. To verify the effectiveness of the distillation module, we conduct ablations on the hyperparameter of distillation loss weight " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{distill}}" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": ", as this affects the imitation learning of PV structural features. As" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 339, + 297, + 531, + 357 + ], + "blocks": [ + { + "bbox": [ + 55, + 463, + 295, + 497 + ], + "lines": [ + { + "bbox": [ + 55, + 463, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 295, + 497 + ], + "type": "text", + "content": "Table 6. Ablations on the construction of PV prompts. \"3D Trans.\" means the depth-based transformation. The first row with " + }, + { + "bbox": [ + 55, + 463, + 295, + 497 + ], + "type": "inline_equation", + "content": "\\frac{1}{r}" + }, + { + "bbox": [ + 55, + 463, + 295, + 497 + ], + "type": "text", + "content": " indicates direct PV coordinate embedding." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 339, + 297, + 531, + 357 + ], + "lines": [ + { + "bbox": [ + 339, + 297, + 531, + 357 + ], + "spans": [ + { + "bbox": [ + 339, + 297, + 531, + 357 + ], + "type": "table", + "html": "
P2BEVP2QAPpedAPdivAPboumAP
39.330.445.538.4
✓39.931.146.039.2
✓39.630.746.138.8
✓✓40.331.847.239.8
", + "image_path": "1371390fca46ceb8263cb3c1796ecbd8703ef199dec52ee3fe4ee5a78ccf21f5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 345, + 387, + 524, + 419 + ], + "blocks": [ + { + "bbox": [ + 313, + 361, + 554, + 383 + ], + "lines": [ + { + "bbox": [ + 313, + 361, + 554, + 383 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 554, + 383 + ], + "type": "text", + "content": "Table 7. Ablations on hybrid injection. P2BEV and P2Q mean the cross-attention for " + }, + { + "bbox": [ + 313, + 361, + 554, + 383 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_{\\mathrm{bev}}" + }, + { + "bbox": [ + 313, + 361, + 554, + 383 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 361, + 554, + 383 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{map}}" + }, + { + "bbox": [ + 313, + 361, + 554, + 383 + ], + "type": "text", + "content": " separately." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 345, + 387, + 524, + 419 + ], + "lines": [ + { + "bbox": [ + 345, + 387, + 524, + 419 + ], + "spans": [ + { + "bbox": [ + 345, + 387, + 524, + 419 + ], + "type": "table", + "html": "
λdistill571015/
mAP37.538.239.338.839.8
", + "image_path": "816f98bf545daf2f9b2a456fabc8f0d59ee24b843fded633758a719c6942bf8c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 422, + 554, + 445 + ], + "lines": [ + { + "bbox": [ + 313, + 422, + 554, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 554, + 445 + ], + "type": "text", + "content": "Table 8. Ablations on " + }, + { + "bbox": [ + 313, + 422, + 554, + 445 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{distill}}" + }, + { + "bbox": [ + 313, + 422, + 554, + 445 + ], + "type": "text", + "content": " for MQ-Distillation Design. The last column refers to the use of the best teacher of PV prompts." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 452, + 554, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 554, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 554, + 489 + ], + "type": "text", + "content": "illustrated in Table 8, the best performance is achieved by setting the weight to 10.0, which refers to the best learning process of the mimic query for explicit structural features." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 501, + 479, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 479, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 479, + 513 + ], + "type": "text", + "content": "5. Conclusion and Future Work" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 521, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 665 + ], + "type": "text", + "content": "In this paper, we propose UIGenMap, an uncertainty-instructed structure injection approach to achieve a more generalized HD map construction. Specifically, we introduce the PV detection branch to obtain explicit structural features, in which the UA-Decoder produces adaptive probabilistic outputs. The UI2DPrompt module is proposed to construct reliable PV prompts and achieve dynamic integration through the hybrid injection designed. For real-time inference, MQ-Distillation is designed to mimic the structural prompts of PV. Comprehensive experiments on the challenging geo-based partitions of nuScenes and Argoverse2 datasets demonstrate the generalization of our approach." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "For limitations, current methods are designed and evaluated primarily on specific datasets. In the future, developing larger and more diverse datasets and benchmarks across various locations remains a crucial area for exploration." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22366" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "content": "This work is supported by National Natural Science Foundation of China under Grant No.62376244. It is also supported by Information Technology Center and State Key Lab of CAD&CG, Zhejiang University." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 168, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 168, + 296, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 168, + 296, + 223 + ], + "spans": [ + { + "bbox": [ + 61, + 168, + 296, + 223 + ], + "type": "text", + "content": "[1] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11621-11631, 2020. 1, 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 224, + 296, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 224, + 296, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 296, + 257 + ], + "type": "text", + "content": "[2] Anh-Quan Cao, Angela Dai, and Raoul de Charette. Pasco: Urban 3d panoptic scene completion with uncertainty awareness. In CVPR, pages 14554–14564, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 258, + 295, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 290 + ], + "type": "text", + "content": "[3] Jie Chang, Zhonghao Lan, Changmao Cheng, and Yichen Wei. Data uncertainty learning in face recognition. In CVPR, pages 5710-5719, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 292, + 294, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 292, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 292, + 294, + 334 + ], + "type": "text", + "content": "[4] Jiacheng Chen, Yuefan Wu, Jiaqi Tan, Hang Ma, and Yasutaka Furukawa. Maptracker: Tracking with strided memory fusion for consistent vector hd mapping. In ECCV, pages 90-107. Springer, 2025. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 335, + 295, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 295, + 368 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 295, + 368 + ], + "type": "text", + "content": "[5] Sehwan Choi, Jungho Kim, Hongjae Shin, and Jun Won Choi. Mask2map: Vectorized hd map construction using bird's eye view segmentation masks. In ECCV, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 369, + 294, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 294, + 402 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 294, + 402 + ], + "type": "text", + "content": "[6] Wenjie Ding, Limeng Qiao, Xi Qiu, and Chi Zhang. Pivotnet: Vectorized pivot learning for end-to-end hd map construction. In ICCV, pages 3672-3682, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 403, + 294, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 294, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 294, + 445 + ], + "type": "text", + "content": "[7] Xunjiang Gu, Guanyu Song, Igor Gilitschenski, Marco Pavone, and Boris Ivanovic. Producing and leveraging online map uncertainty in trajectory prediction. In CVPR, pages 14521-14530, 2024. 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 447, + 294, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 294, + 479 + ], + "type": "text", + "content": "[8] Hongji Guo, Hanjing Wang, and Qiang Ji. Uncertainty-guided probabilistic transformer for complex action recognition. In CVPR, pages 20052-20061, 2022. 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 480, + 294, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 294, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 294, + 513 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 514, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 294, + 557 + ], + "type": "text", + "content": "[10] Yuzhe He, Shuang Liang, Xiaofei Rui, Chengying Cai, and Guowei Wan. Egovm: Achieving precise eg-localization using lightweight vectorized maps. arXiv preprint arXiv:2307.08991, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 559, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 559, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 559, + 294, + 601 + ], + "type": "text", + "content": "[11] Haotian Hu, Fanyi Wang, Yaonong Wang, Laifeng Hu, Jingwei Xu, and Zhiwang Zhang. Admap: Anti-disturbance framework for reconstructing online vectorized hd map. arXiv preprint arXiv:2401.13172, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 603, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 294, + 635 + ], + "type": "text", + "content": "[12] Haoxuanye Ji, Pengpeng Liang, and Erkang Cheng. Enhancing 3d object detection with 2d detection-guided query anchors. In CVPR, pages 21178-21187, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 637, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 637, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 637, + 294, + 689 + ], + "type": "text", + "content": "[13] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, pages 8340-8350, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "text", + "content": "[14] Xiaohui Jiang, Shuai Lin Li, Yingfei Liu, Shihao Wang, Fan Jia, Tiancai Wang, Lijin Han, and Xiangyu Zhang. Far3d:" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Expanding the horizon for surround-view 3d object detection. In AAAI, pages 2561-2569, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 96, + 553, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 126 + ], + "type": "text", + "content": "[15] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? In NeuIPS, 2017. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 129, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 161 + ], + "type": "text", + "content": "[16] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. In NeuIPS, 2017. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 163, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 205 + ], + "type": "text", + "content": "[17] Han Li, Zehao Huang, Zitian Wang, Wenge Rong, Naiyan Wang, and Si Liu. Enhancing 3d lane detection and topology reasoning with 2d lane priors. arXiv preprint arXiv:2406.03105, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "type": "text", + "content": "[18] Qi Li, Yue Wang, Yilun Wang, and Hang Zhao. Hdmapnet: An online hd map construction and evaluation framework. In ICRA, pages 4628-4634. IEEE, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 240, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 282 + ], + "type": "text", + "content": "[19] Siyu Li, Jiacheng Lin, Hao Shi, Jiaming Zhang, Song Wang, You Yao, Zhiyong Li, and Kailun Yang. Dtlmapper: Dual temporal consistent learning for vectorized hd map construction. arXiv preprint arXiv:2405.05518, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "type": "text", + "content": "[20] Siyu Li, Kailun Yang, Hao Shi, Song Wang, You Yao, and Zhiyong Li. Genmapping: Unleashing the potential of inverse perspective mapping for robust online hd map construction. arXiv preprint arXiv:2409.08688, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 328, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 328, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 328, + 553, + 360 + ], + "type": "text", + "content": "[21] Xiaotong Li, Yongxing Dai, Yixiao Ge, Jun Liu, Ying Shan, and Lingyu Duan. Uncertainty modeling for out-of-distribution generalization. In ICLR, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 361, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 553, + 415 + ], + "type": "text", + "content": "[22] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In ECCV, pages 1-18. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 416, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 458 + ], + "type": "text", + "content": "[23] Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In ICCV, pages 14864-14873, 2024. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 460, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 503 + ], + "type": "text", + "content": "[24] Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang, and Zhi Tang. Bevfusion: A simple and robust lidar-camera fusion framework. In NeuIPS, pages 10421-10434, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 505, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 505, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 317, + 505, + 553, + 547 + ], + "type": "text", + "content": "[25] Benchcheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. Maptr: Structured modeling and learning for online vectorized hd map construction. In ICLR, 2022. 1, 2, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "type": "text", + "content": "[26] Benchcheng Liao, Shaoyu Chen, Yunchi Zhang, Bo Jiang, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Maptrv2: An end-to-end framework for online vectorized hd map construction. IJCV, pages 1-23, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 593, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 553, + 635 + ], + "type": "text", + "content": "[27] Adam Lilja, Junsheng Fu, Erik Stenberg, and Lars Hammarstrand. Localization is all you evaluate: Data leakage in online mapping datasets and how to fix it. In CVPR, pages 22150-22159, 2024. 1, 6, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 636, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 679 + ], + "type": "text", + "content": "[28] Xiaolu Liu, Song Wang, Wentong Li, Ruizi Yang, Junbo Chen, and Jianke Zhu. Mgmt: Mask-guided learning for online vectorized hd map construction. In CVPR, pages 14812-14821, 2024. 1, 2, 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[29] Yicheng Liu, Tianyuan Yuan, Yue Wang, Yilun Wang, and Hang Zhao. Vectormapnet: End-to-end vectorized hd map learning. In ICML, pages 22352-22369. PMLR, 2023. 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22367" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[30] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[31] Zihao Liu, Xiaoyu Zhang, Guangwei Liu, Ji Zhao, and Ningyi Xu. Leveraging enhanced queries of point sets for vectorized map construction. arXiv preprint arXiv:2402.17430, 2024. 1, 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 294, + 207 + ], + "type": "text", + "content": "[32] Hao Lu, Yunpeng Zhang, Qing Lian, Dalong Du, and Yingcong Chen. Towards generalizable multi-camera 3d object detection via perspective debiasing. arXiv preprint arXiv:2310.11346, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "text", + "content": "[33] Ao Luo, Fan Yang, Xin Li, Lang Nie, Chunyu Lin, Haoqiang Fan, and Shuaicheng Liu. Gaflow: Incorporating gaussian attention into optical flow. In ICCV, pages 9642–9651, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 255, + 294, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 294, + 289 + ], + "type": "text", + "content": "[34] Jiahuan Pei, Cheng Wang, and György Szarvas. Transformer uncertainty estimation with hierarchical stochastic attention. In AAAI, pages 11147-11155, 2022. 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 290, + 294, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 290, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 56, + 290, + 294, + 323 + ], + "type": "text", + "content": "[35] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, pages 194-210. Springer, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 325, + 294, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 325, + 294, + 356 + ], + "spans": [ + { + "bbox": [ + 56, + 325, + 294, + 356 + ], + "type": "text", + "content": "[36] Fengchun Qiao and Xi Peng. Uncertainty-guided model generalization to unseen domains. In CVPR, pages 6790-6800, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 359, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 359, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 359, + 294, + 392 + ], + "type": "text", + "content": "[37] Limeng Qiao, Wenjie Ding, Xi Qiu, and Chi Zhang. End-to-end vectorized hd-map construction with piecewiseBezier curve. In CVPR, pages 13218-13228, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 394, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 294, + 426 + ], + "type": "text", + "content": "[38] Tixiao Shan and Brendan Englot. Lego-loam: Lightweight and ground-optimized lidar odometry and mapping on variable terrain. In IROS, pages 4758–4765. IEEE, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 428, + 294, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 428, + 294, + 472 + ], + "spans": [ + { + "bbox": [ + 56, + 428, + 294, + 472 + ], + "type": "text", + "content": "[39] Tixiao Shan, Brendan Englot, Drew Meyers, Wei Wang, Carlo Ratti, and Daniela Rus. Lio-sam: Tightly-coupled lidar inertial odometry via smoothing and mapping. In IROS, pages 5135-5142. IEEE, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 475, + 294, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 294, + 517 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 294, + 517 + ], + "type": "text", + "content": "[40] Jingyu Song, Xudong Chen, Liupei Lu, Jie Li, and Katherine A Skinner. Memfusionmap: Working memory fusion for online vectorized hd map construction. arXiv preprint arXiv:2409.18737, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 520, + 286, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 286, + 531 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 286, + 531 + ], + "type": "text", + "content": "[41] A Vaswani. Attention is all you need. In NeuIPS, 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 533, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 294, + 576 + ], + "type": "text", + "content": "[42] Song Wang, Wentong Li, Wenyu Liu, Xiaolu Liu, and Jianke Zhu. Lidar2map: In defense of lidar-based semantic map construction using online camera distillation. In CVPR, pages 5186-5195, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 578, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 294, + 621 + ], + "type": "text", + "content": "[43] Shuo Wang, Fan Jia, Yingfei Liu, Yucheng Zhao, Zehui Chen, Tiancai Wang, Chi Zhang, Xiangyu Zhang, and Feng Zhao. Stream query denoising for vectorized hd map construction. In ECCV, 2024. 2, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "text", + "content": "[44] Song Wang, Zhongdao Wang, Jiawei Yu, Wentong Li, Bailan Feng, Junbo Chen, and Jianke Zhu. Relioco: Towards reliable semantic occupancy prediction via uncertainty learning. arXiv preprint arXiv:2409.18026, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[45] Song Wang, Jiawei Yu, Wentong Li, Wenyu Liu, Xiaolu Liu, Junbo Chen, and Jianke Zhu. Not all voxels are equal: Hardness-aware semantic scene completion with self-distillation. In CVPR, pages 14792-14801, 2024. 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 654 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "type": "text", + "content": "[46] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In NeurIPS, 2021. 1, 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 140, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 173 + ], + "type": "text", + "content": "[47] Qiangqiang Wu, Yan Xia, Jia Wan, and Antoni B Chan. Boosting 3d single object tracking with 2d matching distillation and 3d pre-training. In ECCV, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "text", + "content": "[48] Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, et al. Bevformer v2: Adapting modern image backbones to bird's-eye-view recognition via perspective supervision. In CVPR, pages 17830-17839, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 230, + 553, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 283 + ], + "type": "text", + "content": "[49] Jing Yang, Minyue Jiang, Sen Yang, Xiao Tan, Yingying Li, Errui Ding, Hanli Wang, and Jingdong Wang. Mgmtapnet: Multi-granularity representation learning for end-to-end vectorized hd map construction. arXiv preprint arXiv:2410.07733, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 285, + 553, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 328 + ], + "type": "text", + "content": "[50] Tianyuan Yuan, Yicheng Liu, Yue Wang, Yilun Wang, and Hang Zhao. Streammapnet: Streaming mapping network for vectorized online hd map construction. In WACV, pages 7356-7365, 2024. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 331, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 373 + ], + "type": "text", + "content": "[51] Gongjie Zhang, Jiahao Lin, Shuang Wu, Zhipeng Luo, Yang Xue, Shijian Lu, Zuoguan Wang, et al. Online map vectorization for autonomous driving: A rasterization perspective. In NeuIPS, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 375, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 375, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 375, + 553, + 407 + ], + "type": "text", + "content": "[52] Xiaoyu Zhang, Guangwei Liu, Zihao Liu, Ningyi Xu, Yunhui Liu, and Ji Zhao. Enhancing vectorized map perception with historical rasterized maps. In ECCV, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 409, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 409, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 409, + 553, + 442 + ], + "type": "text", + "content": "[53] Yunpeng Zhang, Wenzhao Zheng, Zheng Zhu, Guan Huang, Jiwen Lu, and Jie Zhou. A simple baseline for multi-camera 3d object detection. In AAAI, pages 3507-3515, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 443, + 553, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 443, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 553, + 485 + ], + "type": "text", + "content": "[54] Zhixin Zhang, Yiyuan Zhang, Xiaohan Ding, Fusheng Jin, and Xiangyu Yue. Online vectorized hd map construction using geometry. arXiv preprint arXiv:2312.03341, 2023. 2, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 487, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 553, + 531 + ], + "type": "text", + "content": "[55] Lijun Zhou, Tao Tang, Pengkun Hao, Zihang He, Kalok Ho, Shuo Gu, Wenbo Hou, Zhihui Hao, Haiyang Sun, Kun Zhan, et al. Ua-track: Uncertainty-aware end-to-end 3d multi-object tracking. arXiv preprint arXiv:2406.02147, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 533, + 553, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 576 + ], + "type": "text", + "content": "[56] Yi Zhou, Hui Zhang, Jiaqian Yu, Yifan Yang, Sangil Jung, Seung-In Park, and ByungIn Yoo. Himap: Hybrid representation learning for end-to-end vectorized hd map construction. In CVPR, pages 15396-15406, 2024. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 577, + 553, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 577, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 577, + 553, + 620 + ], + "type": "text", + "content": "[57] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 1, 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 622, + 553, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 622, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 316, + 622, + 553, + 654 + ], + "type": "text", + "content": "[58] Yanjin Zhu, Xin Zheng, and Jianke Zhu. Mesh-loam: Real-time mesh-based lidar odometry and mapping. TIV, 2024. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22368" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Uncertainty-guided Perturbation for Image Super-Resolution Diffusion Model/10430da4-a80e-4106-a1b5-f1210ade542e_content_list.json b/2025/Uncertainty-guided Perturbation for Image Super-Resolution Diffusion Model/10430da4-a80e-4106-a1b5-f1210ade542e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f44044b8c53340fba1519c644b1f653a009e9415 --- /dev/null +++ b/2025/Uncertainty-guided Perturbation for Image Super-Resolution Diffusion Model/10430da4-a80e-4106-a1b5-f1210ade542e_content_list.json @@ -0,0 +1,1755 @@ +[ + { + "type": "text", + "text": "Uncertainty-guided Perturbation for Image Super-Resolution Diffusion Model", + "text_level": 1, + "bbox": [ + 104, + 130, + 893, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Leheng Zhang Weiyi You Kexuan Shi Shuhang Gu* University of Electronic Science and Technology of China", + "bbox": [ + 266, + 180, + 736, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{lehengzhang12, shuhanggu}@gmail.com", + "bbox": [ + 336, + 219, + 658, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/LabShuHangGU/UPSR", + "bbox": [ + 336, + 236, + 656, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 328, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion-based image super-resolution methods have demonstrated significant advantages over GAN-based approaches, particularly in terms of perceptual quality. Building upon a lengthy Markov chain, diffusion-based methods possess remarkable modeling capacity, enabling them to achieve outstanding performance in real-world scenarios. Unlike previous methods that focus on modifying the noise schedule or sampling process to enhance performance, our approach emphasizes the improved utilization of LR information. We find that different regions of the LR image can be viewed as corresponding to different timesteps in a diffusion process, where flat areas are closer to the target HR distribution but edge and texture regions are farther away. In these flat areas, applying a slight noise is more advantageous for the reconstruction. We associate this characteristic with uncertainty and propose to apply uncertainty estimate to guide region-specific noise level control, a technique we refer to as Uncertainty-guided Noise Weighting. Pixels with lower uncertainty (i.e., flat regions) receive reduced noise to preserve more LR information, therefore improving performance. Furthermore, we modify the network architecture of previous methods to develop our Uncertainty-guided Perturbation SuperResolution (UPSR) model. Extensive experimental results demonstrate that, despite reduced model size and training overhead, the proposed UPSR method outperforms current state-of-the-art methods across various datasets, both quantitatively and qualitatively.", + "bbox": [ + 88, + 318, + 485, + 743 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 772, + 223, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Single image super-resolution (SR), which aims to recover clean the high-resolution (HR) image from its degraded, contaminated low-resolution (LR) counterpart, is a classical problem in the computer vision society. Significant information loss during the degradation process underscores the", + "bbox": [ + 89, + 797, + 483, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5ce0171a3b3d693bfea78f1e1ef31b5ce3268ad5dbc6f5c9d19c2fc6cdbad31d.jpg", + "image_caption": [ + "(a) $\\pmb{y}_0$" + ], + "image_footnote": [], + "bbox": [ + 516, + 285, + 709, + 400 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6523879ee78454190f03b5190d0cca26d6f6f2a1abe839d1dc765e7146f54128.jpg", + "image_caption": [ + "(b) $\\pmb{x}_{T}^{\\mathrm{SR3}} = \\pmb{\\epsilon}$" + ], + "image_footnote": [], + "bbox": [ + 710, + 285, + 903, + 400 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4f5f63e5c0ef62f017072d134c86c0be91a5a13ce95af710fefbc500c022fbc5.jpg", + "image_caption": [ + "(c) $\\mathbf{x}_T^{\\mathrm{ResShift}} = \\mathbf{y}_0 + \\pmb{\\epsilon}$", + "Figure 1. A comparison of initial state setup between different diffusion-based image super-resolution methods, where $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\sigma_{\\mathrm{max}}^2\\mathbf{I})$ . (b) SR3 [27] initiate the diffusion process from pure Gaussian noise, whereas (c) ResShift [44] and (d) our UPSR embed the LR input into the initial noise map. Additionally, we apply uncertainty-guided weighting coefficient $w_{u}(\\pmb{y}_{0})$ to reduce the noise level in flat areas, achieving a more specialized diffusion process for SR to improve performance." + ], + "image_footnote": [], + "bbox": [ + 516, + 416, + 707, + 532 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/53b441a3b3ab62644bab7b1b18b134843afa9381139f2482dce79d8728242380.jpg", + "image_caption": [ + "(d) $\\pmb{x}_T^{\\mathrm{UPSR}} = \\pmb{y}_0 + w_u(\\pmb{y}_0)\\pmb{\\epsilon}$" + ], + "image_footnote": [], + "bbox": [ + 712, + 416, + 905, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "need for robust super-resolution modeling capability to recover missing details and produce visually pleasing results. Existing SR methods have explored a variety of advanced network architectures and complex degradation modeling to improve performance in classical SR [2, 19, 47] and real-world SR [37, 46], respectively.", + "bbox": [ + 511, + 700, + 906, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, diffusion models [8, 12, 30] have demonstrated impressive capability in image synthesis, offering a promising new approach for the real-world image super-resolution task. These methods transform pure Gaussian noise into high-quality images through a predefined Markov chain, and their solid theoretical foundation endows them with exceptional modeling capacity to bridge diverse data distri", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* corresponding author", + "bbox": [ + 107, + 887, + 230, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17980", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "butions. To leverage the modeling capabilities of diffusion models for recovering missing details in LR images, several methods [26, 27] begin the SR process by sampling from a standard Gaussian distribution and gradually refine the noisy inputs into high-quality outcomes. However, starting from pure noise was originally intended for image synthesis tasks, resulting in suboptimal outcomes for superresolution. Additionally, these methods typically require a lengthy sampling process, limiting their practicality for real-world applications.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned issues, ResShift [44] focuses on the construction of the prior distribution, highlighting the importance of LR information during the diffusion process. By embedding the LR image into the initial noise map and progressively recovering the residual between the LR and HR images, ResShift greatly simplifies the diffusion process. Instead of modeling the entire HR image from noise, it only needs to estimate the LR-HR residual, which shortens the sampling process while significantly enhancing super-resolution results. Despite improved performance, challenges remain, as shown in Fig. 1 where most details are obscured by heavy noise. This still poses additional challenges, as exploiting the information from surrounding areas is crucial for recovering missing details in the SR task.", + "bbox": [ + 89, + 246, + 482, + 473 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For the pursuit of designing a more specialized and effective diffusion process for SR to improve performance, we first propose to make better use of the LR information in the prior distribution. In recent studies, little attention has been given to the inherent information in the LR image that flat areas are already close to the target, whereas edge and texture regions are farther away. To leverage this information, we propose to consider different regions in the LR image as being situated at various timesteps of an isotropic diffusion process. This leads to an anisotropic diffusion process, where flat areas are assigned lower noise levels (as $t \\to 0$ ), while edge and texture regions receive larger noise (as $t \\to T$ ). To achieve region-specific noise control, we draw inspiration from uncertainty-driven SR approaches [16, 25] and employ a simple SR network to estimate the uncertainty (variance) of different areas in the input LR image. This uncertainty reflects the difficulty of recovering HR details and is closely related to the distribution disparity between LR and HR, indicating the amount of noise required. We then propose a technique called Uncertainty-guided Noise Weighting, which applies weight coefficients to noise levels in different regions based on their uncertainty estimates. The weight coefficient is positively correlated with the uncertainty estimate, enabling adaptive reduction of noise intensity in flat areas to preserve more details in the initial state of the diffusion process. Equipped with UNW and a modified network architecture, we introduce the UPSR model, which achieves state-of-the-art per", + "bbox": [ + 89, + 477, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "formance while reducing computational costs, as validated on both synthetic and real-world SR datasets.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our major contributions are summarized as follows:", + "bbox": [ + 531, + 122, + 875, + 136 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To achieve a more specialized diffusion pipeline for SR, we propose to adjust the noise level for different areas based on their uncertainty estimate, where the noise in flat areas (low uncertainty) is reduced. This Uncertainty-guided Noise Weighting strategy enables adaptive control across different areas and results in improved performance.", + "- We establish the connection between the residual estimated by a pre-trained SR network and the uncertainty of the LR input. The estimated residual could serve as an approximate measure of uncertainty, enabling effective uncertainty estimation in the proposed uncertainty-guided noise weighting scheme.", + "- By integrating the proposed uncertainty-guided noise weighting scheme and the modified network architecture, our method achieves state-of-the-art performance across various benchmark datasets with smaller model size and less training overhead." + ], + "bbox": [ + 511, + 138, + 903, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 425, + 661, + 440 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image Super-Resolution", + "text_level": 1, + "bbox": [ + 513, + 450, + 733, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The current literature on image super-resolution can be divided into two categories: classical SR and real SR. The former focuses on addressing predefined degradation patterns, e.g., bicubic downsampling. After the era of conventional methods [7, 41, 45], a multitude of works [2, 4, 18, 19, 21, 47, 49] focus on exploring the potential of neural network, ranging from improved convolutional neural networks to transformer variants [33]. These architectural modifications enhance the capacity of modeling complex relationship between LR and HR images, thereby achieving better performance. Despite their success, these methods perform poorly when directly applied to the real SR task, where the degradation model includes a series of unknown and complex noise and blur patterns. BSRGAN [46] incorporates a series of degradation operators to simulate real-world scenarios, providing a robust training baseline. By combining this strategy with a generative adversarial network [6], BSRGAN achieves remarkable fidelity and perceptual quality in the real SR task. Subsequently, RealESRGAN [37] extends this scheme to a high-order degradation model, providing a more robust training data synthesis procedure and better performance.", + "bbox": [ + 511, + 473, + 906, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Diffusion Models", + "text_level": 1, + "bbox": [ + 513, + 816, + 681, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Rooted in nonequilibrium thermodynamics, diffusion probabilistic models [29] have emerged as a new trend in recent research of generative models. DDPM [8] proposes a lengthy parameterized Markov chain to bridge the dis", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17981", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "distribution of high-quality images and the standard Gaussian distribution. Built upon solid theoretic foundations, DDPM simplifies training objectives and achieves a great breakthrough in image synthesis tasks, demonstrating significant generative capacity compared to GAN-based methods [6, 11, 23]. Subsequently, immense amounts of related research have developed various techniques to further improve DDPM by enhancing performance or shortening sampling process. These improvements include deterministic sampling [30], modified noise schedule [24], harnessing latent space [26], second-order ODE solver [12], and improved training dynamics [13].", + "bbox": [ + 89, + 90, + 483, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Diffusion-based Image Super-Resolution", + "text_level": 1, + "bbox": [ + 89, + 279, + 437, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SRDiff [17] and SR3 [27] are the first to apply diffusion models to image super-resolution by taking the LR input as conditional information, demonstrating the efficacy of diffusion model in generating perceptually high-quality SR images. Despite superior performance, SR3 suffers from bias issues and a costly sampling process. LDM-SR [26] trains an autoencoder and performs the diffusion process in its low-dimensional latent space. This approach allows the model to concentrate on perceptually relevant details and significantly improves computational efficiency. ResShift [44] embeds the LR information $\\mathbf{y}_0$ into the initial state $\\mathbf{x}_T$ and built a new prior distribution: $\\mathbf{x}_T \\sim \\mathcal{N}(\\mathbf{x}_T \\mid \\mathbf{y}_0, \\kappa^2 \\eta_T \\mathbf{I})$ . In this way, the diffusion process shifts from generating the HR image $\\mathbf{x}_0$ from pure noise $\\epsilon$ to generating the residual $\\mathbf{x}_0 - \\mathbf{y}_0$ given noisy LR information $\\mathbf{y}_0 + \\epsilon$ , greatly reducing the difficulty of estimation. The forward and backward transition distributions are defined as:", + "bbox": [ + 89, + 301, + 483, + 556 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} q \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {t - 1}, \\boldsymbol {x} _ {0}, \\boldsymbol {y} _ {0}\\right) = (1) \\\\ \\mathcal {N} \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {t - 1} + \\alpha_ {t} \\left(\\boldsymbol {y} _ {0} - \\boldsymbol {x} _ {0}\\right), \\kappa^ {2} \\alpha_ {t} \\boldsymbol {I}\\right) (1) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 560, + 483, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and", + "bbox": [ + 89, + 602, + 120, + 614 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} q \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}, \\boldsymbol {x} _ {0}, \\boldsymbol {y} _ {0}\\right) = \\\\ \\mathcal {N} \\left(\\boldsymbol {x} _ {t - 1} \\mid \\frac {\\eta_ {t - 1}}{\\eta_ {t}} \\boldsymbol {x} _ {t} + \\frac {\\alpha_ {t}}{\\eta_ {t}} \\boldsymbol {x} _ {0}, \\kappa^ {2} \\frac {\\eta_ {t - 1}}{\\eta_ {t}} \\alpha_ {t} \\boldsymbol {I}\\right), \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 617, + 482, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for $t = 1,2,\\dots ,T$ ; $\\eta_t$ and $\\alpha_{t} = \\eta_{t} - \\eta_{t - 1}$ are time-dependent positive parameters that characterize the velocity of mean shift and noise injection at different timesteps. In this work, we adopt the sampling process and network architecture of ResShift as our baseline and propose a series of modifications to the sampling pipeline and network architecture to enhance performance while significantly saving model size and training overhead.", + "bbox": [ + 89, + 691, + 483, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 823, + 225, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Noise Levels in Diffusion-based SR", + "text_level": 1, + "bbox": [ + 89, + 848, + 393, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the context of SDE [31, 32], perturbing the data with Gaussian noise is crucial for facilitating score estima", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d651a5b712f53a3a2c0125b57a1377ec7ba80af9027481e809cd23e1ae853eff.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 583, + 94, + 836, + 204 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bb2fa6b3a956b4dc8aaea47aa8622d1deb4a38037df25e97a3fc45f8923e8955.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 516, + 218, + 709, + 334 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1ae2a69b2d016cf2f269d1794984c7886bf71bc19f31216781a806b90fb7a5a5.jpg", + "image_caption": [ + "Figure 2. (a) The distribution of pixel residual $|y - x|$ computed on ImageNet-Test dataset [44], omitting values where $|y - x| > 0.4$ for clarity. The result exhibits a distinct long-tailed characteristic. (b) The statistical curves of fidelity $|f(y) - x|$ and perceptual quality $|\\phi(f(y)) - \\phi(x)|$ with respect to residual $|y - x|$ under different noise levels. As $|y - x|$ increases, the gap of fidelity remains relatively stable when different noise levels are applied. In contrast, the perceptual quality is more sensitive to the noise level. A larger noise is more requisite in regions with high residual value to achieve better perceptual quality. Meanwhile, we propose weighted noise level $w_u(y)\\sigma_{max}$ which could lead to better results, with details presented in Sec. 3.3." + ], + "image_footnote": [], + "bbox": [ + 709, + 218, + 903, + 334 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion [9], especially for data residing in low-dimensional manifolds. They apply various noise levels to perturb the data to explore the low-density regions and enable the score estimation towards high-density regions. In the super-resolution task, the target is to estimate the gradients from LR images (low density) to HR images (high density) in the real image distribution. Recent diffusion-based methods [43, 44] inject a large initial noise into the LR image and gradually reduce it throughout the reverse diffusion process, as the LR sample progressively moves closer to high-density regions. However, these methods focus solely on the different amounts of noise needed at different timesteps during the diffusion process, without accounting for the fact that different areas of each image require varying noise intensities. To provide a clearer insight, we conduct several experiments to demonstrate how the fidelity and perceptual quality of the output change when different overall noise levels are applied to a diffusion-based SR model.", + "bbox": [ + 511, + 537, + 906, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first analyze the distribution of the residual $|y - x|$ between each pixel pair of the upsampled LR image $y$ and its corresponding HR image $x$ , as illustrated in Fig. 2a. The result clearly indicates that $|y - x|$ follows a long-tailed distribution, with over $95\\%$ of the data concentrated in the range [0.01, 0.16]. Next, we focus on data within this range to", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17982", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/31017c98198a25fbb9bb37d30f00284812426664277c804673b26dfae4239fa9.jpg", + "image_caption": [ + "HR" + ], + "image_footnote": [], + "bbox": [ + 91, + 92, + 187, + 167 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fc7b50b181be7ab0b19f639edd0ca07a47eed10cf9a97e729ee7eee1c411a5fb.jpg", + "image_caption": [ + "LR" + ], + "image_footnote": [], + "bbox": [ + 189, + 92, + 287, + 167 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f6e599077e5024dae8d6d19ee2bbd56835aeb0031c6415ac9b98b961edaa16a2.jpg", + "image_caption": [ + "Real Residual", + "Figure 3. A visualization of the actual residual $|\\pmb{x}^i - \\pmb{y}^i|$ and the estimated residual $|g(\\pmb{y}^i) - \\pmb{y}^i|$ . The real residual exhibits high values in edges and texture regions, indicating the high uncertainty. The residual estimated by SR network is close to the real one and therefore can serve as a rough estimation of uncertainty." + ], + "image_footnote": [], + "bbox": [ + 289, + 93, + 383, + 167 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/dce30906886ae081c3205f27c5141a9bc98c09f9c6715891c15809f4a70e3e3a.jpg", + "image_caption": [ + "Estimated Residual" + ], + "image_footnote": [], + "bbox": [ + 385, + 93, + 480, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "investigate the relationships between the residual $|y - x|$ , fidelity $|f(y) - x|$ , perceptual quality $|\\phi(f(y)) - \\phi(x)|$ of the output of the denoiser $f(\\cdot)$ trained under different noise levels $\\sigma_{max}$ (in a reasonable range). The results are illustrated in Fig. 2b. Due to the blurring and downsampling degradation, the residual $|y - x|$ tends to be low in flat areas and high in regions characterized by edges and textures. As $|y - x|$ grows, the perceptual quality gap between $|\\phi(f(y)) - \\phi(x)|_{\\sigma_{max}=2.5}$ and $|\\phi(f(y)) - \\phi(x)|_{\\sigma_{max}=0.5}$ increases rapidly, while the fidelity gap remains nearly unchanged. This indicates that perceptual quality is more sensitive to the noise level change, with higher noise levels being particularly important in edge and texture areas. These regions lie in the low-density regions of the real image distribution and are affected by the ill-posed nature of superresolution. Applying a low $\\sigma_{max}$ in such areas leads to unreliable score estimation, resulting in perceptually poor, over-smoothed outputs.", + "bbox": [ + 89, + 268, + 483, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The experimental results support our hypothesis that different regions of an image respond differently to noise and require distinct handling based on their content. Flat regions typically closely resemble the corresponding ground truth, requiring only a slight amount of noise; therefore, we consider them as existing at small timesteps $(t\\to 0)$ . Based on this idea, we propose to reduce noise levels applied in flat regions, and the results shown by the red line in Fig. 2b preliminarily verify the effectiveness. In the following subsections, we will dually discuss how to determine the noise level for different types of areas in the LR image.", + "bbox": [ + 89, + 542, + 483, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Uncertainty Estimation", + "text_level": 1, + "bbox": [ + 89, + 724, + 307, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Due to the ill-posed nature of the image super-resolution task, perfectly reconstructing HR images from degraded LR inputs is non-trivial. The high-frequency information in edge and texture areas is severely corrupted during the degradation, exhibiting greater variance compared to flat areas and making accurate prediction more challenging. Several works [16, 25] explore this variance from an uncertainty-based perspective [15]. Given an HR image $\\boldsymbol{x}^i \\in \\mathbb{R}^n$ and its corresponding LR image $\\boldsymbol{y}^i \\in \\mathbb{R}^n$ , the uncertainty $\\psi^i$ of the SR estimate $g(\\boldsymbol{y}^i)$ is related to its resid", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ual with $x^i$ :", + "bbox": [ + 513, + 90, + 596, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} ^ {i} = g \\left(\\boldsymbol {y} ^ {i}\\right) + \\epsilon \\psi \\left(g \\left(\\boldsymbol {y} ^ {i}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 116, + 906, + 135 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon$ represents a standard Laplace or Gaussian distribution when $g(\\cdot)$ is regularized by $L1$ or $L2$ loss function, respectively. If the reconstruction error $|\\pmb{x}^i - g(\\pmb{y}^i)|$ is larger than $|\\pmb{x}^j - g(\\pmb{y}^j)|$ , then it is more likely that $g(\\pmb{y}^i)$ has a higher uncertainty $\\psi(g(\\pmb{y}^i))$ , compared to the uncertainty $\\psi(g(\\pmb{y}^j))$ of $g(\\pmb{y}^j)$ . Similarly, we can associate the uncertainty of $\\pmb{y}^i$ with the residual $|\\pmb{x}^i - \\pmb{y}^i|$ as:", + "bbox": [ + 511, + 143, + 906, + 250 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} ^ {i} = \\boldsymbol {y} ^ {i} + \\hat {\\epsilon} \\psi (\\boldsymbol {y} ^ {i}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 258, + 906, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\epsilon}$ represents an unknown distribution that depends on the degradation pattern. If $g(\\cdot)$ is well-trained, we can assume that $g(\\pmb{y}^i)$ closely approximates $\\pmb{x}^i$ , implying that $|g(\\pmb{y}^i) - \\pmb{y}^i|$ is similar to $|\\pmb{x}^i - \\pmb{y}^i|$ . As illustrated in Fig. 3, the residual estimated by $g(\\cdot)$ is roughly similar to the real residual. Therefore, we propose to leverage the residual $|g(\\pmb{y}^i) - \\pmb{y}^i|$ as the estimate of the uncertainty of $\\pmb{y}^i$ . Specifically, we define the uncertainty estimate as:", + "bbox": [ + 511, + 284, + 906, + 405 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\psi_ {e s t} (\\boldsymbol {y}) = \\frac {1}{2} | g (\\boldsymbol {y}) - \\boldsymbol {y} |. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 412, + 905, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the next subsection, we will take this uncertainty estimate as the criterion to adjust the noise intensity.", + "bbox": [ + 511, + 450, + 905, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Uncertainty-guided Noise Weighting", + "text_level": 1, + "bbox": [ + 511, + 489, + 830, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the discussion in previous sections, we propose to replace the commonly used isotropic Gaussian noise with anisotropic noise that adapts to the image content. Inspired by [25], which deals with areas with higher uncertainty by assigning larger weights to the loss function to impose stronger constraints, we suggest adjusting the noise intensity based on the uncertainty estimate across different regions. Specifically, after obtaining the uncertainty estimate as outlined in Sec. 3.2, we compute the weighting coefficient $w_{u}$ and apply it to modulate noise levels for different areas in the diffusion process. We refer to this strategy as Uncertainty-guided Noise Weighting (UNW). For $y$ with lower uncertainty (e.g. in flat areas), it is more likely that $y$ resides in high-density regions of the HR data distribution. In such cases, a lower noise level is sufficient to achieve perceptual quality comparable to that of a higher noise level, while also preserving more LR details, as discussed in Sec. 3.1. Conversely, if $y$ exhibits high uncertainty (e.g., in edge or texture areas), a higher noise level is requisite to account for the significant distribution disparity and to provide greater chances of reconstructing photorealistic details. Therefore, we model the noise weighting coefficient $w_{u}$ as a monotonically increasing function with respect to the uncertainty:", + "bbox": [ + 511, + 511, + 906, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw _ {u} (\\boldsymbol {y}) := u \\left(\\psi_ {\\text {e s t}} (\\boldsymbol {y})\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 883, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17983", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/174cb11b74ab35961d8efc4ffc7ffad388aa3c9e86fbdac009db44ad3d8b3173.jpg", + "image_caption": [ + "Figure 4. The overall pipeline of the proposed UPSR model. An auxiliary SR network is first employed to estimate the uncertainty of the input $\\mathbf{y}_0$ . Then the weighting coefficient $w_{u}$ computed based on the uncertainty $\\psi_{est}(\\mathbf{y}_0)$ are applied to adjust the noise level in different regions. Meanwhile, both the SR estimate $g(\\mathbf{y}_0)$ and LR input $\\mathbf{y}_0$ are concatenated as the conditional information for the denoiser $f_{\\theta}(\\cdot)$ ." + ], + "image_footnote": [], + "bbox": [ + 200, + 93, + 803, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w_{u}(\\pmb{y}) \\in \\mathbb{R}^{n \\times n}$ represents the diagonal weight matrix used to lower the noise level in low-uncertainty areas while maintaining a higher noise level in high-uncertainty regions, creating a more specialized and adaptive diffusion pipeline for SR. Details of the implementation of $u(\\cdot)$ are presented in the supplementary material.", + "bbox": [ + 88, + 339, + 480, + 430 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Built upon the UNW technique, we introduce a new pipeline for diffusion-based SR, termed Uncertainty-guided Perturbation for SR (UPSR), as illustrated in Fig. 4. Given an LR image $\\pmb{y}_0$ , we first obtain its SR estimate $g(\\pmb{y}_0)$ through an auxiliary SR network $g(\\cdot)$ . Then we estimate the uncertainty of $\\pmb{y}_0$ as $\\psi_{est}(\\pmb{y}_0) = \\frac{1}{2} |g(\\pmb{y}_0) - \\pmb{y}_0|$ and obtain the weighting coefficient as $w_u(\\pmb{y}_0) = u(\\psi_{est}(\\pmb{y}_0))$ . Based on the weighting coefficient, we rewrite the forward transition distribution $q(\\pmb{x}_t \\mid \\pmb{x}_{t-1}, \\pmb{x}_0, \\pmb{y}_0)$ in Eq. 1 as:", + "bbox": [ + 89, + 431, + 483, + 568 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left. \\mathcal {N} \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {t - 1} + \\alpha_ {t} \\left(\\boldsymbol {y} _ {0} - \\boldsymbol {x} _ {0}\\right), \\kappa^ {2} w _ {u} \\left(\\boldsymbol {y} _ {0}\\right) ^ {2} \\alpha_ {t} \\boldsymbol {I}\\right), \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 577, + 482, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and derive the corresponding backward transition distribution $q(\\pmb{x}_{t-1} \\mid \\pmb{x}_t, \\pmb{x}_0, \\pmb{y}_0)$ as:", + "bbox": [ + 89, + 606, + 482, + 638 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {N} \\left(\\boldsymbol {x} _ {t - 1} \\mid \\frac {\\eta_ {t - 1}}{\\eta_ {t}} \\boldsymbol {x} _ {t} + \\frac {\\alpha_ {t}}{\\eta_ {t}} \\boldsymbol {x} _ {0}, \\kappa^ {2} w _ {u} (\\boldsymbol {y} _ {0}) ^ {2} \\frac {\\eta_ {t - 1}}{\\eta_ {t}} \\alpha_ {t} \\boldsymbol {I}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 648, + 482, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the difference is that we leverage the weighting coefficient $w_{u}(\\pmb{y}_{0})$ to control the noise level. Since $g(\\pmb{y}_0)$ is a better estimate of $x_0$ compared to $y_{0}$ , we combine $g(\\pmb{y}_0)$ with $y_{0}$ to provide more accurate conditional information for the denoiser. Details of the derivations of Eq. 7, 8 and the training pipeline are presented in the supplementary materials. Following the previous diffusion-based methods [8, 44], our training objective for the denoiser to predict the target $x_0$ combines both pixel distance $||\\cdot||_2^2$ and LPIPS criterion $L_{per}$ :", + "bbox": [ + 89, + 691, + 482, + 843 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} (\\theta) = \\sum_ {t} [ \\| f _ {\\theta} (\\boldsymbol {x} _ {t}, \\boldsymbol {y} _ {0}, g (\\boldsymbol {y} _ {0}), t) - \\boldsymbol {x} _ {0} \\| _ {2} ^ {2} \\tag {9} \\\\ + \\lambda L _ {p e r} \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {t}, \\boldsymbol {y} _ {0}, g (\\boldsymbol {y} _ {0}), t\\right), \\boldsymbol {x} _ {0}\\right) ], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 852, + 482, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is a hyperparameter to control the trade-off between fidelity and perceptual quality. Optimizing the denoiser $f_{\\theta}(\\cdot)$ through the mixed objective function facilitates fewer diffusion steps while achieving better photo-realistic results in various benchmark real-world datasets.", + "bbox": [ + 511, + 339, + 906, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Network Architecture Modification", + "text_level": 1, + "bbox": [ + 511, + 424, + 821, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Besides altering the noise injection scheme and sampling procedure, we also make several modifications to the network architecture.", + "bbox": [ + 511, + 446, + 905, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We rethink the necessity of leveraging latent space for image super-resolution with only a few diffusion steps. [26] and [44] utilize VQGAN [5] to transfer the diffusion process from pixel space to latent space, for the sake of reducing spatial dimensionality to improve efficiency. However, as the number of sampling steps decreases, the performance-cost ratio of VQGAN gradually decreases due to its own huge computing resource consumption. Moreover, the application of certain perceptual criteria (e.g., LPIPS [48]) requires pixel-level computation, leading to additional decoding costs during the training process. To mitigate its impact, we propose to replace the VQGAN encoder and decoder with PixelUnshuffle operation [28] and a simple nearest neighbor upsampling module, respectively. This approach enables compressing and expanding the spatial dimension with minimal computational cost. Therefore, we can perform the diffusion process directly in pixel space while maintaining a complexity comparable to that in the latent space. In this vein, we significantly reduce the total model size and training overhead without sacrificing performance.", + "bbox": [ + 511, + 492, + 906, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 823, + 645, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setups", + "text_level": 1, + "bbox": [ + 511, + 848, + 710, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For data preparation, we apply randomly cropped $256 \\times 256$ patches from ImageNet [3] as HR training data following", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17984", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5bc0660e2397af8a6b7694c8a72b380f2dfc24c13e5080d60fc45994d572ebee.jpg", + "table_caption": [ + "Table 1. Ablation study on effects of the proposed components, including the SR condition and uncertainty-guided noise weighting (UNW). The best results are highlighted in bold." + ], + "table_footnote": [], + "table_body": "
UNWSR cond.RealSRRealSet
PSNR↑CLIPQA↑MUSIQ↑MANIQA↑NIQE↓CLIPQA↑MUSIQ↑MANIQA↑NIQE↓
26.180.544762.9510.35964.490.614164.3600.37184.42
✓26.120.576064.5120.37174.180.634064.2800.38364.22
✓✓26.440.601064.5410.38184.020.638963.4980.39314.24
", + "bbox": [ + 109, + 130, + 890, + 218 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[26] and [44]. Then the degradation pipeline of RealESR-GAN [37] are adopted to generate degraded $64 \\times 64$ LR inputs. For network architecture, we employ the commonly-used U-Net denoiser $f_{\\theta}(\\cdot)$ with few adjustments to fit our refined diffusion pipeline. For $g(\\cdot)$ , we employ the lightweight version of [47] and pretrain it following the real SR pipeline to produce SR predictions and uncertainty estimates. During the diffusion model training process, we freeze the pretrained $g(\\cdot)$ and optimize $f_{\\theta}(\\cdot)$ for 200k iterations with a batchsize of 32. To quantitatively validate the efficacy of the proposed method, we utilize three full-reference metrics: PSNR, SSIM and LPIPS [48]; and four non-reference metrics CLIPIQA [34], MUSIQ [14], MANIQA [42], and NIQE [22] to evaluate the performance. Among these metrics, PSNR and SSIM reflect the fidelity of the generated super-resolution results, while other metrics assess the perceptual quality of the outputs.", + "bbox": [ + 91, + 234, + 483, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Analysis", + "text_level": 1, + "bbox": [ + 89, + 505, + 194, + 522 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this subsection, we conduct ablation studies on the effectiveness of several key components and analyze the performance-cost trade-off of the proposed method.", + "bbox": [ + 89, + 527, + 483, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of UNW and SR conditioning. In order to show the efficacy of the proposed uncertainty-guided noise weighting scheme and SR conditioning, we use the noise schedule and sampling process of ResShift [44] as the baseline. We first apply the uncertainty-guided noise weighting scheme to adjust the noise intensity across different regions based on the uncertainty estimate. The weighting scheme yields better outcomes in terms of perceptual quality, with enhancements of 0.0355 in CLIPIQA and 1.512 in MUSIQ. We attribute this improvement to the reduced noise intensities applied to flat areas, which allow these clearer regions to provide additional information that aids in more accurately recovering the remaining degraded regions. This validates our idea of reducing noise in flat regions does not degrade perceptual quality, but rather enhances it. Next, we incorporate the SR prediction $g(\\pmb{y}_0)$ and LR input $\\pmb{y}_0$ as the conditional information entered into the denoiser $f_{\\theta}(\\cdot)$ . The additional SR prediction offers more precise supplementary information compared to relying solely on the degraded LR input, resulting in a significant improvement in both fi", + "bbox": [ + 89, + 598, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1e8ea203eb775c33ebcb624b991e02cf40dd4e32bccbfa2d38f0ff1bbb0afa88.jpg", + "table_caption": [ + "Table 2. Model size and computational efficiency comparisons between the proposed UPSR and other diffusion-based methods. Gray numbers denote the parameter counts for auxiliary network, i.e., VQGAN in ResShift and $g(\\cdot)$ in our work. We test the runtimes on $32 \\times 3 \\times 64 \\times 64$ LR input using single RTX4090 GPU and present several results evaluated on the RealSR dataset." + ], + "table_footnote": [], + "table_body": "
ModelParams (M)Runtimes (s)MUSIQMANIQA
LDM-15113.60+55.321.5948.6980.2655
ResShift-15118.59+55.321.9857.7690.3691
ResShift-4118.59+55.321.0055.1890.3337
UPSR-5119.42+2.501.1264.5410.3818
", + "bbox": [ + 517, + 334, + 903, + 412 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1c8609f0990d85f0e69498a1e1dcfe5205787dae23867195198ce58117b26f6e.jpg", + "table_caption": [ + "Table 3. Training overhead comparison between ResShift and UPSR evaluated under a batchsize of 16 per GPU." + ], + "table_footnote": [], + "table_body": "
ModelTraining SpeedMemory Footprint
ResShift1.20 s/iter24.1 G
UPSR0.45 s/iter14.9 G
", + "bbox": [ + 570, + 460, + 846, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "delity and perceptual quality. Specifically, this more robust pipeline further improves PSNR by 0.34 dB and CLIPIQA by 0.0247 in the RealSR dataset. These ablation studies verify the significance of leveraging LR information during diffusion-based SR process.", + "bbox": [ + 511, + 529, + 906, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model Size and Training Overhead Comparison. In section 3.4, we discussed replacing VQGAN in diffusion-based SR model with simpler downsampling and upsampling modules to improve efficiency. To validate the efficacy of the modified architecture, we first evaluate the model size and computational cost of UPSR compared to state-of-the-art diffusion-based methods, as shown in Tab. 2. With the modified architecture and pipeline, UPSR-5 reduces the overall model size by $30\\%$ and achieves better perceptual quality with an inference speed comparable to that of ResShift-4. In contrast, ResShift-4 achieves the acceleration at the expense of performance. We then assess the training overhead of different methods, as shown in Tab. 3. Removing VQGAN results in substantial reductions in training overhead, increasing training speed by $167\\%$ and saving GPU memory footprint by $38\\%$ . These results validate that UPSR strikes a better trade-off between performance and efficiency.", + "bbox": [ + 511, + 628, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17985", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/dbe3b36664d985583725261e14c91ba74bd420ab98667312fdbf165f70ca540b.jpg", + "table_caption": [ + "Table 4. Quantitative results of different methods on one synthetic dataset ImageNet-Test and two real-world datasets RealSR and RealSet. The best and second best results are marked in red and blue. All results of the previous methods are evaluated using the released inference codes and pretrained weights." + ], + "table_footnote": [], + "table_body": "
DatasetsMetricsGAN-based MethodsDiffusion-based Methods
ESRGANRealSR-JPEGBSRGANRealESRGANSwinIRDASRLDM-15ResShift-15ResShift-4UPSR-5
ImageNet -TestPSNR↑20.6723.1124.4224.0423.9924.7524.8524.9425.0223.77
SSIM↑0.44850.59120.65850.66490.66660.67490.66820.67380.68300.6296
LPIPS↓0.48510.32630.25850.25390.23760.24980.26850.23710.20750.2456
CLIPQA↑0.45120.53660.58100.52410.56390.53620.50950.58600.60030.6328
MUSIQ↑43.61546.98154.69652.60953.78948.33746.63953.18252.01959.227
MANIQA↑0.32120.30650.38650.36890.38820.32920.33050.41910.38850.4591
NIQE↓8.335.966.086.075.895.867.216.887.345.24
RealSRPSNR↑27.5727.3426.5125.8326.4327.1927.1826.8025.7726.44
SSIM↑0.77420.76050.77460.77260.78610.78610.78530.76740.74390.7589
LPIPS↓0.41520.39620.26850.27390.25150.31130.30210.34110.34910.2871
CLIPQA↑0.23620.36130.54390.49230.46550.36280.37480.57090.56460.6010
MUSIQ↑29.03736.06963.58759.84959.63545.81848.69857.76955.18964.541
MANIQA↑0.20710.17830.37020.36940.34360.26630.26550.36910.33370.3828
NIQE↓7.736.954.654.684.685.986.225.966.934.02
RealSetCLIPQA↑0.37390.52820.61600.60810.57780.49660.43130.63090.61880.6392
MUSIQ↑42.36650.53965.58364.12563.81755.70848.60259.31958.51663.519
MANIQA↑0.31000.29270.38880.39490.38180.31340.26930.39160.35260.3931
NIQE↓4.934.814.584.384.404.726.475.966.464.23
", + "bbox": [ + 99, + 137, + 898, + 406 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparisons with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 89, + 420, + 465, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We select ImageNet-Test [3, 44] that contains 3,000 images as major dataset for synthetic image super-resolution evaluation. Furthermore, two commonly used real-world datasets, RealSR [1] and RealSet65 [44], are adopted to evaluate the generalizability in real-world scenarios. We make comparisons with several GAN-based methods ESRGAN [36], RealSR-JPEG [10], BSRGAN [46], RealESRGAN [37], SwinIR [19], DASR [20], and two diffusion-based methods LDM [26] and ResShift [44].", + "bbox": [ + 89, + 443, + 483, + 579 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The quantitative results on three benchmark datasets are shown in Tab. 4. We present UPSR with five sampling steps in order to trade off the super-resolution performance and computational consumption. The proposed UPSR achieves better CLIPIQA, MUSIQ, MANIQA and NIQE compared to ResShift, indicating the significant improvement in perceptual quality. Specifically, while reducing $30\\%$ of the overall model size, UPSR-5 still outperforms ResShift-4 by 7.21 and 2.10 in terms of MUSIQ and NIQE in ImageNet-Test dataset. For real-world datasets, our method also yields impressive results. UPSR-5 consistently outperforms ResShift-4 across both the RealSR and RealSet datasets, demonstrating the efficacy of UPSR in addressing unknown degradation. Furthermore, UPSR-5 achieves the best NIQE metric, while recent diffusion-based methods show poorer NIQE performance compared to most GAN-based methods.", + "bbox": [ + 89, + 580, + 483, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Visual Examples", + "text_level": 1, + "bbox": [ + 89, + 832, + 256, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we present several visual examples to illustrate the proposed UNW scheme and the comparisons between different methods.", + "bbox": [ + 89, + 854, + 483, + 898 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/069bfc6d62c3261af3c570caa9f0d953953290d04beff652e8a8567ef5ce3f3d.jpg", + "image_caption": [ + "Figure 5. Visual examples of the proposed UNW strategy. Based on the uncertainty estimate (illustrated as the heatmap), the noise level in most flat areas is reduced to preserve more details for better SR results. Meanwhile, noise in edge areas (e.g., in image (a)) and severely degraded parts (e.g., in image (b)) are maintained relatively heavy to ensure reliable score estimation to produce visually pleasing results." + ], + "image_footnote": [], + "bbox": [ + 531, + 422, + 895, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualization of Weighted Noise. In Fig. 5 we provide visual examples of isotropic and anisotropic noises applied in different images. The proposed UNW strategy enables adaptive perturbation across different areas, leading to the anisotropic diffusion process. The reduced noise level in flat areas, particularly in clear backgrounds, help retain more details in the initial state without sacrificing generative capability. Additionally, these retained details provide more information to reconstruct missing details in surrounding regions, thereby enhancing overall performance. Conversely, in edge regions, the noise level is maintained at the predefined $\\sigma_{max}$ due to their high uncertainty, ensuring sufficient perturbation to prevent the model from generating over-smoothed results.", + "bbox": [ + 511, + 691, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17986", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1f2aaf03c8afccc2c6af213a7e3a19f6c383a311f5fcf00effa1f06d32ab353d.jpg", + "image_caption": [ + "ImageNet val 0079" + ], + "image_footnote": [], + "bbox": [ + 171, + 88, + 366, + 237 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8612017f43e00ec52fb47b1d149174ca1496576723234bb7d75d113186dd1cef.jpg", + "image_caption": [ + "BSRGAN" + ], + "image_footnote": [], + "bbox": [ + 367, + 89, + 478, + 154 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a597c2cae3ae5f90ae8cd28c0081d898454a03fab3cef5792afad59dbab1abce.jpg", + "image_caption": [ + "RealESRGAN" + ], + "image_footnote": [], + "bbox": [ + 480, + 89, + 594, + 154 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ab5c66c7677b34d74b7f0dcc945b2385af73bf01f03b2fb73fa0e5af17493a37.jpg", + "image_caption": [ + "SwinIR" + ], + "image_footnote": [], + "bbox": [ + 596, + 89, + 707, + 154 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8d9fbaae7598fe3023ec8050ddabaabd64b0dccf995ef09e3afaf42bfc0d1850.jpg", + "image_caption": [ + "DASR" + ], + "image_footnote": [], + "bbox": [ + 709, + 89, + 823, + 154 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/420f3298cbe28b597a5d32755eebdcde63fb0f87bb99abe1b65a783c3118ced2.jpg", + "image_caption": [ + "RealSet lion1" + ], + "image_footnote": [], + "bbox": [ + 173, + 253, + 366, + 400 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3b75d8c3fdf20a7d4486776507d46c440de06fd04b13b0655267c1e79378df34.jpg", + "image_caption": [ + "LDM-15" + ], + "image_footnote": [], + "bbox": [ + 367, + 172, + 478, + 237 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a3cd768b1e35f8b9c51707228c16a300953dc05421ce5c0c9ce48c0848ae65e1.jpg", + "image_caption": [ + "ResShift-15" + ], + "image_footnote": [], + "bbox": [ + 480, + 172, + 594, + 237 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f1b264e492ea056ccbad8d058809c648e19d7b5d51dbba6609e1c426b9463552.jpg", + "image_caption": [ + "ResShift-4" + ], + "image_footnote": [], + "bbox": [ + 596, + 172, + 707, + 237 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6ba80c14643f9dc85ddd1053c9d0a23caa5133f6e08d9794b4a5a06a6fe5e787.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 709, + 172, + 823, + 237 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/edc46ab5f7d3db251268e91e68e54d7edb654a6f7255b842ede5c4e925261e94.jpg", + "image_caption": [ + "RealSR_Nikon_047", + "Figure 6. Qualitative comparison between different methods on real-world datasets. Please zoom in for more details." + ], + "image_footnote": [], + "bbox": [ + 173, + 417, + 366, + 565 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cd48bf2bd1e2bdf4a7452ee004db74d6df3aabe63d34cfd5da5e386f2585446c.jpg", + "image_caption": [ + "BSRGAN" + ], + "image_footnote": [], + "bbox": [ + 367, + 253, + 480, + 318 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cd9d6a4c77072601bdd8c7eb2591fb8e750bcfdc20d09e0c0222c08e84d60936.jpg", + "image_caption": [ + "RealESRGAN" + ], + "image_footnote": [], + "bbox": [ + 480, + 253, + 594, + 318 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1485f27b6d3da9538d9e540823c18d08abfd19e984c4aaab3af4d245cf341717.jpg", + "image_caption": [ + "SwinIR" + ], + "image_footnote": [], + "bbox": [ + 596, + 253, + 707, + 318 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b5738fb580fc7c8b6c354ef7251dca7e32b3a00d25ba18abc2d0c67d45d45217.jpg", + "image_caption": [ + "DASR" + ], + "image_footnote": [], + "bbox": [ + 709, + 253, + 823, + 318 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d905cde3131a53209b3501413a8def071a0046cae149960786472d3827f5e0ae.jpg", + "image_caption": [ + "LDM-15" + ], + "image_footnote": [], + "bbox": [ + 367, + 335, + 480, + 401 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3744f14f7292279e0527f13917de22152ddb0eb12375ccf2346610fa34bd486d.jpg", + "image_caption": [ + "ResShift-15" + ], + "image_footnote": [], + "bbox": [ + 480, + 335, + 594, + 401 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a171951f36154a1e5a140d4f21c7d7caec5a8112fc516b63db77dc467e625342.jpg", + "image_caption": [ + "ResShift-4" + ], + "image_footnote": [], + "bbox": [ + 596, + 335, + 707, + 401 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f50e01ccc54e6f7dc071241fb1f11fae7f71adff2a7925511f95fb7103d70033.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 709, + 335, + 823, + 401 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a0b0a6c8975af49113d9d1b7d80ab2f2416a9ce5b637388329328bfa757c8869.jpg", + "image_caption": [ + "BSRGAN" + ], + "image_footnote": [], + "bbox": [ + 367, + 417, + 480, + 483 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ce42b425bef3092cdc3753ead8b94844bbda50db5594ac408835442c37501c50.jpg", + "image_caption": [ + "RealESRGAN" + ], + "image_footnote": [], + "bbox": [ + 480, + 417, + 594, + 483 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6bdf6b5539c279b5d75b8de85011de35c49f64d1c041195a3f6fcdd90fc07512.jpg", + "image_caption": [ + "SwinIR" + ], + "image_footnote": [], + "bbox": [ + 596, + 417, + 707, + 483 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/050394a69fcda9eee7a86f87b84052a38a61f4ad663c03b6e0192923912f32c5.jpg", + "image_caption": [ + "DASR" + ], + "image_footnote": [], + "bbox": [ + 709, + 417, + 823, + 483 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Comparison. In Fig. 6, we provide visual results produced by different methods on real-world datasets. These results validate that the proposed UPSR method is capable of producing visually better outputs. Specifically, UPSR successfully recover finer textures and sharper edges, while other methods yields blurry results. More visual examples are presented in the supplementary material.", + "bbox": [ + 89, + 621, + 482, + 728 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 739, + 209, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose specialized prior distribution and sampling pipeline, along with modifications on the network architecture to develop the Uncertainty-guided Perturbation scheme for Super-Resolution to tackle real-world SR task. Specifically, we propose region-specific processing based on the LR content from a uncertainty-based perspective. In flat areas (low-uncertainty), reducing the noise level could preserve more details in the initial state of the diffusion process and therefore improve performance, whereas edge and", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "texture regions (high-uncertainty) demand stronger noise to bridge their significant disparity with the target distribution. To facilitate region-specific control, we leverage the uncertainty of LR image as the criterion to adjust noise level in different areas. We employ an auxiliary SR network and consider the residual of SR prediction as the approximate uncertainty estimate of LR image due to their similarity. We further incorporate the SR prediction as conditional information and make several architectural modifications to provide a more robust pipeline to further enhance performance. Extensive experiments on both synthetic and real-world datasets validate the efficacy of the proposed UPSR model, even with lower model size and training overhead.", + "bbox": [ + 511, + 622, + 906, + 818 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by National Natural Science Foundation of China (No. 62476051) and Sichuan Natural Science Foundation (No. 2024NSFTD0041).", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17987", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3086-3095, 2019. 7, 15", + "[2] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image superresolution transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22367-22377, 2023. 1, 2", + "[3] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5, 7", + "[4] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 2", + "[5] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 5", + "[6] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2, 3", + "[7] Shuhang Gu, Wangmeng Zuo, Qi Xie, Deyu Meng, Xiangchu Feng, and Lei Zhang. Convolutional sparse coding for image super-resolution. In Proceedings of the IEEE International Conference on Computer Vision, pages 1823-1831, 2015. 2", + "[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2, 5", + "[9] Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research, 6(4), 2005. 3", + "[10] Xiaozhong Ji, Yun Cao, Ying Tai, Chengjie Wang, Jilin Li, and Feiyue Huang. Real-world super-resolution via kernel estimation and noise injection. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 466-467, 2020. 7", + "[11] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 3", + "[12] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. 1, 3", + "[13] Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and improving the training dynamics of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24174-24184, 2024. 3" + ], + "bbox": [ + 93, + 114, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 6", + "[15] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 4", + "[16] Changwoo Lee and Ki-Seok Chung. Gram: Gradient rescaling attention model for data uncertainty estimation in single image super resolution. In 2019 18th IEEE International Conference On Machine Learning And Applications (ICMLA), pages 8-13. IEEE, 2019. 2, 4", + "[17] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. 3", + "[18] Yawei Li, Yuchen Fan, Xiaoyu Xiang, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18278-18289, 2023. 2", + "[19] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 1, 2, 7", + "[20] Jie Liang, Hui Zeng, and Lei Zhang. Efficient and degradation-adaptive network for real-world image superresolution. In European Conference on Computer Vision, pages 574-591. Springer, 2022. 7", + "[21] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 2", + "[22] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 6", + "[23] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. arXiv preprint arXiv:1802.05957, 2018. 3", + "[24] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 3", + "[25] Qian Ning, Weisheng Dong, Xin Li, Jinjian Wu, and Guangming Shi. Uncertainty-driven loss for single image superresolution. Advances in Neural Information Processing Systems, 34:16398-16409, 2021. 2, 4", + "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 5, 6, 7, 13" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17988", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE transactions on pattern analysis and machine intelligence, 45(4):4713-4726, 2022. 1, 2, 3", + "[28] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1874-1883, 2016. 5", + "[29] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2", + "[30] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 3", + "[31] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 3", + "[32] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3", + "[33] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2", + "[34] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 6", + "[35] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, 132(12):5929-5949, 2024. 12", + "[36] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 7", + "[37] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1905-1914, 2021. 1, 2, 6, 7", + "[38] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image superresolution in a single step. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 25796-25805, 2024. 13", + "[39] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. Advances in Neural Information Processing Systems, 37:92529-92553, 2025. 12" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Rui Xie, Chen Zhao, Kai Zhang, Zhenyu Zhang, Jun Zhou, Jian Yang, and Ying Tai. Addsr: Accelerating diffusion-based blind super-resolution with adversarial diffusion distillation. arXiv preprint arXiv:2404.01717, 2024. 12", + "[41] Jianchao Yang, John Wright, Thomas S Huang, and Yi Ma. Image super-resolution via sparse representation. IEEE transactions on image processing, 19(11):2861-2873, 2010. 2", + "[42] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniaq: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1191-1200, 2022. 6", + "[43] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Efficient diffusion model for image restoration by residual shifting. arXiv preprint arXiv:2403.07319, 2024. 3, 13", + "[44] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image superresolution by residual shifting. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 5, 6, 7, 11, 13, 14", + "[45] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In Curves and Surfaces: 7th International Conference, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 2", + "[46] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4791-4800, 2021. 1, 2, 7", + "[47] Leheng Zhang, Yawei Li, Xingyu Zhou, Xiaorui Zhao, and Shuhang Gu. Transcending the limit of local window: Advanced super-resolution transformer with adaptive token dictionary. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856-2865, 2024. 1, 2, 6", + "[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5, 6", + "[49] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 286-301, 2018. 2" + ], + "bbox": [ + 516, + 90, + 903, + 766 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17989", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Uncertainty-guided Perturbation for Image Super-Resolution Diffusion Model/10430da4-a80e-4106-a1b5-f1210ade542e_model.json b/2025/Uncertainty-guided Perturbation for Image Super-Resolution Diffusion Model/10430da4-a80e-4106-a1b5-f1210ade542e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..33a0eb305ea8e87bd6d254d37ef4897666e71d48 --- /dev/null +++ b/2025/Uncertainty-guided Perturbation for Image Super-Resolution Diffusion Model/10430da4-a80e-4106-a1b5-f1210ade542e_model.json @@ -0,0 +1,2563 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.106, + 0.131, + 0.895, + 0.155 + ], + "angle": 0, + "content": "Uncertainty-guided Perturbation for Image Super-Resolution Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.267, + 0.181, + 0.737, + 0.218 + ], + "angle": 0, + "content": "Leheng Zhang Weiyi You Kexuan Shi Shuhang Gu* University of Electronic Science and Technology of China" + }, + { + "type": "text", + "bbox": [ + 0.338, + 0.22, + 0.66, + 0.235 + ], + "angle": 0, + "content": "{lehengzhang12, shuhanggu}@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.337, + 0.237, + 0.658, + 0.251 + ], + "angle": 0, + "content": "https://github.com/LabShuHangGU/UPSR" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.329, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.319, + 0.486, + 0.744 + ], + "angle": 0, + "content": "Diffusion-based image super-resolution methods have demonstrated significant advantages over GAN-based approaches, particularly in terms of perceptual quality. Building upon a lengthy Markov chain, diffusion-based methods possess remarkable modeling capacity, enabling them to achieve outstanding performance in real-world scenarios. Unlike previous methods that focus on modifying the noise schedule or sampling process to enhance performance, our approach emphasizes the improved utilization of LR information. We find that different regions of the LR image can be viewed as corresponding to different timesteps in a diffusion process, where flat areas are closer to the target HR distribution but edge and texture regions are farther away. In these flat areas, applying a slight noise is more advantageous for the reconstruction. We associate this characteristic with uncertainty and propose to apply uncertainty estimate to guide region-specific noise level control, a technique we refer to as Uncertainty-guided Noise Weighting. Pixels with lower uncertainty (i.e., flat regions) receive reduced noise to preserve more LR information, therefore improving performance. Furthermore, we modify the network architecture of previous methods to develop our Uncertainty-guided Perturbation SuperResolution (UPSR) model. Extensive experimental results demonstrate that, despite reduced model size and training overhead, the proposed UPSR method outperforms current state-of-the-art methods across various datasets, both quantitatively and qualitatively." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.773, + 0.224, + 0.789 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.799, + 0.485, + 0.877 + ], + "angle": 0, + "content": "Single image super-resolution (SR), which aims to recover clean the high-resolution (HR) image from its degraded, contaminated low-resolution (LR) counterpart, is a classical problem in the computer vision society. Significant information loss during the degradation process underscores the" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.286, + 0.71, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.403, + 0.631, + 0.416 + ], + "angle": 0, + "content": "(a) \\(\\pmb{y}_0\\)" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.286, + 0.905, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.776, + 0.403, + 0.843, + 0.415 + ], + "angle": 0, + "content": "(b) \\(\\pmb{x}_{T}^{\\mathrm{SR3}} = \\pmb{\\epsilon}\\)" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.417, + 0.709, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.534, + 0.674, + 0.547 + ], + "angle": 0, + "content": "(c) \\( \\mathbf{x}_T^{\\mathrm{ResShift}} = \\mathbf{y}_0 + \\pmb{\\epsilon} \\)" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.417, + 0.906, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.534, + 0.876, + 0.547 + ], + "angle": 0, + "content": "(d) \\(\\pmb{x}_T^{\\mathrm{UPSR}} = \\pmb{y}_0 + w_u(\\pmb{y}_0)\\pmb{\\epsilon}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.557, + 0.908, + 0.67 + ], + "angle": 0, + "content": "Figure 1. A comparison of initial state setup between different diffusion-based image super-resolution methods, where \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\sigma_{\\mathrm{max}}^2\\mathbf{I})\\). (b) SR3 [27] initiate the diffusion process from pure Gaussian noise, whereas (c) ResShift [44] and (d) our UPSR embed the LR input into the initial noise map. Additionally, we apply uncertainty-guided weighting coefficient \\(w_{u}(\\pmb{y}_{0})\\) to reduce the noise level in flat areas, achieving a more specialized diffusion process for SR to improve performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.701, + 0.907, + 0.793 + ], + "angle": 0, + "content": "need for robust super-resolution modeling capability to recover missing details and produce visually pleasing results. Existing SR methods have explored a variety of advanced network architectures and complex degradation modeling to improve performance in classical SR [2, 19, 47] and real-world SR [37, 46], respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Recently, diffusion models [8, 12, 30] have demonstrated impressive capability in image synthesis, offering a promising new approach for the real-world image super-resolution task. These methods transform pure Gaussian noise into high-quality images through a predefined Markov chain, and their solid theoretical foundation endows them with exceptional modeling capacity to bridge diverse data distri" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.888, + 0.232, + 0.901 + ], + "angle": 0, + "content": "* corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17980" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "butions. To leverage the modeling capabilities of diffusion models for recovering missing details in LR images, several methods [26, 27] begin the SR process by sampling from a standard Gaussian distribution and gradually refine the noisy inputs into high-quality outcomes. However, starting from pure noise was originally intended for image synthesis tasks, resulting in suboptimal outcomes for superresolution. Additionally, these methods typically require a lengthy sampling process, limiting their practicality for real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.247, + 0.483, + 0.474 + ], + "angle": 0, + "content": "To address the aforementioned issues, ResShift [44] focuses on the construction of the prior distribution, highlighting the importance of LR information during the diffusion process. By embedding the LR image into the initial noise map and progressively recovering the residual between the LR and HR images, ResShift greatly simplifies the diffusion process. Instead of modeling the entire HR image from noise, it only needs to estimate the LR-HR residual, which shortens the sampling process while significantly enhancing super-resolution results. Despite improved performance, challenges remain, as shown in Fig. 1 where most details are obscured by heavy noise. This still poses additional challenges, as exploiting the information from surrounding areas is crucial for recovering missing details in the SR task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.483, + 0.901 + ], + "angle": 0, + "content": "For the pursuit of designing a more specialized and effective diffusion process for SR to improve performance, we first propose to make better use of the LR information in the prior distribution. In recent studies, little attention has been given to the inherent information in the LR image that flat areas are already close to the target, whereas edge and texture regions are farther away. To leverage this information, we propose to consider different regions in the LR image as being situated at various timesteps of an isotropic diffusion process. This leads to an anisotropic diffusion process, where flat areas are assigned lower noise levels (as \\(t \\to 0\\)), while edge and texture regions receive larger noise (as \\(t \\to T\\)). To achieve region-specific noise control, we draw inspiration from uncertainty-driven SR approaches [16, 25] and employ a simple SR network to estimate the uncertainty (variance) of different areas in the input LR image. This uncertainty reflects the difficulty of recovering HR details and is closely related to the distribution disparity between LR and HR, indicating the amount of noise required. We then propose a technique called Uncertainty-guided Noise Weighting, which applies weight coefficients to noise levels in different regions based on their uncertainty estimates. The weight coefficient is positively correlated with the uncertainty estimate, enabling adaptive reduction of noise intensity in flat areas to preserve more details in the initial state of the diffusion process. Equipped with UNW and a modified network architecture, we introduce the UPSR model, which achieves state-of-the-art per" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "formance while reducing computational costs, as validated on both synthetic and real-world SR datasets." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.123, + 0.877, + 0.137 + ], + "angle": 0, + "content": "Our major contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.139, + 0.905, + 0.242 + ], + "angle": 0, + "content": "- To achieve a more specialized diffusion pipeline for SR, we propose to adjust the noise level for different areas based on their uncertainty estimate, where the noise in flat areas (low uncertainty) is reduced. This Uncertainty-guided Noise Weighting strategy enables adaptive control across different areas and results in improved performance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.244, + 0.905, + 0.333 + ], + "angle": 0, + "content": "- We establish the connection between the residual estimated by a pre-trained SR network and the uncertainty of the LR input. The estimated residual could serve as an approximate measure of uncertainty, enabling effective uncertainty estimation in the proposed uncertainty-guided noise weighting scheme." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.335, + 0.905, + 0.41 + ], + "angle": 0, + "content": "- By integrating the proposed uncertainty-guided noise weighting scheme and the modified network architecture, our method achieves state-of-the-art performance across various benchmark datasets with smaller model size and less training overhead." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.139, + 0.905, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.426, + 0.663, + 0.441 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.452, + 0.735, + 0.468 + ], + "angle": 0, + "content": "2.1. Image Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.474, + 0.907, + 0.807 + ], + "angle": 0, + "content": "The current literature on image super-resolution can be divided into two categories: classical SR and real SR. The former focuses on addressing predefined degradation patterns, e.g., bicubic downsampling. After the era of conventional methods [7, 41, 45], a multitude of works [2, 4, 18, 19, 21, 47, 49] focus on exploring the potential of neural network, ranging from improved convolutional neural networks to transformer variants [33]. These architectural modifications enhance the capacity of modeling complex relationship between LR and HR images, thereby achieving better performance. Despite their success, these methods perform poorly when directly applied to the real SR task, where the degradation model includes a series of unknown and complex noise and blur patterns. BSRGAN [46] incorporates a series of degradation operators to simulate real-world scenarios, providing a robust training baseline. By combining this strategy with a generative adversarial network [6], BSRGAN achieves remarkable fidelity and perceptual quality in the real SR task. Subsequently, RealESRGAN [37] extends this scheme to a high-order degradation model, providing a more robust training data synthesis procedure and better performance." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.818, + 0.682, + 0.833 + ], + "angle": 0, + "content": "2.2. Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Rooted in nonequilibrium thermodynamics, diffusion probabilistic models [29] have emerged as a new trend in recent research of generative models. DDPM [8] proposes a lengthy parameterized Markov chain to bridge the dis" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17981" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.273 + ], + "angle": 0, + "content": "distribution of high-quality images and the standard Gaussian distribution. Built upon solid theoretic foundations, DDPM simplifies training objectives and achieves a great breakthrough in image synthesis tasks, demonstrating significant generative capacity compared to GAN-based methods [6, 11, 23]. Subsequently, immense amounts of related research have developed various techniques to further improve DDPM by enhancing performance or shortening sampling process. These improvements include deterministic sampling [30], modified noise schedule [24], harnessing latent space [26], second-order ODE solver [12], and improved training dynamics [13]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.28, + 0.439, + 0.297 + ], + "angle": 0, + "content": "2.3. Diffusion-based Image Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.485, + 0.558 + ], + "angle": 0, + "content": "SRDiff [17] and SR3 [27] are the first to apply diffusion models to image super-resolution by taking the LR input as conditional information, demonstrating the efficacy of diffusion model in generating perceptually high-quality SR images. Despite superior performance, SR3 suffers from bias issues and a costly sampling process. LDM-SR [26] trains an autoencoder and performs the diffusion process in its low-dimensional latent space. This approach allows the model to concentrate on perceptually relevant details and significantly improves computational efficiency. ResShift [44] embeds the LR information \\( \\mathbf{y}_0 \\) into the initial state \\( \\mathbf{x}_T \\) and built a new prior distribution: \\( \\mathbf{x}_T \\sim \\mathcal{N}(\\mathbf{x}_T \\mid \\mathbf{y}_0, \\kappa^2 \\eta_T \\mathbf{I}) \\). In this way, the diffusion process shifts from generating the HR image \\( \\mathbf{x}_0 \\) from pure noise \\( \\epsilon \\) to generating the residual \\( \\mathbf{x}_0 - \\mathbf{y}_0 \\) given noisy LR information \\( \\mathbf{y}_0 + \\epsilon \\), greatly reducing the difficulty of estimation. The forward and backward transition distributions are defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.561, + 0.485, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} q \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {t - 1}, \\boldsymbol {x} _ {0}, \\boldsymbol {y} _ {0}\\right) = (1) \\\\ \\mathcal {N} \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {t - 1} + \\alpha_ {t} \\left(\\boldsymbol {y} _ {0} - \\boldsymbol {x} _ {0}\\right), \\kappa^ {2} \\alpha_ {t} \\boldsymbol {I}\\right) (1) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.603, + 0.121, + 0.616 + ], + "angle": 0, + "content": "and" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.618, + 0.483, + 0.685 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} q \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}, \\boldsymbol {x} _ {0}, \\boldsymbol {y} _ {0}\\right) = \\\\ \\mathcal {N} \\left(\\boldsymbol {x} _ {t - 1} \\mid \\frac {\\eta_ {t - 1}}{\\eta_ {t}} \\boldsymbol {x} _ {t} + \\frac {\\alpha_ {t}}{\\eta_ {t}} \\boldsymbol {x} _ {0}, \\kappa^ {2} \\frac {\\eta_ {t - 1}}{\\eta_ {t}} \\alpha_ {t} \\boldsymbol {I}\\right), \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.692, + 0.484, + 0.812 + ], + "angle": 0, + "content": "for \\( t = 1,2,\\dots ,T \\); \\( \\eta_t \\) and \\( \\alpha_{t} = \\eta_{t} - \\eta_{t - 1} \\) are time-dependent positive parameters that characterize the velocity of mean shift and noise injection at different timesteps. In this work, we adopt the sampling process and network architecture of ResShift as our baseline and propose a series of modifications to the sampling pipeline and network architecture to enhance performance while significantly saving model size and training overhead." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.824, + 0.226, + 0.841 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.395, + 0.864 + ], + "angle": 0, + "content": "3.1. Noise Levels in Diffusion-based SR" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In the context of SDE [31, 32], perturbing the data with Gaussian noise is crucial for facilitating score estima" + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.095, + 0.838, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.207, + 0.719, + 0.217 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.219, + 0.71, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.219, + 0.905, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.337, + 0.72, + 0.347 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.35, + 0.907, + 0.517 + ], + "angle": 0, + "content": "Figure 2. (a) The distribution of pixel residual \\( |y - x| \\) computed on ImageNet-Test dataset [44], omitting values where \\( |y - x| > 0.4 \\) for clarity. The result exhibits a distinct long-tailed characteristic. (b) The statistical curves of fidelity \\( |f(y) - x| \\) and perceptual quality \\( |\\phi(f(y)) - \\phi(x)| \\) with respect to residual \\( |y - x| \\) under different noise levels. As \\( |y - x| \\) increases, the gap of fidelity remains relatively stable when different noise levels are applied. In contrast, the perceptual quality is more sensitive to the noise level. A larger noise is more requisite in regions with high residual value to achieve better perceptual quality. Meanwhile, we propose weighted noise level \\( w_u(y)\\sigma_{max} \\) which could lead to better results, with details presented in Sec. 3.3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.809 + ], + "angle": 0, + "content": "tion [9], especially for data residing in low-dimensional manifolds. They apply various noise levels to perturb the data to explore the low-density regions and enable the score estimation towards high-density regions. In the super-resolution task, the target is to estimate the gradients from LR images (low density) to HR images (high density) in the real image distribution. Recent diffusion-based methods [43, 44] inject a large initial noise into the LR image and gradually reduce it throughout the reverse diffusion process, as the LR sample progressively moves closer to high-density regions. However, these methods focus solely on the different amounts of noise needed at different timesteps during the diffusion process, without accounting for the fact that different areas of each image require varying noise intensities. To provide a clearer insight, we conduct several experiments to demonstrate how the fidelity and perceptual quality of the output change when different overall noise levels are applied to a diffusion-based SR model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We first analyze the distribution of the residual \\( |y - x| \\) between each pixel pair of the upsampled LR image \\( y \\) and its corresponding HR image \\( x \\), as illustrated in Fig. 2a. The result clearly indicates that \\( |y - x| \\) follows a long-tailed distribution, with over \\( 95\\% \\) of the data concentrated in the range [0.01, 0.16]. Next, we focus on data within this range to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17982" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.093, + 0.189, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.17, + 0.149, + 0.177 + ], + "angle": 0, + "content": "HR" + }, + { + "type": "image", + "bbox": [ + 0.191, + 0.093, + 0.288, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.17, + 0.245, + 0.177 + ], + "angle": 0, + "content": "LR" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.094, + 0.385, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.311, + 0.17, + 0.362, + 0.177 + ], + "angle": 0, + "content": "Real Residual" + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.094, + 0.482, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.399, + 0.17, + 0.469, + 0.177 + ], + "angle": 0, + "content": "Estimated Residual" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.183, + 0.483, + 0.252 + ], + "angle": 0, + "content": "Figure 3. A visualization of the actual residual \\(|\\pmb{x}^i - \\pmb{y}^i|\\) and the estimated residual \\(|g(\\pmb{y}^i) - \\pmb{y}^i|\\). The real residual exhibits high values in edges and texture regions, indicating the high uncertainty. The residual estimated by SR network is close to the real one and therefore can serve as a rough estimation of uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.269, + 0.485, + 0.542 + ], + "angle": 0, + "content": "investigate the relationships between the residual \\( |y - x| \\), fidelity \\( |f(y) - x| \\), perceptual quality \\( |\\phi(f(y)) - \\phi(x)| \\) of the output of the denoiser \\( f(\\cdot) \\) trained under different noise levels \\( \\sigma_{max} \\) (in a reasonable range). The results are illustrated in Fig. 2b. Due to the blurring and downsampling degradation, the residual \\( |y - x| \\) tends to be low in flat areas and high in regions characterized by edges and textures. As \\( |y - x| \\) grows, the perceptual quality gap between \\( |\\phi(f(y)) - \\phi(x)|_{\\sigma_{max}=2.5} \\) and \\( |\\phi(f(y)) - \\phi(x)|_{\\sigma_{max}=0.5} \\) increases rapidly, while the fidelity gap remains nearly unchanged. This indicates that perceptual quality is more sensitive to the noise level change, with higher noise levels being particularly important in edge and texture areas. These regions lie in the low-density regions of the real image distribution and are affected by the ill-posed nature of superresolution. Applying a low \\( \\sigma_{max} \\) in such areas leads to unreliable score estimation, resulting in perceptually poor, over-smoothed outputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.544, + 0.484, + 0.71 + ], + "angle": 0, + "content": "The experimental results support our hypothesis that different regions of an image respond differently to noise and require distinct handling based on their content. Flat regions typically closely resemble the corresponding ground truth, requiring only a slight amount of noise; therefore, we consider them as existing at small timesteps \\((t\\to 0)\\). Based on this idea, we propose to reduce noise levels applied in flat regions, and the results shown by the red line in Fig. 2b preliminarily verify the effectiveness. In the following subsections, we will dually discuss how to determine the noise level for different types of areas in the LR image." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.726, + 0.308, + 0.741 + ], + "angle": 0, + "content": "3.2. Uncertainty Estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Due to the ill-posed nature of the image super-resolution task, perfectly reconstructing HR images from degraded LR inputs is non-trivial. The high-frequency information in edge and texture areas is severely corrupted during the degradation, exhibiting greater variance compared to flat areas and making accurate prediction more challenging. Several works [16, 25] explore this variance from an uncertainty-based perspective [15]. Given an HR image \\( \\boldsymbol{x}^i \\in \\mathbb{R}^n \\) and its corresponding LR image \\( \\boldsymbol{y}^i \\in \\mathbb{R}^n \\), the uncertainty \\( \\psi^i \\) of the SR estimate \\( g(\\boldsymbol{y}^i) \\) is related to its resid" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.091, + 0.597, + 0.106 + ], + "angle": 0, + "content": "ual with \\(x^i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.117, + 0.907, + 0.136 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} ^ {i} = g \\left(\\boldsymbol {y} ^ {i}\\right) + \\epsilon \\psi \\left(g \\left(\\boldsymbol {y} ^ {i}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.144, + 0.907, + 0.25 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) represents a standard Laplace or Gaussian distribution when \\(g(\\cdot)\\) is regularized by \\(L1\\) or \\(L2\\) loss function, respectively. If the reconstruction error \\(|\\pmb{x}^i - g(\\pmb{y}^i)|\\) is larger than \\(|\\pmb{x}^j - g(\\pmb{y}^j)|\\), then it is more likely that \\(g(\\pmb{y}^i)\\) has a higher uncertainty \\(\\psi(g(\\pmb{y}^i))\\), compared to the uncertainty \\(\\psi(g(\\pmb{y}^j))\\) of \\(g(\\pmb{y}^j)\\). Similarly, we can associate the uncertainty of \\(\\pmb{y}^i\\) with the residual \\(|\\pmb{x}^i - \\pmb{y}^i|\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.642, + 0.259, + 0.907, + 0.278 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} ^ {i} = \\boldsymbol {y} ^ {i} + \\hat {\\epsilon} \\psi (\\boldsymbol {y} ^ {i}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.285, + 0.907, + 0.406 + ], + "angle": 0, + "content": "where \\(\\hat{\\epsilon}\\) represents an unknown distribution that depends on the degradation pattern. If \\(g(\\cdot)\\) is well-trained, we can assume that \\(g(\\pmb{y}^i)\\) closely approximates \\(\\pmb{x}^i\\), implying that \\(|g(\\pmb{y}^i) - \\pmb{y}^i|\\) is similar to \\(|\\pmb{x}^i - \\pmb{y}^i|\\). As illustrated in Fig. 3, the residual estimated by \\(g(\\cdot)\\) is roughly similar to the real residual. Therefore, we propose to leverage the residual \\(|g(\\pmb{y}^i) - \\pmb{y}^i|\\) as the estimate of the uncertainty of \\(\\pmb{y}^i\\). Specifically, we define the uncertainty estimate as:" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.414, + 0.906, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\psi_ {e s t} (\\boldsymbol {y}) = \\frac {1}{2} | g (\\boldsymbol {y}) - \\boldsymbol {y} |. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.451, + 0.906, + 0.482 + ], + "angle": 0, + "content": "In the next subsection, we will take this uncertainty estimate as the criterion to adjust the noise intensity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.49, + 0.831, + 0.507 + ], + "angle": 0, + "content": "3.3. Uncertainty-guided Noise Weighting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.512, + 0.907, + 0.875 + ], + "angle": 0, + "content": "Based on the discussion in previous sections, we propose to replace the commonly used isotropic Gaussian noise with anisotropic noise that adapts to the image content. Inspired by [25], which deals with areas with higher uncertainty by assigning larger weights to the loss function to impose stronger constraints, we suggest adjusting the noise intensity based on the uncertainty estimate across different regions. Specifically, after obtaining the uncertainty estimate as outlined in Sec. 3.2, we compute the weighting coefficient \\( w_{u} \\) and apply it to modulate noise levels for different areas in the diffusion process. We refer to this strategy as Uncertainty-guided Noise Weighting (UNW). For \\( y \\) with lower uncertainty (e.g. in flat areas), it is more likely that \\( y \\) resides in high-density regions of the HR data distribution. In such cases, a lower noise level is sufficient to achieve perceptual quality comparable to that of a higher noise level, while also preserving more LR details, as discussed in Sec. 3.1. Conversely, if \\( y \\) exhibits high uncertainty (e.g., in edge or texture areas), a higher noise level is requisite to account for the significant distribution disparity and to provide greater chances of reconstructing photorealistic details. Therefore, we model the noise weighting coefficient \\( w_{u} \\) as a monotonically increasing function with respect to the uncertainty:" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.884, + 0.907, + 0.902 + ], + "angle": 0, + "content": "\\[\nw _ {u} (\\boldsymbol {y}) := u \\left(\\psi_ {\\text {e s t}} (\\boldsymbol {y})\\right), \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17983" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.202, + 0.094, + 0.805, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.282, + 0.908, + 0.327 + ], + "angle": 0, + "content": "Figure 4. The overall pipeline of the proposed UPSR model. An auxiliary SR network is first employed to estimate the uncertainty of the input \\( \\mathbf{y}_0 \\). Then the weighting coefficient \\( w_{u} \\) computed based on the uncertainty \\( \\psi_{est}(\\mathbf{y}_0) \\) are applied to adjust the noise level in different regions. Meanwhile, both the SR estimate \\( g(\\mathbf{y}_0) \\) and LR input \\( \\mathbf{y}_0 \\) are concatenated as the conditional information for the denoiser \\( f_{\\theta}(\\cdot) \\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.34, + 0.482, + 0.431 + ], + "angle": 0, + "content": "where \\( w_{u}(\\pmb{y}) \\in \\mathbb{R}^{n \\times n} \\) represents the diagonal weight matrix used to lower the noise level in low-uncertainty areas while maintaining a higher noise level in high-uncertainty regions, creating a more specialized and adaptive diffusion pipeline for SR. Details of the implementation of \\( u(\\cdot) \\) are presented in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.484, + 0.569 + ], + "angle": 0, + "content": "Built upon the UNW technique, we introduce a new pipeline for diffusion-based SR, termed Uncertainty-guided Perturbation for SR (UPSR), as illustrated in Fig. 4. Given an LR image \\(\\pmb{y}_0\\), we first obtain its SR estimate \\(g(\\pmb{y}_0)\\) through an auxiliary SR network \\(g(\\cdot)\\). Then we estimate the uncertainty of \\(\\pmb{y}_0\\) as \\(\\psi_{est}(\\pmb{y}_0) = \\frac{1}{2} |g(\\pmb{y}_0) - \\pmb{y}_0|\\) and obtain the weighting coefficient as \\(w_u(\\pmb{y}_0) = u(\\psi_{est}(\\pmb{y}_0))\\). Based on the weighting coefficient, we rewrite the forward transition distribution \\(q(\\pmb{x}_t \\mid \\pmb{x}_{t-1}, \\pmb{x}_0, \\pmb{y}_0)\\) in Eq. 1 as:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.578, + 0.483, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\left. \\mathcal {N} \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {t - 1} + \\alpha_ {t} \\left(\\boldsymbol {y} _ {0} - \\boldsymbol {x} _ {0}\\right), \\kappa^ {2} w _ {u} \\left(\\boldsymbol {y} _ {0}\\right) ^ {2} \\alpha_ {t} \\boldsymbol {I}\\right), \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.483, + 0.639 + ], + "angle": 0, + "content": "and derive the corresponding backward transition distribution \\( q(\\pmb{x}_{t-1} \\mid \\pmb{x}_t, \\pmb{x}_0, \\pmb{y}_0) \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.649, + 0.483, + 0.682 + ], + "angle": 0, + "content": "\\[\n\\mathcal {N} \\left(\\boldsymbol {x} _ {t - 1} \\mid \\frac {\\eta_ {t - 1}}{\\eta_ {t}} \\boldsymbol {x} _ {t} + \\frac {\\alpha_ {t}}{\\eta_ {t}} \\boldsymbol {x} _ {0}, \\kappa^ {2} w _ {u} (\\boldsymbol {y} _ {0}) ^ {2} \\frac {\\eta_ {t - 1}}{\\eta_ {t}} \\alpha_ {t} \\boldsymbol {I}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.692, + 0.483, + 0.844 + ], + "angle": 0, + "content": "where the difference is that we leverage the weighting coefficient \\( w_{u}(\\pmb{y}_{0}) \\) to control the noise level. Since \\( g(\\pmb{y}_0) \\) is a better estimate of \\( x_0 \\) compared to \\( y_{0} \\), we combine \\( g(\\pmb{y}_0) \\) with \\( y_{0} \\) to provide more accurate conditional information for the denoiser. Details of the derivations of Eq. 7, 8 and the training pipeline are presented in the supplementary materials. Following the previous diffusion-based methods [8, 44], our training objective for the denoiser to predict the target \\( x_0 \\) combines both pixel distance \\( ||\\cdot||_2^2 \\) and LPIPS criterion \\( L_{per} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.853, + 0.483, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} (\\theta) = \\sum_ {t} [ \\| f _ {\\theta} (\\boldsymbol {x} _ {t}, \\boldsymbol {y} _ {0}, g (\\boldsymbol {y} _ {0}), t) - \\boldsymbol {x} _ {0} \\| _ {2} ^ {2} \\tag {9} \\\\ + \\lambda L _ {p e r} \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {t}, \\boldsymbol {y} _ {0}, g (\\boldsymbol {y} _ {0}), t\\right), \\boldsymbol {x} _ {0}\\right) ], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.907, + 0.416 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a hyperparameter to control the trade-off between fidelity and perceptual quality. Optimizing the denoiser \\(f_{\\theta}(\\cdot)\\) through the mixed objective function facilitates fewer diffusion steps while achieving better photo-realistic results in various benchmark real-world datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.425, + 0.822, + 0.44 + ], + "angle": 0, + "content": "3.4. Network Architecture Modification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.906, + 0.492 + ], + "angle": 0, + "content": "Besides altering the noise injection scheme and sampling procedure, we also make several modifications to the network architecture." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.907, + 0.81 + ], + "angle": 0, + "content": "We rethink the necessity of leveraging latent space for image super-resolution with only a few diffusion steps. [26] and [44] utilize VQGAN [5] to transfer the diffusion process from pixel space to latent space, for the sake of reducing spatial dimensionality to improve efficiency. However, as the number of sampling steps decreases, the performance-cost ratio of VQGAN gradually decreases due to its own huge computing resource consumption. Moreover, the application of certain perceptual criteria (e.g., LPIPS [48]) requires pixel-level computation, leading to additional decoding costs during the training process. To mitigate its impact, we propose to replace the VQGAN encoder and decoder with PixelUnshuffle operation [28] and a simple nearest neighbor upsampling module, respectively. This approach enables compressing and expanding the spatial dimension with minimal computational cost. Therefore, we can perform the diffusion process directly in pixel space while maintaining a complexity comparable to that in the latent space. In this vein, we significantly reduce the total model size and training overhead without sacrificing performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.824, + 0.646, + 0.841 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.712, + 0.866 + ], + "angle": 0, + "content": "4.1. Experimental Setups" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "For data preparation, we apply randomly cropped \\(256 \\times 256\\) patches from ImageNet [3] as HR training data following" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17984" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Ablation study on effects of the proposed components, including the SR condition and uncertainty-guided noise weighting (UNW). The best results are highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.131, + 0.891, + 0.219 + ], + "angle": 0, + "content": "
UNWSR cond.RealSRRealSet
PSNR↑CLIPQA↑MUSIQ↑MANIQA↑NIQE↓CLIPQA↑MUSIQ↑MANIQA↑NIQE↓
26.180.544762.9510.35964.490.614164.3600.37184.42
✓26.120.576064.5120.37174.180.634064.2800.38364.22
✓✓26.440.601064.5410.38184.020.638963.4980.39314.24
" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.236, + 0.485, + 0.495 + ], + "angle": 0, + "content": "[26] and [44]. Then the degradation pipeline of RealESR-GAN [37] are adopted to generate degraded \\(64 \\times 64\\) LR inputs. For network architecture, we employ the commonly-used U-Net denoiser \\(f_{\\theta}(\\cdot)\\) with few adjustments to fit our refined diffusion pipeline. For \\(g(\\cdot)\\), we employ the lightweight version of [47] and pretrain it following the real SR pipeline to produce SR predictions and uncertainty estimates. During the diffusion model training process, we freeze the pretrained \\(g(\\cdot)\\) and optimize \\(f_{\\theta}(\\cdot)\\) for 200k iterations with a batchsize of 32. To quantitatively validate the efficacy of the proposed method, we utilize three full-reference metrics: PSNR, SSIM and LPIPS [48]; and four non-reference metrics CLIPIQA [34], MUSIQ [14], MANIQA [42], and NIQE [22] to evaluate the performance. Among these metrics, PSNR and SSIM reflect the fidelity of the generated super-resolution results, while other metrics assess the perceptual quality of the outputs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.506, + 0.196, + 0.523 + ], + "angle": 0, + "content": "4.2. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.529, + 0.484, + 0.577 + ], + "angle": 0, + "content": "In this subsection, we conduct ablation studies on the effectiveness of several key components and analyze the performance-cost trade-off of the proposed method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Effectiveness of UNW and SR conditioning. In order to show the efficacy of the proposed uncertainty-guided noise weighting scheme and SR conditioning, we use the noise schedule and sampling process of ResShift [44] as the baseline. We first apply the uncertainty-guided noise weighting scheme to adjust the noise intensity across different regions based on the uncertainty estimate. The weighting scheme yields better outcomes in terms of perceptual quality, with enhancements of 0.0355 in CLIPIQA and 1.512 in MUSIQ. We attribute this improvement to the reduced noise intensities applied to flat areas, which allow these clearer regions to provide additional information that aids in more accurately recovering the remaining degraded regions. This validates our idea of reducing noise in flat regions does not degrade perceptual quality, but rather enhances it. Next, we incorporate the SR prediction \\( g(\\pmb{y}_0) \\) and LR input \\( \\pmb{y}_0 \\) as the conditional information entered into the denoiser \\( f_{\\theta}(\\cdot) \\). The additional SR prediction offers more precise supplementary information compared to relying solely on the degraded LR input, resulting in a significant improvement in both fi" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.238, + 0.907, + 0.324 + ], + "angle": 0, + "content": "Table 2. Model size and computational efficiency comparisons between the proposed UPSR and other diffusion-based methods. Gray numbers denote the parameter counts for auxiliary network, i.e., VQGAN in ResShift and \\( g(\\cdot) \\) in our work. We test the runtimes on \\( 32 \\times 3 \\times 64 \\times 64 \\) LR input using single RTX4090 GPU and present several results evaluated on the RealSR dataset." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.335, + 0.905, + 0.413 + ], + "angle": 0, + "content": "
ModelParams (M)Runtimes (s)MUSIQMANIQA
LDM-15113.60+55.321.5948.6980.2655
ResShift-15118.59+55.321.9857.7690.3691
ResShift-4118.59+55.321.0055.1890.3337
UPSR-5119.42+2.501.1264.5410.3818
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.425, + 0.907, + 0.454 + ], + "angle": 0, + "content": "Table 3. Training overhead comparison between ResShift and UPSR evaluated under a batchsize of 16 per GPU." + }, + { + "type": "table", + "bbox": [ + 0.571, + 0.461, + 0.848, + 0.515 + ], + "angle": 0, + "content": "
ModelTraining SpeedMemory Footprint
ResShift1.20 s/iter24.1 G
UPSR0.45 s/iter14.9 G
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.907, + 0.606 + ], + "angle": 0, + "content": "delity and perceptual quality. Specifically, this more robust pipeline further improves PSNR by 0.34 dB and CLIPIQA by 0.0247 in the RealSR dataset. These ablation studies verify the significance of leveraging LR information during diffusion-based SR process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Model Size and Training Overhead Comparison. In section 3.4, we discussed replacing VQGAN in diffusion-based SR model with simpler downsampling and upsampling modules to improve efficiency. To validate the efficacy of the modified architecture, we first evaluate the model size and computational cost of UPSR compared to state-of-the-art diffusion-based methods, as shown in Tab. 2. With the modified architecture and pipeline, UPSR-5 reduces the overall model size by \\(30\\%\\) and achieves better perceptual quality with an inference speed comparable to that of ResShift-4. In contrast, ResShift-4 achieves the acceleration at the expense of performance. We then assess the training overhead of different methods, as shown in Tab. 3. Removing VQGAN results in substantial reductions in training overhead, increasing training speed by \\(167\\%\\) and saving GPU memory footprint by \\(38\\%\\). These results validate that UPSR strikes a better trade-off between performance and efficiency." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17985" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.131 + ], + "angle": 0, + "content": "Table 4. Quantitative results of different methods on one synthetic dataset ImageNet-Test and two real-world datasets RealSR and RealSet. The best and second best results are marked in red and blue. All results of the previous methods are evaluated using the released inference codes and pretrained weights." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.138, + 0.9, + 0.407 + ], + "angle": 0, + "content": "
DatasetsMetricsGAN-based MethodsDiffusion-based Methods
ESRGANRealSR-JPEGBSRGANRealESRGANSwinIRDASRLDM-15ResShift-15ResShift-4UPSR-5
ImageNet -TestPSNR↑20.6723.1124.4224.0423.9924.7524.8524.9425.0223.77
SSIM↑0.44850.59120.65850.66490.66660.67490.66820.67380.68300.6296
LPIPS↓0.48510.32630.25850.25390.23760.24980.26850.23710.20750.2456
CLIPQA↑0.45120.53660.58100.52410.56390.53620.50950.58600.60030.6328
MUSIQ↑43.61546.98154.69652.60953.78948.33746.63953.18252.01959.227
MANIQA↑0.32120.30650.38650.36890.38820.32920.33050.41910.38850.4591
NIQE↓8.335.966.086.075.895.867.216.887.345.24
RealSRPSNR↑27.5727.3426.5125.8326.4327.1927.1826.8025.7726.44
SSIM↑0.77420.76050.77460.77260.78610.78610.78530.76740.74390.7589
LPIPS↓0.41520.39620.26850.27390.25150.31130.30210.34110.34910.2871
CLIPQA↑0.23620.36130.54390.49230.46550.36280.37480.57090.56460.6010
MUSIQ↑29.03736.06963.58759.84959.63545.81848.69857.76955.18964.541
MANIQA↑0.20710.17830.37020.36940.34360.26630.26550.36910.33370.3828
NIQE↓7.736.954.654.684.685.986.225.966.934.02
RealSetCLIPQA↑0.37390.52820.61600.60810.57780.49660.43130.63090.61880.6392
MUSIQ↑42.36650.53965.58364.12563.81755.70848.60259.31958.51663.519
MANIQA↑0.31000.29270.38880.39490.38180.31340.26930.39160.35260.3931
NIQE↓4.934.814.584.384.404.726.475.966.464.23
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.421, + 0.466, + 0.437 + ], + "angle": 0, + "content": "4.3. Comparisons with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.444, + 0.484, + 0.58 + ], + "angle": 0, + "content": "We select ImageNet-Test [3, 44] that contains 3,000 images as major dataset for synthetic image super-resolution evaluation. Furthermore, two commonly used real-world datasets, RealSR [1] and RealSet65 [44], are adopted to evaluate the generalizability in real-world scenarios. We make comparisons with several GAN-based methods ESRGAN [36], RealSR-JPEG [10], BSRGAN [46], RealESRGAN [37], SwinIR [19], DASR [20], and two diffusion-based methods LDM [26] and ResShift [44]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.484, + 0.822 + ], + "angle": 0, + "content": "The quantitative results on three benchmark datasets are shown in Tab. 4. We present UPSR with five sampling steps in order to trade off the super-resolution performance and computational consumption. The proposed UPSR achieves better CLIPIQA, MUSIQ, MANIQA and NIQE compared to ResShift, indicating the significant improvement in perceptual quality. Specifically, while reducing \\(30\\%\\) of the overall model size, UPSR-5 still outperforms ResShift-4 by 7.21 and 2.10 in terms of MUSIQ and NIQE in ImageNet-Test dataset. For real-world datasets, our method also yields impressive results. UPSR-5 consistently outperforms ResShift-4 across both the RealSR and RealSet datasets, demonstrating the efficacy of UPSR in addressing unknown degradation. Furthermore, UPSR-5 achieves the best NIQE metric, while recent diffusion-based methods show poorer NIQE performance compared to most GAN-based methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.258, + 0.849 + ], + "angle": 0, + "content": "4.4. Visual Examples" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.9 + ], + "angle": 0, + "content": "In this section, we present several visual examples to illustrate the proposed UNW scheme and the comparisons between different methods." + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.423, + 0.897, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.577, + 0.907, + 0.675 + ], + "angle": 0, + "content": "Figure 5. Visual examples of the proposed UNW strategy. Based on the uncertainty estimate (illustrated as the heatmap), the noise level in most flat areas is reduced to preserve more details for better SR results. Meanwhile, noise in edge areas (e.g., in image (a)) and severely degraded parts (e.g., in image (b)) are maintained relatively heavy to ensure reliable score estimation to produce visually pleasing results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.693, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Visualization of Weighted Noise. In Fig. 5 we provide visual examples of isotropic and anisotropic noises applied in different images. The proposed UNW strategy enables adaptive perturbation across different areas, leading to the anisotropic diffusion process. The reduced noise level in flat areas, particularly in clear backgrounds, help retain more details in the initial state without sacrificing generative capability. Additionally, these retained details provide more information to reconstruct missing details in surrounding regions, thereby enhancing overall performance. Conversely, in edge regions, the noise level is maintained at the predefined \\(\\sigma_{max}\\) due to their high uncertainty, ensuring sufficient perturbation to prevent the model from generating over-smoothed results." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17986" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.173, + 0.089, + 0.367, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.209, + 0.24, + 0.34, + 0.252 + ], + "angle": 0, + "content": "ImageNet val 0079" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.09, + 0.48, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.158, + 0.457, + 0.169 + ], + "angle": 0, + "content": "BSRGAN" + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.09, + 0.596, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.492, + 0.158, + 0.585, + 0.169 + ], + "angle": 0, + "content": "RealESRGAN" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.09, + 0.708, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.628, + 0.158, + 0.677, + 0.169 + ], + "angle": 0, + "content": "SwinIR" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.09, + 0.824, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.158, + 0.788, + 0.169 + ], + "angle": 0, + "content": "DASR" + }, + { + "type": "image", + "bbox": [ + 0.174, + 0.254, + 0.367, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.229, + 0.404, + 0.319, + 0.417 + ], + "angle": 0, + "content": "RealSet lion1" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.173, + 0.48, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.24, + 0.453, + 0.252 + ], + "angle": 0, + "content": "LDM-15" + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.173, + 0.595, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.24, + 0.576, + 0.251 + ], + "angle": 0, + "content": "ResShift-15" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.173, + 0.708, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.619, + 0.24, + 0.687, + 0.251 + ], + "angle": 0, + "content": "ResShift-4" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.173, + 0.824, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.751, + 0.24, + 0.783, + 0.251 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.174, + 0.419, + 0.367, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.211, + 0.568, + 0.338, + 0.582 + ], + "angle": 0, + "content": "RealSR_Nikon_047" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.254, + 0.481, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.322, + 0.457, + 0.333 + ], + "angle": 0, + "content": "BSRGAN" + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.254, + 0.595, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.492, + 0.322, + 0.585, + 0.333 + ], + "angle": 0, + "content": "RealESRGAN" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.254, + 0.708, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.628, + 0.322, + 0.677, + 0.333 + ], + "angle": 0, + "content": "SwinIR" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.254, + 0.824, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.322, + 0.787, + 0.333 + ], + "angle": 0, + "content": "DASR" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.336, + 0.481, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.396, + 0.404, + 0.453, + 0.416 + ], + "angle": 0, + "content": "LDM-15" + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.336, + 0.595, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.404, + 0.576, + 0.416 + ], + "angle": 0, + "content": "ResShift-15" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.336, + 0.708, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.618, + 0.404, + 0.687, + 0.416 + ], + "angle": 0, + "content": "ResShift-4" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.336, + 0.824, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.404, + 0.783, + 0.416 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.419, + 0.481, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.486, + 0.457, + 0.497 + ], + "angle": 0, + "content": "BSRGAN" + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.419, + 0.595, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.492, + 0.486, + 0.585, + 0.497 + ], + "angle": 0, + "content": "RealESRGAN" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.419, + 0.708, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.628, + 0.486, + 0.677, + 0.497 + ], + "angle": 0, + "content": "SwinIR" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.419, + 0.824, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.486, + 0.787, + 0.497 + ], + "angle": 0, + "content": "DASR" + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.592, + 0.842, + 0.607 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison between different methods on real-world datasets. Please zoom in for more details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.622, + 0.483, + 0.729 + ], + "angle": 0, + "content": "Qualitative Comparison. In Fig. 6, we provide visual results produced by different methods on real-world datasets. These results validate that the proposed UPSR method is capable of producing visually better outputs. Specifically, UPSR successfully recover finer textures and sharper edges, while other methods yields blurry results. More visual examples are presented in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.74, + 0.21, + 0.755 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose specialized prior distribution and sampling pipeline, along with modifications on the network architecture to develop the Uncertainty-guided Perturbation scheme for Super-Resolution to tackle real-world SR task. Specifically, we propose region-specific processing based on the LR content from a uncertainty-based perspective. In flat areas (low-uncertainty), reducing the noise level could preserve more details in the initial state of the diffusion process and therefore improve performance, whereas edge and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.623, + 0.907, + 0.819 + ], + "angle": 0, + "content": "texture regions (high-uncertainty) demand stronger noise to bridge their significant disparity with the target distribution. To facilitate region-specific control, we leverage the uncertainty of LR image as the criterion to adjust noise level in different areas. We employ an auxiliary SR network and consider the residual of SR prediction as the approximate uncertainty estimate of LR image due to their similarity. We further incorporate the SR prediction as conditional information and make several architectural modifications to provide a more robust pipeline to further enhance performance. Extensive experiments on both synthetic and real-world datasets validate the efficacy of the proposed UPSR model, even with lower model size and training overhead." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by National Natural Science Foundation of China (No. 62476051) and Sichuan Natural Science Foundation (No. 2024NSFTD0041)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17987" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.185 + ], + "angle": 0, + "content": "[1] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3086-3095, 2019. 7, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.485, + 0.257 + ], + "angle": 0, + "content": "[2] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image superresolution transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22367-22377, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.483, + 0.314 + ], + "angle": 0, + "content": "[3] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.316, + 0.483, + 0.37 + ], + "angle": 0, + "content": "[4] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.373, + 0.483, + 0.428 + ], + "angle": 0, + "content": "[5] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.431, + 0.483, + 0.485 + ], + "angle": 0, + "content": "[6] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.488, + 0.483, + 0.555 + ], + "angle": 0, + "content": "[7] Shuhang Gu, Wangmeng Zuo, Qi Xie, Deyu Meng, Xiangchu Feng, and Lei Zhang. Convolutional sparse coding for image super-resolution. In Proceedings of the IEEE International Conference on Computer Vision, pages 1823-1831, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.558, + 0.483, + 0.6 + ], + "angle": 0, + "content": "[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.603, + 0.483, + 0.644 + ], + "angle": 0, + "content": "[9] Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research, 6(4), 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.647, + 0.483, + 0.715 + ], + "angle": 0, + "content": "[10] Xiaozhong Ji, Yun Cao, Ying Tai, Chengjie Wang, Jilin Li, and Feiyue Huang. Real-world super-resolution via kernel estimation and noise injection. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 466-467, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.718, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[11] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[12] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and improving the training dynamics of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24174-24184, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[14] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[15] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.196, + 0.905, + 0.264 + ], + "angle": 0, + "content": "[16] Changwoo Lee and Ki-Seok Chung. Gram: Gradient rescaling attention model for data uncertainty estimation in single image super resolution. In 2019 18th IEEE International Conference On Machine Learning And Applications (ICMLA), pages 8-13. IEEE, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.268, + 0.905, + 0.322 + ], + "angle": 0, + "content": "[17] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.326, + 0.905, + 0.407 + ], + "angle": 0, + "content": "[18] Yawei Li, Yuchen Fan, Xiaoyu Xiang, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18278-18289, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.412, + 0.905, + 0.48 + ], + "angle": 0, + "content": "[19] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[20] Jie Liang, Hui Zeng, and Lei Zhang. Efficient and degradation-adaptive network for real-world image superresolution. In European Conference on Computer Vision, pages 574-591. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.542, + 0.905, + 0.61 + ], + "angle": 0, + "content": "[21] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.614, + 0.905, + 0.655 + ], + "angle": 0, + "content": "[22] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.658, + 0.905, + 0.711 + ], + "angle": 0, + "content": "[23] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. arXiv preprint arXiv:1802.05957, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.716, + 0.905, + 0.77 + ], + "angle": 0, + "content": "[24] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[25] Qian Ning, Weisheng Dong, Xin Li, Jinjian Wu, and Guangming Shi. Uncertainty-driven loss for single image superresolution. Advances in Neural Information Processing Systems, 34:16398-16409, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 5, 6, 7, 13" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17988" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[27] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE transactions on pattern analysis and machine intelligence, 45(4):4713-4726, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[28] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1874-1883, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.251, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[29] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.323, + 0.482, + 0.363 + ], + "angle": 0, + "content": "[30] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.367, + 0.482, + 0.408 + ], + "angle": 0, + "content": "[31] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.411, + 0.482, + 0.466 + ], + "angle": 0, + "content": "[32] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.497 + ], + "angle": 0, + "content": "[33] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.5, + 0.482, + 0.555 + ], + "angle": 0, + "content": "[34] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.558, + 0.483, + 0.613 + ], + "angle": 0, + "content": "[35] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, 132(12):5929-5949, 2024. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.616, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[36] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[37] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1905-1914, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[38] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image superresolution in a single step. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 25796-25805, 2024. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[39] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. Advances in Neural Information Processing Systems, 37:92529-92553, 2025. 12" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[40] Rui Xie, Chen Zhao, Kai Zhang, Zhenyu Zhang, Jun Zhou, Jian Yang, and Ying Tai. Addsr: Accelerating diffusion-based blind super-resolution with adversarial diffusion distillation. arXiv preprint arXiv:2404.01717, 2024. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[41] Jianchao Yang, John Wright, Thomas S Huang, and Yi Ma. Image super-resolution via sparse representation. IEEE transactions on image processing, 19(11):2861-2873, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[42] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniaq: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1191-1200, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[43] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Efficient diffusion model for image restoration by residual shifting. arXiv preprint arXiv:2403.07319, 2024. 3, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[44] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image superresolution by residual shifting. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 5, 6, 7, 11, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.472 + ], + "angle": 0, + "content": "[45] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In Curves and Surfaces: 7th International Conference, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.474, + 0.905, + 0.543 + ], + "angle": 0, + "content": "[46] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4791-4800, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.545, + 0.905, + 0.626 + ], + "angle": 0, + "content": "[47] Leheng Zhang, Yawei Li, Xingyu Zhou, Xiaorui Zhao, and Shuhang Gu. Transcending the limit of local window: Advanced super-resolution transformer with adaptive token dictionary. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856-2865, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.629, + 0.905, + 0.697 + ], + "angle": 0, + "content": "[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.7, + 0.905, + 0.767 + ], + "angle": 0, + "content": "[49] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 286-301, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.767 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17989" + } + ] +] \ No newline at end of file diff --git a/2025/Uncertainty-guided Perturbation for Image Super-Resolution Diffusion Model/10430da4-a80e-4106-a1b5-f1210ade542e_origin.pdf b/2025/Uncertainty-guided Perturbation for Image Super-Resolution Diffusion Model/10430da4-a80e-4106-a1b5-f1210ade542e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..45d4f7989c1620de3f4c0a20c135ba82e8f1d2ea --- /dev/null +++ b/2025/Uncertainty-guided Perturbation for Image Super-Resolution Diffusion Model/10430da4-a80e-4106-a1b5-f1210ade542e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:675163c1f76e6161f30ce91dd6a1b7d5b24c0519506d79f942a326ffb4800b7b +size 5648193 diff --git a/2025/Uncertainty-guided Perturbation for Image Super-Resolution Diffusion Model/full.md b/2025/Uncertainty-guided Perturbation for Image Super-Resolution Diffusion Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6dbcc785422a281d9bbb837631c8f0e301e030df --- /dev/null +++ b/2025/Uncertainty-guided Perturbation for Image Super-Resolution Diffusion Model/full.md @@ -0,0 +1,357 @@ +# Uncertainty-guided Perturbation for Image Super-Resolution Diffusion Model + +Leheng Zhang Weiyi You Kexuan Shi Shuhang Gu* University of Electronic Science and Technology of China + +{lehengzhang12, shuhanggu}@gmail.com + +https://github.com/LabShuHangGU/UPSR + +# Abstract + +Diffusion-based image super-resolution methods have demonstrated significant advantages over GAN-based approaches, particularly in terms of perceptual quality. Building upon a lengthy Markov chain, diffusion-based methods possess remarkable modeling capacity, enabling them to achieve outstanding performance in real-world scenarios. Unlike previous methods that focus on modifying the noise schedule or sampling process to enhance performance, our approach emphasizes the improved utilization of LR information. We find that different regions of the LR image can be viewed as corresponding to different timesteps in a diffusion process, where flat areas are closer to the target HR distribution but edge and texture regions are farther away. In these flat areas, applying a slight noise is more advantageous for the reconstruction. We associate this characteristic with uncertainty and propose to apply uncertainty estimate to guide region-specific noise level control, a technique we refer to as Uncertainty-guided Noise Weighting. Pixels with lower uncertainty (i.e., flat regions) receive reduced noise to preserve more LR information, therefore improving performance. Furthermore, we modify the network architecture of previous methods to develop our Uncertainty-guided Perturbation SuperResolution (UPSR) model. Extensive experimental results demonstrate that, despite reduced model size and training overhead, the proposed UPSR method outperforms current state-of-the-art methods across various datasets, both quantitatively and qualitatively. + +# 1. Introduction + +Single image super-resolution (SR), which aims to recover clean the high-resolution (HR) image from its degraded, contaminated low-resolution (LR) counterpart, is a classical problem in the computer vision society. Significant information loss during the degradation process underscores the + +![](images/5ce0171a3b3d693bfea78f1e1ef31b5ce3268ad5dbc6f5c9d19c2fc6cdbad31d.jpg) +(a) $\pmb{y}_0$ + +![](images/6523879ee78454190f03b5190d0cca26d6f6f2a1abe839d1dc765e7146f54128.jpg) +(b) $\pmb{x}_{T}^{\mathrm{SR3}} = \pmb{\epsilon}$ + +![](images/4f5f63e5c0ef62f017072d134c86c0be91a5a13ce95af710fefbc500c022fbc5.jpg) +(c) $\mathbf{x}_T^{\mathrm{ResShift}} = \mathbf{y}_0 + \pmb{\epsilon}$ +Figure 1. A comparison of initial state setup between different diffusion-based image super-resolution methods, where $\epsilon \sim \mathcal{N}(\mathbf{0},\sigma_{\mathrm{max}}^2\mathbf{I})$ . (b) SR3 [27] initiate the diffusion process from pure Gaussian noise, whereas (c) ResShift [44] and (d) our UPSR embed the LR input into the initial noise map. Additionally, we apply uncertainty-guided weighting coefficient $w_{u}(\pmb{y}_{0})$ to reduce the noise level in flat areas, achieving a more specialized diffusion process for SR to improve performance. + +![](images/53b441a3b3ab62644bab7b1b18b134843afa9381139f2482dce79d8728242380.jpg) +(d) $\pmb{x}_T^{\mathrm{UPSR}} = \pmb{y}_0 + w_u(\pmb{y}_0)\pmb{\epsilon}$ + +need for robust super-resolution modeling capability to recover missing details and produce visually pleasing results. Existing SR methods have explored a variety of advanced network architectures and complex degradation modeling to improve performance in classical SR [2, 19, 47] and real-world SR [37, 46], respectively. + +Recently, diffusion models [8, 12, 30] have demonstrated impressive capability in image synthesis, offering a promising new approach for the real-world image super-resolution task. These methods transform pure Gaussian noise into high-quality images through a predefined Markov chain, and their solid theoretical foundation endows them with exceptional modeling capacity to bridge diverse data distri + +butions. To leverage the modeling capabilities of diffusion models for recovering missing details in LR images, several methods [26, 27] begin the SR process by sampling from a standard Gaussian distribution and gradually refine the noisy inputs into high-quality outcomes. However, starting from pure noise was originally intended for image synthesis tasks, resulting in suboptimal outcomes for superresolution. Additionally, these methods typically require a lengthy sampling process, limiting their practicality for real-world applications. + +To address the aforementioned issues, ResShift [44] focuses on the construction of the prior distribution, highlighting the importance of LR information during the diffusion process. By embedding the LR image into the initial noise map and progressively recovering the residual between the LR and HR images, ResShift greatly simplifies the diffusion process. Instead of modeling the entire HR image from noise, it only needs to estimate the LR-HR residual, which shortens the sampling process while significantly enhancing super-resolution results. Despite improved performance, challenges remain, as shown in Fig. 1 where most details are obscured by heavy noise. This still poses additional challenges, as exploiting the information from surrounding areas is crucial for recovering missing details in the SR task. + +For the pursuit of designing a more specialized and effective diffusion process for SR to improve performance, we first propose to make better use of the LR information in the prior distribution. In recent studies, little attention has been given to the inherent information in the LR image that flat areas are already close to the target, whereas edge and texture regions are farther away. To leverage this information, we propose to consider different regions in the LR image as being situated at various timesteps of an isotropic diffusion process. This leads to an anisotropic diffusion process, where flat areas are assigned lower noise levels (as $t \to 0$ ), while edge and texture regions receive larger noise (as $t \to T$ ). To achieve region-specific noise control, we draw inspiration from uncertainty-driven SR approaches [16, 25] and employ a simple SR network to estimate the uncertainty (variance) of different areas in the input LR image. This uncertainty reflects the difficulty of recovering HR details and is closely related to the distribution disparity between LR and HR, indicating the amount of noise required. We then propose a technique called Uncertainty-guided Noise Weighting, which applies weight coefficients to noise levels in different regions based on their uncertainty estimates. The weight coefficient is positively correlated with the uncertainty estimate, enabling adaptive reduction of noise intensity in flat areas to preserve more details in the initial state of the diffusion process. Equipped with UNW and a modified network architecture, we introduce the UPSR model, which achieves state-of-the-art per + +formance while reducing computational costs, as validated on both synthetic and real-world SR datasets. + +Our major contributions are summarized as follows: + +- To achieve a more specialized diffusion pipeline for SR, we propose to adjust the noise level for different areas based on their uncertainty estimate, where the noise in flat areas (low uncertainty) is reduced. This Uncertainty-guided Noise Weighting strategy enables adaptive control across different areas and results in improved performance. +- We establish the connection between the residual estimated by a pre-trained SR network and the uncertainty of the LR input. The estimated residual could serve as an approximate measure of uncertainty, enabling effective uncertainty estimation in the proposed uncertainty-guided noise weighting scheme. +- By integrating the proposed uncertainty-guided noise weighting scheme and the modified network architecture, our method achieves state-of-the-art performance across various benchmark datasets with smaller model size and less training overhead. + +# 2. Related Works + +# 2.1. Image Super-Resolution + +The current literature on image super-resolution can be divided into two categories: classical SR and real SR. The former focuses on addressing predefined degradation patterns, e.g., bicubic downsampling. After the era of conventional methods [7, 41, 45], a multitude of works [2, 4, 18, 19, 21, 47, 49] focus on exploring the potential of neural network, ranging from improved convolutional neural networks to transformer variants [33]. These architectural modifications enhance the capacity of modeling complex relationship between LR and HR images, thereby achieving better performance. Despite their success, these methods perform poorly when directly applied to the real SR task, where the degradation model includes a series of unknown and complex noise and blur patterns. BSRGAN [46] incorporates a series of degradation operators to simulate real-world scenarios, providing a robust training baseline. By combining this strategy with a generative adversarial network [6], BSRGAN achieves remarkable fidelity and perceptual quality in the real SR task. Subsequently, RealESRGAN [37] extends this scheme to a high-order degradation model, providing a more robust training data synthesis procedure and better performance. + +# 2.2. Diffusion Models + +Rooted in nonequilibrium thermodynamics, diffusion probabilistic models [29] have emerged as a new trend in recent research of generative models. DDPM [8] proposes a lengthy parameterized Markov chain to bridge the dis + +distribution of high-quality images and the standard Gaussian distribution. Built upon solid theoretic foundations, DDPM simplifies training objectives and achieves a great breakthrough in image synthesis tasks, demonstrating significant generative capacity compared to GAN-based methods [6, 11, 23]. Subsequently, immense amounts of related research have developed various techniques to further improve DDPM by enhancing performance or shortening sampling process. These improvements include deterministic sampling [30], modified noise schedule [24], harnessing latent space [26], second-order ODE solver [12], and improved training dynamics [13]. + +# 2.3. Diffusion-based Image Super-Resolution + +SRDiff [17] and SR3 [27] are the first to apply diffusion models to image super-resolution by taking the LR input as conditional information, demonstrating the efficacy of diffusion model in generating perceptually high-quality SR images. Despite superior performance, SR3 suffers from bias issues and a costly sampling process. LDM-SR [26] trains an autoencoder and performs the diffusion process in its low-dimensional latent space. This approach allows the model to concentrate on perceptually relevant details and significantly improves computational efficiency. ResShift [44] embeds the LR information $\mathbf{y}_0$ into the initial state $\mathbf{x}_T$ and built a new prior distribution: $\mathbf{x}_T \sim \mathcal{N}(\mathbf{x}_T \mid \mathbf{y}_0, \kappa^2 \eta_T \mathbf{I})$ . In this way, the diffusion process shifts from generating the HR image $\mathbf{x}_0$ from pure noise $\epsilon$ to generating the residual $\mathbf{x}_0 - \mathbf{y}_0$ given noisy LR information $\mathbf{y}_0 + \epsilon$ , greatly reducing the difficulty of estimation. The forward and backward transition distributions are defined as: + +$$ +\begin{array}{l} q \left(\boldsymbol {x} _ {t} \mid \boldsymbol {x} _ {t - 1}, \boldsymbol {x} _ {0}, \boldsymbol {y} _ {0}\right) = (1) \\ \mathcal {N} \left(\boldsymbol {x} _ {t} \mid \boldsymbol {x} _ {t - 1} + \alpha_ {t} \left(\boldsymbol {y} _ {0} - \boldsymbol {x} _ {0}\right), \kappa^ {2} \alpha_ {t} \boldsymbol {I}\right) (1) \\ \end{array} +$$ + +and + +$$ +\begin{array}{l} q \left(\boldsymbol {x} _ {t - 1} \mid \boldsymbol {x} _ {t}, \boldsymbol {x} _ {0}, \boldsymbol {y} _ {0}\right) = \\ \mathcal {N} \left(\boldsymbol {x} _ {t - 1} \mid \frac {\eta_ {t - 1}}{\eta_ {t}} \boldsymbol {x} _ {t} + \frac {\alpha_ {t}}{\eta_ {t}} \boldsymbol {x} _ {0}, \kappa^ {2} \frac {\eta_ {t - 1}}{\eta_ {t}} \alpha_ {t} \boldsymbol {I}\right), \tag {2} \\ \end{array} +$$ + +for $t = 1,2,\dots ,T$ ; $\eta_t$ and $\alpha_{t} = \eta_{t} - \eta_{t - 1}$ are time-dependent positive parameters that characterize the velocity of mean shift and noise injection at different timesteps. In this work, we adopt the sampling process and network architecture of ResShift as our baseline and propose a series of modifications to the sampling pipeline and network architecture to enhance performance while significantly saving model size and training overhead. + +# 3. Methodology + +# 3.1. Noise Levels in Diffusion-based SR + +In the context of SDE [31, 32], perturbing the data with Gaussian noise is crucial for facilitating score estima + +![](images/d651a5b712f53a3a2c0125b57a1377ec7ba80af9027481e809cd23e1ae853eff.jpg) +(a) + +![](images/bb2fa6b3a956b4dc8aaea47aa8622d1deb4a38037df25e97a3fc45f8923e8955.jpg) +(b) + +![](images/1ae2a69b2d016cf2f269d1794984c7886bf71bc19f31216781a806b90fb7a5a5.jpg) +Figure 2. (a) The distribution of pixel residual $|y - x|$ computed on ImageNet-Test dataset [44], omitting values where $|y - x| > 0.4$ for clarity. The result exhibits a distinct long-tailed characteristic. (b) The statistical curves of fidelity $|f(y) - x|$ and perceptual quality $|\phi(f(y)) - \phi(x)|$ with respect to residual $|y - x|$ under different noise levels. As $|y - x|$ increases, the gap of fidelity remains relatively stable when different noise levels are applied. In contrast, the perceptual quality is more sensitive to the noise level. A larger noise is more requisite in regions with high residual value to achieve better perceptual quality. Meanwhile, we propose weighted noise level $w_u(y)\sigma_{max}$ which could lead to better results, with details presented in Sec. 3.3. + +tion [9], especially for data residing in low-dimensional manifolds. They apply various noise levels to perturb the data to explore the low-density regions and enable the score estimation towards high-density regions. In the super-resolution task, the target is to estimate the gradients from LR images (low density) to HR images (high density) in the real image distribution. Recent diffusion-based methods [43, 44] inject a large initial noise into the LR image and gradually reduce it throughout the reverse diffusion process, as the LR sample progressively moves closer to high-density regions. However, these methods focus solely on the different amounts of noise needed at different timesteps during the diffusion process, without accounting for the fact that different areas of each image require varying noise intensities. To provide a clearer insight, we conduct several experiments to demonstrate how the fidelity and perceptual quality of the output change when different overall noise levels are applied to a diffusion-based SR model. + +We first analyze the distribution of the residual $|y - x|$ between each pixel pair of the upsampled LR image $y$ and its corresponding HR image $x$ , as illustrated in Fig. 2a. The result clearly indicates that $|y - x|$ follows a long-tailed distribution, with over $95\%$ of the data concentrated in the range [0.01, 0.16]. Next, we focus on data within this range to + +![](images/31017c98198a25fbb9bb37d30f00284812426664277c804673b26dfae4239fa9.jpg) +HR + +![](images/fc7b50b181be7ab0b19f639edd0ca07a47eed10cf9a97e729ee7eee1c411a5fb.jpg) +LR + +![](images/f6e599077e5024dae8d6d19ee2bbd56835aeb0031c6415ac9b98b961edaa16a2.jpg) +Real Residual +Figure 3. A visualization of the actual residual $|\pmb{x}^i - \pmb{y}^i|$ and the estimated residual $|g(\pmb{y}^i) - \pmb{y}^i|$ . The real residual exhibits high values in edges and texture regions, indicating the high uncertainty. The residual estimated by SR network is close to the real one and therefore can serve as a rough estimation of uncertainty. + +![](images/dce30906886ae081c3205f27c5141a9bc98c09f9c6715891c15809f4a70e3e3a.jpg) +Estimated Residual + +investigate the relationships between the residual $|y - x|$ , fidelity $|f(y) - x|$ , perceptual quality $|\phi(f(y)) - \phi(x)|$ of the output of the denoiser $f(\cdot)$ trained under different noise levels $\sigma_{max}$ (in a reasonable range). The results are illustrated in Fig. 2b. Due to the blurring and downsampling degradation, the residual $|y - x|$ tends to be low in flat areas and high in regions characterized by edges and textures. As $|y - x|$ grows, the perceptual quality gap between $|\phi(f(y)) - \phi(x)|_{\sigma_{max}=2.5}$ and $|\phi(f(y)) - \phi(x)|_{\sigma_{max}=0.5}$ increases rapidly, while the fidelity gap remains nearly unchanged. This indicates that perceptual quality is more sensitive to the noise level change, with higher noise levels being particularly important in edge and texture areas. These regions lie in the low-density regions of the real image distribution and are affected by the ill-posed nature of superresolution. Applying a low $\sigma_{max}$ in such areas leads to unreliable score estimation, resulting in perceptually poor, over-smoothed outputs. + +The experimental results support our hypothesis that different regions of an image respond differently to noise and require distinct handling based on their content. Flat regions typically closely resemble the corresponding ground truth, requiring only a slight amount of noise; therefore, we consider them as existing at small timesteps $(t\to 0)$ . Based on this idea, we propose to reduce noise levels applied in flat regions, and the results shown by the red line in Fig. 2b preliminarily verify the effectiveness. In the following subsections, we will dually discuss how to determine the noise level for different types of areas in the LR image. + +# 3.2. Uncertainty Estimation + +Due to the ill-posed nature of the image super-resolution task, perfectly reconstructing HR images from degraded LR inputs is non-trivial. The high-frequency information in edge and texture areas is severely corrupted during the degradation, exhibiting greater variance compared to flat areas and making accurate prediction more challenging. Several works [16, 25] explore this variance from an uncertainty-based perspective [15]. Given an HR image $\boldsymbol{x}^i \in \mathbb{R}^n$ and its corresponding LR image $\boldsymbol{y}^i \in \mathbb{R}^n$ , the uncertainty $\psi^i$ of the SR estimate $g(\boldsymbol{y}^i)$ is related to its resid + +ual with $x^i$ : + +$$ +\boldsymbol {x} ^ {i} = g \left(\boldsymbol {y} ^ {i}\right) + \epsilon \psi \left(g \left(\boldsymbol {y} ^ {i}\right)\right), \tag {3} +$$ + +where $\epsilon$ represents a standard Laplace or Gaussian distribution when $g(\cdot)$ is regularized by $L1$ or $L2$ loss function, respectively. If the reconstruction error $|\pmb{x}^i - g(\pmb{y}^i)|$ is larger than $|\pmb{x}^j - g(\pmb{y}^j)|$ , then it is more likely that $g(\pmb{y}^i)$ has a higher uncertainty $\psi(g(\pmb{y}^i))$ , compared to the uncertainty $\psi(g(\pmb{y}^j))$ of $g(\pmb{y}^j)$ . Similarly, we can associate the uncertainty of $\pmb{y}^i$ with the residual $|\pmb{x}^i - \pmb{y}^i|$ as: + +$$ +\boldsymbol {x} ^ {i} = \boldsymbol {y} ^ {i} + \hat {\epsilon} \psi (\boldsymbol {y} ^ {i}), \tag {4} +$$ + +where $\hat{\epsilon}$ represents an unknown distribution that depends on the degradation pattern. If $g(\cdot)$ is well-trained, we can assume that $g(\pmb{y}^i)$ closely approximates $\pmb{x}^i$ , implying that $|g(\pmb{y}^i) - \pmb{y}^i|$ is similar to $|\pmb{x}^i - \pmb{y}^i|$ . As illustrated in Fig. 3, the residual estimated by $g(\cdot)$ is roughly similar to the real residual. Therefore, we propose to leverage the residual $|g(\pmb{y}^i) - \pmb{y}^i|$ as the estimate of the uncertainty of $\pmb{y}^i$ . Specifically, we define the uncertainty estimate as: + +$$ +\psi_ {e s t} (\boldsymbol {y}) = \frac {1}{2} | g (\boldsymbol {y}) - \boldsymbol {y} |. \tag {5} +$$ + +In the next subsection, we will take this uncertainty estimate as the criterion to adjust the noise intensity. + +# 3.3. Uncertainty-guided Noise Weighting + +Based on the discussion in previous sections, we propose to replace the commonly used isotropic Gaussian noise with anisotropic noise that adapts to the image content. Inspired by [25], which deals with areas with higher uncertainty by assigning larger weights to the loss function to impose stronger constraints, we suggest adjusting the noise intensity based on the uncertainty estimate across different regions. Specifically, after obtaining the uncertainty estimate as outlined in Sec. 3.2, we compute the weighting coefficient $w_{u}$ and apply it to modulate noise levels for different areas in the diffusion process. We refer to this strategy as Uncertainty-guided Noise Weighting (UNW). For $y$ with lower uncertainty (e.g. in flat areas), it is more likely that $y$ resides in high-density regions of the HR data distribution. In such cases, a lower noise level is sufficient to achieve perceptual quality comparable to that of a higher noise level, while also preserving more LR details, as discussed in Sec. 3.1. Conversely, if $y$ exhibits high uncertainty (e.g., in edge or texture areas), a higher noise level is requisite to account for the significant distribution disparity and to provide greater chances of reconstructing photorealistic details. Therefore, we model the noise weighting coefficient $w_{u}$ as a monotonically increasing function with respect to the uncertainty: + +$$ +w _ {u} (\boldsymbol {y}) := u \left(\psi_ {\text {e s t}} (\boldsymbol {y})\right), \tag {6} +$$ + +![](images/174cb11b74ab35961d8efc4ffc7ffad388aa3c9e86fbdac009db44ad3d8b3173.jpg) +Figure 4. The overall pipeline of the proposed UPSR model. An auxiliary SR network is first employed to estimate the uncertainty of the input $\mathbf{y}_0$ . Then the weighting coefficient $w_{u}$ computed based on the uncertainty $\psi_{est}(\mathbf{y}_0)$ are applied to adjust the noise level in different regions. Meanwhile, both the SR estimate $g(\mathbf{y}_0)$ and LR input $\mathbf{y}_0$ are concatenated as the conditional information for the denoiser $f_{\theta}(\cdot)$ . + +where $w_{u}(\pmb{y}) \in \mathbb{R}^{n \times n}$ represents the diagonal weight matrix used to lower the noise level in low-uncertainty areas while maintaining a higher noise level in high-uncertainty regions, creating a more specialized and adaptive diffusion pipeline for SR. Details of the implementation of $u(\cdot)$ are presented in the supplementary material. + +Built upon the UNW technique, we introduce a new pipeline for diffusion-based SR, termed Uncertainty-guided Perturbation for SR (UPSR), as illustrated in Fig. 4. Given an LR image $\pmb{y}_0$ , we first obtain its SR estimate $g(\pmb{y}_0)$ through an auxiliary SR network $g(\cdot)$ . Then we estimate the uncertainty of $\pmb{y}_0$ as $\psi_{est}(\pmb{y}_0) = \frac{1}{2} |g(\pmb{y}_0) - \pmb{y}_0|$ and obtain the weighting coefficient as $w_u(\pmb{y}_0) = u(\psi_{est}(\pmb{y}_0))$ . Based on the weighting coefficient, we rewrite the forward transition distribution $q(\pmb{x}_t \mid \pmb{x}_{t-1}, \pmb{x}_0, \pmb{y}_0)$ in Eq. 1 as: + +$$ +\left. \mathcal {N} \left(\boldsymbol {x} _ {t} \mid \boldsymbol {x} _ {t - 1} + \alpha_ {t} \left(\boldsymbol {y} _ {0} - \boldsymbol {x} _ {0}\right), \kappa^ {2} w _ {u} \left(\boldsymbol {y} _ {0}\right) ^ {2} \alpha_ {t} \boldsymbol {I}\right), \right. \tag {7} +$$ + +and derive the corresponding backward transition distribution $q(\pmb{x}_{t-1} \mid \pmb{x}_t, \pmb{x}_0, \pmb{y}_0)$ as: + +$$ +\mathcal {N} \left(\boldsymbol {x} _ {t - 1} \mid \frac {\eta_ {t - 1}}{\eta_ {t}} \boldsymbol {x} _ {t} + \frac {\alpha_ {t}}{\eta_ {t}} \boldsymbol {x} _ {0}, \kappa^ {2} w _ {u} (\boldsymbol {y} _ {0}) ^ {2} \frac {\eta_ {t - 1}}{\eta_ {t}} \alpha_ {t} \boldsymbol {I}\right), \tag {8} +$$ + +where the difference is that we leverage the weighting coefficient $w_{u}(\pmb{y}_{0})$ to control the noise level. Since $g(\pmb{y}_0)$ is a better estimate of $x_0$ compared to $y_{0}$ , we combine $g(\pmb{y}_0)$ with $y_{0}$ to provide more accurate conditional information for the denoiser. Details of the derivations of Eq. 7, 8 and the training pipeline are presented in the supplementary materials. Following the previous diffusion-based methods [8, 44], our training objective for the denoiser to predict the target $x_0$ combines both pixel distance $||\cdot||_2^2$ and LPIPS criterion $L_{per}$ : + +$$ +\begin{array}{l} \mathcal {L} (\theta) = \sum_ {t} [ \| f _ {\theta} (\boldsymbol {x} _ {t}, \boldsymbol {y} _ {0}, g (\boldsymbol {y} _ {0}), t) - \boldsymbol {x} _ {0} \| _ {2} ^ {2} \tag {9} \\ + \lambda L _ {p e r} \left(f _ {\theta} \left(\boldsymbol {x} _ {t}, \boldsymbol {y} _ {0}, g (\boldsymbol {y} _ {0}), t\right), \boldsymbol {x} _ {0}\right) ], \\ \end{array} +$$ + +where $\lambda$ is a hyperparameter to control the trade-off between fidelity and perceptual quality. Optimizing the denoiser $f_{\theta}(\cdot)$ through the mixed objective function facilitates fewer diffusion steps while achieving better photo-realistic results in various benchmark real-world datasets. + +# 3.4. Network Architecture Modification + +Besides altering the noise injection scheme and sampling procedure, we also make several modifications to the network architecture. + +We rethink the necessity of leveraging latent space for image super-resolution with only a few diffusion steps. [26] and [44] utilize VQGAN [5] to transfer the diffusion process from pixel space to latent space, for the sake of reducing spatial dimensionality to improve efficiency. However, as the number of sampling steps decreases, the performance-cost ratio of VQGAN gradually decreases due to its own huge computing resource consumption. Moreover, the application of certain perceptual criteria (e.g., LPIPS [48]) requires pixel-level computation, leading to additional decoding costs during the training process. To mitigate its impact, we propose to replace the VQGAN encoder and decoder with PixelUnshuffle operation [28] and a simple nearest neighbor upsampling module, respectively. This approach enables compressing and expanding the spatial dimension with minimal computational cost. Therefore, we can perform the diffusion process directly in pixel space while maintaining a complexity comparable to that in the latent space. In this vein, we significantly reduce the total model size and training overhead without sacrificing performance. + +# 4. Experiments + +# 4.1. Experimental Setups + +For data preparation, we apply randomly cropped $256 \times 256$ patches from ImageNet [3] as HR training data following + +Table 1. Ablation study on effects of the proposed components, including the SR condition and uncertainty-guided noise weighting (UNW). The best results are highlighted in bold. + +
UNWSR cond.RealSRRealSet
PSNR↑CLIPQA↑MUSIQ↑MANIQA↑NIQE↓CLIPQA↑MUSIQ↑MANIQA↑NIQE↓
26.180.544762.9510.35964.490.614164.3600.37184.42
✓26.120.576064.5120.37174.180.634064.2800.38364.22
✓✓26.440.601064.5410.38184.020.638963.4980.39314.24
+ +[26] and [44]. Then the degradation pipeline of RealESR-GAN [37] are adopted to generate degraded $64 \times 64$ LR inputs. For network architecture, we employ the commonly-used U-Net denoiser $f_{\theta}(\cdot)$ with few adjustments to fit our refined diffusion pipeline. For $g(\cdot)$ , we employ the lightweight version of [47] and pretrain it following the real SR pipeline to produce SR predictions and uncertainty estimates. During the diffusion model training process, we freeze the pretrained $g(\cdot)$ and optimize $f_{\theta}(\cdot)$ for 200k iterations with a batchsize of 32. To quantitatively validate the efficacy of the proposed method, we utilize three full-reference metrics: PSNR, SSIM and LPIPS [48]; and four non-reference metrics CLIPIQA [34], MUSIQ [14], MANIQA [42], and NIQE [22] to evaluate the performance. Among these metrics, PSNR and SSIM reflect the fidelity of the generated super-resolution results, while other metrics assess the perceptual quality of the outputs. + +# 4.2. Analysis + +In this subsection, we conduct ablation studies on the effectiveness of several key components and analyze the performance-cost trade-off of the proposed method. + +Effectiveness of UNW and SR conditioning. In order to show the efficacy of the proposed uncertainty-guided noise weighting scheme and SR conditioning, we use the noise schedule and sampling process of ResShift [44] as the baseline. We first apply the uncertainty-guided noise weighting scheme to adjust the noise intensity across different regions based on the uncertainty estimate. The weighting scheme yields better outcomes in terms of perceptual quality, with enhancements of 0.0355 in CLIPIQA and 1.512 in MUSIQ. We attribute this improvement to the reduced noise intensities applied to flat areas, which allow these clearer regions to provide additional information that aids in more accurately recovering the remaining degraded regions. This validates our idea of reducing noise in flat regions does not degrade perceptual quality, but rather enhances it. Next, we incorporate the SR prediction $g(\pmb{y}_0)$ and LR input $\pmb{y}_0$ as the conditional information entered into the denoiser $f_{\theta}(\cdot)$ . The additional SR prediction offers more precise supplementary information compared to relying solely on the degraded LR input, resulting in a significant improvement in both fi + +Table 2. Model size and computational efficiency comparisons between the proposed UPSR and other diffusion-based methods. Gray numbers denote the parameter counts for auxiliary network, i.e., VQGAN in ResShift and $g(\cdot)$ in our work. We test the runtimes on $32 \times 3 \times 64 \times 64$ LR input using single RTX4090 GPU and present several results evaluated on the RealSR dataset. + +
ModelParams (M)Runtimes (s)MUSIQMANIQA
LDM-15113.60+55.321.5948.6980.2655
ResShift-15118.59+55.321.9857.7690.3691
ResShift-4118.59+55.321.0055.1890.3337
UPSR-5119.42+2.501.1264.5410.3818
+ +Table 3. Training overhead comparison between ResShift and UPSR evaluated under a batchsize of 16 per GPU. + +
ModelTraining SpeedMemory Footprint
ResShift1.20 s/iter24.1 G
UPSR0.45 s/iter14.9 G
+ +delity and perceptual quality. Specifically, this more robust pipeline further improves PSNR by 0.34 dB and CLIPIQA by 0.0247 in the RealSR dataset. These ablation studies verify the significance of leveraging LR information during diffusion-based SR process. + +Model Size and Training Overhead Comparison. In section 3.4, we discussed replacing VQGAN in diffusion-based SR model with simpler downsampling and upsampling modules to improve efficiency. To validate the efficacy of the modified architecture, we first evaluate the model size and computational cost of UPSR compared to state-of-the-art diffusion-based methods, as shown in Tab. 2. With the modified architecture and pipeline, UPSR-5 reduces the overall model size by $30\%$ and achieves better perceptual quality with an inference speed comparable to that of ResShift-4. In contrast, ResShift-4 achieves the acceleration at the expense of performance. We then assess the training overhead of different methods, as shown in Tab. 3. Removing VQGAN results in substantial reductions in training overhead, increasing training speed by $167\%$ and saving GPU memory footprint by $38\%$ . These results validate that UPSR strikes a better trade-off between performance and efficiency. + +Table 4. Quantitative results of different methods on one synthetic dataset ImageNet-Test and two real-world datasets RealSR and RealSet. The best and second best results are marked in red and blue. All results of the previous methods are evaluated using the released inference codes and pretrained weights. + +
DatasetsMetricsGAN-based MethodsDiffusion-based Methods
ESRGANRealSR-JPEGBSRGANRealESRGANSwinIRDASRLDM-15ResShift-15ResShift-4UPSR-5
ImageNet -TestPSNR↑20.6723.1124.4224.0423.9924.7524.8524.9425.0223.77
SSIM↑0.44850.59120.65850.66490.66660.67490.66820.67380.68300.6296
LPIPS↓0.48510.32630.25850.25390.23760.24980.26850.23710.20750.2456
CLIPQA↑0.45120.53660.58100.52410.56390.53620.50950.58600.60030.6328
MUSIQ↑43.61546.98154.69652.60953.78948.33746.63953.18252.01959.227
MANIQA↑0.32120.30650.38650.36890.38820.32920.33050.41910.38850.4591
NIQE↓8.335.966.086.075.895.867.216.887.345.24
RealSRPSNR↑27.5727.3426.5125.8326.4327.1927.1826.8025.7726.44
SSIM↑0.77420.76050.77460.77260.78610.78610.78530.76740.74390.7589
LPIPS↓0.41520.39620.26850.27390.25150.31130.30210.34110.34910.2871
CLIPQA↑0.23620.36130.54390.49230.46550.36280.37480.57090.56460.6010
MUSIQ↑29.03736.06963.58759.84959.63545.81848.69857.76955.18964.541
MANIQA↑0.20710.17830.37020.36940.34360.26630.26550.36910.33370.3828
NIQE↓7.736.954.654.684.685.986.225.966.934.02
RealSetCLIPQA↑0.37390.52820.61600.60810.57780.49660.43130.63090.61880.6392
MUSIQ↑42.36650.53965.58364.12563.81755.70848.60259.31958.51663.519
MANIQA↑0.31000.29270.38880.39490.38180.31340.26930.39160.35260.3931
NIQE↓4.934.814.584.384.404.726.475.966.464.23
+ +# 4.3. Comparisons with State-of-the-Art Methods + +We select ImageNet-Test [3, 44] that contains 3,000 images as major dataset for synthetic image super-resolution evaluation. Furthermore, two commonly used real-world datasets, RealSR [1] and RealSet65 [44], are adopted to evaluate the generalizability in real-world scenarios. We make comparisons with several GAN-based methods ESRGAN [36], RealSR-JPEG [10], BSRGAN [46], RealESRGAN [37], SwinIR [19], DASR [20], and two diffusion-based methods LDM [26] and ResShift [44]. + +The quantitative results on three benchmark datasets are shown in Tab. 4. We present UPSR with five sampling steps in order to trade off the super-resolution performance and computational consumption. The proposed UPSR achieves better CLIPIQA, MUSIQ, MANIQA and NIQE compared to ResShift, indicating the significant improvement in perceptual quality. Specifically, while reducing $30\%$ of the overall model size, UPSR-5 still outperforms ResShift-4 by 7.21 and 2.10 in terms of MUSIQ and NIQE in ImageNet-Test dataset. For real-world datasets, our method also yields impressive results. UPSR-5 consistently outperforms ResShift-4 across both the RealSR and RealSet datasets, demonstrating the efficacy of UPSR in addressing unknown degradation. Furthermore, UPSR-5 achieves the best NIQE metric, while recent diffusion-based methods show poorer NIQE performance compared to most GAN-based methods. + +# 4.4. Visual Examples + +In this section, we present several visual examples to illustrate the proposed UNW scheme and the comparisons between different methods. + +![](images/069bfc6d62c3261af3c570caa9f0d953953290d04beff652e8a8567ef5ce3f3d.jpg) +Figure 5. Visual examples of the proposed UNW strategy. Based on the uncertainty estimate (illustrated as the heatmap), the noise level in most flat areas is reduced to preserve more details for better SR results. Meanwhile, noise in edge areas (e.g., in image (a)) and severely degraded parts (e.g., in image (b)) are maintained relatively heavy to ensure reliable score estimation to produce visually pleasing results. + +Visualization of Weighted Noise. In Fig. 5 we provide visual examples of isotropic and anisotropic noises applied in different images. The proposed UNW strategy enables adaptive perturbation across different areas, leading to the anisotropic diffusion process. The reduced noise level in flat areas, particularly in clear backgrounds, help retain more details in the initial state without sacrificing generative capability. Additionally, these retained details provide more information to reconstruct missing details in surrounding regions, thereby enhancing overall performance. Conversely, in edge regions, the noise level is maintained at the predefined $\sigma_{max}$ due to their high uncertainty, ensuring sufficient perturbation to prevent the model from generating over-smoothed results. + +![](images/1f2aaf03c8afccc2c6af213a7e3a19f6c383a311f5fcf00effa1f06d32ab353d.jpg) +ImageNet val 0079 + +![](images/8612017f43e00ec52fb47b1d149174ca1496576723234bb7d75d113186dd1cef.jpg) +BSRGAN + +![](images/a597c2cae3ae5f90ae8cd28c0081d898454a03fab3cef5792afad59dbab1abce.jpg) +RealESRGAN + +![](images/ab5c66c7677b34d74b7f0dcc945b2385af73bf01f03b2fb73fa0e5af17493a37.jpg) +SwinIR + +![](images/8d9fbaae7598fe3023ec8050ddabaabd64b0dccf995ef09e3afaf42bfc0d1850.jpg) +DASR + +![](images/420f3298cbe28b597a5d32755eebdcde63fb0f87bb99abe1b65a783c3118ced2.jpg) +RealSet lion1 + +![](images/3b75d8c3fdf20a7d4486776507d46c440de06fd04b13b0655267c1e79378df34.jpg) +LDM-15 + +![](images/a3cd768b1e35f8b9c51707228c16a300953dc05421ce5c0c9ce48c0848ae65e1.jpg) +ResShift-15 + +![](images/f1b264e492ea056ccbad8d058809c648e19d7b5d51dbba6609e1c426b9463552.jpg) +ResShift-4 + +![](images/6ba80c14643f9dc85ddd1053c9d0a23caa5133f6e08d9794b4a5a06a6fe5e787.jpg) +Ours + +![](images/edc46ab5f7d3db251268e91e68e54d7edb654a6f7255b842ede5c4e925261e94.jpg) +RealSR_Nikon_047 +Figure 6. Qualitative comparison between different methods on real-world datasets. Please zoom in for more details. + +![](images/cd48bf2bd1e2bdf4a7452ee004db74d6df3aabe63d34cfd5da5e386f2585446c.jpg) +BSRGAN + +![](images/cd9d6a4c77072601bdd8c7eb2591fb8e750bcfdc20d09e0c0222c08e84d60936.jpg) +RealESRGAN + +![](images/1485f27b6d3da9538d9e540823c18d08abfd19e984c4aaab3af4d245cf341717.jpg) +SwinIR + +![](images/b5738fb580fc7c8b6c354ef7251dca7e32b3a00d25ba18abc2d0c67d45d45217.jpg) +DASR + +![](images/d905cde3131a53209b3501413a8def071a0046cae149960786472d3827f5e0ae.jpg) +LDM-15 + +![](images/3744f14f7292279e0527f13917de22152ddb0eb12375ccf2346610fa34bd486d.jpg) +ResShift-15 + +![](images/a171951f36154a1e5a140d4f21c7d7caec5a8112fc516b63db77dc467e625342.jpg) +ResShift-4 + +![](images/f50e01ccc54e6f7dc071241fb1f11fae7f71adff2a7925511f95fb7103d70033.jpg) +Ours + +![](images/a0b0a6c8975af49113d9d1b7d80ab2f2416a9ce5b637388329328bfa757c8869.jpg) +BSRGAN + +![](images/ce42b425bef3092cdc3753ead8b94844bbda50db5594ac408835442c37501c50.jpg) +RealESRGAN + +![](images/6bdf6b5539c279b5d75b8de85011de35c49f64d1c041195a3f6fcdd90fc07512.jpg) +SwinIR + +![](images/050394a69fcda9eee7a86f87b84052a38a61f4ad663c03b6e0192923912f32c5.jpg) +DASR + +Qualitative Comparison. In Fig. 6, we provide visual results produced by different methods on real-world datasets. These results validate that the proposed UPSR method is capable of producing visually better outputs. Specifically, UPSR successfully recover finer textures and sharper edges, while other methods yields blurry results. More visual examples are presented in the supplementary material. + +# 5. Conclusion + +In this work, we propose specialized prior distribution and sampling pipeline, along with modifications on the network architecture to develop the Uncertainty-guided Perturbation scheme for Super-Resolution to tackle real-world SR task. Specifically, we propose region-specific processing based on the LR content from a uncertainty-based perspective. In flat areas (low-uncertainty), reducing the noise level could preserve more details in the initial state of the diffusion process and therefore improve performance, whereas edge and + +texture regions (high-uncertainty) demand stronger noise to bridge their significant disparity with the target distribution. To facilitate region-specific control, we leverage the uncertainty of LR image as the criterion to adjust noise level in different areas. We employ an auxiliary SR network and consider the residual of SR prediction as the approximate uncertainty estimate of LR image due to their similarity. We further incorporate the SR prediction as conditional information and make several architectural modifications to provide a more robust pipeline to further enhance performance. Extensive experiments on both synthetic and real-world datasets validate the efficacy of the proposed UPSR model, even with lower model size and training overhead. + +Acknowledgement. This work was supported by National Natural Science Foundation of China (No. 62476051) and Sichuan Natural Science Foundation (No. 2024NSFTD0041). + +# References + +[1] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3086-3095, 2019. 7, 15 +[2] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image superresolution transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22367-22377, 2023. 1, 2 +[3] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5, 7 +[4] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 2 +[5] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 5 +[6] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2, 3 +[7] Shuhang Gu, Wangmeng Zuo, Qi Xie, Deyu Meng, Xiangchu Feng, and Lei Zhang. Convolutional sparse coding for image super-resolution. In Proceedings of the IEEE International Conference on Computer Vision, pages 1823-1831, 2015. 2 +[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2, 5 +[9] Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research, 6(4), 2005. 3 +[10] Xiaozhong Ji, Yun Cao, Ying Tai, Chengjie Wang, Jilin Li, and Feiyue Huang. Real-world super-resolution via kernel estimation and noise injection. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 466-467, 2020. 7 +[11] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 3 +[12] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. 1, 3 +[13] Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and improving the training dynamics of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24174-24184, 2024. 3 + +[14] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 6 +[15] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 4 +[16] Changwoo Lee and Ki-Seok Chung. Gram: Gradient rescaling attention model for data uncertainty estimation in single image super resolution. In 2019 18th IEEE International Conference On Machine Learning And Applications (ICMLA), pages 8-13. IEEE, 2019. 2, 4 +[17] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. 3 +[18] Yawei Li, Yuchen Fan, Xiaoyu Xiang, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18278-18289, 2023. 2 +[19] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 1, 2, 7 +[20] Jie Liang, Hui Zeng, and Lei Zhang. Efficient and degradation-adaptive network for real-world image superresolution. In European Conference on Computer Vision, pages 574-591. Springer, 2022. 7 +[21] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 2 +[22] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a "completely blind" image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 6 +[23] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. arXiv preprint arXiv:1802.05957, 2018. 3 +[24] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 3 +[25] Qian Ning, Weisheng Dong, Xin Li, Jinjian Wu, and Guangming Shi. Uncertainty-driven loss for single image superresolution. Advances in Neural Information Processing Systems, 34:16398-16409, 2021. 2, 4 +[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 5, 6, 7, 13 + +[27] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE transactions on pattern analysis and machine intelligence, 45(4):4713-4726, 2022. 1, 2, 3 +[28] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1874-1883, 2016. 5 +[29] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2 +[30] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 3 +[31] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 3 +[32] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3 +[33] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2 +[34] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 6 +[35] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, 132(12):5929-5949, 2024. 12 +[36] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 7 +[37] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1905-1914, 2021. 1, 2, 6, 7 +[38] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image superresolution in a single step. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 25796-25805, 2024. 13 +[39] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. Advances in Neural Information Processing Systems, 37:92529-92553, 2025. 12 + +[40] Rui Xie, Chen Zhao, Kai Zhang, Zhenyu Zhang, Jun Zhou, Jian Yang, and Ying Tai. Addsr: Accelerating diffusion-based blind super-resolution with adversarial diffusion distillation. arXiv preprint arXiv:2404.01717, 2024. 12 +[41] Jianchao Yang, John Wright, Thomas S Huang, and Yi Ma. Image super-resolution via sparse representation. IEEE transactions on image processing, 19(11):2861-2873, 2010. 2 +[42] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniaq: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1191-1200, 2022. 6 +[43] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Efficient diffusion model for image restoration by residual shifting. arXiv preprint arXiv:2403.07319, 2024. 3, 13 +[44] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image superresolution by residual shifting. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 5, 6, 7, 11, 13, 14 +[45] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In Curves and Surfaces: 7th International Conference, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 2 +[46] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4791-4800, 2021. 1, 2, 7 +[47] Leheng Zhang, Yawei Li, Xingyu Zhou, Xiaorui Zhao, and Shuhang Gu. Transcending the limit of local window: Advanced super-resolution transformer with adaptive token dictionary. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856-2865, 2024. 1, 2, 6 +[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5, 6 +[49] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 286-301, 2018. 2 \ No newline at end of file diff --git a/2025/Uncertainty-guided Perturbation for Image Super-Resolution Diffusion Model/images.zip b/2025/Uncertainty-guided Perturbation for Image Super-Resolution Diffusion Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1c514febce7d74020ccafc1c6c56a309a921f663 --- /dev/null +++ b/2025/Uncertainty-guided Perturbation for Image Super-Resolution Diffusion Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:efb7834bda2c9ebc94034e2405e28cce57ed75a4f786b80c5eb42ca253da59a9 +size 703300 diff --git a/2025/Uncertainty-guided Perturbation for Image Super-Resolution Diffusion Model/layout.json b/2025/Uncertainty-guided Perturbation for Image Super-Resolution Diffusion Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..118dfddd3b6be4d37a2f84a82640bea42706393f --- /dev/null +++ b/2025/Uncertainty-guided Perturbation for Image Super-Resolution Diffusion Model/layout.json @@ -0,0 +1,10437 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 103, + 547, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 103, + 547, + 122 + ], + "spans": [ + { + "bbox": [ + 64, + 103, + 547, + 122 + ], + "type": "text", + "content": "Uncertainty-guided Perturbation for Image Super-Resolution Diffusion Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 163, + 143, + 451, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 163, + 143, + 451, + 172 + ], + "spans": [ + { + "bbox": [ + 163, + 143, + 451, + 172 + ], + "type": "text", + "content": "Leheng Zhang Weiyi You Kexuan Shi Shuhang Gu* University of Electronic Science and Technology of China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 206, + 174, + 403, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 174, + 403, + 186 + ], + "spans": [ + { + "bbox": [ + 206, + 174, + 403, + 186 + ], + "type": "text", + "content": "{lehengzhang12, shuhanggu}@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 206, + 187, + 402, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 187, + 402, + 198 + ], + "spans": [ + { + "bbox": [ + 206, + 187, + 402, + 198 + ], + "type": "text", + "content": "https://github.com/LabShuHangGU/UPSR" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 227, + 201, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 201, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 201, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 252, + 297, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 252, + 297, + 589 + ], + "spans": [ + { + "bbox": [ + 54, + 252, + 297, + 589 + ], + "type": "text", + "content": "Diffusion-based image super-resolution methods have demonstrated significant advantages over GAN-based approaches, particularly in terms of perceptual quality. Building upon a lengthy Markov chain, diffusion-based methods possess remarkable modeling capacity, enabling them to achieve outstanding performance in real-world scenarios. Unlike previous methods that focus on modifying the noise schedule or sampling process to enhance performance, our approach emphasizes the improved utilization of LR information. We find that different regions of the LR image can be viewed as corresponding to different timesteps in a diffusion process, where flat areas are closer to the target HR distribution but edge and texture regions are farther away. In these flat areas, applying a slight noise is more advantageous for the reconstruction. We associate this characteristic with uncertainty and propose to apply uncertainty estimate to guide region-specific noise level control, a technique we refer to as Uncertainty-guided Noise Weighting. Pixels with lower uncertainty (i.e., flat regions) receive reduced noise to preserve more LR information, therefore improving performance. Furthermore, we modify the network architecture of previous methods to develop our Uncertainty-guided Perturbation SuperResolution (UPSR) model. Extensive experimental results demonstrate that, despite reduced model size and training overhead, the proposed UPSR method outperforms current state-of-the-art methods across various datasets, both quantitatively and qualitatively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 612, + 137, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 612, + 137, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 612, + 137, + 624 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 632, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 296, + 694 + ], + "type": "text", + "content": "Single image super-resolution (SR), which aims to recover clean the high-resolution (HR) image from its degraded, contaminated low-resolution (LR) counterpart, is a classical problem in the computer vision society. Significant information loss during the degradation process underscores the" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 226, + 434, + 317 + ], + "blocks": [ + { + "bbox": [ + 316, + 226, + 434, + 317 + ], + "lines": [ + { + "bbox": [ + 316, + 226, + 434, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 226, + 434, + 317 + ], + "type": "image", + "image_path": "5ce0171a3b3d693bfea78f1e1ef31b5ce3268ad5dbc6f5c9d19c2fc6cdbad31d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 319, + 386, + 329 + ], + "lines": [ + { + "bbox": [ + 364, + 319, + 386, + 329 + ], + "spans": [ + { + "bbox": [ + 364, + 319, + 386, + 329 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 364, + 319, + 386, + 329 + ], + "type": "inline_equation", + "content": "\\pmb{y}_0" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 435, + 226, + 553, + 317 + ], + "blocks": [ + { + "bbox": [ + 435, + 226, + 553, + 317 + ], + "lines": [ + { + "bbox": [ + 435, + 226, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 435, + 226, + 553, + 317 + ], + "type": "image", + "image_path": "6523879ee78454190f03b5190d0cca26d6f6f2a1abe839d1dc765e7146f54128.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 474, + 319, + 515, + 328 + ], + "lines": [ + { + "bbox": [ + 474, + 319, + 515, + 328 + ], + "spans": [ + { + "bbox": [ + 474, + 319, + 515, + 328 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 474, + 319, + 515, + 328 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{T}^{\\mathrm{SR3}} = \\pmb{\\epsilon}" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 316, + 330, + 433, + 422 + ], + "blocks": [ + { + "bbox": [ + 316, + 330, + 433, + 422 + ], + "lines": [ + { + "bbox": [ + 316, + 330, + 433, + 422 + ], + "spans": [ + { + "bbox": [ + 316, + 330, + 433, + 422 + ], + "type": "image", + "image_path": "4f5f63e5c0ef62f017072d134c86c0be91a5a13ce95af710fefbc500c022fbc5.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 422, + 412, + 433 + ], + "lines": [ + { + "bbox": [ + 347, + 422, + 412, + 433 + ], + "spans": [ + { + "bbox": [ + 347, + 422, + 412, + 433 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 347, + 422, + 412, + 433 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T^{\\mathrm{ResShift}} = \\mathbf{y}_0 + \\pmb{\\epsilon}" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 441, + 555, + 530 + ], + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 530 + ], + "type": "text", + "content": "Figure 1. A comparison of initial state setup between different diffusion-based image super-resolution methods, where " + }, + { + "bbox": [ + 313, + 441, + 555, + 530 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\sigma_{\\mathrm{max}}^2\\mathbf{I})" + }, + { + "bbox": [ + 313, + 441, + 555, + 530 + ], + "type": "text", + "content": ". (b) SR3 [27] initiate the diffusion process from pure Gaussian noise, whereas (c) ResShift [44] and (d) our UPSR embed the LR input into the initial noise map. Additionally, we apply uncertainty-guided weighting coefficient " + }, + { + "bbox": [ + 313, + 441, + 555, + 530 + ], + "type": "inline_equation", + "content": "w_{u}(\\pmb{y}_{0})" + }, + { + "bbox": [ + 313, + 441, + 555, + 530 + ], + "type": "text", + "content": " to reduce the noise level in flat areas, achieving a more specialized diffusion process for SR to improve performance." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 436, + 330, + 554, + 422 + ], + "blocks": [ + { + "bbox": [ + 436, + 330, + 554, + 422 + ], + "lines": [ + { + "bbox": [ + 436, + 330, + 554, + 422 + ], + "spans": [ + { + "bbox": [ + 436, + 330, + 554, + 422 + ], + "type": "image", + "image_path": "53b441a3b3ab62644bab7b1b18b134843afa9381139f2482dce79d8728242380.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 422, + 536, + 433 + ], + "lines": [ + { + "bbox": [ + 453, + 422, + 536, + 433 + ], + "spans": [ + { + "bbox": [ + 453, + 422, + 536, + 433 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 453, + 422, + 536, + 433 + ], + "type": "inline_equation", + "content": "\\pmb{x}_T^{\\mathrm{UPSR}} = \\pmb{y}_0 + w_u(\\pmb{y}_0)\\pmb{\\epsilon}" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 555, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 555, + 628 + ], + "type": "text", + "content": "need for robust super-resolution modeling capability to recover missing details and produce visually pleasing results. Existing SR methods have explored a variety of advanced network architectures and complex degradation modeling to improve performance in classical SR [2, 19, 47] and real-world SR [37, 46], respectively." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Recently, diffusion models [8, 12, 30] have demonstrated impressive capability in image synthesis, offering a promising new approach for the real-world image super-resolution task. These methods transform pure Gaussian noise into high-quality images through a predefined Markov chain, and their solid theoretical foundation endows them with exceptional modeling capacity to bridge diverse data distri" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 703, + 141, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 703, + 141, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 703, + 141, + 713 + ], + "type": "text", + "content": "* corresponding author" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17980" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "butions. To leverage the modeling capabilities of diffusion models for recovering missing details in LR images, several methods [26, 27] begin the SR process by sampling from a standard Gaussian distribution and gradually refine the noisy inputs into high-quality outcomes. However, starting from pure noise was originally intended for image synthesis tasks, resulting in suboptimal outcomes for superresolution. Additionally, these methods typically require a lengthy sampling process, limiting their practicality for real-world applications." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 195, + 295, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 195, + 295, + 375 + ], + "spans": [ + { + "bbox": [ + 55, + 195, + 295, + 375 + ], + "type": "text", + "content": "To address the aforementioned issues, ResShift [44] focuses on the construction of the prior distribution, highlighting the importance of LR information during the diffusion process. By embedding the LR image into the initial noise map and progressively recovering the residual between the LR and HR images, ResShift greatly simplifies the diffusion process. Instead of modeling the entire HR image from noise, it only needs to estimate the LR-HR residual, which shortens the sampling process while significantly enhancing super-resolution results. Despite improved performance, challenges remain, as shown in Fig. 1 where most details are obscured by heavy noise. This still poses additional challenges, as exploiting the information from surrounding areas is crucial for recovering missing details in the SR task." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 378, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 295, + 713 + ], + "type": "text", + "content": "For the pursuit of designing a more specialized and effective diffusion process for SR to improve performance, we first propose to make better use of the LR information in the prior distribution. In recent studies, little attention has been given to the inherent information in the LR image that flat areas are already close to the target, whereas edge and texture regions are farther away. To leverage this information, we propose to consider different regions in the LR image as being situated at various timesteps of an isotropic diffusion process. This leads to an anisotropic diffusion process, where flat areas are assigned lower noise levels (as " + }, + { + "bbox": [ + 55, + 378, + 295, + 713 + ], + "type": "inline_equation", + "content": "t \\to 0" + }, + { + "bbox": [ + 55, + 378, + 295, + 713 + ], + "type": "text", + "content": "), while edge and texture regions receive larger noise (as " + }, + { + "bbox": [ + 55, + 378, + 295, + 713 + ], + "type": "inline_equation", + "content": "t \\to T" + }, + { + "bbox": [ + 55, + 378, + 295, + 713 + ], + "type": "text", + "content": "). To achieve region-specific noise control, we draw inspiration from uncertainty-driven SR approaches [16, 25] and employ a simple SR network to estimate the uncertainty (variance) of different areas in the input LR image. This uncertainty reflects the difficulty of recovering HR details and is closely related to the distribution disparity between LR and HR, indicating the amount of noise required. We then propose a technique called Uncertainty-guided Noise Weighting, which applies weight coefficients to noise levels in different regions based on their uncertainty estimates. The weight coefficient is positively correlated with the uncertainty estimate, enabling adaptive reduction of noise intensity in flat areas to preserve more details in the initial state of the diffusion process. Equipped with UNW and a modified network architecture, we introduce the UPSR model, which achieves state-of-the-art per" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "formance while reducing computational costs, as validated on both synthetic and real-world SR datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 325, + 97, + 536, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 97, + 536, + 108 + ], + "spans": [ + { + "bbox": [ + 325, + 97, + 536, + 108 + ], + "type": "text", + "content": "Our major contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 110, + 553, + 324 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 110, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 110, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 314, + 110, + 553, + 191 + ], + "type": "text", + "content": "- To achieve a more specialized diffusion pipeline for SR, we propose to adjust the noise level for different areas based on their uncertainty estimate, where the noise in flat areas (low uncertainty) is reduced. This Uncertainty-guided Noise Weighting strategy enables adaptive control across different areas and results in improved performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 193, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 193, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 553, + 263 + ], + "type": "text", + "content": "- We establish the connection between the residual estimated by a pre-trained SR network and the uncertainty of the LR input. The estimated residual could serve as an approximate measure of uncertainty, enabling effective uncertainty estimation in the proposed uncertainty-guided noise weighting scheme." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 265, + 553, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 553, + 324 + ], + "type": "text", + "content": "- By integrating the proposed uncertainty-guided noise weighting scheme and the modified network architecture, our method achieves state-of-the-art performance across various benchmark datasets with smaller model size and less training overhead." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 337, + 405, + 349 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 337, + 405, + 349 + ], + "spans": [ + { + "bbox": [ + 314, + 337, + 405, + 349 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 357, + 449, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 357, + 449, + 370 + ], + "spans": [ + { + "bbox": [ + 314, + 357, + 449, + 370 + ], + "type": "text", + "content": "2.1. Image Super-Resolution" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 375, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 555, + 639 + ], + "type": "text", + "content": "The current literature on image super-resolution can be divided into two categories: classical SR and real SR. The former focuses on addressing predefined degradation patterns, e.g., bicubic downsampling. After the era of conventional methods [7, 41, 45], a multitude of works [2, 4, 18, 19, 21, 47, 49] focus on exploring the potential of neural network, ranging from improved convolutional neural networks to transformer variants [33]. These architectural modifications enhance the capacity of modeling complex relationship between LR and HR images, thereby achieving better performance. Despite their success, these methods perform poorly when directly applied to the real SR task, where the degradation model includes a series of unknown and complex noise and blur patterns. BSRGAN [46] incorporates a series of degradation operators to simulate real-world scenarios, providing a robust training baseline. By combining this strategy with a generative adversarial network [6], BSRGAN achieves remarkable fidelity and perceptual quality in the real SR task. Subsequently, RealESRGAN [37] extends this scheme to a high-order degradation model, providing a more robust training data synthesis procedure and better performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 647, + 417, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 647, + 417, + 659 + ], + "spans": [ + { + "bbox": [ + 314, + 647, + 417, + 659 + ], + "type": "text", + "content": "2.2. Diffusion Models" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "Rooted in nonequilibrium thermodynamics, diffusion probabilistic models [29] have emerged as a new trend in recent research of generative models. DDPM [8] proposes a lengthy parameterized Markov chain to bridge the dis" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "17981" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": "distribution of high-quality images and the standard Gaussian distribution. Built upon solid theoretic foundations, DDPM simplifies training objectives and achieves a great breakthrough in image synthesis tasks, demonstrating significant generative capacity compared to GAN-based methods [6, 11, 23]. Subsequently, immense amounts of related research have developed various techniques to further improve DDPM by enhancing performance or shortening sampling process. These improvements include deterministic sampling [30], modified noise schedule [24], harnessing latent space [26], second-order ODE solver [12], and improved training dynamics [13]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 221, + 268, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 221, + 268, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 268, + 235 + ], + "type": "text", + "content": "2.3. Diffusion-based Image Super-Resolution" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 239, + 296, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 441 + ], + "type": "text", + "content": "SRDiff [17] and SR3 [27] are the first to apply diffusion models to image super-resolution by taking the LR input as conditional information, demonstrating the efficacy of diffusion model in generating perceptually high-quality SR images. Despite superior performance, SR3 suffers from bias issues and a costly sampling process. LDM-SR [26] trains an autoencoder and performs the diffusion process in its low-dimensional latent space. This approach allows the model to concentrate on perceptually relevant details and significantly improves computational efficiency. ResShift [44] embeds the LR information " + }, + { + "bbox": [ + 55, + 239, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_0" + }, + { + "bbox": [ + 55, + 239, + 296, + 441 + ], + "type": "text", + "content": " into the initial state " + }, + { + "bbox": [ + 55, + 239, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 55, + 239, + 296, + 441 + ], + "type": "text", + "content": " and built a new prior distribution: " + }, + { + "bbox": [ + 55, + 239, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T \\sim \\mathcal{N}(\\mathbf{x}_T \\mid \\mathbf{y}_0, \\kappa^2 \\eta_T \\mathbf{I})" + }, + { + "bbox": [ + 55, + 239, + 296, + 441 + ], + "type": "text", + "content": ". In this way, the diffusion process shifts from generating the HR image " + }, + { + "bbox": [ + 55, + 239, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 55, + 239, + 296, + 441 + ], + "type": "text", + "content": " from pure noise " + }, + { + "bbox": [ + 55, + 239, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 239, + 296, + 441 + ], + "type": "text", + "content": " to generating the residual " + }, + { + "bbox": [ + 55, + 239, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0 - \\mathbf{y}_0" + }, + { + "bbox": [ + 55, + 239, + 296, + 441 + ], + "type": "text", + "content": " given noisy LR information " + }, + { + "bbox": [ + 55, + 239, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_0 + \\epsilon" + }, + { + "bbox": [ + 55, + 239, + 296, + 441 + ], + "type": "text", + "content": ", greatly reducing the difficulty of estimation. The forward and backward transition distributions are defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 71, + 444, + 296, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 444, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 71, + 444, + 296, + 472 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} q \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {t - 1}, \\boldsymbol {x} _ {0}, \\boldsymbol {y} _ {0}\\right) = (1) \\\\ \\mathcal {N} \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {t - 1} + \\alpha_ {t} \\left(\\boldsymbol {y} _ {0} - \\boldsymbol {x} _ {0}\\right), \\kappa^ {2} \\alpha_ {t} \\boldsymbol {I}\\right) (1) \\\\ \\end{array}", + "image_path": "d2f8176ea47ff8553353030d044aa76955b80b60d59205020bd692ec97be9d87.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 477, + 74, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 477, + 74, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 477, + 74, + 487 + ], + "type": "text", + "content": "and" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 65, + 489, + 295, + 542 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 489, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 65, + 489, + 295, + 542 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} q \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}, \\boldsymbol {x} _ {0}, \\boldsymbol {y} _ {0}\\right) = \\\\ \\mathcal {N} \\left(\\boldsymbol {x} _ {t - 1} \\mid \\frac {\\eta_ {t - 1}}{\\eta_ {t}} \\boldsymbol {x} _ {t} + \\frac {\\alpha_ {t}}{\\eta_ {t}} \\boldsymbol {x} _ {0}, \\kappa^ {2} \\frac {\\eta_ {t - 1}}{\\eta_ {t}} \\alpha_ {t} \\boldsymbol {I}\\right), \\tag {2} \\\\ \\end{array}", + "image_path": "f7a1841f00e86b3c098942f9732363f56b553eedf2cf8d00c1e08bd5d9a8d035.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 548, + 296, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 296, + 643 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 55, + 548, + 296, + 643 + ], + "type": "inline_equation", + "content": "t = 1,2,\\dots ,T" + }, + { + "bbox": [ + 55, + 548, + 296, + 643 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 55, + 548, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\eta_t" + }, + { + "bbox": [ + 55, + 548, + 296, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 548, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = \\eta_{t} - \\eta_{t - 1}" + }, + { + "bbox": [ + 55, + 548, + 296, + 643 + ], + "type": "text", + "content": " are time-dependent positive parameters that characterize the velocity of mean shift and noise injection at different timesteps. In this work, we adopt the sampling process and network architecture of ResShift as our baseline and propose a series of modifications to the sampling pipeline and network architecture to enhance performance while significantly saving model size and training overhead." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 652, + 138, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 138, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 138, + 666 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 672, + 241, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 241, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 241, + 684 + ], + "type": "text", + "content": "3.1. Noise Levels in Diffusion-based SR" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "In the context of SDE [31, 32], perturbing the data with Gaussian noise is crucial for facilitating score estima" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 357, + 75, + 512, + 162 + ], + "blocks": [ + { + "bbox": [ + 357, + 75, + 512, + 162 + ], + "lines": [ + { + "bbox": [ + 357, + 75, + 512, + 162 + ], + "spans": [ + { + "bbox": [ + 357, + 75, + 512, + 162 + ], + "type": "image", + "image_path": "d651a5b712f53a3a2c0125b57a1377ec7ba80af9027481e809cd23e1ae853eff.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 163, + 440, + 171 + ], + "lines": [ + { + "bbox": [ + 429, + 163, + 440, + 171 + ], + "spans": [ + { + "bbox": [ + 429, + 163, + 440, + 171 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 173, + 434, + 265 + ], + "blocks": [ + { + "bbox": [ + 316, + 173, + 434, + 265 + ], + "lines": [ + { + "bbox": [ + 316, + 173, + 434, + 265 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 434, + 265 + ], + "type": "image", + "image_path": "bb2fa6b3a956b4dc8aaea47aa8622d1deb4a38037df25e97a3fc45f8923e8955.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 266, + 440, + 274 + ], + "lines": [ + { + "bbox": [ + 429, + 266, + 440, + 274 + ], + "spans": [ + { + "bbox": [ + 429, + 266, + 440, + 274 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 434, + 173, + 553, + 265 + ], + "blocks": [ + { + "bbox": [ + 434, + 173, + 553, + 265 + ], + "lines": [ + { + "bbox": [ + 434, + 173, + 553, + 265 + ], + "spans": [ + { + "bbox": [ + 434, + 173, + 553, + 265 + ], + "type": "image", + "image_path": "1ae2a69b2d016cf2f269d1794984c7886bf71bc19f31216781a806b90fb7a5a5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 277, + 555, + 409 + ], + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 409 + ], + "type": "text", + "content": "Figure 2. (a) The distribution of pixel residual " + }, + { + "bbox": [ + 313, + 277, + 555, + 409 + ], + "type": "inline_equation", + "content": "|y - x|" + }, + { + "bbox": [ + 313, + 277, + 555, + 409 + ], + "type": "text", + "content": " computed on ImageNet-Test dataset [44], omitting values where " + }, + { + "bbox": [ + 313, + 277, + 555, + 409 + ], + "type": "inline_equation", + "content": "|y - x| > 0.4" + }, + { + "bbox": [ + 313, + 277, + 555, + 409 + ], + "type": "text", + "content": " for clarity. The result exhibits a distinct long-tailed characteristic. (b) The statistical curves of fidelity " + }, + { + "bbox": [ + 313, + 277, + 555, + 409 + ], + "type": "inline_equation", + "content": "|f(y) - x|" + }, + { + "bbox": [ + 313, + 277, + 555, + 409 + ], + "type": "text", + "content": " and perceptual quality " + }, + { + "bbox": [ + 313, + 277, + 555, + 409 + ], + "type": "inline_equation", + "content": "|\\phi(f(y)) - \\phi(x)|" + }, + { + "bbox": [ + 313, + 277, + 555, + 409 + ], + "type": "text", + "content": " with respect to residual " + }, + { + "bbox": [ + 313, + 277, + 555, + 409 + ], + "type": "inline_equation", + "content": "|y - x|" + }, + { + "bbox": [ + 313, + 277, + 555, + 409 + ], + "type": "text", + "content": " under different noise levels. As " + }, + { + "bbox": [ + 313, + 277, + 555, + 409 + ], + "type": "inline_equation", + "content": "|y - x|" + }, + { + "bbox": [ + 313, + 277, + 555, + 409 + ], + "type": "text", + "content": " increases, the gap of fidelity remains relatively stable when different noise levels are applied. In contrast, the perceptual quality is more sensitive to the noise level. A larger noise is more requisite in regions with high residual value to achieve better perceptual quality. Meanwhile, we propose weighted noise level " + }, + { + "bbox": [ + 313, + 277, + 555, + 409 + ], + "type": "inline_equation", + "content": "w_u(y)\\sigma_{max}" + }, + { + "bbox": [ + 313, + 277, + 555, + 409 + ], + "type": "text", + "content": " which could lead to better results, with details presented in Sec. 3.3." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 426, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 640 + ], + "type": "text", + "content": "tion [9], especially for data residing in low-dimensional manifolds. They apply various noise levels to perturb the data to explore the low-density regions and enable the score estimation towards high-density regions. In the super-resolution task, the target is to estimate the gradients from LR images (low density) to HR images (high density) in the real image distribution. Recent diffusion-based methods [43, 44] inject a large initial noise into the LR image and gradually reduce it throughout the reverse diffusion process, as the LR sample progressively moves closer to high-density regions. However, these methods focus solely on the different amounts of noise needed at different timesteps during the diffusion process, without accounting for the fact that different areas of each image require varying noise intensities. To provide a clearer insight, we conduct several experiments to demonstrate how the fidelity and perceptual quality of the output change when different overall noise levels are applied to a diffusion-based SR model." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "We first analyze the distribution of the residual " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "|y - x|" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " between each pixel pair of the upsampled LR image " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " and its corresponding HR image " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": ", as illustrated in Fig. 2a. The result clearly indicates that " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "|y - x|" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " follows a long-tailed distribution, with over " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " of the data concentrated in the range [0.01, 0.16]. Next, we focus on data within this range to" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17982" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 73, + 115, + 133 + ], + "blocks": [ + { + "bbox": [ + 56, + 73, + 115, + 133 + ], + "lines": [ + { + "bbox": [ + 56, + 73, + 115, + 133 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 115, + 133 + ], + "type": "image", + "image_path": "31017c98198a25fbb9bb37d30f00284812426664277c804673b26dfae4239fa9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 134, + 91, + 140 + ], + "lines": [ + { + "bbox": [ + 82, + 134, + 91, + 140 + ], + "spans": [ + { + "bbox": [ + 82, + 134, + 91, + 140 + ], + "type": "text", + "content": "HR" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 116, + 73, + 176, + 133 + ], + "blocks": [ + { + "bbox": [ + 116, + 73, + 176, + 133 + ], + "lines": [ + { + "bbox": [ + 116, + 73, + 176, + 133 + ], + "spans": [ + { + "bbox": [ + 116, + 73, + 176, + 133 + ], + "type": "image", + "image_path": "fc7b50b181be7ab0b19f639edd0ca07a47eed10cf9a97e729ee7eee1c411a5fb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 134, + 149, + 140 + ], + "lines": [ + { + "bbox": [ + 141, + 134, + 149, + 140 + ], + "spans": [ + { + "bbox": [ + 141, + 134, + 149, + 140 + ], + "type": "text", + "content": "LR" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 177, + 74, + 235, + 133 + ], + "blocks": [ + { + "bbox": [ + 177, + 74, + 235, + 133 + ], + "lines": [ + { + "bbox": [ + 177, + 74, + 235, + 133 + ], + "spans": [ + { + "bbox": [ + 177, + 74, + 235, + 133 + ], + "type": "image", + "image_path": "f6e599077e5024dae8d6d19ee2bbd56835aeb0031c6415ac9b98b961edaa16a2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 134, + 221, + 140 + ], + "lines": [ + { + "bbox": [ + 190, + 134, + 221, + 140 + ], + "spans": [ + { + "bbox": [ + 190, + 134, + 221, + 140 + ], + "type": "text", + "content": "Real Residual" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 144, + 295, + 199 + ], + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 199 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 199 + ], + "type": "text", + "content": "Figure 3. A visualization of the actual residual " + }, + { + "bbox": [ + 55, + 144, + 295, + 199 + ], + "type": "inline_equation", + "content": "|\\pmb{x}^i - \\pmb{y}^i|" + }, + { + "bbox": [ + 55, + 144, + 295, + 199 + ], + "type": "text", + "content": " and the estimated residual " + }, + { + "bbox": [ + 55, + 144, + 295, + 199 + ], + "type": "inline_equation", + "content": "|g(\\pmb{y}^i) - \\pmb{y}^i|" + }, + { + "bbox": [ + 55, + 144, + 295, + 199 + ], + "type": "text", + "content": ". The real residual exhibits high values in edges and texture regions, indicating the high uncertainty. The residual estimated by SR network is close to the real one and therefore can serve as a rough estimation of uncertainty." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 236, + 74, + 294, + 133 + ], + "blocks": [ + { + "bbox": [ + 236, + 74, + 294, + 133 + ], + "lines": [ + { + "bbox": [ + 236, + 74, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 236, + 74, + 294, + 133 + ], + "type": "image", + "image_path": "dce30906886ae081c3205f27c5141a9bc98c09f9c6715891c15809f4a70e3e3a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 134, + 287, + 140 + ], + "lines": [ + { + "bbox": [ + 244, + 134, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 244, + 134, + 287, + 140 + ], + "type": "text", + "content": "Estimated Residual" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 213, + 296, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 296, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 296, + 429 + ], + "type": "text", + "content": "investigate the relationships between the residual " + }, + { + "bbox": [ + 55, + 213, + 296, + 429 + ], + "type": "inline_equation", + "content": "|y - x|" + }, + { + "bbox": [ + 55, + 213, + 296, + 429 + ], + "type": "text", + "content": ", fidelity " + }, + { + "bbox": [ + 55, + 213, + 296, + 429 + ], + "type": "inline_equation", + "content": "|f(y) - x|" + }, + { + "bbox": [ + 55, + 213, + 296, + 429 + ], + "type": "text", + "content": ", perceptual quality " + }, + { + "bbox": [ + 55, + 213, + 296, + 429 + ], + "type": "inline_equation", + "content": "|\\phi(f(y)) - \\phi(x)|" + }, + { + "bbox": [ + 55, + 213, + 296, + 429 + ], + "type": "text", + "content": " of the output of the denoiser " + }, + { + "bbox": [ + 55, + 213, + 296, + 429 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 55, + 213, + 296, + 429 + ], + "type": "text", + "content": " trained under different noise levels " + }, + { + "bbox": [ + 55, + 213, + 296, + 429 + ], + "type": "inline_equation", + "content": "\\sigma_{max}" + }, + { + "bbox": [ + 55, + 213, + 296, + 429 + ], + "type": "text", + "content": " (in a reasonable range). The results are illustrated in Fig. 2b. Due to the blurring and downsampling degradation, the residual " + }, + { + "bbox": [ + 55, + 213, + 296, + 429 + ], + "type": "inline_equation", + "content": "|y - x|" + }, + { + "bbox": [ + 55, + 213, + 296, + 429 + ], + "type": "text", + "content": " tends to be low in flat areas and high in regions characterized by edges and textures. As " + }, + { + "bbox": [ + 55, + 213, + 296, + 429 + ], + "type": "inline_equation", + "content": "|y - x|" + }, + { + "bbox": [ + 55, + 213, + 296, + 429 + ], + "type": "text", + "content": " grows, the perceptual quality gap between " + }, + { + "bbox": [ + 55, + 213, + 296, + 429 + ], + "type": "inline_equation", + "content": "|\\phi(f(y)) - \\phi(x)|_{\\sigma_{max}=2.5}" + }, + { + "bbox": [ + 55, + 213, + 296, + 429 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 213, + 296, + 429 + ], + "type": "inline_equation", + "content": "|\\phi(f(y)) - \\phi(x)|_{\\sigma_{max}=0.5}" + }, + { + "bbox": [ + 55, + 213, + 296, + 429 + ], + "type": "text", + "content": " increases rapidly, while the fidelity gap remains nearly unchanged. This indicates that perceptual quality is more sensitive to the noise level change, with higher noise levels being particularly important in edge and texture areas. These regions lie in the low-density regions of the real image distribution and are affected by the ill-posed nature of superresolution. Applying a low " + }, + { + "bbox": [ + 55, + 213, + 296, + 429 + ], + "type": "inline_equation", + "content": "\\sigma_{max}" + }, + { + "bbox": [ + 55, + 213, + 296, + 429 + ], + "type": "text", + "content": " in such areas leads to unreliable score estimation, resulting in perceptually poor, over-smoothed outputs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 430, + 296, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 296, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 296, + 562 + ], + "type": "text", + "content": "The experimental results support our hypothesis that different regions of an image respond differently to noise and require distinct handling based on their content. Flat regions typically closely resemble the corresponding ground truth, requiring only a slight amount of noise; therefore, we consider them as existing at small timesteps " + }, + { + "bbox": [ + 55, + 430, + 296, + 562 + ], + "type": "inline_equation", + "content": "(t\\to 0)" + }, + { + "bbox": [ + 55, + 430, + 296, + 562 + ], + "type": "text", + "content": ". Based on this idea, we propose to reduce noise levels applied in flat regions, and the results shown by the red line in Fig. 2b preliminarily verify the effectiveness. In the following subsections, we will dually discuss how to determine the noise level for different types of areas in the LR image." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 574, + 188, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 188, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 188, + 586 + ], + "type": "text", + "content": "3.2. Uncertainty Estimation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Due to the ill-posed nature of the image super-resolution task, perfectly reconstructing HR images from degraded LR inputs is non-trivial. The high-frequency information in edge and texture areas is severely corrupted during the degradation, exhibiting greater variance compared to flat areas and making accurate prediction more challenging. Several works [16, 25] explore this variance from an uncertainty-based perspective [15]. Given an HR image " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}^i \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " and its corresponding LR image " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\boldsymbol{y}^i \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": ", the uncertainty " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\psi^i" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " of the SR estimate " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "g(\\boldsymbol{y}^i)" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " is related to its resid" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 72, + 365, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 365, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 365, + 83 + ], + "type": "text", + "content": "ual with " + }, + { + "bbox": [ + 314, + 72, + 365, + 83 + ], + "type": "inline_equation", + "content": "x^i" + }, + { + "bbox": [ + 314, + 72, + 365, + 83 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 380, + 92, + 555, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 92, + 555, + 107 + ], + "spans": [ + { + "bbox": [ + 380, + 92, + 555, + 107 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} ^ {i} = g \\left(\\boldsymbol {y} ^ {i}\\right) + \\epsilon \\psi \\left(g \\left(\\boldsymbol {y} ^ {i}\\right)\\right), \\tag {3}", + "image_path": "d7645c94f2196e9c8ec9c3dc8039978d1624275365a672cd15ca98aa5c4429f2.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 114, + 555, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 114, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 555, + 198 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 114, + 555, + 198 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 114, + 555, + 198 + ], + "type": "text", + "content": " represents a standard Laplace or Gaussian distribution when " + }, + { + "bbox": [ + 313, + 114, + 555, + 198 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 313, + 114, + 555, + 198 + ], + "type": "text", + "content": " is regularized by " + }, + { + "bbox": [ + 313, + 114, + 555, + 198 + ], + "type": "inline_equation", + "content": "L1" + }, + { + "bbox": [ + 313, + 114, + 555, + 198 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 114, + 555, + 198 + ], + "type": "inline_equation", + "content": "L2" + }, + { + "bbox": [ + 313, + 114, + 555, + 198 + ], + "type": "text", + "content": " loss function, respectively. If the reconstruction error " + }, + { + "bbox": [ + 313, + 114, + 555, + 198 + ], + "type": "inline_equation", + "content": "|\\pmb{x}^i - g(\\pmb{y}^i)|" + }, + { + "bbox": [ + 313, + 114, + 555, + 198 + ], + "type": "text", + "content": " is larger than " + }, + { + "bbox": [ + 313, + 114, + 555, + 198 + ], + "type": "inline_equation", + "content": "|\\pmb{x}^j - g(\\pmb{y}^j)|" + }, + { + "bbox": [ + 313, + 114, + 555, + 198 + ], + "type": "text", + "content": ", then it is more likely that " + }, + { + "bbox": [ + 313, + 114, + 555, + 198 + ], + "type": "inline_equation", + "content": "g(\\pmb{y}^i)" + }, + { + "bbox": [ + 313, + 114, + 555, + 198 + ], + "type": "text", + "content": " has a higher uncertainty " + }, + { + "bbox": [ + 313, + 114, + 555, + 198 + ], + "type": "inline_equation", + "content": "\\psi(g(\\pmb{y}^i))" + }, + { + "bbox": [ + 313, + 114, + 555, + 198 + ], + "type": "text", + "content": ", compared to the uncertainty " + }, + { + "bbox": [ + 313, + 114, + 555, + 198 + ], + "type": "inline_equation", + "content": "\\psi(g(\\pmb{y}^j))" + }, + { + "bbox": [ + 313, + 114, + 555, + 198 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 114, + 555, + 198 + ], + "type": "inline_equation", + "content": "g(\\pmb{y}^j)" + }, + { + "bbox": [ + 313, + 114, + 555, + 198 + ], + "type": "text", + "content": ". Similarly, we can associate the uncertainty of " + }, + { + "bbox": [ + 313, + 114, + 555, + 198 + ], + "type": "inline_equation", + "content": "\\pmb{y}^i" + }, + { + "bbox": [ + 313, + 114, + 555, + 198 + ], + "type": "text", + "content": " with the residual " + }, + { + "bbox": [ + 313, + 114, + 555, + 198 + ], + "type": "inline_equation", + "content": "|\\pmb{x}^i - \\pmb{y}^i|" + }, + { + "bbox": [ + 313, + 114, + 555, + 198 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 392, + 205, + 555, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 205, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 392, + 205, + 555, + 220 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} ^ {i} = \\boldsymbol {y} ^ {i} + \\hat {\\epsilon} \\psi (\\boldsymbol {y} ^ {i}), \\tag {4}", + "image_path": "49332fc76e42cf7a3d914ab7bd97358f5a6a6450b0e5a29797623896870df809.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 225, + 555, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 225, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 555, + 321 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 225, + 555, + 321 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}" + }, + { + "bbox": [ + 313, + 225, + 555, + 321 + ], + "type": "text", + "content": " represents an unknown distribution that depends on the degradation pattern. If " + }, + { + "bbox": [ + 313, + 225, + 555, + 321 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 313, + 225, + 555, + 321 + ], + "type": "text", + "content": " is well-trained, we can assume that " + }, + { + "bbox": [ + 313, + 225, + 555, + 321 + ], + "type": "inline_equation", + "content": "g(\\pmb{y}^i)" + }, + { + "bbox": [ + 313, + 225, + 555, + 321 + ], + "type": "text", + "content": " closely approximates " + }, + { + "bbox": [ + 313, + 225, + 555, + 321 + ], + "type": "inline_equation", + "content": "\\pmb{x}^i" + }, + { + "bbox": [ + 313, + 225, + 555, + 321 + ], + "type": "text", + "content": ", implying that " + }, + { + "bbox": [ + 313, + 225, + 555, + 321 + ], + "type": "inline_equation", + "content": "|g(\\pmb{y}^i) - \\pmb{y}^i|" + }, + { + "bbox": [ + 313, + 225, + 555, + 321 + ], + "type": "text", + "content": " is similar to " + }, + { + "bbox": [ + 313, + 225, + 555, + 321 + ], + "type": "inline_equation", + "content": "|\\pmb{x}^i - \\pmb{y}^i|" + }, + { + "bbox": [ + 313, + 225, + 555, + 321 + ], + "type": "text", + "content": ". As illustrated in Fig. 3, the residual estimated by " + }, + { + "bbox": [ + 313, + 225, + 555, + 321 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 313, + 225, + 555, + 321 + ], + "type": "text", + "content": " is roughly similar to the real residual. Therefore, we propose to leverage the residual " + }, + { + "bbox": [ + 313, + 225, + 555, + 321 + ], + "type": "inline_equation", + "content": "|g(\\pmb{y}^i) - \\pmb{y}^i|" + }, + { + "bbox": [ + 313, + 225, + 555, + 321 + ], + "type": "text", + "content": " as the estimate of the uncertainty of " + }, + { + "bbox": [ + 313, + 225, + 555, + 321 + ], + "type": "inline_equation", + "content": "\\pmb{y}^i" + }, + { + "bbox": [ + 313, + 225, + 555, + 321 + ], + "type": "text", + "content": ". Specifically, we define the uncertainty estimate as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 382, + 327, + 554, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 327, + 554, + 352 + ], + "spans": [ + { + "bbox": [ + 382, + 327, + 554, + 352 + ], + "type": "interline_equation", + "content": "\\psi_ {e s t} (\\boldsymbol {y}) = \\frac {1}{2} | g (\\boldsymbol {y}) - \\boldsymbol {y} |. \\tag {5}", + "image_path": "313f4102ea2dac367a59cae902ba7b3e1747b6c48ce210f47bcd378a69a12775.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 357, + 554, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 554, + 381 + ], + "type": "text", + "content": "In the next subsection, we will take this uncertainty estimate as the criterion to adjust the noise intensity." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 388, + 508, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 508, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 508, + 401 + ], + "type": "text", + "content": "3.3. Uncertainty-guided Noise Weighting" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 405, + 555, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 555, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 555, + 693 + ], + "type": "text", + "content": "Based on the discussion in previous sections, we propose to replace the commonly used isotropic Gaussian noise with anisotropic noise that adapts to the image content. Inspired by [25], which deals with areas with higher uncertainty by assigning larger weights to the loss function to impose stronger constraints, we suggest adjusting the noise intensity based on the uncertainty estimate across different regions. Specifically, after obtaining the uncertainty estimate as outlined in Sec. 3.2, we compute the weighting coefficient " + }, + { + "bbox": [ + 313, + 405, + 555, + 693 + ], + "type": "inline_equation", + "content": "w_{u}" + }, + { + "bbox": [ + 313, + 405, + 555, + 693 + ], + "type": "text", + "content": " and apply it to modulate noise levels for different areas in the diffusion process. We refer to this strategy as Uncertainty-guided Noise Weighting (UNW). For " + }, + { + "bbox": [ + 313, + 405, + 555, + 693 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 405, + 555, + 693 + ], + "type": "text", + "content": " with lower uncertainty (e.g. in flat areas), it is more likely that " + }, + { + "bbox": [ + 313, + 405, + 555, + 693 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 405, + 555, + 693 + ], + "type": "text", + "content": " resides in high-density regions of the HR data distribution. In such cases, a lower noise level is sufficient to achieve perceptual quality comparable to that of a higher noise level, while also preserving more LR details, as discussed in Sec. 3.1. Conversely, if " + }, + { + "bbox": [ + 313, + 405, + 555, + 693 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 405, + 555, + 693 + ], + "type": "text", + "content": " exhibits high uncertainty (e.g., in edge or texture areas), a higher noise level is requisite to account for the significant distribution disparity and to provide greater chances of reconstructing photorealistic details. Therefore, we model the noise weighting coefficient " + }, + { + "bbox": [ + 313, + 405, + 555, + 693 + ], + "type": "inline_equation", + "content": "w_{u}" + }, + { + "bbox": [ + 313, + 405, + 555, + 693 + ], + "type": "text", + "content": " as a monotonically increasing function with respect to the uncertainty:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 387, + 700, + 555, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 700, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 387, + 700, + 555, + 714 + ], + "type": "interline_equation", + "content": "w _ {u} (\\boldsymbol {y}) := u \\left(\\psi_ {\\text {e s t}} (\\boldsymbol {y})\\right), \\tag {6}", + "image_path": "50ad02443b63fb00be02b29c9fb31ced009c94ac417cba222417864f249094db.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17983" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 123, + 74, + 492, + 220 + ], + "blocks": [ + { + "bbox": [ + 123, + 74, + 492, + 220 + ], + "lines": [ + { + "bbox": [ + 123, + 74, + 492, + 220 + ], + "spans": [ + { + "bbox": [ + 123, + 74, + 492, + 220 + ], + "type": "image", + "image_path": "174cb11b74ab35961d8efc4ffc7ffad388aa3c9e86fbdac009db44ad3d8b3173.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 223, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 54, + 223, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 54, + 223, + 555, + 258 + ], + "type": "text", + "content": "Figure 4. The overall pipeline of the proposed UPSR model. An auxiliary SR network is first employed to estimate the uncertainty of the input " + }, + { + "bbox": [ + 54, + 223, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_0" + }, + { + "bbox": [ + 54, + 223, + 555, + 258 + ], + "type": "text", + "content": ". Then the weighting coefficient " + }, + { + "bbox": [ + 54, + 223, + 555, + 258 + ], + "type": "inline_equation", + "content": "w_{u}" + }, + { + "bbox": [ + 54, + 223, + 555, + 258 + ], + "type": "text", + "content": " computed based on the uncertainty " + }, + { + "bbox": [ + 54, + 223, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\psi_{est}(\\mathbf{y}_0)" + }, + { + "bbox": [ + 54, + 223, + 555, + 258 + ], + "type": "text", + "content": " are applied to adjust the noise level in different regions. Meanwhile, both the SR estimate " + }, + { + "bbox": [ + 54, + 223, + 555, + 258 + ], + "type": "inline_equation", + "content": "g(\\mathbf{y}_0)" + }, + { + "bbox": [ + 54, + 223, + 555, + 258 + ], + "type": "text", + "content": " and LR input " + }, + { + "bbox": [ + 54, + 223, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_0" + }, + { + "bbox": [ + 54, + 223, + 555, + 258 + ], + "type": "text", + "content": " are concatenated as the conditional information for the denoiser " + }, + { + "bbox": [ + 54, + 223, + 555, + 258 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 54, + 223, + 555, + 258 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 269, + 294, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 269, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 54, + 269, + 294, + 341 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 269, + 294, + 341 + ], + "type": "inline_equation", + "content": "w_{u}(\\pmb{y}) \\in \\mathbb{R}^{n \\times n}" + }, + { + "bbox": [ + 54, + 269, + 294, + 341 + ], + "type": "text", + "content": " represents the diagonal weight matrix used to lower the noise level in low-uncertainty areas while maintaining a higher noise level in high-uncertainty regions, creating a more specialized and adaptive diffusion pipeline for SR. Details of the implementation of " + }, + { + "bbox": [ + 54, + 269, + 294, + 341 + ], + "type": "inline_equation", + "content": "u(\\cdot)" + }, + { + "bbox": [ + 54, + 269, + 294, + 341 + ], + "type": "text", + "content": " are presented in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "type": "text", + "content": "Built upon the UNW technique, we introduce a new pipeline for diffusion-based SR, termed Uncertainty-guided Perturbation for SR (UPSR), as illustrated in Fig. 4. Given an LR image " + }, + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "type": "inline_equation", + "content": "\\pmb{y}_0" + }, + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "type": "text", + "content": ", we first obtain its SR estimate " + }, + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "type": "inline_equation", + "content": "g(\\pmb{y}_0)" + }, + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "type": "text", + "content": " through an auxiliary SR network " + }, + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "type": "text", + "content": ". Then we estimate the uncertainty of " + }, + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "type": "inline_equation", + "content": "\\pmb{y}_0" + }, + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "type": "inline_equation", + "content": "\\psi_{est}(\\pmb{y}_0) = \\frac{1}{2} |g(\\pmb{y}_0) - \\pmb{y}_0|" + }, + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "type": "text", + "content": " and obtain the weighting coefficient as " + }, + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "type": "inline_equation", + "content": "w_u(\\pmb{y}_0) = u(\\psi_{est}(\\pmb{y}_0))" + }, + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "type": "text", + "content": ". Based on the weighting coefficient, we rewrite the forward transition distribution " + }, + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "type": "inline_equation", + "content": "q(\\pmb{x}_t \\mid \\pmb{x}_{t-1}, \\pmb{x}_0, \\pmb{y}_0)" + }, + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "type": "text", + "content": " in Eq. 1 as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 457, + 295, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 457, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 70, + 457, + 295, + 472 + ], + "type": "interline_equation", + "content": "\\left. \\mathcal {N} \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {t - 1} + \\alpha_ {t} \\left(\\boldsymbol {y} _ {0} - \\boldsymbol {x} _ {0}\\right), \\kappa^ {2} w _ {u} \\left(\\boldsymbol {y} _ {0}\\right) ^ {2} \\alpha_ {t} \\boldsymbol {I}\\right), \\right. \\tag {7}", + "image_path": "9fe904eb77bd4d24d7e6fe78211a2f596737dfc13d590cf35ca8748da1be8664.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 480, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 295, + 506 + ], + "type": "text", + "content": "and derive the corresponding backward transition distribution " + }, + { + "bbox": [ + 55, + 480, + 295, + 506 + ], + "type": "inline_equation", + "content": "q(\\pmb{x}_{t-1} \\mid \\pmb{x}_t, \\pmb{x}_0, \\pmb{y}_0)" + }, + { + "bbox": [ + 55, + 480, + 295, + 506 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 514, + 295, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 514, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 61, + 514, + 295, + 540 + ], + "type": "interline_equation", + "content": "\\mathcal {N} \\left(\\boldsymbol {x} _ {t - 1} \\mid \\frac {\\eta_ {t - 1}}{\\eta_ {t}} \\boldsymbol {x} _ {t} + \\frac {\\alpha_ {t}}{\\eta_ {t}} \\boldsymbol {x} _ {0}, \\kappa^ {2} w _ {u} (\\boldsymbol {y} _ {0}) ^ {2} \\frac {\\eta_ {t - 1}}{\\eta_ {t}} \\alpha_ {t} \\boldsymbol {I}\\right), \\tag {8}", + "image_path": "851b6cda622182867b1466c6a85432d4aec00133c5dc505109e86d5316a8464f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "text", + "content": "where the difference is that we leverage the weighting coefficient " + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "inline_equation", + "content": "w_{u}(\\pmb{y}_{0})" + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "text", + "content": " to control the noise level. Since " + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "inline_equation", + "content": "g(\\pmb{y}_0)" + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "text", + "content": " is a better estimate of " + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "text", + "content": " compared to " + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "inline_equation", + "content": "y_{0}" + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "text", + "content": ", we combine " + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "inline_equation", + "content": "g(\\pmb{y}_0)" + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "inline_equation", + "content": "y_{0}" + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "text", + "content": " to provide more accurate conditional information for the denoiser. Details of the derivations of Eq. 7, 8 and the training pipeline are presented in the supplementary materials. Following the previous diffusion-based methods [8, 44], our training objective for the denoiser to predict the target " + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "text", + "content": " combines both pixel distance " + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "inline_equation", + "content": "||\\cdot||_2^2" + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "text", + "content": " and LPIPS criterion " + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "inline_equation", + "content": "L_{per}" + }, + { + "bbox": [ + 55, + 548, + 295, + 668 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 70, + 675, + 295, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 675, + 295, + 717 + ], + "spans": [ + { + "bbox": [ + 70, + 675, + 295, + 717 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} (\\theta) = \\sum_ {t} [ \\| f _ {\\theta} (\\boldsymbol {x} _ {t}, \\boldsymbol {y} _ {0}, g (\\boldsymbol {y} _ {0}), t) - \\boldsymbol {x} _ {0} \\| _ {2} ^ {2} \\tag {9} \\\\ + \\lambda L _ {p e r} \\left(f _ {\\theta} \\left(\\boldsymbol {x} _ {t}, \\boldsymbol {y} _ {0}, g (\\boldsymbol {y} _ {0}), t\\right), \\boldsymbol {x} _ {0}\\right) ], \\\\ \\end{array}", + "image_path": "58d18c911ed4c9f9b0280b49a14cd6984fd70d13244003f7ed446331cefe8883.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 269, + 555, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 555, + 329 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 269, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 269, + 555, + 329 + ], + "type": "text", + "content": " is a hyperparameter to control the trade-off between fidelity and perceptual quality. Optimizing the denoiser " + }, + { + "bbox": [ + 313, + 269, + 555, + 329 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 313, + 269, + 555, + 329 + ], + "type": "text", + "content": " through the mixed objective function facilitates fewer diffusion steps while achieving better photo-realistic results in various benchmark real-world datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 336, + 503, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 503, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 503, + 348 + ], + "type": "text", + "content": "3.4. Network Architecture Modification" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 354, + 554, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 554, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 554, + 389 + ], + "type": "text", + "content": "Besides altering the noise injection scheme and sampling procedure, we also make several modifications to the network architecture." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 390, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 555, + 641 + ], + "type": "text", + "content": "We rethink the necessity of leveraging latent space for image super-resolution with only a few diffusion steps. [26] and [44] utilize VQGAN [5] to transfer the diffusion process from pixel space to latent space, for the sake of reducing spatial dimensionality to improve efficiency. However, as the number of sampling steps decreases, the performance-cost ratio of VQGAN gradually decreases due to its own huge computing resource consumption. Moreover, the application of certain perceptual criteria (e.g., LPIPS [48]) requires pixel-level computation, leading to additional decoding costs during the training process. To mitigate its impact, we propose to replace the VQGAN encoder and decoder with PixelUnshuffle operation [28] and a simple nearest neighbor upsampling module, respectively. This approach enables compressing and expanding the spatial dimension with minimal computational cost. Therefore, we can perform the diffusion process directly in pixel space while maintaining a complexity comparable to that in the latent space. In this vein, we significantly reduce the total model size and training overhead without sacrificing performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 652, + 395, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 395, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 395, + 666 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 672, + 435, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 435, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 435, + 685 + ], + "type": "text", + "content": "4.1. Experimental Setups" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "For data preparation, we apply randomly cropped " + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": " patches from ImageNet [3] as HR training data following" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17984" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 103, + 545, + 173 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Ablation study on effects of the proposed components, including the SR condition and uncertainty-guided noise weighting (UNW). The best results are highlighted in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 103, + 545, + 173 + ], + "lines": [ + { + "bbox": [ + 67, + 103, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 67, + 103, + 545, + 173 + ], + "type": "table", + "html": "
UNWSR cond.RealSRRealSet
PSNR↑CLIPQA↑MUSIQ↑MANIQA↑NIQE↓CLIPQA↑MUSIQ↑MANIQA↑NIQE↓
26.180.544762.9510.35964.490.614164.3600.37184.42
✓26.120.576064.5120.37174.180.634064.2800.38364.22
✓✓26.440.601064.5410.38184.020.638963.4980.39314.24
", + "image_path": "5bc0660e2397af8a6b7694c8a72b380f2dfc24c13e5080d60fc45994d572ebee.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 186, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 296, + 392 + ], + "type": "text", + "content": "[26] and [44]. Then the degradation pipeline of RealESR-GAN [37] are adopted to generate degraded " + }, + { + "bbox": [ + 56, + 186, + 296, + 392 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 56, + 186, + 296, + 392 + ], + "type": "text", + "content": " LR inputs. For network architecture, we employ the commonly-used U-Net denoiser " + }, + { + "bbox": [ + 56, + 186, + 296, + 392 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 56, + 186, + 296, + 392 + ], + "type": "text", + "content": " with few adjustments to fit our refined diffusion pipeline. For " + }, + { + "bbox": [ + 56, + 186, + 296, + 392 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 56, + 186, + 296, + 392 + ], + "type": "text", + "content": ", we employ the lightweight version of [47] and pretrain it following the real SR pipeline to produce SR predictions and uncertainty estimates. During the diffusion model training process, we freeze the pretrained " + }, + { + "bbox": [ + 56, + 186, + 296, + 392 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 56, + 186, + 296, + 392 + ], + "type": "text", + "content": " and optimize " + }, + { + "bbox": [ + 56, + 186, + 296, + 392 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 56, + 186, + 296, + 392 + ], + "type": "text", + "content": " for 200k iterations with a batchsize of 32. To quantitatively validate the efficacy of the proposed method, we utilize three full-reference metrics: PSNR, SSIM and LPIPS [48]; and four non-reference metrics CLIPIQA [34], MUSIQ [14], MANIQA [42], and NIQE [22] to evaluate the performance. Among these metrics, PSNR and SSIM reflect the fidelity of the generated super-resolution results, while other metrics assess the perceptual quality of the outputs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 400, + 119, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 119, + 414 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 119, + 414 + ], + "type": "text", + "content": "4.2. Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 418, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 456 + ], + "type": "text", + "content": "In this subsection, we conduct ablation studies on the effectiveness of several key components and analyze the performance-cost trade-off of the proposed method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 474, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 296, + 714 + ], + "type": "text", + "content": "Effectiveness of UNW and SR conditioning. In order to show the efficacy of the proposed uncertainty-guided noise weighting scheme and SR conditioning, we use the noise schedule and sampling process of ResShift [44] as the baseline. We first apply the uncertainty-guided noise weighting scheme to adjust the noise intensity across different regions based on the uncertainty estimate. The weighting scheme yields better outcomes in terms of perceptual quality, with enhancements of 0.0355 in CLIPIQA and 1.512 in MUSIQ. We attribute this improvement to the reduced noise intensities applied to flat areas, which allow these clearer regions to provide additional information that aids in more accurately recovering the remaining degraded regions. This validates our idea of reducing noise in flat regions does not degrade perceptual quality, but rather enhances it. Next, we incorporate the SR prediction " + }, + { + "bbox": [ + 55, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "g(\\pmb{y}_0)" + }, + { + "bbox": [ + 55, + 474, + 296, + 714 + ], + "type": "text", + "content": " and LR input " + }, + { + "bbox": [ + 55, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{y}_0" + }, + { + "bbox": [ + 55, + 474, + 296, + 714 + ], + "type": "text", + "content": " as the conditional information entered into the denoiser " + }, + { + "bbox": [ + 55, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 55, + 474, + 296, + 714 + ], + "type": "text", + "content": ". The additional SR prediction offers more precise supplementary information compared to relying solely on the degraded LR input, resulting in a significant improvement in both fi" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 265, + 553, + 327 + ], + "blocks": [ + { + "bbox": [ + 313, + 188, + 555, + 256 + ], + "lines": [ + { + "bbox": [ + 313, + 188, + 555, + 256 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 555, + 256 + ], + "type": "text", + "content": "Table 2. Model size and computational efficiency comparisons between the proposed UPSR and other diffusion-based methods. Gray numbers denote the parameter counts for auxiliary network, i.e., VQGAN in ResShift and " + }, + { + "bbox": [ + 313, + 188, + 555, + 256 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 313, + 188, + 555, + 256 + ], + "type": "text", + "content": " in our work. We test the runtimes on " + }, + { + "bbox": [ + 313, + 188, + 555, + 256 + ], + "type": "inline_equation", + "content": "32 \\times 3 \\times 64 \\times 64" + }, + { + "bbox": [ + 313, + 188, + 555, + 256 + ], + "type": "text", + "content": " LR input using single RTX4090 GPU and present several results evaluated on the RealSR dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 265, + 553, + 327 + ], + "lines": [ + { + "bbox": [ + 317, + 265, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 317, + 265, + 553, + 327 + ], + "type": "table", + "html": "
ModelParams (M)Runtimes (s)MUSIQMANIQA
LDM-15113.60+55.321.5948.6980.2655
ResShift-15118.59+55.321.9857.7690.3691
ResShift-4118.59+55.321.0055.1890.3337
UPSR-5119.42+2.501.1264.5410.3818
", + "image_path": "1e8ea203eb775c33ebcb624b991e02cf40dd4e32bccbfa2d38f0ff1bbb0afa88.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 349, + 365, + 518, + 407 + ], + "blocks": [ + { + "bbox": [ + 313, + 336, + 555, + 359 + ], + "lines": [ + { + "bbox": [ + 313, + 336, + 555, + 359 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 555, + 359 + ], + "type": "text", + "content": "Table 3. Training overhead comparison between ResShift and UPSR evaluated under a batchsize of 16 per GPU." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 349, + 365, + 518, + 407 + ], + "lines": [ + { + "bbox": [ + 349, + 365, + 518, + 407 + ], + "spans": [ + { + "bbox": [ + 349, + 365, + 518, + 407 + ], + "type": "table", + "html": "
ModelTraining SpeedMemory Footprint
ResShift1.20 s/iter24.1 G
UPSR0.45 s/iter14.9 G
", + "image_path": "1c8609f0990d85f0e69498a1e1dcfe5205787dae23867195198ce58117b26f6e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 419, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 555, + 479 + ], + "type": "text", + "content": "delity and perceptual quality. Specifically, this more robust pipeline further improves PSNR by 0.34 dB and CLIPIQA by 0.0247 in the RealSR dataset. These ablation studies verify the significance of leveraging LR information during diffusion-based SR process." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": "Model Size and Training Overhead Comparison. In section 3.4, we discussed replacing VQGAN in diffusion-based SR model with simpler downsampling and upsampling modules to improve efficiency. To validate the efficacy of the modified architecture, we first evaluate the model size and computational cost of UPSR compared to state-of-the-art diffusion-based methods, as shown in Tab. 2. With the modified architecture and pipeline, UPSR-5 reduces the overall model size by " + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": " and achieves better perceptual quality with an inference speed comparable to that of ResShift-4. In contrast, ResShift-4 achieves the acceleration at the expense of performance. We then assess the training overhead of different methods, as shown in Tab. 3. Removing VQGAN results in substantial reductions in training overhead, increasing training speed by " + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "inline_equation", + "content": "167\\%" + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": " and saving GPU memory footprint by " + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "inline_equation", + "content": "38\\%" + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": ". These results validate that UPSR strikes a better trade-off between performance and efficiency." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17985" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 109, + 550, + 322 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 4. Quantitative results of different methods on one synthetic dataset ImageNet-Test and two real-world datasets RealSR and RealSet. The best and second best results are marked in red and blue. All results of the previous methods are evaluated using the released inference codes and pretrained weights." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 109, + 550, + 322 + ], + "lines": [ + { + "bbox": [ + 61, + 109, + 550, + 322 + ], + "spans": [ + { + "bbox": [ + 61, + 109, + 550, + 322 + ], + "type": "table", + "html": "
DatasetsMetricsGAN-based MethodsDiffusion-based Methods
ESRGANRealSR-JPEGBSRGANRealESRGANSwinIRDASRLDM-15ResShift-15ResShift-4UPSR-5
ImageNet -TestPSNR↑20.6723.1124.4224.0423.9924.7524.8524.9425.0223.77
SSIM↑0.44850.59120.65850.66490.66660.67490.66820.67380.68300.6296
LPIPS↓0.48510.32630.25850.25390.23760.24980.26850.23710.20750.2456
CLIPQA↑0.45120.53660.58100.52410.56390.53620.50950.58600.60030.6328
MUSIQ↑43.61546.98154.69652.60953.78948.33746.63953.18252.01959.227
MANIQA↑0.32120.30650.38650.36890.38820.32920.33050.41910.38850.4591
NIQE↓8.335.966.086.075.895.867.216.887.345.24
RealSRPSNR↑27.5727.3426.5125.8326.4327.1927.1826.8025.7726.44
SSIM↑0.77420.76050.77460.77260.78610.78610.78530.76740.74390.7589
LPIPS↓0.41520.39620.26850.27390.25150.31130.30210.34110.34910.2871
CLIPQA↑0.23620.36130.54390.49230.46550.36280.37480.57090.56460.6010
MUSIQ↑29.03736.06963.58759.84959.63545.81848.69857.76955.18964.541
MANIQA↑0.20710.17830.37020.36940.34360.26630.26550.36910.33370.3828
NIQE↓7.736.954.654.684.685.986.225.966.934.02
RealSetCLIPQA↑0.37390.52820.61600.60810.57780.49660.43130.63090.61880.6392
MUSIQ↑42.36650.53965.58364.12563.81755.70848.60259.31958.51663.519
MANIQA↑0.31000.29270.38880.39490.38180.31340.26930.39160.35260.3931
NIQE↓4.934.814.584.384.404.726.475.966.464.23
", + "image_path": "dbe3b36664d985583725261e14c91ba74bd420ab98667312fdbf165f70ca540b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 333, + 285, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 333, + 285, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 285, + 346 + ], + "type": "text", + "content": "4.3. Comparisons with State-of-the-Art Methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 351, + 296, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 296, + 459 + ], + "type": "text", + "content": "We select ImageNet-Test [3, 44] that contains 3,000 images as major dataset for synthetic image super-resolution evaluation. Furthermore, two commonly used real-world datasets, RealSR [1] and RealSet65 [44], are adopted to evaluate the generalizability in real-world scenarios. We make comparisons with several GAN-based methods ESRGAN [36], RealSR-JPEG [10], BSRGAN [46], RealESRGAN [37], SwinIR [19], DASR [20], and two diffusion-based methods LDM [26] and ResShift [44]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 460, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 651 + ], + "type": "text", + "content": "The quantitative results on three benchmark datasets are shown in Tab. 4. We present UPSR with five sampling steps in order to trade off the super-resolution performance and computational consumption. The proposed UPSR achieves better CLIPIQA, MUSIQ, MANIQA and NIQE compared to ResShift, indicating the significant improvement in perceptual quality. Specifically, while reducing " + }, + { + "bbox": [ + 55, + 460, + 296, + 651 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 55, + 460, + 296, + 651 + ], + "type": "text", + "content": " of the overall model size, UPSR-5 still outperforms ResShift-4 by 7.21 and 2.10 in terms of MUSIQ and NIQE in ImageNet-Test dataset. For real-world datasets, our method also yields impressive results. UPSR-5 consistently outperforms ResShift-4 across both the RealSR and RealSet datasets, demonstrating the efficacy of UPSR in addressing unknown degradation. Furthermore, UPSR-5 achieves the best NIQE metric, while recent diffusion-based methods show poorer NIQE performance compared to most GAN-based methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 659, + 157, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 157, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 157, + 672 + ], + "type": "text", + "content": "4.4. Visual Examples" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 712 + ], + "type": "text", + "content": "In this section, we present several visual examples to illustrate the proposed UNW scheme and the comparisons between different methods." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 325, + 335, + 548, + 452 + ], + "blocks": [ + { + "bbox": [ + 325, + 335, + 548, + 452 + ], + "lines": [ + { + "bbox": [ + 325, + 335, + 548, + 452 + ], + "spans": [ + { + "bbox": [ + 325, + 335, + 548, + 452 + ], + "type": "image", + "image_path": "069bfc6d62c3261af3c570caa9f0d953953290d04beff652e8a8567ef5ce3f3d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 456, + 555, + 534 + ], + "lines": [ + { + "bbox": [ + 313, + 456, + 555, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 555, + 534 + ], + "type": "text", + "content": "Figure 5. Visual examples of the proposed UNW strategy. Based on the uncertainty estimate (illustrated as the heatmap), the noise level in most flat areas is reduced to preserve more details for better SR results. Meanwhile, noise in edge areas (e.g., in image (a)) and severely degraded parts (e.g., in image (b)) are maintained relatively heavy to ensure reliable score estimation to produce visually pleasing results." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 548, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 555, + 715 + ], + "type": "text", + "content": "Visualization of Weighted Noise. In Fig. 5 we provide visual examples of isotropic and anisotropic noises applied in different images. The proposed UNW strategy enables adaptive perturbation across different areas, leading to the anisotropic diffusion process. The reduced noise level in flat areas, particularly in clear backgrounds, help retain more details in the initial state without sacrificing generative capability. Additionally, these retained details provide more information to reconstruct missing details in surrounding regions, thereby enhancing overall performance. Conversely, in edge regions, the noise level is maintained at the predefined " + }, + { + "bbox": [ + 313, + 548, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\sigma_{max}" + }, + { + "bbox": [ + 313, + 548, + 555, + 715 + ], + "type": "text", + "content": " due to their high uncertainty, ensuring sufficient perturbation to prevent the model from generating over-smoothed results." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "17986" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 105, + 70, + 224, + 188 + ], + "blocks": [ + { + "bbox": [ + 105, + 70, + 224, + 188 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 224, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 224, + 188 + ], + "type": "image", + "image_path": "1f2aaf03c8afccc2c6af213a7e3a19f6c383a311f5fcf00effa1f06d32ab353d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 190, + 208, + 199 + ], + "lines": [ + { + "bbox": [ + 127, + 190, + 208, + 199 + ], + "spans": [ + { + "bbox": [ + 127, + 190, + 208, + 199 + ], + "type": "text", + "content": "ImageNet val 0079" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 225, + 71, + 293, + 122 + ], + "blocks": [ + { + "bbox": [ + 225, + 71, + 293, + 122 + ], + "lines": [ + { + "bbox": [ + 225, + 71, + 293, + 122 + ], + "spans": [ + { + "bbox": [ + 225, + 71, + 293, + 122 + ], + "type": "image", + "image_path": "8612017f43e00ec52fb47b1d149174ca1496576723234bb7d75d113186dd1cef.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 125, + 279, + 133 + ], + "lines": [ + { + "bbox": [ + 239, + 125, + 279, + 133 + ], + "spans": [ + { + "bbox": [ + 239, + 125, + 279, + 133 + ], + "type": "text", + "content": "BSRGAN" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 294, + 71, + 364, + 122 + ], + "blocks": [ + { + "bbox": [ + 294, + 71, + 364, + 122 + ], + "lines": [ + { + "bbox": [ + 294, + 71, + 364, + 122 + ], + "spans": [ + { + "bbox": [ + 294, + 71, + 364, + 122 + ], + "type": "image", + "image_path": "a597c2cae3ae5f90ae8cd28c0081d898454a03fab3cef5792afad59dbab1abce.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 301, + 125, + 358, + 133 + ], + "lines": [ + { + "bbox": [ + 301, + 125, + 358, + 133 + ], + "spans": [ + { + "bbox": [ + 301, + 125, + 358, + 133 + ], + "type": "text", + "content": "RealESRGAN" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 365, + 71, + 433, + 122 + ], + "blocks": [ + { + "bbox": [ + 365, + 71, + 433, + 122 + ], + "lines": [ + { + "bbox": [ + 365, + 71, + 433, + 122 + ], + "spans": [ + { + "bbox": [ + 365, + 71, + 433, + 122 + ], + "type": "image", + "image_path": "ab5c66c7677b34d74b7f0dcc945b2385af73bf01f03b2fb73fa0e5af17493a37.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 125, + 414, + 133 + ], + "lines": [ + { + "bbox": [ + 384, + 125, + 414, + 133 + ], + "spans": [ + { + "bbox": [ + 384, + 125, + 414, + 133 + ], + "type": "text", + "content": "SwinIR" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 434, + 71, + 504, + 122 + ], + "blocks": [ + { + "bbox": [ + 434, + 71, + 504, + 122 + ], + "lines": [ + { + "bbox": [ + 434, + 71, + 504, + 122 + ], + "spans": [ + { + "bbox": [ + 434, + 71, + 504, + 122 + ], + "type": "image", + "image_path": "8d9fbaae7598fe3023ec8050ddabaabd64b0dccf995ef09e3afaf42bfc0d1850.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 125, + 482, + 133 + ], + "lines": [ + { + "bbox": [ + 456, + 125, + 482, + 133 + ], + "spans": [ + { + "bbox": [ + 456, + 125, + 482, + 133 + ], + "type": "text", + "content": "DASR" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 106, + 201, + 224, + 317 + ], + "blocks": [ + { + "bbox": [ + 106, + 201, + 224, + 317 + ], + "lines": [ + { + "bbox": [ + 106, + 201, + 224, + 317 + ], + "spans": [ + { + "bbox": [ + 106, + 201, + 224, + 317 + ], + "type": "image", + "image_path": "420f3298cbe28b597a5d32755eebdcde63fb0f87bb99abe1b65a783c3118ced2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 319, + 195, + 330 + ], + "lines": [ + { + "bbox": [ + 140, + 319, + 195, + 330 + ], + "spans": [ + { + "bbox": [ + 140, + 319, + 195, + 330 + ], + "type": "text", + "content": "RealSet lion1" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 225, + 137, + 293, + 188 + ], + "blocks": [ + { + "bbox": [ + 225, + 137, + 293, + 188 + ], + "lines": [ + { + "bbox": [ + 225, + 137, + 293, + 188 + ], + "spans": [ + { + "bbox": [ + 225, + 137, + 293, + 188 + ], + "type": "image", + "image_path": "3b75d8c3fdf20a7d4486776507d46c440de06fd04b13b0655267c1e79378df34.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 190, + 277, + 199 + ], + "lines": [ + { + "bbox": [ + 241, + 190, + 277, + 199 + ], + "spans": [ + { + "bbox": [ + 241, + 190, + 277, + 199 + ], + "type": "text", + "content": "LDM-15" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 294, + 137, + 364, + 188 + ], + "blocks": [ + { + "bbox": [ + 294, + 137, + 364, + 188 + ], + "lines": [ + { + "bbox": [ + 294, + 137, + 364, + 188 + ], + "spans": [ + { + "bbox": [ + 294, + 137, + 364, + 188 + ], + "type": "image", + "image_path": "a3cd768b1e35f8b9c51707228c16a300953dc05421ce5c0c9ce48c0848ae65e1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 190, + 352, + 198 + ], + "lines": [ + { + "bbox": [ + 306, + 190, + 352, + 198 + ], + "spans": [ + { + "bbox": [ + 306, + 190, + 352, + 198 + ], + "type": "text", + "content": "ResShift-15" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 365, + 137, + 433, + 188 + ], + "blocks": [ + { + "bbox": [ + 365, + 137, + 433, + 188 + ], + "lines": [ + { + "bbox": [ + 365, + 137, + 433, + 188 + ], + "spans": [ + { + "bbox": [ + 365, + 137, + 433, + 188 + ], + "type": "image", + "image_path": "f1b264e492ea056ccbad8d058809c648e19d7b5d51dbba6609e1c426b9463552.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 190, + 420, + 198 + ], + "lines": [ + { + "bbox": [ + 378, + 190, + 420, + 198 + ], + "spans": [ + { + "bbox": [ + 378, + 190, + 420, + 198 + ], + "type": "text", + "content": "ResShift-4" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 434, + 137, + 504, + 188 + ], + "blocks": [ + { + "bbox": [ + 434, + 137, + 504, + 188 + ], + "lines": [ + { + "bbox": [ + 434, + 137, + 504, + 188 + ], + "spans": [ + { + "bbox": [ + 434, + 137, + 504, + 188 + ], + "type": "image", + "image_path": "6ba80c14643f9dc85ddd1053c9d0a23caa5133f6e08d9794b4a5a06a6fe5e787.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 190, + 479, + 198 + ], + "lines": [ + { + "bbox": [ + 459, + 190, + 479, + 198 + ], + "spans": [ + { + "bbox": [ + 459, + 190, + 479, + 198 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 106, + 331, + 224, + 448 + ], + "blocks": [ + { + "bbox": [ + 106, + 331, + 224, + 448 + ], + "lines": [ + { + "bbox": [ + 106, + 331, + 224, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 224, + 448 + ], + "type": "image", + "image_path": "edc46ab5f7d3db251268e91e68e54d7edb654a6f7255b842ede5c4e925261e94.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 449, + 206, + 460 + ], + "lines": [ + { + "bbox": [ + 129, + 449, + 206, + 460 + ], + "spans": [ + { + "bbox": [ + 129, + 449, + 206, + 460 + ], + "type": "text", + "content": "RealSR_Nikon_047" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 93, + 468, + 515, + 480 + ], + "lines": [ + { + "bbox": [ + 93, + 468, + 515, + 480 + ], + "spans": [ + { + "bbox": [ + 93, + 468, + 515, + 480 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison between different methods on real-world datasets. Please zoom in for more details." + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 225, + 201, + 294, + 252 + ], + "blocks": [ + { + "bbox": [ + 225, + 201, + 294, + 252 + ], + "lines": [ + { + "bbox": [ + 225, + 201, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 225, + 201, + 294, + 252 + ], + "type": "image", + "image_path": "cd48bf2bd1e2bdf4a7452ee004db74d6df3aabe63d34cfd5da5e386f2585446c.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 255, + 279, + 263 + ], + "lines": [ + { + "bbox": [ + 239, + 255, + 279, + 263 + ], + "spans": [ + { + "bbox": [ + 239, + 255, + 279, + 263 + ], + "type": "text", + "content": "BSRGAN" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 294, + 201, + 364, + 252 + ], + "blocks": [ + { + "bbox": [ + 294, + 201, + 364, + 252 + ], + "lines": [ + { + "bbox": [ + 294, + 201, + 364, + 252 + ], + "spans": [ + { + "bbox": [ + 294, + 201, + 364, + 252 + ], + "type": "image", + "image_path": "cd9d6a4c77072601bdd8c7eb2591fb8e750bcfdc20d09e0c0222c08e84d60936.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 301, + 255, + 358, + 263 + ], + "lines": [ + { + "bbox": [ + 301, + 255, + 358, + 263 + ], + "spans": [ + { + "bbox": [ + 301, + 255, + 358, + 263 + ], + "type": "text", + "content": "RealESRGAN" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 365, + 201, + 433, + 252 + ], + "blocks": [ + { + "bbox": [ + 365, + 201, + 433, + 252 + ], + "lines": [ + { + "bbox": [ + 365, + 201, + 433, + 252 + ], + "spans": [ + { + "bbox": [ + 365, + 201, + 433, + 252 + ], + "type": "image", + "image_path": "1485f27b6d3da9538d9e540823c18d08abfd19e984c4aaab3af4d245cf341717.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 255, + 414, + 263 + ], + "lines": [ + { + "bbox": [ + 384, + 255, + 414, + 263 + ], + "spans": [ + { + "bbox": [ + 384, + 255, + 414, + 263 + ], + "type": "text", + "content": "SwinIR" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 434, + 201, + 504, + 252 + ], + "blocks": [ + { + "bbox": [ + 434, + 201, + 504, + 252 + ], + "lines": [ + { + "bbox": [ + 434, + 201, + 504, + 252 + ], + "spans": [ + { + "bbox": [ + 434, + 201, + 504, + 252 + ], + "type": "image", + "image_path": "b5738fb580fc7c8b6c354ef7251dca7e32b3a00d25ba18abc2d0c67d45d45217.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 255, + 481, + 263 + ], + "lines": [ + { + "bbox": [ + 456, + 255, + 481, + 263 + ], + "spans": [ + { + "bbox": [ + 456, + 255, + 481, + 263 + ], + "type": "text", + "content": "DASR" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 225, + 266, + 294, + 318 + ], + "blocks": [ + { + "bbox": [ + 225, + 266, + 294, + 318 + ], + "lines": [ + { + "bbox": [ + 225, + 266, + 294, + 318 + ], + "spans": [ + { + "bbox": [ + 225, + 266, + 294, + 318 + ], + "type": "image", + "image_path": "d905cde3131a53209b3501413a8def071a0046cae149960786472d3827f5e0ae.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 319, + 277, + 329 + ], + "lines": [ + { + "bbox": [ + 242, + 319, + 277, + 329 + ], + "spans": [ + { + "bbox": [ + 242, + 319, + 277, + 329 + ], + "type": "text", + "content": "LDM-15" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 294, + 266, + 364, + 318 + ], + "blocks": [ + { + "bbox": [ + 294, + 266, + 364, + 318 + ], + "lines": [ + { + "bbox": [ + 294, + 266, + 364, + 318 + ], + "spans": [ + { + "bbox": [ + 294, + 266, + 364, + 318 + ], + "type": "image", + "image_path": "3744f14f7292279e0527f13917de22152ddb0eb12375ccf2346610fa34bd486d.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 319, + 352, + 329 + ], + "lines": [ + { + "bbox": [ + 306, + 319, + 352, + 329 + ], + "spans": [ + { + "bbox": [ + 306, + 319, + 352, + 329 + ], + "type": "text", + "content": "ResShift-15" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 365, + 266, + 433, + 318 + ], + "blocks": [ + { + "bbox": [ + 365, + 266, + 433, + 318 + ], + "lines": [ + { + "bbox": [ + 365, + 266, + 433, + 318 + ], + "spans": [ + { + "bbox": [ + 365, + 266, + 433, + 318 + ], + "type": "image", + "image_path": "a171951f36154a1e5a140d4f21c7d7caec5a8112fc516b63db77dc467e625342.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 319, + 420, + 329 + ], + "lines": [ + { + "bbox": [ + 378, + 319, + 420, + 329 + ], + "spans": [ + { + "bbox": [ + 378, + 319, + 420, + 329 + ], + "type": "text", + "content": "ResShift-4" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 434, + 266, + 504, + 318 + ], + "blocks": [ + { + "bbox": [ + 434, + 266, + 504, + 318 + ], + "lines": [ + { + "bbox": [ + 434, + 266, + 504, + 318 + ], + "spans": [ + { + "bbox": [ + 434, + 266, + 504, + 318 + ], + "type": "image", + "image_path": "f50e01ccc54e6f7dc071241fb1f11fae7f71adff2a7925511f95fb7103d70033.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 319, + 479, + 329 + ], + "lines": [ + { + "bbox": [ + 459, + 319, + 479, + 329 + ], + "spans": [ + { + "bbox": [ + 459, + 319, + 479, + 329 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 225, + 331, + 294, + 383 + ], + "blocks": [ + { + "bbox": [ + 225, + 331, + 294, + 383 + ], + "lines": [ + { + "bbox": [ + 225, + 331, + 294, + 383 + ], + "spans": [ + { + "bbox": [ + 225, + 331, + 294, + 383 + ], + "type": "image", + "image_path": "a0b0a6c8975af49113d9d1b7d80ab2f2416a9ce5b637388329328bfa757c8869.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 384, + 279, + 393 + ], + "lines": [ + { + "bbox": [ + 239, + 384, + 279, + 393 + ], + "spans": [ + { + "bbox": [ + 239, + 384, + 279, + 393 + ], + "type": "text", + "content": "BSRGAN" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 294, + 331, + 364, + 383 + ], + "blocks": [ + { + "bbox": [ + 294, + 331, + 364, + 383 + ], + "lines": [ + { + "bbox": [ + 294, + 331, + 364, + 383 + ], + "spans": [ + { + "bbox": [ + 294, + 331, + 364, + 383 + ], + "type": "image", + "image_path": "ce42b425bef3092cdc3753ead8b94844bbda50db5594ac408835442c37501c50.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 301, + 384, + 358, + 393 + ], + "lines": [ + { + "bbox": [ + 301, + 384, + 358, + 393 + ], + "spans": [ + { + "bbox": [ + 301, + 384, + 358, + 393 + ], + "type": "text", + "content": "RealESRGAN" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 365, + 331, + 433, + 383 + ], + "blocks": [ + { + "bbox": [ + 365, + 331, + 433, + 383 + ], + "lines": [ + { + "bbox": [ + 365, + 331, + 433, + 383 + ], + "spans": [ + { + "bbox": [ + 365, + 331, + 433, + 383 + ], + "type": "image", + "image_path": "6bdf6b5539c279b5d75b8de85011de35c49f64d1c041195a3f6fcdd90fc07512.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 384, + 414, + 393 + ], + "lines": [ + { + "bbox": [ + 384, + 384, + 414, + 393 + ], + "spans": [ + { + "bbox": [ + 384, + 384, + 414, + 393 + ], + "type": "text", + "content": "SwinIR" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 434, + 331, + 504, + 383 + ], + "blocks": [ + { + "bbox": [ + 434, + 331, + 504, + 383 + ], + "lines": [ + { + "bbox": [ + 434, + 331, + 504, + 383 + ], + "spans": [ + { + "bbox": [ + 434, + 331, + 504, + 383 + ], + "type": "image", + "image_path": "050394a69fcda9eee7a86f87b84052a38a61f4ad663c03b6e0192923912f32c5.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 384, + 481, + 393 + ], + "lines": [ + { + "bbox": [ + 456, + 384, + 481, + 393 + ], + "spans": [ + { + "bbox": [ + 456, + 384, + 481, + 393 + ], + "type": "text", + "content": "DASR" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "bbox": [ + 55, + 492, + 295, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 295, + 577 + ], + "type": "text", + "content": "Qualitative Comparison. In Fig. 6, we provide visual results produced by different methods on real-world datasets. These results validate that the proposed UPSR method is capable of producing visually better outputs. Specifically, UPSR successfully recover finer textures and sharper edges, while other methods yields blurry results. More visual examples are presented in the supplementary material." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 55, + 586, + 128, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 128, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 128, + 597 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "In this work, we propose specialized prior distribution and sampling pipeline, along with modifications on the network architecture to develop the Uncertainty-guided Perturbation scheme for Super-Resolution to tackle real-world SR task. Specifically, we propose region-specific processing based on the LR content from a uncertainty-based perspective. In flat areas (low-uncertainty), reducing the noise level could preserve more details in the initial state of the diffusion process and therefore improve performance, whereas edge and" + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 313, + 493, + 555, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 555, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 555, + 648 + ], + "type": "text", + "content": "texture regions (high-uncertainty) demand stronger noise to bridge their significant disparity with the target distribution. To facilitate region-specific control, we leverage the uncertainty of LR image as the criterion to adjust noise level in different areas. We employ an auxiliary SR network and consider the residual of SR prediction as the approximate uncertainty estimate of LR image due to their similarity. We further incorporate the SR prediction as conditional information and make several architectural modifications to provide a more robust pipeline to further enhance performance. Extensive experiments on both synthetic and real-world datasets validate the efficacy of the proposed UPSR model, even with lower model size and training overhead." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by National Natural Science Foundation of China (No. 62476051) and Sichuan Natural Science Foundation (No. 2024NSFTD0041)." + } + ] + } + ], + "index": 51 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17987" + } + ] + } + ], + "index": 52 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 146 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 146 + ], + "type": "text", + "content": "[1] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3086-3095, 2019. 7, 15" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "text", + "content": "[2] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image superresolution transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22367-22377, 2023. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 205, + 295, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 295, + 248 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 295, + 248 + ], + "type": "text", + "content": "[3] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 250, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 250, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 250, + 295, + 293 + ], + "type": "text", + "content": "[4] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 295, + 295, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 295, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 295, + 295, + 338 + ], + "type": "text", + "content": "[5] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 341, + 295, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 341, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 341, + 295, + 384 + ], + "type": "text", + "content": "[6] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 386, + 295, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 386, + 295, + 439 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 295, + 439 + ], + "type": "text", + "content": "[7] Shuhang Gu, Wangmeng Zuo, Qi Xie, Deyu Meng, Xiangchu Feng, and Lei Zhang. Convolutional sparse coding for image super-resolution. In Proceedings of the IEEE International Conference on Computer Vision, pages 1823-1831, 2015. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 441, + 295, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 441, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 441, + 295, + 475 + ], + "type": "text", + "content": "[8] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 477, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 477, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 477, + 295, + 510 + ], + "type": "text", + "content": "[9] Aapo Hyvarinen and Peter Dayan. Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research, 6(4), 2005. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 512, + 295, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 512, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 57, + 512, + 295, + 566 + ], + "type": "text", + "content": "[10] Xiaozhong Ji, Yun Cao, Ying Tai, Chengjie Wang, Jilin Li, and Feiyue Huang. Real-world super-resolution via kernel estimation and noise injection. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 466-467, 2020. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 568, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 568, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 568, + 295, + 612 + ], + "type": "text", + "content": "[11] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 613, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 656 + ], + "type": "text", + "content": "[12] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in neural information processing systems, 35:26565-26577, 2022. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "text", + "content": "[13] Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and improving the training dynamics of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24174-24184, 2024. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[14] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 119, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 152 + ], + "type": "text", + "content": "[15] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 155, + 553, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 155, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 317, + 155, + 553, + 209 + ], + "type": "text", + "content": "[16] Changwoo Lee and Ki-Seok Chung. Gram: Gradient rescaling attention model for data uncertainty estimation in single image super resolution. In 2019 18th IEEE International Conference On Machine Learning And Applications (ICMLA), pages 8-13. IEEE, 2019. 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 212, + 553, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 212, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 316, + 212, + 553, + 255 + ], + "type": "text", + "content": "[17] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 258, + 553, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 258, + 553, + 322 + ], + "spans": [ + { + "bbox": [ + 317, + 258, + 553, + 322 + ], + "type": "text", + "content": "[18] Yawei Li, Yuchen Fan, Xiaoyu Xiang, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18278-18289, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 326, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 326, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 326, + 553, + 380 + ], + "type": "text", + "content": "[19] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 1, 2, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "text", + "content": "[20] Jie Liang, Hui Zeng, and Lei Zhang. Efficient and degradation-adaptive network for real-world image superresolution. In European Conference on Computer Vision, pages 574-591. Springer, 2022. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 429, + 553, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 429, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 316, + 429, + 553, + 483 + ], + "type": "text", + "content": "[21] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 486, + 553, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 486, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 316, + 486, + 553, + 518 + ], + "type": "text", + "content": "[22] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 521, + 553, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 521, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 317, + 521, + 553, + 563 + ], + "type": "text", + "content": "[23] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. arXiv preprint arXiv:1802.05957, 2018. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 567, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 609 + ], + "type": "text", + "content": "[24] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "type": "text", + "content": "[25] Qian Ning, Weisheng Dong, Xin Li, Jinjian Wu, and Guangming Shi. Uncertainty-driven loss for single image superresolution. Advances in Neural Information Processing Systems, 34:16398-16409, 2021. 2, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "type": "text", + "content": "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 5, 6, 7, 13" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17988" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[27] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE transactions on pattern analysis and machine intelligence, 45(4):4713-4726, 2022. 1, 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 130, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 294, + 196 + ], + "type": "text", + "content": "[28] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1874-1883, 2016. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "type": "text", + "content": "[29] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 255, + 294, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 294, + 287 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 294, + 287 + ], + "type": "text", + "content": "[30] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 290, + 294, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 290, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 56, + 290, + 294, + 323 + ], + "type": "text", + "content": "[31] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 325, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 325, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 325, + 294, + 369 + ], + "type": "text", + "content": "[32] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 393 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 393 + ], + "type": "text", + "content": "[33] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 396, + 294, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 396, + 294, + 439 + ], + "spans": [ + { + "bbox": [ + 56, + 396, + 294, + 439 + ], + "type": "text", + "content": "[34] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 441, + 295, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 441, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 56, + 441, + 295, + 485 + ], + "type": "text", + "content": "[35] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. International Journal of Computer Vision, 132(12):5929-5949, 2024. 12" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 487, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 294, + 542 + ], + "type": "text", + "content": "[36] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 544, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 599 + ], + "type": "text", + "content": "[37] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1905-1914, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 601, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 667 + ], + "type": "text", + "content": "[38] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image superresolution in a single step. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 25796-25805, 2024. 13" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[39] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. Advances in Neural Information Processing Systems, 37:92529-92553, 2025. 12" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 607 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[40] Rui Xie, Chen Zhao, Kai Zhang, Zhenyu Zhang, Jun Zhou, Jian Yang, and Ying Tai. Addsr: Accelerating diffusion-based blind super-resolution with adversarial diffusion distillation. arXiv preprint arXiv:2404.01717, 2024. 12" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[41] Jianchao Yang, John Wright, Thomas S Huang, and Yi Ma. Image super-resolution via sparse representation. IEEE transactions on image processing, 19(11):2861-2873, 2010. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 228 + ], + "type": "text", + "content": "[42] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniaq: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1191-1200, 2022. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 230, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 262 + ], + "type": "text", + "content": "[43] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Efficient diffusion model for image restoration by residual shifting. arXiv preprint arXiv:2403.07319, 2024. 3, 13" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 317 + ], + "type": "text", + "content": "[44] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image superresolution by residual shifting. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 5, 6, 7, 11, 13, 14" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "type": "text", + "content": "[45] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In Curves and Surfaces: 7th International Conference, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 375, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 375, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 375, + 553, + 430 + ], + "type": "text", + "content": "[46] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4791-4800, 2021. 1, 2, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 431, + 553, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 431, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 316, + 431, + 553, + 495 + ], + "type": "text", + "content": "[47] Leheng Zhang, Yawei Li, Xingyu Zhou, Xiaorui Zhao, and Shuhang Gu. Transcending the limit of local window: Advanced super-resolution transformer with adaptive token dictionary. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856-2865, 2024. 1, 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 498, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 552 + ], + "type": "text", + "content": "[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 554, + 553, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 554, + 553, + 607 + ], + "spans": [ + { + "bbox": [ + 316, + 554, + 553, + 607 + ], + "type": "text", + "content": "[49] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 286-301, 2018. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17989" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Understanding Fine-tuning CLIP for Open-vocabulary Semantic Segmentation in Hyperbolic Space/4933d0fc-2175-4f82-8204-a2ef1c424606_content_list.json b/2025/Understanding Fine-tuning CLIP for Open-vocabulary Semantic Segmentation in Hyperbolic Space/4933d0fc-2175-4f82-8204-a2ef1c424606_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e7fc5352990690d898725e32773131efb1758c47 --- /dev/null +++ b/2025/Understanding Fine-tuning CLIP for Open-vocabulary Semantic Segmentation in Hyperbolic Space/4933d0fc-2175-4f82-8204-a2ef1c424606_content_list.json @@ -0,0 +1,1556 @@ +[ + { + "type": "text", + "text": "Understanding Fine-tuning CLIP for Open-vocabulary Semantic Segmentation in Hyperbolic Space", + "text_level": 1, + "bbox": [ + 91, + 128, + 906, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zelin Peng $^{1}$ , Zhengqin Xu $^{1}$ , Zhilin Zeng $^{1}$ , Changsong Wen $^{1}$ , Yu Huang $^{1}$ , Menglin Yang $^{2}$ , Feilong Tang $^{3,4}$ , and Wei Shen $^{1(\\boxtimes)}$", + "bbox": [ + 207, + 202, + 787, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University $^{2}$ Hong Kong University of Science and Technology (Guangzhou) $^{3}$ MBZUAI, $^{4}$ Monash University.", + "bbox": [ + 109, + 239, + 885, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/SJTU-DeepVisionLab/HyperCLIP", + "bbox": [ + 285, + 277, + 704, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 342, + 326, + 358 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "CLIP, a foundational vision-language model, has emerged as a powerful tool for open-vocabulary semantic segmentation. While freezing the text encoder preserves its powerful embeddings, recent studies show that fine-tuning both the text and image encoders jointly significantly enhances segmentation performance, especially for classes from open sets. In this work, we explain this phenomenon from the perspective of hierarchical alignment, since during fine-tuning, the hierarchy level of image embeddings shifts from image-level to pixel-level. We achieve this by leveraging hyperbolic space, which naturally encoders hierarchical structures. Our key observation is that, during fine-tuning, the hyperbolic radius of CLIP's text embeddings decreases, facilitating better alignment with the pixel-level hierarchical structure of visual data. Building on this insight, we propose HyperCLIP, a novel fine-tuning strategy that adjusts the hyperbolic radius of the text embeddings through scaling transformations. By doing so, HyperCLIP equips CLIP with segmentation capability while introducing only a small number of learnable parameters. Our experiments demonstrate that HyperCLIP achieves state-of-the-art performance on open-vocabulary semantic segmentation tasks across three benchmarks, while fine-tuning only approximately $4\\%$ of the total parameters of CLIP. More importantly, we observe that after adjustment, CLIP's text embeddings exhibit a relatively fixed hyperbolic radius across datasets, suggesting that the granularity required for this segmentation task might be quantified using the hyperbolic radius.", + "bbox": [ + 88, + 373, + 485, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 823, + 220, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The goal of open-vocabulary semantic segmentation is to develop a segmentation model capable of labeling each", + "bbox": [ + 89, + 849, + 482, + 880 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ce1ce3e1555f07c3dd7048c05431bbe51de4cea6ceaa505bb78707e8d71bf6a0.jpg", + "image_caption": [ + "Figure 1. Visualization of hyperbolic space by a Poincaré ball model: The height of points reflects the hierarchical level of input data. The black segments refer to the shortest path between the points in the Poincaré ball model. The red arrow line denotes the hyperbolic radius of text embeddings. Embeddings closer to the center have a smaller radius, representing low-level visual information (e.g., pixels), while points farther from the center have a larger radius, representing high-level visual information (e.g., images). During fine-tuning, the hyperbolic radius of text embeddings is adjusted to match the radius associated with pixels, thereby endowing CLIP with segmentation ability (Best viewed in color)." + ], + "image_footnote": [], + "bbox": [ + 524, + 329, + 893, + 614 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pixel in an image with categories that extend beyond a predefined closed set, based on textual descriptions. Vision-language foundation models [7, 12, 13, 15, 22, 23, 26, 28, 32–35, 39, 43, 44, 46, 51, 52, 55, 63], particularly CLIP [46], are frequently employed to provide open-vocabulary recognition capabilities. Consequently, open-vocabulary seman", + "bbox": [ + 511, + 810, + 911, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding Author: wei.shen@sjtu.edu.cn", + "bbox": [ + 112, + 886, + 413, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4562", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tic segmentation fundamentally involves transferring these vision-language foundation models, originally trained with image-level supervision, to achieve pixel-level predictions.", + "bbox": [ + 89, + 90, + 480, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Current methods [9, 59, 62, 65] typically fine-tune CLIP on a closed set with segmentation annotations, i.e., COCO [4], to equip it with the segmentation capability. A prevailing opinion is that simply freezing CLIP's text encoder maximally preserves the powerful text embeddings of a vast array of classes, which is believed to be beneficial for good generalization on open sets. Many studies follow this opinion [31, 37, 62]. However, recent studies [9, 42, 59] offer an alternative perspective on fine-tuning: simultaneously fine-tuning both the image and text encoders results in superior segmentation performance on open sets compared to former approaches. To explain and resolve this seemingly paradoxical phenomenon, we first leverage a non-Euclidean manifold—specifically, hyperbolic space [5]—to understand why fine-tuning the text encoder appears to improve segmentation performance on open sets. Building on this understanding, our objective is to design an explainable fine-tuning strategy with a minimal number of introduced tunable parameters, which is able to equip CLIP with strong segmentation performance while minimizing the loss of its generalization ability.", + "bbox": [ + 91, + 138, + 486, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing research [1] has found that images inherently exhibit hierarchical structures, consisting of multiple levels: pixels, patches, objects, entire scenes, etc. Given that, in open-vocabulary semantic segmentation, CLIP's image embeddings transition from image level to pixel level, we infer that the performance improvement brought by the recent studies [9, 59] could be attributed to adjusting CLIP's text embeddings from their original hierarchical level to better align with the level represented by pixels, thereby enhancing cross-modal alignment. However, most existing fine-tuning methods operate in Euclidean space, which makes it challenging to quantify such a transition through different hierarchical levels. Hyperbolic spaces have gained significant interest in recent years, owing to their ability to naturally and compactly encode hierarchical structures [47, 48, 57], as shown in Fig. 1. The hyperbolic radius, defined as the distance from a point to the center in hyperbolic space, represents its hierarchical level [50]. Points closer to the center (smaller radius) correspond to low-level visual embeddings, e.g., those from pixels, while points near the edge (larger radius) correspond to high-level visual embeddings, e.g., those from images.", + "bbox": [ + 91, + 459, + 485, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by this, we project the text embeddings into hyperbolic space and visualize the changes in their radii before and after fine-tuning, as shown in Fig. 2. This fine-tuning is realized using the state-of-the-art method provided in [9]. Specifically, one can observe that the radius after fine-tuning is smaller than that of pre-trained CLIP. In particular, we identify a pattern during fine-tuning: the hierarchical level of", + "bbox": [ + 89, + 795, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ae42f7ff7a74c0af75347c3ac569e41b75208e3d2c72093f57eb1f21128d9361.jpg", + "image_caption": [ + "Figure 2. Understanding fine-tuning CLIP's text encoder from the perspective of hyperbolic radius. We use hyperbolic radius to illustrate the change in the hierarchical level of 171 classes after fine-tuning. We observe that the hyperbolic radius of each class becomes smaller, suggesting that adjusting the hyperbolic radius equips CLIP with segmentation ability (Best viewed in color)." + ], + "image_footnote": [], + "bbox": [ + 513, + 85, + 906, + 265 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the text embeddings adjusts in conjunction with the hierarchical level of the embeddings provided by the image encoder. Our finding brings a novel perspective for fine-tuning CLIP: Adjusting the hyperbolic radius of text embeddings from edge to center endowed CLIP with segmentation ability.", + "bbox": [ + 511, + 383, + 906, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on the analysis above, we propose a novel finetuning strategy for CLIP in hyperbolic space, termed HyperCLIP. HyperCLIP aims to directly adjust the hyperbolic radius of CLIP's embeddings to facilitate alignment with the appropriate hierarchical level for segmentation tasks. Specifically, we introduce several block-diagonal scaling matrices. Multiplying these matrices with CLIP's feature representations using a Möbius matrix multiplication operation is equivalent to performing scaling transformations that adjust the radius of CLIP's embeddings. This enables the adjustment of their hierarchical levels, ultimately enhancing the model's performance. As a result, HyperCLIP can equip CLIP with segmentation ability with only a small number of introduced learnable parameters.", + "bbox": [ + 511, + 460, + 908, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments demonstrate that our method sets new state-of-the-art results in open-vocabulary semantic segmentation across three benchmarks, while fine-tuning only approximately $4\\%$ of the total parameters of CLIP. More importantly, we observe that after adjustment, CLIP's text embeddings maintain a relatively fixed hyperbolic radius across different datasets. This suggests that the granularity of the segmentation task might be quantified using the hyperbolic radius, warranting further exploration.", + "bbox": [ + 511, + 672, + 910, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 821, + 653, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Open-vocabulary Semantic Segmentation", + "text_level": 1, + "bbox": [ + 511, + 847, + 867, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prior works on open-vocabulary semantic segmentation typically tackle this task by leveraging CLIP [46]. Early efforts,", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4563", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "e.g., [68], directly fine-tune CLIP on mainstream segmentation datasets, like COCO [4]. However, these approaches argue that fully fine-tuning CLIP's encoder significantly diminishes its ability to generalize to unseen classes. To mitigate this issue, some methods [11, 17, 60, 61] take the opposite approach, fine-tuning an additional mask generator [8] for segmentation while keeping CLIP frozen to preserve its generalization capabilities. However, this approach, with its frozen parameter space, lacks segmentation awareness, leading to a misalignment between regions and text descriptions [37]. Other studies [9, 62, 65] propose more sophisticated solutions that fine-tune only specific parameters, e.g., certain layers of CLIP, enabling pixel-level predictions while keeping the majority of CLIP's parameters fixed to minimize the loss of generalization. While these approaches show significant advantages, they often rely on very small learning rates, implicitly promoting minimal deviations from the pre-trained CLIP, which limits segmentation performance. In summary, there is still limited understanding of how fine-tuning influences CLIP's representations and why fine-tuning CLIP's text encoder appears to enhance segmentation performance. In this paper, we aim to provide several new insights that reveal the underlying logic by incorporating a novel perspective: hyperbolic space.", + "bbox": [ + 89, + 90, + 486, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Learning in Hyperbolic Spaces", + "text_level": 1, + "bbox": [ + 89, + 464, + 364, + 481 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike Euclidean geometry, hyperbolic space can be viewed as the continuous analog of a tree [2], enabling a hierarchical understanding of knowledge structures. Previous research has demonstrated the advantages of representing such hierarchical relationships in various domains, including molecular structures [66], 3D data [6, 56], text data [16, 70], and images [18, 27]. Furthermore, several studies have investigated hyperbolic learning in vision tasks to model hierarchical relationships across different modalities [10, 30, 47], showcasing the benefits of capturing the semantic hierarchy among various data types. Nevertheless, while most studies typically leverage the intrinsic hierarchical properties of hyperbolic space, this paper seeks to capture the detailed changes in semantics reflected by the hyperbolic radius. This approach potentially complements current fine-tuning efforts in open-vocabulary tasks by offering a novel hyperbolic perspective.", + "bbox": [ + 89, + 488, + 486, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary Background", + "text_level": 1, + "bbox": [ + 89, + 744, + 323, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Hyperbolic geometry", + "text_level": 1, + "bbox": [ + 89, + 771, + 290, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hyperbolic geometry often has constant negative curvature and is fundamentally different to the Euclidean or spherical geometry which has zero or constant positive curvature, respectively. This enables unique properties such as the divergence of parallel lines and the exponential volume growth towards the boundary. Hyperbolic geometry has five different analytic models to build the hyperbolic space [5]. In this", + "bbox": [ + 89, + 795, + 486, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "work, we utilize the classical Poincaré ball model as its effectiveness in capturing hierarchical structures [16, 20, 41, 47].", + "bbox": [ + 511, + 90, + 908, + 122 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. The Poincaré Ball Model", + "text_level": 1, + "bbox": [ + 511, + 133, + 741, + 148 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Poincaré ball model $(\\mathbb{D}_c^n,g^{\\mathbb{D}_c})$ with radius $1 / \\sqrt{c}$ and negative curvature $-c(c > 0)$ is a Riemannian manifold $\\mathbb{D}_c^n \\coloneqq \\{\\mathbf{x} \\in \\mathbb{R}^n : c\\|\\mathbf{x}\\| < 1\\}$ with the metric $g^{\\mathbb{D}_c} = \\lambda_{c,\\mathbf{x}}^2 g^E$ , where $\\lambda_{c,\\mathbf{x}} = \\frac{2}{1 - c\\|\\mathbf{x}\\|^2}$ is the conformal factor, and $g^E = \\mathbf{I}_n$ is the Euclidean metric tensor. The vectorial structure operations defined in Euclidean space can not be directly used in hyperbolic space due to the differential of the space properties. Therefore, we introduce the framework of gyrovector spaces [53, 54] to provide non-associative algebraic operations for hyperbolic space. According to special relativity, when adding speed vectors belonging to the Poincaré ball of radius $c$ , these added speed vectors remain in the ball by the gyrovector spaces operation, also known as the Möbius operation. For instance, the addition and multiplication operations are defined as follows:", + "bbox": [ + 511, + 156, + 908, + 387 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mobius addition operation. For $\\mathbf{x},\\mathbf{y}\\in \\mathbb{D}_c^n$ , the Mobius addition operation is defined as:", + "bbox": [ + 511, + 391, + 906, + 421 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} \\oplus_ {c} \\mathbf {y} := \\frac {(1 + 2 c \\langle \\mathbf {x} , \\mathbf {y} \\rangle + c \\| \\mathbf {y} \\| ^ {2}) \\mathbf {x} + (1 - c \\| \\mathbf {x} \\| ^ {2}) \\mathbf {y}}{1 + 2 c \\langle \\mathbf {x} , \\mathbf {y} \\rangle + c ^ {2} \\| \\mathbf {x} \\| ^ {2} \\| \\mathbf {y} \\| ^ {2}}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 433, + 906, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mobius scalar multiplication operation. For a scalar $r \\in \\mathbb{R}$ and a vector $\\mathbf{x} \\in \\mathbb{D}_c^n$ , the Möbius scalar multiplication operation is defined as:", + "bbox": [ + 511, + 484, + 908, + 529 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nr \\otimes_ {c} \\mathbf {x} := (1 / \\sqrt {c}) \\tanh \\left(r \\tanh ^ {- 1} (\\sqrt {c} \\| \\mathbf {x} \\|\\right)) \\frac {\\mathbf {x}}{\\| \\mathbf {x} \\|}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 541, + 906, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The operation satisfies $r \\otimes_{c} \\mathbf{0} \\coloneqq \\mathbf{0}$ , $r \\otimes_{c} \\mathbf{x} = \\mathbf{x} \\oplus_{c} \\dots \\oplus_{c} \\mathbf{x}$ ( $r$ additions), $(r' + r) \\otimes_{c} \\mathbf{x} = r \\otimes \\mathbf{x} \\oplus_{c} r' \\otimes_{c} \\mathbf{x}$ , $(r'r) \\otimes_{c} \\mathbf{x} = r \\otimes_{c} (r' \\otimes_{c} \\mathbf{x})$ and $\\left(\\left|r\\right| \\otimes_{c} \\mathbf{x}\\right) / \\left\\|r \\otimes_{c} \\mathbf{x}\\right\\| = \\mathbf{x} / \\left\\|\\mathbf{x}\\right\\|$ .", + "bbox": [ + 511, + 583, + 908, + 630 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mobius matrix multiplication operation. Refer to the relation definition in [27], for a matrix $\\mathbf{M} \\in \\mathbb{R}^{n \\times n}$ and a vector $\\mathbf{x} \\in \\mathbb{D}_c^n$ , if $\\mathbf{Mx} \\neq \\mathbf{0}$ , the Möbius matrix multiplication operation is defined as:", + "bbox": [ + 511, + 632, + 908, + 693 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} \\otimes_ {c} \\mathbf {x} := (\\frac {1}{\\sqrt {c}}) \\tanh \\left(\\frac {\\| \\mathbf {M x} \\|}{\\| \\mathbf {x} \\|} \\tanh ^ {- 1} (\\sqrt {c} \\| \\mathbf {x} \\|)\\right) \\frac {\\mathbf {M x}}{\\| \\mathbf {M x} \\|}.\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 704, + 901, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The distance of $\\mathbf{x},\\mathbf{y}\\in \\mathbb{D}_c^n$ is defined as:", + "bbox": [ + 511, + 750, + 779, + 766 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd _ {c} ^ {\\mathbb {D}} (\\mathbf {x}, \\mathbf {y}) = (2 / \\sqrt {c}) \\tanh ^ {- 1} (\\sqrt {c} \\| - \\mathbf {x} \\oplus_ {c} \\mathbf {y} \\Vert). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 777, + 906, + 796 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The tangent space $\\mathcal{T}_{\\mathbf{x}}^{c}\\mathbb{D}_{c}^{n}$ at a point $\\mathbf{x}\\in \\mathbb{D}_c^n$ is the first order approximation of $\\mathbb{D}_c^n$ , which is an $n$ -dimensional Euclidean space. The tangent space $\\mathcal{T}_{\\mathbf{x}}^{c}\\mathbb{D}_{c}^{n}$ and $\\mathbb{D}_c^n$ are mapped to each other by exponential $(\\mathcal{T}_{\\mathbf{x}}^{c}\\mathbb{D}_{c}^{n}\\mapsto \\mathbb{D}_{c}^{n}:\\exp_{\\mathbf{x}}^{\\mathbb{D},c}(\\cdot))$ and logarithmic $(\\mathbb{D}_c^n\\mapsto \\mathcal{T}_x^c\\mathbb{D}_c^n:\\log_{\\mathbf{x}}^{\\mathbb{D},c}(\\cdot))$ maps, respectively. For any $\\mathbf{x},\\mathbf{y}\\in \\mathbb{D}_c^n$ and $\\mathbf{v}\\in \\mathcal{T}_{\\mathbf{x}}^{c}\\mathbb{D}_{c}^{n}$ , the mapping functions", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4564", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/96a21682e611343295ba4c39cd67d5c983ea22b86ce529b120b03380e9df88be.jpg", + "image_caption": [ + "Figure 3. A schematic representation of HyperCLIP. In the HyperCLIP framework, we fine-tune CLIP in hyperbolic space by minimizing its embeddings' hyperbolic radii using several tunable block-diagonal scaling matrices." + ], + "image_footnote": [], + "bbox": [ + 98, + 89, + 903, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "are given for $\\mathbf{v} \\neq \\mathbf{0}$ and $\\mathbf{y} \\neq \\mathbf{x}$ by:", + "bbox": [ + 89, + 311, + 323, + 329 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\exp_ {\\mathbf {x}} ^ {\\mathbb {D}, c} (\\mathbf {v}) = \\mathbf {x} \\oplus_ {c} \\left(\\tanh \\left(\\sqrt {c} \\frac {\\lambda_ {c , \\mathbf {x}} \\| \\mathbf {v} \\|}{2}\\right) \\frac {\\mathbf {v}}{\\sqrt {c} \\| \\mathbf {v} \\|}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 340, + 483, + 375 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\log_ {\\mathbf {x}} ^ {\\mathbb {D}, c} (\\mathbf {y}) = \\frac {2}{\\sqrt {c} \\lambda_ {c , \\mathbf {x}}} \\tanh ^ {- 1} \\left(\\sqrt {c} \\| - \\mathbf {x} \\oplus_ {c} \\mathbf {y} \\|\\right) \\frac {- \\mathbf {x} \\oplus_ {c} \\mathbf {y}}{\\| - \\mathbf {x} \\oplus_ {c} \\mathbf {y} \\|}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 377, + 491, + 424 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 89, + 440, + 225, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce HyperCLIP for open-vocabulary semantic segmentation. HyperCLIP is based on the Poincaré ball model, which aims to adjust the hyperbolic radius of CLIP's embeddings, thereby equipping CLIP with segmentation ability.", + "bbox": [ + 89, + 467, + 483, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Fine-tuning in Hyperbolic Space", + "text_level": 1, + "bbox": [ + 89, + 556, + 375, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The core technique in HyperCLIP involves linear transformations, i.e., scaling, in the hyperbolic space. One of the primary methods [64] to learn the weights of linear transformations in hyperbolic space is hyperLoRA [64]. Given a feature embedding $\\mathbf{z}$ in Euclidean space, this method first maps $\\mathbf{z}$ to the hyperbolic space at a local reference point, typically the center, using the exponential map (e.g., Eq. 4). Linear transformations with learnable weight matrix $\\mathbf{W}$ is then applied to $\\mathbf{z}$ within the hyperbolic space to obtain the hyperbolic representation. Then, a logarithmic map (e.g., Eq. 5) is applied to map the hyperbolic representation back to the Euclidean space. Formally, in the Poincaré ball model, this process is realized as:", + "bbox": [ + 89, + 582, + 483, + 777 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {W} \\mathbf {z} = \\log_ {\\mathbf {0}} ^ {\\mathbb {D}, c} \\left(\\mathbf {W} \\otimes_ {c} \\exp_ {\\mathbf {0}} ^ {\\mathbb {D}, c} (\\mathbf {z})\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 791, + 483, + 811 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\otimes_{c}$ is the mobius matrix multiplication operation. The primary aim of this method [64] is to capture more intricate hierarchical relationships within hyperbolic space. In contrast, our goal is to directly adjust the hyperbolic radius of CLIP's embeddings, leading to a scaling transformation.", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Hyperbolic Radius Adjustment", + "text_level": 1, + "bbox": [ + 511, + 311, + 790, + 328 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The objective of HyperCLIP is to introduce a scaling transformation to directly adjust the hyperbolic radii of CLIP's embeddings in hyperbolic space, thereby equipping CLIP with segmentation ability. Concretely, it necessitates the utilization of a diagonal matrix, denoted as $\\mathbf{S}$ , to scale a CLIP's feature embedding $\\mathbf{z} \\in \\mathbb{R}^{b \\times d}$ , where $d$ is the feature dimension of $\\mathbf{z}$ , $b$ is the spatial dimension of input images or number of classes. When extending this process to hyperbolic space, $\\mathbf{S}$ can directly adjust the hyperbolic radius of $\\mathbf{z}$ (this deviation can be found in Sec. 4.3). Specifically, we realize the diagonal matrix $\\mathbf{S}$ in a block-wise manner,", + "bbox": [ + 511, + 335, + 906, + 501 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} = \\operatorname {d i a g} \\left(\\mathbf {S} _ {1}, \\mathbf {S} _ {2}, \\dots , \\mathbf {S} _ {K}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 513, + 906, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{S}_k\\in \\mathbb{R}^{n\\times n}$ denotes a square matrix of the $k$ -th block, and the block number $K$ is calculated by $K = d / n$ . By adjusting the size of $n$ , we can achieve a transition between naive scaling transformation $(n = 1$ , i.e., modifying only the diagonal elements) and arbitrary transformation $(n = d$ , i.e., fully fine-tuning in hyperbolic space). The ablation study of $n$ is presented in Sec. 5.3.", + "bbox": [ + 511, + 541, + 906, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overall Fine-tuning. During the fine-tuning phase, HyperCLIP introduces a similar process to Eq. 6 is incorporated into CLIP's encoder. Specifically, HyperCLIP uses the following forward pass:", + "bbox": [ + 511, + 648, + 908, + 709 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} \\mathbf {z} = \\log_ {\\mathbf {0}} ^ {\\mathbb {D}, c} \\left(\\mathbf {S} \\otimes_ {c} \\exp_ {\\mathbf {0}} ^ {\\mathbb {D}, c} (\\mathbf {z})\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 719, + 906, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concurrently, the original components of CLIP load their weights from the pre-trained checkpoint, with their parameters remaining frozen.", + "bbox": [ + 511, + 750, + 906, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Parameter complexity. For an $K$ -block diagonal matrix, the number of parameters is $d^2 / K$ , resulting in a complexity of $\\mathcal{O}\\left(d^2 / K\\right)$ . We can optionally share the block matrix to reduce the number of parameters, i.e., $\\mathbf{S}_k = \\mathbf{S}_j, \\forall k \\neq j$ . This reduces the parameter complexity to $\\mathcal{O}\\left(d^2 / K^2\\right)$ . We can further reduce parameter complexity by setting $n = 1$ which transforms the block-diagonal matrix into a diagonal", + "bbox": [ + 511, + 795, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4565", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/7fb52ec309871fbf9aeae8a6597307907b625f5b4ff2fb4f73a06a7fa4df1564.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelVLMAdditional BackboneFine-tuning SpaceA-847PC-459A-150PC-59PAS-20\\( PAS-20^b \\)
Partial Fine-Tuning
ZS3Net [3]-ResNet-101E---19.438.3-
LSeg [31]CLIP ViT-B/32ResNet-101E----47.4-
OpenSeg [17]ALIGNResNet-101E4.47.917.540.1-63.8
ZegFormer [11]CLIP ViT-B/16ResNet-101E4.99.116.942.886.262.7
ZSseg [61]CLIP ViT-B/16ResNet-101E7.0-20.547.788.4-
OVSeg [37]CLIP ViT-B/16ResNet-101cE7.111.024.853.392.6-
ZegCLIP [69]CLIP ViT-B/16-E---41.293.6-
Selective Fine-Tuning
SED [59]CLIP ConvNeXt-B-E11.418.631.657.394.4-
CAT-Seg [9]CLIP ViT-B/16-E12.019.031.857.594.677.3
Parameter-efficient Fine-Tuning
SAN [62]CLIP ViT-B/16Side AdapterE10.112.627.553.894.0-
OursCLIP ViT-B/16-H12.319.232.158.595.678.9
Partial Fine-Tuning
LSeg [31]CLIP ViT-B/32ViT-L/16E----52.3-
OpenSeg [17]ALIGNEff-B7E8.111.526.444.8-70.2
OVSeg [37]CLIP ViT-L/14Swin-BE9.012.429.655.794.5-
SAN [62]CLIP ViT-L/14-E12.415.732.157.794.6-
ODISE [60]CLIP ViT-L/14Stable DiffusionE11.114.529.957.3--
FC-CLIP [65]CLIP ConvNeXt-L-E14.818.234.158.495.4-
Selective Fine-Tuning
SED [59]CLIP ConvNeXt-L-E13.922.635.260.696.1-
CAT-Seg [9]CLIP ViT-L/14-E16.023.837.963.397.082.5
Parameter-efficient Fine-Tuning
SAN [62]CLIP ViT-L/14Side AdapterE12.415.732.157.794.6-
OursCLIP ViT-L/14-H16.324.138.264.298.384.3
", + "bbox": [ + 91, + 89, + 926, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Comparison with state-of-the-art methods on standard benchmarks. The best-performing results are presented in bold, while the second-best results are underlined. \"E\": Euclidean Space. \"H\": Hyperbolic Space.", + "bbox": [ + 89, + 513, + 906, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "matrix. In this case, the number of parameters becomes $d$ , resulting in a complexity of $\\mathcal{O}(d)$ . Despite all these strategies to improve parameter efficiency, we note that the resulting matrix $\\mathbf{S}$ remains to realize scaling transformations, thereby allowing it to adjust the radius of CLIP's embeddings.", + "bbox": [ + 89, + 556, + 483, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Theoretical Analysis", + "text_level": 1, + "bbox": [ + 89, + 643, + 284, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further demonstrate the effectiveness of the proposed HyperCLIP method, we conduct a theoretical analysis of the scaling transformation. Theorem 1 reveals that the employed scaling matrix can directly adjust the hyperbolic radius, thus enabling transferring through different hierarchical levels. Detailed proof refers to the supplementary.", + "bbox": [ + 89, + 666, + 483, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 1 (Scaling). According to the definitions of the classical Poincaré Ball model and the hyperbolic tangent, for a point $\\mathbf{x} \\in \\mathbb{D}_c^n$ in hyperbolic space, its hyperbolic radius is defined as:", + "bbox": [ + 89, + 758, + 483, + 818 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {R a d} _ {\\mathbf {x}} := d _ {c} ^ {\\mathbb {D}} (\\mathbf {x}, \\mathbf {0}) = \\left(\\frac {2}{\\sqrt {c}}\\right) \\tanh ^ {- 1} \\left(\\sqrt {c} \\| \\mathbf {x} \\|\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 827, + 483, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{0}$ is the center of the hyperbolic space. Let $s$ is a scaling parameter, combine with Eq. 2 and Eq. 9, the radius", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of the point $\\mathbf{x}_s = \\mathcal{T}_{\\mathbf{x}}(s) := s \\otimes_c \\mathbf{x}$ which changed by the scaling transformation $\\mathcal{T}_{\\mathbf{x}}(s)$ is obtained by:", + "bbox": [ + 511, + 556, + 906, + 588 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {R a d} _ {\\mathbf {x}, s} (s) = \\frac {2}{\\sqrt {c}} \\tanh ^ {- 1} (\\sqrt {c} \\| \\mathbf {x} _ {s} \\|) \\\\ = \\frac {2}{\\sqrt {c}} \\tanh ^ {- 1} (\\sqrt {c} \\left| \\left| \\frac {\\tanh \\left(\\frac {s \\sqrt {c}}{2} \\operatorname {R a d} _ {\\mathbf {x}}\\right) \\mathbf {x}}{\\sqrt {c} \\| \\mathbf {x} \\|} \\right| \\right|. \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 598, + 906, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is noticed that the $\\mathrm{Rad}_{\\mathbf{x},s}(s)$ satisfies the monotonicity criteria, therefore, we can adjust the hyperbolic radius by changing the scaling parameter $s$ . Refer to the definition of the Möbius matrix multiplication operation in [27], the scaling matrix $\\mathbf{S}$ also has the ability to adjust the hyperbolic radius via the scaling transformation $\\mathcal{T}_{\\mathbf{x}}(\\mathbf{S})$ .", + "bbox": [ + 511, + 685, + 906, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 791, + 645, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 816, + 702, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Following prior works [9, 59], we use the COCO-Stuff dataset [4] as our training set. This dataset consists of around 118,000 densely annotated images, covering 171 distinct semantic categories. For inference, we compare our", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4566", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/324337a2a28aa44cbac3b98a1ec26cd585cb77f2559418d253825db387c8d8fa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsSpace for Fine-tuning\\( \\mathbf{A P^{Base}}_{50} \\)\\( \\mathbf{A P^{Novel}}_{50} \\)
CLIP-21.636.4
Other Fine-Tuning
CLIM [58]E25.742.5
CAT-Seg [9]E25.242.3
Parameter-efficient Fine-Tuning
LoRA [21]E24.441.5
HyperLoRA [64]H25.142.0
OursH26.342.7
", + "bbox": [ + 93, + 88, + 478, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "method against state-of-the-art approaches across several semantic segmentation benchmarks, including ADE20K [67], PASCAL VOC [14], and PASCAL-Context [40].", + "bbox": [ + 89, + 315, + 483, + 359 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- ADE20K [67] is a well-established semantic segmentation dataset with approximately 20,000 training images and 2,000 validation images. It also includes two separate test sets: A-150, consisting of 150 common categories, and A-847, which contains 847 categories.", + "- PASCAL VOC [14] is a smaller dataset for semantic segmentation, consisting of 1464 training images and 1449 validation images, covering 20 different foreground categories. We refer to this dataset as PAS-20. In line with [9], we also report scores on PAS-20 $^b$ , which includes the \"background\" as the 21st category.", + "- PASCAL-Context [40] is an extended version of the original PASCAL VOC dataset. It provides two test sets: PC-59 and PC-459, containing 59 and 459 categories respectively, for evaluation purposes." + ], + "bbox": [ + 89, + 361, + 483, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metric. Consistent with prior studies [9, 59], we use mean Intersection over Union (mIoU) to evaluate the semantic segmentation performance on these benchmarks.", + "bbox": [ + 89, + 589, + 483, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We implement our method using the Transformer-based CLIP model. Following the protocol established in [9], we evaluate our results on two versions of the CLIP model: ViT-B/16 and ViT-L/14. For training, we follow the same experimental setup as [9]. Besides, we use the Adam optimizer [29] with an initial learning rate of $1 \\times 10^{-5}$ for CLIP, and a weight decay of $10^{-4}$ . Training is conducted with one image per mini-batch. We apply the block-diagonal learnable weight in the self-attention module in each layer of CLIP. We employ the cost-based approach from [9] as our decoder. The curvature $c$ of Poincaré ball model is set to 0.01. All models are trained for 80,000 iterations using 4 NVIDIA A800 GPUs.", + "bbox": [ + 89, + 638, + 483, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparisons with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 89, + 845, + 460, + 862 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, we compare our proposed method with several state-of-the-art methods, as shown in Table 1, using six test sets", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c2da7c2fe3613642e2dc3724d396950cc2a84a5db63bf079730aac22d3811f11.jpg", + "table_caption": [ + "Table 2. Comparisons on open-vocabulary object detection task on the COCO [38] dataset. Results are measured by the box AP at IoU threshold 0.5. \"E\": Euclidean space. \"H\": hyperbolic space." + ], + "table_footnote": [], + "table_body": "
MethodODISE [60]FC-CLIP [65]MAFT+ [25]Ours
PQ23.426.827.129.2
", + "bbox": [ + 517, + 88, + 893, + 125 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3a336029fa794eee3636f82615c109d3e74c800c9e566bc404e32391641f4c42.jpg", + "table_caption": [ + "Table 3. Comparisons on open-vocabulary panoptic segmentation task on the ADE20K [67] dataset. \"PQ\": Panoptic quality." + ], + "table_footnote": [], + "table_body": "
nParam. (M)A-847PC-459A-150PC-59PAS-20b
10.045.99.427.452.474.9
80.3510.115.429.954.275.6
160.7011.417.431.055.976.2
321.4111.718.131.756.676.7
642.8111.918.732.057.877.8
1285.6312.319.232.158.578.9
25611.2512.219.031.958.879.3
", + "bbox": [ + 516, + 169, + 903, + 311 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/904e30cee89766e681aab0fb2396a7fd07f299c9e9a583cbba1ed8237bb1dada.jpg", + "table_caption": [ + "Table 4. Ablation on the size of diagonal blocks $n$ . The base model is ViT-B/16. The best-performing results are presented in bold, while the second-best results are underlined." + ], + "table_footnote": [], + "table_body": "
MethodsOVSeg [37]CAT-Seg [9]SAN [62]Ours
Param. (M)147.239.58.45.6
", + "bbox": [ + 531, + 377, + 883, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 5. Efficiency comparison in terms of learnable parameters for fine-tuning CLIP. The base model is ViT-B/16.", + "bbox": [ + 511, + 440, + 903, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "across three benchmarks. Overall, we achieve the best results. Most existing open-vocabulary semantic segmentation methods follow partial fine-tuning strategies, i.e., fine-tuning CLIP's image encoder. Although these methods provide sufficient flexibility for aligning with text descriptions generated by the text encoder, they fail to adjust the text embeddings to the appropriate hierarchical level for effective text-to-image alignment, often leading to suboptimal segmentation performance. Differently, CAT-Seg [9] simultaneously fine-tunes both the text encoder and image encoder of CLIP, achieving performance comparable to ours on some of the datasets. However, its fine-tuning scheme is manually controlled through different layer combinations, necessitating a careful design to balance generalization and segmentation ability, while ours does not suffer from such an issue. Then, compared to SAN [62], another parameter-efficient fine-tuning method that introduces only a limited number of tunable parameters, our approach significantly outperforms it, achieving improvements of $6.6\\%$ on the PC-459 dataset and $4.7\\%$ on the PC-59 dataset with ViT-B/16 as the base model. These results demonstrate the effectiveness of our method in preserving generalization while mastering segmentation capability.", + "bbox": [ + 511, + 491, + 906, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Efficiency comparison. We compare the efficiency of our method with other approaches, including OVSeg [37], CATSeg [9], and SAN [62], all of which utilize a ViT version of CLIP. The comparison, summarized in Table 5, shows", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4567", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d68b8761010aa08d10241ef82ece80946be3d996a1d0e0aae541bc1df5465765.jpg", + "image_caption": [ + "Figure 4. Comparison of qualitative results on ADE20K [67] with 847 categories. We compare our method with other two state-of-the-art methods, i.e., SAN [62] and CAT-Seg [9]. The results show our method performs more precise segmentation in different scenarios." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 906, + 393 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/024972f2c32ef18df7d932f78b74259fff08062197a1e9d35d974c050f025c6e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodImage encoderText encoderParam. (M)A-847PC-459A-150PC-59PAS-20PAS-20b
FreezeXX04.46.624.849.492.571.9
Parameter-efficient Fine-Tuning
LoRA [21]✓✓7.511.417.628.655.194.276.7
OFT [45]✓✓3.810.918.030.253.793.774.3
VPT [24]✓✓2.25.710.223.754.393.875.1
Adapter [19]✓✓7.510.416.528.854.994.275.2
SSF [36]✓✓0.66.915.228.652.193.272.8
HyperLoRA [64]✓✓7.511.418.229.755.894.377.2
HyperCLIP (Ours)✓X3.410.316.830.557.694.478.0
X✓2.211.117.529.855.193.776.1
✓✓5.612.319.232.158.595.678.9
", + "bbox": [ + 127, + 443, + 864, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Ablation study on the fine-tuning encoder in HyperCLIP and comparison of universal PEFT methods, including \"LoRA\" [21], \"OFT\" [45], \"VPT\" [24], \"Adapter\" [19], \"LST\" [49], \"SSF\" [36] and \"HyperLoRA\" [64] which is a recent PEFT method fine-tunes large language models in hyperbolic space. For a clear comparison, we ignore the segmentation decoder when calculating the number of learnable parameters. The base model is ViT-B/16. $\\checkmark$ : Fine-tuning. $\\times$ : Freezing.", + "bbox": [ + 88, + 662, + 906, + 719 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "that our method uses the fewest trainable parameters while equipping CLIP with segmentation ability. Since we do not introduce any extra architectures, and the logarithmic and exponential maps can be completed within linear complexity [64], the inference overhead is negligible.", + "bbox": [ + 88, + 744, + 480, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Application to other tasks. Here, we further carry out extra experiments in two other tasks. In an open-vocabulary object detection task, CLIP's image embeddings are transitioned from the image level to the object level. We follow a recent work [58] to investigate open-vocabulary object detection on", + "bbox": [ + 88, + 825, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the COCO dataset [38]. As shown in Table 2, the superior results demonstrate that HyperCLIP can effectively generalize to other tasks that require aligning text embeddings with embeddings from arbitrary hierarchical levels, e.g., object-level alignment. Furthermore, the open-vocabulary panoptic segmentation results presented in Table 3 further corroborate this conclusion.", + "bbox": [ + 511, + 744, + 906, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results. Here, we visualize our method's representative example segmentation results against prevailing methods, i.e., CAT-Seg [9] and SAN [62], in the A-847", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4568", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "dataset. As shown in Fig. 4, we can observe that our method can make more accurate predictions on object location and category. Even if objects are small and the background is complicated, our method still performs well.", + "bbox": [ + 89, + 90, + 482, + 151 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 164, + 243, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation of fine-tuning CLIP. We assess the efficacy of various fine-tuning design choices for HyperCLIP. As shown in Table 6, we freeze one of CLIP's encoders and compare the performance with several commonly-used PEFT methods, e.g., LoRA [21]. Compared to fully freezing CLIP's encoders (row 1), fine-tuning either encoder independently allows for minimizing its hyperbolic radius, thereby equipping the corresponding encoder with segmentation ability. However, one modality's hierarchical level remains at the image level, which hinders CLIP from fully acquiring segmentation ability. Consequently, jointly fine-tuning both encoders aligns the hierarchical levels at the pixel level, resulting in optimal performance. We also compare our method with existing PEFT strategies. Most of these methods, which operate in Euclidean space, struggle to capture the transition between different hierarchical levels. Notably, HyperLoRA [64] also updates parameters in hyperbolic space. However, because HyperLoRA does not specifically adjust the hyperbolic radius, it fails to fully endow CLIP with segmentation ability, resulting in inferior performance compared to our method.", + "bbox": [ + 89, + 188, + 485, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation of the size of diagonal blocks. The size of the diagonal block in our scaling transformations determines the number of parameters required to equip CLIP with segmentation ability. To investigate its impact, we conduct an ablation study on the scaling transformations by systematically varying the size of the diagonal block. As shown in Table 4, accuracy improves incrementally on datasets with fewer classes. However, on datasets with more classes (e.g., A-847), the accuracy performance reaches a saturation point around a block size of 128. This suggests that increasing the dimensionality of the diagonal block introduces more trainable parameters, which may lead to over-fitting on a closed set.", + "bbox": [ + 89, + 491, + 483, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of the hyperbolic radius of text embeddings. In order to closely examine the effect of the hyperbolic radius on equipping CLIP with segmentation ability, we apply a mean squared error (MSE) loss during fine-tuning. This MSE loss encourages the average hyperbolic radius of the text embeddings to converge towards different target values. As shown in Fig. 5, the initial value of the average hyperbolic radius is approximately 7.5, which is subsequently minimized to around 6.0 to achieve the optimal value for segmentation. Deviating from this value, either by increasing or decreasing the expected radius, results in diminished performance. This observation offers a new insight: the hyperbolic radius required for a vision task might be relatively fixed, suggesting that it could serve as a metric to quantify the granularity of different vision tasks.", + "bbox": [ + 89, + 674, + 485, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4a39143cac545910b33e6792cd51adfbe3402e02de4d989cc59e31d94370acc9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Möbius.A-847PC-459A-150PC-59PAS-20b
X11.118.230.257.377.5
✓12.319.232.158.578.9
", + "bbox": [ + 532, + 88, + 885, + 155 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Ablation on the necessity of Möbius matrix multiplication. \"Möbius\": Möbius matrix multiplication.", + "bbox": [ + 511, + 166, + 906, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/027af0d30e3bce7c9134c69adcd015dd10e90dbbd331a1282bc33de86be3bacc.jpg", + "image_caption": [ + "Figure 5. Visualization on the effect of different average hyperbolic radii of text embeddings. The controlling of different hyperbolic radii is achieved via a mean squared error (MSE) loss that enforces a constraint between the expected radius and the mean radius of the text embeddings during fine-tuning." + ], + "image_footnote": [], + "bbox": [ + 517, + 205, + 901, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Necessity of Möbius matrix multiplication. We ablate this experiment to validate the influence of the Möbius matrix multiplication operation. In Table 7, “ $\\times$ ” means that we use standard matrix multiplication. It achieves worse performance compared with “√”, which indicates that the Möbius matrix multiplication is essential for minimizing the hyperbolic radius of CLIP's embeddings.", + "bbox": [ + 511, + 512, + 908, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 632, + 633, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose HyperCLIP, a novel fine-tuning strategy for CLIP tailored to open-vocabulary semantic segmentation. Recognizing the hierarchical structure of images, we hypothesize that aligning the hierarchical level of CLIP's text encoder with pixel-level representations is crucial for accurate pixel-level predictions. HyperCLIP achieves this by adjusting the hierarchical level in hyperbolic space through scaling transformations applied to CLIP's embeddings, enabling precise cross-modal alignment. Extensive experiments demonstrate that HyperCLIP sets new state-of-the-art results in open-vocabulary semantic segmentation. More importantly, we observe that after adjustment, CLIP's text embeddings maintain a relatively fixed hyperbolic radius across datasets. This suggests that the segmentation task's granularity may be quantified using the hyperbolic radius, offering a promising direction for future research.", + "bbox": [ + 511, + 657, + 908, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4569", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the NSFC under Grant 62322604, 62176159 and 62401361, and in part by the Shanghai Municipal Science and Technology Major Project under Grant 2021SHZDX0102.", + "bbox": [ + 89, + 113, + 485, + 171 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 196, + 187, + 213 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Irving Biederman. Recognition-by-components: a theory of human image understanding. Psychological review, 94(2): 115, 1987. 2", + "[2] Martin R Bridson and André Haefliger. Metric spaces of non-positive curvature. Springer Science & Business Media, 2013. 3", + "[3] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. Advances in Neural Information Processing Systems, 32, 2019. 5", + "[4] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 2, 3, 5", + "[5] James W Cannon, William J Floyd, Richard Kenyon, Walter R Parry, et al. Hyperbolic geometry. Flavors of geometry, 31 (59-115):2, 1997. 2, 3", + "[6] Jiaxin Chen, Jie Qin, Yuming Shen, Li Liu, Fan Zhu, and Ling Shao. Learning attentive and hierarchical representations for 3d shape recognition. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 105-122. Springer, 2020. 3", + "[7] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 1", + "[8] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 3", + "[9] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Cat-seg: Cost aggregation for open-vocabulary semantic segmentation, 2024. 2, 3, 5, 6, 7", + "[10] Karan Desai, Maximilian Nickel, Tanmay Rajpurohit, Justin Johnson, and Shanmukha Ramakrishna Vedantam. Hyperbolic image-text representations. In International Conference on Machine Learning, pages 7694-7731. PMLR, 2023. 3", + "[11] Jian Ding, Nan Xue, Gui-Song Xia, and Dengxin Dai. Decoupling zero-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11583-11592, 2022. 3, 5", + "[12] Zi-Yi Dou, Aishwarya Kamath, Zhe Gan, Pengchuan Zhang, Jianfeng Wang, Linjie Li, Zicheng Liu, Ce Liu, Yann Le-Cun, Nanyun Peng, et al. Coarse-to-fine vision-language pre-training with fusion in the backbone. Advances in neural information processing systems, 35:32942-32956, 2022. 1" + ], + "bbox": [ + 93, + 222, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuohang Wang, Lijuan Wang, Chenguang Zhu, Pengchuan Zhang, Lu Yuan, Nanyun Peng, et al. An empirical study of training end-to-end vision-and-language transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18166-18176, 2022. 1", + "[14] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 6", + "[15] Zhe Gan, Yen-Chun Chen, Linjie Li, Chen Zhu, Yu Cheng, and Jingjing Liu. Large-scale adversarial training for vision-and-language representation learning. Advances in Neural Information Processing Systems, 33:6616-6628, 2020. 1", + "[16] Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. Advances in neural information processing systems, 31, 2018. 3", + "[17] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 3, 5", + "[18] Yunhui Guo, Xudong Wang, Yubei Chen, and Stella X Yu. Clipped hyperbolic classifiers are super-hyperbolic classifiers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11-20, 2022. 3", + "[19] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790-2799. PMLR, 2019. 7", + "[20] Chengyang Hu, Ke-Yue Zhang, Taiping Yao, Shouhong Ding, and Lizhuang Ma. Rethinking generalizable face anti-spoofing via hierarchical prototype-guided distribution refinement in hyperbolic space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1032-1041, 2024. 3", + "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 6, 7, 8", + "[22] Ming Hu, Kun Yuan, Yaling Shen, Feilong Tang, Xiaohao Xu, Lin Zhou, Wei Li, Ying Chen, Zhongxing Xu, Zelin Peng, et al. Ophclip: Hierarchical retrieval-augmented learning for ophthalmic surgical video-language pretraining. arXiv preprint arXiv:2411.15421, 2024. 1", + "[23] Zhicheng Huang, Zhaoyang Zeng, Yupan Huang, Bei Liu, Dongmei Fu, and Jianlong Fu. Seeing out of the box: End-to-end pre-training for vision-language representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12976-12985, 2021. 1", + "[24] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 7", + "[25] Siyu Jiao, Hongguang Zhu, Jiannan Huang, Yao Zhao, Yunchao Wei, and Humphrey Shi. Collaborative vision-text representation optimizing for open-vocabulary segmentation. In" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4570", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "European Conference on Computer Vision, pages 399-416. Springer, 2024. 6", + "[26] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 1", + "[27] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6418-6428, 2020. 3, 5", + "[28] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International conference on machine learning, pages 5583-5594. PMLR, 2021. 1", + "[29] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[30] Fanjie Kong, Yanbei Chen, Jiarui Cai, and Davide Modolo. Hyperbolic learning with synthetic captions for openworld detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16762-16771, 2024. 3", + "[31] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In International Conference on Learning Representations, 2022. 2, 5", + "[32] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 1", + "[33] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022.", + "[34] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022.", + "[35] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 121-137. Springer, 2020. 1", + "[36] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. Advances in Neural Information Processing Systems, 35:109-123, 2022. 7", + "[37] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 2, 3, 5, 6", + "[38] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6, 7", + "[39] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 1", + "[40] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 891–898, 2014. 6", + "[41] Maximillian Nickel and Douwe Kiela. Learning continuous hierarchies in the lorentz model of hyperbolic geometry. In International conference on machine learning, pages 3779-3788. PMLR, 2018. 3", + "[42] Zelin Peng, Zhengqin Xu, Zhilin Zeng, Yaoming Wang, Lingxi Xie, Qi Tian, and Wei Shen. Parameter-efficient fin-tuning in hyperspherical space for open-vocabulary semantic segmentation. arXiv preprint arXiv:2405.18840, 2024. 2", + "[43] Zelin Peng, Zhengqin Xu, Zhilin Zeng, Lingxi Xie, Qi Tian, and Wei Shen. Parameter efficient fine-tuning via cross block orchestration for segment anything model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3743-3752, 2024. 1", + "[44] Zelin Peng, Zhengqin Xu, Zhilin Zeng, Xiaokang Yang, and Wei Shen. Sam-parser: Fine-tuning sam efficiently by parameter space reconstruction. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4515-4523, 2024. 1", + "[45] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Scholkopf. Controlling text-to-image diffusion by orthogonal finetuning. Advances in Neural Information Processing Systems, 36:79320-79362, 2023. 7", + "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2", + "[47] Sameera Ramasinghe, Violetta Shevchenko, Gil Avraham, and Ajanthan Thalaiyasingam. Accept the modality gap: An exploration in the hyperbolic space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27263-27272, 2024. 2, 3", + "[48] Frederic Sala, Chris De Sa, Albert Gu, and Christopher Ré. Representation tradeoffs for hyperbolic embeddings. In International conference on machine learning, pages 4460-4469. PMLR, 2018. 2" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4571", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Lst: Ladder side-tuning for parameter and memory efficient transfer learning. Advances in Neural Information Processing Systems, 35: 12991-13005, 2022. 7", + "[50] Dídac Surís, Ruoshi Liu, and Carl Vondrick. Learning the predictability of the future. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12607-12617, 2021. 2", + "[51] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 1", + "[52] Feilong Tang, Zile Huang, Chengzhi Liu, Qiang Sun, Harry Yang, and Ser-Nam Lim. Intervening anchor token: Decoding strategy in alleviating hallucinations for mllms. In The Thirteenth International Conference on Learning Representations. 1", + "[53] Abraham A Ungar. Hyperbolic trigonometry and its application in the poincaré ball model of hyperbolic geometry. Computers & Mathematics with Applications, 41(1-2):135-147, 2001. 3", + "[54] Abraham Albert Ungar. A gyrovector space approach to hyperbolic geometry. Synthesis Lectures on Mathematics and Statistics, 1(1):1-194, 2008. 3", + "[55] Jianfeng Wang, Xiaowei Hu, Zhe Gan, Zhengyuan Yang, Xiyang Dai, Zicheng Liu, Yumao Lu, and Lijuan Wang. Ufo: A unified transformer for vision-language representation learning. arXiv preprint arXiv:2111.10023, 2021. 1", + "[56] Sijie Wang, Qiyu Kang, Rui She, Wei Wang, Kai Zhao, Yang Song, and Wee Peng Tay. Hypliloc: Towards effective lidar pose regression with hyperbolic fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5176-5185, 2023. 3", + "[57] Melanie Weber, Manzil Zaheer, Ankit Singh Rawat, Aditya K Menon, and Sanjiv Kumar. Robust large-margin learning in hyperbolic space. Advances in Neural Information Processing Systems, 33:17863-17873, 2020. 2", + "[58] Size Wu, Wenwei Zhang, Lumin Xu, Sheng Jin, Wentao Liu, and Chen Change Loy. Clim: Contrastive language-image mosaic for region representation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6117-6125, 2024. 6, 7", + "[59] Bin Xie, Jiale Cao, Jin Xie, Fahad Shahbaz Khan, and Yanwei Pang. Sed: A simple encoder-decoder for open-vocabulary semantic segmentation, 2023. 2, 5, 6", + "[60] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2955-2966, 2023. 3, 5, 6", + "[61] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 3, 5", + "[62] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xiang Bai. Side adapter network for open-vocabulary semantic segmentation. In Proceedings of the IEEE/CVF Conference" + ], + "bbox": [ + 91, + 90, + 485, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition, pages 2945-2954, 2023. 2, 3, 5, 6, 7", + "[63] Haochen Xue, Feilong Tang, Ming Hu, Yexin Liu, Qidong Huang, Yulong Li, Chengzhi Liu, Zhongxing Xu, Chong Zhang, Chun-Mei Feng, et al. Mmrc: A large-scale benchmark for understanding multimodal large language model in real-world conversation. arXiv preprint arXiv:2502.11903, 2025. 1", + "[64] Menglin Yang, Aosong Feng, Bo Xiong, Jihong Liu, Irwin King, and Rex Ying. Hyperbolic fine-tuning for large language models. arXiv preprint arXiv:2410.04010, 2024. 4, 6, 7, 8", + "[65] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. In NeurIPS, 2023. 2, 3, 5, 6", + "[66] Zhen Yu, Toan Nguyen, Yaniv Gal, Lie Ju, Shekhar S Chandra, Lei Zhang, Paul Bonnington, Victoria Mar, Zhiyong Wang, and Zongyuan Ge. Skin lesion recognition with class-hierarchy regularized hyperbolic embeddings. In International conference on medical image computing and computer-assisted intervention, pages 594-603. Springer, 2022. 3", + "[67] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127:302-321, 2019. 6, 7", + "[68] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712. Springer, 2022. 3", + "[69] Ziqin Zhou, Yinjie Lei, Bowen Zhang, Lingqiao Liu, and Yifan Liu. Zegclip: Towards adapting clip for zero-shot semantic segmentation. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5", + "[70] Yudong Zhu, Di Zhou, Jinghui Xiao, Xin Jiang, Xiao Chen, and Qun Liu. Hypertext: Endowing fasttext with hyperbolic geometry. arXiv preprint arXiv:2010.16143, 2020. 3" + ], + "bbox": [ + 516, + 90, + 906, + 614 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4572", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Understanding Fine-tuning CLIP for Open-vocabulary Semantic Segmentation in Hyperbolic Space/4933d0fc-2175-4f82-8204-a2ef1c424606_model.json b/2025/Understanding Fine-tuning CLIP for Open-vocabulary Semantic Segmentation in Hyperbolic Space/4933d0fc-2175-4f82-8204-a2ef1c424606_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2ab46b91e8adeca9b833c3efb216b7fe4ef9d771 --- /dev/null +++ b/2025/Understanding Fine-tuning CLIP for Open-vocabulary Semantic Segmentation in Hyperbolic Space/4933d0fc-2175-4f82-8204-a2ef1c424606_model.json @@ -0,0 +1,2334 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.13, + 0.908, + 0.178 + ], + "angle": 0, + "content": "Understanding Fine-tuning CLIP for Open-vocabulary Semantic Segmentation in Hyperbolic Space" + }, + { + "type": "text", + "bbox": [ + 0.209, + 0.203, + 0.788, + 0.24 + ], + "angle": 0, + "content": "Zelin Peng\\(^{1}\\), Zhengqin Xu\\(^{1}\\), Zhilin Zeng\\(^{1}\\), Changsong Wen\\(^{1}\\), Yu Huang\\(^{1}\\), Menglin Yang\\(^{2}\\), Feilong Tang\\(^{3,4}\\), and Wei Shen\\(^{1(\\boxtimes)}\\)" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.24, + 0.886, + 0.277 + ], + "angle": 0, + "content": "\\(^{1}\\)MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University \\(^{2}\\)Hong Kong University of Science and Technology (Guangzhou) \\(^{3}\\)MBZUAI, \\(^{4}\\)Monash University." + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.278, + 0.705, + 0.293 + ], + "angle": 0, + "content": "https://github.com/SJTU-DeepVisionLab/HyperCLIP" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.343, + 0.327, + 0.359 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.374, + 0.486, + 0.798 + ], + "angle": 0, + "content": "CLIP, a foundational vision-language model, has emerged as a powerful tool for open-vocabulary semantic segmentation. While freezing the text encoder preserves its powerful embeddings, recent studies show that fine-tuning both the text and image encoders jointly significantly enhances segmentation performance, especially for classes from open sets. In this work, we explain this phenomenon from the perspective of hierarchical alignment, since during fine-tuning, the hierarchy level of image embeddings shifts from image-level to pixel-level. We achieve this by leveraging hyperbolic space, which naturally encoders hierarchical structures. Our key observation is that, during fine-tuning, the hyperbolic radius of CLIP's text embeddings decreases, facilitating better alignment with the pixel-level hierarchical structure of visual data. Building on this insight, we propose HyperCLIP, a novel fine-tuning strategy that adjusts the hyperbolic radius of the text embeddings through scaling transformations. By doing so, HyperCLIP equips CLIP with segmentation capability while introducing only a small number of learnable parameters. Our experiments demonstrate that HyperCLIP achieves state-of-the-art performance on open-vocabulary semantic segmentation tasks across three benchmarks, while fine-tuning only approximately \\(4\\%\\) of the total parameters of CLIP. More importantly, we observe that after adjustment, CLIP's text embeddings exhibit a relatively fixed hyperbolic radius across datasets, suggesting that the granularity required for this segmentation task might be quantified using the hyperbolic radius." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.824, + 0.222, + 0.84 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.85, + 0.483, + 0.881 + ], + "angle": 0, + "content": "The goal of open-vocabulary semantic segmentation is to develop a segmentation model capable of labeling each" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.33, + 0.895, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.63, + 0.909, + 0.785 + ], + "angle": 0, + "content": "Figure 1. Visualization of hyperbolic space by a Poincaré ball model: The height of points reflects the hierarchical level of input data. The black segments refer to the shortest path between the points in the Poincaré ball model. The red arrow line denotes the hyperbolic radius of text embeddings. Embeddings closer to the center have a smaller radius, representing low-level visual information (e.g., pixels), while points farther from the center have a larger radius, representing high-level visual information (e.g., images). During fine-tuning, the hyperbolic radius of text embeddings is adjusted to match the radius associated with pixels, thereby endowing CLIP with segmentation ability (Best viewed in color)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.912, + 0.903 + ], + "angle": 0, + "content": "pixel in an image with categories that extend beyond a predefined closed set, based on textual descriptions. Vision-language foundation models [7, 12, 13, 15, 22, 23, 26, 28, 32–35, 39, 43, 44, 46, 51, 52, 55, 63], particularly CLIP [46], are frequently employed to provide open-vocabulary recognition capabilities. Consequently, open-vocabulary seman" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.887, + 0.415, + 0.901 + ], + "angle": 0, + "content": "Corresponding Author: wei.shen@sjtu.edu.cn" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4562" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.138 + ], + "angle": 0, + "content": "tic segmentation fundamentally involves transferring these vision-language foundation models, originally trained with image-level supervision, to achieve pixel-level predictions." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.14, + 0.487, + 0.457 + ], + "angle": 0, + "content": "Current methods [9, 59, 62, 65] typically fine-tune CLIP on a closed set with segmentation annotations, i.e., COCO [4], to equip it with the segmentation capability. A prevailing opinion is that simply freezing CLIP's text encoder maximally preserves the powerful text embeddings of a vast array of classes, which is believed to be beneficial for good generalization on open sets. Many studies follow this opinion [31, 37, 62]. However, recent studies [9, 42, 59] offer an alternative perspective on fine-tuning: simultaneously fine-tuning both the image and text encoders results in superior segmentation performance on open sets compared to former approaches. To explain and resolve this seemingly paradoxical phenomenon, we first leverage a non-Euclidean manifold—specifically, hyperbolic space [5]—to understand why fine-tuning the text encoder appears to improve segmentation performance on open sets. Building on this understanding, our objective is to design an explainable fine-tuning strategy with a minimal number of introduced tunable parameters, which is able to equip CLIP with strong segmentation performance while minimizing the loss of its generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.46, + 0.486, + 0.793 + ], + "angle": 0, + "content": "Existing research [1] has found that images inherently exhibit hierarchical structures, consisting of multiple levels: pixels, patches, objects, entire scenes, etc. Given that, in open-vocabulary semantic segmentation, CLIP's image embeddings transition from image level to pixel level, we infer that the performance improvement brought by the recent studies [9, 59] could be attributed to adjusting CLIP's text embeddings from their original hierarchical level to better align with the level represented by pixels, thereby enhancing cross-modal alignment. However, most existing fine-tuning methods operate in Euclidean space, which makes it challenging to quantify such a transition through different hierarchical levels. Hyperbolic spaces have gained significant interest in recent years, owing to their ability to naturally and compactly encode hierarchical structures [47, 48, 57], as shown in Fig. 1. The hyperbolic radius, defined as the distance from a point to the center in hyperbolic space, represents its hierarchical level [50]. Points closer to the center (smaller radius) correspond to low-level visual embeddings, e.g., those from pixels, while points near the edge (larger radius) correspond to high-level visual embeddings, e.g., those from images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Motivated by this, we project the text embeddings into hyperbolic space and visualize the changes in their radii before and after fine-tuning, as shown in Fig. 2. This fine-tuning is realized using the state-of-the-art method provided in [9]. Specifically, one can observe that the radius after fine-tuning is smaller than that of pre-trained CLIP. In particular, we identify a pattern during fine-tuning: the hierarchical level of" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.086, + 0.907, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.276, + 0.907, + 0.361 + ], + "angle": 0, + "content": "Figure 2. Understanding fine-tuning CLIP's text encoder from the perspective of hyperbolic radius. We use hyperbolic radius to illustrate the change in the hierarchical level of 171 classes after fine-tuning. We observe that the hyperbolic radius of each class becomes smaller, suggesting that adjusting the hyperbolic radius equips CLIP with segmentation ability (Best viewed in color)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.385, + 0.907, + 0.461 + ], + "angle": 0, + "content": "the text embeddings adjusts in conjunction with the hierarchical level of the embeddings provided by the image encoder. Our finding brings a novel perspective for fine-tuning CLIP: Adjusting the hyperbolic radius of text embeddings from edge to center endowed CLIP with segmentation ability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.909, + 0.674 + ], + "angle": 0, + "content": "Building on the analysis above, we propose a novel finetuning strategy for CLIP in hyperbolic space, termed HyperCLIP. HyperCLIP aims to directly adjust the hyperbolic radius of CLIP's embeddings to facilitate alignment with the appropriate hierarchical level for segmentation tasks. Specifically, we introduce several block-diagonal scaling matrices. Multiplying these matrices with CLIP's feature representations using a Möbius matrix multiplication operation is equivalent to performing scaling transformations that adjust the radius of CLIP's embeddings. This enables the adjustment of their hierarchical levels, ultimately enhancing the model's performance. As a result, HyperCLIP can equip CLIP with segmentation ability with only a small number of introduced learnable parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.911, + 0.81 + ], + "angle": 0, + "content": "Extensive experiments demonstrate that our method sets new state-of-the-art results in open-vocabulary semantic segmentation across three benchmarks, while fine-tuning only approximately \\(4\\%\\) of the total parameters of CLIP. More importantly, we observe that after adjustment, CLIP's text embeddings maintain a relatively fixed hyperbolic radius across different datasets. This suggests that the granularity of the segmentation task might be quantified using the hyperbolic radius, warranting further exploration." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.822, + 0.655, + 0.838 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.868, + 0.864 + ], + "angle": 0, + "content": "2.1. Open-vocabulary Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Prior works on open-vocabulary semantic segmentation typically tackle this task by leveraging CLIP [46]. Early efforts," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4563" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.487, + 0.455 + ], + "angle": 0, + "content": "e.g., [68], directly fine-tune CLIP on mainstream segmentation datasets, like COCO [4]. However, these approaches argue that fully fine-tuning CLIP's encoder significantly diminishes its ability to generalize to unseen classes. To mitigate this issue, some methods [11, 17, 60, 61] take the opposite approach, fine-tuning an additional mask generator [8] for segmentation while keeping CLIP frozen to preserve its generalization capabilities. However, this approach, with its frozen parameter space, lacks segmentation awareness, leading to a misalignment between regions and text descriptions [37]. Other studies [9, 62, 65] propose more sophisticated solutions that fine-tune only specific parameters, e.g., certain layers of CLIP, enabling pixel-level predictions while keeping the majority of CLIP's parameters fixed to minimize the loss of generalization. While these approaches show significant advantages, they often rely on very small learning rates, implicitly promoting minimal deviations from the pre-trained CLIP, which limits segmentation performance. In summary, there is still limited understanding of how fine-tuning influences CLIP's representations and why fine-tuning CLIP's text encoder appears to enhance segmentation performance. In this paper, we aim to provide several new insights that reveal the underlying logic by incorporating a novel perspective: hyperbolic space." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.465, + 0.365, + 0.482 + ], + "angle": 0, + "content": "2.2. Learning in Hyperbolic Spaces" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.487, + 0.732 + ], + "angle": 0, + "content": "Unlike Euclidean geometry, hyperbolic space can be viewed as the continuous analog of a tree [2], enabling a hierarchical understanding of knowledge structures. Previous research has demonstrated the advantages of representing such hierarchical relationships in various domains, including molecular structures [66], 3D data [6, 56], text data [16, 70], and images [18, 27]. Furthermore, several studies have investigated hyperbolic learning in vision tasks to model hierarchical relationships across different modalities [10, 30, 47], showcasing the benefits of capturing the semantic hierarchy among various data types. Nevertheless, while most studies typically leverage the intrinsic hierarchical properties of hyperbolic space, this paper seeks to capture the detailed changes in semantics reflected by the hyperbolic radius. This approach potentially complements current fine-tuning efforts in open-vocabulary tasks by offering a novel hyperbolic perspective." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.746, + 0.325, + 0.763 + ], + "angle": 0, + "content": "3. Preliminary Background" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.291, + 0.789 + ], + "angle": 0, + "content": "3.1. Hyperbolic geometry" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.487, + 0.903 + ], + "angle": 0, + "content": "Hyperbolic geometry often has constant negative curvature and is fundamentally different to the Euclidean or spherical geometry which has zero or constant positive curvature, respectively. This enables unique properties such as the divergence of parallel lines and the exponential volume growth towards the boundary. Hyperbolic geometry has five different analytic models to build the hyperbolic space [5]. In this" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.123 + ], + "angle": 0, + "content": "work, we utilize the classical Poincaré ball model as its effectiveness in capturing hierarchical structures [16, 20, 41, 47]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.134, + 0.742, + 0.149 + ], + "angle": 0, + "content": "3.2. The Poincaré Ball Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.157, + 0.909, + 0.388 + ], + "angle": 0, + "content": "The Poincaré ball model \\((\\mathbb{D}_c^n,g^{\\mathbb{D}_c})\\) with radius \\(1 / \\sqrt{c}\\) and negative curvature \\(-c(c > 0)\\) is a Riemannian manifold \\(\\mathbb{D}_c^n \\coloneqq \\{\\mathbf{x} \\in \\mathbb{R}^n : c\\|\\mathbf{x}\\| < 1\\}\\) with the metric \\(g^{\\mathbb{D}_c} = \\lambda_{c,\\mathbf{x}}^2 g^E\\), where \\(\\lambda_{c,\\mathbf{x}} = \\frac{2}{1 - c\\|\\mathbf{x}\\|^2}\\) is the conformal factor, and \\(g^E = \\mathbf{I}_n\\) is the Euclidean metric tensor. The vectorial structure operations defined in Euclidean space can not be directly used in hyperbolic space due to the differential of the space properties. Therefore, we introduce the framework of gyrovector spaces [53, 54] to provide non-associative algebraic operations for hyperbolic space. According to special relativity, when adding speed vectors belonging to the Poincaré ball of radius \\(c\\), these added speed vectors remain in the ball by the gyrovector spaces operation, also known as the Möbius operation. For instance, the addition and multiplication operations are defined as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.392, + 0.907, + 0.422 + ], + "angle": 0, + "content": "Mobius addition operation. For \\(\\mathbf{x},\\mathbf{y}\\in \\mathbb{D}_c^n\\) , the Mobius addition operation is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.434, + 0.907, + 0.469 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} \\oplus_ {c} \\mathbf {y} := \\frac {(1 + 2 c \\langle \\mathbf {x} , \\mathbf {y} \\rangle + c \\| \\mathbf {y} \\| ^ {2}) \\mathbf {x} + (1 - c \\| \\mathbf {x} \\| ^ {2}) \\mathbf {y}}{1 + 2 c \\langle \\mathbf {x} , \\mathbf {y} \\rangle + c ^ {2} \\| \\mathbf {x} \\| ^ {2} \\| \\mathbf {y} \\| ^ {2}}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.485, + 0.909, + 0.53 + ], + "angle": 0, + "content": "Mobius scalar multiplication operation. For a scalar \\(r \\in \\mathbb{R}\\) and a vector \\(\\mathbf{x} \\in \\mathbb{D}_c^n\\), the Möbius scalar multiplication operation is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.542, + 0.907, + 0.572 + ], + "angle": 0, + "content": "\\[\nr \\otimes_ {c} \\mathbf {x} := (1 / \\sqrt {c}) \\tanh \\left(r \\tanh ^ {- 1} (\\sqrt {c} \\| \\mathbf {x} \\|\\right)) \\frac {\\mathbf {x}}{\\| \\mathbf {x} \\|}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.584, + 0.909, + 0.631 + ], + "angle": 0, + "content": "The operation satisfies \\(r \\otimes_{c} \\mathbf{0} \\coloneqq \\mathbf{0}\\), \\(r \\otimes_{c} \\mathbf{x} = \\mathbf{x} \\oplus_{c} \\dots \\oplus_{c} \\mathbf{x}\\) (\\(r\\) additions), \\((r' + r) \\otimes_{c} \\mathbf{x} = r \\otimes \\mathbf{x} \\oplus_{c} r' \\otimes_{c} \\mathbf{x}\\), \\((r'r) \\otimes_{c} \\mathbf{x} = r \\otimes_{c} (r' \\otimes_{c} \\mathbf{x})\\) and \\(\\left(\\left|r\\right| \\otimes_{c} \\mathbf{x}\\right) / \\left\\|r \\otimes_{c} \\mathbf{x}\\right\\| = \\mathbf{x} / \\left\\|\\mathbf{x}\\right\\|\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.633, + 0.909, + 0.694 + ], + "angle": 0, + "content": "Mobius matrix multiplication operation. Refer to the relation definition in [27], for a matrix \\(\\mathbf{M} \\in \\mathbb{R}^{n \\times n}\\) and a vector \\(\\mathbf{x} \\in \\mathbb{D}_c^n\\), if \\(\\mathbf{Mx} \\neq \\mathbf{0}\\), the Möbius matrix multiplication operation is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.705, + 0.902, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} \\otimes_ {c} \\mathbf {x} := (\\frac {1}{\\sqrt {c}}) \\tanh \\left(\\frac {\\| \\mathbf {M x} \\|}{\\| \\mathbf {x} \\|} \\tanh ^ {- 1} (\\sqrt {c} \\| \\mathbf {x} \\|)\\right) \\frac {\\mathbf {M x}}{\\| \\mathbf {M x} \\|}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.751, + 0.781, + 0.767 + ], + "angle": 0, + "content": "The distance of \\(\\mathbf{x},\\mathbf{y}\\in \\mathbb{D}_c^n\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.778, + 0.907, + 0.797 + ], + "angle": 0, + "content": "\\[\nd _ {c} ^ {\\mathbb {D}} (\\mathbf {x}, \\mathbf {y}) = (2 / \\sqrt {c}) \\tanh ^ {- 1} (\\sqrt {c} \\| - \\mathbf {x} \\oplus_ {c} \\mathbf {y} \\Vert). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The tangent space \\(\\mathcal{T}_{\\mathbf{x}}^{c}\\mathbb{D}_{c}^{n}\\) at a point \\(\\mathbf{x}\\in \\mathbb{D}_c^n\\) is the first order approximation of \\(\\mathbb{D}_c^n\\) , which is an \\(n\\) -dimensional Euclidean space. The tangent space \\(\\mathcal{T}_{\\mathbf{x}}^{c}\\mathbb{D}_{c}^{n}\\) and \\(\\mathbb{D}_c^n\\) are mapped to each other by exponential \\((\\mathcal{T}_{\\mathbf{x}}^{c}\\mathbb{D}_{c}^{n}\\mapsto \\mathbb{D}_{c}^{n}:\\exp_{\\mathbf{x}}^{\\mathbb{D},c}(\\cdot))\\) and logarithmic \\((\\mathbb{D}_c^n\\mapsto \\mathcal{T}_x^c\\mathbb{D}_c^n:\\log_{\\mathbf{x}}^{\\mathbb{D},c}(\\cdot))\\) maps, respectively. For any \\(\\mathbf{x},\\mathbf{y}\\in \\mathbb{D}_c^n\\) and \\(\\mathbf{v}\\in \\mathcal{T}_{\\mathbf{x}}^{c}\\mathbb{D}_{c}^{n}\\) , the mapping functions" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4564" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.904, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.258, + 0.908, + 0.288 + ], + "angle": 0, + "content": "Figure 3. A schematic representation of HyperCLIP. In the HyperCLIP framework, we fine-tune CLIP in hyperbolic space by minimizing its embeddings' hyperbolic radii using several tunable block-diagonal scaling matrices." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.313, + 0.325, + 0.33 + ], + "angle": 0, + "content": "are given for \\(\\mathbf{v} \\neq \\mathbf{0}\\) and \\(\\mathbf{y} \\neq \\mathbf{x}\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.342, + 0.484, + 0.376 + ], + "angle": 0, + "content": "\\[\n\\exp_ {\\mathbf {x}} ^ {\\mathbb {D}, c} (\\mathbf {v}) = \\mathbf {x} \\oplus_ {c} \\left(\\tanh \\left(\\sqrt {c} \\frac {\\lambda_ {c , \\mathbf {x}} \\| \\mathbf {v} \\|}{2}\\right) \\frac {\\mathbf {v}}{\\sqrt {c} \\| \\mathbf {v} \\|}\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.378, + 0.493, + 0.425 + ], + "angle": 0, + "content": "\\[\n\\log_ {\\mathbf {x}} ^ {\\mathbb {D}, c} (\\mathbf {y}) = \\frac {2}{\\sqrt {c} \\lambda_ {c , \\mathbf {x}}} \\tanh ^ {- 1} \\left(\\sqrt {c} \\| - \\mathbf {x} \\oplus_ {c} \\mathbf {y} \\|\\right) \\frac {- \\mathbf {x} \\oplus_ {c} \\mathbf {y}}{\\| - \\mathbf {x} \\oplus_ {c} \\mathbf {y} \\|}. \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.441, + 0.226, + 0.459 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.468, + 0.484, + 0.545 + ], + "angle": 0, + "content": "In this section, we introduce HyperCLIP for open-vocabulary semantic segmentation. HyperCLIP is based on the Poincaré ball model, which aims to adjust the hyperbolic radius of CLIP's embeddings, thereby equipping CLIP with segmentation ability." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.558, + 0.377, + 0.575 + ], + "angle": 0, + "content": "4.1. Fine-tuning in Hyperbolic Space" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.484, + 0.779 + ], + "angle": 0, + "content": "The core technique in HyperCLIP involves linear transformations, i.e., scaling, in the hyperbolic space. One of the primary methods [64] to learn the weights of linear transformations in hyperbolic space is hyperLoRA [64]. Given a feature embedding \\(\\mathbf{z}\\) in Euclidean space, this method first maps \\(\\mathbf{z}\\) to the hyperbolic space at a local reference point, typically the center, using the exponential map (e.g., Eq. 4). Linear transformations with learnable weight matrix \\(\\mathbf{W}\\) is then applied to \\(\\mathbf{z}\\) within the hyperbolic space to obtain the hyperbolic representation. Then, a logarithmic map (e.g., Eq. 5) is applied to map the hyperbolic representation back to the Euclidean space. Formally, in the Poincaré ball model, this process is realized as:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.792, + 0.484, + 0.813 + ], + "angle": 0, + "content": "\\[\n\\mathbf {W} \\mathbf {z} = \\log_ {\\mathbf {0}} ^ {\\mathbb {D}, c} \\left(\\mathbf {W} \\otimes_ {c} \\exp_ {\\mathbf {0}} ^ {\\mathbb {D}, c} (\\mathbf {z})\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.485, + 0.903 + ], + "angle": 0, + "content": "where \\(\\otimes_{c}\\) is the mobius matrix multiplication operation. The primary aim of this method [64] is to capture more intricate hierarchical relationships within hyperbolic space. In contrast, our goal is to directly adjust the hyperbolic radius of CLIP's embeddings, leading to a scaling transformation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.312, + 0.791, + 0.329 + ], + "angle": 0, + "content": "4.2. Hyperbolic Radius Adjustment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.336, + 0.907, + 0.502 + ], + "angle": 0, + "content": "The objective of HyperCLIP is to introduce a scaling transformation to directly adjust the hyperbolic radii of CLIP's embeddings in hyperbolic space, thereby equipping CLIP with segmentation ability. Concretely, it necessitates the utilization of a diagonal matrix, denoted as \\(\\mathbf{S}\\), to scale a CLIP's feature embedding \\(\\mathbf{z} \\in \\mathbb{R}^{b \\times d}\\), where \\(d\\) is the feature dimension of \\(\\mathbf{z}\\), \\(b\\) is the spatial dimension of input images or number of classes. When extending this process to hyperbolic space, \\(\\mathbf{S}\\) can directly adjust the hyperbolic radius of \\(\\mathbf{z}\\) (this deviation can be found in Sec. 4.3). Specifically, we realize the diagonal matrix \\(\\mathbf{S}\\) in a block-wise manner," + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.515, + 0.907, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} = \\operatorname {d i a g} \\left(\\mathbf {S} _ {1}, \\mathbf {S} _ {2}, \\dots , \\mathbf {S} _ {K}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.542, + 0.907, + 0.648 + ], + "angle": 0, + "content": "where \\(\\mathbf{S}_k\\in \\mathbb{R}^{n\\times n}\\) denotes a square matrix of the \\(k\\) -th block, and the block number \\(K\\) is calculated by \\(K = d / n\\) . By adjusting the size of \\(n\\) , we can achieve a transition between naive scaling transformation \\((n = 1\\) , i.e., modifying only the diagonal elements) and arbitrary transformation \\((n = d\\) , i.e., fully fine-tuning in hyperbolic space). The ablation study of \\(n\\) is presented in Sec. 5.3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.649, + 0.909, + 0.71 + ], + "angle": 0, + "content": "Overall Fine-tuning. During the fine-tuning phase, HyperCLIP introduces a similar process to Eq. 6 is incorporated into CLIP's encoder. Specifically, HyperCLIP uses the following forward pass:" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.72, + 0.907, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} \\mathbf {z} = \\log_ {\\mathbf {0}} ^ {\\mathbb {D}, c} \\left(\\mathbf {S} \\otimes_ {c} \\exp_ {\\mathbf {0}} ^ {\\mathbb {D}, c} (\\mathbf {z})\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.908, + 0.795 + ], + "angle": 0, + "content": "Concurrently, the original components of CLIP load their weights from the pre-trained checkpoint, with their parameters remaining frozen." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Parameter complexity. For an \\(K\\)-block diagonal matrix, the number of parameters is \\(d^2 / K\\), resulting in a complexity of \\(\\mathcal{O}\\left(d^2 / K\\right)\\). We can optionally share the block matrix to reduce the number of parameters, i.e., \\(\\mathbf{S}_k = \\mathbf{S}_j, \\forall k \\neq j\\). This reduces the parameter complexity to \\(\\mathcal{O}\\left(d^2 / K^2\\right)\\). We can further reduce parameter complexity by setting \\(n = 1\\) which transforms the block-diagonal matrix into a diagonal" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4565" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.09, + 0.927, + 0.504 + ], + "angle": 0, + "content": "
ModelVLMAdditional BackboneFine-tuning SpaceA-847PC-459A-150PC-59PAS-20\\( PAS-20^b \\)
Partial Fine-Tuning
ZS3Net [3]-ResNet-101E---19.438.3-
LSeg [31]CLIP ViT-B/32ResNet-101E----47.4-
OpenSeg [17]ALIGNResNet-101E4.47.917.540.1-63.8
ZegFormer [11]CLIP ViT-B/16ResNet-101E4.99.116.942.886.262.7
ZSseg [61]CLIP ViT-B/16ResNet-101E7.0-20.547.788.4-
OVSeg [37]CLIP ViT-B/16ResNet-101cE7.111.024.853.392.6-
ZegCLIP [69]CLIP ViT-B/16-E---41.293.6-
Selective Fine-Tuning
SED [59]CLIP ConvNeXt-B-E11.418.631.657.394.4-
CAT-Seg [9]CLIP ViT-B/16-E12.019.031.857.594.677.3
Parameter-efficient Fine-Tuning
SAN [62]CLIP ViT-B/16Side AdapterE10.112.627.553.894.0-
OursCLIP ViT-B/16-H12.319.232.158.595.678.9
Partial Fine-Tuning
LSeg [31]CLIP ViT-B/32ViT-L/16E----52.3-
OpenSeg [17]ALIGNEff-B7E8.111.526.444.8-70.2
OVSeg [37]CLIP ViT-L/14Swin-BE9.012.429.655.794.5-
SAN [62]CLIP ViT-L/14-E12.415.732.157.794.6-
ODISE [60]CLIP ViT-L/14Stable DiffusionE11.114.529.957.3--
FC-CLIP [65]CLIP ConvNeXt-L-E14.818.234.158.495.4-
Selective Fine-Tuning
SED [59]CLIP ConvNeXt-L-E13.922.635.260.696.1-
CAT-Seg [9]CLIP ViT-L/14-E16.023.837.963.397.082.5
Parameter-efficient Fine-Tuning
SAN [62]CLIP ViT-L/14Side AdapterE12.415.732.157.794.6-
OursCLIP ViT-L/14-H16.324.138.264.298.384.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.515, + 0.907, + 0.543 + ], + "angle": 0, + "content": "Table 1. Comparison with state-of-the-art methods on standard benchmarks. The best-performing results are presented in bold, while the second-best results are underlined. \"E\": Euclidean Space. \"H\": Hyperbolic Space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.557, + 0.484, + 0.633 + ], + "angle": 0, + "content": "matrix. In this case, the number of parameters becomes \\(d\\), resulting in a complexity of \\(\\mathcal{O}(d)\\). Despite all these strategies to improve parameter efficiency, we note that the resulting matrix \\(\\mathbf{S}\\) remains to realize scaling transformations, thereby allowing it to adjust the radius of CLIP's embeddings." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.644, + 0.285, + 0.658 + ], + "angle": 0, + "content": "4.3. Theoretical Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.667, + 0.484, + 0.758 + ], + "angle": 0, + "content": "To further demonstrate the effectiveness of the proposed HyperCLIP method, we conduct a theoretical analysis of the scaling transformation. Theorem 1 reveals that the employed scaling matrix can directly adjust the hyperbolic radius, thus enabling transferring through different hierarchical levels. Detailed proof refers to the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.759, + 0.484, + 0.819 + ], + "angle": 0, + "content": "Theorem 1 (Scaling). According to the definitions of the classical Poincaré Ball model and the hyperbolic tangent, for a point \\(\\mathbf{x} \\in \\mathbb{D}_c^n\\) in hyperbolic space, its hyperbolic radius is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.828, + 0.484, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\operatorname {R a d} _ {\\mathbf {x}} := d _ {c} ^ {\\mathbb {D}} (\\mathbf {x}, \\mathbf {0}) = \\left(\\frac {2}{\\sqrt {c}}\\right) \\tanh ^ {- 1} \\left(\\sqrt {c} \\| \\mathbf {x} \\|\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{0}\\) is the center of the hyperbolic space. Let \\(s\\) is a scaling parameter, combine with Eq. 2 and Eq. 9, the radius" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.557, + 0.907, + 0.589 + ], + "angle": 0, + "content": "of the point \\(\\mathbf{x}_s = \\mathcal{T}_{\\mathbf{x}}(s) := s \\otimes_c \\mathbf{x}\\) which changed by the scaling transformation \\(\\mathcal{T}_{\\mathbf{x}}(s)\\) is obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.599, + 0.907, + 0.674 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {R a d} _ {\\mathbf {x}, s} (s) = \\frac {2}{\\sqrt {c}} \\tanh ^ {- 1} (\\sqrt {c} \\| \\mathbf {x} _ {s} \\|) \\\\ = \\frac {2}{\\sqrt {c}} \\tanh ^ {- 1} (\\sqrt {c} \\left| \\left| \\frac {\\tanh \\left(\\frac {s \\sqrt {c}}{2} \\operatorname {R a d} _ {\\mathbf {x}}\\right) \\mathbf {x}}{\\sqrt {c} \\| \\mathbf {x} \\|} \\right| \\right|. \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.907, + 0.777 + ], + "angle": 0, + "content": "It is noticed that the \\(\\mathrm{Rad}_{\\mathbf{x},s}(s)\\) satisfies the monotonicity criteria, therefore, we can adjust the hyperbolic radius by changing the scaling parameter \\(s\\). Refer to the definition of the Möbius matrix multiplication operation in [27], the scaling matrix \\(\\mathbf{S}\\) also has the ability to adjust the hyperbolic radius via the scaling transformation \\(\\mathcal{T}_{\\mathbf{x}}(\\mathbf{S})\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.792, + 0.646, + 0.809 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.817, + 0.704, + 0.833 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Datasets. Following prior works [9, 59], we use the COCO-Stuff dataset [4] as our training set. This dataset consists of around 118,000 densely annotated images, covering 171 distinct semantic categories. For inference, we compare our" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4566" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.48, + 0.246 + ], + "angle": 0, + "content": "
MethodsSpace for Fine-tuning\\( \\mathbf{A P^{Base}}_{50} \\)\\( \\mathbf{A P^{Novel}}_{50} \\)
CLIP-21.636.4
Other Fine-Tuning
CLIM [58]E25.742.5
CAT-Seg [9]E25.242.3
Parameter-efficient Fine-Tuning
LoRA [21]E24.441.5
HyperLoRA [64]H25.142.0
OursH26.342.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.255, + 0.483, + 0.298 + ], + "angle": 0, + "content": "Table 2. Comparisons on open-vocabulary object detection task on the COCO [38] dataset. Results are measured by the box AP at IoU threshold 0.5. \"E\": Euclidean space. \"H\": hyperbolic space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.316, + 0.484, + 0.361 + ], + "angle": 0, + "content": "method against state-of-the-art approaches across several semantic segmentation benchmarks, including ADE20K [67], PASCAL VOC [14], and PASCAL-Context [40]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.362, + 0.483, + 0.437 + ], + "angle": 0, + "content": "- ADE20K [67] is a well-established semantic segmentation dataset with approximately 20,000 training images and 2,000 validation images. It also includes two separate test sets: A-150, consisting of 150 common categories, and A-847, which contains 847 categories." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.438, + 0.484, + 0.528 + ], + "angle": 0, + "content": "- PASCAL VOC [14] is a smaller dataset for semantic segmentation, consisting of 1464 training images and 1449 validation images, covering 20 different foreground categories. We refer to this dataset as PAS-20. In line with [9], we also report scores on PAS-20\\(^b\\), which includes the \"background\" as the 21st category." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.528, + 0.484, + 0.589 + ], + "angle": 0, + "content": "- PASCAL-Context [40] is an extended version of the original PASCAL VOC dataset. It provides two test sets: PC-59 and PC-459, containing 59 and 459 categories respectively, for evaluation purposes." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.362, + 0.484, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.59, + 0.484, + 0.635 + ], + "angle": 0, + "content": "Evaluation Metric. Consistent with prior studies [9, 59], we use mean Intersection over Union (mIoU) to evaluate the semantic segmentation performance on these benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.639, + 0.484, + 0.835 + ], + "angle": 0, + "content": "Implementation Details. We implement our method using the Transformer-based CLIP model. Following the protocol established in [9], we evaluate our results on two versions of the CLIP model: ViT-B/16 and ViT-L/14. For training, we follow the same experimental setup as [9]. Besides, we use the Adam optimizer [29] with an initial learning rate of \\( 1 \\times 10^{-5} \\) for CLIP, and a weight decay of \\( 10^{-4} \\). Training is conducted with one image per mini-batch. We apply the block-diagonal learnable weight in the self-attention module in each layer of CLIP. We employ the cost-based approach from [9] as our decoder. The curvature \\( c \\) of Poincaré ball model is set to 0.01. All models are trained for 80,000 iterations using 4 NVIDIA A800 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.847, + 0.462, + 0.863 + ], + "angle": 0, + "content": "5.2. Comparisons with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Here, we compare our proposed method with several state-of-the-art methods, as shown in Table 1, using six test sets" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.895, + 0.126 + ], + "angle": 0, + "content": "
MethodODISE [60]FC-CLIP [65]MAFT+ [25]Ours
PQ23.426.827.129.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.134, + 0.907, + 0.163 + ], + "angle": 0, + "content": "Table 3. Comparisons on open-vocabulary panoptic segmentation task on the ADE20K [67] dataset. \"PQ\": Panoptic quality." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.17, + 0.905, + 0.312 + ], + "angle": 0, + "content": "
nParam. (M)A-847PC-459A-150PC-59PAS-20b
10.045.99.427.452.474.9
80.3510.115.429.954.275.6
160.7011.417.431.055.976.2
321.4111.718.131.756.676.7
642.8111.918.732.057.877.8
1285.6312.319.232.158.578.9
25611.2512.219.031.958.879.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.322, + 0.905, + 0.363 + ], + "angle": 0, + "content": "Table 4. Ablation on the size of diagonal blocks \\( n \\). The base model is ViT-B/16. The best-performing results are presented in bold, while the second-best results are underlined." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.378, + 0.885, + 0.433 + ], + "angle": 0, + "content": "
MethodsOVSeg [37]CAT-Seg [9]SAN [62]Ours
Param. (M)147.239.58.45.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.441, + 0.905, + 0.469 + ], + "angle": 0, + "content": "Table 5. Efficiency comparison in terms of learnable parameters for fine-tuning CLIP. The base model is ViT-B/16." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.907, + 0.84 + ], + "angle": 0, + "content": "across three benchmarks. Overall, we achieve the best results. Most existing open-vocabulary semantic segmentation methods follow partial fine-tuning strategies, i.e., fine-tuning CLIP's image encoder. Although these methods provide sufficient flexibility for aligning with text descriptions generated by the text encoder, they fail to adjust the text embeddings to the appropriate hierarchical level for effective text-to-image alignment, often leading to suboptimal segmentation performance. Differently, CAT-Seg [9] simultaneously fine-tunes both the text encoder and image encoder of CLIP, achieving performance comparable to ours on some of the datasets. However, its fine-tuning scheme is manually controlled through different layer combinations, necessitating a careful design to balance generalization and segmentation ability, while ours does not suffer from such an issue. Then, compared to SAN [62], another parameter-efficient fine-tuning method that introduces only a limited number of tunable parameters, our approach significantly outperforms it, achieving improvements of \\(6.6\\%\\) on the PC-459 dataset and \\(4.7\\%\\) on the PC-59 dataset with ViT-B/16 as the base model. These results demonstrate the effectiveness of our method in preserving generalization while mastering segmentation capability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Efficiency comparison. We compare the efficiency of our method with other approaches, including OVSeg [37], CATSeg [9], and SAN [62], all of which utilize a ViT version of CLIP. The comparison, summarized in Table 5, shows" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4567" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.402, + 0.907, + 0.431 + ], + "angle": 0, + "content": "Figure 4. Comparison of qualitative results on ADE20K [67] with 847 categories. We compare our method with other two state-of-the-art methods, i.e., SAN [62] and CAT-Seg [9]. The results show our method performs more precise segmentation in different scenarios." + }, + { + "type": "table", + "bbox": [ + 0.129, + 0.444, + 0.866, + 0.656 + ], + "angle": 0, + "content": "
MethodImage encoderText encoderParam. (M)A-847PC-459A-150PC-59PAS-20PAS-20b
FreezeXX04.46.624.849.492.571.9
Parameter-efficient Fine-Tuning
LoRA [21]✓✓7.511.417.628.655.194.276.7
OFT [45]✓✓3.810.918.030.253.793.774.3
VPT [24]✓✓2.25.710.223.754.393.875.1
Adapter [19]✓✓7.510.416.528.854.994.275.2
SSF [36]✓✓0.66.915.228.652.193.272.8
HyperLoRA [64]✓✓7.511.418.229.755.894.377.2
HyperCLIP (Ours)✓X3.410.316.830.557.694.478.0
X✓2.211.117.529.855.193.776.1
✓✓5.612.319.232.158.595.678.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.663, + 0.907, + 0.72 + ], + "angle": 0, + "content": "Table 6. Ablation study on the fine-tuning encoder in HyperCLIP and comparison of universal PEFT methods, including \"LoRA\" [21], \"OFT\" [45], \"VPT\" [24], \"Adapter\" [19], \"LST\" [49], \"SSF\" [36] and \"HyperLoRA\" [64] which is a recent PEFT method fine-tunes large language models in hyperbolic space. For a clear comparison, we ignore the segmentation decoder when calculating the number of learnable parameters. The base model is ViT-B/16. \\(\\checkmark\\): Fine-tuning. \\(\\times\\): Freezing." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.745, + 0.482, + 0.822 + ], + "angle": 0, + "content": "that our method uses the fewest trainable parameters while equipping CLIP with segmentation ability. Since we do not introduce any extra architectures, and the logarithmic and exponential maps can be completed within linear complexity [64], the inference overhead is negligible." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.826, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Application to other tasks. Here, we further carry out extra experiments in two other tasks. In an open-vocabulary object detection task, CLIP's image embeddings are transitioned from the image level to the object level. We follow a recent work [58] to investigate open-vocabulary object detection on" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.745, + 0.907, + 0.852 + ], + "angle": 0, + "content": "the COCO dataset [38]. As shown in Table 2, the superior results demonstrate that HyperCLIP can effectively generalize to other tasks that require aligning text embeddings with embeddings from arbitrary hierarchical levels, e.g., object-level alignment. Furthermore, the open-vocabulary panoptic segmentation results presented in Table 3 further corroborate this conclusion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Qualitative results. Here, we visualize our method's representative example segmentation results against prevailing methods, i.e., CAT-Seg [9] and SAN [62], in the A-847" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4568" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.152 + ], + "angle": 0, + "content": "dataset. As shown in Fig. 4, we can observe that our method can make more accurate predictions on object location and category. Even if objects are small and the background is complicated, our method still performs well." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.165, + 0.245, + 0.182 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.189, + 0.486, + 0.491 + ], + "angle": 0, + "content": "Ablation of fine-tuning CLIP. We assess the efficacy of various fine-tuning design choices for HyperCLIP. As shown in Table 6, we freeze one of CLIP's encoders and compare the performance with several commonly-used PEFT methods, e.g., LoRA [21]. Compared to fully freezing CLIP's encoders (row 1), fine-tuning either encoder independently allows for minimizing its hyperbolic radius, thereby equipping the corresponding encoder with segmentation ability. However, one modality's hierarchical level remains at the image level, which hinders CLIP from fully acquiring segmentation ability. Consequently, jointly fine-tuning both encoders aligns the hierarchical levels at the pixel level, resulting in optimal performance. We also compare our method with existing PEFT strategies. Most of these methods, which operate in Euclidean space, struggle to capture the transition between different hierarchical levels. Notably, HyperLoRA [64] also updates parameters in hyperbolic space. However, because HyperLoRA does not specifically adjust the hyperbolic radius, it fails to fully endow CLIP with segmentation ability, resulting in inferior performance compared to our method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.485, + 0.674 + ], + "angle": 0, + "content": "Ablation of the size of diagonal blocks. The size of the diagonal block in our scaling transformations determines the number of parameters required to equip CLIP with segmentation ability. To investigate its impact, we conduct an ablation study on the scaling transformations by systematically varying the size of the diagonal block. As shown in Table 4, accuracy improves incrementally on datasets with fewer classes. However, on datasets with more classes (e.g., A-847), the accuracy performance reaches a saturation point around a block size of 128. This suggests that increasing the dimensionality of the diagonal block introduces more trainable parameters, which may lead to over-fitting on a closed set." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Effect of the hyperbolic radius of text embeddings. In order to closely examine the effect of the hyperbolic radius on equipping CLIP with segmentation ability, we apply a mean squared error (MSE) loss during fine-tuning. This MSE loss encourages the average hyperbolic radius of the text embeddings to converge towards different target values. As shown in Fig. 5, the initial value of the average hyperbolic radius is approximately 7.5, which is subsequently minimized to around 6.0 to achieve the optimal value for segmentation. Deviating from this value, either by increasing or decreasing the expected radius, results in diminished performance. This observation offers a new insight: the hyperbolic radius required for a vision task might be relatively fixed, suggesting that it could serve as a metric to quantify the granularity of different vision tasks." + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.089, + 0.887, + 0.156 + ], + "angle": 0, + "content": "
Möbius.A-847PC-459A-150PC-59PAS-20b
X11.118.230.257.377.5
✓12.319.232.158.578.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.167, + 0.908, + 0.196 + ], + "angle": 0, + "content": "Table 7. Ablation on the necessity of Möbius matrix multiplication. \"Möbius\": Möbius matrix multiplication." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.207, + 0.903, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.426, + 0.908, + 0.496 + ], + "angle": 0, + "content": "Figure 5. Visualization on the effect of different average hyperbolic radii of text embeddings. The controlling of different hyperbolic radii is achieved via a mean squared error (MSE) loss that enforces a constraint between the expected radius and the mean radius of the text embeddings during fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.909, + 0.62 + ], + "angle": 0, + "content": "Necessity of Möbius matrix multiplication. We ablate this experiment to validate the influence of the Möbius matrix multiplication operation. In Table 7, “\\(\\times\\)” means that we use standard matrix multiplication. It achieves worse performance compared with “√”, which indicates that the Möbius matrix multiplication is essential for minimizing the hyperbolic radius of CLIP's embeddings." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.633, + 0.634, + 0.649 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.909, + 0.9 + ], + "angle": 0, + "content": "In this paper, we propose HyperCLIP, a novel fine-tuning strategy for CLIP tailored to open-vocabulary semantic segmentation. Recognizing the hierarchical structure of images, we hypothesize that aligning the hierarchical level of CLIP's text encoder with pixel-level representations is crucial for accurate pixel-level predictions. HyperCLIP achieves this by adjusting the hierarchical level in hyperbolic space through scaling transformations applied to CLIP's embeddings, enabling precise cross-modal alignment. Extensive experiments demonstrate that HyperCLIP sets new state-of-the-art results in open-vocabulary semantic segmentation. More importantly, we observe that after adjustment, CLIP's text embeddings maintain a relatively fixed hyperbolic radius across datasets. This suggests that the segmentation task's granularity may be quantified using the hyperbolic radius, offering a promising direction for future research." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4569" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.172 + ], + "angle": 0, + "content": "This work was supported by the NSFC under Grant 62322604, 62176159 and 62401361, and in part by the Shanghai Municipal Science and Technology Major Project under Grant 2021SHZDX0102." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.198, + 0.188, + 0.214 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.223, + 0.486, + 0.265 + ], + "angle": 0, + "content": "[1] Irving Biederman. Recognition-by-components: a theory of human image understanding. Psychological review, 94(2): 115, 1987. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.266, + 0.485, + 0.307 + ], + "angle": 0, + "content": "[2] Martin R Bridson and André Haefliger. Metric spaces of non-positive curvature. Springer Science & Business Media, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.309, + 0.483, + 0.35 + ], + "angle": 0, + "content": "[3] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. Advances in Neural Information Processing Systems, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.352, + 0.484, + 0.407 + ], + "angle": 0, + "content": "[4] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.409, + 0.483, + 0.449 + ], + "angle": 0, + "content": "[5] James W Cannon, William J Floyd, Richard Kenyon, Walter R Parry, et al. Hyperbolic geometry. Flavors of geometry, 31 (59-115):2, 1997. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.451, + 0.483, + 0.521 + ], + "angle": 0, + "content": "[6] Jiaxin Chen, Jie Qin, Yuming Shen, Li Liu, Fan Zhu, and Ling Shao. Learning attentive and hierarchical representations for 3d shape recognition. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 105-122. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.522, + 0.485, + 0.589 + ], + "angle": 0, + "content": "[7] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.592, + 0.485, + 0.661 + ], + "angle": 0, + "content": "[8] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.663, + 0.484, + 0.717 + ], + "angle": 0, + "content": "[9] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Cat-seg: Cost aggregation for open-vocabulary semantic segmentation, 2024. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.484, + 0.774 + ], + "angle": 0, + "content": "[10] Karan Desai, Maximilian Nickel, Tanmay Rajpurohit, Justin Johnson, and Shanmukha Ramakrishna Vedantam. Hyperbolic image-text representations. In International Conference on Machine Learning, pages 7694-7731. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.484, + 0.831 + ], + "angle": 0, + "content": "[11] Jian Ding, Nan Xue, Gui-Song Xia, and Dengxin Dai. Decoupling zero-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11583-11592, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Zi-Yi Dou, Aishwarya Kamath, Zhe Gan, Pengchuan Zhang, Jianfeng Wang, Linjie Li, Zicheng Liu, Ce Liu, Yann Le-Cun, Nanyun Peng, et al. Coarse-to-fine vision-language pre-training with fusion in the backbone. Advances in neural information processing systems, 35:32942-32956, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.223, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[13] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuohang Wang, Lijuan Wang, Chenguang Zhu, Pengchuan Zhang, Lu Yuan, Nanyun Peng, et al. An empirical study of training end-to-end vision-and-language transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18166-18176, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.177, + 0.907, + 0.232 + ], + "angle": 0, + "content": "[14] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.907, + 0.288 + ], + "angle": 0, + "content": "[15] Zhe Gan, Yen-Chun Chen, Linjie Li, Chen Zhu, Yu Cheng, and Jingjing Liu. Large-scale adversarial training for vision-and-language representation learning. Advances in Neural Information Processing Systems, 33:6616-6628, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.907, + 0.329 + ], + "angle": 0, + "content": "[16] Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. Advances in neural information processing systems, 31, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.907, + 0.385 + ], + "angle": 0, + "content": "[17] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.907, + 0.442 + ], + "angle": 0, + "content": "[18] Yunhui Guo, Xudong Wang, Yubei Chen, and Stella X Yu. Clipped hyperbolic classifiers are super-hyperbolic classifiers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11-20, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.443, + 0.906, + 0.511 + ], + "angle": 0, + "content": "[19] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790-2799. PMLR, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.512, + 0.907, + 0.594 + ], + "angle": 0, + "content": "[20] Chengyang Hu, Ke-Yue Zhang, Taiping Yao, Shouhong Ding, and Lizhuang Ma. Rethinking generalizable face anti-spoofing via hierarchical prototype-guided distribution refinement in hyperbolic space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1032-1041, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.907, + 0.663 + ], + "angle": 0, + "content": "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.906, + 0.734 + ], + "angle": 0, + "content": "[22] Ming Hu, Kun Yuan, Yaling Shen, Feilong Tang, Xiaohao Xu, Lin Zhou, Wei Li, Ying Chen, Zhongxing Xu, Zelin Peng, et al. Ophclip: Hierarchical retrieval-augmented learning for ophthalmic surgical video-language pretraining. arXiv preprint arXiv:2411.15421, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.907, + 0.803 + ], + "angle": 0, + "content": "[23] Zhicheng Huang, Zhaoyang Zeng, Yupan Huang, Bei Liu, Dongmei Fu, and Jianlong Fu. Seeing out of the box: End-to-end pre-training for vision-language representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12976-12985, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.907, + 0.859 + ], + "angle": 0, + "content": "[24] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[25] Siyu Jiao, Hongguang Zhu, Jiannan Huang, Yao Zhao, Yunchao Wei, and Humphrey Shi. Collaborative vision-text representation optimizing for open-vocabulary segmentation. In" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "4570" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "European Conference on Computer Vision, pages 399-416. Springer, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.191 + ], + "angle": 0, + "content": "[26] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.484, + 0.261 + ], + "angle": 0, + "content": "[27] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6418-6428, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.484, + 0.319 + ], + "angle": 0, + "content": "[28] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International conference on machine learning, pages 5583-5594. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.484, + 0.36 + ], + "angle": 0, + "content": "[29] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.364, + 0.484, + 0.432 + ], + "angle": 0, + "content": "[30] Fanjie Kong, Yanbei Chen, Jiarui Cai, and Davide Modolo. Hyperbolic learning with synthetic captions for openworld detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16762-16771, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.484, + 0.489 + ], + "angle": 0, + "content": "[31] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In International Conference on Learning Representations, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.492, + 0.484, + 0.56 + ], + "angle": 0, + "content": "[32] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.484, + 0.63 + ], + "angle": 0, + "content": "[33] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.484, + 0.715 + ], + "angle": 0, + "content": "[34] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.484, + 0.801 + ], + "angle": 0, + "content": "[35] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 121-137. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[36] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. Advances in Neural Information Processing Systems, 35:109-123, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[37] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.133 + ], + "angle": 0, + "content": "mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.219 + ], + "angle": 0, + "content": "[38] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.907, + 0.276 + ], + "angle": 0, + "content": "[39] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.907, + 0.36 + ], + "angle": 0, + "content": "[40] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 891–898, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.907, + 0.417 + ], + "angle": 0, + "content": "[41] Maximillian Nickel and Douwe Kiela. Learning continuous hierarchies in the lorentz model of hyperbolic geometry. In International conference on machine learning, pages 3779-3788. PMLR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.907, + 0.475 + ], + "angle": 0, + "content": "[42] Zelin Peng, Zhengqin Xu, Zhilin Zeng, Yaoming Wang, Lingxi Xie, Qi Tian, and Wei Shen. Parameter-efficient fin-tuning in hyperspherical space for open-vocabulary semantic segmentation. arXiv preprint arXiv:2405.18840, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.908, + 0.546 + ], + "angle": 0, + "content": "[43] Zelin Peng, Zhengqin Xu, Zhilin Zeng, Lingxi Xie, Qi Tian, and Wei Shen. Parameter efficient fine-tuning via cross block orchestration for segment anything model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3743-3752, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.908, + 0.615 + ], + "angle": 0, + "content": "[44] Zelin Peng, Zhengqin Xu, Zhilin Zeng, Xiaokang Yang, and Wei Shen. Sam-parser: Fine-tuning sam efficiently by parameter space reconstruction. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4515-4523, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.908, + 0.687 + ], + "angle": 0, + "content": "[45] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Scholkopf. Controlling text-to-image diffusion by orthogonal finetuning. Advances in Neural Information Processing Systems, 36:79320-79362, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.908, + 0.771 + ], + "angle": 0, + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.908, + 0.843 + ], + "angle": 0, + "content": "[47] Sameera Ramasinghe, Violetta Shevchenko, Gil Avraham, and Ajanthan Thalaiyasingam. Accept the modality gap: An exploration in the hyperbolic space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27263-27272, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.908, + 0.9 + ], + "angle": 0, + "content": "[48] Frederic Sala, Chris De Sa, Albert Gu, and Christopher Ré. Representation tradeoffs for hyperbolic embeddings. In International conference on machine learning, pages 4460-4469. PMLR, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4571" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[49] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Lst: Ladder side-tuning for parameter and memory efficient transfer learning. Advances in Neural Information Processing Systems, 35: 12991-13005, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[50] Dídac Surís, Ruoshi Liu, and Carl Vondrick. Learning the predictability of the future. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12607-12617, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.248 + ], + "angle": 0, + "content": "[51] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.484, + 0.317 + ], + "angle": 0, + "content": "[52] Feilong Tang, Zile Huang, Chengzhi Liu, Qiang Sun, Harry Yang, and Ser-Nam Lim. Intervening anchor token: Decoding strategy in alleviating hallucinations for mllms. In The Thirteenth International Conference on Learning Representations. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.484, + 0.375 + ], + "angle": 0, + "content": "[53] Abraham A Ungar. Hyperbolic trigonometry and its application in the poincaré ball model of hyperbolic geometry. Computers & Mathematics with Applications, 41(1-2):135-147, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.484, + 0.418 + ], + "angle": 0, + "content": "[54] Abraham Albert Ungar. A gyrovector space approach to hyperbolic geometry. Synthesis Lectures on Mathematics and Statistics, 1(1):1-194, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.484, + 0.476 + ], + "angle": 0, + "content": "[55] Jianfeng Wang, Xiaowei Hu, Zhe Gan, Zhengyuan Yang, Xiyang Dai, Zicheng Liu, Yumao Lu, and Lijuan Wang. Ufo: A unified transformer for vision-language representation learning. arXiv preprint arXiv:2111.10023, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.484, + 0.546 + ], + "angle": 0, + "content": "[56] Sijie Wang, Qiyu Kang, Rui She, Wei Wang, Kai Zhao, Yang Song, and Wee Peng Tay. Hypliloc: Towards effective lidar pose regression with hyperbolic fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5176-5185, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.484, + 0.603 + ], + "angle": 0, + "content": "[57] Melanie Weber, Manzil Zaheer, Ankit Singh Rawat, Aditya K Menon, and Sanjiv Kumar. Robust large-margin learning in hyperbolic space. Advances in Neural Information Processing Systems, 33:17863-17873, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.484, + 0.673 + ], + "angle": 0, + "content": "[58] Size Wu, Wenwei Zhang, Lumin Xu, Sheng Jin, Wentao Liu, and Chen Change Loy. Clim: Contrastive language-image mosaic for region representation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6117-6125, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.484, + 0.716 + ], + "angle": 0, + "content": "[59] Bin Xie, Jiale Cao, Jin Xie, Fahad Shahbaz Khan, and Yanwei Pang. Sed: A simple encoder-decoder for open-vocabulary semantic segmentation, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.484, + 0.788 + ], + "angle": 0, + "content": "[60] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2955-2966, 2023. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[61] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[62] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xiang Bai. Side adapter network for open-vocabulary semantic segmentation. In Proceedings of the IEEE/CVF Conference" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.908, + 0.12 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition, pages 2945-2954, 2023. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.908, + 0.204 + ], + "angle": 0, + "content": "[63] Haochen Xue, Feilong Tang, Ming Hu, Yexin Liu, Qidong Huang, Yulong Li, Chengzhi Liu, Zhongxing Xu, Chong Zhang, Chun-Mei Feng, et al. Mmrc: A large-scale benchmark for understanding multimodal large language model in real-world conversation. arXiv preprint arXiv:2502.11903, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.908, + 0.26 + ], + "angle": 0, + "content": "[64] Menglin Yang, Aosong Feng, Bo Xiong, Jihong Liu, Irwin King, and Rex Ying. Hyperbolic fine-tuning for large language models. arXiv preprint arXiv:2410.04010, 2024. 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.908, + 0.317 + ], + "angle": 0, + "content": "[65] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. In NeurIPS, 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.908, + 0.403 + ], + "angle": 0, + "content": "[66] Zhen Yu, Toan Nguyen, Yaniv Gal, Lie Ju, Shekhar S Chandra, Lei Zhang, Paul Bonnington, Victoria Mar, Zhiyong Wang, and Zongyuan Ge. Skin lesion recognition with class-hierarchy regularized hyperbolic embeddings. In International conference on medical image computing and computer-assisted intervention, pages 594-603. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.908, + 0.459 + ], + "angle": 0, + "content": "[67] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127:302-321, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.46, + 0.908, + 0.502 + ], + "angle": 0, + "content": "[68] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.503, + 0.908, + 0.571 + ], + "angle": 0, + "content": "[69] Ziqin Zhou, Yinjie Lei, Bowen Zhang, Lingqiao Liu, and Yifan Liu. Zegclip: Towards adapting clip for zero-shot semantic segmentation. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.573, + 0.908, + 0.616 + ], + "angle": 0, + "content": "[70] Yudong Zhu, Di Zhou, Jinghui Xiao, Xin Jiang, Xiao Chen, and Qun Liu. Hypertext: Endowing fasttext with hyperbolic geometry. arXiv preprint arXiv:2010.16143, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4572" + } + ] +] \ No newline at end of file diff --git a/2025/Understanding Fine-tuning CLIP for Open-vocabulary Semantic Segmentation in Hyperbolic Space/4933d0fc-2175-4f82-8204-a2ef1c424606_origin.pdf b/2025/Understanding Fine-tuning CLIP for Open-vocabulary Semantic Segmentation in Hyperbolic Space/4933d0fc-2175-4f82-8204-a2ef1c424606_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..df9db9e3dcdc5558476a634f855a2e1bcd519037 --- /dev/null +++ b/2025/Understanding Fine-tuning CLIP for Open-vocabulary Semantic Segmentation in Hyperbolic Space/4933d0fc-2175-4f82-8204-a2ef1c424606_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:479415d8cb6593c990f3e63963d14025a5d28ed4584ba2e6158ae7dcad39be2d +size 1977707 diff --git a/2025/Understanding Fine-tuning CLIP for Open-vocabulary Semantic Segmentation in Hyperbolic Space/full.md b/2025/Understanding Fine-tuning CLIP for Open-vocabulary Semantic Segmentation in Hyperbolic Space/full.md new file mode 100644 index 0000000000000000000000000000000000000000..51c63094502bfacf1f9d7f8593a9857e869183b0 --- /dev/null +++ b/2025/Understanding Fine-tuning CLIP for Open-vocabulary Semantic Segmentation in Hyperbolic Space/full.md @@ -0,0 +1,325 @@ +# Understanding Fine-tuning CLIP for Open-vocabulary Semantic Segmentation in Hyperbolic Space + +Zelin Peng $^{1}$ , Zhengqin Xu $^{1}$ , Zhilin Zeng $^{1}$ , Changsong Wen $^{1}$ , Yu Huang $^{1}$ , Menglin Yang $^{2}$ , Feilong Tang $^{3,4}$ , and Wei Shen $^{1(\boxtimes)}$ + +$^{1}$ MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University $^{2}$ Hong Kong University of Science and Technology (Guangzhou) $^{3}$ MBZUAI, $^{4}$ Monash University. + +https://github.com/SJTU-DeepVisionLab/HyperCLIP + +# Abstract + +CLIP, a foundational vision-language model, has emerged as a powerful tool for open-vocabulary semantic segmentation. While freezing the text encoder preserves its powerful embeddings, recent studies show that fine-tuning both the text and image encoders jointly significantly enhances segmentation performance, especially for classes from open sets. In this work, we explain this phenomenon from the perspective of hierarchical alignment, since during fine-tuning, the hierarchy level of image embeddings shifts from image-level to pixel-level. We achieve this by leveraging hyperbolic space, which naturally encoders hierarchical structures. Our key observation is that, during fine-tuning, the hyperbolic radius of CLIP's text embeddings decreases, facilitating better alignment with the pixel-level hierarchical structure of visual data. Building on this insight, we propose HyperCLIP, a novel fine-tuning strategy that adjusts the hyperbolic radius of the text embeddings through scaling transformations. By doing so, HyperCLIP equips CLIP with segmentation capability while introducing only a small number of learnable parameters. Our experiments demonstrate that HyperCLIP achieves state-of-the-art performance on open-vocabulary semantic segmentation tasks across three benchmarks, while fine-tuning only approximately $4\%$ of the total parameters of CLIP. More importantly, we observe that after adjustment, CLIP's text embeddings exhibit a relatively fixed hyperbolic radius across datasets, suggesting that the granularity required for this segmentation task might be quantified using the hyperbolic radius. + +# 1. Introduction + +The goal of open-vocabulary semantic segmentation is to develop a segmentation model capable of labeling each + +![](images/ce1ce3e1555f07c3dd7048c05431bbe51de4cea6ceaa505bb78707e8d71bf6a0.jpg) +Figure 1. Visualization of hyperbolic space by a Poincaré ball model: The height of points reflects the hierarchical level of input data. The black segments refer to the shortest path between the points in the Poincaré ball model. The red arrow line denotes the hyperbolic radius of text embeddings. Embeddings closer to the center have a smaller radius, representing low-level visual information (e.g., pixels), while points farther from the center have a larger radius, representing high-level visual information (e.g., images). During fine-tuning, the hyperbolic radius of text embeddings is adjusted to match the radius associated with pixels, thereby endowing CLIP with segmentation ability (Best viewed in color). + +pixel in an image with categories that extend beyond a predefined closed set, based on textual descriptions. Vision-language foundation models [7, 12, 13, 15, 22, 23, 26, 28, 32–35, 39, 43, 44, 46, 51, 52, 55, 63], particularly CLIP [46], are frequently employed to provide open-vocabulary recognition capabilities. Consequently, open-vocabulary seman + +tic segmentation fundamentally involves transferring these vision-language foundation models, originally trained with image-level supervision, to achieve pixel-level predictions. + +Current methods [9, 59, 62, 65] typically fine-tune CLIP on a closed set with segmentation annotations, i.e., COCO [4], to equip it with the segmentation capability. A prevailing opinion is that simply freezing CLIP's text encoder maximally preserves the powerful text embeddings of a vast array of classes, which is believed to be beneficial for good generalization on open sets. Many studies follow this opinion [31, 37, 62]. However, recent studies [9, 42, 59] offer an alternative perspective on fine-tuning: simultaneously fine-tuning both the image and text encoders results in superior segmentation performance on open sets compared to former approaches. To explain and resolve this seemingly paradoxical phenomenon, we first leverage a non-Euclidean manifold—specifically, hyperbolic space [5]—to understand why fine-tuning the text encoder appears to improve segmentation performance on open sets. Building on this understanding, our objective is to design an explainable fine-tuning strategy with a minimal number of introduced tunable parameters, which is able to equip CLIP with strong segmentation performance while minimizing the loss of its generalization ability. + +Existing research [1] has found that images inherently exhibit hierarchical structures, consisting of multiple levels: pixels, patches, objects, entire scenes, etc. Given that, in open-vocabulary semantic segmentation, CLIP's image embeddings transition from image level to pixel level, we infer that the performance improvement brought by the recent studies [9, 59] could be attributed to adjusting CLIP's text embeddings from their original hierarchical level to better align with the level represented by pixels, thereby enhancing cross-modal alignment. However, most existing fine-tuning methods operate in Euclidean space, which makes it challenging to quantify such a transition through different hierarchical levels. Hyperbolic spaces have gained significant interest in recent years, owing to their ability to naturally and compactly encode hierarchical structures [47, 48, 57], as shown in Fig. 1. The hyperbolic radius, defined as the distance from a point to the center in hyperbolic space, represents its hierarchical level [50]. Points closer to the center (smaller radius) correspond to low-level visual embeddings, e.g., those from pixels, while points near the edge (larger radius) correspond to high-level visual embeddings, e.g., those from images. + +Motivated by this, we project the text embeddings into hyperbolic space and visualize the changes in their radii before and after fine-tuning, as shown in Fig. 2. This fine-tuning is realized using the state-of-the-art method provided in [9]. Specifically, one can observe that the radius after fine-tuning is smaller than that of pre-trained CLIP. In particular, we identify a pattern during fine-tuning: the hierarchical level of + +![](images/ae42f7ff7a74c0af75347c3ac569e41b75208e3d2c72093f57eb1f21128d9361.jpg) +Figure 2. Understanding fine-tuning CLIP's text encoder from the perspective of hyperbolic radius. We use hyperbolic radius to illustrate the change in the hierarchical level of 171 classes after fine-tuning. We observe that the hyperbolic radius of each class becomes smaller, suggesting that adjusting the hyperbolic radius equips CLIP with segmentation ability (Best viewed in color). + +the text embeddings adjusts in conjunction with the hierarchical level of the embeddings provided by the image encoder. Our finding brings a novel perspective for fine-tuning CLIP: Adjusting the hyperbolic radius of text embeddings from edge to center endowed CLIP with segmentation ability. + +Building on the analysis above, we propose a novel finetuning strategy for CLIP in hyperbolic space, termed HyperCLIP. HyperCLIP aims to directly adjust the hyperbolic radius of CLIP's embeddings to facilitate alignment with the appropriate hierarchical level for segmentation tasks. Specifically, we introduce several block-diagonal scaling matrices. Multiplying these matrices with CLIP's feature representations using a Möbius matrix multiplication operation is equivalent to performing scaling transformations that adjust the radius of CLIP's embeddings. This enables the adjustment of their hierarchical levels, ultimately enhancing the model's performance. As a result, HyperCLIP can equip CLIP with segmentation ability with only a small number of introduced learnable parameters. + +Extensive experiments demonstrate that our method sets new state-of-the-art results in open-vocabulary semantic segmentation across three benchmarks, while fine-tuning only approximately $4\%$ of the total parameters of CLIP. More importantly, we observe that after adjustment, CLIP's text embeddings maintain a relatively fixed hyperbolic radius across different datasets. This suggests that the granularity of the segmentation task might be quantified using the hyperbolic radius, warranting further exploration. + +# 2. Related Work + +# 2.1. Open-vocabulary Semantic Segmentation + +Prior works on open-vocabulary semantic segmentation typically tackle this task by leveraging CLIP [46]. Early efforts, + +e.g., [68], directly fine-tune CLIP on mainstream segmentation datasets, like COCO [4]. However, these approaches argue that fully fine-tuning CLIP's encoder significantly diminishes its ability to generalize to unseen classes. To mitigate this issue, some methods [11, 17, 60, 61] take the opposite approach, fine-tuning an additional mask generator [8] for segmentation while keeping CLIP frozen to preserve its generalization capabilities. However, this approach, with its frozen parameter space, lacks segmentation awareness, leading to a misalignment between regions and text descriptions [37]. Other studies [9, 62, 65] propose more sophisticated solutions that fine-tune only specific parameters, e.g., certain layers of CLIP, enabling pixel-level predictions while keeping the majority of CLIP's parameters fixed to minimize the loss of generalization. While these approaches show significant advantages, they often rely on very small learning rates, implicitly promoting minimal deviations from the pre-trained CLIP, which limits segmentation performance. In summary, there is still limited understanding of how fine-tuning influences CLIP's representations and why fine-tuning CLIP's text encoder appears to enhance segmentation performance. In this paper, we aim to provide several new insights that reveal the underlying logic by incorporating a novel perspective: hyperbolic space. + +# 2.2. Learning in Hyperbolic Spaces + +Unlike Euclidean geometry, hyperbolic space can be viewed as the continuous analog of a tree [2], enabling a hierarchical understanding of knowledge structures. Previous research has demonstrated the advantages of representing such hierarchical relationships in various domains, including molecular structures [66], 3D data [6, 56], text data [16, 70], and images [18, 27]. Furthermore, several studies have investigated hyperbolic learning in vision tasks to model hierarchical relationships across different modalities [10, 30, 47], showcasing the benefits of capturing the semantic hierarchy among various data types. Nevertheless, while most studies typically leverage the intrinsic hierarchical properties of hyperbolic space, this paper seeks to capture the detailed changes in semantics reflected by the hyperbolic radius. This approach potentially complements current fine-tuning efforts in open-vocabulary tasks by offering a novel hyperbolic perspective. + +# 3. Preliminary Background + +# 3.1. Hyperbolic geometry + +Hyperbolic geometry often has constant negative curvature and is fundamentally different to the Euclidean or spherical geometry which has zero or constant positive curvature, respectively. This enables unique properties such as the divergence of parallel lines and the exponential volume growth towards the boundary. Hyperbolic geometry has five different analytic models to build the hyperbolic space [5]. In this + +work, we utilize the classical Poincaré ball model as its effectiveness in capturing hierarchical structures [16, 20, 41, 47]. + +# 3.2. The Poincaré Ball Model + +The Poincaré ball model $(\mathbb{D}_c^n,g^{\mathbb{D}_c})$ with radius $1 / \sqrt{c}$ and negative curvature $-c(c > 0)$ is a Riemannian manifold $\mathbb{D}_c^n \coloneqq \{\mathbf{x} \in \mathbb{R}^n : c\|\mathbf{x}\| < 1\}$ with the metric $g^{\mathbb{D}_c} = \lambda_{c,\mathbf{x}}^2 g^E$ , where $\lambda_{c,\mathbf{x}} = \frac{2}{1 - c\|\mathbf{x}\|^2}$ is the conformal factor, and $g^E = \mathbf{I}_n$ is the Euclidean metric tensor. The vectorial structure operations defined in Euclidean space can not be directly used in hyperbolic space due to the differential of the space properties. Therefore, we introduce the framework of gyrovector spaces [53, 54] to provide non-associative algebraic operations for hyperbolic space. According to special relativity, when adding speed vectors belonging to the Poincaré ball of radius $c$ , these added speed vectors remain in the ball by the gyrovector spaces operation, also known as the Möbius operation. For instance, the addition and multiplication operations are defined as follows: + +Mobius addition operation. For $\mathbf{x},\mathbf{y}\in \mathbb{D}_c^n$ , the Mobius addition operation is defined as: + +$$ +\mathbf {x} \oplus_ {c} \mathbf {y} := \frac {(1 + 2 c \langle \mathbf {x} , \mathbf {y} \rangle + c \| \mathbf {y} \| ^ {2}) \mathbf {x} + (1 - c \| \mathbf {x} \| ^ {2}) \mathbf {y}}{1 + 2 c \langle \mathbf {x} , \mathbf {y} \rangle + c ^ {2} \| \mathbf {x} \| ^ {2} \| \mathbf {y} \| ^ {2}}. \tag {1} +$$ + +Mobius scalar multiplication operation. For a scalar $r \in \mathbb{R}$ and a vector $\mathbf{x} \in \mathbb{D}_c^n$ , the Möbius scalar multiplication operation is defined as: + +$$ +r \otimes_ {c} \mathbf {x} := (1 / \sqrt {c}) \tanh \left(r \tanh ^ {- 1} (\sqrt {c} \| \mathbf {x} \|\right)) \frac {\mathbf {x}}{\| \mathbf {x} \|}. \tag {2} +$$ + +The operation satisfies $r \otimes_{c} \mathbf{0} \coloneqq \mathbf{0}$ , $r \otimes_{c} \mathbf{x} = \mathbf{x} \oplus_{c} \dots \oplus_{c} \mathbf{x}$ ( $r$ additions), $(r' + r) \otimes_{c} \mathbf{x} = r \otimes \mathbf{x} \oplus_{c} r' \otimes_{c} \mathbf{x}$ , $(r'r) \otimes_{c} \mathbf{x} = r \otimes_{c} (r' \otimes_{c} \mathbf{x})$ and $\left(\left|r\right| \otimes_{c} \mathbf{x}\right) / \left\|r \otimes_{c} \mathbf{x}\right\| = \mathbf{x} / \left\|\mathbf{x}\right\|$ . + +Mobius matrix multiplication operation. Refer to the relation definition in [27], for a matrix $\mathbf{M} \in \mathbb{R}^{n \times n}$ and a vector $\mathbf{x} \in \mathbb{D}_c^n$ , if $\mathbf{Mx} \neq \mathbf{0}$ , the Möbius matrix multiplication operation is defined as: + +$$ +\mathbf {M} \otimes_ {c} \mathbf {x} := (\frac {1}{\sqrt {c}}) \tanh \left(\frac {\| \mathbf {M x} \|}{\| \mathbf {x} \|} \tanh ^ {- 1} (\sqrt {c} \| \mathbf {x} \|)\right) \frac {\mathbf {M x}}{\| \mathbf {M x} \|}. +$$ + +The distance of $\mathbf{x},\mathbf{y}\in \mathbb{D}_c^n$ is defined as: + +$$ +d _ {c} ^ {\mathbb {D}} (\mathbf {x}, \mathbf {y}) = (2 / \sqrt {c}) \tanh ^ {- 1} (\sqrt {c} \| - \mathbf {x} \oplus_ {c} \mathbf {y} \Vert). \tag {3} +$$ + +The tangent space $\mathcal{T}_{\mathbf{x}}^{c}\mathbb{D}_{c}^{n}$ at a point $\mathbf{x}\in \mathbb{D}_c^n$ is the first order approximation of $\mathbb{D}_c^n$ , which is an $n$ -dimensional Euclidean space. The tangent space $\mathcal{T}_{\mathbf{x}}^{c}\mathbb{D}_{c}^{n}$ and $\mathbb{D}_c^n$ are mapped to each other by exponential $(\mathcal{T}_{\mathbf{x}}^{c}\mathbb{D}_{c}^{n}\mapsto \mathbb{D}_{c}^{n}:\exp_{\mathbf{x}}^{\mathbb{D},c}(\cdot))$ and logarithmic $(\mathbb{D}_c^n\mapsto \mathcal{T}_x^c\mathbb{D}_c^n:\log_{\mathbf{x}}^{\mathbb{D},c}(\cdot))$ maps, respectively. For any $\mathbf{x},\mathbf{y}\in \mathbb{D}_c^n$ and $\mathbf{v}\in \mathcal{T}_{\mathbf{x}}^{c}\mathbb{D}_{c}^{n}$ , the mapping functions + +![](images/96a21682e611343295ba4c39cd67d5c983ea22b86ce529b120b03380e9df88be.jpg) +Figure 3. A schematic representation of HyperCLIP. In the HyperCLIP framework, we fine-tune CLIP in hyperbolic space by minimizing its embeddings' hyperbolic radii using several tunable block-diagonal scaling matrices. + +are given for $\mathbf{v} \neq \mathbf{0}$ and $\mathbf{y} \neq \mathbf{x}$ by: + +$$ +\exp_ {\mathbf {x}} ^ {\mathbb {D}, c} (\mathbf {v}) = \mathbf {x} \oplus_ {c} \left(\tanh \left(\sqrt {c} \frac {\lambda_ {c , \mathbf {x}} \| \mathbf {v} \|}{2}\right) \frac {\mathbf {v}}{\sqrt {c} \| \mathbf {v} \|}\right), \tag {4} +$$ + +$$ +\log_ {\mathbf {x}} ^ {\mathbb {D}, c} (\mathbf {y}) = \frac {2}{\sqrt {c} \lambda_ {c , \mathbf {x}}} \tanh ^ {- 1} \left(\sqrt {c} \| - \mathbf {x} \oplus_ {c} \mathbf {y} \|\right) \frac {- \mathbf {x} \oplus_ {c} \mathbf {y}}{\| - \mathbf {x} \oplus_ {c} \mathbf {y} \|}. \tag {5} +$$ + +# 4. Methodology + +In this section, we introduce HyperCLIP for open-vocabulary semantic segmentation. HyperCLIP is based on the Poincaré ball model, which aims to adjust the hyperbolic radius of CLIP's embeddings, thereby equipping CLIP with segmentation ability. + +# 4.1. Fine-tuning in Hyperbolic Space + +The core technique in HyperCLIP involves linear transformations, i.e., scaling, in the hyperbolic space. One of the primary methods [64] to learn the weights of linear transformations in hyperbolic space is hyperLoRA [64]. Given a feature embedding $\mathbf{z}$ in Euclidean space, this method first maps $\mathbf{z}$ to the hyperbolic space at a local reference point, typically the center, using the exponential map (e.g., Eq. 4). Linear transformations with learnable weight matrix $\mathbf{W}$ is then applied to $\mathbf{z}$ within the hyperbolic space to obtain the hyperbolic representation. Then, a logarithmic map (e.g., Eq. 5) is applied to map the hyperbolic representation back to the Euclidean space. Formally, in the Poincaré ball model, this process is realized as: + +$$ +\mathbf {W} \mathbf {z} = \log_ {\mathbf {0}} ^ {\mathbb {D}, c} \left(\mathbf {W} \otimes_ {c} \exp_ {\mathbf {0}} ^ {\mathbb {D}, c} (\mathbf {z})\right), \tag {6} +$$ + +where $\otimes_{c}$ is the mobius matrix multiplication operation. The primary aim of this method [64] is to capture more intricate hierarchical relationships within hyperbolic space. In contrast, our goal is to directly adjust the hyperbolic radius of CLIP's embeddings, leading to a scaling transformation. + +# 4.2. Hyperbolic Radius Adjustment + +The objective of HyperCLIP is to introduce a scaling transformation to directly adjust the hyperbolic radii of CLIP's embeddings in hyperbolic space, thereby equipping CLIP with segmentation ability. Concretely, it necessitates the utilization of a diagonal matrix, denoted as $\mathbf{S}$ , to scale a CLIP's feature embedding $\mathbf{z} \in \mathbb{R}^{b \times d}$ , where $d$ is the feature dimension of $\mathbf{z}$ , $b$ is the spatial dimension of input images or number of classes. When extending this process to hyperbolic space, $\mathbf{S}$ can directly adjust the hyperbolic radius of $\mathbf{z}$ (this deviation can be found in Sec. 4.3). Specifically, we realize the diagonal matrix $\mathbf{S}$ in a block-wise manner, + +$$ +\mathbf {S} = \operatorname {d i a g} \left(\mathbf {S} _ {1}, \mathbf {S} _ {2}, \dots , \mathbf {S} _ {K}\right), \tag {7} +$$ + +where $\mathbf{S}_k\in \mathbb{R}^{n\times n}$ denotes a square matrix of the $k$ -th block, and the block number $K$ is calculated by $K = d / n$ . By adjusting the size of $n$ , we can achieve a transition between naive scaling transformation $(n = 1$ , i.e., modifying only the diagonal elements) and arbitrary transformation $(n = d$ , i.e., fully fine-tuning in hyperbolic space). The ablation study of $n$ is presented in Sec. 5.3. + +Overall Fine-tuning. During the fine-tuning phase, HyperCLIP introduces a similar process to Eq. 6 is incorporated into CLIP's encoder. Specifically, HyperCLIP uses the following forward pass: + +$$ +\mathbf {S} \mathbf {z} = \log_ {\mathbf {0}} ^ {\mathbb {D}, c} \left(\mathbf {S} \otimes_ {c} \exp_ {\mathbf {0}} ^ {\mathbb {D}, c} (\mathbf {z})\right). \tag {8} +$$ + +Concurrently, the original components of CLIP load their weights from the pre-trained checkpoint, with their parameters remaining frozen. + +Parameter complexity. For an $K$ -block diagonal matrix, the number of parameters is $d^2 / K$ , resulting in a complexity of $\mathcal{O}\left(d^2 / K\right)$ . We can optionally share the block matrix to reduce the number of parameters, i.e., $\mathbf{S}_k = \mathbf{S}_j, \forall k \neq j$ . This reduces the parameter complexity to $\mathcal{O}\left(d^2 / K^2\right)$ . We can further reduce parameter complexity by setting $n = 1$ which transforms the block-diagonal matrix into a diagonal + +
ModelVLMAdditional BackboneFine-tuning SpaceA-847PC-459A-150PC-59PAS-20\( PAS-20^b \)
Partial Fine-Tuning
ZS3Net [3]-ResNet-101E---19.438.3-
LSeg [31]CLIP ViT-B/32ResNet-101E----47.4-
OpenSeg [17]ALIGNResNet-101E4.47.917.540.1-63.8
ZegFormer [11]CLIP ViT-B/16ResNet-101E4.99.116.942.886.262.7
ZSseg [61]CLIP ViT-B/16ResNet-101E7.0-20.547.788.4-
OVSeg [37]CLIP ViT-B/16ResNet-101cE7.111.024.853.392.6-
ZegCLIP [69]CLIP ViT-B/16-E---41.293.6-
Selective Fine-Tuning
SED [59]CLIP ConvNeXt-B-E11.418.631.657.394.4-
CAT-Seg [9]CLIP ViT-B/16-E12.019.031.857.594.677.3
Parameter-efficient Fine-Tuning
SAN [62]CLIP ViT-B/16Side AdapterE10.112.627.553.894.0-
OursCLIP ViT-B/16-H12.319.232.158.595.678.9
Partial Fine-Tuning
LSeg [31]CLIP ViT-B/32ViT-L/16E----52.3-
OpenSeg [17]ALIGNEff-B7E8.111.526.444.8-70.2
OVSeg [37]CLIP ViT-L/14Swin-BE9.012.429.655.794.5-
SAN [62]CLIP ViT-L/14-E12.415.732.157.794.6-
ODISE [60]CLIP ViT-L/14Stable DiffusionE11.114.529.957.3--
FC-CLIP [65]CLIP ConvNeXt-L-E14.818.234.158.495.4-
Selective Fine-Tuning
SED [59]CLIP ConvNeXt-L-E13.922.635.260.696.1-
CAT-Seg [9]CLIP ViT-L/14-E16.023.837.963.397.082.5
Parameter-efficient Fine-Tuning
SAN [62]CLIP ViT-L/14Side AdapterE12.415.732.157.794.6-
OursCLIP ViT-L/14-H16.324.138.264.298.384.3
+ +Table 1. Comparison with state-of-the-art methods on standard benchmarks. The best-performing results are presented in bold, while the second-best results are underlined. "E": Euclidean Space. "H": Hyperbolic Space. + +matrix. In this case, the number of parameters becomes $d$ , resulting in a complexity of $\mathcal{O}(d)$ . Despite all these strategies to improve parameter efficiency, we note that the resulting matrix $\mathbf{S}$ remains to realize scaling transformations, thereby allowing it to adjust the radius of CLIP's embeddings. + +# 4.3. Theoretical Analysis + +To further demonstrate the effectiveness of the proposed HyperCLIP method, we conduct a theoretical analysis of the scaling transformation. Theorem 1 reveals that the employed scaling matrix can directly adjust the hyperbolic radius, thus enabling transferring through different hierarchical levels. Detailed proof refers to the supplementary. + +Theorem 1 (Scaling). According to the definitions of the classical Poincaré Ball model and the hyperbolic tangent, for a point $\mathbf{x} \in \mathbb{D}_c^n$ in hyperbolic space, its hyperbolic radius is defined as: + +$$ +\operatorname {R a d} _ {\mathbf {x}} := d _ {c} ^ {\mathbb {D}} (\mathbf {x}, \mathbf {0}) = \left(\frac {2}{\sqrt {c}}\right) \tanh ^ {- 1} \left(\sqrt {c} \| \mathbf {x} \|\right), \tag {9} +$$ + +where $\mathbf{0}$ is the center of the hyperbolic space. Let $s$ is a scaling parameter, combine with Eq. 2 and Eq. 9, the radius + +of the point $\mathbf{x}_s = \mathcal{T}_{\mathbf{x}}(s) := s \otimes_c \mathbf{x}$ which changed by the scaling transformation $\mathcal{T}_{\mathbf{x}}(s)$ is obtained by: + +$$ +\begin{array}{l} \operatorname {R a d} _ {\mathbf {x}, s} (s) = \frac {2}{\sqrt {c}} \tanh ^ {- 1} (\sqrt {c} \| \mathbf {x} _ {s} \|) \\ = \frac {2}{\sqrt {c}} \tanh ^ {- 1} (\sqrt {c} \left| \left| \frac {\tanh \left(\frac {s \sqrt {c}}{2} \operatorname {R a d} _ {\mathbf {x}}\right) \mathbf {x}}{\sqrt {c} \| \mathbf {x} \|} \right| \right|. \tag {10} \\ \end{array} +$$ + +It is noticed that the $\mathrm{Rad}_{\mathbf{x},s}(s)$ satisfies the monotonicity criteria, therefore, we can adjust the hyperbolic radius by changing the scaling parameter $s$ . Refer to the definition of the Möbius matrix multiplication operation in [27], the scaling matrix $\mathbf{S}$ also has the ability to adjust the hyperbolic radius via the scaling transformation $\mathcal{T}_{\mathbf{x}}(\mathbf{S})$ . + +# 5. Experiments + +# 5.1. Experimental Setup + +Datasets. Following prior works [9, 59], we use the COCO-Stuff dataset [4] as our training set. This dataset consists of around 118,000 densely annotated images, covering 171 distinct semantic categories. For inference, we compare our + +
MethodsSpace for Fine-tuning\( \mathbf{A P^{Base}}_{50} \)\( \mathbf{A P^{Novel}}_{50} \)
CLIP-21.636.4
Other Fine-Tuning
CLIM [58]E25.742.5
CAT-Seg [9]E25.242.3
Parameter-efficient Fine-Tuning
LoRA [21]E24.441.5
HyperLoRA [64]H25.142.0
OursH26.342.7
+ +method against state-of-the-art approaches across several semantic segmentation benchmarks, including ADE20K [67], PASCAL VOC [14], and PASCAL-Context [40]. + +- ADE20K [67] is a well-established semantic segmentation dataset with approximately 20,000 training images and 2,000 validation images. It also includes two separate test sets: A-150, consisting of 150 common categories, and A-847, which contains 847 categories. +- PASCAL VOC [14] is a smaller dataset for semantic segmentation, consisting of 1464 training images and 1449 validation images, covering 20 different foreground categories. We refer to this dataset as PAS-20. In line with [9], we also report scores on PAS-20 $^b$ , which includes the "background" as the 21st category. +- PASCAL-Context [40] is an extended version of the original PASCAL VOC dataset. It provides two test sets: PC-59 and PC-459, containing 59 and 459 categories respectively, for evaluation purposes. + +Evaluation Metric. Consistent with prior studies [9, 59], we use mean Intersection over Union (mIoU) to evaluate the semantic segmentation performance on these benchmarks. + +Implementation Details. We implement our method using the Transformer-based CLIP model. Following the protocol established in [9], we evaluate our results on two versions of the CLIP model: ViT-B/16 and ViT-L/14. For training, we follow the same experimental setup as [9]. Besides, we use the Adam optimizer [29] with an initial learning rate of $1 \times 10^{-5}$ for CLIP, and a weight decay of $10^{-4}$ . Training is conducted with one image per mini-batch. We apply the block-diagonal learnable weight in the self-attention module in each layer of CLIP. We employ the cost-based approach from [9] as our decoder. The curvature $c$ of Poincaré ball model is set to 0.01. All models are trained for 80,000 iterations using 4 NVIDIA A800 GPUs. + +# 5.2. Comparisons with State-of-the-art Methods + +Here, we compare our proposed method with several state-of-the-art methods, as shown in Table 1, using six test sets + +Table 2. Comparisons on open-vocabulary object detection task on the COCO [38] dataset. Results are measured by the box AP at IoU threshold 0.5. "E": Euclidean space. "H": hyperbolic space. + +
MethodODISE [60]FC-CLIP [65]MAFT+ [25]Ours
PQ23.426.827.129.2
+ +Table 3. Comparisons on open-vocabulary panoptic segmentation task on the ADE20K [67] dataset. "PQ": Panoptic quality. + +
nParam. (M)A-847PC-459A-150PC-59PAS-20b
10.045.99.427.452.474.9
80.3510.115.429.954.275.6
160.7011.417.431.055.976.2
321.4111.718.131.756.676.7
642.8111.918.732.057.877.8
1285.6312.319.232.158.578.9
25611.2512.219.031.958.879.3
+ +Table 4. Ablation on the size of diagonal blocks $n$ . The base model is ViT-B/16. The best-performing results are presented in bold, while the second-best results are underlined. + +
MethodsOVSeg [37]CAT-Seg [9]SAN [62]Ours
Param. (M)147.239.58.45.6
+ +Table 5. Efficiency comparison in terms of learnable parameters for fine-tuning CLIP. The base model is ViT-B/16. + +across three benchmarks. Overall, we achieve the best results. Most existing open-vocabulary semantic segmentation methods follow partial fine-tuning strategies, i.e., fine-tuning CLIP's image encoder. Although these methods provide sufficient flexibility for aligning with text descriptions generated by the text encoder, they fail to adjust the text embeddings to the appropriate hierarchical level for effective text-to-image alignment, often leading to suboptimal segmentation performance. Differently, CAT-Seg [9] simultaneously fine-tunes both the text encoder and image encoder of CLIP, achieving performance comparable to ours on some of the datasets. However, its fine-tuning scheme is manually controlled through different layer combinations, necessitating a careful design to balance generalization and segmentation ability, while ours does not suffer from such an issue. Then, compared to SAN [62], another parameter-efficient fine-tuning method that introduces only a limited number of tunable parameters, our approach significantly outperforms it, achieving improvements of $6.6\%$ on the PC-459 dataset and $4.7\%$ on the PC-59 dataset with ViT-B/16 as the base model. These results demonstrate the effectiveness of our method in preserving generalization while mastering segmentation capability. + +Efficiency comparison. We compare the efficiency of our method with other approaches, including OVSeg [37], CATSeg [9], and SAN [62], all of which utilize a ViT version of CLIP. The comparison, summarized in Table 5, shows + +![](images/d68b8761010aa08d10241ef82ece80946be3d996a1d0e0aae541bc1df5465765.jpg) +Figure 4. Comparison of qualitative results on ADE20K [67] with 847 categories. We compare our method with other two state-of-the-art methods, i.e., SAN [62] and CAT-Seg [9]. The results show our method performs more precise segmentation in different scenarios. + +
MethodImage encoderText encoderParam. (M)A-847PC-459A-150PC-59PAS-20PAS-20b
FreezeXX04.46.624.849.492.571.9
Parameter-efficient Fine-Tuning
LoRA [21]✓✓7.511.417.628.655.194.276.7
OFT [45]✓✓3.810.918.030.253.793.774.3
VPT [24]✓✓2.25.710.223.754.393.875.1
Adapter [19]✓✓7.510.416.528.854.994.275.2
SSF [36]✓✓0.66.915.228.652.193.272.8
HyperLoRA [64]✓✓7.511.418.229.755.894.377.2
HyperCLIP (Ours)✓X3.410.316.830.557.694.478.0
X✓2.211.117.529.855.193.776.1
✓✓5.612.319.232.158.595.678.9
+ +Table 6. Ablation study on the fine-tuning encoder in HyperCLIP and comparison of universal PEFT methods, including "LoRA" [21], "OFT" [45], "VPT" [24], "Adapter" [19], "LST" [49], "SSF" [36] and "HyperLoRA" [64] which is a recent PEFT method fine-tunes large language models in hyperbolic space. For a clear comparison, we ignore the segmentation decoder when calculating the number of learnable parameters. The base model is ViT-B/16. $\checkmark$ : Fine-tuning. $\times$ : Freezing. + +that our method uses the fewest trainable parameters while equipping CLIP with segmentation ability. Since we do not introduce any extra architectures, and the logarithmic and exponential maps can be completed within linear complexity [64], the inference overhead is negligible. + +Application to other tasks. Here, we further carry out extra experiments in two other tasks. In an open-vocabulary object detection task, CLIP's image embeddings are transitioned from the image level to the object level. We follow a recent work [58] to investigate open-vocabulary object detection on + +the COCO dataset [38]. As shown in Table 2, the superior results demonstrate that HyperCLIP can effectively generalize to other tasks that require aligning text embeddings with embeddings from arbitrary hierarchical levels, e.g., object-level alignment. Furthermore, the open-vocabulary panoptic segmentation results presented in Table 3 further corroborate this conclusion. + +Qualitative results. Here, we visualize our method's representative example segmentation results against prevailing methods, i.e., CAT-Seg [9] and SAN [62], in the A-847 + +dataset. As shown in Fig. 4, we can observe that our method can make more accurate predictions on object location and category. Even if objects are small and the background is complicated, our method still performs well. + +# 5.3. Ablation Study + +Ablation of fine-tuning CLIP. We assess the efficacy of various fine-tuning design choices for HyperCLIP. As shown in Table 6, we freeze one of CLIP's encoders and compare the performance with several commonly-used PEFT methods, e.g., LoRA [21]. Compared to fully freezing CLIP's encoders (row 1), fine-tuning either encoder independently allows for minimizing its hyperbolic radius, thereby equipping the corresponding encoder with segmentation ability. However, one modality's hierarchical level remains at the image level, which hinders CLIP from fully acquiring segmentation ability. Consequently, jointly fine-tuning both encoders aligns the hierarchical levels at the pixel level, resulting in optimal performance. We also compare our method with existing PEFT strategies. Most of these methods, which operate in Euclidean space, struggle to capture the transition between different hierarchical levels. Notably, HyperLoRA [64] also updates parameters in hyperbolic space. However, because HyperLoRA does not specifically adjust the hyperbolic radius, it fails to fully endow CLIP with segmentation ability, resulting in inferior performance compared to our method. + +Ablation of the size of diagonal blocks. The size of the diagonal block in our scaling transformations determines the number of parameters required to equip CLIP with segmentation ability. To investigate its impact, we conduct an ablation study on the scaling transformations by systematically varying the size of the diagonal block. As shown in Table 4, accuracy improves incrementally on datasets with fewer classes. However, on datasets with more classes (e.g., A-847), the accuracy performance reaches a saturation point around a block size of 128. This suggests that increasing the dimensionality of the diagonal block introduces more trainable parameters, which may lead to over-fitting on a closed set. + +Effect of the hyperbolic radius of text embeddings. In order to closely examine the effect of the hyperbolic radius on equipping CLIP with segmentation ability, we apply a mean squared error (MSE) loss during fine-tuning. This MSE loss encourages the average hyperbolic radius of the text embeddings to converge towards different target values. As shown in Fig. 5, the initial value of the average hyperbolic radius is approximately 7.5, which is subsequently minimized to around 6.0 to achieve the optimal value for segmentation. Deviating from this value, either by increasing or decreasing the expected radius, results in diminished performance. This observation offers a new insight: the hyperbolic radius required for a vision task might be relatively fixed, suggesting that it could serve as a metric to quantify the granularity of different vision tasks. + +
Möbius.A-847PC-459A-150PC-59PAS-20b
X11.118.230.257.377.5
✓12.319.232.158.578.9
+ +Table 7. Ablation on the necessity of Möbius matrix multiplication. "Möbius": Möbius matrix multiplication. + +![](images/027af0d30e3bce7c9134c69adcd015dd10e90dbbd331a1282bc33de86be3bacc.jpg) +Figure 5. Visualization on the effect of different average hyperbolic radii of text embeddings. The controlling of different hyperbolic radii is achieved via a mean squared error (MSE) loss that enforces a constraint between the expected radius and the mean radius of the text embeddings during fine-tuning. + +Necessity of Möbius matrix multiplication. We ablate this experiment to validate the influence of the Möbius matrix multiplication operation. In Table 7, “ $\times$ ” means that we use standard matrix multiplication. It achieves worse performance compared with “√”, which indicates that the Möbius matrix multiplication is essential for minimizing the hyperbolic radius of CLIP's embeddings. + +# 6. Conclusion + +In this paper, we propose HyperCLIP, a novel fine-tuning strategy for CLIP tailored to open-vocabulary semantic segmentation. Recognizing the hierarchical structure of images, we hypothesize that aligning the hierarchical level of CLIP's text encoder with pixel-level representations is crucial for accurate pixel-level predictions. HyperCLIP achieves this by adjusting the hierarchical level in hyperbolic space through scaling transformations applied to CLIP's embeddings, enabling precise cross-modal alignment. Extensive experiments demonstrate that HyperCLIP sets new state-of-the-art results in open-vocabulary semantic segmentation. More importantly, we observe that after adjustment, CLIP's text embeddings maintain a relatively fixed hyperbolic radius across datasets. This suggests that the segmentation task's granularity may be quantified using the hyperbolic radius, offering a promising direction for future research. + +# Acknowledgments + +This work was supported by the NSFC under Grant 62322604, 62176159 and 62401361, and in part by the Shanghai Municipal Science and Technology Major Project under Grant 2021SHZDX0102. + +# References + +[1] Irving Biederman. Recognition-by-components: a theory of human image understanding. Psychological review, 94(2): 115, 1987. 2 +[2] Martin R Bridson and André Haefliger. Metric spaces of non-positive curvature. Springer Science & Business Media, 2013. 3 +[3] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. Advances in Neural Information Processing Systems, 32, 2019. 5 +[4] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 2, 3, 5 +[5] James W Cannon, William J Floyd, Richard Kenyon, Walter R Parry, et al. Hyperbolic geometry. Flavors of geometry, 31 (59-115):2, 1997. 2, 3 +[6] Jiaxin Chen, Jie Qin, Yuming Shen, Li Liu, Fan Zhu, and Ling Shao. Learning attentive and hierarchical representations for 3d shape recognition. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 105-122. Springer, 2020. 3 +[7] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 1 +[8] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 3 +[9] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Cat-seg: Cost aggregation for open-vocabulary semantic segmentation, 2024. 2, 3, 5, 6, 7 +[10] Karan Desai, Maximilian Nickel, Tanmay Rajpurohit, Justin Johnson, and Shanmukha Ramakrishna Vedantam. Hyperbolic image-text representations. In International Conference on Machine Learning, pages 7694-7731. PMLR, 2023. 3 +[11] Jian Ding, Nan Xue, Gui-Song Xia, and Dengxin Dai. Decoupling zero-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11583-11592, 2022. 3, 5 +[12] Zi-Yi Dou, Aishwarya Kamath, Zhe Gan, Pengchuan Zhang, Jianfeng Wang, Linjie Li, Zicheng Liu, Ce Liu, Yann Le-Cun, Nanyun Peng, et al. Coarse-to-fine vision-language pre-training with fusion in the backbone. Advances in neural information processing systems, 35:32942-32956, 2022. 1 + +[13] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuohang Wang, Lijuan Wang, Chenguang Zhu, Pengchuan Zhang, Lu Yuan, Nanyun Peng, et al. An empirical study of training end-to-end vision-and-language transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18166-18176, 2022. 1 +[14] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 6 +[15] Zhe Gan, Yen-Chun Chen, Linjie Li, Chen Zhu, Yu Cheng, and Jingjing Liu. Large-scale adversarial training for vision-and-language representation learning. Advances in Neural Information Processing Systems, 33:6616-6628, 2020. 1 +[16] Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. Advances in neural information processing systems, 31, 2018. 3 +[17] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 3, 5 +[18] Yunhui Guo, Xudong Wang, Yubei Chen, and Stella X Yu. Clipped hyperbolic classifiers are super-hyperbolic classifiers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11-20, 2022. 3 +[19] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790-2799. PMLR, 2019. 7 +[20] Chengyang Hu, Ke-Yue Zhang, Taiping Yao, Shouhong Ding, and Lizhuang Ma. Rethinking generalizable face anti-spoofing via hierarchical prototype-guided distribution refinement in hyperbolic space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1032-1041, 2024. 3 +[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 6, 7, 8 +[22] Ming Hu, Kun Yuan, Yaling Shen, Feilong Tang, Xiaohao Xu, Lin Zhou, Wei Li, Ying Chen, Zhongxing Xu, Zelin Peng, et al. Ophclip: Hierarchical retrieval-augmented learning for ophthalmic surgical video-language pretraining. arXiv preprint arXiv:2411.15421, 2024. 1 +[23] Zhicheng Huang, Zhaoyang Zeng, Yupan Huang, Bei Liu, Dongmei Fu, and Jianlong Fu. Seeing out of the box: End-to-end pre-training for vision-language representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12976-12985, 2021. 1 +[24] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 7 +[25] Siyu Jiao, Hongguang Zhu, Jiannan Huang, Yao Zhao, Yunchao Wei, and Humphrey Shi. Collaborative vision-text representation optimizing for open-vocabulary segmentation. In + +European Conference on Computer Vision, pages 399-416. Springer, 2024. 6 +[26] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 1 +[27] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6418-6428, 2020. 3, 5 +[28] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International conference on machine learning, pages 5583-5594. PMLR, 2021. 1 +[29] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[30] Fanjie Kong, Yanbei Chen, Jiarui Cai, and Davide Modolo. Hyperbolic learning with synthetic captions for openworld detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16762-16771, 2024. 3 +[31] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In International Conference on Learning Representations, 2022. 2, 5 +[32] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 1 +[33] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. +[34] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. +[35] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 121-137. Springer, 2020. 1 +[36] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. Advances in Neural Information Processing Systems, 35:109-123, 2022. 7 +[37] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with + +mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 2, 3, 5, 6 +[38] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6, 7 +[39] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 1 +[40] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 891–898, 2014. 6 +[41] Maximillian Nickel and Douwe Kiela. Learning continuous hierarchies in the lorentz model of hyperbolic geometry. In International conference on machine learning, pages 3779-3788. PMLR, 2018. 3 +[42] Zelin Peng, Zhengqin Xu, Zhilin Zeng, Yaoming Wang, Lingxi Xie, Qi Tian, and Wei Shen. Parameter-efficient fin-tuning in hyperspherical space for open-vocabulary semantic segmentation. arXiv preprint arXiv:2405.18840, 2024. 2 +[43] Zelin Peng, Zhengqin Xu, Zhilin Zeng, Lingxi Xie, Qi Tian, and Wei Shen. Parameter efficient fine-tuning via cross block orchestration for segment anything model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3743-3752, 2024. 1 +[44] Zelin Peng, Zhengqin Xu, Zhilin Zeng, Xiaokang Yang, and Wei Shen. Sam-parser: Fine-tuning sam efficiently by parameter space reconstruction. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4515-4523, 2024. 1 +[45] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Scholkopf. Controlling text-to-image diffusion by orthogonal finetuning. Advances in Neural Information Processing Systems, 36:79320-79362, 2023. 7 +[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2 +[47] Sameera Ramasinghe, Violetta Shevchenko, Gil Avraham, and Ajanthan Thalaiyasingam. Accept the modality gap: An exploration in the hyperbolic space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27263-27272, 2024. 2, 3 +[48] Frederic Sala, Chris De Sa, Albert Gu, and Christopher Ré. Representation tradeoffs for hyperbolic embeddings. In International conference on machine learning, pages 4460-4469. PMLR, 2018. 2 + +[49] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Lst: Ladder side-tuning for parameter and memory efficient transfer learning. Advances in Neural Information Processing Systems, 35: 12991-13005, 2022. 7 +[50] Dídac Surís, Ruoshi Liu, and Carl Vondrick. Learning the predictability of the future. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12607-12617, 2021. 2 +[51] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 1 +[52] Feilong Tang, Zile Huang, Chengzhi Liu, Qiang Sun, Harry Yang, and Ser-Nam Lim. Intervening anchor token: Decoding strategy in alleviating hallucinations for mllms. In The Thirteenth International Conference on Learning Representations. 1 +[53] Abraham A Ungar. Hyperbolic trigonometry and its application in the poincaré ball model of hyperbolic geometry. Computers & Mathematics with Applications, 41(1-2):135-147, 2001. 3 +[54] Abraham Albert Ungar. A gyrovector space approach to hyperbolic geometry. Synthesis Lectures on Mathematics and Statistics, 1(1):1-194, 2008. 3 +[55] Jianfeng Wang, Xiaowei Hu, Zhe Gan, Zhengyuan Yang, Xiyang Dai, Zicheng Liu, Yumao Lu, and Lijuan Wang. Ufo: A unified transformer for vision-language representation learning. arXiv preprint arXiv:2111.10023, 2021. 1 +[56] Sijie Wang, Qiyu Kang, Rui She, Wei Wang, Kai Zhao, Yang Song, and Wee Peng Tay. Hypliloc: Towards effective lidar pose regression with hyperbolic fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5176-5185, 2023. 3 +[57] Melanie Weber, Manzil Zaheer, Ankit Singh Rawat, Aditya K Menon, and Sanjiv Kumar. Robust large-margin learning in hyperbolic space. Advances in Neural Information Processing Systems, 33:17863-17873, 2020. 2 +[58] Size Wu, Wenwei Zhang, Lumin Xu, Sheng Jin, Wentao Liu, and Chen Change Loy. Clim: Contrastive language-image mosaic for region representation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6117-6125, 2024. 6, 7 +[59] Bin Xie, Jiale Cao, Jin Xie, Fahad Shahbaz Khan, and Yanwei Pang. Sed: A simple encoder-decoder for open-vocabulary semantic segmentation, 2023. 2, 5, 6 +[60] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2955-2966, 2023. 3, 5, 6 +[61] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 3, 5 +[62] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xiang Bai. Side adapter network for open-vocabulary semantic segmentation. In Proceedings of the IEEE/CVF Conference + +on Computer Vision and Pattern Recognition, pages 2945-2954, 2023. 2, 3, 5, 6, 7 +[63] Haochen Xue, Feilong Tang, Ming Hu, Yexin Liu, Qidong Huang, Yulong Li, Chengzhi Liu, Zhongxing Xu, Chong Zhang, Chun-Mei Feng, et al. Mmrc: A large-scale benchmark for understanding multimodal large language model in real-world conversation. arXiv preprint arXiv:2502.11903, 2025. 1 +[64] Menglin Yang, Aosong Feng, Bo Xiong, Jihong Liu, Irwin King, and Rex Ying. Hyperbolic fine-tuning for large language models. arXiv preprint arXiv:2410.04010, 2024. 4, 6, 7, 8 +[65] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. In NeurIPS, 2023. 2, 3, 5, 6 +[66] Zhen Yu, Toan Nguyen, Yaniv Gal, Lie Ju, Shekhar S Chandra, Lei Zhang, Paul Bonnington, Victoria Mar, Zhiyong Wang, and Zongyuan Ge. Skin lesion recognition with class-hierarchy regularized hyperbolic embeddings. In International conference on medical image computing and computer-assisted intervention, pages 594-603. Springer, 2022. 3 +[67] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127:302-321, 2019. 6, 7 +[68] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712. Springer, 2022. 3 +[69] Ziqin Zhou, Yinjie Lei, Bowen Zhang, Lingqiao Liu, and Yifan Liu. Zegclip: Towards adapting clip for zero-shot semantic segmentation. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5 +[70] Yudong Zhu, Di Zhou, Jinghui Xiao, Xin Jiang, Xiao Chen, and Qun Liu. Hypertext: Endowing fasttext with hyperbolic geometry. arXiv preprint arXiv:2010.16143, 2020. 3 \ No newline at end of file diff --git a/2025/Understanding Fine-tuning CLIP for Open-vocabulary Semantic Segmentation in Hyperbolic Space/images.zip b/2025/Understanding Fine-tuning CLIP for Open-vocabulary Semantic Segmentation in Hyperbolic Space/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..34e23a821e49992c0a19276cbf2e38dec9e1de38 --- /dev/null +++ b/2025/Understanding Fine-tuning CLIP for Open-vocabulary Semantic Segmentation in Hyperbolic Space/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4105ef9546bc3b92bf77e9bb3546e59ec486b388233dfe58f41bd54df45ca873 +size 677525 diff --git a/2025/Understanding Fine-tuning CLIP for Open-vocabulary Semantic Segmentation in Hyperbolic Space/layout.json b/2025/Understanding Fine-tuning CLIP for Open-vocabulary Semantic Segmentation in Hyperbolic Space/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fb0d5dd34d087c1d362ef8b6bc8b3fcae0c995de --- /dev/null +++ b/2025/Understanding Fine-tuning CLIP for Open-vocabulary Semantic Segmentation in Hyperbolic Space/layout.json @@ -0,0 +1,8986 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 102, + 555, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 102, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 555, + 140 + ], + "type": "text", + "content": "Understanding Fine-tuning CLIP for Open-vocabulary Semantic Segmentation in Hyperbolic Space" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 160, + 482, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 160, + 482, + 190 + ], + "spans": [ + { + "bbox": [ + 127, + 160, + 482, + 190 + ], + "type": "text", + "content": "Zelin Peng" + }, + { + "bbox": [ + 127, + 160, + 482, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 160, + 482, + 190 + ], + "type": "text", + "content": ", Zhengqin Xu" + }, + { + "bbox": [ + 127, + 160, + 482, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 160, + 482, + 190 + ], + "type": "text", + "content": ", Zhilin Zeng" + }, + { + "bbox": [ + 127, + 160, + 482, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 160, + 482, + 190 + ], + "type": "text", + "content": ", Changsong Wen" + }, + { + "bbox": [ + 127, + 160, + 482, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 160, + 482, + 190 + ], + "type": "text", + "content": ", Yu Huang" + }, + { + "bbox": [ + 127, + 160, + 482, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 160, + 482, + 190 + ], + "type": "text", + "content": ", Menglin Yang" + }, + { + "bbox": [ + 127, + 160, + 482, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 127, + 160, + 482, + 190 + ], + "type": "text", + "content": ", Feilong Tang" + }, + { + "bbox": [ + 127, + 160, + 482, + 190 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 127, + 160, + 482, + 190 + ], + "type": "text", + "content": ", and Wei Shen" + }, + { + "bbox": [ + 127, + 160, + 482, + 190 + ], + "type": "inline_equation", + "content": "^{1(\\boxtimes)}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 190, + 542, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 190, + 542, + 219 + ], + "spans": [ + { + "bbox": [ + 67, + 190, + 542, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 190, + 542, + 219 + ], + "type": "text", + "content": "MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University " + }, + { + "bbox": [ + 67, + 190, + 542, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 190, + 542, + 219 + ], + "type": "text", + "content": "Hong Kong University of Science and Technology (Guangzhou) " + }, + { + "bbox": [ + 67, + 190, + 542, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 190, + 542, + 219 + ], + "type": "text", + "content": "MBZUAI, " + }, + { + "bbox": [ + 67, + 190, + 542, + 219 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 67, + 190, + 542, + 219 + ], + "type": "text", + "content": "Monash University." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 175, + 220, + 431, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 220, + 431, + 232 + ], + "spans": [ + { + "bbox": [ + 175, + 220, + 431, + 232 + ], + "type": "text", + "content": "https://github.com/SJTU-DeepVisionLab/HyperCLIP" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 271, + 200, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 271, + 200, + 284 + ], + "spans": [ + { + "bbox": [ + 151, + 271, + 200, + 284 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 296, + 297, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 296, + 297, + 632 + ], + "spans": [ + { + "bbox": [ + 54, + 296, + 297, + 632 + ], + "type": "text", + "content": "CLIP, a foundational vision-language model, has emerged as a powerful tool for open-vocabulary semantic segmentation. While freezing the text encoder preserves its powerful embeddings, recent studies show that fine-tuning both the text and image encoders jointly significantly enhances segmentation performance, especially for classes from open sets. In this work, we explain this phenomenon from the perspective of hierarchical alignment, since during fine-tuning, the hierarchy level of image embeddings shifts from image-level to pixel-level. We achieve this by leveraging hyperbolic space, which naturally encoders hierarchical structures. Our key observation is that, during fine-tuning, the hyperbolic radius of CLIP's text embeddings decreases, facilitating better alignment with the pixel-level hierarchical structure of visual data. Building on this insight, we propose HyperCLIP, a novel fine-tuning strategy that adjusts the hyperbolic radius of the text embeddings through scaling transformations. By doing so, HyperCLIP equips CLIP with segmentation capability while introducing only a small number of learnable parameters. Our experiments demonstrate that HyperCLIP achieves state-of-the-art performance on open-vocabulary semantic segmentation tasks across three benchmarks, while fine-tuning only approximately " + }, + { + "bbox": [ + 54, + 296, + 297, + 632 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 54, + 296, + 297, + 632 + ], + "type": "text", + "content": " of the total parameters of CLIP. More importantly, we observe that after adjustment, CLIP's text embeddings exhibit a relatively fixed hyperbolic radius across datasets, suggesting that the granularity required for this segmentation task might be quantified using the hyperbolic radius." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 652, + 135, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 135, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 135, + 665 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 673, + 295, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 295, + 697 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 295, + 697 + ], + "type": "text", + "content": "The goal of open-vocabulary semantic segmentation is to develop a segmentation model capable of labeling each" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 321, + 261, + 547, + 487 + ], + "blocks": [ + { + "bbox": [ + 321, + 261, + 547, + 487 + ], + "lines": [ + { + "bbox": [ + 321, + 261, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 321, + 261, + 547, + 487 + ], + "type": "image", + "image_path": "ce1ce3e1555f07c3dd7048c05431bbe51de4cea6ceaa505bb78707e8d71bf6a0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 498, + 556, + 621 + ], + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 621 + ], + "type": "text", + "content": "Figure 1. Visualization of hyperbolic space by a Poincaré ball model: The height of points reflects the hierarchical level of input data. The black segments refer to the shortest path between the points in the Poincaré ball model. The red arrow line denotes the hyperbolic radius of text embeddings. Embeddings closer to the center have a smaller radius, representing low-level visual information (e.g., pixels), while points farther from the center have a larger radius, representing high-level visual information (e.g., images). During fine-tuning, the hyperbolic radius of text embeddings is adjusted to match the radius associated with pixels, thereby endowing CLIP with segmentation ability (Best viewed in color)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 642, + 558, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 558, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 558, + 715 + ], + "type": "text", + "content": "pixel in an image with categories that extend beyond a predefined closed set, based on textual descriptions. Vision-language foundation models [7, 12, 13, 15, 22, 23, 26, 28, 32–35, 39, 43, 44, 46, 51, 52, 55, 63], particularly CLIP [46], are frequently employed to provide open-vocabulary recognition capabilities. Consequently, open-vocabulary seman" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 702, + 253, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 702, + 253, + 713 + ], + "spans": [ + { + "bbox": [ + 69, + 702, + 253, + 713 + ], + "type": "text", + "content": "Corresponding Author: wei.shen@sjtu.edu.cn" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4562" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 109 + ], + "type": "text", + "content": "tic segmentation fundamentally involves transferring these vision-language foundation models, originally trained with image-level supervision, to achieve pixel-level predictions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 110, + 298, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 110, + 298, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 110, + 298, + 361 + ], + "type": "text", + "content": "Current methods [9, 59, 62, 65] typically fine-tune CLIP on a closed set with segmentation annotations, i.e., COCO [4], to equip it with the segmentation capability. A prevailing opinion is that simply freezing CLIP's text encoder maximally preserves the powerful text embeddings of a vast array of classes, which is believed to be beneficial for good generalization on open sets. Many studies follow this opinion [31, 37, 62]. However, recent studies [9, 42, 59] offer an alternative perspective on fine-tuning: simultaneously fine-tuning both the image and text encoders results in superior segmentation performance on open sets compared to former approaches. To explain and resolve this seemingly paradoxical phenomenon, we first leverage a non-Euclidean manifold—specifically, hyperbolic space [5]—to understand why fine-tuning the text encoder appears to improve segmentation performance on open sets. Building on this understanding, our objective is to design an explainable fine-tuning strategy with a minimal number of introduced tunable parameters, which is able to equip CLIP with strong segmentation performance while minimizing the loss of its generalization ability." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 364, + 297, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 364, + 297, + 628 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 297, + 628 + ], + "type": "text", + "content": "Existing research [1] has found that images inherently exhibit hierarchical structures, consisting of multiple levels: pixels, patches, objects, entire scenes, etc. Given that, in open-vocabulary semantic segmentation, CLIP's image embeddings transition from image level to pixel level, we infer that the performance improvement brought by the recent studies [9, 59] could be attributed to adjusting CLIP's text embeddings from their original hierarchical level to better align with the level represented by pixels, thereby enhancing cross-modal alignment. However, most existing fine-tuning methods operate in Euclidean space, which makes it challenging to quantify such a transition through different hierarchical levels. Hyperbolic spaces have gained significant interest in recent years, owing to their ability to naturally and compactly encode hierarchical structures [47, 48, 57], as shown in Fig. 1. The hyperbolic radius, defined as the distance from a point to the center in hyperbolic space, represents its hierarchical level [50]. Points closer to the center (smaller radius) correspond to low-level visual embeddings, e.g., those from pixels, while points near the edge (larger radius) correspond to high-level visual embeddings, e.g., those from images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": "Motivated by this, we project the text embeddings into hyperbolic space and visualize the changes in their radii before and after fine-tuning, as shown in Fig. 2. This fine-tuning is realized using the state-of-the-art method provided in [9]. Specifically, one can observe that the radius after fine-tuning is smaller than that of pre-trained CLIP. In particular, we identify a pattern during fine-tuning: the hierarchical level of" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 314, + 68, + 555, + 210 + ], + "blocks": [ + { + "bbox": [ + 314, + 68, + 555, + 210 + ], + "lines": [ + { + "bbox": [ + 314, + 68, + 555, + 210 + ], + "spans": [ + { + "bbox": [ + 314, + 68, + 555, + 210 + ], + "type": "image", + "image_path": "ae42f7ff7a74c0af75347c3ac569e41b75208e3d2c72093f57eb1f21128d9361.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 218, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 313, + 218, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 555, + 285 + ], + "type": "text", + "content": "Figure 2. Understanding fine-tuning CLIP's text encoder from the perspective of hyperbolic radius. We use hyperbolic radius to illustrate the change in the hierarchical level of 171 classes after fine-tuning. We observe that the hyperbolic radius of each class becomes smaller, suggesting that adjusting the hyperbolic radius equips CLIP with segmentation ability (Best viewed in color)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 304, + 555, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 365 + ], + "type": "text", + "content": "the text embeddings adjusts in conjunction with the hierarchical level of the embeddings provided by the image encoder. Our finding brings a novel perspective for fine-tuning CLIP: Adjusting the hyperbolic radius of text embeddings from edge to center endowed CLIP with segmentation ability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 365, + 556, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 556, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 556, + 533 + ], + "type": "text", + "content": "Building on the analysis above, we propose a novel finetuning strategy for CLIP in hyperbolic space, termed HyperCLIP. HyperCLIP aims to directly adjust the hyperbolic radius of CLIP's embeddings to facilitate alignment with the appropriate hierarchical level for segmentation tasks. Specifically, we introduce several block-diagonal scaling matrices. Multiplying these matrices with CLIP's feature representations using a Möbius matrix multiplication operation is equivalent to performing scaling transformations that adjust the radius of CLIP's embeddings. This enables the adjustment of their hierarchical levels, ultimately enhancing the model's performance. As a result, HyperCLIP can equip CLIP with segmentation ability with only a small number of introduced learnable parameters." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 533, + 557, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 557, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 557, + 641 + ], + "type": "text", + "content": "Extensive experiments demonstrate that our method sets new state-of-the-art results in open-vocabulary semantic segmentation across three benchmarks, while fine-tuning only approximately " + }, + { + "bbox": [ + 313, + 533, + 557, + 641 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 313, + 533, + 557, + 641 + ], + "type": "text", + "content": " of the total parameters of CLIP. More importantly, we observe that after adjustment, CLIP's text embeddings maintain a relatively fixed hyperbolic radius across different datasets. This suggests that the granularity of the segmentation task might be quantified using the hyperbolic radius, warranting further exploration." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 651, + 400, + 663 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 651, + 400, + 663 + ], + "spans": [ + { + "bbox": [ + 314, + 651, + 400, + 663 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 671, + 531, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 531, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 531, + 684 + ], + "type": "text", + "content": "2.1. Open-vocabulary Semantic Segmentation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "Prior works on open-vocabulary semantic segmentation typically tackle this task by leveraging CLIP [46]. Early efforts," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4563" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 298, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 298, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 298, + 360 + ], + "type": "text", + "content": "e.g., [68], directly fine-tune CLIP on mainstream segmentation datasets, like COCO [4]. However, these approaches argue that fully fine-tuning CLIP's encoder significantly diminishes its ability to generalize to unseen classes. To mitigate this issue, some methods [11, 17, 60, 61] take the opposite approach, fine-tuning an additional mask generator [8] for segmentation while keeping CLIP frozen to preserve its generalization capabilities. However, this approach, with its frozen parameter space, lacks segmentation awareness, leading to a misalignment between regions and text descriptions [37]. Other studies [9, 62, 65] propose more sophisticated solutions that fine-tune only specific parameters, e.g., certain layers of CLIP, enabling pixel-level predictions while keeping the majority of CLIP's parameters fixed to minimize the loss of generalization. While these approaches show significant advantages, they often rely on very small learning rates, implicitly promoting minimal deviations from the pre-trained CLIP, which limits segmentation performance. In summary, there is still limited understanding of how fine-tuning influences CLIP's representations and why fine-tuning CLIP's text encoder appears to enhance segmentation performance. In this paper, we aim to provide several new insights that reveal the underlying logic by incorporating a novel perspective: hyperbolic space." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 368, + 223, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 368, + 223, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 368, + 223, + 381 + ], + "type": "text", + "content": "2.2. Learning in Hyperbolic Spaces" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 387, + 298, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 298, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 298, + 579 + ], + "type": "text", + "content": "Unlike Euclidean geometry, hyperbolic space can be viewed as the continuous analog of a tree [2], enabling a hierarchical understanding of knowledge structures. Previous research has demonstrated the advantages of representing such hierarchical relationships in various domains, including molecular structures [66], 3D data [6, 56], text data [16, 70], and images [18, 27]. Furthermore, several studies have investigated hyperbolic learning in vision tasks to model hierarchical relationships across different modalities [10, 30, 47], showcasing the benefits of capturing the semantic hierarchy among various data types. Nevertheless, while most studies typically leverage the intrinsic hierarchical properties of hyperbolic space, this paper seeks to capture the detailed changes in semantics reflected by the hyperbolic radius. This approach potentially complements current fine-tuning efforts in open-vocabulary tasks by offering a novel hyperbolic perspective." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 590, + 198, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 198, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 198, + 604 + ], + "type": "text", + "content": "3. Preliminary Background" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 611, + 178, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 178, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 178, + 624 + ], + "type": "text", + "content": "3.1. Hyperbolic geometry" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 630, + 298, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 298, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 298, + 715 + ], + "type": "text", + "content": "Hyperbolic geometry often has constant negative curvature and is fundamentally different to the Euclidean or spherical geometry which has zero or constant positive curvature, respectively. This enables unique properties such as the divergence of parallel lines and the exponential volume growth towards the boundary. Hyperbolic geometry has five different analytic models to build the hyperbolic space [5]. In this" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 556, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 97 + ], + "type": "text", + "content": "work, we utilize the classical Poincaré ball model as its effectiveness in capturing hierarchical structures [16, 20, 41, 47]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 106, + 454, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 106, + 454, + 118 + ], + "spans": [ + { + "bbox": [ + 313, + 106, + 454, + 118 + ], + "type": "text", + "content": "3.2. The Poincaré Ball Model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 124, + 556, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 124, + 556, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 124, + 556, + 307 + ], + "type": "text", + "content": "The Poincaré ball model " + }, + { + "bbox": [ + 313, + 124, + 556, + 307 + ], + "type": "inline_equation", + "content": "(\\mathbb{D}_c^n,g^{\\mathbb{D}_c})" + }, + { + "bbox": [ + 313, + 124, + 556, + 307 + ], + "type": "text", + "content": " with radius " + }, + { + "bbox": [ + 313, + 124, + 556, + 307 + ], + "type": "inline_equation", + "content": "1 / \\sqrt{c}" + }, + { + "bbox": [ + 313, + 124, + 556, + 307 + ], + "type": "text", + "content": " and negative curvature " + }, + { + "bbox": [ + 313, + 124, + 556, + 307 + ], + "type": "inline_equation", + "content": "-c(c > 0)" + }, + { + "bbox": [ + 313, + 124, + 556, + 307 + ], + "type": "text", + "content": " is a Riemannian manifold " + }, + { + "bbox": [ + 313, + 124, + 556, + 307 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_c^n \\coloneqq \\{\\mathbf{x} \\in \\mathbb{R}^n : c\\|\\mathbf{x}\\| < 1\\}" + }, + { + "bbox": [ + 313, + 124, + 556, + 307 + ], + "type": "text", + "content": " with the metric " + }, + { + "bbox": [ + 313, + 124, + 556, + 307 + ], + "type": "inline_equation", + "content": "g^{\\mathbb{D}_c} = \\lambda_{c,\\mathbf{x}}^2 g^E" + }, + { + "bbox": [ + 313, + 124, + 556, + 307 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 124, + 556, + 307 + ], + "type": "inline_equation", + "content": "\\lambda_{c,\\mathbf{x}} = \\frac{2}{1 - c\\|\\mathbf{x}\\|^2}" + }, + { + "bbox": [ + 313, + 124, + 556, + 307 + ], + "type": "text", + "content": " is the conformal factor, and " + }, + { + "bbox": [ + 313, + 124, + 556, + 307 + ], + "type": "inline_equation", + "content": "g^E = \\mathbf{I}_n" + }, + { + "bbox": [ + 313, + 124, + 556, + 307 + ], + "type": "text", + "content": " is the Euclidean metric tensor. The vectorial structure operations defined in Euclidean space can not be directly used in hyperbolic space due to the differential of the space properties. Therefore, we introduce the framework of gyrovector spaces [53, 54] to provide non-associative algebraic operations for hyperbolic space. According to special relativity, when adding speed vectors belonging to the Poincaré ball of radius " + }, + { + "bbox": [ + 313, + 124, + 556, + 307 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 124, + 556, + 307 + ], + "type": "text", + "content": ", these added speed vectors remain in the ball by the gyrovector spaces operation, also known as the Möbius operation. For instance, the addition and multiplication operations are defined as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 310, + 555, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 310, + 555, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 555, + 334 + ], + "type": "text", + "content": "Mobius addition operation. For " + }, + { + "bbox": [ + 313, + 310, + 555, + 334 + ], + "type": "inline_equation", + "content": "\\mathbf{x},\\mathbf{y}\\in \\mathbb{D}_c^n" + }, + { + "bbox": [ + 313, + 310, + 555, + 334 + ], + "type": "text", + "content": " , the Mobius addition operation is defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 318, + 343, + 555, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 343, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 318, + 343, + 555, + 371 + ], + "type": "interline_equation", + "content": "\\mathbf {x} \\oplus_ {c} \\mathbf {y} := \\frac {(1 + 2 c \\langle \\mathbf {x} , \\mathbf {y} \\rangle + c \\| \\mathbf {y} \\| ^ {2}) \\mathbf {x} + (1 - c \\| \\mathbf {x} \\| ^ {2}) \\mathbf {y}}{1 + 2 c \\langle \\mathbf {x} , \\mathbf {y} \\rangle + c ^ {2} \\| \\mathbf {x} \\| ^ {2} \\| \\mathbf {y} \\| ^ {2}}. \\tag {1}", + "image_path": "6f5454361e0b60228b8a0eb2a59aaa310e28d5a76898b8b90c233059cea95453.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 384, + 556, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 556, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 556, + 419 + ], + "type": "text", + "content": "Mobius scalar multiplication operation. For a scalar " + }, + { + "bbox": [ + 313, + 384, + 556, + 419 + ], + "type": "inline_equation", + "content": "r \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 384, + 556, + 419 + ], + "type": "text", + "content": " and a vector " + }, + { + "bbox": [ + 313, + 384, + 556, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{D}_c^n" + }, + { + "bbox": [ + 313, + 384, + 556, + 419 + ], + "type": "text", + "content": ", the Möbius scalar multiplication operation is defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 337, + 429, + 555, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 429, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 337, + 429, + 555, + 453 + ], + "type": "interline_equation", + "content": "r \\otimes_ {c} \\mathbf {x} := (1 / \\sqrt {c}) \\tanh \\left(r \\tanh ^ {- 1} (\\sqrt {c} \\| \\mathbf {x} \\|\\right)) \\frac {\\mathbf {x}}{\\| \\mathbf {x} \\|}. \\tag {2}", + "image_path": "e2cebbde0837b1f9eca642d03fa8e12b20d0daa399a76266b8c14a11fb96a9a5.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 462, + 556, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 556, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 556, + 499 + ], + "type": "text", + "content": "The operation satisfies " + }, + { + "bbox": [ + 313, + 462, + 556, + 499 + ], + "type": "inline_equation", + "content": "r \\otimes_{c} \\mathbf{0} \\coloneqq \\mathbf{0}" + }, + { + "bbox": [ + 313, + 462, + 556, + 499 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 462, + 556, + 499 + ], + "type": "inline_equation", + "content": "r \\otimes_{c} \\mathbf{x} = \\mathbf{x} \\oplus_{c} \\dots \\oplus_{c} \\mathbf{x}" + }, + { + "bbox": [ + 313, + 462, + 556, + 499 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 462, + 556, + 499 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 462, + 556, + 499 + ], + "type": "text", + "content": " additions), " + }, + { + "bbox": [ + 313, + 462, + 556, + 499 + ], + "type": "inline_equation", + "content": "(r' + r) \\otimes_{c} \\mathbf{x} = r \\otimes \\mathbf{x} \\oplus_{c} r' \\otimes_{c} \\mathbf{x}" + }, + { + "bbox": [ + 313, + 462, + 556, + 499 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 462, + 556, + 499 + ], + "type": "inline_equation", + "content": "(r'r) \\otimes_{c} \\mathbf{x} = r \\otimes_{c} (r' \\otimes_{c} \\mathbf{x})" + }, + { + "bbox": [ + 313, + 462, + 556, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 462, + 556, + 499 + ], + "type": "inline_equation", + "content": "\\left(\\left|r\\right| \\otimes_{c} \\mathbf{x}\\right) / \\left\\|r \\otimes_{c} \\mathbf{x}\\right\\| = \\mathbf{x} / \\left\\|\\mathbf{x}\\right\\|" + }, + { + "bbox": [ + 313, + 462, + 556, + 499 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 501, + 556, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 556, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 556, + 549 + ], + "type": "text", + "content": "Mobius matrix multiplication operation. Refer to the relation definition in [27], for a matrix " + }, + { + "bbox": [ + 313, + 501, + 556, + 549 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\mathbb{R}^{n \\times n}" + }, + { + "bbox": [ + 313, + 501, + 556, + 549 + ], + "type": "text", + "content": " and a vector " + }, + { + "bbox": [ + 313, + 501, + 556, + 549 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{D}_c^n" + }, + { + "bbox": [ + 313, + 501, + 556, + 549 + ], + "type": "text", + "content": ", if " + }, + { + "bbox": [ + 313, + 501, + 556, + 549 + ], + "type": "inline_equation", + "content": "\\mathbf{Mx} \\neq \\mathbf{0}" + }, + { + "bbox": [ + 313, + 501, + 556, + 549 + ], + "type": "text", + "content": ", the Möbius matrix multiplication operation is defined as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 558, + 552, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 558, + 552, + 586 + ], + "spans": [ + { + "bbox": [ + 315, + 558, + 552, + 586 + ], + "type": "interline_equation", + "content": "\\mathbf {M} \\otimes_ {c} \\mathbf {x} := (\\frac {1}{\\sqrt {c}}) \\tanh \\left(\\frac {\\| \\mathbf {M x} \\|}{\\| \\mathbf {x} \\|} \\tanh ^ {- 1} (\\sqrt {c} \\| \\mathbf {x} \\|)\\right) \\frac {\\mathbf {M x}}{\\| \\mathbf {M x} \\|}.", + "image_path": "f4a90e5127307b2c01f9cba5ec9e465cf481c5eae6cdcb745142d2b40b719748.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 477, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 477, + 607 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 477, + 607 + ], + "type": "text", + "content": "The distance of " + }, + { + "bbox": [ + 313, + 594, + 477, + 607 + ], + "type": "inline_equation", + "content": "\\mathbf{x},\\mathbf{y}\\in \\mathbb{D}_c^n" + }, + { + "bbox": [ + 313, + 594, + 477, + 607 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 342, + 616, + 555, + 631 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 616, + 555, + 631 + ], + "spans": [ + { + "bbox": [ + 342, + 616, + 555, + 631 + ], + "type": "interline_equation", + "content": "d _ {c} ^ {\\mathbb {D}} (\\mathbf {x}, \\mathbf {y}) = (2 / \\sqrt {c}) \\tanh ^ {- 1} (\\sqrt {c} \\| - \\mathbf {x} \\oplus_ {c} \\mathbf {y} \\Vert). \\tag {3}", + "image_path": "46bd16904bc621fa6af0200b43747203856d6dac93971dcad0aebe82b5054cff.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "The tangent space " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathbf{x}}^{c}\\mathbb{D}_{c}^{n}" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " at a point " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\in \\mathbb{D}_c^n" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " is the first order approximation of " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_c^n" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " , which is an " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " -dimensional Euclidean space. The tangent space " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathbf{x}}^{c}\\mathbb{D}_{c}^{n}" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_c^n" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " are mapped to each other by exponential " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "(\\mathcal{T}_{\\mathbf{x}}^{c}\\mathbb{D}_{c}^{n}\\mapsto \\mathbb{D}_{c}^{n}:\\exp_{\\mathbf{x}}^{\\mathbb{D},c}(\\cdot))" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " and logarithmic " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "(\\mathbb{D}_c^n\\mapsto \\mathcal{T}_x^c\\mathbb{D}_c^n:\\log_{\\mathbf{x}}^{\\mathbb{D},c}(\\cdot))" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " maps, respectively. For any " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x},\\mathbf{y}\\in \\mathbb{D}_c^n" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{v}\\in \\mathcal{T}_{\\mathbf{x}}^{c}\\mathbb{D}_{c}^{n}" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " , the mapping functions" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4564" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 553, + 194 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 553, + 194 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 553, + 194 + ], + "type": "image", + "image_path": "96a21682e611343295ba4c39cd67d5c983ea22b86ce529b120b03380e9df88be.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 204, + 555, + 228 + ], + "lines": [ + { + "bbox": [ + 54, + 204, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 54, + 204, + 555, + 228 + ], + "type": "text", + "content": "Figure 3. A schematic representation of HyperCLIP. In the HyperCLIP framework, we fine-tune CLIP in hyperbolic space by minimizing its embeddings' hyperbolic radii using several tunable block-diagonal scaling matrices." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 247, + 198, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 198, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 198, + 261 + ], + "type": "text", + "content": "are given for " + }, + { + "bbox": [ + 55, + 247, + 198, + 261 + ], + "type": "inline_equation", + "content": "\\mathbf{v} \\neq \\mathbf{0}" + }, + { + "bbox": [ + 55, + 247, + 198, + 261 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 247, + 198, + 261 + ], + "type": "inline_equation", + "content": "\\mathbf{y} \\neq \\mathbf{x}" + }, + { + "bbox": [ + 55, + 247, + 198, + 261 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 270, + 296, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 270, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 63, + 270, + 296, + 297 + ], + "type": "interline_equation", + "content": "\\exp_ {\\mathbf {x}} ^ {\\mathbb {D}, c} (\\mathbf {v}) = \\mathbf {x} \\oplus_ {c} \\left(\\tanh \\left(\\sqrt {c} \\frac {\\lambda_ {c , \\mathbf {x}} \\| \\mathbf {v} \\|}{2}\\right) \\frac {\\mathbf {v}}{\\sqrt {c} \\| \\mathbf {v} \\|}\\right), \\tag {4}", + "image_path": "f261f32b921987931ac1bba7557dfb09f8b1cb29cc52a53c450d419a1748ce7f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 299, + 301, + 336 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 301, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 301, + 336 + ], + "type": "interline_equation", + "content": "\\log_ {\\mathbf {x}} ^ {\\mathbb {D}, c} (\\mathbf {y}) = \\frac {2}{\\sqrt {c} \\lambda_ {c , \\mathbf {x}}} \\tanh ^ {- 1} \\left(\\sqrt {c} \\| - \\mathbf {x} \\oplus_ {c} \\mathbf {y} \\|\\right) \\frac {- \\mathbf {x} \\oplus_ {c} \\mathbf {y}}{\\| - \\mathbf {x} \\oplus_ {c} \\mathbf {y} \\|}. \\tag {5}", + "image_path": "02ed6b148163535596f053eab66c5809f8617afc62868cb2dbe5e71963802a94.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 349, + 138, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 138, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 138, + 363 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 370, + 296, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 296, + 431 + ], + "type": "text", + "content": "In this section, we introduce HyperCLIP for open-vocabulary semantic segmentation. HyperCLIP is based on the Poincaré ball model, which aims to adjust the hyperbolic radius of CLIP's embeddings, thereby equipping CLIP with segmentation ability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 441, + 230, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 230, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 230, + 455 + ], + "type": "text", + "content": "4.1. Fine-tuning in Hyperbolic Space" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 461, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 296, + 616 + ], + "type": "text", + "content": "The core technique in HyperCLIP involves linear transformations, i.e., scaling, in the hyperbolic space. One of the primary methods [64] to learn the weights of linear transformations in hyperbolic space is hyperLoRA [64]. Given a feature embedding " + }, + { + "bbox": [ + 55, + 461, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 55, + 461, + 296, + 616 + ], + "type": "text", + "content": " in Euclidean space, this method first maps " + }, + { + "bbox": [ + 55, + 461, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 55, + 461, + 296, + 616 + ], + "type": "text", + "content": " to the hyperbolic space at a local reference point, typically the center, using the exponential map (e.g., Eq. 4). Linear transformations with learnable weight matrix " + }, + { + "bbox": [ + 55, + 461, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 55, + 461, + 296, + 616 + ], + "type": "text", + "content": " is then applied to " + }, + { + "bbox": [ + 55, + 461, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 55, + 461, + 296, + 616 + ], + "type": "text", + "content": " within the hyperbolic space to obtain the hyperbolic representation. Then, a logarithmic map (e.g., Eq. 5) is applied to map the hyperbolic representation back to the Euclidean space. Formally, in the Poincaré ball model, this process is realized as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 627, + 296, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 627, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 107, + 627, + 296, + 643 + ], + "type": "interline_equation", + "content": "\\mathbf {W} \\mathbf {z} = \\log_ {\\mathbf {0}} ^ {\\mathbb {D}, c} \\left(\\mathbf {W} \\otimes_ {c} \\exp_ {\\mathbf {0}} ^ {\\mathbb {D}, c} (\\mathbf {z})\\right), \\tag {6}", + "image_path": "14f82480d00121217641a1c372b0e113ba4b105c5c56ef5c7bc904004b10e765.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\otimes_{c}" + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": " is the mobius matrix multiplication operation. The primary aim of this method [64] is to capture more intricate hierarchical relationships within hyperbolic space. In contrast, our goal is to directly adjust the hyperbolic radius of CLIP's embeddings, leading to a scaling transformation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 247, + 484, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 484, + 260 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 484, + 260 + ], + "type": "text", + "content": "4.2. Hyperbolic Radius Adjustment" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 266, + 555, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 555, + 397 + ], + "type": "text", + "content": "The objective of HyperCLIP is to introduce a scaling transformation to directly adjust the hyperbolic radii of CLIP's embeddings in hyperbolic space, thereby equipping CLIP with segmentation ability. Concretely, it necessitates the utilization of a diagonal matrix, denoted as " + }, + { + "bbox": [ + 313, + 266, + 555, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 266, + 555, + 397 + ], + "type": "text", + "content": ", to scale a CLIP's feature embedding " + }, + { + "bbox": [ + 313, + 266, + 555, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\in \\mathbb{R}^{b \\times d}" + }, + { + "bbox": [ + 313, + 266, + 555, + 397 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 266, + 555, + 397 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 266, + 555, + 397 + ], + "type": "text", + "content": " is the feature dimension of " + }, + { + "bbox": [ + 313, + 266, + 555, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 313, + 266, + 555, + 397 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 266, + 555, + 397 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 266, + 555, + 397 + ], + "type": "text", + "content": " is the spatial dimension of input images or number of classes. When extending this process to hyperbolic space, " + }, + { + "bbox": [ + 313, + 266, + 555, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 266, + 555, + 397 + ], + "type": "text", + "content": " can directly adjust the hyperbolic radius of " + }, + { + "bbox": [ + 313, + 266, + 555, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 313, + 266, + 555, + 397 + ], + "type": "text", + "content": " (this deviation can be found in Sec. 4.3). Specifically, we realize the diagonal matrix " + }, + { + "bbox": [ + 313, + 266, + 555, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 266, + 555, + 397 + ], + "type": "text", + "content": " in a block-wise manner," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 376, + 407, + 555, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 407, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 376, + 407, + 555, + 420 + ], + "type": "interline_equation", + "content": "\\mathbf {S} = \\operatorname {d i a g} \\left(\\mathbf {S} _ {1}, \\mathbf {S} _ {2}, \\dots , \\mathbf {S} _ {K}\\right), \\tag {7}", + "image_path": "1b1f368a52e02e5f385ce91e18c5eab33f5a9c71a964541ccb92a33212491120.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 429, + 555, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 555, + 513 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 429, + 555, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_k\\in \\mathbb{R}^{n\\times n}" + }, + { + "bbox": [ + 313, + 429, + 555, + 513 + ], + "type": "text", + "content": " denotes a square matrix of the " + }, + { + "bbox": [ + 313, + 429, + 555, + 513 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 429, + 555, + 513 + ], + "type": "text", + "content": " -th block, and the block number " + }, + { + "bbox": [ + 313, + 429, + 555, + 513 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 429, + 555, + 513 + ], + "type": "text", + "content": " is calculated by " + }, + { + "bbox": [ + 313, + 429, + 555, + 513 + ], + "type": "inline_equation", + "content": "K = d / n" + }, + { + "bbox": [ + 313, + 429, + 555, + 513 + ], + "type": "text", + "content": " . By adjusting the size of " + }, + { + "bbox": [ + 313, + 429, + 555, + 513 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 429, + 555, + 513 + ], + "type": "text", + "content": " , we can achieve a transition between naive scaling transformation " + }, + { + "bbox": [ + 313, + 429, + 555, + 513 + ], + "type": "inline_equation", + "content": "(n = 1" + }, + { + "bbox": [ + 313, + 429, + 555, + 513 + ], + "type": "text", + "content": " , i.e., modifying only the diagonal elements) and arbitrary transformation " + }, + { + "bbox": [ + 313, + 429, + 555, + 513 + ], + "type": "inline_equation", + "content": "(n = d" + }, + { + "bbox": [ + 313, + 429, + 555, + 513 + ], + "type": "text", + "content": " , i.e., fully fine-tuning in hyperbolic space). The ablation study of " + }, + { + "bbox": [ + 313, + 429, + 555, + 513 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 429, + 555, + 513 + ], + "type": "text", + "content": " is presented in Sec. 5.3." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 514, + 556, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 556, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 556, + 562 + ], + "type": "text", + "content": "Overall Fine-tuning. During the fine-tuning phase, HyperCLIP introduces a similar process to Eq. 6 is incorporated into CLIP's encoder. Specifically, HyperCLIP uses the following forward pass:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 371, + 570, + 555, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 570, + 555, + 586 + ], + "spans": [ + { + "bbox": [ + 371, + 570, + 555, + 586 + ], + "type": "interline_equation", + "content": "\\mathbf {S} \\mathbf {z} = \\log_ {\\mathbf {0}} ^ {\\mathbb {D}, c} \\left(\\mathbf {S} \\otimes_ {c} \\exp_ {\\mathbf {0}} ^ {\\mathbb {D}, c} (\\mathbf {z})\\right). \\tag {8}", + "image_path": "994311515e8a1485f12f1accd40a47e18f576fa73b983af732a7ac414d445971.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 594, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 629 + ], + "type": "text", + "content": "Concurrently, the original components of CLIP load their weights from the pre-trained checkpoint, with their parameters remaining frozen." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": "Parameter complexity. For an " + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": "-block diagonal matrix, the number of parameters is " + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "inline_equation", + "content": "d^2 / K" + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": ", resulting in a complexity of " + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{O}\\left(d^2 / K\\right)" + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": ". We can optionally share the block matrix to reduce the number of parameters, i.e., " + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_k = \\mathbf{S}_j, \\forall k \\neq j" + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": ". This reduces the parameter complexity to " + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{O}\\left(d^2 / K^2\\right)" + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": ". We can further reduce parameter complexity by setting " + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "inline_equation", + "content": "n = 1" + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": " which transforms the block-diagonal matrix into a diagonal" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4565" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 71, + 567, + 399 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 567, + 399 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 567, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 567, + 399 + ], + "type": "table", + "html": "
ModelVLMAdditional BackboneFine-tuning SpaceA-847PC-459A-150PC-59PAS-20\\( PAS-20^b \\)
Partial Fine-Tuning
ZS3Net [3]-ResNet-101E---19.438.3-
LSeg [31]CLIP ViT-B/32ResNet-101E----47.4-
OpenSeg [17]ALIGNResNet-101E4.47.917.540.1-63.8
ZegFormer [11]CLIP ViT-B/16ResNet-101E4.99.116.942.886.262.7
ZSseg [61]CLIP ViT-B/16ResNet-101E7.0-20.547.788.4-
OVSeg [37]CLIP ViT-B/16ResNet-101cE7.111.024.853.392.6-
ZegCLIP [69]CLIP ViT-B/16-E---41.293.6-
Selective Fine-Tuning
SED [59]CLIP ConvNeXt-B-E11.418.631.657.394.4-
CAT-Seg [9]CLIP ViT-B/16-E12.019.031.857.594.677.3
Parameter-efficient Fine-Tuning
SAN [62]CLIP ViT-B/16Side AdapterE10.112.627.553.894.0-
OursCLIP ViT-B/16-H12.319.232.158.595.678.9
Partial Fine-Tuning
LSeg [31]CLIP ViT-B/32ViT-L/16E----52.3-
OpenSeg [17]ALIGNEff-B7E8.111.526.444.8-70.2
OVSeg [37]CLIP ViT-L/14Swin-BE9.012.429.655.794.5-
SAN [62]CLIP ViT-L/14-E12.415.732.157.794.6-
ODISE [60]CLIP ViT-L/14Stable DiffusionE11.114.529.957.3--
FC-CLIP [65]CLIP ConvNeXt-L-E14.818.234.158.495.4-
Selective Fine-Tuning
SED [59]CLIP ConvNeXt-L-E13.922.635.260.696.1-
CAT-Seg [9]CLIP ViT-L/14-E16.023.837.963.397.082.5
Parameter-efficient Fine-Tuning
SAN [62]CLIP ViT-L/14Side AdapterE12.415.732.157.794.6-
OursCLIP ViT-L/14-H16.324.138.264.298.384.3
", + "image_path": "7fb52ec309871fbf9aeae8a6597307907b625f5b4ff2fb4f73a06a7fa4df1564.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 407, + 555, + 430 + ], + "lines": [ + { + "bbox": [ + 55, + 407, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 555, + 430 + ], + "type": "text", + "content": "Table 1. Comparison with state-of-the-art methods on standard benchmarks. The best-performing results are presented in bold, while the second-best results are underlined. \"E\": Euclidean Space. \"H\": Hyperbolic Space." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 441, + 296, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 296, + 501 + ], + "type": "text", + "content": "matrix. In this case, the number of parameters becomes " + }, + { + "bbox": [ + 55, + 441, + 296, + 501 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 441, + 296, + 501 + ], + "type": "text", + "content": ", resulting in a complexity of " + }, + { + "bbox": [ + 55, + 441, + 296, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(d)" + }, + { + "bbox": [ + 55, + 441, + 296, + 501 + ], + "type": "text", + "content": ". Despite all these strategies to improve parameter efficiency, we note that the resulting matrix " + }, + { + "bbox": [ + 55, + 441, + 296, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 55, + 441, + 296, + 501 + ], + "type": "text", + "content": " remains to realize scaling transformations, thereby allowing it to adjust the radius of CLIP's embeddings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 510, + 174, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 174, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 174, + 521 + ], + "type": "text", + "content": "4.3. Theoretical Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 528, + 296, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 296, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 296, + 600 + ], + "type": "text", + "content": "To further demonstrate the effectiveness of the proposed HyperCLIP method, we conduct a theoretical analysis of the scaling transformation. Theorem 1 reveals that the employed scaling matrix can directly adjust the hyperbolic radius, thus enabling transferring through different hierarchical levels. Detailed proof refers to the supplementary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 601, + 296, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 296, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 296, + 648 + ], + "type": "text", + "content": "Theorem 1 (Scaling). According to the definitions of the classical Poincaré Ball model and the hyperbolic tangent, for a point " + }, + { + "bbox": [ + 55, + 601, + 296, + 648 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{D}_c^n" + }, + { + "bbox": [ + 55, + 601, + 296, + 648 + ], + "type": "text", + "content": " in hyperbolic space, its hyperbolic radius is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 85, + 655, + 296, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 655, + 296, + 681 + ], + "spans": [ + { + "bbox": [ + 85, + 655, + 296, + 681 + ], + "type": "interline_equation", + "content": "\\operatorname {R a d} _ {\\mathbf {x}} := d _ {c} ^ {\\mathbb {D}} (\\mathbf {x}, \\mathbf {0}) = \\left(\\frac {2}{\\sqrt {c}}\\right) \\tanh ^ {- 1} \\left(\\sqrt {c} \\| \\mathbf {x} \\|\\right), \\tag {9}", + "image_path": "50a01d22adc46d6a52ac0bf40a0c93637fb3746ff12e4531be5d920df917730b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{0}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is the center of the hyperbolic space. Let " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is a scaling parameter, combine with Eq. 2 and Eq. 9, the radius" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 441, + 555, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 466 + ], + "type": "text", + "content": "of the point " + }, + { + "bbox": [ + 313, + 441, + 555, + 466 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_s = \\mathcal{T}_{\\mathbf{x}}(s) := s \\otimes_c \\mathbf{x}" + }, + { + "bbox": [ + 313, + 441, + 555, + 466 + ], + "type": "text", + "content": " which changed by the scaling transformation " + }, + { + "bbox": [ + 313, + 441, + 555, + 466 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathbf{x}}(s)" + }, + { + "bbox": [ + 313, + 441, + 555, + 466 + ], + "type": "text", + "content": " is obtained by:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 322, + 474, + 555, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 474, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 322, + 474, + 555, + 533 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {R a d} _ {\\mathbf {x}, s} (s) = \\frac {2}{\\sqrt {c}} \\tanh ^ {- 1} (\\sqrt {c} \\| \\mathbf {x} _ {s} \\|) \\\\ = \\frac {2}{\\sqrt {c}} \\tanh ^ {- 1} (\\sqrt {c} \\left| \\left| \\frac {\\tanh \\left(\\frac {s \\sqrt {c}}{2} \\operatorname {R a d} _ {\\mathbf {x}}\\right) \\mathbf {x}}{\\sqrt {c} \\| \\mathbf {x} \\|} \\right| \\right|. \\tag {10} \\\\ \\end{array}", + "image_path": "86c448a6b8f62fb088c08d01a4239c4dc444280eb30ffaaf5df88a8a61cc82eb.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 543, + 555, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 555, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 555, + 615 + ], + "type": "text", + "content": "It is noticed that the " + }, + { + "bbox": [ + 313, + 543, + 555, + 615 + ], + "type": "inline_equation", + "content": "\\mathrm{Rad}_{\\mathbf{x},s}(s)" + }, + { + "bbox": [ + 313, + 543, + 555, + 615 + ], + "type": "text", + "content": " satisfies the monotonicity criteria, therefore, we can adjust the hyperbolic radius by changing the scaling parameter " + }, + { + "bbox": [ + 313, + 543, + 555, + 615 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 543, + 555, + 615 + ], + "type": "text", + "content": ". Refer to the definition of the Möbius matrix multiplication operation in [27], the scaling matrix " + }, + { + "bbox": [ + 313, + 543, + 555, + 615 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 543, + 555, + 615 + ], + "type": "text", + "content": " also has the ability to adjust the hyperbolic radius via the scaling transformation " + }, + { + "bbox": [ + 313, + 543, + 555, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathbf{x}}(\\mathbf{S})" + }, + { + "bbox": [ + 313, + 543, + 555, + 615 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 627, + 395, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 627, + 395, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 627, + 395, + 640 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 647, + 430, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 430, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 430, + 659 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "Datasets. Following prior works [9, 59], we use the COCO-Stuff dataset [4] as our training set. This dataset consists of around 118,000 densely annotated images, covering 171 distinct semantic categories. For inference, we compare our" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4566" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 293, + 194 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 293, + 194 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 293, + 194 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 293, + 194 + ], + "type": "table", + "html": "
MethodsSpace for Fine-tuning\\( \\mathbf{A P^{Base}}_{50} \\)\\( \\mathbf{A P^{Novel}}_{50} \\)
CLIP-21.636.4
Other Fine-Tuning
CLIM [58]E25.742.5
CAT-Seg [9]E25.242.3
Parameter-efficient Fine-Tuning
LoRA [21]E24.441.5
HyperLoRA [64]H25.142.0
OursH26.342.7
", + "image_path": "324337a2a28aa44cbac3b98a1ec26cd585cb77f2559418d253825db387c8d8fa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 250, + 296, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 250, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 296, + 285 + ], + "type": "text", + "content": "method against state-of-the-art approaches across several semantic segmentation benchmarks, including ADE20K [67], PASCAL VOC [14], and PASCAL-Context [40]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 286, + 296, + 466 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 286, + 295, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 295, + 346 + ], + "type": "text", + "content": "- ADE20K [67] is a well-established semantic segmentation dataset with approximately 20,000 training images and 2,000 validation images. It also includes two separate test sets: A-150, consisting of 150 common categories, and A-847, which contains 847 categories." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 346, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 296, + 418 + ], + "type": "text", + "content": "- PASCAL VOC [14] is a smaller dataset for semantic segmentation, consisting of 1464 training images and 1449 validation images, covering 20 different foreground categories. We refer to this dataset as PAS-20. In line with [9], we also report scores on PAS-20" + }, + { + "bbox": [ + 55, + 346, + 296, + 418 + ], + "type": "inline_equation", + "content": "^b" + }, + { + "bbox": [ + 55, + 346, + 296, + 418 + ], + "type": "text", + "content": ", which includes the \"background\" as the 21st category." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 418, + 296, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 466 + ], + "type": "text", + "content": "- PASCAL-Context [40] is an extended version of the original PASCAL VOC dataset. It provides two test sets: PC-59 and PC-459, containing 59 and 459 categories respectively, for evaluation purposes." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 467, + 296, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 296, + 502 + ], + "type": "text", + "content": "Evaluation Metric. Consistent with prior studies [9, 59], we use mean Intersection over Union (mIoU) to evaluate the semantic segmentation performance on these benchmarks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 506, + 296, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 296, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 296, + 661 + ], + "type": "text", + "content": "Implementation Details. We implement our method using the Transformer-based CLIP model. Following the protocol established in [9], we evaluate our results on two versions of the CLIP model: ViT-B/16 and ViT-L/14. For training, we follow the same experimental setup as [9]. Besides, we use the Adam optimizer [29] with an initial learning rate of " + }, + { + "bbox": [ + 55, + 506, + 296, + 661 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 55, + 506, + 296, + 661 + ], + "type": "text", + "content": " for CLIP, and a weight decay of " + }, + { + "bbox": [ + 55, + 506, + 296, + 661 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 55, + 506, + 296, + 661 + ], + "type": "text", + "content": ". Training is conducted with one image per mini-batch. We apply the block-diagonal learnable weight in the self-attention module in each layer of CLIP. We employ the cost-based approach from [9] as our decoder. The curvature " + }, + { + "bbox": [ + 55, + 506, + 296, + 661 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 506, + 296, + 661 + ], + "type": "text", + "content": " of Poincaré ball model is set to 0.01. All models are trained for 80,000 iterations using 4 NVIDIA A800 GPUs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 670, + 282, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 282, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 282, + 683 + ], + "type": "text", + "content": "5.2. Comparisons with State-of-the-art Methods" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Here, we compare our proposed method with several state-of-the-art methods, as shown in Table 1, using six test sets" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 547, + 99 + ], + "blocks": [ + { + "bbox": [ + 55, + 201, + 295, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 201, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 295, + 236 + ], + "type": "text", + "content": "Table 2. Comparisons on open-vocabulary object detection task on the COCO [38] dataset. Results are measured by the box AP at IoU threshold 0.5. \"E\": Euclidean space. \"H\": hyperbolic space." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 70, + 547, + 99 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 547, + 99 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 547, + 99 + ], + "type": "table", + "html": "
MethodODISE [60]FC-CLIP [65]MAFT+ [25]Ours
PQ23.426.827.129.2
", + "image_path": "c2da7c2fe3613642e2dc3724d396950cc2a84a5db63bf079730aac22d3811f11.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 316, + 134, + 553, + 247 + ], + "blocks": [ + { + "bbox": [ + 314, + 106, + 555, + 129 + ], + "lines": [ + { + "bbox": [ + 314, + 106, + 555, + 129 + ], + "spans": [ + { + "bbox": [ + 314, + 106, + 555, + 129 + ], + "type": "text", + "content": "Table 3. Comparisons on open-vocabulary panoptic segmentation task on the ADE20K [67] dataset. \"PQ\": Panoptic quality." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 134, + 553, + 247 + ], + "lines": [ + { + "bbox": [ + 316, + 134, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 316, + 134, + 553, + 247 + ], + "type": "table", + "html": "
nParam. (M)A-847PC-459A-150PC-59PAS-20b
10.045.99.427.452.474.9
80.3510.115.429.954.275.6
160.7011.417.431.055.976.2
321.4111.718.131.756.676.7
642.8111.918.732.057.877.8
1285.6312.319.232.158.578.9
25611.2512.219.031.958.879.3
", + "image_path": "3a336029fa794eee3636f82615c109d3e74c800c9e566bc404e32391641f4c42.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 325, + 299, + 541, + 342 + ], + "blocks": [ + { + "bbox": [ + 313, + 255, + 553, + 287 + ], + "lines": [ + { + "bbox": [ + 313, + 255, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 553, + 287 + ], + "type": "text", + "content": "Table 4. Ablation on the size of diagonal blocks " + }, + { + "bbox": [ + 313, + 255, + 553, + 287 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 255, + 553, + 287 + ], + "type": "text", + "content": ". The base model is ViT-B/16. The best-performing results are presented in bold, while the second-best results are underlined." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 299, + 541, + 342 + ], + "lines": [ + { + "bbox": [ + 325, + 299, + 541, + 342 + ], + "spans": [ + { + "bbox": [ + 325, + 299, + 541, + 342 + ], + "type": "table", + "html": "
MethodsOVSeg [37]CAT-Seg [9]SAN [62]Ours
Param. (M)147.239.58.45.6
", + "image_path": "904e30cee89766e681aab0fb2396a7fd07f299c9e9a583cbba1ed8237bb1dada.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 349, + 553, + 371 + ], + "lines": [ + { + "bbox": [ + 313, + 349, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 553, + 371 + ], + "type": "text", + "content": "Table 5. Efficiency comparison in terms of learnable parameters for fine-tuning CLIP. The base model is ViT-B/16." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 389, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 665 + ], + "type": "text", + "content": "across three benchmarks. Overall, we achieve the best results. Most existing open-vocabulary semantic segmentation methods follow partial fine-tuning strategies, i.e., fine-tuning CLIP's image encoder. Although these methods provide sufficient flexibility for aligning with text descriptions generated by the text encoder, they fail to adjust the text embeddings to the appropriate hierarchical level for effective text-to-image alignment, often leading to suboptimal segmentation performance. Differently, CAT-Seg [9] simultaneously fine-tunes both the text encoder and image encoder of CLIP, achieving performance comparable to ours on some of the datasets. However, its fine-tuning scheme is manually controlled through different layer combinations, necessitating a careful design to balance generalization and segmentation ability, while ours does not suffer from such an issue. Then, compared to SAN [62], another parameter-efficient fine-tuning method that introduces only a limited number of tunable parameters, our approach significantly outperforms it, achieving improvements of " + }, + { + "bbox": [ + 313, + 389, + 555, + 665 + ], + "type": "inline_equation", + "content": "6.6\\%" + }, + { + "bbox": [ + 313, + 389, + 555, + 665 + ], + "type": "text", + "content": " on the PC-459 dataset and " + }, + { + "bbox": [ + 313, + 389, + 555, + 665 + ], + "type": "inline_equation", + "content": "4.7\\%" + }, + { + "bbox": [ + 313, + 389, + 555, + 665 + ], + "type": "text", + "content": " on the PC-59 dataset with ViT-B/16 as the base model. These results demonstrate the effectiveness of our method in preserving generalization while mastering segmentation capability." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "Efficiency comparison. We compare the efficiency of our method with other approaches, including OVSeg [37], CATSeg [9], and SAN [62], all of which utilize a ViT version of CLIP. The comparison, summarized in Table 5, shows" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4567" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 555, + 312 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 312 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 312 + ], + "type": "image", + "image_path": "d68b8761010aa08d10241ef82ece80946be3d996a1d0e0aae541bc1df5465765.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 318, + 555, + 341 + ], + "lines": [ + { + "bbox": [ + 55, + 318, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 555, + 341 + ], + "type": "text", + "content": "Figure 4. Comparison of qualitative results on ADE20K [67] with 847 categories. We compare our method with other two state-of-the-art methods, i.e., SAN [62] and CAT-Seg [9]. The results show our method performs more precise segmentation in different scenarios." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 78, + 351, + 529, + 519 + ], + "blocks": [ + { + "bbox": [ + 78, + 351, + 529, + 519 + ], + "lines": [ + { + "bbox": [ + 78, + 351, + 529, + 519 + ], + "spans": [ + { + "bbox": [ + 78, + 351, + 529, + 519 + ], + "type": "table", + "html": "
MethodImage encoderText encoderParam. (M)A-847PC-459A-150PC-59PAS-20PAS-20b
FreezeXX04.46.624.849.492.571.9
Parameter-efficient Fine-Tuning
LoRA [21]✓✓7.511.417.628.655.194.276.7
OFT [45]✓✓3.810.918.030.253.793.774.3
VPT [24]✓✓2.25.710.223.754.393.875.1
Adapter [19]✓✓7.510.416.528.854.994.275.2
SSF [36]✓✓0.66.915.228.652.193.272.8
HyperLoRA [64]✓✓7.511.418.229.755.894.377.2
HyperCLIP (Ours)✓X3.410.316.830.557.694.478.0
X✓2.211.117.529.855.193.776.1
✓✓5.612.319.232.158.595.678.9
", + "image_path": "024972f2c32ef18df7d932f78b74259fff08062197a1e9d35d974c050f025c6e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 525, + 555, + 570 + ], + "lines": [ + { + "bbox": [ + 54, + 525, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 54, + 525, + 555, + 570 + ], + "type": "text", + "content": "Table 6. Ablation study on the fine-tuning encoder in HyperCLIP and comparison of universal PEFT methods, including \"LoRA\" [21], \"OFT\" [45], \"VPT\" [24], \"Adapter\" [19], \"LST\" [49], \"SSF\" [36] and \"HyperLoRA\" [64] which is a recent PEFT method fine-tunes large language models in hyperbolic space. For a clear comparison, we ignore the segmentation decoder when calculating the number of learnable parameters. The base model is ViT-B/16. " + }, + { + "bbox": [ + 54, + 525, + 555, + 570 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 54, + 525, + 555, + 570 + ], + "type": "text", + "content": ": Fine-tuning. " + }, + { + "bbox": [ + 54, + 525, + 555, + 570 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 54, + 525, + 555, + 570 + ], + "type": "text", + "content": ": Freezing." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 590, + 294, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 590, + 294, + 651 + ], + "spans": [ + { + "bbox": [ + 54, + 590, + 294, + 651 + ], + "type": "text", + "content": "that our method uses the fewest trainable parameters while equipping CLIP with segmentation ability. Since we do not introduce any extra architectures, and the logarithmic and exponential maps can be completed within linear complexity [64], the inference overhead is negligible." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 654, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 654, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 654, + 295, + 715 + ], + "type": "text", + "content": "Application to other tasks. Here, we further carry out extra experiments in two other tasks. In an open-vocabulary object detection task, CLIP's image embeddings are transitioned from the image level to the object level. We follow a recent work [58] to investigate open-vocabulary object detection on" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 590, + 555, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 555, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 555, + 674 + ], + "type": "text", + "content": "the COCO dataset [38]. As shown in Table 2, the superior results demonstrate that HyperCLIP can effectively generalize to other tasks that require aligning text embeddings with embeddings from arbitrary hierarchical levels, e.g., object-level alignment. Furthermore, the open-vocabulary panoptic segmentation results presented in Table 3 further corroborate this conclusion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Qualitative results. Here, we visualize our method's representative example segmentation results against prevailing methods, i.e., CAT-Seg [9] and SAN [62], in the A-847" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4568" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": "dataset. As shown in Fig. 4, we can observe that our method can make more accurate predictions on object location and category. Even if objects are small and the background is complicated, our method still performs well." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 130, + 149, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 130, + 149, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 130, + 149, + 144 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 149, + 297, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 149, + 297, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 149, + 297, + 388 + ], + "type": "text", + "content": "Ablation of fine-tuning CLIP. We assess the efficacy of various fine-tuning design choices for HyperCLIP. As shown in Table 6, we freeze one of CLIP's encoders and compare the performance with several commonly-used PEFT methods, e.g., LoRA [21]. Compared to fully freezing CLIP's encoders (row 1), fine-tuning either encoder independently allows for minimizing its hyperbolic radius, thereby equipping the corresponding encoder with segmentation ability. However, one modality's hierarchical level remains at the image level, which hinders CLIP from fully acquiring segmentation ability. Consequently, jointly fine-tuning both encoders aligns the hierarchical levels at the pixel level, resulting in optimal performance. We also compare our method with existing PEFT strategies. Most of these methods, which operate in Euclidean space, struggle to capture the transition between different hierarchical levels. Notably, HyperLoRA [64] also updates parameters in hyperbolic space. However, because HyperLoRA does not specifically adjust the hyperbolic radius, it fails to fully endow CLIP with segmentation ability, resulting in inferior performance compared to our method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 389, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 533 + ], + "type": "text", + "content": "Ablation of the size of diagonal blocks. The size of the diagonal block in our scaling transformations determines the number of parameters required to equip CLIP with segmentation ability. To investigate its impact, we conduct an ablation study on the scaling transformations by systematically varying the size of the diagonal block. As shown in Table 4, accuracy improves incrementally on datasets with fewer classes. However, on datasets with more classes (e.g., A-847), the accuracy performance reaches a saturation point around a block size of 128. This suggests that increasing the dimensionality of the diagonal block introduces more trainable parameters, which may lead to over-fitting on a closed set." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 534, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 297, + 713 + ], + "type": "text", + "content": "Effect of the hyperbolic radius of text embeddings. In order to closely examine the effect of the hyperbolic radius on equipping CLIP with segmentation ability, we apply a mean squared error (MSE) loss during fine-tuning. This MSE loss encourages the average hyperbolic radius of the text embeddings to converge towards different target values. As shown in Fig. 5, the initial value of the average hyperbolic radius is approximately 7.5, which is subsequently minimized to around 6.0 to achieve the optimal value for segmentation. Deviating from this value, either by increasing or decreasing the expected radius, results in diminished performance. This observation offers a new insight: the hyperbolic radius required for a vision task might be relatively fixed, suggesting that it could serve as a metric to quantify the granularity of different vision tasks." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 326, + 70, + 542, + 123 + ], + "blocks": [ + { + "bbox": [ + 326, + 70, + 542, + 123 + ], + "lines": [ + { + "bbox": [ + 326, + 70, + 542, + 123 + ], + "spans": [ + { + "bbox": [ + 326, + 70, + 542, + 123 + ], + "type": "table", + "html": "
Möbius.A-847PC-459A-150PC-59PAS-20b
X11.118.230.257.377.5
✓12.319.232.158.578.9
", + "image_path": "4a39143cac545910b33e6792cd51adfbe3402e02de4d989cc59e31d94370acc9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 132, + 555, + 155 + ], + "lines": [ + { + "bbox": [ + 313, + 132, + 555, + 155 + ], + "spans": [ + { + "bbox": [ + 313, + 132, + 555, + 155 + ], + "type": "text", + "content": "Table 7. Ablation on the necessity of Möbius matrix multiplication. \"Möbius\": Möbius matrix multiplication." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 317, + 163, + 552, + 334 + ], + "blocks": [ + { + "bbox": [ + 317, + 163, + 552, + 334 + ], + "lines": [ + { + "bbox": [ + 317, + 163, + 552, + 334 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 552, + 334 + ], + "type": "image", + "image_path": "027af0d30e3bce7c9134c69adcd015dd10e90dbbd331a1282bc33de86be3bacc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 337, + 555, + 392 + ], + "lines": [ + { + "bbox": [ + 313, + 337, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 555, + 392 + ], + "type": "text", + "content": "Figure 5. Visualization on the effect of different average hyperbolic radii of text embeddings. The controlling of different hyperbolic radii is achieved via a mean squared error (MSE) loss that enforces a constraint between the expected radius and the mean radius of the text embeddings during fine-tuning." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 406, + 556, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 556, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 556, + 491 + ], + "type": "text", + "content": "Necessity of Möbius matrix multiplication. We ablate this experiment to validate the influence of the Möbius matrix multiplication operation. In Table 7, “" + }, + { + "bbox": [ + 313, + 406, + 556, + 491 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 313, + 406, + 556, + 491 + ], + "type": "text", + "content": "” means that we use standard matrix multiplication. It achieves worse performance compared with “√”, which indicates that the Möbius matrix multiplication is essential for minimizing the hyperbolic radius of CLIP's embeddings." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 501, + 388, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 388, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 388, + 514 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 521, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 712 + ], + "type": "text", + "content": "In this paper, we propose HyperCLIP, a novel fine-tuning strategy for CLIP tailored to open-vocabulary semantic segmentation. Recognizing the hierarchical structure of images, we hypothesize that aligning the hierarchical level of CLIP's text encoder with pixel-level representations is crucial for accurate pixel-level predictions. HyperCLIP achieves this by adjusting the hierarchical level in hyperbolic space through scaling transformations applied to CLIP's embeddings, enabling precise cross-modal alignment. Extensive experiments demonstrate that HyperCLIP sets new state-of-the-art results in open-vocabulary semantic segmentation. More importantly, we observe that after adjustment, CLIP's text embeddings maintain a relatively fixed hyperbolic radius across datasets. This suggests that the segmentation task's granularity may be quantified using the hyperbolic radius, offering a promising direction for future research." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4569" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 136 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 136 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 136 + ], + "type": "text", + "content": "This work was supported by the NSFC under Grant 62322604, 62176159 and 62401361, and in part by the Shanghai Municipal Science and Technology Major Project under Grant 2021SHZDX0102." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "spans": [ + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 176, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 176, + 297, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 176, + 297, + 209 + ], + "spans": [ + { + "bbox": [ + 61, + 176, + 297, + 209 + ], + "type": "text", + "content": "[1] Irving Biederman. Recognition-by-components: a theory of human image understanding. Psychological review, 94(2): 115, 1987. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 210, + 296, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 210, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 61, + 210, + 296, + 243 + ], + "type": "text", + "content": "[2] Martin R Bridson and André Haefliger. Metric spaces of non-positive curvature. Springer Science & Business Media, 2013. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 244, + 295, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 244, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 61, + 244, + 295, + 277 + ], + "type": "text", + "content": "[3] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. Advances in Neural Information Processing Systems, 32, 2019. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 278, + 296, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 278, + 296, + 322 + ], + "spans": [ + { + "bbox": [ + 61, + 278, + 296, + 322 + ], + "type": "text", + "content": "[4] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 2, 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 323, + 295, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 323, + 295, + 355 + ], + "spans": [ + { + "bbox": [ + 62, + 323, + 295, + 355 + ], + "type": "text", + "content": "[5] James W Cannon, William J Floyd, Richard Kenyon, Walter R Parry, et al. Hyperbolic geometry. Flavors of geometry, 31 (59-115):2, 1997. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 357, + 295, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 295, + 412 + ], + "type": "text", + "content": "[6] Jiaxin Chen, Jie Qin, Yuming Shen, Li Liu, Fan Zhu, and Ling Shao. Learning attentive and hierarchical representations for 3d shape recognition. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 105-122. Springer, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 413, + 296, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 413, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 62, + 413, + 296, + 466 + ], + "type": "text", + "content": "[7] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 468, + 296, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 296, + 523 + ], + "type": "text", + "content": "[8] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 525, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 296, + 567 + ], + "type": "text", + "content": "[9] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Cat-seg: Cost aggregation for open-vocabulary semantic segmentation, 2024. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 569, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 296, + 613 + ], + "type": "text", + "content": "[10] Karan Desai, Maximilian Nickel, Tanmay Rajpurohit, Justin Johnson, and Shanmukha Ramakrishna Vedantam. Hyperbolic image-text representations. In International Conference on Machine Learning, pages 7694-7731. PMLR, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 614, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 296, + 658 + ], + "type": "text", + "content": "[11] Jian Ding, Nan Xue, Gui-Song Xia, and Dengxin Dai. Decoupling zero-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11583-11592, 2022. 3, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "text", + "content": "[12] Zi-Yi Dou, Aishwarya Kamath, Zhe Gan, Pengchuan Zhang, Jianfeng Wang, Linjie Li, Zicheng Liu, Ce Liu, Yann Le-Cun, Nanyun Peng, et al. Coarse-to-fine vision-language pre-training with fusion in the backbone. Advances in neural information processing systems, 35:32942-32956, 2022. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "type": "text", + "content": "[13] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuohang Wang, Lijuan Wang, Chenguang Zhu, Pengchuan Zhang, Lu Yuan, Nanyun Peng, et al. An empirical study of training end-to-end vision-and-language transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18166-18176, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 140, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 555, + 183 + ], + "type": "text", + "content": "[14] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 184, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 555, + 228 + ], + "type": "text", + "content": "[15] Zhe Gan, Yen-Chun Chen, Linjie Li, Chen Zhu, Yu Cheng, and Jingjing Liu. Large-scale adversarial training for vision-and-language representation learning. Advances in Neural Information Processing Systems, 33:6616-6628, 2020. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 228, + 555, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 555, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 555, + 260 + ], + "type": "text", + "content": "[16] Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. Advances in neural information processing systems, 31, 2018. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 261, + 555, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 555, + 304 + ], + "type": "text", + "content": "[17] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In European Conference on Computer Vision, pages 540-557. Springer, 2022. 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 305, + 555, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 555, + 350 + ], + "type": "text", + "content": "[18] Yunhui Guo, Xudong Wang, Yubei Chen, and Stella X Yu. Clipped hyperbolic classifiers are super-hyperbolic classifiers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11-20, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 350, + 554, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 554, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 554, + 404 + ], + "type": "text", + "content": "[19] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790-2799. PMLR, 2019. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 405, + 555, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 405, + 555, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 405, + 555, + 470 + ], + "type": "text", + "content": "[20] Chengyang Hu, Ke-Yue Zhang, Taiping Yao, Shouhong Ding, and Lizhuang Ma. Rethinking generalizable face anti-spoofing via hierarchical prototype-guided distribution refinement in hyperbolic space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1032-1041, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 471, + 555, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 555, + 525 + ], + "type": "text", + "content": "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 526, + 554, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 554, + 581 + ], + "type": "text", + "content": "[22] Ming Hu, Kun Yuan, Yaling Shen, Feilong Tang, Xiaohao Xu, Lin Zhou, Wei Li, Ying Chen, Zhongxing Xu, Zelin Peng, et al. Ophclip: Hierarchical retrieval-augmented learning for ophthalmic surgical video-language pretraining. arXiv preprint arXiv:2411.15421, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 582, + 555, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 555, + 635 + ], + "type": "text", + "content": "[23] Zhicheng Huang, Zhaoyang Zeng, Yupan Huang, Bei Liu, Dongmei Fu, and Jianlong Fu. Seeing out of the box: End-to-end pre-training for vision-language representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12976-12985, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 636, + 555, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 555, + 680 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 555, + 680 + ], + "type": "text", + "content": "[24] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "type": "text", + "content": "[25] Siyu Jiao, Hongguang Zhu, Jiannan Huang, Yao Zhao, Yunchao Wei, and Humphrey Shi. Collaborative vision-text representation optimizing for open-vocabulary segmentation. In" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4570" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "European Conference on Computer Vision, pages 399-416. Springer, 2024. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "type": "text", + "content": "[26] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 206 + ], + "type": "text", + "content": "[27] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6418-6428, 2020. 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 296, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 296, + 252 + ], + "type": "text", + "content": "[28] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International conference on machine learning, pages 5583-5594. PMLR, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 296, + 285 + ], + "type": "text", + "content": "[29] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 288, + 296, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 288, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 288, + 296, + 342 + ], + "type": "text", + "content": "[30] Fanjie Kong, Yanbei Chen, Jiarui Cai, and Davide Modolo. Hyperbolic learning with synthetic captions for openworld detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16762-16771, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 344, + 296, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 296, + 387 + ], + "type": "text", + "content": "[31] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In International Conference on Learning Representations, 2022. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 389, + 296, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 389, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 389, + 296, + 443 + ], + "type": "text", + "content": "[32] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 296, + 498 + ], + "type": "text", + "content": "[33] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 296, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 296, + 566 + ], + "type": "text", + "content": "[34] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 296, + 634 + ], + "type": "text", + "content": "[35] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. Oscar: Object-semantics aligned pre-training for vision-language tasks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX 16, pages 121-137. Springer, 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 296, + 678 + ], + "type": "text", + "content": "[36] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. Advances in Neural Information Processing Systems, 35:109-123, 2022. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "type": "text", + "content": "[37] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 105 + ], + "type": "text", + "content": "mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 555, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 173 + ], + "type": "text", + "content": "[38] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 555, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 555, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 555, + 218 + ], + "type": "text", + "content": "[39] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 220, + 555, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 555, + 285 + ], + "type": "text", + "content": "[40] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 891–898, 2014. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 287, + 555, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 555, + 330 + ], + "type": "text", + "content": "[41] Maximillian Nickel and Douwe Kiela. Learning continuous hierarchies in the lorentz model of hyperbolic geometry. In International conference on machine learning, pages 3779-3788. PMLR, 2018. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 332, + 555, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 555, + 376 + ], + "type": "text", + "content": "[42] Zelin Peng, Zhengqin Xu, Zhilin Zeng, Yaoming Wang, Lingxi Xie, Qi Tian, and Wei Shen. Parameter-efficient fin-tuning in hyperspherical space for open-vocabulary semantic segmentation. arXiv preprint arXiv:2405.18840, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 377, + 555, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 555, + 432 + ], + "type": "text", + "content": "[43] Zelin Peng, Zhengqin Xu, Zhilin Zeng, Lingxi Xie, Qi Tian, and Wei Shen. Parameter efficient fine-tuning via cross block orchestration for segment anything model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3743-3752, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 434, + 555, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 555, + 487 + ], + "type": "text", + "content": "[44] Zelin Peng, Zhengqin Xu, Zhilin Zeng, Xiaokang Yang, and Wei Shen. Sam-parser: Fine-tuning sam efficiently by parameter space reconstruction. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4515-4523, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 489, + 555, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 555, + 544 + ], + "type": "text", + "content": "[45] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Scholkopf. Controlling text-to-image diffusion by orthogonal finetuning. Advances in Neural Information Processing Systems, 36:79320-79362, 2023. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 545, + 555, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 555, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 555, + 610 + ], + "type": "text", + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 613, + 555, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 555, + 667 + ], + "type": "text", + "content": "[47] Sameera Ramasinghe, Violetta Shevchenko, Gil Avraham, and Ajanthan Thalaiyasingam. Accept the modality gap: An exploration in the hyperbolic space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27263-27272, 2024. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "text", + "content": "[48] Frederic Sala, Chris De Sa, Albert Gu, and Christopher Ré. Representation tradeoffs for hyperbolic embeddings. In International conference on machine learning, pages 4460-4469. PMLR, 2018. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "4571" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "text", + "content": "[49] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Lst: Ladder side-tuning for parameter and memory efficient transfer learning. Advances in Neural Information Processing Systems, 35: 12991-13005, 2022. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "type": "text", + "content": "[50] Dídac Surís, Ruoshi Liu, and Carl Vondrick. Learning the predictability of the future. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12607-12617, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 196 + ], + "type": "text", + "content": "[51] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 296, + 251 + ], + "type": "text", + "content": "[52] Feilong Tang, Zile Huang, Chengzhi Liu, Qiang Sun, Harry Yang, and Ser-Nam Lim. Intervening anchor token: Decoding strategy in alleviating hallucinations for mllms. In The Thirteenth International Conference on Learning Representations. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 296, + 297 + ], + "type": "text", + "content": "[53] Abraham A Ungar. Hyperbolic trigonometry and its application in the poincaré ball model of hyperbolic geometry. Computers & Mathematics with Applications, 41(1-2):135-147, 2001. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 298, + 296, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 296, + 331 + ], + "type": "text", + "content": "[54] Abraham Albert Ungar. A gyrovector space approach to hyperbolic geometry. Synthesis Lectures on Mathematics and Statistics, 1(1):1-194, 2008. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 296, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 296, + 376 + ], + "type": "text", + "content": "[55] Jianfeng Wang, Xiaowei Hu, Zhe Gan, Zhengyuan Yang, Xiyang Dai, Zicheng Liu, Yumao Lu, and Lijuan Wang. Ufo: A unified transformer for vision-language representation learning. arXiv preprint arXiv:2111.10023, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 377, + 296, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 296, + 432 + ], + "type": "text", + "content": "[56] Sijie Wang, Qiyu Kang, Rui She, Wei Wang, Kai Zhao, Yang Song, and Wee Peng Tay. Hypliloc: Towards effective lidar pose regression with hyperbolic fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5176-5185, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 296, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 296, + 477 + ], + "type": "text", + "content": "[57] Melanie Weber, Manzil Zaheer, Ankit Singh Rawat, Aditya K Menon, and Sanjiv Kumar. Robust large-margin learning in hyperbolic space. Advances in Neural Information Processing Systems, 33:17863-17873, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 478, + 296, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 296, + 533 + ], + "type": "text", + "content": "[58] Size Wu, Wenwei Zhang, Lumin Xu, Sheng Jin, Wentao Liu, and Chen Change Loy. Clim: Contrastive language-image mosaic for region representation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6117-6125, 2024. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 534, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 296, + 567 + ], + "type": "text", + "content": "[59] Bin Xie, Jiale Cao, Jin Xie, Fahad Shahbaz Khan, and Yanwei Pang. Sed: A simple encoder-decoder for open-vocabulary semantic segmentation, 2023. 2, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 568, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 296, + 624 + ], + "type": "text", + "content": "[60] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2955-2966, 2023. 3, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 296, + 679 + ], + "type": "text", + "content": "[61] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 3, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 680, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 296, + 714 + ], + "type": "text", + "content": "[62] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xiang Bai. Side adapter network for open-vocabulary semantic segmentation. In Proceedings of the IEEE/CVF Conference" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 487 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 333, + 72, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 555, + 95 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition, pages 2945-2954, 2023. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 555, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 161 + ], + "type": "text", + "content": "[63] Haochen Xue, Feilong Tang, Ming Hu, Yexin Liu, Qidong Huang, Yulong Li, Chengzhi Liu, Zhongxing Xu, Chong Zhang, Chun-Mei Feng, et al. Mmrc: A large-scale benchmark for understanding multimodal large language model in real-world conversation. arXiv preprint arXiv:2502.11903, 2025. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 555, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 555, + 205 + ], + "type": "text", + "content": "[64] Menglin Yang, Aosong Feng, Bo Xiong, Jihong Liu, Irwin King, and Rex Ying. Hyperbolic fine-tuning for large language models. arXiv preprint arXiv:2410.04010, 2024. 4, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 207, + 555, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 555, + 251 + ], + "type": "text", + "content": "[65] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. In NeurIPS, 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 252, + 555, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 555, + 319 + ], + "type": "text", + "content": "[66] Zhen Yu, Toan Nguyen, Yaniv Gal, Lie Ju, Shekhar S Chandra, Lei Zhang, Paul Bonnington, Victoria Mar, Zhiyong Wang, and Zongyuan Ge. Skin lesion recognition with class-hierarchy regularized hyperbolic embeddings. In International conference on medical image computing and computer-assisted intervention, pages 594-603. Springer, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 319, + 555, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 555, + 363 + ], + "type": "text", + "content": "[67] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127:302-321, 2019. 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 364, + 555, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 364, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 364, + 555, + 397 + ], + "type": "text", + "content": "[68] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712. Springer, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 398, + 555, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 398, + 555, + 452 + ], + "spans": [ + { + "bbox": [ + 316, + 398, + 555, + 452 + ], + "type": "text", + "content": "[69] Ziqin Zhou, Yinjie Lei, Bowen Zhang, Lingqiao Liu, and Yifan Liu. Zegclip: Towards adapting clip for zero-shot semantic segmentation. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 453, + 555, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 555, + 487 + ], + "type": "text", + "content": "[70] Yudong Zhu, Di Zhou, Jinghui Xiao, Xin Jiang, Xiao Chen, and Qun Liu. Hypertext: Endowing fasttext with hyperbolic geometry. arXiv preprint arXiv:2010.16143, 2020. 3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4572" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Understanding Multi-Task Activities from Single-Task Videos/b633800a-1331-4862-80d6-5ccbeba54c12_content_list.json b/2025/Understanding Multi-Task Activities from Single-Task Videos/b633800a-1331-4862-80d6-5ccbeba54c12_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..59053590c54c2c24a86122833cc6c24c5064ca43 --- /dev/null +++ b/2025/Understanding Multi-Task Activities from Single-Task Videos/b633800a-1331-4862-80d6-5ccbeba54c12_content_list.json @@ -0,0 +1,1711 @@ +[ + { + "type": "text", + "text": "Understanding Multi-Task Activities from Single-Task Videos", + "text_level": 1, + "bbox": [ + 187, + 130, + 810, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuhan Shen Northeastern University shen.yuh@northeastern.edu", + "bbox": [ + 225, + 181, + 433, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ehsan Elhamifar Northeastern University e.elhamifar@northeastern.edu", + "bbox": [ + 506, + 181, + 756, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce and develop a framework for Multi-Task Temporal Action Segmentation (MT-TAS), a novel paradigm that addresses the challenges of interleaved actions when performing multiple tasks simultaneously. Traditional action segmentation models, trained on single-task videos, struggle to handle task switches and complex scenes inherent in multi-task scenarios. To overcome these challenges, our MT-TAS approach synthesizes multi-task video data from single-task sources using our Multi-task Sequence Blending and Segment Boundary Learning modules. Additionally, we propose to dynamically isolate foreground and background elements within video frames, addressing the intricacies of object layouts in multi-task scenarios and enabling a new two-stage temporal action segmentation framework with Foreground-Aware Action Refinement. Also, we introduce the Multi-task Egocentric Kitchen Activities (MEKA) dataset, containing 12 hours of egocentric multi-task videos, to rigorously benchmark MT-TAS models. Extensive experiments demonstrate that our framework effectively bridges the gap between single-task training and multi-task testing, advancing temporal action segmentation with state-of-the-art performance in complex environments.", + "bbox": [ + 89, + 300, + 483, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 667, + 220, + 683 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Consider a typical morning routine, when you prepare breakfast: you start making oatmeal and while waiting for it to cook, you brew coffee and pack your food for lunch. In fact, in our daily lives, we often perform multiple tasks simultaneously. For computer vision systems to accurately analyze such behaviors and provide useful assistance, they must comprehend these complex and interleaved activities. Temporal action segmentation (TAS) has emerged as a crucial technology for understanding human activities by partitioning videos into meaningful segments, each corresponding to a distinct action or step [17, 22, 24, 31, 33, 48, 49, 64, 73, 84, 88]. Recently, TAS has attracted growing interests", + "bbox": [ + 89, + 694, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/876624b57e93088296560aa0e7567f0fb2210e50cd7c430c726ab5ffeb622451.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 268, + 895, + 368 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/554a9cc7c6b4d3d3b32430a73a9db0d07b0df68fbfd7fc36a6182a46ed326df7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 371, + 633, + 460 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9356039069186c8d6417b05ee8d78549904a653fe932c676beb3732d3ca8899c.jpg", + "image_caption": [ + "single-task videos", + "Figure 1. (a) Multi-task temporal action segmentation involves segmenting actions from interleaved tasks. It requires capturing interruptions and resumptions in the execution of each task. (b) Comparison between frames from single-task and multi-task videos. In multi-task videos, at each time instant, the scene often contains objects from tasks different from the currently executed one (highlighted by purple boxes)." + ], + "image_footnote": [], + "bbox": [ + 638, + 371, + 741, + 463 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f43393b8d019cb9d2ebf1840e71f5897a05230b65306117e98d26f44ec2fd7bd.jpg", + "image_caption": [ + "multi-task videos", + "(b) Comparison between frames from single-task videos v.s. multi-task videos" + ], + "image_footnote": [], + "bbox": [ + 779, + 371, + 883, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "within egocentric videos [5, 36, 52, 54, 65, 77], where activities are captured from a first-person perspective, enabling fine-grained analysis of human actions, which is essential for applications such as real-time personalized assistance, embodied agents and efficient video content retrieval.", + "bbox": [ + 511, + 580, + 906, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we study the new problem of Multi-Task Temporal Action Segmentation (MT-TAS), which explicitly focuses on scenarios where users perform multiple tasks by interleaving their steps within a single video, see Fig. 1. There are two major challenges that existing TAS models face in addressing MT-TAS. First, despite the notable progress through deep learning techniques [1, 6, 14, 17, 41, 42, 45, 49, 84], current TAS models have primarily focused on scenarios where each video contains actions related to a single high-level task, such as assembling a piece of furniture or cooking a specific recipe. This single-task focus limits the applicability of TAS models to the multi-task setting since i) they are biased to predict actions/steps from only one task at the inference time, ii) when performing several tasks, the environment and the scene often contains objects relevant to multiple tasks, which adds complexity", + "bbox": [ + 511, + 657, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Code: https://github.com/Yuhan-Shen/MT-TAS.", + "bbox": [ + 107, + 886, + 441, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19120", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to the context, while single-task videos often contain objects relevant to one task (see Fig. 1(b)), iii) interleaving steps of multiple tasks introduces complexities such as unpredictable task switches and disruptions to the temporal flows of actions within each task.", + "bbox": [ + 89, + 90, + 480, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Second, a significant obstacle in advancing MT-TAS is the lack of appropriate training data. Existing datasets [5, 16, 18, 31, 36, 63, 88] are predominantly collected for single-task scenarios, making them insufficient for capturing complexities of interleaved tasks. Furthermore, many possible step combinations in the multi-task setting makes it impractical to collect and annotate comprehensive real-world videos covering all potential situations (such videos would also be much longer than single-task videos).", + "bbox": [ + 89, + 167, + 480, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Paper Contributions. We study the new problem of MT-TAS. To overcome the aforementioned challenges, we propose an MT-TAS framework that uses only single-task videos for training. We generate synthetic multi-task data using single-task videos and propose a two-stage action segmentation model for accurate predictions. For the data generation, we first develop a Multi-task Sequence Blending (MSB) module that leverages Large Language Models (LLMs) to create multi-task transcripts of interleaved steps and their associated videos. Second, we develop a Segment Boundary Learning (SBL) module to mitigate the abrupt transitions at segment boundaries caused by the synthetic data. Third, we introduce a Dynamic Isolation of Video Elements (DIVE) approach to isolate foreground and background components within video frames by identifying action-relevant objects and combining foregrounds of actions in one task with backgrounds of other tasks to capture the more complex scenes of multi-task videos. By increasing the variation in the background components for the training data, our method becomes more robust to unseen object compositions in multi-task scenarios.", + "bbox": [ + 91, + 309, + 483, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For the action segmentation model, we develop a two-stage framework, where we refine the initial action predictions using a Foreground-Aware Action Refinement (FAAR) module. This approach leverages foreground information to enhance the focus of our model on action-relevant cues (e.g., focusing on the regions of kettle and measuring cup for the action \"transfer water to kettle\" as in Fig. 3(b)), improving accuracy and robustness to background variations during the multi-task testing.", + "bbox": [ + 89, + 628, + 480, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For proper evaluation of MT-TAS models, we gather and annotate the Multi-task Egocentric Kitchen Activities (MEKA) video dataset (we plan to publicly release the dataset), built upon the recent EgoPER [36] dataset. MEKA comprises 100 egocentric videos of users (about 12 hours of footage) performing multiple kitchen recipes from EgoPER in each run by interleaving their steps. This new benchmark captures a variety of interleaved kitchen activities, providing a rigorous testing ground for MT-TAS models. Through", + "bbox": [ + 89, + 763, + 480, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "extensive experiments in both offline and online settings, we demonstrate that our proposed framework successfully bridges the gap between single-task training data and the complexities of real-world multi-task scenarios, advancing research in temporal action segmentation.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 511, + 176, + 663, + 193 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Temporal Action Segmentation. TAS aims to understand the sequential structure of human activities in videos by assigning an action label to each frame, thereby partitioning the video into a sequence of action segments. Based on the level of supervision during training, TAS methods can be categorized into three main groups: unsupervised [2, 15, 16, 20, 33, 34, 61, 66, 76, 79, 80, 88], weakly-supervised [8, 12, 19, 22-24, 38-40, 44, 47, 48, 57-59, 62, 64, 71], and fully-supervised [1, 4, 6, 17, 27, 29, 32, 35, 36, 41, 42, 45, 49, 51, 60, 65, 68-71, 83, 84] approaches. Our work primarily aligns with the fully-supervised methods. However, regardless of the supervision level, previous studies have predominantly focused on single-task TAS, where test videos contain actions from only one high-level activity or task, while we focus on multi-task TAS. A closely related work, UnweaveNet [53], aims to predict the start, continuation, and resumption of multiple activity threads within a video. While UnweaveNet models multiple concurrent activities, its predictions are at the activity or task level and do not capture the fine-grained action steps within each task. In contrast, our approach focuses on recognizing and segmenting detailed action steps across interleaved tasks, providing a more granular understanding of multi-task activities.", + "bbox": [ + 511, + 203, + 903, + 564 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Procedural Video Datasets. To support research in TAS and related tasks, numerous procedural video datasets have been collected and released. These datasets include third-person videos (e.g., Breakfast [30], 50Salads [72], YouTube Instructional [2], COIN [74], CrossTask [88], ProceL [16], and ATA [24]), egocentric videos (e.g., GTEA [18], EGTEA [43], MECCANO [56], EgoProceL [5], HoloAssist [78], and EgoPER [36]), or both views such as Assembly101 [63], $(3 + 1)\\mathrm{ReC}$ [62] and EgoExoLearn [28]. While large-scale video datasets like EPIC-Kitchens [11], Ego4D [25], and Ego-Exo4D [26] exist, they primarily focus on detecting and recognizing atomic action steps rather than modeling the procedural relationships between steps within a task, making them less suitable for TAS research. Also, most existing TAS datasets are limited to single-task scenarios, where each video contains actions related to one high-level activity. In contrast, our collected MEKA dataset specifically addresses multi-task temporal action segmentation by capturing videos where multiple tasks are performed in an interleaved manner, providing a more challenging and realistic benchmark for advancing TAS research.", + "bbox": [ + 511, + 566, + 903, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Object-Aware and Foreground/Background Learning.", + "bbox": [ + 511, + 885, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19121", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning to localize active objects or relevant objects is a prominent topic in video understanding [67, 75, 85, 86]. Furthermore, researchers have explored leveraging object information to improve video representation learning [85], action recognition [87], and action anticipation [86]. Some studies have also focused on disentangling foreground and background elements to enhance video representations [3, 13, 37]. Our work differs by specifically addressing the generalization of TAS models to multi-task videos. Unlike existing approaches that concentrate on short video clips or single-task videos, we propose a method that dynamically isolates foreground and background elements to handle unseen object compositions in multi-task scenarios, thereby mitigating interference from irrelevant objects.", + "bbox": [ + 89, + 90, + 483, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Multi-Task Temporal Action Segmentation", + "text_level": 1, + "bbox": [ + 89, + 321, + 473, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Setting", + "text_level": 1, + "bbox": [ + 89, + 347, + 253, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a video sequence $\\mathcal{V} = (v_{1},\\ldots ,v_{T})$ of $T$ frames, the goal of TAS is to predict the sequence of action labels $A = (a_{1},\\dots,a_{T})$ , where each label $a_{t}$ comes from a predefined set of action classes $\\mathcal{C}$ . Assume we have $O$ tasks, where the set $\\mathcal{C}_o$ consists of actions for the task $o\\in \\{1,\\dots,O\\}$ and $\\mathcal{C} = \\bigcup_{o = 1}^{O}\\mathcal{C}_{o}$ contains actions of all tasks. In a single-task video, actions come from one $\\mathcal{C}_o$ , while a multi-task video has interleaved actions from several tasks, hence, $a_{t}$ may belong to any $\\mathcal{C}_o$ for $o\\in \\{1,\\dots,O\\}$ . Additionally, there may be background segments (e.g., 'answering phone') irrelevant to any of the tasks. A particular challenge in MT-TAS is to distinguish between action segments belonging to different tasks and background segments.", + "bbox": [ + 89, + 371, + 483, + 569 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our setting, the training data consists of single-task videos with full framewise annotations. Our goal is to train a model on single-task videos so that it can effectively generalize to multi-task videos during testing.", + "bbox": [ + 89, + 571, + 483, + 632 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview of Proposed Framework", + "text_level": 1, + "bbox": [ + 89, + 635, + 390, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2 shows the overview of our MS-TAS framework. Our approach first employs Multi-task Sequence Blending (MSB) to generate synthetic multi-task videos by leveraging LLM-guided task transitions, eliminating the need for multi-task data collection. However, the initial synthesized multi-task videos contain abrupt transitions at segment boundaries, which disrupt the temporal flow, and contain discrepancies in object compositions between single-task and multi-task videos. To mitigate these issues, we introduce three components: a Segment Boundary Learning (SBL) module that enhances temporal smoothness between segments, a Dynamic Isolation of Video Elements approach that augments the background variations during training, and a two-stage TAS method with foreground-aware action refinement to improve the predictions of the model. Next, we discuss each component of our framework in details.", + "bbox": [ + 89, + 659, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/06105a7da63057dc2cd121d50dfc8ccf433ef315bfdd074cf2dc27bb4a16576b.jpg", + "image_caption": [ + "Figure 2. Framework Overview. The modules marked with $\\Theta$ are used exclusively for training, and modules marked with $\\diamond$ are used for both training and testing." + ], + "image_footnote": [], + "bbox": [ + 517, + 92, + 898, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Multi-task Sequence Blending (MSB)", + "text_level": 1, + "bbox": [ + 511, + 532, + 834, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of MSB is to create initial multi-task video sequences from single-task videos, to simulate real-world multitasking behaviors. Given single-task training videos $\\{\\mathcal{V}^{(i)}\\}_{i=1}^{M}$ , where each video $\\mathcal{V}^{(i)} = (v_{1}^{(i)}, v_{2}^{(i)}, \\ldots, v_{T_{i}}^{(i)})$ has a task label $\\tau_{i}$ and consists of a sequence of action segments $S^{(i)} = (s_{1}^{(i)}, s_{2}^{(i)}, \\ldots, s_{N_{i}}^{(i)})$ with action labels $A^{(i)} = (a_{1}^{(i)}, a_{2}^{(i)}, \\ldots, a_{N_{i}}^{(i)})$ , a naive approach would be to randomly sample segments from videos to concatenate. However, this ignores the temporal order of actions within a task and more importantly can create unrealistic sequences, e.g., \"scooping jelly from jar\" followed by \"grinding coffee beans\" (this is unrealistic since it is common sense to spread the jelly right away after scooping it from a jar).", + "bbox": [ + 511, + 556, + 906, + 761 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To create plausible multi-task sequences, we: i) preserve the original temporal order of action segments within each task, ii) preserve common sense action ordering and switches among tasks. Specifically, the sequence of action segments $s_j^{(i)}$ from task $\\tau_i$ follows the same order as in the original single-task video. At the end of an action segment, we leverage an LLM, which encodes extensive contextual and commonsense knowledge from real-world data, to decide whether to continue the current task or switch to an-", + "bbox": [ + 511, + 763, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19122", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ec6246c1a902940358b037def5032591c178a36212053ff045e9457870704a0a.jpg", + "image_caption": [ + "(a) Multi-task Sequence Blending" + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 400, + 203 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/05d873dcd92b3f451998b6be9d9e716d31161b854f1ac6792057d68a435fe8b5.jpg", + "image_caption": [ + "(b) Dynamic Isolation of Video Elements", + "Figure 3. Illustration of leveraging LLM in our proposed framework for MSB (a) and DIVE (b)." + ], + "image_footnote": [], + "bbox": [ + 408, + 88, + 903, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "other task (see Fig. 3(a)). The LLM assesses the plausibility of continuing or switching based on the current action $a_{j}^{(i)}$ , the next action of the current task $a_{j+1}^{(i)}$ , and the next action of another task $a_{m}^{(k)}$ , where $\\tau_{k}$ is a different task and $a_{m}^{(k)}$ is its next pending action. For example, after \"turning on the kettle to boil water\" in a tea-making task, it is plausible to switch to a new task during the waiting time. Conversely, after \"scooping jelly from a jar\" in a pinwheel-making task, it makes sense to continue with the next action \"spread jelly on tortilla\" rather than switching tasks.", + "bbox": [ + 88, + 272, + 482, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During training, we randomly sample potential task switch points at the end of action segments and use the LLM to decide whether to continue or switch tasks. For more than two tasks, we first randomly select one unfinished task as the candidate for switching. We then blend the sequence of action segments based on the LLM's decisions, resulting in a new multi-task sequence of actions and frames.", + "bbox": [ + 88, + 435, + 483, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Segment Boundary Learning (SBL)", + "text_level": 1, + "bbox": [ + 89, + 549, + 398, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The multi-task videos synthesized by MSB exhibit abrupt transitions at boundaries where we switch tasks, which disrupt the temporal flow and negatively impact the model's performance. Therefore, we propose the SBL module that learns to recover the feature representation of a frame $f_{t}$ using its nearby frames, excluding frames immediately adjacent to $t$ to avoid trivial learning. This approach allows us to generate features for frames where we concatenate segments from different videos, thus enhancing the temporal smoothness between action segments.", + "bbox": [ + 88, + 570, + 482, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following prior works in TAS [17, 45, 49, 84], we first extract the framewise feature representation $f_{t}$ using an I3D feature extractor [7] applied to a sliding window $V_{t}$ ,", + "bbox": [ + 89, + 722, + 483, + 767 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {t} = \\mathrm {I 3 D} \\left(V _ {t}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 776, + 482, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, for a given frame $t$ , we consider its neighboring frames from the intervals $[t - \\delta - w, t - \\delta]$ and $[t + \\delta, t + \\delta + w]$ , where $w$ defines the window size and $\\delta$ specifies the exclusion margin around $t$ . We train a reconstruction function $\\mathcal{F}_{\\mathrm{SBL}}$ to predict the feature $f_{t}$ from its surrounding frames,", + "bbox": [ + 89, + 801, + 483, + 876 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {f} _ {t} = \\mathcal {F} _ {\\mathrm {S B L}} \\left(f _ {t - \\delta - w: t - \\delta}, f _ {t + \\delta : t + \\delta + w}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 883, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{t - \\delta - w:t - \\delta}$ and $f_{t + \\delta: t + \\delta + w}$ denote the features of frames before and after $t$ within the specified intervals. We optimize the reconstruction function by minimizing the reconstruction loss over non-boundary frames,", + "bbox": [ + 511, + 273, + 906, + 334 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {S B L}} = \\sum_ {t \\notin \\mathcal {B}} \\left\\| \\bar {f} _ {t} - f _ {t} \\right\\| ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 342, + 903, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{B}$ is the set of boundary frames (which come from two different videos, therefore are not continuous or realistic and we should not learn from them). By learning this reconstruction function, the model can generate smooth feature representations at segment boundaries. During training, we replace the original features at boundary frames with the reconstructed features $\\bar{f}_t$ to enhance the temporal smoothness of the synthesized multi-task videos.", + "bbox": [ + 511, + 383, + 906, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Dynamic Isolation of Video Elements (DIVE)", + "text_level": 1, + "bbox": [ + 511, + 512, + 893, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In real multi-task videos, the present objects and their layouts can significantly differ from those in single-task videos. For example, as shown in the top-right frame in Fig. 1(b), in a video of making both tea and pinwheels, we might add honey to a mug while a rolled tortilla is on the cutting board, which is typically unseen in single-task videos. However, the output of MSB, which concatenates different actions from different single-task videos, does not capture such realistic multi-task layouts. Observing that action-relevant objects remain consistent across single-task and multi-task videos, we propose a Dynamic Isolation of Video Elements approach to isolate video elements into foreground and background components, and develop a Foreground-Background Feature Composition module to synthesize more realistic multi-task videos for training.", + "bbox": [ + 511, + 535, + 906, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Detecting Action-Relevant Objects. Given a frame, it is challenging to distinguish between foreground and background solely based on visual cues, especially in cluttered multi-task environments. The definition of \"foreground\" or \"relevant objects\" depends on the action being performed; thus, knowledge of the action context is essential. Fig. 3(b) illustrates how we obtain the foreground and background components. We first query an LLM to obtain a list of relevant objects for each action class $a \\in \\mathcal{C}$ . For example,", + "bbox": [ + 511, + 763, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19123", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the action \"transfer water to kettle\" would be associated with \"kettle\" and \"measuring cup\". Then, we use an open-vocabulary object detector, GroundingDINO [46], to detect these relevant objects along with hands (we add \"left hand\" and \"right hand\" as object names) within each frame. We define the foreground region as the union of detected bounding boxes. We generate foreground and background frames through Gaussian blurring [82] outside and inside this region, respectively. We empirically observed that Gaussian blurring performs slightly better than simply zero masking.", + "bbox": [ + 89, + 90, + 483, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Foreground-Background Feature Composition (FBFC). After generating foreground and background frames, we extract I3D features,", + "bbox": [ + 89, + 244, + 483, + 291 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {t} ^ {\\mathrm {f g}} = \\mathrm {I 3 D} (V _ {t} \\odot M _ {t} ^ {\\mathrm {f g}}), f _ {t} ^ {\\mathrm {b g}} = \\mathrm {I 3 D} (V _ {t} \\odot M _ {t} ^ {\\mathrm {b g}}), \\qquad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 297, + 482, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f_{t}^{\\mathrm{fg}}$ and $f_{t}^{\\mathrm{bg}}$ are foreground and background frame features, respectively, and $\\odot$ is the Gaussian blurring operation with corresponding foreground masks $M_{t}^{\\mathrm{fg}}$ and background masks $M_{t}^{\\mathrm{fg}}$ . We train a decoder $\\mathcal{D}$ to reconstruct the original feature from foreground and background features,", + "bbox": [ + 89, + 323, + 483, + 402 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {f} _ {t} = \\mathcal {D} \\left(f _ {t} ^ {\\mathrm {f g}}, f _ {t} ^ {\\mathrm {b g}}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 410, + 482, + 430 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $f_{t}$ denote the original feature in Eq. (1). We train the decoder by minimizing the composition loss,", + "bbox": [ + 89, + 436, + 483, + 467 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {F B F C}} = \\sum_ {t} \\left\\| \\hat {f} _ {t} - f _ {t} \\right\\| ^ {2}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 476, + 482, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To augment synthetic multi-task features, we mix the background features of the current frame with background features of frames from another task", + "bbox": [ + 89, + 518, + 483, + 561 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {f} _ {t} ^ {\\mathrm {b g}} = \\beta f _ {t} ^ {\\mathrm {b g}} + (1 - \\beta) f _ {t ^ {\\prime}} ^ {\\mathrm {b g}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 571, + 482, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f_{t}^{\\mathrm{bg}}$ is the background feature from the current frame, $f_{t'}^{\\mathrm{bg}}$ is the background feature from another task, and $\\beta \\in [0,1]$ controls the mixing ratio. We then compose multi-task features from the foreground feature and the mixed background feature using the decoder", + "bbox": [ + 89, + 602, + 483, + 680 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {f} _ {t} = \\mathcal {D} \\left(f _ {t} ^ {\\mathrm {f g}}, \\hat {f} _ {t} ^ {\\mathrm {b g}}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 688, + 482, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We replace the training data with these recomposed features, exposing the model to diverse background variations.", + "bbox": [ + 89, + 718, + 483, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Two-Stage Temporal Action Segmentation", + "text_level": 1, + "bbox": [ + 89, + 757, + 450, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further enhance the robustness of our model to background variations during testing, we propose a two-stage action segmentation method incorporating a Foreground-Aware Action Refinement (FAAR) module. By focusing on foreground regions, FAAR mitigates the impact of background distractions and refines the initial TAS predictions.", + "bbox": [ + 89, + 780, + 483, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, with ground-truth actions available, we obtain the action-relevant objects and detect the foreground", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "regions. We extract CLIP features [55] from the foreground frames and apply a lightweight TAS model to predict foreground-aware action probabilities based solely on the foreground frames, $p_t^{\\mathrm{fg}}$ , which are combined with the initial action predictions $p_t$ output by the base TAS model,", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {t} ^ {\\mathrm {f g}} = \\mathcal {R} \\left(g \\left(v _ {t} \\odot m _ {t} ^ {\\mathrm {f g}}\\right)\\right), \\quad p _ {t} ^ {\\text {f i n a l}} = (1 - \\alpha) \\cdot p _ {t} + \\alpha \\cdot p _ {t} ^ {\\mathrm {f g}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 176, + 906, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $m_t^{\\mathrm{fg}}$ is the binary foreground mask, $g$ denotes CLIP feature extractor, $\\mathcal{R}$ denotes the lightweight TAS model, $\\alpha \\in [0,1]$ controls the weight of foreground-aware probabilities, and $p_t^{\\mathrm{final}}$ is the refined action probabilities.", + "bbox": [ + 511, + 203, + 903, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During inference, without ground-truth actions, we first select the top $K$ predicted action classes $\\{\\hat{a}_t^{(1)},\\dots ,\\hat{a}_t^{(K)}\\}$ based on the initial action probabilities $p_t$ from the base TAS model, and retrieve their relevant objects. We then detect the corresponding foreground regions and obtain foreground-aware action probabilities $p_t^{\\mathrm{fg},(k)}$ for each of the top $K$ actions respectively. We compute the final prediction by combining the initial action probabilities with the foreground-aware ones,", + "bbox": [ + 511, + 265, + 906, + 405 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {t} ^ {\\text {f i n a l}} = (1 - \\alpha) \\cdot p _ {t} + \\alpha \\cdot \\sum_ {k = 1} ^ {K} \\theta_ {t} ^ {(k)} p _ {t} ^ {\\mathrm {f g}, (k)}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 417, + 903, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the weights $\\theta_t^{(k)}$ are proportional to the initial probabilities of the top- $K$ actions, i.e., $\\theta_t^{(k)} = \\frac{p_t\\left(\\hat{a}_t^{(k)}\\right)}{\\sum_{j=1}^K p_t\\left(\\hat{a}_t^{(j)}\\right)}$ . This approach refines the predictions by focusing on the most probable actions and their relevant objects, enhancing robustness to background variations in multi-task scenarios.", + "bbox": [ + 511, + 459, + 906, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.7. Training and Inference", + "text_level": 1, + "bbox": [ + 513, + 561, + 725, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. We use the synthetic multi-task videos generated by MSB for training. We train the SBL module $\\mathcal{F}_{\\mathrm{SBL}}$ via $\\mathcal{L}_{\\mathrm{SBL}}$ in Eq. (3) and utilize the reconstructed features at segment boundaries to enhance temporal smoothness. Similarly, we train the FBFC decoder $\\mathcal{D}$ via $\\mathcal{L}_{\\mathrm{FBFC}}$ in Eq. (6), and augment the training data by mixing background features from different tasks as described in Eqs. (7) and (8). We train the base TAS model with FAAR using the standard action loss. The total loss function $\\mathcal{L}$ is:", + "bbox": [ + 511, + 593, + 905, + 728 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {a c t i o n}} + \\lambda_ {\\mathrm {S B L}} \\mathcal {L} _ {\\mathrm {S B L}} + \\lambda_ {\\mathrm {F B F C}} \\mathcal {L} _ {\\mathrm {F B F C}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 742, + 903, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{action}}$ includes the cross-entropy loss and smoothing loss for TAS, applied to the final action probabilities output by the FAAR module. $\\lambda_{\\mathrm{SBL}}$ and $\\lambda_{\\mathrm{FBFC}}$ are the loss weights.", + "bbox": [ + 511, + 771, + 905, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference. During inference, we do not use the MSB, SBL and FBFC modules, as they are designed to augment training. We first obtain initial action predictions from the base TAS model, and then apply the FAAR module, as specified in Eq. (10), to derive the final action predictions.", + "bbox": [ + 511, + 824, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19124", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.8. Incorporating Unlabeled Multi-Task Videos", + "text_level": 1, + "bbox": [ + 89, + 90, + 464, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our proposed framework so far trains an MT-TAS model using only single-task videos. However, we additionally investigate the potential benefits of incorporating unlabeled multi-task videos, when such videos are available. Since annotating long multi-task videos is labor-intensive and time-consuming, leveraging unlabeled multi-task videos offers a practical way to enhance the model performance.", + "bbox": [ + 89, + 112, + 483, + 218 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We formulate this as a domain adaptation problem [9, 10, 50], where labeled single-task and synthetic multi-task videos constitute the source domain, while unlabeled real multi-task videos form the target domain. We aim to learn domain-invariant feature representations through adversarial training, achieved by adding a domain classifier with a Gradient Reversal Layer (GRL) [21] after the base TAS model's early feature layer. The model is trained jointly with action labels on the source domain and domain labels on both domains. During training, the GRL reverses the gradient flow, encouraging indistinguishable feature distributions across domains. Our overall training loss is", + "bbox": [ + 89, + 218, + 483, + 398 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {a c t i o n}} + \\lambda_ {\\mathrm {S B L}} \\mathcal {L} _ {\\mathrm {S B L}} + \\lambda_ {\\mathrm {F B F C}} \\mathcal {L} _ {\\mathrm {F B F C}} + \\lambda_ {\\text {d o m a i n}} \\mathcal {L} _ {\\text {d o m a i n}}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 405, + 480, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{domain}}$ is the framewise binary domain classification loss, and $\\lambda_{\\mathrm{domain}}$ is the loss weight.", + "bbox": [ + 89, + 425, + 482, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. MEKA Dataset", + "text_level": 1, + "bbox": [ + 89, + 467, + 246, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While our proposed framework can be trained using only single-task videos, we require an MT-TAS benchmark for evaluation. To this end, we introduce the Multi-task Ego-centric Kitchen Activities (MEKA) dataset, which extends the EgoPER dataset [36] to multi-task scenarios. EgoPER is an egocentric procedural video dataset designed for temporal action segmentation and error detection. It contains single-task videos of users performing kitchen activities while wearing an egocentric camera (HoloLens 2). The dataset includes five cooking tasks: making pinwheels, quesadillas, oatmeal, coffee, and tea, each associated with a predefined task graph.", + "bbox": [ + 89, + 492, + 483, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To create realistic multi-task scenarios, we generated multi-task transcripts by interleaving action sequences from multiple tasks while ensuring logical coherence within the task graphs. Participants equipped with a HoloLens 2 camera followed these transcripts during recording. While we used similar recording equipment and kitchen tools as in the EgoPER dataset to maintain consistency, variations in the environments introduce additional challenges to the evaluation benchmark.", + "bbox": [ + 89, + 674, + 483, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After data collection, we annotated each video by assigning framewise labels for both actions and tasks. In total, MEKA consists of 100 videos for 12 hours of footage. In each video, participants perform two or three cooking tasks in an interleaved manner. See the supplementary materials for more details about the dataset. We will make the", + "bbox": [ + 89, + 810, + 483, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MEKA dataset publicly available upon publication, including videos and framewise annotations, to facilitate further research in multi-task temporal action segmentation.", + "bbox": [ + 511, + 90, + 906, + 137 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 154, + 648, + 171 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 180, + 705, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. Since MT-TAS is a new task, existing datasets cannot be used for its evaluation. Consequently, we evaluate our approach using single-task videos from EgoPER [36] for training and our collected multi-task videos from MEKA for testing. EgoPER contains 213 normal single-task egocentric videos and MEKA contains 100 multi-task videos for the five recipes in EgoPER. In total, it has 51 classes of action steps. We were also able to adapt the EGTEA dataset [43] for multi-task evaluation. We manually select actions that appear exclusively in either the first or second half of each video, treating these segments as single-task videos, and consider the full videos with interleaved actions as multi-task videos. The supplementary materials provide additional details and the experimental results on the adapted EGTEA dataset.", + "bbox": [ + 511, + 203, + 906, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model Architecture. We evaluate our method with three base TAS models: two offline models (MSTCN [17] and FACT [49]) and one online model (ProTAS [65]). Online TAS differs from offline TAS in that it only uses the frames up to the current moment for prediction. We implement the SBL module $(\\mathcal{F}_{\\mathrm{SBL}})$ by processing interval features through two convolutional layers, followed by temporal global average pooling. The Decoder $\\mathcal{D}$ processes foreground and background features through parallel two-layer MLPs, each reducing feature dimension by half, followed by a linear layer that reconstructs the original feature representation from the concatenated features. We employ a three-layer TCN [17] with a hidden dimension of 64 for FAAR $(\\mathcal{R})$ . For experiments with domain adaptation, we insert a GRL after the feature layer of the initial stage of TAS models, and use a two-layer MLP for binary domain classification.", + "bbox": [ + 511, + 431, + 908, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Following standard practice [17, 49, 65, 81, 84], we extract 2048-dimensional I3D features [7] pretrained on Kinetics [7] as video representations. We evaluate performance using standard TAS metrics: frame-wise accuracy (Acc), frame-wise accuracy excluding background frames (Acc-bg), segment-wise edit score (Edit), and segment-wise F1 scores at multiple IoU thresholds (F1@10,25,50). We adopt the default configurations from MSTCN [17], FACT [49], and ProTAS [65] using their official source codes. We train all models using Adam optimizer with a learning rate of 0.0005. We set the loss weights as $\\lambda_{\\mathrm{SBL}} = \\lambda_{\\mathrm{FBFC}} = \\lambda_{\\mathrm{domain}} = 1$ . For the SBL module, we define boundary frames as the 10 frames centered at the points where segments from different videos are concatenated, and set window size $w$ as 5 and margin", + "bbox": [ + 511, + 674, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19125", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c564bd9dfc416d7b87ced4b04c424de16e6b98cac0088df7ad5e7524d64d879b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MSBSBLFBFCFAARTAS model: MSTCN [17]TAS model: FACT [49]
AccAcc-bgEditF1@{10,25,50}AccAcc-bgEditF1@{10,25,50}
baseline62.349.150.152.649.339.753.741.442.944.641.631.7
✓67.863.764.268.966.856.072.867.873.173.170.058.4
✓✓68.865.964.770.868.858.173.670.474.576.173.662.3
✓✓✓73.764.469.573.069.958.873.871.374.479.676.765.5
✓✓✓✓75.772.174.979.777.667.475.772.876.081.279.369.9
", + "bbox": [ + 101, + 88, + 893, + 214 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ee1a6cc12ff026ca31f80a58a099142861da7e77d621b58ed88820d8aebae381.jpg", + "table_caption": [ + "Table 1. Offline multi-task temporal action segmentation performance on MEKA." + ], + "table_footnote": [], + "table_body": "
MSBSBLFBFCFAARAccEditF1@{10,25,50}
baseline49.124.624.120.012.4
✓51.130.726.522.112.8
✓✓51.133.129.324.514.6
✓✓✓55.237.532.827.118.3
✓✓✓✓67.854.854.649.836.7
", + "bbox": [ + 96, + 241, + 478, + 348 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\delta$ as 2. For Foreground-Background Feature Composition (FBFC), we select $10\\%$ of frames and augment them with Eq. (7) by randomly sampling $\\beta$ from [0.5, 1] during training. See Sec. 5.2 for the qualitative analysis on the value of $\\beta$ . For the FAAR module, we employ a two-stage training process. In the first stage, we train the model without FAAR by setting $\\alpha = 0$ in Eq. (9). In the second stage, we incorporate FAAR and randomly sample $\\alpha$ from [0, 1] during training. For inference, we set $K = 3$ and $\\alpha = 0.3$ in Eq. (10). We include ablation studies on the values of $K$ and $\\alpha$ in Sec. 5.2. Please refer to the supplementary materials for more implementation details.", + "bbox": [ + 88, + 385, + 482, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Experimental Results", + "text_level": 1, + "bbox": [ + 89, + 575, + 294, + 592 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Offline Temporal Action Segmentation. Tab. 1 presents the performance of our approach on the MEKA dataset using two baseline offline TAS models: MSTCN [17] and FACT [49]. We start with the baseline models trained on single-task videos without any of our proposed modules (first row). We then progressively incorporate the Multi-task Sequence Blending (MSB), Segment Boundary Learning (SBL), Foreground-Background Feature Composition (FBFC), and Foreground-Aware Action Refinement (FAAR) modules to assess their individual contributions. In the first row, MSTCN achieves an accuracy of $62.3\\%$ and an F1@50 score of $39.7\\%$ , while FACT attains $53.7\\%$ accuracy and $31.7\\%$ F1@50, indicating that the baseline models struggle to generalize to multi-task videos when trained solely on single-task data. Incorporating MSB significantly improves performance for both models, especially for FACT, due to its explicit modeling of video transcripts during training. Adding SBL further boosts performance. With MSB and SBL, MSTCN's accuracy improves by $6.5\\%$ and F1@50 by $18.4\\%$ , while FACT's accuracy", + "bbox": [ + 91, + 598, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0040405f1705a89fedc47e2cf25308e0e97a16a290b211ba6bd23f27cdb76022.jpg", + "table_caption": [ + "Table 2. Online temporal action segmentation on MEKA." + ], + "table_footnote": [], + "table_body": "
Training SetupAccEditF1@{10,25,50}
MSTCN [17]Offline Temporal Action Segmentation
LMT78.075.978.376.670.4
LST&LMT80.684.784.582.574.6
Oursw/o DA75.774.979.777.667.4
w/ DA76.6+0.977.6+2.782.0+2.380.6+3.072.2+4.8
+LMT82.787.389.788.582.4
ProTAS [65]Online Temporal Action Segmentation
LMT61.541.137.232.822.0
LST&LMT64.549.839.434.522.7
Oursw/o DA67.854.854.649.836.7
w/ DA69.1+1.356.9+2.157.5+2.952.0+2.238.8+2.1
+LMT74.970.667.262.549.3
", + "bbox": [ + 517, + 241, + 903, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. The effect of domain adaptation using unlabeled multi-task videos and comparison with methods using labeled multi-task videos.", + "bbox": [ + 511, + 412, + 903, + 438 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "improves by $19.9\\%$ and F1@50 by $30.6\\%$ . Incorporating FBFC further improves the accuracy by $4.9\\%$ on MSTCN and the F1@50 by $3.2\\%$ on FACT. Finally, integrating all proposed modules leads to the highest performance, resulting in an F1@50 score of $67.4\\%$ for MSTCN and $69.9\\%$ for FACT. Overall, our approach effectively enhances the capability of standard TAS models to handle multi-task videos.", + "bbox": [ + 509, + 460, + 906, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Online Temporal Action Segmentation. We evaluate the online TAS performance of our approach in Tab. 2, using ProTAS [65] as the base TAS model. Without access to future frames, the online model relies solely on past and current frames, making it challenging to anticipate action transitions and accurately segment boundaries. This limitation results in a noticeable performance drop compared to the offline models, echoing the difficulties of online MT-TAS. However, our proposed modules still achieve a notable performance boost. When all modules are combined, the model increases accuracy by $18.7\\%$ and F1@50 by $24.3\\%$ compared to the baseline. Particularly, the integration of FAAR improves the accuracy by $12.6\\%$ and all F1 scores by nearly $20\\%$ , highlighting the value of foreground-aware refinements in scenarios without access to future frames.", + "bbox": [ + 509, + 566, + 908, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Domain Adaptation with Unlabeled Multi-Task Videos. In Tab. 3, we evaluate our domain adaptation approach using unlabeled multi-task videos, as introduced in Sec. 3.8. Domain adaptation consistently improves performance across all metrics, increasing F1@50 by $4.8\\%$ and $2.1\\%$ in offline and online settings, respectively. To thoroughly assess our method, we compare against two baseline se", + "bbox": [ + 509, + 795, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19126", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cd67e2174538760dc4ad57c12e056e00f6cafa5b5a3bc11836a6d91db026c76e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 114, + 90, + 369, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c0e86034b9c548275726504e3c6c29173a7b73eed56da81c37a6534b511e4fd9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 90, + 622, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5b7ee7fa9602c134e99b580adea4ef2bd8b115f5399e9dc22822167b24a51fe5.jpg", + "image_caption": [ + "Figure 4. Ablation Studies on FAAR: (a) comparison of different inputs; (b) effects of $K$ ; (c) effects of $\\alpha$ ." + ], + "image_footnote": [], + "bbox": [ + 622, + 90, + 879, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tups using 3-fold cross-validation on multi-task videos: 1) the model is trained on labeled multi-task (LMT) videos; 2) the model is trained on both labeled single-task and labeled multi-task (LST&LMT) videos. Note that these baselines use labeled multi-task data, while our method employs only unlabeled multi-task videos during training. In offline TAS, our model with domain adaptation outperforms the LMT baseline and achieves comparable results to the LST&LMT baseline. In the more challenging online setting, our models surpass both baselines, demonstrating the ability of our modules to enhance generalization and handle the data-intensive demands of online TAS. We also extend our method by incorporating labeled multi-task videos (+LMT) to provide an upper bound for domain adaptation. The performance gap reveals room for improvement in unsupervised domain adaptation. Furthermore, the performance gains from LST&LMT to Ours+LMT demonstrate that our modules provide complementary benefits even when labeled multi-task data is available.", + "bbox": [ + 88, + 292, + 485, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Studies on FAAR. In Fig. 4, we conduct ablation studies to evaluate different inputs of FAAR and analyze the effects of $K$ and $\\alpha$ in Eq. (10) during inference. First, we change the module's input from foreground frames to background frames or full images. Background frames perform poorly while foreground frames outperform full images, showing the importance of action-relevant foreground information. We then examine how varying $K$ , the number of top predicted action classes considered, and $\\alpha$ , the weighting factor between initial predictions and foreground-aware predictions, influences the performance. The model's performance gradually improves with increasing $K$ , reaching a plateau when $K \\geq 3$ . As $\\alpha$ increases from 0 to 1, the model performs best between 0.3 and 0.4. Notably, while the accuracies at $\\alpha = 0$ (base TAS predictions only) and $\\alpha = 1$ (foreground-aware predictions only) are similar, the Edit and F1@50 scores for foreground-only predictions are significantly lower. This suggests that FAAR emphasizes individual frame accuracy but may lack the temporal consistency needed for segment-wise metrics.", + "bbox": [ + 89, + 580, + 486, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Analysis of FBFC. In Fig. 5, we qualitatively", + "bbox": [ + 89, + 885, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/70d9736337e0a31e03bbbc8769e4c0cedf2830271b2cd3e5c21f9f83fa5759a1.jpg", + "image_caption": [ + "Figure 5. Retrieved nearest multi-task videos by different ratios with FBFC. For simplicity, only middle frames of video clips are displayed." + ], + "image_footnote": [], + "bbox": [ + 526, + 290, + 893, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "analyze how FBFC works with different values of the mixing ratio $\\beta$ . We sample two clips from single-task videos, extract foreground frames from one clip and background frames from another, and then recompose these features using FBFC via Eqs. (7) and (8). To assess the decomposition quality, we retrieve the nearest clip from multi-task testing videos based on feature similarity. With a lower mixing ratio ( $\\beta = 0.2$ ), the recomposed features are more heavily influenced by background features, and the context of the retrieved frames predominantly match the background frames. Conversely, with a higher mixing ratio ( $\\beta = 0.8$ ), the retrieved frames often depict the same action as the foreground frames while sharing some similarities with the context of the background frames. Based on these observations, we sample $\\beta \\in [0.5,1]$ in training to ensure preserving action features while introducing background variations.", + "bbox": [ + 511, + 494, + 906, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "See supplementary materials for more results, qualitative visualization, and discussion on complexity and limitations.", + "bbox": [ + 511, + 737, + 906, + 767 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 784, + 640, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed an MT-TAS framework to tackle the challenges of interleaved actions in multi-task scenarios. Our approach combines synthesized multi-task data, a two-stage TAS model, and the new MEKA benchmark dataset. We validated the effectiveness of our approach through extensive experiments in both offline and online MT-TAS.", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19127", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is sponsored by ARPA-H (1AY2AX000062), DARPA PTG (HR00112220001), NSF (IIS-2115110) and ARO (W911NF2110276). Content does not necessarily reflect the position/policy of the Government.", + "bbox": [ + 89, + 114, + 483, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 188, + 187, + 204 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hyemin Ahn and Dongheui Lee. Refining action segmentation with hierarchical video representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16302-16310, 2021. 1, 2", + "[2] J. B. Alayrac, P. Bojanowski, N. Agrawal, J. Sivic, I. Laptev, and S. Lacoste-Julien. Unsupervised learning from narrated instruction videos. IEEE Conference on Computer Vision and Pattern Recognition, 2016. 2", + "[3] Kyungho Bae, Geo Ahn, Youngrae Kim, and Jinwoo Choi. Devias: Learning disentangled video representations of action and scene for holistic video understanding. In European Conference on Computer Vision, 2024. 3", + "[4] Emad Bahrami, Gianpiero Francesca, and Juergen Gall. How much temporal long-term context is needed for action segmentation? In International Conference on Computer Vision (ICCV), 2023. 2", + "[5] Siddhant Bansal, Chetan Arora, and C.V. Jawahar. My view is the best view: Procedure learning from egocentric videos. In European Conference on Computer Vision (ECCV), 2022. 1, 2", + "[6] Nadine Behrmann, S. Alireza Golestaneh, Zico Kolter, Juergen Gall, and Mehdi Noroozi. Unified fully and timestamp supervised temporal action segmentation via sequence to sequence translation. In ECCV, 2022. 1, 2", + "[7] J. Carreira and A. Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In IEEE Conference on Computer Vision and Pattern Recognition, 2017. 4, 6", + "[8] Chien-Yi Chang, De-An Huang, Yanan Sui, Li Fei-Fei, and Juan Carlos Niebles. D3tw: Discriminative differentiable dynamic time warping for weakly supervised action alignment and segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2", + "[9] Min-Hung Chen, Baopu Li, Yingze Bao, Ghassan Al-Regib, and Zsolt Kira. Action segmentation with joint self-supervised temporal domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9454–9463, 2020. 6", + "[10] Victor G. Turrisi da Costa, Giacomo Zara, Paolo Rota, Thiago Oliveira-Santos, Nicu Sebe, Vittorio Murino, and Elisa Ricci. Dual-head contrastive domain adaptation for video action recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1181-1190, 2022. 6", + "[11] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision (IJCV), 130:33-55, 2022. 2" + ], + "bbox": [ + 93, + 213, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Li Ding and Chenliang Xu. Weakly-supervised action segmentation with iterative soft boundary assignment. IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2", + "[13] Shuangrui Ding, Maomao Li, Tianyu Yang, Rui Qian, Haohang Xu, Qingyi Chen, Jue Wang, and Hongkai Xiong. Motion-aware contrastive video representation learning via foreground-background merging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9716-9726, 2022. 3", + "[14] G. Donahue and E. Elhamifar. Learning to predict activity progress by self-supervised video alignment. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1", + "[15] E. Elhamifar and D. Huynh. Self-supervised multi-task procedure learning from instructional videos. European Conference on Computer Vision, 2020. 2", + "[16] E. Elhamifar and Z. Naing. Unsupervised procedure learning via joint dynamic summarization. International Conference on Computer Vision, 2019. 2", + "[17] Yazan Abu Farha and Jurgen Gall. Ms-tcn: Multi-stage temporal convolutional network for action segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3575-3584, 2019. 1, 2, 4, 6, 7", + "[18] Alireza Fathi, Xiaofeng Ren, and James M Rehg. Learning to recognize objects in egocentric activities. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2011. 2", + "[19] Mohsen Fayyaz and Jurgen Gall. Sct: Set constrained temporal transformer for set supervised action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[20] Daniel Fried, Jean-Baptiste Alayrac, Phil Blunsom, Chris Dyer, Stephen Clark, and Aida Nematzadeh. Learning to segment actions from observation and narration. Annual Meeting of the Association for Computational Linguistics, 2020. 2", + "[21] Y. Ganin and V. Lempitsky. Unsupervised domain adaptation by backpropagation. International Conference on Machine Learning, 2015. 6", + "[22] Reza Ghoddoosian, Isht Dwivedi, Nakul Agarwal, Chiho Choi, and Behzad Dariush. Weakly-supervised online action segmentation in multi-view instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13780-13790, 2022. 1, 2", + "[23] Reza Ghoddosian, Saif Sayed, and Vassilis Athitsos. Hierarchical modeling for task recognition and action segmentation in weakly-labeled instructional videos. IEEE Winter Conference on Applications of Computer Vision, 2022.", + "[24] Reza Ghoddosian, Isht Dwivedi, Nakul Agarwal, and Behzad Dariush. Weakly-supervised action segmentation and unseen error detection in anomalous instructional videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10128-10138, 2023. 1, 2", + "[25] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Mar" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19128", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tin, Tushar Nagarajan, Ilija Radosavovic, Santhosh K. Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Z. Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Christian Fuegen, Abraham Gebreselasie, Cristina Gonzalez, James M. Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolár, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran K. Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Yunyi Zhu, Pablo Arbeláez, David J. Crandall, Dima Damen, Giovanni Maria Farinella, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard A. Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the world in 3,000 hours of egocentric video. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18973-18990, 2021. 2", + "[26] Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, et al. Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19383-19400, 2024. 2", + "[27] Yifei Huang, Yusuke Sugano, and Yoichi Sato. Improving action segmentation via graph-based temporal reasoning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[28] Yifei Huang, Guo Chen, Jilan Xu, Mingfang Zhang, Lijin Yang, Baoqi Pei, Hongjie Zhang, Lu Dong, Yali Wang, Limin Wang, and Yu Qiao. Egoexolearn: A dataset for bridging asynchronous ego- and exo-centric view of procedural activities in real world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22072-22086, 2024. 2", + "[29] Yuchi Ishikawa, Seito Kasai, Yoshimitsu Aoki, and Hirokatsu Kataoka. Alleviating over-segmentation errors by detecting action boundaries. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2322-2331, 2021. 2", + "[30] Hilde Kuehne, Ali Arslan, and Thomas Serre. The language of actions: Recovering the syntax and semantics of goal-directed human activities. IEEE Conference on Computer Vision and Pattern Recognition, 2014. 2", + "[31] H. Kuehne, A. Arslan, and T. Serre. The language of actions: Recovering the syntax and semantics of goal-directed human. IEEE Conference on Computer Vision and Pattern Recognition, 2014. 1, 2", + "[32] H. Kuehne, J. Gall, and T. Serre. An end-to-end generative framework for video segmentation and recognition." + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE Winter Conference on Applications of Computer Vision, 2016. 2", + "[33] Anna Kukleva, Hilde Kuehne, Fadime Sener, and Jurgen Gall. Unsupervised learning of action classes with continuous temporal embedding. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1, 2", + "[34] Sateesh Kumar, Sanjay Haresh, Awais Ahmed, Andrey Konin, M. Zeeshan Zia, and Quoc-Huy Tran. Unsupervised action segmentation by joint representation learning and online clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20174-20185, 2022. 2", + "[35] C. Lea, M. D. Flynn, R. Vidal, A. Reiter, and G. D. Hager. Temporal convolutional networks for action segmentation and detection. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2", + "[36] S. Lee, Z. Lu, Z. Zhang, M. Hoai, and E. Elhamifar. Error detection in egocentric procedural task videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 6", + "[37] Haoxin Li, Yuan Liu, Hanwang Zhang, and Boyang Li. Mitigating and evaluating static bias of action representations in the background and the foreground. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19911-19923, 2023. 3", + "[38] Jun Li and Sinisa Todorovic. Set-constrained viterbi for set-supervised action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[39] J. Li and S. Todorovic. Anchor-constrained viterbi for set-supervised action segmentation. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021.", + "[40] J. Li, P. Lei, and S. Todorovic. Weakly supervised energy-based learning for action segmentation. International Conference on Computer Vision, 2019. 2", + "[41] M. Li, L. Chen, Y. Duarr, Z. Hu, J. Feng, J. Zhou, and J. Lu. Bridge-prompt: Towards ordinal action understanding in instructional videos. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2", + "[42] Shi-Jie Li, Yazan AbuFarha, Yun Liu, Ming-Ming Cheng, and Juergen Gall. Ms-tcn++: Multi-stage temporal convolutional network for action segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2020. 1, 2", + "[43] Yin Li, Miao Liu, and James M. Rehg. In the eye of beholder: Joint learning of gaze and actions in first person video. European Conference on Computer Vision, 2018. 2, 6", + "[44] Zhe Li, Yazan Abu Farha, and Jurgen Gall. Temporal action segmentation from timestamp supervision. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2", + "[45] Daochang Liu, Qiyue Li, Anh-Dung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. International Conference on Computer Vision, 2023. 1, 2, 4", + "[46] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19129", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. *eccv*, 2024. 5", + "[47] Z. Lu and E. Elhamifar. Weakly-supervised action segmentation and alignment via transcript-aware union-of-subspaces learning. International Conference on Computer Vision, 2021. 2", + "[48] Z. Lu and E. Elhamifar. Set-supervised action learning in procedural task videos via pairwise order consistency. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 1, 2", + "[49] Z. Lu and E. Elhamifar. Fact: Frame-action cross-attention temporal modeling for efficient action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 4, 6, 7", + "[50] J. Munro and D. Damen. Multi-modal domain adaptation for fine-grained action recognition. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 6", + "[51] Zhanzhong Pang, Fadime Sener, Shrinivas Ramasubramanian, and Angela Yao. Long-tail temporal action segmentation with group-wise temporal logit adjustment. 2024. 2", + "[52] Sam Perochon and Laurent Oudre. Unsupervised action segmentation of untrimmed egocentric videos. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 1", + "[53] Will Price, Carl Vondrick, and Dima Damen. Unweavenet: Unweaving activity stories. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13770-13779, 2022. 2", + "[54] Camillo Quattrocchi, Antonino Furnari, Daniele Di Mauro, Mario Valerio Giuffrida, and Giovanni Maria Farinella. Synchronization is all you need: Exocentric-to-egocentric transfer for temporal action segmentation with unlabeled synchronized video pairs. European Conference on Computer Vision, 2024. 1", + "[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. International Conference on Machine learning, 2021. 5", + "[56] Francesco Ragusa, Antonino Furnari, Salvatore Livatino, and Giovanni Maria Farinella. The meccano dataset: Understanding human-object interactions from egocentric videos in an industrial-like domain. Winter Conference on Applications of Computer Vision, 2021. 2", + "[57] Rahul Rahaman, Dipika Singhania, Alexandre Thiery, and Angela Yao. A generalized and robust framework for timestamp supervision in temporal action segmentation. In Computer Vision-ECCV 2022: 17th European Conference, 2022. 2", + "[58] A. Richard, H. Kuehne, and J. Gall. Action sets: Weakly supervised action segmentation without ordering constraints. IEEE Conference on Computer Vision and Pattern Recognition, 2018.", + "[59] A. Richard, H. Kuehne, A. Iqbal, and J. Gall. Neuralnetwork-viterbi: A framework for weakly supervised video learning. IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[60] M. Rohrbach, S. Amin, M. Andriluka, and B. Schiele. A database for fine grained activity detection of cooking activities. IEEE Conference on Computer Vision and Pattern Recognition, 2012. 2", + "[61] Saquib Sarfraz, Naila Murray, Vivek Sharma, Ali Diba, Luc Van Gool, and Rainer Stiefelhagen. Temporally-weighted hierarchical clustering for unsupervised action segmentation. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2", + "[62] Saif Sayed, Reza Ghoddoosian, Bhaskar Trivedi, and Vassilis Athitsos. A new dataset and approach for timestamp supervised action segmentation using human object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 3133-3142, 2023. 2", + "[63] Fadime Sener, Dibyadip Chatterjee, Daniel Shelepov, Kun He, Dipika Singhania, Robert Wang, and Angela Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21096-21106, 2022. 2", + "[64] Y. Shen and E. Elhamifar. Semi-weakly-supervised learning of complex actions from instructional task videos. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 1, 2", + "[65] Y. Shen and E. Elhamifar. Progress-aware online action segmentation for egocentric procedural task videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 6, 7", + "[66] Y. Shen, L. Wang, and E. Elhamifar. Learning to segment actions from visual and language instructions via differentiable weak sequence alignment. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2", + "[67] Y. Shen, H. Wang, X. Yang, M. Feiszli, E. Elhamifar, L. Torresani, and E. Mavroudi. Learning to segment referred objects from narrated egocentric videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 3", + "[68] G. A. Sigurdsson, S. Divvala, A. Farhadi, and A. Gupta. Asynchronous temporal fields for action recognition. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2", + "[69] B. Singh, T. K. Marks, M. Jones, O. Tuzel, and M. Shao. A multi-stream bi-directional recurrent neural network for finegrained action detection. IEEE Conference on Computer Vision and Pattern Recognition, 2016.", + "[70] Dipika Singhania, Rahul Rahaman, and Angela Yao. C2f- tcn: A framework for semi-and fully-supervised temporal action segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):11484-11501, 2023.", + "[71] Yaser Souri, Mohsen Fayyaz, Luca Minciullo, Gianpiero Francesca, and Juergen Gall. Fast Weakly Supervised Action Segmentation Using Mutual Consistency. PAMI, 2021. 2", + "[72] Sebastian Stein and Stephen J McKenna. Combining embedded accelerometers with computer vision for recognizing food preparation activities. In Proceedings of the 2013 ACM international joint conference on Pervasive and ubiquitous computing, pages 729-738, 2013. 2" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19130", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[73] Yuhao Su and Ehsan Elhamifar. Two-stage active learning for efficient temporal action segmentation. In European Conference on Computer Vision, pages 161-183. Springer, 2024. 1", + "[74] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[75] Sanket Thakur, Cigdem Bayan, Pietro Morerio, Vittorio Murino, and Alessio Del Bue. Anticipating next active objects for egocentric videos. IEEE Access, 2024. 3", + "[76] Quoc-Huy Tran, Ahmed Mehmood, Muhammad Ahmed, Muhammad Naufil, Anas Zafar, Andrey Konin, and Zeeshan Zia. Permutation-aware activity segmentation via unsupervised frame-to-segment alignment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6426-6436, 2024. 2", + "[77] Huiyu Wang, Mitesh Kumar Singh, and Lorenzo Torresani. Ego-only: Egocentric action detection without exocentric transferring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5250-5261, 2023. 1", + "[78] Jiahao Wang, Guo Chen, Yifei Huang, Limin Wang, and Tong Lu. Memory-and-anticipation transformer for online action understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13824–13835, 2023. 2", + "[79] Zhe Wang, Hao Chen, Xinyu Li, Chunhui Liu, Yuanjun Xiong, Joseph Tighe, and Charless Fowlkes. Sscap: Self-supervised co-occurrence action parsing for unsupervised temporal action segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1819–1828, 2022. 2", + "[80] Ming Xu and Stephen Gould. Temporally consistent unbalanced optimal transport for unsupervised action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2", + "[81] Ziwei Xu, Yogesh S Rawat, Yongkang Wong, Mohan Kankanhalli, and Mubarak Shah. Don't pour cereal into coffee: Differentiable temporal logic for temporal action segmentation. In Advances in Neural Information Processing Systems, 2022. 6", + "[82] Lingfeng Yang, Yueze Wang, Xiang Li, Xinlong Wang, and Jian Yang. Fine-grained visual prompting. Advances in Neural Information Processing Systems, 36, 2023. 5", + "[83] S. Yeung, O. Russakovsky, N. Jin, M. Andriluka, G. Mori, and L. Fei-Fei. Every moment counts: Dense detailed labeling of actions in complex videos. International Journal of Computer Vision, 2018. 2", + "[84] Fangqiu Yi, Hongyu Wen, and Tingting Jiang. Asformer: Transformer for action segmentation. In The British Machine Vision Conference (BMVC), 2021. 1, 2, 4, 6", + "[85] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Helping hands: An object-aware ego-centric video recognition model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13901-13912, 2023. 3" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[86] Ce Zhang, Changcheng Fu, Shijie Wang, Nakul Agarwal, Kwonjoon Lee, Chiho Choi, and Chen Sun. Object-centric video representation for long-term action anticipation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6751-6761, 2024. 3", + "[87] Xingyi Zhou, Anurag Arnab, Chen Sun, and Cordelia Schmid. How can objects help action recognition? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2362, 2023. 3", + "[88] D. Zhukov, J. B. Alayrac, R. G. Cinbis, D. Fouhey, I. Laptev, and J. Sivic. Cross-task weakly supervised learning from instructional videos. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1, 2" + ], + "bbox": [ + 516, + 90, + 903, + 273 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "19131", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Understanding Multi-Task Activities from Single-Task Videos/b633800a-1331-4862-80d6-5ccbeba54c12_model.json b/2025/Understanding Multi-Task Activities from Single-Task Videos/b633800a-1331-4862-80d6-5ccbeba54c12_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5572e7eb435a2b7fa1861589199e943f1c120bb8 --- /dev/null +++ b/2025/Understanding Multi-Task Activities from Single-Task Videos/b633800a-1331-4862-80d6-5ccbeba54c12_model.json @@ -0,0 +1,2677 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.189, + 0.131, + 0.812, + 0.154 + ], + "angle": 0, + "content": "Understanding Multi-Task Activities from Single-Task Videos" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.182, + 0.434, + 0.233 + ], + "angle": 0, + "content": "Yuhan Shen Northeastern University shen.yuh@northeastern.edu" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.182, + 0.757, + 0.233 + ], + "angle": 0, + "content": "Ehsan Elhamifar Northeastern University e.elhamifar@northeastern.edu" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.485, + 0.649 + ], + "angle": 0, + "content": "We introduce and develop a framework for Multi-Task Temporal Action Segmentation (MT-TAS), a novel paradigm that addresses the challenges of interleaved actions when performing multiple tasks simultaneously. Traditional action segmentation models, trained on single-task videos, struggle to handle task switches and complex scenes inherent in multi-task scenarios. To overcome these challenges, our MT-TAS approach synthesizes multi-task video data from single-task sources using our Multi-task Sequence Blending and Segment Boundary Learning modules. Additionally, we propose to dynamically isolate foreground and background elements within video frames, addressing the intricacies of object layouts in multi-task scenarios and enabling a new two-stage temporal action segmentation framework with Foreground-Aware Action Refinement. Also, we introduce the Multi-task Egocentric Kitchen Activities (MEKA) dataset, containing 12 hours of egocentric multi-task videos, to rigorously benchmark MT-TAS models. Extensive experiments demonstrate that our framework effectively bridges the gap between single-task training and multi-task testing, advancing temporal action segmentation with state-of-the-art performance in complex environments." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.669, + 0.222, + 0.684 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.695, + 0.484, + 0.876 + ], + "angle": 0, + "content": "Consider a typical morning routine, when you prepare breakfast: you start making oatmeal and while waiting for it to cook, you brew coffee and pack your food for lunch. In fact, in our daily lives, we often perform multiple tasks simultaneously. For computer vision systems to accurately analyze such behaviors and provide useful assistance, they must comprehend these complex and interleaved activities. Temporal action segmentation (TAS) has emerged as a crucial technology for understanding human activities by partitioning videos into meaningful segments, each corresponding to a distinct action or step [17, 22, 24, 31, 33, 48, 49, 64, 73, 84, 88]. Recently, TAS has attracted growing interests" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.269, + 0.897, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.372, + 0.634, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.464, + 0.663, + 0.471 + ], + "angle": 0, + "content": "single-task videos" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.372, + 0.743, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.78, + 0.372, + 0.884, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.8, + 0.464, + 0.859, + 0.47 + ], + "angle": 0, + "content": "multi-task videos" + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.474, + 0.878, + 0.484 + ], + "angle": 0, + "content": "(b) Comparison between frames from single-task videos v.s. multi-task videos" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.487, + 0.907, + 0.561 + ], + "angle": 0, + "content": "Figure 1. (a) Multi-task temporal action segmentation involves segmenting actions from interleaved tasks. It requires capturing interruptions and resumptions in the execution of each task. (b) Comparison between frames from single-task and multi-task videos. In multi-task videos, at each time instant, the scene often contains objects from tasks different from the currently executed one (highlighted by purple boxes)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.907, + 0.656 + ], + "angle": 0, + "content": "within egocentric videos [5, 36, 52, 54, 65, 77], where activities are captured from a first-person perspective, enabling fine-grained analysis of human actions, which is essential for applications such as real-time personalized assistance, embodied agents and efficient video content retrieval." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we study the new problem of Multi-Task Temporal Action Segmentation (MT-TAS), which explicitly focuses on scenarios where users perform multiple tasks by interleaving their steps within a single video, see Fig. 1. There are two major challenges that existing TAS models face in addressing MT-TAS. First, despite the notable progress through deep learning techniques [1, 6, 14, 17, 41, 42, 45, 49, 84], current TAS models have primarily focused on scenarios where each video contains actions related to a single high-level task, such as assembling a piece of furniture or cooking a specific recipe. This single-task focus limits the applicability of TAS models to the multi-task setting since i) they are biased to predict actions/steps from only one task at the inference time, ii) when performing several tasks, the environment and the scene often contains objects relevant to multiple tasks, which adds complexity" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.442, + 0.9 + ], + "angle": 0, + "content": "\\(^{1}\\)Code: https://github.com/Yuhan-Shen/MT-TAS." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19120" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.166 + ], + "angle": 0, + "content": "to the context, while single-task videos often contain objects relevant to one task (see Fig. 1(b)), iii) interleaving steps of multiple tasks introduces complexities such as unpredictable task switches and disruptions to the temporal flows of actions within each task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.482, + 0.304 + ], + "angle": 0, + "content": "Second, a significant obstacle in advancing MT-TAS is the lack of appropriate training data. Existing datasets [5, 16, 18, 31, 36, 63, 88] are predominantly collected for single-task scenarios, making them insufficient for capturing complexities of interleaved tasks. Furthermore, many possible step combinations in the multi-task setting makes it impractical to collect and annotate comprehensive real-world videos covering all potential situations (such videos would also be much longer than single-task videos)." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.31, + 0.485, + 0.627 + ], + "angle": 0, + "content": "Paper Contributions. We study the new problem of MT-TAS. To overcome the aforementioned challenges, we propose an MT-TAS framework that uses only single-task videos for training. We generate synthetic multi-task data using single-task videos and propose a two-stage action segmentation model for accurate predictions. For the data generation, we first develop a Multi-task Sequence Blending (MSB) module that leverages Large Language Models (LLMs) to create multi-task transcripts of interleaved steps and their associated videos. Second, we develop a Segment Boundary Learning (SBL) module to mitigate the abrupt transitions at segment boundaries caused by the synthetic data. Third, we introduce a Dynamic Isolation of Video Elements (DIVE) approach to isolate foreground and background components within video frames by identifying action-relevant objects and combining foregrounds of actions in one task with backgrounds of other tasks to capture the more complex scenes of multi-task videos. By increasing the variation in the background components for the training data, our method becomes more robust to unseen object compositions in multi-task scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.482, + 0.764 + ], + "angle": 0, + "content": "For the action segmentation model, we develop a two-stage framework, where we refine the initial action predictions using a Foreground-Aware Action Refinement (FAAR) module. This approach leverages foreground information to enhance the focus of our model on action-relevant cues (e.g., focusing on the regions of kettle and measuring cup for the action \"transfer water to kettle\" as in Fig. 3(b)), improving accuracy and robustness to background variations during the multi-task testing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.482, + 0.902 + ], + "angle": 0, + "content": "For proper evaluation of MT-TAS models, we gather and annotate the Multi-task Egocentric Kitchen Activities (MEKA) video dataset (we plan to publicly release the dataset), built upon the recent EgoPER [36] dataset. MEKA comprises 100 egocentric videos of users (about 12 hours of footage) performing multiple kitchen recipes from EgoPER in each run by interleaving their steps. This new benchmark captures a variety of interleaved kitchen activities, providing a rigorous testing ground for MT-TAS models. Through" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "extensive experiments in both offline and online settings, we demonstrate that our proposed framework successfully bridges the gap between single-task training data and the complexities of real-world multi-task scenarios, advancing research in temporal action segmentation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.178, + 0.665, + 0.194 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.204, + 0.905, + 0.565 + ], + "angle": 0, + "content": "Temporal Action Segmentation. TAS aims to understand the sequential structure of human activities in videos by assigning an action label to each frame, thereby partitioning the video into a sequence of action segments. Based on the level of supervision during training, TAS methods can be categorized into three main groups: unsupervised [2, 15, 16, 20, 33, 34, 61, 66, 76, 79, 80, 88], weakly-supervised [8, 12, 19, 22-24, 38-40, 44, 47, 48, 57-59, 62, 64, 71], and fully-supervised [1, 4, 6, 17, 27, 29, 32, 35, 36, 41, 42, 45, 49, 51, 60, 65, 68-71, 83, 84] approaches. Our work primarily aligns with the fully-supervised methods. However, regardless of the supervision level, previous studies have predominantly focused on single-task TAS, where test videos contain actions from only one high-level activity or task, while we focus on multi-task TAS. A closely related work, UnweaveNet [53], aims to predict the start, continuation, and resumption of multiple activity threads within a video. While UnweaveNet models multiple concurrent activities, its predictions are at the activity or task level and do not capture the fine-grained action steps within each task. In contrast, our approach focuses on recognizing and segmenting detailed action steps across interleaved tasks, providing a more granular understanding of multi-task activities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.905, + 0.884 + ], + "angle": 0, + "content": "Procedural Video Datasets. To support research in TAS and related tasks, numerous procedural video datasets have been collected and released. These datasets include third-person videos (e.g., Breakfast [30], 50Salads [72], YouTube Instructional [2], COIN [74], CrossTask [88], ProceL [16], and ATA [24]), egocentric videos (e.g., GTEA [18], EGTEA [43], MECCANO [56], EgoProceL [5], HoloAssist [78], and EgoPER [36]), or both views such as Assembly101 [63], \\((3 + 1)\\mathrm{ReC}\\) [62] and EgoExoLearn [28]. While large-scale video datasets like EPIC-Kitchens [11], Ego4D [25], and Ego-Exo4D [26] exist, they primarily focus on detecting and recognizing atomic action steps rather than modeling the procedural relationships between steps within a task, making them less suitable for TAS research. Also, most existing TAS datasets are limited to single-task scenarios, where each video contains actions related to one high-level activity. In contrast, our collected MEKA dataset specifically addresses multi-task temporal action segmentation by capturing videos where multiple tasks are performed in an interleaved manner, providing a more challenging and realistic benchmark for advancing TAS research." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Object-Aware and Foreground/Background Learning." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19121" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.304 + ], + "angle": 0, + "content": "Learning to localize active objects or relevant objects is a prominent topic in video understanding [67, 75, 85, 86]. Furthermore, researchers have explored leveraging object information to improve video representation learning [85], action recognition [87], and action anticipation [86]. Some studies have also focused on disentangling foreground and background elements to enhance video representations [3, 13, 37]. Our work differs by specifically addressing the generalization of TAS models to multi-task videos. Unlike existing approaches that concentrate on short video clips or single-task videos, we propose a method that dynamically isolates foreground and background elements to handle unseen object compositions in multi-task scenarios, thereby mitigating interference from irrelevant objects." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.322, + 0.475, + 0.34 + ], + "angle": 0, + "content": "3. Multi-Task Temporal Action Segmentation" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.348, + 0.254, + 0.365 + ], + "angle": 0, + "content": "3.1. Problem Setting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.372, + 0.484, + 0.57 + ], + "angle": 0, + "content": "Given a video sequence \\(\\mathcal{V} = (v_{1},\\ldots ,v_{T})\\) of \\(T\\) frames, the goal of TAS is to predict the sequence of action labels \\(A = (a_{1},\\dots,a_{T})\\), where each label \\(a_{t}\\) comes from a predefined set of action classes \\(\\mathcal{C}\\). Assume we have \\(O\\) tasks, where the set \\(\\mathcal{C}_o\\) consists of actions for the task \\(o\\in \\{1,\\dots,O\\}\\) and \\(\\mathcal{C} = \\bigcup_{o = 1}^{O}\\mathcal{C}_{o}\\) contains actions of all tasks. In a single-task video, actions come from one \\(\\mathcal{C}_o\\), while a multi-task video has interleaved actions from several tasks, hence, \\(a_{t}\\) may belong to any \\(\\mathcal{C}_o\\) for \\(o\\in \\{1,\\dots,O\\}\\). Additionally, there may be background segments (e.g., 'answering phone') irrelevant to any of the tasks. A particular challenge in MT-TAS is to distinguish between action segments belonging to different tasks and background segments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.572, + 0.484, + 0.633 + ], + "angle": 0, + "content": "In our setting, the training data consists of single-task videos with full framewise annotations. Our goal is to train a model on single-task videos so that it can effectively generalize to multi-task videos during testing." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.636, + 0.392, + 0.652 + ], + "angle": 0, + "content": "3.2. Overview of Proposed Framework" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Fig. 2 shows the overview of our MS-TAS framework. Our approach first employs Multi-task Sequence Blending (MSB) to generate synthetic multi-task videos by leveraging LLM-guided task transitions, eliminating the need for multi-task data collection. However, the initial synthesized multi-task videos contain abrupt transitions at segment boundaries, which disrupt the temporal flow, and contain discrepancies in object compositions between single-task and multi-task videos. To mitigate these issues, we introduce three components: a Segment Boundary Learning (SBL) module that enhances temporal smoothness between segments, a Dynamic Isolation of Video Elements approach that augments the background variations during training, and a two-stage TAS method with foreground-aware action refinement to improve the predictions of the model. Next, we discuss each component of our framework in details." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.093, + 0.9, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.47, + 0.907, + 0.513 + ], + "angle": 0, + "content": "Figure 2. Framework Overview. The modules marked with \\(\\Theta\\) are used exclusively for training, and modules marked with \\(\\diamond\\) are used for both training and testing." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.534, + 0.836, + 0.55 + ], + "angle": 0, + "content": "3.3. Multi-task Sequence Blending (MSB)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.557, + 0.907, + 0.762 + ], + "angle": 0, + "content": "The goal of MSB is to create initial multi-task video sequences from single-task videos, to simulate real-world multitasking behaviors. Given single-task training videos \\(\\{\\mathcal{V}^{(i)}\\}_{i=1}^{M}\\), where each video \\(\\mathcal{V}^{(i)} = (v_{1}^{(i)}, v_{2}^{(i)}, \\ldots, v_{T_{i}}^{(i)})\\) has a task label \\(\\tau_{i}\\) and consists of a sequence of action segments \\(S^{(i)} = (s_{1}^{(i)}, s_{2}^{(i)}, \\ldots, s_{N_{i}}^{(i)})\\) with action labels \\(A^{(i)} = (a_{1}^{(i)}, a_{2}^{(i)}, \\ldots, a_{N_{i}}^{(i)})\\), a naive approach would be to randomly sample segments from videos to concatenate. However, this ignores the temporal order of actions within a task and more importantly can create unrealistic sequences, e.g., \"scooping jelly from jar\" followed by \"grinding coffee beans\" (this is unrealistic since it is common sense to spread the jelly right away after scooping it from a jar)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To create plausible multi-task sequences, we: i) preserve the original temporal order of action segments within each task, ii) preserve common sense action ordering and switches among tasks. Specifically, the sequence of action segments \\( s_j^{(i)} \\) from task \\( \\tau_i \\) follows the same order as in the original single-task video. At the end of an action segment, we leverage an LLM, which encodes extensive contextual and commonsense knowledge from real-world data, to decide whether to continue the current task or switch to an-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19122" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.401, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.209, + 0.319, + 0.221 + ], + "angle": 0, + "content": "(a) Multi-task Sequence Blending" + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.089, + 0.904, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.209, + 0.756, + 0.221 + ], + "angle": 0, + "content": "(b) Dynamic Isolation of Video Elements" + }, + { + "type": "image_caption", + "bbox": [ + 0.212, + 0.233, + 0.782, + 0.247 + ], + "angle": 0, + "content": "Figure 3. Illustration of leveraging LLM in our proposed framework for MSB (a) and DIVE (b)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.273, + 0.483, + 0.435 + ], + "angle": 0, + "content": "other task (see Fig. 3(a)). The LLM assesses the plausibility of continuing or switching based on the current action \\( a_{j}^{(i)} \\), the next action of the current task \\( a_{j+1}^{(i)} \\), and the next action of another task \\( a_{m}^{(k)} \\), where \\( \\tau_{k} \\) is a different task and \\( a_{m}^{(k)} \\) is its next pending action. For example, after \"turning on the kettle to boil water\" in a tea-making task, it is plausible to switch to a new task during the waiting time. Conversely, after \"scooping jelly from a jar\" in a pinwheel-making task, it makes sense to continue with the next action \"spread jelly on tortilla\" rather than switching tasks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.436, + 0.484, + 0.542 + ], + "angle": 0, + "content": "During training, we randomly sample potential task switch points at the end of action segments and use the LLM to decide whether to continue or switch tasks. For more than two tasks, we first randomly select one unfinished task as the candidate for switching. We then blend the sequence of action segments based on the LLM's decisions, resulting in a new multi-task sequence of actions and frames." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.55, + 0.4, + 0.566 + ], + "angle": 0, + "content": "3.4. Segment Boundary Learning (SBL)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.571, + 0.483, + 0.721 + ], + "angle": 0, + "content": "The multi-task videos synthesized by MSB exhibit abrupt transitions at boundaries where we switch tasks, which disrupt the temporal flow and negatively impact the model's performance. Therefore, we propose the SBL module that learns to recover the feature representation of a frame \\( f_{t} \\) using its nearby frames, excluding frames immediately adjacent to \\( t \\) to avoid trivial learning. This approach allows us to generate features for frames where we concatenate segments from different videos, thus enhancing the temporal smoothness between action segments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.723, + 0.484, + 0.768 + ], + "angle": 0, + "content": "Following prior works in TAS [17, 45, 49, 84], we first extract the framewise feature representation \\( f_{t} \\) using an I3D feature extractor [7] applied to a sliding window \\( V_{t} \\)," + }, + { + "type": "equation", + "bbox": [ + 0.237, + 0.777, + 0.483, + 0.793 + ], + "angle": 0, + "content": "\\[\nf _ {t} = \\mathrm {I 3 D} \\left(V _ {t}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.802, + 0.484, + 0.877 + ], + "angle": 0, + "content": "Then, for a given frame \\(t\\), we consider its neighboring frames from the intervals \\([t - \\delta - w, t - \\delta]\\) and \\([t + \\delta, t + \\delta + w]\\), where \\(w\\) defines the window size and \\(\\delta\\) specifies the exclusion margin around \\(t\\). We train a reconstruction function \\(\\mathcal{F}_{\\mathrm{SBL}}\\) to predict the feature \\(f_{t}\\) from its surrounding frames," + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.884, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\bar {f} _ {t} = \\mathcal {F} _ {\\mathrm {S B L}} \\left(f _ {t - \\delta - w: t - \\delta}, f _ {t + \\delta : t + \\delta + w}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.274, + 0.907, + 0.335 + ], + "angle": 0, + "content": "where \\( f_{t - \\delta - w:t - \\delta} \\) and \\( f_{t + \\delta: t + \\delta + w} \\) denote the features of frames before and after \\( t \\) within the specified intervals. We optimize the reconstruction function by minimizing the reconstruction loss over non-boundary frames," + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.343, + 0.905, + 0.377 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {S B L}} = \\sum_ {t \\notin \\mathcal {B}} \\left\\| \\bar {f} _ {t} - f _ {t} \\right\\| ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.384, + 0.907, + 0.505 + ], + "angle": 0, + "content": "where \\(\\mathcal{B}\\) is the set of boundary frames (which come from two different videos, therefore are not continuous or realistic and we should not learn from them). By learning this reconstruction function, the model can generate smooth feature representations at segment boundaries. During training, we replace the original features at boundary frames with the reconstructed features \\(\\bar{f}_t\\) to enhance the temporal smoothness of the synthesized multi-task videos." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.513, + 0.895, + 0.53 + ], + "angle": 0, + "content": "3.5. Dynamic Isolation of Video Elements (DIVE)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.536, + 0.907, + 0.763 + ], + "angle": 0, + "content": "In real multi-task videos, the present objects and their layouts can significantly differ from those in single-task videos. For example, as shown in the top-right frame in Fig. 1(b), in a video of making both tea and pinwheels, we might add honey to a mug while a rolled tortilla is on the cutting board, which is typically unseen in single-task videos. However, the output of MSB, which concatenates different actions from different single-task videos, does not capture such realistic multi-task layouts. Observing that action-relevant objects remain consistent across single-task and multi-task videos, we propose a Dynamic Isolation of Video Elements approach to isolate video elements into foreground and background components, and develop a Foreground-Background Feature Composition module to synthesize more realistic multi-task videos for training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Detecting Action-Relevant Objects. Given a frame, it is challenging to distinguish between foreground and background solely based on visual cues, especially in cluttered multi-task environments. The definition of \"foreground\" or \"relevant objects\" depends on the action being performed; thus, knowledge of the action context is essential. Fig. 3(b) illustrates how we obtain the foreground and background components. We first query an LLM to obtain a list of relevant objects for each action class \\(a \\in \\mathcal{C}\\). For example," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19123" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.244 + ], + "angle": 0, + "content": "the action \"transfer water to kettle\" would be associated with \"kettle\" and \"measuring cup\". Then, we use an open-vocabulary object detector, GroundingDINO [46], to detect these relevant objects along with hands (we add \"left hand\" and \"right hand\" as object names) within each frame. We define the foreground region as the union of detected bounding boxes. We generate foreground and background frames through Gaussian blurring [82] outside and inside this region, respectively. We empirically observed that Gaussian blurring performs slightly better than simply zero masking." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.246, + 0.484, + 0.292 + ], + "angle": 0, + "content": "Foreground-Background Feature Composition (FBFC). After generating foreground and background frames, we extract I3D features," + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.298, + 0.483, + 0.318 + ], + "angle": 0, + "content": "\\[\nf _ {t} ^ {\\mathrm {f g}} = \\mathrm {I 3 D} (V _ {t} \\odot M _ {t} ^ {\\mathrm {f g}}), f _ {t} ^ {\\mathrm {b g}} = \\mathrm {I 3 D} (V _ {t} \\odot M _ {t} ^ {\\mathrm {b g}}), \\qquad (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.484, + 0.404 + ], + "angle": 0, + "content": "where \\( f_{t}^{\\mathrm{fg}} \\) and \\( f_{t}^{\\mathrm{bg}} \\) are foreground and background frame features, respectively, and \\( \\odot \\) is the Gaussian blurring operation with corresponding foreground masks \\( M_{t}^{\\mathrm{fg}} \\) and background masks \\( M_{t}^{\\mathrm{fg}} \\). We train a decoder \\( \\mathcal{D} \\) to reconstruct the original feature from foreground and background features," + }, + { + "type": "equation", + "bbox": [ + 0.227, + 0.411, + 0.483, + 0.431 + ], + "angle": 0, + "content": "\\[\n\\hat {f} _ {t} = \\mathcal {D} \\left(f _ {t} ^ {\\mathrm {f g}}, f _ {t} ^ {\\mathrm {b g}}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.437, + 0.484, + 0.468 + ], + "angle": 0, + "content": "Let \\( f_{t} \\) denote the original feature in Eq. (1). We train the decoder by minimizing the composition loss," + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.477, + 0.483, + 0.513 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {F B F C}} = \\sum_ {t} \\left\\| \\hat {f} _ {t} - f _ {t} \\right\\| ^ {2}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.519, + 0.484, + 0.563 + ], + "angle": 0, + "content": "To augment synthetic multi-task features, we mix the background features of the current frame with background features of frames from another task" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.572, + 0.483, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\tilde {f} _ {t} ^ {\\mathrm {b g}} = \\beta f _ {t} ^ {\\mathrm {b g}} + (1 - \\beta) f _ {t ^ {\\prime}} ^ {\\mathrm {b g}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.603, + 0.484, + 0.681 + ], + "angle": 0, + "content": "where \\( f_{t}^{\\mathrm{bg}} \\) is the background feature from the current frame, \\( f_{t'}^{\\mathrm{bg}} \\) is the background feature from another task, and \\( \\beta \\in [0,1] \\) controls the mixing ratio. We then compose multi-task features from the foreground feature and the mixed background feature using the decoder" + }, + { + "type": "equation", + "bbox": [ + 0.225, + 0.689, + 0.483, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\tilde {f} _ {t} = \\mathcal {D} \\left(f _ {t} ^ {\\mathrm {f g}}, \\hat {f} _ {t} ^ {\\mathrm {b g}}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.484, + 0.75 + ], + "angle": 0, + "content": "We replace the training data with these recomposed features, exposing the model to diverse background variations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.452, + 0.775 + ], + "angle": 0, + "content": "3.6. Two-Stage Temporal Action Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.871 + ], + "angle": 0, + "content": "To further enhance the robustness of our model to background variations during testing, we propose a two-stage action segmentation method incorporating a Foreground-Aware Action Refinement (FAAR) module. By focusing on foreground regions, FAAR mitigates the impact of background distractions and refines the initial TAS predictions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "During training, with ground-truth actions available, we obtain the action-relevant objects and detect the foreground" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "regions. We extract CLIP features [55] from the foreground frames and apply a lightweight TAS model to predict foreground-aware action probabilities based solely on the foreground frames, \\( p_t^{\\mathrm{fg}} \\), which are combined with the initial action predictions \\( p_t \\) output by the base TAS model," + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.178, + 0.907, + 0.198 + ], + "angle": 0, + "content": "\\[\np _ {t} ^ {\\mathrm {f g}} = \\mathcal {R} \\left(g \\left(v _ {t} \\odot m _ {t} ^ {\\mathrm {f g}}\\right)\\right), \\quad p _ {t} ^ {\\text {f i n a l}} = (1 - \\alpha) \\cdot p _ {t} + \\alpha \\cdot p _ {t} ^ {\\mathrm {f g}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.204, + 0.905, + 0.265 + ], + "angle": 0, + "content": "where \\(m_t^{\\mathrm{fg}}\\) is the binary foreground mask, \\(g\\) denotes CLIP feature extractor, \\(\\mathcal{R}\\) denotes the lightweight TAS model, \\(\\alpha \\in [0,1]\\) controls the weight of foreground-aware probabilities, and \\(p_t^{\\mathrm{final}}\\) is the refined action probabilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.266, + 0.907, + 0.406 + ], + "angle": 0, + "content": "During inference, without ground-truth actions, we first select the top \\(K\\) predicted action classes \\(\\{\\hat{a}_t^{(1)},\\dots ,\\hat{a}_t^{(K)}\\}\\) based on the initial action probabilities \\(p_t\\) from the base TAS model, and retrieve their relevant objects. We then detect the corresponding foreground regions and obtain foreground-aware action probabilities \\(p_t^{\\mathrm{fg},(k)}\\) for each of the top \\(K\\) actions respectively. We compute the final prediction by combining the initial action probabilities with the foreground-aware ones," + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.418, + 0.905, + 0.459 + ], + "angle": 0, + "content": "\\[\np _ {t} ^ {\\text {f i n a l}} = (1 - \\alpha) \\cdot p _ {t} + \\alpha \\cdot \\sum_ {k = 1} ^ {K} \\theta_ {t} ^ {(k)} p _ {t} ^ {\\mathrm {f g}, (k)}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.46, + 0.907, + 0.553 + ], + "angle": 0, + "content": "where the weights \\(\\theta_t^{(k)}\\) are proportional to the initial probabilities of the top- \\(K\\) actions, i.e., \\(\\theta_t^{(k)} = \\frac{p_t\\left(\\hat{a}_t^{(k)}\\right)}{\\sum_{j=1}^K p_t\\left(\\hat{a}_t^{(j)}\\right)}\\). This approach refines the predictions by focusing on the most probable actions and their relevant objects, enhancing robustness to background variations in multi-task scenarios." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.563, + 0.727, + 0.579 + ], + "angle": 0, + "content": "3.7. Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.906, + 0.729 + ], + "angle": 0, + "content": "Training. We use the synthetic multi-task videos generated by MSB for training. We train the SBL module \\(\\mathcal{F}_{\\mathrm{SBL}}\\) via \\(\\mathcal{L}_{\\mathrm{SBL}}\\) in Eq. (3) and utilize the reconstructed features at segment boundaries to enhance temporal smoothness. Similarly, we train the FBFC decoder \\(\\mathcal{D}\\) via \\(\\mathcal{L}_{\\mathrm{FBFC}}\\) in Eq. (6), and augment the training data by mixing background features from different tasks as described in Eqs. (7) and (8). We train the base TAS model with FAAR using the standard action loss. The total loss function \\(\\mathcal{L}\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.743, + 0.905, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {a c t i o n}} + \\lambda_ {\\mathrm {S B L}} \\mathcal {L} _ {\\mathrm {S B L}} + \\lambda_ {\\mathrm {F B F C}} \\mathcal {L} _ {\\mathrm {F B F C}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.772, + 0.906, + 0.817 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{action}}\\) includes the cross-entropy loss and smoothing loss for TAS, applied to the final action probabilities output by the FAAR module. \\(\\lambda_{\\mathrm{SBL}}\\) and \\(\\lambda_{\\mathrm{FBFC}}\\) are the loss weights." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Inference. During inference, we do not use the MSB, SBL and FBFC modules, as they are designed to augment training. We first obtain initial action predictions from the base TAS model, and then apply the FAAR module, as specified in Eq. (10), to derive the final action predictions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19124" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.465, + 0.108 + ], + "angle": 0, + "content": "3.8. Incorporating Unlabeled Multi-Task Videos" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.219 + ], + "angle": 0, + "content": "Our proposed framework so far trains an MT-TAS model using only single-task videos. However, we additionally investigate the potential benefits of incorporating unlabeled multi-task videos, when such videos are available. Since annotating long multi-task videos is labor-intensive and time-consuming, leveraging unlabeled multi-task videos offers a practical way to enhance the model performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.219, + 0.484, + 0.4 + ], + "angle": 0, + "content": "We formulate this as a domain adaptation problem [9, 10, 50], where labeled single-task and synthetic multi-task videos constitute the source domain, while unlabeled real multi-task videos form the target domain. We aim to learn domain-invariant feature representations through adversarial training, achieved by adding a domain classifier with a Gradient Reversal Layer (GRL) [21] after the base TAS model's early feature layer. The model is trained jointly with action labels on the source domain and domain labels on both domains. During training, the GRL reverses the gradient flow, encouraging indistinguishable feature distributions across domains. Our overall training loss is" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.406, + 0.482, + 0.422 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {a c t i o n}} + \\lambda_ {\\mathrm {S B L}} \\mathcal {L} _ {\\mathrm {S B L}} + \\lambda_ {\\mathrm {F B F C}} \\mathcal {L} _ {\\mathrm {F B F C}} + \\lambda_ {\\text {d o m a i n}} \\mathcal {L} _ {\\text {d o m a i n}}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.426, + 0.483, + 0.457 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{domain}}\\) is the framewise binary domain classification loss, and \\(\\lambda_{\\mathrm{domain}}\\) is the loss weight." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.468, + 0.248, + 0.484 + ], + "angle": 0, + "content": "4. MEKA Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.484, + 0.675 + ], + "angle": 0, + "content": "While our proposed framework can be trained using only single-task videos, we require an MT-TAS benchmark for evaluation. To this end, we introduce the Multi-task Ego-centric Kitchen Activities (MEKA) dataset, which extends the EgoPER dataset [36] to multi-task scenarios. EgoPER is an egocentric procedural video dataset designed for temporal action segmentation and error detection. It contains single-task videos of users performing kitchen activities while wearing an egocentric camera (HoloLens 2). The dataset includes five cooking tasks: making pinwheels, quesadillas, oatmeal, coffee, and tea, each associated with a predefined task graph." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.809 + ], + "angle": 0, + "content": "To create realistic multi-task scenarios, we generated multi-task transcripts by interleaving action sequences from multiple tasks while ensuring logical coherence within the task graphs. Participants equipped with a HoloLens 2 camera followed these transcripts during recording. While we used similar recording equipment and kitchen tools as in the EgoPER dataset to maintain consistency, variations in the environments introduce additional challenges to the evaluation benchmark." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.9 + ], + "angle": 0, + "content": "After data collection, we annotated each video by assigning framewise labels for both actions and tasks. In total, MEKA consists of 100 videos for 12 hours of footage. In each video, participants perform two or three cooking tasks in an interleaved manner. See the supplementary materials for more details about the dataset. We will make the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.138 + ], + "angle": 0, + "content": "MEKA dataset publicly available upon publication, including videos and framewise annotations, to facilitate further research in multi-task temporal action segmentation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.155, + 0.649, + 0.172 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.181, + 0.707, + 0.198 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.204, + 0.907, + 0.43 + ], + "angle": 0, + "content": "Datasets. Since MT-TAS is a new task, existing datasets cannot be used for its evaluation. Consequently, we evaluate our approach using single-task videos from EgoPER [36] for training and our collected multi-task videos from MEKA for testing. EgoPER contains 213 normal single-task egocentric videos and MEKA contains 100 multi-task videos for the five recipes in EgoPER. In total, it has 51 classes of action steps. We were also able to adapt the EGTEA dataset [43] for multi-task evaluation. We manually select actions that appear exclusively in either the first or second half of each video, treating these segments as single-task videos, and consider the full videos with interleaved actions as multi-task videos. The supplementary materials provide additional details and the experimental results on the adapted EGTEA dataset." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.432, + 0.909, + 0.674 + ], + "angle": 0, + "content": "Model Architecture. We evaluate our method with three base TAS models: two offline models (MSTCN [17] and FACT [49]) and one online model (ProTAS [65]). Online TAS differs from offline TAS in that it only uses the frames up to the current moment for prediction. We implement the SBL module \\((\\mathcal{F}_{\\mathrm{SBL}})\\) by processing interval features through two convolutional layers, followed by temporal global average pooling. The Decoder \\(\\mathcal{D}\\) processes foreground and background features through parallel two-layer MLPs, each reducing feature dimension by half, followed by a linear layer that reconstructs the original feature representation from the concatenated features. We employ a three-layer TCN [17] with a hidden dimension of 64 for FAAR \\((\\mathcal{R})\\). For experiments with domain adaptation, we insert a GRL after the feature layer of the initial stage of TAS models, and use a two-layer MLP for binary domain classification." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.675, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. Following standard practice [17, 49, 65, 81, 84], we extract 2048-dimensional I3D features [7] pretrained on Kinetics [7] as video representations. We evaluate performance using standard TAS metrics: frame-wise accuracy (Acc), frame-wise accuracy excluding background frames (Acc-bg), segment-wise edit score (Edit), and segment-wise F1 scores at multiple IoU thresholds (F1@10,25,50). We adopt the default configurations from MSTCN [17], FACT [49], and ProTAS [65] using their official source codes. We train all models using Adam optimizer with a learning rate of 0.0005. We set the loss weights as \\(\\lambda_{\\mathrm{SBL}} = \\lambda_{\\mathrm{FBFC}} = \\lambda_{\\mathrm{domain}} = 1\\). For the SBL module, we define boundary frames as the 10 frames centered at the points where segments from different videos are concatenated, and set window size \\(w\\) as 5 and margin" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19125" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.089, + 0.895, + 0.215 + ], + "angle": 0, + "content": "
MSBSBLFBFCFAARTAS model: MSTCN [17]TAS model: FACT [49]
AccAcc-bgEditF1@{10,25,50}AccAcc-bgEditF1@{10,25,50}
baseline62.349.150.152.649.339.753.741.442.944.641.631.7
✓67.863.764.268.966.856.072.867.873.173.170.058.4
✓✓68.865.964.770.868.858.173.670.474.576.173.662.3
✓✓✓73.764.469.573.069.958.873.871.374.479.676.765.5
✓✓✓✓75.772.174.979.777.667.475.772.876.081.279.369.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.255, + 0.218, + 0.741, + 0.232 + ], + "angle": 0, + "content": "Table 1. Offline multi-task temporal action segmentation performance on MEKA." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.242, + 0.48, + 0.349 + ], + "angle": 0, + "content": "
MSBSBLFBFCFAARAccEditF1@{10,25,50}
baseline49.124.624.120.012.4
✓51.130.726.522.112.8
✓✓51.133.129.324.514.6
✓✓✓55.237.532.827.118.3
✓✓✓✓67.854.854.649.836.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.115, + 0.353, + 0.458, + 0.368 + ], + "angle": 0, + "content": "Table 2. Online temporal action segmentation on MEKA." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.386, + 0.483, + 0.568 + ], + "angle": 0, + "content": "\\(\\delta\\) as 2. For Foreground-Background Feature Composition (FBFC), we select \\(10\\%\\) of frames and augment them with Eq. (7) by randomly sampling \\(\\beta\\) from [0.5, 1] during training. See Sec. 5.2 for the qualitative analysis on the value of \\(\\beta\\). For the FAAR module, we employ a two-stage training process. In the first stage, we train the model without FAAR by setting \\(\\alpha = 0\\) in Eq. (9). In the second stage, we incorporate FAAR and randomly sample \\(\\alpha\\) from [0, 1] during training. For inference, we set \\(K = 3\\) and \\(\\alpha = 0.3\\) in Eq. (10). We include ablation studies on the values of \\(K\\) and \\(\\alpha\\) in Sec. 5.2. Please refer to the supplementary materials for more implementation details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.577, + 0.295, + 0.593 + ], + "angle": 0, + "content": "5.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.599, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Offline Temporal Action Segmentation. Tab. 1 presents the performance of our approach on the MEKA dataset using two baseline offline TAS models: MSTCN [17] and FACT [49]. We start with the baseline models trained on single-task videos without any of our proposed modules (first row). We then progressively incorporate the Multi-task Sequence Blending (MSB), Segment Boundary Learning (SBL), Foreground-Background Feature Composition (FBFC), and Foreground-Aware Action Refinement (FAAR) modules to assess their individual contributions. In the first row, MSTCN achieves an accuracy of \\(62.3\\%\\) and an F1@50 score of \\(39.7\\%\\), while FACT attains \\(53.7\\%\\) accuracy and \\(31.7\\%\\) F1@50, indicating that the baseline models struggle to generalize to multi-task videos when trained solely on single-task data. Incorporating MSB significantly improves performance for both models, especially for FACT, due to its explicit modeling of video transcripts during training. Adding SBL further boosts performance. With MSB and SBL, MSTCN's accuracy improves by \\(6.5\\%\\) and F1@50 by \\(18.4\\%\\), while FACT's accuracy" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.242, + 0.905, + 0.408 + ], + "angle": 0, + "content": "
Training SetupAccEditF1@{10,25,50}
MSTCN [17]Offline Temporal Action Segmentation
LMT78.075.978.376.670.4
LST&LMT80.684.784.582.574.6
Oursw/o DA75.774.979.777.667.4
w/ DA76.6+0.977.6+2.782.0+2.380.6+3.072.2+4.8
+LMT82.787.389.788.582.4
ProTAS [65]Online Temporal Action Segmentation
LMT61.541.137.232.822.0
LST&LMT64.549.839.434.522.7
Oursw/o DA67.854.854.649.836.7
w/ DA69.1+1.356.9+2.157.5+2.952.0+2.238.8+2.1
+LMT74.970.667.262.549.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.413, + 0.905, + 0.439 + ], + "angle": 0, + "content": "Table 3. The effect of domain adaptation using unlabeled multi-task videos and comparison with methods using labeled multi-task videos." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.461, + 0.907, + 0.567 + ], + "angle": 0, + "content": "improves by \\(19.9\\%\\) and F1@50 by \\(30.6\\%\\). Incorporating FBFC further improves the accuracy by \\(4.9\\%\\) on MSTCN and the F1@50 by \\(3.2\\%\\) on FACT. Finally, integrating all proposed modules leads to the highest performance, resulting in an F1@50 score of \\(67.4\\%\\) for MSTCN and \\(69.9\\%\\) for FACT. Overall, our approach effectively enhances the capability of standard TAS models to handle multi-task videos." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.568, + 0.909, + 0.793 + ], + "angle": 0, + "content": "Online Temporal Action Segmentation. We evaluate the online TAS performance of our approach in Tab. 2, using ProTAS [65] as the base TAS model. Without access to future frames, the online model relies solely on past and current frames, making it challenging to anticipate action transitions and accurately segment boundaries. This limitation results in a noticeable performance drop compared to the offline models, echoing the difficulties of online MT-TAS. However, our proposed modules still achieve a notable performance boost. When all modules are combined, the model increases accuracy by \\(18.7\\%\\) and F1@50 by \\(24.3\\%\\) compared to the baseline. Particularly, the integration of FAAR improves the accuracy by \\(12.6\\%\\) and all F1 scores by nearly \\(20\\%\\), highlighting the value of foreground-aware refinements in scenarios without access to future frames." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.796, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Domain Adaptation with Unlabeled Multi-Task Videos. In Tab. 3, we evaluate our domain adaptation approach using unlabeled multi-task videos, as introduced in Sec. 3.8. Domain adaptation consistently improves performance across all metrics, increasing F1@50 by \\(4.8\\%\\) and \\(2.1\\%\\) in offline and online settings, respectively. To thoroughly assess our method, we compare against two baseline se" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19126" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.116, + 0.092, + 0.37, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.092, + 0.623, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.092, + 0.88, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.184, + 0.253, + 0.812, + 0.267 + ], + "angle": 0, + "content": "Figure 4. Ablation Studies on FAAR: (a) comparison of different inputs; (b) effects of \\( K \\); (c) effects of \\( \\alpha \\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.293, + 0.486, + 0.581 + ], + "angle": 0, + "content": "tups using 3-fold cross-validation on multi-task videos: 1) the model is trained on labeled multi-task (LMT) videos; 2) the model is trained on both labeled single-task and labeled multi-task (LST&LMT) videos. Note that these baselines use labeled multi-task data, while our method employs only unlabeled multi-task videos during training. In offline TAS, our model with domain adaptation outperforms the LMT baseline and achieves comparable results to the LST&LMT baseline. In the more challenging online setting, our models surpass both baselines, demonstrating the ability of our modules to enhance generalization and handle the data-intensive demands of online TAS. We also extend our method by incorporating labeled multi-task videos (+LMT) to provide an upper bound for domain adaptation. The performance gap reveals room for improvement in unsupervised domain adaptation. Furthermore, the performance gains from LST&LMT to Ours+LMT demonstrate that our modules provide complementary benefits even when labeled multi-task data is available." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.487, + 0.884 + ], + "angle": 0, + "content": "Ablation Studies on FAAR. In Fig. 4, we conduct ablation studies to evaluate different inputs of FAAR and analyze the effects of \\( K \\) and \\( \\alpha \\) in Eq. (10) during inference. First, we change the module's input from foreground frames to background frames or full images. Background frames perform poorly while foreground frames outperform full images, showing the importance of action-relevant foreground information. We then examine how varying \\( K \\), the number of top predicted action classes considered, and \\( \\alpha \\), the weighting factor between initial predictions and foreground-aware predictions, influences the performance. The model's performance gradually improves with increasing \\( K \\), reaching a plateau when \\( K \\geq 3 \\). As \\( \\alpha \\) increases from 0 to 1, the model performs best between 0.3 and 0.4. Notably, while the accuracies at \\( \\alpha = 0 \\) (base TAS predictions only) and \\( \\alpha = 1 \\) (foreground-aware predictions only) are similar, the Edit and F1@50 scores for foreground-only predictions are significantly lower. This suggests that FAAR emphasizes individual frame accuracy but may lack the temporal consistency needed for segment-wise metrics." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Qualitative Analysis of FBFC. In Fig. 5, we qualitatively" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.291, + 0.895, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.442, + 0.907, + 0.468 + ], + "angle": 0, + "content": "Figure 5. Retrieved nearest multi-task videos by different ratios with FBFC. For simplicity, only middle frames of video clips are displayed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.495, + 0.907, + 0.737 + ], + "angle": 0, + "content": "analyze how FBFC works with different values of the mixing ratio \\(\\beta\\). We sample two clips from single-task videos, extract foreground frames from one clip and background frames from another, and then recompose these features using FBFC via Eqs. (7) and (8). To assess the decomposition quality, we retrieve the nearest clip from multi-task testing videos based on feature similarity. With a lower mixing ratio (\\(\\beta = 0.2\\)), the recomposed features are more heavily influenced by background features, and the context of the retrieved frames predominantly match the background frames. Conversely, with a higher mixing ratio (\\(\\beta = 0.8\\)), the retrieved frames often depict the same action as the foreground frames while sharing some similarities with the context of the background frames. Based on these observations, we sample \\(\\beta \\in [0.5,1]\\) in training to ensure preserving action features while introducing background variations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.738, + 0.907, + 0.768 + ], + "angle": 0, + "content": "See supplementary materials for more results, qualitative visualization, and discussion on complexity and limitations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.785, + 0.642, + 0.8 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We proposed an MT-TAS framework to tackle the challenges of interleaved actions in multi-task scenarios. Our approach combines synthesized multi-task data, a two-stage TAS model, and the new MEKA benchmark dataset. We validated the effectiveness of our approach through extensive experiments in both offline and online MT-TAS." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19127" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.177 + ], + "angle": 0, + "content": "This work is sponsored by ARPA-H (1AY2AX000062), DARPA PTG (HR00112220001), NSF (IIS-2115110) and ARO (W911NF2110276). Content does not necessarily reflect the position/policy of the Government." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.189, + 0.188, + 0.205 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.214, + 0.484, + 0.27 + ], + "angle": 0, + "content": "[1] Hyemin Ahn and Dongheui Lee. Refining action segmentation with hierarchical video representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16302-16310, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.271, + 0.484, + 0.325 + ], + "angle": 0, + "content": "[2] J. B. Alayrac, P. Bojanowski, N. Agrawal, J. Sivic, I. Laptev, and S. Lacoste-Julien. Unsupervised learning from narrated instruction videos. IEEE Conference on Computer Vision and Pattern Recognition, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.327, + 0.484, + 0.382 + ], + "angle": 0, + "content": "[3] Kyungho Bae, Geo Ahn, Youngrae Kim, and Jinwoo Choi. Devias: Learning disentangled video representations of action and scene for holistic video understanding. In European Conference on Computer Vision, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.383, + 0.484, + 0.437 + ], + "angle": 0, + "content": "[4] Emad Bahrami, Gianpiero Francesca, and Juergen Gall. How much temporal long-term context is needed for action segmentation? In International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.484, + 0.493 + ], + "angle": 0, + "content": "[5] Siddhant Bansal, Chetan Arora, and C.V. Jawahar. My view is the best view: Procedure learning from egocentric videos. In European Conference on Computer Vision (ECCV), 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.496, + 0.484, + 0.55 + ], + "angle": 0, + "content": "[6] Nadine Behrmann, S. Alireza Golestaneh, Zico Kolter, Juergen Gall, and Mehdi Noroozi. Unified fully and timestamp supervised temporal action segmentation via sequence to sequence translation. In ECCV, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.552, + 0.484, + 0.593 + ], + "angle": 0, + "content": "[7] J. Carreira and A. Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In IEEE Conference on Computer Vision and Pattern Recognition, 2017. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.595, + 0.484, + 0.663 + ], + "angle": 0, + "content": "[8] Chien-Yi Chang, De-An Huang, Yanan Sui, Li Fei-Fei, and Juan Carlos Niebles. D3tw: Discriminative differentiable dynamic time warping for weakly supervised action alignment and segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.665, + 0.484, + 0.734 + ], + "angle": 0, + "content": "[9] Min-Hung Chen, Baopu Li, Yingze Bao, Ghassan Al-Regib, and Zsolt Kira. Action segmentation with joint self-supervised temporal domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9454–9463, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.735, + 0.483, + 0.817 + ], + "angle": 0, + "content": "[10] Victor G. Turrisi da Costa, Giacomo Zara, Paolo Rota, Thiago Oliveira-Santos, Nicu Sebe, Vittorio Murino, and Elisa Ricci. Dual-head contrastive domain adaptation for video action recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1181-1190, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[11] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision (IJCV), 130:33-55, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.214, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[12] Li Ding and Chenliang Xu. Weakly-supervised action segmentation with iterative soft boundary assignment. IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.906, + 0.233 + ], + "angle": 0, + "content": "[13] Shuangrui Ding, Maomao Li, Tianyu Yang, Rui Qian, Haohang Xu, Qingyi Chen, Jue Wang, and Hongkai Xiong. Motion-aware contrastive video representation learning via foreground-background merging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9716-9726, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[14] G. Donahue and E. Elhamifar. Learning to predict activity progress by self-supervised video alignment. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[15] E. Elhamifar and D. Huynh. Self-supervised multi-task procedure learning from instructional videos. European Conference on Computer Vision, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.362 + ], + "angle": 0, + "content": "[16] E. Elhamifar and Z. Naing. Unsupervised procedure learning via joint dynamic summarization. International Conference on Computer Vision, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[17] Yazan Abu Farha and Jurgen Gall. Ms-tcn: Multi-stage temporal convolutional network for action segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3575-3584, 2019. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[18] Alireza Fathi, Xiaofeng Ren, and James M Rehg. Learning to recognize objects in egocentric activities. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[19] Mohsen Fayyaz and Jurgen Gall. Sct: Set constrained temporal transformer for set supervised action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.906, + 0.616 + ], + "angle": 0, + "content": "[20] Daniel Fried, Jean-Baptiste Alayrac, Phil Blunsom, Chris Dyer, Stephen Clark, and Aida Nematzadeh. Learning to segment actions from observation and narration. Annual Meeting of the Association for Computational Linguistics, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.906, + 0.659 + ], + "angle": 0, + "content": "[21] Y. Ganin and V. Lempitsky. Unsupervised domain adaptation by backpropagation. International Conference on Machine Learning, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.906, + 0.731 + ], + "angle": 0, + "content": "[22] Reza Ghoddoosian, Isht Dwivedi, Nakul Agarwal, Chiho Choi, and Behzad Dariush. Weakly-supervised online action segmentation in multi-view instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13780-13790, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.906, + 0.788 + ], + "angle": 0, + "content": "[23] Reza Ghoddosian, Saif Sayed, and Vassilis Athitsos. Hierarchical modeling for task recognition and action segmentation in weakly-labeled instructional videos. IEEE Winter Conference on Applications of Computer Vision, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.906, + 0.858 + ], + "angle": 0, + "content": "[24] Reza Ghoddosian, Isht Dwivedi, Nakul Agarwal, and Behzad Dariush. Weakly-supervised action segmentation and unseen error detection in anomalous instructional videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10128-10138, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[25] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Mar" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19128" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.483, + 0.425 + ], + "angle": 0, + "content": "tin, Tushar Nagarajan, Ilija Radosavovic, Santhosh K. Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Z. Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Christian Fuegen, Abraham Gebreselasie, Cristina Gonzalez, James M. Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolár, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran K. Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Yunyi Zhu, Pablo Arbeláez, David J. Crandall, Dima Damen, Giovanni Maria Farinella, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard A. Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the world in 3,000 hours of egocentric video. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18973-18990, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.483, + 0.525 + ], + "angle": 0, + "content": "[26] Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, et al. Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19383-19400, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.527, + 0.482, + 0.582 + ], + "angle": 0, + "content": "[27] Yifei Huang, Yusuke Sugano, and Yoichi Sato. Improving action segmentation via graph-based temporal reasoning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.585, + 0.482, + 0.682 + ], + "angle": 0, + "content": "[28] Yifei Huang, Guo Chen, Jilan Xu, Mingfang Zhang, Lijin Yang, Baoqi Pei, Hongjie Zhang, Lu Dong, Yali Wang, Limin Wang, and Yu Qiao. Egoexolearn: A dataset for bridging asynchronous ego- and exo-centric view of procedural activities in real world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22072-22086, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.685, + 0.482, + 0.754 + ], + "angle": 0, + "content": "[29] Yuchi Ishikawa, Seito Kasai, Yoshimitsu Aoki, and Hirokatsu Kataoka. Alleviating over-segmentation errors by detecting action boundaries. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2322-2331, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.757, + 0.482, + 0.813 + ], + "angle": 0, + "content": "[30] Hilde Kuehne, Ali Arslan, and Thomas Serre. The language of actions: Recovering the syntax and semantics of goal-directed human activities. IEEE Conference on Computer Vision and Pattern Recognition, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.815, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[31] H. Kuehne, A. Arslan, and T. Serre. The language of actions: Recovering the syntax and semantics of goal-directed human. IEEE Conference on Computer Vision and Pattern Recognition, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[32] H. Kuehne, J. Gall, and T. Serre. An end-to-end generative framework for video segmentation and recognition." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "IEEE Winter Conference on Applications of Computer Vision, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[33] Anna Kukleva, Hilde Kuehne, Fadime Sener, and Jurgen Gall. Unsupervised learning of action classes with continuous temporal embedding. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[34] Sateesh Kumar, Sanjay Haresh, Awais Ahmed, Andrey Konin, M. Zeeshan Zia, and Quoc-Huy Tran. Unsupervised action segmentation by joint representation learning and online clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20174-20185, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[35] C. Lea, M. D. Flynn, R. Vidal, A. Reiter, and G. D. Hager. Temporal convolutional networks for action segmentation and detection. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[36] S. Lee, Z. Lu, Z. Zhang, M. Hoai, and E. Elhamifar. Error detection in egocentric procedural task videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[37] Haoxin Li, Yuan Liu, Hanwang Zhang, and Boyang Li. Mitigating and evaluating static bias of action representations in the background and the foreground. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19911-19923, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[38] Jun Li and Sinisa Todorovic. Set-constrained viterbi for set-supervised action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[39] J. Li and S. Todorovic. Anchor-constrained viterbi for set-supervised action segmentation. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.575 + ], + "angle": 0, + "content": "[40] J. Li, P. Lei, and S. Todorovic. Weakly supervised energy-based learning for action segmentation. International Conference on Computer Vision, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.643 + ], + "angle": 0, + "content": "[41] M. Li, L. Chen, Y. Duarr, Z. Hu, J. Feng, J. Zhou, and J. Lu. Bridge-prompt: Towards ordinal action understanding in instructional videos. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[42] Shi-Jie Li, Yazan AbuFarha, Yun Liu, Ming-Ming Cheng, and Juergen Gall. Ms-tcn++: Multi-stage temporal convolutional network for action segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[43] Yin Li, Miao Liu, and James M. Rehg. In the eye of beholder: Joint learning of gaze and actions in first person video. European Conference on Computer Vision, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[44] Zhe Li, Yazan Abu Farha, and Jurgen Gall. Temporal action segmentation from timestamp supervision. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[45] Daochang Liu, Qiyue Li, Anh-Dung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. International Conference on Computer Vision, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[46] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19129" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. *eccv*, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.121, + 0.483, + 0.175 + ], + "angle": 0, + "content": "[47] Z. Lu and E. Elhamifar. Weakly-supervised action segmentation and alignment via transcript-aware union-of-subspaces learning. International Conference on Computer Vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.483, + 0.232 + ], + "angle": 0, + "content": "[48] Z. Lu and E. Elhamifar. Set-supervised action learning in procedural task videos via pairwise order consistency. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.483, + 0.287 + ], + "angle": 0, + "content": "[49] Z. Lu and E. Elhamifar. Fact: Frame-action cross-attention temporal modeling for efficient action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.289, + 0.483, + 0.329 + ], + "angle": 0, + "content": "[50] J. Munro and D. Damen. Multi-modal domain adaptation for fine-grained action recognition. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.331, + 0.483, + 0.372 + ], + "angle": 0, + "content": "[51] Zhanzhong Pang, Fadime Sener, Shrinivas Ramasubramanian, and Angela Yao. Long-tail temporal action segmentation with group-wise temporal logit adjustment. 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.373, + 0.483, + 0.428 + ], + "angle": 0, + "content": "[52] Sam Perochon and Laurent Oudre. Unsupervised action segmentation of untrimmed egocentric videos. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.429, + 0.483, + 0.483 + ], + "angle": 0, + "content": "[53] Will Price, Carl Vondrick, and Dima Damen. Unweavenet: Unweaving activity stories. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13770-13779, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.485, + 0.483, + 0.566 + ], + "angle": 0, + "content": "[54] Camillo Quattrocchi, Antonino Furnari, Daniele Di Mauro, Mario Valerio Giuffrida, and Giovanni Maria Farinella. Synchronization is all you need: Exocentric-to-egocentric transfer for temporal action segmentation with unlabeled synchronized video pairs. European Conference on Computer Vision, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.568, + 0.483, + 0.65 + ], + "angle": 0, + "content": "[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. International Conference on Machine learning, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.483, + 0.719 + ], + "angle": 0, + "content": "[56] Francesco Ragusa, Antonino Furnari, Salvatore Livatino, and Giovanni Maria Farinella. The meccano dataset: Understanding human-object interactions from egocentric videos in an industrial-like domain. Winter Conference on Applications of Computer Vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.721, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[57] Rahul Rahaman, Dipika Singhania, Alexandre Thiery, and Angela Yao. A generalized and robust framework for timestamp supervision in temporal action segmentation. In Computer Vision-ECCV 2022: 17th European Conference, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[58] A. Richard, H. Kuehne, and J. Gall. Action sets: Weakly supervised action segmentation without ordering constraints. IEEE Conference on Computer Vision and Pattern Recognition, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[59] A. Richard, H. Kuehne, A. Iqbal, and J. Gall. Neuralnetwork-viterbi: A framework for weakly supervised video learning. IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[60] M. Rohrbach, S. Amin, M. Andriluka, and B. Schiele. A database for fine grained activity detection of cooking activities. IEEE Conference on Computer Vision and Pattern Recognition, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.906, + 0.217 + ], + "angle": 0, + "content": "[61] Saquib Sarfraz, Naila Murray, Vivek Sharma, Ali Diba, Luc Van Gool, and Rainer Stiefelhagen. Temporally-weighted hierarchical clustering for unsupervised action segmentation. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.906, + 0.3 + ], + "angle": 0, + "content": "[62] Saif Sayed, Reza Ghoddoosian, Bhaskar Trivedi, and Vassilis Athitsos. A new dataset and approach for timestamp supervised action segmentation using human object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 3133-3142, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.906, + 0.385 + ], + "angle": 0, + "content": "[63] Fadime Sener, Dibyadip Chatterjee, Daniel Shelepov, Kun He, Dipika Singhania, Robert Wang, and Angela Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21096-21106, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.906, + 0.439 + ], + "angle": 0, + "content": "[64] Y. Shen and E. Elhamifar. Semi-weakly-supervised learning of complex actions from instructional task videos. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.906, + 0.495 + ], + "angle": 0, + "content": "[65] Y. Shen and E. Elhamifar. Progress-aware online action segmentation for egocentric procedural task videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.906, + 0.552 + ], + "angle": 0, + "content": "[66] Y. Shen, L. Wang, and E. Elhamifar. Learning to segment actions from visual and language instructions via differentiable weak sequence alignment. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.906, + 0.608 + ], + "angle": 0, + "content": "[67] Y. Shen, H. Wang, X. Yang, M. Feiszli, E. Elhamifar, L. Torresani, and E. Mavroudi. Learning to segment referred objects from narrated egocentric videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.906, + 0.662 + ], + "angle": 0, + "content": "[68] G. A. Sigurdsson, S. Divvala, A. Farhadi, and A. Gupta. Asynchronous temporal fields for action recognition. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.906, + 0.72 + ], + "angle": 0, + "content": "[69] B. Singh, T. K. Marks, M. Jones, O. Tuzel, and M. Shao. A multi-stream bi-directional recurrent neural network for finegrained action detection. IEEE Conference on Computer Vision and Pattern Recognition, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.906, + 0.775 + ], + "angle": 0, + "content": "[70] Dipika Singhania, Rahul Rahaman, and Angela Yao. C2f- tcn: A framework for semi-and fully-supervised temporal action segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):11484-11501, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.906, + 0.829 + ], + "angle": 0, + "content": "[71] Yaser Souri, Mohsen Fayyaz, Luca Minciullo, Gianpiero Francesca, and Juergen Gall. Fast Weakly Supervised Action Segmentation Using Mutual Consistency. PAMI, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[72] Sebastian Stein and Stephen J McKenna. Combining embedded accelerometers with computer vision for recognizing food preparation activities. In Proceedings of the 2013 ACM international joint conference on Pervasive and ubiquitous computing, pages 729-738, 2013. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19130" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[73] Yuhao Su and Ehsan Elhamifar. Two-stage active learning for efficient temporal action segmentation. In European Conference on Computer Vision, pages 161-183. Springer, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[74] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[75] Sanket Thakur, Cigdem Bayan, Pietro Morerio, Vittorio Murino, and Alessio Del Bue. Anticipating next active objects for egocentric videos. IEEE Access, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[76] Quoc-Huy Tran, Ahmed Mehmood, Muhammad Ahmed, Muhammad Naufil, Anas Zafar, Andrey Konin, and Zeeshan Zia. Permutation-aware activity segmentation via unsupervised frame-to-segment alignment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6426-6436, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[77] Huiyu Wang, Mitesh Kumar Singh, and Lorenzo Torresani. Ego-only: Egocentric action detection without exocentric transferring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5250-5261, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[78] Jiahao Wang, Guo Chen, Yifei Huang, Limin Wang, and Tong Lu. Memory-and-anticipation transformer for online action understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13824–13835, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[79] Zhe Wang, Hao Chen, Xinyu Li, Chunhui Liu, Yuanjun Xiong, Joseph Tighe, and Charless Fowlkes. Sscap: Self-supervised co-occurrence action parsing for unsupervised temporal action segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1819–1828, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[80] Ming Xu and Stephen Gould. Temporally consistent unbalanced optimal transport for unsupervised action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.617, + 0.482, + 0.686 + ], + "angle": 0, + "content": "[81] Ziwei Xu, Yogesh S Rawat, Yongkang Wong, Mohan Kankanhalli, and Mubarak Shah. Don't pour cereal into coffee: Differentiable temporal logic for temporal action segmentation. In Advances in Neural Information Processing Systems, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[82] Lingfeng Yang, Yueze Wang, Xiang Li, Xinlong Wang, and Jian Yang. Fine-grained visual prompting. Advances in Neural Information Processing Systems, 36, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[83] S. Yeung, O. Russakovsky, N. Jin, M. Andriluka, G. Mori, and L. Fei-Fei. Every moment counts: Dense detailed labeling of actions in complex videos. International Journal of Computer Vision, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[84] Fangqiu Yi, Hongyu Wen, and Tingting Jiang. Asformer: Transformer for action segmentation. In The British Machine Vision Conference (BMVC), 2021. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[85] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Helping hands: An object-aware ego-centric video recognition model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13901-13912, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[86] Ce Zhang, Changcheng Fu, Shijie Wang, Nakul Agarwal, Kwonjoon Lee, Chiho Choi, and Chen Sun. Object-centric video representation for long-term action anticipation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6751-6761, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[87] Xingyi Zhou, Anurag Arnab, Chen Sun, and Cordelia Schmid. How can objects help action recognition? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2362, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[88] D. Zhukov, J. B. Alayrac, R. G. Cinbis, D. Fouhey, I. Laptev, and J. Sivic. Cross-task weakly supervised learning from instructional videos. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19131" + } + ] +] \ No newline at end of file diff --git a/2025/Understanding Multi-Task Activities from Single-Task Videos/b633800a-1331-4862-80d6-5ccbeba54c12_origin.pdf b/2025/Understanding Multi-Task Activities from Single-Task Videos/b633800a-1331-4862-80d6-5ccbeba54c12_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..37ed01873759eeda527ac7ffd94322b6ff3f0026 --- /dev/null +++ b/2025/Understanding Multi-Task Activities from Single-Task Videos/b633800a-1331-4862-80d6-5ccbeba54c12_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da7e86b3c795e8f981f5cdf13ff6d1a6acc239a8c112100ca684478d59540c76 +size 2592999 diff --git a/2025/Understanding Multi-Task Activities from Single-Task Videos/full.md b/2025/Understanding Multi-Task Activities from Single-Task Videos/full.md new file mode 100644 index 0000000000000000000000000000000000000000..475e096ce97b57f21df6d93a8e3708a9c63ef51a --- /dev/null +++ b/2025/Understanding Multi-Task Activities from Single-Task Videos/full.md @@ -0,0 +1,366 @@ +# Understanding Multi-Task Activities from Single-Task Videos + +Yuhan Shen Northeastern University shen.yuh@northeastern.edu + +Ehsan Elhamifar Northeastern University e.elhamifar@northeastern.edu + +# Abstract + +We introduce and develop a framework for Multi-Task Temporal Action Segmentation (MT-TAS), a novel paradigm that addresses the challenges of interleaved actions when performing multiple tasks simultaneously. Traditional action segmentation models, trained on single-task videos, struggle to handle task switches and complex scenes inherent in multi-task scenarios. To overcome these challenges, our MT-TAS approach synthesizes multi-task video data from single-task sources using our Multi-task Sequence Blending and Segment Boundary Learning modules. Additionally, we propose to dynamically isolate foreground and background elements within video frames, addressing the intricacies of object layouts in multi-task scenarios and enabling a new two-stage temporal action segmentation framework with Foreground-Aware Action Refinement. Also, we introduce the Multi-task Egocentric Kitchen Activities (MEKA) dataset, containing 12 hours of egocentric multi-task videos, to rigorously benchmark MT-TAS models. Extensive experiments demonstrate that our framework effectively bridges the gap between single-task training and multi-task testing, advancing temporal action segmentation with state-of-the-art performance in complex environments. + +# 1. Introduction + +Consider a typical morning routine, when you prepare breakfast: you start making oatmeal and while waiting for it to cook, you brew coffee and pack your food for lunch. In fact, in our daily lives, we often perform multiple tasks simultaneously. For computer vision systems to accurately analyze such behaviors and provide useful assistance, they must comprehend these complex and interleaved activities. Temporal action segmentation (TAS) has emerged as a crucial technology for understanding human activities by partitioning videos into meaningful segments, each corresponding to a distinct action or step [17, 22, 24, 31, 33, 48, 49, 64, 73, 84, 88]. Recently, TAS has attracted growing interests + +![](images/876624b57e93088296560aa0e7567f0fb2210e50cd7c430c726ab5ffeb622451.jpg) + +![](images/554a9cc7c6b4d3d3b32430a73a9db0d07b0df68fbfd7fc36a6182a46ed326df7.jpg) + +![](images/9356039069186c8d6417b05ee8d78549904a653fe932c676beb3732d3ca8899c.jpg) +single-task videos +Figure 1. (a) Multi-task temporal action segmentation involves segmenting actions from interleaved tasks. It requires capturing interruptions and resumptions in the execution of each task. (b) Comparison between frames from single-task and multi-task videos. In multi-task videos, at each time instant, the scene often contains objects from tasks different from the currently executed one (highlighted by purple boxes). + +![](images/f43393b8d019cb9d2ebf1840e71f5897a05230b65306117e98d26f44ec2fd7bd.jpg) +multi-task videos +(b) Comparison between frames from single-task videos v.s. multi-task videos + +within egocentric videos [5, 36, 52, 54, 65, 77], where activities are captured from a first-person perspective, enabling fine-grained analysis of human actions, which is essential for applications such as real-time personalized assistance, embodied agents and efficient video content retrieval. + +In this paper, we study the new problem of Multi-Task Temporal Action Segmentation (MT-TAS), which explicitly focuses on scenarios where users perform multiple tasks by interleaving their steps within a single video, see Fig. 1. There are two major challenges that existing TAS models face in addressing MT-TAS. First, despite the notable progress through deep learning techniques [1, 6, 14, 17, 41, 42, 45, 49, 84], current TAS models have primarily focused on scenarios where each video contains actions related to a single high-level task, such as assembling a piece of furniture or cooking a specific recipe. This single-task focus limits the applicability of TAS models to the multi-task setting since i) they are biased to predict actions/steps from only one task at the inference time, ii) when performing several tasks, the environment and the scene often contains objects relevant to multiple tasks, which adds complexity + +to the context, while single-task videos often contain objects relevant to one task (see Fig. 1(b)), iii) interleaving steps of multiple tasks introduces complexities such as unpredictable task switches and disruptions to the temporal flows of actions within each task. + +Second, a significant obstacle in advancing MT-TAS is the lack of appropriate training data. Existing datasets [5, 16, 18, 31, 36, 63, 88] are predominantly collected for single-task scenarios, making them insufficient for capturing complexities of interleaved tasks. Furthermore, many possible step combinations in the multi-task setting makes it impractical to collect and annotate comprehensive real-world videos covering all potential situations (such videos would also be much longer than single-task videos). + +Paper Contributions. We study the new problem of MT-TAS. To overcome the aforementioned challenges, we propose an MT-TAS framework that uses only single-task videos for training. We generate synthetic multi-task data using single-task videos and propose a two-stage action segmentation model for accurate predictions. For the data generation, we first develop a Multi-task Sequence Blending (MSB) module that leverages Large Language Models (LLMs) to create multi-task transcripts of interleaved steps and their associated videos. Second, we develop a Segment Boundary Learning (SBL) module to mitigate the abrupt transitions at segment boundaries caused by the synthetic data. Third, we introduce a Dynamic Isolation of Video Elements (DIVE) approach to isolate foreground and background components within video frames by identifying action-relevant objects and combining foregrounds of actions in one task with backgrounds of other tasks to capture the more complex scenes of multi-task videos. By increasing the variation in the background components for the training data, our method becomes more robust to unseen object compositions in multi-task scenarios. + +For the action segmentation model, we develop a two-stage framework, where we refine the initial action predictions using a Foreground-Aware Action Refinement (FAAR) module. This approach leverages foreground information to enhance the focus of our model on action-relevant cues (e.g., focusing on the regions of kettle and measuring cup for the action "transfer water to kettle" as in Fig. 3(b)), improving accuracy and robustness to background variations during the multi-task testing. + +For proper evaluation of MT-TAS models, we gather and annotate the Multi-task Egocentric Kitchen Activities (MEKA) video dataset (we plan to publicly release the dataset), built upon the recent EgoPER [36] dataset. MEKA comprises 100 egocentric videos of users (about 12 hours of footage) performing multiple kitchen recipes from EgoPER in each run by interleaving their steps. This new benchmark captures a variety of interleaved kitchen activities, providing a rigorous testing ground for MT-TAS models. Through + +extensive experiments in both offline and online settings, we demonstrate that our proposed framework successfully bridges the gap between single-task training data and the complexities of real-world multi-task scenarios, advancing research in temporal action segmentation. + +# 2. Related Works + +Temporal Action Segmentation. TAS aims to understand the sequential structure of human activities in videos by assigning an action label to each frame, thereby partitioning the video into a sequence of action segments. Based on the level of supervision during training, TAS methods can be categorized into three main groups: unsupervised [2, 15, 16, 20, 33, 34, 61, 66, 76, 79, 80, 88], weakly-supervised [8, 12, 19, 22-24, 38-40, 44, 47, 48, 57-59, 62, 64, 71], and fully-supervised [1, 4, 6, 17, 27, 29, 32, 35, 36, 41, 42, 45, 49, 51, 60, 65, 68-71, 83, 84] approaches. Our work primarily aligns with the fully-supervised methods. However, regardless of the supervision level, previous studies have predominantly focused on single-task TAS, where test videos contain actions from only one high-level activity or task, while we focus on multi-task TAS. A closely related work, UnweaveNet [53], aims to predict the start, continuation, and resumption of multiple activity threads within a video. While UnweaveNet models multiple concurrent activities, its predictions are at the activity or task level and do not capture the fine-grained action steps within each task. In contrast, our approach focuses on recognizing and segmenting detailed action steps across interleaved tasks, providing a more granular understanding of multi-task activities. + +Procedural Video Datasets. To support research in TAS and related tasks, numerous procedural video datasets have been collected and released. These datasets include third-person videos (e.g., Breakfast [30], 50Salads [72], YouTube Instructional [2], COIN [74], CrossTask [88], ProceL [16], and ATA [24]), egocentric videos (e.g., GTEA [18], EGTEA [43], MECCANO [56], EgoProceL [5], HoloAssist [78], and EgoPER [36]), or both views such as Assembly101 [63], $(3 + 1)\mathrm{ReC}$ [62] and EgoExoLearn [28]. While large-scale video datasets like EPIC-Kitchens [11], Ego4D [25], and Ego-Exo4D [26] exist, they primarily focus on detecting and recognizing atomic action steps rather than modeling the procedural relationships between steps within a task, making them less suitable for TAS research. Also, most existing TAS datasets are limited to single-task scenarios, where each video contains actions related to one high-level activity. In contrast, our collected MEKA dataset specifically addresses multi-task temporal action segmentation by capturing videos where multiple tasks are performed in an interleaved manner, providing a more challenging and realistic benchmark for advancing TAS research. + +Object-Aware and Foreground/Background Learning. + +Learning to localize active objects or relevant objects is a prominent topic in video understanding [67, 75, 85, 86]. Furthermore, researchers have explored leveraging object information to improve video representation learning [85], action recognition [87], and action anticipation [86]. Some studies have also focused on disentangling foreground and background elements to enhance video representations [3, 13, 37]. Our work differs by specifically addressing the generalization of TAS models to multi-task videos. Unlike existing approaches that concentrate on short video clips or single-task videos, we propose a method that dynamically isolates foreground and background elements to handle unseen object compositions in multi-task scenarios, thereby mitigating interference from irrelevant objects. + +# 3. Multi-Task Temporal Action Segmentation + +# 3.1. Problem Setting + +Given a video sequence $\mathcal{V} = (v_{1},\ldots ,v_{T})$ of $T$ frames, the goal of TAS is to predict the sequence of action labels $A = (a_{1},\dots,a_{T})$ , where each label $a_{t}$ comes from a predefined set of action classes $\mathcal{C}$ . Assume we have $O$ tasks, where the set $\mathcal{C}_o$ consists of actions for the task $o\in \{1,\dots,O\}$ and $\mathcal{C} = \bigcup_{o = 1}^{O}\mathcal{C}_{o}$ contains actions of all tasks. In a single-task video, actions come from one $\mathcal{C}_o$ , while a multi-task video has interleaved actions from several tasks, hence, $a_{t}$ may belong to any $\mathcal{C}_o$ for $o\in \{1,\dots,O\}$ . Additionally, there may be background segments (e.g., 'answering phone') irrelevant to any of the tasks. A particular challenge in MT-TAS is to distinguish between action segments belonging to different tasks and background segments. + +In our setting, the training data consists of single-task videos with full framewise annotations. Our goal is to train a model on single-task videos so that it can effectively generalize to multi-task videos during testing. + +# 3.2. Overview of Proposed Framework + +Fig. 2 shows the overview of our MS-TAS framework. Our approach first employs Multi-task Sequence Blending (MSB) to generate synthetic multi-task videos by leveraging LLM-guided task transitions, eliminating the need for multi-task data collection. However, the initial synthesized multi-task videos contain abrupt transitions at segment boundaries, which disrupt the temporal flow, and contain discrepancies in object compositions between single-task and multi-task videos. To mitigate these issues, we introduce three components: a Segment Boundary Learning (SBL) module that enhances temporal smoothness between segments, a Dynamic Isolation of Video Elements approach that augments the background variations during training, and a two-stage TAS method with foreground-aware action refinement to improve the predictions of the model. Next, we discuss each component of our framework in details. + +![](images/06105a7da63057dc2cd121d50dfc8ccf433ef315bfdd074cf2dc27bb4a16576b.jpg) +Figure 2. Framework Overview. The modules marked with $\Theta$ are used exclusively for training, and modules marked with $\diamond$ are used for both training and testing. + +# 3.3. Multi-task Sequence Blending (MSB) + +The goal of MSB is to create initial multi-task video sequences from single-task videos, to simulate real-world multitasking behaviors. Given single-task training videos $\{\mathcal{V}^{(i)}\}_{i=1}^{M}$ , where each video $\mathcal{V}^{(i)} = (v_{1}^{(i)}, v_{2}^{(i)}, \ldots, v_{T_{i}}^{(i)})$ has a task label $\tau_{i}$ and consists of a sequence of action segments $S^{(i)} = (s_{1}^{(i)}, s_{2}^{(i)}, \ldots, s_{N_{i}}^{(i)})$ with action labels $A^{(i)} = (a_{1}^{(i)}, a_{2}^{(i)}, \ldots, a_{N_{i}}^{(i)})$ , a naive approach would be to randomly sample segments from videos to concatenate. However, this ignores the temporal order of actions within a task and more importantly can create unrealistic sequences, e.g., "scooping jelly from jar" followed by "grinding coffee beans" (this is unrealistic since it is common sense to spread the jelly right away after scooping it from a jar). + +To create plausible multi-task sequences, we: i) preserve the original temporal order of action segments within each task, ii) preserve common sense action ordering and switches among tasks. Specifically, the sequence of action segments $s_j^{(i)}$ from task $\tau_i$ follows the same order as in the original single-task video. At the end of an action segment, we leverage an LLM, which encodes extensive contextual and commonsense knowledge from real-world data, to decide whether to continue the current task or switch to an- + +![](images/ec6246c1a902940358b037def5032591c178a36212053ff045e9457870704a0a.jpg) +(a) Multi-task Sequence Blending + +![](images/05d873dcd92b3f451998b6be9d9e716d31161b854f1ac6792057d68a435fe8b5.jpg) +(b) Dynamic Isolation of Video Elements +Figure 3. Illustration of leveraging LLM in our proposed framework for MSB (a) and DIVE (b). + +other task (see Fig. 3(a)). The LLM assesses the plausibility of continuing or switching based on the current action $a_{j}^{(i)}$ , the next action of the current task $a_{j+1}^{(i)}$ , and the next action of another task $a_{m}^{(k)}$ , where $\tau_{k}$ is a different task and $a_{m}^{(k)}$ is its next pending action. For example, after "turning on the kettle to boil water" in a tea-making task, it is plausible to switch to a new task during the waiting time. Conversely, after "scooping jelly from a jar" in a pinwheel-making task, it makes sense to continue with the next action "spread jelly on tortilla" rather than switching tasks. + +During training, we randomly sample potential task switch points at the end of action segments and use the LLM to decide whether to continue or switch tasks. For more than two tasks, we first randomly select one unfinished task as the candidate for switching. We then blend the sequence of action segments based on the LLM's decisions, resulting in a new multi-task sequence of actions and frames. + +# 3.4. Segment Boundary Learning (SBL) + +The multi-task videos synthesized by MSB exhibit abrupt transitions at boundaries where we switch tasks, which disrupt the temporal flow and negatively impact the model's performance. Therefore, we propose the SBL module that learns to recover the feature representation of a frame $f_{t}$ using its nearby frames, excluding frames immediately adjacent to $t$ to avoid trivial learning. This approach allows us to generate features for frames where we concatenate segments from different videos, thus enhancing the temporal smoothness between action segments. + +Following prior works in TAS [17, 45, 49, 84], we first extract the framewise feature representation $f_{t}$ using an I3D feature extractor [7] applied to a sliding window $V_{t}$ , + +$$ +f _ {t} = \mathrm {I 3 D} \left(V _ {t}\right). \tag {1} +$$ + +Then, for a given frame $t$ , we consider its neighboring frames from the intervals $[t - \delta - w, t - \delta]$ and $[t + \delta, t + \delta + w]$ , where $w$ defines the window size and $\delta$ specifies the exclusion margin around $t$ . We train a reconstruction function $\mathcal{F}_{\mathrm{SBL}}$ to predict the feature $f_{t}$ from its surrounding frames, + +$$ +\bar {f} _ {t} = \mathcal {F} _ {\mathrm {S B L}} \left(f _ {t - \delta - w: t - \delta}, f _ {t + \delta : t + \delta + w}\right), \tag {2} +$$ + +where $f_{t - \delta - w:t - \delta}$ and $f_{t + \delta: t + \delta + w}$ denote the features of frames before and after $t$ within the specified intervals. We optimize the reconstruction function by minimizing the reconstruction loss over non-boundary frames, + +$$ +\mathcal {L} _ {\mathrm {S B L}} = \sum_ {t \notin \mathcal {B}} \left\| \bar {f} _ {t} - f _ {t} \right\| ^ {2}, \tag {3} +$$ + +where $\mathcal{B}$ is the set of boundary frames (which come from two different videos, therefore are not continuous or realistic and we should not learn from them). By learning this reconstruction function, the model can generate smooth feature representations at segment boundaries. During training, we replace the original features at boundary frames with the reconstructed features $\bar{f}_t$ to enhance the temporal smoothness of the synthesized multi-task videos. + +# 3.5. Dynamic Isolation of Video Elements (DIVE) + +In real multi-task videos, the present objects and their layouts can significantly differ from those in single-task videos. For example, as shown in the top-right frame in Fig. 1(b), in a video of making both tea and pinwheels, we might add honey to a mug while a rolled tortilla is on the cutting board, which is typically unseen in single-task videos. However, the output of MSB, which concatenates different actions from different single-task videos, does not capture such realistic multi-task layouts. Observing that action-relevant objects remain consistent across single-task and multi-task videos, we propose a Dynamic Isolation of Video Elements approach to isolate video elements into foreground and background components, and develop a Foreground-Background Feature Composition module to synthesize more realistic multi-task videos for training. + +Detecting Action-Relevant Objects. Given a frame, it is challenging to distinguish between foreground and background solely based on visual cues, especially in cluttered multi-task environments. The definition of "foreground" or "relevant objects" depends on the action being performed; thus, knowledge of the action context is essential. Fig. 3(b) illustrates how we obtain the foreground and background components. We first query an LLM to obtain a list of relevant objects for each action class $a \in \mathcal{C}$ . For example, + +the action "transfer water to kettle" would be associated with "kettle" and "measuring cup". Then, we use an open-vocabulary object detector, GroundingDINO [46], to detect these relevant objects along with hands (we add "left hand" and "right hand" as object names) within each frame. We define the foreground region as the union of detected bounding boxes. We generate foreground and background frames through Gaussian blurring [82] outside and inside this region, respectively. We empirically observed that Gaussian blurring performs slightly better than simply zero masking. + +Foreground-Background Feature Composition (FBFC). After generating foreground and background frames, we extract I3D features, + +$$ +f _ {t} ^ {\mathrm {f g}} = \mathrm {I 3 D} (V _ {t} \odot M _ {t} ^ {\mathrm {f g}}), f _ {t} ^ {\mathrm {b g}} = \mathrm {I 3 D} (V _ {t} \odot M _ {t} ^ {\mathrm {b g}}), \qquad (4) +$$ + +where $f_{t}^{\mathrm{fg}}$ and $f_{t}^{\mathrm{bg}}$ are foreground and background frame features, respectively, and $\odot$ is the Gaussian blurring operation with corresponding foreground masks $M_{t}^{\mathrm{fg}}$ and background masks $M_{t}^{\mathrm{fg}}$ . We train a decoder $\mathcal{D}$ to reconstruct the original feature from foreground and background features, + +$$ +\hat {f} _ {t} = \mathcal {D} \left(f _ {t} ^ {\mathrm {f g}}, f _ {t} ^ {\mathrm {b g}}\right). \tag {5} +$$ + +Let $f_{t}$ denote the original feature in Eq. (1). We train the decoder by minimizing the composition loss, + +$$ +\mathcal {L} _ {\mathrm {F B F C}} = \sum_ {t} \left\| \hat {f} _ {t} - f _ {t} \right\| ^ {2}. \tag {6} +$$ + +To augment synthetic multi-task features, we mix the background features of the current frame with background features of frames from another task + +$$ +\tilde {f} _ {t} ^ {\mathrm {b g}} = \beta f _ {t} ^ {\mathrm {b g}} + (1 - \beta) f _ {t ^ {\prime}} ^ {\mathrm {b g}}, \tag {7} +$$ + +where $f_{t}^{\mathrm{bg}}$ is the background feature from the current frame, $f_{t'}^{\mathrm{bg}}$ is the background feature from another task, and $\beta \in [0,1]$ controls the mixing ratio. We then compose multi-task features from the foreground feature and the mixed background feature using the decoder + +$$ +\tilde {f} _ {t} = \mathcal {D} \left(f _ {t} ^ {\mathrm {f g}}, \hat {f} _ {t} ^ {\mathrm {b g}}\right). \tag {8} +$$ + +We replace the training data with these recomposed features, exposing the model to diverse background variations. + +# 3.6. Two-Stage Temporal Action Segmentation + +To further enhance the robustness of our model to background variations during testing, we propose a two-stage action segmentation method incorporating a Foreground-Aware Action Refinement (FAAR) module. By focusing on foreground regions, FAAR mitigates the impact of background distractions and refines the initial TAS predictions. + +During training, with ground-truth actions available, we obtain the action-relevant objects and detect the foreground + +regions. We extract CLIP features [55] from the foreground frames and apply a lightweight TAS model to predict foreground-aware action probabilities based solely on the foreground frames, $p_t^{\mathrm{fg}}$ , which are combined with the initial action predictions $p_t$ output by the base TAS model, + +$$ +p _ {t} ^ {\mathrm {f g}} = \mathcal {R} \left(g \left(v _ {t} \odot m _ {t} ^ {\mathrm {f g}}\right)\right), \quad p _ {t} ^ {\text {f i n a l}} = (1 - \alpha) \cdot p _ {t} + \alpha \cdot p _ {t} ^ {\mathrm {f g}}, \tag {9} +$$ + +where $m_t^{\mathrm{fg}}$ is the binary foreground mask, $g$ denotes CLIP feature extractor, $\mathcal{R}$ denotes the lightweight TAS model, $\alpha \in [0,1]$ controls the weight of foreground-aware probabilities, and $p_t^{\mathrm{final}}$ is the refined action probabilities. + +During inference, without ground-truth actions, we first select the top $K$ predicted action classes $\{\hat{a}_t^{(1)},\dots ,\hat{a}_t^{(K)}\}$ based on the initial action probabilities $p_t$ from the base TAS model, and retrieve their relevant objects. We then detect the corresponding foreground regions and obtain foreground-aware action probabilities $p_t^{\mathrm{fg},(k)}$ for each of the top $K$ actions respectively. We compute the final prediction by combining the initial action probabilities with the foreground-aware ones, + +$$ +p _ {t} ^ {\text {f i n a l}} = (1 - \alpha) \cdot p _ {t} + \alpha \cdot \sum_ {k = 1} ^ {K} \theta_ {t} ^ {(k)} p _ {t} ^ {\mathrm {f g}, (k)}, \tag {10} +$$ + +where the weights $\theta_t^{(k)}$ are proportional to the initial probabilities of the top- $K$ actions, i.e., $\theta_t^{(k)} = \frac{p_t\left(\hat{a}_t^{(k)}\right)}{\sum_{j=1}^K p_t\left(\hat{a}_t^{(j)}\right)}$ . This approach refines the predictions by focusing on the most probable actions and their relevant objects, enhancing robustness to background variations in multi-task scenarios. + +# 3.7. Training and Inference + +Training. We use the synthetic multi-task videos generated by MSB for training. We train the SBL module $\mathcal{F}_{\mathrm{SBL}}$ via $\mathcal{L}_{\mathrm{SBL}}$ in Eq. (3) and utilize the reconstructed features at segment boundaries to enhance temporal smoothness. Similarly, we train the FBFC decoder $\mathcal{D}$ via $\mathcal{L}_{\mathrm{FBFC}}$ in Eq. (6), and augment the training data by mixing background features from different tasks as described in Eqs. (7) and (8). We train the base TAS model with FAAR using the standard action loss. The total loss function $\mathcal{L}$ is: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {a c t i o n}} + \lambda_ {\mathrm {S B L}} \mathcal {L} _ {\mathrm {S B L}} + \lambda_ {\mathrm {F B F C}} \mathcal {L} _ {\mathrm {F B F C}}, \tag {11} +$$ + +where $\mathcal{L}_{\mathrm{action}}$ includes the cross-entropy loss and smoothing loss for TAS, applied to the final action probabilities output by the FAAR module. $\lambda_{\mathrm{SBL}}$ and $\lambda_{\mathrm{FBFC}}$ are the loss weights. + +Inference. During inference, we do not use the MSB, SBL and FBFC modules, as they are designed to augment training. We first obtain initial action predictions from the base TAS model, and then apply the FAAR module, as specified in Eq. (10), to derive the final action predictions. + +# 3.8. Incorporating Unlabeled Multi-Task Videos + +Our proposed framework so far trains an MT-TAS model using only single-task videos. However, we additionally investigate the potential benefits of incorporating unlabeled multi-task videos, when such videos are available. Since annotating long multi-task videos is labor-intensive and time-consuming, leveraging unlabeled multi-task videos offers a practical way to enhance the model performance. + +We formulate this as a domain adaptation problem [9, 10, 50], where labeled single-task and synthetic multi-task videos constitute the source domain, while unlabeled real multi-task videos form the target domain. We aim to learn domain-invariant feature representations through adversarial training, achieved by adding a domain classifier with a Gradient Reversal Layer (GRL) [21] after the base TAS model's early feature layer. The model is trained jointly with action labels on the source domain and domain labels on both domains. During training, the GRL reverses the gradient flow, encouraging indistinguishable feature distributions across domains. Our overall training loss is + +$$ +\mathcal {L} = \mathcal {L} _ {\text {a c t i o n}} + \lambda_ {\mathrm {S B L}} \mathcal {L} _ {\mathrm {S B L}} + \lambda_ {\mathrm {F B F C}} \mathcal {L} _ {\mathrm {F B F C}} + \lambda_ {\text {d o m a i n}} \mathcal {L} _ {\text {d o m a i n}}, \tag {12} +$$ + +where $\mathcal{L}_{\mathrm{domain}}$ is the framewise binary domain classification loss, and $\lambda_{\mathrm{domain}}$ is the loss weight. + +# 4. MEKA Dataset + +While our proposed framework can be trained using only single-task videos, we require an MT-TAS benchmark for evaluation. To this end, we introduce the Multi-task Ego-centric Kitchen Activities (MEKA) dataset, which extends the EgoPER dataset [36] to multi-task scenarios. EgoPER is an egocentric procedural video dataset designed for temporal action segmentation and error detection. It contains single-task videos of users performing kitchen activities while wearing an egocentric camera (HoloLens 2). The dataset includes five cooking tasks: making pinwheels, quesadillas, oatmeal, coffee, and tea, each associated with a predefined task graph. + +To create realistic multi-task scenarios, we generated multi-task transcripts by interleaving action sequences from multiple tasks while ensuring logical coherence within the task graphs. Participants equipped with a HoloLens 2 camera followed these transcripts during recording. While we used similar recording equipment and kitchen tools as in the EgoPER dataset to maintain consistency, variations in the environments introduce additional challenges to the evaluation benchmark. + +After data collection, we annotated each video by assigning framewise labels for both actions and tasks. In total, MEKA consists of 100 videos for 12 hours of footage. In each video, participants perform two or three cooking tasks in an interleaved manner. See the supplementary materials for more details about the dataset. We will make the + +MEKA dataset publicly available upon publication, including videos and framewise annotations, to facilitate further research in multi-task temporal action segmentation. + +# 5. Experiments + +# 5.1. Experimental Setup + +Datasets. Since MT-TAS is a new task, existing datasets cannot be used for its evaluation. Consequently, we evaluate our approach using single-task videos from EgoPER [36] for training and our collected multi-task videos from MEKA for testing. EgoPER contains 213 normal single-task egocentric videos and MEKA contains 100 multi-task videos for the five recipes in EgoPER. In total, it has 51 classes of action steps. We were also able to adapt the EGTEA dataset [43] for multi-task evaluation. We manually select actions that appear exclusively in either the first or second half of each video, treating these segments as single-task videos, and consider the full videos with interleaved actions as multi-task videos. The supplementary materials provide additional details and the experimental results on the adapted EGTEA dataset. + +Model Architecture. We evaluate our method with three base TAS models: two offline models (MSTCN [17] and FACT [49]) and one online model (ProTAS [65]). Online TAS differs from offline TAS in that it only uses the frames up to the current moment for prediction. We implement the SBL module $(\mathcal{F}_{\mathrm{SBL}})$ by processing interval features through two convolutional layers, followed by temporal global average pooling. The Decoder $\mathcal{D}$ processes foreground and background features through parallel two-layer MLPs, each reducing feature dimension by half, followed by a linear layer that reconstructs the original feature representation from the concatenated features. We employ a three-layer TCN [17] with a hidden dimension of 64 for FAAR $(\mathcal{R})$ . For experiments with domain adaptation, we insert a GRL after the feature layer of the initial stage of TAS models, and use a two-layer MLP for binary domain classification. + +Implementation Details. Following standard practice [17, 49, 65, 81, 84], we extract 2048-dimensional I3D features [7] pretrained on Kinetics [7] as video representations. We evaluate performance using standard TAS metrics: frame-wise accuracy (Acc), frame-wise accuracy excluding background frames (Acc-bg), segment-wise edit score (Edit), and segment-wise F1 scores at multiple IoU thresholds (F1@10,25,50). We adopt the default configurations from MSTCN [17], FACT [49], and ProTAS [65] using their official source codes. We train all models using Adam optimizer with a learning rate of 0.0005. We set the loss weights as $\lambda_{\mathrm{SBL}} = \lambda_{\mathrm{FBFC}} = \lambda_{\mathrm{domain}} = 1$ . For the SBL module, we define boundary frames as the 10 frames centered at the points where segments from different videos are concatenated, and set window size $w$ as 5 and margin + +
MSBSBLFBFCFAARTAS model: MSTCN [17]TAS model: FACT [49]
AccAcc-bgEditF1@{10,25,50}AccAcc-bgEditF1@{10,25,50}
baseline62.349.150.152.649.339.753.741.442.944.641.631.7
✓67.863.764.268.966.856.072.867.873.173.170.058.4
✓✓68.865.964.770.868.858.173.670.474.576.173.662.3
✓✓✓73.764.469.573.069.958.873.871.374.479.676.765.5
✓✓✓✓75.772.174.979.777.667.475.772.876.081.279.369.9
+ +Table 1. Offline multi-task temporal action segmentation performance on MEKA. + +
MSBSBLFBFCFAARAccEditF1@{10,25,50}
baseline49.124.624.120.012.4
✓51.130.726.522.112.8
✓✓51.133.129.324.514.6
✓✓✓55.237.532.827.118.3
✓✓✓✓67.854.854.649.836.7
+ +$\delta$ as 2. For Foreground-Background Feature Composition (FBFC), we select $10\%$ of frames and augment them with Eq. (7) by randomly sampling $\beta$ from [0.5, 1] during training. See Sec. 5.2 for the qualitative analysis on the value of $\beta$ . For the FAAR module, we employ a two-stage training process. In the first stage, we train the model without FAAR by setting $\alpha = 0$ in Eq. (9). In the second stage, we incorporate FAAR and randomly sample $\alpha$ from [0, 1] during training. For inference, we set $K = 3$ and $\alpha = 0.3$ in Eq. (10). We include ablation studies on the values of $K$ and $\alpha$ in Sec. 5.2. Please refer to the supplementary materials for more implementation details. + +# 5.2. Experimental Results + +Offline Temporal Action Segmentation. Tab. 1 presents the performance of our approach on the MEKA dataset using two baseline offline TAS models: MSTCN [17] and FACT [49]. We start with the baseline models trained on single-task videos without any of our proposed modules (first row). We then progressively incorporate the Multi-task Sequence Blending (MSB), Segment Boundary Learning (SBL), Foreground-Background Feature Composition (FBFC), and Foreground-Aware Action Refinement (FAAR) modules to assess their individual contributions. In the first row, MSTCN achieves an accuracy of $62.3\%$ and an F1@50 score of $39.7\%$ , while FACT attains $53.7\%$ accuracy and $31.7\%$ F1@50, indicating that the baseline models struggle to generalize to multi-task videos when trained solely on single-task data. Incorporating MSB significantly improves performance for both models, especially for FACT, due to its explicit modeling of video transcripts during training. Adding SBL further boosts performance. With MSB and SBL, MSTCN's accuracy improves by $6.5\%$ and F1@50 by $18.4\%$ , while FACT's accuracy + +Table 2. Online temporal action segmentation on MEKA. + +
Training SetupAccEditF1@{10,25,50}
MSTCN [17]Offline Temporal Action Segmentation
LMT78.075.978.376.670.4
LST&LMT80.684.784.582.574.6
Oursw/o DA75.774.979.777.667.4
w/ DA76.6+0.977.6+2.782.0+2.380.6+3.072.2+4.8
+LMT82.787.389.788.582.4
ProTAS [65]Online Temporal Action Segmentation
LMT61.541.137.232.822.0
LST&LMT64.549.839.434.522.7
Oursw/o DA67.854.854.649.836.7
w/ DA69.1+1.356.9+2.157.5+2.952.0+2.238.8+2.1
+LMT74.970.667.262.549.3
+ +Table 3. The effect of domain adaptation using unlabeled multi-task videos and comparison with methods using labeled multi-task videos. + +improves by $19.9\%$ and F1@50 by $30.6\%$ . Incorporating FBFC further improves the accuracy by $4.9\%$ on MSTCN and the F1@50 by $3.2\%$ on FACT. Finally, integrating all proposed modules leads to the highest performance, resulting in an F1@50 score of $67.4\%$ for MSTCN and $69.9\%$ for FACT. Overall, our approach effectively enhances the capability of standard TAS models to handle multi-task videos. + +Online Temporal Action Segmentation. We evaluate the online TAS performance of our approach in Tab. 2, using ProTAS [65] as the base TAS model. Without access to future frames, the online model relies solely on past and current frames, making it challenging to anticipate action transitions and accurately segment boundaries. This limitation results in a noticeable performance drop compared to the offline models, echoing the difficulties of online MT-TAS. However, our proposed modules still achieve a notable performance boost. When all modules are combined, the model increases accuracy by $18.7\%$ and F1@50 by $24.3\%$ compared to the baseline. Particularly, the integration of FAAR improves the accuracy by $12.6\%$ and all F1 scores by nearly $20\%$ , highlighting the value of foreground-aware refinements in scenarios without access to future frames. + +Domain Adaptation with Unlabeled Multi-Task Videos. In Tab. 3, we evaluate our domain adaptation approach using unlabeled multi-task videos, as introduced in Sec. 3.8. Domain adaptation consistently improves performance across all metrics, increasing F1@50 by $4.8\%$ and $2.1\%$ in offline and online settings, respectively. To thoroughly assess our method, we compare against two baseline se + +![](images/cd67e2174538760dc4ad57c12e056e00f6cafa5b5a3bc11836a6d91db026c76e.jpg) + +![](images/c0e86034b9c548275726504e3c6c29173a7b73eed56da81c37a6534b511e4fd9.jpg) + +![](images/5b7ee7fa9602c134e99b580adea4ef2bd8b115f5399e9dc22822167b24a51fe5.jpg) +Figure 4. Ablation Studies on FAAR: (a) comparison of different inputs; (b) effects of $K$ ; (c) effects of $\alpha$ . + +tups using 3-fold cross-validation on multi-task videos: 1) the model is trained on labeled multi-task (LMT) videos; 2) the model is trained on both labeled single-task and labeled multi-task (LST&LMT) videos. Note that these baselines use labeled multi-task data, while our method employs only unlabeled multi-task videos during training. In offline TAS, our model with domain adaptation outperforms the LMT baseline and achieves comparable results to the LST&LMT baseline. In the more challenging online setting, our models surpass both baselines, demonstrating the ability of our modules to enhance generalization and handle the data-intensive demands of online TAS. We also extend our method by incorporating labeled multi-task videos (+LMT) to provide an upper bound for domain adaptation. The performance gap reveals room for improvement in unsupervised domain adaptation. Furthermore, the performance gains from LST&LMT to Ours+LMT demonstrate that our modules provide complementary benefits even when labeled multi-task data is available. + +Ablation Studies on FAAR. In Fig. 4, we conduct ablation studies to evaluate different inputs of FAAR and analyze the effects of $K$ and $\alpha$ in Eq. (10) during inference. First, we change the module's input from foreground frames to background frames or full images. Background frames perform poorly while foreground frames outperform full images, showing the importance of action-relevant foreground information. We then examine how varying $K$ , the number of top predicted action classes considered, and $\alpha$ , the weighting factor between initial predictions and foreground-aware predictions, influences the performance. The model's performance gradually improves with increasing $K$ , reaching a plateau when $K \geq 3$ . As $\alpha$ increases from 0 to 1, the model performs best between 0.3 and 0.4. Notably, while the accuracies at $\alpha = 0$ (base TAS predictions only) and $\alpha = 1$ (foreground-aware predictions only) are similar, the Edit and F1@50 scores for foreground-only predictions are significantly lower. This suggests that FAAR emphasizes individual frame accuracy but may lack the temporal consistency needed for segment-wise metrics. + +Qualitative Analysis of FBFC. In Fig. 5, we qualitatively + +![](images/70d9736337e0a31e03bbbc8769e4c0cedf2830271b2cd3e5c21f9f83fa5759a1.jpg) +Figure 5. Retrieved nearest multi-task videos by different ratios with FBFC. For simplicity, only middle frames of video clips are displayed. + +analyze how FBFC works with different values of the mixing ratio $\beta$ . We sample two clips from single-task videos, extract foreground frames from one clip and background frames from another, and then recompose these features using FBFC via Eqs. (7) and (8). To assess the decomposition quality, we retrieve the nearest clip from multi-task testing videos based on feature similarity. With a lower mixing ratio ( $\beta = 0.2$ ), the recomposed features are more heavily influenced by background features, and the context of the retrieved frames predominantly match the background frames. Conversely, with a higher mixing ratio ( $\beta = 0.8$ ), the retrieved frames often depict the same action as the foreground frames while sharing some similarities with the context of the background frames. Based on these observations, we sample $\beta \in [0.5,1]$ in training to ensure preserving action features while introducing background variations. + +See supplementary materials for more results, qualitative visualization, and discussion on complexity and limitations. + +# 6. Conclusions + +We proposed an MT-TAS framework to tackle the challenges of interleaved actions in multi-task scenarios. Our approach combines synthesized multi-task data, a two-stage TAS model, and the new MEKA benchmark dataset. We validated the effectiveness of our approach through extensive experiments in both offline and online MT-TAS. + +# Acknowledgement + +This work is sponsored by ARPA-H (1AY2AX000062), DARPA PTG (HR00112220001), NSF (IIS-2115110) and ARO (W911NF2110276). Content does not necessarily reflect the position/policy of the Government. + +# References + +[1] Hyemin Ahn and Dongheui Lee. Refining action segmentation with hierarchical video representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16302-16310, 2021. 1, 2 +[2] J. B. Alayrac, P. Bojanowski, N. Agrawal, J. Sivic, I. Laptev, and S. Lacoste-Julien. Unsupervised learning from narrated instruction videos. IEEE Conference on Computer Vision and Pattern Recognition, 2016. 2 +[3] Kyungho Bae, Geo Ahn, Youngrae Kim, and Jinwoo Choi. Devias: Learning disentangled video representations of action and scene for holistic video understanding. In European Conference on Computer Vision, 2024. 3 +[4] Emad Bahrami, Gianpiero Francesca, and Juergen Gall. How much temporal long-term context is needed for action segmentation? In International Conference on Computer Vision (ICCV), 2023. 2 +[5] Siddhant Bansal, Chetan Arora, and C.V. Jawahar. My view is the best view: Procedure learning from egocentric videos. In European Conference on Computer Vision (ECCV), 2022. 1, 2 +[6] Nadine Behrmann, S. Alireza Golestaneh, Zico Kolter, Juergen Gall, and Mehdi Noroozi. Unified fully and timestamp supervised temporal action segmentation via sequence to sequence translation. In ECCV, 2022. 1, 2 +[7] J. Carreira and A. Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In IEEE Conference on Computer Vision and Pattern Recognition, 2017. 4, 6 +[8] Chien-Yi Chang, De-An Huang, Yanan Sui, Li Fei-Fei, and Juan Carlos Niebles. D3tw: Discriminative differentiable dynamic time warping for weakly supervised action alignment and segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2 +[9] Min-Hung Chen, Baopu Li, Yingze Bao, Ghassan Al-Regib, and Zsolt Kira. Action segmentation with joint self-supervised temporal domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9454–9463, 2020. 6 +[10] Victor G. Turrisi da Costa, Giacomo Zara, Paolo Rota, Thiago Oliveira-Santos, Nicu Sebe, Vittorio Murino, and Elisa Ricci. Dual-head contrastive domain adaptation for video action recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1181-1190, 2022. 6 +[11] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision (IJCV), 130:33-55, 2022. 2 + +[12] Li Ding and Chenliang Xu. Weakly-supervised action segmentation with iterative soft boundary assignment. IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2 +[13] Shuangrui Ding, Maomao Li, Tianyu Yang, Rui Qian, Haohang Xu, Qingyi Chen, Jue Wang, and Hongkai Xiong. Motion-aware contrastive video representation learning via foreground-background merging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9716-9726, 2022. 3 +[14] G. Donahue and E. Elhamifar. Learning to predict activity progress by self-supervised video alignment. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1 +[15] E. Elhamifar and D. Huynh. Self-supervised multi-task procedure learning from instructional videos. European Conference on Computer Vision, 2020. 2 +[16] E. Elhamifar and Z. Naing. Unsupervised procedure learning via joint dynamic summarization. International Conference on Computer Vision, 2019. 2 +[17] Yazan Abu Farha and Jurgen Gall. Ms-tcn: Multi-stage temporal convolutional network for action segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3575-3584, 2019. 1, 2, 4, 6, 7 +[18] Alireza Fathi, Xiaofeng Ren, and James M Rehg. Learning to recognize objects in egocentric activities. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2011. 2 +[19] Mohsen Fayyaz and Jurgen Gall. Sct: Set constrained temporal transformer for set supervised action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2 +[20] Daniel Fried, Jean-Baptiste Alayrac, Phil Blunsom, Chris Dyer, Stephen Clark, and Aida Nematzadeh. Learning to segment actions from observation and narration. Annual Meeting of the Association for Computational Linguistics, 2020. 2 +[21] Y. Ganin and V. Lempitsky. Unsupervised domain adaptation by backpropagation. International Conference on Machine Learning, 2015. 6 +[22] Reza Ghoddoosian, Isht Dwivedi, Nakul Agarwal, Chiho Choi, and Behzad Dariush. Weakly-supervised online action segmentation in multi-view instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13780-13790, 2022. 1, 2 +[23] Reza Ghoddosian, Saif Sayed, and Vassilis Athitsos. Hierarchical modeling for task recognition and action segmentation in weakly-labeled instructional videos. IEEE Winter Conference on Applications of Computer Vision, 2022. +[24] Reza Ghoddosian, Isht Dwivedi, Nakul Agarwal, and Behzad Dariush. Weakly-supervised action segmentation and unseen error detection in anomalous instructional videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10128-10138, 2023. 1, 2 +[25] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Mar + +tin, Tushar Nagarajan, Ilija Radosavovic, Santhosh K. Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Z. Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Christian Fuegen, Abraham Gebreselasie, Cristina Gonzalez, James M. Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolár, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran K. Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Yunyi Zhu, Pablo Arbeláez, David J. Crandall, Dima Damen, Giovanni Maria Farinella, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard A. Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the world in 3,000 hours of egocentric video. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18973-18990, 2021. 2 +[26] Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, et al. Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19383-19400, 2024. 2 +[27] Yifei Huang, Yusuke Sugano, and Yoichi Sato. Improving action segmentation via graph-based temporal reasoning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[28] Yifei Huang, Guo Chen, Jilan Xu, Mingfang Zhang, Lijin Yang, Baoqi Pei, Hongjie Zhang, Lu Dong, Yali Wang, Limin Wang, and Yu Qiao. Egoexolearn: A dataset for bridging asynchronous ego- and exo-centric view of procedural activities in real world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22072-22086, 2024. 2 +[29] Yuchi Ishikawa, Seito Kasai, Yoshimitsu Aoki, and Hirokatsu Kataoka. Alleviating over-segmentation errors by detecting action boundaries. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2322-2331, 2021. 2 +[30] Hilde Kuehne, Ali Arslan, and Thomas Serre. The language of actions: Recovering the syntax and semantics of goal-directed human activities. IEEE Conference on Computer Vision and Pattern Recognition, 2014. 2 +[31] H. Kuehne, A. Arslan, and T. Serre. The language of actions: Recovering the syntax and semantics of goal-directed human. IEEE Conference on Computer Vision and Pattern Recognition, 2014. 1, 2 +[32] H. Kuehne, J. Gall, and T. Serre. An end-to-end generative framework for video segmentation and recognition. + +IEEE Winter Conference on Applications of Computer Vision, 2016. 2 +[33] Anna Kukleva, Hilde Kuehne, Fadime Sener, and Jurgen Gall. Unsupervised learning of action classes with continuous temporal embedding. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1, 2 +[34] Sateesh Kumar, Sanjay Haresh, Awais Ahmed, Andrey Konin, M. Zeeshan Zia, and Quoc-Huy Tran. Unsupervised action segmentation by joint representation learning and online clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20174-20185, 2022. 2 +[35] C. Lea, M. D. Flynn, R. Vidal, A. Reiter, and G. D. Hager. Temporal convolutional networks for action segmentation and detection. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2 +[36] S. Lee, Z. Lu, Z. Zhang, M. Hoai, and E. Elhamifar. Error detection in egocentric procedural task videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 6 +[37] Haoxin Li, Yuan Liu, Hanwang Zhang, and Boyang Li. Mitigating and evaluating static bias of action representations in the background and the foreground. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19911-19923, 2023. 3 +[38] Jun Li and Sinisa Todorovic. Set-constrained viterbi for set-supervised action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2 +[39] J. Li and S. Todorovic. Anchor-constrained viterbi for set-supervised action segmentation. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. +[40] J. Li, P. Lei, and S. Todorovic. Weakly supervised energy-based learning for action segmentation. International Conference on Computer Vision, 2019. 2 +[41] M. Li, L. Chen, Y. Duarr, Z. Hu, J. Feng, J. Zhou, and J. Lu. Bridge-prompt: Towards ordinal action understanding in instructional videos. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2 +[42] Shi-Jie Li, Yazan AbuFarha, Yun Liu, Ming-Ming Cheng, and Juergen Gall. Ms-tcn++: Multi-stage temporal convolutional network for action segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2020. 1, 2 +[43] Yin Li, Miao Liu, and James M. Rehg. In the eye of beholder: Joint learning of gaze and actions in first person video. European Conference on Computer Vision, 2018. 2, 6 +[44] Zhe Li, Yazan Abu Farha, and Jurgen Gall. Temporal action segmentation from timestamp supervision. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2 +[45] Daochang Liu, Qiyue Li, Anh-Dung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. International Conference on Computer Vision, 2023. 1, 2, 4 +[46] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun + +Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. *eccv*, 2024. 5 +[47] Z. Lu and E. Elhamifar. Weakly-supervised action segmentation and alignment via transcript-aware union-of-subspaces learning. International Conference on Computer Vision, 2021. 2 +[48] Z. Lu and E. Elhamifar. Set-supervised action learning in procedural task videos via pairwise order consistency. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 1, 2 +[49] Z. Lu and E. Elhamifar. Fact: Frame-action cross-attention temporal modeling for efficient action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 4, 6, 7 +[50] J. Munro and D. Damen. Multi-modal domain adaptation for fine-grained action recognition. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 6 +[51] Zhanzhong Pang, Fadime Sener, Shrinivas Ramasubramanian, and Angela Yao. Long-tail temporal action segmentation with group-wise temporal logit adjustment. 2024. 2 +[52] Sam Perochon and Laurent Oudre. Unsupervised action segmentation of untrimmed egocentric videos. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 1 +[53] Will Price, Carl Vondrick, and Dima Damen. Unweavenet: Unweaving activity stories. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13770-13779, 2022. 2 +[54] Camillo Quattrocchi, Antonino Furnari, Daniele Di Mauro, Mario Valerio Giuffrida, and Giovanni Maria Farinella. Synchronization is all you need: Exocentric-to-egocentric transfer for temporal action segmentation with unlabeled synchronized video pairs. European Conference on Computer Vision, 2024. 1 +[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. International Conference on Machine learning, 2021. 5 +[56] Francesco Ragusa, Antonino Furnari, Salvatore Livatino, and Giovanni Maria Farinella. The meccano dataset: Understanding human-object interactions from egocentric videos in an industrial-like domain. Winter Conference on Applications of Computer Vision, 2021. 2 +[57] Rahul Rahaman, Dipika Singhania, Alexandre Thiery, and Angela Yao. A generalized and robust framework for timestamp supervision in temporal action segmentation. In Computer Vision-ECCV 2022: 17th European Conference, 2022. 2 +[58] A. Richard, H. Kuehne, and J. Gall. Action sets: Weakly supervised action segmentation without ordering constraints. IEEE Conference on Computer Vision and Pattern Recognition, 2018. +[59] A. Richard, H. Kuehne, A. Iqbal, and J. Gall. Neuralnetwork-viterbi: A framework for weakly supervised video learning. IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2 + +[60] M. Rohrbach, S. Amin, M. Andriluka, and B. Schiele. A database for fine grained activity detection of cooking activities. IEEE Conference on Computer Vision and Pattern Recognition, 2012. 2 +[61] Saquib Sarfraz, Naila Murray, Vivek Sharma, Ali Diba, Luc Van Gool, and Rainer Stiefelhagen. Temporally-weighted hierarchical clustering for unsupervised action segmentation. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2 +[62] Saif Sayed, Reza Ghoddoosian, Bhaskar Trivedi, and Vassilis Athitsos. A new dataset and approach for timestamp supervised action segmentation using human object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 3133-3142, 2023. 2 +[63] Fadime Sener, Dibyadip Chatterjee, Daniel Shelepov, Kun He, Dipika Singhania, Robert Wang, and Angela Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21096-21106, 2022. 2 +[64] Y. Shen and E. Elhamifar. Semi-weakly-supervised learning of complex actions from instructional task videos. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 1, 2 +[65] Y. Shen and E. Elhamifar. Progress-aware online action segmentation for egocentric procedural task videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 6, 7 +[66] Y. Shen, L. Wang, and E. Elhamifar. Learning to segment actions from visual and language instructions via differentiable weak sequence alignment. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2 +[67] Y. Shen, H. Wang, X. Yang, M. Feiszli, E. Elhamifar, L. Torresani, and E. Mavroudi. Learning to segment referred objects from narrated egocentric videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 3 +[68] G. A. Sigurdsson, S. Divvala, A. Farhadi, and A. Gupta. Asynchronous temporal fields for action recognition. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2 +[69] B. Singh, T. K. Marks, M. Jones, O. Tuzel, and M. Shao. A multi-stream bi-directional recurrent neural network for finegrained action detection. IEEE Conference on Computer Vision and Pattern Recognition, 2016. +[70] Dipika Singhania, Rahul Rahaman, and Angela Yao. C2f- tcn: A framework for semi-and fully-supervised temporal action segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):11484-11501, 2023. +[71] Yaser Souri, Mohsen Fayyaz, Luca Minciullo, Gianpiero Francesca, and Juergen Gall. Fast Weakly Supervised Action Segmentation Using Mutual Consistency. PAMI, 2021. 2 +[72] Sebastian Stein and Stephen J McKenna. Combining embedded accelerometers with computer vision for recognizing food preparation activities. In Proceedings of the 2013 ACM international joint conference on Pervasive and ubiquitous computing, pages 729-738, 2013. 2 + +[73] Yuhao Su and Ehsan Elhamifar. Two-stage active learning for efficient temporal action segmentation. In European Conference on Computer Vision, pages 161-183. Springer, 2024. 1 +[74] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[75] Sanket Thakur, Cigdem Bayan, Pietro Morerio, Vittorio Murino, and Alessio Del Bue. Anticipating next active objects for egocentric videos. IEEE Access, 2024. 3 +[76] Quoc-Huy Tran, Ahmed Mehmood, Muhammad Ahmed, Muhammad Naufil, Anas Zafar, Andrey Konin, and Zeeshan Zia. Permutation-aware activity segmentation via unsupervised frame-to-segment alignment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6426-6436, 2024. 2 +[77] Huiyu Wang, Mitesh Kumar Singh, and Lorenzo Torresani. Ego-only: Egocentric action detection without exocentric transferring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5250-5261, 2023. 1 +[78] Jiahao Wang, Guo Chen, Yifei Huang, Limin Wang, and Tong Lu. Memory-and-anticipation transformer for online action understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13824–13835, 2023. 2 +[79] Zhe Wang, Hao Chen, Xinyu Li, Chunhui Liu, Yuanjun Xiong, Joseph Tighe, and Charless Fowlkes. Sscap: Self-supervised co-occurrence action parsing for unsupervised temporal action segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1819–1828, 2022. 2 +[80] Ming Xu and Stephen Gould. Temporally consistent unbalanced optimal transport for unsupervised action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2 +[81] Ziwei Xu, Yogesh S Rawat, Yongkang Wong, Mohan Kankanhalli, and Mubarak Shah. Don't pour cereal into coffee: Differentiable temporal logic for temporal action segmentation. In Advances in Neural Information Processing Systems, 2022. 6 +[82] Lingfeng Yang, Yueze Wang, Xiang Li, Xinlong Wang, and Jian Yang. Fine-grained visual prompting. Advances in Neural Information Processing Systems, 36, 2023. 5 +[83] S. Yeung, O. Russakovsky, N. Jin, M. Andriluka, G. Mori, and L. Fei-Fei. Every moment counts: Dense detailed labeling of actions in complex videos. International Journal of Computer Vision, 2018. 2 +[84] Fangqiu Yi, Hongyu Wen, and Tingting Jiang. Asformer: Transformer for action segmentation. In The British Machine Vision Conference (BMVC), 2021. 1, 2, 4, 6 +[85] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Helping hands: An object-aware ego-centric video recognition model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13901-13912, 2023. 3 + +[86] Ce Zhang, Changcheng Fu, Shijie Wang, Nakul Agarwal, Kwonjoon Lee, Chiho Choi, and Chen Sun. Object-centric video representation for long-term action anticipation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6751-6761, 2024. 3 +[87] Xingyi Zhou, Anurag Arnab, Chen Sun, and Cordelia Schmid. How can objects help action recognition? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2362, 2023. 3 +[88] D. Zhukov, J. B. Alayrac, R. G. Cinbis, D. Fouhey, I. Laptev, and J. Sivic. Cross-task weakly supervised learning from instructional videos. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1, 2 \ No newline at end of file diff --git a/2025/Understanding Multi-Task Activities from Single-Task Videos/images.zip b/2025/Understanding Multi-Task Activities from Single-Task Videos/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4ca76f2a4de29984d1e3a191fec4043198dc9ba8 --- /dev/null +++ b/2025/Understanding Multi-Task Activities from Single-Task Videos/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c730c99795a252531fcf2cf83dafc97a10b30fa4c95172dd192036d93b4e9bc2 +size 444041 diff --git a/2025/Understanding Multi-Task Activities from Single-Task Videos/layout.json b/2025/Understanding Multi-Task Activities from Single-Task Videos/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..628b752c8d843e88af494f7ecb63ffd2fffa92db --- /dev/null +++ b/2025/Understanding Multi-Task Activities from Single-Task Videos/layout.json @@ -0,0 +1,10702 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 115, + 103, + 496, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 103, + 496, + 121 + ], + "spans": [ + { + "bbox": [ + 115, + 103, + 496, + 121 + ], + "type": "text", + "content": "Understanding Multi-Task Activities from Single-Task Videos" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 144, + 265, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 144, + 265, + 184 + ], + "spans": [ + { + "bbox": [ + 138, + 144, + 265, + 184 + ], + "type": "text", + "content": "Yuhan Shen Northeastern University shen.yuh@northeastern.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 310, + 144, + 463, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 144, + 463, + 184 + ], + "spans": [ + { + "bbox": [ + 310, + 144, + 463, + 184 + ], + "type": "text", + "content": "Ehsan Elhamifar Northeastern University e.elhamifar@northeastern.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 238, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 296, + 514 + ], + "type": "text", + "content": "We introduce and develop a framework for Multi-Task Temporal Action Segmentation (MT-TAS), a novel paradigm that addresses the challenges of interleaved actions when performing multiple tasks simultaneously. Traditional action segmentation models, trained on single-task videos, struggle to handle task switches and complex scenes inherent in multi-task scenarios. To overcome these challenges, our MT-TAS approach synthesizes multi-task video data from single-task sources using our Multi-task Sequence Blending and Segment Boundary Learning modules. Additionally, we propose to dynamically isolate foreground and background elements within video frames, addressing the intricacies of object layouts in multi-task scenarios and enabling a new two-stage temporal action segmentation framework with Foreground-Aware Action Refinement. Also, we introduce the Multi-task Egocentric Kitchen Activities (MEKA) dataset, containing 12 hours of egocentric multi-task videos, to rigorously benchmark MT-TAS models. Extensive experiments demonstrate that our framework effectively bridges the gap between single-task training and multi-task testing, advancing temporal action segmentation with state-of-the-art performance in complex environments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 529, + 135, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 529, + 135, + 541 + ], + "spans": [ + { + "bbox": [ + 56, + 529, + 135, + 541 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 550, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 296, + 693 + ], + "type": "text", + "content": "Consider a typical morning routine, when you prepare breakfast: you start making oatmeal and while waiting for it to cook, you brew coffee and pack your food for lunch. In fact, in our daily lives, we often perform multiple tasks simultaneously. For computer vision systems to accurately analyze such behaviors and provide useful assistance, they must comprehend these complex and interleaved activities. Temporal action segmentation (TAS) has emerged as a crucial technology for understanding human activities by partitioning videos into meaningful segments, each corresponding to a distinct action or step [17, 22, 24, 31, 33, 48, 49, 64, 73, 84, 88]. Recently, TAS has attracted growing interests" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 213, + 548, + 292 + ], + "blocks": [ + { + "bbox": [ + 317, + 213, + 548, + 292 + ], + "lines": [ + { + "bbox": [ + 317, + 213, + 548, + 292 + ], + "spans": [ + { + "bbox": [ + 317, + 213, + 548, + 292 + ], + "type": "image", + "image_path": "876624b57e93088296560aa0e7567f0fb2210e50cd7c430c726ab5ffeb622451.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 323, + 294, + 388, + 365 + ], + "blocks": [ + { + "bbox": [ + 323, + 294, + 388, + 365 + ], + "lines": [ + { + "bbox": [ + 323, + 294, + 388, + 365 + ], + "spans": [ + { + "bbox": [ + 323, + 294, + 388, + 365 + ], + "type": "image", + "image_path": "554a9cc7c6b4d3d3b32430a73a9db0d07b0df68fbfd7fc36a6182a46ed326df7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 391, + 294, + 454, + 367 + ], + "blocks": [ + { + "bbox": [ + 367, + 367, + 405, + 373 + ], + "lines": [ + { + "bbox": [ + 367, + 367, + 405, + 373 + ], + "spans": [ + { + "bbox": [ + 367, + 367, + 405, + 373 + ], + "type": "text", + "content": "single-task videos" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 294, + 454, + 367 + ], + "lines": [ + { + "bbox": [ + 391, + 294, + 454, + 367 + ], + "spans": [ + { + "bbox": [ + 391, + 294, + 454, + 367 + ], + "type": "image", + "image_path": "9356039069186c8d6417b05ee8d78549904a653fe932c676beb3732d3ca8899c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 385, + 555, + 444 + ], + "lines": [ + { + "bbox": [ + 313, + 385, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 555, + 444 + ], + "type": "text", + "content": "Figure 1. (a) Multi-task temporal action segmentation involves segmenting actions from interleaved tasks. It requires capturing interruptions and resumptions in the execution of each task. (b) Comparison between frames from single-task and multi-task videos. In multi-task videos, at each time instant, the scene often contains objects from tasks different from the currently executed one (highlighted by purple boxes)." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 477, + 294, + 541, + 365 + ], + "blocks": [ + { + "bbox": [ + 477, + 294, + 541, + 365 + ], + "lines": [ + { + "bbox": [ + 477, + 294, + 541, + 365 + ], + "spans": [ + { + "bbox": [ + 477, + 294, + 541, + 365 + ], + "type": "image", + "image_path": "f43393b8d019cb9d2ebf1840e71f5897a05230b65306117e98d26f44ec2fd7bd.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 489, + 367, + 525, + 372 + ], + "lines": [ + { + "bbox": [ + 489, + 367, + 525, + 372 + ], + "spans": [ + { + "bbox": [ + 489, + 367, + 525, + 372 + ], + "type": "text", + "content": "multi-task videos" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 375, + 537, + 383 + ], + "lines": [ + { + "bbox": [ + 323, + 375, + 537, + 383 + ], + "spans": [ + { + "bbox": [ + 323, + 375, + 537, + 383 + ], + "type": "text", + "content": "(b) Comparison between frames from single-task videos v.s. multi-task videos" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 460, + 555, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 519 + ], + "type": "text", + "content": "within egocentric videos [5, 36, 52, 54, 65, 77], where activities are captured from a first-person perspective, enabling fine-grained analysis of human actions, which is essential for applications such as real-time personalized assistance, embodied agents and efficient video content retrieval." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we study the new problem of Multi-Task Temporal Action Segmentation (MT-TAS), which explicitly focuses on scenarios where users perform multiple tasks by interleaving their steps within a single video, see Fig. 1. There are two major challenges that existing TAS models face in addressing MT-TAS. First, despite the notable progress through deep learning techniques [1, 6, 14, 17, 41, 42, 45, 49, 84], current TAS models have primarily focused on scenarios where each video contains actions related to a single high-level task, such as assembling a piece of furniture or cooking a specific recipe. This single-task focus limits the applicability of TAS models to the multi-task setting since i) they are biased to predict actions/steps from only one task at the inference time, ii) when performing several tasks, the environment and the scene often contains objects relevant to multiple tasks, which adds complexity" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 270, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 270, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 270, + 712 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 702, + 270, + 712 + ], + "type": "text", + "content": "Code: https://github.com/Yuhan-Shen/MT-TAS." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19120" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "content": "to the context, while single-task videos often contain objects relevant to one task (see Fig. 1(b)), iii) interleaving steps of multiple tasks introduces complexities such as unpredictable task switches and disruptions to the temporal flows of actions within each task." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 294, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 294, + 240 + ], + "type": "text", + "content": "Second, a significant obstacle in advancing MT-TAS is the lack of appropriate training data. Existing datasets [5, 16, 18, 31, 36, 63, 88] are predominantly collected for single-task scenarios, making them insufficient for capturing complexities of interleaved tasks. Furthermore, many possible step combinations in the multi-task setting makes it impractical to collect and annotate comprehensive real-world videos covering all potential situations (such videos would also be much longer than single-task videos)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 245, + 296, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 245, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 56, + 245, + 296, + 496 + ], + "type": "text", + "content": "Paper Contributions. We study the new problem of MT-TAS. To overcome the aforementioned challenges, we propose an MT-TAS framework that uses only single-task videos for training. We generate synthetic multi-task data using single-task videos and propose a two-stage action segmentation model for accurate predictions. For the data generation, we first develop a Multi-task Sequence Blending (MSB) module that leverages Large Language Models (LLMs) to create multi-task transcripts of interleaved steps and their associated videos. Second, we develop a Segment Boundary Learning (SBL) module to mitigate the abrupt transitions at segment boundaries caused by the synthetic data. Third, we introduce a Dynamic Isolation of Video Elements (DIVE) approach to isolate foreground and background components within video frames by identifying action-relevant objects and combining foregrounds of actions in one task with backgrounds of other tasks to capture the more complex scenes of multi-task videos. By increasing the variation in the background components for the training data, our method becomes more robust to unseen object compositions in multi-task scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 498, + 294, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 294, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 294, + 605 + ], + "type": "text", + "content": "For the action segmentation model, we develop a two-stage framework, where we refine the initial action predictions using a Foreground-Aware Action Refinement (FAAR) module. This approach leverages foreground information to enhance the focus of our model on action-relevant cues (e.g., focusing on the regions of kettle and measuring cup for the action \"transfer water to kettle\" as in Fig. 3(b)), improving accuracy and robustness to background variations during the multi-task testing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 294, + 714 + ], + "type": "text", + "content": "For proper evaluation of MT-TAS models, we gather and annotate the Multi-task Egocentric Kitchen Activities (MEKA) video dataset (we plan to publicly release the dataset), built upon the recent EgoPER [36] dataset. MEKA comprises 100 egocentric videos of users (about 12 hours of footage) performing multiple kitchen recipes from EgoPER in each run by interleaving their steps. This new benchmark captures a variety of interleaved kitchen activities, providing a rigorous testing ground for MT-TAS models. Through" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "extensive experiments in both offline and online settings, we demonstrate that our proposed framework successfully bridges the gap between single-task training data and the complexities of real-world multi-task scenarios, advancing research in temporal action segmentation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 140, + 406, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 406, + 153 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 406, + 153 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 161, + 553, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 161, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 161, + 553, + 447 + ], + "type": "text", + "content": "Temporal Action Segmentation. TAS aims to understand the sequential structure of human activities in videos by assigning an action label to each frame, thereby partitioning the video into a sequence of action segments. Based on the level of supervision during training, TAS methods can be categorized into three main groups: unsupervised [2, 15, 16, 20, 33, 34, 61, 66, 76, 79, 80, 88], weakly-supervised [8, 12, 19, 22-24, 38-40, 44, 47, 48, 57-59, 62, 64, 71], and fully-supervised [1, 4, 6, 17, 27, 29, 32, 35, 36, 41, 42, 45, 49, 51, 60, 65, 68-71, 83, 84] approaches. Our work primarily aligns with the fully-supervised methods. However, regardless of the supervision level, previous studies have predominantly focused on single-task TAS, where test videos contain actions from only one high-level activity or task, while we focus on multi-task TAS. A closely related work, UnweaveNet [53], aims to predict the start, continuation, and resumption of multiple activity threads within a video. While UnweaveNet models multiple concurrent activities, its predictions are at the activity or task level and do not capture the fine-grained action steps within each task. In contrast, our approach focuses on recognizing and segmenting detailed action steps across interleaved tasks, providing a more granular understanding of multi-task activities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 449, + 553, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 553, + 700 + ], + "type": "text", + "content": "Procedural Video Datasets. To support research in TAS and related tasks, numerous procedural video datasets have been collected and released. These datasets include third-person videos (e.g., Breakfast [30], 50Salads [72], YouTube Instructional [2], COIN [74], CrossTask [88], ProceL [16], and ATA [24]), egocentric videos (e.g., GTEA [18], EGTEA [43], MECCANO [56], EgoProceL [5], HoloAssist [78], and EgoPER [36]), or both views such as Assembly101 [63], " + }, + { + "bbox": [ + 313, + 449, + 553, + 700 + ], + "type": "inline_equation", + "content": "(3 + 1)\\mathrm{ReC}" + }, + { + "bbox": [ + 313, + 449, + 553, + 700 + ], + "type": "text", + "content": " [62] and EgoExoLearn [28]. While large-scale video datasets like EPIC-Kitchens [11], Ego4D [25], and Ego-Exo4D [26] exist, they primarily focus on detecting and recognizing atomic action steps rather than modeling the procedural relationships between steps within a task, making them less suitable for TAS research. Also, most existing TAS datasets are limited to single-task scenarios, where each video contains actions related to one high-level activity. In contrast, our collected MEKA dataset specifically addresses multi-task temporal action segmentation by capturing videos where multiple tasks are performed in an interleaved manner, providing a more challenging and realistic benchmark for advancing TAS research." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "content": "Object-Aware and Foreground/Background Learning." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19121" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": "Learning to localize active objects or relevant objects is a prominent topic in video understanding [67, 75, 85, 86]. Furthermore, researchers have explored leveraging object information to improve video representation learning [85], action recognition [87], and action anticipation [86]. Some studies have also focused on disentangling foreground and background elements to enhance video representations [3, 13, 37]. Our work differs by specifically addressing the generalization of TAS models to multi-task videos. Unlike existing approaches that concentrate on short video clips or single-task videos, we propose a method that dynamically isolates foreground and background elements to handle unseen object compositions in multi-task scenarios, thereby mitigating interference from irrelevant objects." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 255, + 290, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 290, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 290, + 269 + ], + "type": "text", + "content": "3. Multi-Task Temporal Action Segmentation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 275, + 155, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 155, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 155, + 289 + ], + "type": "text", + "content": "3.1. Problem Setting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 294, + 296, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 296, + 451 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 296, + 451 + ], + "type": "text", + "content": "Given a video sequence " + }, + { + "bbox": [ + 55, + 294, + 296, + 451 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = (v_{1},\\ldots ,v_{T})" + }, + { + "bbox": [ + 55, + 294, + 296, + 451 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 294, + 296, + 451 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 294, + 296, + 451 + ], + "type": "text", + "content": " frames, the goal of TAS is to predict the sequence of action labels " + }, + { + "bbox": [ + 55, + 294, + 296, + 451 + ], + "type": "inline_equation", + "content": "A = (a_{1},\\dots,a_{T})" + }, + { + "bbox": [ + 55, + 294, + 296, + 451 + ], + "type": "text", + "content": ", where each label " + }, + { + "bbox": [ + 55, + 294, + 296, + 451 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 55, + 294, + 296, + 451 + ], + "type": "text", + "content": " comes from a predefined set of action classes " + }, + { + "bbox": [ + 55, + 294, + 296, + 451 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 294, + 296, + 451 + ], + "type": "text", + "content": ". Assume we have " + }, + { + "bbox": [ + 55, + 294, + 296, + 451 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 55, + 294, + 296, + 451 + ], + "type": "text", + "content": " tasks, where the set " + }, + { + "bbox": [ + 55, + 294, + 296, + 451 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_o" + }, + { + "bbox": [ + 55, + 294, + 296, + 451 + ], + "type": "text", + "content": " consists of actions for the task " + }, + { + "bbox": [ + 55, + 294, + 296, + 451 + ], + "type": "inline_equation", + "content": "o\\in \\{1,\\dots,O\\}" + }, + { + "bbox": [ + 55, + 294, + 296, + 451 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 294, + 296, + 451 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\bigcup_{o = 1}^{O}\\mathcal{C}_{o}" + }, + { + "bbox": [ + 55, + 294, + 296, + 451 + ], + "type": "text", + "content": " contains actions of all tasks. In a single-task video, actions come from one " + }, + { + "bbox": [ + 55, + 294, + 296, + 451 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_o" + }, + { + "bbox": [ + 55, + 294, + 296, + 451 + ], + "type": "text", + "content": ", while a multi-task video has interleaved actions from several tasks, hence, " + }, + { + "bbox": [ + 55, + 294, + 296, + 451 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 55, + 294, + 296, + 451 + ], + "type": "text", + "content": " may belong to any " + }, + { + "bbox": [ + 55, + 294, + 296, + 451 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_o" + }, + { + "bbox": [ + 55, + 294, + 296, + 451 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 294, + 296, + 451 + ], + "type": "inline_equation", + "content": "o\\in \\{1,\\dots,O\\}" + }, + { + "bbox": [ + 55, + 294, + 296, + 451 + ], + "type": "text", + "content": ". Additionally, there may be background segments (e.g., 'answering phone') irrelevant to any of the tasks. A particular challenge in MT-TAS is to distinguish between action segments belonging to different tasks and background segments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 453, + 296, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 296, + 501 + ], + "type": "text", + "content": "In our setting, the training data consists of single-task videos with full framewise annotations. Our goal is to train a model on single-task videos so that it can effectively generalize to multi-task videos during testing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 503, + 239, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 239, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 239, + 516 + ], + "type": "text", + "content": "3.2. Overview of Proposed Framework" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "content": "Fig. 2 shows the overview of our MS-TAS framework. Our approach first employs Multi-task Sequence Blending (MSB) to generate synthetic multi-task videos by leveraging LLM-guided task transitions, eliminating the need for multi-task data collection. However, the initial synthesized multi-task videos contain abrupt transitions at segment boundaries, which disrupt the temporal flow, and contain discrepancies in object compositions between single-task and multi-task videos. To mitigate these issues, we introduce three components: a Segment Boundary Learning (SBL) module that enhances temporal smoothness between segments, a Dynamic Isolation of Video Elements approach that augments the background variations during training, and a two-stage TAS method with foreground-aware action refinement to improve the predictions of the model. Next, we discuss each component of our framework in details." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 73, + 550, + 369 + ], + "blocks": [ + { + "bbox": [ + 317, + 73, + 550, + 369 + ], + "lines": [ + { + "bbox": [ + 317, + 73, + 550, + 369 + ], + "spans": [ + { + "bbox": [ + 317, + 73, + 550, + 369 + ], + "type": "image", + "image_path": "06105a7da63057dc2cd121d50dfc8ccf433ef315bfdd074cf2dc27bb4a16576b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 372, + 555, + 406 + ], + "lines": [ + { + "bbox": [ + 313, + 372, + 555, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 555, + 406 + ], + "type": "text", + "content": "Figure 2. Framework Overview. The modules marked with " + }, + { + "bbox": [ + 313, + 372, + 555, + 406 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 372, + 555, + 406 + ], + "type": "text", + "content": " are used exclusively for training, and modules marked with " + }, + { + "bbox": [ + 313, + 372, + 555, + 406 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 313, + 372, + 555, + 406 + ], + "type": "text", + "content": " are used for both training and testing." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 422, + 511, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 511, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 511, + 435 + ], + "type": "text", + "content": "3.3. Multi-task Sequence Blending (MSB)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 441, + 555, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 603 + ], + "type": "text", + "content": "The goal of MSB is to create initial multi-task video sequences from single-task videos, to simulate real-world multitasking behaviors. Given single-task training videos " + }, + { + "bbox": [ + 313, + 441, + 555, + 603 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{V}^{(i)}\\}_{i=1}^{M}" + }, + { + "bbox": [ + 313, + 441, + 555, + 603 + ], + "type": "text", + "content": ", where each video " + }, + { + "bbox": [ + 313, + 441, + 555, + 603 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^{(i)} = (v_{1}^{(i)}, v_{2}^{(i)}, \\ldots, v_{T_{i}}^{(i)})" + }, + { + "bbox": [ + 313, + 441, + 555, + 603 + ], + "type": "text", + "content": " has a task label " + }, + { + "bbox": [ + 313, + 441, + 555, + 603 + ], + "type": "inline_equation", + "content": "\\tau_{i}" + }, + { + "bbox": [ + 313, + 441, + 555, + 603 + ], + "type": "text", + "content": " and consists of a sequence of action segments " + }, + { + "bbox": [ + 313, + 441, + 555, + 603 + ], + "type": "inline_equation", + "content": "S^{(i)} = (s_{1}^{(i)}, s_{2}^{(i)}, \\ldots, s_{N_{i}}^{(i)})" + }, + { + "bbox": [ + 313, + 441, + 555, + 603 + ], + "type": "text", + "content": " with action labels " + }, + { + "bbox": [ + 313, + 441, + 555, + 603 + ], + "type": "inline_equation", + "content": "A^{(i)} = (a_{1}^{(i)}, a_{2}^{(i)}, \\ldots, a_{N_{i}}^{(i)})" + }, + { + "bbox": [ + 313, + 441, + 555, + 603 + ], + "type": "text", + "content": ", a naive approach would be to randomly sample segments from videos to concatenate. However, this ignores the temporal order of actions within a task and more importantly can create unrealistic sequences, e.g., \"scooping jelly from jar\" followed by \"grinding coffee beans\" (this is unrealistic since it is common sense to spread the jelly right away after scooping it from a jar)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": "To create plausible multi-task sequences, we: i) preserve the original temporal order of action segments within each task, ii) preserve common sense action ordering and switches among tasks. Specifically, the sequence of action segments " + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "inline_equation", + "content": "s_j^{(i)}" + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": " from task " + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\tau_i" + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": " follows the same order as in the original single-task video. At the end of an action segment, we leverage an LLM, which encodes extensive contextual and commonsense knowledge from real-world data, to decide whether to continue the current task or switch to an-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19122" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 245, + 161 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 245, + 161 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 245, + 161 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 245, + 161 + ], + "type": "image", + "image_path": "ec6246c1a902940358b037def5032591c178a36212053ff045e9457870704a0a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 90, + 165, + 195, + 175 + ], + "lines": [ + { + "bbox": [ + 90, + 165, + 195, + 175 + ], + "spans": [ + { + "bbox": [ + 90, + 165, + 195, + 175 + ], + "type": "text", + "content": "(a) Multi-task Sequence Blending" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 250, + 70, + 553, + 163 + ], + "blocks": [ + { + "bbox": [ + 250, + 70, + 553, + 163 + ], + "lines": [ + { + "bbox": [ + 250, + 70, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 250, + 70, + 553, + 163 + ], + "type": "image", + "image_path": "05d873dcd92b3f451998b6be9d9e716d31161b854f1ac6792057d68a435fe8b5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 165, + 462, + 175 + ], + "lines": [ + { + "bbox": [ + 347, + 165, + 462, + 175 + ], + "spans": [ + { + "bbox": [ + 347, + 165, + 462, + 175 + ], + "type": "text", + "content": "(b) Dynamic Isolation of Video Elements" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 129, + 184, + 478, + 195 + ], + "lines": [ + { + "bbox": [ + 129, + 184, + 478, + 195 + ], + "spans": [ + { + "bbox": [ + 129, + 184, + 478, + 195 + ], + "type": "text", + "content": "Figure 3. Illustration of leveraging LLM in our proposed framework for MSB (a) and DIVE (b)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 216, + 295, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 216, + 295, + 344 + ], + "spans": [ + { + "bbox": [ + 54, + 216, + 295, + 344 + ], + "type": "text", + "content": "other task (see Fig. 3(a)). The LLM assesses the plausibility of continuing or switching based on the current action " + }, + { + "bbox": [ + 54, + 216, + 295, + 344 + ], + "type": "inline_equation", + "content": "a_{j}^{(i)}" + }, + { + "bbox": [ + 54, + 216, + 295, + 344 + ], + "type": "text", + "content": ", the next action of the current task " + }, + { + "bbox": [ + 54, + 216, + 295, + 344 + ], + "type": "inline_equation", + "content": "a_{j+1}^{(i)}" + }, + { + "bbox": [ + 54, + 216, + 295, + 344 + ], + "type": "text", + "content": ", and the next action of another task " + }, + { + "bbox": [ + 54, + 216, + 295, + 344 + ], + "type": "inline_equation", + "content": "a_{m}^{(k)}" + }, + { + "bbox": [ + 54, + 216, + 295, + 344 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 216, + 295, + 344 + ], + "type": "inline_equation", + "content": "\\tau_{k}" + }, + { + "bbox": [ + 54, + 216, + 295, + 344 + ], + "type": "text", + "content": " is a different task and " + }, + { + "bbox": [ + 54, + 216, + 295, + 344 + ], + "type": "inline_equation", + "content": "a_{m}^{(k)}" + }, + { + "bbox": [ + 54, + 216, + 295, + 344 + ], + "type": "text", + "content": " is its next pending action. For example, after \"turning on the kettle to boil water\" in a tea-making task, it is plausible to switch to a new task during the waiting time. Conversely, after \"scooping jelly from a jar\" in a pinwheel-making task, it makes sense to continue with the next action \"spread jelly on tortilla\" rather than switching tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 345, + 296, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 345, + 296, + 429 + ], + "spans": [ + { + "bbox": [ + 54, + 345, + 296, + 429 + ], + "type": "text", + "content": "During training, we randomly sample potential task switch points at the end of action segments and use the LLM to decide whether to continue or switch tasks. For more than two tasks, we first randomly select one unfinished task as the candidate for switching. We then blend the sequence of action segments based on the LLM's decisions, resulting in a new multi-task sequence of actions and frames." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 435, + 244, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 244, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 244, + 448 + ], + "type": "text", + "content": "3.4. Segment Boundary Learning (SBL)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 452, + 295, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 452, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 54, + 452, + 295, + 571 + ], + "type": "text", + "content": "The multi-task videos synthesized by MSB exhibit abrupt transitions at boundaries where we switch tasks, which disrupt the temporal flow and negatively impact the model's performance. Therefore, we propose the SBL module that learns to recover the feature representation of a frame " + }, + { + "bbox": [ + 54, + 452, + 295, + 571 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 54, + 452, + 295, + 571 + ], + "type": "text", + "content": " using its nearby frames, excluding frames immediately adjacent to " + }, + { + "bbox": [ + 54, + 452, + 295, + 571 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 452, + 295, + 571 + ], + "type": "text", + "content": " to avoid trivial learning. This approach allows us to generate features for frames where we concatenate segments from different videos, thus enhancing the temporal smoothness between action segments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 572, + 296, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 296, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 296, + 608 + ], + "type": "text", + "content": "Following prior works in TAS [17, 45, 49, 84], we first extract the framewise feature representation " + }, + { + "bbox": [ + 55, + 572, + 296, + 608 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 55, + 572, + 296, + 608 + ], + "type": "text", + "content": " using an I3D feature extractor [7] applied to a sliding window " + }, + { + "bbox": [ + 55, + 572, + 296, + 608 + ], + "type": "inline_equation", + "content": "V_{t}" + }, + { + "bbox": [ + 55, + 572, + 296, + 608 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 145, + 615, + 295, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 615, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 145, + 615, + 295, + 628 + ], + "type": "interline_equation", + "content": "f _ {t} = \\mathrm {I 3 D} \\left(V _ {t}\\right). \\tag {1}", + "image_path": "82b867fda7732f56fbbb9b30e53cfd85f07f6707235dac5429317b38b6729c93.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 635, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 296, + 694 + ], + "type": "text", + "content": "Then, for a given frame " + }, + { + "bbox": [ + 55, + 635, + 296, + 694 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 635, + 296, + 694 + ], + "type": "text", + "content": ", we consider its neighboring frames from the intervals " + }, + { + "bbox": [ + 55, + 635, + 296, + 694 + ], + "type": "inline_equation", + "content": "[t - \\delta - w, t - \\delta]" + }, + { + "bbox": [ + 55, + 635, + 296, + 694 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 635, + 296, + 694 + ], + "type": "inline_equation", + "content": "[t + \\delta, t + \\delta + w]" + }, + { + "bbox": [ + 55, + 635, + 296, + 694 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 635, + 296, + 694 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 635, + 296, + 694 + ], + "type": "text", + "content": " defines the window size and " + }, + { + "bbox": [ + 55, + 635, + 296, + 694 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 635, + 296, + 694 + ], + "type": "text", + "content": " specifies the exclusion margin around " + }, + { + "bbox": [ + 55, + 635, + 296, + 694 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 635, + 296, + 694 + ], + "type": "text", + "content": ". We train a reconstruction function " + }, + { + "bbox": [ + 55, + 635, + 296, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{SBL}}" + }, + { + "bbox": [ + 55, + 635, + 296, + 694 + ], + "type": "text", + "content": " to predict the feature " + }, + { + "bbox": [ + 55, + 635, + 296, + 694 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 55, + 635, + 296, + 694 + ], + "type": "text", + "content": " from its surrounding frames," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 97, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 97, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\bar {f} _ {t} = \\mathcal {F} _ {\\mathrm {S B L}} \\left(f _ {t - \\delta - w: t - \\delta}, f _ {t + \\delta : t + \\delta + w}\\right), \\tag {2}", + "image_path": "f2a78df740e5a6c045df3b4ee6b18e1da1b86d464bf822b161a93a5876f7e683.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 217, + 555, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 265 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 265 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 217, + 555, + 265 + ], + "type": "inline_equation", + "content": "f_{t - \\delta - w:t - \\delta}" + }, + { + "bbox": [ + 313, + 217, + 555, + 265 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 217, + 555, + 265 + ], + "type": "inline_equation", + "content": "f_{t + \\delta: t + \\delta + w}" + }, + { + "bbox": [ + 313, + 217, + 555, + 265 + ], + "type": "text", + "content": " denote the features of frames before and after " + }, + { + "bbox": [ + 313, + 217, + 555, + 265 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 217, + 555, + 265 + ], + "type": "text", + "content": " within the specified intervals. We optimize the reconstruction function by minimizing the reconstruction loss over non-boundary frames," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 383, + 271, + 553, + 298 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 271, + 553, + 298 + ], + "spans": [ + { + "bbox": [ + 383, + 271, + 553, + 298 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {S B L}} = \\sum_ {t \\notin \\mathcal {B}} \\left\\| \\bar {f} _ {t} - f _ {t} \\right\\| ^ {2}, \\tag {3}", + "image_path": "3451c3ec6561ef437ac67f6fd97194653288761659ba6d54b3c01a51128f0f1b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 304, + 555, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 399 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 304, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 304, + 555, + 399 + ], + "type": "text", + "content": " is the set of boundary frames (which come from two different videos, therefore are not continuous or realistic and we should not learn from them). By learning this reconstruction function, the model can generate smooth feature representations at segment boundaries. During training, we replace the original features at boundary frames with the reconstructed features " + }, + { + "bbox": [ + 313, + 304, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\bar{f}_t" + }, + { + "bbox": [ + 313, + 304, + 555, + 399 + ], + "type": "text", + "content": " to enhance the temporal smoothness of the synthesized multi-task videos." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 406, + 547, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 547, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 547, + 419 + ], + "type": "text", + "content": "3.5. Dynamic Isolation of Video Elements (DIVE)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 424, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 555, + 604 + ], + "type": "text", + "content": "In real multi-task videos, the present objects and their layouts can significantly differ from those in single-task videos. For example, as shown in the top-right frame in Fig. 1(b), in a video of making both tea and pinwheels, we might add honey to a mug while a rolled tortilla is on the cutting board, which is typically unseen in single-task videos. However, the output of MSB, which concatenates different actions from different single-task videos, does not capture such realistic multi-task layouts. Observing that action-relevant objects remain consistent across single-task and multi-task videos, we propose a Dynamic Isolation of Video Elements approach to isolate video elements into foreground and background components, and develop a Foreground-Background Feature Composition module to synthesize more realistic multi-task videos for training." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": "Detecting Action-Relevant Objects. Given a frame, it is challenging to distinguish between foreground and background solely based on visual cues, especially in cluttered multi-task environments. The definition of \"foreground\" or \"relevant objects\" depends on the action being performed; thus, knowledge of the action context is essential. Fig. 3(b) illustrates how we obtain the foreground and background components. We first query an LLM to obtain a list of relevant objects for each action class " + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "inline_equation", + "content": "a \\in \\mathcal{C}" + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": ". For example," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19123" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": "the action \"transfer water to kettle\" would be associated with \"kettle\" and \"measuring cup\". Then, we use an open-vocabulary object detector, GroundingDINO [46], to detect these relevant objects along with hands (we add \"left hand\" and \"right hand\" as object names) within each frame. We define the foreground region as the union of detected bounding boxes. We generate foreground and background frames through Gaussian blurring [82] outside and inside this region, respectively. We empirically observed that Gaussian blurring performs slightly better than simply zero masking." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 194, + 296, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 296, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 296, + 231 + ], + "type": "text", + "content": "Foreground-Background Feature Composition (FBFC). After generating foreground and background frames, we extract I3D features," + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 83, + 236, + 295, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 236, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 83, + 236, + 295, + 251 + ], + "type": "interline_equation", + "content": "f _ {t} ^ {\\mathrm {f g}} = \\mathrm {I 3 D} (V _ {t} \\odot M _ {t} ^ {\\mathrm {f g}}), f _ {t} ^ {\\mathrm {b g}} = \\mathrm {I 3 D} (V _ {t} \\odot M _ {t} ^ {\\mathrm {b g}}), \\qquad (4)", + "image_path": "68b274e0859d0f0d91d0e1920bf00a3612d2e0bb5e8f0db2fe21e3831fab5e15.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 256, + 296, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 296, + 319 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 256, + 296, + 319 + ], + "type": "inline_equation", + "content": "f_{t}^{\\mathrm{fg}}" + }, + { + "bbox": [ + 55, + 256, + 296, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 256, + 296, + 319 + ], + "type": "inline_equation", + "content": "f_{t}^{\\mathrm{bg}}" + }, + { + "bbox": [ + 55, + 256, + 296, + 319 + ], + "type": "text", + "content": " are foreground and background frame features, respectively, and " + }, + { + "bbox": [ + 55, + 256, + 296, + 319 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 256, + 296, + 319 + ], + "type": "text", + "content": " is the Gaussian blurring operation with corresponding foreground masks " + }, + { + "bbox": [ + 55, + 256, + 296, + 319 + ], + "type": "inline_equation", + "content": "M_{t}^{\\mathrm{fg}}" + }, + { + "bbox": [ + 55, + 256, + 296, + 319 + ], + "type": "text", + "content": " and background masks " + }, + { + "bbox": [ + 55, + 256, + 296, + 319 + ], + "type": "inline_equation", + "content": "M_{t}^{\\mathrm{fg}}" + }, + { + "bbox": [ + 55, + 256, + 296, + 319 + ], + "type": "text", + "content": ". We train a decoder " + }, + { + "bbox": [ + 55, + 256, + 296, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 256, + 296, + 319 + ], + "type": "text", + "content": " to reconstruct the original feature from foreground and background features," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 325, + 295, + 341 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 325, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 138, + 325, + 295, + 341 + ], + "type": "interline_equation", + "content": "\\hat {f} _ {t} = \\mathcal {D} \\left(f _ {t} ^ {\\mathrm {f g}}, f _ {t} ^ {\\mathrm {b g}}\\right). \\tag {5}", + "image_path": "49674806f951c431532a5c7b3a9fd7d762953f72f5eb6c7cdd9dc3192574ad66.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 346, + 296, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 296, + 370 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 346, + 296, + 370 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 55, + 346, + 296, + 370 + ], + "type": "text", + "content": " denote the original feature in Eq. (1). We train the decoder by minimizing the composition loss," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 122, + 377, + 295, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 377, + 295, + 406 + ], + "spans": [ + { + "bbox": [ + 122, + 377, + 295, + 406 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {F B F C}} = \\sum_ {t} \\left\\| \\hat {f} _ {t} - f _ {t} \\right\\| ^ {2}. \\tag {6}", + "image_path": "24ae0c0eb2f15a96a5c17cee135d6bfaf37cfab86db92cff82ba046877d061fd.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 411, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 296, + 445 + ], + "type": "text", + "content": "To augment synthetic multi-task features, we mix the background features of the current frame with background features of frames from another task" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 121, + 453, + 295, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 453, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 121, + 453, + 295, + 469 + ], + "type": "interline_equation", + "content": "\\tilde {f} _ {t} ^ {\\mathrm {b g}} = \\beta f _ {t} ^ {\\mathrm {b g}} + (1 - \\beta) f _ {t ^ {\\prime}} ^ {\\mathrm {b g}}, \\tag {7}", + "image_path": "81a102565ffef91c1276cbd24a92c88b65608226272fc85e0fd55a76dd66e2c9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 477, + 296, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 477, + 296, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 477, + 296, + 539 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 477, + 296, + 539 + ], + "type": "inline_equation", + "content": "f_{t}^{\\mathrm{bg}}" + }, + { + "bbox": [ + 55, + 477, + 296, + 539 + ], + "type": "text", + "content": " is the background feature from the current frame, " + }, + { + "bbox": [ + 55, + 477, + 296, + 539 + ], + "type": "inline_equation", + "content": "f_{t'}^{\\mathrm{bg}}" + }, + { + "bbox": [ + 55, + 477, + 296, + 539 + ], + "type": "text", + "content": " is the background feature from another task, and " + }, + { + "bbox": [ + 55, + 477, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\beta \\in [0,1]" + }, + { + "bbox": [ + 55, + 477, + 296, + 539 + ], + "type": "text", + "content": " controls the mixing ratio. We then compose multi-task features from the foreground feature and the mixed background feature using the decoder" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 545, + 295, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 545, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 137, + 545, + 295, + 563 + ], + "type": "interline_equation", + "content": "\\tilde {f} _ {t} = \\mathcal {D} \\left(f _ {t} ^ {\\mathrm {f g}}, \\hat {f} _ {t} ^ {\\mathrm {b g}}\\right). \\tag {8}", + "image_path": "5bf7943d9936bd55133253903ea8661e8c05ed5287e80874d4661456eb4c2a8c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 569, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 296, + 594 + ], + "type": "text", + "content": "We replace the training data with these recomposed features, exposing the model to diverse background variations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 600, + 276, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 276, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 276, + 613 + ], + "type": "text", + "content": "3.6. Two-Stage Temporal Action Segmentation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 618, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 689 + ], + "type": "text", + "content": "To further enhance the robustness of our model to background variations during testing, we propose a two-stage action segmentation method incorporating a Foreground-Aware Action Refinement (FAAR) module. By focusing on foreground regions, FAAR mitigates the impact of background distractions and refines the initial TAS predictions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "During training, with ground-truth actions available, we obtain the action-relevant objects and detect the foreground" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "regions. We extract CLIP features [55] from the foreground frames and apply a lightweight TAS model to predict foreground-aware action probabilities based solely on the foreground frames, " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "p_t^{\\mathrm{fg}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": ", which are combined with the initial action predictions " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " output by the base TAS model," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 318, + 140, + 555, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 140, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 318, + 140, + 555, + 156 + ], + "type": "interline_equation", + "content": "p _ {t} ^ {\\mathrm {f g}} = \\mathcal {R} \\left(g \\left(v _ {t} \\odot m _ {t} ^ {\\mathrm {f g}}\\right)\\right), \\quad p _ {t} ^ {\\text {f i n a l}} = (1 - \\alpha) \\cdot p _ {t} + \\alpha \\cdot p _ {t} ^ {\\mathrm {f g}}, \\tag {9}", + "image_path": "93c81ac7d16dc6c80c3fad9e9aa28c583b2ee3b7eb99407627cc038c363d155f.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 161, + 553, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 161, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 161, + 553, + 209 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 161, + 553, + 209 + ], + "type": "inline_equation", + "content": "m_t^{\\mathrm{fg}}" + }, + { + "bbox": [ + 313, + 161, + 553, + 209 + ], + "type": "text", + "content": " is the binary foreground mask, " + }, + { + "bbox": [ + 313, + 161, + 553, + 209 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 161, + 553, + 209 + ], + "type": "text", + "content": " denotes CLIP feature extractor, " + }, + { + "bbox": [ + 313, + 161, + 553, + 209 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 313, + 161, + 553, + 209 + ], + "type": "text", + "content": " denotes the lightweight TAS model, " + }, + { + "bbox": [ + 313, + 161, + 553, + 209 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,1]" + }, + { + "bbox": [ + 313, + 161, + 553, + 209 + ], + "type": "text", + "content": " controls the weight of foreground-aware probabilities, and " + }, + { + "bbox": [ + 313, + 161, + 553, + 209 + ], + "type": "inline_equation", + "content": "p_t^{\\mathrm{final}}" + }, + { + "bbox": [ + 313, + 161, + 553, + 209 + ], + "type": "text", + "content": " is the refined action probabilities." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 210, + 555, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 210, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 555, + 321 + ], + "type": "text", + "content": "During inference, without ground-truth actions, we first select the top " + }, + { + "bbox": [ + 313, + 210, + 555, + 321 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 210, + 555, + 321 + ], + "type": "text", + "content": " predicted action classes " + }, + { + "bbox": [ + 313, + 210, + 555, + 321 + ], + "type": "inline_equation", + "content": "\\{\\hat{a}_t^{(1)},\\dots ,\\hat{a}_t^{(K)}\\}" + }, + { + "bbox": [ + 313, + 210, + 555, + 321 + ], + "type": "text", + "content": " based on the initial action probabilities " + }, + { + "bbox": [ + 313, + 210, + 555, + 321 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 313, + 210, + 555, + 321 + ], + "type": "text", + "content": " from the base TAS model, and retrieve their relevant objects. We then detect the corresponding foreground regions and obtain foreground-aware action probabilities " + }, + { + "bbox": [ + 313, + 210, + 555, + 321 + ], + "type": "inline_equation", + "content": "p_t^{\\mathrm{fg},(k)}" + }, + { + "bbox": [ + 313, + 210, + 555, + 321 + ], + "type": "text", + "content": " for each of the top " + }, + { + "bbox": [ + 313, + 210, + 555, + 321 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 210, + 555, + 321 + ], + "type": "text", + "content": " actions respectively. We compute the final prediction by combining the initial action probabilities with the foreground-aware ones," + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 349, + 331, + 553, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 331, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 349, + 331, + 553, + 363 + ], + "type": "interline_equation", + "content": "p _ {t} ^ {\\text {f i n a l}} = (1 - \\alpha) \\cdot p _ {t} + \\alpha \\cdot \\sum_ {k = 1} ^ {K} \\theta_ {t} ^ {(k)} p _ {t} ^ {\\mathrm {f g}, (k)}, \\tag {10}", + "image_path": "35fd6a45e31ee37cf283c3ac172e949ef8b21a38d4de750c045276b33dec758d.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 364, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 555, + 437 + ], + "type": "text", + "content": "where the weights " + }, + { + "bbox": [ + 313, + 364, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\theta_t^{(k)}" + }, + { + "bbox": [ + 313, + 364, + 555, + 437 + ], + "type": "text", + "content": " are proportional to the initial probabilities of the top- " + }, + { + "bbox": [ + 313, + 364, + 555, + 437 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 364, + 555, + 437 + ], + "type": "text", + "content": " actions, i.e., " + }, + { + "bbox": [ + 313, + 364, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\theta_t^{(k)} = \\frac{p_t\\left(\\hat{a}_t^{(k)}\\right)}{\\sum_{j=1}^K p_t\\left(\\hat{a}_t^{(j)}\\right)}" + }, + { + "bbox": [ + 313, + 364, + 555, + 437 + ], + "type": "text", + "content": ". This approach refines the predictions by focusing on the most probable actions and their relevant objects, enhancing robustness to background variations in multi-task scenarios." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 445, + 444, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 445, + 444, + 458 + ], + "spans": [ + { + "bbox": [ + 314, + 445, + 444, + 458 + ], + "type": "text", + "content": "3.7. Training and Inference" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 470, + 554, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 554, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 554, + 577 + ], + "type": "text", + "content": "Training. We use the synthetic multi-task videos generated by MSB for training. We train the SBL module " + }, + { + "bbox": [ + 313, + 470, + 554, + 577 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{SBL}}" + }, + { + "bbox": [ + 313, + 470, + 554, + 577 + ], + "type": "text", + "content": " via " + }, + { + "bbox": [ + 313, + 470, + 554, + 577 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SBL}}" + }, + { + "bbox": [ + 313, + 470, + 554, + 577 + ], + "type": "text", + "content": " in Eq. (3) and utilize the reconstructed features at segment boundaries to enhance temporal smoothness. Similarly, we train the FBFC decoder " + }, + { + "bbox": [ + 313, + 470, + 554, + 577 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 470, + 554, + 577 + ], + "type": "text", + "content": " via " + }, + { + "bbox": [ + 313, + 470, + 554, + 577 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{FBFC}}" + }, + { + "bbox": [ + 313, + 470, + 554, + 577 + ], + "type": "text", + "content": " in Eq. (6), and augment the training data by mixing background features from different tasks as described in Eqs. (7) and (8). We train the base TAS model with FAAR using the standard action loss. The total loss function " + }, + { + "bbox": [ + 313, + 470, + 554, + 577 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 470, + 554, + 577 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 353, + 588, + 553, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 588, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 353, + 588, + 553, + 601 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {a c t i o n}} + \\lambda_ {\\mathrm {S B L}} \\mathcal {L} _ {\\mathrm {S B L}} + \\lambda_ {\\mathrm {F B F C}} \\mathcal {L} _ {\\mathrm {F B F C}}, \\tag {11}", + "image_path": "68b2a0d241f0108c2f68b917c3941758649e06daa89483d47570238b53ebe0a8.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 611, + 554, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 554, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 554, + 647 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 611, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{action}}" + }, + { + "bbox": [ + 313, + 611, + 554, + 647 + ], + "type": "text", + "content": " includes the cross-entropy loss and smoothing loss for TAS, applied to the final action probabilities output by the FAAR module. " + }, + { + "bbox": [ + 313, + 611, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{SBL}}" + }, + { + "bbox": [ + 313, + 611, + 554, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 611, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{FBFC}}" + }, + { + "bbox": [ + 313, + 611, + 554, + 647 + ], + "type": "text", + "content": " are the loss weights." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": "Inference. During inference, we do not use the MSB, SBL and FBFC modules, as they are designed to augment training. We first obtain initial action predictions from the base TAS model, and then apply the FAAR module, as specified in Eq. (10), to derive the final action predictions." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19124" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 284, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 284, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 284, + 85 + ], + "type": "text", + "content": "3.8. Incorporating Unlabeled Multi-Task Videos" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 173 + ], + "type": "text", + "content": "Our proposed framework so far trains an MT-TAS model using only single-task videos. However, we additionally investigate the potential benefits of incorporating unlabeled multi-task videos, when such videos are available. Since annotating long multi-task videos is labor-intensive and time-consuming, leveraging unlabeled multi-task videos offers a practical way to enhance the model performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 173, + 296, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 173, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 173, + 296, + 316 + ], + "type": "text", + "content": "We formulate this as a domain adaptation problem [9, 10, 50], where labeled single-task and synthetic multi-task videos constitute the source domain, while unlabeled real multi-task videos form the target domain. We aim to learn domain-invariant feature representations through adversarial training, achieved by adding a domain classifier with a Gradient Reversal Layer (GRL) [21] after the base TAS model's early feature layer. The model is trained jointly with action labels on the source domain and domain labels on both domains. During training, the GRL reverses the gradient flow, encouraging indistinguishable feature distributions across domains. Our overall training loss is" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 60, + 321, + 294, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 321, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 60, + 321, + 294, + 334 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {a c t i o n}} + \\lambda_ {\\mathrm {S B L}} \\mathcal {L} _ {\\mathrm {S B L}} + \\lambda_ {\\mathrm {F B F C}} \\mathcal {L} _ {\\mathrm {F B F C}} + \\lambda_ {\\text {d o m a i n}} \\mathcal {L} _ {\\text {d o m a i n}}, \\tag {12}", + "image_path": "905e380b6d8377be93b89ace420df77067df2041c2f4c78f805cb337615b1547.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 337, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 295, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 337, + 295, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{domain}}" + }, + { + "bbox": [ + 55, + 337, + 295, + 361 + ], + "type": "text", + "content": " is the framewise binary domain classification loss, and " + }, + { + "bbox": [ + 55, + 337, + 295, + 361 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{domain}}" + }, + { + "bbox": [ + 55, + 337, + 295, + 361 + ], + "type": "text", + "content": " is the loss weight." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 370, + 151, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 151, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 151, + 383 + ], + "type": "text", + "content": "4. MEKA Dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 390, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 296, + 534 + ], + "type": "text", + "content": "While our proposed framework can be trained using only single-task videos, we require an MT-TAS benchmark for evaluation. To this end, we introduce the Multi-task Ego-centric Kitchen Activities (MEKA) dataset, which extends the EgoPER dataset [36] to multi-task scenarios. EgoPER is an egocentric procedural video dataset designed for temporal action segmentation and error detection. It contains single-task videos of users performing kitchen activities while wearing an egocentric camera (HoloLens 2). The dataset includes five cooking tasks: making pinwheels, quesadillas, oatmeal, coffee, and tea, each associated with a predefined task graph." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 534, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 640 + ], + "type": "text", + "content": "To create realistic multi-task scenarios, we generated multi-task transcripts by interleaving action sequences from multiple tasks while ensuring logical coherence within the task graphs. Participants equipped with a HoloLens 2 camera followed these transcripts during recording. While we used similar recording equipment and kitchen tools as in the EgoPER dataset to maintain consistency, variations in the environments introduce additional challenges to the evaluation benchmark." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 642, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 712 + ], + "type": "text", + "content": "After data collection, we annotated each video by assigning framewise labels for both actions and tasks. In total, MEKA consists of 100 videos for 12 hours of footage. In each video, participants perform two or three cooking tasks in an interleaved manner. See the supplementary materials for more details about the dataset. We will make the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": "MEKA dataset publicly available upon publication, including videos and framewise annotations, to facilitate further research in multi-task temporal action segmentation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 122, + 397, + 136 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 122, + 397, + 136 + ], + "spans": [ + { + "bbox": [ + 313, + 122, + 397, + 136 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 143, + 432, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 143, + 432, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 143, + 432, + 156 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 161, + 555, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 161, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 161, + 555, + 340 + ], + "type": "text", + "content": "Datasets. Since MT-TAS is a new task, existing datasets cannot be used for its evaluation. Consequently, we evaluate our approach using single-task videos from EgoPER [36] for training and our collected multi-task videos from MEKA for testing. EgoPER contains 213 normal single-task egocentric videos and MEKA contains 100 multi-task videos for the five recipes in EgoPER. In total, it has 51 classes of action steps. We were also able to adapt the EGTEA dataset [43] for multi-task evaluation. We manually select actions that appear exclusively in either the first or second half of each video, treating these segments as single-task videos, and consider the full videos with interleaved actions as multi-task videos. The supplementary materials provide additional details and the experimental results on the adapted EGTEA dataset." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 342, + 556, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 556, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 556, + 533 + ], + "type": "text", + "content": "Model Architecture. We evaluate our method with three base TAS models: two offline models (MSTCN [17] and FACT [49]) and one online model (ProTAS [65]). Online TAS differs from offline TAS in that it only uses the frames up to the current moment for prediction. We implement the SBL module " + }, + { + "bbox": [ + 313, + 342, + 556, + 533 + ], + "type": "inline_equation", + "content": "(\\mathcal{F}_{\\mathrm{SBL}})" + }, + { + "bbox": [ + 313, + 342, + 556, + 533 + ], + "type": "text", + "content": " by processing interval features through two convolutional layers, followed by temporal global average pooling. The Decoder " + }, + { + "bbox": [ + 313, + 342, + 556, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 342, + 556, + 533 + ], + "type": "text", + "content": " processes foreground and background features through parallel two-layer MLPs, each reducing feature dimension by half, followed by a linear layer that reconstructs the original feature representation from the concatenated features. We employ a three-layer TCN [17] with a hidden dimension of 64 for FAAR " + }, + { + "bbox": [ + 313, + 342, + 556, + 533 + ], + "type": "inline_equation", + "content": "(\\mathcal{R})" + }, + { + "bbox": [ + 313, + 342, + 556, + 533 + ], + "type": "text", + "content": ". For experiments with domain adaptation, we insert a GRL after the feature layer of the initial stage of TAS models, and use a two-layer MLP for binary domain classification." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": "Implementation Details. Following standard practice [17, 49, 65, 81, 84], we extract 2048-dimensional I3D features [7] pretrained on Kinetics [7] as video representations. We evaluate performance using standard TAS metrics: frame-wise accuracy (Acc), frame-wise accuracy excluding background frames (Acc-bg), segment-wise edit score (Edit), and segment-wise F1 scores at multiple IoU thresholds (F1@10,25,50). We adopt the default configurations from MSTCN [17], FACT [49], and ProTAS [65] using their official source codes. We train all models using Adam optimizer with a learning rate of 0.0005. We set the loss weights as " + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{SBL}} = \\lambda_{\\mathrm{FBFC}} = \\lambda_{\\mathrm{domain}} = 1" + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": ". For the SBL module, we define boundary frames as the 10 frames centered at the points where segments from different videos are concatenated, and set window size " + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": " as 5 and margin" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19125" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 70, + 547, + 170 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 547, + 170 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 547, + 170 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 547, + 170 + ], + "type": "table", + "html": "
MSBSBLFBFCFAARTAS model: MSTCN [17]TAS model: FACT [49]
AccAcc-bgEditF1@{10,25,50}AccAcc-bgEditF1@{10,25,50}
baseline62.349.150.152.649.339.753.741.442.944.641.631.7
✓67.863.764.268.966.856.072.867.873.173.170.058.4
✓✓68.865.964.770.868.858.173.670.474.576.173.662.3
✓✓✓73.764.469.573.069.958.873.871.374.479.676.765.5
✓✓✓✓75.772.174.979.777.667.475.772.876.081.279.369.9
", + "image_path": "c564bd9dfc416d7b87ced4b04c424de16e6b98cac0088df7ad5e7524d64d879b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 191, + 293, + 276 + ], + "blocks": [ + { + "bbox": [ + 156, + 172, + 453, + 183 + ], + "lines": [ + { + "bbox": [ + 156, + 172, + 453, + 183 + ], + "spans": [ + { + "bbox": [ + 156, + 172, + 453, + 183 + ], + "type": "text", + "content": "Table 1. Offline multi-task temporal action segmentation performance on MEKA." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 191, + 293, + 276 + ], + "lines": [ + { + "bbox": [ + 59, + 191, + 293, + 276 + ], + "spans": [ + { + "bbox": [ + 59, + 191, + 293, + 276 + ], + "type": "table", + "html": "
MSBSBLFBFCFAARAccEditF1@{10,25,50}
baseline49.124.624.120.012.4
✓51.130.726.522.112.8
✓✓51.133.129.324.514.6
✓✓✓55.237.532.827.118.3
✓✓✓✓67.854.854.649.836.7
", + "image_path": "ee1a6cc12ff026ca31f80a58a099142861da7e77d621b58ed88820d8aebae381.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 305, + 295, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 305, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 54, + 305, + 295, + 449 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 54, + 305, + 295, + 449 + ], + "type": "text", + "content": " as 2. For Foreground-Background Feature Composition (FBFC), we select " + }, + { + "bbox": [ + 54, + 305, + 295, + 449 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 54, + 305, + 295, + 449 + ], + "type": "text", + "content": " of frames and augment them with Eq. (7) by randomly sampling " + }, + { + "bbox": [ + 54, + 305, + 295, + 449 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 54, + 305, + 295, + 449 + ], + "type": "text", + "content": " from [0.5, 1] during training. See Sec. 5.2 for the qualitative analysis on the value of " + }, + { + "bbox": [ + 54, + 305, + 295, + 449 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 54, + 305, + 295, + 449 + ], + "type": "text", + "content": ". For the FAAR module, we employ a two-stage training process. In the first stage, we train the model without FAAR by setting " + }, + { + "bbox": [ + 54, + 305, + 295, + 449 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 54, + 305, + 295, + 449 + ], + "type": "text", + "content": " in Eq. (9). In the second stage, we incorporate FAAR and randomly sample " + }, + { + "bbox": [ + 54, + 305, + 295, + 449 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 305, + 295, + 449 + ], + "type": "text", + "content": " from [0, 1] during training. For inference, we set " + }, + { + "bbox": [ + 54, + 305, + 295, + 449 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 54, + 305, + 295, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 305, + 295, + 449 + ], + "type": "inline_equation", + "content": "\\alpha = 0.3" + }, + { + "bbox": [ + 54, + 305, + 295, + 449 + ], + "type": "text", + "content": " in Eq. (10). We include ablation studies on the values of " + }, + { + "bbox": [ + 54, + 305, + 295, + 449 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 305, + 295, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 305, + 295, + 449 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 305, + 295, + 449 + ], + "type": "text", + "content": " in Sec. 5.2. Please refer to the supplementary materials for more implementation details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 456, + 180, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 180, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 180, + 469 + ], + "type": "text", + "content": "5.2. Experimental Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 474, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 474, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 474, + 295, + 714 + ], + "type": "text", + "content": "Offline Temporal Action Segmentation. Tab. 1 presents the performance of our approach on the MEKA dataset using two baseline offline TAS models: MSTCN [17] and FACT [49]. We start with the baseline models trained on single-task videos without any of our proposed modules (first row). We then progressively incorporate the Multi-task Sequence Blending (MSB), Segment Boundary Learning (SBL), Foreground-Background Feature Composition (FBFC), and Foreground-Aware Action Refinement (FAAR) modules to assess their individual contributions. In the first row, MSTCN achieves an accuracy of " + }, + { + "bbox": [ + 56, + 474, + 295, + 714 + ], + "type": "inline_equation", + "content": "62.3\\%" + }, + { + "bbox": [ + 56, + 474, + 295, + 714 + ], + "type": "text", + "content": " and an F1@50 score of " + }, + { + "bbox": [ + 56, + 474, + 295, + 714 + ], + "type": "inline_equation", + "content": "39.7\\%" + }, + { + "bbox": [ + 56, + 474, + 295, + 714 + ], + "type": "text", + "content": ", while FACT attains " + }, + { + "bbox": [ + 56, + 474, + 295, + 714 + ], + "type": "inline_equation", + "content": "53.7\\%" + }, + { + "bbox": [ + 56, + 474, + 295, + 714 + ], + "type": "text", + "content": " accuracy and " + }, + { + "bbox": [ + 56, + 474, + 295, + 714 + ], + "type": "inline_equation", + "content": "31.7\\%" + }, + { + "bbox": [ + 56, + 474, + 295, + 714 + ], + "type": "text", + "content": " F1@50, indicating that the baseline models struggle to generalize to multi-task videos when trained solely on single-task data. Incorporating MSB significantly improves performance for both models, especially for FACT, due to its explicit modeling of video transcripts during training. Adding SBL further boosts performance. With MSB and SBL, MSTCN's accuracy improves by " + }, + { + "bbox": [ + 56, + 474, + 295, + 714 + ], + "type": "inline_equation", + "content": "6.5\\%" + }, + { + "bbox": [ + 56, + 474, + 295, + 714 + ], + "type": "text", + "content": " and F1@50 by " + }, + { + "bbox": [ + 56, + 474, + 295, + 714 + ], + "type": "inline_equation", + "content": "18.4\\%" + }, + { + "bbox": [ + 56, + 474, + 295, + 714 + ], + "type": "text", + "content": ", while FACT's accuracy" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 191, + 553, + 323 + ], + "blocks": [ + { + "bbox": [ + 70, + 279, + 280, + 291 + ], + "lines": [ + { + "bbox": [ + 70, + 279, + 280, + 291 + ], + "spans": [ + { + "bbox": [ + 70, + 279, + 280, + 291 + ], + "type": "text", + "content": "Table 2. Online temporal action segmentation on MEKA." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 191, + 553, + 323 + ], + "lines": [ + { + "bbox": [ + 317, + 191, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 317, + 191, + 553, + 323 + ], + "type": "table", + "html": "
Training SetupAccEditF1@{10,25,50}
MSTCN [17]Offline Temporal Action Segmentation
LMT78.075.978.376.670.4
LST&LMT80.684.784.582.574.6
Oursw/o DA75.774.979.777.667.4
w/ DA76.6+0.977.6+2.782.0+2.380.6+3.072.2+4.8
+LMT82.787.389.788.582.4
ProTAS [65]Online Temporal Action Segmentation
LMT61.541.137.232.822.0
LST&LMT64.549.839.434.522.7
Oursw/o DA67.854.854.649.836.7
w/ DA69.1+1.356.9+2.157.5+2.952.0+2.238.8+2.1
+LMT74.970.667.262.549.3
", + "image_path": "0040405f1705a89fedc47e2cf25308e0e97a16a290b211ba6bd23f27cdb76022.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 327, + 553, + 347 + ], + "lines": [ + { + "bbox": [ + 313, + 327, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 553, + 347 + ], + "type": "text", + "content": "Table 3. The effect of domain adaptation using unlabeled multi-task videos and comparison with methods using labeled multi-task videos." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 312, + 365, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 365, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 312, + 365, + 555, + 449 + ], + "type": "text", + "content": "improves by " + }, + { + "bbox": [ + 312, + 365, + 555, + 449 + ], + "type": "inline_equation", + "content": "19.9\\%" + }, + { + "bbox": [ + 312, + 365, + 555, + 449 + ], + "type": "text", + "content": " and F1@50 by " + }, + { + "bbox": [ + 312, + 365, + 555, + 449 + ], + "type": "inline_equation", + "content": "30.6\\%" + }, + { + "bbox": [ + 312, + 365, + 555, + 449 + ], + "type": "text", + "content": ". Incorporating FBFC further improves the accuracy by " + }, + { + "bbox": [ + 312, + 365, + 555, + 449 + ], + "type": "inline_equation", + "content": "4.9\\%" + }, + { + "bbox": [ + 312, + 365, + 555, + 449 + ], + "type": "text", + "content": " on MSTCN and the F1@50 by " + }, + { + "bbox": [ + 312, + 365, + 555, + 449 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 312, + 365, + 555, + 449 + ], + "type": "text", + "content": " on FACT. Finally, integrating all proposed modules leads to the highest performance, resulting in an F1@50 score of " + }, + { + "bbox": [ + 312, + 365, + 555, + 449 + ], + "type": "inline_equation", + "content": "67.4\\%" + }, + { + "bbox": [ + 312, + 365, + 555, + 449 + ], + "type": "text", + "content": " for MSTCN and " + }, + { + "bbox": [ + 312, + 365, + 555, + 449 + ], + "type": "inline_equation", + "content": "69.9\\%" + }, + { + "bbox": [ + 312, + 365, + 555, + 449 + ], + "type": "text", + "content": " for FACT. Overall, our approach effectively enhances the capability of standard TAS models to handle multi-task videos." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 449, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 449, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 312, + 449, + 556, + 628 + ], + "type": "text", + "content": "Online Temporal Action Segmentation. We evaluate the online TAS performance of our approach in Tab. 2, using ProTAS [65] as the base TAS model. Without access to future frames, the online model relies solely on past and current frames, making it challenging to anticipate action transitions and accurately segment boundaries. This limitation results in a noticeable performance drop compared to the offline models, echoing the difficulties of online MT-TAS. However, our proposed modules still achieve a notable performance boost. When all modules are combined, the model increases accuracy by " + }, + { + "bbox": [ + 312, + 449, + 556, + 628 + ], + "type": "inline_equation", + "content": "18.7\\%" + }, + { + "bbox": [ + 312, + 449, + 556, + 628 + ], + "type": "text", + "content": " and F1@50 by " + }, + { + "bbox": [ + 312, + 449, + 556, + 628 + ], + "type": "inline_equation", + "content": "24.3\\%" + }, + { + "bbox": [ + 312, + 449, + 556, + 628 + ], + "type": "text", + "content": " compared to the baseline. Particularly, the integration of FAAR improves the accuracy by " + }, + { + "bbox": [ + 312, + 449, + 556, + 628 + ], + "type": "inline_equation", + "content": "12.6\\%" + }, + { + "bbox": [ + 312, + 449, + 556, + 628 + ], + "type": "text", + "content": " and all F1 scores by nearly " + }, + { + "bbox": [ + 312, + 449, + 556, + 628 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 312, + 449, + 556, + 628 + ], + "type": "text", + "content": ", highlighting the value of foreground-aware refinements in scenarios without access to future frames." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 630, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 630, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 630, + 556, + 714 + ], + "type": "text", + "content": "Domain Adaptation with Unlabeled Multi-Task Videos. In Tab. 3, we evaluate our domain adaptation approach using unlabeled multi-task videos, as introduced in Sec. 3.8. Domain adaptation consistently improves performance across all metrics, increasing F1@50 by " + }, + { + "bbox": [ + 312, + 630, + 556, + 714 + ], + "type": "inline_equation", + "content": "4.8\\%" + }, + { + "bbox": [ + 312, + 630, + 556, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 312, + 630, + 556, + 714 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 312, + 630, + 556, + 714 + ], + "type": "text", + "content": " in offline and online settings, respectively. To thoroughly assess our method, we compare against two baseline se" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19126" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 72, + 226, + 194 + ], + "blocks": [ + { + "bbox": [ + 70, + 72, + 226, + 194 + ], + "lines": [ + { + "bbox": [ + 70, + 72, + 226, + 194 + ], + "spans": [ + { + "bbox": [ + 70, + 72, + 226, + 194 + ], + "type": "image", + "image_path": "cd67e2174538760dc4ad57c12e056e00f6cafa5b5a3bc11836a6d91db026c76e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 227, + 72, + 381, + 194 + ], + "blocks": [ + { + "bbox": [ + 227, + 72, + 381, + 194 + ], + "lines": [ + { + "bbox": [ + 227, + 72, + 381, + 194 + ], + "spans": [ + { + "bbox": [ + 227, + 72, + 381, + 194 + ], + "type": "image", + "image_path": "c0e86034b9c548275726504e3c6c29173a7b73eed56da81c37a6534b511e4fd9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 381, + 72, + 538, + 198 + ], + "blocks": [ + { + "bbox": [ + 381, + 72, + 538, + 198 + ], + "lines": [ + { + "bbox": [ + 381, + 72, + 538, + 198 + ], + "spans": [ + { + "bbox": [ + 381, + 72, + 538, + 198 + ], + "type": "image", + "image_path": "5b7ee7fa9602c134e99b580adea4ef2bd8b115f5399e9dc22822167b24a51fe5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 112, + 200, + 496, + 211 + ], + "lines": [ + { + "bbox": [ + 112, + 200, + 496, + 211 + ], + "spans": [ + { + "bbox": [ + 112, + 200, + 496, + 211 + ], + "type": "text", + "content": "Figure 4. Ablation Studies on FAAR: (a) comparison of different inputs; (b) effects of " + }, + { + "bbox": [ + 112, + 200, + 496, + 211 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 112, + 200, + 496, + 211 + ], + "type": "text", + "content": "; (c) effects of " + }, + { + "bbox": [ + 112, + 200, + 496, + 211 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 112, + 200, + 496, + 211 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 232, + 297, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 232, + 297, + 460 + ], + "spans": [ + { + "bbox": [ + 54, + 232, + 297, + 460 + ], + "type": "text", + "content": "tups using 3-fold cross-validation on multi-task videos: 1) the model is trained on labeled multi-task (LMT) videos; 2) the model is trained on both labeled single-task and labeled multi-task (LST&LMT) videos. Note that these baselines use labeled multi-task data, while our method employs only unlabeled multi-task videos during training. In offline TAS, our model with domain adaptation outperforms the LMT baseline and achieves comparable results to the LST&LMT baseline. In the more challenging online setting, our models surpass both baselines, demonstrating the ability of our modules to enhance generalization and handle the data-intensive demands of online TAS. We also extend our method by incorporating labeled multi-task videos (+LMT) to provide an upper bound for domain adaptation. The performance gap reveals room for improvement in unsupervised domain adaptation. Furthermore, the performance gains from LST&LMT to Ours+LMT demonstrate that our modules provide complementary benefits even when labeled multi-task data is available." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 460, + 298, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 298, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 298, + 700 + ], + "type": "text", + "content": "Ablation Studies on FAAR. In Fig. 4, we conduct ablation studies to evaluate different inputs of FAAR and analyze the effects of " + }, + { + "bbox": [ + 55, + 460, + 298, + 700 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 460, + 298, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 460, + 298, + 700 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 460, + 298, + 700 + ], + "type": "text", + "content": " in Eq. (10) during inference. First, we change the module's input from foreground frames to background frames or full images. Background frames perform poorly while foreground frames outperform full images, showing the importance of action-relevant foreground information. We then examine how varying " + }, + { + "bbox": [ + 55, + 460, + 298, + 700 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 460, + 298, + 700 + ], + "type": "text", + "content": ", the number of top predicted action classes considered, and " + }, + { + "bbox": [ + 55, + 460, + 298, + 700 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 460, + 298, + 700 + ], + "type": "text", + "content": ", the weighting factor between initial predictions and foreground-aware predictions, influences the performance. The model's performance gradually improves with increasing " + }, + { + "bbox": [ + 55, + 460, + 298, + 700 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 460, + 298, + 700 + ], + "type": "text", + "content": ", reaching a plateau when " + }, + { + "bbox": [ + 55, + 460, + 298, + 700 + ], + "type": "inline_equation", + "content": "K \\geq 3" + }, + { + "bbox": [ + 55, + 460, + 298, + 700 + ], + "type": "text", + "content": ". As " + }, + { + "bbox": [ + 55, + 460, + 298, + 700 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 460, + 298, + 700 + ], + "type": "text", + "content": " increases from 0 to 1, the model performs best between 0.3 and 0.4. Notably, while the accuracies at " + }, + { + "bbox": [ + 55, + 460, + 298, + 700 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 55, + 460, + 298, + 700 + ], + "type": "text", + "content": " (base TAS predictions only) and " + }, + { + "bbox": [ + 55, + 460, + 298, + 700 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 55, + 460, + 298, + 700 + ], + "type": "text", + "content": " (foreground-aware predictions only) are similar, the Edit and F1@50 scores for foreground-only predictions are significantly lower. This suggests that FAAR emphasizes individual frame accuracy but may lack the temporal consistency needed for segment-wise metrics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": "Qualitative Analysis of FBFC. In Fig. 5, we qualitatively" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 322, + 230, + 547, + 346 + ], + "blocks": [ + { + "bbox": [ + 322, + 230, + 547, + 346 + ], + "lines": [ + { + "bbox": [ + 322, + 230, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 322, + 230, + 547, + 346 + ], + "type": "image", + "image_path": "70d9736337e0a31e03bbbc8769e4c0cedf2830271b2cd3e5c21f9f83fa5759a1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 350, + 555, + 370 + ], + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 370 + ], + "type": "text", + "content": "Figure 5. Retrieved nearest multi-task videos by different ratios with FBFC. For simplicity, only middle frames of video clips are displayed." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 392, + 555, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 583 + ], + "type": "text", + "content": "analyze how FBFC works with different values of the mixing ratio " + }, + { + "bbox": [ + 313, + 392, + 555, + 583 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 392, + 555, + 583 + ], + "type": "text", + "content": ". We sample two clips from single-task videos, extract foreground frames from one clip and background frames from another, and then recompose these features using FBFC via Eqs. (7) and (8). To assess the decomposition quality, we retrieve the nearest clip from multi-task testing videos based on feature similarity. With a lower mixing ratio (" + }, + { + "bbox": [ + 313, + 392, + 555, + 583 + ], + "type": "inline_equation", + "content": "\\beta = 0.2" + }, + { + "bbox": [ + 313, + 392, + 555, + 583 + ], + "type": "text", + "content": "), the recomposed features are more heavily influenced by background features, and the context of the retrieved frames predominantly match the background frames. Conversely, with a higher mixing ratio (" + }, + { + "bbox": [ + 313, + 392, + 555, + 583 + ], + "type": "inline_equation", + "content": "\\beta = 0.8" + }, + { + "bbox": [ + 313, + 392, + 555, + 583 + ], + "type": "text", + "content": "), the retrieved frames often depict the same action as the foreground frames while sharing some similarities with the context of the background frames. Based on these observations, we sample " + }, + { + "bbox": [ + 313, + 392, + 555, + 583 + ], + "type": "inline_equation", + "content": "\\beta \\in [0.5,1]" + }, + { + "bbox": [ + 313, + 392, + 555, + 583 + ], + "type": "text", + "content": " in training to ensure preserving action features while introducing background variations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 584, + 555, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 584, + 555, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 584, + 555, + 608 + ], + "type": "text", + "content": "See supplementary materials for more results, qualitative visualization, and discussion on complexity and limitations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 621, + 392, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 392, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 392, + 633 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "We proposed an MT-TAS framework to tackle the challenges of interleaved actions in multi-task scenarios. Our approach combines synthesized multi-task data, a two-stage TAS model, and the new MEKA benchmark dataset. We validated the effectiveness of our approach through extensive experiments in both offline and online MT-TAS." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19127" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "content": "This work is sponsored by ARPA-H (1AY2AX000062), DARPA PTG (HR00112220001), NSF (IIS-2115110) and ARO (W911NF2110276). Content does not necessarily reflect the position/policy of the Government." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 149, + 115, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 149, + 115, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 149, + 115, + 162 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 169, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 169, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 169, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 169, + 296, + 213 + ], + "type": "text", + "content": "[1] Hyemin Ahn and Dongheui Lee. Refining action segmentation with hierarchical video representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16302-16310, 2021. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 214, + 296, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 214, + 296, + 257 + ], + "spans": [ + { + "bbox": [ + 61, + 214, + 296, + 257 + ], + "type": "text", + "content": "[2] J. B. Alayrac, P. Bojanowski, N. Agrawal, J. Sivic, I. Laptev, and S. Lacoste-Julien. Unsupervised learning from narrated instruction videos. IEEE Conference on Computer Vision and Pattern Recognition, 2016. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 258, + 296, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 296, + 302 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 296, + 302 + ], + "type": "text", + "content": "[3] Kyungho Bae, Geo Ahn, Youngrae Kim, and Jinwoo Choi. Devias: Learning disentangled video representations of action and scene for holistic video understanding. In European Conference on Computer Vision, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 303, + 296, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 303, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 303, + 296, + 346 + ], + "type": "text", + "content": "[4] Emad Bahrami, Gianpiero Francesca, and Juergen Gall. How much temporal long-term context is needed for action segmentation? In International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 347, + 296, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 296, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 296, + 390 + ], + "type": "text", + "content": "[5] Siddhant Bansal, Chetan Arora, and C.V. Jawahar. My view is the best view: Procedure learning from egocentric videos. In European Conference on Computer Vision (ECCV), 2022. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 392, + 296, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 296, + 435 + ], + "type": "text", + "content": "[6] Nadine Behrmann, S. Alireza Golestaneh, Zico Kolter, Juergen Gall, and Mehdi Noroozi. Unified fully and timestamp supervised temporal action segmentation via sequence to sequence translation. In ECCV, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 437, + 296, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 437, + 296, + 469 + ], + "spans": [ + { + "bbox": [ + 62, + 437, + 296, + 469 + ], + "type": "text", + "content": "[7] J. Carreira and A. Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In IEEE Conference on Computer Vision and Pattern Recognition, 2017. 4, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 471, + 296, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 471, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 471, + 296, + 525 + ], + "type": "text", + "content": "[8] Chien-Yi Chang, De-An Huang, Yanan Sui, Li Fei-Fei, and Juan Carlos Niebles. D3tw: Discriminative differentiable dynamic time warping for weakly supervised action alignment and segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 526, + 296, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 526, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 62, + 526, + 296, + 581 + ], + "type": "text", + "content": "[9] Min-Hung Chen, Baopu Li, Yingze Bao, Ghassan Al-Regib, and Zsolt Kira. Action segmentation with joint self-supervised temporal domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9454–9463, 2020. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 582, + 295, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 582, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 57, + 582, + 295, + 647 + ], + "type": "text", + "content": "[10] Victor G. Turrisi da Costa, Giacomo Zara, Paolo Rota, Thiago Oliveira-Santos, Nicu Sebe, Vittorio Murino, and Elisa Ricci. Dual-head contrastive domain adaptation for video action recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1181-1190, 2022. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "type": "text", + "content": "[11] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision (IJCV), 130:33-55, 2022. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[12] Li Ding and Chenliang Xu. Weakly-supervised action segmentation with iterative soft boundary assignment. IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 554, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 184 + ], + "type": "text", + "content": "[13] Shuangrui Ding, Maomao Li, Tianyu Yang, Rui Qian, Haohang Xu, Qingyi Chen, Jue Wang, and Hongkai Xiong. Motion-aware contrastive video representation learning via foreground-background merging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9716-9726, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 186, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 218 + ], + "type": "text", + "content": "[14] G. Donahue and E. Elhamifar. Learning to predict activity progress by self-supervised video alignment. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 219, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 252 + ], + "type": "text", + "content": "[15] E. Elhamifar and D. Huynh. Self-supervised multi-task procedure learning from instructional videos. European Conference on Computer Vision, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 254, + 553, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 286 + ], + "type": "text", + "content": "[16] E. Elhamifar and Z. Naing. Unsupervised procedure learning via joint dynamic summarization. International Conference on Computer Vision, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 287, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 341 + ], + "type": "text", + "content": "[17] Yazan Abu Farha and Jurgen Gall. Ms-tcn: Multi-stage temporal convolutional network for action segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3575-3584, 2019. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "type": "text", + "content": "[18] Alireza Fathi, Xiaofeng Ren, and James M Rehg. Learning to recognize objects in egocentric activities. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2011. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "text", + "content": "[19] Mohsen Fayyaz and Jurgen Gall. Sct: Set constrained temporal transformer for set supervised action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 434, + 554, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 554, + 487 + ], + "type": "text", + "content": "[20] Daniel Fried, Jean-Baptiste Alayrac, Phil Blunsom, Chris Dyer, Stephen Clark, and Aida Nematzadeh. Learning to segment actions from observation and narration. Annual Meeting of the Association for Computational Linguistics, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 489, + 554, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 554, + 521 + ], + "type": "text", + "content": "[21] Y. Ganin and V. Lempitsky. Unsupervised domain adaptation by backpropagation. International Conference on Machine Learning, 2015. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 523, + 554, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 554, + 578 + ], + "type": "text", + "content": "[22] Reza Ghoddoosian, Isht Dwivedi, Nakul Agarwal, Chiho Choi, and Behzad Dariush. Weakly-supervised online action segmentation in multi-view instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13780-13790, 2022. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 579, + 554, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 554, + 624 + ], + "type": "text", + "content": "[23] Reza Ghoddosian, Saif Sayed, and Vassilis Athitsos. Hierarchical modeling for task recognition and action segmentation in weakly-labeled instructional videos. IEEE Winter Conference on Applications of Computer Vision, 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 624, + 554, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 679 + ], + "type": "text", + "content": "[24] Reza Ghoddosian, Isht Dwivedi, Nakul Agarwal, and Behzad Dariush. Weakly-supervised action segmentation and unseen error detection in anomalous instructional videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10128-10138, 2023. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[25] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Mar" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19128" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 76, + 72, + 295, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 295, + 336 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 295, + 336 + ], + "type": "text", + "content": "tin, Tushar Nagarajan, Ilija Radosavovic, Santhosh K. Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Z. Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Christian Fuegen, Abraham Gebreselasie, Cristina Gonzalez, James M. Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolár, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran K. Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Yunyi Zhu, Pablo Arbeláez, David J. Crandall, Dima Damen, Giovanni Maria Farinella, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard A. Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the world in 3,000 hours of egocentric video. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18973-18990, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 338, + 295, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 295, + 415 + ], + "type": "text", + "content": "[26] Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, et al. Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19383-19400, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 417, + 294, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 417, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 417, + 294, + 460 + ], + "type": "text", + "content": "[27] Yifei Huang, Yusuke Sugano, and Yoichi Sato. Improving action segmentation via graph-based temporal reasoning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 463, + 294, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 463, + 294, + 540 + ], + "spans": [ + { + "bbox": [ + 56, + 463, + 294, + 540 + ], + "type": "text", + "content": "[28] Yifei Huang, Guo Chen, Jilan Xu, Mingfang Zhang, Lijin Yang, Baoqi Pei, Hongjie Zhang, Lu Dong, Yali Wang, Limin Wang, and Yu Qiao. Egoexolearn: A dataset for bridging asynchronous ego- and exo-centric view of procedural activities in real world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22072-22086, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 542, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 542, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 542, + 294, + 597 + ], + "type": "text", + "content": "[29] Yuchi Ishikawa, Seito Kasai, Yoshimitsu Aoki, and Hirokatsu Kataoka. Alleviating over-segmentation errors by detecting action boundaries. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2322-2331, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 599, + 294, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 599, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 56, + 599, + 294, + 643 + ], + "type": "text", + "content": "[30] Hilde Kuehne, Ali Arslan, and Thomas Serre. The language of actions: Recovering the syntax and semantics of goal-directed human activities. IEEE Conference on Computer Vision and Pattern Recognition, 2014. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 645, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 645, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 645, + 294, + 689 + ], + "type": "text", + "content": "[31] H. Kuehne, A. Arslan, and T. Serre. The language of actions: Recovering the syntax and semantics of goal-directed human. IEEE Conference on Computer Vision and Pattern Recognition, 2014. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[32] H. Kuehne, J. Gall, and T. Serre. An end-to-end generative framework for video segmentation and recognition." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "IEEE Winter Conference on Applications of Computer Vision, 2016. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[33] Anna Kukleva, Hilde Kuehne, Fadime Sener, and Jurgen Gall. Unsupervised learning of action classes with continuous temporal embedding. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 141, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 206 + ], + "type": "text", + "content": "[34] Sateesh Kumar, Sanjay Haresh, Awais Ahmed, Andrey Konin, M. Zeeshan Zia, and Quoc-Huy Tran. Unsupervised action segmentation by joint representation learning and online clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20174-20185, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[35] C. Lea, M. D. Flynn, R. Vidal, A. Reiter, and G. D. Hager. Temporal convolutional networks for action segmentation and detection. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 253, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 296 + ], + "type": "text", + "content": "[36] S. Lee, Z. Lu, Z. Zhang, M. Hoai, and E. Elhamifar. Error detection in egocentric procedural task videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "type": "text", + "content": "[37] Haoxin Li, Yuan Liu, Hanwang Zhang, and Boyang Li. Mitigating and evaluating static bias of action representations in the background and the foreground. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19911-19923, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "type": "text", + "content": "[38] Jun Li and Sinisa Todorovic. Set-constrained viterbi for set-supervised action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 388, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 421 + ], + "type": "text", + "content": "[39] J. Li and S. Todorovic. Anchor-constrained viterbi for set-supervised action segmentation. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 422, + 553, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 455 + ], + "type": "text", + "content": "[40] J. Li, P. Lei, and S. Todorovic. Weakly supervised energy-based learning for action segmentation. International Conference on Computer Vision, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 456, + 553, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 509 + ], + "type": "text", + "content": "[41] M. Li, L. Chen, Y. Duarr, Z. Hu, J. Feng, J. Zhou, and J. Lu. Bridge-prompt: Towards ordinal action understanding in instructional videos. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 511, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 553, + 566 + ], + "type": "text", + "content": "[42] Shi-Jie Li, Yazan AbuFarha, Yun Liu, Ming-Ming Cheng, and Juergen Gall. Ms-tcn++: Multi-stage temporal convolutional network for action segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2020. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 567, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 600 + ], + "type": "text", + "content": "[43] Yin Li, Miao Liu, and James M. Rehg. In the eye of beholder: Joint learning of gaze and actions in first person video. European Conference on Computer Vision, 2018. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 601, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 643 + ], + "type": "text", + "content": "[44] Zhe Li, Yazan Abu Farha, and Jurgen Gall. Temporal action segmentation from timestamp supervision. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "type": "text", + "content": "[45] Daochang Liu, Qiyue Li, Anh-Dung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. International Conference on Computer Vision, 2023. 1, 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[46] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19129" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. *eccv*, 2024. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 95, + 295, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 95, + 295, + 138 + ], + "spans": [ + { + "bbox": [ + 57, + 95, + 295, + 138 + ], + "type": "text", + "content": "[47] Z. Lu and E. Elhamifar. Weakly-supervised action segmentation and alignment via transcript-aware union-of-subspaces learning. International Conference on Computer Vision, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 295, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 295, + 183 + ], + "type": "text", + "content": "[48] Z. Lu and E. Elhamifar. Set-supervised action learning in procedural task videos via pairwise order consistency. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 295, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 295, + 227 + ], + "type": "text", + "content": "[49] Z. Lu and E. Elhamifar. Fact: Frame-action cross-attention temporal modeling for efficient action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 295, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 295, + 260 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 295, + 260 + ], + "type": "text", + "content": "[50] J. Munro and D. Damen. Multi-modal domain adaptation for fine-grained action recognition. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 262, + 295, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 262, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 262, + 295, + 294 + ], + "type": "text", + "content": "[51] Zhanzhong Pang, Fadime Sener, Shrinivas Ramasubramanian, and Angela Yao. Long-tail temporal action segmentation with group-wise temporal logit adjustment. 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 295, + 295, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 295, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 295, + 295, + 338 + ], + "type": "text", + "content": "[52] Sam Perochon and Laurent Oudre. Unsupervised action segmentation of untrimmed egocentric videos. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 339, + 295, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 339, + 295, + 382 + ], + "spans": [ + { + "bbox": [ + 56, + 339, + 295, + 382 + ], + "type": "text", + "content": "[53] Will Price, Carl Vondrick, and Dima Damen. Unweavenet: Unweaving activity stories. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13770-13779, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 384, + 295, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 384, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 384, + 295, + 448 + ], + "type": "text", + "content": "[54] Camillo Quattrocchi, Antonino Furnari, Daniele Di Mauro, Mario Valerio Giuffrida, and Giovanni Maria Farinella. Synchronization is all you need: Exocentric-to-egocentric transfer for temporal action segmentation with unlabeled synchronized video pairs. European Conference on Computer Vision, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 449, + 295, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 295, + 514 + ], + "type": "text", + "content": "[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. International Conference on Machine learning, 2021. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 515, + 295, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 295, + 569 + ], + "type": "text", + "content": "[56] Francesco Ragusa, Antonino Furnari, Salvatore Livatino, and Giovanni Maria Farinella. The meccano dataset: Understanding human-object interactions from egocentric videos in an industrial-like domain. Winter Conference on Applications of Computer Vision, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 571, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 571, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 571, + 295, + 624 + ], + "type": "text", + "content": "[57] Rahul Rahaman, Dipika Singhania, Alexandre Thiery, and Angela Yao. A generalized and robust framework for timestamp supervision in temporal action segmentation. In Computer Vision-ECCV 2022: 17th European Conference, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 295, + 669 + ], + "type": "text", + "content": "[58] A. Richard, H. Kuehne, and J. Gall. Action sets: Weakly supervised action segmentation without ordering constraints. IEEE Conference on Computer Vision and Pattern Recognition, 2018." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "text", + "content": "[59] A. Richard, H. Kuehne, A. Iqbal, and J. Gall. Neuralnetwork-viterbi: A framework for weakly supervised video learning. IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[60] M. Rohrbach, S. Amin, M. Andriluka, and B. Schiele. A database for fine grained activity detection of cooking activities. IEEE Conference on Computer Vision and Pattern Recognition, 2012. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 554, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 171 + ], + "type": "text", + "content": "[61] Saquib Sarfraz, Naila Murray, Vivek Sharma, Ali Diba, Luc Van Gool, and Rainer Stiefelhagen. Temporally-weighted hierarchical clustering for unsupervised action segmentation. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 173, + 554, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 554, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 554, + 237 + ], + "type": "text", + "content": "[62] Saif Sayed, Reza Ghoddoosian, Bhaskar Trivedi, and Vassilis Athitsos. A new dataset and approach for timestamp supervised action segmentation using human object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 3133-3142, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 239, + 554, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 554, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 554, + 304 + ], + "type": "text", + "content": "[63] Fadime Sener, Dibyadip Chatterjee, Daniel Shelepov, Kun He, Dipika Singhania, Robert Wang, and Angela Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21096-21106, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 305, + 554, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 554, + 347 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 554, + 347 + ], + "type": "text", + "content": "[64] Y. Shen and E. Elhamifar. Semi-weakly-supervised learning of complex actions from instructional task videos. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 350, + 554, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 554, + 392 + ], + "type": "text", + "content": "[65] Y. Shen and E. Elhamifar. Progress-aware online action segmentation for egocentric procedural task videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 393, + 554, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 554, + 437 + ], + "type": "text", + "content": "[66] Y. Shen, L. Wang, and E. Elhamifar. Learning to segment actions from visual and language instructions via differentiable weak sequence alignment. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 437, + 554, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 554, + 481 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 554, + 481 + ], + "type": "text", + "content": "[67] Y. Shen, H. Wang, X. Yang, M. Feiszli, E. Elhamifar, L. Torresani, and E. Mavroudi. Learning to segment referred objects from narrated egocentric videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 482, + 554, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 554, + 524 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 554, + 524 + ], + "type": "text", + "content": "[68] G. A. Sigurdsson, S. Divvala, A. Farhadi, and A. Gupta. Asynchronous temporal fields for action recognition. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 526, + 554, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 554, + 570 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 554, + 570 + ], + "type": "text", + "content": "[69] B. Singh, T. K. Marks, M. Jones, O. Tuzel, and M. Shao. A multi-stream bi-directional recurrent neural network for finegrained action detection. IEEE Conference on Computer Vision and Pattern Recognition, 2016." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 571, + 554, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 554, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 554, + 613 + ], + "type": "text", + "content": "[70] Dipika Singhania, Rahul Rahaman, and Angela Yao. C2f- tcn: A framework for semi-and fully-supervised temporal action segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):11484-11501, 2023." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 614, + 554, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 554, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 554, + 656 + ], + "type": "text", + "content": "[71] Yaser Souri, Mohsen Fayyaz, Luca Minciullo, Gianpiero Francesca, and Juergen Gall. Fast Weakly Supervised Action Segmentation Using Mutual Consistency. PAMI, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "type": "text", + "content": "[72] Sebastian Stein and Stephen J McKenna. Combining embedded accelerometers with computer vision for recognizing food preparation activities. In Proceedings of the 2013 ACM international joint conference on Pervasive and ubiquitous computing, pages 729-738, 2013. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19130" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "text", + "content": "[73] Yuhao Su and Ehsan Elhamifar. Two-stage active learning for efficient temporal action segmentation. In European Conference on Computer Vision, pages 161-183. Springer, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "text", + "content": "[74] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "type": "text", + "content": "[75] Sanket Thakur, Cigdem Bayan, Pietro Morerio, Vittorio Murino, and Alessio Del Bue. Anticipating next active objects for egocentric videos. IEEE Access, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 274 + ], + "type": "text", + "content": "[76] Quoc-Huy Tran, Ahmed Mehmood, Muhammad Ahmed, Muhammad Naufil, Anas Zafar, Andrey Konin, and Zeeshan Zia. Permutation-aware activity segmentation via unsupervised frame-to-segment alignment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6426-6436, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 319 + ], + "type": "text", + "content": "[77] Huiyu Wang, Mitesh Kumar Singh, and Lorenzo Torresani. Ego-only: Egocentric action detection without exocentric transferring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5250-5261, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "type": "text", + "content": "[78] Jiahao Wang, Guo Chen, Yifei Huang, Limin Wang, and Tong Lu. Memory-and-anticipation transformer for online action understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13824–13835, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 442 + ], + "type": "text", + "content": "[79] Zhe Wang, Hao Chen, Xinyu Li, Chunhui Liu, Yuanjun Xiong, Joseph Tighe, and Charless Fowlkes. Sscap: Self-supervised co-occurrence action parsing for unsupervised temporal action segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1819–1828, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 444, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 294, + 487 + ], + "type": "text", + "content": "[80] Ming Xu and Stephen Gould. Temporally consistent unbalanced optimal transport for unsupervised action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 488, + 294, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 488, + 294, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 488, + 294, + 543 + ], + "type": "text", + "content": "[81] Ziwei Xu, Yogesh S Rawat, Yongkang Wong, Mohan Kankanhalli, and Mubarak Shah. Don't pour cereal into coffee: Differentiable temporal logic for temporal action segmentation. In Advances in Neural Information Processing Systems, 2022. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 544, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 578 + ], + "type": "text", + "content": "[82] Lingfeng Yang, Yueze Wang, Xiang Li, Xinlong Wang, and Jian Yang. Fine-grained visual prompting. Advances in Neural Information Processing Systems, 36, 2023. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 295, + 622 + ], + "type": "text", + "content": "[83] S. Yeung, O. Russakovsky, N. Jin, M. Andriluka, G. Mori, and L. Fei-Fei. Every moment counts: Dense detailed labeling of actions in complex videos. International Journal of Computer Vision, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 657 + ], + "type": "text", + "content": "[84] Fangqiu Yi, Hongyu Wen, and Tingting Jiang. Asformer: Transformer for action segmentation. In The British Machine Vision Conference (BMVC), 2021. 1, 2, 4, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[85] Chuhan Zhang, Ankush Gupta, and Andrew Zisserman. Helping hands: An object-aware ego-centric video recognition model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13901-13912, 2023. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 217 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "text", + "content": "[86] Ce Zhang, Changcheng Fu, Shijie Wang, Nakul Agarwal, Kwonjoon Lee, Chiho Choi, and Chen Sun. Object-centric video representation for long-term action anticipation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6751-6761, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "text", + "content": "[87] Xingyi Zhou, Anurag Arnab, Chen Sun, and Cordelia Schmid. How can objects help action recognition? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2362, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "type": "text", + "content": "[88] D. Zhukov, J. B. Alayrac, R. G. Cinbis, D. Fouhey, I. Laptev, and J. Sivic. Cross-task weakly supervised learning from instructional videos. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1, 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19131" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Understanding Multi-layered Transmission Matrices/0c5698ec-c1a0-4efc-b1d6-15a1169745e5_content_list.json b/2025/Understanding Multi-layered Transmission Matrices/0c5698ec-c1a0-4efc-b1d6-15a1169745e5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7ce3876cd18f5fc3ee0d4cf454429245f8823925 --- /dev/null +++ b/2025/Understanding Multi-layered Transmission Matrices/0c5698ec-c1a0-4efc-b1d6-15a1169745e5_content_list.json @@ -0,0 +1,1666 @@ +[ + { + "type": "text", + "text": "Understanding multi-layered transmission matrices", + "text_level": 1, + "bbox": [ + 238, + 130, + 759, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anat Levin and Marina Alterman \nDepartment of Electrical and Computer Engineering, Technion Haifa, Israel", + "bbox": [ + 250, + 172, + 746, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 261, + 326, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wavefront shaping systems attempt to correct aberrations caused by tissue scattering, by placing a spatial light modulator (SLM) in the optical path and using it to reshape the light wave emitted from a target of interest deep inside the tissue. However, the field-of-view we can correct with one SLM pattern is extremely small, because inherently the aberration is caused by the 3D tissue structure while the SLM is planar. To overcome this, multi-conjugate correction systems built with multiple layered SLMs have been introduced, which attempt to approximate the 3D tissue structure with multiple planar aberrations.", + "bbox": [ + 89, + 294, + 482, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, multi-conjugate systems with a large number of SLMs are not easy to construct. Therefore, it is important to understand how many SLM layers are actually needed for a good correction, and whether there are practical benefits from correction systems with a relatively small number of layers. To help in the design of future systems, this paper analyzes multi-layer corrections. We show that the well-known missing cone problem which fundamentally limits the axial resolution of microscopes, turns into an advantage when designing 3D correction systems. Due to this property less information should be captured by the correction system, facilitating sparser approximations. We also show that even if the number of available layers is insufficient for a reasonable approximation of the full 3D tissue volume, it can significantly expand the field-of-view when compared to single-layer correction systems, and therefore can largely accelerate their operation.", + "bbox": [ + 89, + 460, + 482, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 729, + 220, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scattering is a central challenge in multiple imaging tasks ranging from astronomy to microscopy, and biomedical imaging. It degrades our ability to resolve stars through telescopes, image into the bottom of the sea, and drive through foggy weather. Most importantly in the context of this paper, scattering severely limits our ability to image through biological tissue. This happens due to small variations in the refractive index of the tissue volume, which scatters the propagating light.", + "bbox": [ + 89, + 750, + 482, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Numerous computational optics approaches have been", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "developed for imaging through scattering tissue. These include computational restoration algorithms which attempt to fit the observed data with various models for the aberration and the target of interest [4, 5, 15, 17, 21, 23, 25, 32, 37, 39, 40]. Also, optical systems such as a confocal microscope [14] or an OCT [24], aim to filter out the scattered light and maintain only the ballistic unscattered part. However, as we attempt to see deeper into a scattering layer, ballistic light weakens significantly, and the complexity of aberration models increases. Moreover, biological target of interest such as fluorescent tissue components are very weak and as a result the scattered signal is also of a very low signal to noise ratio (SNR), further limiting digital correction models. A particularly attractive approach for scattering aberration removal is wavefront shaping, where one places a spatial light modulator (SLM) in the optical path of the microscope and uses it to correct the wavefront emitted from a point deep inside the tissue, so that light photons emerging from a single target point are brought into a single sensor point, despite the tissue aberration. Wavefront shaping is a particularly attractive solution for fluorescent imaging because the correction is done optically rather than digitally. Unlike ballistic filtering approaches, all light photons are used. Since the correction is done in the optical path, the limited number of available photons is brought into a single detector pixel and can be measured with higher SNR. As a result, wavefront shaping has the potential to revolutionize tissue imaging and allow us to overcome the fundamental limitations of scattering, and image very deep inside biological targets at a high SNR.", + "bbox": [ + 511, + 262, + 906, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The main limitation for wavefront shaping approaches is that the aberrations of neighboring points inside volumetric tissue vary rapidly, and in practice a single correction mask can only allow imaging of a small local field-of-view. This is because the scattering process happens throughout the 3D volume, while the aberration correction only happens on a planar SLM. Earlier demonstrations of wavefront shaping have considered targets placed at a distance behind a thin scattering layer, and this far-field arrangement has allowed them to correct larger regions with a single planar modulation [12, 15, 34, 39]. Similarly in the context of adaptive-optics, researchers have attempted to correct aber", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23164", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "rations due to imperfect optics or a refractive index mismatch [8, 18, 22]. These aberrations are introduced along the optical path, before the tissue, so the correction can apply to a larger area. However, in most biomedical applications the fluorescent target of interest is embedded within the scattering volume. If our goal is to correct scattering within the tissue itself, the field-of-view which can be corrected by a single planar modulation is extremely small, usually only a few microns [3].", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To correct larger fields of view, multi conjugate correction systems have been proposed. Such systems use multiple planar SLMs placed such that each of them is conjugate to a different plane inside the tissue volume. Each such SLM should realize a correction for the aberration happening inside a thin layer of the volume and the combination of such multi-plane corrections should approximate the 3D variation in refractive index, and allow us to correct it. Fig. 1 visualizes the concept of a multi-conjugate correction [16, 27, 35, 36]. In practice, multi-plane correction systems are more complicated than the simplified visualization of Fig. 1, as one needs to add multiple relay systems between the SLMs. This is first because there are physical limitations on the minimal distance between 2 SLMs, and second because we need to revert the order of the correction layers in the inversion path [19]. This makes multiconductor correction systems complex and cumbersome to build. Previous experimental demonstrations have realized corrections with no more than two SLMs. While two-layer correction systems have demonstrated improved performance compared to single layer correction, two planes clearly cannot model a thick 3D tissue structure.", + "bbox": [ + 91, + 229, + 483, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our goal in this research is to analytically study multislice models and attempt to answer the following questions: How many planes are required to approximate the 3D aberration of a scattering tissue volume? How does the quality of the correction degrade if we use a sparser approximation? If a construction of a practical correction system is limited to only to M layers, what is the maximum field of view that can be effectively corrected?", + "bbox": [ + 89, + 563, + 482, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we use the mathematical model of a transmission matrix, which stores the propagation of coherent wavefronts from one end of a scattering volume to the other. We study how well we can approximate such transmission matrices with a multi-slice model, and how does the error depend on the number of fitted layers.", + "bbox": [ + 89, + 686, + 482, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The straightforward answer to this question is to say that the spacing between layers should match the axial resolution of the microscope. For example, using an $\\mathbf{NA} = 0.5$ objective at wavelength $500nm$ , we get that the spacing between layers should be lower than $2\\mu m$ . Handling a modest tissue thickness of, e.g., $200\\mu m$ already requires 100 slices for a valid approximation. Clearly building a multiconjugate wavefront shaping systems with so many layers", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9bd6169b3d3f269f4ac098e352a507d116a8e3e75e350f687cba249d2cd12475.jpg", + "image_caption": [ + "Figure 1. Layered wavefront shaping: (Top) Consider a simple microscope setup where we use a 4F relay to image the light emerging from a tissue sample. Light is aberrated on its way to the sensor, and we see speckle patterns rather than sharp spots. (Middle) A single planar SLM can correct aberration for one spot (see red beam), but light emerging from nearby spots may not be corrected (the yellow beam generates a speckle pattern on the sensor). (Bottom) The field of view of the correction can possibly be extended with a multi-layer correction. The goal is to form a correction replicating the 3D structure of volumetric aberration inside the tissue. The figure illustrates two wavefronts from nearby spots (red and yellow) passing through the same volumetric correction, and both focus into a sharp spot behind the tissue." + ], + "image_footnote": [], + "bbox": [ + 526, + 79, + 911, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "is impractical.", + "bbox": [ + 513, + 507, + 609, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this research we take inspiration from [20], and offer a thorough analysis of the frequency content of transmission matrices, combining the forward scattering nature of biological tissue with the structure of the missing cone. We show that while the missing cone structure is known as one of the fundamental limitations of 3D microscopy, in the context of transmission matrices this limitation turns into an advantage. Due to the missing cone structure, many of the frequencies of the volumetric aberration do not participate in the transmission matrix, and therefore the number of layers required for approximating a transmission matrix is way smaller than the Nyquist limit. However, there are no free lunches and the number of layers scales with the tissue thickness. To explain a few hundred micron thick tissue one needs a few dozen layers, which is still too much to be realized by a physical correction system.", + "bbox": [ + 511, + 521, + 906, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given this conclusion we aim to understand if there are any practical benefits in realizing a multi-conjugate correction system with a small number of say 2 - 4 layers. We show that while a small number of layers cannot well approximate the 3D structure of the tissue, it can allow us to realize a correction over a limited field-of-view. The area which can be corrected with $M$ aberration layers is significantly larger than the area which can be corrected with one layer, and therefore it can significantly accelerate sequen", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23165", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f0721efb15f9ea6c7b32268b5b6ba3c736c1699d40f6b520e16a78a964fd6112.jpg", + "image_caption": [ + "(a) Full setup" + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 504, + 166 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a5c58ed619bc8c45c2f1d477f8c5034216e4fd639836400a17b0ee3e9b90dc3b.jpg", + "image_caption": [ + "Figure 2. Setup for transmission matrix acquisition: (a) A setup for transmission matrix imaging where a set of basis illuminations at the back of the sample propagate through the volume, and the scattered wavefronts are measured by a camera at the other end. In (b,c) we zoom on the sample illustrating two possible illumination bases. (b) A set of directional wavefronts (Fourier basis). (c) A set of wavefronts focused to points at the edge of the volume (primal basis). Both transmission matrices can be related by a simple Fourier transform. (d) A multi-slice model approximates the volumetric scattering with a set of planar aberrations." + ], + "image_footnote": [], + "bbox": [ + 532, + 93, + 630, + 162 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c49845d22af3999f33d70e1c082ade41175702092ae11814836b23961ce4eec6.jpg", + "image_caption": [ + "(b) Fourier illum.", + "(c) Point illum." + ], + "image_footnote": [], + "bbox": [ + 643, + 93, + 740, + 162 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/737cbeb37f676e96a48d7e780ea731c4776d5f743e73b0ed249eedd22cbb31b1.jpg", + "image_caption": [ + "(d) Layered approx." + ], + "image_footnote": [], + "bbox": [ + 759, + 94, + 856, + 162 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tial correction of the field of view. We validate our findings using real transmission matrices captured in the lab as well as synthetic transmission matrices simulated using accurate wave propagation models.", + "bbox": [ + 89, + 279, + 482, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Problem formulation", + "text_level": 1, + "bbox": [ + 89, + 349, + 290, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Imaging setup: For simplicity, the following derivation considers coherent scalar waves, ignoring polarization. In Fig. 2(a) we visualize a setup for transmission matrix acquisition. A scattering medium, such as tissue, is illuminated by a coherent source $\\mathbf{e}^k$ at one side of the medium. After propagating through the medium, the resulting coherent wavefront $\\mathbf{t}^k$ is recorded. Usually $\\mathbf{t}^k$ are complex valued speckle patterns with a pseudo-random structure. The illumination is translated and multiple wavefronts are recorded and stored as columns of a transmission matrix $\\mathcal{T}$ .", + "bbox": [ + 89, + 369, + 483, + 520 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The transmission matrix: Since the propagation of coherent light is linear, if we want to predict the propagation of any wavefront $\\mathbf{u}$ , we simply express it as a superposition of the recorded incoming illuminations $\\mathbf{u} = \\sum_{k} \\mathbf{u}(k) \\mathbf{e}^{k}$ , where $\\mathbf{u}(k)$ are complex scalars, and multiply it by the transmission matrix, to predict $\\mathcal{T} \\mathbf{u}$ . The illumination wavefronts $\\mathbf{e}^{k}$ can be selected at different bases. For example, the illumination can be translated at the Fourier plane, and then the wavefronts illuminating the sample are plane waves at different angles (Fig. 2(b)). Alternatively one can focus the light to a point at the back plane of the sample (Fig. 2(c)). This configuration is more appropriate for wavefront shaping algorithms whose goal is to recover wavefronts focusing inside the tissue and not far behind it. In both cases, however, the illuminations we can create are limited by the numerical aperture of the illuminating objective, which we mark here as NA. Similarly, the scattered wavefronts are measured via an imaging objective with the same bounded NA. This NA would serve as a central component in the results derived below. For simplicity of resulting equations, for the rest of this manuscript we assume the imaging and illumination objectives are focused at the same plane, such that if we would remove the scattering sample and image the sources directly, we would see the $\\mathbf{e}^{k}$ illuminations as sharp spots on the sensor plane. That is, if we use directional", + "bbox": [ + 89, + 523, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "illuminations, we also measure the Fourier transform of the scattered waves, namely place the sensor at the Fourier plane of the imaging objective. If we use point sources at the illumination end of the medium, we also focus the imaging objective at the same plane, so we measure a speckle image of a relatively small support.", + "bbox": [ + 511, + 279, + 905, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Range and resolution: We assume that in the primal basis the illuminators are spaced within an area of $\\Omega_p \\times \\Omega_p$ , and we sample them at the Nyquist resolution, so for an illumination wavelength $\\lambda$ they are separated by a pitch $\\Delta_p = \\lambda/(2\\mathbf{NA})$ . Equivalently, in the Fourier domain, we illuminate the sample with plane waves of the form", + "bbox": [ + 511, + 369, + 905, + 453 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {e} (\\mathbf {r}) = e ^ {2 \\pi i (\\bar {\\kappa} \\cdot \\mathbf {r})} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 656, + 455, + 903, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{r}$ is a point inside the 3D volume, and $\\bar{\\kappa} = \\frac{1}{\\lambda}\\bar{\\omega}$ where $\\lambda$ is the wavelength and $\\bar{\\omega}$ is a unit norm direction vector, i.e. the last entry of the direction vector $\\omega_{z}$ is $\\omega_{z} = \\sqrt{1 - (\\omega_{x}^{2} + \\omega_{y}^{2})}$ . For simplicity of notation, we assume the refractive index (RI) of the leading medium is 1. We can only use illumination direction inside the numerical aperture $\\sqrt{\\omega_x^2 + \\omega_y^2}\\leq \\mathbf{N}\\mathbf{A}$ (or equivalently $\\sqrt{\\kappa_x^2 + \\kappa_y^2}\\leq \\Omega_f$ with $\\Omega_{f} = \\mathbf{N A} / \\lambda$ ). To sample the transmission matrix at the Nyquist pitch, the $\\kappa_{x},\\kappa_{y}$ components of the illumination are sampled by the Nyquist pitch $\\Delta_{f} = 1 / 2\\Omega_{p}$ .", + "bbox": [ + 511, + 476, + 905, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the above sampling scheme, the transmission matrices at the Fourier and primal bases are related by a simple Fourier transform $\\mathcal{T}^p = \\mathcal{F}^T\\mathcal{T}^f\\mathcal{F}$ . While for wavefront shaping we are interested in primal transmission matrices, the Fourier ones simplify analysis. Importantly, since Fourier and primal transmission matrices are related by an orthonormal transformation, the fitting errors in both bases are equivalent.", + "bbox": [ + 511, + 628, + 905, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The multi-slice model: The relation between an incoming illumination and an outgoing one is determined by the solution to a wave equation which takes into account the variation in refractive index (RI) inside the sample. Rather than a fully continuous model, in a multi-slice model we approximate the volume with a set of planar aberrations, separated by small $z$ distances $\\Delta_z$ , where we have free space propagation between layers, as in Fig. 2(d). Mathematically this can be described as", + "bbox": [ + 511, + 747, + 905, + 872 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mathcal {T}} \\left(\\rho_ {1}, \\dots , \\rho_ {M}\\right) = \\boldsymbol {\\mathcal {P}} _ {O} \\boldsymbol {\\mathcal {D}} \\left(\\rho_ {M}\\right) \\dots \\boldsymbol {\\mathcal {P}} _ {\\Delta z} \\boldsymbol {\\mathcal {D}} \\left(\\rho_ {2}\\right) \\boldsymbol {\\mathcal {P}} _ {\\Delta z} \\boldsymbol {\\mathcal {D}} \\left(\\rho_ {1}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 877, + 903, + 905 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23166", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ffe8fb84265c39b1272d097ee9dc1821b61d3afdaf274043f211763d73a1f15e.jpg", + "image_caption": [ + "(a) Ground truth" + ], + "image_footnote": [], + "bbox": [ + 151, + 88, + 243, + 265 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/95e6e867cb700e3562b6d2eb098aff8864b563efaea60941564317acad376783.jpg", + "image_caption": [ + "(b) 32 layers" + ], + "image_footnote": [], + "bbox": [ + 254, + 99, + 346, + 266 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2f43cc57c9d498714e26fb51adff727a8d6794186eab3ec66cc99c556b26905b.jpg", + "image_caption": [ + "(c) 8 layers", + "Figure 3. Reconstructed volumes: (a) An xz slice from a toy volume, visualizing its refractive index variations. (b-d) Reconstructions with varying layer numbers. The columns of the transmission matrix provide an under-constraint inversion problem, and even when a dense set of layers is allocated in b, the axial resolution of the reconstructed volume is low. As a result the sparse reconstructions in c,d can also well explain the transmission matrix. (e) Numerical evaluation of fitting error as a function of the number of layers. Transmission matrices can be well fitted even if the allocated number of layers is lower than Nyquist predictions. However the required number of layers increases with the tissue thickness. The layer number is less affected when the optical density is increased for the same tissue thickness." + ], + "image_footnote": [], + "bbox": [ + 354, + 99, + 446, + 266 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d1c7c44fb676b59ccce5726adf4f7d480990a8c40a71cf6fa4d29bab9b1ef627.jpg", + "image_caption": [ + "(d) 4 layers" + ], + "image_footnote": [], + "bbox": [ + 454, + 99, + 544, + 266 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/eb9bb615bbe774bfbe881ece5da6a5612783a420231a2c520d5bcb03a87bb5a0.jpg", + "image_caption": [ + "(e) Fitting curves" + ], + "image_footnote": [], + "bbox": [ + 545, + 99, + 841, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\rho_{m}(x,y)$ is the 2D aberration of the $m$ th layer and $\\mathcal{D}(\\rho_m)$ is a diagonal matrix performing element-wise multiplication of the incoming wavefront by the aberration. $\\mathcal{P}_{\\Delta_z}$ describes free space propagation between layers at distance $\\Delta_z$ , which is essentially a convolution with a spherical wavefront. Finally $\\mathcal{P}_O$ propagates the wavefront in free space from the tissue boundary, via the lenses, to the camera sensor.", + "bbox": [ + 88, + 392, + 482, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goal is to understand what plane spacing $\\Delta_z$ will facilitate a good representation of $\\mathcal{T}$ , and how this approximation would degrade when we use sparser sampling.", + "bbox": [ + 86, + 513, + 482, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A common approximation for weak aberrations is:", + "bbox": [ + 109, + 559, + 442, + 571 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\rho} _ {m} (x, y) = \\mu + \\delta \\boldsymbol {\\rho} _ {m} (x, y), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 571, + 480, + 587 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu$ is a scalar whose magnitude is a bit smaller than 1, encoding the ballistic portion of the light which passes through the layer without aberration, and the magnitude of the aberration is significantly smaller than the ballistic term $|\\delta \\rho_{m}(x,y)|\\ll |\\mu |$ . For weak scattering most of the energy does not scatter more than once. It is therefore common to consider a first order approximation to the transmission matrix (often known as the single scattering, first-Born or Rytrov approximation [9, 10, 13, 29, 33]), which reduces to", + "bbox": [ + 89, + 592, + 482, + 717 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mathcal {T}} \\left(\\delta \\rho_ {1}, \\dots , \\delta \\rho_ {M}\\right) \\approx \\mu^ {M} \\boldsymbol {\\mathcal {P}} _ {O} \\boldsymbol {\\mathcal {P}} _ {M \\Delta_ {z}} + \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 719, + 482, + 736 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mu^ {M - 1} \\sum_ {m = 1} ^ {M} \\mathcal {P} _ {O} \\mathcal {P} _ {(M - m) \\Delta_ {z}} \\mathcal {D} (\\delta \\rho_ {m}) \\mathcal {P} _ {m \\Delta_ {z}}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 738, + 480, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The full light propagation in Eq. (2) includes multiplication of the different aberrations and hence the transmission matrix is highly nonlinear. In contrast, the advantage of the weakly scattering model in Eq. (4) is that the resulting transmission matrix is a linear function of the aberration layers. This linear approximation significantly simplifies analysis.", + "bbox": [ + 89, + 779, + 482, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffraction tomography algorithms such as [11] capture a subset of columns from the transmission matrix and use", + "bbox": [ + 89, + 869, + 482, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "gradient descent optimization to seek layers $\\rho_1, \\ldots, \\rho_M$ which best explain the measured wavefronts, assuming that the propagation operators $\\mathcal{P}_{\\Delta_z}, \\mathcal{P}_O$ are known. We will use a similar optimization scheme below, when trying to find a multi-slice representation for a given transmission matrix. To correct aberrations, a physical multi-conjugate wavefront shaping system would display on the SLM the conjugate of the recovered layers.", + "bbox": [ + 511, + 392, + 906, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Evaluating multi-slice approximations", + "text_level": 1, + "bbox": [ + 511, + 527, + 856, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We start with numerical experiments demonstrating the main findings. In Sec. 4 below we derive a theory explaining them. To gain intuition we generated a toy volume composed of spheres with varying refractive indices illustrated in Fig. 3(a), and obtained the corresponding transmission matrix by propagating a set of point sources through the volume. We then used the transmission matrix to solve for a multi-slice approximation, while varying the number of layers. In Fig. 3(b-d) we visualize some of the volumetric reconstructions, assuming a $\\lambda = 500nm$ illumination. Even when the fitted layers are sampled very densely at an interval of $\\Delta_z = 0.5\\mu m$ , we cannot fully recover the target and the fitted volume has a low axial resolution. We analyze this in Sec. 4 below and explain that this results from the well known missing cone problem, a fundamental degeneracy limiting the axial resolution of a microscope. Said differently the transmission matrix, or the full set of propagating waves, only provides a set of under-constraint measurements from which the volume cannot be uniquely recovered. While the missing cone problem is considered a fundamental limitation in microscopy, we argue that in the context of wavefront shaping it actually turns into an advantage. Since the volume reconstruction is not unique we can", + "bbox": [ + 511, + 553, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23167", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c030826722d17eab87af7556ccee4f64461f35318021127ea39c15c82e996dde.jpg", + "image_caption": [ + "(a) Ground truth" + ], + "image_footnote": [], + "bbox": [ + 106, + 93, + 197, + 265 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/44e481b3b0d8ba9707b4df26d4dec5884e8dcb09e2c991e70c435b772a81296e.jpg", + "image_caption": [ + "(b) Dense fit" + ], + "image_footnote": [], + "bbox": [ + 212, + 95, + 303, + 265 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bcfef151987ea8ba01c9b03cba4f97f51e36b7536bf389e24341ff5ee1f529dc.jpg", + "image_caption": [ + "(c) Sparse fit" + ], + "image_footnote": [], + "bbox": [ + 305, + 95, + 395, + 265 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/89bb3c593d97c43e54b7e49fec9b357151b35b0b8b8387c56fc621c6217359b3.jpg", + "image_caption": [ + "(d) Dense fit", + "Figure 4. Reconstructed volumes with varying supports: (a) An xz slice from a toy volume, visualizing its refractive index variations. (b,d) Two reconstructions of the volume using a dense layer sampling. When the support of the measured window $\\Omega_{p}$ shrinks, the lateral extent of the reconstructed region decreases, however, its axial resolution is also reduced. (c,e) A sparse fit with the same input. Despite the low quality of these reconstructions, they predict well the columns of the transmission matrix of the given support $\\Omega_{p}$ . (f) Numerical evaluation of fitting error as a function of the number of layers, while varying the range $\\Omega_{p}$ covered by the transmission matrix. Smaller supports can be fitted with fewer layers." + ], + "image_footnote": [], + "bbox": [ + 395, + 95, + 500, + 265 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a9eab028a53d42a94619d402f561f77c79d549e911fd9c9e2140b5007b22c571.jpg", + "image_caption": [ + "(e) Sparse fit" + ], + "image_footnote": [], + "bbox": [ + 500, + 95, + 591, + 265 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d667e5f775f637a56c3bd675225f8d9c91df8e3aaf10a06697fafb6262f87970.jpg", + "image_caption": [ + "(f) Fitting curves" + ], + "image_footnote": [], + "bbox": [ + 604, + 103, + 890, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "approximate the same transmission matrix with a sparser set of layers.", + "bbox": [ + 88, + 380, + 482, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To quantify the fitting error of the multi-slice model, we denote by $\\pmb{\\varepsilon}_{M}$ the reconstruction error of a given transmission matrix with the best $M$ layers", + "bbox": [ + 89, + 409, + 482, + 450 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mathcal {E}} _ {M} = \\min _ {\\rho_ {1}, \\dots , \\rho_ {M}} \\| \\boldsymbol {\\mathcal {T}} _ {\\text {e x a c t}} - \\boldsymbol {\\mathcal {T}} (\\rho_ {1}, \\dots , \\rho_ {M}) \\| ^ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 450, + 482, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Fig. 3(e) we evaluate $\\pmb{\\mathcal{E}}_M$ as a function of the layer number $M$ . Even the seemingly aliased fit of Fig. 3(d) still provides a low fitting error. However, there are no free lunches and when the number of layers is too low the fitting error is high. When we repeat the simulation with a thicker volume we need even more layers for a good fit. Interestingly, we also repeat the simulation with denser volumes (larger optical depth, or shorter mean free path) of the same thickness, and this has less effect on the required number of layers. In our plots, the case $M = 0$ refers to no correction at all, so effectively we assume that the transmission matrix is diagonal. The similarity between a diagonal transmission matrix and the target transmission matrix is a measure of the amount of ballistic light.", + "bbox": [ + 88, + 477, + 482, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The thickness of the synthetic simulation here is limited due to computational constraints. However, usually wavefront shaping systems attempt to see into a few hundred microns of tissue and we anticipate that for such volumes we need at least a few dozen layers for a good fit. Clearly realizing a physical wavefront shaping system with that many layers is impractical. We argue that while a small number of layers is not enough for focusing light anywhere across the tissue area, it can still help us to focus light over a compact support area, and this area is significantly wider than what we can correct with a single layer.", + "bbox": [ + 88, + 688, + 482, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compact support: To test the effect of the target area, we consider transmission matrices measured with illumination points confined to a limited area of size $\\Omega_p \\times \\Omega_p$ . If we can", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "fit such a transmission matrix with $M$ layers and place the values of these $M$ layers on $M$ SLMs, the resulting wavefront shaping system can correct light from the $\\Omega_p \\times \\Omega_p$ area, but target points outside this area are unlikely to be corrected as they are not included in the fitting process.", + "bbox": [ + 511, + 380, + 903, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To test the effect of support on the number of fitted layers, we return to the numerical simulation of Fig. 3, but fit subsets of columns contained in smaller $\\Omega_p$ ranges. Indeed, smaller supports can be well-fitted with a smaller number of layers. As another way to understand it, we show in Fig. 4 an xz slice from the ground truth refractive index volume and from a few reconstructions. We show that even if we do not limit the layer number $M$ , the recovered RI volume has limited axial resolution and this axial resolution is lower when measured range $\\Omega_p$ decreases. This explains why lower supports can be fitted with fewer layers. For each of the two supports we also show a sparse fit, with the minimal $M$ value that provides a good prediction of the transmission matrix. We explain this observation in Sec. 4. This result is also in agreement with [26], who correct aberration over a relatively small field of view through a few hundred microns of tissue, using only 5 layers.", + "bbox": [ + 511, + 458, + 906, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implications to layered wavefront shaping systems. In Fig. 4(f) we plotted a bar below which the approximation is reasonable. We can see that for e.g. the $40 \\times 40\\mu m$ support we can obtain good focusing with about $M = 6$ layers. It is likely that fitting a full field-of-view of a few hundred microns would require more layers. However, we can see that while with $M = 1$ layer we could only focus over a small area of about $1 \\times 1\\mu m$ , when increasing to $M = 2$ we could focus over a $\\times 4$ larger area of support $2 \\times 2\\mu m$ , and with $M = 3$ we could focus over a $\\times 25$ larger area of support $5 \\times 5\\mu m$ . Computing multiple correction elements is only going to increase the complexity of the wavefront shaping", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23168", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a619d8c80fbac286a52416d0615f50d448ffa7ed9fbf2a4570d53d4dff932fdb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 116, + 89, + 408, + 257 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3ef80ea7a846532437c145aa0d761356e765cebe99d085ea8dc6bb3863484d79.jpg", + "image_caption": [ + "Figure 5. Layer support for a captured transmission matrix: We consider a transmission matrix captured in the lab through a chicken breast layer of thickness $170\\mu m$ . Left: plotting the fitting error as a function of the number of layers. Smaller supports can be fitted with fewer layers. Right: Visualization of a spot focusing behind the tissue computed using the wavefronts of the fitted model. We compare models fitted to different supports. In the top row we need a larger number of layers, but the spot can be scanned over a $25 \\times 25\\mu m$ window using the same layers. In the lower rows we achieve a focused spot using a smaller number of layers, but these layers can scan the focused spot over smaller windows of sizes $6 \\times 6\\mu m$ and $1 \\times 1\\mu m$ ." + ], + "image_footnote": [], + "bbox": [ + 428, + 89, + 883, + 276 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "algorithm in [1] linearly with the number of layers. Alternatively, a wavefront shaping system using a single SLM would need to scan the field-of-view sequentially and correct each local window independently. Since the increase in the support of the window we can explain with e.g. 3 layers is significantly larger than a $\\times 3$ factor, a multi-conjugate wavefront shaping system with a small number of SLMs as in [16, 27, 31, 35] can already have a large practical benefit. While Fig. 4 only considers a simulation, we analyze realistic numbers below.", + "bbox": [ + 88, + 369, + 480, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Lab results: In the supplementary file we validate the theory with multiple real transmission matrices, representing thick volumes with multiple-scattering.", + "bbox": [ + 89, + 523, + 482, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We start with numerically simulated transmission matrices using a more accurate wave-propagation model [6, 7]. We then test lab-captured transmission matrices measuring realistic scattering samples including mouse brain tissue, chicken breast tissue and a parafilm layer. Please refer to supplementary file for additional results.", + "bbox": [ + 89, + 568, + 482, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Fig. 5 we show one fitting result using a multiply-scattering chicken breast layer of thickness $d = 170\\mu m$ . The transmission matrices cover an area with support $\\Omega_{p} \\times \\Omega_{p} = 13 \\times 13\\mu m$ . As in Fig. 4(f), we see that smaller spatial ranges can be fitted with fewer layers.", + "bbox": [ + 89, + 659, + 482, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As another way to assess the quality of the fit, we used columns of the fitted matrices to focus light through the tissue and show the shape of the focused spot and visualize the focusing quality in Fig. 5. To simulate the focused spot we simply visualize the product $\\mathcal{T}_{\\mathrm{exact}} \\cdot \\mathbf{t}_{\\mathrm{fit,M}}^k$ , where $\\mathbf{t}_{\\mathrm{fit,M}}^k$ is a column of the fitted matrix $\\mathcal{T}(\\rho_1, \\ldots, \\rho_M)$ , namely a wavefront generated by the layered approximation. Note that while we only show focusing at one point, the layers are optimized such that they allow us to focus at any point inside the $\\Omega_p \\times \\Omega_p$ window.", + "bbox": [ + 89, + 734, + 482, + 885 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "From the results in Fig. 5 we can see that for the $13\\times$", + "bbox": [ + 109, + 885, + 480, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$13\\mu m$ support measured, we can obtain good focusing with about 3 layers. Fitting a full field-of-view of several hundred microns likely requires additional layers. However, we observe that while using $M = 1$ allows us to focus on a small area of about $3\\times 3\\mu m$ , with $M = 3$ , we can focus on an area $\\times 18$ larger, reaching $13\\times 13\\mu m$ . Therefore, using multiple layers can significantly accelerate a sequential scanning of a wave-front shaping system.", + "bbox": [ + 511, + 369, + 906, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Sampling theory for transmission matrices", + "text_level": 1, + "bbox": [ + 511, + 502, + 893, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our goal in this section is to explain the empirical findings of the previous section by analyzing the information captured by a transmission matrix.", + "bbox": [ + 511, + 523, + 905, + 569 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Nyquist sampling and the missing cone", + "text_level": 1, + "bbox": [ + 511, + 579, + 846, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We start by considering the case of weakly scattering volumes which, as in Eq. (4), can be reasonably approximated by considering the component of light that scattered once. We also assume the transmission matrix is sampled over a very wide support $\\Omega_{p}$ . We discuss the implications of multiple scattering and limited supports in subsequent sections.", + "bbox": [ + 511, + 597, + 905, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We denote by $\\pmb{n}(\\mathbf{r})$ the difference between the RI at a 3D position $\\mathbf{r}$ in the volume to that of the leading medium around it. We denote its 3D Fourier transform by $\\hat{n}(\\kappa)$ .", + "bbox": [ + 511, + 688, + 903, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We start by considering the Fourier representation of the transmission matrix, where it is illuminated and measured by plane waves. Using the weakly scattering (first-Born) approximation, the wavefront scattering toward direction $\\bar{\\omega}^v$ when illuminated by an incoming plane wave at direction $\\bar{\\omega}^i$ is proportional, up to a multiplicative factor, to [28, 30, 38]:", + "bbox": [ + 511, + 733, + 905, + 830 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\propto \\int \\boldsymbol {n} (\\mathbf {r}) e ^ {2 \\pi i \\left(\\bar {\\kappa} ^ {i} - \\bar {\\kappa} ^ {v}\\right) \\cdot \\mathbf {r}} d \\mathbf {r}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 835, + 903, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with $\\bar{\\kappa}^v = 1 / \\lambda \\bar{\\omega}^v$ , $\\bar{\\kappa}^i = 1 / \\lambda \\bar{\\omega}^i$ . This implies that effectively an entry of the transmission matrix is a sample", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23169", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/46bfd089b3b35d5bc046dad013056388f346bc9e16d249fa49982443b4d7092f.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 218, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0a5fc2e4f939bdd085a233905ed2d980356a80886f347bad34b87a953175c55f.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 223, + 88, + 352, + 185 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bd09b39ba8f1ab55a3632705decb26df4068d7436e42bebbaa667ed447d32ce1.jpg", + "image_caption": [ + "(c)", + "Figure 6. Spectrum structure: (a) An $xz$ slice out of the spectrum of $\\hat{n} (\\omega)$ . Entries of a weakly-scattering transmission matrix limited by an aperture NA only lie inside the butterfly area. (b) Zooming on the center right area of (a) (purple square). Assuming the $\\omega_{z}$ axis is cut at $\\pm \\frac{1}{2}\\Lambda^{M}$ , the transmission matrix entries inside the dashed blue area are maintained, and the entries in the dashed red area are lost. (c) The energy at a slice from the spectrum of the volume in Fig. 3(a). Since the volume structure is smooth, most energy is concentrated at the low frequencies. As a result the cropped red region in (b) do not contain a lot of energy." + ], + "image_footnote": [], + "bbox": [ + 357, + 89, + 477, + 183 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "from the 3D Fourier transform of the volume at frequency $\\kappa = \\bar{\\kappa}^i -\\bar{\\kappa}^v$ . Note that throughout this paper we use bars $\\bar{\\omega}$ , $\\bar{\\kappa}$ to denote vectors with norm constraints, and use normal bold fonts to denote standard 3D vectors, such as frequencies $\\kappa$ in the 3D spectrum. Unfortunately, a large portion of the 3D spectrum cannot be measured. Recalling that the illumination and viewing directions are limited by an aperture of width NA, the subset of 3D frequencies captured by the transmission matrix occupies only a butterfly shape, and the rest of the cone is missing, see illustration in Fig. 6(a). Using paraxial approximation it can be shown [28, 38] that the range of frequencies which can be generated is", + "bbox": [ + 89, + 351, + 483, + 547 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\left| \\kappa_ {z} \\right| \\leq \\mathbf {N A} \\left| \\boldsymbol {\\kappa} _ {x y} \\right| - \\frac {\\lambda}{2} \\left| \\boldsymbol {\\kappa} _ {x y} \\right| ^ {2}, \\quad \\text {w i t h} \\quad \\left| \\boldsymbol {\\kappa} _ {x y} \\right| = \\sqrt {\\kappa_ {x} ^ {2} + \\kappa_ {y} ^ {2}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 556, + 483, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The missing cone is a serious problem for diffraction tomography algorithms aiming to recover the 3D structure of the refractive index $n(\\mathbf{r})$ from columns of the transmission matrix, since many of the frequencies of the volume cannot be recovered. We argue that in the context of transmission matrices, the missing cone problem turns into an advantage. The fact that the transmission matrix only captures a limited range of frequencies allows us to approximate it with fewer samples, as derived below.", + "bbox": [ + 89, + 597, + 483, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Minimal sampling: Our goal is to approximate a transmission matrix with layers. We start by asking what spacing between layers will allow an exact reconstruction of the transmission matrix. If we approximate the volume $\\pmb{n}(\\mathbf{r})$ with a set of planes, we essentially sample a continuous signal. Classical Nyquist theory states that if we sample a signal at intervals of $\\Delta_z$ , its Fourier transform includes replicas spaced by $\\Lambda = 1 / \\Delta_z$ . To achieve a good reconstruction, we want replicas to happen only in areas of the spectrum that the transmission matrix does not sample. Since the $\\kappa_x, \\kappa_y$ components of our illuminations are limited by the aperture $|\\kappa_{xy}| \\leq \\Omega_f$ , and given the butterfly shape in Eq. (7), the $\\kappa_z$", + "bbox": [ + 89, + 733, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "frequencies we need to maintain are bounded to", + "bbox": [ + 511, + 90, + 830, + 104 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\left| \\kappa_ {z} \\right| \\leq \\frac {\\mathrm {N A} \\Omega_ {f}}{2}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 658, + 108, + 903, + 136 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Therefore, for an error-free reconstruction, the minimal Fourier range and primal spacing are", + "bbox": [ + 511, + 138, + 903, + 167 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\Lambda^ {*} = \\mathbf {N A} \\boldsymbol {\\Omega} _ {f}, \\quad \\Delta_ {z} ^ {*} = \\frac {1}{\\mathbf {N A} \\boldsymbol {\\Omega} _ {f}} = \\frac {\\lambda}{\\mathbf {N A} ^ {2}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 167, + 903, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For a scattering volume of thickness $d$ , the number of required layers is:", + "bbox": [ + 511, + 199, + 903, + 226 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nM ^ {*} = \\frac {d \\cdot \\mathbf {N A} ^ {2}}{\\lambda}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 658, + 222, + 903, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For typical numbers, consider e.g. $\\lambda = 0.5\\mu m$ and $\\mathbf{N}\\mathbf{A} = 0.5$ , the plane spacing should be relatively small, and $\\Delta_z \\leq 2\\mu m$ . For a modest tissue thickness of $d = 200\\mu m$ , we will need $M = 100$ planes. Clearly it is impractical to realize such a large number of planes by a physical correction system.", + "bbox": [ + 511, + 250, + 906, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Sparse multi-slice approximations", + "text_level": 1, + "bbox": [ + 511, + 344, + 810, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Consider a scattering sample of thickness $d$ , which we try to approximate using $M < M^{*}$ layers separated by $\\Delta_z = \\frac{d}{M}$ . Our goal is to show that the fitting error $\\mathcal{E}_M$ defined in Eq. (5) decays relatively fast with $M$ and we can get a reasonable approximation to the transmission matrix even if the number of layers is lower than the exact prediction in Eq. (10). This is a result of two main properties: (i) The volumes describing realistic tissue samples have more energy in the low frequencies; and (ii) the structure of the missing cone implies that, in any case, a significant portion of the spectrum is not sampled by the transmission matrix.", + "bbox": [ + 511, + 362, + 905, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To understand this, consider a naive selection of $M$ layers $\\rho_{1},\\ldots ,\\rho_{M}$ . Rather than actually solving an optimization problem, we use the ground truth volume $\\hat{n} (\\kappa)$ , and simply set to zero any frequency content of the RI volume $\\hat{n} (\\kappa)$ at $\\kappa_z$ values larger than the possible Nyquist range", + "bbox": [ + 511, + 527, + 903, + 598 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\Lambda^ {M} = \\frac {1}{\\Delta_ {z}} = \\frac {\\lambda M}{d}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 645, + 598, + 903, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and we then Fourier transform $\\hat{n}(\\kappa)$ to $\\pmb{n}(\\mathbf{r})$ and sample planes at spacing $\\Delta_z = d / M$ . The error of this approximation is basically the integral of content above the cut-off $\\Lambda^M$ . Since according to Eq. (11), the cut-off frequency $\\Lambda^M$ scales linearly with the number of layers, the portion of the spectrum which is lost by this low-pass operation scales linearly with the number of layers $M$ . Thus, the naive answer is that the error of a multi-slice approximation decays linearly with $M$ . In practice we show below that the error decays much faster, since large areas from the 3D spectrum of the sample are not used by the transmission matrix.", + "bbox": [ + 511, + 628, + 905, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To gain intuition, consider Fig. 6(b). We note that the butterfly shape is such that low 2D frequencies (i.e. low $|\\kappa_{xy}|$ ), also span less content along the $\\kappa_z$ axes, and hence these frequencies are not lost even with low bandwidths $\\Lambda^M$ . For the higher $|\\kappa_{xy}|$ frequencies, the lower $\\kappa_z$ part, marked in dashed blue in Fig. 6(b) is preserved, while the higher portion marked in red is lost.", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23170", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ddb4ea6488694c3777f31370d0477afa592bcb08ab53079e3459e6c64263bdf.jpg", + "image_caption": [ + "(a) Support primal" + ], + "image_footnote": [], + "bbox": [ + 106, + 88, + 207, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2665e7a357dd917add952742cb4d40055332310bf78b39f5de47e5510472e173.jpg", + "image_caption": [ + "(b) Support freq." + ], + "image_footnote": [], + "bbox": [ + 220, + 88, + 330, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/244545752f597c82df357181f9f1869e68249011bbda892398d06aa4786b500c.jpg", + "image_caption": [ + "(c) Local directions", + "Figure 7. Layer support: (a) As point sources expand while propagating through the volume, the width of the aberration layers should be wider than the measured support $\\Omega_{p}$ . (b) To express compact support transmission matrices in the Fourier basis, we illuminate the volume by a set of plane waves passing through an aperture. (c) The local cone of illumination angles reaching different points inside the volume is much smaller than the actual range of incoming illuminations." + ], + "image_footnote": [], + "bbox": [ + 343, + 88, + 460, + 210 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Another important property of tissue is that its RI is locally smooth, and therefore if we look at its spectrum (Fig. 6(c)), we have much more energy in low frequencies than in high ones. This also agrees with the fact that tissue is forward scattering. Therefore, in the red areas of Fig. 6(b) that are not sampled by the transmission matrix, there is less energy than in the lower frequencies. Thus, despite the fact that the multi-slice approximation sacrifices such frequencies, not much energy is being lost.", + "bbox": [ + 89, + 340, + 483, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In most cases, it is hard to give analytic formulas for the decay of $\\pmb{\\mathcal{E}}_M$ as a function of $M$ , but the supplementary file contains some derivation under simplifying assumptions.", + "bbox": [ + 89, + 479, + 483, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multiple scattering. The results in Fig. 3(e) include multislice fits with both lower and higher optical depth values up to $OD = 4$ , showing that the fitting error still decays rapidly with the number of layers.", + "bbox": [ + 89, + 526, + 483, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While it is hard to give analytical predictions for the multiple scattering case, it appears that when the optical depth of the tissue is not too high, similar results hold. The reason is that for a forward-scattering material such as tissue, light paths only scatter at small angles, and after a small number of scattering events they are not completely scrambled. For an intuition, in the supplementary file we use a Monte-Carlo simulation illustrating that for a forward-scattering material, after a modest number of scattering events, most light paths still remain within the butterfly area of the spectrum. This is usually the regime that wavefront-shaping algorithms like [1, 2] attempt to tackle, as such algorithms attempt to push the depth at which conventional microscopes can see, but they do not yet attempt to image extremely deep where scattering is fully diffused.", + "bbox": [ + 89, + 587, + 483, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Bounded support matrices", + "text_level": 1, + "bbox": [ + 89, + 819, + 331, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As stated above, much fewer layers are needed if we only attempt to fit transmission matrices sampled over a small area of size $\\Omega_p\\times \\Omega_p$", + "bbox": [ + 89, + 839, + 482, + 885 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When fitting layers, we need to account for the fact that a", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "point/plane illumination scattering through a volume is expanding. Thus, the aberration layers required to explain the transmission matrix are wider than the imaged area of size $\\Omega_{p} \\times \\Omega_{p}$ , see Fig. 7(a). Expressing a transmission matrix with a compact support at the frequency basis is equivalent to illuminating the sample with plane waves spanning a wide set of angles, but all the waves pass through a narrow aperture of size $\\Omega_{p} \\times \\Omega_{p}$ . In this case most points inside the volume usually do not receive light from all angles. In Fig. 7(c) we plot the narrow cone of angles arriving at three different points (a ray leaving a point $\\mathbf{p}_{\\mathbf{j}}$ at angle $\\bar{\\omega}$ is included in the cone only if this angle is not cropped by the aperture at the back of the sample). Since the cone of light reaching each point is narrower than the full numerical aperture, a local Fourier transform would not have content over the full butterfly area, but cover an even lower range of axial $\\omega_{z}$ frequencies. Hence, following the analysis in the previous section, it can be sampled with fewer layers. We visualize the local spectra in the supplement.", + "bbox": [ + 511, + 90, + 906, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 511, + 393, + 627, + 410 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper studies how well we can fit transmission matrices with multi-slice models and how many layers are needed in practice. We show that the required number of layers is much lower than the prediction of Nyquist sampling theory. This is because the missing cone problem of 3D microscopy turns into an advantage when analyzing transmission matrices. Moreover, as tissue is forward scattering, its 3D spectrum contains most energy in the low frequencies and much less energy in the higher ones. Hence, even if the sparse sampling acts as a low-pass on the large axial frequencies, not much energy is lost.", + "bbox": [ + 511, + 416, + 906, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We validate this prediction using both synthetic simulations and realistic lab measurements. We show that even though realistic transmission matrices contain backward and side-ward scattering which are not modeled by multi-slice matrices, they provide a reasonable approximation that is good enough for wavefront shaping focusing.", + "bbox": [ + 511, + 585, + 905, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our findings suggest that multi-slice models can be applicable for the design of non-local wavefront shaping systems, allowing for a volumetric correction with a small number of layered SLMs [16, 27, 35, 36]. Even if $2 - 3$ SLM layers may not be enough for correcting the full field-of-view of a wide image through a thick tissue, they can allow us to increase the support of the fitted area by much more than $\\times 2 - 3$ thus they can significantly accelerate a sequential wavefront-shaping correction of a wide image.", + "bbox": [ + 511, + 680, + 905, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments: This research was funded by ERC SpeckleCorr-101043471, ISF 563/24.", + "bbox": [ + 511, + 820, + 905, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Code and data: can be obtained at https:// webee.technion.ac.il/people/anat.levin/ papers/LayerdTmat(cvpr25_code.zip", + "bbox": [ + 511, + 854, + 903, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23171", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dror Aizik and Anat Levin. Non-invasive and noise-robust light focusing using confocal wavefront shaping. Nature communications, 15(5575), 2024.", + "[2] Dror Aizik, Ioannis Gkioulekas, and Anat Levin. Fluorescent wavefront shaping using incoherent iterative phase conjugation. Optica, 9(7):746-754, 2022.", + "[3] M. Alterman, C. Bar, I. Gkioulekas, and A. Levin. Imaging with local speckle intensity correlations: theory and practice. ACM TOG, 2021.", + "[4] YoonSeok Baek, Hilton B. de Aguiar, and Sylvain Gigan. Phase conjugation with spatially incoherent light in complex media. Nature Photonics, 17(12):1114-1119, 2023.", + "[5] Paul Balondrade, Victor Barolle, Nicolas Guigui, Emeric Auriant, Nathan Rougier, Claude Boccara, Mathias Fink, and Alexandre Aubry. Multi-spectral reflection matrix for ultrafast 3d label-free microscopy. Nature Photonics, 18(10): 1097-1104, 2024.", + "[6] Chen Bar, Marina Alterman, Ioannis Gkioulekas, and Anat Levin. A monte carlo framework for rendering speckle statistics in scattering media. ACM TOG, 2019.", + "[7] Chen Bar, Ioannis Gkioulekas, and Anat Levin. Rendering near-field speckle statistics in scattering media. ACM TOG, 2020.", + "[8] Martin Booth. Adaptive optical microscopy: The ongoing quest for a perfect image. Light: Science and Applications, 3:e165, 2014.", + "[9] Bingquan Chen and Jakob J. Stamnes. Validity of diffraction tomography based on the first born and the first rytrov approximations. Appl. Opt., 37(14):2996-3006, 1998.", + "[10] Wonshik Choi, Christopher Fang-Yen, Kamran Badizadegan, Seungeun Oh, Niyom Lue, Ramachandra R Dasari, and Michael S Feld. Tomographic phase microscopy. Nature methods, 4(9):717-719, 2007.", + "[11] Shwetadwip Chowdhury, Michael Chen, Regina Eckert, David Ren, Fan Wu, Nicole Repina, and Laura Waller. High-resolution 3d refractive index microscopy of multiple-scattering samples from intensity images. Optica, 6(9): 1211-1219, 2019.", + "[12] Anat Daniel, Dan Oron, and Yaron Silberberg. Light focusing through scattering media via linear fluorescence variance maximization, and its application for fluorescence imaging. Optics Express, 27(15):21778, 2019.", + "[13] A. J. Devaney. Inverse-scattering theory within the rytov approximation. Opt. Lett., 6(8):374-376, 1981.", + "[14] Amicia D. Elliott. Confocal microscopy: Principles and modern practices. Current Protocols in Cytometry, 92(1): e68, 2020.", + "[15] Brandon Y. Feng, Haiyun Guo, Mingyang Xie, Vivek Boominathan, Manoj K. Sharma, Ashok Veeraraghavan, and Christopher A. Metzler. Neuws: Neural wavefront shaping for guidestar-free imaging through static and dynamic scattering media. Science Advances, 9(26):eadg4671, 2023.", + "[16] T. Furieri, A. Bassi, and S. Bonora. Large field of view aberrations correction with deformable lenses and multi conjugate adaptive optics. Journal of Biophotonics, 16(12): e202300104, 2023." + ], + "bbox": [ + 93, + 114, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Omri Haim, Jeremy Boger-Lombard, and Ori Katz. Image-guided computational holographic wavefront shaping. Nature Photonics, 2024.", + "[18] Karen Hampson, Raphael Turcotte, Donald Miller, Kazuhiro Kurokawa, Jared Males, Na Ji, and Martin Booth. Adaptive optics for high-resolution imaging. Nature Reviews Methods Primers, 1:68, 2021.", + "[19] John W. Hardy. Adaptive Optics for Astronomical Telescopes. 1998.", + "[20] Yi Hua, M. Salman Asif, and Aswin C. Sankaranarayanan. Spatial and axial resolution limits for mask-based lensless cameras. Opt. Express, 31(2):2538-2551, 2023.", + "[21] Seungwon Jeong, Ye-Ryoung Lee, Wonjun Choi, Sungsam Kang, Jin Hee Hong, Jin-Sung Park, Yong-Sik Lim, HongGyu Park, and Wonshik Choi. Focusing of light energy inside a scattering medium by controlling the time-gated multiple light scattering. Nature Photonics, 12(5):277-283, 2018.", + "[22] Na Ji. Adaptive optical fluorescence microscopy. Nature Methods, 14:374-380, 2017.", + "[23] Yonghyeon Jo, Ye-Ryoung Lee, Jin Hee Hong, Dong-Young Kim, Junhwan Kwon, Myunghwan Choi, Moonseok Kim, and Wonshik Choi. Through-skull brain imaging in vivo at visible wavelengths via dimensionality reduction adaptive-optical microscopy. Science Advances, 8(30):eabo4366, 2022.", + "[24] Andrew Jr, Alon Harris, Josh Gross, Ingrida Januleviciene, Aaditya Shah, and Brent Siesky. Optical coherence tomography angiography: An overview of the technology and an assessment of applications for clinical research. British Journal of Ophthalmology, 101, 2016.", + "[25] Sungsam Kang, Pilsung Kang, Seungwon Jeong, Yongwoo Kwon, Taeseok D. Yang, Jin Hee Hong, Moonseok Kim, Kyung-Deok Song, Jin Hyoung Park, Jun Ho Lee, Myoung Joon Kim, Ki Hean Kim, and Wonshik Choi. High-resolution adaptive optical imaging within thick scattering media using closed-loop accumulation of single scattering. Nature Communications, 8(1):2157, 2017.", + "[26] Sungsam Kang, Yongwoo Kwon, Hojun Lee, Seho Kim, Jin Hee Hong, Seokchan Yoon, and Wonshik Choi. Tracing multiple scattering trajectories for deep optical imaging in scattering media. Nature communications, 14(1), 2023.", + "[27] Marie Laslandes, Matthias Salas, Christoph K. Hitzenberger, and Michael Pircher. Increasing the field of view of adaptive optics scanning laser ophthalmoscopy. Biomed. Opt. Express, 8(11):4811-4826, 2017.", + "[28] V. Lauer. New approach to optical diffraction tomography yielding a vector equation of diffraction tomography and a novel tomographic microscope. Journal of Microscopy, 205 (2):165-176, 2002.", + "[29] F. C. Lin and M. A. Fiddy. The born-rytov controversy: I. comparing analytical and approximate expressions for the one-dimensional deterministic case. J. Opt. Soc. Am. A, 9 (7):1102–1110, 1992.", + "[30] Jerome Mertz. Introduction to optical microscopy. Cambridge University Press, 2019.", + "[31] Sagi Monin, Aswin Sankaranarayanan, and Anat Levin. Exponentially-wide etendue displays using a tilting cascade. IEEE ICCP, pages 1-12, 2022." + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23172", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Ulysse Najar, Victor Barolle, Paul Balondrade, Mathias Fink, Claude Boccara, and Alexandre Aubry. Harnessing forward multiple scattering for optical imaging deep inside an opaque medium. Nature Communications, 15(1), 2024.", + "[33] Tan Nguyen, Mikhail Kandel, Marcello Rubessa, Matthew Wheeler, and Gabriel Popescu. Gradient light interference microscopy for 3d imaging of unlabeled specimens. Nature Communications, 8, 2017.", + "[34] Galya Stern and Ori Katz. Noninvasive focusing through scattering layers using speckle correlations. Optics Letters, 44(1):143, 2018.", + "[35] Jörgen Thaung, Per Knutsson, Zoran Popovic, and Mette Owner-Petersen. Dual-conjugate adaptive optics for wide-field high-resolution retinal imaging. Opt. Express, 17(6): 4454-4467, 2009.", + "[36] Tsai wei Wu and Meng Cui. Numerical study of multiconjugate large area wavefront correction for deep tissue microscopy. Opt. Express, 23(6):7463-7470, 2015.", + "[37] Gil Weinberg, Elad Sunray, and Ori Katz. Noninvasive megapixel fluorescence microscopy through scattering layers by a virtual incoherent reflection matrix. Science Advances, 10(47):eadl5218, 2024.", + "[38] Emil Wolf. Three-dimensional structure determination of semi-transparent objects from holographic data. Optics Communications, 1(4):153-156, 1969.", + "[39] Tomer Yeminy and Ori Katz. Guidestar-free image-guided wavefront shaping. Science Advances, 7(21):eabf5364, 2021.", + "[40] Lei Zhu, Fernando Soldevila, Claudio Moretti, Alexandra d'Arco, Antoine Boniface, Xiaopeng Shao, Hilton B. de Aguiar, and Sylvain Gigan. Large field-of-view non-invasive imaging through scattering layers using fluctuating random illumination. Nature Communications, 13:1447, 2022." + ], + "bbox": [ + 91, + 92, + 480, + 558 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23173", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Understanding Multi-layered Transmission Matrices/0c5698ec-c1a0-4efc-b1d6-15a1169745e5_model.json b/2025/Understanding Multi-layered Transmission Matrices/0c5698ec-c1a0-4efc-b1d6-15a1169745e5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..25a2a3f424e3c406540a513eba3f406db1d4bc27 --- /dev/null +++ b/2025/Understanding Multi-layered Transmission Matrices/0c5698ec-c1a0-4efc-b1d6-15a1169745e5_model.json @@ -0,0 +1,2266 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.239, + 0.131, + 0.761, + 0.153 + ], + "angle": 0, + "content": "Understanding multi-layered transmission matrices" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.174, + 0.748, + 0.227 + ], + "angle": 0, + "content": "Anat Levin and Marina Alterman \nDepartment of Electrical and Computer Engineering, Technion Haifa, Israel" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.262, + 0.327, + 0.279 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.295, + 0.483, + 0.461 + ], + "angle": 0, + "content": "Wavefront shaping systems attempt to correct aberrations caused by tissue scattering, by placing a spatial light modulator (SLM) in the optical path and using it to reshape the light wave emitted from a target of interest deep inside the tissue. However, the field-of-view we can correct with one SLM pattern is extremely small, because inherently the aberration is caused by the 3D tissue structure while the SLM is planar. To overcome this, multi-conjugate correction systems built with multiple layered SLMs have been introduced, which attempt to approximate the 3D tissue structure with multiple planar aberrations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.461, + 0.483, + 0.718 + ], + "angle": 0, + "content": "However, multi-conjugate systems with a large number of SLMs are not easy to construct. Therefore, it is important to understand how many SLM layers are actually needed for a good correction, and whether there are practical benefits from correction systems with a relatively small number of layers. To help in the design of future systems, this paper analyzes multi-layer corrections. We show that the well-known missing cone problem which fundamentally limits the axial resolution of microscopes, turns into an advantage when designing 3D correction systems. Due to this property less information should be captured by the correction system, facilitating sparser approximations. We also show that even if the number of available layers is insufficient for a reasonable approximation of the full 3D tissue volume, it can significantly expand the field-of-view when compared to single-layer correction systems, and therefore can largely accelerate their operation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.731, + 0.222, + 0.745 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.885 + ], + "angle": 0, + "content": "Scattering is a central challenge in multiple imaging tasks ranging from astronomy to microscopy, and biomedical imaging. It degrades our ability to resolve stars through telescopes, image into the bottom of the sea, and drive through foggy weather. Most importantly in the context of this paper, scattering severely limits our ability to image through biological tissue. This happens due to small variations in the refractive index of the tissue volume, which scatters the propagating light." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Numerous computational optics approaches have been" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.263, + 0.908, + 0.717 + ], + "angle": 0, + "content": "developed for imaging through scattering tissue. These include computational restoration algorithms which attempt to fit the observed data with various models for the aberration and the target of interest [4, 5, 15, 17, 21, 23, 25, 32, 37, 39, 40]. Also, optical systems such as a confocal microscope [14] or an OCT [24], aim to filter out the scattered light and maintain only the ballistic unscattered part. However, as we attempt to see deeper into a scattering layer, ballistic light weakens significantly, and the complexity of aberration models increases. Moreover, biological target of interest such as fluorescent tissue components are very weak and as a result the scattered signal is also of a very low signal to noise ratio (SNR), further limiting digital correction models. A particularly attractive approach for scattering aberration removal is wavefront shaping, where one places a spatial light modulator (SLM) in the optical path of the microscope and uses it to correct the wavefront emitted from a point deep inside the tissue, so that light photons emerging from a single target point are brought into a single sensor point, despite the tissue aberration. Wavefront shaping is a particularly attractive solution for fluorescent imaging because the correction is done optically rather than digitally. Unlike ballistic filtering approaches, all light photons are used. Since the correction is done in the optical path, the limited number of available photons is brought into a single detector pixel and can be measured with higher SNR. As a result, wavefront shaping has the potential to revolutionize tissue imaging and allow us to overcome the fundamental limitations of scattering, and image very deep inside biological targets at a high SNR." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "The main limitation for wavefront shaping approaches is that the aberrations of neighboring points inside volumetric tissue vary rapidly, and in practice a single correction mask can only allow imaging of a small local field-of-view. This is because the scattering process happens throughout the 3D volume, while the aberration correction only happens on a planar SLM. Earlier demonstrations of wavefront shaping have considered targets placed at a distance behind a thin scattering layer, and this far-field arrangement has allowed them to correct larger regions with a single planar modulation [12, 15, 34, 39]. Similarly in the context of adaptive-optics, researchers have attempted to correct aber" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23164" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "rations due to imperfect optics or a refractive index mismatch [8, 18, 22]. These aberrations are introduced along the optical path, before the tissue, so the correction can apply to a larger area. However, in most biomedical applications the fluorescent target of interest is embedded within the scattering volume. If our goal is to correct scattering within the tissue itself, the field-of-view which can be corrected by a single planar modulation is extremely small, usually only a few microns [3]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.231, + 0.485, + 0.562 + ], + "angle": 0, + "content": "To correct larger fields of view, multi conjugate correction systems have been proposed. Such systems use multiple planar SLMs placed such that each of them is conjugate to a different plane inside the tissue volume. Each such SLM should realize a correction for the aberration happening inside a thin layer of the volume and the combination of such multi-plane corrections should approximate the 3D variation in refractive index, and allow us to correct it. Fig. 1 visualizes the concept of a multi-conjugate correction [16, 27, 35, 36]. In practice, multi-plane correction systems are more complicated than the simplified visualization of Fig. 1, as one needs to add multiple relay systems between the SLMs. This is first because there are physical limitations on the minimal distance between 2 SLMs, and second because we need to revert the order of the correction layers in the inversion path [19]. This makes multiconductor correction systems complex and cumbersome to build. Previous experimental demonstrations have realized corrections with no more than two SLMs. While two-layer correction systems have demonstrated improved performance compared to single layer correction, two planes clearly cannot model a thick 3D tissue structure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.564, + 0.483, + 0.686 + ], + "angle": 0, + "content": "Our goal in this research is to analytically study multislice models and attempt to answer the following questions: How many planes are required to approximate the 3D aberration of a scattering tissue volume? How does the quality of the correction degrade if we use a sparser approximation? If a construction of a practical correction system is limited to only to M layers, what is the maximum field of view that can be effectively corrected?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.483, + 0.779 + ], + "angle": 0, + "content": "To this end, we use the mathematical model of a transmission matrix, which stores the propagation of coherent wavefronts from one end of a scattering volume to the other. We study how well we can approximate such transmission matrices with a multi-slice model, and how does the error depend on the number of fitted layers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The straightforward answer to this question is to say that the spacing between layers should match the axial resolution of the microscope. For example, using an \\(\\mathbf{NA} = 0.5\\) objective at wavelength \\(500nm\\), we get that the spacing between layers should be lower than \\(2\\mu m\\). Handling a modest tissue thickness of, e.g., \\(200\\mu m\\) already requires 100 slices for a valid approximation. Clearly building a multiconjugate wavefront shaping systems with so many layers" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.08, + 0.912, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.334, + 0.908, + 0.501 + ], + "angle": 0, + "content": "Figure 1. Layered wavefront shaping: (Top) Consider a simple microscope setup where we use a 4F relay to image the light emerging from a tissue sample. Light is aberrated on its way to the sensor, and we see speckle patterns rather than sharp spots. (Middle) A single planar SLM can correct aberration for one spot (see red beam), but light emerging from nearby spots may not be corrected (the yellow beam generates a speckle pattern on the sensor). (Bottom) The field of view of the correction can possibly be extended with a multi-layer correction. The goal is to form a correction replicating the 3D structure of volumetric aberration inside the tissue. The figure illustrates two wavefronts from nearby spots (red and yellow) passing through the same volumetric correction, and both focus into a sharp spot behind the tissue." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.508, + 0.611, + 0.522 + ], + "angle": 0, + "content": "is impractical." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.907, + 0.763 + ], + "angle": 0, + "content": "In this research we take inspiration from [20], and offer a thorough analysis of the frequency content of transmission matrices, combining the forward scattering nature of biological tissue with the structure of the missing cone. We show that while the missing cone structure is known as one of the fundamental limitations of 3D microscopy, in the context of transmission matrices this limitation turns into an advantage. Due to the missing cone structure, many of the frequencies of the volumetric aberration do not participate in the transmission matrix, and therefore the number of layers required for approximating a transmission matrix is way smaller than the Nyquist limit. However, there are no free lunches and the number of layers scales with the tissue thickness. To explain a few hundred micron thick tissue one needs a few dozen layers, which is still too much to be realized by a physical correction system." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Given this conclusion we aim to understand if there are any practical benefits in realizing a multi-conjugate correction system with a small number of say 2 - 4 layers. We show that while a small number of layers cannot well approximate the 3D structure of the tissue, it can allow us to realize a correction over a limited field-of-view. The area which can be corrected with \\(M\\) aberration layers is significantly larger than the area which can be corrected with one layer, and therefore it can significantly accelerate sequen" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23165" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.506, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.282, + 0.171, + 0.356, + 0.184 + ], + "angle": 0, + "content": "(a) Full setup" + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.094, + 0.631, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.171, + 0.629, + 0.183 + ], + "angle": 0, + "content": "(b) Fourier illum." + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.094, + 0.741, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.171, + 0.733, + 0.183 + ], + "angle": 0, + "content": "(c) Point illum." + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.095, + 0.857, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.171, + 0.862, + 0.184 + ], + "angle": 0, + "content": "(d) Layered approx." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.19, + 0.907, + 0.252 + ], + "angle": 0, + "content": "Figure 2. Setup for transmission matrix acquisition: (a) A setup for transmission matrix imaging where a set of basis illuminations at the back of the sample propagate through the volume, and the scattered wavefronts are measured by a camera at the other end. In (b,c) we zoom on the sample illustrating two possible illumination bases. (b) A set of directional wavefronts (Fourier basis). (c) A set of wavefronts focused to points at the edge of the volume (primal basis). Both transmission matrices can be related by a simple Fourier transform. (d) A multi-slice model approximates the volumetric scattering with a set of planar aberrations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.28, + 0.483, + 0.34 + ], + "angle": 0, + "content": "tial correction of the field of view. We validate our findings using real transmission matrices captured in the lab as well as synthetic transmission matrices simulated using accurate wave propagation models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.351, + 0.292, + 0.365 + ], + "angle": 0, + "content": "2. Problem formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.37, + 0.484, + 0.521 + ], + "angle": 0, + "content": "Imaging setup: For simplicity, the following derivation considers coherent scalar waves, ignoring polarization. In Fig. 2(a) we visualize a setup for transmission matrix acquisition. A scattering medium, such as tissue, is illuminated by a coherent source \\(\\mathbf{e}^k\\) at one side of the medium. After propagating through the medium, the resulting coherent wavefront \\(\\mathbf{t}^k\\) is recorded. Usually \\(\\mathbf{t}^k\\) are complex valued speckle patterns with a pseudo-random structure. The illumination is translated and multiple wavefronts are recorded and stored as columns of a transmission matrix \\(\\mathcal{T}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The transmission matrix: Since the propagation of coherent light is linear, if we want to predict the propagation of any wavefront \\(\\mathbf{u}\\), we simply express it as a superposition of the recorded incoming illuminations \\(\\mathbf{u} = \\sum_{k} \\mathbf{u}(k) \\mathbf{e}^{k}\\), where \\(\\mathbf{u}(k)\\) are complex scalars, and multiply it by the transmission matrix, to predict \\(\\mathcal{T} \\mathbf{u}\\). The illumination wavefronts \\(\\mathbf{e}^{k}\\) can be selected at different bases. For example, the illumination can be translated at the Fourier plane, and then the wavefronts illuminating the sample are plane waves at different angles (Fig. 2(b)). Alternatively one can focus the light to a point at the back plane of the sample (Fig. 2(c)). This configuration is more appropriate for wavefront shaping algorithms whose goal is to recover wavefronts focusing inside the tissue and not far behind it. In both cases, however, the illuminations we can create are limited by the numerical aperture of the illuminating objective, which we mark here as NA. Similarly, the scattered wavefronts are measured via an imaging objective with the same bounded NA. This NA would serve as a central component in the results derived below. For simplicity of resulting equations, for the rest of this manuscript we assume the imaging and illumination objectives are focused at the same plane, such that if we would remove the scattering sample and image the sources directly, we would see the \\(\\mathbf{e}^{k}\\) illuminations as sharp spots on the sensor plane. That is, if we use directional" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.28, + 0.906, + 0.37 + ], + "angle": 0, + "content": "illuminations, we also measure the Fourier transform of the scattered waves, namely place the sensor at the Fourier plane of the imaging objective. If we use point sources at the illumination end of the medium, we also focus the imaging objective at the same plane, so we measure a speckle image of a relatively small support." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.37, + 0.906, + 0.454 + ], + "angle": 0, + "content": "Range and resolution: We assume that in the primal basis the illuminators are spaced within an area of \\(\\Omega_p \\times \\Omega_p\\), and we sample them at the Nyquist resolution, so for an illumination wavelength \\(\\lambda\\) they are separated by a pitch \\(\\Delta_p = \\lambda/(2\\mathbf{NA})\\). Equivalently, in the Fourier domain, we illuminate the sample with plane waves of the form" + }, + { + "type": "equation", + "bbox": [ + 0.657, + 0.456, + 0.905, + 0.473 + ], + "angle": 0, + "content": "\\[\n\\mathbf {e} (\\mathbf {r}) = e ^ {2 \\pi i (\\bar {\\kappa} \\cdot \\mathbf {r})} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.906, + 0.629 + ], + "angle": 0, + "content": "where \\(\\mathbf{r}\\) is a point inside the 3D volume, and \\(\\bar{\\kappa} = \\frac{1}{\\lambda}\\bar{\\omega}\\) where \\(\\lambda\\) is the wavelength and \\(\\bar{\\omega}\\) is a unit norm direction vector, i.e. the last entry of the direction vector \\(\\omega_{z}\\) is \\(\\omega_{z} = \\sqrt{1 - (\\omega_{x}^{2} + \\omega_{y}^{2})}\\). For simplicity of notation, we assume the refractive index (RI) of the leading medium is 1. We can only use illumination direction inside the numerical aperture \\(\\sqrt{\\omega_x^2 + \\omega_y^2}\\leq \\mathbf{N}\\mathbf{A}\\) (or equivalently \\(\\sqrt{\\kappa_x^2 + \\kappa_y^2}\\leq \\Omega_f\\) with \\(\\Omega_{f} = \\mathbf{N A} / \\lambda\\)). To sample the transmission matrix at the Nyquist pitch, the \\(\\kappa_{x},\\kappa_{y}\\) components of the illumination are sampled by the Nyquist pitch \\(\\Delta_{f} = 1 / 2\\Omega_{p}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.748 + ], + "angle": 0, + "content": "With the above sampling scheme, the transmission matrices at the Fourier and primal bases are related by a simple Fourier transform \\(\\mathcal{T}^p = \\mathcal{F}^T\\mathcal{T}^f\\mathcal{F}\\). While for wavefront shaping we are interested in primal transmission matrices, the Fourier ones simplify analysis. Importantly, since Fourier and primal transmission matrices are related by an orthonormal transformation, the fitting errors in both bases are equivalent." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.906, + 0.873 + ], + "angle": 0, + "content": "The multi-slice model: The relation between an incoming illumination and an outgoing one is determined by the solution to a wave equation which takes into account the variation in refractive index (RI) inside the sample. Rather than a fully continuous model, in a multi-slice model we approximate the volume with a set of planar aberrations, separated by small \\(z\\) distances \\(\\Delta_z\\), where we have free space propagation between layers, as in Fig. 2(d). Mathematically this can be described as" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.878, + 0.905, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mathcal {T}} \\left(\\rho_ {1}, \\dots , \\rho_ {M}\\right) = \\boldsymbol {\\mathcal {P}} _ {O} \\boldsymbol {\\mathcal {D}} \\left(\\rho_ {M}\\right) \\dots \\boldsymbol {\\mathcal {P}} _ {\\Delta z} \\boldsymbol {\\mathcal {D}} \\left(\\rho_ {2}\\right) \\boldsymbol {\\mathcal {P}} _ {\\Delta z} \\boldsymbol {\\mathcal {D}} \\left(\\rho_ {1}\\right), \\tag {2}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23166" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.152, + 0.089, + 0.245, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.268, + 0.245, + 0.278 + ], + "angle": 0, + "content": "(a) Ground truth" + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.101, + 0.348, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.269, + 0.268, + 0.339, + 0.278 + ], + "angle": 0, + "content": "(b) 32 layers" + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.101, + 0.447, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.268, + 0.433, + 0.278 + ], + "angle": 0, + "content": "(c) 8 layers" + }, + { + "type": "image", + "bbox": [ + 0.455, + 0.101, + 0.545, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.268, + 0.532, + 0.278 + ], + "angle": 0, + "content": "(d) 4 layers" + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.1, + 0.842, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.668, + 0.267, + 0.765, + 0.278 + ], + "angle": 0, + "content": "(e) Fitting curves" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.288, + 0.907, + 0.361 + ], + "angle": 0, + "content": "Figure 3. Reconstructed volumes: (a) An xz slice from a toy volume, visualizing its refractive index variations. (b-d) Reconstructions with varying layer numbers. The columns of the transmission matrix provide an under-constraint inversion problem, and even when a dense set of layers is allocated in b, the axial resolution of the reconstructed volume is low. As a result the sparse reconstructions in c,d can also well explain the transmission matrix. (e) Numerical evaluation of fitting error as a function of the number of layers. Transmission matrices can be well fitted even if the allocated number of layers is lower than Nyquist predictions. However the required number of layers increases with the tissue thickness. The layer number is less affected when the optical density is increased for the same tissue thickness." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.393, + 0.483, + 0.512 + ], + "angle": 0, + "content": "where \\(\\rho_{m}(x,y)\\) is the 2D aberration of the \\(m\\)th layer and \\(\\mathcal{D}(\\rho_m)\\) is a diagonal matrix performing element-wise multiplication of the incoming wavefront by the aberration. \\(\\mathcal{P}_{\\Delta_z}\\) describes free space propagation between layers at distance \\(\\Delta_z\\), which is essentially a convolution with a spherical wavefront. Finally \\(\\mathcal{P}_O\\) propagates the wavefront in free space from the tissue boundary, via the lenses, to the camera sensor." + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.514, + 0.483, + 0.559 + ], + "angle": 0, + "content": "Our goal is to understand what plane spacing \\(\\Delta_z\\) will facilitate a good representation of \\(\\mathcal{T}\\), and how this approximation would degrade when we use sparser sampling." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.56, + 0.444, + 0.572 + ], + "angle": 0, + "content": "A common approximation for weak aberrations is:" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.573, + 0.482, + 0.588 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\rho} _ {m} (x, y) = \\mu + \\delta \\boldsymbol {\\rho} _ {m} (x, y), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.593, + 0.483, + 0.718 + ], + "angle": 0, + "content": "where \\(\\mu\\) is a scalar whose magnitude is a bit smaller than 1, encoding the ballistic portion of the light which passes through the layer without aberration, and the magnitude of the aberration is significantly smaller than the ballistic term \\(|\\delta \\rho_{m}(x,y)|\\ll |\\mu |\\) . For weak scattering most of the energy does not scatter more than once. It is therefore common to consider a first order approximation to the transmission matrix (often known as the single scattering, first-Born or Rytrov approximation [9, 10, 13, 29, 33]), which reduces to" + }, + { + "type": "equation", + "bbox": [ + 0.087, + 0.72, + 0.483, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mathcal {T}} \\left(\\delta \\rho_ {1}, \\dots , \\delta \\rho_ {M}\\right) \\approx \\mu^ {M} \\boldsymbol {\\mathcal {P}} _ {O} \\boldsymbol {\\mathcal {P}} _ {M \\Delta_ {z}} + \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.739, + 0.482, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\mu^ {M - 1} \\sum_ {m = 1} ^ {M} \\mathcal {P} _ {O} \\mathcal {P} _ {(M - m) \\Delta_ {z}} \\mathcal {D} (\\delta \\rho_ {m}) \\mathcal {P} _ {m \\Delta_ {z}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.871 + ], + "angle": 0, + "content": "The full light propagation in Eq. (2) includes multiplication of the different aberrations and hence the transmission matrix is highly nonlinear. In contrast, the advantage of the weakly scattering model in Eq. (4) is that the resulting transmission matrix is a linear function of the aberration layers. This linear approximation significantly simplifies analysis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.899 + ], + "angle": 0, + "content": "Diffraction tomography algorithms such as [11] capture a subset of columns from the transmission matrix and use" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.393, + 0.907, + 0.514 + ], + "angle": 0, + "content": "gradient descent optimization to seek layers \\(\\rho_1, \\ldots, \\rho_M\\) which best explain the measured wavefronts, assuming that the propagation operators \\(\\mathcal{P}_{\\Delta_z}, \\mathcal{P}_O\\) are known. We will use a similar optimization scheme below, when trying to find a multi-slice representation for a given transmission matrix. To correct aberrations, a physical multi-conjugate wavefront shaping system would display on the SLM the conjugate of the recovered layers." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.528, + 0.857, + 0.546 + ], + "angle": 0, + "content": "3. Evaluating multi-slice approximations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We start with numerical experiments demonstrating the main findings. In Sec. 4 below we derive a theory explaining them. To gain intuition we generated a toy volume composed of spheres with varying refractive indices illustrated in Fig. 3(a), and obtained the corresponding transmission matrix by propagating a set of point sources through the volume. We then used the transmission matrix to solve for a multi-slice approximation, while varying the number of layers. In Fig. 3(b-d) we visualize some of the volumetric reconstructions, assuming a \\(\\lambda = 500nm\\) illumination. Even when the fitted layers are sampled very densely at an interval of \\(\\Delta_z = 0.5\\mu m\\), we cannot fully recover the target and the fitted volume has a low axial resolution. We analyze this in Sec. 4 below and explain that this results from the well known missing cone problem, a fundamental degeneracy limiting the axial resolution of a microscope. Said differently the transmission matrix, or the full set of propagating waves, only provides a set of under-constraint measurements from which the volume cannot be uniquely recovered. While the missing cone problem is considered a fundamental limitation in microscopy, we argue that in the context of wavefront shaping it actually turns into an advantage. Since the volume reconstruction is not unique we can" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23167" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.094, + 0.199, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.267, + 0.193, + 0.277 + ], + "angle": 0, + "content": "(a) Ground truth" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.096, + 0.304, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.267, + 0.28, + 0.277 + ], + "angle": 0, + "content": "(b) Dense fit" + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.096, + 0.396, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.308, + 0.268, + 0.374, + 0.278 + ], + "angle": 0, + "content": "(c) Sparse fit" + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.096, + 0.5, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.268, + 0.477, + 0.278 + ], + "angle": 0, + "content": "(d) Dense fit" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.096, + 0.592, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.505, + 0.267, + 0.571, + 0.278 + ], + "angle": 0, + "content": "(e) Sparse fit" + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.104, + 0.891, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.722, + 0.267, + 0.808, + 0.278 + ], + "angle": 0, + "content": "(f) Fitting curves" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.288, + 0.907, + 0.351 + ], + "angle": 0, + "content": "Figure 4. Reconstructed volumes with varying supports: (a) An xz slice from a toy volume, visualizing its refractive index variations. (b,d) Two reconstructions of the volume using a dense layer sampling. When the support of the measured window \\(\\Omega_{p}\\) shrinks, the lateral extent of the reconstructed region decreases, however, its axial resolution is also reduced. (c,e) A sparse fit with the same input. Despite the low quality of these reconstructions, they predict well the columns of the transmission matrix of the given support \\(\\Omega_{p}\\). (f) Numerical evaluation of fitting error as a function of the number of layers, while varying the range \\(\\Omega_{p}\\) covered by the transmission matrix. Smaller supports can be fitted with fewer layers." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.381, + 0.483, + 0.41 + ], + "angle": 0, + "content": "approximate the same transmission matrix with a sparser set of layers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.483, + 0.451 + ], + "angle": 0, + "content": "To quantify the fitting error of the multi-slice model, we denote by \\(\\pmb{\\varepsilon}_{M}\\) the reconstruction error of a given transmission matrix with the best \\(M\\) layers" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.452, + 0.483, + 0.475 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mathcal {E}} _ {M} = \\min _ {\\rho_ {1}, \\dots , \\rho_ {M}} \\| \\boldsymbol {\\mathcal {T}} _ {\\text {e x a c t}} - \\boldsymbol {\\mathcal {T}} (\\rho_ {1}, \\dots , \\rho_ {M}) \\| ^ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.478, + 0.483, + 0.688 + ], + "angle": 0, + "content": "In Fig. 3(e) we evaluate \\(\\pmb{\\mathcal{E}}_M\\) as a function of the layer number \\(M\\). Even the seemingly aliased fit of Fig. 3(d) still provides a low fitting error. However, there are no free lunches and when the number of layers is too low the fitting error is high. When we repeat the simulation with a thicker volume we need even more layers for a good fit. Interestingly, we also repeat the simulation with denser volumes (larger optical depth, or shorter mean free path) of the same thickness, and this has less effect on the required number of layers. In our plots, the case \\(M = 0\\) refers to no correction at all, so effectively we assume that the transmission matrix is diagonal. The similarity between a diagonal transmission matrix and the target transmission matrix is a measure of the amount of ballistic light." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.689, + 0.483, + 0.854 + ], + "angle": 0, + "content": "The thickness of the synthetic simulation here is limited due to computational constraints. However, usually wavefront shaping systems attempt to see into a few hundred microns of tissue and we anticipate that for such volumes we need at least a few dozen layers for a good fit. Clearly realizing a physical wavefront shaping system with that many layers is impractical. We argue that while a small number of layers is not enough for focusing light anywhere across the tissue area, it can still help us to focus light over a compact support area, and this area is significantly wider than what we can correct with a single layer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Compact support: To test the effect of the target area, we consider transmission matrices measured with illumination points confined to a limited area of size \\(\\Omega_p \\times \\Omega_p\\). If we can" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.905, + 0.457 + ], + "angle": 0, + "content": "fit such a transmission matrix with \\(M\\) layers and place the values of these \\(M\\) layers on \\(M\\) SLMs, the resulting wavefront shaping system can correct light from the \\(\\Omega_p \\times \\Omega_p\\) area, but target points outside this area are unlikely to be corrected as they are not included in the fitting process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.907, + 0.717 + ], + "angle": 0, + "content": "To test the effect of support on the number of fitted layers, we return to the numerical simulation of Fig. 3, but fit subsets of columns contained in smaller \\(\\Omega_p\\) ranges. Indeed, smaller supports can be well-fitted with a smaller number of layers. As another way to understand it, we show in Fig. 4 an xz slice from the ground truth refractive index volume and from a few reconstructions. We show that even if we do not limit the layer number \\(M\\), the recovered RI volume has limited axial resolution and this axial resolution is lower when measured range \\(\\Omega_p\\) decreases. This explains why lower supports can be fitted with fewer layers. For each of the two supports we also show a sparse fit, with the minimal \\(M\\) value that provides a good prediction of the transmission matrix. We explain this observation in Sec. 4. This result is also in agreement with [26], who correct aberration over a relatively small field of view through a few hundred microns of tissue, using only 5 layers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Implications to layered wavefront shaping systems. In Fig. 4(f) we plotted a bar below which the approximation is reasonable. We can see that for e.g. the \\(40 \\times 40\\mu m\\) support we can obtain good focusing with about \\(M = 6\\) layers. It is likely that fitting a full field-of-view of a few hundred microns would require more layers. However, we can see that while with \\(M = 1\\) layer we could only focus over a small area of about \\(1 \\times 1\\mu m\\), when increasing to \\(M = 2\\) we could focus over a \\(\\times 4\\) larger area of support \\(2 \\times 2\\mu m\\), and with \\(M = 3\\) we could focus over a \\(\\times 25\\) larger area of support \\(5 \\times 5\\mu m\\). Computing multiple correction elements is only going to increase the complexity of the wavefront shaping" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23168" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.09, + 0.41, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.09, + 0.885, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.281, + 0.908, + 0.346 + ], + "angle": 0, + "content": "Figure 5. Layer support for a captured transmission matrix: We consider a transmission matrix captured in the lab through a chicken breast layer of thickness \\(170\\mu m\\). Left: plotting the fitting error as a function of the number of layers. Smaller supports can be fitted with fewer layers. Right: Visualization of a spot focusing behind the tissue computed using the wavefronts of the fitted model. We compare models fitted to different supports. In the top row we need a larger number of layers, but the spot can be scanned over a \\(25 \\times 25\\mu m\\) window using the same layers. In the lower rows we achieve a focused spot using a smaller number of layers, but these layers can scan the focused spot over smaller windows of sizes \\(6 \\times 6\\mu m\\) and \\(1 \\times 1\\mu m\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.371, + 0.482, + 0.522 + ], + "angle": 0, + "content": "algorithm in [1] linearly with the number of layers. Alternatively, a wavefront shaping system using a single SLM would need to scan the field-of-view sequentially and correct each local window independently. Since the increase in the support of the window we can explain with e.g. 3 layers is significantly larger than a \\(\\times 3\\) factor, a multi-conjugate wavefront shaping system with a small number of SLMs as in [16, 27, 31, 35] can already have a large practical benefit. While Fig. 4 only considers a simulation, we analyze realistic numbers below." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.483, + 0.569 + ], + "angle": 0, + "content": "Lab results: In the supplementary file we validate the theory with multiple real transmission matrices, representing thick volumes with multiple-scattering." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.659 + ], + "angle": 0, + "content": "We start with numerically simulated transmission matrices using a more accurate wave-propagation model [6, 7]. We then test lab-captured transmission matrices measuring realistic scattering samples including mouse brain tissue, chicken breast tissue and a parafilm layer. Please refer to supplementary file for additional results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.735 + ], + "angle": 0, + "content": "In Fig. 5 we show one fitting result using a multiply-scattering chicken breast layer of thickness \\( d = 170\\mu m \\). The transmission matrices cover an area with support \\( \\Omega_{p} \\times \\Omega_{p} = 13 \\times 13\\mu m \\). As in Fig. 4(f), we see that smaller spatial ranges can be fitted with fewer layers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.736, + 0.483, + 0.886 + ], + "angle": 0, + "content": "As another way to assess the quality of the fit, we used columns of the fitted matrices to focus light through the tissue and show the shape of the focused spot and visualize the focusing quality in Fig. 5. To simulate the focused spot we simply visualize the product \\(\\mathcal{T}_{\\mathrm{exact}} \\cdot \\mathbf{t}_{\\mathrm{fit,M}}^k\\), where \\(\\mathbf{t}_{\\mathrm{fit,M}}^k\\) is a column of the fitted matrix \\(\\mathcal{T}(\\rho_1, \\ldots, \\rho_M)\\), namely a wavefront generated by the layered approximation. Note that while we only show focusing at one point, the layers are optimized such that they allow us to focus at any point inside the \\(\\Omega_p \\times \\Omega_p\\) window." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.482, + 0.901 + ], + "angle": 0, + "content": "From the results in Fig. 5 we can see that for the \\(13\\times\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.371, + 0.907, + 0.493 + ], + "angle": 0, + "content": "\\(13\\mu m\\) support measured, we can obtain good focusing with about 3 layers. Fitting a full field-of-view of several hundred microns likely requires additional layers. However, we observe that while using \\(M = 1\\) allows us to focus on a small area of about \\(3\\times 3\\mu m\\), with \\(M = 3\\), we can focus on an area \\(\\times 18\\) larger, reaching \\(13\\times 13\\mu m\\). Therefore, using multiple layers can significantly accelerate a sequential scanning of a wave-front shaping system." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.503, + 0.895, + 0.52 + ], + "angle": 0, + "content": "4. Sampling theory for transmission matrices" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.906, + 0.57 + ], + "angle": 0, + "content": "Our goal in this section is to explain the empirical findings of the previous section by analyzing the information captured by a transmission matrix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.58, + 0.848, + 0.596 + ], + "angle": 0, + "content": "4.1. Nyquist sampling and the missing cone" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.906, + 0.688 + ], + "angle": 0, + "content": "We start by considering the case of weakly scattering volumes which, as in Eq. (4), can be reasonably approximated by considering the component of light that scattered once. We also assume the transmission matrix is sampled over a very wide support \\(\\Omega_{p}\\). We discuss the implications of multiple scattering and limited supports in subsequent sections." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.905, + 0.734 + ], + "angle": 0, + "content": "We denote by \\( \\pmb{n}(\\mathbf{r}) \\) the difference between the RI at a 3D position \\( \\mathbf{r} \\) in the volume to that of the leading medium around it. We denote its 3D Fourier transform by \\( \\hat{n}(\\kappa) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.906, + 0.831 + ], + "angle": 0, + "content": "We start by considering the Fourier representation of the transmission matrix, where it is illuminated and measured by plane waves. Using the weakly scattering (first-Born) approximation, the wavefront scattering toward direction \\(\\bar{\\omega}^v\\) when illuminated by an incoming plane wave at direction \\(\\bar{\\omega}^i\\) is proportional, up to a multiplicative factor, to [28, 30, 38]:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.837, + 0.905, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\propto \\int \\boldsymbol {n} (\\mathbf {r}) e ^ {2 \\pi i \\left(\\bar {\\kappa} ^ {i} - \\bar {\\kappa} ^ {v}\\right) \\cdot \\mathbf {r}} d \\mathbf {r}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.87, + 0.907, + 0.902 + ], + "angle": 0, + "content": "with \\(\\bar{\\kappa}^v = 1 / \\lambda \\bar{\\omega}^v\\), \\(\\bar{\\kappa}^i = 1 / \\lambda \\bar{\\omega}^i\\). This implies that effectively an entry of the transmission matrix is a sample" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23169" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.22, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.189, + 0.168, + 0.202 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.089, + 0.353, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.189, + 0.299, + 0.202 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.09, + 0.478, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.407, + 0.189, + 0.428, + 0.202 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.209, + 0.484, + 0.318 + ], + "angle": 0, + "content": "Figure 6. Spectrum structure: (a) An \\(xz\\) slice out of the spectrum of \\(\\hat{n} (\\omega)\\). Entries of a weakly-scattering transmission matrix limited by an aperture NA only lie inside the butterfly area. (b) Zooming on the center right area of (a) (purple square). Assuming the \\(\\omega_{z}\\) axis is cut at \\(\\pm \\frac{1}{2}\\Lambda^{M}\\), the transmission matrix entries inside the dashed blue area are maintained, and the entries in the dashed red area are lost. (c) The energy at a slice from the spectrum of the volume in Fig. 3(a). Since the volume structure is smooth, most energy is concentrated at the low frequencies. As a result the cropped red region in (b) do not contain a lot of energy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.484, + 0.549 + ], + "angle": 0, + "content": "from the 3D Fourier transform of the volume at frequency \\(\\kappa = \\bar{\\kappa}^i -\\bar{\\kappa}^v\\). Note that throughout this paper we use bars \\(\\bar{\\omega}\\), \\(\\bar{\\kappa}\\) to denote vectors with norm constraints, and use normal bold fonts to denote standard 3D vectors, such as frequencies \\(\\kappa\\) in the 3D spectrum. Unfortunately, a large portion of the 3D spectrum cannot be measured. Recalling that the illumination and viewing directions are limited by an aperture of width NA, the subset of 3D frequencies captured by the transmission matrix occupies only a butterfly shape, and the rest of the cone is missing, see illustration in Fig. 6(a). Using paraxial approximation it can be shown [28, 38] that the range of frequencies which can be generated is" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.558, + 0.485, + 0.598 + ], + "angle": 0, + "content": "\\[\n\\left| \\kappa_ {z} \\right| \\leq \\mathbf {N A} \\left| \\boldsymbol {\\kappa} _ {x y} \\right| - \\frac {\\lambda}{2} \\left| \\boldsymbol {\\kappa} _ {x y} \\right| ^ {2}, \\quad \\text {w i t h} \\quad \\left| \\boldsymbol {\\kappa} _ {x y} \\right| = \\sqrt {\\kappa_ {x} ^ {2} + \\kappa_ {y} ^ {2}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.484, + 0.732 + ], + "angle": 0, + "content": "The missing cone is a serious problem for diffraction tomography algorithms aiming to recover the 3D structure of the refractive index \\( n(\\mathbf{r}) \\) from columns of the transmission matrix, since many of the frequencies of the volume cannot be recovered. We argue that in the context of transmission matrices, the missing cone problem turns into an advantage. The fact that the transmission matrix only captures a limited range of frequencies allows us to approximate it with fewer samples, as derived below." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.484, + 0.904 + ], + "angle": 0, + "content": "Minimal sampling: Our goal is to approximate a transmission matrix with layers. We start by asking what spacing between layers will allow an exact reconstruction of the transmission matrix. If we approximate the volume \\( \\pmb{n}(\\mathbf{r}) \\) with a set of planes, we essentially sample a continuous signal. Classical Nyquist theory states that if we sample a signal at intervals of \\( \\Delta_z \\), its Fourier transform includes replicas spaced by \\( \\Lambda = 1 / \\Delta_z \\). To achieve a good reconstruction, we want replicas to happen only in areas of the spectrum that the transmission matrix does not sample. Since the \\( \\kappa_x, \\kappa_y \\) components of our illuminations are limited by the aperture \\( |\\kappa_{xy}| \\leq \\Omega_f \\), and given the butterfly shape in Eq. (7), the \\( \\kappa_z \\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.831, + 0.106 + ], + "angle": 0, + "content": "frequencies we need to maintain are bounded to" + }, + { + "type": "equation", + "bbox": [ + 0.66, + 0.109, + 0.905, + 0.137 + ], + "angle": 0, + "content": "\\[\n\\left| \\kappa_ {z} \\right| \\leq \\frac {\\mathrm {N A} \\Omega_ {f}}{2}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.139, + 0.905, + 0.168 + ], + "angle": 0, + "content": "Therefore, for an error-free reconstruction, the minimal Fourier range and primal spacing are" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.168, + 0.905, + 0.197 + ], + "angle": 0, + "content": "\\[\n\\Lambda^ {*} = \\mathbf {N A} \\boldsymbol {\\Omega} _ {f}, \\quad \\Delta_ {z} ^ {*} = \\frac {1}{\\mathbf {N A} \\boldsymbol {\\Omega} _ {f}} = \\frac {\\lambda}{\\mathbf {N A} ^ {2}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.2, + 0.905, + 0.227 + ], + "angle": 0, + "content": "For a scattering volume of thickness \\(d\\), the number of required layers is:" + }, + { + "type": "equation", + "bbox": [ + 0.659, + 0.223, + 0.905, + 0.25 + ], + "angle": 0, + "content": "\\[\nM ^ {*} = \\frac {d \\cdot \\mathbf {N A} ^ {2}}{\\lambda}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.25, + 0.907, + 0.34 + ], + "angle": 0, + "content": "For typical numbers, consider e.g. \\(\\lambda = 0.5\\mu m\\) and \\(\\mathbf{N}\\mathbf{A} = 0.5\\), the plane spacing should be relatively small, and \\(\\Delta_z \\leq 2\\mu m\\). For a modest tissue thickness of \\(d = 200\\mu m\\), we will need \\(M = 100\\) planes. Clearly it is impractical to realize such a large number of planes by a physical correction system." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.345, + 0.811, + 0.361 + ], + "angle": 0, + "content": "4.2. Sparse multi-slice approximations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.363, + 0.906, + 0.528 + ], + "angle": 0, + "content": "Consider a scattering sample of thickness \\(d\\), which we try to approximate using \\(M < M^{*}\\) layers separated by \\(\\Delta_z = \\frac{d}{M}\\). Our goal is to show that the fitting error \\(\\mathcal{E}_M\\) defined in Eq. (5) decays relatively fast with \\(M\\) and we can get a reasonable approximation to the transmission matrix even if the number of layers is lower than the exact prediction in Eq. (10). This is a result of two main properties: (i) The volumes describing realistic tissue samples have more energy in the low frequencies; and (ii) the structure of the missing cone implies that, in any case, a significant portion of the spectrum is not sampled by the transmission matrix." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.528, + 0.905, + 0.599 + ], + "angle": 0, + "content": "To understand this, consider a naive selection of \\(M\\) layers \\(\\rho_{1},\\ldots ,\\rho_{M}\\). Rather than actually solving an optimization problem, we use the ground truth volume \\(\\hat{n} (\\kappa)\\), and simply set to zero any frequency content of the RI volume \\(\\hat{n} (\\kappa)\\) at \\(\\kappa_z\\) values larger than the possible Nyquist range" + }, + { + "type": "equation", + "bbox": [ + 0.647, + 0.599, + 0.905, + 0.626 + ], + "angle": 0, + "content": "\\[\n\\Lambda^ {M} = \\frac {1}{\\Delta_ {z}} = \\frac {\\lambda M}{d}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.795 + ], + "angle": 0, + "content": "and we then Fourier transform \\(\\hat{n}(\\kappa)\\) to \\(\\pmb{n}(\\mathbf{r})\\) and sample planes at spacing \\(\\Delta_z = d / M\\). The error of this approximation is basically the integral of content above the cut-off \\(\\Lambda^M\\). Since according to Eq. (11), the cut-off frequency \\(\\Lambda^M\\) scales linearly with the number of layers, the portion of the spectrum which is lost by this low-pass operation scales linearly with the number of layers \\(M\\). Thus, the naive answer is that the error of a multi-slice approximation decays linearly with \\(M\\). In practice we show below that the error decays much faster, since large areas from the 3D spectrum of the sample are not used by the transmission matrix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To gain intuition, consider Fig. 6(b). We note that the butterfly shape is such that low 2D frequencies (i.e. low \\(|\\kappa_{xy}|\\)), also span less content along the \\(\\kappa_z\\) axes, and hence these frequencies are not lost even with low bandwidths \\(\\Lambda^M\\). For the higher \\(|\\kappa_{xy}|\\) frequencies, the lower \\(\\kappa_z\\) part, marked in dashed blue in Fig. 6(b) is preserved, while the higher portion marked in red is lost." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23170" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.208, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.213, + 0.209, + 0.225 + ], + "angle": 0, + "content": "(a) Support primal" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.089, + 0.331, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.237, + 0.213, + 0.326, + 0.224 + ], + "angle": 0, + "content": "(b) Support freq." + }, + { + "type": "image", + "bbox": [ + 0.344, + 0.089, + 0.462, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.35, + 0.213, + 0.455, + 0.224 + ], + "angle": 0, + "content": "(c) Local directions" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.234, + 0.485, + 0.32 + ], + "angle": 0, + "content": "Figure 7. Layer support: (a) As point sources expand while propagating through the volume, the width of the aberration layers should be wider than the measured support \\(\\Omega_{p}\\). (b) To express compact support transmission matrices in the Fourier basis, we illuminate the volume by a set of plane waves passing through an aperture. (c) The local cone of illumination angles reaching different points inside the volume is much smaller than the actual range of incoming illuminations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.484, + 0.478 + ], + "angle": 0, + "content": "Another important property of tissue is that its RI is locally smooth, and therefore if we look at its spectrum (Fig. 6(c)), we have much more energy in low frequencies than in high ones. This also agrees with the fact that tissue is forward scattering. Therefore, in the red areas of Fig. 6(b) that are not sampled by the transmission matrix, there is less energy than in the lower frequencies. Thus, despite the fact that the multi-slice approximation sacrifices such frequencies, not much energy is being lost." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.48, + 0.484, + 0.525 + ], + "angle": 0, + "content": "In most cases, it is hard to give analytic formulas for the decay of \\(\\pmb{\\mathcal{E}}_M\\) as a function of \\(M\\), but the supplementary file contains some derivation under simplifying assumptions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.527, + 0.484, + 0.587 + ], + "angle": 0, + "content": "Multiple scattering. The results in Fig. 3(e) include multislice fits with both lower and higher optical depth values up to \\(OD = 4\\), showing that the fitting error still decays rapidly with the number of layers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.588, + 0.484, + 0.816 + ], + "angle": 0, + "content": "While it is hard to give analytical predictions for the multiple scattering case, it appears that when the optical depth of the tissue is not too high, similar results hold. The reason is that for a forward-scattering material such as tissue, light paths only scatter at small angles, and after a small number of scattering events they are not completely scrambled. For an intuition, in the supplementary file we use a Monte-Carlo simulation illustrating that for a forward-scattering material, after a modest number of scattering events, most light paths still remain within the butterfly area of the spectrum. This is usually the regime that wavefront-shaping algorithms like [1, 2] attempt to tackle, as such algorithms attempt to push the depth at which conventional microscopes can see, but they do not yet attempt to image extremely deep where scattering is fully diffused." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.82, + 0.332, + 0.836 + ], + "angle": 0, + "content": "4.3. Bounded support matrices" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.886 + ], + "angle": 0, + "content": "As stated above, much fewer layers are needed if we only attempt to fit transmission matrices sampled over a small area of size \\(\\Omega_p\\times \\Omega_p\\)" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "When fitting layers, we need to account for the fact that a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.38 + ], + "angle": 0, + "content": "point/plane illumination scattering through a volume is expanding. Thus, the aberration layers required to explain the transmission matrix are wider than the imaged area of size \\(\\Omega_{p} \\times \\Omega_{p}\\), see Fig. 7(a). Expressing a transmission matrix with a compact support at the frequency basis is equivalent to illuminating the sample with plane waves spanning a wide set of angles, but all the waves pass through a narrow aperture of size \\(\\Omega_{p} \\times \\Omega_{p}\\). In this case most points inside the volume usually do not receive light from all angles. In Fig. 7(c) we plot the narrow cone of angles arriving at three different points (a ray leaving a point \\(\\mathbf{p}_{\\mathbf{j}}\\) at angle \\(\\bar{\\omega}\\) is included in the cone only if this angle is not cropped by the aperture at the back of the sample). Since the cone of light reaching each point is narrower than the full numerical aperture, a local Fourier transform would not have content over the full butterfly area, but cover an even lower range of axial \\(\\omega_{z}\\) frequencies. Hence, following the analysis in the previous section, it can be sampled with fewer layers. We visualize the local spectra in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.395, + 0.628, + 0.411 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.907, + 0.584 + ], + "angle": 0, + "content": "This paper studies how well we can fit transmission matrices with multi-slice models and how many layers are needed in practice. We show that the required number of layers is much lower than the prediction of Nyquist sampling theory. This is because the missing cone problem of 3D microscopy turns into an advantage when analyzing transmission matrices. Moreover, as tissue is forward scattering, its 3D spectrum contains most energy in the low frequencies and much less energy in the higher ones. Hence, even if the sparse sampling acts as a low-pass on the large axial frequencies, not much energy is lost." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.587, + 0.906, + 0.678 + ], + "angle": 0, + "content": "We validate this prediction using both synthetic simulations and realistic lab measurements. We show that even though realistic transmission matrices contain backward and side-ward scattering which are not modeled by multi-slice matrices, they provide a reasonable approximation that is good enough for wavefront shaping focusing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.681, + 0.906, + 0.817 + ], + "angle": 0, + "content": "Our findings suggest that multi-slice models can be applicable for the design of non-local wavefront shaping systems, allowing for a volumetric correction with a small number of layered SLMs [16, 27, 35, 36]. Even if \\(2 - 3\\) SLM layers may not be enough for correcting the full field-of-view of a wide image through a thick tissue, they can allow us to increase the support of the fitted area by much more than \\(\\times 2 - 3\\) thus they can significantly accelerate a sequential wavefront-shaping correction of a wide image." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.821, + 0.906, + 0.851 + ], + "angle": 0, + "content": "Acknowledgments: This research was funded by ERC SpeckleCorr-101043471, ISF 563/24." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.903 + ], + "angle": 0, + "content": "Code and data: can be obtained at https:// webee.technion.ac.il/people/anat.levin/ papers/LayerdTmat(cvpr25_code.zip" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23171" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.483, + 0.156 + ], + "angle": 0, + "content": "[1] Dror Aizik and Anat Levin. Non-invasive and noise-robust light focusing using confocal wavefront shaping. Nature communications, 15(5575), 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.157, + 0.483, + 0.199 + ], + "angle": 0, + "content": "[2] Dror Aizik, Ioannis Gkioulekas, and Anat Levin. Fluorescent wavefront shaping using incoherent iterative phase conjugation. Optica, 9(7):746-754, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.2, + 0.482, + 0.24 + ], + "angle": 0, + "content": "[3] M. Alterman, C. Bar, I. Gkioulekas, and A. Levin. Imaging with local speckle intensity correlations: theory and practice. ACM TOG, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.482, + 0.283 + ], + "angle": 0, + "content": "[4] YoonSeok Baek, Hilton B. de Aguiar, and Sylvain Gigan. Phase conjugation with spatially incoherent light in complex media. Nature Photonics, 17(12):1114-1119, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.285, + 0.482, + 0.352 + ], + "angle": 0, + "content": "[5] Paul Balondrade, Victor Barolle, Nicolas Guigui, Emeric Auriant, Nathan Rougier, Claude Boccara, Mathias Fink, and Alexandre Aubry. Multi-spectral reflection matrix for ultrafast 3d label-free microscopy. Nature Photonics, 18(10): 1097-1104, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.482, + 0.395 + ], + "angle": 0, + "content": "[6] Chen Bar, Marina Alterman, Ioannis Gkioulekas, and Anat Levin. A monte carlo framework for rendering speckle statistics in scattering media. ACM TOG, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.397, + 0.482, + 0.436 + ], + "angle": 0, + "content": "[7] Chen Bar, Ioannis Gkioulekas, and Anat Levin. Rendering near-field speckle statistics in scattering media. ACM TOG, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.482, + 0.479 + ], + "angle": 0, + "content": "[8] Martin Booth. Adaptive optical microscopy: The ongoing quest for a perfect image. Light: Science and Applications, 3:e165, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.482, + 0.522 + ], + "angle": 0, + "content": "[9] Bingquan Chen and Jakob J. Stamnes. Validity of diffraction tomography based on the first born and the first rytrov approximations. Appl. Opt., 37(14):2996-3006, 1998." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.524, + 0.483, + 0.578 + ], + "angle": 0, + "content": "[10] Wonshik Choi, Christopher Fang-Yen, Kamran Badizadegan, Seungeun Oh, Niyom Lue, Ramachandra R Dasari, and Michael S Feld. Tomographic phase microscopy. Nature methods, 4(9):717-719, 2007." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.58, + 0.482, + 0.647 + ], + "angle": 0, + "content": "[11] Shwetadwip Chowdhury, Michael Chen, Regina Eckert, David Ren, Fan Wu, Nicole Repina, and Laura Waller. High-resolution 3d refractive index microscopy of multiple-scattering samples from intensity images. Optica, 6(9): 1211-1219, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.649, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[12] Anat Daniel, Dan Oron, and Yaron Silberberg. Light focusing through scattering media via linear fluorescence variance maximization, and its application for fluorescence imaging. Optics Express, 27(15):21778, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.482, + 0.733 + ], + "angle": 0, + "content": "[13] A. J. Devaney. Inverse-scattering theory within the rytov approximation. Opt. Lett., 6(8):374-376, 1981." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[14] Amicia D. Elliott. Confocal microscopy: Principles and modern practices. Current Protocols in Cytometry, 92(1): e68, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[15] Brandon Y. Feng, Haiyun Guo, Mingyang Xie, Vivek Boominathan, Manoj K. Sharma, Ashok Veeraraghavan, and Christopher A. Metzler. Neuws: Neural wavefront shaping for guidestar-free imaging through static and dynamic scattering media. Science Advances, 9(26):eadg4671, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[16] T. Furieri, A. Bassi, and S. Bonora. Large field of view aberrations correction with deformable lenses and multi conjugate adaptive optics. Journal of Biophotonics, 16(12): e202300104, 2023." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[17] Omri Haim, Jeremy Boger-Lombard, and Ori Katz. Image-guided computational holographic wavefront shaping. Nature Photonics, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[18] Karen Hampson, Raphael Turcotte, Donald Miller, Kazuhiro Kurokawa, Jared Males, Na Ji, and Martin Booth. Adaptive optics for high-resolution imaging. Nature Reviews Methods Primers, 1:68, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[19] John W. Hardy. Adaptive Optics for Astronomical Telescopes. 1998." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[20] Yi Hua, M. Salman Asif, and Aswin C. Sankaranarayanan. Spatial and axial resolution limits for mask-based lensless cameras. Opt. Express, 31(2):2538-2551, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.329 + ], + "angle": 0, + "content": "[21] Seungwon Jeong, Ye-Ryoung Lee, Wonjun Choi, Sungsam Kang, Jin Hee Hong, Jin-Sung Park, Yong-Sik Lim, HongGyu Park, and Wonshik Choi. Focusing of light energy inside a scattering medium by controlling the time-gated multiple light scattering. Nature Photonics, 12(5):277-283, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[22] Na Ji. Adaptive optical fluorescence microscopy. Nature Methods, 14:374-380, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.359, + 0.905, + 0.439 + ], + "angle": 0, + "content": "[23] Yonghyeon Jo, Ye-Ryoung Lee, Jin Hee Hong, Dong-Young Kim, Junhwan Kwon, Myunghwan Choi, Moonseok Kim, and Wonshik Choi. Through-skull brain imaging in vivo at visible wavelengths via dimensionality reduction adaptive-optical microscopy. Science Advances, 8(30):eabo4366, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[24] Andrew Jr, Alon Harris, Josh Gross, Ingrida Januleviciene, Aaditya Shah, and Brent Siesky. Optical coherence tomography angiography: An overview of the technology and an assessment of applications for clinical research. British Journal of Ophthalmology, 101, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.512, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[25] Sungsam Kang, Pilsung Kang, Seungwon Jeong, Yongwoo Kwon, Taeseok D. Yang, Jin Hee Hong, Moonseok Kim, Kyung-Deok Song, Jin Hyoung Park, Jun Ho Lee, Myoung Joon Kim, Ki Hean Kim, and Wonshik Choi. High-resolution adaptive optical imaging within thick scattering media using closed-loop accumulation of single scattering. Nature Communications, 8(1):2157, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[26] Sungsam Kang, Yongwoo Kwon, Hojun Lee, Seho Kim, Jin Hee Hong, Seokchan Yoon, and Wonshik Choi. Tracing multiple scattering trajectories for deep optical imaging in scattering media. Nature communications, 14(1), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[27] Marie Laslandes, Matthias Salas, Christoph K. Hitzenberger, and Michael Pircher. Increasing the field of view of adaptive optics scanning laser ophthalmoscopy. Biomed. Opt. Express, 8(11):4811-4826, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[28] V. Lauer. New approach to optical diffraction tomography yielding a vector equation of diffraction tomography and a novel tomographic microscope. Journal of Microscopy, 205 (2):165-176, 2002." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[29] F. C. Lin and M. A. Fiddy. The born-rytov controversy: I. comparing analytical and approximate expressions for the one-dimensional deterministic case. J. Opt. Soc. Am. A, 9 (7):1102–1110, 1992." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[30] Jerome Mertz. Introduction to optical microscopy. Cambridge University Press, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[31] Sagi Monin, Aswin Sankaranarayanan, and Anat Levin. Exponentially-wide etendue displays using a tilting cascade. IEEE ICCP, pages 1-12, 2022." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23172" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[32] Ulysse Najar, Victor Barolle, Paul Balondrade, Mathias Fink, Claude Boccara, and Alexandre Aubry. Harnessing forward multiple scattering for optical imaging deep inside an opaque medium. Nature Communications, 15(1), 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[33] Tan Nguyen, Mikhail Kandel, Marcello Rubessa, Matthew Wheeler, and Gabriel Popescu. Gradient light interference microscopy for 3d imaging of unlabeled specimens. Nature Communications, 8, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.206, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[34] Galya Stern and Ori Katz. Noninvasive focusing through scattering layers using speckle correlations. Optics Letters, 44(1):143, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[35] Jörgen Thaung, Per Knutsson, Zoran Popovic, and Mette Owner-Petersen. Dual-conjugate adaptive optics for wide-field high-resolution retinal imaging. Opt. Express, 17(6): 4454-4467, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[36] Tsai wei Wu and Meng Cui. Numerical study of multiconjugate large area wavefront correction for deep tissue microscopy. Opt. Express, 23(6):7463-7470, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[37] Gil Weinberg, Elad Sunray, and Ori Katz. Noninvasive megapixel fluorescence microscopy through scattering layers by a virtual incoherent reflection matrix. Science Advances, 10(47):eadl5218, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[38] Emil Wolf. Three-dimensional structure determination of semi-transparent objects from holographic data. Optics Communications, 1(4):153-156, 1969." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[39] Tomer Yeminy and Ori Katz. Guidestar-free image-guided wavefront shaping. Science Advances, 7(21):eabf5364, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[40] Lei Zhu, Fernando Soldevila, Claudio Moretti, Alexandra d'Arco, Antoine Boniface, Xiaopeng Shao, Hilton B. de Aguiar, and Sylvain Gigan. Large field-of-view non-invasive imaging through scattering layers using fluctuating random illumination. Nature Communications, 13:1447, 2022." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23173" + } + ] +] \ No newline at end of file diff --git a/2025/Understanding Multi-layered Transmission Matrices/0c5698ec-c1a0-4efc-b1d6-15a1169745e5_origin.pdf b/2025/Understanding Multi-layered Transmission Matrices/0c5698ec-c1a0-4efc-b1d6-15a1169745e5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..11e56502bf0d2e38ef4de7dffd91dd71de6ca162 --- /dev/null +++ b/2025/Understanding Multi-layered Transmission Matrices/0c5698ec-c1a0-4efc-b1d6-15a1169745e5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:29891e6026c530a567acf685a43cb8389254a1c4277501b37f81786d78a0ef84 +size 2006481 diff --git a/2025/Understanding Multi-layered Transmission Matrices/full.md b/2025/Understanding Multi-layered Transmission Matrices/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5c8f2aca48f00a88118bf1a06100d2ff469608ae --- /dev/null +++ b/2025/Understanding Multi-layered Transmission Matrices/full.md @@ -0,0 +1,337 @@ +# Understanding multi-layered transmission matrices + +Anat Levin and Marina Alterman +Department of Electrical and Computer Engineering, Technion Haifa, Israel + +# Abstract + +Wavefront shaping systems attempt to correct aberrations caused by tissue scattering, by placing a spatial light modulator (SLM) in the optical path and using it to reshape the light wave emitted from a target of interest deep inside the tissue. However, the field-of-view we can correct with one SLM pattern is extremely small, because inherently the aberration is caused by the 3D tissue structure while the SLM is planar. To overcome this, multi-conjugate correction systems built with multiple layered SLMs have been introduced, which attempt to approximate the 3D tissue structure with multiple planar aberrations. + +However, multi-conjugate systems with a large number of SLMs are not easy to construct. Therefore, it is important to understand how many SLM layers are actually needed for a good correction, and whether there are practical benefits from correction systems with a relatively small number of layers. To help in the design of future systems, this paper analyzes multi-layer corrections. We show that the well-known missing cone problem which fundamentally limits the axial resolution of microscopes, turns into an advantage when designing 3D correction systems. Due to this property less information should be captured by the correction system, facilitating sparser approximations. We also show that even if the number of available layers is insufficient for a reasonable approximation of the full 3D tissue volume, it can significantly expand the field-of-view when compared to single-layer correction systems, and therefore can largely accelerate their operation. + +# 1. Introduction + +Scattering is a central challenge in multiple imaging tasks ranging from astronomy to microscopy, and biomedical imaging. It degrades our ability to resolve stars through telescopes, image into the bottom of the sea, and drive through foggy weather. Most importantly in the context of this paper, scattering severely limits our ability to image through biological tissue. This happens due to small variations in the refractive index of the tissue volume, which scatters the propagating light. + +Numerous computational optics approaches have been + +developed for imaging through scattering tissue. These include computational restoration algorithms which attempt to fit the observed data with various models for the aberration and the target of interest [4, 5, 15, 17, 21, 23, 25, 32, 37, 39, 40]. Also, optical systems such as a confocal microscope [14] or an OCT [24], aim to filter out the scattered light and maintain only the ballistic unscattered part. However, as we attempt to see deeper into a scattering layer, ballistic light weakens significantly, and the complexity of aberration models increases. Moreover, biological target of interest such as fluorescent tissue components are very weak and as a result the scattered signal is also of a very low signal to noise ratio (SNR), further limiting digital correction models. A particularly attractive approach for scattering aberration removal is wavefront shaping, where one places a spatial light modulator (SLM) in the optical path of the microscope and uses it to correct the wavefront emitted from a point deep inside the tissue, so that light photons emerging from a single target point are brought into a single sensor point, despite the tissue aberration. Wavefront shaping is a particularly attractive solution for fluorescent imaging because the correction is done optically rather than digitally. Unlike ballistic filtering approaches, all light photons are used. Since the correction is done in the optical path, the limited number of available photons is brought into a single detector pixel and can be measured with higher SNR. As a result, wavefront shaping has the potential to revolutionize tissue imaging and allow us to overcome the fundamental limitations of scattering, and image very deep inside biological targets at a high SNR. + +The main limitation for wavefront shaping approaches is that the aberrations of neighboring points inside volumetric tissue vary rapidly, and in practice a single correction mask can only allow imaging of a small local field-of-view. This is because the scattering process happens throughout the 3D volume, while the aberration correction only happens on a planar SLM. Earlier demonstrations of wavefront shaping have considered targets placed at a distance behind a thin scattering layer, and this far-field arrangement has allowed them to correct larger regions with a single planar modulation [12, 15, 34, 39]. Similarly in the context of adaptive-optics, researchers have attempted to correct aber + +rations due to imperfect optics or a refractive index mismatch [8, 18, 22]. These aberrations are introduced along the optical path, before the tissue, so the correction can apply to a larger area. However, in most biomedical applications the fluorescent target of interest is embedded within the scattering volume. If our goal is to correct scattering within the tissue itself, the field-of-view which can be corrected by a single planar modulation is extremely small, usually only a few microns [3]. + +To correct larger fields of view, multi conjugate correction systems have been proposed. Such systems use multiple planar SLMs placed such that each of them is conjugate to a different plane inside the tissue volume. Each such SLM should realize a correction for the aberration happening inside a thin layer of the volume and the combination of such multi-plane corrections should approximate the 3D variation in refractive index, and allow us to correct it. Fig. 1 visualizes the concept of a multi-conjugate correction [16, 27, 35, 36]. In practice, multi-plane correction systems are more complicated than the simplified visualization of Fig. 1, as one needs to add multiple relay systems between the SLMs. This is first because there are physical limitations on the minimal distance between 2 SLMs, and second because we need to revert the order of the correction layers in the inversion path [19]. This makes multiconductor correction systems complex and cumbersome to build. Previous experimental demonstrations have realized corrections with no more than two SLMs. While two-layer correction systems have demonstrated improved performance compared to single layer correction, two planes clearly cannot model a thick 3D tissue structure. + +Our goal in this research is to analytically study multislice models and attempt to answer the following questions: How many planes are required to approximate the 3D aberration of a scattering tissue volume? How does the quality of the correction degrade if we use a sparser approximation? If a construction of a practical correction system is limited to only to M layers, what is the maximum field of view that can be effectively corrected? + +To this end, we use the mathematical model of a transmission matrix, which stores the propagation of coherent wavefronts from one end of a scattering volume to the other. We study how well we can approximate such transmission matrices with a multi-slice model, and how does the error depend on the number of fitted layers. + +The straightforward answer to this question is to say that the spacing between layers should match the axial resolution of the microscope. For example, using an $\mathbf{NA} = 0.5$ objective at wavelength $500nm$ , we get that the spacing between layers should be lower than $2\mu m$ . Handling a modest tissue thickness of, e.g., $200\mu m$ already requires 100 slices for a valid approximation. Clearly building a multiconjugate wavefront shaping systems with so many layers + +![](images/9bd6169b3d3f269f4ac098e352a507d116a8e3e75e350f687cba249d2cd12475.jpg) +Figure 1. Layered wavefront shaping: (Top) Consider a simple microscope setup where we use a 4F relay to image the light emerging from a tissue sample. Light is aberrated on its way to the sensor, and we see speckle patterns rather than sharp spots. (Middle) A single planar SLM can correct aberration for one spot (see red beam), but light emerging from nearby spots may not be corrected (the yellow beam generates a speckle pattern on the sensor). (Bottom) The field of view of the correction can possibly be extended with a multi-layer correction. The goal is to form a correction replicating the 3D structure of volumetric aberration inside the tissue. The figure illustrates two wavefronts from nearby spots (red and yellow) passing through the same volumetric correction, and both focus into a sharp spot behind the tissue. + +is impractical. + +In this research we take inspiration from [20], and offer a thorough analysis of the frequency content of transmission matrices, combining the forward scattering nature of biological tissue with the structure of the missing cone. We show that while the missing cone structure is known as one of the fundamental limitations of 3D microscopy, in the context of transmission matrices this limitation turns into an advantage. Due to the missing cone structure, many of the frequencies of the volumetric aberration do not participate in the transmission matrix, and therefore the number of layers required for approximating a transmission matrix is way smaller than the Nyquist limit. However, there are no free lunches and the number of layers scales with the tissue thickness. To explain a few hundred micron thick tissue one needs a few dozen layers, which is still too much to be realized by a physical correction system. + +Given this conclusion we aim to understand if there are any practical benefits in realizing a multi-conjugate correction system with a small number of say 2 - 4 layers. We show that while a small number of layers cannot well approximate the 3D structure of the tissue, it can allow us to realize a correction over a limited field-of-view. The area which can be corrected with $M$ aberration layers is significantly larger than the area which can be corrected with one layer, and therefore it can significantly accelerate sequen + +![](images/f0721efb15f9ea6c7b32268b5b6ba3c736c1699d40f6b520e16a78a964fd6112.jpg) +(a) Full setup + +![](images/a5c58ed619bc8c45c2f1d477f8c5034216e4fd639836400a17b0ee3e9b90dc3b.jpg) +Figure 2. Setup for transmission matrix acquisition: (a) A setup for transmission matrix imaging where a set of basis illuminations at the back of the sample propagate through the volume, and the scattered wavefronts are measured by a camera at the other end. In (b,c) we zoom on the sample illustrating two possible illumination bases. (b) A set of directional wavefronts (Fourier basis). (c) A set of wavefronts focused to points at the edge of the volume (primal basis). Both transmission matrices can be related by a simple Fourier transform. (d) A multi-slice model approximates the volumetric scattering with a set of planar aberrations. + +![](images/c49845d22af3999f33d70e1c082ade41175702092ae11814836b23961ce4eec6.jpg) +(b) Fourier illum. +(c) Point illum. + +![](images/737cbeb37f676e96a48d7e780ea731c4776d5f743e73b0ed249eedd22cbb31b1.jpg) +(d) Layered approx. + +tial correction of the field of view. We validate our findings using real transmission matrices captured in the lab as well as synthetic transmission matrices simulated using accurate wave propagation models. + +# 2. Problem formulation + +Imaging setup: For simplicity, the following derivation considers coherent scalar waves, ignoring polarization. In Fig. 2(a) we visualize a setup for transmission matrix acquisition. A scattering medium, such as tissue, is illuminated by a coherent source $\mathbf{e}^k$ at one side of the medium. After propagating through the medium, the resulting coherent wavefront $\mathbf{t}^k$ is recorded. Usually $\mathbf{t}^k$ are complex valued speckle patterns with a pseudo-random structure. The illumination is translated and multiple wavefronts are recorded and stored as columns of a transmission matrix $\mathcal{T}$ . + +The transmission matrix: Since the propagation of coherent light is linear, if we want to predict the propagation of any wavefront $\mathbf{u}$ , we simply express it as a superposition of the recorded incoming illuminations $\mathbf{u} = \sum_{k} \mathbf{u}(k) \mathbf{e}^{k}$ , where $\mathbf{u}(k)$ are complex scalars, and multiply it by the transmission matrix, to predict $\mathcal{T} \mathbf{u}$ . The illumination wavefronts $\mathbf{e}^{k}$ can be selected at different bases. For example, the illumination can be translated at the Fourier plane, and then the wavefronts illuminating the sample are plane waves at different angles (Fig. 2(b)). Alternatively one can focus the light to a point at the back plane of the sample (Fig. 2(c)). This configuration is more appropriate for wavefront shaping algorithms whose goal is to recover wavefronts focusing inside the tissue and not far behind it. In both cases, however, the illuminations we can create are limited by the numerical aperture of the illuminating objective, which we mark here as NA. Similarly, the scattered wavefronts are measured via an imaging objective with the same bounded NA. This NA would serve as a central component in the results derived below. For simplicity of resulting equations, for the rest of this manuscript we assume the imaging and illumination objectives are focused at the same plane, such that if we would remove the scattering sample and image the sources directly, we would see the $\mathbf{e}^{k}$ illuminations as sharp spots on the sensor plane. That is, if we use directional + +illuminations, we also measure the Fourier transform of the scattered waves, namely place the sensor at the Fourier plane of the imaging objective. If we use point sources at the illumination end of the medium, we also focus the imaging objective at the same plane, so we measure a speckle image of a relatively small support. + +Range and resolution: We assume that in the primal basis the illuminators are spaced within an area of $\Omega_p \times \Omega_p$ , and we sample them at the Nyquist resolution, so for an illumination wavelength $\lambda$ they are separated by a pitch $\Delta_p = \lambda/(2\mathbf{NA})$ . Equivalently, in the Fourier domain, we illuminate the sample with plane waves of the form + +$$ +\mathbf {e} (\mathbf {r}) = e ^ {2 \pi i (\bar {\kappa} \cdot \mathbf {r})} \tag {1} +$$ + +where $\mathbf{r}$ is a point inside the 3D volume, and $\bar{\kappa} = \frac{1}{\lambda}\bar{\omega}$ where $\lambda$ is the wavelength and $\bar{\omega}$ is a unit norm direction vector, i.e. the last entry of the direction vector $\omega_{z}$ is $\omega_{z} = \sqrt{1 - (\omega_{x}^{2} + \omega_{y}^{2})}$ . For simplicity of notation, we assume the refractive index (RI) of the leading medium is 1. We can only use illumination direction inside the numerical aperture $\sqrt{\omega_x^2 + \omega_y^2}\leq \mathbf{N}\mathbf{A}$ (or equivalently $\sqrt{\kappa_x^2 + \kappa_y^2}\leq \Omega_f$ with $\Omega_{f} = \mathbf{N A} / \lambda$ ). To sample the transmission matrix at the Nyquist pitch, the $\kappa_{x},\kappa_{y}$ components of the illumination are sampled by the Nyquist pitch $\Delta_{f} = 1 / 2\Omega_{p}$ . + +With the above sampling scheme, the transmission matrices at the Fourier and primal bases are related by a simple Fourier transform $\mathcal{T}^p = \mathcal{F}^T\mathcal{T}^f\mathcal{F}$ . While for wavefront shaping we are interested in primal transmission matrices, the Fourier ones simplify analysis. Importantly, since Fourier and primal transmission matrices are related by an orthonormal transformation, the fitting errors in both bases are equivalent. + +The multi-slice model: The relation between an incoming illumination and an outgoing one is determined by the solution to a wave equation which takes into account the variation in refractive index (RI) inside the sample. Rather than a fully continuous model, in a multi-slice model we approximate the volume with a set of planar aberrations, separated by small $z$ distances $\Delta_z$ , where we have free space propagation between layers, as in Fig. 2(d). Mathematically this can be described as + +$$ +\boldsymbol {\mathcal {T}} \left(\rho_ {1}, \dots , \rho_ {M}\right) = \boldsymbol {\mathcal {P}} _ {O} \boldsymbol {\mathcal {D}} \left(\rho_ {M}\right) \dots \boldsymbol {\mathcal {P}} _ {\Delta z} \boldsymbol {\mathcal {D}} \left(\rho_ {2}\right) \boldsymbol {\mathcal {P}} _ {\Delta z} \boldsymbol {\mathcal {D}} \left(\rho_ {1}\right), \tag {2} +$$ + +![](images/ffe8fb84265c39b1272d097ee9dc1821b61d3afdaf274043f211763d73a1f15e.jpg) +(a) Ground truth + +![](images/95e6e867cb700e3562b6d2eb098aff8864b563efaea60941564317acad376783.jpg) +(b) 32 layers + +![](images/2f43cc57c9d498714e26fb51adff727a8d6794186eab3ec66cc99c556b26905b.jpg) +(c) 8 layers +Figure 3. Reconstructed volumes: (a) An xz slice from a toy volume, visualizing its refractive index variations. (b-d) Reconstructions with varying layer numbers. The columns of the transmission matrix provide an under-constraint inversion problem, and even when a dense set of layers is allocated in b, the axial resolution of the reconstructed volume is low. As a result the sparse reconstructions in c,d can also well explain the transmission matrix. (e) Numerical evaluation of fitting error as a function of the number of layers. Transmission matrices can be well fitted even if the allocated number of layers is lower than Nyquist predictions. However the required number of layers increases with the tissue thickness. The layer number is less affected when the optical density is increased for the same tissue thickness. + +![](images/d1c7c44fb676b59ccce5726adf4f7d480990a8c40a71cf6fa4d29bab9b1ef627.jpg) +(d) 4 layers + +![](images/eb9bb615bbe774bfbe881ece5da6a5612783a420231a2c520d5bcb03a87bb5a0.jpg) +(e) Fitting curves + +where $\rho_{m}(x,y)$ is the 2D aberration of the $m$ th layer and $\mathcal{D}(\rho_m)$ is a diagonal matrix performing element-wise multiplication of the incoming wavefront by the aberration. $\mathcal{P}_{\Delta_z}$ describes free space propagation between layers at distance $\Delta_z$ , which is essentially a convolution with a spherical wavefront. Finally $\mathcal{P}_O$ propagates the wavefront in free space from the tissue boundary, via the lenses, to the camera sensor. + +Our goal is to understand what plane spacing $\Delta_z$ will facilitate a good representation of $\mathcal{T}$ , and how this approximation would degrade when we use sparser sampling. + +A common approximation for weak aberrations is: + +$$ +\boldsymbol {\rho} _ {m} (x, y) = \mu + \delta \boldsymbol {\rho} _ {m} (x, y), \tag {3} +$$ + +where $\mu$ is a scalar whose magnitude is a bit smaller than 1, encoding the ballistic portion of the light which passes through the layer without aberration, and the magnitude of the aberration is significantly smaller than the ballistic term $|\delta \rho_{m}(x,y)|\ll |\mu |$ . For weak scattering most of the energy does not scatter more than once. It is therefore common to consider a first order approximation to the transmission matrix (often known as the single scattering, first-Born or Rytrov approximation [9, 10, 13, 29, 33]), which reduces to + +$$ +\boldsymbol {\mathcal {T}} \left(\delta \rho_ {1}, \dots , \delta \rho_ {M}\right) \approx \mu^ {M} \boldsymbol {\mathcal {P}} _ {O} \boldsymbol {\mathcal {P}} _ {M \Delta_ {z}} + \tag {4} +$$ + +$$ +\mu^ {M - 1} \sum_ {m = 1} ^ {M} \mathcal {P} _ {O} \mathcal {P} _ {(M - m) \Delta_ {z}} \mathcal {D} (\delta \rho_ {m}) \mathcal {P} _ {m \Delta_ {z}} +$$ + +The full light propagation in Eq. (2) includes multiplication of the different aberrations and hence the transmission matrix is highly nonlinear. In contrast, the advantage of the weakly scattering model in Eq. (4) is that the resulting transmission matrix is a linear function of the aberration layers. This linear approximation significantly simplifies analysis. + +Diffraction tomography algorithms such as [11] capture a subset of columns from the transmission matrix and use + +gradient descent optimization to seek layers $\rho_1, \ldots, \rho_M$ which best explain the measured wavefronts, assuming that the propagation operators $\mathcal{P}_{\Delta_z}, \mathcal{P}_O$ are known. We will use a similar optimization scheme below, when trying to find a multi-slice representation for a given transmission matrix. To correct aberrations, a physical multi-conjugate wavefront shaping system would display on the SLM the conjugate of the recovered layers. + +# 3. Evaluating multi-slice approximations + +We start with numerical experiments demonstrating the main findings. In Sec. 4 below we derive a theory explaining them. To gain intuition we generated a toy volume composed of spheres with varying refractive indices illustrated in Fig. 3(a), and obtained the corresponding transmission matrix by propagating a set of point sources through the volume. We then used the transmission matrix to solve for a multi-slice approximation, while varying the number of layers. In Fig. 3(b-d) we visualize some of the volumetric reconstructions, assuming a $\lambda = 500nm$ illumination. Even when the fitted layers are sampled very densely at an interval of $\Delta_z = 0.5\mu m$ , we cannot fully recover the target and the fitted volume has a low axial resolution. We analyze this in Sec. 4 below and explain that this results from the well known missing cone problem, a fundamental degeneracy limiting the axial resolution of a microscope. Said differently the transmission matrix, or the full set of propagating waves, only provides a set of under-constraint measurements from which the volume cannot be uniquely recovered. While the missing cone problem is considered a fundamental limitation in microscopy, we argue that in the context of wavefront shaping it actually turns into an advantage. Since the volume reconstruction is not unique we can + +![](images/c030826722d17eab87af7556ccee4f64461f35318021127ea39c15c82e996dde.jpg) +(a) Ground truth + +![](images/44e481b3b0d8ba9707b4df26d4dec5884e8dcb09e2c991e70c435b772a81296e.jpg) +(b) Dense fit + +![](images/bcfef151987ea8ba01c9b03cba4f97f51e36b7536bf389e24341ff5ee1f529dc.jpg) +(c) Sparse fit + +![](images/89bb3c593d97c43e54b7e49fec9b357151b35b0b8b8387c56fc621c6217359b3.jpg) +(d) Dense fit +Figure 4. Reconstructed volumes with varying supports: (a) An xz slice from a toy volume, visualizing its refractive index variations. (b,d) Two reconstructions of the volume using a dense layer sampling. When the support of the measured window $\Omega_{p}$ shrinks, the lateral extent of the reconstructed region decreases, however, its axial resolution is also reduced. (c,e) A sparse fit with the same input. Despite the low quality of these reconstructions, they predict well the columns of the transmission matrix of the given support $\Omega_{p}$ . (f) Numerical evaluation of fitting error as a function of the number of layers, while varying the range $\Omega_{p}$ covered by the transmission matrix. Smaller supports can be fitted with fewer layers. + +![](images/a9eab028a53d42a94619d402f561f77c79d549e911fd9c9e2140b5007b22c571.jpg) +(e) Sparse fit + +![](images/d667e5f775f637a56c3bd675225f8d9c91df8e3aaf10a06697fafb6262f87970.jpg) +(f) Fitting curves + +approximate the same transmission matrix with a sparser set of layers. + +To quantify the fitting error of the multi-slice model, we denote by $\pmb{\varepsilon}_{M}$ the reconstruction error of a given transmission matrix with the best $M$ layers + +$$ +\boldsymbol {\mathcal {E}} _ {M} = \min _ {\rho_ {1}, \dots , \rho_ {M}} \| \boldsymbol {\mathcal {T}} _ {\text {e x a c t}} - \boldsymbol {\mathcal {T}} (\rho_ {1}, \dots , \rho_ {M}) \| ^ {2}. \tag {5} +$$ + +In Fig. 3(e) we evaluate $\pmb{\mathcal{E}}_M$ as a function of the layer number $M$ . Even the seemingly aliased fit of Fig. 3(d) still provides a low fitting error. However, there are no free lunches and when the number of layers is too low the fitting error is high. When we repeat the simulation with a thicker volume we need even more layers for a good fit. Interestingly, we also repeat the simulation with denser volumes (larger optical depth, or shorter mean free path) of the same thickness, and this has less effect on the required number of layers. In our plots, the case $M = 0$ refers to no correction at all, so effectively we assume that the transmission matrix is diagonal. The similarity between a diagonal transmission matrix and the target transmission matrix is a measure of the amount of ballistic light. + +The thickness of the synthetic simulation here is limited due to computational constraints. However, usually wavefront shaping systems attempt to see into a few hundred microns of tissue and we anticipate that for such volumes we need at least a few dozen layers for a good fit. Clearly realizing a physical wavefront shaping system with that many layers is impractical. We argue that while a small number of layers is not enough for focusing light anywhere across the tissue area, it can still help us to focus light over a compact support area, and this area is significantly wider than what we can correct with a single layer. + +Compact support: To test the effect of the target area, we consider transmission matrices measured with illumination points confined to a limited area of size $\Omega_p \times \Omega_p$ . If we can + +fit such a transmission matrix with $M$ layers and place the values of these $M$ layers on $M$ SLMs, the resulting wavefront shaping system can correct light from the $\Omega_p \times \Omega_p$ area, but target points outside this area are unlikely to be corrected as they are not included in the fitting process. + +To test the effect of support on the number of fitted layers, we return to the numerical simulation of Fig. 3, but fit subsets of columns contained in smaller $\Omega_p$ ranges. Indeed, smaller supports can be well-fitted with a smaller number of layers. As another way to understand it, we show in Fig. 4 an xz slice from the ground truth refractive index volume and from a few reconstructions. We show that even if we do not limit the layer number $M$ , the recovered RI volume has limited axial resolution and this axial resolution is lower when measured range $\Omega_p$ decreases. This explains why lower supports can be fitted with fewer layers. For each of the two supports we also show a sparse fit, with the minimal $M$ value that provides a good prediction of the transmission matrix. We explain this observation in Sec. 4. This result is also in agreement with [26], who correct aberration over a relatively small field of view through a few hundred microns of tissue, using only 5 layers. + +Implications to layered wavefront shaping systems. In Fig. 4(f) we plotted a bar below which the approximation is reasonable. We can see that for e.g. the $40 \times 40\mu m$ support we can obtain good focusing with about $M = 6$ layers. It is likely that fitting a full field-of-view of a few hundred microns would require more layers. However, we can see that while with $M = 1$ layer we could only focus over a small area of about $1 \times 1\mu m$ , when increasing to $M = 2$ we could focus over a $\times 4$ larger area of support $2 \times 2\mu m$ , and with $M = 3$ we could focus over a $\times 25$ larger area of support $5 \times 5\mu m$ . Computing multiple correction elements is only going to increase the complexity of the wavefront shaping + +![](images/a619d8c80fbac286a52416d0615f50d448ffa7ed9fbf2a4570d53d4dff932fdb.jpg) + +![](images/3ef80ea7a846532437c145aa0d761356e765cebe99d085ea8dc6bb3863484d79.jpg) +Figure 5. Layer support for a captured transmission matrix: We consider a transmission matrix captured in the lab through a chicken breast layer of thickness $170\mu m$ . Left: plotting the fitting error as a function of the number of layers. Smaller supports can be fitted with fewer layers. Right: Visualization of a spot focusing behind the tissue computed using the wavefronts of the fitted model. We compare models fitted to different supports. In the top row we need a larger number of layers, but the spot can be scanned over a $25 \times 25\mu m$ window using the same layers. In the lower rows we achieve a focused spot using a smaller number of layers, but these layers can scan the focused spot over smaller windows of sizes $6 \times 6\mu m$ and $1 \times 1\mu m$ . + +algorithm in [1] linearly with the number of layers. Alternatively, a wavefront shaping system using a single SLM would need to scan the field-of-view sequentially and correct each local window independently. Since the increase in the support of the window we can explain with e.g. 3 layers is significantly larger than a $\times 3$ factor, a multi-conjugate wavefront shaping system with a small number of SLMs as in [16, 27, 31, 35] can already have a large practical benefit. While Fig. 4 only considers a simulation, we analyze realistic numbers below. + +Lab results: In the supplementary file we validate the theory with multiple real transmission matrices, representing thick volumes with multiple-scattering. + +We start with numerically simulated transmission matrices using a more accurate wave-propagation model [6, 7]. We then test lab-captured transmission matrices measuring realistic scattering samples including mouse brain tissue, chicken breast tissue and a parafilm layer. Please refer to supplementary file for additional results. + +In Fig. 5 we show one fitting result using a multiply-scattering chicken breast layer of thickness $d = 170\mu m$ . The transmission matrices cover an area with support $\Omega_{p} \times \Omega_{p} = 13 \times 13\mu m$ . As in Fig. 4(f), we see that smaller spatial ranges can be fitted with fewer layers. + +As another way to assess the quality of the fit, we used columns of the fitted matrices to focus light through the tissue and show the shape of the focused spot and visualize the focusing quality in Fig. 5. To simulate the focused spot we simply visualize the product $\mathcal{T}_{\mathrm{exact}} \cdot \mathbf{t}_{\mathrm{fit,M}}^k$ , where $\mathbf{t}_{\mathrm{fit,M}}^k$ is a column of the fitted matrix $\mathcal{T}(\rho_1, \ldots, \rho_M)$ , namely a wavefront generated by the layered approximation. Note that while we only show focusing at one point, the layers are optimized such that they allow us to focus at any point inside the $\Omega_p \times \Omega_p$ window. + +From the results in Fig. 5 we can see that for the $13\times$ + +$13\mu m$ support measured, we can obtain good focusing with about 3 layers. Fitting a full field-of-view of several hundred microns likely requires additional layers. However, we observe that while using $M = 1$ allows us to focus on a small area of about $3\times 3\mu m$ , with $M = 3$ , we can focus on an area $\times 18$ larger, reaching $13\times 13\mu m$ . Therefore, using multiple layers can significantly accelerate a sequential scanning of a wave-front shaping system. + +# 4. Sampling theory for transmission matrices + +Our goal in this section is to explain the empirical findings of the previous section by analyzing the information captured by a transmission matrix. + +# 4.1. Nyquist sampling and the missing cone + +We start by considering the case of weakly scattering volumes which, as in Eq. (4), can be reasonably approximated by considering the component of light that scattered once. We also assume the transmission matrix is sampled over a very wide support $\Omega_{p}$ . We discuss the implications of multiple scattering and limited supports in subsequent sections. + +We denote by $\pmb{n}(\mathbf{r})$ the difference between the RI at a 3D position $\mathbf{r}$ in the volume to that of the leading medium around it. We denote its 3D Fourier transform by $\hat{n}(\kappa)$ . + +We start by considering the Fourier representation of the transmission matrix, where it is illuminated and measured by plane waves. Using the weakly scattering (first-Born) approximation, the wavefront scattering toward direction $\bar{\omega}^v$ when illuminated by an incoming plane wave at direction $\bar{\omega}^i$ is proportional, up to a multiplicative factor, to [28, 30, 38]: + +$$ +\propto \int \boldsymbol {n} (\mathbf {r}) e ^ {2 \pi i \left(\bar {\kappa} ^ {i} - \bar {\kappa} ^ {v}\right) \cdot \mathbf {r}} d \mathbf {r}, \tag {6} +$$ + +with $\bar{\kappa}^v = 1 / \lambda \bar{\omega}^v$ , $\bar{\kappa}^i = 1 / \lambda \bar{\omega}^i$ . This implies that effectively an entry of the transmission matrix is a sample + +![](images/46bfd089b3b35d5bc046dad013056388f346bc9e16d249fa49982443b4d7092f.jpg) +(a) + +![](images/0a5fc2e4f939bdd085a233905ed2d980356a80886f347bad34b87a953175c55f.jpg) +(b) + +![](images/bd09b39ba8f1ab55a3632705decb26df4068d7436e42bebbaa667ed447d32ce1.jpg) +(c) +Figure 6. Spectrum structure: (a) An $xz$ slice out of the spectrum of $\hat{n} (\omega)$ . Entries of a weakly-scattering transmission matrix limited by an aperture NA only lie inside the butterfly area. (b) Zooming on the center right area of (a) (purple square). Assuming the $\omega_{z}$ axis is cut at $\pm \frac{1}{2}\Lambda^{M}$ , the transmission matrix entries inside the dashed blue area are maintained, and the entries in the dashed red area are lost. (c) The energy at a slice from the spectrum of the volume in Fig. 3(a). Since the volume structure is smooth, most energy is concentrated at the low frequencies. As a result the cropped red region in (b) do not contain a lot of energy. + +from the 3D Fourier transform of the volume at frequency $\kappa = \bar{\kappa}^i -\bar{\kappa}^v$ . Note that throughout this paper we use bars $\bar{\omega}$ , $\bar{\kappa}$ to denote vectors with norm constraints, and use normal bold fonts to denote standard 3D vectors, such as frequencies $\kappa$ in the 3D spectrum. Unfortunately, a large portion of the 3D spectrum cannot be measured. Recalling that the illumination and viewing directions are limited by an aperture of width NA, the subset of 3D frequencies captured by the transmission matrix occupies only a butterfly shape, and the rest of the cone is missing, see illustration in Fig. 6(a). Using paraxial approximation it can be shown [28, 38] that the range of frequencies which can be generated is + +$$ +\left| \kappa_ {z} \right| \leq \mathbf {N A} \left| \boldsymbol {\kappa} _ {x y} \right| - \frac {\lambda}{2} \left| \boldsymbol {\kappa} _ {x y} \right| ^ {2}, \quad \text {w i t h} \quad \left| \boldsymbol {\kappa} _ {x y} \right| = \sqrt {\kappa_ {x} ^ {2} + \kappa_ {y} ^ {2}}. \tag {7} +$$ + +The missing cone is a serious problem for diffraction tomography algorithms aiming to recover the 3D structure of the refractive index $n(\mathbf{r})$ from columns of the transmission matrix, since many of the frequencies of the volume cannot be recovered. We argue that in the context of transmission matrices, the missing cone problem turns into an advantage. The fact that the transmission matrix only captures a limited range of frequencies allows us to approximate it with fewer samples, as derived below. + +Minimal sampling: Our goal is to approximate a transmission matrix with layers. We start by asking what spacing between layers will allow an exact reconstruction of the transmission matrix. If we approximate the volume $\pmb{n}(\mathbf{r})$ with a set of planes, we essentially sample a continuous signal. Classical Nyquist theory states that if we sample a signal at intervals of $\Delta_z$ , its Fourier transform includes replicas spaced by $\Lambda = 1 / \Delta_z$ . To achieve a good reconstruction, we want replicas to happen only in areas of the spectrum that the transmission matrix does not sample. Since the $\kappa_x, \kappa_y$ components of our illuminations are limited by the aperture $|\kappa_{xy}| \leq \Omega_f$ , and given the butterfly shape in Eq. (7), the $\kappa_z$ + +frequencies we need to maintain are bounded to + +$$ +\left| \kappa_ {z} \right| \leq \frac {\mathrm {N A} \Omega_ {f}}{2}. \tag {8} +$$ + +Therefore, for an error-free reconstruction, the minimal Fourier range and primal spacing are + +$$ +\Lambda^ {*} = \mathbf {N A} \boldsymbol {\Omega} _ {f}, \quad \Delta_ {z} ^ {*} = \frac {1}{\mathbf {N A} \boldsymbol {\Omega} _ {f}} = \frac {\lambda}{\mathbf {N A} ^ {2}}. \tag {9} +$$ + +For a scattering volume of thickness $d$ , the number of required layers is: + +$$ +M ^ {*} = \frac {d \cdot \mathbf {N A} ^ {2}}{\lambda}. \tag {10} +$$ + +For typical numbers, consider e.g. $\lambda = 0.5\mu m$ and $\mathbf{N}\mathbf{A} = 0.5$ , the plane spacing should be relatively small, and $\Delta_z \leq 2\mu m$ . For a modest tissue thickness of $d = 200\mu m$ , we will need $M = 100$ planes. Clearly it is impractical to realize such a large number of planes by a physical correction system. + +# 4.2. Sparse multi-slice approximations + +Consider a scattering sample of thickness $d$ , which we try to approximate using $M < M^{*}$ layers separated by $\Delta_z = \frac{d}{M}$ . Our goal is to show that the fitting error $\mathcal{E}_M$ defined in Eq. (5) decays relatively fast with $M$ and we can get a reasonable approximation to the transmission matrix even if the number of layers is lower than the exact prediction in Eq. (10). This is a result of two main properties: (i) The volumes describing realistic tissue samples have more energy in the low frequencies; and (ii) the structure of the missing cone implies that, in any case, a significant portion of the spectrum is not sampled by the transmission matrix. + +To understand this, consider a naive selection of $M$ layers $\rho_{1},\ldots ,\rho_{M}$ . Rather than actually solving an optimization problem, we use the ground truth volume $\hat{n} (\kappa)$ , and simply set to zero any frequency content of the RI volume $\hat{n} (\kappa)$ at $\kappa_z$ values larger than the possible Nyquist range + +$$ +\Lambda^ {M} = \frac {1}{\Delta_ {z}} = \frac {\lambda M}{d}, \tag {11} +$$ + +and we then Fourier transform $\hat{n}(\kappa)$ to $\pmb{n}(\mathbf{r})$ and sample planes at spacing $\Delta_z = d / M$ . The error of this approximation is basically the integral of content above the cut-off $\Lambda^M$ . Since according to Eq. (11), the cut-off frequency $\Lambda^M$ scales linearly with the number of layers, the portion of the spectrum which is lost by this low-pass operation scales linearly with the number of layers $M$ . Thus, the naive answer is that the error of a multi-slice approximation decays linearly with $M$ . In practice we show below that the error decays much faster, since large areas from the 3D spectrum of the sample are not used by the transmission matrix. + +To gain intuition, consider Fig. 6(b). We note that the butterfly shape is such that low 2D frequencies (i.e. low $|\kappa_{xy}|$ ), also span less content along the $\kappa_z$ axes, and hence these frequencies are not lost even with low bandwidths $\Lambda^M$ . For the higher $|\kappa_{xy}|$ frequencies, the lower $\kappa_z$ part, marked in dashed blue in Fig. 6(b) is preserved, while the higher portion marked in red is lost. + +![](images/2ddb4ea6488694c3777f31370d0477afa592bcb08ab53079e3459e6c64263bdf.jpg) +(a) Support primal + +![](images/2665e7a357dd917add952742cb4d40055332310bf78b39f5de47e5510472e173.jpg) +(b) Support freq. + +![](images/244545752f597c82df357181f9f1869e68249011bbda892398d06aa4786b500c.jpg) +(c) Local directions +Figure 7. Layer support: (a) As point sources expand while propagating through the volume, the width of the aberration layers should be wider than the measured support $\Omega_{p}$ . (b) To express compact support transmission matrices in the Fourier basis, we illuminate the volume by a set of plane waves passing through an aperture. (c) The local cone of illumination angles reaching different points inside the volume is much smaller than the actual range of incoming illuminations. + +Another important property of tissue is that its RI is locally smooth, and therefore if we look at its spectrum (Fig. 6(c)), we have much more energy in low frequencies than in high ones. This also agrees with the fact that tissue is forward scattering. Therefore, in the red areas of Fig. 6(b) that are not sampled by the transmission matrix, there is less energy than in the lower frequencies. Thus, despite the fact that the multi-slice approximation sacrifices such frequencies, not much energy is being lost. + +In most cases, it is hard to give analytic formulas for the decay of $\pmb{\mathcal{E}}_M$ as a function of $M$ , but the supplementary file contains some derivation under simplifying assumptions. + +Multiple scattering. The results in Fig. 3(e) include multislice fits with both lower and higher optical depth values up to $OD = 4$ , showing that the fitting error still decays rapidly with the number of layers. + +While it is hard to give analytical predictions for the multiple scattering case, it appears that when the optical depth of the tissue is not too high, similar results hold. The reason is that for a forward-scattering material such as tissue, light paths only scatter at small angles, and after a small number of scattering events they are not completely scrambled. For an intuition, in the supplementary file we use a Monte-Carlo simulation illustrating that for a forward-scattering material, after a modest number of scattering events, most light paths still remain within the butterfly area of the spectrum. This is usually the regime that wavefront-shaping algorithms like [1, 2] attempt to tackle, as such algorithms attempt to push the depth at which conventional microscopes can see, but they do not yet attempt to image extremely deep where scattering is fully diffused. + +# 4.3. Bounded support matrices + +As stated above, much fewer layers are needed if we only attempt to fit transmission matrices sampled over a small area of size $\Omega_p\times \Omega_p$ + +When fitting layers, we need to account for the fact that a + +point/plane illumination scattering through a volume is expanding. Thus, the aberration layers required to explain the transmission matrix are wider than the imaged area of size $\Omega_{p} \times \Omega_{p}$ , see Fig. 7(a). Expressing a transmission matrix with a compact support at the frequency basis is equivalent to illuminating the sample with plane waves spanning a wide set of angles, but all the waves pass through a narrow aperture of size $\Omega_{p} \times \Omega_{p}$ . In this case most points inside the volume usually do not receive light from all angles. In Fig. 7(c) we plot the narrow cone of angles arriving at three different points (a ray leaving a point $\mathbf{p}_{\mathbf{j}}$ at angle $\bar{\omega}$ is included in the cone only if this angle is not cropped by the aperture at the back of the sample). Since the cone of light reaching each point is narrower than the full numerical aperture, a local Fourier transform would not have content over the full butterfly area, but cover an even lower range of axial $\omega_{z}$ frequencies. Hence, following the analysis in the previous section, it can be sampled with fewer layers. We visualize the local spectra in the supplement. + +# 5. Discussion + +This paper studies how well we can fit transmission matrices with multi-slice models and how many layers are needed in practice. We show that the required number of layers is much lower than the prediction of Nyquist sampling theory. This is because the missing cone problem of 3D microscopy turns into an advantage when analyzing transmission matrices. Moreover, as tissue is forward scattering, its 3D spectrum contains most energy in the low frequencies and much less energy in the higher ones. Hence, even if the sparse sampling acts as a low-pass on the large axial frequencies, not much energy is lost. + +We validate this prediction using both synthetic simulations and realistic lab measurements. We show that even though realistic transmission matrices contain backward and side-ward scattering which are not modeled by multi-slice matrices, they provide a reasonable approximation that is good enough for wavefront shaping focusing. + +Our findings suggest that multi-slice models can be applicable for the design of non-local wavefront shaping systems, allowing for a volumetric correction with a small number of layered SLMs [16, 27, 35, 36]. Even if $2 - 3$ SLM layers may not be enough for correcting the full field-of-view of a wide image through a thick tissue, they can allow us to increase the support of the fitted area by much more than $\times 2 - 3$ thus they can significantly accelerate a sequential wavefront-shaping correction of a wide image. + +Acknowledgments: This research was funded by ERC SpeckleCorr-101043471, ISF 563/24. + +Code and data: can be obtained at https:// webee.technion.ac.il/people/anat.levin/ papers/LayerdTmat(cvpr25_code.zip + +# References + +[1] Dror Aizik and Anat Levin. Non-invasive and noise-robust light focusing using confocal wavefront shaping. Nature communications, 15(5575), 2024. +[2] Dror Aizik, Ioannis Gkioulekas, and Anat Levin. Fluorescent wavefront shaping using incoherent iterative phase conjugation. Optica, 9(7):746-754, 2022. +[3] M. Alterman, C. Bar, I. Gkioulekas, and A. Levin. Imaging with local speckle intensity correlations: theory and practice. ACM TOG, 2021. +[4] YoonSeok Baek, Hilton B. de Aguiar, and Sylvain Gigan. Phase conjugation with spatially incoherent light in complex media. Nature Photonics, 17(12):1114-1119, 2023. +[5] Paul Balondrade, Victor Barolle, Nicolas Guigui, Emeric Auriant, Nathan Rougier, Claude Boccara, Mathias Fink, and Alexandre Aubry. Multi-spectral reflection matrix for ultrafast 3d label-free microscopy. Nature Photonics, 18(10): 1097-1104, 2024. +[6] Chen Bar, Marina Alterman, Ioannis Gkioulekas, and Anat Levin. A monte carlo framework for rendering speckle statistics in scattering media. ACM TOG, 2019. +[7] Chen Bar, Ioannis Gkioulekas, and Anat Levin. Rendering near-field speckle statistics in scattering media. ACM TOG, 2020. +[8] Martin Booth. Adaptive optical microscopy: The ongoing quest for a perfect image. Light: Science and Applications, 3:e165, 2014. +[9] Bingquan Chen and Jakob J. Stamnes. Validity of diffraction tomography based on the first born and the first rytrov approximations. Appl. Opt., 37(14):2996-3006, 1998. +[10] Wonshik Choi, Christopher Fang-Yen, Kamran Badizadegan, Seungeun Oh, Niyom Lue, Ramachandra R Dasari, and Michael S Feld. Tomographic phase microscopy. Nature methods, 4(9):717-719, 2007. +[11] Shwetadwip Chowdhury, Michael Chen, Regina Eckert, David Ren, Fan Wu, Nicole Repina, and Laura Waller. High-resolution 3d refractive index microscopy of multiple-scattering samples from intensity images. Optica, 6(9): 1211-1219, 2019. +[12] Anat Daniel, Dan Oron, and Yaron Silberberg. Light focusing through scattering media via linear fluorescence variance maximization, and its application for fluorescence imaging. Optics Express, 27(15):21778, 2019. +[13] A. J. Devaney. Inverse-scattering theory within the rytov approximation. Opt. Lett., 6(8):374-376, 1981. +[14] Amicia D. Elliott. Confocal microscopy: Principles and modern practices. Current Protocols in Cytometry, 92(1): e68, 2020. +[15] Brandon Y. Feng, Haiyun Guo, Mingyang Xie, Vivek Boominathan, Manoj K. Sharma, Ashok Veeraraghavan, and Christopher A. Metzler. Neuws: Neural wavefront shaping for guidestar-free imaging through static and dynamic scattering media. Science Advances, 9(26):eadg4671, 2023. +[16] T. Furieri, A. Bassi, and S. Bonora. Large field of view aberrations correction with deformable lenses and multi conjugate adaptive optics. Journal of Biophotonics, 16(12): e202300104, 2023. + +[17] Omri Haim, Jeremy Boger-Lombard, and Ori Katz. Image-guided computational holographic wavefront shaping. Nature Photonics, 2024. +[18] Karen Hampson, Raphael Turcotte, Donald Miller, Kazuhiro Kurokawa, Jared Males, Na Ji, and Martin Booth. Adaptive optics for high-resolution imaging. Nature Reviews Methods Primers, 1:68, 2021. +[19] John W. Hardy. Adaptive Optics for Astronomical Telescopes. 1998. +[20] Yi Hua, M. Salman Asif, and Aswin C. Sankaranarayanan. Spatial and axial resolution limits for mask-based lensless cameras. Opt. Express, 31(2):2538-2551, 2023. +[21] Seungwon Jeong, Ye-Ryoung Lee, Wonjun Choi, Sungsam Kang, Jin Hee Hong, Jin-Sung Park, Yong-Sik Lim, HongGyu Park, and Wonshik Choi. Focusing of light energy inside a scattering medium by controlling the time-gated multiple light scattering. Nature Photonics, 12(5):277-283, 2018. +[22] Na Ji. Adaptive optical fluorescence microscopy. Nature Methods, 14:374-380, 2017. +[23] Yonghyeon Jo, Ye-Ryoung Lee, Jin Hee Hong, Dong-Young Kim, Junhwan Kwon, Myunghwan Choi, Moonseok Kim, and Wonshik Choi. Through-skull brain imaging in vivo at visible wavelengths via dimensionality reduction adaptive-optical microscopy. Science Advances, 8(30):eabo4366, 2022. +[24] Andrew Jr, Alon Harris, Josh Gross, Ingrida Januleviciene, Aaditya Shah, and Brent Siesky. Optical coherence tomography angiography: An overview of the technology and an assessment of applications for clinical research. British Journal of Ophthalmology, 101, 2016. +[25] Sungsam Kang, Pilsung Kang, Seungwon Jeong, Yongwoo Kwon, Taeseok D. Yang, Jin Hee Hong, Moonseok Kim, Kyung-Deok Song, Jin Hyoung Park, Jun Ho Lee, Myoung Joon Kim, Ki Hean Kim, and Wonshik Choi. High-resolution adaptive optical imaging within thick scattering media using closed-loop accumulation of single scattering. Nature Communications, 8(1):2157, 2017. +[26] Sungsam Kang, Yongwoo Kwon, Hojun Lee, Seho Kim, Jin Hee Hong, Seokchan Yoon, and Wonshik Choi. Tracing multiple scattering trajectories for deep optical imaging in scattering media. Nature communications, 14(1), 2023. +[27] Marie Laslandes, Matthias Salas, Christoph K. Hitzenberger, and Michael Pircher. Increasing the field of view of adaptive optics scanning laser ophthalmoscopy. Biomed. Opt. Express, 8(11):4811-4826, 2017. +[28] V. Lauer. New approach to optical diffraction tomography yielding a vector equation of diffraction tomography and a novel tomographic microscope. Journal of Microscopy, 205 (2):165-176, 2002. +[29] F. C. Lin and M. A. Fiddy. The born-rytov controversy: I. comparing analytical and approximate expressions for the one-dimensional deterministic case. J. Opt. Soc. Am. A, 9 (7):1102–1110, 1992. +[30] Jerome Mertz. Introduction to optical microscopy. Cambridge University Press, 2019. +[31] Sagi Monin, Aswin Sankaranarayanan, and Anat Levin. Exponentially-wide etendue displays using a tilting cascade. IEEE ICCP, pages 1-12, 2022. + +[32] Ulysse Najar, Victor Barolle, Paul Balondrade, Mathias Fink, Claude Boccara, and Alexandre Aubry. Harnessing forward multiple scattering for optical imaging deep inside an opaque medium. Nature Communications, 15(1), 2024. +[33] Tan Nguyen, Mikhail Kandel, Marcello Rubessa, Matthew Wheeler, and Gabriel Popescu. Gradient light interference microscopy for 3d imaging of unlabeled specimens. Nature Communications, 8, 2017. +[34] Galya Stern and Ori Katz. Noninvasive focusing through scattering layers using speckle correlations. Optics Letters, 44(1):143, 2018. +[35] Jörgen Thaung, Per Knutsson, Zoran Popovic, and Mette Owner-Petersen. Dual-conjugate adaptive optics for wide-field high-resolution retinal imaging. Opt. Express, 17(6): 4454-4467, 2009. +[36] Tsai wei Wu and Meng Cui. Numerical study of multiconjugate large area wavefront correction for deep tissue microscopy. Opt. Express, 23(6):7463-7470, 2015. +[37] Gil Weinberg, Elad Sunray, and Ori Katz. Noninvasive megapixel fluorescence microscopy through scattering layers by a virtual incoherent reflection matrix. Science Advances, 10(47):eadl5218, 2024. +[38] Emil Wolf. Three-dimensional structure determination of semi-transparent objects from holographic data. Optics Communications, 1(4):153-156, 1969. +[39] Tomer Yeminy and Ori Katz. Guidestar-free image-guided wavefront shaping. Science Advances, 7(21):eabf5364, 2021. +[40] Lei Zhu, Fernando Soldevila, Claudio Moretti, Alexandra d'Arco, Antoine Boniface, Xiaopeng Shao, Hilton B. de Aguiar, and Sylvain Gigan. Large field-of-view non-invasive imaging through scattering layers using fluctuating random illumination. Nature Communications, 13:1447, 2022. \ No newline at end of file diff --git a/2025/Understanding Multi-layered Transmission Matrices/images.zip b/2025/Understanding Multi-layered Transmission Matrices/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..67a82ef3f7de60eabff71b9dc7889b047bd17530 --- /dev/null +++ b/2025/Understanding Multi-layered Transmission Matrices/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80f5f0546c6fe48175fbf20a432ae13ca138ce0d3b8ecdbb35e51321ae2b79f0 +size 350104 diff --git a/2025/Understanding Multi-layered Transmission Matrices/layout.json b/2025/Understanding Multi-layered Transmission Matrices/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1d2f2bc840d44b0428a6b2865e1c2bd2e2b9619d --- /dev/null +++ b/2025/Understanding Multi-layered Transmission Matrices/layout.json @@ -0,0 +1,10131 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 103, + 465, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 103, + 465, + 121 + ], + "spans": [ + { + "bbox": [ + 146, + 103, + 465, + 121 + ], + "type": "text", + "content": "Understanding multi-layered transmission matrices" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 153, + 137, + 457, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 137, + 457, + 179 + ], + "spans": [ + { + "bbox": [ + 153, + 137, + 457, + 179 + ], + "type": "text", + "content": "Anat Levin and Marina Alterman \nDepartment of Electrical and Computer Engineering, Technion Haifa, Israel" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 207, + 200, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 207, + 200, + 220 + ], + "spans": [ + { + "bbox": [ + 151, + 207, + 200, + 220 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 233, + 295, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 233, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 295, + 365 + ], + "type": "text", + "content": "Wavefront shaping systems attempt to correct aberrations caused by tissue scattering, by placing a spatial light modulator (SLM) in the optical path and using it to reshape the light wave emitted from a target of interest deep inside the tissue. However, the field-of-view we can correct with one SLM pattern is extremely small, because inherently the aberration is caused by the 3D tissue structure while the SLM is planar. To overcome this, multi-conjugate correction systems built with multiple layered SLMs have been introduced, which attempt to approximate the 3D tissue structure with multiple planar aberrations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 365, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 568 + ], + "type": "text", + "content": "However, multi-conjugate systems with a large number of SLMs are not easy to construct. Therefore, it is important to understand how many SLM layers are actually needed for a good correction, and whether there are practical benefits from correction systems with a relatively small number of layers. To help in the design of future systems, this paper analyzes multi-layer corrections. We show that the well-known missing cone problem which fundamentally limits the axial resolution of microscopes, turns into an advantage when designing 3D correction systems. Due to this property less information should be captured by the correction system, facilitating sparser approximations. We also show that even if the number of available layers is insufficient for a reasonable approximation of the full 3D tissue volume, it can significantly expand the field-of-view when compared to single-layer correction systems, and therefore can largely accelerate their operation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 578, + 135, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 135, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 135, + 590 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 700 + ], + "type": "text", + "content": "Scattering is a central challenge in multiple imaging tasks ranging from astronomy to microscopy, and biomedical imaging. It degrades our ability to resolve stars through telescopes, image into the bottom of the sea, and drive through foggy weather. Most importantly in the context of this paper, scattering severely limits our ability to image through biological tissue. This happens due to small variations in the refractive index of the tissue volume, which scatters the propagating light." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "Numerous computational optics approaches have been" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 208, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 208, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 555, + 567 + ], + "type": "text", + "content": "developed for imaging through scattering tissue. These include computational restoration algorithms which attempt to fit the observed data with various models for the aberration and the target of interest [4, 5, 15, 17, 21, 23, 25, 32, 37, 39, 40]. Also, optical systems such as a confocal microscope [14] or an OCT [24], aim to filter out the scattered light and maintain only the ballistic unscattered part. However, as we attempt to see deeper into a scattering layer, ballistic light weakens significantly, and the complexity of aberration models increases. Moreover, biological target of interest such as fluorescent tissue components are very weak and as a result the scattered signal is also of a very low signal to noise ratio (SNR), further limiting digital correction models. A particularly attractive approach for scattering aberration removal is wavefront shaping, where one places a spatial light modulator (SLM) in the optical path of the microscope and uses it to correct the wavefront emitted from a point deep inside the tissue, so that light photons emerging from a single target point are brought into a single sensor point, despite the tissue aberration. Wavefront shaping is a particularly attractive solution for fluorescent imaging because the correction is done optically rather than digitally. Unlike ballistic filtering approaches, all light photons are used. Since the correction is done in the optical path, the limited number of available photons is brought into a single detector pixel and can be measured with higher SNR. As a result, wavefront shaping has the potential to revolutionize tissue imaging and allow us to overcome the fundamental limitations of scattering, and image very deep inside biological targets at a high SNR." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "The main limitation for wavefront shaping approaches is that the aberrations of neighboring points inside volumetric tissue vary rapidly, and in practice a single correction mask can only allow imaging of a small local field-of-view. This is because the scattering process happens throughout the 3D volume, while the aberration correction only happens on a planar SLM. Earlier demonstrations of wavefront shaping have considered targets placed at a distance behind a thin scattering layer, and this far-field arrangement has allowed them to correct larger regions with a single planar modulation [12, 15, 34, 39]. Similarly in the context of adaptive-optics, researchers have attempted to correct aber" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23164" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "rations due to imperfect optics or a refractive index mismatch [8, 18, 22]. These aberrations are introduced along the optical path, before the tissue, so the correction can apply to a larger area. However, in most biomedical applications the fluorescent target of interest is embedded within the scattering volume. If our goal is to correct scattering within the tissue itself, the field-of-view which can be corrected by a single planar modulation is extremely small, usually only a few microns [3]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 182, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 182, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 56, + 182, + 296, + 445 + ], + "type": "text", + "content": "To correct larger fields of view, multi conjugate correction systems have been proposed. Such systems use multiple planar SLMs placed such that each of them is conjugate to a different plane inside the tissue volume. Each such SLM should realize a correction for the aberration happening inside a thin layer of the volume and the combination of such multi-plane corrections should approximate the 3D variation in refractive index, and allow us to correct it. Fig. 1 visualizes the concept of a multi-conjugate correction [16, 27, 35, 36]. In practice, multi-plane correction systems are more complicated than the simplified visualization of Fig. 1, as one needs to add multiple relay systems between the SLMs. This is first because there are physical limitations on the minimal distance between 2 SLMs, and second because we need to revert the order of the correction layers in the inversion path [19]. This makes multiconductor correction systems complex and cumbersome to build. Previous experimental demonstrations have realized corrections with no more than two SLMs. While two-layer correction systems have demonstrated improved performance compared to single layer correction, two planes clearly cannot model a thick 3D tissue structure." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 446, + 295, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 295, + 543 + ], + "type": "text", + "content": "Our goal in this research is to analytically study multislice models and attempt to answer the following questions: How many planes are required to approximate the 3D aberration of a scattering tissue volume? How does the quality of the correction degrade if we use a sparser approximation? If a construction of a practical correction system is limited to only to M layers, what is the maximum field of view that can be effectively corrected?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 544, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 616 + ], + "type": "text", + "content": "To this end, we use the mathematical model of a transmission matrix, which stores the propagation of coherent wavefronts from one end of a scattering volume to the other. We study how well we can approximate such transmission matrices with a multi-slice model, and how does the error depend on the number of fitted layers." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "The straightforward answer to this question is to say that the spacing between layers should match the axial resolution of the microscope. For example, using an " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{NA} = 0.5" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " objective at wavelength " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "500nm" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": ", we get that the spacing between layers should be lower than " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "2\\mu m" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": ". Handling a modest tissue thickness of, e.g., " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "200\\mu m" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " already requires 100 slices for a valid approximation. Clearly building a multiconjugate wavefront shaping systems with so many layers" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 322, + 63, + 558, + 252 + ], + "blocks": [ + { + "bbox": [ + 322, + 63, + 558, + 252 + ], + "lines": [ + { + "bbox": [ + 322, + 63, + 558, + 252 + ], + "spans": [ + { + "bbox": [ + 322, + 63, + 558, + 252 + ], + "type": "image", + "image_path": "9bd6169b3d3f269f4ac098e352a507d116a8e3e75e350f687cba249d2cd12475.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 264, + 555, + 396 + ], + "lines": [ + { + "bbox": [ + 313, + 264, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 555, + 396 + ], + "type": "text", + "content": "Figure 1. Layered wavefront shaping: (Top) Consider a simple microscope setup where we use a 4F relay to image the light emerging from a tissue sample. Light is aberrated on its way to the sensor, and we see speckle patterns rather than sharp spots. (Middle) A single planar SLM can correct aberration for one spot (see red beam), but light emerging from nearby spots may not be corrected (the yellow beam generates a speckle pattern on the sensor). (Bottom) The field of view of the correction can possibly be extended with a multi-layer correction. The goal is to form a correction replicating the 3D structure of volumetric aberration inside the tissue. The figure illustrates two wavefronts from nearby spots (red and yellow) passing through the same volumetric correction, and both focus into a sharp spot behind the tissue." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 402, + 373, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 402, + 373, + 413 + ], + "spans": [ + { + "bbox": [ + 314, + 402, + 373, + 413 + ], + "type": "text", + "content": "is impractical." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 413, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 555, + 604 + ], + "type": "text", + "content": "In this research we take inspiration from [20], and offer a thorough analysis of the frequency content of transmission matrices, combining the forward scattering nature of biological tissue with the structure of the missing cone. We show that while the missing cone structure is known as one of the fundamental limitations of 3D microscopy, in the context of transmission matrices this limitation turns into an advantage. Due to the missing cone structure, many of the frequencies of the volumetric aberration do not participate in the transmission matrix, and therefore the number of layers required for approximating a transmission matrix is way smaller than the Nyquist limit. However, there are no free lunches and the number of layers scales with the tissue thickness. To explain a few hundred micron thick tissue one needs a few dozen layers, which is still too much to be realized by a physical correction system." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Given this conclusion we aim to understand if there are any practical benefits in realizing a multi-conjugate correction system with a small number of say 2 - 4 layers. We show that while a small number of layers cannot well approximate the 3D structure of the tissue, it can allow us to realize a correction over a limited field-of-view. The area which can be corrected with " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": " aberration layers is significantly larger than the area which can be corrected with one layer, and therefore it can significantly accelerate sequen" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23165" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 309, + 132 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 309, + 132 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 309, + 132 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 309, + 132 + ], + "type": "image", + "image_path": "f0721efb15f9ea6c7b32268b5b6ba3c736c1699d40f6b520e16a78a964fd6112.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 172, + 135, + 217, + 145 + ], + "lines": [ + { + "bbox": [ + 172, + 135, + 217, + 145 + ], + "spans": [ + { + "bbox": [ + 172, + 135, + 217, + 145 + ], + "type": "text", + "content": "(a) Full setup" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 326, + 74, + 386, + 129 + ], + "blocks": [ + { + "bbox": [ + 326, + 74, + 386, + 129 + ], + "lines": [ + { + "bbox": [ + 326, + 74, + 386, + 129 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 386, + 129 + ], + "type": "image", + "image_path": "a5c58ed619bc8c45c2f1d477f8c5034216e4fd639836400a17b0ee3e9b90dc3b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 150, + 555, + 199 + ], + "lines": [ + { + "bbox": [ + 55, + 150, + 555, + 199 + ], + "spans": [ + { + "bbox": [ + 55, + 150, + 555, + 199 + ], + "type": "text", + "content": "Figure 2. Setup for transmission matrix acquisition: (a) A setup for transmission matrix imaging where a set of basis illuminations at the back of the sample propagate through the volume, and the scattered wavefronts are measured by a camera at the other end. In (b,c) we zoom on the sample illustrating two possible illumination bases. (b) A set of directional wavefronts (Fourier basis). (c) A set of wavefronts focused to points at the edge of the volume (primal basis). Both transmission matrices can be related by a simple Fourier transform. (d) A multi-slice model approximates the volumetric scattering with a set of planar aberrations." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 394, + 74, + 453, + 129 + ], + "blocks": [ + { + "bbox": [ + 328, + 135, + 384, + 144 + ], + "lines": [ + { + "bbox": [ + 328, + 135, + 384, + 144 + ], + "spans": [ + { + "bbox": [ + 328, + 135, + 384, + 144 + ], + "type": "text", + "content": "(b) Fourier illum." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 394, + 74, + 453, + 129 + ], + "lines": [ + { + "bbox": [ + 394, + 74, + 453, + 129 + ], + "spans": [ + { + "bbox": [ + 394, + 74, + 453, + 129 + ], + "type": "image", + "image_path": "c49845d22af3999f33d70e1c082ade41175702092ae11814836b23961ce4eec6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 135, + 448, + 144 + ], + "lines": [ + { + "bbox": [ + 399, + 135, + 448, + 144 + ], + "spans": [ + { + "bbox": [ + 399, + 135, + 448, + 144 + ], + "type": "text", + "content": "(c) Point illum." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 465, + 75, + 524, + 129 + ], + "blocks": [ + { + "bbox": [ + 465, + 75, + 524, + 129 + ], + "lines": [ + { + "bbox": [ + 465, + 75, + 524, + 129 + ], + "spans": [ + { + "bbox": [ + 465, + 75, + 524, + 129 + ], + "type": "image", + "image_path": "737cbeb37f676e96a48d7e780ea731c4776d5f743e73b0ed249eedd22cbb31b1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 135, + 527, + 145 + ], + "lines": [ + { + "bbox": [ + 463, + 135, + 527, + 145 + ], + "spans": [ + { + "bbox": [ + 463, + 135, + 527, + 145 + ], + "type": "text", + "content": "(d) Layered approx." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 221, + 295, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 221, + 295, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 295, + 269 + ], + "type": "text", + "content": "tial correction of the field of view. We validate our findings using real transmission matrices captured in the lab as well as synthetic transmission matrices simulated using accurate wave propagation models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 277, + 178, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 178, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 178, + 289 + ], + "type": "text", + "content": "2. Problem formulation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 293, + 296, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 296, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 296, + 412 + ], + "type": "text", + "content": "Imaging setup: For simplicity, the following derivation considers coherent scalar waves, ignoring polarization. In Fig. 2(a) we visualize a setup for transmission matrix acquisition. A scattering medium, such as tissue, is illuminated by a coherent source " + }, + { + "bbox": [ + 55, + 293, + 296, + 412 + ], + "type": "inline_equation", + "content": "\\mathbf{e}^k" + }, + { + "bbox": [ + 55, + 293, + 296, + 412 + ], + "type": "text", + "content": " at one side of the medium. After propagating through the medium, the resulting coherent wavefront " + }, + { + "bbox": [ + 55, + 293, + 296, + 412 + ], + "type": "inline_equation", + "content": "\\mathbf{t}^k" + }, + { + "bbox": [ + 55, + 293, + 296, + 412 + ], + "type": "text", + "content": " is recorded. Usually " + }, + { + "bbox": [ + 55, + 293, + 296, + 412 + ], + "type": "inline_equation", + "content": "\\mathbf{t}^k" + }, + { + "bbox": [ + 55, + 293, + 296, + 412 + ], + "type": "text", + "content": " are complex valued speckle patterns with a pseudo-random structure. The illumination is translated and multiple wavefronts are recorded and stored as columns of a transmission matrix " + }, + { + "bbox": [ + 55, + 293, + 296, + 412 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 293, + 296, + 412 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "type": "text", + "content": "The transmission matrix: Since the propagation of coherent light is linear, if we want to predict the propagation of any wavefront " + }, + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "type": "text", + "content": ", we simply express it as a superposition of the recorded incoming illuminations " + }, + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{u} = \\sum_{k} \\mathbf{u}(k) \\mathbf{e}^{k}" + }, + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{u}(k)" + }, + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "type": "text", + "content": " are complex scalars, and multiply it by the transmission matrix, to predict " + }, + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T} \\mathbf{u}" + }, + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "type": "text", + "content": ". The illumination wavefronts " + }, + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{e}^{k}" + }, + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "type": "text", + "content": " can be selected at different bases. For example, the illumination can be translated at the Fourier plane, and then the wavefronts illuminating the sample are plane waves at different angles (Fig. 2(b)). Alternatively one can focus the light to a point at the back plane of the sample (Fig. 2(c)). This configuration is more appropriate for wavefront shaping algorithms whose goal is to recover wavefronts focusing inside the tissue and not far behind it. In both cases, however, the illuminations we can create are limited by the numerical aperture of the illuminating objective, which we mark here as NA. Similarly, the scattered wavefronts are measured via an imaging objective with the same bounded NA. This NA would serve as a central component in the results derived below. For simplicity of resulting equations, for the rest of this manuscript we assume the imaging and illumination objectives are focused at the same plane, such that if we would remove the scattering sample and image the sources directly, we would see the " + }, + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{e}^{k}" + }, + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "type": "text", + "content": " illuminations as sharp spots on the sensor plane. That is, if we use directional" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 221, + 554, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 554, + 293 + ], + "type": "text", + "content": "illuminations, we also measure the Fourier transform of the scattered waves, namely place the sensor at the Fourier plane of the imaging objective. If we use point sources at the illumination end of the medium, we also focus the imaging objective at the same plane, so we measure a speckle image of a relatively small support." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 293, + 554, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 554, + 359 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 554, + 359 + ], + "type": "text", + "content": "Range and resolution: We assume that in the primal basis the illuminators are spaced within an area of " + }, + { + "bbox": [ + 313, + 293, + 554, + 359 + ], + "type": "inline_equation", + "content": "\\Omega_p \\times \\Omega_p" + }, + { + "bbox": [ + 313, + 293, + 554, + 359 + ], + "type": "text", + "content": ", and we sample them at the Nyquist resolution, so for an illumination wavelength " + }, + { + "bbox": [ + 313, + 293, + 554, + 359 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 293, + 554, + 359 + ], + "type": "text", + "content": " they are separated by a pitch " + }, + { + "bbox": [ + 313, + 293, + 554, + 359 + ], + "type": "inline_equation", + "content": "\\Delta_p = \\lambda/(2\\mathbf{NA})" + }, + { + "bbox": [ + 313, + 293, + 554, + 359 + ], + "type": "text", + "content": ". Equivalently, in the Fourier domain, we illuminate the sample with plane waves of the form" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 402, + 361, + 553, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 361, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 402, + 361, + 553, + 374 + ], + "type": "interline_equation", + "content": "\\mathbf {e} (\\mathbf {r}) = e ^ {2 \\pi i (\\bar {\\kappa} \\cdot \\mathbf {r})} \\tag {1}", + "image_path": "db93faaed00d36fb1e00fa32c3f1b069a4e7946fcf158fa628f5c8e1897bb509.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "text", + "content": " is a point inside the 3D volume, and " + }, + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\bar{\\kappa} = \\frac{1}{\\lambda}\\bar{\\omega}" + }, + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "text", + "content": " is the wavelength and " + }, + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\bar{\\omega}" + }, + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "text", + "content": " is a unit norm direction vector, i.e. the last entry of the direction vector " + }, + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\omega_{z}" + }, + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\omega_{z} = \\sqrt{1 - (\\omega_{x}^{2} + \\omega_{y}^{2})}" + }, + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "text", + "content": ". For simplicity of notation, we assume the refractive index (RI) of the leading medium is 1. We can only use illumination direction inside the numerical aperture " + }, + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\sqrt{\\omega_x^2 + \\omega_y^2}\\leq \\mathbf{N}\\mathbf{A}" + }, + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "text", + "content": " (or equivalently " + }, + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\sqrt{\\kappa_x^2 + \\kappa_y^2}\\leq \\Omega_f" + }, + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\Omega_{f} = \\mathbf{N A} / \\lambda" + }, + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "text", + "content": "). To sample the transmission matrix at the Nyquist pitch, the " + }, + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\kappa_{x},\\kappa_{y}" + }, + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "text", + "content": " components of the illumination are sampled by the Nyquist pitch " + }, + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\Delta_{f} = 1 / 2\\Omega_{p}" + }, + { + "bbox": [ + 313, + 377, + 554, + 498 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 498, + 554, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 592 + ], + "type": "text", + "content": "With the above sampling scheme, the transmission matrices at the Fourier and primal bases are related by a simple Fourier transform " + }, + { + "bbox": [ + 313, + 498, + 554, + 592 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^p = \\mathcal{F}^T\\mathcal{T}^f\\mathcal{F}" + }, + { + "bbox": [ + 313, + 498, + 554, + 592 + ], + "type": "text", + "content": ". While for wavefront shaping we are interested in primal transmission matrices, the Fourier ones simplify analysis. Importantly, since Fourier and primal transmission matrices are related by an orthonormal transformation, the fitting errors in both bases are equivalent." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 592, + 554, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 554, + 691 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 554, + 691 + ], + "type": "text", + "content": "The multi-slice model: The relation between an incoming illumination and an outgoing one is determined by the solution to a wave equation which takes into account the variation in refractive index (RI) inside the sample. Rather than a fully continuous model, in a multi-slice model we approximate the volume with a set of planar aberrations, separated by small " + }, + { + "bbox": [ + 313, + 592, + 554, + 691 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 592, + 554, + 691 + ], + "type": "text", + "content": " distances " + }, + { + "bbox": [ + 313, + 592, + 554, + 691 + ], + "type": "inline_equation", + "content": "\\Delta_z" + }, + { + "bbox": [ + 313, + 592, + 554, + 691 + ], + "type": "text", + "content": ", where we have free space propagation between layers, as in Fig. 2(d). Mathematically this can be described as" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 322, + 695, + 553, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 695, + 553, + 717 + ], + "spans": [ + { + "bbox": [ + 322, + 695, + 553, + 717 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mathcal {T}} \\left(\\rho_ {1}, \\dots , \\rho_ {M}\\right) = \\boldsymbol {\\mathcal {P}} _ {O} \\boldsymbol {\\mathcal {D}} \\left(\\rho_ {M}\\right) \\dots \\boldsymbol {\\mathcal {P}} _ {\\Delta z} \\boldsymbol {\\mathcal {D}} \\left(\\rho_ {2}\\right) \\boldsymbol {\\mathcal {P}} _ {\\Delta z} \\boldsymbol {\\mathcal {D}} \\left(\\rho_ {1}\\right), \\tag {2}", + "image_path": "4a9ac7cae2a2332a6d195bbad8bc511073b6d65c022cdb2d76e196853b2f4d8a.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23166" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 93, + 70, + 149, + 210 + ], + "blocks": [ + { + "bbox": [ + 93, + 70, + 149, + 210 + ], + "lines": [ + { + "bbox": [ + 93, + 70, + 149, + 210 + ], + "spans": [ + { + "bbox": [ + 93, + 70, + 149, + 210 + ], + "type": "image", + "image_path": "ffe8fb84265c39b1272d097ee9dc1821b61d3afdaf274043f211763d73a1f15e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 212, + 149, + 220 + ], + "lines": [ + { + "bbox": [ + 94, + 212, + 149, + 220 + ], + "spans": [ + { + "bbox": [ + 94, + 212, + 149, + 220 + ], + "type": "text", + "content": "(a) Ground truth" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 156, + 79, + 212, + 211 + ], + "blocks": [ + { + "bbox": [ + 156, + 79, + 212, + 211 + ], + "lines": [ + { + "bbox": [ + 156, + 79, + 212, + 211 + ], + "spans": [ + { + "bbox": [ + 156, + 79, + 212, + 211 + ], + "type": "image", + "image_path": "95e6e867cb700e3562b6d2eb098aff8864b563efaea60941564317acad376783.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 212, + 207, + 220 + ], + "lines": [ + { + "bbox": [ + 164, + 212, + 207, + 220 + ], + "spans": [ + { + "bbox": [ + 164, + 212, + 207, + 220 + ], + "type": "text", + "content": "(b) 32 layers" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 217, + 79, + 273, + 211 + ], + "blocks": [ + { + "bbox": [ + 217, + 79, + 273, + 211 + ], + "lines": [ + { + "bbox": [ + 217, + 79, + 273, + 211 + ], + "spans": [ + { + "bbox": [ + 217, + 79, + 273, + 211 + ], + "type": "image", + "image_path": "2f43cc57c9d498714e26fb51adff727a8d6794186eab3ec66cc99c556b26905b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 212, + 264, + 220 + ], + "lines": [ + { + "bbox": [ + 227, + 212, + 264, + 220 + ], + "spans": [ + { + "bbox": [ + 227, + 212, + 264, + 220 + ], + "type": "text", + "content": "(c) 8 layers" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 228, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 228, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 555, + 285 + ], + "type": "text", + "content": "Figure 3. Reconstructed volumes: (a) An xz slice from a toy volume, visualizing its refractive index variations. (b-d) Reconstructions with varying layer numbers. The columns of the transmission matrix provide an under-constraint inversion problem, and even when a dense set of layers is allocated in b, the axial resolution of the reconstructed volume is low. As a result the sparse reconstructions in c,d can also well explain the transmission matrix. (e) Numerical evaluation of fitting error as a function of the number of layers. Transmission matrices can be well fitted even if the allocated number of layers is lower than Nyquist predictions. However the required number of layers increases with the tissue thickness. The layer number is less affected when the optical density is increased for the same tissue thickness." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 278, + 79, + 333, + 211 + ], + "blocks": [ + { + "bbox": [ + 278, + 79, + 333, + 211 + ], + "lines": [ + { + "bbox": [ + 278, + 79, + 333, + 211 + ], + "spans": [ + { + "bbox": [ + 278, + 79, + 333, + 211 + ], + "type": "image", + "image_path": "d1c7c44fb676b59ccce5726adf4f7d480990a8c40a71cf6fa4d29bab9b1ef627.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 286, + 212, + 325, + 220 + ], + "lines": [ + { + "bbox": [ + 286, + 212, + 325, + 220 + ], + "spans": [ + { + "bbox": [ + 286, + 212, + 325, + 220 + ], + "type": "text", + "content": "(d) 4 layers" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 334, + 79, + 515, + 209 + ], + "blocks": [ + { + "bbox": [ + 334, + 79, + 515, + 209 + ], + "lines": [ + { + "bbox": [ + 334, + 79, + 515, + 209 + ], + "spans": [ + { + "bbox": [ + 334, + 79, + 515, + 209 + ], + "type": "image", + "image_path": "eb9bb615bbe774bfbe881ece5da6a5612783a420231a2c520d5bcb03a87bb5a0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 211, + 468, + 220 + ], + "lines": [ + { + "bbox": [ + 408, + 211, + 468, + 220 + ], + "spans": [ + { + "bbox": [ + 408, + 211, + 468, + 220 + ], + "type": "text", + "content": "(e) Fitting curves" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 311, + 295, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 311, + 295, + 405 + ], + "spans": [ + { + "bbox": [ + 54, + 311, + 295, + 405 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 311, + 295, + 405 + ], + "type": "inline_equation", + "content": "\\rho_{m}(x,y)" + }, + { + "bbox": [ + 54, + 311, + 295, + 405 + ], + "type": "text", + "content": " is the 2D aberration of the " + }, + { + "bbox": [ + 54, + 311, + 295, + 405 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 54, + 311, + 295, + 405 + ], + "type": "text", + "content": "th layer and " + }, + { + "bbox": [ + 54, + 311, + 295, + 405 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\rho_m)" + }, + { + "bbox": [ + 54, + 311, + 295, + 405 + ], + "type": "text", + "content": " is a diagonal matrix performing element-wise multiplication of the incoming wavefront by the aberration. " + }, + { + "bbox": [ + 54, + 311, + 295, + 405 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\Delta_z}" + }, + { + "bbox": [ + 54, + 311, + 295, + 405 + ], + "type": "text", + "content": " describes free space propagation between layers at distance " + }, + { + "bbox": [ + 54, + 311, + 295, + 405 + ], + "type": "inline_equation", + "content": "\\Delta_z" + }, + { + "bbox": [ + 54, + 311, + 295, + 405 + ], + "type": "text", + "content": ", which is essentially a convolution with a spherical wavefront. Finally " + }, + { + "bbox": [ + 54, + 311, + 295, + 405 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_O" + }, + { + "bbox": [ + 54, + 311, + 295, + 405 + ], + "type": "text", + "content": " propagates the wavefront in free space from the tissue boundary, via the lenses, to the camera sensor." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 407, + 295, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 407, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 53, + 407, + 295, + 442 + ], + "type": "text", + "content": "Our goal is to understand what plane spacing " + }, + { + "bbox": [ + 53, + 407, + 295, + 442 + ], + "type": "inline_equation", + "content": "\\Delta_z" + }, + { + "bbox": [ + 53, + 407, + 295, + 442 + ], + "type": "text", + "content": " will facilitate a good representation of " + }, + { + "bbox": [ + 53, + 407, + 295, + 442 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 53, + 407, + 295, + 442 + ], + "type": "text", + "content": ", and how this approximation would degrade when we use sparser sampling." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 443, + 271, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 443, + 271, + 453 + ], + "spans": [ + { + "bbox": [ + 67, + 443, + 271, + 453 + ], + "type": "text", + "content": "A common approximation for weak aberrations is:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 121, + 453, + 294, + 465 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 453, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 121, + 453, + 294, + 465 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\rho} _ {m} (x, y) = \\mu + \\delta \\boldsymbol {\\rho} _ {m} (x, y), \\tag {3}", + "image_path": "5d2a664593e0364bbbeda842b519470500518f5ceafe6cf9970f34be68fd59c2.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 469, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 295, + 568 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 469, + 295, + 568 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 469, + 295, + 568 + ], + "type": "text", + "content": " is a scalar whose magnitude is a bit smaller than 1, encoding the ballistic portion of the light which passes through the layer without aberration, and the magnitude of the aberration is significantly smaller than the ballistic term " + }, + { + "bbox": [ + 55, + 469, + 295, + 568 + ], + "type": "inline_equation", + "content": "|\\delta \\rho_{m}(x,y)|\\ll |\\mu |" + }, + { + "bbox": [ + 55, + 469, + 295, + 568 + ], + "type": "text", + "content": " . For weak scattering most of the energy does not scatter more than once. It is therefore common to consider a first order approximation to the transmission matrix (often known as the single scattering, first-Born or Rytrov approximation [9, 10, 13, 29, 33]), which reduces to" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 53, + 570, + 295, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 570, + 295, + 583 + ], + "spans": [ + { + "bbox": [ + 53, + 570, + 295, + 583 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mathcal {T}} \\left(\\delta \\rho_ {1}, \\dots , \\delta \\rho_ {M}\\right) \\approx \\mu^ {M} \\boldsymbol {\\mathcal {P}} _ {O} \\boldsymbol {\\mathcal {P}} _ {M \\Delta_ {z}} + \\tag {4}", + "image_path": "923d9b2e658850a780e85691cbe93c939161ad9ad9391b7dcf1a7239b5fd00de.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 132, + 585, + 294, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 585, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 132, + 585, + 294, + 613 + ], + "type": "interline_equation", + "content": "\\mu^ {M - 1} \\sum_ {m = 1} ^ {M} \\mathcal {P} _ {O} \\mathcal {P} _ {(M - m) \\Delta_ {z}} \\mathcal {D} (\\delta \\rho_ {m}) \\mathcal {P} _ {m \\Delta_ {z}}", + "image_path": "97a6fa4085795dbcecba9a04a2f98e50920c13f1ea981a05442f802b516514e5.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 617, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 689 + ], + "type": "text", + "content": "The full light propagation in Eq. (2) includes multiplication of the different aberrations and hence the transmission matrix is highly nonlinear. In contrast, the advantage of the weakly scattering model in Eq. (4) is that the resulting transmission matrix is a linear function of the aberration layers. This linear approximation significantly simplifies analysis." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "text", + "content": "Diffraction tomography algorithms such as [11] capture a subset of columns from the transmission matrix and use" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 311, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 555, + 407 + ], + "type": "text", + "content": "gradient descent optimization to seek layers " + }, + { + "bbox": [ + 313, + 311, + 555, + 407 + ], + "type": "inline_equation", + "content": "\\rho_1, \\ldots, \\rho_M" + }, + { + "bbox": [ + 313, + 311, + 555, + 407 + ], + "type": "text", + "content": " which best explain the measured wavefronts, assuming that the propagation operators " + }, + { + "bbox": [ + 313, + 311, + 555, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\Delta_z}, \\mathcal{P}_O" + }, + { + "bbox": [ + 313, + 311, + 555, + 407 + ], + "type": "text", + "content": " are known. We will use a similar optimization scheme below, when trying to find a multi-slice representation for a given transmission matrix. To correct aberrations, a physical multi-conjugate wavefront shaping system would display on the SLM the conjugate of the recovered layers." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 418, + 524, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 524, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 524, + 432 + ], + "type": "text", + "content": "3. Evaluating multi-slice approximations" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "text", + "content": "We start with numerical experiments demonstrating the main findings. In Sec. 4 below we derive a theory explaining them. To gain intuition we generated a toy volume composed of spheres with varying refractive indices illustrated in Fig. 3(a), and obtained the corresponding transmission matrix by propagating a set of point sources through the volume. We then used the transmission matrix to solve for a multi-slice approximation, while varying the number of layers. In Fig. 3(b-d) we visualize some of the volumetric reconstructions, assuming a " + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\lambda = 500nm" + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "text", + "content": " illumination. Even when the fitted layers are sampled very densely at an interval of " + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\Delta_z = 0.5\\mu m" + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "text", + "content": ", we cannot fully recover the target and the fitted volume has a low axial resolution. We analyze this in Sec. 4 below and explain that this results from the well known missing cone problem, a fundamental degeneracy limiting the axial resolution of a microscope. Said differently the transmission matrix, or the full set of propagating waves, only provides a set of under-constraint measurements from which the volume cannot be uniquely recovered. While the missing cone problem is considered a fundamental limitation in microscopy, we argue that in the context of wavefront shaping it actually turns into an advantage. Since the volume reconstruction is not unique we can" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23167" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 74, + 121, + 210 + ], + "blocks": [ + { + "bbox": [ + 65, + 74, + 121, + 210 + ], + "lines": [ + { + "bbox": [ + 65, + 74, + 121, + 210 + ], + "spans": [ + { + "bbox": [ + 65, + 74, + 121, + 210 + ], + "type": "image", + "image_path": "c030826722d17eab87af7556ccee4f64461f35318021127ea39c15c82e996dde.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 211, + 118, + 219 + ], + "lines": [ + { + "bbox": [ + 67, + 211, + 118, + 219 + ], + "spans": [ + { + "bbox": [ + 67, + 211, + 118, + 219 + ], + "type": "text", + "content": "(a) Ground truth" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 130, + 76, + 186, + 210 + ], + "blocks": [ + { + "bbox": [ + 130, + 76, + 186, + 210 + ], + "lines": [ + { + "bbox": [ + 130, + 76, + 186, + 210 + ], + "spans": [ + { + "bbox": [ + 130, + 76, + 186, + 210 + ], + "type": "image", + "image_path": "44e481b3b0d8ba9707b4df26d4dec5884e8dcb09e2c991e70c435b772a81296e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 211, + 171, + 219 + ], + "lines": [ + { + "bbox": [ + 131, + 211, + 171, + 219 + ], + "spans": [ + { + "bbox": [ + 131, + 211, + 171, + 219 + ], + "type": "text", + "content": "(b) Dense fit" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 187, + 76, + 242, + 210 + ], + "blocks": [ + { + "bbox": [ + 187, + 76, + 242, + 210 + ], + "lines": [ + { + "bbox": [ + 187, + 76, + 242, + 210 + ], + "spans": [ + { + "bbox": [ + 187, + 76, + 242, + 210 + ], + "type": "image", + "image_path": "bcfef151987ea8ba01c9b03cba4f97f51e36b7536bf389e24341ff5ee1f529dc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 188, + 212, + 228, + 220 + ], + "lines": [ + { + "bbox": [ + 188, + 212, + 228, + 220 + ], + "spans": [ + { + "bbox": [ + 188, + 212, + 228, + 220 + ], + "type": "text", + "content": "(c) Sparse fit" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 242, + 76, + 306, + 210 + ], + "blocks": [ + { + "bbox": [ + 242, + 76, + 306, + 210 + ], + "lines": [ + { + "bbox": [ + 242, + 76, + 306, + 210 + ], + "spans": [ + { + "bbox": [ + 242, + 76, + 306, + 210 + ], + "type": "image", + "image_path": "89bb3c593d97c43e54b7e49fec9b357151b35b0b8b8387c56fc621c6217359b3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 212, + 291, + 220 + ], + "lines": [ + { + "bbox": [ + 252, + 212, + 291, + 220 + ], + "spans": [ + { + "bbox": [ + 252, + 212, + 291, + 220 + ], + "type": "text", + "content": "(d) Dense fit" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 228, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 54, + 228, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 54, + 228, + 555, + 277 + ], + "type": "text", + "content": "Figure 4. Reconstructed volumes with varying supports: (a) An xz slice from a toy volume, visualizing its refractive index variations. (b,d) Two reconstructions of the volume using a dense layer sampling. When the support of the measured window " + }, + { + "bbox": [ + 54, + 228, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\Omega_{p}" + }, + { + "bbox": [ + 54, + 228, + 555, + 277 + ], + "type": "text", + "content": " shrinks, the lateral extent of the reconstructed region decreases, however, its axial resolution is also reduced. (c,e) A sparse fit with the same input. Despite the low quality of these reconstructions, they predict well the columns of the transmission matrix of the given support " + }, + { + "bbox": [ + 54, + 228, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\Omega_{p}" + }, + { + "bbox": [ + 54, + 228, + 555, + 277 + ], + "type": "text", + "content": ". (f) Numerical evaluation of fitting error as a function of the number of layers, while varying the range " + }, + { + "bbox": [ + 54, + 228, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\Omega_{p}" + }, + { + "bbox": [ + 54, + 228, + 555, + 277 + ], + "type": "text", + "content": " covered by the transmission matrix. Smaller supports can be fitted with fewer layers." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 76, + 362, + 210 + ], + "blocks": [ + { + "bbox": [ + 306, + 76, + 362, + 210 + ], + "lines": [ + { + "bbox": [ + 306, + 76, + 362, + 210 + ], + "spans": [ + { + "bbox": [ + 306, + 76, + 362, + 210 + ], + "type": "image", + "image_path": "a9eab028a53d42a94619d402f561f77c79d549e911fd9c9e2140b5007b22c571.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 211, + 349, + 220 + ], + "lines": [ + { + "bbox": [ + 309, + 211, + 349, + 220 + ], + "spans": [ + { + "bbox": [ + 309, + 211, + 349, + 220 + ], + "type": "text", + "content": "(e) Sparse fit" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 370, + 82, + 545, + 209 + ], + "blocks": [ + { + "bbox": [ + 370, + 82, + 545, + 209 + ], + "lines": [ + { + "bbox": [ + 370, + 82, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 370, + 82, + 545, + 209 + ], + "type": "image", + "image_path": "d667e5f775f637a56c3bd675225f8d9c91df8e3aaf10a06697fafb6262f87970.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 211, + 494, + 220 + ], + "lines": [ + { + "bbox": [ + 441, + 211, + 494, + 220 + ], + "spans": [ + { + "bbox": [ + 441, + 211, + 494, + 220 + ], + "type": "text", + "content": "(f) Fitting curves" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 301, + 295, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 301, + 295, + 324 + ], + "spans": [ + { + "bbox": [ + 54, + 301, + 295, + 324 + ], + "type": "text", + "content": "approximate the same transmission matrix with a sparser set of layers." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 324, + 295, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 295, + 357 + ], + "type": "text", + "content": "To quantify the fitting error of the multi-slice model, we denote by " + }, + { + "bbox": [ + 55, + 324, + 295, + 357 + ], + "type": "inline_equation", + "content": "\\pmb{\\varepsilon}_{M}" + }, + { + "bbox": [ + 55, + 324, + 295, + 357 + ], + "type": "text", + "content": " the reconstruction error of a given transmission matrix with the best " + }, + { + "bbox": [ + 55, + 324, + 295, + 357 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 324, + 295, + 357 + ], + "type": "text", + "content": " layers" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 357, + 295, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 357, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 89, + 357, + 295, + 376 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mathcal {E}} _ {M} = \\min _ {\\rho_ {1}, \\dots , \\rho_ {M}} \\| \\boldsymbol {\\mathcal {T}} _ {\\text {e x a c t}} - \\boldsymbol {\\mathcal {T}} (\\rho_ {1}, \\dots , \\rho_ {M}) \\| ^ {2}. \\tag {5}", + "image_path": "86d08bdf983606e9682c7a547370dfdedc9c07cfe3f9de223676db56a26eb217.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 54, + 378, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 378, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 54, + 378, + 295, + 544 + ], + "type": "text", + "content": "In Fig. 3(e) we evaluate " + }, + { + "bbox": [ + 54, + 378, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\pmb{\\mathcal{E}}_M" + }, + { + "bbox": [ + 54, + 378, + 295, + 544 + ], + "type": "text", + "content": " as a function of the layer number " + }, + { + "bbox": [ + 54, + 378, + 295, + 544 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 54, + 378, + 295, + 544 + ], + "type": "text", + "content": ". Even the seemingly aliased fit of Fig. 3(d) still provides a low fitting error. However, there are no free lunches and when the number of layers is too low the fitting error is high. When we repeat the simulation with a thicker volume we need even more layers for a good fit. Interestingly, we also repeat the simulation with denser volumes (larger optical depth, or shorter mean free path) of the same thickness, and this has less effect on the required number of layers. In our plots, the case " + }, + { + "bbox": [ + 54, + 378, + 295, + 544 + ], + "type": "inline_equation", + "content": "M = 0" + }, + { + "bbox": [ + 54, + 378, + 295, + 544 + ], + "type": "text", + "content": " refers to no correction at all, so effectively we assume that the transmission matrix is diagonal. The similarity between a diagonal transmission matrix and the target transmission matrix is a measure of the amount of ballistic light." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 54, + 545, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 545, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 54, + 545, + 295, + 676 + ], + "type": "text", + "content": "The thickness of the synthetic simulation here is limited due to computational constraints. However, usually wavefront shaping systems attempt to see into a few hundred microns of tissue and we anticipate that for such volumes we need at least a few dozen layers for a good fit. Clearly realizing a physical wavefront shaping system with that many layers is impractical. We argue that while a small number of layers is not enough for focusing light anywhere across the tissue area, it can still help us to focus light over a compact support area, and this area is significantly wider than what we can correct with a single layer." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Compact support: To test the effect of the target area, we consider transmission matrices measured with illumination points confined to a limited area of size " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\Omega_p \\times \\Omega_p" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": ". If we can" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 301, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 553, + 361 + ], + "type": "text", + "content": "fit such a transmission matrix with " + }, + { + "bbox": [ + 313, + 301, + 553, + 361 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 301, + 553, + 361 + ], + "type": "text", + "content": " layers and place the values of these " + }, + { + "bbox": [ + 313, + 301, + 553, + 361 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 301, + 553, + 361 + ], + "type": "text", + "content": " layers on " + }, + { + "bbox": [ + 313, + 301, + 553, + 361 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 301, + 553, + 361 + ], + "type": "text", + "content": " SLMs, the resulting wavefront shaping system can correct light from the " + }, + { + "bbox": [ + 313, + 301, + 553, + 361 + ], + "type": "inline_equation", + "content": "\\Omega_p \\times \\Omega_p" + }, + { + "bbox": [ + 313, + 301, + 553, + 361 + ], + "type": "text", + "content": " area, but target points outside this area are unlikely to be corrected as they are not included in the fitting process." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 363, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 567 + ], + "type": "text", + "content": "To test the effect of support on the number of fitted layers, we return to the numerical simulation of Fig. 3, but fit subsets of columns contained in smaller " + }, + { + "bbox": [ + 313, + 363, + 555, + 567 + ], + "type": "inline_equation", + "content": "\\Omega_p" + }, + { + "bbox": [ + 313, + 363, + 555, + 567 + ], + "type": "text", + "content": " ranges. Indeed, smaller supports can be well-fitted with a smaller number of layers. As another way to understand it, we show in Fig. 4 an xz slice from the ground truth refractive index volume and from a few reconstructions. We show that even if we do not limit the layer number " + }, + { + "bbox": [ + 313, + 363, + 555, + 567 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 363, + 555, + 567 + ], + "type": "text", + "content": ", the recovered RI volume has limited axial resolution and this axial resolution is lower when measured range " + }, + { + "bbox": [ + 313, + 363, + 555, + 567 + ], + "type": "inline_equation", + "content": "\\Omega_p" + }, + { + "bbox": [ + 313, + 363, + 555, + 567 + ], + "type": "text", + "content": " decreases. This explains why lower supports can be fitted with fewer layers. For each of the two supports we also show a sparse fit, with the minimal " + }, + { + "bbox": [ + 313, + 363, + 555, + 567 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 363, + 555, + 567 + ], + "type": "text", + "content": " value that provides a good prediction of the transmission matrix. We explain this observation in Sec. 4. This result is also in agreement with [26], who correct aberration over a relatively small field of view through a few hundred microns of tissue, using only 5 layers." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "Implications to layered wavefront shaping systems. In Fig. 4(f) we plotted a bar below which the approximation is reasonable. We can see that for e.g. the " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "40 \\times 40\\mu m" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " support we can obtain good focusing with about " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "M = 6" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " layers. It is likely that fitting a full field-of-view of a few hundred microns would require more layers. However, we can see that while with " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "M = 1" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " layer we could only focus over a small area of about " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "1 \\times 1\\mu m" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": ", when increasing to " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "M = 2" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " we could focus over a " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " larger area of support " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "2 \\times 2\\mu m" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": ", and with " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "M = 3" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " we could focus over a " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\times 25" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " larger area of support " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "5 \\times 5\\mu m" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": ". Computing multiple correction elements is only going to increase the complexity of the wavefront shaping" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23168" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 71, + 250, + 204 + ], + "blocks": [ + { + "bbox": [ + 71, + 71, + 250, + 204 + ], + "lines": [ + { + "bbox": [ + 71, + 71, + 250, + 204 + ], + "spans": [ + { + "bbox": [ + 71, + 71, + 250, + 204 + ], + "type": "image", + "image_path": "a619d8c80fbac286a52416d0615f50d448ffa7ed9fbf2a4570d53d4dff932fdb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 262, + 71, + 541, + 219 + ], + "blocks": [ + { + "bbox": [ + 262, + 71, + 541, + 219 + ], + "lines": [ + { + "bbox": [ + 262, + 71, + 541, + 219 + ], + "spans": [ + { + "bbox": [ + 262, + 71, + 541, + 219 + ], + "type": "image", + "image_path": "3ef80ea7a846532437c145aa0d761356e765cebe99d085ea8dc6bb3863484d79.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 222, + 555, + 274 + ], + "lines": [ + { + "bbox": [ + 54, + 222, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 54, + 222, + 555, + 274 + ], + "type": "text", + "content": "Figure 5. Layer support for a captured transmission matrix: We consider a transmission matrix captured in the lab through a chicken breast layer of thickness " + }, + { + "bbox": [ + 54, + 222, + 555, + 274 + ], + "type": "inline_equation", + "content": "170\\mu m" + }, + { + "bbox": [ + 54, + 222, + 555, + 274 + ], + "type": "text", + "content": ". Left: plotting the fitting error as a function of the number of layers. Smaller supports can be fitted with fewer layers. Right: Visualization of a spot focusing behind the tissue computed using the wavefronts of the fitted model. We compare models fitted to different supports. In the top row we need a larger number of layers, but the spot can be scanned over a " + }, + { + "bbox": [ + 54, + 222, + 555, + 274 + ], + "type": "inline_equation", + "content": "25 \\times 25\\mu m" + }, + { + "bbox": [ + 54, + 222, + 555, + 274 + ], + "type": "text", + "content": " window using the same layers. In the lower rows we achieve a focused spot using a smaller number of layers, but these layers can scan the focused spot over smaller windows of sizes " + }, + { + "bbox": [ + 54, + 222, + 555, + 274 + ], + "type": "inline_equation", + "content": "6 \\times 6\\mu m" + }, + { + "bbox": [ + 54, + 222, + 555, + 274 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 222, + 555, + 274 + ], + "type": "inline_equation", + "content": "1 \\times 1\\mu m" + }, + { + "bbox": [ + 54, + 222, + 555, + 274 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 293, + 294, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 293, + 294, + 413 + ], + "spans": [ + { + "bbox": [ + 54, + 293, + 294, + 413 + ], + "type": "text", + "content": "algorithm in [1] linearly with the number of layers. Alternatively, a wavefront shaping system using a single SLM would need to scan the field-of-view sequentially and correct each local window independently. Since the increase in the support of the window we can explain with e.g. 3 layers is significantly larger than a " + }, + { + "bbox": [ + 54, + 293, + 294, + 413 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 54, + 293, + 294, + 413 + ], + "type": "text", + "content": " factor, a multi-conjugate wavefront shaping system with a small number of SLMs as in [16, 27, 31, 35] can already have a large practical benefit. While Fig. 4 only considers a simulation, we analyze realistic numbers below." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 415, + 295, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 450 + ], + "type": "text", + "content": "Lab results: In the supplementary file we validate the theory with multiple real transmission matrices, representing thick volumes with multiple-scattering." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 450, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 521 + ], + "type": "text", + "content": "We start with numerically simulated transmission matrices using a more accurate wave-propagation model [6, 7]. We then test lab-captured transmission matrices measuring realistic scattering samples including mouse brain tissue, chicken breast tissue and a parafilm layer. Please refer to supplementary file for additional results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 522, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 582 + ], + "type": "text", + "content": "In Fig. 5 we show one fitting result using a multiply-scattering chicken breast layer of thickness " + }, + { + "bbox": [ + 55, + 522, + 295, + 582 + ], + "type": "inline_equation", + "content": "d = 170\\mu m" + }, + { + "bbox": [ + 55, + 522, + 295, + 582 + ], + "type": "text", + "content": ". The transmission matrices cover an area with support " + }, + { + "bbox": [ + 55, + 522, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\Omega_{p} \\times \\Omega_{p} = 13 \\times 13\\mu m" + }, + { + "bbox": [ + 55, + 522, + 295, + 582 + ], + "type": "text", + "content": ". As in Fig. 4(f), we see that smaller spatial ranges can be fitted with fewer layers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 701 + ], + "type": "text", + "content": "As another way to assess the quality of the fit, we used columns of the fitted matrices to focus light through the tissue and show the shape of the focused spot and visualize the focusing quality in Fig. 5. To simulate the focused spot we simply visualize the product " + }, + { + "bbox": [ + 55, + 582, + 295, + 701 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{exact}} \\cdot \\mathbf{t}_{\\mathrm{fit,M}}^k" + }, + { + "bbox": [ + 55, + 582, + 295, + 701 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 582, + 295, + 701 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{\\mathrm{fit,M}}^k" + }, + { + "bbox": [ + 55, + 582, + 295, + 701 + ], + "type": "text", + "content": " is a column of the fitted matrix " + }, + { + "bbox": [ + 55, + 582, + 295, + 701 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\rho_1, \\ldots, \\rho_M)" + }, + { + "bbox": [ + 55, + 582, + 295, + 701 + ], + "type": "text", + "content": ", namely a wavefront generated by the layered approximation. Note that while we only show focusing at one point, the layers are optimized such that they allow us to focus at any point inside the " + }, + { + "bbox": [ + 55, + 582, + 295, + 701 + ], + "type": "inline_equation", + "content": "\\Omega_p \\times \\Omega_p" + }, + { + "bbox": [ + 55, + 582, + 295, + 701 + ], + "type": "text", + "content": " window." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "type": "text", + "content": "From the results in Fig. 5 we can see that for the " + }, + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "type": "inline_equation", + "content": "13\\times" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 293, + 555, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 390 + ], + "type": "inline_equation", + "content": "13\\mu m" + }, + { + "bbox": [ + 313, + 293, + 555, + 390 + ], + "type": "text", + "content": " support measured, we can obtain good focusing with about 3 layers. Fitting a full field-of-view of several hundred microns likely requires additional layers. However, we observe that while using " + }, + { + "bbox": [ + 313, + 293, + 555, + 390 + ], + "type": "inline_equation", + "content": "M = 1" + }, + { + "bbox": [ + 313, + 293, + 555, + 390 + ], + "type": "text", + "content": " allows us to focus on a small area of about " + }, + { + "bbox": [ + 313, + 293, + 555, + 390 + ], + "type": "inline_equation", + "content": "3\\times 3\\mu m" + }, + { + "bbox": [ + 313, + 293, + 555, + 390 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 293, + 555, + 390 + ], + "type": "inline_equation", + "content": "M = 3" + }, + { + "bbox": [ + 313, + 293, + 555, + 390 + ], + "type": "text", + "content": ", we can focus on an area " + }, + { + "bbox": [ + 313, + 293, + 555, + 390 + ], + "type": "inline_equation", + "content": "\\times 18" + }, + { + "bbox": [ + 313, + 293, + 555, + 390 + ], + "type": "text", + "content": " larger, reaching " + }, + { + "bbox": [ + 313, + 293, + 555, + 390 + ], + "type": "inline_equation", + "content": "13\\times 13\\mu m" + }, + { + "bbox": [ + 313, + 293, + 555, + 390 + ], + "type": "text", + "content": ". Therefore, using multiple layers can significantly accelerate a sequential scanning of a wave-front shaping system." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 398, + 547, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 547, + 411 + ], + "type": "text", + "content": "4. Sampling theory for transmission matrices" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 415, + 554, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 451 + ], + "type": "text", + "content": "Our goal in this section is to explain the empirical findings of the previous section by analyzing the information captured by a transmission matrix." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 459, + 518, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 518, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 518, + 472 + ], + "type": "text", + "content": "4.1. Nyquist sampling and the missing cone" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 473, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 554, + 544 + ], + "type": "text", + "content": "We start by considering the case of weakly scattering volumes which, as in Eq. (4), can be reasonably approximated by considering the component of light that scattered once. We also assume the transmission matrix is sampled over a very wide support " + }, + { + "bbox": [ + 313, + 473, + 554, + 544 + ], + "type": "inline_equation", + "content": "\\Omega_{p}" + }, + { + "bbox": [ + 313, + 473, + 554, + 544 + ], + "type": "text", + "content": ". We discuss the implications of multiple scattering and limited supports in subsequent sections." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 545, + 553, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 553, + 581 + ], + "type": "text", + "content": "We denote by " + }, + { + "bbox": [ + 313, + 545, + 553, + 581 + ], + "type": "inline_equation", + "content": "\\pmb{n}(\\mathbf{r})" + }, + { + "bbox": [ + 313, + 545, + 553, + 581 + ], + "type": "text", + "content": " the difference between the RI at a 3D position " + }, + { + "bbox": [ + 313, + 545, + 553, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 313, + 545, + 553, + 581 + ], + "type": "text", + "content": " in the volume to that of the leading medium around it. We denote its 3D Fourier transform by " + }, + { + "bbox": [ + 313, + 545, + 553, + 581 + ], + "type": "inline_equation", + "content": "\\hat{n}(\\kappa)" + }, + { + "bbox": [ + 313, + 545, + 553, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 581, + 554, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 554, + 658 + ], + "type": "text", + "content": "We start by considering the Fourier representation of the transmission matrix, where it is illuminated and measured by plane waves. Using the weakly scattering (first-Born) approximation, the wavefront scattering toward direction " + }, + { + "bbox": [ + 313, + 581, + 554, + 658 + ], + "type": "inline_equation", + "content": "\\bar{\\omega}^v" + }, + { + "bbox": [ + 313, + 581, + 554, + 658 + ], + "type": "text", + "content": " when illuminated by an incoming plane wave at direction " + }, + { + "bbox": [ + 313, + 581, + 554, + 658 + ], + "type": "inline_equation", + "content": "\\bar{\\omega}^i" + }, + { + "bbox": [ + 313, + 581, + 554, + 658 + ], + "type": "text", + "content": " is proportional, up to a multiplicative factor, to [28, 30, 38]:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 383, + 662, + 553, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 662, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 383, + 662, + 553, + 685 + ], + "type": "interline_equation", + "content": "\\propto \\int \\boldsymbol {n} (\\mathbf {r}) e ^ {2 \\pi i \\left(\\bar {\\kappa} ^ {i} - \\bar {\\kappa} ^ {v}\\right) \\cdot \\mathbf {r}} d \\mathbf {r}, \\tag {6}", + "image_path": "d55f69c18dadcae85a9f074b3d18d27ce43f3ace2a75c3508589fb89ce95604a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\bar{\\kappa}^v = 1 / \\lambda \\bar{\\omega}^v" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\bar{\\kappa}^i = 1 / \\lambda \\bar{\\omega}^i" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": ". This implies that effectively an entry of the transmission matrix is a sample" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23169" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 134, + 145 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 134, + 145 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 134, + 145 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 134, + 145 + ], + "type": "image", + "image_path": "46bfd089b3b35d5bc046dad013056388f346bc9e16d249fa49982443b4d7092f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 90, + 149, + 102, + 159 + ], + "lines": [ + { + "bbox": [ + 90, + 149, + 102, + 159 + ], + "spans": [ + { + "bbox": [ + 90, + 149, + 102, + 159 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 137, + 70, + 216, + 147 + ], + "blocks": [ + { + "bbox": [ + 137, + 70, + 216, + 147 + ], + "lines": [ + { + "bbox": [ + 137, + 70, + 216, + 147 + ], + "spans": [ + { + "bbox": [ + 137, + 70, + 216, + 147 + ], + "type": "image", + "image_path": "0a5fc2e4f939bdd085a233905ed2d980356a80886f347bad34b87a953175c55f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 169, + 149, + 182, + 159 + ], + "lines": [ + { + "bbox": [ + 169, + 149, + 182, + 159 + ], + "spans": [ + { + "bbox": [ + 169, + 149, + 182, + 159 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 219, + 71, + 292, + 145 + ], + "blocks": [ + { + "bbox": [ + 219, + 71, + 292, + 145 + ], + "lines": [ + { + "bbox": [ + 219, + 71, + 292, + 145 + ], + "spans": [ + { + "bbox": [ + 219, + 71, + 292, + 145 + ], + "type": "image", + "image_path": "bd09b39ba8f1ab55a3632705decb26df4068d7436e42bebbaa667ed447d32ce1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 149, + 261, + 159 + ], + "lines": [ + { + "bbox": [ + 249, + 149, + 261, + 159 + ], + "spans": [ + { + "bbox": [ + 249, + 149, + 261, + 159 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 165, + 296, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 165, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 165, + 296, + 251 + ], + "type": "text", + "content": "Figure 6. Spectrum structure: (a) An " + }, + { + "bbox": [ + 55, + 165, + 296, + 251 + ], + "type": "inline_equation", + "content": "xz" + }, + { + "bbox": [ + 55, + 165, + 296, + 251 + ], + "type": "text", + "content": " slice out of the spectrum of " + }, + { + "bbox": [ + 55, + 165, + 296, + 251 + ], + "type": "inline_equation", + "content": "\\hat{n} (\\omega)" + }, + { + "bbox": [ + 55, + 165, + 296, + 251 + ], + "type": "text", + "content": ". Entries of a weakly-scattering transmission matrix limited by an aperture NA only lie inside the butterfly area. (b) Zooming on the center right area of (a) (purple square). Assuming the " + }, + { + "bbox": [ + 55, + 165, + 296, + 251 + ], + "type": "inline_equation", + "content": "\\omega_{z}" + }, + { + "bbox": [ + 55, + 165, + 296, + 251 + ], + "type": "text", + "content": " axis is cut at " + }, + { + "bbox": [ + 55, + 165, + 296, + 251 + ], + "type": "inline_equation", + "content": "\\pm \\frac{1}{2}\\Lambda^{M}" + }, + { + "bbox": [ + 55, + 165, + 296, + 251 + ], + "type": "text", + "content": ", the transmission matrix entries inside the dashed blue area are maintained, and the entries in the dashed red area are lost. (c) The energy at a slice from the spectrum of the volume in Fig. 3(a). Since the volume structure is smooth, most energy is concentrated at the low frequencies. As a result the cropped red region in (b) do not contain a lot of energy." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 278, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 296, + 434 + ], + "type": "text", + "content": "from the 3D Fourier transform of the volume at frequency " + }, + { + "bbox": [ + 55, + 278, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\kappa = \\bar{\\kappa}^i -\\bar{\\kappa}^v" + }, + { + "bbox": [ + 55, + 278, + 296, + 434 + ], + "type": "text", + "content": ". Note that throughout this paper we use bars " + }, + { + "bbox": [ + 55, + 278, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\bar{\\omega}" + }, + { + "bbox": [ + 55, + 278, + 296, + 434 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 278, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\bar{\\kappa}" + }, + { + "bbox": [ + 55, + 278, + 296, + 434 + ], + "type": "text", + "content": " to denote vectors with norm constraints, and use normal bold fonts to denote standard 3D vectors, such as frequencies " + }, + { + "bbox": [ + 55, + 278, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 55, + 278, + 296, + 434 + ], + "type": "text", + "content": " in the 3D spectrum. Unfortunately, a large portion of the 3D spectrum cannot be measured. Recalling that the illumination and viewing directions are limited by an aperture of width NA, the subset of 3D frequencies captured by the transmission matrix occupies only a butterfly shape, and the rest of the cone is missing, see illustration in Fig. 6(a). Using paraxial approximation it can be shown [28, 38] that the range of frequencies which can be generated is" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 441, + 296, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 296, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 296, + 473 + ], + "type": "interline_equation", + "content": "\\left| \\kappa_ {z} \\right| \\leq \\mathbf {N A} \\left| \\boldsymbol {\\kappa} _ {x y} \\right| - \\frac {\\lambda}{2} \\left| \\boldsymbol {\\kappa} _ {x y} \\right| ^ {2}, \\quad \\text {w i t h} \\quad \\left| \\boldsymbol {\\kappa} _ {x y} \\right| = \\sqrt {\\kappa_ {x} ^ {2} + \\kappa_ {y} ^ {2}}. \\tag {7}", + "image_path": "7878e47905679e47c869b1377ec87d6e9a7f36fc6fe0ae8f239784614031d7ab.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 473, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 296, + 579 + ], + "type": "text", + "content": "The missing cone is a serious problem for diffraction tomography algorithms aiming to recover the 3D structure of the refractive index " + }, + { + "bbox": [ + 55, + 473, + 296, + 579 + ], + "type": "inline_equation", + "content": "n(\\mathbf{r})" + }, + { + "bbox": [ + 55, + 473, + 296, + 579 + ], + "type": "text", + "content": " from columns of the transmission matrix, since many of the frequencies of the volume cannot be recovered. We argue that in the context of transmission matrices, the missing cone problem turns into an advantage. The fact that the transmission matrix only captures a limited range of frequencies allows us to approximate it with fewer samples, as derived below." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 581, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 296, + 715 + ], + "type": "text", + "content": "Minimal sampling: Our goal is to approximate a transmission matrix with layers. We start by asking what spacing between layers will allow an exact reconstruction of the transmission matrix. If we approximate the volume " + }, + { + "bbox": [ + 55, + 581, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{n}(\\mathbf{r})" + }, + { + "bbox": [ + 55, + 581, + 296, + 715 + ], + "type": "text", + "content": " with a set of planes, we essentially sample a continuous signal. Classical Nyquist theory states that if we sample a signal at intervals of " + }, + { + "bbox": [ + 55, + 581, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\Delta_z" + }, + { + "bbox": [ + 55, + 581, + 296, + 715 + ], + "type": "text", + "content": ", its Fourier transform includes replicas spaced by " + }, + { + "bbox": [ + 55, + 581, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\Lambda = 1 / \\Delta_z" + }, + { + "bbox": [ + 55, + 581, + 296, + 715 + ], + "type": "text", + "content": ". To achieve a good reconstruction, we want replicas to happen only in areas of the spectrum that the transmission matrix does not sample. Since the " + }, + { + "bbox": [ + 55, + 581, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\kappa_x, \\kappa_y" + }, + { + "bbox": [ + 55, + 581, + 296, + 715 + ], + "type": "text", + "content": " components of our illuminations are limited by the aperture " + }, + { + "bbox": [ + 55, + 581, + 296, + 715 + ], + "type": "inline_equation", + "content": "|\\kappa_{xy}| \\leq \\Omega_f" + }, + { + "bbox": [ + 55, + 581, + 296, + 715 + ], + "type": "text", + "content": ", and given the butterfly shape in Eq. (7), the " + }, + { + "bbox": [ + 55, + 581, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\kappa_z" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 508, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 508, + 83 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 508, + 83 + ], + "type": "text", + "content": "frequencies we need to maintain are bounded to" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 403, + 86, + 553, + 108 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 86, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 403, + 86, + 553, + 108 + ], + "type": "interline_equation", + "content": "\\left| \\kappa_ {z} \\right| \\leq \\frac {\\mathrm {N A} \\Omega_ {f}}{2}. \\tag {8}", + "image_path": "ff20ac4f3728a942c666ed2435965ab99e7e5144e666b8153e3a2693b902e247.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 110, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 553, + 133 + ], + "type": "text", + "content": "Therefore, for an error-free reconstruction, the minimal Fourier range and primal spacing are" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 356, + 133, + 553, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 133, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 356, + 133, + 553, + 156 + ], + "type": "interline_equation", + "content": "\\Lambda^ {*} = \\mathbf {N A} \\boldsymbol {\\Omega} _ {f}, \\quad \\Delta_ {z} ^ {*} = \\frac {1}{\\mathbf {N A} \\boldsymbol {\\Omega} _ {f}} = \\frac {\\lambda}{\\mathbf {N A} ^ {2}}. \\tag {9}", + "image_path": "217287c9fd3f2915ee3c97ca52efe7743ed1bf50bee160d3ce530bcd47c7fb9b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 158, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 158, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 553, + 179 + ], + "type": "text", + "content": "For a scattering volume of thickness " + }, + { + "bbox": [ + 313, + 158, + 553, + 179 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 158, + 553, + 179 + ], + "type": "text", + "content": ", the number of required layers is:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 403, + 176, + 553, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 176, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 403, + 176, + 553, + 198 + ], + "type": "interline_equation", + "content": "M ^ {*} = \\frac {d \\cdot \\mathbf {N A} ^ {2}}{\\lambda}. \\tag {10}", + "image_path": "6bf8198d9f0025b4235e4a7511cfe4fb7b00c75dfbdd20d7878bea94d5e26f1a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "text", + "content": "For typical numbers, consider e.g. " + }, + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "inline_equation", + "content": "\\lambda = 0.5\\mu m" + }, + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "inline_equation", + "content": "\\mathbf{N}\\mathbf{A} = 0.5" + }, + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "text", + "content": ", the plane spacing should be relatively small, and " + }, + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "inline_equation", + "content": "\\Delta_z \\leq 2\\mu m" + }, + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "text", + "content": ". For a modest tissue thickness of " + }, + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "inline_equation", + "content": "d = 200\\mu m" + }, + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "text", + "content": ", we will need " + }, + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "inline_equation", + "content": "M = 100" + }, + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "text", + "content": " planes. Clearly it is impractical to realize such a large number of planes by a physical correction system." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 273, + 496, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 496, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 496, + 285 + ], + "type": "text", + "content": "4.2. Sparse multi-slice approximations" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 287, + 554, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 554, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 554, + 418 + ], + "type": "text", + "content": "Consider a scattering sample of thickness " + }, + { + "bbox": [ + 313, + 287, + 554, + 418 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 287, + 554, + 418 + ], + "type": "text", + "content": ", which we try to approximate using " + }, + { + "bbox": [ + 313, + 287, + 554, + 418 + ], + "type": "inline_equation", + "content": "M < M^{*}" + }, + { + "bbox": [ + 313, + 287, + 554, + 418 + ], + "type": "text", + "content": " layers separated by " + }, + { + "bbox": [ + 313, + 287, + 554, + 418 + ], + "type": "inline_equation", + "content": "\\Delta_z = \\frac{d}{M}" + }, + { + "bbox": [ + 313, + 287, + 554, + 418 + ], + "type": "text", + "content": ". Our goal is to show that the fitting error " + }, + { + "bbox": [ + 313, + 287, + 554, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_M" + }, + { + "bbox": [ + 313, + 287, + 554, + 418 + ], + "type": "text", + "content": " defined in Eq. (5) decays relatively fast with " + }, + { + "bbox": [ + 313, + 287, + 554, + 418 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 287, + 554, + 418 + ], + "type": "text", + "content": " and we can get a reasonable approximation to the transmission matrix even if the number of layers is lower than the exact prediction in Eq. (10). This is a result of two main properties: (i) The volumes describing realistic tissue samples have more energy in the low frequencies; and (ii) the structure of the missing cone implies that, in any case, a significant portion of the spectrum is not sampled by the transmission matrix." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 418, + 553, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 553, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 553, + 474 + ], + "type": "text", + "content": "To understand this, consider a naive selection of " + }, + { + "bbox": [ + 313, + 418, + 553, + 474 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 418, + 553, + 474 + ], + "type": "text", + "content": " layers " + }, + { + "bbox": [ + 313, + 418, + 553, + 474 + ], + "type": "inline_equation", + "content": "\\rho_{1},\\ldots ,\\rho_{M}" + }, + { + "bbox": [ + 313, + 418, + 553, + 474 + ], + "type": "text", + "content": ". Rather than actually solving an optimization problem, we use the ground truth volume " + }, + { + "bbox": [ + 313, + 418, + 553, + 474 + ], + "type": "inline_equation", + "content": "\\hat{n} (\\kappa)" + }, + { + "bbox": [ + 313, + 418, + 553, + 474 + ], + "type": "text", + "content": ", and simply set to zero any frequency content of the RI volume " + }, + { + "bbox": [ + 313, + 418, + 553, + 474 + ], + "type": "inline_equation", + "content": "\\hat{n} (\\kappa)" + }, + { + "bbox": [ + 313, + 418, + 553, + 474 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 313, + 418, + 553, + 474 + ], + "type": "inline_equation", + "content": "\\kappa_z" + }, + { + "bbox": [ + 313, + 418, + 553, + 474 + ], + "type": "text", + "content": " values larger than the possible Nyquist range" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 395, + 474, + 553, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 474, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 395, + 474, + 553, + 495 + ], + "type": "interline_equation", + "content": "\\Lambda^ {M} = \\frac {1}{\\Delta_ {z}} = \\frac {\\lambda M}{d}, \\tag {11}", + "image_path": "f72ae9abf37355b1ca1b3c1bec369a7a5ae9efd6c7de5e6543c22385fdf3ed02.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "type": "text", + "content": "and we then Fourier transform " + }, + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "type": "inline_equation", + "content": "\\hat{n}(\\kappa)" + }, + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "type": "inline_equation", + "content": "\\pmb{n}(\\mathbf{r})" + }, + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "type": "text", + "content": " and sample planes at spacing " + }, + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "type": "inline_equation", + "content": "\\Delta_z = d / M" + }, + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "type": "text", + "content": ". The error of this approximation is basically the integral of content above the cut-off " + }, + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "type": "inline_equation", + "content": "\\Lambda^M" + }, + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "type": "text", + "content": ". Since according to Eq. (11), the cut-off frequency " + }, + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "type": "inline_equation", + "content": "\\Lambda^M" + }, + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "type": "text", + "content": " scales linearly with the number of layers, the portion of the spectrum which is lost by this low-pass operation scales linearly with the number of layers " + }, + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "type": "text", + "content": ". Thus, the naive answer is that the error of a multi-slice approximation decays linearly with " + }, + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "type": "text", + "content": ". In practice we show below that the error decays much faster, since large areas from the 3D spectrum of the sample are not used by the transmission matrix." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "To gain intuition, consider Fig. 6(b). We note that the butterfly shape is such that low 2D frequencies (i.e. low " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "|\\kappa_{xy}|" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "), also span less content along the " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\kappa_z" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " axes, and hence these frequencies are not lost even with low bandwidths " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\Lambda^M" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": ". For the higher " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "|\\kappa_{xy}|" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " frequencies, the lower " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\kappa_z" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " part, marked in dashed blue in Fig. 6(b) is preserved, while the higher portion marked in red is lost." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23170" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 127, + 167 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 127, + 167 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 127, + 167 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 127, + 167 + ], + "type": "image", + "image_path": "2ddb4ea6488694c3777f31370d0477afa592bcb08ab53079e3459e6c64263bdf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 168, + 127, + 178 + ], + "lines": [ + { + "bbox": [ + 67, + 168, + 127, + 178 + ], + "spans": [ + { + "bbox": [ + 67, + 168, + 127, + 178 + ], + "type": "text", + "content": "(a) Support primal" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 135, + 70, + 202, + 167 + ], + "blocks": [ + { + "bbox": [ + 135, + 70, + 202, + 167 + ], + "lines": [ + { + "bbox": [ + 135, + 70, + 202, + 167 + ], + "spans": [ + { + "bbox": [ + 135, + 70, + 202, + 167 + ], + "type": "image", + "image_path": "2665e7a357dd917add952742cb4d40055332310bf78b39f5de47e5510472e173.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 168, + 199, + 177 + ], + "lines": [ + { + "bbox": [ + 145, + 168, + 199, + 177 + ], + "spans": [ + { + "bbox": [ + 145, + 168, + 199, + 177 + ], + "type": "text", + "content": "(b) Support freq." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 210, + 70, + 282, + 167 + ], + "blocks": [ + { + "bbox": [ + 210, + 70, + 282, + 167 + ], + "lines": [ + { + "bbox": [ + 210, + 70, + 282, + 167 + ], + "spans": [ + { + "bbox": [ + 210, + 70, + 282, + 167 + ], + "type": "image", + "image_path": "244545752f597c82df357181f9f1869e68249011bbda892398d06aa4786b500c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 168, + 278, + 177 + ], + "lines": [ + { + "bbox": [ + 214, + 168, + 278, + 177 + ], + "spans": [ + { + "bbox": [ + 214, + 168, + 278, + 177 + ], + "type": "text", + "content": "(c) Local directions" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 185, + 296, + 253 + ], + "lines": [ + { + "bbox": [ + 55, + 185, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 185, + 296, + 253 + ], + "type": "text", + "content": "Figure 7. Layer support: (a) As point sources expand while propagating through the volume, the width of the aberration layers should be wider than the measured support " + }, + { + "bbox": [ + 55, + 185, + 296, + 253 + ], + "type": "inline_equation", + "content": "\\Omega_{p}" + }, + { + "bbox": [ + 55, + 185, + 296, + 253 + ], + "type": "text", + "content": ". (b) To express compact support transmission matrices in the Fourier basis, we illuminate the volume by a set of plane waves passing through an aperture. (c) The local cone of illumination angles reaching different points inside the volume is much smaller than the actual range of incoming illuminations." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 270, + 296, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 378 + ], + "type": "text", + "content": "Another important property of tissue is that its RI is locally smooth, and therefore if we look at its spectrum (Fig. 6(c)), we have much more energy in low frequencies than in high ones. This also agrees with the fact that tissue is forward scattering. Therefore, in the red areas of Fig. 6(b) that are not sampled by the transmission matrix, there is less energy than in the lower frequencies. Thus, despite the fact that the multi-slice approximation sacrifices such frequencies, not much energy is being lost." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 380, + 296, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 296, + 415 + ], + "type": "text", + "content": "In most cases, it is hard to give analytic formulas for the decay of " + }, + { + "bbox": [ + 55, + 380, + 296, + 415 + ], + "type": "inline_equation", + "content": "\\pmb{\\mathcal{E}}_M" + }, + { + "bbox": [ + 55, + 380, + 296, + 415 + ], + "type": "text", + "content": " as a function of " + }, + { + "bbox": [ + 55, + 380, + 296, + 415 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 380, + 296, + 415 + ], + "type": "text", + "content": ", but the supplementary file contains some derivation under simplifying assumptions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 417, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 417, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 296, + 464 + ], + "type": "text", + "content": "Multiple scattering. The results in Fig. 3(e) include multislice fits with both lower and higher optical depth values up to " + }, + { + "bbox": [ + 55, + 417, + 296, + 464 + ], + "type": "inline_equation", + "content": "OD = 4" + }, + { + "bbox": [ + 55, + 417, + 296, + 464 + ], + "type": "text", + "content": ", showing that the fitting error still decays rapidly with the number of layers." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 465, + 296, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 465, + 296, + 646 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 296, + 646 + ], + "type": "text", + "content": "While it is hard to give analytical predictions for the multiple scattering case, it appears that when the optical depth of the tissue is not too high, similar results hold. The reason is that for a forward-scattering material such as tissue, light paths only scatter at small angles, and after a small number of scattering events they are not completely scrambled. For an intuition, in the supplementary file we use a Monte-Carlo simulation illustrating that for a forward-scattering material, after a modest number of scattering events, most light paths still remain within the butterfly area of the spectrum. This is usually the regime that wavefront-shaping algorithms like [1, 2] attempt to tackle, as such algorithms attempt to push the depth at which conventional microscopes can see, but they do not yet attempt to image extremely deep where scattering is fully diffused." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 649, + 203, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 649, + 203, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 649, + 203, + 662 + ], + "type": "text", + "content": "4.3. Bounded support matrices" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 665, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 701 + ], + "type": "text", + "content": "As stated above, much fewer layers are needed if we only attempt to fit transmission matrices sampled over a small area of size " + }, + { + "bbox": [ + 55, + 665, + 295, + 701 + ], + "type": "inline_equation", + "content": "\\Omega_p\\times \\Omega_p" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "When fitting layers, we need to account for the fact that a" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "content": "point/plane illumination scattering through a volume is expanding. Thus, the aberration layers required to explain the transmission matrix are wider than the imaged area of size " + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\Omega_{p} \\times \\Omega_{p}" + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "content": ", see Fig. 7(a). Expressing a transmission matrix with a compact support at the frequency basis is equivalent to illuminating the sample with plane waves spanning a wide set of angles, but all the waves pass through a narrow aperture of size " + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\Omega_{p} \\times \\Omega_{p}" + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "content": ". In this case most points inside the volume usually do not receive light from all angles. In Fig. 7(c) we plot the narrow cone of angles arriving at three different points (a ray leaving a point " + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{\\mathbf{j}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "content": " at angle " + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\bar{\\omega}" + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "content": " is included in the cone only if this angle is not cropped by the aperture at the back of the sample). Since the cone of light reaching each point is narrower than the full numerical aperture, a local Fourier transform would not have content over the full butterfly area, but cover an even lower range of axial " + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\omega_{z}" + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "content": " frequencies. Hence, following the analysis in the previous section, it can be sampled with fewer layers. We visualize the local spectra in the supplement." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 312, + 384, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 384, + 325 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 384, + 325 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 330, + 555, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 555, + 462 + ], + "type": "text", + "content": "This paper studies how well we can fit transmission matrices with multi-slice models and how many layers are needed in practice. We show that the required number of layers is much lower than the prediction of Nyquist sampling theory. This is because the missing cone problem of 3D microscopy turns into an advantage when analyzing transmission matrices. Moreover, as tissue is forward scattering, its 3D spectrum contains most energy in the low frequencies and much less energy in the higher ones. Hence, even if the sparse sampling acts as a low-pass on the large axial frequencies, not much energy is lost." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 464, + 554, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 554, + 536 + ], + "type": "text", + "content": "We validate this prediction using both synthetic simulations and realistic lab measurements. We show that even though realistic transmission matrices contain backward and side-ward scattering which are not modeled by multi-slice matrices, they provide a reasonable approximation that is good enough for wavefront shaping focusing." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 539, + 554, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 554, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 554, + 647 + ], + "type": "text", + "content": "Our findings suggest that multi-slice models can be applicable for the design of non-local wavefront shaping systems, allowing for a volumetric correction with a small number of layered SLMs [16, 27, 35, 36]. Even if " + }, + { + "bbox": [ + 313, + 539, + 554, + 647 + ], + "type": "inline_equation", + "content": "2 - 3" + }, + { + "bbox": [ + 313, + 539, + 554, + 647 + ], + "type": "text", + "content": " SLM layers may not be enough for correcting the full field-of-view of a wide image through a thick tissue, they can allow us to increase the support of the fitted area by much more than " + }, + { + "bbox": [ + 313, + 539, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\times 2 - 3" + }, + { + "bbox": [ + 313, + 539, + 554, + 647 + ], + "type": "text", + "content": " thus they can significantly accelerate a sequential wavefront-shaping correction of a wide image." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 650, + 554, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 650, + 554, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 650, + 554, + 673 + ], + "type": "text", + "content": "Acknowledgments: This research was funded by ERC SpeckleCorr-101043471, ISF 563/24." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 677, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 715 + ], + "type": "text", + "content": "Code and data: can be obtained at https:// webee.technion.ac.il/people/anat.levin/ papers/LayerdTmat(cvpr25_code.zip" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23171" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 91, + 295, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 295, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 295, + 123 + ], + "type": "text", + "content": "[1] Dror Aizik and Anat Levin. Non-invasive and noise-robust light focusing using confocal wavefront shaping. Nature communications, 15(5575), 2024." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 124, + 295, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 124, + 295, + 157 + ], + "spans": [ + { + "bbox": [ + 61, + 124, + 295, + 157 + ], + "type": "text", + "content": "[2] Dror Aizik, Ioannis Gkioulekas, and Anat Levin. Fluorescent wavefront shaping using incoherent iterative phase conjugation. Optica, 9(7):746-754, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 158, + 294, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 158, + 294, + 190 + ], + "spans": [ + { + "bbox": [ + 62, + 158, + 294, + 190 + ], + "type": "text", + "content": "[3] M. Alterman, C. Bar, I. Gkioulekas, and A. Levin. Imaging with local speckle intensity correlations: theory and practice. ACM TOG, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 191, + 294, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 294, + 224 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 294, + 224 + ], + "type": "text", + "content": "[4] YoonSeok Baek, Hilton B. de Aguiar, and Sylvain Gigan. Phase conjugation with spatially incoherent light in complex media. Nature Photonics, 17(12):1114-1119, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 225, + 294, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 225, + 294, + 278 + ], + "spans": [ + { + "bbox": [ + 62, + 225, + 294, + 278 + ], + "type": "text", + "content": "[5] Paul Balondrade, Victor Barolle, Nicolas Guigui, Emeric Auriant, Nathan Rougier, Claude Boccara, Mathias Fink, and Alexandre Aubry. Multi-spectral reflection matrix for ultrafast 3d label-free microscopy. Nature Photonics, 18(10): 1097-1104, 2024." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 280, + 294, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 294, + 312 + ], + "type": "text", + "content": "[6] Chen Bar, Marina Alterman, Ioannis Gkioulekas, and Anat Levin. A monte carlo framework for rendering speckle statistics in scattering media. ACM TOG, 2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 314, + 294, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 314, + 294, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 314, + 294, + 345 + ], + "type": "text", + "content": "[7] Chen Bar, Ioannis Gkioulekas, and Anat Levin. Rendering near-field speckle statistics in scattering media. ACM TOG, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 347, + 294, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 294, + 379 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 294, + 379 + ], + "type": "text", + "content": "[8] Martin Booth. Adaptive optical microscopy: The ongoing quest for a perfect image. Light: Science and Applications, 3:e165, 2014." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 380, + 294, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 294, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 294, + 413 + ], + "type": "text", + "content": "[9] Bingquan Chen and Jakob J. Stamnes. Validity of diffraction tomography based on the first born and the first rytrov approximations. Appl. Opt., 37(14):2996-3006, 1998." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 415, + 295, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 415, + 295, + 457 + ], + "spans": [ + { + "bbox": [ + 57, + 415, + 295, + 457 + ], + "type": "text", + "content": "[10] Wonshik Choi, Christopher Fang-Yen, Kamran Badizadegan, Seungeun Oh, Niyom Lue, Ramachandra R Dasari, and Michael S Feld. Tomographic phase microscopy. Nature methods, 4(9):717-719, 2007." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 459, + 294, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 459, + 294, + 512 + ], + "spans": [ + { + "bbox": [ + 57, + 459, + 294, + 512 + ], + "type": "text", + "content": "[11] Shwetadwip Chowdhury, Michael Chen, Regina Eckert, David Ren, Fan Wu, Nicole Repina, and Laura Waller. High-resolution 3d refractive index microscopy of multiple-scattering samples from intensity images. Optica, 6(9): 1211-1219, 2019." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 514, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 294, + 557 + ], + "type": "text", + "content": "[12] Anat Daniel, Dan Oron, and Yaron Silberberg. Light focusing through scattering media via linear fluorescence variance maximization, and its application for fluorescence imaging. Optics Express, 27(15):21778, 2019." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 558, + 294, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 294, + 580 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 294, + 580 + ], + "type": "text", + "content": "[13] A. J. Devaney. Inverse-scattering theory within the rytov approximation. Opt. Lett., 6(8):374-376, 1981." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 581, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 294, + 612 + ], + "type": "text", + "content": "[14] Amicia D. Elliott. Confocal microscopy: Principles and modern practices. Current Protocols in Cytometry, 92(1): e68, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 614, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 294, + 669 + ], + "type": "text", + "content": "[15] Brandon Y. Feng, Haiyun Guo, Mingyang Xie, Vivek Boominathan, Manoj K. Sharma, Ashok Veeraraghavan, and Christopher A. Metzler. Neuws: Neural wavefront shaping for guidestar-free imaging through static and dynamic scattering media. Science Advances, 9(26):eadg4671, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[16] T. Furieri, A. Bassi, and S. Bonora. Large field of view aberrations correction with deformable lenses and multi conjugate adaptive optics. Journal of Biophotonics, 16(12): e202300104, 2023." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[17] Omri Haim, Jeremy Boger-Lombard, and Ori Katz. Image-guided computational holographic wavefront shaping. Nature Photonics, 2024." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 107, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 149 + ], + "type": "text", + "content": "[18] Karen Hampson, Raphael Turcotte, Donald Miller, Kazuhiro Kurokawa, Jared Males, Na Ji, and Martin Booth. Adaptive optics for high-resolution imaging. Nature Reviews Methods Primers, 1:68, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 151, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 553, + 172 + ], + "type": "text", + "content": "[19] John W. Hardy. Adaptive Optics for Astronomical Telescopes. 1998." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "type": "text", + "content": "[20] Yi Hua, M. Salman Asif, and Aswin C. Sankaranarayanan. Spatial and axial resolution limits for mask-based lensless cameras. Opt. Express, 31(2):2538-2551, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 206, + 553, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 260 + ], + "type": "text", + "content": "[21] Seungwon Jeong, Ye-Ryoung Lee, Wonjun Choi, Sungsam Kang, Jin Hee Hong, Jin-Sung Park, Yong-Sik Lim, HongGyu Park, and Wonshik Choi. Focusing of light energy inside a scattering medium by controlling the time-gated multiple light scattering. Nature Photonics, 12(5):277-283, 2018." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 261, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 282 + ], + "type": "text", + "content": "[22] Na Ji. Adaptive optical fluorescence microscopy. Nature Methods, 14:374-380, 2017." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 284, + 553, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 284, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 317, + 284, + 553, + 347 + ], + "type": "text", + "content": "[23] Yonghyeon Jo, Ye-Ryoung Lee, Jin Hee Hong, Dong-Young Kim, Junhwan Kwon, Myunghwan Choi, Moonseok Kim, and Wonshik Choi. Through-skull brain imaging in vivo at visible wavelengths via dimensionality reduction adaptive-optical microscopy. Science Advances, 8(30):eabo4366, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 350, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 404 + ], + "type": "text", + "content": "[24] Andrew Jr, Alon Harris, Josh Gross, Ingrida Januleviciene, Aaditya Shah, and Brent Siesky. Optical coherence tomography angiography: An overview of the technology and an assessment of applications for clinical research. British Journal of Ophthalmology, 101, 2016." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 405, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 405, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 317, + 405, + 553, + 480 + ], + "type": "text", + "content": "[25] Sungsam Kang, Pilsung Kang, Seungwon Jeong, Yongwoo Kwon, Taeseok D. Yang, Jin Hee Hong, Moonseok Kim, Kyung-Deok Song, Jin Hyoung Park, Jun Ho Lee, Myoung Joon Kim, Ki Hean Kim, and Wonshik Choi. High-resolution adaptive optical imaging within thick scattering media using closed-loop accumulation of single scattering. Nature Communications, 8(1):2157, 2017." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 482, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 525 + ], + "type": "text", + "content": "[26] Sungsam Kang, Yongwoo Kwon, Hojun Lee, Seho Kim, Jin Hee Hong, Seokchan Yoon, and Wonshik Choi. Tracing multiple scattering trajectories for deep optical imaging in scattering media. Nature communications, 14(1), 2023." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 526, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 569 + ], + "type": "text", + "content": "[27] Marie Laslandes, Matthias Salas, Christoph K. Hitzenberger, and Michael Pircher. Increasing the field of view of adaptive optics scanning laser ophthalmoscopy. Biomed. Opt. Express, 8(11):4811-4826, 2017." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 571, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 553, + 613 + ], + "type": "text", + "content": "[28] V. Lauer. New approach to optical diffraction tomography yielding a vector equation of diffraction tomography and a novel tomographic microscope. Journal of Microscopy, 205 (2):165-176, 2002." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 614, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 553, + 657 + ], + "type": "text", + "content": "[29] F. C. Lin and M. A. Fiddy. The born-rytov controversy: I. comparing analytical and approximate expressions for the one-dimensional deterministic case. J. Opt. Soc. Am. A, 9 (7):1102–1110, 1992." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 658, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 680 + ], + "type": "text", + "content": "[30] Jerome Mertz. Introduction to optical microscopy. Cambridge University Press, 2019." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[31] Sagi Monin, Aswin Sankaranarayanan, and Anat Levin. Exponentially-wide etendue displays using a tilting cascade. IEEE ICCP, pages 1-12, 2022." + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23172" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 442 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[32] Ulysse Najar, Victor Barolle, Paul Balondrade, Mathias Fink, Claude Boccara, and Alexandre Aubry. Harnessing forward multiple scattering for optical imaging deep inside an opaque medium. Nature Communications, 15(1), 2024." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "text", + "content": "[33] Tan Nguyen, Mikhail Kandel, Marcello Rubessa, Matthew Wheeler, and Gabriel Popescu. Gradient light interference microscopy for 3d imaging of unlabeled specimens. Nature Communications, 8, 2017." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 195 + ], + "type": "text", + "content": "[34] Galya Stern and Ori Katz. Noninvasive focusing through scattering layers using speckle correlations. Optics Letters, 44(1):143, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "text", + "content": "[35] Jörgen Thaung, Per Knutsson, Zoran Popovic, and Mette Owner-Petersen. Dual-conjugate adaptive optics for wide-field high-resolution retinal imaging. Opt. Express, 17(6): 4454-4467, 2009." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "type": "text", + "content": "[36] Tsai wei Wu and Meng Cui. Numerical study of multiconjugate large area wavefront correction for deep tissue microscopy. Opt. Express, 23(6):7463-7470, 2015." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "text", + "content": "[37] Gil Weinberg, Elad Sunray, and Ori Katz. Noninvasive megapixel fluorescence microscopy through scattering layers by a virtual incoherent reflection matrix. Science Advances, 10(47):eadl5218, 2024." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 320, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 352 + ], + "type": "text", + "content": "[38] Emil Wolf. Three-dimensional structure determination of semi-transparent objects from holographic data. Optics Communications, 1(4):153-156, 1969." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 354, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 386 + ], + "type": "text", + "content": "[39] Tomer Yeminy and Ori Katz. Guidestar-free image-guided wavefront shaping. Science Advances, 7(21):eabf5364, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "type": "text", + "content": "[40] Lei Zhu, Fernando Soldevila, Claudio Moretti, Alexandra d'Arco, Antoine Boniface, Xiaopeng Shao, Hilton B. de Aguiar, and Sylvain Gigan. Large field-of-view non-invasive imaging through scattering layers using fluctuating random illumination. Nature Communications, 13:1447, 2022." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23173" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Uni-Renderer_ Unifying Rendering and Inverse Rendering Via Dual Stream Diffusion/ef5cd413-b783-4606-bcb0-152603271134_content_list.json b/2025/Uni-Renderer_ Unifying Rendering and Inverse Rendering Via Dual Stream Diffusion/ef5cd413-b783-4606-bcb0-152603271134_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..14315c037f04b45f2b655cd2c94a70df75d96672 --- /dev/null +++ b/2025/Uni-Renderer_ Unifying Rendering and Inverse Rendering Via Dual Stream Diffusion/ef5cd413-b783-4606-bcb0-152603271134_content_list.json @@ -0,0 +1,1734 @@ +[ + { + "type": "text", + "text": "Uni-Renderer: Unifying Rendering and Inverse Rendering Via Dual Stream Diffusion", + "text_level": 1, + "bbox": [ + 116, + 102, + 880, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhifei Chen $^{1,*}$ , Tianshuo Xu $^{1,*}$ , Wenhang Ge $^{1,*}$ , Leyi Wu $^{1}$ , Dongyu Yan $^{1}$ , Jing He $^{1}$ , Luozhou Wang $^{1}$ , Lu Zeng $^{3}$ , Shunsi Zhang $^{3}$ , Yingcong Chen $^{1,2,\\dagger}$ , ${}^{1}\\mathrm{HKUST}(\\mathrm{GZ})$ , ${}^{2}\\mathrm{HKUST}$ , ${}^{3}\\mathrm{Quwan}$", + "bbox": [ + 99, + 174, + 895, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zchen379, txu647}@connect.hkust-gz.edu.cn; yingcongchen@ust.hk", + "bbox": [ + 217, + 231, + 774, + 247 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/33c954137fe5c44b190944d552c103fdb2941765da13a0c5c50d4c0642914194.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 104, + 273, + 892, + 364 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/71eb58cb509cdc30ace89b4995357217a5bb94d9376531626877cf5d150bf5be.jpg", + "image_caption": [ + "Figure 1. Our framework, Uni-renderer, empowers the generative model to function both as a renderer and an inverse renderer by approximating the rendering equation using a data-driven approach. Given intrinsic attributes, Uni-renderer generates photo-realistic images, functioning as a renderer. When provided with a single RGB image, it effectively decomposes the intrinsic properties, functioning as an inverse renderer. Top: Uni-renderer generates smooth variations as renderer. Setting the roughness value to 1.0 results in the \"dog\" case, shown at the top, lacking specular highlights. Conversely, setting the metallic value to 1 makes the \"hat\" case appear metallic. Bottom Left: When functioning as an inverse renderer, Uni-renderer decomposes the intrinsic properties of a single RGB image. Bottom Right: Uni-renderer generates relighting results under different environment lightings." + ], + "image_footnote": [], + "bbox": [ + 102, + 369, + 890, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 680, + 326, + 695 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rendering and inverse rendering are pivotal tasks in both computer vision and graphics. The rendering equation is the core of the two tasks, as an ideal conditional distribution transfer function from intrinsic properties to RGB images. Despite achieving promising results of existing rendering methods, they merely approximate the ideal estimation for a specific scene and come with a high computational cost. Additionally, the inverse conditional distribution transfer is intractable due to the inherent ambiguity. To address these challenges, we propose a data-driven method that jointly models rendering and inverse rendering as two conditional", + "bbox": [ + 88, + 700, + 485, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "generation tasks within a single diffusion framework. Inspired by UniDiffuser, we utilize two distinct time schedules to model both tasks, and with a tailored dual streaming module, we achieve cross-conditioning of two pre-trained diffusion models. This unified approach, named Uni-Renderer, allows the two processes to facilitate each other through a cycle-consistent constrain, mitigating ambiguity by enforcing consistency between intrinsic properties and rendered images. Combined with a meticulously prepared dataset, our method effectively decomposition of intrinsic properties and demonstrating a strong capability to recognize changes during rendering.", + "bbox": [ + 511, + 681, + 908, + 864 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. †Corresponding Author. This project is supported by Quwan Technology.", + "bbox": [ + 112, + 876, + 372, + 902 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26504", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 89, + 222, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Physically based rendering is crucial in computer vision and graphics for producing photorealistic 2D images from 3D models, textures, and lighting setups. This process is essential in animation [43], and architectural visualization [23]. At its core, the rendering equation [18] models the flow of light energy within a scene, offering an ideal conditional distribution transfer function for rendering. Monte Carlo light-transport simulations [34], utilizing path tracing [24], are commonly employed to evaluate the rendering equation. However, its recursive tracing incurs significant computational demands, posing challenges for real-time implementation. Inverse rendering, which aims to deduce geometric, material, and lighting information from images, has long been studied due to its under-constrained nature [10]. This approach allows the reconstructed 3D model to be directly integrated into rendering engines, thereby playing a critical role in downstream applications such as game production [25].", + "bbox": [ + 89, + 114, + 485, + 385 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advancements employ differentiable rendering for 3D representations [5, 45] or focus on direct decomposition of intrinsics using 2D object priors [7, 11]. These approaches use differentiable renderer to integrate intrinsic properties into RGB images, facilitating loss computation against ground-truth images for optimization. For instance, MaterialGAN [11] leverages StyleGAN2 [19] to synthesize realistic material properties using a large-scale, spatially-varying material dataset [7], while other works [4, 22, 44, 46] utilize diffusion models to achieve better results. Nevertheless, the ambiguity between geometry, materials, and lighting still hinders the effectiveness of decomposition. The mapping from observed images back to intrinsic properties is not one-to-one, leading to ambiguity and suboptimal performance. The most similar work to ours is RGB2X [47], which performs both forward and inverse rendering using two separate diffusion models. However, it fails to establish a connection between the two processes, treating them independently. Ambiguity issue still exists.", + "bbox": [ + 89, + 386, + 483, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To alleviate this issue, we propose to jointly learn the two distributions together. By integrating rendering and inverse rendering into a single framework from a multi-task learning perspective, the two processes can facilitate each other. This joint learning allows us to use the inverted intrinsic properties to perform another cycle of rendering, effectively creating a cycle-consistent constrains. This cycle rendering can mitigate the ambiguity problem by enforcing consistency between the intrinsic properties and the rendered images, leading to improved performance.", + "bbox": [ + 89, + 674, + 483, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we explicitly model the rendering process and inverse rendering as two conditional generation tasks. Inspired by UniDiffuser [1], we utilize two distinct time schedules to model both conditional generation tasks within a single diffusion pipeline. By cross-conditioning two pre", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "trained diffusion models through a dual streaming module, we achieve both rendering and inverse rendering in a unified framework.", + "bbox": [ + 511, + 90, + 906, + 133 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, we introduce Uni-renderer which utilizes a pre-trained diffusion model to handle both rendering and inverse rendering in a unified framework. The key contributions of our method are as follows:", + "bbox": [ + 511, + 136, + 906, + 196 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a data-driven method to approximate the rendering equation. By modeling rendering and inverse rendering as two conditional generation tasks, we design a unified diffusion model with a tailored dual stream module for cross-condition generation.", + "- We render a synthetic dataset of different fine-grained material edits using 200K 3D objects with randomized intrinsic attributes by varying one of the rendering elements at a time while keeping the others constant.", + "- Extensive experiments validate the effectiveness of our method, achieving robust disentangling of intrinsic properties and demonstrating a strong capability to recognize changes in rendering." + ], + "bbox": [ + 513, + 200, + 908, + 397 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 412, + 653, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Rendering", + "text_level": 1, + "bbox": [ + 513, + 436, + 633, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Rendering, the process of generating 2D images from 3D models, synthesizes raw data from a 3D scene—including geometry, materials, and lighting—using the rendering equation [18]. This task is typically performed by rendering engines, such as Blender [14] and Unity [12], which serve as the technological intermediaries converting 3D models into 2D images or videos. However, achieving photo-realistic 2D images often involves recursive ray tracing [38] or path tracing [24], techniques that closely approximate the rendering equation but are time-consuming. Although several GPU-based acceleration methods have been proposed [8, 13, 21], and existing engines have incorporated these tools, real-time rendering continues to pose significant challenges. In contrast to traditional rendering approaches, our method considers the diffusion model as an alternative renderer, which utilizes geometry, materials, and lighting as conditional cues to produce photo-realistic 2D images without the need for recursive tracing.", + "bbox": [ + 511, + 459, + 908, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Inverse Rendering", + "text_level": 1, + "bbox": [ + 513, + 742, + 692, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inverse rendering [2, 33], the task of decomposing an image's appearance into intrinsic properties such as geometry, material, and lighting, remains a longstanding and severely ill-posed problem in computer vision and graphics. Some approaches employ differentiable renderers on 3D representations [5, 20] to directly optimize these properties using image losses. Advancements in volume rendering, as utilized in Neural Radiance Fields (NeRFs) [29] for 3D reconstruction from multi-view 2D images, have led most", + "bbox": [ + 511, + 763, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26505", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b8e693381e9e32c516785c82b1b60055ce426559187e9a6dc017121808dd9454.jpg", + "image_caption": [ + "Figure 2. The overview of our pipeline. During training, both attribute and RGB images are input to a unified model with pre-trained VAE encoders. The timestep selector plays a crucial role by adjusting the timesteps for each branch. Specifically, it ensures that one branch (either the attribute or RGB) has a timestep of 0, while the other branch selects a timestep from $t \\in [0, T]$ . This mechanism allows our model to effectively learn the conditional distributions $q(\\mathbf{x}_0|\\mathbf{y}_0)$ , $q(\\mathbf{y}_0|\\mathbf{x}_0)$ in alternating iterations. During rendering and inverse rendering, the corresponding conditions are input to the model with a timestep of 0, and the attributes/RGB images are generated through a sampled noise. (The VAE encoder and decoder are omitted for simplicity.)" + ], + "image_footnote": [], + "bbox": [ + 94, + 98, + 911, + 338 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "methods [9, 26, 27, 45, 48] to leverage multiple images for reconstruction and subsequent estimation of materials and lighting. Additionally, they require per-object optimization and struggle to generalize across different objects, limiting their efficiency in fast estimation. MaterialGAN [11] utilizes a StyleGAN2-based model for generating intrinsic properties and a renderer for novel view synthesis under specified lighting conditions. Similarly, SIC [7] employs an encoder-decoder network for single-image-based inverse rendering, where rendering loss is used for optimization. Nonetheless, these efforts predominantly focus on planar surfaces and still necessitate a renderer to integrate predicted attributes into the final rendered image. [4, 22, 44, 46, 47] utilize diffusion models to achieve better results. However, these methods have not achieved optimal performance due to the inherent ambiguity involved in transferring distributions from images to intrinsic properties. Our method effectively mitigates the ambiguity among intrinsic properties by introducing a dual-stream diffusion pipeline and combined with a meticulously prepared dataset.", + "bbox": [ + 91, + 463, + 483, + 765 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 772, + 259, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Models [15, 32, 35, 39, 40] belong to the class of probabilistic generative models that progressively destruct data by injecting noise, and then subsequently learn to reverse this process in order to generate new samples. Alchemist [37] leverages a pre-trained diffusion model to achieve material control over a given object. [42] adopts a collaborative diffusion pipeline to decompose physically-", + "bbox": [ + 89, + 794, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "based rendering (PBR) material properties from RGB. However, these works focus either on rendering or inverse rendering and fail to yield a unified model to fit in real-world usages. Alternatively, we design a dual stream diffusion model as both a renderer and an inverse renderer. During the inverse rendering stage, it takes a single RGB image as input and disentangles all the intrinsic properties. For the rendering stage, all of the intrinsic properties including material, geometry, and lighting conditions will be taken and rendered via the same model.", + "bbox": [ + 511, + 463, + 908, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 513, + 666, + 624, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a single RGB image $\\mathbf{I}$ , inverse rendering aims at decomposing it into intrinsic attributes including metallic $m$ , roughness $r$ , albedo $\\mathbf{a}$ , surface normal $\\mathbf{n}$ , specular lighting $\\mathbf{s}$ and diffuse lighting $\\mathbf{d}$ , which is formulated as $\\{m, r, \\mathbf{a}, \\mathbf{n}, \\mathbf{s}, \\mathbf{d}\\} = \\text{InverseRenderer}(\\mathbf{I})$ . For later reference, we denote the combination of attributes as $\\mathbf{C}$ . Renderer takes attributes as input and renders them into RGB image denoted as $\\mathbf{I} = \\text{Renderer}(\\mathbf{C})$ . We formulate a dual stream diffusion model as renderer and inverse renderer simultaneously. We start by briefly revisiting physically based rendering (PBR) and Unidiffuser [1] in Section 3.1. Next, we introduce our dual stream diffusion framework in Section 3.2. We elaborate on the data preparation in Section 3.3.", + "bbox": [ + 511, + 703, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26506", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d8995d7a23d05fff4967f6a9ed356893286239cab760d1fe6fa3605d22f9644d.jpg", + "image_caption": [ + "Figure 3. We demonstrate smooth changes via rendering for different metallic and roughness strengths. The rendering is performed giving different combinations of the attributes. When the roughness value was set to 1, the cake and clock case shown in the top left are without specular highlights. When the metallic value was set to 1, the orange and baseball cases appeared to be metallic and revealing object illumination. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 91, + 90, + 903, + 424 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 510, + 230, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Physically Based Rendering. Physically-based rendering is a computer graphics approach that seeks to render images by modeling the interaction between lights and materials in the real world. At its core, the rendering equation [18] describes the light energy flow in the scene, formulated as", + "bbox": [ + 89, + 536, + 482, + 612 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL (\\mathbf {x}, \\omega_ {o}) = \\int_ {\\Omega} f (\\mathbf {x}, \\omega_ {o}, \\omega_ {i}) L _ {i} (\\mathbf {x}, \\omega_ {i}) (\\omega_ {i}, \\mathbf {n}) d \\omega_ {i}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 632, + 482, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\omega_{o}$ is the viewing direction of the outgoing light, $f$ is the BRDF properties, $L_{i}$ is the incident light of direction $\\omega_{i}$ sampled from the upper hemisphere $\\Omega$ , and $\\mathbf{n}$ is the surface normal. The rendering equation is the key to handle both rendering and inverse rendering problems. Given intrinsic properties, the rendering equation generates photo-realistic images. Given an image, a rendering equation can also be used to optimize intrinsic properties.", + "bbox": [ + 89, + 670, + 482, + 791 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unidiffuser. UniDiffuser [1], a unified diffusion model that models all distributions at the same time, injects noise $\\epsilon^x$ and $\\epsilon^y$ to a set of paired image-text data $(\\mathbf{x_0},\\mathbf{y_0})$ and generates noisy data $\\mathbf{x_t^x}$ and $\\mathbf{y_t^y}$ , where $0\\leq t_{x},t_{y}\\leq T$ represent two individual timesteps. It then trains a joint noise prediction network $\\epsilon_{\\theta}(\\mathbf{x_{t^{x}}},\\mathbf{y_{t^{y}}},t_{x},t_{y})$ to predict the noise $\\epsilon^x$ and $\\epsilon^y$ by minimizing the mean squared error loss:", + "bbox": [ + 89, + 794, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\epsilon^ {\\mathbf {x}}, \\epsilon^ {\\mathbf {y}}, \\mathbf {x} _ {0}, \\mathbf {y} _ {0}} {\\mathbb {E}} \\left[ \\left\\| [ \\epsilon^ {\\mathbf {x}}, \\epsilon^ {\\mathbf {y}} ] - \\epsilon_ {\\theta} (\\mathbf {x} _ {\\mathbf {t}} ^ {\\mathbf {x}}, \\mathbf {y} _ {\\mathbf {t}} ^ {\\mathbf {y}}, t _ {x}, t _ {y}) \\| ^ {2} \\right], \\right. \\quad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 531, + 906, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By predicting $\\epsilon_{\\theta}(\\mathbf{x}_{t^{\\mathbf{x}}},\\mathbf{y}_{t^{\\mathbf{y}}},t_x,t_y)$ for any $t_x$ and $t_y$ , UniDiffuser learns all distributions related to $(\\mathbf{x}_0,\\mathbf{y}_0)$ . This includes all conditional distributions: $q(\\mathbf{x}_0|\\mathbf{y}_0)$ , $q(\\mathbf{y}_0|\\mathbf{x}_0)$ , and those conditioned on noisy input, for example $q(\\mathbf{x}_0|\\mathbf{y}_{t^{\\mathbf{y}}})$ and $q(\\mathbf{y}_0|x_{t^x})$ , for $0\\leq t_x,t_y\\leq T$ . Learning a conditional distribution can be seen as learning a distinct task. From a multitask learning perspective, due to limited network bandwidth, learning many tasks simultaneously (i.e., fitting all distributions to a single network) may result in task competition or task conflict, ultimately leading to sub-optimal performance. For rendering and inverse rendering, we exclusively modeled the two conditional distributions $q(\\mathbf{x}_0|\\mathbf{y}_0)$ , $q(\\mathbf{y}_0|\\mathbf{x}_0)$ to resolve the aforementioned issue and enhance the performance.", + "bbox": [ + 511, + 568, + 908, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Uni-Renderer", + "text_level": 1, + "bbox": [ + 511, + 787, + 658, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our design incorporates a pre-trained RGB image model and a PBR model, the two were tightly linked to one another via a dual streaming technique aligning the RGB model's information with the PBR model's information. The overall framework is shown in Figure 2. Next, we will discuss each sub-module in the pipeline in detail.", + "bbox": [ + 511, + 810, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26507", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fce9c77b5886093e0d36e348354bc70aa86f3c35ec6db4f9e16f6816a3d5542b.jpg", + "image_caption": [ + "Ref. Image" + ], + "image_footnote": [], + "bbox": [ + 107, + 88, + 197, + 359 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d076f9198be8b4fd20c611e1ba18c19467acdec790e26e1b3c5aa7b9fb512e14.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 98, + 269, + 150 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/95eb62e66ba26b82f097ebd061a696f92d1ff9a8f4026ce296bb747c867e0f19.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 165, + 269, + 219 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8efbd67ed6f43f507ee577ac9bd41c4cb3eae40eccf25c3abe822882b63d5bf8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 233, + 269, + 287 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2acebc2d319bec8331123d806b54053663f6005974c26bede52fde47f9e93282.jpg", + "image_caption": [ + "Yi.et al" + ], + "image_footnote": [], + "bbox": [ + 210, + 304, + 267, + 357 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/de07a5ca93e501833ef95e4fb995a3d6ff503b3f40ed2f36ba9c2cd735ede5f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 98, + 357, + 150 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/629169e4536c96fc11fc6098d814fb892a8af092f07faa244fe50dfe9f73fc1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 308, + 166, + 341, + 218 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4cd34e57ead8a77ff649f5074f60f0395caa22b8076eb4c3b4c7289a7bb99606.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 233, + 357, + 286 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fe36d1ac3d0b44e56ff68bb5f4217ceabdf5c90cfd3f56473cb6abff4f915403.jpg", + "image_caption": [ + "RGB2X" + ], + "image_footnote": [], + "bbox": [ + 300, + 304, + 354, + 356 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c3c9a53ac0a33108adfe8c3ac65a63c41422d6619f5448796289e4a695092468.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 98, + 446, + 150 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/025b74baa17b9e1770c2fcc396c73d00897e3fe3d16b1a4688ad86bac5cd821f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 166, + 433, + 217 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c48bc7781711d78f4c01d712a2e001983b7a1b2339fc259590611f42ee14f63b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 387, + 232, + 449, + 285 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/acab94daa601a6f9eacb05012c31a294b2fab02c9700d184ea2a62f3ab6b0355.jpg", + "image_caption": [ + "IntrinImgDecomp", + "Figure 4. Albedo comparison. Albedo Comparison of Uni-Renderer with baseline methods. We compared 4 learning-based methods and 2 optimization-based methods. Among all, Uni-renderer yields the most realistic results. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 388, + 304, + 442, + 356 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6070618d1b4ee747ec4277038b8adfe4cfd9eedce46f7db7835fb350a9533b92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 475, + 98, + 532, + 150 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4dd6ce48e766c294e072dfd0d4d3fa7a774561bb2b3fbcf0b6519be056a5bf84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 167, + 519, + 217 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8efc27d847f0ad3a0f2d0f51d31f61ddca6a8140c258071baf80758f0fcc1d2a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 475, + 233, + 535, + 285 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d0c1cd7ea4ec1954b8871c591869e7915a8e6a7c53d6b69758b8c2d15711414d.jpg", + "image_caption": [ + "IntrinImgDiff." + ], + "image_footnote": [], + "bbox": [ + 480, + 304, + 534, + 356 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fadbb50a2b5bc704872812686a38bd24a3e42b934772aa85a52500db989a6ea7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 98, + 622, + 150 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/deddfa13121f4d58715c8fee9367d76b20b5de98fb3ea07cc21c59821ca33574.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 167, + 607, + 217 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/064c89d599ce4dc1967cc22b0ff7c920af7677827df67f3375e51a44efce3ba8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 232, + 625, + 285 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4dd84127bce98846052708e9441ee80cdb5319003450c3533f1b95ede475f947.jpg", + "image_caption": [ + "GaussShader" + ], + "image_footnote": [], + "bbox": [ + 566, + 304, + 622, + 356 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/84be300370926bf4794e720243080b41556bf15ec0c2bb3b97cdf07034927eff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 98, + 704, + 150 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ba724ebcabf3a2cf4040441e20f8b586ebf28b4d3e7c874d2c6cdc98460d1e76.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 167, + 689, + 217 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/77514a80d79ebc0478f20d91a56c558565816321419609f6e57908aeea4ec3a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 648, + 232, + 707, + 285 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5f80d967e15a91c65476f2bdb2cba51031e30fb2e0f53fcd787ed2f0e98808eb.jpg", + "image_caption": [ + "NVdiffRec" + ], + "image_footnote": [], + "bbox": [ + 651, + 304, + 704, + 356 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/29b54965e165b782974368cc94761be3e02b00d854a83b2f785dc8a6cf90251b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 98, + 787, + 150 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3cb57a5a4e34ed4a5af02f323f1a0f141e60fe8dba8ecf44e3ff6de0fbedcbd6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 741, + 167, + 771, + 215 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4344b182ac230dadee907d23f61656dab4ce1f552047722d0a8d2e6a3de2a0b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 727, + 232, + 789, + 285 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2fb697f4a063b8da59c503650be4cece06832ddc10cadc4bfcc6f8ef305c45f3.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 733, + 304, + 784, + 354 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9080c5ef2ccf5f2659382b804e24618a5d9950722d11b01a39dc55c3e5be3a07.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 816, + 98, + 872, + 148 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e9372a335f015cd024f794c2cfcb26e892e6a99e0e2513e01c579e1b95acb84f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 828, + 167, + 857, + 215 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ac965cae4efccbc788eec2d061f14b237f631973b8784b59fcd37a24f23222ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 813, + 232, + 874, + 285 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/25700b7a6a24c0d40c0d8c0399f8449af6f02a435858fe234c8f8d70575e8aa9.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 815, + 303, + 867, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Latent Preparation. Recent advancements have benefited greatly from the pre-trained VAE [36], which downsample an RGB image into lower dimensional latent. We used separate VAE for encoding $[m, r]$ , albedo a, surface normal n and environment lighting s, d before we channel-wisely concatenate them and feed them into the model. It is worth noting that attributes like metallic and roughness, often come with a scalar value, and cannot be directly encoded. Thus, we generate additional masks to binarize the two values into gray-scale images. Instead of having two separate VAE for m and r, We formed channel triplets $a_{\\mathrm{material}}$ , consisting of $[m, r, \\mathbf{m}]$ and process those with the RGB VAE, where m is the mask This design allows us to reduce the bandwidth overhead within the diffusion model and further improve generation quality.", + "bbox": [ + 91, + 455, + 483, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Conditional Distributions Modeling. To achieve rendering and inverse rendering within a single model, we borrowed ideas from [1], and adopted two different timesteps $t_{\\mathrm{attributes}}$ , $t_{\\mathrm{RGB}}$ from three timesteps choices $\\{0, t, T\\}$ to train a joint noise prediction network. It is worth noting that since our unified model only requires modeling two conditional distributions, we ruled out the choices where both $t_{\\mathrm{attributes}}$ and $t_{\\mathrm{RGB}}$ take on $t$ or $T$ and make sure there is always a timestep equal to 0. Formally,", + "bbox": [ + 89, + 685, + 483, + 821 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(t _ {\\text {a t t r i b u t e s}}, t _ {\\mathrm {R G B}}\\right) = \\left\\{ \\begin{array}{l l} (0, \\tilde {t}) & \\text {d u r i n g} \\\\ (\\tilde {t}, 0) & \\text {i n v e r s e} \\end{array} , \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 833, + 483, + 875 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 109, + 886, + 153, + 898 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {t} = \\left\\{ \\begin{array}{l l} t & \\text {w i t h p r o b a b i l i t y} p \\\\ T & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 473, + 906, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The pseudo-code for generating timesteps can be found in supplementary material.", + "bbox": [ + 511, + 526, + 906, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dual Stream Diffusion. We designed a dual diffusion network to facilitate the unified model and adopt the $\\mathbf{x}_0$ -prediction in the diffusion process. Training: During training, as shown in Figure 2, the upper branch of the model dealt with the intrinsic attributes only, leaving only RGB for the lower branch. The timestep selection module will determine if a rendering/inverse rendering iteration has occurred at a probability $p$ . During a rendering iteration, the clean attributes $\\mathbf{C}$ will be used as conditions to denoise the noisy RGB image $\\mathbf{I}$ , the same will be applied to the inverse rendering iteration except that the noisy attributes are being fed to the model. Inference: At the inference stage, we provided attributes as conditions to the model and gradually denoise the RGB, and vice versa. The effectiveness of the dual stream framework is demonstrated in Sec 4.3.", + "bbox": [ + 511, + 561, + 908, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Cycle-Consistent Constrain. To alleviate the inherent ambiguity problem, we introduce a cycle-consistent constraint within our unified framework, where its incorporation is straightforward with our pipeline. During training, we use the model's predicted inverse results to perform an additional cycle of rendering. The re-rendered outputs are then utilized in the loss calculation to further optimize the inverse", + "bbox": [ + 511, + 794, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26508", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a92650899f318613a3665f18c14a5dbb7f0abae4ca24dde9626215b732189dc1.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison. Rendering Comparison of Uni-Renderer with baseline methods. The base images are with a metallic value of 0.5. The comparison is made with a higher metallic value of 1.0. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 480, + 198 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "rendering process. Formally,", + "bbox": [ + 89, + 301, + 282, + 316 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {\\mathbf {x} _ {0}, \\epsilon , t} \\left[ \\| \\mathbf {x} _ {0} - \\hat {\\mathbf {x}} _ {0} \\left(\\hat {\\mathbf {x}} _ {\\mathrm {r g b}}, t, \\mathbf {C}\\right) \\| ^ {2} \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 325, + 483, + 352 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{x}}_0\\left(\\mathbf{x}_t,t,\\mathbf{C}\\right)$ is the model's prediction of the original rendering output $\\mathbf{x}_0$ , conditioned on the noisy input $\\hat{\\mathbf{x}}_{\\mathrm{rgb}}$ , timestep $t$ , and conditioning attribute information $\\mathbf{C}$ . The effectiveness of the cycle-consistent constrain is demonstrated in Sec 4.3.", + "bbox": [ + 89, + 362, + 483, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Data Preparation", + "text_level": 1, + "bbox": [ + 89, + 446, + 261, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We prepare our training data with 3D data from Objaverse [6], consisting of synthetic 3D assets. We sampled 200K assets from Objaverse. For each 3D asset in Objaverse, we rendered the 2D images, and materials map (i.e., metallic, roughness, albedo) by changing the metalness and roughness (range from 0 to 1 with step 0.1). We also randomly select a 2D environment map from a pool of 20K natural scenes from a subset of LHQ-1024 [17] and format them into an RGB style for providing lighting. For each object, we have 121 pairs with different metallic, roughness and lighting. The rendered images are of resolution $1024 \\times 1024$ , and the camera poses are fixed in the front of the object. We also rendered 100 objects that were unseen during training to test the model generalization ability.", + "bbox": [ + 88, + 468, + 482, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 693, + 220, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present qualitative and quantitative analysis in Section 4.1 to demonstrate the rendering capabilities of our model. The inverse rendering results can be found in Section 4.2.", + "bbox": [ + 89, + 718, + 482, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Rendering", + "text_level": 1, + "bbox": [ + 89, + 772, + 209, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Figure 3, we demonstrate the effectiveness of our model in generalizing rendering techniques by utilizing different combinations of material attributes. Using our dual streaming pipeline, we start off by passing a reference image to perform inverse rendering. Then update the metallic, roughness value and finally re-render into an RGB image with updated materials.", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ab732e1f26d54e019fd0ff9561022ff24626dd63d264ebe6d023fad420ebda37.jpg", + "table_caption": [ + "Table 1. Quantitative results on rendering. Metrics for the prompt-only InstructPix2Pix* trained on our data, Subias et.al and our proposed method computing the PSNR, and LPIPS." + ], + "table_footnote": [], + "table_body": "
MethodsMetallicRoughness
PSNR↑LPIPS↓PSNR↑LPIPS↓
InPix2Pix* [3]24.250.103224.430.1056
Subias et.al [41]28.090.095428.130.0817
Ours30.720.076331.680.0695
Ours w/o unified27.330.093229.120.0987
Ours w/o re-render28.720.082428.930.0833
", + "bbox": [ + 527, + 131, + 893, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Roughness. As the roughness increases, the output demonstrates the elimination of specular highlights, replaced by an estimate of the base albedo. Conversely, reducing the roughness amplifies the highlights, as evidenced in the cases of the crow and the lamp in Figure 3.", + "bbox": [ + 511, + 244, + 906, + 319 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metallics. The increase of the metallic strength in both the baseball and oranges in Figure 3 reduces the significance of the base albedo and enhances the lighting effects on the surfaces. In contrast, reducing the metallic strength reverses this effect.", + "bbox": [ + 511, + 323, + 905, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Analysis. We compare our model to other material editing baselines: Subias et al. [41], Prompt-to- Prompt with Null-text Inversion (NTI) [30], and InstructPix2Pix [3] in Figure 5. We fine-tuned the InstructPix2Pix prompt-based approach with our dataset. Subias et al.'s method results in exaggerated material changes as their objective is perceptual, not physically-based material rendering. Null-text inversion and InstructPix2Pix trained on our dataset with a prompt-only approach, they either incorrectly modified the background or produces minimum changes. Our method renders the photo-realistic results, introducing the specular highlights while retaining the geometry and illumination effects. Additionally, the quantitative results with the aforementioned baselines are shown in Table 1.", + "bbox": [ + 511, + 402, + 908, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Inverse Rendering", + "text_level": 1, + "bbox": [ + 511, + 625, + 691, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) as metrics to evaluate the image quality of the albedo, relighting images, and novel view synthesis. We use Cosine Similarity for normal evaluation. Roughness and metallic estimation are evaluated using Mean Squared Error (MSE). Detailed configurations of the comparisons are discussed in the supplementary file.", + "bbox": [ + 511, + 646, + 908, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1. Lighting Estimation", + "text_level": 1, + "bbox": [ + 511, + 775, + 700, + 790 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our model with NVdiffRec [31], Gaussian-Shader [16] in terms of specular and diffuse lighting decomposition, the results can be found in Figure 6 and in Table 3. In Figure 6, we demonstrate the relighting ability of our proposed model given different environmental lighting. To achieve this, we updated the specular, diffuse, and environment maps of the given objects. The rendered outputs", + "bbox": [ + 511, + 794, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26509", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3673a81aa2384c6f115f27b4c35a1042dc7c4cba9f7688b89d686ada9e5d3998.jpg", + "image_caption": [ + "Env. Lighting" + ], + "image_footnote": [], + "bbox": [ + 125, + 89, + 189, + 267 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7f15c99c4863b8ac5c7e7938eba7d8054e2fedf288a3a7fa61182af87f9d55dd.jpg", + "image_caption": [ + "Figure 6. Relighting Comparison The relighting comparison is performed on validation objects. We first inverse render the input RGB to acquire the intrinsics, and then we updated the lighting information to get the relighting results. The leftmost column is the reference environment lighting. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 196, + 93, + 393, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/edbb858b3be3bfc71c987e9383233ab24af2bce5e5f4093672c6fadc1b717b90.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 406, + 93, + 630, + 279 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/76c4853f9385c34bded3a5e958d4a7453a92a4515a8d1623b5b53c9954200c6e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 93, + 870, + 279 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d4aaa1e446c362d21e28876424fcb9ddab1b3c4bbee7553ce8fade38494e7d78.jpg", + "table_caption": [ + "Table 2. Quantitative results on inverse rendering. For quantitative comparison, we have included both data-driven methods and optimization-based baselines. \"Ours w/o unified\" is trained without the unified framework, and \"Ours w/o constrain\" is trained without the cycle-consistent constrain term." + ], + "table_footnote": [], + "table_body": "
MethodAlbedoMetallicRoughnessNormal
PSNR↑SSIM↑LPIPS↓MSE↓MSE↓cos-simi↑
Yi et al. [46]20.930.89600.1024---
IntrinsicAny. [44]22.670.92180.0633---
Wonder3d [28]-----0.834
IntrinImgDecomp. [4]21.910.89340.0659---
IntrinImgDiff. [22]21.830.90600.06320.19200.1315-
RGB2X. [47]18.150.88290.0851--0.871
NvDiffrec [31]13.560.86950.12430.31640.29560.631
GaussianShader [16]16.550.86400.09060.34210.37140.908
Ours23.200.91820.05320.11820.10370.928
Ours w/o unified18.620.88460.08330.16320.13910.867
Ours w/o constrain21.200.89340.06020.13910.13040.922
", + "bbox": [ + 99, + 428, + 498, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "under diverse lighting scenarios show consistent adjustments, preserving the material attributes while accurately reflecting the changes in illumination. We have demonstrated a superior relighting ability compared to other baselines.", + "bbox": [ + 89, + 656, + 485, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2. Materials and geometry", + "text_level": 1, + "bbox": [ + 89, + 726, + 303, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Surface Normal. We present the surface normal comparison results in Figure 7. Our model yields the best performance in normal estimation when compared to Wonder3D [28], RGB2X [47], and Yi et al, [46].", + "bbox": [ + 89, + 744, + 483, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Albedo. For albedo estimation, we have compared 2 optimization-based methods and 6 data-driven methods in terms of PSNR, SSIM and LPIPS. Figure 4 presents the qualitative comparison for albedo estimation, and the quantitative comparison can be found in Table 2. Our method significantly outperforms other methods.", + "bbox": [ + 89, + 810, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/362a2bfcd9ae253e6c5dfeea9eadc1db197fd8c5d6e9569737adf7c70d12b173.jpg", + "table_caption": [ + "Table 3. Quantitative results on relighting. We have compared the inverse rendering results on lighting estimation in our left-out test set, \"Ours w/o unified\" is trained without the unified framework, and 'Ours w/o constrain' is trained without the cycle-consistent constrain." + ], + "table_footnote": [], + "table_body": "
MethodSpecularDiffuseRelighting
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NvDiffrec [31]14.120.88220.178215.200.88340.183321.990.88500.0834
GaussianShader [16]16.820.89120.098217.990.89690.083126.470.88260.0822
Ours22.710.93660.049823.150.96940.037330.840.90320.0763
Ours w/o unified21.950.88160.055421.920.89340.049526.950.88320.0824
Ours w/o constrain22.070.90270.051322.820.92590.040828.120.89340.0924
", + "bbox": [ + 521, + 428, + 921, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metallic and Roughness. Since existing data-driven methods do not include metallic and roughness estimation, we have compared the roughness and metallic estimation with two optimization-based inverse rendering method, including NVdiffRec [31] and GaussianShader [16], the quantitative results can be found in Table 2. More visualization cases can be found in the supplementary material.", + "bbox": [ + 511, + 556, + 906, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.3. Real-World Inversing", + "text_level": 1, + "bbox": [ + 511, + 669, + 712, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Despite being trained on synthetic data, our model is capable of performing inverse rendering in real-world cases. As shown in Figure 8, the \"Phone stand\" case has a more metallic appearance while maintaining some roughness. It is under a white-to-yellow light source thereby showing a white-to-yellow color in lighting. Our model also excels at recovering high-frequency ambient lighting from the image as specular lighting, as demonstrated in the \"kettle\" case.", + "bbox": [ + 511, + 688, + 908, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 818, + 666, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dual Stream Diffusion. To demonstrate the effectiveness of our dual stream diffusion framework, we compared results with two separate diffusion models for performing rendering and inverse rendering. Combining the two models in a sin-", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26510", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/420bf8d99724d8a4298c35a8e3ea6acd429c25fc5a98e51abdb10324020602f1.jpg", + "image_caption": [ + "Figure 7. Normal Comparison. Normal Comparison of Uni-Renderer with others methods. Best viewed in color.", + "Specular", + "Diffuse" + ], + "image_footnote": [], + "bbox": [ + 138, + 99, + 869, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "gle pipeline for both rendering and inverse can benefit each other, as shown in Table 2, 3 where the jointly trained model performance prevails. From a multi-task learning perspective, the unified pipeline benefits each other by sharing the knowledge through model weights.", + "bbox": [ + 89, + 526, + 483, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cycle-Consistent Constraint. We also conducted experiments to show the effectiveness of the cycle-consistent constraint. The constrain can mitigate the ambiguity problem by enforcing consistency between the intrinsic properties and the images, leading to improved performance. The results in Table 2, 3 demonstrated the inverse rendering quality improves with the constrain term being included.", + "bbox": [ + 89, + 604, + 483, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 724, + 209, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a method that allows photo-realistic rendering and inverse rendering in a unified framework. Our method demonstrates strong performance in real-world cases, achieving photo-realistic rendering and inverse rendering that generalizes well beyond synthetic training data. Despite this success, there are still limitations when applying our approach to certain real-world scenarios, primarily due to the domain gap between synthetic training data and real-world images. This gap can lead to challenges in accurately handling complex or unfamiliar objects and environments, where the synthetic-", + "bbox": [ + 89, + 750, + 485, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d2b6057a123f66217f55acd4e9150532d89dfad8625269247b3c3a3bb441e260.jpg", + "image_caption": [ + "Normal", + "Albedo", + "Figure 8. Real World Inversing. We present different real-world inversing cases under different lighting conditions. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 516, + 523, + 684, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to-real adaptation is less effective. To address this, future work will focus on incorporating more real-world data into our training process. By expanding our dataset with real-world examples, we aim to enhance the model's ability to bridge the synthetic-real domain gap, enabling better generalization and higher-quality rendering results across diverse real-world scenes.", + "bbox": [ + 511, + 794, + 908, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26511", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Fan Bao, Shen Nie, Kaiwen Xue, Chongxuan Li, Shi Pu, Yaole Wang, Gang Yue, Yue Cao, Hang Su, and Jun Zhu. One transformer fits all distributions in multi-modal diffusion at scale, 2023. 2, 3, 4, 5", + "[2] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence (TPAMI), 2014. 2", + "[3] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions, 2023. 6", + "[4] Chris Careaga and Yaqiz Aksoy. Intrinsic image decomposition via ordinal shading. ACM Trans. Graph., 43(1), 2023. 2, 3, 7", + "[5] Wenzheng Chen, Joey Litalien, Jun Gao, Zian Wang, Clement Fuji Tsang, Sameh Khamis, Or Litany, and Sanja Fidler. Dibr++: learning to predict lighting and material with a hybrid differentiable renderer. Advances in Neural Information Processing Systems (NeuIPS), 2021. 2", + "[6] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6", + "[7] Valentin Deschaintre, Miika Aittala, Fredo Durand, George Drettakis, and Adrien Bousseau. Single-image svbrdf capture with a rendering-aware deep network. ACM Transactions on Graphics (ToG), 2018. 2, 3", + "[8] Cass Everitt and Mark J Kilgard. Practical and robust stenciled shadow volumes for hardware-accelerated rendering. 2003. 2", + "[9] Yue Fan, Ivan Skorokhodov, Oleg Voynov, Savva Ignatyev, Evgeny Burnaev, Peter Wonka, and Yiqun Wang. Factoredneus: Reconstructing surfaces, illumination, and materials of possibly glossy objects. arXiv preprint arXiv:2305.17929, 2023. 3", + "[10] Roger Grosse, Micah K Johnson, Edward H Adelson, and William T Freeman. Ground truth dataset and baseline evaluations for intrinsic image algorithms. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2009. 2", + "[11] Yu Guo, Cameron Smith, Milos Hasan, Kalyan Sunkavalli, and Shuang Zhao. Materialgan: Reflectance capture using a generative svbrdf model. arXiv preprint arXiv:2010.00114, 2020. 2, 3", + "[12] John K Haas. A history of the unity game engine. Diss. Worcester Polytechnic Institute, 2014. 2", + "[13] Wolfgang Heidrich and Hans-Peter Seidel. Realistic, hardware-accelerated shading and lighting. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, 1999. 2", + "[14] Roland Hess. *Blender foundations: The essential guide to learning blender* 2.5. 2013. 2", + "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3" + ], + "bbox": [ + 91, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. arXiv preprint arXiv:2311.17977, 2023. 6, 7", + "[17] Kaggle user: dimensiOn. Lhq-1024 dataset. https://www.kaggle.com/datasets/dimensiOn/1024, 2023.6", + "[18] James T Kajiya. The rendering equation. In Proceedings of the 13th annual conference on Computer graphics and interactive techniques, 1986. 2, 4", + "[19] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), 2020. 2", + "[20] Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neural 3d mesh renderer. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), 2018. 2", + "[21] Mark J Kilgard and Jeff Bolz. GPU-accelerated path rendering. ACM Transactions on Graphics (TOG), 2012. 2", + "[22] Peter Kocsis, Vincent Sitzmann, and Matthias Nießner. Intrinsic image diffusion for indoor single-view material estimation, 2024. 2, 3, 7", + "[23] Jaroslav Křivánek, Christophe Chevallier, Vladimir Koylazov, Ondrej Karlík, Henrik Wann Jensen, and Thomas Ludwig. Realistic rendering in architecture and product visualization. 2018. 2", + "[24] Eric P Lafortune and Yves D Willems. Bi-directional path tracing. 1993. 2", + "[25] Michael Lewis and Jeffrey Jacobson. Game engines. Communications of the ACM, 2002. 2", + "[26] Ruofan Liang, Huiting Chen, Chunlin Li, Fan Chen, Selvakumar Panneer, and Nandita Vijaykumar. Envidr: Implicit differentiable renderer with neural environment lighting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3", + "[27] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In SIGGRAPH, 2023. 3", + "[28] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, and Wenping Wang. _Wonder3d: Single image to 3d using cross-domain diffusion_, 2023. 7", + "[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 2", + "[30] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models, 2022. 6", + "[31] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting Triangular 3D Models, Materials, and Lighting From Images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8280-8290, 2022. 6, 7" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26512", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 3", + "[33] Merlin Nimier-David, Delio Vicini, Tizian Zeltner, and Wenzel Jakob. Mitsuba 2: A retargetable forward and inverse renderer. ACM Transactions on Graphics (TOG), 2019. 2", + "[34] Matt Pharr, Wenzel Jakob, and Greg Humphreys. Physically based rendering: From theory to implementation. 2023. 2", + "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 3", + "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 5", + "[37] Prafull Sharma, Varun Jampani, Yuanzhen Li, Xuhui Jia, Dmitry Lagun, Fredo Durand, William T. Freeman, and Mark Matthews. Alchemist: Parametric control of material properties with diffusion models, 2023. 3", + "[38] Peter Shirley and R Keith Morley. Realistic ray tracing. AK Peters, Ltd., 2008. 2", + "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3", + "[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3", + "[41] J. Daniel Subias and Manuel Lagunas. In-the-wild material appearance editing using perceptual attributes, 2023. 6", + "[42] Shimon Vainer, Mark Boss, Mathias Parger, Konstantin Kutsy, Dante De Nigris, Ciara Rowles, Nicolas Perony, and Simon Donne. Collaborative control for geometry-conditioned pbr image generation, 2024. 3", + "[43] Catherine Winder and Zahra Dowlatabadi. Producing animation. 2013. 2", + "[44] Chen Xi, Peng Sida, Yang Dongchen, Liu Yuan, Pan Bowen, Lv Chengfei, and Zhou. Xiaowei. Intrinsicanything: Learning diffusion priors for inverse rendering under unknown illumination. arxiv: 2404.11593, 2024. 2, 3, 7", + "[45] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf: Neural incident light field for physically-based material estimation. In European Conference on Computer Vision (ECCV), 2022. 2, 3", + "[46] Renjiao Yi, Chenyang Zhu, and Kai Xu. Weakly-supervised single-view image relighting, 2023. 2, 3, 7", + "[47] Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloš Hasan. Rgb↔x: Image decomposition and synthesis using material- and lighting-aware diffusion models. In Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers '24, page 1-11. ACM, 2024. 2, 3, 7" + ], + "bbox": [ + 89, + 92, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[48] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "bbox": [ + 514, + 92, + 906, + 157 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26513", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Uni-Renderer_ Unifying Rendering and Inverse Rendering Via Dual Stream Diffusion/ef5cd413-b783-4606-bcb0-152603271134_model.json b/2025/Uni-Renderer_ Unifying Rendering and Inverse Rendering Via Dual Stream Diffusion/ef5cd413-b783-4606-bcb0-152603271134_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c9fa75d9781b819dc6a8a5fab1d98a47950e477a --- /dev/null +++ b/2025/Uni-Renderer_ Unifying Rendering and Inverse Rendering Via Dual Stream Diffusion/ef5cd413-b783-4606-bcb0-152603271134_model.json @@ -0,0 +1,2354 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.117, + 0.103, + 0.882, + 0.147 + ], + "angle": 0, + "content": "Uni-Renderer: Unifying Rendering and Inverse Rendering Via Dual Stream Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.175, + 0.897, + 0.23 + ], + "angle": 0, + "content": "Zhifei Chen\\(^{1,*}\\), Tianshuo Xu\\(^{1,*}\\), Wenhang Ge\\(^{1,*}\\), Leyi Wu\\(^{1}\\), Dongyu Yan\\(^{1}\\), Jing He\\(^{1}\\), Luozhou Wang\\(^{1}\\), Lu Zeng\\(^{3}\\), Shunsi Zhang\\(^{3}\\), Yingcong Chen\\(^{1,2,\\dagger}\\), \\({}^{1}\\mathrm{HKUST}(\\mathrm{GZ})\\), \\({}^{2}\\mathrm{HKUST}\\), \\({}^{3}\\mathrm{Quwan}\\)" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.232, + 0.775, + 0.248 + ], + "angle": 0, + "content": "{zchen379, txu647}@connect.hkust-gz.edu.cn; yingcongchen@ust.hk" + }, + { + "type": "image", + "bbox": [ + 0.106, + 0.275, + 0.893, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.371, + 0.892, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.557, + 0.908, + 0.655 + ], + "angle": 0, + "content": "Figure 1. Our framework, Uni-renderer, empowers the generative model to function both as a renderer and an inverse renderer by approximating the rendering equation using a data-driven approach. Given intrinsic attributes, Uni-renderer generates photo-realistic images, functioning as a renderer. When provided with a single RGB image, it effectively decomposes the intrinsic properties, functioning as an inverse renderer. Top: Uni-renderer generates smooth variations as renderer. Setting the roughness value to 1.0 results in the \"dog\" case, shown at the top, lacking specular highlights. Conversely, setting the metallic value to 1 makes the \"hat\" case appear metallic. Bottom Left: When functioning as an inverse renderer, Uni-renderer decomposes the intrinsic properties of a single RGB image. Bottom Right: Uni-renderer generates relighting results under different environment lightings." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.681, + 0.327, + 0.696 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.701, + 0.486, + 0.868 + ], + "angle": 0, + "content": "Rendering and inverse rendering are pivotal tasks in both computer vision and graphics. The rendering equation is the core of the two tasks, as an ideal conditional distribution transfer function from intrinsic properties to RGB images. Despite achieving promising results of existing rendering methods, they merely approximate the ideal estimation for a specific scene and come with a high computational cost. Additionally, the inverse conditional distribution transfer is intractable due to the inherent ambiguity. To address these challenges, we propose a data-driven method that jointly models rendering and inverse rendering as two conditional" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.682, + 0.909, + 0.865 + ], + "angle": 0, + "content": "generation tasks within a single diffusion framework. Inspired by UniDiffuser, we utilize two distinct time schedules to model both tasks, and with a tailored dual streaming module, we achieve cross-conditioning of two pre-trained diffusion models. This unified approach, named Uni-Renderer, allows the two processes to facilitate each other through a cycle-consistent constrain, mitigating ambiguity by enforcing consistency between intrinsic properties and rendered images. Combined with a meticulously prepared dataset, our method effectively decomposition of intrinsic properties and demonstrating a strong capability to recognize changes during rendering." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.877, + 0.373, + 0.903 + ], + "angle": 0, + "content": "* Equal contribution. †Corresponding Author. This project is supported by Quwan Technology." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26504" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.223, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.386 + ], + "angle": 0, + "content": "Physically based rendering is crucial in computer vision and graphics for producing photorealistic 2D images from 3D models, textures, and lighting setups. This process is essential in animation [43], and architectural visualization [23]. At its core, the rendering equation [18] models the flow of light energy within a scene, offering an ideal conditional distribution transfer function for rendering. Monte Carlo light-transport simulations [34], utilizing path tracing [24], are commonly employed to evaluate the rendering equation. However, its recursive tracing incurs significant computational demands, posing challenges for real-time implementation. Inverse rendering, which aims to deduce geometric, material, and lighting information from images, has long been studied due to its under-constrained nature [10]. This approach allows the reconstructed 3D model to be directly integrated into rendering engines, thereby playing a critical role in downstream applications such as game production [25]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.387, + 0.485, + 0.674 + ], + "angle": 0, + "content": "Recent advancements employ differentiable rendering for 3D representations [5, 45] or focus on direct decomposition of intrinsics using 2D object priors [7, 11]. These approaches use differentiable renderer to integrate intrinsic properties into RGB images, facilitating loss computation against ground-truth images for optimization. For instance, MaterialGAN [11] leverages StyleGAN2 [19] to synthesize realistic material properties using a large-scale, spatially-varying material dataset [7], while other works [4, 22, 44, 46] utilize diffusion models to achieve better results. Nevertheless, the ambiguity between geometry, materials, and lighting still hinders the effectiveness of decomposition. The mapping from observed images back to intrinsic properties is not one-to-one, leading to ambiguity and suboptimal performance. The most similar work to ours is RGB2X [47], which performs both forward and inverse rendering using two separate diffusion models. However, it fails to establish a connection between the two processes, treating them independently. Ambiguity issue still exists." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.485, + 0.825 + ], + "angle": 0, + "content": "To alleviate this issue, we propose to jointly learn the two distributions together. By integrating rendering and inverse rendering into a single framework from a multi-task learning perspective, the two processes can facilitate each other. This joint learning allows us to use the inverted intrinsic properties to perform another cycle of rendering, effectively creating a cycle-consistent constrains. This cycle rendering can mitigate the ambiguity problem by enforcing consistency between the intrinsic properties and the rendered images, leading to improved performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.485, + 0.902 + ], + "angle": 0, + "content": "In this work, we explicitly model the rendering process and inverse rendering as two conditional generation tasks. Inspired by UniDiffuser [1], we utilize two distinct time schedules to model both conditional generation tasks within a single diffusion pipeline. By cross-conditioning two pre" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.135 + ], + "angle": 0, + "content": "trained diffusion models through a dual streaming module, we achieve both rendering and inverse rendering in a unified framework." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.137, + 0.907, + 0.197 + ], + "angle": 0, + "content": "To summarize, we introduce Uni-renderer which utilizes a pre-trained diffusion model to handle both rendering and inverse rendering in a unified framework. The key contributions of our method are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.202, + 0.909, + 0.277 + ], + "angle": 0, + "content": "- We propose a data-driven method to approximate the rendering equation. By modeling rendering and inverse rendering as two conditional generation tasks, we design a unified diffusion model with a tailored dual stream module for cross-condition generation." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.277, + 0.908, + 0.337 + ], + "angle": 0, + "content": "- We render a synthetic dataset of different fine-grained material edits using 200K 3D objects with randomized intrinsic attributes by varying one of the rendering elements at a time while keeping the others constant." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.338, + 0.908, + 0.398 + ], + "angle": 0, + "content": "- Extensive experiments validate the effectiveness of our method, achieving robust disentangling of intrinsic properties and demonstrating a strong capability to recognize changes in rendering." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.202, + 0.909, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.413, + 0.655, + 0.428 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.438, + 0.634, + 0.453 + ], + "angle": 0, + "content": "2.1. Rendering" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.909, + 0.732 + ], + "angle": 0, + "content": "Rendering, the process of generating 2D images from 3D models, synthesizes raw data from a 3D scene—including geometry, materials, and lighting—using the rendering equation [18]. This task is typically performed by rendering engines, such as Blender [14] and Unity [12], which serve as the technological intermediaries converting 3D models into 2D images or videos. However, achieving photo-realistic 2D images often involves recursive ray tracing [38] or path tracing [24], techniques that closely approximate the rendering equation but are time-consuming. Although several GPU-based acceleration methods have been proposed [8, 13, 21], and existing engines have incorporated these tools, real-time rendering continues to pose significant challenges. In contrast to traditional rendering approaches, our method considers the diffusion model as an alternative renderer, which utilizes geometry, materials, and lighting as conditional cues to produce photo-realistic 2D images without the need for recursive tracing." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.743, + 0.693, + 0.759 + ], + "angle": 0, + "content": "2.2. Inverse Rendering" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Inverse rendering [2, 33], the task of decomposing an image's appearance into intrinsic properties such as geometry, material, and lighting, remains a longstanding and severely ill-posed problem in computer vision and graphics. Some approaches employ differentiable renderers on 3D representations [5, 20] to directly optimize these properties using image losses. Advancements in volume rendering, as utilized in Neural Radiance Fields (NeRFs) [29] for 3D reconstruction from multi-view 2D images, have led most" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26505" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.099, + 0.912, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.354, + 0.908, + 0.437 + ], + "angle": 0, + "content": "Figure 2. The overview of our pipeline. During training, both attribute and RGB images are input to a unified model with pre-trained VAE encoders. The timestep selector plays a crucial role by adjusting the timesteps for each branch. Specifically, it ensures that one branch (either the attribute or RGB) has a timestep of 0, while the other branch selects a timestep from \\(t \\in [0, T]\\). This mechanism allows our model to effectively learn the conditional distributions \\(q(\\mathbf{x}_0|\\mathbf{y}_0)\\), \\(q(\\mathbf{y}_0|\\mathbf{x}_0)\\) in alternating iterations. During rendering and inverse rendering, the corresponding conditions are input to the model with a timestep of 0, and the attributes/RGB images are generated through a sampled noise. (The VAE encoder and decoder are omitted for simplicity.)" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.464, + 0.485, + 0.766 + ], + "angle": 0, + "content": "methods [9, 26, 27, 45, 48] to leverage multiple images for reconstruction and subsequent estimation of materials and lighting. Additionally, they require per-object optimization and struggle to generalize across different objects, limiting their efficiency in fast estimation. MaterialGAN [11] utilizes a StyleGAN2-based model for generating intrinsic properties and a renderer for novel view synthesis under specified lighting conditions. Similarly, SIC [7] employs an encoder-decoder network for single-image-based inverse rendering, where rendering loss is used for optimization. Nonetheless, these efforts predominantly focus on planar surfaces and still necessitate a renderer to integrate predicted attributes into the final rendered image. [4, 22, 44, 46, 47] utilize diffusion models to achieve better results. However, these methods have not achieved optimal performance due to the inherent ambiguity involved in transferring distributions from images to intrinsic properties. Our method effectively mitigates the ambiguity among intrinsic properties by introducing a dual-stream diffusion pipeline and combined with a meticulously prepared dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.26, + 0.789 + ], + "angle": 0, + "content": "2.3. Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Diffusion Models [15, 32, 35, 39, 40] belong to the class of probabilistic generative models that progressively destruct data by injecting noise, and then subsequently learn to reverse this process in order to generate new samples. Alchemist [37] leverages a pre-trained diffusion model to achieve material control over a given object. [42] adopts a collaborative diffusion pipeline to decompose physically-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.909, + 0.614 + ], + "angle": 0, + "content": "based rendering (PBR) material properties from RGB. However, these works focus either on rendering or inverse rendering and fail to yield a unified model to fit in real-world usages. Alternatively, we design a dual stream diffusion model as both a renderer and an inverse renderer. During the inverse rendering stage, it takes a single RGB image as input and disentangles all the intrinsic properties. For the rendering stage, all of the intrinsic properties including material, geometry, and lighting conditions will be taken and rendered via the same model." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.667, + 0.625, + 0.684 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Given a single RGB image \\(\\mathbf{I}\\), inverse rendering aims at decomposing it into intrinsic attributes including metallic \\(m\\), roughness \\(r\\), albedo \\(\\mathbf{a}\\), surface normal \\(\\mathbf{n}\\), specular lighting \\(\\mathbf{s}\\) and diffuse lighting \\(\\mathbf{d}\\), which is formulated as \\(\\{m, r, \\mathbf{a}, \\mathbf{n}, \\mathbf{s}, \\mathbf{d}\\} = \\text{InverseRenderer}(\\mathbf{I})\\). For later reference, we denote the combination of attributes as \\(\\mathbf{C}\\). Renderer takes attributes as input and renders them into RGB image denoted as \\(\\mathbf{I} = \\text{Renderer}(\\mathbf{C})\\). We formulate a dual stream diffusion model as renderer and inverse renderer simultaneously. We start by briefly revisiting physically based rendering (PBR) and Unidiffuser [1] in Section 3.1. Next, we introduce our dual stream diffusion framework in Section 3.2. We elaborate on the data preparation in Section 3.3." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26506" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.092, + 0.905, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.43, + 0.908, + 0.486 + ], + "angle": 0, + "content": "Figure 3. We demonstrate smooth changes via rendering for different metallic and roughness strengths. The rendering is performed giving different combinations of the attributes. When the roughness value was set to 1, the cake and clock case shown in the top left are without specular highlights. When the metallic value was set to 1, the orange and baseball cases appeared to be metallic and revealing object illumination. Best viewed in color." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.511, + 0.232, + 0.525 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.483, + 0.613 + ], + "angle": 0, + "content": "Physically Based Rendering. Physically-based rendering is a computer graphics approach that seeks to render images by modeling the interaction between lights and materials in the real world. At its core, the rendering equation [18] describes the light energy flow in the scene, formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.633, + 0.483, + 0.665 + ], + "angle": 0, + "content": "\\[\nL (\\mathbf {x}, \\omega_ {o}) = \\int_ {\\Omega} f (\\mathbf {x}, \\omega_ {o}, \\omega_ {i}) L _ {i} (\\mathbf {x}, \\omega_ {i}) (\\omega_ {i}, \\mathbf {n}) d \\omega_ {i}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.671, + 0.483, + 0.792 + ], + "angle": 0, + "content": "where \\(\\omega_{o}\\) is the viewing direction of the outgoing light, \\(f\\) is the BRDF properties, \\(L_{i}\\) is the incident light of direction \\(\\omega_{i}\\) sampled from the upper hemisphere \\(\\Omega\\), and \\(\\mathbf{n}\\) is the surface normal. The rendering equation is the key to handle both rendering and inverse rendering problems. Given intrinsic properties, the rendering equation generates photo-realistic images. Given an image, a rendering equation can also be used to optimize intrinsic properties." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Unidiffuser. UniDiffuser [1], a unified diffusion model that models all distributions at the same time, injects noise \\(\\epsilon^x\\) and \\(\\epsilon^y\\) to a set of paired image-text data \\((\\mathbf{x_0},\\mathbf{y_0})\\) and generates noisy data \\(\\mathbf{x_t^x}\\) and \\(\\mathbf{y_t^y}\\), where \\(0\\leq t_{x},t_{y}\\leq T\\) represent two individual timesteps. It then trains a joint noise prediction network \\(\\epsilon_{\\theta}(\\mathbf{x_{t^{x}}},\\mathbf{y_{t^{y}}},t_{x},t_{y})\\) to predict the noise \\(\\epsilon^x\\) and \\(\\epsilon^y\\) by minimizing the mean squared error loss:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.532, + 0.907, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\underset {\\epsilon^ {\\mathbf {x}}, \\epsilon^ {\\mathbf {y}}, \\mathbf {x} _ {0}, \\mathbf {y} _ {0}} {\\mathbb {E}} \\left[ \\left\\| [ \\epsilon^ {\\mathbf {x}}, \\epsilon^ {\\mathbf {y}} ] - \\epsilon_ {\\theta} (\\mathbf {x} _ {\\mathbf {t}} ^ {\\mathbf {x}}, \\mathbf {y} _ {\\mathbf {t}} ^ {\\mathbf {y}}, t _ {x}, t _ {y}) \\| ^ {2} \\right], \\right. \\quad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.909, + 0.78 + ], + "angle": 0, + "content": "By predicting \\(\\epsilon_{\\theta}(\\mathbf{x}_{t^{\\mathbf{x}}},\\mathbf{y}_{t^{\\mathbf{y}}},t_x,t_y)\\) for any \\(t_x\\) and \\(t_y\\), UniDiffuser learns all distributions related to \\((\\mathbf{x}_0,\\mathbf{y}_0)\\). This includes all conditional distributions: \\(q(\\mathbf{x}_0|\\mathbf{y}_0)\\), \\(q(\\mathbf{y}_0|\\mathbf{x}_0)\\), and those conditioned on noisy input, for example \\(q(\\mathbf{x}_0|\\mathbf{y}_{t^{\\mathbf{y}}})\\) and \\(q(\\mathbf{y}_0|x_{t^x})\\), for \\(0\\leq t_x,t_y\\leq T\\). Learning a conditional distribution can be seen as learning a distinct task. From a multitask learning perspective, due to limited network bandwidth, learning many tasks simultaneously (i.e., fitting all distributions to a single network) may result in task competition or task conflict, ultimately leading to sub-optimal performance. For rendering and inverse rendering, we exclusively modeled the two conditional distributions \\(q(\\mathbf{x}_0|\\mathbf{y}_0)\\), \\(q(\\mathbf{y}_0|\\mathbf{x}_0)\\) to resolve the aforementioned issue and enhance the performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.659, + 0.803 + ], + "angle": 0, + "content": "3.2. Uni-Renderer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Our design incorporates a pre-trained RGB image model and a PBR model, the two were tightly linked to one another via a dual streaming technique aligning the RGB model's information with the PBR model's information. The overall framework is shown in Figure 2. Next, we will discuss each sub-module in the pipeline in detail." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26507" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.108, + 0.089, + 0.199, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.362, + 0.185, + 0.372 + ], + "angle": 0, + "content": "Ref. Image" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.099, + 0.271, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.166, + 0.27, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.234, + 0.27, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.305, + 0.268, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.361, + 0.266, + 0.371 + ], + "angle": 0, + "content": "Yi.et al" + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.099, + 0.358, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.167, + 0.342, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.234, + 0.359, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.301, + 0.305, + 0.356, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.361, + 0.341, + 0.37 + ], + "angle": 0, + "content": "RGB2X" + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.099, + 0.447, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.402, + 0.167, + 0.434, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.233, + 0.45, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.305, + 0.443, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.372, + 0.361, + 0.461, + 0.371 + ], + "angle": 0, + "content": "IntrinImgDecomp" + }, + { + "type": "image", + "bbox": [ + 0.476, + 0.099, + 0.534, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.168, + 0.521, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.476, + 0.234, + 0.537, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.305, + 0.535, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.47, + 0.361, + 0.551, + 0.371 + ], + "angle": 0, + "content": "IntrinImgDiff." + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.099, + 0.623, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.579, + 0.168, + 0.608, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.233, + 0.627, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.305, + 0.623, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.566, + 0.36, + 0.632, + 0.37 + ], + "angle": 0, + "content": "GaussShader" + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.099, + 0.705, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.168, + 0.691, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.233, + 0.708, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.305, + 0.705, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.36, + 0.707, + 0.37 + ], + "angle": 0, + "content": "NVdiffRec" + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.099, + 0.788, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.743, + 0.168, + 0.772, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.233, + 0.79, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.734, + 0.305, + 0.785, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.361, + 0.771, + 0.37 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.817, + 0.099, + 0.874, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.829, + 0.168, + 0.859, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.815, + 0.233, + 0.875, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.816, + 0.304, + 0.869, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.812, + 0.36, + 0.883, + 0.369 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.4, + 0.907, + 0.43 + ], + "angle": 0, + "content": "Figure 4. Albedo comparison. Albedo Comparison of Uni-Renderer with baseline methods. We compared 4 learning-based methods and 2 optimization-based methods. Among all, Uni-renderer yields the most realistic results. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.456, + 0.485, + 0.682 + ], + "angle": 0, + "content": "Latent Preparation. Recent advancements have benefited greatly from the pre-trained VAE [36], which downsample an RGB image into lower dimensional latent. We used separate VAE for encoding \\([m, r]\\), albedo a, surface normal n and environment lighting s, d before we channel-wisely concatenate them and feed them into the model. It is worth noting that attributes like metallic and roughness, often come with a scalar value, and cannot be directly encoded. Thus, we generate additional masks to binarize the two values into gray-scale images. Instead of having two separate VAE for m and r, We formed channel triplets \\(a_{\\mathrm{material}}\\), consisting of \\([m, r, \\mathbf{m}]\\) and process those with the RGB VAE, where m is the mask This design allows us to reduce the bandwidth overhead within the diffusion model and further improve generation quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.484, + 0.822 + ], + "angle": 0, + "content": "Conditional Distributions Modeling. To achieve rendering and inverse rendering within a single model, we borrowed ideas from [1], and adopted two different timesteps \\(t_{\\mathrm{attributes}}\\), \\(t_{\\mathrm{RGB}}\\) from three timesteps choices \\(\\{0, t, T\\}\\) to train a joint noise prediction network. It is worth noting that since our unified model only requires modeling two conditional distributions, we ruled out the choices where both \\(t_{\\mathrm{attributes}}\\) and \\(t_{\\mathrm{RGB}}\\) take on \\(t\\) or \\(T\\) and make sure there is always a timestep equal to 0. Formally," + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.834, + 0.484, + 0.875 + ], + "angle": 0, + "content": "\\[\n\\left(t _ {\\text {a t t r i b u t e s}}, t _ {\\mathrm {R G B}}\\right) = \\left\\{ \\begin{array}{l l} (0, \\tilde {t}) & \\text {d u r i n g} \\\\ (\\tilde {t}, 0) & \\text {i n v e r s e} \\end{array} , \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.887, + 0.155, + 0.899 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.474, + 0.907, + 0.515 + ], + "angle": 0, + "content": "\\[\n\\tilde {t} = \\left\\{ \\begin{array}{l l} t & \\text {w i t h p r o b a b i l i t y} p \\\\ T & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.527, + 0.907, + 0.557 + ], + "angle": 0, + "content": "The pseudo-code for generating timesteps can be found in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.563, + 0.909, + 0.789 + ], + "angle": 0, + "content": "Dual Stream Diffusion. We designed a dual diffusion network to facilitate the unified model and adopt the \\(\\mathbf{x}_0\\)-prediction in the diffusion process. Training: During training, as shown in Figure 2, the upper branch of the model dealt with the intrinsic attributes only, leaving only RGB for the lower branch. The timestep selection module will determine if a rendering/inverse rendering iteration has occurred at a probability \\(p\\). During a rendering iteration, the clean attributes \\(\\mathbf{C}\\) will be used as conditions to denoise the noisy RGB image \\(\\mathbf{I}\\), the same will be applied to the inverse rendering iteration except that the noisy attributes are being fed to the model. Inference: At the inference stage, we provided attributes as conditions to the model and gradually denoise the RGB, and vice versa. The effectiveness of the dual stream framework is demonstrated in Sec 4.3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Cycle-Consistent Constrain. To alleviate the inherent ambiguity problem, we introduce a cycle-consistent constraint within our unified framework, where its incorporation is straightforward with our pipeline. During training, we use the model's predicted inverse results to perform an additional cycle of rendering. The re-rendered outputs are then utilized in the loss calculation to further optimize the inverse" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26508" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.482, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.221, + 0.485, + 0.276 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison. Rendering Comparison of Uni-Renderer with baseline methods. The base images are with a metallic value of 0.5. The comparison is made with a higher metallic value of 1.0. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.303, + 0.284, + 0.318 + ], + "angle": 0, + "content": "rendering process. Formally," + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.327, + 0.484, + 0.353 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {\\mathbf {x} _ {0}, \\epsilon , t} \\left[ \\| \\mathbf {x} _ {0} - \\hat {\\mathbf {x}} _ {0} \\left(\\hat {\\mathbf {x}} _ {\\mathrm {r g b}}, t, \\mathbf {C}\\right) \\| ^ {2} \\right], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.363, + 0.484, + 0.437 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{x}}_0\\left(\\mathbf{x}_t,t,\\mathbf{C}\\right)\\) is the model's prediction of the original rendering output \\(\\mathbf{x}_0\\), conditioned on the noisy input \\(\\hat{\\mathbf{x}}_{\\mathrm{rgb}}\\), timestep \\(t\\), and conditioning attribute information \\(\\mathbf{C}\\). The effectiveness of the cycle-consistent constrain is demonstrated in Sec 4.3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.447, + 0.262, + 0.463 + ], + "angle": 0, + "content": "3.3. Data Preparation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.469, + 0.483, + 0.681 + ], + "angle": 0, + "content": "We prepare our training data with 3D data from Objaverse [6], consisting of synthetic 3D assets. We sampled 200K assets from Objaverse. For each 3D asset in Objaverse, we rendered the 2D images, and materials map (i.e., metallic, roughness, albedo) by changing the metalness and roughness (range from 0 to 1 with step 0.1). We also randomly select a 2D environment map from a pool of 20K natural scenes from a subset of LHQ-1024 [17] and format them into an RGB style for providing lighting. For each object, we have 121 pairs with different metallic, roughness and lighting. The rendered images are of resolution \\(1024 \\times 1024\\), and the camera poses are fixed in the front of the object. We also rendered 100 objects that were unseen during training to test the model generalization ability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.694, + 0.222, + 0.71 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.764 + ], + "angle": 0, + "content": "We present qualitative and quantitative analysis in Section 4.1 to demonstrate the rendering capabilities of our model. The inverse rendering results can be found in Section 4.2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.21, + 0.789 + ], + "angle": 0, + "content": "4.1. Rendering" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In Figure 3, we demonstrate the effectiveness of our model in generalizing rendering techniques by utilizing different combinations of material attributes. Using our dual streaming pipeline, we start off by passing a reference image to perform inverse rendering. Then update the metallic, roughness value and finally re-render into an RGB image with updated materials." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.079, + 0.907, + 0.121 + ], + "angle": 0, + "content": "Table 1. Quantitative results on rendering. Metrics for the prompt-only InstructPix2Pix* trained on our data, Subias et.al and our proposed method computing the PSNR, and LPIPS." + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.132, + 0.895, + 0.228 + ], + "angle": 0, + "content": "
MethodsMetallicRoughness
PSNR↑LPIPS↓PSNR↑LPIPS↓
InPix2Pix* [3]24.250.103224.430.1056
Subias et.al [41]28.090.095428.130.0817
Ours30.720.076331.680.0695
Ours w/o unified27.330.093229.120.0987
Ours w/o re-render28.720.082428.930.0833
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.245, + 0.907, + 0.32 + ], + "angle": 0, + "content": "Roughness. As the roughness increases, the output demonstrates the elimination of specular highlights, replaced by an estimate of the base albedo. Conversely, reducing the roughness amplifies the highlights, as evidenced in the cases of the crow and the lamp in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.324, + 0.906, + 0.399 + ], + "angle": 0, + "content": "Metallics. The increase of the metallic strength in both the baseball and oranges in Figure 3 reduces the significance of the base albedo and enhances the lighting effects on the surfaces. In contrast, reducing the metallic strength reverses this effect." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.909, + 0.614 + ], + "angle": 0, + "content": "Analysis. We compare our model to other material editing baselines: Subias et al. [41], Prompt-to- Prompt with Null-text Inversion (NTI) [30], and InstructPix2Pix [3] in Figure 5. We fine-tuned the InstructPix2Pix prompt-based approach with our dataset. Subias et al.'s method results in exaggerated material changes as their objective is perceptual, not physically-based material rendering. Null-text inversion and InstructPix2Pix trained on our dataset with a prompt-only approach, they either incorrectly modified the background or produces minimum changes. Our method renders the photo-realistic results, introducing the specular highlights while retaining the geometry and illumination effects. Additionally, the quantitative results with the aforementioned baselines are shown in Table 1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.625, + 0.692, + 0.641 + ], + "angle": 0, + "content": "4.2. Inverse Rendering" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.647, + 0.909, + 0.769 + ], + "angle": 0, + "content": "We use Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) as metrics to evaluate the image quality of the albedo, relighting images, and novel view synthesis. We use Cosine Similarity for normal evaluation. Roughness and metallic estimation are evaluated using Mean Squared Error (MSE). Detailed configurations of the comparisons are discussed in the supplementary file." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.776, + 0.702, + 0.791 + ], + "angle": 0, + "content": "4.2.1. Lighting Estimation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We compare our model with NVdiffRec [31], Gaussian-Shader [16] in terms of specular and diffuse lighting decomposition, the results can be found in Figure 6 and in Table 3. In Figure 6, we demonstrate the relighting ability of our proposed model given different environmental lighting. To achieve this, we updated the specular, diffuse, and environment maps of the given objects. The rendered outputs" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26509" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.09, + 0.191, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.27, + 0.185, + 0.279 + ], + "angle": 0, + "content": "Env. Lighting" + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.094, + 0.394, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.094, + 0.631, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.094, + 0.871, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.292, + 0.908, + 0.335 + ], + "angle": 0, + "content": "Figure 6. Relighting Comparison The relighting comparison is performed on validation objects. We first inverse render the input RGB to acquire the intrinsics, and then we updated the lighting information to get the relighting results. The leftmost column is the reference environment lighting. Best viewed in color." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.359, + 0.486, + 0.428 + ], + "angle": 0, + "content": "Table 2. Quantitative results on inverse rendering. For quantitative comparison, we have included both data-driven methods and optimization-based baselines. \"Ours w/o unified\" is trained without the unified framework, and \"Ours w/o constrain\" is trained without the cycle-consistent constrain term." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.429, + 0.499, + 0.63 + ], + "angle": 0, + "content": "
MethodAlbedoMetallicRoughnessNormal
PSNR↑SSIM↑LPIPS↓MSE↓MSE↓cos-simi↑
Yi et al. [46]20.930.89600.1024---
IntrinsicAny. [44]22.670.92180.0633---
Wonder3d [28]-----0.834
IntrinImgDecomp. [4]21.910.89340.0659---
IntrinImgDiff. [22]21.830.90600.06320.19200.1315-
RGB2X. [47]18.150.88290.0851--0.871
NvDiffrec [31]13.560.86950.12430.31640.29560.631
GaussianShader [16]16.550.86400.09060.34210.37140.908
Ours23.200.91820.05320.11820.10370.928
Ours w/o unified18.620.88460.08330.16320.13910.867
Ours w/o constrain21.200.89340.06020.13910.13040.922
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.486, + 0.718 + ], + "angle": 0, + "content": "under diverse lighting scenarios show consistent adjustments, preserving the material attributes while accurately reflecting the changes in illumination. We have demonstrated a superior relighting ability compared to other baselines." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.305, + 0.743 + ], + "angle": 0, + "content": "4.2.2. Materials and geometry" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.746, + 0.485, + 0.807 + ], + "angle": 0, + "content": "Surface Normal. We present the surface normal comparison results in Figure 7. Our model yields the best performance in normal estimation when compared to Wonder3D [28], RGB2X [47], and Yi et al, [46]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Albedo. For albedo estimation, we have compared 2 optimization-based methods and 6 data-driven methods in terms of PSNR, SSIM and LPIPS. Figure 4 presents the qualitative comparison for albedo estimation, and the quantitative comparison can be found in Table 2. Our method significantly outperforms other methods." + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.359, + 0.908, + 0.427 + ], + "angle": 0, + "content": "Table 3. Quantitative results on relighting. We have compared the inverse rendering results on lighting estimation in our left-out test set, \"Ours w/o unified\" is trained without the unified framework, and 'Ours w/o constrain' is trained without the cycle-consistent constrain." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.429, + 0.922, + 0.535 + ], + "angle": 0, + "content": "
MethodSpecularDiffuseRelighting
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NvDiffrec [31]14.120.88220.178215.200.88340.183321.990.88500.0834
GaussianShader [16]16.820.89120.098217.990.89690.083126.470.88260.0822
Ours22.710.93660.049823.150.96940.037330.840.90320.0763
Ours w/o unified21.950.88160.055421.920.89340.049526.950.88320.0824
Ours w/o constrain22.070.90270.051322.820.92590.040828.120.89340.0924
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.558, + 0.908, + 0.664 + ], + "angle": 0, + "content": "Metallic and Roughness. Since existing data-driven methods do not include metallic and roughness estimation, we have compared the roughness and metallic estimation with two optimization-based inverse rendering method, including NVdiffRec [31] and GaussianShader [16], the quantitative results can be found in Table 2. More visualization cases can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.67, + 0.714, + 0.685 + ], + "angle": 0, + "content": "4.2.3. Real-World Inversing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.81 + ], + "angle": 0, + "content": "Despite being trained on synthetic data, our model is capable of performing inverse rendering in real-world cases. As shown in Figure 8, the \"Phone stand\" case has a more metallic appearance while maintaining some roughness. It is under a white-to-yellow light source thereby showing a white-to-yellow color in lighting. Our model also excels at recovering high-frequency ambient lighting from the image as specular lighting, as demonstrated in the \"kettle\" case." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.667, + 0.835 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Dual Stream Diffusion. To demonstrate the effectiveness of our dual stream diffusion framework, we compared results with two separate diffusion models for performing rendering and inverse rendering. Combining the two models in a sin-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26510" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.1, + 0.87, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.158, + 0.486, + 0.839, + 0.502 + ], + "angle": 0, + "content": "Figure 7. Normal Comparison. Normal Comparison of Uni-Renderer with others methods. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.527, + 0.484, + 0.603 + ], + "angle": 0, + "content": "gle pipeline for both rendering and inverse can benefit each other, as shown in Table 2, 3 where the jointly trained model performance prevails. From a multi-task learning perspective, the unified pipeline benefits each other by sharing the knowledge through model weights." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.606, + 0.484, + 0.713 + ], + "angle": 0, + "content": "Cycle-Consistent Constraint. We also conducted experiments to show the effectiveness of the cycle-consistent constraint. The constrain can mitigate the ambiguity problem by enforcing consistency between the intrinsic properties and the images, leading to improved performance. The results in Table 2, 3 demonstrated the inverse rendering quality improves with the constrain term being included." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.725, + 0.21, + 0.74 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.486, + 0.903 + ], + "angle": 0, + "content": "We present a method that allows photo-realistic rendering and inverse rendering in a unified framework. Our method demonstrates strong performance in real-world cases, achieving photo-realistic rendering and inverse rendering that generalizes well beyond synthetic training data. Despite this success, there are still limitations when applying our approach to certain real-world scenarios, primarily due to the domain gap between synthetic training data and real-world images. This gap can lead to challenges in accurately handling complex or unfamiliar objects and environments, where the synthetic-" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.525, + 0.685, + 0.709 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.526, + 0.728, + 0.535 + ], + "angle": 0, + "content": "Normal" + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.526, + 0.78, + 0.535 + ], + "angle": 0, + "content": "Albedo" + }, + { + "type": "image_caption", + "bbox": [ + 0.804, + 0.526, + 0.842, + 0.535 + ], + "angle": 0, + "content": "Specular" + }, + { + "type": "image_caption", + "bbox": [ + 0.857, + 0.526, + 0.892, + 0.535 + ], + "angle": 0, + "content": "Diffuse" + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.72, + 0.907, + 0.761 + ], + "angle": 0, + "content": "Figure 8. Real World Inversing. We present different real-world inversing cases under different lighting conditions. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.9 + ], + "angle": 0, + "content": "to-real adaptation is less effective. To address this, future work will focus on incorporating more real-world data into our training process. By expanding our dataset with real-world examples, we aim to enhance the model's ability to bridge the synthetic-real domain gap, enabling better generalization and higher-quality rendering results across diverse real-world scenes." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26511" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Fan Bao, Shen Nie, Kaiwen Xue, Chongxuan Li, Shi Pu, Yaole Wang, Gang Yue, Yue Cao, Hang Su, and Jun Zhu. One transformer fits all distributions in multi-modal diffusion at scale, 2023. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.171, + 0.484, + 0.213 + ], + "angle": 0, + "content": "[2] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence (TPAMI), 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.213, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[3] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.484, + 0.297 + ], + "angle": 0, + "content": "[4] Chris Careaga and Yaqiz Aksoy. Intrinsic image decomposition via ordinal shading. ACM Trans. Graph., 43(1), 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.484, + 0.368 + ], + "angle": 0, + "content": "[5] Wenzheng Chen, Joey Litalien, Jun Gao, Zian Wang, Clement Fuji Tsang, Sameh Khamis, Or Litany, and Sanja Fidler. Dibr++: learning to predict lighting and material with a hybrid differentiable renderer. Advances in Neural Information Processing Systems (NeuIPS), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.484, + 0.45 + ], + "angle": 0, + "content": "[6] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.451, + 0.484, + 0.507 + ], + "angle": 0, + "content": "[7] Valentin Deschaintre, Miika Aittala, Fredo Durand, George Drettakis, and Adrien Bousseau. Single-image svbrdf capture with a rendering-aware deep network. ACM Transactions on Graphics (ToG), 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.508, + 0.484, + 0.548 + ], + "angle": 0, + "content": "[8] Cass Everitt and Mark J Kilgard. Practical and robust stenciled shadow volumes for hardware-accelerated rendering. 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.55, + 0.484, + 0.618 + ], + "angle": 0, + "content": "[9] Yue Fan, Ivan Skorokhodov, Oleg Voynov, Savva Ignatyev, Evgeny Burnaev, Peter Wonka, and Yiqun Wang. Factoredneus: Reconstructing surfaces, illumination, and materials of possibly glossy objects. arXiv preprint arXiv:2305.17929, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.62, + 0.484, + 0.689 + ], + "angle": 0, + "content": "[10] Roger Grosse, Micah K Johnson, Edward H Adelson, and William T Freeman. Ground truth dataset and baseline evaluations for intrinsic image algorithms. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.484, + 0.745 + ], + "angle": 0, + "content": "[11] Yu Guo, Cameron Smith, Milos Hasan, Kalyan Sunkavalli, and Shuang Zhao. Materialgan: Reflectance capture using a generative svbrdf model. arXiv preprint arXiv:2010.00114, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[12] John K Haas. A history of the unity game engine. Diss. Worcester Polytechnic Institute, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.484, + 0.83 + ], + "angle": 0, + "content": "[13] Wolfgang Heidrich and Hans-Peter Seidel. Realistic, hardware-accelerated shading and lighting. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[14] Roland Hess. *Blender foundations: The essential guide to learning blender* 2.5. 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.148 + ], + "angle": 0, + "content": "[16] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. arXiv preprint arXiv:2311.17977, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.908, + 0.189 + ], + "angle": 0, + "content": "[17] Kaggle user: dimensiOn. Lhq-1024 dataset. https://www.kaggle.com/datasets/dimensiOn/1024, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.19, + 0.907, + 0.232 + ], + "angle": 0, + "content": "[18] James T Kajiya. The rendering equation. In Proceedings of the 13th annual conference on Computer graphics and interactive techniques, 1986. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.908, + 0.3 + ], + "angle": 0, + "content": "[19] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.301, + 0.907, + 0.343 + ], + "angle": 0, + "content": "[20] Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neural 3d mesh renderer. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.907, + 0.371 + ], + "angle": 0, + "content": "[21] Mark J Kilgard and Jeff Bolz. GPU-accelerated path rendering. ACM Transactions on Graphics (TOG), 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.907, + 0.412 + ], + "angle": 0, + "content": "[22] Peter Kocsis, Vincent Sitzmann, and Matthias Nießner. Intrinsic image diffusion for indoor single-view material estimation, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.907, + 0.467 + ], + "angle": 0, + "content": "[23] Jaroslav Křivánek, Christophe Chevallier, Vladimir Koylazov, Ondrej Karlík, Henrik Wann Jensen, and Thomas Ludwig. Realistic rendering in architecture and product visualization. 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.906, + 0.497 + ], + "angle": 0, + "content": "[24] Eric P Lafortune and Yves D Willems. Bi-directional path tracing. 1993. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.907, + 0.525 + ], + "angle": 0, + "content": "[25] Michael Lewis and Jeffrey Jacobson. Game engines. Communications of the ACM, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.907, + 0.594 + ], + "angle": 0, + "content": "[26] Ruofan Liang, Huiting Chen, Chunlin Li, Fan Chen, Selvakumar Panneer, and Nandita Vijaykumar. Envidr: Implicit differentiable renderer with neural environment lighting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.907, + 0.65 + ], + "angle": 0, + "content": "[27] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In SIGGRAPH, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.907, + 0.718 + ], + "angle": 0, + "content": "[28] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, and Wenping Wang. _Wonder3d: Single image to 3d using cross-domain diffusion_, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.907, + 0.775 + ], + "angle": 0, + "content": "[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.906, + 0.817 + ], + "angle": 0, + "content": "[30] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[31] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting Triangular 3D Models, Materials, and Lighting From Images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8280-8290, 2022. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26512" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.486, + 0.146 + ], + "angle": 0, + "content": "[32] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.15, + 0.485, + 0.192 + ], + "angle": 0, + "content": "[33] Merlin Nimier-David, Delio Vicini, Tizian Zeltner, and Wenzel Jakob. Mitsuba 2: A retargetable forward and inverse renderer. ACM Transactions on Graphics (TOG), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.194, + 0.484, + 0.223 + ], + "angle": 0, + "content": "[34] Matt Pharr, Wenzel Jakob, and Greg Humphreys. Physically based rendering: From theory to implementation. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.224, + 0.485, + 0.293 + ], + "angle": 0, + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.296, + 0.484, + 0.337 + ], + "angle": 0, + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.34, + 0.484, + 0.394 + ], + "angle": 0, + "content": "[37] Prafull Sharma, Varun Jampani, Yuanzhen Li, Xuhui Jia, Dmitry Lagun, Fredo Durand, William T. Freeman, and Mark Matthews. Alchemist: Parametric control of material properties with diffusion models, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.397, + 0.484, + 0.424 + ], + "angle": 0, + "content": "[38] Peter Shirley and R Keith Morley. Realistic ray tracing. AK Peters, Ltd., 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.426, + 0.484, + 0.466 + ], + "angle": 0, + "content": "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.47, + 0.484, + 0.525 + ], + "angle": 0, + "content": "[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.528, + 0.482, + 0.556 + ], + "angle": 0, + "content": "[41] J. Daniel Subias and Manuel Lagunas. In-the-wild material appearance editing using perceptual attributes, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.558, + 0.484, + 0.613 + ], + "angle": 0, + "content": "[42] Shimon Vainer, Mark Boss, Mathias Parger, Konstantin Kutsy, Dante De Nigris, Ciara Rowles, Nicolas Perony, and Simon Donne. Collaborative control for geometry-conditioned pbr image generation, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.616, + 0.484, + 0.642 + ], + "angle": 0, + "content": "[43] Catherine Winder and Zahra Dowlatabadi. Producing animation. 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.646, + 0.484, + 0.699 + ], + "angle": 0, + "content": "[44] Chen Xi, Peng Sida, Yang Dongchen, Liu Yuan, Pan Bowen, Lv Chengfei, and Zhou. Xiaowei. Intrinsicanything: Learning diffusion priors for inverse rendering under unknown illumination. arxiv: 2404.11593, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.703, + 0.484, + 0.77 + ], + "angle": 0, + "content": "[45] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf: Neural incident light field for physically-based material estimation. In European Conference on Computer Vision (ECCV), 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[46] Renjiao Yi, Chenyang Zhu, and Kai Xu. Weakly-supervised single-view image relighting, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.805, + 0.484, + 0.899 + ], + "angle": 0, + "content": "[47] Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloš Hasan. Rgb↔x: Image decomposition and synthesis using material- and lighting-aware diffusion models. In Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers '24, page 1-11. ACM, 2024. 2, 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.486, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.908, + 0.159 + ], + "angle": 0, + "content": "[48] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26513" + } + ] +] \ No newline at end of file diff --git a/2025/Uni-Renderer_ Unifying Rendering and Inverse Rendering Via Dual Stream Diffusion/ef5cd413-b783-4606-bcb0-152603271134_origin.pdf b/2025/Uni-Renderer_ Unifying Rendering and Inverse Rendering Via Dual Stream Diffusion/ef5cd413-b783-4606-bcb0-152603271134_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..802e500de23fca21f907212a6820f860e19fcec9 --- /dev/null +++ b/2025/Uni-Renderer_ Unifying Rendering and Inverse Rendering Via Dual Stream Diffusion/ef5cd413-b783-4606-bcb0-152603271134_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:72dd823466b678af3c7f1ba37aa3cf098b62f4dc073867ecf0812ca6ebec9ebf +size 1764188 diff --git a/2025/Uni-Renderer_ Unifying Rendering and Inverse Rendering Via Dual Stream Diffusion/full.md b/2025/Uni-Renderer_ Unifying Rendering and Inverse Rendering Via Dual Stream Diffusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e717697fce6953c919446d23a5eed295752b2af3 --- /dev/null +++ b/2025/Uni-Renderer_ Unifying Rendering and Inverse Rendering Via Dual Stream Diffusion/full.md @@ -0,0 +1,331 @@ +# Uni-Renderer: Unifying Rendering and Inverse Rendering Via Dual Stream Diffusion + +Zhifei Chen $^{1,*}$ , Tianshuo Xu $^{1,*}$ , Wenhang Ge $^{1,*}$ , Leyi Wu $^{1}$ , Dongyu Yan $^{1}$ , Jing He $^{1}$ , Luozhou Wang $^{1}$ , Lu Zeng $^{3}$ , Shunsi Zhang $^{3}$ , Yingcong Chen $^{1,2,\dagger}$ , ${}^{1}\mathrm{HKUST}(\mathrm{GZ})$ , ${}^{2}\mathrm{HKUST}$ , ${}^{3}\mathrm{Quwan}$ + +{zchen379, txu647}@connect.hkust-gz.edu.cn; yingcongchen@ust.hk + +![](images/33c954137fe5c44b190944d552c103fdb2941765da13a0c5c50d4c0642914194.jpg) + +![](images/71eb58cb509cdc30ace89b4995357217a5bb94d9376531626877cf5d150bf5be.jpg) +Figure 1. Our framework, Uni-renderer, empowers the generative model to function both as a renderer and an inverse renderer by approximating the rendering equation using a data-driven approach. Given intrinsic attributes, Uni-renderer generates photo-realistic images, functioning as a renderer. When provided with a single RGB image, it effectively decomposes the intrinsic properties, functioning as an inverse renderer. Top: Uni-renderer generates smooth variations as renderer. Setting the roughness value to 1.0 results in the "dog" case, shown at the top, lacking specular highlights. Conversely, setting the metallic value to 1 makes the "hat" case appear metallic. Bottom Left: When functioning as an inverse renderer, Uni-renderer decomposes the intrinsic properties of a single RGB image. Bottom Right: Uni-renderer generates relighting results under different environment lightings. + +# Abstract + +Rendering and inverse rendering are pivotal tasks in both computer vision and graphics. The rendering equation is the core of the two tasks, as an ideal conditional distribution transfer function from intrinsic properties to RGB images. Despite achieving promising results of existing rendering methods, they merely approximate the ideal estimation for a specific scene and come with a high computational cost. Additionally, the inverse conditional distribution transfer is intractable due to the inherent ambiguity. To address these challenges, we propose a data-driven method that jointly models rendering and inverse rendering as two conditional + +generation tasks within a single diffusion framework. Inspired by UniDiffuser, we utilize two distinct time schedules to model both tasks, and with a tailored dual streaming module, we achieve cross-conditioning of two pre-trained diffusion models. This unified approach, named Uni-Renderer, allows the two processes to facilitate each other through a cycle-consistent constrain, mitigating ambiguity by enforcing consistency between intrinsic properties and rendered images. Combined with a meticulously prepared dataset, our method effectively decomposition of intrinsic properties and demonstrating a strong capability to recognize changes during rendering. + +# 1. Introduction + +Physically based rendering is crucial in computer vision and graphics for producing photorealistic 2D images from 3D models, textures, and lighting setups. This process is essential in animation [43], and architectural visualization [23]. At its core, the rendering equation [18] models the flow of light energy within a scene, offering an ideal conditional distribution transfer function for rendering. Monte Carlo light-transport simulations [34], utilizing path tracing [24], are commonly employed to evaluate the rendering equation. However, its recursive tracing incurs significant computational demands, posing challenges for real-time implementation. Inverse rendering, which aims to deduce geometric, material, and lighting information from images, has long been studied due to its under-constrained nature [10]. This approach allows the reconstructed 3D model to be directly integrated into rendering engines, thereby playing a critical role in downstream applications such as game production [25]. + +Recent advancements employ differentiable rendering for 3D representations [5, 45] or focus on direct decomposition of intrinsics using 2D object priors [7, 11]. These approaches use differentiable renderer to integrate intrinsic properties into RGB images, facilitating loss computation against ground-truth images for optimization. For instance, MaterialGAN [11] leverages StyleGAN2 [19] to synthesize realistic material properties using a large-scale, spatially-varying material dataset [7], while other works [4, 22, 44, 46] utilize diffusion models to achieve better results. Nevertheless, the ambiguity between geometry, materials, and lighting still hinders the effectiveness of decomposition. The mapping from observed images back to intrinsic properties is not one-to-one, leading to ambiguity and suboptimal performance. The most similar work to ours is RGB2X [47], which performs both forward and inverse rendering using two separate diffusion models. However, it fails to establish a connection between the two processes, treating them independently. Ambiguity issue still exists. + +To alleviate this issue, we propose to jointly learn the two distributions together. By integrating rendering and inverse rendering into a single framework from a multi-task learning perspective, the two processes can facilitate each other. This joint learning allows us to use the inverted intrinsic properties to perform another cycle of rendering, effectively creating a cycle-consistent constrains. This cycle rendering can mitigate the ambiguity problem by enforcing consistency between the intrinsic properties and the rendered images, leading to improved performance. + +In this work, we explicitly model the rendering process and inverse rendering as two conditional generation tasks. Inspired by UniDiffuser [1], we utilize two distinct time schedules to model both conditional generation tasks within a single diffusion pipeline. By cross-conditioning two pre + +trained diffusion models through a dual streaming module, we achieve both rendering and inverse rendering in a unified framework. + +To summarize, we introduce Uni-renderer which utilizes a pre-trained diffusion model to handle both rendering and inverse rendering in a unified framework. The key contributions of our method are as follows: + +- We propose a data-driven method to approximate the rendering equation. By modeling rendering and inverse rendering as two conditional generation tasks, we design a unified diffusion model with a tailored dual stream module for cross-condition generation. +- We render a synthetic dataset of different fine-grained material edits using 200K 3D objects with randomized intrinsic attributes by varying one of the rendering elements at a time while keeping the others constant. +- Extensive experiments validate the effectiveness of our method, achieving robust disentangling of intrinsic properties and demonstrating a strong capability to recognize changes in rendering. + +# 2. Related Work + +# 2.1. Rendering + +Rendering, the process of generating 2D images from 3D models, synthesizes raw data from a 3D scene—including geometry, materials, and lighting—using the rendering equation [18]. This task is typically performed by rendering engines, such as Blender [14] and Unity [12], which serve as the technological intermediaries converting 3D models into 2D images or videos. However, achieving photo-realistic 2D images often involves recursive ray tracing [38] or path tracing [24], techniques that closely approximate the rendering equation but are time-consuming. Although several GPU-based acceleration methods have been proposed [8, 13, 21], and existing engines have incorporated these tools, real-time rendering continues to pose significant challenges. In contrast to traditional rendering approaches, our method considers the diffusion model as an alternative renderer, which utilizes geometry, materials, and lighting as conditional cues to produce photo-realistic 2D images without the need for recursive tracing. + +# 2.2. Inverse Rendering + +Inverse rendering [2, 33], the task of decomposing an image's appearance into intrinsic properties such as geometry, material, and lighting, remains a longstanding and severely ill-posed problem in computer vision and graphics. Some approaches employ differentiable renderers on 3D representations [5, 20] to directly optimize these properties using image losses. Advancements in volume rendering, as utilized in Neural Radiance Fields (NeRFs) [29] for 3D reconstruction from multi-view 2D images, have led most + +![](images/b8e693381e9e32c516785c82b1b60055ce426559187e9a6dc017121808dd9454.jpg) +Figure 2. The overview of our pipeline. During training, both attribute and RGB images are input to a unified model with pre-trained VAE encoders. The timestep selector plays a crucial role by adjusting the timesteps for each branch. Specifically, it ensures that one branch (either the attribute or RGB) has a timestep of 0, while the other branch selects a timestep from $t \in [0, T]$ . This mechanism allows our model to effectively learn the conditional distributions $q(\mathbf{x}_0|\mathbf{y}_0)$ , $q(\mathbf{y}_0|\mathbf{x}_0)$ in alternating iterations. During rendering and inverse rendering, the corresponding conditions are input to the model with a timestep of 0, and the attributes/RGB images are generated through a sampled noise. (The VAE encoder and decoder are omitted for simplicity.) + +methods [9, 26, 27, 45, 48] to leverage multiple images for reconstruction and subsequent estimation of materials and lighting. Additionally, they require per-object optimization and struggle to generalize across different objects, limiting their efficiency in fast estimation. MaterialGAN [11] utilizes a StyleGAN2-based model for generating intrinsic properties and a renderer for novel view synthesis under specified lighting conditions. Similarly, SIC [7] employs an encoder-decoder network for single-image-based inverse rendering, where rendering loss is used for optimization. Nonetheless, these efforts predominantly focus on planar surfaces and still necessitate a renderer to integrate predicted attributes into the final rendered image. [4, 22, 44, 46, 47] utilize diffusion models to achieve better results. However, these methods have not achieved optimal performance due to the inherent ambiguity involved in transferring distributions from images to intrinsic properties. Our method effectively mitigates the ambiguity among intrinsic properties by introducing a dual-stream diffusion pipeline and combined with a meticulously prepared dataset. + +# 2.3. Diffusion Models + +Diffusion Models [15, 32, 35, 39, 40] belong to the class of probabilistic generative models that progressively destruct data by injecting noise, and then subsequently learn to reverse this process in order to generate new samples. Alchemist [37] leverages a pre-trained diffusion model to achieve material control over a given object. [42] adopts a collaborative diffusion pipeline to decompose physically- + +based rendering (PBR) material properties from RGB. However, these works focus either on rendering or inverse rendering and fail to yield a unified model to fit in real-world usages. Alternatively, we design a dual stream diffusion model as both a renderer and an inverse renderer. During the inverse rendering stage, it takes a single RGB image as input and disentangles all the intrinsic properties. For the rendering stage, all of the intrinsic properties including material, geometry, and lighting conditions will be taken and rendered via the same model. + +# 3. Approach + +Given a single RGB image $\mathbf{I}$ , inverse rendering aims at decomposing it into intrinsic attributes including metallic $m$ , roughness $r$ , albedo $\mathbf{a}$ , surface normal $\mathbf{n}$ , specular lighting $\mathbf{s}$ and diffuse lighting $\mathbf{d}$ , which is formulated as $\{m, r, \mathbf{a}, \mathbf{n}, \mathbf{s}, \mathbf{d}\} = \text{InverseRenderer}(\mathbf{I})$ . For later reference, we denote the combination of attributes as $\mathbf{C}$ . Renderer takes attributes as input and renders them into RGB image denoted as $\mathbf{I} = \text{Renderer}(\mathbf{C})$ . We formulate a dual stream diffusion model as renderer and inverse renderer simultaneously. We start by briefly revisiting physically based rendering (PBR) and Unidiffuser [1] in Section 3.1. Next, we introduce our dual stream diffusion framework in Section 3.2. We elaborate on the data preparation in Section 3.3. + +![](images/d8995d7a23d05fff4967f6a9ed356893286239cab760d1fe6fa3605d22f9644d.jpg) +Figure 3. We demonstrate smooth changes via rendering for different metallic and roughness strengths. The rendering is performed giving different combinations of the attributes. When the roughness value was set to 1, the cake and clock case shown in the top left are without specular highlights. When the metallic value was set to 1, the orange and baseball cases appeared to be metallic and revealing object illumination. Best viewed in color. + +# 3.1. Preliminaries + +Physically Based Rendering. Physically-based rendering is a computer graphics approach that seeks to render images by modeling the interaction between lights and materials in the real world. At its core, the rendering equation [18] describes the light energy flow in the scene, formulated as + +$$ +L (\mathbf {x}, \omega_ {o}) = \int_ {\Omega} f (\mathbf {x}, \omega_ {o}, \omega_ {i}) L _ {i} (\mathbf {x}, \omega_ {i}) (\omega_ {i}, \mathbf {n}) d \omega_ {i}, \tag {1} +$$ + +where $\omega_{o}$ is the viewing direction of the outgoing light, $f$ is the BRDF properties, $L_{i}$ is the incident light of direction $\omega_{i}$ sampled from the upper hemisphere $\Omega$ , and $\mathbf{n}$ is the surface normal. The rendering equation is the key to handle both rendering and inverse rendering problems. Given intrinsic properties, the rendering equation generates photo-realistic images. Given an image, a rendering equation can also be used to optimize intrinsic properties. + +Unidiffuser. UniDiffuser [1], a unified diffusion model that models all distributions at the same time, injects noise $\epsilon^x$ and $\epsilon^y$ to a set of paired image-text data $(\mathbf{x_0},\mathbf{y_0})$ and generates noisy data $\mathbf{x_t^x}$ and $\mathbf{y_t^y}$ , where $0\leq t_{x},t_{y}\leq T$ represent two individual timesteps. It then trains a joint noise prediction network $\epsilon_{\theta}(\mathbf{x_{t^{x}}},\mathbf{y_{t^{y}}},t_{x},t_{y})$ to predict the noise $\epsilon^x$ and $\epsilon^y$ by minimizing the mean squared error loss: + +$$ +\underset {\epsilon^ {\mathbf {x}}, \epsilon^ {\mathbf {y}}, \mathbf {x} _ {0}, \mathbf {y} _ {0}} {\mathbb {E}} \left[ \left\| [ \epsilon^ {\mathbf {x}}, \epsilon^ {\mathbf {y}} ] - \epsilon_ {\theta} (\mathbf {x} _ {\mathbf {t}} ^ {\mathbf {x}}, \mathbf {y} _ {\mathbf {t}} ^ {\mathbf {y}}, t _ {x}, t _ {y}) \| ^ {2} \right], \right. \quad (2) +$$ + +By predicting $\epsilon_{\theta}(\mathbf{x}_{t^{\mathbf{x}}},\mathbf{y}_{t^{\mathbf{y}}},t_x,t_y)$ for any $t_x$ and $t_y$ , UniDiffuser learns all distributions related to $(\mathbf{x}_0,\mathbf{y}_0)$ . This includes all conditional distributions: $q(\mathbf{x}_0|\mathbf{y}_0)$ , $q(\mathbf{y}_0|\mathbf{x}_0)$ , and those conditioned on noisy input, for example $q(\mathbf{x}_0|\mathbf{y}_{t^{\mathbf{y}}})$ and $q(\mathbf{y}_0|x_{t^x})$ , for $0\leq t_x,t_y\leq T$ . Learning a conditional distribution can be seen as learning a distinct task. From a multitask learning perspective, due to limited network bandwidth, learning many tasks simultaneously (i.e., fitting all distributions to a single network) may result in task competition or task conflict, ultimately leading to sub-optimal performance. For rendering and inverse rendering, we exclusively modeled the two conditional distributions $q(\mathbf{x}_0|\mathbf{y}_0)$ , $q(\mathbf{y}_0|\mathbf{x}_0)$ to resolve the aforementioned issue and enhance the performance. + +# 3.2. Uni-Renderer + +Our design incorporates a pre-trained RGB image model and a PBR model, the two were tightly linked to one another via a dual streaming technique aligning the RGB model's information with the PBR model's information. The overall framework is shown in Figure 2. Next, we will discuss each sub-module in the pipeline in detail. + +![](images/fce9c77b5886093e0d36e348354bc70aa86f3c35ec6db4f9e16f6816a3d5542b.jpg) +Ref. Image + +![](images/d076f9198be8b4fd20c611e1ba18c19467acdec790e26e1b3c5aa7b9fb512e14.jpg) + +![](images/95eb62e66ba26b82f097ebd061a696f92d1ff9a8f4026ce296bb747c867e0f19.jpg) + +![](images/8efbd67ed6f43f507ee577ac9bd41c4cb3eae40eccf25c3abe822882b63d5bf8.jpg) + +![](images/2acebc2d319bec8331123d806b54053663f6005974c26bede52fde47f9e93282.jpg) +Yi.et al + +![](images/de07a5ca93e501833ef95e4fb995a3d6ff503b3f40ed2f36ba9c2cd735ede5f6.jpg) + +![](images/629169e4536c96fc11fc6098d814fb892a8af092f07faa244fe50dfe9f73fc1e.jpg) + +![](images/4cd34e57ead8a77ff649f5074f60f0395caa22b8076eb4c3b4c7289a7bb99606.jpg) + +![](images/fe36d1ac3d0b44e56ff68bb5f4217ceabdf5c90cfd3f56473cb6abff4f915403.jpg) +RGB2X + +![](images/c3c9a53ac0a33108adfe8c3ac65a63c41422d6619f5448796289e4a695092468.jpg) + +![](images/025b74baa17b9e1770c2fcc396c73d00897e3fe3d16b1a4688ad86bac5cd821f.jpg) + +![](images/c48bc7781711d78f4c01d712a2e001983b7a1b2339fc259590611f42ee14f63b.jpg) + +![](images/acab94daa601a6f9eacb05012c31a294b2fab02c9700d184ea2a62f3ab6b0355.jpg) +IntrinImgDecomp +Figure 4. Albedo comparison. Albedo Comparison of Uni-Renderer with baseline methods. We compared 4 learning-based methods and 2 optimization-based methods. Among all, Uni-renderer yields the most realistic results. Best viewed in color. + +![](images/6070618d1b4ee747ec4277038b8adfe4cfd9eedce46f7db7835fb350a9533b92.jpg) + +![](images/4dd6ce48e766c294e072dfd0d4d3fa7a774561bb2b3fbcf0b6519be056a5bf84.jpg) + +![](images/8efc27d847f0ad3a0f2d0f51d31f61ddca6a8140c258071baf80758f0fcc1d2a.jpg) + +![](images/d0c1cd7ea4ec1954b8871c591869e7915a8e6a7c53d6b69758b8c2d15711414d.jpg) +IntrinImgDiff. + +![](images/fadbb50a2b5bc704872812686a38bd24a3e42b934772aa85a52500db989a6ea7.jpg) + +![](images/deddfa13121f4d58715c8fee9367d76b20b5de98fb3ea07cc21c59821ca33574.jpg) + +![](images/064c89d599ce4dc1967cc22b0ff7c920af7677827df67f3375e51a44efce3ba8.jpg) + +![](images/4dd84127bce98846052708e9441ee80cdb5319003450c3533f1b95ede475f947.jpg) +GaussShader + +![](images/84be300370926bf4794e720243080b41556bf15ec0c2bb3b97cdf07034927eff.jpg) + +![](images/ba724ebcabf3a2cf4040441e20f8b586ebf28b4d3e7c874d2c6cdc98460d1e76.jpg) + +![](images/77514a80d79ebc0478f20d91a56c558565816321419609f6e57908aeea4ec3a2.jpg) + +![](images/5f80d967e15a91c65476f2bdb2cba51031e30fb2e0f53fcd787ed2f0e98808eb.jpg) +NVdiffRec + +![](images/29b54965e165b782974368cc94761be3e02b00d854a83b2f785dc8a6cf90251b.jpg) + +![](images/3cb57a5a4e34ed4a5af02f323f1a0f141e60fe8dba8ecf44e3ff6de0fbedcbd6.jpg) + +![](images/4344b182ac230dadee907d23f61656dab4ce1f552047722d0a8d2e6a3de2a0b9.jpg) + +![](images/2fb697f4a063b8da59c503650be4cece06832ddc10cadc4bfcc6f8ef305c45f3.jpg) +Ours + +![](images/9080c5ef2ccf5f2659382b804e24618a5d9950722d11b01a39dc55c3e5be3a07.jpg) + +![](images/e9372a335f015cd024f794c2cfcb26e892e6a99e0e2513e01c579e1b95acb84f.jpg) + +![](images/ac965cae4efccbc788eec2d061f14b237f631973b8784b59fcd37a24f23222ce.jpg) + +![](images/25700b7a6a24c0d40c0d8c0399f8449af6f02a435858fe234c8f8d70575e8aa9.jpg) +Ground Truth + +Latent Preparation. Recent advancements have benefited greatly from the pre-trained VAE [36], which downsample an RGB image into lower dimensional latent. We used separate VAE for encoding $[m, r]$ , albedo a, surface normal n and environment lighting s, d before we channel-wisely concatenate them and feed them into the model. It is worth noting that attributes like metallic and roughness, often come with a scalar value, and cannot be directly encoded. Thus, we generate additional masks to binarize the two values into gray-scale images. Instead of having two separate VAE for m and r, We formed channel triplets $a_{\mathrm{material}}$ , consisting of $[m, r, \mathbf{m}]$ and process those with the RGB VAE, where m is the mask This design allows us to reduce the bandwidth overhead within the diffusion model and further improve generation quality. + +Conditional Distributions Modeling. To achieve rendering and inverse rendering within a single model, we borrowed ideas from [1], and adopted two different timesteps $t_{\mathrm{attributes}}$ , $t_{\mathrm{RGB}}$ from three timesteps choices $\{0, t, T\}$ to train a joint noise prediction network. It is worth noting that since our unified model only requires modeling two conditional distributions, we ruled out the choices where both $t_{\mathrm{attributes}}$ and $t_{\mathrm{RGB}}$ take on $t$ or $T$ and make sure there is always a timestep equal to 0. Formally, + +$$ +\left(t _ {\text {a t t r i b u t e s}}, t _ {\mathrm {R G B}}\right) = \left\{ \begin{array}{l l} (0, \tilde {t}) & \text {d u r i n g} \\ (\tilde {t}, 0) & \text {i n v e r s e} \end{array} , \right. \tag {3} +$$ + +where + +$$ +\tilde {t} = \left\{ \begin{array}{l l} t & \text {w i t h p r o b a b i l i t y} p \\ T & \text {o t h e r w i s e} \end{array} . \right. \tag {4} +$$ + +The pseudo-code for generating timesteps can be found in supplementary material. + +Dual Stream Diffusion. We designed a dual diffusion network to facilitate the unified model and adopt the $\mathbf{x}_0$ -prediction in the diffusion process. Training: During training, as shown in Figure 2, the upper branch of the model dealt with the intrinsic attributes only, leaving only RGB for the lower branch. The timestep selection module will determine if a rendering/inverse rendering iteration has occurred at a probability $p$ . During a rendering iteration, the clean attributes $\mathbf{C}$ will be used as conditions to denoise the noisy RGB image $\mathbf{I}$ , the same will be applied to the inverse rendering iteration except that the noisy attributes are being fed to the model. Inference: At the inference stage, we provided attributes as conditions to the model and gradually denoise the RGB, and vice versa. The effectiveness of the dual stream framework is demonstrated in Sec 4.3. + +Cycle-Consistent Constrain. To alleviate the inherent ambiguity problem, we introduce a cycle-consistent constraint within our unified framework, where its incorporation is straightforward with our pipeline. During training, we use the model's predicted inverse results to perform an additional cycle of rendering. The re-rendered outputs are then utilized in the loss calculation to further optimize the inverse + +![](images/a92650899f318613a3665f18c14a5dbb7f0abae4ca24dde9626215b732189dc1.jpg) +Figure 5. Qualitative comparison. Rendering Comparison of Uni-Renderer with baseline methods. The base images are with a metallic value of 0.5. The comparison is made with a higher metallic value of 1.0. Best viewed in color. + +rendering process. Formally, + +$$ +\mathcal {L} = \mathbb {E} _ {\mathbf {x} _ {0}, \epsilon , t} \left[ \| \mathbf {x} _ {0} - \hat {\mathbf {x}} _ {0} \left(\hat {\mathbf {x}} _ {\mathrm {r g b}}, t, \mathbf {C}\right) \| ^ {2} \right], \tag {5} +$$ + +where $\hat{\mathbf{x}}_0\left(\mathbf{x}_t,t,\mathbf{C}\right)$ is the model's prediction of the original rendering output $\mathbf{x}_0$ , conditioned on the noisy input $\hat{\mathbf{x}}_{\mathrm{rgb}}$ , timestep $t$ , and conditioning attribute information $\mathbf{C}$ . The effectiveness of the cycle-consistent constrain is demonstrated in Sec 4.3. + +# 3.3. Data Preparation + +We prepare our training data with 3D data from Objaverse [6], consisting of synthetic 3D assets. We sampled 200K assets from Objaverse. For each 3D asset in Objaverse, we rendered the 2D images, and materials map (i.e., metallic, roughness, albedo) by changing the metalness and roughness (range from 0 to 1 with step 0.1). We also randomly select a 2D environment map from a pool of 20K natural scenes from a subset of LHQ-1024 [17] and format them into an RGB style for providing lighting. For each object, we have 121 pairs with different metallic, roughness and lighting. The rendered images are of resolution $1024 \times 1024$ , and the camera poses are fixed in the front of the object. We also rendered 100 objects that were unseen during training to test the model generalization ability. + +# 4. Experiments + +We present qualitative and quantitative analysis in Section 4.1 to demonstrate the rendering capabilities of our model. The inverse rendering results can be found in Section 4.2. + +# 4.1. Rendering + +In Figure 3, we demonstrate the effectiveness of our model in generalizing rendering techniques by utilizing different combinations of material attributes. Using our dual streaming pipeline, we start off by passing a reference image to perform inverse rendering. Then update the metallic, roughness value and finally re-render into an RGB image with updated materials. + +Table 1. Quantitative results on rendering. Metrics for the prompt-only InstructPix2Pix* trained on our data, Subias et.al and our proposed method computing the PSNR, and LPIPS. + +
MethodsMetallicRoughness
PSNR↑LPIPS↓PSNR↑LPIPS↓
InPix2Pix* [3]24.250.103224.430.1056
Subias et.al [41]28.090.095428.130.0817
Ours30.720.076331.680.0695
Ours w/o unified27.330.093229.120.0987
Ours w/o re-render28.720.082428.930.0833
+ +Roughness. As the roughness increases, the output demonstrates the elimination of specular highlights, replaced by an estimate of the base albedo. Conversely, reducing the roughness amplifies the highlights, as evidenced in the cases of the crow and the lamp in Figure 3. + +Metallics. The increase of the metallic strength in both the baseball and oranges in Figure 3 reduces the significance of the base albedo and enhances the lighting effects on the surfaces. In contrast, reducing the metallic strength reverses this effect. + +Analysis. We compare our model to other material editing baselines: Subias et al. [41], Prompt-to- Prompt with Null-text Inversion (NTI) [30], and InstructPix2Pix [3] in Figure 5. We fine-tuned the InstructPix2Pix prompt-based approach with our dataset. Subias et al.'s method results in exaggerated material changes as their objective is perceptual, not physically-based material rendering. Null-text inversion and InstructPix2Pix trained on our dataset with a prompt-only approach, they either incorrectly modified the background or produces minimum changes. Our method renders the photo-realistic results, introducing the specular highlights while retaining the geometry and illumination effects. Additionally, the quantitative results with the aforementioned baselines are shown in Table 1. + +# 4.2. Inverse Rendering + +We use Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) as metrics to evaluate the image quality of the albedo, relighting images, and novel view synthesis. We use Cosine Similarity for normal evaluation. Roughness and metallic estimation are evaluated using Mean Squared Error (MSE). Detailed configurations of the comparisons are discussed in the supplementary file. + +# 4.2.1. Lighting Estimation + +We compare our model with NVdiffRec [31], Gaussian-Shader [16] in terms of specular and diffuse lighting decomposition, the results can be found in Figure 6 and in Table 3. In Figure 6, we demonstrate the relighting ability of our proposed model given different environmental lighting. To achieve this, we updated the specular, diffuse, and environment maps of the given objects. The rendered outputs + +![](images/3673a81aa2384c6f115f27b4c35a1042dc7c4cba9f7688b89d686ada9e5d3998.jpg) +Env. Lighting + +![](images/7f15c99c4863b8ac5c7e7938eba7d8054e2fedf288a3a7fa61182af87f9d55dd.jpg) +Figure 6. Relighting Comparison The relighting comparison is performed on validation objects. We first inverse render the input RGB to acquire the intrinsics, and then we updated the lighting information to get the relighting results. The leftmost column is the reference environment lighting. Best viewed in color. + +![](images/edbb858b3be3bfc71c987e9383233ab24af2bce5e5f4093672c6fadc1b717b90.jpg) + +![](images/76c4853f9385c34bded3a5e958d4a7453a92a4515a8d1623b5b53c9954200c6e.jpg) + +Table 2. Quantitative results on inverse rendering. For quantitative comparison, we have included both data-driven methods and optimization-based baselines. "Ours w/o unified" is trained without the unified framework, and "Ours w/o constrain" is trained without the cycle-consistent constrain term. + +
MethodAlbedoMetallicRoughnessNormal
PSNR↑SSIM↑LPIPS↓MSE↓MSE↓cos-simi↑
Yi et al. [46]20.930.89600.1024---
IntrinsicAny. [44]22.670.92180.0633---
Wonder3d [28]-----0.834
IntrinImgDecomp. [4]21.910.89340.0659---
IntrinImgDiff. [22]21.830.90600.06320.19200.1315-
RGB2X. [47]18.150.88290.0851--0.871
NvDiffrec [31]13.560.86950.12430.31640.29560.631
GaussianShader [16]16.550.86400.09060.34210.37140.908
Ours23.200.91820.05320.11820.10370.928
Ours w/o unified18.620.88460.08330.16320.13910.867
Ours w/o constrain21.200.89340.06020.13910.13040.922
+ +under diverse lighting scenarios show consistent adjustments, preserving the material attributes while accurately reflecting the changes in illumination. We have demonstrated a superior relighting ability compared to other baselines. + +# 4.2.2. Materials and geometry + +Surface Normal. We present the surface normal comparison results in Figure 7. Our model yields the best performance in normal estimation when compared to Wonder3D [28], RGB2X [47], and Yi et al, [46]. + +Albedo. For albedo estimation, we have compared 2 optimization-based methods and 6 data-driven methods in terms of PSNR, SSIM and LPIPS. Figure 4 presents the qualitative comparison for albedo estimation, and the quantitative comparison can be found in Table 2. Our method significantly outperforms other methods. + +Table 3. Quantitative results on relighting. We have compared the inverse rendering results on lighting estimation in our left-out test set, "Ours w/o unified" is trained without the unified framework, and 'Ours w/o constrain' is trained without the cycle-consistent constrain. + +
MethodSpecularDiffuseRelighting
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NvDiffrec [31]14.120.88220.178215.200.88340.183321.990.88500.0834
GaussianShader [16]16.820.89120.098217.990.89690.083126.470.88260.0822
Ours22.710.93660.049823.150.96940.037330.840.90320.0763
Ours w/o unified21.950.88160.055421.920.89340.049526.950.88320.0824
Ours w/o constrain22.070.90270.051322.820.92590.040828.120.89340.0924
+ +Metallic and Roughness. Since existing data-driven methods do not include metallic and roughness estimation, we have compared the roughness and metallic estimation with two optimization-based inverse rendering method, including NVdiffRec [31] and GaussianShader [16], the quantitative results can be found in Table 2. More visualization cases can be found in the supplementary material. + +# 4.2.3. Real-World Inversing + +Despite being trained on synthetic data, our model is capable of performing inverse rendering in real-world cases. As shown in Figure 8, the "Phone stand" case has a more metallic appearance while maintaining some roughness. It is under a white-to-yellow light source thereby showing a white-to-yellow color in lighting. Our model also excels at recovering high-frequency ambient lighting from the image as specular lighting, as demonstrated in the "kettle" case. + +# 4.3. Ablation Study + +Dual Stream Diffusion. To demonstrate the effectiveness of our dual stream diffusion framework, we compared results with two separate diffusion models for performing rendering and inverse rendering. Combining the two models in a sin- + +![](images/420bf8d99724d8a4298c35a8e3ea6acd429c25fc5a98e51abdb10324020602f1.jpg) +Figure 7. Normal Comparison. Normal Comparison of Uni-Renderer with others methods. Best viewed in color. +Specular +Diffuse + +gle pipeline for both rendering and inverse can benefit each other, as shown in Table 2, 3 where the jointly trained model performance prevails. From a multi-task learning perspective, the unified pipeline benefits each other by sharing the knowledge through model weights. + +Cycle-Consistent Constraint. We also conducted experiments to show the effectiveness of the cycle-consistent constraint. The constrain can mitigate the ambiguity problem by enforcing consistency between the intrinsic properties and the images, leading to improved performance. The results in Table 2, 3 demonstrated the inverse rendering quality improves with the constrain term being included. + +# 5. Conclusion + +We present a method that allows photo-realistic rendering and inverse rendering in a unified framework. Our method demonstrates strong performance in real-world cases, achieving photo-realistic rendering and inverse rendering that generalizes well beyond synthetic training data. Despite this success, there are still limitations when applying our approach to certain real-world scenarios, primarily due to the domain gap between synthetic training data and real-world images. This gap can lead to challenges in accurately handling complex or unfamiliar objects and environments, where the synthetic- + +![](images/d2b6057a123f66217f55acd4e9150532d89dfad8625269247b3c3a3bb441e260.jpg) +Normal +Albedo +Figure 8. Real World Inversing. We present different real-world inversing cases under different lighting conditions. Best viewed in color. + +to-real adaptation is less effective. To address this, future work will focus on incorporating more real-world data into our training process. By expanding our dataset with real-world examples, we aim to enhance the model's ability to bridge the synthetic-real domain gap, enabling better generalization and higher-quality rendering results across diverse real-world scenes. + +# References + +[1] Fan Bao, Shen Nie, Kaiwen Xue, Chongxuan Li, Shi Pu, Yaole Wang, Gang Yue, Yue Cao, Hang Su, and Jun Zhu. One transformer fits all distributions in multi-modal diffusion at scale, 2023. 2, 3, 4, 5 +[2] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence (TPAMI), 2014. 2 +[3] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions, 2023. 6 +[4] Chris Careaga and Yaqiz Aksoy. Intrinsic image decomposition via ordinal shading. ACM Trans. Graph., 43(1), 2023. 2, 3, 7 +[5] Wenzheng Chen, Joey Litalien, Jun Gao, Zian Wang, Clement Fuji Tsang, Sameh Khamis, Or Litany, and Sanja Fidler. Dibr++: learning to predict lighting and material with a hybrid differentiable renderer. Advances in Neural Information Processing Systems (NeuIPS), 2021. 2 +[6] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6 +[7] Valentin Deschaintre, Miika Aittala, Fredo Durand, George Drettakis, and Adrien Bousseau. Single-image svbrdf capture with a rendering-aware deep network. ACM Transactions on Graphics (ToG), 2018. 2, 3 +[8] Cass Everitt and Mark J Kilgard. Practical and robust stenciled shadow volumes for hardware-accelerated rendering. 2003. 2 +[9] Yue Fan, Ivan Skorokhodov, Oleg Voynov, Savva Ignatyev, Evgeny Burnaev, Peter Wonka, and Yiqun Wang. Factoredneus: Reconstructing surfaces, illumination, and materials of possibly glossy objects. arXiv preprint arXiv:2305.17929, 2023. 3 +[10] Roger Grosse, Micah K Johnson, Edward H Adelson, and William T Freeman. Ground truth dataset and baseline evaluations for intrinsic image algorithms. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2009. 2 +[11] Yu Guo, Cameron Smith, Milos Hasan, Kalyan Sunkavalli, and Shuang Zhao. Materialgan: Reflectance capture using a generative svbrdf model. arXiv preprint arXiv:2010.00114, 2020. 2, 3 +[12] John K Haas. A history of the unity game engine. Diss. Worcester Polytechnic Institute, 2014. 2 +[13] Wolfgang Heidrich and Hans-Peter Seidel. Realistic, hardware-accelerated shading and lighting. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, 1999. 2 +[14] Roland Hess. *Blender foundations: The essential guide to learning blender* 2.5. 2013. 2 +[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3 + +[16] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. arXiv preprint arXiv:2311.17977, 2023. 6, 7 +[17] Kaggle user: dimensiOn. Lhq-1024 dataset. https://www.kaggle.com/datasets/dimensiOn/1024, 2023.6 +[18] James T Kajiya. The rendering equation. In Proceedings of the 13th annual conference on Computer graphics and interactive techniques, 1986. 2, 4 +[19] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), 2020. 2 +[20] Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neural 3d mesh renderer. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), 2018. 2 +[21] Mark J Kilgard and Jeff Bolz. GPU-accelerated path rendering. ACM Transactions on Graphics (TOG), 2012. 2 +[22] Peter Kocsis, Vincent Sitzmann, and Matthias Nießner. Intrinsic image diffusion for indoor single-view material estimation, 2024. 2, 3, 7 +[23] Jaroslav Křivánek, Christophe Chevallier, Vladimir Koylazov, Ondrej Karlík, Henrik Wann Jensen, and Thomas Ludwig. Realistic rendering in architecture and product visualization. 2018. 2 +[24] Eric P Lafortune and Yves D Willems. Bi-directional path tracing. 1993. 2 +[25] Michael Lewis and Jeffrey Jacobson. Game engines. Communications of the ACM, 2002. 2 +[26] Ruofan Liang, Huiting Chen, Chunlin Li, Fan Chen, Selvakumar Panneer, and Nandita Vijaykumar. Envidr: Implicit differentiable renderer with neural environment lighting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3 +[27] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In SIGGRAPH, 2023. 3 +[28] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, and Wenping Wang. _Wonder3d: Single image to 3d using cross-domain diffusion_, 2023. 7 +[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 2 +[30] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models, 2022. 6 +[31] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting Triangular 3D Models, Materials, and Lighting From Images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8280-8290, 2022. 6, 7 + +[32] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 3 +[33] Merlin Nimier-David, Delio Vicini, Tizian Zeltner, and Wenzel Jakob. Mitsuba 2: A retargetable forward and inverse renderer. ACM Transactions on Graphics (TOG), 2019. 2 +[34] Matt Pharr, Wenzel Jakob, and Greg Humphreys. Physically based rendering: From theory to implementation. 2023. 2 +[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 3 +[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 5 +[37] Prafull Sharma, Varun Jampani, Yuanzhen Li, Xuhui Jia, Dmitry Lagun, Fredo Durand, William T. Freeman, and Mark Matthews. Alchemist: Parametric control of material properties with diffusion models, 2023. 3 +[38] Peter Shirley and R Keith Morley. Realistic ray tracing. AK Peters, Ltd., 2008. 2 +[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3 +[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3 +[41] J. Daniel Subias and Manuel Lagunas. In-the-wild material appearance editing using perceptual attributes, 2023. 6 +[42] Shimon Vainer, Mark Boss, Mathias Parger, Konstantin Kutsy, Dante De Nigris, Ciara Rowles, Nicolas Perony, and Simon Donne. Collaborative control for geometry-conditioned pbr image generation, 2024. 3 +[43] Catherine Winder and Zahra Dowlatabadi. Producing animation. 2013. 2 +[44] Chen Xi, Peng Sida, Yang Dongchen, Liu Yuan, Pan Bowen, Lv Chengfei, and Zhou. Xiaowei. Intrinsicanything: Learning diffusion priors for inverse rendering under unknown illumination. arxiv: 2404.11593, 2024. 2, 3, 7 +[45] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf: Neural incident light field for physically-based material estimation. In European Conference on Computer Vision (ECCV), 2022. 2, 3 +[46] Renjiao Yi, Chenyang Zhu, and Kai Xu. Weakly-supervised single-view image relighting, 2023. 2, 3, 7 +[47] Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloš Hasan. Rgb↔x: Image decomposition and synthesis using material- and lighting-aware diffusion models. In Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers '24, page 1-11. ACM, 2024. 2, 3, 7 + +[48] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 \ No newline at end of file diff --git a/2025/Uni-Renderer_ Unifying Rendering and Inverse Rendering Via Dual Stream Diffusion/images.zip b/2025/Uni-Renderer_ Unifying Rendering and Inverse Rendering Via Dual Stream Diffusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..30b6eee7d50b6c49b3fe4d02c34d620ade5c2336 --- /dev/null +++ b/2025/Uni-Renderer_ Unifying Rendering and Inverse Rendering Via Dual Stream Diffusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c3151d40527d775d83ebd5640bf88d3ea2860394a8e156767d76f5e2b708606 +size 705379 diff --git a/2025/Uni-Renderer_ Unifying Rendering and Inverse Rendering Via Dual Stream Diffusion/layout.json b/2025/Uni-Renderer_ Unifying Rendering and Inverse Rendering Via Dual Stream Diffusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dae6add1695e07c3296cdaa806a6e39b43c69556 --- /dev/null +++ b/2025/Uni-Renderer_ Unifying Rendering and Inverse Rendering Via Dual Stream Diffusion/layout.json @@ -0,0 +1,9038 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 71, + 81, + 539, + 116 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 81, + 539, + 116 + ], + "spans": [ + { + "bbox": [ + 71, + 81, + 539, + 116 + ], + "type": "text", + "content": "Uni-Renderer: Unifying Rendering and Inverse Rendering Via Dual Stream Diffusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 138, + 548, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 138, + 548, + 182 + ], + "spans": [ + { + "bbox": [ + 61, + 138, + 548, + 182 + ], + "type": "text", + "content": "Zhifei Chen" + }, + { + "bbox": [ + 61, + 138, + 548, + 182 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 61, + 138, + 548, + 182 + ], + "type": "text", + "content": ", Tianshuo Xu" + }, + { + "bbox": [ + 61, + 138, + 548, + 182 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 61, + 138, + 548, + 182 + ], + "type": "text", + "content": ", Wenhang Ge" + }, + { + "bbox": [ + 61, + 138, + 548, + 182 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 61, + 138, + 548, + 182 + ], + "type": "text", + "content": ", Leyi Wu" + }, + { + "bbox": [ + 61, + 138, + 548, + 182 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 138, + 548, + 182 + ], + "type": "text", + "content": ", Dongyu Yan" + }, + { + "bbox": [ + 61, + 138, + 548, + 182 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 138, + 548, + 182 + ], + "type": "text", + "content": ", Jing He" + }, + { + "bbox": [ + 61, + 138, + 548, + 182 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 138, + 548, + 182 + ], + "type": "text", + "content": ", Luozhou Wang" + }, + { + "bbox": [ + 61, + 138, + 548, + 182 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 138, + 548, + 182 + ], + "type": "text", + "content": ", Lu Zeng" + }, + { + "bbox": [ + 61, + 138, + 548, + 182 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 138, + 548, + 182 + ], + "type": "text", + "content": ", Shunsi Zhang" + }, + { + "bbox": [ + 61, + 138, + 548, + 182 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 138, + 548, + 182 + ], + "type": "text", + "content": ", Yingcong Chen" + }, + { + "bbox": [ + 61, + 138, + 548, + 182 + ], + "type": "inline_equation", + "content": "^{1,2,\\dagger}" + }, + { + "bbox": [ + 61, + 138, + 548, + 182 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 61, + 138, + 548, + 182 + ], + "type": "inline_equation", + "content": "{}^{1}\\mathrm{HKUST}(\\mathrm{GZ})" + }, + { + "bbox": [ + 61, + 138, + 548, + 182 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 61, + 138, + 548, + 182 + ], + "type": "inline_equation", + "content": "{}^{2}\\mathrm{HKUST}" + }, + { + "bbox": [ + 61, + 138, + 548, + 182 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 61, + 138, + 548, + 182 + ], + "type": "inline_equation", + "content": "{}^{3}\\mathrm{Quwan}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 183, + 474, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 183, + 474, + 196 + ], + "spans": [ + { + "bbox": [ + 133, + 183, + 474, + 196 + ], + "type": "text", + "content": "{zchen379, txu647}@connect.hkust-gz.edu.cn; yingcongchen@ust.hk" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 64, + 217, + 546, + 289 + ], + "blocks": [ + { + "bbox": [ + 64, + 217, + 546, + 289 + ], + "lines": [ + { + "bbox": [ + 64, + 217, + 546, + 289 + ], + "spans": [ + { + "bbox": [ + 64, + 217, + 546, + 289 + ], + "type": "image", + "image_path": "33c954137fe5c44b190944d552c103fdb2941765da13a0c5c50d4c0642914194.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 63, + 293, + 545, + 430 + ], + "blocks": [ + { + "bbox": [ + 63, + 293, + 545, + 430 + ], + "lines": [ + { + "bbox": [ + 63, + 293, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 63, + 293, + 545, + 430 + ], + "type": "image", + "image_path": "71eb58cb509cdc30ace89b4995357217a5bb94d9376531626877cf5d150bf5be.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 441, + 555, + 518 + ], + "lines": [ + { + "bbox": [ + 54, + 441, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 54, + 441, + 555, + 518 + ], + "type": "text", + "content": "Figure 1. Our framework, Uni-renderer, empowers the generative model to function both as a renderer and an inverse renderer by approximating the rendering equation using a data-driven approach. Given intrinsic attributes, Uni-renderer generates photo-realistic images, functioning as a renderer. When provided with a single RGB image, it effectively decomposes the intrinsic properties, functioning as an inverse renderer. Top: Uni-renderer generates smooth variations as renderer. Setting the roughness value to 1.0 results in the \"dog\" case, shown at the top, lacking specular highlights. Conversely, setting the metallic value to 1 makes the \"hat\" case appear metallic. Bottom Left: When functioning as an inverse renderer, Uni-renderer decomposes the intrinsic properties of a single RGB image. Bottom Right: Uni-renderer generates relighting results under different environment lightings." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 539, + 200, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 539, + 200, + 551 + ], + "spans": [ + { + "bbox": [ + 151, + 539, + 200, + 551 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 555, + 297, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 555, + 297, + 687 + ], + "spans": [ + { + "bbox": [ + 54, + 555, + 297, + 687 + ], + "type": "text", + "content": "Rendering and inverse rendering are pivotal tasks in both computer vision and graphics. The rendering equation is the core of the two tasks, as an ideal conditional distribution transfer function from intrinsic properties to RGB images. Despite achieving promising results of existing rendering methods, they merely approximate the ideal estimation for a specific scene and come with a high computational cost. Additionally, the inverse conditional distribution transfer is intractable due to the inherent ambiguity. To address these challenges, we propose a data-driven method that jointly models rendering and inverse rendering as two conditional" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 540, + 556, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 556, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 556, + 685 + ], + "type": "text", + "content": "generation tasks within a single diffusion framework. Inspired by UniDiffuser, we utilize two distinct time schedules to model both tasks, and with a tailored dual streaming module, we achieve cross-conditioning of two pre-trained diffusion models. This unified approach, named Uni-Renderer, allows the two processes to facilitate each other through a cycle-consistent constrain, mitigating ambiguity by enforcing consistency between intrinsic properties and rendered images. Combined with a meticulously prepared dataset, our method effectively decomposition of intrinsic properties and demonstrating a strong capability to recognize changes during rendering." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 694, + 228, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 694, + 228, + 715 + ], + "spans": [ + { + "bbox": [ + 69, + 694, + 228, + 715 + ], + "type": "text", + "content": "* Equal contribution. †Corresponding Author. This project is supported by Quwan Technology." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26504" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 305 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 305 + ], + "type": "text", + "content": "Physically based rendering is crucial in computer vision and graphics for producing photorealistic 2D images from 3D models, textures, and lighting setups. This process is essential in animation [43], and architectural visualization [23]. At its core, the rendering equation [18] models the flow of light energy within a scene, offering an ideal conditional distribution transfer function for rendering. Monte Carlo light-transport simulations [34], utilizing path tracing [24], are commonly employed to evaluate the rendering equation. However, its recursive tracing incurs significant computational demands, posing challenges for real-time implementation. Inverse rendering, which aims to deduce geometric, material, and lighting information from images, has long been studied due to its under-constrained nature [10]. This approach allows the reconstructed 3D model to be directly integrated into rendering engines, thereby playing a critical role in downstream applications such as game production [25]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 306, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 296, + 533 + ], + "type": "text", + "content": "Recent advancements employ differentiable rendering for 3D representations [5, 45] or focus on direct decomposition of intrinsics using 2D object priors [7, 11]. These approaches use differentiable renderer to integrate intrinsic properties into RGB images, facilitating loss computation against ground-truth images for optimization. For instance, MaterialGAN [11] leverages StyleGAN2 [19] to synthesize realistic material properties using a large-scale, spatially-varying material dataset [7], while other works [4, 22, 44, 46] utilize diffusion models to achieve better results. Nevertheless, the ambiguity between geometry, materials, and lighting still hinders the effectiveness of decomposition. The mapping from observed images back to intrinsic properties is not one-to-one, leading to ambiguity and suboptimal performance. The most similar work to ours is RGB2X [47], which performs both forward and inverse rendering using two separate diffusion models. However, it fails to establish a connection between the two processes, treating them independently. Ambiguity issue still exists." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 534, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 653 + ], + "type": "text", + "content": "To alleviate this issue, we propose to jointly learn the two distributions together. By integrating rendering and inverse rendering into a single framework from a multi-task learning perspective, the two processes can facilitate each other. This joint learning allows us to use the inverted intrinsic properties to perform another cycle of rendering, effectively creating a cycle-consistent constrains. This cycle rendering can mitigate the ambiguity problem by enforcing consistency between the intrinsic properties and the rendered images, leading to improved performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "In this work, we explicitly model the rendering process and inverse rendering as two conditional generation tasks. Inspired by UniDiffuser [1], we utilize two distinct time schedules to model both conditional generation tasks within a single diffusion pipeline. By cross-conditioning two pre" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 106 + ], + "type": "text", + "content": "trained diffusion models through a dual streaming module, we achieve both rendering and inverse rendering in a unified framework." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 108, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 555, + 156 + ], + "type": "text", + "content": "To summarize, we introduce Uni-renderer which utilizes a pre-trained diffusion model to handle both rendering and inverse rendering in a unified framework. The key contributions of our method are as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 159, + 556, + 315 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 159, + 556, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 159, + 556, + 219 + ], + "spans": [ + { + "bbox": [ + 314, + 159, + 556, + 219 + ], + "type": "text", + "content": "- We propose a data-driven method to approximate the rendering equation. By modeling rendering and inverse rendering as two conditional generation tasks, we design a unified diffusion model with a tailored dual stream module for cross-condition generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 219, + 555, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 219, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 314, + 219, + 555, + 266 + ], + "type": "text", + "content": "- We render a synthetic dataset of different fine-grained material edits using 200K 3D objects with randomized intrinsic attributes by varying one of the rendering elements at a time while keeping the others constant." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 267, + 555, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 267, + 555, + 315 + ], + "spans": [ + { + "bbox": [ + 314, + 267, + 555, + 315 + ], + "type": "text", + "content": "- Extensive experiments validate the effectiveness of our method, achieving robust disentangling of intrinsic properties and demonstrating a strong capability to recognize changes in rendering." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 327, + 400, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 327, + 400, + 338 + ], + "spans": [ + { + "bbox": [ + 314, + 327, + 400, + 338 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 346, + 388, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 346, + 388, + 358 + ], + "spans": [ + { + "bbox": [ + 314, + 346, + 388, + 358 + ], + "type": "text", + "content": "2.1. Rendering" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 364, + 556, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 556, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 556, + 579 + ], + "type": "text", + "content": "Rendering, the process of generating 2D images from 3D models, synthesizes raw data from a 3D scene—including geometry, materials, and lighting—using the rendering equation [18]. This task is typically performed by rendering engines, such as Blender [14] and Unity [12], which serve as the technological intermediaries converting 3D models into 2D images or videos. However, achieving photo-realistic 2D images often involves recursive ray tracing [38] or path tracing [24], techniques that closely approximate the rendering equation but are time-consuming. Although several GPU-based acceleration methods have been proposed [8, 13, 21], and existing engines have incorporated these tools, real-time rendering continues to pose significant challenges. In contrast to traditional rendering approaches, our method considers the diffusion model as an alternative renderer, which utilizes geometry, materials, and lighting as conditional cues to produce photo-realistic 2D images without the need for recursive tracing." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 588, + 424, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 588, + 424, + 601 + ], + "spans": [ + { + "bbox": [ + 314, + 588, + 424, + 601 + ], + "type": "text", + "content": "2.2. Inverse Rendering" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": "Inverse rendering [2, 33], the task of decomposing an image's appearance into intrinsic properties such as geometry, material, and lighting, remains a longstanding and severely ill-posed problem in computer vision and graphics. Some approaches employ differentiable renderers on 3D representations [5, 20] to directly optimize these properties using image losses. Advancements in volume rendering, as utilized in Neural Radiance Fields (NeRFs) [29] for 3D reconstruction from multi-view 2D images, have led most" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26505" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 78, + 558, + 268 + ], + "blocks": [ + { + "bbox": [ + 58, + 78, + 558, + 268 + ], + "lines": [ + { + "bbox": [ + 58, + 78, + 558, + 268 + ], + "spans": [ + { + "bbox": [ + 58, + 78, + 558, + 268 + ], + "type": "image", + "image_path": "b8e693381e9e32c516785c82b1b60055ce426559187e9a6dc017121808dd9454.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 280, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 55, + 280, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 555, + 346 + ], + "type": "text", + "content": "Figure 2. The overview of our pipeline. During training, both attribute and RGB images are input to a unified model with pre-trained VAE encoders. The timestep selector plays a crucial role by adjusting the timesteps for each branch. Specifically, it ensures that one branch (either the attribute or RGB) has a timestep of 0, while the other branch selects a timestep from " + }, + { + "bbox": [ + 55, + 280, + 555, + 346 + ], + "type": "inline_equation", + "content": "t \\in [0, T]" + }, + { + "bbox": [ + 55, + 280, + 555, + 346 + ], + "type": "text", + "content": ". This mechanism allows our model to effectively learn the conditional distributions " + }, + { + "bbox": [ + 55, + 280, + 555, + 346 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_0|\\mathbf{y}_0)" + }, + { + "bbox": [ + 55, + 280, + 555, + 346 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 280, + 555, + 346 + ], + "type": "inline_equation", + "content": "q(\\mathbf{y}_0|\\mathbf{x}_0)" + }, + { + "bbox": [ + 55, + 280, + 555, + 346 + ], + "type": "text", + "content": " in alternating iterations. During rendering and inverse rendering, the corresponding conditions are input to the model with a timestep of 0, and the attributes/RGB images are generated through a sampled noise. (The VAE encoder and decoder are omitted for simplicity.)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 367, + 296, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 367, + 296, + 606 + ], + "spans": [ + { + "bbox": [ + 56, + 367, + 296, + 606 + ], + "type": "text", + "content": "methods [9, 26, 27, 45, 48] to leverage multiple images for reconstruction and subsequent estimation of materials and lighting. Additionally, they require per-object optimization and struggle to generalize across different objects, limiting their efficiency in fast estimation. MaterialGAN [11] utilizes a StyleGAN2-based model for generating intrinsic properties and a renderer for novel view synthesis under specified lighting conditions. Similarly, SIC [7] employs an encoder-decoder network for single-image-based inverse rendering, where rendering loss is used for optimization. Nonetheless, these efforts predominantly focus on planar surfaces and still necessitate a renderer to integrate predicted attributes into the final rendered image. [4, 22, 44, 46, 47] utilize diffusion models to achieve better results. However, these methods have not achieved optimal performance due to the inherent ambiguity involved in transferring distributions from images to intrinsic properties. Our method effectively mitigates the ambiguity among intrinsic properties by introducing a dual-stream diffusion pipeline and combined with a meticulously prepared dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 612, + 159, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 159, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 159, + 624 + ], + "type": "text", + "content": "2.3. Diffusion Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "text", + "content": "Diffusion Models [15, 32, 35, 39, 40] belong to the class of probabilistic generative models that progressively destruct data by injecting noise, and then subsequently learn to reverse this process in order to generate new samples. Alchemist [37] leverages a pre-trained diffusion model to achieve material control over a given object. [42] adopts a collaborative diffusion pipeline to decompose physically-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 367, + 556, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 556, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 556, + 486 + ], + "type": "text", + "content": "based rendering (PBR) material properties from RGB. However, these works focus either on rendering or inverse rendering and fail to yield a unified model to fit in real-world usages. Alternatively, we design a dual stream diffusion model as both a renderer and an inverse renderer. During the inverse rendering stage, it takes a single RGB image as input and disentangles all the intrinsic properties. For the rendering stage, all of the intrinsic properties including material, geometry, and lighting conditions will be taken and rendered via the same model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 528, + 382, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 528, + 382, + 541 + ], + "spans": [ + { + "bbox": [ + 314, + 528, + 382, + 541 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "content": "Given a single RGB image " + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "content": ", inverse rendering aims at decomposing it into intrinsic attributes including metallic " + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "content": ", roughness " + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "content": ", albedo " + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "content": ", surface normal " + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "content": ", specular lighting " + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "content": " and diffuse lighting " + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "content": ", which is formulated as " + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\{m, r, \\mathbf{a}, \\mathbf{n}, \\mathbf{s}, \\mathbf{d}\\} = \\text{InverseRenderer}(\\mathbf{I})" + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "content": ". For later reference, we denote the combination of attributes as " + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "content": ". Renderer takes attributes as input and renders them into RGB image denoted as " + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I} = \\text{Renderer}(\\mathbf{C})" + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "content": ". We formulate a dual stream diffusion model as renderer and inverse renderer simultaneously. We start by briefly revisiting physically based rendering (PBR) and Unidiffuser [1] in Section 3.1. Next, we introduce our dual stream diffusion framework in Section 3.2. We elaborate on the data preparation in Section 3.3." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26506" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 72, + 553, + 336 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 553, + 336 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 553, + 336 + ], + "type": "image", + "image_path": "d8995d7a23d05fff4967f6a9ed356893286239cab760d1fe6fa3605d22f9644d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 340, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 55, + 340, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 555, + 384 + ], + "type": "text", + "content": "Figure 3. We demonstrate smooth changes via rendering for different metallic and roughness strengths. The rendering is performed giving different combinations of the attributes. When the roughness value was set to 1, the cake and clock case shown in the top left are without specular highlights. When the metallic value was set to 1, the orange and baseball cases appeared to be metallic and revealing object illumination. Best viewed in color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 404, + 141, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 141, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 141, + 415 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 425, + 295, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 295, + 485 + ], + "type": "text", + "content": "Physically Based Rendering. Physically-based rendering is a computer graphics approach that seeks to render images by modeling the interaction between lights and materials in the real world. At its core, the rendering equation [18] describes the light energy flow in the scene, formulated as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 501, + 295, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 501, + 295, + 526 + ], + "spans": [ + { + "bbox": [ + 69, + 501, + 295, + 526 + ], + "type": "interline_equation", + "content": "L (\\mathbf {x}, \\omega_ {o}) = \\int_ {\\Omega} f (\\mathbf {x}, \\omega_ {o}, \\omega_ {i}) L _ {i} (\\mathbf {x}, \\omega_ {i}) (\\omega_ {i}, \\mathbf {n}) d \\omega_ {i}, \\tag {1}", + "image_path": "9d3595b7aebc51c85dcde163f011298057584c3633d7eb5b634427c27e6f45b3.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 531, + 295, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 295, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 295, + 627 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 531, + 295, + 627 + ], + "type": "inline_equation", + "content": "\\omega_{o}" + }, + { + "bbox": [ + 55, + 531, + 295, + 627 + ], + "type": "text", + "content": " is the viewing direction of the outgoing light, " + }, + { + "bbox": [ + 55, + 531, + 295, + 627 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 531, + 295, + 627 + ], + "type": "text", + "content": " is the BRDF properties, " + }, + { + "bbox": [ + 55, + 531, + 295, + 627 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 55, + 531, + 295, + 627 + ], + "type": "text", + "content": " is the incident light of direction " + }, + { + "bbox": [ + 55, + 531, + 295, + 627 + ], + "type": "inline_equation", + "content": "\\omega_{i}" + }, + { + "bbox": [ + 55, + 531, + 295, + 627 + ], + "type": "text", + "content": " sampled from the upper hemisphere " + }, + { + "bbox": [ + 55, + 531, + 295, + 627 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 531, + 295, + 627 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 531, + 295, + 627 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 55, + 531, + 295, + 627 + ], + "type": "text", + "content": " is the surface normal. The rendering equation is the key to handle both rendering and inverse rendering problems. Given intrinsic properties, the rendering equation generates photo-realistic images. Given an image, a rendering equation can also be used to optimize intrinsic properties." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": "Unidiffuser. UniDiffuser [1], a unified diffusion model that models all distributions at the same time, injects noise " + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon^x" + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon^y" + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": " to a set of paired image-text data " + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "inline_equation", + "content": "(\\mathbf{x_0},\\mathbf{y_0})" + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": " and generates noisy data " + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x_t^x}" + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{y_t^y}" + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "inline_equation", + "content": "0\\leq t_{x},t_{y}\\leq T" + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": " represent two individual timesteps. It then trains a joint noise prediction network " + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x_{t^{x}}},\\mathbf{y_{t^{y}}},t_{x},t_{y})" + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": " to predict the noise " + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon^x" + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon^y" + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": " by minimizing the mean squared error loss:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 339, + 421, + 555, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 421, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 339, + 421, + 555, + 444 + ], + "type": "interline_equation", + "content": "\\underset {\\epsilon^ {\\mathbf {x}}, \\epsilon^ {\\mathbf {y}}, \\mathbf {x} _ {0}, \\mathbf {y} _ {0}} {\\mathbb {E}} \\left[ \\left\\| [ \\epsilon^ {\\mathbf {x}}, \\epsilon^ {\\mathbf {y}} ] - \\epsilon_ {\\theta} (\\mathbf {x} _ {\\mathbf {t}} ^ {\\mathbf {x}}, \\mathbf {y} _ {\\mathbf {t}} ^ {\\mathbf {y}}, t _ {x}, t _ {y}) \\| ^ {2} \\right], \\right. \\quad (2)", + "image_path": "7b55050d6d2a9eb6c90fd518e1e1337306adfb53dc001792c7482fe1e0591fbc.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 450, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 556, + 617 + ], + "type": "text", + "content": "By predicting " + }, + { + "bbox": [ + 313, + 450, + 556, + 617 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_{t^{\\mathbf{x}}},\\mathbf{y}_{t^{\\mathbf{y}}},t_x,t_y)" + }, + { + "bbox": [ + 313, + 450, + 556, + 617 + ], + "type": "text", + "content": " for any " + }, + { + "bbox": [ + 313, + 450, + 556, + 617 + ], + "type": "inline_equation", + "content": "t_x" + }, + { + "bbox": [ + 313, + 450, + 556, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 450, + 556, + 617 + ], + "type": "inline_equation", + "content": "t_y" + }, + { + "bbox": [ + 313, + 450, + 556, + 617 + ], + "type": "text", + "content": ", UniDiffuser learns all distributions related to " + }, + { + "bbox": [ + 313, + 450, + 556, + 617 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_0,\\mathbf{y}_0)" + }, + { + "bbox": [ + 313, + 450, + 556, + 617 + ], + "type": "text", + "content": ". This includes all conditional distributions: " + }, + { + "bbox": [ + 313, + 450, + 556, + 617 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_0|\\mathbf{y}_0)" + }, + { + "bbox": [ + 313, + 450, + 556, + 617 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 450, + 556, + 617 + ], + "type": "inline_equation", + "content": "q(\\mathbf{y}_0|\\mathbf{x}_0)" + }, + { + "bbox": [ + 313, + 450, + 556, + 617 + ], + "type": "text", + "content": ", and those conditioned on noisy input, for example " + }, + { + "bbox": [ + 313, + 450, + 556, + 617 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_0|\\mathbf{y}_{t^{\\mathbf{y}}})" + }, + { + "bbox": [ + 313, + 450, + 556, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 450, + 556, + 617 + ], + "type": "inline_equation", + "content": "q(\\mathbf{y}_0|x_{t^x})" + }, + { + "bbox": [ + 313, + 450, + 556, + 617 + ], + "type": "text", + "content": ", for " + }, + { + "bbox": [ + 313, + 450, + 556, + 617 + ], + "type": "inline_equation", + "content": "0\\leq t_x,t_y\\leq T" + }, + { + "bbox": [ + 313, + 450, + 556, + 617 + ], + "type": "text", + "content": ". Learning a conditional distribution can be seen as learning a distinct task. From a multitask learning perspective, due to limited network bandwidth, learning many tasks simultaneously (i.e., fitting all distributions to a single network) may result in task competition or task conflict, ultimately leading to sub-optimal performance. For rendering and inverse rendering, we exclusively modeled the two conditional distributions " + }, + { + "bbox": [ + 313, + 450, + 556, + 617 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_0|\\mathbf{y}_0)" + }, + { + "bbox": [ + 313, + 450, + 556, + 617 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 450, + 556, + 617 + ], + "type": "inline_equation", + "content": "q(\\mathbf{y}_0|\\mathbf{x}_0)" + }, + { + "bbox": [ + 313, + 450, + 556, + 617 + ], + "type": "text", + "content": " to resolve the aforementioned issue and enhance the performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 624, + 403, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 403, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 403, + 635 + ], + "type": "text", + "content": "3.2. Uni-Renderer" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": "Our design incorporates a pre-trained RGB image model and a PBR model, the two were tightly linked to one another via a dual streaming technique aligning the RGB model's information with the PBR model's information. The overall framework is shown in Figure 2. Next, we will discuss each sub-module in the pipeline in detail." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26507" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 70, + 121, + 285 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 121, + 285 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 121, + 285 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 121, + 285 + ], + "type": "image", + "image_path": "fce9c77b5886093e0d36e348354bc70aa86f3c35ec6db4f9e16f6816a3d5542b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 286, + 113, + 294 + ], + "lines": [ + { + "bbox": [ + 76, + 286, + 113, + 294 + ], + "spans": [ + { + "bbox": [ + 76, + 286, + 113, + 294 + ], + "type": "text", + "content": "Ref. Image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 129, + 78, + 165, + 119 + ], + "blocks": [ + { + "bbox": [ + 129, + 78, + 165, + 119 + ], + "lines": [ + { + "bbox": [ + 129, + 78, + 165, + 119 + ], + "spans": [ + { + "bbox": [ + 129, + 78, + 165, + 119 + ], + "type": "image", + "image_path": "d076f9198be8b4fd20c611e1ba18c19467acdec790e26e1b3c5aa7b9fb512e14.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 129, + 131, + 165, + 174 + ], + "blocks": [ + { + "bbox": [ + 129, + 131, + 165, + 174 + ], + "lines": [ + { + "bbox": [ + 129, + 131, + 165, + 174 + ], + "spans": [ + { + "bbox": [ + 129, + 131, + 165, + 174 + ], + "type": "image", + "image_path": "95eb62e66ba26b82f097ebd061a696f92d1ff9a8f4026ce296bb747c867e0f19.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 129, + 185, + 165, + 228 + ], + "blocks": [ + { + "bbox": [ + 129, + 185, + 165, + 228 + ], + "lines": [ + { + "bbox": [ + 129, + 185, + 165, + 228 + ], + "spans": [ + { + "bbox": [ + 129, + 185, + 165, + 228 + ], + "type": "image", + "image_path": "8efbd67ed6f43f507ee577ac9bd41c4cb3eae40eccf25c3abe822882b63d5bf8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 129, + 241, + 164, + 283 + ], + "blocks": [ + { + "bbox": [ + 129, + 241, + 164, + 283 + ], + "lines": [ + { + "bbox": [ + 129, + 241, + 164, + 283 + ], + "spans": [ + { + "bbox": [ + 129, + 241, + 164, + 283 + ], + "type": "image", + "image_path": "2acebc2d319bec8331123d806b54053663f6005974c26bede52fde47f9e93282.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 285, + 162, + 293 + ], + "lines": [ + { + "bbox": [ + 133, + 285, + 162, + 293 + ], + "spans": [ + { + "bbox": [ + 133, + 285, + 162, + 293 + ], + "type": "text", + "content": "Yi.et al" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 182, + 78, + 219, + 119 + ], + "blocks": [ + { + "bbox": [ + 182, + 78, + 219, + 119 + ], + "lines": [ + { + "bbox": [ + 182, + 78, + 219, + 119 + ], + "spans": [ + { + "bbox": [ + 182, + 78, + 219, + 119 + ], + "type": "image", + "image_path": "de07a5ca93e501833ef95e4fb995a3d6ff503b3f40ed2f36ba9c2cd735ede5f6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 189, + 132, + 209, + 173 + ], + "blocks": [ + { + "bbox": [ + 189, + 132, + 209, + 173 + ], + "lines": [ + { + "bbox": [ + 189, + 132, + 209, + 173 + ], + "spans": [ + { + "bbox": [ + 189, + 132, + 209, + 173 + ], + "type": "image", + "image_path": "629169e4536c96fc11fc6098d814fb892a8af092f07faa244fe50dfe9f73fc1e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 182, + 185, + 219, + 227 + ], + "blocks": [ + { + "bbox": [ + 182, + 185, + 219, + 227 + ], + "lines": [ + { + "bbox": [ + 182, + 185, + 219, + 227 + ], + "spans": [ + { + "bbox": [ + 182, + 185, + 219, + 227 + ], + "type": "image", + "image_path": "4cd34e57ead8a77ff649f5074f60f0395caa22b8076eb4c3b4c7289a7bb99606.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 184, + 241, + 217, + 282 + ], + "blocks": [ + { + "bbox": [ + 184, + 241, + 217, + 282 + ], + "lines": [ + { + "bbox": [ + 184, + 241, + 217, + 282 + ], + "spans": [ + { + "bbox": [ + 184, + 241, + 217, + 282 + ], + "type": "image", + "image_path": "fe36d1ac3d0b44e56ff68bb5f4217ceabdf5c90cfd3f56473cb6abff4f915403.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 285, + 208, + 293 + ], + "lines": [ + { + "bbox": [ + 189, + 285, + 208, + 293 + ], + "spans": [ + { + "bbox": [ + 189, + 285, + 208, + 293 + ], + "type": "text", + "content": "RGB2X" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 238, + 78, + 273, + 119 + ], + "blocks": [ + { + "bbox": [ + 238, + 78, + 273, + 119 + ], + "lines": [ + { + "bbox": [ + 238, + 78, + 273, + 119 + ], + "spans": [ + { + "bbox": [ + 238, + 78, + 273, + 119 + ], + "type": "image", + "image_path": "c3c9a53ac0a33108adfe8c3ac65a63c41422d6619f5448796289e4a695092468.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 246, + 132, + 265, + 172 + ], + "blocks": [ + { + "bbox": [ + 246, + 132, + 265, + 172 + ], + "lines": [ + { + "bbox": [ + 246, + 132, + 265, + 172 + ], + "spans": [ + { + "bbox": [ + 246, + 132, + 265, + 172 + ], + "type": "image", + "image_path": "025b74baa17b9e1770c2fcc396c73d00897e3fe3d16b1a4688ad86bac5cd821f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 237, + 184, + 275, + 226 + ], + "blocks": [ + { + "bbox": [ + 237, + 184, + 275, + 226 + ], + "lines": [ + { + "bbox": [ + 237, + 184, + 275, + 226 + ], + "spans": [ + { + "bbox": [ + 237, + 184, + 275, + 226 + ], + "type": "image", + "image_path": "c48bc7781711d78f4c01d712a2e001983b7a1b2339fc259590611f42ee14f63b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 238, + 241, + 271, + 282 + ], + "blocks": [ + { + "bbox": [ + 238, + 241, + 271, + 282 + ], + "lines": [ + { + "bbox": [ + 238, + 241, + 271, + 282 + ], + "spans": [ + { + "bbox": [ + 238, + 241, + 271, + 282 + ], + "type": "image", + "image_path": "acab94daa601a6f9eacb05012c31a294b2fab02c9700d184ea2a62f3ab6b0355.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 285, + 282, + 293 + ], + "lines": [ + { + "bbox": [ + 227, + 285, + 282, + 293 + ], + "spans": [ + { + "bbox": [ + 227, + 285, + 282, + 293 + ], + "type": "text", + "content": "IntrinImgDecomp" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 316, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 55, + 316, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 555, + 340 + ], + "type": "text", + "content": "Figure 4. Albedo comparison. Albedo Comparison of Uni-Renderer with baseline methods. We compared 4 learning-based methods and 2 optimization-based methods. Among all, Uni-renderer yields the most realistic results. Best viewed in color." + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 291, + 78, + 326, + 119 + ], + "blocks": [ + { + "bbox": [ + 291, + 78, + 326, + 119 + ], + "lines": [ + { + "bbox": [ + 291, + 78, + 326, + 119 + ], + "spans": [ + { + "bbox": [ + 291, + 78, + 326, + 119 + ], + "type": "image", + "image_path": "6070618d1b4ee747ec4277038b8adfe4cfd9eedce46f7db7835fb350a9533b92.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 299, + 133, + 318, + 172 + ], + "blocks": [ + { + "bbox": [ + 299, + 133, + 318, + 172 + ], + "lines": [ + { + "bbox": [ + 299, + 133, + 318, + 172 + ], + "spans": [ + { + "bbox": [ + 299, + 133, + 318, + 172 + ], + "type": "image", + "image_path": "4dd6ce48e766c294e072dfd0d4d3fa7a774561bb2b3fbcf0b6519be056a5bf84.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 291, + 185, + 328, + 226 + ], + "blocks": [ + { + "bbox": [ + 291, + 185, + 328, + 226 + ], + "lines": [ + { + "bbox": [ + 291, + 185, + 328, + 226 + ], + "spans": [ + { + "bbox": [ + 291, + 185, + 328, + 226 + ], + "type": "image", + "image_path": "8efc27d847f0ad3a0f2d0f51d31f61ddca6a8140c258071baf80758f0fcc1d2a.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 294, + 241, + 327, + 282 + ], + "blocks": [ + { + "bbox": [ + 294, + 241, + 327, + 282 + ], + "lines": [ + { + "bbox": [ + 294, + 241, + 327, + 282 + ], + "spans": [ + { + "bbox": [ + 294, + 241, + 327, + 282 + ], + "type": "image", + "image_path": "d0c1cd7ea4ec1954b8871c591869e7915a8e6a7c53d6b69758b8c2d15711414d.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 287, + 285, + 337, + 293 + ], + "lines": [ + { + "bbox": [ + 287, + 285, + 337, + 293 + ], + "spans": [ + { + "bbox": [ + 287, + 285, + 337, + 293 + ], + "type": "text", + "content": "IntrinImgDiff." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 345, + 78, + 381, + 119 + ], + "blocks": [ + { + "bbox": [ + 345, + 78, + 381, + 119 + ], + "lines": [ + { + "bbox": [ + 345, + 78, + 381, + 119 + ], + "spans": [ + { + "bbox": [ + 345, + 78, + 381, + 119 + ], + "type": "image", + "image_path": "fadbb50a2b5bc704872812686a38bd24a3e42b934772aa85a52500db989a6ea7.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 354, + 133, + 372, + 172 + ], + "blocks": [ + { + "bbox": [ + 354, + 133, + 372, + 172 + ], + "lines": [ + { + "bbox": [ + 354, + 133, + 372, + 172 + ], + "spans": [ + { + "bbox": [ + 354, + 133, + 372, + 172 + ], + "type": "image", + "image_path": "deddfa13121f4d58715c8fee9367d76b20b5de98fb3ea07cc21c59821ca33574.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 345, + 184, + 383, + 226 + ], + "blocks": [ + { + "bbox": [ + 345, + 184, + 383, + 226 + ], + "lines": [ + { + "bbox": [ + 345, + 184, + 383, + 226 + ], + "spans": [ + { + "bbox": [ + 345, + 184, + 383, + 226 + ], + "type": "image", + "image_path": "064c89d599ce4dc1967cc22b0ff7c920af7677827df67f3375e51a44efce3ba8.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 347, + 241, + 381, + 282 + ], + "blocks": [ + { + "bbox": [ + 347, + 241, + 381, + 282 + ], + "lines": [ + { + "bbox": [ + 347, + 241, + 381, + 282 + ], + "spans": [ + { + "bbox": [ + 347, + 241, + 381, + 282 + ], + "type": "image", + "image_path": "4dd84127bce98846052708e9441ee80cdb5319003450c3533f1b95ede475f947.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 346, + 285, + 386, + 293 + ], + "lines": [ + { + "bbox": [ + 346, + 285, + 386, + 293 + ], + "spans": [ + { + "bbox": [ + 346, + 285, + 386, + 293 + ], + "type": "text", + "content": "GaussShader" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 396, + 78, + 431, + 119 + ], + "blocks": [ + { + "bbox": [ + 396, + 78, + 431, + 119 + ], + "lines": [ + { + "bbox": [ + 396, + 78, + 431, + 119 + ], + "spans": [ + { + "bbox": [ + 396, + 78, + 431, + 119 + ], + "type": "image", + "image_path": "84be300370926bf4794e720243080b41556bf15ec0c2bb3b97cdf07034927eff.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 405, + 133, + 422, + 172 + ], + "blocks": [ + { + "bbox": [ + 405, + 133, + 422, + 172 + ], + "lines": [ + { + "bbox": [ + 405, + 133, + 422, + 172 + ], + "spans": [ + { + "bbox": [ + 405, + 133, + 422, + 172 + ], + "type": "image", + "image_path": "ba724ebcabf3a2cf4040441e20f8b586ebf28b4d3e7c874d2c6cdc98460d1e76.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 397, + 184, + 433, + 226 + ], + "blocks": [ + { + "bbox": [ + 397, + 184, + 433, + 226 + ], + "lines": [ + { + "bbox": [ + 397, + 184, + 433, + 226 + ], + "spans": [ + { + "bbox": [ + 397, + 184, + 433, + 226 + ], + "type": "image", + "image_path": "77514a80d79ebc0478f20d91a56c558565816321419609f6e57908aeea4ec3a2.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 399, + 241, + 431, + 282 + ], + "blocks": [ + { + "bbox": [ + 399, + 241, + 431, + 282 + ], + "lines": [ + { + "bbox": [ + 399, + 241, + 431, + 282 + ], + "spans": [ + { + "bbox": [ + 399, + 241, + 431, + 282 + ], + "type": "image", + "image_path": "5f80d967e15a91c65476f2bdb2cba51031e30fb2e0f53fcd787ed2f0e98808eb.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 285, + 432, + 293 + ], + "lines": [ + { + "bbox": [ + 399, + 285, + 432, + 293 + ], + "spans": [ + { + "bbox": [ + 399, + 285, + 432, + 293 + ], + "type": "text", + "content": "NVdiffRec" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 446, + 78, + 482, + 119 + ], + "blocks": [ + { + "bbox": [ + 446, + 78, + 482, + 119 + ], + "lines": [ + { + "bbox": [ + 446, + 78, + 482, + 119 + ], + "spans": [ + { + "bbox": [ + 446, + 78, + 482, + 119 + ], + "type": "image", + "image_path": "29b54965e165b782974368cc94761be3e02b00d854a83b2f785dc8a6cf90251b.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 454, + 133, + 472, + 171 + ], + "blocks": [ + { + "bbox": [ + 454, + 133, + 472, + 171 + ], + "lines": [ + { + "bbox": [ + 454, + 133, + 472, + 171 + ], + "spans": [ + { + "bbox": [ + 454, + 133, + 472, + 171 + ], + "type": "image", + "image_path": "3cb57a5a4e34ed4a5af02f323f1a0f141e60fe8dba8ecf44e3ff6de0fbedcbd6.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 445, + 184, + 483, + 226 + ], + "blocks": [ + { + "bbox": [ + 445, + 184, + 483, + 226 + ], + "lines": [ + { + "bbox": [ + 445, + 184, + 483, + 226 + ], + "spans": [ + { + "bbox": [ + 445, + 184, + 483, + 226 + ], + "type": "image", + "image_path": "4344b182ac230dadee907d23f61656dab4ce1f552047722d0a8d2e6a3de2a0b9.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 449, + 241, + 480, + 281 + ], + "blocks": [ + { + "bbox": [ + 449, + 241, + 480, + 281 + ], + "lines": [ + { + "bbox": [ + 449, + 241, + 480, + 281 + ], + "spans": [ + { + "bbox": [ + 449, + 241, + 480, + 281 + ], + "type": "image", + "image_path": "2fb697f4a063b8da59c503650be4cece06832ddc10cadc4bfcc6f8ef305c45f3.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 285, + 471, + 293 + ], + "lines": [ + { + "bbox": [ + 456, + 285, + 471, + 293 + ], + "spans": [ + { + "bbox": [ + 456, + 285, + 471, + 293 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 500, + 78, + 534, + 118 + ], + "blocks": [ + { + "bbox": [ + 500, + 78, + 534, + 118 + ], + "lines": [ + { + "bbox": [ + 500, + 78, + 534, + 118 + ], + "spans": [ + { + "bbox": [ + 500, + 78, + 534, + 118 + ], + "type": "image", + "image_path": "9080c5ef2ccf5f2659382b804e24618a5d9950722d11b01a39dc55c3e5be3a07.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 507, + 133, + 525, + 171 + ], + "blocks": [ + { + "bbox": [ + 507, + 133, + 525, + 171 + ], + "lines": [ + { + "bbox": [ + 507, + 133, + 525, + 171 + ], + "spans": [ + { + "bbox": [ + 507, + 133, + 525, + 171 + ], + "type": "image", + "image_path": "e9372a335f015cd024f794c2cfcb26e892e6a99e0e2513e01c579e1b95acb84f.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 498, + 184, + 535, + 226 + ], + "blocks": [ + { + "bbox": [ + 498, + 184, + 535, + 226 + ], + "lines": [ + { + "bbox": [ + 498, + 184, + 535, + 226 + ], + "spans": [ + { + "bbox": [ + 498, + 184, + 535, + 226 + ], + "type": "image", + "image_path": "ac965cae4efccbc788eec2d061f14b237f631973b8784b59fcd37a24f23222ce.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 499, + 240, + 531, + 280 + ], + "blocks": [ + { + "bbox": [ + 499, + 240, + 531, + 280 + ], + "lines": [ + { + "bbox": [ + 499, + 240, + 531, + 280 + ], + "spans": [ + { + "bbox": [ + 499, + 240, + 531, + 280 + ], + "type": "image", + "image_path": "25700b7a6a24c0d40c0d8c0399f8449af6f02a435858fe234c8f8d70575e8aa9.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 496, + 285, + 540, + 292 + ], + "lines": [ + { + "bbox": [ + 496, + 285, + 540, + 292 + ], + "spans": [ + { + "bbox": [ + 496, + 285, + 540, + 292 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "bbox": [ + 56, + 361, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 296, + 540 + ], + "type": "text", + "content": "Latent Preparation. Recent advancements have benefited greatly from the pre-trained VAE [36], which downsample an RGB image into lower dimensional latent. We used separate VAE for encoding " + }, + { + "bbox": [ + 56, + 361, + 296, + 540 + ], + "type": "inline_equation", + "content": "[m, r]" + }, + { + "bbox": [ + 56, + 361, + 296, + 540 + ], + "type": "text", + "content": ", albedo a, surface normal n and environment lighting s, d before we channel-wisely concatenate them and feed them into the model. It is worth noting that attributes like metallic and roughness, often come with a scalar value, and cannot be directly encoded. Thus, we generate additional masks to binarize the two values into gray-scale images. Instead of having two separate VAE for m and r, We formed channel triplets " + }, + { + "bbox": [ + 56, + 361, + 296, + 540 + ], + "type": "inline_equation", + "content": "a_{\\mathrm{material}}" + }, + { + "bbox": [ + 56, + 361, + 296, + 540 + ], + "type": "text", + "content": ", consisting of " + }, + { + "bbox": [ + 56, + 361, + 296, + 540 + ], + "type": "inline_equation", + "content": "[m, r, \\mathbf{m}]" + }, + { + "bbox": [ + 56, + 361, + 296, + 540 + ], + "type": "text", + "content": " and process those with the RGB VAE, where m is the mask This design allows us to reduce the bandwidth overhead within the diffusion model and further improve generation quality." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 55, + 543, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 296, + 651 + ], + "type": "text", + "content": "Conditional Distributions Modeling. To achieve rendering and inverse rendering within a single model, we borrowed ideas from [1], and adopted two different timesteps " + }, + { + "bbox": [ + 55, + 543, + 296, + 651 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{attributes}}" + }, + { + "bbox": [ + 55, + 543, + 296, + 651 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 543, + 296, + 651 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{RGB}}" + }, + { + "bbox": [ + 55, + 543, + 296, + 651 + ], + "type": "text", + "content": " from three timesteps choices " + }, + { + "bbox": [ + 55, + 543, + 296, + 651 + ], + "type": "inline_equation", + "content": "\\{0, t, T\\}" + }, + { + "bbox": [ + 55, + 543, + 296, + 651 + ], + "type": "text", + "content": " to train a joint noise prediction network. It is worth noting that since our unified model only requires modeling two conditional distributions, we ruled out the choices where both " + }, + { + "bbox": [ + 55, + 543, + 296, + 651 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{attributes}}" + }, + { + "bbox": [ + 55, + 543, + 296, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 543, + 296, + 651 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{RGB}}" + }, + { + "bbox": [ + 55, + 543, + 296, + 651 + ], + "type": "text", + "content": " take on " + }, + { + "bbox": [ + 55, + 543, + 296, + 651 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 543, + 296, + 651 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 543, + 296, + 651 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 543, + 296, + 651 + ], + "type": "text", + "content": " and make sure there is always a timestep equal to 0. Formally," + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 78, + 660, + 296, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 660, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 78, + 660, + 296, + 693 + ], + "type": "interline_equation", + "content": "\\left(t _ {\\text {a t t r i b u t e s}}, t _ {\\mathrm {R G B}}\\right) = \\left\\{ \\begin{array}{l l} (0, \\tilde {t}) & \\text {d u r i n g} \\\\ (\\tilde {t}, 0) & \\text {i n v e r s e} \\end{array} , \\right. \\tag {3}", + "image_path": "9b4dd2a5adbf6b01eedd06f9e8152179955e73f2e893cb37c35ca69763f5f02c.jpg" + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 67, + 702, + 94, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 94, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 94, + 712 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 372, + 375, + 555, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 375, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 372, + 375, + 555, + 407 + ], + "type": "interline_equation", + "content": "\\tilde {t} = \\left\\{ \\begin{array}{l l} t & \\text {w i t h p r o b a b i l i t y} p \\\\ T & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {4}", + "image_path": "f645c29cad385353001c3c7306699cda31dd18c47fa4599e88c731741a7be651.jpg" + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 313, + 417, + 555, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 555, + 441 + ], + "type": "text", + "content": "The pseudo-code for generating timesteps can be found in supplementary material." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 313, + 445, + 556, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 556, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 556, + 624 + ], + "type": "text", + "content": "Dual Stream Diffusion. We designed a dual diffusion network to facilitate the unified model and adopt the " + }, + { + "bbox": [ + 313, + 445, + 556, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 445, + 556, + 624 + ], + "type": "text", + "content": "-prediction in the diffusion process. Training: During training, as shown in Figure 2, the upper branch of the model dealt with the intrinsic attributes only, leaving only RGB for the lower branch. The timestep selection module will determine if a rendering/inverse rendering iteration has occurred at a probability " + }, + { + "bbox": [ + 313, + 445, + 556, + 624 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 445, + 556, + 624 + ], + "type": "text", + "content": ". During a rendering iteration, the clean attributes " + }, + { + "bbox": [ + 313, + 445, + 556, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 313, + 445, + 556, + 624 + ], + "type": "text", + "content": " will be used as conditions to denoise the noisy RGB image " + }, + { + "bbox": [ + 313, + 445, + 556, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 445, + 556, + 624 + ], + "type": "text", + "content": ", the same will be applied to the inverse rendering iteration except that the noisy attributes are being fed to the model. Inference: At the inference stage, we provided attributes as conditions to the model and gradually denoise the RGB, and vice versa. The effectiveness of the dual stream framework is demonstrated in Sec 4.3." + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "content": "Cycle-Consistent Constrain. To alleviate the inherent ambiguity problem, we introduce a cycle-consistent constraint within our unified framework, where its incorporation is straightforward with our pipeline. During training, we use the model's predicted inverse results to perform an additional cycle of rendering. The re-rendered outputs are then utilized in the loss calculation to further optimize the inverse" + } + ] + } + ], + "index": 50 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26508" + } + ] + } + ], + "index": 51 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 294, + 157 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 294, + 157 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 294, + 157 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 294, + 157 + ], + "type": "image", + "image_path": "a92650899f318613a3665f18c14a5dbb7f0abae4ca24dde9626215b732189dc1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 175, + 296, + 218 + ], + "lines": [ + { + "bbox": [ + 55, + 175, + 296, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 296, + 218 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison. Rendering Comparison of Uni-Renderer with baseline methods. The base images are with a metallic value of 0.5. The comparison is made with a higher metallic value of 1.0. Best viewed in color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 239, + 173, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 173, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 173, + 251 + ], + "type": "text", + "content": "rendering process. Formally," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 95, + 258, + 296, + 279 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 258, + 296, + 279 + ], + "spans": [ + { + "bbox": [ + 95, + 258, + 296, + 279 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {\\mathbf {x} _ {0}, \\epsilon , t} \\left[ \\| \\mathbf {x} _ {0} - \\hat {\\mathbf {x}} _ {0} \\left(\\hat {\\mathbf {x}} _ {\\mathrm {r g b}}, t, \\mathbf {C}\\right) \\| ^ {2} \\right], \\tag {5}", + "image_path": "f5a686a7069f0e7808c51a1638752711d9a4219dc34fbe519c72bac0c1fe4b0b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 287, + 296, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 287, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 296, + 346 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 287, + 296, + 346 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0\\left(\\mathbf{x}_t,t,\\mathbf{C}\\right)" + }, + { + "bbox": [ + 55, + 287, + 296, + 346 + ], + "type": "text", + "content": " is the model's prediction of the original rendering output " + }, + { + "bbox": [ + 55, + 287, + 296, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 55, + 287, + 296, + 346 + ], + "type": "text", + "content": ", conditioned on the noisy input " + }, + { + "bbox": [ + 55, + 287, + 296, + 346 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{\\mathrm{rgb}}" + }, + { + "bbox": [ + 55, + 287, + 296, + 346 + ], + "type": "text", + "content": ", timestep " + }, + { + "bbox": [ + 55, + 287, + 296, + 346 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 287, + 296, + 346 + ], + "type": "text", + "content": ", and conditioning attribute information " + }, + { + "bbox": [ + 55, + 287, + 296, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 55, + 287, + 296, + 346 + ], + "type": "text", + "content": ". The effectiveness of the cycle-consistent constrain is demonstrated in Sec 4.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 354, + 160, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 160, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 160, + 366 + ], + "type": "text", + "content": "3.3. Data Preparation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 371, + 295, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 371, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 54, + 371, + 295, + 539 + ], + "type": "text", + "content": "We prepare our training data with 3D data from Objaverse [6], consisting of synthetic 3D assets. We sampled 200K assets from Objaverse. For each 3D asset in Objaverse, we rendered the 2D images, and materials map (i.e., metallic, roughness, albedo) by changing the metalness and roughness (range from 0 to 1 with step 0.1). We also randomly select a 2D environment map from a pool of 20K natural scenes from a subset of LHQ-1024 [17] and format them into an RGB style for providing lighting. For each object, we have 121 pairs with different metallic, roughness and lighting. The rendered images are of resolution " + }, + { + "bbox": [ + 54, + 371, + 295, + 539 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 54, + 371, + 295, + 539 + ], + "type": "text", + "content": ", and the camera poses are fixed in the front of the object. We also rendered 100 objects that were unseen during training to test the model generalization ability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 549, + 135, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 135, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 135, + 562 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "text", + "content": "We present qualitative and quantitative analysis in Section 4.1 to demonstrate the rendering capabilities of our model. The inverse rendering results can be found in Section 4.2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 612, + 128, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 128, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 128, + 624 + ], + "type": "text", + "content": "4.1. Rendering" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "In Figure 3, we demonstrate the effectiveness of our model in generalizing rendering techniques by utilizing different combinations of material attributes. Using our dual streaming pipeline, we start off by passing a reference image to perform inverse rendering. Then update the metallic, roughness value and finally re-render into an RGB image with updated materials." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 323, + 104, + 547, + 180 + ], + "blocks": [ + { + "bbox": [ + 313, + 62, + 555, + 95 + ], + "lines": [ + { + "bbox": [ + 313, + 62, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 62, + 555, + 95 + ], + "type": "text", + "content": "Table 1. Quantitative results on rendering. Metrics for the prompt-only InstructPix2Pix* trained on our data, Subias et.al and our proposed method computing the PSNR, and LPIPS." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 104, + 547, + 180 + ], + "lines": [ + { + "bbox": [ + 323, + 104, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 323, + 104, + 547, + 180 + ], + "type": "table", + "html": "
MethodsMetallicRoughness
PSNR↑LPIPS↓PSNR↑LPIPS↓
InPix2Pix* [3]24.250.103224.430.1056
Subias et.al [41]28.090.095428.130.0817
Ours30.720.076331.680.0695
Ours w/o unified27.330.093229.120.0987
Ours w/o re-render28.720.082428.930.0833
", + "image_path": "ab732e1f26d54e019fd0ff9561022ff24626dd63d264ebe6d023fad420ebda37.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 194, + 555, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 555, + 253 + ], + "type": "text", + "content": "Roughness. As the roughness increases, the output demonstrates the elimination of specular highlights, replaced by an estimate of the base albedo. Conversely, reducing the roughness amplifies the highlights, as evidenced in the cases of the crow and the lamp in Figure 3." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 256, + 554, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 554, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 554, + 316 + ], + "type": "text", + "content": "Metallics. The increase of the metallic strength in both the baseball and oranges in Figure 3 reduces the significance of the base albedo and enhances the lighting effects on the surfaces. In contrast, reducing the metallic strength reverses this effect." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 319, + 556, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 556, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 556, + 486 + ], + "type": "text", + "content": "Analysis. We compare our model to other material editing baselines: Subias et al. [41], Prompt-to- Prompt with Null-text Inversion (NTI) [30], and InstructPix2Pix [3] in Figure 5. We fine-tuned the InstructPix2Pix prompt-based approach with our dataset. Subias et al.'s method results in exaggerated material changes as their objective is perceptual, not physically-based material rendering. Null-text inversion and InstructPix2Pix trained on our dataset with a prompt-only approach, they either incorrectly modified the background or produces minimum changes. Our method renders the photo-realistic results, introducing the specular highlights while retaining the geometry and illumination effects. Additionally, the quantitative results with the aforementioned baselines are shown in Table 1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 495, + 423, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 423, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 423, + 507 + ], + "type": "text", + "content": "4.2. Inverse Rendering" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 512, + 556, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 512, + 556, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 512, + 556, + 609 + ], + "type": "text", + "content": "We use Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) as metrics to evaluate the image quality of the albedo, relighting images, and novel view synthesis. We use Cosine Similarity for normal evaluation. Roughness and metallic estimation are evaluated using Mean Squared Error (MSE). Detailed configurations of the comparisons are discussed in the supplementary file." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 614, + 429, + 626 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 614, + 429, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 614, + 429, + 626 + ], + "type": "text", + "content": "4.2.1. Lighting Estimation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "content": "We compare our model with NVdiffRec [31], Gaussian-Shader [16] in terms of specular and diffuse lighting decomposition, the results can be found in Figure 6 and in Table 3. In Figure 6, we demonstrate the relighting ability of our proposed model given different environmental lighting. To achieve this, we updated the specular, diffuse, and environment maps of the given objects. The rendered outputs" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26509" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 71, + 116, + 212 + ], + "blocks": [ + { + "bbox": [ + 77, + 71, + 116, + 212 + ], + "lines": [ + { + "bbox": [ + 77, + 71, + 116, + 212 + ], + "spans": [ + { + "bbox": [ + 77, + 71, + 116, + 212 + ], + "type": "image", + "image_path": "3673a81aa2384c6f115f27b4c35a1042dc7c4cba9f7688b89d686ada9e5d3998.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 213, + 113, + 220 + ], + "lines": [ + { + "bbox": [ + 74, + 213, + 113, + 220 + ], + "spans": [ + { + "bbox": [ + 74, + 213, + 113, + 220 + ], + "type": "text", + "content": "Env. Lighting" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 120, + 74, + 241, + 220 + ], + "blocks": [ + { + "bbox": [ + 120, + 74, + 241, + 220 + ], + "lines": [ + { + "bbox": [ + 120, + 74, + 241, + 220 + ], + "spans": [ + { + "bbox": [ + 120, + 74, + 241, + 220 + ], + "type": "image", + "image_path": "7f15c99c4863b8ac5c7e7938eba7d8054e2fedf288a3a7fa61182af87f9d55dd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 231, + 555, + 265 + ], + "lines": [ + { + "bbox": [ + 55, + 231, + 555, + 265 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 555, + 265 + ], + "type": "text", + "content": "Figure 6. Relighting Comparison The relighting comparison is performed on validation objects. We first inverse render the input RGB to acquire the intrinsics, and then we updated the lighting information to get the relighting results. The leftmost column is the reference environment lighting. Best viewed in color." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 249, + 74, + 386, + 221 + ], + "blocks": [ + { + "bbox": [ + 249, + 74, + 386, + 221 + ], + "lines": [ + { + "bbox": [ + 249, + 74, + 386, + 221 + ], + "spans": [ + { + "bbox": [ + 249, + 74, + 386, + 221 + ], + "type": "image", + "image_path": "edbb858b3be3bfc71c987e9383233ab24af2bce5e5f4093672c6fadc1b717b90.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 392, + 74, + 533, + 221 + ], + "blocks": [ + { + "bbox": [ + 392, + 74, + 533, + 221 + ], + "lines": [ + { + "bbox": [ + 392, + 74, + 533, + 221 + ], + "spans": [ + { + "bbox": [ + 392, + 74, + 533, + 221 + ], + "type": "image", + "image_path": "76c4853f9385c34bded3a5e958d4a7453a92a4515a8d1623b5b53c9954200c6e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 61, + 339, + 305, + 498 + ], + "blocks": [ + { + "bbox": [ + 55, + 284, + 297, + 338 + ], + "lines": [ + { + "bbox": [ + 55, + 284, + 297, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 297, + 338 + ], + "type": "text", + "content": "Table 2. Quantitative results on inverse rendering. For quantitative comparison, we have included both data-driven methods and optimization-based baselines. \"Ours w/o unified\" is trained without the unified framework, and \"Ours w/o constrain\" is trained without the cycle-consistent constrain term." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 339, + 305, + 498 + ], + "lines": [ + { + "bbox": [ + 61, + 339, + 305, + 498 + ], + "spans": [ + { + "bbox": [ + 61, + 339, + 305, + 498 + ], + "type": "table", + "html": "
MethodAlbedoMetallicRoughnessNormal
PSNR↑SSIM↑LPIPS↓MSE↓MSE↓cos-simi↑
Yi et al. [46]20.930.89600.1024---
IntrinsicAny. [44]22.670.92180.0633---
Wonder3d [28]-----0.834
IntrinImgDecomp. [4]21.910.89340.0659---
IntrinImgDiff. [22]21.830.90600.06320.19200.1315-
RGB2X. [47]18.150.88290.0851--0.871
NvDiffrec [31]13.560.86950.12430.31640.29560.631
GaussianShader [16]16.550.86400.09060.34210.37140.908
Ours23.200.91820.05320.11820.10370.928
Ours w/o unified18.620.88460.08330.16320.13910.867
Ours w/o constrain21.200.89340.06020.13910.13040.922
", + "image_path": "d4aaa1e446c362d21e28876424fcb9ddab1b3c4bbee7553ce8fade38494e7d78.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 520, + 297, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 297, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 297, + 568 + ], + "type": "text", + "content": "under diverse lighting scenarios show consistent adjustments, preserving the material attributes while accurately reflecting the changes in illumination. We have demonstrated a superior relighting ability compared to other baselines." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 575, + 186, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 186, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 186, + 588 + ], + "type": "text", + "content": "4.2.2. Materials and geometry" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 590, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 296, + 639 + ], + "type": "text", + "content": "Surface Normal. We present the surface normal comparison results in Figure 7. Our model yields the best performance in normal estimation when compared to Wonder3D [28], RGB2X [47], and Yi et al, [46]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 642, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 297, + 714 + ], + "type": "text", + "content": "Albedo. For albedo estimation, we have compared 2 optimization-based methods and 6 data-driven methods in terms of PSNR, SSIM and LPIPS. Figure 4 presents the qualitative comparison for albedo estimation, and the quantitative comparison can be found in Table 2. Our method significantly outperforms other methods." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 319, + 339, + 564, + 423 + ], + "blocks": [ + { + "bbox": [ + 313, + 284, + 555, + 338 + ], + "lines": [ + { + "bbox": [ + 313, + 284, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 555, + 338 + ], + "type": "text", + "content": "Table 3. Quantitative results on relighting. We have compared the inverse rendering results on lighting estimation in our left-out test set, \"Ours w/o unified\" is trained without the unified framework, and 'Ours w/o constrain' is trained without the cycle-consistent constrain." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 339, + 564, + 423 + ], + "lines": [ + { + "bbox": [ + 319, + 339, + 564, + 423 + ], + "spans": [ + { + "bbox": [ + 319, + 339, + 564, + 423 + ], + "type": "table", + "html": "
MethodSpecularDiffuseRelighting
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NvDiffrec [31]14.120.88220.178215.200.88340.183321.990.88500.0834
GaussianShader [16]16.820.89120.098217.990.89690.083126.470.88260.0822
Ours22.710.93660.049823.150.96940.037330.840.90320.0763
Ours w/o unified21.950.88160.055421.920.89340.049526.950.88320.0824
Ours w/o constrain22.070.90270.051322.820.92590.040828.120.89340.0924
", + "image_path": "362a2bfcd9ae253e6c5dfeea9eadc1db197fd8c5d6e9569737adf7c70d12b173.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 441, + 555, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 525 + ], + "type": "text", + "content": "Metallic and Roughness. Since existing data-driven methods do not include metallic and roughness estimation, we have compared the roughness and metallic estimation with two optimization-based inverse rendering method, including NVdiffRec [31] and GaussianShader [16], the quantitative results can be found in Table 2. More visualization cases can be found in the supplementary material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 530, + 436, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 436, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 436, + 542 + ], + "type": "text", + "content": "4.2.3. Real-World Inversing" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 545, + 556, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 641 + ], + "type": "text", + "content": "Despite being trained on synthetic data, our model is capable of performing inverse rendering in real-world cases. As shown in Figure 8, the \"Phone stand\" case has a more metallic appearance while maintaining some roughness. It is under a white-to-yellow light source thereby showing a white-to-yellow color in lighting. Our model also excels at recovering high-frequency ambient lighting from the image as specular lighting, as demonstrated in the \"kettle\" case." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 648, + 408, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 408, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 408, + 661 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "Dual Stream Diffusion. To demonstrate the effectiveness of our dual stream diffusion framework, we compared results with two separate diffusion models for performing rendering and inverse rendering. Combining the two models in a sin-" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26510" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 79, + 532, + 373 + ], + "blocks": [ + { + "bbox": [ + 85, + 79, + 532, + 373 + ], + "lines": [ + { + "bbox": [ + 85, + 79, + 532, + 373 + ], + "spans": [ + { + "bbox": [ + 85, + 79, + 532, + 373 + ], + "type": "image", + "image_path": "420bf8d99724d8a4298c35a8e3ea6acd429c25fc5a98e51abdb10324020602f1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 384, + 513, + 397 + ], + "lines": [ + { + "bbox": [ + 96, + 384, + 513, + 397 + ], + "spans": [ + { + "bbox": [ + 96, + 384, + 513, + 397 + ], + "type": "text", + "content": "Figure 7. Normal Comparison. Normal Comparison of Uni-Renderer with others methods. Best viewed in color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 492, + 416, + 515, + 423 + ], + "lines": [ + { + "bbox": [ + 492, + 416, + 515, + 423 + ], + "spans": [ + { + "bbox": [ + 492, + 416, + 515, + 423 + ], + "type": "text", + "content": "Specular" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 524, + 416, + 545, + 423 + ], + "lines": [ + { + "bbox": [ + 524, + 416, + 545, + 423 + ], + "spans": [ + { + "bbox": [ + 524, + 416, + 545, + 423 + ], + "type": "text", + "content": "Diffuse" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 296, + 477 + ], + "type": "text", + "content": "gle pipeline for both rendering and inverse can benefit each other, as shown in Table 2, 3 where the jointly trained model performance prevails. From a multi-task learning perspective, the unified pipeline benefits each other by sharing the knowledge through model weights." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 479, + 296, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 296, + 564 + ], + "type": "text", + "content": "Cycle-Consistent Constraint. We also conducted experiments to show the effectiveness of the cycle-consistent constraint. The constrain can mitigate the ambiguity problem by enforcing consistency between the intrinsic properties and the images, leading to improved performance. The results in Table 2, 3 demonstrated the inverse rendering quality improves with the constrain term being included." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 574, + 128, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 128, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 128, + 586 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 297, + 715 + ], + "type": "text", + "content": "We present a method that allows photo-realistic rendering and inverse rendering in a unified framework. Our method demonstrates strong performance in real-world cases, achieving photo-realistic rendering and inverse rendering that generalizes well beyond synthetic training data. Despite this success, there are still limitations when applying our approach to certain real-world scenarios, primarily due to the domain gap between synthetic training data and real-world images. This gap can lead to challenges in accurately handling complex or unfamiliar objects and environments, where the synthetic-" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 415, + 419, + 561 + ], + "blocks": [ + { + "bbox": [ + 316, + 415, + 419, + 561 + ], + "lines": [ + { + "bbox": [ + 316, + 415, + 419, + 561 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 419, + 561 + ], + "type": "image", + "image_path": "d2b6057a123f66217f55acd4e9150532d89dfad8625269247b3c3a3bb441e260.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 426, + 416, + 445, + 423 + ], + "lines": [ + { + "bbox": [ + 426, + 416, + 445, + 423 + ], + "spans": [ + { + "bbox": [ + 426, + 416, + 445, + 423 + ], + "type": "text", + "content": "Normal" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 459, + 416, + 477, + 423 + ], + "lines": [ + { + "bbox": [ + 459, + 416, + 477, + 423 + ], + "spans": [ + { + "bbox": [ + 459, + 416, + 477, + 423 + ], + "type": "text", + "content": "Albedo" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 570, + 555, + 602 + ], + "lines": [ + { + "bbox": [ + 316, + 570, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 555, + 602 + ], + "type": "text", + "content": "Figure 8. Real World Inversing. We present different real-world inversing cases under different lighting conditions. Best viewed in color." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 629, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 712 + ], + "type": "text", + "content": "to-real adaptation is less effective. To address this, future work will focus on incorporating more real-world data into our training process. By expanding our dataset with real-world examples, we aim to enhance the model's ability to bridge the synthetic-real domain gap, enabling better generalization and higher-quality rendering results across diverse real-world scenes." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26511" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Fan Bao, Shen Nie, Kaiwen Xue, Chongxuan Li, Shi Pu, Yaole Wang, Gang Yue, Yue Cao, Hang Su, and Jun Zhu. One transformer fits all distributions in multi-modal diffusion at scale, 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 135, + 296, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 135, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 135, + 296, + 168 + ], + "type": "text", + "content": "[2] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence (TPAMI), 2014. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 168, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 168, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 62, + 168, + 296, + 201 + ], + "type": "text", + "content": "[3] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions, 2023. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 202, + 296, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 296, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 296, + 235 + ], + "type": "text", + "content": "[4] Chris Careaga and Yaqiz Aksoy. Intrinsic image decomposition via ordinal shading. ACM Trans. Graph., 43(1), 2023. 2, 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 236, + 296, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 296, + 291 + ], + "type": "text", + "content": "[5] Wenzheng Chen, Joey Litalien, Jun Gao, Zian Wang, Clement Fuji Tsang, Sameh Khamis, Or Litany, and Sanja Fidler. Dibr++: learning to predict lighting and material with a hybrid differentiable renderer. Advances in Neural Information Processing Systems (NeuIPS), 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 291, + 296, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 296, + 356 + ], + "type": "text", + "content": "[6] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 357, + 296, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 296, + 401 + ], + "type": "text", + "content": "[7] Valentin Deschaintre, Miika Aittala, Fredo Durand, George Drettakis, and Adrien Bousseau. Single-image svbrdf capture with a rendering-aware deep network. ACM Transactions on Graphics (ToG), 2018. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 402, + 296, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 402, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 402, + 296, + 434 + ], + "type": "text", + "content": "[8] Cass Everitt and Mark J Kilgard. Practical and robust stenciled shadow volumes for hardware-accelerated rendering. 2003. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 435, + 296, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 435, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 62, + 435, + 296, + 489 + ], + "type": "text", + "content": "[9] Yue Fan, Ivan Skorokhodov, Oleg Voynov, Savva Ignatyev, Evgeny Burnaev, Peter Wonka, and Yiqun Wang. Factoredneus: Reconstructing surfaces, illumination, and materials of possibly glossy objects. arXiv preprint arXiv:2305.17929, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 491, + 296, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 491, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 56, + 491, + 296, + 545 + ], + "type": "text", + "content": "[10] Roger Grosse, Micah K Johnson, Edward H Adelson, and William T Freeman. Ground truth dataset and baseline evaluations for intrinsic image algorithms. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2009. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 546, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 296, + 590 + ], + "type": "text", + "content": "[11] Yu Guo, Cameron Smith, Milos Hasan, Kalyan Sunkavalli, and Shuang Zhao. Materialgan: Reflectance capture using a generative svbrdf model. arXiv preprint arXiv:2010.00114, 2020. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 590, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 296, + 612 + ], + "type": "text", + "content": "[12] John K Haas. A history of the unity game engine. Diss. Worcester Polytechnic Institute, 2014. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 613, + 296, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 296, + 657 + ], + "type": "text", + "content": "[13] Wolfgang Heidrich and Hans-Peter Seidel. Realistic, hardware-accelerated shading and lighting. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, 1999. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 658, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 296, + 679 + ], + "type": "text", + "content": "[14] Roland Hess. *Blender foundations: The essential guide to learning blender* 2.5. 2013. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "type": "text", + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "type": "text", + "content": "[16] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. arXiv preprint arXiv:2311.17977, 2023. 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 118, + 555, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 555, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 555, + 149 + ], + "type": "text", + "content": "[17] Kaggle user: dimensiOn. Lhq-1024 dataset. https://www.kaggle.com/datasets/dimensiOn/1024, 2023.6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 150, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 150, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 150, + 555, + 183 + ], + "type": "text", + "content": "[18] James T Kajiya. The rendering equation. In Proceedings of the 13th annual conference on Computer graphics and interactive techniques, 1986. 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 184, + 555, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 555, + 237 + ], + "type": "text", + "content": "[19] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 238, + 555, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 238, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 238, + 555, + 271 + ], + "type": "text", + "content": "[20] Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neural 3d mesh renderer. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 272, + 555, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 555, + 293 + ], + "type": "text", + "content": "[21] Mark J Kilgard and Jeff Bolz. GPU-accelerated path rendering. ACM Transactions on Graphics (TOG), 2012. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 294, + 555, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 555, + 326 + ], + "type": "text", + "content": "[22] Peter Kocsis, Vincent Sitzmann, and Matthias Nießner. Intrinsic image diffusion for indoor single-view material estimation, 2024. 2, 3, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 327, + 555, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 555, + 369 + ], + "type": "text", + "content": "[23] Jaroslav Křivánek, Christophe Chevallier, Vladimir Koylazov, Ondrej Karlík, Henrik Wann Jensen, and Thomas Ludwig. Realistic rendering in architecture and product visualization. 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 371, + 554, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 554, + 393 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 554, + 393 + ], + "type": "text", + "content": "[24] Eric P Lafortune and Yves D Willems. Bi-directional path tracing. 1993. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 394, + 555, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 555, + 415 + ], + "type": "text", + "content": "[25] Michael Lewis and Jeffrey Jacobson. Game engines. Communications of the ACM, 2002. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 416, + 555, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 555, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 555, + 470 + ], + "type": "text", + "content": "[26] Ruofan Liang, Huiting Chen, Chunlin Li, Fan Chen, Selvakumar Panneer, and Nandita Vijaykumar. Envidr: Implicit differentiable renderer with neural environment lighting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 471, + 555, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 555, + 514 + ], + "type": "text", + "content": "[27] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images. In SIGGRAPH, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 515, + 555, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 555, + 568 + ], + "type": "text", + "content": "[28] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, and Wenping Wang. _Wonder3d: Single image to 3d using cross-domain diffusion_, 2023. 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 570, + 555, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 555, + 613 + ], + "type": "text", + "content": "[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 614, + 554, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 554, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 554, + 647 + ], + "type": "text", + "content": "[30] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models, 2022. 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 647, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 713 + ], + "type": "text", + "content": "[31] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting Triangular 3D Models, Materials, and Lighting From Images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8280-8290, 2022. 6, 7" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26512" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 115 + ], + "type": "text", + "content": "[32] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 118, + 296, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 296, + 152 + ], + "type": "text", + "content": "[33] Merlin Nimier-David, Delio Vicini, Tizian Zeltner, and Wenzel Jakob. Mitsuba 2: A retargetable forward and inverse renderer. ACM Transactions on Graphics (TOG), 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 153, + 296, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 153, + 296, + 176 + ], + "spans": [ + { + "bbox": [ + 57, + 153, + 296, + 176 + ], + "type": "text", + "content": "[34] Matt Pharr, Wenzel Jakob, and Greg Humphreys. Physically based rendering: From theory to implementation. 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 177, + 296, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 177, + 296, + 232 + ], + "spans": [ + { + "bbox": [ + 57, + 177, + 296, + 232 + ], + "type": "text", + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 234, + 296, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 234, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 57, + 234, + 296, + 266 + ], + "type": "text", + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 269, + 296, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 269, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 57, + 269, + 296, + 312 + ], + "type": "text", + "content": "[37] Prafull Sharma, Varun Jampani, Yuanzhen Li, Xuhui Jia, Dmitry Lagun, Fredo Durand, William T. Freeman, and Mark Matthews. Alchemist: Parametric control of material properties with diffusion models, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 314, + 296, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 314, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 57, + 314, + 296, + 335 + ], + "type": "text", + "content": "[38] Peter Shirley and R Keith Morley. Realistic ray tracing. AK Peters, Ltd., 2008. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 337, + 296, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 337, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 57, + 337, + 296, + 369 + ], + "type": "text", + "content": "[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 372, + 296, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 372, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 57, + 372, + 296, + 415 + ], + "type": "text", + "content": "[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 418, + 294, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 418, + 294, + 440 + ], + "spans": [ + { + "bbox": [ + 57, + 418, + 294, + 440 + ], + "type": "text", + "content": "[41] J. Daniel Subias and Manuel Lagunas. In-the-wild material appearance editing using perceptual attributes, 2023. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 441, + 296, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 441, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 57, + 441, + 296, + 485 + ], + "type": "text", + "content": "[42] Shimon Vainer, Mark Boss, Mathias Parger, Konstantin Kutsy, Dante De Nigris, Ciara Rowles, Nicolas Perony, and Simon Donne. Collaborative control for geometry-conditioned pbr image generation, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 487, + 296, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 487, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 57, + 487, + 296, + 508 + ], + "type": "text", + "content": "[43] Catherine Winder and Zahra Dowlatabadi. Producing animation. 2013. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 511, + 296, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 511, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 57, + 511, + 296, + 553 + ], + "type": "text", + "content": "[44] Chen Xi, Peng Sida, Yang Dongchen, Liu Yuan, Pan Bowen, Lv Chengfei, and Zhou. Xiaowei. Intrinsicanything: Learning diffusion priors for inverse rendering under unknown illumination. arxiv: 2404.11593, 2024. 2, 3, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 556, + 296, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 556, + 296, + 609 + ], + "spans": [ + { + "bbox": [ + 57, + 556, + 296, + 609 + ], + "type": "text", + "content": "[45] Yao Yao, Jingyang Zhang, Jingbo Liu, Yihang Qu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf: Neural incident light field for physically-based material estimation. In European Conference on Computer Vision (ECCV), 2022. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 613, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 294, + 635 + ], + "type": "text", + "content": "[46] Renjiao Yi, Chenyang Zhu, and Kai Xu. Weakly-supervised single-view image relighting, 2023. 2, 3, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 637, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 637, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 637, + 296, + 712 + ], + "type": "text", + "content": "[47] Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloš Hasan. Rgb↔x: Image decomposition and synthesis using material- and lighting-aware diffusion models. In Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers '24, page 1-11. ACM, 2024. 2, 3, 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 555, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 555, + 125 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 555, + 125 + ], + "type": "text", + "content": "[48] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26513" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Uni4D_ Unifying Visual Foundation Models for 4D Modeling from a Single Video/b0bfb21a-5579-4e48-8be6-e0a25d4d029f_content_list.json b/2025/Uni4D_ Unifying Visual Foundation Models for 4D Modeling from a Single Video/b0bfb21a-5579-4e48-8be6-e0a25d4d029f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..305da45e85c813852be31a8c087aca8584e12b6c --- /dev/null +++ b/2025/Uni4D_ Unifying Visual Foundation Models for 4D Modeling from a Single Video/b0bfb21a-5579-4e48-8be6-e0a25d4d029f_content_list.json @@ -0,0 +1,1955 @@ +[ + { + "type": "text", + "text": "Uni4D: Unifying Visual Foundation Models for 4D Modeling from a Single Video", + "text_level": 1, + "bbox": [ + 91, + 130, + 906, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "David Yifan Yao Albert J. Zhai Shenlong Wang University of Illinois at Urbana-Champaign", + "bbox": [ + 264, + 180, + 730, + 217 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8d5a68cd8104e8c9ef41a266d3ea84d68c1f4bac745d0a671f92caf9ae9157de.jpg", + "image_caption": [ + "Figure 1. Given a casually captured video, Uni4D harnesses pretrained visual foundation models and multi-stage optimization to jointly estimate camera poses, dynamic geometry, and dense 3D motion. The resulting camera poses and geometry are accurate, consistent, and coherent both temporally and spatially. This is all done without any additional training or fine-tuning." + ], + "image_footnote": [], + "bbox": [ + 106, + 229, + 493, + 445 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c0741c4b9d98bb85d12de55a7b5097b9139c7076824b0261f814e7bcfb06049b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 227, + 890, + 445 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 503, + 326, + 520 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper presents a unified approach to understanding dynamic scenes from casual videos. Large pretrained vision foundation models, such as vision-language, video depth prediction, motion tracking, and segmentation models, offer promising capabilities. However, training a single model for comprehensive 4D understanding remains challenging. We introduce Uni4D, a multi-stage optimization framework that harnesses multiple pretrained models to advance dynamic 3D modeling, including static/dynamic reconstruction, camera pose estimation, and dense 3D motion tracking. Our results show state-of-the-art performance in dynamic 4D modeling with superior visual quality. Notably, Uni4D requires no retraining or fine-tuning, highlighting the effectiveness of repurposing visual foundation models for 4D understanding. Code and more results are available at: https://davidya099.github.io/uni4d.", + "bbox": [ + 86, + 537, + 485, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 813, + 220, + 828 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Over the past two years, many visual foundation models have emerged [4, 21, 23, 25, 27, 31, 44, 58, 63], achieving high accuracy on tasks like depth prediction, segmentation, human parsing, normal estimation, few-view recon", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "struction, and motion tracking. These models leverage supervised learning on large, diverse datasets, achieving impressive accuracy and remarkable generalization. However, these advances have not translated to 4D (time + geometry) modeling, a longstanding challenge in computer vision. We see two main reasons: first, collecting high-quality, ground-truth 4D data from real-world environments remains complex and resource-intensive. Second, 4D understanding is a holistic problem involving interconnected tasks like camera pose estimation, 3D reconstruction, and dynamics tracking. Although each subtask shows progress, data-driven cues remain noisy, and more importantly, unifying them synergistically for holistic modeling remains challenging. This paper seeks to answer: Can we harness the success of visual foundation models for dynamic 4D modeling?", + "bbox": [ + 511, + 506, + 906, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose Uni4D, a novel framework that reconstructs 4D scenes from a single video captured in the wild. Our method integrates data-driven foundation models and conventional model-driven dynamic structure-from-motion, combining data-driven cues and model-based knowledge synergistically. Our intuition is simple: each data-driven visual cue, such as video segmentation, pixel-level motion tracking, and video depth, is a partial projection from the 4D world to the 2D video. The key is to create a 4D scene representation that coherently aligns with each cue while incorporating strong prior knowledge", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1116", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of real-world motion and shape to resolve temporary inconsistencies. To achieve this, we take an energy-minimization perspective, framing the problem as an optimization task that jointly infers camera poses, static and dynamic geometry, and motion. The dynamic 4D modeling framework is complex, as it involves various optimization variables, visual cues, and constraints. To overcome the challenge of joint reasoning, we carefully design a novel three-stage, divide-and-conquer pipeline that progressively incorporates camera poses, static geometry, and dynamic geometry and motion into the optimization framework. Uni4D leverages pretrained foundation models across different tasks, requiring no task-specific retraining or fine-tuning. This design sidesteps the need for 4D ground-truth data, a major challenge in the field. Through incorporating strong priors on geometry and dynamic motion, our method produces realistic 4D scenes that are coherent across space and time while maintaining high accuracy.", + "bbox": [ + 89, + 90, + 480, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We demonstrate the effectiveness of Uni4D on various datasets. As shown in Fig. 1, our method effectively recovers the clean geometry and motion of the scene, as well as camera trajectories, from a single video. As shown in Fig. 2, our model outperforms all dynamic 4D modeling baselines in both camera pose and geometry.", + "bbox": [ + 89, + 363, + 480, + 453 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 465, + 241, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Structure from Motion and SLAM. Structure from Motion (SfM) is a classical problem in computer vision that aims to jointly recover camera parameters and 3D structure from images [2, 10, 41, 47, 48]. Simultaneous localization and mapping (SLAM) is a similar problem that focuses on real-time efficiency in an online setting [11, 13, 38, 53]. The core idea for many approaches to SfM and SLAM is to optimize a form of reprojection error with respect to both camera parameters and 3D points, also known as bundle adjustment [55]. This relies on the key assumption that the scene being viewed is rigid (\"static\"), making these approaches not applicable to dynamic scenes.", + "bbox": [ + 89, + 491, + 482, + 671 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/78379f0a216985cae1c7c7b2d32abd724fde2148f237936d74eb3b1738427295.jpg", + "image_caption": [ + "Figure 2. Uni4D outperforms other recent 4D modeling methods in both camera pose and geometry accuracy on the Sintel dataset." + ], + "image_footnote": [], + "bbox": [ + 148, + 672, + 415, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "4D Understanding. The problem of non-rigid structure from motion is highly ill-posed. Previous works on non-rigid SfM leveraged various forms of priors or weaker assumptions on shape and or motion to make the problem solvable [3, 6, 28, 29, 70]. With the rise of modern segmentation models, many works have studied category-specific articulation and rigidity priors, focusing on object categories such as humans [17, 49, 52, 64], animals [60-62, 66], and vehicles [18, 37]. The generality of such methods is limited by their category-specific nature. Recently, several works have also begun to leverage trained neural networks as priors for 4D understanding in a more freeform manner [30, 65, 68, 69]. However, these methods involve either training a network [65, 69] or optimizing specific layers of an existing network [30, 68], making it difficult to integrate newer models into the pipeline. Our method takes the strategy further and integrates pretrained models in a completely modular manner, fully unleashing their generalization capabilities and allowing for seamless integration of newer models. Note that various recent works have also studied 4D reconstruction in an optimization-based framework [15, 32, 35, 57, 59]. However, these methods have additional rendering capacities and focus on rendering metrics, while we focus on recovering high-quality geometry.", + "bbox": [ + 511, + 90, + 906, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual Foundation Models. In recent years, a number of visual foundation models have been developed, achieving remarkable performance on tasks like depth estimation [5, 20, 23, 44, 63], detection and segmentation [12, 27, 36, 46], human parsing [25], surface normal estimation [4, 14], few-view reconstruction [34, 58], and point tracking [21, 22, 31]. Our insight is that nearly all of these models have the potential to contribute towards holistic 4D understanding, and harnessing them in a unifying framework can advance the state-of-the-art in tasks such as non-rigid structure from motion. To this end, we integrate the following pretrained models in Uni4D: UniDepthv2 [45] for geometry initialization, CoTracker3 [22] for correspondence initialization, and a collection of Recognize Anything Model [67], ChatGPT [1], Grounding-SAM [27, 36], and DEVA [9] for dynamic object segmentation. Through multi-stage optimization with a few regularizing priors, we are able to use these models to extract accurate pose and 4D geometry from monocular video.", + "bbox": [ + 511, + 470, + 908, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 768, + 606, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we are interested in recovering 4D geometry and camera parameters from a monocular casual video. Our model is built on the intuition that the 2D visual cues from the video can be seen as perspective projections of 4D geometry and motion, where the video depth represents a projection of 4D geometry, 2D dense tracking corresponds to a projection of 4D motion, and segmentation reflects a", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1117", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d8ce7c042acb1ab62c4a2977507cfb5462ef20ee69131d51165d7e7bf959e86a.jpg", + "image_caption": [ + "Figure 3. Given a casually captured video, Uni4D exploits visual foundation models to extract dynamic segmentation, video depth, and motion tracks. Static geometry and poses are obtained through tracklet-based structure-from-motion along with camera motion priors. Dynamic geometry is improved through nonrigid bundle adjustment and scene motion priors. A final fusion densifies geometry to obtain high quality 4D reconstruction." + ], + "image_footnote": [], + "bbox": [ + 130, + 80, + 867, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "projection of the 4D dynamic object silhouette. We propose a novel, training-free, foundation-model-based energy minimization scheme that leverages visual cues to jointly infer camera poses, geometry, and motion, enabling generalization across diverse videos.", + "bbox": [ + 89, + 382, + 483, + 457 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Pretrained Visual Cues", + "text_level": 1, + "bbox": [ + 89, + 469, + 305, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first exploit foundation models to extract visual cues, including dynamic segmentation, video depth, and motion tracking. All of the models are pre-trained on large and diverse datasets and exhibit strong generalization abilities.", + "bbox": [ + 89, + 492, + 483, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Segmentation. Recognizing and segmenting dynamic objects is crucial for 4D scene understanding. We leverage the latest advancements in video semantic segmentation and tracking to estimate dynamic objects over time. First, we use RAM [67] to identify semantic classes in the video. These classes are then filtered through GPT-4o [1] to exclude static and background elements (e.g., buildings, poles), retaining only dynamic objects (e.g., humans, animals, vehicles). Next, Grounding-SAM [27, 36] performs segmentation at each keyframe, and DEVA [9] tracks these segments over time, resulting in accurate dynamic video segmentation $\\{\\mathbf{M}_{\\mathbf{t}}\\}_{t = 0}^{T}$", + "bbox": [ + 89, + 575, + 483, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dense Motion Tracking. We use dense pixel tracking to establish correspondences over time, which serve as motion cues to assist inference of geometry reconstruction and dynamic object 3D motion. Unlike traditional 4D reconstruction methods relying on optical flow [65, 68], dense motion tracklets yield more correspondence pairs across large viewpoints and structural changes. Pixel tracking also outperforms sparse matching in density and surpasses flow", + "bbox": [ + 89, + 779, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "propagation in robustness, making it ideal for dynamic 4D reconstruction. We utilize Co-TrackerV3 [22] for its robustness, as it employs a 4D cost volume with an attention mechanism to track 2D points through occlusions, recently proving to be effective for dynamic neural rendering. We apply Co-Tracker bi-directionally on a dense grid every 10 frames to ensure thorough coverage. We filter and classify tracklets using segmentation masks yielding a set of correspondent point trajectories $\\{\\mathbf{Z}_k\\in \\mathbb{R}^{T\\times 2}\\}_{k = 0}^K$ at visible time steps determined by Co-Tracker.", + "bbox": [ + 511, + 382, + 906, + 534 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Depth Estimation. Monocular depth reasoning, while insufficient as a standalone tool for complete 4D geometry recovery, provides strong geometry initialization cues. In our model, we use UniDepthV2 [45], a monocular depth estimation network, to estimate an initial depth map, $\\{\\mathbf{D}_t\\}_{t=0}^T$ , and initial camera intrinsics, $\\mathbf{K}_{\\mathrm{init}}$ .", + "bbox": [ + 511, + 551, + 906, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Energy Formulation", + "text_level": 1, + "bbox": [ + 511, + 652, + 705, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now describe the energy formulation of our proposed dynamic 4D reconstruction framework. Let $\\mathcal{M} = \\{\\mathbf{M}_t\\}_{t=0}^T$ , $\\mathcal{D} = \\{\\mathbf{D}_t\\}_{t=0}^T$ , $\\mathcal{Z} = \\{\\mathbf{Z}_k\\}_{k=0}^K$ be the input dynamic segmentation, monocular depth, dense motion trajectory extracted from the input video $\\mathcal{I} = \\{\\mathbf{I}_t\\}_{t=0}^T$ . Formally, our goal is to obtain camera parameters $\\mathcal{C}$ , namely poses $\\mathcal{T}$ and intrinsics $\\mathbf{K}$ , and a set of 4D point clouds $\\mathcal{P}$ containing both dynamic and static parts separately as $\\mathcal{P} = \\{\\mathbf{P}_{\\mathrm{static}}, \\mathcal{P}_{\\mathrm{dyn}}\\}$ . Here, $\\mathbf{P}_{\\mathrm{static}}$ does not change over time and $\\mathcal{P}_{\\mathrm{dyn}} = \\{\\mathbf{p}_k \\in \\mathbb{R}^{T \\times 3}\\}_k$ is a temporally varying point cloud, where each $\\mathbf{p}_k$ is a dynamic point trajectory. We represent the camera poses as rigid transforms $\\mathcal{T} = \\{\\xi_t \\in \\mathbb{S}\\mathbb{E}(3)\\}_{t=0}^T$ , and parameterize the rotations as $\\mathfrak{so}(3)$ rotation vectors as a minimal representation for easy optimization. We assume all frames share the same intrinsic", + "bbox": [ + 511, + 674, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1118", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/834930e04012363df5016111a3bee8ce3495a8ca2a9da60075df68c2210066de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 169, + 114, + 334, + 185 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/247789a78f903db6aad089be37d4686ba23524b29dc2e5de20bf5239cac878ab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 171, + 186, + 333, + 257 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a33c4e9277c71089a2a2638ef4762215479de867b6a9f52deb7c0149ea7a520e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 171, + 261, + 334, + 330 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6226712897f8335ff927d38bc98b15ef6c348ccf59843de21125380284717855.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 171, + 335, + 334, + 405 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8c5c854b24ae6d177707345b61a44d8ec348c4c04e16912680c77c64b54cda10.jpg", + "image_caption": [ + "Input Video" + ], + "image_footnote": [], + "bbox": [ + 171, + 411, + 334, + 484 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fe38f0cf37f9a93c7f58ef1aae41a21b23169af57cd0db1cf0c864591fee9322.jpg", + "image_caption": [ + "CasualSAM Static" + ], + "image_footnote": [], + "bbox": [ + 356, + 98, + 500, + 229 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b4771ac96413e646bb92cc1089cf5cd59d9c40ef0f49f0bca1808dd5dd81a2a5.jpg", + "image_caption": [ + "CasualSAM 4D" + ], + "image_footnote": [], + "bbox": [ + 534, + 98, + 671, + 226 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f553aeebeeafcf1c8636677221770b21e1c0ed5c91d29dd0cce0d0ca7cf4c942.jpg", + "image_caption": [ + "CasualSAM Bird's Eye" + ], + "image_footnote": [], + "bbox": [ + 715, + 98, + 785, + 226 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/66690b622dd088d3bd6cc54dd4775d3cce24ee0f93180903dcf91e3fc2cdc27a.jpg", + "image_caption": [ + "Monst3R Static" + ], + "image_footnote": [], + "bbox": [ + 359, + 255, + 511, + 369 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e1feff5e9a91584c6778b64d7fdb3472be9a31f26d87c384e28245874c0b4a5d.jpg", + "image_caption": [ + "Monst3R 4D" + ], + "image_footnote": [], + "bbox": [ + 532, + 252, + 684, + 366 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/337696d7026100c9076a1e374816381bb8eea04e2651b08d51770c393f5aaf1d.jpg", + "image_caption": [ + "Monst3R Bird's Eye" + ], + "image_footnote": [], + "bbox": [ + 718, + 247, + 808, + 367 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2bf43608fd6f44f172acd05fa6a475d4f6cb08fabd19cc424f999da22d6a0e3d.jpg", + "image_caption": [ + "Ours Static" + ], + "image_footnote": [], + "bbox": [ + 361, + 383, + 519, + 500 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/73c9f62c338b5cbc2e05ca2f7a75ae1fb92310322955f64b4b109f3f8e341690.jpg", + "image_caption": [ + "Ours 4D", + "Figure 4. Qualitative results of 4D reconstruction on the DAVIS [43] dataset. CasualSAM [68] suffers from slanted geometry, and Monst3R [65] has unclear geometry and does not resolve conflicts from multiple views (note the wall in the bird's-eye view). Both CasualSAM and Monst3R lack clean dynamic reconstruction and segmentation. Uni4D achieves a realistic layout, thanks to joint optimization, and provides accurate dynamic segmentation and reconstruction by leveraging foundation visual models as cues." + ], + "image_footnote": [], + "bbox": [ + 531, + 382, + 689, + 494 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cab83464e20a235e71f8960502452634221a646975c294da08ca627e07266bb1.jpg", + "image_caption": [ + "Ours Bird's Eye" + ], + "image_footnote": [], + "bbox": [ + 728, + 383, + 803, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "matrix $\\mathbf{K}$ where we optimize focal lengths $f_{x}$ and $f_{y}$ . We formulate the 4D joint reasoning problem as minimization of the following energy function:", + "bbox": [ + 89, + 590, + 482, + 636 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\mathrm {B A}} (\\mathcal {C}, \\mathbf {P} _ {\\text {s t a t i c}}) + E _ {\\mathrm {N R}} (\\mathcal {P} _ {\\mathrm {d y n}}) + E _ {\\text {m o t i o n}} (\\mathcal {P} _ {\\mathrm {d y n}}) + E _ {\\text {c a m}} (\\mathcal {T}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 648, + 482, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $E_{\\mathrm{BA}}(\\mathcal{C}, \\mathbf{P}_{\\mathrm{static}})$ is a bundle adjustment term that measures the discrepancy between static-region correspondences and the static 3D structure $\\mathbf{P}_{\\mathrm{static}}$ through perspective reprojection. $E_{\\mathrm{NR}}(\\mathcal{P}_{\\mathrm{dyn}})$ is a non-rigid structure-from-motion energy term that measures the disagreement between the dynamic point cloud and their tracklet correspondences, $E_{\\mathrm{cam}}(\\mathcal{T})$ is a regularization term on the camera motion smoothness, and $E_{\\mathrm{motion}}(\\mathcal{P}_{\\mathrm{dyn}})$ is a regularization term on the dynamic structure and motion. Each energy term is involved in different stages of the optimization process, which will be described in Sec. 3.3.", + "bbox": [ + 89, + 680, + 482, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Static Bundle Adjustment Term. The bundle adjustment energy $E_{\\mathrm{BA}}(\\mathcal{C}, \\mathbf{P}_{\\mathrm{static}})$ measures the consistency between", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the pixel-level correspondences and the 3D structure of static scene elements. Given the input pixel tracks $\\mathcal{Z} = \\{\\mathbf{Z}_k\\}$ and video segmentation $\\mathcal{M}$ , we filter all tracks corresponding to static areas and minimize the distance between the projected pixel location and the observed pixel location:", + "bbox": [ + 511, + 590, + 906, + 666 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\mathrm {B A}} (\\mathcal {C}, \\mathbf {P} _ {\\text {s t a t i c}}; \\mathcal {Z}, \\mathcal {M}) = \\sum_ {\\mathbf {z} _ {k} \\in \\mathcal {M}} \\sum_ {t} \\mathbf {w} _ {\\mathbf {k}, \\mathbf {t}} \\| \\mathbf {z} _ {k, t} - \\pi_ {\\mathbf {K}} \\left(\\mathbf {p} _ {k}, \\boldsymbol {\\xi} _ {t}\\right) \\| _ {2} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 699, + 926, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{z}_{k,t}$ is the $k$ -th 3D point's corresponding pixel track's 2D coordinates at time $t$ , $\\mathbf{w_{k,t}} \\in \\{0,1\\}$ is a visibility indicator and $\\pi_{\\mathbf{K}}$ is the perspective projection function.", + "bbox": [ + 511, + 744, + 906, + 790 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Non-Rigid Bundle Adjustment Term. For dynamic objects, we impose a nonrigid bundle adjustment term, $E_{\\mathrm{NR}}(\\mathcal{P}_{\\mathrm{dyn}})$ , which measures the discrepancy between the dynamic point cloud and pixel tracklets. Here, each pixel tracklet corresponds to a dynamic 3D point sequence,", + "bbox": [ + 511, + 824, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1119", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f37dc1e04d2f370fb41c79d8847603ba08af17dfc7808c0fd6a4c38e891423e0.jpg", + "image_caption": [ + "Figure 5. Qualitative results on Bonn dataset. Both CasualSAM and MonST3R have trailing artifacts and incorrect dynamic estimations. Uni4D provides clear dynamic and static geometry." + ], + "image_footnote": [], + "bbox": [ + 98, + 93, + 472, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\{\\mathbf{p}_{k,t}\\}_{t}$ , optimized for each observed tracklet:", + "bbox": [ + 89, + 306, + 393, + 321 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\mathrm {N R}} \\left(\\mathcal {P} _ {\\mathrm {d y n}}; \\mathcal {C}, \\mathcal {Z}, \\mathcal {M}\\right) = \\sum_ {\\mathbf {z} _ {k} \\in \\mathcal {M}} \\sum_ {t} \\mathbf {w} _ {\\mathbf {k}, \\mathbf {t}} \\left\\| \\mathbf {z} _ {k, t} - \\pi_ {\\mathbf {K}} \\left(\\mathbf {p} _ {k, t}, \\boldsymbol {\\xi} _ {t}\\right) \\right\\| _ {2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 329, + 501, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{p}_{k,t} \\in \\mathbb{R}^3$ is the $k$ -th dynamic point's location at $t$ .", + "bbox": [ + 89, + 375, + 470, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Camera Motion Prior. Considering that our input is a video, we incorporate a temporal smoothness prior on camera poses that penalizes sudden changes in relative pose: $\\pmb{\\xi}_{t\\rightarrow t + 1} = \\pmb{\\xi}_{t + 1}^{-1}\\cdot \\pmb{\\xi}_t$ . We reweight this term based on the magnitude of the relative motion: intuitively, if the relative motion is large, we penalize change rates in relative motion less; if the relative motion is small, we apply a higher penalty on change rate. Formally, we have:", + "bbox": [ + 89, + 407, + 483, + 527 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\mathrm {c a m}} (\\mathcal {T}) = \\sum_ {t} E _ {\\mathrm {r o t}} \\left(\\mathbf {R} _ {t - 1, t, t + 1}\\right) + \\sum_ {t} E _ {\\mathrm {t r a n s}} \\left(\\mathbf {t} _ {t - 1, t, t + 1}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 535, + 480, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $E_{\\mathrm{rot}}(\\mathbf{R}_{t - 1,t,t + 1}) = \\frac{2\\|\\mathrm{rad}(\\mathbf{R}_{t\\to t + 1}) - \\mathrm{rad}(\\mathbf{R}_{t - 1\\to t})\\|}{\\|\\mathrm{rad}(\\mathbf{R}_{t - 1\\to t})\\| + \\|\\mathrm{rad}(\\mathbf{R}_{t\\to t + 1})\\|}$ and $E_{\\mathrm{trans}}(\\mathbf{t}_{t - 1,t,t + 1}) = \\frac{2\\|\\mathbf{t}_{t\\to t + 1} - \\mathbf{t}_{t - 1\\to t}\\|}{\\|\\mathbf{t}_{t - 1\\to t}\\| + \\|\\mathbf{t}_{t\\to t + 1}\\|}$ ; rad converts the rotation matrix into absolute radians.", + "bbox": [ + 89, + 575, + 480, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dynamic Motion Prior. $E_{\\mathrm{motion}}(\\mathcal{P}_{\\mathrm{dyn}})$ is a regularization term that encodes the characteristics of the dynamic structure. It contains two prior terms that are used to regularize the dynamic structure, both of which have demonstrated effectiveness in previous work [39, 50]:", + "bbox": [ + 89, + 648, + 483, + 724 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\text {m o t i o n}} \\left(\\mathcal {P} _ {\\text {d y n}}\\right) = E _ {\\text {a r a p}} \\left(\\mathcal {P} _ {\\text {d y n}}\\right) + E _ {\\text {s m o o t h}} \\left(\\mathcal {P} _ {\\text {d y n}}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 732, + 480, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$E_{\\mathrm{arap}}$ is an as-rigid-as-possible (ARAP) [50] prior that penalizes extreme deformations that compromise local rigidity. Specifically, we obtain the nearest neighbors of each dynamic control point $k$ by applying KNN over the other tracks and enforce that the relative distances between these close-by pairs do not undergo sudden changes:", + "bbox": [ + 89, + 755, + 483, + 845 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\text {a r a p}} = \\sum_ {\\mathbf {t}} \\sum_ {(k, m)} \\mathbf {w} _ {\\mathbf {k m}, \\mathbf {t}} \\| d \\left(\\mathbf {p} _ {\\mathbf {k}, \\mathbf {t}}, \\mathbf {p} _ {m, t}\\right) - d \\left(\\mathbf {p} _ {\\mathbf {k}, \\mathbf {t} + 1}, \\mathbf {p} _ {m, t + 1}\\right) \\| _ {2} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 853, + 506, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d(\\cdot, \\cdot)$ is the L2 distance and $\\mathbf{w}_{\\mathbf{k}\\mathbf{m},\\mathbf{t}} = 1$ if all relevant points are visible.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$E_{\\mathrm{smooth}}$ is a simple smoothness term that promotes temporal smoothness for the dynamic point cloud:", + "bbox": [ + 511, + 121, + 903, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\text {s m o o t h}} = \\sum_ {\\mathbf {t}} \\sum_ {\\mathbf {p} _ {k} \\in \\mathcal {P} _ {\\text {d y n}}} \\mathbf {w} _ {\\mathbf {k}, \\mathbf {t}} \\| \\mathbf {p} _ {k, t} - \\mathbf {p} _ {k, t + 1} \\| _ {2}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 157, + 903, + 190 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Despite simplicity, both motion terms are crucial in our formulation, as they significantly reduce ambiguities in 4D dynamic structure estimation, which is highly ill-posed. Unlike other methods, we do not assume strong model-based motion priors, such as rigid motion [37], articulated motion [61], or a linear motion basis [57].", + "bbox": [ + 511, + 196, + 906, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Inference", + "text_level": 1, + "bbox": [ + 513, + 294, + 622, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Directly minimizing the energy defined in Eq. 1 is nontrivial, as our energy function is highly non-linear and involves millions of free variables. To address this, we developed a three-stage optimization pipeline, enabling us to minimize the energy and estimate the scene variables in a divide-and-conquer fashion.", + "bbox": [ + 511, + 315, + 905, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage 1: Camera Initialization. We start by initializing camera parameters. Combining video depth estimation $\\mathcal{D}$ and dense pixel motion $\\mathcal{Z}$ allows us to establish 2D-to-3D correspondences. This allows us to initialize and tune $\\mathcal{C}$ by minimizing the following energy function with respect to camera parameters only. Specifically, we can unproject each video frame's depth at time $t$ back to 3D and minimize the following energy function:", + "bbox": [ + 511, + 422, + 905, + 545 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathbf {C}} \\sum_ {\\left(t ^ {\\prime}, t\\right)} \\sum_ {\\mathbf {z} _ {k} \\in \\neg \\mathcal {M}} \\| \\mathbf {z} _ {k, t ^ {\\prime}} - \\pi_ {\\mathbf {K}} \\left(\\pi_ {\\mathbf {K}} ^ {- 1} \\left(\\mathbf {z} _ {k, t}, \\mathbf {D} _ {t}, \\boldsymbol {\\xi} _ {t}\\right), \\boldsymbol {\\xi} _ {t ^ {\\prime}}\\right) \\| _ {2} ^ {2} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 550, + 903, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pi_{\\mathbf{K}}^{-1}$ is the unprojection function that maps 2D coordinates into 3D world coordinates using estimated depth $\\mathbf{D}_t$ . We perform this over all pairs within a temporal sliding window of 5 frames, producing a good initial pose estimate as shown in Tab. 4. Given camera initialization $\\hat{\\mathcal{C}}$ , we directly unproject our depth prediction into a common world coordinate system, which provides an initial 4D structure $\\hat{\\mathcal{P}}$ . This is used as initialization for later optimization.", + "bbox": [ + 511, + 590, + 905, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage 2: Bundle Adjustment. Our second stage jointly optimizes camera pose and static geometry by minimizing the static component-related energy in a bundle adjustment fashion. Formally speaking, we solve the following:", + "bbox": [ + 511, + 729, + 905, + 790 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathcal {C}, \\mathbf {P} _ {\\text {s t a t i c}}} E _ {\\mathrm {B A}} \\left(\\mathcal {C}, \\mathbf {P} _ {\\text {s t a t i c}}; \\mathcal {Z}, \\mathcal {M}\\right) + E _ {\\mathrm {c a m}} (\\mathcal {T}) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 797, + 903, + 820 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By enforcing consistency with each other, this improves both the static geometry and the camera pose quality. We perform a final scene integration by unprojecting correspondences into 3D using improved pose and filtering outlier noisy points in 3D.", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1120", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/10cc49294d795eb17d072b528e4bcc702067db7f77b979e3dd24a1c820fed94c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 130, + 87, + 236, + 135 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/44c81bbada89dc995ea93df0de7375a4d7c7d14368c22f6541d55b50ddb8dc54.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 133, + 138, + 235, + 185 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/567f9c46a8bcb0eb644ada89f30f61d76beec164c931a9bd006d99697bf7ebef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 133, + 189, + 235, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cb8fbb9d5c7a86cd616d357bad64832d0e69574f24b8e9c05e046eeb10b7a77b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 132, + 238, + 235, + 284 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d948dd0ff7d74a774ceea539861a9b0ac83859daeb725cc5ae424066c692d138.jpg", + "image_caption": [ + "Input Video" + ], + "image_footnote": [], + "bbox": [ + 130, + 287, + 233, + 334 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b5385671bef8f8af1d9d37ef6bde1ce7695c1f80551782313b3b10be04dda362.jpg", + "image_caption": [ + "CasualSAM Static" + ], + "image_footnote": [], + "bbox": [ + 254, + 88, + 431, + 162 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/506995d07fba5dcfed6a0d409aaba68be698542faabf10452d3d09f83f4e9070.jpg", + "image_caption": [ + "CasualSAM 4D" + ], + "image_footnote": [], + "bbox": [ + 439, + 88, + 645, + 164 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/591ab479303c6367c072c19b78dab2608794cd804efc771dc1f3c68e26d36dc6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 683, + 79, + 836, + 159 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/24c11c9c46de54934afd8a13c58dd489327ef5842ababa0f72f6b2aa4d17221e.jpg", + "image_caption": [ + "MonST3R Static" + ], + "image_footnote": [], + "bbox": [ + 253, + 183, + 436, + 257 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/49ed163bd00d2d62977fe1cc0c8f6b8fd00113a51b6fa8d62158e031c4b3a5ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 452, + 178, + 658, + 255 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/231d89d3a4532147a2c9258877d0cc9c67fcafc5d2062518ed83785bd86cea0c.jpg", + "image_caption": [ + "Ours 4D" + ], + "image_footnote": [], + "bbox": [ + 460, + 268, + 661, + 342 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6475bb994ced197f35942cf9a35d7080d4e585c855a1fd5647a7f206d4fde379.jpg", + "image_caption": [ + "CasualSAM Bird's Eye", + "MonST3R Bird's Eye" + ], + "image_footnote": [], + "bbox": [ + 728, + 172, + 844, + 239 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f4500869aa7c95b53fa95ede09d0b7c12c819f447f550b5d0f3b1efcfbb6eeac.jpg", + "image_caption": [ + "Ours Bird's Eye" + ], + "image_footnote": [], + "bbox": [ + 736, + 260, + 841, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Stage 3: Non-Rigid Bundle Adjustment. Given the estimated camera pose, our third stage focuses on inferring dynamic structure. Note that we freeze camera parameters in this stage, as we find that incorrect geometry and motion evidence often harm camera pose estimation rather than improve it. Additionally, enabling camera pose optimization introduces extra flexibility in this ill-posed problem, harming robustness. Formally speaking, we solve the following:", + "bbox": [ + 89, + 426, + 483, + 547 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathcal {P} _ {\\mathrm {d y n}}} E _ {\\mathrm {N R}} \\left(\\mathcal {P} _ {\\mathrm {d y n}}; \\mathbf {C}, \\mathcal {Z}, \\mathcal {M}\\right) + E _ {\\text {m o t i o n}} \\left(\\mathcal {P} _ {\\mathrm {d y n}}\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 559, + 480, + 584 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We initialize $\\mathcal{P}_{\\mathrm{dyn}}$ using video depth and our optimized camera pose from Stage 2. We scale $E_{\\mathrm{smooth}}$ and $E_{\\mathrm{arap}}$ with constants 10 and 100 respectively which we found empirically led to better dynamics. This energy optimization might still leave some high-energy noisy points, often from incorrect cues, motion boundaries, or occlusions. We filter these outliers based on their energy values in a final step.", + "bbox": [ + 89, + 595, + 483, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fusion. From our energy minimization, we acquire a semi-dense dynamic and static point cloud along with camera parameters. To further densify the point cloud, enabling each pixel to correspond to a 3D point, we perform depth-based interpolation by computing a scale offset. Details can be found in the supplementary.", + "bbox": [ + 89, + 719, + 483, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To unify our outputs into a consistent 4D representation, we use our camera parameters and static masks to project the aligned depth maps into world coordinates. To handle noisy depth values at boundaries, we create an edge mask filter by thresholding gradients of the depth maps.", + "bbox": [ + 89, + 810, + 483, + 885 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This process enables us to update the depth value for", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/13f5c037775fe836d11053ed5c70993c5e8280c1391b9dd51fb5f24a02aadcb1.jpg", + "image_caption": [ + "Figure 6. Qualitative results of 4D reconstruction on DAVIS dataset. CasualSAM [68] distorts the room geometry as evident from the bird's eye view. Dynamic elements have inconsistent geometry over time. MonST3R [65] has noisy static geometry in the room's far corner and incomplete dynamic object geometry evident in the green highlight. Uni4D has the cleanest dynamic object reconstruction and segmentation results, with geometrically accurate room shapes as evident from the bird's eye view.", + "Figure 7. Qualitative results of 4D reconstruction on Sintel dataset. We highlight and contrast performance on (1) dynamic objects and (2) geometry on planar surfaces." + ], + "image_footnote": [], + "bbox": [ + 514, + 426, + 906, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "each high-confidence pixel, which we can then reproject into 3D space to obtain the final dense dynamic geometry reconstruction result, as shown in Fig. 1 and Fig. 3.", + "bbox": [ + 511, + 676, + 906, + 723 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 736, + 645, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Uni4D estimates camera pose, depth, and 3D motion from a single video. We perform experiments evaluating its performance with respect to baselines on these tasks.", + "bbox": [ + 511, + 762, + 905, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 818, + 730, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For optimization, we use the Adam optimizer with ReduceLROnPlateau and EarlyStopping in PyTorch [26, 42]. We perform 600 iterations per sliding window in stage 1, 2000 iterations in stage 2, and 1000 iterations in stage 3.", + "bbox": [ + 511, + 840, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1121", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3e2f1b73546d2317601a1657f82e06ddda6c4111716074ec0c9d7f954ca8d1d9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CategoryMethodSintelTUM-dynamicsBonn
ATE ↓RPE trans ↓RPE rot ↓ATE ↓RPE trans ↓RPE rot ↓ATE ↓RPE trans ↓RPE rot ↓
Pose onlyDPVO*[54]0.1710.0631.2910.0190.0140.4060.0220.0140.913
LEAP-VO*[8]0.0350.0651.6690.0250.0312.8430.0370.0140.844
Joint depth & poseRobust-CVD[30]0.3680.1533.4620.0960.0272.5900.0850.0180.803
CasualSAM[68]0.1370.0390.6300.0360.0180.7450.0240.0140.849
Monst3R[65]0.1080.0430.7290.1080.0221.3710.0230.0110.807
Uni4D0.1100.0320.3380.0120.0040.3350.0170.0100.818
Uni4D*0.0920.0330.1410.0120.0040.3310.0160.0100.817
", + "bbox": [ + 155, + 88, + 836, + 256 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c6af3f41ea4fbab950aa5afa8338e399727bc37706787c6af0f63060c8eb66ed.jpg", + "table_caption": [ + "Table 1. Camera Pose Evaluation on Sintel, TUM-dynamic, and Bonn datasets. We bold and underline the best and second best results respectively. * indicates known camera intrinsic." + ], + "table_footnote": [], + "table_body": "
AlignmentCategoryMethodSintelBonnKITTI
Abs Rel ↓ δ<1.25 ↑Abs Rel ↓ δ<1.25 ↑Abs Rel ↓ δ<1.25 ↑
Per-sequence scale & shiftSingle-frame depthMetric3D[19]0.20571.90.04498.50.03998.8
Depth-pro[5]0.28060.50.04998.60.08094.2
Unidepth[44]0.19872.80.04098.50.03898.8
Video depthDepthCrafter[20]0.23169.00.06597.60.11288.4
Joint video depth & poseRobust-CVD[30]0.35849.70.10889.80.18272.9
CasualSAM[68]0.29256.90.06996.60.11388.3
Monst3r[65]0.35852.10.06095.00.08591.9
Uni4D0.21672.50.03898.30.09889.7
Per-sequence scaleJoint depth & poseMonst3r[65]0.34455.90.04198.20.08991.4
Joint depth & poseUni4D0.28964.90.03898.30.08693.3
", + "bbox": [ + 173, + 301, + 820, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Video depth evaluation on Sintel, Bonn, and KITTI datasets. We bold and underline the best and second best results respectively.", + "bbox": [ + 89, + 516, + 906, + 531 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We initialize the learning rate at $1 \\times 10^{-3}$ for stage 1, and $1 \\times 10^{-2}$ for stages 2 and 3, reducing all to $1 \\times 10^{-4}$ . Our entire framework takes roughly 5 minutes for a 50-frame video on a RTX A6000 GPU. We include a detailed runtime breakdown in the supplementary. We run all baselines on the datasets using their official implementations and hyperparameters. Co-Trackers [22] are initialized at a dense $50 \\times 50$ grid, with a $75 \\times 75$ grid for Sintel to handle its large camera perspective change. All optimization hyperparameters are kept the same for all runs on all datasets.", + "bbox": [ + 89, + 537, + 485, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Pose Estimation", + "text_level": 1, + "bbox": [ + 89, + 732, + 250, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We compare with several recent methods for pose estimation in dynamic scenes. LEAPVO [8] and DPVO [54] are learning-based visual odometry methods. Robust-CVD [30] optimizes for pose and depth deformation through an SFM pipeline. CasualSAM [68] further improves on the idea by directly finetuning network weights along with a novel uncertainty formulation. Monst3R [65] is a very recent work that fine-tunes DUSt3R [58] for 4D reconstruction through PnP [33].", + "bbox": [ + 89, + 763, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Benchmarks and metrics. We evaluate pose estimation on three dynamic datasets: Sintel [7], TUM-dynamics [51], and Bonn [40]. We follow LEAP-VO's evaluation split for Sintel and use all videos from TUM-dynamics and Bonn. Following MonST3R [65], we subsample every 3 frames from the first 270 frames from TUM-dynamics to save compute. We follow the standard pose evaluation process of aligning camera poses with Umeyama alignment [56]. We report Absolute Translation Error (ATE), Relative Translation and Rotation Error (RPE trans and RPE rot).", + "bbox": [ + 511, + 539, + 906, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. As reported in Tab. 1, Uni4D achieves competitive results across all metrics and datasets, highlighting the generalizability and performance of our pipeline. Uni4D is flexible to the availability of camera intrinsics, showing further improvement with known camera intrinsics. Training-based approaches such as LEAP-VO achieve good results on synthetic dataset like Sintel but does not generalize well to real-world datasets. Our method matches performance with MonST3R [65] on Sintel and achieves significantly better results on real-world datasets even compared to methods using ground-truth intrinsics.", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1122", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Video Depth Evaluation", + "text_level": 1, + "bbox": [ + 89, + 90, + 310, + 107 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Baselines. For video depth accuracy evaluations, we focus on the top performing metric depth estimators, namely Metric3Dv2 [19], Depth-Pro [63], DepthCrafter [20] and Unidepth [44]. Metric depth estimators are trained without scale and shift alignment, making them strong baselines for video depth estimation. We also include the same baselines as our pose evaluations for joint 4D modeling approaches, namely CasualSAM [68], RCVD [30], and MonST3R [65].", + "bbox": [ + 89, + 113, + 483, + 236 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "**Benchmarks.** We evaluate video depth estimates on Sintel [7], Bonn [40] and KITTI [16]. We follow standard video depth evaluation protocols [20] of aligning global shift and scale to predicted video depthmaps. We report the absolute relative error (Abs Rel) and percent of inlier points $(\\delta < 1.25)$ . All methods are aligned in disparity space using the same least-squares alignment. We additionally report scale-aligned depth estimates similar to MonST3R.", + "bbox": [ + 89, + 258, + 483, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. By leveraging the strong performance of single-frame metric depth estimation models, Tab. 2 shows Uni4D achieves competitive depth estimation results, with superior performance among joint depth and pose methods, and closely matching the performance of single-frame depth estimation models on some datasets. With per-sequence scale, our model produces more accurate depth estimates across all datasets compared to recently released MonST3R. Our method closely retains depth estimation accuracy of underlying Unidepth depthmaps, while significantly improving its depth consistency shown in Sec. 4.5.", + "bbox": [ + 89, + 405, + 483, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative", + "text_level": 1, + "bbox": [ + 89, + 584, + 212, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further show qualitative results of our reconstructions on the DAVIS [43] dataset. We apply MonST3R's own confidence-guided fusion for their reconstruction. We fuse casualSAM's depthmaps with dynamic masking by thresholding its uncertainty prediction to output 4D reconstruction. We use highlights to indicate dynamic objects at different timesteps. Throughout our qualitative results in Figs. 4,5,6,7, CasualSAM produces warped geometry and poor dynamic segmentations. MonST3R produces poor geometry in far regions and noisy dynamic masks and shapes. Our model produces the cleanest dynamic segmentations, along with the best dynamic and static geometry.", + "bbox": [ + 89, + 607, + 483, + 789 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 801, + 243, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We ablate our multi-stage optimization in Tab. 4, highlighting the importance of each stage. Joint 4D optimization adds complexity that requires strong initialization for optimal convergence. Stage 1 introduces drift, which Stage 2 rectifies, resulting in superior final pose estimates.", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c3abe6cd08cb56ffe4c3f2a6571664e3a5d964e710d9fa4eb0e48ed1b76d773b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 88, + 586, + 119 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ec1b32e6e75c1904c19aa48ce29d046fa377c3c451f069abc03fdf843987f964.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 125, + 584, + 155 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/46c81b931261bd7fcf3ea4d0fa4374cc681d9a854a7d0d1fbbc2dd1f4010df9c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 157, + 584, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8146d9bcc40619c7660947afdfa42a6eb53c54925bbb21c16779482cd3af1e4c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 194, + 584, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f015f7ddf543d49d2921b7e0a7aa0397f1a6a6f97aa42ec529e4344d55b8188b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 229, + 584, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/33130ec47c4ae58385798c67a722d5cc9fb72cd91c723f5c676646fa527e2ddf.jpg", + "image_caption": [ + "Undepth Bird's Eye" + ], + "image_footnote": [], + "bbox": [ + 602, + 89, + 743, + 165 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c2d07328020d760daa3a3b630632d92ccabd6c77cbd36217bd5e49d205766de4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 175, + 674, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c70d37987e4634952b6f696822399651c262dd66e14990eded2cf6621b247df7.jpg", + "image_caption": [ + "Figure 8. Depth Consistency. Direct fusion of Unidepth [44] predictions causes misaligned scene geometry visible as a layered wall in bird's-eye view, and jittering dynamic motion. Through motion priors and alignment, Uni4D produces a thin, crisp wall structure and smooth dynamic motion." + ], + "image_footnote": [], + "bbox": [ + 601, + 196, + 732, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/89d421e2141e8b84a518a2e637656252a945fadb67be716a0fd90169ca401eed.jpg", + "image_caption": [ + "Ours Bird's Eye" + ], + "image_footnote": [], + "bbox": [ + 759, + 87, + 897, + 164 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1712028f1251fdd8f1300cc7b4ecbf2932d6fe6f20664c71f2da842ace651e8c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 810, + 174, + 828, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ebbae18983fa690b97b546328258ddda6a722c4caf4741570065b935c2d0b786.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 750, + 196, + 882, + 261 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b150306567e59caf85819334c052f58871724698268ae16c09049dfeceec8f8a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSintel
SC ↓δSC<0.01 ↑δSC<0.05 ↑
Unidepth [44]0.10931.876.8
Uni4D0.04369.388.1
", + "bbox": [ + 549, + 340, + 867, + 409 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2302e39e9bdb8c9f87c599fbf640158902d229d56f553e3071c9d313f060ac16.jpg", + "table_caption": [ + "Table 3. Video depth consistency on Sintel. Uni4D improves Unidepth in consistency." + ], + "table_footnote": [], + "table_body": "
MethodSintel
ATE ↓RPE trans ↓RPE rot ↓
Uni4D (stage 1 only)0.1500.0510.551
Uni4D (stage 2 only)0.5870.1934.12
Uni4D (full)0.1100.0320.338
", + "bbox": [ + 545, + 443, + 870, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Multi-stage Ablation. We evaluate pose estimation results on Sintel for both stage 1 and stage 2.", + "bbox": [ + 511, + 526, + 905, + 555 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Directly reprojecting Unidepth depth maps leads to flickering geometry. We ablate our fused depthmap quantitatively in Tab. 3, showing that Uni4D significantly improves consistency using Self-Consistency (SC) metrics [24]. SC measures depth errors between estimated and reprojected depth maps in static regions. Furthermore, we qualitatively show the effectiveness of Uni4D in rectifying Unidepth inconsistencies in Fig. 8. Our reconstruction achieves much better geometric and temporal consistency over Unidepth. Please refer to the supplementary for more ablative results regarding choice of foundation models.", + "bbox": [ + 511, + 571, + 906, + 738 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 753, + 633, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents Uni4D, a framework unifying visual foundation models and structured energy minimization for dynamic 4D modeling from casual video. Our key insight is to optimize a 4D representation that aligns with visual cues from foundation models while following motion and geometry priors. Results show state-of-the-art performance with superior visual quality on Sintel, DAVIS, TUM-Dynamics and Bonn datasets, without any retraining or fine-tuning.", + "bbox": [ + 511, + 779, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1123", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 209, + 187, + 224 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 3", + "[2] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building rome in a day. Communications of the ACM, 54 (10):105-112, 2011. 2", + "[3] Antonio Agudo, Melcior Pijoan, and Francesc Moreno-Noguer. Image collection pop-up: 3d reconstruction and clustering of rigid and non-rigid categories. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2607-2615, 2018. 2", + "[4] Gwangbin Bae and Andrew J Davison. Rethinking inductive biases for surface normal estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9535-9545, 2024. 1, 2", + "[5] Aleksei Bochkovskii, Amael Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 2, 7", + "[6] Christoph Bregler, Aaron Hertzmann, and Henning Biermann. Recovering non-rigid 3d shape from image streams. In Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662), pages 690-696. IEEE, 2000. 2", + "[7] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part VI 12, pages 611-625. Springer, 2012. 7, 8", + "[8] Weirong Chen, Le Chen, Rui Wang, and Marc Pollefeys. Leap-vo: Long-term effective any point tracking for visual odometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19844–19853, 2024. 7", + "[9] Ho Kei Cheng, Seoung Wug Oh, Brian Price, Alexander Schwing, and Joon-Young Lee. Tracking anything with decoupled video segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1316–1326, 2023. 2, 3", + "[10] Hainan Cui, Xiang Gao, Shuhan Shen, and Zhanyi Hu. Hsfm: Hybrid structure-from-motion. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1212-1221, 2017. 2" + ], + "bbox": [ + 93, + 234, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This project is supported by the Intel AI SRS gift, Amazon-Illinois AICE grant, Meta Research Grant, IBM IIDAI Grant, and NSF Awards #2331878, #2340254, #2312102, #2414227, and #2404385. We greatly appreciate the NCSA for providing computing resources.", + "[11] Andrew J Davison, Ian D Reid, Nicholas D Molton, and Olivier Stasse. Monoslam: Real-time single camera slam. IEEE transactions on pattern analysis and machine intelligence, 29(6):1052-1067, 2007. 2", + "[12] Matt Deitke, Christopher Clark, Sangho Lee, Rohun Tripathi, Yue Yang, Jae Sung Park, Mohammadreza Salehi, Niklas Muennighoff, Kyle Lo, Luca Soldaini, et al. Molmo and pixmo: Open weights and open data for state-of-the-art multimodal models. arXiv preprint arXiv:2409.17146, 2024. 2", + "[13] Jakob Engel, Thomas Schöps, and Daniel Cremers. Lsdslam: Large-scale direct monocular slam. In European conference on computer vision, pages 834-849. Springer, 2014. 2", + "[14] Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. In European Conference on Computer Vision, pages 241-258. Springer, 2025. 2", + "[15] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2", + "[16] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 8", + "[17] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa*, and Jitendra Malik*. Humans in 4D: Reconstructing and tracking humans with transformers. In International Conference on Computer Vision (ICCV), 2023. 2", + "[18] Zan Gojcic, Or Litany, Andreas Wieser, Leonidas J Guibas, and Tolga Birdal. Weakly supervised learning of rigid 3d scene flow. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5692-5703, 2021. 2", + "[19] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. arXiv preprint arXiv:2404.15506, 2024.7,8", + "[20] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024. 2, 7, 8", + "[21] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv preprint arXiv:2307.07635, 2023. 1, 2", + "[22] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker3: Simpler and better point tracking by pseudolabelling real videos. arXiv preprint arXiv:2410.11831, 2024. 2, 3, 7", + "[23] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurpos" + ], + "bbox": [ + 93, + 90, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1124", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 1, 2", + "[24] Numair Khan, Eric Penner, Douglas Lanman, and Lei Xiao. Temporally consistent online depth estimation using point-based fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9119-9129, 2023. 8", + "[25] Rawal Khirodkar, Timur Bagautdinov, Julieta Martinez, Su Zhaoen, Austin James, Peter Selednik, Stuart Anderson, and Shunsuke Saito. Sapiens: Foundation for human vision models. In European Conference on Computer Vision, pages 206-228. Springer, 2025. 1, 2", + "[26] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[27] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2, 3", + "[28] Chen Kong and Simon Lucey. Prior-less compressible structure from motion. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4123-4131, 2016. 2", + "[29] Chen Kong, Rui Zhu, Hamed Kiani, and Simon Lucey. Structure from category: A generic and prior-less approach. In 2016 Fourth International Conference on 3D Vision (3DV), pages 296-304. IEEE, 2016. 2", + "[30] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. 2, 7, 8", + "[31] Guillaume Le Moing, Jean Ponce, and Cordelia Schmid. Dense optical tracking: connecting the dots. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187-19197, 2024. 1, 2", + "[32] Jiahui Lei, Yijia Weng, Adam Harley, Leonidas Guibas, and Kostas Daniilidis. Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds. arXiv preprint arXiv:2405.17421, 2024. 2", + "[33] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Ep n p: An accurate o (n) solution to the p np problem. International journal of computer vision, 81:155-166, 2009. 7", + "[34] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 2", + "[35] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4273-4284, 2023. 2", + "[36] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with" + ], + "bbox": [ + 91, + 90, + 483, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 3", + "[37] Wei-Chiu Ma, Shenlong Wang, Rui Hu, Yuwen Xiong, and Raquel Urtasun. Deep rigid instance scene flow. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3614-3622, 2019. 2, 5", + "[38] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015. 2", + "[39] Richard A Newcombe, Dieter Fox, and Steven M Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 343-352, 2015. 5", + "[40] Emanuele Palazzolo, Jens Behley, Philipp Lottes, Philippe Giguere, and Cyril Stachniss. Refusion: 3d reconstruction in dynamic environments for rgb-d cameras exploiting residuals. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7855-7862. IEEE, 2019. 7, 8", + "[41] Linfei Pan, Daniel Barath, Marc Pollefeys, and Johannes L Schonberger. Global structure-from-motion revisited. In European Conference on Computer Vision (ECCV), 2024. 2", + "[42] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6", + "[43] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 4, 8", + "[44] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Undepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10106-10116, 2024. 1, 2, 7, 8", + "[45] Luigi Piccinelli, Christos Sakaridis, Yung-Hsu Yang, Mattia Segu, Siyuan Li, Wim Abbeloos, and Luc Van Gool. Undepthv2: Universal monocular metric depth estimation made simpler. arXiv preprint arXiv:2502.20110, 2025. 2, 3", + "[46] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2", + "[47] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2", + "[48] Noah Snavely, Steven M Seitz, and Richard Szeliski. Modeling the world from internet photo collections. International journal of computer vision, 80:189-210, 2008. 2" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1125", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Chonghyuk Song, Gengshan Yang, Kangle Deng, Jun-Yan Zhu, and Deva Ramanan. Total-recon: Deformable scene reconstruction for embodied view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17671-17682, 2023. 2", + "[50] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Symposium on Geometry processing, pages 109–116. CiteSeer, 2007. 5", + "[51] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of rgb-d slam systems. In 2012 IEEE/RSJ international conference on intelligent robots and systems, pages 573-580. IEEE, 2012. 7", + "[52] Yu Sun, Qian Bao, Wu Liu, Tao Mei, and Michael J Black. Trace: 5d temporal regression of avatars with dynamic cameras in 3d environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8856-8866, 2023. 2", + "[53] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 2", + "[54] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. Advances in Neural Information Processing Systems, 36, 2024. 7", + "[55] Bill Triggs, Philip F McLauchlan, Richard I Hartley, and Andrew W Fitzgibbon. Bundle adjustment—a modern synthesis. In Vision Algorithms: Theory and Practice: International Workshop on Vision Algorithms Corfu, Greece, September 21–22, 1999 Proceedings, pages 298–372. Springer, 2000. 2", + "[56] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on Pattern Analysis & Machine Intelligence, 13(04): 376-380, 1991. 7", + "[57] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 2, 5", + "[58] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 1, 2, 7", + "[59] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024. 2", + "[60] Yuefan Wu, Zeyuan Chen, Shaowei Liu, Zhongzheng Ren, and Shenlong Wang. Casa: Category-agnostic skeletal animal reconstruction. Advances in Neural Information Processing Systems, 35:28559-28574, 2022. 2", + "[61] Gengshan Yang, Minh Vo, Natalia Neverova, Deva Ramanan, Andrea Vedaldi, and Hanbyul Joo. Banmo: Building animatable 3d neural models from many casual videos. In" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2863-2873, 2022. 5", + "[62] Gengshan Yang, Shuo Yang, John Z Zhang, Zachary Manchester, and Deva Ramanan. Ppr: Physically plausible reconstruction from monocular videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3914-3924, 2023. 2", + "[63] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 1, 2, 8", + "[64] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[65] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825, 2024. 2, 3, 4, 6, 7, 8", + "[66] John Z Zhang, Shuo Yang, Gengshan Yang, Arun L Bishop, Swaminathan Gurumurthy, Deva Ramanan, and Zachary Manchester. Slomo: A general system for legged robot motion imitation from casual videos. IEEE Robotics and Automation Letters, 2023. 2", + "[67] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1724-1732, 2024. 2, 3", + "[68] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T Freeman. Structure and motion from casual videos. In European Conference on Computer Vision, pages 20-37. Springer, 2022. 2, 3, 4, 6, 7, 8", + "[69] Wang Zhao, Shaohui Liu, Hengkai Guo, Wenping Wang, and Yong-Jin Liu. *Particlesfm: Exploiting dense point trajectories for localizing moving cameras in the wild*. In European Conference on Computer Vision, pages 523-542. Springer, 2022. 2", + "[70] Yingying Zhu, Dong Huang, Fernando De La Torre, and Simon Lucey. Complex non-rigid motion 3d reconstruction by union of subspaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1542-1549, 2014. 2" + ], + "bbox": [ + 516, + 92, + 906, + 752 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1126", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Uni4D_ Unifying Visual Foundation Models for 4D Modeling from a Single Video/b0bfb21a-5579-4e48-8be6-e0a25d4d029f_model.json b/2025/Uni4D_ Unifying Visual Foundation Models for 4D Modeling from a Single Video/b0bfb21a-5579-4e48-8be6-e0a25d4d029f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..81d12b883c9daeb8649eeb0797fee0a13d504f54 --- /dev/null +++ b/2025/Uni4D_ Unifying Visual Foundation Models for 4D Modeling from a Single Video/b0bfb21a-5579-4e48-8be6-e0a25d4d029f_model.json @@ -0,0 +1,2851 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.131, + 0.908, + 0.152 + ], + "angle": 0, + "content": "Uni4D: Unifying Visual Foundation Models for 4D Modeling from a Single Video" + }, + { + "type": "text", + "bbox": [ + 0.266, + 0.181, + 0.731, + 0.218 + ], + "angle": 0, + "content": "David Yifan Yao Albert J. Zhai Shenlong Wang University of Illinois at Urbana-Champaign" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.23, + 0.495, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.228, + 0.891, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.45, + 0.908, + 0.493 + ], + "angle": 0, + "content": "Figure 1. Given a casually captured video, Uni4D harnesses pretrained visual foundation models and multi-stage optimization to jointly estimate camera poses, dynamic geometry, and dense 3D motion. The resulting camera poses and geometry are accurate, consistent, and coherent both temporally and spatially. This is all done without any additional training or fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.505, + 0.327, + 0.521 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.538, + 0.486, + 0.78 + ], + "angle": 0, + "content": "This paper presents a unified approach to understanding dynamic scenes from casual videos. Large pretrained vision foundation models, such as vision-language, video depth prediction, motion tracking, and segmentation models, offer promising capabilities. However, training a single model for comprehensive 4D understanding remains challenging. We introduce Uni4D, a multi-stage optimization framework that harnesses multiple pretrained models to advance dynamic 3D modeling, including static/dynamic reconstruction, camera pose estimation, and dense 3D motion tracking. Our results show state-of-the-art performance in dynamic 4D modeling with superior visual quality. Notably, Uni4D requires no retraining or fine-tuning, highlighting the effectiveness of repurposing visual foundation models for 4D understanding. Code and more results are available at: https://davidya099.github.io/uni4d." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.814, + 0.222, + 0.829 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Over the past two years, many visual foundation models have emerged [4, 21, 23, 25, 27, 31, 44, 58, 63], achieving high accuracy on tasks like depth prediction, segmentation, human parsing, normal estimation, few-view recon" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.907, + 0.734 + ], + "angle": 0, + "content": "struction, and motion tracking. These models leverage supervised learning on large, diverse datasets, achieving impressive accuracy and remarkable generalization. However, these advances have not translated to 4D (time + geometry) modeling, a longstanding challenge in computer vision. We see two main reasons: first, collecting high-quality, ground-truth 4D data from real-world environments remains complex and resource-intensive. Second, 4D understanding is a holistic problem involving interconnected tasks like camera pose estimation, 3D reconstruction, and dynamics tracking. Although each subtask shows progress, data-driven cues remain noisy, and more importantly, unifying them synergistically for holistic modeling remains challenging. This paper seeks to answer: Can we harness the success of visual foundation models for dynamic 4D modeling?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we propose Uni4D, a novel framework that reconstructs 4D scenes from a single video captured in the wild. Our method integrates data-driven foundation models and conventional model-driven dynamic structure-from-motion, combining data-driven cues and model-based knowledge synergistically. Our intuition is simple: each data-driven visual cue, such as video segmentation, pixel-level motion tracking, and video depth, is a partial projection from the 4D world to the 2D video. The key is to create a 4D scene representation that coherently aligns with each cue while incorporating strong prior knowledge" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1116" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.363 + ], + "angle": 0, + "content": "of real-world motion and shape to resolve temporary inconsistencies. To achieve this, we take an energy-minimization perspective, framing the problem as an optimization task that jointly infers camera poses, static and dynamic geometry, and motion. The dynamic 4D modeling framework is complex, as it involves various optimization variables, visual cues, and constraints. To overcome the challenge of joint reasoning, we carefully design a novel three-stage, divide-and-conquer pipeline that progressively incorporates camera poses, static geometry, and dynamic geometry and motion into the optimization framework. Uni4D leverages pretrained foundation models across different tasks, requiring no task-specific retraining or fine-tuning. This design sidesteps the need for 4D ground-truth data, a major challenge in the field. Through incorporating strong priors on geometry and dynamic motion, our method produces realistic 4D scenes that are coherent across space and time while maintaining high accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.482, + 0.454 + ], + "angle": 0, + "content": "We demonstrate the effectiveness of Uni4D on various datasets. As shown in Fig. 1, our method effectively recovers the clean geometry and motion of the scene, as well as camera trajectories, from a single video. As shown in Fig. 2, our model outperforms all dynamic 4D modeling baselines in both camera pose and geometry." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.467, + 0.242, + 0.484 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.483, + 0.672 + ], + "angle": 0, + "content": "Structure from Motion and SLAM. Structure from Motion (SfM) is a classical problem in computer vision that aims to jointly recover camera parameters and 3D structure from images [2, 10, 41, 47, 48]. Simultaneous localization and mapping (SLAM) is a similar problem that focuses on real-time efficiency in an online setting [11, 13, 38, 53]. The core idea for many approaches to SfM and SLAM is to optimize a form of reprojection error with respect to both camera parameters and 3D points, also known as bundle adjustment [55]. This relies on the key assumption that the scene being viewed is rigid (\"static\"), making these approaches not applicable to dynamic scenes." + }, + { + "type": "image", + "bbox": [ + 0.149, + 0.673, + 0.416, + 0.87 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.872, + 0.483, + 0.9 + ], + "angle": 0, + "content": "Figure 2. Uni4D outperforms other recent 4D modeling methods in both camera pose and geometry accuracy on the Sintel dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.456 + ], + "angle": 0, + "content": "4D Understanding. The problem of non-rigid structure from motion is highly ill-posed. Previous works on non-rigid SfM leveraged various forms of priors or weaker assumptions on shape and or motion to make the problem solvable [3, 6, 28, 29, 70]. With the rise of modern segmentation models, many works have studied category-specific articulation and rigidity priors, focusing on object categories such as humans [17, 49, 52, 64], animals [60-62, 66], and vehicles [18, 37]. The generality of such methods is limited by their category-specific nature. Recently, several works have also begun to leverage trained neural networks as priors for 4D understanding in a more freeform manner [30, 65, 68, 69]. However, these methods involve either training a network [65, 69] or optimizing specific layers of an existing network [30, 68], making it difficult to integrate newer models into the pipeline. Our method takes the strategy further and integrates pretrained models in a completely modular manner, fully unleashing their generalization capabilities and allowing for seamless integration of newer models. Note that various recent works have also studied 4D reconstruction in an optimization-based framework [15, 32, 35, 57, 59]. However, these methods have additional rendering capacities and focus on rendering metrics, while we focus on recovering high-quality geometry." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.909, + 0.759 + ], + "angle": 0, + "content": "Visual Foundation Models. In recent years, a number of visual foundation models have been developed, achieving remarkable performance on tasks like depth estimation [5, 20, 23, 44, 63], detection and segmentation [12, 27, 36, 46], human parsing [25], surface normal estimation [4, 14], few-view reconstruction [34, 58], and point tracking [21, 22, 31]. Our insight is that nearly all of these models have the potential to contribute towards holistic 4D understanding, and harnessing them in a unifying framework can advance the state-of-the-art in tasks such as non-rigid structure from motion. To this end, we integrate the following pretrained models in Uni4D: UniDepthv2 [45] for geometry initialization, CoTracker3 [22] for correspondence initialization, and a collection of Recognize Anything Model [67], ChatGPT [1], Grounding-SAM [27, 36], and DEVA [9] for dynamic object segmentation. Through multi-stage optimization with a few regularizing priors, we are able to use these models to extract accurate pose and 4D geometry from monocular video." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.77, + 0.607, + 0.785 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we are interested in recovering 4D geometry and camera parameters from a monocular casual video. Our model is built on the intuition that the 2D visual cues from the video can be seen as perspective projections of 4D geometry and motion, where the video depth represents a projection of 4D geometry, 2D dense tracking corresponds to a projection of 4D motion, and segmentation reflects a" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1117" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.131, + 0.082, + 0.868, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.318, + 0.908, + 0.375 + ], + "angle": 0, + "content": "Figure 3. Given a casually captured video, Uni4D exploits visual foundation models to extract dynamic segmentation, video depth, and motion tracks. Static geometry and poses are obtained through tracklet-based structure-from-motion along with camera motion priors. Dynamic geometry is improved through nonrigid bundle adjustment and scene motion priors. A final fusion densifies geometry to obtain high quality 4D reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.383, + 0.484, + 0.458 + ], + "angle": 0, + "content": "projection of the 4D dynamic object silhouette. We propose a novel, training-free, foundation-model-based energy minimization scheme that leverages visual cues to jointly infer camera poses, geometry, and motion, enabling generalization across diverse videos." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.47, + 0.307, + 0.485 + ], + "angle": 0, + "content": "3.1. Pretrained Visual Cues" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.484, + 0.555 + ], + "angle": 0, + "content": "We first exploit foundation models to extract visual cues, including dynamic segmentation, video depth, and motion tracking. All of the models are pre-trained on large and diverse datasets and exhibit strong generalization abilities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.576, + 0.484, + 0.76 + ], + "angle": 0, + "content": "Video Segmentation. Recognizing and segmenting dynamic objects is crucial for 4D scene understanding. We leverage the latest advancements in video semantic segmentation and tracking to estimate dynamic objects over time. First, we use RAM [67] to identify semantic classes in the video. These classes are then filtered through GPT-4o [1] to exclude static and background elements (e.g., buildings, poles), retaining only dynamic objects (e.g., humans, animals, vehicles). Next, Grounding-SAM [27, 36] performs segmentation at each keyframe, and DEVA [9] tracks these segments over time, resulting in accurate dynamic video segmentation \\(\\{\\mathbf{M}_{\\mathbf{t}}\\}_{t = 0}^{T}\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Dense Motion Tracking. We use dense pixel tracking to establish correspondences over time, which serve as motion cues to assist inference of geometry reconstruction and dynamic object 3D motion. Unlike traditional 4D reconstruction methods relying on optical flow [65, 68], dense motion tracklets yield more correspondence pairs across large viewpoints and structural changes. Pixel tracking also outperforms sparse matching in density and surpasses flow" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.907, + 0.535 + ], + "angle": 0, + "content": "propagation in robustness, making it ideal for dynamic 4D reconstruction. We utilize Co-TrackerV3 [22] for its robustness, as it employs a 4D cost volume with an attention mechanism to track 2D points through occlusions, recently proving to be effective for dynamic neural rendering. We apply Co-Tracker bi-directionally on a dense grid every 10 frames to ensure thorough coverage. We filter and classify tracklets using segmentation masks yielding a set of correspondent point trajectories \\(\\{\\mathbf{Z}_k\\in \\mathbb{R}^{T\\times 2}\\}_{k = 0}^K\\) at visible time steps determined by Co-Tracker." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.553, + 0.907, + 0.645 + ], + "angle": 0, + "content": "Video Depth Estimation. Monocular depth reasoning, while insufficient as a standalone tool for complete 4D geometry recovery, provides strong geometry initialization cues. In our model, we use UniDepthV2 [45], a monocular depth estimation network, to estimate an initial depth map, \\(\\{\\mathbf{D}_t\\}_{t=0}^T\\), and initial camera intrinsics, \\(\\mathbf{K}_{\\mathrm{init}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.653, + 0.707, + 0.669 + ], + "angle": 0, + "content": "3.2. Energy Formulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.902 + ], + "angle": 0, + "content": "We now describe the energy formulation of our proposed dynamic 4D reconstruction framework. Let \\(\\mathcal{M} = \\{\\mathbf{M}_t\\}_{t=0}^T\\), \\(\\mathcal{D} = \\{\\mathbf{D}_t\\}_{t=0}^T\\), \\(\\mathcal{Z} = \\{\\mathbf{Z}_k\\}_{k=0}^K\\) be the input dynamic segmentation, monocular depth, dense motion trajectory extracted from the input video \\(\\mathcal{I} = \\{\\mathbf{I}_t\\}_{t=0}^T\\). Formally, our goal is to obtain camera parameters \\(\\mathcal{C}\\), namely poses \\(\\mathcal{T}\\) and intrinsics \\(\\mathbf{K}\\), and a set of 4D point clouds \\(\\mathcal{P}\\) containing both dynamic and static parts separately as \\(\\mathcal{P} = \\{\\mathbf{P}_{\\mathrm{static}}, \\mathcal{P}_{\\mathrm{dyn}}\\}\\). Here, \\(\\mathbf{P}_{\\mathrm{static}}\\) does not change over time and \\(\\mathcal{P}_{\\mathrm{dyn}} = \\{\\mathbf{p}_k \\in \\mathbb{R}^{T \\times 3}\\}_k\\) is a temporally varying point cloud, where each \\(\\mathbf{p}_k\\) is a dynamic point trajectory. We represent the camera poses as rigid transforms \\(\\mathcal{T} = \\{\\xi_t \\in \\mathbb{S}\\mathbb{E}(3)\\}_{t=0}^T\\), and parameterize the rotations as \\(\\mathfrak{so}(3)\\) rotation vectors as a minimal representation for easy optimization. We assume all frames share the same intrinsic" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1118" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.171, + 0.116, + 0.335, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.172, + 0.188, + 0.334, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.172, + 0.262, + 0.335, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.172, + 0.337, + 0.335, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.172, + 0.412, + 0.335, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.502, + 0.287, + 0.512 + ], + "angle": 0, + "content": "Input Video" + }, + { + "type": "image", + "bbox": [ + 0.357, + 0.099, + 0.501, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.387, + 0.233, + 0.476, + 0.242 + ], + "angle": 0, + "content": "CasualSAM Static" + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.099, + 0.673, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.57, + 0.232, + 0.643, + 0.242 + ], + "angle": 0, + "content": "CasualSAM 4D" + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.099, + 0.787, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.71, + 0.232, + 0.819, + 0.242 + ], + "angle": 0, + "content": "CasualSAM Bird's Eye" + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.256, + 0.512, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.396, + 0.372, + 0.475, + 0.381 + ], + "angle": 0, + "content": "Monst3R Static" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.253, + 0.685, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.371, + 0.631, + 0.38 + ], + "angle": 0, + "content": "Monst3R 4D" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.248, + 0.81, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.716, + 0.371, + 0.815, + 0.382 + ], + "angle": 0, + "content": "Monst3R Bird's Eye" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.385, + 0.521, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.502, + 0.472, + 0.512 + ], + "angle": 0, + "content": "Ours Static" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.383, + 0.691, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.594, + 0.502, + 0.637, + 0.512 + ], + "angle": 0, + "content": "Ours 4D" + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.385, + 0.804, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.738, + 0.502, + 0.815, + 0.513 + ], + "angle": 0, + "content": "Ours Bird's Eye" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.519, + 0.907, + 0.575 + ], + "angle": 0, + "content": "Figure 4. Qualitative results of 4D reconstruction on the DAVIS [43] dataset. CasualSAM [68] suffers from slanted geometry, and Monst3R [65] has unclear geometry and does not resolve conflicts from multiple views (note the wall in the bird's-eye view). Both CasualSAM and Monst3R lack clean dynamic reconstruction and segmentation. Uni4D achieves a realistic layout, thanks to joint optimization, and provides accurate dynamic segmentation and reconstruction by leveraging foundation visual models as cues." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.591, + 0.483, + 0.637 + ], + "angle": 0, + "content": "matrix \\(\\mathbf{K}\\) where we optimize focal lengths \\(f_{x}\\) and \\(f_{y}\\). We formulate the 4D joint reasoning problem as minimization of the following energy function:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.649, + 0.483, + 0.679 + ], + "angle": 0, + "content": "\\[\nE _ {\\mathrm {B A}} (\\mathcal {C}, \\mathbf {P} _ {\\text {s t a t i c}}) + E _ {\\mathrm {N R}} (\\mathcal {P} _ {\\mathrm {d y n}}) + E _ {\\text {m o t i o n}} (\\mathcal {P} _ {\\mathrm {d y n}}) + E _ {\\text {c a m}} (\\mathcal {T}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.681, + 0.483, + 0.846 + ], + "angle": 0, + "content": "where \\( E_{\\mathrm{BA}}(\\mathcal{C}, \\mathbf{P}_{\\mathrm{static}}) \\) is a bundle adjustment term that measures the discrepancy between static-region correspondences and the static 3D structure \\( \\mathbf{P}_{\\mathrm{static}} \\) through perspective reprojection. \\( E_{\\mathrm{NR}}(\\mathcal{P}_{\\mathrm{dyn}}) \\) is a non-rigid structure-from-motion energy term that measures the disagreement between the dynamic point cloud and their tracklet correspondences, \\( E_{\\mathrm{cam}}(\\mathcal{T}) \\) is a regularization term on the camera motion smoothness, and \\( E_{\\mathrm{motion}}(\\mathcal{P}_{\\mathrm{dyn}}) \\) is a regularization term on the dynamic structure and motion. Each energy term is involved in different stages of the optimization process, which will be described in Sec. 3.3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Static Bundle Adjustment Term. The bundle adjustment energy \\( E_{\\mathrm{BA}}(\\mathcal{C}, \\mathbf{P}_{\\mathrm{static}}) \\) measures the consistency between" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.591, + 0.907, + 0.667 + ], + "angle": 0, + "content": "the pixel-level correspondences and the 3D structure of static scene elements. Given the input pixel tracks \\(\\mathcal{Z} = \\{\\mathbf{Z}_k\\}\\) and video segmentation \\(\\mathcal{M}\\), we filter all tracks corresponding to static areas and minimize the distance between the projected pixel location and the observed pixel location:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.7, + 0.928, + 0.745 + ], + "angle": 0, + "content": "\\[\nE _ {\\mathrm {B A}} (\\mathcal {C}, \\mathbf {P} _ {\\text {s t a t i c}}; \\mathcal {Z}, \\mathcal {M}) = \\sum_ {\\mathbf {z} _ {k} \\in \\mathcal {M}} \\sum_ {t} \\mathbf {w} _ {\\mathbf {k}, \\mathbf {t}} \\| \\mathbf {z} _ {k, t} - \\pi_ {\\mathbf {K}} \\left(\\mathbf {p} _ {k}, \\boldsymbol {\\xi} _ {t}\\right) \\| _ {2} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.746, + 0.907, + 0.791 + ], + "angle": 0, + "content": "where \\(\\mathbf{z}_{k,t}\\) is the \\(k\\)-th 3D point's corresponding pixel track's 2D coordinates at time \\(t\\), \\(\\mathbf{w_{k,t}} \\in \\{0,1\\}\\) is a visibility indicator and \\(\\pi_{\\mathbf{K}}\\) is the perspective projection function." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Non-Rigid Bundle Adjustment Term. For dynamic objects, we impose a nonrigid bundle adjustment term, \\( E_{\\mathrm{NR}}(\\mathcal{P}_{\\mathrm{dyn}}) \\), which measures the discrepancy between the dynamic point cloud and pixel tracklets. Here, each pixel tracklet corresponds to a dynamic 3D point sequence," + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1119" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.094, + 0.473, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.242, + 0.483, + 0.286 + ], + "angle": 0, + "content": "Figure 5. Qualitative results on Bonn dataset. Both CasualSAM and MonST3R have trailing artifacts and incorrect dynamic estimations. Uni4D provides clear dynamic and static geometry." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.307, + 0.395, + 0.322 + ], + "angle": 0, + "content": "\\(\\{\\mathbf{p}_{k,t}\\}_{t}\\), optimized for each observed tracklet:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.33, + 0.502, + 0.375 + ], + "angle": 0, + "content": "\\[\nE _ {\\mathrm {N R}} \\left(\\mathcal {P} _ {\\mathrm {d y n}}; \\mathcal {C}, \\mathcal {Z}, \\mathcal {M}\\right) = \\sum_ {\\mathbf {z} _ {k} \\in \\mathcal {M}} \\sum_ {t} \\mathbf {w} _ {\\mathbf {k}, \\mathbf {t}} \\left\\| \\mathbf {z} _ {k, t} - \\pi_ {\\mathbf {K}} \\left(\\mathbf {p} _ {k, t}, \\boldsymbol {\\xi} _ {t}\\right) \\right\\| _ {2} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.375, + 0.472, + 0.391 + ], + "angle": 0, + "content": "where \\(\\mathbf{p}_{k,t} \\in \\mathbb{R}^3\\) is the \\(k\\)-th dynamic point's location at \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.408, + 0.484, + 0.529 + ], + "angle": 0, + "content": "Camera Motion Prior. Considering that our input is a video, we incorporate a temporal smoothness prior on camera poses that penalizes sudden changes in relative pose: \\(\\pmb{\\xi}_{t\\rightarrow t + 1} = \\pmb{\\xi}_{t + 1}^{-1}\\cdot \\pmb{\\xi}_t\\). We reweight this term based on the magnitude of the relative motion: intuitively, if the relative motion is large, we penalize change rates in relative motion less; if the relative motion is small, we apply a higher penalty on change rate. Formally, we have:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.536, + 0.482, + 0.579 + ], + "angle": 0, + "content": "\\[\nE _ {\\mathrm {c a m}} (\\mathcal {T}) = \\sum_ {t} E _ {\\mathrm {r o t}} \\left(\\mathbf {R} _ {t - 1, t, t + 1}\\right) + \\sum_ {t} E _ {\\mathrm {t r a n s}} \\left(\\mathbf {t} _ {t - 1, t, t + 1}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.577, + 0.482, + 0.631 + ], + "angle": 0, + "content": "where \\(E_{\\mathrm{rot}}(\\mathbf{R}_{t - 1,t,t + 1}) = \\frac{2\\|\\mathrm{rad}(\\mathbf{R}_{t\\to t + 1}) - \\mathrm{rad}(\\mathbf{R}_{t - 1\\to t})\\|}{\\|\\mathrm{rad}(\\mathbf{R}_{t - 1\\to t})\\| + \\|\\mathrm{rad}(\\mathbf{R}_{t\\to t + 1})\\|}\\) and \\(E_{\\mathrm{trans}}(\\mathbf{t}_{t - 1,t,t + 1}) = \\frac{2\\|\\mathbf{t}_{t\\to t + 1} - \\mathbf{t}_{t - 1\\to t}\\|}{\\|\\mathbf{t}_{t - 1\\to t}\\| + \\|\\mathbf{t}_{t\\to t + 1}\\|}\\); rad converts the rotation matrix into absolute radians." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.65, + 0.484, + 0.725 + ], + "angle": 0, + "content": "Dynamic Motion Prior. \\( E_{\\mathrm{motion}}(\\mathcal{P}_{\\mathrm{dyn}}) \\) is a regularization term that encodes the characteristics of the dynamic structure. It contains two prior terms that are used to regularize the dynamic structure, both of which have demonstrated effectiveness in previous work [39, 50]:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.733, + 0.482, + 0.75 + ], + "angle": 0, + "content": "\\[\nE _ {\\text {m o t i o n}} \\left(\\mathcal {P} _ {\\text {d y n}}\\right) = E _ {\\text {a r a p}} \\left(\\mathcal {P} _ {\\text {d y n}}\\right) + E _ {\\text {s m o o t h}} \\left(\\mathcal {P} _ {\\text {d y n}}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.756, + 0.484, + 0.847 + ], + "angle": 0, + "content": "\\(E_{\\mathrm{arap}}\\) is an as-rigid-as-possible (ARAP) [50] prior that penalizes extreme deformations that compromise local rigidity. Specifically, we obtain the nearest neighbors of each dynamic control point \\(k\\) by applying KNN over the other tracks and enforce that the relative distances between these close-by pairs do not undergo sudden changes:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.854, + 0.508, + 0.901 + ], + "angle": 0, + "content": "\\[\nE _ {\\text {a r a p}} = \\sum_ {\\mathbf {t}} \\sum_ {(k, m)} \\mathbf {w} _ {\\mathbf {k m}, \\mathbf {t}} \\| d \\left(\\mathbf {p} _ {\\mathbf {k}, \\mathbf {t}}, \\mathbf {p} _ {m, t}\\right) - d \\left(\\mathbf {p} _ {\\mathbf {k}, \\mathbf {t} + 1}, \\mathbf {p} _ {m, t + 1}\\right) \\| _ {2} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "where \\(d(\\cdot, \\cdot)\\) is the L2 distance and \\(\\mathbf{w}_{\\mathbf{k}\\mathbf{m},\\mathbf{t}} = 1\\) if all relevant points are visible." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.905, + 0.152 + ], + "angle": 0, + "content": "\\(E_{\\mathrm{smooth}}\\) is a simple smoothness term that promotes temporal smoothness for the dynamic point cloud:" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.158, + 0.905, + 0.191 + ], + "angle": 0, + "content": "\\[\nE _ {\\text {s m o o t h}} = \\sum_ {\\mathbf {t}} \\sum_ {\\mathbf {p} _ {k} \\in \\mathcal {P} _ {\\text {d y n}}} \\mathbf {w} _ {\\mathbf {k}, \\mathbf {t}} \\| \\mathbf {p} _ {k, t} - \\mathbf {p} _ {k, t + 1} \\| _ {2}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.197, + 0.907, + 0.288 + ], + "angle": 0, + "content": "Despite simplicity, both motion terms are crucial in our formulation, as they significantly reduce ambiguities in 4D dynamic structure estimation, which is highly ill-posed. Unlike other methods, we do not assume strong model-based motion priors, such as rigid motion [37], articulated motion [61], or a linear motion basis [57]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.295, + 0.623, + 0.308 + ], + "angle": 0, + "content": "3.3. Inference" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.316, + 0.906, + 0.407 + ], + "angle": 0, + "content": "Directly minimizing the energy defined in Eq. 1 is nontrivial, as our energy function is highly non-linear and involves millions of free variables. To address this, we developed a three-stage optimization pipeline, enabling us to minimize the energy and estimate the scene variables in a divide-and-conquer fashion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.906, + 0.546 + ], + "angle": 0, + "content": "Stage 1: Camera Initialization. We start by initializing camera parameters. Combining video depth estimation \\(\\mathcal{D}\\) and dense pixel motion \\(\\mathcal{Z}\\) allows us to establish 2D-to-3D correspondences. This allows us to initialize and tune \\(\\mathcal{C}\\) by minimizing the following energy function with respect to camera parameters only. Specifically, we can unproject each video frame's depth at time \\(t\\) back to 3D and minimize the following energy function:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.551, + 0.905, + 0.585 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathbf {C}} \\sum_ {\\left(t ^ {\\prime}, t\\right)} \\sum_ {\\mathbf {z} _ {k} \\in \\neg \\mathcal {M}} \\| \\mathbf {z} _ {k, t ^ {\\prime}} - \\pi_ {\\mathbf {K}} \\left(\\pi_ {\\mathbf {K}} ^ {- 1} \\left(\\mathbf {z} _ {k, t}, \\mathbf {D} _ {t}, \\boldsymbol {\\xi} _ {t}\\right), \\boldsymbol {\\xi} _ {t ^ {\\prime}}\\right) \\| _ {2} ^ {2} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.592, + 0.906, + 0.714 + ], + "angle": 0, + "content": "where \\(\\pi_{\\mathbf{K}}^{-1}\\) is the unprojection function that maps 2D coordinates into 3D world coordinates using estimated depth \\(\\mathbf{D}_t\\). We perform this over all pairs within a temporal sliding window of 5 frames, producing a good initial pose estimate as shown in Tab. 4. Given camera initialization \\(\\hat{\\mathcal{C}}\\), we directly unproject our depth prediction into a common world coordinate system, which provides an initial 4D structure \\(\\hat{\\mathcal{P}}\\). This is used as initialization for later optimization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.731, + 0.906, + 0.791 + ], + "angle": 0, + "content": "Stage 2: Bundle Adjustment. Our second stage jointly optimizes camera pose and static geometry by minimizing the static component-related energy in a bundle adjustment fashion. Formally speaking, we solve the following:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.798, + 0.905, + 0.821 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathcal {C}, \\mathbf {P} _ {\\text {s t a t i c}}} E _ {\\mathrm {B A}} \\left(\\mathcal {C}, \\mathbf {P} _ {\\text {s t a t i c}}; \\mathcal {Z}, \\mathcal {M}\\right) + E _ {\\mathrm {c a m}} (\\mathcal {T}) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "By enforcing consistency with each other, this improves both the static geometry and the camera pose quality. We perform a final scene integration by unprojecting correspondences into 3D using improved pose and filtering outlier noisy points in 3D." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1120" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.088, + 0.237, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.14, + 0.236, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.19, + 0.236, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.133, + 0.239, + 0.236, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.132, + 0.289, + 0.235, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.156, + 0.342, + 0.21, + 0.351 + ], + "angle": 0, + "content": "Input Video" + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.089, + 0.432, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.314, + 0.163, + 0.391, + 0.171 + ], + "angle": 0, + "content": "CasualSAM Static" + }, + { + "type": "image", + "bbox": [ + 0.441, + 0.089, + 0.647, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.163, + 0.604, + 0.17 + ], + "angle": 0, + "content": "CasualSAM 4D" + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.08, + 0.837, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.163, + 0.84, + 0.171 + ], + "angle": 0, + "content": "CasualSAM Bird's Eye" + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.184, + 0.437, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.317, + 0.244, + 0.388, + 0.254 + ], + "angle": 0, + "content": "MonST3R Static" + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.179, + 0.659, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.462, + 0.269, + 0.663, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.341, + 0.595, + 0.349 + ], + "angle": 0, + "content": "Ours 4D" + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.174, + 0.846, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.748, + 0.244, + 0.835, + 0.253 + ], + "angle": 0, + "content": "MonST3R Bird's Eye" + }, + { + "type": "image", + "bbox": [ + 0.737, + 0.261, + 0.842, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.766, + 0.34, + 0.832, + 0.349 + ], + "angle": 0, + "content": "Ours Bird's Eye" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.363, + 0.907, + 0.419 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of 4D reconstruction on DAVIS dataset. CasualSAM [68] distorts the room geometry as evident from the bird's eye view. Dynamic elements have inconsistent geometry over time. MonST3R [65] has noisy static geometry in the room's far corner and incomplete dynamic object geometry evident in the green highlight. Uni4D has the cleanest dynamic object reconstruction and segmentation results, with geometrically accurate room shapes as evident from the bird's eye view." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.484, + 0.549 + ], + "angle": 0, + "content": "Stage 3: Non-Rigid Bundle Adjustment. Given the estimated camera pose, our third stage focuses on inferring dynamic structure. Note that we freeze camera parameters in this stage, as we find that incorrect geometry and motion evidence often harm camera pose estimation rather than improve it. Additionally, enabling camera pose optimization introduces extra flexibility in this ill-posed problem, harming robustness. Formally speaking, we solve the following:" + }, + { + "type": "equation", + "bbox": [ + 0.123, + 0.56, + 0.482, + 0.585 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathcal {P} _ {\\mathrm {d y n}}} E _ {\\mathrm {N R}} \\left(\\mathcal {P} _ {\\mathrm {d y n}}; \\mathbf {C}, \\mathcal {Z}, \\mathcal {M}\\right) + E _ {\\text {m o t i o n}} \\left(\\mathcal {P} _ {\\mathrm {d y n}}\\right) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.484, + 0.701 + ], + "angle": 0, + "content": "We initialize \\(\\mathcal{P}_{\\mathrm{dyn}}\\) using video depth and our optimized camera pose from Stage 2. We scale \\(E_{\\mathrm{smooth}}\\) and \\(E_{\\mathrm{arap}}\\) with constants 10 and 100 respectively which we found empirically led to better dynamics. This energy optimization might still leave some high-energy noisy points, often from incorrect cues, motion boundaries, or occlusions. We filter these outliers based on their energy values in a final step." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.81 + ], + "angle": 0, + "content": "Fusion. From our energy minimization, we acquire a semi-dense dynamic and static point cloud along with camera parameters. To further densify the point cloud, enabling each pixel to correspond to a 3D point, we perform depth-based interpolation by computing a scale offset. Details can be found in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.886 + ], + "angle": 0, + "content": "To unify our outputs into a consistent 4D representation, we use our camera parameters and static masks to project the aligned depth maps into world coordinates. To handle noisy depth values at boundaries, we create an edge mask filter by thresholding gradients of the depth maps." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "This process enables us to update the depth value for" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.428, + 0.907, + 0.62 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.626, + 0.907, + 0.669 + ], + "angle": 0, + "content": "Figure 7. Qualitative results of 4D reconstruction on Sintel dataset. We highlight and contrast performance on (1) dynamic objects and (2) geometry on planar surfaces." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.678, + 0.907, + 0.724 + ], + "angle": 0, + "content": "each high-confidence pixel, which we can then reproject into 3D space to obtain the final dense dynamic geometry reconstruction result, as shown in Fig. 1 and Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.737, + 0.646, + 0.755 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.763, + 0.906, + 0.809 + ], + "angle": 0, + "content": "Uni4D estimates camera pose, depth, and 3D motion from a single video. We perform experiments evaluating its performance with respect to baselines on these tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.731, + 0.835 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.901 + ], + "angle": 0, + "content": "For optimization, we use the Adam optimizer with ReduceLROnPlateau and EarlyStopping in PyTorch [26, 42]. We perform 600 iterations per sliding window in stage 1, 2000 iterations in stage 2, and 1000 iterations in stage 3." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1121" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.156, + 0.089, + 0.837, + 0.257 + ], + "angle": 0, + "content": "
CategoryMethodSintelTUM-dynamicsBonn
ATE ↓RPE trans ↓RPE rot ↓ATE ↓RPE trans ↓RPE rot ↓ATE ↓RPE trans ↓RPE rot ↓
Pose onlyDPVO*[54]0.1710.0631.2910.0190.0140.4060.0220.0140.913
LEAP-VO*[8]0.0350.0651.6690.0250.0312.8430.0370.0140.844
Joint depth & poseRobust-CVD[30]0.3680.1533.4620.0960.0272.5900.0850.0180.803
CasualSAM[68]0.1370.0390.6300.0360.0180.7450.0240.0140.849
Monst3R[65]0.1080.0430.7290.1080.0221.3710.0230.0110.807
Uni4D0.1100.0320.3380.0120.0040.3350.0170.0100.818
Uni4D*0.0920.0330.1410.0120.0040.3310.0160.0100.817
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.26, + 0.907, + 0.289 + ], + "angle": 0, + "content": "Table 1. Camera Pose Evaluation on Sintel, TUM-dynamic, and Bonn datasets. We bold and underline the best and second best results respectively. * indicates known camera intrinsic." + }, + { + "type": "table", + "bbox": [ + 0.174, + 0.302, + 0.821, + 0.515 + ], + "angle": 0, + "content": "
AlignmentCategoryMethodSintelBonnKITTI
Abs Rel ↓ δ<1.25 ↑Abs Rel ↓ δ<1.25 ↑Abs Rel ↓ δ<1.25 ↑
Per-sequence scale & shiftSingle-frame depthMetric3D[19]0.20571.90.04498.50.03998.8
Depth-pro[5]0.28060.50.04998.60.08094.2
Unidepth[44]0.19872.80.04098.50.03898.8
Video depthDepthCrafter[20]0.23169.00.06597.60.11288.4
Joint video depth & poseRobust-CVD[30]0.35849.70.10889.80.18272.9
CasualSAM[68]0.29256.90.06996.60.11388.3
Monst3r[65]0.35852.10.06095.00.08591.9
Uni4D0.21672.50.03898.30.09889.7
Per-sequence scaleJoint depth & poseMonst3r[65]0.34455.90.04198.20.08991.4
Joint depth & poseUni4D0.28964.90.03898.30.08693.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.517, + 0.907, + 0.532 + ], + "angle": 0, + "content": "Table 2. Video depth evaluation on Sintel, Bonn, and KITTI datasets. We bold and underline the best and second best results respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.486, + 0.691 + ], + "angle": 0, + "content": "We initialize the learning rate at \\(1 \\times 10^{-3}\\) for stage 1, and \\(1 \\times 10^{-2}\\) for stages 2 and 3, reducing all to \\(1 \\times 10^{-4}\\). Our entire framework takes roughly 5 minutes for a 50-frame video on a RTX A6000 GPU. We include a detailed runtime breakdown in the supplementary. We run all baselines on the datasets using their official implementations and hyperparameters. Co-Trackers [22] are initialized at a dense \\(50 \\times 50\\) grid, with a \\(75 \\times 75\\) grid for Sintel to handle its large camera perspective change. All optimization hyperparameters are kept the same for all runs on all datasets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.733, + 0.251, + 0.747 + ], + "angle": 0, + "content": "4.2. Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Baselines. We compare with several recent methods for pose estimation in dynamic scenes. LEAPVO [8] and DPVO [54] are learning-based visual odometry methods. Robust-CVD [30] optimizes for pose and depth deformation through an SFM pipeline. CasualSAM [68] further improves on the idea by directly finetuning network weights along with a novel uncertainty formulation. Monst3R [65] is a very recent work that fine-tunes DUSt3R [58] for 4D reconstruction through PnP [33]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.54, + 0.907, + 0.691 + ], + "angle": 0, + "content": "Benchmarks and metrics. We evaluate pose estimation on three dynamic datasets: Sintel [7], TUM-dynamics [51], and Bonn [40]. We follow LEAP-VO's evaluation split for Sintel and use all videos from TUM-dynamics and Bonn. Following MonST3R [65], we subsample every 3 frames from the first 270 frames from TUM-dynamics to save compute. We follow the standard pose evaluation process of aligning camera poses with Umeyama alignment [56]. We report Absolute Translation Error (ATE), Relative Translation and Rotation Error (RPE trans and RPE rot)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Results. As reported in Tab. 1, Uni4D achieves competitive results across all metrics and datasets, highlighting the generalizability and performance of our pipeline. Uni4D is flexible to the availability of camera intrinsics, showing further improvement with known camera intrinsics. Training-based approaches such as LEAP-VO achieve good results on synthetic dataset like Sintel but does not generalize well to real-world datasets. Our method matches performance with MonST3R [65] on Sintel and achieves significantly better results on real-world datasets even compared to methods using ground-truth intrinsics." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1122" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.312, + 0.108 + ], + "angle": 0, + "content": "4.3. Video Depth Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.237 + ], + "angle": 0, + "content": "Baselines. For video depth accuracy evaluations, we focus on the top performing metric depth estimators, namely Metric3Dv2 [19], Depth-Pro [63], DepthCrafter [20] and Unidepth [44]. Metric depth estimators are trained without scale and shift alignment, making them strong baselines for video depth estimation. We also include the same baselines as our pose evaluations for joint 4D modeling approaches, namely CasualSAM [68], RCVD [30], and MonST3R [65]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.26, + 0.484, + 0.381 + ], + "angle": 0, + "content": "**Benchmarks.** We evaluate video depth estimates on Sintel [7], Bonn [40] and KITTI [16]. We follow standard video depth evaluation protocols [20] of aligning global shift and scale to predicted video depthmaps. We report the absolute relative error (Abs Rel) and percent of inlier points \\((\\delta < 1.25)\\). All methods are aligned in disparity space using the same least-squares alignment. We additionally report scale-aligned depth estimates similar to MonST3R." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.406, + 0.484, + 0.572 + ], + "angle": 0, + "content": "Results. By leveraging the strong performance of single-frame metric depth estimation models, Tab. 2 shows Uni4D achieves competitive depth estimation results, with superior performance among joint depth and pose methods, and closely matching the performance of single-frame depth estimation models on some datasets. With per-sequence scale, our model produces more accurate depth estimates across all datasets compared to recently released MonST3R. Our method closely retains depth estimation accuracy of underlying Unidepth depthmaps, while significantly improving its depth consistency shown in Sec. 4.5." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.585, + 0.214, + 0.601 + ], + "angle": 0, + "content": "4.4. Qualitative" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.608, + 0.484, + 0.79 + ], + "angle": 0, + "content": "We further show qualitative results of our reconstructions on the DAVIS [43] dataset. We apply MonST3R's own confidence-guided fusion for their reconstruction. We fuse casualSAM's depthmaps with dynamic masking by thresholding its uncertainty prediction to output 4D reconstruction. We use highlights to indicate dynamic objects at different timesteps. Throughout our qualitative results in Figs. 4,5,6,7, CasualSAM produces warped geometry and poor dynamic segmentations. MonST3R produces poor geometry in far regions and noisy dynamic masks and shapes. Our model produces the cleanest dynamic segmentations, along with the best dynamic and static geometry." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.802, + 0.245, + 0.818 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We ablate our multi-stage optimization in Tab. 4, highlighting the importance of each stage. Joint 4D optimization adds complexity that requires strong initialization for optimal convergence. Stage 1 introduces drift, which Stage 2 rectifies, resulting in superior final pose estimates." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.089, + 0.587, + 0.121 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.125, + 0.586, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.159, + 0.586, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.195, + 0.586, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.23, + 0.586, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.09, + 0.744, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.659, + 0.176, + 0.675, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.175, + 0.716, + 0.188 + ], + "angle": 0, + "content": "Undepth Bird's Eye" + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.198, + 0.733, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.088, + 0.898, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.812, + 0.175, + 0.829, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.84, + 0.175, + 0.872, + 0.188 + ], + "angle": 0, + "content": "Ours Bird's Eye" + }, + { + "type": "image", + "bbox": [ + 0.751, + 0.198, + 0.883, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.267, + 0.907, + 0.336 + ], + "angle": 0, + "content": "Figure 8. Depth Consistency. Direct fusion of Unidepth [44] predictions causes misaligned scene geometry visible as a layered wall in bird's-eye view, and jittering dynamic motion. Through motion priors and alignment, Uni4D produces a thin, crisp wall structure and smooth dynamic motion." + }, + { + "type": "table", + "bbox": [ + 0.55, + 0.341, + 0.868, + 0.41 + ], + "angle": 0, + "content": "
MethodSintel
SC ↓δSC<0.01 ↑δSC<0.05 ↑
Unidepth [44]0.10931.876.8
Uni4D0.04369.388.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.412, + 0.906, + 0.44 + ], + "angle": 0, + "content": "Table 3. Video depth consistency on Sintel. Uni4D improves Unidepth in consistency." + }, + { + "type": "table", + "bbox": [ + 0.547, + 0.444, + 0.871, + 0.526 + ], + "angle": 0, + "content": "
MethodSintel
ATE ↓RPE trans ↓RPE rot ↓
Uni4D (stage 1 only)0.1500.0510.551
Uni4D (stage 2 only)0.5870.1934.12
Uni4D (full)0.1100.0320.338
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.527, + 0.906, + 0.556 + ], + "angle": 0, + "content": "Table 4. Multi-stage Ablation. We evaluate pose estimation results on Sintel for both stage 1 and stage 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.572, + 0.907, + 0.739 + ], + "angle": 0, + "content": "Directly reprojecting Unidepth depth maps leads to flickering geometry. We ablate our fused depthmap quantitatively in Tab. 3, showing that Uni4D significantly improves consistency using Self-Consistency (SC) metrics [24]. SC measures depth errors between estimated and reprojected depth maps in static regions. Furthermore, we qualitatively show the effectiveness of Uni4D in rectifying Unidepth inconsistencies in Fig. 8. Our reconstruction achieves much better geometric and temporal consistency over Unidepth. Please refer to the supplementary for more ablative results regarding choice of foundation models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.754, + 0.634, + 0.77 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "This paper presents Uni4D, a framework unifying visual foundation models and structured energy minimization for dynamic 4D modeling from casual video. Our key insight is to optimize a 4D representation that aligns with visual cues from foundation models while following motion and geometry priors. Results show state-of-the-art performance with superior visual quality on Sintel, DAVIS, TUM-Dynamics and Bonn datasets, without any retraining or fine-tuning." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1123" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.114, + 0.484, + 0.186 + ], + "angle": 0, + "content": "This project is supported by the Intel AI SRS gift, Amazon-Illinois AICE grant, Meta Research Grant, IBM IIDAI Grant, and NSF Awards #2331878, #2340254, #2312102, #2414227, and #2404385. We greatly appreciate the NCSA for providing computing resources." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.21, + 0.188, + 0.226 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.235, + 0.484, + 0.304 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.305, + 0.484, + 0.359 + ], + "angle": 0, + "content": "[2] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building rome in a day. Communications of the ACM, 54 (10):105-112, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.361, + 0.484, + 0.429 + ], + "angle": 0, + "content": "[3] Antonio Agudo, Melcior Pijoan, and Francesc Moreno-Noguer. Image collection pop-up: 3d reconstruction and clustering of rigid and non-rigid categories. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2607-2615, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.43, + 0.484, + 0.485 + ], + "angle": 0, + "content": "[4] Gwangbin Bae and Andrew J Davison. Rethinking inductive biases for surface normal estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9535-9545, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.486, + 0.484, + 0.553 + ], + "angle": 0, + "content": "[5] Aleksei Bochkovskii, Amael Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.555, + 0.484, + 0.623 + ], + "angle": 0, + "content": "[6] Christoph Bregler, Aaron Hertzmann, and Henning Biermann. Recovering non-rigid 3d shape from image streams. In Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662), pages 690-696. IEEE, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.624, + 0.484, + 0.707 + ], + "angle": 0, + "content": "[7] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part VI 12, pages 611-625. Springer, 2012. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.708, + 0.484, + 0.775 + ], + "angle": 0, + "content": "[8] Weirong Chen, Le Chen, Rui Wang, and Marc Pollefeys. Leap-vo: Long-term effective any point tracking for visual odometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19844–19853, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.777, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[9] Ho Kei Cheng, Seoung Wug Oh, Brian Price, Alexander Schwing, and Joon-Young Lee. Tracking anything with decoupled video segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1316–1326, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[10] Hainan Cui, Xiang Gao, Shuhan Shen, and Zhanyi Hu. Hsfm: Hybrid structure-from-motion. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1212-1221, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.235, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[11] Andrew J Davison, Ian D Reid, Nicholas D Molton, and Olivier Stasse. Monoslam: Real-time single camera slam. IEEE transactions on pattern analysis and machine intelligence, 29(6):1052-1067, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.906, + 0.23 + ], + "angle": 0, + "content": "[12] Matt Deitke, Christopher Clark, Sangho Lee, Rohun Tripathi, Yue Yang, Jae Sung Park, Mohammadreza Salehi, Niklas Muennighoff, Kyle Lo, Luca Soldaini, et al. Molmo and pixmo: Open weights and open data for state-of-the-art multimodal models. arXiv preprint arXiv:2409.17146, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.232, + 0.905, + 0.285 + ], + "angle": 0, + "content": "[13] Jakob Engel, Thomas Schöps, and Daniel Cremers. Lsdslam: Large-scale direct monocular slam. In European conference on computer vision, pages 834-849. Springer, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[14] Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. In European Conference on Computer Vision, pages 241-258. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[15] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.467 + ], + "angle": 0, + "content": "[16] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.535 + ], + "angle": 0, + "content": "[17] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa*, and Jitendra Malik*. Humans in 4D: Reconstructing and tracking humans with transformers. In International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.539, + 0.906, + 0.607 + ], + "angle": 0, + "content": "[18] Zan Gojcic, Or Litany, Andreas Wieser, Leonidas J Guibas, and Tolga Birdal. Weakly supervised learning of rigid 3d scene flow. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5692-5703, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.906, + 0.691 + ], + "angle": 0, + "content": "[19] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. arXiv preprint arXiv:2404.15506, 2024.7,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.692, + 0.906, + 0.748 + ], + "angle": 0, + "content": "[20] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.906, + 0.802 + ], + "angle": 0, + "content": "[21] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv preprint arXiv:2307.07635, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.906, + 0.872 + ], + "angle": 0, + "content": "[22] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker3: Simpler and better point tracking by pseudolabelling real videos. arXiv preprint arXiv:2410.11831, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurpos" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.091, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1124" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "ing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.218 + ], + "angle": 0, + "content": "[24] Numair Khan, Eric Penner, Douglas Lanman, and Lei Xiao. Temporally consistent online depth estimation using point-based fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9119-9129, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[25] Rawal Khirodkar, Timur Bagautdinov, Julieta Martinez, Su Zhaoen, Austin James, Peter Selednik, Stuart Anderson, and Shunsuke Saito. Sapiens: Foundation for human vision models. In European Conference on Computer Vision, pages 206-228. Springer, 2025. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.483, + 0.319 + ], + "angle": 0, + "content": "[26] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.483, + 0.401 + ], + "angle": 0, + "content": "[27] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.404, + 0.483, + 0.459 + ], + "angle": 0, + "content": "[28] Chen Kong and Simon Lucey. Prior-less compressible structure from motion. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4123-4131, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.483, + 0.516 + ], + "angle": 0, + "content": "[29] Chen Kong, Rui Zhu, Hamed Kiani, and Simon Lucey. Structure from category: A generic and prior-less approach. In 2016 Fourth International Conference on 3D Vision (3DV), pages 296-304. IEEE, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.484, + 0.573 + ], + "angle": 0, + "content": "[30] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.483, + 0.631 + ], + "angle": 0, + "content": "[31] Guillaume Le Moing, Jean Ponce, and Cordelia Schmid. Dense optical tracking: connecting the dots. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187-19197, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[32] Jiahui Lei, Yijia Weng, Adam Harley, Leonidas Guibas, and Kostas Daniilidis. Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds. arXiv preprint arXiv:2405.17421, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.483, + 0.743 + ], + "angle": 0, + "content": "[33] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Ep n p: An accurate o (n) solution to the p np problem. International journal of computer vision, 81:155-166, 2009. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[34] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.856 + ], + "angle": 0, + "content": "[35] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4273-4284, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[36] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[37] Wei-Chiu Ma, Shenlong Wang, Rui Hu, Yuwen Xiong, and Raquel Urtasun. Deep rigid instance scene flow. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3614-3622, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[38] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[39] Richard A Newcombe, Dieter Fox, and Steven M Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 343-352, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[40] Emanuele Palazzolo, Jens Behley, Philipp Lottes, Philippe Giguere, and Cyril Stachniss. Refusion: 3d reconstruction in dynamic environments for rgb-d cameras exploiting residuals. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7855-7862. IEEE, 2019. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[41] Linfei Pan, Daniel Barath, Marc Pollefeys, and Johannes L Schonberger. Global structure-from-motion revisited. In European Conference on Computer Vision (ECCV), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[42] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[43] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.603, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[44] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Undepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10106-10116, 2024. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[45] Luigi Piccinelli, Christos Sakaridis, Yung-Hsu Yang, Mattia Segu, Siyuan Li, Wim Abbeloos, and Luc Van Gool. Undepthv2: Universal monocular metric depth estimation made simpler. arXiv preprint arXiv:2502.20110, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[46] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[47] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[48] Noah Snavely, Steven M Seitz, and Richard Szeliski. Modeling the world from internet photo collections. International journal of computer vision, 80:189-210, 2008. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1125" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[49] Chonghyuk Song, Gengshan Yang, Kangle Deng, Jun-Yan Zhu, and Deva Ramanan. Total-recon: Deformable scene reconstruction for embodied view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17671-17682, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[50] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Symposium on Geometry processing, pages 109–116. CiteSeer, 2007. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[51] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of rgb-d slam systems. In 2012 IEEE/RSJ international conference on intelligent robots and systems, pages 573-580. IEEE, 2012. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[52] Yu Sun, Qian Bao, Wu Liu, Tao Mei, and Michael J Black. Trace: 5d temporal regression of avatars with dynamic cameras in 3d environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8856-8866, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.402 + ], + "angle": 0, + "content": "[53] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[54] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. Advances in Neural Information Processing Systems, 36, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[55] Bill Triggs, Philip F McLauchlan, Richard I Hartley, and Andrew W Fitzgibbon. Bundle adjustment—a modern synthesis. In Vision Algorithms: Theory and Practice: International Workshop on Vision Algorithms Corfu, Greece, September 21–22, 1999 Proceedings, pages 298–372. Springer, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[56] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on Pattern Analysis & Machine Intelligence, 13(04): 376-380, 1991. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[57] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[58] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.799 + ], + "angle": 0, + "content": "[59] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[60] Yuefan Wu, Zeyuan Chen, Shaowei Liu, Zhongzheng Ren, and Shenlong Wang. Casa: Category-agnostic skeletal animal reconstruction. Advances in Neural Information Processing Systems, 35:28559-28574, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[61] Gengshan Yang, Minh Vo, Natalia Neverova, Deva Ramanan, Andrea Vedaldi, and Hanbyul Joo. Banmo: Building animatable 3d neural models from many casual videos. In" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2863-2873, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[62] Gengshan Yang, Shuo Yang, John Z Zhang, Zachary Manchester, and Deva Ramanan. Ppr: Physically plausible reconstruction from monocular videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3914-3924, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.907, + 0.261 + ], + "angle": 0, + "content": "[63] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[64] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[65] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825, 2024. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.457 + ], + "angle": 0, + "content": "[66] John Z Zhang, Shuo Yang, Gengshan Yang, Arun L Bishop, Swaminathan Gurumurthy, Deva Ramanan, and Zachary Manchester. Slomo: A general system for legged robot motion imitation from casual videos. IEEE Robotics and Automation Letters, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.46, + 0.905, + 0.542 + ], + "angle": 0, + "content": "[67] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1724-1732, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.544, + 0.905, + 0.613 + ], + "angle": 0, + "content": "[68] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T Freeman. Structure and motion from casual videos. In European Conference on Computer Vision, pages 20-37. Springer, 2022. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.615, + 0.905, + 0.682 + ], + "angle": 0, + "content": "[69] Wang Zhao, Shaohui Liu, Hengkai Guo, Wenping Wang, and Yong-Jin Liu. *Particlesfm: Exploiting dense point trajectories for localizing moving cameras in the wild*. In European Conference on Computer Vision, pages 523-542. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.685, + 0.905, + 0.753 + ], + "angle": 0, + "content": "[70] Yingying Zhu, Dong Huang, Fernando De La Torre, and Simon Lucey. Complex non-rigid motion 3d reconstruction by union of subspaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1542-1549, 2014. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.753 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1126" + } + ] +] \ No newline at end of file diff --git a/2025/Uni4D_ Unifying Visual Foundation Models for 4D Modeling from a Single Video/b0bfb21a-5579-4e48-8be6-e0a25d4d029f_origin.pdf b/2025/Uni4D_ Unifying Visual Foundation Models for 4D Modeling from a Single Video/b0bfb21a-5579-4e48-8be6-e0a25d4d029f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..12ec34546ed68a9a23c76abd32a1a1c890cc87b2 --- /dev/null +++ b/2025/Uni4D_ Unifying Visual Foundation Models for 4D Modeling from a Single Video/b0bfb21a-5579-4e48-8be6-e0a25d4d029f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c34dd2794d8cee8f2946600723a9bb01946afd73d34b3d62a939a1ec6cab89b +size 8710331 diff --git a/2025/Uni4D_ Unifying Visual Foundation Models for 4D Modeling from a Single Video/full.md b/2025/Uni4D_ Unifying Visual Foundation Models for 4D Modeling from a Single Video/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a79f936cfaa147e545dec86d6097dba00be76848 --- /dev/null +++ b/2025/Uni4D_ Unifying Visual Foundation Models for 4D Modeling from a Single Video/full.md @@ -0,0 +1,399 @@ +# Uni4D: Unifying Visual Foundation Models for 4D Modeling from a Single Video + +David Yifan Yao Albert J. Zhai Shenlong Wang University of Illinois at Urbana-Champaign + +![](images/8d5a68cd8104e8c9ef41a266d3ea84d68c1f4bac745d0a671f92caf9ae9157de.jpg) +Figure 1. Given a casually captured video, Uni4D harnesses pretrained visual foundation models and multi-stage optimization to jointly estimate camera poses, dynamic geometry, and dense 3D motion. The resulting camera poses and geometry are accurate, consistent, and coherent both temporally and spatially. This is all done without any additional training or fine-tuning. + +![](images/c0741c4b9d98bb85d12de55a7b5097b9139c7076824b0261f814e7bcfb06049b.jpg) + +# Abstract + +This paper presents a unified approach to understanding dynamic scenes from casual videos. Large pretrained vision foundation models, such as vision-language, video depth prediction, motion tracking, and segmentation models, offer promising capabilities. However, training a single model for comprehensive 4D understanding remains challenging. We introduce Uni4D, a multi-stage optimization framework that harnesses multiple pretrained models to advance dynamic 3D modeling, including static/dynamic reconstruction, camera pose estimation, and dense 3D motion tracking. Our results show state-of-the-art performance in dynamic 4D modeling with superior visual quality. Notably, Uni4D requires no retraining or fine-tuning, highlighting the effectiveness of repurposing visual foundation models for 4D understanding. Code and more results are available at: https://davidya099.github.io/uni4d. + +# 1. Introduction + +Over the past two years, many visual foundation models have emerged [4, 21, 23, 25, 27, 31, 44, 58, 63], achieving high accuracy on tasks like depth prediction, segmentation, human parsing, normal estimation, few-view recon + +struction, and motion tracking. These models leverage supervised learning on large, diverse datasets, achieving impressive accuracy and remarkable generalization. However, these advances have not translated to 4D (time + geometry) modeling, a longstanding challenge in computer vision. We see two main reasons: first, collecting high-quality, ground-truth 4D data from real-world environments remains complex and resource-intensive. Second, 4D understanding is a holistic problem involving interconnected tasks like camera pose estimation, 3D reconstruction, and dynamics tracking. Although each subtask shows progress, data-driven cues remain noisy, and more importantly, unifying them synergistically for holistic modeling remains challenging. This paper seeks to answer: Can we harness the success of visual foundation models for dynamic 4D modeling? + +In this paper, we propose Uni4D, a novel framework that reconstructs 4D scenes from a single video captured in the wild. Our method integrates data-driven foundation models and conventional model-driven dynamic structure-from-motion, combining data-driven cues and model-based knowledge synergistically. Our intuition is simple: each data-driven visual cue, such as video segmentation, pixel-level motion tracking, and video depth, is a partial projection from the 4D world to the 2D video. The key is to create a 4D scene representation that coherently aligns with each cue while incorporating strong prior knowledge + +of real-world motion and shape to resolve temporary inconsistencies. To achieve this, we take an energy-minimization perspective, framing the problem as an optimization task that jointly infers camera poses, static and dynamic geometry, and motion. The dynamic 4D modeling framework is complex, as it involves various optimization variables, visual cues, and constraints. To overcome the challenge of joint reasoning, we carefully design a novel three-stage, divide-and-conquer pipeline that progressively incorporates camera poses, static geometry, and dynamic geometry and motion into the optimization framework. Uni4D leverages pretrained foundation models across different tasks, requiring no task-specific retraining or fine-tuning. This design sidesteps the need for 4D ground-truth data, a major challenge in the field. Through incorporating strong priors on geometry and dynamic motion, our method produces realistic 4D scenes that are coherent across space and time while maintaining high accuracy. + +We demonstrate the effectiveness of Uni4D on various datasets. As shown in Fig. 1, our method effectively recovers the clean geometry and motion of the scene, as well as camera trajectories, from a single video. As shown in Fig. 2, our model outperforms all dynamic 4D modeling baselines in both camera pose and geometry. + +# 2. Related Works + +Structure from Motion and SLAM. Structure from Motion (SfM) is a classical problem in computer vision that aims to jointly recover camera parameters and 3D structure from images [2, 10, 41, 47, 48]. Simultaneous localization and mapping (SLAM) is a similar problem that focuses on real-time efficiency in an online setting [11, 13, 38, 53]. The core idea for many approaches to SfM and SLAM is to optimize a form of reprojection error with respect to both camera parameters and 3D points, also known as bundle adjustment [55]. This relies on the key assumption that the scene being viewed is rigid ("static"), making these approaches not applicable to dynamic scenes. + +![](images/78379f0a216985cae1c7c7b2d32abd724fde2148f237936d74eb3b1738427295.jpg) +Figure 2. Uni4D outperforms other recent 4D modeling methods in both camera pose and geometry accuracy on the Sintel dataset. + +4D Understanding. The problem of non-rigid structure from motion is highly ill-posed. Previous works on non-rigid SfM leveraged various forms of priors or weaker assumptions on shape and or motion to make the problem solvable [3, 6, 28, 29, 70]. With the rise of modern segmentation models, many works have studied category-specific articulation and rigidity priors, focusing on object categories such as humans [17, 49, 52, 64], animals [60-62, 66], and vehicles [18, 37]. The generality of such methods is limited by their category-specific nature. Recently, several works have also begun to leverage trained neural networks as priors for 4D understanding in a more freeform manner [30, 65, 68, 69]. However, these methods involve either training a network [65, 69] or optimizing specific layers of an existing network [30, 68], making it difficult to integrate newer models into the pipeline. Our method takes the strategy further and integrates pretrained models in a completely modular manner, fully unleashing their generalization capabilities and allowing for seamless integration of newer models. Note that various recent works have also studied 4D reconstruction in an optimization-based framework [15, 32, 35, 57, 59]. However, these methods have additional rendering capacities and focus on rendering metrics, while we focus on recovering high-quality geometry. + +Visual Foundation Models. In recent years, a number of visual foundation models have been developed, achieving remarkable performance on tasks like depth estimation [5, 20, 23, 44, 63], detection and segmentation [12, 27, 36, 46], human parsing [25], surface normal estimation [4, 14], few-view reconstruction [34, 58], and point tracking [21, 22, 31]. Our insight is that nearly all of these models have the potential to contribute towards holistic 4D understanding, and harnessing them in a unifying framework can advance the state-of-the-art in tasks such as non-rigid structure from motion. To this end, we integrate the following pretrained models in Uni4D: UniDepthv2 [45] for geometry initialization, CoTracker3 [22] for correspondence initialization, and a collection of Recognize Anything Model [67], ChatGPT [1], Grounding-SAM [27, 36], and DEVA [9] for dynamic object segmentation. Through multi-stage optimization with a few regularizing priors, we are able to use these models to extract accurate pose and 4D geometry from monocular video. + +# 3. Method + +In this paper, we are interested in recovering 4D geometry and camera parameters from a monocular casual video. Our model is built on the intuition that the 2D visual cues from the video can be seen as perspective projections of 4D geometry and motion, where the video depth represents a projection of 4D geometry, 2D dense tracking corresponds to a projection of 4D motion, and segmentation reflects a + +![](images/d8ce7c042acb1ab62c4a2977507cfb5462ef20ee69131d51165d7e7bf959e86a.jpg) +Figure 3. Given a casually captured video, Uni4D exploits visual foundation models to extract dynamic segmentation, video depth, and motion tracks. Static geometry and poses are obtained through tracklet-based structure-from-motion along with camera motion priors. Dynamic geometry is improved through nonrigid bundle adjustment and scene motion priors. A final fusion densifies geometry to obtain high quality 4D reconstruction. + +projection of the 4D dynamic object silhouette. We propose a novel, training-free, foundation-model-based energy minimization scheme that leverages visual cues to jointly infer camera poses, geometry, and motion, enabling generalization across diverse videos. + +# 3.1. Pretrained Visual Cues + +We first exploit foundation models to extract visual cues, including dynamic segmentation, video depth, and motion tracking. All of the models are pre-trained on large and diverse datasets and exhibit strong generalization abilities. + +Video Segmentation. Recognizing and segmenting dynamic objects is crucial for 4D scene understanding. We leverage the latest advancements in video semantic segmentation and tracking to estimate dynamic objects over time. First, we use RAM [67] to identify semantic classes in the video. These classes are then filtered through GPT-4o [1] to exclude static and background elements (e.g., buildings, poles), retaining only dynamic objects (e.g., humans, animals, vehicles). Next, Grounding-SAM [27, 36] performs segmentation at each keyframe, and DEVA [9] tracks these segments over time, resulting in accurate dynamic video segmentation $\{\mathbf{M}_{\mathbf{t}}\}_{t = 0}^{T}$ + +Dense Motion Tracking. We use dense pixel tracking to establish correspondences over time, which serve as motion cues to assist inference of geometry reconstruction and dynamic object 3D motion. Unlike traditional 4D reconstruction methods relying on optical flow [65, 68], dense motion tracklets yield more correspondence pairs across large viewpoints and structural changes. Pixel tracking also outperforms sparse matching in density and surpasses flow + +propagation in robustness, making it ideal for dynamic 4D reconstruction. We utilize Co-TrackerV3 [22] for its robustness, as it employs a 4D cost volume with an attention mechanism to track 2D points through occlusions, recently proving to be effective for dynamic neural rendering. We apply Co-Tracker bi-directionally on a dense grid every 10 frames to ensure thorough coverage. We filter and classify tracklets using segmentation masks yielding a set of correspondent point trajectories $\{\mathbf{Z}_k\in \mathbb{R}^{T\times 2}\}_{k = 0}^K$ at visible time steps determined by Co-Tracker. + +Video Depth Estimation. Monocular depth reasoning, while insufficient as a standalone tool for complete 4D geometry recovery, provides strong geometry initialization cues. In our model, we use UniDepthV2 [45], a monocular depth estimation network, to estimate an initial depth map, $\{\mathbf{D}_t\}_{t=0}^T$ , and initial camera intrinsics, $\mathbf{K}_{\mathrm{init}}$ . + +# 3.2. Energy Formulation + +We now describe the energy formulation of our proposed dynamic 4D reconstruction framework. Let $\mathcal{M} = \{\mathbf{M}_t\}_{t=0}^T$ , $\mathcal{D} = \{\mathbf{D}_t\}_{t=0}^T$ , $\mathcal{Z} = \{\mathbf{Z}_k\}_{k=0}^K$ be the input dynamic segmentation, monocular depth, dense motion trajectory extracted from the input video $\mathcal{I} = \{\mathbf{I}_t\}_{t=0}^T$ . Formally, our goal is to obtain camera parameters $\mathcal{C}$ , namely poses $\mathcal{T}$ and intrinsics $\mathbf{K}$ , and a set of 4D point clouds $\mathcal{P}$ containing both dynamic and static parts separately as $\mathcal{P} = \{\mathbf{P}_{\mathrm{static}}, \mathcal{P}_{\mathrm{dyn}}\}$ . Here, $\mathbf{P}_{\mathrm{static}}$ does not change over time and $\mathcal{P}_{\mathrm{dyn}} = \{\mathbf{p}_k \in \mathbb{R}^{T \times 3}\}_k$ is a temporally varying point cloud, where each $\mathbf{p}_k$ is a dynamic point trajectory. We represent the camera poses as rigid transforms $\mathcal{T} = \{\xi_t \in \mathbb{S}\mathbb{E}(3)\}_{t=0}^T$ , and parameterize the rotations as $\mathfrak{so}(3)$ rotation vectors as a minimal representation for easy optimization. We assume all frames share the same intrinsic + +![](images/834930e04012363df5016111a3bee8ce3495a8ca2a9da60075df68c2210066de.jpg) + +![](images/247789a78f903db6aad089be37d4686ba23524b29dc2e5de20bf5239cac878ab.jpg) + +![](images/a33c4e9277c71089a2a2638ef4762215479de867b6a9f52deb7c0149ea7a520e.jpg) + +![](images/6226712897f8335ff927d38bc98b15ef6c348ccf59843de21125380284717855.jpg) + +![](images/8c5c854b24ae6d177707345b61a44d8ec348c4c04e16912680c77c64b54cda10.jpg) +Input Video + +![](images/fe38f0cf37f9a93c7f58ef1aae41a21b23169af57cd0db1cf0c864591fee9322.jpg) +CasualSAM Static + +![](images/b4771ac96413e646bb92cc1089cf5cd59d9c40ef0f49f0bca1808dd5dd81a2a5.jpg) +CasualSAM 4D + +![](images/f553aeebeeafcf1c8636677221770b21e1c0ed5c91d29dd0cce0d0ca7cf4c942.jpg) +CasualSAM Bird's Eye + +![](images/66690b622dd088d3bd6cc54dd4775d3cce24ee0f93180903dcf91e3fc2cdc27a.jpg) +Monst3R Static + +![](images/e1feff5e9a91584c6778b64d7fdb3472be9a31f26d87c384e28245874c0b4a5d.jpg) +Monst3R 4D + +![](images/337696d7026100c9076a1e374816381bb8eea04e2651b08d51770c393f5aaf1d.jpg) +Monst3R Bird's Eye + +![](images/2bf43608fd6f44f172acd05fa6a475d4f6cb08fabd19cc424f999da22d6a0e3d.jpg) +Ours Static + +![](images/73c9f62c338b5cbc2e05ca2f7a75ae1fb92310322955f64b4b109f3f8e341690.jpg) +Ours 4D +Figure 4. Qualitative results of 4D reconstruction on the DAVIS [43] dataset. CasualSAM [68] suffers from slanted geometry, and Monst3R [65] has unclear geometry and does not resolve conflicts from multiple views (note the wall in the bird's-eye view). Both CasualSAM and Monst3R lack clean dynamic reconstruction and segmentation. Uni4D achieves a realistic layout, thanks to joint optimization, and provides accurate dynamic segmentation and reconstruction by leveraging foundation visual models as cues. + +![](images/cab83464e20a235e71f8960502452634221a646975c294da08ca627e07266bb1.jpg) +Ours Bird's Eye + +matrix $\mathbf{K}$ where we optimize focal lengths $f_{x}$ and $f_{y}$ . We formulate the 4D joint reasoning problem as minimization of the following energy function: + +$$ +E _ {\mathrm {B A}} (\mathcal {C}, \mathbf {P} _ {\text {s t a t i c}}) + E _ {\mathrm {N R}} (\mathcal {P} _ {\mathrm {d y n}}) + E _ {\text {m o t i o n}} (\mathcal {P} _ {\mathrm {d y n}}) + E _ {\text {c a m}} (\mathcal {T}) \tag {1} +$$ + +where $E_{\mathrm{BA}}(\mathcal{C}, \mathbf{P}_{\mathrm{static}})$ is a bundle adjustment term that measures the discrepancy between static-region correspondences and the static 3D structure $\mathbf{P}_{\mathrm{static}}$ through perspective reprojection. $E_{\mathrm{NR}}(\mathcal{P}_{\mathrm{dyn}})$ is a non-rigid structure-from-motion energy term that measures the disagreement between the dynamic point cloud and their tracklet correspondences, $E_{\mathrm{cam}}(\mathcal{T})$ is a regularization term on the camera motion smoothness, and $E_{\mathrm{motion}}(\mathcal{P}_{\mathrm{dyn}})$ is a regularization term on the dynamic structure and motion. Each energy term is involved in different stages of the optimization process, which will be described in Sec. 3.3. + +Static Bundle Adjustment Term. The bundle adjustment energy $E_{\mathrm{BA}}(\mathcal{C}, \mathbf{P}_{\mathrm{static}})$ measures the consistency between + +the pixel-level correspondences and the 3D structure of static scene elements. Given the input pixel tracks $\mathcal{Z} = \{\mathbf{Z}_k\}$ and video segmentation $\mathcal{M}$ , we filter all tracks corresponding to static areas and minimize the distance between the projected pixel location and the observed pixel location: + +$$ +E _ {\mathrm {B A}} (\mathcal {C}, \mathbf {P} _ {\text {s t a t i c}}; \mathcal {Z}, \mathcal {M}) = \sum_ {\mathbf {z} _ {k} \in \mathcal {M}} \sum_ {t} \mathbf {w} _ {\mathbf {k}, \mathbf {t}} \| \mathbf {z} _ {k, t} - \pi_ {\mathbf {K}} \left(\mathbf {p} _ {k}, \boldsymbol {\xi} _ {t}\right) \| _ {2} \tag {2} +$$ + +where $\mathbf{z}_{k,t}$ is the $k$ -th 3D point's corresponding pixel track's 2D coordinates at time $t$ , $\mathbf{w_{k,t}} \in \{0,1\}$ is a visibility indicator and $\pi_{\mathbf{K}}$ is the perspective projection function. + +Non-Rigid Bundle Adjustment Term. For dynamic objects, we impose a nonrigid bundle adjustment term, $E_{\mathrm{NR}}(\mathcal{P}_{\mathrm{dyn}})$ , which measures the discrepancy between the dynamic point cloud and pixel tracklets. Here, each pixel tracklet corresponds to a dynamic 3D point sequence, + +![](images/f37dc1e04d2f370fb41c79d8847603ba08af17dfc7808c0fd6a4c38e891423e0.jpg) +Figure 5. Qualitative results on Bonn dataset. Both CasualSAM and MonST3R have trailing artifacts and incorrect dynamic estimations. Uni4D provides clear dynamic and static geometry. + +$\{\mathbf{p}_{k,t}\}_{t}$ , optimized for each observed tracklet: + +$$ +E _ {\mathrm {N R}} \left(\mathcal {P} _ {\mathrm {d y n}}; \mathcal {C}, \mathcal {Z}, \mathcal {M}\right) = \sum_ {\mathbf {z} _ {k} \in \mathcal {M}} \sum_ {t} \mathbf {w} _ {\mathbf {k}, \mathbf {t}} \left\| \mathbf {z} _ {k, t} - \pi_ {\mathbf {K}} \left(\mathbf {p} _ {k, t}, \boldsymbol {\xi} _ {t}\right) \right\| _ {2} \tag {3} +$$ + +where $\mathbf{p}_{k,t} \in \mathbb{R}^3$ is the $k$ -th dynamic point's location at $t$ . + +Camera Motion Prior. Considering that our input is a video, we incorporate a temporal smoothness prior on camera poses that penalizes sudden changes in relative pose: $\pmb{\xi}_{t\rightarrow t + 1} = \pmb{\xi}_{t + 1}^{-1}\cdot \pmb{\xi}_t$ . We reweight this term based on the magnitude of the relative motion: intuitively, if the relative motion is large, we penalize change rates in relative motion less; if the relative motion is small, we apply a higher penalty on change rate. Formally, we have: + +$$ +E _ {\mathrm {c a m}} (\mathcal {T}) = \sum_ {t} E _ {\mathrm {r o t}} \left(\mathbf {R} _ {t - 1, t, t + 1}\right) + \sum_ {t} E _ {\mathrm {t r a n s}} \left(\mathbf {t} _ {t - 1, t, t + 1}\right) \tag {4} +$$ + +where $E_{\mathrm{rot}}(\mathbf{R}_{t - 1,t,t + 1}) = \frac{2\|\mathrm{rad}(\mathbf{R}_{t\to t + 1}) - \mathrm{rad}(\mathbf{R}_{t - 1\to t})\|}{\|\mathrm{rad}(\mathbf{R}_{t - 1\to t})\| + \|\mathrm{rad}(\mathbf{R}_{t\to t + 1})\|}$ and $E_{\mathrm{trans}}(\mathbf{t}_{t - 1,t,t + 1}) = \frac{2\|\mathbf{t}_{t\to t + 1} - \mathbf{t}_{t - 1\to t}\|}{\|\mathbf{t}_{t - 1\to t}\| + \|\mathbf{t}_{t\to t + 1}\|}$ ; rad converts the rotation matrix into absolute radians. + +Dynamic Motion Prior. $E_{\mathrm{motion}}(\mathcal{P}_{\mathrm{dyn}})$ is a regularization term that encodes the characteristics of the dynamic structure. It contains two prior terms that are used to regularize the dynamic structure, both of which have demonstrated effectiveness in previous work [39, 50]: + +$$ +E _ {\text {m o t i o n}} \left(\mathcal {P} _ {\text {d y n}}\right) = E _ {\text {a r a p}} \left(\mathcal {P} _ {\text {d y n}}\right) + E _ {\text {s m o o t h}} \left(\mathcal {P} _ {\text {d y n}}\right). \tag {5} +$$ + +$E_{\mathrm{arap}}$ is an as-rigid-as-possible (ARAP) [50] prior that penalizes extreme deformations that compromise local rigidity. Specifically, we obtain the nearest neighbors of each dynamic control point $k$ by applying KNN over the other tracks and enforce that the relative distances between these close-by pairs do not undergo sudden changes: + +$$ +E _ {\text {a r a p}} = \sum_ {\mathbf {t}} \sum_ {(k, m)} \mathbf {w} _ {\mathbf {k m}, \mathbf {t}} \| d \left(\mathbf {p} _ {\mathbf {k}, \mathbf {t}}, \mathbf {p} _ {m, t}\right) - d \left(\mathbf {p} _ {\mathbf {k}, \mathbf {t} + 1}, \mathbf {p} _ {m, t + 1}\right) \| _ {2} \tag {6} +$$ + +where $d(\cdot, \cdot)$ is the L2 distance and $\mathbf{w}_{\mathbf{k}\mathbf{m},\mathbf{t}} = 1$ if all relevant points are visible. + +$E_{\mathrm{smooth}}$ is a simple smoothness term that promotes temporal smoothness for the dynamic point cloud: + +$$ +E _ {\text {s m o o t h}} = \sum_ {\mathbf {t}} \sum_ {\mathbf {p} _ {k} \in \mathcal {P} _ {\text {d y n}}} \mathbf {w} _ {\mathbf {k}, \mathbf {t}} \| \mathbf {p} _ {k, t} - \mathbf {p} _ {k, t + 1} \| _ {2}. \tag {7} +$$ + +Despite simplicity, both motion terms are crucial in our formulation, as they significantly reduce ambiguities in 4D dynamic structure estimation, which is highly ill-posed. Unlike other methods, we do not assume strong model-based motion priors, such as rigid motion [37], articulated motion [61], or a linear motion basis [57]. + +# 3.3. Inference + +Directly minimizing the energy defined in Eq. 1 is nontrivial, as our energy function is highly non-linear and involves millions of free variables. To address this, we developed a three-stage optimization pipeline, enabling us to minimize the energy and estimate the scene variables in a divide-and-conquer fashion. + +Stage 1: Camera Initialization. We start by initializing camera parameters. Combining video depth estimation $\mathcal{D}$ and dense pixel motion $\mathcal{Z}$ allows us to establish 2D-to-3D correspondences. This allows us to initialize and tune $\mathcal{C}$ by minimizing the following energy function with respect to camera parameters only. Specifically, we can unproject each video frame's depth at time $t$ back to 3D and minimize the following energy function: + +$$ +\min _ {\mathbf {C}} \sum_ {\left(t ^ {\prime}, t\right)} \sum_ {\mathbf {z} _ {k} \in \neg \mathcal {M}} \| \mathbf {z} _ {k, t ^ {\prime}} - \pi_ {\mathbf {K}} \left(\pi_ {\mathbf {K}} ^ {- 1} \left(\mathbf {z} _ {k, t}, \mathbf {D} _ {t}, \boldsymbol {\xi} _ {t}\right), \boldsymbol {\xi} _ {t ^ {\prime}}\right) \| _ {2} ^ {2} \tag {8} +$$ + +where $\pi_{\mathbf{K}}^{-1}$ is the unprojection function that maps 2D coordinates into 3D world coordinates using estimated depth $\mathbf{D}_t$ . We perform this over all pairs within a temporal sliding window of 5 frames, producing a good initial pose estimate as shown in Tab. 4. Given camera initialization $\hat{\mathcal{C}}$ , we directly unproject our depth prediction into a common world coordinate system, which provides an initial 4D structure $\hat{\mathcal{P}}$ . This is used as initialization for later optimization. + +Stage 2: Bundle Adjustment. Our second stage jointly optimizes camera pose and static geometry by minimizing the static component-related energy in a bundle adjustment fashion. Formally speaking, we solve the following: + +$$ +\min _ {\mathcal {C}, \mathbf {P} _ {\text {s t a t i c}}} E _ {\mathrm {B A}} \left(\mathcal {C}, \mathbf {P} _ {\text {s t a t i c}}; \mathcal {Z}, \mathcal {M}\right) + E _ {\mathrm {c a m}} (\mathcal {T}) \tag {9} +$$ + +By enforcing consistency with each other, this improves both the static geometry and the camera pose quality. We perform a final scene integration by unprojecting correspondences into 3D using improved pose and filtering outlier noisy points in 3D. + +![](images/10cc49294d795eb17d072b528e4bcc702067db7f77b979e3dd24a1c820fed94c.jpg) + +![](images/44c81bbada89dc995ea93df0de7375a4d7c7d14368c22f6541d55b50ddb8dc54.jpg) + +![](images/567f9c46a8bcb0eb644ada89f30f61d76beec164c931a9bd006d99697bf7ebef.jpg) + +![](images/cb8fbb9d5c7a86cd616d357bad64832d0e69574f24b8e9c05e046eeb10b7a77b.jpg) + +![](images/d948dd0ff7d74a774ceea539861a9b0ac83859daeb725cc5ae424066c692d138.jpg) +Input Video + +![](images/b5385671bef8f8af1d9d37ef6bde1ce7695c1f80551782313b3b10be04dda362.jpg) +CasualSAM Static + +![](images/506995d07fba5dcfed6a0d409aaba68be698542faabf10452d3d09f83f4e9070.jpg) +CasualSAM 4D + +![](images/591ab479303c6367c072c19b78dab2608794cd804efc771dc1f3c68e26d36dc6.jpg) + +![](images/24c11c9c46de54934afd8a13c58dd489327ef5842ababa0f72f6b2aa4d17221e.jpg) +MonST3R Static + +![](images/49ed163bd00d2d62977fe1cc0c8f6b8fd00113a51b6fa8d62158e031c4b3a5ec.jpg) + +![](images/231d89d3a4532147a2c9258877d0cc9c67fcafc5d2062518ed83785bd86cea0c.jpg) +Ours 4D + +![](images/6475bb994ced197f35942cf9a35d7080d4e585c855a1fd5647a7f206d4fde379.jpg) +CasualSAM Bird's Eye +MonST3R Bird's Eye + +![](images/f4500869aa7c95b53fa95ede09d0b7c12c819f447f550b5d0f3b1efcfbb6eeac.jpg) +Ours Bird's Eye + +Stage 3: Non-Rigid Bundle Adjustment. Given the estimated camera pose, our third stage focuses on inferring dynamic structure. Note that we freeze camera parameters in this stage, as we find that incorrect geometry and motion evidence often harm camera pose estimation rather than improve it. Additionally, enabling camera pose optimization introduces extra flexibility in this ill-posed problem, harming robustness. Formally speaking, we solve the following: + +$$ +\min _ {\mathcal {P} _ {\mathrm {d y n}}} E _ {\mathrm {N R}} \left(\mathcal {P} _ {\mathrm {d y n}}; \mathbf {C}, \mathcal {Z}, \mathcal {M}\right) + E _ {\text {m o t i o n}} \left(\mathcal {P} _ {\mathrm {d y n}}\right) \tag {10} +$$ + +We initialize $\mathcal{P}_{\mathrm{dyn}}$ using video depth and our optimized camera pose from Stage 2. We scale $E_{\mathrm{smooth}}$ and $E_{\mathrm{arap}}$ with constants 10 and 100 respectively which we found empirically led to better dynamics. This energy optimization might still leave some high-energy noisy points, often from incorrect cues, motion boundaries, or occlusions. We filter these outliers based on their energy values in a final step. + +Fusion. From our energy minimization, we acquire a semi-dense dynamic and static point cloud along with camera parameters. To further densify the point cloud, enabling each pixel to correspond to a 3D point, we perform depth-based interpolation by computing a scale offset. Details can be found in the supplementary. + +To unify our outputs into a consistent 4D representation, we use our camera parameters and static masks to project the aligned depth maps into world coordinates. To handle noisy depth values at boundaries, we create an edge mask filter by thresholding gradients of the depth maps. + +This process enables us to update the depth value for + +![](images/13f5c037775fe836d11053ed5c70993c5e8280c1391b9dd51fb5f24a02aadcb1.jpg) +Figure 6. Qualitative results of 4D reconstruction on DAVIS dataset. CasualSAM [68] distorts the room geometry as evident from the bird's eye view. Dynamic elements have inconsistent geometry over time. MonST3R [65] has noisy static geometry in the room's far corner and incomplete dynamic object geometry evident in the green highlight. Uni4D has the cleanest dynamic object reconstruction and segmentation results, with geometrically accurate room shapes as evident from the bird's eye view. +Figure 7. Qualitative results of 4D reconstruction on Sintel dataset. We highlight and contrast performance on (1) dynamic objects and (2) geometry on planar surfaces. + +each high-confidence pixel, which we can then reproject into 3D space to obtain the final dense dynamic geometry reconstruction result, as shown in Fig. 1 and Fig. 3. + +# 4. Experiments + +Uni4D estimates camera pose, depth, and 3D motion from a single video. We perform experiments evaluating its performance with respect to baselines on these tasks. + +# 4.1. Implementation Details + +For optimization, we use the Adam optimizer with ReduceLROnPlateau and EarlyStopping in PyTorch [26, 42]. We perform 600 iterations per sliding window in stage 1, 2000 iterations in stage 2, and 1000 iterations in stage 3. + +
CategoryMethodSintelTUM-dynamicsBonn
ATE ↓RPE trans ↓RPE rot ↓ATE ↓RPE trans ↓RPE rot ↓ATE ↓RPE trans ↓RPE rot ↓
Pose onlyDPVO*[54]0.1710.0631.2910.0190.0140.4060.0220.0140.913
LEAP-VO*[8]0.0350.0651.6690.0250.0312.8430.0370.0140.844
Joint depth & poseRobust-CVD[30]0.3680.1533.4620.0960.0272.5900.0850.0180.803
CasualSAM[68]0.1370.0390.6300.0360.0180.7450.0240.0140.849
Monst3R[65]0.1080.0430.7290.1080.0221.3710.0230.0110.807
Uni4D0.1100.0320.3380.0120.0040.3350.0170.0100.818
Uni4D*0.0920.0330.1410.0120.0040.3310.0160.0100.817
+ +Table 1. Camera Pose Evaluation on Sintel, TUM-dynamic, and Bonn datasets. We bold and underline the best and second best results respectively. * indicates known camera intrinsic. + +
AlignmentCategoryMethodSintelBonnKITTI
Abs Rel ↓ δ<1.25 ↑Abs Rel ↓ δ<1.25 ↑Abs Rel ↓ δ<1.25 ↑
Per-sequence scale & shiftSingle-frame depthMetric3D[19]0.20571.90.04498.50.03998.8
Depth-pro[5]0.28060.50.04998.60.08094.2
Unidepth[44]0.19872.80.04098.50.03898.8
Video depthDepthCrafter[20]0.23169.00.06597.60.11288.4
Joint video depth & poseRobust-CVD[30]0.35849.70.10889.80.18272.9
CasualSAM[68]0.29256.90.06996.60.11388.3
Monst3r[65]0.35852.10.06095.00.08591.9
Uni4D0.21672.50.03898.30.09889.7
Per-sequence scaleJoint depth & poseMonst3r[65]0.34455.90.04198.20.08991.4
Joint depth & poseUni4D0.28964.90.03898.30.08693.3
+ +Table 2. Video depth evaluation on Sintel, Bonn, and KITTI datasets. We bold and underline the best and second best results respectively. + +We initialize the learning rate at $1 \times 10^{-3}$ for stage 1, and $1 \times 10^{-2}$ for stages 2 and 3, reducing all to $1 \times 10^{-4}$ . Our entire framework takes roughly 5 minutes for a 50-frame video on a RTX A6000 GPU. We include a detailed runtime breakdown in the supplementary. We run all baselines on the datasets using their official implementations and hyperparameters. Co-Trackers [22] are initialized at a dense $50 \times 50$ grid, with a $75 \times 75$ grid for Sintel to handle its large camera perspective change. All optimization hyperparameters are kept the same for all runs on all datasets. + +# 4.2. Pose Estimation + +Baselines. We compare with several recent methods for pose estimation in dynamic scenes. LEAPVO [8] and DPVO [54] are learning-based visual odometry methods. Robust-CVD [30] optimizes for pose and depth deformation through an SFM pipeline. CasualSAM [68] further improves on the idea by directly finetuning network weights along with a novel uncertainty formulation. Monst3R [65] is a very recent work that fine-tunes DUSt3R [58] for 4D reconstruction through PnP [33]. + +Benchmarks and metrics. We evaluate pose estimation on three dynamic datasets: Sintel [7], TUM-dynamics [51], and Bonn [40]. We follow LEAP-VO's evaluation split for Sintel and use all videos from TUM-dynamics and Bonn. Following MonST3R [65], we subsample every 3 frames from the first 270 frames from TUM-dynamics to save compute. We follow the standard pose evaluation process of aligning camera poses with Umeyama alignment [56]. We report Absolute Translation Error (ATE), Relative Translation and Rotation Error (RPE trans and RPE rot). + +Results. As reported in Tab. 1, Uni4D achieves competitive results across all metrics and datasets, highlighting the generalizability and performance of our pipeline. Uni4D is flexible to the availability of camera intrinsics, showing further improvement with known camera intrinsics. Training-based approaches such as LEAP-VO achieve good results on synthetic dataset like Sintel but does not generalize well to real-world datasets. Our method matches performance with MonST3R [65] on Sintel and achieves significantly better results on real-world datasets even compared to methods using ground-truth intrinsics. + +# 4.3. Video Depth Evaluation + +Baselines. For video depth accuracy evaluations, we focus on the top performing metric depth estimators, namely Metric3Dv2 [19], Depth-Pro [63], DepthCrafter [20] and Unidepth [44]. Metric depth estimators are trained without scale and shift alignment, making them strong baselines for video depth estimation. We also include the same baselines as our pose evaluations for joint 4D modeling approaches, namely CasualSAM [68], RCVD [30], and MonST3R [65]. + +**Benchmarks.** We evaluate video depth estimates on Sintel [7], Bonn [40] and KITTI [16]. We follow standard video depth evaluation protocols [20] of aligning global shift and scale to predicted video depthmaps. We report the absolute relative error (Abs Rel) and percent of inlier points $(\delta < 1.25)$ . All methods are aligned in disparity space using the same least-squares alignment. We additionally report scale-aligned depth estimates similar to MonST3R. + +Results. By leveraging the strong performance of single-frame metric depth estimation models, Tab. 2 shows Uni4D achieves competitive depth estimation results, with superior performance among joint depth and pose methods, and closely matching the performance of single-frame depth estimation models on some datasets. With per-sequence scale, our model produces more accurate depth estimates across all datasets compared to recently released MonST3R. Our method closely retains depth estimation accuracy of underlying Unidepth depthmaps, while significantly improving its depth consistency shown in Sec. 4.5. + +# 4.4. Qualitative + +We further show qualitative results of our reconstructions on the DAVIS [43] dataset. We apply MonST3R's own confidence-guided fusion for their reconstruction. We fuse casualSAM's depthmaps with dynamic masking by thresholding its uncertainty prediction to output 4D reconstruction. We use highlights to indicate dynamic objects at different timesteps. Throughout our qualitative results in Figs. 4,5,6,7, CasualSAM produces warped geometry and poor dynamic segmentations. MonST3R produces poor geometry in far regions and noisy dynamic masks and shapes. Our model produces the cleanest dynamic segmentations, along with the best dynamic and static geometry. + +# 4.5. Ablation Study + +We ablate our multi-stage optimization in Tab. 4, highlighting the importance of each stage. Joint 4D optimization adds complexity that requires strong initialization for optimal convergence. Stage 1 introduces drift, which Stage 2 rectifies, resulting in superior final pose estimates. + +![](images/c3abe6cd08cb56ffe4c3f2a6571664e3a5d964e710d9fa4eb0e48ed1b76d773b.jpg) + +![](images/ec1b32e6e75c1904c19aa48ce29d046fa377c3c451f069abc03fdf843987f964.jpg) + +![](images/46c81b931261bd7fcf3ea4d0fa4374cc681d9a854a7d0d1fbbc2dd1f4010df9c.jpg) + +![](images/8146d9bcc40619c7660947afdfa42a6eb53c54925bbb21c16779482cd3af1e4c.jpg) + +![](images/f015f7ddf543d49d2921b7e0a7aa0397f1a6a6f97aa42ec529e4344d55b8188b.jpg) + +![](images/33130ec47c4ae58385798c67a722d5cc9fb72cd91c723f5c676646fa527e2ddf.jpg) +Undepth Bird's Eye + +![](images/c2d07328020d760daa3a3b630632d92ccabd6c77cbd36217bd5e49d205766de4.jpg) + +![](images/c70d37987e4634952b6f696822399651c262dd66e14990eded2cf6621b247df7.jpg) +Figure 8. Depth Consistency. Direct fusion of Unidepth [44] predictions causes misaligned scene geometry visible as a layered wall in bird's-eye view, and jittering dynamic motion. Through motion priors and alignment, Uni4D produces a thin, crisp wall structure and smooth dynamic motion. + +![](images/89d421e2141e8b84a518a2e637656252a945fadb67be716a0fd90169ca401eed.jpg) +Ours Bird's Eye + +![](images/1712028f1251fdd8f1300cc7b4ecbf2932d6fe6f20664c71f2da842ace651e8c.jpg) + +![](images/ebbae18983fa690b97b546328258ddda6a722c4caf4741570065b935c2d0b786.jpg) + +
MethodSintel
SC ↓δSC<0.01 ↑δSC<0.05 ↑
Unidepth [44]0.10931.876.8
Uni4D0.04369.388.1
+ +Table 3. Video depth consistency on Sintel. Uni4D improves Unidepth in consistency. + +
MethodSintel
ATE ↓RPE trans ↓RPE rot ↓
Uni4D (stage 1 only)0.1500.0510.551
Uni4D (stage 2 only)0.5870.1934.12
Uni4D (full)0.1100.0320.338
+ +Table 4. Multi-stage Ablation. We evaluate pose estimation results on Sintel for both stage 1 and stage 2. + +Directly reprojecting Unidepth depth maps leads to flickering geometry. We ablate our fused depthmap quantitatively in Tab. 3, showing that Uni4D significantly improves consistency using Self-Consistency (SC) metrics [24]. SC measures depth errors between estimated and reprojected depth maps in static regions. Furthermore, we qualitatively show the effectiveness of Uni4D in rectifying Unidepth inconsistencies in Fig. 8. Our reconstruction achieves much better geometric and temporal consistency over Unidepth. Please refer to the supplementary for more ablative results regarding choice of foundation models. + +# 5. Conclusion + +This paper presents Uni4D, a framework unifying visual foundation models and structured energy minimization for dynamic 4D modeling from casual video. Our key insight is to optimize a 4D representation that aligns with visual cues from foundation models while following motion and geometry priors. Results show state-of-the-art performance with superior visual quality on Sintel, DAVIS, TUM-Dynamics and Bonn datasets, without any retraining or fine-tuning. + +# Acknowledgements + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 3 +[2] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building rome in a day. Communications of the ACM, 54 (10):105-112, 2011. 2 +[3] Antonio Agudo, Melcior Pijoan, and Francesc Moreno-Noguer. Image collection pop-up: 3d reconstruction and clustering of rigid and non-rigid categories. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2607-2615, 2018. 2 +[4] Gwangbin Bae and Andrew J Davison. Rethinking inductive biases for surface normal estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9535-9545, 2024. 1, 2 +[5] Aleksei Bochkovskii, Amael Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 2, 7 +[6] Christoph Bregler, Aaron Hertzmann, and Henning Biermann. Recovering non-rigid 3d shape from image streams. In Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662), pages 690-696. IEEE, 2000. 2 +[7] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part VI 12, pages 611-625. Springer, 2012. 7, 8 +[8] Weirong Chen, Le Chen, Rui Wang, and Marc Pollefeys. Leap-vo: Long-term effective any point tracking for visual odometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19844–19853, 2024. 7 +[9] Ho Kei Cheng, Seoung Wug Oh, Brian Price, Alexander Schwing, and Joon-Young Lee. Tracking anything with decoupled video segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1316–1326, 2023. 2, 3 +[10] Hainan Cui, Xiang Gao, Shuhan Shen, and Zhanyi Hu. Hsfm: Hybrid structure-from-motion. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1212-1221, 2017. 2 + +This project is supported by the Intel AI SRS gift, Amazon-Illinois AICE grant, Meta Research Grant, IBM IIDAI Grant, and NSF Awards #2331878, #2340254, #2312102, #2414227, and #2404385. We greatly appreciate the NCSA for providing computing resources. +[11] Andrew J Davison, Ian D Reid, Nicholas D Molton, and Olivier Stasse. Monoslam: Real-time single camera slam. IEEE transactions on pattern analysis and machine intelligence, 29(6):1052-1067, 2007. 2 +[12] Matt Deitke, Christopher Clark, Sangho Lee, Rohun Tripathi, Yue Yang, Jae Sung Park, Mohammadreza Salehi, Niklas Muennighoff, Kyle Lo, Luca Soldaini, et al. Molmo and pixmo: Open weights and open data for state-of-the-art multimodal models. arXiv preprint arXiv:2409.17146, 2024. 2 +[13] Jakob Engel, Thomas Schöps, and Daniel Cremers. Lsdslam: Large-scale direct monocular slam. In European conference on computer vision, pages 834-849. Springer, 2014. 2 +[14] Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. In European Conference on Computer Vision, pages 241-258. Springer, 2025. 2 +[15] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2 +[16] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 8 +[17] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa*, and Jitendra Malik*. Humans in 4D: Reconstructing and tracking humans with transformers. In International Conference on Computer Vision (ICCV), 2023. 2 +[18] Zan Gojcic, Or Litany, Andreas Wieser, Leonidas J Guibas, and Tolga Birdal. Weakly supervised learning of rigid 3d scene flow. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5692-5703, 2021. 2 +[19] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. arXiv preprint arXiv:2404.15506, 2024.7,8 +[20] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024. 2, 7, 8 +[21] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv preprint arXiv:2307.07635, 2023. 1, 2 +[22] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker3: Simpler and better point tracking by pseudolabelling real videos. arXiv preprint arXiv:2410.11831, 2024. 2, 3, 7 +[23] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurpos + +ing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 1, 2 +[24] Numair Khan, Eric Penner, Douglas Lanman, and Lei Xiao. Temporally consistent online depth estimation using point-based fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9119-9129, 2023. 8 +[25] Rawal Khirodkar, Timur Bagautdinov, Julieta Martinez, Su Zhaoen, Austin James, Peter Selednik, Stuart Anderson, and Shunsuke Saito. Sapiens: Foundation for human vision models. In European Conference on Computer Vision, pages 206-228. Springer, 2025. 1, 2 +[26] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[27] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2, 3 +[28] Chen Kong and Simon Lucey. Prior-less compressible structure from motion. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4123-4131, 2016. 2 +[29] Chen Kong, Rui Zhu, Hamed Kiani, and Simon Lucey. Structure from category: A generic and prior-less approach. In 2016 Fourth International Conference on 3D Vision (3DV), pages 296-304. IEEE, 2016. 2 +[30] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. 2, 7, 8 +[31] Guillaume Le Moing, Jean Ponce, and Cordelia Schmid. Dense optical tracking: connecting the dots. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187-19197, 2024. 1, 2 +[32] Jiahui Lei, Yijia Weng, Adam Harley, Leonidas Guibas, and Kostas Daniilidis. Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds. arXiv preprint arXiv:2405.17421, 2024. 2 +[33] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Ep n p: An accurate o (n) solution to the p np problem. International journal of computer vision, 81:155-166, 2009. 7 +[34] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 2 +[35] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4273-4284, 2023. 2 +[36] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with + +grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 3 +[37] Wei-Chiu Ma, Shenlong Wang, Rui Hu, Yuwen Xiong, and Raquel Urtasun. Deep rigid instance scene flow. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3614-3622, 2019. 2, 5 +[38] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015. 2 +[39] Richard A Newcombe, Dieter Fox, and Steven M Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 343-352, 2015. 5 +[40] Emanuele Palazzolo, Jens Behley, Philipp Lottes, Philippe Giguere, and Cyril Stachniss. Refusion: 3d reconstruction in dynamic environments for rgb-d cameras exploiting residuals. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7855-7862. IEEE, 2019. 7, 8 +[41] Linfei Pan, Daniel Barath, Marc Pollefeys, and Johannes L Schonberger. Global structure-from-motion revisited. In European Conference on Computer Vision (ECCV), 2024. 2 +[42] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6 +[43] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 4, 8 +[44] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Undepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10106-10116, 2024. 1, 2, 7, 8 +[45] Luigi Piccinelli, Christos Sakaridis, Yung-Hsu Yang, Mattia Segu, Siyuan Li, Wim Abbeloos, and Luc Van Gool. Undepthv2: Universal monocular metric depth estimation made simpler. arXiv preprint arXiv:2502.20110, 2025. 2, 3 +[46] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2 +[47] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2 +[48] Noah Snavely, Steven M Seitz, and Richard Szeliski. Modeling the world from internet photo collections. International journal of computer vision, 80:189-210, 2008. 2 + +[49] Chonghyuk Song, Gengshan Yang, Kangle Deng, Jun-Yan Zhu, and Deva Ramanan. Total-recon: Deformable scene reconstruction for embodied view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17671-17682, 2023. 2 +[50] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Symposium on Geometry processing, pages 109–116. CiteSeer, 2007. 5 +[51] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of rgb-d slam systems. In 2012 IEEE/RSJ international conference on intelligent robots and systems, pages 573-580. IEEE, 2012. 7 +[52] Yu Sun, Qian Bao, Wu Liu, Tao Mei, and Michael J Black. Trace: 5d temporal regression of avatars with dynamic cameras in 3d environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8856-8866, 2023. 2 +[53] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 2 +[54] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. Advances in Neural Information Processing Systems, 36, 2024. 7 +[55] Bill Triggs, Philip F McLauchlan, Richard I Hartley, and Andrew W Fitzgibbon. Bundle adjustment—a modern synthesis. In Vision Algorithms: Theory and Practice: International Workshop on Vision Algorithms Corfu, Greece, September 21–22, 1999 Proceedings, pages 298–372. Springer, 2000. 2 +[56] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on Pattern Analysis & Machine Intelligence, 13(04): 376-380, 1991. 7 +[57] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 2, 5 +[58] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 1, 2, 7 +[59] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024. 2 +[60] Yuefan Wu, Zeyuan Chen, Shaowei Liu, Zhongzheng Ren, and Shenlong Wang. Casa: Category-agnostic skeletal animal reconstruction. Advances in Neural Information Processing Systems, 35:28559-28574, 2022. 2 +[61] Gengshan Yang, Minh Vo, Natalia Neverova, Deva Ramanan, Andrea Vedaldi, and Hanbyul Joo. Banmo: Building animatable 3d neural models from many casual videos. In + +Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2863-2873, 2022. 5 +[62] Gengshan Yang, Shuo Yang, John Z Zhang, Zachary Manchester, and Deva Ramanan. Ppr: Physically plausible reconstruction from monocular videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3914-3924, 2023. 2 +[63] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 1, 2, 8 +[64] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[65] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825, 2024. 2, 3, 4, 6, 7, 8 +[66] John Z Zhang, Shuo Yang, Gengshan Yang, Arun L Bishop, Swaminathan Gurumurthy, Deva Ramanan, and Zachary Manchester. Slomo: A general system for legged robot motion imitation from casual videos. IEEE Robotics and Automation Letters, 2023. 2 +[67] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1724-1732, 2024. 2, 3 +[68] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T Freeman. Structure and motion from casual videos. In European Conference on Computer Vision, pages 20-37. Springer, 2022. 2, 3, 4, 6, 7, 8 +[69] Wang Zhao, Shaohui Liu, Hengkai Guo, Wenping Wang, and Yong-Jin Liu. *Particlesfm: Exploiting dense point trajectories for localizing moving cameras in the wild*. In European Conference on Computer Vision, pages 523-542. Springer, 2022. 2 +[70] Yingying Zhu, Dong Huang, Fernando De La Torre, and Simon Lucey. Complex non-rigid motion 3d reconstruction by union of subspaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1542-1549, 2014. 2 \ No newline at end of file diff --git a/2025/Uni4D_ Unifying Visual Foundation Models for 4D Modeling from a Single Video/images.zip b/2025/Uni4D_ Unifying Visual Foundation Models for 4D Modeling from a Single Video/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4ca21808213a6615240aefb2944f49a23d765da9 --- /dev/null +++ b/2025/Uni4D_ Unifying Visual Foundation Models for 4D Modeling from a Single Video/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5878b32ad9019234a310e70caabbb26264ef3dc3eff157101ea46b8b7d99a390 +size 921227 diff --git a/2025/Uni4D_ Unifying Visual Foundation Models for 4D Modeling from a Single Video/layout.json b/2025/Uni4D_ Unifying Visual Foundation Models for 4D Modeling from a Single Video/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a1e3475066d240b66b8f8db22ddd5ce465f653ca --- /dev/null +++ b/2025/Uni4D_ Unifying Visual Foundation Models for 4D Modeling from a Single Video/layout.json @@ -0,0 +1,10344 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 555, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 555, + 120 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 555, + 120 + ], + "type": "text", + "content": "Uni4D: Unifying Visual Foundation Models for 4D Modeling from a Single Video" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 162, + 143, + 447, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 143, + 447, + 172 + ], + "spans": [ + { + "bbox": [ + 162, + 143, + 447, + 172 + ], + "type": "text", + "content": "David Yifan Yao Albert J. Zhai Shenlong Wang University of Illinois at Urbana-Champaign" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 65, + 182, + 302, + 353 + ], + "blocks": [ + { + "bbox": [ + 65, + 182, + 302, + 353 + ], + "lines": [ + { + "bbox": [ + 65, + 182, + 302, + 353 + ], + "spans": [ + { + "bbox": [ + 65, + 182, + 302, + 353 + ], + "type": "image", + "image_path": "8d5a68cd8104e8c9ef41a266d3ea84d68c1f4bac745d0a671f92caf9ae9157de.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 356, + 555, + 390 + ], + "lines": [ + { + "bbox": [ + 54, + 356, + 555, + 390 + ], + "spans": [ + { + "bbox": [ + 54, + 356, + 555, + 390 + ], + "type": "text", + "content": "Figure 1. Given a casually captured video, Uni4D harnesses pretrained visual foundation models and multi-stage optimization to jointly estimate camera poses, dynamic geometry, and dense 3D motion. The resulting camera poses and geometry are accurate, consistent, and coherent both temporally and spatially. This is all done without any additional training or fine-tuning." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 306, + 180, + 545, + 353 + ], + "blocks": [ + { + "bbox": [ + 306, + 180, + 545, + 353 + ], + "lines": [ + { + "bbox": [ + 306, + 180, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 306, + 180, + 545, + 353 + ], + "type": "image", + "image_path": "c0741c4b9d98bb85d12de55a7b5097b9139c7076824b0261f814e7bcfb06049b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 399, + 200, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 399, + 200, + 412 + ], + "spans": [ + { + "bbox": [ + 151, + 399, + 200, + 412 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 426, + 297, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 426, + 297, + 617 + ], + "spans": [ + { + "bbox": [ + 53, + 426, + 297, + 617 + ], + "type": "text", + "content": "This paper presents a unified approach to understanding dynamic scenes from casual videos. Large pretrained vision foundation models, such as vision-language, video depth prediction, motion tracking, and segmentation models, offer promising capabilities. However, training a single model for comprehensive 4D understanding remains challenging. We introduce Uni4D, a multi-stage optimization framework that harnesses multiple pretrained models to advance dynamic 3D modeling, including static/dynamic reconstruction, camera pose estimation, and dense 3D motion tracking. Our results show state-of-the-art performance in dynamic 4D modeling with superior visual quality. Notably, Uni4D requires no retraining or fine-tuning, highlighting the effectiveness of repurposing visual foundation models for 4D understanding. Code and more results are available at: https://davidya099.github.io/uni4d." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 644, + 135, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 644, + 135, + 656 + ], + "spans": [ + { + "bbox": [ + 55, + 644, + 135, + 656 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Over the past two years, many visual foundation models have emerged [4, 21, 23, 25, 27, 31, 44, 58, 63], achieving high accuracy on tasks like depth prediction, segmentation, human parsing, normal estimation, few-view recon" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 401, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 555, + 581 + ], + "type": "text", + "content": "struction, and motion tracking. These models leverage supervised learning on large, diverse datasets, achieving impressive accuracy and remarkable generalization. However, these advances have not translated to 4D (time + geometry) modeling, a longstanding challenge in computer vision. We see two main reasons: first, collecting high-quality, ground-truth 4D data from real-world environments remains complex and resource-intensive. Second, 4D understanding is a holistic problem involving interconnected tasks like camera pose estimation, 3D reconstruction, and dynamics tracking. Although each subtask shows progress, data-driven cues remain noisy, and more importantly, unifying them synergistically for holistic modeling remains challenging. This paper seeks to answer: Can we harness the success of visual foundation models for dynamic 4D modeling?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we propose Uni4D, a novel framework that reconstructs 4D scenes from a single video captured in the wild. Our method integrates data-driven foundation models and conventional model-driven dynamic structure-from-motion, combining data-driven cues and model-based knowledge synergistically. Our intuition is simple: each data-driven visual cue, such as video segmentation, pixel-level motion tracking, and video depth, is a partial projection from the 4D world to the 2D video. The key is to create a 4D scene representation that coherently aligns with each cue while incorporating strong prior knowledge" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1116" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 287 + ], + "type": "text", + "content": "of real-world motion and shape to resolve temporary inconsistencies. To achieve this, we take an energy-minimization perspective, framing the problem as an optimization task that jointly infers camera poses, static and dynamic geometry, and motion. The dynamic 4D modeling framework is complex, as it involves various optimization variables, visual cues, and constraints. To overcome the challenge of joint reasoning, we carefully design a novel three-stage, divide-and-conquer pipeline that progressively incorporates camera poses, static geometry, and dynamic geometry and motion into the optimization framework. Uni4D leverages pretrained foundation models across different tasks, requiring no task-specific retraining or fine-tuning. This design sidesteps the need for 4D ground-truth data, a major challenge in the field. Through incorporating strong priors on geometry and dynamic motion, our method produces realistic 4D scenes that are coherent across space and time while maintaining high accuracy." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 288, + 294, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 294, + 359 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 294, + 359 + ], + "type": "text", + "content": "We demonstrate the effectiveness of Uni4D on various datasets. As shown in Fig. 1, our method effectively recovers the clean geometry and motion of the scene, as well as camera trajectories, from a single video. As shown in Fig. 2, our model outperforms all dynamic 4D modeling baselines in both camera pose and geometry." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 369, + 148, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 148, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 148, + 383 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 389, + 295, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 295, + 532 + ], + "type": "text", + "content": "Structure from Motion and SLAM. Structure from Motion (SfM) is a classical problem in computer vision that aims to jointly recover camera parameters and 3D structure from images [2, 10, 41, 47, 48]. Simultaneous localization and mapping (SLAM) is a similar problem that focuses on real-time efficiency in an online setting [11, 13, 38, 53]. The core idea for many approaches to SfM and SLAM is to optimize a form of reprojection error with respect to both camera parameters and 3D points, also known as bundle adjustment [55]. This relies on the key assumption that the scene being viewed is rigid (\"static\"), making these approaches not applicable to dynamic scenes." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 91, + 533, + 254, + 689 + ], + "blocks": [ + { + "bbox": [ + 91, + 533, + 254, + 689 + ], + "lines": [ + { + "bbox": [ + 91, + 533, + 254, + 689 + ], + "spans": [ + { + "bbox": [ + 91, + 533, + 254, + 689 + ], + "type": "image", + "image_path": "78379f0a216985cae1c7c7b2d32abd724fde2148f237936d74eb3b1738427295.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 690, + 295, + 712 + ], + "lines": [ + { + "bbox": [ + 55, + 690, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 690, + 295, + 712 + ], + "type": "text", + "content": "Figure 2. Uni4D outperforms other recent 4D modeling methods in both camera pose and geometry accuracy on the Sintel dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 361 + ], + "type": "text", + "content": "4D Understanding. The problem of non-rigid structure from motion is highly ill-posed. Previous works on non-rigid SfM leveraged various forms of priors or weaker assumptions on shape and or motion to make the problem solvable [3, 6, 28, 29, 70]. With the rise of modern segmentation models, many works have studied category-specific articulation and rigidity priors, focusing on object categories such as humans [17, 49, 52, 64], animals [60-62, 66], and vehicles [18, 37]. The generality of such methods is limited by their category-specific nature. Recently, several works have also begun to leverage trained neural networks as priors for 4D understanding in a more freeform manner [30, 65, 68, 69]. However, these methods involve either training a network [65, 69] or optimizing specific layers of an existing network [30, 68], making it difficult to integrate newer models into the pipeline. Our method takes the strategy further and integrates pretrained models in a completely modular manner, fully unleashing their generalization capabilities and allowing for seamless integration of newer models. Note that various recent works have also studied 4D reconstruction in an optimization-based framework [15, 32, 35, 57, 59]. However, these methods have additional rendering capacities and focus on rendering metrics, while we focus on recovering high-quality geometry." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 373, + 556, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 556, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 556, + 601 + ], + "type": "text", + "content": "Visual Foundation Models. In recent years, a number of visual foundation models have been developed, achieving remarkable performance on tasks like depth estimation [5, 20, 23, 44, 63], detection and segmentation [12, 27, 36, 46], human parsing [25], surface normal estimation [4, 14], few-view reconstruction [34, 58], and point tracking [21, 22, 31]. Our insight is that nearly all of these models have the potential to contribute towards holistic 4D understanding, and harnessing them in a unifying framework can advance the state-of-the-art in tasks such as non-rigid structure from motion. To this end, we integrate the following pretrained models in Uni4D: UniDepthv2 [45] for geometry initialization, CoTracker3 [22] for correspondence initialization, and a collection of Recognize Anything Model [67], ChatGPT [1], Grounding-SAM [27, 36], and DEVA [9] for dynamic object segmentation. Through multi-stage optimization with a few regularizing priors, we are able to use these models to extract accurate pose and 4D geometry from monocular video." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 609, + 371, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 371, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 371, + 621 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we are interested in recovering 4D geometry and camera parameters from a monocular casual video. Our model is built on the intuition that the 2D visual cues from the video can be seen as perspective projections of 4D geometry and motion, where the video depth represents a projection of 4D geometry, 2D dense tracking corresponds to a projection of 4D motion, and segmentation reflects a" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1117" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 64, + 531, + 248 + ], + "blocks": [ + { + "bbox": [ + 80, + 64, + 531, + 248 + ], + "lines": [ + { + "bbox": [ + 80, + 64, + 531, + 248 + ], + "spans": [ + { + "bbox": [ + 80, + 64, + 531, + 248 + ], + "type": "image", + "image_path": "d8ce7c042acb1ab62c4a2977507cfb5462ef20ee69131d51165d7e7bf959e86a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 555, + 297 + ], + "type": "text", + "content": "Figure 3. Given a casually captured video, Uni4D exploits visual foundation models to extract dynamic segmentation, video depth, and motion tracks. Static geometry and poses are obtained through tracklet-based structure-from-motion along with camera motion priors. Dynamic geometry is improved through nonrigid bundle adjustment and scene motion priors. A final fusion densifies geometry to obtain high quality 4D reconstruction." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 303, + 296, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 296, + 362 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 296, + 362 + ], + "type": "text", + "content": "projection of the 4D dynamic object silhouette. We propose a novel, training-free, foundation-model-based energy minimization scheme that leverages visual cues to jointly infer camera poses, geometry, and motion, enabling generalization across diverse videos." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 372, + 187, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 187, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 187, + 384 + ], + "type": "text", + "content": "3.1. Pretrained Visual Cues" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 390, + 296, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 296, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 296, + 439 + ], + "type": "text", + "content": "We first exploit foundation models to extract visual cues, including dynamic segmentation, video depth, and motion tracking. All of the models are pre-trained on large and diverse datasets and exhibit strong generalization abilities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 456, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 296, + 601 + ], + "type": "text", + "content": "Video Segmentation. Recognizing and segmenting dynamic objects is crucial for 4D scene understanding. We leverage the latest advancements in video semantic segmentation and tracking to estimate dynamic objects over time. First, we use RAM [67] to identify semantic classes in the video. These classes are then filtered through GPT-4o [1] to exclude static and background elements (e.g., buildings, poles), retaining only dynamic objects (e.g., humans, animals, vehicles). Next, Grounding-SAM [27, 36] performs segmentation at each keyframe, and DEVA [9] tracks these segments over time, resulting in accurate dynamic video segmentation " + }, + { + "bbox": [ + 55, + 456, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_{\\mathbf{t}}\\}_{t = 0}^{T}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "content": "Dense Motion Tracking. We use dense pixel tracking to establish correspondences over time, which serve as motion cues to assist inference of geometry reconstruction and dynamic object 3D motion. Unlike traditional 4D reconstruction methods relying on optical flow [65, 68], dense motion tracklets yield more correspondence pairs across large viewpoints and structural changes. Pixel tracking also outperforms sparse matching in density and surpasses flow" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 303, + 555, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 555, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 555, + 423 + ], + "type": "text", + "content": "propagation in robustness, making it ideal for dynamic 4D reconstruction. We utilize Co-TrackerV3 [22] for its robustness, as it employs a 4D cost volume with an attention mechanism to track 2D points through occlusions, recently proving to be effective for dynamic neural rendering. We apply Co-Tracker bi-directionally on a dense grid every 10 frames to ensure thorough coverage. We filter and classify tracklets using segmentation masks yielding a set of correspondent point trajectories " + }, + { + "bbox": [ + 313, + 303, + 555, + 423 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Z}_k\\in \\mathbb{R}^{T\\times 2}\\}_{k = 0}^K" + }, + { + "bbox": [ + 313, + 303, + 555, + 423 + ], + "type": "text", + "content": " at visible time steps determined by Co-Tracker." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 437, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 510 + ], + "type": "text", + "content": "Video Depth Estimation. Monocular depth reasoning, while insufficient as a standalone tool for complete 4D geometry recovery, provides strong geometry initialization cues. In our model, we use UniDepthV2 [45], a monocular depth estimation network, to estimate an initial depth map, " + }, + { + "bbox": [ + 313, + 437, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{D}_t\\}_{t=0}^T" + }, + { + "bbox": [ + 313, + 437, + 555, + 510 + ], + "type": "text", + "content": ", and initial camera intrinsics, " + }, + { + "bbox": [ + 313, + 437, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{\\mathrm{init}}" + }, + { + "bbox": [ + 313, + 437, + 555, + 510 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 517, + 432, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 432, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 432, + 529 + ], + "type": "text", + "content": "3.2. Energy Formulation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": "We now describe the energy formulation of our proposed dynamic 4D reconstruction framework. Let " + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\{\\mathbf{M}_t\\}_{t=0}^T" + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{\\mathbf{D}_t\\}_{t=0}^T" + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Z} = \\{\\mathbf{Z}_k\\}_{k=0}^K" + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": " be the input dynamic segmentation, monocular depth, dense motion trajectory extracted from the input video " + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\{\\mathbf{I}_t\\}_{t=0}^T" + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": ". Formally, our goal is to obtain camera parameters " + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": ", namely poses " + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": " and intrinsics " + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": ", and a set of 4D point clouds " + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": " containing both dynamic and static parts separately as " + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{\\mathbf{P}_{\\mathrm{static}}, \\mathcal{P}_{\\mathrm{dyn}}\\}" + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{static}}" + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": " does not change over time and " + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{dyn}} = \\{\\mathbf{p}_k \\in \\mathbb{R}^{T \\times 3}\\}_k" + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": " is a temporally varying point cloud, where each " + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_k" + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": " is a dynamic point trajectory. We represent the camera poses as rigid transforms " + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{\\xi_t \\in \\mathbb{S}\\mathbb{E}(3)\\}_{t=0}^T" + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": ", and parameterize the rotations as " + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathfrak{so}(3)" + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": " rotation vectors as a minimal representation for easy optimization. We assume all frames share the same intrinsic" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1118" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 104, + 91, + 205, + 147 + ], + "blocks": [ + { + "bbox": [ + 104, + 91, + 205, + 147 + ], + "lines": [ + { + "bbox": [ + 104, + 91, + 205, + 147 + ], + "spans": [ + { + "bbox": [ + 104, + 91, + 205, + 147 + ], + "type": "image", + "image_path": "834930e04012363df5016111a3bee8ce3495a8ca2a9da60075df68c2210066de.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 105, + 148, + 204, + 204 + ], + "blocks": [ + { + "bbox": [ + 105, + 148, + 204, + 204 + ], + "lines": [ + { + "bbox": [ + 105, + 148, + 204, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 204, + 204 + ], + "type": "image", + "image_path": "247789a78f903db6aad089be37d4686ba23524b29dc2e5de20bf5239cac878ab.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 105, + 207, + 205, + 262 + ], + "blocks": [ + { + "bbox": [ + 105, + 207, + 205, + 262 + ], + "lines": [ + { + "bbox": [ + 105, + 207, + 205, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 205, + 262 + ], + "type": "image", + "image_path": "a33c4e9277c71089a2a2638ef4762215479de867b6a9f52deb7c0149ea7a520e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 105, + 266, + 205, + 321 + ], + "blocks": [ + { + "bbox": [ + 105, + 266, + 205, + 321 + ], + "lines": [ + { + "bbox": [ + 105, + 266, + 205, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 205, + 321 + ], + "type": "image", + "image_path": "6226712897f8335ff927d38bc98b15ef6c348ccf59843de21125380284717855.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 105, + 326, + 205, + 384 + ], + "blocks": [ + { + "bbox": [ + 105, + 326, + 205, + 384 + ], + "lines": [ + { + "bbox": [ + 105, + 326, + 205, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 205, + 384 + ], + "type": "image", + "image_path": "8c5c854b24ae6d177707345b61a44d8ec348c4c04e16912680c77c64b54cda10.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 397, + 175, + 405 + ], + "lines": [ + { + "bbox": [ + 138, + 397, + 175, + 405 + ], + "spans": [ + { + "bbox": [ + 138, + 397, + 175, + 405 + ], + "type": "text", + "content": "Input Video" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 218, + 78, + 306, + 182 + ], + "blocks": [ + { + "bbox": [ + 218, + 78, + 306, + 182 + ], + "lines": [ + { + "bbox": [ + 218, + 78, + 306, + 182 + ], + "spans": [ + { + "bbox": [ + 218, + 78, + 306, + 182 + ], + "type": "image", + "image_path": "fe38f0cf37f9a93c7f58ef1aae41a21b23169af57cd0db1cf0c864591fee9322.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 236, + 184, + 291, + 191 + ], + "lines": [ + { + "bbox": [ + 236, + 184, + 291, + 191 + ], + "spans": [ + { + "bbox": [ + 236, + 184, + 291, + 191 + ], + "type": "text", + "content": "CasualSAM Static" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 327, + 78, + 411, + 179 + ], + "blocks": [ + { + "bbox": [ + 327, + 78, + 411, + 179 + ], + "lines": [ + { + "bbox": [ + 327, + 78, + 411, + 179 + ], + "spans": [ + { + "bbox": [ + 327, + 78, + 411, + 179 + ], + "type": "image", + "image_path": "b4771ac96413e646bb92cc1089cf5cd59d9c40ef0f49f0bca1808dd5dd81a2a5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 183, + 393, + 191 + ], + "lines": [ + { + "bbox": [ + 348, + 183, + 393, + 191 + ], + "spans": [ + { + "bbox": [ + 348, + 183, + 393, + 191 + ], + "type": "text", + "content": "CasualSAM 4D" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 438, + 78, + 481, + 179 + ], + "blocks": [ + { + "bbox": [ + 438, + 78, + 481, + 179 + ], + "lines": [ + { + "bbox": [ + 438, + 78, + 481, + 179 + ], + "spans": [ + { + "bbox": [ + 438, + 78, + 481, + 179 + ], + "type": "image", + "image_path": "f553aeebeeafcf1c8636677221770b21e1c0ed5c91d29dd0cce0d0ca7cf4c942.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 434, + 183, + 501, + 191 + ], + "lines": [ + { + "bbox": [ + 434, + 183, + 501, + 191 + ], + "spans": [ + { + "bbox": [ + 434, + 183, + 501, + 191 + ], + "type": "text", + "content": "CasualSAM Bird's Eye" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 220, + 202, + 313, + 293 + ], + "blocks": [ + { + "bbox": [ + 220, + 202, + 313, + 293 + ], + "lines": [ + { + "bbox": [ + 220, + 202, + 313, + 293 + ], + "spans": [ + { + "bbox": [ + 220, + 202, + 313, + 293 + ], + "type": "image", + "image_path": "66690b622dd088d3bd6cc54dd4775d3cce24ee0f93180903dcf91e3fc2cdc27a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 294, + 290, + 301 + ], + "lines": [ + { + "bbox": [ + 242, + 294, + 290, + 301 + ], + "spans": [ + { + "bbox": [ + 242, + 294, + 290, + 301 + ], + "type": "text", + "content": "Monst3R Static" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 326, + 200, + 419, + 290 + ], + "blocks": [ + { + "bbox": [ + 326, + 200, + 419, + 290 + ], + "lines": [ + { + "bbox": [ + 326, + 200, + 419, + 290 + ], + "spans": [ + { + "bbox": [ + 326, + 200, + 419, + 290 + ], + "type": "image", + "image_path": "e1feff5e9a91584c6778b64d7fdb3472be9a31f26d87c384e28245874c0b4a5d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 293, + 386, + 300 + ], + "lines": [ + { + "bbox": [ + 347, + 293, + 386, + 300 + ], + "spans": [ + { + "bbox": [ + 347, + 293, + 386, + 300 + ], + "type": "text", + "content": "Monst3R 4D" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 440, + 196, + 495, + 291 + ], + "blocks": [ + { + "bbox": [ + 440, + 196, + 495, + 291 + ], + "lines": [ + { + "bbox": [ + 440, + 196, + 495, + 291 + ], + "spans": [ + { + "bbox": [ + 440, + 196, + 495, + 291 + ], + "type": "image", + "image_path": "337696d7026100c9076a1e374816381bb8eea04e2651b08d51770c393f5aaf1d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 293, + 498, + 302 + ], + "lines": [ + { + "bbox": [ + 438, + 293, + 498, + 302 + ], + "spans": [ + { + "bbox": [ + 438, + 293, + 498, + 302 + ], + "type": "text", + "content": "Monst3R Bird's Eye" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 221, + 304, + 318, + 396 + ], + "blocks": [ + { + "bbox": [ + 221, + 304, + 318, + 396 + ], + "lines": [ + { + "bbox": [ + 221, + 304, + 318, + 396 + ], + "spans": [ + { + "bbox": [ + 221, + 304, + 318, + 396 + ], + "type": "image", + "image_path": "2bf43608fd6f44f172acd05fa6a475d4f6cb08fabd19cc424f999da22d6a0e3d.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 253, + 397, + 288, + 405 + ], + "lines": [ + { + "bbox": [ + 253, + 397, + 288, + 405 + ], + "spans": [ + { + "bbox": [ + 253, + 397, + 288, + 405 + ], + "type": "text", + "content": "Ours Static" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 325, + 303, + 422, + 392 + ], + "blocks": [ + { + "bbox": [ + 325, + 303, + 422, + 392 + ], + "lines": [ + { + "bbox": [ + 325, + 303, + 422, + 392 + ], + "spans": [ + { + "bbox": [ + 325, + 303, + 422, + 392 + ], + "type": "image", + "image_path": "73c9f62c338b5cbc2e05ca2f7a75ae1fb92310322955f64b4b109f3f8e341690.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 363, + 397, + 389, + 405 + ], + "lines": [ + { + "bbox": [ + 363, + 397, + 389, + 405 + ], + "spans": [ + { + "bbox": [ + 363, + 397, + 389, + 405 + ], + "type": "text", + "content": "Ours 4D" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 411, + 555, + 455 + ], + "lines": [ + { + "bbox": [ + 55, + 411, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 555, + 455 + ], + "type": "text", + "content": "Figure 4. Qualitative results of 4D reconstruction on the DAVIS [43] dataset. CasualSAM [68] suffers from slanted geometry, and Monst3R [65] has unclear geometry and does not resolve conflicts from multiple views (note the wall in the bird's-eye view). Both CasualSAM and Monst3R lack clean dynamic reconstruction and segmentation. Uni4D achieves a realistic layout, thanks to joint optimization, and provides accurate dynamic segmentation and reconstruction by leveraging foundation visual models as cues." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 446, + 304, + 492, + 389 + ], + "blocks": [ + { + "bbox": [ + 446, + 304, + 492, + 389 + ], + "lines": [ + { + "bbox": [ + 446, + 304, + 492, + 389 + ], + "spans": [ + { + "bbox": [ + 446, + 304, + 492, + 389 + ], + "type": "image", + "image_path": "cab83464e20a235e71f8960502452634221a646975c294da08ca627e07266bb1.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 397, + 498, + 406 + ], + "lines": [ + { + "bbox": [ + 451, + 397, + 498, + 406 + ], + "spans": [ + { + "bbox": [ + 451, + 397, + 498, + 406 + ], + "type": "text", + "content": "Ours Bird's Eye" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 468, + 295, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 295, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 295, + 504 + ], + "type": "text", + "content": "matrix " + }, + { + "bbox": [ + 55, + 468, + 295, + 504 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 55, + 468, + 295, + 504 + ], + "type": "text", + "content": " where we optimize focal lengths " + }, + { + "bbox": [ + 55, + 468, + 295, + 504 + ], + "type": "inline_equation", + "content": "f_{x}" + }, + { + "bbox": [ + 55, + 468, + 295, + 504 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 468, + 295, + 504 + ], + "type": "inline_equation", + "content": "f_{y}" + }, + { + "bbox": [ + 55, + 468, + 295, + 504 + ], + "type": "text", + "content": ". We formulate the 4D joint reasoning problem as minimization of the following energy function:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 56, + 514, + 295, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 295, + 537 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 295, + 537 + ], + "type": "interline_equation", + "content": "E _ {\\mathrm {B A}} (\\mathcal {C}, \\mathbf {P} _ {\\text {s t a t i c}}) + E _ {\\mathrm {N R}} (\\mathcal {P} _ {\\mathrm {d y n}}) + E _ {\\text {m o t i o n}} (\\mathcal {P} _ {\\mathrm {d y n}}) + E _ {\\text {c a m}} (\\mathcal {T}) \\tag {1}", + "image_path": "07734948be85449547d19f484a974a2a0d67fd9139e8e71d60552e221bf12af5.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 539, + 295, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 295, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 295, + 670 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 539, + 295, + 670 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{BA}}(\\mathcal{C}, \\mathbf{P}_{\\mathrm{static}})" + }, + { + "bbox": [ + 55, + 539, + 295, + 670 + ], + "type": "text", + "content": " is a bundle adjustment term that measures the discrepancy between static-region correspondences and the static 3D structure " + }, + { + "bbox": [ + 55, + 539, + 295, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{static}}" + }, + { + "bbox": [ + 55, + 539, + 295, + 670 + ], + "type": "text", + "content": " through perspective reprojection. " + }, + { + "bbox": [ + 55, + 539, + 295, + 670 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{NR}}(\\mathcal{P}_{\\mathrm{dyn}})" + }, + { + "bbox": [ + 55, + 539, + 295, + 670 + ], + "type": "text", + "content": " is a non-rigid structure-from-motion energy term that measures the disagreement between the dynamic point cloud and their tracklet correspondences, " + }, + { + "bbox": [ + 55, + 539, + 295, + 670 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{cam}}(\\mathcal{T})" + }, + { + "bbox": [ + 55, + 539, + 295, + 670 + ], + "type": "text", + "content": " is a regularization term on the camera motion smoothness, and " + }, + { + "bbox": [ + 55, + 539, + 295, + 670 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{motion}}(\\mathcal{P}_{\\mathrm{dyn}})" + }, + { + "bbox": [ + 55, + 539, + 295, + 670 + ], + "type": "text", + "content": " is a regularization term on the dynamic structure and motion. Each energy term is involved in different stages of the optimization process, which will be described in Sec. 3.3." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Static Bundle Adjustment Term. The bundle adjustment energy " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{BA}}(\\mathcal{C}, \\mathbf{P}_{\\mathrm{static}})" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " measures the consistency between" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 468, + 555, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 555, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 555, + 528 + ], + "type": "text", + "content": "the pixel-level correspondences and the 3D structure of static scene elements. Given the input pixel tracks " + }, + { + "bbox": [ + 313, + 468, + 555, + 528 + ], + "type": "inline_equation", + "content": "\\mathcal{Z} = \\{\\mathbf{Z}_k\\}" + }, + { + "bbox": [ + 313, + 468, + 555, + 528 + ], + "type": "text", + "content": " and video segmentation " + }, + { + "bbox": [ + 313, + 468, + 555, + 528 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 468, + 555, + 528 + ], + "type": "text", + "content": ", we filter all tracks corresponding to static areas and minimize the distance between the projected pixel location and the observed pixel location:" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 314, + 554, + 567, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 554, + 567, + 590 + ], + "spans": [ + { + "bbox": [ + 314, + 554, + 567, + 590 + ], + "type": "interline_equation", + "content": "E _ {\\mathrm {B A}} (\\mathcal {C}, \\mathbf {P} _ {\\text {s t a t i c}}; \\mathcal {Z}, \\mathcal {M}) = \\sum_ {\\mathbf {z} _ {k} \\in \\mathcal {M}} \\sum_ {t} \\mathbf {w} _ {\\mathbf {k}, \\mathbf {t}} \\| \\mathbf {z} _ {k, t} - \\pi_ {\\mathbf {K}} \\left(\\mathbf {p} _ {k}, \\boldsymbol {\\xi} _ {t}\\right) \\| _ {2} \\tag {2}", + "image_path": "ffa0361e4540f9519ca82d7b182c7bfb11a11fa49852d91d90a21a655d0ae1f0.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 590, + 555, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 555, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 555, + 626 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 590, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{k,t}" + }, + { + "bbox": [ + 313, + 590, + 555, + 626 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 590, + 555, + 626 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 590, + 555, + 626 + ], + "type": "text", + "content": "-th 3D point's corresponding pixel track's 2D coordinates at time " + }, + { + "bbox": [ + 313, + 590, + 555, + 626 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 590, + 555, + 626 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 590, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{w_{k,t}} \\in \\{0,1\\}" + }, + { + "bbox": [ + 313, + 590, + 555, + 626 + ], + "type": "text", + "content": " is a visibility indicator and " + }, + { + "bbox": [ + 313, + 590, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathbf{K}}" + }, + { + "bbox": [ + 313, + 590, + 555, + 626 + ], + "type": "text", + "content": " is the perspective projection function." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "content": "Non-Rigid Bundle Adjustment Term. For dynamic objects, we impose a nonrigid bundle adjustment term, " + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{NR}}(\\mathcal{P}_{\\mathrm{dyn}})" + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "content": ", which measures the discrepancy between the dynamic point cloud and pixel tracklets. Here, each pixel tracklet corresponds to a dynamic 3D point sequence," + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1119" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 74, + 289, + 182 + ], + "blocks": [ + { + "bbox": [ + 60, + 74, + 289, + 182 + ], + "lines": [ + { + "bbox": [ + 60, + 74, + 289, + 182 + ], + "spans": [ + { + "bbox": [ + 60, + 74, + 289, + 182 + ], + "type": "image", + "image_path": "f37dc1e04d2f370fb41c79d8847603ba08af17dfc7808c0fd6a4c38e891423e0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 191, + 295, + 226 + ], + "lines": [ + { + "bbox": [ + 55, + 191, + 295, + 226 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 295, + 226 + ], + "type": "text", + "content": "Figure 5. Qualitative results on Bonn dataset. Both CasualSAM and MonST3R have trailing artifacts and incorrect dynamic estimations. Uni4D provides clear dynamic and static geometry." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 243, + 241, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 241, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 241, + 255 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{p}_{k,t}\\}_{t}" + }, + { + "bbox": [ + 55, + 243, + 241, + 255 + ], + "type": "text", + "content": ", optimized for each observed tracklet:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 261, + 307, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 261, + 307, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 307, + 297 + ], + "type": "interline_equation", + "content": "E _ {\\mathrm {N R}} \\left(\\mathcal {P} _ {\\mathrm {d y n}}; \\mathcal {C}, \\mathcal {Z}, \\mathcal {M}\\right) = \\sum_ {\\mathbf {z} _ {k} \\in \\mathcal {M}} \\sum_ {t} \\mathbf {w} _ {\\mathbf {k}, \\mathbf {t}} \\left\\| \\mathbf {z} _ {k, t} - \\pi_ {\\mathbf {K}} \\left(\\mathbf {p} _ {k, t}, \\boldsymbol {\\xi} _ {t}\\right) \\right\\| _ {2} \\tag {3}", + "image_path": "e2115ce0c3c1f2239067265223eb6a7bcc31b6dd8327bfc85621bdd72f04c976.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 297, + 288, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 288, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 288, + 309 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 297, + 288, + 309 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{k,t} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 297, + 288, + 309 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 55, + 297, + 288, + 309 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 297, + 288, + 309 + ], + "type": "text", + "content": "-th dynamic point's location at " + }, + { + "bbox": [ + 55, + 297, + 288, + 309 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 297, + 288, + 309 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 323, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 418 + ], + "type": "text", + "content": "Camera Motion Prior. Considering that our input is a video, we incorporate a temporal smoothness prior on camera poses that penalizes sudden changes in relative pose: " + }, + { + "bbox": [ + 55, + 323, + 296, + 418 + ], + "type": "inline_equation", + "content": "\\pmb{\\xi}_{t\\rightarrow t + 1} = \\pmb{\\xi}_{t + 1}^{-1}\\cdot \\pmb{\\xi}_t" + }, + { + "bbox": [ + 55, + 323, + 296, + 418 + ], + "type": "text", + "content": ". We reweight this term based on the magnitude of the relative motion: intuitively, if the relative motion is large, we penalize change rates in relative motion less; if the relative motion is small, we apply a higher penalty on change rate. Formally, we have:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 424, + 294, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 294, + 458 + ], + "type": "interline_equation", + "content": "E _ {\\mathrm {c a m}} (\\mathcal {T}) = \\sum_ {t} E _ {\\mathrm {r o t}} \\left(\\mathbf {R} _ {t - 1, t, t + 1}\\right) + \\sum_ {t} E _ {\\mathrm {t r a n s}} \\left(\\mathbf {t} _ {t - 1, t, t + 1}\\right) \\tag {4}", + "image_path": "3c2338c10220a58cb839e81e7ac781f125d657975fa69794826aac6b0b6c460a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 456, + 294, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 294, + 499 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 456, + 294, + 499 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{rot}}(\\mathbf{R}_{t - 1,t,t + 1}) = \\frac{2\\|\\mathrm{rad}(\\mathbf{R}_{t\\to t + 1}) - \\mathrm{rad}(\\mathbf{R}_{t - 1\\to t})\\|}{\\|\\mathrm{rad}(\\mathbf{R}_{t - 1\\to t})\\| + \\|\\mathrm{rad}(\\mathbf{R}_{t\\to t + 1})\\|}" + }, + { + "bbox": [ + 55, + 456, + 294, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 456, + 294, + 499 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{trans}}(\\mathbf{t}_{t - 1,t,t + 1}) = \\frac{2\\|\\mathbf{t}_{t\\to t + 1} - \\mathbf{t}_{t - 1\\to t}\\|}{\\|\\mathbf{t}_{t - 1\\to t}\\| + \\|\\mathbf{t}_{t\\to t + 1}\\|}" + }, + { + "bbox": [ + 55, + 456, + 294, + 499 + ], + "type": "text", + "content": "; rad converts the rotation matrix into absolute radians." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "type": "text", + "content": "Dynamic Motion Prior. " + }, + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{motion}}(\\mathcal{P}_{\\mathrm{dyn}})" + }, + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "type": "text", + "content": " is a regularization term that encodes the characteristics of the dynamic structure. It contains two prior terms that are used to regularize the dynamic structure, both of which have demonstrated effectiveness in previous work [39, 50]:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 580, + 294, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 580, + 294, + 594 + ], + "spans": [ + { + "bbox": [ + 67, + 580, + 294, + 594 + ], + "type": "interline_equation", + "content": "E _ {\\text {m o t i o n}} \\left(\\mathcal {P} _ {\\text {d y n}}\\right) = E _ {\\text {a r a p}} \\left(\\mathcal {P} _ {\\text {d y n}}\\right) + E _ {\\text {s m o o t h}} \\left(\\mathcal {P} _ {\\text {d y n}}\\right). \\tag {5}", + "image_path": "bcf732cc9a5cf9c0f1d1255a4cd3a019de1709f908c5d60e91917c7e409cb13b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 598, + 296, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 296, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 296, + 670 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{arap}}" + }, + { + "bbox": [ + 55, + 598, + 296, + 670 + ], + "type": "text", + "content": " is an as-rigid-as-possible (ARAP) [50] prior that penalizes extreme deformations that compromise local rigidity. Specifically, we obtain the nearest neighbors of each dynamic control point " + }, + { + "bbox": [ + 55, + 598, + 296, + 670 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 598, + 296, + 670 + ], + "type": "text", + "content": " by applying KNN over the other tracks and enforce that the relative distances between these close-by pairs do not undergo sudden changes:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 676, + 310, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 676, + 310, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 676, + 310, + 713 + ], + "type": "interline_equation", + "content": "E _ {\\text {a r a p}} = \\sum_ {\\mathbf {t}} \\sum_ {(k, m)} \\mathbf {w} _ {\\mathbf {k m}, \\mathbf {t}} \\| d \\left(\\mathbf {p} _ {\\mathbf {k}, \\mathbf {t}}, \\mathbf {p} _ {m, t}\\right) - d \\left(\\mathbf {p} _ {\\mathbf {k}, \\mathbf {t} + 1}, \\mathbf {p} _ {m, t + 1}\\right) \\| _ {2} \\tag {6}", + "image_path": "8b7ec7f896d229aaf6d7525ccc1fa7064fb86b718cb79a9eec0063607bd7dff6.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "inline_equation", + "content": "d(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": " is the L2 distance and " + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{\\mathbf{k}\\mathbf{m},\\mathbf{t}} = 1" + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": " if all relevant points are visible." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 96, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 553, + 120 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{smooth}}" + }, + { + "bbox": [ + 313, + 96, + 553, + 120 + ], + "type": "text", + "content": " is a simple smoothness term that promotes temporal smoothness for the dynamic point cloud:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 338, + 125, + 553, + 151 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 125, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 338, + 125, + 553, + 151 + ], + "type": "interline_equation", + "content": "E _ {\\text {s m o o t h}} = \\sum_ {\\mathbf {t}} \\sum_ {\\mathbf {p} _ {k} \\in \\mathcal {P} _ {\\text {d y n}}} \\mathbf {w} _ {\\mathbf {k}, \\mathbf {t}} \\| \\mathbf {p} _ {k, t} - \\mathbf {p} _ {k, t + 1} \\| _ {2}. \\tag {7}", + "image_path": "6397053e124fe8f337a2b898a2a7e60a3a1f2309b0b6bae7da331e94e8803bcf.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 156, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 228 + ], + "type": "text", + "content": "Despite simplicity, both motion terms are crucial in our formulation, as they significantly reduce ambiguities in 4D dynamic structure estimation, which is highly ill-posed. Unlike other methods, we do not assume strong model-based motion priors, such as rigid motion [37], articulated motion [61], or a linear motion basis [57]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 233, + 381, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 233, + 381, + 243 + ], + "spans": [ + { + "bbox": [ + 314, + 233, + 381, + 243 + ], + "type": "text", + "content": "3.3. Inference" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 250, + 554, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 250, + 554, + 322 + ], + "spans": [ + { + "bbox": [ + 313, + 250, + 554, + 322 + ], + "type": "text", + "content": "Directly minimizing the energy defined in Eq. 1 is nontrivial, as our energy function is highly non-linear and involves millions of free variables. To address this, we developed a three-stage optimization pipeline, enabling us to minimize the energy and estimate the scene variables in a divide-and-conquer fashion." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 335, + 554, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 554, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 554, + 432 + ], + "type": "text", + "content": "Stage 1: Camera Initialization. We start by initializing camera parameters. Combining video depth estimation " + }, + { + "bbox": [ + 313, + 335, + 554, + 432 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 335, + 554, + 432 + ], + "type": "text", + "content": " and dense pixel motion " + }, + { + "bbox": [ + 313, + 335, + 554, + 432 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 313, + 335, + 554, + 432 + ], + "type": "text", + "content": " allows us to establish 2D-to-3D correspondences. This allows us to initialize and tune " + }, + { + "bbox": [ + 313, + 335, + 554, + 432 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 335, + 554, + 432 + ], + "type": "text", + "content": " by minimizing the following energy function with respect to camera parameters only. Specifically, we can unproject each video frame's depth at time " + }, + { + "bbox": [ + 313, + 335, + 554, + 432 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 335, + 554, + 432 + ], + "type": "text", + "content": " back to 3D and minimize the following energy function:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 318, + 436, + 553, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 436, + 553, + 463 + ], + "spans": [ + { + "bbox": [ + 318, + 436, + 553, + 463 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathbf {C}} \\sum_ {\\left(t ^ {\\prime}, t\\right)} \\sum_ {\\mathbf {z} _ {k} \\in \\neg \\mathcal {M}} \\| \\mathbf {z} _ {k, t ^ {\\prime}} - \\pi_ {\\mathbf {K}} \\left(\\pi_ {\\mathbf {K}} ^ {- 1} \\left(\\mathbf {z} _ {k, t}, \\mathbf {D} _ {t}, \\boldsymbol {\\xi} _ {t}\\right), \\boldsymbol {\\xi} _ {t ^ {\\prime}}\\right) \\| _ {2} ^ {2} \\tag {8}", + "image_path": "6b366aef273bf117dae8d8997d8f2f68e0bd742e7447cd675f765cf860b6720a.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 468, + 554, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 554, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 554, + 565 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 468, + 554, + 565 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathbf{K}}^{-1}" + }, + { + "bbox": [ + 313, + 468, + 554, + 565 + ], + "type": "text", + "content": " is the unprojection function that maps 2D coordinates into 3D world coordinates using estimated depth " + }, + { + "bbox": [ + 313, + 468, + 554, + 565 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_t" + }, + { + "bbox": [ + 313, + 468, + 554, + 565 + ], + "type": "text", + "content": ". We perform this over all pairs within a temporal sliding window of 5 frames, producing a good initial pose estimate as shown in Tab. 4. Given camera initialization " + }, + { + "bbox": [ + 313, + 468, + 554, + 565 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{C}}" + }, + { + "bbox": [ + 313, + 468, + 554, + 565 + ], + "type": "text", + "content": ", we directly unproject our depth prediction into a common world coordinate system, which provides an initial 4D structure " + }, + { + "bbox": [ + 313, + 468, + 554, + 565 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}" + }, + { + "bbox": [ + 313, + 468, + 554, + 565 + ], + "type": "text", + "content": ". This is used as initialization for later optimization." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 578, + 554, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 554, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 554, + 626 + ], + "type": "text", + "content": "Stage 2: Bundle Adjustment. Our second stage jointly optimizes camera pose and static geometry by minimizing the static component-related energy in a bundle adjustment fashion. Formally speaking, we solve the following:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 344, + 632, + 553, + 650 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 632, + 553, + 650 + ], + "spans": [ + { + "bbox": [ + 344, + 632, + 553, + 650 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathcal {C}, \\mathbf {P} _ {\\text {s t a t i c}}} E _ {\\mathrm {B A}} \\left(\\mathcal {C}, \\mathbf {P} _ {\\text {s t a t i c}}; \\mathcal {Z}, \\mathcal {M}\\right) + E _ {\\mathrm {c a m}} (\\mathcal {T}) \\tag {9}", + "image_path": "f140e8198632b63cae2ee5e4dd29781a0ff138d87ade9e70f0e15805c3622896.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "By enforcing consistency with each other, this improves both the static geometry and the camera pose quality. We perform a final scene integration by unprojecting correspondences into 3D using improved pose and filtering outlier noisy points in 3D." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1120" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 69, + 145, + 107 + ], + "blocks": [ + { + "bbox": [ + 80, + 69, + 145, + 107 + ], + "lines": [ + { + "bbox": [ + 80, + 69, + 145, + 107 + ], + "spans": [ + { + "bbox": [ + 80, + 69, + 145, + 107 + ], + "type": "image", + "image_path": "10cc49294d795eb17d072b528e4bcc702067db7f77b979e3dd24a1c820fed94c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 82, + 110, + 144, + 147 + ], + "blocks": [ + { + "bbox": [ + 82, + 110, + 144, + 147 + ], + "lines": [ + { + "bbox": [ + 82, + 110, + 144, + 147 + ], + "spans": [ + { + "bbox": [ + 82, + 110, + 144, + 147 + ], + "type": "image", + "image_path": "44c81bbada89dc995ea93df0de7375a4d7c7d14368c22f6541d55b50ddb8dc54.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 82, + 150, + 144, + 186 + ], + "blocks": [ + { + "bbox": [ + 82, + 150, + 144, + 186 + ], + "lines": [ + { + "bbox": [ + 82, + 150, + 144, + 186 + ], + "spans": [ + { + "bbox": [ + 82, + 150, + 144, + 186 + ], + "type": "image", + "image_path": "567f9c46a8bcb0eb644ada89f30f61d76beec164c931a9bd006d99697bf7ebef.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 81, + 189, + 144, + 225 + ], + "blocks": [ + { + "bbox": [ + 81, + 189, + 144, + 225 + ], + "lines": [ + { + "bbox": [ + 81, + 189, + 144, + 225 + ], + "spans": [ + { + "bbox": [ + 81, + 189, + 144, + 225 + ], + "type": "image", + "image_path": "cb8fbb9d5c7a86cd616d357bad64832d0e69574f24b8e9c05e046eeb10b7a77b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 80, + 228, + 143, + 265 + ], + "blocks": [ + { + "bbox": [ + 80, + 228, + 143, + 265 + ], + "lines": [ + { + "bbox": [ + 80, + 228, + 143, + 265 + ], + "spans": [ + { + "bbox": [ + 80, + 228, + 143, + 265 + ], + "type": "image", + "image_path": "d948dd0ff7d74a774ceea539861a9b0ac83859daeb725cc5ae424066c692d138.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 95, + 270, + 128, + 277 + ], + "lines": [ + { + "bbox": [ + 95, + 270, + 128, + 277 + ], + "spans": [ + { + "bbox": [ + 95, + 270, + 128, + 277 + ], + "type": "text", + "content": "Input Video" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 156, + 70, + 264, + 129 + ], + "blocks": [ + { + "bbox": [ + 156, + 70, + 264, + 129 + ], + "lines": [ + { + "bbox": [ + 156, + 70, + 264, + 129 + ], + "spans": [ + { + "bbox": [ + 156, + 70, + 264, + 129 + ], + "type": "image", + "image_path": "b5385671bef8f8af1d9d37ef6bde1ce7695c1f80551782313b3b10be04dda362.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 129, + 239, + 135 + ], + "lines": [ + { + "bbox": [ + 192, + 129, + 239, + 135 + ], + "spans": [ + { + "bbox": [ + 192, + 129, + 239, + 135 + ], + "type": "text", + "content": "CasualSAM Static" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 269, + 70, + 395, + 130 + ], + "blocks": [ + { + "bbox": [ + 269, + 70, + 395, + 130 + ], + "lines": [ + { + "bbox": [ + 269, + 70, + 395, + 130 + ], + "spans": [ + { + "bbox": [ + 269, + 70, + 395, + 130 + ], + "type": "image", + "image_path": "506995d07fba5dcfed6a0d409aaba68be698542faabf10452d3d09f83f4e9070.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 129, + 369, + 134 + ], + "lines": [ + { + "bbox": [ + 329, + 129, + 369, + 134 + ], + "spans": [ + { + "bbox": [ + 329, + 129, + 369, + 134 + ], + "type": "text", + "content": "CasualSAM 4D" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 418, + 63, + 512, + 126 + ], + "blocks": [ + { + "bbox": [ + 418, + 63, + 512, + 126 + ], + "lines": [ + { + "bbox": [ + 418, + 63, + 512, + 126 + ], + "spans": [ + { + "bbox": [ + 418, + 63, + 512, + 126 + ], + "type": "image", + "image_path": "591ab479303c6367c072c19b78dab2608794cd804efc771dc1f3c68e26d36dc6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 155, + 145, + 267, + 204 + ], + "blocks": [ + { + "bbox": [ + 155, + 145, + 267, + 204 + ], + "lines": [ + { + "bbox": [ + 155, + 145, + 267, + 204 + ], + "spans": [ + { + "bbox": [ + 155, + 145, + 267, + 204 + ], + "type": "image", + "image_path": "24c11c9c46de54934afd8a13c58dd489327ef5842ababa0f72f6b2aa4d17221e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 193, + 237, + 201 + ], + "lines": [ + { + "bbox": [ + 194, + 193, + 237, + 201 + ], + "spans": [ + { + "bbox": [ + 194, + 193, + 237, + 201 + ], + "type": "text", + "content": "MonST3R Static" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 277, + 141, + 403, + 202 + ], + "blocks": [ + { + "bbox": [ + 277, + 141, + 403, + 202 + ], + "lines": [ + { + "bbox": [ + 277, + 141, + 403, + 202 + ], + "spans": [ + { + "bbox": [ + 277, + 141, + 403, + 202 + ], + "type": "image", + "image_path": "49ed163bd00d2d62977fe1cc0c8f6b8fd00113a51b6fa8d62158e031c4b3a5ec.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 282, + 213, + 405, + 271 + ], + "blocks": [ + { + "bbox": [ + 282, + 213, + 405, + 271 + ], + "lines": [ + { + "bbox": [ + 282, + 213, + 405, + 271 + ], + "spans": [ + { + "bbox": [ + 282, + 213, + 405, + 271 + ], + "type": "image", + "image_path": "231d89d3a4532147a2c9258877d0cc9c67fcafc5d2062518ed83785bd86cea0c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 270, + 364, + 276 + ], + "lines": [ + { + "bbox": [ + 342, + 270, + 364, + 276 + ], + "spans": [ + { + "bbox": [ + 342, + 270, + 364, + 276 + ], + "type": "text", + "content": "Ours 4D" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 446, + 137, + 517, + 190 + ], + "blocks": [ + { + "bbox": [ + 456, + 129, + 514, + 135 + ], + "lines": [ + { + "bbox": [ + 456, + 129, + 514, + 135 + ], + "spans": [ + { + "bbox": [ + 456, + 129, + 514, + 135 + ], + "type": "text", + "content": "CasualSAM Bird's Eye" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 446, + 137, + 517, + 190 + ], + "lines": [ + { + "bbox": [ + 446, + 137, + 517, + 190 + ], + "spans": [ + { + "bbox": [ + 446, + 137, + 517, + 190 + ], + "type": "image", + "image_path": "6475bb994ced197f35942cf9a35d7080d4e585c855a1fd5647a7f206d4fde379.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 457, + 193, + 511, + 200 + ], + "lines": [ + { + "bbox": [ + 457, + 193, + 511, + 200 + ], + "spans": [ + { + "bbox": [ + 457, + 193, + 511, + 200 + ], + "type": "text", + "content": "MonST3R Bird's Eye" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 451, + 206, + 515, + 266 + ], + "blocks": [ + { + "bbox": [ + 451, + 206, + 515, + 266 + ], + "lines": [ + { + "bbox": [ + 451, + 206, + 515, + 266 + ], + "spans": [ + { + "bbox": [ + 451, + 206, + 515, + 266 + ], + "type": "image", + "image_path": "f4500869aa7c95b53fa95ede09d0b7c12c819f447f550b5d0f3b1efcfbb6eeac.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 269, + 509, + 276 + ], + "lines": [ + { + "bbox": [ + 468, + 269, + 509, + 276 + ], + "spans": [ + { + "bbox": [ + 468, + 269, + 509, + 276 + ], + "type": "text", + "content": "Ours Bird's Eye" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "text", + "content": "Stage 3: Non-Rigid Bundle Adjustment. Given the estimated camera pose, our third stage focuses on inferring dynamic structure. Note that we freeze camera parameters in this stage, as we find that incorrect geometry and motion evidence often harm camera pose estimation rather than improve it. Additionally, enabling camera pose optimization introduces extra flexibility in this ill-posed problem, harming robustness. Formally speaking, we solve the following:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 75, + 443, + 294, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 443, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 75, + 443, + 294, + 463 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathcal {P} _ {\\mathrm {d y n}}} E _ {\\mathrm {N R}} \\left(\\mathcal {P} _ {\\mathrm {d y n}}; \\mathbf {C}, \\mathcal {Z}, \\mathcal {M}\\right) + E _ {\\text {m o t i o n}} \\left(\\mathcal {P} _ {\\mathrm {d y n}}\\right) \\tag {10}", + "image_path": "9f14fa00717632d7d20e5ccf1d81a22ea35ef600184a466eb35b9ca2a48868ea.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 472, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 555 + ], + "type": "text", + "content": "We initialize " + }, + { + "bbox": [ + 55, + 472, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{dyn}}" + }, + { + "bbox": [ + 55, + 472, + 296, + 555 + ], + "type": "text", + "content": " using video depth and our optimized camera pose from Stage 2. We scale " + }, + { + "bbox": [ + 55, + 472, + 296, + 555 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{smooth}}" + }, + { + "bbox": [ + 55, + 472, + 296, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 472, + 296, + 555 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{arap}}" + }, + { + "bbox": [ + 55, + 472, + 296, + 555 + ], + "type": "text", + "content": " with constants 10 and 100 respectively which we found empirically led to better dynamics. This energy optimization might still leave some high-energy noisy points, often from incorrect cues, motion boundaries, or occlusions. We filter these outliers based on their energy values in a final step." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 570, + 296, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 641 + ], + "type": "text", + "content": "Fusion. From our energy minimization, we acquire a semi-dense dynamic and static point cloud along with camera parameters. To further densify the point cloud, enabling each pixel to correspond to a 3D point, we perform depth-based interpolation by computing a scale offset. Details can be found in the supplementary." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 642, + 296, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 701 + ], + "type": "text", + "content": "To unify our outputs into a consistent 4D representation, we use our camera parameters and static masks to project the aligned depth maps into world coordinates. To handle noisy depth values at boundaries, we create an edge mask filter by thresholding gradients of the depth maps." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "This process enables us to update the depth value for" + } + ] + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 315, + 338, + 555, + 491 + ], + "blocks": [ + { + "bbox": [ + 55, + 287, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 55, + 287, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 555, + 331 + ], + "type": "text", + "content": "Figure 6. Qualitative results of 4D reconstruction on DAVIS dataset. CasualSAM [68] distorts the room geometry as evident from the bird's eye view. Dynamic elements have inconsistent geometry over time. MonST3R [65] has noisy static geometry in the room's far corner and incomplete dynamic object geometry evident in the green highlight. Uni4D has the cleanest dynamic object reconstruction and segmentation results, with geometrically accurate room shapes as evident from the bird's eye view." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 338, + 555, + 491 + ], + "lines": [ + { + "bbox": [ + 315, + 338, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 315, + 338, + 555, + 491 + ], + "type": "image", + "image_path": "13f5c037775fe836d11053ed5c70993c5e8280c1391b9dd51fb5f24a02aadcb1.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 495, + 555, + 529 + ], + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 529 + ], + "type": "text", + "content": "Figure 7. Qualitative results of 4D reconstruction on Sintel dataset. We highlight and contrast performance on (1) dynamic objects and (2) geometry on planar surfaces." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 536, + 555, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 555, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 555, + 573 + ], + "type": "text", + "content": "each high-confidence pixel, which we can then reproject into 3D space to obtain the final dense dynamic geometry reconstruction result, as shown in Fig. 1 and Fig. 3." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 583, + 395, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 583, + 395, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 583, + 395, + 597 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 604, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 554, + 640 + ], + "type": "text", + "content": "Uni4D estimates camera pose, depth, and 3D motion from a single video. We perform experiments evaluating its performance with respect to baselines on these tasks." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 648, + 447, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 447, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 447, + 661 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": "For optimization, we use the Adam optimizer with ReduceLROnPlateau and EarlyStopping in PyTorch [26, 42]. We perform 600 iterations per sliding window in stage 1, 2000 iterations in stage 2, and 1000 iterations in stage 3." + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1121" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 95, + 70, + 512, + 203 + ], + "blocks": [ + { + "bbox": [ + 95, + 70, + 512, + 203 + ], + "lines": [ + { + "bbox": [ + 95, + 70, + 512, + 203 + ], + "spans": [ + { + "bbox": [ + 95, + 70, + 512, + 203 + ], + "type": "table", + "html": "
CategoryMethodSintelTUM-dynamicsBonn
ATE ↓RPE trans ↓RPE rot ↓ATE ↓RPE trans ↓RPE rot ↓ATE ↓RPE trans ↓RPE rot ↓
Pose onlyDPVO*[54]0.1710.0631.2910.0190.0140.4060.0220.0140.913
LEAP-VO*[8]0.0350.0651.6690.0250.0312.8430.0370.0140.844
Joint depth & poseRobust-CVD[30]0.3680.1533.4620.0960.0272.5900.0850.0180.803
CasualSAM[68]0.1370.0390.6300.0360.0180.7450.0240.0140.849
Monst3R[65]0.1080.0430.7290.1080.0221.3710.0230.0110.807
Uni4D0.1100.0320.3380.0120.0040.3350.0170.0100.818
Uni4D*0.0920.0330.1410.0120.0040.3310.0160.0100.817
", + "image_path": "3e2f1b73546d2317601a1657f82e06ddda6c4111716074ec0c9d7f954ca8d1d9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 106, + 239, + 502, + 407 + ], + "blocks": [ + { + "bbox": [ + 55, + 205, + 555, + 228 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 555, + 228 + ], + "type": "text", + "content": "Table 1. Camera Pose Evaluation on Sintel, TUM-dynamic, and Bonn datasets. We bold and underline the best and second best results respectively. * indicates known camera intrinsic." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 106, + 239, + 502, + 407 + ], + "lines": [ + { + "bbox": [ + 106, + 239, + 502, + 407 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 502, + 407 + ], + "type": "table", + "html": "
AlignmentCategoryMethodSintelBonnKITTI
Abs Rel ↓ δ<1.25 ↑Abs Rel ↓ δ<1.25 ↑Abs Rel ↓ δ<1.25 ↑
Per-sequence scale & shiftSingle-frame depthMetric3D[19]0.20571.90.04498.50.03998.8
Depth-pro[5]0.28060.50.04998.60.08094.2
Unidepth[44]0.19872.80.04098.50.03898.8
Video depthDepthCrafter[20]0.23169.00.06597.60.11288.4
Joint video depth & poseRobust-CVD[30]0.35849.70.10889.80.18272.9
CasualSAM[68]0.29256.90.06996.60.11388.3
Monst3r[65]0.35852.10.06095.00.08591.9
Uni4D0.21672.50.03898.30.09889.7
Per-sequence scaleJoint depth & poseMonst3r[65]0.34455.90.04198.20.08991.4
Joint depth & poseUni4D0.28964.90.03898.30.08693.3
", + "image_path": "c6af3f41ea4fbab950aa5afa8338e399727bc37706787c6af0f63060c8eb66ed.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 409, + 555, + 421 + ], + "lines": [ + { + "bbox": [ + 55, + 409, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 555, + 421 + ], + "type": "text", + "content": "Table 2. Video depth evaluation on Sintel, Bonn, and KITTI datasets. We bold and underline the best and second best results respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 426, + 297, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 297, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 297, + 547 + ], + "type": "text", + "content": "We initialize the learning rate at " + }, + { + "bbox": [ + 55, + 426, + 297, + 547 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-3}" + }, + { + "bbox": [ + 55, + 426, + 297, + 547 + ], + "type": "text", + "content": " for stage 1, and " + }, + { + "bbox": [ + 55, + 426, + 297, + 547 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-2}" + }, + { + "bbox": [ + 55, + 426, + 297, + 547 + ], + "type": "text", + "content": " for stages 2 and 3, reducing all to " + }, + { + "bbox": [ + 55, + 426, + 297, + 547 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 55, + 426, + 297, + 547 + ], + "type": "text", + "content": ". Our entire framework takes roughly 5 minutes for a 50-frame video on a RTX A6000 GPU. We include a detailed runtime breakdown in the supplementary. We run all baselines on the datasets using their official implementations and hyperparameters. Co-Trackers [22] are initialized at a dense " + }, + { + "bbox": [ + 55, + 426, + 297, + 547 + ], + "type": "inline_equation", + "content": "50 \\times 50" + }, + { + "bbox": [ + 55, + 426, + 297, + 547 + ], + "type": "text", + "content": " grid, with a " + }, + { + "bbox": [ + 55, + 426, + 297, + 547 + ], + "type": "inline_equation", + "content": "75 \\times 75" + }, + { + "bbox": [ + 55, + 426, + 297, + 547 + ], + "type": "text", + "content": " grid for Sintel to handle its large camera perspective change. All optimization hyperparameters are kept the same for all runs on all datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 580, + 153, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 153, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 153, + 591 + ], + "type": "text", + "content": "4.2. Pose Estimation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "Baselines. We compare with several recent methods for pose estimation in dynamic scenes. LEAPVO [8] and DPVO [54] are learning-based visual odometry methods. Robust-CVD [30] optimizes for pose and depth deformation through an SFM pipeline. CasualSAM [68] further improves on the idea by directly finetuning network weights along with a novel uncertainty formulation. Monst3R [65] is a very recent work that fine-tunes DUSt3R [58] for 4D reconstruction through PnP [33]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 427, + 555, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 555, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 555, + 547 + ], + "type": "text", + "content": "Benchmarks and metrics. We evaluate pose estimation on three dynamic datasets: Sintel [7], TUM-dynamics [51], and Bonn [40]. We follow LEAP-VO's evaluation split for Sintel and use all videos from TUM-dynamics and Bonn. Following MonST3R [65], we subsample every 3 frames from the first 270 frames from TUM-dynamics to save compute. We follow the standard pose evaluation process of aligning camera poses with Umeyama alignment [56]. We report Absolute Translation Error (ATE), Relative Translation and Rotation Error (RPE trans and RPE rot)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Results. As reported in Tab. 1, Uni4D achieves competitive results across all metrics and datasets, highlighting the generalizability and performance of our pipeline. Uni4D is flexible to the availability of camera intrinsics, showing further improvement with known camera intrinsics. Training-based approaches such as LEAP-VO achieve good results on synthetic dataset like Sintel but does not generalize well to real-world datasets. Our method matches performance with MonST3R [65] on Sintel and achieves significantly better results on real-world datasets even compared to methods using ground-truth intrinsics." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1122" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 190, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 190, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 190, + 85 + ], + "type": "text", + "content": "4.3. Video Depth Evaluation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 187 + ], + "type": "text", + "content": "Baselines. For video depth accuracy evaluations, we focus on the top performing metric depth estimators, namely Metric3Dv2 [19], Depth-Pro [63], DepthCrafter [20] and Unidepth [44]. Metric depth estimators are trained without scale and shift alignment, making them strong baselines for video depth estimation. We also include the same baselines as our pose evaluations for joint 4D modeling approaches, namely CasualSAM [68], RCVD [30], and MonST3R [65]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 205, + 296, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 296, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 296, + 301 + ], + "type": "text", + "content": "**Benchmarks.** We evaluate video depth estimates on Sintel [7], Bonn [40] and KITTI [16]. We follow standard video depth evaluation protocols [20] of aligning global shift and scale to predicted video depthmaps. We report the absolute relative error (Abs Rel) and percent of inlier points " + }, + { + "bbox": [ + 55, + 205, + 296, + 301 + ], + "type": "inline_equation", + "content": "(\\delta < 1.25)" + }, + { + "bbox": [ + 55, + 205, + 296, + 301 + ], + "type": "text", + "content": ". All methods are aligned in disparity space using the same least-squares alignment. We additionally report scale-aligned depth estimates similar to MonST3R." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 321, + 296, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 321, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 321, + 296, + 453 + ], + "type": "text", + "content": "Results. By leveraging the strong performance of single-frame metric depth estimation models, Tab. 2 shows Uni4D achieves competitive depth estimation results, with superior performance among joint depth and pose methods, and closely matching the performance of single-frame depth estimation models on some datasets. With per-sequence scale, our model produces more accurate depth estimates across all datasets compared to recently released MonST3R. Our method closely retains depth estimation accuracy of underlying Unidepth depthmaps, while significantly improving its depth consistency shown in Sec. 4.5." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 463, + 130, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 130, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 130, + 475 + ], + "type": "text", + "content": "4.4. Qualitative" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 481, + 296, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 296, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 296, + 625 + ], + "type": "text", + "content": "We further show qualitative results of our reconstructions on the DAVIS [43] dataset. We apply MonST3R's own confidence-guided fusion for their reconstruction. We fuse casualSAM's depthmaps with dynamic masking by thresholding its uncertainty prediction to output 4D reconstruction. We use highlights to indicate dynamic objects at different timesteps. Throughout our qualitative results in Figs. 4,5,6,7, CasualSAM produces warped geometry and poor dynamic segmentations. MonST3R produces poor geometry in far regions and noisy dynamic masks and shapes. Our model produces the cleanest dynamic segmentations, along with the best dynamic and static geometry." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 635, + 149, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 149, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 149, + 647 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "We ablate our multi-stage optimization in Tab. 4, highlighting the importance of each stage. Joint 4D optimization adds complexity that requires strong initialization for optimal convergence. Stage 1 introduces drift, which Stage 2 rectifies, resulting in superior final pose estimates." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 70, + 359, + 95 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 359, + 95 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 359, + 95 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 359, + 95 + ], + "type": "image", + "image_path": "c3abe6cd08cb56ffe4c3f2a6571664e3a5d964e710d9fa4eb0e48ed1b76d773b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 99, + 358, + 123 + ], + "blocks": [ + { + "bbox": [ + 316, + 99, + 358, + 123 + ], + "lines": [ + { + "bbox": [ + 316, + 99, + 358, + 123 + ], + "spans": [ + { + "bbox": [ + 316, + 99, + 358, + 123 + ], + "type": "image", + "image_path": "ec1b32e6e75c1904c19aa48ce29d046fa377c3c451f069abc03fdf843987f964.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 125, + 358, + 152 + ], + "blocks": [ + { + "bbox": [ + 316, + 125, + 358, + 152 + ], + "lines": [ + { + "bbox": [ + 316, + 125, + 358, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 125, + 358, + 152 + ], + "type": "image", + "image_path": "46c81b931261bd7fcf3ea4d0fa4374cc681d9a854a7d0d1fbbc2dd1f4010df9c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 154, + 358, + 178 + ], + "blocks": [ + { + "bbox": [ + 316, + 154, + 358, + 178 + ], + "lines": [ + { + "bbox": [ + 316, + 154, + 358, + 178 + ], + "spans": [ + { + "bbox": [ + 316, + 154, + 358, + 178 + ], + "type": "image", + "image_path": "8146d9bcc40619c7660947afdfa42a6eb53c54925bbb21c16779482cd3af1e4c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 316, + 182, + 358, + 207 + ], + "blocks": [ + { + "bbox": [ + 316, + 182, + 358, + 207 + ], + "lines": [ + { + "bbox": [ + 316, + 182, + 358, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 182, + 358, + 207 + ], + "type": "image", + "image_path": "f015f7ddf543d49d2921b7e0a7aa0397f1a6a6f97aa42ec529e4344d55b8188b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 369, + 71, + 455, + 131 + ], + "blocks": [ + { + "bbox": [ + 369, + 71, + 455, + 131 + ], + "lines": [ + { + "bbox": [ + 369, + 71, + 455, + 131 + ], + "spans": [ + { + "bbox": [ + 369, + 71, + 455, + 131 + ], + "type": "image", + "image_path": "33130ec47c4ae58385798c67a722d5cc9fb72cd91c723f5c676646fa527e2ddf.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 138, + 438, + 148 + ], + "lines": [ + { + "bbox": [ + 419, + 138, + 438, + 148 + ], + "spans": [ + { + "bbox": [ + 419, + 138, + 438, + 148 + ], + "type": "text", + "content": "Undepth Bird's Eye" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 403, + 139, + 413, + 148 + ], + "blocks": [ + { + "bbox": [ + 403, + 139, + 413, + 148 + ], + "lines": [ + { + "bbox": [ + 403, + 139, + 413, + 148 + ], + "spans": [ + { + "bbox": [ + 403, + 139, + 413, + 148 + ], + "type": "image", + "image_path": "c2d07328020d760daa3a3b630632d92ccabd6c77cbd36217bd5e49d205766de4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 368, + 156, + 448, + 207 + ], + "blocks": [ + { + "bbox": [ + 368, + 156, + 448, + 207 + ], + "lines": [ + { + "bbox": [ + 368, + 156, + 448, + 207 + ], + "spans": [ + { + "bbox": [ + 368, + 156, + 448, + 207 + ], + "type": "image", + "image_path": "c70d37987e4634952b6f696822399651c262dd66e14990eded2cf6621b247df7.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 211, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 313, + 211, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 555, + 266 + ], + "type": "text", + "content": "Figure 8. Depth Consistency. Direct fusion of Unidepth [44] predictions causes misaligned scene geometry visible as a layered wall in bird's-eye view, and jittering dynamic motion. Through motion priors and alignment, Uni4D produces a thin, crisp wall structure and smooth dynamic motion." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 465, + 69, + 549, + 130 + ], + "blocks": [ + { + "bbox": [ + 465, + 69, + 549, + 130 + ], + "lines": [ + { + "bbox": [ + 465, + 69, + 549, + 130 + ], + "spans": [ + { + "bbox": [ + 465, + 69, + 549, + 130 + ], + "type": "image", + "image_path": "89d421e2141e8b84a518a2e637656252a945fadb67be716a0fd90169ca401eed.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 514, + 138, + 533, + 148 + ], + "lines": [ + { + "bbox": [ + 514, + 138, + 533, + 148 + ], + "spans": [ + { + "bbox": [ + 514, + 138, + 533, + 148 + ], + "type": "text", + "content": "Ours Bird's Eye" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 496, + 138, + 507, + 148 + ], + "blocks": [ + { + "bbox": [ + 496, + 138, + 507, + 148 + ], + "lines": [ + { + "bbox": [ + 496, + 138, + 507, + 148 + ], + "spans": [ + { + "bbox": [ + 496, + 138, + 507, + 148 + ], + "type": "image", + "image_path": "1712028f1251fdd8f1300cc7b4ecbf2932d6fe6f20664c71f2da842ace651e8c.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 459, + 156, + 540, + 207 + ], + "blocks": [ + { + "bbox": [ + 459, + 156, + 540, + 207 + ], + "lines": [ + { + "bbox": [ + 459, + 156, + 540, + 207 + ], + "spans": [ + { + "bbox": [ + 459, + 156, + 540, + 207 + ], + "type": "image", + "image_path": "ebbae18983fa690b97b546328258ddda6a722c4caf4741570065b935c2d0b786.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "table", + "bbox": [ + 336, + 270, + 531, + 324 + ], + "blocks": [ + { + "bbox": [ + 336, + 270, + 531, + 324 + ], + "lines": [ + { + "bbox": [ + 336, + 270, + 531, + 324 + ], + "spans": [ + { + "bbox": [ + 336, + 270, + 531, + 324 + ], + "type": "table", + "html": "
MethodSintel
SC ↓δSC<0.01 ↑δSC<0.05 ↑
Unidepth [44]0.10931.876.8
Uni4D0.04369.388.1
", + "image_path": "b150306567e59caf85819334c052f58871724698268ae16c09049dfeceec8f8a.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_body" + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 334, + 351, + 533, + 416 + ], + "blocks": [ + { + "bbox": [ + 314, + 326, + 554, + 348 + ], + "lines": [ + { + "bbox": [ + 314, + 326, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 314, + 326, + 554, + 348 + ], + "type": "text", + "content": "Table 3. Video depth consistency on Sintel. Uni4D improves Unidepth in consistency." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 334, + 351, + 533, + 416 + ], + "lines": [ + { + "bbox": [ + 334, + 351, + 533, + 416 + ], + "spans": [ + { + "bbox": [ + 334, + 351, + 533, + 416 + ], + "type": "table", + "html": "
MethodSintel
ATE ↓RPE trans ↓RPE rot ↓
Uni4D (stage 1 only)0.1500.0510.551
Uni4D (stage 2 only)0.5870.1934.12
Uni4D (full)0.1100.0320.338
", + "image_path": "2302e39e9bdb8c9f87c599fbf640158902d229d56f553e3071c9d313f060ac16.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 417, + 554, + 440 + ], + "lines": [ + { + "bbox": [ + 313, + 417, + 554, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 554, + 440 + ], + "type": "text", + "content": "Table 4. Multi-stage Ablation. We evaluate pose estimation results on Sintel for both stage 1 and stage 2." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 453, + 555, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 555, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 555, + 585 + ], + "type": "text", + "content": "Directly reprojecting Unidepth depth maps leads to flickering geometry. We ablate our fused depthmap quantitatively in Tab. 3, showing that Uni4D significantly improves consistency using Self-Consistency (SC) metrics [24]. SC measures depth errors between estimated and reprojected depth maps in static regions. Furthermore, we qualitatively show the effectiveness of Uni4D in rectifying Unidepth inconsistencies in Fig. 8. Our reconstruction achieves much better geometric and temporal consistency over Unidepth. Please refer to the supplementary for more ablative results regarding choice of foundation models." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 597, + 388, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 388, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 388, + 609 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "This paper presents Uni4D, a framework unifying visual foundation models and structured energy minimization for dynamic 4D modeling from casual video. Our key insight is to optimize a 4D representation that aligns with visual cues from foundation models while following motion and geometry priors. Results show state-of-the-art performance with superior visual quality on Sintel, DAVIS, TUM-Dynamics and Bonn datasets, without any retraining or fine-tuning." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1123" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 166, + 115, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 166, + 115, + 178 + ], + "spans": [ + { + "bbox": [ + 56, + 166, + 115, + 178 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 186, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 186, + 296, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 186, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 61, + 186, + 296, + 240 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 241, + 296, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 241, + 296, + 284 + ], + "spans": [ + { + "bbox": [ + 61, + 241, + 296, + 284 + ], + "type": "text", + "content": "[2] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building rome in a day. Communications of the ACM, 54 (10):105-112, 2011. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 285, + 296, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 296, + 339 + ], + "type": "text", + "content": "[3] Antonio Agudo, Melcior Pijoan, and Francesc Moreno-Noguer. Image collection pop-up: 3d reconstruction and clustering of rigid and non-rigid categories. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2607-2615, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 340, + 296, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 340, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 340, + 296, + 384 + ], + "type": "text", + "content": "[4] Gwangbin Bae and Andrew J Davison. Rethinking inductive biases for surface normal estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9535-9545, 2024. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 384, + 296, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 384, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 62, + 384, + 296, + 437 + ], + "type": "text", + "content": "[5] Aleksei Bochkovskii, Amael Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 439, + 296, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 439, + 296, + 493 + ], + "spans": [ + { + "bbox": [ + 62, + 439, + 296, + 493 + ], + "type": "text", + "content": "[6] Christoph Bregler, Aaron Hertzmann, and Henning Biermann. Recovering non-rigid 3d shape from image streams. In Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662), pages 690-696. IEEE, 2000. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 494, + 296, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 494, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 62, + 494, + 296, + 559 + ], + "type": "text", + "content": "[7] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part VI 12, pages 611-625. Springer, 2012. 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 560, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 560, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 62, + 560, + 296, + 613 + ], + "type": "text", + "content": "[8] Weirong Chen, Le Chen, Rui Wang, and Marc Pollefeys. Leap-vo: Long-term effective any point tracking for visual odometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19844–19853, 2024. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 615, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 615, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 62, + 615, + 296, + 669 + ], + "type": "text", + "content": "[9] Ho Kei Cheng, Seoung Wug Oh, Brian Price, Alexander Schwing, and Joon-Young Lee. Tracking anything with decoupled video segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1316–1326, 2023. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[10] Hainan Cui, Xiang Gao, Shuhan Shen, and Zhanyi Hu. Hsfm: Hybrid structure-from-motion. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1212-1221, 2017. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 72, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 55, + 90, + 296, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 147 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 147 + ], + "type": "text", + "content": "This project is supported by the Intel AI SRS gift, Amazon-Illinois AICE grant, Meta Research Grant, IBM IIDAI Grant, and NSF Awards #2331878, #2340254, #2312102, #2414227, and #2404385. We greatly appreciate the NCSA for providing computing resources." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[11] Andrew J Davison, Ian D Reid, Nicholas D Molton, and Olivier Stasse. Monoslam: Real-time single camera slam. IEEE transactions on pattern analysis and machine intelligence, 29(6):1052-1067, 2007. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 554, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 182 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 182 + ], + "type": "text", + "content": "[12] Matt Deitke, Christopher Clark, Sangho Lee, Rohun Tripathi, Yue Yang, Jae Sung Park, Mohammadreza Salehi, Niklas Muennighoff, Kyle Lo, Luca Soldaini, et al. Molmo and pixmo: Open weights and open data for state-of-the-art multimodal models. arXiv preprint arXiv:2409.17146, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 183, + 553, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 183, + 553, + 225 + ], + "spans": [ + { + "bbox": [ + 317, + 183, + 553, + 225 + ], + "type": "text", + "content": "[13] Jakob Engel, Thomas Schöps, and Daniel Cremers. Lsdslam: Large-scale direct monocular slam. In European conference on computer vision, pages 834-849. Springer, 2014. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 228, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 282 + ], + "type": "text", + "content": "[14] Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. In European Conference on Computer Vision, pages 241-258. Springer, 2025. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "type": "text", + "content": "[15] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 327, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 369 + ], + "type": "text", + "content": "[16] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 372, + 553, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 423 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 423 + ], + "type": "text", + "content": "[17] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa*, and Jitendra Malik*. Humans in 4D: Reconstructing and tracking humans with transformers. In International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 426, + 554, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 426, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 317, + 426, + 554, + 480 + ], + "type": "text", + "content": "[18] Zan Gojcic, Or Litany, Andreas Wieser, Leonidas J Guibas, and Tolga Birdal. Weakly supervised learning of rigid 3d scene flow. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5692-5703, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 482, + 554, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 554, + 547 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 554, + 547 + ], + "type": "text", + "content": "[19] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. arXiv preprint arXiv:2404.15506, 2024.7,8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 548, + 554, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 554, + 592 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 554, + 592 + ], + "type": "text", + "content": "[20] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024. 2, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 593, + 554, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 554, + 635 + ], + "type": "text", + "content": "[21] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv preprint arXiv:2307.07635, 2023. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 636, + 554, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 554, + 690 + ], + "type": "text", + "content": "[22] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker3: Simpler and better point tracking by pseudolabelling real videos. arXiv preprint arXiv:2410.11831, 2024. 2, 3, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[23] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurpos" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1124" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "text", + "content": "ing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 172 + ], + "type": "text", + "content": "[24] Numair Khan, Eric Penner, Douglas Lanman, and Lei Xiao. Temporally consistent online depth estimation using point-based fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9119-9129, 2023. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 295, + 229 + ], + "type": "text", + "content": "[25] Rawal Khirodkar, Timur Bagautdinov, Julieta Martinez, Su Zhaoen, Austin James, Peter Selednik, Stuart Anderson, and Shunsuke Saito. Sapiens: Foundation for human vision models. In European Conference on Computer Vision, pages 206-228. Springer, 2025. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 295, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 295, + 252 + ], + "type": "text", + "content": "[26] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 295, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 295, + 317 + ], + "type": "text", + "content": "[27] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 295, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 295, + 363 + ], + "type": "text", + "content": "[28] Chen Kong and Simon Lucey. Prior-less compressible structure from motion. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4123-4131, 2016. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 295, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 295, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 295, + 408 + ], + "type": "text", + "content": "[29] Chen Kong, Rui Zhu, Hamed Kiani, and Simon Lucey. Structure from category: A generic and prior-less approach. In 2016 Fourth International Conference on 3D Vision (3DV), pages 296-304. IEEE, 2016. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 410, + 296, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 296, + 453 + ], + "type": "text", + "content": "[30] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. 2, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 455, + 295, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 295, + 499 + ], + "type": "text", + "content": "[31] Guillaume Le Moing, Jean Ponce, and Cordelia Schmid. Dense optical tracking: connecting the dots. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19187-19197, 2024. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 500, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 295, + 544 + ], + "type": "text", + "content": "[32] Jiahui Lei, Yijia Weng, Adam Harley, Leonidas Guibas, and Kostas Daniilidis. Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds. arXiv preprint arXiv:2405.17421, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 295, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 295, + 588 + ], + "type": "text", + "content": "[33] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Ep n p: An accurate o (n) solution to the p np problem. International journal of computer vision, 81:155-166, 2009. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 295, + 623 + ], + "type": "text", + "content": "[34] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 295, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 677 + ], + "type": "text", + "content": "[35] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4273-4284, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 680, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 295, + 714 + ], + "type": "text", + "content": "[36] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[37] Wei-Chiu Ma, Shenlong Wang, Rui Hu, Yuwen Xiong, and Raquel Urtasun. Deep rigid instance scene flow. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3614-3622, 2019. 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "text", + "content": "[38] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 186, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 240 + ], + "type": "text", + "content": "[39] Richard A Newcombe, Dieter Fox, and Steven M Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 343-352, 2015. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 241, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 308 + ], + "type": "text", + "content": "[40] Emanuele Palazzolo, Jens Behley, Philipp Lottes, Philippe Giguere, and Cyril Stachniss. Refusion: 3d reconstruction in dynamic environments for rgb-d cameras exploiting residuals. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7855-7862. IEEE, 2019. 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "text", + "content": "[41] Linfei Pan, Daniel Barath, Marc Pollefeys, and Johannes L Schonberger. Global structure-from-motion revisited. In European Conference on Computer Vision (ECCV), 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 343, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 408 + ], + "type": "text", + "content": "[42] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 410, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 553, + 475 + ], + "type": "text", + "content": "[43] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 4, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 477, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 477, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 477, + 553, + 533 + ], + "type": "text", + "content": "[44] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Undepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10106-10116, 2024. 1, 2, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 534, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 553, + 578 + ], + "type": "text", + "content": "[45] Luigi Piccinelli, Christos Sakaridis, Yung-Hsu Yang, Mattia Segu, Siyuan Li, Wim Abbeloos, and Luc Van Gool. Undepthv2: Universal monocular metric depth estimation made simpler. arXiv preprint arXiv:2502.20110, 2025. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 579, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 633 + ], + "type": "text", + "content": "[46] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "type": "text", + "content": "[47] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "type": "text", + "content": "[48] Noah Snavely, Steven M Seitz, and Richard Szeliski. Modeling the world from internet photo collections. International journal of computer vision, 80:189-210, 2008. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1125" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[49] Chonghyuk Song, Gengshan Yang, Kangle Deng, Jun-Yan Zhu, and Deva Ramanan. Total-recon: Deformable scene reconstruction for embodied view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17671-17682, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 161 + ], + "type": "text", + "content": "[50] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Symposium on Geometry processing, pages 109–116. CiteSeer, 2007. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "text", + "content": "[51] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of rgb-d slam systems. In 2012 IEEE/RSJ international conference on intelligent robots and systems, pages 573-580. IEEE, 2012. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 274 + ], + "type": "text", + "content": "[52] Yu Sun, Qian Bao, Wu Liu, Tao Mei, and Michael J Black. Trace: 5d temporal regression of avatars with dynamic cameras in 3d environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8856-8866, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 318 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 318 + ], + "type": "text", + "content": "[53] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 34:16558-16569, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 353 + ], + "type": "text", + "content": "[54] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. Advances in Neural Information Processing Systems, 36, 2024. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 420 + ], + "type": "text", + "content": "[55] Bill Triggs, Philip F McLauchlan, Richard I Hartley, and Andrew W Fitzgibbon. Bundle adjustment—a modern synthesis. In Vision Algorithms: Theory and Practice: International Workshop on Vision Algorithms Corfu, Greece, September 21–22, 1999 Proceedings, pages 298–372. Springer, 2000. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "type": "text", + "content": "[56] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on Pattern Analysis & Machine Intelligence, 13(04): 376-380, 1991. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "text", + "content": "[57] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 566 + ], + "type": "text", + "content": "[58] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 1, 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 632 + ], + "type": "text", + "content": "[59] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "type": "text", + "content": "[60] Yuefan Wu, Zeyuan Chen, Shaowei Liu, Zhongzheng Ren, and Shenlong Wang. Casa: Category-agnostic skeletal animal reconstruction. Advances in Neural Information Processing Systems, 35:28559-28574, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "text", + "content": "[61] Gengshan Yang, Minh Vo, Natalia Neverova, Deva Ramanan, Andrea Vedaldi, and Hanbyul Joo. Banmo: Building animatable 3d neural models from many casual videos. In" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 596 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2863-2873, 2022. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[62] Gengshan Yang, Shuo Yang, John Z Zhang, Zachary Manchester, and Deva Ramanan. Ppr: Physically plausible reconstruction from monocular videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3914-3924, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 555, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 206 + ], + "type": "text", + "content": "[63] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 1, 2, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[64] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 253, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 306 + ], + "type": "text", + "content": "[65] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825, 2024. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 308, + 553, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 361 + ], + "type": "text", + "content": "[66] John Z Zhang, Shuo Yang, Gengshan Yang, Arun L Bishop, Swaminathan Gurumurthy, Deva Ramanan, and Zachary Manchester. Slomo: A general system for legged robot motion imitation from casual videos. IEEE Robotics and Automation Letters, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 364, + 553, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 364, + 553, + 429 + ], + "spans": [ + { + "bbox": [ + 316, + 364, + 553, + 429 + ], + "type": "text", + "content": "[67] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1724-1732, 2024. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 430, + 553, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 430, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 316, + 430, + 553, + 485 + ], + "type": "text", + "content": "[68] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T Freeman. Structure and motion from casual videos. In European Conference on Computer Vision, pages 20-37. Springer, 2022. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 487, + 553, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 553, + 540 + ], + "type": "text", + "content": "[69] Wang Zhao, Shaohui Liu, Hengkai Guo, Wenping Wang, and Yong-Jin Liu. *Particlesfm: Exploiting dense point trajectories for localizing moving cameras in the wild*. In European Conference on Computer Vision, pages 523-542. Springer, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 542, + 553, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 542, + 553, + 596 + ], + "spans": [ + { + "bbox": [ + 316, + 542, + 553, + 596 + ], + "type": "text", + "content": "[70] Yingying Zhu, Dong Huang, Fernando De La Torre, and Simon Lucey. Complex non-rigid motion 3d reconstruction by union of subspaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1542-1549, 2014. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1126" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/UniAP_ Unifying Inter- and Intra-Layer Automatic Parallelism by Mixed Integer Quadratic Programming/b1ddd4d6-5b56-44ba-82ed-6323538fa782_content_list.json b/2025/UniAP_ Unifying Inter- and Intra-Layer Automatic Parallelism by Mixed Integer Quadratic Programming/b1ddd4d6-5b56-44ba-82ed-6323538fa782_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3856f121856fe6a97ae49ac881d4d1cf9fae1adb --- /dev/null +++ b/2025/UniAP_ Unifying Inter- and Intra-Layer Automatic Parallelism by Mixed Integer Quadratic Programming/b1ddd4d6-5b56-44ba-82ed-6323538fa782_content_list.json @@ -0,0 +1,2035 @@ +[ + { + "type": "text", + "text": "UniAP: Unifying Inter- and Intra-Layer Automatic Parallelism by Mixed Integer Quadratic Programming", + "text_level": 1, + "bbox": [ + 91, + 128, + 906, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao Lin\\*, Ke Wu\\*, Jie Li\\*, Jun Li, Wu-Jun Li† \nNational Key Laboratory for Novel Software Technology School of Computer Science \nNanjing University, Nanjing 210023, China", + "bbox": [ + 271, + 202, + 723, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{hao.lin, ke.wu, jie-li, lijun}@smail.nju.edu.cn, liwujun@nju.edu.cn", + "bbox": [ + 109, + 275, + 883, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Distributed learning is commonly used for training deep learning models, especially large models. In distributed learning, manual parallelism (MP) methods demand considerable human effort and have limited flexibility. Hence, automatic parallelism (AP) methods have recently been proposed for automating the parallel strategy optimization process. Existing AP methods suffer from sub-optimal solutions because they do not jointly optimize the two categories of parallel strategies (i.e., inter-layer parallelism and intra-layer parallelism). In this paper, we propose a novel AP method called UniAP, which unifies inter- and intra-layer automatic parallelism by mixed integer quadratic programming. To the best of our knowledge, UniAP is the first parallel method that can jointly optimize the two categories of parallel strategies to find an optimal solution. Experimental results show that UniAP outperforms state-of-the-art methods by up to $3.80 \\times$ in throughput and reduces strategy optimization time by up to $107 \\times$ across five Transformer-based models.", + "bbox": [ + 88, + 358, + 485, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 660, + 223, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Distributed learning (also called parallel learning) on clusters with several machines or GPUs is commonly used for training deep learning models, especially for some large models with billions of parameters [1, 40, 41]. Several parallel strategies, including pipeline parallelism (PP), data parallelism (DP), tensor parallelism (TP), and fully sharded data parallelism (FSDP), have been proposed for distributed learning. These parallel strategies can be divided into two main categories: inter-layer parallelism and intra-layer parallelism. Inter-layer parallelism [7, 9, 10, 15, 20, 21, 28, 29], which includes PP, partitions the model into disjoint sets", + "bbox": [ + 89, + 685, + 485, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "without partitioning tensors in each layer. Intra-layer parallelism [8, 22, 30, 34], which includes DP, TP, and FSDP, partitions tensors in a layer along one or more axes.", + "bbox": [ + 511, + 328, + 906, + 373 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The parallel method1 in one specific distributed learning method or system typically adopts one parallel strategy or a combination of several parallel strategies. Existing parallel methods can be divided into two categories: manual parallelism (MP) methods and automatic parallelism (AP) methods. In MP methods [30, 37, 44], one or several parallel strategies are manually optimized by researchers or developers. MP methods require extensive domain knowledge in deep learning models and hardware architectures. With the rapid development of deep learning models and the increasing diversity of modern hardware architectures [11, 12], MP methods demand considerable human effort and have limited flexibility.", + "bbox": [ + 511, + 375, + 908, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the two limitations of MP methods, AP methods [14, 28, 47] have recently been proposed for automating the parallel strategy optimization process. Although existing AP methods have achieved promising progress, they optimize the two categories of parallel strategies separately rather than jointly. More specifically, some methods optimize only one category of parallel strategies [2, 16, 17, 23, 36, 42, 46], and the others optimize inter- and intra-layer parallelism hierarchically [9, 14, 28, 38, 39, 47]. Hence, existing AP methods suffer from sub-optimal solutions.", + "bbox": [ + 511, + 571, + 910, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose a novel AP method called UniAP for distributed learning. The contributions of UniAP are outlined as follows:", + "bbox": [ + 511, + 724, + 906, + 768 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- UniAP unifies inter- and intra-layer automatic parallelism by mixed integer quadratic programming (MIQP) [19].", + "- To the best of our knowledge, UniAP is the first parallel method that can optimize the two categories of parallel strategies jointly rather than separately to find an optimal solution." + ], + "bbox": [ + 511, + 771, + 903, + 859 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 863, + 220, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 875, + 235, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "To avoid confusion, we treat 'parallel method' and 'parallel strategy' as two different terminologies in this paper.", + "bbox": [ + 511, + 875, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20947", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "- Experimental results show that UniAP outperforms state-of-the-art methods by up to $3.80 \\times$ in throughput and reduces strategy optimization time by up to $107 \\times$ across five Transformer-based models.", + "bbox": [ + 89, + 90, + 486, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background", + "text_level": 1, + "bbox": [ + 89, + 167, + 220, + 183 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Parallel strategy", + "text_level": 1, + "bbox": [ + 89, + 191, + 253, + 209 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pipeline parallelism (PP) In PP [15], each worker (machine or GPU) holds a subset of model layers. Adjacent layers on different workers need to transfer activations in the forward propagation (FP) step and gradients in the backward propagation (BP) step.", + "bbox": [ + 89, + 214, + 483, + 290 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Data parallelism (DP) In DP [22], each worker holds a replica of the whole model and partitions training samples. In each iteration, each worker computes gradients and synchronizes them with the other workers using all-reduce collective communication (CC). All workers will have the same model parameters after the synchronization step.", + "bbox": [ + 89, + 291, + 483, + 382 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Tensor parallelism (TP) In TP [30], each worker holds a replica of training samples and partitions within model layers. In each iteration, each worker computes its local outputs in FP and its local gradients in BP. To synchronize outputs and gradients, all workers will perform all-reduce CC in FP and BP steps according to the partition scheme.", + "bbox": [ + 89, + 382, + 483, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fully sharded data parallelism (FSDP) FSDP [8, 32] partitions optimizer states, parameters and gradients of the model into separate workers. During the FP and BP step of each iteration, FSDP performs an all-gather CC to obtain the complete parameters for the relevant layer, respectively. After computing gradients, FSDP conducts a reduce-scatter CC to distribute the global gradients among the workers.", + "bbox": [ + 89, + 473, + 483, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Parallel method", + "text_level": 1, + "bbox": [ + 89, + 590, + 250, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Manual parallelism (MP) MP refers to the parallel methods in which human experts design and optimize the parallel strategies. Representative MP methods include Megatron-LM [30], Mesh-TensorFlow [37], and GSPMD [44]. Megatron-LM manually designs TP and PP strategies for training Transformer-based models and exhibits superior efficiency. Mesh-TensorFlow and GSPMD require human effort to designate and tune the intra-layer parallel strategy. These methods rely on expert design and have little flexibility, challenging their automatic application to other models.", + "bbox": [ + 89, + 612, + 483, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inter-layer-only AP or intra-layer-only AP For interlayer-only AP, GPipe [15] and vPipe [46] employ a balanced partition algorithm and a dynamic layer partitioning middleware to partition pipelines, respectively. For intra-layer-only AP, OptCNN [16], TensorOpt [2], and Tofu [42] employ dynamic programming methods to optimize DP and TP strategies together. FlexFlow [17] and Automap [36] use the Monte Carlo method to find the optimal DP and TP strat", + "bbox": [ + 89, + 779, + 485, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/daccd7bd1041350ba01fa7ba9cd1e1eaecb29d6e64a83d9930b70150b159382a.jpg", + "image_caption": [ + "Figure 1. Parallel methods for optimizing parallel strategies for a three-layer model. The different arrangements of slices with varying transparency within the same layer block indicate different intra-layer parallelism strategies adopted by layers. The different arrangements of gray blocks which wrap the layer blocks indicate different inter-layer parallelism strategies." + ], + "image_footnote": [], + "bbox": [ + 516, + 92, + 903, + 328 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "egy. Colossal-Auto [23] utilizes integer programming (IP) techniques to generate intra-layer parallelism and activation checkpointing strategies. All these methods optimize only one category of parallel strategies.", + "bbox": [ + 511, + 438, + 906, + 497 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inter- and intra-layer AP PipeDream [28], DAPPLE [9], and PipeTransformer [14] use dynamic programming to determine optimal strategies for both DP and PP. DNN-partitioning [38] adopts IP and dynamic programming to explore DP and PP strategies. DT-FM [45] combines genetic algorithm and the local search strategy to explore DP and PP strategies in a layer-wise manner. Piper [39] and Alpa [47] adopt a parallel method considering DP, TP, and PP. Galvatron [25] uses dynamic programming to determine DP, TP, and FSDP strategies in a single pipeline stage. As for PP, it partitions stages and determines micro-batch size using naive greedy algorithms. All these methods are hierarchical, which will result in sub-optimal solutions.", + "bbox": [ + 511, + 498, + 908, + 695 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 709, + 604, + 723 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we introduce our proposed method called UniAP, which jointly optimizes the two categories of parallel strategies, including PP, DP, TP, and FSDP, to find an optimal solution. Figure 1 illustrates the difference between UniAP and other automatic parallelism methods. Inter-layer-only and intra-layer-only AP methods optimize (search) from a set of candidate inter-layer-only and intra-layer-only parallel strategies, respectively. Hierarchical AP methods first adopt greedy or dynamic programming to propose candidate inter-layer parallel strategies. Then, they optimize the intra-layer parallel strategy for every fixed inter-layer parallel strategy.", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20948", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e9250057b878b58ef36dc760bdaa16ab25a6caa0041ff5f300fe1e13b4c9702e.jpg", + "image_caption": [ + "Figure 2. Flowchart of UniAP." + ], + "image_footnote": [], + "bbox": [ + 102, + 95, + 475, + 195 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "UniAP has the largest strategy space for exploration (joint optimization).", + "bbox": [ + 89, + 241, + 482, + 270 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2 illustrates the flowchart of UniAP. UniAP first profiles the runtime information for the user's hardware environment and the deep learning model. After that, UniAP estimates inter- and intra-layer costs given the computation graph and profiling results with its cost models. The estimated costs and the computation graph are then transformed into an MIQP problem. The objective function of the MIQP is to maximize the training throughput, or in other words, to minimize the training time per iteration (TPI). By iteratively applying the cost model and MIQP with different parameters, UniAP determines the minimal TPI and its corresponding parallel strategies. We name this process the Unified Optimization Process (UOP). Finally, UniAP interprets the parallel strategies into the execution plan for the designated model.", + "bbox": [ + 89, + 272, + 483, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Profiling", + "text_level": 1, + "bbox": [ + 89, + 510, + 194, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "UniAP collects runtime information about the hardware environment and deep learning model during profiling. For the hardware environment, UniAP evaluates the efficiency of all-reduce and point-to-point (P2P) communication for different device subsets. For example, when profiling a node with 4 GPUs, UniAP measures the all-reduce efficiency for various DP, TP, and FSDP combinations across these GPUs. Additionally, UniAP ranks these GPUs from 0 to 3 and evaluates the speed of P2P for two pipeline options: $(0\\to 2$ and $1\\rightarrow 3)$ and $(0\\to 1,1\\to 2$ and $2\\to 3)$ . Furthermore, UniAP estimates the computation-communication overlap coefficient (CCOC) [25, 33].", + "bbox": [ + 89, + 532, + 483, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "UniAP acquires two types of information for the deep learning model: computation time and memory usage. On one hand, UniAP distinguishes the forward computation time per sample for different types of hidden layers. On the other hand, UniAP collects memory usage information for each layer, including the memory occupied by parameters and the memory usage of activation per sample in different TP sizes.", + "bbox": [ + 89, + 715, + 483, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Cost model", + "text_level": 1, + "bbox": [ + 89, + 847, + 215, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "UniAP employs two primary cost models, namely the time cost model and the memory cost model.", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Time cost model To estimate computation time, UniAP first calculates the forward computation time by multiplying the batch size with the forward computation time per sample obtained from profiling. Users can obtain a more precise result by profiling the forward time on the specified batch size. For Transformer-based models that mainly consist of the MatMul operator, the computation time in the BP stages is roughly twice that of the FP stages [21, 25, 30]. Additionally, UniAP estimates the communication time by dividing the size of transmitting tensors by the profiled communication efficiency for different communication primitives. To accommodate overlapping, UniAP multiplies the profiled CCOC by the overlapping interval of computation and communication. To model the communication time between pipeline stages, UniAP calculates the cross-stage cost between consecutive stages by the summation of P2P costs.", + "bbox": [ + 511, + 90, + 906, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Memory cost model UniAP estimates memory consumption for each layer with its memory cost model. This estimation consists of three steps for a given layer. First, it computes the activation memory cost $m_{a}$ by multiplying the batch size and the profiled activation memory cost per sample of the TP size used by the strategy. Next, UniAP calculates the memory cost of model states $m_{s}$ for each layer based on their parameter size $ps$ , TP size $ts$ , FSDP size $fs$ , and a constant $c_{dtype}$ dependent on the data type. Formally, we have", + "bbox": [ + 511, + 333, + 908, + 481 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nm _ {s} = \\frac {c _ {d t y p e} \\times p s}{t s \\times f s}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 481, + 906, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For example, if we choose precision as FP32, then $c_{dtype} = (4 + 4 + 4 + 4) / 4 = 4$ since the learnable parameters, gradients, momentum and variance consume equal size of memory. If we opt for mixed precision with FP16 activated, then $c_{dtype} = (4 + 4 + 4 + 2 + 2) / 2 = 8$ . Finally, UniAP aggregates the activation memory cost $m_a$ , memory cost of model states $m_s$ , and context memory cost $m_c$ to a constant matrix $M$ , where $M_{uk}$ denotes the memory cost for the $k$ -th intra-layer strategy of layer $u$ on a single device.", + "bbox": [ + 511, + 513, + 908, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Mixed integer quadratic programming", + "text_level": 1, + "bbox": [ + 511, + 656, + 848, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The estimated costs and the computation graph are then transformed into an MIQP problem. Its formulation includes an objective function and several constraints.", + "bbox": [ + 511, + 678, + 906, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3.1. Objective function", + "text_level": 1, + "bbox": [ + 511, + 729, + 689, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The objective function tries to minimize TPI. In this paper, we have chosen GPipe as our PP strategy for illustration.2 Figure 3 depicts the time cost decomposition of a GPipe-style PP with non-negligible communication costs. The time needed to apply gradients at the end of each iteration is not included, as it depends on the optimizer and is insignificant compared to the total time spent on FP and BP.", + "bbox": [ + 511, + 750, + 908, + 854 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "${}^{2}$ UniAP is also compatible with other PP strategies. For example,users need to modify only the memory constraint in Section 3.3.2 to adapt to synchronous 1F1B pipeline [9, 29].", + "bbox": [ + 511, + 862, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20949", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/183a476c7327e98af73d2b0f588b882911e8b9011441b9aab9d066f331a2f2ba.jpg", + "image_caption": [ + "Figure 3. Time cost decomposition of a GPipe-style PP." + ], + "image_footnote": [], + "bbox": [ + 101, + 90, + 566, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/03761f9a11a6fb9e0f102a74ba3ed15114a72e5c16d162b70f2a658c9a47ae1c.jpg", + "image_caption": [ + "Figure 4. A contiguous set." + ], + "image_footnote": [], + "bbox": [ + 589, + 117, + 898, + 219 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We denote the cost for computation stages as $\\mathbb{P} = \\{p_1, p_2, \\ldots, p_{deg}\\}$ and the cost for communication stages as $\\mathbb{O} = \\{o_1, o_2, \\ldots, o_{deg-1}\\}$ . Here, $deg$ represents the number of computation stages, which corresponds to the pipeline parallel size. In Figure 3, $fp_i$ and $bp_i$ denote forward and backward computation time for computation stage $i$ , respectively. $fo_j$ and $bo_j$ denote forward and backward communication time for communication stage $j$ , respectively. Hence, we have $pi = fp_i + bp_i$ and $oj = fo_j + bo_j$ .", + "bbox": [ + 89, + 262, + 483, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In a GPipe-style pipeline, we use $c$ to denote the number of micro-batches. As illustrated in Figure 3, a mini-batch is uniformly split into four micro-batches, and the total TPI is determined by the latency of all computation and communication stages and the latency of the slowest stage. We further denote TPI in GPipe as $tpi_{gpipe}$ . Given that a stage with a higher FP computation cost leads to a higher BP computation cost with high probability, we can write the objective function of GPipe-style pipeline as follows:", + "bbox": [ + 89, + 398, + 483, + 535 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min t p i _ {g p i p e} = \\sum_ {i = 1} ^ {d e g} p _ {i} + \\sum_ {j = 1} ^ {d e g - 1} o _ {j} \\tag {2} \\\\ + (c - 1) \\max (\\mathbb {P} \\cup \\mathbb {O}). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 547, + 483, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2. Constraint", + "text_level": 1, + "bbox": [ + 89, + 623, + 212, + 637 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first introduce additional notations before presenting the constraints. For a given layer $u \\in \\mathbb{V}$ , $\\mathbb{S}_u$ represents its set of intra-layer parallel strategies, $A_{uk}$ denotes the $k$ -th intra-layer execution cost obtained from our time cost model. Additionally, we use $S_{uk} \\in \\{0,1\\}$ to indicate whether the $k$ -th parallel strategy is selected for the layer $u$ , and use $P_{ui} \\in \\{0,1\\}$ to indicate whether layer $u$ is placed on the $i$ -th computation stage. Each edge $\\langle u,v\\rangle \\in \\mathbb{E}$ is assigned a resharding cost denoted by $R_{uv}$ if the vertices are located within the same pipeline stage. Alternatively, if the vertices are located across consecutive stages, the resharding cost between them is denoted by $R_{uv}'$ . These two resharding costs are constant matrices derived from our time cost model.", + "bbox": [ + 89, + 643, + 483, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Computation-stage constraint To compute the total cost for a single computation stage $i$ , all computation and communication costs associated with that stage must be aggregated and assigned to $p_i$ . This constraint can be formulated as", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "follows:", + "bbox": [ + 513, + 263, + 571, + 275 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\sum_ {u \\in \\mathbb {V}} \\boldsymbol {P} _ {u i} \\boldsymbol {S} _ {u} ^ {\\top} \\boldsymbol {A} _ {u} + \\sum_ {\\langle u, v \\rangle \\in \\mathbb {E}} \\boldsymbol {P} _ {u i} \\boldsymbol {P} _ {v i} \\left(\\boldsymbol {S} _ {u} ^ {\\top} \\boldsymbol {R} _ {u v} \\boldsymbol {S} _ {v}\\right) = p _ {i}, \\tag {3} \\\\ \\forall i \\in \\{1, \\dots , d e g \\}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 284, + 906, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "On the left side of Equation (3), the first polynomial term represents the cost of choosing specific intra-layer strategies for layers placed in stage $i$ . The second term represents total resharding costs within stage $i$ .", + "bbox": [ + 511, + 349, + 905, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Communication-stage constraint To calculate the total cost for a single communication stage $j$ , we should aggregate the P2P costs incurred between consecutive stages and assign them to $o_j$ . This constraint can be formulated as follows:", + "bbox": [ + 511, + 410, + 905, + 470 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\sum_ {\\langle u, v \\rangle \\in \\mathbb {E}} \\boldsymbol {P} _ {u j} \\boldsymbol {P} _ {v (j + 1)} \\left(\\boldsymbol {S} _ {u} ^ {\\top} \\boldsymbol {R} _ {u v} ^ {\\prime} \\boldsymbol {S} _ {v}\\right) = o _ {j}, \\tag {4} \\\\ \\forall j \\in \\{1, \\dots , d e g - 1 \\}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 479, + 906, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Memory constraint We need to guarantee that no devices (GPUs) will encounter out-of-memory (OOM) exceptions during training process. This constraint can be formulated as follows:", + "bbox": [ + 511, + 545, + 906, + 604 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {u \\in \\mathbb {V}} \\boldsymbol {P} _ {u i} \\boldsymbol {S} _ {u} ^ {\\top} \\boldsymbol {M} _ {u} \\leqslant m, \\forall i \\in \\{1, \\dots , d e g \\}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 613, + 906, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $m$ denotes the memory limit for each device. In the case of homogeneous computing devices, the value of $m$ remains constant throughout all stages. But the value of $m$ varies in the case of heterogeneous computing devices.", + "bbox": [ + 511, + 655, + 905, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Order-preserving constraint PP is not a single-program multiple-data (SPMD) parallel strategy [15]. Hence, we need an order-preserving constraint to ensure that the subgraphs of $\\mathcal{G}$ are contiguous. We adopt the definition of contiguous from previous work [38, 39].", + "bbox": [ + 511, + 715, + 905, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 3.1. A set $\\mathbb{W} \\subseteq \\mathbb{V}$ is contiguous if there do not exist nodes $u \\in \\mathbb{W}, v \\in \\mathbb{V} \\setminus \\mathbb{W}$ , and $w \\in \\mathbb{W}$ such that $v$ is reachable from $u$ and $w$ is reachable from $v$ .", + "bbox": [ + 511, + 800, + 905, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 4 illustrates an example of a contiguous set $\\mathbb{W}$ , in which we cannot find any reachable node pairs $\\langle u, v \\rangle$ and $\\langle v, w \\rangle$ where $u, w \\in \\mathbb{W}$ and $v \\in \\mathbb{V} \\setminus \\mathbb{W}$ .", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20950", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our case, our model will not be assigned to different pipeline stages in a disordered manner if we ensure that all subgraphs on each computation stage are contiguous. After reformulating this constraint in linear form, we have", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} _ {v i} \\geqslant \\mathbf {P} _ {v i},\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 162, + 192, + 176 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\forall v \\in \\mathbb {V}, \\forall i \\in \\{1, 2, \\dots , d e g \\}, \\tag {6a}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 181, + 483, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} _ {v i} \\leqslant \\mathbf {Z} _ {u i},\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 200, + 192, + 215 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\forall u, v \\in \\mathbb {V}, \\forall \\langle u, v \\rangle \\in \\mathbb {E}, \\forall i \\in \\{1, 2, \\dots , d e g \\}, \\tag {6b}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 218, + 483, + 234 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} _ {v i} \\leqslant \\mathbf {P} _ {v i} - \\mathbf {P} _ {u i} + 1,\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 237, + 263, + 253 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\forall u, v \\in \\mathbb {V}, \\forall \\langle u, v \\rangle \\in \\mathbb {E}, \\forall i \\in \\{1, 2, \\dots , d e g \\}. \\tag {6c}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 256, + 483, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\mathbf{Z}$ is an auxiliary variable shaped like $\\mathbf{P}$ . Given the node set $\\mathbb{V}_i$ that contains nodes in stage $i$ , we define $\\mathbf{Z}_{vi} = 1$ if a node $w \\in \\mathbb{V}_i$ is reachable from $v$ ( $v \\in \\mathbb{V}$ ). Otherwise, $\\mathbf{Z}_{vi} = 0$ . Detailed proof can be found in Appendix B.", + "bbox": [ + 89, + 282, + 483, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Layer-placement constraint All layers should be placed on exactly one pipeline stage and at least one layer should be placed on each pipeline stage. This constraint can be formulated as follows:", + "bbox": [ + 89, + 344, + 483, + 402 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {i = 1} ^ {\\text {d e g}} \\boldsymbol {P} _ {u i} = 1, \\quad \\forall u \\in \\mathbb {V}, \\tag {7a}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 412, + 483, + 452 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {u \\in \\mathbb {V}} \\boldsymbol {P} _ {u i} \\geqslant 1, \\quad \\forall i \\in \\{1, \\dots , d e g \\}, \\tag {7b}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 455, + 483, + 487 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {P} _ {u i} \\in \\{0, 1 \\}, \\quad \\forall u \\in \\mathbb {V}, i \\in \\{1, \\dots , d e g \\}. \\tag {7c}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 491, + 483, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Strategy-selection constraint Each layer must and can choose only one strategy. This constraint can be formulated as follows:", + "bbox": [ + 89, + 517, + 483, + 561 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {k = 1} ^ {| \\mathbb {S} _ {u} |} \\boldsymbol {S} _ {u k} = 1, \\quad \\forall u \\in \\mathbb {V}, \\tag {8a}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 570, + 483, + 613 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {S} _ {u k} \\in \\{0, 1 \\}, \\quad \\forall u \\in \\mathbb {V}, k \\in \\{1, \\dots , | \\mathbb {S} _ {u} | \\}. \\tag {8b}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 616, + 483, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The MIQP formulation for UniAP includes the objective function in Equation (2) and all the constraints from Euqation (3) - (8b).", + "bbox": [ + 89, + 642, + 483, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Unified optimization process", + "text_level": 1, + "bbox": [ + 89, + 696, + 346, + 714 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "UOP integrates the cost model and MIQP based on the profiling results and the computation graph to return the optimal parallel strategy and the corresponding TPI. Algorithm 1 summarizes the whole process. In Algorithm 1, we denote intra-layer cost as $A$ , inter-layer cost as $R$ , cross-stage cost as $R'$ , and memory cost as $M$ . The CostModeling process calculates these four costs based on the cost model described in Section 3.2.", + "bbox": [ + 89, + 719, + 483, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, UOP optimizes intra-layer-only parallelism for cases in which pipeline parallelism is not adopted. Several works [23, 47] have used quadratic integer programming (QIP) to optimize intra-layer-only parallel strategy and", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 Unified Optimization Process", + "bbox": [ + 514, + 90, + 799, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: Profiling results $PR$ , strategy dictionary $SD$ , minibatch size $B$ , computation graph $\\mathcal{G}$ , and the number of GPUs $n$ .", + "bbox": [ + 527, + 109, + 906, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: Optimal cost $cost^{*}$ , pipeline parallel size $deg^{*}$ , the number of micro-batches $c^{*}$ , layer placement $P^{*}$ , and intra-layer strategy $S^{*}$ .", + "bbox": [ + 527, + 155, + 906, + 200 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd e g ^ {*} = 1;\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 200, + 601, + 215 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nc ^ {*} = B;\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 217, + 588, + 229 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA, R, \\_ , M = \\operatorname {C o s t M o d e l i n g} (P R, S D [ 1 ], \\mathcal {G}, B);\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 231, + 874, + 244 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nc o s t ^ {*}, P ^ {*}, S ^ {*} = \\mathbb {Q} \\operatorname {I P} (\\boldsymbol {A}, \\boldsymbol {R}, \\boldsymbol {M});\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 246, + 741, + 260 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Get all factors for $n$ except 1 and insert them to set $\\mathbb{F}$ ;", + "bbox": [ + 529, + 261, + 885, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Get all factors for $B$ except 1 and insert them to set $\\mathbb{B}$", + "bbox": [ + 529, + 277, + 888, + 290 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {f o r} d e g \\text {i n} \\mathbb {F} \\text {d o}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 291, + 637, + 305 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf o r c i n \\mathbb {B} d o\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 306, + 635, + 319 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {M i c r o - b a t c h s i z e} b = B / c;\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 321, + 741, + 335 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {A}, \\boldsymbol {R}, \\boldsymbol {R} ^ {\\prime}, \\boldsymbol {M} = \\text {C o s t M o d e l i n g} (P R, S D [ d e g ], \\mathcal {G}, \\\\ b); \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 337, + 906, + 364 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nc o s t, \\boldsymbol {P}, \\boldsymbol {S} = \\operatorname {M I Q P} (\\boldsymbol {A}, \\boldsymbol {R}, \\boldsymbol {R} ^ {\\prime}, \\boldsymbol {M}, \\deg , c);\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 367, + 826, + 381 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {i f} \\quad c o s t < c o s t ^ {*} \\text {t h e n}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 383, + 702, + 395 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nc o s t ^ {*}, d e g ^ {*}, c ^ {*}, P ^ {*}, S ^ {*} = c o s t, d e g, c, P, S;\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 397, + 864, + 412 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ne n d i f\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 412, + 607, + 425 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ne n d \\mathrm {f o r}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 428, + 601, + 440 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ne n d \\mathrm {f o r}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 443, + 584, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "achieved promising results. UniAP provides a QIP formulation for intra-layer-only parallelism in Appendix C.", + "bbox": [ + 511, + 487, + 906, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, UOP enumerates all factors of $n$ except 1 as the pipeline parallel size $deg$ . For each $deg$ , UOP enumerates all factors of $B$ except 1 as the number of micro-batches $c$ . These enumerations aim to achieve load balance on a homogeneous cluster.", + "bbox": [ + 511, + 517, + 905, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each candidate $deg$ and $c$ , UOP formulates the cost for a training iteration to an MIQP expression. It then waits for the MIQP solver to return the optimal cost and parallel strategy.", + "bbox": [ + 511, + 594, + 905, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, UOP returns the minimum cost $cost^*$ and its corresponding pipeline parallel size $deg^*$ , number of microbatches $c^*$ , layer placement $P^*$ , and intra-layer strategies $S^*$ . We provide visualization for a candidate solution to UOP in Appendix D.", + "bbox": [ + 511, + 655, + 906, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Complexity analysis", + "text_level": 1, + "bbox": [ + 511, + 741, + 704, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $|\\mathbb{V}|$ , $|\\mathbb{S}|$ , and $n$ denote the number of layers, parallel strategies, and GPUs, respectively. As illustrated in Algorithm 1, UniAP enumerates all factors of $n$ except 1 as the $deg$ in the outer loop and all factors of $B$ except 1 in the inner loop. The time complexity for the two loops in the UniAP algorithm is $\\mathcal{O}(\\sqrt{Bn})$ . Within the inner loop body, UniAP calls CostModeling to model the cost of each stage for each parallel strategy. Furthermore, the optimization time limit of the MIQP solver can be set as a constant", + "bbox": [ + 511, + 763, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20951", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "hyperparameter when UniAP calls it. Therefore, the overall computational complexity of UniAP is $\\mathcal{O}(|\\mathbb{V}||\\mathbb{S}|\\sqrt{Bn})$", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 137, + 215, + 154 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiment setup", + "text_level": 1, + "bbox": [ + 89, + 162, + 264, + 179 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct experiments on four different kinds of environments to validate the effectiveness and universality of our method. ENVA refers to a node (machine) with 1 Xeon 6248 CPU, 8 V100-SXM2 32GB GPUs, and 472GB memory. ENVB refers to 4 nodes interconnected with 10Gbps networks and each node has 2 Xeon E5-2620 v4 CPUs, 4 TITAN Xp 12GB GPUs, and 125GB memory. ENVC refers to a node with 8 A100 40GB PCIe GPUs. ENVD is a cloud cluster with 16 nodes and each node has 4 DCUs (a type of non-NVIDIA GPU). In the following text, we specify the number of GPUs after the environment name (e.g., ENVB-8) only when partial nodes of the environment are used.", + "bbox": [ + 89, + 184, + 485, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate UniAP with five Transformer-based models, BERT-Huge [5], T5-Large [31], ViT-Huge [6], SwinHuge [24], and Llama [40, 41]. We follow the common practice of training these transformer-based models. To eliminate factors that affect training throughput, we turn off techniques orthogonal to parallel strategies, such as activation checkpointing [3]. However, we integrate FP16 mixed precision training [26] for the largest model, Llama, in some cases to successfully orchestrate the process. More details on these models are provided in Appendix E.", + "bbox": [ + 89, + 366, + 483, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Several solvers can be used for QIP and MIQP optimization. In our implementation, we choose Gurobi [13], with detailed configurations provided in Appendix E. Exploring other solvers is left for future work.", + "bbox": [ + 89, + 518, + 483, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The experimental evaluation concentrates on two primary metrics: training throughput and strategy optimization time. The former is calculated by averaging throughput from the 10th to the 60th iteration of training, while the latter is determined by measuring the time of the UOP. More details are provided in Appendix E. Besides, we provide the accuracy of our cost model in Appendix F.", + "bbox": [ + 89, + 579, + 483, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Training throughput and strategy optimization time", + "text_level": 1, + "bbox": [ + 89, + 696, + 482, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare the training throughput and strategy optimization time of UniAP with those of the baselines on ENVA, ENVB-8, ENVC, and ENVD-32. For experiments on ENVA, we set the mini-batch size to be 32, 16, 128, and 128 for BERT, T5, ViT, and Swin, respectively. For experiments on ENVB-8, we set the mini-batch size to be 16, 8, 64, and 32 for these four models, respectively. For the Llama model [40, 41] run on ENVC, we set the mini-batch size to be 8. For experiments on ENVA, ENVB-8, and ENVC, we choose Galvatron [25] and Alma [47] as baselines because they have achieved SOTA performance. Specifically,", + "bbox": [ + 89, + 734, + 485, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Galvatron has surpassed other methods, including PyTorch DDP [22], Megatron-LM [30], FSDP [8, 32], GPipe [15], and DeepSpeed 3D [27] in terms of training throughput, as reported in the original paper [25]. Furthermore, Alpa utilizes the Just-In-Time (JIT) compilation feature in JAX and outperforms Megatron-LM and DeepSpeed.", + "bbox": [ + 511, + 90, + 906, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For benchmarks on ENVD-32, pytorch libraries and frameworks need specialized adaptations due to the heterogeneity of DCUs against NVIDIA GPUs. Thus, this part of experiments is based on adapted pytorch libraries and Megatron-DeepSpeed[27] framework provided by the platform developers, and we select MP methods including Megatron [30] and DeepSpeed (ZeRO-3) [34] as our baseline methods. We set the mini-batch size as 8 for Llama-7B and 4 for Llama-13B. The remaining configurations align with those for Llama-7B in Section 4.1, such as activating FP16 mixed precision training.", + "bbox": [ + 511, + 183, + 908, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 shows the training throughput and strategy optimization time on ENVA, ENVB-8, and ENVC. On ENVA, UniAP and Galvatron get the same optimal strategy for BERT-Huge, T5-Large, and ViT-Huge, outperforming Alpa in terms of training throughput and strategy optimization time. In addition, UniAP finds a solution for Swin-Huge, while Galvatron encounters CUDA OOM issues. In particular, UniAP achieves a maximum optimization speedup that is $17 \\times$ faster than Galvatron and hundreds of times faster than Alpa on BERT-Huge. This is mainly due to the ability of the MIQP solver to search for an optimal strategy on multiple threads, while the dynamic programming based methods like Galvatron and Alpa run on a single thread due to their strong data dependency.", + "bbox": [ + 511, + 351, + 908, + 563 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On ENVB-8, UniAP consistently demonstrates competitive or larger training throughput compared to Galvatron and Alpa. We attribute the performance improvement to UniAP's larger strategy space. A detailed study is provided in Section 4.4. Furthermore, UniAP's strategy optimization time is also significantly shorter than the two baseline methods.", + "bbox": [ + 511, + 564, + 908, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On ENVC, UniAP shows an optimization speedup of $11.83 \\times$ and a training speedup of $3.80 \\times$ compared to Galvatron on Llama. We examine the parallel strategy they adopted. UniAP employs an 8-stage PP with a micro-batch size of 1, whereas Galvatron employs a 4-stage PP with a micro-batch size of 3 (with the final micro-batch containing 2 samples). Within each PP stage, Galvatron utilizes a 2-way TP. For ENVC, which is equipped with 8 A100 40GB PCIe GPUs, minimizing communication volume is critical. Since PP has significantly less inter-device communication volume than TP, the parallel strategy discovered by UniAP is more reasonable than Galvatron in this case.", + "bbox": [ + 511, + 657, + 908, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 shows the results on ENVD-32. On this environment, UniAP consistently identifies the fastest parallel strategies while expending considerably less time on strategy optimization compared to Megatron (about $157 \\times$ faster for", + "bbox": [ + 511, + 839, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20952", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/43821e5e3eb06fadba1b6ed9361354782a15a94aaa847b3ab317f6cf365e7786.jpg", + "table_caption": [], + "table_footnote": [ + "1) T5-Large tested on ENVB is restricted to 16/16 layers to avoid out-of-memory (OOM) exceptions.", + "2) MEM×: OOM exceptions during strategy optimization.", + "3) CUDA×: CUDA OOM exceptions during model training.", + "4) The official implementation for Swin-Huge and Llama in Alpa is absent. We have endeavored to adopt the code, yet encountered compilation errors. Consequently, experiments related to these models on Alpa are marked as N/A." + ], + "table_body": "
Env.ModelTraining throughput (samples/s)Minimum speedupMaximum speedup
GalvatronAlpaUniAP
ENVABERT-Huge33.46 ± 0.2831.56 ± 0.0433.46 ± 0.281.001.06
T5-Large23.29 ± 0.04MEM×2)23.29 ± 0.041.001.00
ViT-Huge109.51 ± 0.0797.66 ± 1.42109.51 ± 0.071.001.12
Swin-HugeCUDA×3)N/Å4)67.96 ± 0.12N/Å4)N/Å4)
ENVB-8BERT-Huge6.27 ± 0.178.95 ± 0.0610.77 ± 0.131.201.71
T5-Large1)8.06 ± 0.06MEM×2)7.98 ± 0.050.990.99
ViT-Huge32.20 ± 0.1738.74 ± 0.2045.58 ± 0.541.181.41
Swin-Huge13.90 ± 0.17N/Å4)19.08 ± 0.101.371.37
ENVCLlama-7B1.22 ± 0.01N/Å4)4.63 ± 0.0073.803.80
Env.ModelStrategy optimization time (min.)Minimum speedupMaximum speedup
GalvatronAlpaUniAP
ENVABERT-Huge6.44 ± 0.588>400.37 ± 0.00217.29>107.41
T5-Large12.41 ± 0.122MEM×2)0.89 ± 0.00713.9813.98
ViT-Huge6.29 ± 0.464>400.57 ± 0.00910.95>69.60
Swin-Huge11.88 ± 0.666N/Å4)2.16 ± 0.0045.495.49
ENVB-8BERT-Huge2.04 ± 0.010>401.51 ± 0.0051.34>26.32
T5-Large1)2.64 ± 0.110MEM×2)0.91 ± 0.0052.902.90
ViT-Huge2.37 ± 0.180>401.11 ± 0.0112.14>36.01
Swin-Huge4.29 ± 0.320N/Å4)2.29 ± 0.0101.871.87
ENVCLlama-7B6.84 ± 0.055N/Å4)0.58 ± 0.00611.8311.83
", + "bbox": [ + 178, + 88, + 820, + 464 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8178c288e8c8274ea3f1b842870bfa4c802ffbae6fd173dfcc76569c5e4db385.jpg", + "table_caption": [ + "Table 1. Training throughput and strategy optimization time on ENVA, ENVB-8, and ENVC." + ], + "table_footnote": [ + "1) $\\mathrm{SOL} \\times$ : No solution after strategy optimization." + ], + "table_body": "
ModelTraining throughput (samples/s)Strategy optimization time (min.)
MegatronDeepSpeedUniAPMegatronDeepSpeedUniAP
Llama-7B2.01 ± 0.005\\( SOL \\times^{1)} \\)2.01 ± 0.005>8.0 hours\\( SOL \\times^{1)} \\)3.07 ± 0.121
Llama-13B0.82 ± 0.001\\( SOL \\times^{1)} \\)0.82 ± 0.001>2.5 hours\\( SOL \\times^{1)} \\)1.95 ± 0.076
", + "bbox": [ + 178, + 560, + 818, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Training throughput and strategy optimization time on ENVD-32.", + "bbox": [ + 274, + 657, + 718, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Llama-7B and $77 \\times$ faster for Llama-13B). Please note that the strategy optimization time Megatron needs for Llama-7B surpasses that for Llama-13B. We attribute this phenomenon to the variations in the mini-batch size. Specifically, the escalation from a mini-batch size of 4 (employed in Llama-13B) to 8 (employed in Llama-7B) leads to a rise in the number of candidate parallel strategies, as well as the parallel strategies that will successfully train the model. As a result, the strategy optimization time will become increasingly long when the mini-batch size exceeds 8 in the pretraining scenario [1, 40, 41]. In this situation, UniAP will identify the optimal parallel strategy much faster.", + "bbox": [ + 88, + 680, + 485, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, our experiments highlight a limitation encountered with DeepSpeed (ZeRO-3). It requires the mini", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "batch size to be divisible evenly by the total number of computation devices. This specific prerequisite prevents DeepSpeed from successfully launching the training process with 32 DCUs.", + "bbox": [ + 511, + 680, + 906, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To facilitate further discussions, we provide a case study of the optimal parallel strategy in Appendix G.", + "bbox": [ + 511, + 742, + 906, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Scalability", + "text_level": 1, + "bbox": [ + 511, + 786, + 630, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We study the scalability of UniAP using BERT, T5, ViT, and Swin on ENVB. We set the mini-batch sizes for each data point as 8, 4, 32, and 16 times the number of nodes (denoted as '#nodes'). The experimental results are shown in Figure 5. In Figure 5a, the training throughput of the optimal strategy demonstrate near-linearity as the number of nodes and mini", + "bbox": [ + 511, + 809, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20953", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/66722c8fe49e8b37355d8087928ffce42c0cd8255580b2159f01564323db92f2.jpg", + "table_caption": [], + "table_footnote": [ + "1) 1 st.: The training throughput achieved by the fastest parallel strategy.", + "2) 2 nd.: The training throughput achieved by the second fastest parallel strategy.", + "3) Slowest: The training throughput achieved by the slowest parallel strategy.", + "4) Median: The median value of training throughputs across parallel strategies that will successfully train the model.", + "5) #infeasible: The number of parallel strategies that will encounter exceptions such as CUDA OOM during model training.", + "6) #candidate: The total number of candidate parallel strategies." + ], + "table_body": "
ModelBatch sizeTraining throughput (samples/s)#infeasible5)#candidate6)
1 st.1)2 nd.2)Slowest3)Median4)
Llama-7B82.011.920.220.824164
Llama-13B40.820.580.270.424248
", + "bbox": [ + 161, + 88, + 836, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/004e96860ab7f6150ce8899db71deaa4401cda5a214c3fa308f9c6d4b73b679f.jpg", + "image_caption": [ + "(a) Training throughput." + ], + "image_footnote": [], + "bbox": [ + 102, + 268, + 281, + 351 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a83ef687e1bdd872c0500a20f62216c10a3a2e292f630d8d6dfa4a39e8fc7f70.jpg", + "image_caption": [ + "(b) Strategy optimization time.", + "Figure 5. Scalability on training throughput and strategy optimization time." + ], + "image_footnote": [], + "bbox": [ + 292, + 268, + 478, + 352 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0f57b362b266a39694d18e25280700986792601dea87ce1deaac3ae0173fbf87.jpg", + "table_caption": [ + "Table 3. Statistics on the candidate parallel strategies for Megatron." + ], + "table_footnote": [], + "table_body": "
Number of DCUsTraining throughput (samples/s)Strategy optimization time(min.)
160.1726 ± 0.00050.75 ± 0.0100
320.3393 ± 0.00041.64 ± 0.0321
640.6656 ± 0.00231.55 ± 0.0058
", + "bbox": [ + 101, + 417, + 473, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Scalability on training throughput among DCU clusters.", + "bbox": [ + 93, + 513, + 478, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "batch size increase. In Figure 5b, the strategy optimization time matches the complexity analysis in Section 3.5.", + "bbox": [ + 89, + 545, + 482, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additionally, we use DCU nodes in ENVD to scale UniAP to larger clusters on Llama-7B, with precision of FP32. We conducted experiments on clusters with 16, 32, and 64 DCUs, where we set the mini-batch size to $1/8$ of the number of DCUs (i.e., for a cluster with 16 DCUs, we set the minibatch size to 2). The experimental results are shown in Table 4, further demonstrating the near-linear scalability of UniAP.", + "bbox": [ + 89, + 577, + 483, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Further study on the significance of UniAP", + "text_level": 1, + "bbox": [ + 89, + 696, + 455, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, we investigate the significance of UniAP by answering two questions. The first question is why we need AP rather than MP. We examine the statistics on the candidate parallel strategies for Megatron, as shown in Table 3. With MP, most users are unfamiliar with the complex hyperparameter detail of the candidate parallel strategies and typically choose a parallel strategy at random, which will result in a $64.1\\%$ (41/64) chance of failing to train Llama-7B and an $87.5\\%$ (42/48) chance of failing to train Llama-13B even when the hardware resources are sufficient. Even if they successfully eliminate all infeasible parallel strategies, they still only have a $50\\%$ chance of identifying a parallel strata", + "bbox": [ + 88, + 719, + 485, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6a3ed920a1cd6fc2b02003b23020df2c91b40ad6c6921d0d8b5b17713eb59e0a.jpg", + "image_caption": [ + "Figure 6. Ablation study. $\\mathrm{SOL} \\times$ indicates no feasible solution during strategy optimization, and $\\mathrm{CUDA} \\times$ indicates CUDA OOM exceptions during training." + ], + "image_footnote": [], + "bbox": [ + 529, + 272, + 885, + 381 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "egy with training throughput better than 0.82 samples/s for Llama-7B and 0.42 samples/s for Llama-13B. When the users can't identify the fastest parallel strategy out of hundreds of candidates, they may sacrifice at least $4.4\\%$ of the training throughput for Llama-7B and $29.2\\%$ for Llama-13B. In such circumstances, UniAP is able to identify the fastest strategy, as shown in Table 2.", + "bbox": [ + 511, + 444, + 906, + 550 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The second question is why UniAP achieves better performance than other AP methods. We investigate the importance of the strategy space for the optimality of parallel strategies with an ablation study. Specifically, we constrain the strategy space to inter-layer-only and intra-layer-only strategies and evaluate the training throughput of the resulting optimal strategy on ENVB. We set the mini-batch sizes to be 16, 12, 64, and 32, respectively. Results are shown in Figure 6. We can find that constraining the strategy space compromises the optimality of parallel strategies or gets strategies that encounter OOM across different models. Hence, unifying inter- and intra-layer AP for joint optimization makes UniAP outperform other AP methods.", + "bbox": [ + 511, + 551, + 908, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 767, + 633, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel AP method called UniAP to unify inter- and intra-layer AP by MIQP. To the best of our knowledge, UniAP is the first parallel method that can jointly optimize the two categories of parallel strategies to find an optimal solution. Experimental results show that UniAP can outperform other state-of-the-art baselines to achieve the best performance.", + "bbox": [ + 511, + 794, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20954", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 89, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by National Key R&D Program of China (No.2020YFA0713900), NSFC Project (No.12326615), the Key Major Project of the Pengcheng Laboratory (No.PCL2024A06).", + "bbox": [ + 89, + 114, + 485, + 172 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 200, + 187, + 215 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language Models are Few-Shot Learners. In Advances in Neural Information Processing Systems 33, pages 1877-1901, 2020. 1, 7", + "[2] Zhenkun Cai, Xiao Yan, Kaihao Ma, Yidi Wu, Yuzhen Huang, James Cheng, Teng Su, and Fan Yu. TensorOpt: Exploring the Tradeoffs in Distributed DNN Training With AutoParallelism. IEEE Transactions on Parallel and Distributed Systems, 33(8):1967-1981, 2022. 1, 2", + "[3] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training Deep Nets with Sublinear Memory Cost. CoRR, abs/1604.06174, 2016. 6", + "[4] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. PaLM: Scaling Language Modeling with Pathways. Journal of Machine Learning Research, 24:240:1-240:113, 2023. 17", + "[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In *The North American Chapter of the Association for Computational Linguistics: Human Language Technologies*, pages 4171–4186, 2019. 6, 14", + "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + ], + "bbox": [ + 99, + 224, + 486, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In International Conference on Learning Representations, 2021. 6, 14", + "[7] Nan Du, Yanping Huang, Andrew M. Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yanqi Zhou, Adams Wei Yu, Orhan Firat, Barret Zoph, Liam Fedus, Maarten P. Bosma, Zongwei Zhou, Tao Wang, Yu Emma Wang, Kellie Webster, Marie Pellat, Kevin Robinson, Kathleen S. Meier-Hellstern, Toju Duke, Lucas Dixon, Kun Zhang, Quoc V. Le, Yonghui Wu, Zhifeng Chen, and Claire Cui. GLaM: Efficient Scaling of Language Models with Mixture- of-Experts. In International Conference on Machine Learning, pages 5547-5569, 2022. 1", + "[8] FairScale authors. FairScale: A general purpose modular PyTorch library for high performance and large scale training. https://github.com/facebookresearch/fairscale, 2021. 1, 2, 6, 14", + "[9] Shiqing Fan, Yi Rong, Chen Meng, Zongyan Cao, Siyu Wang, Zhen Zheng, Chuan Wu, Guoping Long, Jun Yang, Lixue Xia, Lansong Diao, Xiaoyong Liu, and Wei Lin. DAPPLE: A Pipelined Data Parallel Approach for Training Large Models. In Symposium on Principles and Practice of Parallel Programming, pages 431-445, 2021. 1, 2, 3", + "[10] William Fedus, Barret Zoph, and Noam Shazeer. Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity. The Journal of Machine Learning Research, 23(120):5232-5270, 2022. 1", + "[11] Michael J. Flynn. Very High-speed Computing Systems. Proceedings of the IEEE, 54(12):1901-1909, 1966. 1", + "[12] Michael J. Flynn. Some Computer Organizations and Their Effectiveness. IEEE Transactions on Computers, C-21(9): 948–960, 1972. 1", + "[13] Gurobi Optimization, LLC. Gurobi Optimizer Reference Manual. https://www.gurobi.com, 2023. 6, 14", + "[14] Chaoyang He, Shen Li, Mahdi Soltanolkotabi, and Salman Avestimehr. PipeTransformer: Automated Elastic Pipelining for Distributed Training of Large-scale Models. In International Conference on Machine Learning, pages 4150-4159, 2021. 1, 2", + "[15] Yanping Huang, Youlong Cheng, Ankur Bapna, Orhan First, Dehao Chen, Mia Xu Chen, HyoukJoong Lee, Jiquan Ngiam, Quoc V. Le, Yonghui Wu, and Zhifeng Chen. GPipe: Efficient Training of Giant Neural Networks using Pipeline Parallelism. In Advances in Neural Information Processing Systems 32, pages 103-112, 2019. 1, 2, 4, 6, 14", + "[16] Zhihao Jia, Sina Lin, Charles R. Qi, and Alex Aiken. Exploring Hidden Dimensions in Parallelizing Convolutional Neural Networks. In International Conference on Machine Learning, pages 2279-2288, 2018. 1, 2", + "[17] Zhihao Jia, Matei Zaharia, and Alex Aiken. Beyond Data and Model Parallelism for Deep Neural Networks. In Machine Learning and Systems, pages 1-13, 2019. 1, 2", + "[18] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. In International Conference on Learning Representations, 2015. 14" + ], + "bbox": [ + 514, + 92, + 908, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20955", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[19] Rafael Lazimy. Mixed-integer Quadratic Programming. Math. Program., 22(1):332-349, 1982. 1", + "[20] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. GShard: Scaling Giant Models with Conditional Computation and Automatic Sharding. In International Conference on Learning Representations, 2021. 1", + "[21] Shigang Li and Torsten Hoefler. Chimera: Efficiently Training Large-Scale Neural Networks with Bidirectional Pipelines. In International Conference for High Performance Computing, Networking, Storage and Analysis, pages 27:1-27:14, 2021. 1, 3", + "[22] Shen Li, Yanli Zhao, Rohan Varma, Omkar Salpekar, Pieter Noordhuis, Teng Li, Adam Paszke, Jeff Smith, Brian Vaughan, Pritam Damania, and Soumith Chintala. PyTorch Distributed: Experiences on Accelerating Data Parallel Training. Proceedings of the VLDB Endowment, 13(12):3005-3018, 2020. 1, 2, 6, 14", + "[23] Yuliang Liu, Shenggui Li, Jiarui Fang, Yanjun Shao, Boyuan Yao, and Yang You. Colossal-Auto: Unified Automation of Parallelization and Activation Checkpoint for Large-scale Models. CoRR, abs/2302.02599, 2023. 1, 2, 5", + "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. In International Conference on Computer Vision, pages 9992-10002, 2021. 6, 14", + "[25] Xupeng Miao, Yujie Wang, Youhe Jiang, Chunan Shi, Xiaonan Nie, Hailin Zhang, and Bin Cui. Galvatron: Efficient Transformer Training over Multiple GPUs Using Automatic Parallelism. Proceedings of the VLDB Endowment, 16(3): 470-479, 2022. 2, 3, 6, 15", + "[26] Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory F. Diamos, Erich Elsen, David García, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, and Hao Wu. Mixed Precision Training. In International Conference on Learning Representations, 2018. 6", + "[27] Microsoft. Megatron-DeepSpeed. https://github.com/microsoft/Megatron-DeepSpeed/tree/main, 2021.6", + "[28] Deepak Narayanan, Aaron Harlap, Amar Phanishayee, Vivek Seshadri, Nikhil R. Devanur, Gregory R. Ganger, Phillip B. Gibbons, and Matei Zaharia. PipeDream: Generalized Pipeline Parallelism for DNN Training. In Symposium on Operating Systems Principles, pages 1-15, 2019. 1, 2", + "[29] Deepak Narayanan, Amar Phanishayee, Kaiyu Shi, Xie Chen, and Matei Zaharia. Memory-Efficient Pipeline-Parallel DNN Training. In International Conference on Machine Learning, pages 7937–7947, 2021. 1, 3", + "[30] Deepak Narayanan, Mohammad Shoeybi, Jared Casper, Patrick LeGresley, Mostofa Patwary, Vijay Korthikanti, Dmitri Vainbrand, Prethvi Kashinkunti, Julie Bernauer, Bryan Catanzaro, Amar Phanishayee, and Matei Zaharia. Efficient Large-scale Language Model Training on GPU Clusters Using Megatron-LM. In International Conference for High Performance Computing, Networking, Storage and Analysis, pages 58:1-58:15, 2021. 1, 2, 3, 6, 14" + ], + "bbox": [ + 89, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 6, 14", + "[32] Samyam Rajbhandari, Jeff Rasley, Olatunjri Ruwase, and Yuxiong He. ZeRO: Memory Optimizations Toward Training Trillion Parameter Models. In International Conference for High Performance Computing, Networking, Storage and Analysis, pages 20:1-20:16, 2020. 2, 6", + "[33] Saeed Rashidi, Matthew Denton, Srinivas Sridharan, Sudarshan Srinivasan, Amoghavarsha Suresh, Jade Nie, and Tushar Krishna. Enabling Compute-Communication Overlap in Distributed Deep Learning Training Platforms. In International Symposium on Computer Architecture, pages 540–553, 2021. 3", + "[34] Jeff Rasley, Samyam Rajbhandari, Olatunjri Ruwase, and Yuxiong He. DeepSpeed: System Optimizations Enable Training Deep Learning Models with Over 100 Billion Parameters. In The 26th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 3505-3506, 2020. 1, 6", + "[35] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision, 115(3):211-252, 2015. 14", + "[36] Michael Schaarschmidt, Dominik Grewe, Dimitrios Vytiniotis, Adam Paszke, Georg Stefan Schmid, Tamara Norman, James Molloy, Jonathan Godwin, Norman Alexander Rink, Vinod Nair, and Dan Belov. Automap: Towards Ergonomic Automated Parallelism for ML Models. CoRR, abs/2112.02958, 2021. 1, 2", + "[37] Noam Shazeer, Youlong Cheng, Niki Parmar, Dustin Tran, Ashish Vaswani, Penporn Koanantakool, Peter Hawkins, HyoukJoong Lee, Mingsheng Hong, Cliff Young, Ryan Sepassi, and Blake A. Hechtman. Mesh-TensorFlow: Deep Learning for Supercomputers. In Advances in Neural Information Processing Systems 31, pages 10435-10444, 2018. 1, 2", + "[38] Jakub Tarnawski, Amar Phanishayee, Nikhil R. Devanur, Divya Mahajan, and Fanny Nina Paravecino. Efficient Algorithms for Device Placement of DNN Graph Operators. In Advances in Neural Information Processing Systems 33, pages 15451-15463, 2020. 1, 2, 4, 12", + "[39] Jakub Tarnawski, Deepak Narayanan, and Amar Phanishayee. Piper: Multidimensional Planner for DNN Parallelization. In Advances in Neural Information Processing Systems 34, pages 24829-24840, 2021. 1, 2, 4", + "[40] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurélien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. LLaMA: Open and Efficient Foundation Language Models. CoRR, abs/2302.13971, 2023. 1, 6, 7, 14", + "[41] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas" + ], + "bbox": [ + 514, + 92, + 908, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20956", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Blecher, Cristian Canton-Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, MarieAnne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open Foundation and Fine-Tuned Chat Models. CoRR, abs/2307.09288, 2023. 1, 6, 7, 14", + "[42] Minjie Wang, Chien-Chin Huang, and Jinyang Li. Supporting Very Large Models using Automatic Dataflow Graph Partitioning. In Proceedings of the Fourteenth EuroSys Conference, pages 26:1-26:17, 2019. 1, 2", + "[43] Wikipedia Foundation. Wikipedia Downloads. https:// dumps.wikipedia.org, 2023. 14", + "[44] Yuanzhong Xu, HyoukJoong Lee, Dehao Chen, Blake A. Hechtman, Yanping Huang, Rahul Joshi, Maxim Krikun, Dmitry Lepikhin, Andy Ly, Marcello Maggioni, Ruoming Pang, Noam Shazeer, Shibo Wang, Tao Wang, Yonghui Wu, and Zhifeng Chen. GSPMD: General and Scalable Parallelization for ML Computation Graphs. CoRR, abs/2105.04663, 2021. 1, 2", + "[45] Binhang Yuan, Yongjun He, Jared Davis, Tianyi Zhang, Tri Dao, Beidi Chen, Percy Liang, Christopher Ré, and Ce Zhang. Decentralized Training of Foundation Models in Heterogeneous Environments. In Advances in Neural Information Processing Systems 35, pages 25464-25477, 2022. 2", + "[46] Shixiong Zhao, Fanxin Li, Xusheng Chen, Xiuxian Guan, Jianyu Jiang, Dong Huang, Yuhao Qing, Sen Wang, Peng Wang, Gong Zhang, Cheng Li, Ping Luo, and Heming Cui. vPipe: A Virtualized Acceleration System for Achieving Efficient and Scalable Pipeline Parallel DNN Training. IEEE Transactions on Parallel and Distributed Systems, 33(3):489-506, 2022. 1, 2", + "[47] Lianmin Zheng, Zhuohan Li, Hao Zhang, Yonghao Zhuang, Zhifeng Chen, Yanping Huang, Yida Wang, Yuanzhong Xu, Danyang Zhuo, Eric P. Xing, Joseph E. Gonzalez, and Ion Stoica. Alpa: Automating Inter- and Intra-Operator Parallelism for Distributed Deep Learning. In USENIX Symposium on Operating Systems Design and Implementation, pages 559–578, 2022. 1, 2, 5, 6" + ], + "bbox": [ + 91, + 90, + 485, + 776 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20957", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/UniAP_ Unifying Inter- and Intra-Layer Automatic Parallelism by Mixed Integer Quadratic Programming/b1ddd4d6-5b56-44ba-82ed-6323538fa782_model.json b/2025/UniAP_ Unifying Inter- and Intra-Layer Automatic Parallelism by Mixed Integer Quadratic Programming/b1ddd4d6-5b56-44ba-82ed-6323538fa782_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1aa6058dc66503ff64039b488f2b0d660fdef4eb --- /dev/null +++ b/2025/UniAP_ Unifying Inter- and Intra-Layer Automatic Parallelism by Mixed Integer Quadratic Programming/b1ddd4d6-5b56-44ba-82ed-6323538fa782_model.json @@ -0,0 +1,2708 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.13, + 0.908, + 0.178 + ], + "angle": 0, + "content": "UniAP: Unifying Inter- and Intra-Layer Automatic Parallelism by Mixed Integer Quadratic Programming" + }, + { + "type": "text", + "bbox": [ + 0.272, + 0.203, + 0.725, + 0.275 + ], + "angle": 0, + "content": "Hao Lin\\*, Ke Wu\\*, Jie Li\\*, Jun Li, Wu-Jun Li† \nNational Key Laboratory for Novel Software Technology School of Computer Science \nNanjing University, Nanjing 210023, China" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.276, + 0.884, + 0.293 + ], + "angle": 0, + "content": "{hao.lin, ke.wu, jie-li, lijun}@smail.nju.edu.cn, liwujun@nju.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.359, + 0.486, + 0.631 + ], + "angle": 0, + "content": "Distributed learning is commonly used for training deep learning models, especially large models. In distributed learning, manual parallelism (MP) methods demand considerable human effort and have limited flexibility. Hence, automatic parallelism (AP) methods have recently been proposed for automating the parallel strategy optimization process. Existing AP methods suffer from sub-optimal solutions because they do not jointly optimize the two categories of parallel strategies (i.e., inter-layer parallelism and intra-layer parallelism). In this paper, we propose a novel AP method called UniAP, which unifies inter- and intra-layer automatic parallelism by mixed integer quadratic programming. To the best of our knowledge, UniAP is the first parallel method that can jointly optimize the two categories of parallel strategies to find an optimal solution. Experimental results show that UniAP outperforms state-of-the-art methods by up to \\(3.80 \\times\\) in throughput and reduces strategy optimization time by up to \\(107 \\times\\) across five Transformer-based models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.661, + 0.224, + 0.677 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.486, + 0.854 + ], + "angle": 0, + "content": "Distributed learning (also called parallel learning) on clusters with several machines or GPUs is commonly used for training deep learning models, especially for some large models with billions of parameters [1, 40, 41]. Several parallel strategies, including pipeline parallelism (PP), data parallelism (DP), tensor parallelism (TP), and fully sharded data parallelism (FSDP), have been proposed for distributed learning. These parallel strategies can be divided into two main categories: inter-layer parallelism and intra-layer parallelism. Inter-layer parallelism [7, 9, 10, 15, 20, 21, 28, 29], which includes PP, partitions the model into disjoint sets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.329, + 0.908, + 0.374 + ], + "angle": 0, + "content": "without partitioning tensors in each layer. Intra-layer parallelism [8, 22, 30, 34], which includes DP, TP, and FSDP, partitions tensors in a layer along one or more axes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.376, + 0.909, + 0.571 + ], + "angle": 0, + "content": "The parallel method1 in one specific distributed learning method or system typically adopts one parallel strategy or a combination of several parallel strategies. Existing parallel methods can be divided into two categories: manual parallelism (MP) methods and automatic parallelism (AP) methods. In MP methods [30, 37, 44], one or several parallel strategies are manually optimized by researchers or developers. MP methods require extensive domain knowledge in deep learning models and hardware architectures. With the rapid development of deep learning models and the increasing diversity of modern hardware architectures [11, 12], MP methods demand considerable human effort and have limited flexibility." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.573, + 0.911, + 0.725 + ], + "angle": 0, + "content": "To address the two limitations of MP methods, AP methods [14, 28, 47] have recently been proposed for automating the parallel strategy optimization process. Although existing AP methods have achieved promising progress, they optimize the two categories of parallel strategies separately rather than jointly. More specifically, some methods optimize only one category of parallel strategies [2, 16, 17, 23, 36, 42, 46], and the others optimize inter- and intra-layer parallelism hierarchically [9, 14, 28, 38, 39, 47]. Hence, existing AP methods suffer from sub-optimal solutions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.726, + 0.908, + 0.77 + ], + "angle": 0, + "content": "In this paper, we propose a novel AP method called UniAP for distributed learning. The contributions of UniAP are outlined as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.772, + 0.905, + 0.802 + ], + "angle": 0, + "content": "- UniAP unifies inter- and intra-layer automatic parallelism by mixed integer quadratic programming (MIQP) [19]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.802, + 0.905, + 0.861 + ], + "angle": 0, + "content": "- To the best of our knowledge, UniAP is the first parallel method that can optimize the two categories of parallel strategies jointly rather than separately to find an optimal solution." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.772, + 0.905, + 0.861 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.864, + 0.221, + 0.876 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.876, + 0.236, + 0.888 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.864, + 0.236, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.908, + 0.902 + ], + "angle": 0, + "content": "To avoid confusion, we treat 'parallel method' and 'parallel strategy' as two different terminologies in this paper." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20947" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.488, + 0.152 + ], + "angle": 0, + "content": "- Experimental results show that UniAP outperforms state-of-the-art methods by up to \\(3.80 \\times\\) in throughput and reduces strategy optimization time by up to \\(107 \\times\\) across five Transformer-based models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.168, + 0.221, + 0.184 + ], + "angle": 0, + "content": "2. Background" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.193, + 0.254, + 0.21 + ], + "angle": 0, + "content": "2.1. Parallel strategy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.215, + 0.485, + 0.291 + ], + "angle": 0, + "content": "Pipeline parallelism (PP) In PP [15], each worker (machine or GPU) holds a subset of model layers. Adjacent layers on different workers need to transfer activations in the forward propagation (FP) step and gradients in the backward propagation (BP) step." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.292, + 0.484, + 0.383 + ], + "angle": 0, + "content": "Data parallelism (DP) In DP [22], each worker holds a replica of the whole model and partitions training samples. In each iteration, each worker computes gradients and synchronizes them with the other workers using all-reduce collective communication (CC). All workers will have the same model parameters after the synchronization step." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.383, + 0.484, + 0.473 + ], + "angle": 0, + "content": "Tensor parallelism (TP) In TP [30], each worker holds a replica of training samples and partitions within model layers. In each iteration, each worker computes its local outputs in FP and its local gradients in BP. To synchronize outputs and gradients, all workers will perform all-reduce CC in FP and BP steps according to the partition scheme." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.474, + 0.484, + 0.581 + ], + "angle": 0, + "content": "Fully sharded data parallelism (FSDP) FSDP [8, 32] partitions optimizer states, parameters and gradients of the model into separate workers. During the FP and BP step of each iteration, FSDP performs an all-gather CC to obtain the complete parameters for the relevant layer, respectively. After computing gradients, FSDP conducts a reduce-scatter CC to distribute the global gradients among the workers." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.591, + 0.25, + 0.605 + ], + "angle": 0, + "content": "2.2. Parallel method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.484, + 0.779 + ], + "angle": 0, + "content": "Manual parallelism (MP) MP refers to the parallel methods in which human experts design and optimize the parallel strategies. Representative MP methods include Megatron-LM [30], Mesh-TensorFlow [37], and GSPMD [44]. Megatron-LM manually designs TP and PP strategies for training Transformer-based models and exhibits superior efficiency. Mesh-TensorFlow and GSPMD require human effort to designate and tune the intra-layer parallel strategy. These methods rely on expert design and have little flexibility, challenging their automatic application to other models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Inter-layer-only AP or intra-layer-only AP For interlayer-only AP, GPipe [15] and vPipe [46] employ a balanced partition algorithm and a dynamic layer partitioning middleware to partition pipelines, respectively. For intra-layer-only AP, OptCNN [16], TensorOpt [2], and Tofu [42] employ dynamic programming methods to optimize DP and TP strategies together. FlexFlow [17] and Automap [36] use the Monte Carlo method to find the optimal DP and TP strat" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.341, + 0.908, + 0.425 + ], + "angle": 0, + "content": "Figure 1. Parallel methods for optimizing parallel strategies for a three-layer model. The different arrangements of slices with varying transparency within the same layer block indicate different intra-layer parallelism strategies adopted by layers. The different arrangements of gray blocks which wrap the layer blocks indicate different inter-layer parallelism strategies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.907, + 0.498 + ], + "angle": 0, + "content": "egy. Colossal-Auto [23] utilizes integer programming (IP) techniques to generate intra-layer parallelism and activation checkpointing strategies. All these methods optimize only one category of parallel strategies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.499, + 0.909, + 0.696 + ], + "angle": 0, + "content": "Inter- and intra-layer AP PipeDream [28], DAPPLE [9], and PipeTransformer [14] use dynamic programming to determine optimal strategies for both DP and PP. DNN-partitioning [38] adopts IP and dynamic programming to explore DP and PP strategies. DT-FM [45] combines genetic algorithm and the local search strategy to explore DP and PP strategies in a layer-wise manner. Piper [39] and Alpa [47] adopt a parallel method considering DP, TP, and PP. Galvatron [25] uses dynamic programming to determine DP, TP, and FSDP strategies in a single pipeline stage. As for PP, it partitions stages and determines micro-batch size using naive greedy algorithms. All these methods are hierarchical, which will result in sub-optimal solutions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.71, + 0.605, + 0.724 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this section, we introduce our proposed method called UniAP, which jointly optimizes the two categories of parallel strategies, including PP, DP, TP, and FSDP, to find an optimal solution. Figure 1 illustrates the difference between UniAP and other automatic parallelism methods. Inter-layer-only and intra-layer-only AP methods optimize (search) from a set of candidate inter-layer-only and intra-layer-only parallel strategies, respectively. Hierarchical AP methods first adopt greedy or dynamic programming to propose candidate inter-layer parallel strategies. Then, they optimize the intra-layer parallel strategy for every fixed inter-layer parallel strategy." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20948" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.096, + 0.477, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.195, + 0.212, + 0.381, + 0.226 + ], + "angle": 0, + "content": "Figure 2. Flowchart of UniAP." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.242, + 0.483, + 0.271 + ], + "angle": 0, + "content": "UniAP has the largest strategy space for exploration (joint optimization)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.484, + 0.5 + ], + "angle": 0, + "content": "Figure 2 illustrates the flowchart of UniAP. UniAP first profiles the runtime information for the user's hardware environment and the deep learning model. After that, UniAP estimates inter- and intra-layer costs given the computation graph and profiling results with its cost models. The estimated costs and the computation graph are then transformed into an MIQP problem. The objective function of the MIQP is to maximize the training throughput, or in other words, to minimize the training time per iteration (TPI). By iteratively applying the cost model and MIQP with different parameters, UniAP determines the minimal TPI and its corresponding parallel strategies. We name this process the Unified Optimization Process (UOP). Finally, UniAP interprets the parallel strategies into the execution plan for the designated model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.511, + 0.196, + 0.527 + ], + "angle": 0, + "content": "3.1. Profiling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.484, + 0.715 + ], + "angle": 0, + "content": "UniAP collects runtime information about the hardware environment and deep learning model during profiling. For the hardware environment, UniAP evaluates the efficiency of all-reduce and point-to-point (P2P) communication for different device subsets. For example, when profiling a node with 4 GPUs, UniAP measures the all-reduce efficiency for various DP, TP, and FSDP combinations across these GPUs. Additionally, UniAP ranks these GPUs from 0 to 3 and evaluates the speed of P2P for two pipeline options: \\((0\\to 2\\) and \\(1\\rightarrow 3)\\) and \\((0\\to 1,1\\to 2\\) and \\(2\\to 3)\\). Furthermore, UniAP estimates the computation-communication overlap coefficient (CCOC) [25, 33]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.716, + 0.484, + 0.835 + ], + "angle": 0, + "content": "UniAP acquires two types of information for the deep learning model: computation time and memory usage. On one hand, UniAP distinguishes the forward computation time per sample for different types of hidden layers. On the other hand, UniAP collects memory usage information for each layer, including the memory occupied by parameters and the memory usage of activation per sample in different TP sizes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.216, + 0.863 + ], + "angle": 0, + "content": "3.2. Cost model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "UniAP employs two primary cost models, namely the time cost model and the memory cost model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.332 + ], + "angle": 0, + "content": "Time cost model To estimate computation time, UniAP first calculates the forward computation time by multiplying the batch size with the forward computation time per sample obtained from profiling. Users can obtain a more precise result by profiling the forward time on the specified batch size. For Transformer-based models that mainly consist of the MatMul operator, the computation time in the BP stages is roughly twice that of the FP stages [21, 25, 30]. Additionally, UniAP estimates the communication time by dividing the size of transmitting tensors by the profiled communication efficiency for different communication primitives. To accommodate overlapping, UniAP multiplies the profiled CCOC by the overlapping interval of computation and communication. To model the communication time between pipeline stages, UniAP calculates the cross-stage cost between consecutive stages by the summation of P2P costs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.909, + 0.482 + ], + "angle": 0, + "content": "Memory cost model UniAP estimates memory consumption for each layer with its memory cost model. This estimation consists of three steps for a given layer. First, it computes the activation memory cost \\( m_{a} \\) by multiplying the batch size and the profiled activation memory cost per sample of the TP size used by the strategy. Next, UniAP calculates the memory cost of model states \\( m_{s} \\) for each layer based on their parameter size \\( ps \\), TP size \\( ts \\), FSDP size \\( fs \\), and a constant \\( c_{dtype} \\) dependent on the data type. Formally, we have" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.482, + 0.907, + 0.51 + ], + "angle": 0, + "content": "\\[\nm _ {s} = \\frac {c _ {d t y p e} \\times p s}{t s \\times f s}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.514, + 0.909, + 0.65 + ], + "angle": 0, + "content": "For example, if we choose precision as FP32, then \\( c_{dtype} = (4 + 4 + 4 + 4) / 4 = 4 \\) since the learnable parameters, gradients, momentum and variance consume equal size of memory. If we opt for mixed precision with FP16 activated, then \\( c_{dtype} = (4 + 4 + 4 + 2 + 2) / 2 = 8 \\). Finally, UniAP aggregates the activation memory cost \\( m_a \\), memory cost of model states \\( m_s \\), and context memory cost \\( m_c \\) to a constant matrix \\( M \\), where \\( M_{uk} \\) denotes the memory cost for the \\( k \\)-th intra-layer strategy of layer \\( u \\) on a single device." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.657, + 0.849, + 0.674 + ], + "angle": 0, + "content": "3.3. Mixed integer quadratic programming" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.679, + 0.908, + 0.725 + ], + "angle": 0, + "content": "The estimated costs and the computation graph are then transformed into an MIQP problem. Its formulation includes an objective function and several constraints." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.731, + 0.691, + 0.746 + ], + "angle": 0, + "content": "3.3.1. Objective function" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.856 + ], + "angle": 0, + "content": "The objective function tries to minimize TPI. In this paper, we have chosen GPipe as our PP strategy for illustration.2 Figure 3 depicts the time cost decomposition of a GPipe-style PP with non-negligible communication costs. The time needed to apply gradients at the end of each iteration is not included, as it depends on the optimizer and is insignificant compared to the total time spent on FP and BP." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.863, + 0.908, + 0.901 + ], + "angle": 0, + "content": "\\( {}^{2} \\) UniAP is also compatible with other PP strategies. For example,users need to modify only the memory constraint in Section 3.3.2 to adapt to synchronous 1F1B pipeline [9, 29]." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "20949" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.092, + 0.568, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.17, + 0.231, + 0.503, + 0.246 + ], + "angle": 0, + "content": "Figure 3. Time cost decomposition of a GPipe-style PP." + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.118, + 0.9, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.66, + 0.231, + 0.825, + 0.246 + ], + "angle": 0, + "content": "Figure 4. A contiguous set." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.263, + 0.484, + 0.4 + ], + "angle": 0, + "content": "We denote the cost for computation stages as \\(\\mathbb{P} = \\{p_1, p_2, \\ldots, p_{deg}\\}\\) and the cost for communication stages as \\(\\mathbb{O} = \\{o_1, o_2, \\ldots, o_{deg-1}\\}\\). Here, \\(deg\\) represents the number of computation stages, which corresponds to the pipeline parallel size. In Figure 3, \\(fp_i\\) and \\(bp_i\\) denote forward and backward computation time for computation stage \\(i\\), respectively. \\(fo_j\\) and \\(bo_j\\) denote forward and backward communication time for communication stage \\(j\\), respectively. Hence, we have \\(pi = fp_i + bp_i\\) and \\(oj = fo_j + bo_j\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.484, + 0.536 + ], + "angle": 0, + "content": "In a GPipe-style pipeline, we use \\( c \\) to denote the number of micro-batches. As illustrated in Figure 3, a mini-batch is uniformly split into four micro-batches, and the total TPI is determined by the latency of all computation and communication stages and the latency of the slowest stage. We further denote TPI in GPipe as \\( tpi_{gpipe} \\). Given that a stage with a higher FP computation cost leads to a higher BP computation cost with high probability, we can write the objective function of GPipe-style pipeline as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.548, + 0.484, + 0.612 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min t p i _ {g p i p e} = \\sum_ {i = 1} ^ {d e g} p _ {i} + \\sum_ {j = 1} ^ {d e g - 1} o _ {j} \\tag {2} \\\\ + (c - 1) \\max (\\mathbb {P} \\cup \\mathbb {O}). \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.624, + 0.214, + 0.638 + ], + "angle": 0, + "content": "3.3.2. Constraint" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.839 + ], + "angle": 0, + "content": "We first introduce additional notations before presenting the constraints. For a given layer \\( u \\in \\mathbb{V} \\), \\( \\mathbb{S}_u \\) represents its set of intra-layer parallel strategies, \\( A_{uk} \\) denotes the \\( k \\)-th intra-layer execution cost obtained from our time cost model. Additionally, we use \\( S_{uk} \\in \\{0,1\\} \\) to indicate whether the \\( k \\)-th parallel strategy is selected for the layer \\( u \\), and use \\( P_{ui} \\in \\{0,1\\} \\) to indicate whether layer \\( u \\) is placed on the \\( i \\)-th computation stage. Each edge \\( \\langle u,v\\rangle \\in \\mathbb{E} \\) is assigned a resharding cost denoted by \\( R_{uv} \\) if the vertices are located within the same pipeline stage. Alternatively, if the vertices are located across consecutive stages, the resharding cost between them is denoted by \\( R_{uv}' \\). These two resharding costs are constant matrices derived from our time cost model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Computation-stage constraint To compute the total cost for a single computation stage \\(i\\), all computation and communication costs associated with that stage must be aggregated and assigned to \\(p_i\\). This constraint can be formulated as" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.264, + 0.573, + 0.276 + ], + "angle": 0, + "content": "follows:" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.285, + 0.907, + 0.341 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\sum_ {u \\in \\mathbb {V}} \\boldsymbol {P} _ {u i} \\boldsymbol {S} _ {u} ^ {\\top} \\boldsymbol {A} _ {u} + \\sum_ {\\langle u, v \\rangle \\in \\mathbb {E}} \\boldsymbol {P} _ {u i} \\boldsymbol {P} _ {v i} \\left(\\boldsymbol {S} _ {u} ^ {\\top} \\boldsymbol {R} _ {u v} \\boldsymbol {S} _ {v}\\right) = p _ {i}, \\tag {3} \\\\ \\forall i \\in \\{1, \\dots , d e g \\}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.35, + 0.906, + 0.41 + ], + "angle": 0, + "content": "On the left side of Equation (3), the first polynomial term represents the cost of choosing specific intra-layer strategies for layers placed in stage \\( i \\). The second term represents total resharding costs within stage \\( i \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.411, + 0.906, + 0.471 + ], + "angle": 0, + "content": "Communication-stage constraint To calculate the total cost for a single communication stage \\( j \\), we should aggregate the P2P costs incurred between consecutive stages and assign them to \\( o_j \\). This constraint can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.481, + 0.907, + 0.537 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\sum_ {\\langle u, v \\rangle \\in \\mathbb {E}} \\boldsymbol {P} _ {u j} \\boldsymbol {P} _ {v (j + 1)} \\left(\\boldsymbol {S} _ {u} ^ {\\top} \\boldsymbol {R} _ {u v} ^ {\\prime} \\boldsymbol {S} _ {v}\\right) = o _ {j}, \\tag {4} \\\\ \\forall j \\in \\{1, \\dots , d e g - 1 \\}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.546, + 0.908, + 0.605 + ], + "angle": 0, + "content": "Memory constraint We need to guarantee that no devices (GPUs) will encounter out-of-memory (OOM) exceptions during training process. This constraint can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.614, + 0.907, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\sum_ {u \\in \\mathbb {V}} \\boldsymbol {P} _ {u i} \\boldsymbol {S} _ {u} ^ {\\top} \\boldsymbol {M} _ {u} \\leqslant m, \\forall i \\in \\{1, \\dots , d e g \\}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.906, + 0.717 + ], + "angle": 0, + "content": "Here, \\( m \\) denotes the memory limit for each device. In the case of homogeneous computing devices, the value of \\( m \\) remains constant throughout all stages. But the value of \\( m \\) varies in the case of heterogeneous computing devices." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.717, + 0.906, + 0.793 + ], + "angle": 0, + "content": "Order-preserving constraint PP is not a single-program multiple-data (SPMD) parallel strategy [15]. Hence, we need an order-preserving constraint to ensure that the subgraphs of \\(\\mathcal{G}\\) are contiguous. We adopt the definition of contiguous from previous work [38, 39]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.801, + 0.906, + 0.846 + ], + "angle": 0, + "content": "Definition 3.1. A set \\(\\mathbb{W} \\subseteq \\mathbb{V}\\) is contiguous if there do not exist nodes \\(u \\in \\mathbb{W}, v \\in \\mathbb{V} \\setminus \\mathbb{W}\\), and \\(w \\in \\mathbb{W}\\) such that \\(v\\) is reachable from \\(u\\) and \\(w\\) is reachable from \\(v\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Figure 4 illustrates an example of a contiguous set \\(\\mathbb{W}\\), in which we cannot find any reachable node pairs \\(\\langle u, v \\rangle\\) and \\(\\langle v, w \\rangle\\) where \\(u, w \\in \\mathbb{W}\\) and \\(v \\in \\mathbb{V} \\setminus \\mathbb{W}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20950" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "In our case, our model will not be assigned to different pipeline stages in a disordered manner if we ensure that all subgraphs on each computation stage are contiguous. After reformulating this constraint in linear form, we have" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.163, + 0.193, + 0.178 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} _ {v i} \\geqslant \\mathbf {P} _ {v i},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.182, + 0.484, + 0.198 + ], + "angle": 0, + "content": "\\[\n\\forall v \\in \\mathbb {V}, \\forall i \\in \\{1, 2, \\dots , d e g \\}, \\tag {6a}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.201, + 0.193, + 0.216 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} _ {v i} \\leqslant \\mathbf {Z} _ {u i},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.219, + 0.484, + 0.235 + ], + "angle": 0, + "content": "\\[\n\\forall u, v \\in \\mathbb {V}, \\forall \\langle u, v \\rangle \\in \\mathbb {E}, \\forall i \\in \\{1, 2, \\dots , d e g \\}, \\tag {6b}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.238, + 0.264, + 0.254 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} _ {v i} \\leqslant \\mathbf {P} _ {v i} - \\mathbf {P} _ {u i} + 1,\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.257, + 0.484, + 0.274 + ], + "angle": 0, + "content": "\\[\n\\forall u, v \\in \\mathbb {V}, \\forall \\langle u, v \\rangle \\in \\mathbb {E}, \\forall i \\in \\{1, 2, \\dots , d e g \\}. \\tag {6c}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.283, + 0.484, + 0.344 + ], + "angle": 0, + "content": "Here, \\( \\mathbf{Z} \\) is an auxiliary variable shaped like \\( \\mathbf{P} \\). Given the node set \\( \\mathbb{V}_i \\) that contains nodes in stage \\( i \\), we define \\( \\mathbf{Z}_{vi} = 1 \\) if a node \\( w \\in \\mathbb{V}_i \\) is reachable from \\( v \\) (\\( v \\in \\mathbb{V} \\)). Otherwise, \\( \\mathbf{Z}_{vi} = 0 \\). Detailed proof can be found in Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.345, + 0.484, + 0.404 + ], + "angle": 0, + "content": "Layer-placement constraint All layers should be placed on exactly one pipeline stage and at least one layer should be placed on each pipeline stage. This constraint can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.413, + 0.484, + 0.453 + ], + "angle": 0, + "content": "\\[\n\\sum_ {i = 1} ^ {\\text {d e g}} \\boldsymbol {P} _ {u i} = 1, \\quad \\forall u \\in \\mathbb {V}, \\tag {7a}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.457, + 0.484, + 0.488 + ], + "angle": 0, + "content": "\\[\n\\sum_ {u \\in \\mathbb {V}} \\boldsymbol {P} _ {u i} \\geqslant 1, \\quad \\forall i \\in \\{1, \\dots , d e g \\}, \\tag {7b}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.492, + 0.484, + 0.509 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {P} _ {u i} \\in \\{0, 1 \\}, \\quad \\forall u \\in \\mathbb {V}, i \\in \\{1, \\dots , d e g \\}. \\tag {7c}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.484, + 0.563 + ], + "angle": 0, + "content": "Strategy-selection constraint Each layer must and can choose only one strategy. This constraint can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.123, + 0.571, + 0.484, + 0.614 + ], + "angle": 0, + "content": "\\[\n\\sum_ {k = 1} ^ {| \\mathbb {S} _ {u} |} \\boldsymbol {S} _ {u k} = 1, \\quad \\forall u \\in \\mathbb {V}, \\tag {8a}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.121, + 0.617, + 0.484, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {S} _ {u k} \\in \\{0, 1 \\}, \\quad \\forall u \\in \\mathbb {V}, k \\in \\{1, \\dots , | \\mathbb {S} _ {u} | \\}. \\tag {8b}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.485, + 0.689 + ], + "angle": 0, + "content": "The MIQP formulation for UniAP includes the objective function in Equation (2) and all the constraints from Euqation (3) - (8b)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.347, + 0.715 + ], + "angle": 0, + "content": "3.4. Unified optimization process" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.839 + ], + "angle": 0, + "content": "UOP integrates the cost model and MIQP based on the profiling results and the computation graph to return the optimal parallel strategy and the corresponding TPI. Algorithm 1 summarizes the whole process. In Algorithm 1, we denote intra-layer cost as \\( A \\), inter-layer cost as \\( R \\), cross-stage cost as \\( R' \\), and memory cost as \\( M \\). The CostModeling process calculates these four costs based on the cost model described in Section 3.2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "First, UOP optimizes intra-layer-only parallelism for cases in which pipeline parallelism is not adopted. Several works [23, 47] have used quadratic integer programming (QIP) to optimize intra-layer-only parallel strategy and" + }, + { + "type": "code_caption", + "bbox": [ + 0.515, + 0.092, + 0.8, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 Unified Optimization Process" + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.111, + 0.908, + 0.154 + ], + "angle": 0, + "content": "Input: Profiling results \\(PR\\), strategy dictionary \\(SD\\), minibatch size \\(B\\), computation graph \\(\\mathcal{G}\\), and the number of GPUs \\(n\\)." + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.156, + 0.907, + 0.201 + ], + "angle": 0, + "content": "Output: Optimal cost \\(cost^{*}\\), pipeline parallel size \\(deg^{*}\\), the number of micro-batches \\(c^{*}\\), layer placement \\(P^{*}\\), and intra-layer strategy \\(S^{*}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.202, + 0.602, + 0.217 + ], + "angle": 0, + "content": "\\[\nd e g ^ {*} = 1;\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.218, + 0.589, + 0.23 + ], + "angle": 0, + "content": "\\[\nc ^ {*} = B;\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.232, + 0.875, + 0.246 + ], + "angle": 0, + "content": "\\[\nA, R, \\_ , M = \\operatorname {C o s t M o d e l i n g} (P R, S D [ 1 ], \\mathcal {G}, B);\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.247, + 0.742, + 0.261 + ], + "angle": 0, + "content": "\\[\nc o s t ^ {*}, P ^ {*}, S ^ {*} = \\mathbb {Q} \\operatorname {I P} (\\boldsymbol {A}, \\boldsymbol {R}, \\boldsymbol {M});\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.262, + 0.886, + 0.277 + ], + "angle": 0, + "content": "Get all factors for \\( n \\) except 1 and insert them to set \\( \\mathbb{F} \\);" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.278, + 0.889, + 0.291 + ], + "angle": 0, + "content": "Get all factors for \\(B\\) except 1 and insert them to set \\(\\mathbb{B}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.292, + 0.638, + 0.306 + ], + "angle": 0, + "content": "\\[\n\\text {f o r} d e g \\text {i n} \\mathbb {F} \\text {d o}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.308, + 0.637, + 0.32 + ], + "angle": 0, + "content": "\\[\nf o r c i n \\mathbb {B} d o\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.323, + 0.743, + 0.337 + ], + "angle": 0, + "content": "\\[\n\\text {M i c r o - b a t c h s i z e} b = B / c;\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.338, + 0.907, + 0.366 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {A}, \\boldsymbol {R}, \\boldsymbol {R} ^ {\\prime}, \\boldsymbol {M} = \\text {C o s t M o d e l i n g} (P R, S D [ d e g ], \\mathcal {G}, \\\\ b); \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.368, + 0.827, + 0.382 + ], + "angle": 0, + "content": "\\[\nc o s t, \\boldsymbol {P}, \\boldsymbol {S} = \\operatorname {M I Q P} (\\boldsymbol {A}, \\boldsymbol {R}, \\boldsymbol {R} ^ {\\prime}, \\boldsymbol {M}, \\deg , c);\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.384, + 0.703, + 0.396 + ], + "angle": 0, + "content": "\\[\n\\text {i f} \\quad c o s t < c o s t ^ {*} \\text {t h e n}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.398, + 0.866, + 0.413 + ], + "angle": 0, + "content": "\\[\nc o s t ^ {*}, d e g ^ {*}, c ^ {*}, P ^ {*}, S ^ {*} = c o s t, d e g, c, P, S;\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.414, + 0.608, + 0.426 + ], + "angle": 0, + "content": "\\[\ne n d i f\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.429, + 0.602, + 0.441 + ], + "angle": 0, + "content": "\\[\ne n d \\mathrm {f o r}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.444, + 0.586, + 0.456 + ], + "angle": 0, + "content": "\\[\ne n d \\mathrm {f o r}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.907, + 0.518 + ], + "angle": 0, + "content": "achieved promising results. UniAP provides a QIP formulation for intra-layer-only parallelism in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.906, + 0.594 + ], + "angle": 0, + "content": "Then, UOP enumerates all factors of \\( n \\) except 1 as the pipeline parallel size \\( deg \\). For each \\( deg \\), UOP enumerates all factors of \\( B \\) except 1 as the number of micro-batches \\( c \\). These enumerations aim to achieve load balance on a homogeneous cluster." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.906, + 0.655 + ], + "angle": 0, + "content": "For each candidate \\( deg \\) and \\( c \\), UOP formulates the cost for a training iteration to an MIQP expression. It then waits for the MIQP solver to return the optimal cost and parallel strategy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.907, + 0.731 + ], + "angle": 0, + "content": "Finally, UOP returns the minimum cost \\( cost^* \\) and its corresponding pipeline parallel size \\( deg^* \\), number of microbatches \\( c^* \\), layer placement \\( P^* \\), and intra-layer strategies \\( S^* \\). We provide visualization for a candidate solution to UOP in Appendix D." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.742, + 0.705, + 0.759 + ], + "angle": 0, + "content": "3.5. Complexity analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Let \\( |\\mathbb{V}| \\), \\( |\\mathbb{S}| \\), and \\( n \\) denote the number of layers, parallel strategies, and GPUs, respectively. As illustrated in Algorithm 1, UniAP enumerates all factors of \\( n \\) except 1 as the \\( deg \\) in the outer loop and all factors of \\( B \\) except 1 in the inner loop. The time complexity for the two loops in the UniAP algorithm is \\( \\mathcal{O}(\\sqrt{Bn}) \\). Within the inner loop body, UniAP calls CostModeling to model the cost of each stage for each parallel strategy. Furthermore, the optimization time limit of the MIQP solver can be set as a constant" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20951" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "hyperparameter when UniAP calls it. Therefore, the overall computational complexity of UniAP is \\(\\mathcal{O}(|\\mathbb{V}||\\mathbb{S}|\\sqrt{Bn})\\)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.138, + 0.216, + 0.155 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.163, + 0.265, + 0.18 + ], + "angle": 0, + "content": "4.1. Experiment setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.185, + 0.486, + 0.366 + ], + "angle": 0, + "content": "We conduct experiments on four different kinds of environments to validate the effectiveness and universality of our method. ENVA refers to a node (machine) with 1 Xeon 6248 CPU, 8 V100-SXM2 32GB GPUs, and 472GB memory. ENVB refers to 4 nodes interconnected with 10Gbps networks and each node has 2 Xeon E5-2620 v4 CPUs, 4 TITAN Xp 12GB GPUs, and 125GB memory. ENVC refers to a node with 8 A100 40GB PCIe GPUs. ENVD is a cloud cluster with 16 nodes and each node has 4 DCUs (a type of non-NVIDIA GPU). In the following text, we specify the number of GPUs after the environment name (e.g., ENVB-8) only when partial nodes of the environment are used." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.367, + 0.485, + 0.519 + ], + "angle": 0, + "content": "We evaluate UniAP with five Transformer-based models, BERT-Huge [5], T5-Large [31], ViT-Huge [6], SwinHuge [24], and Llama [40, 41]. We follow the common practice of training these transformer-based models. To eliminate factors that affect training throughput, we turn off techniques orthogonal to parallel strategies, such as activation checkpointing [3]. However, we integrate FP16 mixed precision training [26] for the largest model, Llama, in some cases to successfully orchestrate the process. More details on these models are provided in Appendix E." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.519, + 0.485, + 0.579 + ], + "angle": 0, + "content": "Several solvers can be used for QIP and MIQP optimization. In our implementation, we choose Gurobi [13], with detailed configurations provided in Appendix E. Exploring other solvers is left for future work." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.485, + 0.687 + ], + "angle": 0, + "content": "The experimental evaluation concentrates on two primary metrics: training throughput and strategy optimization time. The former is calculated by averaging throughput from the 10th to the 60th iteration of training, while the latter is determined by measuring the time of the UOP. More details are provided in Appendix E. Besides, we provide the accuracy of our cost model in Appendix F." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.697, + 0.483, + 0.729 + ], + "angle": 0, + "content": "4.2. Training throughput and strategy optimization time" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.902 + ], + "angle": 0, + "content": "We compare the training throughput and strategy optimization time of UniAP with those of the baselines on ENVA, ENVB-8, ENVC, and ENVD-32. For experiments on ENVA, we set the mini-batch size to be 32, 16, 128, and 128 for BERT, T5, ViT, and Swin, respectively. For experiments on ENVB-8, we set the mini-batch size to be 16, 8, 64, and 32 for these four models, respectively. For the Llama model [40, 41] run on ENVC, we set the mini-batch size to be 8. For experiments on ENVA, ENVB-8, and ENVC, we choose Galvatron [25] and Alma [47] as baselines because they have achieved SOTA performance. Specifically," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.183 + ], + "angle": 0, + "content": "Galvatron has surpassed other methods, including PyTorch DDP [22], Megatron-LM [30], FSDP [8, 32], GPipe [15], and DeepSpeed 3D [27] in terms of training throughput, as reported in the original paper [25]. Furthermore, Alpa utilizes the Just-In-Time (JIT) compilation feature in JAX and outperforms Megatron-LM and DeepSpeed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.184, + 0.909, + 0.35 + ], + "angle": 0, + "content": "For benchmarks on ENVD-32, pytorch libraries and frameworks need specialized adaptations due to the heterogeneity of DCUs against NVIDIA GPUs. Thus, this part of experiments is based on adapted pytorch libraries and Megatron-DeepSpeed[27] framework provided by the platform developers, and we select MP methods including Megatron [30] and DeepSpeed (ZeRO-3) [34] as our baseline methods. We set the mini-batch size as 8 for Llama-7B and 4 for Llama-13B. The remaining configurations align with those for Llama-7B in Section 4.1, such as activating FP16 mixed precision training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.352, + 0.909, + 0.564 + ], + "angle": 0, + "content": "Table 1 shows the training throughput and strategy optimization time on ENVA, ENVB-8, and ENVC. On ENVA, UniAP and Galvatron get the same optimal strategy for BERT-Huge, T5-Large, and ViT-Huge, outperforming Alpa in terms of training throughput and strategy optimization time. In addition, UniAP finds a solution for Swin-Huge, while Galvatron encounters CUDA OOM issues. In particular, UniAP achieves a maximum optimization speedup that is \\(17 \\times\\) faster than Galvatron and hundreds of times faster than Alpa on BERT-Huge. This is mainly due to the ability of the MIQP solver to search for an optimal strategy on multiple threads, while the dynamic programming based methods like Galvatron and Alpa run on a single thread due to their strong data dependency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.909, + 0.656 + ], + "angle": 0, + "content": "On ENVB-8, UniAP consistently demonstrates competitive or larger training throughput compared to Galvatron and Alpa. We attribute the performance improvement to UniAP's larger strategy space. A detailed study is provided in Section 4.4. Furthermore, UniAP's strategy optimization time is also significantly shorter than the two baseline methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.909, + 0.838 + ], + "angle": 0, + "content": "On ENVC, UniAP shows an optimization speedup of \\(11.83 \\times\\) and a training speedup of \\(3.80 \\times\\) compared to Galvatron on Llama. We examine the parallel strategy they adopted. UniAP employs an 8-stage PP with a micro-batch size of 1, whereas Galvatron employs a 4-stage PP with a micro-batch size of 3 (with the final micro-batch containing 2 samples). Within each PP stage, Galvatron utilizes a 2-way TP. For ENVC, which is equipped with 8 A100 40GB PCIe GPUs, minimizing communication volume is critical. Since PP has significantly less inter-device communication volume than TP, the parallel strategy discovered by UniAP is more reasonable than Galvatron in this case." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Table 2 shows the results on ENVD-32. On this environment, UniAP consistently identifies the fastest parallel strategies while expending considerably less time on strategy optimization compared to Megatron (about \\(157 \\times\\) faster for" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20952" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.179, + 0.089, + 0.821, + 0.465 + ], + "angle": 0, + "content": "
Env.ModelTraining throughput (samples/s)Minimum speedupMaximum speedup
GalvatronAlpaUniAP
ENVABERT-Huge33.46 ± 0.2831.56 ± 0.0433.46 ± 0.281.001.06
T5-Large23.29 ± 0.04MEM×2)23.29 ± 0.041.001.00
ViT-Huge109.51 ± 0.0797.66 ± 1.42109.51 ± 0.071.001.12
Swin-HugeCUDA×3)N/Å4)67.96 ± 0.12N/Å4)N/Å4)
ENVB-8BERT-Huge6.27 ± 0.178.95 ± 0.0610.77 ± 0.131.201.71
T5-Large1)8.06 ± 0.06MEM×2)7.98 ± 0.050.990.99
ViT-Huge32.20 ± 0.1738.74 ± 0.2045.58 ± 0.541.181.41
Swin-Huge13.90 ± 0.17N/Å4)19.08 ± 0.101.371.37
ENVCLlama-7B1.22 ± 0.01N/Å4)4.63 ± 0.0073.803.80
Env.ModelStrategy optimization time (min.)Minimum speedupMaximum speedup
GalvatronAlpaUniAP
ENVABERT-Huge6.44 ± 0.588>400.37 ± 0.00217.29>107.41
T5-Large12.41 ± 0.122MEM×2)0.89 ± 0.00713.9813.98
ViT-Huge6.29 ± 0.464>400.57 ± 0.00910.95>69.60
Swin-Huge11.88 ± 0.666N/Å4)2.16 ± 0.0045.495.49
ENVB-8BERT-Huge2.04 ± 0.010>401.51 ± 0.0051.34>26.32
T5-Large1)2.64 ± 0.110MEM×2)0.91 ± 0.0052.902.90
ViT-Huge2.37 ± 0.180>401.11 ± 0.0112.14>36.01
Swin-Huge4.29 ± 0.320N/Å4)2.29 ± 0.0101.871.87
ENVCLlama-7B6.84 ± 0.055N/Å4)0.58 ± 0.00611.8311.83
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.191, + 0.467, + 0.718, + 0.479 + ], + "angle": 0, + "content": "1) T5-Large tested on ENVB is restricted to 16/16 layers to avoid out-of-memory (OOM) exceptions." + }, + { + "type": "table_footnote", + "bbox": [ + 0.191, + 0.479, + 0.49, + 0.491 + ], + "angle": 0, + "content": "2) MEM×: OOM exceptions during strategy optimization." + }, + { + "type": "table_footnote", + "bbox": [ + 0.191, + 0.491, + 0.504, + 0.503 + ], + "angle": 0, + "content": "3) CUDA×: CUDA OOM exceptions during model training." + }, + { + "type": "table_footnote", + "bbox": [ + 0.191, + 0.503, + 0.818, + 0.527 + ], + "angle": 0, + "content": "4) The official implementation for Swin-Huge and Llama in Alpa is absent. We have endeavored to adopt the code, yet encountered compilation errors. Consequently, experiments related to these models on Alpa are marked as N/A." + }, + { + "type": "list", + "bbox": [ + 0.191, + 0.467, + 0.818, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.223, + 0.533, + 0.772, + 0.546 + ], + "angle": 0, + "content": "Table 1. Training throughput and strategy optimization time on ENVA, ENVB-8, and ENVC." + }, + { + "type": "table", + "bbox": [ + 0.179, + 0.561, + 0.82, + 0.639 + ], + "angle": 0, + "content": "
ModelTraining throughput (samples/s)Strategy optimization time (min.)
MegatronDeepSpeedUniAPMegatronDeepSpeedUniAP
Llama-7B2.01 ± 0.005\\( SOL \\times^{1)} \\)2.01 ± 0.005>8.0 hours\\( SOL \\times^{1)} \\)3.07 ± 0.121
Llama-13B0.82 ± 0.001\\( SOL \\times^{1)} \\)0.82 ± 0.001>2.5 hours\\( SOL \\times^{1)} \\)1.95 ± 0.076
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.19, + 0.64, + 0.452, + 0.653 + ], + "angle": 0, + "content": "1) \\( \\mathrm{SOL} \\times \\): No solution after strategy optimization." + }, + { + "type": "table_caption", + "bbox": [ + 0.275, + 0.659, + 0.719, + 0.673 + ], + "angle": 0, + "content": "Table 2. Training throughput and strategy optimization time on ENVD-32." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.681, + 0.486, + 0.864 + ], + "angle": 0, + "content": "Llama-7B and \\(77 \\times\\) faster for Llama-13B). Please note that the strategy optimization time Megatron needs for Llama-7B surpasses that for Llama-13B. We attribute this phenomenon to the variations in the mini-batch size. Specifically, the escalation from a mini-batch size of 4 (employed in Llama-13B) to 8 (employed in Llama-7B) leads to a rise in the number of candidate parallel strategies, as well as the parallel strategies that will successfully train the model. As a result, the strategy optimization time will become increasingly long when the mini-batch size exceeds 8 in the pretraining scenario [1, 40, 41]. In this situation, UniAP will identify the optimal parallel strategy much faster." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Furthermore, our experiments highlight a limitation encountered with DeepSpeed (ZeRO-3). It requires the mini" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.681, + 0.907, + 0.741 + ], + "angle": 0, + "content": "batch size to be divisible evenly by the total number of computation devices. This specific prerequisite prevents DeepSpeed from successfully launching the training process with 32 DCUs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.743, + 0.907, + 0.774 + ], + "angle": 0, + "content": "To facilitate further discussions, we provide a case study of the optimal parallel strategy in Appendix G." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.787, + 0.631, + 0.803 + ], + "angle": 0, + "content": "4.3. Scalability" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We study the scalability of UniAP using BERT, T5, ViT, and Swin on ENVB. We set the mini-batch sizes for each data point as 8, 4, 32, and 16 times the number of nodes (denoted as '#nodes'). The experimental results are shown in Figure 5. In Figure 5a, the training throughput of the optimal strategy demonstrate near-linearity as the number of nodes and mini" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20953" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.162, + 0.089, + 0.837, + 0.168 + ], + "angle": 0, + "content": "
ModelBatch sizeTraining throughput (samples/s)#infeasible5)#candidate6)
1 st.1)2 nd.2)Slowest3)Median4)
Llama-7B82.011.920.220.824164
Llama-13B40.820.580.270.424248
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.173, + 0.169, + 0.553, + 0.181 + ], + "angle": 0, + "content": "1) 1 st.: The training throughput achieved by the fastest parallel strategy." + }, + { + "type": "table_footnote", + "bbox": [ + 0.173, + 0.181, + 0.596, + 0.193 + ], + "angle": 0, + "content": "2) 2 nd.: The training throughput achieved by the second fastest parallel strategy." + }, + { + "type": "table_footnote", + "bbox": [ + 0.174, + 0.193, + 0.578, + 0.205 + ], + "angle": 0, + "content": "3) Slowest: The training throughput achieved by the slowest parallel strategy." + }, + { + "type": "table_footnote", + "bbox": [ + 0.174, + 0.205, + 0.783, + 0.216 + ], + "angle": 0, + "content": "4) Median: The median value of training throughputs across parallel strategies that will successfully train the model." + }, + { + "type": "table_footnote", + "bbox": [ + 0.174, + 0.216, + 0.819, + 0.228 + ], + "angle": 0, + "content": "5) #infeasible: The number of parallel strategies that will encounter exceptions such as CUDA OOM during model training." + }, + { + "type": "table_footnote", + "bbox": [ + 0.174, + 0.228, + 0.507, + 0.24 + ], + "angle": 0, + "content": "6) #candidate: The total number of candidate parallel strategies." + }, + { + "type": "list", + "bbox": [ + 0.173, + 0.169, + 0.819, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.297, + 0.246, + 0.698, + 0.261 + ], + "angle": 0, + "content": "Table 3. Statistics on the candidate parallel strategies for Megatron." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.269, + 0.282, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.355, + 0.255, + 0.368 + ], + "angle": 0, + "content": "(a) Training throughput." + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.269, + 0.479, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.303, + 0.355, + 0.465, + 0.368 + ], + "angle": 0, + "content": "(b) Strategy optimization time." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.371, + 0.486, + 0.398 + ], + "angle": 0, + "content": "Figure 5. Scalability on training throughput and strategy optimization time." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.419, + 0.475, + 0.505 + ], + "angle": 0, + "content": "
Number of DCUsTraining throughput (samples/s)Strategy optimization time(min.)
160.1726 ± 0.00050.75 ± 0.0100
320.3393 ± 0.00041.64 ± 0.0321
640.6656 ± 0.00231.55 ± 0.0058
" + }, + { + "type": "table_caption", + "bbox": [ + 0.094, + 0.514, + 0.479, + 0.529 + ], + "angle": 0, + "content": "Table 4. Scalability on training throughput among DCU clusters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.546, + 0.483, + 0.576 + ], + "angle": 0, + "content": "batch size increase. In Figure 5b, the strategy optimization time matches the complexity analysis in Section 3.5." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.578, + 0.484, + 0.683 + ], + "angle": 0, + "content": "Additionally, we use DCU nodes in ENVD to scale UniAP to larger clusters on Llama-7B, with precision of FP32. We conducted experiments on clusters with 16, 32, and 64 DCUs, where we set the mini-batch size to \\(1/8\\) of the number of DCUs (i.e., for a cluster with 16 DCUs, we set the minibatch size to 2). The experimental results are shown in Table 4, further demonstrating the near-linear scalability of UniAP." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.697, + 0.456, + 0.713 + ], + "angle": 0, + "content": "4.4. Further study on the significance of UniAP" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Finally, we investigate the significance of UniAP by answering two questions. The first question is why we need AP rather than MP. We examine the statistics on the candidate parallel strategies for Megatron, as shown in Table 3. With MP, most users are unfamiliar with the complex hyperparameter detail of the candidate parallel strategies and typically choose a parallel strategy at random, which will result in a \\(64.1\\%\\) (41/64) chance of failing to train Llama-7B and an \\(87.5\\%\\) (42/48) chance of failing to train Llama-13B even when the hardware resources are sufficient. Even if they successfully eliminate all infeasible parallel strategies, they still only have a \\(50\\%\\) chance of identifying a parallel strata" + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.273, + 0.887, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.392, + 0.907, + 0.435 + ], + "angle": 0, + "content": "Figure 6. Ablation study. \\(\\mathrm{SOL} \\times\\) indicates no feasible solution during strategy optimization, and \\(\\mathrm{CUDA} \\times\\) indicates CUDA OOM exceptions during training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.445, + 0.907, + 0.551 + ], + "angle": 0, + "content": "egy with training throughput better than 0.82 samples/s for Llama-7B and 0.42 samples/s for Llama-13B. When the users can't identify the fastest parallel strategy out of hundreds of candidates, they may sacrifice at least \\(4.4\\%\\) of the training throughput for Llama-7B and \\(29.2\\%\\) for Llama-13B. In such circumstances, UniAP is able to identify the fastest strategy, as shown in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.909, + 0.75 + ], + "angle": 0, + "content": "The second question is why UniAP achieves better performance than other AP methods. We investigate the importance of the strategy space for the optimality of parallel strategies with an ablation study. Specifically, we constrain the strategy space to inter-layer-only and intra-layer-only strategies and evaluate the training throughput of the resulting optimal strategy on ENVB. We set the mini-batch sizes to be 16, 12, 64, and 32, respectively. Results are shown in Figure 6. We can find that constraining the strategy space compromises the optimality of parallel strategies or gets strategies that encounter OOM across different models. Hence, unifying inter- and intra-layer AP for joint optimization makes UniAP outperform other AP methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.768, + 0.634, + 0.784 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a novel AP method called UniAP to unify inter- and intra-layer AP by MIQP. To the best of our knowledge, UniAP is the first parallel method that can jointly optimize the two categories of parallel strategies to find an optimal solution. Experimental results show that UniAP can outperform other state-of-the-art baselines to achieve the best performance." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20954" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.271, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.173 + ], + "angle": 0, + "content": "This work is supported by National Key R&D Program of China (No.2020YFA0713900), NSFC Project (No.12326615), the Key Major Project of the Pengcheng Laboratory (No.PCL2024A06)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.201, + 0.188, + 0.217 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.226, + 0.486, + 0.379 + ], + "angle": 0, + "content": "[1] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language Models are Few-Shot Learners. In Advances in Neural Information Processing Systems 33, pages 1877-1901, 2020. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.381, + 0.485, + 0.45 + ], + "angle": 0, + "content": "[2] Zhenkun Cai, Xiao Yan, Kaihao Ma, Yidi Wu, Yuzhen Huang, James Cheng, Teng Su, and Fan Yu. TensorOpt: Exploring the Tradeoffs in Distributed DNN Training With AutoParallelism. IEEE Transactions on Parallel and Distributed Systems, 33(8):1967-1981, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.452, + 0.484, + 0.493 + ], + "angle": 0, + "content": "[3] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training Deep Nets with Sublinear Memory Cost. CoRR, abs/1604.06174, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.496, + 0.487, + 0.786 + ], + "angle": 0, + "content": "[4] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. PaLM: Scaling Language Modeling with Pathways. Journal of Machine Learning Research, 24:240:1-240:113, 2023. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.788, + 0.485, + 0.87 + ], + "angle": 0, + "content": "[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In *The North American Chapter of the Association for Computational Linguistics: Human Language Technologies*, pages 4171–4186, 2019. 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.873, + 0.485, + 0.901 + ], + "angle": 0, + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.226, + 0.487, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.162 + ], + "angle": 0, + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In International Conference on Learning Representations, 2021. 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.164, + 0.909, + 0.302 + ], + "angle": 0, + "content": "[7] Nan Du, Yanping Huang, Andrew M. Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yanqi Zhou, Adams Wei Yu, Orhan Firat, Barret Zoph, Liam Fedus, Maarten P. Bosma, Zongwei Zhou, Tao Wang, Yu Emma Wang, Kellie Webster, Marie Pellat, Kevin Robinson, Kathleen S. Meier-Hellstern, Toju Duke, Lucas Dixon, Kun Zhang, Quoc V. Le, Yonghui Wu, Zhifeng Chen, and Claire Cui. GLaM: Efficient Scaling of Language Models with Mixture- of-Experts. In International Conference on Machine Learning, pages 5547-5569, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.304, + 0.908, + 0.357 + ], + "angle": 0, + "content": "[8] FairScale authors. FairScale: A general purpose modular PyTorch library for high performance and large scale training. https://github.com/facebookresearch/fairscale, 2021. 1, 2, 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.361, + 0.907, + 0.444 + ], + "angle": 0, + "content": "[9] Shiqing Fan, Yi Rong, Chen Meng, Zongyan Cao, Siyu Wang, Zhen Zheng, Chuan Wu, Guoping Long, Jun Yang, Lixue Xia, Lansong Diao, Xiaoyong Liu, and Wei Lin. DAPPLE: A Pipelined Data Parallel Approach for Training Large Models. In Symposium on Principles and Practice of Parallel Programming, pages 431-445, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.446, + 0.907, + 0.5 + ], + "angle": 0, + "content": "[10] William Fedus, Barret Zoph, and Noam Shazeer. Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity. The Journal of Machine Learning Research, 23(120):5232-5270, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.502, + 0.907, + 0.53 + ], + "angle": 0, + "content": "[11] Michael J. Flynn. Very High-speed Computing Systems. Proceedings of the IEEE, 54(12):1901-1909, 1966. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.532, + 0.906, + 0.572 + ], + "angle": 0, + "content": "[12] Michael J. Flynn. Some Computer Organizations and Their Effectiveness. IEEE Transactions on Computers, C-21(9): 948–960, 1972. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.575, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[13] Gurobi Optimization, LLC. Gurobi Optimizer Reference Manual. https://www.gurobi.com, 2023. 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.907, + 0.671 + ], + "angle": 0, + "content": "[14] Chaoyang He, Shen Li, Mahdi Soltanolkotabi, and Salman Avestimehr. PipeTransformer: Automated Elastic Pipelining for Distributed Training of Large-scale Models. In International Conference on Machine Learning, pages 4150-4159, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.908, + 0.758 + ], + "angle": 0, + "content": "[15] Yanping Huang, Youlong Cheng, Ankur Bapna, Orhan First, Dehao Chen, Mia Xu Chen, HyoukJoong Lee, Jiquan Ngiam, Quoc V. Le, Yonghui Wu, and Zhifeng Chen. GPipe: Efficient Training of Giant Neural Networks using Pipeline Parallelism. In Advances in Neural Information Processing Systems 32, pages 103-112, 2019. 1, 2, 4, 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.76, + 0.908, + 0.814 + ], + "angle": 0, + "content": "[16] Zhihao Jia, Sina Lin, Charles R. Qi, and Alex Aiken. Exploring Hidden Dimensions in Parallelizing Convolutional Neural Networks. In International Conference on Machine Learning, pages 2279-2288, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.817, + 0.907, + 0.858 + ], + "angle": 0, + "content": "[17] Zhihao Jia, Matei Zaharia, and Alex Aiken. Beyond Data and Model Parallelism for Deep Neural Networks. In Machine Learning and Systems, pages 1-13, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[18] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. In International Conference on Learning Representations, 2015. 14" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.909, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20955" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.092, + 0.486, + 0.12 + ], + "angle": 0, + "content": "[19] Rafael Lazimy. Mixed-integer Quadratic Programming. Math. Program., 22(1):332-349, 1982. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.122, + 0.486, + 0.205 + ], + "angle": 0, + "content": "[20] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. GShard: Scaling Giant Models with Conditional Computation and Automatic Sharding. In International Conference on Learning Representations, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.485, + 0.275 + ], + "angle": 0, + "content": "[21] Shigang Li and Torsten Hoefler. Chimera: Efficiently Training Large-Scale Neural Networks with Bidirectional Pipelines. In International Conference for High Performance Computing, Networking, Storage and Analysis, pages 27:1-27:14, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.486, + 0.36 + ], + "angle": 0, + "content": "[22] Shen Li, Yanli Zhao, Rohan Varma, Omkar Salpekar, Pieter Noordhuis, Teng Li, Adam Paszke, Jeff Smith, Brian Vaughan, Pritam Damania, and Soumith Chintala. PyTorch Distributed: Experiences on Accelerating Data Parallel Training. Proceedings of the VLDB Endowment, 13(12):3005-3018, 2020. 1, 2, 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.362, + 0.485, + 0.417 + ], + "angle": 0, + "content": "[23] Yuliang Liu, Shenggui Li, Jiarui Fang, Yanjun Shao, Boyuan Yao, and Yang You. Colossal-Auto: Unified Automation of Parallelization and Activation Checkpoint for Large-scale Models. CoRR, abs/2302.02599, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.486, + 0.488 + ], + "angle": 0, + "content": "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. In International Conference on Computer Vision, pages 9992-10002, 2021. 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.486, + 0.559 + ], + "angle": 0, + "content": "[25] Xupeng Miao, Yujie Wang, Youhe Jiang, Chunan Shi, Xiaonan Nie, Hailin Zhang, and Bin Cui. Galvatron: Efficient Transformer Training over Multiple GPUs Using Automatic Parallelism. Proceedings of the VLDB Endowment, 16(3): 470-479, 2022. 2, 3, 6, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.486, + 0.631 + ], + "angle": 0, + "content": "[26] Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory F. Diamos, Erich Elsen, David García, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, and Hao Wu. Mixed Precision Training. In International Conference on Learning Representations, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.486, + 0.66 + ], + "angle": 0, + "content": "[27] Microsoft. Megatron-DeepSpeed. https://github.com/microsoft/Megatron-DeepSpeed/tree/main, 2021.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.662, + 0.485, + 0.732 + ], + "angle": 0, + "content": "[28] Deepak Narayanan, Aaron Harlap, Amar Phanishayee, Vivek Seshadri, Nikhil R. Devanur, Gregory R. Ganger, Phillip B. Gibbons, and Matei Zaharia. PipeDream: Generalized Pipeline Parallelism for DNN Training. In Symposium on Operating Systems Principles, pages 1-15, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.485, + 0.789 + ], + "angle": 0, + "content": "[29] Deepak Narayanan, Amar Phanishayee, Kaiyu Shi, Xie Chen, and Matei Zaharia. Memory-Efficient Pipeline-Parallel DNN Training. In International Conference on Machine Learning, pages 7937–7947, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.486, + 0.901 + ], + "angle": 0, + "content": "[30] Deepak Narayanan, Mohammad Shoeybi, Jared Casper, Patrick LeGresley, Mostofa Patwary, Vijay Korthikanti, Dmitri Vainbrand, Prethvi Kashinkunti, Julie Bernauer, Bryan Catanzaro, Amar Phanishayee, and Matei Zaharia. Efficient Large-scale Language Model Training on GPU Clusters Using Megatron-LM. In International Conference for High Performance Computing, Networking, Storage and Analysis, pages 58:1-58:15, 2021. 1, 2, 3, 6, 14" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.907, + 0.162 + ], + "angle": 0, + "content": "[31] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.908, + 0.234 + ], + "angle": 0, + "content": "[32] Samyam Rajbhandari, Jeff Rasley, Olatunjri Ruwase, and Yuxiong He. ZeRO: Memory Optimizations Toward Training Trillion Parameter Models. In International Conference for High Performance Computing, Networking, Storage and Analysis, pages 20:1-20:16, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.909, + 0.317 + ], + "angle": 0, + "content": "[33] Saeed Rashidi, Matthew Denton, Srinivas Sridharan, Sudarshan Srinivasan, Amoghavarsha Suresh, Jade Nie, and Tushar Krishna. Enabling Compute-Communication Overlap in Distributed Deep Learning Training Platforms. In International Symposium on Computer Architecture, pages 540–553, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.908, + 0.389 + ], + "angle": 0, + "content": "[34] Jeff Rasley, Samyam Rajbhandari, Olatunjri Ruwase, and Yuxiong He. DeepSpeed: System Optimizations Enable Training Deep Learning Models with Over 100 Billion Parameters. In The 26th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 3505-3506, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.908, + 0.473 + ], + "angle": 0, + "content": "[35] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision, 115(3):211-252, 2015. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.476, + 0.908, + 0.558 + ], + "angle": 0, + "content": "[36] Michael Schaarschmidt, Dominik Grewe, Dimitrios Vytiniotis, Adam Paszke, Georg Stefan Schmid, Tamara Norman, James Molloy, Jonathan Godwin, Norman Alexander Rink, Vinod Nair, and Dan Belov. Automap: Towards Ergonomic Automated Parallelism for ML Models. CoRR, abs/2112.02958, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.561, + 0.908, + 0.644 + ], + "angle": 0, + "content": "[37] Noam Shazeer, Youlong Cheng, Niki Parmar, Dustin Tran, Ashish Vaswani, Penporn Koanantakool, Peter Hawkins, HyoukJoong Lee, Mingsheng Hong, Cliff Young, Ryan Sepassi, and Blake A. Hechtman. Mesh-TensorFlow: Deep Learning for Supercomputers. In Advances in Neural Information Processing Systems 31, pages 10435-10444, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.908, + 0.714 + ], + "angle": 0, + "content": "[38] Jakub Tarnawski, Amar Phanishayee, Nikhil R. Devanur, Divya Mahajan, and Fanny Nina Paravecino. Efficient Algorithms for Device Placement of DNN Graph Operators. In Advances in Neural Information Processing Systems 33, pages 15451-15463, 2020. 1, 2, 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.717, + 0.908, + 0.773 + ], + "angle": 0, + "content": "[39] Jakub Tarnawski, Deepak Narayanan, and Amar Phanishayee. Piper: Multidimensional Planner for DNN Parallelization. In Advances in Neural Information Processing Systems 34, pages 24829-24840, 2021. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.908, + 0.857 + ], + "angle": 0, + "content": "[40] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurélien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. LLaMA: Open and Efficient Foundation Language Models. CoRR, abs/2302.13971, 2023. 1, 6, 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[41] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.909, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20956" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.486, + 0.327 + ], + "angle": 0, + "content": "Blecher, Cristian Canton-Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, MarieAnne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open Foundation and Fine-Tuned Chat Models. CoRR, abs/2307.09288, 2023. 1, 6, 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.329, + 0.486, + 0.384 + ], + "angle": 0, + "content": "[42] Minjie Wang, Chien-Chin Huang, and Jinyang Li. Supporting Very Large Models using Automatic Dataflow Graph Partitioning. In Proceedings of the Fourteenth EuroSys Conference, pages 26:1-26:17, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.485, + 0.413 + ], + "angle": 0, + "content": "[43] Wikipedia Foundation. Wikipedia Downloads. https:// dumps.wikipedia.org, 2023. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.415, + 0.486, + 0.51 + ], + "angle": 0, + "content": "[44] Yuanzhong Xu, HyoukJoong Lee, Dehao Chen, Blake A. Hechtman, Yanping Huang, Rahul Joshi, Maxim Krikun, Dmitry Lepikhin, Andy Ly, Marcello Maggioni, Ruoming Pang, Noam Shazeer, Shibo Wang, Tao Wang, Yonghui Wu, and Zhifeng Chen. GSPMD: General and Scalable Parallelization for ML Computation Graphs. CoRR, abs/2105.04663, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.513, + 0.486, + 0.581 + ], + "angle": 0, + "content": "[45] Binhang Yuan, Yongjun He, Jared Davis, Tianyi Zhang, Tri Dao, Beidi Chen, Percy Liang, Christopher Ré, and Ce Zhang. Decentralized Training of Foundation Models in Heterogeneous Environments. In Advances in Neural Information Processing Systems 35, pages 25464-25477, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.583, + 0.486, + 0.679 + ], + "angle": 0, + "content": "[46] Shixiong Zhao, Fanxin Li, Xusheng Chen, Xiuxian Guan, Jianyu Jiang, Dong Huang, Yuhao Qing, Sen Wang, Peng Wang, Gong Zhang, Cheng Li, Ping Luo, and Heming Cui. vPipe: A Virtualized Acceleration System for Achieving Efficient and Scalable Pipeline Parallel DNN Training. IEEE Transactions on Parallel and Distributed Systems, 33(3):489-506, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.681, + 0.486, + 0.777 + ], + "angle": 0, + "content": "[47] Lianmin Zheng, Zhuohan Li, Hao Zhang, Yonghao Zhuang, Zhifeng Chen, Yanping Huang, Yida Wang, Yuanzhong Xu, Danyang Zhuo, Eric P. Xing, Joseph E. Gonzalez, and Ion Stoica. Alpa: Automating Inter- and Intra-Operator Parallelism for Distributed Deep Learning. In USENIX Symposium on Operating Systems Design and Implementation, pages 559–578, 2022. 1, 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.777 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20957" + } + ] +] \ No newline at end of file diff --git a/2025/UniAP_ Unifying Inter- and Intra-Layer Automatic Parallelism by Mixed Integer Quadratic Programming/b1ddd4d6-5b56-44ba-82ed-6323538fa782_origin.pdf b/2025/UniAP_ Unifying Inter- and Intra-Layer Automatic Parallelism by Mixed Integer Quadratic Programming/b1ddd4d6-5b56-44ba-82ed-6323538fa782_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a7ad841b917181053a68c89cdadc2b247040430e --- /dev/null +++ b/2025/UniAP_ Unifying Inter- and Intra-Layer Automatic Parallelism by Mixed Integer Quadratic Programming/b1ddd4d6-5b56-44ba-82ed-6323538fa782_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6741fe72af70f0fe559a6ba44bb97f90e0c1e5cb08e63a218a17f608e0589f34 +size 657658 diff --git a/2025/UniAP_ Unifying Inter- and Intra-Layer Automatic Parallelism by Mixed Integer Quadratic Programming/full.md b/2025/UniAP_ Unifying Inter- and Intra-Layer Automatic Parallelism by Mixed Integer Quadratic Programming/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d09fe706d32e173a5283bf2f3b925a9ababd3930 --- /dev/null +++ b/2025/UniAP_ Unifying Inter- and Intra-Layer Automatic Parallelism by Mixed Integer Quadratic Programming/full.md @@ -0,0 +1,436 @@ +# UniAP: Unifying Inter- and Intra-Layer Automatic Parallelism by Mixed Integer Quadratic Programming + +Hao Lin\*, Ke Wu\*, Jie Li\*, Jun Li, Wu-Jun Li† +National Key Laboratory for Novel Software Technology School of Computer Science +Nanjing University, Nanjing 210023, China + +{hao.lin, ke.wu, jie-li, lijun}@smail.nju.edu.cn, liwujun@nju.edu.cn + +# Abstract + +Distributed learning is commonly used for training deep learning models, especially large models. In distributed learning, manual parallelism (MP) methods demand considerable human effort and have limited flexibility. Hence, automatic parallelism (AP) methods have recently been proposed for automating the parallel strategy optimization process. Existing AP methods suffer from sub-optimal solutions because they do not jointly optimize the two categories of parallel strategies (i.e., inter-layer parallelism and intra-layer parallelism). In this paper, we propose a novel AP method called UniAP, which unifies inter- and intra-layer automatic parallelism by mixed integer quadratic programming. To the best of our knowledge, UniAP is the first parallel method that can jointly optimize the two categories of parallel strategies to find an optimal solution. Experimental results show that UniAP outperforms state-of-the-art methods by up to $3.80 \times$ in throughput and reduces strategy optimization time by up to $107 \times$ across five Transformer-based models. + +# 1. Introduction + +Distributed learning (also called parallel learning) on clusters with several machines or GPUs is commonly used for training deep learning models, especially for some large models with billions of parameters [1, 40, 41]. Several parallel strategies, including pipeline parallelism (PP), data parallelism (DP), tensor parallelism (TP), and fully sharded data parallelism (FSDP), have been proposed for distributed learning. These parallel strategies can be divided into two main categories: inter-layer parallelism and intra-layer parallelism. Inter-layer parallelism [7, 9, 10, 15, 20, 21, 28, 29], which includes PP, partitions the model into disjoint sets + +without partitioning tensors in each layer. Intra-layer parallelism [8, 22, 30, 34], which includes DP, TP, and FSDP, partitions tensors in a layer along one or more axes. + +The parallel method1 in one specific distributed learning method or system typically adopts one parallel strategy or a combination of several parallel strategies. Existing parallel methods can be divided into two categories: manual parallelism (MP) methods and automatic parallelism (AP) methods. In MP methods [30, 37, 44], one or several parallel strategies are manually optimized by researchers or developers. MP methods require extensive domain knowledge in deep learning models and hardware architectures. With the rapid development of deep learning models and the increasing diversity of modern hardware architectures [11, 12], MP methods demand considerable human effort and have limited flexibility. + +To address the two limitations of MP methods, AP methods [14, 28, 47] have recently been proposed for automating the parallel strategy optimization process. Although existing AP methods have achieved promising progress, they optimize the two categories of parallel strategies separately rather than jointly. More specifically, some methods optimize only one category of parallel strategies [2, 16, 17, 23, 36, 42, 46], and the others optimize inter- and intra-layer parallelism hierarchically [9, 14, 28, 38, 39, 47]. Hence, existing AP methods suffer from sub-optimal solutions. + +In this paper, we propose a novel AP method called UniAP for distributed learning. The contributions of UniAP are outlined as follows: + +- UniAP unifies inter- and intra-layer automatic parallelism by mixed integer quadratic programming (MIQP) [19]. +- To the best of our knowledge, UniAP is the first parallel method that can optimize the two categories of parallel strategies jointly rather than separately to find an optimal solution. + +- Experimental results show that UniAP outperforms state-of-the-art methods by up to $3.80 \times$ in throughput and reduces strategy optimization time by up to $107 \times$ across five Transformer-based models. + +# 2. Background + +# 2.1. Parallel strategy + +Pipeline parallelism (PP) In PP [15], each worker (machine or GPU) holds a subset of model layers. Adjacent layers on different workers need to transfer activations in the forward propagation (FP) step and gradients in the backward propagation (BP) step. + +Data parallelism (DP) In DP [22], each worker holds a replica of the whole model and partitions training samples. In each iteration, each worker computes gradients and synchronizes them with the other workers using all-reduce collective communication (CC). All workers will have the same model parameters after the synchronization step. + +Tensor parallelism (TP) In TP [30], each worker holds a replica of training samples and partitions within model layers. In each iteration, each worker computes its local outputs in FP and its local gradients in BP. To synchronize outputs and gradients, all workers will perform all-reduce CC in FP and BP steps according to the partition scheme. + +Fully sharded data parallelism (FSDP) FSDP [8, 32] partitions optimizer states, parameters and gradients of the model into separate workers. During the FP and BP step of each iteration, FSDP performs an all-gather CC to obtain the complete parameters for the relevant layer, respectively. After computing gradients, FSDP conducts a reduce-scatter CC to distribute the global gradients among the workers. + +# 2.2. Parallel method + +Manual parallelism (MP) MP refers to the parallel methods in which human experts design and optimize the parallel strategies. Representative MP methods include Megatron-LM [30], Mesh-TensorFlow [37], and GSPMD [44]. Megatron-LM manually designs TP and PP strategies for training Transformer-based models and exhibits superior efficiency. Mesh-TensorFlow and GSPMD require human effort to designate and tune the intra-layer parallel strategy. These methods rely on expert design and have little flexibility, challenging their automatic application to other models. + +Inter-layer-only AP or intra-layer-only AP For interlayer-only AP, GPipe [15] and vPipe [46] employ a balanced partition algorithm and a dynamic layer partitioning middleware to partition pipelines, respectively. For intra-layer-only AP, OptCNN [16], TensorOpt [2], and Tofu [42] employ dynamic programming methods to optimize DP and TP strategies together. FlexFlow [17] and Automap [36] use the Monte Carlo method to find the optimal DP and TP strat + +![](images/daccd7bd1041350ba01fa7ba9cd1e1eaecb29d6e64a83d9930b70150b159382a.jpg) +Figure 1. Parallel methods for optimizing parallel strategies for a three-layer model. The different arrangements of slices with varying transparency within the same layer block indicate different intra-layer parallelism strategies adopted by layers. The different arrangements of gray blocks which wrap the layer blocks indicate different inter-layer parallelism strategies. + +egy. Colossal-Auto [23] utilizes integer programming (IP) techniques to generate intra-layer parallelism and activation checkpointing strategies. All these methods optimize only one category of parallel strategies. + +Inter- and intra-layer AP PipeDream [28], DAPPLE [9], and PipeTransformer [14] use dynamic programming to determine optimal strategies for both DP and PP. DNN-partitioning [38] adopts IP and dynamic programming to explore DP and PP strategies. DT-FM [45] combines genetic algorithm and the local search strategy to explore DP and PP strategies in a layer-wise manner. Piper [39] and Alpa [47] adopt a parallel method considering DP, TP, and PP. Galvatron [25] uses dynamic programming to determine DP, TP, and FSDP strategies in a single pipeline stage. As for PP, it partitions stages and determines micro-batch size using naive greedy algorithms. All these methods are hierarchical, which will result in sub-optimal solutions. + +# 3. Method + +In this section, we introduce our proposed method called UniAP, which jointly optimizes the two categories of parallel strategies, including PP, DP, TP, and FSDP, to find an optimal solution. Figure 1 illustrates the difference between UniAP and other automatic parallelism methods. Inter-layer-only and intra-layer-only AP methods optimize (search) from a set of candidate inter-layer-only and intra-layer-only parallel strategies, respectively. Hierarchical AP methods first adopt greedy or dynamic programming to propose candidate inter-layer parallel strategies. Then, they optimize the intra-layer parallel strategy for every fixed inter-layer parallel strategy. + +![](images/e9250057b878b58ef36dc760bdaa16ab25a6caa0041ff5f300fe1e13b4c9702e.jpg) +Figure 2. Flowchart of UniAP. + +UniAP has the largest strategy space for exploration (joint optimization). + +Figure 2 illustrates the flowchart of UniAP. UniAP first profiles the runtime information for the user's hardware environment and the deep learning model. After that, UniAP estimates inter- and intra-layer costs given the computation graph and profiling results with its cost models. The estimated costs and the computation graph are then transformed into an MIQP problem. The objective function of the MIQP is to maximize the training throughput, or in other words, to minimize the training time per iteration (TPI). By iteratively applying the cost model and MIQP with different parameters, UniAP determines the minimal TPI and its corresponding parallel strategies. We name this process the Unified Optimization Process (UOP). Finally, UniAP interprets the parallel strategies into the execution plan for the designated model. + +# 3.1. Profiling + +UniAP collects runtime information about the hardware environment and deep learning model during profiling. For the hardware environment, UniAP evaluates the efficiency of all-reduce and point-to-point (P2P) communication for different device subsets. For example, when profiling a node with 4 GPUs, UniAP measures the all-reduce efficiency for various DP, TP, and FSDP combinations across these GPUs. Additionally, UniAP ranks these GPUs from 0 to 3 and evaluates the speed of P2P for two pipeline options: $(0\to 2$ and $1\rightarrow 3)$ and $(0\to 1,1\to 2$ and $2\to 3)$ . Furthermore, UniAP estimates the computation-communication overlap coefficient (CCOC) [25, 33]. + +UniAP acquires two types of information for the deep learning model: computation time and memory usage. On one hand, UniAP distinguishes the forward computation time per sample for different types of hidden layers. On the other hand, UniAP collects memory usage information for each layer, including the memory occupied by parameters and the memory usage of activation per sample in different TP sizes. + +# 3.2. Cost model + +UniAP employs two primary cost models, namely the time cost model and the memory cost model. + +Time cost model To estimate computation time, UniAP first calculates the forward computation time by multiplying the batch size with the forward computation time per sample obtained from profiling. Users can obtain a more precise result by profiling the forward time on the specified batch size. For Transformer-based models that mainly consist of the MatMul operator, the computation time in the BP stages is roughly twice that of the FP stages [21, 25, 30]. Additionally, UniAP estimates the communication time by dividing the size of transmitting tensors by the profiled communication efficiency for different communication primitives. To accommodate overlapping, UniAP multiplies the profiled CCOC by the overlapping interval of computation and communication. To model the communication time between pipeline stages, UniAP calculates the cross-stage cost between consecutive stages by the summation of P2P costs. + +Memory cost model UniAP estimates memory consumption for each layer with its memory cost model. This estimation consists of three steps for a given layer. First, it computes the activation memory cost $m_{a}$ by multiplying the batch size and the profiled activation memory cost per sample of the TP size used by the strategy. Next, UniAP calculates the memory cost of model states $m_{s}$ for each layer based on their parameter size $ps$ , TP size $ts$ , FSDP size $fs$ , and a constant $c_{dtype}$ dependent on the data type. Formally, we have + +$$ +m _ {s} = \frac {c _ {d t y p e} \times p s}{t s \times f s}. \tag {1} +$$ + +For example, if we choose precision as FP32, then $c_{dtype} = (4 + 4 + 4 + 4) / 4 = 4$ since the learnable parameters, gradients, momentum and variance consume equal size of memory. If we opt for mixed precision with FP16 activated, then $c_{dtype} = (4 + 4 + 4 + 2 + 2) / 2 = 8$ . Finally, UniAP aggregates the activation memory cost $m_a$ , memory cost of model states $m_s$ , and context memory cost $m_c$ to a constant matrix $M$ , where $M_{uk}$ denotes the memory cost for the $k$ -th intra-layer strategy of layer $u$ on a single device. + +# 3.3. Mixed integer quadratic programming + +The estimated costs and the computation graph are then transformed into an MIQP problem. Its formulation includes an objective function and several constraints. + +# 3.3.1. Objective function + +The objective function tries to minimize TPI. In this paper, we have chosen GPipe as our PP strategy for illustration.2 Figure 3 depicts the time cost decomposition of a GPipe-style PP with non-negligible communication costs. The time needed to apply gradients at the end of each iteration is not included, as it depends on the optimizer and is insignificant compared to the total time spent on FP and BP. + +![](images/183a476c7327e98af73d2b0f588b882911e8b9011441b9aab9d066f331a2f2ba.jpg) +Figure 3. Time cost decomposition of a GPipe-style PP. + +![](images/03761f9a11a6fb9e0f102a74ba3ed15114a72e5c16d162b70f2a658c9a47ae1c.jpg) +Figure 4. A contiguous set. + +We denote the cost for computation stages as $\mathbb{P} = \{p_1, p_2, \ldots, p_{deg}\}$ and the cost for communication stages as $\mathbb{O} = \{o_1, o_2, \ldots, o_{deg-1}\}$ . Here, $deg$ represents the number of computation stages, which corresponds to the pipeline parallel size. In Figure 3, $fp_i$ and $bp_i$ denote forward and backward computation time for computation stage $i$ , respectively. $fo_j$ and $bo_j$ denote forward and backward communication time for communication stage $j$ , respectively. Hence, we have $pi = fp_i + bp_i$ and $oj = fo_j + bo_j$ . + +In a GPipe-style pipeline, we use $c$ to denote the number of micro-batches. As illustrated in Figure 3, a mini-batch is uniformly split into four micro-batches, and the total TPI is determined by the latency of all computation and communication stages and the latency of the slowest stage. We further denote TPI in GPipe as $tpi_{gpipe}$ . Given that a stage with a higher FP computation cost leads to a higher BP computation cost with high probability, we can write the objective function of GPipe-style pipeline as follows: + +$$ +\begin{array}{l} \min t p i _ {g p i p e} = \sum_ {i = 1} ^ {d e g} p _ {i} + \sum_ {j = 1} ^ {d e g - 1} o _ {j} \tag {2} \\ + (c - 1) \max (\mathbb {P} \cup \mathbb {O}). \\ \end{array} +$$ + +# 3.3.2. Constraint + +We first introduce additional notations before presenting the constraints. For a given layer $u \in \mathbb{V}$ , $\mathbb{S}_u$ represents its set of intra-layer parallel strategies, $A_{uk}$ denotes the $k$ -th intra-layer execution cost obtained from our time cost model. Additionally, we use $S_{uk} \in \{0,1\}$ to indicate whether the $k$ -th parallel strategy is selected for the layer $u$ , and use $P_{ui} \in \{0,1\}$ to indicate whether layer $u$ is placed on the $i$ -th computation stage. Each edge $\langle u,v\rangle \in \mathbb{E}$ is assigned a resharding cost denoted by $R_{uv}$ if the vertices are located within the same pipeline stage. Alternatively, if the vertices are located across consecutive stages, the resharding cost between them is denoted by $R_{uv}'$ . These two resharding costs are constant matrices derived from our time cost model. + +Computation-stage constraint To compute the total cost for a single computation stage $i$ , all computation and communication costs associated with that stage must be aggregated and assigned to $p_i$ . This constraint can be formulated as + +follows: + +$$ +\begin{array}{l} \sum_ {u \in \mathbb {V}} \boldsymbol {P} _ {u i} \boldsymbol {S} _ {u} ^ {\top} \boldsymbol {A} _ {u} + \sum_ {\langle u, v \rangle \in \mathbb {E}} \boldsymbol {P} _ {u i} \boldsymbol {P} _ {v i} \left(\boldsymbol {S} _ {u} ^ {\top} \boldsymbol {R} _ {u v} \boldsymbol {S} _ {v}\right) = p _ {i}, \tag {3} \\ \forall i \in \{1, \dots , d e g \}. \\ \end{array} +$$ + +On the left side of Equation (3), the first polynomial term represents the cost of choosing specific intra-layer strategies for layers placed in stage $i$ . The second term represents total resharding costs within stage $i$ . + +Communication-stage constraint To calculate the total cost for a single communication stage $j$ , we should aggregate the P2P costs incurred between consecutive stages and assign them to $o_j$ . This constraint can be formulated as follows: + +$$ +\begin{array}{l} \sum_ {\langle u, v \rangle \in \mathbb {E}} \boldsymbol {P} _ {u j} \boldsymbol {P} _ {v (j + 1)} \left(\boldsymbol {S} _ {u} ^ {\top} \boldsymbol {R} _ {u v} ^ {\prime} \boldsymbol {S} _ {v}\right) = o _ {j}, \tag {4} \\ \forall j \in \{1, \dots , d e g - 1 \}. \\ \end{array} +$$ + +Memory constraint We need to guarantee that no devices (GPUs) will encounter out-of-memory (OOM) exceptions during training process. This constraint can be formulated as follows: + +$$ +\sum_ {u \in \mathbb {V}} \boldsymbol {P} _ {u i} \boldsymbol {S} _ {u} ^ {\top} \boldsymbol {M} _ {u} \leqslant m, \forall i \in \{1, \dots , d e g \}. \tag {5} +$$ + +Here, $m$ denotes the memory limit for each device. In the case of homogeneous computing devices, the value of $m$ remains constant throughout all stages. But the value of $m$ varies in the case of heterogeneous computing devices. + +Order-preserving constraint PP is not a single-program multiple-data (SPMD) parallel strategy [15]. Hence, we need an order-preserving constraint to ensure that the subgraphs of $\mathcal{G}$ are contiguous. We adopt the definition of contiguous from previous work [38, 39]. + +Definition 3.1. A set $\mathbb{W} \subseteq \mathbb{V}$ is contiguous if there do not exist nodes $u \in \mathbb{W}, v \in \mathbb{V} \setminus \mathbb{W}$ , and $w \in \mathbb{W}$ such that $v$ is reachable from $u$ and $w$ is reachable from $v$ . + +Figure 4 illustrates an example of a contiguous set $\mathbb{W}$ , in which we cannot find any reachable node pairs $\langle u, v \rangle$ and $\langle v, w \rangle$ where $u, w \in \mathbb{W}$ and $v \in \mathbb{V} \setminus \mathbb{W}$ . + +In our case, our model will not be assigned to different pipeline stages in a disordered manner if we ensure that all subgraphs on each computation stage are contiguous. After reformulating this constraint in linear form, we have + +$$ +\mathbf {Z} _ {v i} \geqslant \mathbf {P} _ {v i}, +$$ + +$$ +\forall v \in \mathbb {V}, \forall i \in \{1, 2, \dots , d e g \}, \tag {6a} +$$ + +$$ +\mathbf {Z} _ {v i} \leqslant \mathbf {Z} _ {u i}, +$$ + +$$ +\forall u, v \in \mathbb {V}, \forall \langle u, v \rangle \in \mathbb {E}, \forall i \in \{1, 2, \dots , d e g \}, \tag {6b} +$$ + +$$ +\mathbf {Z} _ {v i} \leqslant \mathbf {P} _ {v i} - \mathbf {P} _ {u i} + 1, +$$ + +$$ +\forall u, v \in \mathbb {V}, \forall \langle u, v \rangle \in \mathbb {E}, \forall i \in \{1, 2, \dots , d e g \}. \tag {6c} +$$ + +Here, $\mathbf{Z}$ is an auxiliary variable shaped like $\mathbf{P}$ . Given the node set $\mathbb{V}_i$ that contains nodes in stage $i$ , we define $\mathbf{Z}_{vi} = 1$ if a node $w \in \mathbb{V}_i$ is reachable from $v$ ( $v \in \mathbb{V}$ ). Otherwise, $\mathbf{Z}_{vi} = 0$ . Detailed proof can be found in Appendix B. + +Layer-placement constraint All layers should be placed on exactly one pipeline stage and at least one layer should be placed on each pipeline stage. This constraint can be formulated as follows: + +$$ +\sum_ {i = 1} ^ {\text {d e g}} \boldsymbol {P} _ {u i} = 1, \quad \forall u \in \mathbb {V}, \tag {7a} +$$ + +$$ +\sum_ {u \in \mathbb {V}} \boldsymbol {P} _ {u i} \geqslant 1, \quad \forall i \in \{1, \dots , d e g \}, \tag {7b} +$$ + +$$ +\boldsymbol {P} _ {u i} \in \{0, 1 \}, \quad \forall u \in \mathbb {V}, i \in \{1, \dots , d e g \}. \tag {7c} +$$ + +Strategy-selection constraint Each layer must and can choose only one strategy. This constraint can be formulated as follows: + +$$ +\sum_ {k = 1} ^ {| \mathbb {S} _ {u} |} \boldsymbol {S} _ {u k} = 1, \quad \forall u \in \mathbb {V}, \tag {8a} +$$ + +$$ +\boldsymbol {S} _ {u k} \in \{0, 1 \}, \quad \forall u \in \mathbb {V}, k \in \{1, \dots , | \mathbb {S} _ {u} | \}. \tag {8b} +$$ + +The MIQP formulation for UniAP includes the objective function in Equation (2) and all the constraints from Euqation (3) - (8b). + +# 3.4. Unified optimization process + +UOP integrates the cost model and MIQP based on the profiling results and the computation graph to return the optimal parallel strategy and the corresponding TPI. Algorithm 1 summarizes the whole process. In Algorithm 1, we denote intra-layer cost as $A$ , inter-layer cost as $R$ , cross-stage cost as $R'$ , and memory cost as $M$ . The CostModeling process calculates these four costs based on the cost model described in Section 3.2. + +First, UOP optimizes intra-layer-only parallelism for cases in which pipeline parallelism is not adopted. Several works [23, 47] have used quadratic integer programming (QIP) to optimize intra-layer-only parallel strategy and + +Algorithm 1 Unified Optimization Process + +Input: Profiling results $PR$ , strategy dictionary $SD$ , minibatch size $B$ , computation graph $\mathcal{G}$ , and the number of GPUs $n$ . + +Output: Optimal cost $cost^{*}$ , pipeline parallel size $deg^{*}$ , the number of micro-batches $c^{*}$ , layer placement $P^{*}$ , and intra-layer strategy $S^{*}$ . + +$$ +d e g ^ {*} = 1; +$$ + +$$ +c ^ {*} = B; +$$ + +$$ +A, R, \_ , M = \operatorname {C o s t M o d e l i n g} (P R, S D [ 1 ], \mathcal {G}, B); +$$ + +$$ +c o s t ^ {*}, P ^ {*}, S ^ {*} = \mathbb {Q} \operatorname {I P} (\boldsymbol {A}, \boldsymbol {R}, \boldsymbol {M}); +$$ + +Get all factors for $n$ except 1 and insert them to set $\mathbb{F}$ ; + +Get all factors for $B$ except 1 and insert them to set $\mathbb{B}$ + +$$ +\text {f o r} d e g \text {i n} \mathbb {F} \text {d o} +$$ + +$$ +f o r c i n \mathbb {B} d o +$$ + +$$ +\text {M i c r o - b a t c h s i z e} b = B / c; +$$ + +$$ +\begin{array}{l} \boldsymbol {A}, \boldsymbol {R}, \boldsymbol {R} ^ {\prime}, \boldsymbol {M} = \text {C o s t M o d e l i n g} (P R, S D [ d e g ], \mathcal {G}, \\ b); \end{array} +$$ + +$$ +c o s t, \boldsymbol {P}, \boldsymbol {S} = \operatorname {M I Q P} (\boldsymbol {A}, \boldsymbol {R}, \boldsymbol {R} ^ {\prime}, \boldsymbol {M}, \deg , c); +$$ + +$$ +\text {i f} \quad c o s t < c o s t ^ {*} \text {t h e n} +$$ + +$$ +c o s t ^ {*}, d e g ^ {*}, c ^ {*}, P ^ {*}, S ^ {*} = c o s t, d e g, c, P, S; +$$ + +$$ +e n d i f +$$ + +$$ +e n d \mathrm {f o r} +$$ + +$$ +e n d \mathrm {f o r} +$$ + +achieved promising results. UniAP provides a QIP formulation for intra-layer-only parallelism in Appendix C. + +Then, UOP enumerates all factors of $n$ except 1 as the pipeline parallel size $deg$ . For each $deg$ , UOP enumerates all factors of $B$ except 1 as the number of micro-batches $c$ . These enumerations aim to achieve load balance on a homogeneous cluster. + +For each candidate $deg$ and $c$ , UOP formulates the cost for a training iteration to an MIQP expression. It then waits for the MIQP solver to return the optimal cost and parallel strategy. + +Finally, UOP returns the minimum cost $cost^*$ and its corresponding pipeline parallel size $deg^*$ , number of microbatches $c^*$ , layer placement $P^*$ , and intra-layer strategies $S^*$ . We provide visualization for a candidate solution to UOP in Appendix D. + +# 3.5. Complexity analysis + +Let $|\mathbb{V}|$ , $|\mathbb{S}|$ , and $n$ denote the number of layers, parallel strategies, and GPUs, respectively. As illustrated in Algorithm 1, UniAP enumerates all factors of $n$ except 1 as the $deg$ in the outer loop and all factors of $B$ except 1 in the inner loop. The time complexity for the two loops in the UniAP algorithm is $\mathcal{O}(\sqrt{Bn})$ . Within the inner loop body, UniAP calls CostModeling to model the cost of each stage for each parallel strategy. Furthermore, the optimization time limit of the MIQP solver can be set as a constant + +hyperparameter when UniAP calls it. Therefore, the overall computational complexity of UniAP is $\mathcal{O}(|\mathbb{V}||\mathbb{S}|\sqrt{Bn})$ + +# 4. Experiment + +# 4.1. Experiment setup + +We conduct experiments on four different kinds of environments to validate the effectiveness and universality of our method. ENVA refers to a node (machine) with 1 Xeon 6248 CPU, 8 V100-SXM2 32GB GPUs, and 472GB memory. ENVB refers to 4 nodes interconnected with 10Gbps networks and each node has 2 Xeon E5-2620 v4 CPUs, 4 TITAN Xp 12GB GPUs, and 125GB memory. ENVC refers to a node with 8 A100 40GB PCIe GPUs. ENVD is a cloud cluster with 16 nodes and each node has 4 DCUs (a type of non-NVIDIA GPU). In the following text, we specify the number of GPUs after the environment name (e.g., ENVB-8) only when partial nodes of the environment are used. + +We evaluate UniAP with five Transformer-based models, BERT-Huge [5], T5-Large [31], ViT-Huge [6], SwinHuge [24], and Llama [40, 41]. We follow the common practice of training these transformer-based models. To eliminate factors that affect training throughput, we turn off techniques orthogonal to parallel strategies, such as activation checkpointing [3]. However, we integrate FP16 mixed precision training [26] for the largest model, Llama, in some cases to successfully orchestrate the process. More details on these models are provided in Appendix E. + +Several solvers can be used for QIP and MIQP optimization. In our implementation, we choose Gurobi [13], with detailed configurations provided in Appendix E. Exploring other solvers is left for future work. + +The experimental evaluation concentrates on two primary metrics: training throughput and strategy optimization time. The former is calculated by averaging throughput from the 10th to the 60th iteration of training, while the latter is determined by measuring the time of the UOP. More details are provided in Appendix E. Besides, we provide the accuracy of our cost model in Appendix F. + +# 4.2. Training throughput and strategy optimization time + +We compare the training throughput and strategy optimization time of UniAP with those of the baselines on ENVA, ENVB-8, ENVC, and ENVD-32. For experiments on ENVA, we set the mini-batch size to be 32, 16, 128, and 128 for BERT, T5, ViT, and Swin, respectively. For experiments on ENVB-8, we set the mini-batch size to be 16, 8, 64, and 32 for these four models, respectively. For the Llama model [40, 41] run on ENVC, we set the mini-batch size to be 8. For experiments on ENVA, ENVB-8, and ENVC, we choose Galvatron [25] and Alma [47] as baselines because they have achieved SOTA performance. Specifically, + +Galvatron has surpassed other methods, including PyTorch DDP [22], Megatron-LM [30], FSDP [8, 32], GPipe [15], and DeepSpeed 3D [27] in terms of training throughput, as reported in the original paper [25]. Furthermore, Alpa utilizes the Just-In-Time (JIT) compilation feature in JAX and outperforms Megatron-LM and DeepSpeed. + +For benchmarks on ENVD-32, pytorch libraries and frameworks need specialized adaptations due to the heterogeneity of DCUs against NVIDIA GPUs. Thus, this part of experiments is based on adapted pytorch libraries and Megatron-DeepSpeed[27] framework provided by the platform developers, and we select MP methods including Megatron [30] and DeepSpeed (ZeRO-3) [34] as our baseline methods. We set the mini-batch size as 8 for Llama-7B and 4 for Llama-13B. The remaining configurations align with those for Llama-7B in Section 4.1, such as activating FP16 mixed precision training. + +Table 1 shows the training throughput and strategy optimization time on ENVA, ENVB-8, and ENVC. On ENVA, UniAP and Galvatron get the same optimal strategy for BERT-Huge, T5-Large, and ViT-Huge, outperforming Alpa in terms of training throughput and strategy optimization time. In addition, UniAP finds a solution for Swin-Huge, while Galvatron encounters CUDA OOM issues. In particular, UniAP achieves a maximum optimization speedup that is $17 \times$ faster than Galvatron and hundreds of times faster than Alpa on BERT-Huge. This is mainly due to the ability of the MIQP solver to search for an optimal strategy on multiple threads, while the dynamic programming based methods like Galvatron and Alpa run on a single thread due to their strong data dependency. + +On ENVB-8, UniAP consistently demonstrates competitive or larger training throughput compared to Galvatron and Alpa. We attribute the performance improvement to UniAP's larger strategy space. A detailed study is provided in Section 4.4. Furthermore, UniAP's strategy optimization time is also significantly shorter than the two baseline methods. + +On ENVC, UniAP shows an optimization speedup of $11.83 \times$ and a training speedup of $3.80 \times$ compared to Galvatron on Llama. We examine the parallel strategy they adopted. UniAP employs an 8-stage PP with a micro-batch size of 1, whereas Galvatron employs a 4-stage PP with a micro-batch size of 3 (with the final micro-batch containing 2 samples). Within each PP stage, Galvatron utilizes a 2-way TP. For ENVC, which is equipped with 8 A100 40GB PCIe GPUs, minimizing communication volume is critical. Since PP has significantly less inter-device communication volume than TP, the parallel strategy discovered by UniAP is more reasonable than Galvatron in this case. + +Table 2 shows the results on ENVD-32. On this environment, UniAP consistently identifies the fastest parallel strategies while expending considerably less time on strategy optimization compared to Megatron (about $157 \times$ faster for + +
Env.ModelTraining throughput (samples/s)Minimum speedupMaximum speedup
GalvatronAlpaUniAP
ENVABERT-Huge33.46 ± 0.2831.56 ± 0.0433.46 ± 0.281.001.06
T5-Large23.29 ± 0.04MEM×2)23.29 ± 0.041.001.00
ViT-Huge109.51 ± 0.0797.66 ± 1.42109.51 ± 0.071.001.12
Swin-HugeCUDA×3)N/Å4)67.96 ± 0.12N/Å4)N/Å4)
ENVB-8BERT-Huge6.27 ± 0.178.95 ± 0.0610.77 ± 0.131.201.71
T5-Large1)8.06 ± 0.06MEM×2)7.98 ± 0.050.990.99
ViT-Huge32.20 ± 0.1738.74 ± 0.2045.58 ± 0.541.181.41
Swin-Huge13.90 ± 0.17N/Å4)19.08 ± 0.101.371.37
ENVCLlama-7B1.22 ± 0.01N/Å4)4.63 ± 0.0073.803.80
Env.ModelStrategy optimization time (min.)Minimum speedupMaximum speedup
GalvatronAlpaUniAP
ENVABERT-Huge6.44 ± 0.588>400.37 ± 0.00217.29>107.41
T5-Large12.41 ± 0.122MEM×2)0.89 ± 0.00713.9813.98
ViT-Huge6.29 ± 0.464>400.57 ± 0.00910.95>69.60
Swin-Huge11.88 ± 0.666N/Å4)2.16 ± 0.0045.495.49
ENVB-8BERT-Huge2.04 ± 0.010>401.51 ± 0.0051.34>26.32
T5-Large1)2.64 ± 0.110MEM×2)0.91 ± 0.0052.902.90
ViT-Huge2.37 ± 0.180>401.11 ± 0.0112.14>36.01
Swin-Huge4.29 ± 0.320N/Å4)2.29 ± 0.0101.871.87
ENVCLlama-7B6.84 ± 0.055N/Å4)0.58 ± 0.00611.8311.83
+ +1) T5-Large tested on ENVB is restricted to 16/16 layers to avoid out-of-memory (OOM) exceptions. +2) MEM×: OOM exceptions during strategy optimization. +3) CUDA×: CUDA OOM exceptions during model training. +4) The official implementation for Swin-Huge and Llama in Alpa is absent. We have endeavored to adopt the code, yet encountered compilation errors. Consequently, experiments related to these models on Alpa are marked as N/A. + +Table 1. Training throughput and strategy optimization time on ENVA, ENVB-8, and ENVC. + +
ModelTraining throughput (samples/s)Strategy optimization time (min.)
MegatronDeepSpeedUniAPMegatronDeepSpeedUniAP
Llama-7B2.01 ± 0.005\( SOL \times^{1)} \)2.01 ± 0.005>8.0 hours\( SOL \times^{1)} \)3.07 ± 0.121
Llama-13B0.82 ± 0.001\( SOL \times^{1)} \)0.82 ± 0.001>2.5 hours\( SOL \times^{1)} \)1.95 ± 0.076
+ +1) $\mathrm{SOL} \times$ : No solution after strategy optimization. + +Table 2. Training throughput and strategy optimization time on ENVD-32. + +Llama-7B and $77 \times$ faster for Llama-13B). Please note that the strategy optimization time Megatron needs for Llama-7B surpasses that for Llama-13B. We attribute this phenomenon to the variations in the mini-batch size. Specifically, the escalation from a mini-batch size of 4 (employed in Llama-13B) to 8 (employed in Llama-7B) leads to a rise in the number of candidate parallel strategies, as well as the parallel strategies that will successfully train the model. As a result, the strategy optimization time will become increasingly long when the mini-batch size exceeds 8 in the pretraining scenario [1, 40, 41]. In this situation, UniAP will identify the optimal parallel strategy much faster. + +Furthermore, our experiments highlight a limitation encountered with DeepSpeed (ZeRO-3). It requires the mini + +batch size to be divisible evenly by the total number of computation devices. This specific prerequisite prevents DeepSpeed from successfully launching the training process with 32 DCUs. + +To facilitate further discussions, we provide a case study of the optimal parallel strategy in Appendix G. + +# 4.3. Scalability + +We study the scalability of UniAP using BERT, T5, ViT, and Swin on ENVB. We set the mini-batch sizes for each data point as 8, 4, 32, and 16 times the number of nodes (denoted as '#nodes'). The experimental results are shown in Figure 5. In Figure 5a, the training throughput of the optimal strategy demonstrate near-linearity as the number of nodes and mini + +
ModelBatch sizeTraining throughput (samples/s)#infeasible5)#candidate6)
1 st.1)2 nd.2)Slowest3)Median4)
Llama-7B82.011.920.220.824164
Llama-13B40.820.580.270.424248
+ +1) 1 st.: The training throughput achieved by the fastest parallel strategy. +2) 2 nd.: The training throughput achieved by the second fastest parallel strategy. +3) Slowest: The training throughput achieved by the slowest parallel strategy. +4) Median: The median value of training throughputs across parallel strategies that will successfully train the model. +5) #infeasible: The number of parallel strategies that will encounter exceptions such as CUDA OOM during model training. +6) #candidate: The total number of candidate parallel strategies. + +![](images/004e96860ab7f6150ce8899db71deaa4401cda5a214c3fa308f9c6d4b73b679f.jpg) +(a) Training throughput. + +![](images/a83ef687e1bdd872c0500a20f62216c10a3a2e292f630d8d6dfa4a39e8fc7f70.jpg) +(b) Strategy optimization time. +Figure 5. Scalability on training throughput and strategy optimization time. + +Table 3. Statistics on the candidate parallel strategies for Megatron. + +
Number of DCUsTraining throughput (samples/s)Strategy optimization time(min.)
160.1726 ± 0.00050.75 ± 0.0100
320.3393 ± 0.00041.64 ± 0.0321
640.6656 ± 0.00231.55 ± 0.0058
+ +Table 4. Scalability on training throughput among DCU clusters. + +batch size increase. In Figure 5b, the strategy optimization time matches the complexity analysis in Section 3.5. + +Additionally, we use DCU nodes in ENVD to scale UniAP to larger clusters on Llama-7B, with precision of FP32. We conducted experiments on clusters with 16, 32, and 64 DCUs, where we set the mini-batch size to $1/8$ of the number of DCUs (i.e., for a cluster with 16 DCUs, we set the minibatch size to 2). The experimental results are shown in Table 4, further demonstrating the near-linear scalability of UniAP. + +# 4.4. Further study on the significance of UniAP + +Finally, we investigate the significance of UniAP by answering two questions. The first question is why we need AP rather than MP. We examine the statistics on the candidate parallel strategies for Megatron, as shown in Table 3. With MP, most users are unfamiliar with the complex hyperparameter detail of the candidate parallel strategies and typically choose a parallel strategy at random, which will result in a $64.1\%$ (41/64) chance of failing to train Llama-7B and an $87.5\%$ (42/48) chance of failing to train Llama-13B even when the hardware resources are sufficient. Even if they successfully eliminate all infeasible parallel strategies, they still only have a $50\%$ chance of identifying a parallel strata + +![](images/6a3ed920a1cd6fc2b02003b23020df2c91b40ad6c6921d0d8b5b17713eb59e0a.jpg) +Figure 6. Ablation study. $\mathrm{SOL} \times$ indicates no feasible solution during strategy optimization, and $\mathrm{CUDA} \times$ indicates CUDA OOM exceptions during training. + +egy with training throughput better than 0.82 samples/s for Llama-7B and 0.42 samples/s for Llama-13B. When the users can't identify the fastest parallel strategy out of hundreds of candidates, they may sacrifice at least $4.4\%$ of the training throughput for Llama-7B and $29.2\%$ for Llama-13B. In such circumstances, UniAP is able to identify the fastest strategy, as shown in Table 2. + +The second question is why UniAP achieves better performance than other AP methods. We investigate the importance of the strategy space for the optimality of parallel strategies with an ablation study. Specifically, we constrain the strategy space to inter-layer-only and intra-layer-only strategies and evaluate the training throughput of the resulting optimal strategy on ENVB. We set the mini-batch sizes to be 16, 12, 64, and 32, respectively. Results are shown in Figure 6. We can find that constraining the strategy space compromises the optimality of parallel strategies or gets strategies that encounter OOM across different models. Hence, unifying inter- and intra-layer AP for joint optimization makes UniAP outperform other AP methods. + +# 5. Conclusion + +In this paper, we propose a novel AP method called UniAP to unify inter- and intra-layer AP by MIQP. To the best of our knowledge, UniAP is the first parallel method that can jointly optimize the two categories of parallel strategies to find an optimal solution. Experimental results show that UniAP can outperform other state-of-the-art baselines to achieve the best performance. + +# 6. Acknowledgement + +This work is supported by National Key R&D Program of China (No.2020YFA0713900), NSFC Project (No.12326615), the Key Major Project of the Pengcheng Laboratory (No.PCL2024A06). + +# References + +[1] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language Models are Few-Shot Learners. In Advances in Neural Information Processing Systems 33, pages 1877-1901, 2020. 1, 7 +[2] Zhenkun Cai, Xiao Yan, Kaihao Ma, Yidi Wu, Yuzhen Huang, James Cheng, Teng Su, and Fan Yu. TensorOpt: Exploring the Tradeoffs in Distributed DNN Training With AutoParallelism. IEEE Transactions on Parallel and Distributed Systems, 33(8):1967-1981, 2022. 1, 2 +[3] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training Deep Nets with Sublinear Memory Cost. CoRR, abs/1604.06174, 2016. 6 +[4] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. PaLM: Scaling Language Modeling with Pathways. Journal of Machine Learning Research, 24:240:1-240:113, 2023. 17 +[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In *The North American Chapter of the Association for Computational Linguistics: Human Language Technologies*, pages 4171–4186, 2019. 6, 14 +[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, + +Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In International Conference on Learning Representations, 2021. 6, 14 +[7] Nan Du, Yanping Huang, Andrew M. Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yanqi Zhou, Adams Wei Yu, Orhan Firat, Barret Zoph, Liam Fedus, Maarten P. Bosma, Zongwei Zhou, Tao Wang, Yu Emma Wang, Kellie Webster, Marie Pellat, Kevin Robinson, Kathleen S. Meier-Hellstern, Toju Duke, Lucas Dixon, Kun Zhang, Quoc V. Le, Yonghui Wu, Zhifeng Chen, and Claire Cui. GLaM: Efficient Scaling of Language Models with Mixture- of-Experts. In International Conference on Machine Learning, pages 5547-5569, 2022. 1 +[8] FairScale authors. FairScale: A general purpose modular PyTorch library for high performance and large scale training. https://github.com/facebookresearch/fairscale, 2021. 1, 2, 6, 14 +[9] Shiqing Fan, Yi Rong, Chen Meng, Zongyan Cao, Siyu Wang, Zhen Zheng, Chuan Wu, Guoping Long, Jun Yang, Lixue Xia, Lansong Diao, Xiaoyong Liu, and Wei Lin. DAPPLE: A Pipelined Data Parallel Approach for Training Large Models. In Symposium on Principles and Practice of Parallel Programming, pages 431-445, 2021. 1, 2, 3 +[10] William Fedus, Barret Zoph, and Noam Shazeer. Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity. The Journal of Machine Learning Research, 23(120):5232-5270, 2022. 1 +[11] Michael J. Flynn. Very High-speed Computing Systems. Proceedings of the IEEE, 54(12):1901-1909, 1966. 1 +[12] Michael J. Flynn. Some Computer Organizations and Their Effectiveness. IEEE Transactions on Computers, C-21(9): 948–960, 1972. 1 +[13] Gurobi Optimization, LLC. Gurobi Optimizer Reference Manual. https://www.gurobi.com, 2023. 6, 14 +[14] Chaoyang He, Shen Li, Mahdi Soltanolkotabi, and Salman Avestimehr. PipeTransformer: Automated Elastic Pipelining for Distributed Training of Large-scale Models. In International Conference on Machine Learning, pages 4150-4159, 2021. 1, 2 +[15] Yanping Huang, Youlong Cheng, Ankur Bapna, Orhan First, Dehao Chen, Mia Xu Chen, HyoukJoong Lee, Jiquan Ngiam, Quoc V. Le, Yonghui Wu, and Zhifeng Chen. GPipe: Efficient Training of Giant Neural Networks using Pipeline Parallelism. In Advances in Neural Information Processing Systems 32, pages 103-112, 2019. 1, 2, 4, 6, 14 +[16] Zhihao Jia, Sina Lin, Charles R. Qi, and Alex Aiken. Exploring Hidden Dimensions in Parallelizing Convolutional Neural Networks. In International Conference on Machine Learning, pages 2279-2288, 2018. 1, 2 +[17] Zhihao Jia, Matei Zaharia, and Alex Aiken. Beyond Data and Model Parallelism for Deep Neural Networks. In Machine Learning and Systems, pages 1-13, 2019. 1, 2 +[18] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. In International Conference on Learning Representations, 2015. 14 + +[19] Rafael Lazimy. Mixed-integer Quadratic Programming. Math. Program., 22(1):332-349, 1982. 1 +[20] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. GShard: Scaling Giant Models with Conditional Computation and Automatic Sharding. In International Conference on Learning Representations, 2021. 1 +[21] Shigang Li and Torsten Hoefler. Chimera: Efficiently Training Large-Scale Neural Networks with Bidirectional Pipelines. In International Conference for High Performance Computing, Networking, Storage and Analysis, pages 27:1-27:14, 2021. 1, 3 +[22] Shen Li, Yanli Zhao, Rohan Varma, Omkar Salpekar, Pieter Noordhuis, Teng Li, Adam Paszke, Jeff Smith, Brian Vaughan, Pritam Damania, and Soumith Chintala. PyTorch Distributed: Experiences on Accelerating Data Parallel Training. Proceedings of the VLDB Endowment, 13(12):3005-3018, 2020. 1, 2, 6, 14 +[23] Yuliang Liu, Shenggui Li, Jiarui Fang, Yanjun Shao, Boyuan Yao, and Yang You. Colossal-Auto: Unified Automation of Parallelization and Activation Checkpoint for Large-scale Models. CoRR, abs/2302.02599, 2023. 1, 2, 5 +[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. In International Conference on Computer Vision, pages 9992-10002, 2021. 6, 14 +[25] Xupeng Miao, Yujie Wang, Youhe Jiang, Chunan Shi, Xiaonan Nie, Hailin Zhang, and Bin Cui. Galvatron: Efficient Transformer Training over Multiple GPUs Using Automatic Parallelism. Proceedings of the VLDB Endowment, 16(3): 470-479, 2022. 2, 3, 6, 15 +[26] Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory F. Diamos, Erich Elsen, David García, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, and Hao Wu. Mixed Precision Training. In International Conference on Learning Representations, 2018. 6 +[27] Microsoft. Megatron-DeepSpeed. https://github.com/microsoft/Megatron-DeepSpeed/tree/main, 2021.6 +[28] Deepak Narayanan, Aaron Harlap, Amar Phanishayee, Vivek Seshadri, Nikhil R. Devanur, Gregory R. Ganger, Phillip B. Gibbons, and Matei Zaharia. PipeDream: Generalized Pipeline Parallelism for DNN Training. In Symposium on Operating Systems Principles, pages 1-15, 2019. 1, 2 +[29] Deepak Narayanan, Amar Phanishayee, Kaiyu Shi, Xie Chen, and Matei Zaharia. Memory-Efficient Pipeline-Parallel DNN Training. In International Conference on Machine Learning, pages 7937–7947, 2021. 1, 3 +[30] Deepak Narayanan, Mohammad Shoeybi, Jared Casper, Patrick LeGresley, Mostofa Patwary, Vijay Korthikanti, Dmitri Vainbrand, Prethvi Kashinkunti, Julie Bernauer, Bryan Catanzaro, Amar Phanishayee, and Matei Zaharia. Efficient Large-scale Language Model Training on GPU Clusters Using Megatron-LM. In International Conference for High Performance Computing, Networking, Storage and Analysis, pages 58:1-58:15, 2021. 1, 2, 3, 6, 14 + +[31] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 6, 14 +[32] Samyam Rajbhandari, Jeff Rasley, Olatunjri Ruwase, and Yuxiong He. ZeRO: Memory Optimizations Toward Training Trillion Parameter Models. In International Conference for High Performance Computing, Networking, Storage and Analysis, pages 20:1-20:16, 2020. 2, 6 +[33] Saeed Rashidi, Matthew Denton, Srinivas Sridharan, Sudarshan Srinivasan, Amoghavarsha Suresh, Jade Nie, and Tushar Krishna. Enabling Compute-Communication Overlap in Distributed Deep Learning Training Platforms. In International Symposium on Computer Architecture, pages 540–553, 2021. 3 +[34] Jeff Rasley, Samyam Rajbhandari, Olatunjri Ruwase, and Yuxiong He. DeepSpeed: System Optimizations Enable Training Deep Learning Models with Over 100 Billion Parameters. In The 26th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 3505-3506, 2020. 1, 6 +[35] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision, 115(3):211-252, 2015. 14 +[36] Michael Schaarschmidt, Dominik Grewe, Dimitrios Vytiniotis, Adam Paszke, Georg Stefan Schmid, Tamara Norman, James Molloy, Jonathan Godwin, Norman Alexander Rink, Vinod Nair, and Dan Belov. Automap: Towards Ergonomic Automated Parallelism for ML Models. CoRR, abs/2112.02958, 2021. 1, 2 +[37] Noam Shazeer, Youlong Cheng, Niki Parmar, Dustin Tran, Ashish Vaswani, Penporn Koanantakool, Peter Hawkins, HyoukJoong Lee, Mingsheng Hong, Cliff Young, Ryan Sepassi, and Blake A. Hechtman. Mesh-TensorFlow: Deep Learning for Supercomputers. In Advances in Neural Information Processing Systems 31, pages 10435-10444, 2018. 1, 2 +[38] Jakub Tarnawski, Amar Phanishayee, Nikhil R. Devanur, Divya Mahajan, and Fanny Nina Paravecino. Efficient Algorithms for Device Placement of DNN Graph Operators. In Advances in Neural Information Processing Systems 33, pages 15451-15463, 2020. 1, 2, 4, 12 +[39] Jakub Tarnawski, Deepak Narayanan, and Amar Phanishayee. Piper: Multidimensional Planner for DNN Parallelization. In Advances in Neural Information Processing Systems 34, pages 24829-24840, 2021. 1, 2, 4 +[40] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurélien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. LLaMA: Open and Efficient Foundation Language Models. CoRR, abs/2302.13971, 2023. 1, 6, 7, 14 +[41] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas + +Blecher, Cristian Canton-Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, MarieAnne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open Foundation and Fine-Tuned Chat Models. CoRR, abs/2307.09288, 2023. 1, 6, 7, 14 +[42] Minjie Wang, Chien-Chin Huang, and Jinyang Li. Supporting Very Large Models using Automatic Dataflow Graph Partitioning. In Proceedings of the Fourteenth EuroSys Conference, pages 26:1-26:17, 2019. 1, 2 +[43] Wikipedia Foundation. Wikipedia Downloads. https:// dumps.wikipedia.org, 2023. 14 +[44] Yuanzhong Xu, HyoukJoong Lee, Dehao Chen, Blake A. Hechtman, Yanping Huang, Rahul Joshi, Maxim Krikun, Dmitry Lepikhin, Andy Ly, Marcello Maggioni, Ruoming Pang, Noam Shazeer, Shibo Wang, Tao Wang, Yonghui Wu, and Zhifeng Chen. GSPMD: General and Scalable Parallelization for ML Computation Graphs. CoRR, abs/2105.04663, 2021. 1, 2 +[45] Binhang Yuan, Yongjun He, Jared Davis, Tianyi Zhang, Tri Dao, Beidi Chen, Percy Liang, Christopher Ré, and Ce Zhang. Decentralized Training of Foundation Models in Heterogeneous Environments. In Advances in Neural Information Processing Systems 35, pages 25464-25477, 2022. 2 +[46] Shixiong Zhao, Fanxin Li, Xusheng Chen, Xiuxian Guan, Jianyu Jiang, Dong Huang, Yuhao Qing, Sen Wang, Peng Wang, Gong Zhang, Cheng Li, Ping Luo, and Heming Cui. vPipe: A Virtualized Acceleration System for Achieving Efficient and Scalable Pipeline Parallel DNN Training. IEEE Transactions on Parallel and Distributed Systems, 33(3):489-506, 2022. 1, 2 +[47] Lianmin Zheng, Zhuohan Li, Hao Zhang, Yonghao Zhuang, Zhifeng Chen, Yanping Huang, Yida Wang, Yuanzhong Xu, Danyang Zhuo, Eric P. Xing, Joseph E. Gonzalez, and Ion Stoica. Alpa: Automating Inter- and Intra-Operator Parallelism for Distributed Deep Learning. In USENIX Symposium on Operating Systems Design and Implementation, pages 559–578, 2022. 1, 2, 5, 6 \ No newline at end of file diff --git a/2025/UniAP_ Unifying Inter- and Intra-Layer Automatic Parallelism by Mixed Integer Quadratic Programming/images.zip b/2025/UniAP_ Unifying Inter- and Intra-Layer Automatic Parallelism by Mixed Integer Quadratic Programming/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..904e2cfa47773c7ad9bed96946fbf76c978d18d2 --- /dev/null +++ b/2025/UniAP_ Unifying Inter- and Intra-Layer Automatic Parallelism by Mixed Integer Quadratic Programming/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1bdbe4440d4692d142a591656231f7ed48c947203e4502c8a6348642964c3aad +size 503493 diff --git a/2025/UniAP_ Unifying Inter- and Intra-Layer Automatic Parallelism by Mixed Integer Quadratic Programming/layout.json b/2025/UniAP_ Unifying Inter- and Intra-Layer Automatic Parallelism by Mixed Integer Quadratic Programming/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d41d77c2e98ade43ed641cdf9505704511cddae6 --- /dev/null +++ b/2025/UniAP_ Unifying Inter- and Intra-Layer Automatic Parallelism by Mixed Integer Quadratic Programming/layout.json @@ -0,0 +1,10762 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 102, + 555, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 102, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 555, + 140 + ], + "type": "text", + "content": "UniAP: Unifying Inter- and Intra-Layer Automatic Parallelism by Mixed Integer Quadratic Programming" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 166, + 160, + 443, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 160, + 443, + 217 + ], + "spans": [ + { + "bbox": [ + 166, + 160, + 443, + 217 + ], + "type": "text", + "content": "Hao Lin\\*, Ke Wu\\*, Jie Li\\*, Jun Li, Wu-Jun Li† \nNational Key Laboratory for Novel Software Technology School of Computer Science \nNanjing University, Nanjing 210023, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 218, + 541, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 218, + 541, + 232 + ], + "spans": [ + { + "bbox": [ + 67, + 218, + 541, + 232 + ], + "type": "text", + "content": "{hao.lin, ke.wu, jie-li, lijun}@smail.nju.edu.cn, liwujun@nju.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 284, + 297, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 284, + 297, + 499 + ], + "spans": [ + { + "bbox": [ + 54, + 284, + 297, + 499 + ], + "type": "text", + "content": "Distributed learning is commonly used for training deep learning models, especially large models. In distributed learning, manual parallelism (MP) methods demand considerable human effort and have limited flexibility. Hence, automatic parallelism (AP) methods have recently been proposed for automating the parallel strategy optimization process. Existing AP methods suffer from sub-optimal solutions because they do not jointly optimize the two categories of parallel strategies (i.e., inter-layer parallelism and intra-layer parallelism). In this paper, we propose a novel AP method called UniAP, which unifies inter- and intra-layer automatic parallelism by mixed integer quadratic programming. To the best of our knowledge, UniAP is the first parallel method that can jointly optimize the two categories of parallel strategies to find an optimal solution. Experimental results show that UniAP outperforms state-of-the-art methods by up to " + }, + { + "bbox": [ + 54, + 284, + 297, + 499 + ], + "type": "inline_equation", + "content": "3.80 \\times" + }, + { + "bbox": [ + 54, + 284, + 297, + 499 + ], + "type": "text", + "content": " in throughput and reduces strategy optimization time by up to " + }, + { + "bbox": [ + 54, + 284, + 297, + 499 + ], + "type": "inline_equation", + "content": "107 \\times" + }, + { + "bbox": [ + 54, + 284, + 297, + 499 + ], + "type": "text", + "content": " across five Transformer-based models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 523, + 137, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 523, + 137, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 523, + 137, + 536 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 543, + 297, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 297, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 297, + 676 + ], + "type": "text", + "content": "Distributed learning (also called parallel learning) on clusters with several machines or GPUs is commonly used for training deep learning models, especially for some large models with billions of parameters [1, 40, 41]. Several parallel strategies, including pipeline parallelism (PP), data parallelism (DP), tensor parallelism (TP), and fully sharded data parallelism (FSDP), have been proposed for distributed learning. These parallel strategies can be divided into two main categories: inter-layer parallelism and intra-layer parallelism. Inter-layer parallelism [7, 9, 10, 15, 20, 21, 28, 29], which includes PP, partitions the model into disjoint sets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 260, + 555, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 555, + 296 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 555, + 296 + ], + "type": "text", + "content": "without partitioning tensors in each layer. Intra-layer parallelism [8, 22, 30, 34], which includes DP, TP, and FSDP, partitions tensors in a layer along one or more axes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 297, + 556, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 556, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 556, + 452 + ], + "type": "text", + "content": "The parallel method1 in one specific distributed learning method or system typically adopts one parallel strategy or a combination of several parallel strategies. Existing parallel methods can be divided into two categories: manual parallelism (MP) methods and automatic parallelism (AP) methods. In MP methods [30, 37, 44], one or several parallel strategies are manually optimized by researchers or developers. MP methods require extensive domain knowledge in deep learning models and hardware architectures. With the rapid development of deep learning models and the increasing diversity of modern hardware architectures [11, 12], MP methods demand considerable human effort and have limited flexibility." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 453, + 557, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 557, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 557, + 574 + ], + "type": "text", + "content": "To address the two limitations of MP methods, AP methods [14, 28, 47] have recently been proposed for automating the parallel strategy optimization process. Although existing AP methods have achieved promising progress, they optimize the two categories of parallel strategies separately rather than jointly. More specifically, some methods optimize only one category of parallel strategies [2, 16, 17, 23, 36, 42, 46], and the others optimize inter- and intra-layer parallelism hierarchically [9, 14, 28, 38, 39, 47]. Hence, existing AP methods suffer from sub-optimal solutions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 574, + 555, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 555, + 609 + ], + "type": "text", + "content": "In this paper, we propose a novel AP method called UniAP for distributed learning. The contributions of UniAP are outlined as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 611, + 553, + 681 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 313, + 611, + 553, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 553, + 635 + ], + "type": "text", + "content": "- UniAP unifies inter- and intra-layer automatic parallelism by mixed integer quadratic programming (MIQP) [19]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 635, + 553, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 553, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 553, + 681 + ], + "type": "text", + "content": "- To the best of our knowledge, UniAP is the first parallel method that can optimize the two categories of parallel strategies jointly rather than separately to find an optimal solution." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 684, + 135, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 684, + 135, + 693 + ], + "spans": [ + { + "bbox": [ + 66, + 684, + 135, + 693 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 67, + 693, + 144, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 693, + 144, + 703 + ], + "spans": [ + { + "bbox": [ + 67, + 693, + 144, + 703 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 693, + 555, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 555, + 714 + ], + "type": "text", + "content": "To avoid confusion, we treat 'parallel method' and 'parallel strategy' as two different terminologies in this paper." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20947" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 298, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 298, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 298, + 120 + ], + "type": "text", + "content": "- Experimental results show that UniAP outperforms state-of-the-art methods by up to " + }, + { + "bbox": [ + 55, + 72, + 298, + 120 + ], + "type": "inline_equation", + "content": "3.80 \\times" + }, + { + "bbox": [ + 55, + 72, + 298, + 120 + ], + "type": "text", + "content": " in throughput and reduces strategy optimization time by up to " + }, + { + "bbox": [ + 55, + 72, + 298, + 120 + ], + "type": "inline_equation", + "content": "107 \\times" + }, + { + "bbox": [ + 55, + 72, + 298, + 120 + ], + "type": "text", + "content": " across five Transformer-based models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 135, + 145 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 135, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 135, + 145 + ], + "type": "text", + "content": "2. Background" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 152, + 155, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 152, + 155, + 166 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 155, + 166 + ], + "type": "text", + "content": "2.1. Parallel strategy" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 170, + 296, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 170, + 296, + 230 + ], + "spans": [ + { + "bbox": [ + 55, + 170, + 296, + 230 + ], + "type": "text", + "content": "Pipeline parallelism (PP) In PP [15], each worker (machine or GPU) holds a subset of model layers. Adjacent layers on different workers need to transfer activations in the forward propagation (FP) step and gradients in the backward propagation (BP) step." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 231, + 296, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 231, + 296, + 303 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 296, + 303 + ], + "type": "text", + "content": "Data parallelism (DP) In DP [22], each worker holds a replica of the whole model and partitions training samples. In each iteration, each worker computes gradients and synchronizes them with the other workers using all-reduce collective communication (CC). All workers will have the same model parameters after the synchronization step." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 303, + 296, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 296, + 374 + ], + "type": "text", + "content": "Tensor parallelism (TP) In TP [30], each worker holds a replica of training samples and partitions within model layers. In each iteration, each worker computes its local outputs in FP and its local gradients in BP. To synchronize outputs and gradients, all workers will perform all-reduce CC in FP and BP steps according to the partition scheme." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 375, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 296, + 460 + ], + "type": "text", + "content": "Fully sharded data parallelism (FSDP) FSDP [8, 32] partitions optimizer states, parameters and gradients of the model into separate workers. During the FP and BP step of each iteration, FSDP performs an all-gather CC to obtain the complete parameters for the relevant layer, respectively. After computing gradients, FSDP conducts a reduce-scatter CC to distribute the global gradients among the workers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 468, + 153, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 153, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 153, + 479 + ], + "type": "text", + "content": "2.2. Parallel method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 485, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 296, + 616 + ], + "type": "text", + "content": "Manual parallelism (MP) MP refers to the parallel methods in which human experts design and optimize the parallel strategies. Representative MP methods include Megatron-LM [30], Mesh-TensorFlow [37], and GSPMD [44]. Megatron-LM manually designs TP and PP strategies for training Transformer-based models and exhibits superior efficiency. Mesh-TensorFlow and GSPMD require human effort to designate and tune the intra-layer parallel strategy. These methods rely on expert design and have little flexibility, challenging their automatic application to other models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 617, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 297, + 714 + ], + "type": "text", + "content": "Inter-layer-only AP or intra-layer-only AP For interlayer-only AP, GPipe [15] and vPipe [46] employ a balanced partition algorithm and a dynamic layer partitioning middleware to partition pipelines, respectively. For intra-layer-only AP, OptCNN [16], TensorOpt [2], and Tofu [42] employ dynamic programming methods to optimize DP and TP strategies together. FlexFlow [17] and Automap [36] use the Monte Carlo method to find the optimal DP and TP strat" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 73, + 553, + 260 + ], + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 260 + ], + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 260 + ], + "type": "image", + "image_path": "daccd7bd1041350ba01fa7ba9cd1e1eaecb29d6e64a83d9930b70150b159382a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 270, + 555, + 336 + ], + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 336 + ], + "type": "text", + "content": "Figure 1. Parallel methods for optimizing parallel strategies for a three-layer model. The different arrangements of slices with varying transparency within the same layer block indicate different intra-layer parallelism strategies adopted by layers. The different arrangements of gray blocks which wrap the layer blocks indicate different inter-layer parallelism strategies." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 347, + 555, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 394 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 394 + ], + "type": "text", + "content": "egy. Colossal-Auto [23] utilizes integer programming (IP) techniques to generate intra-layer parallelism and activation checkpointing strategies. All these methods optimize only one category of parallel strategies." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 395, + 556, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 556, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 556, + 551 + ], + "type": "text", + "content": "Inter- and intra-layer AP PipeDream [28], DAPPLE [9], and PipeTransformer [14] use dynamic programming to determine optimal strategies for both DP and PP. DNN-partitioning [38] adopts IP and dynamic programming to explore DP and PP strategies. DT-FM [45] combines genetic algorithm and the local search strategy to explore DP and PP strategies in a layer-wise manner. Piper [39] and Alpa [47] adopt a parallel method considering DP, TP, and PP. Galvatron [25] uses dynamic programming to determine DP, TP, and FSDP strategies in a single pipeline stage. As for PP, it partitions stages and determines micro-batch size using naive greedy algorithms. All these methods are hierarchical, which will result in sub-optimal solutions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 562, + 370, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 562, + 370, + 573 + ], + "spans": [ + { + "bbox": [ + 314, + 562, + 370, + 573 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "In this section, we introduce our proposed method called UniAP, which jointly optimizes the two categories of parallel strategies, including PP, DP, TP, and FSDP, to find an optimal solution. Figure 1 illustrates the difference between UniAP and other automatic parallelism methods. Inter-layer-only and intra-layer-only AP methods optimize (search) from a set of candidate inter-layer-only and intra-layer-only parallel strategies, respectively. Hierarchical AP methods first adopt greedy or dynamic programming to propose candidate inter-layer parallel strategies. Then, they optimize the intra-layer parallel strategy for every fixed inter-layer parallel strategy." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20948" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 76, + 291, + 155 + ], + "blocks": [ + { + "bbox": [ + 63, + 76, + 291, + 155 + ], + "lines": [ + { + "bbox": [ + 63, + 76, + 291, + 155 + ], + "spans": [ + { + "bbox": [ + 63, + 76, + 291, + 155 + ], + "type": "image", + "image_path": "e9250057b878b58ef36dc760bdaa16ab25a6caa0041ff5f300fe1e13b4c9702e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 119, + 167, + 233, + 178 + ], + "lines": [ + { + "bbox": [ + 119, + 167, + 233, + 178 + ], + "spans": [ + { + "bbox": [ + 119, + 167, + 233, + 178 + ], + "type": "text", + "content": "Figure 2. Flowchart of UniAP." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 191, + 295, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 191, + 295, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 295, + 214 + ], + "type": "text", + "content": "UniAP has the largest strategy space for exploration (joint optimization)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 216, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 296, + 396 + ], + "type": "text", + "content": "Figure 2 illustrates the flowchart of UniAP. UniAP first profiles the runtime information for the user's hardware environment and the deep learning model. After that, UniAP estimates inter- and intra-layer costs given the computation graph and profiling results with its cost models. The estimated costs and the computation graph are then transformed into an MIQP problem. The objective function of the MIQP is to maximize the training throughput, or in other words, to minimize the training time per iteration (TPI). By iteratively applying the cost model and MIQP with different parameters, UniAP determines the minimal TPI and its corresponding parallel strategies. We name this process the Unified Optimization Process (UOP). Finally, UniAP interprets the parallel strategies into the execution plan for the designated model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 404, + 119, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 119, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 119, + 417 + ], + "type": "text", + "content": "3.1. Profiling" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 422, + 296, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 566 + ], + "type": "text", + "content": "UniAP collects runtime information about the hardware environment and deep learning model during profiling. For the hardware environment, UniAP evaluates the efficiency of all-reduce and point-to-point (P2P) communication for different device subsets. For example, when profiling a node with 4 GPUs, UniAP measures the all-reduce efficiency for various DP, TP, and FSDP combinations across these GPUs. Additionally, UniAP ranks these GPUs from 0 to 3 and evaluates the speed of P2P for two pipeline options: " + }, + { + "bbox": [ + 55, + 422, + 296, + 566 + ], + "type": "inline_equation", + "content": "(0\\to 2" + }, + { + "bbox": [ + 55, + 422, + 296, + 566 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 422, + 296, + 566 + ], + "type": "inline_equation", + "content": "1\\rightarrow 3)" + }, + { + "bbox": [ + 55, + 422, + 296, + 566 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 422, + 296, + 566 + ], + "type": "inline_equation", + "content": "(0\\to 1,1\\to 2" + }, + { + "bbox": [ + 55, + 422, + 296, + 566 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 422, + 296, + 566 + ], + "type": "inline_equation", + "content": "2\\to 3)" + }, + { + "bbox": [ + 55, + 422, + 296, + 566 + ], + "type": "text", + "content": ". Furthermore, UniAP estimates the computation-communication overlap coefficient (CCOC) [25, 33]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 567, + 296, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 296, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 296, + 661 + ], + "type": "text", + "content": "UniAP acquires two types of information for the deep learning model: computation time and memory usage. On one hand, UniAP distinguishes the forward computation time per sample for different types of hidden layers. On the other hand, UniAP collects memory usage information for each layer, including the memory occupied by parameters and the memory usage of activation per sample in different TP sizes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 671, + 132, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 132, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 132, + 683 + ], + "type": "text", + "content": "3.2. Cost model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "UniAP employs two primary cost models, namely the time cost model and the memory cost model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 262 + ], + "type": "text", + "content": "Time cost model To estimate computation time, UniAP first calculates the forward computation time by multiplying the batch size with the forward computation time per sample obtained from profiling. Users can obtain a more precise result by profiling the forward time on the specified batch size. For Transformer-based models that mainly consist of the MatMul operator, the computation time in the BP stages is roughly twice that of the FP stages [21, 25, 30]. Additionally, UniAP estimates the communication time by dividing the size of transmitting tensors by the profiled communication efficiency for different communication primitives. To accommodate overlapping, UniAP multiplies the profiled CCOC by the overlapping interval of computation and communication. To model the communication time between pipeline stages, UniAP calculates the cross-stage cost between consecutive stages by the summation of P2P costs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 264, + 556, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 556, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 556, + 381 + ], + "type": "text", + "content": "Memory cost model UniAP estimates memory consumption for each layer with its memory cost model. This estimation consists of three steps for a given layer. First, it computes the activation memory cost " + }, + { + "bbox": [ + 313, + 264, + 556, + 381 + ], + "type": "inline_equation", + "content": "m_{a}" + }, + { + "bbox": [ + 313, + 264, + 556, + 381 + ], + "type": "text", + "content": " by multiplying the batch size and the profiled activation memory cost per sample of the TP size used by the strategy. Next, UniAP calculates the memory cost of model states " + }, + { + "bbox": [ + 313, + 264, + 556, + 381 + ], + "type": "inline_equation", + "content": "m_{s}" + }, + { + "bbox": [ + 313, + 264, + 556, + 381 + ], + "type": "text", + "content": " for each layer based on their parameter size " + }, + { + "bbox": [ + 313, + 264, + 556, + 381 + ], + "type": "inline_equation", + "content": "ps" + }, + { + "bbox": [ + 313, + 264, + 556, + 381 + ], + "type": "text", + "content": ", TP size " + }, + { + "bbox": [ + 313, + 264, + 556, + 381 + ], + "type": "inline_equation", + "content": "ts" + }, + { + "bbox": [ + 313, + 264, + 556, + 381 + ], + "type": "text", + "content": ", FSDP size " + }, + { + "bbox": [ + 313, + 264, + 556, + 381 + ], + "type": "inline_equation", + "content": "fs" + }, + { + "bbox": [ + 313, + 264, + 556, + 381 + ], + "type": "text", + "content": ", and a constant " + }, + { + "bbox": [ + 313, + 264, + 556, + 381 + ], + "type": "inline_equation", + "content": "c_{dtype}" + }, + { + "bbox": [ + 313, + 264, + 556, + 381 + ], + "type": "text", + "content": " dependent on the data type. Formally, we have" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 393, + 381, + 555, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 381, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 393, + 381, + 555, + 403 + ], + "type": "interline_equation", + "content": "m _ {s} = \\frac {c _ {d t y p e} \\times p s}{t s \\times f s}. \\tag {1}", + "image_path": "3e41acc454cda3bddbeb79ed144dc49ed9b9f2164658526904fabcdda35c50cc.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 407, + 556, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 556, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 556, + 514 + ], + "type": "text", + "content": "For example, if we choose precision as FP32, then " + }, + { + "bbox": [ + 313, + 407, + 556, + 514 + ], + "type": "inline_equation", + "content": "c_{dtype} = (4 + 4 + 4 + 4) / 4 = 4" + }, + { + "bbox": [ + 313, + 407, + 556, + 514 + ], + "type": "text", + "content": " since the learnable parameters, gradients, momentum and variance consume equal size of memory. If we opt for mixed precision with FP16 activated, then " + }, + { + "bbox": [ + 313, + 407, + 556, + 514 + ], + "type": "inline_equation", + "content": "c_{dtype} = (4 + 4 + 4 + 2 + 2) / 2 = 8" + }, + { + "bbox": [ + 313, + 407, + 556, + 514 + ], + "type": "text", + "content": ". Finally, UniAP aggregates the activation memory cost " + }, + { + "bbox": [ + 313, + 407, + 556, + 514 + ], + "type": "inline_equation", + "content": "m_a" + }, + { + "bbox": [ + 313, + 407, + 556, + 514 + ], + "type": "text", + "content": ", memory cost of model states " + }, + { + "bbox": [ + 313, + 407, + 556, + 514 + ], + "type": "inline_equation", + "content": "m_s" + }, + { + "bbox": [ + 313, + 407, + 556, + 514 + ], + "type": "text", + "content": ", and context memory cost " + }, + { + "bbox": [ + 313, + 407, + 556, + 514 + ], + "type": "inline_equation", + "content": "m_c" + }, + { + "bbox": [ + 313, + 407, + 556, + 514 + ], + "type": "text", + "content": " to a constant matrix " + }, + { + "bbox": [ + 313, + 407, + 556, + 514 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 407, + 556, + 514 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 407, + 556, + 514 + ], + "type": "inline_equation", + "content": "M_{uk}" + }, + { + "bbox": [ + 313, + 407, + 556, + 514 + ], + "type": "text", + "content": " denotes the memory cost for the " + }, + { + "bbox": [ + 313, + 407, + 556, + 514 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 407, + 556, + 514 + ], + "type": "text", + "content": "-th intra-layer strategy of layer " + }, + { + "bbox": [ + 313, + 407, + 556, + 514 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 313, + 407, + 556, + 514 + ], + "type": "text", + "content": " on a single device." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 520, + 519, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 519, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 519, + 533 + ], + "type": "text", + "content": "3.3. Mixed integer quadratic programming" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 537, + 555, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 555, + 574 + ], + "type": "text", + "content": "The estimated costs and the computation graph are then transformed into an MIQP problem. Its formulation includes an objective function and several constraints." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 578, + 422, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 422, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 422, + 590 + ], + "type": "text", + "content": "3.3.1. Objective function" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 556, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 677 + ], + "type": "text", + "content": "The objective function tries to minimize TPI. In this paper, we have chosen GPipe as our PP strategy for illustration.2 Figure 3 depicts the time cost decomposition of a GPipe-style PP with non-negligible communication costs. The time needed to apply gradients at the end of each iteration is not included, as it depends on the optimizer and is insignificant compared to the total time spent on FP and BP." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "text", + "content": " UniAP is also compatible with other PP strategies. For example,users need to modify only the memory constraint in Section 3.3.2 to adapt to synchronous 1F1B pipeline [9, 29]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20949" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 72, + 347, + 175 + ], + "blocks": [ + { + "bbox": [ + 62, + 72, + 347, + 175 + ], + "lines": [ + { + "bbox": [ + 62, + 72, + 347, + 175 + ], + "spans": [ + { + "bbox": [ + 62, + 72, + 347, + 175 + ], + "type": "image", + "image_path": "183a476c7327e98af73d2b0f588b882911e8b9011441b9aab9d066f331a2f2ba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 182, + 307, + 194 + ], + "lines": [ + { + "bbox": [ + 104, + 182, + 307, + 194 + ], + "spans": [ + { + "bbox": [ + 104, + 182, + 307, + 194 + ], + "type": "text", + "content": "Figure 3. Time cost decomposition of a GPipe-style PP." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 361, + 93, + 550, + 174 + ], + "blocks": [ + { + "bbox": [ + 361, + 93, + 550, + 174 + ], + "lines": [ + { + "bbox": [ + 361, + 93, + 550, + 174 + ], + "spans": [ + { + "bbox": [ + 361, + 93, + 550, + 174 + ], + "type": "image", + "image_path": "03761f9a11a6fb9e0f102a74ba3ed15114a72e5c16d162b70f2a658c9a47ae1c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 182, + 504, + 194 + ], + "lines": [ + { + "bbox": [ + 403, + 182, + 504, + 194 + ], + "spans": [ + { + "bbox": [ + 403, + 182, + 504, + 194 + ], + "type": "text", + "content": "Figure 4. A contiguous set." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 208, + 296, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 208, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 296, + 316 + ], + "type": "text", + "content": "We denote the cost for computation stages as " + }, + { + "bbox": [ + 55, + 208, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\mathbb{P} = \\{p_1, p_2, \\ldots, p_{deg}\\}" + }, + { + "bbox": [ + 55, + 208, + 296, + 316 + ], + "type": "text", + "content": " and the cost for communication stages as " + }, + { + "bbox": [ + 55, + 208, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\mathbb{O} = \\{o_1, o_2, \\ldots, o_{deg-1}\\}" + }, + { + "bbox": [ + 55, + 208, + 296, + 316 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 208, + 296, + 316 + ], + "type": "inline_equation", + "content": "deg" + }, + { + "bbox": [ + 55, + 208, + 296, + 316 + ], + "type": "text", + "content": " represents the number of computation stages, which corresponds to the pipeline parallel size. In Figure 3, " + }, + { + "bbox": [ + 55, + 208, + 296, + 316 + ], + "type": "inline_equation", + "content": "fp_i" + }, + { + "bbox": [ + 55, + 208, + 296, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 208, + 296, + 316 + ], + "type": "inline_equation", + "content": "bp_i" + }, + { + "bbox": [ + 55, + 208, + 296, + 316 + ], + "type": "text", + "content": " denote forward and backward computation time for computation stage " + }, + { + "bbox": [ + 55, + 208, + 296, + 316 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 208, + 296, + 316 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 55, + 208, + 296, + 316 + ], + "type": "inline_equation", + "content": "fo_j" + }, + { + "bbox": [ + 55, + 208, + 296, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 208, + 296, + 316 + ], + "type": "inline_equation", + "content": "bo_j" + }, + { + "bbox": [ + 55, + 208, + 296, + 316 + ], + "type": "text", + "content": " denote forward and backward communication time for communication stage " + }, + { + "bbox": [ + 55, + 208, + 296, + 316 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 208, + 296, + 316 + ], + "type": "text", + "content": ", respectively. Hence, we have " + }, + { + "bbox": [ + 55, + 208, + 296, + 316 + ], + "type": "inline_equation", + "content": "pi = fp_i + bp_i" + }, + { + "bbox": [ + 55, + 208, + 296, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 208, + 296, + 316 + ], + "type": "inline_equation", + "content": "oj = fo_j + bo_j" + }, + { + "bbox": [ + 55, + 208, + 296, + 316 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 316, + 296, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 424 + ], + "type": "text", + "content": "In a GPipe-style pipeline, we use " + }, + { + "bbox": [ + 55, + 316, + 296, + 424 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 316, + 296, + 424 + ], + "type": "text", + "content": " to denote the number of micro-batches. As illustrated in Figure 3, a mini-batch is uniformly split into four micro-batches, and the total TPI is determined by the latency of all computation and communication stages and the latency of the slowest stage. We further denote TPI in GPipe as " + }, + { + "bbox": [ + 55, + 316, + 296, + 424 + ], + "type": "inline_equation", + "content": "tpi_{gpipe}" + }, + { + "bbox": [ + 55, + 316, + 296, + 424 + ], + "type": "text", + "content": ". Given that a stage with a higher FP computation cost leads to a higher BP computation cost with high probability, we can write the objective function of GPipe-style pipeline as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 97, + 434, + 296, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 434, + 296, + 484 + ], + "spans": [ + { + "bbox": [ + 97, + 434, + 296, + 484 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min t p i _ {g p i p e} = \\sum_ {i = 1} ^ {d e g} p _ {i} + \\sum_ {j = 1} ^ {d e g - 1} o _ {j} \\tag {2} \\\\ + (c - 1) \\max (\\mathbb {P} \\cup \\mathbb {O}). \\\\ \\end{array}", + "image_path": "28b8d942e0e176e5e21991b5c6e56d336c618045a6ab15636f26a3b891519ae1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 494, + 130, + 505 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 130, + 505 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 130, + 505 + ], + "type": "text", + "content": "3.3.2. Constraint" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 510, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 664 + ], + "type": "text", + "content": "We first introduce additional notations before presenting the constraints. For a given layer " + }, + { + "bbox": [ + 55, + 510, + 296, + 664 + ], + "type": "inline_equation", + "content": "u \\in \\mathbb{V}" + }, + { + "bbox": [ + 55, + 510, + 296, + 664 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 510, + 296, + 664 + ], + "type": "inline_equation", + "content": "\\mathbb{S}_u" + }, + { + "bbox": [ + 55, + 510, + 296, + 664 + ], + "type": "text", + "content": " represents its set of intra-layer parallel strategies, " + }, + { + "bbox": [ + 55, + 510, + 296, + 664 + ], + "type": "inline_equation", + "content": "A_{uk}" + }, + { + "bbox": [ + 55, + 510, + 296, + 664 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 55, + 510, + 296, + 664 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 510, + 296, + 664 + ], + "type": "text", + "content": "-th intra-layer execution cost obtained from our time cost model. Additionally, we use " + }, + { + "bbox": [ + 55, + 510, + 296, + 664 + ], + "type": "inline_equation", + "content": "S_{uk} \\in \\{0,1\\}" + }, + { + "bbox": [ + 55, + 510, + 296, + 664 + ], + "type": "text", + "content": " to indicate whether the " + }, + { + "bbox": [ + 55, + 510, + 296, + 664 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 510, + 296, + 664 + ], + "type": "text", + "content": "-th parallel strategy is selected for the layer " + }, + { + "bbox": [ + 55, + 510, + 296, + 664 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 55, + 510, + 296, + 664 + ], + "type": "text", + "content": ", and use " + }, + { + "bbox": [ + 55, + 510, + 296, + 664 + ], + "type": "inline_equation", + "content": "P_{ui} \\in \\{0,1\\}" + }, + { + "bbox": [ + 55, + 510, + 296, + 664 + ], + "type": "text", + "content": " to indicate whether layer " + }, + { + "bbox": [ + 55, + 510, + 296, + 664 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 55, + 510, + 296, + 664 + ], + "type": "text", + "content": " is placed on the " + }, + { + "bbox": [ + 55, + 510, + 296, + 664 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 510, + 296, + 664 + ], + "type": "text", + "content": "-th computation stage. Each edge " + }, + { + "bbox": [ + 55, + 510, + 296, + 664 + ], + "type": "inline_equation", + "content": "\\langle u,v\\rangle \\in \\mathbb{E}" + }, + { + "bbox": [ + 55, + 510, + 296, + 664 + ], + "type": "text", + "content": " is assigned a resharding cost denoted by " + }, + { + "bbox": [ + 55, + 510, + 296, + 664 + ], + "type": "inline_equation", + "content": "R_{uv}" + }, + { + "bbox": [ + 55, + 510, + 296, + 664 + ], + "type": "text", + "content": " if the vertices are located within the same pipeline stage. Alternatively, if the vertices are located across consecutive stages, the resharding cost between them is denoted by " + }, + { + "bbox": [ + 55, + 510, + 296, + 664 + ], + "type": "inline_equation", + "content": "R_{uv}'" + }, + { + "bbox": [ + 55, + 510, + 296, + 664 + ], + "type": "text", + "content": ". These two resharding costs are constant matrices derived from our time cost model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Computation-stage constraint To compute the total cost for a single computation stage " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": ", all computation and communication costs associated with that stage must be aggregated and assigned to " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": ". This constraint can be formulated as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 209, + 350, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 209, + 350, + 218 + ], + "spans": [ + { + "bbox": [ + 314, + 209, + 350, + 218 + ], + "type": "text", + "content": "follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 335, + 225, + 555, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 225, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 335, + 225, + 555, + 270 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\sum_ {u \\in \\mathbb {V}} \\boldsymbol {P} _ {u i} \\boldsymbol {S} _ {u} ^ {\\top} \\boldsymbol {A} _ {u} + \\sum_ {\\langle u, v \\rangle \\in \\mathbb {E}} \\boldsymbol {P} _ {u i} \\boldsymbol {P} _ {v i} \\left(\\boldsymbol {S} _ {u} ^ {\\top} \\boldsymbol {R} _ {u v} \\boldsymbol {S} _ {v}\\right) = p _ {i}, \\tag {3} \\\\ \\forall i \\in \\{1, \\dots , d e g \\}. \\\\ \\end{array}", + "image_path": "05403ba1f9e6497617bf48e3e2402a02d75522e762220bfb8d6c26d326684483.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 277, + 554, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 554, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 554, + 324 + ], + "type": "text", + "content": "On the left side of Equation (3), the first polynomial term represents the cost of choosing specific intra-layer strategies for layers placed in stage " + }, + { + "bbox": [ + 313, + 277, + 554, + 324 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 277, + 554, + 324 + ], + "type": "text", + "content": ". The second term represents total resharding costs within stage " + }, + { + "bbox": [ + 313, + 277, + 554, + 324 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 277, + 554, + 324 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 325, + 554, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 554, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 554, + 373 + ], + "type": "text", + "content": "Communication-stage constraint To calculate the total cost for a single communication stage " + }, + { + "bbox": [ + 313, + 325, + 554, + 373 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 325, + 554, + 373 + ], + "type": "text", + "content": ", we should aggregate the P2P costs incurred between consecutive stages and assign them to " + }, + { + "bbox": [ + 313, + 325, + 554, + 373 + ], + "type": "inline_equation", + "content": "o_j" + }, + { + "bbox": [ + 313, + 325, + 554, + 373 + ], + "type": "text", + "content": ". This constraint can be formulated as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 365, + 380, + 555, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 380, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 365, + 380, + 555, + 425 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\sum_ {\\langle u, v \\rangle \\in \\mathbb {E}} \\boldsymbol {P} _ {u j} \\boldsymbol {P} _ {v (j + 1)} \\left(\\boldsymbol {S} _ {u} ^ {\\top} \\boldsymbol {R} _ {u v} ^ {\\prime} \\boldsymbol {S} _ {v}\\right) = o _ {j}, \\tag {4} \\\\ \\forall j \\in \\{1, \\dots , d e g - 1 \\}. \\\\ \\end{array}", + "image_path": "7ec49dd1627e10c5ba985b043afc52bd3e29d6ee7ba0b8996febb7a0fc70a918.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 432, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 555, + 479 + ], + "type": "text", + "content": "Memory constraint We need to guarantee that no devices (GPUs) will encounter out-of-memory (OOM) exceptions during training process. This constraint can be formulated as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 351, + 486, + 555, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 486, + 555, + 512 + ], + "spans": [ + { + "bbox": [ + 351, + 486, + 555, + 512 + ], + "type": "interline_equation", + "content": "\\sum_ {u \\in \\mathbb {V}} \\boldsymbol {P} _ {u i} \\boldsymbol {S} _ {u} ^ {\\top} \\boldsymbol {M} _ {u} \\leqslant m, \\forall i \\in \\{1, \\dots , d e g \\}. \\tag {5}", + "image_path": "f920ffb7067544354388bb928d98c20053e54af043ce5cf4c739283b1a90d2f0.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 519, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 554, + 567 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 519, + 554, + 567 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 519, + 554, + 567 + ], + "type": "text", + "content": " denotes the memory limit for each device. In the case of homogeneous computing devices, the value of " + }, + { + "bbox": [ + 313, + 519, + 554, + 567 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 519, + 554, + 567 + ], + "type": "text", + "content": " remains constant throughout all stages. But the value of " + }, + { + "bbox": [ + 313, + 519, + 554, + 567 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 519, + 554, + 567 + ], + "type": "text", + "content": " varies in the case of heterogeneous computing devices." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 567, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 554, + 628 + ], + "type": "text", + "content": "Order-preserving constraint PP is not a single-program multiple-data (SPMD) parallel strategy [15]. Hence, we need an order-preserving constraint to ensure that the subgraphs of " + }, + { + "bbox": [ + 313, + 567, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 567, + 554, + 628 + ], + "type": "text", + "content": " are contiguous. We adopt the definition of contiguous from previous work [38, 39]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 634, + 554, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 634, + 554, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 634, + 554, + 670 + ], + "type": "text", + "content": "Definition 3.1. A set " + }, + { + "bbox": [ + 313, + 634, + 554, + 670 + ], + "type": "inline_equation", + "content": "\\mathbb{W} \\subseteq \\mathbb{V}" + }, + { + "bbox": [ + 313, + 634, + 554, + 670 + ], + "type": "text", + "content": " is contiguous if there do not exist nodes " + }, + { + "bbox": [ + 313, + 634, + 554, + 670 + ], + "type": "inline_equation", + "content": "u \\in \\mathbb{W}, v \\in \\mathbb{V} \\setminus \\mathbb{W}" + }, + { + "bbox": [ + 313, + 634, + 554, + 670 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 634, + 554, + 670 + ], + "type": "inline_equation", + "content": "w \\in \\mathbb{W}" + }, + { + "bbox": [ + 313, + 634, + 554, + 670 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 313, + 634, + 554, + 670 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 634, + 554, + 670 + ], + "type": "text", + "content": " is reachable from " + }, + { + "bbox": [ + 313, + 634, + 554, + 670 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 313, + 634, + 554, + 670 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 634, + 554, + 670 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 634, + 554, + 670 + ], + "type": "text", + "content": " is reachable from " + }, + { + "bbox": [ + 313, + 634, + 554, + 670 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 634, + 554, + 670 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "Figure 4 illustrates an example of a contiguous set " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{W}" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": ", in which we cannot find any reachable node pairs " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\langle u, v \\rangle" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\langle v, w \\rangle" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "u, w \\in \\mathbb{W}" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "v \\in \\mathbb{V} \\setminus \\mathbb{W}" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20950" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "In our case, our model will not be assigned to different pipeline stages in a disordered manner if we ensure that all subgraphs on each computation stage are contiguous. After reformulating this constraint in linear form, we have" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 129, + 118, + 140 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 129, + 118, + 140 + ], + "spans": [ + { + "bbox": [ + 70, + 129, + 118, + 140 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} _ {v i} \\geqslant \\mathbf {P} _ {v i},", + "image_path": "c5484b1b034b476f69e8621c132e949510c9273be63dffa348197a5f69292ef5.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 72, + 144, + 296, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 144, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 72, + 144, + 296, + 156 + ], + "type": "interline_equation", + "content": "\\forall v \\in \\mathbb {V}, \\forall i \\in \\{1, 2, \\dots , d e g \\}, \\tag {6a}", + "image_path": "c9b85713f265de636e8476bac431527b985f33b228c7c86da4b6da332410da11.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 71, + 159, + 118, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 159, + 118, + 171 + ], + "spans": [ + { + "bbox": [ + 71, + 159, + 118, + 171 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} _ {v i} \\leqslant \\mathbf {Z} _ {u i},", + "image_path": "622864c91f0b68a2a4816e894b77e3fd64bbcbc91bfedac3c7df7c2664253b10.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 73, + 173, + 296, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 173, + 296, + 186 + ], + "spans": [ + { + "bbox": [ + 73, + 173, + 296, + 186 + ], + "type": "interline_equation", + "content": "\\forall u, v \\in \\mathbb {V}, \\forall \\langle u, v \\rangle \\in \\mathbb {E}, \\forall i \\in \\{1, 2, \\dots , d e g \\}, \\tag {6b}", + "image_path": "05cb1dace6cdbc681aab8f8f4d3c7afa0c4ce26ab97a58f616ec84c68d64e5dd.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 71, + 188, + 161, + 201 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 188, + 161, + 201 + ], + "spans": [ + { + "bbox": [ + 71, + 188, + 161, + 201 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} _ {v i} \\leqslant \\mathbf {P} _ {v i} - \\mathbf {P} _ {u i} + 1,", + "image_path": "8c586c922755f804a2decdaf373c65559ec73f16147e8017c2f8755902ef6090.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 73, + 203, + 296, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 203, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 73, + 203, + 296, + 217 + ], + "type": "interline_equation", + "content": "\\forall u, v \\in \\mathbb {V}, \\forall \\langle u, v \\rangle \\in \\mathbb {E}, \\forall i \\in \\{1, 2, \\dots , d e g \\}. \\tag {6c}", + "image_path": "e1226a546cde57689d9a4a4889e6dcb1c8a43272857b072103bcb60a3e57be81.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 224, + 296, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 296, + 272 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 296, + 272 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 224, + 296, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 55, + 224, + 296, + 272 + ], + "type": "text", + "content": " is an auxiliary variable shaped like " + }, + { + "bbox": [ + 55, + 224, + 296, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 55, + 224, + 296, + 272 + ], + "type": "text", + "content": ". Given the node set " + }, + { + "bbox": [ + 55, + 224, + 296, + 272 + ], + "type": "inline_equation", + "content": "\\mathbb{V}_i" + }, + { + "bbox": [ + 55, + 224, + 296, + 272 + ], + "type": "text", + "content": " that contains nodes in stage " + }, + { + "bbox": [ + 55, + 224, + 296, + 272 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 224, + 296, + 272 + ], + "type": "text", + "content": ", we define " + }, + { + "bbox": [ + 55, + 224, + 296, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{vi} = 1" + }, + { + "bbox": [ + 55, + 224, + 296, + 272 + ], + "type": "text", + "content": " if a node " + }, + { + "bbox": [ + 55, + 224, + 296, + 272 + ], + "type": "inline_equation", + "content": "w \\in \\mathbb{V}_i" + }, + { + "bbox": [ + 55, + 224, + 296, + 272 + ], + "type": "text", + "content": " is reachable from " + }, + { + "bbox": [ + 55, + 224, + 296, + 272 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 224, + 296, + 272 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 224, + 296, + 272 + ], + "type": "inline_equation", + "content": "v \\in \\mathbb{V}" + }, + { + "bbox": [ + 55, + 224, + 296, + 272 + ], + "type": "text", + "content": "). Otherwise, " + }, + { + "bbox": [ + 55, + 224, + 296, + 272 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{vi} = 0" + }, + { + "bbox": [ + 55, + 224, + 296, + 272 + ], + "type": "text", + "content": ". Detailed proof can be found in Appendix B." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 273, + 296, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 273, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 296, + 319 + ], + "type": "text", + "content": "Layer-placement constraint All layers should be placed on exactly one pipeline stage and at least one layer should be placed on each pipeline stage. This constraint can be formulated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 77, + 327, + 296, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 327, + 296, + 358 + ], + "spans": [ + { + "bbox": [ + 77, + 327, + 296, + 358 + ], + "type": "interline_equation", + "content": "\\sum_ {i = 1} ^ {\\text {d e g}} \\boldsymbol {P} _ {u i} = 1, \\quad \\forall u \\in \\mathbb {V}, \\tag {7a}", + "image_path": "37654767a864dc6650291e623f3e89741e8f261dc80a203fd890077507a05f67.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 77, + 361, + 296, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 361, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 77, + 361, + 296, + 386 + ], + "type": "interline_equation", + "content": "\\sum_ {u \\in \\mathbb {V}} \\boldsymbol {P} _ {u i} \\geqslant 1, \\quad \\forall i \\in \\{1, \\dots , d e g \\}, \\tag {7b}", + "image_path": "ab2d5e73c9ff8ff58b99e327b0c1dce1d591c3eb72e8edba1c34047b9a82c3b1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 75, + 389, + 296, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 389, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 75, + 389, + 296, + 403 + ], + "type": "interline_equation", + "content": "\\boldsymbol {P} _ {u i} \\in \\{0, 1 \\}, \\quad \\forall u \\in \\mathbb {V}, i \\in \\{1, \\dots , d e g \\}. \\tag {7c}", + "image_path": "c177d299f94bb977b4261f29ade271a9e036f35148a14126fb8a21490707b674.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 410, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 296, + 445 + ], + "type": "text", + "content": "Strategy-selection constraint Each layer must and can choose only one strategy. This constraint can be formulated as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 75, + 452, + 296, + 486 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 452, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 75, + 452, + 296, + 486 + ], + "type": "interline_equation", + "content": "\\sum_ {k = 1} ^ {| \\mathbb {S} _ {u} |} \\boldsymbol {S} _ {u k} = 1, \\quad \\forall u \\in \\mathbb {V}, \\tag {8a}", + "image_path": "9fe522cfc0b4fb49148473575663617b14ccd16184ef13ba1bbce353dcd85bdb.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 74, + 488, + 296, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 488, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 74, + 488, + 296, + 502 + ], + "type": "interline_equation", + "content": "\\boldsymbol {S} _ {u k} \\in \\{0, 1 \\}, \\quad \\forall u \\in \\mathbb {V}, k \\in \\{1, \\dots , | \\mathbb {S} _ {u} | \\}. \\tag {8b}", + "image_path": "836354bc60eec584150dc517f04484911b0159b47c3f43de23069ddeaee50792.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 509, + 296, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 296, + 545 + ], + "type": "text", + "content": "The MIQP formulation for UniAP includes the objective function in Equation (2) and all the constraints from Euqation (3) - (8b)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 552, + 212, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 212, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 212, + 566 + ], + "type": "text", + "content": "3.4. Unified optimization process" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 570, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 664 + ], + "type": "text", + "content": "UOP integrates the cost model and MIQP based on the profiling results and the computation graph to return the optimal parallel strategy and the corresponding TPI. Algorithm 1 summarizes the whole process. In Algorithm 1, we denote intra-layer cost as " + }, + { + "bbox": [ + 55, + 570, + 296, + 664 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 570, + 296, + 664 + ], + "type": "text", + "content": ", inter-layer cost as " + }, + { + "bbox": [ + 55, + 570, + 296, + 664 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 570, + 296, + 664 + ], + "type": "text", + "content": ", cross-stage cost as " + }, + { + "bbox": [ + 55, + 570, + 296, + 664 + ], + "type": "inline_equation", + "content": "R'" + }, + { + "bbox": [ + 55, + 570, + 296, + 664 + ], + "type": "text", + "content": ", and memory cost as " + }, + { + "bbox": [ + 55, + 570, + 296, + 664 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 570, + 296, + 664 + ], + "type": "text", + "content": ". The CostModeling process calculates these four costs based on the cost model described in Section 3.2." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "First, UOP optimizes intra-layer-only parallelism for cases in which pipeline parallelism is not adopted. Several works [23, 47] have used quadratic integer programming (QIP) to optimize intra-layer-only parallel strategy and" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 72, + 489, + 83 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 489, + 83 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 489, + 83 + ], + "type": "text", + "content": "Algorithm 1 Unified Optimization Process" + } + ] + } + ], + "index": 19, + "type": "text" + }, + { + "bbox": [ + 323, + 87, + 555, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 87, + 555, + 121 + ], + "spans": [ + { + "bbox": [ + 323, + 87, + 555, + 121 + ], + "type": "text", + "content": "Input: Profiling results " + }, + { + "bbox": [ + 323, + 87, + 555, + 121 + ], + "type": "inline_equation", + "content": "PR" + }, + { + "bbox": [ + 323, + 87, + 555, + 121 + ], + "type": "text", + "content": ", strategy dictionary " + }, + { + "bbox": [ + 323, + 87, + 555, + 121 + ], + "type": "inline_equation", + "content": "SD" + }, + { + "bbox": [ + 323, + 87, + 555, + 121 + ], + "type": "text", + "content": ", minibatch size " + }, + { + "bbox": [ + 323, + 87, + 555, + 121 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 323, + 87, + 555, + 121 + ], + "type": "text", + "content": ", computation graph " + }, + { + "bbox": [ + 323, + 87, + 555, + 121 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 323, + 87, + 555, + 121 + ], + "type": "text", + "content": ", and the number of GPUs " + }, + { + "bbox": [ + 323, + 87, + 555, + 121 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 323, + 87, + 555, + 121 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 323, + 123, + 555, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 123, + 555, + 159 + ], + "spans": [ + { + "bbox": [ + 323, + 123, + 555, + 159 + ], + "type": "text", + "content": "Output: Optimal cost " + }, + { + "bbox": [ + 323, + 123, + 555, + 159 + ], + "type": "inline_equation", + "content": "cost^{*}" + }, + { + "bbox": [ + 323, + 123, + 555, + 159 + ], + "type": "text", + "content": ", pipeline parallel size " + }, + { + "bbox": [ + 323, + 123, + 555, + 159 + ], + "type": "inline_equation", + "content": "deg^{*}" + }, + { + "bbox": [ + 323, + 123, + 555, + 159 + ], + "type": "text", + "content": ", the number of micro-batches " + }, + { + "bbox": [ + 323, + 123, + 555, + 159 + ], + "type": "inline_equation", + "content": "c^{*}" + }, + { + "bbox": [ + 323, + 123, + 555, + 159 + ], + "type": "text", + "content": ", layer placement " + }, + { + "bbox": [ + 323, + 123, + 555, + 159 + ], + "type": "inline_equation", + "content": "P^{*}" + }, + { + "bbox": [ + 323, + 123, + 555, + 159 + ], + "type": "text", + "content": ", and intra-layer strategy " + }, + { + "bbox": [ + 323, + 123, + 555, + 159 + ], + "type": "inline_equation", + "content": "S^{*}" + }, + { + "bbox": [ + 323, + 123, + 555, + 159 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 324, + 159, + 368, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 159, + 368, + 171 + ], + "spans": [ + { + "bbox": [ + 324, + 159, + 368, + 171 + ], + "type": "interline_equation", + "content": "d e g ^ {*} = 1;", + "image_path": "edb772d1b0cf14d2f4fd0b257e5835164a5f8200737db09b7b19d846e475c160.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 325, + 172, + 360, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 172, + 360, + 182 + ], + "spans": [ + { + "bbox": [ + 325, + 172, + 360, + 182 + ], + "type": "interline_equation", + "content": "c ^ {*} = B;", + "image_path": "09bc992afc10a07ca2444f8f3de0a21af07d3b1077cf84c4fe6192ac08b183d6.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 325, + 183, + 535, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 183, + 535, + 194 + ], + "spans": [ + { + "bbox": [ + 325, + 183, + 535, + 194 + ], + "type": "interline_equation", + "content": "A, R, \\_ , M = \\operatorname {C o s t M o d e l i n g} (P R, S D [ 1 ], \\mathcal {G}, B);", + "image_path": "611ef99142ca4a89f248c1dc9648580405e5d7d7d8788a58f3c33f3c98ce178f.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 325, + 195, + 454, + 206 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 195, + 454, + 206 + ], + "spans": [ + { + "bbox": [ + 325, + 195, + 454, + 206 + ], + "type": "interline_equation", + "content": "c o s t ^ {*}, P ^ {*}, S ^ {*} = \\mathbb {Q} \\operatorname {I P} (\\boldsymbol {A}, \\boldsymbol {R}, \\boldsymbol {M});", + "image_path": "2d7a697f9cb73c83c71e05be1fdd4873b8eec81ac9c8cb97a25effe9afc7a3e7.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 324, + 207, + 542, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 207, + 542, + 219 + ], + "spans": [ + { + "bbox": [ + 324, + 207, + 542, + 219 + ], + "type": "text", + "content": "Get all factors for " + }, + { + "bbox": [ + 324, + 207, + 542, + 219 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 324, + 207, + 542, + 219 + ], + "type": "text", + "content": " except 1 and insert them to set " + }, + { + "bbox": [ + 324, + 207, + 542, + 219 + ], + "type": "inline_equation", + "content": "\\mathbb{F}" + }, + { + "bbox": [ + 324, + 207, + 542, + 219 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 324, + 220, + 544, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 220, + 544, + 230 + ], + "spans": [ + { + "bbox": [ + 324, + 220, + 544, + 230 + ], + "type": "text", + "content": "Get all factors for " + }, + { + "bbox": [ + 324, + 220, + 544, + 230 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 324, + 220, + 544, + 230 + ], + "type": "text", + "content": " except 1 and insert them to set " + }, + { + "bbox": [ + 324, + 220, + 544, + 230 + ], + "type": "inline_equation", + "content": "\\mathbb{B}" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 325, + 231, + 390, + 242 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 231, + 390, + 242 + ], + "spans": [ + { + "bbox": [ + 325, + 231, + 390, + 242 + ], + "type": "interline_equation", + "content": "\\text {f o r} d e g \\text {i n} \\mathbb {F} \\text {d o}", + "image_path": "928b8b0844fedf1e5ee85e724d64a4f9bf45e2259d158a6974c4f5e26540fb34.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 335, + 243, + 389, + 253 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 243, + 389, + 253 + ], + "spans": [ + { + "bbox": [ + 335, + 243, + 389, + 253 + ], + "type": "interline_equation", + "content": "f o r c i n \\mathbb {B} d o", + "image_path": "34e297343a5fa2b224620cb7f4c209c2609eb473e5c53ab3ea7d4e1c99477d57.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 345, + 255, + 454, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 255, + 454, + 266 + ], + "spans": [ + { + "bbox": [ + 345, + 255, + 454, + 266 + ], + "type": "interline_equation", + "content": "\\text {M i c r o - b a t c h s i z e} b = B / c;", + "image_path": "ebc5f113981a07e793e457f33adaf7fbae0491a0712f1c2304f90e3afd692f77.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 345, + 267, + 555, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 267, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 345, + 267, + 555, + 289 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {A}, \\boldsymbol {R}, \\boldsymbol {R} ^ {\\prime}, \\boldsymbol {M} = \\text {C o s t M o d e l i n g} (P R, S D [ d e g ], \\mathcal {G}, \\\\ b); \\end{array}", + "image_path": "c72a0aabf71942be8a716383546169612257a7db436b6b4dcd66e14695101685.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 345, + 291, + 506, + 302 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 291, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 345, + 291, + 506, + 302 + ], + "type": "interline_equation", + "content": "c o s t, \\boldsymbol {P}, \\boldsymbol {S} = \\operatorname {M I Q P} (\\boldsymbol {A}, \\boldsymbol {R}, \\boldsymbol {R} ^ {\\prime}, \\boldsymbol {M}, \\deg , c);", + "image_path": "5909c408f4691c8bce2bed517060b01f6c74d808a18cd3f082d598e66efdafa8.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 345, + 304, + 430, + 313 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 304, + 430, + 313 + ], + "spans": [ + { + "bbox": [ + 345, + 304, + 430, + 313 + ], + "type": "interline_equation", + "content": "\\text {i f} \\quad c o s t < c o s t ^ {*} \\text {t h e n}", + "image_path": "0dd07ea990bef3634306e0eb4299f96b72693b136b7a7ae81eda835bc1c597b4.jpg" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 356, + 315, + 529, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 315, + 529, + 327 + ], + "spans": [ + { + "bbox": [ + 356, + 315, + 529, + 327 + ], + "type": "interline_equation", + "content": "c o s t ^ {*}, d e g ^ {*}, c ^ {*}, P ^ {*}, S ^ {*} = c o s t, d e g, c, P, S;", + "image_path": "015bbec8c0b4c813af0787b025283f622b37b610aea55afc274b55ab15bf80e9.jpg" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 345, + 327, + 372, + 337 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 327, + 372, + 337 + ], + "spans": [ + { + "bbox": [ + 345, + 327, + 372, + 337 + ], + "type": "interline_equation", + "content": "e n d i f", + "image_path": "e67c68c76b19681fef2284ef32ba18a4f999bad42229701d8803ad5b7a33b97d.jpg" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 335, + 339, + 368, + 349 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 339, + 368, + 349 + ], + "spans": [ + { + "bbox": [ + 335, + 339, + 368, + 349 + ], + "type": "interline_equation", + "content": "e n d \\mathrm {f o r}", + "image_path": "297aa4b41660bb07da630d0be1bb42091aaab270c613c538528271db6b8c9535.jpg" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 325, + 351, + 358, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 351, + 358, + 361 + ], + "spans": [ + { + "bbox": [ + 325, + 351, + 358, + 361 + ], + "type": "interline_equation", + "content": "e n d \\mathrm {f o r}", + "image_path": "6b37b78500a694582b3a49f7bb89cf0f2d4c27f8e79577bafe4627f180d07fd7.jpg" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 386, + 555, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 555, + 410 + ], + "type": "text", + "content": "achieved promising results. UniAP provides a QIP formulation for intra-layer-only parallelism in Appendix C." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 410, + 554, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 554, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 554, + 470 + ], + "type": "text", + "content": "Then, UOP enumerates all factors of " + }, + { + "bbox": [ + 313, + 410, + 554, + 470 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 410, + 554, + 470 + ], + "type": "text", + "content": " except 1 as the pipeline parallel size " + }, + { + "bbox": [ + 313, + 410, + 554, + 470 + ], + "type": "inline_equation", + "content": "deg" + }, + { + "bbox": [ + 313, + 410, + 554, + 470 + ], + "type": "text", + "content": ". For each " + }, + { + "bbox": [ + 313, + 410, + 554, + 470 + ], + "type": "inline_equation", + "content": "deg" + }, + { + "bbox": [ + 313, + 410, + 554, + 470 + ], + "type": "text", + "content": ", UOP enumerates all factors of " + }, + { + "bbox": [ + 313, + 410, + 554, + 470 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 410, + 554, + 470 + ], + "type": "text", + "content": " except 1 as the number of micro-batches " + }, + { + "bbox": [ + 313, + 410, + 554, + 470 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 410, + 554, + 470 + ], + "type": "text", + "content": ". These enumerations aim to achieve load balance on a homogeneous cluster." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 471, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 554, + 518 + ], + "type": "text", + "content": "For each candidate " + }, + { + "bbox": [ + 313, + 471, + 554, + 518 + ], + "type": "inline_equation", + "content": "deg" + }, + { + "bbox": [ + 313, + 471, + 554, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 471, + 554, + 518 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 471, + 554, + 518 + ], + "type": "text", + "content": ", UOP formulates the cost for a training iteration to an MIQP expression. It then waits for the MIQP solver to return the optimal cost and parallel strategy." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 519, + 555, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 555, + 578 + ], + "type": "text", + "content": "Finally, UOP returns the minimum cost " + }, + { + "bbox": [ + 313, + 519, + 555, + 578 + ], + "type": "inline_equation", + "content": "cost^*" + }, + { + "bbox": [ + 313, + 519, + 555, + 578 + ], + "type": "text", + "content": " and its corresponding pipeline parallel size " + }, + { + "bbox": [ + 313, + 519, + 555, + 578 + ], + "type": "inline_equation", + "content": "deg^*" + }, + { + "bbox": [ + 313, + 519, + 555, + 578 + ], + "type": "text", + "content": ", number of microbatches " + }, + { + "bbox": [ + 313, + 519, + 555, + 578 + ], + "type": "inline_equation", + "content": "c^*" + }, + { + "bbox": [ + 313, + 519, + 555, + 578 + ], + "type": "text", + "content": ", layer placement " + }, + { + "bbox": [ + 313, + 519, + 555, + 578 + ], + "type": "inline_equation", + "content": "P^*" + }, + { + "bbox": [ + 313, + 519, + 555, + 578 + ], + "type": "text", + "content": ", and intra-layer strategies " + }, + { + "bbox": [ + 313, + 519, + 555, + 578 + ], + "type": "inline_equation", + "content": "S^*" + }, + { + "bbox": [ + 313, + 519, + 555, + 578 + ], + "type": "text", + "content": ". We provide visualization for a candidate solution to UOP in Appendix D." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 313, + 587, + 431, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 431, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 431, + 601 + ], + "type": "text", + "content": "3.5. Complexity analysis" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "inline_equation", + "content": "|\\mathbb{V}|" + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "inline_equation", + "content": "|\\mathbb{S}|" + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": " denote the number of layers, parallel strategies, and GPUs, respectively. As illustrated in Algorithm 1, UniAP enumerates all factors of " + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": " except 1 as the " + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "inline_equation", + "content": "deg" + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": " in the outer loop and all factors of " + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": " except 1 in the inner loop. The time complexity for the two loops in the UniAP algorithm is " + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\sqrt{Bn})" + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": ". Within the inner loop body, UniAP calls CostModeling to model the cost of each stage for each parallel strategy. Furthermore, the optimization time limit of the MIQP solver can be set as a constant" + } + ] + } + ], + "index": 43 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20951" + } + ] + } + ], + "index": 44 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "hyperparameter when UniAP calls it. Therefore, the overall computational complexity of UniAP is " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(|\\mathbb{V}||\\mathbb{S}|\\sqrt{Bn})" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 109, + 132, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 109, + 132, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 109, + 132, + 122 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 129, + 162, + 142 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 129, + 162, + 142 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 162, + 142 + ], + "type": "text", + "content": "4.1. Experiment setup" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 146, + 297, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 146, + 297, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 297, + 289 + ], + "type": "text", + "content": "We conduct experiments on four different kinds of environments to validate the effectiveness and universality of our method. ENVA refers to a node (machine) with 1 Xeon 6248 CPU, 8 V100-SXM2 32GB GPUs, and 472GB memory. ENVB refers to 4 nodes interconnected with 10Gbps networks and each node has 2 Xeon E5-2620 v4 CPUs, 4 TITAN Xp 12GB GPUs, and 125GB memory. ENVC refers to a node with 8 A100 40GB PCIe GPUs. ENVD is a cloud cluster with 16 nodes and each node has 4 DCUs (a type of non-NVIDIA GPU). In the following text, we specify the number of GPUs after the environment name (e.g., ENVB-8) only when partial nodes of the environment are used." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 290, + 296, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 296, + 411 + ], + "type": "text", + "content": "We evaluate UniAP with five Transformer-based models, BERT-Huge [5], T5-Large [31], ViT-Huge [6], SwinHuge [24], and Llama [40, 41]. We follow the common practice of training these transformer-based models. To eliminate factors that affect training throughput, we turn off techniques orthogonal to parallel strategies, such as activation checkpointing [3]. However, we integrate FP16 mixed precision training [26] for the largest model, Llama, in some cases to successfully orchestrate the process. More details on these models are provided in Appendix E." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 411, + 296, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 296, + 458 + ], + "type": "text", + "content": "Several solvers can be used for QIP and MIQP optimization. In our implementation, we choose Gurobi [13], with detailed configurations provided in Appendix E. Exploring other solvers is left for future work." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 459, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 296, + 544 + ], + "type": "text", + "content": "The experimental evaluation concentrates on two primary metrics: training throughput and strategy optimization time. The former is calculated by averaging throughput from the 10th to the 60th iteration of training, while the latter is determined by measuring the time of the UOP. More details are provided in Appendix E. Besides, we provide the accuracy of our cost model in Appendix F." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 552, + 295, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 295, + 577 + ], + "type": "text", + "content": "4.2. Training throughput and strategy optimization time" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "content": "We compare the training throughput and strategy optimization time of UniAP with those of the baselines on ENVA, ENVB-8, ENVC, and ENVD-32. For experiments on ENVA, we set the mini-batch size to be 32, 16, 128, and 128 for BERT, T5, ViT, and Swin, respectively. For experiments on ENVB-8, we set the mini-batch size to be 16, 8, 64, and 32 for these four models, respectively. For the Llama model [40, 41] run on ENVC, we set the mini-batch size to be 8. For experiments on ENVA, ENVB-8, and ENVC, we choose Galvatron [25] and Alma [47] as baselines because they have achieved SOTA performance. Specifically," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": "Galvatron has surpassed other methods, including PyTorch DDP [22], Megatron-LM [30], FSDP [8, 32], GPipe [15], and DeepSpeed 3D [27] in terms of training throughput, as reported in the original paper [25]. Furthermore, Alpa utilizes the Just-In-Time (JIT) compilation feature in JAX and outperforms Megatron-LM and DeepSpeed." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 145, + 556, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 145, + 556, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 145, + 556, + 277 + ], + "type": "text", + "content": "For benchmarks on ENVD-32, pytorch libraries and frameworks need specialized adaptations due to the heterogeneity of DCUs against NVIDIA GPUs. Thus, this part of experiments is based on adapted pytorch libraries and Megatron-DeepSpeed[27] framework provided by the platform developers, and we select MP methods including Megatron [30] and DeepSpeed (ZeRO-3) [34] as our baseline methods. We set the mini-batch size as 8 for Llama-7B and 4 for Llama-13B. The remaining configurations align with those for Llama-7B in Section 4.1, such as activating FP16 mixed precision training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 278, + 556, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 556, + 446 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 556, + 446 + ], + "type": "text", + "content": "Table 1 shows the training throughput and strategy optimization time on ENVA, ENVB-8, and ENVC. On ENVA, UniAP and Galvatron get the same optimal strategy for BERT-Huge, T5-Large, and ViT-Huge, outperforming Alpa in terms of training throughput and strategy optimization time. In addition, UniAP finds a solution for Swin-Huge, while Galvatron encounters CUDA OOM issues. In particular, UniAP achieves a maximum optimization speedup that is " + }, + { + "bbox": [ + 313, + 278, + 556, + 446 + ], + "type": "inline_equation", + "content": "17 \\times" + }, + { + "bbox": [ + 313, + 278, + 556, + 446 + ], + "type": "text", + "content": " faster than Galvatron and hundreds of times faster than Alpa on BERT-Huge. This is mainly due to the ability of the MIQP solver to search for an optimal strategy on multiple threads, while the dynamic programming based methods like Galvatron and Alpa run on a single thread due to their strong data dependency." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 447, + 556, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 556, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 556, + 519 + ], + "type": "text", + "content": "On ENVB-8, UniAP consistently demonstrates competitive or larger training throughput compared to Galvatron and Alpa. We attribute the performance improvement to UniAP's larger strategy space. A detailed study is provided in Section 4.4. Furthermore, UniAP's strategy optimization time is also significantly shorter than the two baseline methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 521, + 556, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 663 + ], + "type": "text", + "content": "On ENVC, UniAP shows an optimization speedup of " + }, + { + "bbox": [ + 313, + 521, + 556, + 663 + ], + "type": "inline_equation", + "content": "11.83 \\times" + }, + { + "bbox": [ + 313, + 521, + 556, + 663 + ], + "type": "text", + "content": " and a training speedup of " + }, + { + "bbox": [ + 313, + 521, + 556, + 663 + ], + "type": "inline_equation", + "content": "3.80 \\times" + }, + { + "bbox": [ + 313, + 521, + 556, + 663 + ], + "type": "text", + "content": " compared to Galvatron on Llama. We examine the parallel strategy they adopted. UniAP employs an 8-stage PP with a micro-batch size of 1, whereas Galvatron employs a 4-stage PP with a micro-batch size of 3 (with the final micro-batch containing 2 samples). Within each PP stage, Galvatron utilizes a 2-way TP. For ENVC, which is equipped with 8 A100 40GB PCIe GPUs, minimizing communication volume is critical. Since PP has significantly less inter-device communication volume than TP, the parallel strategy discovered by UniAP is more reasonable than Galvatron in this case." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": "Table 2 shows the results on ENVD-32. On this environment, UniAP consistently identifies the fastest parallel strategies while expending considerably less time on strategy optimization compared to Megatron (about " + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "inline_equation", + "content": "157 \\times" + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": " faster for" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20952" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 109, + 70, + 502, + 368 + ], + "blocks": [ + { + "bbox": [ + 109, + 70, + 502, + 368 + ], + "lines": [ + { + "bbox": [ + 109, + 70, + 502, + 368 + ], + "spans": [ + { + "bbox": [ + 109, + 70, + 502, + 368 + ], + "type": "table", + "html": "
Env.ModelTraining throughput (samples/s)Minimum speedupMaximum speedup
GalvatronAlpaUniAP
ENVABERT-Huge33.46 ± 0.2831.56 ± 0.0433.46 ± 0.281.001.06
T5-Large23.29 ± 0.04MEM×2)23.29 ± 0.041.001.00
ViT-Huge109.51 ± 0.0797.66 ± 1.42109.51 ± 0.071.001.12
Swin-HugeCUDA×3)N/Å4)67.96 ± 0.12N/Å4)N/Å4)
ENVB-8BERT-Huge6.27 ± 0.178.95 ± 0.0610.77 ± 0.131.201.71
T5-Large1)8.06 ± 0.06MEM×2)7.98 ± 0.050.990.99
ViT-Huge32.20 ± 0.1738.74 ± 0.2045.58 ± 0.541.181.41
Swin-Huge13.90 ± 0.17N/Å4)19.08 ± 0.101.371.37
ENVCLlama-7B1.22 ± 0.01N/Å4)4.63 ± 0.0073.803.80
Env.ModelStrategy optimization time (min.)Minimum speedupMaximum speedup
GalvatronAlpaUniAP
ENVABERT-Huge6.44 ± 0.588>400.37 ± 0.00217.29>107.41
T5-Large12.41 ± 0.122MEM×2)0.89 ± 0.00713.9813.98
ViT-Huge6.29 ± 0.464>400.57 ± 0.00910.95>69.60
Swin-Huge11.88 ± 0.666N/Å4)2.16 ± 0.0045.495.49
ENVB-8BERT-Huge2.04 ± 0.010>401.51 ± 0.0051.34>26.32
T5-Large1)2.64 ± 0.110MEM×2)0.91 ± 0.0052.902.90
ViT-Huge2.37 ± 0.180>401.11 ± 0.0112.14>36.01
Swin-Huge4.29 ± 0.320N/Å4)2.29 ± 0.0101.871.87
ENVCLlama-7B6.84 ± 0.055N/Å4)0.58 ± 0.00611.8311.83
", + "image_path": "43821e5e3eb06fadba1b6ed9361354782a15a94aaa847b3ab317f6cf365e7786.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 116, + 369, + 439, + 379 + ], + "lines": [ + { + "bbox": [ + 116, + 369, + 439, + 379 + ], + "spans": [ + { + "bbox": [ + 116, + 369, + 439, + 379 + ], + "type": "text", + "content": "1) T5-Large tested on ENVB is restricted to 16/16 layers to avoid out-of-memory (OOM) exceptions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 116, + 379, + 299, + 388 + ], + "lines": [ + { + "bbox": [ + 116, + 379, + 299, + 388 + ], + "spans": [ + { + "bbox": [ + 116, + 379, + 299, + 388 + ], + "type": "text", + "content": "2) MEM×: OOM exceptions during strategy optimization." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 116, + 388, + 308, + 398 + ], + "lines": [ + { + "bbox": [ + 116, + 388, + 308, + 398 + ], + "spans": [ + { + "bbox": [ + 116, + 388, + 308, + 398 + ], + "type": "text", + "content": "3) CUDA×: CUDA OOM exceptions during model training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 116, + 398, + 500, + 417 + ], + "lines": [ + { + "bbox": [ + 116, + 398, + 500, + 417 + ], + "spans": [ + { + "bbox": [ + 116, + 398, + 500, + 417 + ], + "type": "text", + "content": "4) The official implementation for Swin-Huge and Llama in Alpa is absent. We have endeavored to adopt the code, yet encountered compilation errors. Consequently, experiments related to these models on Alpa are marked as N/A." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 109, + 444, + 501, + 506 + ], + "blocks": [ + { + "bbox": [ + 136, + 422, + 472, + 432 + ], + "lines": [ + { + "bbox": [ + 136, + 422, + 472, + 432 + ], + "spans": [ + { + "bbox": [ + 136, + 422, + 472, + 432 + ], + "type": "text", + "content": "Table 1. Training throughput and strategy optimization time on ENVA, ENVB-8, and ENVC." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 109, + 444, + 501, + 506 + ], + "lines": [ + { + "bbox": [ + 109, + 444, + 501, + 506 + ], + "spans": [ + { + "bbox": [ + 109, + 444, + 501, + 506 + ], + "type": "table", + "html": "
ModelTraining throughput (samples/s)Strategy optimization time (min.)
MegatronDeepSpeedUniAPMegatronDeepSpeedUniAP
Llama-7B2.01 ± 0.005\\( SOL \\times^{1)} \\)2.01 ± 0.005>8.0 hours\\( SOL \\times^{1)} \\)3.07 ± 0.121
Llama-13B0.82 ± 0.001\\( SOL \\times^{1)} \\)0.82 ± 0.001>2.5 hours\\( SOL \\times^{1)} \\)1.95 ± 0.076
", + "image_path": "8178c288e8c8274ea3f1b842870bfa4c802ffbae6fd173dfcc76569c5e4db385.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 116, + 506, + 276, + 517 + ], + "lines": [ + { + "bbox": [ + 116, + 506, + 276, + 517 + ], + "spans": [ + { + "bbox": [ + 116, + 506, + 276, + 517 + ], + "type": "text", + "content": "1) " + }, + { + "bbox": [ + 116, + 506, + 276, + 517 + ], + "type": "inline_equation", + "content": "\\mathrm{SOL} \\times" + }, + { + "bbox": [ + 116, + 506, + 276, + 517 + ], + "type": "text", + "content": ": No solution after strategy optimization." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 7 + }, + { + "bbox": [ + 168, + 521, + 440, + 533 + ], + "lines": [ + { + "bbox": [ + 168, + 521, + 440, + 533 + ], + "spans": [ + { + "bbox": [ + 168, + 521, + 440, + 533 + ], + "type": "text", + "content": "Table 2. Training throughput and strategy optimization time on ENVD-32." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 539, + 297, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 539, + 297, + 684 + ], + "spans": [ + { + "bbox": [ + 54, + 539, + 297, + 684 + ], + "type": "text", + "content": "Llama-7B and " + }, + { + "bbox": [ + 54, + 539, + 297, + 684 + ], + "type": "inline_equation", + "content": "77 \\times" + }, + { + "bbox": [ + 54, + 539, + 297, + 684 + ], + "type": "text", + "content": " faster for Llama-13B). Please note that the strategy optimization time Megatron needs for Llama-7B surpasses that for Llama-13B. We attribute this phenomenon to the variations in the mini-batch size. Specifically, the escalation from a mini-batch size of 4 (employed in Llama-13B) to 8 (employed in Llama-7B) leads to a rise in the number of candidate parallel strategies, as well as the parallel strategies that will successfully train the model. As a result, the strategy optimization time will become increasingly long when the mini-batch size exceeds 8 in the pretraining scenario [1, 40, 41]. In this situation, UniAP will identify the optimal parallel strategy much faster." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Furthermore, our experiments highlight a limitation encountered with DeepSpeed (ZeRO-3). It requires the mini" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 539, + 555, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 555, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 555, + 586 + ], + "type": "text", + "content": "batch size to be divisible evenly by the total number of computation devices. This specific prerequisite prevents DeepSpeed from successfully launching the training process with 32 DCUs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 588, + 555, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 555, + 613 + ], + "type": "text", + "content": "To facilitate further discussions, we provide a case study of the optimal parallel strategy in Appendix G." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 623, + 386, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 386, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 386, + 635 + ], + "type": "text", + "content": "4.3. Scalability" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "type": "text", + "content": "We study the scalability of UniAP using BERT, T5, ViT, and Swin on ENVB. We set the mini-batch sizes for each data point as 8, 4, 32, and 16 times the number of nodes (denoted as '#nodes'). The experimental results are shown in Figure 5. In Figure 5a, the training throughput of the optimal strategy demonstrate near-linearity as the number of nodes and mini" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20953" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 99, + 70, + 512, + 133 + ], + "blocks": [ + { + "bbox": [ + 99, + 70, + 512, + 133 + ], + "lines": [ + { + "bbox": [ + 99, + 70, + 512, + 133 + ], + "spans": [ + { + "bbox": [ + 99, + 70, + 512, + 133 + ], + "type": "table", + "html": "
ModelBatch sizeTraining throughput (samples/s)#infeasible5)#candidate6)
1 st.1)2 nd.2)Slowest3)Median4)
Llama-7B82.011.920.220.824164
Llama-13B40.820.580.270.424248
", + "image_path": "66722c8fe49e8b37355d8087928ffce42c0cd8255580b2159f01564323db92f2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 105, + 133, + 338, + 143 + ], + "lines": [ + { + "bbox": [ + 105, + 133, + 338, + 143 + ], + "spans": [ + { + "bbox": [ + 105, + 133, + 338, + 143 + ], + "type": "text", + "content": "1) 1 st.: The training throughput achieved by the fastest parallel strategy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 105, + 143, + 364, + 152 + ], + "lines": [ + { + "bbox": [ + 105, + 143, + 364, + 152 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 364, + 152 + ], + "type": "text", + "content": "2) 2 nd.: The training throughput achieved by the second fastest parallel strategy." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 106, + 152, + 353, + 162 + ], + "lines": [ + { + "bbox": [ + 106, + 152, + 353, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 152, + 353, + 162 + ], + "type": "text", + "content": "3) Slowest: The training throughput achieved by the slowest parallel strategy." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 106, + 162, + 479, + 171 + ], + "lines": [ + { + "bbox": [ + 106, + 162, + 479, + 171 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 479, + 171 + ], + "type": "text", + "content": "4) Median: The median value of training throughputs across parallel strategies that will successfully train the model." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 106, + 171, + 501, + 180 + ], + "lines": [ + { + "bbox": [ + 106, + 171, + 501, + 180 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 501, + 180 + ], + "type": "text", + "content": "5) #infeasible: The number of parallel strategies that will encounter exceptions such as CUDA OOM during model training." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 106, + 180, + 310, + 190 + ], + "lines": [ + { + "bbox": [ + 106, + 180, + 310, + 190 + ], + "spans": [ + { + "bbox": [ + 106, + 180, + 310, + 190 + ], + "type": "text", + "content": "6) #candidate: The total number of candidate parallel strategies." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 63, + 213, + 172, + 278 + ], + "blocks": [ + { + "bbox": [ + 63, + 213, + 172, + 278 + ], + "lines": [ + { + "bbox": [ + 63, + 213, + 172, + 278 + ], + "spans": [ + { + "bbox": [ + 63, + 213, + 172, + 278 + ], + "type": "image", + "image_path": "004e96860ab7f6150ce8899db71deaa4401cda5a214c3fa308f9c6d4b73b679f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 281, + 156, + 291 + ], + "lines": [ + { + "bbox": [ + 78, + 281, + 156, + 291 + ], + "spans": [ + { + "bbox": [ + 78, + 281, + 156, + 291 + ], + "type": "text", + "content": "(a) Training throughput." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 179, + 213, + 293, + 279 + ], + "blocks": [ + { + "bbox": [ + 179, + 213, + 293, + 279 + ], + "lines": [ + { + "bbox": [ + 179, + 213, + 293, + 279 + ], + "spans": [ + { + "bbox": [ + 179, + 213, + 293, + 279 + ], + "type": "image", + "image_path": "a83ef687e1bdd872c0500a20f62216c10a3a2e292f630d8d6dfa4a39e8fc7f70.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 185, + 281, + 284, + 291 + ], + "lines": [ + { + "bbox": [ + 185, + 281, + 284, + 291 + ], + "spans": [ + { + "bbox": [ + 185, + 281, + 284, + 291 + ], + "type": "text", + "content": "(b) Strategy optimization time." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 293, + 297, + 315 + ], + "lines": [ + { + "bbox": [ + 55, + 293, + 297, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 297, + 315 + ], + "type": "text", + "content": "Figure 5. Scalability on training throughput and strategy optimization time." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 62, + 331, + 290, + 399 + ], + "blocks": [ + { + "bbox": [ + 181, + 194, + 427, + 206 + ], + "lines": [ + { + "bbox": [ + 181, + 194, + 427, + 206 + ], + "spans": [ + { + "bbox": [ + 181, + 194, + 427, + 206 + ], + "type": "text", + "content": "Table 3. Statistics on the candidate parallel strategies for Megatron." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 331, + 290, + 399 + ], + "lines": [ + { + "bbox": [ + 62, + 331, + 290, + 399 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 290, + 399 + ], + "type": "table", + "html": "
Number of DCUsTraining throughput (samples/s)Strategy optimization time(min.)
160.1726 ± 0.00050.75 ± 0.0100
320.3393 ± 0.00041.64 ± 0.0321
640.6656 ± 0.00231.55 ± 0.0058
", + "image_path": "0f57b362b266a39694d18e25280700986792601dea87ce1deaac3ae0173fbf87.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 407, + 293, + 418 + ], + "lines": [ + { + "bbox": [ + 57, + 407, + 293, + 418 + ], + "spans": [ + { + "bbox": [ + 57, + 407, + 293, + 418 + ], + "type": "text", + "content": "Table 4. Scalability on training throughput among DCU clusters." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 432, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 295, + 456 + ], + "type": "text", + "content": "batch size increase. In Figure 5b, the strategy optimization time matches the complexity analysis in Section 3.5." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 457, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 296, + 540 + ], + "type": "text", + "content": "Additionally, we use DCU nodes in ENVD to scale UniAP to larger clusters on Llama-7B, with precision of FP32. We conducted experiments on clusters with 16, 32, and 64 DCUs, where we set the mini-batch size to " + }, + { + "bbox": [ + 55, + 457, + 296, + 540 + ], + "type": "inline_equation", + "content": "1/8" + }, + { + "bbox": [ + 55, + 457, + 296, + 540 + ], + "type": "text", + "content": " of the number of DCUs (i.e., for a cluster with 16 DCUs, we set the minibatch size to 2). The experimental results are shown in Table 4, further demonstrating the near-linear scalability of UniAP." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 552, + 279, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 279, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 279, + 564 + ], + "type": "text", + "content": "4.4. Further study on the significance of UniAP" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "text", + "content": "Finally, we investigate the significance of UniAP by answering two questions. The first question is why we need AP rather than MP. We examine the statistics on the candidate parallel strategies for Megatron, as shown in Table 3. With MP, most users are unfamiliar with the complex hyperparameter detail of the candidate parallel strategies and typically choose a parallel strategy at random, which will result in a " + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "64.1\\%" + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "text", + "content": " (41/64) chance of failing to train Llama-7B and an " + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "87.5\\%" + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "text", + "content": " (42/48) chance of failing to train Llama-13B even when the hardware resources are sufficient. Even if they successfully eliminate all infeasible parallel strategies, they still only have a " + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "text", + "content": " chance of identifying a parallel strata" + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 324, + 216, + 542, + 302 + ], + "blocks": [ + { + "bbox": [ + 324, + 216, + 542, + 302 + ], + "lines": [ + { + "bbox": [ + 324, + 216, + 542, + 302 + ], + "spans": [ + { + "bbox": [ + 324, + 216, + 542, + 302 + ], + "type": "image", + "image_path": "6a3ed920a1cd6fc2b02003b23020df2c91b40ad6c6921d0d8b5b17713eb59e0a.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 310, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 313, + 310, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 555, + 344 + ], + "type": "text", + "content": "Figure 6. Ablation study. " + }, + { + "bbox": [ + 313, + 310, + 555, + 344 + ], + "type": "inline_equation", + "content": "\\mathrm{SOL} \\times" + }, + { + "bbox": [ + 313, + 310, + 555, + 344 + ], + "type": "text", + "content": " indicates no feasible solution during strategy optimization, and " + }, + { + "bbox": [ + 313, + 310, + 555, + 344 + ], + "type": "inline_equation", + "content": "\\mathrm{CUDA} \\times" + }, + { + "bbox": [ + 313, + 310, + 555, + 344 + ], + "type": "text", + "content": " indicates CUDA OOM exceptions during training." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 352, + 555, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 555, + 436 + ], + "type": "text", + "content": "egy with training throughput better than 0.82 samples/s for Llama-7B and 0.42 samples/s for Llama-13B. When the users can't identify the fastest parallel strategy out of hundreds of candidates, they may sacrifice at least " + }, + { + "bbox": [ + 313, + 352, + 555, + 436 + ], + "type": "inline_equation", + "content": "4.4\\%" + }, + { + "bbox": [ + 313, + 352, + 555, + 436 + ], + "type": "text", + "content": " of the training throughput for Llama-7B and " + }, + { + "bbox": [ + 313, + 352, + 555, + 436 + ], + "type": "inline_equation", + "content": "29.2\\%" + }, + { + "bbox": [ + 313, + 352, + 555, + 436 + ], + "type": "text", + "content": " for Llama-13B. In such circumstances, UniAP is able to identify the fastest strategy, as shown in Table 2." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 437, + 556, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 556, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 556, + 594 + ], + "type": "text", + "content": "The second question is why UniAP achieves better performance than other AP methods. We investigate the importance of the strategy space for the optimality of parallel strategies with an ablation study. Specifically, we constrain the strategy space to inter-layer-only and intra-layer-only strategies and evaluate the training throughput of the resulting optimal strategy on ENVB. We set the mini-batch sizes to be 16, 12, 64, and 32, respectively. Results are shown in Figure 6. We can find that constraining the strategy space compromises the optimality of parallel strategies or gets strategies that encounter OOM across different models. Hence, unifying inter- and intra-layer AP for joint optimization makes UniAP outperform other AP methods." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 608, + 388, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 608, + 388, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 608, + 388, + 620 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we propose a novel AP method called UniAP to unify inter- and intra-layer AP by MIQP. To the best of our knowledge, UniAP is the first parallel method that can jointly optimize the two categories of parallel strategies to find an optimal solution. Experimental results show that UniAP can outperform other state-of-the-art baselines to achieve the best performance." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20954" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 137 + ], + "type": "text", + "content": "This work is supported by National Key R&D Program of China (No.2020YFA0713900), NSFC Project (No.12326615), the Key Major Project of the Pengcheng Laboratory (No.PCL2024A06)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 159, + 115, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 159, + 115, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 159, + 115, + 171 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 178, + 298, + 713 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 61, + 178, + 297, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 178, + 297, + 300 + ], + "spans": [ + { + "bbox": [ + 61, + 178, + 297, + 300 + ], + "type": "text", + "content": "[1] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language Models are Few-Shot Learners. In Advances in Neural Information Processing Systems 33, pages 1877-1901, 2020. 1, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 301, + 296, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 301, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 61, + 301, + 296, + 356 + ], + "type": "text", + "content": "[2] Zhenkun Cai, Xiao Yan, Kaihao Ma, Yidi Wu, Yuzhen Huang, James Cheng, Teng Su, and Fan Yu. TensorOpt: Exploring the Tradeoffs in Distributed DNN Training With AutoParallelism. IEEE Transactions on Parallel and Distributed Systems, 33(8):1967-1981, 2022. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 357, + 296, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 357, + 296, + 390 + ], + "spans": [ + { + "bbox": [ + 61, + 357, + 296, + 390 + ], + "type": "text", + "content": "[3] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training Deep Nets with Sublinear Memory Cost. CoRR, abs/1604.06174, 2016. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 392, + 298, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 392, + 298, + 622 + ], + "spans": [ + { + "bbox": [ + 61, + 392, + 298, + 622 + ], + "type": "text", + "content": "[4] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. PaLM: Scaling Language Modeling with Pathways. Journal of Machine Learning Research, 24:240:1-240:113, 2023. 17" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 624, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 624, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 61, + 624, + 296, + 689 + ], + "type": "text", + "content": "[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In *The North American Chapter of the Association for Computational Linguistics: Human Language Technologies*, pages 4171–4186, 2019. 6, 14" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 691, + 296, + 713 + ], + "type": "text", + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 556, + 713 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 128 + ], + "type": "text", + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In International Conference on Learning Representations, 2021. 6, 14" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 320, + 129, + 556, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 129, + 556, + 239 + ], + "spans": [ + { + "bbox": [ + 320, + 129, + 556, + 239 + ], + "type": "text", + "content": "[7] Nan Du, Yanping Huang, Andrew M. Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yanqi Zhou, Adams Wei Yu, Orhan Firat, Barret Zoph, Liam Fedus, Maarten P. Bosma, Zongwei Zhou, Tao Wang, Yu Emma Wang, Kellie Webster, Marie Pellat, Kevin Robinson, Kathleen S. Meier-Hellstern, Toju Duke, Lucas Dixon, Kun Zhang, Quoc V. Le, Yonghui Wu, Zhifeng Chen, and Claire Cui. GLaM: Efficient Scaling of Language Models with Mixture- of-Experts. In International Conference on Machine Learning, pages 5547-5569, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 240, + 555, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 240, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 320, + 240, + 555, + 282 + ], + "type": "text", + "content": "[8] FairScale authors. FairScale: A general purpose modular PyTorch library for high performance and large scale training. https://github.com/facebookresearch/fairscale, 2021. 1, 2, 6, 14" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 285, + 555, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 285, + 555, + 351 + ], + "spans": [ + { + "bbox": [ + 320, + 285, + 555, + 351 + ], + "type": "text", + "content": "[9] Shiqing Fan, Yi Rong, Chen Meng, Zongyan Cao, Siyu Wang, Zhen Zheng, Chuan Wu, Guoping Long, Jun Yang, Lixue Xia, Lansong Diao, Xiaoyong Liu, and Wei Lin. DAPPLE: A Pipelined Data Parallel Approach for Training Large Models. In Symposium on Principles and Practice of Parallel Programming, pages 431-445, 2021. 1, 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 353, + 555, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 353, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 353, + 555, + 396 + ], + "type": "text", + "content": "[10] William Fedus, Barret Zoph, and Noam Shazeer. Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity. The Journal of Machine Learning Research, 23(120):5232-5270, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 397, + 555, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 397, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 315, + 397, + 555, + 419 + ], + "type": "text", + "content": "[11] Michael J. Flynn. Very High-speed Computing Systems. Proceedings of the IEEE, 54(12):1901-1909, 1966. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 421, + 554, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 421, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 315, + 421, + 554, + 453 + ], + "type": "text", + "content": "[12] Michael J. Flynn. Some Computer Organizations and Their Effectiveness. IEEE Transactions on Computers, C-21(9): 948–960, 1972. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 455, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 455, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 315, + 455, + 553, + 476 + ], + "type": "text", + "content": "[13] Gurobi Optimization, LLC. Gurobi Optimizer Reference Manual. https://www.gurobi.com, 2023. 6, 14" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 478, + 555, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 555, + 531 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 555, + 531 + ], + "type": "text", + "content": "[14] Chaoyang He, Shen Li, Mahdi Soltanolkotabi, and Salman Avestimehr. PipeTransformer: Automated Elastic Pipelining for Distributed Training of Large-scale Models. In International Conference on Machine Learning, pages 4150-4159, 2021. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 534, + 555, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 555, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 555, + 600 + ], + "type": "text", + "content": "[15] Yanping Huang, Youlong Cheng, Ankur Bapna, Orhan First, Dehao Chen, Mia Xu Chen, HyoukJoong Lee, Jiquan Ngiam, Quoc V. Le, Yonghui Wu, and Zhifeng Chen. GPipe: Efficient Training of Giant Neural Networks using Pipeline Parallelism. In Advances in Neural Information Processing Systems 32, pages 103-112, 2019. 1, 2, 4, 6, 14" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 601, + 555, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 601, + 555, + 644 + ], + "spans": [ + { + "bbox": [ + 315, + 601, + 555, + 644 + ], + "type": "text", + "content": "[16] Zhihao Jia, Sina Lin, Charles R. Qi, and Alex Aiken. Exploring Hidden Dimensions in Parallelizing Convolutional Neural Networks. In International Conference on Machine Learning, pages 2279-2288, 2018. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 647, + 555, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 647, + 555, + 679 + ], + "spans": [ + { + "bbox": [ + 315, + 647, + 555, + 679 + ], + "type": "text", + "content": "[17] Zhihao Jia, Matei Zaharia, and Alex Aiken. Beyond Data and Model Parallelism for Deep Neural Networks. In Machine Learning and Systems, pages 1-13, 2019. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[18] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. In International Conference on Learning Representations, 2015. 14" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20955" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 95 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 95 + ], + "type": "text", + "content": "[19] Rafael Lazimy. Mixed-integer Quadratic Programming. Math. Program., 22(1):332-349, 1982. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 297, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 297, + 162 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 297, + 162 + ], + "type": "text", + "content": "[20] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. GShard: Scaling Giant Models with Conditional Computation and Automatic Sharding. In International Conference on Learning Representations, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 217 + ], + "type": "text", + "content": "[21] Shigang Li and Torsten Hoefler. Chimera: Efficiently Training Large-Scale Neural Networks with Bidirectional Pipelines. In International Conference for High Performance Computing, Networking, Storage and Analysis, pages 27:1-27:14, 2021. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 297, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 297, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 297, + 285 + ], + "type": "text", + "content": "[22] Shen Li, Yanli Zhao, Rohan Varma, Omkar Salpekar, Pieter Noordhuis, Teng Li, Adam Paszke, Jeff Smith, Brian Vaughan, Pritam Damania, and Soumith Chintala. PyTorch Distributed: Experiences on Accelerating Data Parallel Training. Proceedings of the VLDB Endowment, 13(12):3005-3018, 2020. 1, 2, 6, 14" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 286, + 296, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 286, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 57, + 286, + 296, + 330 + ], + "type": "text", + "content": "[23] Yuliang Liu, Shenggui Li, Jiarui Fang, Yanjun Shao, Boyuan Yao, and Yang You. Colossal-Auto: Unified Automation of Parallelization and Activation Checkpoint for Large-scale Models. CoRR, abs/2302.02599, 2023. 1, 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 297, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 297, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 297, + 386 + ], + "type": "text", + "content": "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. In International Conference on Computer Vision, pages 9992-10002, 2021. 6, 14" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 297, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 297, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 297, + 442 + ], + "type": "text", + "content": "[25] Xupeng Miao, Yujie Wang, Youhe Jiang, Chunan Shi, Xiaonan Nie, Hailin Zhang, and Bin Cui. Galvatron: Efficient Transformer Training over Multiple GPUs Using Automatic Parallelism. Proceedings of the VLDB Endowment, 16(3): 470-479, 2022. 2, 3, 6, 15" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 445, + 297, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 297, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 297, + 499 + ], + "type": "text", + "content": "[26] Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory F. Diamos, Erich Elsen, David García, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, and Hao Wu. Mixed Precision Training. In International Conference on Learning Representations, 2018. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 501, + 297, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 297, + 522 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 297, + 522 + ], + "type": "text", + "content": "[27] Microsoft. Megatron-DeepSpeed. https://github.com/microsoft/Megatron-DeepSpeed/tree/main, 2021.6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 524, + 296, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 524, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 56, + 524, + 296, + 579 + ], + "type": "text", + "content": "[28] Deepak Narayanan, Aaron Harlap, Amar Phanishayee, Vivek Seshadri, Nikhil R. Devanur, Gregory R. Ganger, Phillip B. Gibbons, and Matei Zaharia. PipeDream: Generalized Pipeline Parallelism for DNN Training. In Symposium on Operating Systems Principles, pages 1-15, 2019. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 581, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 296, + 624 + ], + "type": "text", + "content": "[29] Deepak Narayanan, Amar Phanishayee, Kaiyu Shi, Xie Chen, and Matei Zaharia. Memory-Efficient Pipeline-Parallel DNN Training. In International Conference on Machine Learning, pages 7937–7947, 2021. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 297, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 297, + 713 + ], + "type": "text", + "content": "[30] Deepak Narayanan, Mohammad Shoeybi, Jared Casper, Patrick LeGresley, Mostofa Patwary, Vijay Korthikanti, Dmitri Vainbrand, Prethvi Kashinkunti, Julie Bernauer, Bryan Catanzaro, Amar Phanishayee, and Matei Zaharia. Efficient Large-scale Language Model Training on GPU Clusters Using Megatron-LM. In International Conference for High Performance Computing, Networking, Storage and Analysis, pages 58:1-58:15, 2021. 1, 2, 3, 6, 14" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 556, + 713 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 315, + 73, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 555, + 128 + ], + "type": "text", + "content": "[31] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 6, 14" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 555, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 185 + ], + "type": "text", + "content": "[32] Samyam Rajbhandari, Jeff Rasley, Olatunjri Ruwase, and Yuxiong He. ZeRO: Memory Optimizations Toward Training Trillion Parameter Models. In International Conference for High Performance Computing, Networking, Storage and Analysis, pages 20:1-20:16, 2020. 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 186, + 556, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 556, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 556, + 251 + ], + "type": "text", + "content": "[33] Saeed Rashidi, Matthew Denton, Srinivas Sridharan, Sudarshan Srinivasan, Amoghavarsha Suresh, Jade Nie, and Tushar Krishna. Enabling Compute-Communication Overlap in Distributed Deep Learning Training Platforms. In International Symposium on Computer Architecture, pages 540–553, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 253, + 555, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 555, + 308 + ], + "type": "text", + "content": "[34] Jeff Rasley, Samyam Rajbhandari, Olatunjri Ruwase, and Yuxiong He. DeepSpeed: System Optimizations Enable Training Deep Learning Models with Over 100 Billion Parameters. In The 26th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 3505-3506, 2020. 1, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 309, + 555, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 555, + 374 + ], + "type": "text", + "content": "[35] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision, 115(3):211-252, 2015. 14" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 376, + 555, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 376, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 317, + 376, + 555, + 441 + ], + "type": "text", + "content": "[36] Michael Schaarschmidt, Dominik Grewe, Dimitrios Vytiniotis, Adam Paszke, Georg Stefan Schmid, Tamara Norman, James Molloy, Jonathan Godwin, Norman Alexander Rink, Vinod Nair, and Dan Belov. Automap: Towards Ergonomic Automated Parallelism for ML Models. CoRR, abs/2112.02958, 2021. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 444, + 555, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 444, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 444, + 555, + 510 + ], + "type": "text", + "content": "[37] Noam Shazeer, Youlong Cheng, Niki Parmar, Dustin Tran, Ashish Vaswani, Penporn Koanantakool, Peter Hawkins, HyoukJoong Lee, Mingsheng Hong, Cliff Young, Ryan Sepassi, and Blake A. Hechtman. Mesh-TensorFlow: Deep Learning for Supercomputers. In Advances in Neural Information Processing Systems 31, pages 10435-10444, 2018. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 512, + 555, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 555, + 565 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 555, + 565 + ], + "type": "text", + "content": "[38] Jakub Tarnawski, Amar Phanishayee, Nikhil R. Devanur, Divya Mahajan, and Fanny Nina Paravecino. Efficient Algorithms for Device Placement of DNN Graph Operators. In Advances in Neural Information Processing Systems 33, pages 15451-15463, 2020. 1, 2, 4, 12" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 567, + 555, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 567, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 567, + 555, + 612 + ], + "type": "text", + "content": "[39] Jakub Tarnawski, Deepak Narayanan, and Amar Phanishayee. Piper: Multidimensional Planner for DNN Parallelization. In Advances in Neural Information Processing Systems 34, pages 24829-24840, 2021. 1, 2, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 613, + 555, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 555, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 555, + 678 + ], + "type": "text", + "content": "[40] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurélien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. LLaMA: Open and Efficient Foundation Language Models. CoRR, abs/2302.13971, 2023. 1, 6, 7, 14" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "type": "text", + "content": "[41] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "20956" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 615 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 76, + 72, + 297, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 297, + 258 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 297, + 258 + ], + "type": "text", + "content": "Blecher, Cristian Canton-Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, MarieAnne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open Foundation and Fine-Tuned Chat Models. CoRR, abs/2307.09288, 2023. 1, 6, 7, 14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 260, + 297, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 260, + 297, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 260, + 297, + 304 + ], + "type": "text", + "content": "[42] Minjie Wang, Chien-Chin Huang, and Jinyang Li. Supporting Very Large Models using Automatic Dataflow Graph Partitioning. In Proceedings of the Fourteenth EuroSys Conference, pages 26:1-26:17, 2019. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 304, + 296, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 296, + 327 + ], + "type": "text", + "content": "[43] Wikipedia Foundation. Wikipedia Downloads. https:// dumps.wikipedia.org, 2023. 14" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 328, + 297, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 328, + 297, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 328, + 297, + 403 + ], + "type": "text", + "content": "[44] Yuanzhong Xu, HyoukJoong Lee, Dehao Chen, Blake A. Hechtman, Yanping Huang, Rahul Joshi, Maxim Krikun, Dmitry Lepikhin, Andy Ly, Marcello Maggioni, Ruoming Pang, Noam Shazeer, Shibo Wang, Tao Wang, Yonghui Wu, and Zhifeng Chen. GSPMD: General and Scalable Parallelization for ML Computation Graphs. CoRR, abs/2105.04663, 2021. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 406, + 297, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 406, + 297, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 406, + 297, + 460 + ], + "type": "text", + "content": "[45] Binhang Yuan, Yongjun He, Jared Davis, Tianyi Zhang, Tri Dao, Beidi Chen, Percy Liang, Christopher Ré, and Ce Zhang. Decentralized Training of Foundation Models in Heterogeneous Environments. In Advances in Neural Information Processing Systems 35, pages 25464-25477, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 461, + 297, + 537 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 461, + 297, + 537 + ], + "spans": [ + { + "bbox": [ + 56, + 461, + 297, + 537 + ], + "type": "text", + "content": "[46] Shixiong Zhao, Fanxin Li, Xusheng Chen, Xiuxian Guan, Jianyu Jiang, Dong Huang, Yuhao Qing, Sen Wang, Peng Wang, Gong Zhang, Cheng Li, Ping Luo, and Heming Cui. vPipe: A Virtualized Acceleration System for Achieving Efficient and Scalable Pipeline Parallel DNN Training. IEEE Transactions on Parallel and Distributed Systems, 33(3):489-506, 2022. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 539, + 297, + 615 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 539, + 297, + 615 + ], + "spans": [ + { + "bbox": [ + 56, + 539, + 297, + 615 + ], + "type": "text", + "content": "[47] Lianmin Zheng, Zhuohan Li, Hao Zhang, Yonghao Zhuang, Zhifeng Chen, Yanping Huang, Yida Wang, Yuanzhong Xu, Danyang Zhuo, Eric P. Xing, Joseph E. Gonzalez, and Ion Stoica. Alpa: Automating Inter- and Intra-Operator Parallelism for Distributed Deep Learning. In USENIX Symposium on Operating Systems Design and Implementation, pages 559–578, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20957" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file